diff --git a/.gitattributes b/.gitattributes index a39a0f9f30cfd545bc514a35a906149ebd0959a9..fb5e0401171bc15356788136a132a56fba7bb5ba 100644 --- a/.gitattributes +++ b/.gitattributes @@ -7655,3 +7655,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/Structural[[:space:]]Multiplane[[:space:]]Image_[[:space:]]Bridging[[:space:]]Neural[[:space:]]View[[:space:]]Synthesis[[:space:]]and[[:space:]]3D[[:space:]]Reconstruction/d01abaa5-0135-433b-83fa-882e39a28acd_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Structure[[:space:]]Aggregation[[:space:]]for[[:space:]]Cross-Spectral[[:space:]]Stereo[[:space:]]Image[[:space:]]Guided[[:space:]]Denoising/f81a83a0-2faa-4877-8a8a-0a610504614c_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Structured[[:space:]]3D[[:space:]]Features[[:space:]]for[[:space:]]Reconstructing[[:space:]]Controllable[[:space:]]Avatars/e451a1ab-dc78-421c-b81e-748a81f9e9ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Structured[[:space:]]Kernel[[:space:]]Estimation[[:space:]]for[[:space:]]Photon-Limited[[:space:]]Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Structured[[:space:]]Sparsity[[:space:]]Learning[[:space:]]for[[:space:]]Efficient[[:space:]]Video[[:space:]]Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyLess_[[:space:]]Boosting[[:space:]]the[[:space:]]Transferability[[:space:]]of[[:space:]]Adversarial[[:space:]]Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Style[[:space:]]Projected[[:space:]]Clustering[[:space:]]for[[:space:]]Domain[[:space:]]Generalized[[:space:]]Semantic[[:space:]]Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleAdv_[[:space:]]Meta[[:space:]]Style[[:space:]]Adversarial[[:space:]]Training[[:space:]]for[[:space:]]Cross-Domain[[:space:]]Few-Shot[[:space:]]Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleGAN[[:space:]]Salon_[[:space:]]Multi-View[[:space:]]Latent[[:space:]]Optimization[[:space:]]for[[:space:]]Pose-Invariant[[:space:]]Hairstyle[[:space:]]Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleGene_[[:space:]]Crossover[[:space:]]and[[:space:]]Mutation[[:space:]]of[[:space:]]Region-Level[[:space:]]Facial[[:space:]]Genes[[:space:]]for[[:space:]]Kinship[[:space:]]Face[[:space:]]Synthesis/88326347-639e-4732-90cc-676114c49600_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleIPSB_[[:space:]]Identity-Preserving[[:space:]]Semantic[[:space:]]Basis[[:space:]]of[[:space:]]StyleGAN[[:space:]]for[[:space:]]High[[:space:]]Fidelity[[:space:]]Face[[:space:]]Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleRF_[[:space:]]Zero-Shot[[:space:]]3D[[:space:]]Style[[:space:]]Transfer[[:space:]]of[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleRes_[[:space:]]Transforming[[:space:]]the[[:space:]]Residuals[[:space:]]for[[:space:]]Real[[:space:]]Image[[:space:]]Editing[[:space:]]With[[:space:]]StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleSync_[[:space:]]High-Fidelity[[:space:]]Generalized[[:space:]]and[[:space:]]Personalized[[:space:]]Lip[[:space:]]Sync[[:space:]]in[[:space:]]Style-Based[[:space:]]Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SunStage_[[:space:]]Portrait[[:space:]]Reconstruction[[:space:]]and[[:space:]]Relighting[[:space:]]Using[[:space:]]the[[:space:]]Sun[[:space:]]as[[:space:]]a[[:space:]]Light[[:space:]]Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Super-CLEVR_[[:space:]]A[[:space:]]Virtual[[:space:]]Benchmark[[:space:]]To[[:space:]]Diagnose[[:space:]]Domain[[:space:]]Robustness[[:space:]]in[[:space:]]Visual[[:space:]]Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Super-Resolution[[:space:]]Neural[[:space:]]Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SuperDisco_[[:space:]]Super-Class[[:space:]]Discovery[[:space:]]Improves[[:space:]]Visual[[:space:]]Recognition[[:space:]]for[[:space:]]the[[:space:]]Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Superclass[[:space:]]Learning[[:space:]]With[[:space:]]Representation[[:space:]]Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Supervised[[:space:]]Masked[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]for[[:space:]]Few-Shot[[:space:]]Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SurfelNeRF_[[:space:]]Neural[[:space:]]Surfel[[:space:]]Radiance[[:space:]]Fields[[:space:]]for[[:space:]]Online[[:space:]]Photorealistic[[:space:]]Reconstruction[[:space:]]of[[:space:]]Indoor[[:space:]]Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Swept-Angle[[:space:]]Synthetic[[:space:]]Wavelength[[:space:]]Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Switchable[[:space:]]Representation[[:space:]]Learning[[:space:]]Framework[[:space:]]With[[:space:]]Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Symmetric[[:space:]]Shape-Preserving[[:space:]]Autoencoder[[:space:]]for[[:space:]]Unsupervised[[:space:]]Real[[:space:]]Scene[[:space:]]Point[[:space:]]Cloud[[:space:]]Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SynthVSR_[[:space:]]Scaling[[:space:]]Up[[:space:]]Visual[[:space:]]Speech[[:space:]]Recognition[[:space:]]With[[:space:]]Synthetic[[:space:]]Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Synthesizing[[:space:]]Photorealistic[[:space:]]Virtual[[:space:]]Humans[[:space:]]Through[[:space:]]Cross-Modal[[:space:]]Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/System-Status-Aware[[:space:]]Adaptive[[:space:]]Network[[:space:]]for[[:space:]]Online[[:space:]]Streaming[[:space:]]Video[[:space:]]Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/T-SEA_[[:space:]]Transfer-Based[[:space:]]Self-Ensemble[[:space:]]Attack[[:space:]]on[[:space:]]Object[[:space:]]Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TAPS3D_[[:space:]]Text-Guided[[:space:]]3D[[:space:]]Textured[[:space:]]Shape[[:space:]]Generation[[:space:]]From[[:space:]]Pseudo[[:space:]]Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TBP-Former_[[:space:]]Learning[[:space:]]Temporal[[:space:]]Bird's-Eye-View[[:space:]]Pyramid[[:space:]]for[[:space:]]Joint[[:space:]]Perception[[:space:]]and[[:space:]]Prediction[[:space:]]in[[:space:]]Vision-Centric[[:space:]]Autonomous[[:space:]]Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TINC_[[:space:]]Tree-Structured[[:space:]]Implicit[[:space:]]Neural[[:space:]]Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TIPI_[[:space:]]Test[[:space:]]Time[[:space:]]Adaptation[[:space:]]With[[:space:]]Transformation[[:space:]]Invariance/5131537e-5515-4827-8458-d96e395e71bb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TMO_[[:space:]]Textured[[:space:]]Mesh[[:space:]]Acquisition[[:space:]]of[[:space:]]Objects[[:space:]]With[[:space:]]a[[:space:]]Mobile[[:space:]]Device[[:space:]]by[[:space:]]Using[[:space:]]Differentiable[[:space:]]Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TOPLight_[[:space:]]Lightweight[[:space:]]Neural[[:space:]]Networks[[:space:]]With[[:space:]]Task-Oriented[[:space:]]Pretraining[[:space:]]for[[:space:]]Visible-Infrared[[:space:]]Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TRACE_[[:space:]]5D[[:space:]]Temporal[[:space:]]Regression[[:space:]]of[[:space:]]Avatars[[:space:]]With[[:space:]]Dynamic[[:space:]]Cameras[[:space:]]in[[:space:]]3D[[:space:]]Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TTA-COPE_[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Category-Level[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TWINS_[[:space:]]A[[:space:]]Fine-Tuning[[:space:]]Framework[[:space:]]for[[:space:]]Improved[[:space:]]Transferability[[:space:]]of[[:space:]]Adversarial[[:space:]]Robustness[[:space:]]and[[:space:]]Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Taming[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Audio-Driven[[:space:]]Co-Speech[[:space:]]Gesture[[:space:]]Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tangentially[[:space:]]Elongated[[:space:]]Gaussian[[:space:]]Belief[[:space:]]Propagation[[:space:]]for[[:space:]]Event-Based[[:space:]]Incremental[[:space:]]Optical[[:space:]]Flow[[:space:]]Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TarViS_[[:space:]]A[[:space:]]Unified[[:space:]]Approach[[:space:]]for[[:space:]]Target-Based[[:space:]]Video[[:space:]]Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Target-Referenced[[:space:]]Reactive[[:space:]]Grasping[[:space:]]for[[:space:]]Dynamic[[:space:]]Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Task[[:space:]]Difficulty[[:space:]]Aware[[:space:]]Parameter[[:space:]]Allocation[[:space:]]&[[:space:]]Regularization[[:space:]]for[[:space:]]Lifelong[[:space:]]Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Task[[:space:]]Residual[[:space:]]for[[:space:]]Tuning[[:space:]]Vision-Language[[:space:]]Models/7e947285-2912-4e13-91da-b9b910f110fa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Task-Specific[[:space:]]Fine-Tuning[[:space:]]via[[:space:]]Variational[[:space:]]Information[[:space:]]Bottleneck[[:space:]]for[[:space:]]Weakly-Supervised[[:space:]]Pathology[[:space:]]Whole[[:space:]]Slide[[:space:]]Image[[:space:]]Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TeSLA_[[:space:]]Test-Time[[:space:]]Self-Learning[[:space:]]With[[:space:]]Automatic[[:space:]]Adversarial[[:space:]]Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Teacher-Generated[[:space:]]Spatial-Attention[[:space:]]Labels[[:space:]]Boost[[:space:]]Robustness[[:space:]]and[[:space:]]Accuracy[[:space:]]of[[:space:]]Contrastive[[:space:]]Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Teaching[[:space:]]Matters_[[:space:]]Investigating[[:space:]]the[[:space:]]Role[[:space:]]of[[:space:]]Supervision[[:space:]]in[[:space:]]Vision[[:space:]]Transformers/25874b86-4464-466c-bbd4-91f62653cf36_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Teaching[[:space:]]Structured[[:space:]]Vision[[:space:]]&[[:space:]]Language[[:space:]]Concepts[[:space:]]to[[:space:]]Vision[[:space:]]&[[:space:]]Language[[:space:]]Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Teleidoscopic[[:space:]]Imaging[[:space:]]System[[:space:]]for[[:space:]]Microscale[[:space:]]3D[[:space:]]Shape[[:space:]]Reconstruction/d72f0f54-2daa-4009-a052-255877447821_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tell[[:space:]]Me[[:space:]]What[[:space:]]Happened_[[:space:]]Unifying[[:space:]]Text-Guided[[:space:]]Video[[:space:]]Completion[[:space:]]via[[:space:]]Multimodal[[:space:]]Masked[[:space:]]Video[[:space:]]Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TempSAL[[:space:]]-[[:space:]]Uncovering[[:space:]]Temporal[[:space:]]Information[[:space:]]for[[:space:]]Deep[[:space:]]Saliency[[:space:]]Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporal[[:space:]]Attention[[:space:]]Unit_[[:space:]]Towards[[:space:]]Efficient[[:space:]]Spatiotemporal[[:space:]]Predictive[[:space:]]Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporal[[:space:]]Consistent[[:space:]]3D[[:space:]]LiDAR[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Semantic[[:space:]]Perception[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporal[[:space:]]Interpolation[[:space:]]Is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]for[[:space:]]Dynamic[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/22a9031d-e022-4324-b56f-e7ab749525c7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporally[[:space:]]Consistent[[:space:]]Online[[:space:]]Depth[[:space:]]Estimation[[:space:]]Using[[:space:]]Point-Based[[:space:]]Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TensoIR_[[:space:]]Tensorial[[:space:]]Inverse[[:space:]]Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tensor4D_[[:space:]]Efficient[[:space:]]Neural[[:space:]]4D[[:space:]]Decomposition[[:space:]]for[[:space:]]High-Fidelity[[:space:]]Dynamic[[:space:]]Reconstruction[[:space:]]and[[:space:]]Rendering/eeed7500-cf04-43cc-8f38-007884849085_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Test[[:space:]]Time[[:space:]]Adaptation[[:space:]]With[[:space:]]Regularized[[:space:]]Loss[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Salient[[:space:]]Object[[:space:]]Detection/2441aa83-753b-4443-ba38-3993bb6d5100_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Test[[:space:]]of[[:space:]]Time_[[:space:]]Instilling[[:space:]]Video-Language[[:space:]]Models[[:space:]]With[[:space:]]a[[:space:]]Sense[[:space:]]of[[:space:]]Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TexPose_[[:space:]]Neural[[:space:]]Texture[[:space:]]Learning[[:space:]]for[[:space:]]Self-Supervised[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text[[:space:]]With[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Augmented[[:space:]]Transformer[[:space:]]for[[:space:]]Video[[:space:]]Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text-Guided[[:space:]]Unsupervised[[:space:]]Latent[[:space:]]Transformation[[:space:]]for[[:space:]]Multi-Attribute[[:space:]]Image[[:space:]]Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text-Visual[[:space:]]Prompting[[:space:]]for[[:space:]]Efficient[[:space:]]2D[[:space:]]Temporal[[:space:]]Video[[:space:]]Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text2Scene_[[:space:]]Text-Driven[[:space:]]Indoor[[:space:]]Scene[[:space:]]Stylization[[:space:]]With[[:space:]]Part-Aware[[:space:]]Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Texts[[:space:]]as[[:space:]]Images[[:space:]]in[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Multi-Label[[:space:]]Image[[:space:]]Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Texture-Guided[[:space:]]Saliency[[:space:]]Distilling[[:space:]]for[[:space:]]Unsupervised[[:space:]]Salient[[:space:]]Object[[:space:]]Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Best[[:space:]]Defense[[:space:]]Is[[:space:]]a[[:space:]]Good[[:space:]]Offense_[[:space:]]Adversarial[[:space:]]Augmentation[[:space:]]Against[[:space:]]Adversarial[[:space:]]Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_content_list.json b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..340f38a16331fc390b60e1125862c47b6fb558c8 --- /dev/null +++ b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_content_list.json @@ -0,0 +1,1671 @@ +[ + { + "type": "text", + "text": "Structured Kernel Estimation for Photon-Limited Deconvolution", + "text_level": 1, + "bbox": [ + 156, + 130, + 810, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yash Sanghvi, Zhiyuan Mao, Stanley H. Chan \nSchool of Electrical and Computer Engineering, Purdue University \n{ysanghvi, maol14, stanchan}@purdue.edu", + "bbox": [ + 218, + 181, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images taken in a low light condition with the presence of camera shake suffer from motion blur and photon shot noise. While state-of-the-art image restoration networks show promising results, they are largely limited to well-illuminated scenes and their performance drops significantly when photon shot noise is strong.", + "bbox": [ + 75, + 300, + 470, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a new blur estimation technique customized for photon-limited conditions. The proposed method employs a gradient-based backpropagation method to estimate the blur kernel. By modeling the blur kernel using a low-dimensional representation with the key points on the motion trajectory, we significantly reduce the search space and improve the regularity of the kernel estimation problem. When plugged into an iterative framework, our novel low-dimensional representation provides improved kernel estimates and hence significantly better deconvolution performance when compared to end-to-end trained neural networks. The source code and pretrained mdoels are available at https://github.com/sanghviyashiitb/structured-kernel-cvpr23", + "bbox": [ + 75, + 391, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 645, + 209, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photon-Limited Blind Deconvolution: This paper studies the photon-limited blind deconvolution problem. Blind deconvolution refers to simultaneously recovering both the blur kernel and latent clean image from a blurred image and \"photon-limited\" refers to presence of photon-shot noise in images taken in low-illumination / short exposure. The corresponding forward model is as follows:", + "bbox": [ + 75, + 669, + 468, + 776 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\operatorname {P o i s s o n} (\\alpha \\mathbf {h} * \\mathbf {x}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 784, + 468, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this equation, $\\mathbf{y} \\in \\mathbb{R}^N$ is the blurred-noisy image, $\\mathbf{x} \\in \\mathbb{R}^N$ is the latent clean image, and $\\mathbf{h} \\in \\mathbb{R}^M$ is the blur kernel. We assume that $\\mathbf{x}$ is normalized to $[0,1]$ and the entries of $\\mathbf{h}$ are non-negative and sum up to 1. The constant $\\alpha$ represents the average number of photons per pixel and is inversely proportional to the amount of Poisson noise.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Iterative Kernel Estimation: Blind image deconvolution has been studied for decades with many successful algorithms including the latest deep neural networks [8, 24, 34, 42, 43]. Arguably, the adaptation from the traditional Gaussian noise model to the photon-limited Poisson noise model can be done by retraining the existing networks with appropriate data. However, the restoration is not guaranteed to perform well because the end-to-end networks seldom explicitly take the forward image formation model into account.", + "bbox": [ + 496, + 268, + 893, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, people have started to recognize the importance of blur kernel estimation for photon-limited conditions. One of these works is by Sanghvi et. al [30], where they propose an iterative kernel estimation method to backpropagate the gradient of an unsupervised reburring function, hence to update the blur kernel. However, as we can see in Figure 1, their performance is still limited when the photon shot noise is strong.", + "bbox": [ + 496, + 424, + 893, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Structured Kernel Estimation: Inspired by [30], we believe that the iterative kernel estimation process and the unsupervised reburbling loss are useful. However, instead of searching for the kernel directly (which can easily lead to local minima because the search space is too big), we propose to search in a low-dimensional space by imposing structure to the motion blur kernel.", + "bbox": [ + 496, + 547, + 893, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To construct such a low-dimensional space, we frame the blur kernel in terms of trajectory of the camera motion. Motion trajectory is often a continuous but irregular path in the two-dimensional plane. To specify the trajectory, we introduce the concept of key point estimation where we identify a set of anchor points of the kernel. By interpolating the path along these anchor points, we can then reproduce the kernel. Since the number of anchor points is significantly lower than the number of pixels in a kernel, we can reduce the dimensionality of the kernel estimation problem.", + "bbox": [ + 496, + 656, + 893, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key contribution of this paper is as follows: We propose a new kernel estimation method called Kernel Trajectory Network (KTN). KTN models the blur kernel in a low-dimensional and differentiable space by specifying key points of the motion trajectory. Plugging this low-dimensional representation in an iterative framework im", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5079f0c6931497d4b8bd3e86b22c2fe6e199338969df4cebf626e158736c8ec6.jpg", + "image_caption": [ + "Blurred and Noisy" + ], + "image_footnote": [], + "bbox": [ + 88, + 87, + 246, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b07d992d729dcf2e550b4bcfe3926aedbba947de08b7e374eeb6be4ce4953268.jpg", + "image_caption": [ + "MPR-Net [42]" + ], + "image_footnote": [], + "bbox": [ + 246, + 87, + 403, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5940cc1d5f81fe8a449c784dfa1e71bab26ee5d425f9cb2aa5d1805b47700350.jpg", + "image_caption": [ + "Sanghvi et. al [30]", + "Figure 1. The proposed Kernel Trajectory Network (KTN) on real noisy blurred image from Photon-Limited Deblurring Dataset (PLDD) [29] The result corresponding to MPR-Net was generated by retraining the network with GoPro dataset [24] corrupted by Poisson noise. The inset images for \"Sanghvi et. al\" and \"Ours\" represent the estimated kernel and the inset image for \"Ground-Truth\" represents the kernel captured using a point source, as provided in PLDD." + ], + "image_footnote": [], + "bbox": [ + 406, + 87, + 563, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/916b6af1dccd3a5c9a2a7e6a6a12c066e94ee6b9b71952b48b40eb136d5061c0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 563, + 87, + 720, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22178167ccf11a52cb61ffb0adf6436f6c94fe55393547dfb3246b9fd5da9016.jpg", + "image_caption": [ + "Ground-Truth" + ], + "image_footnote": [], + "bbox": [ + 722, + 87, + 880, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proves the regularity of the kernel estimation problem. This leads to substantially better blur kernel estimates in photon-limited regimes where existing methods fail.", + "bbox": [ + 75, + 321, + 468, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 386, + 218, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Blind Deconvolution: Classical approaches to the (noiseless) blind deconvolution problem [6, 7, 22, 32, 40] use a joint optimization framework in which both the kernel and image are updated in an alternating fashion in order to minimize a cost function with kernel and image priors. For high noise regimes, a combination of $\\ell_1 + \\mathrm{TV}$ prior has been used in [2]. Levin et. al [16] pointed out that this joint optimization framework for the blind deconvolution problem favours the no-blur degenerate solution i.e. $(\\mathbf{x}^{*},\\mathbf{h}^{*}) = (\\mathbf{y},\\mathbf{I})$ where $\\mathbf{I}$ is the identity operator. Some methods model the blur kernel in terms of the camera trajectory and then recover both the trajectory and the clean image using optimization [12, 38, 39] and supervised-learning techniques [11, 33, 46].", + "bbox": [ + 75, + 412, + 468, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the non-blind case, i.e., when the blur kernel is assumed to be known, the Poisson deconvolution problem has been studied for decades starting from Richardson-Lucy algorithm [21, 26]. More contemporary methods include Plug-and-Play [28, 29], PURE-LET [18], and MAP-based optimization methods [10, 13].", + "bbox": [ + 75, + 627, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Learning Methods. Recent years, many deep learning-based methods [5, 31] have been proposed for the blind image deblurring task. The most common strategy is to train a network end-to-end on large-scale datasets, such as the GoPro [24] and the RealBlur [27] datasets. Notably, many recent works [8, 24, 34, 42, 43] improve the performance of deblurring networks by adopting the multiscale strategies, where the training follows a coarse-to-fine setting that resembles the iterative approach. Generative Adversarial Network (GAN) based deblurring methods [3,14,15,44] are also shown to produce visually appealing images. Zamir et al. [41] and Wang et al. [37] adapt the", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "popular vision transformers to the image restoration problems and demonstrate competitive performance on the deblurring task.", + "bbox": [ + 496, + 321, + 890, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Networks and Iterative Methods: While neural networks have shown state-of-the-art performance on the deblurring task, another class of methods incorporating iterative methods with deep learning have shown promising results. Algorithm unrolling [23], where an iterative method is unrolled for fixed iterations and trained end-to-end has been applied to image deblurring [1,19]. In SelfDeblur [25], authors use Deep-Image-Prior [35] to represent the image and blur kernel and obtain state-of-the-art blind deconvolution performance.", + "bbox": [ + 496, + 368, + 892, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 534, + 589, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Kernel as Structured Motion Estimation", + "text_level": 1, + "bbox": [ + 500, + 559, + 844, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera motion blur can be modeled as a latent clean image $\\mathbf{x}$ convolved with a blur kernel $\\mathbf{h}$ . If we assume the blur kernel lies in a window of size $32 \\times 32$ , then $\\mathbf{h} \\in \\mathbb{R}^{1024}$ . However, in this high dimensional space, only few entries of the blur kernel $\\mathbf{h}$ are non-zero. Additionally, the kernel is generated from a two-dimensional trajectory which suggests that a simple sparsity prior is not sufficient. Given the difficulty of the photon-limited deconvolution problem, we need to impose a stronger prior on the kernel. To this end, we propose a differentiable and low-dimensional representation of the blur kernel, which we will use as the search space in our kernel estimation algorithm.", + "bbox": [ + 496, + 583, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We take the two-dimensional trajectory of the camera during the exposure time and divide it into $K$ \"key points\". Each key point represents either the start, the end or a change in direction of the camera trajectory as seen in Figure 2. Given the $K$ key points as points mapped out in $x - y$ space, we can interpolate them using cubic splines to form a continuous trajectory in 2D. To convert this continuous trajectory to an equivalent blur kernel, we assume a point source image and move it through the given trajectory. The", + "bbox": [ + 496, + 765, + 892, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/06516385496c988531e9722e022d28179d445ffe9c2380da388f80478ecb0ecc.jpg", + "image_caption": [ + "Figure 2. Blur as Structured Motion Estimation: In our formulation, we view the blur kernel as the continuous camera trajectory reduced to $K$ key points, as shown in top half of the figure. We learn a differentiable representation from the vectorized $K$ key points to a blur kernel using a neural network. This lower dimensional and differentiable representation is leveraged to estimate a better blur kernel and avoiding local minima during inference." + ], + "image_footnote": [], + "bbox": [ + 71, + 80, + 477, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resulting frames are then averaged to give the corresponding blur kernel as shown in Figure 2.", + "bbox": [ + 75, + 518, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the formulation of blur kernel $\\mathbf{h}$ in terms of $K$ key points, we now need to put this representation to a differentiable form since we intend to use it in an iterative scheme. To achieve this, we learn the transformations from the key points to the blur kernels using a neural network, which will be referred to as Kernel-Trajectory Network (KTN), and represent it using a differentiable function $T(\\cdot)$ . Why differentiability is important to us will become clear to the reader in the next subsection.", + "bbox": [ + 75, + 549, + 468, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the Kernel-Trajectory Network, we generate training data as follows. First, for a fixed $K$ , we get $K$ key points by starting from $(0,0)$ and choosing the next $K - 1$ points by successively adding a random vector with a uniformly chosen random direction i.e. $U[0,360]$ and uniformly chosen length from $U[0,100 / (K - 1)]$ . Next, the set of key points are converted to a continuous smooth trajectory using bicubic interpolation. Then, we move a point source image through the given trajectory using the warpPerspective function in OpenCV, and average the resulting frames.", + "bbox": [ + 75, + 686, + 468, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the process defined above, we generate 60,000 blur kernels and their corresponding key point representations. For the Kernel-Trajectory Network $T(\\cdot)$ , we take a", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "U-Net like network with the first half replaced by 3-fully connected layers and train it with the generated data using $\\ell_2$ -loss. For further architectural details on the KTN, we refer the reader to the supplementary document.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Proposed Iterative Scheme", + "text_level": 1, + "bbox": [ + 500, + 157, + 741, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We described in the previous subsection how to obtain a low-dimensional and differentiable representation $T(\\cdot)$ for the blur kernel and now we are ready to present the full iterative scheme in detail. The proposed iterative scheme can be divided into three stages which are summarized as follows. We first generate an initial estimate of the direction and magnitude of the blur. This is used as initialization for a gradient-based scheme in Stage I which searches the appropriate kernel representation in the latent space $\\mathbf{z}$ . This is followed by Stage II where we fine-tune the kernel obtained from Stage I using a similar process.", + "bbox": [ + 496, + 181, + 890, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initialization Before starting the iterative scheme, we need a light-weight initialization method. This is important because of multiple local minima in the kernel estimation process.", + "bbox": [ + 496, + 348, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose to initialize the method with a rectilinear motion kernel, parameterized by length $\\rho$ and orientation $\\theta$ . To determine the length and orientation of the kernel, we use a minor variation of the kernel estimation in PolyBlur [9]. In this variation, the \"blur-only image\" $G(\\mathbf{y})$ is used as the input and $\\rho, \\theta$ for the initial kernel are estimated using the minimum of the directional gradients. We refer the reader to Section II in the supplementary document for further details on the initialization. Explanation on the \"blur-only image\" is provided when we describe Stage I of the scheme.", + "bbox": [ + 496, + 407, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage I: Kernel Estimation in Latent Space Given an initial kernel, we choose initial latent $\\mathbf{z}^0$ by dividing the rectilinear kernel into $K$ key points. Following the framework in [30], we run a gradient descent based scheme which optimizes the following cost function:", + "bbox": [ + 496, + 559, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {z}) \\stackrel {\\text {d e f}} {=} \\underbrace {\\| G (\\mathbf {y}) - \\mathbf {h} _ {\\mathbf {z}} * F (\\mathbf {y} , \\mathbf {h} _ {\\mathbf {z}})) \\| _ {2} ^ {2}} _ {\\text {R e b l u r i n g L o s s}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 640, + 890, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_{\\mathbf{z}} \\stackrel{\\mathrm{def}}{=} T(\\mathbf{z})$ represents the kernel output from Kernel-Trajectory network $T(.)$ given the vectorized key points representation $\\mathbf{z}$ . $F(.)$ represents the Poisson non-blind deconvolution solver which takes both noisy-blurred image and a blur kernel as the input. $G(\\mathbf{y})$ represents a denoiser which is trained to remove only the noise from noisy-blurred image. The overall cost function represents reburring loss i.e. how well the kernel estimate and corresponding image estimate $\\mathbf{h}_{\\mathbf{z}} * F(\\mathbf{y}, \\mathbf{h}_{\\mathbf{z}})$ match the blur-only image $G(\\mathbf{y})$ .", + "bbox": [ + 496, + 686, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To minimize the cost function in (2), we use a simple gradient descent based iterative update for $\\mathbf{z}$ as follows:", + "bbox": [ + 496, + 825, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {k + 1} = \\mathbf {z} ^ {k} - \\delta \\underbrace {\\nabla_ {\\mathbf {z}} \\mathcal {L} \\left(\\mathbf {z} ^ {k}\\right)} _ {\\text {b a c k p r o p a g a t i o n}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 859, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e571aa7677f0e24e0426bfbfdf5b4f78fcaf734bd154df6d5a5cabf0a45236d3.jpg", + "image_caption": [ + "Figure 3. Flowchart describing first stage of the proposed scheme. We estimate the motion kernel of the blurry noisy image in lower dimensional latent space $\\mathbf{z}$ where the blur kernel is represented by $T(\\mathbf{z})$ and by minimizing the reburring loss $\\mathcal{L}$ as defined in equation 2" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 816, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta > 0$ is the step size and $\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)$ represents the gradient of the cost function $\\mathcal{L}$ with respect to $\\mathbf{z}$ evaluated $\\mathbf{z}^k$ . It should be noted that the cost function is evaluated using the non-blind solver $F(.)$ and Kernel-Trajectory Network $T(.)$ - two neural network computations. Therefore, we can compute the gradient $\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)$ using autodifferentiation tools provided in PyTorch by backpropagating the gradients through $F(.)$ and then $T(.)$", + "bbox": [ + 75, + 402, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage II: Kernel Fine-tuning In the second stage, using the kernel estimate of Stage I, we fine-tune the kernel by \"opening up\" the search space to the entire kernel vector instead of parametrizing by $T(\\cdot)$ . Specifically, we optimize the following loss function", + "bbox": [ + 75, + 523, + 468, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {h}) \\stackrel {\\text {d e f}} {=} \\| G (\\mathbf {y}) - \\mathbf {h} * F (\\mathbf {y}, \\mathbf {h})) \\| _ {2} ^ {2} + \\gamma \\| \\mathbf {h} \\| _ {1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 611, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note the presence of the second term which acts as an $\\ell_1$ -norm sparsity prior. Also the kernel vector $\\mathbf{h}$ is being optimized instead of the latent key point vector $\\mathbf{z}$ . Using variable splitting as used in Half-Quadratic Splitting (HQS), we convert the optimization problem in (4) to as follows:", + "bbox": [ + 75, + 642, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {h}, \\mathbf {v}) = \\| G (\\mathbf {y}) - \\mathbf {h} * F (\\mathbf {y}, \\mathbf {h})) \\| _ {2} ^ {2} + \\gamma \\| \\mathbf {h} \\| _ {1} + \\frac {\\mu}{2} \\| \\mathbf {h} - \\mathbf {v} \\| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 729, + 485, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for some hyperparameter $\\mu > 0$ . This leads us to the following iterative updates", + "bbox": [ + 75, + 781, + 468, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} ^ {k + 1} = \\mathbf {h} ^ {k} - \\delta \\cdot \\left\\{\\nabla_ {\\mathbf {h}} \\mathcal {L} \\left(\\mathbf {h} ^ {k}\\right) + \\mu \\left(\\mathbf {h} ^ {k} - \\mathbf {v} ^ {k}\\right) \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 821, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {v} ^ {k + 1} = \\max \\left(\\left| \\mathbf {h} ^ {k + 1} \\right| - \\gamma / \\mu , 0\\right) \\cdot \\operatorname {s i g n} \\left(\\mathbf {h} ^ {k + 1}\\right) \\\\ \\stackrel {\\text {d e f}} {=} \\mathcal {S} _ {\\gamma / \\mu} (\\mathbf {h} ^ {k + 1}). \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 842, + 468, + 886 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Iterative Poisson Deconvolution Scheme", + "text_level": 1, + "bbox": [ + 501, + 402, + 857, + 417 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Noisy-blurry $\\mathbf{y}$ , Photon-Level $\\alpha$ , denoiser $G(\\cdot)$ , non-blind solver $F(\\cdot)$ , Kernel-Trajectory-Network $T(\\cdot)$ .", + "2: Initialize $\\mathbf{z}^0$ using method described in Algorithm 1 in supplementary", + "3: for $k = 0,1,2,\\dots$ do % Stage I begins here", + "4: $\\mathbf{h}_{\\mathbf{z}}^{k}\\gets T(\\mathbf{z}^{k})$", + "5: $\\mathcal{L}(\\mathbf{z})\\gets \\| G(\\mathbf{y}) - \\mathbf{h}_{\\mathbf{z}}^{k}\\circledast F(\\mathbf{y},\\mathbf{h}_{\\mathbf{z}}^{k})\\|_{2}^{2}$", + "6: Calculate $\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)$ using automatic differentiation", + "7: $\\mathbf{z}^{k + 1}\\gets \\mathbf{z}^k -\\delta \\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)$", + "8: end for", + "9: $\\mathbf{h}^0, \\mathbf{v}^0 \\gets T(\\mathbf{z}^\\infty), \\mu \\gets 2.0, \\gamma \\gets 10^{-4}$", + "10: for $k = 0,1,2,\\dots$ do % Stage II begins here", + "11: $\\mathcal{L}(\\mathbf{h})\\gets \\| G(\\mathbf{y}) - \\mathbf{h}\\circledast F(\\mathbf{y},\\mathbf{h})\\| _2^2$", + "12: Calculate $\\nabla_{\\mathbf{h}}\\mathcal{L}(\\mathbf{h}^k)$ using automatic differentiation", + "13: $\\mathbf{h}^{k + 1}\\gets \\mathbf{h}^k -\\delta \\big(\\nabla_{\\mathbf{h}}\\mathcal{L}(\\mathbf{h}^k) + \\mu (\\mathbf{h}^k -\\mathbf{v}^k)\\big)$", + "14: $\\mathbf{v}^{k + 1}\\gets \\mathcal{S}_{\\gamma /\\mu}(\\mathbf{h}^{k + 1})$", + "15: $\\mu \\gets 1.01\\mu$", + "16: end for", + "17: return $\\mathbf{h}^{(\\infty)}$ and $\\mathbf{x}^{(\\infty)} = F(\\mathbf{y},\\mathbf{h}^{(\\infty)})$" + ], + "bbox": [ + 506, + 420, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 739, + 630, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Training", + "text_level": 1, + "bbox": [ + 500, + 768, + 604, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While our overall method is not end-to-end trained, it contains pre-trained components - the non-blind solver $F(\\cdot)$ and denoiser $G(\\cdot)$ . The architectures of $F(\\cdot)$ and $G(\\cdot)$ are inherited from PhD-Net [29] which takes as input a noisy-blurred image and kernel. However, for the denoiser $G(\\cdot)$ , we fix the kernel input to identity operator since it is trained to remove only the noise from noisy-blurred image $\\mathbf{y}$ .", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9866", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$F(.)$ and $G(.)$ are trained using synthetic data as follows. We take clean images from Flickr2K dataset [20] and the blur kernels from the code in Boracchi and Foi [4]. The blurred images are also corrupted using Poisson shot noise with photon levels $\\alpha$ uniformly sampled from [1, 60]. The non-blind solver $F(.)$ is trained using kernel and noisy blurred image as input, and clean image as the target. The denoiser $G(.)$ is trained with similar procedure but with blurred-noisy image as the only input blur-only images as the target. The training processes, along with other experiments described in this paper are implemented using PyTorch on a NVIDIA Titan Xp GPU.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For quantitative comparison of the method presented, we retrain the following state-of-the-art networks for Poisson Noise: Scale Recurrent Network (SRN) [34], Deep-Deblur [24], DHMPN [43], MPR-Net [42], and MIMO-UNet+ [8]. We perform this retraining in the following two different ways. First, we use synthetic data training as described for $F(.)$ and $G(.)$ . Second, for testing on realistic blur, we retrain the networks using the GoPro dataset [24] as it is often used to train neural networks in contemporary deblurring literature. We add the Poisson noise with the same distribution as the synthetic datasets to the blurred images. While retraining the networks, we use the respective loss functions from the original papers for sake of a fair comparison.", + "bbox": [ + 75, + 273, + 472, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 482, + 310, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We quantitatively evaluate the proposed method on three different datasets, and compare it with state-of-the-art deblurring methods. In addition to the end-to-end trained methods described previously, we also compare our approach to the following Poisson deblurring methods: Poisson-Plug-and-Play [28], and PURE-LET [18]. Even though these methods assume the blur kernel to be known, we include them in the quantitative comparison since they are specifically designed for Poisson noise. For all of the methods described above, we compare the restored image's quality using PSNR, SSIM, and Learned Perceptual Image Patch Similarity (LPIPS-Alex, LPIPS-VGG) [45]. We include the latter as another metric in our evaluation since failure of MSE/SSIM to assess image quality has been well documented in [36, 45]", + "bbox": [ + 75, + 506, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BSD100: First, we evaluate our method on synthetic blur as follows. We collect 100 random images from the BSD-500 dataset, blur them synthetically with motion kernels from the Levin dataset [17] followed by adding Poisson noise at photon-levels $\\alpha = 10, 20,$ and $40$ . The results of the quantitative evaluation are provided in Table 1. Since the blur is synthetic, ground-truth kernel is known and hence, can be used to simultaneously evaluate Poisson non-blind deblurring methods i.e, Poisson Plug-and-Play, PURE-LET, and PhD-Net. The last method is the non-blind solver $F(.)$ and serves as an upper bound on the deconvolution perfor", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mance.", + "bbox": [ + 500, + 93, + 552, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Levin Dataset: Next, we evaluate our method on the Levin dataset [17] which contains 32 real blurred images along with the ground truth kernels, as measured through a point source. We evaluate our method on this dataset with addition of Poisson noise at photon levels $\\alpha = 10, 20$ and 40 and the results are shown in Table 2. For a fair comparison, end-to-end trained methods are retrained using synthetically blurred data (as described in Section IV-A) for evaluation on BSD100 and Levin dataset.", + "bbox": [ + 496, + 107, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RealBlur-J [27]: To demonstrate that our method is able to handle realistic blur, we evaluate our performance on randomly selected 50 patches of size $256 \\times 256$ from the Real-Blur-J [27] dataset. Note that we reduce the size of the tested image because our method is based on a single-blur convolutional model. Such model may not be applicable for a large image with spatially varying blur and local motion of objects. However, for a smaller patch of a larger image, the single-blur-kernel model of deconvolution is a much more valid assumption.", + "bbox": [ + 496, + 243, + 892, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure a fair comparison, we evaluate end-to-end networks by retraining on both the synthetic and GoPro dataset. As shown in Table 3, we find that end-to-end networks perform consistently better on the RealBlur dataset when trained using the GoPro dataset instead of synthetic blur. This can be explained by the fact both GoPro and RealBlur have realistic blur which is not necessarily captured by a single blur convolutional model.", + "bbox": [ + 496, + 395, + 892, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 500, + 527, + 720, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Color Reconstruction We show reconstructions on examples from the real-blur dataset in Figure 4. While our method is grayscale, we perform colour reconstruction by estimating the kernel from the luminance-channel. Given the estimated kernel, we deblur each channel of the image using the non-blind solver and then combine the different channels into a single RGB-image. Note that all qualitative examples in this paper for end-to-end trained networks are trained using the GoPro dataset, since they provide the better visual result.", + "bbox": [ + 496, + 551, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photon-Limited Deblurring Dataset We also show qualitative examples from photon-limited deblurring dataset [29] which contains 30 images' raw sensor data, blurred by camera shake and taken in extremely low-illumination. For reconstructing these images, we take the average of the R, G, B channels of the Bayer patter image, average it and then reconstruct it using the given method. The qualitative results for this dataset can be found in Figure 5. We also show the estimated kernels, along with estimated kernels from [30,40], in Figure 6.", + "bbox": [ + 496, + 703, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, instead of using the reburring loss directly, we find the scheme is more numerically stable if we take the gradients of the image first and then estimate the reburring", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/27e028122713e790b09e6a3ee9f160fba3e408eb29884858493a79d64d775067.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method\nPhoton Level, MetricSRN [34]DHMPN [43]Deep-\nDeblur [24]MIMO-Unet+ [8]MPRNet [42]OursP4IP [28]PURE-LET [18]PhD-Net [29]
α = 10PSNR ↑20.7120.8921.1721.0421.0921.5719.2622.4923.00
SSIM ↑0.3860.3910.4010.3560.3930.4710.3480.4850.500
LPIPS-Alex ↓0.6810.7020.6560.7330.6780.5600.7330.5880.544
LPIPS-VGG ↓0.6460.6520.6270.6830.6410.5870.6740.6070.567
α = 20PSNR ↑20.7921.0321.3021.3621.2521.9319.4522.9423.63
SSIM ↑0.3920.4010.4100.3960.4050.4830.3530.5160.540
LPIPS-Alex ↓0.6830.6880.6660.6600.6670.5420.7260.5260.500
LPIPS-VGG ↓0.6390.6400.6210.6630.6310.5780.6680.5840.539
α = 40PSNR ↑20.8921.1521.4321.6321.4121.6220.1823.4824.38
SSIM ↑0.4090.4180.4250.4410.4280.5270.3720.5610.593
LPIPS-Alex ↓0.6770.6730.6730.5860.6470.4880.7060.4670.446
LPIPS-VGG ↓0.6290.6260.6120.6390.6130.5490.6600.5570.503
Blind?✓✓✓✓✓✓XXX
End-To-End Trained?✓✓✓✓✓XXX✓
", + "bbox": [ + 86, + 87, + 875, + 339 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/03af7ded342a7af4fe279da322890292351bf1b9d1a1b4d00d3dc0681fed9ed4.jpg", + "table_caption": [ + "Table 1. Performance on BSD100 Dataset with Synthetic Blur. ↑ represents metrics where higher means better and vice versa for ↓. LPIPS-Alex and LPIPS-VGG represent the perceptual measures from [45]. The best performing blind deconvolution method for each metric and photon level is shown in bold. The non-blind deconvolution methods are shown for reference in grey columns." + ], + "table_footnote": [], + "table_body": "
Method\nPhoton Level, MetricSRN\n[34]DHMPN\n[43]Deep-\nDeblur [24]MIMO-\nUNet+ [8]MPRNet\n[42]OursP4IP\n[28]PURE-LET\n[18]PhD-Net\n[29]
α = 10PSNR ↑20.2620.5020.9321.2521.0422.0119.9221.6322.41
SSIM ↑0.5100.5090.5240.5160.5330.6110.4630.5900.638
LPIPS-Alex ↓0.5070.5210.4960.5940.4790.3400.5460.3710.341
LPIPS-VGG ↓0.5310.5260.5180.6610.5110.4770.5550.5220.466
α = 20PSNR ↑20.4920.3921.1121.6421.3322.7219.5321.7922.78
SSIM ↑0.5230.5210.5360.5540.5510.6410.4420.6070.667
LPIPS-Alex ↓0.4960.5020.4920.4850.4590.3040.5330.3390.304
LPIPS-VGG ↓0.5150.5140.5010.6100.4930.4480.5540.5100.447
α = 40PSNR ↑20.5920.5021.2021.8821.5422.3217.3221.7822.96
SSIM ↑0.5350.5320.5450.5830.5670.6470.3620.6140.687
LPIPS-Alex ↓0.4910.4940.4940.4280.4470.2730.4870.3240.263
LPIPS-VGG ↓0.5060.5060.4930.5570.4790.4440.5600.5070.432
", + "bbox": [ + 86, + 410, + 875, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance on Levin dataset with realistic camera shake blur [16]. The best performing blind deconvolution method for each metric and photon level is shown in bold and non-blind deconvolution methods are shown for reference in grey columns.", + "bbox": [ + 75, + 645, + 892, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "loss. This can be explained by the fact that unlike simulated data, the photon level is not known exactly and is estimated using the sensor data itself by a simple heuristic. For further details on how to use the sensor data, we refer the reader to [29].", + "bbox": [ + 75, + 700, + 470, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 230, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 4, we provide an ablation study by running the scheme for different number of key points i.e. $K = 4,6$ and 8 and without KTN $(K = 0)$ on RealBlur dataset. Through this study, we demonstrate the effect of the Kernel Trajectory Network has on the iterative scheme. As expected, changing the search space for kernel estimation im", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proves the performance significantly across all metrics. Increasing the number of key points used for representing kernels also steadily improves the performance of the scheme, which can be explained by the fact there are larger degrees of freedom.", + "bbox": [ + 496, + 700, + 892, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 797, + 619, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we use an iterative framework for the photon-limited blind deconvolution problem. More specifically, we use a non-blind solver which can deconvolve Poisson corrupted and blurred images given a blur kernel. To mitigate ill-posedness of the kernel estimation in such", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9868", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7e2b47fad625dfc0af46d2df94382e7f113097e26ed3c69b2c9675f605f5adc6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method →SRN [34]DHMPN [43]Deep-Deblur [24]MIMO-UNet+ [8]MPRNet [42]Ours
Training → Photon lvl, MetricSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.
α = 10PSNR ↑25.7227.6425.7227.5825.9827.5726.2026.7826.2628.1626.61
SSIM ↑0.6120.7060.6030.6960.5770.7190.5310.5710.6410.7290.738
LPIPS-Alex ↓0.4380.3100.4540.3290.4410.2970.4840.3960.4010.2880.277
LPIPS-VGG ↓0.5080.4540.5090.4720.4960.4400.5490.5080.4960.4270.416
α = 20PSNR ↑25.3727.9125.4628.0225.9527.8126.6927.5326.5128.2927.23
SSIM ↑0.6580.7750.6550.7640.6360.7780.6300.6780.7150.7930.793
LPIPS-Alex ↓0.4260.2750.4290.2880.4270.2650.4010.3130.3600.2560.241
LPIPS-VGG ↓0.4920.4210.4960.4370.4850.4100.4950.4460.4660.4020.382
α = 40PSNR ↑25.6728.3425.7228.2726.2228.1327.2428.1426.8528.7227.11
SSIM ↑0.6650.7680.6530.7600.6260.7710.6750.7120.7160.7880.782
LPIPS-Alex ↓0.4150.2680.4180.2680.4180.2580.3470.2670.3430.2450.221
LPIPS-VGG ↓0.4820.4050.4810.4130.4700.3960.4570.4040.4440.3860.360
", + "bbox": [ + 81, + 88, + 895, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance on RealBlur-J Dataset with realistic blur [27]: Bold and underline refer to overall best performing method and best synthetic performance method. It should be noted that methods that are not trained end-to-end are usually at disadvantage when comparing on metrics like PSNR. However, it can be seen that our reconstruction is generally preferred by other perceptual metrics.", + "bbox": [ + 75, + 345, + 893, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d477d47fe21454f99a2b516ece3f9c65b8ef0f98124e3a0440e82af0fc1fcffe.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 76, + 400, + 395, + 647 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1fe634750551cf68b9f01a42f99ad3e89f2879db2b3f929d0ffbb730c3dd874.jpg", + "image_caption": [ + "SRN" + ], + "image_footnote": [], + "bbox": [ + 406, + 402, + 560, + 517 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/445b30191774076574eb350b0cee49a750b59b4b536b7ac2b4c1a84ed2354d5f.jpg", + "image_caption": [ + "DHMPN" + ], + "image_footnote": [], + "bbox": [ + 570, + 402, + 720, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c500ff00a4dc6b8f72224939a03e1d2a306db948f6d50a9a990b684600a55f92.jpg", + "image_caption": [ + "MPR-Net" + ], + "image_footnote": [], + "bbox": [ + 728, + 402, + 880, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01eeeb6a7fb5af23b7dfbbee07c8502fe05cd54c6fbd3378b59038c7ec7ef2df.jpg", + "image_caption": [ + "MIMO-UNet+", + "Figure 4. Qualitative example on the Real-Blur Dataset: For a more extensive set of results, we refer the reader to the supplementary document." + ], + "image_footnote": [], + "bbox": [ + 405, + 535, + 555, + 650 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d104b36501a56302ca101fc25328c14d316e4b1c6d137a66c29284e9abf6b2aa.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 576, + 535, + 725, + 650 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74c1624539f8accb5d5a5f5bad0f34b1f5785a75b7d1c9fe9265c6123bea0c66.jpg", + "image_caption": [ + "Ground-Truth" + ], + "image_footnote": [], + "bbox": [ + 730, + 535, + 880, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "high noise regimes, we propose a novel low-dimensional representation to represent the motion blur. By using this novel low-dimensional and differentiable representation as a search space, we show state-of-the-art deconvolution performance and outperform end-to-end trained image restoration networks by a significant margin.", + "bbox": [ + 75, + 734, + 470, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We believe this is a promising direction of research for both deblurring and general blind inverse problems i.e., inverse problems where the forward operator is not fully known. Future work could involve a supervised version", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of this scheme which does not involve backpropagation through a network as it would greatly reduce the computational cost. Another research direction could be to apply this framework to the problem of spatially varying blur.", + "bbox": [ + 496, + 734, + 893, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 825, + 648, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The work is supported, in part, by the United States National Science Foundation under the grants IIS-2133032 and ECCS-2030570.", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9869", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9b34d8706aede597c516a6f23b594d7796ad6d1f7c3e40d9dc239a1517364bc.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 73, + 85, + 395, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c2621458c21df398131d0c54c8cdc6f1d1ea518fe497cdacf6fa5ccf1dc2acc.jpg", + "image_caption": [ + "SRN" + ], + "image_footnote": [], + "bbox": [ + 406, + 87, + 558, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a4cf93e9228ec0dbe29c259ef329f18dbff332fb2341659b203e89940509fe7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 403, + 220, + 553, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b2fa0c32b8c9bb64cd66c36aed3fc4081b49f59f046cc81f3a42caed119caab.jpg", + "image_caption": [ + "DHMPN" + ], + "image_footnote": [], + "bbox": [ + 568, + 88, + 718, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ac61e326f309512c40e1a27f5ca5d357924882ba0b84335f1960e334e9bc9f9.jpg", + "image_caption": [ + "Non-Blind" + ], + "image_footnote": [], + "bbox": [ + 573, + 220, + 723, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2caa0a0f7b704c6880cd861047c5281cb22bdccfa39c911cdd23e5b02baca836.jpg", + "image_caption": [ + "MPR-Net" + ], + "image_footnote": [], + "bbox": [ + 728, + 88, + 877, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87cea74457ad089bdb0184845c96dcf2839bf4fc6f27c3150c45940d6f5c0dfe.jpg", + "image_caption": [ + "Ground-Truth" + ], + "image_footnote": [], + "bbox": [ + 728, + 220, + 877, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04c5112c3539aadd764bc7636eaa06b7f689b2c7f75d1489a8c9abb9c875c3b2.jpg", + "image_caption": [ + "Figure 5. Visual comparisons on Photon-Limited Deblurring Dataset. Qualitative results on realistic blurred and photon-limited images from the Photon-Limited Deblurring dataset [29]. The inset image for \"Ours\" and \"Non-Blind\" represent the estimated and ground-truth kernel respectively. For a more extensive set of qualitative results, we refer the reader to the supplementary document." + ], + "image_footnote": [], + "bbox": [ + 73, + 436, + 169, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4eda77e898ff5560a734069390f76ec4f816bf259395d90b9054410b4b9992c5.jpg", + "image_caption": [ + "Two-Phase [40]", + "Figure 6. Estimated Kernels for different methods: We show the estimated kernels from two examples from the PLDD dataset. Two-Phase [40] uses blur-only image $G(\\mathbf{y})$ as input, and ground-truth kernel is estimated using a point-source." + ], + "image_footnote": [], + "bbox": [ + 73, + 513, + 169, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/796a2e733cc9ee6a57512c46ffdf93b664f08bdf8140adc0d2654802183b58b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 173, + 438, + 267, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12ae9125af3590569a74e92f794ce0a247097cbe046f465aa11c3564cb9c2774.jpg", + "image_caption": [ + "Sanghvi et.al [30]" + ], + "image_footnote": [], + "bbox": [ + 173, + 513, + 267, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/966b37f00bf6d80d025a0a9e92f876cc927e7caa3e357175c7e73dc40a0f339b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 438, + 364, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42da63d1576319cf4b7b2539241e3b87fe573a3fa2f5e9a5be3516f60f7560c7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 269, + 513, + 364, + 588 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78ea8fb7ed6b39ae08f508e6aec9efc37a55d00159203a8099e8a9c516accaa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 438, + 462, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41d4728cefd2b726977b20bb9c5ce3ba039abedcb952e7e837cbedf75990e067.jpg", + "image_caption": [ + "GroundTruth" + ], + "image_footnote": [], + "bbox": [ + 367, + 513, + 462, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 715, + 173, + 731 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chirag Agarwal, Shahin Khobahi, Arindam Bose, Mojtaba Soltanalian, and Dan Schonfeld. Deep-uri: A model-aware approach to blind deconvolution based on deep unfolded Richardson-Lucy network. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3299-3303, 2020. 2", + "[2] Jérémy Anger, Maurício Delbracio, and Gabriele Facciolo. Efficient blind deblurring under high noise levels. In 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 123-128. IEEE, 2019. 2", + "[3] Muhammad Asim, Fahad Shamshad, and Ali Ahmed." + ], + "bbox": [ + 84, + 742, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c9282bfd3d213c6850cc8130c4f5b6e9f468382d589f252304641b92cdbcde87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Key Points ↓\nPhoton Level ↓\nMetric →PSNRSSIMLPIPS-\nAlexLPIPS-\nVGG
α = 10K = 025.440.7190.2950.434
K = 425.900.7300.2860.423
K = 626.380.7350.2790.422
K = 826.730.7400.2720.414
α = 20K = 025.610.7670.2670.407
K = 426.930.7850.2420.385
K = 626.780.7840.2460.392
K = 827.260.7950.2400.384
α = 40K = 025.820.7590.2480.390
K = 426.880.7710.2280.374
K = 627.370.7800.2160.366
K = 827.290.7850.2140.361
", + "bbox": [ + 524, + 438, + 870, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Study for effect of Kernel Trajectory Network $T(\\cdot)$ : Reconstruction metrics for different number of key points. $K = 0$ represents the variant of the scheme which does not use the Kernel-Trajectory Network and estimates the kernel directly from Stage II.", + "bbox": [ + 500, + 678, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging, 6:1493-1506, 2020. 2", + "[4] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE Transactions on Image Processing, 21(8):3502-3517, 2012. 5", + "[5] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV)," + ], + "bbox": [ + 508, + 781, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9870", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pages 221-235. Springer, 2016. 2", + "[6] Tony F Chan and Chiu-Kwong Wong. Total variation blind deconvolution. IEEE Transactions on Image Processing, 7(3):370-375, 1998. 2", + "[7] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Transactions on Graphics, pages 1-8, 2009. 2", + "[8] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1, 2, 5, 6, 7", + "[9] Mauricio Delbracio, Ignacio Garcia-Dorado, SungJoon Choi, Damien Kelly, and Peyman Milanfar. Polyblur: Removing mild blur by polynomial reburring. IEEE Transactions on Computational Imaging, 7:837-848, 2021. 3", + "[10] Mario AT Figueiredo and Jose M Bioucas-Dias. Deconvolution of Poissonian images using variable splitting and augmented Lagrangian optimization. In Proceedings of the IEEE/SP Workshop on Statistical Signal Processing, pages 733-736, 2009. 2", + "[11] Dong Gong, Jie Yang, Lingqiao Liu, Yanning Zhang, Ian Reid, Chunhua Shen, Anton Van Den Hengel, and Qinfeng Shi. From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2319-2328, 2017. 2", + "[12] Ankit Gupta, Neel Joshi, C Lawrence Zitnick, Michael Cohen, and Brian Curless. Single image deblurring using motion density functions. In European Conference on Computer Vision (ECCV), pages 171-184. Springer, 2010. 2", + "[13] Zachary T Harmany, Roummel F Marcia, and Rebecca M Willett. This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and practice. IEEE Transactions on Image Processing, 21(3):1084-1096, 2011. 2", + "[14] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jií Matas. DeblurGAN: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 2", + "[15] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In The IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 2", + "[16] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems (NeurIPS), 19, 2006. 2, 6", + "[17] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1964-1971, 2009. 5", + "[18] Jizhou Li, Florian Luisier, and Thierry Blu. PURE-LET image deconvolution. IEEE Transactions on Image Processing, 27(1):92–105, 2017. 2, 5, 6", + "[19] Yuelong Li, Mohammad Tofighi, Vishal Monga, and Yonina C Eldar. An algorithm unrolling approach to deep image deblurring. In IEEE International Conference on Acoustics," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Speech and Signal Processing (ICASSP), pages 7675-7679. IEEE, 2019. 2", + "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 5", + "[21] Leon B Lucy. An iterative technique for the rectification of observed distributions. The Astronomical Journal, 79:745, 1974. 2", + "[22] Zhiyuan Mao, Nicholas Chimitt, and Stanley H. Chan. Image reconstruction of static and dynamic scenes through anisoplanatic turbulence. IEEE Transactions on Computational Imaging, 6:1415-1428, 2020. 2", + "[23] Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 2", + "[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2, 5, 6, 7", + "[25] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3338-3347, 2020. 2", + "[26] William Hadley Richardson. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1):55-59, 1972. 2", + "[27] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Proceedings of the European Conference on Computer Vision (ECCV), pages 184-201. Springer, 2020. 2, 5, 7", + "[28] Arie Rond, Raja Giryes, and Michael Elad. Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation, 41:96-108, 2016. 2, 5, 6", + "[29] Yash Sanghvi, Abhiram Gnanasambandam, and Stanley H Chan. Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging (TCI), 8:851-864, 2022. 2, 4, 5, 6, 8", + "[30] Yash Sanghvi, Abhiram Gnanasambandam, Zhiyuan Mao, and Stanley H. Chan. Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging, 8:1051-1062, 2022. 1, 2, 3, 5, 8", + "[31] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(7):1439-1451, 2015. 2", + "[32] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. ACM Transactions on Graphics, 27(3):1-10, 2008. 2", + "[33] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9871", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2", + "[34] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 1, 2, 5, 6, 7", + "[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 2", + "[36] Zhou Wang and Alan C Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 5", + "[37] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general U-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, June 2022. 2", + "[38] Oliver Whyte, Josef Sivic, and Andrew Zisserman. Deblurring shaken and partially saturated images. International Journal of Computer Vision, 110(2):185-201, 2014. 2", + "[39] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International Journal of Computer Vision, 98(2):168-186, 2012. 2", + "[40] Li Xu and Jiaya Jia. Two-phase kernel estimation for robust motion deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), pages 157-170. Springer, 2010. 2, 5, 8", + "[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 1, 2, 5, 6, 7", + "[43] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1, 2, 5, 6, 7", + "[44] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[46] Youjian Zhang, Chaoyue Wang, Stephen J Maybank, and Dacheng Tao. Exposure trajectory recovery from motion blur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7490-7504, 2021. 2", + "bbox": [ + 501, + 90, + 890, + 147 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9872", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_model.json b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6891a0b5a624335a2b95970f0c558c9fe77a944a --- /dev/null +++ b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_model.json @@ -0,0 +1,2519 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.812, + 0.151 + ], + "angle": 0, + "content": "Structured Kernel Estimation for Photon-Limited Deconvolution" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.182, + 0.75, + 0.234 + ], + "angle": 0, + "content": "Yash Sanghvi, Zhiyuan Mao, Stanley H. Chan \nSchool of Electrical and Computer Engineering, Purdue University \n{ysanghvi, maol14, stanchan}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.471, + 0.391 + ], + "angle": 0, + "content": "Images taken in a low light condition with the presence of camera shake suffer from motion blur and photon shot noise. While state-of-the-art image restoration networks show promising results, they are largely limited to well-illuminated scenes and their performance drops significantly when photon shot noise is strong." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.474, + 0.619 + ], + "angle": 0, + "content": "In this paper, we propose a new blur estimation technique customized for photon-limited conditions. The proposed method employs a gradient-based backpropagation method to estimate the blur kernel. By modeling the blur kernel using a low-dimensional representation with the key points on the motion trajectory, we significantly reduce the search space and improve the regularity of the kernel estimation problem. When plugged into an iterative framework, our novel low-dimensional representation provides improved kernel estimates and hence significantly better deconvolution performance when compared to end-to-end trained neural networks. The source code and pretrained mdoels are available at https://github.com/sanghviyashiitb/structured-kernel-cvpr23" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.646, + 0.21, + 0.66 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Photon-Limited Blind Deconvolution: This paper studies the photon-limited blind deconvolution problem. Blind deconvolution refers to simultaneously recovering both the blur kernel and latent clean image from a blurred image and \"photon-limited\" refers to presence of photon-shot noise in images taken in low-illumination / short exposure. The corresponding forward model is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.785, + 0.47, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\operatorname {P o i s s o n} (\\alpha \\mathbf {h} * \\mathbf {x}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this equation, \\(\\mathbf{y} \\in \\mathbb{R}^N\\) is the blurred-noisy image, \\(\\mathbf{x} \\in \\mathbb{R}^N\\) is the latent clean image, and \\(\\mathbf{h} \\in \\mathbb{R}^M\\) is the blur kernel. We assume that \\(\\mathbf{x}\\) is normalized to \\([0,1]\\) and the entries of \\(\\mathbf{h}\\) are non-negative and sum up to 1. The constant \\(\\alpha\\) represents the average number of photons per pixel and is inversely proportional to the amount of Poisson noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.42 + ], + "angle": 0, + "content": "Deep Iterative Kernel Estimation: Blind image deconvolution has been studied for decades with many successful algorithms including the latest deep neural networks [8, 24, 34, 42, 43]. Arguably, the adaptation from the traditional Gaussian noise model to the photon-limited Poisson noise model can be done by retraining the existing networks with appropriate data. However, the restoration is not guaranteed to perform well because the end-to-end networks seldom explicitly take the forward image formation model into account." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.895, + 0.546 + ], + "angle": 0, + "content": "Recently, people have started to recognize the importance of blur kernel estimation for photon-limited conditions. One of these works is by Sanghvi et. al [30], where they propose an iterative kernel estimation method to backpropagate the gradient of an unsupervised reburring function, hence to update the blur kernel. However, as we can see in Figure 1, their performance is still limited when the photon shot noise is strong." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.895, + 0.654 + ], + "angle": 0, + "content": "Structured Kernel Estimation: Inspired by [30], we believe that the iterative kernel estimation process and the unsupervised reburbling loss are useful. However, instead of searching for the kernel directly (which can easily lead to local minima because the search space is too big), we propose to search in a low-dimensional space by imposing structure to the motion blur kernel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.808 + ], + "angle": 0, + "content": "To construct such a low-dimensional space, we frame the blur kernel in terms of trajectory of the camera motion. Motion trajectory is often a continuous but irregular path in the two-dimensional plane. To specify the trajectory, we introduce the concept of key point estimation where we identify a set of anchor points of the kernel. By interpolating the path along these anchor points, we can then reproduce the kernel. Since the number of anchor points is significantly lower than the number of pixels in a kernel, we can reduce the dimensionality of the kernel estimation problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The key contribution of this paper is as follows: We propose a new kernel estimation method called Kernel Trajectory Network (KTN). KTN models the blur kernel in a low-dimensional and differentiable space by specifying key points of the motion trajectory. Plugging this low-dimensional representation in an iterative framework im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.088, + 0.247, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.214, + 0.224, + 0.227 + ], + "angle": 0, + "content": "Blurred and Noisy" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.088, + 0.405, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.214, + 0.375, + 0.227 + ], + "angle": 0, + "content": "MPR-Net [42]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.088, + 0.565, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.214, + 0.541, + 0.228 + ], + "angle": 0, + "content": "Sanghvi et. al [30]" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.088, + 0.722, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.215, + 0.662, + 0.226 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.088, + 0.882, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.214, + 0.845, + 0.226 + ], + "angle": 0, + "content": "Ground-Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.894, + 0.298 + ], + "angle": 0, + "content": "Figure 1. The proposed Kernel Trajectory Network (KTN) on real noisy blurred image from Photon-Limited Deblurring Dataset (PLDD) [29] The result corresponding to MPR-Net was generated by retraining the network with GoPro dataset [24] corrupted by Poisson noise. The inset images for \"Sanghvi et. al\" and \"Ours\" represent the estimated kernel and the inset image for \"Ground-Truth\" represents the kernel captured using a point source, as provided in PLDD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.369 + ], + "angle": 0, + "content": "proves the regularity of the kernel estimation problem. This leads to substantially better blur kernel estimates in photon-limited regimes where existing methods fail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.387, + 0.219, + 0.403 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Traditional Blind Deconvolution: Classical approaches to the (noiseless) blind deconvolution problem [6, 7, 22, 32, 40] use a joint optimization framework in which both the kernel and image are updated in an alternating fashion in order to minimize a cost function with kernel and image priors. For high noise regimes, a combination of \\(\\ell_1 + \\mathrm{TV}\\) prior has been used in [2]. Levin et. al [16] pointed out that this joint optimization framework for the blind deconvolution problem favours the no-blur degenerate solution i.e. \\((\\mathbf{x}^{*},\\mathbf{h}^{*}) = (\\mathbf{y},\\mathbf{I})\\) where \\(\\mathbf{I}\\) is the identity operator. Some methods model the blur kernel in terms of the camera trajectory and then recover both the trajectory and the clean image using optimization [12, 38, 39] and supervised-learning techniques [11, 33, 46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.718 + ], + "angle": 0, + "content": "For the non-blind case, i.e., when the blur kernel is assumed to be known, the Poisson deconvolution problem has been studied for decades starting from Richardson-Lucy algorithm [21, 26]. More contemporary methods include Plug-and-Play [28, 29], PURE-LET [18], and MAP-based optimization methods [10, 13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Deep Learning Methods. Recent years, many deep learning-based methods [5, 31] have been proposed for the blind image deblurring task. The most common strategy is to train a network end-to-end on large-scale datasets, such as the GoPro [24] and the RealBlur [27] datasets. Notably, many recent works [8, 24, 34, 42, 43] improve the performance of deblurring networks by adopting the multiscale strategies, where the training follows a coarse-to-fine setting that resembles the iterative approach. Generative Adversarial Network (GAN) based deblurring methods [3,14,15,44] are also shown to produce visually appealing images. Zamir et al. [41] and Wang et al. [37] adapt the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.367 + ], + "angle": 0, + "content": "popular vision transformers to the image restoration problems and demonstrate competitive performance on the deblurring task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Neural Networks and Iterative Methods: While neural networks have shown state-of-the-art performance on the deblurring task, another class of methods incorporating iterative methods with deep learning have shown promising results. Algorithm unrolling [23], where an iterative method is unrolled for fixed iterations and trained end-to-end has been applied to image deblurring [1,19]. In SelfDeblur [25], authors use Deep-Image-Prior [35] to represent the image and blur kernel and obtain state-of-the-art blind deconvolution performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.591, + 0.55 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.845, + 0.575 + ], + "angle": 0, + "content": "3.1. Kernel as Structured Motion Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Camera motion blur can be modeled as a latent clean image \\(\\mathbf{x}\\) convolved with a blur kernel \\(\\mathbf{h}\\). If we assume the blur kernel lies in a window of size \\(32 \\times 32\\), then \\(\\mathbf{h} \\in \\mathbb{R}^{1024}\\). However, in this high dimensional space, only few entries of the blur kernel \\(\\mathbf{h}\\) are non-zero. Additionally, the kernel is generated from a two-dimensional trajectory which suggests that a simple sparsity prior is not sufficient. Given the difficulty of the photon-limited deconvolution problem, we need to impose a stronger prior on the kernel. To this end, we propose a differentiable and low-dimensional representation of the blur kernel, which we will use as the search space in our kernel estimation algorithm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We take the two-dimensional trajectory of the camera during the exposure time and divide it into \\( K \\) \"key points\". Each key point represents either the start, the end or a change in direction of the camera trajectory as seen in Figure 2. Given the \\( K \\) key points as points mapped out in \\( x - y \\) space, we can interpolate them using cubic splines to form a continuous trajectory in 2D. To convert this continuous trajectory to an equivalent blur kernel, we assume a point source image and move it through the given trajectory. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9864" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.073, + 0.082, + 0.478, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.489 + ], + "angle": 0, + "content": "Figure 2. Blur as Structured Motion Estimation: In our formulation, we view the blur kernel as the continuous camera trajectory reduced to \\( K \\) key points, as shown in top half of the figure. We learn a differentiable representation from the vectorized \\( K \\) key points to a blur kernel using a neural network. This lower dimensional and differentiable representation is leveraged to estimate a better blur kernel and avoiding local minima during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.549 + ], + "angle": 0, + "content": "resulting frames are then averaged to give the corresponding blur kernel as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Given the formulation of blur kernel \\(\\mathbf{h}\\) in terms of \\(K\\) key points, we now need to put this representation to a differentiable form since we intend to use it in an iterative scheme. To achieve this, we learn the transformations from the key points to the blur kernels using a neural network, which will be referred to as Kernel-Trajectory Network (KTN), and represent it using a differentiable function \\(T(\\cdot)\\). Why differentiability is important to us will become clear to the reader in the next subsection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.853 + ], + "angle": 0, + "content": "To train the Kernel-Trajectory Network, we generate training data as follows. First, for a fixed \\( K \\), we get \\( K \\) key points by starting from \\( (0,0) \\) and choosing the next \\( K - 1 \\) points by successively adding a random vector with a uniformly chosen random direction i.e. \\( U[0,360] \\) and uniformly chosen length from \\( U[0,100 / (K - 1)] \\). Next, the set of key points are converted to a continuous smooth trajectory using bicubic interpolation. Then, we move a point source image through the given trajectory using the warpPerspective function in OpenCV, and average the resulting frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Using the process defined above, we generate 60,000 blur kernels and their corresponding key point representations. For the Kernel-Trajectory Network \\( T(\\cdot) \\), we take a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "U-Net like network with the first half replaced by 3-fully connected layers and train it with the generated data using \\(\\ell_2\\)-loss. For further architectural details on the KTN, we refer the reader to the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.159, + 0.742, + 0.175 + ], + "angle": 0, + "content": "3.2. Proposed Iterative Scheme" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.349 + ], + "angle": 0, + "content": "We described in the previous subsection how to obtain a low-dimensional and differentiable representation \\( T(\\cdot) \\) for the blur kernel and now we are ready to present the full iterative scheme in detail. The proposed iterative scheme can be divided into three stages which are summarized as follows. We first generate an initial estimate of the direction and magnitude of the blur. This is used as initialization for a gradient-based scheme in Stage I which searches the appropriate kernel representation in the latent space \\( \\mathbf{z} \\). This is followed by Stage II where we fine-tune the kernel obtained from Stage I using a similar process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.408 + ], + "angle": 0, + "content": "Initialization Before starting the iterative scheme, we need a light-weight initialization method. This is important because of multiple local minima in the kernel estimation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.56 + ], + "angle": 0, + "content": "We choose to initialize the method with a rectilinear motion kernel, parameterized by length \\(\\rho\\) and orientation \\(\\theta\\). To determine the length and orientation of the kernel, we use a minor variation of the kernel estimation in PolyBlur [9]. In this variation, the \"blur-only image\" \\(G(\\mathbf{y})\\) is used as the input and \\(\\rho, \\theta\\) for the initial kernel are estimated using the minimum of the directional gradients. We refer the reader to Section II in the supplementary document for further details on the initialization. Explanation on the \"blur-only image\" is provided when we describe Stage I of the scheme." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Stage I: Kernel Estimation in Latent Space Given an initial kernel, we choose initial latent \\(\\mathbf{z}^0\\) by dividing the rectilinear kernel into \\(K\\) key points. Following the framework in [30], we run a gradient descent based scheme which optimizes the following cost function:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.641, + 0.892, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {z}) \\stackrel {\\text {d e f}} {=} \\underbrace {\\| G (\\mathbf {y}) - \\mathbf {h} _ {\\mathbf {z}} * F (\\mathbf {y} , \\mathbf {h} _ {\\mathbf {z}})) \\| _ {2} ^ {2}} _ {\\text {R e b l u r i n g L o s s}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.827 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_{\\mathbf{z}} \\stackrel{\\mathrm{def}}{=} T(\\mathbf{z})\\) represents the kernel output from Kernel-Trajectory network \\(T(.)\\) given the vectorized key points representation \\(\\mathbf{z}\\). \\(F(.)\\) represents the Poisson non-blind deconvolution solver which takes both noisy-blurred image and a blur kernel as the input. \\(G(\\mathbf{y})\\) represents a denoiser which is trained to remove only the noise from noisy-blurred image. The overall cost function represents reburring loss i.e. how well the kernel estimate and corresponding image estimate \\(\\mathbf{h}_{\\mathbf{z}} * F(\\mathbf{y}, \\mathbf{h}_{\\mathbf{z}})\\) match the blur-only image \\(G(\\mathbf{y})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.856 + ], + "angle": 0, + "content": "To minimize the cost function in (2), we use a simple gradient descent based iterative update for \\(\\mathbf{z}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.861, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {k + 1} = \\mathbf {z} ^ {k} - \\delta \\underbrace {\\nabla_ {\\mathbf {z}} \\mathcal {L} \\left(\\mathbf {z} ^ {k}\\right)} _ {\\text {b a c k p r o p a g a t i o n}} \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.818, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Figure 3. Flowchart describing first stage of the proposed scheme. We estimate the motion kernel of the blurry noisy image in lower dimensional latent space \\(\\mathbf{z}\\) where the blur kernel is represented by \\(T(\\mathbf{z})\\) and by minimizing the reburring loss \\(\\mathcal{L}\\) as defined in equation 2" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.525 + ], + "angle": 0, + "content": "where \\(\\delta > 0\\) is the step size and \\(\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)\\) represents the gradient of the cost function \\(\\mathcal{L}\\) with respect to \\(\\mathbf{z}\\) evaluated \\(\\mathbf{z}^k\\). It should be noted that the cost function is evaluated using the non-blind solver \\(F(.)\\) and Kernel-Trajectory Network \\(T(.)\\) - two neural network computations. Therefore, we can compute the gradient \\(\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)\\) using autodifferentiation tools provided in PyTorch by backpropagating the gradients through \\(F(.)\\) and then \\(T(.)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Stage II: Kernel Fine-tuning In the second stage, using the kernel estimate of Stage I, we fine-tune the kernel by \"opening up\" the search space to the entire kernel vector instead of parametrizing by \\( T(\\cdot) \\). Specifically, we optimize the following loss function" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.612, + 0.47, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {h}) \\stackrel {\\text {d e f}} {=} \\| G (\\mathbf {y}) - \\mathbf {h} * F (\\mathbf {y}, \\mathbf {h})) \\| _ {2} ^ {2} + \\gamma \\| \\mathbf {h} \\| _ {1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Note the presence of the second term which acts as an \\(\\ell_1\\)-norm sparsity prior. Also the kernel vector \\(\\mathbf{h}\\) is being optimized instead of the latent key point vector \\(\\mathbf{z}\\). Using variable splitting as used in Half-Quadratic Splitting (HQS), we convert the optimization problem in (4) to as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.73, + 0.486, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {h}, \\mathbf {v}) = \\| G (\\mathbf {y}) - \\mathbf {h} * F (\\mathbf {y}, \\mathbf {h})) \\| _ {2} ^ {2} + \\gamma \\| \\mathbf {h} \\| _ {1} + \\frac {\\mu}{2} \\| \\mathbf {h} - \\mathbf {v} \\| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.782, + 0.469, + 0.813 + ], + "angle": 0, + "content": "for some hyperparameter \\(\\mu > 0\\). This leads us to the following iterative updates" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.822, + 0.469, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} ^ {k + 1} = \\mathbf {h} ^ {k} - \\delta \\cdot \\left\\{\\nabla_ {\\mathbf {h}} \\mathcal {L} \\left(\\mathbf {h} ^ {k}\\right) + \\mu \\left(\\mathbf {h} ^ {k} - \\mathbf {v} ^ {k}\\right) \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.843, + 0.469, + 0.887 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {v} ^ {k + 1} = \\max \\left(\\left| \\mathbf {h} ^ {k + 1} \\right| - \\gamma / \\mu , 0\\right) \\cdot \\operatorname {s i g n} \\left(\\mathbf {h} ^ {k + 1}\\right) \\\\ \\stackrel {\\text {d e f}} {=} \\mathcal {S} _ {\\gamma / \\mu} (\\mathbf {h} ^ {k + 1}). \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.404, + 0.858, + 0.418 + ], + "angle": 0, + "content": "Algorithm 1 Iterative Poisson Deconvolution Scheme" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.421, + 0.892, + 0.466 + ], + "angle": 0, + "content": "1: Input: Noisy-blurry \\(\\mathbf{y}\\), Photon-Level \\(\\alpha\\), denoiser \\(G(\\cdot)\\), non-blind solver \\(F(\\cdot)\\), Kernel-Trajectory-Network \\(T(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.466, + 0.892, + 0.496 + ], + "angle": 0, + "content": "2: Initialize \\(\\mathbf{z}^0\\) using method described in Algorithm 1 in supplementary" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.496, + 0.81, + 0.51 + ], + "angle": 0, + "content": "3: for \\( k = 0,1,2,\\dots \\) do % Stage I begins here" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.51, + 0.635, + 0.524 + ], + "angle": 0, + "content": "4: \\(\\mathbf{h}_{\\mathbf{z}}^{k}\\gets T(\\mathbf{z}^{k})\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.524, + 0.775, + 0.539 + ], + "angle": 0, + "content": "5: \\(\\mathcal{L}(\\mathbf{z})\\gets \\| G(\\mathbf{y}) - \\mathbf{h}_{\\mathbf{z}}^{k}\\circledast F(\\mathbf{y},\\mathbf{h}_{\\mathbf{z}}^{k})\\|_{2}^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.539, + 0.857, + 0.551 + ], + "angle": 0, + "content": "6: Calculate \\(\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)\\) using automatic differentiation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.551, + 0.707, + 0.566 + ], + "angle": 0, + "content": "7: \\(\\mathbf{z}^{k + 1}\\gets \\mathbf{z}^k -\\delta \\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.578, + 0.577 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.579, + 0.768, + 0.593 + ], + "angle": 0, + "content": "9: \\(\\mathbf{h}^0, \\mathbf{v}^0 \\gets T(\\mathbf{z}^\\infty), \\mu \\gets 2.0, \\gamma \\gets 10^{-4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.594, + 0.8, + 0.607 + ], + "angle": 0, + "content": "10: for \\( k = 0,1,2,\\dots \\) do % Stage II begins here" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.607, + 0.76, + 0.621 + ], + "angle": 0, + "content": "11: \\(\\mathcal{L}(\\mathbf{h})\\gets \\| G(\\mathbf{y}) - \\mathbf{h}\\circledast F(\\mathbf{y},\\mathbf{h})\\| _2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.621, + 0.858, + 0.634 + ], + "angle": 0, + "content": "12: Calculate \\(\\nabla_{\\mathbf{h}}\\mathcal{L}(\\mathbf{h}^k)\\) using automatic differentiation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.634, + 0.817, + 0.65 + ], + "angle": 0, + "content": "13: \\(\\mathbf{h}^{k + 1}\\gets \\mathbf{h}^k -\\delta \\big(\\nabla_{\\mathbf{h}}\\mathcal{L}(\\mathbf{h}^k) + \\mu (\\mathbf{h}^k -\\mathbf{v}^k)\\big)\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.65, + 0.684, + 0.665 + ], + "angle": 0, + "content": "14: \\(\\mathbf{v}^{k + 1}\\gets \\mathcal{S}_{\\gamma /\\mu}(\\mathbf{h}^{k + 1})\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.665, + 0.624, + 0.678 + ], + "angle": 0, + "content": "15: \\(\\mu \\gets 1.01\\mu\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.679, + 0.578, + 0.689 + ], + "angle": 0, + "content": "16: end for" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.69, + 0.754, + 0.706 + ], + "angle": 0, + "content": "17: return \\(\\mathbf{h}^{(\\infty)}\\) and \\(\\mathbf{x}^{(\\infty)} = F(\\mathbf{y},\\mathbf{h}^{(\\infty)})\\)" + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.421, + 0.892, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.632, + 0.759 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.605, + 0.786 + ], + "angle": 0, + "content": "4.1. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "While our overall method is not end-to-end trained, it contains pre-trained components - the non-blind solver \\( F(\\cdot) \\) and denoiser \\( G(\\cdot) \\). The architectures of \\( F(\\cdot) \\) and \\( G(\\cdot) \\) are inherited from PhD-Net [29] which takes as input a noisy-blurred image and kernel. However, for the denoiser \\( G(\\cdot) \\), we fix the kernel input to identity operator since it is trained to remove only the noise from noisy-blurred image \\( \\mathbf{y} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9866" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "\\(F(.)\\) and \\(G(.)\\) are trained using synthetic data as follows. We take clean images from Flickr2K dataset [20] and the blur kernels from the code in Boracchi and Foi [4]. The blurred images are also corrupted using Poisson shot noise with photon levels \\(\\alpha\\) uniformly sampled from [1, 60]. The non-blind solver \\(F(.)\\) is trained using kernel and noisy blurred image as input, and clean image as the target. The denoiser \\(G(.)\\) is trained with similar procedure but with blurred-noisy image as the only input blur-only images as the target. The training processes, along with other experiments described in this paper are implemented using PyTorch on a NVIDIA Titan Xp GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.471 + ], + "angle": 0, + "content": "For quantitative comparison of the method presented, we retrain the following state-of-the-art networks for Poisson Noise: Scale Recurrent Network (SRN) [34], Deep-Deblur [24], DHMPN [43], MPR-Net [42], and MIMO-UNet+ [8]. We perform this retraining in the following two different ways. First, we use synthetic data training as described for \\( F(.) \\) and \\( G(.) \\). Second, for testing on realistic blur, we retrain the networks using the GoPro dataset [24] as it is often used to train neural networks in contemporary deblurring literature. We add the Poisson noise with the same distribution as the synthetic datasets to the blurred images. While retraining the networks, we use the respective loss functions from the original papers for sake of a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.312, + 0.5 + ], + "angle": 0, + "content": "4.2. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.735 + ], + "angle": 0, + "content": "We quantitatively evaluate the proposed method on three different datasets, and compare it with state-of-the-art deblurring methods. In addition to the end-to-end trained methods described previously, we also compare our approach to the following Poisson deblurring methods: Poisson-Plug-and-Play [28], and PURE-LET [18]. Even though these methods assume the blur kernel to be known, we include them in the quantitative comparison since they are specifically designed for Poisson noise. For all of the methods described above, we compare the restored image's quality using PSNR, SSIM, and Learned Perceptual Image Patch Similarity (LPIPS-Alex, LPIPS-VGG) [45]. We include the latter as another metric in our evaluation since failure of MSE/SSIM to assess image quality has been well documented in [36, 45]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "BSD100: First, we evaluate our method on synthetic blur as follows. We collect 100 random images from the BSD-500 dataset, blur them synthetically with motion kernels from the Levin dataset [17] followed by adding Poisson noise at photon-levels \\(\\alpha = 10, 20,\\) and \\(40\\). The results of the quantitative evaluation are provided in Table 1. Since the blur is synthetic, ground-truth kernel is known and hence, can be used to simultaneously evaluate Poisson non-blind deblurring methods i.e, Poisson Plug-and-Play, PURE-LET, and PhD-Net. The last method is the non-blind solver \\(F(.)\\) and serves as an upper bound on the deconvolution perfor" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.094, + 0.553, + 0.105 + ], + "angle": 0, + "content": "mance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Levin Dataset: Next, we evaluate our method on the Levin dataset [17] which contains 32 real blurred images along with the ground truth kernels, as measured through a point source. We evaluate our method on this dataset with addition of Poisson noise at photon levels \\(\\alpha = 10, 20\\) and 40 and the results are shown in Table 2. For a fair comparison, end-to-end trained methods are retrained using synthetically blurred data (as described in Section IV-A) for evaluation on BSD100 and Levin dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.893, + 0.395 + ], + "angle": 0, + "content": "RealBlur-J [27]: To demonstrate that our method is able to handle realistic blur, we evaluate our performance on randomly selected 50 patches of size \\(256 \\times 256\\) from the Real-Blur-J [27] dataset. Note that we reduce the size of the tested image because our method is based on a single-blur convolutional model. Such model may not be applicable for a large image with spatially varying blur and local motion of objects. However, for a smaller patch of a larger image, the single-blur-kernel model of deconvolution is a much more valid assumption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.516 + ], + "angle": 0, + "content": "To ensure a fair comparison, we evaluate end-to-end networks by retraining on both the synthetic and GoPro dataset. As shown in Table 3, we find that end-to-end networks perform consistently better on the RealBlur dataset when trained using the GoPro dataset instead of synthetic blur. This can be explained by the fact both GoPro and RealBlur have realistic blur which is not necessarily captured by a single blur convolutional model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.722, + 0.545 + ], + "angle": 0, + "content": "4.3. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Color Reconstruction We show reconstructions on examples from the real-blur dataset in Figure 4. While our method is grayscale, we perform colour reconstruction by estimating the kernel from the luminance-channel. Given the estimated kernel, we deblur each channel of the image using the non-blind solver and then combine the different channels into a single RGB-image. Note that all qualitative examples in this paper for end-to-end trained networks are trained using the GoPro dataset, since they provide the better visual result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Photon-Limited Deblurring Dataset We also show qualitative examples from photon-limited deblurring dataset [29] which contains 30 images' raw sensor data, blurred by camera shake and taken in extremely low-illumination. For reconstructing these images, we take the average of the R, G, B channels of the Bayer patter image, average it and then reconstruct it using the given method. The qualitative results for this dataset can be found in Figure 5. We also show the estimated kernels, along with estimated kernels from [30,40], in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "However, instead of using the reburring loss directly, we find the scheme is more numerically stable if we take the gradients of the image first and then estimate the reburring" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9867" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.088, + 0.877, + 0.34 + ], + "angle": 0, + "content": "
Method\nPhoton Level, MetricSRN [34]DHMPN [43]Deep-\nDeblur [24]MIMO-Unet+ [8]MPRNet [42]OursP4IP [28]PURE-LET [18]PhD-Net [29]
α = 10PSNR ↑20.7120.8921.1721.0421.0921.5719.2622.4923.00
SSIM ↑0.3860.3910.4010.3560.3930.4710.3480.4850.500
LPIPS-Alex ↓0.6810.7020.6560.7330.6780.5600.7330.5880.544
LPIPS-VGG ↓0.6460.6520.6270.6830.6410.5870.6740.6070.567
α = 20PSNR ↑20.7921.0321.3021.3621.2521.9319.4522.9423.63
SSIM ↑0.3920.4010.4100.3960.4050.4830.3530.5160.540
LPIPS-Alex ↓0.6830.6880.6660.6600.6670.5420.7260.5260.500
LPIPS-VGG ↓0.6390.6400.6210.6630.6310.5780.6680.5840.539
α = 40PSNR ↑20.8921.1521.4321.6321.4121.6220.1823.4824.38
SSIM ↑0.4090.4180.4250.4410.4280.5270.3720.5610.593
LPIPS-Alex ↓0.6770.6730.6730.5860.6470.4880.7060.4670.446
LPIPS-VGG ↓0.6290.6260.6120.6390.6130.5490.6600.5570.503
Blind?✓✓✓✓✓✓XXX
End-To-End Trained?✓✓✓✓✓XXX✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.355, + 0.894, + 0.399 + ], + "angle": 0, + "content": "Table 1. Performance on BSD100 Dataset with Synthetic Blur. ↑ represents metrics where higher means better and vice versa for ↓. LPIPS-Alex and LPIPS-VGG represent the perceptual measures from [45]. The best performing blind deconvolution method for each metric and photon level is shown in bold. The non-blind deconvolution methods are shown for reference in grey columns." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.411, + 0.877, + 0.63 + ], + "angle": 0, + "content": "
Method\nPhoton Level, MetricSRN\n[34]DHMPN\n[43]Deep-\nDeblur [24]MIMO-\nUNet+ [8]MPRNet\n[42]OursP4IP\n[28]PURE-LET\n[18]PhD-Net\n[29]
α = 10PSNR ↑20.2620.5020.9321.2521.0422.0119.9221.6322.41
SSIM ↑0.5100.5090.5240.5160.5330.6110.4630.5900.638
LPIPS-Alex ↓0.5070.5210.4960.5940.4790.3400.5460.3710.341
LPIPS-VGG ↓0.5310.5260.5180.6610.5110.4770.5550.5220.466
α = 20PSNR ↑20.4920.3921.1121.6421.3322.7219.5321.7922.78
SSIM ↑0.5230.5210.5360.5540.5510.6410.4420.6070.667
LPIPS-Alex ↓0.4960.5020.4920.4850.4590.3040.5330.3390.304
LPIPS-VGG ↓0.5150.5140.5010.6100.4930.4480.5540.5100.447
α = 40PSNR ↑20.5920.5021.2021.8821.5422.3217.3221.7822.96
SSIM ↑0.5350.5320.5450.5830.5670.6470.3620.6140.687
LPIPS-Alex ↓0.4910.4940.4940.4280.4470.2730.4870.3240.263
LPIPS-VGG ↓0.5060.5060.4930.5570.4790.4440.5600.5070.432
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.646, + 0.893, + 0.677 + ], + "angle": 0, + "content": "Table 2. Performance on Levin dataset with realistic camera shake blur [16]. The best performing blind deconvolution method for each metric and photon level is shown in bold and non-blind deconvolution methods are shown for reference in grey columns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.471, + 0.778 + ], + "angle": 0, + "content": "loss. This can be explained by the fact that unlike simulated data, the photon level is not known exactly and is estimated using the sensor data itself by a simple heuristic. For further details on how to use the sensor data, we refer the reader to [29]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.231, + 0.803 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In Table 4, we provide an ablation study by running the scheme for different number of key points i.e. \\( K = 4,6 \\) and 8 and without KTN \\( (K = 0) \\) on RealBlur dataset. Through this study, we demonstrate the effect of the Kernel Trajectory Network has on the iterative scheme. As expected, changing the search space for kernel estimation im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.893, + 0.777 + ], + "angle": 0, + "content": "proves the performance significantly across all metrics. Increasing the number of key points used for representing kernels also steadily improves the performance of the scheme, which can be explained by the fact there are larger degrees of freedom." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.62, + 0.814 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this paper, we use an iterative framework for the photon-limited blind deconvolution problem. More specifically, we use a non-blind solver which can deconvolve Poisson corrupted and blurred images given a blur kernel. To mitigate ill-posedness of the kernel estimation in such" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9868" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.897, + 0.329 + ], + "angle": 0, + "content": "
Method →SRN [34]DHMPN [43]Deep-Deblur [24]MIMO-UNet+ [8]MPRNet [42]Ours
Training → Photon lvl, MetricSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.
α = 10PSNR ↑25.7227.6425.7227.5825.9827.5726.2026.7826.2628.1626.61
SSIM ↑0.6120.7060.6030.6960.5770.7190.5310.5710.6410.7290.738
LPIPS-Alex ↓0.4380.3100.4540.3290.4410.2970.4840.3960.4010.2880.277
LPIPS-VGG ↓0.5080.4540.5090.4720.4960.4400.5490.5080.4960.4270.416
α = 20PSNR ↑25.3727.9125.4628.0225.9527.8126.6927.5326.5128.2927.23
SSIM ↑0.6580.7750.6550.7640.6360.7780.6300.6780.7150.7930.793
LPIPS-Alex ↓0.4260.2750.4290.2880.4270.2650.4010.3130.3600.2560.241
LPIPS-VGG ↓0.4920.4210.4960.4370.4850.4100.4950.4460.4660.4020.382
α = 40PSNR ↑25.6728.3425.7228.2726.2228.1327.2428.1426.8528.7227.11
SSIM ↑0.6650.7680.6530.7600.6260.7710.6750.7120.7160.7880.782
LPIPS-Alex ↓0.4150.2680.4180.2680.4180.2580.3470.2670.3430.2450.221
LPIPS-VGG ↓0.4820.4050.4810.4130.4700.3960.4570.4040.4440.3860.360
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.347, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Table 3. Performance on RealBlur-J Dataset with realistic blur [27]: Bold and underline refer to overall best performing method and best synthetic performance method. It should be noted that methods that are not trained end-to-end are usually at disadvantage when comparing on metrics like PSNR. However, it can be seen that our reconstruction is generally preferred by other perceptual metrics." + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.401, + 0.396, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.651, + 0.255, + 0.665 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.403, + 0.561, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.522, + 0.503, + 0.534 + ], + "angle": 0, + "content": "SRN" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.403, + 0.721, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.522, + 0.676, + 0.534 + ], + "angle": 0, + "content": "DHMPN" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.403, + 0.881, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.522, + 0.839, + 0.534 + ], + "angle": 0, + "content": "MPR-Net" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.536, + 0.556, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.654, + 0.53, + 0.666 + ], + "angle": 0, + "content": "MIMO-UNet+" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.536, + 0.726, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.654, + 0.671, + 0.666 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.536, + 0.881, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.654, + 0.852, + 0.666 + ], + "angle": 0, + "content": "Ground-Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.681, + 0.895, + 0.709 + ], + "angle": 0, + "content": "Figure 4. Qualitative example on the Real-Blur Dataset: For a more extensive set of results, we refer the reader to the supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.827 + ], + "angle": 0, + "content": "high noise regimes, we propose a novel low-dimensional representation to represent the motion blur. By using this novel low-dimensional and differentiable representation as a search space, we show state-of-the-art deconvolution performance and outperform end-to-end trained image restoration networks by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We believe this is a promising direction of research for both deblurring and general blind inverse problems i.e., inverse problems where the forward operator is not fully known. Future work could involve a supervised version" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.798 + ], + "angle": 0, + "content": "of this scheme which does not involve backpropagation through a network as it would greatly reduce the computational cost. Another research direction could be to apply this framework to the problem of spatially varying blur." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.826, + 0.649, + 0.843 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The work is supported, in part, by the United States National Science Foundation under the grants IIS-2133032 and ECCS-2030570." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9869" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.075, + 0.087, + 0.396, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.337, + 0.252, + 0.351 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.088, + 0.56, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.208, + 0.501, + 0.22 + ], + "angle": 0, + "content": "SRN" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.222, + 0.554, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.339, + 0.498, + 0.352 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.089, + 0.719, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.208, + 0.674, + 0.22 + ], + "angle": 0, + "content": "DHMPN" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.221, + 0.724, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.339, + 0.686, + 0.352 + ], + "angle": 0, + "content": "Non-Blind" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.089, + 0.879, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.208, + 0.836, + 0.22 + ], + "angle": 0, + "content": "MPR-Net" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.221, + 0.878, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.339, + 0.849, + 0.352 + ], + "angle": 0, + "content": "Ground-Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.893, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons on Photon-Limited Deblurring Dataset. Qualitative results on realistic blurred and photon-limited images from the Photon-Limited Deblurring dataset [29]. The inset image for \"Ours\" and \"Non-Blind\" represent the estimated and ground-truth kernel respectively. For a more extensive set of qualitative results, we refer the reader to the supplementary document." + }, + { + "type": "image", + "bbox": [ + 0.075, + 0.438, + 0.171, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.075, + 0.514, + 0.171, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.592, + 0.16, + 0.622 + ], + "angle": 0, + "content": "Two-Phase [40]" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.439, + 0.268, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.514, + 0.268, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.592, + 0.254, + 0.621 + ], + "angle": 0, + "content": "Sanghvi et.al [30]" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.439, + 0.365, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.514, + 0.365, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.599, + 0.337, + 0.612 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.439, + 0.464, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.514, + 0.464, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.592, + 0.447, + 0.62 + ], + "angle": 0, + "content": "GroundTruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Figure 6. Estimated Kernels for different methods: We show the estimated kernels from two examples from the PLDD dataset. Two-Phase [40] uses blur-only image \\( G(\\mathbf{y}) \\) as input, and ground-truth kernel is estimated using a point-source." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.717, + 0.174, + 0.732 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.743, + 0.47, + 0.826 + ], + "angle": 0, + "content": "[1] Chirag Agarwal, Shahin Khobahi, Arindam Bose, Mojtaba Soltanalian, and Dan Schonfeld. Deep-uri: A model-aware approach to blind deconvolution based on deep unfolded Richardson-Lucy network. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3299-3303, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.829, + 0.47, + 0.884 + ], + "angle": 0, + "content": "[2] Jérémy Anger, Maurício Delbracio, and Gabriele Facciolo. Efficient blind deblurring under high noise levels. In 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 123-128. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.887, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[3] Muhammad Asim, Fahad Shamshad, and Ali Ahmed." + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.743, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.439, + 0.871, + 0.668 + ], + "angle": 0, + "content": "
Key Points ↓\nPhoton Level ↓\nMetric →PSNRSSIMLPIPS-\nAlexLPIPS-\nVGG
α = 10K = 025.440.7190.2950.434
K = 425.900.7300.2860.423
K = 626.380.7350.2790.422
K = 826.730.7400.2720.414
α = 20K = 025.610.7670.2670.407
K = 426.930.7850.2420.385
K = 626.780.7840.2460.392
K = 827.260.7950.2400.384
α = 40K = 025.820.7590.2480.390
K = 426.880.7710.2280.374
K = 627.370.7800.2160.366
K = 827.290.7850.2140.361
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.679, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Table 4. Ablation Study for effect of Kernel Trajectory Network \\( T(\\cdot) \\): Reconstruction metrics for different number of key points. \\( K = 0 \\) represents the variant of the scheme which does not use the Kernel-Trajectory Network and estimates the kernel directly from Stage II." + }, + { + "type": "ref_text", + "bbox": [ + 0.532, + 0.782, + 0.892, + 0.823 + ], + "angle": 0, + "content": "Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging, 6:1493-1506, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.828, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[4] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE Transactions on Image Processing, 21(8):3502-3517, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[5] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV)," + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.782, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9870" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.316, + 0.107 + ], + "angle": 0, + "content": "pages 221-235. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.109, + 0.469, + 0.148 + ], + "angle": 0, + "content": "[6] Tony F Chan and Chiu-Kwong Wong. Total variation blind deconvolution. IEEE Transactions on Image Processing, 7(3):370-375, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.152, + 0.468, + 0.178 + ], + "angle": 0, + "content": "[7] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Transactions on Graphics, pages 1-8, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.181, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[8] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.251, + 0.469, + 0.306 + ], + "angle": 0, + "content": "[9] Mauricio Delbracio, Ignacio Garcia-Dorado, SungJoon Choi, Damien Kelly, and Peyman Milanfar. Polyblur: Removing mild blur by polynomial reburring. IEEE Transactions on Computational Imaging, 7:837-848, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[10] Mario AT Figueiredo and Jose M Bioucas-Dias. Deconvolution of Poissonian images using variable splitting and augmented Lagrangian optimization. In Proceedings of the IEEE/SP Workshop on Statistical Signal Processing, pages 733-736, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[11] Dong Gong, Jie Yang, Lingqiao Liu, Yanning Zhang, Ian Reid, Chunhua Shen, Anton Van Den Hengel, and Qinfeng Shi. From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2319-2328, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[12] Ankit Gupta, Neel Joshi, C Lawrence Zitnick, Michael Cohen, and Brian Curless. Single image deblurring using motion density functions. In European Conference on Computer Vision (ECCV), pages 171-184. Springer, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[13] Zachary T Harmany, Roummel F Marcia, and Rebecca M Willett. This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and practice. IEEE Transactions on Image Processing, 21(3):1084-1096, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[14] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jií Matas. DeblurGAN: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[15] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In The IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[16] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems (NeurIPS), 19, 2006. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[17] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1964-1971, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[18] Jizhou Li, Florian Luisier, and Thierry Blu. PURE-LET image deconvolution. IEEE Transactions on Image Processing, 27(1):92–105, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[19] Yuelong Li, Mohammad Tofighi, Vishal Monga, and Yonina C Eldar. An algorithm unrolling approach to deep image deblurring. In IEEE International Conference on Acoustics," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Speech and Signal Processing (ICASSP), pages 7675-7679. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[21] Leon B Lucy. An iterative technique for the rectification of observed distributions. The Astronomical Journal, 79:745, 1974. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[22] Zhiyuan Mao, Nicholas Chimitt, and Stanley H. Chan. Image reconstruction of static and dynamic scenes through anisoplanatic turbulence. IEEE Transactions on Computational Imaging, 6:1415-1428, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[23] Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[25] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3338-3347, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[26] William Hadley Richardson. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1):55-59, 1972. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[27] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Proceedings of the European Conference on Computer Vision (ECCV), pages 184-201. Springer, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[28] Arie Rond, Raja Giryes, and Michael Elad. Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation, 41:96-108, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[29] Yash Sanghvi, Abhiram Gnanasambandam, and Stanley H Chan. Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging (TCI), 8:851-864, 2022. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[30] Yash Sanghvi, Abhiram Gnanasambandam, Zhiyuan Mao, and Stanley H. Chan. Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging, 8:1051-1062, 2022. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[31] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(7):1439-1451, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[32] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. ACM Transactions on Graphics, 27(3):1-10, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9871" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[34] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[36] Zhou Wang and Alan C Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[37] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general U-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[38] Oliver Whyte, Josef Sivic, and Andrew Zisserman. Deblurring shaken and partially saturated images. International Journal of Computer Vision, 110(2):185-201, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[39] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International Journal of Computer Vision, 98(2):168-186, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[40] Li Xu and Jiaya Jia. Two-phase kernel estimation for robust motion deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), pages 157-170. Springer, 2010. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[43] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[44] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[46] Youjian Zhang, Chaoyue Wang, Stephen J Maybank, and Dacheng Tao. Exposure trajectory recovery from motion blur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7490-7504, 2021. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9872" + } + ] +] \ No newline at end of file diff --git a/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_origin.pdf b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..21a8c568dca2ba736e410f7dc4b5b95d7a00e6ca --- /dev/null +++ b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/6df892c8-2e1a-4696-b021-641afb67c068_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91a27a954ac0f97f9869bde2660c4da2ebdfc0c6aa6a049c65a1578f3ae0fdbd +size 3617950 diff --git a/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/full.md b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e9d7f66a1dd5091dab4fac054df8428b2d6f01f3 --- /dev/null +++ b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/full.md @@ -0,0 +1,359 @@ +# Structured Kernel Estimation for Photon-Limited Deconvolution + +Yash Sanghvi, Zhiyuan Mao, Stanley H. Chan +School of Electrical and Computer Engineering, Purdue University +{ysanghvi, maol14, stanchan}@purdue.edu + +# Abstract + +Images taken in a low light condition with the presence of camera shake suffer from motion blur and photon shot noise. While state-of-the-art image restoration networks show promising results, they are largely limited to well-illuminated scenes and their performance drops significantly when photon shot noise is strong. + +In this paper, we propose a new blur estimation technique customized for photon-limited conditions. The proposed method employs a gradient-based backpropagation method to estimate the blur kernel. By modeling the blur kernel using a low-dimensional representation with the key points on the motion trajectory, we significantly reduce the search space and improve the regularity of the kernel estimation problem. When plugged into an iterative framework, our novel low-dimensional representation provides improved kernel estimates and hence significantly better deconvolution performance when compared to end-to-end trained neural networks. The source code and pretrained mdoels are available at https://github.com/sanghviyashiitb/structured-kernel-cvpr23 + +# 1. Introduction + +Photon-Limited Blind Deconvolution: This paper studies the photon-limited blind deconvolution problem. Blind deconvolution refers to simultaneously recovering both the blur kernel and latent clean image from a blurred image and "photon-limited" refers to presence of photon-shot noise in images taken in low-illumination / short exposure. The corresponding forward model is as follows: + +$$ +\mathbf {y} = \operatorname {P o i s s o n} (\alpha \mathbf {h} * \mathbf {x}). \tag {1} +$$ + +In this equation, $\mathbf{y} \in \mathbb{R}^N$ is the blurred-noisy image, $\mathbf{x} \in \mathbb{R}^N$ is the latent clean image, and $\mathbf{h} \in \mathbb{R}^M$ is the blur kernel. We assume that $\mathbf{x}$ is normalized to $[0,1]$ and the entries of $\mathbf{h}$ are non-negative and sum up to 1. The constant $\alpha$ represents the average number of photons per pixel and is inversely proportional to the amount of Poisson noise. + +Deep Iterative Kernel Estimation: Blind image deconvolution has been studied for decades with many successful algorithms including the latest deep neural networks [8, 24, 34, 42, 43]. Arguably, the adaptation from the traditional Gaussian noise model to the photon-limited Poisson noise model can be done by retraining the existing networks with appropriate data. However, the restoration is not guaranteed to perform well because the end-to-end networks seldom explicitly take the forward image formation model into account. + +Recently, people have started to recognize the importance of blur kernel estimation for photon-limited conditions. One of these works is by Sanghvi et. al [30], where they propose an iterative kernel estimation method to backpropagate the gradient of an unsupervised reburring function, hence to update the blur kernel. However, as we can see in Figure 1, their performance is still limited when the photon shot noise is strong. + +Structured Kernel Estimation: Inspired by [30], we believe that the iterative kernel estimation process and the unsupervised reburbling loss are useful. However, instead of searching for the kernel directly (which can easily lead to local minima because the search space is too big), we propose to search in a low-dimensional space by imposing structure to the motion blur kernel. + +To construct such a low-dimensional space, we frame the blur kernel in terms of trajectory of the camera motion. Motion trajectory is often a continuous but irregular path in the two-dimensional plane. To specify the trajectory, we introduce the concept of key point estimation where we identify a set of anchor points of the kernel. By interpolating the path along these anchor points, we can then reproduce the kernel. Since the number of anchor points is significantly lower than the number of pixels in a kernel, we can reduce the dimensionality of the kernel estimation problem. + +The key contribution of this paper is as follows: We propose a new kernel estimation method called Kernel Trajectory Network (KTN). KTN models the blur kernel in a low-dimensional and differentiable space by specifying key points of the motion trajectory. Plugging this low-dimensional representation in an iterative framework im + +![](images/5079f0c6931497d4b8bd3e86b22c2fe6e199338969df4cebf626e158736c8ec6.jpg) +Blurred and Noisy + +![](images/b07d992d729dcf2e550b4bcfe3926aedbba947de08b7e374eeb6be4ce4953268.jpg) +MPR-Net [42] + +![](images/5940cc1d5f81fe8a449c784dfa1e71bab26ee5d425f9cb2aa5d1805b47700350.jpg) +Sanghvi et. al [30] +Figure 1. The proposed Kernel Trajectory Network (KTN) on real noisy blurred image from Photon-Limited Deblurring Dataset (PLDD) [29] The result corresponding to MPR-Net was generated by retraining the network with GoPro dataset [24] corrupted by Poisson noise. The inset images for "Sanghvi et. al" and "Ours" represent the estimated kernel and the inset image for "Ground-Truth" represents the kernel captured using a point source, as provided in PLDD. + +![](images/916b6af1dccd3a5c9a2a7e6a6a12c066e94ee6b9b71952b48b40eb136d5061c0.jpg) +Ours + +![](images/22178167ccf11a52cb61ffb0adf6436f6c94fe55393547dfb3246b9fd5da9016.jpg) +Ground-Truth + +proves the regularity of the kernel estimation problem. This leads to substantially better blur kernel estimates in photon-limited regimes where existing methods fail. + +# 2. Related Work + +Traditional Blind Deconvolution: Classical approaches to the (noiseless) blind deconvolution problem [6, 7, 22, 32, 40] use a joint optimization framework in which both the kernel and image are updated in an alternating fashion in order to minimize a cost function with kernel and image priors. For high noise regimes, a combination of $\ell_1 + \mathrm{TV}$ prior has been used in [2]. Levin et. al [16] pointed out that this joint optimization framework for the blind deconvolution problem favours the no-blur degenerate solution i.e. $(\mathbf{x}^{*},\mathbf{h}^{*}) = (\mathbf{y},\mathbf{I})$ where $\mathbf{I}$ is the identity operator. Some methods model the blur kernel in terms of the camera trajectory and then recover both the trajectory and the clean image using optimization [12, 38, 39] and supervised-learning techniques [11, 33, 46]. + +For the non-blind case, i.e., when the blur kernel is assumed to be known, the Poisson deconvolution problem has been studied for decades starting from Richardson-Lucy algorithm [21, 26]. More contemporary methods include Plug-and-Play [28, 29], PURE-LET [18], and MAP-based optimization methods [10, 13]. + +Deep Learning Methods. Recent years, many deep learning-based methods [5, 31] have been proposed for the blind image deblurring task. The most common strategy is to train a network end-to-end on large-scale datasets, such as the GoPro [24] and the RealBlur [27] datasets. Notably, many recent works [8, 24, 34, 42, 43] improve the performance of deblurring networks by adopting the multiscale strategies, where the training follows a coarse-to-fine setting that resembles the iterative approach. Generative Adversarial Network (GAN) based deblurring methods [3,14,15,44] are also shown to produce visually appealing images. Zamir et al. [41] and Wang et al. [37] adapt the + +popular vision transformers to the image restoration problems and demonstrate competitive performance on the deblurring task. + +Neural Networks and Iterative Methods: While neural networks have shown state-of-the-art performance on the deblurring task, another class of methods incorporating iterative methods with deep learning have shown promising results. Algorithm unrolling [23], where an iterative method is unrolled for fixed iterations and trained end-to-end has been applied to image deblurring [1,19]. In SelfDeblur [25], authors use Deep-Image-Prior [35] to represent the image and blur kernel and obtain state-of-the-art blind deconvolution performance. + +# 3. Method + +# 3.1. Kernel as Structured Motion Estimation + +Camera motion blur can be modeled as a latent clean image $\mathbf{x}$ convolved with a blur kernel $\mathbf{h}$ . If we assume the blur kernel lies in a window of size $32 \times 32$ , then $\mathbf{h} \in \mathbb{R}^{1024}$ . However, in this high dimensional space, only few entries of the blur kernel $\mathbf{h}$ are non-zero. Additionally, the kernel is generated from a two-dimensional trajectory which suggests that a simple sparsity prior is not sufficient. Given the difficulty of the photon-limited deconvolution problem, we need to impose a stronger prior on the kernel. To this end, we propose a differentiable and low-dimensional representation of the blur kernel, which we will use as the search space in our kernel estimation algorithm. + +We take the two-dimensional trajectory of the camera during the exposure time and divide it into $K$ "key points". Each key point represents either the start, the end or a change in direction of the camera trajectory as seen in Figure 2. Given the $K$ key points as points mapped out in $x - y$ space, we can interpolate them using cubic splines to form a continuous trajectory in 2D. To convert this continuous trajectory to an equivalent blur kernel, we assume a point source image and move it through the given trajectory. The + +![](images/06516385496c988531e9722e022d28179d445ffe9c2380da388f80478ecb0ecc.jpg) +Figure 2. Blur as Structured Motion Estimation: In our formulation, we view the blur kernel as the continuous camera trajectory reduced to $K$ key points, as shown in top half of the figure. We learn a differentiable representation from the vectorized $K$ key points to a blur kernel using a neural network. This lower dimensional and differentiable representation is leveraged to estimate a better blur kernel and avoiding local minima during inference. + +resulting frames are then averaged to give the corresponding blur kernel as shown in Figure 2. + +Given the formulation of blur kernel $\mathbf{h}$ in terms of $K$ key points, we now need to put this representation to a differentiable form since we intend to use it in an iterative scheme. To achieve this, we learn the transformations from the key points to the blur kernels using a neural network, which will be referred to as Kernel-Trajectory Network (KTN), and represent it using a differentiable function $T(\cdot)$ . Why differentiability is important to us will become clear to the reader in the next subsection. + +To train the Kernel-Trajectory Network, we generate training data as follows. First, for a fixed $K$ , we get $K$ key points by starting from $(0,0)$ and choosing the next $K - 1$ points by successively adding a random vector with a uniformly chosen random direction i.e. $U[0,360]$ and uniformly chosen length from $U[0,100 / (K - 1)]$ . Next, the set of key points are converted to a continuous smooth trajectory using bicubic interpolation. Then, we move a point source image through the given trajectory using the warpPerspective function in OpenCV, and average the resulting frames. + +Using the process defined above, we generate 60,000 blur kernels and their corresponding key point representations. For the Kernel-Trajectory Network $T(\cdot)$ , we take a + +U-Net like network with the first half replaced by 3-fully connected layers and train it with the generated data using $\ell_2$ -loss. For further architectural details on the KTN, we refer the reader to the supplementary document. + +# 3.2. Proposed Iterative Scheme + +We described in the previous subsection how to obtain a low-dimensional and differentiable representation $T(\cdot)$ for the blur kernel and now we are ready to present the full iterative scheme in detail. The proposed iterative scheme can be divided into three stages which are summarized as follows. We first generate an initial estimate of the direction and magnitude of the blur. This is used as initialization for a gradient-based scheme in Stage I which searches the appropriate kernel representation in the latent space $\mathbf{z}$ . This is followed by Stage II where we fine-tune the kernel obtained from Stage I using a similar process. + +Initialization Before starting the iterative scheme, we need a light-weight initialization method. This is important because of multiple local minima in the kernel estimation process. + +We choose to initialize the method with a rectilinear motion kernel, parameterized by length $\rho$ and orientation $\theta$ . To determine the length and orientation of the kernel, we use a minor variation of the kernel estimation in PolyBlur [9]. In this variation, the "blur-only image" $G(\mathbf{y})$ is used as the input and $\rho, \theta$ for the initial kernel are estimated using the minimum of the directional gradients. We refer the reader to Section II in the supplementary document for further details on the initialization. Explanation on the "blur-only image" is provided when we describe Stage I of the scheme. + +Stage I: Kernel Estimation in Latent Space Given an initial kernel, we choose initial latent $\mathbf{z}^0$ by dividing the rectilinear kernel into $K$ key points. Following the framework in [30], we run a gradient descent based scheme which optimizes the following cost function: + +$$ +\mathcal {L} (\mathbf {z}) \stackrel {\text {d e f}} {=} \underbrace {\| G (\mathbf {y}) - \mathbf {h} _ {\mathbf {z}} * F (\mathbf {y} , \mathbf {h} _ {\mathbf {z}})) \| _ {2} ^ {2}} _ {\text {R e b l u r i n g L o s s}}, \tag {2} +$$ + +where $\mathbf{h}_{\mathbf{z}} \stackrel{\mathrm{def}}{=} T(\mathbf{z})$ represents the kernel output from Kernel-Trajectory network $T(.)$ given the vectorized key points representation $\mathbf{z}$ . $F(.)$ represents the Poisson non-blind deconvolution solver which takes both noisy-blurred image and a blur kernel as the input. $G(\mathbf{y})$ represents a denoiser which is trained to remove only the noise from noisy-blurred image. The overall cost function represents reburring loss i.e. how well the kernel estimate and corresponding image estimate $\mathbf{h}_{\mathbf{z}} * F(\mathbf{y}, \mathbf{h}_{\mathbf{z}})$ match the blur-only image $G(\mathbf{y})$ . + +To minimize the cost function in (2), we use a simple gradient descent based iterative update for $\mathbf{z}$ as follows: + +$$ +\mathbf {z} ^ {k + 1} = \mathbf {z} ^ {k} - \delta \underbrace {\nabla_ {\mathbf {z}} \mathcal {L} \left(\mathbf {z} ^ {k}\right)} _ {\text {b a c k p r o p a g a t i o n}} \tag {3} +$$ + +![](images/e571aa7677f0e24e0426bfbfdf5b4f78fcaf734bd154df6d5a5cabf0a45236d3.jpg) +Figure 3. Flowchart describing first stage of the proposed scheme. We estimate the motion kernel of the blurry noisy image in lower dimensional latent space $\mathbf{z}$ where the blur kernel is represented by $T(\mathbf{z})$ and by minimizing the reburring loss $\mathcal{L}$ as defined in equation 2 + +where $\delta > 0$ is the step size and $\nabla_{\mathbf{z}}\mathcal{L}(\mathbf{z}^k)$ represents the gradient of the cost function $\mathcal{L}$ with respect to $\mathbf{z}$ evaluated $\mathbf{z}^k$ . It should be noted that the cost function is evaluated using the non-blind solver $F(.)$ and Kernel-Trajectory Network $T(.)$ - two neural network computations. Therefore, we can compute the gradient $\nabla_{\mathbf{z}}\mathcal{L}(\mathbf{z}^k)$ using autodifferentiation tools provided in PyTorch by backpropagating the gradients through $F(.)$ and then $T(.)$ + +Stage II: Kernel Fine-tuning In the second stage, using the kernel estimate of Stage I, we fine-tune the kernel by "opening up" the search space to the entire kernel vector instead of parametrizing by $T(\cdot)$ . Specifically, we optimize the following loss function + +$$ +\mathcal {L} (\mathbf {h}) \stackrel {\text {d e f}} {=} \| G (\mathbf {y}) - \mathbf {h} * F (\mathbf {y}, \mathbf {h})) \| _ {2} ^ {2} + \gamma \| \mathbf {h} \| _ {1}. \tag {4} +$$ + +Note the presence of the second term which acts as an $\ell_1$ -norm sparsity prior. Also the kernel vector $\mathbf{h}$ is being optimized instead of the latent key point vector $\mathbf{z}$ . Using variable splitting as used in Half-Quadratic Splitting (HQS), we convert the optimization problem in (4) to as follows: + +$$ +\mathcal {L} (\mathbf {h}, \mathbf {v}) = \| G (\mathbf {y}) - \mathbf {h} * F (\mathbf {y}, \mathbf {h})) \| _ {2} ^ {2} + \gamma \| \mathbf {h} \| _ {1} + \frac {\mu}{2} \| \mathbf {h} - \mathbf {v} \| _ {2} ^ {2} \tag {5} +$$ + +for some hyperparameter $\mu > 0$ . This leads us to the following iterative updates + +$$ +\mathbf {h} ^ {k + 1} = \mathbf {h} ^ {k} - \delta \cdot \left\{\nabla_ {\mathbf {h}} \mathcal {L} \left(\mathbf {h} ^ {k}\right) + \mu \left(\mathbf {h} ^ {k} - \mathbf {v} ^ {k}\right) \right\}, \tag {6} +$$ + +$$ +\begin{array}{l} \mathbf {v} ^ {k + 1} = \max \left(\left| \mathbf {h} ^ {k + 1} \right| - \gamma / \mu , 0\right) \cdot \operatorname {s i g n} \left(\mathbf {h} ^ {k + 1}\right) \\ \stackrel {\text {d e f}} {=} \mathcal {S} _ {\gamma / \mu} (\mathbf {h} ^ {k + 1}). \tag {7} \\ \end{array} +$$ + +# Algorithm 1 Iterative Poisson Deconvolution Scheme + +1: Input: Noisy-blurry $\mathbf{y}$ , Photon-Level $\alpha$ , denoiser $G(\cdot)$ , non-blind solver $F(\cdot)$ , Kernel-Trajectory-Network $T(\cdot)$ . +2: Initialize $\mathbf{z}^0$ using method described in Algorithm 1 in supplementary +3: for $k = 0,1,2,\dots$ do % Stage I begins here +4: $\mathbf{h}_{\mathbf{z}}^{k}\gets T(\mathbf{z}^{k})$ +5: $\mathcal{L}(\mathbf{z})\gets \| G(\mathbf{y}) - \mathbf{h}_{\mathbf{z}}^{k}\circledast F(\mathbf{y},\mathbf{h}_{\mathbf{z}}^{k})\|_{2}^{2}$ +6: Calculate $\nabla_{\mathbf{z}}\mathcal{L}(\mathbf{z}^k)$ using automatic differentiation +7: $\mathbf{z}^{k + 1}\gets \mathbf{z}^k -\delta \nabla_{\mathbf{z}}\mathcal{L}(\mathbf{z}^k)$ +8: end for +9: $\mathbf{h}^0, \mathbf{v}^0 \gets T(\mathbf{z}^\infty), \mu \gets 2.0, \gamma \gets 10^{-4}$ +10: for $k = 0,1,2,\dots$ do % Stage II begins here +11: $\mathcal{L}(\mathbf{h})\gets \| G(\mathbf{y}) - \mathbf{h}\circledast F(\mathbf{y},\mathbf{h})\| _2^2$ +12: Calculate $\nabla_{\mathbf{h}}\mathcal{L}(\mathbf{h}^k)$ using automatic differentiation +13: $\mathbf{h}^{k + 1}\gets \mathbf{h}^k -\delta \big(\nabla_{\mathbf{h}}\mathcal{L}(\mathbf{h}^k) + \mu (\mathbf{h}^k -\mathbf{v}^k)\big)$ +14: $\mathbf{v}^{k + 1}\gets \mathcal{S}_{\gamma /\mu}(\mathbf{h}^{k + 1})$ +15: $\mu \gets 1.01\mu$ +16: end for +17: return $\mathbf{h}^{(\infty)}$ and $\mathbf{x}^{(\infty)} = F(\mathbf{y},\mathbf{h}^{(\infty)})$ + +# 4. Experiments + +# 4.1. Training + +While our overall method is not end-to-end trained, it contains pre-trained components - the non-blind solver $F(\cdot)$ and denoiser $G(\cdot)$ . The architectures of $F(\cdot)$ and $G(\cdot)$ are inherited from PhD-Net [29] which takes as input a noisy-blurred image and kernel. However, for the denoiser $G(\cdot)$ , we fix the kernel input to identity operator since it is trained to remove only the noise from noisy-blurred image $\mathbf{y}$ . + +$F(.)$ and $G(.)$ are trained using synthetic data as follows. We take clean images from Flickr2K dataset [20] and the blur kernels from the code in Boracchi and Foi [4]. The blurred images are also corrupted using Poisson shot noise with photon levels $\alpha$ uniformly sampled from [1, 60]. The non-blind solver $F(.)$ is trained using kernel and noisy blurred image as input, and clean image as the target. The denoiser $G(.)$ is trained with similar procedure but with blurred-noisy image as the only input blur-only images as the target. The training processes, along with other experiments described in this paper are implemented using PyTorch on a NVIDIA Titan Xp GPU. + +For quantitative comparison of the method presented, we retrain the following state-of-the-art networks for Poisson Noise: Scale Recurrent Network (SRN) [34], Deep-Deblur [24], DHMPN [43], MPR-Net [42], and MIMO-UNet+ [8]. We perform this retraining in the following two different ways. First, we use synthetic data training as described for $F(.)$ and $G(.)$ . Second, for testing on realistic blur, we retrain the networks using the GoPro dataset [24] as it is often used to train neural networks in contemporary deblurring literature. We add the Poisson noise with the same distribution as the synthetic datasets to the blurred images. While retraining the networks, we use the respective loss functions from the original papers for sake of a fair comparison. + +# 4.2. Quantitative Comparison + +We quantitatively evaluate the proposed method on three different datasets, and compare it with state-of-the-art deblurring methods. In addition to the end-to-end trained methods described previously, we also compare our approach to the following Poisson deblurring methods: Poisson-Plug-and-Play [28], and PURE-LET [18]. Even though these methods assume the blur kernel to be known, we include them in the quantitative comparison since they are specifically designed for Poisson noise. For all of the methods described above, we compare the restored image's quality using PSNR, SSIM, and Learned Perceptual Image Patch Similarity (LPIPS-Alex, LPIPS-VGG) [45]. We include the latter as another metric in our evaluation since failure of MSE/SSIM to assess image quality has been well documented in [36, 45] + +BSD100: First, we evaluate our method on synthetic blur as follows. We collect 100 random images from the BSD-500 dataset, blur them synthetically with motion kernels from the Levin dataset [17] followed by adding Poisson noise at photon-levels $\alpha = 10, 20,$ and $40$ . The results of the quantitative evaluation are provided in Table 1. Since the blur is synthetic, ground-truth kernel is known and hence, can be used to simultaneously evaluate Poisson non-blind deblurring methods i.e, Poisson Plug-and-Play, PURE-LET, and PhD-Net. The last method is the non-blind solver $F(.)$ and serves as an upper bound on the deconvolution perfor + +mance. + +Levin Dataset: Next, we evaluate our method on the Levin dataset [17] which contains 32 real blurred images along with the ground truth kernels, as measured through a point source. We evaluate our method on this dataset with addition of Poisson noise at photon levels $\alpha = 10, 20$ and 40 and the results are shown in Table 2. For a fair comparison, end-to-end trained methods are retrained using synthetically blurred data (as described in Section IV-A) for evaluation on BSD100 and Levin dataset. + +RealBlur-J [27]: To demonstrate that our method is able to handle realistic blur, we evaluate our performance on randomly selected 50 patches of size $256 \times 256$ from the Real-Blur-J [27] dataset. Note that we reduce the size of the tested image because our method is based on a single-blur convolutional model. Such model may not be applicable for a large image with spatially varying blur and local motion of objects. However, for a smaller patch of a larger image, the single-blur-kernel model of deconvolution is a much more valid assumption. + +To ensure a fair comparison, we evaluate end-to-end networks by retraining on both the synthetic and GoPro dataset. As shown in Table 3, we find that end-to-end networks perform consistently better on the RealBlur dataset when trained using the GoPro dataset instead of synthetic blur. This can be explained by the fact both GoPro and RealBlur have realistic blur which is not necessarily captured by a single blur convolutional model. + +# 4.3. Qualitative Comparison + +Color Reconstruction We show reconstructions on examples from the real-blur dataset in Figure 4. While our method is grayscale, we perform colour reconstruction by estimating the kernel from the luminance-channel. Given the estimated kernel, we deblur each channel of the image using the non-blind solver and then combine the different channels into a single RGB-image. Note that all qualitative examples in this paper for end-to-end trained networks are trained using the GoPro dataset, since they provide the better visual result. + +Photon-Limited Deblurring Dataset We also show qualitative examples from photon-limited deblurring dataset [29] which contains 30 images' raw sensor data, blurred by camera shake and taken in extremely low-illumination. For reconstructing these images, we take the average of the R, G, B channels of the Bayer patter image, average it and then reconstruct it using the given method. The qualitative results for this dataset can be found in Figure 5. We also show the estimated kernels, along with estimated kernels from [30,40], in Figure 6. + +However, instead of using the reburring loss directly, we find the scheme is more numerically stable if we take the gradients of the image first and then estimate the reburring + +
Method +Photon Level, MetricSRN [34]DHMPN [43]Deep- +Deblur [24]MIMO-Unet+ [8]MPRNet [42]OursP4IP [28]PURE-LET [18]PhD-Net [29]
α = 10PSNR ↑20.7120.8921.1721.0421.0921.5719.2622.4923.00
SSIM ↑0.3860.3910.4010.3560.3930.4710.3480.4850.500
LPIPS-Alex ↓0.6810.7020.6560.7330.6780.5600.7330.5880.544
LPIPS-VGG ↓0.6460.6520.6270.6830.6410.5870.6740.6070.567
α = 20PSNR ↑20.7921.0321.3021.3621.2521.9319.4522.9423.63
SSIM ↑0.3920.4010.4100.3960.4050.4830.3530.5160.540
LPIPS-Alex ↓0.6830.6880.6660.6600.6670.5420.7260.5260.500
LPIPS-VGG ↓0.6390.6400.6210.6630.6310.5780.6680.5840.539
α = 40PSNR ↑20.8921.1521.4321.6321.4121.6220.1823.4824.38
SSIM ↑0.4090.4180.4250.4410.4280.5270.3720.5610.593
LPIPS-Alex ↓0.6770.6730.6730.5860.6470.4880.7060.4670.446
LPIPS-VGG ↓0.6290.6260.6120.6390.6130.5490.6600.5570.503
Blind?✓✓✓✓✓✓XXX
End-To-End Trained?✓✓✓✓✓XXX✓
+ +Table 1. Performance on BSD100 Dataset with Synthetic Blur. ↑ represents metrics where higher means better and vice versa for ↓. LPIPS-Alex and LPIPS-VGG represent the perceptual measures from [45]. The best performing blind deconvolution method for each metric and photon level is shown in bold. The non-blind deconvolution methods are shown for reference in grey columns. + +
Method +Photon Level, MetricSRN +[34]DHMPN +[43]Deep- +Deblur [24]MIMO- +UNet+ [8]MPRNet +[42]OursP4IP +[28]PURE-LET +[18]PhD-Net +[29]
α = 10PSNR ↑20.2620.5020.9321.2521.0422.0119.9221.6322.41
SSIM ↑0.5100.5090.5240.5160.5330.6110.4630.5900.638
LPIPS-Alex ↓0.5070.5210.4960.5940.4790.3400.5460.3710.341
LPIPS-VGG ↓0.5310.5260.5180.6610.5110.4770.5550.5220.466
α = 20PSNR ↑20.4920.3921.1121.6421.3322.7219.5321.7922.78
SSIM ↑0.5230.5210.5360.5540.5510.6410.4420.6070.667
LPIPS-Alex ↓0.4960.5020.4920.4850.4590.3040.5330.3390.304
LPIPS-VGG ↓0.5150.5140.5010.6100.4930.4480.5540.5100.447
α = 40PSNR ↑20.5920.5021.2021.8821.5422.3217.3221.7822.96
SSIM ↑0.5350.5320.5450.5830.5670.6470.3620.6140.687
LPIPS-Alex ↓0.4910.4940.4940.4280.4470.2730.4870.3240.263
LPIPS-VGG ↓0.5060.5060.4930.5570.4790.4440.5600.5070.432
+ +Table 2. Performance on Levin dataset with realistic camera shake blur [16]. The best performing blind deconvolution method for each metric and photon level is shown in bold and non-blind deconvolution methods are shown for reference in grey columns. + +loss. This can be explained by the fact that unlike simulated data, the photon level is not known exactly and is estimated using the sensor data itself by a simple heuristic. For further details on how to use the sensor data, we refer the reader to [29]. + +# 4.4. Ablation Study + +In Table 4, we provide an ablation study by running the scheme for different number of key points i.e. $K = 4,6$ and 8 and without KTN $(K = 0)$ on RealBlur dataset. Through this study, we demonstrate the effect of the Kernel Trajectory Network has on the iterative scheme. As expected, changing the search space for kernel estimation im + +proves the performance significantly across all metrics. Increasing the number of key points used for representing kernels also steadily improves the performance of the scheme, which can be explained by the fact there are larger degrees of freedom. + +# 5. Conclusion + +In this paper, we use an iterative framework for the photon-limited blind deconvolution problem. More specifically, we use a non-blind solver which can deconvolve Poisson corrupted and blurred images given a blur kernel. To mitigate ill-posedness of the kernel estimation in such + +
Method →SRN [34]DHMPN [43]Deep-Deblur [24]MIMO-UNet+ [8]MPRNet [42]Ours
Training → Photon lvl, MetricSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.
α = 10PSNR ↑25.7227.6425.7227.5825.9827.5726.2026.7826.2628.1626.61
SSIM ↑0.6120.7060.6030.6960.5770.7190.5310.5710.6410.7290.738
LPIPS-Alex ↓0.4380.3100.4540.3290.4410.2970.4840.3960.4010.2880.277
LPIPS-VGG ↓0.5080.4540.5090.4720.4960.4400.5490.5080.4960.4270.416
α = 20PSNR ↑25.3727.9125.4628.0225.9527.8126.6927.5326.5128.2927.23
SSIM ↑0.6580.7750.6550.7640.6360.7780.6300.6780.7150.7930.793
LPIPS-Alex ↓0.4260.2750.4290.2880.4270.2650.4010.3130.3600.2560.241
LPIPS-VGG ↓0.4920.4210.4960.4370.4850.4100.4950.4460.4660.4020.382
α = 40PSNR ↑25.6728.3425.7228.2726.2228.1327.2428.1426.8528.7227.11
SSIM ↑0.6650.7680.6530.7600.6260.7710.6750.7120.7160.7880.782
LPIPS-Alex ↓0.4150.2680.4180.2680.4180.2580.3470.2670.3430.2450.221
LPIPS-VGG ↓0.4820.4050.4810.4130.4700.3960.4570.4040.4440.3860.360
+ +Table 3. Performance on RealBlur-J Dataset with realistic blur [27]: Bold and underline refer to overall best performing method and best synthetic performance method. It should be noted that methods that are not trained end-to-end are usually at disadvantage when comparing on metrics like PSNR. However, it can be seen that our reconstruction is generally preferred by other perceptual metrics. + +![](images/d477d47fe21454f99a2b516ece3f9c65b8ef0f98124e3a0440e82af0fc1fcffe.jpg) +Input + +![](images/f1fe634750551cf68b9f01a42f99ad3e89f2879db2b3f929d0ffbb730c3dd874.jpg) +SRN + +![](images/445b30191774076574eb350b0cee49a750b59b4b536b7ac2b4c1a84ed2354d5f.jpg) +DHMPN + +![](images/c500ff00a4dc6b8f72224939a03e1d2a306db948f6d50a9a990b684600a55f92.jpg) +MPR-Net + +![](images/01eeeb6a7fb5af23b7dfbbee07c8502fe05cd54c6fbd3378b59038c7ec7ef2df.jpg) +MIMO-UNet+ +Figure 4. Qualitative example on the Real-Blur Dataset: For a more extensive set of results, we refer the reader to the supplementary document. + +![](images/d104b36501a56302ca101fc25328c14d316e4b1c6d137a66c29284e9abf6b2aa.jpg) +Ours + +![](images/74c1624539f8accb5d5a5f5bad0f34b1f5785a75b7d1c9fe9265c6123bea0c66.jpg) +Ground-Truth + +high noise regimes, we propose a novel low-dimensional representation to represent the motion blur. By using this novel low-dimensional and differentiable representation as a search space, we show state-of-the-art deconvolution performance and outperform end-to-end trained image restoration networks by a significant margin. + +We believe this is a promising direction of research for both deblurring and general blind inverse problems i.e., inverse problems where the forward operator is not fully known. Future work could involve a supervised version + +of this scheme which does not involve backpropagation through a network as it would greatly reduce the computational cost. Another research direction could be to apply this framework to the problem of spatially varying blur. + +# Acknowledgement + +The work is supported, in part, by the United States National Science Foundation under the grants IIS-2133032 and ECCS-2030570. + +![](images/c9b34d8706aede597c516a6f23b594d7796ad6d1f7c3e40d9dc239a1517364bc.jpg) +Input + +![](images/3c2621458c21df398131d0c54c8cdc6f1d1ea518fe497cdacf6fa5ccf1dc2acc.jpg) +SRN + +![](images/0a4cf93e9228ec0dbe29c259ef329f18dbff332fb2341659b203e89940509fe7.jpg) +Ours + +![](images/4b2fa0c32b8c9bb64cd66c36aed3fc4081b49f59f046cc81f3a42caed119caab.jpg) +DHMPN + +![](images/8ac61e326f309512c40e1a27f5ca5d357924882ba0b84335f1960e334e9bc9f9.jpg) +Non-Blind + +![](images/2caa0a0f7b704c6880cd861047c5281cb22bdccfa39c911cdd23e5b02baca836.jpg) +MPR-Net + +![](images/87cea74457ad089bdb0184845c96dcf2839bf4fc6f27c3150c45940d6f5c0dfe.jpg) +Ground-Truth + +![](images/04c5112c3539aadd764bc7636eaa06b7f689b2c7f75d1489a8c9abb9c875c3b2.jpg) +Figure 5. Visual comparisons on Photon-Limited Deblurring Dataset. Qualitative results on realistic blurred and photon-limited images from the Photon-Limited Deblurring dataset [29]. The inset image for "Ours" and "Non-Blind" represent the estimated and ground-truth kernel respectively. For a more extensive set of qualitative results, we refer the reader to the supplementary document. + +![](images/4eda77e898ff5560a734069390f76ec4f816bf259395d90b9054410b4b9992c5.jpg) +Two-Phase [40] +Figure 6. Estimated Kernels for different methods: We show the estimated kernels from two examples from the PLDD dataset. Two-Phase [40] uses blur-only image $G(\mathbf{y})$ as input, and ground-truth kernel is estimated using a point-source. + +![](images/796a2e733cc9ee6a57512c46ffdf93b664f08bdf8140adc0d2654802183b58b5.jpg) + +![](images/12ae9125af3590569a74e92f794ce0a247097cbe046f465aa11c3564cb9c2774.jpg) +Sanghvi et.al [30] + +![](images/966b37f00bf6d80d025a0a9e92f876cc927e7caa3e357175c7e73dc40a0f339b.jpg) + +![](images/42da63d1576319cf4b7b2539241e3b87fe573a3fa2f5e9a5be3516f60f7560c7.jpg) +Ours + +![](images/78ea8fb7ed6b39ae08f508e6aec9efc37a55d00159203a8099e8a9c516accaa8.jpg) + +![](images/41d4728cefd2b726977b20bb9c5ce3ba039abedcb952e7e837cbedf75990e067.jpg) +GroundTruth + +# References + +[1] Chirag Agarwal, Shahin Khobahi, Arindam Bose, Mojtaba Soltanalian, and Dan Schonfeld. Deep-uri: A model-aware approach to blind deconvolution based on deep unfolded Richardson-Lucy network. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3299-3303, 2020. 2 +[2] Jérémy Anger, Maurício Delbracio, and Gabriele Facciolo. Efficient blind deblurring under high noise levels. In 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 123-128. IEEE, 2019. 2 +[3] Muhammad Asim, Fahad Shamshad, and Ali Ahmed. + +
Key Points ↓ +Photon Level ↓ +Metric →PSNRSSIMLPIPS- +AlexLPIPS- +VGG
α = 10K = 025.440.7190.2950.434
K = 425.900.7300.2860.423
K = 626.380.7350.2790.422
K = 826.730.7400.2720.414
α = 20K = 025.610.7670.2670.407
K = 426.930.7850.2420.385
K = 626.780.7840.2460.392
K = 827.260.7950.2400.384
α = 40K = 025.820.7590.2480.390
K = 426.880.7710.2280.374
K = 627.370.7800.2160.366
K = 827.290.7850.2140.361
+ +Table 4. Ablation Study for effect of Kernel Trajectory Network $T(\cdot)$ : Reconstruction metrics for different number of key points. $K = 0$ represents the variant of the scheme which does not use the Kernel-Trajectory Network and estimates the kernel directly from Stage II. + +Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging, 6:1493-1506, 2020. 2 +[4] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE Transactions on Image Processing, 21(8):3502-3517, 2012. 5 +[5] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV), + +pages 221-235. Springer, 2016. 2 +[6] Tony F Chan and Chiu-Kwong Wong. Total variation blind deconvolution. IEEE Transactions on Image Processing, 7(3):370-375, 1998. 2 +[7] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Transactions on Graphics, pages 1-8, 2009. 2 +[8] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1, 2, 5, 6, 7 +[9] Mauricio Delbracio, Ignacio Garcia-Dorado, SungJoon Choi, Damien Kelly, and Peyman Milanfar. Polyblur: Removing mild blur by polynomial reburring. IEEE Transactions on Computational Imaging, 7:837-848, 2021. 3 +[10] Mario AT Figueiredo and Jose M Bioucas-Dias. Deconvolution of Poissonian images using variable splitting and augmented Lagrangian optimization. In Proceedings of the IEEE/SP Workshop on Statistical Signal Processing, pages 733-736, 2009. 2 +[11] Dong Gong, Jie Yang, Lingqiao Liu, Yanning Zhang, Ian Reid, Chunhua Shen, Anton Van Den Hengel, and Qinfeng Shi. From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2319-2328, 2017. 2 +[12] Ankit Gupta, Neel Joshi, C Lawrence Zitnick, Michael Cohen, and Brian Curless. Single image deblurring using motion density functions. In European Conference on Computer Vision (ECCV), pages 171-184. Springer, 2010. 2 +[13] Zachary T Harmany, Roummel F Marcia, and Rebecca M Willett. This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and practice. IEEE Transactions on Image Processing, 21(3):1084-1096, 2011. 2 +[14] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jií Matas. DeblurGAN: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 2 +[15] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In The IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 2 +[16] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems (NeurIPS), 19, 2006. 2, 6 +[17] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1964-1971, 2009. 5 +[18] Jizhou Li, Florian Luisier, and Thierry Blu. PURE-LET image deconvolution. IEEE Transactions on Image Processing, 27(1):92–105, 2017. 2, 5, 6 +[19] Yuelong Li, Mohammad Tofighi, Vishal Monga, and Yonina C Eldar. An algorithm unrolling approach to deep image deblurring. In IEEE International Conference on Acoustics, + +Speech and Signal Processing (ICASSP), pages 7675-7679. IEEE, 2019. 2 +[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 5 +[21] Leon B Lucy. An iterative technique for the rectification of observed distributions. The Astronomical Journal, 79:745, 1974. 2 +[22] Zhiyuan Mao, Nicholas Chimitt, and Stanley H. Chan. Image reconstruction of static and dynamic scenes through anisoplanatic turbulence. IEEE Transactions on Computational Imaging, 6:1415-1428, 2020. 2 +[23] Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 2 +[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2, 5, 6, 7 +[25] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3338-3347, 2020. 2 +[26] William Hadley Richardson. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1):55-59, 1972. 2 +[27] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Proceedings of the European Conference on Computer Vision (ECCV), pages 184-201. Springer, 2020. 2, 5, 7 +[28] Arie Rond, Raja Giryes, and Michael Elad. Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation, 41:96-108, 2016. 2, 5, 6 +[29] Yash Sanghvi, Abhiram Gnanasambandam, and Stanley H Chan. Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging (TCI), 8:851-864, 2022. 2, 4, 5, 6, 8 +[30] Yash Sanghvi, Abhiram Gnanasambandam, Zhiyuan Mao, and Stanley H. Chan. Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging, 8:1051-1062, 2022. 1, 2, 3, 5, 8 +[31] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(7):1439-1451, 2015. 2 +[32] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. ACM Transactions on Graphics, 27(3):1-10, 2008. 2 +[33] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion + +blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2 +[34] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 1, 2, 5, 6, 7 +[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 2 +[36] Zhou Wang and Alan C Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 5 +[37] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general U-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, June 2022. 2 +[38] Oliver Whyte, Josef Sivic, and Andrew Zisserman. Deblurring shaken and partially saturated images. International Journal of Computer Vision, 110(2):185-201, 2014. 2 +[39] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International Journal of Computer Vision, 98(2):168-186, 2012. 2 +[40] Li Xu and Jiaya Jia. Two-phase kernel estimation for robust motion deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), pages 157-170. Springer, 2010. 2, 5, 8 +[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 1, 2, 5, 6, 7 +[43] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1, 2, 5, 6, 7 +[44] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2 +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6 + +[46] Youjian Zhang, Chaoyue Wang, Stephen J Maybank, and Dacheng Tao. Exposure trajectory recovery from motion blur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7490-7504, 2021. 2 \ No newline at end of file diff --git a/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/images.zip b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5f3e227e3f4ba4e7d09a24471322e320b1f145d --- /dev/null +++ b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:015802db61681388ebc93739569ff5cf24fde87e3f753ccddf9859fb0550f733 +size 781914 diff --git a/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/layout.json b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9bf513ddc8de336181c159250dc7b6bb86e1ae7c --- /dev/null +++ b/2023/Structured Kernel Estimation for Photon-Limited Deconvolution/layout.json @@ -0,0 +1,9950 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 496, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 496, + 119 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 496, + 119 + ], + "type": "text", + "content": "Structured Kernel Estimation for Photon-Limited Deconvolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 144, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 144, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 144, + 459, + 185 + ], + "type": "text", + "content": "Yash Sanghvi, Zhiyuan Mao, Stanley H. Chan \nSchool of Electrical and Computer Engineering, Purdue University \n{ysanghvi, maol14, stanchan}@purdue.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 288, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 288, + 309 + ], + "type": "text", + "content": "Images taken in a low light condition with the presence of camera shake suffer from motion blur and photon shot noise. While state-of-the-art image restoration networks show promising results, they are largely limited to well-illuminated scenes and their performance drops significantly when photon shot noise is strong." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 310, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 290, + 490 + ], + "type": "text", + "content": "In this paper, we propose a new blur estimation technique customized for photon-limited conditions. The proposed method employs a gradient-based backpropagation method to estimate the blur kernel. By modeling the blur kernel using a low-dimensional representation with the key points on the motion trajectory, we significantly reduce the search space and improve the regularity of the kernel estimation problem. When plugged into an iterative framework, our novel low-dimensional representation provides improved kernel estimates and hence significantly better deconvolution performance when compared to end-to-end trained neural networks. The source code and pretrained mdoels are available at https://github.com/sanghviyashiitb/structured-kernel-cvpr23" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 511, + 128, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 128, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 128, + 522 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 530, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 615 + ], + "type": "text", + "content": "Photon-Limited Blind Deconvolution: This paper studies the photon-limited blind deconvolution problem. Blind deconvolution refers to simultaneously recovering both the blur kernel and latent clean image from a blurred image and \"photon-limited\" refers to presence of photon-shot noise in images taken in low-illumination / short exposure. The corresponding forward model is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 621, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 621, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 118, + 621, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\operatorname {P o i s s o n} (\\alpha \\mathbf {h} * \\mathbf {x}). \\tag {1}", + "image_path": "9e021d359d6cdfacc55259d3d1c83f6caba19dc55a34d5ce3875c06565cce70f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the blurred-noisy image, " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the latent clean image, and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is the blur kernel. We assume that " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is normalized to " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " and the entries of " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " are non-negative and sum up to 1. The constant " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " represents the average number of photons per pixel and is inversely proportional to the amount of Poisson noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 213, + 547, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 332 + ], + "type": "text", + "content": "Deep Iterative Kernel Estimation: Blind image deconvolution has been studied for decades with many successful algorithms including the latest deep neural networks [8, 24, 34, 42, 43]. Arguably, the adaptation from the traditional Gaussian noise model to the photon-limited Poisson noise model can be done by retraining the existing networks with appropriate data. However, the restoration is not guaranteed to perform well because the end-to-end networks seldom explicitly take the forward image formation model into account." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 336, + 547, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 547, + 432 + ], + "type": "text", + "content": "Recently, people have started to recognize the importance of blur kernel estimation for photon-limited conditions. One of these works is by Sanghvi et. al [30], where they propose an iterative kernel estimation method to backpropagate the gradient of an unsupervised reburring function, hence to update the blur kernel. However, as we can see in Figure 1, their performance is still limited when the photon shot noise is strong." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 517 + ], + "type": "text", + "content": "Structured Kernel Estimation: Inspired by [30], we believe that the iterative kernel estimation process and the unsupervised reburbling loss are useful. However, instead of searching for the kernel directly (which can easily lead to local minima because the search space is too big), we propose to search in a low-dimensional space by imposing structure to the motion blur kernel." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 520, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 639 + ], + "type": "text", + "content": "To construct such a low-dimensional space, we frame the blur kernel in terms of trajectory of the camera motion. Motion trajectory is often a continuous but irregular path in the two-dimensional plane. To specify the trajectory, we introduce the concept of key point estimation where we identify a set of anchor points of the kernel. By interpolating the path along these anchor points, we can then reproduce the kernel. Since the number of anchor points is significantly lower than the number of pixels in a kernel, we can reduce the dimensionality of the kernel estimation problem." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "The key contribution of this paper is as follows: We propose a new kernel estimation method called Kernel Trajectory Network (KTN). KTN models the blur kernel in a low-dimensional and differentiable space by specifying key points of the motion trajectory. Plugging this low-dimensional representation in an iterative framework im" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9863" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 69, + 151, + 167 + ], + "blocks": [ + { + "bbox": [ + 54, + 69, + 151, + 167 + ], + "lines": [ + { + "bbox": [ + 54, + 69, + 151, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 69, + 151, + 167 + ], + "type": "image", + "image_path": "5079f0c6931497d4b8bd3e86b22c2fe6e199338969df4cebf626e158736c8ec6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 169, + 137, + 179 + ], + "lines": [ + { + "bbox": [ + 69, + 169, + 137, + 179 + ], + "spans": [ + { + "bbox": [ + 69, + 169, + 137, + 179 + ], + "type": "text", + "content": "Blurred and Noisy" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 69, + 247, + 167 + ], + "blocks": [ + { + "bbox": [ + 151, + 69, + 247, + 167 + ], + "lines": [ + { + "bbox": [ + 151, + 69, + 247, + 167 + ], + "spans": [ + { + "bbox": [ + 151, + 69, + 247, + 167 + ], + "type": "image", + "image_path": "b07d992d729dcf2e550b4bcfe3926aedbba947de08b7e374eeb6be4ce4953268.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 169, + 229, + 179 + ], + "lines": [ + { + "bbox": [ + 170, + 169, + 229, + 179 + ], + "spans": [ + { + "bbox": [ + 170, + 169, + 229, + 179 + ], + "type": "text", + "content": "MPR-Net [42]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 69, + 345, + 167 + ], + "blocks": [ + { + "bbox": [ + 249, + 69, + 345, + 167 + ], + "lines": [ + { + "bbox": [ + 249, + 69, + 345, + 167 + ], + "spans": [ + { + "bbox": [ + 249, + 69, + 345, + 167 + ], + "type": "image", + "image_path": "5940cc1d5f81fe8a449c784dfa1e71bab26ee5d425f9cb2aa5d1805b47700350.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 169, + 331, + 180 + ], + "lines": [ + { + "bbox": [ + 263, + 169, + 331, + 180 + ], + "spans": [ + { + "bbox": [ + 263, + 169, + 331, + 180 + ], + "type": "text", + "content": "Sanghvi et. al [30]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "type": "text", + "content": "Figure 1. The proposed Kernel Trajectory Network (KTN) on real noisy blurred image from Photon-Limited Deblurring Dataset (PLDD) [29] The result corresponding to MPR-Net was generated by retraining the network with GoPro dataset [24] corrupted by Poisson noise. The inset images for \"Sanghvi et. al\" and \"Ours\" represent the estimated kernel and the inset image for \"Ground-Truth\" represents the kernel captured using a point source, as provided in PLDD." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 345, + 69, + 441, + 167 + ], + "blocks": [ + { + "bbox": [ + 345, + 69, + 441, + 167 + ], + "lines": [ + { + "bbox": [ + 345, + 69, + 441, + 167 + ], + "spans": [ + { + "bbox": [ + 345, + 69, + 441, + 167 + ], + "type": "image", + "image_path": "916b6af1dccd3a5c9a2a7e6a6a12c066e94ee6b9b71952b48b40eb136d5061c0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 170, + 405, + 178 + ], + "lines": [ + { + "bbox": [ + 383, + 170, + 405, + 178 + ], + "spans": [ + { + "bbox": [ + 383, + 170, + 405, + 178 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 442, + 69, + 539, + 167 + ], + "blocks": [ + { + "bbox": [ + 442, + 69, + 539, + 167 + ], + "lines": [ + { + "bbox": [ + 442, + 69, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 442, + 69, + 539, + 167 + ], + "type": "image", + "image_path": "22178167ccf11a52cb61ffb0adf6436f6c94fe55393547dfb3246b9fd5da9016.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 169, + 517, + 178 + ], + "lines": [ + { + "bbox": [ + 465, + 169, + 517, + 178 + ], + "spans": [ + { + "bbox": [ + 465, + 169, + 517, + 178 + ], + "type": "text", + "content": "Ground-Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 292 + ], + "type": "text", + "content": "proves the regularity of the kernel estimation problem. This leads to substantially better blur kernel estimates in photon-limited regimes where existing methods fail." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 306, + 134, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 134, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 134, + 319 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": "Traditional Blind Deconvolution: Classical approaches to the (noiseless) blind deconvolution problem [6, 7, 22, 32, 40] use a joint optimization framework in which both the kernel and image are updated in an alternating fashion in order to minimize a cost function with kernel and image priors. For high noise regimes, a combination of " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\ell_1 + \\mathrm{TV}" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": " prior has been used in [2]. Levin et. al [16] pointed out that this joint optimization framework for the blind deconvolution problem favours the no-blur degenerate solution i.e. " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}^{*},\\mathbf{h}^{*}) = (\\mathbf{y},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 327, + 287, + 495 + ], + "type": "text", + "content": " is the identity operator. Some methods model the blur kernel in terms of the camera trajectory and then recover both the trajectory and the clean image using optimization [12, 38, 39] and supervised-learning techniques [11, 33, 46]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 568 + ], + "type": "text", + "content": "For the non-blind case, i.e., when the blur kernel is assumed to be known, the Poisson deconvolution problem has been studied for decades starting from Richardson-Lucy algorithm [21, 26]. More contemporary methods include Plug-and-Play [28, 29], PURE-LET [18], and MAP-based optimization methods [10, 13]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Deep Learning Methods. Recent years, many deep learning-based methods [5, 31] have been proposed for the blind image deblurring task. The most common strategy is to train a network end-to-end on large-scale datasets, such as the GoPro [24] and the RealBlur [27] datasets. Notably, many recent works [8, 24, 34, 42, 43] improve the performance of deblurring networks by adopting the multiscale strategies, where the training follows a coarse-to-fine setting that resembles the iterative approach. Generative Adversarial Network (GAN) based deblurring methods [3,14,15,44] are also shown to produce visually appealing images. Zamir et al. [41] and Wang et al. [37] adapt the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "text", + "content": "popular vision transformers to the image restoration problems and demonstrate competitive performance on the deblurring task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 292, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 546, + 411 + ], + "type": "text", + "content": "Neural Networks and Iterative Methods: While neural networks have shown state-of-the-art performance on the deblurring task, another class of methods incorporating iterative methods with deep learning have shown promising results. Algorithm unrolling [23], where an iterative method is unrolled for fixed iterations and trained end-to-end has been applied to image deblurring [1,19]. In SelfDeblur [25], authors use Deep-Image-Prior [35] to represent the image and blur kernel and obtain state-of-the-art blind deconvolution performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 423, + 361, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 361, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 361, + 435 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 443, + 517, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 517, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 517, + 455 + ], + "type": "text", + "content": "3.1. Kernel as Structured Motion Estimation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": "Camera motion blur can be modeled as a latent clean image " + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": " convolved with a blur kernel " + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": ". If we assume the blur kernel lies in a window of size " + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{1024}" + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": ". However, in this high dimensional space, only few entries of the blur kernel " + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 304, + 462, + 545, + 605 + ], + "type": "text", + "content": " are non-zero. Additionally, the kernel is generated from a two-dimensional trajectory which suggests that a simple sparsity prior is not sufficient. Given the difficulty of the photon-limited deconvolution problem, we need to impose a stronger prior on the kernel. To this end, we propose a differentiable and low-dimensional representation of the blur kernel, which we will use as the search space in our kernel estimation algorithm." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "text", + "content": "We take the two-dimensional trajectory of the camera during the exposure time and divide it into " + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "text", + "content": " \"key points\". Each key point represents either the start, the end or a change in direction of the camera trajectory as seen in Figure 2. Given the " + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "text", + "content": " key points as points mapped out in " + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 304, + 606, + 546, + 714 + ], + "type": "text", + "content": " space, we can interpolate them using cubic splines to form a continuous trajectory in 2D. To convert this continuous trajectory to an equivalent blur kernel, we assume a point source image and move it through the given trajectory. The" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9864" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 64, + 292, + 305 + ], + "blocks": [ + { + "bbox": [ + 44, + 64, + 292, + 305 + ], + "lines": [ + { + "bbox": [ + 44, + 64, + 292, + 305 + ], + "spans": [ + { + "bbox": [ + 44, + 64, + 292, + 305 + ], + "type": "image", + "image_path": "06516385496c988531e9722e022d28179d445ffe9c2380da388f80478ecb0ecc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "type": "text", + "content": "Figure 2. Blur as Structured Motion Estimation: In our formulation, we view the blur kernel as the continuous camera trajectory reduced to " + }, + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "type": "text", + "content": " key points, as shown in top half of the figure. We learn a differentiable representation from the vectorized " + }, + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 309, + 288, + 387 + ], + "type": "text", + "content": " key points to a blur kernel using a neural network. This lower dimensional and differentiable representation is leveraged to estimate a better blur kernel and avoiding local minima during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "content": "resulting frames are then averaged to give the corresponding blur kernel as shown in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "text", + "content": "Given the formulation of blur kernel " + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "text", + "content": " key points, we now need to put this representation to a differentiable form since we intend to use it in an iterative scheme. To achieve this, we learn the transformations from the key points to the blur kernels using a neural network, which will be referred to as Kernel-Trajectory Network (KTN), and represent it using a differentiable function " + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 46, + 435, + 287, + 543 + ], + "type": "text", + "content": ". Why differentiability is important to us will become clear to the reader in the next subsection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": "To train the Kernel-Trajectory Network, we generate training data as follows. First, for a fixed " + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": ", we get " + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": " key points by starting from " + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": " and choosing the next " + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": " points by successively adding a random vector with a uniformly chosen random direction i.e. " + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "inline_equation", + "content": "U[0,360]" + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": " and uniformly chosen length from " + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "inline_equation", + "content": "U[0,100 / (K - 1)]" + }, + { + "bbox": [ + 46, + 544, + 287, + 675 + ], + "type": "text", + "content": ". Next, the set of key points are converted to a continuous smooth trajectory using bicubic interpolation. Then, we move a point source image through the given trajectory using the warpPerspective function in OpenCV, and average the resulting frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Using the process defined above, we generate 60,000 blur kernels and their corresponding key point representations. For the Kernel-Trajectory Network " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", we take a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "U-Net like network with the first half replaced by 3-fully connected layers and train it with the generated data using " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "-loss. For further architectural details on the KTN, we refer the reader to the supplementary document." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 125, + 454, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 125, + 454, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 454, + 138 + ], + "type": "text", + "content": "3.2. Proposed Iterative Scheme" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "type": "text", + "content": "We described in the previous subsection how to obtain a low-dimensional and differentiable representation " + }, + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "type": "text", + "content": " for the blur kernel and now we are ready to present the full iterative scheme in detail. The proposed iterative scheme can be divided into three stages which are summarized as follows. We first generate an initial estimate of the direction and magnitude of the blur. This is used as initialization for a gradient-based scheme in Stage I which searches the appropriate kernel representation in the latent space " + }, + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 144, + 545, + 276 + ], + "type": "text", + "content": ". This is followed by Stage II where we fine-tune the kernel obtained from Stage I using a similar process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 276, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 323 + ], + "type": "text", + "content": "Initialization Before starting the iterative scheme, we need a light-weight initialization method. This is important because of multiple local minima in the kernel estimation process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": "We choose to initialize the method with a rectilinear motion kernel, parameterized by length " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " and orientation " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": ". To determine the length and orientation of the kernel, we use a minor variation of the kernel estimation in PolyBlur [9]. In this variation, the \"blur-only image\" " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "G(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " is used as the input and " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\rho, \\theta" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " for the initial kernel are estimated using the minimum of the directional gradients. We refer the reader to Section II in the supplementary document for further details on the initialization. Explanation on the \"blur-only image\" is provided when we describe Stage I of the scheme." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "type": "text", + "content": "Stage I: Kernel Estimation in Latent Space Given an initial kernel, we choose initial latent " + }, + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^0" + }, + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "type": "text", + "content": " by dividing the rectilinear kernel into " + }, + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 443, + 545, + 502 + ], + "type": "text", + "content": " key points. Following the framework in [30], we run a gradient descent based scheme which optimizes the following cost function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 507, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 507, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 349, + 507, + 545, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {z}) \\stackrel {\\text {d e f}} {=} \\underbrace {\\| G (\\mathbf {y}) - \\mathbf {h} _ {\\mathbf {z}} * F (\\mathbf {y} , \\mathbf {h} _ {\\mathbf {z}})) \\| _ {2} ^ {2}} _ {\\text {R e b l u r i n g L o s s}}, \\tag {2}", + "image_path": "b65435aa502cd9a3ea01d287d0e9936301e5145077d43ee63e005cbbc3b23a94.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathbf{z}} \\stackrel{\\mathrm{def}}{=} T(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " represents the kernel output from Kernel-Trajectory network " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "T(.)" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " given the vectorized key points representation " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " represents the Poisson non-blind deconvolution solver which takes both noisy-blurred image and a blur kernel as the input. " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "G(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " represents a denoiser which is trained to remove only the noise from noisy-blurred image. The overall cost function represents reburring loss i.e. how well the kernel estimate and corresponding image estimate " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathbf{z}} * F(\\mathbf{y}, \\mathbf{h}_{\\mathbf{z}})" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": " match the blur-only image " + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "inline_equation", + "content": "G(\\mathbf{y})" + }, + { + "bbox": [ + 304, + 544, + 545, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "content": "To minimize the cost function in (2), we use a simple gradient descent based iterative update for " + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 681, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 681, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 370, + 681, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {k + 1} = \\mathbf {z} ^ {k} - \\delta \\underbrace {\\nabla_ {\\mathbf {z}} \\mathcal {L} \\left(\\mathbf {z} ^ {k}\\right)} _ {\\text {b a c k p r o p a g a t i o n}} \\tag {3}", + "image_path": "9820551689fc424f69fa7e8e6fc9d9c1ed4056573cde130d64b5966b56d998e4.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9865" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 500, + 259 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 500, + 259 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 500, + 259 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 500, + 259 + ], + "type": "image", + "image_path": "e571aa7677f0e24e0426bfbfdf5b4f78fcaf734bd154df6d5a5cabf0a45236d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "text", + "content": "Figure 3. Flowchart describing first stage of the proposed scheme. We estimate the motion kernel of the blurry noisy image in lower dimensional latent space " + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "text", + "content": " where the blur kernel is represented by " + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "inline_equation", + "content": "T(\\mathbf{z})" + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "text", + "content": " and by minimizing the reburring loss " + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "text", + "content": " as defined in equation 2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " is the step size and " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " represents the gradient of the cost function " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " evaluated " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^k" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": ". It should be noted that the cost function is evaluated using the non-blind solver " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " and Kernel-Trajectory Network " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "T(.)" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " - two neural network computations. Therefore, we can compute the gradient " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " using autodifferentiation tools provided in PyTorch by backpropagating the gradients through " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": " and then " + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "inline_equation", + "content": "T(.)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": "Stage II: Kernel Fine-tuning In the second stage, using the kernel estimate of Stage I, we fine-tune the kernel by \"opening up\" the search space to the entire kernel vector instead of parametrizing by " + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": ". Specifically, we optimize the following loss function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 484, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 484, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 75, + 484, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {h}) \\stackrel {\\text {d e f}} {=} \\| G (\\mathbf {y}) - \\mathbf {h} * F (\\mathbf {y}, \\mathbf {h})) \\| _ {2} ^ {2} + \\gamma \\| \\mathbf {h} \\| _ {1}. \\tag {4}", + "image_path": "bc6b7620332c220acb9ce8acc84f902387d7843c10452efa07f459549fb310c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "text", + "content": "Note the presence of the second term which acts as an " + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "text", + "content": "-norm sparsity prior. Also the kernel vector " + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "text", + "content": " is being optimized instead of the latent key point vector " + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 509, + 287, + 570 + ], + "type": "text", + "content": ". Using variable splitting as used in Half-Quadratic Splitting (HQS), we convert the optimization problem in (4) to as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 578, + 297, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 297, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 297, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {h}, \\mathbf {v}) = \\| G (\\mathbf {y}) - \\mathbf {h} * F (\\mathbf {y}, \\mathbf {h})) \\| _ {2} ^ {2} + \\gamma \\| \\mathbf {h} \\| _ {1} + \\frac {\\mu}{2} \\| \\mathbf {h} - \\mathbf {v} \\| _ {2} ^ {2} \\tag {5}", + "image_path": "58c13c2b0a7411858171246809cf92a90b956053262456a627dd76f3caf89f09.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "type": "text", + "content": "for some hyperparameter " + }, + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mu > 0" + }, + { + "bbox": [ + 46, + 619, + 287, + 643 + ], + "type": "text", + "content": ". This leads us to the following iterative updates" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 651, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 651, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 72, + 651, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {h} ^ {k + 1} = \\mathbf {h} ^ {k} - \\delta \\cdot \\left\\{\\nabla_ {\\mathbf {h}} \\mathcal {L} \\left(\\mathbf {h} ^ {k}\\right) + \\mu \\left(\\mathbf {h} ^ {k} - \\mathbf {v} ^ {k}\\right) \\right\\}, \\tag {6}", + "image_path": "8b7d4259930223c3a4d4e191631f673711f439a33e1009d58af84194948386f3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 667, + 287, + 702 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 667, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 75, + 667, + 287, + 702 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {v} ^ {k + 1} = \\max \\left(\\left| \\mathbf {h} ^ {k + 1} \\right| - \\gamma / \\mu , 0\\right) \\cdot \\operatorname {s i g n} \\left(\\mathbf {h} ^ {k + 1}\\right) \\\\ \\stackrel {\\text {d e f}} {=} \\mathcal {S} _ {\\gamma / \\mu} (\\mathbf {h} ^ {k + 1}). \\tag {7} \\\\ \\end{array}", + "image_path": "382652d0a1d32f7fd4184b96b354ea4d62672ea1b498499bef9c4a8bf9b55070.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 319, + 525, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 525, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 525, + 331 + ], + "type": "text", + "content": "Algorithm 1 Iterative Poisson Deconvolution Scheme" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 333, + 545, + 559 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "content": "1: Input: Noisy-blurry " + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "content": ", Photon-Level " + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "content": ", denoiser " + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "content": ", non-blind solver " + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "content": ", Kernel-Trajectory-Network " + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 311, + 333, + 545, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 369, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 369, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 369, + 545, + 392 + ], + "type": "text", + "content": "2: Initialize " + }, + { + "bbox": [ + 312, + 369, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^0" + }, + { + "bbox": [ + 312, + 369, + 545, + 392 + ], + "type": "text", + "content": " using method described in Algorithm 1 in supplementary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 392, + 495, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 392, + 495, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 392, + 495, + 403 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 312, + 392, + 495, + 403 + ], + "type": "inline_equation", + "content": "k = 0,1,2,\\dots" + }, + { + "bbox": [ + 312, + 392, + 495, + 403 + ], + "type": "text", + "content": " do % Stage I begins here" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 403, + 388, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 403, + 388, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 403, + 388, + 415 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 403, + 388, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathbf{z}}^{k}\\gets T(\\mathbf{z}^{k})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 415, + 474, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 474, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 474, + 426 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 415, + 474, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{z})\\gets \\| G(\\mathbf{y}) - \\mathbf{h}_{\\mathbf{z}}^{k}\\circledast F(\\mathbf{y},\\mathbf{h}_{\\mathbf{z}}^{k})\\|_{2}^{2}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 426, + 524, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 524, + 436 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 524, + 436 + ], + "type": "text", + "content": "6: Calculate " + }, + { + "bbox": [ + 312, + 426, + 524, + 436 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)" + }, + { + "bbox": [ + 312, + 426, + 524, + 436 + ], + "type": "text", + "content": " using automatic differentiation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 436, + 432, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 436, + 432, + 448 + ], + "spans": [ + { + "bbox": [ + 312, + 436, + 432, + 448 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 312, + 436, + 432, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{k + 1}\\gets \\mathbf{z}^k -\\delta \\nabla_{\\mathbf{z}}\\mathcal{L}(\\mathbf{z}^k)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 449, + 353, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 353, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 353, + 456 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 458, + 470, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 458, + 470, + 469 + ], + "spans": [ + { + "bbox": [ + 312, + 458, + 470, + 469 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 312, + 458, + 470, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^0, \\mathbf{v}^0 \\gets T(\\mathbf{z}^\\infty), \\mu \\gets 2.0, \\gamma \\gets 10^{-4}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 470, + 489, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 470, + 489, + 480 + ], + "spans": [ + { + "bbox": [ + 310, + 470, + 489, + 480 + ], + "type": "text", + "content": "10: for " + }, + { + "bbox": [ + 310, + 470, + 489, + 480 + ], + "type": "inline_equation", + "content": "k = 0,1,2,\\dots" + }, + { + "bbox": [ + 310, + 470, + 489, + 480 + ], + "type": "text", + "content": " do % Stage II begins here" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 310, + 480, + 465, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 480, + 465, + 491 + ], + "spans": [ + { + "bbox": [ + 310, + 480, + 465, + 491 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 310, + 480, + 465, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{h})\\gets \\| G(\\mathbf{y}) - \\mathbf{h}\\circledast F(\\mathbf{y},\\mathbf{h})\\| _2^2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 491, + 525, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 491, + 525, + 502 + ], + "spans": [ + { + "bbox": [ + 310, + 491, + 525, + 502 + ], + "type": "text", + "content": "12: Calculate " + }, + { + "bbox": [ + 310, + 491, + 525, + 502 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{h}}\\mathcal{L}(\\mathbf{h}^k)" + }, + { + "bbox": [ + 310, + 491, + 525, + 502 + ], + "type": "text", + "content": " using automatic differentiation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 502, + 500, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 502, + 500, + 514 + ], + "spans": [ + { + "bbox": [ + 310, + 502, + 500, + 514 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 310, + 502, + 500, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{k + 1}\\gets \\mathbf{h}^k -\\delta \\big(\\nabla_{\\mathbf{h}}\\mathcal{L}(\\mathbf{h}^k) + \\mu (\\mathbf{h}^k -\\mathbf{v}^k)\\big)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 310, + 514, + 418, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 514, + 418, + 526 + ], + "spans": [ + { + "bbox": [ + 310, + 514, + 418, + 526 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 310, + 514, + 418, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{k + 1}\\gets \\mathcal{S}_{\\gamma /\\mu}(\\mathbf{h}^{k + 1})" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 310, + 526, + 381, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 526, + 381, + 536 + ], + "spans": [ + { + "bbox": [ + 310, + 526, + 381, + 536 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 310, + 526, + 381, + 536 + ], + "type": "inline_equation", + "content": "\\mu \\gets 1.01\\mu" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 310, + 537, + 353, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 537, + 353, + 545 + ], + "spans": [ + { + "bbox": [ + 310, + 537, + 353, + 545 + ], + "type": "text", + "content": "16: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 310, + 546, + 461, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 546, + 461, + 559 + ], + "spans": [ + { + "bbox": [ + 310, + 546, + 461, + 559 + ], + "type": "text", + "content": "17: return " + }, + { + "bbox": [ + 310, + 546, + 461, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{(\\infty)}" + }, + { + "bbox": [ + 310, + 546, + 461, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 310, + 546, + 461, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(\\infty)} = F(\\mathbf{y},\\mathbf{h}^{(\\infty)})" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 586, + 386, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 386, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 386, + 601 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 609, + 370, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 370, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 370, + 622 + ], + "type": "text", + "content": "4.1. Training" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "While our overall method is not end-to-end trained, it contains pre-trained components - the non-blind solver " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and denoiser " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ". The architectures of " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " are inherited from PhD-Net [29] which takes as input a noisy-blurred image and kernel. However, for the denoiser " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", we fix the kernel input to identity operator since it is trained to remove only the noise from noisy-blurred image " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9866" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "G(.)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " are trained using synthetic data as follows. We take clean images from Flickr2K dataset [20] and the blur kernels from the code in Boracchi and Foi [4]. The blurred images are also corrupted using Poisson shot noise with photon levels " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " uniformly sampled from [1, 60]. The non-blind solver " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " is trained using kernel and noisy blurred image as input, and clean image as the target. The denoiser " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "G(.)" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " is trained with similar procedure but with blurred-noisy image as the only input blur-only images as the target. The training processes, along with other experiments described in this paper are implemented using PyTorch on a NVIDIA Titan Xp GPU." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "type": "text", + "content": "For quantitative comparison of the method presented, we retrain the following state-of-the-art networks for Poisson Noise: Scale Recurrent Network (SRN) [34], Deep-Deblur [24], DHMPN [43], MPR-Net [42], and MIMO-UNet+ [8]. We perform this retraining in the following two different ways. First, we use synthetic data training as described for " + }, + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "type": "inline_equation", + "content": "G(.)" + }, + { + "bbox": [ + 46, + 217, + 289, + 373 + ], + "type": "text", + "content": ". Second, for testing on realistic blur, we retrain the networks using the GoPro dataset [24] as it is often used to train neural networks in contemporary deblurring literature. We add the Poisson noise with the same distribution as the synthetic datasets to the blurred images. While retraining the networks, we use the respective loss functions from the original papers for sake of a fair comparison." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 382, + 190, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 190, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 190, + 396 + ], + "type": "text", + "content": "4.2. Quantitative Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 401, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 582 + ], + "type": "text", + "content": "We quantitatively evaluate the proposed method on three different datasets, and compare it with state-of-the-art deblurring methods. In addition to the end-to-end trained methods described previously, we also compare our approach to the following Poisson deblurring methods: Poisson-Plug-and-Play [28], and PURE-LET [18]. Even though these methods assume the blur kernel to be known, we include them in the quantitative comparison since they are specifically designed for Poisson noise. For all of the methods described above, we compare the restored image's quality using PSNR, SSIM, and Learned Perceptual Image Patch Similarity (LPIPS-Alex, LPIPS-VGG) [45]. We include the latter as another metric in our evaluation since failure of MSE/SSIM to assess image quality has been well documented in [36, 45]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "BSD100: First, we evaluate our method on synthetic blur as follows. We collect 100 random images from the BSD-500 dataset, blur them synthetically with motion kernels from the Levin dataset [17] followed by adding Poisson noise at photon-levels " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 10, 20," + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "40" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": ". The results of the quantitative evaluation are provided in Table 1. Since the blur is synthetic, ground-truth kernel is known and hence, can be used to simultaneously evaluate Poisson non-blind deblurring methods i.e, Poisson Plug-and-Play, PURE-LET, and PhD-Net. The last method is the non-blind solver " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and serves as an upper bound on the deconvolution perfor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 74, + 338, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 74, + 338, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 74, + 338, + 83 + ], + "type": "text", + "content": "mance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 85, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 191 + ], + "type": "text", + "content": "Levin Dataset: Next, we evaluate our method on the Levin dataset [17] which contains 32 real blurred images along with the ground truth kernels, as measured through a point source. We evaluate our method on this dataset with addition of Poisson noise at photon levels " + }, + { + "bbox": [ + 304, + 85, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\alpha = 10, 20" + }, + { + "bbox": [ + 304, + 85, + 545, + 191 + ], + "type": "text", + "content": " and 40 and the results are shown in Table 2. For a fair comparison, end-to-end trained methods are retrained using synthetically blurred data (as described in Section IV-A) for evaluation on BSD100 and Levin dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 193, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 546, + 312 + ], + "type": "text", + "content": "RealBlur-J [27]: To demonstrate that our method is able to handle realistic blur, we evaluate our performance on randomly selected 50 patches of size " + }, + { + "bbox": [ + 304, + 193, + 546, + 312 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 193, + 546, + 312 + ], + "type": "text", + "content": " from the Real-Blur-J [27] dataset. Note that we reduce the size of the tested image because our method is based on a single-blur convolutional model. Such model may not be applicable for a large image with spatially varying blur and local motion of objects. However, for a smaller patch of a larger image, the single-blur-kernel model of deconvolution is a much more valid assumption." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 313, + 546, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 408 + ], + "type": "text", + "content": "To ensure a fair comparison, we evaluate end-to-end networks by retraining on both the synthetic and GoPro dataset. As shown in Table 3, we find that end-to-end networks perform consistently better on the RealBlur dataset when trained using the GoPro dataset instead of synthetic blur. This can be explained by the fact both GoPro and RealBlur have realistic blur which is not necessarily captured by a single blur convolutional model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 418, + 441, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 441, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 441, + 431 + ], + "type": "text", + "content": "4.3. Qualitative Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 555 + ], + "type": "text", + "content": "Color Reconstruction We show reconstructions on examples from the real-blur dataset in Figure 4. While our method is grayscale, we perform colour reconstruction by estimating the kernel from the luminance-channel. Given the estimated kernel, we deblur each channel of the image using the non-blind solver and then combine the different channels into a single RGB-image. Note that all qualitative examples in this paper for end-to-end trained networks are trained using the GoPro dataset, since they provide the better visual result." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 677 + ], + "type": "text", + "content": "Photon-Limited Deblurring Dataset We also show qualitative examples from photon-limited deblurring dataset [29] which contains 30 images' raw sensor data, blurred by camera shake and taken in extremely low-illumination. For reconstructing these images, we take the average of the R, G, B channels of the Bayer patter image, average it and then reconstruct it using the given method. The qualitative results for this dataset can be found in Figure 5. We also show the estimated kernels, along with estimated kernels from [30,40], in Figure 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "However, instead of using the reburring loss directly, we find the scheme is more numerically stable if we take the gradients of the image first and then estimate the reburring" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9867" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 69, + 536, + 269 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 536, + 269 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 536, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 536, + 269 + ], + "type": "table", + "html": "
Method\nPhoton Level, MetricSRN [34]DHMPN [43]Deep-\nDeblur [24]MIMO-Unet+ [8]MPRNet [42]OursP4IP [28]PURE-LET [18]PhD-Net [29]
α = 10PSNR ↑20.7120.8921.1721.0421.0921.5719.2622.4923.00
SSIM ↑0.3860.3910.4010.3560.3930.4710.3480.4850.500
LPIPS-Alex ↓0.6810.7020.6560.7330.6780.5600.7330.5880.544
LPIPS-VGG ↓0.6460.6520.6270.6830.6410.5870.6740.6070.567
α = 20PSNR ↑20.7921.0321.3021.3621.2521.9319.4522.9423.63
SSIM ↑0.3920.4010.4100.3960.4050.4830.3530.5160.540
LPIPS-Alex ↓0.6830.6880.6660.6600.6670.5420.7260.5260.500
LPIPS-VGG ↓0.6390.6400.6210.6630.6310.5780.6680.5840.539
α = 40PSNR ↑20.8921.1521.4321.6321.4121.6220.1823.4824.38
SSIM ↑0.4090.4180.4250.4410.4280.5270.3720.5610.593
LPIPS-Alex ↓0.6770.6730.6730.5860.6470.4880.7060.4670.446
LPIPS-VGG ↓0.6290.6260.6120.6390.6130.5490.6600.5570.503
Blind?✓✓✓✓✓✓XXX
End-To-End Trained?✓✓✓✓✓XXX✓
", + "image_path": "27e028122713e790b09e6a3ee9f160fba3e408eb29884858493a79d64d775067.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 325, + 536, + 498 + ], + "blocks": [ + { + "bbox": [ + 46, + 281, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 316 + ], + "type": "text", + "content": "Table 1. Performance on BSD100 Dataset with Synthetic Blur. ↑ represents metrics where higher means better and vice versa for ↓. LPIPS-Alex and LPIPS-VGG represent the perceptual measures from [45]. The best performing blind deconvolution method for each metric and photon level is shown in bold. The non-blind deconvolution methods are shown for reference in grey columns." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 325, + 536, + 498 + ], + "lines": [ + { + "bbox": [ + 53, + 325, + 536, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 536, + 498 + ], + "type": "table", + "html": "
Method\nPhoton Level, MetricSRN\n[34]DHMPN\n[43]Deep-\nDeblur [24]MIMO-\nUNet+ [8]MPRNet\n[42]OursP4IP\n[28]PURE-LET\n[18]PhD-Net\n[29]
α = 10PSNR ↑20.2620.5020.9321.2521.0422.0119.9221.6322.41
SSIM ↑0.5100.5090.5240.5160.5330.6110.4630.5900.638
LPIPS-Alex ↓0.5070.5210.4960.5940.4790.3400.5460.3710.341
LPIPS-VGG ↓0.5310.5260.5180.6610.5110.4770.5550.5220.466
α = 20PSNR ↑20.4920.3921.1121.6421.3322.7219.5321.7922.78
SSIM ↑0.5230.5210.5360.5540.5510.6410.4420.6070.667
LPIPS-Alex ↓0.4960.5020.4920.4850.4590.3040.5330.3390.304
LPIPS-VGG ↓0.5150.5140.5010.6100.4930.4480.5540.5100.447
α = 40PSNR ↑20.5920.5021.2021.8821.5422.3217.3221.7822.96
SSIM ↑0.5350.5320.5450.5830.5670.6470.3620.6140.687
LPIPS-Alex ↓0.4910.4940.4940.4280.4470.2730.4870.3240.263
LPIPS-VGG ↓0.5060.5060.4930.5570.4790.4440.5600.5070.432
", + "image_path": "03af7ded342a7af4fe279da322890292351bf1b9d1a1b4d00d3dc0681fed9ed4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 511, + 546, + 536 + ], + "lines": [ + { + "bbox": [ + 46, + 511, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 546, + 536 + ], + "type": "text", + "content": "Table 2. Performance on Levin dataset with realistic camera shake blur [16]. The best performing blind deconvolution method for each metric and photon level is shown in bold and non-blind deconvolution methods are shown for reference in grey columns." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 616 + ], + "type": "text", + "content": "loss. This can be explained by the fact that unlike simulated data, the photon level is not known exactly and is estimated using the sensor data itself by a simple heuristic. For further details on how to use the sensor data, we refer the reader to [29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 141, + 635 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "In Table 4, we provide an ablation study by running the scheme for different number of key points i.e. " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "K = 4,6" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " and 8 and without KTN " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "(K = 0)" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " on RealBlur dataset. Through this study, we demonstrate the effect of the Kernel Trajectory Network has on the iterative scheme. As expected, changing the search space for kernel estimation im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 555, + 546, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 546, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 546, + 615 + ], + "type": "text", + "content": "proves the performance significantly across all metrics. Increasing the number of key points used for representing kernels also steadily improves the performance of the scheme, which can be explained by the fact there are larger degrees of freedom." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 632, + 379, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 379, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 379, + 644 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we use an iterative framework for the photon-limited blind deconvolution problem. More specifically, we use a non-blind solver which can deconvolve Poisson corrupted and blurred images given a blur kernel. To mitigate ill-posedness of the kernel estimation in such" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9868" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 548, + 260 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 548, + 260 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 548, + 260 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 548, + 260 + ], + "type": "table", + "html": "
Method →SRN [34]DHMPN [43]Deep-Deblur [24]MIMO-UNet+ [8]MPRNet [42]Ours
Training → Photon lvl, MetricSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.GoProSynth.
α = 10PSNR ↑25.7227.6425.7227.5825.9827.5726.2026.7826.2628.1626.61
SSIM ↑0.6120.7060.6030.6960.5770.7190.5310.5710.6410.7290.738
LPIPS-Alex ↓0.4380.3100.4540.3290.4410.2970.4840.3960.4010.2880.277
LPIPS-VGG ↓0.5080.4540.5090.4720.4960.4400.5490.5080.4960.4270.416
α = 20PSNR ↑25.3727.9125.4628.0225.9527.8126.6927.5326.5128.2927.23
SSIM ↑0.6580.7750.6550.7640.6360.7780.6300.6780.7150.7930.793
LPIPS-Alex ↓0.4260.2750.4290.2880.4270.2650.4010.3130.3600.2560.241
LPIPS-VGG ↓0.4920.4210.4960.4370.4850.4100.4950.4460.4660.4020.382
α = 40PSNR ↑25.6728.3425.7228.2726.2228.1327.2428.1426.8528.7227.11
SSIM ↑0.6650.7680.6530.7600.6260.7710.6750.7120.7160.7880.782
LPIPS-Alex ↓0.4150.2680.4180.2680.4180.2580.3470.2670.3430.2450.221
LPIPS-VGG ↓0.4820.4050.4810.4130.4700.3960.4570.4040.4440.3860.360
", + "image_path": "7e2b47fad625dfc0af46d2df94382e7f113097e26ed3c69b2c9675f605f5adc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 309 + ], + "type": "text", + "content": "Table 3. Performance on RealBlur-J Dataset with realistic blur [27]: Bold and underline refer to overall best performing method and best synthetic performance method. It should be noted that methods that are not trained end-to-end are usually at disadvantage when comparing on metrics like PSNR. However, it can be seen that our reconstruction is generally preferred by other perceptual metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 317, + 242, + 513 + ], + "blocks": [ + { + "bbox": [ + 47, + 317, + 242, + 513 + ], + "lines": [ + { + "bbox": [ + 47, + 317, + 242, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 242, + 513 + ], + "type": "image", + "image_path": "d477d47fe21454f99a2b516ece3f9c65b8ef0f98124e3a0440e82af0fc1fcffe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 515, + 156, + 526 + ], + "lines": [ + { + "bbox": [ + 133, + 515, + 156, + 526 + ], + "spans": [ + { + "bbox": [ + 133, + 515, + 156, + 526 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 319, + 343, + 410 + ], + "blocks": [ + { + "bbox": [ + 249, + 319, + 343, + 410 + ], + "lines": [ + { + "bbox": [ + 249, + 319, + 343, + 410 + ], + "spans": [ + { + "bbox": [ + 249, + 319, + 343, + 410 + ], + "type": "image", + "image_path": "f1fe634750551cf68b9f01a42f99ad3e89f2879db2b3f929d0ffbb730c3dd874.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 413, + 307, + 422 + ], + "lines": [ + { + "bbox": [ + 287, + 413, + 307, + 422 + ], + "spans": [ + { + "bbox": [ + 287, + 413, + 307, + 422 + ], + "type": "text", + "content": "SRN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 349, + 319, + 441, + 411 + ], + "blocks": [ + { + "bbox": [ + 349, + 319, + 441, + 411 + ], + "lines": [ + { + "bbox": [ + 349, + 319, + 441, + 411 + ], + "spans": [ + { + "bbox": [ + 349, + 319, + 441, + 411 + ], + "type": "image", + "image_path": "445b30191774076574eb350b0cee49a750b59b4b536b7ac2b4c1a84ed2354d5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 413, + 413, + 422 + ], + "lines": [ + { + "bbox": [ + 376, + 413, + 413, + 422 + ], + "spans": [ + { + "bbox": [ + 376, + 413, + 413, + 422 + ], + "type": "text", + "content": "DHMPN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 446, + 319, + 539, + 411 + ], + "blocks": [ + { + "bbox": [ + 446, + 319, + 539, + 411 + ], + "lines": [ + { + "bbox": [ + 446, + 319, + 539, + 411 + ], + "spans": [ + { + "bbox": [ + 446, + 319, + 539, + 411 + ], + "type": "image", + "image_path": "c500ff00a4dc6b8f72224939a03e1d2a306db948f6d50a9a990b684600a55f92.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 413, + 513, + 422 + ], + "lines": [ + { + "bbox": [ + 473, + 413, + 513, + 422 + ], + "spans": [ + { + "bbox": [ + 473, + 413, + 513, + 422 + ], + "type": "text", + "content": "MPR-Net" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 248, + 424, + 340, + 515 + ], + "blocks": [ + { + "bbox": [ + 248, + 424, + 340, + 515 + ], + "lines": [ + { + "bbox": [ + 248, + 424, + 340, + 515 + ], + "spans": [ + { + "bbox": [ + 248, + 424, + 340, + 515 + ], + "type": "image", + "image_path": "01eeeb6a7fb5af23b7dfbbee07c8502fe05cd54c6fbd3378b59038c7ec7ef2df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 517, + 324, + 527 + ], + "lines": [ + { + "bbox": [ + 263, + 517, + 324, + 527 + ], + "spans": [ + { + "bbox": [ + 263, + 517, + 324, + 527 + ], + "type": "text", + "content": "MIMO-UNet+" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 539, + 547, + 561 + ], + "lines": [ + { + "bbox": [ + 46, + 539, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 547, + 561 + ], + "type": "text", + "content": "Figure 4. Qualitative example on the Real-Blur Dataset: For a more extensive set of results, we refer the reader to the supplementary document." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 424, + 444, + 515 + ], + "blocks": [ + { + "bbox": [ + 353, + 424, + 444, + 515 + ], + "lines": [ + { + "bbox": [ + 353, + 424, + 444, + 515 + ], + "spans": [ + { + "bbox": [ + 353, + 424, + 444, + 515 + ], + "type": "image", + "image_path": "d104b36501a56302ca101fc25328c14d316e4b1c6d137a66c29284e9abf6b2aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 517, + 410, + 527 + ], + "lines": [ + { + "bbox": [ + 387, + 517, + 410, + 527 + ], + "spans": [ + { + "bbox": [ + 387, + 517, + 410, + 527 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 447, + 424, + 539, + 515 + ], + "blocks": [ + { + "bbox": [ + 447, + 424, + 539, + 515 + ], + "lines": [ + { + "bbox": [ + 447, + 424, + 539, + 515 + ], + "spans": [ + { + "bbox": [ + 447, + 424, + 539, + 515 + ], + "type": "image", + "image_path": "74c1624539f8accb5d5a5f5bad0f34b1f5785a75b7d1c9fe9265c6123bea0c66.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 517, + 521, + 527 + ], + "lines": [ + { + "bbox": [ + 463, + 517, + 521, + 527 + ], + "spans": [ + { + "bbox": [ + 463, + 517, + 521, + 527 + ], + "type": "text", + "content": "Ground-Truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 582, + 288, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 654 + ], + "type": "text", + "content": "high noise regimes, we propose a novel low-dimensional representation to represent the motion blur. By using this novel low-dimensional and differentiable representation as a search space, we show state-of-the-art deconvolution performance and outperform end-to-end trained image restoration networks by a significant margin." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We believe this is a promising direction of research for both deblurring and general blind inverse problems i.e., inverse problems where the forward operator is not fully known. Future work could involve a supervised version" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 632 + ], + "type": "text", + "content": "of this scheme which does not involve backpropagation through a network as it would greatly reduce the computational cost. Another research direction could be to apply this framework to the problem of spatially varying blur." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 654, + 397, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 397, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 397, + 667 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "The work is supported, in part, by the United States National Science Foundation under the grants IIS-2133032 and ECCS-2030570." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9869" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 45, + 68, + 242, + 264 + ], + "blocks": [ + { + "bbox": [ + 45, + 68, + 242, + 264 + ], + "lines": [ + { + "bbox": [ + 45, + 68, + 242, + 264 + ], + "spans": [ + { + "bbox": [ + 45, + 68, + 242, + 264 + ], + "type": "image", + "image_path": "c9b34d8706aede597c516a6f23b594d7796ad6d1f7c3e40d9dc239a1517364bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 266, + 154, + 277 + ], + "lines": [ + { + "bbox": [ + 131, + 266, + 154, + 277 + ], + "spans": [ + { + "bbox": [ + 131, + 266, + 154, + 277 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 249, + 69, + 342, + 162 + ], + "blocks": [ + { + "bbox": [ + 249, + 69, + 342, + 162 + ], + "lines": [ + { + "bbox": [ + 249, + 69, + 342, + 162 + ], + "spans": [ + { + "bbox": [ + 249, + 69, + 342, + 162 + ], + "type": "image", + "image_path": "3c2621458c21df398131d0c54c8cdc6f1d1ea518fe497cdacf6fa5ccf1dc2acc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 164, + 306, + 174 + ], + "lines": [ + { + "bbox": [ + 285, + 164, + 306, + 174 + ], + "spans": [ + { + "bbox": [ + 285, + 164, + 306, + 174 + ], + "type": "text", + "content": "SRN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 175, + 339, + 266 + ], + "blocks": [ + { + "bbox": [ + 247, + 175, + 339, + 266 + ], + "lines": [ + { + "bbox": [ + 247, + 175, + 339, + 266 + ], + "spans": [ + { + "bbox": [ + 247, + 175, + 339, + 266 + ], + "type": "image", + "image_path": "0a4cf93e9228ec0dbe29c259ef329f18dbff332fb2341659b203e89940509fe7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 268, + 304, + 278 + ], + "lines": [ + { + "bbox": [ + 280, + 268, + 304, + 278 + ], + "spans": [ + { + "bbox": [ + 280, + 268, + 304, + 278 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 348, + 70, + 440, + 162 + ], + "blocks": [ + { + "bbox": [ + 348, + 70, + 440, + 162 + ], + "lines": [ + { + "bbox": [ + 348, + 70, + 440, + 162 + ], + "spans": [ + { + "bbox": [ + 348, + 70, + 440, + 162 + ], + "type": "image", + "image_path": "4b2fa0c32b8c9bb64cd66c36aed3fc4081b49f59f046cc81f3a42caed119caab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 164, + 412, + 174 + ], + "lines": [ + { + "bbox": [ + 375, + 164, + 412, + 174 + ], + "spans": [ + { + "bbox": [ + 375, + 164, + 412, + 174 + ], + "type": "text", + "content": "DHMPN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 175, + 443, + 266 + ], + "blocks": [ + { + "bbox": [ + 351, + 175, + 443, + 266 + ], + "lines": [ + { + "bbox": [ + 351, + 175, + 443, + 266 + ], + "spans": [ + { + "bbox": [ + 351, + 175, + 443, + 266 + ], + "type": "image", + "image_path": "8ac61e326f309512c40e1a27f5ca5d357924882ba0b84335f1960e334e9bc9f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 268, + 419, + 278 + ], + "lines": [ + { + "bbox": [ + 374, + 268, + 419, + 278 + ], + "spans": [ + { + "bbox": [ + 374, + 268, + 419, + 278 + ], + "type": "text", + "content": "Non-Blind" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 446, + 70, + 537, + 162 + ], + "blocks": [ + { + "bbox": [ + 446, + 70, + 537, + 162 + ], + "lines": [ + { + "bbox": [ + 446, + 70, + 537, + 162 + ], + "spans": [ + { + "bbox": [ + 446, + 70, + 537, + 162 + ], + "type": "image", + "image_path": "2caa0a0f7b704c6880cd861047c5281cb22bdccfa39c911cdd23e5b02baca836.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 164, + 511, + 174 + ], + "lines": [ + { + "bbox": [ + 471, + 164, + 511, + 174 + ], + "spans": [ + { + "bbox": [ + 471, + 164, + 511, + 174 + ], + "type": "text", + "content": "MPR-Net" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 446, + 175, + 537, + 266 + ], + "blocks": [ + { + "bbox": [ + 446, + 175, + 537, + 266 + ], + "lines": [ + { + "bbox": [ + 446, + 175, + 537, + 266 + ], + "spans": [ + { + "bbox": [ + 446, + 175, + 537, + 266 + ], + "type": "image", + "image_path": "87cea74457ad089bdb0184845c96dcf2839bf4fc6f27c3150c45940d6f5c0dfe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 268, + 519, + 278 + ], + "lines": [ + { + "bbox": [ + 463, + 268, + 519, + 278 + ], + "spans": [ + { + "bbox": [ + 463, + 268, + 519, + 278 + ], + "type": "text", + "content": "Ground-Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 45, + 346, + 104, + 406 + ], + "blocks": [ + { + "bbox": [ + 46, + 294, + 546, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 546, + 328 + ], + "type": "text", + "content": "Figure 5. Visual comparisons on Photon-Limited Deblurring Dataset. Qualitative results on realistic blurred and photon-limited images from the Photon-Limited Deblurring dataset [29]. The inset image for \"Ours\" and \"Non-Blind\" represent the estimated and ground-truth kernel respectively. For a more extensive set of qualitative results, we refer the reader to the supplementary document." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 346, + 104, + 406 + ], + "lines": [ + { + "bbox": [ + 45, + 346, + 104, + 406 + ], + "spans": [ + { + "bbox": [ + 45, + 346, + 104, + 406 + ], + "type": "image", + "image_path": "04c5112c3539aadd764bc7636eaa06b7f689b2c7f75d1489a8c9abb9c875c3b2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 45, + 407, + 104, + 466 + ], + "blocks": [ + { + "bbox": [ + 45, + 407, + 104, + 466 + ], + "lines": [ + { + "bbox": [ + 45, + 407, + 104, + 466 + ], + "spans": [ + { + "bbox": [ + 45, + 407, + 104, + 466 + ], + "type": "image", + "image_path": "4eda77e898ff5560a734069390f76ec4f816bf259395d90b9054410b4b9992c5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 468, + 97, + 492 + ], + "lines": [ + { + "bbox": [ + 52, + 468, + 97, + 492 + ], + "spans": [ + { + "bbox": [ + 52, + 468, + 97, + 492 + ], + "type": "text", + "content": "Two-Phase [40]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 502, + 287, + 546 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 546 + ], + "type": "text", + "content": "Figure 6. Estimated Kernels for different methods: We show the estimated kernels from two examples from the PLDD dataset. Two-Phase [40] uses blur-only image " + }, + { + "bbox": [ + 46, + 502, + 287, + 546 + ], + "type": "inline_equation", + "content": "G(\\mathbf{y})" + }, + { + "bbox": [ + 46, + 502, + 287, + 546 + ], + "type": "text", + "content": " as input, and ground-truth kernel is estimated using a point-source." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 106, + 347, + 164, + 406 + ], + "blocks": [ + { + "bbox": [ + 106, + 347, + 164, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 164, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 164, + 406 + ], + "type": "image", + "image_path": "796a2e733cc9ee6a57512c46ffdf93b664f08bdf8140adc0d2654802183b58b5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 106, + 407, + 164, + 466 + ], + "blocks": [ + { + "bbox": [ + 106, + 407, + 164, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 407, + 164, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 164, + 466 + ], + "type": "image", + "image_path": "12ae9125af3590569a74e92f794ce0a247097cbe046f465aa11c3564cb9c2774.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 468, + 155, + 491 + ], + "lines": [ + { + "bbox": [ + 113, + 468, + 155, + 491 + ], + "spans": [ + { + "bbox": [ + 113, + 468, + 155, + 491 + ], + "type": "text", + "content": "Sanghvi et.al [30]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 165, + 347, + 223, + 406 + ], + "blocks": [ + { + "bbox": [ + 165, + 347, + 223, + 406 + ], + "lines": [ + { + "bbox": [ + 165, + 347, + 223, + 406 + ], + "spans": [ + { + "bbox": [ + 165, + 347, + 223, + 406 + ], + "type": "image", + "image_path": "966b37f00bf6d80d025a0a9e92f876cc927e7caa3e357175c7e73dc40a0f339b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 165, + 407, + 223, + 466 + ], + "blocks": [ + { + "bbox": [ + 165, + 407, + 223, + 466 + ], + "lines": [ + { + "bbox": [ + 165, + 407, + 223, + 466 + ], + "spans": [ + { + "bbox": [ + 165, + 407, + 223, + 466 + ], + "type": "image", + "image_path": "42da63d1576319cf4b7b2539241e3b87fe573a3fa2f5e9a5be3516f60f7560c7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 474, + 206, + 484 + ], + "lines": [ + { + "bbox": [ + 182, + 474, + 206, + 484 + ], + "spans": [ + { + "bbox": [ + 182, + 474, + 206, + 484 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 225, + 347, + 283, + 406 + ], + "blocks": [ + { + "bbox": [ + 225, + 347, + 283, + 406 + ], + "lines": [ + { + "bbox": [ + 225, + 347, + 283, + 406 + ], + "spans": [ + { + "bbox": [ + 225, + 347, + 283, + 406 + ], + "type": "image", + "image_path": "78ea8fb7ed6b39ae08f508e6aec9efc37a55d00159203a8099e8a9c516accaa8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 225, + 407, + 283, + 466 + ], + "blocks": [ + { + "bbox": [ + 225, + 407, + 283, + 466 + ], + "lines": [ + { + "bbox": [ + 225, + 407, + 283, + 466 + ], + "spans": [ + { + "bbox": [ + 225, + 407, + 283, + 466 + ], + "type": "image", + "image_path": "41d4728cefd2b726977b20bb9c5ce3ba039abedcb952e7e837cbedf75990e067.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 468, + 273, + 491 + ], + "lines": [ + { + "bbox": [ + 235, + 468, + 273, + 491 + ], + "spans": [ + { + "bbox": [ + 235, + 468, + 273, + 491 + ], + "type": "text", + "content": "GroundTruth" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 48, + 567, + 106, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 106, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 106, + 579 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 52, + 588, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 52, + 588, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 588, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 52, + 588, + 287, + 654 + ], + "type": "text", + "content": "[1] Chirag Agarwal, Shahin Khobahi, Arindam Bose, Mojtaba Soltanalian, and Dan Schonfeld. Deep-uri: A model-aware approach to blind deconvolution based on deep unfolded Richardson-Lucy network. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3299-3303, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 52, + 656, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 656, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 52, + 656, + 287, + 700 + ], + "type": "text", + "content": "[2] Jérémy Anger, Maurício Delbracio, and Gabriele Facciolo. Efficient blind deblurring under high noise levels. In 2019 11th International Symposium on Image and Signal Processing and Analysis (ISPA), pages 123-128. IEEE, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 52, + 702, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 702, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 52, + 702, + 287, + 712 + ], + "type": "text", + "content": "[3] Muhammad Asim, Fahad Shamshad, and Ali Ahmed." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + }, + { + "type": "table", + "bbox": [ + 321, + 347, + 533, + 529 + ], + "blocks": [ + { + "bbox": [ + 321, + 347, + 533, + 529 + ], + "lines": [ + { + "bbox": [ + 321, + 347, + 533, + 529 + ], + "spans": [ + { + "bbox": [ + 321, + 347, + 533, + 529 + ], + "type": "table", + "html": "
Key Points ↓\nPhoton Level ↓\nMetric →PSNRSSIMLPIPS-\nAlexLPIPS-\nVGG
α = 10K = 025.440.7190.2950.434
K = 425.900.7300.2860.423
K = 626.380.7350.2790.422
K = 826.730.7400.2720.414
α = 20K = 025.610.7670.2670.407
K = 426.930.7850.2420.385
K = 626.780.7840.2460.392
K = 827.260.7950.2400.384
α = 40K = 025.820.7590.2480.390
K = 426.880.7710.2280.374
K = 627.370.7800.2160.366
K = 827.290.7850.2140.361
", + "image_path": "c9282bfd3d213c6850cc8130c4f5b6e9f468382d589f252304641b92cdbcde87.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "lines": [ + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "type": "text", + "content": "Table 4. Ablation Study for effect of Kernel Trajectory Network " + }, + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "type": "text", + "content": ": Reconstruction metrics for different number of key points. " + }, + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 306, + 537, + 545, + 594 + ], + "type": "text", + "content": " represents the variant of the scheme which does not use the Kernel-Trajectory Network and estimates the kernel directly from Stage II." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 311, + 619, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 325, + 619, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 619, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 325, + 619, + 545, + 651 + ], + "type": "text", + "content": "Blind image deconvolution using deep generative priors. IEEE Transactions on Computational Imaging, 6:1493-1506, 2020. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 311, + 655, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 655, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 311, + 655, + 545, + 689 + ], + "type": "text", + "content": "[4] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE Transactions on Image Processing, 21(8):3502-3517, 2012. 5" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "type": "text", + "content": "[5] Ayan Chakrabarti. A neural approach to blind motion deblurring. In European Conference on Computer Vision (ECCV)," + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9870" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 193, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 193, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 193, + 84 + ], + "type": "text", + "content": "pages 221-235. Springer, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 86, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 86, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 86, + 287, + 117 + ], + "type": "text", + "content": "[6] Tony F Chan and Chiu-Kwong Wong. Total variation blind deconvolution. IEEE Transactions on Image Processing, 7(3):370-375, 1998. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 120, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 120, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 120, + 286, + 140 + ], + "type": "text", + "content": "[7] Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Transactions on Graphics, pages 1-8, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 143, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 287, + 196 + ], + "type": "text", + "content": "[8] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4641-4650, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 287, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 287, + 242 + ], + "type": "text", + "content": "[9] Mauricio Delbracio, Ignacio Garcia-Dorado, SungJoon Choi, Damien Kelly, and Peyman Milanfar. Polyblur: Removing mild blur by polynomial reburring. IEEE Transactions on Computational Imaging, 7:837-848, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "text", + "content": "[10] Mario AT Figueiredo and Jose M Bioucas-Dias. Deconvolution of Poissonian images using variable splitting and augmented Lagrangian optimization. In Proceedings of the IEEE/SP Workshop on Statistical Signal Processing, pages 733-736, 2009. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 365 + ], + "type": "text", + "content": "[11] Dong Gong, Jie Yang, Lingqiao Liu, Yanning Zhang, Ian Reid, Chunhua Shen, Anton Van Den Hengel, and Qinfeng Shi. From motion blur to motion flow: A deep learning solution for removing heterogeneous motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2319-2328, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 410 + ], + "type": "text", + "content": "[12] Ankit Gupta, Neel Joshi, C Lawrence Zitnick, Michael Cohen, and Brian Curless. Single image deblurring using motion density functions. In European Conference on Computer Vision (ECCV), pages 171-184. Springer, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[13] Zachary T Harmany, Roummel F Marcia, and Rebecca M Willett. This is SPIRAL-TAP: Sparse Poisson intensity reconstruction algorithms—theory and practice. IEEE Transactions on Image Processing, 21(3):1084-1096, 2011. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[14] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jií Matas. DeblurGAN: Blind motion deblurring using conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8183-8192, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[15] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In The IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 589 + ], + "type": "text", + "content": "[16] Anat Levin. Blind motion deblurring using image statistics. Advances in Neural Information Processing Systems (NeurIPS), 19, 2006. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 644 + ], + "type": "text", + "content": "[17] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1964-1971, 2009. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[18] Jizhou Li, Florian Luisier, and Thierry Blu. PURE-LET image deconvolution. IEEE Transactions on Image Processing, 27(1):92–105, 2017. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[19] Yuelong Li, Mohammad Tofighi, Vishal Monga, and Yonina C Eldar. An algorithm unrolling approach to deep image deblurring. In IEEE International Conference on Acoustics," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Speech and Signal Processing (ICASSP), pages 7675-7679. IEEE, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "type": "text", + "content": "[20] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[21] Leon B Lucy. An iterative technique for the rectification of observed distributions. The Astronomical Journal, 79:745, 1974. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "text", + "content": "[22] Zhiyuan Mao, Nicholas Chimitt, and Stanley H. Chan. Image reconstruction of static and dynamic scenes through anisoplanatic turbulence. IEEE Transactions on Computational Imaging, 6:1415-1428, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[23] Vishal Monga, Yuelong Li, and Yonina C Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "text", + "content": "[24] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 375 + ], + "type": "text", + "content": "[25] Dongwei Ren, Kai Zhang, Qilong Wang, Qinghua Hu, and Wangmeng Zuo. Neural blind deconvolution using deep priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3338-3347, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 377, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 409 + ], + "type": "text", + "content": "[26] William Hadley Richardson. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1):55-59, 1972. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[27] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Proceedings of the European Conference on Computer Vision (ECCV), pages 184-201. Springer, 2020. 2, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 510 + ], + "type": "text", + "content": "[28] Arie Rond, Raja Giryes, and Michael Elad. Poisson inverse problems by the plug-and-play scheme. Journal of Visual Communication and Image Representation, 41:96-108, 2016. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[29] Yash Sanghvi, Abhiram Gnanasambandam, and Stanley H Chan. Photon limited non-blind deblurring using algorithm unrolling. IEEE Transactions on Computational Imaging (TCI), 8:851-864, 2022. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "text", + "content": "[30] Yash Sanghvi, Abhiram Gnanasambandam, Zhiyuan Mao, and Stanley H. Chan. Photon-limited blind deconvolution using unsupervised iterative kernel estimation. IEEE Transactions on Computational Imaging, 8:1051-1062, 2022. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "text", + "content": "[31] Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(7):1439-1451, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[32] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. ACM Transactions on Graphics, 27(3):1-10, 2008. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[33] Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learning a convolutional neural network for non-uniform motion" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9871" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "blur removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 769-777, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[34] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8174-8182, 2018. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "text", + "content": "[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 239 + ], + "type": "text", + "content": "[36] Zhou Wang and Alan C Bovik. Mean squared error: Love it or leave it? a new look at signal fidelity measures. IEEE Signal Processing Magazine, 26(1):98-117, 2009. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "type": "text", + "content": "[37] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general U-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, June 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 338 + ], + "type": "text", + "content": "[38] Oliver Whyte, Josef Sivic, and Andrew Zisserman. Deblurring shaken and partially saturated images. International Journal of Computer Vision, 110(2):185-201, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "text", + "content": "[39] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International Journal of Computer Vision, 98(2):168-186, 2012. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "type": "text", + "content": "[40] Li Xu and Jiaya Jia. Two-phase kernel estimation for robust motion deblurring. In Proceedings of the European Conference on Computer Vision (ECCV), pages 157-170. Springer, 2010. 2, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "type": "text", + "content": "[41] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 547 + ], + "type": "text", + "content": "[42] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[43] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5978-5986, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[44] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Bjorn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2737-2746, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 117 + ], + "type": "text", + "content": "[46] Youjian Zhang, Chaoyue Wang, Stephen J Maybank, and Dacheng Tao. Exposure trajectory recovery from motion blur. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7490-7504, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9872" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_content_list.json b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1f2f48f3f8fe08449a50c8584cdc31ed7be66151 --- /dev/null +++ b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_content_list.json @@ -0,0 +1,1553 @@ +[ + { + "type": "text", + "text": "Structured Sparsity Learning for Efficient Video Super-Resolution", + "text_level": 1, + "bbox": [ + 148, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Xia $^{1}$ , Jingwen He $^{2}$ , Yulun Zhang $^{3}$ , Yitong Wang $^{4}$ , Yapeng Tian $^{5}$ , Wenming Yang $^{1*}$ , and Luc Van Gool $^{3}$", + "bbox": [ + 228, + 179, + 712, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, $^{2}$ Shanghai AI Laboratory, $^{3}$ ETH Zürich,", + "bbox": [ + 228, + 215, + 738, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 ByteDance Inc, 5 University of Texas at Dallas", + "bbox": [ + 292, + 233, + 674, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL.", + "bbox": [ + 75, + 300, + 472, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 638, + 209, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) aims to generate a high-resolution (HR) video from its corresponding low-resolution (LR) observation by filling in missing details. With the popularity of intelligent edge devices such as smartphones and small drones, performing VSR on these devices is in high demand. Although a variety of VSR networks [20, 24, 29, 44, 51] can achieve great performance, these models are usually difficult to be deployed on edge devices with limited computation and memory resources.", + "bbox": [ + 75, + 652, + 468, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate this issue, we explore a new direction for effective and efficient VSR. To reduce the redundancy of Conv kernels [4,5,36,38] obtaining a more efficient VSR network, we develop a neural network pruning scheme for the VSR task for the first time. Since structured pruning [14,23,46,57] (focusing on filter pruning) can achieve", + "bbox": [ + 75, + 789, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an actual acceleration [41,46] superior to unstructured pruning [11,12] (focusing on weight-element pruning), we adopt structured pruning principle to develop our VSR pruning scheme. Given a powerful VSR network, our pruning scheme can find submodels under presetting pruning rate without significantly compromising performance.", + "bbox": [ + 500, + 287, + 890, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Structured pruning is a general concept, and designing a concrete pruning scheme for VSR networks is challenging. (1) Recurrent networks are widely used in VSR models to extract temporal features, consisting of residual blocks (e.g., BasicVSR [2] has 60 residual blocks). However, it is hard to prune the residual blocks because the skip and residual connections ought to share the same indices [23] (Fig. 1 (a)). As shown in Fig. 1 (b), quite a few structured pruning schemes [23,34] do not prune the last Conv layer of the residual blocks, which restricts the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] introduce regularization and prune the same indices on skip and residual connections to keep channel alignment (local pruning scheme,i.e., each layer pruning the same ratio of filters). However, ASSL and SRPN still cannot achieve the potential of pruning residual blocks on recurrent networks. The recurrent networks take the previous output as later input (Fig. 2 (a)). This requires the pruned indices of the first and last Conv's in recurrent networks to be the same. But ASSL and SRPN cannot guarantee filter indices are aligned. Besides, many SR methods [45, 56] have shown that the information contained in front Conv layers can help the restoration feature extraction of later Conv layers. Thus, we design a Residual Sparsity Connection (RSC) for VSR recurrent networks, which preserves all channels of the input and output feature maps and selects the important channels for operation (Fig. 1 (d)). Compared with other pruning schemes [57, 58], RSC does not require the pruned indices of the first and last Conv's of recurrent networks to be the same, can preserve the information contained in all layers, and liberates the pruning space of the last Conv of the residual blocks without adding extra calculations. Notably, RSC can prune residual blocks globally (i.e., the filters in various layers are compared together to remove unimportant ones).", + "bbox": [ + 496, + 387, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22638", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c09dc091ece02e6550193c068b31ab201953219c277666bcd98d51b974ff2a0.jpg", + "image_caption": [ + "(d) Residual Sparsity Connection Scheme (RSC, Ours).", + "Figure 1. Illustration of different schemes for pruning residual blocks of recurrent networks. (a) Structure of the residual block in the VSR network. (b) The residual block pruning schemes [7, 23, 42] do not prune the last Conv. (c) ASSL [57] and SRPN [58] prunes the same indices on skip and residual connections to keep channel alignment, which abandons some channels of input and output feature maps. (d) RSC preserves all channels of input and output feature maps, which does not need to align the pruned indices on the first and last Conv's in recurrent networks, can fully use restoration information, and can prune the first and last Conv's of residual blocks without restrictions." + ], + "image_footnote": [], + "bbox": [ + 129, + 85, + 834, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(2) We observe that the upsampling network accounts for $22\\%$ of the total calculations in BasicVSR [2], which is necessary to be pruned to reduce redundancy. Since the pixel-shuffle [37] operation in VSR networks converts the channels to space, pruning the pixel-shuffle without any restrictions would cause the channel-space conversion to fail. Thus, we specially design a pixel-shuffle pruning scheme by taking four consecutive filters as the pruning unit for $2\\times$ pixel-shuffle. (3) Furthermore, we observe that the error of pruning VSR networks would accumulate with propagation steps increasing along with recurrent networks, which limits the efficiency and performance of pruning. Thus, we further introduce Temporal Finetuning (TF) to constrain the pruning error accumulation in recurrent networks. Overall, our main contributions are threefold:", + "bbox": [ + 75, + 426, + 468, + 651 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our work is necessary and timely. There is an urgent need to compress VSR models for deployment. To the best of our knowledge, we are one of the first to design a structured pruning scheme for VSR.", + "- We propose an integral VSR pruning scheme called Structured Sparsity Learning (SSL) for various components of VSR models, such as residual blocks, recurrent networks, and pixel-shuffle operations.", + "- We employ SSL to train VSR models, which surpass recent pruning schemes and lightweight VSR models." + ], + "bbox": [ + 94, + 654, + 468, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 821, + 218, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Super-Resolution", + "text_level": 1, + "bbox": [ + 76, + 847, + 294, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VSR models can exploit additional information from neighboring LR frames for restoration [3, 8, 16, 17, 28, 48,", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[49,52-54]. Earlier VSR methods [1,39,49] estimate the optical flow between LR frames and perform spatial warping for alignment. Later methods resort to a more sophisticated approach of implicit alignment. Instead of image-level motion alignment, TDAN [40] and EDVR [44] work at the feature level. TDAN [40] first adopted deformable Conv [6] in VSR to align the features of different frames. EDVR [44] extended TDAN by introducing coarse-to-fine deformable alignment and a new spatial-temporal attention fusion module. RSDN [16] adopted a recurrent detail-structural block and a hidden state adaptation module to reduce the effect of appearance changes and error accumulation. Recently, BasicVSR [2] found that bidirectional propagation coupled with a simple optical flow-based feature alignment can further improve performance. Similarly, Yi et al. [50] used the bidirectional propagation framework to exploit LR frames and estimated hidden states from the past, present, and future. To compress the VSR model, Xiao et al. designed a space-time knowledge distillation scheme [47]. However, these VSR methods require high computational costs impeding their application on resource-limited devices. Different from previous methods, we focus on designing SSL to compress VSR models by pruning redundant filters.", + "bbox": [ + 496, + 426, + 890, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Network Pruning", + "text_level": 1, + "bbox": [ + 500, + 787, + 669, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Network pruning [4, 5, 34, 36, 38] is widely used to remove a set of redundant parameters for network acceleration. Pruning methods can be divided into two branches, structured pruning [10, 14, 23, 46] and unstructured pruning [11, 12]. Structured pruning methods prune the network at the level of filters, channels, and even layers, which", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22639", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can obtain regular sparsity after pruning. This is beneficial for acceleration. In contrast, unstructured methods focus on pruning weights, leading up to much irregular sparsity. This is beneficial for compression but tends not to yield an actual acceleration [41, 46]. Specifically, Li et al. [23] applied the $\\mathrm{L}_1$ -norm to measure the importance of different filters and then removed the less important ones. Afterward, Liu et al. [30] added a sparsity-inducing penalty term on scaling factors of the batch normalization layers to enforce the channels with lower scaling factors to be the less informative ones. Recently, ASSL [57] and SRPN [58] utilized aligned structured sparsity learning for structured pruning of residual blocks. In addition, Luo et al. [32] developed a residual block pruning scheme for image classification using the Convs on skip connections. However, the residual blocks of VSR networks do not have such Convs. Lin et al. [26] conducted runtime neural network pruning according to the input image. Besides, Wang et al. [43] designed an unstructured pruning scheme for single image SR tasks by using sparse Conv to skip redundant computations. Since we cannot directly apply a general structured pruning method for VSR, we explore the properties of VSR networks and develop a VSR pruning scheme in this paper.", + "bbox": [ + 76, + 90, + 472, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 446, + 210, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 468, + 187, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 (a) shows VSR networks based on the bidirectional recurrent structures, such as BasicVSR [2]. Given a LR frame $I_{t}$ , the forward network concatenates $I_{t}$ and the previous hidden state $H_{F,t-1}$ to extract features from $I_{t}$ and aggregate the reference information from $H_{F,t-1}$ . Similarly, the backward network extracts features from $I_{t}$ and aggregates the reference information from the future hidden state $H_{B,t+1}$ . Note that both the forward and backward networks consist of numerous residual blocks. Then, the features generated by forward and backward networks are fed into the upsampling network, which consists of multiple pixel-shuffle operations and Convs, to obtain the recovered frame $SR_{t}$ . However, SOTA VSR networks [2,3,50] need massive computation and memory resources, limiting their deployment on resource-limited devices.", + "bbox": [ + 76, + 492, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To pursue more efficient VSR networks, we specially design a VSR structured pruning scheme called Structured Sparsity Learning (SSL), according to the properties of VSR networks. Specifically, SSL has three stages, including pretraining, pruning, and finetuning. In the pretraining stage, we train a powerful VSR network. Since current VSR networks do not use BatchNorm [15], we introduce a scaling factor in pretrained VSR models to tune the sparsity of each channel and filter. In the pruning stage, we select the unimportant filters according to the pruning criterion and apply sparsity-inducing regularization on corresponding scaling factors. In addition, we propose a Residual Spar", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sity Connection (RSC) scheme to liberate the restrictions on pruning residual blocks of recurrent networks and preserve all restoration information contained in channels of feature maps for better performance. Moreover, for the upsampling networks, we specially develop a pruning scheme for the pixel-shuffle operation to guarantee the accuracy of channel-space conversion after pruning. Besides, we observe that the error of the hidden state would be amplified with the propagation in recurrent networks after pruning. Thus, in the finetuning stage, we design Temporal Finetuning (TF) to alleviate the error accumulation.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Structured Sparsity Learning", + "text_level": 1, + "bbox": [ + 498, + 270, + 764, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structured Sparsity Learning (SSL) is a structured pruning scheme specially designed for VSR. It can reduce the redundancy of neural filters and obtain more efficient VSR submodels. Next, we will explain our SSL in detail.", + "bbox": [ + 496, + 291, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Scaling Factor. Structured pruning aims to remove Conv filters based on a designed importance criterion. In the classification task, quite a few works use scale parameters of BatchNorm [15] to control the throughput of each filter. Zero scale parameters make the value of corresponding channels vanish. As a result, they contribute nothing to the subsequent Convs and can be removed. By regularizing the scale parameter, we can assess and tune the importance of each filter. However, the BatchNorm is not useful for SR tasks [25], and SOTA VSR networks [2,3,50] do not utilize it. Therefore, it is infeasible to apply the existing pruning schemes directly. In our pruning scheme, as shown in Fig. 1 (d) and Fig. 2 (b), we multiply the scaling factors $\\gamma$ before or after Convs. Then, we perform regularization on scaling factors to enforce sparsity for pruning.", + "bbox": [ + 496, + 353, + 892, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Pruning Criterion and Regularization Form. To remove the redundant filters, we need to select unimportant scaling factors $\\gamma$ to induce sparsity. In previous works, ASSL [57] and SRPN [58] had to adopt a local pruning scheme (namely, scaling factors are only compared within the same layer, and each layer has the same pruning ratio) to guarantee that skip and residual connections keep the same number of filters and indices for the adding operation. Given that the importance of the Conv filters in various layers is different and our RSC does not have restrictions as ASSL, we can adopt the global pruning scheme (i.e., scaling factors of different layers are compared together).", + "bbox": [ + 496, + 580, + 893, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the pruning criterion, we adopt the simple and practical $\\mathbf{L}_1$ -norm. Specifically, given Conv kernel $\\boldsymbol{W_i} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times K_h \\times K_w}$ in the $i$ -th layer, we calculate the absolute weight sum of $k$ -th Conv filter $\\boldsymbol{W_i}[k,\\ldots] \\in \\mathbb{R}^{C_{in} \\times K_h \\times K_w}$ with $s_{i,k} = \\sum |\\boldsymbol{W_i}[k,\\ldots]|$ . In particular, for our RSC in Fig. 1 (d), we require to additionally prune the input channels for the first Conv, and calculate its $\\mathbf{L}_1$ -norm score with $s_{i,k}' = \\sum |\\boldsymbol{W_i}[:k,\\ldots]|$ , where $\\boldsymbol{W_i}[:k,\\ldots] \\in \\mathbb{R}^{C_{out} \\times K_h \\times K_w}$ . Moreover, for the Conv", + "bbox": [ + 496, + 762, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22640", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d2a50c42dc03702202f93e98afc7957105627ef1628089e5cf202066e17c24f1.jpg", + "image_caption": [ + "Figure 2. (a) The basic architecture of the VSR methods with the bidirectional recurrent network. The forward and backward networks both consist of numerous residual blocks. The upsampling networks contain multiple pixel-shuffle operations and Convs. (b) The pruning scheme for the pixel-shuffle. For the $2 \\times$ upsampling pixel-shuffle [37] operation, we take four channels with consecutive indices as the pruning unit to guarantee the accuracy of channel-space conversion after pruning. (c) The application of SSL on VSR models." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 464, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5af617e3fde59078383e8811d28ea970988c43f10e1e8b079b44aa28aa105ee4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 88, + 887, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "before the pixel-shuffle operation, we take four consecutive filters as a pruning unit and calculate the score with $s_{i,k} = \\sum |\\mathbf{W}_i[4k:4(k + 1),\\ldots ]|$ . Then, given the pruning ratio $p$ and the total number of filters or channels $N$ , we sort all their $\\mathbf{L}_1$ -norm scores $s$ together and choose the $N\\times p$ filters with the smallest $\\mathbf{L}_1$ -norm values as unimportant filters or channels, denoted as set $S$ (Fig. 2 (c)).", + "bbox": [ + 75, + 359, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After identifying the unimportant filters and channels set $S$ , we apply sparsity-inducing regularization (SIR) to the corresponding scaling factors, denoted as set $S_{sf}$ . It is notable that we do not enforce sparsity-inducing regularization to the important filters and channels since they will remain in the network. Specifically, we employ $L_{2}$ regularization on the scaling factors to enforce sparsity (Fig. 2 (c)):", + "bbox": [ + 75, + 465, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S I R} = \\alpha_ {\\gamma} \\sum_ {\\gamma \\in S _ {s f}} \\gamma^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 579, + 468, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is the scalar selected from $\\gamma \\in \\mathbb{R}^C$ corresponding to unimportant filters or channels; $\\alpha_{\\gamma}$ is a scalar. We increment $\\alpha_{\\gamma}$ by a presetting constant $\\Delta$ every $T_{1}$ iterations. When $\\alpha_{\\gamma}$ reaches the pre-defined upper limit $\\tau$ , we keep $\\alpha_{\\gamma}$ constant and continue training $T_{2}$ iterations.", + "bbox": [ + 75, + 613, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(3) Pruning Scheme for Residual Blocks of Recurrent Networks. VSR recurrent networks consist of residual blocks. Residual blocks are difficult to prune because the addition operations require the pruned filter indices between the skip and residual connections to be the same. As shown in Fig. 1 (b), quite a few pruning schemes [23, 30, 34] simply skipped the pruning of the last Conv in residual blocks, which restricted the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] pruned the last Conv in the residual block by using regularization. However, recurrent networks take the previous output as later input. Thus, pruning recurrent networks require the pruned indices of the first and last Convs to be the same. ASSL and SRPN cannot guarantee the indices of important filters", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the first and last Convs of recurrent networks are aligned. Besides, ASSL and SRPN have to remove some channels in output feature maps and adopt a local pruning scheme, which limits pruning space and restoration information utilization (RDN [56] and ESRGAN [45] have shown restoration information of front layers can guide feature extraction of later layers). To break the restriction of pruning and fully use restoration information contained in channels of front layers, we propose the RSC to prune residual blocks in recurrent networks (Fig. 1 (d)). As we can see, RSC preserves all channels of input $F_{i}^{\\prime}$ and output $F_{i+1}^{\\prime}$ in the residual blocks. For the first Conv, we select the important channels (the indices not in $S$ ) to participate in the Conv operation, which can be expressed as Eq. 2. After the last Conv, we obtained $F_{i+1}$ and add $F_{i+1}$ to $F_{i}^{\\prime}$ on the corresponding channel indices to obtain $F_{i+1}^{\\prime}$ (Eq. 3 and Eq. 4). Furthermore, we do not prune the $1 \\times 1$ Conv of upsampling networks (Fig. 2 (a)) to aggregate all preserved restoration information in $H_{F,t}, H_{B,t} \\in \\mathbb{R}^{C \\times H \\times W}$ .", + "bbox": [ + 496, + 359, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {i} = \\boldsymbol {F} _ {i} ^ {\\prime} \\otimes (\\gamma_ {j - 1} \\boldsymbol {W} _ {i} \\gamma_ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 654, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {i + 1} = \\boldsymbol {F} _ {i} \\otimes \\left(\\boldsymbol {W} _ {i + 1} \\gamma_ {j + 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 674, + 890, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {i + 1} ^ {\\prime} = \\boldsymbol {F} _ {i + 1} + \\boldsymbol {F} _ {i} ^ {\\prime}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 691, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ indicates Conv. $F_{i}^{\\prime}, F_{i+1}^{\\prime} \\in \\mathbb{R}^{C \\times H \\times W}$ are the input and output feature maps of the residual block, respectively. $F_{i}, F_{i+1} \\in \\mathbb{R}^{C_{p} \\times H \\times W}$ are intermediate feature maps. $W_{i}, W_{i+1} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times K_{h} \\times K_{w}}$ are weights of Conv kernels. $\\gamma_{j-1} \\in \\mathbb{R}^{C_{in}}$ and $\\gamma_{j}, \\gamma_{j+1} \\in \\mathbb{R}^{C_{out}}$ are scaling factors to apply sparsity-inducing regularization. $F_{i+1}^{\\prime}$ prunes some channels, and $F_{i}$ keeps all channels. In Eq. 4, $F_{i}^{\\prime}$ adds $F_{i+1}$ on corresponding kept channels. It is notable that our RSC does not introduce extra parameters and computational cost.", + "bbox": [ + 496, + 718, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(4) Pruning Scheme for Pixel-Shuffle. The upsampling network of the VSR network uses Conv to increase chan", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22641", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nels of feature maps and adopts the pixel-shuffle [37] operation to convert the channels to space realizing upsampling. As shown in Fig. 2 (b), given the input feature map $F_{i} \\in \\mathbb{R}^{C \\times H \\times W}$ , we expand its channels $4 \\times$ by a Conv with weight $W_{i}$ to obtain $F_{i+1}' \\in \\mathbb{R}^{4C \\times H \\times W}$ . Then, the pixel-shuffle operation takes four channels as a group to convert $F_{i+1}'$ to $F_{i+1} \\in \\mathbb{R}^{C \\times 2H \\times 2W}$ realize $2 \\times$ upsampling. We observe that, if we prune the Conv before pixel-shuffle without any restriction, the pruned feature maps will be spatially disordered after passing the pixel-shuffle operation and lead to performance drop. To address the problem, we specially design a strong and simple pruning scheme for the pixel-shuffle operation. Given the input feature map, we take four filters as a pruning unit to evaluate their importance and then impose the scaling factor $\\gamma_{j}$ on filters to enforce sparsity (described in (2) pruning criterion and regularization form):", + "bbox": [ + 76, + 90, + 472, + 333 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} _ {i} ^ {\\prime} = \\boldsymbol {W} _ {i} [ 4 k: 4 (k + 1), \\dots ] \\gamma_ {j} [ k ], k \\in [ 0, C _ {i n}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 349, + 468, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{i}\\in \\mathbb{R}^{4C_{in}\\times C_{in}\\times K_{h}\\times K_{w}}$ is the weights of Conv kernel. $\\gamma_{j}$ is the scaling factor.", + "bbox": [ + 75, + 369, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(5) Temporal Finetuning. We observe that the pruned VSR network generates a minor error in hidden state $H_{F}$ and $H_{B}$ (Fig. 2 (a)), which will be amplified as the hidden state propagates along with recurrent networks. To solve the issue, we introduce Temporal Finetuning (TF) by enforcing the hidden states of pruned networks to align with the accurate hidden states of unpruned networks:", + "bbox": [ + 75, + 401, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t f} = \\left\\| H _ {F, T} - H _ {F, T} ^ {\\prime} \\right\\| + \\left\\| H _ {B, 0} - H _ {B, 0} ^ {\\prime} \\right\\|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 510, + 468, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is the number of input frames, and $H_{F,T}$ and $H_{F,T}^{\\prime}$ are the final hidden states after $T$ frames forward propagation in pruned and original VSR networks, separately. Similarly, $H_{B,0}$ and $H_{B,0}^{\\prime}$ are the final hidden states after backward propagation in the pruned and original VSR networks.", + "bbox": [ + 75, + 529, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the whole VSR network, we use the Charbonnier loss [2, 44], which can be formulated as:", + "bbox": [ + 75, + 604, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sqrt {\\left\\| S R _ {t} - H R _ {t} \\right\\| ^ {2} + \\varepsilon^ {2}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 638, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varepsilon$ is set to $10^{-6}$ . $SR_{t}$ and $HR_{t}$ are $t$ -th reconstructed and HR frames, respectively. The overall loss function for pruned network finetuning is designed as:", + "bbox": [ + 75, + 681, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {t f}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 736, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Arm VSR Models with SSL", + "text_level": 1, + "bbox": [ + 76, + 760, + 326, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our SSL can be used for VSR networks. BasicVSR [2] and BasicVSR++ [3] are two strong SOTA VSR methods. The dense deformable Conv of BasicVSR++ requires reading a large amount of irregular memory data, which is unsuitable for deployment on resource-limited devices without GPUs. Besides, the Second-Order Grid Propagation of BasicVSR++ will force a delay of two frames, which further impedes its usage in real-time devices.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our study, we use BasicVSR for VSR pruning, which is more suitable for applications on edge devices. In addition, we further propose unidirectional BasicVSR (BasicVSR-uni), obtained by removing the backward network, for online inference. Since the SpyNet of BasicVSR is used for flow estimation, we do not apply our pruning scheme SSL on it. In the pruning stage, as shown in Fig. 2 (c), we first add the scaling factor to the Conv and residual blocks as described in Sec. 3.2. Then we use the pruning criterion to select unimportant filters globally and apply sparsity-inducing regularization to the corresponding scaling factor. Afterward, we remove the unimportant Conv filters and finetune the pruned VSR network with $T_{3}$ iterations. We provide more details in Alg. 1 of supplementary.", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 308, + 632, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 328, + 705, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt two widely used datasets for training: REDS [33] and Vimeo-90K [49]. For REDS, we use REDS4 containing 4 clips as our test set. Additionally, we adopt REDSval4 as our validation set, which contains 4 clips selected from the REDS validation set. The remaining clips of REDS are used for training. In addition, we utilize Vid4 [27], UDM10 [52], and Vimeo-90K-T [49] as test sets along with Vimeo-90K. We train and test models with $4 \\times$ downsampling using two degradations Bicubic (BI) and Blur Downsampling (BD) as BasicVSR did. For BI, the MATLAB function \"imresize\" is used for downsampling. For BD, we blur the HR images by a Gaussian filter with $\\sigma = 1.6$ , followed by a subsampling every four pixels.", + "bbox": [ + 496, + 348, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We pretrain the unidirectional BasicVSR (BasicVSR-uni) as done for BasicVSR. In sparsity-inducing regularization, the iterations $T_{1}$ and $T_{2}$ are set to 5 and 3, 375 separately. The scalars $\\Delta$ and $\\tau$ are set to $10^{-4}$ and 0.1, respectively. Note that we fix the parameters of the flow estimator in sparsity-inducing regularization. In the pruned VSR network finetuning, we set $T_{3}$ to 300,000. We adopt the Adam optimizer [22] and Cosine Annealing scheme [31]. The initial learning rate of the flow estimator is $2.5 \\times 10^{-5}$ . The learning rate for all other modules is $2 \\times 10^{-4}$ . The patch size of input LR frames is $64 \\times 64$ . Experiments are conducted on a server with PyTorch 1.10 and V100 GPUs.", + "bbox": [ + 496, + 545, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative and Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 498, + 729, + 862, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since BasicVSR violates causality and cannot be evaluated online, we construct the unidirectional BasicVSR (BasicVSR-uni) by removing the backward network for online inference. We compare the proposed SSL with three other pruning schemes at pruning ratio $p = 0.5$ : $\\mathrm{L}_1$ -norm pruning [23] (which simply removes filters with the smallest $\\mathrm{L}_1$ -norms and is the most prevailing filter pruning method now), and ASSL [57]. We apply these pruning schemes on BasicVSR, thus obtaining $\\mathrm{L}_1$ -norm-bi, ASSL-bi, and SSL-bi separately. In addition, we use these", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22642", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3ef2c569b81766602acd9eae5dc375f70695367d1690342761e966b5c32bed2f.jpg", + "table_caption": [ + "Table 1. Quantitative comparison (average PSNR/SSIM). Pruning schemes applied on bidirectional and unidirectional BasicVSR (\"bi\" and \"uni\") and marked in rouse and gray, respectively. * means the space-time knowledge distillation scheme [47]. We mark the best results among comparing pruning schemes in bold. The FLOPs and runtime are computed based on an LR size of $180 \\times 320$ ." + ], + "table_footnote": [], + "table_body": "
MethodsParams (M)FLOPs (G)Runtime (ms)BI degradationinBD degradationin
REDS4 [33]Vimeo-90K-T [49]Vid4 [27]UDM10 [52]Vimeo-90K-T [49]Vid4 [27]
Bicubic---26.14/0.729231.32/0.868423.78/0.634728.47/0.825331.30/0.868721.80/0.5246
VESPCN [1]-----25.35/0.7557---
SPMC [39]-----25.88/0.7752---
TOFlow [49]1.4274.9161027.98/0.799033.08/0.905425.89/0.765136.26/0.943834.62/0.9212-
DUF [21]5.81645.897428.63/0.8251--38.48/0.960536.87/0.944727.38/0.8329
RBPN [13]12.28516.0150730.09/0.859037.07/0.943527.12/0.818038.66/0.959637.20/0.9458-
EDVR-M [44]3.3304.211830.53/0.869937.09/0.944627.10/0.818639.40/0.966337.33/0.948427.45/0.8406
PFNL [51]3.0940.029529.63/0.850236.14/0.936326.73/0.802938.74/0.9627-27.16/0.8355
TGA [18]5.8694.1236----37.59/0.951627.63/0.8423
RLSP [9]4.282.349---38.48/0.960636.49/0.940327.48/0.8388
RSDN [16]6.2355.794---39.35/0.965337.23/0.947127.92/0.8505
RRN [19]3.4108.745---38.96/0.9644-27.69/0.8488
FastDVDnet* [47]2.664.3--36.12/0.934826.14/0.8029---
BasicVSR [2]4.9338.55731.42/0.890937.18/0.945027.24/0.825139.96/0.969437.53/0.949827.96/0.8553
BasicVSR-lite1.385.52430.56/0.873836.57/0.939726.86/0.812538.98/0.964536.78/0.943127.27/0.8327
L1-norm-bi [23]1.385.52430.66/0.876636.69/0.940626.87/0.812139.04/0.965036.84/0.943727.29/0.8335
ASSL-bi [57]1.385.52430.74/0.877036.75/0.941427.01/0.817639.15/0.966036.93/0.945027.40/0.8400
SSL-bi (Ours)1.092.12431.06/0.893336.82/0.941927.15/0.820839.35/0.966537.06/0.945827.56/0.8431
BasicVSR.uni [2]2.6218.13930.56/0.869836.95/0.942927.01/0.816439.25/0.964537.25/0.947227.57/0.8424
BasicVSR.uni-lite0.762.41829.95/0.856136.38/0.937226.68/0.801238.24/0.958636.38/0.938826.87/0.8157
L1-norm.uni [23]0.762.41829.97/0.857036.45/0.938126.70/0.803138.43/0.960136.53/0.940526.89/0.8187
ASSL.uni [57]0.762.41830.02/0.858936.49/0.938526.76/0.805138.48/0.960336.61/0.941627.02/0.8236
SSL.uni (Ours)0.563.91830.24/0.863336.56/0.939227.01/0.814838.68/0.961536.77/0.942927.18/0.8296
", + "bbox": [ + 80, + 135, + 883, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pruning schemes on BasicVSR-uni, obtaining $\\mathrm{L}_1$ -norm-uni, ASSL-uni, and SSL-uni. Moreover, we reduce the channels of BasicVSR and BasicVSR-uni to obtain lightweight VSR models BasicVSR-lite and BasicVSR-uni-lite, respectively. Furthermore, we compare our pruned BasciVSR and BasicVSR-uni with other lightweight VSR networks, including TOFlow [49], EDVR-M [44], RLSP [9], RSDN [16], etc. Since we only prune VSR networks, the parameters and FLOPs of the optical flow network, SPyNet [35] (Params 1.4M, Flops 19.6G) are not included.", + "bbox": [ + 75, + 446, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results measures performance (PSNR and SSIM), the number of parameters, runtime, and FLOPs on the different methods, which are shown in Tab. 1. (1) Compared with competitive lightweight VSR networks, our SSL-bi obtains 0.53dB gain on REDS4 over EDVR-M. Note that, different from careful network designs like EDVR-M, we merely prune the BasicVSR, a simple backbone with 60 residual blocks, obtaining superior performance while only consuming less the FLOPs of EDVR-M. (2) Our SSL-bi surpasses the BasicVSR-lite by 0.5dB, and SSL-uni surpasses the BasicVSR-uni-lite by 0.29dB. This demonstrates the effectiveness of applying SSL for offline and online VSR network pruning. (3) We also adapt some SOTA pruning schemes, such as the $\\mathrm{L}_1$ -norm and ASSL, to VSR networks for comparison. As a result, our SSL achieves superior performance on BasicVSR and BasicVSR-uni to other pruning schemes. This shows that SSL can make better use of the sparsity of the network and increases the efficiency of the learned network parameters. (4) Besides, comparing our pruning scheme and other VSR", + "bbox": [ + 75, + 598, + 473, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model compression method (space-time knowledge distillation scheme [47], FastDVDnet\\*), our SSL-uni (0.5M parameters) surpasses the FastDVDnet\\* (2.6M parameters) by 0.87dB on Vid4, which further demonstrates the effectiveness of our structured pruning scheme. Moreover, with our SSL strategy, we do not have to train a teacher network as knowledge distillation [47] methods did.", + "bbox": [ + 496, + 446, + 893, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative results are shown in Fig. 3. Our SSL-bi achieves the best visual quality containing more realistic details. These visual comparisons are consistent with the quantitative results, showing the superiority of SSL. SSL can learn to remove the redundant filters to compress a network to a smaller one while maintaining the most restoration ability. More visual results are given in supplementary.", + "bbox": [ + 496, + 553, + 893, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 667, + 653, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Validation of Components in SSL. We conduct an ablation study to demonstrate the effectiveness of our SSL by progressively adding components. The results are shown in Tab. 2. $\\mathrm{SSL}_1$ uses the aligned pruning [57] scheme for residual blocks, while $\\mathrm{SSL}_4$ adopts our RSC. Comparing $\\mathrm{SSL}_1$ and $\\mathrm{SSL}_4$ , we can see that our RSC is superior to the advanced residual block pruning scheme. That is because RSC can break the pruning restrictions and preserve all information contained in feature channels for VSR. For $\\mathrm{SSL}_2$ , we halve the number of filters in upsampling networks to keep the same model size as other models. As we can see, $\\mathrm{SSL}_3$ outperforms $\\mathrm{SSL}_2$ . This is because introducing the pruning scheme for the pixel-shuffle operation can increase the available pruning space in upsampling networks. Com", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22643", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e1e950f6eab04306eb44be8b7e1640b4c40d37e72d431ec1fd673765aae7bef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 87, + 250, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cc4f60e2109697bcca6ce6b674b4ffa11985003a79a2f16156e8594c4a1d407.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 87, + 357, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dba4a77c028d9d529a681fcded9c98ad951202094c3ab1ee5d5ccf49af43aaa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 87, + 462, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b1328866a8540c1574ef6d66c8b68088c743ebf0077c19a2e47791446191a5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 87, + 568, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50c2613c605d3375c0ae07ca7a5408b251558f7fc91f99ab0f00e269ce35ae1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 87, + 671, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eced6bf3057191da520f1a2bc9b4bed789f989d1c534a7b981f8371626a3bbf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 87, + 776, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6ee23f1e068ae0a76e014e27f2056e0480d06df3d3e9bb7451175ea75e6706c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 87, + 880, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/686cccca37926081a6b88ef9a29e47dfce73b3d3f0a48ac3e9dc1c7ce4b72282.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 217, + 250, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faa4bbedc0ac47773e1ecb3414a097bc90f520ce75ed9453377e47e4eef07e78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 217, + 357, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a7d28957bda741f235a68048a56956092538fe5ca4c6c30277cd22081da738d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 217, + 462, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0abf7d97cfbe794371efb3f3fc42e4a5c37d4ba2f034335ca79d3b758af303ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 217, + 568, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f038b3ff1ed95467138f139464eb5f81242761a22155e0fdb2259b8c48b96a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 217, + 671, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95290f6b17d20ef08b14e7fc727cb399123615d14f4e97af736a13c6c2fce0e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 217, + 776, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e64b9e6c2e4a4c045ff716aed8a0b88bc5360090db6f4725f48e06cc172788e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 217, + 880, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/432c1403ba3381c7d3acdc007f2cec353f2377e34eabd25919098d92b2ddbff4.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison between various VSR and pruning schemes on REDS4 [33], Vid4 [27], and Viceo90K-T [49], separately." + ], + "image_footnote": [], + "bbox": [ + 86, + 345, + 250, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47095df1b9ae219a3886a79eff9f9d64083e0890ada0880320b0ed476392f4ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 345, + 357, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ef4a4f21fd5d56d2720b1417e9ce901c5d4ad61b421267a4597958ae522f418.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 345, + 462, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62c2281011ef9812ff54bf6fc4f84c25495d68a4f48d72ef3d4877beb9d67cb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 345, + 568, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c933ec84728b6a910177a3045ac093baf200f5b04fa650031318f0b000a69d33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 345, + 671, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/384dbb92e86ece277980850b0bfc3fba1fd482c100f4697a6375e83609cfb2ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 345, + 776, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b969d2963a012b9467b8693a4c9b444428dc6644f2b4f16a54a56f69bf2298d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 345, + 880, + 470 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b5c2febe44a9f1d996165cfb529f68395e77f82d6e63bdebb98924c2cf0b9832.jpg", + "table_caption": [ + "Table 2. Validation of the components in our SSL. PSNR (dB) results evaluated on REDS4 [33] $(4\\times)$ . The backbone is BasicVSR [2], and the pruning ratio is set to 0.5." + ], + "table_footnote": [], + "table_body": "
MethodsSSL1SSL2SSL3SSL4 (Ours)
Aligned Pruning [57]✓
Residual Sparsity Connection✓✓✓
Pixel-Shuffle Pruning✓✓✓
Temporal Finetuning✓✓
PSNR (dB)30.8630.8230.9831.06
", + "bbox": [ + 96, + 546, + 446, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "paring $\\mathrm{SSL}_4$ and $\\mathrm{SSL}_3$ , we can see that adopting TF can bring a 0.08 dB improvement, reducing the error accumulation of hidden states in the recurrent network after pruning. Pruning Methods with Various Pruning Ratios. To further demonstrate SSL's effectiveness, we compare it with widely used pruning schemes, including $\\mathrm{L}_1$ -norm [23], ASSL [57] at different pruning ratios. We use these pruning schemes to obtain numerous submodels with different FLOPs. Besides, we also adjust the channels of BasicVSR to obtain BasicVSR-lite with different FLOPs. The results are shown in Fig. 4. (1) Our SSL achieves the best performance compared with other pruning schemes at different pruning ratios and FLOPs. Note that SSL even surpasses the BasicVSR-lite in the same model size by 0.73 dB on submodels with around 3.2G FLOPs. This demonstrates the superiority of SSL on VSR. (2) With the pruning ratio increasing and FLOPs decreasing, the performance advantage", + "bbox": [ + 75, + 643, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1779c4201c4432d2a345eaeb9451ac046396d655eb03b085be043015fa6296ea.jpg", + "image_caption": [ + "Figure 4. PSNR (dB) comparison on REDS4 $(\\times 4)$ between SSL and three other methods obtaining the same small network." + ], + "image_footnote": [], + "bbox": [ + 535, + 513, + 831, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "brought by our SSL becomes more evident compared with BasicVSR-lite, $\\mathrm{L}_1$ -norm and ASSL.", + "bbox": [ + 498, + 670, + 890, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Different Pruning Criteria. We explore the influence of different pruning criteria on the pruned VSR model at different pruning ratios. Specifically, we select and remove the unimportant filters globally (namely, comparing all filters from all layers together) with minimum $\\mathrm{L}_1$ -norm scores, which is expressed as \"Min + Global\". In addition, we select and remove the unimportant filters locally (namely, filters are compared with each other in the same layer, and each layer has the same pruning ratio) with maximum $\\mathrm{L}_1$ -norm scores, which is expressed as \"Max + Local\". Similarly, we determine \"Max + Global\" and \"Min + local\". Furthermore, we randomly remove the unimportant filters as \"Rand\". Then, we com", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22644", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95e82c17eb1c142ca96a6d8dca50f3231f37faa9824076d7be2bbef84ede7c66.jpg", + "image_caption": [ + "(a) Forward Network." + ], + "image_footnote": [], + "bbox": [ + 91, + 104, + 385, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c3b9deeb97a0bd1ef720d97f313941a265e2208fc196a0d0f7dceebe351149b.jpg", + "image_caption": [ + "(b) Backward Network." + ], + "image_footnote": [], + "bbox": [ + 405, + 104, + 699, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/455419633ce7bf3c45f4eff3b6f9586de8d8dc576e771a09e2ed9f2a368ed9c9.jpg", + "image_caption": [ + "(c) Upsampling Network." + ], + "image_footnote": [], + "bbox": [ + 714, + 109, + 844, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ef00972cee5a0213f50621e26b702ff10e3408f937a993f48bc9bdec41298c1.jpg", + "table_caption": [ + "Table 3. PSNR (dB) comparison on REDS4 $(4\\times)$ for our pruning scheme (SSL) with different pruning criteria and pruning ratios. The unpruned model is BasicVSR [2] baseline." + ], + "table_footnote": [], + "table_body": "
Pruning RatiosMin Global (Ours)Max GlobalMin LocalMax LocalRand
0.331.2830.5931.2230.9731.10
0.531.0628.9030.8330.3830.68
0.730.4125.8930.2029.5130.09
", + "bbox": [ + 98, + 327, + 447, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pare all pruning criteria at 0.3, 0.5, and 0.7 pruning ratios. The results are shown in Tab. 3. (1) The \"Min + Global\" pruning criterion achieves the best performance at different pruning ratios. It implies that the filters with minimum $\\mathrm{L}_1$ -norm scores are relatively unimportant for VSR. Besides, this shows that the importance of filters in different layers are different and it is better to compare them together to select unimportant filters (pruning globally) for VSR. (2) The performance of \"Max + Global\" and \"Max + local\" are both inferior to \"Rand\" for the removal of more important filters. This implies that filters with large $\\mathrm{L}_1$ -norm scores are more important than those with small ones for VSR networks.", + "bbox": [ + 75, + 415, + 468, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pruning Ratios of Different Layers. We take BasicVSR pruned by SSL at 0.5 pruning ratio as an example. We visualize pruning ratios in different layers and show results in Fig. 5. (1) In the forward and backward networks, the pruning ratios of the first Conv input channels (corresponding to the $\\gamma_{j-1}$ in Fig. 1 (d)) are lower than the pruning ratios of second Conv output filters (corresponding to the $\\gamma_{j+1}$ in Fig. 1 (d)). It implies that VSR networks tend to aggregate information from the numerous input channels into several important output channels. (2) The residual blocks in deeper position of forward and backward networks tend to have minor pruning ratio. This means that residual blocks in deeper position contribute more to VSR. (3) The average pruning ratio of upsampling network is 0.2 (less than 0.5), indicating that upsampling network plays a quite important role in VSR performance. Previous works [2,9,25,44,45,51,55] have paid much attention to the design of feature extraction modules. In latter VSR research, paying more attention to upsampling network design is likely to improve VSR performance more.", + "bbox": [ + 75, + 598, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71ab01863011609003bf89f8870521b6feff264b427b36a9a18591f86b85ea06.jpg", + "image_caption": [ + "Figure 5. (a), (b), and (c) show the pruning ratios of Conv layers in forward, backward and upsampling networks, respectively.", + "Figure 6. The SSL pruning process of Convs in the BasicVSR [2]." + ], + "image_footnote": [], + "bbox": [ + 527, + 287, + 859, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularization Visualization. To understand how sparsity-inducing regularization works, we plot the average scaling factors in the BasicVSR [2] during applying regularization at 0.5 pruning ratio (Fig. 6). The average scaling factor is split into two parts, pruned and kept. As seen, the average scaling factor $\\gamma$ of the pruned filters decreases as the corresponding penalty term $\\alpha_{\\gamma}$ and iterations become larger. Besides, it is interesting that the average scaling factor of the kept filters will increase without any regularization term to enforce them to be larger. It means that, as the unimportant filters are removed, the network will strengthen the kept filters to compensate for the performance, which is similar to the compensation effect in the human brain.", + "bbox": [ + 496, + 446, + 890, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 651, + 617, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a structured pruning scheme called SSL for efficient VSR in resource-limited situations. Specifically, for the difficulty of pruning residual blocks of recurrent networks, we propose the RSC. Compared with previous pruning schemes for residual blocks, RSC does not have pruning restrictions as other pruning schemes and can fully utilize restoration information in all channels for better performance. In addition, for the pixel-shuffle operation in the upsampling network, we specially design a pruning scheme by grouping filters to guarantee the accuracy of channel-space conversion after pruning. Furthermore, we propose Temporal Finetuning to reduce the error accumulation in recurrent networks. We apply SSL on the BasicVSR, and SSL achieves superior performance to that of recent SOTA methods, quantitatively and qualitatively.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22645", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017. 2, 6", + "[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8", + "[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. arXiv preprint arXiv:2104.13371, 2021. 2, 3, 5", + "[4] Jian Cheng, Pei-song Wang, Gang Li, Qing-hao Hu, and Han-qing Lu. Recent advances in efficient computation of deep convolutional neural networks. Frontiers of Information Technology & Electronic Engineering, 2018. 1, 2", + "[5] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. Model compression and acceleration for deep neural networks: The principles, progress, and challenges. IEEE Signal Processing Magazine, 2018. 1, 2", + "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 2", + "[7] Xiaohan Ding, Guiguang Ding, Yuchen Guo, Jungong Han, and Chenggang Yan. Approximated oracle filter pruning for destructive cnn width optimization. In ICML, 2019. 2", + "[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 2", + "[9] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 6, 8", + "[10] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. *ICLR*, 2018. 2", + "[11] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1, 2", + "[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1, 2", + "[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019. 6", + "[14] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 1, 2", + "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 3", + "[16] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2, 6", + "[17] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022. 2", + "[18] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020. 6", + "[19] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020. 6", + "[20] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 1", + "[21] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 6", + "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[23] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 1, 2, 3, 4, 5, 6, 7", + "[24] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019. 1", + "[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 3, 8", + "[26] Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning. NeurIPS, 2017. 3", + "[27] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. TPAMI, 2013. 5, 6, 7", + "[28] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2", + "[29] Hongying Liu, Peng Zhao, Zhubo Ruan, Fanhua Shang, and Yuanyuan Liu. Large motion video super-resolution with dual subnet and multi-stage communicated upsampling. arXiv preprint arXiv:2103.11744, 2021. 1", + "[30] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3, 4", + "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5", + "[32] Jian-Hao Luo and Jianxin Wu. Neural network pruning with residual-connections and limited-data. In CVPR, 2020. 3", + "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In CVPRW, 2019. 5, 6, 7", + "[34] Junghun Oh, Heewon Kim, Seungjun Nah, Cheeun Hong, Jonghyun Choi, and Kyoung Mu Lee. Attentive fine-grained structured sparsity for image restoration. In CVPR, 2022. 1, 2, 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22646", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 6", + "[36] Russell Reed. Pruning algorithms-a survey. IEEE transactions on Neural Networks, 1993. 1, 2", + "[37] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 4, 5", + "[38] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. Efficient processing of deep neural networks. Synthesis Lectures on Computer Architecture, 2020. 1, 2", + "[39] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 2, 6", + "[40] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 2", + "[41] Huan Wang, Xinyi Hu, Qiming Zhang, Yuehai Wang, Lu Yu, and Haoji Hu. Structured pruning for efficient convolutional neural networks via incremental regularization. IEEE Journal of Selected Topics in Signal Processing, 2019. 1, 3", + "[42] Huan Wang, Can Qin, Yulun Zhang, and Yun Fu. Neural pruning via growing regularization. *ICLR*, 2021. 2", + "[43] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4926, 2021. 3", + "[44] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 1, 2, 5, 6, 8", + "[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In ECCVW, 2018. 1, 4, 8", + "[46] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. NeurIPS, 2016. 1, 2, 3", + "[47] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, 2021. 2, 6", + "[48] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In ACM MM, 2020. 2", + "[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6, 7", + "[50] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, 2021. 2, 3", + "[51] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 1, 6, 8" + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2, 5, 6", + "[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, 2022. 2", + "[54] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In ICCV, 2019. 2", + "[55] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 8", + "[56] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 1, 4", + "[57] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Aligned structured sparsity learning for efficient image superresolution. NeurIPS, 2021. 1, 2, 3, 4, 5, 6, 7", + "[58] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 1, 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 402 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22647", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_model.json b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f6e078e4fc14386faf5dbe1c7b531509c71080c2 --- /dev/null +++ b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "Structured Sparsity Learning for Efficient Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.18, + 0.713, + 0.217 + ], + "angle": 0, + "content": "Bin Xia\\(^{1}\\), Jingwen He\\(^{2}\\), Yulun Zhang\\(^{3}\\), Yitong Wang\\(^{4}\\), Yapeng Tian\\(^{5}\\), Wenming Yang\\(^{1*}\\), and Luc Van Gool\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.217, + 0.74, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\) Tsinghua University, \\(^{2}\\) Shanghai AI Laboratory, \\(^{3}\\) ETH Zürich," + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.234, + 0.675, + 0.252 + ], + "angle": 0, + "content": "4 ByteDance Inc, 5 University of Texas at Dallas" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.302 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.633 + ], + "angle": 0, + "content": "The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.639, + 0.21, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Video super-resolution (VSR) aims to generate a high-resolution (HR) video from its corresponding low-resolution (LR) observation by filling in missing details. With the popularity of intelligent edge devices such as smartphones and small drones, performing VSR on these devices is in high demand. Although a variety of VSR networks [20, 24, 29, 44, 51] can achieve great performance, these models are usually difficult to be deployed on edge devices with limited computation and memory resources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.471, + 0.88 + ], + "angle": 0, + "content": "To alleviate this issue, we explore a new direction for effective and efficient VSR. To reduce the redundancy of Conv kernels [4,5,36,38] obtaining a more efficient VSR network, we develop a neural network pruning scheme for the VSR task for the first time. Since structured pruning [14,23,46,57] (focusing on filter pruning) can achieve" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.38 + ], + "angle": 0, + "content": "an actual acceleration [41,46] superior to unstructured pruning [11,12] (focusing on weight-element pruning), we adopt structured pruning principle to develop our VSR pruning scheme. Given a powerful VSR network, our pruning scheme can find submodels under presetting pruning rate without significantly compromising performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Structured pruning is a general concept, and designing a concrete pruning scheme for VSR networks is challenging. (1) Recurrent networks are widely used in VSR models to extract temporal features, consisting of residual blocks (e.g., BasicVSR [2] has 60 residual blocks). However, it is hard to prune the residual blocks because the skip and residual connections ought to share the same indices [23] (Fig. 1 (a)). As shown in Fig. 1 (b), quite a few structured pruning schemes [23,34] do not prune the last Conv layer of the residual blocks, which restricts the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] introduce regularization and prune the same indices on skip and residual connections to keep channel alignment (local pruning scheme,i.e., each layer pruning the same ratio of filters). However, ASSL and SRPN still cannot achieve the potential of pruning residual blocks on recurrent networks. The recurrent networks take the previous output as later input (Fig. 2 (a)). This requires the pruned indices of the first and last Conv's in recurrent networks to be the same. But ASSL and SRPN cannot guarantee filter indices are aligned. Besides, many SR methods [45, 56] have shown that the information contained in front Conv layers can help the restoration feature extraction of later Conv layers. Thus, we design a Residual Sparsity Connection (RSC) for VSR recurrent networks, which preserves all channels of the input and output feature maps and selects the important channels for operation (Fig. 1 (d)). Compared with other pruning schemes [57, 58], RSC does not require the pruned indices of the first and last Conv's of recurrent networks to be the same, can preserve the information contained in all layers, and liberates the pruning space of the last Conv of the residual blocks without adding extra calculations. Notably, RSC can prune residual blocks globally (i.e., the filters in various layers are compared together to remove unimportant ones)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22638" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.13, + 0.087, + 0.836, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.334, + 0.625, + 0.344 + ], + "angle": 0, + "content": "(d) Residual Sparsity Connection Scheme (RSC, Ours)." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Figure 1. Illustration of different schemes for pruning residual blocks of recurrent networks. (a) Structure of the residual block in the VSR network. (b) The residual block pruning schemes [7, 23, 42] do not prune the last Conv. (c) ASSL [57] and SRPN [58] prunes the same indices on skip and residual connections to keep channel alignment, which abandons some channels of input and output feature maps. (d) RSC preserves all channels of input and output feature maps, which does not need to align the pruned indices on the first and last Conv's in recurrent networks, can fully use restoration information, and can prune the first and last Conv's of residual blocks without restrictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.652 + ], + "angle": 0, + "content": "(2) We observe that the upsampling network accounts for \\(22\\%\\) of the total calculations in BasicVSR [2], which is necessary to be pruned to reduce redundancy. Since the pixel-shuffle [37] operation in VSR networks converts the channels to space, pruning the pixel-shuffle without any restrictions would cause the channel-space conversion to fail. Thus, we specially design a pixel-shuffle pruning scheme by taking four consecutive filters as the pruning unit for \\(2\\times\\) pixel-shuffle. (3) Furthermore, we observe that the error of pruning VSR networks would accumulate with propagation steps increasing along with recurrent networks, which limits the efficiency and performance of pruning. Thus, we further introduce Temporal Finetuning (TF) to constrain the pruning error accumulation in recurrent networks. Overall, our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.655, + 0.469, + 0.715 + ], + "angle": 0, + "content": "- Our work is necessary and timely. There is an urgent need to compress VSR models for deployment. To the best of our knowledge, we are one of the first to design a structured pruning scheme for VSR." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.719, + 0.469, + 0.779 + ], + "angle": 0, + "content": "- We propose an integral VSR pruning scheme called Structured Sparsity Learning (SSL) for various components of VSR models, such as residual blocks, recurrent networks, and pixel-shuffle operations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.783, + 0.468, + 0.813 + ], + "angle": 0, + "content": "- We employ SSL to train VSR models, which surpass recent pruning schemes and lightweight VSR models." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.655, + 0.469, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.219, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.295, + 0.863 + ], + "angle": 0, + "content": "2.1. Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "VSR models can exploit additional information from neighboring LR frames for restoration [3, 8, 16, 17, 28, 48," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[49,52-54]. Earlier VSR methods [1,39,49] estimate the optical flow between LR frames and perform spatial warping for alignment. Later methods resort to a more sophisticated approach of implicit alignment. Instead of image-level motion alignment, TDAN [40] and EDVR [44] work at the feature level. TDAN [40] first adopted deformable Conv [6] in VSR to align the features of different frames. EDVR [44] extended TDAN by introducing coarse-to-fine deformable alignment and a new spatial-temporal attention fusion module. RSDN [16] adopted a recurrent detail-structural block and a hidden state adaptation module to reduce the effect of appearance changes and error accumulation. Recently, BasicVSR [2] found that bidirectional propagation coupled with a simple optical flow-based feature alignment can further improve performance. Similarly, Yi et al. [50] used the bidirectional propagation framework to exploit LR frames and estimated hidden states from the past, present, and future. To compress the VSR model, Xiao et al. designed a space-time knowledge distillation scheme [47]. However, these VSR methods require high computational costs impeding their application on resource-limited devices. Different from previous methods, we focus on designing SSL to compress VSR models by pruning redundant filters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.788, + 0.671, + 0.804 + ], + "angle": 0, + "content": "2.2. Network Pruning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Network pruning [4, 5, 34, 36, 38] is widely used to remove a set of redundant parameters for network acceleration. Pruning methods can be divided into two branches, structured pruning [10, 14, 23, 46] and unstructured pruning [11, 12]. Structured pruning methods prune the network at the level of filters, channels, and even layers, which" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.441 + ], + "angle": 0, + "content": "can obtain regular sparsity after pruning. This is beneficial for acceleration. In contrast, unstructured methods focus on pruning weights, leading up to much irregular sparsity. This is beneficial for compression but tends not to yield an actual acceleration [41, 46]. Specifically, Li et al. [23] applied the \\( \\mathrm{L}_1 \\)-norm to measure the importance of different filters and then removed the less important ones. Afterward, Liu et al. [30] added a sparsity-inducing penalty term on scaling factors of the batch normalization layers to enforce the channels with lower scaling factors to be the less informative ones. Recently, ASSL [57] and SRPN [58] utilized aligned structured sparsity learning for structured pruning of residual blocks. In addition, Luo et al. [32] developed a residual block pruning scheme for image classification using the Convs on skip connections. However, the residual blocks of VSR networks do not have such Convs. Lin et al. [26] conducted runtime neural network pruning according to the input image. Besides, Wang et al. [43] designed an unstructured pruning scheme for single image SR tasks by using sparse Conv to skip redundant computations. Since we cannot directly apply a general structured pruning method for VSR, we explore the properties of VSR networks and develop a VSR pruning scheme in this paper." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.447, + 0.212, + 0.464 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.469, + 0.188, + 0.484 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.493, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Figure 2 (a) shows VSR networks based on the bidirectional recurrent structures, such as BasicVSR [2]. Given a LR frame \\( I_{t} \\), the forward network concatenates \\( I_{t} \\) and the previous hidden state \\( H_{F,t-1} \\) to extract features from \\( I_{t} \\) and aggregate the reference information from \\( H_{F,t-1} \\). Similarly, the backward network extracts features from \\( I_{t} \\) and aggregates the reference information from the future hidden state \\( H_{B,t+1} \\). Note that both the forward and backward networks consist of numerous residual blocks. Then, the features generated by forward and backward networks are fed into the upsampling network, which consists of multiple pixel-shuffle operations and Convs, to obtain the recovered frame \\( SR_{t} \\). However, SOTA VSR networks [2,3,50] need massive computation and memory resources, limiting their deployment on resource-limited devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To pursue more efficient VSR networks, we specially design a VSR structured pruning scheme called Structured Sparsity Learning (SSL), according to the properties of VSR networks. Specifically, SSL has three stages, including pretraining, pruning, and finetuning. In the pretraining stage, we train a powerful VSR network. Since current VSR networks do not use BatchNorm [15], we introduce a scaling factor in pretrained VSR models to tune the sparsity of each channel and filter. In the pruning stage, we select the unimportant filters according to the pruning criterion and apply sparsity-inducing regularization on corresponding scaling factors. In addition, we propose a Residual Spar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "sity Connection (RSC) scheme to liberate the restrictions on pruning residual blocks of recurrent networks and preserve all restoration information contained in channels of feature maps for better performance. Moreover, for the upsampling networks, we specially develop a pruning scheme for the pixel-shuffle operation to guarantee the accuracy of channel-space conversion after pruning. Besides, we observe that the error of the hidden state would be amplified with the propagation in recurrent networks after pruning. Thus, in the finetuning stage, we design Temporal Finetuning (TF) to alleviate the error accumulation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.271, + 0.766, + 0.287 + ], + "angle": 0, + "content": "3.2. Structured Sparsity Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Structured Sparsity Learning (SSL) is a structured pruning scheme specially designed for VSR. It can reduce the redundancy of neural filters and obtain more efficient VSR submodels. Next, we will explain our SSL in detail." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.893, + 0.58 + ], + "angle": 0, + "content": "(1) Scaling Factor. Structured pruning aims to remove Conv filters based on a designed importance criterion. In the classification task, quite a few works use scale parameters of BatchNorm [15] to control the throughput of each filter. Zero scale parameters make the value of corresponding channels vanish. As a result, they contribute nothing to the subsequent Convs and can be removed. By regularizing the scale parameter, we can assess and tune the importance of each filter. However, the BatchNorm is not useful for SR tasks [25], and SOTA VSR networks [2,3,50] do not utilize it. Therefore, it is infeasible to apply the existing pruning schemes directly. In our pruning scheme, as shown in Fig. 1 (d) and Fig. 2 (b), we multiply the scaling factors \\(\\gamma\\) before or after Convs. Then, we perform regularization on scaling factors to enforce sparsity for pruning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.894, + 0.763 + ], + "angle": 0, + "content": "(2) Pruning Criterion and Regularization Form. To remove the redundant filters, we need to select unimportant scaling factors \\(\\gamma\\) to induce sparsity. In previous works, ASSL [57] and SRPN [58] had to adopt a local pruning scheme (namely, scaling factors are only compared within the same layer, and each layer has the same pruning ratio) to guarantee that skip and residual connections keep the same number of filters and indices for the adding operation. Given that the importance of the Conv filters in various layers is different and our RSC does not have restrictions as ASSL, we can adopt the global pruning scheme (i.e., scaling factors of different layers are compared together)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For the pruning criterion, we adopt the simple and practical \\(\\mathbf{L}_1\\)-norm. Specifically, given Conv kernel \\(\\boldsymbol{W_i} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times K_h \\times K_w}\\) in the \\(i\\)-th layer, we calculate the absolute weight sum of \\(k\\)-th Conv filter \\(\\boldsymbol{W_i}[k,\\ldots] \\in \\mathbb{R}^{C_{in} \\times K_h \\times K_w}\\) with \\(s_{i,k} = \\sum |\\boldsymbol{W_i}[k,\\ldots]|\\). In particular, for our RSC in Fig. 1 (d), we require to additionally prune the input channels for the first Conv, and calculate its \\(\\mathbf{L}_1\\)-norm score with \\(s_{i,k}' = \\sum |\\boldsymbol{W_i}[:k,\\ldots]|\\), where \\(\\boldsymbol{W_i}[:k,\\ldots] \\in \\mathbb{R}^{C_{out} \\times K_h \\times K_w}\\). Moreover, for the Conv" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.465, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.089, + 0.888, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Figure 2. (a) The basic architecture of the VSR methods with the bidirectional recurrent network. The forward and backward networks both consist of numerous residual blocks. The upsampling networks contain multiple pixel-shuffle operations and Convs. (b) The pruning scheme for the pixel-shuffle. For the \\(2 \\times\\) upsampling pixel-shuffle [37] operation, we take four channels with consecutive indices as the pruning unit to guarantee the accuracy of channel-space conversion after pruning. (c) The application of SSL on VSR models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.466 + ], + "angle": 0, + "content": "before the pixel-shuffle operation, we take four consecutive filters as a pruning unit and calculate the score with \\( s_{i,k} = \\sum |\\mathbf{W}_i[4k:4(k + 1),\\ldots ]| \\). Then, given the pruning ratio \\( p \\) and the total number of filters or channels \\( N \\), we sort all their \\( \\mathbf{L}_1 \\)-norm scores \\( s \\) together and choose the \\( N\\times p \\) filters with the smallest \\( \\mathbf{L}_1 \\)-norm values as unimportant filters or channels, denoted as set \\( S \\) (Fig. 2 (c))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.572 + ], + "angle": 0, + "content": "After identifying the unimportant filters and channels set \\( S \\), we apply sparsity-inducing regularization (SIR) to the corresponding scaling factors, denoted as set \\( S_{sf} \\). It is notable that we do not enforce sparsity-inducing regularization to the important filters and channels since they will remain in the network. Specifically, we employ \\( L_{2} \\) regularization on the scaling factors to enforce sparsity (Fig. 2 (c)):" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.58, + 0.47, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S I R} = \\alpha_ {\\gamma} \\sum_ {\\gamma \\in S _ {s f}} \\gamma^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.689 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the scalar selected from \\(\\gamma \\in \\mathbb{R}^C\\) corresponding to unimportant filters or channels; \\(\\alpha_{\\gamma}\\) is a scalar. We increment \\(\\alpha_{\\gamma}\\) by a presetting constant \\(\\Delta\\) every \\(T_{1}\\) iterations. When \\(\\alpha_{\\gamma}\\) reaches the pre-defined upper limit \\(\\tau\\), we keep \\(\\alpha_{\\gamma}\\) constant and continue training \\(T_{2}\\) iterations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "(3) Pruning Scheme for Residual Blocks of Recurrent Networks. VSR recurrent networks consist of residual blocks. Residual blocks are difficult to prune because the addition operations require the pruned filter indices between the skip and residual connections to be the same. As shown in Fig. 1 (b), quite a few pruning schemes [23, 30, 34] simply skipped the pruning of the last Conv in residual blocks, which restricted the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] pruned the last Conv in the residual block by using regularization. However, recurrent networks take the previous output as later input. Thus, pruning recurrent networks require the pruned indices of the first and last Convs to be the same. ASSL and SRPN cannot guarantee the indices of important filters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.648 + ], + "angle": 0, + "content": "in the first and last Convs of recurrent networks are aligned. Besides, ASSL and SRPN have to remove some channels in output feature maps and adopt a local pruning scheme, which limits pruning space and restoration information utilization (RDN [56] and ESRGAN [45] have shown restoration information of front layers can guide feature extraction of later layers). To break the restriction of pruning and fully use restoration information contained in channels of front layers, we propose the RSC to prune residual blocks in recurrent networks (Fig. 1 (d)). As we can see, RSC preserves all channels of input \\( F_{i}^{\\prime} \\) and output \\( F_{i+1}^{\\prime} \\) in the residual blocks. For the first Conv, we select the important channels (the indices not in \\( S \\)) to participate in the Conv operation, which can be expressed as Eq. 2. After the last Conv, we obtained \\( F_{i+1} \\) and add \\( F_{i+1} \\) to \\( F_{i}^{\\prime} \\) on the corresponding channel indices to obtain \\( F_{i+1}^{\\prime} \\) (Eq. 3 and Eq. 4). Furthermore, we do not prune the \\( 1 \\times 1 \\) Conv of upsampling networks (Fig. 2 (a)) to aggregate all preserved restoration information in \\( H_{F,t}, H_{B,t} \\in \\mathbb{R}^{C \\times H \\times W} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.655, + 0.892, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {i} = \\boldsymbol {F} _ {i} ^ {\\prime} \\otimes (\\gamma_ {j - 1} \\boldsymbol {W} _ {i} \\gamma_ {j}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.675, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {i + 1} = \\boldsymbol {F} _ {i} \\otimes \\left(\\boldsymbol {W} _ {i + 1} \\gamma_ {j + 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.693, + 0.891, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {i + 1} ^ {\\prime} = \\boldsymbol {F} _ {i + 1} + \\boldsymbol {F} _ {i} ^ {\\prime}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\(\\otimes\\) indicates Conv. \\(F_{i}^{\\prime}, F_{i+1}^{\\prime} \\in \\mathbb{R}^{C \\times H \\times W}\\) are the input and output feature maps of the residual block, respectively. \\(F_{i}, F_{i+1} \\in \\mathbb{R}^{C_{p} \\times H \\times W}\\) are intermediate feature maps. \\(W_{i}, W_{i+1} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times K_{h} \\times K_{w}}\\) are weights of Conv kernels. \\(\\gamma_{j-1} \\in \\mathbb{R}^{C_{in}}\\) and \\(\\gamma_{j}, \\gamma_{j+1} \\in \\mathbb{R}^{C_{out}}\\) are scaling factors to apply sparsity-inducing regularization. \\(F_{i+1}^{\\prime}\\) prunes some channels, and \\(F_{i}\\) keeps all channels. In Eq. 4, \\(F_{i}^{\\prime}\\) adds \\(F_{i+1}\\) on corresponding kept channels. It is notable that our RSC does not introduce extra parameters and computational cost." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "(4) Pruning Scheme for Pixel-Shuffle. The upsampling network of the VSR network uses Conv to increase chan" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22641" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "nels of feature maps and adopts the pixel-shuffle [37] operation to convert the channels to space realizing upsampling. As shown in Fig. 2 (b), given the input feature map \\( F_{i} \\in \\mathbb{R}^{C \\times H \\times W} \\), we expand its channels \\( 4 \\times \\) by a Conv with weight \\( W_{i} \\) to obtain \\( F_{i+1}' \\in \\mathbb{R}^{4C \\times H \\times W} \\). Then, the pixel-shuffle operation takes four channels as a group to convert \\( F_{i+1}' \\) to \\( F_{i+1} \\in \\mathbb{R}^{C \\times 2H \\times 2W} \\) realize \\( 2 \\times \\) upsampling. We observe that, if we prune the Conv before pixel-shuffle without any restriction, the pruned feature maps will be spatially disordered after passing the pixel-shuffle operation and lead to performance drop. To address the problem, we specially design a strong and simple pruning scheme for the pixel-shuffle operation. Given the input feature map, we take four filters as a pruning unit to evaluate their importance and then impose the scaling factor \\( \\gamma_{j} \\) on filters to enforce sparsity (described in (2) pruning criterion and regularization form):" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.35, + 0.47, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} _ {i} ^ {\\prime} = \\boldsymbol {W} _ {i} [ 4 k: 4 (k + 1), \\dots ] \\gamma_ {j} [ k ], k \\in [ 0, C _ {i n}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.401 + ], + "angle": 0, + "content": "where \\(W_{i}\\in \\mathbb{R}^{4C_{in}\\times C_{in}\\times K_{h}\\times K_{w}}\\) is the weights of Conv kernel. \\(\\gamma_{j}\\) is the scaling factor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.506 + ], + "angle": 0, + "content": "(5) Temporal Finetuning. We observe that the pruned VSR network generates a minor error in hidden state \\( H_{F} \\) and \\( H_{B} \\) (Fig. 2 (a)), which will be amplified as the hidden state propagates along with recurrent networks. To solve the issue, we introduce Temporal Finetuning (TF) by enforcing the hidden states of pruned networks to align with the accurate hidden states of unpruned networks:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.511, + 0.469, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t f} = \\left\\| H _ {F, T} - H _ {F, T} ^ {\\prime} \\right\\| + \\left\\| H _ {B, 0} - H _ {B, 0} ^ {\\prime} \\right\\|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.605 + ], + "angle": 0, + "content": "where \\(T\\) is the number of input frames, and \\(H_{F,T}\\) and \\(H_{F,T}^{\\prime}\\) are the final hidden states after \\(T\\) frames forward propagation in pruned and original VSR networks, separately. Similarly, \\(H_{B,0}\\) and \\(H_{B,0}^{\\prime}\\) are the final hidden states after backward propagation in the pruned and original VSR networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.635 + ], + "angle": 0, + "content": "To train the whole VSR network, we use the Charbonnier loss [2, 44], which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.64, + 0.469, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sqrt {\\left\\| S R _ {t} - H R _ {t} \\right\\| ^ {2} + \\varepsilon^ {2}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.728 + ], + "angle": 0, + "content": "where \\(\\varepsilon\\) is set to \\(10^{-6}\\). \\(SR_{t}\\) and \\(HR_{t}\\) are \\(t\\)-th reconstructed and HR frames, respectively. The overall loss function for pruned network finetuning is designed as:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.737, + 0.469, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {t f}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.761, + 0.327, + 0.775 + ], + "angle": 0, + "content": "3.3. Arm VSR Models with SSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our SSL can be used for VSR networks. BasicVSR [2] and BasicVSR++ [3] are two strong SOTA VSR methods. The dense deformable Conv of BasicVSR++ requires reading a large amount of irregular memory data, which is unsuitable for deployment on resource-limited devices without GPUs. Besides, the Second-Order Grid Propagation of BasicVSR++ will force a delay of two frames, which further impedes its usage in real-time devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "In our study, we use BasicVSR for VSR pruning, which is more suitable for applications on edge devices. In addition, we further propose unidirectional BasicVSR (BasicVSR-uni), obtained by removing the backward network, for online inference. Since the SpyNet of BasicVSR is used for flow estimation, we do not apply our pruning scheme SSL on it. In the pruning stage, as shown in Fig. 2 (c), we first add the scaling factor to the Conv and residual blocks as described in Sec. 3.2. Then we use the pruning criterion to select unimportant filters globally and apply sparsity-inducing regularization to the corresponding scaling factor. Afterward, we remove the unimportant Conv filters and finetune the pruned VSR network with \\( T_{3} \\) iterations. We provide more details in Alg. 1 of supplementary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.309, + 0.633, + 0.325 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.329, + 0.707, + 0.345 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.545 + ], + "angle": 0, + "content": "We adopt two widely used datasets for training: REDS [33] and Vimeo-90K [49]. For REDS, we use REDS4 containing 4 clips as our test set. Additionally, we adopt REDSval4 as our validation set, which contains 4 clips selected from the REDS validation set. The remaining clips of REDS are used for training. In addition, we utilize Vid4 [27], UDM10 [52], and Vimeo-90K-T [49] as test sets along with Vimeo-90K. We train and test models with \\(4 \\times\\) downsampling using two degradations Bicubic (BI) and Blur Downsampling (BD) as BasicVSR did. For BI, the MATLAB function \"imresize\" is used for downsampling. For BD, we blur the HR images by a Gaussian filter with \\(\\sigma = 1.6\\), followed by a subsampling every four pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.727 + ], + "angle": 0, + "content": "We pretrain the unidirectional BasicVSR (BasicVSR-uni) as done for BasicVSR. In sparsity-inducing regularization, the iterations \\( T_{1} \\) and \\( T_{2} \\) are set to 5 and 3, 375 separately. The scalars \\( \\Delta \\) and \\( \\tau \\) are set to \\( 10^{-4} \\) and 0.1, respectively. Note that we fix the parameters of the flow estimator in sparsity-inducing regularization. In the pruned VSR network finetuning, we set \\( T_{3} \\) to 300,000. We adopt the Adam optimizer [22] and Cosine Annealing scheme [31]. The initial learning rate of the flow estimator is \\( 2.5 \\times 10^{-5} \\). The learning rate for all other modules is \\( 2 \\times 10^{-4} \\). The patch size of input LR frames is \\( 64 \\times 64 \\). Experiments are conducted on a server with PyTorch 1.10 and V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.731, + 0.864, + 0.747 + ], + "angle": 0, + "content": "4.2. Quantitative and Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since BasicVSR violates causality and cannot be evaluated online, we construct the unidirectional BasicVSR (BasicVSR-uni) by removing the backward network for online inference. We compare the proposed SSL with three other pruning schemes at pruning ratio \\( p = 0.5 \\): \\( \\mathrm{L}_1 \\)-norm pruning [23] (which simply removes filters with the smallest \\( \\mathrm{L}_1 \\)-norms and is the most prevailing filter pruning method now), and ASSL [57]. We apply these pruning schemes on BasicVSR, thus obtaining \\( \\mathrm{L}_1 \\)-norm-bi, ASSL-bi, and SSL-bi separately. In addition, we use these" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22642" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison (average PSNR/SSIM). Pruning schemes applied on bidirectional and unidirectional BasicVSR (\"bi\" and \"uni\") and marked in rouse and gray, respectively. * means the space-time knowledge distillation scheme [47]. We mark the best results among comparing pruning schemes in bold. The FLOPs and runtime are computed based on an LR size of \\(180 \\times 320\\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.136, + 0.885, + 0.442 + ], + "angle": 0, + "content": "
MethodsParams (M)FLOPs (G)Runtime (ms)BI degradationinBD degradationin
REDS4 [33]Vimeo-90K-T [49]Vid4 [27]UDM10 [52]Vimeo-90K-T [49]Vid4 [27]
Bicubic---26.14/0.729231.32/0.868423.78/0.634728.47/0.825331.30/0.868721.80/0.5246
VESPCN [1]-----25.35/0.7557---
SPMC [39]-----25.88/0.7752---
TOFlow [49]1.4274.9161027.98/0.799033.08/0.905425.89/0.765136.26/0.943834.62/0.9212-
DUF [21]5.81645.897428.63/0.8251--38.48/0.960536.87/0.944727.38/0.8329
RBPN [13]12.28516.0150730.09/0.859037.07/0.943527.12/0.818038.66/0.959637.20/0.9458-
EDVR-M [44]3.3304.211830.53/0.869937.09/0.944627.10/0.818639.40/0.966337.33/0.948427.45/0.8406
PFNL [51]3.0940.029529.63/0.850236.14/0.936326.73/0.802938.74/0.9627-27.16/0.8355
TGA [18]5.8694.1236----37.59/0.951627.63/0.8423
RLSP [9]4.282.349---38.48/0.960636.49/0.940327.48/0.8388
RSDN [16]6.2355.794---39.35/0.965337.23/0.947127.92/0.8505
RRN [19]3.4108.745---38.96/0.9644-27.69/0.8488
FastDVDnet* [47]2.664.3--36.12/0.934826.14/0.8029---
BasicVSR [2]4.9338.55731.42/0.890937.18/0.945027.24/0.825139.96/0.969437.53/0.949827.96/0.8553
BasicVSR-lite1.385.52430.56/0.873836.57/0.939726.86/0.812538.98/0.964536.78/0.943127.27/0.8327
L1-norm-bi [23]1.385.52430.66/0.876636.69/0.940626.87/0.812139.04/0.965036.84/0.943727.29/0.8335
ASSL-bi [57]1.385.52430.74/0.877036.75/0.941427.01/0.817639.15/0.966036.93/0.945027.40/0.8400
SSL-bi (Ours)1.092.12431.06/0.893336.82/0.941927.15/0.820839.35/0.966537.06/0.945827.56/0.8431
BasicVSR.uni [2]2.6218.13930.56/0.869836.95/0.942927.01/0.816439.25/0.964537.25/0.947227.57/0.8424
BasicVSR.uni-lite0.762.41829.95/0.856136.38/0.937226.68/0.801238.24/0.958636.38/0.938826.87/0.8157
L1-norm.uni [23]0.762.41829.97/0.857036.45/0.938126.70/0.803138.43/0.960136.53/0.940526.89/0.8187
ASSL.uni [57]0.762.41830.02/0.858936.49/0.938526.76/0.805138.48/0.960336.61/0.941627.02/0.8236
SSL.uni (Ours)0.563.91830.24/0.863336.56/0.939227.01/0.814838.68/0.961536.77/0.942927.18/0.8296
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.599 + ], + "angle": 0, + "content": "pruning schemes on BasicVSR-uni, obtaining \\( \\mathrm{L}_1 \\)-norm-uni, ASSL-uni, and SSL-uni. Moreover, we reduce the channels of BasicVSR and BasicVSR-uni to obtain lightweight VSR models BasicVSR-lite and BasicVSR-uni-lite, respectively. Furthermore, we compare our pruned BasciVSR and BasicVSR-uni with other lightweight VSR networks, including TOFlow [49], EDVR-M [44], RLSP [9], RSDN [16], etc. Since we only prune VSR networks, the parameters and FLOPs of the optical flow network, SPyNet [35] (Params 1.4M, Flops 19.6G) are not included." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.474, + 0.904 + ], + "angle": 0, + "content": "The quantitative results measures performance (PSNR and SSIM), the number of parameters, runtime, and FLOPs on the different methods, which are shown in Tab. 1. (1) Compared with competitive lightweight VSR networks, our SSL-bi obtains 0.53dB gain on REDS4 over EDVR-M. Note that, different from careful network designs like EDVR-M, we merely prune the BasicVSR, a simple backbone with 60 residual blocks, obtaining superior performance while only consuming less the FLOPs of EDVR-M. (2) Our SSL-bi surpasses the BasicVSR-lite by 0.5dB, and SSL-uni surpasses the BasicVSR-uni-lite by 0.29dB. This demonstrates the effectiveness of applying SSL for offline and online VSR network pruning. (3) We also adapt some SOTA pruning schemes, such as the \\(\\mathrm{L}_1\\)-norm and ASSL, to VSR networks for comparison. As a result, our SSL achieves superior performance on BasicVSR and BasicVSR-uni to other pruning schemes. This shows that SSL can make better use of the sparsity of the network and increases the efficiency of the learned network parameters. (4) Besides, comparing our pruning scheme and other VSR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.553 + ], + "angle": 0, + "content": "model compression method (space-time knowledge distillation scheme [47], FastDVDnet\\*), our SSL-uni (0.5M parameters) surpasses the FastDVDnet\\* (2.6M parameters) by 0.87dB on Vid4, which further demonstrates the effectiveness of our structured pruning scheme. Moreover, with our SSL strategy, we do not have to train a teacher network as knowledge distillation [47] methods did." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.661 + ], + "angle": 0, + "content": "The qualitative results are shown in Fig. 3. Our SSL-bi achieves the best visual quality containing more realistic details. These visual comparisons are consistent with the quantitative results, showing the superiority of SSL. SSL can learn to remove the redundant filters to compress a network to a smaller one while maintaining the most restoration ability. More visual results are given in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.655, + 0.685 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The Validation of Components in SSL. We conduct an ablation study to demonstrate the effectiveness of our SSL by progressively adding components. The results are shown in Tab. 2. \\(\\mathrm{SSL}_1\\) uses the aligned pruning [57] scheme for residual blocks, while \\(\\mathrm{SSL}_4\\) adopts our RSC. Comparing \\(\\mathrm{SSL}_1\\) and \\(\\mathrm{SSL}_4\\), we can see that our RSC is superior to the advanced residual block pruning scheme. That is because RSC can break the pruning restrictions and preserve all information contained in feature channels for VSR. For \\(\\mathrm{SSL}_2\\), we halve the number of filters in upsampling networks to keep the same model size as other models. As we can see, \\(\\mathrm{SSL}_3\\) outperforms \\(\\mathrm{SSL}_2\\). This is because introducing the pruning scheme for the pixel-shuffle operation can increase the available pruning space in upsampling networks. Com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22643" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.251, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.088, + 0.358, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.088, + 0.464, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.088, + 0.569, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.088, + 0.672, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.088, + 0.777, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.088, + 0.882, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.218, + 0.251, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.218, + 0.358, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.218, + 0.464, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.218, + 0.569, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.218, + 0.672, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.218, + 0.777, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.218, + 0.882, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.346, + 0.251, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.346, + 0.358, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.346, + 0.464, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.346, + 0.569, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.346, + 0.673, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.346, + 0.777, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.346, + 0.882, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.479, + 0.891, + 0.493 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison between various VSR and pruning schemes on REDS4 [33], Vid4 [27], and Viceo90K-T [49], separately." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Table 2. Validation of the components in our SSL. PSNR (dB) results evaluated on REDS4 [33] \\((4\\times)\\). The backbone is BasicVSR [2], and the pruning ratio is set to 0.5." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.547, + 0.447, + 0.64 + ], + "angle": 0, + "content": "
MethodsSSL1SSL2SSL3SSL4 (Ours)
Aligned Pruning [57]✓
Residual Sparsity Connection✓✓✓
Pixel-Shuffle Pruning✓✓✓
Temporal Finetuning✓✓
PSNR (dB)30.8630.8230.9831.06
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.902 + ], + "angle": 0, + "content": "paring \\(\\mathrm{SSL}_4\\) and \\(\\mathrm{SSL}_3\\), we can see that adopting TF can bring a 0.08 dB improvement, reducing the error accumulation of hidden states in the recurrent network after pruning. Pruning Methods with Various Pruning Ratios. To further demonstrate SSL's effectiveness, we compare it with widely used pruning schemes, including \\(\\mathrm{L}_1\\)-norm [23], ASSL [57] at different pruning ratios. We use these pruning schemes to obtain numerous submodels with different FLOPs. Besides, we also adjust the channels of BasicVSR to obtain BasicVSR-lite with different FLOPs. The results are shown in Fig. 4. (1) Our SSL achieves the best performance compared with other pruning schemes at different pruning ratios and FLOPs. Note that SSL even surpasses the BasicVSR-lite in the same model size by 0.73 dB on submodels with around 3.2G FLOPs. This demonstrates the superiority of SSL on VSR. (2) With the pruning ratio increasing and FLOPs decreasing, the performance advantage" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.515, + 0.833, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.661 + ], + "angle": 0, + "content": "Figure 4. PSNR (dB) comparison on REDS4 \\((\\times 4)\\) between SSL and three other methods obtaining the same small network." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.671, + 0.891, + 0.7 + ], + "angle": 0, + "content": "brought by our SSL becomes more evident compared with BasicVSR-lite, \\(\\mathrm{L}_1\\)-norm and ASSL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison with Different Pruning Criteria. We explore the influence of different pruning criteria on the pruned VSR model at different pruning ratios. Specifically, we select and remove the unimportant filters globally (namely, comparing all filters from all layers together) with minimum \\( \\mathrm{L}_1 \\)-norm scores, which is expressed as \"Min + Global\". In addition, we select and remove the unimportant filters locally (namely, filters are compared with each other in the same layer, and each layer has the same pruning ratio) with maximum \\( \\mathrm{L}_1 \\)-norm scores, which is expressed as \"Max + Local\". Similarly, we determine \"Max + Global\" and \"Min + local\". Furthermore, we randomly remove the unimportant filters as \"Rand\". Then, we com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22644" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.106, + 0.387, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.251, + 0.299, + 0.261 + ], + "angle": 0, + "content": "(a) Forward Network." + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.106, + 0.7, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.251, + 0.618, + 0.262 + ], + "angle": 0, + "content": "(b) Backward Network." + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.11, + 0.846, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.251, + 0.846, + 0.263 + ], + "angle": 0, + "content": "(c) Upsampling Network." + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.264, + 0.858, + 0.278 + ], + "angle": 0, + "content": "Figure 5. (a), (b), and (c) show the pruning ratios of Conv layers in forward, backward and upsampling networks, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.285, + 0.47, + 0.327 + ], + "angle": 0, + "content": "Table 3. PSNR (dB) comparison on REDS4 \\((4\\times)\\) for our pruning scheme (SSL) with different pruning criteria and pruning ratios. The unpruned model is BasicVSR [2] baseline." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.328, + 0.449, + 0.414 + ], + "angle": 0, + "content": "
Pruning RatiosMin Global (Ours)Max GlobalMin LocalMax LocalRand
0.331.2830.5931.2230.9731.10
0.531.0628.9030.8330.3830.68
0.730.4125.8930.2029.5130.09
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.597 + ], + "angle": 0, + "content": "pare all pruning criteria at 0.3, 0.5, and 0.7 pruning ratios. The results are shown in Tab. 3. (1) The \"Min + Global\" pruning criterion achieves the best performance at different pruning ratios. It implies that the filters with minimum \\( \\mathrm{L}_1 \\)-norm scores are relatively unimportant for VSR. Besides, this shows that the importance of filters in different layers are different and it is better to compare them together to select unimportant filters (pruning globally) for VSR. (2) The performance of \"Max + Global\" and \"Max + local\" are both inferior to \"Rand\" for the removal of more important filters. This implies that filters with large \\( \\mathrm{L}_1 \\)-norm scores are more important than those with small ones for VSR networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Pruning Ratios of Different Layers. We take BasicVSR pruned by SSL at 0.5 pruning ratio as an example. We visualize pruning ratios in different layers and show results in Fig. 5. (1) In the forward and backward networks, the pruning ratios of the first Conv input channels (corresponding to the \\(\\gamma_{j-1}\\) in Fig. 1 (d)) are lower than the pruning ratios of second Conv output filters (corresponding to the \\(\\gamma_{j+1}\\) in Fig. 1 (d)). It implies that VSR networks tend to aggregate information from the numerous input channels into several important output channels. (2) The residual blocks in deeper position of forward and backward networks tend to have minor pruning ratio. This means that residual blocks in deeper position contribute more to VSR. (3) The average pruning ratio of upsampling network is 0.2 (less than 0.5), indicating that upsampling network plays a quite important role in VSR performance. Previous works [2,9,25,44,45,51,55] have paid much attention to the design of feature extraction modules. In latter VSR research, paying more attention to upsampling network design is likely to improve VSR performance more." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.289, + 0.86, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.428, + 0.891, + 0.442 + ], + "angle": 0, + "content": "Figure 6. The SSL pruning process of Convs in the BasicVSR [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Regularization Visualization. To understand how sparsity-inducing regularization works, we plot the average scaling factors in the BasicVSR [2] during applying regularization at 0.5 pruning ratio (Fig. 6). The average scaling factor is split into two parts, pruned and kept. As seen, the average scaling factor \\(\\gamma\\) of the pruned filters decreases as the corresponding penalty term \\(\\alpha_{\\gamma}\\) and iterations become larger. Besides, it is interesting that the average scaling factor of the kept filters will increase without any regularization term to enforce them to be larger. It means that, as the unimportant filters are removed, the network will strengthen the kept filters to compensate for the performance, which is similar to the compensation effect in the human brain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.619, + 0.668 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a structured pruning scheme called SSL for efficient VSR in resource-limited situations. Specifically, for the difficulty of pruning residual blocks of recurrent networks, we propose the RSC. Compared with previous pruning schemes for residual blocks, RSC does not have pruning restrictions as other pruning schemes and can fully utilize restoration information in all channels for better performance. In addition, for the pixel-shuffle operation in the upsampling network, we specially design a pruning scheme by grouping filters to guarantee the accuracy of channel-space conversion after pruning. Furthermore, we propose Temporal Finetuning to reduce the error accumulation in recurrent networks. We apply SSL on the BasicVSR, and SSL achieves superior performance to that of recent SOTA methods, quantitatively and qualitatively." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22645" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. arXiv preprint arXiv:2104.13371, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[4] Jian Cheng, Pei-song Wang, Gang Li, Qing-hao Hu, and Han-qing Lu. Recent advances in efficient computation of deep convolutional neural networks. Frontiers of Information Technology & Electronic Engineering, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. Model compression and acceleration for deep neural networks: The principles, progress, and challenges. IEEE Signal Processing Magazine, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.437, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[7] Xiaohan Ding, Guiguang Ding, Yuchen Guo, Jungong Han, and Chenggang Yan. Approximated oracle filter pruning for destructive cnn width optimization. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.52 + ], + "angle": 0, + "content": "[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[9] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.565, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[10] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. *ICLR*, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[11] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.662, + 0.471, + 0.703 + ], + "angle": 0, + "content": "[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.704, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.746, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[14] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[16] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[17] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[18] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[19] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[20] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.273, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[21] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[23] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[24] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[26] Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[27] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. TPAMI, 2013. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[28] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[29] Hongying Liu, Peng Zhao, Zhubo Ruan, Fanhua Shang, and Yuanyuan Liu. Large motion video super-resolution with dual subnet and multi-stage communicated upsampling. arXiv preprint arXiv:2103.11744, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[30] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[32] Jian-Hao Luo and Jianxin Wu. Neural network pruning with residual-connections and limited-data. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In CVPRW, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Junghun Oh, Heewon Kim, Seungjun Nah, Cheeun Hong, Jonghyun Choi, and Kyoung Mu Lee. Attentive fine-grained structured sparsity for image restoration. In CVPR, 2022. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22646" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[35] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[36] Russell Reed. Pruning algorithms-a survey. IEEE transactions on Neural Networks, 1993. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.152, + 0.468, + 0.22 + ], + "angle": 0, + "content": "[37] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.223, + 0.468, + 0.264 + ], + "angle": 0, + "content": "[38] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. Efficient processing of deep neural networks. Synthesis Lectures on Computer Architecture, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.267, + 0.468, + 0.307 + ], + "angle": 0, + "content": "[39] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.311, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[40] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.354, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[41] Huan Wang, Xinyi Hu, Qiming Zhang, Yuehai Wang, Lu Yu, and Haoji Hu. Structured pruning for efficient convolutional neural networks via incremental regularization. IEEE Journal of Selected Topics in Signal Processing, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.412, + 0.468, + 0.439 + ], + "angle": 0, + "content": "[42] Huan Wang, Can Qin, Yulun Zhang, and Yun Fu. Neural pruning via growing regularization. *ICLR*, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.468, + 0.51 + ], + "angle": 0, + "content": "[43] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4926, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.513, + 0.468, + 0.566 + ], + "angle": 0, + "content": "[44] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.571, + 0.468, + 0.624 + ], + "angle": 0, + "content": "[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In ECCVW, 2018. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.628, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[46] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. NeurIPS, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.672, + 0.468, + 0.711 + ], + "angle": 0, + "content": "[47] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[48] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In ACM MM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.759, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.468, + 0.842 + ], + "angle": 0, + "content": "[50] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[51] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 1, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[52] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[54] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[55] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[56] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[57] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Aligned structured sparsity learning for efficient image superresolution. NeurIPS, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[58] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22647" + } + ] +] \ No newline at end of file diff --git a/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_origin.pdf b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc580c8547c58fc7f24b3d7f38ea78d5cf31dab8 --- /dev/null +++ b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/ed077dac-8227-4e0e-9bc4-a7377d89e3b1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff9fbd4b818cf9c1a492e75270d360d2748389b19895faefb604cf39df68c245 +size 2849766 diff --git a/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/full.md b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cac0bfe053d336dae8135e273781bdf3d7f7d411 --- /dev/null +++ b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/full.md @@ -0,0 +1,310 @@ +# Structured Sparsity Learning for Efficient Video Super-Resolution + +Bin Xia $^{1}$ , Jingwen He $^{2}$ , Yulun Zhang $^{3}$ , Yitong Wang $^{4}$ , Yapeng Tian $^{5}$ , Wenming Yang $^{1*}$ , and Luc Van Gool $^{3}$ + +$^{1}$ Tsinghua University, $^{2}$ Shanghai AI Laboratory, $^{3}$ ETH Zürich, + +4 ByteDance Inc, 5 University of Texas at Dallas + +# Abstract + +The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL. + +# 1. Introduction + +Video super-resolution (VSR) aims to generate a high-resolution (HR) video from its corresponding low-resolution (LR) observation by filling in missing details. With the popularity of intelligent edge devices such as smartphones and small drones, performing VSR on these devices is in high demand. Although a variety of VSR networks [20, 24, 29, 44, 51] can achieve great performance, these models are usually difficult to be deployed on edge devices with limited computation and memory resources. + +To alleviate this issue, we explore a new direction for effective and efficient VSR. To reduce the redundancy of Conv kernels [4,5,36,38] obtaining a more efficient VSR network, we develop a neural network pruning scheme for the VSR task for the first time. Since structured pruning [14,23,46,57] (focusing on filter pruning) can achieve + +an actual acceleration [41,46] superior to unstructured pruning [11,12] (focusing on weight-element pruning), we adopt structured pruning principle to develop our VSR pruning scheme. Given a powerful VSR network, our pruning scheme can find submodels under presetting pruning rate without significantly compromising performance. + +Structured pruning is a general concept, and designing a concrete pruning scheme for VSR networks is challenging. (1) Recurrent networks are widely used in VSR models to extract temporal features, consisting of residual blocks (e.g., BasicVSR [2] has 60 residual blocks). However, it is hard to prune the residual blocks because the skip and residual connections ought to share the same indices [23] (Fig. 1 (a)). As shown in Fig. 1 (b), quite a few structured pruning schemes [23,34] do not prune the last Conv layer of the residual blocks, which restricts the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] introduce regularization and prune the same indices on skip and residual connections to keep channel alignment (local pruning scheme,i.e., each layer pruning the same ratio of filters). However, ASSL and SRPN still cannot achieve the potential of pruning residual blocks on recurrent networks. The recurrent networks take the previous output as later input (Fig. 2 (a)). This requires the pruned indices of the first and last Conv's in recurrent networks to be the same. But ASSL and SRPN cannot guarantee filter indices are aligned. Besides, many SR methods [45, 56] have shown that the information contained in front Conv layers can help the restoration feature extraction of later Conv layers. Thus, we design a Residual Sparsity Connection (RSC) for VSR recurrent networks, which preserves all channels of the input and output feature maps and selects the important channels for operation (Fig. 1 (d)). Compared with other pruning schemes [57, 58], RSC does not require the pruned indices of the first and last Conv's of recurrent networks to be the same, can preserve the information contained in all layers, and liberates the pruning space of the last Conv of the residual blocks without adding extra calculations. Notably, RSC can prune residual blocks globally (i.e., the filters in various layers are compared together to remove unimportant ones). + +![](images/9c09dc091ece02e6550193c068b31ab201953219c277666bcd98d51b974ff2a0.jpg) +(d) Residual Sparsity Connection Scheme (RSC, Ours). +Figure 1. Illustration of different schemes for pruning residual blocks of recurrent networks. (a) Structure of the residual block in the VSR network. (b) The residual block pruning schemes [7, 23, 42] do not prune the last Conv. (c) ASSL [57] and SRPN [58] prunes the same indices on skip and residual connections to keep channel alignment, which abandons some channels of input and output feature maps. (d) RSC preserves all channels of input and output feature maps, which does not need to align the pruned indices on the first and last Conv's in recurrent networks, can fully use restoration information, and can prune the first and last Conv's of residual blocks without restrictions. + +(2) We observe that the upsampling network accounts for $22\%$ of the total calculations in BasicVSR [2], which is necessary to be pruned to reduce redundancy. Since the pixel-shuffle [37] operation in VSR networks converts the channels to space, pruning the pixel-shuffle without any restrictions would cause the channel-space conversion to fail. Thus, we specially design a pixel-shuffle pruning scheme by taking four consecutive filters as the pruning unit for $2\times$ pixel-shuffle. (3) Furthermore, we observe that the error of pruning VSR networks would accumulate with propagation steps increasing along with recurrent networks, which limits the efficiency and performance of pruning. Thus, we further introduce Temporal Finetuning (TF) to constrain the pruning error accumulation in recurrent networks. Overall, our main contributions are threefold: + +- Our work is necessary and timely. There is an urgent need to compress VSR models for deployment. To the best of our knowledge, we are one of the first to design a structured pruning scheme for VSR. +- We propose an integral VSR pruning scheme called Structured Sparsity Learning (SSL) for various components of VSR models, such as residual blocks, recurrent networks, and pixel-shuffle operations. +- We employ SSL to train VSR models, which surpass recent pruning schemes and lightweight VSR models. + +# 2. Related Work + +# 2.1. Video Super-Resolution + +VSR models can exploit additional information from neighboring LR frames for restoration [3, 8, 16, 17, 28, 48, + +[49,52-54]. Earlier VSR methods [1,39,49] estimate the optical flow between LR frames and perform spatial warping for alignment. Later methods resort to a more sophisticated approach of implicit alignment. Instead of image-level motion alignment, TDAN [40] and EDVR [44] work at the feature level. TDAN [40] first adopted deformable Conv [6] in VSR to align the features of different frames. EDVR [44] extended TDAN by introducing coarse-to-fine deformable alignment and a new spatial-temporal attention fusion module. RSDN [16] adopted a recurrent detail-structural block and a hidden state adaptation module to reduce the effect of appearance changes and error accumulation. Recently, BasicVSR [2] found that bidirectional propagation coupled with a simple optical flow-based feature alignment can further improve performance. Similarly, Yi et al. [50] used the bidirectional propagation framework to exploit LR frames and estimated hidden states from the past, present, and future. To compress the VSR model, Xiao et al. designed a space-time knowledge distillation scheme [47]. However, these VSR methods require high computational costs impeding their application on resource-limited devices. Different from previous methods, we focus on designing SSL to compress VSR models by pruning redundant filters. + +# 2.2. Network Pruning + +Network pruning [4, 5, 34, 36, 38] is widely used to remove a set of redundant parameters for network acceleration. Pruning methods can be divided into two branches, structured pruning [10, 14, 23, 46] and unstructured pruning [11, 12]. Structured pruning methods prune the network at the level of filters, channels, and even layers, which + +can obtain regular sparsity after pruning. This is beneficial for acceleration. In contrast, unstructured methods focus on pruning weights, leading up to much irregular sparsity. This is beneficial for compression but tends not to yield an actual acceleration [41, 46]. Specifically, Li et al. [23] applied the $\mathrm{L}_1$ -norm to measure the importance of different filters and then removed the less important ones. Afterward, Liu et al. [30] added a sparsity-inducing penalty term on scaling factors of the batch normalization layers to enforce the channels with lower scaling factors to be the less informative ones. Recently, ASSL [57] and SRPN [58] utilized aligned structured sparsity learning for structured pruning of residual blocks. In addition, Luo et al. [32] developed a residual block pruning scheme for image classification using the Convs on skip connections. However, the residual blocks of VSR networks do not have such Convs. Lin et al. [26] conducted runtime neural network pruning according to the input image. Besides, Wang et al. [43] designed an unstructured pruning scheme for single image SR tasks by using sparse Conv to skip redundant computations. Since we cannot directly apply a general structured pruning method for VSR, we explore the properties of VSR networks and develop a VSR pruning scheme in this paper. + +# 3. Methodology + +# 3.1. Overview + +Figure 2 (a) shows VSR networks based on the bidirectional recurrent structures, such as BasicVSR [2]. Given a LR frame $I_{t}$ , the forward network concatenates $I_{t}$ and the previous hidden state $H_{F,t-1}$ to extract features from $I_{t}$ and aggregate the reference information from $H_{F,t-1}$ . Similarly, the backward network extracts features from $I_{t}$ and aggregates the reference information from the future hidden state $H_{B,t+1}$ . Note that both the forward and backward networks consist of numerous residual blocks. Then, the features generated by forward and backward networks are fed into the upsampling network, which consists of multiple pixel-shuffle operations and Convs, to obtain the recovered frame $SR_{t}$ . However, SOTA VSR networks [2,3,50] need massive computation and memory resources, limiting their deployment on resource-limited devices. + +To pursue more efficient VSR networks, we specially design a VSR structured pruning scheme called Structured Sparsity Learning (SSL), according to the properties of VSR networks. Specifically, SSL has three stages, including pretraining, pruning, and finetuning. In the pretraining stage, we train a powerful VSR network. Since current VSR networks do not use BatchNorm [15], we introduce a scaling factor in pretrained VSR models to tune the sparsity of each channel and filter. In the pruning stage, we select the unimportant filters according to the pruning criterion and apply sparsity-inducing regularization on corresponding scaling factors. In addition, we propose a Residual Spar + +sity Connection (RSC) scheme to liberate the restrictions on pruning residual blocks of recurrent networks and preserve all restoration information contained in channels of feature maps for better performance. Moreover, for the upsampling networks, we specially develop a pruning scheme for the pixel-shuffle operation to guarantee the accuracy of channel-space conversion after pruning. Besides, we observe that the error of the hidden state would be amplified with the propagation in recurrent networks after pruning. Thus, in the finetuning stage, we design Temporal Finetuning (TF) to alleviate the error accumulation. + +# 3.2. Structured Sparsity Learning + +Structured Sparsity Learning (SSL) is a structured pruning scheme specially designed for VSR. It can reduce the redundancy of neural filters and obtain more efficient VSR submodels. Next, we will explain our SSL in detail. + +(1) Scaling Factor. Structured pruning aims to remove Conv filters based on a designed importance criterion. In the classification task, quite a few works use scale parameters of BatchNorm [15] to control the throughput of each filter. Zero scale parameters make the value of corresponding channels vanish. As a result, they contribute nothing to the subsequent Convs and can be removed. By regularizing the scale parameter, we can assess and tune the importance of each filter. However, the BatchNorm is not useful for SR tasks [25], and SOTA VSR networks [2,3,50] do not utilize it. Therefore, it is infeasible to apply the existing pruning schemes directly. In our pruning scheme, as shown in Fig. 1 (d) and Fig. 2 (b), we multiply the scaling factors $\gamma$ before or after Convs. Then, we perform regularization on scaling factors to enforce sparsity for pruning. + +(2) Pruning Criterion and Regularization Form. To remove the redundant filters, we need to select unimportant scaling factors $\gamma$ to induce sparsity. In previous works, ASSL [57] and SRPN [58] had to adopt a local pruning scheme (namely, scaling factors are only compared within the same layer, and each layer has the same pruning ratio) to guarantee that skip and residual connections keep the same number of filters and indices for the adding operation. Given that the importance of the Conv filters in various layers is different and our RSC does not have restrictions as ASSL, we can adopt the global pruning scheme (i.e., scaling factors of different layers are compared together). + +For the pruning criterion, we adopt the simple and practical $\mathbf{L}_1$ -norm. Specifically, given Conv kernel $\boldsymbol{W_i} \in \mathbb{R}^{C_{out} \times C_{in} \times K_h \times K_w}$ in the $i$ -th layer, we calculate the absolute weight sum of $k$ -th Conv filter $\boldsymbol{W_i}[k,\ldots] \in \mathbb{R}^{C_{in} \times K_h \times K_w}$ with $s_{i,k} = \sum |\boldsymbol{W_i}[k,\ldots]|$ . In particular, for our RSC in Fig. 1 (d), we require to additionally prune the input channels for the first Conv, and calculate its $\mathbf{L}_1$ -norm score with $s_{i,k}' = \sum |\boldsymbol{W_i}[:k,\ldots]|$ , where $\boldsymbol{W_i}[:k,\ldots] \in \mathbb{R}^{C_{out} \times K_h \times K_w}$ . Moreover, for the Conv + +![](images/d2a50c42dc03702202f93e98afc7957105627ef1628089e5cf202066e17c24f1.jpg) +Figure 2. (a) The basic architecture of the VSR methods with the bidirectional recurrent network. The forward and backward networks both consist of numerous residual blocks. The upsampling networks contain multiple pixel-shuffle operations and Convs. (b) The pruning scheme for the pixel-shuffle. For the $2 \times$ upsampling pixel-shuffle [37] operation, we take four channels with consecutive indices as the pruning unit to guarantee the accuracy of channel-space conversion after pruning. (c) The application of SSL on VSR models. + +![](images/5af617e3fde59078383e8811d28ea970988c43f10e1e8b079b44aa28aa105ee4.jpg) + +before the pixel-shuffle operation, we take four consecutive filters as a pruning unit and calculate the score with $s_{i,k} = \sum |\mathbf{W}_i[4k:4(k + 1),\ldots ]|$ . Then, given the pruning ratio $p$ and the total number of filters or channels $N$ , we sort all their $\mathbf{L}_1$ -norm scores $s$ together and choose the $N\times p$ filters with the smallest $\mathbf{L}_1$ -norm values as unimportant filters or channels, denoted as set $S$ (Fig. 2 (c)). + +After identifying the unimportant filters and channels set $S$ , we apply sparsity-inducing regularization (SIR) to the corresponding scaling factors, denoted as set $S_{sf}$ . It is notable that we do not enforce sparsity-inducing regularization to the important filters and channels since they will remain in the network. Specifically, we employ $L_{2}$ regularization on the scaling factors to enforce sparsity (Fig. 2 (c)): + +$$ +\mathcal {L} _ {S I R} = \alpha_ {\gamma} \sum_ {\gamma \in S _ {s f}} \gamma^ {2}, \tag {1} +$$ + +where $\gamma$ is the scalar selected from $\gamma \in \mathbb{R}^C$ corresponding to unimportant filters or channels; $\alpha_{\gamma}$ is a scalar. We increment $\alpha_{\gamma}$ by a presetting constant $\Delta$ every $T_{1}$ iterations. When $\alpha_{\gamma}$ reaches the pre-defined upper limit $\tau$ , we keep $\alpha_{\gamma}$ constant and continue training $T_{2}$ iterations. + +(3) Pruning Scheme for Residual Blocks of Recurrent Networks. VSR recurrent networks consist of residual blocks. Residual blocks are difficult to prune because the addition operations require the pruned filter indices between the skip and residual connections to be the same. As shown in Fig. 1 (b), quite a few pruning schemes [23, 30, 34] simply skipped the pruning of the last Conv in residual blocks, which restricted the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] pruned the last Conv in the residual block by using regularization. However, recurrent networks take the previous output as later input. Thus, pruning recurrent networks require the pruned indices of the first and last Convs to be the same. ASSL and SRPN cannot guarantee the indices of important filters + +in the first and last Convs of recurrent networks are aligned. Besides, ASSL and SRPN have to remove some channels in output feature maps and adopt a local pruning scheme, which limits pruning space and restoration information utilization (RDN [56] and ESRGAN [45] have shown restoration information of front layers can guide feature extraction of later layers). To break the restriction of pruning and fully use restoration information contained in channels of front layers, we propose the RSC to prune residual blocks in recurrent networks (Fig. 1 (d)). As we can see, RSC preserves all channels of input $F_{i}^{\prime}$ and output $F_{i+1}^{\prime}$ in the residual blocks. For the first Conv, we select the important channels (the indices not in $S$ ) to participate in the Conv operation, which can be expressed as Eq. 2. After the last Conv, we obtained $F_{i+1}$ and add $F_{i+1}$ to $F_{i}^{\prime}$ on the corresponding channel indices to obtain $F_{i+1}^{\prime}$ (Eq. 3 and Eq. 4). Furthermore, we do not prune the $1 \times 1$ Conv of upsampling networks (Fig. 2 (a)) to aggregate all preserved restoration information in $H_{F,t}, H_{B,t} \in \mathbb{R}^{C \times H \times W}$ . + +$$ +\boldsymbol {F} _ {i} = \boldsymbol {F} _ {i} ^ {\prime} \otimes (\gamma_ {j - 1} \boldsymbol {W} _ {i} \gamma_ {j}), \tag {2} +$$ + +$$ +\boldsymbol {F} _ {i + 1} = \boldsymbol {F} _ {i} \otimes \left(\boldsymbol {W} _ {i + 1} \gamma_ {j + 1}\right), \tag {3} +$$ + +$$ +\boldsymbol {F} _ {i + 1} ^ {\prime} = \boldsymbol {F} _ {i + 1} + \boldsymbol {F} _ {i} ^ {\prime}, \tag {4} +$$ + +where $\otimes$ indicates Conv. $F_{i}^{\prime}, F_{i+1}^{\prime} \in \mathbb{R}^{C \times H \times W}$ are the input and output feature maps of the residual block, respectively. $F_{i}, F_{i+1} \in \mathbb{R}^{C_{p} \times H \times W}$ are intermediate feature maps. $W_{i}, W_{i+1} \in \mathbb{R}^{C_{out} \times C_{in} \times K_{h} \times K_{w}}$ are weights of Conv kernels. $\gamma_{j-1} \in \mathbb{R}^{C_{in}}$ and $\gamma_{j}, \gamma_{j+1} \in \mathbb{R}^{C_{out}}$ are scaling factors to apply sparsity-inducing regularization. $F_{i+1}^{\prime}$ prunes some channels, and $F_{i}$ keeps all channels. In Eq. 4, $F_{i}^{\prime}$ adds $F_{i+1}$ on corresponding kept channels. It is notable that our RSC does not introduce extra parameters and computational cost. + +(4) Pruning Scheme for Pixel-Shuffle. The upsampling network of the VSR network uses Conv to increase chan + +nels of feature maps and adopts the pixel-shuffle [37] operation to convert the channels to space realizing upsampling. As shown in Fig. 2 (b), given the input feature map $F_{i} \in \mathbb{R}^{C \times H \times W}$ , we expand its channels $4 \times$ by a Conv with weight $W_{i}$ to obtain $F_{i+1}' \in \mathbb{R}^{4C \times H \times W}$ . Then, the pixel-shuffle operation takes four channels as a group to convert $F_{i+1}'$ to $F_{i+1} \in \mathbb{R}^{C \times 2H \times 2W}$ realize $2 \times$ upsampling. We observe that, if we prune the Conv before pixel-shuffle without any restriction, the pruned feature maps will be spatially disordered after passing the pixel-shuffle operation and lead to performance drop. To address the problem, we specially design a strong and simple pruning scheme for the pixel-shuffle operation. Given the input feature map, we take four filters as a pruning unit to evaluate their importance and then impose the scaling factor $\gamma_{j}$ on filters to enforce sparsity (described in (2) pruning criterion and regularization form): + +$$ +\boldsymbol {W} _ {i} ^ {\prime} = \boldsymbol {W} _ {i} [ 4 k: 4 (k + 1), \dots ] \gamma_ {j} [ k ], k \in [ 0, C _ {i n}), \tag {5} +$$ + +where $W_{i}\in \mathbb{R}^{4C_{in}\times C_{in}\times K_{h}\times K_{w}}$ is the weights of Conv kernel. $\gamma_{j}$ is the scaling factor. + +(5) Temporal Finetuning. We observe that the pruned VSR network generates a minor error in hidden state $H_{F}$ and $H_{B}$ (Fig. 2 (a)), which will be amplified as the hidden state propagates along with recurrent networks. To solve the issue, we introduce Temporal Finetuning (TF) by enforcing the hidden states of pruned networks to align with the accurate hidden states of unpruned networks: + +$$ +\mathcal {L} _ {t f} = \left\| H _ {F, T} - H _ {F, T} ^ {\prime} \right\| + \left\| H _ {B, 0} - H _ {B, 0} ^ {\prime} \right\|, \tag {6} +$$ + +where $T$ is the number of input frames, and $H_{F,T}$ and $H_{F,T}^{\prime}$ are the final hidden states after $T$ frames forward propagation in pruned and original VSR networks, separately. Similarly, $H_{B,0}$ and $H_{B,0}^{\prime}$ are the final hidden states after backward propagation in the pruned and original VSR networks. + +To train the whole VSR network, we use the Charbonnier loss [2, 44], which can be formulated as: + +$$ +\mathcal {L} _ {r e c} = \frac {1}{T} \sum_ {t = 1} ^ {T} \sqrt {\left\| S R _ {t} - H R _ {t} \right\| ^ {2} + \varepsilon^ {2}}, \tag {7} +$$ + +where $\varepsilon$ is set to $10^{-6}$ . $SR_{t}$ and $HR_{t}$ are $t$ -th reconstructed and HR frames, respectively. The overall loss function for pruned network finetuning is designed as: + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\text {t f}}. \tag {8} +$$ + +# 3.3. Arm VSR Models with SSL + +Our SSL can be used for VSR networks. BasicVSR [2] and BasicVSR++ [3] are two strong SOTA VSR methods. The dense deformable Conv of BasicVSR++ requires reading a large amount of irregular memory data, which is unsuitable for deployment on resource-limited devices without GPUs. Besides, the Second-Order Grid Propagation of BasicVSR++ will force a delay of two frames, which further impedes its usage in real-time devices. + +In our study, we use BasicVSR for VSR pruning, which is more suitable for applications on edge devices. In addition, we further propose unidirectional BasicVSR (BasicVSR-uni), obtained by removing the backward network, for online inference. Since the SpyNet of BasicVSR is used for flow estimation, we do not apply our pruning scheme SSL on it. In the pruning stage, as shown in Fig. 2 (c), we first add the scaling factor to the Conv and residual blocks as described in Sec. 3.2. Then we use the pruning criterion to select unimportant filters globally and apply sparsity-inducing regularization to the corresponding scaling factor. Afterward, we remove the unimportant Conv filters and finetune the pruned VSR network with $T_{3}$ iterations. We provide more details in Alg. 1 of supplementary. + +# 4. Experiments + +# 4.1. Experimental Settings + +We adopt two widely used datasets for training: REDS [33] and Vimeo-90K [49]. For REDS, we use REDS4 containing 4 clips as our test set. Additionally, we adopt REDSval4 as our validation set, which contains 4 clips selected from the REDS validation set. The remaining clips of REDS are used for training. In addition, we utilize Vid4 [27], UDM10 [52], and Vimeo-90K-T [49] as test sets along with Vimeo-90K. We train and test models with $4 \times$ downsampling using two degradations Bicubic (BI) and Blur Downsampling (BD) as BasicVSR did. For BI, the MATLAB function "imresize" is used for downsampling. For BD, we blur the HR images by a Gaussian filter with $\sigma = 1.6$ , followed by a subsampling every four pixels. + +We pretrain the unidirectional BasicVSR (BasicVSR-uni) as done for BasicVSR. In sparsity-inducing regularization, the iterations $T_{1}$ and $T_{2}$ are set to 5 and 3, 375 separately. The scalars $\Delta$ and $\tau$ are set to $10^{-4}$ and 0.1, respectively. Note that we fix the parameters of the flow estimator in sparsity-inducing regularization. In the pruned VSR network finetuning, we set $T_{3}$ to 300,000. We adopt the Adam optimizer [22] and Cosine Annealing scheme [31]. The initial learning rate of the flow estimator is $2.5 \times 10^{-5}$ . The learning rate for all other modules is $2 \times 10^{-4}$ . The patch size of input LR frames is $64 \times 64$ . Experiments are conducted on a server with PyTorch 1.10 and V100 GPUs. + +# 4.2. Quantitative and Qualitative Comparisons + +Since BasicVSR violates causality and cannot be evaluated online, we construct the unidirectional BasicVSR (BasicVSR-uni) by removing the backward network for online inference. We compare the proposed SSL with three other pruning schemes at pruning ratio $p = 0.5$ : $\mathrm{L}_1$ -norm pruning [23] (which simply removes filters with the smallest $\mathrm{L}_1$ -norms and is the most prevailing filter pruning method now), and ASSL [57]. We apply these pruning schemes on BasicVSR, thus obtaining $\mathrm{L}_1$ -norm-bi, ASSL-bi, and SSL-bi separately. In addition, we use these + +Table 1. Quantitative comparison (average PSNR/SSIM). Pruning schemes applied on bidirectional and unidirectional BasicVSR ("bi" and "uni") and marked in rouse and gray, respectively. * means the space-time knowledge distillation scheme [47]. We mark the best results among comparing pruning schemes in bold. The FLOPs and runtime are computed based on an LR size of $180 \times 320$ . + +
MethodsParams (M)FLOPs (G)Runtime (ms)BI degradationinBD degradationin
REDS4 [33]Vimeo-90K-T [49]Vid4 [27]UDM10 [52]Vimeo-90K-T [49]Vid4 [27]
Bicubic---26.14/0.729231.32/0.868423.78/0.634728.47/0.825331.30/0.868721.80/0.5246
VESPCN [1]-----25.35/0.7557---
SPMC [39]-----25.88/0.7752---
TOFlow [49]1.4274.9161027.98/0.799033.08/0.905425.89/0.765136.26/0.943834.62/0.9212-
DUF [21]5.81645.897428.63/0.8251--38.48/0.960536.87/0.944727.38/0.8329
RBPN [13]12.28516.0150730.09/0.859037.07/0.943527.12/0.818038.66/0.959637.20/0.9458-
EDVR-M [44]3.3304.211830.53/0.869937.09/0.944627.10/0.818639.40/0.966337.33/0.948427.45/0.8406
PFNL [51]3.0940.029529.63/0.850236.14/0.936326.73/0.802938.74/0.9627-27.16/0.8355
TGA [18]5.8694.1236----37.59/0.951627.63/0.8423
RLSP [9]4.282.349---38.48/0.960636.49/0.940327.48/0.8388
RSDN [16]6.2355.794---39.35/0.965337.23/0.947127.92/0.8505
RRN [19]3.4108.745---38.96/0.9644-27.69/0.8488
FastDVDnet* [47]2.664.3--36.12/0.934826.14/0.8029---
BasicVSR [2]4.9338.55731.42/0.890937.18/0.945027.24/0.825139.96/0.969437.53/0.949827.96/0.8553
BasicVSR-lite1.385.52430.56/0.873836.57/0.939726.86/0.812538.98/0.964536.78/0.943127.27/0.8327
L1-norm-bi [23]1.385.52430.66/0.876636.69/0.940626.87/0.812139.04/0.965036.84/0.943727.29/0.8335
ASSL-bi [57]1.385.52430.74/0.877036.75/0.941427.01/0.817639.15/0.966036.93/0.945027.40/0.8400
SSL-bi (Ours)1.092.12431.06/0.893336.82/0.941927.15/0.820839.35/0.966537.06/0.945827.56/0.8431
BasicVSR.uni [2]2.6218.13930.56/0.869836.95/0.942927.01/0.816439.25/0.964537.25/0.947227.57/0.8424
BasicVSR.uni-lite0.762.41829.95/0.856136.38/0.937226.68/0.801238.24/0.958636.38/0.938826.87/0.8157
L1-norm.uni [23]0.762.41829.97/0.857036.45/0.938126.70/0.803138.43/0.960136.53/0.940526.89/0.8187
ASSL.uni [57]0.762.41830.02/0.858936.49/0.938526.76/0.805138.48/0.960336.61/0.941627.02/0.8236
SSL.uni (Ours)0.563.91830.24/0.863336.56/0.939227.01/0.814838.68/0.961536.77/0.942927.18/0.8296
+ +pruning schemes on BasicVSR-uni, obtaining $\mathrm{L}_1$ -norm-uni, ASSL-uni, and SSL-uni. Moreover, we reduce the channels of BasicVSR and BasicVSR-uni to obtain lightweight VSR models BasicVSR-lite and BasicVSR-uni-lite, respectively. Furthermore, we compare our pruned BasciVSR and BasicVSR-uni with other lightweight VSR networks, including TOFlow [49], EDVR-M [44], RLSP [9], RSDN [16], etc. Since we only prune VSR networks, the parameters and FLOPs of the optical flow network, SPyNet [35] (Params 1.4M, Flops 19.6G) are not included. + +The quantitative results measures performance (PSNR and SSIM), the number of parameters, runtime, and FLOPs on the different methods, which are shown in Tab. 1. (1) Compared with competitive lightweight VSR networks, our SSL-bi obtains 0.53dB gain on REDS4 over EDVR-M. Note that, different from careful network designs like EDVR-M, we merely prune the BasicVSR, a simple backbone with 60 residual blocks, obtaining superior performance while only consuming less the FLOPs of EDVR-M. (2) Our SSL-bi surpasses the BasicVSR-lite by 0.5dB, and SSL-uni surpasses the BasicVSR-uni-lite by 0.29dB. This demonstrates the effectiveness of applying SSL for offline and online VSR network pruning. (3) We also adapt some SOTA pruning schemes, such as the $\mathrm{L}_1$ -norm and ASSL, to VSR networks for comparison. As a result, our SSL achieves superior performance on BasicVSR and BasicVSR-uni to other pruning schemes. This shows that SSL can make better use of the sparsity of the network and increases the efficiency of the learned network parameters. (4) Besides, comparing our pruning scheme and other VSR + +model compression method (space-time knowledge distillation scheme [47], FastDVDnet\*), our SSL-uni (0.5M parameters) surpasses the FastDVDnet\* (2.6M parameters) by 0.87dB on Vid4, which further demonstrates the effectiveness of our structured pruning scheme. Moreover, with our SSL strategy, we do not have to train a teacher network as knowledge distillation [47] methods did. + +The qualitative results are shown in Fig. 3. Our SSL-bi achieves the best visual quality containing more realistic details. These visual comparisons are consistent with the quantitative results, showing the superiority of SSL. SSL can learn to remove the redundant filters to compress a network to a smaller one while maintaining the most restoration ability. More visual results are given in supplementary. + +# 5. Ablation Study + +The Validation of Components in SSL. We conduct an ablation study to demonstrate the effectiveness of our SSL by progressively adding components. The results are shown in Tab. 2. $\mathrm{SSL}_1$ uses the aligned pruning [57] scheme for residual blocks, while $\mathrm{SSL}_4$ adopts our RSC. Comparing $\mathrm{SSL}_1$ and $\mathrm{SSL}_4$ , we can see that our RSC is superior to the advanced residual block pruning scheme. That is because RSC can break the pruning restrictions and preserve all information contained in feature channels for VSR. For $\mathrm{SSL}_2$ , we halve the number of filters in upsampling networks to keep the same model size as other models. As we can see, $\mathrm{SSL}_3$ outperforms $\mathrm{SSL}_2$ . This is because introducing the pruning scheme for the pixel-shuffle operation can increase the available pruning space in upsampling networks. Com + +![](images/3e1e950f6eab04306eb44be8b7e1640b4c40d37e72d431ec1fd673765aae7bef.jpg) + +![](images/7cc4f60e2109697bcca6ce6b674b4ffa11985003a79a2f16156e8594c4a1d407.jpg) + +![](images/dba4a77c028d9d529a681fcded9c98ad951202094c3ab1ee5d5ccf49af43aaa9.jpg) + +![](images/1b1328866a8540c1574ef6d66c8b68088c743ebf0077c19a2e47791446191a5a.jpg) + +![](images/50c2613c605d3375c0ae07ca7a5408b251558f7fc91f99ab0f00e269ce35ae1e.jpg) + +![](images/eced6bf3057191da520f1a2bc9b4bed789f989d1c534a7b981f8371626a3bbf4.jpg) + +![](images/d6ee23f1e068ae0a76e014e27f2056e0480d06df3d3e9bb7451175ea75e6706c.jpg) + +![](images/686cccca37926081a6b88ef9a29e47dfce73b3d3f0a48ac3e9dc1c7ce4b72282.jpg) + +![](images/faa4bbedc0ac47773e1ecb3414a097bc90f520ce75ed9453377e47e4eef07e78.jpg) + +![](images/6a7d28957bda741f235a68048a56956092538fe5ca4c6c30277cd22081da738d.jpg) + +![](images/0abf7d97cfbe794371efb3f3fc42e4a5c37d4ba2f034335ca79d3b758af303ca.jpg) + +![](images/6f038b3ff1ed95467138f139464eb5f81242761a22155e0fdb2259b8c48b96a9.jpg) + +![](images/95290f6b17d20ef08b14e7fc727cb399123615d14f4e97af736a13c6c2fce0e1.jpg) + +![](images/8e64b9e6c2e4a4c045ff716aed8a0b88bc5360090db6f4725f48e06cc172788e.jpg) + +![](images/432c1403ba3381c7d3acdc007f2cec353f2377e34eabd25919098d92b2ddbff4.jpg) +Figure 3. Qualitative comparison between various VSR and pruning schemes on REDS4 [33], Vid4 [27], and Viceo90K-T [49], separately. + +![](images/47095df1b9ae219a3886a79eff9f9d64083e0890ada0880320b0ed476392f4ba.jpg) + +![](images/1ef4a4f21fd5d56d2720b1417e9ce901c5d4ad61b421267a4597958ae522f418.jpg) + +![](images/62c2281011ef9812ff54bf6fc4f84c25495d68a4f48d72ef3d4877beb9d67cb4.jpg) + +![](images/c933ec84728b6a910177a3045ac093baf200f5b04fa650031318f0b000a69d33.jpg) + +![](images/384dbb92e86ece277980850b0bfc3fba1fd482c100f4697a6375e83609cfb2ec.jpg) + +![](images/8b969d2963a012b9467b8693a4c9b444428dc6644f2b4f16a54a56f69bf2298d.jpg) + +Table 2. Validation of the components in our SSL. PSNR (dB) results evaluated on REDS4 [33] $(4\times)$ . The backbone is BasicVSR [2], and the pruning ratio is set to 0.5. + +
MethodsSSL1SSL2SSL3SSL4 (Ours)
Aligned Pruning [57]✓
Residual Sparsity Connection✓✓✓
Pixel-Shuffle Pruning✓✓✓
Temporal Finetuning✓✓
PSNR (dB)30.8630.8230.9831.06
+ +paring $\mathrm{SSL}_4$ and $\mathrm{SSL}_3$ , we can see that adopting TF can bring a 0.08 dB improvement, reducing the error accumulation of hidden states in the recurrent network after pruning. Pruning Methods with Various Pruning Ratios. To further demonstrate SSL's effectiveness, we compare it with widely used pruning schemes, including $\mathrm{L}_1$ -norm [23], ASSL [57] at different pruning ratios. We use these pruning schemes to obtain numerous submodels with different FLOPs. Besides, we also adjust the channels of BasicVSR to obtain BasicVSR-lite with different FLOPs. The results are shown in Fig. 4. (1) Our SSL achieves the best performance compared with other pruning schemes at different pruning ratios and FLOPs. Note that SSL even surpasses the BasicVSR-lite in the same model size by 0.73 dB on submodels with around 3.2G FLOPs. This demonstrates the superiority of SSL on VSR. (2) With the pruning ratio increasing and FLOPs decreasing, the performance advantage + +![](images/1779c4201c4432d2a345eaeb9451ac046396d655eb03b085be043015fa6296ea.jpg) +Figure 4. PSNR (dB) comparison on REDS4 $(\times 4)$ between SSL and three other methods obtaining the same small network. + +brought by our SSL becomes more evident compared with BasicVSR-lite, $\mathrm{L}_1$ -norm and ASSL. + +Comparison with Different Pruning Criteria. We explore the influence of different pruning criteria on the pruned VSR model at different pruning ratios. Specifically, we select and remove the unimportant filters globally (namely, comparing all filters from all layers together) with minimum $\mathrm{L}_1$ -norm scores, which is expressed as "Min + Global". In addition, we select and remove the unimportant filters locally (namely, filters are compared with each other in the same layer, and each layer has the same pruning ratio) with maximum $\mathrm{L}_1$ -norm scores, which is expressed as "Max + Local". Similarly, we determine "Max + Global" and "Min + local". Furthermore, we randomly remove the unimportant filters as "Rand". Then, we com + +![](images/95e82c17eb1c142ca96a6d8dca50f3231f37faa9824076d7be2bbef84ede7c66.jpg) +(a) Forward Network. + +![](images/6c3b9deeb97a0bd1ef720d97f313941a265e2208fc196a0d0f7dceebe351149b.jpg) +(b) Backward Network. + +![](images/455419633ce7bf3c45f4eff3b6f9586de8d8dc576e771a09e2ed9f2a368ed9c9.jpg) +(c) Upsampling Network. + +Table 3. PSNR (dB) comparison on REDS4 $(4\times)$ for our pruning scheme (SSL) with different pruning criteria and pruning ratios. The unpruned model is BasicVSR [2] baseline. + +
Pruning RatiosMin Global (Ours)Max GlobalMin LocalMax LocalRand
0.331.2830.5931.2230.9731.10
0.531.0628.9030.8330.3830.68
0.730.4125.8930.2029.5130.09
+ +pare all pruning criteria at 0.3, 0.5, and 0.7 pruning ratios. The results are shown in Tab. 3. (1) The "Min + Global" pruning criterion achieves the best performance at different pruning ratios. It implies that the filters with minimum $\mathrm{L}_1$ -norm scores are relatively unimportant for VSR. Besides, this shows that the importance of filters in different layers are different and it is better to compare them together to select unimportant filters (pruning globally) for VSR. (2) The performance of "Max + Global" and "Max + local" are both inferior to "Rand" for the removal of more important filters. This implies that filters with large $\mathrm{L}_1$ -norm scores are more important than those with small ones for VSR networks. + +Pruning Ratios of Different Layers. We take BasicVSR pruned by SSL at 0.5 pruning ratio as an example. We visualize pruning ratios in different layers and show results in Fig. 5. (1) In the forward and backward networks, the pruning ratios of the first Conv input channels (corresponding to the $\gamma_{j-1}$ in Fig. 1 (d)) are lower than the pruning ratios of second Conv output filters (corresponding to the $\gamma_{j+1}$ in Fig. 1 (d)). It implies that VSR networks tend to aggregate information from the numerous input channels into several important output channels. (2) The residual blocks in deeper position of forward and backward networks tend to have minor pruning ratio. This means that residual blocks in deeper position contribute more to VSR. (3) The average pruning ratio of upsampling network is 0.2 (less than 0.5), indicating that upsampling network plays a quite important role in VSR performance. Previous works [2,9,25,44,45,51,55] have paid much attention to the design of feature extraction modules. In latter VSR research, paying more attention to upsampling network design is likely to improve VSR performance more. + +![](images/71ab01863011609003bf89f8870521b6feff264b427b36a9a18591f86b85ea06.jpg) +Figure 5. (a), (b), and (c) show the pruning ratios of Conv layers in forward, backward and upsampling networks, respectively. +Figure 6. The SSL pruning process of Convs in the BasicVSR [2]. + +Regularization Visualization. To understand how sparsity-inducing regularization works, we plot the average scaling factors in the BasicVSR [2] during applying regularization at 0.5 pruning ratio (Fig. 6). The average scaling factor is split into two parts, pruned and kept. As seen, the average scaling factor $\gamma$ of the pruned filters decreases as the corresponding penalty term $\alpha_{\gamma}$ and iterations become larger. Besides, it is interesting that the average scaling factor of the kept filters will increase without any regularization term to enforce them to be larger. It means that, as the unimportant filters are removed, the network will strengthen the kept filters to compensate for the performance, which is similar to the compensation effect in the human brain. + +# 6. Conclusion + +In this work, we propose a structured pruning scheme called SSL for efficient VSR in resource-limited situations. Specifically, for the difficulty of pruning residual blocks of recurrent networks, we propose the RSC. Compared with previous pruning schemes for residual blocks, RSC does not have pruning restrictions as other pruning schemes and can fully utilize restoration information in all channels for better performance. In addition, for the pixel-shuffle operation in the upsampling network, we specially design a pruning scheme by grouping filters to guarantee the accuracy of channel-space conversion after pruning. Furthermore, we propose Temporal Finetuning to reduce the error accumulation in recurrent networks. We apply SSL on the BasicVSR, and SSL achieves superior performance to that of recent SOTA methods, quantitatively and qualitatively. + +# References + +[1] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017. 2, 6 +[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8 +[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. arXiv preprint arXiv:2104.13371, 2021. 2, 3, 5 +[4] Jian Cheng, Pei-song Wang, Gang Li, Qing-hao Hu, and Han-qing Lu. Recent advances in efficient computation of deep convolutional neural networks. Frontiers of Information Technology & Electronic Engineering, 2018. 1, 2 +[5] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. Model compression and acceleration for deep neural networks: The principles, progress, and challenges. IEEE Signal Processing Magazine, 2018. 1, 2 +[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 2 +[7] Xiaohan Ding, Guiguang Ding, Yuchen Guo, Jungong Han, and Chenggang Yan. Approximated oracle filter pruning for destructive cnn width optimization. In ICML, 2019. 2 +[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 2 +[9] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 6, 8 +[10] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. *ICLR*, 2018. 2 +[11] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1, 2 +[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1, 2 +[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019. 6 +[14] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 1, 2 +[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 3 +[16] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2, 6 +[17] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look + +back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022. 2 +[18] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020. 6 +[19] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020. 6 +[20] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 1 +[21] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 6 +[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[23] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 1, 2, 3, 4, 5, 6, 7 +[24] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019. 1 +[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 3, 8 +[26] Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning. NeurIPS, 2017. 3 +[27] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. TPAMI, 2013. 5, 6, 7 +[28] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2 +[29] Hongying Liu, Peng Zhao, Zhubo Ruan, Fanhua Shang, and Yuanyuan Liu. Large motion video super-resolution with dual subnet and multi-stage communicated upsampling. arXiv preprint arXiv:2103.11744, 2021. 1 +[30] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3, 4 +[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5 +[32] Jian-Hao Luo and Jianxin Wu. Neural network pruning with residual-connections and limited-data. In CVPR, 2020. 3 +[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In CVPRW, 2019. 5, 6, 7 +[34] Junghun Oh, Heewon Kim, Seungjun Nah, Cheeun Hong, Jonghyun Choi, and Kyoung Mu Lee. Attentive fine-grained structured sparsity for image restoration. In CVPR, 2022. 1, 2, 4 + +[35] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 6 +[36] Russell Reed. Pruning algorithms-a survey. IEEE transactions on Neural Networks, 1993. 1, 2 +[37] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 4, 5 +[38] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. Efficient processing of deep neural networks. Synthesis Lectures on Computer Architecture, 2020. 1, 2 +[39] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 2, 6 +[40] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 2 +[41] Huan Wang, Xinyi Hu, Qiming Zhang, Yuehai Wang, Lu Yu, and Haoji Hu. Structured pruning for efficient convolutional neural networks via incremental regularization. IEEE Journal of Selected Topics in Signal Processing, 2019. 1, 3 +[42] Huan Wang, Can Qin, Yulun Zhang, and Yun Fu. Neural pruning via growing regularization. *ICLR*, 2021. 2 +[43] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4926, 2021. 3 +[44] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 1, 2, 5, 6, 8 +[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In ECCVW, 2018. 1, 4, 8 +[46] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. NeurIPS, 2016. 1, 2, 3 +[47] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, 2021. 2, 6 +[48] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In ACM MM, 2020. 2 +[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6, 7 +[50] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, 2021. 2, 3 +[51] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 1, 6, 8 + +[52] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2, 5, 6 +[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, 2022. 2 +[54] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In ICCV, 2019. 2 +[55] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 8 +[56] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 1, 4 +[57] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Aligned structured sparsity learning for efficient image superresolution. NeurIPS, 2021. 1, 2, 3, 4, 5, 6, 7 +[58] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 1, 2, 3, 4 \ No newline at end of file diff --git a/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/images.zip b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ade749351cac51884b248111009d47afd561e0fd --- /dev/null +++ b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43eee8afc3ace686eb06f662da819c19799d645d53a3e02b6cc0f930668b9c32 +size 686630 diff --git a/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/layout.json b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d764f258ec337c729f3b5ffed938e75f20e6bba --- /dev/null +++ b/2023/Structured Sparsity Learning for Efficient Video Super-Resolution/layout.json @@ -0,0 +1,9703 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 501, + 121 + ], + "type": "text", + "content": "Structured Sparsity Learning for Efficient Video Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": "Bin Xia" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": ", Jingwen He" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": ", Yulun Zhang" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": ", Yitong Wang" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": ", Yapeng Tian" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": ", Wenming Yang" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "text", + "content": ", and Luc Van Gool" + }, + { + "bbox": [ + 140, + 142, + 436, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "spans": [ + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "text", + "content": " Tsinghua University, " + }, + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "text", + "content": " Shanghai AI Laboratory, " + }, + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 171, + 452, + 185 + ], + "type": "text", + "content": " ETH Zürich," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 185, + 413, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 185, + 413, + 199 + ], + "spans": [ + { + "bbox": [ + 179, + 185, + 413, + 199 + ], + "type": "text", + "content": "4 ByteDance Inc, 5 University of Texas at Dallas" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 501 + ], + "type": "text", + "content": "The high computational costs of video super-resolution (VSR) models hinder their deployment on resource-limited devices, e.g., smartphones and drones. Existing VSR models contain considerable redundant filters, which drag down the inference efficiency. To prune these unimportant filters, we develop a structured pruning scheme called Structured Sparsity Learning (SSL) according to the properties of VSR. In SSL, we design pruning schemes for several key components in VSR models, including residual blocks, recurrent networks, and upsampling networks. Specifically, we develop a Residual Sparsity Connection (RSC) scheme for residual blocks of recurrent networks to liberate pruning restrictions and preserve the restoration information. For upsampling networks, we design a pixel-shuffle pruning scheme to guarantee the accuracy of feature channel-space conversion. In addition, we observe that pruning error would be amplified as the hidden states propagate along with recurrent networks. To alleviate the issue, we design Temporal Finetuning (TF). Extensive experiments show that SSL can significantly outperform recent methods quantitatively and qualitatively. The code is available at https://github.com/Zj-BinXia/SSL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 506, + 128, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 128, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 128, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": "Video super-resolution (VSR) aims to generate a high-resolution (HR) video from its corresponding low-resolution (LR) observation by filling in missing details. With the popularity of intelligent edge devices such as smartphones and small drones, performing VSR on these devices is in high demand. Although a variety of VSR networks [20, 24, 29, 44, 51] can achieve great performance, these models are usually difficult to be deployed on edge devices with limited computation and memory resources." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 625, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 288, + 696 + ], + "type": "text", + "content": "To alleviate this issue, we explore a new direction for effective and efficient VSR. To reduce the redundancy of Conv kernels [4,5,36,38] obtaining a more efficient VSR network, we develop a neural network pruning scheme for the VSR task for the first time. Since structured pruning [14,23,46,57] (focusing on filter pruning) can achieve" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 228, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 300 + ], + "type": "text", + "content": "an actual acceleration [41,46] superior to unstructured pruning [11,12] (focusing on weight-element pruning), we adopt structured pruning principle to develop our VSR pruning scheme. Given a powerful VSR network, our pruning scheme can find submodels under presetting pruning rate without significantly compromising performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 307, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 715 + ], + "type": "text", + "content": "Structured pruning is a general concept, and designing a concrete pruning scheme for VSR networks is challenging. (1) Recurrent networks are widely used in VSR models to extract temporal features, consisting of residual blocks (e.g., BasicVSR [2] has 60 residual blocks). However, it is hard to prune the residual blocks because the skip and residual connections ought to share the same indices [23] (Fig. 1 (a)). As shown in Fig. 1 (b), quite a few structured pruning schemes [23,34] do not prune the last Conv layer of the residual blocks, which restricts the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] introduce regularization and prune the same indices on skip and residual connections to keep channel alignment (local pruning scheme,i.e., each layer pruning the same ratio of filters). However, ASSL and SRPN still cannot achieve the potential of pruning residual blocks on recurrent networks. The recurrent networks take the previous output as later input (Fig. 2 (a)). This requires the pruned indices of the first and last Conv's in recurrent networks to be the same. But ASSL and SRPN cannot guarantee filter indices are aligned. Besides, many SR methods [45, 56] have shown that the information contained in front Conv layers can help the restoration feature extraction of later Conv layers. Thus, we design a Residual Sparsity Connection (RSC) for VSR recurrent networks, which preserves all channels of the input and output feature maps and selects the important channels for operation (Fig. 1 (d)). Compared with other pruning schemes [57, 58], RSC does not require the pruned indices of the first and last Conv's of recurrent networks to be the same, can preserve the information contained in all layers, and liberates the pruning space of the last Conv of the residual blocks without adding extra calculations. Notably, RSC can prune residual blocks globally (i.e., the filters in various layers are compared together to remove unimportant ones)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22638" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 68, + 511, + 266 + ], + "blocks": [ + { + "bbox": [ + 79, + 68, + 511, + 266 + ], + "lines": [ + { + "bbox": [ + 79, + 68, + 511, + 266 + ], + "spans": [ + { + "bbox": [ + 79, + 68, + 511, + 266 + ], + "type": "image", + "image_path": "9c09dc091ece02e6550193c068b31ab201953219c277666bcd98d51b974ff2a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 264, + 382, + 272 + ], + "lines": [ + { + "bbox": [ + 223, + 264, + 382, + 272 + ], + "spans": [ + { + "bbox": [ + 223, + 264, + 382, + 272 + ], + "type": "text", + "content": "(d) Residual Sparsity Connection Scheme (RSC, Ours)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 272, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 545, + 327 + ], + "type": "text", + "content": "Figure 1. Illustration of different schemes for pruning residual blocks of recurrent networks. (a) Structure of the residual block in the VSR network. (b) The residual block pruning schemes [7, 23, 42] do not prune the last Conv. (c) ASSL [57] and SRPN [58] prunes the same indices on skip and residual connections to keep channel alignment, which abandons some channels of input and output feature maps. (d) RSC preserves all channels of input and output feature maps, which does not need to align the pruned indices on the first and last Conv's in recurrent networks, can fully use restoration information, and can prune the first and last Conv's of residual blocks without restrictions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "type": "text", + "content": "(2) We observe that the upsampling network accounts for " + }, + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "type": "text", + "content": " of the total calculations in BasicVSR [2], which is necessary to be pruned to reduce redundancy. Since the pixel-shuffle [37] operation in VSR networks converts the channels to space, pruning the pixel-shuffle without any restrictions would cause the channel-space conversion to fail. Thus, we specially design a pixel-shuffle pruning scheme by taking four consecutive filters as the pruning unit for " + }, + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 46, + 338, + 287, + 516 + ], + "type": "text", + "content": " pixel-shuffle. (3) Furthermore, we observe that the error of pruning VSR networks would accumulate with propagation steps increasing along with recurrent networks, which limits the efficiency and performance of pruning. Thus, we further introduce Temporal Finetuning (TF) to constrain the pruning error accumulation in recurrent networks. Overall, our main contributions are threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 518, + 287, + 643 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 518, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 518, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 59, + 518, + 287, + 566 + ], + "type": "text", + "content": "- Our work is necessary and timely. There is an urgent need to compress VSR models for deployment. To the best of our knowledge, we are one of the first to design a structured pruning scheme for VSR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 569, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 569, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 569, + 287, + 616 + ], + "type": "text", + "content": "- We propose an integral VSR pruning scheme called Structured Sparsity Learning (SSL) for various components of VSR models, such as residual blocks, recurrent networks, and pixel-shuffle operations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 620, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 620, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 58, + 620, + 286, + 643 + ], + "type": "text", + "content": "- We employ SSL to train VSR models, which surpass recent pruning schemes and lightweight VSR models." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 180, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 180, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 180, + 683 + ], + "type": "text", + "content": "2.1. Video Super-Resolution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "VSR models can exploit additional information from neighboring LR frames for restoration [3, 8, 16, 17, 28, 48," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 338, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 612 + ], + "type": "text", + "content": "[49,52-54]. Earlier VSR methods [1,39,49] estimate the optical flow between LR frames and perform spatial warping for alignment. Later methods resort to a more sophisticated approach of implicit alignment. Instead of image-level motion alignment, TDAN [40] and EDVR [44] work at the feature level. TDAN [40] first adopted deformable Conv [6] in VSR to align the features of different frames. EDVR [44] extended TDAN by introducing coarse-to-fine deformable alignment and a new spatial-temporal attention fusion module. RSDN [16] adopted a recurrent detail-structural block and a hidden state adaptation module to reduce the effect of appearance changes and error accumulation. Recently, BasicVSR [2] found that bidirectional propagation coupled with a simple optical flow-based feature alignment can further improve performance. Similarly, Yi et al. [50] used the bidirectional propagation framework to exploit LR frames and estimated hidden states from the past, present, and future. To compress the VSR model, Xiao et al. designed a space-time knowledge distillation scheme [47]. However, these VSR methods require high computational costs impeding their application on resource-limited devices. Different from previous methods, we focus on designing SSL to compress VSR models by pruning redundant filters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 624, + 410, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 410, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 410, + 636 + ], + "type": "text", + "content": "2.2. Network Pruning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "Network pruning [4, 5, 34, 36, 38] is widely used to remove a set of redundant parameters for network acceleration. Pruning methods can be divided into two branches, structured pruning [10, 14, 23, 46] and unstructured pruning [11, 12]. Structured pruning methods prune the network at the level of filters, channels, and even layers, which" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22639" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 349 + ], + "type": "text", + "content": "can obtain regular sparsity after pruning. This is beneficial for acceleration. In contrast, unstructured methods focus on pruning weights, leading up to much irregular sparsity. This is beneficial for compression but tends not to yield an actual acceleration [41, 46]. Specifically, Li et al. [23] applied the " + }, + { + "bbox": [ + 47, + 72, + 289, + 349 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 47, + 72, + 289, + 349 + ], + "type": "text", + "content": "-norm to measure the importance of different filters and then removed the less important ones. Afterward, Liu et al. [30] added a sparsity-inducing penalty term on scaling factors of the batch normalization layers to enforce the channels with lower scaling factors to be the less informative ones. Recently, ASSL [57] and SRPN [58] utilized aligned structured sparsity learning for structured pruning of residual blocks. In addition, Luo et al. [32] developed a residual block pruning scheme for image classification using the Convs on skip connections. However, the residual blocks of VSR networks do not have such Convs. Lin et al. [26] conducted runtime neural network pruning according to the input image. Besides, Wang et al. [43] designed an unstructured pruning scheme for single image SR tasks by using sparse Conv to skip redundant computations. Since we cannot directly apply a general structured pruning method for VSR, we explore the properties of VSR networks and develop a VSR pruning scheme in this paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 354, + 129, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 129, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 129, + 367 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 371, + 115, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 115, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 115, + 383 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": "Figure 2 (a) shows VSR networks based on the bidirectional recurrent structures, such as BasicVSR [2]. Given a LR frame " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": ", the forward network concatenates " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": " and the previous hidden state " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "H_{F,t-1}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": " to extract features from " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": " and aggregate the reference information from " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "H_{F,t-1}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": ". Similarly, the backward network extracts features from " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": " and aggregates the reference information from the future hidden state " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "H_{B,t+1}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": ". Note that both the forward and backward networks consist of numerous residual blocks. Then, the features generated by forward and backward networks are fed into the upsampling network, which consists of multiple pixel-shuffle operations and Convs, to obtain the recovered frame " + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "inline_equation", + "content": "SR_{t}" + }, + { + "bbox": [ + 47, + 390, + 287, + 568 + ], + "type": "text", + "content": ". However, SOTA VSR networks [2,3,50] need massive computation and memory resources, limiting their deployment on resource-limited devices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "To pursue more efficient VSR networks, we specially design a VSR structured pruning scheme called Structured Sparsity Learning (SSL), according to the properties of VSR networks. Specifically, SSL has three stages, including pretraining, pruning, and finetuning. In the pretraining stage, we train a powerful VSR network. Since current VSR networks do not use BatchNorm [15], we introduce a scaling factor in pretrained VSR models to tune the sparsity of each channel and filter. In the pruning stage, we select the unimportant filters according to the pruning criterion and apply sparsity-inducing regularization on corresponding scaling factors. In addition, we propose a Residual Spar" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "sity Connection (RSC) scheme to liberate the restrictions on pruning residual blocks of recurrent networks and preserve all restoration information contained in channels of feature maps for better performance. Moreover, for the upsampling networks, we specially develop a pruning scheme for the pixel-shuffle operation to guarantee the accuracy of channel-space conversion after pruning. Besides, we observe that the error of the hidden state would be amplified with the propagation in recurrent networks after pruning. Thus, in the finetuning stage, we design Temporal Finetuning (TF) to alleviate the error accumulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 214, + 468, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 214, + 468, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 468, + 227 + ], + "type": "text", + "content": "3.2. Structured Sparsity Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 278 + ], + "type": "text", + "content": "Structured Sparsity Learning (SSL) is a structured pruning scheme specially designed for VSR. It can reduce the redundancy of neural filters and obtain more efficient VSR submodels. Next, we will explain our SSL in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 280, + 546, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 546, + 459 + ], + "type": "text", + "content": "(1) Scaling Factor. Structured pruning aims to remove Conv filters based on a designed importance criterion. In the classification task, quite a few works use scale parameters of BatchNorm [15] to control the throughput of each filter. Zero scale parameters make the value of corresponding channels vanish. As a result, they contribute nothing to the subsequent Convs and can be removed. By regularizing the scale parameter, we can assess and tune the importance of each filter. However, the BatchNorm is not useful for SR tasks [25], and SOTA VSR networks [2,3,50] do not utilize it. Therefore, it is infeasible to apply the existing pruning schemes directly. In our pruning scheme, as shown in Fig. 1 (d) and Fig. 2 (b), we multiply the scaling factors " + }, + { + "bbox": [ + 304, + 280, + 546, + 459 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 280, + 546, + 459 + ], + "type": "text", + "content": " before or after Convs. Then, we perform regularization on scaling factors to enforce sparsity for pruning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "content": "(2) Pruning Criterion and Regularization Form. To remove the redundant filters, we need to select unimportant scaling factors " + }, + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "content": " to induce sparsity. In previous works, ASSL [57] and SRPN [58] had to adopt a local pruning scheme (namely, scaling factors are only compared within the same layer, and each layer has the same pruning ratio) to guarantee that skip and residual connections keep the same number of filters and indices for the adding operation. Given that the importance of the Conv filters in various layers is different and our RSC does not have restrictions as ASSL, we can adopt the global pruning scheme (i.e., scaling factors of different layers are compared together)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": "For the pruning criterion, we adopt the simple and practical " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": "-norm. Specifically, given Conv kernel " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W_i} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times K_h \\times K_w}" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": "-th layer, we calculate the absolute weight sum of " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": "-th Conv filter " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W_i}[k,\\ldots] \\in \\mathbb{R}^{C_{in} \\times K_h \\times K_w}" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "s_{i,k} = \\sum |\\boldsymbol{W_i}[k,\\ldots]|" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": ". In particular, for our RSC in Fig. 1 (d), we require to additionally prune the input channels for the first Conv, and calculate its " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": "-norm score with " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "s_{i,k}' = \\sum |\\boldsymbol{W_i}[:k,\\ldots]|" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W_i}[:k,\\ldots] \\in \\mathbb{R}^{C_{out} \\times K_h \\times K_w}" + }, + { + "bbox": [ + 304, + 604, + 547, + 715 + ], + "type": "text", + "content": ". Moreover, for the Conv" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22640" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 284, + 232 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 284, + 232 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 284, + 232 + ], + "type": "image", + "image_path": "d2a50c42dc03702202f93e98afc7957105627ef1628089e5cf202066e17c24f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": "Figure 2. (a) The basic architecture of the VSR methods with the bidirectional recurrent network. The forward and backward networks both consist of numerous residual blocks. The upsampling networks contain multiple pixel-shuffle operations and Convs. (b) The pruning scheme for the pixel-shuffle. For the " + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": " upsampling pixel-shuffle [37] operation, we take four channels with consecutive indices as the pruning unit to guarantee the accuracy of channel-space conversion after pruning. (c) The application of SSL on VSR models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 285, + 70, + 543, + 232 + ], + "blocks": [ + { + "bbox": [ + 285, + 70, + 543, + 232 + ], + "lines": [ + { + "bbox": [ + 285, + 70, + 543, + 232 + ], + "spans": [ + { + "bbox": [ + 285, + 70, + 543, + 232 + ], + "type": "image", + "image_path": "5af617e3fde59078383e8811d28ea970988c43f10e1e8b079b44aa28aa105ee4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": "before the pixel-shuffle operation, we take four consecutive filters as a pruning unit and calculate the score with " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "s_{i,k} = \\sum |\\mathbf{W}_i[4k:4(k + 1),\\ldots ]|" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": ". Then, given the pruning ratio " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": " and the total number of filters or channels " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": ", we sort all their " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": "-norm scores " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": " together and choose the " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "N\\times p" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": " filters with the smallest " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": "-norm values as unimportant filters or channels, denoted as set " + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 285, + 287, + 369 + ], + "type": "text", + "content": " (Fig. 2 (c))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": "After identifying the unimportant filters and channels set " + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": ", we apply sparsity-inducing regularization (SIR) to the corresponding scaling factors, denoted as set " + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "inline_equation", + "content": "S_{sf}" + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": ". It is notable that we do not enforce sparsity-inducing regularization to the important filters and channels since they will remain in the network. Specifically, we employ " + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 369, + 287, + 453 + ], + "type": "text", + "content": " regularization on the scaling factors to enforce sparsity (Fig. 2 (c)):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 459, + 287, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 459, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 122, + 459, + 287, + 486 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S I R} = \\alpha_ {\\gamma} \\sum_ {\\gamma \\in S _ {s f}} \\gamma^ {2}, \\tag {1}", + "image_path": "cd84d1f65f7c1b396ecb96f96d390f0972923777759a483b988e8ece1694f0d2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " is the scalar selected from " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " corresponding to unimportant filters or channels; " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\alpha_{\\gamma}" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " is a scalar. We increment " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\alpha_{\\gamma}" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " by a presetting constant " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " every " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " iterations. When " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\alpha_{\\gamma}" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " reaches the pre-defined upper limit " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": ", we keep " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\alpha_{\\gamma}" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " constant and continue training " + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 46, + 486, + 287, + 545 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "(3) Pruning Scheme for Residual Blocks of Recurrent Networks. VSR recurrent networks consist of residual blocks. Residual blocks are difficult to prune because the addition operations require the pruned filter indices between the skip and residual connections to be the same. As shown in Fig. 1 (b), quite a few pruning schemes [23, 30, 34] simply skipped the pruning of the last Conv in residual blocks, which restricted the pruning space. Recently, as shown in Fig. 1 (c), ASSL [57] and SRPN [58] pruned the last Conv in the residual block by using regularization. However, recurrent networks take the previous output as later input. Thus, pruning recurrent networks require the pruned indices of the first and last Convs to be the same. ASSL and SRPN cannot guarantee the indices of important filters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": "in the first and last Convs of recurrent networks are aligned. Besides, ASSL and SRPN have to remove some channels in output feature maps and adopt a local pruning scheme, which limits pruning space and restoration information utilization (RDN [56] and ESRGAN [45] have shown restoration information of front layers can guide feature extraction of later layers). To break the restriction of pruning and fully use restoration information contained in channels of front layers, we propose the RSC to prune residual blocks in recurrent networks (Fig. 1 (d)). As we can see, RSC preserves all channels of input " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "F_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " and output " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "F_{i+1}^{\\prime}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " in the residual blocks. For the first Conv, we select the important channels (the indices not in " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": ") to participate in the Conv operation, which can be expressed as Eq. 2. After the last Conv, we obtained " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "F_{i+1}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " and add " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "F_{i+1}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "F_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " on the corresponding channel indices to obtain " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "F_{i+1}^{\\prime}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " (Eq. 3 and Eq. 4). Furthermore, we do not prune the " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": " Conv of upsampling networks (Fig. 2 (a)) to aggregate all preserved restoration information in " + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "inline_equation", + "content": "H_{F,t}, H_{B,t} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 285, + 545, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 366, + 518, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 518, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 366, + 518, + 545, + 533 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {i} = \\boldsymbol {F} _ {i} ^ {\\prime} \\otimes (\\gamma_ {j - 1} \\boldsymbol {W} _ {i} \\gamma_ {j}), \\tag {2}", + "image_path": "b2fbe13fb5d9e73da901381f55bb7e8336079babad3c82c9ec79a29a79a66972.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 366, + 534, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 534, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 366, + 534, + 545, + 547 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {i + 1} = \\boldsymbol {F} _ {i} \\otimes \\left(\\boldsymbol {W} _ {i + 1} \\gamma_ {j + 1}\\right), \\tag {3}", + "image_path": "e5fdf0f4a54e859029dada962ca3a7915ca70b4b508481e898d243a930cf1ed3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 366, + 548, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 548, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 366, + 548, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {i + 1} ^ {\\prime} = \\boldsymbol {F} _ {i + 1} + \\boldsymbol {F} _ {i} ^ {\\prime}, \\tag {4}", + "image_path": "09bed41727434d2a903d7428636adaedb2c3e57463d3043024aec1b51fb01a98.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " indicates Conv. " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{i}^{\\prime}, F_{i+1}^{\\prime} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " are the input and output feature maps of the residual block, respectively. " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{i}, F_{i+1} \\in \\mathbb{R}^{C_{p} \\times H \\times W}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " are intermediate feature maps. " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "W_{i}, W_{i+1} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times K_{h} \\times K_{w}}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " are weights of Conv kernels. " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\gamma_{j-1} \\in \\mathbb{R}^{C_{in}}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\gamma_{j}, \\gamma_{j+1} \\in \\mathbb{R}^{C_{out}}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " are scaling factors to apply sparsity-inducing regularization. " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{i+1}^{\\prime}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " prunes some channels, and " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " keeps all channels. In Eq. 4, " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " adds " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "F_{i+1}" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " on corresponding kept channels. It is notable that our RSC does not introduce extra parameters and computational cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "(4) Pruning Scheme for Pixel-Shuffle. The upsampling network of the VSR network uses Conv to increase chan" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22641" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": "nels of feature maps and adopts the pixel-shuffle [37] operation to convert the channels to space realizing upsampling. As shown in Fig. 2 (b), given the input feature map " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "F_{i} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": ", we expand its channels " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": " by a Conv with weight " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "F_{i+1}' \\in \\mathbb{R}^{4C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": ". Then, the pixel-shuffle operation takes four channels as a group to convert " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "F_{i+1}'" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "F_{i+1} \\in \\mathbb{R}^{C \\times 2H \\times 2W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": " realize " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": " upsampling. We observe that, if we prune the Conv before pixel-shuffle without any restriction, the pruned feature maps will be spatially disordered after passing the pixel-shuffle operation and lead to performance drop. To address the problem, we specially design a strong and simple pruning scheme for the pixel-shuffle operation. Given the input feature map, we take four filters as a pruning unit to evaluate their importance and then impose the scaling factor " + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 264 + ], + "type": "text", + "content": " on filters to enforce sparsity (described in (2) pruning criterion and regularization form):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 277, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 277, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 63, + 277, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} _ {i} ^ {\\prime} = \\boldsymbol {W} _ {i} [ 4 k: 4 (k + 1), \\dots ] \\gamma_ {j} [ k ], k \\in [ 0, C _ {i n}), \\tag {5}", + "image_path": "a17428e45891cff0e52fd8d942e7695870c7ec0806cf724152e8873097483c74.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "type": "inline_equation", + "content": "W_{i}\\in \\mathbb{R}^{4C_{in}\\times C_{in}\\times K_{h}\\times K_{w}}" + }, + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "type": "text", + "content": " is the weights of Conv kernel. " + }, + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 46, + 293, + 287, + 317 + ], + "type": "text", + "content": " is the scaling factor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": "(5) Temporal Finetuning. We observe that the pruned VSR network generates a minor error in hidden state " + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "inline_equation", + "content": "H_{F}" + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "inline_equation", + "content": "H_{B}" + }, + { + "bbox": [ + 46, + 318, + 287, + 400 + ], + "type": "text", + "content": " (Fig. 2 (a)), which will be amplified as the hidden state propagates along with recurrent networks. To solve the issue, we introduce Temporal Finetuning (TF) by enforcing the hidden states of pruned networks to align with the accurate hidden states of unpruned networks:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 404, + 287, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 404, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 75, + 404, + 287, + 419 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t f} = \\left\\| H _ {F, T} - H _ {F, T} ^ {\\prime} \\right\\| + \\left\\| H _ {B, 0} - H _ {B, 0} ^ {\\prime} \\right\\|, \\tag {6}", + "image_path": "6a4ddaa64f933293b51a921147da946e413426ce8e892482f9a10cceb1585870.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": " is the number of input frames, and " + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "inline_equation", + "content": "H_{F,T}" + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "inline_equation", + "content": "H_{F,T}^{\\prime}" + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": " are the final hidden states after " + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": " frames forward propagation in pruned and original VSR networks, separately. Similarly, " + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "inline_equation", + "content": "H_{B,0}" + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "inline_equation", + "content": "H_{B,0}^{\\prime}" + }, + { + "bbox": [ + 46, + 419, + 287, + 479 + ], + "type": "text", + "content": " are the final hidden states after backward propagation in the pruned and original VSR networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 502 + ], + "type": "text", + "content": "To train the whole VSR network, we use the Charbonnier loss [2, 44], which can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 506, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 506, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 86, + 506, + 287, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sqrt {\\left\\| S R _ {t} - H R _ {t} \\right\\| ^ {2} + \\varepsilon^ {2}}, \\tag {7}", + "image_path": "d09f8e6d6591a4b4822b37438d715c57861d82aee63f1f281fc6ced10ce54a55.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "SR_{t}" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "HR_{t}" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 540, + 287, + 576 + ], + "type": "text", + "content": "-th reconstructed and HR frames, respectively. The overall loss function for pruned network finetuning is designed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 583, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 583, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 126, + 583, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {t f}}. \\tag {8}", + "image_path": "278c0dc0d3fbb6276f508664aba5f2126c756bbcd92e69a83eff3feedc058e50.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 602, + 200, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 200, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 200, + 613 + ], + "type": "text", + "content": "3.3. Arm VSR Models with SSL" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Our SSL can be used for VSR networks. BasicVSR [2] and BasicVSR++ [3] are two strong SOTA VSR methods. The dense deformable Conv of BasicVSR++ requires reading a large amount of irregular memory data, which is unsuitable for deployment on resource-limited devices without GPUs. Besides, the Second-Order Grid Propagation of BasicVSR++ will force a delay of two frames, which further impedes its usage in real-time devices." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "In our study, we use BasicVSR for VSR pruning, which is more suitable for applications on edge devices. In addition, we further propose unidirectional BasicVSR (BasicVSR-uni), obtained by removing the backward network, for online inference. Since the SpyNet of BasicVSR is used for flow estimation, we do not apply our pruning scheme SSL on it. In the pruning stage, as shown in Fig. 2 (c), we first add the scaling factor to the Conv and residual blocks as described in Sec. 3.2. Then we use the pruning criterion to select unimportant filters globally and apply sparsity-inducing regularization to the corresponding scaling factor. Afterward, we remove the unimportant Conv filters and finetune the pruned VSR network with " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "T_{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " iterations. We provide more details in Alg. 1 of supplementary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 244, + 387, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 387, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 387, + 257 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 260, + 432, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 432, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 432, + 273 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "type": "text", + "content": "We adopt two widely used datasets for training: REDS [33] and Vimeo-90K [49]. For REDS, we use REDS4 containing 4 clips as our test set. Additionally, we adopt REDSval4 as our validation set, which contains 4 clips selected from the REDS validation set. The remaining clips of REDS are used for training. In addition, we utilize Vid4 [27], UDM10 [52], and Vimeo-90K-T [49] as test sets along with Vimeo-90K. We train and test models with " + }, + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "type": "text", + "content": " downsampling using two degradations Bicubic (BI) and Blur Downsampling (BD) as BasicVSR did. For BI, the MATLAB function \"imresize\" is used for downsampling. For BD, we blur the HR images by a Gaussian filter with " + }, + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\sigma = 1.6" + }, + { + "bbox": [ + 304, + 276, + 545, + 431 + ], + "type": "text", + "content": ", followed by a subsampling every four pixels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": "We pretrain the unidirectional BasicVSR (BasicVSR-uni) as done for BasicVSR. In sparsity-inducing regularization, the iterations " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": " are set to 5 and 3, 375 separately. The scalars " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": " and 0.1, respectively. Note that we fix the parameters of the flow estimator in sparsity-inducing regularization. In the pruned VSR network finetuning, we set " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "T_{3}" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": " to 300,000. We adopt the Adam optimizer [22] and Cosine Annealing scheme [31]. The initial learning rate of the flow estimator is " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": ". The learning rate for all other modules is " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": ". The patch size of input LR frames is " + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 432, + 545, + 575 + ], + "type": "text", + "content": ". Experiments are conducted on a server with PyTorch 1.10 and V100 GPUs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 578, + 528, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 528, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 528, + 591 + ], + "type": "text", + "content": "4.2. Quantitative and Qualitative Comparisons" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Since BasicVSR violates causality and cannot be evaluated online, we construct the unidirectional BasicVSR (BasicVSR-uni) by removing the backward network for online inference. We compare the proposed SSL with three other pruning schemes at pruning ratio " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "-norm pruning [23] (which simply removes filters with the smallest " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "-norms and is the most prevailing filter pruning method now), and ASSL [57]. We apply these pruning schemes on BasicVSR, thus obtaining " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "-norm-bi, ASSL-bi, and SSL-bi separately. In addition, we use these" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22642" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 107, + 541, + 350 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative comparison (average PSNR/SSIM). Pruning schemes applied on bidirectional and unidirectional BasicVSR (\"bi\" and \"uni\") and marked in rouse and gray, respectively. * means the space-time knowledge distillation scheme [47]. We mark the best results among comparing pruning schemes in bold. The FLOPs and runtime are computed based on an LR size of " + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "180 \\times 320" + }, + { + "bbox": [ + 47, + 70, + 547, + 105 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 107, + 541, + 350 + ], + "lines": [ + { + "bbox": [ + 49, + 107, + 541, + 350 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 541, + 350 + ], + "type": "table", + "html": "
MethodsParams (M)FLOPs (G)Runtime (ms)BI degradationinBD degradationin
REDS4 [33]Vimeo-90K-T [49]Vid4 [27]UDM10 [52]Vimeo-90K-T [49]Vid4 [27]
Bicubic---26.14/0.729231.32/0.868423.78/0.634728.47/0.825331.30/0.868721.80/0.5246
VESPCN [1]-----25.35/0.7557---
SPMC [39]-----25.88/0.7752---
TOFlow [49]1.4274.9161027.98/0.799033.08/0.905425.89/0.765136.26/0.943834.62/0.9212-
DUF [21]5.81645.897428.63/0.8251--38.48/0.960536.87/0.944727.38/0.8329
RBPN [13]12.28516.0150730.09/0.859037.07/0.943527.12/0.818038.66/0.959637.20/0.9458-
EDVR-M [44]3.3304.211830.53/0.869937.09/0.944627.10/0.818639.40/0.966337.33/0.948427.45/0.8406
PFNL [51]3.0940.029529.63/0.850236.14/0.936326.73/0.802938.74/0.9627-27.16/0.8355
TGA [18]5.8694.1236----37.59/0.951627.63/0.8423
RLSP [9]4.282.349---38.48/0.960636.49/0.940327.48/0.8388
RSDN [16]6.2355.794---39.35/0.965337.23/0.947127.92/0.8505
RRN [19]3.4108.745---38.96/0.9644-27.69/0.8488
FastDVDnet* [47]2.664.3--36.12/0.934826.14/0.8029---
BasicVSR [2]4.9338.55731.42/0.890937.18/0.945027.24/0.825139.96/0.969437.53/0.949827.96/0.8553
BasicVSR-lite1.385.52430.56/0.873836.57/0.939726.86/0.812538.98/0.964536.78/0.943127.27/0.8327
L1-norm-bi [23]1.385.52430.66/0.876636.69/0.940626.87/0.812139.04/0.965036.84/0.943727.29/0.8335
ASSL-bi [57]1.385.52430.74/0.877036.75/0.941427.01/0.817639.15/0.966036.93/0.945027.40/0.8400
SSL-bi (Ours)1.092.12431.06/0.893336.82/0.941927.15/0.820839.35/0.966537.06/0.945827.56/0.8431
BasicVSR.uni [2]2.6218.13930.56/0.869836.95/0.942927.01/0.816439.25/0.964537.25/0.947227.57/0.8424
BasicVSR.uni-lite0.762.41829.95/0.856136.38/0.937226.68/0.801238.24/0.958636.38/0.938826.87/0.8157
L1-norm.uni [23]0.762.41829.97/0.857036.45/0.938126.70/0.803138.43/0.960136.53/0.940526.89/0.8187
ASSL.uni [57]0.762.41830.02/0.858936.49/0.938526.76/0.805138.48/0.960336.61/0.941627.02/0.8236
SSL.uni (Ours)0.563.91830.24/0.863336.56/0.939227.01/0.814838.68/0.961536.77/0.942927.18/0.8296
", + "image_path": "3ef2c569b81766602acd9eae5dc375f70695367d1690342761e966b5c32bed2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": "pruning schemes on BasicVSR-uni, obtaining " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": "-norm-uni, ASSL-uni, and SSL-uni. Moreover, we reduce the channels of BasicVSR and BasicVSR-uni to obtain lightweight VSR models BasicVSR-lite and BasicVSR-uni-lite, respectively. Furthermore, we compare our pruned BasciVSR and BasicVSR-uni with other lightweight VSR networks, including TOFlow [49], EDVR-M [44], RLSP [9], RSDN [16], etc. Since we only prune VSR networks, the parameters and FLOPs of the optical flow network, SPyNet [35] (Params 1.4M, Flops 19.6G) are not included." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": "The quantitative results measures performance (PSNR and SSIM), the number of parameters, runtime, and FLOPs on the different methods, which are shown in Tab. 1. (1) Compared with competitive lightweight VSR networks, our SSL-bi obtains 0.53dB gain on REDS4 over EDVR-M. Note that, different from careful network designs like EDVR-M, we merely prune the BasicVSR, a simple backbone with 60 residual blocks, obtaining superior performance while only consuming less the FLOPs of EDVR-M. (2) Our SSL-bi surpasses the BasicVSR-lite by 0.5dB, and SSL-uni surpasses the BasicVSR-uni-lite by 0.29dB. This demonstrates the effectiveness of applying SSL for offline and online VSR network pruning. (3) We also adapt some SOTA pruning schemes, such as the " + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": "-norm and ASSL, to VSR networks for comparison. As a result, our SSL achieves superior performance on BasicVSR and BasicVSR-uni to other pruning schemes. This shows that SSL can make better use of the sparsity of the network and increases the efficiency of the learned network parameters. (4) Besides, comparing our pruning scheme and other VSR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "type": "text", + "content": "model compression method (space-time knowledge distillation scheme [47], FastDVDnet\\*), our SSL-uni (0.5M parameters) surpasses the FastDVDnet\\* (2.6M parameters) by 0.87dB on Vid4, which further demonstrates the effectiveness of our structured pruning scheme. Moreover, with our SSL strategy, we do not have to train a teacher network as knowledge distillation [47] methods did." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 438, + 547, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 523 + ], + "type": "text", + "content": "The qualitative results are shown in Fig. 3. Our SSL-bi achieves the best visual quality containing more realistic details. These visual comparisons are consistent with the quantitative results, showing the superiority of SSL. SSL can learn to remove the redundant filters to compress a network to a smaller one while maintaining the most restoration ability. More visual results are given in supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 529, + 400, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 400, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 400, + 542 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "The Validation of Components in SSL. We conduct an ablation study to demonstrate the effectiveness of our SSL by progressively adding components. The results are shown in Tab. 2. " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_1" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " uses the aligned pruning [57] scheme for residual blocks, while " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_4" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " adopts our RSC. Comparing " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_1" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_4" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", we can see that our RSC is superior to the advanced residual block pruning scheme. That is because RSC can break the pruning restrictions and preserve all information contained in feature channels for VSR. For " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_2" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", we halve the number of filters in upsampling networks to keep the same model size as other models. As we can see, " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_3" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_2" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". This is because introducing the pruning scheme for the pixel-shuffle operation can increase the available pruning space in upsampling networks. Com" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22643" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 153, + 170 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 153, + 170 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 153, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 153, + 170 + ], + "type": "image", + "image_path": "3e1e950f6eab04306eb44be8b7e1640b4c40d37e72d431ec1fd673765aae7bef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 69, + 219, + 171 + ], + "blocks": [ + { + "bbox": [ + 156, + 69, + 219, + 171 + ], + "lines": [ + { + "bbox": [ + 156, + 69, + 219, + 171 + ], + "spans": [ + { + "bbox": [ + 156, + 69, + 219, + 171 + ], + "type": "image", + "image_path": "7cc4f60e2109697bcca6ce6b674b4ffa11985003a79a2f16156e8594c4a1d407.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 221, + 69, + 283, + 171 + ], + "blocks": [ + { + "bbox": [ + 221, + 69, + 283, + 171 + ], + "lines": [ + { + "bbox": [ + 221, + 69, + 283, + 171 + ], + "spans": [ + { + "bbox": [ + 221, + 69, + 283, + 171 + ], + "type": "image", + "image_path": "dba4a77c028d9d529a681fcded9c98ad951202094c3ab1ee5d5ccf49af43aaa9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 285, + 69, + 348, + 171 + ], + "blocks": [ + { + "bbox": [ + 285, + 69, + 348, + 171 + ], + "lines": [ + { + "bbox": [ + 285, + 69, + 348, + 171 + ], + "spans": [ + { + "bbox": [ + 285, + 69, + 348, + 171 + ], + "type": "image", + "image_path": "1b1328866a8540c1574ef6d66c8b68088c743ebf0077c19a2e47791446191a5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 69, + 411, + 171 + ], + "blocks": [ + { + "bbox": [ + 350, + 69, + 411, + 171 + ], + "lines": [ + { + "bbox": [ + 350, + 69, + 411, + 171 + ], + "spans": [ + { + "bbox": [ + 350, + 69, + 411, + 171 + ], + "type": "image", + "image_path": "50c2613c605d3375c0ae07ca7a5408b251558f7fc91f99ab0f00e269ce35ae1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 413, + 69, + 475, + 171 + ], + "blocks": [ + { + "bbox": [ + 413, + 69, + 475, + 171 + ], + "lines": [ + { + "bbox": [ + 413, + 69, + 475, + 171 + ], + "spans": [ + { + "bbox": [ + 413, + 69, + 475, + 171 + ], + "type": "image", + "image_path": "eced6bf3057191da520f1a2bc9b4bed789f989d1c534a7b981f8371626a3bbf4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 477, + 69, + 539, + 171 + ], + "blocks": [ + { + "bbox": [ + 477, + 69, + 539, + 171 + ], + "lines": [ + { + "bbox": [ + 477, + 69, + 539, + 171 + ], + "spans": [ + { + "bbox": [ + 477, + 69, + 539, + 171 + ], + "type": "image", + "image_path": "d6ee23f1e068ae0a76e014e27f2056e0480d06df3d3e9bb7451175ea75e6706c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 172, + 153, + 270 + ], + "blocks": [ + { + "bbox": [ + 53, + 172, + 153, + 270 + ], + "lines": [ + { + "bbox": [ + 53, + 172, + 153, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 153, + 270 + ], + "type": "image", + "image_path": "686cccca37926081a6b88ef9a29e47dfce73b3d3f0a48ac3e9dc1c7ce4b72282.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 156, + 172, + 219, + 269 + ], + "blocks": [ + { + "bbox": [ + 156, + 172, + 219, + 269 + ], + "lines": [ + { + "bbox": [ + 156, + 172, + 219, + 269 + ], + "spans": [ + { + "bbox": [ + 156, + 172, + 219, + 269 + ], + "type": "image", + "image_path": "faa4bbedc0ac47773e1ecb3414a097bc90f520ce75ed9453377e47e4eef07e78.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 172, + 283, + 269 + ], + "blocks": [ + { + "bbox": [ + 221, + 172, + 283, + 269 + ], + "lines": [ + { + "bbox": [ + 221, + 172, + 283, + 269 + ], + "spans": [ + { + "bbox": [ + 221, + 172, + 283, + 269 + ], + "type": "image", + "image_path": "6a7d28957bda741f235a68048a56956092538fe5ca4c6c30277cd22081da738d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 285, + 172, + 348, + 269 + ], + "blocks": [ + { + "bbox": [ + 285, + 172, + 348, + 269 + ], + "lines": [ + { + "bbox": [ + 285, + 172, + 348, + 269 + ], + "spans": [ + { + "bbox": [ + 285, + 172, + 348, + 269 + ], + "type": "image", + "image_path": "0abf7d97cfbe794371efb3f3fc42e4a5c37d4ba2f034335ca79d3b758af303ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 350, + 172, + 411, + 269 + ], + "blocks": [ + { + "bbox": [ + 350, + 172, + 411, + 269 + ], + "lines": [ + { + "bbox": [ + 350, + 172, + 411, + 269 + ], + "spans": [ + { + "bbox": [ + 350, + 172, + 411, + 269 + ], + "type": "image", + "image_path": "6f038b3ff1ed95467138f139464eb5f81242761a22155e0fdb2259b8c48b96a9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 413, + 172, + 475, + 265 + ], + "blocks": [ + { + "bbox": [ + 413, + 172, + 475, + 265 + ], + "lines": [ + { + "bbox": [ + 413, + 172, + 475, + 265 + ], + "spans": [ + { + "bbox": [ + 413, + 172, + 475, + 265 + ], + "type": "image", + "image_path": "95290f6b17d20ef08b14e7fc727cb399123615d14f4e97af736a13c6c2fce0e1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 477, + 172, + 539, + 265 + ], + "blocks": [ + { + "bbox": [ + 477, + 172, + 539, + 265 + ], + "lines": [ + { + "bbox": [ + 477, + 172, + 539, + 265 + ], + "spans": [ + { + "bbox": [ + 477, + 172, + 539, + 265 + ], + "type": "image", + "image_path": "8e64b9e6c2e4a4c045ff716aed8a0b88bc5360090db6f4725f48e06cc172788e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 53, + 274, + 153, + 373 + ], + "blocks": [ + { + "bbox": [ + 53, + 274, + 153, + 373 + ], + "lines": [ + { + "bbox": [ + 53, + 274, + 153, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 153, + 373 + ], + "type": "image", + "image_path": "432c1403ba3381c7d3acdc007f2cec353f2377e34eabd25919098d92b2ddbff4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 379, + 545, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 379, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 545, + 390 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison between various VSR and pruning schemes on REDS4 [33], Vid4 [27], and Viceo90K-T [49], separately." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 156, + 274, + 219, + 373 + ], + "blocks": [ + { + "bbox": [ + 156, + 274, + 219, + 373 + ], + "lines": [ + { + "bbox": [ + 156, + 274, + 219, + 373 + ], + "spans": [ + { + "bbox": [ + 156, + 274, + 219, + 373 + ], + "type": "image", + "image_path": "47095df1b9ae219a3886a79eff9f9d64083e0890ada0880320b0ed476392f4ba.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 221, + 274, + 283, + 373 + ], + "blocks": [ + { + "bbox": [ + 221, + 274, + 283, + 373 + ], + "lines": [ + { + "bbox": [ + 221, + 274, + 283, + 373 + ], + "spans": [ + { + "bbox": [ + 221, + 274, + 283, + 373 + ], + "type": "image", + "image_path": "1ef4a4f21fd5d56d2720b1417e9ce901c5d4ad61b421267a4597958ae522f418.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 285, + 274, + 348, + 373 + ], + "blocks": [ + { + "bbox": [ + 285, + 274, + 348, + 373 + ], + "lines": [ + { + "bbox": [ + 285, + 274, + 348, + 373 + ], + "spans": [ + { + "bbox": [ + 285, + 274, + 348, + 373 + ], + "type": "image", + "image_path": "62c2281011ef9812ff54bf6fc4f84c25495d68a4f48d72ef3d4877beb9d67cb4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 350, + 274, + 411, + 373 + ], + "blocks": [ + { + "bbox": [ + 350, + 274, + 411, + 373 + ], + "lines": [ + { + "bbox": [ + 350, + 274, + 411, + 373 + ], + "spans": [ + { + "bbox": [ + 350, + 274, + 411, + 373 + ], + "type": "image", + "image_path": "c933ec84728b6a910177a3045ac093baf200f5b04fa650031318f0b000a69d33.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 413, + 274, + 475, + 373 + ], + "blocks": [ + { + "bbox": [ + 413, + 274, + 475, + 373 + ], + "lines": [ + { + "bbox": [ + 413, + 274, + 475, + 373 + ], + "spans": [ + { + "bbox": [ + 413, + 274, + 475, + 373 + ], + "type": "image", + "image_path": "384dbb92e86ece277980850b0bfc3fba1fd482c100f4697a6375e83609cfb2ec.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 477, + 274, + 539, + 373 + ], + "blocks": [ + { + "bbox": [ + 477, + 274, + 539, + 373 + ], + "lines": [ + { + "bbox": [ + 477, + 274, + 539, + 373 + ], + "spans": [ + { + "bbox": [ + 477, + 274, + 539, + 373 + ], + "type": "image", + "image_path": "8b969d2963a012b9467b8693a4c9b444428dc6644f2b4f16a54a56f69bf2298d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 59, + 433, + 273, + 506 + ], + "blocks": [ + { + "bbox": [ + 46, + 399, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 431 + ], + "type": "text", + "content": "Table 2. Validation of the components in our SSL. PSNR (dB) results evaluated on REDS4 [33] " + }, + { + "bbox": [ + 46, + 399, + 287, + 431 + ], + "type": "inline_equation", + "content": "(4\\times)" + }, + { + "bbox": [ + 46, + 399, + 287, + 431 + ], + "type": "text", + "content": ". The backbone is BasicVSR [2], and the pruning ratio is set to 0.5." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 433, + 273, + 506 + ], + "lines": [ + { + "bbox": [ + 59, + 433, + 273, + 506 + ], + "spans": [ + { + "bbox": [ + 59, + 433, + 273, + 506 + ], + "type": "table", + "html": "
MethodsSSL1SSL2SSL3SSL4 (Ours)
Aligned Pruning [57]✓
Residual Sparsity Connection✓✓✓
Pixel-Shuffle Pruning✓✓✓
Temporal Finetuning✓✓
PSNR (dB)30.8630.8230.9831.06
", + "image_path": "b5c2febe44a9f1d996165cfb529f68395e77f82d6e63bdebb98924c2cf0b9832.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": "paring " + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_4" + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_3" + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": ", we can see that adopting TF can bring a 0.08 dB improvement, reducing the error accumulation of hidden states in the recurrent network after pruning. Pruning Methods with Various Pruning Ratios. To further demonstrate SSL's effectiveness, we compare it with widely used pruning schemes, including " + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": "-norm [23], ASSL [57] at different pruning ratios. We use these pruning schemes to obtain numerous submodels with different FLOPs. Besides, we also adjust the channels of BasicVSR to obtain BasicVSR-lite with different FLOPs. The results are shown in Fig. 4. (1) Our SSL achieves the best performance compared with other pruning schemes at different pruning ratios and FLOPs. Note that SSL even surpasses the BasicVSR-lite in the same model size by 0.73 dB on submodels with around 3.2G FLOPs. This demonstrates the superiority of SSL on VSR. (2) With the pruning ratio increasing and FLOPs decreasing, the performance advantage" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 328, + 407, + 509, + 502 + ], + "blocks": [ + { + "bbox": [ + 328, + 407, + 509, + 502 + ], + "lines": [ + { + "bbox": [ + 328, + 407, + 509, + 502 + ], + "spans": [ + { + "bbox": [ + 328, + 407, + 509, + 502 + ], + "type": "image", + "image_path": "1779c4201c4432d2a345eaeb9451ac046396d655eb03b085be043015fa6296ea.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 502, + 545, + 523 + ], + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 523 + ], + "type": "text", + "content": "Figure 4. PSNR (dB) comparison on REDS4 " + }, + { + "bbox": [ + 305, + 502, + 545, + 523 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 305, + 502, + 545, + 523 + ], + "type": "text", + "content": " between SSL and three other methods obtaining the same small network." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 531, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 531, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 305, + 531, + 545, + 554 + ], + "type": "text", + "content": "brought by our SSL becomes more evident compared with BasicVSR-lite, " + }, + { + "bbox": [ + 305, + 531, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 305, + 531, + 545, + 554 + ], + "type": "text", + "content": "-norm and ASSL." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Comparison with Different Pruning Criteria. We explore the influence of different pruning criteria on the pruned VSR model at different pruning ratios. Specifically, we select and remove the unimportant filters globally (namely, comparing all filters from all layers together) with minimum " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "-norm scores, which is expressed as \"Min + Global\". In addition, we select and remove the unimportant filters locally (namely, filters are compared with each other in the same layer, and each layer has the same pruning ratio) with maximum " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "-norm scores, which is expressed as \"Max + Local\". Similarly, we determine \"Max + Global\" and \"Min + local\". Furthermore, we randomly remove the unimportant filters as \"Rand\". Then, we com" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22644" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 83, + 236, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 83, + 236, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 83, + 236, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 236, + 194 + ], + "type": "image", + "image_path": "95e82c17eb1c142ca96a6d8dca50f3231f37faa9824076d7be2bbef84ede7c66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 198, + 182, + 206 + ], + "lines": [ + { + "bbox": [ + 115, + 198, + 182, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 198, + 182, + 206 + ], + "type": "text", + "content": "(a) Forward Network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 248, + 83, + 428, + 194 + ], + "blocks": [ + { + "bbox": [ + 248, + 83, + 428, + 194 + ], + "lines": [ + { + "bbox": [ + 248, + 83, + 428, + 194 + ], + "spans": [ + { + "bbox": [ + 248, + 83, + 428, + 194 + ], + "type": "image", + "image_path": "6c3b9deeb97a0bd1ef720d97f313941a265e2208fc196a0d0f7dceebe351149b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 198, + 378, + 207 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 378, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 378, + 207 + ], + "type": "text", + "content": "(b) Backward Network." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 437, + 87, + 517, + 196 + ], + "blocks": [ + { + "bbox": [ + 437, + 87, + 517, + 196 + ], + "lines": [ + { + "bbox": [ + 437, + 87, + 517, + 196 + ], + "spans": [ + { + "bbox": [ + 437, + 87, + 517, + 196 + ], + "type": "image", + "image_path": "455419633ce7bf3c45f4eff3b6f9586de8d8dc576e771a09e2ed9f2a368ed9c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 198, + 517, + 208 + ], + "lines": [ + { + "bbox": [ + 438, + 198, + 517, + 208 + ], + "spans": [ + { + "bbox": [ + 438, + 198, + 517, + 208 + ], + "type": "text", + "content": "(c) Upsampling Network." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 60, + 259, + 274, + 327 + ], + "blocks": [ + { + "bbox": [ + 47, + 225, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 225, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 287, + 258 + ], + "type": "text", + "content": "Table 3. PSNR (dB) comparison on REDS4 " + }, + { + "bbox": [ + 47, + 225, + 287, + 258 + ], + "type": "inline_equation", + "content": "(4\\times)" + }, + { + "bbox": [ + 47, + 225, + 287, + 258 + ], + "type": "text", + "content": " for our pruning scheme (SSL) with different pruning criteria and pruning ratios. The unpruned model is BasicVSR [2] baseline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 259, + 274, + 327 + ], + "lines": [ + { + "bbox": [ + 60, + 259, + 274, + 327 + ], + "spans": [ + { + "bbox": [ + 60, + 259, + 274, + 327 + ], + "type": "table", + "html": "
Pruning RatiosMin Global (Ours)Max GlobalMin LocalMax LocalRand
0.331.2830.5931.2230.9731.10
0.531.0628.9030.8330.3830.68
0.730.4125.8930.2029.5130.09
", + "image_path": "7ef00972cee5a0213f50621e26b702ff10e3408f937a993f48bc9bdec41298c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "content": "pare all pruning criteria at 0.3, 0.5, and 0.7 pruning ratios. The results are shown in Tab. 3. (1) The \"Min + Global\" pruning criterion achieves the best performance at different pruning ratios. It implies that the filters with minimum " + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "content": "-norm scores are relatively unimportant for VSR. Besides, this shows that the importance of filters in different layers are different and it is better to compare them together to select unimportant filters (pruning globally) for VSR. (2) The performance of \"Max + Global\" and \"Max + local\" are both inferior to \"Rand\" for the removal of more important filters. This implies that filters with large " + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 46, + 329, + 287, + 472 + ], + "type": "text", + "content": "-norm scores are more important than those with small ones for VSR networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": "Pruning Ratios of Different Layers. We take BasicVSR pruned by SSL at 0.5 pruning ratio as an example. We visualize pruning ratios in different layers and show results in Fig. 5. (1) In the forward and backward networks, the pruning ratios of the first Conv input channels (corresponding to the " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{j-1}" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": " in Fig. 1 (d)) are lower than the pruning ratios of second Conv output filters (corresponding to the " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{j+1}" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": " in Fig. 1 (d)). It implies that VSR networks tend to aggregate information from the numerous input channels into several important output channels. (2) The residual blocks in deeper position of forward and backward networks tend to have minor pruning ratio. This means that residual blocks in deeper position contribute more to VSR. (3) The average pruning ratio of upsampling network is 0.2 (less than 0.5), indicating that upsampling network plays a quite important role in VSR performance. Previous works [2,9,25,44,45,51,55] have paid much attention to the design of feature extraction modules. In latter VSR research, paying more attention to upsampling network design is likely to improve VSR performance more." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 228, + 526, + 338 + ], + "blocks": [ + { + "bbox": [ + 67, + 209, + 525, + 220 + ], + "lines": [ + { + "bbox": [ + 67, + 209, + 525, + 220 + ], + "spans": [ + { + "bbox": [ + 67, + 209, + 525, + 220 + ], + "type": "text", + "content": "Figure 5. (a), (b), and (c) show the pruning ratios of Conv layers in forward, backward and upsampling networks, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 228, + 526, + 338 + ], + "lines": [ + { + "bbox": [ + 323, + 228, + 526, + 338 + ], + "spans": [ + { + "bbox": [ + 323, + 228, + 526, + 338 + ], + "type": "image", + "image_path": "71ab01863011609003bf89f8870521b6feff264b427b36a9a18591f86b85ea06.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 338, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 350 + ], + "type": "text", + "content": "Figure 6. The SSL pruning process of Convs in the BasicVSR [2]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "type": "text", + "content": "Regularization Visualization. To understand how sparsity-inducing regularization works, we plot the average scaling factors in the BasicVSR [2] during applying regularization at 0.5 pruning ratio (Fig. 6). The average scaling factor is split into two parts, pruned and kept. As seen, the average scaling factor " + }, + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "type": "text", + "content": " of the pruned filters decreases as the corresponding penalty term " + }, + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{\\gamma}" + }, + { + "bbox": [ + 304, + 354, + 545, + 510 + ], + "type": "text", + "content": " and iterations become larger. Besides, it is interesting that the average scaling factor of the kept filters will increase without any regularization term to enforce them to be larger. It means that, as the unimportant filters are removed, the network will strengthen the kept filters to compensate for the performance, which is similar to the compensation effect in the human brain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 516, + 378, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 378, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 378, + 529 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose a structured pruning scheme called SSL for efficient VSR in resource-limited situations. Specifically, for the difficulty of pruning residual blocks of recurrent networks, we propose the RSC. Compared with previous pruning schemes for residual blocks, RSC does not have pruning restrictions as other pruning schemes and can fully utilize restoration information in all channels for better performance. In addition, for the pixel-shuffle operation in the upsampling network, we specially design a pruning scheme by grouping filters to guarantee the accuracy of channel-space conversion after pruning. Furthermore, we propose Temporal Finetuning to reduce the error accumulation in recurrent networks. We apply SSL on the BasicVSR, and SSL achieves superior performance to that of recent SOTA methods, quantitatively and qualitatively." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22645" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In CVPR, 2017. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "type": "text", + "content": "[3] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. arXiv preprint arXiv:2104.13371, 2021. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] Jian Cheng, Pei-song Wang, Gang Li, Qing-hao Hu, and Han-qing Lu. Recent advances in efficient computation of deep convolutional neural networks. Frontiers of Information Technology & Electronic Engineering, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 268, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 288, + 312 + ], + "type": "text", + "content": "[5] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. Model compression and acceleration for deep neural networks: The principles, progress, and challenges. IEEE Signal Processing Magazine, 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 345 + ], + "type": "text", + "content": "[6] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 378 + ], + "type": "text", + "content": "[7] Xiaohan Ding, Guiguang Ding, Yuchen Guo, Jungong Han, and Chenggang Yan. Approximated oracle filter pruning for destructive cnn width optimization. In ICML, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "type": "text", + "content": "[8] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 445 + ], + "type": "text", + "content": "[9] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In ICCVW, 2019. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 288, + 479 + ], + "type": "text", + "content": "[10] Xitong Gao, Yiren Zhao, Lukasz Dudziak, Robert Mullins, and Cheng-zhong Xu. Dynamic channel pruning: Feature boosting and suppression. *ICLR*, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 288, + 523 + ], + "type": "text", + "content": "[11] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 524, + 288, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 288, + 556 + ], + "type": "text", + "content": "[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. NeurIPS, 28, 2015. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "type": "text", + "content": "[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Recurrent back-projection network for video superresolution. In CVPR, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 623 + ], + "type": "text", + "content": "[14] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "text", + "content": "[15] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[16] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In ECCV, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[17] Takashi Isobe, Xu Jia, Xin Tao, Changlin Li, Ruihuang Li, Yongjie Shi, Jing Mu, Huchuan Lu, and Yu-Wing Tai. Look" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "back and forth: Video super-resolution with explicit temporal difference modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[18] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In CVPR, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[19] Takashi Isobe, Fang Zhu, Xu Jia, and Shengjin Wang. Revisiting temporal modeling for video super-resolution. BMVC, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 215 + ], + "type": "text", + "content": "[20] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 216, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 216, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 545, + 259 + ], + "type": "text", + "content": "[21] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In CVPR, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 260, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 545, + 293 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 326 + ], + "type": "text", + "content": "[23] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "text", + "content": "[24] Sheng Li, Fengxiang He, Bo Du, Lefei Zhang, Yonghao Xu, and Dacheng Tao. Fast spatio-temporal residual network for video super-resolution. In CVPR, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "type": "text", + "content": "[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 3, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 393, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 415 + ], + "type": "text", + "content": "[26] Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 416, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 437 + ], + "type": "text", + "content": "[27] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. TPAMI, 2013. 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 480 + ], + "type": "text", + "content": "[28] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[29] Hongying Liu, Peng Zhao, Zhubo Ruan, Fanhua Shang, and Yuanyuan Liu. Large motion video super-resolution with dual subnet and multi-stage communicated upsampling. arXiv preprint arXiv:2103.11744, 2021. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "text", + "content": "[30] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 3, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 603, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 624 + ], + "type": "text", + "content": "[32] Jian-Hao Luo and Jianxin Wu. Neural network pruning with residual-connections and limited-data. In CVPR, 2020. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and super-resolution: Dataset and study. In CVPRW, 2019. 5, 6, 7" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[34] Junghun Oh, Heewon Kim, Seungjun Nah, Cheeun Hong, Jonghyun Choi, and Kyoung Mu Lee. Attentive fine-grained structured sparsity for image restoration. In CVPR, 2022. 1, 2, 4" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22646" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 95 + ], + "type": "text", + "content": "[35] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 286, + 118 + ], + "type": "text", + "content": "[36] Russell Reed. Pruning algorithms-a survey. IEEE transactions on Neural Networks, 1993. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 120, + 286, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 120, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 120, + 286, + 174 + ], + "type": "text", + "content": "[37] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 176, + 286, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 176, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 176, + 286, + 209 + ], + "type": "text", + "content": "[38] Vivienne Sze, Yu-Hsin Chen, Tien-Ju Yang, and Joel S Emer. Efficient processing of deep neural networks. Synthesis Lectures on Computer Architecture, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 211, + 286, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 211, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 211, + 286, + 243 + ], + "type": "text", + "content": "[39] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In ICCV, 2017. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 246, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 246, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 49, + 246, + 286, + 278 + ], + "type": "text", + "content": "[40] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 280, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 280, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 280, + 286, + 323 + ], + "type": "text", + "content": "[41] Huan Wang, Xinyi Hu, Qiming Zhang, Yuehai Wang, Lu Yu, and Haoji Hu. Structured pruning for efficient convolutional neural networks via incremental regularization. IEEE Journal of Selected Topics in Signal Processing, 2019. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 326, + 286, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 286, + 347 + ], + "type": "text", + "content": "[42] Huan Wang, Can Qin, Yulun Zhang, and Yun Fu. Neural pruning via growing regularization. *ICLR*, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 350, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 286, + 403 + ], + "type": "text", + "content": "[43] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4917-4926, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 406, + 286, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 406, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 406, + 286, + 448 + ], + "type": "text", + "content": "[44] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 452, + 286, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 452, + 286, + 494 + ], + "spans": [ + { + "bbox": [ + 49, + 452, + 286, + 494 + ], + "type": "text", + "content": "[45] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In ECCVW, 2018. 1, 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 497, + 286, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 497, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 49, + 497, + 286, + 528 + ], + "type": "text", + "content": "[46] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. NeurIPS, 2016. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 532, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 532, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 532, + 286, + 563 + ], + "type": "text", + "content": "[47] Zeyu Xiao, Xueyang Fu, Jie Huang, Zhen Cheng, and Zhiwei Xiong. Space-time distillation for video super-resolution. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 567, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 286, + 597 + ], + "type": "text", + "content": "[48] Zeyu Xiao, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Space-time video super-resolution using temporal profiles. In ACM MM, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 601, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 286, + 632 + ], + "type": "text", + "content": "[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 635, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 666 + ], + "type": "text", + "content": "[50] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video super-resolution. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[51] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 1, 6, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 319 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[52] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In ICCV, 2019. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[53] Jiyang Yu, Jingen Liu, Liefeng Bo, and Tao Mei. Memory-augmented non-local attention for video super-resolution. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "text", + "content": "[54] Haochen Zhang, Dong Liu, and Zhiwei Xiong. Two-stream action recognition-oriented video super-resolution. In ICCV, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[55] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[56] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 1, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[57] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Aligned structured sparsity learning for efficient image superresolution. NeurIPS, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 319 + ], + "type": "text", + "content": "[58] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22647" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_content_list.json b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..06708da69af1aeb420f681d74f1ab150a9c245ab --- /dev/null +++ b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_content_list.json @@ -0,0 +1,1667 @@ +[ + { + "type": "text", + "text": "StyLess: Boosting the Transferability of Adversarial Examples", + "text_level": 1, + "bbox": [ + 168, + 130, + 800, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaisheng Liang Bin Xiao* The Hong Kong Polytechnic University", + "bbox": [ + 321, + 181, + 640, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cskliang, csbxiao}@comp.polyu.edu.hk", + "bbox": [ + 320, + 219, + 643, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial attacks can mislead deep neural networks (DNNs) by adding imperceptible perturbations to benign examples. The attack transferability enables adversarial examples to attack black-box DNNs with unknown architectures or parameters, which poses threats to many real-world applications. We find that existing transferable attacks do not distinguish between style and content features during optimization, limiting their attack transferability. To improve attack transferability, we propose a novel attack method called style-less perturbation (StyLess). Specifically, instead of using a vanilla network as the surrogate model, we advocate using stylized networks, which encode different style features by perturbing an adaptive instance normalization. Our method can prevent adversarial examples from using non-robust style features and help generate transferable perturbations. Comprehensive experiments show that our method can significantly improve the transferability of adversarial examples. Furthermore, our approach is generic and can outperform state-of-the-art transferable attacks when combined with other attack techniques.", + "bbox": [ + 75, + 300, + 473, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) [14, 24] are currently effective methods for solving various challenging tasks such as computer vision, and natural language processing. Although DNNs have amazing accuracy, especially for computer vision tasks such as image classification, they are also known to be vulnerable to adversarial examples [12, 43]. Adversarial examples are malicious images obtained by adding imperceptible perturbations to benign images. Notably, the transferability of adversarial examples is an intriguing phenomenon, which refers to the property that the same adversarial example can successfully attack different black-box DNNs [5, 30, 34, 51].", + "bbox": [ + 75, + 655, + 468, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It has been observed that image style can be decoupled from image content, and style transfer techniques allow us", + "bbox": [ + 76, + 835, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e121519470c08256f3fa8e7871280b7004d4e0a1f88e0e784aafcfa05fbe2db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 271, + 883, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b184b7496c37c27594dbd1c96ddb100e8ad8204cc0a9fcb9db059ac1deee6572.jpg", + "image_caption": [ + "Benign Sample" + ], + "image_footnote": [], + "bbox": [ + 508, + 449, + 625, + 539 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c69e622c4e8ff9912cb4e41cdf21686622998cb2e660aea10b2dd4308582521.jpg", + "image_caption": [ + "Stylized Output" + ], + "image_footnote": [], + "bbox": [ + 638, + 449, + 754, + 539 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fc1b5901b3f78bfe106d5fd1f94eef84d8c0d87fcd0657060e53eb3de1a844f.jpg", + "image_caption": [ + "StyLess Attack", + "Figure 1. An overview of our StyLess attack. We create stylized model $\\bar{F}$ by injecting synthesized style features into the surrogate model $(F = F_{2} \\circ F_{1})$ using an adaptive IN layer. StyLess reduces the use of non-robust style features in the vanilla surrogate model $F$ , ultimately improving attack transferability." + ], + "image_footnote": [], + "bbox": [ + 767, + 449, + 883, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to generate stylized images based on arbitrary style images [18]. Image style refers to the unique visual characteristics of an image, including its colors, textures, and lighting. For instance, two photos of the same object taken by different photographers can have very different styles. Robust DNNs should rely more on content features of data than style features. This inspired us to improve attack transferability from the perspective of avoiding non-robust features. We believe that style features of DNNs are non-robust for building transferable attacks. However, existing attacks do not distinguish between the surrogate models' style and content features, which may reduce attack transferability.", + "bbox": [ + 496, + 642, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose using stylized surrogate models to control style features, which can significantly improve transferability. We refer to the original surrogate model as the \"vanilla model.\" The proposed stylized model is created by adding an adaptive instance normalization (IN) layer to the vanilla", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code is available at https://github.com/uhiu/StyLess", + "bbox": [ + 94, + 887, + 397, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model. By adjusting the parameters of the inserted IN layer, we can easily transform the style features of the surrogate models. To compare the stylized and vanilla surrogate models, we analyzed their network losses during optimization. Surprisingly, we found that the adversarial loss of the vanilla model increased much faster than those of the stylized models, resulting in a widening loss gap. This phenomenon reveals that as the attack iteration progresses, current attack methods only focus on maximizing the loss of the vanilla model, leading to increased use of the style features of the vanilla surrogate model. However, we believe that style features are non-robust for transferable attacks, and relying too much on them may reduce attack transferability. To enhance transferability, we aim to limit the use of non-robust style features and close the loss gap.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above findings, we propose a novel method called StyLess to improve the transferability of adversarial examples. Our method uses multiple synthesized style features to compete with the original style features during the iterative optimization of attack. The process is illustrated in Figure 1. We encode various synthesized style features into a surrogate model via an IN layer to achieve stylized surrogate models. Instead of using only the vanilla surrogate model, we use the gradients of both the stylized surrogate models and the vanilla one to update adversarial examples. The front part of the surrogate model works as a style encoder, and the IN layer simulates synthesized style features. Although we can use a decoder to explicitly generate the final stylized samples, it is unnecessary for the proposed attack method. Experimental results demonstrate that StyLess can enhance the transferability of state-of-the-art adversarial attacks on both unsecured and secured black-box DNNs. Our main contributions are summarized as follows:", + "bbox": [ + 75, + 316, + 472, + 588 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel perspective for interpreting attack transferability: the original style features may hinder transferability. We verify that current iterative attacks increasingly use the style features of the surrogate model during the optimization process.", + "- We propose a novel attack called StyLess to enhance transferability by minimizing the use of original style features. To achieve this, we insert an IN layer to create stylized surrogate models and use gradients from both stylized and vanilla models.", + "- We conducted comprehensive experiments on various black-box DNNs to demonstrate that StyLess can significantly improve attack transferability. Furthermore, we show that StyLess is a generic approach that can be combined with existing attack techniques." + ], + "bbox": [ + 94, + 590, + 468, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 218, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attacks. Adversarial attacks reveal the vulnerability of current DNNs [43]. The classic adversarial attack methods are gradient-based, such as FGSM [12] and", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "I-FGSM [25]. C&W [4] considers optimizing the distance between adversarial examples and benign samples, and proposed optimization-based attacks. Adversarial attacks can also be performed in the physical world [10, 39]. As for defending against adversarial examples, adversarial training is a popular defense method that uses adversarial examples as extra training data to improve robustness [32].", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Increasing Attack Transferability. An intriguing property of adversarial attacks is the transferability. Ensemble-based attack [30] uses multiple surrogate networks instead of one network. Ghost networks [27] generates different surrogate networks by perturbing skip connection and dropout layers. Optimization methods, such as MI [5], uses a momentum-based optimization, while VT [47] introduces gradient variance to control the stability of the localized gradients. RAP [35] generates adversarial examples located in a flat loss region. Data augmentation methods, such as DI [51], uses image transformation like resizing and padding, while TI [6] considers translating image pixels. SI [29] calculates gradients with the help of several scaled benign samples. Admix [48] calculates iterative gradients by mixing the benign images with randomly sampled images.", + "bbox": [ + 496, + 199, + 892, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various network architectures and features exhibit different relationships with adversarial attacks. DNNs' linearity is believed to cause adversarial vulnerability [12], and LinBP [13] skips the nonlinear activation during the backpropagation. SGM [49] uses more gradients through skip connections in residual networks. To better leverage the intermediate layers, one can train auxiliary classifiers based on feature spaces [20, 21], maximize the distance between natural images and their adversarial examples in feature spaces [54], or fine-tune the existing adversarial examples in intermediate layer level by ILA [17, 26].", + "bbox": [ + 496, + 428, + 892, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Style Transfer and Instance Normalization. Style transfer can change the style of an image to match the style of another one [9, 11, 23]. Fast feedforward networks can perform stylization with arbitrary styles in a single forward pass [18, 28]. Interestingly, style transfer has a wide range of applications. AdvCam [8] uses natural styles to hide non- $L_{p}$ restricted perturbations. FSA [52] generates natural-looking adversarial examples by using optimized style changes. Style transfer has also been used to improve network robustness by exploring additional feature information [33]. Latent style transformations can detect adversarial attacks [46]. AMT-GAN [15] proposes an adversarial makeup transfer to protect facial privacy by preserving stronger black-box transferability.", + "bbox": [ + 496, + 595, + 892, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The family of instance normalization (IN) including batch normalization [22], layer normalization [1], instance normalization [45], and group normalization [50]. Normalizations are mainly used to reduce the covariate shift, and speed model training. Recently, normalizations have been found to be related to robustness. It has been shown that", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "batch normalization makes DNNs use more non-robust but useful features [3, 19]. AdvBN proposed adding an extra batch normalization into network training to increase training loss adversarially, which enables the network to resist various domain shifts [40]. Adjusting batch normalization statistics such as the running mean and variance in the inference phase, which are estimated during training, improves robustness and defense common corruption [2, 38].", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among existing style-based attacks, FSA [52] differentiates style features and content features, which is similar to our method. However, there are three significant differences between FSA and our approach: 1) FSA proposes to hide adversarial perturbations in the optimized style, while we avoid relying on any style. 2) FSA aims at enhancing the natural looking of non- $\\ell_p$ restricted attacks, while we focus on the transferability of $\\ell_p$ restricted adversarial examples. 3) Both FSA and our work are inspired by AdaIN [18], but we use IN layer differently. FSA perturbs the IN layer to search malicious styles and requires a decoder. But we use randomized IN layers to augment attacks and don't need to train a decoder.", + "bbox": [ + 75, + 213, + 470, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 424, + 210, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Threat Model", + "text_level": 1, + "bbox": [ + 76, + 450, + 218, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attack objective. Given a benign image $x$ with label $y$ , transfer-based attacks aim to generate an adversarial perturbation based on a white-box surrogate network $F$ . The general attack objective can be formulated as follows:", + "bbox": [ + 75, + 474, + 468, + 535 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\delta} \\mathcal {L} (F (x + \\delta), y) \\quad \\text {s . t .} \\quad \\| \\delta \\| \\leq \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 547, + 468, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ denotes the adversarial loss, $\\delta$ is the adversarial perturbation, and $\\epsilon$ is the maximum perturbation size.", + "bbox": [ + 75, + 582, + 468, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A popular framework to solve the above problem is iterative fast gradient sign method (I-FGSM) [12, 25]:", + "bbox": [ + 76, + 614, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {a d v} ^ {t + 1} = x _ {a d v} ^ {t} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {a d v} ^ {t}\\right), y\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 656, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is the learning rate, and a clip function will be used on $x_{adv}^{t + 1}$ to ensure $\\| x_{adv}^{t + 1} - x\\| \\leq \\epsilon$", + "bbox": [ + 75, + 686, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attacker capability. We follow the same setting in previous work that attackers have a surrogate model and some test samples, but cannot access target models, and don't know network architectures, training data, or defense strategies. It should be noted that our method doesn't require any additional datasets. Our approach involves style features, which can be extracted from an arbitrary image or synthesized without any style image.", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transferable attacks as black-box attacks. Transferable attacks use the surrogate model $F$ to create adversarial examples that can fool unseen target models. In this way, these attacks can be viewed as black-box attacks.", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Motivation", + "text_level": 1, + "bbox": [ + 500, + 90, + 620, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing transferable attacks often rely on the gradient of the adversarial loss function $\\mathcal{L}$ (Equation 2) without considering the impact of different components of $\\mathcal{L}$ . However, these approaches have limitations because transferable attacks should minimize the use of non-robust features of the surrogate model. Interestingly, for image classification task, style features of images are typically less robust than content features. Based on this observation, we propose to enhance attack transferability by explicitly reducing the use of style features of the surrogate model within the loss $\\mathcal{L}$ .", + "bbox": [ + 496, + 113, + 890, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key idea is simulating various surrogate models without the style features of the given vanilla surrogate model. We discovered that inserting an IN layer into the vanilla surrogate model enables us to create new surrogate models that we refer to as stylized surrogate models. Sometimes we omit the word \"surrogate.\" Stylized models can explicitly manipulate style features without compromising model accuracy. Recall that our goal is to construct adversarial examples that can mislead unseen target models, which should include these stylized models. However, existing methods, such as MI and I, only focus on maximizing the loss of the vanilla model.", + "bbox": [ + 496, + 265, + 890, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 indicates that MI and I(-FGSM) have limited attack transferability on stylized models since the vanilla model's adversarial loss increases much faster than stylized models', resulting in a widening loss gap. In the following sections, we will demonstrate how our method addresses this issue by maximizing the loss of both the stylized and vanilla models, which significantly improves transferability.", + "bbox": [ + 496, + 446, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Stylized Surrogate Models", + "text_level": 1, + "bbox": [ + 500, + 560, + 738, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section will give the definition of our stylized surrogate models, which encode various style features by inserting an IN layer. Then we will analyze the stylized loss gap $(\\Delta \\mathcal{L})$ between the vanilla and stylized surrogate models. Specifically, we will illustrate the increasing $\\Delta \\mathcal{L}$ limits transferability and our idea to decrease $\\Delta \\mathcal{L}$ .", + "bbox": [ + 496, + 584, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1 Encoding Styles by Stylized Models", + "text_level": 1, + "bbox": [ + 500, + 681, + 800, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a classifier $F = F_{2} \\circ F_{1}$ as the surrogate model, we define a stylized surrogate model as", + "bbox": [ + 496, + 704, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {F} _ {x _ {s}} = F _ {2} \\circ \\mathrm {I N} _ {x _ {s}} \\circ F _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 744, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{s}$ is a style input, $\\mathrm{IN}_{x_s}$ is an IN layer instantiated by $x_{s}$ . In general, an IN layer is defined as", + "bbox": [ + 496, + 773, + 888, + 803 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I N} (x; \\mu , \\sigma) = \\sigma \\cdot \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 811, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ are the network parameters of layer IN, and $\\mu(x)$ and $\\sigma(x)$ are the mean and variance of input $x$ . According to adaptive instance normalization (AdaIN) [18], to", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8165", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3a4f1f6154ac9eb82e4d946815110808987b22bd0e05522feb72b5946012cdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 271, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57ef0de4d49b9bdbb8839d88f796e1e89860a296ac4f5c1c09c51f49238c7752.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 89, + 450, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c00564f5812bb613c7a341be67caeb0ffee7a5d35392dc08f953cc55d0f4f0a6.jpg", + "image_caption": [ + "(a) Baseline attack MI vs ours $\\mathrm{MI + StyLess}$", + "Figure 2. Illustration of the loss gap between vanilla network and stylized network, with RN50 as surrogate model. The greater the loss, the better the attack performance. The widening loss gap in the baseline means that attack performance on stylized models is lagging behind. The proposed StyLess can close the gap." + ], + "image_footnote": [], + "bbox": [ + 94, + 236, + 272, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2684c3128e67515e02360ce3985f115f4c886f044307a715c14891482fd808e9.jpg", + "image_caption": [ + "(b) Baseline attack I(-FGSM) vs ours I+StyLess" + ], + "image_footnote": [], + "bbox": [ + 272, + 236, + 450, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stylize an input $x$ with a given style input $x_{s}$ , we only need to instantiate the IN as", + "bbox": [ + 75, + 484, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\operatorname {I N} _ {x _ {s}} (x) = \\operatorname {I N} (x) \\right| _ {\\mu = \\mu \\left(x _ {s}\\right), \\sigma = \\sigma \\left(x _ {s}\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 529, + 468, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our stylized model $\\bar{F}_{x_s}$ has encoded style features. Based on AdaIN, the $F_{1}$ from the classifier $F = F_{2} \\circ F_{1}$ works as an encoder for style transfer, and given a style input $x_{s}$ , we can get a stylized image as $\\bar{x} = D \\circ \\mathrm{IN}_{x_s} \\circ F_1(x)$ where $D$ denotes a decoder. Thus, $\\mathrm{IN}_{x_s} \\circ F_1(x)$ has encoded the style features of $x_{s}$ to $\\bar{F}_{x_s}$ .", + "bbox": [ + 75, + 560, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Stylized Loss Gap Limits transferability", + "text_level": 1, + "bbox": [ + 76, + 662, + 410, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To validate the performance of adversarial attacks on these stylized surrogate models, we define a stylized loss gap as", + "bbox": [ + 75, + 688, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathcal {L} = \\mathbb {E} _ {x _ {s} \\in \\mathcal {D}} [ \\mathcal {L} (F (x), y) - \\mathcal {L} (\\bar {F} _ {x _ {s}} (x), y) ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 732, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F$ and $\\bar{F}_{x_s}$ are vanilla and stylized model, respectively; $x_{s}$ is style input.", + "bbox": [ + 75, + 762, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An increasing loss gap $\\Delta \\mathcal{L}$ can be observed in Figure 2 which limits attack transferability, as explained below. Hypothetically, let's say we can decouple style-dependent loss from content-dependent loss in $\\mathcal{L}$ as follows:", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} ^ {c} + \\mathcal {L} _ {x} ^ {s}, \\\\ \\bar {\\bar {z}} = z ^ {c} - z ^ {s} \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 867, + 468, + 891 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} ^ {c} + \\mathcal {L} _ {x _ {s}} ^ {s},\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 888, + 325, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is the vanilla loss, and $\\bar{\\mathcal{L}}$ is the stylized one; $\\mathcal{L}^c$ is the common content-dependent loss; $\\mathcal{L}_x^s$ is input sample $x$ -specific style-dependent loss, while $\\mathcal{L}_{x_s}^s$ is style sample $x_{s}$ -specific style-dependent loss. In this case, $\\Delta \\mathcal{L} = \\mathcal{L}_x^s -\\mathcal{L}_{x_s}^s$", + "bbox": [ + 500, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, the non-robust features are related to $\\mathcal{L}_x^s$ and $\\mathcal{L}_{x_s}^s$ , while $\\mathcal{L}_c$ is supposed to be shared by other unseen DNNs. MI and I-(FGSM) have increasing gaps in Figure 2 means that attackers gradually focus on optimizing the $\\mathcal{L}_x^s$ part, which only belonged to the vanilla surrogate model $F$ . Therefore, the loss gap limits the transferability of adversarial examples in unseen stylized models.", + "bbox": [ + 498, + 151, + 890, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To decrease $\\Delta \\mathcal{L}$ and boost transferability, we consider involving $\\mathcal{L}_{x_s}^s$ as a competitor of $\\mathcal{L}_x^s$ in optimization process to suppress the growth of $\\mathcal{L}_x^s$ . In general, we can assume that all these losses in Equation 7 are non-negative, and they satisfy: $\\mathcal{L}^c\\gg \\mathcal{L}_x^s$ , $\\mathcal{L}^c\\gg \\mathcal{L}_{x_s}^s$ . Also, there is a upper bound $B$ for these losses: $\\mathcal{L} < B$ , $\\bar{\\mathcal{L}} < B$ , as the adversarial perturbations are required to smaller than a given $\\epsilon$ . If we only maximize the vanilla loss $\\mathcal{L}$ , both $\\mathcal{L}^c$ and $\\mathcal{L}_x^s$ are likely to be increased. To this end, we propose to maximize $\\mathbb{E}_{x_s\\in \\mathcal{D}}\\mathcal{L}_{x_s}^s +\\mathcal{L}_x^s +\\mathcal{L}^c$ , which involves multiple style-dependent losses $\\mathcal{L}_{x_s}^s (x_s\\in \\mathcal{D})$ to compete with $\\mathcal{L}_x^s$ and leads to a decrease of $\\Delta \\mathcal{L}$ .", + "bbox": [ + 498, + 257, + 892, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Proposed Style-Less Perturbations (StyLess)", + "text_level": 1, + "bbox": [ + 500, + 446, + 875, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above analysis, we propose style-less perturbations (StyLess) method to increase attack transferability by optimizing stylized loss and vanilla loss together:", + "bbox": [ + 498, + 470, + 890, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\delta} \\mathbb {E} _ {x _ {s} \\in \\mathcal {D}} \\mathcal {L} \\left(\\bar {F} _ {x _ {s}} (x + \\delta), y\\right) + \\mathcal {L} \\left(F (x + \\delta), y\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 525, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key to generate multiple stylized models $\\bar{F}_{x_s}$ is synthesizing style statistics $\\mu, \\sigma$ for Equation 5 to obtain parameterized IN layers. We propose using scaling and interpolation to simulate multiple style features, formulated as", + "bbox": [ + 498, + 558, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\beta (\\lambda \\mu_ {x} + (1 - \\lambda) \\mu_ {s}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 628, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\gamma (\\lambda \\sigma_ {x} + (1 - \\lambda) \\sigma_ {s}),\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 648, + 782, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_x, \\sigma_x$ is the mean and variance of $F_1(x)$ , relating to the benign content input $x$ , while $\\mu_s, \\sigma_s$ are for style input $x_s$ similarly. $x_s$ is an arbitrary image. $\\lambda$ is a scalar that controls the interpolation of two styles. $\\beta$ and $\\gamma$ are $c$ -dimensional vectors that scale the synthesized style, where $c$ refers to the number of channels in the style feature.", + "bbox": [ + 498, + 672, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We summarize the proposed StyLess attack in Algorithm 1. To obtain the parameters of the IN layer without training, we select a random image as $x_{s}$ , and generate a pair of $\\mu$ and $\\sigma$ using Equation 9. With $\\mu$ and $\\sigma$ , we can create a stylized model $\\bar{F}_{x_s}(x)$ using Equation 3 and 5. $\\bar{F}_{x_s}(x)$ need to be equal to $F(x)$ or label $y$ , otherwise we will regenerate $\\mu$ and $\\sigma$ by altering $\\beta, \\gamma$ and $\\lambda$ . In each iteration, we use $N$ stylized models to update adversarial examples based on the proposed objective function in Equation 8.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8166", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Style-Less Perturbations (StyLess) Algorithm \nInput: Surrogate model $F$ , benign example $x$ , iteration number $T$ , maximum perturbation $\\epsilon$ , data augmentation $\\phi(\\cdot)$ , decay factor $\\eta$ . Scale factors $\\beta, \\gamma$ and interpolation factor $\\lambda$ . The number of stylized models $N$ . \nOutput: An adversarial example $x_{adv}$ \n1: $x_{adv} = x$ , $g_0 = 0$ , $\\alpha = \\epsilon/2$ . \n2: for $t = 0 \\rightarrow T - 1$ do: \n3: Augment input $x_{adv} = \\phi(x_{adv})$ \n4: Obtain gradient $\\tilde{g}_{t+1}$ with respect to $x_{adv}$ using $F$ \n5: repeat \n6: Synthesize a style statistic by Equation 9 \n7: Obtain a stylized model $\\bar{F}_{x_s}$ by Equation 3 \n8: Get gradient $\\tilde{g}$ with respect to $x_{adv}$ using $\\bar{F}_{x_s}$ \n9: Update $\\tilde{g}_{t+1} = \\tilde{g}_{t+1} + \\tilde{g}$ \n10: until obtain $N$ stylized models \n11: Calculate momentum $g_{t+1} = \\eta \\cdot g_t + \\tilde{g}_{t+1} / \\| \\tilde{g}_{t+1} \\|_1$ \n12: Update example $x_{adv} = x_{adv} + \\alpha \\cdot \\text{sign}(g_{t+1})$ \n13: end for \n14: return $x_{adv}$ .", + "bbox": [ + 76, + 89, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 426, + 209, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 452, + 266, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We use ImageNet [36] for experiments. Specifically, we use 1000 images which are randomly selected from ImageNet validation set by Wang et al. [47]. Similar dataset settings have been widely used in previous work [5, 13, 29, 47, 49, 51]. These images include all categories; almost all are correctly classified by the target DNNs.", + "bbox": [ + 75, + 476, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. We evaluate the generated adversarial examples on different black-box DNNs, including both unsecured and secured models. Unsecured models are trained on ImageNet using traditional methods, while secured models are based on adversarial training. The unsecured models include VGG19 [41], AlexNet [24], ResNet50 (RN50) [14], WideResNet101 (WRN101) [53], DenseNet121 (DN121) [16], InceptionV3 (IncV3) [42], MnasNet [44], MobileNetV2 (MNv2) [37], ShuffleNetV2 (SNv2) [31] and ViT [7]. Their pre-training parameters are obtained from PyTorch official. The secured DNNs are IncV3ens3 (ensemble of 3 InceptionV3 networks), IncV3ens4 (ensemble of 4 InceptionV3 networks) and IncResV2ens3 (ensemble of 3 IncResV2 networks). These models were adversarily trained and widely used in previous work [5,30,47,49]. As for the surrogate models, we use VGG19 [41], RN50 [14], WRN101 [53] and DN121 [16].", + "bbox": [ + 75, + 566, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use I-FGSM [12] as the initial baseline. Unless otherwise specified, the attacks are untargeted and $l_{\\infty}$ -restricted. The maximum perturbation size is set to $\\epsilon = 16 / 255$ . We compare StyLess primarily with six transferable attacks: MI [5], DI [51], TI [6], SI [29],", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Admix (AI) [48], and an ensemble-based approach [30]. We set the optimization step size to $\\alpha = \\epsilon /2$ , and the number of iterations to $T = 50$ . The momentum decay in MI is $\\mu = 1$ . For DI, SI and AI, we follow the official settings described in the corresponding papers. For StyLess, we simulate 10 stylized models in each iteration, denoted by $N = 10$ . To generate a stylized model for a given $x$ , we randomly sample $\\lambda$ from $[0,0.2]$ , and $\\beta, \\gamma$ from $[0,2]$ , and ensure that $\\bar{F}_{x_s}(x)$ is equal to $F(x)$ or the real label. The IN layer is inserted after the first bottleneck block for RN50 and WRN101, and after the first dense block for DN121.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Attacking Unsecured Models", + "text_level": 1, + "bbox": [ + 498, + 267, + 758, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare StyLess with other attacks on various unsecured DNNs using three surrogate models. Table 1 shows that StyLess is a powerful and generic method that can be combined with existing attack methods to further improve attack transferability. Specifically, we compare StyLess with I, MI, DI, TI, SI and Admix (AI). For the most challenging case in the table, attacking the black-box IncV3 (let's take $\\mathrm{RN50} \\Rightarrow \\mathrm{IncV3}$ attack as an example), StyLess significantly improves the attack success rate of baseline attacks (I and MI): $46.2\\% \\rightarrow 68.3\\%$ (I), $59.2\\% \\rightarrow 78.9\\%$ (MI). StyLess also demonstrates its capabilities when using other DNNs as the surrogate network. For instance, DN121 $\\Rightarrow$ SNv2 attack, StyLess significantly improves the baseline: $69.3\\% \\rightarrow 91.4\\%$ (I), $77.4\\% \\rightarrow 95.1\\%$ (MI).", + "bbox": [ + 496, + 291, + 890, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "StyLess can be combined with other attack techniques. Previous work has shown that combining various attack methods results in powerful and transferable attacks. StyLess can be integrated with existing combination-based attacks to enhance attack transferability. We report the results of four combinations of existing attacks: MDI, MTDI, MTDSI, and MTDAI. StyLess further enhances these four attack methods' attack success rate by $+8.0\\%$ , $+8.2\\%$ , $+3.0\\%$ , and $+3.3\\%$ (in the case of $\\mathrm{RN}50 \\Rightarrow \\mathrm{IncV}3$ attack). We evaluate attack performance using various surrogate models, including RN50, WRN101 and DN121. For instance, when combining MTDAI with StyLess, our method enhances the transferability of MTDAI: WRN101 $\\Rightarrow$ IncV3 attack: $+4.7\\%$ , WRN101 $\\Rightarrow$ MNv2 attack: $+1.5\\%$ , WRN101 $\\Rightarrow$ SNv2 attack: $+3.4\\%$ , DN121 $\\Rightarrow$ IncV3 attack: $+4.3\\%$ , DN121 $\\Rightarrow$ MNv2 attack: $+0.6\\%$ , DN121 $\\Rightarrow$ SNv2 attack: $+3.8\\%$ . The results show that StyLess is an efficient and generic approach for improving attack transferability.", + "bbox": [ + 496, + 503, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Attacking Secured Models", + "text_level": 1, + "bbox": [ + 498, + 786, + 738, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate StyLess on three widely-used secured models: IncV3ens3, IncV3ens4 and IncResV2ens, as shown in Table 2. We present the results of I, MI, MDI, MTDI, and MTDSI on the three secured models using two surrogate models: RN50 and WRN101. These secured networks are more robust than the unsecured DNNs we men", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8167", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/19ddfe3d92ad4c37566afcc3cf0b6bd4c2ad41660e6360cc020e6952cae06eee.jpg", + "table_caption": [ + "Table 1. Attacking unsecured black-box models with StyLess." + ], + "table_footnote": [], + "table_body": "
SourceAttackVGG19RN50WRN101DN121IncV3MNv2SNv2
RN50I / +Ours72.8 / 88.8100 / 10080.8 / 97.183.0 / 97.846.2 / 68.377.1 / 91.960.6 / 77.8
MI / +Ours82.9 / 94.1100 / 10083.9 / 97.287.5 / 98.759.2 / 78.983.8 / 93.272.3 / 83.5
MDI / +Ours97.5 / 99.2100 / 10098.2 / 99.899.4 / 10089.5 / 97.598.1 / 99.988.5 / 96.9
MTDI / +Ours98.6 / 99.7100 / 10099.2 / 10099.8 / 10090.2 / 98.498.7 / 10090.8 / 98.1
MTDSI / +Ours98.6 / 99.2100 / 10099.5 / 10099.8 / 10096.2 / 99.298.7 / 99.996.2 / 98.2
MTDAI / +Ours99.3 / 99.6100 / 10099.8 / 10099.9 / 10095.5 / 98.899.7 / 10095.7 / 99.0
WRN101I / +Ours64.8 / 79.688.6 / 98.7100 / 10077.7 / 94.243.2 / 66.668.1 / 84.758.1 / 73.2
MI / +Ours75.6 / 86.489.8 / 98.7100 / 10083.9 / 96.457.6 / 73.776.1 / 87.070.2 / 81.2
MDI / +Ours92.2 / 98.599.0 / 99.9100 / 10097.1 / 99.986.3 / 96.393.6 / 99.086.8 / 95.8
MTDI / +Ours92.5 / 98.699.2 / 100100 / 10097.9 / 99.689.2 / 97.995.4 / 99.388.4 / 97.0
MTDSI / +Ours95.9 / 98.999.6 / 100100 / 10099.7 / 10095.8 / 99.597.7 / 99.994.5 / 98.4
MTDAI / +Ours96.5 / 99.299.9 / 99.9100 / 10099.4 / 10093.7 / 98.498.1 / 99.694.6 / 98.0
DN121I / +Ours79.7 / 97.787.0 / 99.474.7 / 98.0100 / 10055.7 / 88.781.3 / 97.069.3 / 91.4
MI / +Ours86.9 / 99.189.4 / 99.778.2 / 98.6100 / 10066.0 / 95.786.5 / 98.777.4 / 95.1
MDI / +Ours96.7 / 99.998.9 / 10095.4 / 99.8100 / 10090.5 / 99.097.5 / 10090.9 / 98.6
MTDI / +Ours97.5 / 99.999.1 / 10095.9 / 99.8100 / 10092.7 / 99.496.9 / 10092.2 / 98.6
MTDSI / +Ours97.4 / 99.999.3 / 10098.0 / 99.8100 / 10096.5 / 99.898.9 / 10095.1 / 99.1
MTDAI / +Ours99.2 / 99.999.8 / 10098.1 / 99.9100 / 10095.1 / 99.499.4 / 10095.2 / 99.0
", + "bbox": [ + 99, + 107, + 874, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f656f5acfc0c984e6d62cb435692c9608089a7391f38216d71c6b7b049c4bea1.jpg", + "table_caption": [ + "Table 2. Attacking three secured black-box models with StyLess. The surrogate model is RN50 or WRN101." + ], + "table_footnote": [], + "table_body": "
AttackRN50⇒
IncV3ens3IncV3ens4IncResV2ens
I / +Ours21.6 / 34.618.9 / 32.014.1 / 21.5
MI/ +Ours31.5 / 47.129.3 / 42.420.8 / 31.0
MDI / +Ours59.6 / 78.153.5 / 69.838.3 / 57.3
MTDI / +Ours69.2 / 89.663.8 / 81.854.6 / 72.2
MTDSI / +Ours88.0 / 93.184.7 / 91.377.8 / 84.5
WRN101⇒
I / +Ours23.4 / 40.320.9 / 34.615.5 / 25.6
MI/ +Ours35.0 / 51.230.4 / 46.624.7 / 35.9
MDI / +Ours64.0 / 81.558.8 / 76.246.9 / 66.4
MTDI / +Ours75.5 / 91.470.5 / 87.463.6 / 81.0
MTDSI / +Ours91.6 / 97.588.6 / 96.183.0 / 92.5
", + "bbox": [ + 83, + 444, + 460, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tioned above. Taking $\\mathrm{RN50} \\Rightarrow \\mathrm{IncV3}_{\\mathrm{ens3}}$ attack as an example, I, MI, MDI, and MTDI achieve attack success rates of only $21.6\\%$ , $31.5\\%$ , $59.6\\%$ , and $69.2\\%$ , respectively. StyLess effectively improves the performance of these attacks. Notably, MTDSI has been improved by our method: $91.6\\% \\rightarrow 97.5\\%$ in WRN101 $\\Rightarrow$ IncV3ens3 attack; $88.6\\% \\rightarrow$ $96.1\\%$ in WRN101 $\\Rightarrow$ IncV3ens4 attack; $83.0\\% \\rightarrow 92.5\\%$ in WRN101 $\\Rightarrow$ IncResV2ens attack. StyLess demonstrates its great power to enhance transferability in breaking these challenging secured DNNs.", + "bbox": [ + 75, + 681, + 468, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 shows a comparison of StyLess and LinBP. We consider LinBP and its combination with existing attacks such as ILA, SGM, and MDI on various target DNNs. Our method exhibits the best attack transferability.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbf30828713e1c2f8a39b49983f4f7d50063c040818e9006083e276d10e45445.jpg", + "image_caption": [ + "Figure 3. Comparing with LinBP on different black-box models and defenses. The surrogate model is RN50." + ], + "image_footnote": [], + "bbox": [ + 513, + 414, + 875, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Combining with Ensemble-Based Method", + "text_level": 1, + "bbox": [ + 500, + 633, + 856, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generating adversarial examples based on multiple surrogate networks simultaneously can improve attack performance in practice [30]. This ensemble-based method has been shown to be a powerful attack, and StyLess can further improve it. We use an ensemble of RN50, WRN101, and DN121 as the integrated surrogate model.", + "bbox": [ + 496, + 657, + 890, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared our StyLess with ensemble-based MI and MTDI in Table 3. The considerable strength of the ensemble-based method can be seen when comparing with the results in Table 1 and 2. For example, in RN50 $\\Rightarrow$ IncV3 attack, the baseline attack MI only gets $59.2\\%$ , while it gets $81.0\\%$ attack success rate in the case of RN50+WRN101+DN121 $\\Rightarrow$ IncV3 attack. Noted that we generally use $\\epsilon = 16$ , which is a standard setting. As we can see, when $\\epsilon = 16$ , the ensemble-based MTDI achieves an average attack success rate of less than $90\\%$ on the three", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8168", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2fb2777cf91eeb99c67f3b725da9b27aa549a7005849aa042df58899bec4fa2d.jpg", + "table_caption": [ + "Table 3. Combining with ensemble-based attack method." + ], + "table_footnote": [], + "table_body": "
εAttackRN50+WRN101+DN121⇒
VGG19RN50WRN101DN121IncV3MNv2SNv2IncV3ens3IncV3ens4IncResV2ens
4MI56.896.494.897.235.762.348.318.415.89.4
MTDI79.297.996.698.867.584.571.636.532.122.3
+StyLess90.499.599.499.983.294.385.450.042.929.1
8MI79.299.698.799.458.383.368.031.426.818.2
MTDI95.299.999.599.989.597.590.467.661.846.9
+StyLess99.110010010097.599.697.285.177.466.1
16MI92.310010010081.093.285.053.046.335.0
MTDI99.210010010097.510097.992.786.480.7
+StyLess10010010010099.810010098.597.594.3
", + "bbox": [ + 91, + 107, + 879, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "secured networks: IncV3ens3, IncV3ens4, and IncResV2ens. StyLess can further boost the transferability of ensemble-based MTDI: $92.7\\% \\rightarrow 98.5\\%$ for IncV3ens3, $86.4\\% \\rightarrow 97.5\\%$ for IncV3ens4, $80.7\\% \\rightarrow 94.3\\%$ for IncResV2ens. These results show that StyLess is a different type of attack, and can work perfectly with the ensemble-based method.", + "bbox": [ + 75, + 297, + 467, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also report the experimental results with different $\\epsilon$ . We use $\\epsilon = 4$ or 8 to increase the difficulty to attack. For instance, in the case of ensemble-based $\\mathrm{MTDI} \\Rightarrow \\mathrm{IncV3}_{\\mathrm{ens3}}$ attack, attack success rate drops from $92.7\\%$ to $67.6\\%$ when $\\epsilon = 8$ instead of $\\epsilon = 16$ , and StyLess can help ensemble-based MTDI gains $+17.5\\%$ ( $67.6\\% \\rightarrow 85.1\\%$ ), which is a huge improvement. When $\\epsilon = 4$ , the results also demonstrate the advantages of StyLess. This shows that StyLess consistently delivers strong transferability when faced with a more robust network that is difficult to attack.", + "bbox": [ + 75, + 390, + 467, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Attacking the Google Cloud Vision API", + "text_level": 1, + "bbox": [ + 76, + 544, + 415, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the Google Cloud Vision API as an example of real-world applications to evaluate the transferability of attacks. This API allows us to use various vision features, such as image labeling and face detection. As a black-box model, regular users like us cannot access its network architecture, training data, or defense mechanism. In this section, we will focus on attacking its image labeling feature.", + "bbox": [ + 75, + 566, + 467, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To utilize the image labeling feature of the API, we need to upload images and obtain the predicted labels. We use 1000 images from ImageNet as the target images, as described in the experimental setup. Due to the fact that the API does not support all 1000 categories in ImageNet, our objective is to mislead the API's original top-1 prediction.", + "bbox": [ + 75, + 674, + 467, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use ResNet50 as the surrogate network to compare the baseline method MTDSI with our method. Experimental results show that MTDSI achieves an attack success rate of $75.6\\%$ on the Google Cloud Vision API. Our approach, MTDSI-SyLess, achieves an attack success rate of $85.2\\%$ , which is a $9.6\\%$ improvement over the baseline. Our results demonstrate that transfer-based black-box attacks pose a severe threat to real-world applications, and StyLess can effectively boost attack transferability.", + "bbox": [ + 75, + 763, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 297, + 651, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this session, we will present four ablation studies: 1) The position of the inserted IN layer; 2) The number of the generated stylized models; 3) The clean losses of stylized models and attack transferability; 4) The most important statistic of style features.", + "bbox": [ + 496, + 323, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Which position to insert the IN layer? As shown in Figure 4, we use features from different layers of the vanilla surrogate networks as the encoder $F_{1}$ for style transfer. We evaluate the attack success rates using two surrogate networks: WRN101 and DN121. After injecting synthesized styles into different network layers, we report the attack success rates on various DNNs, including AlexNet, VGG19, RN50, WRN101, and DN121. We observe a trend that the best attack success rates are usually achieved in the shallow layers of the surrogate networks. For example, when using WRN101 as the surrogate network, injecting the synthesized styles in the layers before layer ten is a good choice. Using intermediate layers such as layers 40 to 80 is also acceptable, but the attack performance may be unstable or even worse when injecting the styles in the last few layers. The results of using DN121 as the surrogate network also indicate that the last few layers are the worst choices, and the shallow layers are the optimal.", + "bbox": [ + 496, + 398, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ea263319dba8a4b7ae047d615b6e042abcdaacace7e706f94f8cc297766e967.jpg", + "image_caption": [ + "Figure 4. Ablation study on using which network layer to synthesize styles. The surrogate networks are WRN101 and DN121." + ], + "image_footnote": [], + "bbox": [ + 501, + 676, + 696, + 815 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08b72aef0a42964ce780b7380181ac3a300bd88b3cf4ceb742349c0424ed1186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 676, + 888, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How many stylized models should be created in each iteration? In Figure 5, we vary the number of stylized models generated in each attack iteration from zero to", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8169", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ten. We conduct experiments by combining StyLess with two baseline attacks: MI and DI. The first figure evaluates the attack success rates on three unsecured networks for our $\\mathrm{MI} + \\mathrm{StyLess}$ attack. In the second figure, we test $\\mathrm{DI} + \\mathrm{StyLess}$ on three secured models. When the number of stylized models is 0, StyLess is not involved, so it is the vanilla baseline attack. As we can see, when the number increases from 0 to 1, the attack success rate starts to grow, which means StyLess starts to work. There is an anomaly when DI begins to combine with StyLess. If the number is less than three, the attack success rate on IncResV2 $_{ens}$ is slightly worse than the baseline (around $25\\%$ ). When the number increases by more than three, the attack success rate becomes higher than the baseline. This may be because IncResV2 $_{ens}$ is very strong, and more synthesized style features need to be injected into the surrogate model. According to the figure, StyLess works quite well when six to ten stylized models are used in an attack iteration.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26c2de9ed8b7a09ab3e7993e4a02aabf49c0486747e0cca07ade759cd9524c1a.jpg", + "image_caption": [ + "Figure 5. Ablation study on the number of stylized models in an attack iteration. The surrogate network is RN50." + ], + "image_footnote": [], + "bbox": [ + 81, + 373, + 274, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95a270d533d13e1510ccc974bb86c4626743fc8fe389ab9642f34e0d39c2373e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 373, + 467, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How network loss affects attack success rate? In Figure 6, we demonstrate how varying strengths of style injection can affect network loss, which in turn impacts attack performance on a surrogate network (in this case, VGG19). The strength of synthesized style features is denoted by the number of stars, with more stars indicating greater style strength that alters the style features of the surrogate model. Generally, injecting synthesized style features should not significantly affect the clean loss, as an increase in network loss typically leads to a decrease in clean accuracy. Overly corrupted stylized surrogate models can also result in bad gradients for attack methods. Therefore, there is an upper bound on clean loss when generating stylized models. The red line in the figure represents the situation of overly corrupting, in which the clean loss exceeds the estimated bound (indicated by a yellow line in the right figure), and the attack success rate drops significantly. This demonstrates the importance of maintaining relatively good clean accuracy when creating stylized networks.", + "bbox": [ + 75, + 566, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Which statistic of style features matters most? In Figure 7, we compare the effects of the mean and variance of style features on StyLess. Here the interpolation fac", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbe6534e33657c4c91dce20b9f2181a5b968269717ac629b442d06b57e96d2dc.jpg", + "image_caption": [ + "Figure 6. Study how synthesized style features affect the clean loss, which in return impacts the attack success rate. The number of stars indicates the degree of change in original style features." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 697, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52dd2e04af075b6d32dd07abb33c26013e3a1709f9def9de45a65c7a6ad60182.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tor is $\\lambda = 0$ . Equation 9 shows that $\\beta$ involves the mean in IN, while $\\gamma$ affects the variance. The attack success rate represents the average success rates of attacks on five DNNs: VGG19, AlexNet, RN50, WRN101, and DN121. The results show that $\\gamma$ plays a more important role than $\\beta$ . Specifically, when RN50 is used as the surrogate network, modifying $\\beta$ alone barely improves the baseline attack, while involving $\\gamma$ alone enhances the attack success rate by around $5\\%$ . A similar observation can be made for VGG19. From the perspective of gradient calculations, this also makes sense. When we backpropagate through an IN layer, we have $\\frac{\\partial}{\\partial x} \\mathrm{IN} = \\sigma$ , which also indicates that the variance matters most for adversarial attacks.", + "bbox": [ + 496, + 297, + 892, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59079058ed5ab83c5bc5bccaab88489a161886b76ab926ba5444b02608e316e3.jpg", + "image_caption": [ + "Figure 7. The effect of different statistics of style." + ], + "image_footnote": [], + "bbox": [ + 524, + 506, + 854, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 619, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we analyze the mechanism of attack transferability in terms of style features. We demonstrate that existing attack methods increasingly use the style features of surrogate models during the iterative optimization, which hampers attack transferability. To address this issue, we propose a novel attack method called StyLess to enhance transferability by reducing reliance on original style features. StyLess uses stylized surrogate models instead of a vanilla surrogate model. Experimental results show that StyLess outperforms existing attacks by a large margin, and can be combined with other attack methods. Notably, StyLess is a different paradigm from previous transferable attack methods, and we hope it will shed light on the interpretation of adversarial attacks in the future.", + "bbox": [ + 496, + 689, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8170", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 2", + "[2] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Revisiting batch normalization for improving corruption robustness. In WACV, 2021. 3", + "[3] Philipp Benz, Chaoning Zhang, and In So Kweon. Batch normalization increases adversarial vulnerability and decreases adversarial transferability: A non-robust feature perspective. In ICCV, 2021. 3", + "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 2", + "[5] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 5", + "[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 2, 5", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5", + "[8] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A. K. Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2", + "[9] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In ICLR, 2017. 2", + "[10] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2", + "[12] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 3, 5", + "[13] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples. In NeurIPS, 2020. 2, 5", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5", + "[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In CVPR, 2022. 2", + "[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5", + "[17] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 2, 3", + "[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 3", + "[20] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 2", + "[21] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 2", + "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2", + "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2", + "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 1, 5", + "[25] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 2, 3", + "[26] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 2", + "[27] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning transferable adversarial examples via ghost networks. In AAAI, 2020. 2", + "[28] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 2017. 2", + "[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 2, 5", + "[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 1, 2, 5, 6", + "[31] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5", + "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2", + "[33] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. Stylized adversarial defense. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2", + "[34] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 1", + "[35] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8171", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115:211-252, 2015. 5", + "[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, 2018. 5", + "[38] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, 2020. 3", + "[39] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM Conference on Computer and Communications Security (CCS), pages 1528-1540, 2016. 2", + "[40] Manli Shu, Zuxuan Wu, Micah Goldblum, and Tom Goldstein. Encoding robustness to image style via adversarial feature perturbations. NeurIPS, 2021. 3", + "[41] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5", + "[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 5", + "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2", + "[44] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, 2019. 5", + "[45] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 2", + "[46] Shuo Wang, Surya Nepal, Alsharif Abuadbba, Carsten Rudolph, and Marthie Grobler. Adversarial detection by latent style transformations. IEEE Transactions on Information Forensics and Security (TIFS), 17:1099-1114, 2022. 2", + "[47] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 2, 5", + "[48] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 2, 5", + "[49] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 2, 5", + "[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 2", + "[51] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 2, 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2, 3", + "[53] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5", + "[54] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018. 2" + ], + "bbox": [ + 501, + 90, + 890, + 205 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8172", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_model.json b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..254eb04010ecb4a4250cc83db8e6a05ade6a9b27 --- /dev/null +++ b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.131, + 0.802, + 0.154 + ], + "angle": 0, + "content": "StyLess: Boosting the Transferability of Adversarial Examples" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.182, + 0.642, + 0.217 + ], + "angle": 0, + "content": "Kaisheng Liang Bin Xiao* The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.22, + 0.645, + 0.234 + ], + "angle": 0, + "content": "{cskliang, csbxiao}@comp.polyu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.604 + ], + "angle": 0, + "content": "Adversarial attacks can mislead deep neural networks (DNNs) by adding imperceptible perturbations to benign examples. The attack transferability enables adversarial examples to attack black-box DNNs with unknown architectures or parameters, which poses threats to many real-world applications. We find that existing transferable attacks do not distinguish between style and content features during optimization, limiting their attack transferability. To improve attack transferability, we propose a novel attack method called style-less perturbation (StyLess). Specifically, instead of using a vanilla network as the surrogate model, we advocate using stylized networks, which encode different style features by perturbing an adaptive instance normalization. Our method can prevent adversarial examples from using non-robust style features and help generate transferable perturbations. Comprehensive experiments show that our method can significantly improve the transferability of adversarial examples. Furthermore, our approach is generic and can outperform state-of-the-art transferable attacks when combined with other attack techniques." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.836 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) [14, 24] are currently effective methods for solving various challenging tasks such as computer vision, and natural language processing. Although DNNs have amazing accuracy, especially for computer vision tasks such as image classification, they are also known to be vulnerable to adversarial examples [12, 43]. Adversarial examples are malicious images obtained by adding imperceptible perturbations to benign images. Notably, the transferability of adversarial examples is an intriguing phenomenon, which refers to the property that the same adversarial example can successfully attack different black-box DNNs [5, 30, 34, 51]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.837, + 0.47, + 0.868 + ], + "angle": 0, + "content": "It has been observed that image style can be decoupled from image content, and style transfer techniques allow us" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.272, + 0.885, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.45, + 0.627, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.541, + 0.608, + 0.552 + ], + "angle": 0, + "content": "Benign Sample" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.45, + 0.755, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.541, + 0.738, + 0.552 + ], + "angle": 0, + "content": "Stylized Output" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.45, + 0.885, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.541, + 0.866, + 0.552 + ], + "angle": 0, + "content": "StyLess Attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.559, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Figure 1. An overview of our StyLess attack. We create stylized model \\(\\bar{F}\\) by injecting synthesized style features into the surrogate model \\((F = F_{2} \\circ F_{1})\\) using an adaptive IN layer. StyLess reduces the use of non-robust style features in the vanilla surrogate model \\(F\\), ultimately improving attack transferability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.824 + ], + "angle": 0, + "content": "to generate stylized images based on arbitrary style images [18]. Image style refers to the unique visual characteristics of an image, including its colors, textures, and lighting. For instance, two photos of the same object taken by different photographers can have very different styles. Robust DNNs should rely more on content features of data than style features. This inspired us to improve attack transferability from the perspective of avoiding non-robust features. We believe that style features of DNNs are non-robust for building transferable attacks. However, existing attacks do not distinguish between the surrogate models' style and content features, which may reduce attack transferability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We propose using stylized surrogate models to control style features, which can significantly improve transferability. We refer to the original surrogate model as the \"vanilla model.\" The proposed stylized model is created by adding an adaptive instance normalization (IN) layer to the vanilla" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.398, + 0.9 + ], + "angle": 0, + "content": "1Our code is available at https://github.com/uhiu/StyLess" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.398, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8163" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "model. By adjusting the parameters of the inserted IN layer, we can easily transform the style features of the surrogate models. To compare the stylized and vanilla surrogate models, we analyzed their network losses during optimization. Surprisingly, we found that the adversarial loss of the vanilla model increased much faster than those of the stylized models, resulting in a widening loss gap. This phenomenon reveals that as the attack iteration progresses, current attack methods only focus on maximizing the loss of the vanilla model, leading to increased use of the style features of the vanilla surrogate model. However, we believe that style features are non-robust for transferable attacks, and relying too much on them may reduce attack transferability. To enhance transferability, we aim to limit the use of non-robust style features and close the loss gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.589 + ], + "angle": 0, + "content": "Based on the above findings, we propose a novel method called StyLess to improve the transferability of adversarial examples. Our method uses multiple synthesized style features to compete with the original style features during the iterative optimization of attack. The process is illustrated in Figure 1. We encode various synthesized style features into a surrogate model via an IN layer to achieve stylized surrogate models. Instead of using only the vanilla surrogate model, we use the gradients of both the stylized surrogate models and the vanilla one to update adversarial examples. The front part of the surrogate model works as a style encoder, and the IN layer simulates synthesized style features. Although we can use a decoder to explicitly generate the final stylized samples, it is unnecessary for the proposed attack method. Experimental results demonstrate that StyLess can enhance the transferability of state-of-the-art adversarial attacks on both unsecured and secured black-box DNNs. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.591, + 0.47, + 0.667 + ], + "angle": 0, + "content": "- We introduce a novel perspective for interpreting attack transferability: the original style features may hinder transferability. We verify that current iterative attacks increasingly use the style features of the surrogate model during the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.667, + 0.47, + 0.742 + ], + "angle": 0, + "content": "- We propose a novel attack called StyLess to enhance transferability by minimizing the use of original style features. To achieve this, we insert an IN layer to create stylized surrogate models and use gradients from both stylized and vanilla models." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.743, + 0.47, + 0.82 + ], + "angle": 0, + "content": "- We conducted comprehensive experiments on various black-box DNNs to demonstrate that StyLess can significantly improve attack transferability. Furthermore, we show that StyLess is a generic approach that can be combined with existing attack techniques." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.591, + 0.47, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.22, + 0.847 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Adversarial Attacks. Adversarial attacks reveal the vulnerability of current DNNs [43]. The classic adversarial attack methods are gradient-based, such as FGSM [12] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "I-FGSM [25]. C&W [4] considers optimizing the distance between adversarial examples and benign samples, and proposed optimization-based attacks. Adversarial attacks can also be performed in the physical world [10, 39]. As for defending against adversarial examples, adversarial training is a popular defense method that uses adversarial examples as extra training data to improve robustness [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.893, + 0.427 + ], + "angle": 0, + "content": "Increasing Attack Transferability. An intriguing property of adversarial attacks is the transferability. Ensemble-based attack [30] uses multiple surrogate networks instead of one network. Ghost networks [27] generates different surrogate networks by perturbing skip connection and dropout layers. Optimization methods, such as MI [5], uses a momentum-based optimization, while VT [47] introduces gradient variance to control the stability of the localized gradients. RAP [35] generates adversarial examples located in a flat loss region. Data augmentation methods, such as DI [51], uses image transformation like resizing and padding, while TI [6] considers translating image pixels. SI [29] calculates gradients with the help of several scaled benign samples. Admix [48] calculates iterative gradients by mixing the benign images with randomly sampled images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.893, + 0.595 + ], + "angle": 0, + "content": "Various network architectures and features exhibit different relationships with adversarial attacks. DNNs' linearity is believed to cause adversarial vulnerability [12], and LinBP [13] skips the nonlinear activation during the backpropagation. SGM [49] uses more gradients through skip connections in residual networks. To better leverage the intermediate layers, one can train auxiliary classifiers based on feature spaces [20, 21], maximize the distance between natural images and their adversarial examples in feature spaces [54], or fine-tune the existing adversarial examples in intermediate layer level by ILA [17, 26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.809 + ], + "angle": 0, + "content": "Style Transfer and Instance Normalization. Style transfer can change the style of an image to match the style of another one [9, 11, 23]. Fast feedforward networks can perform stylization with arbitrary styles in a single forward pass [18, 28]. Interestingly, style transfer has a wide range of applications. AdvCam [8] uses natural styles to hide non-\\(L_{p}\\) restricted perturbations. FSA [52] generates natural-looking adversarial examples by using optimized style changes. Style transfer has also been used to improve network robustness by exploring additional feature information [33]. Latent style transformations can detect adversarial attacks [46]. AMT-GAN [15] proposes an adversarial makeup transfer to protect facial privacy by preserving stronger black-box transferability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The family of instance normalization (IN) including batch normalization [22], layer normalization [1], instance normalization [45], and group normalization [50]. Normalizations are mainly used to reduce the covariate shift, and speed model training. Recently, normalizations have been found to be related to robustness. It has been shown that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "batch normalization makes DNNs use more non-robust but useful features [3, 19]. AdvBN proposed adding an extra batch normalization into network training to increase training loss adversarially, which enables the network to resist various domain shifts [40]. Adjusting batch normalization statistics such as the running mean and variance in the inference phase, which are estimated during training, improves robustness and defense common corruption [2, 38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.471, + 0.41 + ], + "angle": 0, + "content": "Among existing style-based attacks, FSA [52] differentiates style features and content features, which is similar to our method. However, there are three significant differences between FSA and our approach: 1) FSA proposes to hide adversarial perturbations in the optimized style, while we avoid relying on any style. 2) FSA aims at enhancing the natural looking of non-\\(\\ell_p\\) restricted attacks, while we focus on the transferability of \\(\\ell_p\\) restricted adversarial examples. 3) Both FSA and our work are inspired by AdaIN [18], but we use IN layer differently. FSA perturbs the IN layer to search malicious styles and requires a decoder. But we use randomized IN layers to augment attacks and don't need to train a decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.212, + 0.442 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.22, + 0.466 + ], + "angle": 0, + "content": "3.1. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Attack objective. Given a benign image \\( x \\) with label \\( y \\), transfer-based attacks aim to generate an adversarial perturbation based on a white-box surrogate network \\( F \\). The general attack objective can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.549, + 0.469, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\delta} \\mathcal {L} (F (x + \\delta), y) \\quad \\text {s . t .} \\quad \\| \\delta \\| \\leq \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.614 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) denotes the adversarial loss, \\(\\delta\\) is the adversarial perturbation, and \\(\\epsilon\\) is the maximum perturbation size." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.615, + 0.469, + 0.645 + ], + "angle": 0, + "content": "A popular framework to solve the above problem is iterative fast gradient sign method (I-FGSM) [12, 25]:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.657, + 0.469, + 0.677 + ], + "angle": 0, + "content": "\\[\nx _ {a d v} ^ {t + 1} = x _ {a d v} ^ {t} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {a d v} ^ {t}\\right), y\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.719 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the learning rate, and a clip function will be used on \\(x_{adv}^{t + 1}\\) to ensure \\(\\| x_{adv}^{t + 1} - x\\| \\leq \\epsilon\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Attacker capability. We follow the same setting in previous work that attackers have a surrogate model and some test samples, but cannot access target models, and don't know network architectures, training data, or defense strategies. It should be noted that our method doesn't require any additional datasets. Our approach involves style features, which can be extracted from an arbitrary image or synthesized without any style image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Transferable attacks as black-box attacks. Transferable attacks use the surrogate model \\( F \\) to create adversarial examples that can fool unseen target models. In this way, these attacks can be viewed as black-box attacks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.622, + 0.106 + ], + "angle": 0, + "content": "3.2. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Existing transferable attacks often rely on the gradient of the adversarial loss function \\(\\mathcal{L}\\) (Equation 2) without considering the impact of different components of \\(\\mathcal{L}\\). However, these approaches have limitations because transferable attacks should minimize the use of non-robust features of the surrogate model. Interestingly, for image classification task, style features of images are typically less robust than content features. Based on this observation, we propose to enhance attack transferability by explicitly reducing the use of style features of the surrogate model within the loss \\(\\mathcal{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Our key idea is simulating various surrogate models without the style features of the given vanilla surrogate model. We discovered that inserting an IN layer into the vanilla surrogate model enables us to create new surrogate models that we refer to as stylized surrogate models. Sometimes we omit the word \"surrogate.\" Stylized models can explicitly manipulate style features without compromising model accuracy. Recall that our goal is to construct adversarial examples that can mislead unseen target models, which should include these stylized models. However, existing methods, such as MI and I, only focus on maximizing the loss of the vanilla model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Figure 2 indicates that MI and I(-FGSM) have limited attack transferability on stylized models since the vanilla model's adversarial loss increases much faster than stylized models', resulting in a widening loss gap. In the following sections, we will demonstrate how our method addresses this issue by maximizing the loss of both the stylized and vanilla models, which significantly improves transferability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.74, + 0.577 + ], + "angle": 0, + "content": "3.3. Stylized Surrogate Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.891, + 0.675 + ], + "angle": 0, + "content": "This section will give the definition of our stylized surrogate models, which encode various style features by inserting an IN layer. Then we will analyze the stylized loss gap \\((\\Delta \\mathcal{L})\\) between the vanilla and stylized surrogate models. Specifically, we will illustrate the increasing \\(\\Delta \\mathcal{L}\\) limits transferability and our idea to decrease \\(\\Delta \\mathcal{L}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.802, + 0.697 + ], + "angle": 0, + "content": "3.3.1 Encoding Styles by Stylized Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.891, + 0.736 + ], + "angle": 0, + "content": "Given a classifier \\( F = F_{2} \\circ F_{1} \\) as the surrogate model, we define a stylized surrogate model as" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.746, + 0.891, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\bar {F} _ {x _ {s}} = F _ {2} \\circ \\mathrm {I N} _ {x _ {s}} \\circ F _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.774, + 0.89, + 0.804 + ], + "angle": 0, + "content": "where \\(x_{s}\\) is a style input, \\(\\mathrm{IN}_{x_s}\\) is an IN layer instantiated by \\(x_{s}\\). In general, an IN layer is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.813, + 0.891, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I N} (x; \\mu , \\sigma) = \\sigma \\cdot \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) are the network parameters of layer IN, and \\(\\mu(x)\\) and \\(\\sigma(x)\\) are the mean and variance of input \\(x\\). According to adaptive instance normalization (AdaIN) [18], to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8165" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.272, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.09, + 0.452, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.223, + 0.388, + 0.235 + ], + "angle": 0, + "content": "(a) Baseline attack MI vs ours \\(\\mathrm{MI + StyLess}\\)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.237, + 0.273, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.237, + 0.451, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.369, + 0.402, + 0.382 + ], + "angle": 0, + "content": "(b) Baseline attack I(-FGSM) vs ours I+StyLess" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.471, + 0.464 + ], + "angle": 0, + "content": "Figure 2. Illustration of the loss gap between vanilla network and stylized network, with RN50 as surrogate model. The greater the loss, the better the attack performance. The widening loss gap in the baseline means that attack performance on stylized models is lagging behind. The proposed StyLess can close the gap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.516 + ], + "angle": 0, + "content": "stylize an input \\( x \\) with a given style input \\( x_{s} \\), we only need to instantiate the IN as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.53, + 0.47, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\left. \\operatorname {I N} _ {x _ {s}} (x) = \\operatorname {I N} (x) \\right| _ {\\mu = \\mu \\left(x _ {s}\\right), \\sigma = \\sigma \\left(x _ {s}\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.469, + 0.653 + ], + "angle": 0, + "content": "Our stylized model \\(\\bar{F}_{x_s}\\) has encoded style features. Based on AdaIN, the \\(F_{1}\\) from the classifier \\(F = F_{2} \\circ F_{1}\\) works as an encoder for style transfer, and given a style input \\(x_{s}\\), we can get a stylized image as \\(\\bar{x} = D \\circ \\mathrm{IN}_{x_s} \\circ F_1(x)\\) where \\(D\\) denotes a decoder. Thus, \\(\\mathrm{IN}_{x_s} \\circ F_1(x)\\) has encoded the style features of \\(x_{s}\\) to \\(\\bar{F}_{x_s}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.411, + 0.68 + ], + "angle": 0, + "content": "3.3.2 Stylized Loss Gap Limits transferability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.721 + ], + "angle": 0, + "content": "To validate the performance of adversarial attacks on these stylized surrogate models, we define a stylized loss gap as" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.733, + 0.47, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathcal {L} = \\mathbb {E} _ {x _ {s} \\in \\mathcal {D}} [ \\mathcal {L} (F (x), y) - \\mathcal {L} (\\bar {F} _ {x _ {s}} (x), y) ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.469, + 0.794 + ], + "angle": 0, + "content": "where \\(F\\) and \\(\\bar{F}_{x_s}\\) are vanilla and stylized model, respectively; \\(x_{s}\\) is style input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.855 + ], + "angle": 0, + "content": "An increasing loss gap \\(\\Delta \\mathcal{L}\\) can be observed in Figure 2 which limits attack transferability, as explained below. Hypothetically, let's say we can decouple style-dependent loss from content-dependent loss in \\(\\mathcal{L}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.868, + 0.469, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} ^ {c} + \\mathcal {L} _ {x} ^ {s}, \\\\ \\bar {\\bar {z}} = z ^ {c} - z ^ {s} \\end{array} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.89, + 0.326, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} ^ {c} + \\mathcal {L} _ {x _ {s}} ^ {s},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.152 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is the vanilla loss, and \\(\\bar{\\mathcal{L}}\\) is the stylized one; \\(\\mathcal{L}^c\\) is the common content-dependent loss; \\(\\mathcal{L}_x^s\\) is input sample \\(x\\)-specific style-dependent loss, while \\(\\mathcal{L}_{x_s}^s\\) is style sample \\(x_{s}\\)-specific style-dependent loss. In this case, \\(\\Delta \\mathcal{L} = \\mathcal{L}_x^s -\\mathcal{L}_{x_s}^s\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.152, + 0.892, + 0.257 + ], + "angle": 0, + "content": "In other words, the non-robust features are related to \\(\\mathcal{L}_x^s\\) and \\(\\mathcal{L}_{x_s}^s\\), while \\(\\mathcal{L}_c\\) is supposed to be shared by other unseen DNNs. MI and I-(FGSM) have increasing gaps in Figure 2 means that attackers gradually focus on optimizing the \\(\\mathcal{L}_x^s\\) part, which only belonged to the vanilla surrogate model \\(F\\). Therefore, the loss gap limits the transferability of adversarial examples in unseen stylized models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.258, + 0.893, + 0.438 + ], + "angle": 0, + "content": "To decrease \\(\\Delta \\mathcal{L}\\) and boost transferability, we consider involving \\(\\mathcal{L}_{x_s}^s\\) as a competitor of \\(\\mathcal{L}_x^s\\) in optimization process to suppress the growth of \\(\\mathcal{L}_x^s\\). In general, we can assume that all these losses in Equation 7 are non-negative, and they satisfy: \\(\\mathcal{L}^c\\gg \\mathcal{L}_x^s\\), \\(\\mathcal{L}^c\\gg \\mathcal{L}_{x_s}^s\\). Also, there is a upper bound \\(B\\) for these losses: \\(\\mathcal{L} < B\\), \\(\\bar{\\mathcal{L}} < B\\), as the adversarial perturbations are required to smaller than a given \\(\\epsilon\\). If we only maximize the vanilla loss \\(\\mathcal{L}\\), both \\(\\mathcal{L}^c\\) and \\(\\mathcal{L}_x^s\\) are likely to be increased. To this end, we propose to maximize \\(\\mathbb{E}_{x_s\\in \\mathcal{D}}\\mathcal{L}_{x_s}^s +\\mathcal{L}_x^s +\\mathcal{L}^c\\), which involves multiple style-dependent losses \\(\\mathcal{L}_{x_s}^s (x_s\\in \\mathcal{D})\\) to compete with \\(\\mathcal{L}_x^s\\) and leads to a decrease of \\(\\Delta \\mathcal{L}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.448, + 0.877, + 0.464 + ], + "angle": 0, + "content": "3.4. Proposed Style-Less Perturbations (StyLess)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Based on the above analysis, we propose style-less perturbations (StyLess) method to increase attack transferability by optimizing stylized loss and vanilla loss together:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.526, + 0.892, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\delta} \\mathbb {E} _ {x _ {s} \\in \\mathcal {D}} \\mathcal {L} \\left(\\bar {F} _ {x _ {s}} (x + \\delta), y\\right) + \\mathcal {L} \\left(F (x + \\delta), y\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.892, + 0.62 + ], + "angle": 0, + "content": "The key to generate multiple stylized models \\(\\bar{F}_{x_s}\\) is synthesizing style statistics \\(\\mu, \\sigma\\) for Equation 5 to obtain parameterized IN layers. We propose using scaling and interpolation to simulate multiple style features, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.629, + 0.891, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mu = \\beta (\\lambda \\mu_ {x} + (1 - \\lambda) \\mu_ {s}), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.65, + 0.783, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\gamma (\\lambda \\sigma_ {x} + (1 - \\lambda) \\sigma_ {s}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.892, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mu_x, \\sigma_x\\) is the mean and variance of \\(F_1(x)\\), relating to the benign content input \\(x\\), while \\(\\mu_s, \\sigma_s\\) are for style input \\(x_s\\) similarly. \\(x_s\\) is an arbitrary image. \\(\\lambda\\) is a scalar that controls the interpolation of two styles. \\(\\beta\\) and \\(\\gamma\\) are \\(c\\)-dimensional vectors that scale the synthesized style, where \\(c\\) refers to the number of channels in the style feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We summarize the proposed StyLess attack in Algorithm 1. To obtain the parameters of the IN layer without training, we select a random image as \\( x_{s} \\), and generate a pair of \\( \\mu \\) and \\( \\sigma \\) using Equation 9. With \\( \\mu \\) and \\( \\sigma \\), we can create a stylized model \\( \\bar{F}_{x_s}(x) \\) using Equation 3 and 5. \\( \\bar{F}_{x_s}(x) \\) need to be equal to \\( F(x) \\) or label \\( y \\), otherwise we will regenerate \\( \\mu \\) and \\( \\sigma \\) by altering \\( \\beta, \\gamma \\) and \\( \\lambda \\). In each iteration, we use \\( N \\) stylized models to update adversarial examples based on the proposed objective function in Equation 8." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8166" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.09, + 0.47, + 0.401 + ], + "angle": 0, + "content": "Algorithm 1 Style-Less Perturbations (StyLess) Algorithm \nInput: Surrogate model \\(F\\), benign example \\(x\\), iteration number \\(T\\), maximum perturbation \\(\\epsilon\\), data augmentation \\(\\phi(\\cdot)\\), decay factor \\(\\eta\\). Scale factors \\(\\beta, \\gamma\\) and interpolation factor \\(\\lambda\\). The number of stylized models \\(N\\). \nOutput: An adversarial example \\(x_{adv}\\) \n1: \\(x_{adv} = x\\), \\(g_0 = 0\\), \\(\\alpha = \\epsilon/2\\). \n2: for \\(t = 0 \\rightarrow T - 1\\) do: \n3: Augment input \\(x_{adv} = \\phi(x_{adv})\\) \n4: Obtain gradient \\(\\tilde{g}_{t+1}\\) with respect to \\(x_{adv}\\) using \\(F\\) \n5: repeat \n6: Synthesize a style statistic by Equation 9 \n7: Obtain a stylized model \\(\\bar{F}_{x_s}\\) by Equation 3 \n8: Get gradient \\(\\tilde{g}\\) with respect to \\(x_{adv}\\) using \\(\\bar{F}_{x_s}\\) \n9: Update \\(\\tilde{g}_{t+1} = \\tilde{g}_{t+1} + \\tilde{g}\\) \n10: until obtain \\(N\\) stylized models \n11: Calculate momentum \\(g_{t+1} = \\eta \\cdot g_t + \\tilde{g}_{t+1} / \\| \\tilde{g}_{t+1} \\|_1\\) \n12: Update example \\(x_{adv} = x_{adv} + \\alpha \\cdot \\text{sign}(g_{t+1})\\) \n13: end for \n14: return \\(x_{adv}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.428, + 0.21, + 0.444 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.267, + 0.469 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.469, + 0.567 + ], + "angle": 0, + "content": "Dataset. We use ImageNet [36] for experiments. Specifically, we use 1000 images which are randomly selected from ImageNet validation set by Wang et al. [47]. Similar dataset settings have been widely used in previous work [5, 13, 29, 47, 49, 51]. These images include all categories; almost all are correctly classified by the target DNNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Models. We evaluate the generated adversarial examples on different black-box DNNs, including both unsecured and secured models. Unsecured models are trained on ImageNet using traditional methods, while secured models are based on adversarial training. The unsecured models include VGG19 [41], AlexNet [24], ResNet50 (RN50) [14], WideResNet101 (WRN101) [53], DenseNet121 (DN121) [16], InceptionV3 (IncV3) [42], MnasNet [44], MobileNetV2 (MNv2) [37], ShuffleNetV2 (SNv2) [31] and ViT [7]. Their pre-training parameters are obtained from PyTorch official. The secured DNNs are IncV3ens3 (ensemble of 3 InceptionV3 networks), IncV3ens4 (ensemble of 4 InceptionV3 networks) and IncResV2ens3 (ensemble of 3 IncResV2 networks). These models were adversarily trained and widely used in previous work [5,30,47,49]. As for the surrogate models, we use VGG19 [41], RN50 [14], WRN101 [53] and DN121 [16]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We use I-FGSM [12] as the initial baseline. Unless otherwise specified, the attacks are untargeted and \\( l_{\\infty} \\)-restricted. The maximum perturbation size is set to \\( \\epsilon = 16 / 255 \\). We compare StyLess primarily with six transferable attacks: MI [5], DI [51], TI [6], SI [29]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Admix (AI) [48], and an ensemble-based approach [30]. We set the optimization step size to \\(\\alpha = \\epsilon /2\\), and the number of iterations to \\(T = 50\\). The momentum decay in MI is \\(\\mu = 1\\). For DI, SI and AI, we follow the official settings described in the corresponding papers. For StyLess, we simulate 10 stylized models in each iteration, denoted by \\(N = 10\\). To generate a stylized model for a given \\(x\\), we randomly sample \\(\\lambda\\) from \\([0,0.2]\\), and \\(\\beta, \\gamma\\) from \\([0,2]\\), and ensure that \\(\\bar{F}_{x_s}(x)\\) is equal to \\(F(x)\\) or the real label. The IN layer is inserted after the first bottleneck block for RN50 and WRN101, and after the first dense block for DN121." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.268, + 0.759, + 0.284 + ], + "angle": 0, + "content": "4.2. Attacking Unsecured Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.503 + ], + "angle": 0, + "content": "We compare StyLess with other attacks on various unsecured DNNs using three surrogate models. Table 1 shows that StyLess is a powerful and generic method that can be combined with existing attack methods to further improve attack transferability. Specifically, we compare StyLess with I, MI, DI, TI, SI and Admix (AI). For the most challenging case in the table, attacking the black-box IncV3 (let's take \\(\\mathrm{RN50} \\Rightarrow \\mathrm{IncV3}\\) attack as an example), StyLess significantly improves the attack success rate of baseline attacks (I and MI): \\(46.2\\% \\rightarrow 68.3\\%\\) (I), \\(59.2\\% \\rightarrow 78.9\\%\\) (MI). StyLess also demonstrates its capabilities when using other DNNs as the surrogate network. For instance, DN121 \\(\\Rightarrow\\) SNv2 attack, StyLess significantly improves the baseline: \\(69.3\\% \\rightarrow 91.4\\%\\) (I), \\(77.4\\% \\rightarrow 95.1\\%\\) (MI)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.777 + ], + "angle": 0, + "content": "StyLess can be combined with other attack techniques. Previous work has shown that combining various attack methods results in powerful and transferable attacks. StyLess can be integrated with existing combination-based attacks to enhance attack transferability. We report the results of four combinations of existing attacks: MDI, MTDI, MTDSI, and MTDAI. StyLess further enhances these four attack methods' attack success rate by \\(+8.0\\%\\), \\(+8.2\\%\\), \\(+3.0\\%\\), and \\(+3.3\\%\\) (in the case of \\(\\mathrm{RN}50 \\Rightarrow \\mathrm{IncV}3\\) attack). We evaluate attack performance using various surrogate models, including RN50, WRN101 and DN121. For instance, when combining MTDAI with StyLess, our method enhances the transferability of MTDAI: WRN101 \\(\\Rightarrow\\) IncV3 attack: \\(+4.7\\%\\), WRN101 \\(\\Rightarrow\\) MNv2 attack: \\(+1.5\\%\\), WRN101 \\(\\Rightarrow\\) SNv2 attack: \\(+3.4\\%\\), DN121 \\(\\Rightarrow\\) IncV3 attack: \\(+4.3\\%\\), DN121 \\(\\Rightarrow\\) MNv2 attack: \\(+0.6\\%\\), DN121 \\(\\Rightarrow\\) SNv2 attack: \\(+3.8\\%\\). The results show that StyLess is an efficient and generic approach for improving attack transferability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.74, + 0.803 + ], + "angle": 0, + "content": "4.3. Attacking Secured Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We evaluate StyLess on three widely-used secured models: IncV3ens3, IncV3ens4 and IncResV2ens, as shown in Table 2. We present the results of I, MI, MDI, MTDI, and MTDSI on the three secured models using two surrogate models: RN50 and WRN101. These secured networks are more robust than the unsecured DNNs we men" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8167" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.09, + 0.669, + 0.104 + ], + "angle": 0, + "content": "Table 1. Attacking unsecured black-box models with StyLess." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.108, + 0.875, + 0.389 + ], + "angle": 0, + "content": "
SourceAttackVGG19RN50WRN101DN121IncV3MNv2SNv2
RN50I / +Ours72.8 / 88.8100 / 10080.8 / 97.183.0 / 97.846.2 / 68.377.1 / 91.960.6 / 77.8
MI / +Ours82.9 / 94.1100 / 10083.9 / 97.287.5 / 98.759.2 / 78.983.8 / 93.272.3 / 83.5
MDI / +Ours97.5 / 99.2100 / 10098.2 / 99.899.4 / 10089.5 / 97.598.1 / 99.988.5 / 96.9
MTDI / +Ours98.6 / 99.7100 / 10099.2 / 10099.8 / 10090.2 / 98.498.7 / 10090.8 / 98.1
MTDSI / +Ours98.6 / 99.2100 / 10099.5 / 10099.8 / 10096.2 / 99.298.7 / 99.996.2 / 98.2
MTDAI / +Ours99.3 / 99.6100 / 10099.8 / 10099.9 / 10095.5 / 98.899.7 / 10095.7 / 99.0
WRN101I / +Ours64.8 / 79.688.6 / 98.7100 / 10077.7 / 94.243.2 / 66.668.1 / 84.758.1 / 73.2
MI / +Ours75.6 / 86.489.8 / 98.7100 / 10083.9 / 96.457.6 / 73.776.1 / 87.070.2 / 81.2
MDI / +Ours92.2 / 98.599.0 / 99.9100 / 10097.1 / 99.986.3 / 96.393.6 / 99.086.8 / 95.8
MTDI / +Ours92.5 / 98.699.2 / 100100 / 10097.9 / 99.689.2 / 97.995.4 / 99.388.4 / 97.0
MTDSI / +Ours95.9 / 98.999.6 / 100100 / 10099.7 / 10095.8 / 99.597.7 / 99.994.5 / 98.4
MTDAI / +Ours96.5 / 99.299.9 / 99.9100 / 10099.4 / 10093.7 / 98.498.1 / 99.694.6 / 98.0
DN121I / +Ours79.7 / 97.787.0 / 99.474.7 / 98.0100 / 10055.7 / 88.781.3 / 97.069.3 / 91.4
MI / +Ours86.9 / 99.189.4 / 99.778.2 / 98.6100 / 10066.0 / 95.786.5 / 98.777.4 / 95.1
MDI / +Ours96.7 / 99.998.9 / 10095.4 / 99.8100 / 10090.5 / 99.097.5 / 10090.9 / 98.6
MTDI / +Ours97.5 / 99.999.1 / 10095.9 / 99.8100 / 10092.7 / 99.496.9 / 10092.2 / 98.6
MTDSI / +Ours97.4 / 99.999.3 / 10098.0 / 99.8100 / 10096.5 / 99.898.9 / 10095.1 / 99.1
MTDAI / +Ours99.2 / 99.999.8 / 10098.1 / 99.9100 / 10095.1 / 99.499.4 / 10095.2 / 99.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.41, + 0.468, + 0.438 + ], + "angle": 0, + "content": "Table 2. Attacking three secured black-box models with StyLess. The surrogate model is RN50 or WRN101." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.445, + 0.462, + 0.655 + ], + "angle": 0, + "content": "
AttackRN50⇒
IncV3ens3IncV3ens4IncResV2ens
I / +Ours21.6 / 34.618.9 / 32.014.1 / 21.5
MI/ +Ours31.5 / 47.129.3 / 42.420.8 / 31.0
MDI / +Ours59.6 / 78.153.5 / 69.838.3 / 57.3
MTDI / +Ours69.2 / 89.663.8 / 81.854.6 / 72.2
MTDSI / +Ours88.0 / 93.184.7 / 91.377.8 / 84.5
WRN101⇒
I / +Ours23.4 / 40.320.9 / 34.615.5 / 25.6
MI/ +Ours35.0 / 51.230.4 / 46.624.7 / 35.9
MDI / +Ours64.0 / 81.558.8 / 76.246.9 / 66.4
MTDI / +Ours75.5 / 91.470.5 / 87.463.6 / 81.0
MTDSI / +Ours91.6 / 97.588.6 / 96.183.0 / 92.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.835 + ], + "angle": 0, + "content": "tioned above. Taking \\(\\mathrm{RN50} \\Rightarrow \\mathrm{IncV3}_{\\mathrm{ens3}}\\) attack as an example, I, MI, MDI, and MTDI achieve attack success rates of only \\(21.6\\%\\), \\(31.5\\%\\), \\(59.6\\%\\), and \\(69.2\\%\\), respectively. StyLess effectively improves the performance of these attacks. Notably, MTDSI has been improved by our method: \\(91.6\\% \\rightarrow 97.5\\%\\) in WRN101 \\(\\Rightarrow\\) IncV3ens3 attack; \\(88.6\\% \\rightarrow\\) \\(96.1\\%\\) in WRN101 \\(\\Rightarrow\\) IncV3ens4 attack; \\(83.0\\% \\rightarrow 92.5\\%\\) in WRN101 \\(\\Rightarrow\\) IncResV2ens attack. StyLess demonstrates its great power to enhance transferability in breaking these challenging secured DNNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Figure 3 shows a comparison of StyLess and LinBP. We consider LinBP and its combination with existing attacks such as ILA, SGM, and MDI on various target DNNs. Our method exhibits the best attack transferability." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.415, + 0.877, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.594, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Figure 3. Comparing with LinBP on different black-box models and defenses. The surrogate model is RN50." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.857, + 0.65 + ], + "angle": 0, + "content": "4.4. Combining with Ensemble-Based Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Generating adversarial examples based on multiple surrogate networks simultaneously can improve attack performance in practice [30]. This ensemble-based method has been shown to be a powerful attack, and StyLess can further improve it. We use an ensemble of RN50, WRN101, and DN121 as the integrated surrogate model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We compared our StyLess with ensemble-based MI and MTDI in Table 3. The considerable strength of the ensemble-based method can be seen when comparing with the results in Table 1 and 2. For example, in RN50 \\(\\Rightarrow\\) IncV3 attack, the baseline attack MI only gets \\(59.2\\%\\), while it gets \\(81.0\\%\\) attack success rate in the case of RN50+WRN101+DN121 \\(\\Rightarrow\\) IncV3 attack. Noted that we generally use \\(\\epsilon = 16\\), which is a standard setting. As we can see, when \\(\\epsilon = 16\\), the ensemble-based MTDI achieves an average attack success rate of less than \\(90\\%\\) on the three" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8168" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.09, + 0.655, + 0.104 + ], + "angle": 0, + "content": "Table 3. Combining with ensemble-based attack method." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.108, + 0.88, + 0.278 + ], + "angle": 0, + "content": "
εAttackRN50+WRN101+DN121⇒
VGG19RN50WRN101DN121IncV3MNv2SNv2IncV3ens3IncV3ens4IncResV2ens
4MI56.896.494.897.235.762.348.318.415.89.4
MTDI79.297.996.698.867.584.571.636.532.122.3
+StyLess90.499.599.499.983.294.385.450.042.929.1
8MI79.299.698.799.458.383.368.031.426.818.2
MTDI95.299.999.599.989.597.590.467.661.846.9
+StyLess99.110010010097.599.697.285.177.466.1
16MI92.310010010081.093.285.053.046.335.0
MTDI99.210010010097.510097.992.786.480.7
+StyLess10010010010099.810010098.597.594.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.468, + 0.389 + ], + "angle": 0, + "content": "secured networks: IncV3ens3, IncV3ens4, and IncResV2ens. StyLess can further boost the transferability of ensemble-based MTDI: \\(92.7\\% \\rightarrow 98.5\\%\\) for IncV3ens3, \\(86.4\\% \\rightarrow 97.5\\%\\) for IncV3ens4, \\(80.7\\% \\rightarrow 94.3\\%\\) for IncResV2ens. These results show that StyLess is a different type of attack, and can work perfectly with the ensemble-based method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.468, + 0.541 + ], + "angle": 0, + "content": "We also report the experimental results with different \\(\\epsilon\\). We use \\(\\epsilon = 4\\) or 8 to increase the difficulty to attack. For instance, in the case of ensemble-based \\(\\mathrm{MTDI} \\Rightarrow \\mathrm{IncV3}_{\\mathrm{ens3}}\\) attack, attack success rate drops from \\(92.7\\%\\) to \\(67.6\\%\\) when \\(\\epsilon = 8\\) instead of \\(\\epsilon = 16\\), and StyLess can help ensemble-based MTDI gains \\(+17.5\\%\\) (\\(67.6\\% \\rightarrow 85.1\\%\\)), which is a huge improvement. When \\(\\epsilon = 4\\), the results also demonstrate the advantages of StyLess. This shows that StyLess consistently delivers strong transferability when faced with a more robust network that is difficult to attack." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.416, + 0.561 + ], + "angle": 0, + "content": "4.5. Attacking the Google Cloud Vision API" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.674 + ], + "angle": 0, + "content": "We use the Google Cloud Vision API as an example of real-world applications to evaluate the transferability of attacks. This API allows us to use various vision features, such as image labeling and face detection. As a black-box model, regular users like us cannot access its network architecture, training data, or defense mechanism. In this section, we will focus on attacking its image labeling feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.764 + ], + "angle": 0, + "content": "To utilize the image labeling feature of the API, we need to upload images and obtain the predicted labels. We use 1000 images from ImageNet as the target images, as described in the experimental setup. Due to the fact that the API does not support all 1000 categories in ImageNet, our objective is to mislead the API's original top-1 prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.902 + ], + "angle": 0, + "content": "We use ResNet50 as the surrogate network to compare the baseline method MTDSI with our method. Experimental results show that MTDSI achieves an attack success rate of \\(75.6\\%\\) on the Google Cloud Vision API. Our approach, MTDSI-SyLess, achieves an attack success rate of \\(85.2\\%\\), which is a \\(9.6\\%\\) improvement over the baseline. Our results demonstrate that transfer-based black-box attacks pose a severe threat to real-world applications, and StyLess can effectively boost attack transferability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.653, + 0.315 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.399 + ], + "angle": 0, + "content": "In this session, we will present four ablation studies: 1) The position of the inserted IN layer; 2) The number of the generated stylized models; 3) The clean losses of stylized models and attack transferability; 4) The most important statistic of style features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Which position to insert the IN layer? As shown in Figure 4, we use features from different layers of the vanilla surrogate networks as the encoder \\( F_{1} \\) for style transfer. We evaluate the attack success rates using two surrogate networks: WRN101 and DN121. After injecting synthesized styles into different network layers, we report the attack success rates on various DNNs, including AlexNet, VGG19, RN50, WRN101, and DN121. We observe a trend that the best attack success rates are usually achieved in the shallow layers of the surrogate networks. For example, when using WRN101 as the surrogate network, injecting the synthesized styles in the layers before layer ten is a good choice. Using intermediate layers such as layers 40 to 80 is also acceptable, but the attack performance may be unstable or even worse when injecting the styles in the last few layers. The results of using DN121 as the surrogate network also indicate that the last few layers are the worst choices, and the shallow layers are the optimal." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.678, + 0.697, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.678, + 0.889, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.815, + 0.892, + 0.843 + ], + "angle": 0, + "content": "Figure 4. Ablation study on using which network layer to synthesize styles. The surrogate networks are WRN101 and DN121." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "How many stylized models should be created in each iteration? In Figure 5, we vary the number of stylized models generated in each attack iteration from zero to" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "ten. We conduct experiments by combining StyLess with two baseline attacks: MI and DI. The first figure evaluates the attack success rates on three unsecured networks for our \\(\\mathrm{MI} + \\mathrm{StyLess}\\) attack. In the second figure, we test \\(\\mathrm{DI} + \\mathrm{StyLess}\\) on three secured models. When the number of stylized models is 0, StyLess is not involved, so it is the vanilla baseline attack. As we can see, when the number increases from 0 to 1, the attack success rate starts to grow, which means StyLess starts to work. There is an anomaly when DI begins to combine with StyLess. If the number is less than three, the attack success rate on IncResV2\\(_{ens}\\) is slightly worse than the baseline (around \\(25\\%\\)). When the number increases by more than three, the attack success rate becomes higher than the baseline. This may be because IncResV2\\(_{ens}\\) is very strong, and more synthesized style features need to be injected into the surrogate model. According to the figure, StyLess works quite well when six to ten stylized models are used in an attack iteration." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.374, + 0.275, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.374, + 0.468, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Figure 5. Ablation study on the number of stylized models in an attack iteration. The surrogate network is RN50." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.854 + ], + "angle": 0, + "content": "How network loss affects attack success rate? In Figure 6, we demonstrate how varying strengths of style injection can affect network loss, which in turn impacts attack performance on a surrogate network (in this case, VGG19). The strength of synthesized style features is denoted by the number of stars, with more stars indicating greater style strength that alters the style features of the surrogate model. Generally, injecting synthesized style features should not significantly affect the clean loss, as an increase in network loss typically leads to a decrease in clean accuracy. Overly corrupted stylized surrogate models can also result in bad gradients for attack methods. Therefore, there is an upper bound on clean loss when generating stylized models. The red line in the figure represents the situation of overly corrupting, in which the clean loss exceeds the estimated bound (indicated by a yellow line in the right figure), and the attack success rate drops significantly. This demonstrates the importance of maintaining relatively good clean accuracy when creating stylized networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Which statistic of style features matters most? In Figure 7, we compare the effects of the mean and variance of style features on StyLess. Here the interpolation fac" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.699, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.891, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.243, + 0.893, + 0.286 + ], + "angle": 0, + "content": "Figure 6. Study how synthesized style features affect the clean loss, which in return impacts the attack success rate. The number of stars indicates the degree of change in original style features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.893, + 0.495 + ], + "angle": 0, + "content": "tor is \\(\\lambda = 0\\). Equation 9 shows that \\(\\beta\\) involves the mean in IN, while \\(\\gamma\\) affects the variance. The attack success rate represents the average success rates of attacks on five DNNs: VGG19, AlexNet, RN50, WRN101, and DN121. The results show that \\(\\gamma\\) plays a more important role than \\(\\beta\\). Specifically, when RN50 is used as the surrogate network, modifying \\(\\beta\\) alone barely improves the baseline attack, while involving \\(\\gamma\\) alone enhances the attack success rate by around \\(5\\%\\). A similar observation can be made for VGG19. From the perspective of gradient calculations, this also makes sense. When we backpropagate through an IN layer, we have \\(\\frac{\\partial}{\\partial x} \\mathrm{IN} = \\sigma\\), which also indicates that the variance matters most for adversarial attacks." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.507, + 0.856, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.64, + 0.844, + 0.654 + ], + "angle": 0, + "content": "Figure 7. The effect of different statistics of style." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.62, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.9 + ], + "angle": 0, + "content": "In this work, we analyze the mechanism of attack transferability in terms of style features. We demonstrate that existing attack methods increasingly use the style features of surrogate models during the iterative optimization, which hampers attack transferability. To address this issue, we propose a novel attack method called StyLess to enhance transferability by reducing reliance on original style features. StyLess uses stylized surrogate models instead of a vanilla surrogate model. Experimental results show that StyLess outperforms existing attacks by a large margin, and can be combined with other attack methods. Notably, StyLess is a different paradigm from previous transferable attack methods, and we hope it will shed light on the interpretation of adversarial attacks in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8170" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[2] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Revisiting batch normalization for improving corruption robustness. In WACV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[3] Philipp Benz, Chaoning Zhang, and In So Kweon. Batch normalization increases adversarial vulnerability and decreases adversarial transferability: A non-robust feature perspective. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.469, + 0.283 + ], + "angle": 0, + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.469, + 0.324 + ], + "angle": 0, + "content": "[5] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[8] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A. K. Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[9] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.523, + 0.469, + 0.578 + ], + "angle": 0, + "content": "[10] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[12] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[13] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples. In NeurIPS, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[17] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[20] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[21] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[25] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[26] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[27] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning transferable adversarial examples via ghost networks. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[28] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[31] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[33] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. Stylized adversarial defense. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[34] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[35] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8171" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115:211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[38] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[39] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM Conference on Computer and Communications Security (CCS), pages 1528-1540, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[40] Manli Shu, Zuxuan Wu, Micah Goldblum, and Tom Goldstein. Encoding robustness to image style via adversarial feature perturbations. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.472, + 0.411 + ], + "angle": 0, + "content": "[41] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.579 + ], + "angle": 0, + "content": "[44] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.472, + 0.621 + ], + "angle": 0, + "content": "[45] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.472, + 0.677 + ], + "angle": 0, + "content": "[46] Shuo Wang, Surya Nepal, Alsharif Abuadbba, Carsten Rudolph, and Marthie Grobler. Adversarial detection by latent style transformations. IEEE Transactions on Information Forensics and Security (TIFS), 17:1099-1114, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[47] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.472, + 0.76 + ], + "angle": 0, + "content": "[48] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[49] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[51] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[52] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[53] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[54] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8172" + } + ] +] \ No newline at end of file diff --git a/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_origin.pdf b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..338e1de4f116773f250109e1fa78b909f045853e --- /dev/null +++ b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/6f1fccc6-efbf-4f64-b140-75911087cba7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aac24e64cc039d7441816c14409716dd207d43792cf54d3850eb63daef4786db +size 3018467 diff --git a/2023/StyLess_ Boosting the Transferability of Adversarial Examples/full.md b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/full.md new file mode 100644 index 0000000000000000000000000000000000000000..68c268e49dcdffb909e5a4ce16d8eaf46414af87 --- /dev/null +++ b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/full.md @@ -0,0 +1,351 @@ +# StyLess: Boosting the Transferability of Adversarial Examples + +Kaisheng Liang Bin Xiao* The Hong Kong Polytechnic University + +{cskliang, csbxiao}@comp.polyu.edu.hk + +# Abstract + +Adversarial attacks can mislead deep neural networks (DNNs) by adding imperceptible perturbations to benign examples. The attack transferability enables adversarial examples to attack black-box DNNs with unknown architectures or parameters, which poses threats to many real-world applications. We find that existing transferable attacks do not distinguish between style and content features during optimization, limiting their attack transferability. To improve attack transferability, we propose a novel attack method called style-less perturbation (StyLess). Specifically, instead of using a vanilla network as the surrogate model, we advocate using stylized networks, which encode different style features by perturbing an adaptive instance normalization. Our method can prevent adversarial examples from using non-robust style features and help generate transferable perturbations. Comprehensive experiments show that our method can significantly improve the transferability of adversarial examples. Furthermore, our approach is generic and can outperform state-of-the-art transferable attacks when combined with other attack techniques. + +# 1. Introduction + +Deep neural networks (DNNs) [14, 24] are currently effective methods for solving various challenging tasks such as computer vision, and natural language processing. Although DNNs have amazing accuracy, especially for computer vision tasks such as image classification, they are also known to be vulnerable to adversarial examples [12, 43]. Adversarial examples are malicious images obtained by adding imperceptible perturbations to benign images. Notably, the transferability of adversarial examples is an intriguing phenomenon, which refers to the property that the same adversarial example can successfully attack different black-box DNNs [5, 30, 34, 51]. + +It has been observed that image style can be decoupled from image content, and style transfer techniques allow us + +![](images/4e121519470c08256f3fa8e7871280b7004d4e0a1f88e0e784aafcfa05fbe2db.jpg) + +![](images/b184b7496c37c27594dbd1c96ddb100e8ad8204cc0a9fcb9db059ac1deee6572.jpg) +Benign Sample + +![](images/4c69e622c4e8ff9912cb4e41cdf21686622998cb2e660aea10b2dd4308582521.jpg) +Stylized Output + +![](images/5fc1b5901b3f78bfe106d5fd1f94eef84d8c0d87fcd0657060e53eb3de1a844f.jpg) +StyLess Attack +Figure 1. An overview of our StyLess attack. We create stylized model $\bar{F}$ by injecting synthesized style features into the surrogate model $(F = F_{2} \circ F_{1})$ using an adaptive IN layer. StyLess reduces the use of non-robust style features in the vanilla surrogate model $F$ , ultimately improving attack transferability. + +to generate stylized images based on arbitrary style images [18]. Image style refers to the unique visual characteristics of an image, including its colors, textures, and lighting. For instance, two photos of the same object taken by different photographers can have very different styles. Robust DNNs should rely more on content features of data than style features. This inspired us to improve attack transferability from the perspective of avoiding non-robust features. We believe that style features of DNNs are non-robust for building transferable attacks. However, existing attacks do not distinguish between the surrogate models' style and content features, which may reduce attack transferability. + +We propose using stylized surrogate models to control style features, which can significantly improve transferability. We refer to the original surrogate model as the "vanilla model." The proposed stylized model is created by adding an adaptive instance normalization (IN) layer to the vanilla + +model. By adjusting the parameters of the inserted IN layer, we can easily transform the style features of the surrogate models. To compare the stylized and vanilla surrogate models, we analyzed their network losses during optimization. Surprisingly, we found that the adversarial loss of the vanilla model increased much faster than those of the stylized models, resulting in a widening loss gap. This phenomenon reveals that as the attack iteration progresses, current attack methods only focus on maximizing the loss of the vanilla model, leading to increased use of the style features of the vanilla surrogate model. However, we believe that style features are non-robust for transferable attacks, and relying too much on them may reduce attack transferability. To enhance transferability, we aim to limit the use of non-robust style features and close the loss gap. + +Based on the above findings, we propose a novel method called StyLess to improve the transferability of adversarial examples. Our method uses multiple synthesized style features to compete with the original style features during the iterative optimization of attack. The process is illustrated in Figure 1. We encode various synthesized style features into a surrogate model via an IN layer to achieve stylized surrogate models. Instead of using only the vanilla surrogate model, we use the gradients of both the stylized surrogate models and the vanilla one to update adversarial examples. The front part of the surrogate model works as a style encoder, and the IN layer simulates synthesized style features. Although we can use a decoder to explicitly generate the final stylized samples, it is unnecessary for the proposed attack method. Experimental results demonstrate that StyLess can enhance the transferability of state-of-the-art adversarial attacks on both unsecured and secured black-box DNNs. Our main contributions are summarized as follows: + +- We introduce a novel perspective for interpreting attack transferability: the original style features may hinder transferability. We verify that current iterative attacks increasingly use the style features of the surrogate model during the optimization process. +- We propose a novel attack called StyLess to enhance transferability by minimizing the use of original style features. To achieve this, we insert an IN layer to create stylized surrogate models and use gradients from both stylized and vanilla models. +- We conducted comprehensive experiments on various black-box DNNs to demonstrate that StyLess can significantly improve attack transferability. Furthermore, we show that StyLess is a generic approach that can be combined with existing attack techniques. + +# 2. Related Work + +Adversarial Attacks. Adversarial attacks reveal the vulnerability of current DNNs [43]. The classic adversarial attack methods are gradient-based, such as FGSM [12] and + +I-FGSM [25]. C&W [4] considers optimizing the distance between adversarial examples and benign samples, and proposed optimization-based attacks. Adversarial attacks can also be performed in the physical world [10, 39]. As for defending against adversarial examples, adversarial training is a popular defense method that uses adversarial examples as extra training data to improve robustness [32]. + +Increasing Attack Transferability. An intriguing property of adversarial attacks is the transferability. Ensemble-based attack [30] uses multiple surrogate networks instead of one network. Ghost networks [27] generates different surrogate networks by perturbing skip connection and dropout layers. Optimization methods, such as MI [5], uses a momentum-based optimization, while VT [47] introduces gradient variance to control the stability of the localized gradients. RAP [35] generates adversarial examples located in a flat loss region. Data augmentation methods, such as DI [51], uses image transformation like resizing and padding, while TI [6] considers translating image pixels. SI [29] calculates gradients with the help of several scaled benign samples. Admix [48] calculates iterative gradients by mixing the benign images with randomly sampled images. + +Various network architectures and features exhibit different relationships with adversarial attacks. DNNs' linearity is believed to cause adversarial vulnerability [12], and LinBP [13] skips the nonlinear activation during the backpropagation. SGM [49] uses more gradients through skip connections in residual networks. To better leverage the intermediate layers, one can train auxiliary classifiers based on feature spaces [20, 21], maximize the distance between natural images and their adversarial examples in feature spaces [54], or fine-tune the existing adversarial examples in intermediate layer level by ILA [17, 26]. + +Style Transfer and Instance Normalization. Style transfer can change the style of an image to match the style of another one [9, 11, 23]. Fast feedforward networks can perform stylization with arbitrary styles in a single forward pass [18, 28]. Interestingly, style transfer has a wide range of applications. AdvCam [8] uses natural styles to hide non- $L_{p}$ restricted perturbations. FSA [52] generates natural-looking adversarial examples by using optimized style changes. Style transfer has also been used to improve network robustness by exploring additional feature information [33]. Latent style transformations can detect adversarial attacks [46]. AMT-GAN [15] proposes an adversarial makeup transfer to protect facial privacy by preserving stronger black-box transferability. + +The family of instance normalization (IN) including batch normalization [22], layer normalization [1], instance normalization [45], and group normalization [50]. Normalizations are mainly used to reduce the covariate shift, and speed model training. Recently, normalizations have been found to be related to robustness. It has been shown that + +batch normalization makes DNNs use more non-robust but useful features [3, 19]. AdvBN proposed adding an extra batch normalization into network training to increase training loss adversarially, which enables the network to resist various domain shifts [40]. Adjusting batch normalization statistics such as the running mean and variance in the inference phase, which are estimated during training, improves robustness and defense common corruption [2, 38]. + +Among existing style-based attacks, FSA [52] differentiates style features and content features, which is similar to our method. However, there are three significant differences between FSA and our approach: 1) FSA proposes to hide adversarial perturbations in the optimized style, while we avoid relying on any style. 2) FSA aims at enhancing the natural looking of non- $\ell_p$ restricted attacks, while we focus on the transferability of $\ell_p$ restricted adversarial examples. 3) Both FSA and our work are inspired by AdaIN [18], but we use IN layer differently. FSA perturbs the IN layer to search malicious styles and requires a decoder. But we use randomized IN layers to augment attacks and don't need to train a decoder. + +# 3. Methodology + +# 3.1. Threat Model + +Attack objective. Given a benign image $x$ with label $y$ , transfer-based attacks aim to generate an adversarial perturbation based on a white-box surrogate network $F$ . The general attack objective can be formulated as follows: + +$$ +\max _ {\delta} \mathcal {L} (F (x + \delta), y) \quad \text {s . t .} \quad \| \delta \| \leq \epsilon , \tag {1} +$$ + +where $\mathcal{L}$ denotes the adversarial loss, $\delta$ is the adversarial perturbation, and $\epsilon$ is the maximum perturbation size. + +A popular framework to solve the above problem is iterative fast gradient sign method (I-FGSM) [12, 25]: + +$$ +x _ {a d v} ^ {t + 1} = x _ {a d v} ^ {t} + \alpha \cdot \operatorname {s i g n} \left(\nabla_ {x} \mathcal {L} \left(F \left(x _ {a d v} ^ {t}\right), y\right)\right), \tag {2} +$$ + +where $\alpha$ is the learning rate, and a clip function will be used on $x_{adv}^{t + 1}$ to ensure $\| x_{adv}^{t + 1} - x\| \leq \epsilon$ + +Attacker capability. We follow the same setting in previous work that attackers have a surrogate model and some test samples, but cannot access target models, and don't know network architectures, training data, or defense strategies. It should be noted that our method doesn't require any additional datasets. Our approach involves style features, which can be extracted from an arbitrary image or synthesized without any style image. + +Transferable attacks as black-box attacks. Transferable attacks use the surrogate model $F$ to create adversarial examples that can fool unseen target models. In this way, these attacks can be viewed as black-box attacks. + +# 3.2. Motivation + +Existing transferable attacks often rely on the gradient of the adversarial loss function $\mathcal{L}$ (Equation 2) without considering the impact of different components of $\mathcal{L}$ . However, these approaches have limitations because transferable attacks should minimize the use of non-robust features of the surrogate model. Interestingly, for image classification task, style features of images are typically less robust than content features. Based on this observation, we propose to enhance attack transferability by explicitly reducing the use of style features of the surrogate model within the loss $\mathcal{L}$ . + +Our key idea is simulating various surrogate models without the style features of the given vanilla surrogate model. We discovered that inserting an IN layer into the vanilla surrogate model enables us to create new surrogate models that we refer to as stylized surrogate models. Sometimes we omit the word "surrogate." Stylized models can explicitly manipulate style features without compromising model accuracy. Recall that our goal is to construct adversarial examples that can mislead unseen target models, which should include these stylized models. However, existing methods, such as MI and I, only focus on maximizing the loss of the vanilla model. + +Figure 2 indicates that MI and I(-FGSM) have limited attack transferability on stylized models since the vanilla model's adversarial loss increases much faster than stylized models', resulting in a widening loss gap. In the following sections, we will demonstrate how our method addresses this issue by maximizing the loss of both the stylized and vanilla models, which significantly improves transferability. + +# 3.3. Stylized Surrogate Models + +This section will give the definition of our stylized surrogate models, which encode various style features by inserting an IN layer. Then we will analyze the stylized loss gap $(\Delta \mathcal{L})$ between the vanilla and stylized surrogate models. Specifically, we will illustrate the increasing $\Delta \mathcal{L}$ limits transferability and our idea to decrease $\Delta \mathcal{L}$ . + +# 3.3.1 Encoding Styles by Stylized Models + +Given a classifier $F = F_{2} \circ F_{1}$ as the surrogate model, we define a stylized surrogate model as + +$$ +\bar {F} _ {x _ {s}} = F _ {2} \circ \mathrm {I N} _ {x _ {s}} \circ F _ {1}, \tag {3} +$$ + +where $x_{s}$ is a style input, $\mathrm{IN}_{x_s}$ is an IN layer instantiated by $x_{s}$ . In general, an IN layer is defined as + +$$ +\operatorname {I N} (x; \mu , \sigma) = \sigma \cdot \left(\frac {x - \mu (x)}{\sigma (x)}\right) + \mu , \tag {4} +$$ + +where $\mu$ and $\sigma$ are the network parameters of layer IN, and $\mu(x)$ and $\sigma(x)$ are the mean and variance of input $x$ . According to adaptive instance normalization (AdaIN) [18], to + +![](images/d3a4f1f6154ac9eb82e4d946815110808987b22bd0e05522feb72b5946012cdf.jpg) + +![](images/57ef0de4d49b9bdbb8839d88f796e1e89860a296ac4f5c1c09c51f49238c7752.jpg) + +![](images/c00564f5812bb613c7a341be67caeb0ffee7a5d35392dc08f953cc55d0f4f0a6.jpg) +(a) Baseline attack MI vs ours $\mathrm{MI + StyLess}$ +Figure 2. Illustration of the loss gap between vanilla network and stylized network, with RN50 as surrogate model. The greater the loss, the better the attack performance. The widening loss gap in the baseline means that attack performance on stylized models is lagging behind. The proposed StyLess can close the gap. + +![](images/2684c3128e67515e02360ce3985f115f4c886f044307a715c14891482fd808e9.jpg) +(b) Baseline attack I(-FGSM) vs ours I+StyLess + +stylize an input $x$ with a given style input $x_{s}$ , we only need to instantiate the IN as + +$$ +\left. \operatorname {I N} _ {x _ {s}} (x) = \operatorname {I N} (x) \right| _ {\mu = \mu \left(x _ {s}\right), \sigma = \sigma \left(x _ {s}\right)} \tag {5} +$$ + +Our stylized model $\bar{F}_{x_s}$ has encoded style features. Based on AdaIN, the $F_{1}$ from the classifier $F = F_{2} \circ F_{1}$ works as an encoder for style transfer, and given a style input $x_{s}$ , we can get a stylized image as $\bar{x} = D \circ \mathrm{IN}_{x_s} \circ F_1(x)$ where $D$ denotes a decoder. Thus, $\mathrm{IN}_{x_s} \circ F_1(x)$ has encoded the style features of $x_{s}$ to $\bar{F}_{x_s}$ . + +# 3.3.2 Stylized Loss Gap Limits transferability + +To validate the performance of adversarial attacks on these stylized surrogate models, we define a stylized loss gap as + +$$ +\Delta \mathcal {L} = \mathbb {E} _ {x _ {s} \in \mathcal {D}} [ \mathcal {L} (F (x), y) - \mathcal {L} (\bar {F} _ {x _ {s}} (x), y) ], \tag {6} +$$ + +where $F$ and $\bar{F}_{x_s}$ are vanilla and stylized model, respectively; $x_{s}$ is style input. + +An increasing loss gap $\Delta \mathcal{L}$ can be observed in Figure 2 which limits attack transferability, as explained below. Hypothetically, let's say we can decouple style-dependent loss from content-dependent loss in $\mathcal{L}$ as follows: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} ^ {c} + \mathcal {L} _ {x} ^ {s}, \\ \bar {\bar {z}} = z ^ {c} - z ^ {s} \end{array} \tag {7} +$$ + +$$ +\mathcal {L} = \mathcal {L} ^ {c} + \mathcal {L} _ {x _ {s}} ^ {s}, +$$ + +where $\mathcal{L}$ is the vanilla loss, and $\bar{\mathcal{L}}$ is the stylized one; $\mathcal{L}^c$ is the common content-dependent loss; $\mathcal{L}_x^s$ is input sample $x$ -specific style-dependent loss, while $\mathcal{L}_{x_s}^s$ is style sample $x_{s}$ -specific style-dependent loss. In this case, $\Delta \mathcal{L} = \mathcal{L}_x^s -\mathcal{L}_{x_s}^s$ + +In other words, the non-robust features are related to $\mathcal{L}_x^s$ and $\mathcal{L}_{x_s}^s$ , while $\mathcal{L}_c$ is supposed to be shared by other unseen DNNs. MI and I-(FGSM) have increasing gaps in Figure 2 means that attackers gradually focus on optimizing the $\mathcal{L}_x^s$ part, which only belonged to the vanilla surrogate model $F$ . Therefore, the loss gap limits the transferability of adversarial examples in unseen stylized models. + +To decrease $\Delta \mathcal{L}$ and boost transferability, we consider involving $\mathcal{L}_{x_s}^s$ as a competitor of $\mathcal{L}_x^s$ in optimization process to suppress the growth of $\mathcal{L}_x^s$ . In general, we can assume that all these losses in Equation 7 are non-negative, and they satisfy: $\mathcal{L}^c\gg \mathcal{L}_x^s$ , $\mathcal{L}^c\gg \mathcal{L}_{x_s}^s$ . Also, there is a upper bound $B$ for these losses: $\mathcal{L} < B$ , $\bar{\mathcal{L}} < B$ , as the adversarial perturbations are required to smaller than a given $\epsilon$ . If we only maximize the vanilla loss $\mathcal{L}$ , both $\mathcal{L}^c$ and $\mathcal{L}_x^s$ are likely to be increased. To this end, we propose to maximize $\mathbb{E}_{x_s\in \mathcal{D}}\mathcal{L}_{x_s}^s +\mathcal{L}_x^s +\mathcal{L}^c$ , which involves multiple style-dependent losses $\mathcal{L}_{x_s}^s (x_s\in \mathcal{D})$ to compete with $\mathcal{L}_x^s$ and leads to a decrease of $\Delta \mathcal{L}$ . + +# 3.4. Proposed Style-Less Perturbations (StyLess) + +Based on the above analysis, we propose style-less perturbations (StyLess) method to increase attack transferability by optimizing stylized loss and vanilla loss together: + +$$ +\max _ {\delta} \mathbb {E} _ {x _ {s} \in \mathcal {D}} \mathcal {L} \left(\bar {F} _ {x _ {s}} (x + \delta), y\right) + \mathcal {L} \left(F (x + \delta), y\right). \tag {8} +$$ + +The key to generate multiple stylized models $\bar{F}_{x_s}$ is synthesizing style statistics $\mu, \sigma$ for Equation 5 to obtain parameterized IN layers. We propose using scaling and interpolation to simulate multiple style features, formulated as + +$$ +\mu = \beta (\lambda \mu_ {x} + (1 - \lambda) \mu_ {s}), \tag {9} +$$ + +$$ +\sigma = \gamma (\lambda \sigma_ {x} + (1 - \lambda) \sigma_ {s}), +$$ + +where $\mu_x, \sigma_x$ is the mean and variance of $F_1(x)$ , relating to the benign content input $x$ , while $\mu_s, \sigma_s$ are for style input $x_s$ similarly. $x_s$ is an arbitrary image. $\lambda$ is a scalar that controls the interpolation of two styles. $\beta$ and $\gamma$ are $c$ -dimensional vectors that scale the synthesized style, where $c$ refers to the number of channels in the style feature. + +We summarize the proposed StyLess attack in Algorithm 1. To obtain the parameters of the IN layer without training, we select a random image as $x_{s}$ , and generate a pair of $\mu$ and $\sigma$ using Equation 9. With $\mu$ and $\sigma$ , we can create a stylized model $\bar{F}_{x_s}(x)$ using Equation 3 and 5. $\bar{F}_{x_s}(x)$ need to be equal to $F(x)$ or label $y$ , otherwise we will regenerate $\mu$ and $\sigma$ by altering $\beta, \gamma$ and $\lambda$ . In each iteration, we use $N$ stylized models to update adversarial examples based on the proposed objective function in Equation 8. + +Algorithm 1 Style-Less Perturbations (StyLess) Algorithm +Input: Surrogate model $F$ , benign example $x$ , iteration number $T$ , maximum perturbation $\epsilon$ , data augmentation $\phi(\cdot)$ , decay factor $\eta$ . Scale factors $\beta, \gamma$ and interpolation factor $\lambda$ . The number of stylized models $N$ . +Output: An adversarial example $x_{adv}$ +1: $x_{adv} = x$ , $g_0 = 0$ , $\alpha = \epsilon/2$ . +2: for $t = 0 \rightarrow T - 1$ do: +3: Augment input $x_{adv} = \phi(x_{adv})$ +4: Obtain gradient $\tilde{g}_{t+1}$ with respect to $x_{adv}$ using $F$ +5: repeat +6: Synthesize a style statistic by Equation 9 +7: Obtain a stylized model $\bar{F}_{x_s}$ by Equation 3 +8: Get gradient $\tilde{g}$ with respect to $x_{adv}$ using $\bar{F}_{x_s}$ +9: Update $\tilde{g}_{t+1} = \tilde{g}_{t+1} + \tilde{g}$ +10: until obtain $N$ stylized models +11: Calculate momentum $g_{t+1} = \eta \cdot g_t + \tilde{g}_{t+1} / \| \tilde{g}_{t+1} \|_1$ +12: Update example $x_{adv} = x_{adv} + \alpha \cdot \text{sign}(g_{t+1})$ +13: end for +14: return $x_{adv}$ . + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. We use ImageNet [36] for experiments. Specifically, we use 1000 images which are randomly selected from ImageNet validation set by Wang et al. [47]. Similar dataset settings have been widely used in previous work [5, 13, 29, 47, 49, 51]. These images include all categories; almost all are correctly classified by the target DNNs. + +Models. We evaluate the generated adversarial examples on different black-box DNNs, including both unsecured and secured models. Unsecured models are trained on ImageNet using traditional methods, while secured models are based on adversarial training. The unsecured models include VGG19 [41], AlexNet [24], ResNet50 (RN50) [14], WideResNet101 (WRN101) [53], DenseNet121 (DN121) [16], InceptionV3 (IncV3) [42], MnasNet [44], MobileNetV2 (MNv2) [37], ShuffleNetV2 (SNv2) [31] and ViT [7]. Their pre-training parameters are obtained from PyTorch official. The secured DNNs are IncV3ens3 (ensemble of 3 InceptionV3 networks), IncV3ens4 (ensemble of 4 InceptionV3 networks) and IncResV2ens3 (ensemble of 3 IncResV2 networks). These models were adversarily trained and widely used in previous work [5,30,47,49]. As for the surrogate models, we use VGG19 [41], RN50 [14], WRN101 [53] and DN121 [16]. + +Implementation Details. We use I-FGSM [12] as the initial baseline. Unless otherwise specified, the attacks are untargeted and $l_{\infty}$ -restricted. The maximum perturbation size is set to $\epsilon = 16 / 255$ . We compare StyLess primarily with six transferable attacks: MI [5], DI [51], TI [6], SI [29], + +Admix (AI) [48], and an ensemble-based approach [30]. We set the optimization step size to $\alpha = \epsilon /2$ , and the number of iterations to $T = 50$ . The momentum decay in MI is $\mu = 1$ . For DI, SI and AI, we follow the official settings described in the corresponding papers. For StyLess, we simulate 10 stylized models in each iteration, denoted by $N = 10$ . To generate a stylized model for a given $x$ , we randomly sample $\lambda$ from $[0,0.2]$ , and $\beta, \gamma$ from $[0,2]$ , and ensure that $\bar{F}_{x_s}(x)$ is equal to $F(x)$ or the real label. The IN layer is inserted after the first bottleneck block for RN50 and WRN101, and after the first dense block for DN121. + +# 4.2. Attacking Unsecured Models + +We compare StyLess with other attacks on various unsecured DNNs using three surrogate models. Table 1 shows that StyLess is a powerful and generic method that can be combined with existing attack methods to further improve attack transferability. Specifically, we compare StyLess with I, MI, DI, TI, SI and Admix (AI). For the most challenging case in the table, attacking the black-box IncV3 (let's take $\mathrm{RN50} \Rightarrow \mathrm{IncV3}$ attack as an example), StyLess significantly improves the attack success rate of baseline attacks (I and MI): $46.2\% \rightarrow 68.3\%$ (I), $59.2\% \rightarrow 78.9\%$ (MI). StyLess also demonstrates its capabilities when using other DNNs as the surrogate network. For instance, DN121 $\Rightarrow$ SNv2 attack, StyLess significantly improves the baseline: $69.3\% \rightarrow 91.4\%$ (I), $77.4\% \rightarrow 95.1\%$ (MI). + +StyLess can be combined with other attack techniques. Previous work has shown that combining various attack methods results in powerful and transferable attacks. StyLess can be integrated with existing combination-based attacks to enhance attack transferability. We report the results of four combinations of existing attacks: MDI, MTDI, MTDSI, and MTDAI. StyLess further enhances these four attack methods' attack success rate by $+8.0\%$ , $+8.2\%$ , $+3.0\%$ , and $+3.3\%$ (in the case of $\mathrm{RN}50 \Rightarrow \mathrm{IncV}3$ attack). We evaluate attack performance using various surrogate models, including RN50, WRN101 and DN121. For instance, when combining MTDAI with StyLess, our method enhances the transferability of MTDAI: WRN101 $\Rightarrow$ IncV3 attack: $+4.7\%$ , WRN101 $\Rightarrow$ MNv2 attack: $+1.5\%$ , WRN101 $\Rightarrow$ SNv2 attack: $+3.4\%$ , DN121 $\Rightarrow$ IncV3 attack: $+4.3\%$ , DN121 $\Rightarrow$ MNv2 attack: $+0.6\%$ , DN121 $\Rightarrow$ SNv2 attack: $+3.8\%$ . The results show that StyLess is an efficient and generic approach for improving attack transferability. + +# 4.3. Attacking Secured Models + +We evaluate StyLess on three widely-used secured models: IncV3ens3, IncV3ens4 and IncResV2ens, as shown in Table 2. We present the results of I, MI, MDI, MTDI, and MTDSI on the three secured models using two surrogate models: RN50 and WRN101. These secured networks are more robust than the unsecured DNNs we men + +Table 1. Attacking unsecured black-box models with StyLess. + +
SourceAttackVGG19RN50WRN101DN121IncV3MNv2SNv2
RN50I / +Ours72.8 / 88.8100 / 10080.8 / 97.183.0 / 97.846.2 / 68.377.1 / 91.960.6 / 77.8
MI / +Ours82.9 / 94.1100 / 10083.9 / 97.287.5 / 98.759.2 / 78.983.8 / 93.272.3 / 83.5
MDI / +Ours97.5 / 99.2100 / 10098.2 / 99.899.4 / 10089.5 / 97.598.1 / 99.988.5 / 96.9
MTDI / +Ours98.6 / 99.7100 / 10099.2 / 10099.8 / 10090.2 / 98.498.7 / 10090.8 / 98.1
MTDSI / +Ours98.6 / 99.2100 / 10099.5 / 10099.8 / 10096.2 / 99.298.7 / 99.996.2 / 98.2
MTDAI / +Ours99.3 / 99.6100 / 10099.8 / 10099.9 / 10095.5 / 98.899.7 / 10095.7 / 99.0
WRN101I / +Ours64.8 / 79.688.6 / 98.7100 / 10077.7 / 94.243.2 / 66.668.1 / 84.758.1 / 73.2
MI / +Ours75.6 / 86.489.8 / 98.7100 / 10083.9 / 96.457.6 / 73.776.1 / 87.070.2 / 81.2
MDI / +Ours92.2 / 98.599.0 / 99.9100 / 10097.1 / 99.986.3 / 96.393.6 / 99.086.8 / 95.8
MTDI / +Ours92.5 / 98.699.2 / 100100 / 10097.9 / 99.689.2 / 97.995.4 / 99.388.4 / 97.0
MTDSI / +Ours95.9 / 98.999.6 / 100100 / 10099.7 / 10095.8 / 99.597.7 / 99.994.5 / 98.4
MTDAI / +Ours96.5 / 99.299.9 / 99.9100 / 10099.4 / 10093.7 / 98.498.1 / 99.694.6 / 98.0
DN121I / +Ours79.7 / 97.787.0 / 99.474.7 / 98.0100 / 10055.7 / 88.781.3 / 97.069.3 / 91.4
MI / +Ours86.9 / 99.189.4 / 99.778.2 / 98.6100 / 10066.0 / 95.786.5 / 98.777.4 / 95.1
MDI / +Ours96.7 / 99.998.9 / 10095.4 / 99.8100 / 10090.5 / 99.097.5 / 10090.9 / 98.6
MTDI / +Ours97.5 / 99.999.1 / 10095.9 / 99.8100 / 10092.7 / 99.496.9 / 10092.2 / 98.6
MTDSI / +Ours97.4 / 99.999.3 / 10098.0 / 99.8100 / 10096.5 / 99.898.9 / 10095.1 / 99.1
MTDAI / +Ours99.2 / 99.999.8 / 10098.1 / 99.9100 / 10095.1 / 99.499.4 / 10095.2 / 99.0
+ +Table 2. Attacking three secured black-box models with StyLess. The surrogate model is RN50 or WRN101. + +
AttackRN50⇒
IncV3ens3IncV3ens4IncResV2ens
I / +Ours21.6 / 34.618.9 / 32.014.1 / 21.5
MI/ +Ours31.5 / 47.129.3 / 42.420.8 / 31.0
MDI / +Ours59.6 / 78.153.5 / 69.838.3 / 57.3
MTDI / +Ours69.2 / 89.663.8 / 81.854.6 / 72.2
MTDSI / +Ours88.0 / 93.184.7 / 91.377.8 / 84.5
WRN101⇒
I / +Ours23.4 / 40.320.9 / 34.615.5 / 25.6
MI/ +Ours35.0 / 51.230.4 / 46.624.7 / 35.9
MDI / +Ours64.0 / 81.558.8 / 76.246.9 / 66.4
MTDI / +Ours75.5 / 91.470.5 / 87.463.6 / 81.0
MTDSI / +Ours91.6 / 97.588.6 / 96.183.0 / 92.5
+ +tioned above. Taking $\mathrm{RN50} \Rightarrow \mathrm{IncV3}_{\mathrm{ens3}}$ attack as an example, I, MI, MDI, and MTDI achieve attack success rates of only $21.6\%$ , $31.5\%$ , $59.6\%$ , and $69.2\%$ , respectively. StyLess effectively improves the performance of these attacks. Notably, MTDSI has been improved by our method: $91.6\% \rightarrow 97.5\%$ in WRN101 $\Rightarrow$ IncV3ens3 attack; $88.6\% \rightarrow$ $96.1\%$ in WRN101 $\Rightarrow$ IncV3ens4 attack; $83.0\% \rightarrow 92.5\%$ in WRN101 $\Rightarrow$ IncResV2ens attack. StyLess demonstrates its great power to enhance transferability in breaking these challenging secured DNNs. + +Figure 3 shows a comparison of StyLess and LinBP. We consider LinBP and its combination with existing attacks such as ILA, SGM, and MDI on various target DNNs. Our method exhibits the best attack transferability. + +![](images/cbf30828713e1c2f8a39b49983f4f7d50063c040818e9006083e276d10e45445.jpg) +Figure 3. Comparing with LinBP on different black-box models and defenses. The surrogate model is RN50. + +# 4.4. Combining with Ensemble-Based Method + +Generating adversarial examples based on multiple surrogate networks simultaneously can improve attack performance in practice [30]. This ensemble-based method has been shown to be a powerful attack, and StyLess can further improve it. We use an ensemble of RN50, WRN101, and DN121 as the integrated surrogate model. + +We compared our StyLess with ensemble-based MI and MTDI in Table 3. The considerable strength of the ensemble-based method can be seen when comparing with the results in Table 1 and 2. For example, in RN50 $\Rightarrow$ IncV3 attack, the baseline attack MI only gets $59.2\%$ , while it gets $81.0\%$ attack success rate in the case of RN50+WRN101+DN121 $\Rightarrow$ IncV3 attack. Noted that we generally use $\epsilon = 16$ , which is a standard setting. As we can see, when $\epsilon = 16$ , the ensemble-based MTDI achieves an average attack success rate of less than $90\%$ on the three + +Table 3. Combining with ensemble-based attack method. + +
εAttackRN50+WRN101+DN121⇒
VGG19RN50WRN101DN121IncV3MNv2SNv2IncV3ens3IncV3ens4IncResV2ens
4MI56.896.494.897.235.762.348.318.415.89.4
MTDI79.297.996.698.867.584.571.636.532.122.3
+StyLess90.499.599.499.983.294.385.450.042.929.1
8MI79.299.698.799.458.383.368.031.426.818.2
MTDI95.299.999.599.989.597.590.467.661.846.9
+StyLess99.110010010097.599.697.285.177.466.1
16MI92.310010010081.093.285.053.046.335.0
MTDI99.210010010097.510097.992.786.480.7
+StyLess10010010010099.810010098.597.594.3
+ +secured networks: IncV3ens3, IncV3ens4, and IncResV2ens. StyLess can further boost the transferability of ensemble-based MTDI: $92.7\% \rightarrow 98.5\%$ for IncV3ens3, $86.4\% \rightarrow 97.5\%$ for IncV3ens4, $80.7\% \rightarrow 94.3\%$ for IncResV2ens. These results show that StyLess is a different type of attack, and can work perfectly with the ensemble-based method. + +We also report the experimental results with different $\epsilon$ . We use $\epsilon = 4$ or 8 to increase the difficulty to attack. For instance, in the case of ensemble-based $\mathrm{MTDI} \Rightarrow \mathrm{IncV3}_{\mathrm{ens3}}$ attack, attack success rate drops from $92.7\%$ to $67.6\%$ when $\epsilon = 8$ instead of $\epsilon = 16$ , and StyLess can help ensemble-based MTDI gains $+17.5\%$ ( $67.6\% \rightarrow 85.1\%$ ), which is a huge improvement. When $\epsilon = 4$ , the results also demonstrate the advantages of StyLess. This shows that StyLess consistently delivers strong transferability when faced with a more robust network that is difficult to attack. + +# 4.5. Attacking the Google Cloud Vision API + +We use the Google Cloud Vision API as an example of real-world applications to evaluate the transferability of attacks. This API allows us to use various vision features, such as image labeling and face detection. As a black-box model, regular users like us cannot access its network architecture, training data, or defense mechanism. In this section, we will focus on attacking its image labeling feature. + +To utilize the image labeling feature of the API, we need to upload images and obtain the predicted labels. We use 1000 images from ImageNet as the target images, as described in the experimental setup. Due to the fact that the API does not support all 1000 categories in ImageNet, our objective is to mislead the API's original top-1 prediction. + +We use ResNet50 as the surrogate network to compare the baseline method MTDSI with our method. Experimental results show that MTDSI achieves an attack success rate of $75.6\%$ on the Google Cloud Vision API. Our approach, MTDSI-SyLess, achieves an attack success rate of $85.2\%$ , which is a $9.6\%$ improvement over the baseline. Our results demonstrate that transfer-based black-box attacks pose a severe threat to real-world applications, and StyLess can effectively boost attack transferability. + +# 4.6. Ablation Study + +In this session, we will present four ablation studies: 1) The position of the inserted IN layer; 2) The number of the generated stylized models; 3) The clean losses of stylized models and attack transferability; 4) The most important statistic of style features. + +Which position to insert the IN layer? As shown in Figure 4, we use features from different layers of the vanilla surrogate networks as the encoder $F_{1}$ for style transfer. We evaluate the attack success rates using two surrogate networks: WRN101 and DN121. After injecting synthesized styles into different network layers, we report the attack success rates on various DNNs, including AlexNet, VGG19, RN50, WRN101, and DN121. We observe a trend that the best attack success rates are usually achieved in the shallow layers of the surrogate networks. For example, when using WRN101 as the surrogate network, injecting the synthesized styles in the layers before layer ten is a good choice. Using intermediate layers such as layers 40 to 80 is also acceptable, but the attack performance may be unstable or even worse when injecting the styles in the last few layers. The results of using DN121 as the surrogate network also indicate that the last few layers are the worst choices, and the shallow layers are the optimal. + +![](images/1ea263319dba8a4b7ae047d615b6e042abcdaacace7e706f94f8cc297766e967.jpg) +Figure 4. Ablation study on using which network layer to synthesize styles. The surrogate networks are WRN101 and DN121. + +![](images/08b72aef0a42964ce780b7380181ac3a300bd88b3cf4ceb742349c0424ed1186.jpg) + +How many stylized models should be created in each iteration? In Figure 5, we vary the number of stylized models generated in each attack iteration from zero to + +ten. We conduct experiments by combining StyLess with two baseline attacks: MI and DI. The first figure evaluates the attack success rates on three unsecured networks for our $\mathrm{MI} + \mathrm{StyLess}$ attack. In the second figure, we test $\mathrm{DI} + \mathrm{StyLess}$ on three secured models. When the number of stylized models is 0, StyLess is not involved, so it is the vanilla baseline attack. As we can see, when the number increases from 0 to 1, the attack success rate starts to grow, which means StyLess starts to work. There is an anomaly when DI begins to combine with StyLess. If the number is less than three, the attack success rate on IncResV2 $_{ens}$ is slightly worse than the baseline (around $25\%$ ). When the number increases by more than three, the attack success rate becomes higher than the baseline. This may be because IncResV2 $_{ens}$ is very strong, and more synthesized style features need to be injected into the surrogate model. According to the figure, StyLess works quite well when six to ten stylized models are used in an attack iteration. + +![](images/26c2de9ed8b7a09ab3e7993e4a02aabf49c0486747e0cca07ade759cd9524c1a.jpg) +Figure 5. Ablation study on the number of stylized models in an attack iteration. The surrogate network is RN50. + +![](images/95a270d533d13e1510ccc974bb86c4626743fc8fe389ab9642f34e0d39c2373e.jpg) + +How network loss affects attack success rate? In Figure 6, we demonstrate how varying strengths of style injection can affect network loss, which in turn impacts attack performance on a surrogate network (in this case, VGG19). The strength of synthesized style features is denoted by the number of stars, with more stars indicating greater style strength that alters the style features of the surrogate model. Generally, injecting synthesized style features should not significantly affect the clean loss, as an increase in network loss typically leads to a decrease in clean accuracy. Overly corrupted stylized surrogate models can also result in bad gradients for attack methods. Therefore, there is an upper bound on clean loss when generating stylized models. The red line in the figure represents the situation of overly corrupting, in which the clean loss exceeds the estimated bound (indicated by a yellow line in the right figure), and the attack success rate drops significantly. This demonstrates the importance of maintaining relatively good clean accuracy when creating stylized networks. + +Which statistic of style features matters most? In Figure 7, we compare the effects of the mean and variance of style features on StyLess. Here the interpolation fac + +![](images/dbe6534e33657c4c91dce20b9f2181a5b968269717ac629b442d06b57e96d2dc.jpg) +Figure 6. Study how synthesized style features affect the clean loss, which in return impacts the attack success rate. The number of stars indicates the degree of change in original style features. + +![](images/52dd2e04af075b6d32dd07abb33c26013e3a1709f9def9de45a65c7a6ad60182.jpg) + +tor is $\lambda = 0$ . Equation 9 shows that $\beta$ involves the mean in IN, while $\gamma$ affects the variance. The attack success rate represents the average success rates of attacks on five DNNs: VGG19, AlexNet, RN50, WRN101, and DN121. The results show that $\gamma$ plays a more important role than $\beta$ . Specifically, when RN50 is used as the surrogate network, modifying $\beta$ alone barely improves the baseline attack, while involving $\gamma$ alone enhances the attack success rate by around $5\%$ . A similar observation can be made for VGG19. From the perspective of gradient calculations, this also makes sense. When we backpropagate through an IN layer, we have $\frac{\partial}{\partial x} \mathrm{IN} = \sigma$ , which also indicates that the variance matters most for adversarial attacks. + +![](images/59079058ed5ab83c5bc5bccaab88489a161886b76ab926ba5444b02608e316e3.jpg) +Figure 7. The effect of different statistics of style. + +# 5. Conclusion + +In this work, we analyze the mechanism of attack transferability in terms of style features. We demonstrate that existing attack methods increasingly use the style features of surrogate models during the iterative optimization, which hampers attack transferability. To address this issue, we propose a novel attack method called StyLess to enhance transferability by reducing reliance on original style features. StyLess uses stylized surrogate models instead of a vanilla surrogate model. Experimental results show that StyLess outperforms existing attacks by a large margin, and can be combined with other attack methods. Notably, StyLess is a different paradigm from previous transferable attack methods, and we hope it will shed light on the interpretation of adversarial attacks in the future. + +# References + +[1] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 2 +[2] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Revisiting batch normalization for improving corruption robustness. In WACV, 2021. 3 +[3] Philipp Benz, Chaoning Zhang, and In So Kweon. Batch normalization increases adversarial vulnerability and decreases adversarial transferability: A non-robust feature perspective. In ICCV, 2021. 3 +[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 2 +[5] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 5 +[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 2, 5 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5 +[8] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A. K. Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2 +[9] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In ICLR, 2017. 2 +[10] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2 +[12] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 3, 5 +[13] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples. In NeurIPS, 2020. 2, 5 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5 +[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In CVPR, 2022. 2 +[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5 +[17] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 2 + +[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 2, 3 +[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 3 +[20] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 2 +[21] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 2 +[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2 +[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2 +[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 1, 5 +[25] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 2, 3 +[26] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 2 +[27] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning transferable adversarial examples via ghost networks. In AAAI, 2020. 2 +[28] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 2017. 2 +[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 2, 5 +[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 1, 2, 5, 6 +[31] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5 +[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2 +[33] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. Stylized adversarial defense. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2 +[34] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 1 +[35] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 2 + +[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115:211-252, 2015. 5 +[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, 2018. 5 +[38] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, 2020. 3 +[39] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM Conference on Computer and Communications Security (CCS), pages 1528-1540, 2016. 2 +[40] Manli Shu, Zuxuan Wu, Micah Goldblum, and Tom Goldstein. Encoding robustness to image style via adversarial feature perturbations. NeurIPS, 2021. 3 +[41] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5 +[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 5 +[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2 +[44] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, 2019. 5 +[45] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 2 +[46] Shuo Wang, Surya Nepal, Alsharif Abuadbba, Carsten Rudolph, and Marthie Grobler. Adversarial detection by latent style transformations. IEEE Transactions on Information Forensics and Security (TIFS), 17:1099-1114, 2022. 2 +[47] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 2, 5 +[48] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 2, 5 +[49] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 2, 5 +[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 2 +[51] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 2, 5 + +[52] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2, 3 +[53] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5 +[54] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018. 2 \ No newline at end of file diff --git a/2023/StyLess_ Boosting the Transferability of Adversarial Examples/images.zip b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..21919662af89108dcc63dc064ca4f4c94c9ce2da --- /dev/null +++ b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d131df81f1cd5f98d2e73360e4ca08b4b6e78a606e379d9bae422ea0441aafda +size 598267 diff --git a/2023/StyLess_ Boosting the Transferability of Adversarial Examples/layout.json b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5860c83b2784820df295e9bc1ae2b376c72a392 --- /dev/null +++ b/2023/StyLess_ Boosting the Transferability of Adversarial Examples/layout.json @@ -0,0 +1,11460 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 103, + 490, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 103, + 490, + 121 + ], + "spans": [ + { + "bbox": [ + 103, + 103, + 490, + 121 + ], + "type": "text", + "content": "StyLess: Boosting the Transferability of Adversarial Examples" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 144, + 392, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 144, + 392, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 144, + 392, + 171 + ], + "type": "text", + "content": "Kaisheng Liang Bin Xiao* The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 174, + 394, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 174, + 394, + 185 + ], + "spans": [ + { + "bbox": [ + 196, + 174, + 394, + 185 + ], + "type": "text", + "content": "{cskliang, csbxiao}@comp.polyu.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 478 + ], + "type": "text", + "content": "Adversarial attacks can mislead deep neural networks (DNNs) by adding imperceptible perturbations to benign examples. The attack transferability enables adversarial examples to attack black-box DNNs with unknown architectures or parameters, which poses threats to many real-world applications. We find that existing transferable attacks do not distinguish between style and content features during optimization, limiting their attack transferability. To improve attack transferability, we propose a novel attack method called style-less perturbation (StyLess). Specifically, instead of using a vanilla network as the surrogate model, we advocate using stylized networks, which encode different style features by perturbing an adaptive instance normalization. Our method can prevent adversarial examples from using non-robust style features and help generate transferable perturbations. Comprehensive experiments show that our method can significantly improve the transferability of adversarial examples. Furthermore, our approach is generic and can outperform state-of-the-art transferable attacks when combined with other attack techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 519, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 662 + ], + "type": "text", + "content": "Deep neural networks (DNNs) [14, 24] are currently effective methods for solving various challenging tasks such as computer vision, and natural language processing. Although DNNs have amazing accuracy, especially for computer vision tasks such as image classification, they are also known to be vulnerable to adversarial examples [12, 43]. Adversarial examples are malicious images obtained by adding imperceptible perturbations to benign images. Notably, the transferability of adversarial examples is an intriguing phenomenon, which refers to the property that the same adversarial example can successfully attack different black-box DNNs [5, 30, 34, 51]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 662, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 687 + ], + "type": "text", + "content": "It has been observed that image style can be decoupled from image content, and style transfer techniques allow us" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 541, + 352 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 541, + 352 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 541, + 352 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 541, + 352 + ], + "type": "image", + "image_path": "4e121519470c08256f3fa8e7871280b7004d4e0a1f88e0e784aafcfa05fbe2db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 356, + 383, + 427 + ], + "blocks": [ + { + "bbox": [ + 311, + 356, + 383, + 427 + ], + "lines": [ + { + "bbox": [ + 311, + 356, + 383, + 427 + ], + "spans": [ + { + "bbox": [ + 311, + 356, + 383, + 427 + ], + "type": "image", + "image_path": "b184b7496c37c27594dbd1c96ddb100e8ad8204cc0a9fcb9db059ac1deee6572.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 428, + 372, + 437 + ], + "lines": [ + { + "bbox": [ + 323, + 428, + 372, + 437 + ], + "spans": [ + { + "bbox": [ + 323, + 428, + 372, + 437 + ], + "type": "text", + "content": "Benign Sample" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 391, + 356, + 462, + 427 + ], + "blocks": [ + { + "bbox": [ + 391, + 356, + 462, + 427 + ], + "lines": [ + { + "bbox": [ + 391, + 356, + 462, + 427 + ], + "spans": [ + { + "bbox": [ + 391, + 356, + 462, + 427 + ], + "type": "image", + "image_path": "4c69e622c4e8ff9912cb4e41cdf21686622998cb2e660aea10b2dd4308582521.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 428, + 451, + 437 + ], + "lines": [ + { + "bbox": [ + 400, + 428, + 451, + 437 + ], + "spans": [ + { + "bbox": [ + 400, + 428, + 451, + 437 + ], + "type": "text", + "content": "Stylized Output" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 356, + 541, + 427 + ], + "blocks": [ + { + "bbox": [ + 470, + 356, + 541, + 427 + ], + "lines": [ + { + "bbox": [ + 470, + 356, + 541, + 427 + ], + "spans": [ + { + "bbox": [ + 470, + 356, + 541, + 427 + ], + "type": "image", + "image_path": "5fc1b5901b3f78bfe106d5fd1f94eef84d8c0d87fcd0657060e53eb3de1a844f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 428, + 529, + 437 + ], + "lines": [ + { + "bbox": [ + 481, + 428, + 529, + 437 + ], + "spans": [ + { + "bbox": [ + 481, + 428, + 529, + 437 + ], + "type": "text", + "content": "StyLess Attack" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "text", + "content": "Figure 1. An overview of our StyLess attack. We create stylized model " + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\bar{F}" + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "text", + "content": " by injecting synthesized style features into the surrogate model " + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "inline_equation", + "content": "(F = F_{2} \\circ F_{1})" + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "text", + "content": " using an adaptive IN layer. StyLess reduces the use of non-robust style features in the vanilla surrogate model " + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 306, + 442, + 545, + 498 + ], + "type": "text", + "content": ", ultimately improving attack transferability." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 652 + ], + "type": "text", + "content": "to generate stylized images based on arbitrary style images [18]. Image style refers to the unique visual characteristics of an image, including its colors, textures, and lighting. For instance, two photos of the same object taken by different photographers can have very different styles. Robust DNNs should rely more on content features of data than style features. This inspired us to improve attack transferability from the perspective of avoiding non-robust features. We believe that style features of DNNs are non-robust for building transferable attacks. However, existing attacks do not distinguish between the surrogate models' style and content features, which may reduce attack transferability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "We propose using stylized surrogate models to control style features, which can significantly improve transferability. We refer to the original surrogate model as the \"vanilla model.\" The proposed stylized model is created by adding an adaptive instance normalization (IN) layer to the vanilla" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 58, + 703, + 243, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 243, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 243, + 712 + ], + "type": "text", + "content": "1Our code is available at https://github.com/uhiu/StyLess" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8163" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "model. By adjusting the parameters of the inserted IN layer, we can easily transform the style features of the surrogate models. To compare the stylized and vanilla surrogate models, we analyzed their network losses during optimization. Surprisingly, we found that the adversarial loss of the vanilla model increased much faster than those of the stylized models, resulting in a widening loss gap. This phenomenon reveals that as the attack iteration progresses, current attack methods only focus on maximizing the loss of the vanilla model, leading to increased use of the style features of the vanilla surrogate model. However, we believe that style features are non-robust for transferable attacks, and relying too much on them may reduce attack transferability. To enhance transferability, we aim to limit the use of non-robust style features and close the loss gap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 289, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 466 + ], + "type": "text", + "content": "Based on the above findings, we propose a novel method called StyLess to improve the transferability of adversarial examples. Our method uses multiple synthesized style features to compete with the original style features during the iterative optimization of attack. The process is illustrated in Figure 1. We encode various synthesized style features into a surrogate model via an IN layer to achieve stylized surrogate models. Instead of using only the vanilla surrogate model, we use the gradients of both the stylized surrogate models and the vanilla one to update adversarial examples. The front part of the surrogate model works as a style encoder, and the IN layer simulates synthesized style features. Although we can use a decoder to explicitly generate the final stylized samples, it is unnecessary for the proposed attack method. Experimental results demonstrate that StyLess can enhance the transferability of state-of-the-art adversarial attacks on both unsecured and secured black-box DNNs. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 468, + 287, + 649 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 468, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 468, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 58, + 468, + 287, + 528 + ], + "type": "text", + "content": "- We introduce a novel perspective for interpreting attack transferability: the original style features may hinder transferability. We verify that current iterative attacks increasingly use the style features of the surrogate model during the optimization process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 528, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 528, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 58, + 528, + 287, + 587 + ], + "type": "text", + "content": "- We propose a novel attack called StyLess to enhance transferability by minimizing the use of original style features. To achieve this, we insert an IN layer to create stylized surrogate models and use gradients from both stylized and vanilla models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 588, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 588, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 58, + 588, + 287, + 649 + ], + "type": "text", + "content": "- We conducted comprehensive experiments on various black-box DNNs to demonstrate that StyLess can significantly improve attack transferability. Furthermore, we show that StyLess is a generic approach that can be combined with existing attack techniques." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 134, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "Adversarial Attacks. Adversarial attacks reveal the vulnerability of current DNNs [43]. The classic adversarial attack methods are gradient-based, such as FGSM [12] and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "I-FGSM [25]. C&W [4] considers optimizing the distance between adversarial examples and benign samples, and proposed optimization-based attacks. Adversarial attacks can also be performed in the physical world [10, 39]. As for defending against adversarial examples, adversarial training is a popular defense method that uses adversarial examples as extra training data to improve robustness [32]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 158, + 546, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 546, + 338 + ], + "type": "text", + "content": "Increasing Attack Transferability. An intriguing property of adversarial attacks is the transferability. Ensemble-based attack [30] uses multiple surrogate networks instead of one network. Ghost networks [27] generates different surrogate networks by perturbing skip connection and dropout layers. Optimization methods, such as MI [5], uses a momentum-based optimization, while VT [47] introduces gradient variance to control the stability of the localized gradients. RAP [35] generates adversarial examples located in a flat loss region. Data augmentation methods, such as DI [51], uses image transformation like resizing and padding, while TI [6] considers translating image pixels. SI [29] calculates gradients with the help of several scaled benign samples. Admix [48] calculates iterative gradients by mixing the benign images with randomly sampled images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 546, + 471 + ], + "type": "text", + "content": "Various network architectures and features exhibit different relationships with adversarial attacks. DNNs' linearity is believed to cause adversarial vulnerability [12], and LinBP [13] skips the nonlinear activation during the backpropagation. SGM [49] uses more gradients through skip connections in residual networks. To better leverage the intermediate layers, one can train auxiliary classifiers based on feature spaces [20, 21], maximize the distance between natural images and their adversarial examples in feature spaces [54], or fine-tune the existing adversarial examples in intermediate layer level by ILA [17, 26]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 640 + ], + "type": "text", + "content": "Style Transfer and Instance Normalization. Style transfer can change the style of an image to match the style of another one [9, 11, 23]. Fast feedforward networks can perform stylization with arbitrary styles in a single forward pass [18, 28]. Interestingly, style transfer has a wide range of applications. AdvCam [8] uses natural styles to hide non-" + }, + { + "bbox": [ + 304, + 472, + 546, + 640 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 304, + 472, + 546, + 640 + ], + "type": "text", + "content": " restricted perturbations. FSA [52] generates natural-looking adversarial examples by using optimized style changes. Style transfer has also been used to improve network robustness by exploring additional feature information [33]. Latent style transformations can detect adversarial attacks [46]. AMT-GAN [15] proposes an adversarial makeup transfer to protect facial privacy by preserving stronger black-box transferability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "The family of instance normalization (IN) including batch normalization [22], layer normalization [1], instance normalization [45], and group normalization [50]. Normalizations are mainly used to reduce the covariate shift, and speed model training. Recently, normalizations have been found to be related to robustness. It has been shown that" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8164" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "batch normalization makes DNNs use more non-robust but useful features [3, 19]. AdvBN proposed adding an extra batch normalization into network training to increase training loss adversarially, which enables the network to resist various domain shifts [40]. Adjusting batch normalization statistics such as the running mean and variance in the inference phase, which are estimated during training, improves robustness and defense common corruption [2, 38]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "type": "text", + "content": "Among existing style-based attacks, FSA [52] differentiates style features and content features, which is similar to our method. However, there are three significant differences between FSA and our approach: 1) FSA proposes to hide adversarial perturbations in the optimized style, while we avoid relying on any style. 2) FSA aims at enhancing the natural looking of non-" + }, + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "type": "text", + "content": " restricted attacks, while we focus on the transferability of " + }, + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 46, + 169, + 288, + 324 + ], + "type": "text", + "content": " restricted adversarial examples. 3) Both FSA and our work are inspired by AdaIN [18], but we use IN layer differently. FSA perturbs the IN layer to search malicious styles and requires a decoder. But we use randomized IN layers to augment attacks and don't need to train a decoder." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 336, + 129, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 129, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 129, + 350 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 357, + 134, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 134, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 134, + 369 + ], + "type": "text", + "content": "3.1. Threat Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": "Attack objective. Given a benign image " + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": ", transfer-based attacks aim to generate an adversarial perturbation based on a white-box surrogate network " + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 376, + 287, + 424 + ], + "type": "text", + "content": ". The general attack objective can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 434, + 287, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 434, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 91, + 434, + 287, + 452 + ], + "type": "interline_equation", + "content": "\\max _ {\\delta} \\mathcal {L} (F (x + \\delta), y) \\quad \\text {s . t .} \\quad \\| \\delta \\| \\leq \\epsilon , \\tag {1}", + "image_path": "a8072b89df2822d865053e3b1c6dd9fd6160b3bfd6a37e11c3b75fe2f00dc044.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "content": " denotes the adversarial loss, " + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "content": " is the adversarial perturbation, and " + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 461, + 287, + 486 + ], + "type": "text", + "content": " is the maximum perturbation size." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 487, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 287, + 510 + ], + "type": "text", + "content": "A popular framework to solve the above problem is iterative fast gradient sign method (I-FGSM) [12, 25]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 520, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 520, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 72, + 520, + 287, + 536 + ], + "type": "interline_equation", + "content": "x _ {a d v} ^ {t + 1} = x _ {a d v} ^ {t} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} \\mathcal {L} \\left(F \\left(x _ {a d v} ^ {t}\\right), y\\right)\\right), \\tag {2}", + "image_path": "964507363dc0ca654c20cab52f1885b0ff18deda63a5b4df3de9aad3ef5b4905.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "text", + "content": " is the learning rate, and a clip function will be used on " + }, + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "inline_equation", + "content": "x_{adv}^{t + 1}" + }, + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "text", + "content": " to ensure " + }, + { + "bbox": [ + 46, + 544, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\| x_{adv}^{t + 1} - x\\| \\leq \\epsilon" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "Attacker capability. We follow the same setting in previous work that attackers have a surrogate model and some test samples, but cannot access target models, and don't know network architectures, training data, or defense strategies. It should be noted that our method doesn't require any additional datasets. Our approach involves style features, which can be extracted from an arbitrary image or synthesized without any style image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "Transferable attacks as black-box attacks. Transferable attacks use the surrogate model " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": " to create adversarial examples that can fool unseen target models. In this way, these attacks can be viewed as black-box attacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 380, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 380, + 83 + ], + "type": "text", + "content": "3.2. Motivation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": "Existing transferable attacks often rely on the gradient of the adversarial loss function " + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": " (Equation 2) without considering the impact of different components of " + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": ". However, these approaches have limitations because transferable attacks should minimize the use of non-robust features of the surrogate model. Interestingly, for image classification task, style features of images are typically less robust than content features. Based on this observation, we propose to enhance attack transferability by explicitly reducing the use of style features of the surrogate model within the loss " + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "content": "Our key idea is simulating various surrogate models without the style features of the given vanilla surrogate model. We discovered that inserting an IN layer into the vanilla surrogate model enables us to create new surrogate models that we refer to as stylized surrogate models. Sometimes we omit the word \"surrogate.\" Stylized models can explicitly manipulate style features without compromising model accuracy. Recall that our goal is to construct adversarial examples that can mislead unseen target models, which should include these stylized models. However, existing methods, such as MI and I, only focus on maximizing the loss of the vanilla model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 437 + ], + "type": "text", + "content": "Figure 2 indicates that MI and I(-FGSM) have limited attack transferability on stylized models since the vanilla model's adversarial loss increases much faster than stylized models', resulting in a widening loss gap. In the following sections, we will demonstrate how our method addresses this issue by maximizing the loss of both the stylized and vanilla models, which significantly improves transferability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 444, + 452, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 452, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 452, + 456 + ], + "type": "text", + "content": "3.3. Stylized Surrogate Models" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": "This section will give the definition of our stylized surrogate models, which encode various style features by inserting an IN layer. Then we will analyze the stylized loss gap " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathcal{L})" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": " between the vanilla and stylized surrogate models. Specifically, we will illustrate the increasing " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": " limits transferability and our idea to decrease " + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 304, + 463, + 545, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 540, + 490, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 490, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 490, + 552 + ], + "type": "text", + "content": "3.3.1 Encoding Styles by Stylized Models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 558, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 582 + ], + "type": "text", + "content": "Given a classifier " + }, + { + "bbox": [ + 304, + 558, + 545, + 582 + ], + "type": "inline_equation", + "content": "F = F_{2} \\circ F_{1}" + }, + { + "bbox": [ + 304, + 558, + 545, + 582 + ], + "type": "text", + "content": " as the surrogate model, we define a stylized surrogate model as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 380, + 590, + 545, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 590, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 380, + 590, + 545, + 604 + ], + "type": "interline_equation", + "content": "\\bar {F} _ {x _ {s}} = F _ {2} \\circ \\mathrm {I N} _ {x _ {s}} \\circ F _ {1}, \\tag {3}", + "image_path": "d65446402ba8526b881b3a1344801cc93eba59f40de91c016d8a967b3b1d1a7c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "text", + "content": " is a style input, " + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{IN}_{x_s}" + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "text", + "content": " is an IN layer instantiated by " + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 613, + 544, + 636 + ], + "type": "text", + "content": ". In general, an IN layer is defined as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 350, + 643, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 643, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 350, + 643, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\operatorname {I N} (x; \\mu , \\sigma) = \\sigma \\cdot \\left(\\frac {x - \\mu (x)}{\\sigma (x)}\\right) + \\mu , \\tag {4}", + "image_path": "f7abecebab46a9cbad3756c6e6c001ac85d38b3e484f2618be6ff146da08c578.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are the network parameters of layer IN, and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu(x)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma(x)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are the mean and variance of input " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". According to adaptive instance normalization (AdaIN) [18], to" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8165" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 166, + 173 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 166, + 173 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 166, + 173 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 166, + 173 + ], + "type": "image", + "image_path": "d3a4f1f6154ac9eb82e4d946815110808987b22bd0e05522feb72b5946012cdf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 71, + 276, + 173 + ], + "blocks": [ + { + "bbox": [ + 167, + 71, + 276, + 173 + ], + "lines": [ + { + "bbox": [ + 167, + 71, + 276, + 173 + ], + "spans": [ + { + "bbox": [ + 167, + 71, + 276, + 173 + ], + "type": "image", + "image_path": "57ef0de4d49b9bdbb8839d88f796e1e89860a296ac4f5c1c09c51f49238c7752.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 187, + 167, + 289 + ], + "blocks": [ + { + "bbox": [ + 97, + 176, + 237, + 186 + ], + "lines": [ + { + "bbox": [ + 97, + 176, + 237, + 186 + ], + "spans": [ + { + "bbox": [ + 97, + 176, + 237, + 186 + ], + "type": "text", + "content": "(a) Baseline attack MI vs ours " + }, + { + "bbox": [ + 97, + 176, + 237, + 186 + ], + "type": "inline_equation", + "content": "\\mathrm{MI + StyLess}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 187, + 167, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 187, + 167, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 187, + 167, + 289 + ], + "type": "image", + "image_path": "c00564f5812bb613c7a341be67caeb0ffee7a5d35392dc08f953cc55d0f4f0a6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 288, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 288, + 367 + ], + "type": "text", + "content": "Figure 2. Illustration of the loss gap between vanilla network and stylized network, with RN50 as surrogate model. The greater the loss, the better the attack performance. The widening loss gap in the baseline means that attack performance on stylized models is lagging behind. The proposed StyLess can close the gap." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 167, + 187, + 276, + 288 + ], + "blocks": [ + { + "bbox": [ + 167, + 187, + 276, + 288 + ], + "lines": [ + { + "bbox": [ + 167, + 187, + 276, + 288 + ], + "spans": [ + { + "bbox": [ + 167, + 187, + 276, + 288 + ], + "type": "image", + "image_path": "2684c3128e67515e02360ce3985f115f4c886f044307a715c14891482fd808e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 292, + 246, + 302 + ], + "lines": [ + { + "bbox": [ + 89, + 292, + 246, + 302 + ], + "spans": [ + { + "bbox": [ + 89, + 292, + 246, + 302 + ], + "type": "text", + "content": "(b) Baseline attack I(-FGSM) vs ours I+StyLess" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "type": "text", + "content": "stylize an input " + }, + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "type": "text", + "content": " with a given style input " + }, + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 384, + 287, + 408 + ], + "type": "text", + "content": ", we only need to instantiate the IN as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 95, + 419, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 419, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 95, + 419, + 287, + 435 + ], + "type": "interline_equation", + "content": "\\left. \\operatorname {I N} _ {x _ {s}} (x) = \\operatorname {I N} (x) \\right| _ {\\mu = \\mu \\left(x _ {s}\\right), \\sigma = \\sigma \\left(x _ {s}\\right)} \\tag {5}", + "image_path": "a8f0cd2da457b5020e0fde7d4fac5db84c23283e051fc86e013bc50a9d4db691.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": "Our stylized model " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " has encoded style features. Based on AdaIN, the " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " from the classifier " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "F = F_{2} \\circ F_{1}" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " works as an encoder for style transfer, and given a style input " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": ", we can get a stylized image as " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\bar{x} = D \\circ \\mathrm{IN}_{x_s} \\circ F_1(x)" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " denotes a decoder. Thus, " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{IN}_{x_s} \\circ F_1(x)" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " has encoded the style features of " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}" + }, + { + "bbox": [ + 46, + 444, + 287, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 525, + 251, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 251, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 251, + 538 + ], + "type": "text", + "content": "3.3.2 Stylized Loss Gap Limits transferability" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 571 + ], + "type": "text", + "content": "To validate the performance of adversarial attacks on these stylized surrogate models, we define a stylized loss gap as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 78, + 580, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 580, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 78, + 580, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\Delta \\mathcal {L} = \\mathbb {E} _ {x _ {s} \\in \\mathcal {D}} [ \\mathcal {L} (F (x), y) - \\mathcal {L} (\\bar {F} _ {x _ {s}} (x), y) ], \\tag {6}", + "image_path": "54a6b4d3a41fcbee2c737438b9648e2c2adcb8ab8a0d8ed5296279f5d03b1c0e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}" + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "text", + "content": " are vanilla and stylized model, respectively; " + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 604, + 287, + 628 + ], + "type": "text", + "content": " is style input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "An increasing loss gap " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " can be observed in Figure 2 which limits attack transferability, as explained below. Hypothetically, let's say we can decouple style-dependent loss from content-dependent loss in " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 687, + 287, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 687, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 133, + 687, + 287, + 706 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} ^ {c} + \\mathcal {L} _ {x} ^ {s}, \\\\ \\bar {\\bar {z}} = z ^ {c} - z ^ {s} \\end{array} \\tag {7}", + "image_path": "fb18c470c5632b492252dcbaa305e1888ae72fefe9263319105b78b4b9a876e6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 135, + 704, + 199, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 704, + 199, + 716 + ], + "spans": [ + { + "bbox": [ + 135, + 704, + 199, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} ^ {c} + \\mathcal {L} _ {x _ {s}} ^ {s},", + "image_path": "87dd94d50424f2eb08f0bf578dfb669c90250f9c4ee8b334af84194feb043336.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the vanilla loss, and " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the stylized one; " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^c" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the common content-dependent loss; " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": " is input sample " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": "-specific style-dependent loss, while " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{x_s}^s" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": " is style sample " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "text", + "content": "-specific style-dependent loss. In this case, " + }, + { + "bbox": [ + 306, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L} = \\mathcal{L}_x^s -\\mathcal{L}_{x_s}^s" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "content": "In other words, the non-robust features are related to " + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{x_s}^s" + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "content": " is supposed to be shared by other unseen DNNs. MI and I-(FGSM) have increasing gaps in Figure 2 means that attackers gradually focus on optimizing the " + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "content": " part, which only belonged to the vanilla surrogate model " + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 120, + 545, + 203 + ], + "type": "text", + "content": ". Therefore, the loss gap limits the transferability of adversarial examples in unseen stylized models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": "To decrease " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " and boost transferability, we consider involving " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{x_s}^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " as a competitor of " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " in optimization process to suppress the growth of " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ". In general, we can assume that all these losses in Equation 7 are non-negative, and they satisfy: " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^c\\gg \\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^c\\gg \\mathcal{L}_{x_s}^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ". Also, there is a upper bound " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " for these losses: " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L} < B" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{L}} < B" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ", as the adversarial perturbations are required to smaller than a given " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ". If we only maximize the vanilla loss " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ", both " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^c" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " are likely to be increased. To this end, we propose to maximize " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{x_s\\in \\mathcal{D}}\\mathcal{L}_{x_s}^s +\\mathcal{L}_x^s +\\mathcal{L}^c" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": ", which involves multiple style-dependent losses " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{x_s}^s (x_s\\in \\mathcal{D})" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " to compete with " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_x^s" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": " and leads to a decrease of " + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{L}" + }, + { + "bbox": [ + 305, + 204, + 546, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 354, + 536, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 536, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 536, + 367 + ], + "type": "text", + "content": "3.4. Proposed Style-Less Perturbations (StyLess)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 373, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 409 + ], + "type": "text", + "content": "Based on the above analysis, we propose style-less perturbations (StyLess) method to increase attack transferability by optimizing stylized loss and vanilla loss together:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 416, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 416, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 321, + 416, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\max _ {\\delta} \\mathbb {E} _ {x _ {s} \\in \\mathcal {D}} \\mathcal {L} \\left(\\bar {F} _ {x _ {s}} (x + \\delta), y\\right) + \\mathcal {L} \\left(F (x + \\delta), y\\right). \\tag {8}", + "image_path": "31d3fda9637e149221c31dec3fc646dd6f7eaab2f526f1fb3555967a97217c57.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": "The key to generate multiple stylized models " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": " is synthesizing style statistics " + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mu, \\sigma" + }, + { + "bbox": [ + 305, + 442, + 545, + 491 + ], + "type": "text", + "content": " for Equation 5 to obtain parameterized IN layers. We propose using scaling and interpolation to simulate multiple style features, formulated as" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 370, + 498, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 498, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 370, + 498, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\mu = \\beta (\\lambda \\mu_ {x} + (1 - \\lambda) \\mu_ {s}), \\tag {9}", + "image_path": "3bcdea3b0d718f7f3ba7c029ac8c9753600462eac62335e7c3ad28661df24106.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 372, + 514, + 479, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 514, + 479, + 526 + ], + "spans": [ + { + "bbox": [ + 372, + 514, + 479, + 526 + ], + "type": "interline_equation", + "content": "\\sigma = \\gamma (\\lambda \\sigma_ {x} + (1 - \\lambda) \\sigma_ {s}),", + "image_path": "e4177370a60a7ebcefc90041e3c27dbca24d229d6fbca724d23ca68a08873e99.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mu_x, \\sigma_x" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " is the mean and variance of " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "F_1(x)" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": ", relating to the benign content input " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\mu_s, \\sigma_s" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " are for style input " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "x_s" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " similarly. " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "x_s" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " is an arbitrary image. " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " is a scalar that controls the interpolation of two styles. " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": "-dimensional vectors that scale the synthesized style, where " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " refers to the number of channels in the style feature." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We summarize the proposed StyLess attack in Algorithm 1. To obtain the parameters of the IN layer without training, we select a random image as " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", and generate a pair of " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " using Equation 9. With " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", we can create a stylized model " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}(x)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " using Equation 3 and 5. " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}(x)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " need to be equal to " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "F(x)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " or label " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", otherwise we will regenerate " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " by altering " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". In each iteration, we use " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " stylized models to update adversarial examples based on the proposed objective function in Equation 8." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8166" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 71, + 287, + 317 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": "Algorithm 1 Style-Less Perturbations (StyLess) Algorithm \nInput: Surrogate model " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", benign example " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", iteration number " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", maximum perturbation " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", data augmentation " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", decay factor " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ". Scale factors " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " and interpolation factor " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ". The number of stylized models " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ". \nOutput: An adversarial example " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv} = x" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "g_0 = 0" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon/2" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": ". \n2: for " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "t = 0 \\rightarrow T - 1" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " do: \n3: Augment input " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv} = \\phi(x_{adv})" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n4: Obtain gradient " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{t+1}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n5: repeat \n6: Synthesize a style statistic by Equation 9 \n7: Obtain a stylized model " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " by Equation 3 \n8: Get gradient " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\tilde{g}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n9: Update " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\tilde{g}_{t+1} = \\tilde{g}_{t+1} + \\tilde{g}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n10: until obtain " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " stylized models \n11: Calculate momentum " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "g_{t+1} = \\eta \\cdot g_t + \\tilde{g}_{t+1} / \\| \\tilde{g}_{t+1} \\|_1" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n12: Update example " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv} = x_{adv} + \\alpha \\cdot \\text{sign}(g_{t+1})" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": " \n13: end for \n14: return " + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 47, + 71, + 287, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 338, + 128, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 128, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 128, + 351 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 358, + 163, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 163, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 163, + 371 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 449 + ], + "type": "text", + "content": "Dataset. We use ImageNet [36] for experiments. Specifically, we use 1000 images which are randomly selected from ImageNet validation set by Wang et al. [47]. Similar dataset settings have been widely used in previous work [5, 13, 29, 47, 49, 51]. These images include all categories; almost all are correctly classified by the target DNNs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 652 + ], + "type": "text", + "content": "Models. We evaluate the generated adversarial examples on different black-box DNNs, including both unsecured and secured models. Unsecured models are trained on ImageNet using traditional methods, while secured models are based on adversarial training. The unsecured models include VGG19 [41], AlexNet [24], ResNet50 (RN50) [14], WideResNet101 (WRN101) [53], DenseNet121 (DN121) [16], InceptionV3 (IncV3) [42], MnasNet [44], MobileNetV2 (MNv2) [37], ShuffleNetV2 (SNv2) [31] and ViT [7]. Their pre-training parameters are obtained from PyTorch official. The secured DNNs are IncV3ens3 (ensemble of 3 InceptionV3 networks), IncV3ens4 (ensemble of 4 InceptionV3 networks) and IncResV2ens3 (ensemble of 3 IncResV2 networks). These models were adversarily trained and widely used in previous work [5,30,47,49]. As for the surrogate models, we use VGG19 [41], RN50 [14], WRN101 [53] and DN121 [16]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Implementation Details. We use I-FGSM [12] as the initial baseline. Unless otherwise specified, the attacks are untargeted and " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "-restricted. The maximum perturbation size is set to " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 16 / 255" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". We compare StyLess primarily with six transferable attacks: MI [5], DI [51], TI [6], SI [29]," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "Admix (AI) [48], and an ensemble-based approach [30]. We set the optimization step size to " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon /2" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ", and the number of iterations to " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "T = 50" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ". The momentum decay in MI is " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ". For DI, SI and AI, we follow the official settings described in the corresponding papers. For StyLess, we simulate 10 stylized models in each iteration, denoted by " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ". To generate a stylized model for a given " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ", we randomly sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "[0,0.2]" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "[0,2]" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ", and ensure that " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\bar{F}_{x_s}(x)" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " is equal to " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "F(x)" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " or the real label. The IN layer is inserted after the first bottleneck block for RN50 and WRN101, and after the first dense block for DN121." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 212, + 464, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 464, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 464, + 224 + ], + "type": "text", + "content": "4.2. Attacking Unsecured Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": "We compare StyLess with other attacks on various unsecured DNNs using three surrogate models. Table 1 shows that StyLess is a powerful and generic method that can be combined with existing attack methods to further improve attack transferability. Specifically, we compare StyLess with I, MI, DI, TI, SI and Admix (AI). For the most challenging case in the table, attacking the black-box IncV3 (let's take " + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\mathrm{RN50} \\Rightarrow \\mathrm{IncV3}" + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": " attack as an example), StyLess significantly improves the attack success rate of baseline attacks (I and MI): " + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "inline_equation", + "content": "46.2\\% \\rightarrow 68.3\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": " (I), " + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "inline_equation", + "content": "59.2\\% \\rightarrow 78.9\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": " (MI). StyLess also demonstrates its capabilities when using other DNNs as the surrogate network. For instance, DN121 " + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": " SNv2 attack, StyLess significantly improves the baseline: " + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "inline_equation", + "content": "69.3\\% \\rightarrow 91.4\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": " (I), " + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "inline_equation", + "content": "77.4\\% \\rightarrow 95.1\\%" + }, + { + "bbox": [ + 304, + 231, + 545, + 398 + ], + "type": "text", + "content": " (MI)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": "StyLess can be combined with other attack techniques. Previous work has shown that combining various attack methods results in powerful and transferable attacks. StyLess can be integrated with existing combination-based attacks to enhance attack transferability. We report the results of four combinations of existing attacks: MDI, MTDI, MTDSI, and MTDAI. StyLess further enhances these four attack methods' attack success rate by " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+8.0\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+8.2\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+3.0\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+3.3\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " (in the case of " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{RN}50 \\Rightarrow \\mathrm{IncV}3" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " attack). We evaluate attack performance using various surrogate models, including RN50, WRN101 and DN121. For instance, when combining MTDAI with StyLess, our method enhances the transferability of MTDAI: WRN101 " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " IncV3 attack: " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+4.7\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", WRN101 " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " MNv2 attack: " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", WRN101 " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " SNv2 attack: " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+3.4\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", DN121 " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " IncV3 attack: " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+4.3\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", DN121 " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " MNv2 attack: " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+0.6\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ", DN121 " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": " SNv2 attack: " + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "inline_equation", + "content": "+3.8\\%" + }, + { + "bbox": [ + 304, + 399, + 545, + 615 + ], + "type": "text", + "content": ". The results show that StyLess is an efficient and generic approach for improving attack transferability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 623, + 452, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 452, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 452, + 635 + ], + "type": "text", + "content": "4.3. Attacking Secured Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "We evaluate StyLess on three widely-used secured models: IncV3ens3, IncV3ens4 and IncResV2ens, as shown in Table 2. We present the results of I, MI, MDI, MTDI, and MTDSI on the three secured models using two surrogate models: RN50 and WRN101. These secured networks are more robust than the unsecured DNNs we men" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8167" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 85, + 535, + 308 + ], + "blocks": [ + { + "bbox": [ + 183, + 71, + 409, + 82 + ], + "lines": [ + { + "bbox": [ + 183, + 71, + 409, + 82 + ], + "spans": [ + { + "bbox": [ + 183, + 71, + 409, + 82 + ], + "type": "text", + "content": "Table 1. Attacking unsecured black-box models with StyLess." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 85, + 535, + 308 + ], + "lines": [ + { + "bbox": [ + 61, + 85, + 535, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 535, + 308 + ], + "type": "table", + "html": "
SourceAttackVGG19RN50WRN101DN121IncV3MNv2SNv2
RN50I / +Ours72.8 / 88.8100 / 10080.8 / 97.183.0 / 97.846.2 / 68.377.1 / 91.960.6 / 77.8
MI / +Ours82.9 / 94.1100 / 10083.9 / 97.287.5 / 98.759.2 / 78.983.8 / 93.272.3 / 83.5
MDI / +Ours97.5 / 99.2100 / 10098.2 / 99.899.4 / 10089.5 / 97.598.1 / 99.988.5 / 96.9
MTDI / +Ours98.6 / 99.7100 / 10099.2 / 10099.8 / 10090.2 / 98.498.7 / 10090.8 / 98.1
MTDSI / +Ours98.6 / 99.2100 / 10099.5 / 10099.8 / 10096.2 / 99.298.7 / 99.996.2 / 98.2
MTDAI / +Ours99.3 / 99.6100 / 10099.8 / 10099.9 / 10095.5 / 98.899.7 / 10095.7 / 99.0
WRN101I / +Ours64.8 / 79.688.6 / 98.7100 / 10077.7 / 94.243.2 / 66.668.1 / 84.758.1 / 73.2
MI / +Ours75.6 / 86.489.8 / 98.7100 / 10083.9 / 96.457.6 / 73.776.1 / 87.070.2 / 81.2
MDI / +Ours92.2 / 98.599.0 / 99.9100 / 10097.1 / 99.986.3 / 96.393.6 / 99.086.8 / 95.8
MTDI / +Ours92.5 / 98.699.2 / 100100 / 10097.9 / 99.689.2 / 97.995.4 / 99.388.4 / 97.0
MTDSI / +Ours95.9 / 98.999.6 / 100100 / 10099.7 / 10095.8 / 99.597.7 / 99.994.5 / 98.4
MTDAI / +Ours96.5 / 99.299.9 / 99.9100 / 10099.4 / 10093.7 / 98.498.1 / 99.694.6 / 98.0
DN121I / +Ours79.7 / 97.787.0 / 99.474.7 / 98.0100 / 10055.7 / 88.781.3 / 97.069.3 / 91.4
MI / +Ours86.9 / 99.189.4 / 99.778.2 / 98.6100 / 10066.0 / 95.786.5 / 98.777.4 / 95.1
MDI / +Ours96.7 / 99.998.9 / 10095.4 / 99.8100 / 10090.5 / 99.097.5 / 10090.9 / 98.6
MTDI / +Ours97.5 / 99.999.1 / 10095.9 / 99.8100 / 10092.7 / 99.496.9 / 10092.2 / 98.6
MTDSI / +Ours97.4 / 99.999.3 / 10098.0 / 99.8100 / 10096.5 / 99.898.9 / 10095.1 / 99.1
MTDAI / +Ours99.2 / 99.999.8 / 10098.1 / 99.9100 / 10095.1 / 99.499.4 / 10095.2 / 99.0
", + "image_path": "19ddfe3d92ad4c37566afcc3cf0b6bd4c2ad41660e6360cc020e6952cae06eee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 352, + 282, + 518 + ], + "blocks": [ + { + "bbox": [ + 47, + 324, + 286, + 346 + ], + "lines": [ + { + "bbox": [ + 47, + 324, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 286, + 346 + ], + "type": "text", + "content": "Table 2. Attacking three secured black-box models with StyLess. The surrogate model is RN50 or WRN101." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 352, + 282, + 518 + ], + "lines": [ + { + "bbox": [ + 51, + 352, + 282, + 518 + ], + "spans": [ + { + "bbox": [ + 51, + 352, + 282, + 518 + ], + "type": "table", + "html": "
AttackRN50⇒
IncV3ens3IncV3ens4IncResV2ens
I / +Ours21.6 / 34.618.9 / 32.014.1 / 21.5
MI/ +Ours31.5 / 47.129.3 / 42.420.8 / 31.0
MDI / +Ours59.6 / 78.153.5 / 69.838.3 / 57.3
MTDI / +Ours69.2 / 89.663.8 / 81.854.6 / 72.2
MTDSI / +Ours88.0 / 93.184.7 / 91.377.8 / 84.5
WRN101⇒
I / +Ours23.4 / 40.320.9 / 34.615.5 / 25.6
MI/ +Ours35.0 / 51.230.4 / 46.624.7 / 35.9
MDI / +Ours64.0 / 81.558.8 / 76.246.9 / 66.4
MTDI / +Ours75.5 / 91.470.5 / 87.463.6 / 81.0
MTDSI / +Ours91.6 / 97.588.6 / 96.183.0 / 92.5
", + "image_path": "f656f5acfc0c984e6d62cb435692c9608089a7391f38216d71c6b7b049c4bea1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": "tioned above. Taking " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{RN50} \\Rightarrow \\mathrm{IncV3}_{\\mathrm{ens3}}" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " attack as an example, I, MI, MDI, and MTDI achieve attack success rates of only " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "21.6\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "31.5\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "59.6\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": ", respectively. StyLess effectively improves the performance of these attacks. Notably, MTDSI has been improved by our method: " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "91.6\\% \\rightarrow 97.5\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " in WRN101 " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " IncV3ens3 attack; " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "88.6\\% \\rightarrow" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "96.1\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " in WRN101 " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " IncV3ens4 attack; " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "83.0\\% \\rightarrow 92.5\\%" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " in WRN101 " + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 46, + 540, + 287, + 661 + ], + "type": "text", + "content": " IncResV2ens attack. StyLess demonstrates its great power to enhance transferability in breaking these challenging secured DNNs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Figure 3 shows a comparison of StyLess and LinBP. We consider LinBP and its combination with existing attacks such as ILA, SGM, and MDI on various target DNNs. Our method exhibits the best attack transferability." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 328, + 536, + 459 + ], + "blocks": [ + { + "bbox": [ + 314, + 328, + 536, + 459 + ], + "lines": [ + { + "bbox": [ + 314, + 328, + 536, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 328, + 536, + 459 + ], + "type": "image", + "image_path": "cbf30828713e1c2f8a39b49983f4f7d50063c040818e9006083e276d10e45445.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 470, + 545, + 493 + ], + "lines": [ + { + "bbox": [ + 305, + 470, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 545, + 493 + ], + "type": "text", + "content": "Figure 3. Comparing with LinBP on different black-box models and defenses. The surrogate model is RN50." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 502, + 524, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 524, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 524, + 514 + ], + "type": "text", + "content": "4.4. Combining with Ensemble-Based Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "type": "text", + "content": "Generating adversarial examples based on multiple surrogate networks simultaneously can improve attack performance in practice [30]. This ensemble-based method has been shown to be a powerful attack, and StyLess can further improve it. We use an ensemble of RN50, WRN101, and DN121 as the integrated surrogate model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "We compared our StyLess with ensemble-based MI and MTDI in Table 3. The considerable strength of the ensemble-based method can be seen when comparing with the results in Table 1 and 2. For example, in RN50 " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " IncV3 attack, the baseline attack MI only gets " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "59.2\\%" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", while it gets " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "81.0\\%" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " attack success rate in the case of RN50+WRN101+DN121 " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " IncV3 attack. Noted that we generally use " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 16" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", which is a standard setting. As we can see, when " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 16" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", the ensemble-based MTDI achieves an average attack success rate of less than " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " on the three" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8168" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 85, + 538, + 220 + ], + "blocks": [ + { + "bbox": [ + 192, + 71, + 400, + 82 + ], + "lines": [ + { + "bbox": [ + 192, + 71, + 400, + 82 + ], + "spans": [ + { + "bbox": [ + 192, + 71, + 400, + 82 + ], + "type": "text", + "content": "Table 3. Combining with ensemble-based attack method." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 85, + 538, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 85, + 538, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 85, + 538, + 220 + ], + "type": "table", + "html": "
εAttackRN50+WRN101+DN121⇒
VGG19RN50WRN101DN121IncV3MNv2SNv2IncV3ens3IncV3ens4IncResV2ens
4MI56.896.494.897.235.762.348.318.415.89.4
MTDI79.297.996.698.867.584.571.636.532.122.3
+StyLess90.499.599.499.983.294.385.450.042.929.1
8MI79.299.698.799.458.383.368.031.426.818.2
MTDI95.299.999.599.989.597.590.467.661.846.9
+StyLess99.110010010097.599.697.285.177.466.1
16MI92.310010010081.093.285.053.046.335.0
MTDI99.210010010097.510097.992.786.480.7
+StyLess10010010010099.810010098.597.594.3
", + "image_path": "2fb2777cf91eeb99c67f3b725da9b27aa549a7005849aa042df58899bec4fa2d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "text", + "content": "secured networks: IncV3ens3, IncV3ens4, and IncResV2ens. StyLess can further boost the transferability of ensemble-based MTDI: " + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "inline_equation", + "content": "92.7\\% \\rightarrow 98.5\\%" + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "text", + "content": " for IncV3ens3, " + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "inline_equation", + "content": "86.4\\% \\rightarrow 97.5\\%" + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "text", + "content": " for IncV3ens4, " + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "inline_equation", + "content": "80.7\\% \\rightarrow 94.3\\%" + }, + { + "bbox": [ + 46, + 236, + 286, + 308 + ], + "type": "text", + "content": " for IncResV2ens. These results show that StyLess is a different type of attack, and can work perfectly with the ensemble-based method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": "We also report the experimental results with different " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\epsilon = 4" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": " or 8 to increase the difficulty to attack. For instance, in the case of ensemble-based " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\mathrm{MTDI} \\Rightarrow \\mathrm{IncV3}_{\\mathrm{ens3}}" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": " attack, attack success rate drops from " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "92.7\\%" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "67.6\\%" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\epsilon = 8" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\epsilon = 16" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": ", and StyLess can help ensemble-based MTDI gains " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "+17.5\\%" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "67.6\\% \\rightarrow 85.1\\%" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": "), which is a huge improvement. When " + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\epsilon = 4" + }, + { + "bbox": [ + 46, + 309, + 286, + 428 + ], + "type": "text", + "content": ", the results also demonstrate the advantages of StyLess. This shows that StyLess consistently delivers strong transferability when faced with a more robust network that is difficult to attack." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 254, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 254, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 254, + 444 + ], + "type": "text", + "content": "4.5. Attacking the Google Cloud Vision API" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 533 + ], + "type": "text", + "content": "We use the Google Cloud Vision API as an example of real-world applications to evaluate the transferability of attacks. This API allows us to use various vision features, such as image labeling and face detection. As a black-box model, regular users like us cannot access its network architecture, training data, or defense mechanism. In this section, we will focus on attacking its image labeling feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 605 + ], + "type": "text", + "content": "To utilize the image labeling feature of the API, we need to upload images and obtain the predicted labels. We use 1000 images from ImageNet as the target images, as described in the experimental setup. Due to the fact that the API does not support all 1000 categories in ImageNet, our objective is to mislead the API's original top-1 prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": "We use ResNet50 as the surrogate network to compare the baseline method MTDSI with our method. Experimental results show that MTDSI achieves an attack success rate of " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "75.6\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " on the Google Cloud Vision API. Our approach, MTDSI-SyLess, achieves an attack success rate of " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": ", which is a " + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 714 + ], + "type": "text", + "content": " improvement over the baseline. Our results demonstrate that transfer-based black-box attacks pose a severe threat to real-world applications, and StyLess can effectively boost attack transferability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 236, + 399, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 399, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 399, + 249 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 256, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 316 + ], + "type": "text", + "content": "In this session, we will present four ablation studies: 1) The position of the inserted IN layer; 2) The number of the generated stylized models; 3) The clean losses of stylized models and attack transferability; 4) The most important statistic of style features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 316, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 533 + ], + "type": "text", + "content": "Which position to insert the IN layer? As shown in Figure 4, we use features from different layers of the vanilla surrogate networks as the encoder " + }, + { + "bbox": [ + 304, + 316, + 545, + 533 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 316, + 545, + 533 + ], + "type": "text", + "content": " for style transfer. We evaluate the attack success rates using two surrogate networks: WRN101 and DN121. After injecting synthesized styles into different network layers, we report the attack success rates on various DNNs, including AlexNet, VGG19, RN50, WRN101, and DN121. We observe a trend that the best attack success rates are usually achieved in the shallow layers of the surrogate networks. For example, when using WRN101 as the surrogate network, injecting the synthesized styles in the layers before layer ten is a good choice. Using intermediate layers such as layers 40 to 80 is also acceptable, but the attack performance may be unstable or even worse when injecting the styles in the last few layers. The results of using DN121 as the surrogate network also indicate that the last few layers are the worst choices, and the shallow layers are the optimal." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 536, + 426, + 646 + ], + "blocks": [ + { + "bbox": [ + 307, + 536, + 426, + 646 + ], + "lines": [ + { + "bbox": [ + 307, + 536, + 426, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 426, + 646 + ], + "type": "image", + "image_path": "1ea263319dba8a4b7ae047d615b6e042abcdaacace7e706f94f8cc297766e967.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 645, + 545, + 667 + ], + "lines": [ + { + "bbox": [ + 304, + 645, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 545, + 667 + ], + "type": "text", + "content": "Figure 4. Ablation study on using which network layer to synthesize styles. The surrogate networks are WRN101 and DN121." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 536, + 544, + 646 + ], + "blocks": [ + { + "bbox": [ + 426, + 536, + 544, + 646 + ], + "lines": [ + { + "bbox": [ + 426, + 536, + 544, + 646 + ], + "spans": [ + { + "bbox": [ + 426, + 536, + 544, + 646 + ], + "type": "image", + "image_path": "08b72aef0a42964ce780b7380181ac3a300bd88b3cf4ceb742349c0424ed1186.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "How many stylized models should be created in each iteration? In Figure 5, we vary the number of stylized models generated in each attack iteration from zero to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8169" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "ten. We conduct experiments by combining StyLess with two baseline attacks: MI and DI. The first figure evaluates the attack success rates on three unsecured networks for our " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{MI} + \\mathrm{StyLess}" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " attack. In the second figure, we test " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{DI} + \\mathrm{StyLess}" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " on three secured models. When the number of stylized models is 0, StyLess is not involved, so it is the vanilla baseline attack. As we can see, when the number increases from 0 to 1, the attack success rate starts to grow, which means StyLess starts to work. There is an anomaly when DI begins to combine with StyLess. If the number is less than three, the attack success rate on IncResV2" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "_{ens}" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " is slightly worse than the baseline (around " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "). When the number increases by more than three, the attack success rate becomes higher than the baseline. This may be because IncResV2" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "_{ens}" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " is very strong, and more synthesized style features need to be injected into the surrogate model. According to the figure, StyLess works quite well when six to ten stylized models are used in an attack iteration." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 296, + 168, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 296, + 168, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 296, + 168, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 296, + 168, + 407 + ], + "type": "image", + "image_path": "26c2de9ed8b7a09ab3e7993e4a02aabf49c0486747e0cca07ade759cd9524c1a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 417, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 440 + ], + "type": "text", + "content": "Figure 5. Ablation study on the number of stylized models in an attack iteration. The surrogate network is RN50." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 169, + 296, + 286, + 407 + ], + "blocks": [ + { + "bbox": [ + 169, + 296, + 286, + 407 + ], + "lines": [ + { + "bbox": [ + 169, + 296, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 169, + 296, + 286, + 407 + ], + "type": "image", + "image_path": "95a270d533d13e1510ccc974bb86c4626743fc8fe389ab9642f34e0d39c2373e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 676 + ], + "type": "text", + "content": "How network loss affects attack success rate? In Figure 6, we demonstrate how varying strengths of style injection can affect network loss, which in turn impacts attack performance on a surrogate network (in this case, VGG19). The strength of synthesized style features is denoted by the number of stars, with more stars indicating greater style strength that alters the style features of the surrogate model. Generally, injecting synthesized style features should not significantly affect the clean loss, as an increase in network loss typically leads to a decrease in clean accuracy. Overly corrupted stylized surrogate models can also result in bad gradients for attack methods. Therefore, there is an upper bound on clean loss when generating stylized models. The red line in the figure represents the situation of overly corrupting, in which the clean loss exceeds the estimated bound (indicated by a yellow line in the right figure), and the attack success rate drops significantly. This demonstrates the importance of maintaining relatively good clean accuracy when creating stylized networks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Which statistic of style features matters most? In Figure 7, we compare the effects of the mean and variance of style features on StyLess. Here the interpolation fac" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 427, + 182 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 427, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 427, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 427, + 182 + ], + "type": "image", + "image_path": "dbe6534e33657c4c91dce20b9f2181a5b968269717ac629b442d06b57e96d2dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 192, + 546, + 226 + ], + "lines": [ + { + "bbox": [ + 304, + 192, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 546, + 226 + ], + "type": "text", + "content": "Figure 6. Study how synthesized style features affect the clean loss, which in return impacts the attack success rate. The number of stars indicates the degree of change in original style features." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 545, + 182 + ], + "type": "image", + "image_path": "52dd2e04af075b6d32dd07abb33c26013e3a1709f9def9de45a65c7a6ad60182.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": "tor is " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": ". Equation 9 shows that " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": " involves the mean in IN, while " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": " affects the variance. The attack success rate represents the average success rates of attacks on five DNNs: VGG19, AlexNet, RN50, WRN101, and DN121. The results show that " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": " plays a more important role than " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": ". Specifically, when RN50 is used as the surrogate network, modifying " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": " alone barely improves the baseline attack, while involving " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": " alone enhances the attack success rate by around " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": ". A similar observation can be made for VGG19. From the perspective of gradient calculations, this also makes sense. When we backpropagate through an IN layer, we have " + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\frac{\\partial}{\\partial x} \\mathrm{IN} = \\sigma" + }, + { + "bbox": [ + 304, + 236, + 546, + 392 + ], + "type": "text", + "content": ", which also indicates that the variance matters most for adversarial attacks." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 401, + 523, + 502 + ], + "blocks": [ + { + "bbox": [ + 321, + 401, + 523, + 502 + ], + "lines": [ + { + "bbox": [ + 321, + 401, + 523, + 502 + ], + "spans": [ + { + "bbox": [ + 321, + 401, + 523, + 502 + ], + "type": "image", + "image_path": "59079058ed5ab83c5bc5bccaab88489a161886b76ab926ba5444b02608e316e3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 506, + 516, + 517 + ], + "lines": [ + { + "bbox": [ + 334, + 506, + 516, + 517 + ], + "spans": [ + { + "bbox": [ + 334, + 506, + 516, + 517 + ], + "type": "text", + "content": "Figure 7. The effect of different statistics of style." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 526, + 379, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 379, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 379, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 546, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 712 + ], + "type": "text", + "content": "In this work, we analyze the mechanism of attack transferability in terms of style features. We demonstrate that existing attack methods increasingly use the style features of surrogate models during the iterative optimization, which hampers attack transferability. To address this issue, we propose a novel attack method called StyLess to enhance transferability by reducing reliance on original style features. StyLess uses stylized surrogate models instead of a vanilla surrogate model. Experimental results show that StyLess outperforms existing attacks by a large margin, and can be combined with other attack methods. Notably, StyLess is a different paradigm from previous transferable attack methods, and we hope it will shed light on the interpretation of adversarial attacks in the future." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8170" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 146 + ], + "type": "text", + "content": "[2] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Revisiting batch normalization for improving corruption robustness. In WACV, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 190 + ], + "type": "text", + "content": "[3] Philipp Benz, Chaoning Zhang, and In So Kweon. Batch normalization increases adversarial vulnerability and decreases adversarial transferability: A non-robust feature perspective. In ICCV, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "type": "text", + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy (SP), pages 39-57. IEEE, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 225, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 287, + 256 + ], + "type": "text", + "content": "[5] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In CVPR, 2018. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "text", + "content": "[6] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 292, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 287, + 356 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "text", + "content": "[8] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A. K. Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 413 + ], + "type": "text", + "content": "[9] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In ICLR, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 414, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 414, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 414, + 287, + 457 + ], + "type": "text", + "content": "[10] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In CVPR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 459, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 490 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 522 + ], + "type": "text", + "content": "[12] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1, 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 556 + ], + "type": "text", + "content": "[13] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples. In NeurIPS, 2020. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[15] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In CVPR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "text", + "content": "[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[17] Qian Huang, Isay Katsman, Zeqi Gu, Horace He, Serge J. Belongie, and Ser-Nam Lim. Enhancing adversarial example transferability with an intermediate level attack. In ICCV, 2019. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "text", + "content": "[19] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 182 + ], + "type": "text", + "content": "[20] Nathan Inkawich, Kevin J. Liang, Lawrence Carin, and Yiran Chen. Transferable perturbations of deep feature distributions. In ICLR, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "text", + "content": "[21] Nathan Inkawich, Kevin J. Liang, Binghui Wang, Matthew Inkawich, Lawrence Carin, and Yiran Chen. Perturbing across the feature hierarchy to improve standard and strict blackbox attack transferability. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 259 + ], + "type": "text", + "content": "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[23] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "text", + "content": "[25] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In ICLR (Workshop), 2017. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 361, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 380 + ], + "type": "text", + "content": "[26] Qizhang Li, Yiwen Guo, and Hao Chen. Yet another intermediate-level attack. In ECCV, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "text", + "content": "[27] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning transferable adversarial examples via ghost networks. In AAAI, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 416, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 447 + ], + "type": "text", + "content": "[28] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "text", + "content": "[29] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In ICLR, 2020. 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "text", + "content": "[30] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples and black-box attacks. In ICLR, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 548 + ], + "type": "text", + "content": "[31] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "text", + "content": "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[33] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. Stylized adversarial defense. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[34] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277, 2016. 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[35] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8171" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115:211-252, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 160 + ], + "type": "text", + "content": "[37] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[38] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "text", + "content": "[39] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM Conference on Computer and Communications Security (CCS), pages 1528-1540, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "type": "text", + "content": "[40] Manli Shu, Zuxuan Wu, Micah Goldblum, and Tom Goldstein. Encoding robustness to image style via adversarial feature perturbations. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 325 + ], + "type": "text", + "content": "[41] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 369 + ], + "type": "text", + "content": "[42] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 413 + ], + "type": "text", + "content": "[43] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[44] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 491 + ], + "type": "text", + "content": "[45] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 493, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 288, + 536 + ], + "type": "text", + "content": "[46] Shuo Wang, Surya Nepal, Alsharif Abuadbba, Carsten Rudolph, and Marthie Grobler. Adversarial detection by latent style transformations. IEEE Transactions on Information Forensics and Security (TIFS), 17:1099-1114, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 568 + ], + "type": "text", + "content": "[47] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 601 + ], + "type": "text", + "content": "[48] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In ICCV, 2021. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 645 + ], + "type": "text", + "content": "[49] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip connections matter: On the transferability of adversarial examples generated with resnets. In ICLR, 2020. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 667 + ], + "type": "text", + "content": "[50] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[51] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving transferability of adversarial examples with input diversity. In CVPR, 2019. 1, 2, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 163 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[52] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[53] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 129, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 163 + ], + "type": "text", + "content": "[54] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, 2018. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8172" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_content_list.json b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e564a147939fe9baddb4dd57003ad7c1edfc1507 --- /dev/null +++ b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_content_list.json @@ -0,0 +1,1692 @@ +[ + { + "type": "text", + "text": "Style Projected Clustering for Domain Generalized Semantic Segmentation", + "text_level": 1, + "bbox": [ + 106, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Huang $^{1*}$ Chang Chen $^{2\\dagger}$ Yong Li $^{2}$ Jiacheng Li $^{1}$", + "bbox": [ + 258, + 179, + 707, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Li² Fenglong Song² Youliang Yan² Zhiwei Xiong¹", + "bbox": [ + 225, + 198, + 740, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Science and Technology of China $^{2}$ Huawei Noah's Ark Lab", + "bbox": [ + 173, + 215, + 797, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{weih527,jcee}@mail.ustc.edu.cn,zwxiong@ustc.edu.cn,", + "bbox": [ + 245, + 236, + 720, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenchang25,liyong156,licheng89,songfenglong,yanyouliang}@huawei.com", + "bbox": [ + 176, + 253, + 787, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing semantic segmentation methods improve generalization capability, by regularizing various images to a canonical feature space. While this process contributes to generalization, it weakens the representation inevitably. In contrast to existing methods, we instead utilize the difference between images to build a better representation space, where the distinct style features are extracted and stored as the bases of representation. Then, the generalization to unseen image styles is achieved by projecting features to this known space. Specifically, we realize the style projection as a weighted combination of stored bases, where the similarity distances are adopted as the weighting factors. Based on the same concept, we extend this process to the decision part of model and promote the generalization of semantic prediction. By measuring the similarity distances to semantic bases (i.e., prototypes), we replace the common deterministic prediction with semantic clustering. Comprehensive experiments demonstrate the advantage of proposed method to the state of the art, up to $3.6\\%$ mIoU improvement in average on unseen scenarios. Code and models are available at https://gitee.com/mindspore/models/tree/master/research/cv/SPC-Net.", + "bbox": [ + 76, + 335, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 699, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain generalization methods aim to promote the performance of model (trained on source datasets), when applying it to unseen scenarios (target domains) [9, 19, 29, 36, 62, 74, 75]. Recently, domain generalization for semantic segmentation (DGSS) has attracted increasingly more attention due to the rise of safety-critical applications, such as autonomous driving [3, 12, 22, 45].", + "bbox": [ + 75, + 724, + 468, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing DGSS methods improve the pixel-wise generalization performance by learning domain-agnostic rep", + "bbox": [ + 76, + 832, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86d35739b66d16476b3e032ae64bd899880ffd8e7193c0469049fb93a840fe31.jpg", + "image_caption": [ + "Figure 1. Illustration of instance normalization/whitening (IN/IW) [5, 20, 40] and our proposed style projected clustering method. IN and IW regularize image features from different domains to a canonical space (a-c). Our method builds style and semantic representation spaces based on the data from known domains (d)." + ], + "image_footnote": [], + "bbox": [ + 500, + 301, + 893, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resentations [5, 16, 20, 25, 40, 42, 66, 72]. Researches in this line share the similar goal in general, that is to capture the domain-invariant characteristics of object contents, and eliminates the domain-specific ones (i.e., image styles). As two representatives, Instance Normalization (IN) [56] and Instance Whitening (IW) [17] regularize image features from different domains to a canonical space, as illustrated in Fig. 1(a) and 1(b). Specifically, IN achieves center-level feature alignment via channel-wise feature normalization [33,40], and IW realizes uniform feature distribution by removing linear correlation between channels [5,41]. Moreover, the combination of these two methods is proposed in [42] for a better generalization, as shown in Fig. 1(c).", + "bbox": [ + 496, + 611, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, feature regularization inevitably weakens the representation capability, as a part of feature information is eliminated. Theoretically, it works under a strong assumption that the eliminated information is strictly the domain-specific ones. Yet in practice, the perfect disentanglement between image style and content is difficult to", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during W. Huang's internship at Noah's Ark Lab.", + "bbox": [ + 94, + 875, + 468, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3061", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "achieve. It means that a part of content features will also be eliminated in the process of feature regularization, and thus degrades the segmentation performance.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of seeking common ground by feature regularization, we aim to address DGSS in a different way. In this paper, we propose style projection as an alternative, which utilizes the features from different domains as bases to build a better representation space, as shown in Fig. 1(d). The motivation of style projection comes from a basic concept of generalization, that is to represent unseen data based on the known ones. Specifically, following the common practice, we adopt the statistics (i.e., mean and variance) of features in channel dimension to represent image styles. The image styles from source domains are iteratively extracted and stored as the bases of representation. Then, we project the style of given unseen images into this representation space to promote generalization. This projection process is implemented as a weighted combination of stored style bases, where the similarity distance between styles are adopted as the weighting factors, i.e., $\\lambda_{1}$ and $\\lambda_{2}$ shown in Fig. 1(d).", + "bbox": [ + 75, + 137, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the projected style features, we further devise the decision part of model, which is elaborated for semantic segmentation. Typically, existing methods learn a parametric function to map pixel-wise features to semantic predictions. We replace this deterministic prediction with semantic clustering, where the class of each pixel is predicted by the minimal similarity distance to semantic bases, as shown in Fig. 1(d). Notably, it follows the same concept of style projection, that is to predict unseen data based on the known ones. More concretely, to facilitate the performance of semantic clustering, we propose a variant of contrastive loss to align the semantic bases of same classes and enhance discriminability between different classes.", + "bbox": [ + 75, + 393, + 470, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on single- and multi-source settings to demonstrate the superior generalization of our method over existing DGSS methods. In addition, we visually analyze the effective representation of our proposed method for unseen images in both style and semantic spaces.", + "bbox": [ + 75, + 590, + 468, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions of this paper are summarized as follows:", + "bbox": [ + 96, + 681, + 459, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Beyond existing feature regularization methods, we propose style projected clustering, pointing out a new avenue to address DGSS.", + "- We propose style projection, which projects unseen styles into the style representation space built on known domains for a better representation.", + "- We propose semantic clustering to predict the class of each pixel in unseen images by the similarity distance to semantic bases, which further improves the generalization capability for unseen domains.", + "- Our proposed method outperforms the current state of the arts on multiple DGSS benchmarks." + ], + "bbox": [ + 96, + 705, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain adaptation and generalization. To reduce the burden of pixel-wise annotations on target domains, domain adaptation (DA) technologies are proposed to narrow the domain gap between source and target domains via image translation [14, 24, 37], feature alignment [55, 60, 61], self-training [2, 39, 77] and meta-learning [13, 34] strategies. However, these DA methods require the access of data on target domains. Domain generalization (DG) aims to address a more practical problem where the target domain cannot be accessed. Numerous DG works have been proposed for image classification via style augmentation [19, 59, 68, 75], domain alignment [29, 31], feature disentanglement [27, 44] and meta-learning [9, 26, 28].", + "bbox": [ + 496, + 114, + 890, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization for semantic segmentation. Similar to image classification, DG for semantic segmentation (DGSS) methods are proposed to learn domain-agnostic representations, including style augmentation [16, 25, 43, 72], feature normalization/whitening [5, 40, 42, 66] and meta-learning [20]. To avoid overfitting on source domains, DRPC [72] and FSDR [16] adopt style augmentations in the image space to extend the number of source samples, while WildNet [25] realizes it in the feature space with the aid of ImageNet [8]. Alternatively, normalization and whitening are investigated to achieve distribution alignment between different domains. IBN-Net [40] and RobustNet [5] adopt instance normalization and whitening, respectively, to eliminate the specific style information of each domain. Furthermore, SAN-SAW [42] proposes semantic-aware instance normalization and whitening to enhance the distinguishability between classes. In addition, PintheMem [20] combines the memory-guided network with the meta-learning strategy and obtains competitive performances. Different from these DGSS methods, our method embraces the differences from multiple known domains and takes advantage of their diversity to build a better representation space, realizing the representation of unseen images by the known data.", + "bbox": [ + 496, + 311, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prototype learning. Inspired by the cognitive psychology that human use the knowledge learned in the past to judge the class of unknown things [51,69], prototype-based classification methods have attracted increasing attention, where the class of unknown images is determined by its nearest neighbors in the feature space [7, 10]. Owing to its excellent interpretability and generalizability, prototype learning shows good potential in many fields, such as few-shot learning [1, 52], zero-shot learning [67, 71], unsupervised learning [30, 65]. Recently, prototype learning is also introduced in the dense prediction task, including supervised [76], few-shot [54, 63] and domain adaptive [53, 73] semantic segmentation. To facilitate the learning of prototypes, metric learning [23, 50, 64] is often adopted to pull samples belonging to the same class together and push those of different classes away from each other in the embedding (i.e., feature) space.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3062", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/69d247d644081a0f51d0c428aadc22f240ddd97d5ea3345bb41629a57fd7cc1f.jpg", + "image_caption": [ + "Figure 2. The framework of style projected clustering, which consists of two components, i.e., style projection and semantic clustering. We iteratively extract the style and semantic information of seen domains as style bases $(p_m^\\mu, p_m^\\sigma)$ and semantic bases $p_m^c$ . In style projection, we first calculate the similarity between the unseen style $(\\mu_m, \\sigma_m)$ from the shallow feature $F_m^s$ and style bases $(p_m^\\mu, p_m^\\sigma)$ as weighted factors $\\lambda_m$ . Then, the weighted combination of style bases $(\\mu_m', \\sigma_m')$ is projected on $F_m^n$ to obtain the projected feature $F_m^r$ . In semantic clustering, we calculate the similarity between pixel embeddings in the deep feature $F_m^d$ and semantic bases $p_m^c$ . Then, the class of each pixel is determined by the nearest semantic base. During the training phase, the cross-entropy loss $\\mathcal{L}_{CE}$ , variance loss $\\mathcal{L}_{var}$ and discrimination loss $\\mathcal{L}_{dis}$ are adopted to supervise the learning of style and semantic bases." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to these methods, we adopt the form of prototypes (i.e. bases) to represent semantics. Yet these semantic bases are learned in a different way to facilitate domain generalization, by using a new variant of contrastive loss.", + "bbox": [ + 75, + 454, + 470, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Style Projected Clustering", + "text_level": 1, + "bbox": [ + 76, + 527, + 323, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of our proposed method is depicted in Fig. 2, which consists of two components, i.e., style projection and semantic clustering. In style projection, we project the unseen style into the style representation space built on style bases, according to the similarity between the unseen style and style bases. In semantic clustering, we estimate the similarity between pixel embeddings and semantic bases (i.e., prototypes) to determine the class of pixels in unseen images by the nearest semantic base.", + "bbox": [ + 75, + 551, + 470, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 696, + 282, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the domain generalized semantic segmentation problem, we are given $M$ source domains $\\mathcal{S} = \\{\\mathcal{S}_1,\\mathcal{S}_2,\\dots,\\mathcal{S}_M\\}$ that are from multiple datasets with different data distributions. The $m$ -th source domain $\\mathcal{S}_m$ can be represented as $\\mathcal{S}_m = \\{(x_m,y_m)\\}$ , where $x_{m}\\in \\mathbb{R}^{H\\times W\\times 3}$ is an image from the $m$ -th source domain, $y_{m}\\in \\mathbb{R}^{H\\times W\\times C}$ is the corresponding pixel-wise label, $C$ is the number of semantic classes, $H$ and $W$ are the height and width of the image $x_{m}$ , respectively. In this work, our goal is to train a semantic segmentation model $\\phi$ to obtain the best generalization performance on multiple target domains $\\mathcal{T}$ which cannot be accessed during the training phase.", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Style Projection", + "text_level": 1, + "bbox": [ + 500, + 453, + 660, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The style difference of images is the main factor leading to the domain shift, which limits the generalization ability of the learned model. Pioneering works [11, 18, 40, 75] have demonstrated that the feature distribution shift caused by style differences lies mainly in shallow layers of networks. It also shows that the shallow feature distribution of networks can reflect the style information of the input image $x_{m}$ . Thus, existing works always adopt the channel-wise mean and variance of the shallow feature to represent the style distribution of $x_{m}$ [18, 25]. Following these works, let $F_{m}^{s} \\in \\mathbb{R}^{D \\times H_{s} \\times W_{s}}$ be the shallow feature of $x_{m}$ from the network $\\phi$ , where $D$ denotes the number of channels. The channel-wise mean $\\mu_{m} \\in \\mathbb{R}^{D}$ and variance $\\sigma_{m} \\in \\mathbb{R}^{D}$ of the feature $F_{m}^{s}$ can be calculated as follows:", + "bbox": [ + 496, + 477, + 893, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {m} = \\frac {1}{H _ {s} W _ {s}} \\sum_ {h = 1} ^ {H _ {s}} \\sum_ {w = 1} ^ {W _ {s}} F _ {m} ^ {s}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 694, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {m} = \\sqrt {\\frac {1}{H _ {s} W _ {s}} \\sum_ {h = 1} ^ {H _ {s}} \\sum_ {w = 1} ^ {W _ {s}} (F _ {m} ^ {s} - \\mu_ {m}) ^ {2}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 729, + 846, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To eliminate the specific style information of images, instance normalization [40] is adopted to standardize the feature $F_{m}^{s}$ to a standard distribution (i.e., zeros mean and one standard deviation) as follows:", + "bbox": [ + 500, + 767, + 892, + 827 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {m} ^ {n} = \\frac {F _ {m} ^ {s} - \\mu_ {m}}{\\sigma_ {m} + \\epsilon}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 833, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{m}^{n}$ stands for the normalized feature, and $\\epsilon$ is a small value to avoid division by zero.", + "bbox": [ + 498, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3063", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although instance normalization achieves to remove the specific style information of images, it also eliminates the natural differences between domains, which weakens the representation for target domains and produces limited generalization performance. Therefore, to preserve the specific style information of each domain, we propose style bases $P_{sty} = \\{(p_m^\\mu ,p_m^\\sigma)\\}_{m = 1}^M$ to store the style information of source domains, and then leverage the preserved style bases $P_{sty}$ to build a style representation space, realizing the projection of unseen style, as shown in Fig. 2. Specifically, we first leverage Wasserstein distance [57] to estimate the style distribution discrepancy between the input image $x_{m}$ and the $m$ -th style bases $(p_m^\\mu ,p_m^\\sigma)$ as follows:", + "bbox": [ + 76, + 90, + 472, + 289 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {m} = | | \\mu_ {m} - p _ {m} ^ {\\mu} | | _ {2} ^ {2} + (\\sigma_ {m} ^ {2} + p _ {m} ^ {\\sigma 2} - 2 \\sigma_ {m} p _ {m} ^ {\\sigma}), \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 294, + 468, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{m}$ denotes the distribution distance between the current image $x_{m}$ and the $m$ -th source domain. Then, we use the reciprocal of $d_{m}$ to characterize the similarity between $x_{m}$ and $m$ -th style bases as follows:", + "bbox": [ + 76, + 320, + 468, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {m} = \\frac {\\exp \\left(1 / \\left(1 + d _ {m}\\right)\\right)}{\\sum_ {m = 1} ^ {M} \\exp \\left(1 / \\left(1 + d _ {m}\\right)\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 383, + 468, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the softmax operation is utilized to make the sum of $\\lambda = \\{\\lambda_m|m = 1,2,\\dots,M\\}$ equal to 1. Based on the estimated similarity $\\lambda$ , we can obtain the projected style $(\\mu_{m}^{\\prime},\\sigma_{m}^{\\prime})$ by the weighted sum of style bases as follows:", + "bbox": [ + 76, + 425, + 468, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {m} ^ {\\prime} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} p _ {m} ^ {\\mu}, \\sigma_ {m} ^ {\\prime} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} p _ {m} ^ {\\sigma}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 491, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, following previous works [11, 18, 19, 25], we inject the projected style $(\\mu_{m}^{\\prime},\\sigma_{m}^{\\prime})$ into the normalized feature $F_{m}^{n}$ to obtain the projected feature as follows:", + "bbox": [ + 76, + 537, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {m} ^ {r} = \\sigma_ {m} ^ {\\prime} F _ {m} ^ {n} + \\mu_ {m} ^ {\\prime}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 589, + 468, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the training phase, we adopt the momentum update strategy to achieve the online collection of style information as follows:", + "bbox": [ + 76, + 614, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {m} ^ {\\mu} = \\alpha p _ {m} ^ {\\mu} + (1 - \\alpha) \\mu_ {m}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 662, + 468, + 685 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {m} ^ {\\sigma} = \\alpha p _ {m} ^ {\\sigma} + (1 - \\alpha) \\sigma_ {m},\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 683, + 361, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1]$ is a momentum coefficient. In addition, we randomly initialize $P_{sty}$ to start training, where $p_m^\\mu$ and $p_m^\\sigma$ are initialized with zero-mean and one-mean distribution, respectively. By Eq. 7, we realize the style statistic of source domains and store it as style bases efficiently.", + "bbox": [ + 76, + 704, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After style projection, the projected feature $F_{m}^{r}$ is input into the next layer of the network $\\phi$ . Our style projection is designed as a plug-and-play module that can be applied behind any network layer. However, as the layer is deeper, the style information loosens while the semantic information plays a more important role. Thus, in this work, style projection is only used in the first two layers of $\\phi$ to obtain the best generalization performance.", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic Clustering", + "text_level": 1, + "bbox": [ + 500, + 90, + 692, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the final pixel-wise predictions, we further propose semantic clustering on the deep feature extracted by the network $\\phi$ . Let $F_{m}^{d}\\in \\mathbb{R}^{D\\times H_{d}\\times W_{d}}$ be the deep feature of the input image $x_{m}$ from $\\phi$ . Existing DGSS methods generically apply a learnable segmentation classifier $\\phi_{cls}$ on $F_{m}^{d}$ for the dense prediction. However, the parameters of $\\phi_{cls}$ is learned on the deep features of source domains $S$ , and thus its generalization ability on target domain $\\mathcal{T}$ is limited. In addition, the semantic information between different domains is implicitly encoded in the same parameter space, which causes the specific semantic information of domains to be eliminated.", + "bbox": [ + 496, + 113, + 890, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the concept of style bases, we introduce semantic bases $P_{sem} = \\{p_m^c\\}_{c,m=1}^{C,M}$ to preserve the semantic information of each domain and each class, where $p_m^c \\in \\mathbb{R}^D$ is the cluster center of training pixel embeddings belonging to the $c$ -th class from the $m$ -th source domain in the feature space. Following the prototype theory [7, 10, 76], the class of each pixel embedding $e \\in F_m^d$ can be determined by its nearest semantic bases as follows:", + "bbox": [ + 496, + 295, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc (e) = c ^ {*}, \\text {w i t h} \\left(c ^ {*}, m ^ {*}\\right) = \\operatorname {a r g m i n} _ {c, m} \\left\\{d _ {m} ^ {c} \\right\\} _ {c, m = 1} ^ {C, M}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 424, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_m^c = -\\cos(e, p_m^c)$ is the negative cosine distance used to estimate the similarity between the current embedding $e$ and semantic bases $p_m^c$ . In this work, the pixel embedding $e$ and semantic bases $p_m^c$ are both $l_2$ -normalized. Therefore, the similarity distance can be simply formulated as $d_m^c = -ep_m^c$ . Different from the learnable segmentation classifier $\\phi_{cls}$ , $P_{sem}$ not only explicitly captures characteristic properties of each class from each domain, but also determines the class of pixels in unseen images without introducing extra learnable parameters.", + "bbox": [ + 496, + 450, + 890, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To facilitate the training of the network $\\phi$ during the training phase, we estimate the probability value of pixel embedding $e$ belonging to class $c$ as follows:", + "bbox": [ + 496, + 602, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv (c | e) = \\frac {\\exp (- d ^ {c})}{\\sum_ {c = 1} ^ {C} \\exp (- d ^ {c})}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 652, + 890, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d^{c} = \\min_{m}\\{d_{m}^{c}\\}_{m = 1}^{M}$ denotes the similarity between $e$ and its closet semantic base belonging to class $c$ . Then, we adopt the standard cross-entropy loss to supervise the training of the network $\\phi$ as follows:", + "bbox": [ + 496, + 695, + 890, + 756 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} = - \\frac {1}{H _ {d} W _ {d}} \\sum_ {h = 1} ^ {H _ {d}} \\sum_ {w = 1} ^ {W _ {d}} \\sum_ {c = 1} ^ {C} y _ {m} \\log (v (c | e)), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 762, + 890, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y_{m}$ is the pixel-wise label corresponding to the input image $x_{m}$ .", + "bbox": [ + 496, + 810, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the naive cross-entropy loss only optimizes the relative relations between intra-class and inter-class distance, which ignores the absolute distance constraint between pixel embeddings and semantic bases. That is to say,", + "bbox": [ + 496, + 840, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3064", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we expect that the pixel embedding belonging to class $c$ is closer to the $c$ -th semantic base and is farther away from the semantic bases belonging to other classes. Inspired by metric learning [21, 23], we further propose variance and discrimination terms as two extra training objectives. The former is an intra-class cluster that pulls the pixel embedding $e_{m}^{c}$ belonging to class $c$ from the $m$ -th source domain towards the semantic bases $p_{m}^{c}$ :", + "bbox": [ + 75, + 90, + 472, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v a r} = \\frac {1}{M C} \\sum_ {m = 1} ^ {M} \\sum_ {c = 1} ^ {C} \\left(1 - e _ {m} ^ {c} p _ {m} ^ {c}\\right) ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 220, + 468, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The latter is designed in a contrastive learning way which encourages the current cluster center $\\overline{e}_m^c$ is closer to the $c$ -th semantic bases $p_{c+}$ (i.e., positive keys) and to be far away from semantic bases belonging to other class $p_{c-}$ (i.e., negative keys):", + "bbox": [ + 76, + 268, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} = \\frac {1}{M} \\sum_ {p _ {c +}} - \\log \\frac {\\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c +} / \\tau\\right)}{\\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c +} / \\tau\\right) + \\sum_ {p _ {c -}} \\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c -} / \\tau\\right)}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 352, + 486, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{e}_m^c$ is the cluster center (i.e., mean embedding) of pixel embedding $e_m^c$ in the current feature $F_m^d$ , and $\\tau$ is a temperature hyper-parameter. By Eq. 12, we realize the alignment of semantic bases belonging to the same class $c$ from different domains. Different from existing pixel-wise contrastive learning paradigm [64], the positive and negative samples in Eq. 12 are semantic bases rather than pixel embeddings. Thus, we don't need to construct a memory bank to store sufficient embedding samples, which also significantly reduces the computational cost.", + "bbox": [ + 75, + 415, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve the online collocation of semantic information from source domains, we adopt the same momentum update strategy to update semantic bases $P_{sem}$ as follows:", + "bbox": [ + 76, + 566, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {m} ^ {c} = \\alpha p _ {m} ^ {c} + (1 - \\alpha) \\bar {e} _ {m} ^ {c}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 622, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ the momentum coefficient. Like style bases, we also randomly initialize the semantic bases $P_{sem}$ with zero-mean distribution to start our training.", + "bbox": [ + 76, + 648, + 468, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 703, + 290, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training phase, we combine above three loss terms for the end-to-end training as follows:", + "bbox": [ + 76, + 726, + 468, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\beta \\mathcal {L} _ {\\text {v a r}} + \\gamma \\mathcal {L} _ {\\text {d i s}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 768, + 468, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\gamma$ are weighting coefficients to balance these three terms. For each training iteration, in addition to the parameter update of the network $\\phi$ , the style and semantic bases are also updated online by Eq. 7 and Eq. 13.", + "bbox": [ + 76, + 794, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the inference phase, we leverage Eq. 8 to obtain final pixel-wise predictions by the nonparametric cluster of pixel embeddings outputted from the learned network $\\phi$ .", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 630, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 114, + 601, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic datasets. GTAV [47] contains 24966 images with a resolution of $1914 \\times 1052$ captured from the GTA-V game engine. Synthia [48] contains 9400 images with a resolution of $1280 \\times 760$ generated from virtual urban scenes.", + "bbox": [ + 496, + 137, + 890, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-world datasets. IDD [58] contains 10004 images with an average resolution of $1678 \\times 968$ captured from Indian roads. Cityscapes [6] contains 5000 fine annotated images with a resolution of $2048 \\times 1024$ captured from 50 different cities primarily in Germany. BDD100K [70] contains 10000 image with a resolution of $1280 \\times 720$ captures from different locations in US. Mapillary [38] contains 25000 images with an average resolution of $1920 \\times 1080$ captured from all around the world.", + "bbox": [ + 496, + 199, + 890, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 344, + 718, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the previous work [5], we adopt DeepLabV3+ [4] with ResNet-50, ResNet-101 [15], MobileNetV2 [49] and ShuffleNetV2 [35] backbones as our segmentation networks, where all backbones are pre-trained on ImageNet [8]. During the training phase, we adopt the SGD optimizer [46] with a momentum of 0.9 and weight decay of $5e - 4$ . The initial learning rate is set to 0.01 and is decreased using the polynomial scheduling with a power of 0.9. We train all models for 40K iterations, except for the three-source setting, the model is trained for 100K iterations. In addition to some common data augmentations used in [5], we adopt extra strong style augmentations to enrich the style information of urban-scene images [32], which aims to enhance the proposed style projection ability in networks. More details can be found in our supplementary materials.", + "bbox": [ + 496, + 368, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 500, + 604, + 591, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison methods. We extensively compare our proposed method against existing DGSS methods, which can be classified into three groups, including style augmentation (WildNet [25]), feature normalization/whitening (IBNNet [40], RobustNet [5] and SAN-SAW [42]), and meta-learning (MLDG [26] and PintheMem [20]). Since SAN-SAW [42] and WildNet [25] are only implemented on the single-source setting in their paper, we reproduce them on the multi-source setting to make a comparison. In particular, WildNet [25] utilizes the external dataset (i.e., ImageNet) to extend the style and content information of source domains. Thus, we re-implement it by replacing the external dataset with the source dataset for a fair comparison, which is marked with * in our tables.", + "bbox": [ + 496, + 628, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-source setting. To demonstrate the effectiveness of our proposed method, we first conduct contrast experiments on the multi-source DGSS setting, where multiple source domains can be efficiently used to build a diverse", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3065", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/70204edbfc3f84e6bb1cd57dbf4376eba88795412d97f9709b7dbffda5a1b339.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPublicationCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaAvg.-SAvg.-A
\\( Baseline^{\\dagger} \\)-35.4625.0931.9430.8368.4867.9968.2445.79
\\( IBN-Net^{\\dagger} [40] \\)ECCV 201835.5532.1838.0935.2769.7266.9068.3148.49
\\( RobustNet^{\\dagger} [5] \\)CVPR 202137.6934.0938.4936.7668.2668.7768.5249.46
\\( Baseline^{\\ddagger} \\)-33.4229.0732.1931.5669.6363.9366.7845.65
\\( MLDG^{\\ddagger} [26] \\)AAAI 201838.8431.9535.6035.4664.6151.6958.1544.54
\\( PintheMem^{\\ddagger} [20] \\)CVPR 202244.5138.0742.7041.7665.8554.4960.1749.12
Baseline-36.0328.1532.6132.2669.3067.6168.4646.65
SAN-SAW [42]CVPR 202242.1337.7442.9140.9363.9862.5863.2849.87
WildNet [25]CVPR 202243.6539.9043.2842.2868.0563.9866.0251.77
WildNet* [25]CVPR 202239.3334.7641.0638.3869.7062.1165.9149.39
Ours-46.3643.1848.2345.9272.4674.8773.6757.02
", + "bbox": [ + 89, + 88, + 877, + 282 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c93484a4b1a68749defa8875755f73956d4cf48a01bc75abb54eeeb75d6cb3a.jpg", + "table_caption": [ + "Table 1. Source $(\\mathbf{G} + \\mathbf{S})\\rightarrow$ Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) datasets. The best and second best results are highlighted and underlined. Avg.-T, Avg.-S and Avg.-A denote the average results on target, source and all domains, respectively. Results with the $\\dagger$ and $\\ddagger$ sign are from [5] and [20], respectively. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodsCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaIDDAvg.-SAvg.-A
Baseline‡52.5147.4754.7051.5670.3167.1371.5669.6760.61
IBN-Net‡[40]54.3948.9156.0653.1270.7363.6871.0268.4860.80
RobustNet‡[5]54.7049.0056.9053.5370.0666.4071.0269.1661.35
MLDG‡[26]54.7648.5255.9453.0769.5359.7967.7365.6859.38
PintheMem‡[20]56.5750.1858.3155.0269.9962.9967.5866.8560.94
Baseline54.1646.2455.5751.9968.3565.1270.0767.8559.92
SAN-SAW [42]54.8946.5056.3852.5964.4964.7666.3765.2158.90
WildNet [25]55.5850.3157.9354.6167.6561.3570.0766.3660.48
WildNet* [25]53.6148.9256.1852.9070.9859.6964.5265.0658.98
Ours57.9153.2661.6157.5974.6478.3576.0776.3566.97
", + "bbox": [ + 109, + 371, + 857, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Source $(\\mathbf{G} + \\mathbf{S} + \\mathbf{I})\\rightarrow$ Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) and one real (IDD) datasets. Results with the $\\ddagger$ sign are from [20].", + "bbox": [ + 75, + 549, + 892, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representation space. As listed in Table 1, we quantitatively compare our results with existing DGSS methods on both target and source datasets, where all networks with the ResNet-50 backbone are trained with two synthetic datasets (i.e., GTAV and Synthia). Remarkably, compared with the state-of-the-art method (i.e., WildNet [20]), our method not only shows superior generalization capability on target datasets (up to $3.6\\%$ mIoU in average), but also significantly improve the performance on source datasets (up to $7.6\\%$ mIoU), which demonstrates our method can enhance the representation ability of the learned model on both source and target domains. Furthermore, We provide visual prediction results for qualitative comparisons as shown in Fig. 3. Our method obtains the best visual results on different target datasets. Following [20], we add one real dataset (i.e., IDD) to source domains to further verify the superiority of our method on more source datasets. As listed in Table 2, our method also outperforms existing methods on both source and target domains by a large margin.", + "bbox": [ + 75, + 597, + 472, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-source setting. We further implement our method in", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the single-source setting to make a comprehensive comparison, where all network with the ResNet-50 backbone are trained with one synthetic (i.e., GTAV) dataset. As listed in Table 3, our method shows superior generalization performances over existing DGSS methods. Compared with the naive baseline, our method brings approximately $14\\%$ mIoU gains in average on target datasets.", + "bbox": [ + 496, + 597, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different backbones. To demonstrate the wide applicability of our method, we compare our results with classic DGSS methods (i.e., IBN-Net [40] and RobustNet [5]) with different backbones. As listed in Table 4, our method shows superior performances on both large (i.e., ResNet-101) and lightweight (i.e., MobileNet and ShuffleNet) backbones.", + "bbox": [ + 496, + 703, + 893, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 801, + 665, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies with the ResNet-50 backbone on two source domains (i.e., GTAV and Synthia) as following.", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposed strategies. As listed in Table 5, our method shows the best generalization capability when two strategies", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3066", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9539866d645f4702f1a17dda1aef83bff39617833c8facdb157c2d84ccdc47fc.jpg", + "image_caption": [ + "Figure 3. Source $(\\mathbf{G} + \\mathbf{S}) \\rightarrow$ Target (C, B, M): Visualization comparison with existing DGSS methods on three different target domains." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a0027777b31df1a31f11c978b22119e47806ddea99264a1b2387a30e836a1f5d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Source $(\\mathbf{G})\\rightarrow$ Target (C, B, M): Mean IoU(\\%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with the one synthetic (GTAV) dataset. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison." + ], + "table_body": "
MethodsCBMAvg.-T
Baseline28.9525.1428.1827.42
IBN-Net [40]33.8532.3037.7534.63
RobustNet [5]36.5835.2040.3337.37
Baseline31.6026.7029.0029.10
MLDG [26]36.7032.1032.2033.67
PintheMem [20]41.0034.6037.4037.67
Baseline29.3225.7128.3327.79
SAN-SAW [42]39.7537.3441.8639.65
Baseline35.1629.7131.2932.05
WildNet [25]44.6238.4246.0943.04
Baseline32.0126.0429.3529.13
WildNet* [25]40.1034.8239.3838.10
Ours44.1040.4645.5143.36
", + "bbox": [ + 99, + 309, + 444, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are adopted at the same time. Remarkably, compared with the first and second lines, we can find that style projection can approximately bring $12\\%$ mIoU gains in average over the baseline, which fully demonstrates its effectiveness for the generalization on unseen domains.", + "bbox": [ + 75, + 652, + 468, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different ways of style projection. As listed in Table 6, we investigate the effect of different ways of style projection. There are two intuitive ways as follows. One way is using the naive instance normalization to project images from different domains into a normalized feature space (i.e., Normalization). The other way is using the extracted style bases to directly substitute the unseen style (i.e., Substitution). We can find that the weighted combination of style bases can effectively enhance the representation of unseen style, producing better generalization on unseen domains.", + "bbox": [ + 75, + 731, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss terms. As listed in Table 7, we conduct ablation ex", + "bbox": [ + 76, + 885, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3e5188acc96ae35d43fba4736c114d340b55743491bba08f1640f64fadf70ae.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Source $(\\mathbf{G} + \\mathbf{S})\\rightarrow$ Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods with different backbones." + ], + "table_body": "
MethodsCBMAvg.-T
MobileNetBaseline29.1620.2727.1925.24
IBN-Net [40]29.5826.0226.3227.31
RobustNet [5]30.6725.0228.2727.99
Ours39.8834.8338.9137.87
ShuffleNetBaseline29.4826.2731.3529.03
IBN-Net [40]32.6129.5533.2031.79
RobustNet [5]33.1531.9834.8533.33
Ours38.9734.6239.6637.75
ResNet-101Baseline34.7129.3237.7433.92
IBN-Net [40]39.1834.0039.3237.50
RobustNet [5]39.9634.9441.7238.87
Ours47.9343.6248.7946.78
", + "bbox": [ + 514, + 309, + 874, + 513 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c26501ff9bd3ea4e32f0f00c44d1cfb190a880090b6320e8dfad3c233c51b8c0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Ablation results for each strategy used in our method. Sty.-Pro. and Sem.-Clu. indicate style projection and semantic clustering, respectively." + ], + "table_body": "
Sty.-Pro.Sem.-Clu.CBMAvg.-T
36.0328.1532.6132.26
✓44.8742.4246.3744.55
✓39.0130.6035.1934.93
✓✓46.3643.1848.2345.92
", + "bbox": [ + 504, + 564, + 883, + 648 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/99def1dc6a8d64ff02aad51bddff06c5ec8c5f50697e0d57a0016169f965e663.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Ablation results for different ways of style projection." + ], + "table_body": "
MethodsCBMAvg.-T
Normalization43.8340.9544.9243.23
Substitution45.0042.7945.1644.32
Ours46.3643.1848.2345.92
", + "bbox": [ + 527, + 712, + 859, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "periments to demonstrate the effectiveness of two complementary loss functions in Eq. 11 and Eq. 12. Compared with the naive cross-entropy loss, adding any complementary loss can bring the performance gain, which verifies each of them can effectively supplement the main loss $\\mathcal{L}_{CE}$ .", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3067", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a23371a55bed0d7a7f6a040083ad6d6529c0bf35f9e1a0c23e08662f69ea4a17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LCELvarLdisCBMAvg.-T
✓44.0041.8245.9743.93
✓✓45.5742.7846.6144.99
✓✓45.9242.4247.0845.14
✓✓✓46.3643.1848.2345.92
", + "bbox": [ + 84, + 88, + 459, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/24f86139da956e1c32af99905a3d9b662ed4f93055faf75a58adec75d58e9753.jpg", + "table_caption": [ + "Table 7. Ablation results for each loss term." + ], + "table_footnote": [ + "Table 8. Comparison of computational cost. Tested with the image size of $2048 \\times 1024$ on one NVIDIA Tesla V100 GPU. We average the inference time over 500 trials." + ], + "table_body": "
Methods# of ParamsGFLOPsTime (ms)
Baseline45.08M277.777.82
IBN-Net [40]45.08M277.828.74
RobustNet [5]45.08M277.789.48
MLDG [26]45.08M277.779.67
PintheMem [20]45.28M278.3111.64
SAN-SAW [42]25.63M421.8657.58
WildNet [25]45.21M277.168.61
Ours45.22M286.099.98
", + "bbox": [ + 89, + 199, + 452, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Analysis", + "text_level": 1, + "bbox": [ + 76, + 402, + 302, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distribution analysis. We adopt the t-SNE visualization tool to analyze the effectiveness of our proposed style projection and semantic clustering strategies. As shown in Fig 4, we show the variations of style distribution between different domains before and after style projection. We can find that the style distribution of different domains is well separated before style projection (Fig. 4(a)), while their style distribution is approximately constrained between two style bases after style projection (Fig. 4(b)), which demonstrates style projection successfully projects unseen styles into the style representation space built on style bases.", + "bbox": [ + 75, + 430, + 468, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we visualize the semantic distribution between different classes and domains as shown in Fig 5. From Fig. 5(a), we can find that pixel samples belonging to the same class are well clustered while those belonging to different classes are well separated. In addition, the preserved semantic bases are approximately located in the cluster center of pixel samples. From Fig. 5(b), we can find that these pixel samples from different domains are well clustered according to their classes, which demonstrates our semantic clustering successfully achieves the class prediction between different domains by the preserved semantic bases.", + "bbox": [ + 75, + 598, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complexity of networks. As listed in Table 8, we compare the number of parameters and computational cost with existing DGSS methods. Since we need to store style and semantic bases and estimate the similarity between them and unseen images, the number of parameters and computational cost in our method are slightly higher than the naive baseline. However, our inference time is competitive to exiting DGSS methods due to the efficient implementation of distance measures by matrix multiplications.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bff090d503a3576984a0bd43676e2de63e23e6b6491b7d787cafdb091bd9b6bd.jpg", + "image_caption": [ + "(a) Before projection" + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 694, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e68ce15f23383c76af465946d99195077ea4ba8eee8c000c6de99c98af946612.jpg", + "image_caption": [ + "(b) After projection" + ], + "image_footnote": [], + "bbox": [ + 699, + 88, + 887, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7682256312cef08236365d3fb7e0c670a5ebd0dd3fe73e394d9c1faa8f8a79a.jpg", + "image_caption": [ + "Figure 4. t-SNE visualization of style statistics between different domains before (a) and after (b) style projection, where the style statistics (concatenation of mean and variance) is computed from the first layer's feature map of the ResNet-50 trained on two synthetic datasets. Triangles indicate the preserved style bases.", + "(a) Class distribution", + "Figure 5. t-SNE visualization of semantic statistics between different classes (a) and domains (b), where the semantic statistics is computed from the last layer's feature map. Triangles indicate the preserved semantic bases." + ], + "image_footnote": [], + "bbox": [ + 506, + 338, + 696, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea1420d302b5320496147ee66b87378e9d5514a36061bad6bc4db5eb194418c4.jpg", + "image_caption": [ + "(b) Domain distribution" + ], + "image_footnote": [], + "bbox": [ + 699, + 339, + 888, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 617, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel style projected clustering method for domain generalized semantic segmentation, which achieves the style and semantic representation of unseen images based on known data. In particular, style projection projects arbitrary unseen styles into the style representation space of source domains and achieves the retention of specific style information between different domains. Semantic clustering predicts the class of each pixel by the minimal similarity distance to semantic bases, which realizes the semantic representation for unseen images and promotes the generalization ability. Through the evaluation on multiple urban-scene datasets, we demonstrate the superior generalization performance of our proposed method over existing DGSS methods.", + "bbox": [ + 496, + 606, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. And we gratefully acknowledge the support of MindSpore (https://www.mindspore.cn/).", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3068", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In ICML, 2019. 2", + "[2] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, 2021. 2", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1", + "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 5", + "[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 1, 2, 5, 6, 7, 8", + "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5", + "[7] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1):21-27, 1967. 2, 4", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5", + "[9] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 1, 2", + "[10] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):417-435, 2012. 2, 4", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 3, 4", + "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1", + "[13] Xiaoqing Guo, Chen Yang, Baopu Li, and Yixuan Yuan. Metacorrection: Domain-aware meta loss correction for unsupervised domain adaptation in semantic segmentation. In CVPR, 2021. 2", + "[14] Jianzhong He, Xu Jia, Shuaijun Chen, and Jianzhuang Liu. Multi-source domain adaptation with collaborative learning for semantic segmentation. In CVPR, 2021. 2", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + ], + "bbox": [ + 78, + 116, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 1, 2", + "[17] Lei Huang, Dawei Yang, Bo Lang, and Jia Deng. Decorated batch normalization. In CVPR, 2018. 1", + "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3, 4", + "[19] Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. Style neophile: Constantly seeking novel styles for domain generalization. In CVPR, 2022. 1, 2, 4", + "[20] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8", + "[21] Shu Kong and Charless C Fowlkes. Recurrent pixel embedding for instance grouping. In CVPR, 2018. 5", + "[22] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In WACV, 2021. 1", + "[23] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2, 5", + "[24] Seunghun Lee, Sunghyun Cho, and Sunghoon Im. Dranet: Disentangling representation and adaptation networks for unsupervised cross-domain adaptation. In CVPR, 2021. 2", + "[25] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[26] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2018. 2, 5, 6, 7, 8", + "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, 2017. 2", + "[28] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In ICCV, 2019. 2", + "[29] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 1, 2", + "[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2020. 2", + "[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2", + "[32] Road Augmentation Library. https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library.5", + "[33] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable learning-to-normalize via switchable normalization. In ICLR, 2018. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Xinyu Luo, Jiaming Zhang, Kailun Yang, Alina Roitberg, Kunyu Peng, and Rainer Stiefelhagen. Towards robust semantic segmentation of accident scenes via multi-source mixed sampling and meta-learning. In CVPR, 2022. 2", + "[35] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5", + "[36] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, 2013. 1", + "[37] Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In CVPR, 2018. 2", + "[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5", + "[39] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 2", + "[40] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 1, 2, 3, 5, 6, 7, 8", + "[41] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 1", + "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8", + "[43] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2", + "[44] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, 2020. 2", + "[45] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 1", + "[46] Tiago Ramalho and Marta Garnelo. Adaptive posterior learning: few-shot learning with a surprise-based memory module. In ICLR, 2018. 5", + "[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 5", + "[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 5", + "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5", + "[50] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015. 2", + "[51] Herbert A Simon and Allen Newell. Human problem solving: The state of the theory in 1970. American Psychologist, 26(2):145, 1971. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2", + "[53] Haitao Tian, Shiru Qu, and Pierre Payeur. A prototypical knowledge oriented adaptation framework for semantic segmentation. IEEE Transactions on Image Processing, 31:149-163, 2021. 2", + "[54] Zhuotao Tian, Xin Lai, Li Jiang, Shu Liu, Michelle Shu, Hengshuang Zhao, and Jiaya Jia. Generalized few-shot semantic segmentation. In CVPR, 2022. 2", + "[55] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018. 2", + "[56] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In CVPR, 2017. 1", + "[57] SS Vallender. Calculation of the wasserstein distance between probability distributions on the line. Theory of Probability & Its Applications, 18(4):784-786, 1974. 4", + "[58] Girish Varma, Anbumani Subramanian, Anoop Namboodiri, Manmohan Chandraker, and CV Jawahar. Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In WACV, 2019. 5", + "[59] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In NeurIPS, 2018. 2", + "[60] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 2", + "[61] Haoran Wang, Tong Shen, Wei Zhang, Ling-Yu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In ECCV, 2020. 2", + "[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1", + "[63] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 2", + "[64] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In ICCV, 2021. 2, 5", + "[65] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2", + "[66] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 1, 2", + "[67] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. In NeurIPS, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3070", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In ICLR, 2021. 2", + "[69] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 2", + "[70] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5", + "[71] Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Episode-based prototype generating network for zero-shot learning. In CVPR, 2020. 2", + "[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 1, 2", + "[73] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 2", + "[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1", + "[75] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2020. 1, 2, 3", + "[76] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 2, 4", + "[77] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 2" + ], + "bbox": [ + 78, + 90, + 467, + 601 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_model.json b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef4480092d38f5d2ca26be4c9b1e6c4b154d5d56 --- /dev/null +++ b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Style Projected Clustering for Domain Generalized Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.18, + 0.709, + 0.199 + ], + "angle": 0, + "content": "Wei Huang\\(^{1*}\\) Chang Chen\\(^{2\\dagger}\\) Yong Li\\(^{2}\\) Jiacheng Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.199, + 0.741, + 0.217 + ], + "angle": 0, + "content": "Cheng Li² Fenglong Song² Youliang Yan² Zhiwei Xiong¹" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.217, + 0.798, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Science and Technology of China \\(^{2}\\)Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.237, + 0.721, + 0.252 + ], + "angle": 0, + "content": "{weih527,jcee}@mail.ustc.edu.cn,zwxiong@ustc.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.255, + 0.788, + 0.27 + ], + "angle": 0, + "content": "{chenchang25,liyong156,licheng89,songfenglong,yanyouliang}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.474, + 0.671 + ], + "angle": 0, + "content": "Existing semantic segmentation methods improve generalization capability, by regularizing various images to a canonical feature space. While this process contributes to generalization, it weakens the representation inevitably. In contrast to existing methods, we instead utilize the difference between images to build a better representation space, where the distinct style features are extracted and stored as the bases of representation. Then, the generalization to unseen image styles is achieved by projecting features to this known space. Specifically, we realize the style projection as a weighted combination of stored bases, where the similarity distances are adopted as the weighting factors. Based on the same concept, we extend this process to the decision part of model and promote the generalization of semantic prediction. By measuring the similarity distances to semantic bases (i.e., prototypes), we replace the common deterministic prediction with semantic clustering. Comprehensive experiments demonstrate the advantage of proposed method to the state of the art, up to \\(3.6\\%\\) mIoU improvement in average on unseen scenarios. Code and models are available at https://gitee.com/mindspore/models/tree/master/research/cv/SPC-Net." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.7, + 0.21, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.832 + ], + "angle": 0, + "content": "Domain generalization methods aim to promote the performance of model (trained on source datasets), when applying it to unseen scenarios (target domains) [9, 19, 29, 36, 62, 74, 75]. Recently, domain generalization for semantic segmentation (DGSS) has attracted increasingly more attention due to the rise of safety-critical applications, such as autonomous driving [3, 12, 22, 45]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.833, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Existing DGSS methods improve the pixel-wise generalization performance by learning domain-agnostic rep" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.302, + 0.895, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.528, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Illustration of instance normalization/whitening (IN/IW) [5, 20, 40] and our proposed style projected clustering method. IN and IW regularize image features from different domains to a canonical space (a-c). Our method builds style and semantic representation spaces based on the data from known domains (d)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.809 + ], + "angle": 0, + "content": "resentations [5, 16, 20, 25, 40, 42, 66, 72]. Researches in this line share the similar goal in general, that is to capture the domain-invariant characteristics of object contents, and eliminates the domain-specific ones (i.e., image styles). As two representatives, Instance Normalization (IN) [56] and Instance Whitening (IW) [17] regularize image features from different domains to a canonical space, as illustrated in Fig. 1(a) and 1(b). Specifically, IN achieves center-level feature alignment via channel-wise feature normalization [33,40], and IW realizes uniform feature distribution by removing linear correlation between channels [5,41]. Moreover, the combination of these two methods is proposed in [42] for a better generalization, as shown in Fig. 1(c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Nevertheless, feature regularization inevitably weakens the representation capability, as a part of feature information is eliminated. Theoretically, it works under a strong assumption that the eliminated information is strictly the domain-specific ones. Yet in practice, the perfect disentanglement between image style and content is difficult to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.469, + 0.888 + ], + "angle": 0, + "content": "*This work was done during W. Huang's internship at Noah's Ark Lab." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3061" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "achieve. It means that a part of content features will also be eliminated in the process of feature regularization, and thus degrades the segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.394 + ], + "angle": 0, + "content": "Instead of seeking common ground by feature regularization, we aim to address DGSS in a different way. In this paper, we propose style projection as an alternative, which utilizes the features from different domains as bases to build a better representation space, as shown in Fig. 1(d). The motivation of style projection comes from a basic concept of generalization, that is to represent unseen data based on the known ones. Specifically, following the common practice, we adopt the statistics (i.e., mean and variance) of features in channel dimension to represent image styles. The image styles from source domains are iteratively extracted and stored as the bases of representation. Then, we project the style of given unseen images into this representation space to promote generalization. This projection process is implemented as a weighted combination of stored style bases, where the similarity distance between styles are adopted as the weighting factors, i.e., \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) shown in Fig. 1(d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.59 + ], + "angle": 0, + "content": "Based on the projected style features, we further devise the decision part of model, which is elaborated for semantic segmentation. Typically, existing methods learn a parametric function to map pixel-wise features to semantic predictions. We replace this deterministic prediction with semantic clustering, where the class of each pixel is predicted by the minimal similarity distance to semantic bases, as shown in Fig. 1(d). Notably, it follows the same concept of style projection, that is to predict unseen data based on the known ones. More concretely, to facilitate the performance of semantic clustering, we propose a variant of contrastive loss to align the semantic bases of same classes and enhance discriminability between different classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.68 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on single- and multi-source settings to demonstrate the superior generalization of our method over existing DGSS methods. In addition, we visually analyze the effective representation of our proposed method for unseen images in both style and semantic spaces." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.682, + 0.46, + 0.697 + ], + "angle": 0, + "content": "Contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.707, + 0.468, + 0.751 + ], + "angle": 0, + "content": "- Beyond existing feature regularization methods, we propose style projected clustering, pointing out a new avenue to address DGSS." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.757, + 0.468, + 0.801 + ], + "angle": 0, + "content": "- We propose style projection, which projects unseen styles into the style representation space built on known domains for a better representation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.807, + 0.468, + 0.866 + ], + "angle": 0, + "content": "- We propose semantic clustering to predict the class of each pixel in unseen images by the similarity distance to semantic bases, which further improves the generalization capability for unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Our proposed method outperforms the current state of the arts on multiple DGSS benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.707, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.312 + ], + "angle": 0, + "content": "Domain adaptation and generalization. To reduce the burden of pixel-wise annotations on target domains, domain adaptation (DA) technologies are proposed to narrow the domain gap between source and target domains via image translation [14, 24, 37], feature alignment [55, 60, 61], self-training [2, 39, 77] and meta-learning [13, 34] strategies. However, these DA methods require the access of data on target domains. Domain generalization (DG) aims to address a more practical problem where the target domain cannot be accessed. Numerous DG works have been proposed for image classification via style augmentation [19, 59, 68, 75], domain alignment [29, 31], feature disentanglement [27, 44] and meta-learning [9, 26, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Domain generalization for semantic segmentation. Similar to image classification, DG for semantic segmentation (DGSS) methods are proposed to learn domain-agnostic representations, including style augmentation [16, 25, 43, 72], feature normalization/whitening [5, 40, 42, 66] and meta-learning [20]. To avoid overfitting on source domains, DRPC [72] and FSDR [16] adopt style augmentations in the image space to extend the number of source samples, while WildNet [25] realizes it in the feature space with the aid of ImageNet [8]. Alternatively, normalization and whitening are investigated to achieve distribution alignment between different domains. IBN-Net [40] and RobustNet [5] adopt instance normalization and whitening, respectively, to eliminate the specific style information of each domain. Furthermore, SAN-SAW [42] proposes semantic-aware instance normalization and whitening to enhance the distinguishability between classes. In addition, PintheMem [20] combines the memory-guided network with the meta-learning strategy and obtains competitive performances. Different from these DGSS methods, our method embraces the differences from multiple known domains and takes advantage of their diversity to build a better representation space, realizing the representation of unseen images by the known data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Prototype learning. Inspired by the cognitive psychology that human use the knowledge learned in the past to judge the class of unknown things [51,69], prototype-based classification methods have attracted increasing attention, where the class of unknown images is determined by its nearest neighbors in the feature space [7, 10]. Owing to its excellent interpretability and generalizability, prototype learning shows good potential in many fields, such as few-shot learning [1, 52], zero-shot learning [67, 71], unsupervised learning [30, 65]. Recently, prototype learning is also introduced in the dense prediction task, including supervised [76], few-shot [54, 63] and domain adaptive [53, 73] semantic segmentation. To facilitate the learning of prototypes, metric learning [23, 50, 64] is often adopted to pull samples belonging to the same class together and push those of different classes away from each other in the embedding (i.e., feature) space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.342, + 0.897, + 0.442 + ], + "angle": 0, + "content": "Figure 2. The framework of style projected clustering, which consists of two components, i.e., style projection and semantic clustering. We iteratively extract the style and semantic information of seen domains as style bases \\((p_m^\\mu, p_m^\\sigma)\\) and semantic bases \\(p_m^c\\). In style projection, we first calculate the similarity between the unseen style \\((\\mu_m, \\sigma_m)\\) from the shallow feature \\(F_m^s\\) and style bases \\((p_m^\\mu, p_m^\\sigma)\\) as weighted factors \\(\\lambda_m\\). Then, the weighted combination of style bases \\((\\mu_m', \\sigma_m')\\) is projected on \\(F_m^n\\) to obtain the projected feature \\(F_m^r\\). In semantic clustering, we calculate the similarity between pixel embeddings in the deep feature \\(F_m^d\\) and semantic bases \\(p_m^c\\). Then, the class of each pixel is determined by the nearest semantic base. During the training phase, the cross-entropy loss \\(\\mathcal{L}_{CE}\\), variance loss \\(\\mathcal{L}_{var}\\) and discrimination loss \\(\\mathcal{L}_{dis}\\) are adopted to supervise the learning of style and semantic bases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.471, + 0.519 + ], + "angle": 0, + "content": "Similar to these methods, we adopt the form of prototypes (i.e. bases) to represent semantics. Yet these semantic bases are learned in a different way to facilitate domain generalization, by using a new variant of contrastive loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.325, + 0.548 + ], + "angle": 0, + "content": "3. Style Projected Clustering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.471, + 0.691 + ], + "angle": 0, + "content": "The overall architecture of our proposed method is depicted in Fig. 2, which consists of two components, i.e., style projection and semantic clustering. In style projection, we project the unseen style into the style representation space built on style bases, according to the similarity between the unseen style and style bases. In semantic clustering, we estimate the similarity between pixel embeddings and semantic bases (i.e., prototypes) to determine the class of pixels in unseen images by the nearest semantic base." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.283, + 0.713 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In the domain generalized semantic segmentation problem, we are given \\(M\\) source domains \\(\\mathcal{S} = \\{\\mathcal{S}_1,\\mathcal{S}_2,\\dots,\\mathcal{S}_M\\}\\) that are from multiple datasets with different data distributions. The \\(m\\)-th source domain \\(\\mathcal{S}_m\\) can be represented as \\(\\mathcal{S}_m = \\{(x_m,y_m)\\}\\), where \\(x_{m}\\in \\mathbb{R}^{H\\times W\\times 3}\\) is an image from the \\(m\\)-th source domain, \\(y_{m}\\in \\mathbb{R}^{H\\times W\\times C}\\) is the corresponding pixel-wise label, \\(C\\) is the number of semantic classes, \\(H\\) and \\(W\\) are the height and width of the image \\(x_{m}\\), respectively. In this work, our goal is to train a semantic segmentation model \\(\\phi\\) to obtain the best generalization performance on multiple target domains \\(\\mathcal{T}\\) which cannot be accessed during the training phase." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.454, + 0.661, + 0.472 + ], + "angle": 0, + "content": "3.2. Style Projection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.691 + ], + "angle": 0, + "content": "The style difference of images is the main factor leading to the domain shift, which limits the generalization ability of the learned model. Pioneering works [11, 18, 40, 75] have demonstrated that the feature distribution shift caused by style differences lies mainly in shallow layers of networks. It also shows that the shallow feature distribution of networks can reflect the style information of the input image \\( x_{m} \\). Thus, existing works always adopt the channel-wise mean and variance of the shallow feature to represent the style distribution of \\( x_{m} \\) [18, 25]. Following these works, let \\( F_{m}^{s} \\in \\mathbb{R}^{D \\times H_{s} \\times W_{s}} \\) be the shallow feature of \\( x_{m} \\) from the network \\( \\phi \\), where \\( D \\) denotes the number of channels. The channel-wise mean \\( \\mu_{m} \\in \\mathbb{R}^{D} \\) and variance \\( \\sigma_{m} \\in \\mathbb{R}^{D} \\) of the feature \\( F_{m}^{s} \\) can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.695, + 0.891, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mu_ {m} = \\frac {1}{H _ {s} W _ {s}} \\sum_ {h = 1} ^ {H _ {s}} \\sum_ {w = 1} ^ {W _ {s}} F _ {m} ^ {s}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.731, + 0.848, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {m} = \\sqrt {\\frac {1}{H _ {s} W _ {s}} \\sum_ {h = 1} ^ {H _ {s}} \\sum_ {w = 1} ^ {W _ {s}} (F _ {m} ^ {s} - \\mu_ {m}) ^ {2}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.768, + 0.893, + 0.828 + ], + "angle": 0, + "content": "To eliminate the specific style information of images, instance normalization [40] is adopted to standardize the feature \\( F_{m}^{s} \\) to a standard distribution (i.e., zeros mean and one standard deviation) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.834, + 0.892, + 0.866 + ], + "angle": 0, + "content": "\\[\nF _ {m} ^ {n} = \\frac {F _ {m} ^ {s} - \\mu_ {m}}{\\sigma_ {m} + \\epsilon}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(F_{m}^{n}\\) stands for the normalized feature, and \\(\\epsilon\\) is a small value to avoid division by zero." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3063" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.29 + ], + "angle": 0, + "content": "Although instance normalization achieves to remove the specific style information of images, it also eliminates the natural differences between domains, which weakens the representation for target domains and produces limited generalization performance. Therefore, to preserve the specific style information of each domain, we propose style bases \\( P_{sty} = \\{(p_m^\\mu ,p_m^\\sigma)\\}_{m = 1}^M \\) to store the style information of source domains, and then leverage the preserved style bases \\( P_{sty} \\) to build a style representation space, realizing the projection of unseen style, as shown in Fig. 2. Specifically, we first leverage Wasserstein distance [57] to estimate the style distribution discrepancy between the input image \\( x_{m} \\) and the \\( m \\)-th style bases \\( (p_m^\\mu ,p_m^\\sigma) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.295, + 0.47, + 0.314 + ], + "angle": 0, + "content": "\\[\nd _ {m} = | | \\mu_ {m} - p _ {m} ^ {\\mu} | | _ {2} ^ {2} + (\\sigma_ {m} ^ {2} + p _ {m} ^ {\\sigma 2} - 2 \\sigma_ {m} p _ {m} ^ {\\sigma}), \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.47, + 0.381 + ], + "angle": 0, + "content": "where \\( d_{m} \\) denotes the distribution distance between the current image \\( x_{m} \\) and the \\( m \\)-th source domain. Then, we use the reciprocal of \\( d_{m} \\) to characterize the similarity between \\( x_{m} \\) and \\( m \\)-th style bases as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.385, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {m} = \\frac {\\exp \\left(1 / \\left(1 + d _ {m}\\right)\\right)}{\\sum_ {m = 1} ^ {M} \\exp \\left(1 / \\left(1 + d _ {m}\\right)\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.47, + 0.487 + ], + "angle": 0, + "content": "where the softmax operation is utilized to make the sum of \\(\\lambda = \\{\\lambda_m|m = 1,2,\\dots,M\\}\\) equal to 1. Based on the estimated similarity \\(\\lambda\\) , we can obtain the projected style \\((\\mu_{m}^{\\prime},\\sigma_{m}^{\\prime})\\) by the weighted sum of style bases as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.492, + 0.47, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mu_ {m} ^ {\\prime} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} p _ {m} ^ {\\mu}, \\sigma_ {m} ^ {\\prime} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} p _ {m} ^ {\\sigma}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.538, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Finally, following previous works [11, 18, 19, 25], we inject the projected style \\((\\mu_{m}^{\\prime},\\sigma_{m}^{\\prime})\\) into the normalized feature \\(F_{m}^{n}\\) to obtain the projected feature as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.59, + 0.469, + 0.609 + ], + "angle": 0, + "content": "\\[\nF _ {m} ^ {r} = \\sigma_ {m} ^ {\\prime} F _ {m} ^ {n} + \\mu_ {m} ^ {\\prime}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.615, + 0.469, + 0.658 + ], + "angle": 0, + "content": "During the training phase, we adopt the momentum update strategy to achieve the online collection of style information as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.664, + 0.469, + 0.686 + ], + "angle": 0, + "content": "\\[\np _ {m} ^ {\\mu} = \\alpha p _ {m} ^ {\\mu} + (1 - \\alpha) \\mu_ {m}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.684, + 0.362, + 0.7 + ], + "angle": 0, + "content": "\\[\np _ {m} ^ {\\sigma} = \\alpha p _ {m} ^ {\\sigma} + (1 - \\alpha) \\sigma_ {m},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.469, + 0.78 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1]\\) is a momentum coefficient. In addition, we randomly initialize \\(P_{sty}\\) to start training, where \\(p_m^\\mu\\) and \\(p_m^\\sigma\\) are initialized with zero-mean and one-mean distribution, respectively. By Eq. 7, we realize the style statistic of source domains and store it as style bases efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "After style projection, the projected feature \\( F_{m}^{r} \\) is input into the next layer of the network \\( \\phi \\). Our style projection is designed as a plug-and-play module that can be applied behind any network layer. However, as the layer is deeper, the style information loosens while the semantic information plays a more important role. Thus, in this work, style projection is only used in the first two layers of \\( \\phi \\) to obtain the best generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.694, + 0.108 + ], + "angle": 0, + "content": "3.3. Semantic Clustering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.295 + ], + "angle": 0, + "content": "To obtain the final pixel-wise predictions, we further propose semantic clustering on the deep feature extracted by the network \\(\\phi\\). Let \\(F_{m}^{d}\\in \\mathbb{R}^{D\\times H_{d}\\times W_{d}}\\) be the deep feature of the input image \\(x_{m}\\) from \\(\\phi\\). Existing DGSS methods generically apply a learnable segmentation classifier \\(\\phi_{cls}\\) on \\(F_{m}^{d}\\) for the dense prediction. However, the parameters of \\(\\phi_{cls}\\) is learned on the deep features of source domains \\(S\\), and thus its generalization ability on target domain \\(\\mathcal{T}\\) is limited. In addition, the semantic information between different domains is implicitly encoded in the same parameter space, which causes the specific semantic information of domains to be eliminated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Based on the concept of style bases, we introduce semantic bases \\( P_{sem} = \\{p_m^c\\}_{c,m=1}^{C,M} \\) to preserve the semantic information of each domain and each class, where \\( p_m^c \\in \\mathbb{R}^D \\) is the cluster center of training pixel embeddings belonging to the \\( c \\)-th class from the \\( m \\)-th source domain in the feature space. Following the prototype theory [7, 10, 76], the class of each pixel embedding \\( e \\in F_m^d \\) can be determined by its nearest semantic bases as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.425, + 0.892, + 0.446 + ], + "angle": 0, + "content": "\\[\nc (e) = c ^ {*}, \\text {w i t h} \\left(c ^ {*}, m ^ {*}\\right) = \\operatorname {a r g m i n} _ {c, m} \\left\\{d _ {m} ^ {c} \\right\\} _ {c, m = 1} ^ {C, M}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.603 + ], + "angle": 0, + "content": "where \\( d_m^c = -\\cos(e, p_m^c) \\) is the negative cosine distance used to estimate the similarity between the current embedding \\( e \\) and semantic bases \\( p_m^c \\). In this work, the pixel embedding \\( e \\) and semantic bases \\( p_m^c \\) are both \\( l_2 \\)-normalized. Therefore, the similarity distance can be simply formulated as \\( d_m^c = -ep_m^c \\). Different from the learnable segmentation classifier \\( \\phi_{cls} \\), \\( P_{sem} \\) not only explicitly captures characteristic properties of each class from each domain, but also determines the class of pixels in unseen images without introducing extra learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.649 + ], + "angle": 0, + "content": "To facilitate the training of the network \\(\\phi\\) during the training phase, we estimate the probability value of pixel embedding \\(e\\) belonging to class \\(c\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.654, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\nv (c | e) = \\frac {\\exp (- d ^ {c})}{\\sum_ {c = 1} ^ {C} \\exp (- d ^ {c})}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.757 + ], + "angle": 0, + "content": "where \\(d^{c} = \\min_{m}\\{d_{m}^{c}\\}_{m = 1}^{M}\\) denotes the similarity between \\(e\\) and its closet semantic base belonging to class \\(c\\). Then, we adopt the standard cross-entropy loss to supervise the training of the network \\(\\phi\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.763, + 0.892, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} = - \\frac {1}{H _ {d} W _ {d}} \\sum_ {h = 1} ^ {H _ {d}} \\sum_ {w = 1} ^ {W _ {d}} \\sum_ {c = 1} ^ {C} y _ {m} \\log (v (c | e)), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.841 + ], + "angle": 0, + "content": "where \\( y_{m} \\) is the pixel-wise label corresponding to the input image \\( x_{m} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.902 + ], + "angle": 0, + "content": "However, the naive cross-entropy loss only optimizes the relative relations between intra-class and inter-class distance, which ignores the absolute distance constraint between pixel embeddings and semantic bases. That is to say," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.214 + ], + "angle": 0, + "content": "we expect that the pixel embedding belonging to class \\( c \\) is closer to the \\( c \\)-th semantic base and is farther away from the semantic bases belonging to other classes. Inspired by metric learning [21, 23], we further propose variance and discrimination terms as two extra training objectives. The former is an intra-class cluster that pulls the pixel embedding \\( e_{m}^{c} \\) belonging to class \\( c \\) from the \\( m \\)-th source domain towards the semantic bases \\( p_{m}^{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.221, + 0.47, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v a r} = \\frac {1}{M C} \\sum_ {m = 1} ^ {M} \\sum_ {c = 1} ^ {C} \\left(1 - e _ {m} ^ {c} p _ {m} ^ {c}\\right) ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.27, + 0.47, + 0.346 + ], + "angle": 0, + "content": "The latter is designed in a contrastive learning way which encourages the current cluster center \\(\\overline{e}_m^c\\) is closer to the \\(c\\)-th semantic bases \\(p_{c+}\\) (i.e., positive keys) and to be far away from semantic bases belonging to other class \\(p_{c-}\\) (i.e., negative keys):" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.353, + 0.488, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} = \\frac {1}{M} \\sum_ {p _ {c +}} - \\log \\frac {\\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c +} / \\tau\\right)}{\\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c +} / \\tau\\right) + \\sum_ {p _ {c -}} \\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c -} / \\tau\\right)}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.566 + ], + "angle": 0, + "content": "where \\(\\bar{e}_m^c\\) is the cluster center (i.e., mean embedding) of pixel embedding \\(e_m^c\\) in the current feature \\(F_m^d\\), and \\(\\tau\\) is a temperature hyper-parameter. By Eq. 12, we realize the alignment of semantic bases belonging to the same class \\(c\\) from different domains. Different from existing pixel-wise contrastive learning paradigm [64], the positive and negative samples in Eq. 12 are semantic bases rather than pixel embeddings. Thus, we don't need to construct a memory bank to store sufficient embedding samples, which also significantly reduces the computational cost." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.567, + 0.47, + 0.613 + ], + "angle": 0, + "content": "To achieve the online collocation of semantic information from source domains, we adopt the same momentum update strategy to update semantic bases \\( P_{sem} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.623, + 0.469, + 0.641 + ], + "angle": 0, + "content": "\\[\np _ {m} ^ {c} = \\alpha p _ {m} ^ {c} + (1 - \\alpha) \\bar {e} _ {m} ^ {c}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.65, + 0.47, + 0.696 + ], + "angle": 0, + "content": "where \\(\\alpha\\) the momentum coefficient. Like style bases, we also randomly initialize the semantic bases \\(P_{sem}\\) with zero-mean distribution to start our training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.704, + 0.292, + 0.72 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.47, + 0.757 + ], + "angle": 0, + "content": "During the training phase, we combine above three loss terms for the end-to-end training as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.769, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\beta \\mathcal {L} _ {\\text {v a r}} + \\gamma \\mathcal {L} _ {\\text {d i s}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.856 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\gamma\\) are weighting coefficients to balance these three terms. For each training iteration, in addition to the parameter update of the network \\(\\phi\\), the style and semantic bases are also updated online by Eq. 7 and Eq. 13." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "During the inference phase, we leverage Eq. 8 to obtain final pixel-wise predictions by the nonparametric cluster of pixel embeddings outputted from the learned network \\(\\phi\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.632, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.602, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.199 + ], + "angle": 0, + "content": "Synthetic datasets. GTAV [47] contains 24966 images with a resolution of \\(1914 \\times 1052\\) captured from the GTA-V game engine. Synthia [48] contains 9400 images with a resolution of \\(1280 \\times 760\\) generated from virtual urban scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Real-world datasets. IDD [58] contains 10004 images with an average resolution of \\(1678 \\times 968\\) captured from Indian roads. Cityscapes [6] contains 5000 fine annotated images with a resolution of \\(2048 \\times 1024\\) captured from 50 different cities primarily in Germany. BDD100K [70] contains 10000 image with a resolution of \\(1280 \\times 720\\) captures from different locations in US. Mapillary [38] contains 25000 images with an average resolution of \\(1920 \\times 1080\\) captured from all around the world." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.719, + 0.361 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Following the previous work [5], we adopt DeepLabV3+ [4] with ResNet-50, ResNet-101 [15], MobileNetV2 [49] and ShuffleNetV2 [35] backbones as our segmentation networks, where all backbones are pre-trained on ImageNet [8]. During the training phase, we adopt the SGD optimizer [46] with a momentum of 0.9 and weight decay of \\(5e - 4\\). The initial learning rate is set to 0.01 and is decreased using the polynomial scheduling with a power of 0.9. We train all models for 40K iterations, except for the three-source setting, the model is trained for 100K iterations. In addition to some common data augmentations used in [5], we adopt extra strong style augmentations to enrich the style information of urban-scene images [32], which aims to enhance the proposed style projection ability in networks. More details can be found in our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.593, + 0.62 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Comparison methods. We extensively compare our proposed method against existing DGSS methods, which can be classified into three groups, including style augmentation (WildNet [25]), feature normalization/whitening (IBNNet [40], RobustNet [5] and SAN-SAW [42]), and meta-learning (MLDG [26] and PintheMem [20]). Since SAN-SAW [42] and WildNet [25] are only implemented on the single-source setting in their paper, we reproduce them on the multi-source setting to make a comparison. In particular, WildNet [25] utilizes the external dataset (i.e., ImageNet) to extend the style and content information of source domains. Thus, we re-implement it by replacing the external dataset with the source dataset for a fair comparison, which is marked with * in our tables." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-source setting. To demonstrate the effectiveness of our proposed method, we first conduct contrast experiments on the multi-source DGSS setting, where multiple source domains can be efficiently used to build a diverse" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3065" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.089, + 0.878, + 0.284 + ], + "angle": 0, + "content": "
MethodsPublicationCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaAvg.-SAvg.-A
\\( Baseline^{\\dagger} \\)-35.4625.0931.9430.8368.4867.9968.2445.79
\\( IBN-Net^{\\dagger} [40] \\)ECCV 201835.5532.1838.0935.2769.7266.9068.3148.49
\\( RobustNet^{\\dagger} [5] \\)CVPR 202137.6934.0938.4936.7668.2668.7768.5249.46
\\( Baseline^{\\ddagger} \\)-33.4229.0732.1931.5669.6363.9366.7845.65
\\( MLDG^{\\ddagger} [26] \\)AAAI 201838.8431.9535.6035.4664.6151.6958.1544.54
\\( PintheMem^{\\ddagger} [20] \\)CVPR 202244.5138.0742.7041.7665.8554.4960.1749.12
Baseline-36.0328.1532.6132.2669.3067.6168.4646.65
SAN-SAW [42]CVPR 202242.1337.7442.9140.9363.9862.5863.2849.87
WildNet [25]CVPR 202243.6539.9043.2842.2868.0563.9866.0251.77
WildNet* [25]CVPR 202239.3334.7641.0638.3869.7062.1165.9149.39
Ours-46.3643.1848.2345.9272.4674.8773.6757.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.288, + 0.895, + 0.359 + ], + "angle": 0, + "content": "Table 1. Source \\((\\mathbf{G} + \\mathbf{S})\\rightarrow\\) Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) datasets. The best and second best results are highlighted and underlined. Avg.-T, Avg.-S and Avg.-A denote the average results on target, source and all domains, respectively. Results with the \\(\\dagger\\) and \\(\\ddagger\\) sign are from [5] and [20], respectively. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.372, + 0.859, + 0.546 + ], + "angle": 0, + "content": "
MethodsCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaIDDAvg.-SAvg.-A
Baseline‡52.5147.4754.7051.5670.3167.1371.5669.6760.61
IBN-Net‡[40]54.3948.9156.0653.1270.7363.6871.0268.4860.80
RobustNet‡[5]54.7049.0056.9053.5370.0666.4071.0269.1661.35
MLDG‡[26]54.7648.5255.9453.0769.5359.7967.7365.6859.38
PintheMem‡[20]56.5750.1858.3155.0269.9962.9967.5866.8560.94
Baseline54.1646.2455.5751.9968.3565.1270.0767.8559.92
SAN-SAW [42]54.8946.5056.3852.5964.4964.7666.3765.2158.90
WildNet [25]55.5850.3157.9354.6167.6561.3570.0766.3660.48
WildNet* [25]53.6148.9256.1852.9070.9859.6964.5265.0658.98
Ours57.9153.2661.6157.5974.6478.3576.0776.3566.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.55, + 0.893, + 0.581 + ], + "angle": 0, + "content": "Table 2. Source \\((\\mathbf{G} + \\mathbf{S} + \\mathbf{I})\\rightarrow\\) Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) and one real (IDD) datasets. Results with the \\(\\ddagger\\) sign are from [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.473, + 0.884 + ], + "angle": 0, + "content": "representation space. As listed in Table 1, we quantitatively compare our results with existing DGSS methods on both target and source datasets, where all networks with the ResNet-50 backbone are trained with two synthetic datasets (i.e., GTAV and Synthia). Remarkably, compared with the state-of-the-art method (i.e., WildNet [20]), our method not only shows superior generalization capability on target datasets (up to \\(3.6\\%\\) mIoU in average), but also significantly improve the performance on source datasets (up to \\(7.6\\%\\) mIoU), which demonstrates our method can enhance the representation ability of the learned model on both source and target domains. Furthermore, We provide visual prediction results for qualitative comparisons as shown in Fig. 3. Our method obtains the best visual results on different target datasets. Following [20], we add one real dataset (i.e., IDD) to source domains to further verify the superiority of our method on more source datasets. As listed in Table 2, our method also outperforms existing methods on both source and target domains by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Single-source setting. We further implement our method in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.703 + ], + "angle": 0, + "content": "the single-source setting to make a comprehensive comparison, where all network with the ResNet-50 backbone are trained with one synthetic (i.e., GTAV) dataset. As listed in Table 3, our method shows superior generalization performances over existing DGSS methods. Compared with the naive baseline, our method brings approximately \\(14\\%\\) mIoU gains in average on target datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Different backbones. To demonstrate the wide applicability of our method, we compare our results with classic DGSS methods (i.e., IBN-Net [40] and RobustNet [5]) with different backbones. As listed in Table 4, our method shows superior performances on both large (i.e., ResNet-101) and lightweight (i.e., MobileNet and ShuffleNet) backbones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.666, + 0.817 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.871 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies with the ResNet-50 backbone on two source domains (i.e., GTAV and Synthia) as following." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Proposed strategies. As listed in Table 5, our method shows the best generalization capability when two strategies" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.279, + 0.885, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Source \\((\\mathbf{G} + \\mathbf{S}) \\rightarrow\\) Target (C, B, M): Visualization comparison with existing DGSS methods on three different target domains." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.31, + 0.445, + 0.542 + ], + "angle": 0, + "content": "
MethodsCBMAvg.-T
Baseline28.9525.1428.1827.42
IBN-Net [40]33.8532.3037.7534.63
RobustNet [5]36.5835.2040.3337.37
Baseline31.6026.7029.0029.10
MLDG [26]36.7032.1032.2033.67
PintheMem [20]41.0034.6037.4037.67
Baseline29.3225.7128.3327.79
SAN-SAW [42]39.7537.3441.8639.65
Baseline35.1629.7131.2932.05
WildNet [25]44.6238.4246.0943.04
Baseline32.0126.0429.3529.13
WildNet* [25]40.1034.8239.3838.10
Ours44.1040.4645.5143.36
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.63 + ], + "angle": 0, + "content": "Table 3. Source \\((\\mathbf{G})\\rightarrow\\) Target (C, B, M): Mean IoU(\\%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with the one synthetic (GTAV) dataset. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.47, + 0.729 + ], + "angle": 0, + "content": "are adopted at the same time. Remarkably, compared with the first and second lines, we can find that style projection can approximately bring \\(12\\%\\) mIoU gains in average over the baseline, which fully demonstrates its effectiveness for the generalization on unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Different ways of style projection. As listed in Table 6, we investigate the effect of different ways of style projection. There are two intuitive ways as follows. One way is using the naive instance normalization to project images from different domains into a normalized feature space (i.e., Normalization). The other way is using the extracted style bases to directly substitute the unseen style (i.e., Substitution). We can find that the weighted combination of style bases can effectively enhance the representation of unseen style, producing better generalization on unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Loss terms. As listed in Table 7, we conduct ablation ex" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.31, + 0.875, + 0.515 + ], + "angle": 0, + "content": "
MethodsCBMAvg.-T
MobileNetBaseline29.1620.2727.1925.24
IBN-Net [40]29.5826.0226.3227.31
RobustNet [5]30.6725.0228.2727.99
Ours39.8834.8338.9137.87
ShuffleNetBaseline29.4826.2731.3529.03
IBN-Net [40]32.6129.5533.2031.79
RobustNet [5]33.1531.9834.8533.33
Ours38.9734.6239.6637.75
ResNet-101Baseline34.7129.3237.7433.92
IBN-Net [40]39.1834.0039.3237.50
RobustNet [5]39.9634.9441.7238.87
Ours47.9343.6248.7946.78
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.519, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Table 4. Source \\((\\mathbf{G} + \\mathbf{S})\\rightarrow\\) Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods with different backbones." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.565, + 0.885, + 0.649 + ], + "angle": 0, + "content": "
Sty.-Pro.Sem.-Clu.CBMAvg.-T
36.0328.1532.6132.26
✓44.8742.4246.3744.55
✓39.0130.6035.1934.93
✓✓46.3643.1848.2345.92
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.653, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Table 5. Ablation results for each strategy used in our method. Sty.-Pro. and Sem.-Clu. indicate style projection and semantic clustering, respectively." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.713, + 0.86, + 0.784 + ], + "angle": 0, + "content": "
MethodsCBMAvg.-T
Normalization43.8340.9544.9243.23
Substitution45.0042.7945.1644.32
Ours46.3643.1848.2345.92
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.51, + 0.788, + 0.881, + 0.802 + ], + "angle": 0, + "content": "Table 6. Ablation results for different ways of style projection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "periments to demonstrate the effectiveness of two complementary loss functions in Eq. 11 and Eq. 12. Compared with the naive cross-entropy loss, adding any complementary loss can bring the performance gain, which verifies each of them can effectively supplement the main loss \\(\\mathcal{L}_{CE}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3067" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.46, + 0.173 + ], + "angle": 0, + "content": "
LCELvarLdisCBMAvg.-T
✓44.0041.8245.9743.93
✓✓45.5742.7846.6144.99
✓✓45.9242.4247.0845.14
✓✓✓46.3643.1848.2345.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.177, + 0.404, + 0.19 + ], + "angle": 0, + "content": "Table 7. Ablation results for each loss term." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.2, + 0.454, + 0.338 + ], + "angle": 0, + "content": "
Methods# of ParamsGFLOPsTime (ms)
Baseline45.08M277.777.82
IBN-Net [40]45.08M277.828.74
RobustNet [5]45.08M277.789.48
MLDG [26]45.08M277.779.67
PintheMem [20]45.28M278.3111.64
SAN-SAW [42]25.63M421.8657.58
WildNet [25]45.21M277.168.61
Ours45.22M286.099.98
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Table 8. Comparison of computational cost. Tested with the image size of \\(2048 \\times 1024\\) on one NVIDIA Tesla V100 GPU. We average the inference time over 500 trials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.303, + 0.421 + ], + "angle": 0, + "content": "5. Discussion and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.597 + ], + "angle": 0, + "content": "Distribution analysis. We adopt the t-SNE visualization tool to analyze the effectiveness of our proposed style projection and semantic clustering strategies. As shown in Fig 4, we show the variations of style distribution between different domains before and after style projection. We can find that the style distribution of different domains is well separated before style projection (Fig. 4(a)), while their style distribution is approximately constrained between two style bases after style projection (Fig. 4(b)), which demonstrates style projection successfully projects unseen styles into the style representation space built on style bases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Furthermore, we visualize the semantic distribution between different classes and domains as shown in Fig 5. From Fig. 5(a), we can find that pixel samples belonging to the same class are well clustered while those belonging to different classes are well separated. In addition, the preserved semantic bases are approximately located in the cluster center of pixel samples. From Fig. 5(b), we can find that these pixel samples from different domains are well clustered according to their classes, which demonstrates our semantic clustering successfully achieves the class prediction between different domains by the preserved semantic bases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Complexity of networks. As listed in Table 8, we compare the number of parameters and computational cost with existing DGSS methods. Since we need to store style and semantic bases and estimate the similarity between them and unseen images, the number of parameters and computational cost in our method are slightly higher than the naive baseline. However, our inference time is competitive to exiting DGSS methods due to the efficient implementation of distance measures by matrix multiplications." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.696, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.248, + 0.655, + 0.26 + ], + "angle": 0, + "content": "(a) Before projection" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.089, + 0.888, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.248, + 0.844, + 0.26 + ], + "angle": 0, + "content": "(b) After projection" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Figure 4. t-SNE visualization of style statistics between different domains before (a) and after (b) style projection, where the style statistics (concatenation of mean and variance) is computed from the first layer's feature map of the ResNet-50 trained on two synthetic datasets. Triangles indicate the preserved style bases." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.339, + 0.697, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.499, + 0.653, + 0.51 + ], + "angle": 0, + "content": "(a) Class distribution" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.34, + 0.889, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.499, + 0.854, + 0.51 + ], + "angle": 0, + "content": "(b) Domain distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 5. t-SNE visualization of semantic statistics between different classes (a) and domains (b), where the semantic statistics is computed from the last layer's feature map. Triangles indicate the preserved semantic bases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.619, + 0.597 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.818 + ], + "angle": 0, + "content": "In this paper, we propose a novel style projected clustering method for domain generalized semantic segmentation, which achieves the style and semantic representation of unseen images based on known data. In particular, style projection projects arbitrary unseen styles into the style representation space of source domains and achieves the retention of specific style information between different domains. Semantic clustering predicts the class of each pixel by the minimal similarity distance to semantic bases, which realizes the semantic representation for unseen images and promotes the generalization ability. Through the evaluation on multiple urban-scene datasets, we demonstrate the superior generalization performance of our proposed method over existing DGSS methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. And we gratefully acknowledge the support of MindSpore (https://www.mindspore.cn/)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3068" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[2] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.321, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.393, + 0.468, + 0.461 + ], + "angle": 0, + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.465, + 0.468, + 0.506 + ], + "angle": 0, + "content": "[7] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1):21-27, 1967. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[9] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.468, + 0.665 + ], + "angle": 0, + "content": "[10] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):417-435, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.669, + 0.468, + 0.709 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.468, + 0.753 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.468, + 0.811 + ], + "angle": 0, + "content": "[13] Xiaoqing Guo, Chen Yang, Baopu Li, and Yixuan Yuan. Metacorrection: Domain-aware meta loss correction for unsupervised domain adaptation in semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.815, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[14] Jianzhong He, Xu Jia, Shuaijun Chen, and Jianzhuang Liu. Multi-source domain adaptation with collaborative learning for semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Lei Huang, Dawei Yang, Bo Lang, and Jia Deng. Decorated batch normalization. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[19] Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. Style neophile: Constantly seeking novel styles for domain generalization. In CVPR, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[20] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[21] Shu Kong and Charless C Fowlkes. Recurrent pixel embedding for instance grouping. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[22] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In WACV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[23] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[24] Seunghun Lee, Sunghyun Cho, and Sunghoon Im. Dranet: Disentangling representation and adaptation networks for unsupervised cross-domain adaptation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[25] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[26] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2018. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[28] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[29] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[32] Road Augmentation Library. https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[33] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable learning-to-normalize via switchable normalization. In ICLR, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3069" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[34] Xinyu Luo, Jiaming Zhang, Kailun Yang, Alina Roitberg, Kunyu Peng, and Rainer Stiefelhagen. Towards robust semantic segmentation of accident scenes via multi-source mixed sampling and meta-learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[35] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[36] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[37] Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[39] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.372, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[40] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.414, + 0.469, + 0.454 + ], + "angle": 0, + "content": "[41] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.456, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[43] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.553, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[44] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[45] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[46] Tiago Ramalho and Marta Garnelo. Adaptive posterior learning: few-shot learning with a surprise-based memory module. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.721, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[50] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[51] Herbert A Simon and Allen Newell. Human problem solving: The state of the theory in 1970. American Psychologist, 26(2):145, 1971. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[52] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[53] Haitao Tian, Shiru Qu, and Pierre Payeur. A prototypical knowledge oriented adaptation framework for semantic segmentation. IEEE Transactions on Image Processing, 31:149-163, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[54] Zhuotao Tian, Xin Lai, Li Jiang, Shu Liu, Michelle Shu, Hengshuang Zhao, and Jiaya Jia. Generalized few-shot semantic segmentation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[55] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[56] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[57] SS Vallender. Calculation of the wasserstein distance between probability distributions on the line. Theory of Probability & Its Applications, 18(4):784-786, 1974. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[58] Girish Varma, Anbumani Subramanian, Anoop Namboodiri, Manmohan Chandraker, and CV Jawahar. Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In WACV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[59] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[60] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[61] Haoran Wang, Tong Shen, Wei Zhang, Ling-Yu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[63] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[64] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In ICCV, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[65] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[66] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[67] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. In NeurIPS, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[68] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[69] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[70] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.304 + ], + "angle": 0, + "content": "[71] Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Episode-based prototype generating network for zero-shot learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[73] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.517 + ], + "angle": 0, + "content": "[75] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.518, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[76] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.561, + 0.468, + 0.602 + ], + "angle": 0, + "content": "[77] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3071" + } + ] +] \ No newline at end of file diff --git a/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_origin.pdf b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57089ce424d932c07e2fb415938dcf6798e9803d --- /dev/null +++ b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/98057063-f492-4f6c-b6b6-c6b86b75d3df_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8db71a95c5c1d9d66b196fbfa2fd01b60bff2c14b50059477cf41165d18313a4 +size 3462078 diff --git a/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/full.md b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08d6d3c5fef3e32de18f85647bd839f3d13a3fe5 --- /dev/null +++ b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/full.md @@ -0,0 +1,371 @@ +# Style Projected Clustering for Domain Generalized Semantic Segmentation + +Wei Huang $^{1*}$ Chang Chen $^{2\dagger}$ Yong Li $^{2}$ Jiacheng Li $^{1}$ + +Cheng Li² Fenglong Song² Youliang Yan² Zhiwei Xiong¹ + +$^{1}$ University of Science and Technology of China $^{2}$ Huawei Noah's Ark Lab + +{weih527,jcee}@mail.ustc.edu.cn,zwxiong@ustc.edu.cn, + +{chenchang25,liyong156,licheng89,songfenglong,yanyouliang}@huawei.com + +# Abstract + +Existing semantic segmentation methods improve generalization capability, by regularizing various images to a canonical feature space. While this process contributes to generalization, it weakens the representation inevitably. In contrast to existing methods, we instead utilize the difference between images to build a better representation space, where the distinct style features are extracted and stored as the bases of representation. Then, the generalization to unseen image styles is achieved by projecting features to this known space. Specifically, we realize the style projection as a weighted combination of stored bases, where the similarity distances are adopted as the weighting factors. Based on the same concept, we extend this process to the decision part of model and promote the generalization of semantic prediction. By measuring the similarity distances to semantic bases (i.e., prototypes), we replace the common deterministic prediction with semantic clustering. Comprehensive experiments demonstrate the advantage of proposed method to the state of the art, up to $3.6\%$ mIoU improvement in average on unseen scenarios. Code and models are available at https://gitee.com/mindspore/models/tree/master/research/cv/SPC-Net. + +# 1. Introduction + +Domain generalization methods aim to promote the performance of model (trained on source datasets), when applying it to unseen scenarios (target domains) [9, 19, 29, 36, 62, 74, 75]. Recently, domain generalization for semantic segmentation (DGSS) has attracted increasingly more attention due to the rise of safety-critical applications, such as autonomous driving [3, 12, 22, 45]. + +Existing DGSS methods improve the pixel-wise generalization performance by learning domain-agnostic rep + +![](images/86d35739b66d16476b3e032ae64bd899880ffd8e7193c0469049fb93a840fe31.jpg) +Figure 1. Illustration of instance normalization/whitening (IN/IW) [5, 20, 40] and our proposed style projected clustering method. IN and IW regularize image features from different domains to a canonical space (a-c). Our method builds style and semantic representation spaces based on the data from known domains (d). + +resentations [5, 16, 20, 25, 40, 42, 66, 72]. Researches in this line share the similar goal in general, that is to capture the domain-invariant characteristics of object contents, and eliminates the domain-specific ones (i.e., image styles). As two representatives, Instance Normalization (IN) [56] and Instance Whitening (IW) [17] regularize image features from different domains to a canonical space, as illustrated in Fig. 1(a) and 1(b). Specifically, IN achieves center-level feature alignment via channel-wise feature normalization [33,40], and IW realizes uniform feature distribution by removing linear correlation between channels [5,41]. Moreover, the combination of these two methods is proposed in [42] for a better generalization, as shown in Fig. 1(c). + +Nevertheless, feature regularization inevitably weakens the representation capability, as a part of feature information is eliminated. Theoretically, it works under a strong assumption that the eliminated information is strictly the domain-specific ones. Yet in practice, the perfect disentanglement between image style and content is difficult to + +achieve. It means that a part of content features will also be eliminated in the process of feature regularization, and thus degrades the segmentation performance. + +Instead of seeking common ground by feature regularization, we aim to address DGSS in a different way. In this paper, we propose style projection as an alternative, which utilizes the features from different domains as bases to build a better representation space, as shown in Fig. 1(d). The motivation of style projection comes from a basic concept of generalization, that is to represent unseen data based on the known ones. Specifically, following the common practice, we adopt the statistics (i.e., mean and variance) of features in channel dimension to represent image styles. The image styles from source domains are iteratively extracted and stored as the bases of representation. Then, we project the style of given unseen images into this representation space to promote generalization. This projection process is implemented as a weighted combination of stored style bases, where the similarity distance between styles are adopted as the weighting factors, i.e., $\lambda_{1}$ and $\lambda_{2}$ shown in Fig. 1(d). + +Based on the projected style features, we further devise the decision part of model, which is elaborated for semantic segmentation. Typically, existing methods learn a parametric function to map pixel-wise features to semantic predictions. We replace this deterministic prediction with semantic clustering, where the class of each pixel is predicted by the minimal similarity distance to semantic bases, as shown in Fig. 1(d). Notably, it follows the same concept of style projection, that is to predict unseen data based on the known ones. More concretely, to facilitate the performance of semantic clustering, we propose a variant of contrastive loss to align the semantic bases of same classes and enhance discriminability between different classes. + +We conduct comprehensive experiments on single- and multi-source settings to demonstrate the superior generalization of our method over existing DGSS methods. In addition, we visually analyze the effective representation of our proposed method for unseen images in both style and semantic spaces. + +Contributions of this paper are summarized as follows: + +- Beyond existing feature regularization methods, we propose style projected clustering, pointing out a new avenue to address DGSS. +- We propose style projection, which projects unseen styles into the style representation space built on known domains for a better representation. +- We propose semantic clustering to predict the class of each pixel in unseen images by the similarity distance to semantic bases, which further improves the generalization capability for unseen domains. +- Our proposed method outperforms the current state of the arts on multiple DGSS benchmarks. + +# 2. Related Work + +Domain adaptation and generalization. To reduce the burden of pixel-wise annotations on target domains, domain adaptation (DA) technologies are proposed to narrow the domain gap between source and target domains via image translation [14, 24, 37], feature alignment [55, 60, 61], self-training [2, 39, 77] and meta-learning [13, 34] strategies. However, these DA methods require the access of data on target domains. Domain generalization (DG) aims to address a more practical problem where the target domain cannot be accessed. Numerous DG works have been proposed for image classification via style augmentation [19, 59, 68, 75], domain alignment [29, 31], feature disentanglement [27, 44] and meta-learning [9, 26, 28]. + +Domain generalization for semantic segmentation. Similar to image classification, DG for semantic segmentation (DGSS) methods are proposed to learn domain-agnostic representations, including style augmentation [16, 25, 43, 72], feature normalization/whitening [5, 40, 42, 66] and meta-learning [20]. To avoid overfitting on source domains, DRPC [72] and FSDR [16] adopt style augmentations in the image space to extend the number of source samples, while WildNet [25] realizes it in the feature space with the aid of ImageNet [8]. Alternatively, normalization and whitening are investigated to achieve distribution alignment between different domains. IBN-Net [40] and RobustNet [5] adopt instance normalization and whitening, respectively, to eliminate the specific style information of each domain. Furthermore, SAN-SAW [42] proposes semantic-aware instance normalization and whitening to enhance the distinguishability between classes. In addition, PintheMem [20] combines the memory-guided network with the meta-learning strategy and obtains competitive performances. Different from these DGSS methods, our method embraces the differences from multiple known domains and takes advantage of their diversity to build a better representation space, realizing the representation of unseen images by the known data. + +Prototype learning. Inspired by the cognitive psychology that human use the knowledge learned in the past to judge the class of unknown things [51,69], prototype-based classification methods have attracted increasing attention, where the class of unknown images is determined by its nearest neighbors in the feature space [7, 10]. Owing to its excellent interpretability and generalizability, prototype learning shows good potential in many fields, such as few-shot learning [1, 52], zero-shot learning [67, 71], unsupervised learning [30, 65]. Recently, prototype learning is also introduced in the dense prediction task, including supervised [76], few-shot [54, 63] and domain adaptive [53, 73] semantic segmentation. To facilitate the learning of prototypes, metric learning [23, 50, 64] is often adopted to pull samples belonging to the same class together and push those of different classes away from each other in the embedding (i.e., feature) space. + +![](images/69d247d644081a0f51d0c428aadc22f240ddd97d5ea3345bb41629a57fd7cc1f.jpg) +Figure 2. The framework of style projected clustering, which consists of two components, i.e., style projection and semantic clustering. We iteratively extract the style and semantic information of seen domains as style bases $(p_m^\mu, p_m^\sigma)$ and semantic bases $p_m^c$ . In style projection, we first calculate the similarity between the unseen style $(\mu_m, \sigma_m)$ from the shallow feature $F_m^s$ and style bases $(p_m^\mu, p_m^\sigma)$ as weighted factors $\lambda_m$ . Then, the weighted combination of style bases $(\mu_m', \sigma_m')$ is projected on $F_m^n$ to obtain the projected feature $F_m^r$ . In semantic clustering, we calculate the similarity between pixel embeddings in the deep feature $F_m^d$ and semantic bases $p_m^c$ . Then, the class of each pixel is determined by the nearest semantic base. During the training phase, the cross-entropy loss $\mathcal{L}_{CE}$ , variance loss $\mathcal{L}_{var}$ and discrimination loss $\mathcal{L}_{dis}$ are adopted to supervise the learning of style and semantic bases. + +Similar to these methods, we adopt the form of prototypes (i.e. bases) to represent semantics. Yet these semantic bases are learned in a different way to facilitate domain generalization, by using a new variant of contrastive loss. + +# 3. Style Projected Clustering + +The overall architecture of our proposed method is depicted in Fig. 2, which consists of two components, i.e., style projection and semantic clustering. In style projection, we project the unseen style into the style representation space built on style bases, according to the similarity between the unseen style and style bases. In semantic clustering, we estimate the similarity between pixel embeddings and semantic bases (i.e., prototypes) to determine the class of pixels in unseen images by the nearest semantic base. + +# 3.1. Problem Formulation + +In the domain generalized semantic segmentation problem, we are given $M$ source domains $\mathcal{S} = \{\mathcal{S}_1,\mathcal{S}_2,\dots,\mathcal{S}_M\}$ that are from multiple datasets with different data distributions. The $m$ -th source domain $\mathcal{S}_m$ can be represented as $\mathcal{S}_m = \{(x_m,y_m)\}$ , where $x_{m}\in \mathbb{R}^{H\times W\times 3}$ is an image from the $m$ -th source domain, $y_{m}\in \mathbb{R}^{H\times W\times C}$ is the corresponding pixel-wise label, $C$ is the number of semantic classes, $H$ and $W$ are the height and width of the image $x_{m}$ , respectively. In this work, our goal is to train a semantic segmentation model $\phi$ to obtain the best generalization performance on multiple target domains $\mathcal{T}$ which cannot be accessed during the training phase. + +# 3.2. Style Projection + +The style difference of images is the main factor leading to the domain shift, which limits the generalization ability of the learned model. Pioneering works [11, 18, 40, 75] have demonstrated that the feature distribution shift caused by style differences lies mainly in shallow layers of networks. It also shows that the shallow feature distribution of networks can reflect the style information of the input image $x_{m}$ . Thus, existing works always adopt the channel-wise mean and variance of the shallow feature to represent the style distribution of $x_{m}$ [18, 25]. Following these works, let $F_{m}^{s} \in \mathbb{R}^{D \times H_{s} \times W_{s}}$ be the shallow feature of $x_{m}$ from the network $\phi$ , where $D$ denotes the number of channels. The channel-wise mean $\mu_{m} \in \mathbb{R}^{D}$ and variance $\sigma_{m} \in \mathbb{R}^{D}$ of the feature $F_{m}^{s}$ can be calculated as follows: + +$$ +\mu_ {m} = \frac {1}{H _ {s} W _ {s}} \sum_ {h = 1} ^ {H _ {s}} \sum_ {w = 1} ^ {W _ {s}} F _ {m} ^ {s}, \tag {1} +$$ + +$$ +\sigma_ {m} = \sqrt {\frac {1}{H _ {s} W _ {s}} \sum_ {h = 1} ^ {H _ {s}} \sum_ {w = 1} ^ {W _ {s}} (F _ {m} ^ {s} - \mu_ {m}) ^ {2}}. +$$ + +To eliminate the specific style information of images, instance normalization [40] is adopted to standardize the feature $F_{m}^{s}$ to a standard distribution (i.e., zeros mean and one standard deviation) as follows: + +$$ +F _ {m} ^ {n} = \frac {F _ {m} ^ {s} - \mu_ {m}}{\sigma_ {m} + \epsilon}, \tag {2} +$$ + +where $F_{m}^{n}$ stands for the normalized feature, and $\epsilon$ is a small value to avoid division by zero. + +Although instance normalization achieves to remove the specific style information of images, it also eliminates the natural differences between domains, which weakens the representation for target domains and produces limited generalization performance. Therefore, to preserve the specific style information of each domain, we propose style bases $P_{sty} = \{(p_m^\mu ,p_m^\sigma)\}_{m = 1}^M$ to store the style information of source domains, and then leverage the preserved style bases $P_{sty}$ to build a style representation space, realizing the projection of unseen style, as shown in Fig. 2. Specifically, we first leverage Wasserstein distance [57] to estimate the style distribution discrepancy between the input image $x_{m}$ and the $m$ -th style bases $(p_m^\mu ,p_m^\sigma)$ as follows: + +$$ +d _ {m} = | | \mu_ {m} - p _ {m} ^ {\mu} | | _ {2} ^ {2} + (\sigma_ {m} ^ {2} + p _ {m} ^ {\sigma 2} - 2 \sigma_ {m} p _ {m} ^ {\sigma}), \qquad (3) +$$ + +where $d_{m}$ denotes the distribution distance between the current image $x_{m}$ and the $m$ -th source domain. Then, we use the reciprocal of $d_{m}$ to characterize the similarity between $x_{m}$ and $m$ -th style bases as follows: + +$$ +\lambda_ {m} = \frac {\exp \left(1 / \left(1 + d _ {m}\right)\right)}{\sum_ {m = 1} ^ {M} \exp \left(1 / \left(1 + d _ {m}\right)\right)}, \tag {4} +$$ + +where the softmax operation is utilized to make the sum of $\lambda = \{\lambda_m|m = 1,2,\dots,M\}$ equal to 1. Based on the estimated similarity $\lambda$ , we can obtain the projected style $(\mu_{m}^{\prime},\sigma_{m}^{\prime})$ by the weighted sum of style bases as follows: + +$$ +\mu_ {m} ^ {\prime} = \sum_ {m = 1} ^ {M} \lambda_ {m} p _ {m} ^ {\mu}, \sigma_ {m} ^ {\prime} = \sum_ {m = 1} ^ {M} \lambda_ {m} p _ {m} ^ {\sigma}. \tag {5} +$$ + +Finally, following previous works [11, 18, 19, 25], we inject the projected style $(\mu_{m}^{\prime},\sigma_{m}^{\prime})$ into the normalized feature $F_{m}^{n}$ to obtain the projected feature as follows: + +$$ +F _ {m} ^ {r} = \sigma_ {m} ^ {\prime} F _ {m} ^ {n} + \mu_ {m} ^ {\prime}. \tag {6} +$$ + +During the training phase, we adopt the momentum update strategy to achieve the online collection of style information as follows: + +$$ +p _ {m} ^ {\mu} = \alpha p _ {m} ^ {\mu} + (1 - \alpha) \mu_ {m}, \tag {7} +$$ + +$$ +p _ {m} ^ {\sigma} = \alpha p _ {m} ^ {\sigma} + (1 - \alpha) \sigma_ {m}, +$$ + +where $\alpha \in [0,1]$ is a momentum coefficient. In addition, we randomly initialize $P_{sty}$ to start training, where $p_m^\mu$ and $p_m^\sigma$ are initialized with zero-mean and one-mean distribution, respectively. By Eq. 7, we realize the style statistic of source domains and store it as style bases efficiently. + +After style projection, the projected feature $F_{m}^{r}$ is input into the next layer of the network $\phi$ . Our style projection is designed as a plug-and-play module that can be applied behind any network layer. However, as the layer is deeper, the style information loosens while the semantic information plays a more important role. Thus, in this work, style projection is only used in the first two layers of $\phi$ to obtain the best generalization performance. + +# 3.3. Semantic Clustering + +To obtain the final pixel-wise predictions, we further propose semantic clustering on the deep feature extracted by the network $\phi$ . Let $F_{m}^{d}\in \mathbb{R}^{D\times H_{d}\times W_{d}}$ be the deep feature of the input image $x_{m}$ from $\phi$ . Existing DGSS methods generically apply a learnable segmentation classifier $\phi_{cls}$ on $F_{m}^{d}$ for the dense prediction. However, the parameters of $\phi_{cls}$ is learned on the deep features of source domains $S$ , and thus its generalization ability on target domain $\mathcal{T}$ is limited. In addition, the semantic information between different domains is implicitly encoded in the same parameter space, which causes the specific semantic information of domains to be eliminated. + +Based on the concept of style bases, we introduce semantic bases $P_{sem} = \{p_m^c\}_{c,m=1}^{C,M}$ to preserve the semantic information of each domain and each class, where $p_m^c \in \mathbb{R}^D$ is the cluster center of training pixel embeddings belonging to the $c$ -th class from the $m$ -th source domain in the feature space. Following the prototype theory [7, 10, 76], the class of each pixel embedding $e \in F_m^d$ can be determined by its nearest semantic bases as follows: + +$$ +c (e) = c ^ {*}, \text {w i t h} \left(c ^ {*}, m ^ {*}\right) = \operatorname {a r g m i n} _ {c, m} \left\{d _ {m} ^ {c} \right\} _ {c, m = 1} ^ {C, M}, \tag {8} +$$ + +where $d_m^c = -\cos(e, p_m^c)$ is the negative cosine distance used to estimate the similarity between the current embedding $e$ and semantic bases $p_m^c$ . In this work, the pixel embedding $e$ and semantic bases $p_m^c$ are both $l_2$ -normalized. Therefore, the similarity distance can be simply formulated as $d_m^c = -ep_m^c$ . Different from the learnable segmentation classifier $\phi_{cls}$ , $P_{sem}$ not only explicitly captures characteristic properties of each class from each domain, but also determines the class of pixels in unseen images without introducing extra learnable parameters. + +To facilitate the training of the network $\phi$ during the training phase, we estimate the probability value of pixel embedding $e$ belonging to class $c$ as follows: + +$$ +v (c | e) = \frac {\exp (- d ^ {c})}{\sum_ {c = 1} ^ {C} \exp (- d ^ {c})}, \tag {9} +$$ + +where $d^{c} = \min_{m}\{d_{m}^{c}\}_{m = 1}^{M}$ denotes the similarity between $e$ and its closet semantic base belonging to class $c$ . Then, we adopt the standard cross-entropy loss to supervise the training of the network $\phi$ as follows: + +$$ +\mathcal {L} _ {C E} = - \frac {1}{H _ {d} W _ {d}} \sum_ {h = 1} ^ {H _ {d}} \sum_ {w = 1} ^ {W _ {d}} \sum_ {c = 1} ^ {C} y _ {m} \log (v (c | e)), \tag {10} +$$ + +where $y_{m}$ is the pixel-wise label corresponding to the input image $x_{m}$ . + +However, the naive cross-entropy loss only optimizes the relative relations between intra-class and inter-class distance, which ignores the absolute distance constraint between pixel embeddings and semantic bases. That is to say, + +we expect that the pixel embedding belonging to class $c$ is closer to the $c$ -th semantic base and is farther away from the semantic bases belonging to other classes. Inspired by metric learning [21, 23], we further propose variance and discrimination terms as two extra training objectives. The former is an intra-class cluster that pulls the pixel embedding $e_{m}^{c}$ belonging to class $c$ from the $m$ -th source domain towards the semantic bases $p_{m}^{c}$ : + +$$ +\mathcal {L} _ {v a r} = \frac {1}{M C} \sum_ {m = 1} ^ {M} \sum_ {c = 1} ^ {C} \left(1 - e _ {m} ^ {c} p _ {m} ^ {c}\right) ^ {2}. \tag {11} +$$ + +The latter is designed in a contrastive learning way which encourages the current cluster center $\overline{e}_m^c$ is closer to the $c$ -th semantic bases $p_{c+}$ (i.e., positive keys) and to be far away from semantic bases belonging to other class $p_{c-}$ (i.e., negative keys): + +$$ +\mathcal {L} _ {d i s} = \frac {1}{M} \sum_ {p _ {c +}} - \log \frac {\exp \left(\bar {e} _ {m} ^ {c} p _ {c +} / \tau\right)}{\exp \left(\bar {e} _ {m} ^ {c} p _ {c +} / \tau\right) + \sum_ {p _ {c -}} \exp \left(\bar {e} _ {m} ^ {c} p _ {c -} / \tau\right)}, \tag {12} +$$ + +where $\bar{e}_m^c$ is the cluster center (i.e., mean embedding) of pixel embedding $e_m^c$ in the current feature $F_m^d$ , and $\tau$ is a temperature hyper-parameter. By Eq. 12, we realize the alignment of semantic bases belonging to the same class $c$ from different domains. Different from existing pixel-wise contrastive learning paradigm [64], the positive and negative samples in Eq. 12 are semantic bases rather than pixel embeddings. Thus, we don't need to construct a memory bank to store sufficient embedding samples, which also significantly reduces the computational cost. + +To achieve the online collocation of semantic information from source domains, we adopt the same momentum update strategy to update semantic bases $P_{sem}$ as follows: + +$$ +p _ {m} ^ {c} = \alpha p _ {m} ^ {c} + (1 - \alpha) \bar {e} _ {m} ^ {c}, \tag {13} +$$ + +where $\alpha$ the momentum coefficient. Like style bases, we also randomly initialize the semantic bases $P_{sem}$ with zero-mean distribution to start our training. + +# 3.4. Training and Inference + +During the training phase, we combine above three loss terms for the end-to-end training as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {C E} + \beta \mathcal {L} _ {\text {v a r}} + \gamma \mathcal {L} _ {\text {d i s}}, \tag {14} +$$ + +where $\beta$ and $\gamma$ are weighting coefficients to balance these three terms. For each training iteration, in addition to the parameter update of the network $\phi$ , the style and semantic bases are also updated online by Eq. 7 and Eq. 13. + +During the inference phase, we leverage Eq. 8 to obtain final pixel-wise predictions by the nonparametric cluster of pixel embeddings outputted from the learned network $\phi$ . + +# 4. Experiments + +# 4.1. Datasets + +Synthetic datasets. GTAV [47] contains 24966 images with a resolution of $1914 \times 1052$ captured from the GTA-V game engine. Synthia [48] contains 9400 images with a resolution of $1280 \times 760$ generated from virtual urban scenes. + +Real-world datasets. IDD [58] contains 10004 images with an average resolution of $1678 \times 968$ captured from Indian roads. Cityscapes [6] contains 5000 fine annotated images with a resolution of $2048 \times 1024$ captured from 50 different cities primarily in Germany. BDD100K [70] contains 10000 image with a resolution of $1280 \times 720$ captures from different locations in US. Mapillary [38] contains 25000 images with an average resolution of $1920 \times 1080$ captured from all around the world. + +# 4.2. Implementation Details + +Following the previous work [5], we adopt DeepLabV3+ [4] with ResNet-50, ResNet-101 [15], MobileNetV2 [49] and ShuffleNetV2 [35] backbones as our segmentation networks, where all backbones are pre-trained on ImageNet [8]. During the training phase, we adopt the SGD optimizer [46] with a momentum of 0.9 and weight decay of $5e - 4$ . The initial learning rate is set to 0.01 and is decreased using the polynomial scheduling with a power of 0.9. We train all models for 40K iterations, except for the three-source setting, the model is trained for 100K iterations. In addition to some common data augmentations used in [5], we adopt extra strong style augmentations to enrich the style information of urban-scene images [32], which aims to enhance the proposed style projection ability in networks. More details can be found in our supplementary materials. + +# 4.3. Results + +Comparison methods. We extensively compare our proposed method against existing DGSS methods, which can be classified into three groups, including style augmentation (WildNet [25]), feature normalization/whitening (IBNNet [40], RobustNet [5] and SAN-SAW [42]), and meta-learning (MLDG [26] and PintheMem [20]). Since SAN-SAW [42] and WildNet [25] are only implemented on the single-source setting in their paper, we reproduce them on the multi-source setting to make a comparison. In particular, WildNet [25] utilizes the external dataset (i.e., ImageNet) to extend the style and content information of source domains. Thus, we re-implement it by replacing the external dataset with the source dataset for a fair comparison, which is marked with * in our tables. + +Multi-source setting. To demonstrate the effectiveness of our proposed method, we first conduct contrast experiments on the multi-source DGSS setting, where multiple source domains can be efficiently used to build a diverse + +
MethodsPublicationCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaAvg.-SAvg.-A
\( Baseline^{\dagger} \)-35.4625.0931.9430.8368.4867.9968.2445.79
\( IBN-Net^{\dagger} [40] \)ECCV 201835.5532.1838.0935.2769.7266.9068.3148.49
\( RobustNet^{\dagger} [5] \)CVPR 202137.6934.0938.4936.7668.2668.7768.5249.46
\( Baseline^{\ddagger} \)-33.4229.0732.1931.5669.6363.9366.7845.65
\( MLDG^{\ddagger} [26] \)AAAI 201838.8431.9535.6035.4664.6151.6958.1544.54
\( PintheMem^{\ddagger} [20] \)CVPR 202244.5138.0742.7041.7665.8554.4960.1749.12
Baseline-36.0328.1532.6132.2669.3067.6168.4646.65
SAN-SAW [42]CVPR 202242.1337.7442.9140.9363.9862.5863.2849.87
WildNet [25]CVPR 202243.6539.9043.2842.2868.0563.9866.0251.77
WildNet* [25]CVPR 202239.3334.7641.0638.3869.7062.1165.9149.39
Ours-46.3643.1848.2345.9272.4674.8773.6757.02
+ +Table 1. Source $(\mathbf{G} + \mathbf{S})\rightarrow$ Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) datasets. The best and second best results are highlighted and underlined. Avg.-T, Avg.-S and Avg.-A denote the average results on target, source and all domains, respectively. Results with the $\dagger$ and $\ddagger$ sign are from [5] and [20], respectively. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison. + +
MethodsCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaIDDAvg.-SAvg.-A
Baseline‡52.5147.4754.7051.5670.3167.1371.5669.6760.61
IBN-Net‡[40]54.3948.9156.0653.1270.7363.6871.0268.4860.80
RobustNet‡[5]54.7049.0056.9053.5370.0666.4071.0269.1661.35
MLDG‡[26]54.7648.5255.9453.0769.5359.7967.7365.6859.38
PintheMem‡[20]56.5750.1858.3155.0269.9962.9967.5866.8560.94
Baseline54.1646.2455.5751.9968.3565.1270.0767.8559.92
SAN-SAW [42]54.8946.5056.3852.5964.4964.7666.3765.2158.90
WildNet [25]55.5850.3157.9354.6167.6561.3570.0766.3660.48
WildNet* [25]53.6148.9256.1852.9070.9859.6964.5265.0658.98
Ours57.9153.2661.6157.5974.6478.3576.0776.3566.97
+ +Table 2. Source $(\mathbf{G} + \mathbf{S} + \mathbf{I})\rightarrow$ Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) and one real (IDD) datasets. Results with the $\ddagger$ sign are from [20]. + +representation space. As listed in Table 1, we quantitatively compare our results with existing DGSS methods on both target and source datasets, where all networks with the ResNet-50 backbone are trained with two synthetic datasets (i.e., GTAV and Synthia). Remarkably, compared with the state-of-the-art method (i.e., WildNet [20]), our method not only shows superior generalization capability on target datasets (up to $3.6\%$ mIoU in average), but also significantly improve the performance on source datasets (up to $7.6\%$ mIoU), which demonstrates our method can enhance the representation ability of the learned model on both source and target domains. Furthermore, We provide visual prediction results for qualitative comparisons as shown in Fig. 3. Our method obtains the best visual results on different target datasets. Following [20], we add one real dataset (i.e., IDD) to source domains to further verify the superiority of our method on more source datasets. As listed in Table 2, our method also outperforms existing methods on both source and target domains by a large margin. + +Single-source setting. We further implement our method in + +the single-source setting to make a comprehensive comparison, where all network with the ResNet-50 backbone are trained with one synthetic (i.e., GTAV) dataset. As listed in Table 3, our method shows superior generalization performances over existing DGSS methods. Compared with the naive baseline, our method brings approximately $14\%$ mIoU gains in average on target datasets. + +Different backbones. To demonstrate the wide applicability of our method, we compare our results with classic DGSS methods (i.e., IBN-Net [40] and RobustNet [5]) with different backbones. As listed in Table 4, our method shows superior performances on both large (i.e., ResNet-101) and lightweight (i.e., MobileNet and ShuffleNet) backbones. + +# 4.4. Ablation Studies + +We conduct comprehensive ablation studies with the ResNet-50 backbone on two source domains (i.e., GTAV and Synthia) as following. + +Proposed strategies. As listed in Table 5, our method shows the best generalization capability when two strategies + +![](images/9539866d645f4702f1a17dda1aef83bff39617833c8facdb157c2d84ccdc47fc.jpg) +Figure 3. Source $(\mathbf{G} + \mathbf{S}) \rightarrow$ Target (C, B, M): Visualization comparison with existing DGSS methods on three different target domains. + +
MethodsCBMAvg.-T
Baseline28.9525.1428.1827.42
IBN-Net [40]33.8532.3037.7534.63
RobustNet [5]36.5835.2040.3337.37
Baseline31.6026.7029.0029.10
MLDG [26]36.7032.1032.2033.67
PintheMem [20]41.0034.6037.4037.67
Baseline29.3225.7128.3327.79
SAN-SAW [42]39.7537.3441.8639.65
Baseline35.1629.7131.2932.05
WildNet [25]44.6238.4246.0943.04
Baseline32.0126.0429.3529.13
WildNet* [25]40.1034.8239.3838.10
Ours44.1040.4645.5143.36
+ +Table 3. Source $(\mathbf{G})\rightarrow$ Target (C, B, M): Mean IoU(\%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with the one synthetic (GTAV) dataset. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison. + +are adopted at the same time. Remarkably, compared with the first and second lines, we can find that style projection can approximately bring $12\%$ mIoU gains in average over the baseline, which fully demonstrates its effectiveness for the generalization on unseen domains. + +Different ways of style projection. As listed in Table 6, we investigate the effect of different ways of style projection. There are two intuitive ways as follows. One way is using the naive instance normalization to project images from different domains into a normalized feature space (i.e., Normalization). The other way is using the extracted style bases to directly substitute the unseen style (i.e., Substitution). We can find that the weighted combination of style bases can effectively enhance the representation of unseen style, producing better generalization on unseen domains. + +Loss terms. As listed in Table 7, we conduct ablation ex + +
MethodsCBMAvg.-T
MobileNetBaseline29.1620.2727.1925.24
IBN-Net [40]29.5826.0226.3227.31
RobustNet [5]30.6725.0228.2727.99
Ours39.8834.8338.9137.87
ShuffleNetBaseline29.4826.2731.3529.03
IBN-Net [40]32.6129.5533.2031.79
RobustNet [5]33.1531.9834.8533.33
Ours38.9734.6239.6637.75
ResNet-101Baseline34.7129.3237.7433.92
IBN-Net [40]39.1834.0039.3237.50
RobustNet [5]39.9634.9441.7238.87
Ours47.9343.6248.7946.78
+ +Table 4. Source $(\mathbf{G} + \mathbf{S})\rightarrow$ Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods with different backbones. + +
Sty.-Pro.Sem.-Clu.CBMAvg.-T
36.0328.1532.6132.26
✓44.8742.4246.3744.55
✓39.0130.6035.1934.93
✓✓46.3643.1848.2345.92
+ +Table 5. Ablation results for each strategy used in our method. Sty.-Pro. and Sem.-Clu. indicate style projection and semantic clustering, respectively. + +
MethodsCBMAvg.-T
Normalization43.8340.9544.9243.23
Substitution45.0042.7945.1644.32
Ours46.3643.1848.2345.92
+ +Table 6. Ablation results for different ways of style projection. + +periments to demonstrate the effectiveness of two complementary loss functions in Eq. 11 and Eq. 12. Compared with the naive cross-entropy loss, adding any complementary loss can bring the performance gain, which verifies each of them can effectively supplement the main loss $\mathcal{L}_{CE}$ . + +
LCELvarLdisCBMAvg.-T
✓44.0041.8245.9743.93
✓✓45.5742.7846.6144.99
✓✓45.9242.4247.0845.14
✓✓✓46.3643.1848.2345.92
+ +Table 7. Ablation results for each loss term. + +
Methods# of ParamsGFLOPsTime (ms)
Baseline45.08M277.777.82
IBN-Net [40]45.08M277.828.74
RobustNet [5]45.08M277.789.48
MLDG [26]45.08M277.779.67
PintheMem [20]45.28M278.3111.64
SAN-SAW [42]25.63M421.8657.58
WildNet [25]45.21M277.168.61
Ours45.22M286.099.98
+ +Table 8. Comparison of computational cost. Tested with the image size of $2048 \times 1024$ on one NVIDIA Tesla V100 GPU. We average the inference time over 500 trials. + +# 5. Discussion and Analysis + +Distribution analysis. We adopt the t-SNE visualization tool to analyze the effectiveness of our proposed style projection and semantic clustering strategies. As shown in Fig 4, we show the variations of style distribution between different domains before and after style projection. We can find that the style distribution of different domains is well separated before style projection (Fig. 4(a)), while their style distribution is approximately constrained between two style bases after style projection (Fig. 4(b)), which demonstrates style projection successfully projects unseen styles into the style representation space built on style bases. + +Furthermore, we visualize the semantic distribution between different classes and domains as shown in Fig 5. From Fig. 5(a), we can find that pixel samples belonging to the same class are well clustered while those belonging to different classes are well separated. In addition, the preserved semantic bases are approximately located in the cluster center of pixel samples. From Fig. 5(b), we can find that these pixel samples from different domains are well clustered according to their classes, which demonstrates our semantic clustering successfully achieves the class prediction between different domains by the preserved semantic bases. + +Complexity of networks. As listed in Table 8, we compare the number of parameters and computational cost with existing DGSS methods. Since we need to store style and semantic bases and estimate the similarity between them and unseen images, the number of parameters and computational cost in our method are slightly higher than the naive baseline. However, our inference time is competitive to exiting DGSS methods due to the efficient implementation of distance measures by matrix multiplications. + +![](images/bff090d503a3576984a0bd43676e2de63e23e6b6491b7d787cafdb091bd9b6bd.jpg) +(a) Before projection + +![](images/e68ce15f23383c76af465946d99195077ea4ba8eee8c000c6de99c98af946612.jpg) +(b) After projection + +![](images/d7682256312cef08236365d3fb7e0c670a5ebd0dd3fe73e394d9c1faa8f8a79a.jpg) +Figure 4. t-SNE visualization of style statistics between different domains before (a) and after (b) style projection, where the style statistics (concatenation of mean and variance) is computed from the first layer's feature map of the ResNet-50 trained on two synthetic datasets. Triangles indicate the preserved style bases. +(a) Class distribution +Figure 5. t-SNE visualization of semantic statistics between different classes (a) and domains (b), where the semantic statistics is computed from the last layer's feature map. Triangles indicate the preserved semantic bases. + +![](images/ea1420d302b5320496147ee66b87378e9d5514a36061bad6bc4db5eb194418c4.jpg) +(b) Domain distribution + +# 6. Conclusion + +In this paper, we propose a novel style projected clustering method for domain generalized semantic segmentation, which achieves the style and semantic representation of unseen images based on known data. In particular, style projection projects arbitrary unseen styles into the style representation space of source domains and achieves the retention of specific style information between different domains. Semantic clustering predicts the class of each pixel by the minimal similarity distance to semantic bases, which realizes the semantic representation for unseen images and promotes the generalization ability. Through the evaluation on multiple urban-scene datasets, we demonstrate the superior generalization performance of our proposed method over existing DGSS methods. + +Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. And we gratefully acknowledge the support of MindSpore (https://www.mindspore.cn/). + +# References + +[1] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In ICML, 2019. 2 +[2] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, 2021. 2 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1 +[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 5 +[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 1, 2, 5, 6, 7, 8 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5 +[7] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1):21-27, 1967. 2, 4 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5 +[9] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 1, 2 +[10] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):417-435, 2012. 2, 4 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 3, 4 +[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1 +[13] Xiaoqing Guo, Chen Yang, Baopu Li, and Yixuan Yuan. Metacorrection: Domain-aware meta loss correction for unsupervised domain adaptation in semantic segmentation. In CVPR, 2021. 2 +[14] Jianzhong He, Xu Jia, Shuaijun Chen, and Jianzhuang Liu. Multi-source domain adaptation with collaborative learning for semantic segmentation. In CVPR, 2021. 2 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 + +[16] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 1, 2 +[17] Lei Huang, Dawei Yang, Bo Lang, and Jia Deng. Decorated batch normalization. In CVPR, 2018. 1 +[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3, 4 +[19] Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. Style neophile: Constantly seeking novel styles for domain generalization. In CVPR, 2022. 1, 2, 4 +[20] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8 +[21] Shu Kong and Charless C Fowlkes. Recurrent pixel embedding for instance grouping. In CVPR, 2018. 5 +[22] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In WACV, 2021. 1 +[23] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2, 5 +[24] Seunghun Lee, Sunghyun Cho, and Sunghoon Im. Dranet: Disentangling representation and adaptation networks for unsupervised cross-domain adaptation. In CVPR, 2021. 2 +[25] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[26] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2018. 2, 5, 6, 7, 8 +[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, 2017. 2 +[28] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In ICCV, 2019. 2 +[29] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 1, 2 +[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2020. 2 +[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2 +[32] Road Augmentation Library. https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library.5 +[33] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable learning-to-normalize via switchable normalization. In ICLR, 2018. 1 + +[34] Xinyu Luo, Jiaming Zhang, Kailun Yang, Alina Roitberg, Kunyu Peng, and Rainer Stiefelhagen. Towards robust semantic segmentation of accident scenes via multi-source mixed sampling and meta-learning. In CVPR, 2022. 2 +[35] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5 +[36] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, 2013. 1 +[37] Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In CVPR, 2018. 2 +[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5 +[39] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 2 +[40] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 1, 2, 3, 5, 6, 7, 8 +[41] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 1 +[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8 +[43] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2 +[44] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, 2020. 2 +[45] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 1 +[46] Tiago Ramalho and Marta Garnelo. Adaptive posterior learning: few-shot learning with a surprise-based memory module. In ICLR, 2018. 5 +[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 5 +[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 5 +[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5 +[50] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015. 2 +[51] Herbert A Simon and Allen Newell. Human problem solving: The state of the theory in 1970. American Psychologist, 26(2):145, 1971. 2 + +[52] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2 +[53] Haitao Tian, Shiru Qu, and Pierre Payeur. A prototypical knowledge oriented adaptation framework for semantic segmentation. IEEE Transactions on Image Processing, 31:149-163, 2021. 2 +[54] Zhuotao Tian, Xin Lai, Li Jiang, Shu Liu, Michelle Shu, Hengshuang Zhao, and Jiaya Jia. Generalized few-shot semantic segmentation. In CVPR, 2022. 2 +[55] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018. 2 +[56] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In CVPR, 2017. 1 +[57] SS Vallender. Calculation of the wasserstein distance between probability distributions on the line. Theory of Probability & Its Applications, 18(4):784-786, 1974. 4 +[58] Girish Varma, Anbumani Subramanian, Anoop Namboodiri, Manmohan Chandraker, and CV Jawahar. Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In WACV, 2019. 5 +[59] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In NeurIPS, 2018. 2 +[60] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 2 +[61] Haoran Wang, Tong Shen, Wei Zhang, Ling-Yu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In ECCV, 2020. 2 +[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1 +[63] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 2 +[64] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In ICCV, 2021. 2, 5 +[65] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2 +[66] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 1, 2 +[67] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. In NeurIPS, 2020. 2 + +[68] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In ICLR, 2021. 2 +[69] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 2 +[70] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5 +[71] Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Episode-based prototype generating network for zero-shot learning. In CVPR, 2020. 2 +[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 1, 2 +[73] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 2 +[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 +[75] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2020. 1, 2, 3 +[76] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 2, 4 +[77] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 2 \ No newline at end of file diff --git a/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/images.zip b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9203da2e2816315a807fb58fbb1334caca150ed2 --- /dev/null +++ b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92bd2961438b0b56edbffae016aca3a0baff13921c8bae7fa6c0792f9b433e81 +size 795802 diff --git a/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/layout.json b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2036eadd166c5cb219156ddb3660b9c4b1e8e13 --- /dev/null +++ b/2023/Style Projected Clustering for Domain Generalized Semantic Segmentation/layout.json @@ -0,0 +1,10826 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 529, + 121 + ], + "type": "text", + "content": "Style Projected Clustering for Domain Generalized Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "spans": [ + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "text", + "content": "Wei Huang" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "text", + "content": " Chang Chen" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "text", + "content": " Yong Li" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "text", + "content": " Jiacheng Li" + }, + { + "bbox": [ + 158, + 142, + 433, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 157, + 453, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 157, + 453, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 157, + 453, + 171 + ], + "type": "text", + "content": "Cheng Li² Fenglong Song² Youliang Yan² Zhiwei Xiong¹" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 171, + 488, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 171, + 488, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 488, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 171, + 488, + 186 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 106, + 171, + 488, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 171, + 488, + 186 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 150, + 187, + 441, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 187, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 150, + 187, + 441, + 199 + ], + "type": "text", + "content": "{weih527,jcee}@mail.ustc.edu.cn,zwxiong@ustc.edu.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 201, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 201, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 108, + 201, + 482, + 213 + ], + "type": "text", + "content": "{chenchang25,liyong156,licheng89,songfenglong,yanyouliang}@huawei.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 266, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 531 + ], + "type": "text", + "content": "Existing semantic segmentation methods improve generalization capability, by regularizing various images to a canonical feature space. While this process contributes to generalization, it weakens the representation inevitably. In contrast to existing methods, we instead utilize the difference between images to build a better representation space, where the distinct style features are extracted and stored as the bases of representation. Then, the generalization to unseen image styles is achieved by projecting features to this known space. Specifically, we realize the style projection as a weighted combination of stored bases, where the similarity distances are adopted as the weighting factors. Based on the same concept, we extend this process to the decision part of model and promote the generalization of semantic prediction. By measuring the similarity distances to semantic bases (i.e., prototypes), we replace the common deterministic prediction with semantic clustering. Comprehensive experiments demonstrate the advantage of proposed method to the state of the art, up to " + }, + { + "bbox": [ + 47, + 266, + 290, + 531 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 47, + 266, + 290, + 531 + ], + "type": "text", + "content": " mIoU improvement in average on unseen scenarios. Code and models are available at https://gitee.com/mindspore/models/tree/master/research/cv/SPC-Net." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 554, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 658 + ], + "type": "text", + "content": "Domain generalization methods aim to promote the performance of model (trained on source datasets), when applying it to unseen scenarios (target domains) [9, 19, 29, 36, 62, 74, 75]. Recently, domain generalization for semantic segmentation (DGSS) has attracted increasingly more attention due to the rise of safety-critical applications, such as autonomous driving [3, 12, 22, 45]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 659, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 287, + 684 + ], + "type": "text", + "content": "Existing DGSS methods improve the pixel-wise generalization performance by learning domain-agnostic rep" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 239, + 547, + 415 + ], + "blocks": [ + { + "bbox": [ + 306, + 239, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 306, + 239, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 547, + 415 + ], + "type": "image", + "image_path": "86d35739b66d16476b3e032ae64bd899880ffd8e7193c0469049fb93a840fe31.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 418, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 547, + 473 + ], + "type": "text", + "content": "Figure 1. Illustration of instance normalization/whitening (IN/IW) [5, 20, 40] and our proposed style projected clustering method. IN and IW regularize image features from different domains to a canonical space (a-c). Our method builds style and semantic representation spaces based on the data from known domains (d)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 484, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 640 + ], + "type": "text", + "content": "resentations [5, 16, 20, 25, 40, 42, 66, 72]. Researches in this line share the similar goal in general, that is to capture the domain-invariant characteristics of object contents, and eliminates the domain-specific ones (i.e., image styles). As two representatives, Instance Normalization (IN) [56] and Instance Whitening (IW) [17] regularize image features from different domains to a canonical space, as illustrated in Fig. 1(a) and 1(b). Specifically, IN achieves center-level feature alignment via channel-wise feature normalization [33,40], and IW realizes uniform feature distribution by removing linear correlation between channels [5,41]. Moreover, the combination of these two methods is proposed in [42] for a better generalization, as shown in Fig. 1(c)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Nevertheless, feature regularization inevitably weakens the representation capability, as a part of feature information is eliminated. Theoretically, it works under a strong assumption that the eliminated information is strictly the domain-specific ones. Yet in practice, the perfect disentanglement between image style and content is difficult to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 287, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 287, + 703 + ], + "type": "text", + "content": "*This work was done during W. Huang's internship at Noah's Ark Lab." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3061" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "achieve. It means that a part of content features will also be eliminated in the process of feature regularization, and thus degrades the segmentation performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": "Instead of seeking common ground by feature regularization, we aim to address DGSS in a different way. In this paper, we propose style projection as an alternative, which utilizes the features from different domains as bases to build a better representation space, as shown in Fig. 1(d). The motivation of style projection comes from a basic concept of generalization, that is to represent unseen data based on the known ones. Specifically, following the common practice, we adopt the statistics (i.e., mean and variance) of features in channel dimension to represent image styles. The image styles from source domains are iteratively extracted and stored as the bases of representation. Then, we project the style of given unseen images into this representation space to promote generalization. This projection process is implemented as a weighted combination of stored style bases, where the similarity distance between styles are adopted as the weighting factors, i.e., " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " shown in Fig. 1(d)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 467 + ], + "type": "text", + "content": "Based on the projected style features, we further devise the decision part of model, which is elaborated for semantic segmentation. Typically, existing methods learn a parametric function to map pixel-wise features to semantic predictions. We replace this deterministic prediction with semantic clustering, where the class of each pixel is predicted by the minimal similarity distance to semantic bases, as shown in Fig. 1(d). Notably, it follows the same concept of style projection, that is to predict unseen data based on the known ones. More concretely, to facilitate the performance of semantic clustering, we propose a variant of contrastive loss to align the semantic bases of same classes and enhance discriminability between different classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 538 + ], + "type": "text", + "content": "We conduct comprehensive experiments on single- and multi-source settings to demonstrate the superior generalization of our method over existing DGSS methods. In addition, we visually analyze the effective representation of our proposed method for unseen images in both style and semantic spaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 540, + 281, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 540, + 281, + 552 + ], + "spans": [ + { + "bbox": [ + 59, + 540, + 281, + 552 + ], + "type": "text", + "content": "Contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 559, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 59, + 559, + 286, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 559, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 59, + 559, + 286, + 594 + ], + "type": "text", + "content": "- Beyond existing feature regularization methods, we propose style projected clustering, pointing out a new avenue to address DGSS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 599, + 286, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 599, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 59, + 599, + 286, + 634 + ], + "type": "text", + "content": "- We propose style projection, which projects unseen styles into the style representation space built on known domains for a better representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 639, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 639, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 639, + 286, + 685 + ], + "type": "text", + "content": "- We propose semantic clustering to predict the class of each pixel in unseen images by the similarity distance to semantic bases, which further improves the generalization capability for unseen domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 689, + 286, + 712 + ], + "type": "text", + "content": "- Our proposed method outperforms the current state of the arts on multiple DGSS benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 247 + ], + "type": "text", + "content": "Domain adaptation and generalization. To reduce the burden of pixel-wise annotations on target domains, domain adaptation (DA) technologies are proposed to narrow the domain gap between source and target domains via image translation [14, 24, 37], feature alignment [55, 60, 61], self-training [2, 39, 77] and meta-learning [13, 34] strategies. However, these DA methods require the access of data on target domains. Domain generalization (DG) aims to address a more practical problem where the target domain cannot be accessed. Numerous DG works have been proposed for image classification via style augmentation [19, 59, 68, 75], domain alignment [29, 31], feature disentanglement [27, 44] and meta-learning [9, 26, 28]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 247, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 521 + ], + "type": "text", + "content": "Domain generalization for semantic segmentation. Similar to image classification, DG for semantic segmentation (DGSS) methods are proposed to learn domain-agnostic representations, including style augmentation [16, 25, 43, 72], feature normalization/whitening [5, 40, 42, 66] and meta-learning [20]. To avoid overfitting on source domains, DRPC [72] and FSDR [16] adopt style augmentations in the image space to extend the number of source samples, while WildNet [25] realizes it in the feature space with the aid of ImageNet [8]. Alternatively, normalization and whitening are investigated to achieve distribution alignment between different domains. IBN-Net [40] and RobustNet [5] adopt instance normalization and whitening, respectively, to eliminate the specific style information of each domain. Furthermore, SAN-SAW [42] proposes semantic-aware instance normalization and whitening to enhance the distinguishability between classes. In addition, PintheMem [20] combines the memory-guided network with the meta-learning strategy and obtains competitive performances. Different from these DGSS methods, our method embraces the differences from multiple known domains and takes advantage of their diversity to build a better representation space, realizing the representation of unseen images by the known data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Prototype learning. Inspired by the cognitive psychology that human use the knowledge learned in the past to judge the class of unknown things [51,69], prototype-based classification methods have attracted increasing attention, where the class of unknown images is determined by its nearest neighbors in the feature space [7, 10]. Owing to its excellent interpretability and generalizability, prototype learning shows good potential in many fields, such as few-shot learning [1, 52], zero-shot learning [67, 71], unsupervised learning [30, 65]. Recently, prototype learning is also introduced in the dense prediction task, including supervised [76], few-shot [54, 63] and domain adaptive [53, 73] semantic segmentation. To facilitate the learning of prototypes, metric learning [23, 50, 64] is often adopted to pull samples belonging to the same class together and push those of different classes away from each other in the embedding (i.e., feature) space." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3062" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 265 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 265 + ], + "type": "image", + "image_path": "69d247d644081a0f51d0c428aadc22f240ddd97d5ea3345bb41629a57fd7cc1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": "Figure 2. The framework of style projected clustering, which consists of two components, i.e., style projection and semantic clustering. We iteratively extract the style and semantic information of seen domains as style bases " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "(p_m^\\mu, p_m^\\sigma)" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " and semantic bases " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "p_m^c" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": ". In style projection, we first calculate the similarity between the unseen style " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "(\\mu_m, \\sigma_m)" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " from the shallow feature " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "F_m^s" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " and style bases " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "(p_m^\\mu, p_m^\\sigma)" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " as weighted factors " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\lambda_m" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": ". Then, the weighted combination of style bases " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "(\\mu_m', \\sigma_m')" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " is projected on " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "F_m^n" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " to obtain the projected feature " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "F_m^r" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": ". In semantic clustering, we calculate the similarity between pixel embeddings in the deep feature " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "F_m^d" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " and semantic bases " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "p_m^c" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": ". Then, the class of each pixel is determined by the nearest semantic base. During the training phase, the cross-entropy loss " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": ", variance loss " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{var}" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " and discrimination loss " + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 46, + 270, + 548, + 350 + ], + "type": "text", + "content": " are adopted to supervise the learning of style and semantic bases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 360, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 288, + 411 + ], + "type": "text", + "content": "Similar to these methods, we adopt the form of prototypes (i.e. bases) to represent semantics. Yet these semantic bases are learned in a different way to facilitate domain generalization, by using a new variant of contrastive loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 198, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 198, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 198, + 434 + ], + "type": "text", + "content": "3. Style Projected Clustering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 547 + ], + "type": "text", + "content": "The overall architecture of our proposed method is depicted in Fig. 2, which consists of two components, i.e., style projection and semantic clustering. In style projection, we project the unseen style into the style representation space built on style bases, according to the similarity between the unseen style and style bases. In semantic clustering, we estimate the similarity between pixel embeddings and semantic bases (i.e., prototypes) to determine the class of pixels in unseen images by the nearest semantic base." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 552, + 173, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 173, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 173, + 564 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "In the domain generalized semantic segmentation problem, we are given " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " source domains " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\mathcal{S}_1,\\mathcal{S}_2,\\dots,\\mathcal{S}_M\\}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " that are from multiple datasets with different data distributions. The " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "-th source domain " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_m" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " can be represented as " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_m = \\{(x_m,y_m)\\}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "x_{m}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " is an image from the " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "-th source domain, " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "y_{m}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " is the corresponding pixel-wise label, " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " is the number of semantic classes, " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " are the height and width of the image " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", respectively. In this work, our goal is to train a semantic segmentation model " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " to obtain the best generalization performance on multiple target domains " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " which cannot be accessed during the training phase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 359, + 404, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 359, + 404, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 359, + 404, + 373 + ], + "type": "text", + "content": "3.2. Style Projection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": "The style difference of images is the main factor leading to the domain shift, which limits the generalization ability of the learned model. Pioneering works [11, 18, 40, 75] have demonstrated that the feature distribution shift caused by style differences lies mainly in shallow layers of networks. It also shows that the shallow feature distribution of networks can reflect the style information of the input image " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": ". Thus, existing works always adopt the channel-wise mean and variance of the shallow feature to represent the style distribution of " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " [18, 25]. Following these works, let " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "F_{m}^{s} \\in \\mathbb{R}^{D \\times H_{s} \\times W_{s}}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " be the shallow feature of " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " from the network " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " denotes the number of channels. The channel-wise mean " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\mu_{m} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "\\sigma_{m} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " of the feature " + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "inline_equation", + "content": "F_{m}^{s}" + }, + { + "bbox": [ + 304, + 378, + 547, + 547 + ], + "type": "text", + "content": " can be calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 550, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 550, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 331, + 550, + 545, + 582 + ], + "type": "interline_equation", + "content": "\\mu_ {m} = \\frac {1}{H _ {s} W _ {s}} \\sum_ {h = 1} ^ {H _ {s}} \\sum_ {w = 1} ^ {W _ {s}} F _ {m} ^ {s}, \\tag {1}", + "image_path": "30a615b55c46eff6be27e4860c83a3338b3ad1855640bdd9c04c648e85b930de.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 333, + 578, + 518, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 578, + 518, + 604 + ], + "spans": [ + { + "bbox": [ + 333, + 578, + 518, + 604 + ], + "type": "interline_equation", + "content": "\\sigma_ {m} = \\sqrt {\\frac {1}{H _ {s} W _ {s}} \\sum_ {h = 1} ^ {H _ {s}} \\sum_ {w = 1} ^ {W _ {s}} (F _ {m} ^ {s} - \\mu_ {m}) ^ {2}}.", + "image_path": "e5a13be7b0aa96e30d7f5535876a20071e0038fd9994f10779a3682b5267ab5b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 608, + 546, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 546, + 655 + ], + "type": "text", + "content": "To eliminate the specific style information of images, instance normalization [40] is adopted to standardize the feature " + }, + { + "bbox": [ + 306, + 608, + 546, + 655 + ], + "type": "inline_equation", + "content": "F_{m}^{s}" + }, + { + "bbox": [ + 306, + 608, + 546, + 655 + ], + "type": "text", + "content": " to a standard distribution (i.e., zeros mean and one standard deviation) as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 660, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 660, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 388, + 660, + 545, + 685 + ], + "type": "interline_equation", + "content": "F _ {m} ^ {n} = \\frac {F _ {m} ^ {s} - \\mu_ {m}}{\\sigma_ {m} + \\epsilon}, \\tag {2}", + "image_path": "e887cae7302d541a70e899c40a619f6329eba19766403ed465c07b4e7a43e88a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "F_{m}^{n}" + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "content": " stands for the normalized feature, and " + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "content": " is a small value to avoid division by zero." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3063" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "content": "Although instance normalization achieves to remove the specific style information of images, it also eliminates the natural differences between domains, which weakens the representation for target domains and produces limited generalization performance. Therefore, to preserve the specific style information of each domain, we propose style bases " + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "inline_equation", + "content": "P_{sty} = \\{(p_m^\\mu ,p_m^\\sigma)\\}_{m = 1}^M" + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "content": " to store the style information of source domains, and then leverage the preserved style bases " + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "inline_equation", + "content": "P_{sty}" + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "content": " to build a style representation space, realizing the projection of unseen style, as shown in Fig. 2. Specifically, we first leverage Wasserstein distance [57] to estimate the style distribution discrepancy between the input image " + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "content": "-th style bases " + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "inline_equation", + "content": "(p_m^\\mu ,p_m^\\sigma)" + }, + { + "bbox": [ + 47, + 72, + 289, + 229 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 233, + 287, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 233, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 63, + 233, + 287, + 248 + ], + "type": "interline_equation", + "content": "d _ {m} = | | \\mu_ {m} - p _ {m} ^ {\\mu} | | _ {2} ^ {2} + (\\sigma_ {m} ^ {2} + p _ {m} ^ {\\sigma 2} - 2 \\sigma_ {m} p _ {m} ^ {\\sigma}), \\qquad (3)", + "image_path": "811a16440780378ebfa2b823379b6295f62bb1a275afda9212a464b28af5ce4c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": " denotes the distribution distance between the current image " + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": "-th source domain. Then, we use the reciprocal of " + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": " to characterize the similarity between " + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 254, + 287, + 301 + ], + "type": "text", + "content": "-th style bases as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 304, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 304, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 99, + 304, + 287, + 334 + ], + "type": "interline_equation", + "content": "\\lambda_ {m} = \\frac {\\exp \\left(1 / \\left(1 + d _ {m}\\right)\\right)}{\\sum_ {m = 1} ^ {M} \\exp \\left(1 / \\left(1 + d _ {m}\\right)\\right)}, \\tag {4}", + "image_path": "15ef2a6a143ea1a4d8b76f53178e2125b5fd3043ccadbe842c833ecb6fc9e16d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "text", + "content": "where the softmax operation is utilized to make the sum of " + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\lambda = \\{\\lambda_m|m = 1,2,\\dots,M\\}" + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "text", + "content": " equal to 1. Based on the estimated similarity " + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "text", + "content": " , we can obtain the projected style " + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "inline_equation", + "content": "(\\mu_{m}^{\\prime},\\sigma_{m}^{\\prime})" + }, + { + "bbox": [ + 47, + 337, + 287, + 385 + ], + "type": "text", + "content": " by the weighted sum of style bases as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 389, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 389, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 85, + 389, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\mu_ {m} ^ {\\prime} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} p _ {m} ^ {\\mu}, \\sigma_ {m} ^ {\\prime} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} p _ {m} ^ {\\sigma}. \\tag {5}", + "image_path": "9fccf0801549779afd227a5781a247a89a886ff2a7c7c1b7ad8dc71ac9619f8f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "text", + "content": "Finally, following previous works [11, 18, 19, 25], we inject the projected style " + }, + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "inline_equation", + "content": "(\\mu_{m}^{\\prime},\\sigma_{m}^{\\prime})" + }, + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "text", + "content": " into the normalized feature " + }, + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "inline_equation", + "content": "F_{m}^{n}" + }, + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "text", + "content": " to obtain the projected feature as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 467, + 287, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 467, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 121, + 467, + 287, + 482 + ], + "type": "interline_equation", + "content": "F _ {m} ^ {r} = \\sigma_ {m} ^ {\\prime} F _ {m} ^ {n} + \\mu_ {m} ^ {\\prime}. \\tag {6}", + "image_path": "da5c9edbf35a8391ac422f632d4e2f7da02db9c1be427ecdec2ed872ab798936.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 487, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 287, + 521 + ], + "type": "text", + "content": "During the training phase, we adopt the momentum update strategy to achieve the online collection of style information as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 112, + 525, + 287, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 525, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 112, + 525, + 287, + 543 + ], + "type": "interline_equation", + "content": "p _ {m} ^ {\\mu} = \\alpha p _ {m} ^ {\\mu} + (1 - \\alpha) \\mu_ {m}, \\tag {7}", + "image_path": "72583bfd378e8540a630c726ee9af71cd1ef5fc8fb7d75eb6698c638c6358a85.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 541, + 221, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 541, + 221, + 554 + ], + "spans": [ + { + "bbox": [ + 113, + 541, + 221, + 554 + ], + "type": "interline_equation", + "content": "p _ {m} ^ {\\sigma} = \\alpha p _ {m} ^ {\\sigma} + (1 - \\alpha) \\sigma_ {m},", + "image_path": "d1dda68c0fab4aed2b034e6875f76318a8e3ed0b9792b9f0cc685b27f97cc2c2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": " is a momentum coefficient. In addition, we randomly initialize " + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "inline_equation", + "content": "P_{sty}" + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": " to start training, where " + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "inline_equation", + "content": "p_m^\\mu" + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "inline_equation", + "content": "p_m^\\sigma" + }, + { + "bbox": [ + 47, + 558, + 287, + 617 + ], + "type": "text", + "content": " are initialized with zero-mean and one-mean distribution, respectively. By Eq. 7, we realize the style statistic of source domains and store it as style bases efficiently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "After style projection, the projected feature " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_{m}^{r}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " is input into the next layer of the network " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ". Our style projection is designed as a plug-and-play module that can be applied behind any network layer. However, as the layer is deeper, the style information loosens while the semantic information plays a more important role. Thus, in this work, style projection is only used in the first two layers of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " to obtain the best generalization performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 424, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 424, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 424, + 85 + ], + "type": "text", + "content": "3.3. Semantic Clustering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": "To obtain the final pixel-wise predictions, we further propose semantic clustering on the deep feature extracted by the network " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "F_{m}^{d}\\in \\mathbb{R}^{D\\times H_{d}\\times W_{d}}" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " be the deep feature of the input image " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": ". Existing DGSS methods generically apply a learnable segmentation classifier " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\phi_{cls}" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "F_{m}^{d}" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " for the dense prediction. However, the parameters of " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\phi_{cls}" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " is learned on the deep features of source domains " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": ", and thus its generalization ability on target domain " + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": " is limited. In addition, the semantic information between different domains is implicitly encoded in the same parameter space, which causes the specific semantic information of domains to be eliminated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "content": "Based on the concept of style bases, we introduce semantic bases " + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "inline_equation", + "content": "P_{sem} = \\{p_m^c\\}_{c,m=1}^{C,M}" + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "content": " to preserve the semantic information of each domain and each class, where " + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "inline_equation", + "content": "p_m^c \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "content": " is the cluster center of training pixel embeddings belonging to the " + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "content": "-th class from the " + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "content": "-th source domain in the feature space. Following the prototype theory [7, 10, 76], the class of each pixel embedding " + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "inline_equation", + "content": "e \\in F_m^d" + }, + { + "bbox": [ + 304, + 234, + 545, + 331 + ], + "type": "text", + "content": " can be determined by its nearest semantic bases as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 336, + 545, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 336, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 318, + 336, + 545, + 353 + ], + "type": "interline_equation", + "content": "c (e) = c ^ {*}, \\text {w i t h} \\left(c ^ {*}, m ^ {*}\\right) = \\operatorname {a r g m i n} _ {c, m} \\left\\{d _ {m} ^ {c} \\right\\} _ {c, m = 1} ^ {C, M}, \\tag {8}", + "image_path": "45f314e5a2ebc14b6a6ffb2aba7224bc045f64ab55b469df68e4bfd588a2eed1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "d_m^c = -\\cos(e, p_m^c)" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": " is the negative cosine distance used to estimate the similarity between the current embedding " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": " and semantic bases " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "p_m^c" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": ". In this work, the pixel embedding " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": " and semantic bases " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "p_m^c" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": " are both " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": "-normalized. Therefore, the similarity distance can be simply formulated as " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "d_m^c = -ep_m^c" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": ". Different from the learnable segmentation classifier " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\phi_{cls}" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "inline_equation", + "content": "P_{sem}" + }, + { + "bbox": [ + 304, + 357, + 545, + 477 + ], + "type": "text", + "content": " not only explicitly captures characteristic properties of each class from each domain, but also determines the class of pixels in unseen images without introducing extra learnable parameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "text", + "content": "To facilitate the training of the network " + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "text", + "content": " during the training phase, we estimate the probability value of pixel embedding " + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 477, + 545, + 514 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 369, + 517, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 517, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 369, + 517, + 545, + 547 + ], + "type": "interline_equation", + "content": "v (c | e) = \\frac {\\exp (- d ^ {c})}{\\sum_ {c = 1} ^ {C} \\exp (- d ^ {c})}, \\tag {9}", + "image_path": "22027b023ff69878ab9c7c05f6727007d87b6de17c348fbba30a230fabdd6905.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "inline_equation", + "content": "d^{c} = \\min_{m}\\{d_{m}^{c}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "text", + "content": " denotes the similarity between " + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "text", + "content": " and its closet semantic base belonging to class " + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "text", + "content": ". Then, we adopt the standard cross-entropy loss to supervise the training of the network " + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 551, + 545, + 599 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 324, + 604, + 545, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 604, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 324, + 604, + 545, + 638 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} = - \\frac {1}{H _ {d} W _ {d}} \\sum_ {h = 1} ^ {H _ {d}} \\sum_ {w = 1} ^ {W _ {d}} \\sum_ {c = 1} ^ {C} y _ {m} \\log (v (c | e)), \\tag {10}", + "image_path": "0fcb9544dc15e92c1af4eaa51936071d0b7b8aaeffb01814c0514f0cab8be2da.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "content": " is the pixel-wise label corresponding to the input image " + }, + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 304, + 642, + 545, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": "However, the naive cross-entropy loss only optimizes the relative relations between intra-class and inter-class distance, which ignores the absolute distance constraint between pixel embeddings and semantic bases. That is to say," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3064" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "we expect that the pixel embedding belonging to class " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": " is closer to the " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "-th semantic base and is farther away from the semantic bases belonging to other classes. Inspired by metric learning [21, 23], we further propose variance and discrimination terms as two extra training objectives. The former is an intra-class cluster that pulls the pixel embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "e_{m}^{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "-th source domain towards the semantic bases " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "p_{m}^{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 175, + 287, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 175, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 91, + 175, + 287, + 208 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v a r} = \\frac {1}{M C} \\sum_ {m = 1} ^ {M} \\sum_ {c = 1} ^ {C} \\left(1 - e _ {m} ^ {c} p _ {m} ^ {c}\\right) ^ {2}. \\tag {11}", + "image_path": "c7a7ef442b67bfa11b77794eb346a6f05854d553fe5a2d183d061f68d048cea0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": "The latter is designed in a contrastive learning way which encourages the current cluster center " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\overline{e}_m^c" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": " is closer to the " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": "-th semantic bases " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "p_{c+}" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": " (i.e., positive keys) and to be far away from semantic bases belonging to other class " + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "inline_equation", + "content": "p_{c-}" + }, + { + "bbox": [ + 47, + 213, + 287, + 274 + ], + "type": "text", + "content": " (i.e., negative keys):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 279, + 298, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 298, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 298, + 322 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} = \\frac {1}{M} \\sum_ {p _ {c +}} - \\log \\frac {\\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c +} / \\tau\\right)}{\\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c +} / \\tau\\right) + \\sum_ {p _ {c -}} \\exp \\left(\\bar {e} _ {m} ^ {c} p _ {c -} / \\tau\\right)}, \\tag {12}", + "image_path": "df5fdabdccdd0273cc7d4b8d0978d75f8f2856d2fbbd7d605b6a7b1c45e30701.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\bar{e}_m^c" + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "content": " is the cluster center (i.e., mean embedding) of pixel embedding " + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "inline_equation", + "content": "e_m^c" + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "content": " in the current feature " + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "inline_equation", + "content": "F_m^d" + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "content": " is a temperature hyper-parameter. By Eq. 12, we realize the alignment of semantic bases belonging to the same class " + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 329, + 287, + 448 + ], + "type": "text", + "content": " from different domains. Different from existing pixel-wise contrastive learning paradigm [64], the positive and negative samples in Eq. 12 are semantic bases rather than pixel embeddings. Thus, we don't need to construct a memory bank to store sufficient embedding samples, which also significantly reduces the computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 449, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 485 + ], + "type": "text", + "content": "To achieve the online collocation of semantic information from source domains, we adopt the same momentum update strategy to update semantic bases " + }, + { + "bbox": [ + 47, + 449, + 287, + 485 + ], + "type": "inline_equation", + "content": "P_{sem}" + }, + { + "bbox": [ + 47, + 449, + 287, + 485 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 493, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 493, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 113, + 493, + 287, + 507 + ], + "type": "interline_equation", + "content": "p _ {m} ^ {c} = \\alpha p _ {m} ^ {c} + (1 - \\alpha) \\bar {e} _ {m} ^ {c}, \\tag {13}", + "image_path": "64938e7522aac110ba3899f9e661882a47f8c694d5469b8d538cc68c109e848e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "type": "text", + "content": " the momentum coefficient. Like style bases, we also randomly initialize the semantic bases " + }, + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "type": "inline_equation", + "content": "P_{sem}" + }, + { + "bbox": [ + 47, + 514, + 287, + 551 + ], + "type": "text", + "content": " with zero-mean distribution to start our training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 557, + 178, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 178, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 178, + 570 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 599 + ], + "type": "text", + "content": "During the training phase, we combine above three loss terms for the end-to-end training as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 609, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 609, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 97, + 609, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {C E} + \\beta \\mathcal {L} _ {\\text {v a r}} + \\gamma \\mathcal {L} _ {\\text {d i s}}, \\tag {14}", + "image_path": "ab4792226dc7e766728cce301841b6d5d624543430cbcf78d0d4127423a583d8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "text", + "content": " are weighting coefficients to balance these three terms. For each training iteration, in addition to the parameter update of the network " + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 629, + 287, + 677 + ], + "type": "text", + "content": ", the style and semantic bases are also updated online by Eq. 7 and Eq. 13." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "During the inference phase, we leverage Eq. 8 to obtain final pixel-wise predictions by the nonparametric cluster of pixel embeddings outputted from the learned network " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 386, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 386, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 386, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 91, + 368, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 368, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 368, + 102 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "type": "text", + "content": "Synthetic datasets. GTAV [47] contains 24966 images with a resolution of " + }, + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "type": "inline_equation", + "content": "1914 \\times 1052" + }, + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "type": "text", + "content": " captured from the GTA-V game engine. Synthia [48] contains 9400 images with a resolution of " + }, + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "type": "inline_equation", + "content": "1280 \\times 760" + }, + { + "bbox": [ + 304, + 109, + 545, + 157 + ], + "type": "text", + "content": " generated from virtual urban scenes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "content": "Real-world datasets. IDD [58] contains 10004 images with an average resolution of " + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "inline_equation", + "content": "1678 \\times 968" + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "content": " captured from Indian roads. Cityscapes [6] contains 5000 fine annotated images with a resolution of " + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "content": " captured from 50 different cities primarily in Germany. BDD100K [70] contains 10000 image with a resolution of " + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "content": " captures from different locations in US. Mapillary [38] contains 25000 images with an average resolution of " + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "content": " captured from all around the world." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 273, + 440, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 440, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 440, + 285 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 292, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 472 + ], + "type": "text", + "content": "Following the previous work [5], we adopt DeepLabV3+ [4] with ResNet-50, ResNet-101 [15], MobileNetV2 [49] and ShuffleNetV2 [35] backbones as our segmentation networks, where all backbones are pre-trained on ImageNet [8]. During the training phase, we adopt the SGD optimizer [46] with a momentum of 0.9 and weight decay of " + }, + { + "bbox": [ + 304, + 292, + 545, + 472 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 304, + 292, + 545, + 472 + ], + "type": "text", + "content": ". The initial learning rate is set to 0.01 and is decreased using the polynomial scheduling with a power of 0.9. We train all models for 40K iterations, except for the three-source setting, the model is trained for 100K iterations. In addition to some common data augmentations used in [5], we adopt extra strong style augmentations to enrich the style information of urban-scene images [32], which aims to enhance the proposed style projection ability in networks. More details can be found in our supplementary materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 498, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 664 + ], + "type": "text", + "content": "Comparison methods. We extensively compare our proposed method against existing DGSS methods, which can be classified into three groups, including style augmentation (WildNet [25]), feature normalization/whitening (IBNNet [40], RobustNet [5] and SAN-SAW [42]), and meta-learning (MLDG [26] and PintheMem [20]). Since SAN-SAW [42] and WildNet [25] are only implemented on the single-source setting in their paper, we reproduce them on the multi-source setting to make a comparison. In particular, WildNet [25] utilizes the external dataset (i.e., ImageNet) to extend the style and content information of source domains. Thus, we re-implement it by replacing the external dataset with the source dataset for a fair comparison, which is marked with * in our tables." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Multi-source setting. To demonstrate the effectiveness of our proposed method, we first conduct contrast experiments on the multi-source DGSS setting, where multiple source domains can be efficiently used to build a diverse" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3065" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 537, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 537, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 537, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 537, + 224 + ], + "type": "table", + "html": "
MethodsPublicationCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaAvg.-SAvg.-A
\\( Baseline^{\\dagger} \\)-35.4625.0931.9430.8368.4867.9968.2445.79
\\( IBN-Net^{\\dagger} [40] \\)ECCV 201835.5532.1838.0935.2769.7266.9068.3148.49
\\( RobustNet^{\\dagger} [5] \\)CVPR 202137.6934.0938.4936.7668.2668.7768.5249.46
\\( Baseline^{\\ddagger} \\)-33.4229.0732.1931.5669.6363.9366.7845.65
\\( MLDG^{\\ddagger} [26] \\)AAAI 201838.8431.9535.6035.4664.6151.6958.1544.54
\\( PintheMem^{\\ddagger} [20] \\)CVPR 202244.5138.0742.7041.7665.8554.4960.1749.12
Baseline-36.0328.1532.6132.2669.3067.6168.4646.65
SAN-SAW [42]CVPR 202242.1337.7442.9140.9363.9862.5863.2849.87
WildNet [25]CVPR 202243.6539.9043.2842.2868.0563.9866.0251.77
WildNet* [25]CVPR 202239.3334.7641.0638.3869.7062.1165.9149.39
Ours-46.3643.1848.2345.9272.4674.8773.6757.02
", + "image_path": "70204edbfc3f84e6bb1cd57dbf4376eba88795412d97f9709b7dbffda5a1b339.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 294, + 525, + 432 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "text", + "content": "Table 1. Source " + }, + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "inline_equation", + "content": "(\\mathbf{G} + \\mathbf{S})\\rightarrow" + }, + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "text", + "content": " Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) datasets. The best and second best results are highlighted and underlined. Avg.-T, Avg.-S and Avg.-A denote the average results on target, source and all domains, respectively. Results with the " + }, + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 228, + 547, + 284 + ], + "type": "text", + "content": " sign are from [5] and [20], respectively. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 294, + 525, + 432 + ], + "lines": [ + { + "bbox": [ + 67, + 294, + 525, + 432 + ], + "spans": [ + { + "bbox": [ + 67, + 294, + 525, + 432 + ], + "type": "table", + "html": "
MethodsCityscapesBDD100KMapillaryAvg.-TGTAVSynthiaIDDAvg.-SAvg.-A
Baseline‡52.5147.4754.7051.5670.3167.1371.5669.6760.61
IBN-Net‡[40]54.3948.9156.0653.1270.7363.6871.0268.4860.80
RobustNet‡[5]54.7049.0056.9053.5370.0666.4071.0269.1661.35
MLDG‡[26]54.7648.5255.9453.0769.5359.7967.7365.6859.38
PintheMem‡[20]56.5750.1858.3155.0269.9962.9967.5866.8560.94
Baseline54.1646.2455.5751.9968.3565.1270.0767.8559.92
SAN-SAW [42]54.8946.5056.3852.5964.4964.7666.3765.2158.90
WildNet [25]55.5850.3157.9354.6167.6561.3570.0766.3660.48
WildNet* [25]53.6148.9256.1852.9070.9859.6964.5265.0658.98
Ours57.9153.2661.6157.5974.6478.3576.0776.3566.97
", + "image_path": "2c93484a4b1a68749defa8875755f73956d4cf48a01bc75abb54eeeb75d6cb3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "type": "text", + "content": "Table 2. Source " + }, + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "type": "inline_equation", + "content": "(\\mathbf{G} + \\mathbf{S} + \\mathbf{I})\\rightarrow" + }, + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "type": "text", + "content": " Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with two synthetic (GTAV, Synthia) and one real (IDD) datasets. Results with the " + }, + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 435, + 546, + 460 + ], + "type": "text", + "content": " sign are from [20]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "type": "text", + "content": "representation space. As listed in Table 1, we quantitatively compare our results with existing DGSS methods on both target and source datasets, where all networks with the ResNet-50 backbone are trained with two synthetic datasets (i.e., GTAV and Synthia). Remarkably, compared with the state-of-the-art method (i.e., WildNet [20]), our method not only shows superior generalization capability on target datasets (up to " + }, + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "type": "text", + "content": " mIoU in average), but also significantly improve the performance on source datasets (up to " + }, + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 46, + 473, + 289, + 700 + ], + "type": "text", + "content": " mIoU), which demonstrates our method can enhance the representation ability of the learned model on both source and target domains. Furthermore, We provide visual prediction results for qualitative comparisons as shown in Fig. 3. Our method obtains the best visual results on different target datasets. Following [20], we add one real dataset (i.e., IDD) to source domains to further verify the superiority of our method on more source datasets. As listed in Table 2, our method also outperforms existing methods on both source and target domains by a large margin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Single-source setting. We further implement our method in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "content": "the single-source setting to make a comprehensive comparison, where all network with the ResNet-50 backbone are trained with one synthetic (i.e., GTAV) dataset. As listed in Table 3, our method shows superior generalization performances over existing DGSS methods. Compared with the naive baseline, our method brings approximately " + }, + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "content": " mIoU gains in average on target datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 557, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 629 + ], + "type": "text", + "content": "Different backbones. To demonstrate the wide applicability of our method, we compare our results with classic DGSS methods (i.e., IBN-Net [40] and RobustNet [5]) with different backbones. As listed in Table 4, our method shows superior performances on both large (i.e., ResNet-101) and lightweight (i.e., MobileNet and ShuffleNet) backbones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 635, + 407, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 407, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 407, + 647 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies with the ResNet-50 backbone on two source domains (i.e., GTAV and Synthia) as following." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "Proposed strategies. As listed in Table 5, our method shows the best generalization capability when two strategies" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3066" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 213 + ], + "type": "image", + "image_path": "9539866d645f4702f1a17dda1aef83bff39617833c8facdb157c2d84ccdc47fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 220, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 220, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 220, + 541, + 232 + ], + "type": "text", + "content": "Figure 3. Source " + }, + { + "bbox": [ + 50, + 220, + 541, + 232 + ], + "type": "inline_equation", + "content": "(\\mathbf{G} + \\mathbf{S}) \\rightarrow" + }, + { + "bbox": [ + 50, + 220, + 541, + 232 + ], + "type": "text", + "content": " Target (C, B, M): Visualization comparison with existing DGSS methods on three different target domains." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 245, + 272, + 429 + ], + "blocks": [ + { + "bbox": [ + 61, + 245, + 272, + 429 + ], + "lines": [ + { + "bbox": [ + 61, + 245, + 272, + 429 + ], + "spans": [ + { + "bbox": [ + 61, + 245, + 272, + 429 + ], + "type": "table", + "html": "
MethodsCBMAvg.-T
Baseline28.9525.1428.1827.42
IBN-Net [40]33.8532.3037.7534.63
RobustNet [5]36.5835.2040.3337.37
Baseline31.6026.7029.0029.10
MLDG [26]36.7032.1032.2033.67
PintheMem [20]41.0034.6037.4037.67
Baseline29.3225.7128.3327.79
SAN-SAW [42]39.7537.3441.8639.65
Baseline35.1629.7131.2932.05
WildNet [25]44.6238.4246.0943.04
Baseline32.0126.0429.3529.13
WildNet* [25]40.1034.8239.3838.10
Ours44.1040.4645.5143.36
", + "image_path": "a0027777b31df1a31f11c978b22119e47806ddea99264a1b2387a30e836a1f5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 432, + 287, + 498 + ], + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 498 + ], + "type": "text", + "content": "Table 3. Source " + }, + { + "bbox": [ + 46, + 432, + 287, + 498 + ], + "type": "inline_equation", + "content": "(\\mathbf{G})\\rightarrow" + }, + { + "bbox": [ + 46, + 432, + 287, + 498 + ], + "type": "text", + "content": " Target (C, B, M): Mean IoU(\\%) comparison of existing DGSS methods, where all networks with the ResNet-50 backbone are trained with the one synthetic (GTAV) dataset. * indicates that we replace the external dataset (i.e., ImageNet) used in WildNet [25] with the source dataset for a fair comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 517, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 577 + ], + "type": "text", + "content": "are adopted at the same time. Remarkably, compared with the first and second lines, we can find that style projection can approximately bring " + }, + { + "bbox": [ + 46, + 517, + 287, + 577 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 517, + 287, + 577 + ], + "type": "text", + "content": " mIoU gains in average over the baseline, which fully demonstrates its effectiveness for the generalization on unseen domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 579, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 700 + ], + "type": "text", + "content": "Different ways of style projection. As listed in Table 6, we investigate the effect of different ways of style projection. There are two intuitive ways as follows. One way is using the naive instance normalization to project images from different domains into a normalized feature space (i.e., Normalization). The other way is using the extracted style bases to directly substitute the unseen style (i.e., Substitution). We can find that the weighted combination of style bases can effectively enhance the representation of unseen style, producing better generalization on unseen domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": "Loss terms. As listed in Table 7, we conduct ablation ex" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 245, + 535, + 407 + ], + "blocks": [ + { + "bbox": [ + 315, + 245, + 535, + 407 + ], + "lines": [ + { + "bbox": [ + 315, + 245, + 535, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 535, + 407 + ], + "type": "table", + "html": "
MethodsCBMAvg.-T
MobileNetBaseline29.1620.2727.1925.24
IBN-Net [40]29.5826.0226.3227.31
RobustNet [5]30.6725.0228.2727.99
Ours39.8834.8338.9137.87
ShuffleNetBaseline29.4826.2731.3529.03
IBN-Net [40]32.6129.5533.2031.79
RobustNet [5]33.1531.9834.8533.33
Ours38.9734.6239.6637.75
ResNet-101Baseline34.7129.3237.7433.92
IBN-Net [40]39.1834.0039.3237.50
RobustNet [5]39.9634.9441.7238.87
Ours47.9343.6248.7946.78
", + "image_path": "f3e5188acc96ae35d43fba4736c114d340b55743491bba08f1640f64fadf70ae.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 411, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 434 + ], + "type": "text", + "content": "Table 4. Source " + }, + { + "bbox": [ + 306, + 411, + 545, + 434 + ], + "type": "inline_equation", + "content": "(\\mathbf{G} + \\mathbf{S})\\rightarrow" + }, + { + "bbox": [ + 306, + 411, + 545, + 434 + ], + "type": "text", + "content": " Target (C, B, M): Mean IoU(%) comparison of existing DGSS methods with different backbones." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 447, + 541, + 514 + ], + "blocks": [ + { + "bbox": [ + 309, + 447, + 541, + 514 + ], + "lines": [ + { + "bbox": [ + 309, + 447, + 541, + 514 + ], + "spans": [ + { + "bbox": [ + 309, + 447, + 541, + 514 + ], + "type": "table", + "html": "
Sty.-Pro.Sem.-Clu.CBMAvg.-T
36.0328.1532.6132.26
✓44.8742.4246.3744.55
✓39.0130.6035.1934.93
✓✓46.3643.1848.2345.92
", + "image_path": "c26501ff9bd3ea4e32f0f00c44d1cfb190a880090b6320e8dfad3c233c51b8c0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 517, + 545, + 551 + ], + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 551 + ], + "type": "text", + "content": "Table 5. Ablation results for each strategy used in our method. Sty.-Pro. and Sem.-Clu. indicate style projection and semantic clustering, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 323, + 564, + 526, + 620 + ], + "blocks": [ + { + "bbox": [ + 323, + 564, + 526, + 620 + ], + "lines": [ + { + "bbox": [ + 323, + 564, + 526, + 620 + ], + "spans": [ + { + "bbox": [ + 323, + 564, + 526, + 620 + ], + "type": "table", + "html": "
MethodsCBMAvg.-T
Normalization43.8340.9544.9243.23
Substitution45.0042.7945.1644.32
Ours46.3643.1848.2345.92
", + "image_path": "99def1dc6a8d64ff02aad51bddff06c5ec8c5f50697e0d57a0016169f965e663.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 312, + 624, + 539, + 635 + ], + "lines": [ + { + "bbox": [ + 312, + 624, + 539, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 539, + 635 + ], + "type": "text", + "content": "Table 6. Ablation results for different ways of style projection." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "periments to demonstrate the effectiveness of two complementary loss functions in Eq. 11 and Eq. 12. Compared with the naive cross-entropy loss, adding any complementary loss can bring the performance gain, which verifies each of them can effectively supplement the main loss " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3067" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 281, + 137 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 281, + 137 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 281, + 137 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 281, + 137 + ], + "type": "table", + "html": "
LCELvarLdisCBMAvg.-T
✓44.0041.8245.9743.93
✓✓45.5742.7846.6144.99
✓✓45.9242.4247.0845.14
✓✓✓46.3643.1848.2345.92
", + "image_path": "a23371a55bed0d7a7f6a040083ad6d6529c0bf35f9e1a0c23e08662f69ea4a17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 158, + 277, + 267 + ], + "blocks": [ + { + "bbox": [ + 87, + 140, + 247, + 150 + ], + "lines": [ + { + "bbox": [ + 87, + 140, + 247, + 150 + ], + "spans": [ + { + "bbox": [ + 87, + 140, + 247, + 150 + ], + "type": "text", + "content": "Table 7. Ablation results for each loss term." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 158, + 277, + 267 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 277, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 277, + 267 + ], + "type": "table", + "html": "
Methods# of ParamsGFLOPsTime (ms)
Baseline45.08M277.777.82
IBN-Net [40]45.08M277.828.74
RobustNet [5]45.08M277.789.48
MLDG [26]45.08M277.779.67
PintheMem [20]45.28M278.3111.64
SAN-SAW [42]25.63M421.8657.58
WildNet [25]45.21M277.168.61
Ours45.22M286.099.98
", + "image_path": "24f86139da956e1c32af99905a3d9b662ed4f93055faf75a58adec75d58e9753.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 271, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 304 + ], + "type": "text", + "content": "Table 8. Comparison of computational cost. Tested with the image size of " + }, + { + "bbox": [ + 46, + 271, + 287, + 304 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 46, + 271, + 287, + 304 + ], + "type": "text", + "content": " on one NVIDIA Tesla V100 GPU. We average the inference time over 500 trials." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 319, + 185, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 185, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 185, + 333 + ], + "type": "text", + "content": "5. Discussion and Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "type": "text", + "content": "Distribution analysis. We adopt the t-SNE visualization tool to analyze the effectiveness of our proposed style projection and semantic clustering strategies. As shown in Fig 4, we show the variations of style distribution between different domains before and after style projection. We can find that the style distribution of different domains is well separated before style projection (Fig. 4(a)), while their style distribution is approximately constrained between two style bases after style projection (Fig. 4(b)), which demonstrates style projection successfully projects unseen styles into the style representation space built on style bases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 604 + ], + "type": "text", + "content": "Furthermore, we visualize the semantic distribution between different classes and domains as shown in Fig 5. From Fig. 5(a), we can find that pixel samples belonging to the same class are well clustered while those belonging to different classes are well separated. In addition, the preserved semantic bases are approximately located in the cluster center of pixel samples. From Fig. 5(b), we can find that these pixel samples from different domains are well clustered according to their classes, which demonstrates our semantic clustering successfully achieves the class prediction between different domains by the preserved semantic bases." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Complexity of networks. As listed in Table 8, we compare the number of parameters and computational cost with existing DGSS methods. Since we need to store style and semantic bases and estimate the similarity between them and unseen images, the number of parameters and computational cost in our method are slightly higher than the naive baseline. However, our inference time is competitive to exiting DGSS methods due to the efficient implementation of distance measures by matrix multiplications." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 425, + 194 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 425, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 425, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 425, + 194 + ], + "type": "image", + "image_path": "bff090d503a3576984a0bd43676e2de63e23e6b6491b7d787cafdb091bd9b6bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 196, + 400, + 205 + ], + "lines": [ + { + "bbox": [ + 331, + 196, + 400, + 205 + ], + "spans": [ + { + "bbox": [ + 331, + 196, + 400, + 205 + ], + "type": "text", + "content": "(a) Before projection" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 70, + 543, + 194 + ], + "blocks": [ + { + "bbox": [ + 428, + 70, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 428, + 70, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 428, + 70, + 543, + 194 + ], + "type": "image", + "image_path": "e68ce15f23383c76af465946d99195077ea4ba8eee8c000c6de99c98af946612.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 196, + 516, + 205 + ], + "lines": [ + { + "bbox": [ + 452, + 196, + 516, + 205 + ], + "spans": [ + { + "bbox": [ + 452, + 196, + 516, + 205 + ], + "type": "text", + "content": "(b) After projection" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 268, + 426, + 393 + ], + "blocks": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "type": "text", + "content": "Figure 4. t-SNE visualization of style statistics between different domains before (a) and after (b) style projection, where the style statistics (concatenation of mean and variance) is computed from the first layer's feature map of the ResNet-50 trained on two synthetic datasets. Triangles indicate the preserved style bases." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 268, + 426, + 393 + ], + "lines": [ + { + "bbox": [ + 310, + 268, + 426, + 393 + ], + "spans": [ + { + "bbox": [ + 310, + 268, + 426, + 393 + ], + "type": "image", + "image_path": "d7682256312cef08236365d3fb7e0c670a5ebd0dd3fe73e394d9c1faa8f8a79a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 395, + 399, + 403 + ], + "lines": [ + { + "bbox": [ + 331, + 395, + 399, + 403 + ], + "spans": [ + { + "bbox": [ + 331, + 395, + 399, + 403 + ], + "type": "text", + "content": "(a) Class distribution" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 407, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 451 + ], + "type": "text", + "content": "Figure 5. t-SNE visualization of semantic statistics between different classes (a) and domains (b), where the semantic statistics is computed from the last layer's feature map. Triangles indicate the preserved semantic bases." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 428, + 269, + 544, + 393 + ], + "blocks": [ + { + "bbox": [ + 428, + 269, + 544, + 393 + ], + "lines": [ + { + "bbox": [ + 428, + 269, + 544, + 393 + ], + "spans": [ + { + "bbox": [ + 428, + 269, + 544, + 393 + ], + "type": "image", + "image_path": "ea1420d302b5320496147ee66b87378e9d5514a36061bad6bc4db5eb194418c4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 395, + 522, + 403 + ], + "lines": [ + { + "bbox": [ + 445, + 395, + 522, + 403 + ], + "spans": [ + { + "bbox": [ + 445, + 395, + 522, + 403 + ], + "type": "text", + "content": "(b) Domain distribution" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 480, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 647 + ], + "type": "text", + "content": "In this paper, we propose a novel style projected clustering method for domain generalized semantic segmentation, which achieves the style and semantic representation of unseen images based on known data. In particular, style projection projects arbitrary unseen styles into the style representation space of source domains and achieves the retention of specific style information between different domains. Semantic clustering predicts the class of each pixel by the minimal similarity distance to semantic bases, which realizes the semantic representation for unseen images and promotes the generalization ability. Through the evaluation on multiple urban-scene datasets, we demonstrate the superior generalization performance of our proposed method over existing DGSS methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grants 62131003 and 62021001. And we gratefully acknowledge the support of MindSpore (https://www.mindspore.cn/)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3068" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "text", + "content": "[1] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In ICML, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 127, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 287, + 159 + ], + "type": "text", + "content": "[2] Nikita Araslanov and Stefan Roth. Self-supervised augmentation consistency for adapting semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 162, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 286, + 205 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 208, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 286, + 251 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 254, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 254, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 254, + 286, + 308 + ], + "type": "text", + "content": "[5] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 311, + 286, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 311, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 311, + 286, + 365 + ], + "type": "text", + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 368, + 286, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 368, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 368, + 286, + 400 + ], + "type": "text", + "content": "[7] Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1):21-27, 1967. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 435 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 437, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 286, + 470 + ], + "type": "text", + "content": "[9] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 472, + 286, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 286, + 526 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 286, + 526 + ], + "type": "text", + "content": "[10] Salvador Garcia, Joaquin Derrac, Jose Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):417-435, 2012. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 529, + 286, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 286, + 561 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 564, + 286, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 286, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 286, + 596 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 599, + 286, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 286, + 642 + ], + "type": "text", + "content": "[13] Xiaoqing Guo, Chen Yang, Baopu Li, and Yixuan Yuan. Metacorrection: Domain-aware meta loss correction for unsupervised domain adaptation in semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 645, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 645, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 645, + 286, + 678 + ], + "type": "text", + "content": "[14] Jianzhong He, Xu Jia, Shuaijun Chen, and Jianzhuang Liu. Multi-source domain adaptation with collaborative learning for semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 128 + ], + "type": "text", + "content": "[17] Lei Huang, Dawei Yang, Bo Lang, and Jia Deng. Decorated batch normalization. In CVPR, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 162 + ], + "type": "text", + "content": "[18] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "text", + "content": "[19] Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. Style neophile: Constantly seeking novel styles for domain generalization. In CVPR, 2022. 1, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 239 + ], + "type": "text", + "content": "[20] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 242, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 264 + ], + "type": "text", + "content": "[21] Shu Kong and Charless C Fowlkes. Recurrent pixel embedding for instance grouping. In CVPR, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 266, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 319 + ], + "type": "text", + "content": "[22] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In WACV, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 321, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 353 + ], + "type": "text", + "content": "[23] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In ICCV, 2019. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 355, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 387 + ], + "type": "text", + "content": "[24] Seunghun Lee, Sunghyun Cho, and Sunghoon Im. Dranet: Disentangling representation and adaptation networks for unsupervised cross-domain adaptation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[25] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "type": "text", + "content": "[26] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2018. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 468, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 498 + ], + "type": "text", + "content": "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[28] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In ICCV, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 566 + ], + "type": "text", + "content": "[29] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "text", + "content": "[30] Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 602, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 643 + ], + "type": "text", + "content": "[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[32] Road Augmentation Library. https://github.com/UjjwalSaxena/Automold--Road-Augmentation-Library.5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[33] Ping Luo, Jiamin Ren, Zhanglin Peng, Ruimao Zhang, and Jingyu Li. Differentiable learning-to-normalize via switchable normalization. In ICLR, 2018. 1" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3069" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[34] Xinyu Luo, Jiaming Zhang, Kailun Yang, Alina Roitberg, Kunyu Peng, and Rainer Stiefelhagen. Towards robust semantic segmentation of accident scenes via multi-source mixed sampling and meta-learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[35] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 183 + ], + "type": "text", + "content": "[36] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 216 + ], + "type": "text", + "content": "[37] Zak Murez, Soheil Kolouri, David Kriegman, Ravi Ramamoorthi, and Kyungnam Kim. Image to image translation for domain adaptation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 249 + ], + "type": "text", + "content": "[38] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 287, + 293 + ], + "type": "text", + "content": "[39] Fei Pan, Inkyu Shin, Francois Rameau, Seokju Lee, and In So Kweon. Unsupervised intra-domain adaptation for semantic segmentation through self-supervision. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 294, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 294, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 49, + 294, + 287, + 326 + ], + "type": "text", + "content": "[40] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 327, + 287, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 287, + 359 + ], + "type": "text", + "content": "[41] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 361, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 392 + ], + "type": "text", + "content": "[42] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 393, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 393, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 287, + 436 + ], + "type": "text", + "content": "[43] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 437, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 437, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 49, + 437, + 287, + 469 + ], + "type": "text", + "content": "[44] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 471, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 502 + ], + "type": "text", + "content": "[45] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 504, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 287, + 535 + ], + "type": "text", + "content": "[46] Tiago Ramalho and Marta Garnelo. Adaptive posterior learning: few-shot learning with a surprise-based memory module. In ICLR, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 537, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 568 + ], + "type": "text", + "content": "[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 571, + 287, + 613 + ], + "type": "text", + "content": "[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 646 + ], + "type": "text", + "content": "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "type": "text", + "content": "[50] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[51] Herbert A Simon and Allen Newell. Human problem solving: The state of the theory in 1970. American Psychologist, 26(2):145, 1971. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[52] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[53] Haitao Tian, Shiru Qu, and Pierre Payeur. A prototypical knowledge oriented adaptation framework for semantic segmentation. IEEE Transactions on Image Processing, 31:149-163, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "text", + "content": "[54] Zhuotao Tian, Xin Lai, Li Jiang, Shu Liu, Michelle Shu, Hengshuang Zhao, and Jiaya Jia. Generalized few-shot semantic segmentation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "text", + "content": "[55] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[56] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In CVPR, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "text", + "content": "[57] SS Vallender. Calculation of the wasserstein distance between probability distributions on the line. Theory of Probability & Its Applications, 18(4):784-786, 1974. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "text", + "content": "[58] Girish Varma, Anbumani Subramanian, Anoop Namboodiri, Manmohan Chandraker, and CV Jawahar. Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In WACV, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 343, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 386 + ], + "type": "text", + "content": "[59] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "text", + "content": "[60] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "type": "text", + "content": "[61] Haoran Wang, Tong Shen, Wei Zhang, Ling-Yu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "text", + "content": "[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 566 + ], + "type": "text", + "content": "[63] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In ICCV, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "text", + "content": "[64] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. In ICCV, 2021. 2, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 633 + ], + "type": "text", + "content": "[65] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[66] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[67] Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3070" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 476 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[68] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In ICLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "text", + "content": "[69] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "type": "text", + "content": "[70] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 240 + ], + "type": "text", + "content": "[71] Yunlong Yu, Zhong Ji, Jungong Han, and Zhongfei Zhang. Episode-based prototype generating network for zero-shot learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 296 + ], + "type": "text", + "content": "[72] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 286, + 341 + ], + "type": "text", + "content": "[73] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 342, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 286, + 385 + ], + "type": "text", + "content": "[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 409 + ], + "type": "text", + "content": "[75] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2020. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 410, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 410, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 410, + 286, + 442 + ], + "type": "text", + "content": "[76] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking semantic segmentation: A prototype view. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 444, + 286, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 444, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 444, + 286, + 476 + ], + "type": "text", + "content": "[77] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3071" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_content_list.json b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..85fb374c9e630517290483860a49a7684b075f7b --- /dev/null +++ b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_content_list.json @@ -0,0 +1,1314 @@ +[ + { + "type": "text", + "text": "StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot Learning", + "text_level": 1, + "bbox": [ + 76, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuqian Fu $^{1}$ , Yu Xie $^{2}$ , Yanwei Fu $^{3}$ , Yu-Gang Jiang $^{1*}$", + "bbox": [ + 299, + 179, + 709, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University $^{2}$ Purple Mountain Laboratories, Nanjing, China. $^{3}$ School of Data Science, Fudan University $\\{\\text { fuyq20, yxie18, yanweifu, ygj } \\} @$ fudan.edu.cn", + "bbox": [ + 86, + 198, + 916, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 310, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN [10] empirically shows that spanning the style distribution of the source data helps alleviate this issue. However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles \"real\" and \"easy\", which still fall into the original set of the source styles. Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. Particularly, our style attack method synthesizes both \"virtual\" and \"hard\" adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. Code is available at https://github.com/lovelyqian/StyleAdv-CDFSL.", + "bbox": [ + 76, + 318, + 473, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 787, + 207, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper studies the task of Cross-Domain Few-Shot Learning (CD-FSL) which addresses the Few-Shot Learning (FSL) problem across different domains. As a general recipe for FSL, episode-based meta-learning strategy", + "bbox": [ + 75, + 813, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "has also been adopted for training CD-FSL models, e.g., FWT [48], LRP [42], ATA [51], and wave-SAN [10]. Generally, to mimic the low-sample regime in testing stage, meta learning samples episodes for training the model. Each episode contains a small labeled support set and an unlabeled query set. Models learn meta knowledge by predicting the categories of images contained in the query set according to the support set. The learned meta knowledge generalizes the models to novel target classes directly.", + "bbox": [ + 500, + 287, + 893, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Empirically, we find that the changes of visual appearances between source and target data is one of the key causes that leads to the domain gap in CD-FSL. Interestingly, waveSAN [10], our former work, shows that the domain gap issue can be alleviated by augmenting the visual styles of source images. Particularly, wave-SAN proposes to augment the styles, in the form of Adaptive Instance Normalization (AdaIN) [22], by randomly sampling two source episodes and exchanging their styles. However, despite the efficacy of wave-SAN, such a naive style generation method suffers from two limitations: 1) The swap operation makes the styles always be limited in the \"real\" style set of the source dataset; 2) The limited real styles further lead to the generated styles too \"easy\" to learn. Therefore, a natural question is whether we can synthesize \"virtual\" and \"hard\" styles for learning a more robust CD-FSL model? Formally, we use \"real/virtual\" to indicate whether the styles are originally presented in the set of source styles, and define \"easy/hard\" as whether the new styles make meta tasks more difficult.", + "bbox": [ + 496, + 426, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To that end, we draw inspiration from the adversarial training, and propose a novel meta Style Adversarial training method (StyleAdv) for CD-FSL. StyleAdv plays the minimax game in two iterative optimization loops of metaintraining. Particularly, the inner loop generates adversarial styles from the original source styles by adding perturbations. The synthesized adversarial styles are supposed to be more challenging for the current model to recognize, thus, increasing the loss. Whilst the outer loop optimizes the whole network by minimizing the losses of recognizing the images with both original and adversarial styles. Our ultimate goal is to enable learning a model that is robust to various styles, be-", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates corresponding author", + "bbox": [ + 94, + 886, + 269, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24575", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yond the relatively limited and simple styles from the source data. This can potentially improve the generalization ability on novel target domains with visual appearance shifts.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Formally, we introduce a novel style adversarial attack method to support the inner loop of StyleAdv. Inspired yet different from the previous attack methods [14,34], our style attack method perturbs and synthesizes the styles rather than image pixels or features. Technically, we first extract the style from the input feature map, and include the extracted style in the forward computation chain to obtain its gradient for each training step. After that, we synthesize the new style by adding a certain ratio of gradient to the original style. Styles synthesized by our style adversarial attack method have the good properties of \"hard\" and \"virtual\". Particularly, since we perturb styles in the opposite direction of the training gradients, our generation leads to the \"hard\" styles. Our attack method results in totally \"virtual\" styles that are quite different from the original source styles.", + "bbox": [ + 75, + 138, + 470, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Critically, our style attack method makes progressive style synthesizing, with changing style perturbation ratios, which makes it significantly different from vanilla adversarial attacking methods. Specifically, we propose a novel progressive style synthesizing strategy. The naive solution of directly plugging-in perturbations is to attack each block of the feature embedding module individually, which however, may results in large deviations of features from the high-level block. Thus, our strategy is to make the synthesizing signal of the current block be accumulated by adversarial styles from previous blocks. On the other hand, rather than attacking the models by fixing the attacking ratio, we synthesize new styles by randomly sampling the perturbation ratio from a candidate pool. This facilitates the diversity of the synthesized adversarial styles. Experimental results have demonstrated the efficacy of our method: 1) our style adversarial attack method does synthesize more challenging styles, thus, pushing the limits of the source visual distribution; 2) our StyleAdv significantly improves the base model and outperforms all other CD-FSL competitors.", + "bbox": [ + 75, + 367, + 470, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We highlight our StyleAdv is model-agnostic and complementary to other existing FSL or CD-FSL models, e.g., GNN [12] and FWT [48]. More importantly, to benefit from the large-scale pretrained models, e.g., DINO [2], we further explore adapting our StyleAdv to improve the Vision Transformer (ViT) [5] backbone in a non-parametric way. Experimentally, we show that StyleAdv not only improves CNN-based FSL/CD-FSL methods, but also improves the large-scale pretrained ViT model.", + "bbox": [ + 75, + 671, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we summarize our contributions. 1) A novel meta style adversarial training method, termed StyleAdv, is proposed for CD-FSL. By first perturbing the original styles and then forcing the model to learn from such adversarial styles, StyleAdv improves the robustness of CD-FSL models. 2) We present a novel style attack method with the novel", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "progressive synthesizing strategy in changing attacking ratios. Diverse \"virtual\" and \"hard\" styles thus are generated. 3) Our method is complementary to existing FSL and CD-FSL methods; and we validate our idea on both CNN-based and ViT-based backbones. 4) Extensive results on eight unseen target datasets indicate that our StyleAdv outperforms previous CD-FSL methods, building a new SOTA result.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 640, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-Domain Few-Shot Learning. FSL which aims at freeing the model from reliance on massive labeled data has been studied for many years [12, 25, 39, 41, 43, 45, 46, 56, 58]. Particularly, some recent works, e.g., CLIP [38], CoOp [65], CLIP-Adapter [11], Tip-Adapter [59], and PMF [19] explore promoting the FSL with large-scale pretrained models. Particularly, PMF contributes a simple pipeline and builds a SOTA for FSL. As an extended task from FSL, CD-FSL [1, 7–10, 15, 16, 23, 29, 32, 37, 42, 48, 51, 60, 67] mainly solves the FSL across different domains. Typical meta-learning based CD-FSL methods include FWT [48], LRP [42], ATA [51], AFA [20], and wave-SAN [10]. Specifically, FWT and LRP tackle CD-FSL by refining batch normalization layers and using the explanation model to guide training. ATA, AFA, and wave-SAN propose to augment the image pixels, features, and visual styles, respectively. Several transfer-learning based CD-FSL methods, e.g., BSCD-FSL (also known as Fine-tune) [16], BSR [33], and NSAE [32] have also been explored. These methods reveal that finetuning helps improving the performances on target datasets. Other works that introduce extra data or require multiple domain datasets for training include STARTUP [37], Meta-FDMixup [8], Me-D2N [9], TGDM [67], TriAE [15], and DSL [21].", + "bbox": [ + 496, + 234, + 893, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Attack. The adversarial attack aims at misleading models by adding some bespoke perturbations to input data. To generate the perturbations effectively, lots of adversarial attack methods have been proposed [6, 14, 26, 27, 34, 36, 55, 57]. Most of the works [6, 14, 34, 36] attack the image pixels. Specifically, FGSM [14] and PGD [34] are two most classical and famous attack algorithms. Several works [26, 27, 62] attack the feature space. Critically, few methods [57] attack styles. Different from these works that aim to mislead the models, we perturb the styles to tackle the visual shift issue for CD-FSL.", + "bbox": [ + 496, + 598, + 893, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Few-Shot Learning. Several attempts [13, 28, 30, 40, 52] that explore adversarial learning for FSL have been made. Among them, MDAT [30], AQ [13], and MetaAdv [52] first attack the input image and then train the model using the attacked images to improve the defense ability against adversarial samples. Shen et al. [40] attacks the feature of the episode to improve the generalization capability of FSL models. Note that ATA [51] and AFA [20], two CD-FSL methods, also adopt the adversarial learning.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24576", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, we are greatly different from them. ATA and AFA perturb image pixels or features, while we aim at bridging the visual gap by generating diverse hard styles.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Style Augmentation for Domain Shift Problem. Augmenting the style distribution for narrowing the domain shift issue has been explored in domain generation [31, 54, 66], image segmentation [3, 63], person re-ID [61], and CD-FSL [10]. Concretely, MixStyle [66], AdvStyle [63], DSU [31], and wave-SAN [10] synthesize styles without extra parameters via mixing, attacking, sampling from a Gaussian distribution, and swapping. MaxStyle [3] and L2D [54] require additional network modules and complex auxiliary tasks to help generate the new styles. Typically, AdvStyle [63] is the most related work to us. Thus, we highlight the key differences: 1) AdvStyle attacks styles on the image, while we attack styles on multiple feature spaces with a progressive attacking method; 2) AdvStyle uses the same task loss (segmentation) for attacking and optimization; in contrast, we use the classical classification loss to attack the styles, while utilize the task loss (FSL) to optimize the whole network.", + "bbox": [ + 75, + 138, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. StyleAdv: Meta Style Adversarial Training", + "text_level": 1, + "bbox": [ + 76, + 412, + 460, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Formulation. Episode $\\mathcal{T} = ((S, Q), Y)$ is randomly sampled as the input of each meta-task, where $Y$ represents the global class labels of the episode images with respect to $\\mathcal{C}^{tr}$ . Typically, each meta-task is formulated as an $N$ -way $K$ -shot problem. That is, for each episode $\\mathcal{T}$ , $N$ classes with $K$ labeled images are sampled as the support set $S$ , and the same $N$ classes with another $M$ images are used to constitute the query set $Q$ . The FSL or CD-FSL models predict the probability $P$ that the images in $Q$ belong to $N$ categories according to $S$ . Formally, we have $|S| = NK$ , $|Q| = NM$ , $|P| = NM \\times N$ .", + "bbox": [ + 75, + 439, + 468, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FGSM and PGD Attackers. We briefly summarize the algorithms for FGSM [14] and PGD [34], two most famous attacking methods. Given image $x$ with label $y$ , FGSM attacks the $x$ by adding a ratio $\\epsilon$ of signed gradients with respect to the $x$ resulting in the adversarial image $x^{adv}$ as,", + "bbox": [ + 75, + 607, + 468, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {a d v} = x + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} J (\\theta , x, y)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 696, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J(\\cdot)$ and $\\theta$ denote the object function and the learnable parameters of a classification model. PGD can be regarded as a variant of FGSM. Different from the FGSM that only attacks once, PGD attacks the image in an iterative way and sets a random start (abbreviated as RT) for $x$ as,", + "bbox": [ + 75, + 729, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0} ^ {a d v} = x + k _ {R T} \\cdot \\mathcal {N} (0, I), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 820, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} ^ {a d v} = x _ {t - 1} ^ {a d v} + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} J (\\theta , x, y)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 854, + 468, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k_{RT},\\epsilon$ are hyper-parameters. $\\mathcal{N}$ is Gaussian noises.", + "bbox": [ + 76, + 885, + 460, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of Meta Style Adversarial Learning", + "text_level": 1, + "bbox": [ + 500, + 90, + 885, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the performance degradation caused by the changing visual appearance, we tackle CD-FSL by promoting the robustness of models on recognizing various styles. Thus, we expose our FSL model to some challenging virtual styles beyond the image styles that existed in the source dataset. To that end, we present the novel StyleAdv adversarial training method. Critically, rather than adding perturbations to image pixels, we particularly focus on adversarially perturbing the styles. The overall framework of our StyleAdv is illustrated in Figure 1. Our StyleAdv contains a CNN/ViT backbone $E$ , a global FC classifier $f_{cls}$ , and a FSL classifier $f_{fsl}$ with learnable parameters $\\theta_E, \\theta_{cls}, \\theta_{fsl}$ , respectively. Besides, our core style attack method, a novel style extraction module, and the AdaIN are also included.", + "bbox": [ + 496, + 113, + 893, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, we learn the StyleAdv by solving a minimax game. Specifically, the minimax game shall involve two iterative optimization loops in each meta-train step. Particularly,", + "bbox": [ + 496, + 325, + 893, + 371 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Inner loop: synthesizing new adversarial styles by attacking the original source styles; the generated styles will increase the loss of the current network.", + "- Outer loop: optimizing the whole network by classifying source images with both original and adversarial styles; this process will decrease the loss." + ], + "bbox": [ + 500, + 381, + 893, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Style Extraction from CNNs and ViTs", + "text_level": 1, + "bbox": [ + 500, + 493, + 826, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptive Instance Normalization (AdaIN). We recap the vanilla AdaIN [22] proposed for CNN in style transfer. Particularly, AdaIN reveals that the instance-level mean and standard deviation (abbreviated as mean and std) convey the style information of the input image. Denoting the mean and std as $\\mu$ and $\\sigma$ , AdaIN (denoted as $\\mathcal{A}$ ) reveals that the style of $F$ can be transferred to that of $F_{tgt}$ by replacing the original style $(\\mu, \\sigma)$ with the target style $(\\mu_{tgt}, \\sigma_{tgt})$ :", + "bbox": [ + 496, + 516, + 893, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (F, \\mu_ {t g t}, \\sigma_ {t g t}) = \\sigma_ {t g t} \\frac {F - \\mu (F)}{\\sigma (F)} + \\mu_ {t g t}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 648, + 890, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Style Extraction for CNN Features. As shown in the upper part of Figure 1 (b), let $F \\in \\mathcal{R}^{B \\times C \\times H \\times W}$ indicates the input feature batch, where $B, C, H,$ and $W$ denote the batch size, channel, height, and width of the feature $F$ , respectively. As in AdaIN, the mean $\\mu$ and std $\\sigma$ of $F$ are defined as:", + "bbox": [ + 496, + 688, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu (\\mathrm {F}) _ {\\mathrm {b}, \\mathrm {c}} = \\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} F _ {b, c, h, w}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 773, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma (\\mathrm {F}) _ {\\mathrm {b}, \\mathrm {c}} = \\sqrt {\\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left(F _ {b , c , h , w} - \\mu_ {\\mathrm {b , c}} (F)\\right) ^ {2} + \\epsilon}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 829, + 890, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu ,\\sigma \\in \\mathcal{R}^{B\\times C}$", + "bbox": [ + 500, + 883, + 643, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24577", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c14f659c72ee51e20f7668e69c4d9eb75f8585782999416c1fc704fb586f91c.jpg", + "image_caption": [ + "(a) Overall Framework of Our StyleAdv Method" + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 578, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/454d49a5abffa07b993dd1a1099de2eb23c3ae59eab586f9c57962ca8559b7e8.jpg", + "image_caption": [ + "(b) Style Extraction For CNN/ViT Features", + "Figure 1. (a): Overview of StyleAdv method. The inner loop synthesizes adversarial styles, while the outer loop optimizes the whole network. (b): Style extraction for CNN-based and ViT-based features (illustration with $B = 1$ )." + ], + "image_footnote": [], + "bbox": [ + 586, + 87, + 887, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meta Information Extraction for ViT Features. We explore extracting the meta information of the ViT features as the manner of CNN. Intuitively, such meta information can be regarded as a unique \"style\" of ViTs. As shown in Figure 1 (b), we take an input batch data with image split into $P \\times P$ patches as an example. The ViT encoder will encode the batch patches into a class (cls) token ( $F_{cls} \\in \\mathcal{R}^{B \\times C}$ ) and a patch tokens ( $F_0 \\in \\mathcal{R}^{B \\times P^2 \\times C}$ ). To be compatible with AdaIN, we reshape the $F_0$ as $F \\in \\mathcal{R}^{B \\times C \\times P \\times P}$ . At this point, we can calculate the meta information for patch tokens $F$ as in Eq. 5 and Eq. 6. Essentially, note that the transformer integrates the positional embedding into the patch representation, the spatial relations thus could be considered still hold in the patch tokens. This supports us to reform the patch tokens $F_0$ as a spatial feature map $F$ . To some extent, this can also be achieved by applying the convolution on the input data via a kernel of size $P \\times P$ (as indicated by dashed arrows in Figure 1 (b)).", + "bbox": [ + 75, + 440, + 472, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inner Loop: Style Adversarial Attack Method", + "text_level": 1, + "bbox": [ + 76, + 724, + 467, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a novel style adversarial attack method - Fast Style Gradient Sign Method (Style-FGSM) to accomplish the inner loop. As shown in Figure 1, given an input source episode $(\\mathcal{T},Y)$ , we first forward it into the backbone $E$ and the FC classifier $f_{cls}$ producing the global classification loss $\\mathcal{L}_{cls}$ (as illustrated in the ① paths). During this process, a key step is to make the gradient of the style available. To achieve that, let $F_{\\mathcal{T}}$ denotes the features of $\\mathcal{T}$ , we obtain the style $(\\mu ,\\sigma)$ of $F_{\\mathcal{T}}$ as in Sec. 3.2. After that, we reform the original episode feature as $\\mathcal{A}(F_{\\mathcal{T}},\\mu ,\\sigma)$ . And the reformed", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature is actually used for the forward propagation. In this way, we include $\\mu$ and $\\sigma$ in our forward computation chain; and thus, we could access the gradients of them.", + "bbox": [ + 496, + 440, + 892, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the gradients in ② paths, we then attack $\\mu$ and $\\sigma$ as FGSM does - adding a small ratio $\\epsilon$ of the signed gradients with respect to $\\mu$ and $\\sigma$ , respectively.", + "bbox": [ + 496, + 484, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu^ {a d v} = \\mu + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\mu} J \\left(\\theta_ {E}, \\theta_ {f _ {c l s}}, \\mathcal {A} \\left(F _ {\\tau}, \\mu , \\sigma\\right), Y\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 542, + 892, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {a d v} = \\sigma + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\sigma} J \\left(\\theta_ {E}, \\theta_ {f _ {c l s}}, \\mathcal {A} \\left(F _ {\\mathcal {T}}, \\mu , \\sigma\\right), Y\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 571, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $J(\\cdot)$ is the cross-entropy loss between classification predictions and ground truth $Y$ , i.e., $\\mathcal{L}_{cls}$ . Inspired by the random start of PGD, we also add random noises $k_{RT} \\cdot \\mathcal{N}(0, I)$ to $\\mu$ and $\\sigma$ before attacking. $\\mathcal{N}(0, I)$ refers to Gaussian noises and $k_{RT}$ is a hyper-parameter. Our Style-FGSM enables us to generate both \"virtual\" and \"hard\" styles.", + "bbox": [ + 496, + 598, + 893, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Progressive Style Synthesizing Strategy: To prevent the high-level adversarial feature from deviating, we propose to apply our style-FGSM in a progressive strategy. Concretely, the embedding module $E$ has three blocks $E_{1}$ , $E_{2}$ , and $E_{3}$ , with the corresponding features $F_{1}$ , $F_{2}$ , and $F_{3}$ . For the first block, we use $(\\mu_{1}, \\sigma_{1})$ to denote the original styles of $F_{1}$ . The adversarial styles $(\\mu_{1}^{adv}, \\sigma_{1}^{adv})$ are obtained directly as in Eq. 7 and Eq. 8. For subsequent blocks, the attack signals on the current block $i$ are those accumulated from the block 1 to block $i - 1$ . Take the second block as an example, the block feature $F_{2}$ is not simply extracted by $E_{2}(F_{1})$ . Instead, we have $F_{2}' = E_{2}(F_{1}^{adv})$ , where $F_{1}^{adv} = \\mathcal{A}(F_{1}, \\mu_{1}^{adv}, \\sigma_{1}^{adv})$ . Attacking on $F_{2}'$ results in the adversarial styles $(\\mu_{2}^{adv}, \\sigma_{2}^{adv})$ . Accordingly, we gen", + "bbox": [ + 496, + 689, + 895, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24578", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "erate $(\\mu_3^{adv},\\sigma_3^{adv})$ for the last block. The illustration of the progressive attacking strategy is attached in the Appendix. Changing Style Perturbation Ratios: Different from the vanilla FGSM [14] or PGD [34], our style attacking algorithm is expected to synthesize new styles with diversity. Thus, instead of using a fixed attacking ratio $\\epsilon$ , we randomly sample $\\epsilon$ from a candidate list $\\epsilon_{list}$ as the current attacking ratio. Despite the randomness of $\\epsilon$ , we still synthesize styles in a more challenging direction, $\\epsilon$ only affects the extent.", + "bbox": [ + 75, + 90, + 468, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Outer Loop: Optimize the StyleAdv Network", + "text_level": 1, + "bbox": [ + 76, + 236, + 460, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each meta-train iteration with clean episode $\\mathcal{T}$ as input, our inner loop produces adversarial styles $(\\mu_1^{adv},\\sigma_1^{adv})$ $(\\mu_{2}^{adv},\\sigma_{2}^{adv})$ , and $(\\mu_3^{adv},\\sigma_3^{adv})$ . As in Figure 1, the goal of the outer loop is to optimize the whole StyleAdv with both the clean feature $F$ and the style attacked feature $F^{adv}$ utilized as the training data. Typically, the clean episode feature $F$ can be obtained directly as $E(T)$ as in 3 paths.", + "bbox": [ + 75, + 258, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In 4 paths, we obtain the $F^{adv}$ by transferring the original style of $F$ to the corresponding adversarial attacked styles. Similar with the progressive style-FGSM, we have $F_{1}^{adv} = \\mathcal{A}(E_{1}(\\mathcal{T}),\\mu_{1}^{adv},\\sigma_{1}^{adv})$ , $F_{2}^{adv} = \\mathcal{A}(E_{2}(F_{1}^{adv}),\\mu_{2}^{adv},\\sigma_{2}^{adv})$ , and $F_{3}^{adv} = \\mathcal{A}(E_{3}(F_{2}^{adv}),\\mu_{3}^{adv},\\sigma_{3}^{adv})$ . Finally, $F^{adv}$ is obtained by applying an average pooling layer to $F_{3}^{adv}$ . A skip probability $p_{skip}$ is set to decide whether to skip the current attacking. Conducting FSL tasks for both the clean feature $F$ and style attacked feature $F^{adv}$ results in two FSL predictions $P_{fsl}$ , $P_{fsl}^{adv}$ , and two FSL classification losses $\\mathcal{L}_{fsl}$ , $\\mathcal{L}_{fsl}^{adv}$ .", + "bbox": [ + 75, + 364, + 470, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, despite the styles of $F^{adv}$ shifts from that of $F$ , we encourage that the semantic content should be still consistent as in wave-SAN [10]. Thus we add a consistent constraint to the predictions of $P_{fsl}$ and $P_{fsl}^{adv}$ resulting in the consistent loss $\\mathcal{L}_{cons}$ as,", + "bbox": [ + 76, + 532, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s}} = \\mathrm {K L} \\left(P _ {f s l}, P _ {f s l} ^ {\\text {a d v}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 618, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{KL}()$ is Kullback-Leibler divergence loss. In addition, we have the global classification loss $\\mathcal{L}_{cls}$ . This ensures that $\\theta_{cls}$ is optimized to provide correct gradients for style-FGSM. The final meta-objective of StyleAdv is as,", + "bbox": [ + 75, + 647, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {f s l} + \\mathcal {L} _ {f s l} ^ {a d v} + \\mathcal {L} _ {c o n s} + \\mathcal {L} _ {c l s}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 718, + 468, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that our StyleAdv is model-agnostic and orthogonal to existing FSL and CD-FSL methods.", + "bbox": [ + 75, + 747, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Network Inference", + "text_level": 1, + "bbox": [ + 76, + 786, + 256, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Applying StyleAdv Directly for Inference. Our StyleAdv facilitates making CD-FSL model more robust to style shifts. Once the model is meta-trained, we can employ it for inference directly by feeding the testing episode into the $E$ and the $f_{cls}$ . The class with the highest probability will be taken as the predicted result.", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finetuning StyleAdv Using Target Examples. As indicated in previous works [16, 32, 33, 51], finetuning CD-FSL models on target examples helps improve the model performance. Thus, to further promote the performance of StyleAdv, we also equip it with the fintuning strategy forming an upgraded version (\"StyleAdv-FT\"). Specifically, as in ATA-FT [51], for each novel testing episode, we augment the novel support set to form pseudo episodes as training data for tuning the meta-trained model.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 248, + 633, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We take two CD-FSL benchmarks proposed in BSCD-FSL [16] and FWT [48]. Both of them take mini-Imagenet [39] as the source dataset. Two disjoint sets split from mini-Imagenet form $\\mathcal{D}^{tr}$ and $\\mathcal{D}^{eval}$ . Totally eight datasets including ChestX [53], ISIC [4, 47], EuroSAT [18], CropDisease [35], CUB [50], Cars [24], Places [64], and Plantae [49] are taken as novel target datasets. The former four datasets included in BSCD-FSL's benchmark cover medical images varying from X-ray to dermoscopic skin lesions, and natural images from satellite pictures to plant disease photos. While the latter four datasets that focus on more fine-grained concepts such as birds and cars are contained in FWT. These eight target datasets serve as testing set $\\mathcal{D}^{te}$ , respectively.", + "bbox": [ + 496, + 276, + 893, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Modules. For typical CNN based network, following previous CD-FSL methods [10, 42, 48, 51], ResNet-10 [17] is selected as the embedding module while GNN [12] is selected as the FSL classifier; For the emerging ViT based network, following PMF [19], we use the ViT-small [5] and the ProtoNet [41] as the embedding module and the FSL classifier, respectively. Note that, the ViT-small is pretrained on ImageNet1K by DINO [2] as in PMF. The $f_{cls}$ is built by a fully connected layer.", + "bbox": [ + 496, + 489, + 893, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The 5-way 1-shot and 5-way 5-shot settings are conducted. Taking ResNet10 as backbone, we meta train the network for 200 epochs, each epoch contains 120 meta tasks. Adam with a learning rate of 0.001 is utilized as the optimizer. Taking ViT-small as backbone, the meta train stage takes 20 epoch, each epoch contains 2000 meta tasks. The SGD with a initial learning rate of 5e-5 and 0.001 are used for optimize the $E(\\cdot)$ and the $f_{cls}$ , respectively. The $\\epsilon_{list}$ , $k_{RT}$ of Style-FGSM attacker are set as [0.8, 0.08, 0.008], $\\frac{16}{255}$ . The probability $p_{skip}$ of random skipping the attacking is chosen from {0.2, 0.4}. We evaluate our network with 1000 randomly sampled episodes and report average accuracy (\\%) with a 95% confidence interval. Both the results of our \"StyleAdv\" and \"StyleAdv-FT\" are reported. The details of the finetuning are attached in Appendix. ResNet-10 based models are trained and tested on a single GeForce GTX 1080, while ViT-small based models require a single NVIDIA GeForce RTX 3090.", + "bbox": [ + 496, + 628, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24579", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1a24a42a3b4105cb2081cae3c81e264932a8c6f108cf5b4dda2f2fba6e8cfa28.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Results of 5-way 1-shot/5-shot tasks. \"FT\" means whether the finetuning stage is employed. \"LargeP\" represents if large pretrained models are used for model initialization. \"RN10\" is short for \"ResNet-10\". * denotes results are reported by us. Results perform best are bolded. Whether based on ResNet-10 or ViT-small, our method outperforms other competitors significantly." + ], + "table_body": "
1-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--22.00±0.4632.02±0.6663.69±1.0364.48±1.0845.69±0.6831.79±0.5153.10±0.8035.60±0.5643.55
FWT [48]RN10--22.04±0.4431.58±0.6762.36±1.0566.36±1.0447.47±0.7531.61±0.5355.77±0.7935.95±0.5844.14
LRP [42]RN10--22.11±0.2030.94±0.3054.99±0.5059.23±0.5048.29±0.5132.78±0.3954.83±0.5637.49±0.4342.58
ATA [51]RN10--22.10±0.2033.21±0.4061.35±0.5067.47±0.5045.00±0.5033.61±0.4053.57±0.5034.42±0.4043.84
AFA [20]RN10--22.92±0.2033.21±0.3063.12±0.5067.61±0.5046.86±0.5034.25±0.4054.04±0.6036.76±0.4044.85
wave-SAN [10]RN10--22.93±0.4933.35±0.7169.64±1.0970.80±1.0650.25±0.7433.55±0.6157.75±0.8240.71±0.6647.37
StyleAdv (ours)RN10--22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
ATA-FT [51]RN10Y-22.15±0.2034.94±0.4068.62±0.5075.41±0.5046.23±0.5037.15±0.4054.18±0.5037.38±0.4047.01
StyleAdv-FT (ours)RN10Y-22.64±0.3535.76±0.5272.92±0.7580.69±0.2848.49±0.7235.09±0.5558.58±0.8341.13±0.6749.41
PMF* [19]ViT-smallYDINO/IN1K21.73±0.3030.36±0.3670.74±0.6380.79±0.6278.13±0.6637.24±0.5771.11±0.7153.60±0.6655.46
StyleAdv (ours)ViT-small-DINO/IN1K22.92±0.3233.05±0.4472.15±0.6581.22±0.6184.01±0.5840.48±0.5772.64±0.6755.52±0.6657.75
StyleAdv-FT (ours)ViT-smallYDINO/IN1K22.92±0.3233.99±0.4674.93±0.5884.11±0.5784.01±0.5840.48±0.5772.64±0.6755.52±0.6658.57
5-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--25.27±0.4643.94±0.6783.64±0.7787.96±0.6762.25±0.6544.28±0.6370.84±0.6552.53±0.5958.84
FWT [48]RN10--25.18±0.4543.17±0.7083.01±0.7987.11±0.6766.98±0.6844.90±0.6473.94±0.6753.85±0.6259.77
LRP [42]RN10--24.53±0.3044.14±0.4077.14±0.4086.15±0.4064.44±0.4846.20±0.4674.45±0.4754.46±0.4658.94
ATA [51]RN10--24.32±0.4044.91±0.4083.75±0.4090.59±0.3066.22±0.5049.14±0.4075.48±0.4052.69±0.4060.89
AFA [20]RN10--25.02±0.2046.01±0.4085.58±0.4088.06±0.3068.25±0.5049.28±0.5076.21±0.5054.26±0.4061.58
wave-SAN [10]RN10--25.63±0.4944.93±0.6785.22±0.7189.70±0.6470.31±0.6746.11±0.6676.88±0.6357.72±0.6462.06
StyleAdv (ours)RN10--26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
Fine-tune [16]RN10Y-25.97±0.4148.11±0.6479.08±0.6189.25±0.5164.14±0.7752.08±0.7470.06±0.7459.27±0.7061.00
ATA-FT [51]RN10Y-25.08±0.2049.79±0.4089.64±0.3095.44±0.2069.83±0.5054.28±0.5076.64±0.4058.08±0.4064.85
NSAE [32]RN10Y-27.10±0.4454.05±0.6383.96±0.5793.14±0.4768.51±0.7654.91±0.7471.02±0.7259.55±0.7464.03
BSR [33]RN10Y-26.84±0.4454.42±0.6680.89±0.6192.17±0.4569.38±0.7657.49±0.7271.09±0.6861.07±0.7664.17
StyleAdv-FT (ours)RN10Y-26.24±0.3553.05±0.5491.64±0.4396.51±0.2870.90±0.6356.44±0.6879.35±0.6164.10±0.6467.28
PMF [19]ViT-smallYDINO/IN1K27.2750.1285.9892.96-----
StyleAdv (ours)ViT-small-DINO/IN1K26.97±0.3347.73±0.4488.57±0.3494.85±0.3195.82±0.2761.73±0.6288.33±0.4075.55±0.5472.44
StyleAdv-FT (ours)ViT-smallYDINO/IN1K26.97±0.3351.23±0.5190.12±0.3395.99±0.2795.82±0.2766.02±0.6488.33±0.4078.01±0.5474.06
", + "bbox": [ + 76, + 88, + 893, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with the SOTAs", + "text_level": 1, + "bbox": [ + 76, + 497, + 334, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our StyleAdv/StyleAdv-FT against several most representative and competitive CD-FSL methods. Concretly, with the ResNet-10 (abbreviated as RN10) as backbone, totally nine methods including GNN [12], FWT [48], LRP [42], ATA [51], AFA [20], wave-SAN [10], Finetune [16], NSAE [32], and BSR [33] are introduced as our competitors. Among them, the former six competitors are meta-learning based method that used for inference directly, thus we compare our \"StyleAdv\" against them for a fair comparison. Typically, the GNN [12] works as a base model. The Fine-tune [16], NSAE [32], BSR [33], and ATA-FT [51] (formed by finetuning ATA) all require finetuning model during inference, thus our \"StyleAdv-FT\" is used. With the ViT as backbone, the most recent and competitive PMF (SOTA method for FSL) is compared. For fair comparisons, we follow the same pipeline proposed in PMF [19]. Note that we promote CD-FSL models with only one single source domain. Those methods that use extra training datasets, e.g., STARTUP [37], meta-FDMixup [8], and DSL [21] are not considered. The comparison results are given in Table 1.", + "bbox": [ + 75, + 521, + 472, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all results, our method outperforms all the listed CD-FSL competitors significantly and builds a new state of the art. Our StyleAdv-FT (ViT-small) on average achieves $58.57\\%$ and $74.06\\%$ on 5-way 1-shot and 5-shot, respectively. Our StyleAdv (RN10) and StyleAdv-FT (RN10) also", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "beats all the meta-learning based or transfer-learning (finetuning) based methods. Besides of the state-of-the-art accuracy, we also have other worth-mentioning observations. 1) We show that our StyleAdv method is a general solution for both CNN-based models and ViT-based models. Typically, based on ResNet10, our StyleAdv and StyleAdv-FT improve the base GNN by up to $4.93\\%$ and $8.44\\%$ on 5-shot setting. Based on ViT-small, at most cases, our StyleAdv-FT outperforms the PMF by a clear margin. More results of building StyleAdv upon other FSL or CD-FSL methods can be found in the Appendix. 2) Comparing FWT, LRP, ATA, AFA, waveSAN, and our StyleAdv, we find that StyleAdv performs best, followed by wave-SAN, then comes the AFA, ATA, FWT, and LRP. This phenomenon indicates that tackling CD-FSL by solving the visual shift problem is indeed more effective than other perspectives, e.g., adversarial training by perturbing the image features (AFA) or image pixels (ATA), transforms the normalization layers in FWT, and explanation guided training in LRP. 3) For the comparison between StyleAdv and wave-SAN that both tackles the visual styles, we notice that StyleAdv outperforms the wave-SAN in most cases. This demonstrates that the styles generated by our StyleAdv are more conducive to learning robust CD-FSL models than the style augmentation method proposed in wave-SAN. This justifies our idea of synthesizing more challenging (\"hard and virtual\") styles.", + "bbox": [ + 496, + 498, + 895, + 892 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24580", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4) Overall, the large-scale pretrained model promotes the CD-FSL obviously. Take 1-shot as an example, StyleAdv-FT (ViT-small) boosts the StyleAdv-FT (RN10) by $9.16\\%$ on average. However, we show that the performance improvement varies greatly on different target domains. Generally, for target datasets with relative small domain gap, e.g., CUB and Plantae, models benefit a lot; otherwise, the improvement is limited. 5) We also find that under the cross-domain scenarios, finetuning model on target domain, e.g., NSAE, BSR do show an advantage over purely meta-learning based methods, e.g., FWT, LRP, and wave-SAN. However, to finetune model using extremely few examples, e.g., 5-way 1-shot is much harder than on relatively larger shots. This may explain why those finetune-based methods do not conduct experiments on 1-shot setting.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Style-FGSM Attacker. To show the advantages of our progressive style synthesizing strategy and attacking with changing perturbation ratios, we compare our Style-FGSM against several variants and report the results in Figure 2. Specifically, for Figure 2 (a), we compare our style-FGSM against the variant that attacks the blocks individually. Results show that attacking in a progressive way exceeds the naive individual strategy in most cases. For Figure 2 (b), to demonstrate how the performance will be affected by fixed attacking ratios, we also conduct experiments with different $\\epsilon_{list}$ . Since we set the $\\epsilon_{list}$ as [0.8, 0.08, 0.008], three different choices including [0.8], [0.08], and [0.008] are selected. From the results, we first notice that the best result can be reached by a single fixed ratio. However, sampling the attacking ratio from a pool of candidates achieves the best result in most cases.", + "bbox": [ + 75, + 316, + 473, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cddee983b9fc92f8c19d546c7064fd9cc0641f9698878d69e9b84345c247e2d8.jpg", + "image_caption": [ + "Figure 2. Effectiveness of the progressive style synthesizing strategy and the changing style perturbation ratios. The 5-way 1-shot results are reported. Models are built on ResNet10 and GNN." + ], + "image_footnote": [], + "bbox": [ + 116, + 575, + 436, + 863 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51da6b7adbc4530030ff63f7ce1b527353b7445d2c4f2ddcfb218fe7407a2093.jpg", + "image_caption": [ + "Figure 3. Visualization of wave-SAN and StyleAdv. (a): synthesized images; (b): meta-training losses; (c): T-SNE results." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 848, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. More Analysis", + "text_level": 1, + "bbox": [ + 500, + 297, + 648, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of Hard Style Generation. To help understand the \"hard\" style generation of our method intuitively, as in Figure 3, we make several visualizations comparing StyleAdv against the wave-SAN. 1) As in Figure 3 (a), we show the stylized images generated by wave-SAN and our StyleAdv. The visualization is achieved by applying the style augmentation methods to input images. Specifically, for wave-SAN, the style is swapped with another randomly sampled source image; for StyleAdv, the results of attacking style with $\\epsilon = 0.08$ are given. We observe that wave-SAN tends to exchange the global visual appearance, e.g., the color of the input image randomly. By contrast, StyleAdv prefers to disturb the important regions that are key to recognizing the image category. For example, the fur of the cat and the key parts (face and feet) of the dogs. These observations intuitively support our claim that our StyleAdv synthesize more harder styles than wave-SAN. 2) To quantitatively evaluate whether our StyleAdv introduces more challenging styles into the training stage, as in Figure 3 (b), we visualize the meta-training loss. Results reveal that the perturbed losses of wave-SAN oscillate around the original loss, while StyleAdv increases the original loss obviously. These phenomenons further validate that we perturb data towards a more difficult direction thus pushing the limits of style generation to a large extent. 3) To further show the advantages of StyleAdv over wave-SAN, as shown in Figure 3 (c), we visualize the high-level features extracted by the meta-trained wave-SAN and StyleAdv. Five classes (denoted by different colors) of mini-Imagenet are selected. T-SNE is used for reducing the feature dimensions. Results demonstrate that StyleAdv enlarges the inter-class distances making classes more distinguishable.", + "bbox": [ + 496, + 323, + 895, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why Attack Styles Instead of Images or Features? A natural question may be why we choose to attack styles instead of other targets, e.g., the input image as in AQ [13], MDAT [30], and ATA [51] or the features as in Shen et al. [40] and AFA [20]? To answer this question, we compare our StyleAdv which attacks styles against attacking images", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24581", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43090eb3851e9b8ca4351b05486360e19a6adebccc1138eec84ddc08271753f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Attack TargetChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotImage22.71±0.3533.00±0.5367.00±0.8272.65±0.7548.15±0.7234.40±0.6057.89±0.8339.85±0.6446.96
Feature22.55±0.3532.95±0.5368.71±0.8170.86±0.7846.52±0.7034.07±0.5456.68±0.8139.62±0.6246.50
Style (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotImage24.92±0.3642.63±0.4784.18±0.5490.31±0.4766.37±0.6547.46±0.6775.94±0.6257.33±0.6561.14
Feature25.55±0.3743.71±0.5084.22±0.5591.71±0.4467.31±0.6750.26±0.6776.46±0.6557.39±0.6362.08
Style (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
", + "bbox": [ + 78, + 89, + 893, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/746e4dcbd8c969ffeaf283d8f714568384a3d4de4880ed9a7cf5b84ad1201e75.jpg", + "table_caption": [ + "Table 2. Comparison results (%) of attacking image, feature, and styles. Models build upon ResNet10 and GNN classifier." + ], + "table_footnote": [], + "table_body": "
Augment MethodChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotStyleGaus†22.37±0.3531.48±0.5265.71±0.8269.25±0.8046.32±0.7232.69±0.5455.48±0.7937.27±0.6145.07
MixStyle [66]22.43±0.3533.21±0.5367.35±0.8068.80±0.8247.08±0.7333.39±0.5856.12±0.7838.03±0.6245.80
AdvStyle [63]22.04±0.3630.83±0.5265.19±0.8264.96±0.8147.43±0.7231.90±0.5253.95±0.7935.81±0.5944.01
DSU [31]22.35±0.3631.43±0.5164.55±0.8364.73±0.8147.74±0.7231.61±0.5354.81±0.8137.19±0.6144.30
Style-FGSM (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotStyleGaus†24.97±0.3741.74±0.4881.88±0.6189.71±0.4965.98±0.6745.03±0.6472.66±0.6856.66±0.6559.83
MixStyle [66]25.04±0.3643.77±0.5382.67±0.5888.90±0.5265.73±0.6645.91±0.6375.90±0.6356.59±0.6260.56
AdvStyle [63]25.03±0.3543.15±0.5083.09±0.5788.44±0.5266.42±0.6744.85±0.6474.14±0.6554.89±0.6460.00
DSU [31]25.02±0.3645.19±0.5280.30±0.6386.30±0.5667.94±0.6645.65±0.6375.17±0.6454.31±0.6259.99
Style-FGSM (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
", + "bbox": [ + 78, + 222, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Different style augmentation methods are compared. \"StyleGaus†\" means adding random Gaussian noises to the styles, where † represents it is proposed by us. \"MixStyle [66]\", \"AdvStyle [63]\" and \"DSU [31]\" are adapted from other tasks, e.g., domain generation. Results (%) conducted under 5-way 1-shot/5-shot settings. Methods are built upon the ResNet10 and GNN.", + "bbox": [ + 75, + 378, + 893, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and features by modifying the attack targets of our method. The 5-way 1-shot/5-shot results are given in Table 2. We highlight several points. 1) We notice that attacking image, feature, and style all improve the base GNN model (given in Table 1) which shows that all of them boost the generalization ability of the model by adversarial attacks. Interestingly, the results of our \"Attack Image\"/\"Attack Feature\" even outperform the well-designed CD-FSL methods ATA [51] and AFA [20] (shown in Table 1); 2) Our method has clear advantages over attacking images and features. This again indicates the superiority of tackling visual styles for narrowing the domain gap issue for CD-FSL.", + "bbox": [ + 75, + 429, + 472, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is Style-FGSM Better than Other Style Augmentation Methods? To show the advantages of our Style-FGSM against other style augmentation methods, we introduce several competitors including \"StyleGaus\", MixStyle [66], AdvStyle [63], and DSU [31]. Typically, \"StyleGaus\" that adds random Gaussian noises into the styles is introduced as a simple but reasonable baseline. MixStyle [66], AdvStyle [63], and DSU [31] which are initially designed for other tasks, e.g., segmentation and domain generation are also adapted. The results are reported in Table 3. Comparing the results of StyleGuas with that reported in Table 1, we find that perturbing the styles on the feature level by simply adding random noises also improves the base GNN and even surpasses a few CD-FSL competitors on some target datasets. This phenomenon is consistent with the insight that augmenting the style distributions helps boost the CD-FSL methods. As for the comparison between our Style-FGSM and other advanced style augmentation competitors, we find that Style-FGSM performs better than all the MixStyle, AdvStyle, and", + "bbox": [ + 75, + 613, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DSU on both 1-shot and 5-shot settings. Typically, MixStyle and DSU both generate virtual styles, but their new styles are still relatively easy. This shows that our hard styles boost the model to a larger extent. AdvStyle generates both virtual and hard (adversarial) styles. However, it is still inferior to us. This indicates the advantages of our method that attacks in latent feature space and adopts two individual tasks for attacking and optimization.", + "bbox": [ + 496, + 429, + 890, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 569, + 617, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel model-agnostic StyleAdv for CD-FSL. Critically, to narrow the domain gap which is typically in the form of visual shifts, StyleAdv solves the minimax game of style adversarial learning: first adds perturbations to the source styles increasing the loss of the current model, then optimizes the model by forcing it to recognize both the clean and style perturbed data. Besides, a novel progressive style adversarial attack method termed style-FGSM is presented by us. Style-FGSM synthesizes diverse \"hard\" and \"virtual\" styles via adding the signed gradients to original clean styles. These generated styles support the max step of StyleAdv. Intuitively, by exposing the CD-FSL to adversarial styles which are more challenging than those limited real styles that exist in the source dataset, the generalization ability of the model is boosted. Our StyleAdv improves both CNN-based and ViT-based models. Extensive experiments indicate that our StyleAdv build new SOTAs.", + "bbox": [ + 496, + 595, + 893, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project was supported by National Key R&D Program of China (No. 2021ZD0112804) and NSFC under Grant No. 62076067.", + "bbox": [ + 498, + 854, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24582", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] John Cai, Bill Cai, and Shen Sheng Mei. Damsl: Domain agnostic meta score-based learning. In CVPR, 2021. 2", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 5", + "[3] Chen Chen, Zeju Li, Cheng Ouyang, Matt Sinclair, Wenjia Bai, and Daniel Rueckert. Maxstyle: Adversarial style composition for robust medical image segmentation. arXiv preprint, 2022. 3", + "[4] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint, 2019. 5", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint, 2020. 2, 5", + "[6] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A Kai Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2", + "[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. TIP, 2022. 2", + "[8] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In ACM Multimedia, 2021. 2, 6", + "[9] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Me-d2n: Multi-expert domain decompositional network for cross-domain few-shot learning. In ACM Multimedia, 2022. 2", + "[10] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Wave-san: Wavelet based style augmentation network for cross-domain few-shot learning. arXiv preprint, 2022. 1, 2, 3, 5, 6, 12", + "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint, 2021. 2", + "[12] Victor Garcia and Joan Bruna. Few-shot learning with graph neural networks. arXiv preprint, 2017. 2, 5, 6, 12, 13", + "[13] Micah Goldblum, Liam Fowl, and Tom Goldstein. Adversarially robust few-shot learning: A meta-learning approach. NeurIPS, 2020. 2, 7", + "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint, 2014. 2, 3, 5", + "[15] Jiechao Guan, Manli Zhang, and Zhiwu Lu. Large-scale cross-domain few-shot learning. In ACCV, 2020. 2", + "[16] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2, 5, 6", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 2019. 5", + "[19] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In CVPR, 2022. 2, 5, 6, 12, 13", + "[20] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, 2022. 2, 6, 7, 8", + "[21] Zhengdong Hu, Yifan Sun, and Yi Yang. Switch to generalize: Domain-switch learning for cross-domain few-shot classification. In ICLR, 2021. 2, 6", + "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 3", + "[23] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard Radke. Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. NeurIPS, 2021. 2", + "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, 2013. 5", + "[25] Manoj Kumar, Varun Kumar, Hadrien Glaude, Cyprien de Lichy, Aman Alok, and Rahul Gupta. Protoda: Efficient transfer learning for few-shot intent classification. In SLT Workshop, 2021. 2", + "[26] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. NeurIPS, 2019. 2", + "[27] Bo Li and Yevgeniy Vorobeychik. Feature cross-substitution in adversarial classification. NeurIPS, 2014. 2", + "[28] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, 2020. 2", + "[29] Pan Li, Shaogang Gong, Yanwei Fu, and Chengjie Wang. Ranking distance calibration for cross-domain few-shot learning. arXiv preprint, 2021. 2", + "[30] Wenbin Li, Lei Wang, Xingxing Zhang, Jing Huo, Yang Gao, and Jiebo Luo. Defensive few-shot adversarial learning. arXiv preprint, 2019. 2, 7", + "[31] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint, 2022. 3, 8", + "[32] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In ICCV, 2021. 2, 5, 6", + "[33] Bingyu Liu, Zhen Zhao, Zhenpeng Li, Jianan Jiang, Yuhong Guo, and Jieping Ye. Feature transformation ensemble model with batch spectral regularization for cross-domain few-shot classification. arXiv preprint, 2020. 2, 5, 6" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24583", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint, 2017. 2, 3, 5", + "[35] Sharada P Mohanty, David P Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in plant science, 2016. 5", + "[36] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016. 2", + "[37] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. arXiv preprint, 2020. 2, 6", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[39] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2, 5", + "[40] Wei Shen, Ziqiang Shi, and Jun Sun. Learning from adversarial features for few-shot classification. arXiv preprint, 2019. 2, 7", + "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2, 5", + "[42] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. arXiv preprint, 2020. 1, 2, 5, 6", + "[43] Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In CVPR, 2019. 2", + "[44] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, 2018. 12, 13", + "[45] Hao Tang, Zechao Li, Zhimao Peng, and Jinhui Tang. Blockmix: meta regularization and self-calibrated inference for metric-based meta-learning. In ACM Multimedia, 2020. 2", + "[46] Hao Tang, Chengcheng Yuan, Zechao Li, and Jinhui Tang. Learning attention-guided pyramidal features for few-shot fine-grained recognition. PR. 2", + "[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 2018. 5", + "[48] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1, 2, 5, 6, 12, 13", + "[49] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 5", + "[50] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. arXiv preprint, 2021. 1, 2, 5, 6, 7, 8", + "[52] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. arXiv preprint, 2021. 2", + "[53] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 5", + "[54] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In ICCV, 2021. 3", + "[55] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 2", + "[56] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In CVPR, 2021. 2", + "[57] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2", + "[58] Ji Zhang, Jingkuan Song, Lianli Gao, Ye Liu, and Heng Tao Shen. Progressive meta-learning with curriculum. TCVT, 2022. 2", + "[59] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint, 2021. 2", + "[60] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In ICLR. 2", + "[61] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, 2019. 3", + "[62] Zhedong Zheng and Yi Yang. Rectifying pseudo label learning via uncertainty estimation for domain adaptive semantic segmentation. IJCV, 2021. 2", + "[63] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint, 2022. 3, 8", + "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 2017. 5", + "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2", + "[66] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 3, 8, 12", + "[67] Linhai Zhuo, Yuqian Fu, Jingjing Chen, Yixin Cao, and YuGang Jiang. Tgdm: Target guided dynamic mixup for cross-domain few-shot learning. In ACM Multimedia, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 887 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24584", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_model.json b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f866ab4541dc5612b6078c0c11c3a0d4c9ca3f60 --- /dev/null +++ b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.892, + 0.153 + ], + "angle": 0, + "content": "StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.18, + 0.71, + 0.199 + ], + "angle": 0, + "content": "Yuqian Fu\\(^{1}\\), Yu Xie\\(^{2}\\), Yanwei Fu\\(^{3}\\), Yu-Gang Jiang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.199, + 0.917, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University \\(^{2}\\)Purple Mountain Laboratories, Nanjing, China. \\(^{3}\\)School of Data Science, Fudan University \\(\\{\\text { fuyq20, yxie18, yanweifu, ygj } \\} @\\)fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.312, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.474, + 0.759 + ], + "angle": 0, + "content": "Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN [10] empirically shows that spanning the style distribution of the source data helps alleviate this issue. However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles \"real\" and \"easy\", which still fall into the original set of the source styles. Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. Particularly, our style attack method synthesizes both \"virtual\" and \"hard\" adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. Code is available at https://github.com/lovelyqian/StyleAdv-CDFSL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.789, + 0.208, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.471, + 0.876 + ], + "angle": 0, + "content": "This paper studies the task of Cross-Domain Few-Shot Learning (CD-FSL) which addresses the Few-Shot Learning (FSL) problem across different domains. As a general recipe for FSL, episode-based meta-learning strategy" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.895, + 0.424 + ], + "angle": 0, + "content": "has also been adopted for training CD-FSL models, e.g., FWT [48], LRP [42], ATA [51], and wave-SAN [10]. Generally, to mimic the low-sample regime in testing stage, meta learning samples episodes for training the model. Each episode contains a small labeled support set and an unlabeled query set. Models learn meta knowledge by predicting the categories of images contained in the query set according to the support set. The learned meta knowledge generalizes the models to novel target classes directly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Empirically, we find that the changes of visual appearances between source and target data is one of the key causes that leads to the domain gap in CD-FSL. Interestingly, waveSAN [10], our former work, shows that the domain gap issue can be alleviated by augmenting the visual styles of source images. Particularly, wave-SAN proposes to augment the styles, in the form of Adaptive Instance Normalization (AdaIN) [22], by randomly sampling two source episodes and exchanging their styles. However, despite the efficacy of wave-SAN, such a naive style generation method suffers from two limitations: 1) The swap operation makes the styles always be limited in the \"real\" style set of the source dataset; 2) The limited real styles further lead to the generated styles too \"easy\" to learn. Therefore, a natural question is whether we can synthesize \"virtual\" and \"hard\" styles for learning a more robust CD-FSL model? Formally, we use \"real/virtual\" to indicate whether the styles are originally presented in the set of source styles, and define \"easy/hard\" as whether the new styles make meta tasks more difficult." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "To that end, we draw inspiration from the adversarial training, and propose a novel meta Style Adversarial training method (StyleAdv) for CD-FSL. StyleAdv plays the minimax game in two iterative optimization loops of metaintraining. Particularly, the inner loop generates adversarial styles from the original source styles by adding perturbations. The synthesized adversarial styles are supposed to be more challenging for the current model to recognize, thus, increasing the loss. Whilst the outer loop optimizes the whole network by minimizing the losses of recognizing the images with both original and adversarial styles. Our ultimate goal is to enable learning a model that is robust to various styles, be-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.27, + 0.9 + ], + "angle": 0, + "content": "* indicates corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "24575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "yond the relatively limited and simple styles from the source data. This can potentially improve the generalization ability on novel target domains with visual appearance shifts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.471, + 0.366 + ], + "angle": 0, + "content": "Formally, we introduce a novel style adversarial attack method to support the inner loop of StyleAdv. Inspired yet different from the previous attack methods [14,34], our style attack method perturbs and synthesizes the styles rather than image pixels or features. Technically, we first extract the style from the input feature map, and include the extracted style in the forward computation chain to obtain its gradient for each training step. After that, we synthesize the new style by adding a certain ratio of gradient to the original style. Styles synthesized by our style adversarial attack method have the good properties of \"hard\" and \"virtual\". Particularly, since we perturb styles in the opposite direction of the training gradients, our generation leads to the \"hard\" styles. Our attack method results in totally \"virtual\" styles that are quite different from the original source styles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.471, + 0.671 + ], + "angle": 0, + "content": "Critically, our style attack method makes progressive style synthesizing, with changing style perturbation ratios, which makes it significantly different from vanilla adversarial attacking methods. Specifically, we propose a novel progressive style synthesizing strategy. The naive solution of directly plugging-in perturbations is to attack each block of the feature embedding module individually, which however, may results in large deviations of features from the high-level block. Thus, our strategy is to make the synthesizing signal of the current block be accumulated by adversarial styles from previous blocks. On the other hand, rather than attacking the models by fixing the attacking ratio, we synthesize new styles by randomly sampling the perturbation ratio from a candidate pool. This facilitates the diversity of the synthesized adversarial styles. Experimental results have demonstrated the efficacy of our method: 1) our style adversarial attack method does synthesize more challenging styles, thus, pushing the limits of the source visual distribution; 2) our StyleAdv significantly improves the base model and outperforms all other CD-FSL competitors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.809 + ], + "angle": 0, + "content": "We highlight our StyleAdv is model-agnostic and complementary to other existing FSL or CD-FSL models, e.g., GNN [12] and FWT [48]. More importantly, to benefit from the large-scale pretrained models, e.g., DINO [2], we further explore adapting our StyleAdv to improve the Vision Transformer (ViT) [5] backbone in a non-parametric way. Experimentally, we show that StyleAdv not only improves CNN-based FSL/CD-FSL methods, but also improves the large-scale pretrained ViT model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Finally, we summarize our contributions. 1) A novel meta style adversarial training method, termed StyleAdv, is proposed for CD-FSL. By first perturbing the original styles and then forcing the model to learn from such adversarial styles, StyleAdv improves the robustness of CD-FSL models. 2) We present a novel style attack method with the novel" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "progressive synthesizing strategy in changing attacking ratios. Diverse \"virtual\" and \"hard\" styles thus are generated. 3) Our method is complementary to existing FSL and CD-FSL methods; and we validate our idea on both CNN-based and ViT-based backbones. 4) Extensive results on eight unseen target datasets indicate that our StyleAdv outperforms previous CD-FSL methods, building a new SOTA result." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.211, + 0.642, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Cross-Domain Few-Shot Learning. FSL which aims at freeing the model from reliance on massive labeled data has been studied for many years [12, 25, 39, 41, 43, 45, 46, 56, 58]. Particularly, some recent works, e.g., CLIP [38], CoOp [65], CLIP-Adapter [11], Tip-Adapter [59], and PMF [19] explore promoting the FSL with large-scale pretrained models. Particularly, PMF contributes a simple pipeline and builds a SOTA for FSL. As an extended task from FSL, CD-FSL [1, 7–10, 15, 16, 23, 29, 32, 37, 42, 48, 51, 60, 67] mainly solves the FSL across different domains. Typical meta-learning based CD-FSL methods include FWT [48], LRP [42], ATA [51], AFA [20], and wave-SAN [10]. Specifically, FWT and LRP tackle CD-FSL by refining batch normalization layers and using the explanation model to guide training. ATA, AFA, and wave-SAN propose to augment the image pixels, features, and visual styles, respectively. Several transfer-learning based CD-FSL methods, e.g., BSCD-FSL (also known as Fine-tune) [16], BSR [33], and NSAE [32] have also been explored. These methods reveal that finetuning helps improving the performances on target datasets. Other works that introduce extra data or require multiple domain datasets for training include STARTUP [37], Meta-FDMixup [8], Me-D2N [9], TGDM [67], TriAE [15], and DSL [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.764 + ], + "angle": 0, + "content": "Adversarial Attack. The adversarial attack aims at misleading models by adding some bespoke perturbations to input data. To generate the perturbations effectively, lots of adversarial attack methods have been proposed [6, 14, 26, 27, 34, 36, 55, 57]. Most of the works [6, 14, 34, 36] attack the image pixels. Specifically, FGSM [14] and PGD [34] are two most classical and famous attack algorithms. Several works [26, 27, 62] attack the feature space. Critically, few methods [57] attack styles. Different from these works that aim to mislead the models, we perturb the styles to tackle the visual shift issue for CD-FSL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Adversarial Few-Shot Learning. Several attempts [13, 28, 30, 40, 52] that explore adversarial learning for FSL have been made. Among them, MDAT [30], AQ [13], and MetaAdv [52] first attack the input image and then train the model using the attacked images to improve the defense ability against adversarial samples. Shen et al. [40] attacks the feature of the episode to improve the generalization capability of FSL models. Note that ATA [51] and AFA [20], two CD-FSL methods, also adopt the adversarial learning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "24576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "However, we are greatly different from them. ATA and AFA perturb image pixels or features, while we aim at bridging the visual gap by generating diverse hard styles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Style Augmentation for Domain Shift Problem. Augmenting the style distribution for narrowing the domain shift issue has been explored in domain generation [31, 54, 66], image segmentation [3, 63], person re-ID [61], and CD-FSL [10]. Concretely, MixStyle [66], AdvStyle [63], DSU [31], and wave-SAN [10] synthesize styles without extra parameters via mixing, attacking, sampling from a Gaussian distribution, and swapping. MaxStyle [3] and L2D [54] require additional network modules and complex auxiliary tasks to help generate the new styles. Typically, AdvStyle [63] is the most related work to us. Thus, we highlight the key differences: 1) AdvStyle attacks styles on the image, while we attack styles on multiple feature spaces with a progressive attacking method; 2) AdvStyle uses the same task loss (segmentation) for attacking and optimization; in contrast, we use the classical classification loss to attack the styles, while utilize the task loss (FSL) to optimize the whole network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.414, + 0.462, + 0.431 + ], + "angle": 0, + "content": "3. StyleAdv: Meta Style Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.607 + ], + "angle": 0, + "content": "Task Formulation. Episode \\( \\mathcal{T} = ((S, Q), Y) \\) is randomly sampled as the input of each meta-task, where \\( Y \\) represents the global class labels of the episode images with respect to \\( \\mathcal{C}^{tr} \\). Typically, each meta-task is formulated as an \\( N \\)-way \\( K \\)-shot problem. That is, for each episode \\( \\mathcal{T} \\), \\( N \\) classes with \\( K \\) labeled images are sampled as the support set \\( S \\), and the same \\( N \\) classes with another \\( M \\) images are used to constitute the query set \\( Q \\). The FSL or CD-FSL models predict the probability \\( P \\) that the images in \\( Q \\) belong to \\( N \\) categories according to \\( S \\). Formally, we have \\( |S| = NK \\), \\( |Q| = NM \\), \\( |P| = NM \\times N \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.685 + ], + "angle": 0, + "content": "FGSM and PGD Attackers. We briefly summarize the algorithms for FGSM [14] and PGD [34], two most famous attacking methods. Given image \\( x \\) with label \\( y \\), FGSM attacks the \\( x \\) by adding a ratio \\( \\epsilon \\) of signed gradients with respect to the \\( x \\) resulting in the adversarial image \\( x^{adv} \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.698, + 0.47, + 0.716 + ], + "angle": 0, + "content": "\\[\nx ^ {a d v} = x + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} J (\\theta , x, y)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.807 + ], + "angle": 0, + "content": "where \\( J(\\cdot) \\) and \\( \\theta \\) denote the object function and the learnable parameters of a classification model. PGD can be regarded as a variant of FGSM. Different from the FGSM that only attacks once, PGD attacks the image in an iterative way and sets a random start (abbreviated as RT) for \\( x \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.821, + 0.47, + 0.839 + ], + "angle": 0, + "content": "\\[\nx _ {0} ^ {a d v} = x + k _ {R T} \\cdot \\mathcal {N} (0, I), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.856, + 0.47, + 0.875 + ], + "angle": 0, + "content": "\\[\nx _ {t} ^ {a d v} = x _ {t - 1} ^ {a d v} + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} J (\\theta , x, y)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.462, + 0.901 + ], + "angle": 0, + "content": "where \\(k_{RT},\\epsilon\\) are hyper-parameters. \\(\\mathcal{N}\\) is Gaussian noises." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.887, + 0.108 + ], + "angle": 0, + "content": "3.1. Overview of Meta Style Adversarial Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.325 + ], + "angle": 0, + "content": "To alleviate the performance degradation caused by the changing visual appearance, we tackle CD-FSL by promoting the robustness of models on recognizing various styles. Thus, we expose our FSL model to some challenging virtual styles beyond the image styles that existed in the source dataset. To that end, we present the novel StyleAdv adversarial training method. Critically, rather than adding perturbations to image pixels, we particularly focus on adversarially perturbing the styles. The overall framework of our StyleAdv is illustrated in Figure 1. Our StyleAdv contains a CNN/ViT backbone \\( E \\), a global FC classifier \\( f_{cls} \\), and a FSL classifier \\( f_{fsl} \\) with learnable parameters \\( \\theta_E, \\theta_{cls}, \\theta_{fsl} \\), respectively. Besides, our core style attack method, a novel style extraction module, and the AdaIN are also included." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.895, + 0.372 + ], + "angle": 0, + "content": "Overall, we learn the StyleAdv by solving a minimax game. Specifically, the minimax game shall involve two iterative optimization loops in each meta-train step. Particularly," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.382, + 0.894, + 0.427 + ], + "angle": 0, + "content": "- Inner loop: synthesizing new adversarial styles by attacking the original source styles; the generated styles will increase the loss of the current network." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.439, + 0.894, + 0.483 + ], + "angle": 0, + "content": "- Outer loop: optimizing the whole network by classifying source images with both original and adversarial styles; this process will decrease the loss." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.382, + 0.894, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.494, + 0.828, + 0.51 + ], + "angle": 0, + "content": "3.2. Style Extraction from CNNs and ViTs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.894, + 0.64 + ], + "angle": 0, + "content": "Adaptive Instance Normalization (AdaIN). We recap the vanilla AdaIN [22] proposed for CNN in style transfer. Particularly, AdaIN reveals that the instance-level mean and standard deviation (abbreviated as mean and std) convey the style information of the input image. Denoting the mean and std as \\(\\mu\\) and \\(\\sigma\\), AdaIN (denoted as \\(\\mathcal{A}\\)) reveals that the style of \\(F\\) can be transferred to that of \\(F_{tgt}\\) by replacing the original style \\((\\mu, \\sigma)\\) with the target style \\((\\mu_{tgt}, \\sigma_{tgt})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.65, + 0.892, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (F, \\mu_ {t g t}, \\sigma_ {t g t}) = \\sigma_ {t g t} \\frac {F - \\mu (F)}{\\sigma (F)} + \\mu_ {t g t}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.894, + 0.764 + ], + "angle": 0, + "content": "Style Extraction for CNN Features. As shown in the upper part of Figure 1 (b), let \\( F \\in \\mathcal{R}^{B \\times C \\times H \\times W} \\) indicates the input feature batch, where \\( B, C, H, \\) and \\( W \\) denote the batch size, channel, height, and width of the feature \\( F \\), respectively. As in AdaIN, the mean \\( \\mu \\) and std \\( \\sigma \\) of \\( F \\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.775, + 0.892, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mu (\\mathrm {F}) _ {\\mathrm {b}, \\mathrm {c}} = \\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} F _ {b, c, h, w}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.83, + 0.892, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\sigma (\\mathrm {F}) _ {\\mathrm {b}, \\mathrm {c}} = \\sqrt {\\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left(F _ {b , c , h , w} - \\mu_ {\\mathrm {b , c}} (F)\\right) ^ {2} + \\epsilon}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.645, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mu ,\\sigma \\in \\mathcal{R}^{B\\times C}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "24577" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.579, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.376, + 0.482, + 0.391 + ], + "angle": 0, + "content": "(a) Overall Framework of Our StyleAdv Method" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.088, + 0.888, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.376, + 0.862, + 0.39 + ], + "angle": 0, + "content": "(b) Style Extraction For CNN/ViT Features" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.893, + 0.428 + ], + "angle": 0, + "content": "Figure 1. (a): Overview of StyleAdv method. The inner loop synthesizes adversarial styles, while the outer loop optimizes the whole network. (b): Style extraction for CNN-based and ViT-based features (illustration with \\( B = 1 \\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.473, + 0.715 + ], + "angle": 0, + "content": "Meta Information Extraction for ViT Features. We explore extracting the meta information of the ViT features as the manner of CNN. Intuitively, such meta information can be regarded as a unique \"style\" of ViTs. As shown in Figure 1 (b), we take an input batch data with image split into \\( P \\times P \\) patches as an example. The ViT encoder will encode the batch patches into a class (cls) token (\\( F_{cls} \\in \\mathcal{R}^{B \\times C} \\)) and a patch tokens (\\( F_0 \\in \\mathcal{R}^{B \\times P^2 \\times C} \\)). To be compatible with AdaIN, we reshape the \\( F_0 \\) as \\( F \\in \\mathcal{R}^{B \\times C \\times P \\times P} \\). At this point, we can calculate the meta information for patch tokens \\( F \\) as in Eq. 5 and Eq. 6. Essentially, note that the transformer integrates the positional embedding into the patch representation, the spatial relations thus could be considered still hold in the patch tokens. This supports us to reform the patch tokens \\( F_0 \\) as a spatial feature map \\( F \\). To some extent, this can also be achieved by applying the convolution on the input data via a kernel of size \\( P \\times P \\) (as indicated by dashed arrows in Figure 1 (b))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.468, + 0.741 + ], + "angle": 0, + "content": "3.3. Inner Loop: Style Adversarial Attack Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We propose a novel style adversarial attack method - Fast Style Gradient Sign Method (Style-FGSM) to accomplish the inner loop. As shown in Figure 1, given an input source episode \\((\\mathcal{T},Y)\\), we first forward it into the backbone \\(E\\) and the FC classifier \\(f_{cls}\\) producing the global classification loss \\(\\mathcal{L}_{cls}\\) (as illustrated in the ① paths). During this process, a key step is to make the gradient of the style available. To achieve that, let \\(F_{\\mathcal{T}}\\) denotes the features of \\(\\mathcal{T}\\), we obtain the style \\((\\mu ,\\sigma)\\) of \\(F_{\\mathcal{T}}\\) as in Sec. 3.2. After that, we reform the original episode feature as \\(\\mathcal{A}(F_{\\mathcal{T}},\\mu ,\\sigma)\\). And the reformed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.893, + 0.486 + ], + "angle": 0, + "content": "feature is actually used for the forward propagation. In this way, we include \\(\\mu\\) and \\(\\sigma\\) in our forward computation chain; and thus, we could access the gradients of them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.533 + ], + "angle": 0, + "content": "With the gradients in ② paths, we then attack \\(\\mu\\) and \\(\\sigma\\) as FGSM does - adding a small ratio \\(\\epsilon\\) of the signed gradients with respect to \\(\\mu\\) and \\(\\sigma\\), respectively." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.544, + 0.893, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mu^ {a d v} = \\mu + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\mu} J \\left(\\theta_ {E}, \\theta_ {f _ {c l s}}, \\mathcal {A} \\left(F _ {\\tau}, \\mu , \\sigma\\right), Y\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.573, + 0.892, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {a d v} = \\sigma + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\sigma} J \\left(\\theta_ {E}, \\theta_ {f _ {c l s}}, \\mathcal {A} \\left(F _ {\\mathcal {T}}, \\mu , \\sigma\\right), Y\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.689 + ], + "angle": 0, + "content": "where the \\(J(\\cdot)\\) is the cross-entropy loss between classification predictions and ground truth \\(Y\\), i.e., \\(\\mathcal{L}_{cls}\\). Inspired by the random start of PGD, we also add random noises \\(k_{RT} \\cdot \\mathcal{N}(0, I)\\) to \\(\\mu\\) and \\(\\sigma\\) before attacking. \\(\\mathcal{N}(0, I)\\) refers to Gaussian noises and \\(k_{RT}\\) is a hyper-parameter. Our Style-FGSM enables us to generate both \"virtual\" and \"hard\" styles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Progressive Style Synthesizing Strategy: To prevent the high-level adversarial feature from deviating, we propose to apply our style-FGSM in a progressive strategy. Concretely, the embedding module \\(E\\) has three blocks \\(E_{1}\\), \\(E_{2}\\), and \\(E_{3}\\), with the corresponding features \\(F_{1}\\), \\(F_{2}\\), and \\(F_{3}\\). For the first block, we use \\((\\mu_{1}, \\sigma_{1})\\) to denote the original styles of \\(F_{1}\\). The adversarial styles \\((\\mu_{1}^{adv}, \\sigma_{1}^{adv})\\) are obtained directly as in Eq. 7 and Eq. 8. For subsequent blocks, the attack signals on the current block \\(i\\) are those accumulated from the block 1 to block \\(i - 1\\). Take the second block as an example, the block feature \\(F_{2}\\) is not simply extracted by \\(E_{2}(F_{1})\\). Instead, we have \\(F_{2}' = E_{2}(F_{1}^{adv})\\), where \\(F_{1}^{adv} = \\mathcal{A}(F_{1}, \\mu_{1}^{adv}, \\sigma_{1}^{adv})\\). Attacking on \\(F_{2}'\\) results in the adversarial styles \\((\\mu_{2}^{adv}, \\sigma_{2}^{adv})\\). Accordingly, we gen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "24578" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.227 + ], + "angle": 0, + "content": "erate \\((\\mu_3^{adv},\\sigma_3^{adv})\\) for the last block. The illustration of the progressive attacking strategy is attached in the Appendix. Changing Style Perturbation Ratios: Different from the vanilla FGSM [14] or PGD [34], our style attacking algorithm is expected to synthesize new styles with diversity. Thus, instead of using a fixed attacking ratio \\(\\epsilon\\), we randomly sample \\(\\epsilon\\) from a candidate list \\(\\epsilon_{list}\\) as the current attacking ratio. Despite the randomness of \\(\\epsilon\\), we still synthesize styles in a more challenging direction, \\(\\epsilon\\) only affects the extent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.237, + 0.462, + 0.254 + ], + "angle": 0, + "content": "3.4. Outer Loop: Optimize the StyleAdv Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.47, + 0.365 + ], + "angle": 0, + "content": "For each meta-train iteration with clean episode \\(\\mathcal{T}\\) as input, our inner loop produces adversarial styles \\((\\mu_1^{adv},\\sigma_1^{adv})\\) \\((\\mu_{2}^{adv},\\sigma_{2}^{adv})\\) , and \\((\\mu_3^{adv},\\sigma_3^{adv})\\) . As in Figure 1, the goal of the outer loop is to optimize the whole StyleAdv with both the clean feature \\(F\\) and the style attacked feature \\(F^{adv}\\) utilized as the training data. Typically, the clean episode feature \\(F\\) can be obtained directly as \\(E(T)\\) as in 3 paths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.472, + 0.534 + ], + "angle": 0, + "content": "In 4 paths, we obtain the \\(F^{adv}\\) by transferring the original style of \\(F\\) to the corresponding adversarial attacked styles. Similar with the progressive style-FGSM, we have \\(F_{1}^{adv} = \\mathcal{A}(E_{1}(\\mathcal{T}),\\mu_{1}^{adv},\\sigma_{1}^{adv})\\), \\(F_{2}^{adv} = \\mathcal{A}(E_{2}(F_{1}^{adv}),\\mu_{2}^{adv},\\sigma_{2}^{adv})\\), and \\(F_{3}^{adv} = \\mathcal{A}(E_{3}(F_{2}^{adv}),\\mu_{3}^{adv},\\sigma_{3}^{adv})\\). Finally, \\(F^{adv}\\) is obtained by applying an average pooling layer to \\(F_{3}^{adv}\\). A skip probability \\(p_{skip}\\) is set to decide whether to skip the current attacking. Conducting FSL tasks for both the clean feature \\(F\\) and style attacked feature \\(F^{adv}\\) results in two FSL predictions \\(P_{fsl}\\), \\(P_{fsl}^{adv}\\), and two FSL classification losses \\(\\mathcal{L}_{fsl}\\), \\(\\mathcal{L}_{fsl}^{adv}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Further, despite the styles of \\( F^{adv} \\) shifts from that of \\( F \\), we encourage that the semantic content should be still consistent as in wave-SAN [10]. Thus we add a consistent constraint to the predictions of \\( P_{fsl} \\) and \\( P_{fsl}^{adv} \\) resulting in the consistent loss \\( \\mathcal{L}_{cons} \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.619, + 0.47, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s}} = \\mathrm {K L} \\left(P _ {f s l}, P _ {f s l} ^ {\\text {a d v}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.47, + 0.709 + ], + "angle": 0, + "content": "where \\(\\mathrm{KL}()\\) is Kullback-Leibler divergence loss. In addition, we have the global classification loss \\(\\mathcal{L}_{cls}\\). This ensures that \\(\\theta_{cls}\\) is optimized to provide correct gradients for style-FGSM. The final meta-objective of StyleAdv is as," + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.719, + 0.47, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {f s l} + \\mathcal {L} _ {f s l} ^ {a d v} + \\mathcal {L} _ {c o n s} + \\mathcal {L} _ {c l s}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Note that our StyleAdv is model-agnostic and orthogonal to existing FSL and CD-FSL methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.258, + 0.803 + ], + "angle": 0, + "content": "3.5. Network Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Applying StyleAdv Directly for Inference. Our StyleAdv facilitates making CD-FSL model more robust to style shifts. Once the model is meta-trained, we can employ it for inference directly by feeding the testing episode into the \\( E \\) and the \\( f_{cls} \\). The class with the highest probability will be taken as the predicted result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "Finetuning StyleAdv Using Target Examples. As indicated in previous works [16, 32, 33, 51], finetuning CD-FSL models on target examples helps improve the model performance. Thus, to further promote the performance of StyleAdv, we also equip it with the fintuning strategy forming an upgraded version (\"StyleAdv-FT\"). Specifically, as in ATA-FT [51], for each novel testing episode, we augment the novel support set to form pseudo episodes as training data for tuning the meta-trained model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.249, + 0.634, + 0.266 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Datasets. We take two CD-FSL benchmarks proposed in BSCD-FSL [16] and FWT [48]. Both of them take mini-Imagenet [39] as the source dataset. Two disjoint sets split from mini-Imagenet form \\(\\mathcal{D}^{tr}\\) and \\(\\mathcal{D}^{eval}\\). Totally eight datasets including ChestX [53], ISIC [4, 47], EuroSAT [18], CropDisease [35], CUB [50], Cars [24], Places [64], and Plantae [49] are taken as novel target datasets. The former four datasets included in BSCD-FSL's benchmark cover medical images varying from X-ray to dermoscopic skin lesions, and natural images from satellite pictures to plant disease photos. While the latter four datasets that focus on more fine-grained concepts such as birds and cars are contained in FWT. These eight target datasets serve as testing set \\(\\mathcal{D}^{te}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Network Modules. For typical CNN based network, following previous CD-FSL methods [10, 42, 48, 51], ResNet-10 [17] is selected as the embedding module while GNN [12] is selected as the FSL classifier; For the emerging ViT based network, following PMF [19], we use the ViT-small [5] and the ProtoNet [41] as the embedding module and the FSL classifier, respectively. Note that, the ViT-small is pretrained on ImageNet1K by DINO [2] as in PMF. The \\( f_{cls} \\) is built by a fully connected layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. The 5-way 1-shot and 5-way 5-shot settings are conducted. Taking ResNet10 as backbone, we meta train the network for 200 epochs, each epoch contains 120 meta tasks. Adam with a learning rate of 0.001 is utilized as the optimizer. Taking ViT-small as backbone, the meta train stage takes 20 epoch, each epoch contains 2000 meta tasks. The SGD with a initial learning rate of 5e-5 and 0.001 are used for optimize the \\( E(\\cdot) \\) and the \\( f_{cls} \\), respectively. The \\( \\epsilon_{list} \\), \\( k_{RT} \\) of Style-FGSM attacker are set as [0.8, 0.08, 0.008], \\( \\frac{16}{255} \\). The probability \\( p_{skip} \\) of random skipping the attacking is chosen from {0.2, 0.4}. We evaluate our network with 1000 randomly sampled episodes and report average accuracy (\\%) with a 95% confidence interval. Both the results of our \"StyleAdv\" and \"StyleAdv-FT\" are reported. The details of the finetuning are attached in Appendix. ResNet-10 based models are trained and tested on a single GeForce GTX 1080, while ViT-small based models require a single NVIDIA GeForce RTX 3090." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "24579" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.446 + ], + "angle": 0, + "content": "
1-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--22.00±0.4632.02±0.6663.69±1.0364.48±1.0845.69±0.6831.79±0.5153.10±0.8035.60±0.5643.55
FWT [48]RN10--22.04±0.4431.58±0.6762.36±1.0566.36±1.0447.47±0.7531.61±0.5355.77±0.7935.95±0.5844.14
LRP [42]RN10--22.11±0.2030.94±0.3054.99±0.5059.23±0.5048.29±0.5132.78±0.3954.83±0.5637.49±0.4342.58
ATA [51]RN10--22.10±0.2033.21±0.4061.35±0.5067.47±0.5045.00±0.5033.61±0.4053.57±0.5034.42±0.4043.84
AFA [20]RN10--22.92±0.2033.21±0.3063.12±0.5067.61±0.5046.86±0.5034.25±0.4054.04±0.6036.76±0.4044.85
wave-SAN [10]RN10--22.93±0.4933.35±0.7169.64±1.0970.80±1.0650.25±0.7433.55±0.6157.75±0.8240.71±0.6647.37
StyleAdv (ours)RN10--22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
ATA-FT [51]RN10Y-22.15±0.2034.94±0.4068.62±0.5075.41±0.5046.23±0.5037.15±0.4054.18±0.5037.38±0.4047.01
StyleAdv-FT (ours)RN10Y-22.64±0.3535.76±0.5272.92±0.7580.69±0.2848.49±0.7235.09±0.5558.58±0.8341.13±0.6749.41
PMF* [19]ViT-smallYDINO/IN1K21.73±0.3030.36±0.3670.74±0.6380.79±0.6278.13±0.6637.24±0.5771.11±0.7153.60±0.6655.46
StyleAdv (ours)ViT-small-DINO/IN1K22.92±0.3233.05±0.4472.15±0.6581.22±0.6184.01±0.5840.48±0.5772.64±0.6755.52±0.6657.75
StyleAdv-FT (ours)ViT-smallYDINO/IN1K22.92±0.3233.99±0.4674.93±0.5884.11±0.5784.01±0.5840.48±0.5772.64±0.6755.52±0.6658.57
5-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--25.27±0.4643.94±0.6783.64±0.7787.96±0.6762.25±0.6544.28±0.6370.84±0.6552.53±0.5958.84
FWT [48]RN10--25.18±0.4543.17±0.7083.01±0.7987.11±0.6766.98±0.6844.90±0.6473.94±0.6753.85±0.6259.77
LRP [42]RN10--24.53±0.3044.14±0.4077.14±0.4086.15±0.4064.44±0.4846.20±0.4674.45±0.4754.46±0.4658.94
ATA [51]RN10--24.32±0.4044.91±0.4083.75±0.4090.59±0.3066.22±0.5049.14±0.4075.48±0.4052.69±0.4060.89
AFA [20]RN10--25.02±0.2046.01±0.4085.58±0.4088.06±0.3068.25±0.5049.28±0.5076.21±0.5054.26±0.4061.58
wave-SAN [10]RN10--25.63±0.4944.93±0.6785.22±0.7189.70±0.6470.31±0.6746.11±0.6676.88±0.6357.72±0.6462.06
StyleAdv (ours)RN10--26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
Fine-tune [16]RN10Y-25.97±0.4148.11±0.6479.08±0.6189.25±0.5164.14±0.7752.08±0.7470.06±0.7459.27±0.7061.00
ATA-FT [51]RN10Y-25.08±0.2049.79±0.4089.64±0.3095.44±0.2069.83±0.5054.28±0.5076.64±0.4058.08±0.4064.85
NSAE [32]RN10Y-27.10±0.4454.05±0.6383.96±0.5793.14±0.4768.51±0.7654.91±0.7471.02±0.7259.55±0.7464.03
BSR [33]RN10Y-26.84±0.4454.42±0.6680.89±0.6192.17±0.4569.38±0.7657.49±0.7271.09±0.6861.07±0.7664.17
StyleAdv-FT (ours)RN10Y-26.24±0.3553.05±0.5491.64±0.4396.51±0.2870.90±0.6356.44±0.6879.35±0.6164.10±0.6467.28
PMF [19]ViT-smallYDINO/IN1K27.2750.1285.9892.96-----
StyleAdv (ours)ViT-small-DINO/IN1K26.97±0.3347.73±0.4488.57±0.3494.85±0.3195.82±0.2761.73±0.6288.33±0.4075.55±0.5472.44
StyleAdv-FT (ours)ViT-smallYDINO/IN1K26.97±0.3351.23±0.5190.12±0.3395.99±0.2795.82±0.2766.02±0.6488.33±0.4078.01±0.5474.06
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.448, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Table 1. Results of 5-way 1-shot/5-shot tasks. \"FT\" means whether the finetuning stage is employed. \"LargeP\" represents if large pretrained models are used for model initialization. \"RN10\" is short for \"ResNet-10\". * denotes results are reported by us. Results perform best are bolded. Whether based on ResNet-10 or ViT-small, our method outperforms other competitors significantly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.336, + 0.515 + ], + "angle": 0, + "content": "4.1. Comparison with the SOTAs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.473, + 0.825 + ], + "angle": 0, + "content": "We compare our StyleAdv/StyleAdv-FT against several most representative and competitive CD-FSL methods. Concretly, with the ResNet-10 (abbreviated as RN10) as backbone, totally nine methods including GNN [12], FWT [48], LRP [42], ATA [51], AFA [20], wave-SAN [10], Finetune [16], NSAE [32], and BSR [33] are introduced as our competitors. Among them, the former six competitors are meta-learning based method that used for inference directly, thus we compare our \"StyleAdv\" against them for a fair comparison. Typically, the GNN [12] works as a base model. The Fine-tune [16], NSAE [32], BSR [33], and ATA-FT [51] (formed by finetuning ATA) all require finetuning model during inference, thus our \"StyleAdv-FT\" is used. With the ViT as backbone, the most recent and competitive PMF (SOTA method for FSL) is compared. For fair comparisons, we follow the same pipeline proposed in PMF [19]. Note that we promote CD-FSL models with only one single source domain. Those methods that use extra training datasets, e.g., STARTUP [37], meta-FDMixup [8], and DSL [21] are not considered. The comparison results are given in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "For all results, our method outperforms all the listed CD-FSL competitors significantly and builds a new state of the art. Our StyleAdv-FT (ViT-small) on average achieves \\(58.57\\%\\) and \\(74.06\\%\\) on 5-way 1-shot and 5-shot, respectively. Our StyleAdv (RN10) and StyleAdv-FT (RN10) also" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.897, + 0.893 + ], + "angle": 0, + "content": "beats all the meta-learning based or transfer-learning (finetuning) based methods. Besides of the state-of-the-art accuracy, we also have other worth-mentioning observations. 1) We show that our StyleAdv method is a general solution for both CNN-based models and ViT-based models. Typically, based on ResNet10, our StyleAdv and StyleAdv-FT improve the base GNN by up to \\(4.93\\%\\) and \\(8.44\\%\\) on 5-shot setting. Based on ViT-small, at most cases, our StyleAdv-FT outperforms the PMF by a clear margin. More results of building StyleAdv upon other FSL or CD-FSL methods can be found in the Appendix. 2) Comparing FWT, LRP, ATA, AFA, waveSAN, and our StyleAdv, we find that StyleAdv performs best, followed by wave-SAN, then comes the AFA, ATA, FWT, and LRP. This phenomenon indicates that tackling CD-FSL by solving the visual shift problem is indeed more effective than other perspectives, e.g., adversarial training by perturbing the image features (AFA) or image pixels (ATA), transforms the normalization layers in FWT, and explanation guided training in LRP. 3) For the comparison between StyleAdv and wave-SAN that both tackles the visual styles, we notice that StyleAdv outperforms the wave-SAN in most cases. This demonstrates that the styles generated by our StyleAdv are more conducive to learning robust CD-FSL models than the style augmentation method proposed in wave-SAN. This justifies our idea of synthesizing more challenging (\"hard and virtual\") styles." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "24580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.317 + ], + "angle": 0, + "content": "4) Overall, the large-scale pretrained model promotes the CD-FSL obviously. Take 1-shot as an example, StyleAdv-FT (ViT-small) boosts the StyleAdv-FT (RN10) by \\(9.16\\%\\) on average. However, we show that the performance improvement varies greatly on different target domains. Generally, for target datasets with relative small domain gap, e.g., CUB and Plantae, models benefit a lot; otherwise, the improvement is limited. 5) We also find that under the cross-domain scenarios, finetuning model on target domain, e.g., NSAE, BSR do show an advantage over purely meta-learning based methods, e.g., FWT, LRP, and wave-SAN. However, to finetune model using extremely few examples, e.g., 5-way 1-shot is much harder than on relatively larger shots. This may explain why those finetune-based methods do not conduct experiments on 1-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.56 + ], + "angle": 0, + "content": "Effectiveness of Style-FGSM Attacker. To show the advantages of our progressive style synthesizing strategy and attacking with changing perturbation ratios, we compare our Style-FGSM against several variants and report the results in Figure 2. Specifically, for Figure 2 (a), we compare our style-FGSM against the variant that attacks the blocks individually. Results show that attacking in a progressive way exceeds the naive individual strategy in most cases. For Figure 2 (b), to demonstrate how the performance will be affected by fixed attacking ratios, we also conduct experiments with different \\(\\epsilon_{list}\\). Since we set the \\(\\epsilon_{list}\\) as [0.8, 0.08, 0.008], three different choices including [0.8], [0.08], and [0.008] are selected. From the results, we first notice that the best result can be reached by a single fixed ratio. However, sampling the attacking ratio from a pool of candidates achieves the best result in most cases." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.577, + 0.437, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.863, + 0.472, + 0.904 + ], + "angle": 0, + "content": "Figure 2. Effectiveness of the progressive style synthesizing strategy and the changing style perturbation ratios. The 5-way 1-shot results are reported. Models are built on ResNet10 and GNN." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.849, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.266, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Visualization of wave-SAN and StyleAdv. (a): synthesized images; (b): meta-training losses; (c): T-SNE results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.649, + 0.314 + ], + "angle": 0, + "content": "4.2. More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.896, + 0.809 + ], + "angle": 0, + "content": "Visualization of Hard Style Generation. To help understand the \"hard\" style generation of our method intuitively, as in Figure 3, we make several visualizations comparing StyleAdv against the wave-SAN. 1) As in Figure 3 (a), we show the stylized images generated by wave-SAN and our StyleAdv. The visualization is achieved by applying the style augmentation methods to input images. Specifically, for wave-SAN, the style is swapped with another randomly sampled source image; for StyleAdv, the results of attacking style with \\(\\epsilon = 0.08\\) are given. We observe that wave-SAN tends to exchange the global visual appearance, e.g., the color of the input image randomly. By contrast, StyleAdv prefers to disturb the important regions that are key to recognizing the image category. For example, the fur of the cat and the key parts (face and feet) of the dogs. These observations intuitively support our claim that our StyleAdv synthesize more harder styles than wave-SAN. 2) To quantitatively evaluate whether our StyleAdv introduces more challenging styles into the training stage, as in Figure 3 (b), we visualize the meta-training loss. Results reveal that the perturbed losses of wave-SAN oscillate around the original loss, while StyleAdv increases the original loss obviously. These phenomenons further validate that we perturb data towards a more difficult direction thus pushing the limits of style generation to a large extent. 3) To further show the advantages of StyleAdv over wave-SAN, as shown in Figure 3 (c), we visualize the high-level features extracted by the meta-trained wave-SAN and StyleAdv. Five classes (denoted by different colors) of mini-Imagenet are selected. T-SNE is used for reducing the feature dimensions. Results demonstrate that StyleAdv enlarges the inter-class distances making classes more distinguishable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Why Attack Styles Instead of Images or Features? A natural question may be why we choose to attack styles instead of other targets, e.g., the input image as in AQ [13], MDAT [30], and ATA [51] or the features as in Shen et al. [40] and AFA [20]? To answer this question, we compare our StyleAdv which attacks styles against attacking images" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.518, + 0.938 + ], + "angle": 0, + "content": "24581" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.894, + 0.203 + ], + "angle": 0, + "content": "
Attack TargetChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotImage22.71±0.3533.00±0.5367.00±0.8272.65±0.7548.15±0.7234.40±0.6057.89±0.8339.85±0.6446.96
Feature22.55±0.3532.95±0.5368.71±0.8170.86±0.7846.52±0.7034.07±0.5456.68±0.8139.62±0.6246.50
Style (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotImage24.92±0.3642.63±0.4784.18±0.5490.31±0.4766.37±0.6547.46±0.6775.94±0.6257.33±0.6561.14
Feature25.55±0.3743.71±0.5084.22±0.5591.71±0.4467.31±0.6750.26±0.6776.46±0.6557.39±0.6362.08
Style (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.204, + 0.844, + 0.217 + ], + "angle": 0, + "content": "Table 2. Comparison results (%) of attacking image, feature, and styles. Models build upon ResNet10 and GNN classifier." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.223, + 0.892, + 0.379 + ], + "angle": 0, + "content": "
Augment MethodChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotStyleGaus†22.37±0.3531.48±0.5265.71±0.8269.25±0.8046.32±0.7232.69±0.5455.48±0.7937.27±0.6145.07
MixStyle [66]22.43±0.3533.21±0.5367.35±0.8068.80±0.8247.08±0.7333.39±0.5856.12±0.7838.03±0.6245.80
AdvStyle [63]22.04±0.3630.83±0.5265.19±0.8264.96±0.8147.43±0.7231.90±0.5253.95±0.7935.81±0.5944.01
DSU [31]22.35±0.3631.43±0.5164.55±0.8364.73±0.8147.74±0.7231.61±0.5354.81±0.8137.19±0.6144.30
Style-FGSM (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotStyleGaus†24.97±0.3741.74±0.4881.88±0.6189.71±0.4965.98±0.6745.03±0.6472.66±0.6856.66±0.6559.83
MixStyle [66]25.04±0.3643.77±0.5382.67±0.5888.90±0.5265.73±0.6645.91±0.6375.90±0.6356.59±0.6260.56
AdvStyle [63]25.03±0.3543.15±0.5083.09±0.5788.44±0.5266.42±0.6744.85±0.6474.14±0.6554.89±0.6460.00
DSU [31]25.02±0.3645.19±0.5280.30±0.6386.30±0.5667.94±0.6645.65±0.6375.17±0.6454.31±0.6259.99
Style-FGSM (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.38, + 0.894, + 0.422 + ], + "angle": 0, + "content": "Table 3. Different style augmentation methods are compared. \"StyleGaus†\" means adding random Gaussian noises to the styles, where † represents it is proposed by us. \"MixStyle [66]\", \"AdvStyle [63]\" and \"DSU [31]\" are adapted from other tasks, e.g., domain generation. Results (%) conducted under 5-way 1-shot/5-shot settings. Methods are built upon the ResNet10 and GNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.473, + 0.61 + ], + "angle": 0, + "content": "and features by modifying the attack targets of our method. The 5-way 1-shot/5-shot results are given in Table 2. We highlight several points. 1) We notice that attacking image, feature, and style all improve the base GNN model (given in Table 1) which shows that all of them boost the generalization ability of the model by adversarial attacks. Interestingly, the results of our \"Attack Image\"/\"Attack Feature\" even outperform the well-designed CD-FSL methods ATA [51] and AFA [20] (shown in Table 1); 2) Our method has clear advantages over attacking images and features. This again indicates the superiority of tackling visual styles for narrowing the domain gap issue for CD-FSL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Is Style-FGSM Better than Other Style Augmentation Methods? To show the advantages of our Style-FGSM against other style augmentation methods, we introduce several competitors including \"StyleGaus\", MixStyle [66], AdvStyle [63], and DSU [31]. Typically, \"StyleGaus\" that adds random Gaussian noises into the styles is introduced as a simple but reasonable baseline. MixStyle [66], AdvStyle [63], and DSU [31] which are initially designed for other tasks, e.g., segmentation and domain generation are also adapted. The results are reported in Table 3. Comparing the results of StyleGuas with that reported in Table 1, we find that perturbing the styles on the feature level by simply adding random noises also improves the base GNN and even surpasses a few CD-FSL competitors on some target datasets. This phenomenon is consistent with the insight that augmenting the style distributions helps boost the CD-FSL methods. As for the comparison between our Style-FGSM and other advanced style augmentation competitors, we find that Style-FGSM performs better than all the MixStyle, AdvStyle, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.551 + ], + "angle": 0, + "content": "DSU on both 1-shot and 5-shot settings. Typically, MixStyle and DSU both generate virtual styles, but their new styles are still relatively easy. This shows that our hard styles boost the model to a larger extent. AdvStyle generates both virtual and hard (adversarial) styles. However, it is still inferior to us. This indicates the advantages of our method that attacks in latent feature space and adopts two individual tasks for attacking and optimization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.57, + 0.618, + 0.585 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.853 + ], + "angle": 0, + "content": "This paper presents a novel model-agnostic StyleAdv for CD-FSL. Critically, to narrow the domain gap which is typically in the form of visual shifts, StyleAdv solves the minimax game of style adversarial learning: first adds perturbations to the source styles increasing the loss of the current model, then optimizes the model by forcing it to recognize both the clean and style perturbed data. Besides, a novel progressive style adversarial attack method termed style-FGSM is presented by us. Style-FGSM synthesizes diverse \"hard\" and \"virtual\" styles via adding the signed gradients to original clean styles. These generated styles support the max step of StyleAdv. Intuitively, by exposing the CD-FSL to adversarial styles which are more challenging than those limited real styles that exist in the source dataset, the generalization ability of the model is boosted. Our StyleAdv improves both CNN-based and ViT-based models. Extensive experiments indicate that our StyleAdv build new SOTAs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This project was supported by National Key R&D Program of China (No. 2021ZD0112804) and NSFC under Grant No. 62076067." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "24582" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] John Cai, Bill Cai, and Shen Sheng Mei. Damsl: Domain agnostic meta score-based learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.472, + 0.258 + ], + "angle": 0, + "content": "[3] Chen Chen, Zeju Li, Cheng Ouyang, Matt Sinclair, Wenjia Bai, and Daniel Rueckert. Maxstyle: Adversarial style composition for robust medical image segmentation. arXiv preprint, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[4] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.472, + 0.425 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.428, + 0.472, + 0.469 + ], + "angle": 0, + "content": "[6] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A Kai Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.471, + 0.472, + 0.513 + ], + "angle": 0, + "content": "[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. TIP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.472, + 0.555 + ], + "angle": 0, + "content": "[8] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In ACM Multimedia, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.472, + 0.612 + ], + "angle": 0, + "content": "[9] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Me-d2n: Multi-expert domain decompositional network for cross-domain few-shot learning. In ACM Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.472, + 0.669 + ], + "angle": 0, + "content": "[10] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Wave-san: Wavelet based style augmentation network for cross-domain few-shot learning. arXiv preprint, 2022. 1, 2, 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.472, + 0.727 + ], + "angle": 0, + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.472, + 0.756 + ], + "angle": 0, + "content": "[12] Victor Garcia and Joan Bruna. Few-shot learning with graph neural networks. arXiv preprint, 2017. 2, 5, 6, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[13] Micah Goldblum, Liam Fowl, and Tom Goldstein. Adversarially robust few-shot learning: A meta-learning approach. NeurIPS, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.472, + 0.842 + ], + "angle": 0, + "content": "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint, 2014. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.844, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[15] Jiechao Guan, Manli Zhang, and Zhiwu Lu. Large-scale cross-domain few-shot learning. In ACCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[16] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[19] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In CVPR, 2022. 2, 5, 6, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[20] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[21] Zhengdong Hu, Yifan Sun, and Yi Yang. Switch to generalize: Domain-switch learning for cross-domain few-shot classification. In ICLR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.46 + ], + "angle": 0, + "content": "[23] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard Radke. Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.894, + 0.503 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[25] Manoj Kumar, Varun Kumar, Hadrien Glaude, Cyprien de Lichy, Aman Alok, and Rahul Gupta. Protoda: Efficient transfer learning for few-shot intent classification. In SLT Workshop, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.894, + 0.588 + ], + "angle": 0, + "content": "[26] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[27] Bo Li and Yevgeniy Vorobeychik. Feature cross-substitution in adversarial classification. NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.894, + 0.659 + ], + "angle": 0, + "content": "[28] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[29] Pan Li, Shaogang Gong, Yanwei Fu, and Chengjie Wang. Ranking distance calibration for cross-domain few-shot learning. arXiv preprint, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.894, + 0.746 + ], + "angle": 0, + "content": "[30] Wenbin Li, Lei Wang, Xingxing Zhang, Jing Huo, Yang Gao, and Jiebo Luo. Defensive few-shot adversarial learning. arXiv preprint, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[31] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[32] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In ICCV, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[33] Bingyu Liu, Zhen Zhao, Zhenpeng Li, Jianan Jiang, Yuhong Guo, and Jieping Ye. Feature transformation ensemble model with batch spectral regularization for cross-domain few-shot classification. arXiv preprint, 2020. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "24583" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[35] Sharada P Mohanty, David P Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in plant science, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.472, + 0.235 + ], + "angle": 0, + "content": "[36] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.471, + 0.279 + ], + "angle": 0, + "content": "[37] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. arXiv preprint, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.472, + 0.349 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.352, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[39] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[40] Wei Shen, Ziqiang Shi, and Jun Sun. Learning from adversarial features for few-shot classification. arXiv preprint, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.425, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.511 + ], + "angle": 0, + "content": "[42] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. arXiv preprint, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[43] Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.472, + 0.61 + ], + "angle": 0, + "content": "[44] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, 2018. 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[45] Hao Tang, Zechao Li, Zhimao Peng, and Jinhui Tang. Blockmix: meta regularization and self-calibrated inference for metric-based meta-learning. In ACM Multimedia, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[46] Hao Tang, Chengcheng Yuan, Zechao Li, and Jinhui Tang. Learning attention-guided pyramidal features for few-shot fine-grained recognition. PR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.472, + 0.756 + ], + "angle": 0, + "content": "[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[48] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1, 2, 5, 6, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[49] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[50] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[51] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. arXiv preprint, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[52] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. arXiv preprint, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[53] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[54] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[55] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[56] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.446 + ], + "angle": 0, + "content": "[57] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[58] Ji Zhang, Jingkuan Song, Lianli Gao, Ye Liu, and Heng Tao Shen. Progressive meta-learning with curriculum. TCVT, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.894, + 0.546 + ], + "angle": 0, + "content": "[59] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[60] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In ICLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.893, + 0.631 + ], + "angle": 0, + "content": "[61] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.893, + 0.674 + ], + "angle": 0, + "content": "[62] Zhedong Zheng and Yi Yang. Rectifying pseudo label learning via uncertainty estimation for domain adaptive semantic segmentation. IJCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.893, + 0.717 + ], + "angle": 0, + "content": "[63] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.893, + 0.759 + ], + "angle": 0, + "content": "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.893, + 0.8 + ], + "angle": 0, + "content": "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[66] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 3, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.888 + ], + "angle": 0, + "content": "[67] Linhai Zhuo, Yuqian Fu, Jingjing Chen, Yixin Cao, and YuGang Jiang. Tgdm: Target guided dynamic mixup for cross-domain few-shot learning. In ACM Multimedia, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "24584" + } + ] +] \ No newline at end of file diff --git a/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_origin.pdf b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f4545dd4b4dae0f53892c23a3859c72d3ac86779 --- /dev/null +++ b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/8636d260-66e4-4ec3-b4dc-94bb000bc5ce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6b3cbd9a0dcf588bdd479a236eaa352788e4b1ef9a0e9d0aecd6763d47a541c +size 915607 diff --git a/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/full.md b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a6e141a12e6eae9468cf1db81e63930db476eb6e --- /dev/null +++ b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/full.md @@ -0,0 +1,281 @@ +# StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot Learning + +Yuqian Fu $^{1}$ , Yu Xie $^{2}$ , Yanwei Fu $^{3}$ , Yu-Gang Jiang $^{1*}$ + +$^{1}$ Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University $^{2}$ Purple Mountain Laboratories, Nanjing, China. $^{3}$ School of Data Science, Fudan University $\{\text { fuyq20, yxie18, yanweifu, ygj } \} @$ fudan.edu.cn + +# Abstract + +Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN [10] empirically shows that spanning the style distribution of the source data helps alleviate this issue. However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles "real" and "easy", which still fall into the original set of the source styles. Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. Particularly, our style attack method synthesizes both "virtual" and "hard" adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. Code is available at https://github.com/lovelyqian/StyleAdv-CDFSL. + +# 1. Introduction + +This paper studies the task of Cross-Domain Few-Shot Learning (CD-FSL) which addresses the Few-Shot Learning (FSL) problem across different domains. As a general recipe for FSL, episode-based meta-learning strategy + +has also been adopted for training CD-FSL models, e.g., FWT [48], LRP [42], ATA [51], and wave-SAN [10]. Generally, to mimic the low-sample regime in testing stage, meta learning samples episodes for training the model. Each episode contains a small labeled support set and an unlabeled query set. Models learn meta knowledge by predicting the categories of images contained in the query set according to the support set. The learned meta knowledge generalizes the models to novel target classes directly. + +Empirically, we find that the changes of visual appearances between source and target data is one of the key causes that leads to the domain gap in CD-FSL. Interestingly, waveSAN [10], our former work, shows that the domain gap issue can be alleviated by augmenting the visual styles of source images. Particularly, wave-SAN proposes to augment the styles, in the form of Adaptive Instance Normalization (AdaIN) [22], by randomly sampling two source episodes and exchanging their styles. However, despite the efficacy of wave-SAN, such a naive style generation method suffers from two limitations: 1) The swap operation makes the styles always be limited in the "real" style set of the source dataset; 2) The limited real styles further lead to the generated styles too "easy" to learn. Therefore, a natural question is whether we can synthesize "virtual" and "hard" styles for learning a more robust CD-FSL model? Formally, we use "real/virtual" to indicate whether the styles are originally presented in the set of source styles, and define "easy/hard" as whether the new styles make meta tasks more difficult. + +To that end, we draw inspiration from the adversarial training, and propose a novel meta Style Adversarial training method (StyleAdv) for CD-FSL. StyleAdv plays the minimax game in two iterative optimization loops of metaintraining. Particularly, the inner loop generates adversarial styles from the original source styles by adding perturbations. The synthesized adversarial styles are supposed to be more challenging for the current model to recognize, thus, increasing the loss. Whilst the outer loop optimizes the whole network by minimizing the losses of recognizing the images with both original and adversarial styles. Our ultimate goal is to enable learning a model that is robust to various styles, be- + +yond the relatively limited and simple styles from the source data. This can potentially improve the generalization ability on novel target domains with visual appearance shifts. + +Formally, we introduce a novel style adversarial attack method to support the inner loop of StyleAdv. Inspired yet different from the previous attack methods [14,34], our style attack method perturbs and synthesizes the styles rather than image pixels or features. Technically, we first extract the style from the input feature map, and include the extracted style in the forward computation chain to obtain its gradient for each training step. After that, we synthesize the new style by adding a certain ratio of gradient to the original style. Styles synthesized by our style adversarial attack method have the good properties of "hard" and "virtual". Particularly, since we perturb styles in the opposite direction of the training gradients, our generation leads to the "hard" styles. Our attack method results in totally "virtual" styles that are quite different from the original source styles. + +Critically, our style attack method makes progressive style synthesizing, with changing style perturbation ratios, which makes it significantly different from vanilla adversarial attacking methods. Specifically, we propose a novel progressive style synthesizing strategy. The naive solution of directly plugging-in perturbations is to attack each block of the feature embedding module individually, which however, may results in large deviations of features from the high-level block. Thus, our strategy is to make the synthesizing signal of the current block be accumulated by adversarial styles from previous blocks. On the other hand, rather than attacking the models by fixing the attacking ratio, we synthesize new styles by randomly sampling the perturbation ratio from a candidate pool. This facilitates the diversity of the synthesized adversarial styles. Experimental results have demonstrated the efficacy of our method: 1) our style adversarial attack method does synthesize more challenging styles, thus, pushing the limits of the source visual distribution; 2) our StyleAdv significantly improves the base model and outperforms all other CD-FSL competitors. + +We highlight our StyleAdv is model-agnostic and complementary to other existing FSL or CD-FSL models, e.g., GNN [12] and FWT [48]. More importantly, to benefit from the large-scale pretrained models, e.g., DINO [2], we further explore adapting our StyleAdv to improve the Vision Transformer (ViT) [5] backbone in a non-parametric way. Experimentally, we show that StyleAdv not only improves CNN-based FSL/CD-FSL methods, but also improves the large-scale pretrained ViT model. + +Finally, we summarize our contributions. 1) A novel meta style adversarial training method, termed StyleAdv, is proposed for CD-FSL. By first perturbing the original styles and then forcing the model to learn from such adversarial styles, StyleAdv improves the robustness of CD-FSL models. 2) We present a novel style attack method with the novel + +progressive synthesizing strategy in changing attacking ratios. Diverse "virtual" and "hard" styles thus are generated. 3) Our method is complementary to existing FSL and CD-FSL methods; and we validate our idea on both CNN-based and ViT-based backbones. 4) Extensive results on eight unseen target datasets indicate that our StyleAdv outperforms previous CD-FSL methods, building a new SOTA result. + +# 2. Related Work + +Cross-Domain Few-Shot Learning. FSL which aims at freeing the model from reliance on massive labeled data has been studied for many years [12, 25, 39, 41, 43, 45, 46, 56, 58]. Particularly, some recent works, e.g., CLIP [38], CoOp [65], CLIP-Adapter [11], Tip-Adapter [59], and PMF [19] explore promoting the FSL with large-scale pretrained models. Particularly, PMF contributes a simple pipeline and builds a SOTA for FSL. As an extended task from FSL, CD-FSL [1, 7–10, 15, 16, 23, 29, 32, 37, 42, 48, 51, 60, 67] mainly solves the FSL across different domains. Typical meta-learning based CD-FSL methods include FWT [48], LRP [42], ATA [51], AFA [20], and wave-SAN [10]. Specifically, FWT and LRP tackle CD-FSL by refining batch normalization layers and using the explanation model to guide training. ATA, AFA, and wave-SAN propose to augment the image pixels, features, and visual styles, respectively. Several transfer-learning based CD-FSL methods, e.g., BSCD-FSL (also known as Fine-tune) [16], BSR [33], and NSAE [32] have also been explored. These methods reveal that finetuning helps improving the performances on target datasets. Other works that introduce extra data or require multiple domain datasets for training include STARTUP [37], Meta-FDMixup [8], Me-D2N [9], TGDM [67], TriAE [15], and DSL [21]. + +Adversarial Attack. The adversarial attack aims at misleading models by adding some bespoke perturbations to input data. To generate the perturbations effectively, lots of adversarial attack methods have been proposed [6, 14, 26, 27, 34, 36, 55, 57]. Most of the works [6, 14, 34, 36] attack the image pixels. Specifically, FGSM [14] and PGD [34] are two most classical and famous attack algorithms. Several works [26, 27, 62] attack the feature space. Critically, few methods [57] attack styles. Different from these works that aim to mislead the models, we perturb the styles to tackle the visual shift issue for CD-FSL. + +Adversarial Few-Shot Learning. Several attempts [13, 28, 30, 40, 52] that explore adversarial learning for FSL have been made. Among them, MDAT [30], AQ [13], and MetaAdv [52] first attack the input image and then train the model using the attacked images to improve the defense ability against adversarial samples. Shen et al. [40] attacks the feature of the episode to improve the generalization capability of FSL models. Note that ATA [51] and AFA [20], two CD-FSL methods, also adopt the adversarial learning. + +However, we are greatly different from them. ATA and AFA perturb image pixels or features, while we aim at bridging the visual gap by generating diverse hard styles. + +Style Augmentation for Domain Shift Problem. Augmenting the style distribution for narrowing the domain shift issue has been explored in domain generation [31, 54, 66], image segmentation [3, 63], person re-ID [61], and CD-FSL [10]. Concretely, MixStyle [66], AdvStyle [63], DSU [31], and wave-SAN [10] synthesize styles without extra parameters via mixing, attacking, sampling from a Gaussian distribution, and swapping. MaxStyle [3] and L2D [54] require additional network modules and complex auxiliary tasks to help generate the new styles. Typically, AdvStyle [63] is the most related work to us. Thus, we highlight the key differences: 1) AdvStyle attacks styles on the image, while we attack styles on multiple feature spaces with a progressive attacking method; 2) AdvStyle uses the same task loss (segmentation) for attacking and optimization; in contrast, we use the classical classification loss to attack the styles, while utilize the task loss (FSL) to optimize the whole network. + +# 3. StyleAdv: Meta Style Adversarial Training + +Task Formulation. Episode $\mathcal{T} = ((S, Q), Y)$ is randomly sampled as the input of each meta-task, where $Y$ represents the global class labels of the episode images with respect to $\mathcal{C}^{tr}$ . Typically, each meta-task is formulated as an $N$ -way $K$ -shot problem. That is, for each episode $\mathcal{T}$ , $N$ classes with $K$ labeled images are sampled as the support set $S$ , and the same $N$ classes with another $M$ images are used to constitute the query set $Q$ . The FSL or CD-FSL models predict the probability $P$ that the images in $Q$ belong to $N$ categories according to $S$ . Formally, we have $|S| = NK$ , $|Q| = NM$ , $|P| = NM \times N$ . + +FGSM and PGD Attackers. We briefly summarize the algorithms for FGSM [14] and PGD [34], two most famous attacking methods. Given image $x$ with label $y$ , FGSM attacks the $x$ by adding a ratio $\epsilon$ of signed gradients with respect to the $x$ resulting in the adversarial image $x^{adv}$ as, + +$$ +x ^ {a d v} = x + \epsilon \cdot \operatorname {s i g n} \left(\nabla_ {x} J (\theta , x, y)\right), \tag {1} +$$ + +where $J(\cdot)$ and $\theta$ denote the object function and the learnable parameters of a classification model. PGD can be regarded as a variant of FGSM. Different from the FGSM that only attacks once, PGD attacks the image in an iterative way and sets a random start (abbreviated as RT) for $x$ as, + +$$ +x _ {0} ^ {a d v} = x + k _ {R T} \cdot \mathcal {N} (0, I), \tag {2} +$$ + +$$ +x _ {t} ^ {a d v} = x _ {t - 1} ^ {a d v} + \epsilon \cdot \operatorname {s i g n} \left(\nabla_ {x} J (\theta , x, y)\right), \tag {3} +$$ + +where $k_{RT},\epsilon$ are hyper-parameters. $\mathcal{N}$ is Gaussian noises. + +# 3.1. Overview of Meta Style Adversarial Learning + +To alleviate the performance degradation caused by the changing visual appearance, we tackle CD-FSL by promoting the robustness of models on recognizing various styles. Thus, we expose our FSL model to some challenging virtual styles beyond the image styles that existed in the source dataset. To that end, we present the novel StyleAdv adversarial training method. Critically, rather than adding perturbations to image pixels, we particularly focus on adversarially perturbing the styles. The overall framework of our StyleAdv is illustrated in Figure 1. Our StyleAdv contains a CNN/ViT backbone $E$ , a global FC classifier $f_{cls}$ , and a FSL classifier $f_{fsl}$ with learnable parameters $\theta_E, \theta_{cls}, \theta_{fsl}$ , respectively. Besides, our core style attack method, a novel style extraction module, and the AdaIN are also included. + +Overall, we learn the StyleAdv by solving a minimax game. Specifically, the minimax game shall involve two iterative optimization loops in each meta-train step. Particularly, + +- Inner loop: synthesizing new adversarial styles by attacking the original source styles; the generated styles will increase the loss of the current network. +- Outer loop: optimizing the whole network by classifying source images with both original and adversarial styles; this process will decrease the loss. + +# 3.2. Style Extraction from CNNs and ViTs + +Adaptive Instance Normalization (AdaIN). We recap the vanilla AdaIN [22] proposed for CNN in style transfer. Particularly, AdaIN reveals that the instance-level mean and standard deviation (abbreviated as mean and std) convey the style information of the input image. Denoting the mean and std as $\mu$ and $\sigma$ , AdaIN (denoted as $\mathcal{A}$ ) reveals that the style of $F$ can be transferred to that of $F_{tgt}$ by replacing the original style $(\mu, \sigma)$ with the target style $(\mu_{tgt}, \sigma_{tgt})$ : + +$$ +\mathcal {A} (F, \mu_ {t g t}, \sigma_ {t g t}) = \sigma_ {t g t} \frac {F - \mu (F)}{\sigma (F)} + \mu_ {t g t}. \tag {4} +$$ + +Style Extraction for CNN Features. As shown in the upper part of Figure 1 (b), let $F \in \mathcal{R}^{B \times C \times H \times W}$ indicates the input feature batch, where $B, C, H,$ and $W$ denote the batch size, channel, height, and width of the feature $F$ , respectively. As in AdaIN, the mean $\mu$ and std $\sigma$ of $F$ are defined as: + +$$ +\mu (\mathrm {F}) _ {\mathrm {b}, \mathrm {c}} = \frac {1}{H W} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} F _ {b, c, h, w}, \tag {5} +$$ + +$$ +\sigma (\mathrm {F}) _ {\mathrm {b}, \mathrm {c}} = \sqrt {\frac {1}{H W} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \left(F _ {b , c , h , w} - \mu_ {\mathrm {b , c}} (F)\right) ^ {2} + \epsilon}, \tag {6} +$$ + +where $\mu ,\sigma \in \mathcal{R}^{B\times C}$ + +![](images/7c14f659c72ee51e20f7668e69c4d9eb75f8585782999416c1fc704fb586f91c.jpg) +(a) Overall Framework of Our StyleAdv Method + +![](images/454d49a5abffa07b993dd1a1099de2eb23c3ae59eab586f9c57962ca8559b7e8.jpg) +(b) Style Extraction For CNN/ViT Features +Figure 1. (a): Overview of StyleAdv method. The inner loop synthesizes adversarial styles, while the outer loop optimizes the whole network. (b): Style extraction for CNN-based and ViT-based features (illustration with $B = 1$ ). + +Meta Information Extraction for ViT Features. We explore extracting the meta information of the ViT features as the manner of CNN. Intuitively, such meta information can be regarded as a unique "style" of ViTs. As shown in Figure 1 (b), we take an input batch data with image split into $P \times P$ patches as an example. The ViT encoder will encode the batch patches into a class (cls) token ( $F_{cls} \in \mathcal{R}^{B \times C}$ ) and a patch tokens ( $F_0 \in \mathcal{R}^{B \times P^2 \times C}$ ). To be compatible with AdaIN, we reshape the $F_0$ as $F \in \mathcal{R}^{B \times C \times P \times P}$ . At this point, we can calculate the meta information for patch tokens $F$ as in Eq. 5 and Eq. 6. Essentially, note that the transformer integrates the positional embedding into the patch representation, the spatial relations thus could be considered still hold in the patch tokens. This supports us to reform the patch tokens $F_0$ as a spatial feature map $F$ . To some extent, this can also be achieved by applying the convolution on the input data via a kernel of size $P \times P$ (as indicated by dashed arrows in Figure 1 (b)). + +# 3.3. Inner Loop: Style Adversarial Attack Method + +We propose a novel style adversarial attack method - Fast Style Gradient Sign Method (Style-FGSM) to accomplish the inner loop. As shown in Figure 1, given an input source episode $(\mathcal{T},Y)$ , we first forward it into the backbone $E$ and the FC classifier $f_{cls}$ producing the global classification loss $\mathcal{L}_{cls}$ (as illustrated in the ① paths). During this process, a key step is to make the gradient of the style available. To achieve that, let $F_{\mathcal{T}}$ denotes the features of $\mathcal{T}$ , we obtain the style $(\mu ,\sigma)$ of $F_{\mathcal{T}}$ as in Sec. 3.2. After that, we reform the original episode feature as $\mathcal{A}(F_{\mathcal{T}},\mu ,\sigma)$ . And the reformed + +feature is actually used for the forward propagation. In this way, we include $\mu$ and $\sigma$ in our forward computation chain; and thus, we could access the gradients of them. + +With the gradients in ② paths, we then attack $\mu$ and $\sigma$ as FGSM does - adding a small ratio $\epsilon$ of the signed gradients with respect to $\mu$ and $\sigma$ , respectively. + +$$ +\mu^ {a d v} = \mu + \epsilon \cdot \operatorname {s i g n} \left(\nabla_ {\mu} J \left(\theta_ {E}, \theta_ {f _ {c l s}}, \mathcal {A} \left(F _ {\tau}, \mu , \sigma\right), Y\right)\right), \tag {7} +$$ + +$$ +\sigma^ {a d v} = \sigma + \epsilon \cdot \operatorname {s i g n} \left(\nabla_ {\sigma} J \left(\theta_ {E}, \theta_ {f _ {c l s}}, \mathcal {A} \left(F _ {\mathcal {T}}, \mu , \sigma\right), Y\right)\right), \tag {8} +$$ + +where the $J(\cdot)$ is the cross-entropy loss between classification predictions and ground truth $Y$ , i.e., $\mathcal{L}_{cls}$ . Inspired by the random start of PGD, we also add random noises $k_{RT} \cdot \mathcal{N}(0, I)$ to $\mu$ and $\sigma$ before attacking. $\mathcal{N}(0, I)$ refers to Gaussian noises and $k_{RT}$ is a hyper-parameter. Our Style-FGSM enables us to generate both "virtual" and "hard" styles. + +Progressive Style Synthesizing Strategy: To prevent the high-level adversarial feature from deviating, we propose to apply our style-FGSM in a progressive strategy. Concretely, the embedding module $E$ has three blocks $E_{1}$ , $E_{2}$ , and $E_{3}$ , with the corresponding features $F_{1}$ , $F_{2}$ , and $F_{3}$ . For the first block, we use $(\mu_{1}, \sigma_{1})$ to denote the original styles of $F_{1}$ . The adversarial styles $(\mu_{1}^{adv}, \sigma_{1}^{adv})$ are obtained directly as in Eq. 7 and Eq. 8. For subsequent blocks, the attack signals on the current block $i$ are those accumulated from the block 1 to block $i - 1$ . Take the second block as an example, the block feature $F_{2}$ is not simply extracted by $E_{2}(F_{1})$ . Instead, we have $F_{2}' = E_{2}(F_{1}^{adv})$ , where $F_{1}^{adv} = \mathcal{A}(F_{1}, \mu_{1}^{adv}, \sigma_{1}^{adv})$ . Attacking on $F_{2}'$ results in the adversarial styles $(\mu_{2}^{adv}, \sigma_{2}^{adv})$ . Accordingly, we gen + +erate $(\mu_3^{adv},\sigma_3^{adv})$ for the last block. The illustration of the progressive attacking strategy is attached in the Appendix. Changing Style Perturbation Ratios: Different from the vanilla FGSM [14] or PGD [34], our style attacking algorithm is expected to synthesize new styles with diversity. Thus, instead of using a fixed attacking ratio $\epsilon$ , we randomly sample $\epsilon$ from a candidate list $\epsilon_{list}$ as the current attacking ratio. Despite the randomness of $\epsilon$ , we still synthesize styles in a more challenging direction, $\epsilon$ only affects the extent. + +# 3.4. Outer Loop: Optimize the StyleAdv Network + +For each meta-train iteration with clean episode $\mathcal{T}$ as input, our inner loop produces adversarial styles $(\mu_1^{adv},\sigma_1^{adv})$ $(\mu_{2}^{adv},\sigma_{2}^{adv})$ , and $(\mu_3^{adv},\sigma_3^{adv})$ . As in Figure 1, the goal of the outer loop is to optimize the whole StyleAdv with both the clean feature $F$ and the style attacked feature $F^{adv}$ utilized as the training data. Typically, the clean episode feature $F$ can be obtained directly as $E(T)$ as in 3 paths. + +In 4 paths, we obtain the $F^{adv}$ by transferring the original style of $F$ to the corresponding adversarial attacked styles. Similar with the progressive style-FGSM, we have $F_{1}^{adv} = \mathcal{A}(E_{1}(\mathcal{T}),\mu_{1}^{adv},\sigma_{1}^{adv})$ , $F_{2}^{adv} = \mathcal{A}(E_{2}(F_{1}^{adv}),\mu_{2}^{adv},\sigma_{2}^{adv})$ , and $F_{3}^{adv} = \mathcal{A}(E_{3}(F_{2}^{adv}),\mu_{3}^{adv},\sigma_{3}^{adv})$ . Finally, $F^{adv}$ is obtained by applying an average pooling layer to $F_{3}^{adv}$ . A skip probability $p_{skip}$ is set to decide whether to skip the current attacking. Conducting FSL tasks for both the clean feature $F$ and style attacked feature $F^{adv}$ results in two FSL predictions $P_{fsl}$ , $P_{fsl}^{adv}$ , and two FSL classification losses $\mathcal{L}_{fsl}$ , $\mathcal{L}_{fsl}^{adv}$ . + +Further, despite the styles of $F^{adv}$ shifts from that of $F$ , we encourage that the semantic content should be still consistent as in wave-SAN [10]. Thus we add a consistent constraint to the predictions of $P_{fsl}$ and $P_{fsl}^{adv}$ resulting in the consistent loss $\mathcal{L}_{cons}$ as, + +$$ +\mathcal {L} _ {\text {c o n s}} = \mathrm {K L} \left(P _ {f s l}, P _ {f s l} ^ {\text {a d v}}\right), \tag {9} +$$ + +where $\mathrm{KL}()$ is Kullback-Leibler divergence loss. In addition, we have the global classification loss $\mathcal{L}_{cls}$ . This ensures that $\theta_{cls}$ is optimized to provide correct gradients for style-FGSM. The final meta-objective of StyleAdv is as, + +$$ +\mathcal {L} = \mathcal {L} _ {f s l} + \mathcal {L} _ {f s l} ^ {a d v} + \mathcal {L} _ {c o n s} + \mathcal {L} _ {c l s}. \tag {10} +$$ + +Note that our StyleAdv is model-agnostic and orthogonal to existing FSL and CD-FSL methods. + +# 3.5. Network Inference + +Applying StyleAdv Directly for Inference. Our StyleAdv facilitates making CD-FSL model more robust to style shifts. Once the model is meta-trained, we can employ it for inference directly by feeding the testing episode into the $E$ and the $f_{cls}$ . The class with the highest probability will be taken as the predicted result. + +Finetuning StyleAdv Using Target Examples. As indicated in previous works [16, 32, 33, 51], finetuning CD-FSL models on target examples helps improve the model performance. Thus, to further promote the performance of StyleAdv, we also equip it with the fintuning strategy forming an upgraded version ("StyleAdv-FT"). Specifically, as in ATA-FT [51], for each novel testing episode, we augment the novel support set to form pseudo episodes as training data for tuning the meta-trained model. + +# 4. Experiments + +Datasets. We take two CD-FSL benchmarks proposed in BSCD-FSL [16] and FWT [48]. Both of them take mini-Imagenet [39] as the source dataset. Two disjoint sets split from mini-Imagenet form $\mathcal{D}^{tr}$ and $\mathcal{D}^{eval}$ . Totally eight datasets including ChestX [53], ISIC [4, 47], EuroSAT [18], CropDisease [35], CUB [50], Cars [24], Places [64], and Plantae [49] are taken as novel target datasets. The former four datasets included in BSCD-FSL's benchmark cover medical images varying from X-ray to dermoscopic skin lesions, and natural images from satellite pictures to plant disease photos. While the latter four datasets that focus on more fine-grained concepts such as birds and cars are contained in FWT. These eight target datasets serve as testing set $\mathcal{D}^{te}$ , respectively. + +Network Modules. For typical CNN based network, following previous CD-FSL methods [10, 42, 48, 51], ResNet-10 [17] is selected as the embedding module while GNN [12] is selected as the FSL classifier; For the emerging ViT based network, following PMF [19], we use the ViT-small [5] and the ProtoNet [41] as the embedding module and the FSL classifier, respectively. Note that, the ViT-small is pretrained on ImageNet1K by DINO [2] as in PMF. The $f_{cls}$ is built by a fully connected layer. + +Implementation Details. The 5-way 1-shot and 5-way 5-shot settings are conducted. Taking ResNet10 as backbone, we meta train the network for 200 epochs, each epoch contains 120 meta tasks. Adam with a learning rate of 0.001 is utilized as the optimizer. Taking ViT-small as backbone, the meta train stage takes 20 epoch, each epoch contains 2000 meta tasks. The SGD with a initial learning rate of 5e-5 and 0.001 are used for optimize the $E(\cdot)$ and the $f_{cls}$ , respectively. The $\epsilon_{list}$ , $k_{RT}$ of Style-FGSM attacker are set as [0.8, 0.08, 0.008], $\frac{16}{255}$ . The probability $p_{skip}$ of random skipping the attacking is chosen from {0.2, 0.4}. We evaluate our network with 1000 randomly sampled episodes and report average accuracy (\%) with a 95% confidence interval. Both the results of our "StyleAdv" and "StyleAdv-FT" are reported. The details of the finetuning are attached in Appendix. ResNet-10 based models are trained and tested on a single GeForce GTX 1080, while ViT-small based models require a single NVIDIA GeForce RTX 3090. + +
1-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--22.00±0.4632.02±0.6663.69±1.0364.48±1.0845.69±0.6831.79±0.5153.10±0.8035.60±0.5643.55
FWT [48]RN10--22.04±0.4431.58±0.6762.36±1.0566.36±1.0447.47±0.7531.61±0.5355.77±0.7935.95±0.5844.14
LRP [42]RN10--22.11±0.2030.94±0.3054.99±0.5059.23±0.5048.29±0.5132.78±0.3954.83±0.5637.49±0.4342.58
ATA [51]RN10--22.10±0.2033.21±0.4061.35±0.5067.47±0.5045.00±0.5033.61±0.4053.57±0.5034.42±0.4043.84
AFA [20]RN10--22.92±0.2033.21±0.3063.12±0.5067.61±0.5046.86±0.5034.25±0.4054.04±0.6036.76±0.4044.85
wave-SAN [10]RN10--22.93±0.4933.35±0.7169.64±1.0970.80±1.0650.25±0.7433.55±0.6157.75±0.8240.71±0.6647.37
StyleAdv (ours)RN10--22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
ATA-FT [51]RN10Y-22.15±0.2034.94±0.4068.62±0.5075.41±0.5046.23±0.5037.15±0.4054.18±0.5037.38±0.4047.01
StyleAdv-FT (ours)RN10Y-22.64±0.3535.76±0.5272.92±0.7580.69±0.2848.49±0.7235.09±0.5558.58±0.8341.13±0.6749.41
PMF* [19]ViT-smallYDINO/IN1K21.73±0.3030.36±0.3670.74±0.6380.79±0.6278.13±0.6637.24±0.5771.11±0.7153.60±0.6655.46
StyleAdv (ours)ViT-small-DINO/IN1K22.92±0.3233.05±0.4472.15±0.6581.22±0.6184.01±0.5840.48±0.5772.64±0.6755.52±0.6657.75
StyleAdv-FT (ours)ViT-smallYDINO/IN1K22.92±0.3233.99±0.4674.93±0.5884.11±0.5784.01±0.5840.48±0.5772.64±0.6755.52±0.6658.57
5-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--25.27±0.4643.94±0.6783.64±0.7787.96±0.6762.25±0.6544.28±0.6370.84±0.6552.53±0.5958.84
FWT [48]RN10--25.18±0.4543.17±0.7083.01±0.7987.11±0.6766.98±0.6844.90±0.6473.94±0.6753.85±0.6259.77
LRP [42]RN10--24.53±0.3044.14±0.4077.14±0.4086.15±0.4064.44±0.4846.20±0.4674.45±0.4754.46±0.4658.94
ATA [51]RN10--24.32±0.4044.91±0.4083.75±0.4090.59±0.3066.22±0.5049.14±0.4075.48±0.4052.69±0.4060.89
AFA [20]RN10--25.02±0.2046.01±0.4085.58±0.4088.06±0.3068.25±0.5049.28±0.5076.21±0.5054.26±0.4061.58
wave-SAN [10]RN10--25.63±0.4944.93±0.6785.22±0.7189.70±0.6470.31±0.6746.11±0.6676.88±0.6357.72±0.6462.06
StyleAdv (ours)RN10--26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
Fine-tune [16]RN10Y-25.97±0.4148.11±0.6479.08±0.6189.25±0.5164.14±0.7752.08±0.7470.06±0.7459.27±0.7061.00
ATA-FT [51]RN10Y-25.08±0.2049.79±0.4089.64±0.3095.44±0.2069.83±0.5054.28±0.5076.64±0.4058.08±0.4064.85
NSAE [32]RN10Y-27.10±0.4454.05±0.6383.96±0.5793.14±0.4768.51±0.7654.91±0.7471.02±0.7259.55±0.7464.03
BSR [33]RN10Y-26.84±0.4454.42±0.6680.89±0.6192.17±0.4569.38±0.7657.49±0.7271.09±0.6861.07±0.7664.17
StyleAdv-FT (ours)RN10Y-26.24±0.3553.05±0.5491.64±0.4396.51±0.2870.90±0.6356.44±0.6879.35±0.6164.10±0.6467.28
PMF [19]ViT-smallYDINO/IN1K27.2750.1285.9892.96-----
StyleAdv (ours)ViT-small-DINO/IN1K26.97±0.3347.73±0.4488.57±0.3494.85±0.3195.82±0.2761.73±0.6288.33±0.4075.55±0.5472.44
StyleAdv-FT (ours)ViT-smallYDINO/IN1K26.97±0.3351.23±0.5190.12±0.3395.99±0.2795.82±0.2766.02±0.6488.33±0.4078.01±0.5474.06
+ +Table 1. Results of 5-way 1-shot/5-shot tasks. "FT" means whether the finetuning stage is employed. "LargeP" represents if large pretrained models are used for model initialization. "RN10" is short for "ResNet-10". * denotes results are reported by us. Results perform best are bolded. Whether based on ResNet-10 or ViT-small, our method outperforms other competitors significantly. + +# 4.1. Comparison with the SOTAs + +We compare our StyleAdv/StyleAdv-FT against several most representative and competitive CD-FSL methods. Concretly, with the ResNet-10 (abbreviated as RN10) as backbone, totally nine methods including GNN [12], FWT [48], LRP [42], ATA [51], AFA [20], wave-SAN [10], Finetune [16], NSAE [32], and BSR [33] are introduced as our competitors. Among them, the former six competitors are meta-learning based method that used for inference directly, thus we compare our "StyleAdv" against them for a fair comparison. Typically, the GNN [12] works as a base model. The Fine-tune [16], NSAE [32], BSR [33], and ATA-FT [51] (formed by finetuning ATA) all require finetuning model during inference, thus our "StyleAdv-FT" is used. With the ViT as backbone, the most recent and competitive PMF (SOTA method for FSL) is compared. For fair comparisons, we follow the same pipeline proposed in PMF [19]. Note that we promote CD-FSL models with only one single source domain. Those methods that use extra training datasets, e.g., STARTUP [37], meta-FDMixup [8], and DSL [21] are not considered. The comparison results are given in Table 1. + +For all results, our method outperforms all the listed CD-FSL competitors significantly and builds a new state of the art. Our StyleAdv-FT (ViT-small) on average achieves $58.57\%$ and $74.06\%$ on 5-way 1-shot and 5-shot, respectively. Our StyleAdv (RN10) and StyleAdv-FT (RN10) also + +beats all the meta-learning based or transfer-learning (finetuning) based methods. Besides of the state-of-the-art accuracy, we also have other worth-mentioning observations. 1) We show that our StyleAdv method is a general solution for both CNN-based models and ViT-based models. Typically, based on ResNet10, our StyleAdv and StyleAdv-FT improve the base GNN by up to $4.93\%$ and $8.44\%$ on 5-shot setting. Based on ViT-small, at most cases, our StyleAdv-FT outperforms the PMF by a clear margin. More results of building StyleAdv upon other FSL or CD-FSL methods can be found in the Appendix. 2) Comparing FWT, LRP, ATA, AFA, waveSAN, and our StyleAdv, we find that StyleAdv performs best, followed by wave-SAN, then comes the AFA, ATA, FWT, and LRP. This phenomenon indicates that tackling CD-FSL by solving the visual shift problem is indeed more effective than other perspectives, e.g., adversarial training by perturbing the image features (AFA) or image pixels (ATA), transforms the normalization layers in FWT, and explanation guided training in LRP. 3) For the comparison between StyleAdv and wave-SAN that both tackles the visual styles, we notice that StyleAdv outperforms the wave-SAN in most cases. This demonstrates that the styles generated by our StyleAdv are more conducive to learning robust CD-FSL models than the style augmentation method proposed in wave-SAN. This justifies our idea of synthesizing more challenging ("hard and virtual") styles. + +4) Overall, the large-scale pretrained model promotes the CD-FSL obviously. Take 1-shot as an example, StyleAdv-FT (ViT-small) boosts the StyleAdv-FT (RN10) by $9.16\%$ on average. However, we show that the performance improvement varies greatly on different target domains. Generally, for target datasets with relative small domain gap, e.g., CUB and Plantae, models benefit a lot; otherwise, the improvement is limited. 5) We also find that under the cross-domain scenarios, finetuning model on target domain, e.g., NSAE, BSR do show an advantage over purely meta-learning based methods, e.g., FWT, LRP, and wave-SAN. However, to finetune model using extremely few examples, e.g., 5-way 1-shot is much harder than on relatively larger shots. This may explain why those finetune-based methods do not conduct experiments on 1-shot setting. + +Effectiveness of Style-FGSM Attacker. To show the advantages of our progressive style synthesizing strategy and attacking with changing perturbation ratios, we compare our Style-FGSM against several variants and report the results in Figure 2. Specifically, for Figure 2 (a), we compare our style-FGSM against the variant that attacks the blocks individually. Results show that attacking in a progressive way exceeds the naive individual strategy in most cases. For Figure 2 (b), to demonstrate how the performance will be affected by fixed attacking ratios, we also conduct experiments with different $\epsilon_{list}$ . Since we set the $\epsilon_{list}$ as [0.8, 0.08, 0.008], three different choices including [0.8], [0.08], and [0.008] are selected. From the results, we first notice that the best result can be reached by a single fixed ratio. However, sampling the attacking ratio from a pool of candidates achieves the best result in most cases. + +![](images/cddee983b9fc92f8c19d546c7064fd9cc0641f9698878d69e9b84345c247e2d8.jpg) +Figure 2. Effectiveness of the progressive style synthesizing strategy and the changing style perturbation ratios. The 5-way 1-shot results are reported. Models are built on ResNet10 and GNN. + +![](images/51da6b7adbc4530030ff63f7ce1b527353b7445d2c4f2ddcfb218fe7407a2093.jpg) +Figure 3. Visualization of wave-SAN and StyleAdv. (a): synthesized images; (b): meta-training losses; (c): T-SNE results. + +# 4.2. More Analysis + +Visualization of Hard Style Generation. To help understand the "hard" style generation of our method intuitively, as in Figure 3, we make several visualizations comparing StyleAdv against the wave-SAN. 1) As in Figure 3 (a), we show the stylized images generated by wave-SAN and our StyleAdv. The visualization is achieved by applying the style augmentation methods to input images. Specifically, for wave-SAN, the style is swapped with another randomly sampled source image; for StyleAdv, the results of attacking style with $\epsilon = 0.08$ are given. We observe that wave-SAN tends to exchange the global visual appearance, e.g., the color of the input image randomly. By contrast, StyleAdv prefers to disturb the important regions that are key to recognizing the image category. For example, the fur of the cat and the key parts (face and feet) of the dogs. These observations intuitively support our claim that our StyleAdv synthesize more harder styles than wave-SAN. 2) To quantitatively evaluate whether our StyleAdv introduces more challenging styles into the training stage, as in Figure 3 (b), we visualize the meta-training loss. Results reveal that the perturbed losses of wave-SAN oscillate around the original loss, while StyleAdv increases the original loss obviously. These phenomenons further validate that we perturb data towards a more difficult direction thus pushing the limits of style generation to a large extent. 3) To further show the advantages of StyleAdv over wave-SAN, as shown in Figure 3 (c), we visualize the high-level features extracted by the meta-trained wave-SAN and StyleAdv. Five classes (denoted by different colors) of mini-Imagenet are selected. T-SNE is used for reducing the feature dimensions. Results demonstrate that StyleAdv enlarges the inter-class distances making classes more distinguishable. + +Why Attack Styles Instead of Images or Features? A natural question may be why we choose to attack styles instead of other targets, e.g., the input image as in AQ [13], MDAT [30], and ATA [51] or the features as in Shen et al. [40] and AFA [20]? To answer this question, we compare our StyleAdv which attacks styles against attacking images + +
Attack TargetChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotImage22.71±0.3533.00±0.5367.00±0.8272.65±0.7548.15±0.7234.40±0.6057.89±0.8339.85±0.6446.96
Feature22.55±0.3532.95±0.5368.71±0.8170.86±0.7846.52±0.7034.07±0.5456.68±0.8139.62±0.6246.50
Style (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotImage24.92±0.3642.63±0.4784.18±0.5490.31±0.4766.37±0.6547.46±0.6775.94±0.6257.33±0.6561.14
Feature25.55±0.3743.71±0.5084.22±0.5591.71±0.4467.31±0.6750.26±0.6776.46±0.6557.39±0.6362.08
Style (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
+ +Table 2. Comparison results (%) of attacking image, feature, and styles. Models build upon ResNet10 and GNN classifier. + +
Augment MethodChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotStyleGaus†22.37±0.3531.48±0.5265.71±0.8269.25±0.8046.32±0.7232.69±0.5455.48±0.7937.27±0.6145.07
MixStyle [66]22.43±0.3533.21±0.5367.35±0.8068.80±0.8247.08±0.7333.39±0.5856.12±0.7838.03±0.6245.80
AdvStyle [63]22.04±0.3630.83±0.5265.19±0.8264.96±0.8147.43±0.7231.90±0.5253.95±0.7935.81±0.5944.01
DSU [31]22.35±0.3631.43±0.5164.55±0.8364.73±0.8147.74±0.7231.61±0.5354.81±0.8137.19±0.6144.30
Style-FGSM (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotStyleGaus†24.97±0.3741.74±0.4881.88±0.6189.71±0.4965.98±0.6745.03±0.6472.66±0.6856.66±0.6559.83
MixStyle [66]25.04±0.3643.77±0.5382.67±0.5888.90±0.5265.73±0.6645.91±0.6375.90±0.6356.59±0.6260.56
AdvStyle [63]25.03±0.3543.15±0.5083.09±0.5788.44±0.5266.42±0.6744.85±0.6474.14±0.6554.89±0.6460.00
DSU [31]25.02±0.3645.19±0.5280.30±0.6386.30±0.5667.94±0.6645.65±0.6375.17±0.6454.31±0.6259.99
Style-FGSM (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
+ +Table 3. Different style augmentation methods are compared. "StyleGaus†" means adding random Gaussian noises to the styles, where † represents it is proposed by us. "MixStyle [66]", "AdvStyle [63]" and "DSU [31]" are adapted from other tasks, e.g., domain generation. Results (%) conducted under 5-way 1-shot/5-shot settings. Methods are built upon the ResNet10 and GNN. + +and features by modifying the attack targets of our method. The 5-way 1-shot/5-shot results are given in Table 2. We highlight several points. 1) We notice that attacking image, feature, and style all improve the base GNN model (given in Table 1) which shows that all of them boost the generalization ability of the model by adversarial attacks. Interestingly, the results of our "Attack Image"/"Attack Feature" even outperform the well-designed CD-FSL methods ATA [51] and AFA [20] (shown in Table 1); 2) Our method has clear advantages over attacking images and features. This again indicates the superiority of tackling visual styles for narrowing the domain gap issue for CD-FSL. + +Is Style-FGSM Better than Other Style Augmentation Methods? To show the advantages of our Style-FGSM against other style augmentation methods, we introduce several competitors including "StyleGaus", MixStyle [66], AdvStyle [63], and DSU [31]. Typically, "StyleGaus" that adds random Gaussian noises into the styles is introduced as a simple but reasonable baseline. MixStyle [66], AdvStyle [63], and DSU [31] which are initially designed for other tasks, e.g., segmentation and domain generation are also adapted. The results are reported in Table 3. Comparing the results of StyleGuas with that reported in Table 1, we find that perturbing the styles on the feature level by simply adding random noises also improves the base GNN and even surpasses a few CD-FSL competitors on some target datasets. This phenomenon is consistent with the insight that augmenting the style distributions helps boost the CD-FSL methods. As for the comparison between our Style-FGSM and other advanced style augmentation competitors, we find that Style-FGSM performs better than all the MixStyle, AdvStyle, and + +DSU on both 1-shot and 5-shot settings. Typically, MixStyle and DSU both generate virtual styles, but their new styles are still relatively easy. This shows that our hard styles boost the model to a larger extent. AdvStyle generates both virtual and hard (adversarial) styles. However, it is still inferior to us. This indicates the advantages of our method that attacks in latent feature space and adopts two individual tasks for attacking and optimization. + +# 5. Conclusion + +This paper presents a novel model-agnostic StyleAdv for CD-FSL. Critically, to narrow the domain gap which is typically in the form of visual shifts, StyleAdv solves the minimax game of style adversarial learning: first adds perturbations to the source styles increasing the loss of the current model, then optimizes the model by forcing it to recognize both the clean and style perturbed data. Besides, a novel progressive style adversarial attack method termed style-FGSM is presented by us. Style-FGSM synthesizes diverse "hard" and "virtual" styles via adding the signed gradients to original clean styles. These generated styles support the max step of StyleAdv. Intuitively, by exposing the CD-FSL to adversarial styles which are more challenging than those limited real styles that exist in the source dataset, the generalization ability of the model is boosted. Our StyleAdv improves both CNN-based and ViT-based models. Extensive experiments indicate that our StyleAdv build new SOTAs. + +Acknowledgement. This project was supported by National Key R&D Program of China (No. 2021ZD0112804) and NSFC under Grant No. 62076067. + +# References + +[1] John Cai, Bill Cai, and Shen Sheng Mei. Damsl: Domain agnostic meta score-based learning. In CVPR, 2021. 2 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 5 +[3] Chen Chen, Zeju Li, Cheng Ouyang, Matt Sinclair, Wenjia Bai, and Daniel Rueckert. Maxstyle: Adversarial style composition for robust medical image segmentation. arXiv preprint, 2022. 3 +[4] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint, 2019. 5 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint, 2020. 2, 5 +[6] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A Kai Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2 +[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. TIP, 2022. 2 +[8] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In ACM Multimedia, 2021. 2, 6 +[9] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Me-d2n: Multi-expert domain decompositional network for cross-domain few-shot learning. In ACM Multimedia, 2022. 2 +[10] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Wave-san: Wavelet based style augmentation network for cross-domain few-shot learning. arXiv preprint, 2022. 1, 2, 3, 5, 6, 12 +[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint, 2021. 2 +[12] Victor Garcia and Joan Bruna. Few-shot learning with graph neural networks. arXiv preprint, 2017. 2, 5, 6, 12, 13 +[13] Micah Goldblum, Liam Fowl, and Tom Goldstein. Adversarially robust few-shot learning: A meta-learning approach. NeurIPS, 2020. 2, 7 +[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint, 2014. 2, 3, 5 +[15] Jiechao Guan, Manli Zhang, and Zhiwu Lu. Large-scale cross-domain few-shot learning. In ACCV, 2020. 2 +[16] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and + +Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2, 5, 6 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 2019. 5 +[19] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In CVPR, 2022. 2, 5, 6, 12, 13 +[20] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, 2022. 2, 6, 7, 8 +[21] Zhengdong Hu, Yifan Sun, and Yi Yang. Switch to generalize: Domain-switch learning for cross-domain few-shot classification. In ICLR, 2021. 2, 6 +[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 3 +[23] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard Radke. Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. NeurIPS, 2021. 2 +[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, 2013. 5 +[25] Manoj Kumar, Varun Kumar, Hadrien Glaude, Cyprien de Lichy, Aman Alok, and Rahul Gupta. Protoda: Efficient transfer learning for few-shot intent classification. In SLT Workshop, 2021. 2 +[26] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. NeurIPS, 2019. 2 +[27] Bo Li and Yevgeniy Vorobeychik. Feature cross-substitution in adversarial classification. NeurIPS, 2014. 2 +[28] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, 2020. 2 +[29] Pan Li, Shaogang Gong, Yanwei Fu, and Chengjie Wang. Ranking distance calibration for cross-domain few-shot learning. arXiv preprint, 2021. 2 +[30] Wenbin Li, Lei Wang, Xingxing Zhang, Jing Huo, Yang Gao, and Jiebo Luo. Defensive few-shot adversarial learning. arXiv preprint, 2019. 2, 7 +[31] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint, 2022. 3, 8 +[32] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In ICCV, 2021. 2, 5, 6 +[33] Bingyu Liu, Zhen Zhao, Zhenpeng Li, Jianan Jiang, Yuhong Guo, and Jieping Ye. Feature transformation ensemble model with batch spectral regularization for cross-domain few-shot classification. arXiv preprint, 2020. 2, 5, 6 + +[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint, 2017. 2, 3, 5 +[35] Sharada P Mohanty, David P Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in plant science, 2016. 5 +[36] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016. 2 +[37] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. arXiv preprint, 2020. 2, 6 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[39] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2, 5 +[40] Wei Shen, Ziqiang Shi, and Jun Sun. Learning from adversarial features for few-shot classification. arXiv preprint, 2019. 2, 7 +[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2, 5 +[42] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. arXiv preprint, 2020. 1, 2, 5, 6 +[43] Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In CVPR, 2019. 2 +[44] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, 2018. 12, 13 +[45] Hao Tang, Zechao Li, Zhimao Peng, and Jinhui Tang. Blockmix: meta regularization and self-calibrated inference for metric-based meta-learning. In ACM Multimedia, 2020. 2 +[46] Hao Tang, Chengcheng Yuan, Zechao Li, and Jinhui Tang. Learning attention-guided pyramidal features for few-shot fine-grained recognition. PR. 2 +[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 2018. 5 +[48] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1, 2, 5, 6, 12, 13 +[49] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 5 +[50] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 + +[51] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. arXiv preprint, 2021. 1, 2, 5, 6, 7, 8 +[52] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. arXiv preprint, 2021. 2 +[53] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 5 +[54] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In ICCV, 2021. 3 +[55] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 2 +[56] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In CVPR, 2021. 2 +[57] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2 +[58] Ji Zhang, Jingkuan Song, Lianli Gao, Ye Liu, and Heng Tao Shen. Progressive meta-learning with curriculum. TCVT, 2022. 2 +[59] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint, 2021. 2 +[60] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In ICLR. 2 +[61] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, 2019. 3 +[62] Zhedong Zheng and Yi Yang. Rectifying pseudo label learning via uncertainty estimation for domain adaptive semantic segmentation. IJCV, 2021. 2 +[63] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint, 2022. 3, 8 +[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 2017. 5 +[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2 +[66] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 3, 8, 12 +[67] Linhai Zhuo, Yuqian Fu, Jingjing Chen, Yixin Cao, and YuGang Jiang. Tgdm: Target guided dynamic mixup for cross-domain few-shot learning. In ACM Multimedia, 2022. 2 \ No newline at end of file diff --git a/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/images.zip b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b01e649f76b89283237d3351bad735acb15f7db5 --- /dev/null +++ b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f2b1b49f7f74a2b80271a37c80bd66f4485d7a75ae22f8ac86d6565cd5beed6 +size 703691 diff --git a/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/layout.json b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..74b90e0311f3e40061ff5319d4efeb4915505e08 --- /dev/null +++ b/2023/StyleAdv_ Meta Style Adversarial Training for Cross-Domain Few-Shot Learning/layout.json @@ -0,0 +1,9450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 545, + 121 + ], + "type": "text", + "content": "StyleAdv: Meta Style Adversarial Training for Cross-Domain Few-Shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "spans": [ + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "text", + "content": "Yuqian Fu" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "text", + "content": ", Yu Xie" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "text", + "content": ", Yanwei Fu" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "text", + "content": ", Yu-Gang Jiang" + }, + { + "bbox": [ + 183, + 142, + 434, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "text", + "content": "Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University " + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "text", + "content": "Purple Mountain Laboratories, Nanjing, China. " + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "text", + "content": "School of Data Science, Fudan University " + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "inline_equation", + "content": "\\{\\text { fuyq20, yxie18, yanweifu, ygj } \\} @" + }, + { + "bbox": [ + 53, + 157, + 561, + 200 + ], + "type": "text", + "content": "fudan.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 252, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 290, + 601 + ], + "type": "text", + "content": "Cross-Domain Few-Shot Learning (CD-FSL) is a recently emerging task that tackles few-shot learning across different domains. It aims at transferring prior knowledge learned on the source dataset to novel target datasets. The CD-FSL task is especially challenged by the huge domain gap between different datasets. Critically, such a domain gap actually comes from the changes of visual styles, and wave-SAN [10] empirically shows that spanning the style distribution of the source data helps alleviate this issue. However, wave-SAN simply swaps styles of two images. Such a vanilla operation makes the generated styles \"real\" and \"easy\", which still fall into the original set of the source styles. Thus, inspired by vanilla adversarial learning, a novel model-agnostic meta Style Adversarial training (StyleAdv) method together with a novel style adversarial attack method is proposed for CD-FSL. Particularly, our style attack method synthesizes both \"virtual\" and \"hard\" adversarial styles for model training. This is achieved by perturbing the original style with the signed style gradients. By continually attacking styles and forcing the model to recognize these challenging adversarial styles, our model is gradually robust to the visual styles, thus boosting the generalization ability for novel target datasets. Besides the typical CNN-based backbone, we also employ our StyleAdv method on large-scale pretrained vision transformer. Extensive experiments conducted on eight various target datasets show the effectiveness of our method. Whether built upon ResNet or ViT, we achieve the new state of the art for CD-FSL. Code is available at https://github.com/lovelyqian/StyleAdv-CDFSL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 624, + 127, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 127, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 127, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 644, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 288, + 693 + ], + "type": "text", + "content": "This paper studies the task of Cross-Domain Few-Shot Learning (CD-FSL) which addresses the Few-Shot Learning (FSL) problem across different domains. As a general recipe for FSL, episode-based meta-learning strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 547, + 335 + ], + "type": "text", + "content": "has also been adopted for training CD-FSL models, e.g., FWT [48], LRP [42], ATA [51], and wave-SAN [10]. Generally, to mimic the low-sample regime in testing stage, meta learning samples episodes for training the model. Each episode contains a small labeled support set and an unlabeled query set. Models learn meta knowledge by predicting the categories of images contained in the query set according to the support set. The learned meta knowledge generalizes the models to novel target classes directly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 338, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 567 + ], + "type": "text", + "content": "Empirically, we find that the changes of visual appearances between source and target data is one of the key causes that leads to the domain gap in CD-FSL. Interestingly, waveSAN [10], our former work, shows that the domain gap issue can be alleviated by augmenting the visual styles of source images. Particularly, wave-SAN proposes to augment the styles, in the form of Adaptive Instance Normalization (AdaIN) [22], by randomly sampling two source episodes and exchanging their styles. However, despite the efficacy of wave-SAN, such a naive style generation method suffers from two limitations: 1) The swap operation makes the styles always be limited in the \"real\" style set of the source dataset; 2) The limited real styles further lead to the generated styles too \"easy\" to learn. Therefore, a natural question is whether we can synthesize \"virtual\" and \"hard\" styles for learning a more robust CD-FSL model? Formally, we use \"real/virtual\" to indicate whether the styles are originally presented in the set of source styles, and define \"easy/hard\" as whether the new styles make meta tasks more difficult." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "To that end, we draw inspiration from the adversarial training, and propose a novel meta Style Adversarial training method (StyleAdv) for CD-FSL. StyleAdv plays the minimax game in two iterative optimization loops of metaintraining. Particularly, the inner loop generates adversarial styles from the original source styles by adding perturbations. The synthesized adversarial styles are supposed to be more challenging for the current model to recognize, thus, increasing the loss. Whilst the outer loop optimizes the whole network by minimizing the losses of recognizing the images with both original and adversarial styles. Our ultimate goal is to enable learning a model that is robust to various styles, be-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 165, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 165, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 165, + 712 + ], + "type": "text", + "content": "* indicates corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "24575" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "yond the relatively limited and simple styles from the source data. This can potentially improve the generalization ability on novel target domains with visual appearance shifts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 289 + ], + "type": "text", + "content": "Formally, we introduce a novel style adversarial attack method to support the inner loop of StyleAdv. Inspired yet different from the previous attack methods [14,34], our style attack method perturbs and synthesizes the styles rather than image pixels or features. Technically, we first extract the style from the input feature map, and include the extracted style in the forward computation chain to obtain its gradient for each training step. After that, we synthesize the new style by adding a certain ratio of gradient to the original style. Styles synthesized by our style adversarial attack method have the good properties of \"hard\" and \"virtual\". Particularly, since we perturb styles in the opposite direction of the training gradients, our generation leads to the \"hard\" styles. Our attack method results in totally \"virtual\" styles that are quite different from the original source styles." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 288, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 288, + 531 + ], + "type": "text", + "content": "Critically, our style attack method makes progressive style synthesizing, with changing style perturbation ratios, which makes it significantly different from vanilla adversarial attacking methods. Specifically, we propose a novel progressive style synthesizing strategy. The naive solution of directly plugging-in perturbations is to attack each block of the feature embedding module individually, which however, may results in large deviations of features from the high-level block. Thus, our strategy is to make the synthesizing signal of the current block be accumulated by adversarial styles from previous blocks. On the other hand, rather than attacking the models by fixing the attacking ratio, we synthesize new styles by randomly sampling the perturbation ratio from a candidate pool. This facilitates the diversity of the synthesized adversarial styles. Experimental results have demonstrated the efficacy of our method: 1) our style adversarial attack method does synthesize more challenging styles, thus, pushing the limits of the source visual distribution; 2) our StyleAdv significantly improves the base model and outperforms all other CD-FSL competitors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 640 + ], + "type": "text", + "content": "We highlight our StyleAdv is model-agnostic and complementary to other existing FSL or CD-FSL models, e.g., GNN [12] and FWT [48]. More importantly, to benefit from the large-scale pretrained models, e.g., DINO [2], we further explore adapting our StyleAdv to improve the Vision Transformer (ViT) [5] backbone in a non-parametric way. Experimentally, we show that StyleAdv not only improves CNN-based FSL/CD-FSL methods, but also improves the large-scale pretrained ViT model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Finally, we summarize our contributions. 1) A novel meta style adversarial training method, termed StyleAdv, is proposed for CD-FSL. By first perturbing the original styles and then forcing the model to learn from such adversarial styles, StyleAdv improves the robustness of CD-FSL models. 2) We present a novel style attack method with the novel" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "progressive synthesizing strategy in changing attacking ratios. Diverse \"virtual\" and \"hard\" styles thus are generated. 3) Our method is complementary to existing FSL and CD-FSL methods; and we validate our idea on both CNN-based and ViT-based backbones. 4) Extensive results on eight unseen target datasets indicate that our StyleAdv outperforms previous CD-FSL methods, building a new SOTA result." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 186, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 547, + 474 + ], + "type": "text", + "content": "Cross-Domain Few-Shot Learning. FSL which aims at freeing the model from reliance on massive labeled data has been studied for many years [12, 25, 39, 41, 43, 45, 46, 56, 58]. Particularly, some recent works, e.g., CLIP [38], CoOp [65], CLIP-Adapter [11], Tip-Adapter [59], and PMF [19] explore promoting the FSL with large-scale pretrained models. Particularly, PMF contributes a simple pipeline and builds a SOTA for FSL. As an extended task from FSL, CD-FSL [1, 7–10, 15, 16, 23, 29, 32, 37, 42, 48, 51, 60, 67] mainly solves the FSL across different domains. Typical meta-learning based CD-FSL methods include FWT [48], LRP [42], ATA [51], AFA [20], and wave-SAN [10]. Specifically, FWT and LRP tackle CD-FSL by refining batch normalization layers and using the explanation model to guide training. ATA, AFA, and wave-SAN propose to augment the image pixels, features, and visual styles, respectively. Several transfer-learning based CD-FSL methods, e.g., BSCD-FSL (also known as Fine-tune) [16], BSR [33], and NSAE [32] have also been explored. These methods reveal that finetuning helps improving the performances on target datasets. Other works that introduce extra data or require multiple domain datasets for training include STARTUP [37], Meta-FDMixup [8], Me-D2N [9], TGDM [67], TriAE [15], and DSL [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": "Adversarial Attack. The adversarial attack aims at misleading models by adding some bespoke perturbations to input data. To generate the perturbations effectively, lots of adversarial attack methods have been proposed [6, 14, 26, 27, 34, 36, 55, 57]. Most of the works [6, 14, 34, 36] attack the image pixels. Specifically, FGSM [14] and PGD [34] are two most classical and famous attack algorithms. Several works [26, 27, 62] attack the feature space. Critically, few methods [57] attack styles. Different from these works that aim to mislead the models, we perturb the styles to tackle the visual shift issue for CD-FSL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Adversarial Few-Shot Learning. Several attempts [13, 28, 30, 40, 52] that explore adversarial learning for FSL have been made. Among them, MDAT [30], AQ [13], and MetaAdv [52] first attack the input image and then train the model using the attacked images to improve the defense ability against adversarial samples. Shen et al. [40] attacks the feature of the episode to improve the generalization capability of FSL models. Note that ATA [51] and AFA [20], two CD-FSL methods, also adopt the adversarial learning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "24576" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "However, we are greatly different from them. ATA and AFA perturb image pixels or features, while we aim at bridging the visual gap by generating diverse hard styles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 289, + 312 + ], + "type": "text", + "content": "Style Augmentation for Domain Shift Problem. Augmenting the style distribution for narrowing the domain shift issue has been explored in domain generation [31, 54, 66], image segmentation [3, 63], person re-ID [61], and CD-FSL [10]. Concretely, MixStyle [66], AdvStyle [63], DSU [31], and wave-SAN [10] synthesize styles without extra parameters via mixing, attacking, sampling from a Gaussian distribution, and swapping. MaxStyle [3] and L2D [54] require additional network modules and complex auxiliary tasks to help generate the new styles. Typically, AdvStyle [63] is the most related work to us. Thus, we highlight the key differences: 1) AdvStyle attacks styles on the image, while we attack styles on multiple feature spaces with a progressive attacking method; 2) AdvStyle uses the same task loss (segmentation) for attacking and optimization; in contrast, we use the classical classification loss to attack the styles, while utilize the task loss (FSL) to optimize the whole network." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 327, + 282, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 282, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 282, + 341 + ], + "type": "text", + "content": "3. StyleAdv: Meta Style Adversarial Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": "Task Formulation. Episode " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = ((S, Q), Y)" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " is randomly sampled as the input of each meta-task, where " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " represents the global class labels of the episode images with respect to " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{tr}" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ". Typically, each meta-task is formulated as an " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": "-shot problem. That is, for each episode " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " classes with " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " labeled images are sampled as the support set " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ", and the same " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " classes with another " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " images are used to constitute the query set " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ". The FSL or CD-FSL models predict the probability " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " that the images in " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " belong to " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": " categories according to " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ". Formally, we have " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "|S| = NK" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "|Q| = NM" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "inline_equation", + "content": "|P| = NM \\times N" + }, + { + "bbox": [ + 46, + 348, + 287, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": "FGSM and PGD Attackers. We briefly summarize the algorithms for FGSM [14] and PGD [34], two most famous attacking methods. Given image " + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": ", FGSM attacks the " + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": " by adding a ratio " + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": " of signed gradients with respect to the " + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": " resulting in the adversarial image " + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "inline_equation", + "content": "x^{adv}" + }, + { + "bbox": [ + 46, + 481, + 287, + 542 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 552, + 287, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 552, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 93, + 552, + 287, + 567 + ], + "type": "interline_equation", + "content": "x ^ {a d v} = x + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} J (\\theta , x, y)\\right), \\tag {1}", + "image_path": "47a5b1a70fd4a5994dd2fa2b354f9f8551a8a395585f080c6314f4d84d7ba43f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "J(\\cdot)" + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "text", + "content": " denote the object function and the learnable parameters of a classification model. PGD can be regarded as a variant of FGSM. Different from the FGSM that only attacks once, PGD attacks the image in an iterative way and sets a random start (abbreviated as RT) for " + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 578, + 287, + 639 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 650, + 287, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 650, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 110, + 650, + 287, + 664 + ], + "type": "interline_equation", + "content": "x _ {0} ^ {a d v} = x + k _ {R T} \\cdot \\mathcal {N} (0, I), \\tag {2}", + "image_path": "c7cf7af1be4a1f3223a13ead88554cf7ca74cc06df8f8212867bc9fce72b9d47.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 677, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 677, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 86, + 677, + 287, + 693 + ], + "type": "interline_equation", + "content": "x _ {t} ^ {a d v} = x _ {t - 1} ^ {a d v} + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} J (\\theta , x, y)\\right), \\tag {3}", + "image_path": "524b30a3df12a0d68f611eb3ebad75f1bf0940956b2d3bc0a95feb98d59a95ce.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "type": "inline_equation", + "content": "k_{RT},\\epsilon" + }, + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "type": "text", + "content": " are hyper-parameters. " + }, + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 47, + 701, + 282, + 713 + ], + "type": "text", + "content": " is Gaussian noises." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 542, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 542, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 542, + 85 + ], + "type": "text", + "content": "3.1. Overview of Meta Style Adversarial Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "text", + "content": "To alleviate the performance degradation caused by the changing visual appearance, we tackle CD-FSL by promoting the robustness of models on recognizing various styles. Thus, we expose our FSL model to some challenging virtual styles beyond the image styles that existed in the source dataset. To that end, we present the novel StyleAdv adversarial training method. Critically, rather than adding perturbations to image pixels, we particularly focus on adversarially perturbing the styles. The overall framework of our StyleAdv is illustrated in Figure 1. Our StyleAdv contains a CNN/ViT backbone " + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "text", + "content": ", a global FC classifier " + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "text", + "content": ", and a FSL classifier " + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "inline_equation", + "content": "f_{fsl}" + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\theta_E, \\theta_{cls}, \\theta_{fsl}" + }, + { + "bbox": [ + 304, + 90, + 547, + 257 + ], + "type": "text", + "content": ", respectively. Besides, our core style attack method, a novel style extraction module, and the AdaIN are also included." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 258, + 547, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 294 + ], + "type": "text", + "content": "Overall, we learn the StyleAdv by solving a minimax game. Specifically, the minimax game shall involve two iterative optimization loops in each meta-train step. Particularly," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 302, + 547, + 382 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 302, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 302, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 302, + 547, + 338 + ], + "type": "text", + "content": "- Inner loop: synthesizing new adversarial styles by attacking the original source styles; the generated styles will increase the loss of the current network." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 347, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 547, + 382 + ], + "type": "text", + "content": "- Outer loop: optimizing the whole network by classifying source images with both original and adversarial styles; this process will decrease the loss." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 391, + 506, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 506, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 506, + 403 + ], + "type": "text", + "content": "3.2. Style Extraction from CNNs and ViTs" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": "Adaptive Instance Normalization (AdaIN). We recap the vanilla AdaIN [22] proposed for CNN in style transfer. Particularly, AdaIN reveals that the instance-level mean and standard deviation (abbreviated as mean and std) convey the style information of the input image. Denoting the mean and std as " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": ", AdaIN (denoted as " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": ") reveals that the style of " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": " can be transferred to that of " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "F_{tgt}" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": " by replacing the original style " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "(\\mu, \\sigma)" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": " with the target style " + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "inline_equation", + "content": "(\\mu_{tgt}, \\sigma_{tgt})" + }, + { + "bbox": [ + 304, + 409, + 547, + 506 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 514, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 514, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 339, + 514, + 545, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (F, \\mu_ {t g t}, \\sigma_ {t g t}) = \\sigma_ {t g t} \\frac {F - \\mu (F)}{\\sigma (F)} + \\mu_ {t g t}. \\tag {4}", + "image_path": "3ac53fc0377cef9d60d42570f9f691f1d6e2e032e89732e09b3a0bf33a53f8b5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": "Style Extraction for CNN Features. As shown in the upper part of Figure 1 (b), let " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "F \\in \\mathcal{R}^{B \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": " indicates the input feature batch, where " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "B, C, H," + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": " denote the batch size, channel, height, and width of the feature " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": ", respectively. As in AdaIN, the mean " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": " and std " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 545, + 547, + 605 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 354, + 613, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 613, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 354, + 613, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mu (\\mathrm {F}) _ {\\mathrm {b}, \\mathrm {c}} = \\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} F _ {b, c, h, w}, \\tag {5}", + "image_path": "ac4f28032614cdf7807d9ab3844ef076c3a49cc56896944b5842aa6ff8c5c181.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 657, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\sigma (\\mathrm {F}) _ {\\mathrm {b}, \\mathrm {c}} = \\sqrt {\\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left(F _ {b , c , h , w} - \\mu_ {\\mathrm {b , c}} (F)\\right) ^ {2} + \\epsilon}, \\tag {6}", + "image_path": "94e7b575a294153cabf789927846f46611fe5a54819b391c7ccf7310435afb69.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 700, + 394, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 394, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 394, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 700, + 394, + 713 + ], + "type": "inline_equation", + "content": "\\mu ,\\sigma \\in \\mathcal{R}^{B\\times C}" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "24577" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 354, + 293 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 354, + 293 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 354, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 354, + 293 + ], + "type": "image", + "image_path": "7c14f659c72ee51e20f7668e69c4d9eb75f8585782999416c1fc704fb586f91c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 297, + 294, + 309 + ], + "lines": [ + { + "bbox": [ + 113, + 297, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 113, + 297, + 294, + 309 + ], + "type": "text", + "content": "(a) Overall Framework of Our StyleAdv Method" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 359, + 69, + 543, + 293 + ], + "blocks": [ + { + "bbox": [ + 359, + 69, + 543, + 293 + ], + "lines": [ + { + "bbox": [ + 359, + 69, + 543, + 293 + ], + "spans": [ + { + "bbox": [ + 359, + 69, + 543, + 293 + ], + "type": "image", + "image_path": "454d49a5abffa07b993dd1a1099de2eb23c3ae59eab586f9c57962ca8559b7e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 297, + 527, + 308 + ], + "lines": [ + { + "bbox": [ + 365, + 297, + 527, + 308 + ], + "spans": [ + { + "bbox": [ + 365, + 297, + 527, + 308 + ], + "type": "text", + "content": "(b) Style Extraction For CNN/ViT Features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 316, + 546, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 546, + 338 + ], + "type": "text", + "content": "Figure 1. (a): Overview of StyleAdv method. The inner loop synthesizes adversarial styles, while the outer loop optimizes the whole network. (b): Style extraction for CNN-based and ViT-based features (illustration with " + }, + { + "bbox": [ + 46, + 316, + 546, + 338 + ], + "type": "inline_equation", + "content": "B = 1" + }, + { + "bbox": [ + 46, + 316, + 546, + 338 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": "Meta Information Extraction for ViT Features. We explore extracting the meta information of the ViT features as the manner of CNN. Intuitively, such meta information can be regarded as a unique \"style\" of ViTs. As shown in Figure 1 (b), we take an input batch data with image split into " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": " patches as an example. The ViT encoder will encode the batch patches into a class (cls) token (" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F_{cls} \\in \\mathcal{R}^{B \\times C}" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": ") and a patch tokens (" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F_0 \\in \\mathcal{R}^{B \\times P^2 \\times C}" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": "). To be compatible with AdaIN, we reshape the " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F \\in \\mathcal{R}^{B \\times C \\times P \\times P}" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": ". At this point, we can calculate the meta information for patch tokens " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": " as in Eq. 5 and Eq. 6. Essentially, note that the transformer integrates the positional embedding into the patch representation, the spatial relations thus could be considered still hold in the patch tokens. This supports us to reform the patch tokens " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": " as a spatial feature map " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": ". To some extent, this can also be achieved by applying the convolution on the input data via a kernel of size " + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 46, + 349, + 289, + 566 + ], + "type": "text", + "content": " (as indicated by dashed arrows in Figure 1 (b))." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 286, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 286, + 586 + ], + "type": "text", + "content": "3.3. Inner Loop: Style Adversarial Attack Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "We propose a novel style adversarial attack method - Fast Style Gradient Sign Method (Style-FGSM) to accomplish the inner loop. As shown in Figure 1, given an input source episode " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{T},Y)" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", we first forward it into the backbone " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and the FC classifier " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " producing the global classification loss " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " (as illustrated in the ① paths). During this process, a key step is to make the gradient of the style available. To achieve that, let " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_{\\mathcal{T}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " denotes the features of " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", we obtain the style " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mu ,\\sigma)" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_{\\mathcal{T}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " as in Sec. 3.2. After that, we reform the original episode feature as " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(F_{\\mathcal{T}},\\mu ,\\sigma)" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". And the reformed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "type": "text", + "content": "feature is actually used for the forward propagation. In this way, we include " + }, + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 349, + 546, + 384 + ], + "type": "text", + "content": " in our forward computation chain; and thus, we could access the gradients of them." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "content": "With the gradients in ② paths, we then attack " + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "content": " as FGSM does - adding a small ratio " + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "content": " of the signed gradients with respect to " + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 384, + 545, + 422 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 430, + 546, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 430, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 430, + 546, + 444 + ], + "type": "interline_equation", + "content": "\\mu^ {a d v} = \\mu + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\mu} J \\left(\\theta_ {E}, \\theta_ {f _ {c l s}}, \\mathcal {A} \\left(F _ {\\tau}, \\mu , \\sigma\\right), Y\\right)\\right), \\tag {7}", + "image_path": "1c562bbafe59bb611ee5c6b58f76682b97c458871f08abcb2d25728b86e4701a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 453, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 453, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 310, + 453, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\sigma^ {a d v} = \\sigma + \\epsilon \\cdot \\operatorname {s i g n} \\left(\\nabla_ {\\sigma} J \\left(\\theta_ {E}, \\theta_ {f _ {c l s}}, \\mathcal {A} \\left(F _ {\\mathcal {T}}, \\mu , \\sigma\\right), Y\\right)\\right), \\tag {8}", + "image_path": "296b7d599040c8d7fce2f2ed5296576afe6d6515a843ba05f0efaf6bb478171a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "J(\\cdot)" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": " is the cross-entropy loss between classification predictions and ground truth " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": ". Inspired by the random start of PGD, we also add random noises " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "k_{RT} \\cdot \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": " before attacking. " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": " refers to Gaussian noises and " + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "inline_equation", + "content": "k_{RT}" + }, + { + "bbox": [ + 304, + 474, + 547, + 545 + ], + "type": "text", + "content": " is a hyper-parameter. Our Style-FGSM enables us to generate both \"virtual\" and \"hard\" styles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": "Progressive Style Synthesizing Strategy: To prevent the high-level adversarial feature from deviating, we propose to apply our style-FGSM in a progressive strategy. Concretely, the embedding module " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": " has three blocks " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{3}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", with the corresponding features " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{2}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{3}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ". For the first block, we use " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\mu_{1}, \\sigma_{1})" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": " to denote the original styles of " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ". The adversarial styles " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\mu_{1}^{adv}, \\sigma_{1}^{adv})" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": " are obtained directly as in Eq. 7 and Eq. 8. For subsequent blocks, the attack signals on the current block " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": " are those accumulated from the block 1 to block " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "i - 1" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ". Take the second block as an example, the block feature " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{2}" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": " is not simply extracted by " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{2}(F_{1})" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ". Instead, we have " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{2}' = E_{2}(F_{1}^{adv})" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{1}^{adv} = \\mathcal{A}(F_{1}, \\mu_{1}^{adv}, \\sigma_{1}^{adv})" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ". Attacking on " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "F_{2}'" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": " results in the adversarial styles " + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "inline_equation", + "content": "(\\mu_{2}^{adv}, \\sigma_{2}^{adv})" + }, + { + "bbox": [ + 304, + 546, + 548, + 714 + ], + "type": "text", + "content": ". Accordingly, we gen" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "24578" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": "erate " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "(\\mu_3^{adv},\\sigma_3^{adv})" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " for the last block. The illustration of the progressive attacking strategy is attached in the Appendix. Changing Style Perturbation Ratios: Different from the vanilla FGSM [14] or PGD [34], our style attacking algorithm is expected to synthesize new styles with diversity. Thus, instead of using a fixed attacking ratio " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": ", we randomly sample " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " from a candidate list " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "\\epsilon_{list}" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " as the current attacking ratio. Despite the randomness of " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": ", we still synthesize styles in a more challenging direction, " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " only affects the extent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 187, + 282, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 187, + 282, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 187, + 282, + 201 + ], + "type": "text", + "content": "3.4. Outer Loop: Optimize the StyleAdv Network" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": "For each meta-train iteration with clean episode " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " as input, our inner loop produces adversarial styles " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "(\\mu_1^{adv},\\sigma_1^{adv})" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "(\\mu_{2}^{adv},\\sigma_{2}^{adv})" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "(\\mu_3^{adv},\\sigma_3^{adv})" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " . As in Figure 1, the goal of the outer loop is to optimize the whole StyleAdv with both the clean feature " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " and the style attacked feature " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "F^{adv}" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " utilized as the training data. Typically, the clean episode feature " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " can be obtained directly as " + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "inline_equation", + "content": "E(T)" + }, + { + "bbox": [ + 46, + 205, + 287, + 289 + ], + "type": "text", + "content": " as in 3 paths." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": "In 4 paths, we obtain the " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F^{adv}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": " by transferring the original style of " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": " to the corresponding adversarial attacked styles. Similar with the progressive style-FGSM, we have " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F_{1}^{adv} = \\mathcal{A}(E_{1}(\\mathcal{T}),\\mu_{1}^{adv},\\sigma_{1}^{adv})" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F_{2}^{adv} = \\mathcal{A}(E_{2}(F_{1}^{adv}),\\mu_{2}^{adv},\\sigma_{2}^{adv})" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F_{3}^{adv} = \\mathcal{A}(E_{3}(F_{2}^{adv}),\\mu_{3}^{adv},\\sigma_{3}^{adv})" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F^{adv}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": " is obtained by applying an average pooling layer to " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F_{3}^{adv}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ". A skip probability " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "p_{skip}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": " is set to decide whether to skip the current attacking. Conducting FSL tasks for both the clean feature " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": " and style attacked feature " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "F^{adv}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": " results in two FSL predictions " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "P_{fsl}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "P_{fsl}^{adv}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ", and two FSL classification losses " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fsl}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fsl}^{adv}" + }, + { + "bbox": [ + 46, + 289, + 288, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": "Further, despite the styles of " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "F^{adv}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": " shifts from that of " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": ", we encourage that the semantic content should be still consistent as in wave-SAN [10]. Thus we add a consistent constraint to the predictions of " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "P_{fsl}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "P_{fsl}^{adv}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": " resulting in the consistent loss " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cons}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 490, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 490, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 113, + 490, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s}} = \\mathrm {K L} \\left(P _ {f s l}, P _ {f s l} ^ {\\text {a d v}}\\right), \\tag {9}", + "image_path": "4ce0b82eed34a20409f5b81c58ea581101df553cf64343ead84e3fd3f79cb62c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}()" + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "text", + "content": " is Kullback-Leibler divergence loss. In addition, we have the global classification loss " + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "text", + "content": ". This ensures that " + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\theta_{cls}" + }, + { + "bbox": [ + 46, + 513, + 287, + 561 + ], + "type": "text", + "content": " is optimized to provide correct gradients for style-FGSM. The final meta-objective of StyleAdv is as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 569, + 287, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 569, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 96, + 569, + 287, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {f s l} + \\mathcal {L} _ {f s l} ^ {a d v} + \\mathcal {L} _ {c o n s} + \\mathcal {L} _ {c l s}. \\tag {10}", + "image_path": "f5bd1db96433286288b679b294db7b2d53be4bbbfa4b22673b7552214853ca93.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 592, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 616 + ], + "type": "text", + "content": "Note that our StyleAdv is model-agnostic and orthogonal to existing FSL and CD-FSL methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 157, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 157, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 157, + 635 + ], + "type": "text", + "content": "3.5. Network Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Applying StyleAdv Directly for Inference. Our StyleAdv facilitates making CD-FSL model more robust to style shifts. Once the model is meta-trained, we can employ it for inference directly by feeding the testing episode into the " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": ". The class with the highest probability will be taken as the predicted result." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "Finetuning StyleAdv Using Target Examples. As indicated in previous works [16, 32, 33, 51], finetuning CD-FSL models on target examples helps improve the model performance. Thus, to further promote the performance of StyleAdv, we also equip it with the fintuning strategy forming an upgraded version (\"StyleAdv-FT\"). Specifically, as in ATA-FT [51], for each novel testing episode, we augment the novel support set to form pseudo episodes as training data for tuning the meta-trained model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 197, + 388, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 197, + 388, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 197, + 388, + 210 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "text", + "content": "Datasets. We take two CD-FSL benchmarks proposed in BSCD-FSL [16] and FWT [48]. Both of them take mini-Imagenet [39] as the source dataset. Two disjoint sets split from mini-Imagenet form " + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{tr}" + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{eval}" + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "text", + "content": ". Totally eight datasets including ChestX [53], ISIC [4, 47], EuroSAT [18], CropDisease [35], CUB [50], Cars [24], Places [64], and Plantae [49] are taken as novel target datasets. The former four datasets included in BSCD-FSL's benchmark cover medical images varying from X-ray to dermoscopic skin lesions, and natural images from satellite pictures to plant disease photos. While the latter four datasets that focus on more fine-grained concepts such as birds and cars are contained in FWT. These eight target datasets serve as testing set " + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{te}" + }, + { + "bbox": [ + 304, + 219, + 547, + 386 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 388, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 496 + ], + "type": "text", + "content": "Network Modules. For typical CNN based network, following previous CD-FSL methods [10, 42, 48, 51], ResNet-10 [17] is selected as the embedding module while GNN [12] is selected as the FSL classifier; For the emerging ViT based network, following PMF [19], we use the ViT-small [5] and the ProtoNet [41] as the embedding module and the FSL classifier, respectively. Note that, the ViT-small is pretrained on ImageNet1K by DINO [2] as in PMF. The " + }, + { + "bbox": [ + 304, + 388, + 547, + 496 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 304, + 388, + 547, + 496 + ], + "type": "text", + "content": " is built by a fully connected layer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": "Implementation Details. The 5-way 1-shot and 5-way 5-shot settings are conducted. Taking ResNet10 as backbone, we meta train the network for 200 epochs, each epoch contains 120 meta tasks. Adam with a learning rate of 0.001 is utilized as the optimizer. Taking ViT-small as backbone, the meta train stage takes 20 epoch, each epoch contains 2000 meta tasks. The SGD with a initial learning rate of 5e-5 and 0.001 are used for optimize the " + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": ", respectively. The " + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{list}" + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "inline_equation", + "content": "k_{RT}" + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": " of Style-FGSM attacker are set as [0.8, 0.08, 0.008], " + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\frac{16}{255}" + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": ". The probability " + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "inline_equation", + "content": "p_{skip}" + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": " of random skipping the attacking is chosen from {0.2, 0.4}. We evaluate our network with 1000 randomly sampled episodes and report average accuracy (\\%) with a 95% confidence interval. Both the results of our \"StyleAdv\" and \"StyleAdv-FT\" are reported. The details of the finetuning are attached in Appendix. ResNet-10 based models are trained and tested on a single GeForce GTX 1080, while ViT-small based models require a single NVIDIA GeForce RTX 3090." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "24579" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 547, + 353 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 353 + ], + "type": "table", + "html": "
1-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--22.00±0.4632.02±0.6663.69±1.0364.48±1.0845.69±0.6831.79±0.5153.10±0.8035.60±0.5643.55
FWT [48]RN10--22.04±0.4431.58±0.6762.36±1.0566.36±1.0447.47±0.7531.61±0.5355.77±0.7935.95±0.5844.14
LRP [42]RN10--22.11±0.2030.94±0.3054.99±0.5059.23±0.5048.29±0.5132.78±0.3954.83±0.5637.49±0.4342.58
ATA [51]RN10--22.10±0.2033.21±0.4061.35±0.5067.47±0.5045.00±0.5033.61±0.4053.57±0.5034.42±0.4043.84
AFA [20]RN10--22.92±0.2033.21±0.3063.12±0.5067.61±0.5046.86±0.5034.25±0.4054.04±0.6036.76±0.4044.85
wave-SAN [10]RN10--22.93±0.4933.35±0.7169.64±1.0970.80±1.0650.25±0.7433.55±0.6157.75±0.8240.71±0.6647.37
StyleAdv (ours)RN10--22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
ATA-FT [51]RN10Y-22.15±0.2034.94±0.4068.62±0.5075.41±0.5046.23±0.5037.15±0.4054.18±0.5037.38±0.4047.01
StyleAdv-FT (ours)RN10Y-22.64±0.3535.76±0.5272.92±0.7580.69±0.2848.49±0.7235.09±0.5558.58±0.8341.13±0.6749.41
PMF* [19]ViT-smallYDINO/IN1K21.73±0.3030.36±0.3670.74±0.6380.79±0.6278.13±0.6637.24±0.5771.11±0.7153.60±0.6655.46
StyleAdv (ours)ViT-small-DINO/IN1K22.92±0.3233.05±0.4472.15±0.6581.22±0.6184.01±0.5840.48±0.5772.64±0.6755.52±0.6657.75
StyleAdv-FT (ours)ViT-smallYDINO/IN1K22.92±0.3233.99±0.4674.93±0.5884.11±0.5784.01±0.5840.48±0.5772.64±0.6755.52±0.6658.57
5-shotBackboneFTLargePChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
GNN [12]RN10--25.27±0.4643.94±0.6783.64±0.7787.96±0.6762.25±0.6544.28±0.6370.84±0.6552.53±0.5958.84
FWT [48]RN10--25.18±0.4543.17±0.7083.01±0.7987.11±0.6766.98±0.6844.90±0.6473.94±0.6753.85±0.6259.77
LRP [42]RN10--24.53±0.3044.14±0.4077.14±0.4086.15±0.4064.44±0.4846.20±0.4674.45±0.4754.46±0.4658.94
ATA [51]RN10--24.32±0.4044.91±0.4083.75±0.4090.59±0.3066.22±0.5049.14±0.4075.48±0.4052.69±0.4060.89
AFA [20]RN10--25.02±0.2046.01±0.4085.58±0.4088.06±0.3068.25±0.5049.28±0.5076.21±0.5054.26±0.4061.58
wave-SAN [10]RN10--25.63±0.4944.93±0.6785.22±0.7189.70±0.6470.31±0.6746.11±0.6676.88±0.6357.72±0.6462.06
StyleAdv (ours)RN10--26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
Fine-tune [16]RN10Y-25.97±0.4148.11±0.6479.08±0.6189.25±0.5164.14±0.7752.08±0.7470.06±0.7459.27±0.7061.00
ATA-FT [51]RN10Y-25.08±0.2049.79±0.4089.64±0.3095.44±0.2069.83±0.5054.28±0.5076.64±0.4058.08±0.4064.85
NSAE [32]RN10Y-27.10±0.4454.05±0.6383.96±0.5793.14±0.4768.51±0.7654.91±0.7471.02±0.7259.55±0.7464.03
BSR [33]RN10Y-26.84±0.4454.42±0.6680.89±0.6192.17±0.4569.38±0.7657.49±0.7271.09±0.6861.07±0.7664.17
StyleAdv-FT (ours)RN10Y-26.24±0.3553.05±0.5491.64±0.4396.51±0.2870.90±0.6356.44±0.6879.35±0.6164.10±0.6467.28
PMF [19]ViT-smallYDINO/IN1K27.2750.1285.9892.96-----
StyleAdv (ours)ViT-small-DINO/IN1K26.97±0.3347.73±0.4488.57±0.3494.85±0.3195.82±0.2761.73±0.6288.33±0.4075.55±0.5472.44
StyleAdv-FT (ours)ViT-smallYDINO/IN1K26.97±0.3351.23±0.5190.12±0.3395.99±0.2795.82±0.2766.02±0.6488.33±0.4078.01±0.5474.06
", + "image_path": "1a24a42a3b4105cb2081cae3c81e264932a8c6f108cf5b4dda2f2fba6e8cfa28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 354, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 547, + 389 + ], + "type": "text", + "content": "Table 1. Results of 5-way 1-shot/5-shot tasks. \"FT\" means whether the finetuning stage is employed. \"LargeP\" represents if large pretrained models are used for model initialization. \"RN10\" is short for \"ResNet-10\". * denotes results are reported by us. Results perform best are bolded. Whether based on ResNet-10 or ViT-small, our method outperforms other competitors significantly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 394, + 205, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 205, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 205, + 407 + ], + "type": "text", + "content": "4.1. Comparison with the SOTAs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 289, + 653 + ], + "type": "text", + "content": "We compare our StyleAdv/StyleAdv-FT against several most representative and competitive CD-FSL methods. Concretly, with the ResNet-10 (abbreviated as RN10) as backbone, totally nine methods including GNN [12], FWT [48], LRP [42], ATA [51], AFA [20], wave-SAN [10], Finetune [16], NSAE [32], and BSR [33] are introduced as our competitors. Among them, the former six competitors are meta-learning based method that used for inference directly, thus we compare our \"StyleAdv\" against them for a fair comparison. Typically, the GNN [12] works as a base model. The Fine-tune [16], NSAE [32], BSR [33], and ATA-FT [51] (formed by finetuning ATA) all require finetuning model during inference, thus our \"StyleAdv-FT\" is used. With the ViT as backbone, the most recent and competitive PMF (SOTA method for FSL) is compared. For fair comparisons, we follow the same pipeline proposed in PMF [19]. Note that we promote CD-FSL models with only one single source domain. Those methods that use extra training datasets, e.g., STARTUP [37], meta-FDMixup [8], and DSL [21] are not considered. The comparison results are given in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "For all results, our method outperforms all the listed CD-FSL competitors significantly and builds a new state of the art. Our StyleAdv-FT (ViT-small) on average achieves " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "58.57\\%" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "inline_equation", + "content": "74.06\\%" + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": " on 5-way 1-shot and 5-shot, respectively. Our StyleAdv (RN10) and StyleAdv-FT (RN10) also" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "type": "text", + "content": "beats all the meta-learning based or transfer-learning (finetuning) based methods. Besides of the state-of-the-art accuracy, we also have other worth-mentioning observations. 1) We show that our StyleAdv method is a general solution for both CNN-based models and ViT-based models. Typically, based on ResNet10, our StyleAdv and StyleAdv-FT improve the base GNN by up to " + }, + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "type": "inline_equation", + "content": "4.93\\%" + }, + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "type": "inline_equation", + "content": "8.44\\%" + }, + { + "bbox": [ + 304, + 395, + 548, + 707 + ], + "type": "text", + "content": " on 5-shot setting. Based on ViT-small, at most cases, our StyleAdv-FT outperforms the PMF by a clear margin. More results of building StyleAdv upon other FSL or CD-FSL methods can be found in the Appendix. 2) Comparing FWT, LRP, ATA, AFA, waveSAN, and our StyleAdv, we find that StyleAdv performs best, followed by wave-SAN, then comes the AFA, ATA, FWT, and LRP. This phenomenon indicates that tackling CD-FSL by solving the visual shift problem is indeed more effective than other perspectives, e.g., adversarial training by perturbing the image features (AFA) or image pixels (ATA), transforms the normalization layers in FWT, and explanation guided training in LRP. 3) For the comparison between StyleAdv and wave-SAN that both tackles the visual styles, we notice that StyleAdv outperforms the wave-SAN in most cases. This demonstrates that the styles generated by our StyleAdv are more conducive to learning robust CD-FSL models than the style augmentation method proposed in wave-SAN. This justifies our idea of synthesizing more challenging (\"hard and virtual\") styles." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "24580" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "4) Overall, the large-scale pretrained model promotes the CD-FSL obviously. Take 1-shot as an example, StyleAdv-FT (ViT-small) boosts the StyleAdv-FT (RN10) by " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "9.16\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " on average. However, we show that the performance improvement varies greatly on different target domains. Generally, for target datasets with relative small domain gap, e.g., CUB and Plantae, models benefit a lot; otherwise, the improvement is limited. 5) We also find that under the cross-domain scenarios, finetuning model on target domain, e.g., NSAE, BSR do show an advantage over purely meta-learning based methods, e.g., FWT, LRP, and wave-SAN. However, to finetune model using extremely few examples, e.g., 5-way 1-shot is much harder than on relatively larger shots. This may explain why those finetune-based methods do not conduct experiments on 1-shot setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": "Effectiveness of Style-FGSM Attacker. To show the advantages of our progressive style synthesizing strategy and attacking with changing perturbation ratios, we compare our Style-FGSM against several variants and report the results in Figure 2. Specifically, for Figure 2 (a), we compare our style-FGSM against the variant that attacks the blocks individually. Results show that attacking in a progressive way exceeds the naive individual strategy in most cases. For Figure 2 (b), to demonstrate how the performance will be affected by fixed attacking ratios, we also conduct experiments with different " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon_{list}" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": ". Since we set the " + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon_{list}" + }, + { + "bbox": [ + 46, + 251, + 290, + 443 + ], + "type": "text", + "content": " as [0.8, 0.08, 0.008], three different choices including [0.8], [0.08], and [0.008] are selected. From the results, we first notice that the best result can be reached by a single fixed ratio. However, sampling the attacking ratio from a pool of candidates achieves the best result in most cases." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 456, + 267, + 684 + ], + "blocks": [ + { + "bbox": [ + 71, + 456, + 267, + 684 + ], + "lines": [ + { + "bbox": [ + 71, + 456, + 267, + 684 + ], + "spans": [ + { + "bbox": [ + 71, + 456, + 267, + 684 + ], + "type": "image", + "image_path": "cddee983b9fc92f8c19d546c7064fd9cc0641f9698878d69e9b84345c247e2d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "type": "text", + "content": "Figure 2. Effectiveness of the progressive style synthesizing strategy and the changing style perturbation ratios. The 5-way 1-shot results are reported. Models are built on ResNet10 and GNN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 519, + 205 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 519, + 205 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 519, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 519, + 205 + ], + "type": "image", + "image_path": "51da6b7adbc4530030ff63f7ce1b527353b7445d2c4f2ddcfb218fe7407a2093.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 210, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 547, + 232 + ], + "type": "text", + "content": "Figure 3. Visualization of wave-SAN and StyleAdv. (a): synthesized images; (b): meta-training losses; (c): T-SNE results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 236, + 397, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 397, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 397, + 248 + ], + "type": "text", + "content": "4.2. More Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 256, + 548, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 548, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 548, + 640 + ], + "type": "text", + "content": "Visualization of Hard Style Generation. To help understand the \"hard\" style generation of our method intuitively, as in Figure 3, we make several visualizations comparing StyleAdv against the wave-SAN. 1) As in Figure 3 (a), we show the stylized images generated by wave-SAN and our StyleAdv. The visualization is achieved by applying the style augmentation methods to input images. Specifically, for wave-SAN, the style is swapped with another randomly sampled source image; for StyleAdv, the results of attacking style with " + }, + { + "bbox": [ + 304, + 256, + 548, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.08" + }, + { + "bbox": [ + 304, + 256, + 548, + 640 + ], + "type": "text", + "content": " are given. We observe that wave-SAN tends to exchange the global visual appearance, e.g., the color of the input image randomly. By contrast, StyleAdv prefers to disturb the important regions that are key to recognizing the image category. For example, the fur of the cat and the key parts (face and feet) of the dogs. These observations intuitively support our claim that our StyleAdv synthesize more harder styles than wave-SAN. 2) To quantitatively evaluate whether our StyleAdv introduces more challenging styles into the training stage, as in Figure 3 (b), we visualize the meta-training loss. Results reveal that the perturbed losses of wave-SAN oscillate around the original loss, while StyleAdv increases the original loss obviously. These phenomenons further validate that we perturb data towards a more difficult direction thus pushing the limits of style generation to a large extent. 3) To further show the advantages of StyleAdv over wave-SAN, as shown in Figure 3 (c), we visualize the high-level features extracted by the meta-trained wave-SAN and StyleAdv. Five classes (denoted by different colors) of mini-Imagenet are selected. T-SNE is used for reducing the feature dimensions. Results demonstrate that StyleAdv enlarges the inter-class distances making classes more distinguishable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Why Attack Styles Instead of Images or Features? A natural question may be why we choose to attack styles instead of other targets, e.g., the input image as in AQ [13], MDAT [30], and ATA [51] or the features as in Shen et al. [40] and AFA [20]? To answer this question, we compare our StyleAdv which attacks styles against attacking images" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "24581" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 547, + 160 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 547, + 160 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 547, + 160 + ], + "type": "table", + "html": "
Attack TargetChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotImage22.71±0.3533.00±0.5367.00±0.8272.65±0.7548.15±0.7234.40±0.6057.89±0.8339.85±0.6446.96
Feature22.55±0.3532.95±0.5368.71±0.8170.86±0.7846.52±0.7034.07±0.5456.68±0.8139.62±0.6246.50
Style (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotImage24.92±0.3642.63±0.4784.18±0.5490.31±0.4766.37±0.6547.46±0.6775.94±0.6257.33±0.6561.14
Feature25.55±0.3743.71±0.5084.22±0.5591.71±0.4467.31±0.6750.26±0.6776.46±0.6557.39±0.6362.08
Style (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
", + "image_path": "43090eb3851e9b8ca4351b05486360e19a6adebccc1138eec84ddc08271753f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 176, + 545, + 300 + ], + "blocks": [ + { + "bbox": [ + 77, + 161, + 516, + 171 + ], + "lines": [ + { + "bbox": [ + 77, + 161, + 516, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 161, + 516, + 171 + ], + "type": "text", + "content": "Table 2. Comparison results (%) of attacking image, feature, and styles. Models build upon ResNet10 and GNN classifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 176, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 48, + 176, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 545, + 300 + ], + "type": "table", + "html": "
Augment MethodChestXISICEuroSATCropDiseaseCUBCarsPlacesPlantaeAverage
1-shotStyleGaus†22.37±0.3531.48±0.5265.71±0.8269.25±0.8046.32±0.7232.69±0.5455.48±0.7937.27±0.6145.07
MixStyle [66]22.43±0.3533.21±0.5367.35±0.8068.80±0.8247.08±0.7333.39±0.5856.12±0.7838.03±0.6245.80
AdvStyle [63]22.04±0.3630.83±0.5265.19±0.8264.96±0.8147.43±0.7231.90±0.5253.95±0.7935.81±0.5944.01
DSU [31]22.35±0.3631.43±0.5164.55±0.8364.73±0.8147.74±0.7231.61±0.5354.81±0.8137.19±0.6144.30
Style-FGSM (ours)22.64±0.3533.96±0.5770.94±0.8274.13±0.7848.49±0.7234.64±0.5758.58±0.8341.13±0.6748.06
5-shotStyleGaus†24.97±0.3741.74±0.4881.88±0.6189.71±0.4965.98±0.6745.03±0.6472.66±0.6856.66±0.6559.83
MixStyle [66]25.04±0.3643.77±0.5382.67±0.5888.90±0.5265.73±0.6645.91±0.6375.90±0.6356.59±0.6260.56
AdvStyle [63]25.03±0.3543.15±0.5083.09±0.5788.44±0.5266.42±0.6744.85±0.6474.14±0.6554.89±0.6460.00
DSU [31]25.02±0.3645.19±0.5280.30±0.6386.30±0.5667.94±0.6645.65±0.6375.17±0.6454.31±0.6259.99
Style-FGSM (ours)26.07±0.3745.77±0.5186.58±0.5493.65±0.3968.72±0.6750.13±0.6877.73±0.6261.52±0.6863.77
", + "image_path": "746e4dcbd8c969ffeaf283d8f714568384a3d4de4880ed9a7cf5b84ad1201e75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 334 + ], + "type": "text", + "content": "Table 3. Different style augmentation methods are compared. \"StyleGaus†\" means adding random Gaussian noises to the styles, where † represents it is proposed by us. \"MixStyle [66]\", \"AdvStyle [63]\" and \"DSU [31]\" are adapted from other tasks, e.g., domain generation. Results (%) conducted under 5-way 1-shot/5-shot settings. Methods are built upon the ResNet10 and GNN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 340, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 289, + 483 + ], + "type": "text", + "content": "and features by modifying the attack targets of our method. The 5-way 1-shot/5-shot results are given in Table 2. We highlight several points. 1) We notice that attacking image, feature, and style all improve the base GNN model (given in Table 1) which shows that all of them boost the generalization ability of the model by adversarial attacks. Interestingly, the results of our \"Attack Image\"/\"Attack Feature\" even outperform the well-designed CD-FSL methods ATA [51] and AFA [20] (shown in Table 1); 2) Our method has clear advantages over attacking images and features. This again indicates the superiority of tackling visual styles for narrowing the domain gap issue for CD-FSL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "content": "Is Style-FGSM Better than Other Style Augmentation Methods? To show the advantages of our Style-FGSM against other style augmentation methods, we introduce several competitors including \"StyleGaus\", MixStyle [66], AdvStyle [63], and DSU [31]. Typically, \"StyleGaus\" that adds random Gaussian noises into the styles is introduced as a simple but reasonable baseline. MixStyle [66], AdvStyle [63], and DSU [31] which are initially designed for other tasks, e.g., segmentation and domain generation are also adapted. The results are reported in Table 3. Comparing the results of StyleGuas with that reported in Table 1, we find that perturbing the styles on the feature level by simply adding random noises also improves the base GNN and even surpasses a few CD-FSL competitors on some target datasets. This phenomenon is consistent with the insight that augmenting the style distributions helps boost the CD-FSL methods. As for the comparison between our Style-FGSM and other advanced style augmentation competitors, we find that Style-FGSM performs better than all the MixStyle, AdvStyle, and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 340, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 436 + ], + "type": "text", + "content": "DSU on both 1-shot and 5-shot settings. Typically, MixStyle and DSU both generate virtual styles, but their new styles are still relatively easy. This shows that our hard styles boost the model to a larger extent. AdvStyle generates both virtual and hard (adversarial) styles. However, it is still inferior to us. This indicates the advantages of our method that attacks in latent feature space and adopts two individual tasks for attacking and optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 451, + 378, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 451, + 378, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 451, + 378, + 463 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 675 + ], + "type": "text", + "content": "This paper presents a novel model-agnostic StyleAdv for CD-FSL. Critically, to narrow the domain gap which is typically in the form of visual shifts, StyleAdv solves the minimax game of style adversarial learning: first adds perturbations to the source styles increasing the loss of the current model, then optimizes the model by forcing it to recognize both the clean and style perturbed data. Besides, a novel progressive style adversarial attack method termed style-FGSM is presented by us. Style-FGSM synthesizes diverse \"hard\" and \"virtual\" styles via adding the signed gradients to original clean styles. These generated styles support the max step of StyleAdv. Intuitively, by exposing the CD-FSL to adversarial styles which are more challenging than those limited real styles that exist in the source dataset, the generalization ability of the model is boosted. Our StyleAdv improves both CNN-based and ViT-based models. Extensive experiments indicate that our StyleAdv build new SOTAs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement. This project was supported by National Key R&D Program of China (No. 2021ZD0112804) and NSFC under Grant No. 62076067." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "24582" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] John Cai, Bill Cai, and Shen Sheng Mei. Damsl: Domain agnostic meta score-based learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 288, + 158 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 204 + ], + "type": "text", + "content": "[3] Chen Chen, Zeju Li, Cheng Ouyang, Matt Sinclair, Wenjia Bai, and Daniel Rueckert. Maxstyle: Adversarial style composition for robust medical image segmentation. arXiv preprint, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 270 + ], + "type": "text", + "content": "[4] Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 271, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 288, + 336 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint, 2020. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 338, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 371 + ], + "type": "text", + "content": "[6] Ranjie Duan, Xingjun Ma, Yisen Wang, James Bailey, A Kai Qin, and Yun Yang. Adversarial camouflage: Hiding physical-world attacks with natural styles. In CVPR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 288, + 406 + ], + "type": "text", + "content": "[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. TIP, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 407, + 288, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 439 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 439 + ], + "type": "text", + "content": "[8] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In ACM Multimedia, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 441, + 288, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 288, + 484 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 288, + 484 + ], + "type": "text", + "content": "[9] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Me-d2n: Multi-expert domain decompositional network for cross-domain few-shot learning. In ACM Multimedia, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 486, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 288, + 529 + ], + "type": "text", + "content": "[10] Yuqian Fu, Yu Xie, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Wave-san: Wavelet based style augmentation network for cross-domain few-shot learning. arXiv preprint, 2022. 1, 2, 3, 5, 6, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 532, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 575 + ], + "type": "text", + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 576, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 288, + 598 + ], + "type": "text", + "content": "[12] Victor Garcia and Joan Bruna. Few-shot learning with graph neural networks. arXiv preprint, 2017. 2, 5, 6, 12, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 632 + ], + "type": "text", + "content": "[13] Micah Goldblum, Liam Fowl, and Tom Goldstein. Adversarially robust few-shot learning: A meta-learning approach. NeurIPS, 2020. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "type": "text", + "content": "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint, 2014. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 668, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 288, + 690 + ], + "type": "text", + "content": "[15] Jiechao Guan, Manli Zhang, and Zhiwu Lu. Large-scale cross-domain few-shot learning. In ACCV, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[16] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, 2020. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 128 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 130, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 547, + 173 + ], + "type": "text", + "content": "[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens., 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "text", + "content": "[19] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In CVPR, 2022. 2, 5, 6, 12, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 251 + ], + "type": "text", + "content": "[20] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 253, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 547, + 285 + ], + "type": "text", + "content": "[21] Zhengdong Hu, Yifan Sun, and Yi Yang. Switch to generalize: Domain-switch learning for cross-domain few-shot classification. In ICLR, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 286, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 319 + ], + "type": "text", + "content": "[22] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 321, + 547, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 364 + ], + "type": "text", + "content": "[23] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard Radke. Dynamic distillation network for cross-domain few-shot recognition with unlabeled data. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 365, + 547, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 547, + 398 + ], + "type": "text", + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, 2013. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 399, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 547, + 442 + ], + "type": "text", + "content": "[25] Manoj Kumar, Varun Kumar, Hadrien Glaude, Cyprien de Lichy, Aman Alok, and Rahul Gupta. Protoda: Efficient transfer learning for few-shot intent classification. In SLT Workshop, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 444, + 547, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 547, + 465 + ], + "type": "text", + "content": "[26] Cassidy Laidlaw and Soheil Feizi. Functional adversarial attacks. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 467, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 547, + 487 + ], + "type": "text", + "content": "[27] Bo Li and Yevgeniy Vorobeychik. Feature cross-substitution in adversarial classification. NeurIPS, 2014. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 490, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 547, + 521 + ], + "type": "text", + "content": "[28] Kai Li, Yulun Zhang, Kunpeng Li, and Yun Fu. Adversarial feature hallucination networks for few-shot learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 523, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 547, + 555 + ], + "type": "text", + "content": "[29] Pan Li, Shaogang Gong, Yanwei Fu, and Chengjie Wang. Ranking distance calibration for cross-domain few-shot learning. arXiv preprint, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 557, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 547, + 590 + ], + "type": "text", + "content": "[30] Wenbin Li, Lei Wang, Xingxing Zhang, Jing Huo, Yang Gao, and Jiebo Luo. Defensive few-shot adversarial learning. arXiv preprint, 2019. 2, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 591, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 547, + 624 + ], + "type": "text", + "content": "[31] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Ling-Yu Duan. Uncertainty modeling for out-of-distribution generalization. arXiv preprint, 2022. 3, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 624, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 668 + ], + "type": "text", + "content": "[32] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In ICCV, 2021. 2, 5, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[33] Bingyu Liu, Zhen Zhao, Zhenpeng Li, Jianan Jiang, Yuhong Guo, and Jieping Ye. Feature transformation ensemble model with batch spectral regularization for cross-domain few-shot classification. arXiv preprint, 2020. 2, 5, 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "24583" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint, 2017. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 152 + ], + "type": "text", + "content": "[35] Sharada P Mohanty, David P Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in plant science, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 288, + 186 + ], + "type": "text", + "content": "[36] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In CVPR, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 288, + 220 + ], + "type": "text", + "content": "[37] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. arXiv preprint, 2020. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 222, + 288, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 288, + 276 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 278, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 278, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 278, + 288, + 300 + ], + "type": "text", + "content": "[39] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 302, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 288, + 334 + ], + "type": "text", + "content": "[40] Wei Shen, Ziqiang Shi, and Jun Sun. Learning from adversarial features for few-shot classification. arXiv preprint, 2019. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 336, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 336, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 336, + 288, + 358 + ], + "type": "text", + "content": "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 360, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 404 + ], + "type": "text", + "content": "[42] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. arXiv preprint, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 406, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 288, + 437 + ], + "type": "text", + "content": "[43] Qianru Sun, Yaoyao Liu, Tat-Seng Chua, and Bernt Schiele. Meta-transfer learning for few-shot learning. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 440, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 288, + 483 + ], + "type": "text", + "content": "[44] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, 2018. 12, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 486, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 288, + 518 + ], + "type": "text", + "content": "[45] Hao Tang, Zechao Li, Zhimao Peng, and Jinhui Tang. Blockmix: meta regularization and self-calibrated inference for metric-based meta-learning. In ACM Multimedia, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 520, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 288, + 553 + ], + "type": "text", + "content": "[46] Hao Tang, Chengcheng Yuan, Zechao Li, and Jinhui Tang. Learning attention-guided pyramidal features for few-shot fine-grained recognition. PR. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 555, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 598 + ], + "type": "text", + "content": "[47] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Scientific data, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 600, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 633 + ], + "type": "text", + "content": "[48] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1, 2, 5, 6, 12, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[49] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[50] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 703 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[51] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. arXiv preprint, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[52] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. arXiv preprint, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "text", + "content": "[53] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In CVPR, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 208, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 546, + 239 + ], + "type": "text", + "content": "[54] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In ICCV, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 274 + ], + "type": "text", + "content": "[55] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 276, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 546, + 319 + ], + "type": "text", + "content": "[56] Chengming Xu, Yanwei Fu, Chen Liu, Chengjie Wang, Jilin Li, Feiyue Huang, Li Zhang, and Xiangyang Xue. Learning dynamic alignment via meta-filter for few-shot learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 320, + 546, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 353 + ], + "type": "text", + "content": "[57] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In AAAI, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 354, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 386 + ], + "type": "text", + "content": "[58] Ji Zhang, Jingkuan Song, Lianli Gao, Ye Liu, and Heng Tao Shen. Progressive meta-learning with curriculum. TCVT, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 388, + 547, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 432 + ], + "type": "text", + "content": "[59] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 434, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 546, + 465 + ], + "type": "text", + "content": "[60] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In ICLR. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 467, + 546, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 546, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 546, + 499 + ], + "type": "text", + "content": "[61] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In CVPR, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 501, + 546, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 546, + 533 + ], + "type": "text", + "content": "[62] Zhedong Zheng and Yi Yang. Rectifying pseudo label learning via uncertainty estimation for domain adaptive semantic segmentation. IJCV, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 535, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 546, + 567 + ], + "type": "text", + "content": "[63] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint, 2022. 3, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 568, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 546, + 601 + ], + "type": "text", + "content": "[64] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 2017. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 602, + 546, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 546, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 546, + 633 + ], + "type": "text", + "content": "[65] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 635, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 667 + ], + "type": "text", + "content": "[66] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In ICLR, 2021. 3, 8, 12" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 670, + 547, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 703 + ], + "type": "text", + "content": "[67] Linhai Zhuo, Yuqian Fu, Jingjing Chen, Yixin Cao, and YuGang Jiang. Tgdm: Target guided dynamic mixup for cross-domain few-shot learning. In ACM Multimedia, 2022. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "24584" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_content_list.json b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a7b96fd721f15f315a1121812343a290c4091712 --- /dev/null +++ b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_content_list.json @@ -0,0 +1,1670 @@ +[ + { + "type": "text", + "text": "StyleGAN Salon: Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer", + "text_level": 1, + "bbox": [ + 130, + 130, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sasikarn Khwanmuang\\* Pakkapon Phongthawee\\* Patsorn Sangkloy\\* Supasorn Suwajanakorn", + "bbox": [ + 86, + 210, + 901, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VISTEC, Thailand Phranakhon Rajabhat University, Thailand", + "bbox": [ + 220, + 233, + 769, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sasikarn.k_s18, pakkapon.p_s19, supasorn.s}@vistec.ac.th patsn.s@pnru.ac.th", + "bbox": [ + 132, + 258, + 849, + 272 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1facb3b704e8e0370a119f7004f0d6c0b61389ee356caa02c58565d5813fa96d.jpg", + "image_caption": [ + "Input Face", + "Figure 1. Our method can transfer the hairstyle from any reference hair image in the top row to Tom Holland [33], in the second row." + ], + "image_footnote": [], + "bbox": [ + 173, + 305, + 890, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 515, + 313, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our paper seeks to transfer the hairstyle of a reference image to an input photo for virtual hair try-on. We target a variety of challenges scenarios, such as transforming a long hairstyle with bangs to a pixie cut, which requires removing the existing hair and inferring how the forehead would look, or transferring partially visible hair from a hat-wearing person in a different pose. Past solutions leverage StyleGAN for hallucinating any missing parts and producing a seamless face-hair composite through so-called GAN inversion or projection. However, there remains a challenge in controlling the hallucinations to accurately transfer hairstyle and preserve the face shape and identity of the input. To overcome this, we propose a multi-view optimization framework that uses two different views of reference composites to semantically guide occluded or ambiguous regions. Our optimization shares information between two poses, which allows us to produce high fidelity and realistic results from incomplete references. Our framework produces high-quality results and outperforms prior work in a user study that consists of significantly more challenging hair transfer scenarios than previously studied. Project page: https://stylegan-salon.github.io/.", + "bbox": [ + 75, + 551, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 515, + 630, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What makes Jennifer Aniston keep her same hairstyle for over three decades? Perhaps she likes the classic, or perhaps changing her hairstyle is a decision too high-stakes that she could later regret. Unlike garments or makeup, trying on a new hairstyle is not easy, and being able to imagine yourself in different hairstyles could be an indispensable tool.", + "bbox": [ + 496, + 545, + 893, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches for hairstyle transfer, StyleYourHair [20], Barbershop [46], LOHO [30], and Michiganan [35], allow users to manipulate multiple hair attributes of an input image, such as appearance, shape, or color by providing a reference image for each different attribute. These methods [20, 30, 46] rely on a generative adversarial network [11], specifically StyleGAN2 [19], which can synthesize highly realistic face images. Their key idea, which also forms the basis of our method, is to leverage the realistic face distribution learned by StyleGAN and search for a hairstyle-transfer output whose latent code lies within the learned distribution using optimization (commonly known as GAN projection or inversion).", + "bbox": [ + 496, + 638, + 895, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our extensive study on these state-of-the-art techniques still reveal several unsolved challenges for in-the-wild hairstyle transfer. One of the main challenges is when the reference hair comes from a person with a very different head", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pose or facial shape. In this case, the transfer result often degrades significantly [30, 46]. HairFIT [8] and the recently proposed StyleYourHair [20] both attempt to solve this using an additional alignment step to align the pose of the target hair to the input face. In HairFIT [8], this is done explicitly via a flow-based warping module for hair segmentation, but this requires training on multi-view datasets [21, 26]. StyleYourHair [20] avoids this issue and also improves upon HairFIT's results by optimizing for an aligned pose within StyleGAN2's latent space using distances between the detected facial keypoints. While StyleYourHair can handle a certain degree of misalignment, it often struggles to preserve details of the reference hair texture, especially for intricate and non-straight hairstyles, e.g., in Figure 3.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, we have observed that there is a trade-off between hallucinating new details, which is crucial for handling highly different poses, and preserving the original texture from the reference images. These two goals are often at odds with each others. This trade-off is also evident in the results from StyleYourHair, as reported in [20] and as shown in table 3, where BarberShop [46] can still produce better results, for example, when the pose difference is small.", + "bbox": [ + 75, + 303, + 472, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tackle this dilemma by performing a multi-stage optimization. This serves two purposes: first, to hallucinate new details necessary for aligning the poses, and second, to recover face-hair details from the original input images. We preserve these details in a form of two guide images in both viewpoints, which will be jointly optimized to allow new details to be filled while retaining face-hair texture in the original pose. Our pose alignment is done both explicitly via 3D projection [6] on RGB images, and implicitly via latent code(s) sharing during our multi-view optimization.", + "bbox": [ + 75, + 425, + 472, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 96, + 577, + 395, + 592 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose StyleGAN Salon: a pose-invariant hairstyle transfer pipeline that is flexible enough to handle a variety of challenging scenarios including, but not limited to, bangs/hat removal and background inpainting.", + "2. Unlike previous works, our method operates entirely on RGB images, which are more flexible than segmentation masks. This allows us to first draft the output and then refine them via multi-stage optimization.", + "3. We introduce multi-view optimization for hairstyle transfer which incorporates 3D information to align the poses for both face and hair images. Our method leverages both views to help preserve details from the original images.", + "4. We thoroughly analyze the results in several experiments, including a user study with detailed breakdown into various challenging scenarios. Our method shows superior results over existing works in all scenarios." + ], + "bbox": [ + 88, + 603, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Adversarial Networks. Beginning with [11, 18], StyleGANs [17-19] have shown great success in 2D image generation by learning realistic training data distribution corresponding to a fixed low-dimensional distribution, called latent code. The learned latent code can be applied to a downstream task such as image manipulation [13, 15, 32, 40]. Recent work on [6, 12, 44] also explore 3D aware architecture in StyleGAN, resulting in multi-view consistent images.", + "bbox": [ + 496, + 114, + 893, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For hairstyle transfers, MichiGAN [35] uses a conditional hair generation network that can control hair shape, structure, and appearance. Recently, HairFIT [8], a pose-invariant hairstyle transfer network, aligns the reference hair to match the input face pose using a optical flow-based hair alignment module, but requires training on multi-view dataset [21, 26].", + "bbox": [ + 496, + 236, + 893, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGAN Latent Space and Projection Techniques. To generate an image, StyleGAN2 first maps a random latent code $z \\sim N(0, I)$ to an intermediate latent code $w \\in \\mathbb{R}^{512}$ in a new space called $\\mathcal{W}$ . This code $w$ is then replicated (18x) and used as input to each layer of the generator that controls details at different scales through weight demodulation. Collectively, these replicated latent codes are referred to as $w^{+} \\in \\mathbb{R}^{18 \\times 512}$ in the extended latent space $\\mathcal{W}^{+}$ .", + "bbox": [ + 496, + 327, + 893, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pivotal Tuning Inversion (PTI) [29] further improves the projection quality by directly tuning generator weights around the optimized latent code $w$ using their regularization technique. Other techniques [1,2,5] optimize the latent code $w^{+}$ in each layer separately to better match the input image.", + "bbox": [ + 496, + 446, + 893, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PIE [37] introduces hierarchical optimization for semantic attributes editing that first optimizes the latent code in the $\\mathcal{W}$ space, then transfers the code to the $\\mathcal{W}^+$ space and continues the optimization in that space. We adopt similar hierarchical optimization that uses both $\\mathcal{W}$ and $\\mathcal{W}^+$ . Part of our method is also inspired by PULSE [25], which reconstructs a high-resolution image from a small reference image (32x32) by searching for the closest image in the StyleGAN latent space that resembles the low-resolution reference.", + "bbox": [ + 496, + 523, + 893, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some methods [7, 41, 45] project multiple images into latent space simultaneously. However, all of their inputs are complete, whereas our method requires hair information from the reference hair image and additional information from the input face image.", + "bbox": [ + 496, + 659, + 893, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGAN-Based Hairstyle Transfer. LOHO [30] adopts loss functions from Image2StyleGAN++ [2] to combine face-hair inputs into a hairstyle transfer output with their orthogonalization optimization technique which reduces conflicts between multiple loss functions. Barbershop [46] first predicts semantic regions of both inputs and uses them to create a \"target segmentation mask\" of the output with rule (in the original paper) or GAN-based inpainting (in their official code). Then, they optimize two separate latent codes in $\\mathcal{W}^+$ space, one for matching the face and the other for hair, while conforming to the target segmentation mask. To", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8610", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "preserve the original details, the latent optimization is done in their proposed $\\mathcal{F} / \\mathcal{S}$ space, which replaces the first seven blocks of $\\mathcal{W}^+$ space with the corresponding activation maps of StyleGAN's convolution layers. To improve Barbershop's capacity to handle unaligned input poses, StyleYourHair [20] first aligns the reference hair to match the pose of the input face via the proposed local-style-matching loss. However, this alignment often leads to an unrealistic hair shape or inaccurate hair texture results. In contrast, our method simultaneously optimizes for two guides in face and hair views, resulting in a better hair texture.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of estimating the final output from the segmentation mask [20, 46], our multi-view optimization uses face-hair composites in RGB space to overcome this problem, and produces results that better preserve the input facial structure and hairstyle across a wider, more challenging range of scenarios. Concurrent work, HairNet [47], arrives at a similar goal of removing the target segmentation mask via a two-step process that involves baldification using StyleFlow [3] followed by another network to transfer the hairstyle.", + "bbox": [ + 75, + 257, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, handling pose differences is crucial for successful hairstyle transfer; our method incorporates 3D rotation to preserve the geometric consistency and systematically evaluate this aspect.", + "bbox": [ + 75, + 393, + 472, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. StyleGAN Salon", + "text_level": 1, + "bbox": [ + 76, + 467, + 238, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given two input images $I_h$ and $I_f$ , the goal is to transfer the hair from the reference hair image $I_h$ into the face in the image $I_f$ , while preserving all remaining details in $I_f$ , including identity, clothing, and background.", + "bbox": [ + 75, + 492, + 470, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key idea of our approach is to guide the optimization on the learned latent space of StyleGAN2 with two \"guide\" images, which represent rough composites of the final output based on simple cut-and-paste in the viewpoint of $I_{h}$ and $I_{f}$ . We leverage EG3D to construct these guided images in a geometrically consistent way, described in Section 3.1.", + "bbox": [ + 75, + 553, + 472, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization on StyleGAN2's latent space is commonly performed on either the original latent space $\\mathcal{W}$ or the extended latent space $\\mathcal{W}^+$ . Optimizing on $\\mathcal{W}$ space generally leads to more realistic results by staying within the original latent space, whereas optimizing on $\\mathcal{W}^+$ allows a closer match to the reference [37]. In hairstyle transfer, it is important to stay faithful to the input images and preserve important details such as hair texture, face identity, and background. However, optimizing on $\\mathcal{W}^+$ will lead to poor results because our guide images are rough and unrealistic composites. Thus, we propose to optimize on $\\mathcal{W}$ space followed by $\\mathcal{W}^+$ space, similar to a technique in PIE [37] used for editing semantic attributes of an image.", + "bbox": [ + 75, + 643, + 472, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our optimization incorporates both guide images from the two viewpoints, detailed in Section 3.2. Section 3.3 and Section 3.4 cover details of our optimization on $\\mathcal{W}$ and $\\mathcal{W}^+$ , respectively. Finally, we also optimize StyleGAN2 weights", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while freezing the latent codes (Section 3.5) using PTI [29] to further improve detail fidelity. Figure 2 shows an overview of our complete pipeline.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Constructing the Guide Images", + "text_level": 1, + "bbox": [ + 498, + 145, + 777, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The purpose of our guide images is to provide an initial estimate of how the hair would look on $I_{f}$ . We achieve this using a cut-and-paste composite of the face and background from $I_{f}$ and the hair from $I_{h}$ . To better handle a potentially large shift in viewpoint between $I_{f}$ and $I_{h}$ , we propose to leverage EG3D [6] to help generate geometrically consistent guide images. We argue that using multi-view guide images, each in the pose of $I_{f}$ and $I_{h}$ , helps preserve details that could otherwise be lost from using a single viewpoint alone.", + "bbox": [ + 496, + 167, + 893, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A straightforward approach is to simply project $I_h$ into the EG3D [6] latent space, and use their proposed neural rendering pipeline to render it in the view of $I_f$ (and vice versa). However, while the projection can produce a geometrically consistent 3D shape, we found that the resulting texture is not very accurate. To address this, we replace any visible regions in the texture from $I_h$ 's viewpoint with the original pixels of $I_h$ , while leaving the rest of the texture as the projected texture from EG3D's rendering.", + "bbox": [ + 496, + 304, + 893, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also additionally apply uniform scaling and translation to match the faces' widths and centers, which are computed based on detected facial keypoints [22]. These keypoints, along with semantic regions from [42], are also used to handle various corner cases, such as re-painting unwanted hair regions. We refer to Appendix A for more details of this operation.", + "bbox": [ + 496, + 439, + 893, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The entire process is done for both $I_{f}$ and $I_{h}$ viewpoints, resulting in a pair of guide images $I_{\\mathrm{guide}} = [I_{\\mathrm{guide}}^{\\mathrm{face}}, I_{\\mathrm{guide}}^{\\mathrm{hair}}]$ . We emphasize that by leveraging our multi-view optimization (Section 3.2), these guide images do not need to be precise or realistic to produce convincing final results, as demonstrated in the first row of Figure 4.", + "bbox": [ + 496, + 546, + 893, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-View Latent Optimization", + "text_level": 1, + "bbox": [ + 498, + 643, + 784, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our guide images provide complementary information about the target hair and face, albeit from different poses. However, each guide image is only fully accurate in regions with the original pixels seen in the original viewpoint and not warped by EG3D. Thus, our optimization goal is to combine information from both guide images to generate a final output that accurately captures the realistic hair from $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ , as well as other details from $I_{\\mathrm{guide}}^{\\mathrm{face}}$ . We achieve this using multiple loss functions that attend to both viewpoints with different spatial emphasis. Specifically, we perform multi-view latent optimization on $w / w^{+}$ (and stochastic noise maps $n$ ) that fits both guide images:", + "bbox": [ + 496, + 667, + 893, + 851 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\{w / w ^ {+}, n \\}} \\sum_ {\\mathrm {i}} ^ {[ \\text {f a c e , h a i r} ]} \\mathcal {L} _ {\\text {l o s s}} \\left(O ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 859, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8611", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d7cdeef94bda516a58d021f9cf161129d9929addd054f442dd842c985723a89.jpg", + "image_caption": [ + "Figure 2. Overview of StyleGAN Salon: We first align the input face $I_{f}$ and reference hair $I_{h}$ and use them to construct guide images $I_{\\mathrm{guide}}$ , in two different viewpoints, which specifies the target appearance for each output region. (see Section 3)" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $O^{(\\mathrm{i})}$ denotes the output of the StyleGAN2's generator that takes a latent code $w^{(\\mathrm{i})}$ or $w^{+\\mathrm{(i)}}$ and a stochastic noise input $n^{(\\mathrm{i})}$ . $\\mathcal{L}_{\\mathrm{loss}}$ is a sum of all our loss functions:", + "bbox": [ + 75, + 311, + 470, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {l o s s}} = \\sum_ {\\mathrm {j}} ^ {[ \\mathrm {f}, \\mathrm {h}, \\mathrm {b g} ]} \\lambda_ {\\mathrm {p}} ^ {(\\mathrm {j})} \\mathcal {L} _ {\\text {p e r}} ^ {(\\mathrm {j})} (2) \\\\ + \\lambda_ {\\mathrm {g}} \\mathcal {L} _ {\\text {g l o b a l}} + \\lambda_ {\\mathrm {i}} \\mathcal {L} _ {\\text {i n i}} + \\lambda_ {\\varepsilon} \\mathcal {L} _ {\\varepsilon} + \\lambda_ {s} \\mathcal {L} _ {\\text {s i m}}, (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 367, + 468, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{(\\cdot)}$ are balancing weights, and [f, h, bg] refer to face, hair, and background. We next explain each loss function.", + "bbox": [ + 75, + 441, + 468, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 491, + 235, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of our $\\mathcal{W}$ and $\\mathcal{W}^+$ optimization is to generate an output image that fits the corresponding $I_{\\mathrm{guide}}$ without appearing unrealistic. The main challenge is that our guide images are unrealistic and can contain various unwanted artifacts. We design our loss functions to be tolerant of the imprecise nature of our guide images.", + "bbox": [ + 75, + 513, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Perceptual Loss: This loss is based on Learned Perceptual Image Patch Similarity (LPIPS) [43], which compares two input images in a deep feature space. Similar to [20, 30, 46], we apply a binary mask to indicate regions of interest in the deep feature space of the two input images. We use $L_{\\mathrm{LPIPS}}(I_1, I_2; M)$ to denote this LPIPS computation between two images $(I_1, I_2)$ with mask $M$ . However, we observe that applying this masking operation after the feature computation is insufficient to disregard unwanted regions. This is due to the fact that LPIPS is a patch-based similarity loss, and regions outside of the mask can still affect the loss computation. Based on this principle, we propose to apply additional pre-masking to mask out ambiguous regions in $I_{\\mathrm{guide}}$ that should not be trusted. Specifically, our Masked Perceptual Loss is the following:", + "bbox": [ + 75, + 604, + 470, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e r}} ^ {\\mathrm {i}, \\mathrm {r o i}} = \\Lambda_ {\\mathrm {r o i}} ^ {(\\mathrm {i})} L _ {\\mathrm {L P I P S}} \\left(O ^ {(\\mathrm {i})} \\odot \\neg M _ {\\mathrm {r o n i}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})} \\odot \\neg M _ {\\mathrm {r o n i}} ^ {(\\mathrm {i})}; M _ {\\mathrm {r o i}} ^ {(\\mathrm {i})}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 840, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{\\mathrm{roi}}$ is the mask for region of interest (ROI), and $M_{\\mathrm{roni}}$ is the mask for regions of not interest (RONI). $(.)\\odot \\neg M_{\\mathrm{roni}}$", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a simple element-wise multiplication, which effectively excludes these regions from the loss by setting them to 0 in both input images. And $\\Lambda_{\\mathrm{roi}}^{\\mathrm{(i)}}$ is a balancing weight in (i) viewpoint, which is set higher when the region of interest covers the original pixels of $I_{f}$ (or $I_{h}$ ). For instance, the loss that attends to face in $I_{\\mathrm{guide}}^{\\mathrm{face}}$ has a higher weight than the face loss in $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ . We consider 3 regions of interest: $L_{\\mathrm{per}}^{\\mathrm{f}}$ , $L_{\\mathrm{per}}^{\\mathrm{h}}$ , and $L_{\\mathrm{per}}^{\\mathrm{bg}}$ for face, hair, and background regions. We refer to Section 3.3 and 3.4 for details on these mask generation for $\\mathcal{W}$ and $\\mathcal{W}^{+}$ optimization, respectively.", + "bbox": [ + 496, + 311, + 893, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global loss: This loss function attempts to match the overall appearance of the output to $I_{\\mathrm{guide}}$ :", + "bbox": [ + 498, + 469, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g l o b a l}} = L _ {\\text {M S E}} ^ {3 2} (O, I _ {\\text {g u i d e}}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 507, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{\\mathrm{MSE}}^{32}$ is the mean square error computed on $32 \\times 32$ downsampled input images. The goal of the downsampling is to reduce the effect of matching visible seams from higher resolution $I_{\\mathrm{guide}}$ because those seams become imperceptible once downsampled. This loss is inspired by a similar idea used for image super-resolution by Menon et al. [25].", + "bbox": [ + 496, + 532, + 893, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initial Value Loss: To prevent the output from deviating too far from the learned distribution and becoming unrealistic, we force the optimized latent code(s) $w$ and $w^{+}$ to be close to its initial value $w_{0}$ through L2 loss, as used in [2]:", + "bbox": [ + 496, + 625, + 893, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i n i}} = \\left\\| w _ {1: 1 8} - w _ {0} \\right\\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 693, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the estimated mean of $\\mathcal{W}$ latent space as initial values for optimizing $w$ . This is computed by averaging many latent codes $w_{i} = f(z_{i})$ drawn randomly through $z\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and the StyleGAN's mapping network $f$ . Then the optimized latent code $w$ from the first stage becomes the initial value for the optimization in $\\mathcal{W}^+$ extended latent space.", + "bbox": [ + 496, + 719, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noise Regularization Loss: We also use the noise regularization loss $(\\mathcal{L}_{\\varepsilon})$ proposed in StyleGAN2 [19] to ensure that the noise maps will be used for capturing the stochastic variations, and not for encoding the content meant to be captured by the latent code. We refer to StyleGAN2 for the details of this loss.", + "bbox": [ + 496, + 810, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Similarity Loss: To enable information sharing between the two guide images ( $I_{\\mathrm{guide}}^{\\mathrm{face}}$ and $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ ) from two viewpoints, we force both their latent codes to be close together by using the following L2 loss, similarly to [7].", + "bbox": [ + 76, + 90, + 470, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m}} = \\left\\| w ^ {(\\text {f a c e})} - w ^ {(\\text {h a i r})} \\right\\| _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 162, + 470, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ can be in $\\mathcal{W}$ space during $\\mathcal{W}$ optimization or replaced with $w^{+}$ during $\\mathcal{W}^+$ optimization.", + "bbox": [ + 76, + 191, + 470, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Sharing Latent Code", + "text_level": 1, + "bbox": [ + 76, + 242, + 277, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In StyleGAN2, latent codes in early layers have been shown to correspond to high level concepts (e.g. human pose) whereas the remaining layers capture low-level information (e.g. colors) [15, 18]. This motivates the optimization in $\\mathcal{W}^+$ extended space, as each layer can be optimized for different aspects of the image. Similarly, we optimize latent codes to match guide images of the same person and hairstyle but in different poses. Rather than optimizing two latent codes independently, we can enforce information sharing between the two poses by constraining the latent codes of the last few layers ( $w_{l:18}$ ) to be the same. We set $l$ to be 4 in all our experiments, loosely based on the head rotation experiments shown in GANSpace [15].", + "bbox": [ + 75, + 266, + 468, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sharing $w$ Latent Code: Optimization in the space of StyleGAN2 latent code $\\mathcal{W}$ is generally done as a single latent code $(\\mathbb{R}^{512})$ , which are duplicated (18x) and fed into all the layers. In our case, we also optimize for a single latent code $w$ but only duplicate them for the first $l$ layers $(w_{1:l})$ . The remaining layers $(w_{l:18})$ are interpolated from the current latent codes of each guide image with a coefficient $\\alpha$ .", + "bbox": [ + 76, + 463, + 470, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {l: 1 8} = \\alpha w _ {1: l} ^ {\\text {(f a c e)}} + (1 - \\alpha) w _ {1: l} ^ {\\text {(h a i r)}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 579, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We randomize this coefficient to make the latent codes stay within the space of $\\mathcal{W}$ . This random interpolation forces the latent code for each guide image to be partly similar while still producing realistic results from $\\mathcal{W}$ latent space.", + "bbox": [ + 76, + 609, + 468, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sharing $w^{+}$ Latent Code: Optimization in $\\mathcal{W}^{+}$ extended space is straightforward, as latent code for each layer can already be separately optimized. We simply share the last few layers $(w_{l:18})$ of both $w^{+}$ latent codes during optimization. We refer to the Appendix A.5 for more details on the optimization process.", + "bbox": [ + 76, + 670, + 470, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. $\\mathcal{W}$ Optimization: Hallucinate Missing Details", + "text_level": 1, + "bbox": [ + 76, + 770, + 457, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This stage aims to hallucinate details not currently visible in the guide images. The output images of this stage are $O_{1}^{\\mathrm{face}}$ and $O_{1}^{\\mathrm{hair}}$ , which are essentially the projection of the unrealistic $I_{\\mathrm{guide}}^{\\mathrm{face}}$ and $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ into the real image distribution learned in StyleGAN2 [19]. This is done by optimizing on $W$ latent space to fit each guide image with the objective function (Equation 2).", + "bbox": [ + 76, + 794, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While our guide images ( $I_{\\mathrm{guide}}^{\\mathrm{face}}$ and $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ ) capture the overall appearance of the desired outputs, they still lack details in certain regions. These unknown regions may correspond to unseen facial features that were occluded by the hair in $I_{f}$ or incomplete reference hair from $I_{h}$ caused by, e.g., a hat or image cropping. These details will need to be hallucinated and seamlessly blended with the rest of the image.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To accomplish this, we design our pre-masking in Masked Perceptual Loss $(\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{f}},\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{h}}$ , and $\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{bg}}$ ; Equation 4) to include all the unknown regions that need to be hallucinated. Specifically, $M_{\\mathrm{roni}}^{\\mathrm{f}}$ represents the face region occluded by the hair in $I_{f}$ , e.g., the forehead behind the bangs or the ears that should become visible in the final output. Analogously, $M_{\\mathrm{roni}}^{\\mathrm{h}}$ represents the hair region occluded by other objects (e.g., a hat) or not visible in $I_{h}$ due to image cropping. We set $M_{\\mathrm{roi}}^{\\mathrm{f}}$ and $M_{\\mathrm{roi}}^{\\mathrm{h}}$ to be segmentation masks for face and hair regions $(M_{\\mathrm{f}}$ and $M_{\\mathrm{h}})$ , respectively. Both $M_{\\mathrm{roi}}^{\\mathrm{bg}}$ and $\\neg M_{\\mathrm{roni}}^{\\mathrm{bg}}$ is set to be the background regions in $I_{f}$ not covered by the transferred hair $(M_{\\mathrm{bg}})$ . These masks are constructed by composing different semantic regions (union, intersection, etc.) from $I_{f}$ and $I_{h}$ , detailed in Appendix A.3", + "bbox": [ + 496, + 196, + 892, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that $M_{\\mathrm{roi}}$ is applied on the deep feature maps, similar to [20, 30, 46], while $\\neg M_{\\mathrm{roni}}$ is applied on raw RGB images. Using both masking techniques is crucial to our natural and seamless blending. The additional pre-masking allows the hair shape (or face shape) in the output to be different from $I_{\\mathrm{guide}}$ 's and freely grow outward outside $M_{\\mathrm{roi}}$ if this leads to a more natural result (see Figure 4).", + "bbox": [ + 496, + 410, + 893, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. $\\mathcal{W}^+$ Optimization: Recover Face-Hair Details", + "text_level": 1, + "bbox": [ + 498, + 523, + 883, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output of the previous stage, $O_1^{\\mathrm{face}}$ and $O_1^{\\mathrm{hair}}$ , may still look different from the input person and not yet capture the hair details from the reference. This stage aims at refining these output images to better reproduce the hair details from $I_h$ and the rest from $I_f$ . The optimization is done in the extended $\\mathcal{W}^+$ space with respect to $\\{w^{+},n\\}$ . In other words, we allow $w\\in \\mathbb{R}^{512}$ that was previously replicated to each layer to be optimized separately as $w^{+}\\in \\mathbb{R}^{18\\times 512}$ . We denote $O_2^{\\mathrm{face}}$ and $O_2^{\\mathrm{hair}}$ to be the output of this stage generated by StyleGAN2 from our optimized latent codes in $\\mathcal{W}^+$ extended latent space.", + "bbox": [ + 496, + 546, + 893, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We update the target images for the optimization from $I_{\\mathrm{guide}}^{\\mathrm{face}}$ and $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ to the more complete versions based on $O_1^{\\mathrm{face}}$ and $O_1^{\\mathrm{hair}}$ from the first stage. However, because the first stage aims at hallucinating new details into the pre-masking regions, it may lose the original texture details. To address this, we replace the known regions in $O_1^{\\mathrm{face}}$ and $O_1^{\\mathrm{hair}}$ with the correct details from the original $I_{\\mathrm{guide}}$ by setting", + "bbox": [ + 496, + 713, + 892, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {n e w} \\text {g u i d e}} \\leftarrow I _ {\\text {g u i d e}} \\odot M _ {\\mathrm {c}} + O _ {1} \\odot \\neg M _ {\\mathrm {c}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 830, + 892, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{c} = M_{\\mathrm{f}} \\cup M_{\\mathrm{h}} \\cup M_{\\mathrm{bg}}$ , which corresponds to the regions of interest in our Masked Perceptual Loss where we want to match with $I_{\\mathrm{guide}}$ . In short, we simply copy the", + "bbox": [ + 496, + 854, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hallucinated parts from $O_1^{\\mathrm{face}}$ and $O_1^{\\mathrm{hair}}$ , and combine them with the part in $I_{\\mathrm{guide}}$ that are known to be correct.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our Masked Perceptual Loss, we use the same masking procedure as the previous stage (Section 3.3). However, we observe that the projection from EG3D, which produces $I_{\\mathrm{guide}}$ , generally is inferior to the known details from the original images. For example, $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ contains more details on the hair than $I_{\\mathrm{guide}}^{\\mathrm{face}}$ due to the latter being hallucinated from the 3D rotation by EG3D's rendering pipeline. To de-prioritize these inaccurate details, we blur both $I_{\\mathrm{guide}}$ and $O$ when computing the perceptual loss for the less accurate regions of interest. Concretely, we only blur images when computing $\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{f}}$ for $I_{\\mathrm{guide}}^{\\mathrm{hair}}$ optimization and when computing $\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{h}}$ for $I_{\\mathrm{guide}}^{\\mathrm{face}}$ optimization.", + "bbox": [ + 75, + 121, + 473, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Pivotal Tuning Inversion", + "text_level": 1, + "bbox": [ + 76, + 315, + 302, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The purpose of this stage is to further refine the output images $(O_2^{\\mathrm{face}}$ and $O_2^{\\mathrm{hair}})$ by allowing the optimization of the StyleGAN's weights $\\theta$ (while fixing the optimized latent code $w_{\\mathrm{optimized}}$ ). We closely follow the proposed optimization in PTI [29] with the same objective function. However, we update the reconstruction loss to reflect the goal of our task, which is to match the original details from $I_{f}$ and $I_{h}$ .", + "bbox": [ + 75, + 338, + 472, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p t i}} = \\sum_ {\\mathrm {i}} ^ {[ \\text {f a c e}, \\text {h a r i} ]} \\left(L _ {\\mathrm {L P I P S}} \\left(O _ {\\text {t u n e}} ^ {(\\mathrm {i})} \\odot M _ {\\text {r a w}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})} \\odot M _ {\\text {r a w}} ^ {(\\mathrm {i})}; M _ {\\text {r a w}} ^ {(\\mathrm {i})}\\right) \\right. \\\\ + L _ {\\mathrm {M S E}} ^ {3 2} \\left(O _ {\\text {t u n e}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})}; \\neg M _ {\\text {r a w}} ^ {(\\mathrm {i})}\\right)), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 454, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $O_{\\mathrm{tune}}$ is the generated image using the tuned weights $\\theta$ , the mask $M_{\\mathrm{raw}}^{\\mathrm{face}}$ is $(M_{\\mathrm{bg}} \\cup M_{\\mathrm{f}})$ , and $M_{\\mathrm{raw}}^{\\mathrm{hair}}$ is $M_{\\mathrm{h}}$ .", + "bbox": [ + 75, + 527, + 468, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 571, + 209, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our method to state-of-the-art StyleYourHair [20], Barbershop [46], and LOHO [30]. Our evaluation criteria are i) user preference via a user study on a wide variety of scenarios, ii) hairstyle transfer quality, iii) hair reconstruction quality, and iv) how well the input face shape is preserved. Section 4.5 presents ablation studies on our multi-view sharing latent optimization, optimization stage, and loss functions.", + "bbox": [ + 75, + 597, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the official code of LOHO [30] and Barbershop [46] with the default configurations. For StyleYourHair [20], we use the configuration where the hair reference is never flipped.", + "bbox": [ + 75, + 717, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 786, + 299, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide a qualitative comparison to StyleYourHair [20], Barbershop [46], and LOHO [30] in Figure 3. We observe that LOHO and Barbershop often struggle to fit the reference hair accurately when the poses are not well aligned (rows 3-5), resulting in various artifacts, such as remnants of the original hair or wrong placement of the", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54f7e8d6d84270eb96811594f93d8920ee9993f23871d939a8afb48c3990d428.jpg", + "image_caption": [ + "LOHO Barbershop StyleYourHair Ours", + "Figure 3. Comparison to current state-of-the-art methods for transferring hair from 1st column-top, to the face of 1st column-bottom. Our method can accurately transfer hairstyle even when the input face and hair are misaligned (3rd, 4th row). It can also hallucinate missing details such as a forehead that was previously occluded (1st row), or shorten the hair (2nd row). In general, LOHO and Barbershop struggle with pose misalignment, while StyleYourHair struggles to preserve the input's face shape and hair details." + ], + "image_footnote": [], + "bbox": [ + 501, + 103, + 890, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "target hair. Barbershop performs well in preserving hair texture when the poses are similar. However, it falls short when handling challenging cases, such as transitioning from a long to a short hairstyle or removing bangs.", + "bbox": [ + 496, + 564, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "StyleYourHair can produce more realistic results in unaligned cases than LOHO and Barbershop. However, the hair details often look different from the reference hairstyle (row 5). We also observe that StyleYourHair may perform poorly when the pose difference becomes too large (row 3), as also shown in our user study for pose in Table 1. In contrast, our method can transfer hairstyles convincingly, regardless of the misalignment, while still preserving the original face shape and hair details. We refer to Appendix F for more results.", + "bbox": [ + 496, + 625, + 893, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. User Study", + "text_level": 1, + "bbox": [ + 500, + 771, + 622, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results can be misleading and biased, so we conducted a user study using Amazon Mechanical Turk on hairstyle transfer results using randomly selected pairs grouped into various difficulty levels. We compare results of our method with the current state of the art: StyleYourHair, Barbershop, and LOHO. Each participant was shown an input face, marked as 'Face', and a reference hair, marked", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e3bc4fc796cf86a1b682b079a9ab73667e966bcc2efe2f2404ad9234d9873630.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test DatasetsFFHQ Scenario Breakdown
AllFFHQ-PFFHQ-SEasyMediumDifficult
-✓---✓✓✓✓✓---
--✓---✓--✓✓-
---✓---✓-✓-✓
----✓---✓-✓✓
LOHO8.89.68.412.05.34.013.38.05.35.34.010.75.313.314.7
Barbershop16.413.817.817.321.313.321.318.717.317.320.012.018.714.721.3
StyleYourHair18.512.421.625.321.310.722.733.310.724.028.014.732.014.721.3
Ours56.264.252.245.352.072.042.740.066.753.348.062.744.057.342.7
", + "bbox": [ + 81, + 88, + 893, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as 'Hair', and asked to pick only one output that best accomplishes the task of transferring the hairstyle from image 'Hair' to the person in image 'Face'. The output row consists of four images from each method in random order. Each task was evaluated by 3 different participants. All images were in 256x256 resolution.", + "bbox": [ + 75, + 340, + 473, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Datasets", + "text_level": 1, + "bbox": [ + 76, + 465, + 191, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the user study, we construct two challenging benchmarks: FFHQ-P and FFHQ-S, from the test set of Flickr-Faces-HQ dataset (FFHQ) [18].", + "bbox": [ + 75, + 494, + 470, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FFHQ-P contains 150 random input face-hair pairs from FFHQ, covering yaw differences in ranges of $[0 - 15)$ , $[15 - 30)$ , ..., $[75 - 90)$ , with 25 pairs in each range.", + "bbox": [ + 75, + 544, + 470, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FFHQ-S contains 300 pairs, categorized into 12 different config with varying levels of difficulty (Table 1). Each config contains 25 random input pairs and is a combination of four possible scenarios (see details in Appendix B):", + "bbox": [ + 75, + 592, + 470, + 654 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pose Misalignment: When the yaw difference is between [15-30) or [30-45), a single checkmark or double checkmarks are used, respectively in Table 1.", + "- Needs Face Inpainting: This includes scenarios that require hallucinating parts of the original face, e.g., inpainting the forehead to remove bangs. This is challenging because the identity can easily change from the hallucination. We detect such scenarios based on the face/hair regions in $I_{f}$ and $I_{h}$ (Appendix B).", + "- Needs BG Inpainting: This includes scenarios where the hair shape becomes smaller and requires background inpainting. We detect such scenarios automatically based on the hair regions in $I_{f}$ and $I_{h}$ (Appendix B).", + "- $I_h$ Contains Hat: This represents scenarios where the hair reference is not fully visible in $I_h$ . We detect such scenarios based on the hat region in $I_h$ ." + ], + "bbox": [ + 94, + 656, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f43cd28a010c7ef423bdc8b0d2321181046c263bd5bd414b1e495f95fffde2ed.jpg", + "table_caption": [ + "Table 1. User study results on hairstyle transfer (percentage of user preferring each method). Our method outperforms state-of-the-art hairstyle transfer methods on FFHQ datasets in all challenging scenarios. A total of 450 pairs are used in this study, 150 pairs in FFHQ-P and 300 in FFHQ-S. For each pair, we asked 3 unique participants to select the best result for hairstyle transfer." + ], + "table_footnote": [], + "table_body": "
Pose Difference Range (FFHQ-P)
[0,15)[15,30)[30,45)[45,60)[60,75)[75,90)
LOHO10.75.310.712.08.010.7
Barbershop24.013.312.08.010.714.7
StyleYourHair17.316.012.06.714.78.0
Ours48.065.365.373.366.766.7
", + "bbox": [ + 504, + 339, + 893, + 434 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca198d1b7ad9600198b18122dc8f7b81c4b982c27a6fcb2a634ac86b7d68426f.jpg", + "table_caption": [ + "Table 2. User study on pose-invariant hairstyle transfer. Our method outperforms others on all pose difference ranges." + ], + "table_footnote": [], + "table_body": "
Hair Reconstruction MetricsFace Shape
PSNR ↑SSIM ↑LPIPS ↓FID ↓RMSE ↓
LOHO25.760.860.0710.4015.84
Barbershop29.180.890.0510.4617.25
StyleYourHair26.890.870.0910.9314.37
Ours27.840.880.0710.8510.89
", + "bbox": [ + 504, + 481, + 892, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Hair reconstruction results in the self-transfer experiment (Section 4.3), and RMSEs between facial landmarks detected on the input and output images.(Section 4.4).", + "bbox": [ + 498, + 582, + 893, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Results", + "text_level": 1, + "bbox": [ + 500, + 645, + 606, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using all test pairs, the participants preferred our results $56.2\\%$ of the time, whereas StyleYourHair's, Barbershop's, and LOHO's results were selected for $18.5\\%$ , $16.4\\%$ and $8.8\\%$ , respectively. Our method perform the best in both FFHQ-P $(64.2\\%)$ and FFHQ-S $(52.2\\%)$ , and in all configurations in Table 1-2.", + "bbox": [ + 496, + 670, + 893, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quality of Hair Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 770, + 777, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following MichiGAN and LOHO, we perform a self-transfer experiment where we set both the input face and input reference hair to be the same and evaluate the reconstruction accuracy using various metrics: PSNR, SSIM, IP-IPS, FID [14]. For every method, the hair region of the output will be blended back to the input image to ensure that the difference in score only comes from the hair.", + "bbox": [ + 496, + 794, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8615", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea698b7ad3b32866c78395879f28b92d81bedb8aeb868b081837daf97ab6eb0a.jpg", + "image_caption": [ + "Figure 4. Ablation study of optimization stages and pre-masking. These ablated versions of our pipeline may produce results that (A) look realistic, but fail to match the $I_{h}$ 's hairstyle; (B, C, D) have unlikely hair shapes or structures; (E) have sharp boundaries; and (F) contain an incomplete background." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 893, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The scores of our method, LOHO, Barbershop, and StyleYourHair are reported in Table 3. Unsurprisingly, when there is no misalignment between input face and hair, Barbershop generally achieves excellent hair reconstruction quality. Like ours, StyleYourHair is specifically designed to be pose-invariant, but their hair reconstruction quality seems to suffer greatly, as shown in our comparison (row 5 of Figure 3). Our method achieves better performance at hair reconstruction than StyleYourHair, but also suffers from similar drawbacks.", + "bbox": [ + 75, + 323, + 473, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. How Well Is Face Shape Preserved?", + "text_level": 1, + "bbox": [ + 76, + 467, + 387, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also propose a novel evaluation metric for hairstyle transfer that focuses on the ability to preserve facial shape of the original person. We accomplish this by comparing detected keypoints of the input face $I_{f}$ and those on the output using a simple Root-Mean-Square Error (RMSE). These keypoints were detected using an off-the-shelf library Dlib [22], and we only used keypoints on the facial contour (ID 0-16) for this evaluation.", + "bbox": [ + 75, + 491, + 470, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For this evaluation, we randomize additional 1,550 FFHQ test pairs into our FFHQ-P and FFHQ-S datasets (2,000 images in total). Table 3 shows that our method outperforms others with an RMSE of 10.89, which is $24.2\\%$ lower than StyleYourHair's (the second best), $36.9\\%$ lower than Barbershop's and $31.3\\%$ lower than LOHO's.", + "bbox": [ + 75, + 612, + 470, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 710, + 241, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we assess the importance of each component in our pipeline. The results are shown in Figure 4. We test 4 ablation configurations: i) by using only $I_{\\mathrm{guide}}^{\\mathrm{face}}$ from a single view, ii) without using the hallucination stage (Section 3.3), iii) without using the latent sharing structure (Section 3.2.2), and iv) without using pre-masking in $\\mathcal{L}_{\\mathrm{per}}$ (Section 3.2.1).", + "bbox": [ + 75, + 734, + 468, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimizing with a single view guide $I_{\\mathrm{guide}}^{\\mathrm{face}}$ (Config i) yields inaccurate hairstyles (Figure 4-A). Without optimization in W (Config ii), the method produces unrealistic hair results with various artifacts, e.g., sharp edges, unnatural hair shapes (Figure 4-B) or structures (Figure 4-C). This is", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/729380b90aa568c3d848755e24b9051b8719b8d4f73e847ef78a1e113c13e20c.jpg", + "image_caption": [ + "Figure 5. Failure cases; (A) the semantic regions are incorrect; (B) the facial keypoint is incorrect; (C) the EG3D projection is incorrect; (D) the reference hair color is too similar to the background color; (E) lighting looks unnatural; (F) eccentric hairstyle." + ], + "image_footnote": [], + "bbox": [ + 501, + 320, + 629, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e400c49cc39ffac05055b02a5ea0fa5ddf4be1fd2187e6a0ffe2c5c562fcbc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 321, + 759, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f24128bd84b12ef1ccfb0e41c4a40c9b326905223e08c6503fb302d87eacc44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 321, + 890, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "because the second stage, which has higher image fitting capability, tries to fit the initial rough estimation $I_{\\mathrm{guide}}$ . Without the latent sharing structure, the hair detail cannot be shared from the reference hair view ( $\\mathcal{L}_{\\mathrm{sim}}$ alone is not strong enough to force consistency) (Figure 4-A), resulting in inaccurate hair colors or structures (Figure 4-C).", + "bbox": [ + 496, + 541, + 893, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Without pre-masking in $\\mathcal{L}_{\\mathrm{per}}$ , the boundary of the face and hair regions is forced to be the same as in $I_{\\mathrm{guide}}$ , leading to visible and sharp seams between the face and hair or between different facial features. Here the optimizer fails to refine the face-hair boundary to make the results look natural (Figure 4-D), and the original details, such as the background from $I_{f}$ cannot be seamlessly blended (Figure 4-E).", + "bbox": [ + 496, + 632, + 893, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 753, + 756, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a flexible and effective hairstyle transfer system that can handle a variety of challenging in-the-wild scenarios and produce perceptually convincing hairstyle transfer results. Our user study shows that human evaluators prefer our results over previous methods across all tested scenarios. Nevertheless, our method can still fail in certain scenarios, for example, when the face and hair are too eccentric (Figure 5). We refer to Appendix G for more details.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8616", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE international conference on computer vision, pages 4432-4441, 2019. 2", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 4", + "[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 3", + "[4] Ironhammer america. File:elizabeth-olsen-1632123202.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth-olsen-1632123202.jpg/, 2022.17", + "[5] Peter Baylies. Stylegan-encoder @ONLINE. https://github.com/pbaylies/stylegan-encoder/, 2019.2", + "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3, 11", + "[7] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2, 5", + "[8] Chaeyeon Chung, Taewoo Kim, Hyelin Nam, Seunghwan Choi, Gyojung Gu, Sunghyun Park, and Jaegul Choo. Hairfit: Pose-invariant hairstyle transfer via flow-based hair alignment and semantic-region-aware inpainting. arXiv preprint arXiv:2206.08585, 2022. 2", + "[9] Gordon Correll. File: julianne moore (15011443428) (2). jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_(15011443428)_ (2). jpg/, 2014. 17", + "[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7865-7874, 2022. 16", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. 1, 2", + "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2", + "[13] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only chang-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 2", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Proc. NeurIPS, 2020. 2, 5", + "[16] jerryhouuu. Face-yaw-roll-pitch-from-posedestimation-using-opencv @ONLINE. https://github.com/jerryhouuu/Face-Yaw-Roll-Pitch-from-Pose-Estimation-using-OpenCV/, 2018.13", + "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2", + "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5, 7", + "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 4, 5, 12, 13", + "[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21", + "[21] Taewoo Kim, Chaeyeon Chung, Sunghyun Park, Gyojung Gu, Keonmin Nam, Wonzo Choe, Jaesung Lee, and Jaegul Choo. K-hairstyle: A large-scale korean hairstyle dataset for virtual hair editing and hairstyle classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1299-1303. IEEE, 2021. 2", + "[22] Davis E. King. Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research, 10:1755–1758, 2009. 3, 8, 11", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 13", + "[24] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 11", + "[25] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2445, 2020. 2, 4", + "[26] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 2", + "[27] nicolas genin. File:julianne moore 2009 venice film festival.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_2009_Venice_Film_Festival.jpg/, 2009.17" + ], + "bbox": [ + 503, + 93, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 13", + "[29] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 6, 11, 12, 13", + "[30] Rohit Saha, Brendan Duke, Florian Shkurti, Graham Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In CVPR, 2021. 1, 2, 4, 5, 6, 19, 20, 21", + "[31] David Shankbone. File:dwayne johnson at the 2009 tribeca film festival.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Dwayne_Johnson_at_the_2009_Tribeca_Film_Festival.jpg/, 2009.17", + "[32] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 2", + "[33] Gage Skidmore. Tom holland speaking at the 2016 san diego comic-con international in san diego, california. @ONLINE. https://commons.wikimedia.org/wiki/File: Tom_Holland_by_Gage_Skidmore.jpg/, 2016.1", + "[34] Gage Skidmore. File:elizabeth olsen by gage skidmore 2.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth_Olsen_by_Gage_Skidmore_2.jpg/, 2019.17", + "[35] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. arXiv preprint arXiv:2010.16417, 2020. 1, 2", + "[36] tenasia10. File:blackpinkLisa gmp 240622.png trends @ONLINE. https://commons.wikimedia.org/wiki/File:Blackpink_Lisa_GMP_240622.png/, 2022.17", + "[37] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3", + "[38] GABI THORNE. Jennifer aniston reacts to tiktok's best (and most absurd) trends @ONLINE. https://www.allure.com/story/jennifer-aniston-reacts-to-tiktok-trends/, 2022.17", + "[39] Albert Truuväärt. File:re kaja kallas (cropped).jpg @ONLINE. https://commons.wikimedia.org/wiki/File:RE_Kaja_Kallas_(cropped). jpg/, 2011.17", + "[40] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 5914–5922, 2019. 2", + "[41] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting gans with consecutive images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13910-13918, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. 3, 11", + "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4", + "[44] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proc. ECCV, 2022. 2", + "[45] Tao Zhou, Changqing Zhang, Chen Gong, Harish Bhaskar, and Jie Yang. Multiview latent space learning with feature redundancy minimization. IEEE transactions on cybernetics, 50(4):1655-1668, 2018. 2", + "[46] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. arXiv preprint arXiv:2106.01505, 2021. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21", + "[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 651-667. Springer, 2022. 3", + "[48] zllrunnings. face-parsing.pytorch @ONLINE. https: //github.com/zllrunning/face-parsing. PyTorch/, 2019.11" + ], + "bbox": [ + 501, + 92, + 893, + 487 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_model.json b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a5071c5496e50c09ee2d21bc49d0388ba86b78a7 --- /dev/null +++ b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.842, + 0.177 + ], + "angle": 0, + "content": "StyleGAN Salon: Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.212, + 0.903, + 0.231 + ], + "angle": 0, + "content": "Sasikarn Khwanmuang\\* Pakkapon Phongthawee\\* Patsorn Sangkloy\\* Supasorn Suwajanakorn" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.234, + 0.77, + 0.253 + ], + "angle": 0, + "content": "VISTEC, Thailand Phranakhon Rajabhat University, Thailand" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.259, + 0.851, + 0.273 + ], + "angle": 0, + "content": "{sasikarn.k_s18, pakkapon.p_s19, supasorn.s}@vistec.ac.th patsn.s@pnru.ac.th" + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.384, + 0.172, + 0.396 + ], + "angle": 0, + "content": "Input Face" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.306, + 0.892, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.49, + 0.877, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Our method can transfer the hairstyle from any reference hair image in the top row to Tom Holland [33], in the second row." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.516, + 0.314, + 0.531 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.474, + 0.884 + ], + "angle": 0, + "content": "Our paper seeks to transfer the hairstyle of a reference image to an input photo for virtual hair try-on. We target a variety of challenges scenarios, such as transforming a long hairstyle with bangs to a pixie cut, which requires removing the existing hair and inferring how the forehead would look, or transferring partially visible hair from a hat-wearing person in a different pose. Past solutions leverage StyleGAN for hallucinating any missing parts and producing a seamless face-hair composite through so-called GAN inversion or projection. However, there remains a challenge in controlling the hallucinations to accurately transfer hairstyle and preserve the face shape and identity of the input. To overcome this, we propose a multi-view optimization framework that uses two different views of reference composites to semantically guide occluded or ambiguous regions. Our optimization shares information between two poses, which allows us to produce high fidelity and realistic results from incomplete references. Our framework produces high-quality results and outperforms prior work in a user study that consists of significantly more challenging hair transfer scenarios than previously studied. Project page: https://stylegan-salon.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.516, + 0.631, + 0.531 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.637 + ], + "angle": 0, + "content": "What makes Jennifer Aniston keep her same hairstyle for over three decades? Perhaps she likes the classic, or perhaps changing her hairstyle is a decision too high-stakes that she could later regret. Unlike garments or makeup, trying on a new hairstyle is not easy, and being able to imagine yourself in different hairstyles could be an indispensable tool." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.897, + 0.837 + ], + "angle": 0, + "content": "Recent approaches for hairstyle transfer, StyleYourHair [20], Barbershop [46], LOHO [30], and Michiganan [35], allow users to manipulate multiple hair attributes of an input image, such as appearance, shape, or color by providing a reference image for each different attribute. These methods [20, 30, 46] rely on a generative adversarial network [11], specifically StyleGAN2 [19], which can synthesize highly realistic face images. Their key idea, which also forms the basis of our method, is to leverage the realistic face distribution learned by StyleGAN and search for a hairstyle-transfer output whose latent code lies within the learned distribution using optimization (commonly known as GAN projection or inversion)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our extensive study on these state-of-the-art techniques still reveal several unsolved challenges for in-the-wild hairstyle transfer. One of the main challenges is when the reference hair comes from a person with a very different head" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "pose or facial shape. In this case, the transfer result often degrades significantly [30, 46]. HairFIT [8] and the recently proposed StyleYourHair [20] both attempt to solve this using an additional alignment step to align the pose of the target hair to the input face. In HairFIT [8], this is done explicitly via a flow-based warping module for hair segmentation, but this requires training on multi-view datasets [21, 26]. StyleYourHair [20] avoids this issue and also improves upon HairFIT's results by optimizing for an aligned pose within StyleGAN2's latent space using distances between the detected facial keypoints. While StyleYourHair can handle a certain degree of misalignment, it often struggles to preserve details of the reference hair texture, especially for intricate and non-straight hairstyles, e.g., in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.426 + ], + "angle": 0, + "content": "In general, we have observed that there is a trade-off between hallucinating new details, which is crucial for handling highly different poses, and preserving the original texture from the reference images. These two goals are often at odds with each others. This trade-off is also evident in the results from StyleYourHair, as reported in [20] and as shown in table 3, where BarberShop [46] can still produce better results, for example, when the pose difference is small." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.577 + ], + "angle": 0, + "content": "We tackle this dilemma by performing a multi-stage optimization. This serves two purposes: first, to hallucinate new details necessary for aligning the poses, and second, to recover face-hair details from the original input images. We preserve these details in a form of two guide images in both viewpoints, which will be jointly optimized to allow new details to be filled while retaining face-hair texture in the original pose. Our pose alignment is done both explicitly via 3D projection [6] on RGB images, and implicitly via latent code(s) sharing during our multi-view optimization." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.578, + 0.397, + 0.593 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.604, + 0.47, + 0.667 + ], + "angle": 0, + "content": "1. We propose StyleGAN Salon: a pose-invariant hairstyle transfer pipeline that is flexible enough to handle a variety of challenging scenarios including, but not limited to, bangs/hat removal and background inpainting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.678, + 0.47, + 0.739 + ], + "angle": 0, + "content": "2. Unlike previous works, our method operates entirely on RGB images, which are more flexible than segmentation masks. This allows us to first draft the output and then refine them via multi-stage optimization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.751, + 0.469, + 0.828 + ], + "angle": 0, + "content": "3. We introduce multi-view optimization for hairstyle transfer which incorporates 3D information to align the poses for both face and hair images. Our method leverages both views to help preserve details from the original images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "4. We thoroughly analyze the results in several experiments, including a user study with detailed breakdown into various challenging scenarios. Our method shows superior results over existing works in all scenarios." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.604, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.237 + ], + "angle": 0, + "content": "Generative Adversarial Networks. Beginning with [11, 18], StyleGANs [17-19] have shown great success in 2D image generation by learning realistic training data distribution corresponding to a fixed low-dimensional distribution, called latent code. The learned latent code can be applied to a downstream task such as image manipulation [13, 15, 32, 40]. Recent work on [6, 12, 44] also explore 3D aware architecture in StyleGAN, resulting in multi-view consistent images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.895, + 0.327 + ], + "angle": 0, + "content": "For hairstyle transfers, MichiGAN [35] uses a conditional hair generation network that can control hair shape, structure, and appearance. Recently, HairFIT [8], a pose-invariant hairstyle transfer network, aligns the reference hair to match the input face pose using a optical flow-based hair alignment module, but requires training on multi-view dataset [21, 26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.895, + 0.448 + ], + "angle": 0, + "content": "StyleGAN Latent Space and Projection Techniques. To generate an image, StyleGAN2 first maps a random latent code \\( z \\sim N(0, I) \\) to an intermediate latent code \\( w \\in \\mathbb{R}^{512} \\) in a new space called \\( \\mathcal{W} \\). This code \\( w \\) is then replicated (18x) and used as input to each layer of the generator that controls details at different scales through weight demodulation. Collectively, these replicated latent codes are referred to as \\( w^{+} \\in \\mathbb{R}^{18 \\times 512} \\) in the extended latent space \\( \\mathcal{W}^{+} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.894, + 0.524 + ], + "angle": 0, + "content": "Pivotal Tuning Inversion (PTI) [29] further improves the projection quality by directly tuning generator weights around the optimized latent code \\( w \\) using their regularization technique. Other techniques [1,2,5] optimize the latent code \\( w^{+} \\) in each layer separately to better match the input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.659 + ], + "angle": 0, + "content": "PIE [37] introduces hierarchical optimization for semantic attributes editing that first optimizes the latent code in the \\(\\mathcal{W}\\) space, then transfers the code to the \\(\\mathcal{W}^+\\) space and continues the optimization in that space. We adopt similar hierarchical optimization that uses both \\(\\mathcal{W}\\) and \\(\\mathcal{W}^+\\). Part of our method is also inspired by PULSE [25], which reconstructs a high-resolution image from a small reference image (32x32) by searching for the closest image in the StyleGAN latent space that resembles the low-resolution reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.735 + ], + "angle": 0, + "content": "Some methods [7, 41, 45] project multiple images into latent space simultaneously. However, all of their inputs are complete, whereas our method requires hair information from the reference hair image and additional information from the input face image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "StyleGAN-Based Hairstyle Transfer. LOHO [30] adopts loss functions from Image2StyleGAN++ [2] to combine face-hair inputs into a hairstyle transfer output with their orthogonalization optimization technique which reduces conflicts between multiple loss functions. Barbershop [46] first predicts semantic regions of both inputs and uses them to create a \"target segmentation mask\" of the output with rule (in the original paper) or GAN-based inpainting (in their official code). Then, they optimize two separate latent codes in \\(\\mathcal{W}^+\\) space, one for matching the face and the other for hair, while conforming to the target segmentation mask. To" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8610" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "preserve the original details, the latent optimization is done in their proposed \\(\\mathcal{F} / \\mathcal{S}\\) space, which replaces the first seven blocks of \\(\\mathcal{W}^+\\) space with the corresponding activation maps of StyleGAN's convolution layers. To improve Barbershop's capacity to handle unaligned input poses, StyleYourHair [20] first aligns the reference hair to match the pose of the input face via the proposed local-style-matching loss. However, this alignment often leads to an unrealistic hair shape or inaccurate hair texture results. In contrast, our method simultaneously optimizes for two guides in face and hair views, resulting in a better hair texture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.394 + ], + "angle": 0, + "content": "Instead of estimating the final output from the segmentation mask [20, 46], our multi-view optimization uses face-hair composites in RGB space to overcome this problem, and produces results that better preserve the input facial structure and hairstyle across a wider, more challenging range of scenarios. Concurrent work, HairNet [47], arrives at a similar goal of removing the target segmentation mask via a two-step process that involves baldification using StyleFlow [3] followed by another network to transfer the hairstyle." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Nevertheless, handling pose differences is crucial for successful hairstyle transfer; our method incorporates 3D rotation to preserve the geometric consistency and systematically evaluate this aspect." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.24, + 0.485 + ], + "angle": 0, + "content": "3. StyleGAN Salon" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.472, + 0.554 + ], + "angle": 0, + "content": "Given two input images \\( I_h \\) and \\( I_f \\), the goal is to transfer the hair from the reference hair image \\( I_h \\) into the face in the image \\( I_f \\), while preserving all remaining details in \\( I_f \\), including identity, clothing, and background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.645 + ], + "angle": 0, + "content": "The key idea of our approach is to guide the optimization on the learned latent space of StyleGAN2 with two \"guide\" images, which represent rough composites of the final output based on simple cut-and-paste in the viewpoint of \\( I_{h} \\) and \\( I_{f} \\). We leverage EG3D to construct these guided images in a geometrically consistent way, described in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.84 + ], + "angle": 0, + "content": "Optimization on StyleGAN2's latent space is commonly performed on either the original latent space \\(\\mathcal{W}\\) or the extended latent space \\(\\mathcal{W}^+\\). Optimizing on \\(\\mathcal{W}\\) space generally leads to more realistic results by staying within the original latent space, whereas optimizing on \\(\\mathcal{W}^+\\) allows a closer match to the reference [37]. In hairstyle transfer, it is important to stay faithful to the input images and preserve important details such as hair texture, face identity, and background. However, optimizing on \\(\\mathcal{W}^+\\) will lead to poor results because our guide images are rough and unrealistic composites. Thus, we propose to optimize on \\(\\mathcal{W}\\) space followed by \\(\\mathcal{W}^+\\) space, similar to a technique in PIE [37] used for editing semantic attributes of an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Our optimization incorporates both guide images from the two viewpoints, detailed in Section 3.2. Section 3.3 and Section 3.4 cover details of our optimization on \\(\\mathcal{W}\\) and \\(\\mathcal{W}^+\\), respectively. Finally, we also optimize StyleGAN2 weights" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "while freezing the latent codes (Section 3.5) using PTI [29] to further improve detail fidelity. Figure 2 shows an overview of our complete pipeline." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.146, + 0.779, + 0.163 + ], + "angle": 0, + "content": "3.1. Constructing the Guide Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.895, + 0.305 + ], + "angle": 0, + "content": "The purpose of our guide images is to provide an initial estimate of how the hair would look on \\( I_{f} \\). We achieve this using a cut-and-paste composite of the face and background from \\( I_{f} \\) and the hair from \\( I_{h} \\). To better handle a potentially large shift in viewpoint between \\( I_{f} \\) and \\( I_{h} \\), we propose to leverage EG3D [6] to help generate geometrically consistent guide images. We argue that using multi-view guide images, each in the pose of \\( I_{f} \\) and \\( I_{h} \\), helps preserve details that could otherwise be lost from using a single viewpoint alone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.895, + 0.44 + ], + "angle": 0, + "content": "A straightforward approach is to simply project \\( I_h \\) into the EG3D [6] latent space, and use their proposed neural rendering pipeline to render it in the view of \\( I_f \\) (and vice versa). However, while the projection can produce a geometrically consistent 3D shape, we found that the resulting texture is not very accurate. To address this, we replace any visible regions in the texture from \\( I_h \\)'s viewpoint with the original pixels of \\( I_h \\), while leaving the rest of the texture as the projected texture from EG3D's rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.895, + 0.546 + ], + "angle": 0, + "content": "We also additionally apply uniform scaling and translation to match the faces' widths and centers, which are computed based on detected facial keypoints [22]. These keypoints, along with semantic regions from [42], are also used to handle various corner cases, such as re-painting unwanted hair regions. We refer to Appendix A for more details of this operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.637 + ], + "angle": 0, + "content": "The entire process is done for both \\( I_{f} \\) and \\( I_{h} \\) viewpoints, resulting in a pair of guide images \\( I_{\\mathrm{guide}} = [I_{\\mathrm{guide}}^{\\mathrm{face}}, I_{\\mathrm{guide}}^{\\mathrm{hair}}] \\). We emphasize that by leveraging our multi-view optimization (Section 3.2), these guide images do not need to be precise or realistic to produce convincing final results, as demonstrated in the first row of Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.645, + 0.785, + 0.661 + ], + "angle": 0, + "content": "3.2. Multi-View Latent Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.895, + 0.852 + ], + "angle": 0, + "content": "Our guide images provide complementary information about the target hair and face, albeit from different poses. However, each guide image is only fully accurate in regions with the original pixels seen in the original viewpoint and not warped by EG3D. Thus, our optimization goal is to combine information from both guide images to generate a final output that accurately captures the realistic hair from \\( I_{\\mathrm{guide}}^{\\mathrm{hair}} \\), as well as other details from \\( I_{\\mathrm{guide}}^{\\mathrm{face}} \\). We achieve this using multiple loss functions that attend to both viewpoints with different spatial emphasis. Specifically, we perform multi-view latent optimization on \\( w / w^{+} \\) (and stochastic noise maps \\( n \\)) that fits both guide images:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.861, + 0.894, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\{w / w ^ {+}, n \\}} \\sum_ {\\mathrm {i}} ^ {[ \\text {f a c e , h a i r} ]} \\mathcal {L} _ {\\text {l o s s}} \\left(O ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})}\\right), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.287 + ], + "angle": 0, + "content": "Figure 2. Overview of StyleGAN Salon: We first align the input face \\( I_{f} \\) and reference hair \\( I_{h} \\) and use them to construct guide images \\( I_{\\mathrm{guide}} \\), in two different viewpoints, which specifies the target appearance for each output region. (see Section 3)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.358 + ], + "angle": 0, + "content": "where \\(O^{(\\mathrm{i})}\\) denotes the output of the StyleGAN2's generator that takes a latent code \\(w^{(\\mathrm{i})}\\) or \\(w^{+\\mathrm{(i)}}\\) and a stochastic noise input \\(n^{(\\mathrm{i})}\\). \\(\\mathcal{L}_{\\mathrm{loss}}\\) is a sum of all our loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.368, + 0.47, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {l o s s}} = \\sum_ {\\mathrm {j}} ^ {[ \\mathrm {f}, \\mathrm {h}, \\mathrm {b g} ]} \\lambda_ {\\mathrm {p}} ^ {(\\mathrm {j})} \\mathcal {L} _ {\\text {p e r}} ^ {(\\mathrm {j})} (2) \\\\ + \\lambda_ {\\mathrm {g}} \\mathcal {L} _ {\\text {g l o b a l}} + \\lambda_ {\\mathrm {i}} \\mathcal {L} _ {\\text {i n i}} + \\lambda_ {\\varepsilon} \\mathcal {L} _ {\\varepsilon} + \\lambda_ {s} \\mathcal {L} _ {\\text {s i m}}, (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.473 + ], + "angle": 0, + "content": "where \\(\\lambda_{(\\cdot)}\\) are balancing weights, and [f, h, bg] refer to face, hair, and background. We next explain each loss function." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.492, + 0.236, + 0.505 + ], + "angle": 0, + "content": "3.2.1 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.605 + ], + "angle": 0, + "content": "The objective of our \\(\\mathcal{W}\\) and \\(\\mathcal{W}^+\\) optimization is to generate an output image that fits the corresponding \\(I_{\\mathrm{guide}}\\) without appearing unrealistic. The main challenge is that our guide images are unrealistic and can contain various unwanted artifacts. We design our loss functions to be tolerant of the imprecise nature of our guide images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.471, + 0.833 + ], + "angle": 0, + "content": "Masked Perceptual Loss: This loss is based on Learned Perceptual Image Patch Similarity (LPIPS) [43], which compares two input images in a deep feature space. Similar to [20, 30, 46], we apply a binary mask to indicate regions of interest in the deep feature space of the two input images. We use \\( L_{\\mathrm{LPIPS}}(I_1, I_2; M) \\) to denote this LPIPS computation between two images \\( (I_1, I_2) \\) with mask \\( M \\). However, we observe that applying this masking operation after the feature computation is insufficient to disregard unwanted regions. This is due to the fact that LPIPS is a patch-based similarity loss, and regions outside of the mask can still affect the loss computation. Based on this principle, we propose to apply additional pre-masking to mask out ambiguous regions in \\( I_{\\mathrm{guide}} \\) that should not be trusted. Specifically, our Masked Perceptual Loss is the following:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.842, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e r}} ^ {\\mathrm {i}, \\mathrm {r o i}} = \\Lambda_ {\\mathrm {r o i}} ^ {(\\mathrm {i})} L _ {\\mathrm {L P I P S}} \\left(O ^ {(\\mathrm {i})} \\odot \\neg M _ {\\mathrm {r o n i}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})} \\odot \\neg M _ {\\mathrm {r o n i}} ^ {(\\mathrm {i})}; M _ {\\mathrm {r o i}} ^ {(\\mathrm {i})}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(M_{\\mathrm{roi}}\\) is the mask for region of interest (ROI), and \\(M_{\\mathrm{roni}}\\) is the mask for regions of not interest (RONI). \\((.)\\odot \\neg M_{\\mathrm{roni}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.469 + ], + "angle": 0, + "content": "is a simple element-wise multiplication, which effectively excludes these regions from the loss by setting them to 0 in both input images. And \\(\\Lambda_{\\mathrm{roi}}^{\\mathrm{(i)}}\\) is a balancing weight in (i) viewpoint, which is set higher when the region of interest covers the original pixels of \\(I_{f}\\) (or \\(I_{h}\\)). For instance, the loss that attends to face in \\(I_{\\mathrm{guide}}^{\\mathrm{face}}\\) has a higher weight than the face loss in \\(I_{\\mathrm{guide}}^{\\mathrm{hair}}\\). We consider 3 regions of interest: \\(L_{\\mathrm{per}}^{\\mathrm{f}}\\), \\(L_{\\mathrm{per}}^{\\mathrm{h}}\\), and \\(L_{\\mathrm{per}}^{\\mathrm{bg}}\\) for face, hair, and background regions. We refer to Section 3.3 and 3.4 for details on these mask generation for \\(\\mathcal{W}\\) and \\(\\mathcal{W}^{+}\\)optimization, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Global loss: This loss function attempts to match the overall appearance of the output to \\( I_{\\mathrm{guide}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.508, + 0.892, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g l o b a l}} = L _ {\\text {M S E}} ^ {3 2} (O, I _ {\\text {g u i d e}}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.894, + 0.624 + ], + "angle": 0, + "content": "\\(L_{\\mathrm{MSE}}^{32}\\) is the mean square error computed on \\(32 \\times 32\\) downsampled input images. The goal of the downsampling is to reduce the effect of matching visible seams from higher resolution \\(I_{\\mathrm{guide}}\\) because those seams become imperceptible once downsampled. This loss is inspired by a similar idea used for image super-resolution by Menon et al. [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.685 + ], + "angle": 0, + "content": "Initial Value Loss: To prevent the output from deviating too far from the learned distribution and becoming unrealistic, we force the optimized latent code(s) \\( w \\) and \\( w^{+} \\) to be close to its initial value \\( w_{0} \\) through L2 loss, as used in [2]:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.694, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i n i}} = \\left\\| w _ {1: 1 8} - w _ {0} \\right\\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We use the estimated mean of \\(\\mathcal{W}\\) latent space as initial values for optimizing \\(w\\). This is computed by averaging many latent codes \\(w_{i} = f(z_{i})\\) drawn randomly through \\(z\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and the StyleGAN's mapping network \\(f\\). Then the optimized latent code \\(w\\) from the first stage becomes the initial value for the optimization in \\(\\mathcal{W}^+\\) extended latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.899 + ], + "angle": 0, + "content": "Noise Regularization Loss: We also use the noise regularization loss \\((\\mathcal{L}_{\\varepsilon})\\) proposed in StyleGAN2 [19] to ensure that the noise maps will be used for capturing the stochastic variations, and not for encoding the content meant to be captured by the latent code. We refer to StyleGAN2 for the details of this loss." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8612" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "Latent Similarity Loss: To enable information sharing between the two guide images (\\(I_{\\mathrm{guide}}^{\\mathrm{face}}\\) and \\(I_{\\mathrm{guide}}^{\\mathrm{hair}}\\)) from two viewpoints, we force both their latent codes to be close together by using the following L2 loss, similarly to [7]." + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.163, + 0.471, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m}} = \\left\\| w ^ {(\\text {f a c e})} - w ^ {(\\text {h a i r})} \\right\\| _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.193, + 0.471, + 0.224 + ], + "angle": 0, + "content": "where \\(w\\) can be in \\(\\mathcal{W}\\) space during \\(\\mathcal{W}\\) optimization or replaced with \\(w^{+}\\) during \\(\\mathcal{W}^+\\) optimization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.243, + 0.278, + 0.259 + ], + "angle": 0, + "content": "3.2.2 Sharing Latent Code" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.47, + 0.463 + ], + "angle": 0, + "content": "In StyleGAN2, latent codes in early layers have been shown to correspond to high level concepts (e.g. human pose) whereas the remaining layers capture low-level information (e.g. colors) [15, 18]. This motivates the optimization in \\(\\mathcal{W}^+\\) extended space, as each layer can be optimized for different aspects of the image. Similarly, we optimize latent codes to match guide images of the same person and hairstyle but in different poses. Rather than optimizing two latent codes independently, we can enforce information sharing between the two poses by constraining the latent codes of the last few layers (\\(w_{l:18}\\)) to be the same. We set \\(l\\) to be 4 in all our experiments, loosely based on the head rotation experiments shown in GANSpace [15]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.464, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Sharing \\( w \\) Latent Code: Optimization in the space of StyleGAN2 latent code \\( \\mathcal{W} \\) is generally done as a single latent code \\( (\\mathbb{R}^{512}) \\), which are duplicated (18x) and fed into all the layers. In our case, we also optimize for a single latent code \\( w \\) but only duplicate them for the first \\( l \\) layers \\( (w_{1:l}) \\). The remaining layers \\( (w_{l:18}) \\) are interpolated from the current latent codes of each guide image with a coefficient \\( \\alpha \\)." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.58, + 0.47, + 0.6 + ], + "angle": 0, + "content": "\\[\nw _ {l: 1 8} = \\alpha w _ {1: l} ^ {\\text {(f a c e)}} + (1 - \\alpha) w _ {1: l} ^ {\\text {(h a i r)}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.61, + 0.47, + 0.67 + ], + "angle": 0, + "content": "We randomize this coefficient to make the latent codes stay within the space of \\(\\mathcal{W}\\). This random interpolation forces the latent code for each guide image to be partly similar while still producing realistic results from \\(\\mathcal{W}\\) latent space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.671, + 0.471, + 0.762 + ], + "angle": 0, + "content": "Sharing \\( w^{+} \\)Latent Code: Optimization in \\( \\mathcal{W}^{+} \\)extended space is straightforward, as latent code for each layer can already be separately optimized. We simply share the last few layers \\( (w_{l:18}) \\) of both \\( w^{+} \\)latent codes during optimization. We refer to the Appendix A.5 for more details on the optimization process." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.458, + 0.788 + ], + "angle": 0, + "content": "3.3. \\(\\mathcal{W}\\) Optimization: Hallucinate Missing Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "This stage aims to hallucinate details not currently visible in the guide images. The output images of this stage are \\( O_{1}^{\\mathrm{face}} \\) and \\( O_{1}^{\\mathrm{hair}} \\), which are essentially the projection of the unrealistic \\( I_{\\mathrm{guide}}^{\\mathrm{face}} \\) and \\( I_{\\mathrm{guide}}^{\\mathrm{hair}} \\) into the real image distribution learned in StyleGAN2 [19]. This is done by optimizing on \\( W \\) latent space to fit each guide image with the objective function (Equation 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.894, + 0.198 + ], + "angle": 0, + "content": "While our guide images (\\(I_{\\mathrm{guide}}^{\\mathrm{face}}\\) and \\(I_{\\mathrm{guide}}^{\\mathrm{hair}}\\)) capture the overall appearance of the desired outputs, they still lack details in certain regions. These unknown regions may correspond to unseen facial features that were occluded by the hair in \\(I_{f}\\) or incomplete reference hair from \\(I_{h}\\) caused by, e.g., a hat or image cropping. These details will need to be hallucinated and seamlessly blended with the rest of the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.893, + 0.411 + ], + "angle": 0, + "content": "To accomplish this, we design our pre-masking in Masked Perceptual Loss \\((\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{f}},\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{h}}\\) , and \\(\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{bg}}\\) ; Equation 4) to include all the unknown regions that need to be hallucinated. Specifically, \\(M_{\\mathrm{roni}}^{\\mathrm{f}}\\) represents the face region occluded by the hair in \\(I_{f}\\) , e.g., the forehead behind the bangs or the ears that should become visible in the final output. Analogously, \\(M_{\\mathrm{roni}}^{\\mathrm{h}}\\) represents the hair region occluded by other objects (e.g., a hat) or not visible in \\(I_{h}\\) due to image cropping. We set \\(M_{\\mathrm{roi}}^{\\mathrm{f}}\\) and \\(M_{\\mathrm{roi}}^{\\mathrm{h}}\\) to be segmentation masks for face and hair regions \\((M_{\\mathrm{f}}\\) and \\(M_{\\mathrm{h}})\\) , respectively. Both \\(M_{\\mathrm{roi}}^{\\mathrm{bg}}\\) and \\(\\neg M_{\\mathrm{roni}}^{\\mathrm{bg}}\\) is set to be the background regions in \\(I_{f}\\) not covered by the transferred hair \\((M_{\\mathrm{bg}})\\) . These masks are constructed by composing different semantic regions (union, intersection, etc.) from \\(I_{f}\\) and \\(I_{h}\\) , detailed in Appendix A.3" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.894, + 0.516 + ], + "angle": 0, + "content": "We note that \\( M_{\\mathrm{roi}} \\) is applied on the deep feature maps, similar to [20, 30, 46], while \\( \\neg M_{\\mathrm{roni}} \\) is applied on raw RGB images. Using both masking techniques is crucial to our natural and seamless blending. The additional pre-masking allows the hair shape (or face shape) in the output to be different from \\( I_{\\mathrm{guide}} \\)'s and freely grow outward outside \\( M_{\\mathrm{roi}} \\) if this leads to a more natural result (see Figure 4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.524, + 0.884, + 0.54 + ], + "angle": 0, + "content": "3.4. \\(\\mathcal{W}^+\\) Optimization: Recover Face-Hair Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.894, + 0.713 + ], + "angle": 0, + "content": "The output of the previous stage, \\(O_1^{\\mathrm{face}}\\) and \\(O_1^{\\mathrm{hair}}\\), may still look different from the input person and not yet capture the hair details from the reference. This stage aims at refining these output images to better reproduce the hair details from \\(I_h\\) and the rest from \\(I_f\\). The optimization is done in the extended \\(\\mathcal{W}^+\\) space with respect to \\(\\{w^{+},n\\}\\). In other words, we allow \\(w\\in \\mathbb{R}^{512}\\) that was previously replicated to each layer to be optimized separately as \\(w^{+}\\in \\mathbb{R}^{18\\times 512}\\). We denote \\(O_2^{\\mathrm{face}}\\) and \\(O_2^{\\mathrm{hair}}\\) to be the output of this stage generated by StyleGAN2 from our optimized latent codes in \\(\\mathcal{W}^+\\) extended latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.893, + 0.822 + ], + "angle": 0, + "content": "We update the target images for the optimization from \\( I_{\\mathrm{guide}}^{\\mathrm{face}} \\) and \\( I_{\\mathrm{guide}}^{\\mathrm{hair}} \\) to the more complete versions based on \\( O_1^{\\mathrm{face}} \\) and \\( O_1^{\\mathrm{hair}} \\) from the first stage. However, because the first stage aims at hallucinating new details into the pre-masking regions, it may lose the original texture details. To address this, we replace the known regions in \\( O_1^{\\mathrm{face}} \\) and \\( O_1^{\\mathrm{hair}} \\) with the correct details from the original \\( I_{\\mathrm{guide}} \\) by setting" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.831, + 0.893, + 0.848 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {n e w} \\text {g u i d e}} \\leftarrow I _ {\\text {g u i d e}} \\odot M _ {\\mathrm {c}} + O _ {1} \\odot \\neg M _ {\\mathrm {c}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\( M_{c} = M_{\\mathrm{f}} \\cup M_{\\mathrm{h}} \\cup M_{\\mathrm{bg}} \\), which corresponds to the regions of interest in our Masked Perceptual Loss where we want to match with \\( I_{\\mathrm{guide}} \\). In short, we simply copy the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.121 + ], + "angle": 0, + "content": "hallucinated parts from \\(O_1^{\\mathrm{face}}\\) and \\(O_1^{\\mathrm{hair}}\\), and combine them with the part in \\(I_{\\mathrm{guide}}\\) that are known to be correct." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.474, + 0.31 + ], + "angle": 0, + "content": "For our Masked Perceptual Loss, we use the same masking procedure as the previous stage (Section 3.3). However, we observe that the projection from EG3D, which produces \\( I_{\\mathrm{guide}} \\), generally is inferior to the known details from the original images. For example, \\( I_{\\mathrm{guide}}^{\\mathrm{hair}} \\) contains more details on the hair than \\( I_{\\mathrm{guide}}^{\\mathrm{face}} \\) due to the latter being hallucinated from the 3D rotation by EG3D's rendering pipeline. To de-prioritize these inaccurate details, we blur both \\( I_{\\mathrm{guide}} \\) and \\( O \\) when computing the perceptual loss for the less accurate regions of interest. Concretely, we only blur images when computing \\( \\mathcal{L}_{\\mathrm{per}}^{\\mathrm{f}} \\) for \\( I_{\\mathrm{guide}}^{\\mathrm{hair}} \\) optimization and when computing \\( \\mathcal{L}_{\\mathrm{per}}^{\\mathrm{h}} \\) for \\( I_{\\mathrm{guide}}^{\\mathrm{face}} \\) optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.316, + 0.303, + 0.332 + ], + "angle": 0, + "content": "3.5. Pivotal Tuning Inversion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.446 + ], + "angle": 0, + "content": "The purpose of this stage is to further refine the output images \\((O_2^{\\mathrm{face}}\\) and \\(O_2^{\\mathrm{hair}})\\) by allowing the optimization of the StyleGAN's weights \\(\\theta\\) (while fixing the optimized latent code \\(w_{\\mathrm{optimized}}\\)). We closely follow the proposed optimization in PTI [29] with the same objective function. However, we update the reconstruction loss to reflect the goal of our task, which is to match the original details from \\(I_{f}\\) and \\(I_{h}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.455, + 0.47, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p t i}} = \\sum_ {\\mathrm {i}} ^ {[ \\text {f a c e}, \\text {h a r i} ]} \\left(L _ {\\mathrm {L P I P S}} \\left(O _ {\\text {t u n e}} ^ {(\\mathrm {i})} \\odot M _ {\\text {r a w}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})} \\odot M _ {\\text {r a w}} ^ {(\\mathrm {i})}; M _ {\\text {r a w}} ^ {(\\mathrm {i})}\\right) \\right. \\\\ + L _ {\\mathrm {M S E}} ^ {3 2} \\left(O _ {\\text {t u n e}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})}; \\neg M _ {\\text {r a w}} ^ {(\\mathrm {i})}\\right)), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.561 + ], + "angle": 0, + "content": "where \\(O_{\\mathrm{tune}}\\) is the generated image using the tuned weights \\(\\theta\\), the mask \\(M_{\\mathrm{raw}}^{\\mathrm{face}}\\) is \\((M_{\\mathrm{bg}} \\cup M_{\\mathrm{f}})\\), and \\(M_{\\mathrm{raw}}^{\\mathrm{hair}}\\) is \\(M_{\\mathrm{h}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.21, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.717 + ], + "angle": 0, + "content": "In this section, we compare our method to state-of-the-art StyleYourHair [20], Barbershop [46], and LOHO [30]. Our evaluation criteria are i) user preference via a user study on a wide variety of scenarios, ii) hairstyle transfer quality, iii) hair reconstruction quality, and iv) how well the input face shape is preserved. Section 4.5 presents ablation studies on our multi-view sharing latent optimization, optimization stage, and loss functions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We use the official code of LOHO [30] and Barbershop [46] with the default configurations. For StyleYourHair [20], we use the configuration where the hair reference is never flipped." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.3, + 0.804 + ], + "angle": 0, + "content": "4.1. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We provide a qualitative comparison to StyleYourHair [20], Barbershop [46], and LOHO [30] in Figure 3. We observe that LOHO and Barbershop often struggle to fit the reference hair accurately when the poses are not well aligned (rows 3-5), resulting in various artifacts, such as remnants of the original hair or wrong placement of the" + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.093, + 0.861, + 0.103 + ], + "angle": 0, + "content": "LOHO Barbershop StyleYourHair Ours" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.104, + 0.892, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.444, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 3. Comparison to current state-of-the-art methods for transferring hair from 1st column-top, to the face of 1st column-bottom. Our method can accurately transfer hairstyle even when the input face and hair are misaligned (3rd, 4th row). It can also hallucinate missing details such as a forehead that was previously occluded (1st row), or shorten the hair (2nd row). In general, LOHO and Barbershop struggle with pose misalignment, while StyleYourHair struggles to preserve the input's face shape and hair details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.626 + ], + "angle": 0, + "content": "target hair. Barbershop performs well in preserving hair texture when the poses are similar. However, it falls short when handling challenging cases, such as transitioning from a long to a short hairstyle or removing bangs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.894, + 0.762 + ], + "angle": 0, + "content": "StyleYourHair can produce more realistic results in unaligned cases than LOHO and Barbershop. However, the hair details often look different from the reference hairstyle (row 5). We also observe that StyleYourHair may perform poorly when the pose difference becomes too large (row 3), as also shown in our user study for pose in Table 1. In contrast, our method can transfer hairstyles convincingly, regardless of the misalignment, while still preserving the original face shape and hair details. We refer to Appendix F for more results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.623, + 0.789 + ], + "angle": 0, + "content": "4.2. User Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Qualitative results can be misleading and biased, so we conducted a user study using Amazon Mechanical Turk on hairstyle transfer results using randomly selected pairs grouped into various difficulty levels. We compare results of our method with the current state of the art: StyleYourHair, Barbershop, and LOHO. Each participant was shown an input face, marked as 'Face', and a reference hair, marked" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8614" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.263 + ], + "angle": 0, + "content": "
Test DatasetsFFHQ Scenario Breakdown
AllFFHQ-PFFHQ-SEasyMediumDifficult
-✓---✓✓✓✓✓---
--✓---✓--✓✓-
---✓---✓-✓-✓
----✓---✓-✓✓
LOHO8.89.68.412.05.34.013.38.05.35.34.010.75.313.314.7
Barbershop16.413.817.817.321.313.321.318.717.317.320.012.018.714.721.3
StyleYourHair18.512.421.625.321.310.722.733.310.724.028.014.732.014.721.3
Ours56.264.252.245.352.072.042.740.066.753.348.062.744.057.342.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.273, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Table 1. User study results on hairstyle transfer (percentage of user preferring each method). Our method outperforms state-of-the-art hairstyle transfer methods on FFHQ datasets in all challenging scenarios. A total of 450 pairs are used in this study, 150 pairs in FFHQ-P and 300 in FFHQ-S. For each pair, we asked 3 unique participants to select the best result for hairstyle transfer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.475, + 0.434 + ], + "angle": 0, + "content": "as 'Hair', and asked to pick only one output that best accomplishes the task of transferring the hairstyle from image 'Hair' to the person in image 'Face'. The output row consists of four images from each method in random order. Each task was evaluated by 3 different participants. All images were in 256x256 resolution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.192, + 0.482 + ], + "angle": 0, + "content": "4.2.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.471, + 0.543 + ], + "angle": 0, + "content": "For the user study, we construct two challenging benchmarks: FFHQ-P and FFHQ-S, from the test set of Flickr-Faces-HQ dataset (FFHQ) [18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.471, + 0.59 + ], + "angle": 0, + "content": "FFHQ-P contains 150 random input face-hair pairs from FFHQ, covering yaw differences in ranges of \\([0 - 15)\\), \\([15 - 30)\\), ..., \\([75 - 90)\\), with 25 pairs in each range." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.471, + 0.655 + ], + "angle": 0, + "content": "FFHQ-S contains 300 pairs, categorized into 12 different config with varying levels of difficulty (Table 1). Each config contains 25 random input pairs and is a combination of four possible scenarios (see details in Appendix B):" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.657, + 0.472, + 0.702 + ], + "angle": 0, + "content": "- Pose Misalignment: When the yaw difference is between [15-30) or [30-45), a single checkmark or double checkmarks are used, respectively in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.702, + 0.472, + 0.793 + ], + "angle": 0, + "content": "- Needs Face Inpainting: This includes scenarios that require hallucinating parts of the original face, e.g., inpainting the forehead to remove bangs. This is challenging because the identity can easily change from the hallucination. We detect such scenarios based on the face/hair regions in \\( I_{f} \\) and \\( I_{h} \\) (Appendix B)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.793, + 0.472, + 0.853 + ], + "angle": 0, + "content": "- Needs BG Inpainting: This includes scenarios where the hair shape becomes smaller and requires background inpainting. We detect such scenarios automatically based on the hair regions in \\(I_{f}\\) and \\(I_{h}\\) (Appendix B)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.853, + 0.472, + 0.899 + ], + "angle": 0, + "content": "- \\( I_h \\) Contains Hat: This represents scenarios where the hair reference is not fully visible in \\( I_h \\). We detect such scenarios based on the hat region in \\( I_h \\)." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.657, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.34, + 0.895, + 0.435 + ], + "angle": 0, + "content": "
Pose Difference Range (FFHQ-P)
[0,15)[15,30)[30,45)[45,60)[60,75)[75,90)
LOHO10.75.310.712.08.010.7
Barbershop24.013.312.08.010.714.7
StyleYourHair17.316.012.06.714.78.0
Ours48.065.365.373.366.766.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.444, + 0.894, + 0.473 + ], + "angle": 0, + "content": "Table 2. User study on pose-invariant hairstyle transfer. Our method outperforms others on all pose difference ranges." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.482, + 0.893, + 0.573 + ], + "angle": 0, + "content": "
Hair Reconstruction MetricsFace Shape
PSNR ↑SSIM ↑LPIPS ↓FID ↓RMSE ↓
LOHO25.760.860.0710.4015.84
Barbershop29.180.890.0510.4617.25
StyleYourHair26.890.870.0910.9314.37
Ours27.840.880.0710.8510.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.583, + 0.894, + 0.626 + ], + "angle": 0, + "content": "Table 3. Hair reconstruction results in the self-transfer experiment (Section 4.3), and RMSEs between facial landmarks detected on the input and output images.(Section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.607, + 0.66 + ], + "angle": 0, + "content": "4.2.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Using all test pairs, the participants preferred our results \\(56.2\\%\\) of the time, whereas StyleYourHair's, Barbershop's, and LOHO's results were selected for \\(18.5\\%\\), \\(16.4\\%\\) and \\(8.8\\%\\), respectively. Our method perform the best in both FFHQ-P \\((64.2\\%)\\) and FFHQ-S \\((52.2\\%)\\), and in all configurations in Table 1-2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.779, + 0.789 + ], + "angle": 0, + "content": "4.3. Quality of Hair Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Following MichiGAN and LOHO, we perform a self-transfer experiment where we set both the input face and input reference hair to be the same and evaluate the reconstruction accuracy using various metrics: PSNR, SSIM, IP-IPS, FID [14]. For every method, the hair region of the output will be blended back to the input image to ensure that the difference in score only comes from the hair." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.894, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Ablation study of optimization stages and pre-masking. These ablated versions of our pipeline may produce results that (A) look realistic, but fail to match the \\( I_{h} \\)'s hairstyle; (B, C, D) have unlikely hair shapes or structures; (E) have sharp boundaries; and (F) contain an incomplete background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.475, + 0.462 + ], + "angle": 0, + "content": "The scores of our method, LOHO, Barbershop, and StyleYourHair are reported in Table 3. Unsurprisingly, when there is no misalignment between input face and hair, Barbershop generally achieves excellent hair reconstruction quality. Like ours, StyleYourHair is specifically designed to be pose-invariant, but their hair reconstruction quality seems to suffer greatly, as shown in our comparison (row 5 of Figure 3). Our method achieves better performance at hair reconstruction than StyleYourHair, but also suffers from similar drawbacks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.388, + 0.485 + ], + "angle": 0, + "content": "4.4. How Well Is Face Shape Preserved?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.612 + ], + "angle": 0, + "content": "We also propose a novel evaluation metric for hairstyle transfer that focuses on the ability to preserve facial shape of the original person. We accomplish this by comparing detected keypoints of the input face \\( I_{f} \\) and those on the output using a simple Root-Mean-Square Error (RMSE). These keypoints were detected using an off-the-shelf library Dlib [22], and we only used keypoints on the facial contour (ID 0-16) for this evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.472, + 0.704 + ], + "angle": 0, + "content": "For this evaluation, we randomize additional 1,550 FFHQ test pairs into our FFHQ-P and FFHQ-S datasets (2,000 images in total). Table 3 shows that our method outperforms others with an RMSE of 10.89, which is \\(24.2\\%\\) lower than StyleYourHair's (the second best), \\(36.9\\%\\) lower than Barbershop's and \\(31.3\\%\\) lower than LOHO's." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.242, + 0.726 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.826 + ], + "angle": 0, + "content": "Here we assess the importance of each component in our pipeline. The results are shown in Figure 4. We test 4 ablation configurations: i) by using only \\( I_{\\mathrm{guide}}^{\\mathrm{face}} \\) from a single view, ii) without using the hallucination stage (Section 3.3), iii) without using the latent sharing structure (Section 3.2.2), and iv) without using pre-masking in \\( \\mathcal{L}_{\\mathrm{per}} \\) (Section 3.2.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Optimizing with a single view guide \\(I_{\\mathrm{guide}}^{\\mathrm{face}}\\) (Config i) yields inaccurate hairstyles (Figure 4-A). Without optimization in W (Config ii), the method produces unrealistic hair results with various artifacts, e.g., sharp edges, unnatural hair shapes (Figure 4-B) or structures (Figure 4-C). This is" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.321, + 0.63, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.322, + 0.761, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.322, + 0.892, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.457, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Figure 5. Failure cases; (A) the semantic regions are incorrect; (B) the facial keypoint is incorrect; (C) the EG3D projection is incorrect; (D) the reference hair color is too similar to the background color; (E) lighting looks unnatural; (F) eccentric hairstyle." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.632 + ], + "angle": 0, + "content": "because the second stage, which has higher image fitting capability, tries to fit the initial rough estimation \\( I_{\\mathrm{guide}} \\). Without the latent sharing structure, the hair detail cannot be shared from the reference hair view (\\( \\mathcal{L}_{\\mathrm{sim}} \\) alone is not strong enough to force consistency) (Figure 4-A), resulting in inaccurate hair colors or structures (Figure 4-C)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.895, + 0.74 + ], + "angle": 0, + "content": "Without pre-masking in \\(\\mathcal{L}_{\\mathrm{per}}\\), the boundary of the face and hair regions is forced to be the same as in \\(I_{\\mathrm{guide}}\\), leading to visible and sharp seams between the face and hair or between different facial features. Here the optimizer fails to refine the face-hair boundary to make the results look natural (Figure 4-D), and the original details, such as the background from \\(I_{f}\\) cannot be seamlessly blended (Figure 4-E)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.754, + 0.757, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We have presented a flexible and effective hairstyle transfer system that can handle a variety of challenging in-the-wild scenarios and produce perceptually convincing hairstyle transfer results. Our user study shows that human evaluators prefer our results over previous methods across all tested scenarios. Nevertheless, our method can still fail in certain scenarios, for example, when the face and hair are too eccentric (Figure 5). We refer to Appendix G for more details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8616" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.471, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE international conference on computer vision, pages 4432-4441, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.472, + 0.299 + ], + "angle": 0, + "content": "[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[4] Ironhammer america. File:elizabeth-olsen-1632123202.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth-olsen-1632123202.jpg/, 2022.17" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.472, + 0.399 + ], + "angle": 0, + "content": "[5] Peter Baylies. Stylegan-encoder @ONLINE. https://github.com/pbaylies/stylegan-encoder/, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[8] Chaeyeon Chung, Taewoo Kim, Hyelin Nam, Seunghwan Choi, Gyojung Gu, Sunghyun Park, and Jaegul Choo. Hairfit: Pose-invariant hairstyle transfer via flow-based hair alignment and semantic-region-aware inpainting. arXiv preprint arXiv:2206.08585, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[9] Gordon Correll. File: julianne moore (15011443428) (2). jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_(15011443428)_ (2). jpg/, 2014. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7865-7874, 2022. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only chang-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.893, + 0.119 + ], + "angle": 0, + "content": "ing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Proc. NeurIPS, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[16] jerryhouuu. Face-yaw-roll-pitch-from-posedestimation-using-opencv @ONLINE. https://github.com/jerryhouuu/Face-Yaw-Roll-Pitch-from-Pose-Estimation-using-OpenCV/, 2018.13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.893, + 0.343 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.893, + 0.399 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.893, + 0.468 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 4, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.893, + 0.537 + ], + "angle": 0, + "content": "[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.893, + 0.62 + ], + "angle": 0, + "content": "[21] Taewoo Kim, Chaeyeon Chung, Sunghyun Park, Gyojung Gu, Keonmin Nam, Wonzo Choe, Jaesung Lee, and Jaegul Choo. K-hairstyle: A large-scale korean hairstyle dataset for virtual hair editing and hairstyle classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1299-1303. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.893, + 0.65 + ], + "angle": 0, + "content": "[22] Davis E. King. Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research, 10:1755–1758, 2009. 3, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.893, + 0.69 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.893, + 0.734 + ], + "angle": 0, + "content": "[24] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[25] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2445, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[26] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] nicolas genin. File:julianne moore 2009 venice film festival.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_2009_Venice_Film_Festival.jpg/, 2009.17" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.162, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[29] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 6, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[30] Rohit Saha, Brendan Duke, Florian Shkurti, Graham Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In CVPR, 2021. 1, 2, 4, 5, 6, 19, 20, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[31] David Shankbone. File:dwayne johnson at the 2009 tribeca film festival.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Dwayne_Johnson_at_the_2009_Tribeca_Film_Festival.jpg/, 2009.17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[32] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[33] Gage Skidmore. Tom holland speaking at the 2016 san diego comic-con international in san diego, california. @ONLINE. https://commons.wikimedia.org/wiki/File: Tom_Holland_by_Gage_Skidmore.jpg/, 2016.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[34] Gage Skidmore. File:elizabeth olsen by gage skidmore 2.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth_Olsen_by_Gage_Skidmore_2.jpg/, 2019.17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[35] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. arXiv preprint arXiv:2010.16417, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[36] tenasia10. File:blackpinkLisa gmp 240622.png trends @ONLINE. https://commons.wikimedia.org/wiki/File:Blackpink_Lisa_GMP_240622.png/, 2022.17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[37] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[38] GABI THORNE. Jennifer aniston reacts to tiktok's best (and most absurd) trends @ONLINE. https://www.allure.com/story/jennifer-aniston-reacts-to-tiktok-trends/, 2022.17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[39] Albert Truuväärt. File:re kaja kallas (cropped).jpg @ONLINE. https://commons.wikimedia.org/wiki/File:RE_Kaja_Kallas_(cropped). jpg/, 2011.17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[40] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 5914–5922, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[41] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting gans with consecutive images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13910-13918, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[44] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proc. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.894, + 0.317 + ], + "angle": 0, + "content": "[45] Tao Zhou, Changqing Zhang, Chen Gong, Harish Bhaskar, and Jie Yang. Multiview latent space learning with feature redundancy minimization. IEEE transactions on cybernetics, 50(4):1655-1668, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[46] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. arXiv preprint arXiv:2106.01505, 2021. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 651-667. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[48] zllrunnings. face-parsing.pytorch @ONLINE. https: //github.com/zllrunning/face-parsing. PyTorch/, 2019.11" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8618" + } + ] +] \ No newline at end of file diff --git a/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_origin.pdf b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b865058ef8c2907d4c9bde30413e72849328a0b3 --- /dev/null +++ b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/ee5e8b03-df5c-443b-b56f-dfc54bf4046b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:148ea7b38a6787c426b43745ce9cfb59f8fdca6f52cfe0b6c4f6f04fac2b291e +size 7271308 diff --git a/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/full.md b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..80bcf70c29cc71489fbab5d096fdad2444a3317b --- /dev/null +++ b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/full.md @@ -0,0 +1,333 @@ +# StyleGAN Salon: Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer + +Sasikarn Khwanmuang\* Pakkapon Phongthawee\* Patsorn Sangkloy\* Supasorn Suwajanakorn + +VISTEC, Thailand Phranakhon Rajabhat University, Thailand + +{sasikarn.k_s18, pakkapon.p_s19, supasorn.s}@vistec.ac.th patsn.s@pnru.ac.th + +![](images/1facb3b704e8e0370a119f7004f0d6c0b61389ee356caa02c58565d5813fa96d.jpg) +Input Face +Figure 1. Our method can transfer the hairstyle from any reference hair image in the top row to Tom Holland [33], in the second row. + +# Abstract + +Our paper seeks to transfer the hairstyle of a reference image to an input photo for virtual hair try-on. We target a variety of challenges scenarios, such as transforming a long hairstyle with bangs to a pixie cut, which requires removing the existing hair and inferring how the forehead would look, or transferring partially visible hair from a hat-wearing person in a different pose. Past solutions leverage StyleGAN for hallucinating any missing parts and producing a seamless face-hair composite through so-called GAN inversion or projection. However, there remains a challenge in controlling the hallucinations to accurately transfer hairstyle and preserve the face shape and identity of the input. To overcome this, we propose a multi-view optimization framework that uses two different views of reference composites to semantically guide occluded or ambiguous regions. Our optimization shares information between two poses, which allows us to produce high fidelity and realistic results from incomplete references. Our framework produces high-quality results and outperforms prior work in a user study that consists of significantly more challenging hair transfer scenarios than previously studied. Project page: https://stylegan-salon.github.io/. + +# 1. Introduction + +What makes Jennifer Aniston keep her same hairstyle for over three decades? Perhaps she likes the classic, or perhaps changing her hairstyle is a decision too high-stakes that she could later regret. Unlike garments or makeup, trying on a new hairstyle is not easy, and being able to imagine yourself in different hairstyles could be an indispensable tool. + +Recent approaches for hairstyle transfer, StyleYourHair [20], Barbershop [46], LOHO [30], and Michiganan [35], allow users to manipulate multiple hair attributes of an input image, such as appearance, shape, or color by providing a reference image for each different attribute. These methods [20, 30, 46] rely on a generative adversarial network [11], specifically StyleGAN2 [19], which can synthesize highly realistic face images. Their key idea, which also forms the basis of our method, is to leverage the realistic face distribution learned by StyleGAN and search for a hairstyle-transfer output whose latent code lies within the learned distribution using optimization (commonly known as GAN projection or inversion). + +Our extensive study on these state-of-the-art techniques still reveal several unsolved challenges for in-the-wild hairstyle transfer. One of the main challenges is when the reference hair comes from a person with a very different head + +pose or facial shape. In this case, the transfer result often degrades significantly [30, 46]. HairFIT [8] and the recently proposed StyleYourHair [20] both attempt to solve this using an additional alignment step to align the pose of the target hair to the input face. In HairFIT [8], this is done explicitly via a flow-based warping module for hair segmentation, but this requires training on multi-view datasets [21, 26]. StyleYourHair [20] avoids this issue and also improves upon HairFIT's results by optimizing for an aligned pose within StyleGAN2's latent space using distances between the detected facial keypoints. While StyleYourHair can handle a certain degree of misalignment, it often struggles to preserve details of the reference hair texture, especially for intricate and non-straight hairstyles, e.g., in Figure 3. + +In general, we have observed that there is a trade-off between hallucinating new details, which is crucial for handling highly different poses, and preserving the original texture from the reference images. These two goals are often at odds with each others. This trade-off is also evident in the results from StyleYourHair, as reported in [20] and as shown in table 3, where BarberShop [46] can still produce better results, for example, when the pose difference is small. + +We tackle this dilemma by performing a multi-stage optimization. This serves two purposes: first, to hallucinate new details necessary for aligning the poses, and second, to recover face-hair details from the original input images. We preserve these details in a form of two guide images in both viewpoints, which will be jointly optimized to allow new details to be filled while retaining face-hair texture in the original pose. Our pose alignment is done both explicitly via 3D projection [6] on RGB images, and implicitly via latent code(s) sharing during our multi-view optimization. + +In summary, our contributions are as follows: + +1. We propose StyleGAN Salon: a pose-invariant hairstyle transfer pipeline that is flexible enough to handle a variety of challenging scenarios including, but not limited to, bangs/hat removal and background inpainting. +2. Unlike previous works, our method operates entirely on RGB images, which are more flexible than segmentation masks. This allows us to first draft the output and then refine them via multi-stage optimization. +3. We introduce multi-view optimization for hairstyle transfer which incorporates 3D information to align the poses for both face and hair images. Our method leverages both views to help preserve details from the original images. +4. We thoroughly analyze the results in several experiments, including a user study with detailed breakdown into various challenging scenarios. Our method shows superior results over existing works in all scenarios. + +# 2. Related Work + +Generative Adversarial Networks. Beginning with [11, 18], StyleGANs [17-19] have shown great success in 2D image generation by learning realistic training data distribution corresponding to a fixed low-dimensional distribution, called latent code. The learned latent code can be applied to a downstream task such as image manipulation [13, 15, 32, 40]. Recent work on [6, 12, 44] also explore 3D aware architecture in StyleGAN, resulting in multi-view consistent images. + +For hairstyle transfers, MichiGAN [35] uses a conditional hair generation network that can control hair shape, structure, and appearance. Recently, HairFIT [8], a pose-invariant hairstyle transfer network, aligns the reference hair to match the input face pose using a optical flow-based hair alignment module, but requires training on multi-view dataset [21, 26]. + +StyleGAN Latent Space and Projection Techniques. To generate an image, StyleGAN2 first maps a random latent code $z \sim N(0, I)$ to an intermediate latent code $w \in \mathbb{R}^{512}$ in a new space called $\mathcal{W}$ . This code $w$ is then replicated (18x) and used as input to each layer of the generator that controls details at different scales through weight demodulation. Collectively, these replicated latent codes are referred to as $w^{+} \in \mathbb{R}^{18 \times 512}$ in the extended latent space $\mathcal{W}^{+}$ . + +Pivotal Tuning Inversion (PTI) [29] further improves the projection quality by directly tuning generator weights around the optimized latent code $w$ using their regularization technique. Other techniques [1,2,5] optimize the latent code $w^{+}$ in each layer separately to better match the input image. + +PIE [37] introduces hierarchical optimization for semantic attributes editing that first optimizes the latent code in the $\mathcal{W}$ space, then transfers the code to the $\mathcal{W}^+$ space and continues the optimization in that space. We adopt similar hierarchical optimization that uses both $\mathcal{W}$ and $\mathcal{W}^+$ . Part of our method is also inspired by PULSE [25], which reconstructs a high-resolution image from a small reference image (32x32) by searching for the closest image in the StyleGAN latent space that resembles the low-resolution reference. + +Some methods [7, 41, 45] project multiple images into latent space simultaneously. However, all of their inputs are complete, whereas our method requires hair information from the reference hair image and additional information from the input face image. + +StyleGAN-Based Hairstyle Transfer. LOHO [30] adopts loss functions from Image2StyleGAN++ [2] to combine face-hair inputs into a hairstyle transfer output with their orthogonalization optimization technique which reduces conflicts between multiple loss functions. Barbershop [46] first predicts semantic regions of both inputs and uses them to create a "target segmentation mask" of the output with rule (in the original paper) or GAN-based inpainting (in their official code). Then, they optimize two separate latent codes in $\mathcal{W}^+$ space, one for matching the face and the other for hair, while conforming to the target segmentation mask. To + +preserve the original details, the latent optimization is done in their proposed $\mathcal{F} / \mathcal{S}$ space, which replaces the first seven blocks of $\mathcal{W}^+$ space with the corresponding activation maps of StyleGAN's convolution layers. To improve Barbershop's capacity to handle unaligned input poses, StyleYourHair [20] first aligns the reference hair to match the pose of the input face via the proposed local-style-matching loss. However, this alignment often leads to an unrealistic hair shape or inaccurate hair texture results. In contrast, our method simultaneously optimizes for two guides in face and hair views, resulting in a better hair texture. + +Instead of estimating the final output from the segmentation mask [20, 46], our multi-view optimization uses face-hair composites in RGB space to overcome this problem, and produces results that better preserve the input facial structure and hairstyle across a wider, more challenging range of scenarios. Concurrent work, HairNet [47], arrives at a similar goal of removing the target segmentation mask via a two-step process that involves baldification using StyleFlow [3] followed by another network to transfer the hairstyle. + +Nevertheless, handling pose differences is crucial for successful hairstyle transfer; our method incorporates 3D rotation to preserve the geometric consistency and systematically evaluate this aspect. + +# 3. StyleGAN Salon + +Given two input images $I_h$ and $I_f$ , the goal is to transfer the hair from the reference hair image $I_h$ into the face in the image $I_f$ , while preserving all remaining details in $I_f$ , including identity, clothing, and background. + +The key idea of our approach is to guide the optimization on the learned latent space of StyleGAN2 with two "guide" images, which represent rough composites of the final output based on simple cut-and-paste in the viewpoint of $I_{h}$ and $I_{f}$ . We leverage EG3D to construct these guided images in a geometrically consistent way, described in Section 3.1. + +Optimization on StyleGAN2's latent space is commonly performed on either the original latent space $\mathcal{W}$ or the extended latent space $\mathcal{W}^+$ . Optimizing on $\mathcal{W}$ space generally leads to more realistic results by staying within the original latent space, whereas optimizing on $\mathcal{W}^+$ allows a closer match to the reference [37]. In hairstyle transfer, it is important to stay faithful to the input images and preserve important details such as hair texture, face identity, and background. However, optimizing on $\mathcal{W}^+$ will lead to poor results because our guide images are rough and unrealistic composites. Thus, we propose to optimize on $\mathcal{W}$ space followed by $\mathcal{W}^+$ space, similar to a technique in PIE [37] used for editing semantic attributes of an image. + +Our optimization incorporates both guide images from the two viewpoints, detailed in Section 3.2. Section 3.3 and Section 3.4 cover details of our optimization on $\mathcal{W}$ and $\mathcal{W}^+$ , respectively. Finally, we also optimize StyleGAN2 weights + +while freezing the latent codes (Section 3.5) using PTI [29] to further improve detail fidelity. Figure 2 shows an overview of our complete pipeline. + +# 3.1. Constructing the Guide Images + +The purpose of our guide images is to provide an initial estimate of how the hair would look on $I_{f}$ . We achieve this using a cut-and-paste composite of the face and background from $I_{f}$ and the hair from $I_{h}$ . To better handle a potentially large shift in viewpoint between $I_{f}$ and $I_{h}$ , we propose to leverage EG3D [6] to help generate geometrically consistent guide images. We argue that using multi-view guide images, each in the pose of $I_{f}$ and $I_{h}$ , helps preserve details that could otherwise be lost from using a single viewpoint alone. + +A straightforward approach is to simply project $I_h$ into the EG3D [6] latent space, and use their proposed neural rendering pipeline to render it in the view of $I_f$ (and vice versa). However, while the projection can produce a geometrically consistent 3D shape, we found that the resulting texture is not very accurate. To address this, we replace any visible regions in the texture from $I_h$ 's viewpoint with the original pixels of $I_h$ , while leaving the rest of the texture as the projected texture from EG3D's rendering. + +We also additionally apply uniform scaling and translation to match the faces' widths and centers, which are computed based on detected facial keypoints [22]. These keypoints, along with semantic regions from [42], are also used to handle various corner cases, such as re-painting unwanted hair regions. We refer to Appendix A for more details of this operation. + +The entire process is done for both $I_{f}$ and $I_{h}$ viewpoints, resulting in a pair of guide images $I_{\mathrm{guide}} = [I_{\mathrm{guide}}^{\mathrm{face}}, I_{\mathrm{guide}}^{\mathrm{hair}}]$ . We emphasize that by leveraging our multi-view optimization (Section 3.2), these guide images do not need to be precise or realistic to produce convincing final results, as demonstrated in the first row of Figure 4. + +# 3.2. Multi-View Latent Optimization + +Our guide images provide complementary information about the target hair and face, albeit from different poses. However, each guide image is only fully accurate in regions with the original pixels seen in the original viewpoint and not warped by EG3D. Thus, our optimization goal is to combine information from both guide images to generate a final output that accurately captures the realistic hair from $I_{\mathrm{guide}}^{\mathrm{hair}}$ , as well as other details from $I_{\mathrm{guide}}^{\mathrm{face}}$ . We achieve this using multiple loss functions that attend to both viewpoints with different spatial emphasis. Specifically, we perform multi-view latent optimization on $w / w^{+}$ (and stochastic noise maps $n$ ) that fits both guide images: + +$$ +\min _ {\{w / w ^ {+}, n \}} \sum_ {\mathrm {i}} ^ {[ \text {f a c e , h a i r} ]} \mathcal {L} _ {\text {l o s s}} \left(O ^ {(\mathrm {i})}, I _ {\text {g u i d e}} ^ {(\mathrm {i})}\right), \tag {1} +$$ + +![](images/4d7cdeef94bda516a58d021f9cf161129d9929addd054f442dd842c985723a89.jpg) +Figure 2. Overview of StyleGAN Salon: We first align the input face $I_{f}$ and reference hair $I_{h}$ and use them to construct guide images $I_{\mathrm{guide}}$ , in two different viewpoints, which specifies the target appearance for each output region. (see Section 3) + +where $O^{(\mathrm{i})}$ denotes the output of the StyleGAN2's generator that takes a latent code $w^{(\mathrm{i})}$ or $w^{+\mathrm{(i)}}$ and a stochastic noise input $n^{(\mathrm{i})}$ . $\mathcal{L}_{\mathrm{loss}}$ is a sum of all our loss functions: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {l o s s}} = \sum_ {\mathrm {j}} ^ {[ \mathrm {f}, \mathrm {h}, \mathrm {b g} ]} \lambda_ {\mathrm {p}} ^ {(\mathrm {j})} \mathcal {L} _ {\text {p e r}} ^ {(\mathrm {j})} (2) \\ + \lambda_ {\mathrm {g}} \mathcal {L} _ {\text {g l o b a l}} + \lambda_ {\mathrm {i}} \mathcal {L} _ {\text {i n i}} + \lambda_ {\varepsilon} \mathcal {L} _ {\varepsilon} + \lambda_ {s} \mathcal {L} _ {\text {s i m}}, (3) \\ \end{array} +$$ + +where $\lambda_{(\cdot)}$ are balancing weights, and [f, h, bg] refer to face, hair, and background. We next explain each loss function. + +# 3.2.1 Loss Functions + +The objective of our $\mathcal{W}$ and $\mathcal{W}^+$ optimization is to generate an output image that fits the corresponding $I_{\mathrm{guide}}$ without appearing unrealistic. The main challenge is that our guide images are unrealistic and can contain various unwanted artifacts. We design our loss functions to be tolerant of the imprecise nature of our guide images. + +Masked Perceptual Loss: This loss is based on Learned Perceptual Image Patch Similarity (LPIPS) [43], which compares two input images in a deep feature space. Similar to [20, 30, 46], we apply a binary mask to indicate regions of interest in the deep feature space of the two input images. We use $L_{\mathrm{LPIPS}}(I_1, I_2; M)$ to denote this LPIPS computation between two images $(I_1, I_2)$ with mask $M$ . However, we observe that applying this masking operation after the feature computation is insufficient to disregard unwanted regions. This is due to the fact that LPIPS is a patch-based similarity loss, and regions outside of the mask can still affect the loss computation. Based on this principle, we propose to apply additional pre-masking to mask out ambiguous regions in $I_{\mathrm{guide}}$ that should not be trusted. Specifically, our Masked Perceptual Loss is the following: + +$$ +\mathcal {L} _ {\text {p e r}} ^ {\mathrm {i}, \mathrm {r o i}} = \Lambda_ {\mathrm {r o i}} ^ {(\mathrm {i})} L _ {\mathrm {L P I P S}} \left(O ^ {(\mathrm {i})} \odot \neg M _ {\mathrm {r o n i}} ^ {(\mathrm {i})}, I _ {\text {g u i d e}} ^ {(\mathrm {i})} \odot \neg M _ {\mathrm {r o n i}} ^ {(\mathrm {i})}; M _ {\mathrm {r o i}} ^ {(\mathrm {i})}\right), \tag {4} +$$ + +where $M_{\mathrm{roi}}$ is the mask for region of interest (ROI), and $M_{\mathrm{roni}}$ is the mask for regions of not interest (RONI). $(.)\odot \neg M_{\mathrm{roni}}$ + +is a simple element-wise multiplication, which effectively excludes these regions from the loss by setting them to 0 in both input images. And $\Lambda_{\mathrm{roi}}^{\mathrm{(i)}}$ is a balancing weight in (i) viewpoint, which is set higher when the region of interest covers the original pixels of $I_{f}$ (or $I_{h}$ ). For instance, the loss that attends to face in $I_{\mathrm{guide}}^{\mathrm{face}}$ has a higher weight than the face loss in $I_{\mathrm{guide}}^{\mathrm{hair}}$ . We consider 3 regions of interest: $L_{\mathrm{per}}^{\mathrm{f}}$ , $L_{\mathrm{per}}^{\mathrm{h}}$ , and $L_{\mathrm{per}}^{\mathrm{bg}}$ for face, hair, and background regions. We refer to Section 3.3 and 3.4 for details on these mask generation for $\mathcal{W}$ and $\mathcal{W}^{+}$ optimization, respectively. + +Global loss: This loss function attempts to match the overall appearance of the output to $I_{\mathrm{guide}}$ : + +$$ +\mathcal {L} _ {\text {g l o b a l}} = L _ {\text {M S E}} ^ {3 2} (O, I _ {\text {g u i d e}}). \tag {5} +$$ + +$L_{\mathrm{MSE}}^{32}$ is the mean square error computed on $32 \times 32$ downsampled input images. The goal of the downsampling is to reduce the effect of matching visible seams from higher resolution $I_{\mathrm{guide}}$ because those seams become imperceptible once downsampled. This loss is inspired by a similar idea used for image super-resolution by Menon et al. [25]. + +Initial Value Loss: To prevent the output from deviating too far from the learned distribution and becoming unrealistic, we force the optimized latent code(s) $w$ and $w^{+}$ to be close to its initial value $w_{0}$ through L2 loss, as used in [2]: + +$$ +\mathcal {L} _ {\mathrm {i n i}} = \left\| w _ {1: 1 8} - w _ {0} \right\| _ {2} ^ {2}. \tag {6} +$$ + +We use the estimated mean of $\mathcal{W}$ latent space as initial values for optimizing $w$ . This is computed by averaging many latent codes $w_{i} = f(z_{i})$ drawn randomly through $z\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and the StyleGAN's mapping network $f$ . Then the optimized latent code $w$ from the first stage becomes the initial value for the optimization in $\mathcal{W}^+$ extended latent space. + +Noise Regularization Loss: We also use the noise regularization loss $(\mathcal{L}_{\varepsilon})$ proposed in StyleGAN2 [19] to ensure that the noise maps will be used for capturing the stochastic variations, and not for encoding the content meant to be captured by the latent code. We refer to StyleGAN2 for the details of this loss. + +Latent Similarity Loss: To enable information sharing between the two guide images ( $I_{\mathrm{guide}}^{\mathrm{face}}$ and $I_{\mathrm{guide}}^{\mathrm{hair}}$ ) from two viewpoints, we force both their latent codes to be close together by using the following L2 loss, similarly to [7]. + +$$ +\mathcal {L} _ {\text {s i m}} = \left\| w ^ {(\text {f a c e})} - w ^ {(\text {h a i r})} \right\| _ {2} ^ {2}, \tag {7} +$$ + +where $w$ can be in $\mathcal{W}$ space during $\mathcal{W}$ optimization or replaced with $w^{+}$ during $\mathcal{W}^+$ optimization. + +# 3.2.2 Sharing Latent Code + +In StyleGAN2, latent codes in early layers have been shown to correspond to high level concepts (e.g. human pose) whereas the remaining layers capture low-level information (e.g. colors) [15, 18]. This motivates the optimization in $\mathcal{W}^+$ extended space, as each layer can be optimized for different aspects of the image. Similarly, we optimize latent codes to match guide images of the same person and hairstyle but in different poses. Rather than optimizing two latent codes independently, we can enforce information sharing between the two poses by constraining the latent codes of the last few layers ( $w_{l:18}$ ) to be the same. We set $l$ to be 4 in all our experiments, loosely based on the head rotation experiments shown in GANSpace [15]. + +Sharing $w$ Latent Code: Optimization in the space of StyleGAN2 latent code $\mathcal{W}$ is generally done as a single latent code $(\mathbb{R}^{512})$ , which are duplicated (18x) and fed into all the layers. In our case, we also optimize for a single latent code $w$ but only duplicate them for the first $l$ layers $(w_{1:l})$ . The remaining layers $(w_{l:18})$ are interpolated from the current latent codes of each guide image with a coefficient $\alpha$ . + +$$ +w _ {l: 1 8} = \alpha w _ {1: l} ^ {\text {(f a c e)}} + (1 - \alpha) w _ {1: l} ^ {\text {(h a i r)}}. \tag {8} +$$ + +We randomize this coefficient to make the latent codes stay within the space of $\mathcal{W}$ . This random interpolation forces the latent code for each guide image to be partly similar while still producing realistic results from $\mathcal{W}$ latent space. + +Sharing $w^{+}$ Latent Code: Optimization in $\mathcal{W}^{+}$ extended space is straightforward, as latent code for each layer can already be separately optimized. We simply share the last few layers $(w_{l:18})$ of both $w^{+}$ latent codes during optimization. We refer to the Appendix A.5 for more details on the optimization process. + +# 3.3. $\mathcal{W}$ Optimization: Hallucinate Missing Details + +This stage aims to hallucinate details not currently visible in the guide images. The output images of this stage are $O_{1}^{\mathrm{face}}$ and $O_{1}^{\mathrm{hair}}$ , which are essentially the projection of the unrealistic $I_{\mathrm{guide}}^{\mathrm{face}}$ and $I_{\mathrm{guide}}^{\mathrm{hair}}$ into the real image distribution learned in StyleGAN2 [19]. This is done by optimizing on $W$ latent space to fit each guide image with the objective function (Equation 2). + +While our guide images ( $I_{\mathrm{guide}}^{\mathrm{face}}$ and $I_{\mathrm{guide}}^{\mathrm{hair}}$ ) capture the overall appearance of the desired outputs, they still lack details in certain regions. These unknown regions may correspond to unseen facial features that were occluded by the hair in $I_{f}$ or incomplete reference hair from $I_{h}$ caused by, e.g., a hat or image cropping. These details will need to be hallucinated and seamlessly blended with the rest of the image. + +To accomplish this, we design our pre-masking in Masked Perceptual Loss $(\mathcal{L}_{\mathrm{per}}^{\mathrm{f}},\mathcal{L}_{\mathrm{per}}^{\mathrm{h}}$ , and $\mathcal{L}_{\mathrm{per}}^{\mathrm{bg}}$ ; Equation 4) to include all the unknown regions that need to be hallucinated. Specifically, $M_{\mathrm{roni}}^{\mathrm{f}}$ represents the face region occluded by the hair in $I_{f}$ , e.g., the forehead behind the bangs or the ears that should become visible in the final output. Analogously, $M_{\mathrm{roni}}^{\mathrm{h}}$ represents the hair region occluded by other objects (e.g., a hat) or not visible in $I_{h}$ due to image cropping. We set $M_{\mathrm{roi}}^{\mathrm{f}}$ and $M_{\mathrm{roi}}^{\mathrm{h}}$ to be segmentation masks for face and hair regions $(M_{\mathrm{f}}$ and $M_{\mathrm{h}})$ , respectively. Both $M_{\mathrm{roi}}^{\mathrm{bg}}$ and $\neg M_{\mathrm{roni}}^{\mathrm{bg}}$ is set to be the background regions in $I_{f}$ not covered by the transferred hair $(M_{\mathrm{bg}})$ . These masks are constructed by composing different semantic regions (union, intersection, etc.) from $I_{f}$ and $I_{h}$ , detailed in Appendix A.3 + +We note that $M_{\mathrm{roi}}$ is applied on the deep feature maps, similar to [20, 30, 46], while $\neg M_{\mathrm{roni}}$ is applied on raw RGB images. Using both masking techniques is crucial to our natural and seamless blending. The additional pre-masking allows the hair shape (or face shape) in the output to be different from $I_{\mathrm{guide}}$ 's and freely grow outward outside $M_{\mathrm{roi}}$ if this leads to a more natural result (see Figure 4). + +# 3.4. $\mathcal{W}^+$ Optimization: Recover Face-Hair Details + +The output of the previous stage, $O_1^{\mathrm{face}}$ and $O_1^{\mathrm{hair}}$ , may still look different from the input person and not yet capture the hair details from the reference. This stage aims at refining these output images to better reproduce the hair details from $I_h$ and the rest from $I_f$ . The optimization is done in the extended $\mathcal{W}^+$ space with respect to $\{w^{+},n\}$ . In other words, we allow $w\in \mathbb{R}^{512}$ that was previously replicated to each layer to be optimized separately as $w^{+}\in \mathbb{R}^{18\times 512}$ . We denote $O_2^{\mathrm{face}}$ and $O_2^{\mathrm{hair}}$ to be the output of this stage generated by StyleGAN2 from our optimized latent codes in $\mathcal{W}^+$ extended latent space. + +We update the target images for the optimization from $I_{\mathrm{guide}}^{\mathrm{face}}$ and $I_{\mathrm{guide}}^{\mathrm{hair}}$ to the more complete versions based on $O_1^{\mathrm{face}}$ and $O_1^{\mathrm{hair}}$ from the first stage. However, because the first stage aims at hallucinating new details into the pre-masking regions, it may lose the original texture details. To address this, we replace the known regions in $O_1^{\mathrm{face}}$ and $O_1^{\mathrm{hair}}$ with the correct details from the original $I_{\mathrm{guide}}$ by setting + +$$ +I _ {\text {n e w} \text {g u i d e}} \leftarrow I _ {\text {g u i d e}} \odot M _ {\mathrm {c}} + O _ {1} \odot \neg M _ {\mathrm {c}}, \tag {9} +$$ + +where $M_{c} = M_{\mathrm{f}} \cup M_{\mathrm{h}} \cup M_{\mathrm{bg}}$ , which corresponds to the regions of interest in our Masked Perceptual Loss where we want to match with $I_{\mathrm{guide}}$ . In short, we simply copy the + +hallucinated parts from $O_1^{\mathrm{face}}$ and $O_1^{\mathrm{hair}}$ , and combine them with the part in $I_{\mathrm{guide}}$ that are known to be correct. + +For our Masked Perceptual Loss, we use the same masking procedure as the previous stage (Section 3.3). However, we observe that the projection from EG3D, which produces $I_{\mathrm{guide}}$ , generally is inferior to the known details from the original images. For example, $I_{\mathrm{guide}}^{\mathrm{hair}}$ contains more details on the hair than $I_{\mathrm{guide}}^{\mathrm{face}}$ due to the latter being hallucinated from the 3D rotation by EG3D's rendering pipeline. To de-prioritize these inaccurate details, we blur both $I_{\mathrm{guide}}$ and $O$ when computing the perceptual loss for the less accurate regions of interest. Concretely, we only blur images when computing $\mathcal{L}_{\mathrm{per}}^{\mathrm{f}}$ for $I_{\mathrm{guide}}^{\mathrm{hair}}$ optimization and when computing $\mathcal{L}_{\mathrm{per}}^{\mathrm{h}}$ for $I_{\mathrm{guide}}^{\mathrm{face}}$ optimization. + +# 3.5. Pivotal Tuning Inversion + +The purpose of this stage is to further refine the output images $(O_2^{\mathrm{face}}$ and $O_2^{\mathrm{hair}})$ by allowing the optimization of the StyleGAN's weights $\theta$ (while fixing the optimized latent code $w_{\mathrm{optimized}}$ ). We closely follow the proposed optimization in PTI [29] with the same objective function. However, we update the reconstruction loss to reflect the goal of our task, which is to match the original details from $I_{f}$ and $I_{h}$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {p t i}} = \sum_ {\mathrm {i}} ^ {[ \text {f a c e}, \text {h a r i} ]} \left(L _ {\mathrm {L P I P S}} \left(O _ {\text {t u n e}} ^ {(\mathrm {i})} \odot M _ {\text {r a w}} ^ {(\mathrm {i})}, I _ {\text {g u i d e}} ^ {(\mathrm {i})} \odot M _ {\text {r a w}} ^ {(\mathrm {i})}; M _ {\text {r a w}} ^ {(\mathrm {i})}\right) \right. \\ + L _ {\mathrm {M S E}} ^ {3 2} \left(O _ {\text {t u n e}} ^ {(\mathrm {i})}, I _ {\text {g u i d e}} ^ {(\mathrm {i})}; \neg M _ {\text {r a w}} ^ {(\mathrm {i})}\right)), \tag {10} \\ \end{array} +$$ + +where $O_{\mathrm{tune}}$ is the generated image using the tuned weights $\theta$ , the mask $M_{\mathrm{raw}}^{\mathrm{face}}$ is $(M_{\mathrm{bg}} \cup M_{\mathrm{f}})$ , and $M_{\mathrm{raw}}^{\mathrm{hair}}$ is $M_{\mathrm{h}}$ . + +# 4. Experiments + +In this section, we compare our method to state-of-the-art StyleYourHair [20], Barbershop [46], and LOHO [30]. Our evaluation criteria are i) user preference via a user study on a wide variety of scenarios, ii) hairstyle transfer quality, iii) hair reconstruction quality, and iv) how well the input face shape is preserved. Section 4.5 presents ablation studies on our multi-view sharing latent optimization, optimization stage, and loss functions. + +We use the official code of LOHO [30] and Barbershop [46] with the default configurations. For StyleYourHair [20], we use the configuration where the hair reference is never flipped. + +# 4.1. Qualitative Comparison + +We provide a qualitative comparison to StyleYourHair [20], Barbershop [46], and LOHO [30] in Figure 3. We observe that LOHO and Barbershop often struggle to fit the reference hair accurately when the poses are not well aligned (rows 3-5), resulting in various artifacts, such as remnants of the original hair or wrong placement of the + +![](images/54f7e8d6d84270eb96811594f93d8920ee9993f23871d939a8afb48c3990d428.jpg) +LOHO Barbershop StyleYourHair Ours +Figure 3. Comparison to current state-of-the-art methods for transferring hair from 1st column-top, to the face of 1st column-bottom. Our method can accurately transfer hairstyle even when the input face and hair are misaligned (3rd, 4th row). It can also hallucinate missing details such as a forehead that was previously occluded (1st row), or shorten the hair (2nd row). In general, LOHO and Barbershop struggle with pose misalignment, while StyleYourHair struggles to preserve the input's face shape and hair details. + +target hair. Barbershop performs well in preserving hair texture when the poses are similar. However, it falls short when handling challenging cases, such as transitioning from a long to a short hairstyle or removing bangs. + +StyleYourHair can produce more realistic results in unaligned cases than LOHO and Barbershop. However, the hair details often look different from the reference hairstyle (row 5). We also observe that StyleYourHair may perform poorly when the pose difference becomes too large (row 3), as also shown in our user study for pose in Table 1. In contrast, our method can transfer hairstyles convincingly, regardless of the misalignment, while still preserving the original face shape and hair details. We refer to Appendix F for more results. + +# 4.2. User Study + +Qualitative results can be misleading and biased, so we conducted a user study using Amazon Mechanical Turk on hairstyle transfer results using randomly selected pairs grouped into various difficulty levels. We compare results of our method with the current state of the art: StyleYourHair, Barbershop, and LOHO. Each participant was shown an input face, marked as 'Face', and a reference hair, marked + +
Test DatasetsFFHQ Scenario Breakdown
AllFFHQ-PFFHQ-SEasyMediumDifficult
-✓---✓✓✓✓✓---
--✓---✓--✓✓-
---✓---✓-✓-✓
----✓---✓-✓✓
LOHO8.89.68.412.05.34.013.38.05.35.34.010.75.313.314.7
Barbershop16.413.817.817.321.313.321.318.717.317.320.012.018.714.721.3
StyleYourHair18.512.421.625.321.310.722.733.310.724.028.014.732.014.721.3
Ours56.264.252.245.352.072.042.740.066.753.348.062.744.057.342.7
+ +as 'Hair', and asked to pick only one output that best accomplishes the task of transferring the hairstyle from image 'Hair' to the person in image 'Face'. The output row consists of four images from each method in random order. Each task was evaluated by 3 different participants. All images were in 256x256 resolution. + +# 4.2.1 Datasets + +For the user study, we construct two challenging benchmarks: FFHQ-P and FFHQ-S, from the test set of Flickr-Faces-HQ dataset (FFHQ) [18]. + +FFHQ-P contains 150 random input face-hair pairs from FFHQ, covering yaw differences in ranges of $[0 - 15)$ , $[15 - 30)$ , ..., $[75 - 90)$ , with 25 pairs in each range. + +FFHQ-S contains 300 pairs, categorized into 12 different config with varying levels of difficulty (Table 1). Each config contains 25 random input pairs and is a combination of four possible scenarios (see details in Appendix B): + +- Pose Misalignment: When the yaw difference is between [15-30) or [30-45), a single checkmark or double checkmarks are used, respectively in Table 1. +- Needs Face Inpainting: This includes scenarios that require hallucinating parts of the original face, e.g., inpainting the forehead to remove bangs. This is challenging because the identity can easily change from the hallucination. We detect such scenarios based on the face/hair regions in $I_{f}$ and $I_{h}$ (Appendix B). +- Needs BG Inpainting: This includes scenarios where the hair shape becomes smaller and requires background inpainting. We detect such scenarios automatically based on the hair regions in $I_{f}$ and $I_{h}$ (Appendix B). +- $I_h$ Contains Hat: This represents scenarios where the hair reference is not fully visible in $I_h$ . We detect such scenarios based on the hat region in $I_h$ . + +Table 1. User study results on hairstyle transfer (percentage of user preferring each method). Our method outperforms state-of-the-art hairstyle transfer methods on FFHQ datasets in all challenging scenarios. A total of 450 pairs are used in this study, 150 pairs in FFHQ-P and 300 in FFHQ-S. For each pair, we asked 3 unique participants to select the best result for hairstyle transfer. + +
Pose Difference Range (FFHQ-P)
[0,15)[15,30)[30,45)[45,60)[60,75)[75,90)
LOHO10.75.310.712.08.010.7
Barbershop24.013.312.08.010.714.7
StyleYourHair17.316.012.06.714.78.0
Ours48.065.365.373.366.766.7
+ +Table 2. User study on pose-invariant hairstyle transfer. Our method outperforms others on all pose difference ranges. + +
Hair Reconstruction MetricsFace Shape
PSNR ↑SSIM ↑LPIPS ↓FID ↓RMSE ↓
LOHO25.760.860.0710.4015.84
Barbershop29.180.890.0510.4617.25
StyleYourHair26.890.870.0910.9314.37
Ours27.840.880.0710.8510.89
+ +Table 3. Hair reconstruction results in the self-transfer experiment (Section 4.3), and RMSEs between facial landmarks detected on the input and output images.(Section 4.4). + +# 4.2.2 Results + +Using all test pairs, the participants preferred our results $56.2\%$ of the time, whereas StyleYourHair's, Barbershop's, and LOHO's results were selected for $18.5\%$ , $16.4\%$ and $8.8\%$ , respectively. Our method perform the best in both FFHQ-P $(64.2\%)$ and FFHQ-S $(52.2\%)$ , and in all configurations in Table 1-2. + +# 4.3. Quality of Hair Reconstruction + +Following MichiGAN and LOHO, we perform a self-transfer experiment where we set both the input face and input reference hair to be the same and evaluate the reconstruction accuracy using various metrics: PSNR, SSIM, IP-IPS, FID [14]. For every method, the hair region of the output will be blended back to the input image to ensure that the difference in score only comes from the hair. + +![](images/ea698b7ad3b32866c78395879f28b92d81bedb8aeb868b081837daf97ab6eb0a.jpg) +Figure 4. Ablation study of optimization stages and pre-masking. These ablated versions of our pipeline may produce results that (A) look realistic, but fail to match the $I_{h}$ 's hairstyle; (B, C, D) have unlikely hair shapes or structures; (E) have sharp boundaries; and (F) contain an incomplete background. + +The scores of our method, LOHO, Barbershop, and StyleYourHair are reported in Table 3. Unsurprisingly, when there is no misalignment between input face and hair, Barbershop generally achieves excellent hair reconstruction quality. Like ours, StyleYourHair is specifically designed to be pose-invariant, but their hair reconstruction quality seems to suffer greatly, as shown in our comparison (row 5 of Figure 3). Our method achieves better performance at hair reconstruction than StyleYourHair, but also suffers from similar drawbacks. + +# 4.4. How Well Is Face Shape Preserved? + +We also propose a novel evaluation metric for hairstyle transfer that focuses on the ability to preserve facial shape of the original person. We accomplish this by comparing detected keypoints of the input face $I_{f}$ and those on the output using a simple Root-Mean-Square Error (RMSE). These keypoints were detected using an off-the-shelf library Dlib [22], and we only used keypoints on the facial contour (ID 0-16) for this evaluation. + +For this evaluation, we randomize additional 1,550 FFHQ test pairs into our FFHQ-P and FFHQ-S datasets (2,000 images in total). Table 3 shows that our method outperforms others with an RMSE of 10.89, which is $24.2\%$ lower than StyleYourHair's (the second best), $36.9\%$ lower than Barbershop's and $31.3\%$ lower than LOHO's. + +# 4.5. Ablation Studies + +Here we assess the importance of each component in our pipeline. The results are shown in Figure 4. We test 4 ablation configurations: i) by using only $I_{\mathrm{guide}}^{\mathrm{face}}$ from a single view, ii) without using the hallucination stage (Section 3.3), iii) without using the latent sharing structure (Section 3.2.2), and iv) without using pre-masking in $\mathcal{L}_{\mathrm{per}}$ (Section 3.2.1). + +Optimizing with a single view guide $I_{\mathrm{guide}}^{\mathrm{face}}$ (Config i) yields inaccurate hairstyles (Figure 4-A). Without optimization in W (Config ii), the method produces unrealistic hair results with various artifacts, e.g., sharp edges, unnatural hair shapes (Figure 4-B) or structures (Figure 4-C). This is + +![](images/729380b90aa568c3d848755e24b9051b8719b8d4f73e847ef78a1e113c13e20c.jpg) +Figure 5. Failure cases; (A) the semantic regions are incorrect; (B) the facial keypoint is incorrect; (C) the EG3D projection is incorrect; (D) the reference hair color is too similar to the background color; (E) lighting looks unnatural; (F) eccentric hairstyle. + +![](images/3e400c49cc39ffac05055b02a5ea0fa5ddf4be1fd2187e6a0ffe2c5c562fcbc7.jpg) + +![](images/5f24128bd84b12ef1ccfb0e41c4a40c9b326905223e08c6503fb302d87eacc44.jpg) + +because the second stage, which has higher image fitting capability, tries to fit the initial rough estimation $I_{\mathrm{guide}}$ . Without the latent sharing structure, the hair detail cannot be shared from the reference hair view ( $\mathcal{L}_{\mathrm{sim}}$ alone is not strong enough to force consistency) (Figure 4-A), resulting in inaccurate hair colors or structures (Figure 4-C). + +Without pre-masking in $\mathcal{L}_{\mathrm{per}}$ , the boundary of the face and hair regions is forced to be the same as in $I_{\mathrm{guide}}$ , leading to visible and sharp seams between the face and hair or between different facial features. Here the optimizer fails to refine the face-hair boundary to make the results look natural (Figure 4-D), and the original details, such as the background from $I_{f}$ cannot be seamlessly blended (Figure 4-E). + +# 5. Conclusion and Limitations + +We have presented a flexible and effective hairstyle transfer system that can handle a variety of challenging in-the-wild scenarios and produce perceptually convincing hairstyle transfer results. Our user study shows that human evaluators prefer our results over previous methods across all tested scenarios. Nevertheless, our method can still fail in certain scenarios, for example, when the face and hair are too eccentric (Figure 5). We refer to Appendix G for more details. + +# References + +[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE international conference on computer vision, pages 4432-4441, 2019. 2 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 4 +[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 3 +[4] Ironhammer america. File:elizabeth-olsen-1632123202.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth-olsen-1632123202.jpg/, 2022.17 +[5] Peter Baylies. Stylegan-encoder @ONLINE. https://github.com/pbaylies/stylegan-encoder/, 2019.2 +[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3, 11 +[7] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2, 5 +[8] Chaeyeon Chung, Taewoo Kim, Hyelin Nam, Seunghwan Choi, Gyojung Gu, Sunghyun Park, and Jaegul Choo. Hairfit: Pose-invariant hairstyle transfer via flow-based hair alignment and semantic-region-aware inpainting. arXiv preprint arXiv:2206.08585, 2022. 2 +[9] Gordon Correll. File: julianne moore (15011443428) (2). jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_(15011443428)_ (2). jpg/, 2014. 17 +[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7865-7874, 2022. 16 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. 1, 2 +[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2 +[13] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only chang- + +ing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 2 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Proc. NeurIPS, 2020. 2, 5 +[16] jerryhouuu. Face-yaw-roll-pitch-from-posedestimation-using-opencv @ONLINE. https://github.com/jerryhouuu/Face-Yaw-Roll-Pitch-from-Pose-Estimation-using-OpenCV/, 2018.13 +[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 +[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5, 7 +[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 4, 5, 12, 13 +[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21 +[21] Taewoo Kim, Chaeyeon Chung, Sunghyun Park, Gyojung Gu, Keonmin Nam, Wonzo Choe, Jaesung Lee, and Jaegul Choo. K-hairstyle: A large-scale korean hairstyle dataset for virtual hair editing and hairstyle classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1299-1303. IEEE, 2021. 2 +[22] Davis E. King. Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research, 10:1755–1758, 2009. 3, 8, 11 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 13 +[24] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 11 +[25] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2445, 2020. 2, 4 +[26] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 2 +[27] nicolas genin. File:julianne moore 2009 venice film festival.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_2009_Venice_Film_Festival.jpg/, 2009.17 + +[28] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 13 +[29] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 6, 11, 12, 13 +[30] Rohit Saha, Brendan Duke, Florian Shkurti, Graham Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In CVPR, 2021. 1, 2, 4, 5, 6, 19, 20, 21 +[31] David Shankbone. File:dwayne johnson at the 2009 tribeca film festival.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Dwayne_Johnson_at_the_2009_Tribeca_Film_Festival.jpg/, 2009.17 +[32] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 2 +[33] Gage Skidmore. Tom holland speaking at the 2016 san diego comic-con international in san diego, california. @ONLINE. https://commons.wikimedia.org/wiki/File: Tom_Holland_by_Gage_Skidmore.jpg/, 2016.1 +[34] Gage Skidmore. File:elizabeth olsen by gage skidmore 2.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth_Olsen_by_Gage_Skidmore_2.jpg/, 2019.17 +[35] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. arXiv preprint arXiv:2010.16417, 2020. 1, 2 +[36] tenasia10. File:blackpinkLisa gmp 240622.png trends @ONLINE. https://commons.wikimedia.org/wiki/File:Blackpink_Lisa_GMP_240622.png/, 2022.17 +[37] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3 +[38] GABI THORNE. Jennifer aniston reacts to tiktok's best (and most absurd) trends @ONLINE. https://www.allure.com/story/jennifer-aniston-reacts-to-tiktok-trends/, 2022.17 +[39] Albert Truuväärt. File:re kaja kallas (cropped).jpg @ONLINE. https://commons.wikimedia.org/wiki/File:RE_Kaja_Kallas_(cropped). jpg/, 2011.17 +[40] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 5914–5922, 2019. 2 +[41] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting gans with consecutive images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13910-13918, 2021. 2 + +[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. 3, 11 +[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4 +[44] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proc. ECCV, 2022. 2 +[45] Tao Zhou, Changqing Zhang, Chen Gong, Harish Bhaskar, and Jie Yang. Multiview latent space learning with feature redundancy minimization. IEEE transactions on cybernetics, 50(4):1655-1668, 2018. 2 +[46] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. arXiv preprint arXiv:2106.01505, 2021. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21 +[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 651-667. Springer, 2022. 3 +[48] zllrunnings. face-parsing.pytorch @ONLINE. https: //github.com/zllrunning/face-parsing. PyTorch/, 2019.11 \ No newline at end of file diff --git a/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/images.zip b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b31b3047b7d8611145e0841e06c64ab7374191a --- /dev/null +++ b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83fe8f761fb13f827aaf935c599f2c6d70ff27870f31be2de6c89bef99f13d08 +size 570994 diff --git a/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/layout.json b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a288bec8e31c5c4381b383a5a36e2494b1bf9e0 --- /dev/null +++ b/2023/StyleGAN Salon_ Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer/layout.json @@ -0,0 +1,11299 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 515, + 140 + ], + "type": "text", + "content": "StyleGAN Salon: Multi-View Latent Optimization for Pose-Invariant Hairstyle Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 167, + 552, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 167, + 552, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 167, + 552, + 182 + ], + "type": "text", + "content": "Sasikarn Khwanmuang\\* Pakkapon Phongthawee\\* Patsorn Sangkloy\\* Supasorn Suwajanakorn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 185, + 471, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 185, + 471, + 200 + ], + "spans": [ + { + "bbox": [ + 135, + 185, + 471, + 200 + ], + "type": "text", + "content": "VISTEC, Thailand Phranakhon Rajabhat University, Thailand" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 205, + 520, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 205, + 520, + 216 + ], + "spans": [ + { + "bbox": [ + 81, + 205, + 520, + 216 + ], + "type": "text", + "content": "{sasikarn.k_s18, pakkapon.p_s19, supasorn.s}@vistec.ac.th patsn.s@pnru.ac.th" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 106, + 242, + 545, + 387 + ], + "blocks": [ + { + "bbox": [ + 63, + 304, + 105, + 313 + ], + "lines": [ + { + "bbox": [ + 63, + 304, + 105, + 313 + ], + "spans": [ + { + "bbox": [ + 63, + 304, + 105, + 313 + ], + "type": "text", + "content": "Input Face" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 242, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 545, + 387 + ], + "type": "image", + "image_path": "1facb3b704e8e0370a119f7004f0d6c0b61389ee356caa02c58565d5813fa96d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 536, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 536, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 536, + 399 + ], + "type": "text", + "content": "Figure 1. Our method can transfer the hairstyle from any reference hair image in the top row to Tom Holland [33], in the second row." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 408, + 192, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 408, + 192, + 420 + ], + "spans": [ + { + "bbox": [ + 143, + 408, + 192, + 420 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 437, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 290, + 700 + ], + "type": "text", + "content": "Our paper seeks to transfer the hairstyle of a reference image to an input photo for virtual hair try-on. We target a variety of challenges scenarios, such as transforming a long hairstyle with bangs to a pixie cut, which requires removing the existing hair and inferring how the forehead would look, or transferring partially visible hair from a hat-wearing person in a different pose. Past solutions leverage StyleGAN for hallucinating any missing parts and producing a seamless face-hair composite through so-called GAN inversion or projection. However, there remains a challenge in controlling the hallucinations to accurately transfer hairstyle and preserve the face shape and identity of the input. To overcome this, we propose a multi-view optimization framework that uses two different views of reference composites to semantically guide occluded or ambiguous regions. Our optimization shares information between two poses, which allows us to produce high fidelity and realistic results from incomplete references. Our framework produces high-quality results and outperforms prior work in a user study that consists of significantly more challenging hair transfer scenarios than previously studied. Project page: https://stylegan-salon.github.io/." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 408, + 386, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 386, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 386, + 420 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 432, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 504 + ], + "type": "text", + "content": "What makes Jennifer Aniston keep her same hairstyle for over three decades? Perhaps she likes the classic, or perhaps changing her hairstyle is a decision too high-stakes that she could later regret. Unlike garments or makeup, trying on a new hairstyle is not easy, and being able to imagine yourself in different hairstyles could be an indispensable tool." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 548, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 548, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 548, + 662 + ], + "type": "text", + "content": "Recent approaches for hairstyle transfer, StyleYourHair [20], Barbershop [46], LOHO [30], and Michiganan [35], allow users to manipulate multiple hair attributes of an input image, such as appearance, shape, or color by providing a reference image for each different attribute. These methods [20, 30, 46] rely on a generative adversarial network [11], specifically StyleGAN2 [19], which can synthesize highly realistic face images. Their key idea, which also forms the basis of our method, is to leverage the realistic face distribution learned by StyleGAN and search for a hairstyle-transfer output whose latent code lies within the learned distribution using optimization (commonly known as GAN projection or inversion)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Our extensive study on these state-of-the-art techniques still reveal several unsolved challenges for in-the-wild hairstyle transfer. One of the main challenges is when the reference hair comes from a person with a very different head" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8609" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "pose or facial shape. In this case, the transfer result often degrades significantly [30, 46]. HairFIT [8] and the recently proposed StyleYourHair [20] both attempt to solve this using an additional alignment step to align the pose of the target hair to the input face. In HairFIT [8], this is done explicitly via a flow-based warping module for hair segmentation, but this requires training on multi-view datasets [21, 26]. StyleYourHair [20] avoids this issue and also improves upon HairFIT's results by optimizing for an aligned pose within StyleGAN2's latent space using distances between the detected facial keypoints. While StyleYourHair can handle a certain degree of misalignment, it often struggles to preserve details of the reference hair texture, especially for intricate and non-straight hairstyles, e.g., in Figure 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 289, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 337 + ], + "type": "text", + "content": "In general, we have observed that there is a trade-off between hallucinating new details, which is crucial for handling highly different poses, and preserving the original texture from the reference images. These two goals are often at odds with each others. This trade-off is also evident in the results from StyleYourHair, as reported in [20] and as shown in table 3, where BarberShop [46] can still produce better results, for example, when the pose difference is small." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 456 + ], + "type": "text", + "content": "We tackle this dilemma by performing a multi-stage optimization. This serves two purposes: first, to hallucinate new details necessary for aligning the poses, and second, to recover face-hair details from the original input images. We preserve these details in a form of two guide images in both viewpoints, which will be jointly optimized to allow new details to be filled while retaining face-hair texture in the original pose. Our pose alignment is done both explicitly via 3D projection [6] on RGB images, and implicitly via latent code(s) sharing during our multi-view optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 457, + 242, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 457, + 242, + 469 + ], + "spans": [ + { + "bbox": [ + 59, + 457, + 242, + 469 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 478, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 478, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 287, + 528 + ], + "type": "text", + "content": "1. We propose StyleGAN Salon: a pose-invariant hairstyle transfer pipeline that is flexible enough to handle a variety of challenging scenarios including, but not limited to, bangs/hat removal and background inpainting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 536, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 536, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 54, + 536, + 287, + 585 + ], + "type": "text", + "content": "2. Unlike previous works, our method operates entirely on RGB images, which are more flexible than segmentation masks. This allows us to first draft the output and then refine them via multi-stage optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 287, + 655 + ], + "type": "text", + "content": "3. We introduce multi-view optimization for hairstyle transfer which incorporates 3D information to align the poses for both face and hair images. Our method leverages both views to help preserve details from the original images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 287, + 714 + ], + "type": "text", + "content": "4. We thoroughly analyze the results in several experiments, including a user study with detailed breakdown into various challenging scenarios. Our method shows superior results over existing works in all scenarios." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 547, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 187 + ], + "type": "text", + "content": "Generative Adversarial Networks. Beginning with [11, 18], StyleGANs [17-19] have shown great success in 2D image generation by learning realistic training data distribution corresponding to a fixed low-dimensional distribution, called latent code. The learned latent code can be applied to a downstream task such as image manipulation [13, 15, 32, 40]. Recent work on [6, 12, 44] also explore 3D aware architecture in StyleGAN, resulting in multi-view consistent images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 187, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 547, + 258 + ], + "type": "text", + "content": "For hairstyle transfers, MichiGAN [35] uses a conditional hair generation network that can control hair shape, structure, and appearance. Recently, HairFIT [8], a pose-invariant hairstyle transfer network, aligns the reference hair to match the input face pose using a optical flow-based hair alignment module, but requires training on multi-view dataset [21, 26]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": "StyleGAN Latent Space and Projection Techniques. To generate an image, StyleGAN2 first maps a random latent code " + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "inline_equation", + "content": "z \\sim N(0, I)" + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": " to an intermediate latent code " + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": " in a new space called " + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": ". This code " + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": " is then replicated (18x) and used as input to each layer of the generator that controls details at different scales through weight demodulation. Collectively, these replicated latent codes are referred to as " + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "inline_equation", + "content": "w^{+} \\in \\mathbb{R}^{18 \\times 512}" + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": " in the extended latent space " + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{+}" + }, + { + "bbox": [ + 304, + 259, + 547, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "content": "Pivotal Tuning Inversion (PTI) [29] further improves the projection quality by directly tuning generator weights around the optimized latent code " + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "content": " using their regularization technique. Other techniques [1,2,5] optimize the latent code " + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "content": " in each layer separately to better match the input image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "text", + "content": "PIE [37] introduces hierarchical optimization for semantic attributes editing that first optimizes the latent code in the " + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "text", + "content": " space, then transfers the code to the " + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "text", + "content": " space and continues the optimization in that space. We adopt similar hierarchical optimization that uses both " + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 415, + 547, + 521 + ], + "type": "text", + "content": ". Part of our method is also inspired by PULSE [25], which reconstructs a high-resolution image from a small reference image (32x32) by searching for the closest image in the StyleGAN latent space that resembles the low-resolution reference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 522, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 582 + ], + "type": "text", + "content": "Some methods [7, 41, 45] project multiple images into latent space simultaneously. However, all of their inputs are complete, whereas our method requires hair information from the reference hair image and additional information from the input face image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "StyleGAN-Based Hairstyle Transfer. LOHO [30] adopts loss functions from Image2StyleGAN++ [2] to combine face-hair inputs into a hairstyle transfer output with their orthogonalization optimization technique which reduces conflicts between multiple loss functions. Barbershop [46] first predicts semantic regions of both inputs and uses them to create a \"target segmentation mask\" of the output with rule (in the original paper) or GAN-based inpainting (in their official code). Then, they optimize two separate latent codes in " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " space, one for matching the face and the other for hair, while conforming to the target segmentation mask. To" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8610" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "preserve the original details, the latent optimization is done in their proposed " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{F} / \\mathcal{S}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " space, which replaces the first seven blocks of " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " space with the corresponding activation maps of StyleGAN's convolution layers. To improve Barbershop's capacity to handle unaligned input poses, StyleYourHair [20] first aligns the reference hair to match the pose of the input face via the proposed local-style-matching loss. However, this alignment often leads to an unrealistic hair shape or inaccurate hair texture results. In contrast, our method simultaneously optimizes for two guides in face and hair views, resulting in a better hair texture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": "Instead of estimating the final output from the segmentation mask [20, 46], our multi-view optimization uses face-hair composites in RGB space to overcome this problem, and produces results that better preserve the input facial structure and hairstyle across a wider, more challenging range of scenarios. Concurrent work, HairNet [47], arrives at a similar goal of removing the target segmentation mask via a two-step process that involves baldification using StyleFlow [3] followed by another network to transfer the hairstyle." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 360 + ], + "type": "text", + "content": "Nevertheless, handling pose differences is crucial for successful hairstyle transfer; our method incorporates 3D rotation to preserve the geometric consistency and systematically evaluate this aspect." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 146, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 146, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 146, + 384 + ], + "type": "text", + "content": "3. StyleGAN Salon" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "content": "Given two input images " + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "content": ", the goal is to transfer the hair from the reference hair image " + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "content": " into the face in the image " + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "content": ", while preserving all remaining details in " + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 46, + 390, + 288, + 438 + ], + "type": "text", + "content": ", including identity, clothing, and background." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "type": "text", + "content": "The key idea of our approach is to guide the optimization on the learned latent space of StyleGAN2 with two \"guide\" images, which represent rough composites of the final output based on simple cut-and-paste in the viewpoint of " + }, + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 438, + 289, + 510 + ], + "type": "text", + "content": ". We leverage EG3D to construct these guided images in a geometrically consistent way, described in Section 3.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": "Optimization on StyleGAN2's latent space is commonly performed on either the original latent space " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": " or the extended latent space " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": ". Optimizing on " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": " space generally leads to more realistic results by staying within the original latent space, whereas optimizing on " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": " allows a closer match to the reference [37]. In hairstyle transfer, it is important to stay faithful to the input images and preserve important details such as hair texture, face identity, and background. However, optimizing on " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": " will lead to poor results because our guide images are rough and unrealistic composites. Thus, we propose to optimize on " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": " space followed by " + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 510, + 289, + 665 + ], + "type": "text", + "content": " space, similar to a technique in PIE [37] used for editing semantic attributes of an image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "Our optimization incorporates both guide images from the two viewpoints, detailed in Section 3.2. Section 3.3 and Section 3.4 cover details of our optimization on " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": ", respectively. Finally, we also optimize StyleGAN2 weights" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "while freezing the latent codes (Section 3.5) using PTI [29] to further improve detail fidelity. Figure 2 shows an overview of our complete pipeline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 115, + 476, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 115, + 476, + 129 + ], + "spans": [ + { + "bbox": [ + 305, + 115, + 476, + 129 + ], + "type": "text", + "content": "3.1. Constructing the Guide Images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": "The purpose of our guide images is to provide an initial estimate of how the hair would look on " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": ". We achieve this using a cut-and-paste composite of the face and background from " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": " and the hair from " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": ". To better handle a potentially large shift in viewpoint between " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": ", we propose to leverage EG3D [6] to help generate geometrically consistent guide images. We argue that using multi-view guide images, each in the pose of " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 133, + 547, + 241 + ], + "type": "text", + "content": ", helps preserve details that could otherwise be lost from using a single viewpoint alone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "text", + "content": "A straightforward approach is to simply project " + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "text", + "content": " into the EG3D [6] latent space, and use their proposed neural rendering pipeline to render it in the view of " + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "text", + "content": " (and vice versa). However, while the projection can produce a geometrically consistent 3D shape, we found that the resulting texture is not very accurate. To address this, we replace any visible regions in the texture from " + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "text", + "content": "'s viewpoint with the original pixels of " + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 304, + 241, + 547, + 348 + ], + "type": "text", + "content": ", while leaving the rest of the texture as the projected texture from EG3D's rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 348, + 547, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 432 + ], + "type": "text", + "content": "We also additionally apply uniform scaling and translation to match the faces' widths and centers, which are computed based on detected facial keypoints [22]. These keypoints, along with semantic regions from [42], are also used to handle various corner cases, such as re-painting unwanted hair regions. We refer to Appendix A for more details of this operation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "text", + "content": "The entire process is done for both " + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "text", + "content": " viewpoints, resulting in a pair of guide images " + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}} = [I_{\\mathrm{guide}}^{\\mathrm{face}}, I_{\\mathrm{guide}}^{\\mathrm{hair}}]" + }, + { + "bbox": [ + 304, + 433, + 547, + 504 + ], + "type": "text", + "content": ". We emphasize that by leveraging our multi-view optimization (Section 3.2), these guide images do not need to be precise or realistic to produce convincing final results, as demonstrated in the first row of Figure 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 510, + 480, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 480, + 523 + ], + "type": "text", + "content": "3.2. Multi-View Latent Optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "text", + "content": "Our guide images provide complementary information about the target hair and face, albeit from different poses. However, each guide image is only fully accurate in regions with the original pixels seen in the original viewpoint and not warped by EG3D. Thus, our optimization goal is to combine information from both guide images to generate a final output that accurately captures the realistic hair from " + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "text", + "content": ", as well as other details from " + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "text", + "content": ". We achieve this using multiple loss functions that attend to both viewpoints with different spatial emphasis. Specifically, we perform multi-view latent optimization on " + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "inline_equation", + "content": "w / w^{+}" + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "text", + "content": " (and stochastic noise maps " + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 529, + 547, + 674 + ], + "type": "text", + "content": ") that fits both guide images:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 681, + 547, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 681, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 354, + 681, + 547, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {\\{w / w ^ {+}, n \\}} \\sum_ {\\mathrm {i}} ^ {[ \\text {f a c e , h a i r} ]} \\mathcal {L} _ {\\text {l o s s}} \\left(O ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})}\\right), \\tag {1}", + "image_path": "03508c741751adf4d3a00f78c1035055e5d945893eb8a6d5b20f744d94b65362.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8611" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 202 + ], + "type": "image", + "image_path": "4d7cdeef94bda516a58d021f9cf161129d9929addd054f442dd842c985723a89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": "Figure 2. Overview of StyleGAN Salon: We first align the input face " + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": " and reference hair " + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": " and use them to construct guide images " + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 46, + 204, + 547, + 227 + ], + "type": "text", + "content": ", in two different viewpoints, which specifies the target appearance for each output region. (see Section 3)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "inline_equation", + "content": "O^{(\\mathrm{i})}" + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "content": " denotes the output of the StyleGAN2's generator that takes a latent code " + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "inline_equation", + "content": "w^{(\\mathrm{i})}" + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "inline_equation", + "content": "w^{+\\mathrm{(i)}}" + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "content": " and a stochastic noise input " + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "inline_equation", + "content": "n^{(\\mathrm{i})}" + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loss}}" + }, + { + "bbox": [ + 46, + 247, + 288, + 283 + ], + "type": "text", + "content": " is a sum of all our loss functions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 291, + 287, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 291, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 291, + 287, + 342 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {l o s s}} = \\sum_ {\\mathrm {j}} ^ {[ \\mathrm {f}, \\mathrm {h}, \\mathrm {b g} ]} \\lambda_ {\\mathrm {p}} ^ {(\\mathrm {j})} \\mathcal {L} _ {\\text {p e r}} ^ {(\\mathrm {j})} (2) \\\\ + \\lambda_ {\\mathrm {g}} \\mathcal {L} _ {\\text {g l o b a l}} + \\lambda_ {\\mathrm {i}} \\mathcal {L} _ {\\text {i n i}} + \\lambda_ {\\varepsilon} \\mathcal {L} _ {\\varepsilon} + \\lambda_ {s} \\mathcal {L} _ {\\text {s i m}}, (3) \\\\ \\end{array}", + "image_path": "67d6ff4bb8c740df27ce712033d64ff192afc0e6c681b734682e33e770e1ef2c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\lambda_{(\\cdot)}" + }, + { + "bbox": [ + 46, + 350, + 287, + 374 + ], + "type": "text", + "content": " are balancing weights, and [f, h, bg] refer to face, hair, and background. We next explain each loss function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 389, + 144, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 144, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 144, + 399 + ], + "type": "text", + "content": "3.2.1 Loss Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": "The objective of our " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": " optimization is to generate an output image that fits the corresponding " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": " without appearing unrealistic. The main challenge is that our guide images are unrealistic and can contain various unwanted artifacts. We design our loss functions to be tolerant of the imprecise nature of our guide images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "text", + "content": "Masked Perceptual Loss: This loss is based on Learned Perceptual Image Patch Similarity (LPIPS) [43], which compares two input images in a deep feature space. Similar to [20, 30, 46], we apply a binary mask to indicate regions of interest in the deep feature space of the two input images. We use " + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{LPIPS}}(I_1, I_2; M)" + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "text", + "content": " to denote this LPIPS computation between two images " + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "inline_equation", + "content": "(I_1, I_2)" + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "text", + "content": " with mask " + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "text", + "content": ". However, we observe that applying this masking operation after the feature computation is insufficient to disregard unwanted regions. This is due to the fact that LPIPS is a patch-based similarity loss, and regions outside of the mask can still affect the loss computation. Based on this principle, we propose to apply additional pre-masking to mask out ambiguous regions in " + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 46, + 479, + 288, + 659 + ], + "type": "text", + "content": " that should not be trusted. Specifically, our Masked Perceptual Loss is the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 666, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 287, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e r}} ^ {\\mathrm {i}, \\mathrm {r o i}} = \\Lambda_ {\\mathrm {r o i}} ^ {(\\mathrm {i})} L _ {\\mathrm {L P I P S}} \\left(O ^ {(\\mathrm {i})} \\odot \\neg M _ {\\mathrm {r o n i}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})} \\odot \\neg M _ {\\mathrm {r o n i}} ^ {(\\mathrm {i})}; M _ {\\mathrm {r o i}} ^ {(\\mathrm {i})}\\right), \\tag {4}", + "image_path": "ddb1535be792fd5305f2e5ec841e4533f471e1e9a896450ff5f88caee36e1934.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roi}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the mask for region of interest (ROI), and " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roni}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the mask for regions of not interest (RONI). " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "(.)\\odot \\neg M_{\\mathrm{roni}}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": "is a simple element-wise multiplication, which effectively excludes these regions from the loss by setting them to 0 in both input images. And " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\Lambda_{\\mathrm{roi}}^{\\mathrm{(i)}}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": " is a balancing weight in (i) viewpoint, which is set higher when the region of interest covers the original pixels of " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": "). For instance, the loss that attends to face in " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": " has a higher weight than the face loss in " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": ". We consider 3 regions of interest: " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{per}}^{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{per}}^{\\mathrm{h}}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{per}}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": " for face, hair, and background regions. We refer to Section 3.3 and 3.4 for details on these mask generation for " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{+}" + }, + { + "bbox": [ + 304, + 247, + 547, + 371 + ], + "type": "text", + "content": "optimization, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "type": "text", + "content": "Global loss: This loss function attempts to match the overall appearance of the output to " + }, + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 305, + 372, + 545, + 396 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 402, + 545, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 402, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 372, + 402, + 545, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g l o b a l}} = L _ {\\text {M S E}} ^ {3 2} (O, I _ {\\text {g u i d e}}). \\tag {5}", + "image_path": "a39618df0c6e4a3e79b11c67e03280af1bbd2cb28a531f4e67bc4c0031428bd7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MSE}}^{32}" + }, + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "text", + "content": " is the mean square error computed on " + }, + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "text", + "content": " downsampled input images. The goal of the downsampling is to reduce the effect of matching visible seams from higher resolution " + }, + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 304, + 422, + 547, + 494 + ], + "type": "text", + "content": " because those seams become imperceptible once downsampled. This loss is inspired by a similar idea used for image super-resolution by Menon et al. [25]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "text", + "content": "Initial Value Loss: To prevent the output from deviating too far from the learned distribution and becoming unrealistic, we force the optimized latent code(s) " + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "text", + "content": " to be close to its initial value " + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "inline_equation", + "content": "w_{0}" + }, + { + "bbox": [ + 304, + 495, + 547, + 542 + ], + "type": "text", + "content": " through L2 loss, as used in [2]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 549, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 549, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 378, + 549, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i n i}} = \\left\\| w _ {1: 1 8} - w _ {0} \\right\\| _ {2} ^ {2}. \\tag {6}", + "image_path": "0d89e937f5b7b4b5459e8be53914926ac0ec4bd5b60a99663f7438204099b52d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": "We use the estimated mean of " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " latent space as initial values for optimizing " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ". This is computed by averaging many latent codes " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "w_{i} = f(z_{i})" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " drawn randomly through " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "z\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " and the StyleGAN's mapping network " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ". Then the optimized latent code " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " from the first stage becomes the initial value for the optimization in " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " extended latent space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": "Noise Regularization Loss: We also use the noise regularization loss " + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\varepsilon})" + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": " proposed in StyleGAN2 [19] to ensure that the noise maps will be used for capturing the stochastic variations, and not for encoding the content meant to be captured by the latent code. We refer to StyleGAN2 for the details of this loss." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8612" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": "Latent Similarity Loss: To enable information sharing between the two guide images (" + }, + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": ") from two viewpoints, we force both their latent codes to be close together by using the following L2 loss, similarly to [7]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 129, + 288, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 129, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 110, + 129, + 288, + 144 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m}} = \\left\\| w ^ {(\\text {f a c e})} - w ^ {(\\text {h a i r})} \\right\\| _ {2} ^ {2}, \\tag {7}", + "image_path": "d9dfb27c44a158c7300ce84390c1478bd5fbb8060f54a0aa9956c5e333a4396e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "content": " can be in " + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "content": " space during " + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "content": " optimization or replaced with " + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "content": " during " + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 47, + 152, + 288, + 177 + ], + "type": "text", + "content": " optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 192, + 170, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 170, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 170, + 205 + ], + "type": "text", + "content": "3.2.2 Sharing Latent Code" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "text", + "content": "In StyleGAN2, latent codes in early layers have been shown to correspond to high level concepts (e.g. human pose) whereas the remaining layers capture low-level information (e.g. colors) [15, 18]. This motivates the optimization in " + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "text", + "content": " extended space, as each layer can be optimized for different aspects of the image. Similarly, we optimize latent codes to match guide images of the same person and hairstyle but in different poses. Rather than optimizing two latent codes independently, we can enforce information sharing between the two poses by constraining the latent codes of the last few layers (" + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "inline_equation", + "content": "w_{l:18}" + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "text", + "content": ") to be the same. We set " + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 211, + 287, + 366 + ], + "type": "text", + "content": " to be 4 in all our experiments, loosely based on the head rotation experiments shown in GANSpace [15]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": "Sharing " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": " Latent Code: Optimization in the space of StyleGAN2 latent code " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": " is generally done as a single latent code " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{512})" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": ", which are duplicated (18x) and fed into all the layers. In our case, we also optimize for a single latent code " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": " but only duplicate them for the first " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": " layers " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "(w_{1:l})" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": ". The remaining layers " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "(w_{l:18})" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": " are interpolated from the current latent codes of each guide image with a coefficient " + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 367, + 288, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 459, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 459, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 97, + 459, + 287, + 475 + ], + "type": "interline_equation", + "content": "w _ {l: 1 8} = \\alpha w _ {1: l} ^ {\\text {(f a c e)}} + (1 - \\alpha) w _ {1: l} ^ {\\text {(h a i r)}}. \\tag {8}", + "image_path": "fc837aafd7c59021a22cb80a53dae95bf88ded801b33b6f4d86cfe1d234827e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "type": "text", + "content": "We randomize this coefficient to make the latent codes stay within the space of " + }, + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "type": "text", + "content": ". This random interpolation forces the latent code for each guide image to be partly similar while still producing realistic results from " + }, + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 483, + 287, + 530 + ], + "type": "text", + "content": " latent space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "text", + "content": "Sharing " + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "text", + "content": "Latent Code: Optimization in " + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{+}" + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "text", + "content": "extended space is straightforward, as latent code for each layer can already be separately optimized. We simply share the last few layers " + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "inline_equation", + "content": "(w_{l:18})" + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "text", + "content": " of both " + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 47, + 531, + 288, + 603 + ], + "type": "text", + "content": "latent codes during optimization. We refer to the Appendix A.5 for more details on the optimization process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 610, + 280, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 280, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 280, + 624 + ], + "type": "text", + "content": "3.3. " + }, + { + "bbox": [ + 47, + 610, + 280, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 610, + 280, + 624 + ], + "type": "text", + "content": " Optimization: Hallucinate Missing Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "content": "This stage aims to hallucinate details not currently visible in the guide images. The output images of this stage are " + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "O_{1}^{\\mathrm{face}}" + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "O_{1}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "content": ", which are essentially the projection of the unrealistic " + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "content": " into the real image distribution learned in StyleGAN2 [19]. This is done by optimizing on " + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 629, + 287, + 714 + ], + "type": "text", + "content": " latent space to fit each guide image with the objective function (Equation 2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "While our guide images (" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ") capture the overall appearance of the desired outputs, they still lack details in certain regions. These unknown regions may correspond to unseen facial features that were occluded by the hair in " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " or incomplete reference hair from " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " caused by, e.g., a hat or image cropping. These details will need to be hallucinated and seamlessly blended with the rest of the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": "To accomplish this, we design our pre-masking in Masked Perceptual Loss " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{f}},\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{h}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " ; Equation 4) to include all the unknown regions that need to be hallucinated. Specifically, " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roni}}^{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " represents the face region occluded by the hair in " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " , e.g., the forehead behind the bangs or the ears that should become visible in the final output. Analogously, " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roni}}^{\\mathrm{h}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " represents the hair region occluded by other objects (e.g., a hat) or not visible in " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " due to image cropping. We set " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roi}}^{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roi}}^{\\mathrm{h}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " to be segmentation masks for face and hair regions " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "(M_{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{h}})" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " , respectively. Both " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roi}}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "\\neg M_{\\mathrm{roni}}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " is set to be the background regions in " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " not covered by the transferred hair " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "(M_{\\mathrm{bg}})" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " . These masks are constructed by composing different semantic regions (union, intersection, etc.) from " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 304, + 156, + 546, + 325 + ], + "type": "text", + "content": " , detailed in Appendix A.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "text", + "content": "We note that " + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roi}}" + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "text", + "content": " is applied on the deep feature maps, similar to [20, 30, 46], while " + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "inline_equation", + "content": "\\neg M_{\\mathrm{roni}}" + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "text", + "content": " is applied on raw RGB images. Using both masking techniques is crucial to our natural and seamless blending. The additional pre-masking allows the hair shape (or face shape) in the output to be different from " + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "text", + "content": "'s and freely grow outward outside " + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{roi}}" + }, + { + "bbox": [ + 304, + 325, + 547, + 408 + ], + "type": "text", + "content": " if this leads to a more natural result (see Figure 4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 415, + 541, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 541, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 541, + 427 + ], + "type": "text", + "content": "3.4. " + }, + { + "bbox": [ + 305, + 415, + 541, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 305, + 415, + 541, + 427 + ], + "type": "text", + "content": " Optimization: Recover Face-Hair Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": "The output of the previous stage, " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": ", may still look different from the input person and not yet capture the hair details from the reference. This stage aims at refining these output images to better reproduce the hair details from " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " and the rest from " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "I_f" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": ". The optimization is done in the extended " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " space with respect to " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\{w^{+},n\\}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": ". In other words, we allow " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "w\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " that was previously replicated to each layer to be optimized separately as " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "w^{+}\\in \\mathbb{R}^{18\\times 512}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "O_2^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "O_2^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " to be the output of this stage generated by StyleGAN2 from our optimized latent codes in " + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": " extended latent space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": "We update the target images for the optimization from " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " to the more complete versions based on " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " from the first stage. However, because the first stage aims at hallucinating new details into the pre-masking regions, it may lose the original texture details. To address this, we replace the known regions in " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{face}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{hair}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " with the correct details from the original " + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 304, + 565, + 546, + 651 + ], + "type": "text", + "content": " by setting" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 346, + 658, + 546, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 658, + 546, + 671 + ], + "spans": [ + { + "bbox": [ + 346, + 658, + 546, + 671 + ], + "type": "interline_equation", + "content": "I _ {\\text {n e w} \\text {g u i d e}} \\leftarrow I _ {\\text {g u i d e}} \\odot M _ {\\mathrm {c}} + O _ {1} \\odot \\neg M _ {\\mathrm {c}}, \\tag {9}", + "image_path": "73773c46f2bbb53800cc29f002a0084c868c05f6c838d66de874d1a2b65560e6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "M_{c} = M_{\\mathrm{f}} \\cup M_{\\mathrm{h}} \\cup M_{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": ", which corresponds to the regions of interest in our Masked Perceptual Loss where we want to match with " + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": ". In short, we simply copy the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8613" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "hallucinated parts from " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "O_1^{\\mathrm{hair}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": ", and combine them with the part in " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " that are known to be correct." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": "For our Masked Perceptual Loss, we use the same masking procedure as the previous stage (Section 3.3). However, we observe that the projection from EG3D, which produces " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": ", generally is inferior to the known details from the original images. For example, " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " contains more details on the hair than " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " due to the latter being hallucinated from the 3D rotation by EG3D's rendering pipeline. To de-prioritize these inaccurate details, we blur both " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " when computing the perceptual loss for the less accurate regions of interest. Concretely, we only blur images when computing " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{f}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " optimization and when computing " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{per}}^{\\mathrm{h}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 96, + 290, + 245 + ], + "type": "text", + "content": " optimization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 250, + 185, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 185, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 185, + 262 + ], + "type": "text", + "content": "3.5. Pivotal Tuning Inversion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": "The purpose of this stage is to further refine the output images " + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "inline_equation", + "content": "(O_2^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "inline_equation", + "content": "O_2^{\\mathrm{hair}})" + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": " by allowing the optimization of the StyleGAN's weights " + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": " (while fixing the optimized latent code " + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{optimized}}" + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": "). We closely follow the proposed optimization in PTI [29] with the same objective function. However, we update the reconstruction loss to reflect the goal of our task, which is to match the original details from " + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 46, + 268, + 289, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 360, + 287, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 360, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 49, + 360, + 287, + 413 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p t i}} = \\sum_ {\\mathrm {i}} ^ {[ \\text {f a c e}, \\text {h a r i} ]} \\left(L _ {\\mathrm {L P I P S}} \\left(O _ {\\text {t u n e}} ^ {(\\mathrm {i})} \\odot M _ {\\text {r a w}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})} \\odot M _ {\\text {r a w}} ^ {(\\mathrm {i})}; M _ {\\text {r a w}} ^ {(\\mathrm {i})}\\right) \\right. \\\\ + L _ {\\mathrm {M S E}} ^ {3 2} \\left(O _ {\\text {t u n e}} ^ {(\\mathrm {i})}, I _ {\\text {g u i d e}} ^ {(\\mathrm {i})}; \\neg M _ {\\text {r a w}} ^ {(\\mathrm {i})}\\right)), \\tag {10} \\\\ \\end{array}", + "image_path": "55ebc4777338eb45be923707c7d1d7b0b27f07003198cc62c14a6654d6e4466e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{tune}}" + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": " is the generated image using the tuned weights " + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": ", the mask " + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{raw}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "inline_equation", + "content": "(M_{\\mathrm{bg}} \\cup M_{\\mathrm{f}})" + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{raw}}^{\\mathrm{hair}}" + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{h}}" + }, + { + "bbox": [ + 46, + 418, + 287, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 128, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 128, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 128, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 567 + ], + "type": "text", + "content": "In this section, we compare our method to state-of-the-art StyleYourHair [20], Barbershop [46], and LOHO [30]. Our evaluation criteria are i) user preference via a user study on a wide variety of scenarios, ii) hairstyle transfer quality, iii) hair reconstruction quality, and iv) how well the input face shape is preserved. Section 4.5 presents ablation studies on our multi-view sharing latent optimization, optimization stage, and loss functions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 568, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 617 + ], + "type": "text", + "content": "We use the official code of LOHO [30] and Barbershop [46] with the default configurations. For StyleYourHair [20], we use the configuration where the hair reference is never flipped." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 183, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 183, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 183, + 636 + ], + "type": "text", + "content": "4.1. Qualitative Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "We provide a qualitative comparison to StyleYourHair [20], Barbershop [46], and LOHO [30] in Figure 3. We observe that LOHO and Barbershop often struggle to fit the reference hair accurately when the poses are not well aligned (rows 3-5), resulting in various artifacts, such as remnants of the original hair or wrong placement of the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 82, + 545, + 345 + ], + "blocks": [ + { + "bbox": [ + 350, + 73, + 526, + 81 + ], + "lines": [ + { + "bbox": [ + 350, + 73, + 526, + 81 + ], + "spans": [ + { + "bbox": [ + 350, + 73, + 526, + 81 + ], + "type": "text", + "content": "LOHO Barbershop StyleYourHair Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 82, + 545, + 345 + ], + "lines": [ + { + "bbox": [ + 307, + 82, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 82, + 545, + 345 + ], + "type": "image", + "image_path": "54f7e8d6d84270eb96811594f93d8920ee9993f23871d939a8afb48c3990d428.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 351, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 440 + ], + "type": "text", + "content": "Figure 3. Comparison to current state-of-the-art methods for transferring hair from 1st column-top, to the face of 1st column-bottom. Our method can accurately transfer hairstyle even when the input face and hair are misaligned (3rd, 4th row). It can also hallucinate missing details such as a forehead that was previously occluded (1st row), or shorten the hair (2nd row). In general, LOHO and Barbershop struggle with pose misalignment, while StyleYourHair struggles to preserve the input's face shape and hair details." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 495 + ], + "type": "text", + "content": "target hair. Barbershop performs well in preserving hair texture when the poses are similar. However, it falls short when handling challenging cases, such as transitioning from a long to a short hairstyle or removing bangs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 547, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 603 + ], + "type": "text", + "content": "StyleYourHair can produce more realistic results in unaligned cases than LOHO and Barbershop. However, the hair details often look different from the reference hairstyle (row 5). We also observe that StyleYourHair may perform poorly when the pose difference becomes too large (row 3), as also shown in our user study for pose in Table 1. In contrast, our method can transfer hairstyles convincingly, regardless of the misalignment, while still preserving the original face shape and hair details. We refer to Appendix F for more results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 611, + 381, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 381, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 381, + 624 + ], + "type": "text", + "content": "4.2. User Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Qualitative results can be misleading and biased, so we conducted a user study using Amazon Mechanical Turk on hairstyle transfer results using randomly selected pairs grouped into various difficulty levels. We compare results of our method with the current state of the art: StyleYourHair, Barbershop, and LOHO. Each participant was shown an input face, marked as 'Face', and a reference hair, marked" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8614" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 208 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 208 + ], + "type": "table", + "html": "
Test DatasetsFFHQ Scenario Breakdown
AllFFHQ-PFFHQ-SEasyMediumDifficult
-✓---✓✓✓✓✓---
--✓---✓--✓✓-
---✓---✓-✓-✓
----✓---✓-✓✓
LOHO8.89.68.412.05.34.013.38.05.35.34.010.75.313.314.7
Barbershop16.413.817.817.321.313.321.318.717.317.320.012.018.714.721.3
StyleYourHair18.512.421.625.321.310.722.733.310.724.028.014.732.014.721.3
Ours56.264.252.245.352.072.042.740.066.753.348.062.744.057.342.7
", + "image_path": "e3bc4fc796cf86a1b682b079a9ab73667e966bcc2efe2f2404ad9234d9873630.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 290, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 343 + ], + "type": "text", + "content": "as 'Hair', and asked to pick only one output that best accomplishes the task of transferring the hairstyle from image 'Hair' to the person in image 'Face'. The output row consists of four images from each method in random order. Each task was evaluated by 3 different participants. All images were in 256x256 resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 117, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 117, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 117, + 381 + ], + "type": "text", + "content": "4.2.1 Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 392, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 430 + ], + "type": "text", + "content": "For the user study, we construct two challenging benchmarks: FFHQ-P and FFHQ-S, from the test set of Flickr-Faces-HQ dataset (FFHQ) [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "text", + "content": "FFHQ-P contains 150 random input face-hair pairs from FFHQ, covering yaw differences in ranges of " + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "inline_equation", + "content": "[0 - 15)" + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "inline_equation", + "content": "[15 - 30)" + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "text", + "content": ", ..., " + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "inline_equation", + "content": "[75 - 90)" + }, + { + "bbox": [ + 46, + 431, + 288, + 467 + ], + "type": "text", + "content": ", with 25 pairs in each range." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 469, + 288, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 288, + 518 + ], + "type": "text", + "content": "FFHQ-S contains 300 pairs, categorized into 12 different config with varying levels of difficulty (Table 1). Each config contains 25 random input pairs and is a combination of four possible scenarios (see details in Appendix B):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 520, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 58, + 520, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 520, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 58, + 520, + 288, + 555 + ], + "type": "text", + "content": "- Pose Misalignment: When the yaw difference is between [15-30) or [30-45), a single checkmark or double checkmarks are used, respectively in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "type": "text", + "content": "- Needs Face Inpainting: This includes scenarios that require hallucinating parts of the original face, e.g., inpainting the forehead to remove bangs. This is challenging because the identity can easily change from the hallucination. We detect such scenarios based on the face/hair regions in " + }, + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 58, + 555, + 288, + 628 + ], + "type": "text", + "content": " (Appendix B)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "type": "text", + "content": "- Needs BG Inpainting: This includes scenarios where the hair shape becomes smaller and requires background inpainting. We detect such scenarios automatically based on the hair regions in " + }, + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 58, + 628, + 288, + 675 + ], + "type": "text", + "content": " (Appendix B)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "text", + "content": " Contains Hat: This represents scenarios where the hair reference is not fully visible in " + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "text", + "content": ". We detect such scenarios based on the hat region in " + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 58, + 675, + 288, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 309, + 269, + 547, + 344 + ], + "blocks": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 547, + 251 + ], + "type": "text", + "content": "Table 1. User study results on hairstyle transfer (percentage of user preferring each method). Our method outperforms state-of-the-art hairstyle transfer methods on FFHQ datasets in all challenging scenarios. A total of 450 pairs are used in this study, 150 pairs in FFHQ-P and 300 in FFHQ-S. For each pair, we asked 3 unique participants to select the best result for hairstyle transfer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 269, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 309, + 269, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 547, + 344 + ], + "type": "table", + "html": "
Pose Difference Range (FFHQ-P)
[0,15)[15,30)[30,45)[45,60)[60,75)[75,90)
LOHO10.75.310.712.08.010.7
Barbershop24.013.312.08.010.714.7
StyleYourHair17.316.012.06.714.78.0
Ours48.065.365.373.366.766.7
", + "image_path": "f43cd28a010c7ef423bdc8b0d2321181046c263bd5bd414b1e495f95fffde2ed.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 381, + 546, + 453 + ], + "blocks": [ + { + "bbox": [ + 305, + 351, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 351, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 547, + 374 + ], + "type": "text", + "content": "Table 2. User study on pose-invariant hairstyle transfer. Our method outperforms others on all pose difference ranges." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 381, + 546, + 453 + ], + "lines": [ + { + "bbox": [ + 309, + 381, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 309, + 381, + 546, + 453 + ], + "type": "table", + "html": "
Hair Reconstruction MetricsFace Shape
PSNR ↑SSIM ↑LPIPS ↓FID ↓RMSE ↓
LOHO25.760.860.0710.4015.84
Barbershop29.180.890.0510.4617.25
StyleYourHair26.890.870.0910.9314.37
Ours27.840.880.0710.8510.89
", + "image_path": "ca198d1b7ad9600198b18122dc8f7b81c4b982c27a6fcb2a634ac86b7d68426f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 461, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 305, + 461, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 547, + 495 + ], + "type": "text", + "content": "Table 3. Hair reconstruction results in the self-transfer experiment (Section 4.3), and RMSEs between facial landmarks detected on the input and output images.(Section 4.4)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 511, + 371, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 371, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 371, + 522 + ], + "type": "text", + "content": "4.2.2 Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": "Using all test pairs, the participants preferred our results " + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "inline_equation", + "content": "56.2\\%" + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": " of the time, whereas StyleYourHair's, Barbershop's, and LOHO's results were selected for " + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "inline_equation", + "content": "18.5\\%" + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "inline_equation", + "content": "16.4\\%" + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": ", respectively. Our method perform the best in both FFHQ-P " + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "inline_equation", + "content": "(64.2\\%)" + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": " and FFHQ-S " + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "inline_equation", + "content": "(52.2\\%)" + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": ", and in all configurations in Table 1-2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 610, + 476, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 476, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 476, + 624 + ], + "type": "text", + "content": "4.3. Quality of Hair Reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 548, + 715 + ], + "type": "text", + "content": "Following MichiGAN and LOHO, we perform a self-transfer experiment where we set both the input face and input reference hair to be the same and evaluate the reconstruction accuracy using various metrics: PSNR, SSIM, IP-IPS, FID [14]. For every method, the hair region of the output will be blended back to the input image to ensure that the difference in score only comes from the hair." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8615" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 547, + 215 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 547, + 215 + ], + "type": "image", + "image_path": "ea698b7ad3b32866c78395879f28b92d81bedb8aeb868b081837daf97ab6eb0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": "Figure 4. Ablation study of optimization stages and pre-masking. These ablated versions of our pipeline may produce results that (A) look realistic, but fail to match the " + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "inline_equation", + "content": "I_{h}" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": "'s hairstyle; (B, C, D) have unlikely hair shapes or structures; (E) have sharp boundaries; and (F) contain an incomplete background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 290, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 365 + ], + "type": "text", + "content": "The scores of our method, LOHO, Barbershop, and StyleYourHair are reported in Table 3. Unsurprisingly, when there is no misalignment between input face and hair, Barbershop generally achieves excellent hair reconstruction quality. Like ours, StyleYourHair is specifically designed to be pose-invariant, but their hair reconstruction quality seems to suffer greatly, as shown in our comparison (row 5 of Figure 3). Our method achieves better performance at hair reconstruction than StyleYourHair, but also suffers from similar drawbacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 237, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 237, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 237, + 384 + ], + "type": "text", + "content": "4.4. How Well Is Face Shape Preserved?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 288, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 484 + ], + "type": "text", + "content": "We also propose a novel evaluation metric for hairstyle transfer that focuses on the ability to preserve facial shape of the original person. We accomplish this by comparing detected keypoints of the input face " + }, + { + "bbox": [ + 46, + 389, + 288, + 484 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 389, + 288, + 484 + ], + "type": "text", + "content": " and those on the output using a simple Root-Mean-Square Error (RMSE). These keypoints were detected using an off-the-shelf library Dlib [22], and we only used keypoints on the facial contour (ID 0-16) for this evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "text", + "content": "For this evaluation, we randomize additional 1,550 FFHQ test pairs into our FFHQ-P and FFHQ-S datasets (2,000 images in total). Table 3 shows that our method outperforms others with an RMSE of 10.89, which is " + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "inline_equation", + "content": "24.2\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "text", + "content": " lower than StyleYourHair's (the second best), " + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "inline_equation", + "content": "36.9\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "text", + "content": " lower than Barbershop's and " + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 46, + 485, + 288, + 557 + ], + "type": "text", + "content": " lower than LOHO's." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 148, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 148, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 148, + 574 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": "Here we assess the importance of each component in our pipeline. The results are shown in Figure 4. We test 4 ablation configurations: i) by using only " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": " from a single view, ii) without using the hallucination stage (Section 3.3), iii) without using the latent sharing structure (Section 3.2.2), and iv) without using pre-masking in " + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{per}}" + }, + { + "bbox": [ + 46, + 582, + 287, + 654 + ], + "type": "text", + "content": " (Section 3.2.1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "Optimizing with a single view guide " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}^{\\mathrm{face}}" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " (Config i) yields inaccurate hairstyles (Figure 4-A). Without optimization in W (Config ii), the method produces unrealistic hair results with various artifacts, e.g., sharp edges, unnatural hair shapes (Figure 4-B) or structures (Figure 4-C). This is" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 254, + 385, + 357 + ], + "blocks": [ + { + "bbox": [ + 307, + 254, + 385, + 357 + ], + "lines": [ + { + "bbox": [ + 307, + 254, + 385, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 385, + 357 + ], + "type": "image", + "image_path": "729380b90aa568c3d848755e24b9051b8719b8d4f73e847ef78a1e113c13e20c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 361, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 547, + 407 + ], + "type": "text", + "content": "Figure 5. Failure cases; (A) the semantic regions are incorrect; (B) the facial keypoint is incorrect; (C) the EG3D projection is incorrect; (D) the reference hair color is too similar to the background color; (E) lighting looks unnatural; (F) eccentric hairstyle." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 255, + 465, + 357 + ], + "blocks": [ + { + "bbox": [ + 386, + 255, + 465, + 357 + ], + "lines": [ + { + "bbox": [ + 386, + 255, + 465, + 357 + ], + "spans": [ + { + "bbox": [ + 386, + 255, + 465, + 357 + ], + "type": "image", + "image_path": "3e400c49cc39ffac05055b02a5ea0fa5ddf4be1fd2187e6a0ffe2c5c562fcbc7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 466, + 255, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 466, + 255, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 466, + 255, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 466, + 255, + 545, + 357 + ], + "type": "image", + "image_path": "5f24128bd84b12ef1ccfb0e41c4a40c9b326905223e08c6503fb302d87eacc44.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "type": "text", + "content": "because the second stage, which has higher image fitting capability, tries to fit the initial rough estimation " + }, + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "type": "text", + "content": ". Without the latent sharing structure, the hair detail cannot be shared from the reference hair view (" + }, + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sim}}" + }, + { + "bbox": [ + 304, + 429, + 547, + 500 + ], + "type": "text", + "content": " alone is not strong enough to force consistency) (Figure 4-A), resulting in inaccurate hair colors or structures (Figure 4-C)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "content": "Without pre-masking in " + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{per}}" + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "content": ", the boundary of the face and hair regions is forced to be the same as in " + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{guide}}" + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "content": ", leading to visible and sharp seams between the face and hair or between different facial features. Here the optimizer fails to refine the face-hair boundary to make the results look natural (Figure 4-D), and the original details, such as the background from " + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 304, + 501, + 547, + 586 + ], + "type": "text", + "content": " cannot be seamlessly blended (Figure 4-E)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 597, + 463, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 463, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 463, + 609 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We have presented a flexible and effective hairstyle transfer system that can handle a variety of challenging in-the-wild scenarios and produce perceptually convincing hairstyle transfer results. Our user study shows that human evaluators prefer our results over previous methods across all tested scenarios. Nevertheless, our method can still fail in certain scenarios, for example, when the face and hair are too eccentric (Figure 5). We refer to Appendix G for more details." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8616" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE international conference on computer vision, pages 4432-4441, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 236 + ], + "type": "text", + "content": "[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Ironhammer america. File:elizabeth-olsen-1632123202.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth-olsen-1632123202.jpg/, 2022.17" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 316 + ], + "type": "text", + "content": "[5] Peter Baylies. Stylegan-encoder @ONLINE. https://github.com/pbaylies/stylegan-encoder/, 2019.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "text", + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "text", + "content": "[7] Prashanth Chandran, Sebastian Winberg, Gaspard Zoss, Jérémy Riviere, Markus Gross, Paulo Gotardo, and Derek Bradley. Rendering with style: combining traditional and neural approaches for high-quality face rendering. ACM Transactions on Graphics (ToG), 40(6):1-14, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 288, + 485 + ], + "type": "text", + "content": "[8] Chaeyeon Chung, Taewoo Kim, Hyelin Nam, Seunghwan Choi, Gyojung Gu, Sunghyun Park, and Jaegul Choo. Hairfit: Pose-invariant hairstyle transfer via flow-based hair alignment and semantic-region-aware inpainting. arXiv preprint arXiv:2206.08585, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 532 + ], + "type": "text", + "content": "[9] Gordon Correll. File: julianne moore (15011443428) (2). jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_(15011443428)_ (2). jpg/, 2014. 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 586 + ], + "type": "text", + "content": "[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7865-7874, 2022. 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[13] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only chang-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 74, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 546, + 94 + ], + "type": "text", + "content": "ing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 148 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 547, + 182 + ], + "type": "text", + "content": "[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Proc. NeurIPS, 2020. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 227 + ], + "type": "text", + "content": "[16] jerryhouuu. Face-yaw-roll-pitch-from-posedestimation-using-opencv @ONLINE. https://github.com/jerryhouuu/Face-Yaw-Roll-Pitch-from-Pose-Estimation-using-OpenCV/, 2018.13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 546, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 546, + 271 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 273, + 546, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 546, + 316 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 317, + 546, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 546, + 370 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 4, 5, 12, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 372, + 546, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 546, + 425 + ], + "type": "text", + "content": "[20] Taewoo Kim, Chaeyeon Chung, Yoonseo Kim, Sunghyun Park, Kangyeol Kim, and Jaegul Choo. Style your hair: Latent optimization for pose-invariant hairstyle transfer via local-style-aware hair alignment. arXiv preprint arXiv:2208.07765, 2022. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 546, + 491 + ], + "type": "text", + "content": "[21] Taewoo Kim, Chaeyeon Chung, Sunghyun Park, Gyojung Gu, Keonmin Nam, Wonzo Choe, Jaesung Lee, and Jaegul Choo. K-hairstyle: A large-scale korean hairstyle dataset for virtual hair editing and hairstyle classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1299-1303. IEEE, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 493, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 546, + 514 + ], + "type": "text", + "content": "[22] Davis E. King. Dlib-ml: A machine learning toolkit. Journal of Machine Learning Research, 10:1755–1758, 2009. 3, 8, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 546, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 546, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 546, + 546 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 546, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 546, + 581 + ], + "type": "text", + "content": "[24] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 546, + 635 + ], + "type": "text", + "content": "[25] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2445, 2020. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 546, + 669 + ], + "type": "text", + "content": "[26] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[27] nicolas genin. File:julianne moore 2009 venice film festival.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Julianne_Moore_2009_Venice_Film_Festival.jpg/, 2009.17" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8617" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 128, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 128, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 128, + 287, + 161 + ], + "type": "text", + "content": "[29] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 6, 11, 12, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 194 + ], + "type": "text", + "content": "[30] Rohit Saha, Brendan Duke, Florian Shkurti, Graham Taylor, and Parham Aarabi. Loho: Latent optimization of hairstyles via orthogonalization. In CVPR, 2021. 1, 2, 4, 5, 6, 19, 20, 21" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 248 + ], + "type": "text", + "content": "[31] David Shankbone. File:dwayne johnson at the 2009 tribeca film festival.jpg @ONLINE. https://commons.wikimedia.org/wiki/File:Dwayne_Johnson_at_the_2009_Tribeca_Film_Festival.jpg/, 2009.17" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 282 + ], + "type": "text", + "content": "[32] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 288, + 327 + ], + "type": "text", + "content": "[33] Gage Skidmore. Tom holland speaking at the 2016 san diego comic-con international in san diego, california. @ONLINE. https://commons.wikimedia.org/wiki/File: Tom_Holland_by_Gage_Skidmore.jpg/, 2016.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[34] Gage Skidmore. File:elizabeth olsen by gage skidmore 2.jpg @ ONLINE. https://commons.wikimedia.org/wiki/File:Elizabeth_Olsen_by_Gage_Skidmore_2.jpg/, 2019.17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 372, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 288, + 415 + ], + "type": "text", + "content": "[35] Zhentao Tan, Mengelei Chai, Dongdong Chen, Jing Liao, Qi Chu, Lu Yuan, Sergey Tulyakov, and Nenghai Yu. Michigan: multi-input-conditioned hair image generation for portrait editing. arXiv preprint arXiv:2010.16417, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 288, + 458 + ], + "type": "text", + "content": "[36] tenasia10. File:blackpinkLisa gmp 240622.png trends @ONLINE. https://commons.wikimedia.org/wiki/File:Blackpink_Lisa_GMP_240622.png/, 2022.17" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 513 + ], + "type": "text", + "content": "[37] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 557 + ], + "type": "text", + "content": "[38] GABI THORNE. Jennifer aniston reacts to tiktok's best (and most absurd) trends @ONLINE. https://www.allure.com/story/jennifer-aniston-reacts-to-tiktok-trends/, 2022.17" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 601 + ], + "type": "text", + "content": "[39] Albert Truuväärt. File:re kaja kallas (cropped).jpg @ONLINE. https://commons.wikimedia.org/wiki/File:RE_Kaja_Kallas_(cropped). jpg/, 2011.17" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 657 + ], + "type": "text", + "content": "[40] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE International Conference on Computer Vision, pages 5914–5922, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[41] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting gans with consecutive images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13910-13918, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 386 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[42] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. 3, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "type": "text", + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 205 + ], + "type": "text", + "content": "[44] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proc. ECCV, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 207, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 547, + 251 + ], + "type": "text", + "content": "[45] Tao Zhou, Changqing Zhang, Chen Gong, Harish Bhaskar, and Jie Yang. Multiview latent space learning with feature redundancy minimization. IEEE transactions on cybernetics, 50(4):1655-1668, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 252, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 547, + 296 + ], + "type": "text", + "content": "[46] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. arXiv preprint arXiv:2106.01505, 2021. 1, 2, 3, 4, 5, 6, 14, 19, 20, 21" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "type": "text", + "content": "[47] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Hairnet: Hairstyle transfer with pose changes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 651-667. Springer, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 547, + 386 + ], + "type": "text", + "content": "[48] zllrunnings. face-parsing.pytorch @ONLINE. https: //github.com/zllrunning/face-parsing. PyTorch/, 2019.11" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8618" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_content_list.json b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c3d4fd9da7154f73aae488b51bf1e763c4301c5 --- /dev/null +++ b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_content_list.json @@ -0,0 +1,1827 @@ +[ + { + "type": "text", + "text": "StyleGene: Crossover and Mutation of Region-level Facial Genes for Kinship Face Synthesis", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Li $^{1}$ , Xianxu Hou $^{2,5}$ , Zepeng Huang $^{1}$ , Linlin Shen $^{1,2,3,4*}$", + "bbox": [ + 263, + 202, + 753, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Computer Vision Institute, College of Computer Science and Software Engineering, Shenzhen University", + "bbox": [ + 88, + 220, + 923, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University", + "bbox": [ + 106, + 239, + 903, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Shenzhen Institute of Artificial Intelligence and Robotics for Society", + "bbox": [ + 228, + 257, + 779, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University", + "bbox": [ + 156, + 273, + 852, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ School of AI and Advanced Computing, Xi'an Jiaotong-Liverpool University", + "bbox": [ + 197, + 292, + 813, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{haoli2021, huangzepeng2021}@email.szu.edu.cn, hxianxu@gmail.com, llshen@szu.edu.cn", + "bbox": [ + 135, + 311, + 867, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1998cdf32cf9536c7e74081fe95dd73bc74e42df801218b6b2bea9a271e98947.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 349, + 251, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a7df84691f49bdb9b63d832f714d4d1b92ea94c942f0b07725b369ac10c9d00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 349, + 374, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f31204d2347063184e1d1459fe49d2eccd35eee75bd54927bbc8b477a174888b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 398, + 374, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17e0862662d2c685105cf6163b174d61d9c8c3c826b98aacae3743d7078b4016.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 349, + 496, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10696e267e50c0c4a6f3b5dd3896f8289d409d9761cb5ff199f79ca94789a8af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 349, + 620, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d38b3d919d04a9e26884ac9f9bd669931c3efdf65a81b77e3f4f34f5e5702705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 349, + 743, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/541c075681020995d83fa1ddcb7b3a9dc3131b633d251594629047f8d417fbfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 349, + 867, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3798f1f0b9786875d1362048c91fa574a90ec491b8bec8353046565558faedf5.jpg", + "image_caption": [ + "Figure 1. Our StyleGene method synthesizes kinship faces with resemblance to parents, exhibiting diversity and reasonable variations. The first row is the input grandparents, and the second and third rows are their descendants generated by our method." + ], + "image_footnote": [], + "bbox": [ + 192, + 500, + 313, + 589 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e81a98ba803287af67189da48ee6173fffe5c5144b06556c7760bd15e4742ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 500, + 434, + 589 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/176d95cd0ac1500b717cc8261922de7fa8b77f9758c897614d7eef8fd181790b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 500, + 558, + 589 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91501a59c42703cd6a68d80b266ca003d0d04ef4ff667a580a2f3adc9e56cc72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 500, + 679, + 589 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a8d890bf2051b30ae387258cd847eab50828c718365ac1e5c23cfbef3c42d3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 500, + 803, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 635, + 313, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-fidelity kinship face synthesis has many potential applications, such as kinship verification, missing child identification, and social media analysis. However, it is challenging to synthesize high-quality descendant faces with genetic relations due to the lack of large-scale, high-quality annotated kinship data. This paper proposes RFG (Region-level Facial Gene) extraction framework to address this issue. We propose to use IGE (Image-based Gene Encoder), LGE (Latent-based Gene Encoder) and Gene Decoder to learn the RFGs of a given face image, and the relationships between RFGs and the latent space of StyleGAN2. As cycle-like losses are designed to measure the $\\mathcal{L}_2$ distances between the output of Gene Decoder and image encoder, and that between the output of LGE and IGE, only", + "bbox": [ + 75, + 667, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "face images are required to train our framework, i.e. no paired kinship face data is required. Based upon the proposed RFGs, a crossover and mutation module is further designed to inherit the facial parts of parents. A Gene Pool has also been used to introduce the variations into the mutation of RFGs. The diversity of the faces of descendants can thus be significantly increased. Qualitative, quantitative, and subjective experiments on FIW, TSKinFace, and FF-Databases clearly show that the quality and diversity of kinship faces generated by our approach are much better than the existing state-of-the-art methods.", + "bbox": [ + 496, + 637, + 895, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 829, + 632, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can identify kinship through photographs based on the resemblance between parents and children. Many works have investigated this intrinsic relation in the fields", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 47 + ], + "page_idx": 0 + }, + { + "type": "aside_text", + "text": "Grandparents", + "bbox": [ + 102, + 348, + 122, + 414 + ], + "page_idx": 0 + }, + { + "type": "aside_text", + "text": "Parents", + "bbox": [ + 106, + 429, + 122, + 467 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20960", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of kinship verification [9, 32, 42] and genetics [4, 5, 8, 19]. With the popularity of face synthesis and editing technology in recent years, high-fidelity kinship face synthesis has also attracted much attention. This task, aiming to synthesize the faces of descendants based on the appearance of the parents, has many potential applications, such as finding long-lost children, crime investigations, kinship verification, and multimedia social applications.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, many efforts have been made to make use of generative models [7, 12, 15, 16, 27, 29, 38, 43, 45, 48] for kinship face synthesis. These works can be categorized into two paradigms: one-stage and two-stage. The one-stage paradigm [12, 29, 38, 45] treats this problem as an image-to-image translation task and trains a one-to-one kinship face generator with paired data. However, these approaches can only produce low-resolution images and the resultant images can be blurry and lack diversity. Further, it would be quite difficult to obtain annotated kinship data. By contrast, the two-stage paradigm [7, 15, 16, 27, 43, 48] first extracts the genetic representation and assembles them into children's representation based on the parents' faces. To obtain genetic representation, existing methods try to learn the inheritance and variation of facial appearances by training deep neural networks [14, 27, 48] or via a knowledge rule [7]. However, the learned genetic representation is prone to overfitting due to the lack of high-quality kinship annotated training data, resulting in a lack of diversity in the generated children. In addition, these methods cannot provide fine-grained attributes representation, and thus the generated facial attributes lack interpretability.", + "bbox": [ + 75, + 212, + 470, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, the facial genetic process is abstracted as the exchange and mutation of the parents' facial parts. We propose an Image-based Gene Encoder (IGE) to construct an independent representation for each facial part, called a Region-level Facial Gene (RFG), which is used to control the synthesis of facial regions. We further simulate the crossover and mutation process to assemble the RFGs of descendants by using those of the parents, and our proposed Gene Pool used in the mutation process can significantly increase the diversity of the generated descendants. We use the pre-trained StyleGAN2 [24] as the generator to synthesize high-fidelity faces. To achieve this, we use a Gene Decoder to map RFGs to the $\\mathcal{W}^+$ space of StyleGAN2. Since IGE requires a facial parsing mask to generate the RFG, we additionally train a Latent-based Gene Encoder (LGE) to directly map the latent code of StyleGAN2 to RFGs. Thus, facial parsing mask is not required for the RFG extraction in the inference stage. The main contributions of this paper are summarized as follows:", + "bbox": [ + 75, + 546, + 468, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose StyleGene to synthesize high-fidelity kinship faces with controllable facial genetic regions, via modeling the facial genetic relations based on the proposed region-level facial genes.", + "bbox": [ + 94, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel genetic strategy is further introduced by simulating the crossover and mutation process to generate the RFGs of descendants. We introduce a Gene Pool into the mutation process to significantly increase the diversity of the kinship face.", + "- We validate the effectiveness of our approach on several benchmarks, demonstrating the superiority of our StyleGene framework over other state-of-the-art methods, in terms of the quality and diversity of the generated kinship faces." + ], + "bbox": [ + 517, + 90, + 890, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 251, + 640, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Manipulation in Latent Space of StyleGAN", + "text_level": 1, + "bbox": [ + 500, + 276, + 866, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative Adversarial Networks (GANs) [17] have been widely used in face generation [21, 33, 47] and editing [3, 13, 50]. In particular, the StyleGAN [22-24] has attracted much attention due to its ability to synthesize high-fidelity images. The StyleGAN generator first maps a latent code $z \\in \\mathcal{Z}$ drawn from a normal distribution to an intermediate latent code $w \\in \\mathcal{W}$ by a mapping function to control image generation. Previous studies [20, 23, 36] have demonstrated that $\\mathcal{W}$ space has learned facial attribute semantics and different layers of latent code control different levels of image attributes. Based on this finding, many works [6, 49] tried to invert the real image to $\\mathcal{W}$ space for semantic face editing. Some recent works [1, 2, 34] show that the $\\mathcal{W}^+$ space extended from the $\\mathcal{W}$ space has lower reconstruction errors. In this paper, we build a link between our proposed RFGs and the $\\mathcal{W}^+$ space of StyleGAN2 [24], based on which we can achieve fine-grained control over the synthesis of facial regions.", + "bbox": [ + 496, + 299, + 890, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Kinship Face Synthesis", + "text_level": 1, + "bbox": [ + 500, + 579, + 714, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Kinship face synthesis aims to synthesize the face images of descendants given the images of parents. The challenge is to learn facial genetic relations with limited kinship data. Table 1 summarizes the key differences between ours and existing methods. Early works [14,29,48] design the image level mapping from parents to children via supervised learning. However, limited by the quality and scale of training data, the quality of images produced by these methods is usually low and prone to overfitting. Recent works", + "bbox": [ + 496, + 602, + 890, + 739 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/619564a493f248d4a0957da529c873f0641ac47d1d5c10a6f7ffc74b0fa8cf55.jpg", + "table_caption": [ + "Table 1. Comparison between StyleGene and existing kinship face synthesis methods. $N_{r}$ denotes the number of controllable regions." + ], + "table_footnote": [], + "table_body": "
MethodsStagesKinship annotationImage resolutionDiversity ControllingNr
DNA-Net [14]two✓128Noise-
KinshipGAN [29]one✓128Noise-
ChildPredictor [48]two✓128Noise-
CDFS [45]oneX256Noise5
StyleDNA [27]two✓1024Noise-
ChildGAN [7]twoX1024Interpolation5
StyleGene (Ours)twoX1024Gene Pool34
", + "bbox": [ + 504, + 777, + 890, + 910 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20961", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc7f40045aaf7efd468d9d4d75083ccb3579bd24f1bb559f06b95902e917bbdb.jpg", + "image_caption": [ + "(a) Training Stage", + "Figure 2. The overall framework of our method. The image-based gene encoder $E_{img}$ learns to encode an independent RFG representation for each facial region in the training stage. The gene decoder $D$ maps RFGs $g$ to the $\\mathcal{W}^+$ space of StyleGAN2. For ease of use, our latent-based gene encoder $E_{lat}$ maps latent codes $w$ (obtained by an image inversion encoder $E_{inv}$ ) to RFGs. In the inference stage, the RFGs of both parents are first extracted using latent-based gene encoder, and then the RFGs of the descendants are assembled by crossover and mutation module. The RFGs will be mapped back to the $\\mathcal{W}^+$ space using gene decoder, and finally the high-fidelity face is generated by a pre-trained StyleGAN2 generator $G$ ." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 344, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a26cf294bb0568f61cab7d458c0929d1593aa00aea2de332e1c2a37f317300e9.jpg", + "image_caption": [ + "(b) Inference Stage" + ], + "image_footnote": [], + "bbox": [ + 348, + 90, + 880, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[7,27] based on StyleGAN can generate high-resolution images by interpolating in the latent space to control the generation of descendants. However, StyleDNA [27] learns the mapping from parents to children in $\\mathcal{W}$ space by supervised learning, resulting in poor diversity due to overfitting. In addition, the inherited regions of children are uncontrollable. ChildGAN [7] calculates the direction vectors of different facial regions through annotated landmarks, and controls the genetic regions by interpolation. However, due to attribute entanglement in $\\mathcal{W}$ space, finding the disentangled direction vectors for small facial regions is difficult, which limits the diversity of generated descendants.", + "bbox": [ + 75, + 422, + 468, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this problem, our proposed StyleGene first learns region-level facial genes (RFGs) to control the synthesis of face regions. Then we further model kinship relations based on RFGs, and leverage the pre-trained StyleGAN2 generator to generate kinship face.", + "bbox": [ + 75, + 604, + 468, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 694, + 166, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our approach for high-fidelity kinship face synthesis. Given a pair of parental face images, $x_{f}$ and $x_{m}$ , our goal is to synthesize face images of their descendants. We consider the facial genetic process as the inheritance of facial parts from the father and mother.", + "bbox": [ + 75, + 719, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of the proposed framework is shown in Fig. 2. In training, we build models to extract Region-level Facial Genes (RFGs) from StyleGAN2 latent space and then learn to map the disentangled genes representations back to the corresponding latent code. In particular, we use an Image-based Gene Encoder (IGE) to directly extract RFGs from the input face with region annotations.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then a Gene Decoder is trained to map the obtained RFGs to StyleGAN2 latent space. In addition, we use a GAN inversion method to embed the same face to the $\\mathcal{W}^+$ latent space of StyleGAN2, and a Latent-based Gene Encoder (LGE) is used to decompose the obtained latent code into the corresponding RFGs.", + "bbox": [ + 496, + 422, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In inference, we first use LGE to extract the RFGs of parental faces, which are then used to obtain children's RFGs via crossover and mutation process. In addition, we build a Gene Pool to simulate the genetic variations.", + "bbox": [ + 496, + 513, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Region-level Facial Gene", + "text_level": 1, + "bbox": [ + 500, + 587, + 725, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to build facial genetic relations, a key step of the proposed framework is to extract disentangled representations of different face regions. We propose to train IGE to directly extract RFGs by using fine-grained annotated facial parts. We then use a Gene Decoder to transform the obtained gene representations into the StyleGAN2 latent space, from which we can reconstruct the original face image.", + "bbox": [ + 496, + 611, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-grained Facial Parts Segmentation. We follow DatasetGAN [46] pipeline to generate face images and corresponding segmentation masks. DatasetGAN trains a shallow decoder to achieve semantic segmentation based on the features of StyleGAN by only using a few annotated data. In this work, we adapt their approach to StyleGAN2 to produce pixel-level labels for $N$ ( $N = 34$ ) facial parts. These fine-grained facial parts are then used to extract the corresponding facial genes.", + "bbox": [ + 496, + 733, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-based Gene Encoder. Fig. 4 shows the overview of our IGE $E_{img}(\\cdot, \\cdot)$ . The goal of IGE is to learn region-", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20962", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e7a1117b91cd3e7a3e27b567ab99d676447c2453473644e38652985b32bdb204.jpg", + "image_caption": [ + "Figure 3. The proposed Latent-based Gene Encoder (a) and Gene Decoder (b) architecture. We mainly use the Mixer layer [39] to deal with the nonlinear transformation between high-dimensional vectors." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 869, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "level facial gene representation $\\pmb{g}_i$ for each facial part.", + "bbox": [ + 75, + 281, + 434, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a face image $x \\in \\mathbb{R}^{3 \\times 512 \\times 512}$ and the corresponding mask $m = \\{m_1, \\dots, m_N\\}$ , we first transform $x$ to a feature map $f(x)$ with $64 \\times 64$ resolution by a series of convolutional layers $f$ . Then we extract the region feature $r_i$ for the $i$ -th face part by masking the obtained $f(x)$ with the corresponding binary mask $m_i$ , i.e., $r_i = f(x) \\odot m_i$ , where $\\odot$ is an element-wise matrix multiplication operator. Inspired by variational autoencoder [25], the posterior distribution for each region $r_i$ , mapped by a mapping function $h_i$ , is modeled as a multivariate Gaussian distribution. It is defined by", + "bbox": [ + 75, + 295, + 468, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {g}} _ {i} \\sim q (\\hat {\\boldsymbol {g}} _ {i} | \\boldsymbol {r} _ {i}) = \\mathcal {N} (\\hat {\\boldsymbol {g}} _ {i}; \\hat {\\boldsymbol {\\mu}} _ {i}, \\hat {\\boldsymbol {\\sigma}} _ {i} ^ {2} \\boldsymbol {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 467, + 468, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{\\mu}}_i,\\hat{\\pmb{\\sigma}}_i\\in \\mathbb{R}^{18\\times 512}$ are the multi-dimensional output of $h_i(r_i)$ , representing the mean and standard deviation in a diagonal matrix form, respectively. According to [25], backpropagation is made differentiable via the reparameterization trick, thus $\\hat{\\pmb{g}}_i\\in \\mathbb{R}^{18\\times 512}$ can be sampled by", + "bbox": [ + 75, + 489, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {g}} _ {i} = \\hat {\\boldsymbol {\\mu}} _ {i} + \\hat {\\boldsymbol {\\sigma}} _ {i} \\odot \\boldsymbol {\\epsilon}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 571, + 468, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, since we use $N$ facial parts, all the RFG can be denoted by $\\hat{\\pmb{g}} = [\\hat{\\pmb{g}}_1,\\dots ,\\hat{\\pmb{g}}_N]\\in \\mathbb{R}^{18\\times 512\\times N}$ . We simplify this process denoted as $\\hat{\\pmb{g}} = E_{img}(x,m)$", + "bbox": [ + 75, + 594, + 468, + 640 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33e0bdde9596cca5ece169deeed617eff434d09b6aedb8bdb8514fe6fe57696d.jpg", + "image_caption": [ + "Figure 4. The proposed Image-based Gene Encoder." + ], + "image_footnote": [], + "bbox": [ + 101, + 657, + 441, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent-based Gene Encoder. As shown above, the proposed IGE needs a face mask to extract the representations", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each facial region. To eliminate the dependency, we train a LGE $E_{lat}(\\cdot)$ to directly extract RFGs from StyleGAN2 latent space. As shown in Fig. 3(a), our LGE mainly consists of a fully connected layer and several Mixer Layers [39], and maps the latent code $\\boldsymbol{w}$ of a face image $x$ to the mean $\\boldsymbol{\\mu}_i$ and standard deviation $\\sigma_i$ for each pre-defined face region. This process can be formulated by", + "bbox": [ + 496, + 281, + 890, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {\\mu}, \\boldsymbol {\\sigma} \\right\\} = E _ {l a t} (\\boldsymbol {w}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 392, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu} = [\\pmb{\\mu}_1, \\dots, \\pmb{\\mu}_N]$ and $\\pmb{\\sigma} = [\\pmb{\\sigma}_1, \\dots, \\pmb{\\sigma}_N]$ , and we use a pre-trained GAN inversion model $E_{inv}(\\cdot)$ [40] as the image encoder to get the latent code of a face image, i.e. $\\pmb{w} = E_{inv}(x)$ . Then the region-level gene $\\pmb{g}_i$ for the $i$ -th facial region can be sampled by", + "bbox": [ + 496, + 412, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {i} = \\boldsymbol {\\mu} _ {i} + \\boldsymbol {\\sigma} _ {i} \\odot \\boldsymbol {\\epsilon}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 494, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the RFGs of $N$ facial parts can be denoted by $\\pmb{g} = [g_1,\\dots ,g_N]$", + "bbox": [ + 496, + 517, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gene Decoder. We then use Gene Decoder $D(\\cdot)$ to map the region-level facial genes into the StyleGAN2 latent space. Thus, we can reconstruct the face images from the obtained gene representation. Specifically, the Gene Decoder first processes the obtained RFGs with several Mixer Layers [39]. Then all the regional features are concatenated and further processed by a fully connected layer to obtain the latent code $\\hat{\\boldsymbol{w}}\\in \\mathbb{R}^{18\\times 512}$ . This process can be formulated by", + "bbox": [ + 496, + 547, + 890, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {w}} = D (\\hat {\\boldsymbol {g}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 683, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{g}} = [\\hat{\\pmb{g}}_1,\\dots ,\\hat{\\pmb{g}}_N]$ is the RFGs of $N$ facial parts.", + "bbox": [ + 496, + 705, + 861, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Functions. As shown in Fig. 2(a), given an input face image $x$ , the corresponding region mask $m$ and the latent code $w = E_{inv}(x)$ , we train our models by using following loss functions.", + "bbox": [ + 496, + 720, + 890, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To build a bridge between RFGs and latent space, we first constrain the IGE $E_{img}$ and Gene Decoder $D$ to reconstruct the latent code of the input image. The latent reconstruct loss is defined by", + "bbox": [ + 496, + 782, + 890, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l a t e n t}} = \\left\\| \\boldsymbol {w} - \\hat {\\boldsymbol {w}} \\right\\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 848, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{w}} = D(E_{img}(x,m))$", + "bbox": [ + 496, + 869, + 692, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, we use LGE $E_{lat}$ to remove the reliance on", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20963", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "face annotations. To decouple the latent code into the RFGs obtained from IGE, the LGE is trained by", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g e n e}} = \\sum_ {i = 1} ^ {N} \\left[ \\| \\boldsymbol {\\mu} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {i} \\| _ {2} + \\| \\boldsymbol {\\sigma} _ {i} - \\hat {\\boldsymbol {\\sigma}} _ {i} \\| _ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 125, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{\\mu}}_i$ and $\\hat{\\pmb{\\sigma}}_i$ are the outputs of $E_{img}(x,m)$ for the $i$ -th facial region, and $\\pmb{\\mu}_i, \\pmb{\\sigma}_i$ are the outputs of $E_{lat}(\\pmb{w})$ for the $i$ -th facial region.", + "bbox": [ + 76, + 167, + 468, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All modules are trained jointly, and the full loss is defined by", + "bbox": [ + 76, + 213, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {l a t e n t}} + \\lambda \\mathcal {L} _ {\\text {g e n e}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 244, + 468, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter, and we set it to be 1 in our experiments.", + "bbox": [ + 76, + 267, + 468, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that this training process does not rely on kinship data, alleviating the problem of insufficient high-quality annotated kinship datasets.", + "bbox": [ + 76, + 297, + 468, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Crossover & Mutation", + "text_level": 1, + "bbox": [ + 76, + 353, + 285, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As parent-offspring resemblance is often reflected in the local area of the face appearance [5, 31]. Thus, after obtaining the region-level facial genes (RFGs), we simulate a crossover process to generate the RFGs of descendants by combining the RFGs of the parents. Additionally, we further build a gene pool to simulate the gene mutation process, which can significantly increase the diversity of synthesized faces.", + "bbox": [ + 76, + 377, + 468, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gene Crossover. The gene crossover process aims to generate the gene representation of the descendants with a linear combination of the RFGs of the parents. In particular, given a pair of parental face images, $x_{f}$ and $x_{m}$ , we first apply Image Encoder and LGE to obtain a set of RFGs $g^{f} = [g_{1}^{f},\\dots ,g_{N}^{f}]$ and $g^{m} = [g_{1}^{m},\\dots ,g_{N}^{m}]$ from $N$ facial parts, respectively. Then, the RFGs of descendants $g^{c} = [g_{1}^{c},\\dots ,g_{N}^{c}]$ can be calculated by", + "bbox": [ + 76, + 498, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {i} ^ {c} = \\alpha_ {i} \\boldsymbol {g} _ {i} ^ {f} + \\beta_ {i} \\boldsymbol {g} _ {i} ^ {m}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 622, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ and $\\beta_{i}$ are randomly generated weights for each region and $\\alpha_{i} + \\beta_{i} = 1$ when mutation is not applied.", + "bbox": [ + 76, + 643, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 5 shows four example faces generated using RFGs inherited from their parents based on Eq. 9, when different combinations of $\\alpha_{i}$ and $\\beta_{i}$ are applied. The bigger value of $\\alpha_{i}$ , the more similar to father for the $i$ -th facial region. The figure shows that our approach can precisely and independently control the similarity of each facial region with their parents, by setting different values of $\\alpha_{i}$ and $\\beta_{i}$ .", + "bbox": [ + 76, + 672, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gene Mutation. In genetics, the gene pool is the set of all genes of all individuals in a population [30]. For humans, all races share the same gene pool. We introduce the gene pool concept to better simulate the genetic variation in facial appearances. We define the gene pool $\\mathcal{P}$ as the sets formed by grouping a large number of RFGs, i.e. $\\mathcal{P} = \\{\\pmb{g}^1,\\dots ,\\pmb{g}^p,\\dots ,\\pmb{g}^P\\}$ , where $\\pmb{g}^{p} = [g_{1}^{p},\\dots ,g_{N}^{p}]$ . As shown in the supplementary materials, we divide the gene", + "bbox": [ + 76, + 780, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fbabb2e7520161118d4c57c048608d9f45b82dda96ae8645c81d6a77e5961ff.jpg", + "image_caption": [ + "Figure 5. Example of different weights used in crossover of RFGs. While the redder regions look more similar to the father, and the bluer regions look more similar to the mother." + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 867, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pool into nine groups by age, two by gender, and seven by race. Note that the RFG in the gene pool can independently control different facial traits. And its phenotype is related to the grouping condition.", + "bbox": [ + 496, + 385, + 890, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a pair of parental face images and information about generated descendants, such as age, gender, and race. We first query the gene pool to extract a subset $\\mathcal{P}_c\\subseteq \\mathcal{P}$ that satisfies the descendants' information, which will be used to simulate gene mutations for better diversity in the descendants. Then we use Image Encoder and LGE to extract RFGs $\\pmb{g}^{f} = [g_{1}^{f},\\dots ,g_{N}^{f}]$ and $\\pmb{g}^{m} = [g_{1}^{m},\\dots ,g_{N}^{m}]$ of parents. For mutation, we randomly select $\\eta$ percent of facial regions whose RFGs are sampled from the gene pool. Given a $\\pmb{g}^p$ sampled from $\\mathcal{P}_c$ , a one-hot vector $t = \\{t_i\\in \\{0,1\\} ,i = 1,2,\\dots ,N\\}$ is randomly generated to denote whether the $\\pmb{g}_i^c$ of descendant is fully copied from $\\pmb{g}_i^p (t_i = 1)$ , or not $(t_i = 0)$ . When $\\pmb{g}_i^c$ is inherited from parents $(t_i = 0)$ , we also introduce mutation by introducing $\\pmb{g}^p$ into the crossover process defined in Eq. 9. Finally, we can extend Eq. 9 as follows to obtain the RFGs of the descendants $\\pmb{g}^c = [\\pmb {g}_c^1,\\dots ,\\pmb {g}_c^N ]$", + "bbox": [ + 496, + 446, + 892, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {i} ^ {c} = \\left\\{ \\begin{array}{l l} \\boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 1 \\\\ \\alpha_ {i} \\boldsymbol {g} _ {i} ^ {f} + \\beta_ {i} \\boldsymbol {g} _ {i} ^ {m} + \\gamma \\boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 0 \\end{array} , \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 710, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the intensity of mutation, $\\alpha_{i}$ and $\\beta_{i}$ are randomly generated weights and $\\alpha_{i} + \\beta_{i} = 1 - \\gamma$ , $g_{i}^{p} = S(\\mathcal{P}_{c})$ and $S(\\cdot)$ is the random sampling operator. Once the RFG of descendant $g^{c}$ is generated, the learned Gene Decoder is applied to map $g^{c}$ to the latent space of StyleGAN2, i.e. $\\boldsymbol{w}_{c} = D(\\boldsymbol{g}^{c})$ , $\\boldsymbol{w}_{c} \\in \\mathbb{R}^{18 \\times 512}$ , which is further processed for face synthesis.", + "bbox": [ + 496, + 748, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the latent code in $\\mathcal{W}^+$ latent space of StyleGAN2 usually consists of 18 layers, Richardson et al. [34] demonstrate that the first 8 layers of the latent code mainly con", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20964", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tribute to the ID information of the synthesized faces, and higher layers mainly control the skin color and microstructure of the synthesized faces. While the appearance of facial regions for descendants is mainly decided by $g^{c}$ , other attributes like skin, hair color, image lighting, and background, can thus be inherited by inclusion of the latent codes of parents. Therefore, we further fuse the latent codes of parents $\\boldsymbol{w}_{f}$ and $\\boldsymbol{w}_{m}$ . We keep the first $l$ layer of $\\boldsymbol{w}_{c}$ , and the higher layers are fused from $\\boldsymbol{w}_{f}$ and $\\boldsymbol{w}_{m}$ . The final latent code $\\boldsymbol{w}_{c} = [\\boldsymbol{w}_{c}^{1},\\dots ,\\boldsymbol{w}_{c}^{18}]$ of the descendant can be obtained by", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {w} _ {c} ^ {i} = \\frac {1}{2} \\left(\\boldsymbol {w} _ {f} ^ {i} + \\boldsymbol {w} _ {m} ^ {i}\\right), \\quad i \\in \\{l + 1, \\dots , 1 8 \\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 258, + 468, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 300, + 209, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 325, + 282, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our training set consists of three parts, all images of CelebAHQ [21], 50,000 faces sampled from MS-Celeb-1M [18], and 10,000 faces generated by StyleGAN2, to make the model adaptive to different image qualities. CelebA-HQ is a high-quality face dataset consisting of 30,000 aligned face images with $1024 \\times 1024$ resolution. The MS-Celeb-1M is a large-scale face recognition dataset containing about 10 million images, including many noisy face images. We evaluate our model on the FIW [35], TSKinFace [32], and FF-Database [48] datasets. FIW dataset contains the faces of 1,000 families with 1,997 sets of father-mother-child relations. The TSKinFace and FF-Database datasets provide 1,015 and 3,744 sets of father-mother-child kinship face images, respectively. We align and crop the images in these two datasets to $256 \\times 256$ . Our Gene Pool is built based on the FFHQ [23] dataset, which consists of 70,000 high-quality face images, and has a better diversity among human races and ages.", + "bbox": [ + 75, + 349, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method against the state-of-art kinship face synthesis baselines, i.e. DNA-Net [14], ChildGAN [7], ChildPredictor [48], and StyleDNA [27]. Since the source codes of StyleDNA and ChildPredictor are available, we compare our approach with them through all of the evaluations. However, due to the unavailability of the source codes, we can only visually compare the visual quality of the generated faces available in the papers of DNA-Net and ChildGAN.", + "bbox": [ + 75, + 622, + 468, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We use AdamW [28] optimizer with batch size of 32. The initialized learning rate is 0.001, which is divided by 2 every 10 epochs, and we stop the training at the 30th epoch.", + "bbox": [ + 75, + 758, + 468, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 76, + 830, + 284, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Disentanglement of RFGs. Fig. 6 shows two examples of facial region editing using our proposed RFG, where the nose, eyes, jaw and lips of source faces shown in the first", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "column, are replaced sequentially by the corresponding regions of reference faces shown in the first row. One can observe from the examples that, when a certain region is edited, all other facial regions are kept intact, which clearly shows the disentanglement capability of our RFG.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64dd6f01136f79727a0bee14bc3c13014445cf72ed33d2982c92a1936d28a084.jpg", + "image_caption": [ + "Figure 6. Disentangled editing of facial regions." + ], + "image_footnote": [], + "bbox": [ + 511, + 176, + 877, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Gene Pool. Based on the proposed idea of RFG, our approach further introduces a gene pool to increase the variations among the generated faces of descendants. Given a couple of parents shown in the first row, Fig. 7 shows the faces of five children generated by our approach with/without the involvement of gene pool, together with that synthesized by StyleDNA and ChildPredictor. While the faces generated by StyleDNA and ChildPredictor (4th and 5th rows) all look very similar, that synthesized by our approach are much more diverse and the introduction of gene pool (1st row) can further increase variations among children.", + "bbox": [ + 496, + 391, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96f0f3e296a85428c5ad729aa58a58ca200d4f642ca56e531efff9add0b86872.jpg", + "image_caption": [ + "Figure 7. Visual comparison of diversity with StyleDNA and ChildPredictor. The first row shows the father and mother. The next two rows are generated by our method with and without Gene Pool. The last two rows are generated by StyleDNA [27] and ChildPredictor [48], respectively." + ], + "image_footnote": [], + "bbox": [ + 509, + 584, + 877, + 830 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20965", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75637e407a67186632e1bf62e521eeb75f1c1afc3c63b8723c1e5bd7853f6ae1.jpg", + "image_caption": [ + "Father", + "Mother", + "Real Child", + "Ours", + "StyleDN", + "A ChildPre", + "Victor ChildGAN", + "DNA-Net" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 531, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac2707b07ecf1ccedd5f827ff9f2560a641eec8068aa0b9c6ec92f0cebe1b8b5.jpg", + "image_caption": [ + "Father", + "Mother", + "Real 0", + "d Ours", + "StyleDNA ChildPredictor", + "(b) FF-Database", + "Figure 8. Comparison of children faces synthesized by StyleGene and baselines. Left most three columns of (a) FIW and (b) FF-Database depict the father, mother, and real children, and right five or three columns depict the faces synthesized by ours, StyleDNA [27], ChildPredictor [48], ChildGAN [7] and DNA-Net [14]." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 879, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with the state-of-the-art. Fig. 8(a) and (b) show qualitative results on FIW and FF-Database, respectively. We first compare StyleGene with StyleDNA, DNA-Net, ChildGAN, and ChildPredictor on FIW with low-resolution facial images. Note that the results of ChildGAN and DNA-Net are directly taken from their papers. As can be seen, DNA-Net, ChildGAN, and ChildPredictor struggle to yield high-quality facial images. Thanks to a powerful generator, our method can generate high-fidelity faces with hair. Though ChildGAN also uses the StyleGAN generator, its results are significantly affected by the quality of the input. Note that all comparisons have considered the facial information of parents. However, DNA-Net, ChildPredictor, and StyleDNA model facial genetics by learning the implicit mapping between parents and children, without considering regional similarity, which significantly compromised the similarity between the generated faces and their parents. Although ChildGAN considered parents' facial features, the results were too similar to their parents. Thanks to the region-level facial genes and the mutation introduced by gene pool, our results can maintain the local similarity to parents' faces, and at the same time, present reasonable global variations and diversity. When experimenting with faces of different races as shown in Fig. 8(b), our approach can still well preserve the race similarity of children with their parents.", + "bbox": [ + 76, + 358, + 472, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 76, + 768, + 294, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Kinship Verification. We now use kinship verification accuracy to assess whether a genetic relationship exists between the synthetic descendant and parents. Higher accuracy means more realistic synthetic descendants. We use ArcFace pre-trained on the MS-Celeb-1M dataset as the backbone to create a one-versus-one kinship classifier [42], which is fine-tuned on the FF-Database training set. Then", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we performed a cross-database evaluation.", + "bbox": [ + 500, + 358, + 781, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we randomly sample 100 families with one kid from the FF-database test set, FIW, and TSKinFace datasets, respectively, and different methods randomly generate 40 children for each family. We randomly sample 8,000 positive and 8,000 negative pairs for each method from each dataset and calculate in Table 2 the kinship verification accuracy. One can observe that our method achieves substantially higher accuracy than StyleDNA and ChildPredictor, i.e. as high as $81.74\\%$ , $80.38\\%$ , and $62.29\\%$ accuracies are achieved on TSKinFace, FF-Database, and FIW datasets, respectively. The results suggest that the kinship relations of faces generated by our approach are very close to the real children.", + "bbox": [ + 496, + 373, + 890, + 568 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23587e21d80e1aef7eb66b02c86cb8b868c388d042b985f0f2b9a02dff337cca.jpg", + "table_caption": [ + "Table 2. Kinship verification accuracy $(\\%)$ on the TSKinFace [32], FF-Database [48], and FIW [35] dataset." + ], + "table_footnote": [], + "table_body": "
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA [27]53.1555.1149.47
ChildPredictor [48]58.2459.6251.81
StyleGene (Ours)81.7480.3862.29
", + "bbox": [ + 504, + 611, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diversity Evaluation. We used the LPIPS [44] metric to measure the diversity of synthetic descendants, which calculates the L1 distance between pairs of image features extracted by AlexNet [26] pre-trained on the ImageNet [10] dataset. We use the same test data of kinship verification. First, we calculated the distance among the 40 synthesized descendants for each family. Then we take the average of 100 families as the LPIPS score and list them in Table 3. As shown in the table, our method achieves the highest LPIPS across all three datasets, i.e. 0.3270, 0.3418, and 0.3279 LPIPS on TSKinFace, FF-Database, and FIW datasets are achieved, which is significantly higher than that of StyleDNA and ChildPredictor.", + "bbox": [ + 496, + 689, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distribution of Synthesized Children. We are now try-", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20966", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fcc33b211d841c3ea0c1f99ca4c387d7370f231777d1a17c9ed839e6bff3a715.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of the diversity of the generated descendants. * means we cropped the face i.e. no hair." + ], + "table_footnote": [], + "table_body": "
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA* [27]0.07560.07630.0736
ChildPredictor [48]0.16970.17230.1750
StyleGene (Ours)*0.17480.17350.1740
StyleDNA [27]0.15590.15420.1573
StyleGene (Ours)0.32700.34180.3279
", + "bbox": [ + 81, + 122, + 468, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing to model the distribution of the children synthesized by different approaches, and compare them with that of real children. Specifically, we randomly sample 50 families from the TSKinFace dataset, each consisting of a father, mother, son, and daughter. Based on the given parents, we also applied different synthesis methods to generate two children, i.e. a son and a daughter, for each family. After facial features are extracted using ArcFace [11] and reduced to one dimension using t-SNE [41], Kernel Density Estimation (KDE) [37] is further applied to estimate the features' probability density function. Fig. 9 shows the distribution of the 100 real children (blue) and that of children generated by different approaches. As can be seen, the distribution of faces generated by our approach (red) overlaps most with the real children (blue). In contrast, the distributions of faces synthesized by StyleDNA (black) and ChildPredictor (yellow) are located on both sides of the real ones. The distribution for StyleDNA even looks like a mixture of two Gaussian distributions, which is significantly different from that of real children.", + "bbox": [ + 75, + 236, + 470, + 537 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f488ad36f0d2931814796176ee795e854e0fa8ea59acee6f85d2994f8d580d10.jpg", + "image_caption": [ + "Figure 9. Distribution of real children (blue) and children generated using StyleDNA (black), ChildPredictor (yellow), and ours (red). Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 98, + 550, + 450, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. To further demonstrate the effectiveness of the proposed method, we conducted a user study involving 59 participants. The faces of children generated by our StyleGene, StyleDNA and ChildPredictor based on 20 parents, are present to each of the 59 participants, who are asked to rank the quality of the faces, in terms of realness and the similarity with given parents. In total $1,180(59 \\times 20)$ votes are received for each of the three approaches and the average rankings are listed in Table 4. As shown in the table, our StyleGene receives the highest rank among the com", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pared approaches, which suggests that the faces of children synthesized by our approach are the best, in terms of both quality and similarity with parents.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/997645d8ab3582f708ad7098e38be642085565f1fbe5c8b907a7cbd32f1fe7e2.jpg", + "table_caption": [ + "Table 4. The rank of different approaches in user study." + ], + "table_footnote": [], + "table_body": "
ChildPredictorStyleDNAStyleGene (Ours)
Avg. rank2.462.221.32
", + "bbox": [ + 521, + 159, + 870, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 212, + 648, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies on the effectiveness of LGE, IGE and GP, in terms of kinship verification accuracy (ACC) and diversity (LPIPS) using FF-database with the same configuration as in Section 4.3. As shown in Table 5, the integration of IGE (2nd row) can improve both ACC and LPIPS of the synthesized descendant. As expected, the use of GP (3rd row) increases the diversity of synthesized faces. When both IGE and GP are integrated with LGE (4th row), the diversity can be significantly increased to 0.1735, with a comparable ACC $(80.38\\%)$ .", + "bbox": [ + 496, + 234, + 890, + 387 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a635b334709e42d5d046ff799b5a61e06fcc1d661411737d28b71c8598b63c36.jpg", + "table_caption": [ + "Table 5. Ablation study on FF-Database." + ], + "table_footnote": [], + "table_body": "
LGEIGEGPACC (%)LPIPS
1✓XX79.190.0625
2✓✓X79.660.0646
3✓X✓80.210.0839
4✓✓✓80.380.1735
", + "bbox": [ + 547, + 412, + 841, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Parameter sensitivity", + "text_level": 1, + "bbox": [ + 500, + 511, + 697, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test different values of $\\eta$ , $\\gamma$ , and $l$ to see how the performance of synthesis varies. Due to the page limit, details about the results of different parameters are presented in supplementary. In our experiments, we choose $\\eta = 40\\%$ , $\\gamma = 0.47$ , and $l = 8$ to achieve a balance between the diversity and fidelity of the generated faces.", + "bbox": [ + 496, + 534, + 890, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 636, + 619, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed StyleGene, to extract RFGs (Region-level Facial Genes) for kinship face synthesis. While crossover and mutation of RFGs are proposed to model the facial genetic process, Gene Pool is further designed to increase the diversity among generated faces. Quantitative, qualitative, and subjective experimental results show that the realness, similarity with parents, and diversity of kinship faces generated by our approach are much better than existing state-of-the-art methods.", + "bbox": [ + 496, + 661, + 890, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was jointly supported by National Natural Science Foundation of China (62206180 and 82261138629), Guangdong Basic and Applied Basic Research Foundation (2023A1515010688 and 2022A1515011018), and Shenzhen Municipal Science and Technology Innovation Council (JCYJ20220531101412030).", + "bbox": [ + 496, + 803, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20967", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2", + "[3] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2", + "[4] Peter Claes, Jasmien Roosenboom, Julie D White, Tomek Swigut, Dzemila Sero, Jiarui Li, Myoung Keun Lee, Arslan Zaidi, Brooke C Mattern, Corey Liebowitz, et al. Genome-wide mapping of global-to-local genetic effects on human facial shape. Nature genetics, 50(3):414-423, 2018. 2", + "[5] Joanne B Cole, Mange Manyama, Jacinda R Larson, Denise K Liberton, Tracey M Ferrara, Sheri L Riccardi, Mao Li, Washington Mio, Ophir D Klein, Stephanie A Santorico, et al. Human facial shape and size heritability and genetic correlations. Genetics, 205(2):967-978, 2017. 2, 5", + "[6] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020. 2", + "[7] Xiao Cui, Wengang Zhou, Yang Hu, Weilun Wang, and Houqiang Li. Heredity-aware child face image generation with latent space disentanglement. arXiv preprint arXiv:2108.11080, 2021. 2, 3, 6, 7", + "[8] Maria F Dal Martello and Laurence T Maloney. Lateralization of kin recognition signals in the human face. Journal of Vision, 10(8):9-9, 2010. 2", + "[9] Afshin Dehghan, Enrique G Ortiz, Ruben Villegas, and Mubarak Shah. Who do i look like? determining parent-offspring resemblance via gated autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1757-1764, 2014. 2", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 8", + "[12] Itir Önal Ertugrul and Hamdi Dibeklioglu. What will your future child look like? modeling and synthesis of hereditary patterns of facial dynamics. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 33-40. IEEE, 2017. 2", + "[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2", + "[14] Pengyu Gao, Joseph Robinson, Jiaxuan Zhu, Chao Xia, MIng Shao, and Siyu Xia. Dna-net: Age and gender aware kin face synthesizer. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2, 6, 7", + "[15] Pengyu Gao, Siyu Xia, Joseph Robinson, Junkang Zhang, Chao Xia, Ming Shao, and Yun Fu. What will your child look like? dna-net: Age and gender aware kin face synthesizer. arXiv preprint arXiv:1911.07014, 2019. 2", + "[16] Fady S Ghatas and ElSayed E Hemayed. Gankin: generating kin faces using disentangled gan. SN Applied Sciences, 2(2):1-10, 2020. 2", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[18] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In European conference on computer vision, pages 87–102. Springer, 2016. 6", + "[19] Hanne Hoskens, Dongjing Liu, Sahin Naqvi, Myoung Keun Lee, Ryan J Eller, Karlijne Indencleef, Julie D White, Jiarui Li, Maarten HD Larmouseau, Greet Hens, et al. 3d facial phenotyping by biometric sibling matching used in contemporary genomic methodologies. PLoS genetics, 17(5):e1009528, 2021. 2", + "[20] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 2", + "[21] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2, 6", + "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 6", + "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4", + "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7", + "[27] Che-Hsien Lin, Hung-Chun Chen, Li-Chen Cheng, Shu-Chuan Hsu, Jun-Cheng Chen, and Chih-Yu Wang. *Styledna: A high-fidelity age and gender aware kinship face synthesizer*. In 2021 16th IEEE International Conference on Auto" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20968", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matic Face and Gesture Recognition (FG 2021), pages 1-8. IEEE, 2021. 2,3,6,7,8", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[29] Savas Ozkan and Akin Ozkan. Kinshipgan: Synthesizing of kinship faces from family photos by regularizing a deep face network. In 2018 25th IEEE international conference on image processing (ICIP), pages 2142-2146. IEEE, 2018. 2", + "[30] Svante Paäbo. The diverse origins of the human gene pool. Nature Reviews Genetics, 16(6):313-314, 2015. 5", + "[31] Shouneng Peng, Jingze Tan, Sile Hu, Hang Zhou, Jing Guo, Li Jin, and Kun Tang. Detecting genetic association of common human facial morphological variation using high density 3d image registration. PLoS computational biology, 9(12):e1003375, 2013. 5", + "[32] Xiaoqian Qin, Xiaoyang Tan, and Songcan Chen. Tri-subject kinship verification: Understanding the core of a family. IEEE Transactions on Multimedia, 17(10):1855-1867, 2015. 2, 6, 7", + "[33] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2", + "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2, 5", + "[35] Joseph P Robinson, Ming Shao, Yue Wu, and Yun Fu. Families in the wild (fw) large-scale kinship image database and benchmarks. In Proceedings of the 24th ACM international conference on Multimedia, pages 242-246, 2016. 6, 7", + "[36] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2", + "[37] Bernard W Silverman. Density estimation for statistics and data analysis. Routledge, 2018. 8", + "[38] Raunak Sinha, Mayank Vatsa, and Richa Singh. Familygan: Generating kin face images using generative adversarial networks. In European conference on computer vision, pages 297-311. Springer, 2020. 2", + "[39] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. 4", + "[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4", + "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Jun Yu, Mengyan Li, Xinlong Hao, and Guochen Xie. Deep fusion siamese network for automatic kinship verification. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 892-899. IEEE, 2020. 2, 7", + "[43] Ishtiak Zaman and David Crandall. Genetic-gan: Synthesizing images between two domains by genetic crossover. In European Conference on Computer Vision, pages 312-326. Springer, 2020. 2", + "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[45] Yong Zhang, Le Li, Zhilei Liu, Baoyuan Wu, Yanbo Fan, and Zhifeng Li. Controllable descendant face synthesis. arXiv preprint arXiv:2002.11376, 2020. 2", + "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3", + "[47] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2", + "[48] Yuzhi Zhao, Lai-Man Po, Xuehui Wang, Qiong Yan, Wei Shen, Yujia Zhang, Wei Liu, Chun-Kit Wong, Chiu-Sing Pang, Weifeng Ou, et al. Childpredictor: A child face prediction framework with disentangled learning. IEEE Transactions on Multimedia, 2022. 2, 6, 7, 8", + "[49] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2", + "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20969", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_model.json b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_model.json new file mode 100644 index 0000000000000000000000000000000000000000..81997a1db6c3014f91df6ea44f31d4f93d70e347 --- /dev/null +++ b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.048 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.877, + 0.178 + ], + "angle": 0, + "content": "StyleGene: Crossover and Mutation of Region-level Facial Genes for Kinship Face Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.203, + 0.754, + 0.221 + ], + "angle": 0, + "content": "Hao Li\\(^{1}\\), Xianxu Hou\\(^{2,5}\\), Zepeng Huang\\(^{1}\\), Linlin Shen\\(^{1,2,3,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.222, + 0.924, + 0.239 + ], + "angle": 0, + "content": "1Computer Vision Institute, College of Computer Science and Software Engineering, Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.24, + 0.904, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.258, + 0.78, + 0.275 + ], + "angle": 0, + "content": "3Shenzhen Institute of Artificial Intelligence and Robotics for Society" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.275, + 0.854, + 0.293 + ], + "angle": 0, + "content": "\\(^{4}\\)Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.293, + 0.814, + 0.311 + ], + "angle": 0, + "content": "\\(^{5}\\)School of AI and Advanced Computing, Xi'an Jiaotong-Liverpool University" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.313, + 0.869, + 0.327 + ], + "angle": 0, + "content": "{haoli2021, huangzepeng2021}@email.szu.edu.cn, hxianxu@gmail.com, llshen@szu.edu.cn" + }, + { + "type": "aside_text", + "bbox": [ + 0.104, + 0.349, + 0.124, + 0.415 + ], + "angle": 90, + "content": "Grandparents" + }, + { + "type": "aside_text", + "bbox": [ + 0.107, + 0.43, + 0.124, + 0.468 + ], + "angle": 90, + "content": "Parents" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.351, + 0.252, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.351, + 0.375, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.4, + 0.375, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.351, + 0.497, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.351, + 0.621, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.351, + 0.744, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.351, + 0.868, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.501, + 0.314, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.501, + 0.436, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.501, + 0.559, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.501, + 0.681, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.501, + 0.805, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.596, + 0.893, + 0.625 + ], + "angle": 0, + "content": "Figure 1. Our StyleGene method synthesizes kinship faces with resemblance to parents, exhibiting diversity and reasonable variations. The first row is the input grandparents, and the second and third rows are their descendants generated by our method." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.636, + 0.314, + 0.652 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.473, + 0.88 + ], + "angle": 0, + "content": "High-fidelity kinship face synthesis has many potential applications, such as kinship verification, missing child identification, and social media analysis. However, it is challenging to synthesize high-quality descendant faces with genetic relations due to the lack of large-scale, high-quality annotated kinship data. This paper proposes RFG (Region-level Facial Gene) extraction framework to address this issue. We propose to use IGE (Image-based Gene Encoder), LGE (Latent-based Gene Encoder) and Gene Decoder to learn the RFGs of a given face image, and the relationships between RFGs and the latent space of StyleGAN2. As cycle-like losses are designed to measure the \\(\\mathcal{L}_2\\) distances between the output of Gene Decoder and image encoder, and that between the output of LGE and IGE, only" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.896, + 0.804 + ], + "angle": 0, + "content": "face images are required to train our framework, i.e. no paired kinship face data is required. Based upon the proposed RFGs, a crossover and mutation module is further designed to inherit the facial parts of parents. A Gene Pool has also been used to introduce the variations into the mutation of RFGs. The diversity of the faces of descendants can thus be significantly increased. Qualitative, quantitative, and subjective experiments on FIW, TSKinFace, and FF-Databases clearly show that the quality and diversity of kinship faces generated by our approach are much better than the existing state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.83, + 0.633, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Humans can identify kinship through photographs based on the resemblance between parents and children. Many works have investigated this intrinsic relation in the fields" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20960" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "of kinship verification [9, 32, 42] and genetics [4, 5, 8, 19]. With the popularity of face synthesis and editing technology in recent years, high-fidelity kinship face synthesis has also attracted much attention. This task, aiming to synthesize the faces of descendants based on the appearance of the parents, has many potential applications, such as finding long-lost children, crime investigations, kinship verification, and multimedia social applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.546 + ], + "angle": 0, + "content": "In recent years, many efforts have been made to make use of generative models [7, 12, 15, 16, 27, 29, 38, 43, 45, 48] for kinship face synthesis. These works can be categorized into two paradigms: one-stage and two-stage. The one-stage paradigm [12, 29, 38, 45] treats this problem as an image-to-image translation task and trains a one-to-one kinship face generator with paired data. However, these approaches can only produce low-resolution images and the resultant images can be blurry and lack diversity. Further, it would be quite difficult to obtain annotated kinship data. By contrast, the two-stage paradigm [7, 15, 16, 27, 43, 48] first extracts the genetic representation and assembles them into children's representation based on the parents' faces. To obtain genetic representation, existing methods try to learn the inheritance and variation of facial appearances by training deep neural networks [14, 27, 48] or via a knowledge rule [7]. However, the learned genetic representation is prone to overfitting due to the lack of high-quality kinship annotated training data, resulting in a lack of diversity in the generated children. In addition, these methods cannot provide fine-grained attributes representation, and thus the generated facial attributes lack interpretability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.833 + ], + "angle": 0, + "content": "In this paper, the facial genetic process is abstracted as the exchange and mutation of the parents' facial parts. We propose an Image-based Gene Encoder (IGE) to construct an independent representation for each facial part, called a Region-level Facial Gene (RFG), which is used to control the synthesis of facial regions. We further simulate the crossover and mutation process to assemble the RFGs of descendants by using those of the parents, and our proposed Gene Pool used in the mutation process can significantly increase the diversity of the generated descendants. We use the pre-trained StyleGAN2 [24] as the generator to synthesize high-fidelity faces. To achieve this, we use a Gene Decoder to map RFGs to the \\(\\mathcal{W}^+\\) space of StyleGAN2. Since IGE requires a facial parsing mask to generate the RFG, we additionally train a Latent-based Gene Encoder (LGE) to directly map the latent code of StyleGAN2 to RFGs. Thus, facial parsing mask is not required for the RFG extraction in the inference stage. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose StyleGene to synthesize high-fidelity kinship faces with controllable facial genetic regions, via modeling the facial genetic relations based on the proposed region-level facial genes." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "- A novel genetic strategy is further introduced by simulating the crossover and mutation process to generate the RFGs of descendants. We introduce a Gene Pool into the mutation process to significantly increase the diversity of the kinship face." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.171, + 0.892, + 0.246 + ], + "angle": 0, + "content": "- We validate the effectiveness of our approach on several benchmarks, demonstrating the superiority of our StyleGene framework over other state-of-the-art methods, in terms of the quality and diversity of the generated kinship faces." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.642, + 0.267 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.277, + 0.867, + 0.293 + ], + "angle": 0, + "content": "2.1. Manipulation in Latent Space of StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Generative Adversarial Networks (GANs) [17] have been widely used in face generation [21, 33, 47] and editing [3, 13, 50]. In particular, the StyleGAN [22-24] has attracted much attention due to its ability to synthesize high-fidelity images. The StyleGAN generator first maps a latent code \\( z \\in \\mathcal{Z} \\) drawn from a normal distribution to an intermediate latent code \\( w \\in \\mathcal{W} \\) by a mapping function to control image generation. Previous studies [20, 23, 36] have demonstrated that \\( \\mathcal{W} \\) space has learned facial attribute semantics and different layers of latent code control different levels of image attributes. Based on this finding, many works [6, 49] tried to invert the real image to \\( \\mathcal{W} \\) space for semantic face editing. Some recent works [1, 2, 34] show that the \\( \\mathcal{W}^+ \\) space extended from the \\( \\mathcal{W} \\) space has lower reconstruction errors. In this paper, we build a link between our proposed RFGs and the \\( \\mathcal{W}^+ \\) space of StyleGAN2 [24], based on which we can achieve fine-grained control over the synthesis of facial regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.715, + 0.596 + ], + "angle": 0, + "content": "2.2. Kinship Face Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Kinship face synthesis aims to synthesize the face images of descendants given the images of parents. The challenge is to learn facial genetic relations with limited kinship data. Table 1 summarizes the key differences between ours and existing methods. Early works [14,29,48] design the image level mapping from parents to children via supervised learning. However, limited by the quality and scale of training data, the quality of images produced by these methods is usually low and prone to overfitting. Recent works" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.745, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Table 1. Comparison between StyleGene and existing kinship face synthesis methods. \\(N_{r}\\) denotes the number of controllable regions." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.778, + 0.892, + 0.911 + ], + "angle": 0, + "content": "
MethodsStagesKinship annotationImage resolutionDiversity ControllingNr
DNA-Net [14]two✓128Noise-
KinshipGAN [29]one✓128Noise-
ChildPredictor [48]two✓128Noise-
CDFS [45]oneX256Noise5
StyleDNA [27]two✓1024Noise-
ChildGAN [7]twoX1024Interpolation5
StyleGene (Ours)twoX1024Gene Pool34
" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20961" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.345, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.308, + 0.264, + 0.321 + ], + "angle": 0, + "content": "(a) Training Stage" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.091, + 0.882, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.307, + 0.668, + 0.32 + ], + "angle": 0, + "content": "(b) Inference Stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Figure 2. The overall framework of our method. The image-based gene encoder \\( E_{img} \\) learns to encode an independent RFG representation for each facial region in the training stage. The gene decoder \\( D \\) maps RFGs \\( g \\) to the \\( \\mathcal{W}^+ \\) space of StyleGAN2. For ease of use, our latent-based gene encoder \\( E_{lat} \\) maps latent codes \\( w \\) (obtained by an image inversion encoder \\( E_{inv} \\)) to RFGs. In the inference stage, the RFGs of both parents are first extracted using latent-based gene encoder, and then the RFGs of the descendants are assembled by crossover and mutation module. The RFGs will be mapped back to the \\( \\mathcal{W}^+ \\) space using gene decoder, and finally the high-fidelity face is generated by a pre-trained StyleGAN2 generator \\( G \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[7,27] based on StyleGAN can generate high-resolution images by interpolating in the latent space to control the generation of descendants. However, StyleDNA [27] learns the mapping from parents to children in \\(\\mathcal{W}\\) space by supervised learning, resulting in poor diversity due to overfitting. In addition, the inherited regions of children are uncontrollable. ChildGAN [7] calculates the direction vectors of different facial regions through annotated landmarks, and controls the genetic regions by interpolation. However, due to attribute entanglement in \\(\\mathcal{W}\\) space, finding the disentangled direction vectors for small facial regions is difficult, which limits the diversity of generated descendants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.469, + 0.68 + ], + "angle": 0, + "content": "To address this problem, our proposed StyleGene first learns region-level facial genes (RFGs) to control the synthesis of face regions. Then we further model kinship relations based on RFGs, and leverage the pre-trained StyleGAN2 generator to generate kinship face." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.167, + 0.709 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.794 + ], + "angle": 0, + "content": "In this section, we present our approach for high-fidelity kinship face synthesis. Given a pair of parental face images, \\( x_{f} \\) and \\( x_{m} \\), our goal is to synthesize face images of their descendants. We consider the facial genetic process as the inheritance of facial parts from the father and mother." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The overview of the proposed framework is shown in Fig. 2. In training, we build models to extract Region-level Facial Genes (RFGs) from StyleGAN2 latent space and then learn to map the disentangled genes representations back to the corresponding latent code. In particular, we use an Image-based Gene Encoder (IGE) to directly extract RFGs from the input face with region annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Then a Gene Decoder is trained to map the obtained RFGs to StyleGAN2 latent space. In addition, we use a GAN inversion method to embed the same face to the \\(\\mathcal{W}^+\\) latent space of StyleGAN2, and a Latent-based Gene Encoder (LGE) is used to decompose the obtained latent code into the corresponding RFGs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.576 + ], + "angle": 0, + "content": "In inference, we first use LGE to extract the RFGs of parental faces, which are then used to obtain children's RFGs via crossover and mutation process. In addition, we build a Gene Pool to simulate the genetic variations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.726, + 0.604 + ], + "angle": 0, + "content": "3.1. Region-level Facial Gene" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.734 + ], + "angle": 0, + "content": "In order to build facial genetic relations, a key step of the proposed framework is to extract disentangled representations of different face regions. We propose to train IGE to directly extract RFGs by using fine-grained annotated facial parts. We then use a Gene Decoder to transform the obtained gene representations into the StyleGAN2 latent space, from which we can reconstruct the original face image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Fine-grained Facial Parts Segmentation. We follow DatasetGAN [46] pipeline to generate face images and corresponding segmentation masks. DatasetGAN trains a shallow decoder to achieve semantic segmentation based on the features of StyleGAN by only using a few annotated data. In this work, we adapt their approach to StyleGAN2 to produce pixel-level labels for \\(N\\) (\\(N = 34\\)) facial parts. These fine-grained facial parts are then used to extract the corresponding facial genes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image-based Gene Encoder. Fig. 4 shows the overview of our IGE \\( E_{img}(\\cdot, \\cdot) \\). The goal of IGE is to learn region-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20962" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.87, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Figure 3. The proposed Latent-based Gene Encoder (a) and Gene Decoder (b) architecture. We mainly use the Mixer layer [39] to deal with the nonlinear transformation between high-dimensional vectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.435, + 0.296 + ], + "angle": 0, + "content": "level facial gene representation \\( \\pmb{g}_i \\) for each facial part." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.463 + ], + "angle": 0, + "content": "Given a face image \\( x \\in \\mathbb{R}^{3 \\times 512 \\times 512} \\) and the corresponding mask \\( m = \\{m_1, \\dots, m_N\\} \\), we first transform \\( x \\) to a feature map \\( f(x) \\) with \\( 64 \\times 64 \\) resolution by a series of convolutional layers \\( f \\). Then we extract the region feature \\( r_i \\) for the \\( i \\)-th face part by masking the obtained \\( f(x) \\) with the corresponding binary mask \\( m_i \\), i.e., \\( r_i = f(x) \\odot m_i \\), where \\( \\odot \\) is an element-wise matrix multiplication operator. Inspired by variational autoencoder [25], the posterior distribution for each region \\( r_i \\), mapped by a mapping function \\( h_i \\), is modeled as a multivariate Gaussian distribution. It is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.468, + 0.469, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {g}} _ {i} \\sim q (\\hat {\\boldsymbol {g}} _ {i} | \\boldsymbol {r} _ {i}) = \\mathcal {N} (\\hat {\\boldsymbol {g}} _ {i}; \\hat {\\boldsymbol {\\mu}} _ {i}, \\hat {\\boldsymbol {\\sigma}} _ {i} ^ {2} \\boldsymbol {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.469, + 0.567 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{\\mu}}_i,\\hat{\\pmb{\\sigma}}_i\\in \\mathbb{R}^{18\\times 512}\\) are the multi-dimensional output of \\(h_i(r_i)\\), representing the mean and standard deviation in a diagonal matrix form, respectively. According to [25], backpropagation is made differentiable via the reparameterization trick, thus \\(\\hat{\\pmb{g}}_i\\in \\mathbb{R}^{18\\times 512}\\) can be sampled by" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.573, + 0.469, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {g}} _ {i} = \\hat {\\boldsymbol {\\mu}} _ {i} + \\hat {\\boldsymbol {\\sigma}} _ {i} \\odot \\boldsymbol {\\epsilon}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.469, + 0.641 + ], + "angle": 0, + "content": "Finally, since we use \\(N\\) facial parts, all the RFG can be denoted by \\(\\hat{\\pmb{g}} = [\\hat{\\pmb{g}}_1,\\dots ,\\hat{\\pmb{g}}_N]\\in \\mathbb{R}^{18\\times 512\\times N}\\). We simplify this process denoted as \\(\\hat{\\pmb{g}} = E_{img}(x,m)\\)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.659, + 0.442, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.85, + 0.428, + 0.864 + ], + "angle": 0, + "content": "Figure 4. The proposed Image-based Gene Encoder." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Latent-based Gene Encoder. As shown above, the proposed IGE needs a face mask to extract the representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.387 + ], + "angle": 0, + "content": "for each facial region. To eliminate the dependency, we train a LGE \\( E_{lat}(\\cdot) \\) to directly extract RFGs from StyleGAN2 latent space. As shown in Fig. 3(a), our LGE mainly consists of a fully connected layer and several Mixer Layers [39], and maps the latent code \\( \\boldsymbol{w} \\) of a face image \\( x \\) to the mean \\( \\boldsymbol{\\mu}_i \\) and standard deviation \\( \\sigma_i \\) for each pre-defined face region. This process can be formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.393, + 0.891, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {\\mu}, \\boldsymbol {\\sigma} \\right\\} = E _ {l a t} (\\boldsymbol {w}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.49 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu} = [\\pmb{\\mu}_1, \\dots, \\pmb{\\mu}_N]\\) and \\(\\pmb{\\sigma} = [\\pmb{\\sigma}_1, \\dots, \\pmb{\\sigma}_N]\\), and we use a pre-trained GAN inversion model \\(E_{inv}(\\cdot)\\) [40] as the image encoder to get the latent code of a face image, i.e. \\(\\pmb{w} = E_{inv}(x)\\). Then the region-level gene \\(\\pmb{g}_i\\) for the \\(i\\)-th facial region can be sampled by" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.496, + 0.891, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {i} = \\boldsymbol {\\mu} _ {i} + \\boldsymbol {\\sigma} _ {i} \\odot \\boldsymbol {\\epsilon}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Finally, the RFGs of \\(N\\) facial parts can be denoted by \\(\\pmb{g} = [g_1,\\dots ,g_N]\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.683 + ], + "angle": 0, + "content": "Gene Decoder. We then use Gene Decoder \\(D(\\cdot)\\) to map the region-level facial genes into the StyleGAN2 latent space. Thus, we can reconstruct the face images from the obtained gene representation. Specifically, the Gene Decoder first processes the obtained RFGs with several Mixer Layers [39]. Then all the regional features are concatenated and further processed by a fully connected layer to obtain the latent code \\(\\hat{\\boldsymbol{w}}\\in \\mathbb{R}^{18\\times 512}\\). This process can be formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.684, + 0.891, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {w}} = D (\\hat {\\boldsymbol {g}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.707, + 0.862, + 0.722 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{g}} = [\\hat{\\pmb{g}}_1,\\dots ,\\hat{\\pmb{g}}_N]\\) is the RFGs of \\(N\\) facial parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.722, + 0.891, + 0.782 + ], + "angle": 0, + "content": "Loss Functions. As shown in Fig. 2(a), given an input face image \\( x \\), the corresponding region mask \\( m \\) and the latent code \\( w = E_{inv}(x) \\), we train our models by using following loss functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.783, + 0.892, + 0.843 + ], + "angle": 0, + "content": "To build a bridge between RFGs and latent space, we first constrain the IGE \\( E_{img} \\) and Gene Decoder \\( D \\) to reconstruct the latent code of the input image. The latent reconstruct loss is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.849, + 0.891, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l a t e n t}} = \\left\\| \\boldsymbol {w} - \\hat {\\boldsymbol {w}} \\right\\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.694, + 0.886 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{w}} = D(E_{img}(x,m))\\)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, we use LGE \\( E_{lat} \\) to remove the reliance on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20963" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "face annotations. To decouple the latent code into the RFGs obtained from IGE, the LGE is trained by" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.125, + 0.469, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g e n e}} = \\sum_ {i = 1} ^ {N} \\left[ \\| \\boldsymbol {\\mu} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {i} \\| _ {2} + \\| \\boldsymbol {\\sigma} _ {i} - \\hat {\\boldsymbol {\\sigma}} _ {i} \\| _ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.168, + 0.469, + 0.214 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{\\mu}}_i\\) and \\(\\hat{\\pmb{\\sigma}}_i\\) are the outputs of \\(E_{img}(x,m)\\) for the \\(i\\)-th facial region, and \\(\\pmb{\\mu}_i, \\pmb{\\sigma}_i\\) are the outputs of \\(E_{lat}(\\pmb{w})\\) for the \\(i\\)-th facial region." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.214, + 0.469, + 0.243 + ], + "angle": 0, + "content": "All modules are trained jointly, and the full loss is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.245, + 0.469, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {l a t e n t}} + \\lambda \\mathcal {L} _ {\\text {g e n e}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.268, + 0.469, + 0.299 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter, and we set it to be 1 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.299, + 0.469, + 0.344 + ], + "angle": 0, + "content": "Note that this training process does not rely on kinship data, alleviating the problem of insufficient high-quality annotated kinship datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.354, + 0.286, + 0.369 + ], + "angle": 0, + "content": "3.2. Crossover & Mutation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.378, + 0.469, + 0.496 + ], + "angle": 0, + "content": "As parent-offspring resemblance is often reflected in the local area of the face appearance [5, 31]. Thus, after obtaining the region-level facial genes (RFGs), we simulate a crossover process to generate the RFGs of descendants by combining the RFGs of the parents. Additionally, we further build a gene pool to simulate the gene mutation process, which can significantly increase the diversity of synthesized faces." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.499, + 0.469, + 0.621 + ], + "angle": 0, + "content": "Gene Crossover. The gene crossover process aims to generate the gene representation of the descendants with a linear combination of the RFGs of the parents. In particular, given a pair of parental face images, \\( x_{f} \\) and \\( x_{m} \\), we first apply Image Encoder and LGE to obtain a set of RFGs \\( g^{f} = [g_{1}^{f},\\dots ,g_{N}^{f}] \\) and \\( g^{m} = [g_{1}^{m},\\dots ,g_{N}^{m}] \\) from \\( N \\) facial parts, respectively. Then, the RFGs of descendants \\( g^{c} = [g_{1}^{c},\\dots ,g_{N}^{c}] \\) can be calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.623, + 0.469, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {i} ^ {c} = \\alpha_ {i} \\boldsymbol {g} _ {i} ^ {f} + \\beta_ {i} \\boldsymbol {g} _ {i} ^ {m}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.644, + 0.469, + 0.674 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) are randomly generated weights for each region and \\(\\alpha_{i} + \\beta_{i} = 1\\) when mutation is not applied." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Fig. 5 shows four example faces generated using RFGs inherited from their parents based on Eq. 9, when different combinations of \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) are applied. The bigger value of \\(\\alpha_{i}\\), the more similar to father for the \\(i\\)-th facial region. The figure shows that our approach can precisely and independently control the similarity of each facial region with their parents, by setting different values of \\(\\alpha_{i}\\) and \\(\\beta_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Gene Mutation. In genetics, the gene pool is the set of all genes of all individuals in a population [30]. For humans, all races share the same gene pool. We introduce the gene pool concept to better simulate the genetic variation in facial appearances. We define the gene pool \\(\\mathcal{P}\\) as the sets formed by grouping a large number of RFGs, i.e. \\(\\mathcal{P} = \\{\\pmb{g}^1,\\dots ,\\pmb{g}^p,\\dots ,\\pmb{g}^P\\}\\), where \\(\\pmb{g}^{p} = [g_{1}^{p},\\dots ,g_{N}^{p}]\\). As shown in the supplementary materials, we divide the gene" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.868, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.325, + 0.892, + 0.367 + ], + "angle": 0, + "content": "Figure 5. Example of different weights used in crossover of RFGs. While the redder regions look more similar to the father, and the bluer regions look more similar to the mother." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.447 + ], + "angle": 0, + "content": "pool into nine groups by age, two by gender, and seven by race. Note that the RFG in the gene pool can independently control different facial traits. And its phenotype is related to the grouping condition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.707 + ], + "angle": 0, + "content": "Given a pair of parental face images and information about generated descendants, such as age, gender, and race. We first query the gene pool to extract a subset \\(\\mathcal{P}_c\\subseteq \\mathcal{P}\\) that satisfies the descendants' information, which will be used to simulate gene mutations for better diversity in the descendants. Then we use Image Encoder and LGE to extract RFGs \\(\\pmb{g}^{f} = [g_{1}^{f},\\dots ,g_{N}^{f}]\\) and \\(\\pmb{g}^{m} = [g_{1}^{m},\\dots ,g_{N}^{m}]\\) of parents. For mutation, we randomly select \\(\\eta\\) percent of facial regions whose RFGs are sampled from the gene pool. Given a \\(\\pmb{g}^p\\) sampled from \\(\\mathcal{P}_c\\), a one-hot vector \\(t = \\{t_i\\in \\{0,1\\} ,i = 1,2,\\dots ,N\\}\\) is randomly generated to denote whether the \\(\\pmb{g}_i^c\\) of descendant is fully copied from \\(\\pmb{g}_i^p (t_i = 1)\\), or not \\((t_i = 0)\\). When \\(\\pmb{g}_i^c\\) is inherited from parents \\((t_i = 0)\\), we also introduce mutation by introducing \\(\\pmb{g}^p\\) into the crossover process defined in Eq. 9. Finally, we can extend Eq. 9 as follows to obtain the RFGs of the descendants \\(\\pmb{g}^c = [\\pmb {g}_c^1,\\dots ,\\pmb {g}_c^N ]\\)" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.711, + 0.891, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {i} ^ {c} = \\left\\{ \\begin{array}{l l} \\boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 1 \\\\ \\alpha_ {i} \\boldsymbol {g} _ {i} ^ {f} + \\beta_ {i} \\boldsymbol {g} _ {i} ^ {m} + \\gamma \\boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 0 \\end{array} , \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.853 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the intensity of mutation, \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) are randomly generated weights and \\(\\alpha_{i} + \\beta_{i} = 1 - \\gamma\\), \\(g_{i}^{p} = S(\\mathcal{P}_{c})\\) and \\(S(\\cdot)\\) is the random sampling operator. Once the RFG of descendant \\(g^{c}\\) is generated, the learned Gene Decoder is applied to map \\(g^{c}\\) to the latent space of StyleGAN2, i.e. \\(\\boldsymbol{w}_{c} = D(\\boldsymbol{g}^{c})\\), \\(\\boldsymbol{w}_{c} \\in \\mathbb{R}^{18 \\times 512}\\), which is further processed for face synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While the latent code in \\(\\mathcal{W}^+\\) latent space of StyleGAN2 usually consists of 18 layers, Richardson et al. [34] demonstrate that the first 8 layers of the latent code mainly con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20964" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "tribute to the ID information of the synthesized faces, and higher layers mainly control the skin color and microstructure of the synthesized faces. While the appearance of facial regions for descendants is mainly decided by \\( g^{c} \\), other attributes like skin, hair color, image lighting, and background, can thus be inherited by inclusion of the latent codes of parents. Therefore, we further fuse the latent codes of parents \\( \\boldsymbol{w}_{f} \\) and \\( \\boldsymbol{w}_{m} \\). We keep the first \\( l \\) layer of \\( \\boldsymbol{w}_{c} \\), and the higher layers are fused from \\( \\boldsymbol{w}_{f} \\) and \\( \\boldsymbol{w}_{m} \\). The final latent code \\( \\boldsymbol{w}_{c} = [\\boldsymbol{w}_{c}^{1},\\dots ,\\boldsymbol{w}_{c}^{18}] \\) of the descendant can be obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.26, + 0.47, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {w} _ {c} ^ {i} = \\frac {1}{2} \\left(\\boldsymbol {w} _ {f} ^ {i} + \\boldsymbol {w} _ {m} ^ {i}\\right), \\quad i \\in \\{l + 1, \\dots , 1 8 \\}, \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.301, + 0.21, + 0.318 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.326, + 0.284, + 0.342 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.621 + ], + "angle": 0, + "content": "Datasets. Our training set consists of three parts, all images of CelebAHQ [21], 50,000 faces sampled from MS-Celeb-1M [18], and 10,000 faces generated by StyleGAN2, to make the model adaptive to different image qualities. CelebA-HQ is a high-quality face dataset consisting of 30,000 aligned face images with \\(1024 \\times 1024\\) resolution. The MS-Celeb-1M is a large-scale face recognition dataset containing about 10 million images, including many noisy face images. We evaluate our model on the FIW [35], TSKinFace [32], and FF-Database [48] datasets. FIW dataset contains the faces of 1,000 families with 1,997 sets of father-mother-child relations. The TSKinFace and FF-Database datasets provide 1,015 and 3,744 sets of father-mother-child kinship face images, respectively. We align and crop the images in these two datasets to \\(256 \\times 256\\). Our Gene Pool is built based on the FFHQ [23] dataset, which consists of 70,000 high-quality face images, and has a better diversity among human races and ages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.757 + ], + "angle": 0, + "content": "Baselines. We compare our method against the state-of-art kinship face synthesis baselines, i.e. DNA-Net [14], ChildGAN [7], ChildPredictor [48], and StyleDNA [27]. Since the source codes of StyleDNA and ChildPredictor are available, we compare our approach with them through all of the evaluations. However, due to the unavailability of the source codes, we can only visually compare the visual quality of the generated faces available in the papers of DNA-Net and ChildGAN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Training Details. We use AdamW [28] optimizer with batch size of 32. The initialized learning rate is 0.001, which is divided by 2 every 10 epochs, and we stop the training at the 30th epoch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.285, + 0.848 + ], + "angle": 0, + "content": "4.2. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Disentanglement of RFGs. Fig. 6 shows two examples of facial region editing using our proposed RFG, where the nose, eyes, jaw and lips of source faces shown in the first" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "column, are replaced sequentially by the corresponding regions of reference faces shown in the first row. One can observe from the examples that, when a certain region is edited, all other facial regions are kept intact, which clearly shows the disentanglement capability of our RFG." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.178, + 0.878, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.37, + 0.839, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Disentangled editing of facial regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Effectiveness of Gene Pool. Based on the proposed idea of RFG, our approach further introduces a gene pool to increase the variations among the generated faces of descendants. Given a couple of parents shown in the first row, Fig. 7 shows the faces of five children generated by our approach with/without the involvement of gene pool, together with that synthesized by StyleDNA and ChildPredictor. While the faces generated by StyleDNA and ChildPredictor (4th and 5th rows) all look very similar, that synthesized by our approach are much more diverse and the introduction of gene pool (1st row) can further increase variations among children." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.585, + 0.878, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.837, + 0.892, + 0.907 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of diversity with StyleDNA and ChildPredictor. The first row shows the father and mother. The next two rows are generated by our method with and without Gene Pool. The last two rows are generated by StyleDNA [27] and ChildPredictor [48], respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.532, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.273, + 0.132, + 0.283 + ], + "angle": 0, + "content": "Father" + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.273, + 0.187, + 0.282 + ], + "angle": 0, + "content": "Mother" + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.273, + 0.249, + 0.283 + ], + "angle": 0, + "content": "Real Child" + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.273, + 0.29, + 0.282 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.273, + 0.339, + 0.283 + ], + "angle": 0, + "content": "StyleDN" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.273, + 0.4, + 0.283 + ], + "angle": 0, + "content": "A ChildPre" + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.273, + 0.479, + 0.283 + ], + "angle": 0, + "content": "Victor ChildGAN" + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.273, + 0.532, + 0.282 + ], + "angle": 0, + "content": "DNA-Net" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.88, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.273, + 0.582, + 0.282 + ], + "angle": 0, + "content": "Father" + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.273, + 0.639, + 0.282 + ], + "angle": 0, + "content": "Mother" + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.273, + 0.68, + 0.282 + ], + "angle": 0, + "content": "Real 0" + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.273, + 0.748, + 0.282 + ], + "angle": 0, + "content": "d Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.273, + 0.89, + 0.283 + ], + "angle": 0, + "content": "StyleDNA ChildPredictor" + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.286, + 0.756, + 0.297 + ], + "angle": 0, + "content": "(b) FF-Database" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 8. Comparison of children faces synthesized by StyleGene and baselines. Left most three columns of (a) FIW and (b) FF-Database depict the father, mother, and real children, and right five or three columns depict the faces synthesized by ours, StyleDNA [27], ChildPredictor [48], ChildGAN [7] and DNA-Net [14]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.473, + 0.751 + ], + "angle": 0, + "content": "Comparison with the state-of-the-art. Fig. 8(a) and (b) show qualitative results on FIW and FF-Database, respectively. We first compare StyleGene with StyleDNA, DNA-Net, ChildGAN, and ChildPredictor on FIW with low-resolution facial images. Note that the results of ChildGAN and DNA-Net are directly taken from their papers. As can be seen, DNA-Net, ChildGAN, and ChildPredictor struggle to yield high-quality facial images. Thanks to a powerful generator, our method can generate high-fidelity faces with hair. Though ChildGAN also uses the StyleGAN generator, its results are significantly affected by the quality of the input. Note that all comparisons have considered the facial information of parents. However, DNA-Net, ChildPredictor, and StyleDNA model facial genetics by learning the implicit mapping between parents and children, without considering regional similarity, which significantly compromised the similarity between the generated faces and their parents. Although ChildGAN considered parents' facial features, the results were too similar to their parents. Thanks to the region-level facial genes and the mutation introduced by gene pool, our results can maintain the local similarity to parents' faces, and at the same time, present reasonable global variations and diversity. When experimenting with faces of different races as shown in Fig. 8(b), our approach can still well preserve the race similarity of children with their parents." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.295, + 0.785 + ], + "angle": 0, + "content": "4.3. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Kinship Verification. We now use kinship verification accuracy to assess whether a genetic relationship exists between the synthetic descendant and parents. Higher accuracy means more realistic synthetic descendants. We use ArcFace pre-trained on the MS-Celeb-1M dataset as the backbone to create a one-versus-one kinship classifier [42], which is fine-tuned on the FF-Database training set. Then" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.359, + 0.782, + 0.373 + ], + "angle": 0, + "content": "we performed a cross-database evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Specifically, we randomly sample 100 families with one kid from the FF-database test set, FIW, and TSKinFace datasets, respectively, and different methods randomly generate 40 children for each family. We randomly sample 8,000 positive and 8,000 negative pairs for each method from each dataset and calculate in Table 2 the kinship verification accuracy. One can observe that our method achieves substantially higher accuracy than StyleDNA and ChildPredictor, i.e. as high as \\(81.74\\%\\), \\(80.38\\%\\), and \\(62.29\\%\\) accuracies are achieved on TSKinFace, FF-Database, and FIW datasets, respectively. The results suggest that the kinship relations of faces generated by our approach are very close to the real children." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.578, + 0.89, + 0.607 + ], + "angle": 0, + "content": "Table 2. Kinship verification accuracy \\((\\%)\\) on the TSKinFace [32], FF-Database [48], and FIW [35] dataset." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.612, + 0.892, + 0.687 + ], + "angle": 0, + "content": "
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA [27]53.1555.1149.47
ChildPredictor [48]58.2459.6251.81
StyleGene (Ours)81.7480.3862.29
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Diversity Evaluation. We used the LPIPS [44] metric to measure the diversity of synthetic descendants, which calculates the L1 distance between pairs of image features extracted by AlexNet [26] pre-trained on the ImageNet [10] dataset. We use the same test data of kinship verification. First, we calculated the distance among the 40 synthesized descendants for each family. Then we take the average of 100 families as the LPIPS score and list them in Table 3. As shown in the table, our method achieves the highest LPIPS across all three datasets, i.e. 0.3270, 0.3418, and 0.3279 LPIPS on TSKinFace, FF-Database, and FIW datasets are achieved, which is significantly higher than that of StyleDNA and ChildPredictor." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Distribution of Synthesized Children. We are now try-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20966" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of the diversity of the generated descendants. * means we cropped the face i.e. no hair." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.123, + 0.47, + 0.224 + ], + "angle": 0, + "content": "
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA* [27]0.07560.07630.0736
ChildPredictor [48]0.16970.17230.1750
StyleGene (Ours)*0.17480.17350.1740
StyleDNA [27]0.15590.15420.1573
StyleGene (Ours)0.32700.34180.3279
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.471, + 0.538 + ], + "angle": 0, + "content": "ing to model the distribution of the children synthesized by different approaches, and compare them with that of real children. Specifically, we randomly sample 50 families from the TSKinFace dataset, each consisting of a father, mother, son, and daughter. Based on the given parents, we also applied different synthesis methods to generate two children, i.e. a son and a daughter, for each family. After facial features are extracted using ArcFace [11] and reduced to one dimension using t-SNE [41], Kernel Density Estimation (KDE) [37] is further applied to estimate the features' probability density function. Fig. 9 shows the distribution of the 100 real children (blue) and that of children generated by different approaches. As can be seen, the distribution of faces generated by our approach (red) overlaps most with the real children (blue). In contrast, the distributions of faces synthesized by StyleDNA (black) and ChildPredictor (yellow) are located on both sides of the real ones. The distribution for StyleDNA even looks like a mixture of two Gaussian distributions, which is significantly different from that of real children." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.551, + 0.451, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Figure 9. Distribution of real children (blue) and children generated using StyleDNA (black), ChildPredictor (yellow), and ours (red). Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "User Study. To further demonstrate the effectiveness of the proposed method, we conducted a user study involving 59 participants. The faces of children generated by our StyleGene, StyleDNA and ChildPredictor based on 20 parents, are present to each of the 59 participants, who are asked to rank the quality of the faces, in terms of realness and the similarity with given parents. In total \\(1,180(59 \\times 20)\\) votes are received for each of the three approaches and the average rankings are listed in Table 4. As shown in the table, our StyleGene receives the highest rank among the com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "pared approaches, which suggests that the faces of children synthesized by our approach are the best, in terms of both quality and similarity with parents." + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.141, + 0.86, + 0.156 + ], + "angle": 0, + "content": "Table 4. The rank of different approaches in user study." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.16, + 0.871, + 0.201 + ], + "angle": 0, + "content": "
ChildPredictorStyleDNAStyleGene (Ours)
Avg. rank2.462.221.32
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.213, + 0.65, + 0.229 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.388 + ], + "angle": 0, + "content": "In this section, we perform ablation studies on the effectiveness of LGE, IGE and GP, in terms of kinship verification accuracy (ACC) and diversity (LPIPS) using FF-database with the same configuration as in Section 4.3. As shown in Table 5, the integration of IGE (2nd row) can improve both ACC and LPIPS of the synthesized descendant. As expected, the use of GP (3rd row) increases the diversity of synthesized faces. When both IGE and GP are integrated with LGE (4th row), the diversity can be significantly increased to 0.1735, with a comparable ACC \\((80.38\\%)\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.574, + 0.394, + 0.818, + 0.408 + ], + "angle": 0, + "content": "Table 5. Ablation study on FF-Database." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.413, + 0.842, + 0.5 + ], + "angle": 0, + "content": "
LGEIGEGPACC (%)LPIPS
1✓XX79.190.0625
2✓✓X79.660.0646
3✓X✓80.210.0839
4✓✓✓80.380.1735
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.699, + 0.528 + ], + "angle": 0, + "content": "4.5. Parameter sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.627 + ], + "angle": 0, + "content": "We test different values of \\(\\eta\\), \\(\\gamma\\), and \\(l\\) to see how the performance of synthesis varies. Due to the page limit, details about the results of different parameters are presented in supplementary. In our experiments, we choose \\(\\eta = 40\\%\\), \\(\\gamma = 0.47\\), and \\(l = 8\\) to achieve a balance between the diversity and fidelity of the generated faces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.62, + 0.652 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.798 + ], + "angle": 0, + "content": "In this paper, we have proposed StyleGene, to extract RFGs (Region-level Facial Genes) for kinship face synthesis. While crossover and mutation of RFGs are proposed to model the facial genetic process, Gene Pool is further designed to increase the diversity among generated faces. Quantitative, qualitative, and subjective experimental results show that the realness, similarity with parents, and diversity of kinship faces generated by our approach are much better than existing state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.892, + 0.888 + ], + "angle": 0, + "content": "Acknowledgement. This work was jointly supported by National Natural Science Foundation of China (62206180 and 82261138629), Guangdong Basic and Applied Basic Research Foundation (2023A1515010688 and 2022A1515011018), and Shenzhen Municipal Science and Technology Innovation Council (JCYJ20220531101412030)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20967" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.31 + ], + "angle": 0, + "content": "[3] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[4] Peter Claes, Jasmien Roosenboom, Julie D White, Tomek Swigut, Dzemila Sero, Jiarui Li, Myoung Keun Lee, Arslan Zaidi, Brooke C Mattern, Corey Liebowitz, et al. Genome-wide mapping of global-to-local genetic effects on human facial shape. Nature genetics, 50(3):414-423, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[5] Joanne B Cole, Mange Manyama, Jacinda R Larson, Denise K Liberton, Tracey M Ferrara, Sheri L Riccardi, Mao Li, Washington Mio, Ophir D Klein, Stephanie A Santorico, et al. Human facial shape and size heritability and genetic correlations. Genetics, 205(2):967-978, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[6] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[7] Xiao Cui, Wengang Zhou, Yang Hu, Weilun Wang, and Houqiang Li. Heredity-aware child face image generation with latent space disentanglement. arXiv preprint arXiv:2108.11080, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.472, + 0.606 + ], + "angle": 0, + "content": "[8] Maria F Dal Martello and Laurence T Maloney. Lateralization of kin recognition signals in the human face. Journal of Vision, 10(8):9-9, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[9] Afshin Dehghan, Enrique G Ortiz, Ruben Villegas, and Mubarak Shah. Who do i look like? determining parent-offspring resemblance via gated autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1757-1764, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.734, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[12] Itir Önal Ertugrul and Hamdi Dibeklioglu. What will your future child look like? modeling and synthesis of hereditary patterns of facial dynamics. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 33-40. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[14] Pengyu Gao, Joseph Robinson, Jiaxuan Zhu, Chao Xia, MIng Shao, and Siyu Xia. Dna-net: Age and gender aware kin face synthesizer. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Pengyu Gao, Siyu Xia, Joseph Robinson, Junkang Zhang, Chao Xia, Ming Shao, and Yun Fu. What will your child look like? dna-net: Age and gender aware kin face synthesizer. arXiv preprint arXiv:1911.07014, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[16] Fady S Ghatas and ElSayed E Hemayed. Gankin: generating kin faces using disentangled gan. SN Applied Sciences, 2(2):1-10, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[18] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In European conference on computer vision, pages 87–102. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[19] Hanne Hoskens, Dongjing Liu, Sahin Naqvi, Myoung Keun Lee, Ryan J Eller, Karlijne Indencleef, Julie D White, Jiarui Li, Maarten HD Larmouseau, Greet Hens, et al. 3d facial phenotyping by biometric sibling matching used in contemporary genomic methodologies. PLoS genetics, 17(5):e1009528, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[20] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[21] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[27] Che-Hsien Lin, Hung-Chun Chen, Li-Chen Cheng, Shu-Chuan Hsu, Jun-Cheng Chen, and Chih-Yu Wang. *Styledna: A high-fidelity age and gender aware kinship face synthesizer*. In 2021 16th IEEE International Conference on Auto" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20968" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "matic Face and Gesture Recognition (FG 2021), pages 1-8. IEEE, 2021. 2,3,6,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.468, + 0.151 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.153, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[29] Savas Ozkan and Akin Ozkan. Kinshipgan: Synthesizing of kinship faces from family photos by regularizing a deep face network. In 2018 25th IEEE international conference on image processing (ICIP), pages 2142-2146. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.468, + 0.252 + ], + "angle": 0, + "content": "[30] Svante Paäbo. The diverse origins of the human gene pool. Nature Reviews Genetics, 16(6):313-314, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.469, + 0.323 + ], + "angle": 0, + "content": "[31] Shouneng Peng, Jingze Tan, Sile Hu, Hang Zhou, Jing Guo, Li Jin, and Kun Tang. Detecting genetic association of common human facial morphological variation using high density 3d image registration. PLoS computational biology, 9(12):e1003375, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.326, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[32] Xiaoqian Qin, Xiaoyang Tan, and Songcan Chen. Tri-subject kinship verification: Understanding the core of a family. IEEE Transactions on Multimedia, 17(10):1855-1867, 2015. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[33] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[35] Joseph P Robinson, Ming Shao, Yue Wu, and Yun Fu. Families in the wild (fw) large-scale kinship image database and benchmarks. In Proceedings of the 24th ACM international conference on Multimedia, pages 242-246, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[36] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.656 + ], + "angle": 0, + "content": "[37] Bernard W Silverman. Density estimation for statistics and data analysis. Routledge, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[38] Raunak Sinha, Mayank Vatsa, and Richa Singh. Familygan: Generating kin face images using generative adversarial networks. In European conference on computer vision, pages 297-311. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[39] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[42] Jun Yu, Mengyan Li, Xinlong Hao, and Guochen Xie. Deep fusion siamese network for automatic kinship verification. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 892-899. IEEE, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[43] Ishtiak Zaman and David Crandall. Genetic-gan: Synthesizing images between two domains by genetic crossover. In European Conference on Computer Vision, pages 312-326. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[45] Yong Zhang, Le Li, Zhilei Liu, Baoyuan Wu, Yanbo Fan, and Zhifeng Li. Controllable descendant face synthesis. arXiv preprint arXiv:2002.11376, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[47] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[48] Yuzhi Zhao, Lai-Man Po, Xuehui Wang, Qiong Yan, Wei Shen, Yujia Zhang, Wei Liu, Chun-Kit Wong, Chiu-Sing Pang, Weifeng Ou, et al. Childpredictor: A child face prediction framework with disentangled learning. IEEE Transactions on Multimedia, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[49] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.601, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20969" + } + ] +] \ No newline at end of file diff --git a/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_origin.pdf b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db7a4bc2f855ac074826d28aa3044ae91468419a --- /dev/null +++ b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/88326347-639e-4732-90cc-676114c49600_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc5743b6285193cddf35e6614e0c2570df0bf896bbe24c9cabb21e99c919f4e3 +size 7259355 diff --git a/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/full.md b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..54019b29c400b3252b4cdb62c5b52386cebbc973 --- /dev/null +++ b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/full.md @@ -0,0 +1,371 @@ +# StyleGene: Crossover and Mutation of Region-level Facial Genes for Kinship Face Synthesis + +Hao Li $^{1}$ , Xianxu Hou $^{2,5}$ , Zepeng Huang $^{1}$ , Linlin Shen $^{1,2,3,4*}$ + +1Computer Vision Institute, College of Computer Science and Software Engineering, Shenzhen University + +$^{2}$ National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University + +3Shenzhen Institute of Artificial Intelligence and Robotics for Society + +$^{4}$ Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University + +$^{5}$ School of AI and Advanced Computing, Xi'an Jiaotong-Liverpool University + +{haoli2021, huangzepeng2021}@email.szu.edu.cn, hxianxu@gmail.com, llshen@szu.edu.cn + +![](images/1998cdf32cf9536c7e74081fe95dd73bc74e42df801218b6b2bea9a271e98947.jpg) + +![](images/1a7df84691f49bdb9b63d832f714d4d1b92ea94c942f0b07725b369ac10c9d00.jpg) + +![](images/f31204d2347063184e1d1459fe49d2eccd35eee75bd54927bbc8b477a174888b.jpg) + +![](images/17e0862662d2c685105cf6163b174d61d9c8c3c826b98aacae3743d7078b4016.jpg) + +![](images/10696e267e50c0c4a6f3b5dd3896f8289d409d9761cb5ff199f79ca94789a8af.jpg) + +![](images/d38b3d919d04a9e26884ac9f9bd669931c3efdf65a81b77e3f4f34f5e5702705.jpg) + +![](images/541c075681020995d83fa1ddcb7b3a9dc3131b633d251594629047f8d417fbfc.jpg) + +![](images/3798f1f0b9786875d1362048c91fa574a90ec491b8bec8353046565558faedf5.jpg) +Figure 1. Our StyleGene method synthesizes kinship faces with resemblance to parents, exhibiting diversity and reasonable variations. The first row is the input grandparents, and the second and third rows are their descendants generated by our method. + +![](images/2e81a98ba803287af67189da48ee6173fffe5c5144b06556c7760bd15e4742ab.jpg) + +![](images/176d95cd0ac1500b717cc8261922de7fa8b77f9758c897614d7eef8fd181790b.jpg) + +![](images/91501a59c42703cd6a68d80b266ca003d0d04ef4ff667a580a2f3adc9e56cc72.jpg) + +![](images/2a8d890bf2051b30ae387258cd847eab50828c718365ac1e5c23cfbef3c42d3a.jpg) + +# Abstract + +High-fidelity kinship face synthesis has many potential applications, such as kinship verification, missing child identification, and social media analysis. However, it is challenging to synthesize high-quality descendant faces with genetic relations due to the lack of large-scale, high-quality annotated kinship data. This paper proposes RFG (Region-level Facial Gene) extraction framework to address this issue. We propose to use IGE (Image-based Gene Encoder), LGE (Latent-based Gene Encoder) and Gene Decoder to learn the RFGs of a given face image, and the relationships between RFGs and the latent space of StyleGAN2. As cycle-like losses are designed to measure the $\mathcal{L}_2$ distances between the output of Gene Decoder and image encoder, and that between the output of LGE and IGE, only + +face images are required to train our framework, i.e. no paired kinship face data is required. Based upon the proposed RFGs, a crossover and mutation module is further designed to inherit the facial parts of parents. A Gene Pool has also been used to introduce the variations into the mutation of RFGs. The diversity of the faces of descendants can thus be significantly increased. Qualitative, quantitative, and subjective experiments on FIW, TSKinFace, and FF-Databases clearly show that the quality and diversity of kinship faces generated by our approach are much better than the existing state-of-the-art methods. + +# 1. Introduction + +Humans can identify kinship through photographs based on the resemblance between parents and children. Many works have investigated this intrinsic relation in the fields + +of kinship verification [9, 32, 42] and genetics [4, 5, 8, 19]. With the popularity of face synthesis and editing technology in recent years, high-fidelity kinship face synthesis has also attracted much attention. This task, aiming to synthesize the faces of descendants based on the appearance of the parents, has many potential applications, such as finding long-lost children, crime investigations, kinship verification, and multimedia social applications. + +In recent years, many efforts have been made to make use of generative models [7, 12, 15, 16, 27, 29, 38, 43, 45, 48] for kinship face synthesis. These works can be categorized into two paradigms: one-stage and two-stage. The one-stage paradigm [12, 29, 38, 45] treats this problem as an image-to-image translation task and trains a one-to-one kinship face generator with paired data. However, these approaches can only produce low-resolution images and the resultant images can be blurry and lack diversity. Further, it would be quite difficult to obtain annotated kinship data. By contrast, the two-stage paradigm [7, 15, 16, 27, 43, 48] first extracts the genetic representation and assembles them into children's representation based on the parents' faces. To obtain genetic representation, existing methods try to learn the inheritance and variation of facial appearances by training deep neural networks [14, 27, 48] or via a knowledge rule [7]. However, the learned genetic representation is prone to overfitting due to the lack of high-quality kinship annotated training data, resulting in a lack of diversity in the generated children. In addition, these methods cannot provide fine-grained attributes representation, and thus the generated facial attributes lack interpretability. + +In this paper, the facial genetic process is abstracted as the exchange and mutation of the parents' facial parts. We propose an Image-based Gene Encoder (IGE) to construct an independent representation for each facial part, called a Region-level Facial Gene (RFG), which is used to control the synthesis of facial regions. We further simulate the crossover and mutation process to assemble the RFGs of descendants by using those of the parents, and our proposed Gene Pool used in the mutation process can significantly increase the diversity of the generated descendants. We use the pre-trained StyleGAN2 [24] as the generator to synthesize high-fidelity faces. To achieve this, we use a Gene Decoder to map RFGs to the $\mathcal{W}^+$ space of StyleGAN2. Since IGE requires a facial parsing mask to generate the RFG, we additionally train a Latent-based Gene Encoder (LGE) to directly map the latent code of StyleGAN2 to RFGs. Thus, facial parsing mask is not required for the RFG extraction in the inference stage. The main contributions of this paper are summarized as follows: + +- We propose StyleGene to synthesize high-fidelity kinship faces with controllable facial genetic regions, via modeling the facial genetic relations based on the proposed region-level facial genes. + +- A novel genetic strategy is further introduced by simulating the crossover and mutation process to generate the RFGs of descendants. We introduce a Gene Pool into the mutation process to significantly increase the diversity of the kinship face. +- We validate the effectiveness of our approach on several benchmarks, demonstrating the superiority of our StyleGene framework over other state-of-the-art methods, in terms of the quality and diversity of the generated kinship faces. + +# 2. Related Work + +# 2.1. Manipulation in Latent Space of StyleGAN + +Generative Adversarial Networks (GANs) [17] have been widely used in face generation [21, 33, 47] and editing [3, 13, 50]. In particular, the StyleGAN [22-24] has attracted much attention due to its ability to synthesize high-fidelity images. The StyleGAN generator first maps a latent code $z \in \mathcal{Z}$ drawn from a normal distribution to an intermediate latent code $w \in \mathcal{W}$ by a mapping function to control image generation. Previous studies [20, 23, 36] have demonstrated that $\mathcal{W}$ space has learned facial attribute semantics and different layers of latent code control different levels of image attributes. Based on this finding, many works [6, 49] tried to invert the real image to $\mathcal{W}$ space for semantic face editing. Some recent works [1, 2, 34] show that the $\mathcal{W}^+$ space extended from the $\mathcal{W}$ space has lower reconstruction errors. In this paper, we build a link between our proposed RFGs and the $\mathcal{W}^+$ space of StyleGAN2 [24], based on which we can achieve fine-grained control over the synthesis of facial regions. + +# 2.2. Kinship Face Synthesis + +Kinship face synthesis aims to synthesize the face images of descendants given the images of parents. The challenge is to learn facial genetic relations with limited kinship data. Table 1 summarizes the key differences between ours and existing methods. Early works [14,29,48] design the image level mapping from parents to children via supervised learning. However, limited by the quality and scale of training data, the quality of images produced by these methods is usually low and prone to overfitting. Recent works + +Table 1. Comparison between StyleGene and existing kinship face synthesis methods. $N_{r}$ denotes the number of controllable regions. + +
MethodsStagesKinship annotationImage resolutionDiversity ControllingNr
DNA-Net [14]two✓128Noise-
KinshipGAN [29]one✓128Noise-
ChildPredictor [48]two✓128Noise-
CDFS [45]oneX256Noise5
StyleDNA [27]two✓1024Noise-
ChildGAN [7]twoX1024Interpolation5
StyleGene (Ours)twoX1024Gene Pool34
+ +![](images/dc7f40045aaf7efd468d9d4d75083ccb3579bd24f1bb559f06b95902e917bbdb.jpg) +(a) Training Stage +Figure 2. The overall framework of our method. The image-based gene encoder $E_{img}$ learns to encode an independent RFG representation for each facial region in the training stage. The gene decoder $D$ maps RFGs $g$ to the $\mathcal{W}^+$ space of StyleGAN2. For ease of use, our latent-based gene encoder $E_{lat}$ maps latent codes $w$ (obtained by an image inversion encoder $E_{inv}$ ) to RFGs. In the inference stage, the RFGs of both parents are first extracted using latent-based gene encoder, and then the RFGs of the descendants are assembled by crossover and mutation module. The RFGs will be mapped back to the $\mathcal{W}^+$ space using gene decoder, and finally the high-fidelity face is generated by a pre-trained StyleGAN2 generator $G$ . + +![](images/a26cf294bb0568f61cab7d458c0929d1593aa00aea2de332e1c2a37f317300e9.jpg) +(b) Inference Stage + +[7,27] based on StyleGAN can generate high-resolution images by interpolating in the latent space to control the generation of descendants. However, StyleDNA [27] learns the mapping from parents to children in $\mathcal{W}$ space by supervised learning, resulting in poor diversity due to overfitting. In addition, the inherited regions of children are uncontrollable. ChildGAN [7] calculates the direction vectors of different facial regions through annotated landmarks, and controls the genetic regions by interpolation. However, due to attribute entanglement in $\mathcal{W}$ space, finding the disentangled direction vectors for small facial regions is difficult, which limits the diversity of generated descendants. + +To address this problem, our proposed StyleGene first learns region-level facial genes (RFGs) to control the synthesis of face regions. Then we further model kinship relations based on RFGs, and leverage the pre-trained StyleGAN2 generator to generate kinship face. + +# 3. Method + +In this section, we present our approach for high-fidelity kinship face synthesis. Given a pair of parental face images, $x_{f}$ and $x_{m}$ , our goal is to synthesize face images of their descendants. We consider the facial genetic process as the inheritance of facial parts from the father and mother. + +The overview of the proposed framework is shown in Fig. 2. In training, we build models to extract Region-level Facial Genes (RFGs) from StyleGAN2 latent space and then learn to map the disentangled genes representations back to the corresponding latent code. In particular, we use an Image-based Gene Encoder (IGE) to directly extract RFGs from the input face with region annotations. + +Then a Gene Decoder is trained to map the obtained RFGs to StyleGAN2 latent space. In addition, we use a GAN inversion method to embed the same face to the $\mathcal{W}^+$ latent space of StyleGAN2, and a Latent-based Gene Encoder (LGE) is used to decompose the obtained latent code into the corresponding RFGs. + +In inference, we first use LGE to extract the RFGs of parental faces, which are then used to obtain children's RFGs via crossover and mutation process. In addition, we build a Gene Pool to simulate the genetic variations. + +# 3.1. Region-level Facial Gene + +In order to build facial genetic relations, a key step of the proposed framework is to extract disentangled representations of different face regions. We propose to train IGE to directly extract RFGs by using fine-grained annotated facial parts. We then use a Gene Decoder to transform the obtained gene representations into the StyleGAN2 latent space, from which we can reconstruct the original face image. + +Fine-grained Facial Parts Segmentation. We follow DatasetGAN [46] pipeline to generate face images and corresponding segmentation masks. DatasetGAN trains a shallow decoder to achieve semantic segmentation based on the features of StyleGAN by only using a few annotated data. In this work, we adapt their approach to StyleGAN2 to produce pixel-level labels for $N$ ( $N = 34$ ) facial parts. These fine-grained facial parts are then used to extract the corresponding facial genes. + +Image-based Gene Encoder. Fig. 4 shows the overview of our IGE $E_{img}(\cdot, \cdot)$ . The goal of IGE is to learn region- + +![](images/e7a1117b91cd3e7a3e27b567ab99d676447c2453473644e38652985b32bdb204.jpg) +Figure 3. The proposed Latent-based Gene Encoder (a) and Gene Decoder (b) architecture. We mainly use the Mixer layer [39] to deal with the nonlinear transformation between high-dimensional vectors. + +level facial gene representation $\pmb{g}_i$ for each facial part. + +Given a face image $x \in \mathbb{R}^{3 \times 512 \times 512}$ and the corresponding mask $m = \{m_1, \dots, m_N\}$ , we first transform $x$ to a feature map $f(x)$ with $64 \times 64$ resolution by a series of convolutional layers $f$ . Then we extract the region feature $r_i$ for the $i$ -th face part by masking the obtained $f(x)$ with the corresponding binary mask $m_i$ , i.e., $r_i = f(x) \odot m_i$ , where $\odot$ is an element-wise matrix multiplication operator. Inspired by variational autoencoder [25], the posterior distribution for each region $r_i$ , mapped by a mapping function $h_i$ , is modeled as a multivariate Gaussian distribution. It is defined by + +$$ +\hat {\boldsymbol {g}} _ {i} \sim q (\hat {\boldsymbol {g}} _ {i} | \boldsymbol {r} _ {i}) = \mathcal {N} (\hat {\boldsymbol {g}} _ {i}; \hat {\boldsymbol {\mu}} _ {i}, \hat {\boldsymbol {\sigma}} _ {i} ^ {2} \boldsymbol {I}), \tag {1} +$$ + +where $\hat{\pmb{\mu}}_i,\hat{\pmb{\sigma}}_i\in \mathbb{R}^{18\times 512}$ are the multi-dimensional output of $h_i(r_i)$ , representing the mean and standard deviation in a diagonal matrix form, respectively. According to [25], backpropagation is made differentiable via the reparameterization trick, thus $\hat{\pmb{g}}_i\in \mathbb{R}^{18\times 512}$ can be sampled by + +$$ +\hat {\boldsymbol {g}} _ {i} = \hat {\boldsymbol {\mu}} _ {i} + \hat {\boldsymbol {\sigma}} _ {i} \odot \boldsymbol {\epsilon}, \boldsymbol {\epsilon} \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I}), \tag {2} +$$ + +Finally, since we use $N$ facial parts, all the RFG can be denoted by $\hat{\pmb{g}} = [\hat{\pmb{g}}_1,\dots ,\hat{\pmb{g}}_N]\in \mathbb{R}^{18\times 512\times N}$ . We simplify this process denoted as $\hat{\pmb{g}} = E_{img}(x,m)$ + +![](images/33e0bdde9596cca5ece169deeed617eff434d09b6aedb8bdb8514fe6fe57696d.jpg) +Figure 4. The proposed Image-based Gene Encoder. + +Latent-based Gene Encoder. As shown above, the proposed IGE needs a face mask to extract the representations + +for each facial region. To eliminate the dependency, we train a LGE $E_{lat}(\cdot)$ to directly extract RFGs from StyleGAN2 latent space. As shown in Fig. 3(a), our LGE mainly consists of a fully connected layer and several Mixer Layers [39], and maps the latent code $\boldsymbol{w}$ of a face image $x$ to the mean $\boldsymbol{\mu}_i$ and standard deviation $\sigma_i$ for each pre-defined face region. This process can be formulated by + +$$ +\left\{\boldsymbol {\mu}, \boldsymbol {\sigma} \right\} = E _ {l a t} (\boldsymbol {w}), \tag {3} +$$ + +where $\pmb{\mu} = [\pmb{\mu}_1, \dots, \pmb{\mu}_N]$ and $\pmb{\sigma} = [\pmb{\sigma}_1, \dots, \pmb{\sigma}_N]$ , and we use a pre-trained GAN inversion model $E_{inv}(\cdot)$ [40] as the image encoder to get the latent code of a face image, i.e. $\pmb{w} = E_{inv}(x)$ . Then the region-level gene $\pmb{g}_i$ for the $i$ -th facial region can be sampled by + +$$ +\boldsymbol {g} _ {i} = \boldsymbol {\mu} _ {i} + \boldsymbol {\sigma} _ {i} \odot \boldsymbol {\epsilon}, \boldsymbol {\epsilon} \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I}). \tag {4} +$$ + +Finally, the RFGs of $N$ facial parts can be denoted by $\pmb{g} = [g_1,\dots ,g_N]$ + +Gene Decoder. We then use Gene Decoder $D(\cdot)$ to map the region-level facial genes into the StyleGAN2 latent space. Thus, we can reconstruct the face images from the obtained gene representation. Specifically, the Gene Decoder first processes the obtained RFGs with several Mixer Layers [39]. Then all the regional features are concatenated and further processed by a fully connected layer to obtain the latent code $\hat{\boldsymbol{w}}\in \mathbb{R}^{18\times 512}$ . This process can be formulated by + +$$ +\hat {\boldsymbol {w}} = D (\hat {\boldsymbol {g}}), \tag {5} +$$ + +where $\hat{\pmb{g}} = [\hat{\pmb{g}}_1,\dots ,\hat{\pmb{g}}_N]$ is the RFGs of $N$ facial parts. + +Loss Functions. As shown in Fig. 2(a), given an input face image $x$ , the corresponding region mask $m$ and the latent code $w = E_{inv}(x)$ , we train our models by using following loss functions. + +To build a bridge between RFGs and latent space, we first constrain the IGE $E_{img}$ and Gene Decoder $D$ to reconstruct the latent code of the input image. The latent reconstruct loss is defined by + +$$ +\mathcal {L} _ {\text {l a t e n t}} = \left\| \boldsymbol {w} - \hat {\boldsymbol {w}} \right\| _ {2}, \tag {6} +$$ + +where $\hat{\pmb{w}} = D(E_{img}(x,m))$ + +In addition, we use LGE $E_{lat}$ to remove the reliance on + +face annotations. To decouple the latent code into the RFGs obtained from IGE, the LGE is trained by + +$$ +\mathcal {L} _ {\text {g e n e}} = \sum_ {i = 1} ^ {N} \left[ \| \boldsymbol {\mu} _ {i} - \hat {\boldsymbol {\mu}} _ {i} \| _ {2} + \| \boldsymbol {\sigma} _ {i} - \hat {\boldsymbol {\sigma}} _ {i} \| _ {2} \right], \tag {7} +$$ + +where $\hat{\pmb{\mu}}_i$ and $\hat{\pmb{\sigma}}_i$ are the outputs of $E_{img}(x,m)$ for the $i$ -th facial region, and $\pmb{\mu}_i, \pmb{\sigma}_i$ are the outputs of $E_{lat}(\pmb{w})$ for the $i$ -th facial region. + +All modules are trained jointly, and the full loss is defined by + +$$ +\mathcal {L} = \mathcal {L} _ {\text {l a t e n t}} + \lambda \mathcal {L} _ {\text {g e n e}}, \tag {8} +$$ + +where $\lambda$ is a hyperparameter, and we set it to be 1 in our experiments. + +Note that this training process does not rely on kinship data, alleviating the problem of insufficient high-quality annotated kinship datasets. + +# 3.2. Crossover & Mutation + +As parent-offspring resemblance is often reflected in the local area of the face appearance [5, 31]. Thus, after obtaining the region-level facial genes (RFGs), we simulate a crossover process to generate the RFGs of descendants by combining the RFGs of the parents. Additionally, we further build a gene pool to simulate the gene mutation process, which can significantly increase the diversity of synthesized faces. + +Gene Crossover. The gene crossover process aims to generate the gene representation of the descendants with a linear combination of the RFGs of the parents. In particular, given a pair of parental face images, $x_{f}$ and $x_{m}$ , we first apply Image Encoder and LGE to obtain a set of RFGs $g^{f} = [g_{1}^{f},\dots ,g_{N}^{f}]$ and $g^{m} = [g_{1}^{m},\dots ,g_{N}^{m}]$ from $N$ facial parts, respectively. Then, the RFGs of descendants $g^{c} = [g_{1}^{c},\dots ,g_{N}^{c}]$ can be calculated by + +$$ +\boldsymbol {g} _ {i} ^ {c} = \alpha_ {i} \boldsymbol {g} _ {i} ^ {f} + \beta_ {i} \boldsymbol {g} _ {i} ^ {m}, \tag {9} +$$ + +where $\alpha_{i}$ and $\beta_{i}$ are randomly generated weights for each region and $\alpha_{i} + \beta_{i} = 1$ when mutation is not applied. + +Fig. 5 shows four example faces generated using RFGs inherited from their parents based on Eq. 9, when different combinations of $\alpha_{i}$ and $\beta_{i}$ are applied. The bigger value of $\alpha_{i}$ , the more similar to father for the $i$ -th facial region. The figure shows that our approach can precisely and independently control the similarity of each facial region with their parents, by setting different values of $\alpha_{i}$ and $\beta_{i}$ . + +Gene Mutation. In genetics, the gene pool is the set of all genes of all individuals in a population [30]. For humans, all races share the same gene pool. We introduce the gene pool concept to better simulate the genetic variation in facial appearances. We define the gene pool $\mathcal{P}$ as the sets formed by grouping a large number of RFGs, i.e. $\mathcal{P} = \{\pmb{g}^1,\dots ,\pmb{g}^p,\dots ,\pmb{g}^P\}$ , where $\pmb{g}^{p} = [g_{1}^{p},\dots ,g_{N}^{p}]$ . As shown in the supplementary materials, we divide the gene + +![](images/5fbabb2e7520161118d4c57c048608d9f45b82dda96ae8645c81d6a77e5961ff.jpg) +Figure 5. Example of different weights used in crossover of RFGs. While the redder regions look more similar to the father, and the bluer regions look more similar to the mother. + +pool into nine groups by age, two by gender, and seven by race. Note that the RFG in the gene pool can independently control different facial traits. And its phenotype is related to the grouping condition. + +Given a pair of parental face images and information about generated descendants, such as age, gender, and race. We first query the gene pool to extract a subset $\mathcal{P}_c\subseteq \mathcal{P}$ that satisfies the descendants' information, which will be used to simulate gene mutations for better diversity in the descendants. Then we use Image Encoder and LGE to extract RFGs $\pmb{g}^{f} = [g_{1}^{f},\dots ,g_{N}^{f}]$ and $\pmb{g}^{m} = [g_{1}^{m},\dots ,g_{N}^{m}]$ of parents. For mutation, we randomly select $\eta$ percent of facial regions whose RFGs are sampled from the gene pool. Given a $\pmb{g}^p$ sampled from $\mathcal{P}_c$ , a one-hot vector $t = \{t_i\in \{0,1\} ,i = 1,2,\dots ,N\}$ is randomly generated to denote whether the $\pmb{g}_i^c$ of descendant is fully copied from $\pmb{g}_i^p (t_i = 1)$ , or not $(t_i = 0)$ . When $\pmb{g}_i^c$ is inherited from parents $(t_i = 0)$ , we also introduce mutation by introducing $\pmb{g}^p$ into the crossover process defined in Eq. 9. Finally, we can extend Eq. 9 as follows to obtain the RFGs of the descendants $\pmb{g}^c = [\pmb {g}_c^1,\dots ,\pmb {g}_c^N ]$ + +$$ +\boldsymbol {g} _ {i} ^ {c} = \left\{ \begin{array}{l l} \boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 1 \\ \alpha_ {i} \boldsymbol {g} _ {i} ^ {f} + \beta_ {i} \boldsymbol {g} _ {i} ^ {m} + \gamma \boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 0 \end{array} , \right. \tag {10} +$$ + +where $\gamma$ is the intensity of mutation, $\alpha_{i}$ and $\beta_{i}$ are randomly generated weights and $\alpha_{i} + \beta_{i} = 1 - \gamma$ , $g_{i}^{p} = S(\mathcal{P}_{c})$ and $S(\cdot)$ is the random sampling operator. Once the RFG of descendant $g^{c}$ is generated, the learned Gene Decoder is applied to map $g^{c}$ to the latent space of StyleGAN2, i.e. $\boldsymbol{w}_{c} = D(\boldsymbol{g}^{c})$ , $\boldsymbol{w}_{c} \in \mathbb{R}^{18 \times 512}$ , which is further processed for face synthesis. + +While the latent code in $\mathcal{W}^+$ latent space of StyleGAN2 usually consists of 18 layers, Richardson et al. [34] demonstrate that the first 8 layers of the latent code mainly con + +tribute to the ID information of the synthesized faces, and higher layers mainly control the skin color and microstructure of the synthesized faces. While the appearance of facial regions for descendants is mainly decided by $g^{c}$ , other attributes like skin, hair color, image lighting, and background, can thus be inherited by inclusion of the latent codes of parents. Therefore, we further fuse the latent codes of parents $\boldsymbol{w}_{f}$ and $\boldsymbol{w}_{m}$ . We keep the first $l$ layer of $\boldsymbol{w}_{c}$ , and the higher layers are fused from $\boldsymbol{w}_{f}$ and $\boldsymbol{w}_{m}$ . The final latent code $\boldsymbol{w}_{c} = [\boldsymbol{w}_{c}^{1},\dots ,\boldsymbol{w}_{c}^{18}]$ of the descendant can be obtained by + +$$ +\boldsymbol {w} _ {c} ^ {i} = \frac {1}{2} \left(\boldsymbol {w} _ {f} ^ {i} + \boldsymbol {w} _ {m} ^ {i}\right), \quad i \in \{l + 1, \dots , 1 8 \}, \tag {11} +$$ + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. Our training set consists of three parts, all images of CelebAHQ [21], 50,000 faces sampled from MS-Celeb-1M [18], and 10,000 faces generated by StyleGAN2, to make the model adaptive to different image qualities. CelebA-HQ is a high-quality face dataset consisting of 30,000 aligned face images with $1024 \times 1024$ resolution. The MS-Celeb-1M is a large-scale face recognition dataset containing about 10 million images, including many noisy face images. We evaluate our model on the FIW [35], TSKinFace [32], and FF-Database [48] datasets. FIW dataset contains the faces of 1,000 families with 1,997 sets of father-mother-child relations. The TSKinFace and FF-Database datasets provide 1,015 and 3,744 sets of father-mother-child kinship face images, respectively. We align and crop the images in these two datasets to $256 \times 256$ . Our Gene Pool is built based on the FFHQ [23] dataset, which consists of 70,000 high-quality face images, and has a better diversity among human races and ages. + +Baselines. We compare our method against the state-of-art kinship face synthesis baselines, i.e. DNA-Net [14], ChildGAN [7], ChildPredictor [48], and StyleDNA [27]. Since the source codes of StyleDNA and ChildPredictor are available, we compare our approach with them through all of the evaluations. However, due to the unavailability of the source codes, we can only visually compare the visual quality of the generated faces available in the papers of DNA-Net and ChildGAN. + +Training Details. We use AdamW [28] optimizer with batch size of 32. The initialized learning rate is 0.001, which is divided by 2 every 10 epochs, and we stop the training at the 30th epoch. + +# 4.2. Qualitative evaluation + +Disentanglement of RFGs. Fig. 6 shows two examples of facial region editing using our proposed RFG, where the nose, eyes, jaw and lips of source faces shown in the first + +column, are replaced sequentially by the corresponding regions of reference faces shown in the first row. One can observe from the examples that, when a certain region is edited, all other facial regions are kept intact, which clearly shows the disentanglement capability of our RFG. + +![](images/64dd6f01136f79727a0bee14bc3c13014445cf72ed33d2982c92a1936d28a084.jpg) +Figure 6. Disentangled editing of facial regions. + +Effectiveness of Gene Pool. Based on the proposed idea of RFG, our approach further introduces a gene pool to increase the variations among the generated faces of descendants. Given a couple of parents shown in the first row, Fig. 7 shows the faces of five children generated by our approach with/without the involvement of gene pool, together with that synthesized by StyleDNA and ChildPredictor. While the faces generated by StyleDNA and ChildPredictor (4th and 5th rows) all look very similar, that synthesized by our approach are much more diverse and the introduction of gene pool (1st row) can further increase variations among children. + +![](images/96f0f3e296a85428c5ad729aa58a58ca200d4f642ca56e531efff9add0b86872.jpg) +Figure 7. Visual comparison of diversity with StyleDNA and ChildPredictor. The first row shows the father and mother. The next two rows are generated by our method with and without Gene Pool. The last two rows are generated by StyleDNA [27] and ChildPredictor [48], respectively. + +![](images/75637e407a67186632e1bf62e521eeb75f1c1afc3c63b8723c1e5bd7853f6ae1.jpg) +Father +Mother +Real Child +Ours +StyleDN +A ChildPre +Victor ChildGAN +DNA-Net + +![](images/ac2707b07ecf1ccedd5f827ff9f2560a641eec8068aa0b9c6ec92f0cebe1b8b5.jpg) +Father +Mother +Real 0 +d Ours +StyleDNA ChildPredictor +(b) FF-Database +Figure 8. Comparison of children faces synthesized by StyleGene and baselines. Left most three columns of (a) FIW and (b) FF-Database depict the father, mother, and real children, and right five or three columns depict the faces synthesized by ours, StyleDNA [27], ChildPredictor [48], ChildGAN [7] and DNA-Net [14]. + +Comparison with the state-of-the-art. Fig. 8(a) and (b) show qualitative results on FIW and FF-Database, respectively. We first compare StyleGene with StyleDNA, DNA-Net, ChildGAN, and ChildPredictor on FIW with low-resolution facial images. Note that the results of ChildGAN and DNA-Net are directly taken from their papers. As can be seen, DNA-Net, ChildGAN, and ChildPredictor struggle to yield high-quality facial images. Thanks to a powerful generator, our method can generate high-fidelity faces with hair. Though ChildGAN also uses the StyleGAN generator, its results are significantly affected by the quality of the input. Note that all comparisons have considered the facial information of parents. However, DNA-Net, ChildPredictor, and StyleDNA model facial genetics by learning the implicit mapping between parents and children, without considering regional similarity, which significantly compromised the similarity between the generated faces and their parents. Although ChildGAN considered parents' facial features, the results were too similar to their parents. Thanks to the region-level facial genes and the mutation introduced by gene pool, our results can maintain the local similarity to parents' faces, and at the same time, present reasonable global variations and diversity. When experimenting with faces of different races as shown in Fig. 8(b), our approach can still well preserve the race similarity of children with their parents. + +# 4.3. Quantitative evaluation + +Kinship Verification. We now use kinship verification accuracy to assess whether a genetic relationship exists between the synthetic descendant and parents. Higher accuracy means more realistic synthetic descendants. We use ArcFace pre-trained on the MS-Celeb-1M dataset as the backbone to create a one-versus-one kinship classifier [42], which is fine-tuned on the FF-Database training set. Then + +we performed a cross-database evaluation. + +Specifically, we randomly sample 100 families with one kid from the FF-database test set, FIW, and TSKinFace datasets, respectively, and different methods randomly generate 40 children for each family. We randomly sample 8,000 positive and 8,000 negative pairs for each method from each dataset and calculate in Table 2 the kinship verification accuracy. One can observe that our method achieves substantially higher accuracy than StyleDNA and ChildPredictor, i.e. as high as $81.74\%$ , $80.38\%$ , and $62.29\%$ accuracies are achieved on TSKinFace, FF-Database, and FIW datasets, respectively. The results suggest that the kinship relations of faces generated by our approach are very close to the real children. + +Table 2. Kinship verification accuracy $(\%)$ on the TSKinFace [32], FF-Database [48], and FIW [35] dataset. + +
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA [27]53.1555.1149.47
ChildPredictor [48]58.2459.6251.81
StyleGene (Ours)81.7480.3862.29
+ +Diversity Evaluation. We used the LPIPS [44] metric to measure the diversity of synthetic descendants, which calculates the L1 distance between pairs of image features extracted by AlexNet [26] pre-trained on the ImageNet [10] dataset. We use the same test data of kinship verification. First, we calculated the distance among the 40 synthesized descendants for each family. Then we take the average of 100 families as the LPIPS score and list them in Table 3. As shown in the table, our method achieves the highest LPIPS across all three datasets, i.e. 0.3270, 0.3418, and 0.3279 LPIPS on TSKinFace, FF-Database, and FIW datasets are achieved, which is significantly higher than that of StyleDNA and ChildPredictor. + +Distribution of Synthesized Children. We are now try- + +Table 3. Quantitative comparison of the diversity of the generated descendants. * means we cropped the face i.e. no hair. + +
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA* [27]0.07560.07630.0736
ChildPredictor [48]0.16970.17230.1750
StyleGene (Ours)*0.17480.17350.1740
StyleDNA [27]0.15590.15420.1573
StyleGene (Ours)0.32700.34180.3279
+ +ing to model the distribution of the children synthesized by different approaches, and compare them with that of real children. Specifically, we randomly sample 50 families from the TSKinFace dataset, each consisting of a father, mother, son, and daughter. Based on the given parents, we also applied different synthesis methods to generate two children, i.e. a son and a daughter, for each family. After facial features are extracted using ArcFace [11] and reduced to one dimension using t-SNE [41], Kernel Density Estimation (KDE) [37] is further applied to estimate the features' probability density function. Fig. 9 shows the distribution of the 100 real children (blue) and that of children generated by different approaches. As can be seen, the distribution of faces generated by our approach (red) overlaps most with the real children (blue). In contrast, the distributions of faces synthesized by StyleDNA (black) and ChildPredictor (yellow) are located on both sides of the real ones. The distribution for StyleDNA even looks like a mixture of two Gaussian distributions, which is significantly different from that of real children. + +![](images/f488ad36f0d2931814796176ee795e854e0fa8ea59acee6f85d2994f8d580d10.jpg) +Figure 9. Distribution of real children (blue) and children generated using StyleDNA (black), ChildPredictor (yellow), and ours (red). Best viewed in color. + +User Study. To further demonstrate the effectiveness of the proposed method, we conducted a user study involving 59 participants. The faces of children generated by our StyleGene, StyleDNA and ChildPredictor based on 20 parents, are present to each of the 59 participants, who are asked to rank the quality of the faces, in terms of realness and the similarity with given parents. In total $1,180(59 \times 20)$ votes are received for each of the three approaches and the average rankings are listed in Table 4. As shown in the table, our StyleGene receives the highest rank among the com + +pared approaches, which suggests that the faces of children synthesized by our approach are the best, in terms of both quality and similarity with parents. + +Table 4. The rank of different approaches in user study. + +
ChildPredictorStyleDNAStyleGene (Ours)
Avg. rank2.462.221.32
+ +# 4.4. Ablation study + +In this section, we perform ablation studies on the effectiveness of LGE, IGE and GP, in terms of kinship verification accuracy (ACC) and diversity (LPIPS) using FF-database with the same configuration as in Section 4.3. As shown in Table 5, the integration of IGE (2nd row) can improve both ACC and LPIPS of the synthesized descendant. As expected, the use of GP (3rd row) increases the diversity of synthesized faces. When both IGE and GP are integrated with LGE (4th row), the diversity can be significantly increased to 0.1735, with a comparable ACC $(80.38\%)$ . + +Table 5. Ablation study on FF-Database. + +
LGEIGEGPACC (%)LPIPS
1✓XX79.190.0625
2✓✓X79.660.0646
3✓X✓80.210.0839
4✓✓✓80.380.1735
+ +# 4.5. Parameter sensitivity + +We test different values of $\eta$ , $\gamma$ , and $l$ to see how the performance of synthesis varies. Due to the page limit, details about the results of different parameters are presented in supplementary. In our experiments, we choose $\eta = 40\%$ , $\gamma = 0.47$ , and $l = 8$ to achieve a balance between the diversity and fidelity of the generated faces. + +# 5. Conclusion + +In this paper, we have proposed StyleGene, to extract RFGs (Region-level Facial Genes) for kinship face synthesis. While crossover and mutation of RFGs are proposed to model the facial genetic process, Gene Pool is further designed to increase the diversity among generated faces. Quantitative, qualitative, and subjective experimental results show that the realness, similarity with parents, and diversity of kinship faces generated by our approach are much better than existing state-of-the-art methods. + +Acknowledgement. This work was jointly supported by National Natural Science Foundation of China (62206180 and 82261138629), Guangdong Basic and Applied Basic Research Foundation (2023A1515010688 and 2022A1515011018), and Shenzhen Municipal Science and Technology Innovation Council (JCYJ20220531101412030). + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2 +[3] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2 +[4] Peter Claes, Jasmien Roosenboom, Julie D White, Tomek Swigut, Dzemila Sero, Jiarui Li, Myoung Keun Lee, Arslan Zaidi, Brooke C Mattern, Corey Liebowitz, et al. Genome-wide mapping of global-to-local genetic effects on human facial shape. Nature genetics, 50(3):414-423, 2018. 2 +[5] Joanne B Cole, Mange Manyama, Jacinda R Larson, Denise K Liberton, Tracey M Ferrara, Sheri L Riccardi, Mao Li, Washington Mio, Ophir D Klein, Stephanie A Santorico, et al. Human facial shape and size heritability and genetic correlations. Genetics, 205(2):967-978, 2017. 2, 5 +[6] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020. 2 +[7] Xiao Cui, Wengang Zhou, Yang Hu, Weilun Wang, and Houqiang Li. Heredity-aware child face image generation with latent space disentanglement. arXiv preprint arXiv:2108.11080, 2021. 2, 3, 6, 7 +[8] Maria F Dal Martello and Laurence T Maloney. Lateralization of kin recognition signals in the human face. Journal of Vision, 10(8):9-9, 2010. 2 +[9] Afshin Dehghan, Enrique G Ortiz, Ruben Villegas, and Mubarak Shah. Who do i look like? determining parent-offspring resemblance via gated autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1757-1764, 2014. 2 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 8 +[12] Itir Önal Ertugrul and Hamdi Dibeklioglu. What will your future child look like? modeling and synthesis of hereditary patterns of facial dynamics. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 33-40. IEEE, 2017. 2 +[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip + +guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2 +[14] Pengyu Gao, Joseph Robinson, Jiaxuan Zhu, Chao Xia, MIng Shao, and Siyu Xia. Dna-net: Age and gender aware kin face synthesizer. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2, 6, 7 +[15] Pengyu Gao, Siyu Xia, Joseph Robinson, Junkang Zhang, Chao Xia, Ming Shao, and Yun Fu. What will your child look like? dna-net: Age and gender aware kin face synthesizer. arXiv preprint arXiv:1911.07014, 2019. 2 +[16] Fady S Ghatas and ElSayed E Hemayed. Gankin: generating kin faces using disentangled gan. SN Applied Sciences, 2(2):1-10, 2020. 2 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[18] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In European conference on computer vision, pages 87–102. Springer, 2016. 6 +[19] Hanne Hoskens, Dongjing Liu, Sahin Naqvi, Myoung Keun Lee, Ryan J Eller, Karlijne Indencleef, Julie D White, Jiarui Li, Maarten HD Larmouseau, Greet Hens, et al. 3d facial phenotyping by biometric sibling matching used in contemporary genomic methodologies. PLoS genetics, 17(5):e1009528, 2021. 2 +[20] Ali Jahanian, Lucy Chai, and Phillip Isola. "On the" steerability" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 2 +[21] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2, 6 +[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 6 +[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7 +[27] Che-Hsien Lin, Hung-Chun Chen, Li-Chen Cheng, Shu-Chuan Hsu, Jun-Cheng Chen, and Chih-Yu Wang. *Styledna: A high-fidelity age and gender aware kinship face synthesizer*. In 2021 16th IEEE International Conference on Auto + +matic Face and Gesture Recognition (FG 2021), pages 1-8. IEEE, 2021. 2,3,6,7,8 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[29] Savas Ozkan and Akin Ozkan. Kinshipgan: Synthesizing of kinship faces from family photos by regularizing a deep face network. In 2018 25th IEEE international conference on image processing (ICIP), pages 2142-2146. IEEE, 2018. 2 +[30] Svante Paäbo. The diverse origins of the human gene pool. Nature Reviews Genetics, 16(6):313-314, 2015. 5 +[31] Shouneng Peng, Jingze Tan, Sile Hu, Hang Zhou, Jing Guo, Li Jin, and Kun Tang. Detecting genetic association of common human facial morphological variation using high density 3d image registration. PLoS computational biology, 9(12):e1003375, 2013. 5 +[32] Xiaoqian Qin, Xiaoyang Tan, and Songcan Chen. Tri-subject kinship verification: Understanding the core of a family. IEEE Transactions on Multimedia, 17(10):1855-1867, 2015. 2, 6, 7 +[33] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2 +[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2, 5 +[35] Joseph P Robinson, Ming Shao, Yue Wu, and Yun Fu. Families in the wild (fw) large-scale kinship image database and benchmarks. In Proceedings of the 24th ACM international conference on Multimedia, pages 242-246, 2016. 6, 7 +[36] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2 +[37] Bernard W Silverman. Density estimation for statistics and data analysis. Routledge, 2018. 8 +[38] Raunak Sinha, Mayank Vatsa, and Richa Singh. Familygan: Generating kin face images using generative adversarial networks. In European conference on computer vision, pages 297-311. Springer, 2020. 2 +[39] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. 4 +[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4 +[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8 + +[42] Jun Yu, Mengyan Li, Xinlong Hao, and Guochen Xie. Deep fusion siamese network for automatic kinship verification. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 892-899. IEEE, 2020. 2, 7 +[43] Ishtiak Zaman and David Crandall. Genetic-gan: Synthesizing images between two domains by genetic crossover. In European Conference on Computer Vision, pages 312-326. Springer, 2020. 2 +[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[45] Yong Zhang, Le Li, Zhilei Liu, Baoyuan Wu, Yanbo Fan, and Zhifeng Li. Controllable descendant face synthesis. arXiv preprint arXiv:2002.11376, 2020. 2 +[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3 +[47] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2 +[48] Yuzhi Zhao, Lai-Man Po, Xuehui Wang, Qiong Yan, Wei Shen, Yujia Zhang, Wei Liu, Chun-Kit Wong, Chiu-Sing Pang, Weifeng Ou, et al. Childpredictor: A child face prediction framework with disentangled learning. IEEE Transactions on Multimedia, 2022. 2, 6, 7, 8 +[49] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2 +[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/images.zip b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1d37beeddd0e2da532993d25ee029961c0320de0 --- /dev/null +++ b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e15089dfa690a0fb0fecc5acb87c11e5c57f379977e8e43285ebbedab0b79bb3 +size 691130 diff --git a/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/layout.json b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..314ab1d8b706b10ee402c48c7749146aaaf9b8af --- /dev/null +++ b/2023/StyleGene_ Crossover and Mutation of Region-Level Facial Genes for Kinship Face Synthesis/layout.json @@ -0,0 +1,10977 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "text", + "content": "StyleGene: Crossover and Mutation of Region-level Facial Genes for Kinship Face Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "spans": [ + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "text", + "content": "Hao Li" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "text", + "content": ", Xianxu Hou" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "inline_equation", + "content": "^{2,5}" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "text", + "content": ", Zepeng Huang" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "text", + "content": ", Linlin Shen" + }, + { + "bbox": [ + 161, + 160, + 461, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 175, + 565, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 175, + 565, + 189 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 565, + 189 + ], + "type": "text", + "content": "1Computer Vision Institute, College of Computer Science and Software Engineering, Shenzhen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 190, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 190, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 65, + 190, + 553, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 190, + 553, + 204 + ], + "type": "text", + "content": "National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 204, + 477, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 204, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 140, + 204, + 477, + 217 + ], + "type": "text", + "content": "3Shenzhen Institute of Artificial Intelligence and Robotics for Society" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 217, + 522, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 217, + 522, + 232 + ], + "spans": [ + { + "bbox": [ + 96, + 217, + 522, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 217, + 522, + 232 + ], + "type": "text", + "content": "Guangdong Key Laboratory of Intelligent Information Processing, Shenzhen University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 232, + 498, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 232, + 498, + 246 + ], + "spans": [ + { + "bbox": [ + 121, + 232, + 498, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 121, + 232, + 498, + 246 + ], + "type": "text", + "content": "School of AI and Advanced Computing, Xi'an Jiaotong-Liverpool University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 247, + 531, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 247, + 531, + 258 + ], + "spans": [ + { + "bbox": [ + 83, + 247, + 531, + 258 + ], + "type": "text", + "content": "{haoli2021, huangzepeng2021}@email.szu.edu.cn, hxianxu@gmail.com, llshen@szu.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 80, + 277, + 154, + 386 + ], + "blocks": [ + { + "bbox": [ + 80, + 277, + 154, + 386 + ], + "lines": [ + { + "bbox": [ + 80, + 277, + 154, + 386 + ], + "spans": [ + { + "bbox": [ + 80, + 277, + 154, + 386 + ], + "type": "image", + "image_path": "1998cdf32cf9536c7e74081fe95dd73bc74e42df801218b6b2bea9a271e98947.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 156, + 277, + 229, + 316 + ], + "blocks": [ + { + "bbox": [ + 156, + 277, + 229, + 316 + ], + "lines": [ + { + "bbox": [ + 156, + 277, + 229, + 316 + ], + "spans": [ + { + "bbox": [ + 156, + 277, + 229, + 316 + ], + "type": "image", + "image_path": "1a7df84691f49bdb9b63d832f714d4d1b92ea94c942f0b07725b369ac10c9d00.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 156, + 316, + 229, + 386 + ], + "blocks": [ + { + "bbox": [ + 156, + 316, + 229, + 386 + ], + "lines": [ + { + "bbox": [ + 156, + 316, + 229, + 386 + ], + "spans": [ + { + "bbox": [ + 156, + 316, + 229, + 386 + ], + "type": "image", + "image_path": "f31204d2347063184e1d1459fe49d2eccd35eee75bd54927bbc8b477a174888b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 231, + 277, + 304, + 386 + ], + "blocks": [ + { + "bbox": [ + 231, + 277, + 304, + 386 + ], + "lines": [ + { + "bbox": [ + 231, + 277, + 304, + 386 + ], + "spans": [ + { + "bbox": [ + 231, + 277, + 304, + 386 + ], + "type": "image", + "image_path": "17e0862662d2c685105cf6163b174d61d9c8c3c826b98aacae3743d7078b4016.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 305, + 277, + 380, + 386 + ], + "blocks": [ + { + "bbox": [ + 305, + 277, + 380, + 386 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 380, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 380, + 386 + ], + "type": "image", + "image_path": "10696e267e50c0c4a6f3b5dd3896f8289d409d9761cb5ff199f79ca94789a8af.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 381, + 277, + 455, + 386 + ], + "blocks": [ + { + "bbox": [ + 381, + 277, + 455, + 386 + ], + "lines": [ + { + "bbox": [ + 381, + 277, + 455, + 386 + ], + "spans": [ + { + "bbox": [ + 381, + 277, + 455, + 386 + ], + "type": "image", + "image_path": "d38b3d919d04a9e26884ac9f9bd669931c3efdf65a81b77e3f4f34f5e5702705.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 457, + 277, + 531, + 386 + ], + "blocks": [ + { + "bbox": [ + 457, + 277, + 531, + 386 + ], + "lines": [ + { + "bbox": [ + 457, + 277, + 531, + 386 + ], + "spans": [ + { + "bbox": [ + 457, + 277, + 531, + 386 + ], + "type": "image", + "image_path": "541c075681020995d83fa1ddcb7b3a9dc3131b633d251594629047f8d417fbfc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 118, + 396, + 192, + 467 + ], + "blocks": [ + { + "bbox": [ + 118, + 396, + 192, + 467 + ], + "lines": [ + { + "bbox": [ + 118, + 396, + 192, + 467 + ], + "spans": [ + { + "bbox": [ + 118, + 396, + 192, + 467 + ], + "type": "image", + "image_path": "3798f1f0b9786875d1362048c91fa574a90ec491b8bec8353046565558faedf5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 472, + 546, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 546, + 495 + ], + "type": "text", + "content": "Figure 1. Our StyleGene method synthesizes kinship faces with resemblance to parents, exhibiting diversity and reasonable variations. The first row is the input grandparents, and the second and third rows are their descendants generated by our method." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 194, + 396, + 266, + 467 + ], + "blocks": [ + { + "bbox": [ + 194, + 396, + 266, + 467 + ], + "lines": [ + { + "bbox": [ + 194, + 396, + 266, + 467 + ], + "spans": [ + { + "bbox": [ + 194, + 396, + 266, + 467 + ], + "type": "image", + "image_path": "2e81a98ba803287af67189da48ee6173fffe5c5144b06556c7760bd15e4742ab.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 269, + 396, + 342, + 467 + ], + "blocks": [ + { + "bbox": [ + 269, + 396, + 342, + 467 + ], + "lines": [ + { + "bbox": [ + 269, + 396, + 342, + 467 + ], + "spans": [ + { + "bbox": [ + 269, + 396, + 342, + 467 + ], + "type": "image", + "image_path": "176d95cd0ac1500b717cc8261922de7fa8b77f9758c897614d7eef8fd181790b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 343, + 396, + 416, + 467 + ], + "blocks": [ + { + "bbox": [ + 343, + 396, + 416, + 467 + ], + "lines": [ + { + "bbox": [ + 343, + 396, + 416, + 467 + ], + "spans": [ + { + "bbox": [ + 343, + 396, + 416, + 467 + ], + "type": "image", + "image_path": "91501a59c42703cd6a68d80b266ca003d0d04ef4ff667a580a2f3adc9e56cc72.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 419, + 396, + 492, + 467 + ], + "blocks": [ + { + "bbox": [ + 419, + 396, + 492, + 467 + ], + "lines": [ + { + "bbox": [ + 419, + 396, + 492, + 467 + ], + "spans": [ + { + "bbox": [ + 419, + 396, + 492, + 467 + ], + "type": "image", + "image_path": "2a8d890bf2051b30ae387258cd847eab50828c718365ac1e5c23cfbef3c42d3a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 143, + 503, + 192, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 503, + 192, + 516 + ], + "spans": [ + { + "bbox": [ + 143, + 503, + 192, + 516 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 529, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 289, + 696 + ], + "type": "text", + "content": "High-fidelity kinship face synthesis has many potential applications, such as kinship verification, missing child identification, and social media analysis. However, it is challenging to synthesize high-quality descendant faces with genetic relations due to the lack of large-scale, high-quality annotated kinship data. This paper proposes RFG (Region-level Facial Gene) extraction framework to address this issue. We propose to use IGE (Image-based Gene Encoder), LGE (Latent-based Gene Encoder) and Gene Decoder to learn the RFGs of a given face image, and the relationships between RFGs and the latent space of StyleGAN2. As cycle-like losses are designed to measure the " + }, + { + "bbox": [ + 46, + 529, + 289, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 529, + 289, + 696 + ], + "type": "text", + "content": " distances between the output of Gene Decoder and image encoder, and that between the output of LGE and IGE, only" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 505, + 548, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 548, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 548, + 636 + ], + "type": "text", + "content": "face images are required to train our framework, i.e. no paired kinship face data is required. Based upon the proposed RFGs, a crossover and mutation module is further designed to inherit the facial parts of parents. A Gene Pool has also been used to introduce the variations into the mutation of RFGs. The diversity of the faces of descendants can thus be significantly increased. Qualitative, quantitative, and subjective experiments on FIW, TSKinFace, and FF-Databases clearly show that the quality and diversity of kinship faces generated by our approach are much better than the existing state-of-the-art methods." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 657, + 387, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 387, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 387, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Humans can identify kinship through photographs based on the resemblance between parents and children. Many works have investigated this intrinsic relation in the fields" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 276, + 75, + 328 + ], + "type": "aside_text", + "angle": 90, + "lines": [ + { + "bbox": [ + 63, + 276, + 75, + 328 + ], + "spans": [ + { + "bbox": [ + 63, + 276, + 75, + 328 + ], + "type": "text", + "content": "Grandparents" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 65, + 340, + 75, + 370 + ], + "type": "aside_text", + "angle": 90, + "lines": [ + { + "bbox": [ + 65, + 340, + 75, + 370 + ], + "spans": [ + { + "bbox": [ + 65, + 340, + 75, + 370 + ], + "type": "text", + "content": "Parents" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20960" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "of kinship verification [9, 32, 42] and genetics [4, 5, 8, 19]. With the popularity of face synthesis and editing technology in recent years, high-fidelity kinship face synthesis has also attracted much attention. This task, aiming to synthesize the faces of descendants based on the appearance of the parents, has many potential applications, such as finding long-lost children, crime investigations, kinship verification, and multimedia social applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 432 + ], + "type": "text", + "content": "In recent years, many efforts have been made to make use of generative models [7, 12, 15, 16, 27, 29, 38, 43, 45, 48] for kinship face synthesis. These works can be categorized into two paradigms: one-stage and two-stage. The one-stage paradigm [12, 29, 38, 45] treats this problem as an image-to-image translation task and trains a one-to-one kinship face generator with paired data. However, these approaches can only produce low-resolution images and the resultant images can be blurry and lack diversity. Further, it would be quite difficult to obtain annotated kinship data. By contrast, the two-stage paradigm [7, 15, 16, 27, 43, 48] first extracts the genetic representation and assembles them into children's representation based on the parents' faces. To obtain genetic representation, existing methods try to learn the inheritance and variation of facial appearances by training deep neural networks [14, 27, 48] or via a knowledge rule [7]. However, the learned genetic representation is prone to overfitting due to the lack of high-quality kinship annotated training data, resulting in a lack of diversity in the generated children. In addition, these methods cannot provide fine-grained attributes representation, and thus the generated facial attributes lack interpretability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 433, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 659 + ], + "type": "text", + "content": "In this paper, the facial genetic process is abstracted as the exchange and mutation of the parents' facial parts. We propose an Image-based Gene Encoder (IGE) to construct an independent representation for each facial part, called a Region-level Facial Gene (RFG), which is used to control the synthesis of facial regions. We further simulate the crossover and mutation process to assemble the RFGs of descendants by using those of the parents, and our proposed Gene Pool used in the mutation process can significantly increase the diversity of the generated descendants. We use the pre-trained StyleGAN2 [24] as the generator to synthesize high-fidelity faces. To achieve this, we use a Gene Decoder to map RFGs to the " + }, + { + "bbox": [ + 46, + 433, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 433, + 287, + 659 + ], + "type": "text", + "content": " space of StyleGAN2. Since IGE requires a facial parsing mask to generate the RFG, we additionally train a Latent-based Gene Encoder (LGE) to directly map the latent code of StyleGAN2 to RFGs. Thus, facial parsing mask is not required for the RFG extraction in the inference stage. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 287, + 714 + ], + "type": "text", + "content": "- We propose StyleGene to synthesize high-fidelity kinship faces with controllable facial genetic regions, via modeling the facial genetic relations based on the proposed region-level facial genes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 72, + 545, + 194 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 133 + ], + "type": "text", + "content": "- A novel genetic strategy is further introduced by simulating the crossover and mutation process to generate the RFGs of descendants. We introduce a Gene Pool into the mutation process to significantly increase the diversity of the kinship face." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 135, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 135, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 135, + 545, + 194 + ], + "type": "text", + "content": "- We validate the effectiveness of our approach on several benchmarks, demonstrating the superiority of our StyleGene framework over other state-of-the-art methods, in terms of the quality and diversity of the generated kinship faces." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 199, + 392, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 392, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 392, + 211 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 219, + 530, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 219, + 530, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 530, + 232 + ], + "type": "text", + "content": "2.1. Manipulation in Latent Space of StyleGAN" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": "Generative Adversarial Networks (GANs) [17] have been widely used in face generation [21, 33, 47] and editing [3, 13, 50]. In particular, the StyleGAN [22-24] has attracted much attention due to its ability to synthesize high-fidelity images. The StyleGAN generator first maps a latent code " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " drawn from a normal distribution to an intermediate latent code " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "w \\in \\mathcal{W}" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " by a mapping function to control image generation. Previous studies [20, 23, 36] have demonstrated that " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " space has learned facial attribute semantics and different layers of latent code control different levels of image attributes. Based on this finding, many works [6, 49] tried to invert the real image to " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " space for semantic face editing. Some recent works [1, 2, 34] show that the " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " space extended from the " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " space has lower reconstruction errors. In this paper, we build a link between our proposed RFGs and the " + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 237, + 545, + 452 + ], + "type": "text", + "content": " space of StyleGAN2 [24], based on which we can achieve fine-grained control over the synthesis of facial regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 459, + 437, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 437, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 437, + 472 + ], + "type": "text", + "content": "2.2. Kinship Face Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 477, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 586 + ], + "type": "text", + "content": "Kinship face synthesis aims to synthesize the face images of descendants given the images of parents. The challenge is to learn facial genetic relations with limited kinship data. Table 1 summarizes the key differences between ours and existing methods. Early works [14,29,48] design the image level mapping from parents to children via supervised learning. However, limited by the quality and scale of training data, the quality of images produced by these methods is usually low and prone to overfitting. Recent works" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 616, + 545, + 721 + ], + "blocks": [ + { + "bbox": [ + 305, + 590, + 545, + 612 + ], + "lines": [ + { + "bbox": [ + 305, + 590, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 545, + 612 + ], + "type": "text", + "content": "Table 1. Comparison between StyleGene and existing kinship face synthesis methods. " + }, + { + "bbox": [ + 305, + 590, + 545, + 612 + ], + "type": "inline_equation", + "content": "N_{r}" + }, + { + "bbox": [ + 305, + 590, + 545, + 612 + ], + "type": "text", + "content": " denotes the number of controllable regions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 616, + 545, + 721 + ], + "lines": [ + { + "bbox": [ + 309, + 616, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 309, + 616, + 545, + 721 + ], + "type": "table", + "html": "
MethodsStagesKinship annotationImage resolutionDiversity ControllingNr
DNA-Net [14]two✓128Noise-
KinshipGAN [29]one✓128Noise-
ChildPredictor [48]two✓128Noise-
CDFS [45]oneX256Noise5
StyleDNA [27]two✓1024Noise-
ChildGAN [7]twoX1024Interpolation5
StyleGene (Ours)twoX1024Gene Pool34
", + "image_path": "619564a493f248d4a0957da529c873f0641ac47d1d5c10a6f7ffc74b0fa8cf55.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20961" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 211, + 240 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 211, + 240 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 211, + 240 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 211, + 240 + ], + "type": "image", + "image_path": "dc7f40045aaf7efd468d9d4d75083ccb3579bd24f1bb559f06b95902e917bbdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 243, + 161, + 254 + ], + "lines": [ + { + "bbox": [ + 102, + 243, + 161, + 254 + ], + "spans": [ + { + "bbox": [ + 102, + 243, + 161, + 254 + ], + "type": "text", + "content": "(a) Training Stage" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": "Figure 2. The overall framework of our method. The image-based gene encoder " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " learns to encode an independent RFG representation for each facial region in the training stage. The gene decoder " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " maps RFGs " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " space of StyleGAN2. For ease of use, our latent-based gene encoder " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "E_{lat}" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " maps latent codes " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " (obtained by an image inversion encoder " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "E_{inv}" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": ") to RFGs. In the inference stage, the RFGs of both parents are first extracted using latent-based gene encoder, and then the RFGs of the descendants are assembled by crossover and mutation module. The RFGs will be mapped back to the " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": " space using gene decoder, and finally the high-fidelity face is generated by a pre-trained StyleGAN2 generator " + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 255, + 545, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 213, + 72, + 539, + 239 + ], + "blocks": [ + { + "bbox": [ + 213, + 72, + 539, + 239 + ], + "lines": [ + { + "bbox": [ + 213, + 72, + 539, + 239 + ], + "spans": [ + { + "bbox": [ + 213, + 72, + 539, + 239 + ], + "type": "image", + "image_path": "a26cf294bb0568f61cab7d458c0929d1593aa00aea2de332e1c2a37f317300e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 243, + 408, + 253 + ], + "lines": [ + { + "bbox": [ + 346, + 243, + 408, + 253 + ], + "spans": [ + { + "bbox": [ + 346, + 243, + 408, + 253 + ], + "type": "text", + "content": "(b) Inference Stage" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "type": "text", + "content": "[7,27] based on StyleGAN can generate high-resolution images by interpolating in the latent space to control the generation of descendants. However, StyleDNA [27] learns the mapping from parents to children in " + }, + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "type": "text", + "content": " space by supervised learning, resulting in poor diversity due to overfitting. In addition, the inherited regions of children are uncontrollable. ChildGAN [7] calculates the direction vectors of different facial regions through annotated landmarks, and controls the genetic regions by interpolation. However, due to attribute entanglement in " + }, + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 335, + 287, + 478 + ], + "type": "text", + "content": " space, finding the disentangled direction vectors for small facial regions is difficult, which limits the diversity of generated descendants." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 538 + ], + "type": "text", + "content": "To address this problem, our proposed StyleGene first learns region-level facial genes (RFGs) to control the synthesis of face regions. Then we further model kinship relations based on RFGs, and leverage the pre-trained StyleGAN2 generator to generate kinship face." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 550, + 102, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 102, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 102, + 561 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "type": "text", + "content": "In this section, we present our approach for high-fidelity kinship face synthesis. Given a pair of parental face images, " + }, + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "type": "inline_equation", + "content": "x_{f}" + }, + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 46, + 570, + 287, + 628 + ], + "type": "text", + "content": ", our goal is to synthesize face images of their descendants. We consider the facial genetic process as the inheritance of facial parts from the father and mother." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "The overview of the proposed framework is shown in Fig. 2. In training, we build models to extract Region-level Facial Genes (RFGs) from StyleGAN2 latent space and then learn to map the disentangled genes representations back to the corresponding latent code. In particular, we use an Image-based Gene Encoder (IGE) to directly extract RFGs from the input face with region annotations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "type": "text", + "content": "Then a Gene Decoder is trained to map the obtained RFGs to StyleGAN2 latent space. In addition, we use a GAN inversion method to embed the same face to the " + }, + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 335, + 545, + 407 + ], + "type": "text", + "content": " latent space of StyleGAN2, and a Latent-based Gene Encoder (LGE) is used to decompose the obtained latent code into the corresponding RFGs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 407, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 456 + ], + "type": "text", + "content": "In inference, we first use LGE to extract the RFGs of parental faces, which are then used to obtain children's RFGs via crossover and mutation process. In addition, we build a Gene Pool to simulate the genetic variations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 465, + 444, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 444, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 444, + 478 + ], + "type": "text", + "content": "3.1. Region-level Facial Gene" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 581 + ], + "type": "text", + "content": "In order to build facial genetic relations, a key step of the proposed framework is to extract disentangled representations of different face regions. We propose to train IGE to directly extract RFGs by using fine-grained annotated facial parts. We then use a Gene Decoder to transform the obtained gene representations into the StyleGAN2 latent space, from which we can reconstruct the original face image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": "Fine-grained Facial Parts Segmentation. We follow DatasetGAN [46] pipeline to generate face images and corresponding segmentation masks. DatasetGAN trains a shallow decoder to achieve semantic segmentation based on the features of StyleGAN by only using a few annotated data. In this work, we adapt their approach to StyleGAN2 to produce pixel-level labels for " + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "inline_equation", + "content": "N = 34" + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": ") facial parts. These fine-grained facial parts are then used to extract the corresponding facial genes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Image-based Gene Encoder. Fig. 4 shows the overview of our IGE " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{img}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": ". The goal of IGE is to learn region-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20962" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 532, + 186 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 532, + 186 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 532, + 186 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 532, + 186 + ], + "type": "image", + "image_path": "e7a1117b91cd3e7a3e27b567ab99d676447c2453473644e38652985b32bdb204.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "type": "text", + "content": "Figure 3. The proposed Latent-based Gene Encoder (a) and Gene Decoder (b) architecture. We mainly use the Mixer layer [39] to deal with the nonlinear transformation between high-dimensional vectors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 223, + 266, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 266, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 266, + 234 + ], + "type": "text", + "content": "level facial gene representation " + }, + { + "bbox": [ + 46, + 223, + 266, + 234 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i" + }, + { + "bbox": [ + 46, + 223, + 266, + 234 + ], + "type": "text", + "content": " for each facial part." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": "Given a face image " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times 512 \\times 512}" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " and the corresponding mask " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "m = \\{m_1, \\dots, m_N\\}" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": ", we first transform " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " to a feature map " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " resolution by a series of convolutional layers " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": ". Then we extract the region feature " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": "-th face part by masking the obtained " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " with the corresponding binary mask " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "r_i = f(x) \\odot m_i" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": " is an element-wise matrix multiplication operator. Inspired by variational autoencoder [25], the posterior distribution for each region " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": ", mapped by a mapping function " + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 46, + 234, + 287, + 366 + ], + "type": "text", + "content": ", is modeled as a multivariate Gaussian distribution. It is defined by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 370, + 287, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 370, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 97, + 370, + 287, + 385 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {g}} _ {i} \\sim q (\\hat {\\boldsymbol {g}} _ {i} | \\boldsymbol {r} _ {i}) = \\mathcal {N} (\\hat {\\boldsymbol {g}} _ {i}; \\hat {\\boldsymbol {\\mu}} _ {i}, \\hat {\\boldsymbol {\\sigma}} _ {i} ^ {2} \\boldsymbol {I}), \\tag {1}", + "image_path": "9f491868e067eebb1ec98d0d0d5b0f40f2ac4c5d211fcd5baca13c5ead0bc928.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\mu}}_i,\\hat{\\pmb{\\sigma}}_i\\in \\mathbb{R}^{18\\times 512}" + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "text", + "content": " are the multi-dimensional output of " + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "inline_equation", + "content": "h_i(r_i)" + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "text", + "content": ", representing the mean and standard deviation in a diagonal matrix form, respectively. According to [25], backpropagation is made differentiable via the reparameterization trick, thus " + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{g}}_i\\in \\mathbb{R}^{18\\times 512}" + }, + { + "bbox": [ + 46, + 388, + 287, + 449 + ], + "type": "text", + "content": " can be sampled by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 453, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 453, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 99, + 453, + 287, + 466 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {g}} _ {i} = \\hat {\\boldsymbol {\\mu}} _ {i} + \\hat {\\boldsymbol {\\sigma}} _ {i} \\odot \\boldsymbol {\\epsilon}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), \\tag {2}", + "image_path": "050bf15e0513e15f8d8ad0af7c25667c4e530aa55f4f5b5b3cb58c07637db3b1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "text", + "content": "Finally, since we use " + }, + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "text", + "content": " facial parts, all the RFG can be denoted by " + }, + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{g}} = [\\hat{\\pmb{g}}_1,\\dots ,\\hat{\\pmb{g}}_N]\\in \\mathbb{R}^{18\\times 512\\times N}" + }, + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "text", + "content": ". We simplify this process denoted as " + }, + { + "bbox": [ + 46, + 471, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{g}} = E_{img}(x,m)" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 521, + 270, + 670 + ], + "blocks": [ + { + "bbox": [ + 62, + 521, + 270, + 670 + ], + "lines": [ + { + "bbox": [ + 62, + 521, + 270, + 670 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 270, + 670 + ], + "type": "image", + "image_path": "33e0bdde9596cca5ece169deeed617eff434d09b6aedb8bdb8514fe6fe57696d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 673, + 261, + 684 + ], + "lines": [ + { + "bbox": [ + 71, + 673, + 261, + 684 + ], + "spans": [ + { + "bbox": [ + 71, + 673, + 261, + 684 + ], + "type": "text", + "content": "Figure 4. The proposed Image-based Gene Encoder." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Latent-based Gene Encoder. As shown above, the proposed IGE needs a face mask to extract the representations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "content": "for each facial region. To eliminate the dependency, we train a LGE " + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "inline_equation", + "content": "E_{lat}(\\cdot)" + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "content": " to directly extract RFGs from StyleGAN2 latent space. As shown in Fig. 3(a), our LGE mainly consists of a fully connected layer and several Mixer Layers [39], and maps the latent code " + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "content": " of a face image " + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "content": " to the mean " + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}_i" + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "content": " and standard deviation " + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 304, + 223, + 545, + 306 + ], + "type": "text", + "content": " for each pre-defined face region. This process can be formulated by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 311, + 545, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 311, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 383, + 311, + 545, + 323 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {\\mu}, \\boldsymbol {\\sigma} \\right\\} = E _ {l a t} (\\boldsymbol {w}), \\tag {3}", + "image_path": "b47726e51c1257a35bc4eecd052b12430244ee6455904a0128ab438539e5debf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} = [\\pmb{\\mu}_1, \\dots, \\pmb{\\mu}_N]" + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\pmb{\\sigma} = [\\pmb{\\sigma}_1, \\dots, \\pmb{\\sigma}_N]" + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": ", and we use a pre-trained GAN inversion model " + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "inline_equation", + "content": "E_{inv}(\\cdot)" + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": " [40] as the image encoder to get the latent code of a face image, i.e. " + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\pmb{w} = E_{inv}(x)" + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": ". Then the region-level gene " + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i" + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 327, + 545, + 388 + ], + "type": "text", + "content": "-th facial region can be sampled by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 392, + 545, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 392, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 358, + 392, + 545, + 406 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {i} = \\boldsymbol {\\mu} _ {i} + \\boldsymbol {\\sigma} _ {i} \\odot \\boldsymbol {\\epsilon}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}). \\tag {4}", + "image_path": "204111935f5cedcea6b30450e87c53adccf101b605e58f6f6303ccd4e00b9e37.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 410, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 434 + ], + "type": "text", + "content": "Finally, the RFGs of " + }, + { + "bbox": [ + 304, + 410, + 545, + 434 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 410, + 545, + 434 + ], + "type": "text", + "content": " facial parts can be denoted by " + }, + { + "bbox": [ + 304, + 410, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{g} = [g_1,\\dots ,g_N]" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": "Gene Decoder. We then use Gene Decoder " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": " to map the region-level facial genes into the StyleGAN2 latent space. Thus, we can reconstruct the face images from the obtained gene representation. Specifically, the Gene Decoder first processes the obtained RFGs with several Mixer Layers [39]. Then all the regional features are concatenated and further processed by a fully connected layer to obtain the latent code " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{w}}\\in \\mathbb{R}^{18\\times 512}" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": ". This process can be formulated by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 399, + 541, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 541, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 399, + 541, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {w}} = D (\\hat {\\boldsymbol {g}}), \\tag {5}", + "image_path": "a8e3a04b5c1c4b46d8b1bfcedde69c076d4a7c60b5308fdc4e05abda7743e010.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{g}} = [\\hat{\\pmb{g}}_1,\\dots ,\\hat{\\pmb{g}}_N]" + }, + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "type": "text", + "content": " is the RFGs of " + }, + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 559, + 527, + 571 + ], + "type": "text", + "content": " facial parts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": "Loss Functions. As shown in Fig. 2(a), given an input face image " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": ", the corresponding region mask " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": " and the latent code " + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "inline_equation", + "content": "w = E_{inv}(x)" + }, + { + "bbox": [ + 304, + 571, + 545, + 619 + ], + "type": "text", + "content": ", we train our models by using following loss functions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "type": "text", + "content": "To build a bridge between RFGs and latent space, we first constrain the IGE " + }, + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "type": "text", + "content": " and Gene Decoder " + }, + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 620, + 545, + 667 + ], + "type": "text", + "content": " to reconstruct the latent code of the input image. The latent reconstruct loss is defined by" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 378, + 672, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 672, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 378, + 672, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l a t e n t}} = \\left\\| \\boldsymbol {w} - \\hat {\\boldsymbol {w}} \\right\\| _ {2}, \\tag {6}", + "image_path": "cda9ca2025051b5a4781c9d63b8b4ce0a9babecbd3f5466db2f9bce8989ee75f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 689, + 424, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 424, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 424, + 701 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 424, + 701 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{w}} = D(E_{img}(x,m))" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In addition, we use LGE " + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{lat}" + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": " to remove the reliance on" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20963" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "face annotations. To decouple the latent code into the RFGs obtained from IGE, the LGE is trained by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 99, + 287, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 99, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 64, + 99, + 287, + 131 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g e n e}} = \\sum_ {i = 1} ^ {N} \\left[ \\| \\boldsymbol {\\mu} _ {i} - \\hat {\\boldsymbol {\\mu}} _ {i} \\| _ {2} + \\| \\boldsymbol {\\sigma} _ {i} - \\hat {\\boldsymbol {\\sigma}} _ {i} \\| _ {2} \\right], \\tag {7}", + "image_path": "19736393dc14541ba1141d7cd8b9bf0a957f47b9b9a0fb28d6c6b3c829b4aaaa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\mu}}_i" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\sigma}}_i" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": " are the outputs of " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "E_{img}(x,m)" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": "-th facial region, and " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i, \\pmb{\\sigma}_i" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": " are the outputs of " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "E_{lat}(\\pmb{w})" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 133, + 287, + 169 + ], + "type": "text", + "content": "-th facial region." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 169, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 169, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 287, + 192 + ], + "type": "text", + "content": "All modules are trained jointly, and the full loss is defined by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 194, + 287, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 194, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 118, + 194, + 287, + 206 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {l a t e n t}} + \\lambda \\mathcal {L} _ {\\text {g e n e}}, \\tag {8}", + "image_path": "172d30bc7d2d1018aee6e362e014e600127b4bb499533f2641d9a95d5f22a253.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "text", + "content": " is a hyperparameter, and we set it to be 1 in our experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 236, + 287, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 272 + ], + "type": "text", + "content": "Note that this training process does not rely on kinship data, alleviating the problem of insufficient high-quality annotated kinship datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 280, + 175, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 175, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 175, + 292 + ], + "type": "text", + "content": "3.2. Crossover & Mutation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 299, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 287, + 392 + ], + "type": "text", + "content": "As parent-offspring resemblance is often reflected in the local area of the face appearance [5, 31]. Thus, after obtaining the region-level facial genes (RFGs), we simulate a crossover process to generate the RFGs of descendants by combining the RFGs of the parents. Additionally, we further build a gene pool to simulate the gene mutation process, which can significantly increase the diversity of synthesized faces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": "Gene Crossover. The gene crossover process aims to generate the gene representation of the descendants with a linear combination of the RFGs of the parents. In particular, given a pair of parental face images, " + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "inline_equation", + "content": "x_{f}" + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": ", we first apply Image Encoder and LGE to obtain a set of RFGs " + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "inline_equation", + "content": "g^{f} = [g_{1}^{f},\\dots ,g_{N}^{f}]" + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "inline_equation", + "content": "g^{m} = [g_{1}^{m},\\dots ,g_{N}^{m}]" + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": " facial parts, respectively. Then, the RFGs of descendants " + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "inline_equation", + "content": "g^{c} = [g_{1}^{c},\\dots ,g_{N}^{c}]" + }, + { + "bbox": [ + 47, + 395, + 287, + 491 + ], + "type": "text", + "content": " can be calculated by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 124, + 493, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 493, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 124, + 493, + 287, + 507 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {i} ^ {c} = \\alpha_ {i} \\boldsymbol {g} _ {i} ^ {f} + \\beta_ {i} \\boldsymbol {g} _ {i} ^ {m}, \\tag {9}", + "image_path": "4a865192d80443c03b2de1f0b87851ea26d48c1f57cb8b81a5a4c5e510ee46d6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "text", + "content": " are randomly generated weights for each region and " + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\alpha_{i} + \\beta_{i} = 1" + }, + { + "bbox": [ + 47, + 510, + 287, + 533 + ], + "type": "text", + "content": " when mutation is not applied." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": "Fig. 5 shows four example faces generated using RFGs inherited from their parents based on Eq. 9, when different combinations of " + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": " are applied. The bigger value of " + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": ", the more similar to father for the " + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": "-th facial region. The figure shows that our approach can precisely and independently control the similarity of each facial region with their parents, by setting different values of " + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 47, + 533, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "content": "Gene Mutation. In genetics, the gene pool is the set of all genes of all individuals in a population [30]. For humans, all races share the same gene pool. We introduce the gene pool concept to better simulate the genetic variation in facial appearances. We define the gene pool " + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "content": " as the sets formed by grouping a large number of RFGs, i.e. " + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb{g}^1,\\dots ,\\pmb{g}^p,\\dots ,\\pmb{g}^P\\}" + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{g}^{p} = [g_{1}^{p},\\dots ,g_{N}^{p}]" + }, + { + "bbox": [ + 47, + 618, + 287, + 713 + ], + "type": "text", + "content": ". As shown in the supplementary materials, we divide the gene" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 531, + 255 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 531, + 255 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 531, + 255 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 531, + 255 + ], + "type": "image", + "image_path": "5fbabb2e7520161118d4c57c048608d9f45b82dda96ae8645c81d6a77e5961ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 257, + 545, + 290 + ], + "lines": [ + { + "bbox": [ + 306, + 257, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 545, + 290 + ], + "type": "text", + "content": "Figure 5. Example of different weights used in crossover of RFGs. While the redder regions look more similar to the father, and the bluer regions look more similar to the mother." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 305, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 354 + ], + "type": "text", + "content": "pool into nine groups by age, two by gender, and seven by race. Note that the RFG in the gene pool can independently control different facial traits. And its phenotype is related to the grouping condition." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": "Given a pair of parental face images and information about generated descendants, such as age, gender, and race. We first query the gene pool to extract a subset " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c\\subseteq \\mathcal{P}" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " that satisfies the descendants' information, which will be used to simulate gene mutations for better diversity in the descendants. Then we use Image Encoder and LGE to extract RFGs " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}^{f} = [g_{1}^{f},\\dots ,g_{N}^{f}]" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}^{m} = [g_{1}^{m},\\dots ,g_{N}^{m}]" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " of parents. For mutation, we randomly select " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " percent of facial regions whose RFGs are sampled from the gene pool. Given a " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}^p" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": ", a one-hot vector " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "t = \\{t_i\\in \\{0,1\\} ,i = 1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " is randomly generated to denote whether the " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i^c" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " of descendant is fully copied from " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i^p (t_i = 1)" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": ", or not " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "(t_i = 0)" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}_i^c" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " is inherited from parents " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "(t_i = 0)" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": ", we also introduce mutation by introducing " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}^p" + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "text", + "content": " into the crossover process defined in Eq. 9. Finally, we can extend Eq. 9 as follows to obtain the RFGs of the descendants " + }, + { + "bbox": [ + 304, + 354, + 546, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{g}^c = [\\pmb {g}_c^1,\\dots ,\\pmb {g}_c^N ]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 563, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 563, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 339, + 563, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {i} ^ {c} = \\left\\{ \\begin{array}{l l} \\boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 1 \\\\ \\alpha_ {i} \\boldsymbol {g} _ {i} ^ {f} + \\beta_ {i} \\boldsymbol {g} _ {i} ^ {m} + \\gamma \\boldsymbol {g} _ {i} ^ {p}, & t _ {i} = 0 \\end{array} , \\right. \\tag {10}", + "image_path": "c37cb27a96192408cdb56427849141c69b4587de812f506962937f039c0e288a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " is the intensity of mutation, " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " are randomly generated weights and " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\alpha_{i} + \\beta_{i} = 1 - \\gamma" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "g_{i}^{p} = S(\\mathcal{P}_{c})" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " is the random sampling operator. Once the RFG of descendant " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "g^{c}" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " is generated, the learned Gene Decoder is applied to map " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "g^{c}" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": " to the latent space of StyleGAN2, i.e. " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{c} = D(\\boldsymbol{g}^{c})" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{c} \\in \\mathbb{R}^{18 \\times 512}" + }, + { + "bbox": [ + 304, + 593, + 545, + 675 + ], + "type": "text", + "content": ", which is further processed for face synthesis." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "While the latent code in " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " latent space of StyleGAN2 usually consists of 18 layers, Richardson et al. [34] demonstrate that the first 8 layers of the latent code mainly con" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20964" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "tribute to the ID information of the synthesized faces, and higher layers mainly control the skin color and microstructure of the synthesized faces. While the appearance of facial regions for descendants is mainly decided by " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "g^{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ", other attributes like skin, hair color, image lighting, and background, can thus be inherited by inclusion of the latent codes of parents. Therefore, we further fuse the latent codes of parents " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{f}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{m}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ". We keep the first " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " layer of " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ", and the higher layers are fused from " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{f}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{m}" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": ". The final latent code " + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_{c} = [\\boldsymbol{w}_{c}^{1},\\dots ,\\boldsymbol{w}_{c}^{18}]" + }, + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": " of the descendant can be obtained by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 205, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 205, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 287, + 228 + ], + "type": "interline_equation", + "content": "\\boldsymbol {w} _ {c} ^ {i} = \\frac {1}{2} \\left(\\boldsymbol {w} _ {f} ^ {i} + \\boldsymbol {w} _ {m} ^ {i}\\right), \\quad i \\in \\{l + 1, \\dots , 1 8 \\}, \\tag {11}", + "image_path": "ab23fad24e348c350551140040f01345d9aa3168c9d7db44170decb3ed9e1c78.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 238, + 128, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 128, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 128, + 251 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 258, + 173, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 173, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 173, + 270 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "type": "text", + "content": "Datasets. Our training set consists of three parts, all images of CelebAHQ [21], 50,000 faces sampled from MS-Celeb-1M [18], and 10,000 faces generated by StyleGAN2, to make the model adaptive to different image qualities. CelebA-HQ is a high-quality face dataset consisting of 30,000 aligned face images with " + }, + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "type": "text", + "content": " resolution. The MS-Celeb-1M is a large-scale face recognition dataset containing about 10 million images, including many noisy face images. We evaluate our model on the FIW [35], TSKinFace [32], and FF-Database [48] datasets. FIW dataset contains the faces of 1,000 families with 1,997 sets of father-mother-child relations. The TSKinFace and FF-Database datasets provide 1,015 and 3,744 sets of father-mother-child kinship face images, respectively. We align and crop the images in these two datasets to " + }, + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 277, + 287, + 491 + ], + "type": "text", + "content": ". Our Gene Pool is built based on the FFHQ [23] dataset, which consists of 70,000 high-quality face images, and has a better diversity among human races and ages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 493, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 599 + ], + "type": "text", + "content": "Baselines. We compare our method against the state-of-art kinship face synthesis baselines, i.e. DNA-Net [14], ChildGAN [7], ChildPredictor [48], and StyleDNA [27]. Since the source codes of StyleDNA and ChildPredictor are available, we compare our approach with them through all of the evaluations. However, due to the unavailability of the source codes, we can only visually compare the visual quality of the generated faces available in the papers of DNA-Net and ChildGAN." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 601, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 649 + ], + "type": "text", + "content": "Training Details. We use AdamW [28] optimizer with batch size of 32. The initialized learning rate is 0.001, which is divided by 2 every 10 epochs, and we stop the training at the 30th epoch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 174, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 174, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 174, + 671 + ], + "type": "text", + "content": "4.2. Qualitative evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Disentanglement of RFGs. Fig. 6 shows two examples of facial region editing using our proposed RFG, where the nose, eyes, jaw and lips of source faces shown in the first" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "column, are replaced sequentially by the corresponding regions of reference faces shown in the first row. One can observe from the examples that, when a certain region is edited, all other facial regions are kept intact, which clearly shows the disentanglement capability of our RFG." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 140, + 537, + 289 + ], + "blocks": [ + { + "bbox": [ + 313, + 140, + 537, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 537, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 537, + 289 + ], + "type": "image", + "image_path": "64dd6f01136f79727a0bee14bc3c13014445cf72ed33d2982c92a1936d28a084.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 293, + 513, + 304 + ], + "lines": [ + { + "bbox": [ + 337, + 293, + 513, + 304 + ], + "spans": [ + { + "bbox": [ + 337, + 293, + 513, + 304 + ], + "type": "text", + "content": "Figure 6. Disentangled editing of facial regions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 310, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 453 + ], + "type": "text", + "content": "Effectiveness of Gene Pool. Based on the proposed idea of RFG, our approach further introduces a gene pool to increase the variations among the generated faces of descendants. Given a couple of parents shown in the first row, Fig. 7 shows the faces of five children generated by our approach with/without the involvement of gene pool, together with that synthesized by StyleDNA and ChildPredictor. While the faces generated by StyleDNA and ChildPredictor (4th and 5th rows) all look very similar, that synthesized by our approach are much more diverse and the introduction of gene pool (1st row) can further increase variations among children." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 463, + 537, + 658 + ], + "blocks": [ + { + "bbox": [ + 312, + 463, + 537, + 658 + ], + "lines": [ + { + "bbox": [ + 312, + 463, + 537, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 463, + 537, + 658 + ], + "type": "image", + "image_path": "96f0f3e296a85428c5ad729aa58a58ca200d4f642ca56e531efff9add0b86872.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 662, + 545, + 718 + ], + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 718 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 718 + ], + "type": "text", + "content": "Figure 7. Visual comparison of diversity with StyleDNA and ChildPredictor. The first row shows the father and mother. The next two rows are generated by our method with and without Gene Pool. The last two rows are generated by StyleDNA [27] and ChildPredictor [48], respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20965" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 325, + 215 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 325, + 215 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 325, + 215 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 325, + 215 + ], + "type": "image", + "image_path": "75637e407a67186632e1bf62e521eeb75f1c1afc3c63b8723c1e5bd7853f6ae1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 216, + 80, + 224 + ], + "lines": [ + { + "bbox": [ + 59, + 216, + 80, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 216, + 80, + 224 + ], + "type": "text", + "content": "Father" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 216, + 114, + 223 + ], + "lines": [ + { + "bbox": [ + 91, + 216, + 114, + 223 + ], + "spans": [ + { + "bbox": [ + 91, + 216, + 114, + 223 + ], + "type": "text", + "content": "Mother" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 216, + 152, + 224 + ], + "lines": [ + { + "bbox": [ + 119, + 216, + 152, + 224 + ], + "spans": [ + { + "bbox": [ + 119, + 216, + 152, + 224 + ], + "type": "text", + "content": "Real Child" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 216, + 177, + 223 + ], + "lines": [ + { + "bbox": [ + 156, + 216, + 177, + 223 + ], + "spans": [ + { + "bbox": [ + 156, + 216, + 177, + 223 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 216, + 207, + 224 + ], + "lines": [ + { + "bbox": [ + 181, + 216, + 207, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 216, + 207, + 224 + ], + "type": "text", + "content": "StyleDN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 216, + 244, + 224 + ], + "lines": [ + { + "bbox": [ + 211, + 216, + 244, + 224 + ], + "spans": [ + { + "bbox": [ + 211, + 216, + 244, + 224 + ], + "type": "text", + "content": "A ChildPre" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 216, + 293, + 224 + ], + "lines": [ + { + "bbox": [ + 247, + 216, + 293, + 224 + ], + "spans": [ + { + "bbox": [ + 247, + 216, + 293, + 224 + ], + "type": "text", + "content": "Victor ChildGAN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 216, + 325, + 223 + ], + "lines": [ + { + "bbox": [ + 296, + 216, + 325, + 223 + ], + "spans": [ + { + "bbox": [ + 296, + 216, + 325, + 223 + ], + "type": "text", + "content": "DNA-Net" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 538, + 216 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 538, + 216 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 538, + 216 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 538, + 216 + ], + "type": "image", + "image_path": "ac2707b07ecf1ccedd5f827ff9f2560a641eec8068aa0b9c6ec92f0cebe1b8b5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 216, + 356, + 223 + ], + "lines": [ + { + "bbox": [ + 335, + 216, + 356, + 223 + ], + "spans": [ + { + "bbox": [ + 335, + 216, + 356, + 223 + ], + "type": "text", + "content": "Father" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 368, + 216, + 391, + 223 + ], + "lines": [ + { + "bbox": [ + 368, + 216, + 391, + 223 + ], + "spans": [ + { + "bbox": [ + 368, + 216, + 391, + 223 + ], + "type": "text", + "content": "Mother" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 216, + 416, + 223 + ], + "lines": [ + { + "bbox": [ + 394, + 216, + 416, + 223 + ], + "spans": [ + { + "bbox": [ + 394, + 216, + 416, + 223 + ], + "type": "text", + "content": "Real 0" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 216, + 457, + 223 + ], + "lines": [ + { + "bbox": [ + 428, + 216, + 457, + 223 + ], + "spans": [ + { + "bbox": [ + 428, + 216, + 457, + 223 + ], + "type": "text", + "content": "d Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 467, + 216, + 544, + 224 + ], + "lines": [ + { + "bbox": [ + 467, + 216, + 544, + 224 + ], + "spans": [ + { + "bbox": [ + 467, + 216, + 544, + 224 + ], + "type": "text", + "content": "StyleDNA ChildPredictor" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 409, + 226, + 462, + 235 + ], + "lines": [ + { + "bbox": [ + 409, + 226, + 462, + 235 + ], + "spans": [ + { + "bbox": [ + 409, + 226, + 462, + 235 + ], + "type": "text", + "content": "(b) FF-Database" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 237, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 546, + 270 + ], + "type": "text", + "content": "Figure 8. Comparison of children faces synthesized by StyleGene and baselines. Left most three columns of (a) FIW and (b) FF-Database depict the father, mother, and real children, and right five or three columns depict the faces synthesized by ours, StyleDNA [27], ChildPredictor [48], ChildGAN [7] and DNA-Net [14]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 284, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 289, + 594 + ], + "type": "text", + "content": "Comparison with the state-of-the-art. Fig. 8(a) and (b) show qualitative results on FIW and FF-Database, respectively. We first compare StyleGene with StyleDNA, DNA-Net, ChildGAN, and ChildPredictor on FIW with low-resolution facial images. Note that the results of ChildGAN and DNA-Net are directly taken from their papers. As can be seen, DNA-Net, ChildGAN, and ChildPredictor struggle to yield high-quality facial images. Thanks to a powerful generator, our method can generate high-fidelity faces with hair. Though ChildGAN also uses the StyleGAN generator, its results are significantly affected by the quality of the input. Note that all comparisons have considered the facial information of parents. However, DNA-Net, ChildPredictor, and StyleDNA model facial genetics by learning the implicit mapping between parents and children, without considering regional similarity, which significantly compromised the similarity between the generated faces and their parents. Although ChildGAN considered parents' facial features, the results were too similar to their parents. Thanks to the region-level facial genes and the mutation introduced by gene pool, our results can maintain the local similarity to parents' faces, and at the same time, present reasonable global variations and diversity. When experimenting with faces of different races as shown in Fig. 8(b), our approach can still well preserve the race similarity of children with their parents." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 609, + 180, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 180, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 180, + 621 + ], + "type": "text", + "content": "4.3. Quantitative evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Kinship Verification. We now use kinship verification accuracy to assess whether a genetic relationship exists between the synthetic descendant and parents. Higher accuracy means more realistic synthetic descendants. We use ArcFace pre-trained on the MS-Celeb-1M dataset as the backbone to create a one-versus-one kinship classifier [42], which is fine-tuned on the FF-Database training set. Then" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 284, + 478, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 478, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 478, + 295 + ], + "type": "text", + "content": "we performed a cross-database evaluation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "text", + "content": "Specifically, we randomly sample 100 families with one kid from the FF-database test set, FIW, and TSKinFace datasets, respectively, and different methods randomly generate 40 children for each family. We randomly sample 8,000 positive and 8,000 negative pairs for each method from each dataset and calculate in Table 2 the kinship verification accuracy. One can observe that our method achieves substantially higher accuracy than StyleDNA and ChildPredictor, i.e. as high as " + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "inline_equation", + "content": "81.74\\%" + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "inline_equation", + "content": "80.38\\%" + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "inline_equation", + "content": "62.29\\%" + }, + { + "bbox": [ + 304, + 296, + 545, + 450 + ], + "type": "text", + "content": " accuracies are achieved on TSKinFace, FF-Database, and FIW datasets, respectively. The results suggest that the kinship relations of faces generated by our approach are very close to the real children." + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 309, + 484, + 545, + 544 + ], + "blocks": [ + { + "bbox": [ + 306, + 457, + 544, + 480 + ], + "lines": [ + { + "bbox": [ + 306, + 457, + 544, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 544, + 480 + ], + "type": "text", + "content": "Table 2. Kinship verification accuracy " + }, + { + "bbox": [ + 306, + 457, + 544, + 480 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 306, + 457, + 544, + 480 + ], + "type": "text", + "content": " on the TSKinFace [32], FF-Database [48], and FIW [35] dataset." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 484, + 545, + 544 + ], + "lines": [ + { + "bbox": [ + 309, + 484, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 309, + 484, + 545, + 544 + ], + "type": "table", + "html": "
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA [27]53.1555.1149.47
ChildPredictor [48]58.2459.6251.81
StyleGene (Ours)81.7480.3862.29
", + "image_path": "23587e21d80e1aef7eb66b02c86cb8b868c388d042b985f0f2b9a02dff337cca.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 546, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 700 + ], + "type": "text", + "content": "Diversity Evaluation. We used the LPIPS [44] metric to measure the diversity of synthetic descendants, which calculates the L1 distance between pairs of image features extracted by AlexNet [26] pre-trained on the ImageNet [10] dataset. We use the same test data of kinship verification. First, we calculated the distance among the 40 synthesized descendants for each family. Then we take the average of 100 families as the LPIPS score and list them in Table 3. As shown in the table, our method achieves the highest LPIPS across all three datasets, i.e. 0.3270, 0.3418, and 0.3279 LPIPS on TSKinFace, FF-Database, and FIW datasets are achieved, which is significantly higher than that of StyleDNA and ChildPredictor." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Distribution of Synthesized Children. We are now try-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20966" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 97, + 287, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of the diversity of the generated descendants. * means we cropped the face i.e. no hair." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 97, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 50, + 97, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 50, + 97, + 287, + 177 + ], + "type": "table", + "html": "
MethodsTSKinFaceFF-DatabaseFIW
StyleDNA* [27]0.07560.07630.0736
ChildPredictor [48]0.16970.17230.1750
StyleGene (Ours)*0.17480.17350.1740
StyleDNA [27]0.15590.15420.1573
StyleGene (Ours)0.32700.34180.3279
", + "image_path": "fcc33b211d841c3ea0c1f99ca4c387d7370f231777d1a17c9ed839e6bff3a715.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 187, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 288, + 426 + ], + "type": "text", + "content": "ing to model the distribution of the children synthesized by different approaches, and compare them with that of real children. Specifically, we randomly sample 50 families from the TSKinFace dataset, each consisting of a father, mother, son, and daughter. Based on the given parents, we also applied different synthesis methods to generate two children, i.e. a son and a daughter, for each family. After facial features are extracted using ArcFace [11] and reduced to one dimension using t-SNE [41], Kernel Density Estimation (KDE) [37] is further applied to estimate the features' probability density function. Fig. 9 shows the distribution of the 100 real children (blue) and that of children generated by different approaches. As can be seen, the distribution of faces generated by our approach (red) overlaps most with the real children (blue). In contrast, the distributions of faces synthesized by StyleDNA (black) and ChildPredictor (yellow) are located on both sides of the real ones. The distribution for StyleDNA even looks like a mixture of two Gaussian distributions, which is significantly different from that of real children." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 436, + 276, + 545 + ], + "blocks": [ + { + "bbox": [ + 60, + 436, + 276, + 545 + ], + "lines": [ + { + "bbox": [ + 60, + 436, + 276, + 545 + ], + "spans": [ + { + "bbox": [ + 60, + 436, + 276, + 545 + ], + "type": "image", + "image_path": "f488ad36f0d2931814796176ee795e854e0fa8ea59acee6f85d2994f8d580d10.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 553, + 287, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 586 + ], + "type": "text", + "content": "Figure 9. Distribution of real children (blue) and children generated using StyleDNA (black), ChildPredictor (yellow), and ours (red). Best viewed in color." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "User Study. To further demonstrate the effectiveness of the proposed method, we conducted a user study involving 59 participants. The faces of children generated by our StyleGene, StyleDNA and ChildPredictor based on 20 parents, are present to each of the 59 participants, who are asked to rank the quality of the faces, in terms of realness and the similarity with given parents. In total " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "1,180(59 \\times 20)" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " votes are received for each of the three approaches and the average rankings are listed in Table 4. As shown in the table, our StyleGene receives the highest rank among the com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "pared approaches, which suggests that the faces of children synthesized by our approach are the best, in terms of both quality and similarity with parents." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 126, + 533, + 159 + ], + "blocks": [ + { + "bbox": [ + 324, + 111, + 526, + 123 + ], + "lines": [ + { + "bbox": [ + 324, + 111, + 526, + 123 + ], + "spans": [ + { + "bbox": [ + 324, + 111, + 526, + 123 + ], + "type": "text", + "content": "Table 4. The rank of different approaches in user study." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 126, + 533, + 159 + ], + "lines": [ + { + "bbox": [ + 319, + 126, + 533, + 159 + ], + "spans": [ + { + "bbox": [ + 319, + 126, + 533, + 159 + ], + "type": "table", + "html": "
ChildPredictorStyleDNAStyleGene (Ours)
Avg. rank2.462.221.32
", + "image_path": "997645d8ab3582f708ad7098e38be642085565f1fbe5c8b907a7cbd32f1fe7e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 168, + 397, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 397, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 397, + 181 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 186, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 307 + ], + "type": "text", + "content": "In this section, we perform ablation studies on the effectiveness of LGE, IGE and GP, in terms of kinship verification accuracy (ACC) and diversity (LPIPS) using FF-database with the same configuration as in Section 4.3. As shown in Table 5, the integration of IGE (2nd row) can improve both ACC and LPIPS of the synthesized descendant. As expected, the use of GP (3rd row) increases the diversity of synthesized faces. When both IGE and GP are integrated with LGE (4th row), the diversity can be significantly increased to 0.1735, with a comparable ACC " + }, + { + "bbox": [ + 304, + 186, + 545, + 307 + ], + "type": "inline_equation", + "content": "(80.38\\%)" + }, + { + "bbox": [ + 304, + 186, + 545, + 307 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 335, + 327, + 515, + 396 + ], + "blocks": [ + { + "bbox": [ + 351, + 312, + 500, + 323 + ], + "lines": [ + { + "bbox": [ + 351, + 312, + 500, + 323 + ], + "spans": [ + { + "bbox": [ + 351, + 312, + 500, + 323 + ], + "type": "text", + "content": "Table 5. Ablation study on FF-Database." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 327, + 515, + 396 + ], + "lines": [ + { + "bbox": [ + 335, + 327, + 515, + 396 + ], + "spans": [ + { + "bbox": [ + 335, + 327, + 515, + 396 + ], + "type": "table", + "html": "
LGEIGEGPACC (%)LPIPS
1✓XX79.190.0625
2✓✓X79.660.0646
3✓X✓80.210.0839
4✓✓✓80.380.1735
", + "image_path": "a635b334709e42d5d046ff799b5a61e06fcc1d661411737d28b71c8598b63c36.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 405, + 427, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 427, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 427, + 418 + ], + "type": "text", + "content": "4.5. Parameter sensitivity" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": "We test different values of " + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": " to see how the performance of synthesis varies. Due to the page limit, details about the results of different parameters are presented in supplementary. In our experiments, we choose " + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\eta = 40\\%" + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\gamma = 0.47" + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "inline_equation", + "content": "l = 8" + }, + { + "bbox": [ + 304, + 423, + 545, + 496 + ], + "type": "text", + "content": " to achieve a balance between the diversity and fidelity of the generated faces." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 504, + 379, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 379, + 516 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 524, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 632 + ], + "type": "text", + "content": "In this paper, we have proposed StyleGene, to extract RFGs (Region-level Facial Genes) for kinship face synthesis. While crossover and mutation of RFGs are proposed to model the facial genetic process, Gene Pool is further designed to increase the diversity among generated faces. Quantitative, qualitative, and subjective experimental results show that the realness, similarity with parents, and diversity of kinship faces generated by our approach are much better than existing state-of-the-art methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 636, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 545, + 703 + ], + "type": "text", + "content": "Acknowledgement. This work was jointly supported by National Natural Science Foundation of China (62206180 and 82261138629), Guangdong Basic and Applied Basic Research Foundation (2023A1515010688 and 2022A1515011018), and Shenzhen Municipal Science and Technology Innovation Council (JCYJ20220531101412030)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20967" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 245 + ], + "type": "text", + "content": "[3] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 302 + ], + "type": "text", + "content": "[4] Peter Claes, Jasmien Roosenboom, Julie D White, Tomek Swigut, Dzemila Sero, Jiarui Li, Myoung Keun Lee, Arslan Zaidi, Brooke C Mattern, Corey Liebowitz, et al. Genome-wide mapping of global-to-local genetic effects on human facial shape. Nature genetics, 50(3):414-423, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "text", + "content": "[5] Joanne B Cole, Mange Manyama, Jacinda R Larson, Denise K Liberton, Tracey M Ferrara, Sheri L Riccardi, Mao Li, Washington Mio, Ophir D Klein, Stephanie A Santorico, et al. Human facial shape and size heritability and genetic correlations. Genetics, 205(2):967-978, 2017. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 402 + ], + "type": "text", + "content": "[6] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: Uncovering the local semantics of gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5771-5780, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 446 + ], + "type": "text", + "content": "[7] Xiao Cui, Wengang Zhou, Yang Hu, Weilun Wang, and Houqiang Li. Heredity-aware child face image generation with latent space disentanglement. arXiv preprint arXiv:2108.11080, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "text", + "content": "[8] Maria F Dal Martello and Laurence T Maloney. Lateralization of kin recognition signals in the human face. Journal of Vision, 10(8):9-9, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "text", + "content": "[9] Afshin Dehghan, Enrique G Ortiz, Ruben Villegas, and Mubarak Shah. Who do i look like? determining parent-offspring resemblance via gated autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1757-1764, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 581 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 691 + ], + "type": "text", + "content": "[12] Itir Önal Ertugrul and Hamdi Dibeklioglu. What will your future child look like? modeling and synthesis of hereditary patterns of facial dynamics. In 2017 12th IEEE international conference on automatic face & gesture recognition (FG 2017), pages 33-40. IEEE, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[14] Pengyu Gao, Joseph Robinson, Jiaxuan Zhu, Chao Xia, MIng Shao, and Siyu Xia. Dna-net: Age and gender aware kin face synthesizer. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[15] Pengyu Gao, Siyu Xia, Joseph Robinson, Junkang Zhang, Chao Xia, Ming Shao, and Yun Fu. What will your child look like? dna-net: Age and gender aware kin face synthesizer. arXiv preprint arXiv:1911.07014, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[16] Fady S Ghatas and ElSayed E Hemayed. Gankin: generating kin faces using disentangled gan. SN Applied Sciences, 2(2):1-10, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 275 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "type": "text", + "content": "[18] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In European conference on computer vision, pages 87–102. Springer, 2016. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 321, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 387 + ], + "type": "text", + "content": "[19] Hanne Hoskens, Dongjing Liu, Sahin Naqvi, Myoung Keun Lee, Ryan J Eller, Karlijne Indencleef, Julie D White, Jiarui Li, Maarten HD Larmouseau, Greet Hens, et al. 3d facial phenotyping by biometric sibling matching used in contemporary genomic methodologies. PLoS genetics, 17(5):e1009528, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[20] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 455 + ], + "type": "text", + "content": "[21] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[22] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 601 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 624 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "text", + "content": "[27] Che-Hsien Lin, Hung-Chun Chen, Li-Chen Cheng, Shu-Chuan Hsu, Jun-Cheng Chen, and Chih-Yu Wang. *Styledna: A high-fidelity age and gender aware kinship face synthesizer*. In 2021 16th IEEE International Conference on Auto" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20968" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "matic Face and Gesture Recognition (FG 2021), pages 1-8. IEEE, 2021. 2,3,6,7,8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 286, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 286, + 119 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 121, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 287, + 174 + ], + "type": "text", + "content": "[29] Savas Ozkan and Akin Ozkan. Kinshipgan: Synthesizing of kinship faces from family photos by regularizing a deep face network. In 2018 25th IEEE international conference on image processing (ICIP), pages 2142-2146. IEEE, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 286, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 286, + 199 + ], + "type": "text", + "content": "[30] Svante Paäbo. The diverse origins of the human gene pool. Nature Reviews Genetics, 16(6):313-314, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 255 + ], + "type": "text", + "content": "[31] Shouneng Peng, Jingze Tan, Sile Hu, Hang Zhou, Jing Guo, Li Jin, and Kun Tang. Detecting genetic association of common human facial morphological variation using high density 3d image registration. PLoS computational biology, 9(12):e1003375, 2013. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 287, + 300 + ], + "type": "text", + "content": "[32] Xiaoqian Qin, Xiaoyang Tan, and Songcan Chen. Tri-subject kinship verification: Understanding the core of a family. IEEE Transactions on Multimedia, 17(10):1855-1867, 2015. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 346 + ], + "type": "text", + "content": "[33] Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 404 + ], + "type": "text", + "content": "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 450 + ], + "type": "text", + "content": "[35] Joseph P Robinson, Ming Shao, Yue Wu, and Yun Fu. Families in the wild (fw) large-scale kinship image database and benchmarks. In Proceedings of the 24th ACM international conference on Multimedia, pages 242-246, 2016. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 287, + 495 + ], + "type": "text", + "content": "[36] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9243-9252, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 519 + ], + "type": "text", + "content": "[37] Bernard W Silverman. Density estimation for statistics and data analysis. Routledge, 2018. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 565 + ], + "type": "text", + "content": "[38] Raunak Sinha, Mayank Vatsa, and Richa Singh. Familygan: Generating kin face images using generative adversarial networks. In European conference on computer vision, pages 297-311. Springer, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 632 + ], + "type": "text", + "content": "[39] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[41] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 529 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[42] Jun Yu, Mengyan Li, Xinlong Hao, and Guochen Xie. Deep fusion siamese network for automatic kinship verification. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 892-899. IEEE, 2020. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[43] Ishtiak Zaman and David Crandall. Genetic-gan: Synthesizing images between two domains by genetic crossover. In European Conference on Computer Vision, pages 312-326. Springer, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[45] Yong Zhang, Le Li, Zhilei Liu, Baoyuan Wu, Yanbo Fan, and Zhifeng Li. Controllable descendant face synthesis. arXiv preprint arXiv:2002.11376, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 329 + ], + "type": "text", + "content": "[46] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10145-10155, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 374 + ], + "type": "text", + "content": "[47] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "text", + "content": "[48] Yuzhi Zhao, Lai-Man Po, Xuehui Wang, Qiong Yan, Wei Shen, Yujia Zhang, Wei Liu, Chun-Kit Wong, Chiu-Sing Pang, Weifeng Ou, et al. Childpredictor: A child face prediction framework with disentangled learning. IEEE Transactions on Multimedia, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 431, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 473 + ], + "type": "text", + "content": "[49] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "text", + "content": "[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20969" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_content_list.json b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5045e285768bd905ffa57381bfaf08820fa524ef --- /dev/null +++ b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "StyleIPSB: Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping", + "text_level": 1, + "bbox": [ + 163, + 128, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diqiong Jiang $^{1}$ , Dan Song $^{2*}$ , Ruofeng Tong $^{1*}$ , Min Tang $^{1}$ $^{1}$ Zhejiang University, China \n $^{2}$ Tianjin University, China", + "bbox": [ + 256, + 202, + 712, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3722d226273fccf98276d40e8e3aa3a301103da5b224495b7c844d434c1282f4.jpg", + "image_caption": [ + "Figure 1. Compared with the existing high-fidelity face swapping methods MageFS [51] and FSLDS [47], our face swapping framework based on the identity-preserving semantic basis (StyleIPSB) can preserve the pore-level detail and source image's identity." + ], + "image_footnote": [], + "bbox": [ + 107, + 260, + 890, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 579, + 313, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent researches reveal that StyleGAN can generate highly realistic images, inspiring researchers to use pretrained StyleGAN to generate high-fidelity swapped faces. However, existing methods fail to meet the expectations in two essential aspects of high-fidelity face swapping. Their results are blurry without pore-level details and fail to preserve identity for challenging cases. To overcome the above artifacts, we innovatively construct a series of identity-preserving semantic bases of StyleGAN (called StyleIPSB) in respect of pose, expression, and illumination. Each basis of StyleIPSB controls one specific semantic attribute and disentangles with the others. The StyleIPSB constrains style code in the subspace of $W+$ space to preserve pore-level details and gives us a novel tool for high-fidelity face swapping, and we propose a three-stage framework for face swapping with StyleIPSB. Firstly, we transform the", + "bbox": [ + 73, + 611, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "target facial images' attributes to the source image. We learn the mapping from 3D Morphable Model (3DMM) parameters, which capture the prominent semantic variance, to the coordinates of StyleIPSB that show higher identity-preserving and fidelity. Secondly, to transform detailed attributes which 3DMM does not capture, we learn the residual attribute between the reenacted face and the target face. Finally, the face is blended into the background of the target image. Extensive results and comparisons demonstrate that StyleIPSB can effectively preserve identity and pore-level details. The results of face swapping can achieve state-of-the-art performance. We will release our code at https://github.com/a686432/StyleIPSB", + "bbox": [ + 500, + 580, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 811, + 632, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Facial image manipulation [36, 37, 48, 50] is a task of transforming specific attributes from the source image to the target image while persevering other attributes unchanged. Face swapping is one of the essential parts of facial im", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-Corresponding authors", + "bbox": [ + 94, + 862, + 243, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ The research is supported in part by the National Natural Science Foundation of China (61972342, 61832016, 61972341, 61902277)", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "352", + "bbox": [ + 485, + 944, + 511, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age manipulation, which has attracted lots of interest in the computer vision and graphics community. Face swapping aims to generate an image with the source image's identity and the target image's attributes (e.g., expression, pose, background, hair, etc.). It has wide applications in the film industry and computer games.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current face swapping methods are mainly divided into two categories: 3D model-based methods and 2D image-based methods. 3D model-based methods [11,17,40] firstly reconstruct the 3D face models based on 3DMM from the source image and target image and transfer the nonidentity parameters of the target face model to the source face model. Then they render the transferred 3D model and blend it into the target image. Although such methods can transfer coarse facial attributes such as pose, expression, and illumination, they have difficulty in generating realistic hair and teeth accessories.", + "bbox": [ + 75, + 184, + 468, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the development of generative adversarial networks, 2D image-based methods [6, 7, 22, 27, 28, 45] can synthesize photo-realistic images. The generated face images have convincing and detailed facial attributes, such as mouth, teeth, and eyebrows. Recent works [21, 46, 47, 51] employ the pre-trained StyleGAN decoder to further improve the fidelity and synthesize pore-level details. However, as shown in Fig. 1, despite using the pre-trained StyleGAN model, their results fail to generate pore-level details and identity-preserving in challenge conditions. Overall, 2D image-based methods generate more realistic images than 3D model-based methods, but the identity-preserving and pore-level details of the images still need improvement.", + "bbox": [ + 75, + 352, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenges of blurry images and identity-preserving in face swapping, according to our observations, the cause of the blurred images is that the regressed style code is out of $\\mathrm{W + }$ space. Additionally, as mentioned in [18], identity embedding is a non-smooth space, so finding the identity-preserving optimized direction is challenging. To address these problems, our method constrains the regressed style code with identity-preserving semantic bases of StyleGAN (i.e., the proposed StyleIPSB). StyleIPSB stays within the $\\mathrm{W + }$ space, and the identity is preserved when changing its coordinates.", + "bbox": [ + 75, + 550, + 467, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The advantages of the proposed StyleIPSB are summarized as follows: (1) StyleIPSB constitutes a linear space, which is the subspace of the $\\mathrm{W + }$ space of StyleGAN. By ensuring the regressed style code within the $\\mathrm{W + }$ space of StyleGAN, we can more easily generate images with pore-level details. (2) When changing the coordinates of the StyleIPSB, the identity remains preserved as much as possible. (3) StyleIPSB can represent various poses, expressions, and illuminations. To construct the basis that satisfies the above properties, we propose a novel identity-preserving distance metric to find the orthogonal semantic directions, which are further assembled to StyleIPSB.", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleIPSB also cooperates well with 3DMM to control facial attributes. StyleRig [39] builds the mapping of the 3DMM parameter space and $\\mathrm{W + }$ space of StyleGAN, which can change the facial attribute of the generated image by the 3DMM parameters. StyleRig only can manipulate images generated by StyleGAN. Pie [38] designs a non-linear optimization problem to edit the real-world image based on StyleRig, but the optimization operation is time-consuming. GIF [12] generates face images by the FLAME [23] parametric control. However, the generated images are easy to contain artifacts and change identity. Other face manipulation methods [4, 29, 37] use the network directly to find the edit direction. Still, without the guidance of 3DMM, they can only generate some basic expressions (e.g., smile) and fail to cover various expressions. In this paper, we propose the StyleGAN-3DMM mapping network, which transforms the semantic information of 3DMM parameters into StyleIPSB coordinates. The StyleGAN-3DMM mapping network reduces the gap between 3DMM and StyleIPSB. It shows that StyleIPSB is very compatible with 3DMM.", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose a face swapping framework based on StyleIPSB and achieve state-of-the-art results. The main contributions of this paper lie in the following three aspects:", + "bbox": [ + 496, + 393, + 890, + 454 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel method of establishing identity-preserving semantic bases of StyleGAN called StyleIPSB. The face image, generated by the linear space of StyleIPSB, remains pore-level details and identity-preserving.", + "- The proposed StyleGAN-3DMM mapping network serves as the bridge to narrow the gap between 3DMM and StyleIPSB, which can take advantage of the prominent semantic variance of 3DMM and the identity-preserving and high-fidelity of styleIPSB.", + "- We propose the face swapping framework based on StyleIPSB and StyleGAN-3DMM mapping network. Extensive results show our method outperforms others in detail-preserving and identity-preserving." + ], + "bbox": [ + 517, + 465, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 715, + 643, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Modification Using StyleGAN", + "text_level": 1, + "bbox": [ + 500, + 739, + 816, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGAN [14-16] is a powerful image synthesis model that can generate a wide variety of high-quality face images. Some methods [5, 31, 34, 35] find the particular attribute editing direction in $\\mathbf{W}+$ space using the corresponding attribute classifier. However, their attribute editing is limited by the classifier's ability and cannot control diverse expressions and illumination. To facilitate attribute manipulation in an unsupervised manner, GANSpace [13] and SeFa [36] perform decomposition to find primary directions", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "353", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the latent space and explore the interpretable directions among the primary directions. CLIP2StyleGAN [4] discovery and label of StyleGAN edit-direction based on CLIP image space. However, they only find a few types of expressions and illumination and fail to discover a complete set of expressions and illumination. Some works [12,38,39] map the rigging information to face manipulation, but our StyleIPSB can generate more detail and identity-preserving results.", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Face swap", + "text_level": 1, + "bbox": [ + 76, + 234, + 194, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the early years, many works [24-26, 40] apply 3D face models for face swapping. Face2face [40] applies an efficient deformation transfer to track the source video facial expressions and re-render the synthesized faces with retrieved and warped mouth interiors. Ma et al. [25] reconstructed high-resolution facial geometry and appearance by capturing an individual-specific face model with fine-scale details. Those 3D model-based methods are difficult to generate high-fidelity face images, especially realistic hair, mouth, and teeth. In recent years, conditional GAN architecture has been widely used in face-swapping. Many works [7, 19, 21, 22, 28, 46, 49] use neural networks to generate high-fidelity images. SimSwap [7] and FaceShifter [22] use a face recognition network to extract the identity embedding and use a decoder to fuse the identity embedding. Recently, many methods [21, 45-47, 51] use pretrained StyleGAN as the image generator to improve the quality of generated images further. However, the images generated by their method still suffer from the lack of details and face identity shift. Our approach is better at retaining pore-level details and identify-preserving.", + "bbox": [ + 75, + 257, + 472, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 587, + 169, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section reveals the details of our proposed method in three aspects. The first part describes the process of conducting StyleIPSB. The second part describes how to use StyleIPSB to transfer the facial attributes except for identity from the target image to the source face. The third introduces the overall framework of face swap based on StyleIPSB.", + "bbox": [ + 75, + 612, + 468, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. StyleIPSB Construction", + "text_level": 1, + "bbox": [ + 76, + 726, + 297, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This subsection aims to find a semantic basis in the $\\mathrm{W + }$ space of StyleGAN that satisfies the conditions mentioned in the introduction. In short, StyleIPSB should meet the following properties: a subspace of $\\mathrm{W + }$ space, identity-preserving, and representation ability. The previous methods fail to satisfy the above three conditions. For example, as shown in Tab. 1, GIF [12] does not meet the requirements in the $\\mathrm{W + }$ subspace, which would lead to generated image distortion. Some methods [12,35,36] do not concern the identity preservation and suffer the identity shift when", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0611b8c72e8e5061b27d148d175448e25fe0db41ae49fd485550ce7b9cff68bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSubspaceIDRepresentation
InterFaceGAN [34]✓✓
GANSpace [13]✓
SeFa [36]✓
GIF [12]✓
Our✓✓✓
", + "bbox": [ + 503, + 89, + 890, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. The conditions of the basis satisfied by different methods. We compare our method with InterFaceGAN [34], GANSpace [13], GIF [12] and SeFa [36].", + "bbox": [ + 498, + 209, + 893, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "editing. InterFaceGAN [34] cannot edit various expressions and illumination. Unlike existing methods, our method considers all three conditions when we build StyleIPSB.", + "bbox": [ + 496, + 277, + 890, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared with the previous methods, which built the basis by directly performing decomposition on eigenvalue in $\\mathrm{W + }$ space, we added semantic metrics and identity loss in our decomposition. Therefore, StyleIPSB is not only in the $\\mathrm{W + }$ subspace but can also represent the complete set of semantic information and have the property of identity preservation. So adding semantic information and maintaining identity when building StyleIPSB is the crucial point to this subsection. The following introduces the detailed process of conducting StyleIPSB.", + "bbox": [ + 496, + 321, + 892, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Formulation** The StyleGAN network, denoted by $G$ , is a mapping from style code $w$ to image $I$ , $G: R^n \\to R^{H \\times W \\times 3}$ , $w \\mapsto I$ . The 3DMM Fitting network, denoted by $M$ , regresses the 3DMM parameters $p$ from image $I$ . The parameters include pose parameter $p_p$ , expression parameter $p_e$ and illumination parameter $p_i$ . In this paper, we use DECA [10] as our 3DMM Fitting network.", + "bbox": [ + 496, + 473, + 893, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first define the distance metric of the pose, expression, and illumination. The distance metrics measure the attribute difference between two images generated by style codes $w_{1}, w_{2}$ . In addition, we add identity loss to the metric. Therefore, the distance metric increases if the difference between two images' attributes increases and the identity remains similar. Then we decompose the Hessian matrix to find the direction with the fastest distance metric change in the $\\mathrm{W+}$ space. Therefore, the attributes change fast but identity changes slowly along the direction we found. The distance metrics are shown as follows:", + "bbox": [ + 496, + 579, + 893, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {p} (w _ {1}, w _ {2}) = \\frac {| | M (G (w _ {1})) _ {p} - M (G (w _ {2})) _ {p} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 752, + 846, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {e} \\left(w _ {1}, w _ {2}\\right) = \\frac {\\left| \\left| M \\left(G \\left(w _ {1}\\right)\\right) _ {e} - M \\left(G \\left(w _ {2}\\right)\\right) _ {e} \\right| \\right| ^ {2}}{L _ {i d} \\left(G \\left(w _ {1}\\right) , G \\left(w _ {2}\\right)\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 787, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} (w _ {1}, w _ {2}) = \\frac {| | M (G (w _ {1})) _ {i} - M (G (w _ {2})) _ {i} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 825, + 841, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{id}$ measures the similarity of the identities of two images. Its value represents the similarity of two identity", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "354", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5dcf237d6c2b1d96e9b931f7e8214e2d4c460176a5b93452d1633f902bd0c7b6.jpg", + "image_caption": [ + "Figure 2. Our framework is divided into three stages. Stage one transfers the target image's expression, pose, and illumination attributes to the source image through the semantic base. Stage two transfers detailed non-identity attributes. Stage three blends the transferred image into the background of the target image." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 883, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "embeddings obtained by the pre-trained face recognition network [8] from two input face images.", + "bbox": [ + 75, + 521, + 470, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i d} \\left(I _ {1}, I _ {2}\\right) = 1 - \\frac {I D \\left(I _ {1}\\right) \\cdot I D \\left(I _ {2}\\right)}{\\left| \\left| I D \\left(I _ {1}\\right) \\right| \\right| _ {2} \\left| \\left| I D \\left(I _ {2}\\right) \\right| \\right| _ {2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 560, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we use the Hessian matrix to encode the local distance information on the image manifold up to second-order approximation.", + "bbox": [ + 75, + 604, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {2} (w _ {0}, w _ {0} + \\delta w) \\approx | | \\delta w | | _ {H} ^ {2} = \\delta_ {w} ^ {T} H (w _ {0}) \\delta w \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 659, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $w_0$ used to construct the semantic base is randomly sampled, and $w_0 + \\delta w$ is the point near $w_0$ . We decompose the Hessian matrix $H(w_0) = \\sum_i \\lambda_i v_i v_i^T$ and find the first $m$ principal components of $v$ as our basis $V$ . Therefore, the basis $V$ contains the direction vectors, along which the distance metric changes first $m$ fastest.", + "bbox": [ + 75, + 689, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the algorithm [42] to calculate and decompose the Hessian matrix. StyleGAN has 18 different levels, and the style code of each level itself contains certain semantic features (for example, the low level includes the features of face shape and pose, and the high level has the skin color and illumination, etc.). We construct the semantic base of the pose on the first three levels of the style code and the expression on the 4th to 10th levels. The semantic base of", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the illumination is on the levels after the 10th. We sampled $w_{0}$ 100 times and generated 100 sets of bases. And then, we average these bases and performed Schmidt orthogonalization to obtain the semantic base of the pose $V_{p}$ , expression $V_{e}$ , and illumination $V_{i}$ , respectively. Finally, the pose, expression, and illumination base are combined as StyleIPSB $V = [V_{p}, V_{e}, V_{i}]$", + "bbox": [ + 498, + 520, + 892, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3DMM-StyleGAN Mapping", + "text_level": 1, + "bbox": [ + 500, + 637, + 751, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the source and target images are fed into the 3DMM fitting network to regress the 3DMM parameters $\\alpha^s, \\alpha^t$ . The 3DMM-StyleGAN Mapping network, which contains three six-layer multilayer perceptrons(MLPs), maps 3DMM parameters $\\alpha$ to StyleIPSB coordinates of pose $\\beta_{s \\rightarrow t}^p$ , expression $\\beta_{s \\rightarrow t}^e$ and illumination $\\beta_{s \\rightarrow t}^i$ , respectively. The following equations calculate the transferred style code $w_{s \\rightarrow t}$ :", + "bbox": [ + 498, + 661, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s \\rightarrow t} = w _ {s} + \\beta_ {s \\rightarrow t} V \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 796, + 890, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{s\\rightarrow t} = [\\beta_{s\\rightarrow t}^p,\\beta_{s\\rightarrow t}^e,\\beta_{s\\rightarrow t}^i ]$ and $w_{s}$ is the corresponding style code of the source image, which is obtained from the source image by the StyleGAN encoder [41]. Finally, $w_{s\\rightarrow t}$ passes through StyleGAN to get the attribute transferred image.", + "bbox": [ + 496, + 824, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "355", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/18f48963d7dbc37f62adb92016013c716c9a7b0ba2c6802bd6d3976268529d5b.jpg", + "image_caption": [ + "Figure 3. The network structure of the image blend. Stage three blends transformed image to the target image by mixing the VGG encoder feature with the StyleGAN feature through masked spatial feature transforms (MSFT)." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 571, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/566fe69ac52d9fc12ef4c970e107361e7398ae1b87abe8b8c8cfb52277d1ff73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 90, + 915, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3DMM-StyleGAN mapping network aims at precisely controlling the generated attribute through the 3DMM parameters while keeping the identity unchanged. Therefore, we design the attribute loss $L_{attr}$ to make the generated face image have the same attribute as the pose, expression, and illumination as target 3DMM parameters. And we add identity loss $L_{id}$ to preserve the identity of the source image. The total loss is shown as:", + "bbox": [ + 75, + 362, + 468, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {i d} + \\varepsilon_ {\\text {a t t r}} L _ {\\text {a t t r}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 503, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{id}$ measure the identity similarity of two images and defined in Eq. (2). $\\varepsilon_{attr}$ is the weight of two loss functions. $L_{attr}$ measures the difference between the attributes of the transferred face and the target face. We reconstruct the 3D face from the transferred face and the target face, and then compare their difference in 3D face geometries and rendered images:", + "bbox": [ + 75, + 530, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {a t t r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) = L _ {g e o} \\left(\\alpha^ {s}, \\alpha^ {s \\rightarrow t}\\right) (6) \\\\ + L _ {r e n d e r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 648, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha^s, \\alpha^t$ is the 3DMM parameters of the source and target images. And $\\alpha^{s\\rightarrow t}$ is 3DMM parameters of the transferred image where $\\alpha^{s\\rightarrow t} = M(G(w_{s\\rightarrow t}))$ .", + "bbox": [ + 75, + 700, + 468, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define that $G_{3DMM}(\\alpha_s, \\alpha_e, \\alpha_p)$ can generate 3D geometry from 3DMM parameters, and $R(\\alpha_s, \\alpha_e, \\alpha_a, \\alpha_i, \\alpha_p)$ can generate rendered images from 3DMM parameters, where $\\alpha_s, \\alpha_e, \\alpha_a, \\alpha_i, \\alpha_p$ represent shape parameters, pose parameters, expression parameters, albedo parameters, and lighting parameters, respectively. Because we only compare attribute differences between transferred and target images, we reconstruct the 3D face by the same 3DMM shape parameters.", + "bbox": [ + 75, + 748, + 468, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The geometric term $L_{geo}$ uses the $L_{2}$ loss between two", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "face meshes:", + "bbox": [ + 500, + 363, + 588, + 376 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g e o} \\left(\\alpha^ {s}, \\alpha^ {s \\rightarrow t}\\right) =\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 382, + 625, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N} \\left\\| G _ {3 D M M} \\left(\\alpha_ {s} ^ {s}, \\alpha_ {e} ^ {t}, \\alpha_ {p} ^ {t}\\right) - G _ {3 D M M} \\left(\\alpha_ {s} ^ {s \\rightarrow t}, \\alpha_ {e} ^ {s \\rightarrow t}, \\alpha_ {p} ^ {s \\rightarrow t}\\right)\\right\\| _ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 402, + 900, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of vertices of the face mesh.", + "bbox": [ + 500, + 443, + 841, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The render term $L_{\\text{render}}$ uses the $L_1$ loss between two rendered images:", + "bbox": [ + 498, + 458, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e n d e r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) =\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 494, + 669, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| R \\left(\\alpha_ {s} ^ {s}, \\alpha_ {e} ^ {t}, \\alpha_ {a} ^ {t}, \\alpha_ {i} ^ {t}, \\alpha_ {p} ^ {t}\\right) - R \\left(\\alpha_ {s} ^ {s \\rightarrow t}, \\alpha_ {e} ^ {s \\rightarrow t}, \\alpha_ {a} ^ {t}, \\alpha_ {i} ^ {s \\rightarrow t}, \\alpha_ {p} ^ {s \\rightarrow t}\\right)\\right\\| _ {1} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 515, + 906, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Detailed Attribute Transformation", + "text_level": 1, + "bbox": [ + 500, + 555, + 800, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous subsection, we transfer the attributes of the target face to the source face, including pose, expression, and illumination. In this subsection, to transfer attributes beyond the 3DMM expressive capabilities, we use DAT-MLP to transfer more detailed attributes.", + "bbox": [ + 496, + 578, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we project the target images into the $W+$ space using the StyleGAN encoder to obtain the style code $w_{t}$ . Then, DAT-MLP regresses $\\lambda$ from style codes $w_{t}, w_{s \\to t}$ using six-layer MLP. Finally, we use the following formula to get the latent vector $w_{s \\to t}^{\\prime}$ that contains non-identity attributes of the target image:", + "bbox": [ + 496, + 654, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s \\rightarrow t} ^ {\\prime} = \\lambda w _ {s \\rightarrow t} + (1 - \\lambda) w _ {t} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 752, + 890, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda \\in R^{18\\times 512}$ is constrained to be between 0 and 1 through the sigmoid activation layer. We want to transfer attributes of $w_{t}$ into $w_{s\\rightarrow t}^{\\prime}$ as many as possible. The loss function of the second stage is as follows:", + "bbox": [ + 496, + 776, + 890, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {i d} + \\varepsilon_ {p} L _ {p} + \\varepsilon_ {\\text {a t t r}} L _ {\\text {a t t r}} \\left(\\alpha^ {s}, \\alpha^ {t}, M \\left(G \\left(w _ {s \\rightarrow t} ^ {\\prime}\\right)\\right)\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 845, + 890, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function is divided into three items. The first item is face recognition loss, which is used to constrain", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "356", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the identity of the generated face to be consistent with the source image. The second item $L_{p}$ represents the perceptual distance between the generated image and the target image. It can help transfer residual attributes other than pose, expression, and illumination. The third term $L_{attr}$ constrains the attributes of the generated face consistent with the target image. $\\varepsilon_{p}, \\varepsilon_{attr}$ represent the weight of the loss function.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Eq. (9) shows that the style code $w_{s \\to t}^{\\prime}$ is in a bounding box of two style codes $w_{s \\to t}$ and $w_{t}$ . We consider that most of the bounding box is still in the subspace of $\\mathbf{W}+$ space because the attribute of two style codes $w_{s \\to t}$ and $w_{t}$ is very close and generated image retains pore-level details in practice.", + "bbox": [ + 75, + 198, + 467, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Blend Image", + "text_level": 1, + "bbox": [ + 76, + 296, + 210, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This subsection introduces how to blend the generated image into the target image. We propose Masked Spatial Feature Transform (MSFT) module to fuse the feature. Unlike traditional Spatial Feature Transform (SFT) [44], MSFT only fuses the feature of the masked regions.", + "bbox": [ + 75, + 319, + 467, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The detail of this module is shown in Fig. 3. The VG-GFace network [30] extracts various level features of the target image, which is injected into the StyleGAN through the MSFT module. We use Gaussian filtering to filter the masked image obtained by the face segmentation algorithm to make the boundary smoother. The StyleGAN feature is retained in the facial area, while the background area is from the mixed feature. We mix features of StyleGAN and VG-GFace using SFT. The background loss and the perceptual loss make the swapped image have the same background as the target image:", + "bbox": [ + 75, + 395, + 467, + 561 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {b} + \\varepsilon_ {p} L _ {p} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 571, + 467, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$L_{b}$ is the background loss used to measure the difference in the background between the generated image and the target image. $L_{p}$ represents the perceptual distance between the generated image and the target image. $\\varepsilon_{p}$ is the weight of the loss function. The background loss is expressed as:", + "bbox": [ + 75, + 598, + 467, + 674 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b} = \\left\\| M \\odot \\left(I _ {r} - I _ {t}\\right) \\right\\| \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 684, + 467, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the binarized image of the background region obtained by face segmentation algorithm [2].", + "bbox": [ + 75, + 710, + 467, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 753, + 200, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As mentioned above, StyleIPSB has the following advantages: (1) StyleIPSB can represent various poses, expressions, and illumination properties. (2) When the style code is modified along one base, it only changes the specific attribute while remaining identity unchanged. (3) StyleIPSB can preserve the pore-level details. (4) StyleIPSB cooperates well with 3DMM to control facial attributes. Therefore, in this section, we mainly evaluate", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the effectiveness of our method in the following aspects: (1) Evaluating the properties of StyleIPSB. (2) Evaluating the performance of 3DMM controlling facial attributes with StyleIPSB. (3) Comparison of face swapping results with other methods. First of all, we introduce the dataset and training detail.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Training Detail", + "text_level": 1, + "bbox": [ + 498, + 191, + 748, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We used the FFHQ database [15] to train the 3DMM StyleGAN Mapping module. Flickr-Faces-HQ (FFHQ) consists of 70,000 high-quality face images at $1024 \\times 1024$ resolution. CelebAMask-HQ [20] is a large-scale face image dataset with 30,000 high-resolution face images selected from the CelebA dataset by following CelebA-HQ. FaceForensics++ [32] is a forensics dataset consisting of 1,000 original video sequences.", + "bbox": [ + 496, + 215, + 890, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training environment and hyperparameter We train the network on a GTX 3090 using the Pytorch framework. In training, our optimizer is Adam, the weight decay is 5e-5, and the batch size is 2. In order to effectively train the above three stages, we first pre-train the three stages separately and finally train them together. In the three-stage pre-training, the learning rate is 2e-4; in the final together training stage, we adjust the learning rate to 2e-5.", + "bbox": [ + 496, + 337, + 890, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. The Properties of StyleIPSB", + "text_level": 1, + "bbox": [ + 498, + 468, + 751, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The representation and the disentanglement of StyleIPSB. We conduct experiments on the representation of StyleIPSB to evaluate whether it can represent various poses, expressions, and illumination properties while identity is unchanged. Fig. 4 shows that StyleIPSB can express pitch, yaw rotations, and lighting with different colors, directions, and intensities. In terms of expressions, StyleIPSB can not only control whether the mouth is open or not, the eyes are open or closed but also control different eyeball orientations and raising eyebrows. Supplementary material reveals the effect of modifying coordinates in different base directions.", + "bbox": [ + 496, + 492, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the disentanglement of StyleIPSB, we present the quantitative results, which show the identity change when manipulating the pose. As shown in Fig. 5, the horizontal axis represents the yaw angle between the anchored face and the edited image, and the vertical axis represents identity loss as defined in Eq. (2) between the anchored face and the edited image. Our StyleIPSB outperforms Ganspace [13], which is also a basis constructed in $\\mathrm{W}+$ space using unsupervised learning. We achieve comparable results with the supervised method. InterfaceGAN [34]. It needs a dataset that contains left and right-facing faces. StyleIPSB uses unsupervised learning and can represent more attributes than InterfaceGAN. Fig. 5 also shows that if we introduce identity loss into the distance metric when building StyleIPSB, the identity-preserving perfor", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "357", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1aa025ba8784bf394be246e01680491d6dbe57a2b39ae544324aee6023a8405e.jpg", + "image_caption": [ + "Figure 4. StyleIPSB can modify various poses, expressions, and illumination by adjusting its coordinates while keeping the identity unchanged. The first row is the results of changing coordinates of pose and illumination basis. The second row shows the results of the changing coordinate of the expression basis." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44a43f0c63f5c4a7e169767e3b14210bcd67331ee3f2ad8acb67bd51a0b4df18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 879, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74d3cc1660c4dc68a20388f456e659b6cea66dd5801fa352102f74622f665620.jpg", + "image_caption": [ + "Figure 5. Effects of different methods of changing face pose attributes on face identity information" + ], + "image_footnote": [], + "bbox": [ + 140, + 305, + 405, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c089c26ec461eae45fd799aa97ef7ab692ea01effddfcd2492fe8e31f100346.jpg", + "image_caption": [ + "Figure 6. We compare our method with GANSpace [13] and InterfaceGAN [34] in the editing of facial expression." + ], + "image_footnote": [], + "bbox": [ + 86, + 523, + 459, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance of StyleIPSB is enhanced.", + "bbox": [ + 76, + 762, + 300, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Editing results of StyleIPSB. We compare StyleIPSB with other StyleGAN-based face edit methods. Fig. 6 compares the edit results of expression (smile) with other methods. The results show that our approach has better performance in identity-preserving. StyleIPSB preserves the face shape when expression changes, while face shape is deformed in other methods.", + "bbox": [ + 75, + 777, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 shows examples of editing a face image using the", + "bbox": [ + 96, + 885, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9c8577d075d33d9cfce7f2509e21022b6b6d20d31d6e8b831b7d46354e21428.jpg", + "image_caption": [ + "Figure 7. 3D face models control the attribute of generated images. The top is the rendered images of 3D face geometry. The bottom is the edited images controlled by the 3DMM parameters of the top 3D face geometry." + ], + "image_footnote": [], + "bbox": [ + 503, + 301, + 893, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3DMM parameters of FLAME [23]. The first row is the 3D face models generated by the 3DMM parameters. The second row is the edited image controlled by those 3DMM parameters. These results show that StyleIPSB cooperates well with 3DMM to control the facial attributes of the generated image.", + "bbox": [ + 498, + 502, + 890, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Result of Face Swapping", + "text_level": 1, + "bbox": [ + 500, + 604, + 725, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This subsection evaluates the performance of face swapping based on StyleIPSB. First, the ablation experiments evaluate the influence of StyleIPSB on attribute transfer performance. Then, we compare our face swapping with other methods.", + "bbox": [ + 498, + 627, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation experiment In the experiment, \"no basis\" means that the 3DMM-StyleGAN mapping network directly regresses the $w_{s\\rightarrow t}$ instead of $\\beta_{s\\rightarrow t}$ . As shown in Fig. 8, using StyleIPSB can constrain the generated style code on the subspace of $\\mathrm{W + }$ space and preserve the pore-level details.", + "bbox": [ + 498, + 703, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows the quantitative comparison results on the performance of StyleIPSB in stage one. FID (Fréchet inception distance) is used to evaluate the image quality generated by the image generation model. \"Exp\" and \"Pose\" represent the Euclidean distance between the target image and the transferred image expression and pose parameters, respectively. ID Similarity means the cosine similarity be", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "358", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64c5a4b960f30cf8e918e02bb555b5c6976bb01effbe381d4e1f84908d7ce792.jpg", + "image_caption": [ + "Figure 8. The qualitative ablation experiments on StyleIPSB. The results show we can generate pore-level details with StyleIPSB." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 477, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e45ae4dd88c9f2d377d6407544f009e2b20ac8a3e7d6b6acebc983304e24df0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FID ↓Exp ↓Pose ↓ID similarity ↑
No Basis26.063.730.0740.65
With basis22.153.370.0780.67
", + "bbox": [ + 80, + 215, + 467, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tween the identity embedding of the source image and the transferred image. D3FR [9] is used to extract the 3D face parameters, and CosFace [43] is used to extract the identity embedding. The test set is 10,000 image pairs randomly sampled from the CelebAMask-HQ [20]. The Tab. 2 shows that using StyleIPSB can effectively reduce the value of FID because StyleIPSB is in the subspace of $\\mathrm{W + }$ space. In addition, the performance of transferring expressions with StyleIPSB is better in identity-preserving because the identity is preserved while changing the coordinates of StyleIPSB.", + "bbox": [ + 75, + 337, + 468, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows the quantitative comparison results of different methods on the FaceForensics++ dataset. The dataset is classified into 885 identities, and the identity retrieval indicates the top-1 matching rate of swapped and the source image. We apply CosFace [43] to extract identity embedding. Expressions are measured in the same way as Tab. 2, The pose measurement method uses a pose estimator [33]. We compare our results with FaceSwap [3], DeepFakes [1], FaceShifter [22], MegaFace [51] and FSLDS [47]. As shown in the Tab. 3, our method outperforms in identity-preserving and expression transfer thanks to StyleIPSB's identity-preserving property and the ability to present various expressions. Our performance of pose transfer does not outperform other methods, which may be because StyleIPSB only contains two bases expressing poses in raw and pitch directions, which makes it hard to express very accurate poses. However, the error of the pose is only a few degrees, so our results still look accurate visually.", + "bbox": [ + 75, + 505, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 9 shows the results of face swapping under challenging conditions. The results show that our method can preserve the unique noise patterns in portraiture in the red boxes. In the second row of the figure, our method successfully keeps pore-level details and identity while the source and target pose is quite different. Tab. 4 shows that our method has the better image quality and identity preservation ability than other StyleGAN-based methods. The met-", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b5a1cd0dad706ba3016ed74a373f0330e62689c904edb69639068e1c8fbecdb.jpg", + "table_caption": [ + "Table 2. The quantitative ablation experiments on StyleIPSB." + ], + "table_footnote": [], + "table_body": "
MethodID Retri.(%) ↑Exp Err. ↓Pose Err. ↓
FaceSwap [3]72.692.892.58
Deepfakes [1]88.393.334.64
FaceShifter [22]90.682.822.55
MegaFS [51]90.832.922.64
FSLDS [47]90.052.792.46
Ours95.052.233.58
", + "bbox": [ + 501, + 87, + 924, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2ded8812dfc81488de538a16dfd65610e613944713eff49e9a5fb321e16f224f.jpg", + "table_caption": [ + "Table 3. The quantitative experiments on FaceForensics++ dataset with other methods." + ], + "table_footnote": [], + "table_body": "
FID ↓Exp ↓Pose ↓ID similarity ↑
MageFS [51]22.032.850.0430.4837
RAFS [46]13.253.15-0.5232
FSLDS [47]10.012.990.0530.4761
Ours9.372.750.0780.5378
", + "bbox": [ + 501, + 251, + 906, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The quantitative experiments on the CelebAHQ dataset with other methods.", + "bbox": [ + 500, + 357, + 890, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9e7794e558308f9f9ed6750e05305bc63718632f5b667951318ebcf5e70593e.jpg", + "image_caption": [ + "Figure 9. We compare our method with MageFS [51] and FSLDS [47] in challenging conditions." + ], + "image_footnote": [], + "bbox": [ + 501, + 398, + 898, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07e01d192ddee6770729044c541eadfe73a4f218c9d244cc613347cd446d194b.jpg", + "image_caption": [ + "Figure 10. limitation of our method" + ], + "image_footnote": [], + "bbox": [ + 503, + 565, + 890, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ric is the same as in Tab. 2. We follow RAFS and compare $100\\mathrm{k}$ swapped faces for a fair comparison.", + "bbox": [ + 500, + 695, + 890, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 500, + 739, + 748, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have developed a new semantic basis for face swapping, called StyleIPSB, that is specifically designed to preserve identity and pore-level details. Our experiments have demonstrated that StyleIPSB outperforms other state-of-the-art methods. Despite this, there is still potential for further improvement, as shown in Fig. 10. (1) Occlusion is limited by the mask. (2) The glasses in the source image cannot be removed. (3) Light and shadow cannot be perfectly restored in the case of complex illumination.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "359", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Deepfakes. https://github.com/ondyari/FaceForensics/tree/master/dataset/DeepFakes.8", + "[2] face-parsing.pytorch. https://github.com/zllrunning/face-parsing.PyTorch.6", + "[3] Faceswap. https://github.com/ondyari/FaceForensics/tree/master/dataset/FaceSwapKowalski.8", + "[4] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2StyleGAN: Unsupervised extraction of StyleGAN edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3", + "[5] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (TOG), 40(3):1-21, 2021. 2", + "[6] Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, and Gang Hua. Towards open-set identity preserving face synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6713-6722, 2018. 2", + "[7] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003-2011, 2020. 2, 3", + "[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 4", + "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 8", + "[10] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3", + "[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2", + "[12] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. GIF: Generative interpretable faces. In 2020 International Conference on 3D Vision (3DV), pages 868-878. IEEE, 2020. 2, 3", + "[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANspace: Discovering interpretable GAN controls. arXiv preprint arXiv:2004.02546, 2020. 2, 3, 6, 7", + "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 6", + "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2", + "[17] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2", + "[18] Jiseob Kim, Jihoon Lee, and Byoung-Tak Zhang. Smooth-Swap: A simple enhancement for face-swapping with smoothness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10779-10788, 2022. 2", + "[19] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 3677-3685, 2017. 3", + "[20] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 8", + "[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. FaceInpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2, 3", + "[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 2, 3, 8", + "[23] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7", + "[24] Yuan Lin, Shengjin Wang, Qian Lin, and Feng Tang. Face swapping under large pose variations: A 3d model based approach. In 2012 IEEE International Conference on Multimedia and Expo, pages 333-338. IEEE, 2012. 3", + "[25] Luming Ma and Zhigang Deng. Real-time hierarchical facial performance capture. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 1-10, 2019. 3", + "[26] Luming Ma and Zhigang Deng. Real-time face video swapping from a single portrait. In Symposium on Interactive 3D Graphics and Games, pages 1-10, 2020. 3", + "[27] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 2", + "[28] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. arXiv preprint arXiv:2202.12972, 2022. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "360", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2", + "[30] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In Proceedings of the British Machine Vision Conference (BMVC). British Machine Vision Association, 2015. 6", + "[31] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleClip: Text-driven manipulation of StyleGAN imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2", + "[32] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 6", + "[33] Nataniel Ruiz, Eunji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 8", + "[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 2, 3, 6, 7", + "[35] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 3", + "[36] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3", + "[37] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. SemanticStyleGAN: Learning compositional generative priors for controllable image synthesis and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11254-11264, 2022. 1, 2", + "[38] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3", + "[39] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. StyleRig: Rigging StyleGAN for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2, 3", + "[40] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of RGB videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4", + "[42] Binxu Wang and Carlos R Ponce. The geometry of deep generative image models and its applications. arXiv preprint arXiv:2101.06006, 2021. 4", + "[43] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 8", + "[44] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6", + "[45] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 2, 3", + "[46] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2, 3, 8", + "[47] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7642-7651, 2022. 1, 2, 3, 8", + "[48] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. TransEditor: Transformer-based dual-space GAN for highly controllable facial editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2022. 1", + "[49] Zhiliang Xu, Zhibin Hong, Changxing Ding, Zhen Zhu, Junyu Han, Jingtuo Liu, and Errui Ding. MobileFaceSwap: A lightweight framework for video face swapping. arXiv preprint arXiv:2201.03808, 2022. 3", + "[50] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. StyleHeat: One-shot high-resolution editable talking face generation via pretrained StyleGAN. arXiv preprint arXiv:2203.04036, 2022.1", + "[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2, 3, 8" + ], + "bbox": [ + 501, + 92, + 890, + 796 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "361", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_model.json b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe6878cefdf6147909d6efed45a51d740838e40a --- /dev/null +++ b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.13, + 0.806, + 0.178 + ], + "angle": 0, + "content": "StyleIPSB: Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.203, + 0.713, + 0.241 + ], + "angle": 0, + "content": "Diqiong Jiang \\(^{1}\\), Dan Song \\(^{2*}\\), Ruofeng Tong \\(^{1*}\\), Min Tang \\(^{1}\\) \n\\(^{1}\\)Zhejiang University, China \n\\(^{2}\\)Tianjin University, China" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.261, + 0.891, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.539, + 0.894, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Compared with the existing high-fidelity face swapping methods MageFS [51] and FSLDS [47], our face swapping framework based on the identity-preserving semantic basis (StyleIPSB) can preserve the pore-level detail and source image's identity." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.58, + 0.314, + 0.597 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.612, + 0.472, + 0.854 + ], + "angle": 0, + "content": "Recent researches reveal that StyleGAN can generate highly realistic images, inspiring researchers to use pretrained StyleGAN to generate high-fidelity swapped faces. However, existing methods fail to meet the expectations in two essential aspects of high-fidelity face swapping. Their results are blurry without pore-level details and fail to preserve identity for challenging cases. To overcome the above artifacts, we innovatively construct a series of identity-preserving semantic bases of StyleGAN (called StyleIPSB) in respect of pose, expression, and illumination. Each basis of StyleIPSB controls one specific semantic attribute and disentangles with the others. The StyleIPSB constrains style code in the subspace of \\( W+ \\) space to preserve pore-level details and gives us a novel tool for high-fidelity face swapping, and we propose a three-stage framework for face swapping with StyleIPSB. Firstly, we transform the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.581, + 0.895, + 0.779 + ], + "angle": 0, + "content": "target facial images' attributes to the source image. We learn the mapping from 3D Morphable Model (3DMM) parameters, which capture the prominent semantic variance, to the coordinates of StyleIPSB that show higher identity-preserving and fidelity. Secondly, to transform detailed attributes which 3DMM does not capture, we learn the residual attribute between the reenacted face and the target face. Finally, the face is blended into the background of the target image. Extensive results and comparisons demonstrate that StyleIPSB can effectively preserve identity and pore-level details. The results of face swapping can achieve state-of-the-art performance. We will release our code at https://github.com/a686432/StyleIPSB" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.813, + 0.633, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Facial image manipulation [36, 37, 48, 50] is a task of transforming specific attributes from the source image to the target image while persevering other attributes unchanged. Face swapping is one of the essential parts of facial im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.245, + 0.876 + ], + "angle": 0, + "content": "*Co-Corresponding authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.469, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) The research is supported in part by the National Natural Science Foundation of China (61972342, 61832016, 61972341, 61902277)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.958 + ], + "angle": 0, + "content": "352" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.183 + ], + "angle": 0, + "content": "age manipulation, which has attracted lots of interest in the computer vision and graphics community. Face swapping aims to generate an image with the source image's identity and the target image's attributes (e.g., expression, pose, background, hair, etc.). It has wide applications in the film industry and computer games." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.35 + ], + "angle": 0, + "content": "Current face swapping methods are mainly divided into two categories: 3D model-based methods and 2D image-based methods. 3D model-based methods [11,17,40] firstly reconstruct the 3D face models based on 3DMM from the source image and target image and transfer the nonidentity parameters of the target face model to the source face model. Then they render the transferred 3D model and blend it into the target image. Although such methods can transfer coarse facial attributes such as pose, expression, and illumination, they have difficulty in generating realistic hair and teeth accessories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.549 + ], + "angle": 0, + "content": "With the development of generative adversarial networks, 2D image-based methods [6, 7, 22, 27, 28, 45] can synthesize photo-realistic images. The generated face images have convincing and detailed facial attributes, such as mouth, teeth, and eyebrows. Recent works [21, 46, 47, 51] employ the pre-trained StyleGAN decoder to further improve the fidelity and synthesize pore-level details. However, as shown in Fig. 1, despite using the pre-trained StyleGAN model, their results fail to generate pore-level details and identity-preserving in challenge conditions. Overall, 2D image-based methods generate more realistic images than 3D model-based methods, but the identity-preserving and pore-level details of the images still need improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.468, + 0.717 + ], + "angle": 0, + "content": "To tackle the challenges of blurry images and identity-preserving in face swapping, according to our observations, the cause of the blurred images is that the regressed style code is out of \\(\\mathrm{W + }\\) space. Additionally, as mentioned in [18], identity embedding is a non-smooth space, so finding the identity-preserving optimized direction is challenging. To address these problems, our method constrains the regressed style code with identity-preserving semantic bases of StyleGAN (i.e., the proposed StyleIPSB). StyleIPSB stays within the \\(\\mathrm{W + }\\) space, and the identity is preserved when changing its coordinates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The advantages of the proposed StyleIPSB are summarized as follows: (1) StyleIPSB constitutes a linear space, which is the subspace of the \\(\\mathrm{W + }\\) space of StyleGAN. By ensuring the regressed style code within the \\(\\mathrm{W + }\\) space of StyleGAN, we can more easily generate images with pore-level details. (2) When changing the coordinates of the StyleIPSB, the identity remains preserved as much as possible. (3) StyleIPSB can represent various poses, expressions, and illuminations. To construct the basis that satisfies the above properties, we propose a novel identity-preserving distance metric to find the orthogonal semantic directions, which are further assembled to StyleIPSB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": "StyleIPSB also cooperates well with 3DMM to control facial attributes. StyleRig [39] builds the mapping of the 3DMM parameter space and \\(\\mathrm{W + }\\) space of StyleGAN, which can change the facial attribute of the generated image by the 3DMM parameters. StyleRig only can manipulate images generated by StyleGAN. Pie [38] designs a non-linear optimization problem to edit the real-world image based on StyleRig, but the optimization operation is time-consuming. GIF [12] generates face images by the FLAME [23] parametric control. However, the generated images are easy to contain artifacts and change identity. Other face manipulation methods [4, 29, 37] use the network directly to find the edit direction. Still, without the guidance of 3DMM, they can only generate some basic expressions (e.g., smile) and fail to cover various expressions. In this paper, we propose the StyleGAN-3DMM mapping network, which transforms the semantic information of 3DMM parameters into StyleIPSB coordinates. The StyleGAN-3DMM mapping network reduces the gap between 3DMM and StyleIPSB. It shows that StyleIPSB is very compatible with 3DMM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.455 + ], + "angle": 0, + "content": "In summary, we propose a face swapping framework based on StyleIPSB and achieve state-of-the-art results. The main contributions of this paper lie in the following three aspects:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.466, + 0.892, + 0.542 + ], + "angle": 0, + "content": "- We propose a novel method of establishing identity-preserving semantic bases of StyleGAN called StyleIPSB. The face image, generated by the linear space of StyleIPSB, remains pore-level details and identity-preserving." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.553, + 0.892, + 0.63 + ], + "angle": 0, + "content": "- The proposed StyleGAN-3DMM mapping network serves as the bridge to narrow the gap between 3DMM and StyleIPSB, which can take advantage of the prominent semantic variance of 3DMM and the identity-preserving and high-fidelity of styleIPSB." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.641, + 0.892, + 0.702 + ], + "angle": 0, + "content": "- We propose the face swapping framework based on StyleIPSB and StyleGAN-3DMM mapping network. Extensive results show our method outperforms others in detail-preserving and identity-preserving." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.466, + 0.892, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.645, + 0.732 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.818, + 0.758 + ], + "angle": 0, + "content": "2.1. Image Modification Using StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "StyleGAN [14-16] is a powerful image synthesis model that can generate a wide variety of high-quality face images. Some methods [5, 31, 34, 35] find the particular attribute editing direction in \\(\\mathbf{W}+\\) space using the corresponding attribute classifier. However, their attribute editing is limited by the classifier's ability and cannot control diverse expressions and illumination. To facilitate attribute manipulation in an unsupervised manner, GANSpace [13] and SeFa [36] perform decomposition to find primary directions" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "353" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "in the latent space and explore the interpretable directions among the primary directions. CLIP2StyleGAN [4] discovery and label of StyleGAN edit-direction based on CLIP image space. However, they only find a few types of expressions and illumination and fail to discover a complete set of expressions and illumination. Some works [12,38,39] map the rigging information to face manipulation, but our StyleIPSB can generate more detail and identity-preserving results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.235, + 0.195, + 0.252 + ], + "angle": 0, + "content": "2.2. Face swap" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.576 + ], + "angle": 0, + "content": "In the early years, many works [24-26, 40] apply 3D face models for face swapping. Face2face [40] applies an efficient deformation transfer to track the source video facial expressions and re-render the synthesized faces with retrieved and warped mouth interiors. Ma et al. [25] reconstructed high-resolution facial geometry and appearance by capturing an individual-specific face model with fine-scale details. Those 3D model-based methods are difficult to generate high-fidelity face images, especially realistic hair, mouth, and teeth. In recent years, conditional GAN architecture has been widely used in face-swapping. Many works [7, 19, 21, 22, 28, 46, 49] use neural networks to generate high-fidelity images. SimSwap [7] and FaceShifter [22] use a face recognition network to extract the identity embedding and use a decoder to fuse the identity embedding. Recently, many methods [21, 45-47, 51] use pretrained StyleGAN as the image generator to improve the quality of generated images further. However, the images generated by their method still suffer from the lack of details and face identity shift. Our approach is better at retaining pore-level details and identify-preserving." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.588, + 0.17, + 0.603 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.72 + ], + "angle": 0, + "content": "This section reveals the details of our proposed method in three aspects. The first part describes the process of conducting StyleIPSB. The second part describes how to use StyleIPSB to transfer the facial attributes except for identity from the target image to the source face. The third introduces the overall framework of face swap based on StyleIPSB." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.727, + 0.298, + 0.743 + ], + "angle": 0, + "content": "3.1. StyleIPSB Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "This subsection aims to find a semantic basis in the \\(\\mathrm{W + }\\) space of StyleGAN that satisfies the conditions mentioned in the introduction. In short, StyleIPSB should meet the following properties: a subspace of \\(\\mathrm{W + }\\) space, identity-preserving, and representation ability. The previous methods fail to satisfy the above three conditions. For example, as shown in Tab. 1, GIF [12] does not meet the requirements in the \\(\\mathrm{W + }\\) subspace, which would lead to generated image distortion. Some methods [12,35,36] do not concern the identity preservation and suffer the identity shift when" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.892, + 0.2 + ], + "angle": 0, + "content": "
MethodSubspaceIDRepresentation
InterFaceGAN [34]✓✓
GANSpace [13]✓
SeFa [36]✓
GIF [12]✓
Our✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.21, + 0.894, + 0.253 + ], + "angle": 0, + "content": "Table 1. The conditions of the basis satisfied by different methods. We compare our method with InterFaceGAN [34], GANSpace [13], GIF [12] and SeFa [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.323 + ], + "angle": 0, + "content": "editing. InterFaceGAN [34] cannot edit various expressions and illumination. Unlike existing methods, our method considers all three conditions when we build StyleIPSB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.893, + 0.474 + ], + "angle": 0, + "content": "Compared with the previous methods, which built the basis by directly performing decomposition on eigenvalue in \\(\\mathrm{W + }\\) space, we added semantic metrics and identity loss in our decomposition. Therefore, StyleIPSB is not only in the \\(\\mathrm{W + }\\) subspace but can also represent the complete set of semantic information and have the property of identity preservation. So adding semantic information and maintaining identity when building StyleIPSB is the crucial point to this subsection. The following introduces the detailed process of conducting StyleIPSB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.894, + 0.58 + ], + "angle": 0, + "content": "**Formulation** The StyleGAN network, denoted by \\( G \\), is a mapping from style code \\( w \\) to image \\( I \\), \\( G: R^n \\to R^{H \\times W \\times 3} \\), \\( w \\mapsto I \\). The 3DMM Fitting network, denoted by \\( M \\), regresses the 3DMM parameters \\( p \\) from image \\( I \\). The parameters include pose parameter \\( p_p \\), expression parameter \\( p_e \\) and illumination parameter \\( p_i \\). In this paper, we use DECA [10] as our 3DMM Fitting network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.894, + 0.745 + ], + "angle": 0, + "content": "We first define the distance metric of the pose, expression, and illumination. The distance metrics measure the attribute difference between two images generated by style codes \\( w_{1}, w_{2} \\). In addition, we add identity loss to the metric. Therefore, the distance metric increases if the difference between two images' attributes increases and the identity remains similar. Then we decompose the Hessian matrix to find the direction with the fastest distance metric change in the \\( \\mathrm{W+} \\) space. Therefore, the attributes change fast but identity changes slowly along the direction we found. The distance metrics are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.753, + 0.847, + 0.789 + ], + "angle": 0, + "content": "\\[\nD _ {p} (w _ {1}, w _ {2}) = \\frac {| | M (G (w _ {1})) _ {p} - M (G (w _ {2})) _ {p} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.789, + 0.892, + 0.826 + ], + "angle": 0, + "content": "\\[\nD _ {e} \\left(w _ {1}, w _ {2}\\right) = \\frac {\\left| \\left| M \\left(G \\left(w _ {1}\\right)\\right) _ {e} - M \\left(G \\left(w _ {2}\\right)\\right) _ {e} \\right| \\right| ^ {2}}{L _ {i d} \\left(G \\left(w _ {1}\\right) , G \\left(w _ {2}\\right)\\right)} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.827, + 0.842, + 0.864 + ], + "angle": 0, + "content": "\\[\nD _ {i} (w _ {1}, w _ {2}) = \\frac {| | M (G (w _ {1})) _ {i} - M (G (w _ {2})) _ {i} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "where \\(L_{id}\\) measures the similarity of the identities of two images. Its value represents the similarity of two identity" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "354" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.884, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.44, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Figure 2. Our framework is divided into three stages. Stage one transfers the target image's expression, pose, and illumination attributes to the source image through the semantic base. Stage two transfers detailed non-identity attributes. Stage three blends the transferred image into the background of the target image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.471, + 0.552 + ], + "angle": 0, + "content": "embeddings obtained by the pre-trained face recognition network [8] from two input face images." + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.561, + 0.47, + 0.597 + ], + "angle": 0, + "content": "\\[\nL _ {i d} \\left(I _ {1}, I _ {2}\\right) = 1 - \\frac {I D \\left(I _ {1}\\right) \\cdot I D \\left(I _ {2}\\right)}{\\left| \\left| I D \\left(I _ {1}\\right) \\right| \\right| _ {2} \\left| \\left| I D \\left(I _ {2}\\right) \\right| \\right| _ {2}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Then, we use the Hessian matrix to encode the local distance information on the image manifold up to second-order approximation." + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.66, + 0.47, + 0.68 + ], + "angle": 0, + "content": "\\[\nD ^ {2} (w _ {0}, w _ {0} + \\delta w) \\approx | | \\delta w | | _ {H} ^ {2} = \\delta_ {w} ^ {T} H (w _ {0}) \\delta w \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.78 + ], + "angle": 0, + "content": "The \\( w_0 \\) used to construct the semantic base is randomly sampled, and \\( w_0 + \\delta w \\) is the point near \\( w_0 \\). We decompose the Hessian matrix \\( H(w_0) = \\sum_i \\lambda_i v_i v_i^T \\) and find the first \\( m \\) principal components of \\( v \\) as our basis \\( V \\). Therefore, the basis \\( V \\) contains the direction vectors, along which the distance metric changes first \\( m \\) fastest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We use the algorithm [42] to calculate and decompose the Hessian matrix. StyleGAN has 18 different levels, and the style code of each level itself contains certain semantic features (for example, the low level includes the features of face shape and pose, and the high level has the skin color and illumination, etc.). We construct the semantic base of the pose on the first three levels of the style code and the expression on the 4th to 10th levels. The semantic base of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.893, + 0.629 + ], + "angle": 0, + "content": "the illumination is on the levels after the 10th. We sampled \\( w_{0} \\) 100 times and generated 100 sets of bases. And then, we average these bases and performed Schmidt orthogonalization to obtain the semantic base of the pose \\( V_{p} \\), expression \\( V_{e} \\), and illumination \\( V_{i} \\), respectively. Finally, the pose, expression, and illumination base are combined as StyleIPSB \\( V = [V_{p}, V_{e}, V_{i}] \\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.638, + 0.753, + 0.656 + ], + "angle": 0, + "content": "3.2. 3DMM-StyleGAN Mapping" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.662, + 0.892, + 0.784 + ], + "angle": 0, + "content": "As shown in Fig. 2, the source and target images are fed into the 3DMM fitting network to regress the 3DMM parameters \\(\\alpha^s, \\alpha^t\\). The 3DMM-StyleGAN Mapping network, which contains three six-layer multilayer perceptrons(MLPs), maps 3DMM parameters \\(\\alpha\\) to StyleIPSB coordinates of pose \\(\\beta_{s \\rightarrow t}^p\\), expression \\(\\beta_{s \\rightarrow t}^e\\) and illumination \\(\\beta_{s \\rightarrow t}^i\\), respectively. The following equations calculate the transferred style code \\(w_{s \\rightarrow t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.797, + 0.892, + 0.813 + ], + "angle": 0, + "content": "\\[\nw _ {s \\rightarrow t} = w _ {s} + \\beta_ {s \\rightarrow t} V \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\beta_{s\\rightarrow t} = [\\beta_{s\\rightarrow t}^p,\\beta_{s\\rightarrow t}^e,\\beta_{s\\rightarrow t}^i ]\\) and \\(w_{s}\\) is the corresponding style code of the source image, which is obtained from the source image by the StyleGAN encoder [41]. Finally, \\(w_{s\\rightarrow t}\\) passes through StyleGAN to get the attribute transferred image." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "355" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.573, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.092, + 0.916, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.893, + 0.325 + ], + "angle": 0, + "content": "Figure 3. The network structure of the image blend. Stage three blends transformed image to the target image by mixing the VGG encoder feature with the StyleGAN feature through masked spatial feature transforms (MSFT)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.483 + ], + "angle": 0, + "content": "3DMM-StyleGAN mapping network aims at precisely controlling the generated attribute through the 3DMM parameters while keeping the identity unchanged. Therefore, we design the attribute loss \\( L_{attr} \\) to make the generated face image have the same attribute as the pose, expression, and illumination as target 3DMM parameters. And we add identity loss \\( L_{id} \\) to preserve the identity of the source image. The total loss is shown as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.505, + 0.469, + 0.52 + ], + "angle": 0, + "content": "\\[\nL = L _ {i d} + \\varepsilon_ {\\text {a t t r}} L _ {\\text {a t t r}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.636 + ], + "angle": 0, + "content": "where \\( L_{id} \\) measure the identity similarity of two images and defined in Eq. (2). \\( \\varepsilon_{attr} \\) is the weight of two loss functions. \\( L_{attr} \\) measures the difference between the attributes of the transferred face and the target face. We reconstruct the 3D face from the transferred face and the target face, and then compare their difference in 3D face geometries and rendered images:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.65, + 0.469, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {a t t r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) = L _ {g e o} \\left(\\alpha^ {s}, \\alpha^ {s \\rightarrow t}\\right) (6) \\\\ + L _ {r e n d e r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.469, + 0.747 + ], + "angle": 0, + "content": "where \\(\\alpha^s, \\alpha^t\\) is the 3DMM parameters of the source and target images. And \\(\\alpha^{s\\rightarrow t}\\) is 3DMM parameters of the transferred image where \\(\\alpha^{s\\rightarrow t} = M(G(w_{s\\rightarrow t}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.883 + ], + "angle": 0, + "content": "We define that \\( G_{3DMM}(\\alpha_s, \\alpha_e, \\alpha_p) \\) can generate 3D geometry from 3DMM parameters, and \\( R(\\alpha_s, \\alpha_e, \\alpha_a, \\alpha_i, \\alpha_p) \\) can generate rendered images from 3DMM parameters, where \\( \\alpha_s, \\alpha_e, \\alpha_a, \\alpha_i, \\alpha_p \\) represent shape parameters, pose parameters, expression parameters, albedo parameters, and lighting parameters, respectively. Because we only compare attribute differences between transferred and target images, we reconstruct the 3D face by the same 3DMM shape parameters." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The geometric term \\( L_{geo} \\) uses the \\( L_{2} \\) loss between two" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.589, + 0.377 + ], + "angle": 0, + "content": "face meshes:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.383, + 0.627, + 0.4 + ], + "angle": 0, + "content": "\\[\nL _ {g e o} \\left(\\alpha^ {s}, \\alpha^ {s \\rightarrow t}\\right) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.403, + 0.901, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N} \\left\\| G _ {3 D M M} \\left(\\alpha_ {s} ^ {s}, \\alpha_ {e} ^ {t}, \\alpha_ {p} ^ {t}\\right) - G _ {3 D M M} \\left(\\alpha_ {s} ^ {s \\rightarrow t}, \\alpha_ {e} ^ {s \\rightarrow t}, \\alpha_ {p} ^ {s \\rightarrow t}\\right)\\right\\| _ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.444, + 0.843, + 0.458 + ], + "angle": 0, + "content": "where \\(N\\) is the number of vertices of the face mesh." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.489 + ], + "angle": 0, + "content": "The render term \\( L_{\\text{render}} \\) uses the \\( L_1 \\) loss between two rendered images:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.496, + 0.671, + 0.513 + ], + "angle": 0, + "content": "\\[\nL _ {r e n d e r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.516, + 0.907, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\left\\| R \\left(\\alpha_ {s} ^ {s}, \\alpha_ {e} ^ {t}, \\alpha_ {a} ^ {t}, \\alpha_ {i} ^ {t}, \\alpha_ {p} ^ {t}\\right) - R \\left(\\alpha_ {s} ^ {s \\rightarrow t}, \\alpha_ {e} ^ {s \\rightarrow t}, \\alpha_ {a} ^ {t}, \\alpha_ {i} ^ {s \\rightarrow t}, \\alpha_ {p} ^ {s \\rightarrow t}\\right)\\right\\| _ {1} \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.802, + 0.571 + ], + "angle": 0, + "content": "3.3. Detailed Attribute Transformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.654 + ], + "angle": 0, + "content": "In the previous subsection, we transfer the attributes of the target face to the source face, including pose, expression, and illumination. In this subsection, to transfer attributes beyond the 3DMM expressive capabilities, we use DAT-MLP to transfer more detailed attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.745 + ], + "angle": 0, + "content": "First, we project the target images into the \\(W+\\) space using the StyleGAN encoder to obtain the style code \\(w_{t}\\). Then, DAT-MLP regresses \\(\\lambda\\) from style codes \\(w_{t}, w_{s \\to t}\\) using six-layer MLP. Finally, we use the following formula to get the latent vector \\(w_{s \\to t}^{\\prime}\\) that contains non-identity attributes of the target image:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.753, + 0.891, + 0.771 + ], + "angle": 0, + "content": "\\[\nw _ {s \\rightarrow t} ^ {\\prime} = \\lambda w _ {s \\rightarrow t} + (1 - \\lambda) w _ {t} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.892, + 0.838 + ], + "angle": 0, + "content": "where \\(\\lambda \\in R^{18\\times 512}\\) is constrained to be between 0 and 1 through the sigmoid activation layer. We want to transfer attributes of \\(w_{t}\\) into \\(w_{s\\rightarrow t}^{\\prime}\\) as many as possible. The loss function of the second stage is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.846, + 0.891, + 0.864 + ], + "angle": 0, + "content": "\\[\nL = L _ {i d} + \\varepsilon_ {p} L _ {p} + \\varepsilon_ {\\text {a t t r}} L _ {\\text {a t t r}} \\left(\\alpha^ {s}, \\alpha^ {t}, M \\left(G \\left(w _ {s \\rightarrow t} ^ {\\prime}\\right)\\right)\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The loss function is divided into three items. The first item is face recognition loss, which is used to constrain" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "the identity of the generated face to be consistent with the source image. The second item \\( L_{p} \\) represents the perceptual distance between the generated image and the target image. It can help transfer residual attributes other than pose, expression, and illumination. The third term \\( L_{attr} \\) constrains the attributes of the generated face consistent with the target image. \\( \\varepsilon_{p}, \\varepsilon_{attr} \\) represent the weight of the loss function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.288 + ], + "angle": 0, + "content": "The Eq. (9) shows that the style code \\( w_{s \\to t}^{\\prime} \\) is in a bounding box of two style codes \\( w_{s \\to t} \\) and \\( w_{t} \\). We consider that most of the bounding box is still in the subspace of \\( \\mathbf{W}+ \\) space because the attribute of two style codes \\( w_{s \\to t} \\) and \\( w_{t} \\) is very close and generated image retains pore-level details in practice." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.297, + 0.212, + 0.312 + ], + "angle": 0, + "content": "3.4. Blend Image" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.395 + ], + "angle": 0, + "content": "This subsection introduces how to blend the generated image into the target image. We propose Masked Spatial Feature Transform (MSFT) module to fuse the feature. Unlike traditional Spatial Feature Transform (SFT) [44], MSFT only fuses the feature of the masked regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.468, + 0.562 + ], + "angle": 0, + "content": "The detail of this module is shown in Fig. 3. The VG-GFace network [30] extracts various level features of the target image, which is injected into the StyleGAN through the MSFT module. We use Gaussian filtering to filter the masked image obtained by the face segmentation algorithm to make the boundary smoother. The StyleGAN feature is retained in the facial area, while the background area is from the mixed feature. We mix features of StyleGAN and VG-GFace using SFT. The background loss and the perceptual loss make the swapped image have the same background as the target image:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.573, + 0.468, + 0.589 + ], + "angle": 0, + "content": "\\[\nL = L _ {b} + \\varepsilon_ {p} L _ {p} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.675 + ], + "angle": 0, + "content": "\\(L_{b}\\) is the background loss used to measure the difference in the background between the generated image and the target image. \\(L_{p}\\) represents the perceptual distance between the generated image and the target image. \\(\\varepsilon_{p}\\) is the weight of the loss function. The background loss is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.685, + 0.468, + 0.701 + ], + "angle": 0, + "content": "\\[\nL _ {b} = \\left\\| M \\odot \\left(I _ {r} - I _ {t}\\right) \\right\\| \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.468, + 0.743 + ], + "angle": 0, + "content": "where \\(M\\) is the binarized image of the background region obtained by face segmentation algorithm [2]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.202, + 0.772 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As mentioned above, StyleIPSB has the following advantages: (1) StyleIPSB can represent various poses, expressions, and illumination properties. (2) When the style code is modified along one base, it only changes the specific attribute while remaining identity unchanged. (3) StyleIPSB can preserve the pore-level details. (4) StyleIPSB cooperates well with 3DMM to control facial attributes. Therefore, in this section, we mainly evaluate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "the effectiveness of our method in the following aspects: (1) Evaluating the properties of StyleIPSB. (2) Evaluating the performance of 3DMM controlling facial attributes with StyleIPSB. (3) Comparison of face swapping results with other methods. First of all, we introduce the dataset and training detail." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.193, + 0.749, + 0.209 + ], + "angle": 0, + "content": "4.1. Datasets and Training Detail" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Dataset. We used the FFHQ database [15] to train the 3DMM StyleGAN Mapping module. Flickr-Faces-HQ (FFHQ) consists of 70,000 high-quality face images at \\(1024 \\times 1024\\) resolution. CelebAMask-HQ [20] is a large-scale face image dataset with 30,000 high-resolution face images selected from the CelebA dataset by following CelebA-HQ. FaceForensics++ [32] is a forensics dataset consisting of 1,000 original video sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Training environment and hyperparameter We train the network on a GTX 3090 using the Pytorch framework. In training, our optimizer is Adam, the weight decay is 5e-5, and the batch size is 2. In order to effectively train the above three stages, we first pre-train the three stages separately and finally train them together. In the three-stage pre-training, the learning rate is 2e-4; in the final together training stage, we adjust the learning rate to 2e-5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.469, + 0.753, + 0.485 + ], + "angle": 0, + "content": "4.2. The Properties of StyleIPSB" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.673 + ], + "angle": 0, + "content": "The representation and the disentanglement of StyleIPSB. We conduct experiments on the representation of StyleIPSB to evaluate whether it can represent various poses, expressions, and illumination properties while identity is unchanged. Fig. 4 shows that StyleIPSB can express pitch, yaw rotations, and lighting with different colors, directions, and intensities. In terms of expressions, StyleIPSB can not only control whether the mouth is open or not, the eyes are open or closed but also control different eyeball orientations and raising eyebrows. Supplementary material reveals the effect of modifying coordinates in different base directions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate the disentanglement of StyleIPSB, we present the quantitative results, which show the identity change when manipulating the pose. As shown in Fig. 5, the horizontal axis represents the yaw angle between the anchored face and the edited image, and the vertical axis represents identity loss as defined in Eq. (2) between the anchored face and the edited image. Our StyleIPSB outperforms Ganspace [13], which is also a basis constructed in \\(\\mathrm{W}+\\) space using unsupervised learning. We achieve comparable results with the supervised method. InterfaceGAN [34]. It needs a dataset that contains left and right-facing faces. StyleIPSB uses unsupervised learning and can represent more attributes than InterfaceGAN. Fig. 5 also shows that if we introduce identity loss into the distance metric when building StyleIPSB, the identity-preserving perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.88, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.893, + 0.279 + ], + "angle": 0, + "content": "Figure 4. StyleIPSB can modify various poses, expressions, and illumination by adjusting its coordinates while keeping the identity unchanged. The first row is the results of changing coordinates of pose and illumination basis. The second row shows the results of the changing coordinate of the expression basis." + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.306, + 0.406, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.48, + 0.471, + 0.509 + ], + "angle": 0, + "content": "Figure 5. Effects of different methods of changing face pose attributes on face identity information" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.524, + 0.46, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.706, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Figure 6. We compare our method with GANSpace [13] and InterfaceGAN [34] in the editing of facial expression." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.763, + 0.301, + 0.779 + ], + "angle": 0, + "content": "mance of StyleIPSB is enhanced." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Editing results of StyleIPSB. We compare StyleIPSB with other StyleGAN-based face edit methods. Fig. 6 compares the edit results of expression (smile) with other methods. The results show that our approach has better performance in identity-preserving. StyleIPSB preserves the face shape when expression changes, while face shape is deformed in other methods." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Fig. 7 shows examples of editing a face image using the" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.303, + 0.895, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.42, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Figure 7. 3D face models control the attribute of generated images. The top is the rendered images of 3D face geometry. The bottom is the edited images controlled by the 3DMM parameters of the top 3D face geometry." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.503, + 0.892, + 0.595 + ], + "angle": 0, + "content": "3DMM parameters of FLAME [23]. The first row is the 3D face models generated by the 3DMM parameters. The second row is the edited image controlled by those 3DMM parameters. These results show that StyleIPSB cooperates well with 3DMM to control the facial attributes of the generated image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.727, + 0.623 + ], + "angle": 0, + "content": "4.3. Result of Face Swapping" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.703 + ], + "angle": 0, + "content": "This subsection evaluates the performance of face swapping based on StyleIPSB. First, the ablation experiments evaluate the influence of StyleIPSB on attribute transfer performance. Then, we compare our face swapping with other methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Ablation experiment In the experiment, \"no basis\" means that the 3DMM-StyleGAN mapping network directly regresses the \\( w_{s\\rightarrow t} \\) instead of \\( \\beta_{s\\rightarrow t} \\). As shown in Fig. 8, using StyleIPSB can constrain the generated style code on the subspace of \\( \\mathrm{W + } \\) space and preserve the pore-level details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Tab. 2 shows the quantitative comparison results on the performance of StyleIPSB in stage one. FID (Fréchet inception distance) is used to evaluate the image quality generated by the image generation model. \"Exp\" and \"Pose\" represent the Euclidean distance between the target image and the transferred image expression and pose parameters, respectively. ID Similarity means the cosine similarity be" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.478, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.17, + 0.47, + 0.199 + ], + "angle": 0, + "content": "Figure 8. The qualitative ablation experiments on StyleIPSB. The results show we can generate pore-level details with StyleIPSB." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.217, + 0.468, + 0.282 + ], + "angle": 0, + "content": "
FID ↓Exp ↓Pose ↓ID similarity ↑
No Basis26.063.730.0740.65
With basis22.153.370.0780.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.293, + 0.455, + 0.307 + ], + "angle": 0, + "content": "Table 2. The quantitative ablation experiments on StyleIPSB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.504 + ], + "angle": 0, + "content": "tween the identity embedding of the source image and the transferred image. D3FR [9] is used to extract the 3D face parameters, and CosFace [43] is used to extract the identity embedding. The test set is 10,000 image pairs randomly sampled from the CelebAMask-HQ [20]. The Tab. 2 shows that using StyleIPSB can effectively reduce the value of FID because StyleIPSB is in the subspace of \\(\\mathrm{W + }\\) space. In addition, the performance of transferring expressions with StyleIPSB is better in identity-preserving because the identity is preserved while changing the coordinates of StyleIPSB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Tab. 3 shows the quantitative comparison results of different methods on the FaceForensics++ dataset. The dataset is classified into 885 identities, and the identity retrieval indicates the top-1 matching rate of swapped and the source image. We apply CosFace [43] to extract identity embedding. Expressions are measured in the same way as Tab. 2, The pose measurement method uses a pose estimator [33]. We compare our results with FaceSwap [3], DeepFakes [1], FaceShifter [22], MegaFace [51] and FSLDS [47]. As shown in the Tab. 3, our method outperforms in identity-preserving and expression transfer thanks to StyleIPSB's identity-preserving property and the ability to present various expressions. Our performance of pose transfer does not outperform other methods, which may be because StyleIPSB only contains two bases expressing poses in raw and pitch directions, which makes it hard to express very accurate poses. However, the error of the pose is only a few degrees, so our results still look accurate visually." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Fig. 9 shows the results of face swapping under challenging conditions. The results show that our method can preserve the unique noise patterns in portraiture in the red boxes. In the second row of the figure, our method successfully keeps pore-level details and identity while the source and target pose is quite different. Tab. 4 shows that our method has the better image quality and identity preservation ability than other StyleGAN-based methods. The met-" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.088, + 0.925, + 0.199 + ], + "angle": 0, + "content": "
MethodID Retri.(%) ↑Exp Err. ↓Pose Err. ↓
FaceSwap [3]72.692.892.58
Deepfakes [1]88.393.334.64
FaceShifter [22]90.682.822.55
MegaFS [51]90.832.922.64
FSLDS [47]90.052.792.46
Ours95.052.233.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.209, + 0.892, + 0.236 + ], + "angle": 0, + "content": "Table 3. The quantitative experiments on FaceForensics++ dataset with other methods." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.252, + 0.907, + 0.348 + ], + "angle": 0, + "content": "
FID ↓Exp ↓Pose ↓ID similarity ↑
MageFS [51]22.032.850.0430.4837
RAFS [46]13.253.15-0.5232
FSLDS [47]10.012.990.0530.4761
Ours9.372.750.0780.5378
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.358, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Table 4. The quantitative experiments on the CelebAHQ dataset with other methods." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.4, + 0.899, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.523, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Figure 9. We compare our method with MageFS [51] and FSLDS [47] in challenging conditions." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.566, + 0.891, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.655, + 0.804, + 0.669 + ], + "angle": 0, + "content": "Figure 10. limitation of our method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.696, + 0.891, + 0.727 + ], + "angle": 0, + "content": "ric is the same as in Tab. 2. We follow RAFS and compare \\(100\\mathrm{k}\\) swapped faces for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.749, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We have developed a new semantic basis for face swapping, called StyleIPSB, that is specifically designed to preserve identity and pore-level details. Our experiments have demonstrated that StyleIPSB outperforms other state-of-the-art methods. Despite this, there is still potential for further improvement, as shown in Fig. 10. (1) Occlusion is limited by the mask. (2) The glasses in the source image cannot be removed. (3) Light and shadow cannot be perfectly restored in the case of complex illumination." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "359" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Deepfakes. https://github.com/ondyari/FaceForensics/tree/master/dataset/DeepFakes.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.185 + ], + "angle": 0, + "content": "[2] face-parsing.pytorch. https://github.com/zllrunning/face-parsing.PyTorch.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.227 + ], + "angle": 0, + "content": "[3] Faceswap. https://github.com/ondyari/FaceForensics/tree/master/dataset/FaceSwapKowalski.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.468, + 0.284 + ], + "angle": 0, + "content": "[4] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2StyleGAN: Unsupervised extraction of StyleGAN edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[5] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (TOG), 40(3):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.468, + 0.41 + ], + "angle": 0, + "content": "[6] Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, and Gang Hua. Towards open-set identity preserving face synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6713-6722, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.412, + 0.468, + 0.466 + ], + "angle": 0, + "content": "[7] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003-2011, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.468, + 0.618 + ], + "angle": 0, + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[10] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[12] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. GIF: Generative interpretable faces. In 2020 International Conference on 3D Vision (3DV), pages 868-878. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANspace: Discovering interpretable GAN controls. arXiv preprint arXiv:2004.02546, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[17] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[18] Jiseob Kim, Jihoon Lee, and Byoung-Tak Zhang. Smooth-Swap: A simple enhancement for face-swapping with smoothness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10779-10788, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 3677-3685, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. FaceInpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[23] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[24] Yuan Lin, Shengjin Wang, Qian Lin, and Feng Tang. Face swapping under large pose variations: A 3d model based approach. In 2012 IEEE International Conference on Multimedia and Expo, pages 333-338. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Luming Ma and Zhigang Deng. Real-time hierarchical facial performance capture. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 1-10, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[26] Luming Ma and Zhigang Deng. Real-time face video swapping from a single portrait. In Symposium on Interactive 3D Graphics and Games, pages 1-10, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[27] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. arXiv preprint arXiv:2202.12972, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "360" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[29] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[30] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In Proceedings of the British Machine Vision Conference (BMVC). British Machine Vision Association, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[31] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleClip: Text-driven manipulation of StyleGAN imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[32] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Eunji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[35] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.608 + ], + "angle": 0, + "content": "[36] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[37] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. SemanticStyleGAN: Learning compositional generative priors for controllable image synthesis and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11254-11264, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[38] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[39] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. StyleRig: Rigging StyleGAN for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of RGB videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Binxu Wang and Carlos R Ponce. The geometry of deep generative image models and its applications. arXiv preprint arXiv:2101.06006, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[43] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[44] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[45] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[46] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.46, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[47] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7642-7651, 2022. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[48] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. TransEditor: Transformer-based dual-space GAN for highly controllable facial editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.614, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[49] Zhiliang Xu, Zhibin Hong, Changxing Ding, Zhen Zhu, Junyu Han, Jingtuo Liu, and Errui Ding. MobileFaceSwap: A lightweight framework for video face swapping. arXiv preprint arXiv:2201.03808, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.671, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[50] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. StyleHeat: One-shot high-resolution editable talking face generation via pretrained StyleGAN. arXiv preprint arXiv:2203.04036, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "361" + } + ] +] \ No newline at end of file diff --git a/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_origin.pdf b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13cd1282a5ba30f4acb91fc302ab751efbde1f97 --- /dev/null +++ b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/9436e560-374c-40b1-8fb0-26b0a7182b10_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:885b7ca18d83b766efefe7e05b5ae56a785b8413bd078cdfa5fa30ac30d522bd +size 1551853 diff --git a/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/full.md b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..74940d6c846ee4607a8b56c66bfcdf5f66d0ab3d --- /dev/null +++ b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/full.md @@ -0,0 +1,335 @@ +# StyleIPSB: Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping + +Diqiong Jiang $^{1}$ , Dan Song $^{2*}$ , Ruofeng Tong $^{1*}$ , Min Tang $^{1}$ $^{1}$ Zhejiang University, China + $^{2}$ Tianjin University, China + +![](images/3722d226273fccf98276d40e8e3aa3a301103da5b224495b7c844d434c1282f4.jpg) +Figure 1. Compared with the existing high-fidelity face swapping methods MageFS [51] and FSLDS [47], our face swapping framework based on the identity-preserving semantic basis (StyleIPSB) can preserve the pore-level detail and source image's identity. + +# Abstract + +Recent researches reveal that StyleGAN can generate highly realistic images, inspiring researchers to use pretrained StyleGAN to generate high-fidelity swapped faces. However, existing methods fail to meet the expectations in two essential aspects of high-fidelity face swapping. Their results are blurry without pore-level details and fail to preserve identity for challenging cases. To overcome the above artifacts, we innovatively construct a series of identity-preserving semantic bases of StyleGAN (called StyleIPSB) in respect of pose, expression, and illumination. Each basis of StyleIPSB controls one specific semantic attribute and disentangles with the others. The StyleIPSB constrains style code in the subspace of $W+$ space to preserve pore-level details and gives us a novel tool for high-fidelity face swapping, and we propose a three-stage framework for face swapping with StyleIPSB. Firstly, we transform the + +target facial images' attributes to the source image. We learn the mapping from 3D Morphable Model (3DMM) parameters, which capture the prominent semantic variance, to the coordinates of StyleIPSB that show higher identity-preserving and fidelity. Secondly, to transform detailed attributes which 3DMM does not capture, we learn the residual attribute between the reenacted face and the target face. Finally, the face is blended into the background of the target image. Extensive results and comparisons demonstrate that StyleIPSB can effectively preserve identity and pore-level details. The results of face swapping can achieve state-of-the-art performance. We will release our code at https://github.com/a686432/StyleIPSB + +# 1. Introduction + +Facial image manipulation [36, 37, 48, 50] is a task of transforming specific attributes from the source image to the target image while persevering other attributes unchanged. Face swapping is one of the essential parts of facial im + +age manipulation, which has attracted lots of interest in the computer vision and graphics community. Face swapping aims to generate an image with the source image's identity and the target image's attributes (e.g., expression, pose, background, hair, etc.). It has wide applications in the film industry and computer games. + +Current face swapping methods are mainly divided into two categories: 3D model-based methods and 2D image-based methods. 3D model-based methods [11,17,40] firstly reconstruct the 3D face models based on 3DMM from the source image and target image and transfer the nonidentity parameters of the target face model to the source face model. Then they render the transferred 3D model and blend it into the target image. Although such methods can transfer coarse facial attributes such as pose, expression, and illumination, they have difficulty in generating realistic hair and teeth accessories. + +With the development of generative adversarial networks, 2D image-based methods [6, 7, 22, 27, 28, 45] can synthesize photo-realistic images. The generated face images have convincing and detailed facial attributes, such as mouth, teeth, and eyebrows. Recent works [21, 46, 47, 51] employ the pre-trained StyleGAN decoder to further improve the fidelity and synthesize pore-level details. However, as shown in Fig. 1, despite using the pre-trained StyleGAN model, their results fail to generate pore-level details and identity-preserving in challenge conditions. Overall, 2D image-based methods generate more realistic images than 3D model-based methods, but the identity-preserving and pore-level details of the images still need improvement. + +To tackle the challenges of blurry images and identity-preserving in face swapping, according to our observations, the cause of the blurred images is that the regressed style code is out of $\mathrm{W + }$ space. Additionally, as mentioned in [18], identity embedding is a non-smooth space, so finding the identity-preserving optimized direction is challenging. To address these problems, our method constrains the regressed style code with identity-preserving semantic bases of StyleGAN (i.e., the proposed StyleIPSB). StyleIPSB stays within the $\mathrm{W + }$ space, and the identity is preserved when changing its coordinates. + +The advantages of the proposed StyleIPSB are summarized as follows: (1) StyleIPSB constitutes a linear space, which is the subspace of the $\mathrm{W + }$ space of StyleGAN. By ensuring the regressed style code within the $\mathrm{W + }$ space of StyleGAN, we can more easily generate images with pore-level details. (2) When changing the coordinates of the StyleIPSB, the identity remains preserved as much as possible. (3) StyleIPSB can represent various poses, expressions, and illuminations. To construct the basis that satisfies the above properties, we propose a novel identity-preserving distance metric to find the orthogonal semantic directions, which are further assembled to StyleIPSB. + +StyleIPSB also cooperates well with 3DMM to control facial attributes. StyleRig [39] builds the mapping of the 3DMM parameter space and $\mathrm{W + }$ space of StyleGAN, which can change the facial attribute of the generated image by the 3DMM parameters. StyleRig only can manipulate images generated by StyleGAN. Pie [38] designs a non-linear optimization problem to edit the real-world image based on StyleRig, but the optimization operation is time-consuming. GIF [12] generates face images by the FLAME [23] parametric control. However, the generated images are easy to contain artifacts and change identity. Other face manipulation methods [4, 29, 37] use the network directly to find the edit direction. Still, without the guidance of 3DMM, they can only generate some basic expressions (e.g., smile) and fail to cover various expressions. In this paper, we propose the StyleGAN-3DMM mapping network, which transforms the semantic information of 3DMM parameters into StyleIPSB coordinates. The StyleGAN-3DMM mapping network reduces the gap between 3DMM and StyleIPSB. It shows that StyleIPSB is very compatible with 3DMM. + +In summary, we propose a face swapping framework based on StyleIPSB and achieve state-of-the-art results. The main contributions of this paper lie in the following three aspects: + +- We propose a novel method of establishing identity-preserving semantic bases of StyleGAN called StyleIPSB. The face image, generated by the linear space of StyleIPSB, remains pore-level details and identity-preserving. +- The proposed StyleGAN-3DMM mapping network serves as the bridge to narrow the gap between 3DMM and StyleIPSB, which can take advantage of the prominent semantic variance of 3DMM and the identity-preserving and high-fidelity of styleIPSB. +- We propose the face swapping framework based on StyleIPSB and StyleGAN-3DMM mapping network. Extensive results show our method outperforms others in detail-preserving and identity-preserving. + +# 2. Related works + +# 2.1. Image Modification Using StyleGAN + +StyleGAN [14-16] is a powerful image synthesis model that can generate a wide variety of high-quality face images. Some methods [5, 31, 34, 35] find the particular attribute editing direction in $\mathbf{W}+$ space using the corresponding attribute classifier. However, their attribute editing is limited by the classifier's ability and cannot control diverse expressions and illumination. To facilitate attribute manipulation in an unsupervised manner, GANSpace [13] and SeFa [36] perform decomposition to find primary directions + +in the latent space and explore the interpretable directions among the primary directions. CLIP2StyleGAN [4] discovery and label of StyleGAN edit-direction based on CLIP image space. However, they only find a few types of expressions and illumination and fail to discover a complete set of expressions and illumination. Some works [12,38,39] map the rigging information to face manipulation, but our StyleIPSB can generate more detail and identity-preserving results. + +# 2.2. Face swap + +In the early years, many works [24-26, 40] apply 3D face models for face swapping. Face2face [40] applies an efficient deformation transfer to track the source video facial expressions and re-render the synthesized faces with retrieved and warped mouth interiors. Ma et al. [25] reconstructed high-resolution facial geometry and appearance by capturing an individual-specific face model with fine-scale details. Those 3D model-based methods are difficult to generate high-fidelity face images, especially realistic hair, mouth, and teeth. In recent years, conditional GAN architecture has been widely used in face-swapping. Many works [7, 19, 21, 22, 28, 46, 49] use neural networks to generate high-fidelity images. SimSwap [7] and FaceShifter [22] use a face recognition network to extract the identity embedding and use a decoder to fuse the identity embedding. Recently, many methods [21, 45-47, 51] use pretrained StyleGAN as the image generator to improve the quality of generated images further. However, the images generated by their method still suffer from the lack of details and face identity shift. Our approach is better at retaining pore-level details and identify-preserving. + +# 3. Method + +This section reveals the details of our proposed method in three aspects. The first part describes the process of conducting StyleIPSB. The second part describes how to use StyleIPSB to transfer the facial attributes except for identity from the target image to the source face. The third introduces the overall framework of face swap based on StyleIPSB. + +# 3.1. StyleIPSB Construction + +This subsection aims to find a semantic basis in the $\mathrm{W + }$ space of StyleGAN that satisfies the conditions mentioned in the introduction. In short, StyleIPSB should meet the following properties: a subspace of $\mathrm{W + }$ space, identity-preserving, and representation ability. The previous methods fail to satisfy the above three conditions. For example, as shown in Tab. 1, GIF [12] does not meet the requirements in the $\mathrm{W + }$ subspace, which would lead to generated image distortion. Some methods [12,35,36] do not concern the identity preservation and suffer the identity shift when + +
MethodSubspaceIDRepresentation
InterFaceGAN [34]✓✓
GANSpace [13]✓
SeFa [36]✓
GIF [12]✓
Our✓✓✓
+ +Table 1. The conditions of the basis satisfied by different methods. We compare our method with InterFaceGAN [34], GANSpace [13], GIF [12] and SeFa [36]. + +editing. InterFaceGAN [34] cannot edit various expressions and illumination. Unlike existing methods, our method considers all three conditions when we build StyleIPSB. + +Compared with the previous methods, which built the basis by directly performing decomposition on eigenvalue in $\mathrm{W + }$ space, we added semantic metrics and identity loss in our decomposition. Therefore, StyleIPSB is not only in the $\mathrm{W + }$ subspace but can also represent the complete set of semantic information and have the property of identity preservation. So adding semantic information and maintaining identity when building StyleIPSB is the crucial point to this subsection. The following introduces the detailed process of conducting StyleIPSB. + +**Formulation** The StyleGAN network, denoted by $G$ , is a mapping from style code $w$ to image $I$ , $G: R^n \to R^{H \times W \times 3}$ , $w \mapsto I$ . The 3DMM Fitting network, denoted by $M$ , regresses the 3DMM parameters $p$ from image $I$ . The parameters include pose parameter $p_p$ , expression parameter $p_e$ and illumination parameter $p_i$ . In this paper, we use DECA [10] as our 3DMM Fitting network. + +We first define the distance metric of the pose, expression, and illumination. The distance metrics measure the attribute difference between two images generated by style codes $w_{1}, w_{2}$ . In addition, we add identity loss to the metric. Therefore, the distance metric increases if the difference between two images' attributes increases and the identity remains similar. Then we decompose the Hessian matrix to find the direction with the fastest distance metric change in the $\mathrm{W+}$ space. Therefore, the attributes change fast but identity changes slowly along the direction we found. The distance metrics are shown as follows: + +$$ +D _ {p} (w _ {1}, w _ {2}) = \frac {| | M (G (w _ {1})) _ {p} - M (G (w _ {2})) _ {p} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))} +$$ + +$$ +D _ {e} \left(w _ {1}, w _ {2}\right) = \frac {\left| \left| M \left(G \left(w _ {1}\right)\right) _ {e} - M \left(G \left(w _ {2}\right)\right) _ {e} \right| \right| ^ {2}}{L _ {i d} \left(G \left(w _ {1}\right) , G \left(w _ {2}\right)\right)} \tag {1} +$$ + +$$ +D _ {i} (w _ {1}, w _ {2}) = \frac {| | M (G (w _ {1})) _ {i} - M (G (w _ {2})) _ {i} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))} +$$ + +where $L_{id}$ measures the similarity of the identities of two images. Its value represents the similarity of two identity + +![](images/5dcf237d6c2b1d96e9b931f7e8214e2d4c460176a5b93452d1633f902bd0c7b6.jpg) +Figure 2. Our framework is divided into three stages. Stage one transfers the target image's expression, pose, and illumination attributes to the source image through the semantic base. Stage two transfers detailed non-identity attributes. Stage three blends the transferred image into the background of the target image. + +embeddings obtained by the pre-trained face recognition network [8] from two input face images. + +$$ +L _ {i d} \left(I _ {1}, I _ {2}\right) = 1 - \frac {I D \left(I _ {1}\right) \cdot I D \left(I _ {2}\right)}{\left| \left| I D \left(I _ {1}\right) \right| \right| _ {2} \left| \left| I D \left(I _ {2}\right) \right| \right| _ {2}} \tag {2} +$$ + +Then, we use the Hessian matrix to encode the local distance information on the image manifold up to second-order approximation. + +$$ +D ^ {2} (w _ {0}, w _ {0} + \delta w) \approx | | \delta w | | _ {H} ^ {2} = \delta_ {w} ^ {T} H (w _ {0}) \delta w \quad (3) +$$ + +The $w_0$ used to construct the semantic base is randomly sampled, and $w_0 + \delta w$ is the point near $w_0$ . We decompose the Hessian matrix $H(w_0) = \sum_i \lambda_i v_i v_i^T$ and find the first $m$ principal components of $v$ as our basis $V$ . Therefore, the basis $V$ contains the direction vectors, along which the distance metric changes first $m$ fastest. + +We use the algorithm [42] to calculate and decompose the Hessian matrix. StyleGAN has 18 different levels, and the style code of each level itself contains certain semantic features (for example, the low level includes the features of face shape and pose, and the high level has the skin color and illumination, etc.). We construct the semantic base of the pose on the first three levels of the style code and the expression on the 4th to 10th levels. The semantic base of + +the illumination is on the levels after the 10th. We sampled $w_{0}$ 100 times and generated 100 sets of bases. And then, we average these bases and performed Schmidt orthogonalization to obtain the semantic base of the pose $V_{p}$ , expression $V_{e}$ , and illumination $V_{i}$ , respectively. Finally, the pose, expression, and illumination base are combined as StyleIPSB $V = [V_{p}, V_{e}, V_{i}]$ + +# 3.2. 3DMM-StyleGAN Mapping + +As shown in Fig. 2, the source and target images are fed into the 3DMM fitting network to regress the 3DMM parameters $\alpha^s, \alpha^t$ . The 3DMM-StyleGAN Mapping network, which contains three six-layer multilayer perceptrons(MLPs), maps 3DMM parameters $\alpha$ to StyleIPSB coordinates of pose $\beta_{s \rightarrow t}^p$ , expression $\beta_{s \rightarrow t}^e$ and illumination $\beta_{s \rightarrow t}^i$ , respectively. The following equations calculate the transferred style code $w_{s \rightarrow t}$ : + +$$ +w _ {s \rightarrow t} = w _ {s} + \beta_ {s \rightarrow t} V \tag {4} +$$ + +where $\beta_{s\rightarrow t} = [\beta_{s\rightarrow t}^p,\beta_{s\rightarrow t}^e,\beta_{s\rightarrow t}^i ]$ and $w_{s}$ is the corresponding style code of the source image, which is obtained from the source image by the StyleGAN encoder [41]. Finally, $w_{s\rightarrow t}$ passes through StyleGAN to get the attribute transferred image. + +![](images/18f48963d7dbc37f62adb92016013c716c9a7b0ba2c6802bd6d3976268529d5b.jpg) +Figure 3. The network structure of the image blend. Stage three blends transformed image to the target image by mixing the VGG encoder feature with the StyleGAN feature through masked spatial feature transforms (MSFT). + +![](images/566fe69ac52d9fc12ef4c970e107361e7398ae1b87abe8b8c8cfb52277d1ff73.jpg) + +3DMM-StyleGAN mapping network aims at precisely controlling the generated attribute through the 3DMM parameters while keeping the identity unchanged. Therefore, we design the attribute loss $L_{attr}$ to make the generated face image have the same attribute as the pose, expression, and illumination as target 3DMM parameters. And we add identity loss $L_{id}$ to preserve the identity of the source image. The total loss is shown as: + +$$ +L = L _ {i d} + \varepsilon_ {\text {a t t r}} L _ {\text {a t t r}} \tag {5} +$$ + +where $L_{id}$ measure the identity similarity of two images and defined in Eq. (2). $\varepsilon_{attr}$ is the weight of two loss functions. $L_{attr}$ measures the difference between the attributes of the transferred face and the target face. We reconstruct the 3D face from the transferred face and the target face, and then compare their difference in 3D face geometries and rendered images: + +$$ +\begin{array}{l} L _ {a t t r} \left(\alpha^ {s}, \alpha^ {t}, \alpha^ {s \rightarrow t}\right) = L _ {g e o} \left(\alpha^ {s}, \alpha^ {s \rightarrow t}\right) (6) \\ + L _ {r e n d e r} \left(\alpha^ {s}, \alpha^ {t}, \alpha^ {s \rightarrow t}\right) (6) \\ \end{array} +$$ + +where $\alpha^s, \alpha^t$ is the 3DMM parameters of the source and target images. And $\alpha^{s\rightarrow t}$ is 3DMM parameters of the transferred image where $\alpha^{s\rightarrow t} = M(G(w_{s\rightarrow t}))$ . + +We define that $G_{3DMM}(\alpha_s, \alpha_e, \alpha_p)$ can generate 3D geometry from 3DMM parameters, and $R(\alpha_s, \alpha_e, \alpha_a, \alpha_i, \alpha_p)$ can generate rendered images from 3DMM parameters, where $\alpha_s, \alpha_e, \alpha_a, \alpha_i, \alpha_p$ represent shape parameters, pose parameters, expression parameters, albedo parameters, and lighting parameters, respectively. Because we only compare attribute differences between transferred and target images, we reconstruct the 3D face by the same 3DMM shape parameters. + +The geometric term $L_{geo}$ uses the $L_{2}$ loss between two + +face meshes: + +$$ +L _ {g e o} \left(\alpha^ {s}, \alpha^ {s \rightarrow t}\right) = +$$ + +$$ +\frac {1}{N} \left\| G _ {3 D M M} \left(\alpha_ {s} ^ {s}, \alpha_ {e} ^ {t}, \alpha_ {p} ^ {t}\right) - G _ {3 D M M} \left(\alpha_ {s} ^ {s \rightarrow t}, \alpha_ {e} ^ {s \rightarrow t}, \alpha_ {p} ^ {s \rightarrow t}\right)\right\| _ {2} \tag {7} +$$ + +where $N$ is the number of vertices of the face mesh. + +The render term $L_{\text{render}}$ uses the $L_1$ loss between two rendered images: + +$$ +L _ {r e n d e r} \left(\alpha^ {s}, \alpha^ {t}, \alpha^ {s \rightarrow t}\right) = +$$ + +$$ +\left\| R \left(\alpha_ {s} ^ {s}, \alpha_ {e} ^ {t}, \alpha_ {a} ^ {t}, \alpha_ {i} ^ {t}, \alpha_ {p} ^ {t}\right) - R \left(\alpha_ {s} ^ {s \rightarrow t}, \alpha_ {e} ^ {s \rightarrow t}, \alpha_ {a} ^ {t}, \alpha_ {i} ^ {s \rightarrow t}, \alpha_ {p} ^ {s \rightarrow t}\right)\right\| _ {1} \tag {8} +$$ + +# 3.3. Detailed Attribute Transformation + +In the previous subsection, we transfer the attributes of the target face to the source face, including pose, expression, and illumination. In this subsection, to transfer attributes beyond the 3DMM expressive capabilities, we use DAT-MLP to transfer more detailed attributes. + +First, we project the target images into the $W+$ space using the StyleGAN encoder to obtain the style code $w_{t}$ . Then, DAT-MLP regresses $\lambda$ from style codes $w_{t}, w_{s \to t}$ using six-layer MLP. Finally, we use the following formula to get the latent vector $w_{s \to t}^{\prime}$ that contains non-identity attributes of the target image: + +$$ +w _ {s \rightarrow t} ^ {\prime} = \lambda w _ {s \rightarrow t} + (1 - \lambda) w _ {t} \tag {9} +$$ + +where $\lambda \in R^{18\times 512}$ is constrained to be between 0 and 1 through the sigmoid activation layer. We want to transfer attributes of $w_{t}$ into $w_{s\rightarrow t}^{\prime}$ as many as possible. The loss function of the second stage is as follows: + +$$ +L = L _ {i d} + \varepsilon_ {p} L _ {p} + \varepsilon_ {\text {a t t r}} L _ {\text {a t t r}} \left(\alpha^ {s}, \alpha^ {t}, M \left(G \left(w _ {s \rightarrow t} ^ {\prime}\right)\right)\right) \tag {10} +$$ + +The loss function is divided into three items. The first item is face recognition loss, which is used to constrain + +the identity of the generated face to be consistent with the source image. The second item $L_{p}$ represents the perceptual distance between the generated image and the target image. It can help transfer residual attributes other than pose, expression, and illumination. The third term $L_{attr}$ constrains the attributes of the generated face consistent with the target image. $\varepsilon_{p}, \varepsilon_{attr}$ represent the weight of the loss function. + +The Eq. (9) shows that the style code $w_{s \to t}^{\prime}$ is in a bounding box of two style codes $w_{s \to t}$ and $w_{t}$ . We consider that most of the bounding box is still in the subspace of $\mathbf{W}+$ space because the attribute of two style codes $w_{s \to t}$ and $w_{t}$ is very close and generated image retains pore-level details in practice. + +# 3.4. Blend Image + +This subsection introduces how to blend the generated image into the target image. We propose Masked Spatial Feature Transform (MSFT) module to fuse the feature. Unlike traditional Spatial Feature Transform (SFT) [44], MSFT only fuses the feature of the masked regions. + +The detail of this module is shown in Fig. 3. The VG-GFace network [30] extracts various level features of the target image, which is injected into the StyleGAN through the MSFT module. We use Gaussian filtering to filter the masked image obtained by the face segmentation algorithm to make the boundary smoother. The StyleGAN feature is retained in the facial area, while the background area is from the mixed feature. We mix features of StyleGAN and VG-GFace using SFT. The background loss and the perceptual loss make the swapped image have the same background as the target image: + +$$ +L = L _ {b} + \varepsilon_ {p} L _ {p} \tag {11} +$$ + +$L_{b}$ is the background loss used to measure the difference in the background between the generated image and the target image. $L_{p}$ represents the perceptual distance between the generated image and the target image. $\varepsilon_{p}$ is the weight of the loss function. The background loss is expressed as: + +$$ +L _ {b} = \left\| M \odot \left(I _ {r} - I _ {t}\right) \right\| \tag {12} +$$ + +where $M$ is the binarized image of the background region obtained by face segmentation algorithm [2]. + +# 4. Experiment + +As mentioned above, StyleIPSB has the following advantages: (1) StyleIPSB can represent various poses, expressions, and illumination properties. (2) When the style code is modified along one base, it only changes the specific attribute while remaining identity unchanged. (3) StyleIPSB can preserve the pore-level details. (4) StyleIPSB cooperates well with 3DMM to control facial attributes. Therefore, in this section, we mainly evaluate + +the effectiveness of our method in the following aspects: (1) Evaluating the properties of StyleIPSB. (2) Evaluating the performance of 3DMM controlling facial attributes with StyleIPSB. (3) Comparison of face swapping results with other methods. First of all, we introduce the dataset and training detail. + +# 4.1. Datasets and Training Detail + +Dataset. We used the FFHQ database [15] to train the 3DMM StyleGAN Mapping module. Flickr-Faces-HQ (FFHQ) consists of 70,000 high-quality face images at $1024 \times 1024$ resolution. CelebAMask-HQ [20] is a large-scale face image dataset with 30,000 high-resolution face images selected from the CelebA dataset by following CelebA-HQ. FaceForensics++ [32] is a forensics dataset consisting of 1,000 original video sequences. + +Training environment and hyperparameter We train the network on a GTX 3090 using the Pytorch framework. In training, our optimizer is Adam, the weight decay is 5e-5, and the batch size is 2. In order to effectively train the above three stages, we first pre-train the three stages separately and finally train them together. In the three-stage pre-training, the learning rate is 2e-4; in the final together training stage, we adjust the learning rate to 2e-5. + +# 4.2. The Properties of StyleIPSB + +The representation and the disentanglement of StyleIPSB. We conduct experiments on the representation of StyleIPSB to evaluate whether it can represent various poses, expressions, and illumination properties while identity is unchanged. Fig. 4 shows that StyleIPSB can express pitch, yaw rotations, and lighting with different colors, directions, and intensities. In terms of expressions, StyleIPSB can not only control whether the mouth is open or not, the eyes are open or closed but also control different eyeball orientations and raising eyebrows. Supplementary material reveals the effect of modifying coordinates in different base directions. + +To evaluate the disentanglement of StyleIPSB, we present the quantitative results, which show the identity change when manipulating the pose. As shown in Fig. 5, the horizontal axis represents the yaw angle between the anchored face and the edited image, and the vertical axis represents identity loss as defined in Eq. (2) between the anchored face and the edited image. Our StyleIPSB outperforms Ganspace [13], which is also a basis constructed in $\mathrm{W}+$ space using unsupervised learning. We achieve comparable results with the supervised method. InterfaceGAN [34]. It needs a dataset that contains left and right-facing faces. StyleIPSB uses unsupervised learning and can represent more attributes than InterfaceGAN. Fig. 5 also shows that if we introduce identity loss into the distance metric when building StyleIPSB, the identity-preserving perfor + +![](images/1aa025ba8784bf394be246e01680491d6dbe57a2b39ae544324aee6023a8405e.jpg) +Figure 4. StyleIPSB can modify various poses, expressions, and illumination by adjusting its coordinates while keeping the identity unchanged. The first row is the results of changing coordinates of pose and illumination basis. The second row shows the results of the changing coordinate of the expression basis. + +![](images/44a43f0c63f5c4a7e169767e3b14210bcd67331ee3f2ad8acb67bd51a0b4df18.jpg) + +![](images/74d3cc1660c4dc68a20388f456e659b6cea66dd5801fa352102f74622f665620.jpg) +Figure 5. Effects of different methods of changing face pose attributes on face identity information + +![](images/3c089c26ec461eae45fd799aa97ef7ab692ea01effddfcd2492fe8e31f100346.jpg) +Figure 6. We compare our method with GANSpace [13] and InterfaceGAN [34] in the editing of facial expression. + +mance of StyleIPSB is enhanced. + +Editing results of StyleIPSB. We compare StyleIPSB with other StyleGAN-based face edit methods. Fig. 6 compares the edit results of expression (smile) with other methods. The results show that our approach has better performance in identity-preserving. StyleIPSB preserves the face shape when expression changes, while face shape is deformed in other methods. + +Fig. 7 shows examples of editing a face image using the + +![](images/e9c8577d075d33d9cfce7f2509e21022b6b6d20d31d6e8b831b7d46354e21428.jpg) +Figure 7. 3D face models control the attribute of generated images. The top is the rendered images of 3D face geometry. The bottom is the edited images controlled by the 3DMM parameters of the top 3D face geometry. + +3DMM parameters of FLAME [23]. The first row is the 3D face models generated by the 3DMM parameters. The second row is the edited image controlled by those 3DMM parameters. These results show that StyleIPSB cooperates well with 3DMM to control the facial attributes of the generated image. + +# 4.3. Result of Face Swapping + +This subsection evaluates the performance of face swapping based on StyleIPSB. First, the ablation experiments evaluate the influence of StyleIPSB on attribute transfer performance. Then, we compare our face swapping with other methods. + +Ablation experiment In the experiment, "no basis" means that the 3DMM-StyleGAN mapping network directly regresses the $w_{s\rightarrow t}$ instead of $\beta_{s\rightarrow t}$ . As shown in Fig. 8, using StyleIPSB can constrain the generated style code on the subspace of $\mathrm{W + }$ space and preserve the pore-level details. + +Tab. 2 shows the quantitative comparison results on the performance of StyleIPSB in stage one. FID (Fréchet inception distance) is used to evaluate the image quality generated by the image generation model. "Exp" and "Pose" represent the Euclidean distance between the target image and the transferred image expression and pose parameters, respectively. ID Similarity means the cosine similarity be + +![](images/64c5a4b960f30cf8e918e02bb555b5c6976bb01effbe381d4e1f84908d7ce792.jpg) +Figure 8. The qualitative ablation experiments on StyleIPSB. The results show we can generate pore-level details with StyleIPSB. + +
FID ↓Exp ↓Pose ↓ID similarity ↑
No Basis26.063.730.0740.65
With basis22.153.370.0780.67
+ +tween the identity embedding of the source image and the transferred image. D3FR [9] is used to extract the 3D face parameters, and CosFace [43] is used to extract the identity embedding. The test set is 10,000 image pairs randomly sampled from the CelebAMask-HQ [20]. The Tab. 2 shows that using StyleIPSB can effectively reduce the value of FID because StyleIPSB is in the subspace of $\mathrm{W + }$ space. In addition, the performance of transferring expressions with StyleIPSB is better in identity-preserving because the identity is preserved while changing the coordinates of StyleIPSB. + +Tab. 3 shows the quantitative comparison results of different methods on the FaceForensics++ dataset. The dataset is classified into 885 identities, and the identity retrieval indicates the top-1 matching rate of swapped and the source image. We apply CosFace [43] to extract identity embedding. Expressions are measured in the same way as Tab. 2, The pose measurement method uses a pose estimator [33]. We compare our results with FaceSwap [3], DeepFakes [1], FaceShifter [22], MegaFace [51] and FSLDS [47]. As shown in the Tab. 3, our method outperforms in identity-preserving and expression transfer thanks to StyleIPSB's identity-preserving property and the ability to present various expressions. Our performance of pose transfer does not outperform other methods, which may be because StyleIPSB only contains two bases expressing poses in raw and pitch directions, which makes it hard to express very accurate poses. However, the error of the pose is only a few degrees, so our results still look accurate visually. + +Fig. 9 shows the results of face swapping under challenging conditions. The results show that our method can preserve the unique noise patterns in portraiture in the red boxes. In the second row of the figure, our method successfully keeps pore-level details and identity while the source and target pose is quite different. Tab. 4 shows that our method has the better image quality and identity preservation ability than other StyleGAN-based methods. The met- + +Table 2. The quantitative ablation experiments on StyleIPSB. + +
MethodID Retri.(%) ↑Exp Err. ↓Pose Err. ↓
FaceSwap [3]72.692.892.58
Deepfakes [1]88.393.334.64
FaceShifter [22]90.682.822.55
MegaFS [51]90.832.922.64
FSLDS [47]90.052.792.46
Ours95.052.233.58
+ +Table 3. The quantitative experiments on FaceForensics++ dataset with other methods. + +
FID ↓Exp ↓Pose ↓ID similarity ↑
MageFS [51]22.032.850.0430.4837
RAFS [46]13.253.15-0.5232
FSLDS [47]10.012.990.0530.4761
Ours9.372.750.0780.5378
+ +Table 4. The quantitative experiments on the CelebAHQ dataset with other methods. + +![](images/e9e7794e558308f9f9ed6750e05305bc63718632f5b667951318ebcf5e70593e.jpg) +Figure 9. We compare our method with MageFS [51] and FSLDS [47] in challenging conditions. + +![](images/07e01d192ddee6770729044c541eadfe73a4f218c9d244cc613347cd446d194b.jpg) +Figure 10. limitation of our method + +ric is the same as in Tab. 2. We follow RAFS and compare $100\mathrm{k}$ swapped faces for a fair comparison. + +# 5. Conclusion and Limitation + +We have developed a new semantic basis for face swapping, called StyleIPSB, that is specifically designed to preserve identity and pore-level details. Our experiments have demonstrated that StyleIPSB outperforms other state-of-the-art methods. Despite this, there is still potential for further improvement, as shown in Fig. 10. (1) Occlusion is limited by the mask. (2) The glasses in the source image cannot be removed. (3) Light and shadow cannot be perfectly restored in the case of complex illumination. + +# References + +[1] Deepfakes. https://github.com/ondyari/FaceForensics/tree/master/dataset/DeepFakes.8 +[2] face-parsing.pytorch. https://github.com/zllrunning/face-parsing.PyTorch.6 +[3] Faceswap. https://github.com/ondyari/FaceForensics/tree/master/dataset/FaceSwapKowalski.8 +[4] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2StyleGAN: Unsupervised extraction of StyleGAN edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3 +[5] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (TOG), 40(3):1-21, 2021. 2 +[6] Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, and Gang Hua. Towards open-set identity preserving face synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6713-6722, 2018. 2 +[7] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003-2011, 2020. 2, 3 +[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 4 +[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 8 +[10] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3 +[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2 +[12] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. GIF: Generative interpretable faces. In 2020 International Conference on 3D Vision (3DV), pages 868-878. IEEE, 2020. 2, 3 +[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANspace: Discovering interpretable GAN controls. arXiv preprint arXiv:2004.02546, 2020. 2, 3, 6, 7 +[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 + +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 6 +[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2 +[17] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2 +[18] Jiseob Kim, Jihoon Lee, and Byoung-Tak Zhang. Smooth-Swap: A simple enhancement for face-swapping with smoothness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10779-10788, 2022. 2 +[19] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 3677-3685, 2017. 3 +[20] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 8 +[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. FaceInpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2, 3 +[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 2, 3, 8 +[23] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7 +[24] Yuan Lin, Shengjin Wang, Qian Lin, and Feng Tang. Face swapping under large pose variations: A 3d model based approach. In 2012 IEEE International Conference on Multimedia and Expo, pages 333-338. IEEE, 2012. 3 +[25] Luming Ma and Zhigang Deng. Real-time hierarchical facial performance capture. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 1-10, 2019. 3 +[26] Luming Ma and Zhigang Deng. Real-time face video swapping from a single portrait. In Symposium on Interactive 3D Graphics and Games, pages 1-10, 2020. 3 +[27] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 2 +[28] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. arXiv preprint arXiv:2202.12972, 2022. 2, 3 + +[29] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2 +[30] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In Proceedings of the British Machine Vision Conference (BMVC). British Machine Vision Association, 2015. 6 +[31] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleClip: Text-driven manipulation of StyleGAN imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2 +[32] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 6 +[33] Nataniel Ruiz, Eunji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 8 +[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 2, 3, 6, 7 +[35] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 3 +[36] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3 +[37] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. SemanticStyleGAN: Learning compositional generative priors for controllable image synthesis and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11254-11264, 2022. 1, 2 +[38] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3 +[39] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. StyleRig: Rigging StyleGAN for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2, 3 +[40] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of RGB videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3 + +[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4 +[42] Binxu Wang and Carlos R Ponce. The geometry of deep generative image models and its applications. arXiv preprint arXiv:2101.06006, 2021. 4 +[43] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 8 +[44] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6 +[45] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 2, 3 +[46] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2, 3, 8 +[47] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7642-7651, 2022. 1, 2, 3, 8 +[48] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. TransEditor: Transformer-based dual-space GAN for highly controllable facial editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2022. 1 +[49] Zhiliang Xu, Zhibin Hong, Changxing Ding, Zhen Zhu, Junyu Han, Jingtuo Liu, and Errui Ding. MobileFaceSwap: A lightweight framework for video face swapping. arXiv preprint arXiv:2201.03808, 2022. 3 +[50] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. StyleHeat: One-shot high-resolution editable talking face generation via pretrained StyleGAN. arXiv preprint arXiv:2203.04036, 2022.1 +[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2, 3, 8 \ No newline at end of file diff --git a/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/images.zip b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5e54ff2814cd4220b260a9f3bdf9dff6b7758db9 --- /dev/null +++ b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:769126a7929e794ee2930e07e83d6191fe16e6fa9284a25758bb48b10e936ca8 +size 682394 diff --git a/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/layout.json b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..669d59436c01d19e3084ecdec04e2ff084c5309e --- /dev/null +++ b/2023/StyleIPSB_ Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping/layout.json @@ -0,0 +1,8793 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 102, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 102, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 493, + 140 + ], + "type": "text", + "content": "StyleIPSB: Identity-Preserving Semantic Basis of StyleGAN for High Fidelity Face Swapping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "spans": [ + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "content": "Diqiong Jiang " + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "content": ", Dan Song " + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "content": ", Ruofeng Tong " + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "content": ", Min Tang " + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "content": "Zhejiang University, China \n" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 160, + 436, + 190 + ], + "type": "text", + "content": "Tianjin University, China" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 66, + 206, + 545, + 417 + ], + "blocks": [ + { + "bbox": [ + 66, + 206, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 66, + 206, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 66, + 206, + 545, + 417 + ], + "type": "image", + "image_path": "3722d226273fccf98276d40e8e3aa3a301103da5b224495b7c844d434c1282f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 426, + 547, + 450 + ], + "lines": [ + { + "bbox": [ + 45, + 426, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 547, + 450 + ], + "type": "text", + "content": "Figure 1. Compared with the existing high-fidelity face swapping methods MageFS [51] and FSLDS [47], our face swapping framework based on the identity-preserving semantic basis (StyleIPSB) can preserve the pore-level detail and source image's identity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 459, + 192, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 459, + 192, + 472 + ], + "spans": [ + { + "bbox": [ + 143, + 459, + 192, + 472 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 484, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 484, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 45, + 484, + 288, + 676 + ], + "type": "text", + "content": "Recent researches reveal that StyleGAN can generate highly realistic images, inspiring researchers to use pretrained StyleGAN to generate high-fidelity swapped faces. However, existing methods fail to meet the expectations in two essential aspects of high-fidelity face swapping. Their results are blurry without pore-level details and fail to preserve identity for challenging cases. To overcome the above artifacts, we innovatively construct a series of identity-preserving semantic bases of StyleGAN (called StyleIPSB) in respect of pose, expression, and illumination. Each basis of StyleIPSB controls one specific semantic attribute and disentangles with the others. The StyleIPSB constrains style code in the subspace of " + }, + { + "bbox": [ + 45, + 484, + 288, + 676 + ], + "type": "inline_equation", + "content": "W+" + }, + { + "bbox": [ + 45, + 484, + 288, + 676 + ], + "type": "text", + "content": " space to preserve pore-level details and gives us a novel tool for high-fidelity face swapping, and we propose a three-stage framework for face swapping with StyleIPSB. Firstly, we transform the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 460, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 547, + 616 + ], + "type": "text", + "content": "target facial images' attributes to the source image. We learn the mapping from 3D Morphable Model (3DMM) parameters, which capture the prominent semantic variance, to the coordinates of StyleIPSB that show higher identity-preserving and fidelity. Secondly, to transform detailed attributes which 3DMM does not capture, we learn the residual attribute between the reenacted face and the target face. Finally, the face is blended into the background of the target image. Extensive results and comparisons demonstrate that StyleIPSB can effectively preserve identity and pore-level details. The results of face swapping can achieve state-of-the-art performance. We will release our code at https://github.com/a686432/StyleIPSB" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 643, + 387, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 387, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 387, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Facial image manipulation [36, 37, 48, 50] is a task of transforming specific attributes from the source image to the target image while persevering other attributes unchanged. Face swapping is one of the essential parts of facial im" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 149, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 149, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 149, + 693 + ], + "type": "text", + "content": "*Co-Corresponding authors" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": " The research is supported in part by the National Natural Science Foundation of China (61972342, 61832016, 61972341, 61902277)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 748, + 313, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 758 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 758 + ], + "type": "text", + "content": "352" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "age manipulation, which has attracted lots of interest in the computer vision and graphics community. Face swapping aims to generate an image with the source image's identity and the target image's attributes (e.g., expression, pose, background, hair, etc.). It has wide applications in the film industry and computer games." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 277 + ], + "type": "text", + "content": "Current face swapping methods are mainly divided into two categories: 3D model-based methods and 2D image-based methods. 3D model-based methods [11,17,40] firstly reconstruct the 3D face models based on 3DMM from the source image and target image and transfer the nonidentity parameters of the target face model to the source face model. Then they render the transferred 3D model and blend it into the target image. Although such methods can transfer coarse facial attributes such as pose, expression, and illumination, they have difficulty in generating realistic hair and teeth accessories." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 434 + ], + "type": "text", + "content": "With the development of generative adversarial networks, 2D image-based methods [6, 7, 22, 27, 28, 45] can synthesize photo-realistic images. The generated face images have convincing and detailed facial attributes, such as mouth, teeth, and eyebrows. Recent works [21, 46, 47, 51] employ the pre-trained StyleGAN decoder to further improve the fidelity and synthesize pore-level details. However, as shown in Fig. 1, despite using the pre-trained StyleGAN model, their results fail to generate pore-level details and identity-preserving in challenge conditions. Overall, 2D image-based methods generate more realistic images than 3D model-based methods, but the identity-preserving and pore-level details of the images still need improvement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "type": "text", + "content": "To tackle the challenges of blurry images and identity-preserving in face swapping, according to our observations, the cause of the blurred images is that the regressed style code is out of " + }, + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "type": "text", + "content": " space. Additionally, as mentioned in [18], identity embedding is a non-smooth space, so finding the identity-preserving optimized direction is challenging. To address these problems, our method constrains the regressed style code with identity-preserving semantic bases of StyleGAN (i.e., the proposed StyleIPSB). StyleIPSB stays within the " + }, + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 436, + 286, + 567 + ], + "type": "text", + "content": " space, and the identity is preserved when changing its coordinates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "The advantages of the proposed StyleIPSB are summarized as follows: (1) StyleIPSB constitutes a linear space, which is the subspace of the " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " space of StyleGAN. By ensuring the regressed style code within the " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " space of StyleGAN, we can more easily generate images with pore-level details. (2) When changing the coordinates of the StyleIPSB, the identity remains preserved as much as possible. (3) StyleIPSB can represent various poses, expressions, and illuminations. To construct the basis that satisfies the above properties, we propose a novel identity-preserving distance metric to find the orthogonal semantic directions, which are further assembled to StyleIPSB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "StyleIPSB also cooperates well with 3DMM to control facial attributes. StyleRig [39] builds the mapping of the 3DMM parameter space and " + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": " space of StyleGAN, which can change the facial attribute of the generated image by the 3DMM parameters. StyleRig only can manipulate images generated by StyleGAN. Pie [38] designs a non-linear optimization problem to edit the real-world image based on StyleRig, but the optimization operation is time-consuming. GIF [12] generates face images by the FLAME [23] parametric control. However, the generated images are easy to contain artifacts and change identity. Other face manipulation methods [4, 29, 37] use the network directly to find the edit direction. Still, without the guidance of 3DMM, they can only generate some basic expressions (e.g., smile) and fail to cover various expressions. In this paper, we propose the StyleGAN-3DMM mapping network, which transforms the semantic information of 3DMM parameters into StyleIPSB coordinates. The StyleGAN-3DMM mapping network reduces the gap between 3DMM and StyleIPSB. It shows that StyleIPSB is very compatible with 3DMM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 312, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 360 + ], + "type": "text", + "content": "In summary, we propose a face swapping framework based on StyleIPSB and achieve state-of-the-art results. The main contributions of this paper lie in the following three aspects:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 369, + 545, + 555 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 369, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 369, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 545, + 429 + ], + "type": "text", + "content": "- We propose a novel method of establishing identity-preserving semantic bases of StyleGAN called StyleIPSB. The face image, generated by the linear space of StyleIPSB, remains pore-level details and identity-preserving." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 437, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 545, + 498 + ], + "type": "text", + "content": "- The proposed StyleGAN-3DMM mapping network serves as the bridge to narrow the gap between 3DMM and StyleIPSB, which can take advantage of the prominent semantic variance of 3DMM and the identity-preserving and high-fidelity of styleIPSB." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 507, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 507, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 507, + 545, + 555 + ], + "type": "text", + "content": "- We propose the face swapping framework based on StyleIPSB and StyleGAN-3DMM mapping network. Extensive results show our method outperforms others in detail-preserving and identity-preserving." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 567, + 394, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 394, + 579 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 394, + 579 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 500, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 500, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 500, + 600 + ], + "type": "text", + "content": "2.1. Image Modification Using StyleGAN" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "StyleGAN [14-16] is a powerful image synthesis model that can generate a wide variety of high-quality face images. Some methods [5, 31, 34, 35] find the particular attribute editing direction in " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}+" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " space using the corresponding attribute classifier. However, their attribute editing is limited by the classifier's ability and cannot control diverse expressions and illumination. To facilitate attribute manipulation in an unsupervised manner, GANSpace [13] and SeFa [36] perform decomposition to find primary directions" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "353" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "in the latent space and explore the interpretable directions among the primary directions. CLIP2StyleGAN [4] discovery and label of StyleGAN edit-direction based on CLIP image space. However, they only find a few types of expressions and illumination and fail to discover a complete set of expressions and illumination. Some works [12,38,39] map the rigging information to face manipulation, but our StyleIPSB can generate more detail and identity-preserving results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 186, + 119, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 119, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 119, + 199 + ], + "type": "text", + "content": "2.2. Face swap" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 456 + ], + "type": "text", + "content": "In the early years, many works [24-26, 40] apply 3D face models for face swapping. Face2face [40] applies an efficient deformation transfer to track the source video facial expressions and re-render the synthesized faces with retrieved and warped mouth interiors. Ma et al. [25] reconstructed high-resolution facial geometry and appearance by capturing an individual-specific face model with fine-scale details. Those 3D model-based methods are difficult to generate high-fidelity face images, especially realistic hair, mouth, and teeth. In recent years, conditional GAN architecture has been widely used in face-swapping. Many works [7, 19, 21, 22, 28, 46, 49] use neural networks to generate high-fidelity images. SimSwap [7] and FaceShifter [22] use a face recognition network to extract the identity embedding and use a decoder to fuse the identity embedding. Recently, many methods [21, 45-47, 51] use pretrained StyleGAN as the image generator to improve the quality of generated images further. However, the images generated by their method still suffer from the lack of details and face identity shift. Our approach is better at retaining pore-level details and identify-preserving." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 465, + 104, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 104, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 104, + 477 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 570 + ], + "type": "text", + "content": "This section reveals the details of our proposed method in three aspects. The first part describes the process of conducting StyleIPSB. The second part describes how to use StyleIPSB to transfer the facial attributes except for identity from the target image to the source face. The third introduces the overall framework of face swap based on StyleIPSB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 182, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 182, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 182, + 588 + ], + "type": "text", + "content": "3.1. StyleIPSB Construction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "This subsection aims to find a semantic basis in the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " space of StyleGAN that satisfies the conditions mentioned in the introduction. In short, StyleIPSB should meet the following properties: a subspace of " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " space, identity-preserving, and representation ability. The previous methods fail to satisfy the above three conditions. For example, as shown in Tab. 1, GIF [12] does not meet the requirements in the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " subspace, which would lead to generated image distortion. Some methods [12,35,36] do not concern the identity preservation and suffer the identity shift when" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 158 + ], + "type": "table", + "html": "
MethodSubspaceIDRepresentation
InterFaceGAN [34]✓✓
GANSpace [13]✓
SeFa [36]✓
GIF [12]✓
Our✓✓✓
", + "image_path": "0611b8c72e8e5061b27d148d175448e25fe0db41ae49fd485550ce7b9cff68bb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 166, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 547, + 200 + ], + "type": "text", + "content": "Table 1. The conditions of the basis satisfied by different methods. We compare our method with InterFaceGAN [34], GANSpace [13], GIF [12] and SeFa [36]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 220, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 255 + ], + "type": "text", + "content": "editing. InterFaceGAN [34] cannot edit various expressions and illumination. Unlike existing methods, our method considers all three conditions when we build StyleIPSB." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "type": "text", + "content": "Compared with the previous methods, which built the basis by directly performing decomposition on eigenvalue in " + }, + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "type": "text", + "content": " space, we added semantic metrics and identity loss in our decomposition. Therefore, StyleIPSB is not only in the " + }, + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 255, + 546, + 375 + ], + "type": "text", + "content": " subspace but can also represent the complete set of semantic information and have the property of identity preservation. So adding semantic information and maintaining identity when building StyleIPSB is the crucial point to this subsection. The following introduces the detailed process of conducting StyleIPSB." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": "**Formulation** The StyleGAN network, denoted by " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ", is a mapping from style code " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": " to image " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "G: R^n \\to R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "w \\mapsto I" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ". The 3DMM Fitting network, denoted by " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ", regresses the 3DMM parameters " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": " from image " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ". The parameters include pose parameter " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "p_p" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ", expression parameter " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "p_e" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": " and illumination parameter " + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 375, + 547, + 459 + ], + "type": "text", + "content": ". In this paper, we use DECA [10] as our 3DMM Fitting network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "text", + "content": "We first define the distance metric of the pose, expression, and illumination. The distance metrics measure the attribute difference between two images generated by style codes " + }, + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "inline_equation", + "content": "w_{1}, w_{2}" + }, + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "text", + "content": ". In addition, we add identity loss to the metric. Therefore, the distance metric increases if the difference between two images' attributes increases and the identity remains similar. Then we decompose the Hessian matrix to find the direction with the fastest distance metric change in the " + }, + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{W+}" + }, + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "text", + "content": " space. Therefore, the attributes change fast but identity changes slowly along the direction we found. The distance metrics are shown as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 596, + 518, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 596, + 518, + 624 + ], + "spans": [ + { + "bbox": [ + 321, + 596, + 518, + 624 + ], + "type": "interline_equation", + "content": "D _ {p} (w _ {1}, w _ {2}) = \\frac {| | M (G (w _ {1})) _ {p} - M (G (w _ {2})) _ {p} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))}", + "image_path": "3d46e2202ef5ff35c035e5e0bef444388c1889bd5d79754d0135012445aa6dfd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 624, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 545, + 654 + ], + "type": "interline_equation", + "content": "D _ {e} \\left(w _ {1}, w _ {2}\\right) = \\frac {\\left| \\left| M \\left(G \\left(w _ {1}\\right)\\right) _ {e} - M \\left(G \\left(w _ {2}\\right)\\right) _ {e} \\right| \\right| ^ {2}}{L _ {i d} \\left(G \\left(w _ {1}\\right) , G \\left(w _ {2}\\right)\\right)} \\tag {1}", + "image_path": "f655eca7cd18cfc35785873bed69dc9b09bf9bec7e3b3f26ce93137f2758e751.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 654, + 515, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 654, + 515, + 684 + ], + "spans": [ + { + "bbox": [ + 321, + 654, + 515, + 684 + ], + "type": "interline_equation", + "content": "D _ {i} (w _ {1}, w _ {2}) = \\frac {| | M (G (w _ {1})) _ {i} - M (G (w _ {2})) _ {i} | | ^ {2}}{L _ {i d} (G (w _ {1}) , G (w _ {2}))}", + "image_path": "5c35f0e627316e2047db03962f00ef6ba03bfe91a964d6f8d47311f03091c68e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": " measures the similarity of the identities of two images. Its value represents the similarity of two identity" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "354" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 541, + 343 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 541, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 541, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 541, + 343 + ], + "type": "image", + "image_path": "5dcf237d6c2b1d96e9b931f7e8214e2d4c460176a5b93452d1633f902bd0c7b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 348, + 546, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 348, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 546, + 383 + ], + "type": "text", + "content": "Figure 2. Our framework is divided into three stages. Stage one transfers the target image's expression, pose, and illumination attributes to the source image through the semantic base. Stage two transfers detailed non-identity attributes. Stage three blends the transferred image into the background of the target image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 413, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 288, + 437 + ], + "type": "text", + "content": "embeddings obtained by the pre-trained face recognition network [8] from two input face images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 444, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 444, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 78, + 444, + 287, + 472 + ], + "type": "interline_equation", + "content": "L _ {i d} \\left(I _ {1}, I _ {2}\\right) = 1 - \\frac {I D \\left(I _ {1}\\right) \\cdot I D \\left(I _ {2}\\right)}{\\left| \\left| I D \\left(I _ {1}\\right) \\right| \\right| _ {2} \\left| \\left| I D \\left(I _ {2}\\right) \\right| \\right| _ {2}} \\tag {2}", + "image_path": "4750d96d27714c1cdd615f9797c155bfdde859928160678989f132265ad1d866.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 515 + ], + "type": "text", + "content": "Then, we use the Hessian matrix to encode the local distance information on the image manifold up to second-order approximation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 522, + 287, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 522, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 74, + 522, + 287, + 538 + ], + "type": "interline_equation", + "content": "D ^ {2} (w _ {0}, w _ {0} + \\delta w) \\approx | | \\delta w | | _ {H} ^ {2} = \\delta_ {w} ^ {T} H (w _ {0}) \\delta w \\quad (3)", + "image_path": "afe6836a4aec6a713b4e15c9f1aaae2b42018fcc92dbaaabf5b707e5f597e45e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " used to construct the semantic base is randomly sampled, and " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "w_0 + \\delta w" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " is the point near " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": ". We decompose the Hessian matrix " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "H(w_0) = \\sum_i \\lambda_i v_i v_i^T" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " and find the first " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " principal components of " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " as our basis " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": ". Therefore, the basis " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " contains the direction vectors, along which the distance metric changes first " + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 546, + 287, + 617 + ], + "type": "text", + "content": " fastest." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "We use the algorithm [42] to calculate and decompose the Hessian matrix. StyleGAN has 18 different levels, and the style code of each level itself contains certain semantic features (for example, the low level includes the features of face shape and pose, and the high level has the skin color and illumination, etc.). We construct the semantic base of the pose on the first three levels of the style code and the expression on the 4th to 10th levels. The semantic base of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "text", + "content": "the illumination is on the levels after the 10th. We sampled " + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "inline_equation", + "content": "w_{0}" + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "text", + "content": " 100 times and generated 100 sets of bases. And then, we average these bases and performed Schmidt orthogonalization to obtain the semantic base of the pose " + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "inline_equation", + "content": "V_{p}" + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "text", + "content": ", and illumination " + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "text", + "content": ", respectively. Finally, the pose, expression, and illumination base are combined as StyleIPSB " + }, + { + "bbox": [ + 305, + 412, + 546, + 498 + ], + "type": "inline_equation", + "content": "V = [V_{p}, V_{e}, V_{i}]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 505, + 460, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 505, + 460, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 505, + 460, + 519 + ], + "type": "text", + "content": "3.2. 3DMM-StyleGAN Mapping" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": "As shown in Fig. 2, the source and target images are fed into the 3DMM fitting network to regress the 3DMM parameters " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\alpha^s, \\alpha^t" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": ". The 3DMM-StyleGAN Mapping network, which contains three six-layer multilayer perceptrons(MLPs), maps 3DMM parameters " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": " to StyleIPSB coordinates of pose " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\beta_{s \\rightarrow t}^p" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\beta_{s \\rightarrow t}^e" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": " and illumination " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\beta_{s \\rightarrow t}^i" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": ", respectively. The following equations calculate the transferred style code " + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "inline_equation", + "content": "w_{s \\rightarrow t}" + }, + { + "bbox": [ + 305, + 524, + 545, + 620 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 631, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 631, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 380, + 631, + 545, + 643 + ], + "type": "interline_equation", + "content": "w _ {s \\rightarrow t} = w _ {s} + \\beta_ {s \\rightarrow t} V \\tag {4}", + "image_path": "47f0cfcdfb5d663c04ea202aa1b0e00e0e60ea2c46a353f9b22c0556b87272ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{s\\rightarrow t} = [\\beta_{s\\rightarrow t}^p,\\beta_{s\\rightarrow t}^e,\\beta_{s\\rightarrow t}^i ]" + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "content": " is the corresponding style code of the source image, which is obtained from the source image by the StyleGAN encoder [41]. Finally, " + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow t}" + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "content": " passes through StyleGAN to get the attribute transferred image." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "355" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 350, + 224 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 350, + 224 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 350, + 224 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 350, + 224 + ], + "type": "image", + "image_path": "18f48963d7dbc37f62adb92016013c716c9a7b0ba2c6802bd6d3976268529d5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 546, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 546, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 546, + 257 + ], + "type": "text", + "content": "Figure 3. The network structure of the image blend. Stage three blends transformed image to the target image by mixing the VGG encoder feature with the StyleGAN feature through masked spatial feature transforms (MSFT)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 352, + 72, + 560, + 224 + ], + "blocks": [ + { + "bbox": [ + 352, + 72, + 560, + 224 + ], + "lines": [ + { + "bbox": [ + 352, + 72, + 560, + 224 + ], + "spans": [ + { + "bbox": [ + 352, + 72, + 560, + 224 + ], + "type": "image", + "image_path": "566fe69ac52d9fc12ef4c970e107361e7398ae1b87abe8b8c8cfb52277d1ff73.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "type": "text", + "content": "3DMM-StyleGAN mapping network aims at precisely controlling the generated attribute through the 3DMM parameters while keeping the identity unchanged. Therefore, we design the attribute loss " + }, + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "type": "text", + "content": " to make the generated face image have the same attribute as the pose, expression, and illumination as target 3DMM parameters. And we add identity loss " + }, + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 46, + 287, + 287, + 382 + ], + "type": "text", + "content": " to preserve the identity of the source image. The total loss is shown as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 399, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 399, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 121, + 399, + 287, + 411 + ], + "type": "interline_equation", + "content": "L = L _ {i d} + \\varepsilon_ {\\text {a t t r}} L _ {\\text {a t t r}} \\tag {5}", + "image_path": "f8beb58f42b9151b9e993a4bd424ca5c6c0ab89640b7a5d082a96c11e2d77fd2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "text", + "content": " measure the identity similarity of two images and defined in Eq. (2). " + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\varepsilon_{attr}" + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "text", + "content": " is the weight of two loss functions. " + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 46, + 420, + 287, + 503 + ], + "type": "text", + "content": " measures the difference between the attributes of the transferred face and the target face. We reconstruct the 3D face from the transferred face and the target face, and then compare their difference in 3D face geometries and rendered images:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 514, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 514, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 74, + 514, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {a t t r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) = L _ {g e o} \\left(\\alpha^ {s}, \\alpha^ {s \\rightarrow t}\\right) (6) \\\\ + L _ {r e n d e r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) (6) \\\\ \\end{array}", + "image_path": "9c9c2c400a60ebecb7709c7e539f0b95cc40ca740c8517fc741d34451ca4980d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "inline_equation", + "content": "\\alpha^s, \\alpha^t" + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "text", + "content": " is the 3DMM parameters of the source and target images. And " + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "inline_equation", + "content": "\\alpha^{s\\rightarrow t}" + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "text", + "content": " is 3DMM parameters of the transferred image where " + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "inline_equation", + "content": "\\alpha^{s\\rightarrow t} = M(G(w_{s\\rightarrow t}))" + }, + { + "bbox": [ + 46, + 555, + 287, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": "We define that " + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "inline_equation", + "content": "G_{3DMM}(\\alpha_s, \\alpha_e, \\alpha_p)" + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": " can generate 3D geometry from 3DMM parameters, and " + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "inline_equation", + "content": "R(\\alpha_s, \\alpha_e, \\alpha_a, \\alpha_i, \\alpha_p)" + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": " can generate rendered images from 3DMM parameters, where " + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\alpha_s, \\alpha_e, \\alpha_a, \\alpha_i, \\alpha_p" + }, + { + "bbox": [ + 46, + 593, + 287, + 699 + ], + "type": "text", + "content": " represent shape parameters, pose parameters, expression parameters, albedo parameters, and lighting parameters, respectively. Because we only compare attribute differences between transferred and target images, we reconstruct the 3D face by the same 3DMM shape parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "The geometric term " + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{geo}" + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": " uses the " + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": " loss between two" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 288, + 360, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 360, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 360, + 298 + ], + "type": "text", + "content": "face meshes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 303, + 383, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 383, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 383, + 316 + ], + "type": "interline_equation", + "content": "L _ {g e o} \\left(\\alpha^ {s}, \\alpha^ {s \\rightarrow t}\\right) =", + "image_path": "e671ef0a223cebf3b6946177f11ec4c8c21e306196241a91df80e63055cb5ec2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 319, + 551, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 551, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 551, + 350 + ], + "type": "interline_equation", + "content": "\\frac {1}{N} \\left\\| G _ {3 D M M} \\left(\\alpha_ {s} ^ {s}, \\alpha_ {e} ^ {t}, \\alpha_ {p} ^ {t}\\right) - G _ {3 D M M} \\left(\\alpha_ {s} ^ {s \\rightarrow t}, \\alpha_ {e} ^ {s \\rightarrow t}, \\alpha_ {p} ^ {s \\rightarrow t}\\right)\\right\\| _ {2} \\tag {7}", + "image_path": "eaf7db71737f2c7a4255a28bb6b73176b0a08df5492ce2f510907c1d37be54d1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 351, + 515, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 515, + 362 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 515, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 351, + 515, + 362 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 351, + 515, + 362 + ], + "type": "text", + "content": " is the number of vertices of the face mesh." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": "The render term " + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "inline_equation", + "content": "L_{\\text{render}}" + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": " uses the " + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": " loss between two rendered images:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 392, + 410, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 410, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 410, + 406 + ], + "type": "interline_equation", + "content": "L _ {r e n d e r} \\left(\\alpha^ {s}, \\alpha^ {t}, \\alpha^ {s \\rightarrow t}\\right) =", + "image_path": "5ae7637e7c92b8106d578e801a555d72c7acbbb480f304e1ebb88dbe9a980c59.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 408, + 555, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 408, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 309, + 408, + 555, + 434 + ], + "type": "interline_equation", + "content": "\\left\\| R \\left(\\alpha_ {s} ^ {s}, \\alpha_ {e} ^ {t}, \\alpha_ {a} ^ {t}, \\alpha_ {i} ^ {t}, \\alpha_ {p} ^ {t}\\right) - R \\left(\\alpha_ {s} ^ {s \\rightarrow t}, \\alpha_ {e} ^ {s \\rightarrow t}, \\alpha_ {a} ^ {t}, \\alpha_ {i} ^ {s \\rightarrow t}, \\alpha_ {p} ^ {s \\rightarrow t}\\right)\\right\\| _ {1} \\tag {8}", + "image_path": "471dbb36ba906e442e295408f40262a001a9b06d970bfd864153f6609a67fb70.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 440, + 490, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 490, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 490, + 452 + ], + "type": "text", + "content": "3.3. Detailed Attribute Transformation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 458, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 517 + ], + "type": "text", + "content": "In the previous subsection, we transfer the attributes of the target face to the source face, including pose, expression, and illumination. In this subsection, to transfer attributes beyond the 3DMM expressive capabilities, we use DAT-MLP to transfer more detailed attributes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "content": "First, we project the target images into the " + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "inline_equation", + "content": "W+" + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "content": " space using the StyleGAN encoder to obtain the style code " + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "content": ". Then, DAT-MLP regresses " + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "content": " from style codes " + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "inline_equation", + "content": "w_{t}, w_{s \\to t}" + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "content": " using six-layer MLP. Finally, we use the following formula to get the latent vector " + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "inline_equation", + "content": "w_{s \\to t}^{\\prime}" + }, + { + "bbox": [ + 304, + 518, + 545, + 590 + ], + "type": "text", + "content": " that contains non-identity attributes of the target image:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 365, + 596, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 596, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 365, + 596, + 545, + 610 + ], + "type": "interline_equation", + "content": "w _ {s \\rightarrow t} ^ {\\prime} = \\lambda w _ {s \\rightarrow t} + (1 - \\lambda) w _ {t} \\tag {9}", + "image_path": "a074c94591900ace1fea7cc95980fcd656d79196242f2e9eb4189e68e5887910.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\lambda \\in R^{18\\times 512}" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " is constrained to be between 0 and 1 through the sigmoid activation layer. We want to transfer attributes of " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow t}^{\\prime}" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " as many as possible. The loss function of the second stage is as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 670, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 310, + 670, + 545, + 684 + ], + "type": "interline_equation", + "content": "L = L _ {i d} + \\varepsilon_ {p} L _ {p} + \\varepsilon_ {\\text {a t t r}} L _ {\\text {a t t r}} \\left(\\alpha^ {s}, \\alpha^ {t}, M \\left(G \\left(w _ {s \\rightarrow t} ^ {\\prime}\\right)\\right)\\right) \\tag {10}", + "image_path": "4d077723d3a4c03a237b27205975cf67032386d8bc31a711ca0724fab7a16a6d.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The loss function is divided into three items. The first item is face recognition loss, which is used to constrain" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "356" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "the identity of the generated face to be consistent with the source image. The second item " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " represents the perceptual distance between the generated image and the target image. It can help transfer residual attributes other than pose, expression, and illumination. The third term " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "L_{attr}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " constrains the attributes of the generated face consistent with the target image. " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\varepsilon_{p}, \\varepsilon_{attr}" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " represent the weight of the loss function." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": "The Eq. (9) shows that the style code " + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "w_{s \\to t}^{\\prime}" + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": " is in a bounding box of two style codes " + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "w_{s \\to t}" + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": ". We consider that most of the bounding box is still in the subspace of " + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{W}+" + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": " space because the attribute of two style codes " + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "w_{s \\to t}" + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 157, + 286, + 228 + ], + "type": "text", + "content": " is very close and generated image retains pore-level details in practice." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 235, + 129, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 129, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 129, + 247 + ], + "type": "text", + "content": "3.4. Blend Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 253, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 312 + ], + "type": "text", + "content": "This subsection introduces how to blend the generated image into the target image. We propose Masked Spatial Feature Transform (MSFT) module to fuse the feature. Unlike traditional Spatial Feature Transform (SFT) [44], MSFT only fuses the feature of the masked regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 313, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 286, + 445 + ], + "type": "text", + "content": "The detail of this module is shown in Fig. 3. The VG-GFace network [30] extracts various level features of the target image, which is injected into the StyleGAN through the MSFT module. We use Gaussian filtering to filter the masked image obtained by the face segmentation algorithm to make the boundary smoother. The StyleGAN feature is retained in the facial area, while the background area is from the mixed feature. We mix features of StyleGAN and VG-GFace using SFT. The background loss and the perceptual loss make the swapped image have the same background as the target image:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 453, + 286, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 453, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 133, + 453, + 286, + 466 + ], + "type": "interline_equation", + "content": "L = L _ {b} + \\varepsilon_ {p} L _ {p} \\tag {11}", + "image_path": "dc22ae85a8823c59b3261bf9afec13f3c36398aff028471243453d73dfacbffc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "inline_equation", + "content": "L_{b}" + }, + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "text", + "content": " is the background loss used to measure the difference in the background between the generated image and the target image. " + }, + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "text", + "content": " represents the perceptual distance between the generated image and the target image. " + }, + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "inline_equation", + "content": "\\varepsilon_{p}" + }, + { + "bbox": [ + 46, + 474, + 286, + 534 + ], + "type": "text", + "content": " is the weight of the loss function. The background loss is expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 542, + 286, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 542, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 118, + 542, + 286, + 555 + ], + "type": "interline_equation", + "content": "L _ {b} = \\left\\| M \\odot \\left(I _ {r} - I _ {t}\\right) \\right\\| \\tag {12}", + "image_path": "3405dbafd9f0baa5d76895400c5a030d14016c6e1191b83159d9a14bb2252c62.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 563, + 286, + 588 + ], + "type": "text", + "content": " is the binarized image of the background region obtained by face segmentation algorithm [2]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 597, + 123, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 123, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 123, + 611 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "As mentioned above, StyleIPSB has the following advantages: (1) StyleIPSB can represent various poses, expressions, and illumination properties. (2) When the style code is modified along one base, it only changes the specific attribute while remaining identity unchanged. (3) StyleIPSB can preserve the pore-level details. (4) StyleIPSB cooperates well with 3DMM to control facial attributes. Therefore, in this section, we mainly evaluate" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "the effectiveness of our method in the following aspects: (1) Evaluating the properties of StyleIPSB. (2) Evaluating the performance of 3DMM controlling facial attributes with StyleIPSB. (3) Comparison of face swapping results with other methods. First of all, we introduce the dataset and training detail." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 152, + 458, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 458, + 165 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 458, + 165 + ], + "type": "text", + "content": "4.1. Datasets and Training Detail" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 171, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 266 + ], + "type": "text", + "content": "Dataset. We used the FFHQ database [15] to train the 3DMM StyleGAN Mapping module. Flickr-Faces-HQ (FFHQ) consists of 70,000 high-quality face images at " + }, + { + "bbox": [ + 304, + 171, + 545, + 266 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 171, + 545, + 266 + ], + "type": "text", + "content": " resolution. CelebAMask-HQ [20] is a large-scale face image dataset with 30,000 high-resolution face images selected from the CelebA dataset by following CelebA-HQ. FaceForensics++ [32] is a forensics dataset consisting of 1,000 original video sequences." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 267, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 363 + ], + "type": "text", + "content": "Training environment and hyperparameter We train the network on a GTX 3090 using the Pytorch framework. In training, our optimizer is Adam, the weight decay is 5e-5, and the batch size is 2. In order to effectively train the above three stages, we first pre-train the three stages separately and finally train them together. In the three-stage pre-training, the learning rate is 2e-4; in the final together training stage, we adjust the learning rate to 2e-5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 371, + 460, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 460, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 460, + 384 + ], + "type": "text", + "content": "4.2. The Properties of StyleIPSB" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 390, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 533 + ], + "type": "text", + "content": "The representation and the disentanglement of StyleIPSB. We conduct experiments on the representation of StyleIPSB to evaluate whether it can represent various poses, expressions, and illumination properties while identity is unchanged. Fig. 4 shows that StyleIPSB can express pitch, yaw rotations, and lighting with different colors, directions, and intensities. In terms of expressions, StyleIPSB can not only control whether the mouth is open or not, the eyes are open or closed but also control different eyeball orientations and raising eyebrows. Supplementary material reveals the effect of modifying coordinates in different base directions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "To evaluate the disentanglement of StyleIPSB, we present the quantitative results, which show the identity change when manipulating the pose. As shown in Fig. 5, the horizontal axis represents the yaw angle between the anchored face and the edited image, and the vertical axis represents identity loss as defined in Eq. (2) between the anchored face and the edited image. Our StyleIPSB outperforms Ganspace [13], which is also a basis constructed in " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{W}+" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " space using unsupervised learning. We achieve comparable results with the supervised method. InterfaceGAN [34]. It needs a dataset that contains left and right-facing faces. StyleIPSB uses unsupervised learning and can represent more attributes than InterfaceGAN. Fig. 5 also shows that if we introduce identity loss into the distance metric when building StyleIPSB, the identity-preserving perfor" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "357" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 167 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 167 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 167 + ], + "type": "image", + "image_path": "1aa025ba8784bf394be246e01680491d6dbe57a2b39ae544324aee6023a8405e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 546, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 546, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 546, + 220 + ], + "type": "text", + "content": "Figure 4. StyleIPSB can modify various poses, expressions, and illumination by adjusting its coordinates while keeping the identity unchanged. The first row is the results of changing coordinates of pose and illumination basis. The second row shows the results of the changing coordinate of the expression basis." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 538, + 167 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 538, + 167 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 538, + 167 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 538, + 167 + ], + "type": "image", + "image_path": "44a43f0c63f5c4a7e169767e3b14210bcd67331ee3f2ad8acb67bd51a0b4df18.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 242, + 248, + 370 + ], + "blocks": [ + { + "bbox": [ + 86, + 242, + 248, + 370 + ], + "lines": [ + { + "bbox": [ + 86, + 242, + 248, + 370 + ], + "spans": [ + { + "bbox": [ + 86, + 242, + 248, + 370 + ], + "type": "image", + "image_path": "74d3cc1660c4dc68a20388f456e659b6cea66dd5801fa352102f74622f665620.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 380, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 47, + 380, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 288, + 403 + ], + "type": "text", + "content": "Figure 5. Effects of different methods of changing face pose attributes on face identity information" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 415, + 281, + 550 + ], + "blocks": [ + { + "bbox": [ + 53, + 415, + 281, + 550 + ], + "lines": [ + { + "bbox": [ + 53, + 415, + 281, + 550 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 281, + 550 + ], + "type": "image", + "image_path": "3c089c26ec461eae45fd799aa97ef7ab692ea01effddfcd2492fe8e31f100346.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 559, + 287, + 583 + ], + "lines": [ + { + "bbox": [ + 47, + 559, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 287, + 583 + ], + "type": "text", + "content": "Figure 6. We compare our method with GANSpace [13] and InterfaceGAN [34] in the editing of facial expression." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 184, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 184, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 184, + 616 + ], + "type": "text", + "content": "mance of StyleIPSB is enhanced." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 700 + ], + "type": "text", + "content": "Editing results of StyleIPSB. We compare StyleIPSB with other StyleGAN-based face edit methods. Fig. 6 compares the edit results of expression (smile) with other methods. The results show that our approach has better performance in identity-preserving. StyleIPSB preserves the face shape when expression changes, while face shape is deformed in other methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 714 + ], + "type": "text", + "content": "Fig. 7 shows examples of editing a face image using the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 239, + 547, + 323 + ], + "blocks": [ + { + "bbox": [ + 308, + 239, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 308, + 239, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 547, + 323 + ], + "type": "image", + "image_path": "e9c8577d075d33d9cfce7f2509e21022b6b6d20d31d6e8b831b7d46354e21428.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 332, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 305, + 332, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 546, + 376 + ], + "type": "text", + "content": "Figure 7. 3D face models control the attribute of generated images. The top is the rendered images of 3D face geometry. The bottom is the edited images controlled by the 3DMM parameters of the top 3D face geometry." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 398, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 398, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 545, + 471 + ], + "type": "text", + "content": "3DMM parameters of FLAME [23]. The first row is the 3D face models generated by the 3DMM parameters. The second row is the edited image controlled by those 3DMM parameters. These results show that StyleIPSB cooperates well with 3DMM to control the facial attributes of the generated image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 479, + 444, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 444, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 444, + 493 + ], + "type": "text", + "content": "4.3. Result of Face Swapping" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 497, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 556 + ], + "type": "text", + "content": "This subsection evaluates the performance of face swapping based on StyleIPSB. First, the ablation experiments evaluate the influence of StyleIPSB on attribute transfer performance. Then, we compare our face swapping with other methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "text", + "content": "Ablation experiment In the experiment, \"no basis\" means that the 3DMM-StyleGAN mapping network directly regresses the " + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow t}" + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\beta_{s\\rightarrow t}" + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "text", + "content": ". As shown in Fig. 8, using StyleIPSB can constrain the generated style code on the subspace of " + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 305, + 557, + 545, + 629 + ], + "type": "text", + "content": " space and preserve the pore-level details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Tab. 2 shows the quantitative comparison results on the performance of StyleIPSB in stage one. FID (Fréchet inception distance) is used to evaluate the image quality generated by the image generation model. \"Exp\" and \"Pose\" represent the Euclidean distance between the target image and the transferred image expression and pose parameters, respectively. ID Similarity means the cosine similarity be" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "358" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 292, + 126 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 292, + 126 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 292, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 292, + 126 + ], + "type": "image", + "image_path": "64c5a4b960f30cf8e918e02bb555b5c6976bb01effbe381d4e1f84908d7ce792.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 134, + 287, + 157 + ], + "lines": [ + { + "bbox": [ + 47, + 134, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 287, + 157 + ], + "type": "text", + "content": "Figure 8. The qualitative ablation experiments on StyleIPSB. The results show we can generate pore-level details with StyleIPSB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 171, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 49, + 171, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 49, + 171, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 286, + 223 + ], + "type": "table", + "html": "
FID ↓Exp ↓Pose ↓ID similarity ↑
No Basis26.063.730.0740.65
With basis22.153.370.0780.67
", + "image_path": "1e45ae4dd88c9f2d377d6407544f009e2b20ac8a3e7d6b6acebc983304e24df0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 399 + ], + "type": "text", + "content": "tween the identity embedding of the source image and the transferred image. D3FR [9] is used to extract the 3D face parameters, and CosFace [43] is used to extract the identity embedding. The test set is 10,000 image pairs randomly sampled from the CelebAMask-HQ [20]. The Tab. 2 shows that using StyleIPSB can effectively reduce the value of FID because StyleIPSB is in the subspace of " + }, + { + "bbox": [ + 46, + 267, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 267, + 287, + 399 + ], + "type": "text", + "content": " space. In addition, the performance of transferring expressions with StyleIPSB is better in identity-preserving because the identity is preserved while changing the coordinates of StyleIPSB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "content": "Tab. 3 shows the quantitative comparison results of different methods on the FaceForensics++ dataset. The dataset is classified into 885 identities, and the identity retrieval indicates the top-1 matching rate of swapped and the source image. We apply CosFace [43] to extract identity embedding. Expressions are measured in the same way as Tab. 2, The pose measurement method uses a pose estimator [33]. We compare our results with FaceSwap [3], DeepFakes [1], FaceShifter [22], MegaFace [51] and FSLDS [47]. As shown in the Tab. 3, our method outperforms in identity-preserving and expression transfer thanks to StyleIPSB's identity-preserving property and the ability to present various expressions. Our performance of pose transfer does not outperform other methods, which may be because StyleIPSB only contains two bases expressing poses in raw and pitch directions, which makes it hard to express very accurate poses. However, the error of the pose is only a few degrees, so our results still look accurate visually." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Fig. 9 shows the results of face swapping under challenging conditions. The results show that our method can preserve the unique noise patterns in portraiture in the red boxes. In the second row of the figure, our method successfully keeps pore-level details and identity while the source and target pose is quite different. Tab. 4 shows that our method has the better image quality and identity preservation ability than other StyleGAN-based methods. The met-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 69, + 566, + 157 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 278, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 278, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 278, + 243 + ], + "type": "text", + "content": "Table 2. The quantitative ablation experiments on StyleIPSB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 69, + 566, + 157 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 566, + 157 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 566, + 157 + ], + "type": "table", + "html": "
MethodID Retri.(%) ↑Exp Err. ↓Pose Err. ↓
FaceSwap [3]72.692.892.58
Deepfakes [1]88.393.334.64
FaceShifter [22]90.682.822.55
MegaFS [51]90.832.922.64
FSLDS [47]90.052.792.46
Ours95.052.233.58
", + "image_path": "9b5a1cd0dad706ba3016ed74a373f0330e62689c904edb69639068e1c8fbecdb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 199, + 555, + 275 + ], + "blocks": [ + { + "bbox": [ + 306, + 165, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 306, + 165, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 165, + 545, + 186 + ], + "type": "text", + "content": "Table 3. The quantitative experiments on FaceForensics++ dataset with other methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 199, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 307, + 199, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 555, + 275 + ], + "type": "table", + "html": "
FID ↓Exp ↓Pose ↓ID similarity ↑
MageFS [51]22.032.850.0430.4837
RAFS [46]13.253.15-0.5232
FSLDS [47]10.012.990.0530.4761
Ours9.372.750.0780.5378
", + "image_path": "2ded8812dfc81488de538a16dfd65610e613944713eff49e9a5fb321e16f224f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 283, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 306, + 283, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 545, + 304 + ], + "type": "text", + "content": "Table 4. The quantitative experiments on the CelebAHQ dataset with other methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 316, + 550, + 406 + ], + "blocks": [ + { + "bbox": [ + 307, + 316, + 550, + 406 + ], + "lines": [ + { + "bbox": [ + 307, + 316, + 550, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 550, + 406 + ], + "type": "image", + "image_path": "e9e7794e558308f9f9ed6750e05305bc63718632f5b667951318ebcf5e70593e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 414, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 306, + 414, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 545, + 437 + ], + "type": "text", + "content": "Figure 9. We compare our method with MageFS [51] and FSLDS [47] in challenging conditions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 448, + 545, + 510 + ], + "blocks": [ + { + "bbox": [ + 308, + 448, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 308, + 448, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 448, + 545, + 510 + ], + "type": "image", + "image_path": "07e01d192ddee6770729044c541eadfe73a4f218c9d244cc613347cd446d194b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 518, + 492, + 529 + ], + "lines": [ + { + "bbox": [ + 360, + 518, + 492, + 529 + ], + "spans": [ + { + "bbox": [ + 360, + 518, + 492, + 529 + ], + "type": "text", + "content": "Figure 10. limitation of our method" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 551, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 545, + 575 + ], + "type": "text", + "content": "ric is the same as in Tab. 2. We follow RAFS and compare " + }, + { + "bbox": [ + 306, + 551, + 545, + 575 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 306, + 551, + 545, + 575 + ], + "type": "text", + "content": " swapped faces for a fair comparison." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 586, + 458, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 458, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 458, + 597 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We have developed a new semantic basis for face swapping, called StyleIPSB, that is specifically designed to preserve identity and pore-level details. Our experiments have demonstrated that StyleIPSB outperforms other state-of-the-art methods. Despite this, there is still potential for further improvement, as shown in Fig. 10. (1) Occlusion is limited by the mask. (2) The glasses in the source image cannot be removed. (3) Light and shadow cannot be perfectly restored in the case of complex illumination." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "359" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Deepfakes. https://github.com/ondyari/FaceForensics/tree/master/dataset/DeepFakes.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 146 + ], + "type": "text", + "content": "[2] face-parsing.pytorch. https://github.com/zllrunning/face-parsing.PyTorch.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 179 + ], + "type": "text", + "content": "[3] Faceswap. https://github.com/ondyari/FaceForensics/tree/master/dataset/FaceSwapKowalski.8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 181, + 286, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 286, + 224 + ], + "type": "text", + "content": "[4] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2StyleGAN: Unsupervised extraction of StyleGAN edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 225, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 286, + 278 + ], + "type": "text", + "content": "[5] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (TOG), 40(3):1-21, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 286, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 286, + 324 + ], + "type": "text", + "content": "[6] Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, and Gang Hua. Towards open-set identity preserving face synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6713-6722, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 326, + 286, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 286, + 369 + ], + "type": "text", + "content": "[7] Renwang Chen, Xuanhong Chen, Bingbing Ni, and Yanhao Ge. Simswap: An efficient framework for high fidelity face swapping. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2003-2011, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 370, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 286, + 423 + ], + "type": "text", + "content": "[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 286, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 286, + 489 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 286, + 489 + ], + "type": "text", + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 534 + ], + "type": "text", + "content": "[10] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 590 + ], + "type": "text", + "content": "[11] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "text", + "content": "[12] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. GIF: Generative interpretable faces. In 2020 International Conference on 3D Vision (3DV), pages 868-878. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 668 + ], + "type": "text", + "content": "[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANspace: Discovering interpretable GAN controls. arXiv preprint arXiv:2004.02546, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[17] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 292 + ], + "type": "text", + "content": "[18] Jiseob Kim, Jihoon Lee, and Byoung-Tak Zhang. Smooth-Swap: A simple enhancement for face-swapping with smoothness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10779-10788, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 293, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 337 + ], + "type": "text", + "content": "[19] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 3677-3685, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[20] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 435 + ], + "type": "text", + "content": "[21] Jia Li, Zhaoyang Li, Jie Cao, Xingguang Song, and Ran He. FaceInpainter: High fidelity face adaptation to heterogeneous domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5089-5098, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[22] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 2, 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[23] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[24] Yuan Lin, Shengjin Wang, Qian Lin, and Feng Tang. Face swapping under large pose variations: A 3d model based approach. In 2012 IEEE International Conference on Multimedia and Expo, pages 333-338. IEEE, 2012. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[25] Luming Ma and Zhigang Deng. Real-time hierarchical facial performance capture. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 1-10, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "text", + "content": "[26] Luming Ma and Zhigang Deng. Real-time face video swapping from a single portrait. In Symposium on Interactive 3D Graphics and Games, pages 1-10, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[27] Yuval Nirkin, Yosi Keller, and Tal Hassner. Fsgan: Subject agnostic face swapping and reenactment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7184-7193, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[28] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. arXiv preprint arXiv:2202.12972, 2022. 2, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "360" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[29] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "text", + "content": "[30] Omkar M Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In Proceedings of the British Machine Vision Conference (BMVC). British Machine Vision Association, 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[31] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleClip: Text-driven manipulation of StyleGAN imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[32] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Eunji Chong, and James M Rehg. Fine-grained head pose estimation without keypoints. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 2074-2083, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[34] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[35] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: Interpreting the disentangled face representation learned by GANs. IEEE transactions on pattern analysis and machine intelligence, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 481 + ], + "type": "text", + "content": "[36] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "text", + "content": "[37] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. SemanticStyleGAN: Learning compositional generative priors for controllable image synthesis and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11254-11264, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "text", + "content": "[38] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "type": "text", + "content": "[39] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. StyleRig: Rigging StyleGAN for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[40] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of RGB videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 631 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[41] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[42] Binxu Wang and Carlos R Ponce. The geometry of deep generative image models and its applications. arXiv preprint arXiv:2101.06006, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[43] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 262 + ], + "type": "text", + "content": "[44] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 263, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 317 + ], + "type": "text", + "content": "[45] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. arXiv preprint arXiv:2106.09965, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 363 + ], + "type": "text", + "content": "[46] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 364, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 364, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 545, + 419 + ], + "type": "text", + "content": "[47] Yangyang Xu, Bailin Deng, Junle Wang, Yanqing Jing, Jia Pan, and Shengfeng He. High-resolution face swapping via latent semantics disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7642-7651, 2022. 1, 2, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 485 + ], + "type": "text", + "content": "[48] Yanbo Xu, Yueqin Yin, Liming Jiang, Qianyi Wu, Chengyao Zheng, Chen Change Loy, Bo Dai, and Wayne Wu. TransEditor: Transformer-based dual-space GAN for highly controllable facial editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7683-7692, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 486, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 545, + 530 + ], + "type": "text", + "content": "[49] Zhiliang Xu, Zhibin Hong, Changxing Ding, Zhen Zhu, Junyu Han, Jingtuo Liu, and Errui Ding. MobileFaceSwap: A lightweight framework for video face swapping. arXiv preprint arXiv:2201.03808, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 531, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 531, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 531, + 545, + 586 + ], + "type": "text", + "content": "[50] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. StyleHeat: One-shot high-resolution editable talking face generation via pretrained StyleGAN. arXiv preprint arXiv:2203.04036, 2022.1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 586, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 545, + 631 + ], + "type": "text", + "content": "[51] Yuhao Zhu, Qi Li, Jian Wang, Cheng-Zhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4834-4844, 2021. 1, 2, 3, 8" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "361" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_content_list.json b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2cdf3f3abfc5983f07b7dd5106f6ff48e0ec391b --- /dev/null +++ b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_content_list.json @@ -0,0 +1,1689 @@ +[ + { + "type": "text", + "text": "StyleRF: Zero-shot 3D Style Transfer of Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 161, + 130, + 808, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kunhao Liu $^{1}$ Fangneng Zhan $^{2}$ Yiwen Chen $^{1}$ Jiahui Zhang $^{1}$", + "bbox": [ + 222, + 165, + 743, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingchen Yu1 Abdulmotaleb El Saddik3,5 Shijian Lu1* Eric Xing4,5", + "bbox": [ + 189, + 185, + 779, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanyang Technological University $^{2}$ Max Planck Institute for Informatics", + "bbox": [ + 258, + 212, + 710, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Ottawa 4Carnegie Mellon University 5MBZUAI", + "bbox": [ + 282, + 229, + 681, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9535ab4afae93e92cde991a6fd7118d091507d72044ea8a75fac6ea3747ff12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 272, + 279, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/206733db5d51c349e342db393d70e334782f4ca31661a7f780c24b649e3d50f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 273, + 496, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce57bbde87f035ae8eac471049f40853abb79ac1ec8960262aa8d4647e9f2a30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 273, + 696, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5871b8bdcf5616393f246b73870d6a3cdade9a12310ce0abc216dbb591b2daf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 268, + 888, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e9414d224bb9426eb476c87148eb153b0f21ab67b97060e53e771ab9632bfdb.jpg", + "image_caption": [ + "Multi-view Content Images" + ], + "image_footnote": [], + "bbox": [ + 81, + 396, + 310, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bef773278881f74423435479aa4582b7f12771053d7461efd4aeccbbd671a711.jpg", + "image_caption": [ + "Reference Style Images + Stylized Novel Views" + ], + "image_footnote": [], + "bbox": [ + 313, + 411, + 501, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2956a5060b129f79a16e54be3b673dcf3013ac1d6db6926ccc1a1cf942f43381.jpg", + "image_caption": [ + "Figure 1. Zero-shot 3D Style Transfer. Given a set of Multi-view Content Images of a 3D scene, StyleRF can transfer arbitrary Reference Styles to the 3D scene in a zero-shot manner, rendering high-quality Stylized Novel Views with superb multi-view consistency." + ], + "image_footnote": [], + "bbox": [ + 504, + 411, + 694, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59ebdac68649afb08d4027be86917ffce20a35ece0784cbfa339339b2cf61a29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 411, + 887, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 604, + 313, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D style transfer aims to render stylized novel views of a 3D scene with multi-view consistency. However, most existing work suffers from a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. We propose StyleRF (Style Radiance Fields), an innovative 3D style transfer technique that resolves the three-way dilemma by performing style transformation within the feature space of a radiance field. StyleRF employs an explicit grid of high-level features to represent 3D scenes, with which high-fidelity geometry can be reliably restored via volume rendering. In addition, it transforms the grid features according to the reference style which directly leads to high-quality zero-shot style transfer. StyleRF consists of two innovative designs. The first is sampling-invariant content transformation that makes the transformation invari", + "bbox": [ + 73, + 635, + 473, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ant to the holistic statistics of the sampled 3D points and accordingly ensures multi-view consistency. The second is deferred style transformation of 2D feature maps which is equivalent to the transformation of 3D points but greatly reduces memory footprint without degrading multi-view consistency. Extensive experiments show that StyleRF achieves superior 3D stylization quality with precise geometry reconstruction and it can generalize to various new styles in a zero-shot manner. Project website: https://kunhao-liu.github.io/StyleRF/", + "bbox": [ + 500, + 604, + 893, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 781, + 632, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a set of multi-view images of a 3D scene and an image capturing a target style, 3D style transfer aims to generate novel views of the 3D scene that have the target style consistently across the generated views (Fig. 1). Neural style transfer has been investigated extensively, and state-of-the-art methods allow transferring arbitrary styles in a", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Shijian Lu is the corresponding author.", + "bbox": [ + 94, + 886, + 308, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zero-shot manner. However, most existing work focuses on style transfer across 2D images [15, 21, 24] but cannot extend to a 3D scene that has arbitrary new views. Prior studies [19, 22, 37, 39] have shown that naively combining 3D novel view synthesis and 2D style transfer often leads to multi-view inconsistency or poor stylization quality, and 3D style transfer should optimize novel view synthesis and style transfer jointly.", + "bbox": [ + 75, + 90, + 470, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the current 3D style transfer is facing a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to new styles. Different approaches have been investigated to resolve the three-way dilemma. For example, multiple style transfer [11, 22] requires a set of pre-defined styles but cannot generalize to unseen new styles. Point-cloud-based style transfer [19, 37] requires a pre-trained depth estimation module that is prone to inaccurate geometry reconstruction. Zero-shot style transfer with neural radiance fields (NeRF) [8] cannot capture detailed style patterns and textures as it implicitly injects the style information into neural network parameters. Optimization-based style transfer [17, 39, 63] suffers from slow optimization and cannot scale with new styles.", + "bbox": [ + 75, + 214, + 468, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce StyleRF to resolve the three-way dilemma by performing style transformation in the feature space of a radiance field. A radiance field is a continuous volume that can restore more precise geometry than point clouds or meshes. In addition, transforming a radiance field in the feature space is more expressive with better stylization quality than implicit methods [8], and it can also generalize to arbitrary styles. We construct a 3D scene representation with a grid of deep features to enable feature transformation. In addition, multi-view consistent style transformation in the feature space could be achieved by either transforming the whole feature grid or transforming the sampled 3D points. We adopt the latter as the former incurs much more computational cost during training to stylize the whole feature grid in every iteration, whereas the latter can reduce computational cost through decreasing the size of training patch and the number of sampled points. However, applying off-the-shelf style transformations to a batch of sampled 3D points impairs the multi-view consistency as they are conditioned on the holistic statistics of the batch. Beyond that, transforming every sampled 3D point is memory-intensive since NeRF needs to query hundreds of sampled points along each ray for rendering a single pixel.", + "bbox": [ + 75, + 444, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST), the former eliminating the dependency on holistic statistics of sampled point batch and the latter deferring style transformation to 2D feature maps for better efficiency. In SICT, we introduce volume-adaptive normalization that learns the mean and variance of the", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "whole volume instead of computing them from a sampled batch. In addition, we apply channel-wise self-attention to transform each 3D point independently to make it conditioned on the feature of that point regardless of the holistic statistics of the sampled batch. In DST, we defer the style transformation to the volume-rendered 2D feature maps based on the observation that the style transformation of each point is the same. By formulating the style transformation by pure matrix multiplication and adaptive bias addition, transforming 2D feature maps is mathematically equivalent to transforming 3D point features but it saves computation and memory greatly. Thanks to the memory-efficient representation of 3D scenes and deferred style transformation, our network can train with $256 \\times 256$ patches directly without requiring sub-sampling like previous NeRF-based 3D style transfer methods [8, 11, 22].", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work can be summarized in three aspects. First, we introduce StyleRF, an innovative zero-shot 3D style transfer framework that can generate zero-shot high-quality 3D stylization via style transformation within the feature space of a radiance field. Second, we design sampling-invariant content transformation and deferred style transformation, the former achieving multiview consistent transformation by eliminating dependency on holistic statistics of sampled point batch while the latter greatly improves stylization efficiency by deferring style transformation to 2D feature maps. Third, extensive experiments show that StyleRF achieves superior 3D style transfer with accurate geometry reconstruction, high-quality stylization, and great generalization to new styles.", + "bbox": [ + 496, + 333, + 890, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 558, + 640, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural scene representations. 3D scene representation has been extensively studied in recent years with different ways of representations such as volumes [23,27,46,49,59], point clouds [1, 45], meshes [25, 55], depth maps [20, 30], and implicit functions [7, 34, 41, 61]. These methods adopt differentiable rendering which enables model optimization by using 2D multi-view images. Among them, Neural Radiance Field (NeRF) [36] can render a complex 3D scene with high fidelity and accurate geometry. It represents scenes with an implicit coordinate function that maps each 3D coordinate to a density value and a color value, and employs volume rendering to generate images of novel views. However, the implicit coordinate function is represented by a large multilayer perceptron (MLP) that is often hard to optimize and slow to infer. Serval studies adopt a hybrid representation [3, 4, 10, 14, 31, 33, 38, 44, 52, 62] to speed up the reconstruction and rendering. They employ explicit data structures such as discrete voxel grids [14,52], decomposed tensors [3, 4, 13], hash maps [38], etc. to store features or spherical harmonics, enabling fast convergence and inference.", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63c43f0464217d0a98f4d7a93a99827ff2252e30e31db7f33fadfe04ceff4228.jpg", + "image_caption": [ + "Figure 2. The framework of StyleRF. For a batch of sampled points along a ray $\\mathbf{r}$ , the corresponding features $F_{i}, i \\in [1,2,\\dots,N]$ are first extracted, each of which is transformed to $\\bar{F}_{i}$ independently via Sampling-Invariant Content Transformation, regardless of the holistic statistics of the point batch. $\\bar{F}_{i}$ is then transformed to a feature map $\\bar{F}_{c}$ via Volume Rendering. After that, the Deferred Style Transformation transforms $\\bar{F}_{c}$ to the feature map $F_{cs}$ adaptively using the sum weight of the sampled points $w_{\\mathbf{r}}$ along the ray $\\mathbf{r}$ and the style information $T, \\mu(F_{s})$ , and $\\sigma(F_{s})$ . Finally, a stylized novel view is generated via a CNN decoder." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 880, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although most existing work extracts features as middle-level representations of scenes, the extracted features are usually an intermediate output of neural networks which have little semantic meanings and are not suitable for the style transfer task. We introduce decomposed tensors [4] to store high-level features extracted by pre-trained CNNs, which enables transformations in feature space as well as efficient training and inference. Though [3,16,40] also render feature maps instead of RGB maps, they are computationally intensive and usually work with low-resolution feature maps. StyleRF can instead render full-resolution feature maps (the same as the output RGB images) efficiently and it uses high-level features largely for transformation only.", + "bbox": [ + 75, + 339, + 472, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural style transfer. Neural style transfer aims at rendering a new image that contains the content structure of one image and the style patterns of another. The seminal work in [15] shows that multi-level feature statistics extracted from intermediate layers of pre-trained CNNs could be used as a representation of the style of an artistic image, but it treats style transfer as a slow and iterative optimization task. [9,21,24,28,29,32,43,50,58] utilize feed-forward networks to approximate the optimization procedure to speed up rendering. Among them, [9,21,28,29,32,43,50,58] can achieve zero-shot style transfer by applying transformations to the high-level features extracted by pre-trained CNNs, where the feature transformations can be achieved by matching second-order statistics [21, 29], linear transformation [28,58], self-attention transformation [9,32,43], etc. Video style transfer extends style transfer to videos for injecting target styles consistently across adjacent video frames. Several studies leverage optical flow [5,18,47,56,57] as temporal constraints to estimate the movement of video contents. They can produce smooth videos, but have little knowledge of the underlying 3D geometry and cannot render consistent frames in arbitrary views [19,37].", + "bbox": [ + 75, + 537, + 472, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Huang et al. first tackle stylizing complex 3D scenes [19]. They construct a 3D scene by back-projecting", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image features into the 3D space to form a point cloud and then perform style transformation on the features of 3D points. Their method can achieve zero-shot style transfer, but requires an error-prone pre-trained depth estimator to model scene geometry. [37] also constructs a point cloud for stylization but it mainly focuses on monocular images. Instead, [6, 8, 11, 22, 39, 63] use NeRF [36] as the 3D representation which can reconstruct scene geometry more faithfully. [6] is a photorealistic style transfer method that can only transfer the color tone of style images. [39, 63] achieve 3D style transfer via optimization and can produce visually high-quality stylization, but they require a time-consuming optimization procedure for every reference style. [11, 22] employ latent codes to represent a set of pre-defined styles, but cannot generalize to unseen styles. [8] can achieve arbitrary style transfer by implicitly instilling the style information into MLP parameters. However, it can only transfer the color tone of style images but cannot capture detailed style patterns. StyleRF can transfer arbitrary style in a zero-shot manner, and it can capture style details such as strokes and textures as well.", + "bbox": [ + 496, + 338, + 893, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 676, + 591, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of StyleRF is shown in Fig. 2. For a batch of sampled points along a ray $\\mathbf{r}$ , the corresponding features $F_{i}, i \\in [1,2,\\dots,N]$ are first extracted from the feature grid described in Sec. 3.1, each of which is transformed to $\\bar{F}_{i}$ independently via Sampling-Invariant Content Transformation (SICT) described in Sec. 3.2.1, regardless of the holistic statistics of the point batch. $\\bar{F}_{i}$ is then transformed to a feature map $\\bar{F}_{c}$ via Volume Rendering. After that, the Deferred Style Transformation (DST) described in Sec. 3.2.2 transforms $\\bar{F}_{c}$ to the feature map $F_{cs}$ adaptively using the sum weight of the sampled points $w_{\\mathbf{r}}$ along the ray $\\mathbf{r}$ and the style information $T, \\mu(F_{s})$ , and $\\sigma(F_{s})$ . Finally, a stylized novel view is generated via a CNN decoder.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8340", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Grid 3D Representation", + "text_level": 1, + "bbox": [ + 76, + 90, + 361, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model a 3D scene with deep features, we use a continuous volumetric field of density and radiance. Different from the original NeRF [36], for every queried 3D position $x \\in \\mathbb{R}^3$ , we get a volume density $\\sigma(x)$ and a multi-channel feature $F(x) \\in \\mathbb{R}^C$ instead of an RGB color, where $C$ is the number of the feature channels. Then we can get the feature of any rays $\\mathbf{r}$ passing through the volume by integrating sampled points along the ray via approximated volume rendering [36]:", + "bbox": [ + 75, + 113, + 470, + 250 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} w _ {i} F _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 258, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i} = \\exp \\left(-\\sum_{j = 1}^{i - 1}\\sigma_{j}\\delta_{j}\\right)$ $(1 - \\exp (-\\sigma_i\\delta_i))$ (2)", + "bbox": [ + 91, + 309, + 468, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{i}, F_{i}$ denotes the volume density and feature of sampled point $i$ , $w_{i}$ denotes the weight of $F_{i}$ in the ray $\\mathbf{r}$ , and $\\delta_{i}$ is the distance between adjacent samples. We disable view-dependency effect for better multi-view consistency.", + "bbox": [ + 75, + 369, + 468, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we can generate feature maps capturing high-level features and map them to RGB space using a 2D CNN decoder. However, unlike [3,16,40], we render full-resolution feature maps which have the same resolution as the final RGB images rather than down-sampled feature maps. Rendering full-resolution feature maps has two unique features: 1) it discards up-sampling operations which cause multiview inconsistency in general [16], 2) it removes aliasing when rendering low-resolution feature maps [2] which causes severe flickering effects in stylized RGB videos.", + "bbox": [ + 75, + 431, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Directly using 3D voxel grid to store features is memory-intensive. We thus adopt vector-matrix tensor decomposition [4] that relaxes the low-rank constraints for two modes of a 3D tensor and factorizes tensors into compact vector and matrix factors, which lowers the space complexity from $\\mathcal{O}(n^3)$ to $\\mathcal{O}(n^2)$ , massively reducing the memory footprint. We employ a density grid to store volume density and a feature grid to store multi-channel features respectively.", + "bbox": [ + 75, + 580, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Feature Transformation for Style Transfer", + "text_level": 1, + "bbox": [ + 76, + 710, + 437, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we have the feature grid representation of a scene, we can tackle the task of stylizing 3D scenes. Given a reference style image, our goal is to render stylized novel views of the 3D scene with multi-view consistency. To achieve this, we apply transformations to the features of the grid.", + "bbox": [ + 75, + 734, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One plausible solution to this task is to apply style transfer to the feature grid directly. This solution is efficient in evaluations as it can render any stylized views with a single style transfer process only. However, it is impractical to train such transformation as it needs to stylize the whole feature grid in every iteration. Another solution is", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88920ccc6bd8c9b413e7be146402b7e0097d6e49ddd7d117819e8c001ec9822a.jpg", + "image_caption": [ + "(a) Vanilla IN." + ], + "image_footnote": [], + "bbox": [ + 539, + 88, + 705, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d3d63ecf03bf88cb02ea408e43c8c83ded6f017991cfecd795dc26e9f49e533.jpg", + "image_caption": [ + "(b) Volume-adaptive IN.", + "Figure 3. Comparison between vanilla instance normalization (IN) in (a) and volume-adaptive IN in (b). During evaluation, volume-adaptive IN uses learned mean and standard-deviation, discarding dependency over the sampled point batch's holistic statistics (indicated by the red arrows in the left graph)." + ], + "image_footnote": [], + "bbox": [ + 728, + 87, + 885, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to apply an off-the-shelf zero-shot style transfer method to the features of the sampled 3D points. While this solution can reduce computational cost through decreasing the size of training patch and the number of sampled points, it has two problems: 1) vanilla zero-shot style transformation is conditioned on holistic statistics of the sampled point batch [21, 28, 32], which violates multi-view consistency in volume rendering as the feature transformation of a specific 3D point will vary across different sampled points; 2) volume rendering requires sampling hundreds of points along a single ray, which makes transformation on the point batch memory-intensive.", + "bbox": [ + 496, + 306, + 890, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the observation that style transformation is conditioned on both content information and style information, we decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST). After the decomposition, SICT will be conditioned solely on the content information while DST conditioned solely on the style information, more details to be elaborated in the ensuing subsections.", + "bbox": [ + 496, + 488, + 890, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Sampling-invariant Content Transformation", + "text_level": 1, + "bbox": [ + 500, + 621, + 862, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a batch of sampled points, we can get their corresponding features $F_{i} \\in \\mathbb{R}^{C}, i \\in [1,2,\\dots,N]$ from the feature grid, where $N$ is the number of the sampled points along a ray and $C$ is the number of the feature channels. The goal of SICT is to transform the extracted features $F_{i}$ so that they can be better stylized. We formulate SICT as a channel-wise self-attention operation to the features after instance normalization (IN) [54]. Specifically, we formulate $Q(\\text{query}), K(\\text{key})$ , and $V(\\text{value})$ as:", + "bbox": [ + 496, + 643, + 890, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = q \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 791, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK = k \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 810, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = v \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 828, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q, k, v$ are $1 \\times 1$ convolution layers which reduce the channel number from $C$ to $C'$ for computational efficiency, and Norm denotes the IN. However, as shown in Fig. 3,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8341", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/212ea898387ba8c5d3e5d7e2190e1dce3eb3444f467bcf3b2e6bd0372ff953b6.jpg", + "image_caption": [ + "Figure 4. Deferred style transformation. We apply the style transformation to the volume-rendered feature maps $\\bar{F}_c$ according to the style feature maps $F_s$ . To ensure multi-view consistency, we modulate the bias (e.g. the mean value of the style feature maps $\\mu(F_s)$ ) with the sum weight of sampled points along each ray $w_{\\mathbf{r}}$ ." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 433, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vanilla IN calculates per-dimension mean and standard-deviation of the batch of sampled points, which varies with different sampled points and incurs multi-view inconsistency accordingly. Thus we design volume-adaptive IN which, during training, keeps running estimates of the computed mean and standard-deviation, and uses them for normalization during evaluations (instead of computing from the sampled point batch). Through volume-adaptive IN, we can ensure that the content transformation is consistent regardless of the sampled point batch's holistic statistics.", + "bbox": [ + 75, + 380, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Channel-wise self-attention can thus be implemented by:", + "bbox": [ + 96, + 532, + 467, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {F} _ {i} = V \\otimes \\operatorname {S o f t m a x} (\\widetilde {\\operatorname {c o v}} (Q, K)), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 563, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes matrix multiplication and $\\widetilde{cov}(Q, K) \\in \\mathbb{R}^{N \\times C' \\times C'}$ denotes the covariance matrix in the channel dimension.", + "bbox": [ + 76, + 588, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Deferred Style Transformation", + "text_level": 1, + "bbox": [ + 76, + 648, + 346, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After applying SICT to the features of each 3D point, we apply DST to the volume-rendered 2D feature maps $\\bar{F}_c$ rather than 3D point features $\\bar{F}_i$ . To ensure multi-view consistency, we formulate the transformation as matrix multiplication and adaptive bias addition as illustrated in Fig. 4.", + "bbox": [ + 75, + 672, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first extract feature maps $F_{s}$ of the reference style $S$ using a pre-trained VGG [51], and then generate the style transformation matrix $T \\in \\mathbb{R}^{C' \\times C'}$ using feature covariance $cov(F_{s})$ following [28]. Next, we apply matrix multiplication with $T$ to the feature maps $\\bar{F}_{c}$ and use a $1 \\times 1$ convolution layer conv without bias to restore the channel number from $C'$ to $C$ . Though these operations can partially instill style information, they are not expressive enough without bias addition containing style information [58]. Thus following [21], we multiply the feature", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maps with the standard-deviation value $\\sigma(F_s)$ and add the mean value $\\mu(F_s)$ . To ensure it is equivalent when applying the transformation to either 3D point features or 2D feature maps, we adaptively modulate the mean value $\\mu(F_s)$ with the sum weight of sampled points along each ray $w_{\\mathbf{r}}$ . DST can be mathematically formulated by:", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = \\operatorname {c o n v} \\left(T \\otimes \\bar {F} _ {c}\\right) \\times \\sigma \\left(F _ {s}\\right) + w _ {\\mathbf {r}} \\times \\mu \\left(F _ {s}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 193, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad \\bar {F} _ {c} = \\sum_ {i = 1} ^ {N} w _ {i} \\bar {F} _ {i}, w _ {\\mathbf {r}} = \\sum_ {i = 1} ^ {N} w _ {i}, \\mathbf {r} \\in \\mathcal {R} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 222, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{i}$ denotes the weight of sampled point $i$ (Eq. (2)), $\\bar{F}_i$ denotes the feature of sample $i$ after SICT, and $\\mathcal{R}$ is the set of rays in each training batch.", + "bbox": [ + 500, + 268, + 890, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note $conv$ is a $1 \\times 1$ convolution layer without bias, so it is basically a matrix multiplication operation. And $\\sigma(S), \\mu(S)$ are scalars. Together with the adaptive bias modulation $w_{\\mathbf{r}}$ , Eq. (7) can be reformulated by:", + "bbox": [ + 498, + 315, + 890, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = \\sum_ {i = 1} ^ {N} w _ {i} \\left(\\underbrace {\\operatorname {c o n v} \\left(T \\otimes \\bar {F} _ {i}\\right) \\times \\sigma \\left(F _ {s}\\right) + \\mu \\left(F _ {s}\\right)} _ {\\text {(i)}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 386, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where part (i) can be seen as applying style transformation on every 3D point feature independently before volume rendering. This proves that applying DST on 2D feature maps is equivalent to applying the transformation on 3D points' features, maintaining multi-view consistency. The full derivation of Eq. (9) is provided in the appendix.", + "bbox": [ + 496, + 459, + 890, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we adopt a 2D CNN decoder to project the stylized feature maps $F_{cs}$ to RGB space to generate the final stylized novel view images.", + "bbox": [ + 496, + 551, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Two-stage Model Training", + "text_level": 1, + "bbox": [ + 500, + 604, + 738, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training of our model is divided into the feature grid training stage and the stylization training stage, the former is trained with the target of novel view synthesis, and the latter is trained with the target of style transfer.", + "bbox": [ + 496, + 628, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature grid training stage (First stage). We first learn the feature grid 3D representation for the novel view synthesis task, in preparation for performing feature transformation for style transfer. We train the feature grid and the 2D CNN decoder simultaneously, with the supervision of both RGB images and their bilinearly up-sampled feature maps extracted from ReLU3_1 layer of pre-trained VGG [51]. By aligning the VGG features with the feature grid, the reconstructed features acquire semantic information. We use density grid pre-trained solely on RGB images since the supervising feature maps are not strictly multi-view consistent. The training objective is the mean square error (MSE) between the predicted and ground truth feature maps and RGB images. Following [19, 37], we use perceptual", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss [24] as additional supervision to increase reconstructed image quality. The overall loss function is:", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {g r i d} = \\sum_ {r \\in \\mathcal {R}} \\left\\| \\hat {F} (\\mathbf {r}) - F (\\mathbf {r}) \\right\\| _ {2} ^ {2} + \\left\\| \\hat {I} _ {\\mathcal {R}} - I _ {\\mathcal {R}} \\right\\| _ {2} ^ {2} \\\\ + \\sum_ {l \\in l _ {p}} \\left\\| \\mathcal {F} ^ {l} (\\hat {I} _ {\\mathcal {R}}) - \\mathcal {F} ^ {l} (I _ {\\mathcal {R}}) \\right\\| _ {2} ^ {2}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 132, + 468, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the set of rays in each training batch, $\\hat{F} (\\mathbf{r}), F(\\mathbf{r})$ are the predicted and ground truth feature of ray $\\mathbf{r}$ , $\\hat{I}_{\\mathcal{R}}, I_{\\mathcal{R}}$ are the predicted and ground truth RGB image, $l_{p}$ denotes the set of VGG layers that compute perceptual loss, $\\mathcal{F}^l$ denotes the feature maps of the $l$ th layer of pre-trained VGG network.", + "bbox": [ + 75, + 220, + 467, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stylization training stage (Second stage). Our model learns to stylize novel views in the second stage. We freeze the feature grid, train the style transfer module, and finetune the CNN decoder. Thanks to the memory-efficient representation of 3D scenes and DST, unlike [8, 11, 48], our model can be trained directly on $256 \\times 256$ patches, making patch sub-sampling algorithm [8, 11, 22, 48] unnecessary. We use the same loss as [21] where the content loss $\\mathcal{L}_c$ is the MSE of the feature maps and the style loss $\\mathcal{L}_s$ is the MSE of the channel-wise feature mean and standard-deviation:", + "bbox": [ + 75, + 311, + 467, + 462 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t y l i z a t i o n}} = \\mathcal {L} _ {c} + \\lambda \\mathcal {L} _ {s}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 473, + 467, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ balances the content preservation and the stylization effect.", + "bbox": [ + 75, + 498, + 467, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 541, + 207, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate StyleRF extensively with qualitative experiments in Sec. 4.1, quantitative experiments in Sec. 4.2 and ablation studies in Sec. 4.3. We demonstrate two applications of StyleRF in Sec. 4.4. The implementation details are provided in the appendix.", + "bbox": [ + 75, + 566, + 467, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Experiments", + "text_level": 1, + "bbox": [ + 76, + 650, + 302, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate StyleRF over two public datasets including LLFF [35] that contains real scenes with complex geometry structures and Synthetic NeRF [36] that contains $360^{\\circ}$ views of objects. In addition, we benchmark StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8] with their released codes. We perform comparisons on LLFF dataset [35].", + "bbox": [ + 75, + 672, + 467, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 shows qualitative comparisons. We can see that StyleRF achieves clearly better stylization with more precise geometry reconstruction. Specifically, StyleRF can generate high-definition stylization with realistic textures and patterns of style images. The superior stylization is largely attributed to our transformation design that allows working in the feature space with full-resolution feature maps. As illustrated in the highlight boxes, StyleRF", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8b7a4060c8e13d9e558d954735d13d099f165606fddb47c935219e927a760c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodShort-range ConsistencyLong-range Consistency
LPIPSRMSELPIPSRMSE
AdaIN [21]0.1520.1230.2200.186
CCPL [60]0.1100.1060.1910.174
ReReVST [57]0.0980.0800.1860.146
LSNV [19]0.0930.0920.1810.155
Hyper [8]0.0840.0680.1310.101
Ours0.0720.0820.1490.137
", + "bbox": [ + 504, + 88, + 887, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results on consistency. We compare StyleRF with the state-of-the-art on consistency using LPIPS $(\\downarrow)$ and RMSE $(\\downarrow)$ .", + "bbox": [ + 498, + 250, + 890, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "can successfully restore the intricate geometry of complex scenes thanks to its radiance field representations. In addition, only StyleRF faithfully transfers the squareness texture in the second style image. Furthermore, StyleRF can robustly generalize to new styles in a zero-shot manner and can adapt well to $360^{\\circ}$ dataset as illustrated in Fig. 1. As a comparison, LSNV [19] fails to capture fine-level geometry like the bones of the T-Rex and the petals of the flower while Hyper [8] produces very blurry stylization.", + "bbox": [ + 496, + 289, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 498, + 435, + 692, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D style transfer is a very new and under-explored task and there are few metrics for quantitative evaluation of stylization quality. Hence, we manage to evaluate the multiview consistency only. In our experiments, we warp one view to the other according to the optical flow [53] using softmax splatting [42], and then computed the masked RMSE score and LPIPS score [64] to measure the stylization consistency. Following [8, 11, 19], we compute the short-range and long-range consistency scores which compare adjacent views and far-away views respectively. We compare StyleRF against two state-of-the-art zero-shot 3D style transfer methods Hyper [8] and LSNV [19], one SOTA single-frame-based video style transfer method CCPL [60], one SOTA multi-frames-based video style transfer method ReReVST [57], and one classical image style transfer method AdaIN [21].", + "bbox": [ + 496, + 459, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It can be seen from Tab. 1 that StyleRF significantly outperforms image style transfer approach [21] and video style transfer approach [57, 60] which capture little information about the underlying 3D geometry. In addition, StyleRF achieves better consistency than point-cloud-based 3D style transfer [19] as well. Note Hyper [8] achieves slightly better LPIPS and RMSE scores than our method, largely because it produces over-smooth results and inadequate stylization as shown in Fig. 5.", + "bbox": [ + 496, + 700, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 847, + 663, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We design two innovative techniques to improve the stylization quality and maintain the multi-view consistency.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8343", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee025dc9e4e1f6cfb462963d8fc32858f3e5c559245d7aa4c7010c15c28ec177.jpg", + "image_caption": [ + "Figure 5. Comparison of StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8]. For each of the two sample Scenes and reference Styles, StyleRF produces clearly better 3D style transfer and depth estimation. Check zoom-in for details." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 903, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The first is volume-adaptive instance normalization which uses the learned mean and variance of the whole volume during inference, eliminating the dependency on holistic statistics of the sampled point batch. The second is the adaptive bias addition in DST, which improves the stylization quality using bias capturing style information. We evaluate the two designs to examine how they contribute to the overall stylization of our method.", + "bbox": [ + 75, + 705, + 470, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Volume-adaptive instance normalization. We compare our volume-adaptive instance normalization (IN) with vanilla IN and StyleRF without IN. As Fig. 6 (c) shows, vanilla IN produces severe block-shape artifacts as the", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "transformation of each batch is conditioned on the holistic statistics of itself, thus each batch (i.e. block in the image) produces inconsistent stylization which leads to the artifacts. However, if we discard IN as shown in Fig. 6 (d), the multi-view consistency can maintain but the stylization quality compromises a lot, failing to capture the correct color tone of the reference style image. This is because IN removes the original style information of the content image which facilitates the transfer of the reference style [21].", + "bbox": [ + 496, + 705, + 893, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive bias addition. As illustrated in Fig. 6 (b), the stylization quality degrades a lot if we eliminate the adaptive bias addition in DST (Sec. 3.2.2), producing unnatural", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8344", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d9688b3ec0995eb5feab5b0bea28bc1f57424825c892392b6475b2fc78e4535.jpg", + "image_caption": [ + "Ground Truth Novel View" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 207, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4988ff3367a052dbcde5bf5cccf0a0318c2a12081b6610131b78124fcea707e0.jpg", + "image_caption": [ + "Reference Style Image" + ], + "image_footnote": [], + "bbox": [ + 107, + 176, + 183, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57d8076e5926cfcec12eab018620e7ca1e2cf67ecd60f60f9c25de28c0da1665.jpg", + "image_caption": [ + "(a) Ours" + ], + "image_footnote": [], + "bbox": [ + 215, + 90, + 339, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74077c824818391c72be93b2ff2578f227fa1f824657e1c5339ceef321b5f7ca.jpg", + "image_caption": [ + "(b) w/o Adaptive Bias" + ], + "image_footnote": [], + "bbox": [ + 339, + 90, + 465, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46d4595847b45fd388cf1584ad246981845658df3319c3c0c431849ae73133ab.jpg", + "image_caption": [ + "(c) Vanilla IN" + ], + "image_footnote": [], + "bbox": [ + 215, + 176, + 338, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f2cf7efd36889bc1754a0c64a25a05e0d400c4507f2a6a5ab163267feec3767.jpg", + "image_caption": [ + "(d) w/o IN" + ], + "image_footnote": [], + "bbox": [ + 339, + 176, + 464, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93c09f22dce827bbd0c6e32cb7aab03a59b216b7c5f37bde10227a14000938ee.jpg", + "image_caption": [ + "Figure 6. Ablation studies. (a) shows the stylization of our full pipeline. (b) shows the stylization without the adaptive bias. (c) shows the stylization when replacing the volume-adaptive instance normalization (IN) with vanilla IN. (d) shows the stylization without any IN.", + "Figure 7. Multi-style interpolation. StyleRF can smoothly interpolate between arbitrary styles by interpolating features of the scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 349, + 467, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stylization compared to the stylization of our full pipeline in Fig. 6 (a). This is because bias usually contains crucial style information such as the overall color tone [58]. StyleRF employs bias addition that is adaptively modulated by the weight of each ray, improving the stylization quality and keeping multi-view consistency concurrently.", + "bbox": [ + 75, + 590, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 76, + 694, + 210, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "StyleRF can be easily extended along different directions with different applications. We provide two possible extensions in the ensuing subsections.", + "bbox": [ + 75, + 717, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-style interpolation. StyleRF can smoothly interpolate different styles thanks to its high-level feature representation of a 3D scene. As illustrated in Fig. 7, we linearly interpolate the feature maps of a specific view by using four different styles at four corners. Unlike previous NeRF-based 3D style transfer that supports style interpolation by interpolating one-hot latent vectors [11], StyleRF can interpolate arbitrary numbers of unseen new styles by interpolating features of the scene, yielding more smooth", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/811621453ecb598dee2ad002244040ce4d61eda8bbd4b6d8d0022da849410e12.jpg", + "image_caption": [ + "Ground Truth Novel View", + "Generated Mask" + ], + "image_footnote": [], + "bbox": [ + 506, + 99, + 630, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d70f897c0e5096e80ad8d2cfc9fa20eb8ff7345dd54cff87a100fe05ec6564e3.jpg", + "image_caption": [ + "Spatial Composition", + "Figure 8. Compositional 3D style transfer. Given the 3D-consistent segmentation masks, StyleRF can create infinite combinations of styles by spatial composition." + ], + "image_footnote": [], + "bbox": [ + 638, + 99, + 758, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea120450f43283200e27d9f472aefb4f871ab7655d0c5eb3b3bf741242516c3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 99, + 885, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and harmonious interpolation. Hence, StyleRF can not only transfer arbitrary styles in a zero-shot manner but also generate non-existent stylization via multi-style interpolation.", + "bbox": [ + 498, + 321, + 890, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compositional 3D style transfer. Thanks to its precise geometry reconstruction, StyleRF can be seamlessly integrated with NeRF-based object segmentation [12, 26, 65] for compositional 3D style transfer. As shown in Fig. 8, we apply 3D-consistent segmentation masks to the feature maps and apply different styles to stylize the contents inside and outside the masks separately. We can see that the edges of the masks can be blended more softly by applying the segmentation masks to the feature maps instead of RGB images. Due to its zero-shot nature, StyleRF can create infinite combinations of styles without additional training, producing numerous artistic creations and inspirations.", + "bbox": [ + 498, + 369, + 890, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 566, + 617, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present StyleRF, a novel zero-shot 3D style transfer method that balances the three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. By representing the 3D scene with an explicit grid of high-level features, we can faithfully restore high-fidelity geometry through volume rendering. Then we perform style transfer on the feature space of the scene, leading to high-quality zero-shot stylization results. We innovatively design sampling-invariant content transformation to maintain multiview consistency and deferred style transformation to increase efficiency. We demonstrate that StyleRF achieves superior 3D stylization quality than previous zero-shot 3D style transfer methods, and can be extended to various interesting applications for artistic 3D creations.", + "bbox": [ + 496, + 592, + 890, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 645, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with project number RT18/22.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8345", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2, 3, 4", + "[5] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 3", + "[6] Yaosen Chen, Qi Yuan, Zhiqiang Li, Yuegen Liu, Wei Wang, Chaoping Xie, Xuming Wen, and Qien Yu. Upst-nerf: Universal photorealistic style transfer of neural radiance fields for 3d scene. In *arxiv*, 2022. 3", + "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2", + "[8] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2, 3, 6, 7", + "[9] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 3", + "[10] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14304-14313, 2021. 2", + "[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2, 3, 6, 8", + "[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776, 2022. 8", + "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 2", + "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3", + "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3, 4", + "[17] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2", + "[18] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 783-791, 2017. 3", + "[19] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13869-13878, 2021. 2, 3, 5, 6, 7", + "[20] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2", + "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3, 4, 5, 6, 7", + "[22] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2, 3, 6", + "[23] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2", + "[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2, 3, 6", + "[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8346", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 8", + "[27] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2", + "[28] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2019. 3, 4, 5", + "[29] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3", + "[30] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 2", + "[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[32] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 3, 4", + "[33] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 2", + "[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2", + "[35] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6", + "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4, 6", + "[37] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2, 3, 5", + "[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2, 3", + "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3, 4", + "[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2", + "[42] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020. 6", + "[43] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 3", + "[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2", + "[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[46] Charles R Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J Guibas. Volumetric and multi-view cnns for object classification on 3d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2016. 2", + "[47] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. International Journal of Computer Vision, 126(11):1199-1219, 2018. 3", + "[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6", + "[49] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2", + "[50] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8242-8250, 2018. 3", + "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + ], + "bbox": [ + 501, + 90, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8347", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 6", + "[54] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4", + "[55] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2", + "[56] Wenjing Wang, Jizheng Xu, Li Zhang, Yue Wang, and Jiaying Liu. Consistent video style transfer via compound regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12233-12240, 2020. 3", + "[57] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 2020. 3, 6", + "[58] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 3, 5, 8", + "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2", + "[60] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 6", + "[61] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2", + "[62] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. In *The Eleventh International Conference on Learning Representations*, 2023. 2", + "[63] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2, 3", + "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[65] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 8" + ], + "bbox": [ + 78, + 90, + 467, + 854 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8348", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_model.json b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_model.json new file mode 100644 index 0000000000000000000000000000000000000000..266dd38c1ef31dac2849f888d161be7b382a6616 --- /dev/null +++ b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.131, + 0.809, + 0.152 + ], + "angle": 0, + "content": "StyleRF: Zero-shot 3D Style Transfer of Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.166, + 0.744, + 0.185 + ], + "angle": 0, + "content": "Kunhao Liu\\(^{1}\\) Fangneng Zhan\\(^{2}\\) Yiwen Chen\\(^{1}\\) Jiahui Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.186, + 0.78, + 0.205 + ], + "angle": 0, + "content": "Yingchen Yu1 Abdulmotaleb El Saddik3,5 Shijian Lu1* Eric Xing4,5" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.213, + 0.712, + 0.228 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanyang Technological University \\(^{2}\\)Max Planck Institute for Informatics" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.23, + 0.682, + 0.246 + ], + "angle": 0, + "content": "3University of Ottawa 4Carnegie Mellon University 5MBZUAI" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.273, + 0.28, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.275, + 0.497, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.275, + 0.697, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.269, + 0.889, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.397, + 0.312, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.537, + 0.29, + 0.551 + ], + "angle": 0, + "content": "Multi-view Content Images" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.412, + 0.503, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.412, + 0.695, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.412, + 0.888, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.537, + 0.757, + 0.551 + ], + "angle": 0, + "content": "Reference Style Images + Stylized Novel Views" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.563, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Figure 1. Zero-shot 3D Style Transfer. Given a set of Multi-view Content Images of a 3D scene, StyleRF can transfer arbitrary Reference Styles to the 3D scene in a zero-shot manner, rendering high-quality Stylized Novel Views with superb multi-view consistency." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.605, + 0.314, + 0.62 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.636, + 0.474, + 0.879 + ], + "angle": 0, + "content": "3D style transfer aims to render stylized novel views of a 3D scene with multi-view consistency. However, most existing work suffers from a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. We propose StyleRF (Style Radiance Fields), an innovative 3D style transfer technique that resolves the three-way dilemma by performing style transformation within the feature space of a radiance field. StyleRF employs an explicit grid of high-level features to represent 3D scenes, with which high-fidelity geometry can be reliably restored via volume rendering. In addition, it transforms the grid features according to the reference style which directly leads to high-quality zero-shot style transfer. StyleRF consists of two innovative designs. The first is sampling-invariant content transformation that makes the transformation invari" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.606, + 0.895, + 0.757 + ], + "angle": 0, + "content": "ant to the holistic statistics of the sampled 3D points and accordingly ensures multi-view consistency. The second is deferred style transformation of 2D feature maps which is equivalent to the transformation of 3D points but greatly reduces memory footprint without degrading multi-view consistency. Extensive experiments show that StyleRF achieves superior 3D stylization quality with precise geometry reconstruction and it can generalize to various new styles in a zero-shot manner. Project website: https://kunhao-liu.github.io/StyleRF/" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.782, + 0.633, + 0.798 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given a set of multi-view images of a 3D scene and an image capturing a target style, 3D style transfer aims to generate novel views of the 3D scene that have the target style consistently across the generated views (Fig. 1). Neural style transfer has been investigated extensively, and state-of-the-art methods allow transferring arbitrary styles in a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.31, + 0.901 + ], + "angle": 0, + "content": "* Shijian Lu is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8338" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.213 + ], + "angle": 0, + "content": "zero-shot manner. However, most existing work focuses on style transfer across 2D images [15, 21, 24] but cannot extend to a 3D scene that has arbitrary new views. Prior studies [19, 22, 37, 39] have shown that naively combining 3D novel view synthesis and 2D style transfer often leads to multi-view inconsistency or poor stylization quality, and 3D style transfer should optimize novel view synthesis and style transfer jointly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.47, + 0.442 + ], + "angle": 0, + "content": "However, the current 3D style transfer is facing a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to new styles. Different approaches have been investigated to resolve the three-way dilemma. For example, multiple style transfer [11, 22] requires a set of pre-defined styles but cannot generalize to unseen new styles. Point-cloud-based style transfer [19, 37] requires a pre-trained depth estimation module that is prone to inaccurate geometry reconstruction. Zero-shot style transfer with neural radiance fields (NeRF) [8] cannot capture detailed style patterns and textures as it implicitly injects the style information into neural network parameters. Optimization-based style transfer [17, 39, 63] suffers from slow optimization and cannot scale with new styles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.793 + ], + "angle": 0, + "content": "In this work, we introduce StyleRF to resolve the three-way dilemma by performing style transformation in the feature space of a radiance field. A radiance field is a continuous volume that can restore more precise geometry than point clouds or meshes. In addition, transforming a radiance field in the feature space is more expressive with better stylization quality than implicit methods [8], and it can also generalize to arbitrary styles. We construct a 3D scene representation with a grid of deep features to enable feature transformation. In addition, multi-view consistent style transformation in the feature space could be achieved by either transforming the whole feature grid or transforming the sampled 3D points. We adopt the latter as the former incurs much more computational cost during training to stylize the whole feature grid in every iteration, whereas the latter can reduce computational cost through decreasing the size of training patch and the number of sampled points. However, applying off-the-shelf style transformations to a batch of sampled 3D points impairs the multi-view consistency as they are conditioned on the holistic statistics of the batch. Beyond that, transforming every sampled 3D point is memory-intensive since NeRF needs to query hundreds of sampled points along each ray for rendering a single pixel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST), the former eliminating the dependency on holistic statistics of sampled point batch and the latter deferring style transformation to 2D feature maps for better efficiency. In SICT, we introduce volume-adaptive normalization that learns the mean and variance of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "whole volume instead of computing them from a sampled batch. In addition, we apply channel-wise self-attention to transform each 3D point independently to make it conditioned on the feature of that point regardless of the holistic statistics of the sampled batch. In DST, we defer the style transformation to the volume-rendered 2D feature maps based on the observation that the style transformation of each point is the same. By formulating the style transformation by pure matrix multiplication and adaptive bias addition, transforming 2D feature maps is mathematically equivalent to transforming 3D point features but it saves computation and memory greatly. Thanks to the memory-efficient representation of 3D scenes and deferred style transformation, our network can train with \\(256 \\times 256\\) patches directly without requiring sub-sampling like previous NeRF-based 3D style transfer methods [8, 11, 22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.545 + ], + "angle": 0, + "content": "The contributions of this work can be summarized in three aspects. First, we introduce StyleRF, an innovative zero-shot 3D style transfer framework that can generate zero-shot high-quality 3D stylization via style transformation within the feature space of a radiance field. Second, we design sampling-invariant content transformation and deferred style transformation, the former achieving multiview consistent transformation by eliminating dependency on holistic statistics of sampled point batch while the latter greatly improves stylization efficiency by deferring style transformation to 2D feature maps. Third, extensive experiments show that StyleRF achieves superior 3D style transfer with accurate geometry reconstruction, high-quality stylization, and great generalization to new styles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.642, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Neural scene representations. 3D scene representation has been extensively studied in recent years with different ways of representations such as volumes [23,27,46,49,59], point clouds [1, 45], meshes [25, 55], depth maps [20, 30], and implicit functions [7, 34, 41, 61]. These methods adopt differentiable rendering which enables model optimization by using 2D multi-view images. Among them, Neural Radiance Field (NeRF) [36] can render a complex 3D scene with high fidelity and accurate geometry. It represents scenes with an implicit coordinate function that maps each 3D coordinate to a density value and a color value, and employs volume rendering to generate images of novel views. However, the implicit coordinate function is represented by a large multilayer perceptron (MLP) that is often hard to optimize and slow to infer. Serval studies adopt a hybrid representation [3, 4, 10, 14, 31, 33, 38, 44, 52, 62] to speed up the reconstruction and rendering. They employ explicit data structures such as discrete voxel grids [14,52], decomposed tensors [3, 4, 13], hash maps [38], etc. to store features or spherical harmonics, enabling fast convergence and inference." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.881, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.896, + 0.327 + ], + "angle": 0, + "content": "Figure 2. The framework of StyleRF. For a batch of sampled points along a ray \\(\\mathbf{r}\\), the corresponding features \\(F_{i}, i \\in [1,2,\\dots,N]\\) are first extracted, each of which is transformed to \\(\\bar{F}_{i}\\) independently via Sampling-Invariant Content Transformation, regardless of the holistic statistics of the point batch. \\(\\bar{F}_{i}\\) is then transformed to a feature map \\(\\bar{F}_{c}\\) via Volume Rendering. After that, the Deferred Style Transformation transforms \\(\\bar{F}_{c}\\) to the feature map \\(F_{cs}\\) adaptively using the sum weight of the sampled points \\(w_{\\mathbf{r}}\\) along the ray \\(\\mathbf{r}\\) and the style information \\(T, \\mu(F_{s})\\), and \\(\\sigma(F_{s})\\). Finally, a stylized novel view is generated via a CNN decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.536 + ], + "angle": 0, + "content": "Although most existing work extracts features as middle-level representations of scenes, the extracted features are usually an intermediate output of neural networks which have little semantic meanings and are not suitable for the style transfer task. We introduce decomposed tensors [4] to store high-level features extracted by pre-trained CNNs, which enables transformations in feature space as well as efficient training and inference. Though [3,16,40] also render feature maps instead of RGB maps, they are computationally intensive and usually work with low-resolution feature maps. StyleRF can instead render full-resolution feature maps (the same as the output RGB images) efficiently and it uses high-level features largely for transformation only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Neural style transfer. Neural style transfer aims at rendering a new image that contains the content structure of one image and the style patterns of another. The seminal work in [15] shows that multi-level feature statistics extracted from intermediate layers of pre-trained CNNs could be used as a representation of the style of an artistic image, but it treats style transfer as a slow and iterative optimization task. [9,21,24,28,29,32,43,50,58] utilize feed-forward networks to approximate the optimization procedure to speed up rendering. Among them, [9,21,28,29,32,43,50,58] can achieve zero-shot style transfer by applying transformations to the high-level features extracted by pre-trained CNNs, where the feature transformations can be achieved by matching second-order statistics [21, 29], linear transformation [28,58], self-attention transformation [9,32,43], etc. Video style transfer extends style transfer to videos for injecting target styles consistently across adjacent video frames. Several studies leverage optical flow [5,18,47,56,57] as temporal constraints to estimate the movement of video contents. They can produce smooth videos, but have little knowledge of the underlying 3D geometry and cannot render consistent frames in arbitrary views [19,37]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Huang et al. first tackle stylizing complex 3D scenes [19]. They construct a 3D scene by back-projecting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.656 + ], + "angle": 0, + "content": "image features into the 3D space to form a point cloud and then perform style transformation on the features of 3D points. Their method can achieve zero-shot style transfer, but requires an error-prone pre-trained depth estimator to model scene geometry. [37] also constructs a point cloud for stylization but it mainly focuses on monocular images. Instead, [6, 8, 11, 22, 39, 63] use NeRF [36] as the 3D representation which can reconstruct scene geometry more faithfully. [6] is a photorealistic style transfer method that can only transfer the color tone of style images. [39, 63] achieve 3D style transfer via optimization and can produce visually high-quality stylization, but they require a time-consuming optimization procedure for every reference style. [11, 22] employ latent codes to represent a set of pre-defined styles, but cannot generalize to unseen styles. [8] can achieve arbitrary style transfer by implicitly instilling the style information into MLP parameters. However, it can only transfer the color tone of style images but cannot capture detailed style patterns. StyleRF can transfer arbitrary style in a zero-shot manner, and it can capture style details such as strokes and textures as well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.593, + 0.693 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The overview of StyleRF is shown in Fig. 2. For a batch of sampled points along a ray \\(\\mathbf{r}\\), the corresponding features \\(F_{i}, i \\in [1,2,\\dots,N]\\) are first extracted from the feature grid described in Sec. 3.1, each of which is transformed to \\(\\bar{F}_{i}\\) independently via Sampling-Invariant Content Transformation (SICT) described in Sec. 3.2.1, regardless of the holistic statistics of the point batch. \\(\\bar{F}_{i}\\) is then transformed to a feature map \\(\\bar{F}_{c}\\) via Volume Rendering. After that, the Deferred Style Transformation (DST) described in Sec. 3.2.2 transforms \\(\\bar{F}_{c}\\) to the feature map \\(F_{cs}\\) adaptively using the sum weight of the sampled points \\(w_{\\mathbf{r}}\\) along the ray \\(\\mathbf{r}\\) and the style information \\(T, \\mu(F_{s})\\), and \\(\\sigma(F_{s})\\). Finally, a stylized novel view is generated via a CNN decoder." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8340" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.362, + 0.108 + ], + "angle": 0, + "content": "3.1. Feature Grid 3D Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.471, + 0.25 + ], + "angle": 0, + "content": "To model a 3D scene with deep features, we use a continuous volumetric field of density and radiance. Different from the original NeRF [36], for every queried 3D position \\( x \\in \\mathbb{R}^3 \\), we get a volume density \\( \\sigma(x) \\) and a multi-channel feature \\( F(x) \\in \\mathbb{R}^C \\) instead of an RGB color, where \\( C \\) is the number of the feature channels. Then we can get the feature of any rays \\( \\mathbf{r} \\) passing through the volume by integrating sampled points along the ray via approximated volume rendering [36]:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.259, + 0.47, + 0.302 + ], + "angle": 0, + "content": "\\[\nF (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} w _ {i} F _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.31, + 0.469, + 0.37 + ], + "angle": 0, + "content": "where \\(w_{i} = \\exp \\left(-\\sum_{j = 1}^{i - 1}\\sigma_{j}\\delta_{j}\\right)\\) \\((1 - \\exp (-\\sigma_i\\delta_i))\\) (2)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.469, + 0.431 + ], + "angle": 0, + "content": "where \\(\\sigma_{i}, F_{i}\\) denotes the volume density and feature of sampled point \\(i\\), \\(w_{i}\\) denotes the weight of \\(F_{i}\\) in the ray \\(\\mathbf{r}\\), and \\(\\delta_{i}\\) is the distance between adjacent samples. We disable view-dependency effect for better multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.581 + ], + "angle": 0, + "content": "Then we can generate feature maps capturing high-level features and map them to RGB space using a 2D CNN decoder. However, unlike [3,16,40], we render full-resolution feature maps which have the same resolution as the final RGB images rather than down-sampled feature maps. Rendering full-resolution feature maps has two unique features: 1) it discards up-sampling operations which cause multiview inconsistency in general [16], 2) it removes aliasing when rendering low-resolution feature maps [2] which causes severe flickering effects in stylized RGB videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.704 + ], + "angle": 0, + "content": "Directly using 3D voxel grid to store features is memory-intensive. We thus adopt vector-matrix tensor decomposition [4] that relaxes the low-rank constraints for two modes of a 3D tensor and factorizes tensors into compact vector and matrix factors, which lowers the space complexity from \\(\\mathcal{O}(n^3)\\) to \\(\\mathcal{O}(n^2)\\), massively reducing the memory footprint. We employ a density grid to store volume density and a feature grid to store multi-channel features respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.438, + 0.728 + ], + "angle": 0, + "content": "3.2. Feature Transformation for Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Once we have the feature grid representation of a scene, we can tackle the task of stylizing 3D scenes. Given a reference style image, our goal is to render stylized novel views of the 3D scene with multi-view consistency. To achieve this, we apply transformations to the features of the grid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "One plausible solution to this task is to apply style transfer to the feature grid directly. This solution is efficient in evaluations as it can render any stylized views with a single style transfer process only. However, it is impractical to train such transformation as it needs to stylize the whole feature grid in every iteration. Another solution is" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.707, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.203, + 0.655, + 0.215 + ], + "angle": 0, + "content": "(a) Vanilla IN." + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.088, + 0.886, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.203, + 0.874, + 0.215 + ], + "angle": 0, + "content": "(b) Volume-adaptive IN." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.227, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Comparison between vanilla instance normalization (IN) in (a) and volume-adaptive IN in (b). During evaluation, volume-adaptive IN uses learned mean and standard-deviation, discarding dependency over the sampled point batch's holistic statistics (indicated by the red arrows in the left graph)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.488 + ], + "angle": 0, + "content": "to apply an off-the-shelf zero-shot style transfer method to the features of the sampled 3D points. While this solution can reduce computational cost through decreasing the size of training patch and the number of sampled points, it has two problems: 1) vanilla zero-shot style transformation is conditioned on holistic statistics of the sampled point batch [21, 28, 32], which violates multi-view consistency in volume rendering as the feature transformation of a specific 3D point will vary across different sampled points; 2) volume rendering requires sampling hundreds of points along a single ray, which makes transformation on the point batch memory-intensive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Motivated by the observation that style transformation is conditioned on both content information and style information, we decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST). After the decomposition, SICT will be conditioned solely on the content information while DST conditioned solely on the style information, more details to be elaborated in the ensuing subsections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.622, + 0.864, + 0.636 + ], + "angle": 0, + "content": "3.2.1 Sampling-invariant Content Transformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Given a batch of sampled points, we can get their corresponding features \\( F_{i} \\in \\mathbb{R}^{C}, i \\in [1,2,\\dots,N] \\) from the feature grid, where \\( N \\) is the number of the sampled points along a ray and \\( C \\) is the number of the feature channels. The goal of SICT is to transform the extracted features \\( F_{i} \\) so that they can be better stylized. We formulate SICT as a channel-wise self-attention operation to the features after instance normalization (IN) [54]. Specifically, we formulate \\( Q(\\text{query}), K(\\text{key}) \\), and \\( V(\\text{value}) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.792, + 0.891, + 0.808 + ], + "angle": 0, + "content": "\\[\nQ = q \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.811, + 0.891, + 0.826 + ], + "angle": 0, + "content": "\\[\nK = k \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.829, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\nV = v \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(q, k, v\\) are \\(1 \\times 1\\) convolution layers which reduce the channel number from \\(C\\) to \\(C'\\) for computational efficiency, and Norm denotes the IN. However, as shown in Fig. 3," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8341" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.09, + 0.434, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.371 + ], + "angle": 0, + "content": "Figure 4. Deferred style transformation. We apply the style transformation to the volume-rendered feature maps \\(\\bar{F}_c\\) according to the style feature maps \\(F_s\\). To ensure multi-view consistency, we modulate the bias (e.g. the mean value of the style feature maps \\(\\mu(F_s)\\)) with the sum weight of sampled points along each ray \\(w_{\\mathbf{r}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.532 + ], + "angle": 0, + "content": "vanilla IN calculates per-dimension mean and standard-deviation of the batch of sampled points, which varies with different sampled points and incurs multi-view inconsistency accordingly. Thus we design volume-adaptive IN which, during training, keeps running estimates of the computed mean and standard-deviation, and uses them for normalization during evaluations (instead of computing from the sampled point batch). Through volume-adaptive IN, we can ensure that the content transformation is consistent regardless of the sampled point batch's holistic statistics." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.533, + 0.468, + 0.548 + ], + "angle": 0, + "content": "Channel-wise self-attention can thus be implemented by:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.564, + 0.469, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\bar {F} _ {i} = V \\otimes \\operatorname {S o f t m a x} (\\widetilde {\\operatorname {c o v}} (Q, K)), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.589, + 0.47, + 0.634 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes matrix multiplication and \\(\\widetilde{cov}(Q, K) \\in \\mathbb{R}^{N \\times C' \\times C'}\\) denotes the covariance matrix in the channel dimension." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.347, + 0.665 + ], + "angle": 0, + "content": "3.2.2 Deferred Style Transformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.749 + ], + "angle": 0, + "content": "After applying SICT to the features of each 3D point, we apply DST to the volume-rendered 2D feature maps \\(\\bar{F}_c\\) rather than 3D point features \\(\\bar{F}_i\\). To ensure multi-view consistency, we formulate the transformation as matrix multiplication and adaptive bias addition as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Specifically, we first extract feature maps \\( F_{s} \\) of the reference style \\( S \\) using a pre-trained VGG [51], and then generate the style transformation matrix \\( T \\in \\mathbb{R}^{C' \\times C'} \\) using feature covariance \\( cov(F_{s}) \\) following [28]. Next, we apply matrix multiplication with \\( T \\) to the feature maps \\( \\bar{F}_{c} \\) and use a \\( 1 \\times 1 \\) convolution layer conv without bias to restore the channel number from \\( C' \\) to \\( C \\). Though these operations can partially instill style information, they are not expressive enough without bias addition containing style information [58]. Thus following [21], we multiply the feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "maps with the standard-deviation value \\(\\sigma(F_s)\\) and add the mean value \\(\\mu(F_s)\\). To ensure it is equivalent when applying the transformation to either 3D point features or 2D feature maps, we adaptively modulate the mean value \\(\\mu(F_s)\\) with the sum weight of sampled points along each ray \\(w_{\\mathbf{r}}\\). DST can be mathematically formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.194, + 0.892, + 0.212 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = \\operatorname {c o n v} \\left(T \\otimes \\bar {F} _ {c}\\right) \\times \\sigma \\left(F _ {s}\\right) + w _ {\\mathbf {r}} \\times \\mu \\left(F _ {s}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.223, + 0.892, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad \\bar {F} _ {c} = \\sum_ {i = 1} ^ {N} w _ {i} \\bar {F} _ {i}, w _ {\\mathbf {r}} = \\sum_ {i = 1} ^ {N} w _ {i}, \\mathbf {r} \\in \\mathcal {R} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.891, + 0.315 + ], + "angle": 0, + "content": "where \\(w_{i}\\) denotes the weight of sampled point \\(i\\) (Eq. (2)), \\(\\bar{F}_i\\) denotes the feature of sample \\(i\\) after SICT, and \\(\\mathcal{R}\\) is the set of rays in each training batch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Note \\( conv \\) is a \\( 1 \\times 1 \\) convolution layer without bias, so it is basically a matrix multiplication operation. And \\( \\sigma(S), \\mu(S) \\) are scalars. Together with the adaptive bias modulation \\( w_{\\mathbf{r}} \\), Eq. (7) can be reformulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.387, + 0.892, + 0.451 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = \\sum_ {i = 1} ^ {N} w _ {i} \\left(\\underbrace {\\operatorname {c o n v} \\left(T \\otimes \\bar {F} _ {i}\\right) \\times \\sigma \\left(F _ {s}\\right) + \\mu \\left(F _ {s}\\right)} _ {\\text {(i)}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.551 + ], + "angle": 0, + "content": "where part (i) can be seen as applying style transformation on every 3D point feature independently before volume rendering. This proves that applying DST on 2D feature maps is equivalent to applying the transformation on 3D points' features, maintaining multi-view consistency. The full derivation of Eq. (9) is provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Finally, we adopt a 2D CNN decoder to project the stylized feature maps \\( F_{cs} \\) to RGB space to generate the final stylized novel view images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.74, + 0.622 + ], + "angle": 0, + "content": "3.3. Two-stage Model Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The training of our model is divided into the feature grid training stage and the stylization training stage, the former is trained with the target of novel view synthesis, and the latter is trained with the target of style transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Feature grid training stage (First stage). We first learn the feature grid 3D representation for the novel view synthesis task, in preparation for performing feature transformation for style transfer. We train the feature grid and the 2D CNN decoder simultaneously, with the supervision of both RGB images and their bilinearly up-sampled feature maps extracted from ReLU3_1 layer of pre-trained VGG [51]. By aligning the VGG features with the feature grid, the reconstructed features acquire semantic information. We use density grid pre-trained solely on RGB images since the supervising feature maps are not strictly multi-view consistent. The training objective is the mean square error (MSE) between the predicted and ground truth feature maps and RGB images. Following [19, 37], we use perceptual" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8342" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "loss [24] as additional supervision to increase reconstructed image quality. The overall loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.133, + 0.47, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {g r i d} = \\sum_ {r \\in \\mathcal {R}} \\left\\| \\hat {F} (\\mathbf {r}) - F (\\mathbf {r}) \\right\\| _ {2} ^ {2} + \\left\\| \\hat {I} _ {\\mathcal {R}} - I _ {\\mathcal {R}} \\right\\| _ {2} ^ {2} \\\\ + \\sum_ {l \\in l _ {p}} \\left\\| \\mathcal {F} ^ {l} (\\hat {I} _ {\\mathcal {R}}) - \\mathcal {F} ^ {l} (I _ {\\mathcal {R}}) \\right\\| _ {2} ^ {2}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.468, + 0.311 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the set of rays in each training batch, \\(\\hat{F} (\\mathbf{r}), F(\\mathbf{r})\\) are the predicted and ground truth feature of ray \\(\\mathbf{r}\\), \\(\\hat{I}_{\\mathcal{R}}, I_{\\mathcal{R}}\\) are the predicted and ground truth RGB image, \\(l_{p}\\) denotes the set of VGG layers that compute perceptual loss, \\(\\mathcal{F}^l\\) denotes the feature maps of the \\(l\\)th layer of pre-trained VGG network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.468, + 0.463 + ], + "angle": 0, + "content": "Stylization training stage (Second stage). Our model learns to stylize novel views in the second stage. We freeze the feature grid, train the style transfer module, and finetune the CNN decoder. Thanks to the memory-efficient representation of 3D scenes and DST, unlike [8, 11, 48], our model can be trained directly on \\(256 \\times 256\\) patches, making patch sub-sampling algorithm [8, 11, 22, 48] unnecessary. We use the same loss as [21] where the content loss \\(\\mathcal{L}_c\\) is the MSE of the feature maps and the style loss \\(\\mathcal{L}_s\\) is the MSE of the channel-wise feature mean and standard-deviation:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.474, + 0.468, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t y l i z a t i o n}} = \\mathcal {L} _ {c} + \\lambda \\mathcal {L} _ {s}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.468, + 0.528 + ], + "angle": 0, + "content": "where \\(\\lambda\\) balances the content preservation and the stylization effect." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.209, + 0.559 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.468, + 0.643 + ], + "angle": 0, + "content": "We evaluate StyleRF extensively with qualitative experiments in Sec. 4.1, quantitative experiments in Sec. 4.2 and ablation studies in Sec. 4.3. We demonstrate two applications of StyleRF in Sec. 4.4. The implementation details are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.303, + 0.667 + ], + "angle": 0, + "content": "4.1. Qualitative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.78 + ], + "angle": 0, + "content": "We evaluate StyleRF over two public datasets including LLFF [35] that contains real scenes with complex geometry structures and Synthetic NeRF [36] that contains \\(360^{\\circ}\\) views of objects. In addition, we benchmark StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8] with their released codes. We perform comparisons on LLFF dataset [35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 shows qualitative comparisons. We can see that StyleRF achieves clearly better stylization with more precise geometry reconstruction. Specifically, StyleRF can generate high-definition stylization with realistic textures and patterns of style images. The superior stylization is largely attributed to our transformation design that allows working in the feature space with full-resolution feature maps. As illustrated in the highlight boxes, StyleRF" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.888, + 0.24 + ], + "angle": 0, + "content": "
MethodShort-range ConsistencyLong-range Consistency
LPIPSRMSELPIPSRMSE
AdaIN [21]0.1520.1230.2200.186
CCPL [60]0.1100.1060.1910.174
ReReVST [57]0.0980.0800.1860.146
LSNV [19]0.0930.0920.1810.155
Hyper [8]0.0840.0680.1310.101
Ours0.0720.0820.1490.137
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.25, + 0.892, + 0.279 + ], + "angle": 0, + "content": "Table 1. Results on consistency. We compare StyleRF with the state-of-the-art on consistency using LPIPS \\((\\downarrow)\\) and RMSE \\((\\downarrow)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.427 + ], + "angle": 0, + "content": "can successfully restore the intricate geometry of complex scenes thanks to its radiance field representations. In addition, only StyleRF faithfully transfers the squareness texture in the second style image. Furthermore, StyleRF can robustly generalize to new styles in a zero-shot manner and can adapt well to \\(360^{\\circ}\\) dataset as illustrated in Fig. 1. As a comparison, LSNV [19] fails to capture fine-level geometry like the bones of the T-Rex and the petals of the flower while Hyper [8] produces very blurry stylization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.436, + 0.694, + 0.452 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.701 + ], + "angle": 0, + "content": "3D style transfer is a very new and under-explored task and there are few metrics for quantitative evaluation of stylization quality. Hence, we manage to evaluate the multiview consistency only. In our experiments, we warp one view to the other according to the optical flow [53] using softmax splatting [42], and then computed the masked RMSE score and LPIPS score [64] to measure the stylization consistency. Following [8, 11, 19], we compute the short-range and long-range consistency scores which compare adjacent views and far-away views respectively. We compare StyleRF against two state-of-the-art zero-shot 3D style transfer methods Hyper [8] and LSNV [19], one SOTA single-frame-based video style transfer method CCPL [60], one SOTA multi-frames-based video style transfer method ReReVST [57], and one classical image style transfer method AdaIN [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.838 + ], + "angle": 0, + "content": "It can be seen from Tab. 1 that StyleRF significantly outperforms image style transfer approach [21] and video style transfer approach [57, 60] which capture little information about the underlying 3D geometry. In addition, StyleRF achieves better consistency than point-cloud-based 3D style transfer [19] as well. Note Hyper [8] achieves slightly better LPIPS and RMSE scores than our method, largely because it produces over-smooth results and inadequate stylization as shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.665, + 0.862 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We design two innovative techniques to improve the stylization quality and maintain the multi-view consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8343" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.086, + 0.905, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.666, + 0.893, + 0.695 + ], + "angle": 0, + "content": "Figure 5. Comparison of StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8]. For each of the two sample Scenes and reference Styles, StyleRF produces clearly better 3D style transfer and depth estimation. Check zoom-in for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.471, + 0.829 + ], + "angle": 0, + "content": "The first is volume-adaptive instance normalization which uses the learned mean and variance of the whole volume during inference, eliminating the dependency on holistic statistics of the sampled point batch. The second is the adaptive bias addition in DST, which improves the stylization quality using bias capturing style information. We evaluate the two designs to examine how they contribute to the overall stylization of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Volume-adaptive instance normalization. We compare our volume-adaptive instance normalization (IN) with vanilla IN and StyleRF without IN. As Fig. 6 (c) shows, vanilla IN produces severe block-shape artifacts as the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.895, + 0.844 + ], + "angle": 0, + "content": "transformation of each batch is conditioned on the holistic statistics of itself, thus each batch (i.e. block in the image) produces inconsistent stylization which leads to the artifacts. However, if we discard IN as shown in Fig. 6 (d), the multi-view consistency can maintain but the stylization quality compromises a lot, failing to capture the correct color tone of the reference style image. This is because IN removes the original style information of the content image which facilitates the transfer of the reference style [21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Adaptive bias addition. As illustrated in Fig. 6 (b), the stylization quality degrades a lot if we eliminate the adaptive bias addition in DST (Sec. 3.2.2), producing unnatural" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8344" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.09, + 0.209, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.165, + 0.208, + 0.175 + ], + "angle": 0, + "content": "Ground Truth Novel View" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.178, + 0.184, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.253, + 0.201, + 0.263 + ], + "angle": 0, + "content": "Reference Style Image" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.091, + 0.34, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.165, + 0.299, + 0.176 + ], + "angle": 0, + "content": "(a) Ours" + }, + { + "type": "image", + "bbox": [ + 0.341, + 0.091, + 0.466, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.165, + 0.454, + 0.176 + ], + "angle": 0, + "content": "(b) w/o Adaptive Bias" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.178, + 0.339, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.252, + 0.309, + 0.263 + ], + "angle": 0, + "content": "(c) Vanilla IN" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.178, + 0.465, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.252, + 0.428, + 0.263 + ], + "angle": 0, + "content": "(d) w/o IN" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.345 + ], + "angle": 0, + "content": "Figure 6. Ablation studies. (a) shows the stylization of our full pipeline. (b) shows the stylization without the adaptive bias. (c) shows the stylization when replacing the volume-adaptive instance normalization (IN) with vanilla IN. (d) shows the stylization without any IN." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.35, + 0.468, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Figure 7. Multi-style interpolation. StyleRF can smoothly interpolate between arbitrary styles by interpolating features of the scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.682 + ], + "angle": 0, + "content": "stylization compared to the stylization of our full pipeline in Fig. 6 (a). This is because bias usually contains crucial style information such as the overall color tone [58]. StyleRF employs bias addition that is adaptively modulated by the weight of each ray, improving the stylization quality and keeping multi-view consistency concurrently." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.212, + 0.71 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.763 + ], + "angle": 0, + "content": "StyleRF can be easily extended along different directions with different applications. We provide two possible extensions in the ensuing subsections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Multi-style interpolation. StyleRF can smoothly interpolate different styles thanks to its high-level feature representation of a 3D scene. As illustrated in Fig. 7, we linearly interpolate the feature maps of a specific view by using four different styles at four corners. Unlike previous NeRF-based 3D style transfer that supports style interpolation by interpolating one-hot latent vectors [11], StyleRF can interpolate arbitrary numbers of unseen new styles by interpolating features of the scene, yielding more smooth" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.089, + 0.632, + 0.099 + ], + "angle": 0, + "content": "Ground Truth Novel View" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.1, + 0.632, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.246, + 0.612, + 0.256 + ], + "angle": 0, + "content": "Generated Mask" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.1, + 0.759, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.246, + 0.812, + 0.257 + ], + "angle": 0, + "content": "Spatial Composition" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.1, + 0.887, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Figure 8. Compositional 3D style transfer. Given the 3D-consistent segmentation masks, StyleRF can create infinite combinations of styles by spatial composition." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.323, + 0.892, + 0.369 + ], + "angle": 0, + "content": "and harmonious interpolation. Hence, StyleRF can not only transfer arbitrary styles in a zero-shot manner but also generate non-existent stylization via multi-style interpolation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Compositional 3D style transfer. Thanks to its precise geometry reconstruction, StyleRF can be seamlessly integrated with NeRF-based object segmentation [12, 26, 65] for compositional 3D style transfer. As shown in Fig. 8, we apply 3D-consistent segmentation masks to the feature maps and apply different styles to stylize the contents inside and outside the masks separately. We can see that the edges of the masks can be blended more softly by applying the segmentation masks to the feature maps instead of RGB images. Due to its zero-shot nature, StyleRF can create infinite combinations of styles without additional training, producing numerous artistic creations and inspirations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.618, + 0.582 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.82 + ], + "angle": 0, + "content": "In this paper, we present StyleRF, a novel zero-shot 3D style transfer method that balances the three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. By representing the 3D scene with an explicit grid of high-level features, we can faithfully restore high-fidelity geometry through volume rendering. Then we perform style transfer on the feature space of the scene, leading to high-quality zero-shot stylization results. We innovatively design sampling-invariant content transformation to maintain multiview consistency and deferred style transformation to increase efficiency. We demonstrate that StyleRF achieves superior 3D stylization quality than previous zero-shot 3D style transfer methods, and can be extended to various interesting applications for artistic 3D creations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.647, + 0.848 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This project is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with project number RT18/22." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8345" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[6] Yaosen Chen, Qi Yuan, Zhiqiang Li, Yuegen Liu, Wei Wang, Chaoping Xie, Xuming Wen, and Qien Yu. Upst-nerf: Universal photorealistic style transfer of neural radiance fields for 3d scene. In *arxiv*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[8] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.47, + 0.697 + ], + "angle": 0, + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[10] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14304-14313, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.772, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[17] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[18] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 783-791, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13869-13878, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[20] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[22] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8346" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[26] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[28] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2019. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[29] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[30] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[32] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.47, + 0.567 + ], + "angle": 0, + "content": "[33] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.637 + ], + "angle": 0, + "content": "[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[37] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[39] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.893, + 0.203 + ], + "angle": 0, + "content": "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.328 + ], + "angle": 0, + "content": "[42] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.328, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[43] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.893, + 0.442 + ], + "angle": 0, + "content": "[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.51, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[46] Charles R Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J Guibas. Volumetric and multi-view cnns for object classification on 3d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.893, + 0.635 + ], + "angle": 0, + "content": "[47] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. International Journal of Computer Vision, 126(11):1199-1219, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.635, + 0.893, + 0.692 + ], + "angle": 0, + "content": "[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.893, + 0.734 + ], + "angle": 0, + "content": "[49] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.893, + 0.79 + ], + "angle": 0, + "content": "[50] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8242-8250, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.893, + 0.831 + ], + "angle": 0, + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8347" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.135 + ], + "angle": 0, + "content": "[53] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[54] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.246 + ], + "angle": 0, + "content": "[55] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.316 + ], + "angle": 0, + "content": "[56] Wenjing Wang, Jizheng Xu, Li Zhang, Yue Wang, and Jiaying Liu. Consistent video style transfer via compound regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12233-12240, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.319, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[57] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[58] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.468, + 0.501 + ], + "angle": 0, + "content": "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.503, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[60] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.56, + 0.468, + 0.627 + ], + "angle": 0, + "content": "[61] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.468, + 0.684 + ], + "angle": 0, + "content": "[62] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. In *The Eleventh International Conference on Learning Representations*, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.468, + 0.741 + ], + "angle": 0, + "content": "[63] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.743, + 0.468, + 0.812 + ], + "angle": 0, + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.814, + 0.468, + 0.855 + ], + "angle": 0, + "content": "[65] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8348" + } + ] +] \ No newline at end of file diff --git a/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_origin.pdf b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05b6454eb645ac41d41bbf78cb857735f74a85bc --- /dev/null +++ b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/5ea66819-1f3c-4dd7-a1a5-de9269668193_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc23b75d25f07173e3de9e4447d610065f77128ab1e3b40fccb8ded3d52a5ff4 +size 9018888 diff --git a/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/full.md b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4e35ad59d5d3ff865a25c14bdfef3da6c460c662 --- /dev/null +++ b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/full.md @@ -0,0 +1,345 @@ +# StyleRF: Zero-shot 3D Style Transfer of Neural Radiance Fields + +Kunhao Liu $^{1}$ Fangneng Zhan $^{2}$ Yiwen Chen $^{1}$ Jiahui Zhang $^{1}$ + +Yingchen Yu1 Abdulmotaleb El Saddik3,5 Shijian Lu1* Eric Xing4,5 + +$^{1}$ Nanyang Technological University $^{2}$ Max Planck Institute for Informatics + +3University of Ottawa 4Carnegie Mellon University 5MBZUAI + +![](images/e9535ab4afae93e92cde991a6fd7118d091507d72044ea8a75fac6ea3747ff12.jpg) + +![](images/206733db5d51c349e342db393d70e334782f4ca31661a7f780c24b649e3d50f1.jpg) + +![](images/ce57bbde87f035ae8eac471049f40853abb79ac1ec8960262aa8d4647e9f2a30.jpg) + +![](images/c5871b8bdcf5616393f246b73870d6a3cdade9a12310ce0abc216dbb591b2daf.jpg) + +![](images/6e9414d224bb9426eb476c87148eb153b0f21ab67b97060e53e771ab9632bfdb.jpg) +Multi-view Content Images + +![](images/bef773278881f74423435479aa4582b7f12771053d7461efd4aeccbbd671a711.jpg) +Reference Style Images + Stylized Novel Views + +![](images/2956a5060b129f79a16e54be3b673dcf3013ac1d6db6926ccc1a1cf942f43381.jpg) +Figure 1. Zero-shot 3D Style Transfer. Given a set of Multi-view Content Images of a 3D scene, StyleRF can transfer arbitrary Reference Styles to the 3D scene in a zero-shot manner, rendering high-quality Stylized Novel Views with superb multi-view consistency. + +![](images/59ebdac68649afb08d4027be86917ffce20a35ece0784cbfa339339b2cf61a29.jpg) + +# Abstract + +3D style transfer aims to render stylized novel views of a 3D scene with multi-view consistency. However, most existing work suffers from a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. We propose StyleRF (Style Radiance Fields), an innovative 3D style transfer technique that resolves the three-way dilemma by performing style transformation within the feature space of a radiance field. StyleRF employs an explicit grid of high-level features to represent 3D scenes, with which high-fidelity geometry can be reliably restored via volume rendering. In addition, it transforms the grid features according to the reference style which directly leads to high-quality zero-shot style transfer. StyleRF consists of two innovative designs. The first is sampling-invariant content transformation that makes the transformation invari + +ant to the holistic statistics of the sampled 3D points and accordingly ensures multi-view consistency. The second is deferred style transformation of 2D feature maps which is equivalent to the transformation of 3D points but greatly reduces memory footprint without degrading multi-view consistency. Extensive experiments show that StyleRF achieves superior 3D stylization quality with precise geometry reconstruction and it can generalize to various new styles in a zero-shot manner. Project website: https://kunhao-liu.github.io/StyleRF/ + +# 1. Introduction + +Given a set of multi-view images of a 3D scene and an image capturing a target style, 3D style transfer aims to generate novel views of the 3D scene that have the target style consistently across the generated views (Fig. 1). Neural style transfer has been investigated extensively, and state-of-the-art methods allow transferring arbitrary styles in a + +zero-shot manner. However, most existing work focuses on style transfer across 2D images [15, 21, 24] but cannot extend to a 3D scene that has arbitrary new views. Prior studies [19, 22, 37, 39] have shown that naively combining 3D novel view synthesis and 2D style transfer often leads to multi-view inconsistency or poor stylization quality, and 3D style transfer should optimize novel view synthesis and style transfer jointly. + +However, the current 3D style transfer is facing a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to new styles. Different approaches have been investigated to resolve the three-way dilemma. For example, multiple style transfer [11, 22] requires a set of pre-defined styles but cannot generalize to unseen new styles. Point-cloud-based style transfer [19, 37] requires a pre-trained depth estimation module that is prone to inaccurate geometry reconstruction. Zero-shot style transfer with neural radiance fields (NeRF) [8] cannot capture detailed style patterns and textures as it implicitly injects the style information into neural network parameters. Optimization-based style transfer [17, 39, 63] suffers from slow optimization and cannot scale with new styles. + +In this work, we introduce StyleRF to resolve the three-way dilemma by performing style transformation in the feature space of a radiance field. A radiance field is a continuous volume that can restore more precise geometry than point clouds or meshes. In addition, transforming a radiance field in the feature space is more expressive with better stylization quality than implicit methods [8], and it can also generalize to arbitrary styles. We construct a 3D scene representation with a grid of deep features to enable feature transformation. In addition, multi-view consistent style transformation in the feature space could be achieved by either transforming the whole feature grid or transforming the sampled 3D points. We adopt the latter as the former incurs much more computational cost during training to stylize the whole feature grid in every iteration, whereas the latter can reduce computational cost through decreasing the size of training patch and the number of sampled points. However, applying off-the-shelf style transformations to a batch of sampled 3D points impairs the multi-view consistency as they are conditioned on the holistic statistics of the batch. Beyond that, transforming every sampled 3D point is memory-intensive since NeRF needs to query hundreds of sampled points along each ray for rendering a single pixel. + +We decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST), the former eliminating the dependency on holistic statistics of sampled point batch and the latter deferring style transformation to 2D feature maps for better efficiency. In SICT, we introduce volume-adaptive normalization that learns the mean and variance of the + +whole volume instead of computing them from a sampled batch. In addition, we apply channel-wise self-attention to transform each 3D point independently to make it conditioned on the feature of that point regardless of the holistic statistics of the sampled batch. In DST, we defer the style transformation to the volume-rendered 2D feature maps based on the observation that the style transformation of each point is the same. By formulating the style transformation by pure matrix multiplication and adaptive bias addition, transforming 2D feature maps is mathematically equivalent to transforming 3D point features but it saves computation and memory greatly. Thanks to the memory-efficient representation of 3D scenes and deferred style transformation, our network can train with $256 \times 256$ patches directly without requiring sub-sampling like previous NeRF-based 3D style transfer methods [8, 11, 22]. + +The contributions of this work can be summarized in three aspects. First, we introduce StyleRF, an innovative zero-shot 3D style transfer framework that can generate zero-shot high-quality 3D stylization via style transformation within the feature space of a radiance field. Second, we design sampling-invariant content transformation and deferred style transformation, the former achieving multiview consistent transformation by eliminating dependency on holistic statistics of sampled point batch while the latter greatly improves stylization efficiency by deferring style transformation to 2D feature maps. Third, extensive experiments show that StyleRF achieves superior 3D style transfer with accurate geometry reconstruction, high-quality stylization, and great generalization to new styles. + +# 2. Related Work + +Neural scene representations. 3D scene representation has been extensively studied in recent years with different ways of representations such as volumes [23,27,46,49,59], point clouds [1, 45], meshes [25, 55], depth maps [20, 30], and implicit functions [7, 34, 41, 61]. These methods adopt differentiable rendering which enables model optimization by using 2D multi-view images. Among them, Neural Radiance Field (NeRF) [36] can render a complex 3D scene with high fidelity and accurate geometry. It represents scenes with an implicit coordinate function that maps each 3D coordinate to a density value and a color value, and employs volume rendering to generate images of novel views. However, the implicit coordinate function is represented by a large multilayer perceptron (MLP) that is often hard to optimize and slow to infer. Serval studies adopt a hybrid representation [3, 4, 10, 14, 31, 33, 38, 44, 52, 62] to speed up the reconstruction and rendering. They employ explicit data structures such as discrete voxel grids [14,52], decomposed tensors [3, 4, 13], hash maps [38], etc. to store features or spherical harmonics, enabling fast convergence and inference. + +![](images/63c43f0464217d0a98f4d7a93a99827ff2252e30e31db7f33fadfe04ceff4228.jpg) +Figure 2. The framework of StyleRF. For a batch of sampled points along a ray $\mathbf{r}$ , the corresponding features $F_{i}, i \in [1,2,\dots,N]$ are first extracted, each of which is transformed to $\bar{F}_{i}$ independently via Sampling-Invariant Content Transformation, regardless of the holistic statistics of the point batch. $\bar{F}_{i}$ is then transformed to a feature map $\bar{F}_{c}$ via Volume Rendering. After that, the Deferred Style Transformation transforms $\bar{F}_{c}$ to the feature map $F_{cs}$ adaptively using the sum weight of the sampled points $w_{\mathbf{r}}$ along the ray $\mathbf{r}$ and the style information $T, \mu(F_{s})$ , and $\sigma(F_{s})$ . Finally, a stylized novel view is generated via a CNN decoder. + +Although most existing work extracts features as middle-level representations of scenes, the extracted features are usually an intermediate output of neural networks which have little semantic meanings and are not suitable for the style transfer task. We introduce decomposed tensors [4] to store high-level features extracted by pre-trained CNNs, which enables transformations in feature space as well as efficient training and inference. Though [3,16,40] also render feature maps instead of RGB maps, they are computationally intensive and usually work with low-resolution feature maps. StyleRF can instead render full-resolution feature maps (the same as the output RGB images) efficiently and it uses high-level features largely for transformation only. + +Neural style transfer. Neural style transfer aims at rendering a new image that contains the content structure of one image and the style patterns of another. The seminal work in [15] shows that multi-level feature statistics extracted from intermediate layers of pre-trained CNNs could be used as a representation of the style of an artistic image, but it treats style transfer as a slow and iterative optimization task. [9,21,24,28,29,32,43,50,58] utilize feed-forward networks to approximate the optimization procedure to speed up rendering. Among them, [9,21,28,29,32,43,50,58] can achieve zero-shot style transfer by applying transformations to the high-level features extracted by pre-trained CNNs, where the feature transformations can be achieved by matching second-order statistics [21, 29], linear transformation [28,58], self-attention transformation [9,32,43], etc. Video style transfer extends style transfer to videos for injecting target styles consistently across adjacent video frames. Several studies leverage optical flow [5,18,47,56,57] as temporal constraints to estimate the movement of video contents. They can produce smooth videos, but have little knowledge of the underlying 3D geometry and cannot render consistent frames in arbitrary views [19,37]. + +Huang et al. first tackle stylizing complex 3D scenes [19]. They construct a 3D scene by back-projecting + +image features into the 3D space to form a point cloud and then perform style transformation on the features of 3D points. Their method can achieve zero-shot style transfer, but requires an error-prone pre-trained depth estimator to model scene geometry. [37] also constructs a point cloud for stylization but it mainly focuses on monocular images. Instead, [6, 8, 11, 22, 39, 63] use NeRF [36] as the 3D representation which can reconstruct scene geometry more faithfully. [6] is a photorealistic style transfer method that can only transfer the color tone of style images. [39, 63] achieve 3D style transfer via optimization and can produce visually high-quality stylization, but they require a time-consuming optimization procedure for every reference style. [11, 22] employ latent codes to represent a set of pre-defined styles, but cannot generalize to unseen styles. [8] can achieve arbitrary style transfer by implicitly instilling the style information into MLP parameters. However, it can only transfer the color tone of style images but cannot capture detailed style patterns. StyleRF can transfer arbitrary style in a zero-shot manner, and it can capture style details such as strokes and textures as well. + +# 3. Method + +The overview of StyleRF is shown in Fig. 2. For a batch of sampled points along a ray $\mathbf{r}$ , the corresponding features $F_{i}, i \in [1,2,\dots,N]$ are first extracted from the feature grid described in Sec. 3.1, each of which is transformed to $\bar{F}_{i}$ independently via Sampling-Invariant Content Transformation (SICT) described in Sec. 3.2.1, regardless of the holistic statistics of the point batch. $\bar{F}_{i}$ is then transformed to a feature map $\bar{F}_{c}$ via Volume Rendering. After that, the Deferred Style Transformation (DST) described in Sec. 3.2.2 transforms $\bar{F}_{c}$ to the feature map $F_{cs}$ adaptively using the sum weight of the sampled points $w_{\mathbf{r}}$ along the ray $\mathbf{r}$ and the style information $T, \mu(F_{s})$ , and $\sigma(F_{s})$ . Finally, a stylized novel view is generated via a CNN decoder. + +# 3.1. Feature Grid 3D Representation + +To model a 3D scene with deep features, we use a continuous volumetric field of density and radiance. Different from the original NeRF [36], for every queried 3D position $x \in \mathbb{R}^3$ , we get a volume density $\sigma(x)$ and a multi-channel feature $F(x) \in \mathbb{R}^C$ instead of an RGB color, where $C$ is the number of the feature channels. Then we can get the feature of any rays $\mathbf{r}$ passing through the volume by integrating sampled points along the ray via approximated volume rendering [36]: + +$$ +F (\mathbf {r}) = \sum_ {i = 1} ^ {N} w _ {i} F _ {i}, \tag {1} +$$ + +where $w_{i} = \exp \left(-\sum_{j = 1}^{i - 1}\sigma_{j}\delta_{j}\right)$ $(1 - \exp (-\sigma_i\delta_i))$ (2) + +where $\sigma_{i}, F_{i}$ denotes the volume density and feature of sampled point $i$ , $w_{i}$ denotes the weight of $F_{i}$ in the ray $\mathbf{r}$ , and $\delta_{i}$ is the distance between adjacent samples. We disable view-dependency effect for better multi-view consistency. + +Then we can generate feature maps capturing high-level features and map them to RGB space using a 2D CNN decoder. However, unlike [3,16,40], we render full-resolution feature maps which have the same resolution as the final RGB images rather than down-sampled feature maps. Rendering full-resolution feature maps has two unique features: 1) it discards up-sampling operations which cause multiview inconsistency in general [16], 2) it removes aliasing when rendering low-resolution feature maps [2] which causes severe flickering effects in stylized RGB videos. + +Directly using 3D voxel grid to store features is memory-intensive. We thus adopt vector-matrix tensor decomposition [4] that relaxes the low-rank constraints for two modes of a 3D tensor and factorizes tensors into compact vector and matrix factors, which lowers the space complexity from $\mathcal{O}(n^3)$ to $\mathcal{O}(n^2)$ , massively reducing the memory footprint. We employ a density grid to store volume density and a feature grid to store multi-channel features respectively. + +# 3.2. Feature Transformation for Style Transfer + +Once we have the feature grid representation of a scene, we can tackle the task of stylizing 3D scenes. Given a reference style image, our goal is to render stylized novel views of the 3D scene with multi-view consistency. To achieve this, we apply transformations to the features of the grid. + +One plausible solution to this task is to apply style transfer to the feature grid directly. This solution is efficient in evaluations as it can render any stylized views with a single style transfer process only. However, it is impractical to train such transformation as it needs to stylize the whole feature grid in every iteration. Another solution is + +![](images/88920ccc6bd8c9b413e7be146402b7e0097d6e49ddd7d117819e8c001ec9822a.jpg) +(a) Vanilla IN. + +![](images/5d3d63ecf03bf88cb02ea408e43c8c83ded6f017991cfecd795dc26e9f49e533.jpg) +(b) Volume-adaptive IN. +Figure 3. Comparison between vanilla instance normalization (IN) in (a) and volume-adaptive IN in (b). During evaluation, volume-adaptive IN uses learned mean and standard-deviation, discarding dependency over the sampled point batch's holistic statistics (indicated by the red arrows in the left graph). + +to apply an off-the-shelf zero-shot style transfer method to the features of the sampled 3D points. While this solution can reduce computational cost through decreasing the size of training patch and the number of sampled points, it has two problems: 1) vanilla zero-shot style transformation is conditioned on holistic statistics of the sampled point batch [21, 28, 32], which violates multi-view consistency in volume rendering as the feature transformation of a specific 3D point will vary across different sampled points; 2) volume rendering requires sampling hundreds of points along a single ray, which makes transformation on the point batch memory-intensive. + +Motivated by the observation that style transformation is conditioned on both content information and style information, we decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST). After the decomposition, SICT will be conditioned solely on the content information while DST conditioned solely on the style information, more details to be elaborated in the ensuing subsections. + +# 3.2.1 Sampling-invariant Content Transformation + +Given a batch of sampled points, we can get their corresponding features $F_{i} \in \mathbb{R}^{C}, i \in [1,2,\dots,N]$ from the feature grid, where $N$ is the number of the sampled points along a ray and $C$ is the number of the feature channels. The goal of SICT is to transform the extracted features $F_{i}$ so that they can be better stylized. We formulate SICT as a channel-wise self-attention operation to the features after instance normalization (IN) [54]. Specifically, we formulate $Q(\text{query}), K(\text{key})$ , and $V(\text{value})$ as: + +$$ +Q = q \left(\operatorname {N o r m} \left(F _ {i}\right)\right), \tag {3} +$$ + +$$ +K = k \left(\operatorname {N o r m} \left(F _ {i}\right)\right), \tag {4} +$$ + +$$ +V = v \left(\operatorname {N o r m} \left(F _ {i}\right)\right), \tag {5} +$$ + +where $q, k, v$ are $1 \times 1$ convolution layers which reduce the channel number from $C$ to $C'$ for computational efficiency, and Norm denotes the IN. However, as shown in Fig. 3, + +![](images/212ea898387ba8c5d3e5d7e2190e1dce3eb3444f467bcf3b2e6bd0372ff953b6.jpg) +Figure 4. Deferred style transformation. We apply the style transformation to the volume-rendered feature maps $\bar{F}_c$ according to the style feature maps $F_s$ . To ensure multi-view consistency, we modulate the bias (e.g. the mean value of the style feature maps $\mu(F_s)$ ) with the sum weight of sampled points along each ray $w_{\mathbf{r}}$ . + +vanilla IN calculates per-dimension mean and standard-deviation of the batch of sampled points, which varies with different sampled points and incurs multi-view inconsistency accordingly. Thus we design volume-adaptive IN which, during training, keeps running estimates of the computed mean and standard-deviation, and uses them for normalization during evaluations (instead of computing from the sampled point batch). Through volume-adaptive IN, we can ensure that the content transformation is consistent regardless of the sampled point batch's holistic statistics. + +Channel-wise self-attention can thus be implemented by: + +$$ +\bar {F} _ {i} = V \otimes \operatorname {S o f t m a x} (\widetilde {\operatorname {c o v}} (Q, K)), \tag {6} +$$ + +where $\otimes$ denotes matrix multiplication and $\widetilde{cov}(Q, K) \in \mathbb{R}^{N \times C' \times C'}$ denotes the covariance matrix in the channel dimension. + +# 3.2.2 Deferred Style Transformation + +After applying SICT to the features of each 3D point, we apply DST to the volume-rendered 2D feature maps $\bar{F}_c$ rather than 3D point features $\bar{F}_i$ . To ensure multi-view consistency, we formulate the transformation as matrix multiplication and adaptive bias addition as illustrated in Fig. 4. + +Specifically, we first extract feature maps $F_{s}$ of the reference style $S$ using a pre-trained VGG [51], and then generate the style transformation matrix $T \in \mathbb{R}^{C' \times C'}$ using feature covariance $cov(F_{s})$ following [28]. Next, we apply matrix multiplication with $T$ to the feature maps $\bar{F}_{c}$ and use a $1 \times 1$ convolution layer conv without bias to restore the channel number from $C'$ to $C$ . Though these operations can partially instill style information, they are not expressive enough without bias addition containing style information [58]. Thus following [21], we multiply the feature + +maps with the standard-deviation value $\sigma(F_s)$ and add the mean value $\mu(F_s)$ . To ensure it is equivalent when applying the transformation to either 3D point features or 2D feature maps, we adaptively modulate the mean value $\mu(F_s)$ with the sum weight of sampled points along each ray $w_{\mathbf{r}}$ . DST can be mathematically formulated by: + +$$ +F _ {c s} = \operatorname {c o n v} \left(T \otimes \bar {F} _ {c}\right) \times \sigma \left(F _ {s}\right) + w _ {\mathbf {r}} \times \mu \left(F _ {s}\right), \tag {7} +$$ + +$$ +\text {w h e r e} \quad \bar {F} _ {c} = \sum_ {i = 1} ^ {N} w _ {i} \bar {F} _ {i}, w _ {\mathbf {r}} = \sum_ {i = 1} ^ {N} w _ {i}, \mathbf {r} \in \mathcal {R} \tag {8} +$$ + +where $w_{i}$ denotes the weight of sampled point $i$ (Eq. (2)), $\bar{F}_i$ denotes the feature of sample $i$ after SICT, and $\mathcal{R}$ is the set of rays in each training batch. + +Note $conv$ is a $1 \times 1$ convolution layer without bias, so it is basically a matrix multiplication operation. And $\sigma(S), \mu(S)$ are scalars. Together with the adaptive bias modulation $w_{\mathbf{r}}$ , Eq. (7) can be reformulated by: + +$$ +F _ {c s} = \sum_ {i = 1} ^ {N} w _ {i} \left(\underbrace {\operatorname {c o n v} \left(T \otimes \bar {F} _ {i}\right) \times \sigma \left(F _ {s}\right) + \mu \left(F _ {s}\right)} _ {\text {(i)}}\right), \tag {9} +$$ + +where part (i) can be seen as applying style transformation on every 3D point feature independently before volume rendering. This proves that applying DST on 2D feature maps is equivalent to applying the transformation on 3D points' features, maintaining multi-view consistency. The full derivation of Eq. (9) is provided in the appendix. + +Finally, we adopt a 2D CNN decoder to project the stylized feature maps $F_{cs}$ to RGB space to generate the final stylized novel view images. + +# 3.3. Two-stage Model Training + +The training of our model is divided into the feature grid training stage and the stylization training stage, the former is trained with the target of novel view synthesis, and the latter is trained with the target of style transfer. + +Feature grid training stage (First stage). We first learn the feature grid 3D representation for the novel view synthesis task, in preparation for performing feature transformation for style transfer. We train the feature grid and the 2D CNN decoder simultaneously, with the supervision of both RGB images and their bilinearly up-sampled feature maps extracted from ReLU3_1 layer of pre-trained VGG [51]. By aligning the VGG features with the feature grid, the reconstructed features acquire semantic information. We use density grid pre-trained solely on RGB images since the supervising feature maps are not strictly multi-view consistent. The training objective is the mean square error (MSE) between the predicted and ground truth feature maps and RGB images. Following [19, 37], we use perceptual + +loss [24] as additional supervision to increase reconstructed image quality. The overall loss function is: + +$$ +\begin{array}{l} \mathcal {L} _ {g r i d} = \sum_ {r \in \mathcal {R}} \left\| \hat {F} (\mathbf {r}) - F (\mathbf {r}) \right\| _ {2} ^ {2} + \left\| \hat {I} _ {\mathcal {R}} - I _ {\mathcal {R}} \right\| _ {2} ^ {2} \\ + \sum_ {l \in l _ {p}} \left\| \mathcal {F} ^ {l} (\hat {I} _ {\mathcal {R}}) - \mathcal {F} ^ {l} (I _ {\mathcal {R}}) \right\| _ {2} ^ {2}, \tag {10} \\ \end{array} +$$ + +where $\mathcal{R}$ is the set of rays in each training batch, $\hat{F} (\mathbf{r}), F(\mathbf{r})$ are the predicted and ground truth feature of ray $\mathbf{r}$ , $\hat{I}_{\mathcal{R}}, I_{\mathcal{R}}$ are the predicted and ground truth RGB image, $l_{p}$ denotes the set of VGG layers that compute perceptual loss, $\mathcal{F}^l$ denotes the feature maps of the $l$ th layer of pre-trained VGG network. + +Stylization training stage (Second stage). Our model learns to stylize novel views in the second stage. We freeze the feature grid, train the style transfer module, and finetune the CNN decoder. Thanks to the memory-efficient representation of 3D scenes and DST, unlike [8, 11, 48], our model can be trained directly on $256 \times 256$ patches, making patch sub-sampling algorithm [8, 11, 22, 48] unnecessary. We use the same loss as [21] where the content loss $\mathcal{L}_c$ is the MSE of the feature maps and the style loss $\mathcal{L}_s$ is the MSE of the channel-wise feature mean and standard-deviation: + +$$ +\mathcal {L} _ {\text {s t y l i z a t i o n}} = \mathcal {L} _ {c} + \lambda \mathcal {L} _ {s}, \tag {11} +$$ + +where $\lambda$ balances the content preservation and the stylization effect. + +# 4. Experiments + +We evaluate StyleRF extensively with qualitative experiments in Sec. 4.1, quantitative experiments in Sec. 4.2 and ablation studies in Sec. 4.3. We demonstrate two applications of StyleRF in Sec. 4.4. The implementation details are provided in the appendix. + +# 4.1. Qualitative Experiments + +We evaluate StyleRF over two public datasets including LLFF [35] that contains real scenes with complex geometry structures and Synthetic NeRF [36] that contains $360^{\circ}$ views of objects. In addition, we benchmark StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8] with their released codes. We perform comparisons on LLFF dataset [35]. + +Fig. 5 shows qualitative comparisons. We can see that StyleRF achieves clearly better stylization with more precise geometry reconstruction. Specifically, StyleRF can generate high-definition stylization with realistic textures and patterns of style images. The superior stylization is largely attributed to our transformation design that allows working in the feature space with full-resolution feature maps. As illustrated in the highlight boxes, StyleRF + +
MethodShort-range ConsistencyLong-range Consistency
LPIPSRMSELPIPSRMSE
AdaIN [21]0.1520.1230.2200.186
CCPL [60]0.1100.1060.1910.174
ReReVST [57]0.0980.0800.1860.146
LSNV [19]0.0930.0920.1810.155
Hyper [8]0.0840.0680.1310.101
Ours0.0720.0820.1490.137
+ +Table 1. Results on consistency. We compare StyleRF with the state-of-the-art on consistency using LPIPS $(\downarrow)$ and RMSE $(\downarrow)$ . + +can successfully restore the intricate geometry of complex scenes thanks to its radiance field representations. In addition, only StyleRF faithfully transfers the squareness texture in the second style image. Furthermore, StyleRF can robustly generalize to new styles in a zero-shot manner and can adapt well to $360^{\circ}$ dataset as illustrated in Fig. 1. As a comparison, LSNV [19] fails to capture fine-level geometry like the bones of the T-Rex and the petals of the flower while Hyper [8] produces very blurry stylization. + +# 4.2. Quantitative Results + +3D style transfer is a very new and under-explored task and there are few metrics for quantitative evaluation of stylization quality. Hence, we manage to evaluate the multiview consistency only. In our experiments, we warp one view to the other according to the optical flow [53] using softmax splatting [42], and then computed the masked RMSE score and LPIPS score [64] to measure the stylization consistency. Following [8, 11, 19], we compute the short-range and long-range consistency scores which compare adjacent views and far-away views respectively. We compare StyleRF against two state-of-the-art zero-shot 3D style transfer methods Hyper [8] and LSNV [19], one SOTA single-frame-based video style transfer method CCPL [60], one SOTA multi-frames-based video style transfer method ReReVST [57], and one classical image style transfer method AdaIN [21]. + +It can be seen from Tab. 1 that StyleRF significantly outperforms image style transfer approach [21] and video style transfer approach [57, 60] which capture little information about the underlying 3D geometry. In addition, StyleRF achieves better consistency than point-cloud-based 3D style transfer [19] as well. Note Hyper [8] achieves slightly better LPIPS and RMSE scores than our method, largely because it produces over-smooth results and inadequate stylization as shown in Fig. 5. + +# 4.3. Ablation Studies + +We design two innovative techniques to improve the stylization quality and maintain the multi-view consistency. + +![](images/ee025dc9e4e1f6cfb462963d8fc32858f3e5c559245d7aa4c7010c15c28ec177.jpg) +Figure 5. Comparison of StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8]. For each of the two sample Scenes and reference Styles, StyleRF produces clearly better 3D style transfer and depth estimation. Check zoom-in for details. + +The first is volume-adaptive instance normalization which uses the learned mean and variance of the whole volume during inference, eliminating the dependency on holistic statistics of the sampled point batch. The second is the adaptive bias addition in DST, which improves the stylization quality using bias capturing style information. We evaluate the two designs to examine how they contribute to the overall stylization of our method. + +Volume-adaptive instance normalization. We compare our volume-adaptive instance normalization (IN) with vanilla IN and StyleRF without IN. As Fig. 6 (c) shows, vanilla IN produces severe block-shape artifacts as the + +transformation of each batch is conditioned on the holistic statistics of itself, thus each batch (i.e. block in the image) produces inconsistent stylization which leads to the artifacts. However, if we discard IN as shown in Fig. 6 (d), the multi-view consistency can maintain but the stylization quality compromises a lot, failing to capture the correct color tone of the reference style image. This is because IN removes the original style information of the content image which facilitates the transfer of the reference style [21]. + +Adaptive bias addition. As illustrated in Fig. 6 (b), the stylization quality degrades a lot if we eliminate the adaptive bias addition in DST (Sec. 3.2.2), producing unnatural + +![](images/9d9688b3ec0995eb5feab5b0bea28bc1f57424825c892392b6475b2fc78e4535.jpg) +Ground Truth Novel View + +![](images/4988ff3367a052dbcde5bf5cccf0a0318c2a12081b6610131b78124fcea707e0.jpg) +Reference Style Image + +![](images/57d8076e5926cfcec12eab018620e7ca1e2cf67ecd60f60f9c25de28c0da1665.jpg) +(a) Ours + +![](images/74077c824818391c72be93b2ff2578f227fa1f824657e1c5339ceef321b5f7ca.jpg) +(b) w/o Adaptive Bias + +![](images/46d4595847b45fd388cf1584ad246981845658df3319c3c0c431849ae73133ab.jpg) +(c) Vanilla IN + +![](images/1f2cf7efd36889bc1754a0c64a25a05e0d400c4507f2a6a5ab163267feec3767.jpg) +(d) w/o IN + +![](images/93c09f22dce827bbd0c6e32cb7aab03a59b216b7c5f37bde10227a14000938ee.jpg) +Figure 6. Ablation studies. (a) shows the stylization of our full pipeline. (b) shows the stylization without the adaptive bias. (c) shows the stylization when replacing the volume-adaptive instance normalization (IN) with vanilla IN. (d) shows the stylization without any IN. +Figure 7. Multi-style interpolation. StyleRF can smoothly interpolate between arbitrary styles by interpolating features of the scene. + +stylization compared to the stylization of our full pipeline in Fig. 6 (a). This is because bias usually contains crucial style information such as the overall color tone [58]. StyleRF employs bias addition that is adaptively modulated by the weight of each ray, improving the stylization quality and keeping multi-view consistency concurrently. + +# 4.4. Applications + +StyleRF can be easily extended along different directions with different applications. We provide two possible extensions in the ensuing subsections. + +Multi-style interpolation. StyleRF can smoothly interpolate different styles thanks to its high-level feature representation of a 3D scene. As illustrated in Fig. 7, we linearly interpolate the feature maps of a specific view by using four different styles at four corners. Unlike previous NeRF-based 3D style transfer that supports style interpolation by interpolating one-hot latent vectors [11], StyleRF can interpolate arbitrary numbers of unseen new styles by interpolating features of the scene, yielding more smooth + +![](images/811621453ecb598dee2ad002244040ce4d61eda8bbd4b6d8d0022da849410e12.jpg) +Ground Truth Novel View +Generated Mask + +![](images/d70f897c0e5096e80ad8d2cfc9fa20eb8ff7345dd54cff87a100fe05ec6564e3.jpg) +Spatial Composition +Figure 8. Compositional 3D style transfer. Given the 3D-consistent segmentation masks, StyleRF can create infinite combinations of styles by spatial composition. + +![](images/ea120450f43283200e27d9f472aefb4f871ab7655d0c5eb3b3bf741242516c3d.jpg) + +and harmonious interpolation. Hence, StyleRF can not only transfer arbitrary styles in a zero-shot manner but also generate non-existent stylization via multi-style interpolation. + +Compositional 3D style transfer. Thanks to its precise geometry reconstruction, StyleRF can be seamlessly integrated with NeRF-based object segmentation [12, 26, 65] for compositional 3D style transfer. As shown in Fig. 8, we apply 3D-consistent segmentation masks to the feature maps and apply different styles to stylize the contents inside and outside the masks separately. We can see that the edges of the masks can be blended more softly by applying the segmentation masks to the feature maps instead of RGB images. Due to its zero-shot nature, StyleRF can create infinite combinations of styles without additional training, producing numerous artistic creations and inspirations. + +# 5. Conclusion + +In this paper, we present StyleRF, a novel zero-shot 3D style transfer method that balances the three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. By representing the 3D scene with an explicit grid of high-level features, we can faithfully restore high-fidelity geometry through volume rendering. Then we perform style transfer on the feature space of the scene, leading to high-quality zero-shot stylization results. We innovatively design sampling-invariant content transformation to maintain multiview consistency and deferred style transformation to increase efficiency. We demonstrate that StyleRF achieves superior 3D stylization quality than previous zero-shot 3D style transfer methods, and can be extended to various interesting applications for artistic 3D creations. + +# Acknowledgement + +This project is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with project number RT18/22. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2, 3, 4 +[5] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 3 +[6] Yaosen Chen, Qi Yuan, Zhiqiang Li, Yuegen Liu, Wei Wang, Chaoping Xie, Xuming Wen, and Qien Yu. Upst-nerf: Universal photorealistic style transfer of neural radiance fields for 3d scene. In *arxiv*, 2022. 3 +[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2 +[8] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2, 3, 6, 7 +[9] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 3 +[10] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14304-14313, 2021. 2 +[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2, 3, 6, 8 +[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776, 2022. 8 +[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit + +radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 2 +[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3 +[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3, 4 +[17] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2 +[18] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 783-791, 2017. 3 +[19] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13869-13878, 2021. 2, 3, 5, 6, 7 +[20] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2 +[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3, 4, 5, 6, 7 +[22] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2, 3, 6 +[23] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2 +[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2, 3, 6 +[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2 + +[26] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 8 +[27] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2 +[28] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2019. 3, 4, 5 +[29] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3 +[30] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 2 +[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[32] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 3, 4 +[33] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 2 +[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2 +[35] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4, 6 +[37] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2, 3, 5 +[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2 + +[39] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2, 3 +[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3, 4 +[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2 +[42] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020. 6 +[43] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 3 +[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2 +[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[46] Charles R Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J Guibas. Volumetric and multi-view cnns for object classification on 3d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2016. 2 +[47] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. International Journal of Computer Vision, 126(11):1199-1219, 2018. 3 +[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6 +[49] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2 +[50] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8242-8250, 2018. 3 +[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 + +[53] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 6 +[54] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4 +[55] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2 +[56] Wenjing Wang, Jizheng Xu, Li Zhang, Yue Wang, and Jiaying Liu. Consistent video style transfer via compound regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12233-12240, 2020. 3 +[57] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 2020. 3, 6 +[58] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 3, 5, 8 +[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2 +[60] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 6 +[61] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2 +[62] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. In *The Eleventh International Conference on Learning Representations*, 2023. 2 +[63] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2, 3 +[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[65] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 8 \ No newline at end of file diff --git a/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/images.zip b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba3eb6fad48c65b281808ff971ff284be0c686ee --- /dev/null +++ b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23a0f4a752641e81b64ed8e1b119e7f1b7deeb13c3a647805b829d3818631aa8 +size 884305 diff --git a/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/layout.json b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c1d2c4ddab366779d2e75155cd00a5df2e7cb6bd --- /dev/null +++ b/2023/StyleRF_ Zero-Shot 3D Style Transfer of Neural Radiance Fields/layout.json @@ -0,0 +1,9660 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 495, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 495, + 120 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 495, + 120 + ], + "type": "text", + "content": "StyleRF: Zero-shot 3D Style Transfer of Neural Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "spans": [ + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "text", + "content": "Kunhao Liu" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "text", + "content": " Fangneng Zhan" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "text", + "content": " Yiwen Chen" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "text", + "content": " Jiahui Zhang" + }, + { + "bbox": [ + 136, + 131, + 455, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 147, + 477, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 147, + 477, + 162 + ], + "spans": [ + { + "bbox": [ + 116, + 147, + 477, + 162 + ], + "type": "text", + "content": "Yingchen Yu1 Abdulmotaleb El Saddik3,5 Shijian Lu1* Eric Xing4,5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 158, + 168, + 435, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 168, + 435, + 180 + ], + "spans": [ + { + "bbox": [ + 158, + 168, + 435, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 168, + 435, + 180 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 158, + 168, + 435, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 168, + 435, + 180 + ], + "type": "text", + "content": "Max Planck Institute for Informatics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 182, + 417, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 182, + 417, + 194 + ], + "spans": [ + { + "bbox": [ + 173, + 182, + 417, + 194 + ], + "type": "text", + "content": "3University of Ottawa 4Carnegie Mellon University 5MBZUAI" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 216, + 171, + 310 + ], + "blocks": [ + { + "bbox": [ + 56, + 216, + 171, + 310 + ], + "lines": [ + { + "bbox": [ + 56, + 216, + 171, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 171, + 310 + ], + "type": "image", + "image_path": "e9535ab4afae93e92cde991a6fd7118d091507d72044ea8a75fac6ea3747ff12.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 195, + 217, + 304, + 301 + ], + "blocks": [ + { + "bbox": [ + 195, + 217, + 304, + 301 + ], + "lines": [ + { + "bbox": [ + 195, + 217, + 304, + 301 + ], + "spans": [ + { + "bbox": [ + 195, + 217, + 304, + 301 + ], + "type": "image", + "image_path": "206733db5d51c349e342db393d70e334782f4ca31661a7f780c24b649e3d50f1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 217, + 426, + 309 + ], + "blocks": [ + { + "bbox": [ + 319, + 217, + 426, + 309 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 426, + 309 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 426, + 309 + ], + "type": "image", + "image_path": "ce57bbde87f035ae8eac471049f40853abb79ac1ec8960262aa8d4647e9f2a30.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 449, + 213, + 544, + 314 + ], + "blocks": [ + { + "bbox": [ + 449, + 213, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 449, + 213, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 449, + 213, + 544, + 314 + ], + "type": "image", + "image_path": "c5871b8bdcf5616393f246b73870d6a3cdade9a12310ce0abc216dbb591b2daf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 314, + 190, + 414 + ], + "blocks": [ + { + "bbox": [ + 50, + 314, + 190, + 414 + ], + "lines": [ + { + "bbox": [ + 50, + 314, + 190, + 414 + ], + "spans": [ + { + "bbox": [ + 50, + 314, + 190, + 414 + ], + "type": "image", + "image_path": "6e9414d224bb9426eb476c87148eb153b0f21ab67b97060e53e771ab9632bfdb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 425, + 177, + 436 + ], + "lines": [ + { + "bbox": [ + 64, + 425, + 177, + 436 + ], + "spans": [ + { + "bbox": [ + 64, + 425, + 177, + 436 + ], + "type": "text", + "content": "Multi-view Content Images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 192, + 326, + 307, + 417 + ], + "blocks": [ + { + "bbox": [ + 192, + 326, + 307, + 417 + ], + "lines": [ + { + "bbox": [ + 192, + 326, + 307, + 417 + ], + "spans": [ + { + "bbox": [ + 192, + 326, + 307, + 417 + ], + "type": "image", + "image_path": "bef773278881f74423435479aa4582b7f12771053d7461efd4aeccbbd671a711.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 425, + 463, + 436 + ], + "lines": [ + { + "bbox": [ + 272, + 425, + 463, + 436 + ], + "spans": [ + { + "bbox": [ + 272, + 425, + 463, + 436 + ], + "type": "text", + "content": "Reference Style Images + Stylized Novel Views" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 326, + 425, + 418 + ], + "blocks": [ + { + "bbox": [ + 309, + 326, + 425, + 418 + ], + "lines": [ + { + "bbox": [ + 309, + 326, + 425, + 418 + ], + "spans": [ + { + "bbox": [ + 309, + 326, + 425, + 418 + ], + "type": "image", + "image_path": "2956a5060b129f79a16e54be3b673dcf3013ac1d6db6926ccc1a1cf942f43381.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 445, + 546, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 546, + 469 + ], + "type": "text", + "content": "Figure 1. Zero-shot 3D Style Transfer. Given a set of Multi-view Content Images of a 3D scene, StyleRF can transfer arbitrary Reference Styles to the 3D scene in a zero-shot manner, rendering high-quality Stylized Novel Views with superb multi-view consistency." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 326, + 543, + 418 + ], + "blocks": [ + { + "bbox": [ + 427, + 326, + 543, + 418 + ], + "lines": [ + { + "bbox": [ + 427, + 326, + 543, + 418 + ], + "spans": [ + { + "bbox": [ + 427, + 326, + 543, + 418 + ], + "type": "image", + "image_path": "59ebdac68649afb08d4027be86917ffce20a35ece0784cbfa339339b2cf61a29.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 143, + 479, + 192, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 503, + 290, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 503, + 290, + 696 + ], + "spans": [ + { + "bbox": [ + 45, + 503, + 290, + 696 + ], + "type": "text", + "content": "3D style transfer aims to render stylized novel views of a 3D scene with multi-view consistency. However, most existing work suffers from a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. We propose StyleRF (Style Radiance Fields), an innovative 3D style transfer technique that resolves the three-way dilemma by performing style transformation within the feature space of a radiance field. StyleRF employs an explicit grid of high-level features to represent 3D scenes, with which high-fidelity geometry can be reliably restored via volume rendering. In addition, it transforms the grid features according to the reference style which directly leads to high-quality zero-shot style transfer. StyleRF consists of two innovative designs. The first is sampling-invariant content transformation that makes the transformation invari" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 479, + 547, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 547, + 599 + ], + "type": "text", + "content": "ant to the holistic statistics of the sampled 3D points and accordingly ensures multi-view consistency. The second is deferred style transformation of 2D feature maps which is equivalent to the transformation of 3D points but greatly reduces memory footprint without degrading multi-view consistency. Extensive experiments show that StyleRF achieves superior 3D stylization quality with precise geometry reconstruction and it can generalize to various new styles in a zero-shot manner. Project website: https://kunhao-liu.github.io/StyleRF/" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 619, + 387, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 619, + 387, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 619, + 387, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Given a set of multi-view images of a 3D scene and an image capturing a target style, 3D style transfer aims to generate novel views of the 3D scene that have the target style consistently across the generated views (Fig. 1). Neural style transfer has been investigated extensively, and state-of-the-art methods allow transferring arbitrary styles in a" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 713 + ], + "type": "text", + "content": "* Shijian Lu is the corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8338" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": "zero-shot manner. However, most existing work focuses on style transfer across 2D images [15, 21, 24] but cannot extend to a 3D scene that has arbitrary new views. Prior studies [19, 22, 37, 39] have shown that naively combining 3D novel view synthesis and 2D style transfer often leads to multi-view inconsistency or poor stylization quality, and 3D style transfer should optimize novel view synthesis and style transfer jointly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 350 + ], + "type": "text", + "content": "However, the current 3D style transfer is facing a three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to new styles. Different approaches have been investigated to resolve the three-way dilemma. For example, multiple style transfer [11, 22] requires a set of pre-defined styles but cannot generalize to unseen new styles. Point-cloud-based style transfer [19, 37] requires a pre-trained depth estimation module that is prone to inaccurate geometry reconstruction. Zero-shot style transfer with neural radiance fields (NeRF) [8] cannot capture detailed style patterns and textures as it implicitly injects the style information into neural network parameters. Optimization-based style transfer [17, 39, 63] suffers from slow optimization and cannot scale with new styles." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 628 + ], + "type": "text", + "content": "In this work, we introduce StyleRF to resolve the three-way dilemma by performing style transformation in the feature space of a radiance field. A radiance field is a continuous volume that can restore more precise geometry than point clouds or meshes. In addition, transforming a radiance field in the feature space is more expressive with better stylization quality than implicit methods [8], and it can also generalize to arbitrary styles. We construct a 3D scene representation with a grid of deep features to enable feature transformation. In addition, multi-view consistent style transformation in the feature space could be achieved by either transforming the whole feature grid or transforming the sampled 3D points. We adopt the latter as the former incurs much more computational cost during training to stylize the whole feature grid in every iteration, whereas the latter can reduce computational cost through decreasing the size of training patch and the number of sampled points. However, applying off-the-shelf style transformations to a batch of sampled 3D points impairs the multi-view consistency as they are conditioned on the holistic statistics of the batch. Beyond that, transforming every sampled 3D point is memory-intensive since NeRF needs to query hundreds of sampled points along each ray for rendering a single pixel." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "We decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST), the former eliminating the dependency on holistic statistics of sampled point batch and the latter deferring style transformation to 2D feature maps for better efficiency. In SICT, we introduce volume-adaptive normalization that learns the mean and variance of the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "whole volume instead of computing them from a sampled batch. In addition, we apply channel-wise self-attention to transform each 3D point independently to make it conditioned on the feature of that point regardless of the holistic statistics of the sampled batch. In DST, we defer the style transformation to the volume-rendered 2D feature maps based on the observation that the style transformation of each point is the same. By formulating the style transformation by pure matrix multiplication and adaptive bias addition, transforming 2D feature maps is mathematically equivalent to transforming 3D point features but it saves computation and memory greatly. Thanks to the memory-efficient representation of 3D scenes and deferred style transformation, our network can train with " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " patches directly without requiring sub-sampling like previous NeRF-based 3D style transfer methods [8, 11, 22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 264, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 431 + ], + "type": "text", + "content": "The contributions of this work can be summarized in three aspects. First, we introduce StyleRF, an innovative zero-shot 3D style transfer framework that can generate zero-shot high-quality 3D stylization via style transformation within the feature space of a radiance field. Second, we design sampling-invariant content transformation and deferred style transformation, the former achieving multiview consistent transformation by eliminating dependency on holistic statistics of sampled point batch while the latter greatly improves stylization efficiency by deferring style transformation to 2D feature maps. Third, extensive experiments show that StyleRF achieves superior 3D style transfer with accurate geometry reconstruction, high-quality stylization, and great generalization to new styles." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 442, + 392, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 392, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 392, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "Neural scene representations. 3D scene representation has been extensively studied in recent years with different ways of representations such as volumes [23,27,46,49,59], point clouds [1, 45], meshes [25, 55], depth maps [20, 30], and implicit functions [7, 34, 41, 61]. These methods adopt differentiable rendering which enables model optimization by using 2D multi-view images. Among them, Neural Radiance Field (NeRF) [36] can render a complex 3D scene with high fidelity and accurate geometry. It represents scenes with an implicit coordinate function that maps each 3D coordinate to a density value and a color value, and employs volume rendering to generate images of novel views. However, the implicit coordinate function is represented by a large multilayer perceptron (MLP) that is often hard to optimize and slow to infer. Serval studies adopt a hybrid representation [3, 4, 10, 14, 31, 33, 38, 44, 52, 62] to speed up the reconstruction and rendering. They employ explicit data structures such as discrete voxel grids [14,52], decomposed tensors [3, 4, 13], hash maps [38], etc. to store features or spherical harmonics, enabling fast convergence and inference." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8339" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 539, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 539, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 539, + 194 + ], + "type": "image", + "image_path": "63c43f0464217d0a98f4d7a93a99827ff2252e30e31db7f33fadfe04ceff4228.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": "Figure 2. The framework of StyleRF. For a batch of sampled points along a ray " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": ", the corresponding features " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "F_{i}, i \\in [1,2,\\dots,N]" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " are first extracted, each of which is transformed to " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\bar{F}_{i}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " independently via Sampling-Invariant Content Transformation, regardless of the holistic statistics of the point batch. " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\bar{F}_{i}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " is then transformed to a feature map " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\bar{F}_{c}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " via Volume Rendering. After that, the Deferred Style Transformation transforms " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\bar{F}_{c}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " to the feature map " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " adaptively using the sum weight of the sampled points " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " along the ray " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": " and the style information " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "T, \\mu(F_{s})" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "inline_equation", + "content": "\\sigma(F_{s})" + }, + { + "bbox": [ + 46, + 201, + 548, + 258 + ], + "type": "text", + "content": ". Finally, a stylized novel view is generated via a CNN decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 289, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 424 + ], + "type": "text", + "content": "Although most existing work extracts features as middle-level representations of scenes, the extracted features are usually an intermediate output of neural networks which have little semantic meanings and are not suitable for the style transfer task. We introduce decomposed tensors [4] to store high-level features extracted by pre-trained CNNs, which enables transformations in feature space as well as efficient training and inference. Though [3,16,40] also render feature maps instead of RGB maps, they are computationally intensive and usually work with low-resolution feature maps. StyleRF can instead render full-resolution feature maps (the same as the output RGB images) efficiently and it uses high-level features largely for transformation only." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 688 + ], + "type": "text", + "content": "Neural style transfer. Neural style transfer aims at rendering a new image that contains the content structure of one image and the style patterns of another. The seminal work in [15] shows that multi-level feature statistics extracted from intermediate layers of pre-trained CNNs could be used as a representation of the style of an artistic image, but it treats style transfer as a slow and iterative optimization task. [9,21,24,28,29,32,43,50,58] utilize feed-forward networks to approximate the optimization procedure to speed up rendering. Among them, [9,21,28,29,32,43,50,58] can achieve zero-shot style transfer by applying transformations to the high-level features extracted by pre-trained CNNs, where the feature transformations can be achieved by matching second-order statistics [21, 29], linear transformation [28,58], self-attention transformation [9,32,43], etc. Video style transfer extends style transfer to videos for injecting target styles consistently across adjacent video frames. Several studies leverage optical flow [5,18,47,56,57] as temporal constraints to estimate the movement of video contents. They can produce smooth videos, but have little knowledge of the underlying 3D geometry and cannot render consistent frames in arbitrary views [19,37]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Huang et al. first tackle stylizing complex 3D scenes [19]. They construct a 3D scene by back-projecting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 519 + ], + "type": "text", + "content": "image features into the 3D space to form a point cloud and then perform style transformation on the features of 3D points. Their method can achieve zero-shot style transfer, but requires an error-prone pre-trained depth estimator to model scene geometry. [37] also constructs a point cloud for stylization but it mainly focuses on monocular images. Instead, [6, 8, 11, 22, 39, 63] use NeRF [36] as the 3D representation which can reconstruct scene geometry more faithfully. [6] is a photorealistic style transfer method that can only transfer the color tone of style images. [39, 63] achieve 3D style transfer via optimization and can produce visually high-quality stylization, but they require a time-consuming optimization procedure for every reference style. [11, 22] employ latent codes to represent a set of pre-defined styles, but cannot generalize to unseen styles. [8] can achieve arbitrary style transfer by implicitly instilling the style information into MLP parameters. However, it can only transfer the color tone of style images but cannot capture detailed style patterns. StyleRF can transfer arbitrary style in a zero-shot manner, and it can capture style details such as strokes and textures as well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 536, + 362, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 362, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 362, + 548 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "The overview of StyleRF is shown in Fig. 2. For a batch of sampled points along a ray " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ", the corresponding features " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "F_{i}, i \\in [1,2,\\dots,N]" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " are first extracted from the feature grid described in Sec. 3.1, each of which is transformed to " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\bar{F}_{i}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " independently via Sampling-Invariant Content Transformation (SICT) described in Sec. 3.2.1, regardless of the holistic statistics of the point batch. " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\bar{F}_{i}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " is then transformed to a feature map " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\bar{F}_{c}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " via Volume Rendering. After that, the Deferred Style Transformation (DST) described in Sec. 3.2.2 transforms " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\bar{F}_{c}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " to the feature map " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " adaptively using the sum weight of the sampled points " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{r}}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " along the ray " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " and the style information " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "T, \\mu(F_{s})" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma(F_{s})" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ". Finally, a stylized novel view is generated via a CNN decoder." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8340" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 221, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 221, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 221, + 85 + ], + "type": "text", + "content": "3.1. Feature Grid 3D Representation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "content": "To model a 3D scene with deep features, we use a continuous volumetric field of density and radiance. Different from the original NeRF [36], for every queried 3D position " + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "content": ", we get a volume density " + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\sigma(x)" + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "content": " and a multi-channel feature " + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "inline_equation", + "content": "F(x) \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "content": " instead of an RGB color, where " + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "content": " is the number of the feature channels. Then we can get the feature of any rays " + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 90, + 288, + 198 + ], + "type": "text", + "content": " passing through the volume by integrating sampled points along the ray via approximated volume rendering [36]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 129, + 205, + 287, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 205, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 129, + 205, + 287, + 239 + ], + "type": "interline_equation", + "content": "F (\\mathbf {r}) = \\sum_ {i = 1} ^ {N} w _ {i} F _ {i}, \\tag {1}", + "image_path": "47b629260809aeef5aaed1b213c146b5d23b5a41e964a4ccc5b4e2c435346e90.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 245, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 287, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 245, + 287, + 293 + ], + "type": "inline_equation", + "content": "w_{i} = \\exp \\left(-\\sum_{j = 1}^{i - 1}\\sigma_{j}\\delta_{j}\\right)" + }, + { + "bbox": [ + 56, + 245, + 287, + 293 + ], + "type": "inline_equation", + "content": "(1 - \\exp (-\\sigma_i\\delta_i))" + }, + { + "bbox": [ + 56, + 245, + 287, + 293 + ], + "type": "text", + "content": " (2)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "inline_equation", + "content": "\\sigma_{i}, F_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": " denotes the volume density and feature of sampled point " + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": " denotes the weight of " + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": " in the ray " + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 46, + 293, + 287, + 341 + ], + "type": "text", + "content": " is the distance between adjacent samples. We disable view-dependency effect for better multi-view consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 460 + ], + "type": "text", + "content": "Then we can generate feature maps capturing high-level features and map them to RGB space using a 2D CNN decoder. However, unlike [3,16,40], we render full-resolution feature maps which have the same resolution as the final RGB images rather than down-sampled feature maps. Rendering full-resolution feature maps has two unique features: 1) it discards up-sampling operations which cause multiview inconsistency in general [16], 2) it removes aliasing when rendering low-resolution feature maps [2] which causes severe flickering effects in stylized RGB videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "type": "text", + "content": "Directly using 3D voxel grid to store features is memory-intensive. We thus adopt vector-matrix tensor decomposition [4] that relaxes the low-rank constraints for two modes of a 3D tensor and factorizes tensors into compact vector and matrix factors, which lowers the space complexity from " + }, + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^3)" + }, + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^2)" + }, + { + "bbox": [ + 46, + 460, + 287, + 557 + ], + "type": "text", + "content": ", massively reducing the memory footprint. We employ a density grid to store volume density and a feature grid to store multi-channel features respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 268, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 268, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 268, + 576 + ], + "type": "text", + "content": "3.2. Feature Transformation for Style Transfer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "Once we have the feature grid representation of a scene, we can tackle the task of stylizing 3D scenes. Given a reference style image, our goal is to render stylized novel views of the 3D scene with multi-view consistency. To achieve this, we apply transformations to the features of the grid." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "One plausible solution to this task is to apply style transfer to the feature grid directly. This solution is efficient in evaluations as it can render any stylized views with a single style transfer process only. However, it is impractical to train such transformation as it needs to stylize the whole feature grid in every iteration. Another solution is" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 432, + 158 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 432, + 158 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 432, + 158 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 432, + 158 + ], + "type": "image", + "image_path": "88920ccc6bd8c9b413e7be146402b7e0097d6e49ddd7d117819e8c001ec9822a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 160, + 400, + 170 + ], + "lines": [ + { + "bbox": [ + 352, + 160, + 400, + 170 + ], + "spans": [ + { + "bbox": [ + 352, + 160, + 400, + 170 + ], + "type": "text", + "content": "(a) Vanilla IN." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 446, + 69, + 542, + 158 + ], + "blocks": [ + { + "bbox": [ + 446, + 69, + 542, + 158 + ], + "lines": [ + { + "bbox": [ + 446, + 69, + 542, + 158 + ], + "spans": [ + { + "bbox": [ + 446, + 69, + 542, + 158 + ], + "type": "image", + "image_path": "5d3d63ecf03bf88cb02ea408e43c8c83ded6f017991cfecd795dc26e9f49e533.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 160, + 534, + 170 + ], + "lines": [ + { + "bbox": [ + 454, + 160, + 534, + 170 + ], + "spans": [ + { + "bbox": [ + 454, + 160, + 534, + 170 + ], + "type": "text", + "content": "(b) Volume-adaptive IN." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 179, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 179, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 545, + 236 + ], + "type": "text", + "content": "Figure 3. Comparison between vanilla instance normalization (IN) in (a) and volume-adaptive IN in (b). During evaluation, volume-adaptive IN uses learned mean and standard-deviation, discarding dependency over the sampled point batch's holistic statistics (indicated by the red arrows in the left graph)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 386 + ], + "type": "text", + "content": "to apply an off-the-shelf zero-shot style transfer method to the features of the sampled 3D points. While this solution can reduce computational cost through decreasing the size of training patch and the number of sampled points, it has two problems: 1) vanilla zero-shot style transformation is conditioned on holistic statistics of the sampled point batch [21, 28, 32], which violates multi-view consistency in volume rendering as the feature transformation of a specific 3D point will vary across different sampled points; 2) volume rendering requires sampling hundreds of points along a single ray, which makes transformation on the point batch memory-intensive." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": "Motivated by the observation that style transformation is conditioned on both content information and style information, we decompose the style transformation into sampling-invariant content transformation (SICT) and deferred style transformation (DST). After the decomposition, SICT will be conditioned solely on the content information while DST conditioned solely on the style information, more details to be elaborated in the ensuing subsections." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 492, + 528, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 528, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 528, + 503 + ], + "type": "text", + "content": "3.2.1 Sampling-invariant Content Transformation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": "Given a batch of sampled points, we can get their corresponding features " + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "inline_equation", + "content": "F_{i} \\in \\mathbb{R}^{C}, i \\in [1,2,\\dots,N]" + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": " from the feature grid, where " + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": " is the number of the sampled points along a ray and " + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": " is the number of the feature channels. The goal of SICT is to transform the extracted features " + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": " so that they can be better stylized. We formulate SICT as a channel-wise self-attention operation to the features after instance normalization (IN) [54]. Specifically, we formulate " + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "inline_equation", + "content": "Q(\\text{query}), K(\\text{key})" + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "inline_equation", + "content": "V(\\text{value})" + }, + { + "bbox": [ + 304, + 510, + 545, + 618 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 383, + 627, + 545, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 627, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 383, + 627, + 545, + 639 + ], + "type": "interline_equation", + "content": "Q = q \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {3}", + "image_path": "4b882d40f94b70b7515849073dd7a8b3900ca65bf65e897874827ea2f5b5dadd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 383, + 642, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 642, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 383, + 642, + 545, + 654 + ], + "type": "interline_equation", + "content": "K = k \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {4}", + "image_path": "6498090c269d8887731f5ee93ec40592ac5129ad017cbf54bb29eb3c87fce335.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 384, + 656, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 656, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 384, + 656, + 545, + 669 + ], + "type": "interline_equation", + "content": "V = v \\left(\\operatorname {N o r m} \\left(F _ {i}\\right)\\right), \\tag {5}", + "image_path": "a6a542c49dec5f631108f16ee9485f7117ae8ae0e4a10e3818a46646f0ae7af3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "q, k, v" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " convolution layers which reduce the channel number from " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " for computational efficiency, and Norm denotes the IN. However, as shown in Fig. 3," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "8341" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 265, + 228 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 265, + 228 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 265, + 228 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 265, + 228 + ], + "type": "image", + "image_path": "212ea898387ba8c5d3e5d7e2190e1dce3eb3444f467bcf3b2e6bd0372ff953b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "text", + "content": "Figure 4. Deferred style transformation. We apply the style transformation to the volume-rendered feature maps " + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "inline_equation", + "content": "\\bar{F}_c" + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "text", + "content": " according to the style feature maps " + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "text", + "content": ". To ensure multi-view consistency, we modulate the bias (e.g. the mean value of the style feature maps " + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "inline_equation", + "content": "\\mu(F_s)" + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "text", + "content": ") with the sum weight of sampled points along each ray " + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 236, + 288, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 421 + ], + "type": "text", + "content": "vanilla IN calculates per-dimension mean and standard-deviation of the batch of sampled points, which varies with different sampled points and incurs multi-view inconsistency accordingly. Thus we design volume-adaptive IN which, during training, keeps running estimates of the computed mean and standard-deviation, and uses them for normalization during evaluations (instead of computing from the sampled point batch). Through volume-adaptive IN, we can ensure that the content transformation is consistent regardless of the sampled point batch's holistic statistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 422, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 422, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 422, + 286, + 434 + ], + "type": "text", + "content": "Channel-wise self-attention can thus be implemented by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 446, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 446, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 97, + 446, + 287, + 460 + ], + "type": "interline_equation", + "content": "\\bar {F} _ {i} = V \\otimes \\operatorname {S o f t m a x} (\\widetilde {\\operatorname {c o v}} (Q, K)), \\tag {6}", + "image_path": "baf8202e84d3d5a2dcac5c2f238281a3674a61e9464db4e390cc70f46d3aa3bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "type": "text", + "content": " denotes matrix multiplication and " + }, + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\widetilde{cov}(Q, K) \\in \\mathbb{R}^{N \\times C' \\times C'}" + }, + { + "bbox": [ + 47, + 466, + 287, + 502 + ], + "type": "text", + "content": " denotes the covariance matrix in the channel dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 514, + 212, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 212, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 212, + 526 + ], + "type": "text", + "content": "3.2.2 Deferred Style Transformation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "content": "After applying SICT to the features of each 3D point, we apply DST to the volume-rendered 2D feature maps " + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\bar{F}_c" + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "content": " rather than 3D point features " + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\bar{F}_i" + }, + { + "bbox": [ + 46, + 533, + 287, + 593 + ], + "type": "text", + "content": ". To ensure multi-view consistency, we formulate the transformation as matrix multiplication and adaptive bias addition as illustrated in Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Specifically, we first extract feature maps " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " of the reference style " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " using a pre-trained VGG [51], and then generate the style transformation matrix " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{C' \\times C'}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " using feature covariance " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "cov(F_{s})" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " following [28]. Next, we apply matrix multiplication with " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " to the feature maps " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\bar{F}_{c}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " and use a " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " convolution layer conv without bias to restore the channel number from " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". Though these operations can partially instill style information, they are not expressive enough without bias addition containing style information [58]. Thus following [21], we multiply the feature" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "maps with the standard-deviation value " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\sigma(F_s)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and add the mean value " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mu(F_s)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". To ensure it is equivalent when applying the transformation to either 3D point features or 2D feature maps, we adaptively modulate the mean value " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mu(F_s)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " with the sum weight of sampled points along each ray " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{r}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". DST can be mathematically formulated by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 153, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 153, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 330, + 153, + 545, + 167 + ], + "type": "interline_equation", + "content": "F _ {c s} = \\operatorname {c o n v} \\left(T \\otimes \\bar {F} _ {c}\\right) \\times \\sigma \\left(F _ {s}\\right) + w _ {\\mathbf {r}} \\times \\mu \\left(F _ {s}\\right), \\tag {7}", + "image_path": "b80670e2360ab84f0a6033efaa1a04e277fe65f975e1e9c1d0af01498d81ed1a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 176, + 545, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 176, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 334, + 176, + 545, + 209 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad \\bar {F} _ {c} = \\sum_ {i = 1} ^ {N} w _ {i} \\bar {F} _ {i}, w _ {\\mathbf {r}} = \\sum_ {i = 1} ^ {N} w _ {i}, \\mathbf {r} \\in \\mathcal {R} \\tag {8}", + "image_path": "9e1ea763544799f283d8dac03fdb300b33f0888e1db54eb8c60d363f13d5d485.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "content": " denotes the weight of sampled point " + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "content": " (Eq. (2)), " + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\bar{F}_i" + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "content": " denotes the feature of sample " + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "content": " after SICT, and " + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 306, + 213, + 545, + 249 + ], + "type": "text", + "content": " is the set of rays in each training batch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "text", + "content": "Note " + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "inline_equation", + "content": "conv" + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "text", + "content": " convolution layer without bias, so it is basically a matrix multiplication operation. And " + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\sigma(S), \\mu(S)" + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "text", + "content": " are scalars. Together with the adaptive bias modulation " + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{r}}" + }, + { + "bbox": [ + 305, + 250, + 545, + 297 + ], + "type": "text", + "content": ", Eq. (7) can be reformulated by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 306, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 306, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 306, + 545, + 357 + ], + "type": "interline_equation", + "content": "F _ {c s} = \\sum_ {i = 1} ^ {N} w _ {i} \\left(\\underbrace {\\operatorname {c o n v} \\left(T \\otimes \\bar {F} _ {i}\\right) \\times \\sigma \\left(F _ {s}\\right) + \\mu \\left(F _ {s}\\right)} _ {\\text {(i)}}\\right), \\tag {9}", + "image_path": "4455c4948998e362bc02bba4365558448c9e1162ffa14f990956fc9a4c6e08af.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 364, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 436 + ], + "type": "text", + "content": "where part (i) can be seen as applying style transformation on every 3D point feature independently before volume rendering. This proves that applying DST on 2D feature maps is equivalent to applying the transformation on 3D points' features, maintaining multi-view consistency. The full derivation of Eq. (9) is provided in the appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "type": "text", + "content": "Finally, we adopt a 2D CNN decoder to project the stylized feature maps " + }, + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "type": "text", + "content": " to RGB space to generate the final stylized novel view images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 479, + 452, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 452, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 452, + 492 + ], + "type": "text", + "content": "3.3. Two-stage Model Training" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "content": "The training of our model is divided into the feature grid training stage and the stylization training stage, the former is trained with the target of novel view synthesis, and the latter is trained with the target of style transfer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Feature grid training stage (First stage). We first learn the feature grid 3D representation for the novel view synthesis task, in preparation for performing feature transformation for style transfer. We train the feature grid and the 2D CNN decoder simultaneously, with the supervision of both RGB images and their bilinearly up-sampled feature maps extracted from ReLU3_1 layer of pre-trained VGG [51]. By aligning the VGG features with the feature grid, the reconstructed features acquire semantic information. We use density grid pre-trained solely on RGB images since the supervising feature maps are not strictly multi-view consistent. The training objective is the mean square error (MSE) between the predicted and ground truth feature maps and RGB images. Following [19, 37], we use perceptual" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8342" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "loss [24] as additional supervision to increase reconstructed image quality. The overall loss function is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 105, + 287, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 105, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 105, + 287, + 167 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {g r i d} = \\sum_ {r \\in \\mathcal {R}} \\left\\| \\hat {F} (\\mathbf {r}) - F (\\mathbf {r}) \\right\\| _ {2} ^ {2} + \\left\\| \\hat {I} _ {\\mathcal {R}} - I _ {\\mathcal {R}} \\right\\| _ {2} ^ {2} \\\\ + \\sum_ {l \\in l _ {p}} \\left\\| \\mathcal {F} ^ {l} (\\hat {I} _ {\\mathcal {R}}) - \\mathcal {F} ^ {l} (I _ {\\mathcal {R}}) \\right\\| _ {2} ^ {2}, \\tag {10} \\\\ \\end{array}", + "image_path": "6604572925915cd13ec8e94e281a4be851b6665ed0d931908964ad8bb84b6414.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": " is the set of rays in each training batch, " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\hat{F} (\\mathbf{r}), F(\\mathbf{r})" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": " are the predicted and ground truth feature of ray " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathcal{R}}, I_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": " are the predicted and ground truth RGB image, " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": " denotes the set of VGG layers that compute perceptual loss, " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^l" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": " denotes the feature maps of the " + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 175, + 286, + 246 + ], + "type": "text", + "content": "th layer of pre-trained VGG network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": "Stylization training stage (Second stage). Our model learns to stylize novel views in the second stage. We freeze the feature grid, train the style transfer module, and finetune the CNN decoder. Thanks to the memory-efficient representation of 3D scenes and DST, unlike [8, 11, 48], our model can be trained directly on " + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": " patches, making patch sub-sampling algorithm [8, 11, 22, 48] unnecessary. We use the same loss as [21] where the content loss " + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": " is the MSE of the feature maps and the style loss " + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 247, + 286, + 366 + ], + "type": "text", + "content": " is the MSE of the channel-wise feature mean and standard-deviation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 375, + 286, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 375, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 113, + 375, + 286, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t y l i z a t i o n}} = \\mathcal {L} _ {c} + \\lambda \\mathcal {L} _ {s}, \\tag {11}", + "image_path": "5177f1afb8dc7e5db3a1b7e670ac77c2b3904e7d4495c21f18dc644ba393d28b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 395, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 286, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 395, + 286, + 418 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 395, + 286, + 418 + ], + "type": "text", + "content": " balances the content preservation and the stylization effect." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 429, + 127, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 127, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 127, + 442 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 509 + ], + "type": "text", + "content": "We evaluate StyleRF extensively with qualitative experiments in Sec. 4.1, quantitative experiments in Sec. 4.2 and ablation studies in Sec. 4.3. We demonstrate two applications of StyleRF in Sec. 4.4. The implementation details are provided in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 515, + 185, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 185, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 185, + 528 + ], + "type": "text", + "content": "4.1. Qualitative Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "content": "We evaluate StyleRF over two public datasets including LLFF [35] that contains real scenes with complex geometry structures and Synthetic NeRF [36] that contains " + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "content": " views of objects. In addition, we benchmark StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8] with their released codes. We perform comparisons on LLFF dataset [35]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "Fig. 5 shows qualitative comparisons. We can see that StyleRF achieves clearly better stylization with more precise geometry reconstruction. Specifically, StyleRF can generate high-definition stylization with realistic textures and patterns of style images. The superior stylization is largely attributed to our transformation design that allows working in the feature space with full-resolution feature maps. As illustrated in the highlight boxes, StyleRF" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 190 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 543, + 190 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 190 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 190 + ], + "type": "table", + "html": "
MethodShort-range ConsistencyLong-range Consistency
LPIPSRMSELPIPSRMSE
AdaIN [21]0.1520.1230.2200.186
CCPL [60]0.1100.1060.1910.174
ReReVST [57]0.0980.0800.1860.146
LSNV [19]0.0930.0920.1810.155
Hyper [8]0.0840.0680.1310.101
Ours0.0720.0820.1490.137
", + "image_path": "d8b7a4060c8e13d9e558d954735d13d099f165606fddb47c935219e927a760c0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "text", + "content": "Table 1. Results on consistency. We compare StyleRF with the state-of-the-art on consistency using LPIPS " + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "text", + "content": " and RMSE " + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 305, + 198, + 545, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 229, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 338 + ], + "type": "text", + "content": "can successfully restore the intricate geometry of complex scenes thanks to its radiance field representations. In addition, only StyleRF faithfully transfers the squareness texture in the second style image. Furthermore, StyleRF can robustly generalize to new styles in a zero-shot manner and can adapt well to " + }, + { + "bbox": [ + 304, + 229, + 545, + 338 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 229, + 545, + 338 + ], + "type": "text", + "content": " dataset as illustrated in Fig. 1. As a comparison, LSNV [19] fails to capture fine-level geometry like the bones of the T-Rex and the petals of the flower while Hyper [8] produces very blurry stylization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 345, + 424, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 424, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 424, + 357 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 364, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 555 + ], + "type": "text", + "content": "3D style transfer is a very new and under-explored task and there are few metrics for quantitative evaluation of stylization quality. Hence, we manage to evaluate the multiview consistency only. In our experiments, we warp one view to the other according to the optical flow [53] using softmax splatting [42], and then computed the masked RMSE score and LPIPS score [64] to measure the stylization consistency. Following [8, 11, 19], we compute the short-range and long-range consistency scores which compare adjacent views and far-away views respectively. We compare StyleRF against two state-of-the-art zero-shot 3D style transfer methods Hyper [8] and LSNV [19], one SOTA single-frame-based video style transfer method CCPL [60], one SOTA multi-frames-based video style transfer method ReReVST [57], and one classical image style transfer method AdaIN [21]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 555, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 663 + ], + "type": "text", + "content": "It can be seen from Tab. 1 that StyleRF significantly outperforms image style transfer approach [21] and video style transfer approach [57, 60] which capture little information about the underlying 3D geometry. In addition, StyleRF achieves better consistency than point-cloud-based 3D style transfer [19] as well. Note Hyper [8] achieves slightly better LPIPS and RMSE scores than our method, largely because it produces over-smooth results and inadequate stylization as shown in Fig. 5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 671, + 406, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 406, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 406, + 682 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We design two innovative techniques to improve the stylization quality and maintain the multi-view consistency." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8343" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 553, + 521 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 553, + 521 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 553, + 521 + ], + "type": "image", + "image_path": "ee025dc9e4e1f6cfb462963d8fc32858f3e5c559245d7aa4c7010c15c28ec177.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 527, + 546, + 550 + ], + "lines": [ + { + "bbox": [ + 46, + 527, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 546, + 550 + ], + "type": "text", + "content": "Figure 5. Comparison of StyleRF with two state-of-the-art zero-shot 3D style transfer methods LSNV [19] and Hyper [8]. For each of the two sample Scenes and reference Styles, StyleRF produces clearly better 3D style transfer and depth estimation. Check zoom-in for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 559, + 288, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 656 + ], + "type": "text", + "content": "The first is volume-adaptive instance normalization which uses the learned mean and variance of the whole volume during inference, eliminating the dependency on holistic statistics of the sampled point batch. The second is the adaptive bias addition in DST, which improves the stylization quality using bias capturing style information. We evaluate the two designs to examine how they contribute to the overall stylization of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Volume-adaptive instance normalization. We compare our volume-adaptive instance normalization (IN) with vanilla IN and StyleRF without IN. As Fig. 6 (c) shows, vanilla IN produces severe block-shape artifacts as the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 559, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 547, + 668 + ], + "type": "text", + "content": "transformation of each batch is conditioned on the holistic statistics of itself, thus each batch (i.e. block in the image) produces inconsistent stylization which leads to the artifacts. However, if we discard IN as shown in Fig. 6 (d), the multi-view consistency can maintain but the stylization quality compromises a lot, failing to capture the correct color tone of the reference style image. This is because IN removes the original style information of the content image which facilitates the transfer of the reference style [21]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Adaptive bias addition. As illustrated in Fig. 6 (b), the stylization quality degrades a lot if we eliminate the adaptive bias addition in DST (Sec. 3.2.2), producing unnatural" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8344" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 127, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 127, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 127, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 127, + 129 + ], + "type": "image", + "image_path": "9d9688b3ec0995eb5feab5b0bea28bc1f57424825c892392b6475b2fc78e4535.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 130, + 127, + 138 + ], + "lines": [ + { + "bbox": [ + 51, + 130, + 127, + 138 + ], + "spans": [ + { + "bbox": [ + 51, + 130, + 127, + 138 + ], + "type": "text", + "content": "Ground Truth Novel View" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 140, + 112, + 198 + ], + "blocks": [ + { + "bbox": [ + 66, + 140, + 112, + 198 + ], + "lines": [ + { + "bbox": [ + 66, + 140, + 112, + 198 + ], + "spans": [ + { + "bbox": [ + 66, + 140, + 112, + 198 + ], + "type": "image", + "image_path": "4988ff3367a052dbcde5bf5cccf0a0318c2a12081b6610131b78124fcea707e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 200, + 123, + 208 + ], + "lines": [ + { + "bbox": [ + 56, + 200, + 123, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 123, + 208 + ], + "type": "text", + "content": "Reference Style Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 72, + 208, + 129 + ], + "blocks": [ + { + "bbox": [ + 132, + 72, + 208, + 129 + ], + "lines": [ + { + "bbox": [ + 132, + 72, + 208, + 129 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 208, + 129 + ], + "type": "image", + "image_path": "57d8076e5926cfcec12eab018620e7ca1e2cf67ecd60f60f9c25de28c0da1665.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 130, + 182, + 139 + ], + "lines": [ + { + "bbox": [ + 158, + 130, + 182, + 139 + ], + "spans": [ + { + "bbox": [ + 158, + 130, + 182, + 139 + ], + "type": "text", + "content": "(a) Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 208, + 72, + 285, + 129 + ], + "blocks": [ + { + "bbox": [ + 208, + 72, + 285, + 129 + ], + "lines": [ + { + "bbox": [ + 208, + 72, + 285, + 129 + ], + "spans": [ + { + "bbox": [ + 208, + 72, + 285, + 129 + ], + "type": "image", + "image_path": "74077c824818391c72be93b2ff2578f227fa1f824657e1c5339ceef321b5f7ca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 130, + 277, + 139 + ], + "lines": [ + { + "bbox": [ + 214, + 130, + 277, + 139 + ], + "spans": [ + { + "bbox": [ + 214, + 130, + 277, + 139 + ], + "type": "text", + "content": "(b) w/o Adaptive Bias" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 132, + 140, + 207, + 198 + ], + "blocks": [ + { + "bbox": [ + 132, + 140, + 207, + 198 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 207, + 198 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 207, + 198 + ], + "type": "image", + "image_path": "46d4595847b45fd388cf1584ad246981845658df3319c3c0c431849ae73133ab.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 199, + 189, + 208 + ], + "lines": [ + { + "bbox": [ + 151, + 199, + 189, + 208 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 189, + 208 + ], + "type": "text", + "content": "(c) Vanilla IN" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 208, + 140, + 284, + 198 + ], + "blocks": [ + { + "bbox": [ + 208, + 140, + 284, + 198 + ], + "lines": [ + { + "bbox": [ + 208, + 140, + 284, + 198 + ], + "spans": [ + { + "bbox": [ + 208, + 140, + 284, + 198 + ], + "type": "image", + "image_path": "1f2cf7efd36889bc1754a0c64a25a05e0d400c4507f2a6a5ab163267feec3767.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 199, + 261, + 208 + ], + "lines": [ + { + "bbox": [ + 230, + 199, + 261, + 208 + ], + "spans": [ + { + "bbox": [ + 230, + 199, + 261, + 208 + ], + "type": "text", + "content": "(d) w/o IN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 277, + 286, + 414 + ], + "blocks": [ + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 273 + ], + "type": "text", + "content": "Figure 6. Ablation studies. (a) shows the stylization of our full pipeline. (b) shows the stylization without the adaptive bias. (c) shows the stylization when replacing the volume-adaptive instance normalization (IN) with vanilla IN. (d) shows the stylization without any IN." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 277, + 286, + 414 + ], + "lines": [ + { + "bbox": [ + 50, + 277, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 50, + 277, + 286, + 414 + ], + "type": "image", + "image_path": "93c09f22dce827bbd0c6e32cb7aab03a59b216b7c5f37bde10227a14000938ee.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 424, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 456 + ], + "type": "text", + "content": "Figure 7. Multi-style interpolation. StyleRF can smoothly interpolate between arbitrary styles by interpolating features of the scene." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 540 + ], + "type": "text", + "content": "stylization compared to the stylization of our full pipeline in Fig. 6 (a). This is because bias usually contains crucial style information such as the overall color tone [58]. StyleRF employs bias addition that is adaptively modulated by the weight of each ray, improving the stylization quality and keeping multi-view consistency concurrently." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 550, + 129, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 129, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 129, + 562 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 604 + ], + "type": "text", + "content": "StyleRF can be easily extended along different directions with different applications. We provide two possible extensions in the ensuing subsections." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Multi-style interpolation. StyleRF can smoothly interpolate different styles thanks to its high-level feature representation of a 3D scene. As illustrated in Fig. 7, we linearly interpolate the feature maps of a specific view by using four different styles at four corners. Unlike previous NeRF-based 3D style transfer that supports style interpolation by interpolating one-hot latent vectors [11], StyleRF can interpolate arbitrary numbers of unseen new styles by interpolating features of the scene, yielding more smooth" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 79, + 386, + 193 + ], + "blocks": [ + { + "bbox": [ + 311, + 70, + 386, + 78 + ], + "lines": [ + { + "bbox": [ + 311, + 70, + 386, + 78 + ], + "spans": [ + { + "bbox": [ + 311, + 70, + 386, + 78 + ], + "type": "text", + "content": "Ground Truth Novel View" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 79, + 386, + 193 + ], + "lines": [ + { + "bbox": [ + 310, + 79, + 386, + 193 + ], + "spans": [ + { + "bbox": [ + 310, + 79, + 386, + 193 + ], + "type": "image", + "image_path": "811621453ecb598dee2ad002244040ce4d61eda8bbd4b6d8d0022da849410e12.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 194, + 374, + 202 + ], + "lines": [ + { + "bbox": [ + 323, + 194, + 374, + 202 + ], + "spans": [ + { + "bbox": [ + 323, + 194, + 374, + 202 + ], + "type": "text", + "content": "Generated Mask" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 391, + 79, + 464, + 193 + ], + "blocks": [ + { + "bbox": [ + 391, + 79, + 464, + 193 + ], + "lines": [ + { + "bbox": [ + 391, + 79, + 464, + 193 + ], + "spans": [ + { + "bbox": [ + 391, + 79, + 464, + 193 + ], + "type": "image", + "image_path": "d70f897c0e5096e80ad8d2cfc9fa20eb8ff7345dd54cff87a100fe05ec6564e3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 194, + 496, + 203 + ], + "lines": [ + { + "bbox": [ + 435, + 194, + 496, + 203 + ], + "spans": [ + { + "bbox": [ + 435, + 194, + 496, + 203 + ], + "type": "text", + "content": "Spatial Composition" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 212, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 245 + ], + "type": "text", + "content": "Figure 8. Compositional 3D style transfer. Given the 3D-consistent segmentation masks, StyleRF can create infinite combinations of styles by spatial composition." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 466, + 79, + 542, + 193 + ], + "blocks": [ + { + "bbox": [ + 466, + 79, + 542, + 193 + ], + "lines": [ + { + "bbox": [ + 466, + 79, + 542, + 193 + ], + "spans": [ + { + "bbox": [ + 466, + 79, + 542, + 193 + ], + "type": "image", + "image_path": "ea120450f43283200e27d9f472aefb4f871ab7655d0c5eb3b3bf741242516c3d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 255, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 292 + ], + "type": "text", + "content": "and harmonious interpolation. Hence, StyleRF can not only transfer arbitrary styles in a zero-shot manner but also generate non-existent stylization via multi-style interpolation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 293, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 436 + ], + "type": "text", + "content": "Compositional 3D style transfer. Thanks to its precise geometry reconstruction, StyleRF can be seamlessly integrated with NeRF-based object segmentation [12, 26, 65] for compositional 3D style transfer. As shown in Fig. 8, we apply 3D-consistent segmentation masks to the feature maps and apply different styles to stylize the contents inside and outside the masks separately. We can see that the edges of the masks can be blended more softly by applying the segmentation masks to the feature maps instead of RGB images. Due to its zero-shot nature, StyleRF can create infinite combinations of styles without additional training, producing numerous artistic creations and inspirations." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 378, + 460 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 469, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 649 + ], + "type": "text", + "content": "In this paper, we present StyleRF, a novel zero-shot 3D style transfer method that balances the three-way dilemma over accurate geometry reconstruction, high-quality stylization, and being generalizable to arbitrary new styles. By representing the 3D scene with an explicit grid of high-level features, we can faithfully restore high-fidelity geometry through volume rendering. Then we perform style transfer on the feature space of the scene, leading to high-quality zero-shot stylization results. We innovatively design sampling-invariant content transformation to maintain multiview consistency and deferred style transformation to increase efficiency. We demonstrate that StyleRF achieves superior 3D stylization quality than previous zero-shot 3D style transfer methods, and can be extended to various interesting applications for artistic 3D creations." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 658, + 395, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 395, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 395, + 671 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "This project is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with project number RT18/22." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8345" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 201 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 270 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[5] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 395 + ], + "type": "text", + "content": "[6] Yaosen Chen, Qi Yuan, Zhiqiang Li, Yuegen Liu, Wei Wang, Chaoping Xie, Xuming Wen, and Qien Yu. Upst-nerf: Universal photorealistic style transfer of neural radiance fields for 3d scene. In *arxiv*, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 441 + ], + "type": "text", + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 497 + ], + "type": "text", + "content": "[8] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1475-1484, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 552 + ], + "type": "text", + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM international conference on multimedia, pages 2719-2727, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[10] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14304-14313, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "text", + "content": "[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. arXiv preprint arXiv:2209.08776, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 196 + ], + "type": "text", + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "type": "text", + "content": "[17] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "text", + "content": "[18] Haozhi Huang, Hao Wang, Wenhan Luo, Lin Ma, Wenhao Jiang, Xiaolong Zhu, Zhifeng Li, and Wei Liu. Real-time neural style transfer for videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 783-791, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 409 + ], + "type": "text", + "content": "[19] Hsin-Ping Huang, Hung-Yu Tseng, Saurabh Saini, Maneesh Singh, and Ming-Hsuan Yang. Learning to stylize novel views. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13869-13878, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[20] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "text", + "content": "[21] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "type": "text", + "content": "[22] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[23] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8346" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[26] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. arXiv preprint arXiv:2205.15585, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[27] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 194 + ], + "type": "text", + "content": "[28] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2019. 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "text", + "content": "[29] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[30] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "type": "text", + "content": "[31] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 403 + ], + "type": "text", + "content": "[32] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 403, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 287, + 449 + ], + "type": "text", + "content": "[33] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 504 + ], + "type": "text", + "content": "[34] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 558 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 669 + ], + "type": "text", + "content": "[37] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[39] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 546, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 546, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 546, + 160 + ], + "type": "text", + "content": "[40] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 161, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 546, + 217 + ], + "type": "text", + "content": "[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "text", + "content": "[42] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5437-5446, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 259, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 259, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 546, + 304 + ], + "type": "text", + "content": "[43] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 304, + 546, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 350 + ], + "type": "text", + "content": "[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 546, + 403 + ], + "type": "text", + "content": "[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 403, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 546, + 459 + ], + "type": "text", + "content": "[46] Charles R Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J Guibas. Volumetric and multi-view cnns for object classification on 3d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 459, + 546, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 546, + 502 + ], + "type": "text", + "content": "[47] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. International Journal of Computer Vision, 126(11):1199-1219, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 502, + 546, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 546, + 548 + ], + "type": "text", + "content": "[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 546, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 546, + 581 + ], + "type": "text", + "content": "[49] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 581, + 546, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 546, + 625 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 546, + 625 + ], + "type": "text", + "content": "[50] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8242-8250, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 546, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 546, + 658 + ], + "type": "text", + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 712 + ], + "type": "text", + "content": "[52] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8347" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 677 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[53] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[54] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 194 + ], + "type": "text", + "content": "[55] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 250 + ], + "type": "text", + "content": "[56] Wenjing Wang, Jizheng Xu, Li Zhang, Yue Wang, and Jiaying Liu. Consistent video style transfer via compound regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12233-12240, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 252, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 286, + 285 + ], + "type": "text", + "content": "[57] Wenjing Wang, Shuai Yang, Jizheng Xu, and Jiaying Liu. Consistent video style transfer via relaxation and regularization. IEEE Trans. Image Process., 2020. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 341 + ], + "type": "text", + "content": "[58] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 3, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 342, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 286, + 396 + ], + "type": "text", + "content": "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 398, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 398, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 398, + 286, + 442 + ], + "type": "text", + "content": "[60] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 443, + 286, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 443, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 49, + 443, + 286, + 496 + ], + "type": "text", + "content": "[61] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 498, + 286, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 286, + 541 + ], + "type": "text", + "content": "[62] Fangneng Zhan, Lingjie Liu, Adam Kortylewski, and Christian Theobalt. General neural gauge fields. In *The Eleventh International Conference on Learning Representations*, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 544, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 286, + 586 + ], + "type": "text", + "content": "[63] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 588, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 588, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 588, + 286, + 643 + ], + "type": "text", + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 644, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 644, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 644, + 286, + 677 + ], + "type": "text", + "content": "[65] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, 2021. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8348" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_content_list.json b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9dafe8cf4eecf58b70c3deadc56774d0f50c4514 --- /dev/null +++ b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_content_list.json @@ -0,0 +1,1440 @@ +[ + { + "type": "text", + "text": "StyleRes: Transforming the Residuals for Real Image Editing with StyleGAN", + "text_level": 1, + "bbox": [ + 96, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hamza Pehlivan Yusuf Dalva Aysegul Dundar Bilkent University", + "bbox": [ + 261, + 180, + 707, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hamza.pehlivan,yusuf.dalva}@bilkent.edu.tr", + "bbox": [ + 295, + 219, + 673, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adundar@cs.bilkent.edu.tr", + "bbox": [ + 372, + 237, + 596, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. Inverting real images into StyleGAN's latent space is an extensively studied problem, yet the trade-off between the image reconstruction fidelity and image editing quality remains an open challenge. The low-rate latent spaces are limited in their expressiveness power for high-fidelity reconstruction. On the other hand, high-rate latent spaces result in degradation in editing quality. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We train the framework to extract residual features and transform them via a novel architecture pipeline and cycle consistency losses. We run extensive experiments and compare our method with state-of-the-art inversion methods. Qualitative metrics and visual comparisons show significant improvements. Code: https://github.com/hamzapehlivan/StyleRes", + "bbox": [ + 75, + 316, + 473, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Adversarial Networks (GANs) achieve high quality synthesis of various objects that are hard to distinguish from real images [14,21,22,41,43]. These networks also have an important property that they organize their latent space in a semantically meaningful way; as such, via latent editing, one can manipulate an attribute of a generated image. This property makes GANs a promising technology for image attribute editing and not only for generated images but also for real images. However, for real images, one also needs to find the corresponding latent code that will generate the particular real image. For this purpose, different GAN inversion methods are proposed, aiming to project real images to pretrained GAN latent space [15,30,31,33,37]. Even though this is an extensively stud", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6a110052faad7df5ba714c0095b57d515f516b8c8916a12ecbf6a57ea694544.jpg", + "image_caption": [ + "Figure 1. Comparison of our method with e4e, HFGI, and HyperStyle for the pose, bob cut hairstyle, smile removal, and color change edits. Our method achieves high fidelity to the input and high quality edits." + ], + "image_footnote": [], + "bbox": [ + 506, + 295, + 906, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ied problem with significant progress, the trade-off between image reconstruction fidelity and image editing quality remains an open challenge.", + "bbox": [ + 496, + 750, + 892, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The trade-off between image reconstruction fidelity and image editing quality is referred to as the distortion-editability trade-off [32]. Both are essential for real image editing. However, it is shown that the low-rate latent spaces are limited in their expressiveness power, and not every image can be inverted with high fidelity reconstruction [1, 27, 32, 34]. For that reason, higher bit encodings", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1828", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and more expressive style spaces are explored for image inversion [1, 2]. Although with these techniques, images can be reconstructed with better fidelity, the editing quality decreases since there is no guarantee that projected codes will naturally lie in the generator's latent manifold.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a framework that achieves high fidelity input reconstruction and significantly improved editability compared to the state-of-the-art. We learn residual features in higher-rate latent codes that are missing in the reconstruction of encoded features. This enables us to reconstruct image details and background information which are difficult to reconstruct via low rate latent encodings. Our architecture is single stage and learns the residuals based on the encoded features from the encoder and generated features from the pretrained GANs. We also learn a module to transform the higher-latent codes if needed based on the generated features (e.g. when the low-rate latent codes are manipulated). This way, when low-rate latent codes are edited for attribute manipulation, the decoded features can adapt to the edits to reconstruct details. While the attributes are not edited, the encoder can be trained with image reconstruction and adversarial losses. On the other hand, when the image is edited, we cannot use image reconstruction loss to regularize the network to preserve the details. To guide the network to learn correct transformations based on the generated features, we train the model with adversarial loss and cycle consistency constraint; that is, after we edit the latent code and generate an image, we reverse the edit and aim at reconstructing the original image. Since we do not want our method to be limited to predefined edits, during training, we simulate edits by randomly interpolating them with sampled latent codes.", + "bbox": [ + 75, + 167, + 472, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The closest to our approach is HFGI [34], which also learns higher-rate encodings. Our framework is different as we learn a single-stage architecture designed to learn features that are missing from low-rate encodings and we learn how to transform them based on edits. As shown in Fig. 1, our framework achieves significantly better results than HFGI and other methods in editing quality. In summary, our main contributions are:", + "bbox": [ + 75, + 575, + 468, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a single-stage framework that achieves high-fidelity input embedding and editing. Our framework achieves that with a novel encoder architecture.", + "- We propose to guide the image projection with cycle consistency and adversarial losses. We edit encoded images by taking a step toward a randomly sampled latent code. We expect to reconstruct the original image when the edit is reversed. This way, edited images preserve the details of the input image, and edits become high quality.", + "- We conduct extensive experiments to show the effectiveness of our framework and achieve significant improve" + ], + "bbox": [ + 76, + 708, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ments over state-of-the-art for both reconstruction and real image attribute manipulations.", + "bbox": [ + 511, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 135, + 640, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GAN Inversion. GAN inversion methods aim at projecting a real image into GANs embedding so that from the embedding, GAN can generate the given image. Currently, the biggest motivation of the inversion is the ability to edit the image via semantically rich disentangled features of GANs; therefore models aim for high reconstruction and editing quality. Inversion methods can be categorized into two; 1) methods that directly optimize the latent vector to minimize the reconstruction error between the output and target image [1,2,9,22,28], 2) methods that learn encoders to reconstruct images over training images [32,34,45]. Optimization based methods require per-image optimization and iterative passes on GANs, which take several minutes per image. Additionally, overfitting on a single image results in latent codes that do not lie in GAN's natural latent distribution, leading to poor editing quality. Among these methods, PTI [28] takes a different approach, and instead of searching for a latent code that will reconstruct the image most accurately, PTI fine-tunes the generator in order to insert encoded latent code into well-behaved regions of the latent space. It shows better editability; however, the method still suffers from long run-time optimizations and tunings. Encoder based methods leverage the knowledge of training sets while projecting images. They output results with less fidelity to the input, but edits on the projected latents are better quality. They also operate mostly in real-time. They either project the latent code with a single feed-forward pass [27,32,45], or two stage feed-forward passes where the first encoder learns to embed an image, the second learns to reconstruct the missing details [34], or iterative forward passes where the network tries to minimize the reconstruction loss at each pass by taking the original image and reconstructed output as inputs [4,5]. In this work, we propose a novel encoder architecture that achieves significantly better results than state-of-the-art via a single feed-forward pass on the input image.", + "bbox": [ + 496, + 160, + 893, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Translation. There is quite an interest in image translation algorithms that can change an attribute of an image while preserving a given content [12, 25, 35], especially for faces editing [3, 8, 10, 13, 17, 24, 29, 36, 38, 42]. These algorithms set an encoder-decoder architecture and train the models with reconstruction and GAN losses [8, 19, 24, 39, 46]. Even though they are successful at image manipulation, they are trained for a given task and rely on predefined translation datasets. On the other hand, it is shown that GANs that are trained to synthesize objects in an unconditional way can be used for attribute manipulation [6, 21, 22] via their semantically rich disentangled features. This makes pretrained GANs promising technol-", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1829", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d12b3a79f2e1d372a601a2a9b5437bbdf542e752b780479378e718f2567b974.jpg", + "image_caption": [ + "(a1) Input" + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 187, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e90e53b430e865f52e850830a6770d5c91149fb4a94c272f2a2f3fba976c720f.jpg", + "image_caption": [ + "(b1) Low-rate Inversion (e4e)" + ], + "image_footnote": [], + "bbox": [ + 191, + 88, + 279, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64d5da3d2c3d32ae117b8b30ff23edfbacb02d053761e744b20d9ab7ff6ba50d.jpg", + "image_caption": [ + "(c1) Low-rate Edit (e4e)" + ], + "image_footnote": [], + "bbox": [ + 281, + 88, + 370, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f7508805408e03985480a1ccdd29648729d411517e28c06bc28396b48289a1b.jpg", + "image_caption": [ + "(d1) Low-rate Inversion High-rate Inversion" + ], + "image_footnote": [], + "bbox": [ + 372, + 88, + 460, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75eb5812a2381097f15a8345007b2d4bb12e17377e6e23ff6a6babff13e0e523.jpg", + "image_caption": [ + "(1) Low-rate Edit High-rate Inversion" + ], + "image_footnote": [], + "bbox": [ + 99, + 172, + 279, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0415f2b0cfb09985f92a6cbe2afabbd9542faef8ab6a3e374623e051ead97ce5.jpg", + "image_caption": [ + "(1) Low-rate Edit High-rate Transformation" + ], + "image_footnote": [], + "bbox": [ + 279, + 172, + 370, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49ef69017312bf2abdbf04d01381c1c2cb8cc7c74b4513e5631e1a03b726315f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 204, + 464, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85685a9346cdd43ea4940b70bc302c9da019708b51777694bf8db148a0e12af1.jpg", + "image_caption": [ + "(a2) Input" + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 594, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3ba937bcb007e3cbd1586cb384e5ac6bde53fb5cba5df00935f2d6bbc62743aa.jpg", + "image_caption": [ + "(b3) Low-rate Inversion (e4e)" + ], + "image_footnote": [], + "bbox": [ + 596, + 89, + 684, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14fea19cdefd9de980c839b807c52745dc3fcaabd3b9d6dfd63903c8239fa957.jpg", + "image_caption": [ + "(c2) Low-rate Edit (e4e)" + ], + "image_footnote": [], + "bbox": [ + 686, + 89, + 774, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d2eb7cd88e13b3cd2610f5cca8db0676538c2d3e341dcdae278c5738662a3562.jpg", + "image_caption": [ + "(d2) Low-rate Inversion High-rate Inversion" + ], + "image_footnote": [], + "bbox": [ + 777, + 88, + 867, + 157 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e95e929c5586ac277dae561c506543f2693c56bb4d7c89b24602bad2721e8140.jpg", + "image_caption": [ + "Figure 2. When images are encoded to $\\mathrm{W + }$ space, as shown in (b), the reconstructions miss many image details. However, the edits are in good quality (c). Additional features can be learned in higher-rate latent codes. For example, skip connections from encoder to generator in higher resolution features can enable high-fidelity reconstruction to input images as shown in (d). However, if the high-rate features do not transform with the edits, they result in ghosting effects as shown in (e). In this work, we propose high-rate encoding and transformation for successful inversions and edits (f)." + ], + "image_footnote": [], + "bbox": [ + 527, + 172, + 583, + 243 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd886fff90b9f737e70ab6a673353f738bf2bfd857bf94ecd707d80dddff59ac.jpg", + "image_caption": [ + "62)Low-rate Edit High-rate Inversion", + "(f2) Low-rate Edit High-rate Transformation" + ], + "image_footnote": [], + "bbox": [ + 596, + 172, + 684, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7df925b8e6d83d187cd5a93187f4c93968d11d74ff1df7d42317835a3eb69acb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 172, + 776, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f424ffd6f3e78a0d3f6725991688d26ff06ef70235ad99bf01e59f965cef20e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 175, + 844, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ogy for image editing, given that via their ability of image generation, they can achieve many downstream tasks simultaneously. Many methods have been proposed for finding latent directions to edit images [15, 30, 31, 33, 37]. These directions are explored both in supervised [3, 30] and unsupervised ways [15, 31, 33, 37] resulting in exploration of attribute manipulations beyond the predefined attributes of labeled datasets. However, finding these directions is only one part of the problem; to achieve real image editing, there is a need for a successful image inversion method which is the topic of this paper.", + "bbox": [ + 75, + 367, + 472, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 549, + 168, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Section 3.1, we describe the motivation for our approach. The model architecture and the training procedure are presented in Sections 3.2 and 3.3, respectively.", + "bbox": [ + 75, + 575, + 468, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 633, + 197, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current inversion methods aim at learning an encoder to project an input image into StyleGAN's natural latent space so that the inverted image is editable. However, it is observed by previous works that when images are encoded into GAN's natural space (low-rate W or W+ space), their reconstructions suffer from low fidelity to input images, as shown in Fig. 2(b). On the other hand, if the image is encoded to higher feature maps of pretrained GANs, they will not be editable since they are not inverted to the semantic space of GANs. On the other hand, low-rate inversion is editable, as shown in Fig. 2(c).", + "bbox": [ + 75, + 657, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN, during training and inference, relies on direct noise inputs to generate stochastic details. These additional noise inputs provide stochastic image details such as for face images, the exact placement of hairs, stubble, freckles, or skin pores which are not modeled by $\\mathrm{W}+$ space.", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inversion to $\\mathrm{W + }$ space alone, that mechanism is discarded. One can tune the noise maps as well via the reconstruction loss; however, then there will be no mechanism to adapt those stochastic details to attribute manipulation. For example, the freckles should move as the pose is manipulated, but with noise optimization alone, it will not be possible.", + "bbox": [ + 496, + 367, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose to embed images to both low-rate and high-rate embeddings. Consistent with the design of StyleGAN, we aim at encoding the overall composition and high-level aspects of the image to $\\mathrm{W + }$ space (low-rate encoding). Our goal is to embed the stochastic image details and the diverse background details, which are difficult to reconstruct from $\\mathrm{W + }$ space to higher latent codes. However, this setting also requires a mechanism for encoded image details to adopt manipulations in the image. Otherwise, it will cause a ghosting effect, as shown in Fig 2(e). For example, in Fig 2(e1), the smile is removed by $\\mathrm{W + }$ code edit; however, the higher-rate encodings stay the same and cause artifacts around the mouth area. In Fig. 2(e2), the pose is edited but higher detail encodings did not move and caused blur.", + "bbox": [ + 496, + 473, + 892, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we design an encoder architecture and training pipeline to achieve both learning residual features (the ones $\\mathrm{W}+$ space could not reconstruct) and how to transform them to be consistent with the manipulations.", + "bbox": [ + 496, + 685, + 890, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. StyleRes Architecture", + "text_level": 1, + "bbox": [ + 500, + 756, + 702, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method utilizes an encoder $E_0$ that can embed images into $W+$ latent space and a StyleGAN generator $G$ . In our setup, we utilize a pretrained encoder for $E_0$ [32] and StyleGAN2 generator [22] and fix them in our training. Because it is difficult to preserve image details only from low-rate latent codes, we also extract high level features from the encoder. First, the high and low rate features are extracted as $F_0$ , $W^+ = E_0(x)$ using the encoder $E_0$", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1830", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3de877285e98036d1c3895e5a4e5cdf66bebd1f3469d5800068931c14ee4be6.jpg", + "image_caption": [ + "Figure 3. StyleRes encodes missing features for high-fidelity reconstruction of given input via the first encoder, $E_{1}$ . Those encoded features are the ones which could not be encoded to low-rate W+ space via $E_{0}$ due to the information bottleneck. Through the second encoder, $E_{2}$ , StyleRes learns to transform features based on the manipulated features. During training, latent codes are edited by interpolating encoded W+’s with randomly generated ones by StyleGAN's mapping network. During inference, they are edited with semantically meaningful directions discovered by methods such as InterfaceGAN and GANSpace. Note that StyleGAN generator is shown as two parts just for the ease of visualizing the diagram. First part includes the layers that generate features to $64 \\times 64$ and the second part generates the higher resolution features and final image." + ], + "image_footnote": [], + "bbox": [ + 76, + 92, + 893, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the input image $x$ as shown in Fig. 3. $F_{0}$ has the spatial dimension of $64 \\times 64$ and provides us with more image details. Next, from $F_{0}$ , our aim is to encode residuals that are missing from the image reconstruction. For this goal, we set an encoder, $E_{1}$ , which takes $F_{0}$ as input. Since the goal of $E_{1}$ is to encode residual features, we also feed the generated features with $W^{+}$ from the StyleGAN generator, $G_{W} = G_{0\\rightarrow n}(W)$ , where the arrow operator indicates the indices of convolutional layers used from $G$ .", + "bbox": [ + 75, + 457, + 472, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {a} = E _ {1} \\left(F _ {0}, G _ {W +}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 611, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the inputs of $F_{0}$ and $G_{W+}$ , $E_{1}$ can learn what the missing features are by comparing the encoded features $F_{0}$ and generated ones, $G_{W+}$ , so far. While $E_{1}$ can learn the residual features, it is not guided on how to transform them if images are edited. For that purpose, we train $E_{2}$ , which takes $F_{a}$ and edited features from the generator. Since we do not target predefined edits (e.g., smile, pose, age), we simulate the edits by taking random directions. Specifically, we sample a $z$ vector from the normal distribution and obtain $W_{r}^{+}$ by StyleGAN's mapping network, $M$ ; $W_{r}^{+} = M(z)$ . Next, we take a step towards $W_{r}^{+}$ to obtain a mixed style code $W_{\\alpha}^{+}$ .", + "bbox": [ + 75, + 637, + 472, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\alpha} ^ {+} = W ^ {+} + \\alpha \\frac {W _ {r} ^ {+} - W ^ {+}}{1 0} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 830, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ controls the degree of the edit. During the training, we assign $\\alpha$ to 0 (no edit) or a value in the range of", + "bbox": [ + 76, + 869, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(4,5) with $50\\%$ chance. We adjust $F_{a}$ to the altered generator features $G_{\\alpha} = G_{0 \\to n}(W_{\\alpha})$ using a second encoder $E_{2}$ and obtain the final feature map F:", + "bbox": [ + 498, + 457, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = E _ {2} \\left(F _ {a}, G _ {\\alpha}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 517, + 890, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, $F$ and $G_{\\alpha}$ are summed, and given as an input to next convolutional layers in the generator. Architectural details of $E_{1}$ and $E_{2}$ are given in Supplementary.", + "bbox": [ + 496, + 537, + 893, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Phases", + "text_level": 1, + "bbox": [ + 500, + 589, + 661, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train our model with the capabilities of high-fidelity inversion and high-quality editing, we use two training phases, as shown in Fig. 4.", + "bbox": [ + 496, + 613, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "No Editing Path. In this path, we reconstruct images with no editing on encoded $W^{+}$ . This refers to the case where $\\alpha = 0$ in Eq. 2. This is also the training path other inversion methods solely rely on. While this path can teach the network to reconstruct images faithfully, it does not include any edits and cannot guide the network to high-quality editability.", + "bbox": [ + 496, + 659, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cycle Translation Path. In this path, we edit images by setting $\\alpha$ a value in the range of (4, 5), which we found to work best in our ablation studies. Via the edit, the generator outputs image $x_{i}^{\\prime}$ . Next, we feed this intermediate output image to the encoder and reverse the edit by inverting the addition presented in Eq. 2. The generator reconstructs $x''$ , which is supposed to match the input image $x$ . The cycle translation path is important because there is no ground-truth output for the edited image, $x_{i}^{\\prime}$ . Adversarial loss can", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1831", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5332370ca3472db546b2e80c75b86e8a03b5a141fbbce49fe66d376fa4a163d1.jpg", + "image_caption": [ + "(a) No Editing: Reconstruction Loss" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 356, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/edc70d8352bb6ec52392320332ed22397dca7b3fa788a2e951500470de8545a1.jpg", + "image_caption": [ + "(b) Cycle-consistency: Reconstruction Loss", + "Figure 4. We train StyleRes with (a) no editing based reconstruction and (b) cycle consistency based reconstruction losses. Many details from the inference pipeline are omitted for brevity. With reconstruction losses, the model learns to preserve the image details. We additionally apply adversarial losses on $x_{i}^{\\prime}$ ; this way, when the image is edited, the transformation network learn to output realistic images. With cycle consistency based reconstruction, the network is regularized to keep the input details also during edits. We use additional losses as explained in Section 3.4." + ], + "image_footnote": [], + "bbox": [ + 398, + 90, + 890, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "guide $x_{i}^{\\prime}$ to look realistic but will not guide the network to keep the input image details if edited.", + "bbox": [ + 75, + 338, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objectives", + "text_level": 1, + "bbox": [ + 76, + 377, + 266, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction Losses. For both no editing and cycle consistency path outputs, our goal is to reconstruct the input image. To supervise this behavior, we use $L_{2}$ loss, perceptual loss, and identity loss between the input and output images. The first reconstruction loss is as follows, where $x$ is the input image, $x'$ and $x''$ are output images, as given in Fig. 4:", + "bbox": [ + 75, + 400, + 470, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c - l 2} = \\left| \\left| x ^ {\\prime} - x \\right| \\right| _ {2} + \\left| \\left| x ^ {\\prime \\prime} - x \\right| \\right| _ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 532, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use perceptual losses from VGG $(\\Phi)$ at different feature layers $(j)$ between these images from the loss objective as given in Eq. 5.", + "bbox": [ + 76, + 563, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c - p} = \\left| \\left| \\Phi_ {j} \\left(x ^ {\\prime}\\right) - \\Phi_ {j} (x) \\right| \\right| _ {2} + \\left| \\left| \\Phi_ {j} \\left(x ^ {\\prime \\prime}\\right) - \\Phi_ {j} (x) \\right| \\right| _ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 631, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Identity loss is calculated with a pre-trained network $A$ . $A$ is an ArcFace model [11] when training on the face domain and a domain specific ResNet-50 model [32] for our training on car class.", + "bbox": [ + 76, + 656, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c - i d} = (1 - \\langle A (x), A \\left(x ^ {\\prime}\\right) \\rangle) + (1 - \\langle A (x), A \\left(x ^ {\\prime \\prime}\\right) \\rangle) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 727, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Losses. We also use adversarial losses to guide the network to output images that match the real image distribution. This objective improves realistic image inversions and edits. We load the pretrained discriminator from StyleGAN training, $D$ , and train the discriminator together with the encoders.", + "bbox": [ + 76, + 753, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {a d v} = 2 \\log D (x) + \\log \\left(1 - D \\left(x ^ {\\prime}\\right)\\right) \\tag {7} \\\\ + \\log \\left(1 - D \\left(x _ {i} ^ {\\prime}\\right)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 867, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Regularizer. To prevent our encoder from deviating much from the original StyleGAN space, we regularize the residual features to be small:", + "bbox": [ + 498, + 338, + 892, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {F} = \\sum_ {F \\in \\phi} \\| F \\| _ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 393, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Objective. We use the overall objectives given below. The hyper parameters are provided in Supplementary.", + "bbox": [ + 498, + 436, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {E _ {1}, E _ {2}} \\max _ {D} \\lambda_ {a} \\mathcal {L} _ {a d v} + \\lambda_ {r 1} \\mathcal {L} _ {r e c - l 2} + \\lambda_ {r 2} \\mathcal {L} _ {r e c - p} \\tag {9} \\\\ + \\lambda_ {r 3} \\mathcal {L} _ {r e c - i d} + \\lambda_ {f} \\mathcal {L} _ {F} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 491, + 890, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 558, + 633, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Set-up. For datasets and attribute editing, we follow the previous work [34]. For the human face domain, we train the model on the FFHQ [21] dataset and evaluate it on the CelebA-HQ [20] dataset. For the car domain, we use Stanford Cars [23] for training and evaluation. We run extensive experiments with directions explored with InterfaceGAN [30], GANSpace [15], StyleClip [26], and GradCtrl [7] methods.", + "bbox": [ + 496, + 583, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. We report metrics for reconstruction and editing qualities. For reconstruction, we report Frechet Inception Distance (FID) metric [16], which looks at the realism by comparing the target image distribution and reconstructed images, Learned Perceptual Image Patch Similarity (LPIPS) [44] and Structural Similarity Index Measure (SSIM), which compares the target and output pairs at the feature level of a pretrained deep network and pixel level, respectively. Additionally, FIDs are calculated for adding and removing the smile attributes on the CelebAHQ dataset. By using the ground-truth attributes, we add smile to images that do not have smile attribute and calculate FIDs between smile addition edited images and smiling", + "bbox": [ + 496, + 704, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1832", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/97bf85cbdf4c310efa39a4b361107c0245a5f88bc739cec5115373316276cc71.jpg", + "image_caption": [ + "Figure 5. Qualitative results of inversion and editing. For each method, the first column shows inversion, and the second shows InterfaceGAN [30] and GANSpace [15] edits." + ], + "image_footnote": [], + "bbox": [ + 81, + 98, + 893, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ground-truth images. The same set-up is used for smile removal. On the Stanford cars dataset, we change grass and color attribute of images, and since there is no ground-truth attribute, we calculate the FIDs between the edited and original images.", + "bbox": [ + 75, + 709, + 470, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with state-of-the-art image inversion methods pSp [27], e4e [32], ReStyle [4], HyperStyle [5], HFGI [34], StyleTransformer [18], and FeatureStyle [40]. We use the author's released models. Hence, for the Stanford car dataset, some methods are omitted from comparisons if the models are not released. Among those, we only train HFGI for car model with au", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "thor's released code since we base our main comparisons with it.", + "bbox": [ + 498, + 709, + 890, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. In Table 1, we provide reconstruction and editing scores on the CelebA-HQ dataset. Our method achieves better results than all competing methods on all metrics. Most significantly, we achieve better FID scores on both reconstruction and editing qualities. While FeatureStyle achieves comparable SSIM and LPIPS scores on reconstruction, the editing FIDs are worse than our model. As shown in Table 2, we achieve significantly better results than previous methods on the Stanford Car dataset as well. We also compare the runtime of our method", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1833", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/983909fc62436f9c50afef4969a58c0a7ea90929d952c72176200d258472f9af.jpg", + "table_caption": [ + "Table 1. Quantitative results of reconstruction and editing on CelebA-HQ dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for smile addition (+) and removal (-)." + ], + "table_footnote": [], + "table_body": "
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSSmile(+)Smile(-)
pSp [27]23.860.750.1732.4734.0
e4e [32]30.220.710.2138.5839.68
ReStyle [4]24.820.730.2030.3533.69
HyperStyle [5]16.080.830.1126.4325.26
HFGI [34]12.170.850.1325.2227.10
StyleTransformer [18]21.820.750.1734.3234.61
FeatureStyle [40]11.330.900.1027.2026.15
StyleRes (Ours)7.040.900.0923.5221.80
", + "bbox": [ + 81, + 156, + 467, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ceb05a457aead6a7bd5033b9792fe04a720b404ac97d75de9b6b9e17d0a38b4.jpg", + "table_caption": [ + "Table 2. Quantitative results of reconstruction and editing on the Stanford Cars Dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for grass addition and color change." + ], + "table_footnote": [], + "table_body": "
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSGrassColor
e4e [32]14.040.500.3218.0229.79
ReStyle [4]13.380.570.3016.0121.34
HyperStyle [5]11.640.630.2817.1326.30
HFGI [34]9.410.830.1614.8426.65
StyleTransformer [18]14.010.570.2819.4719.94
StyleRes (Ours)7.600.830.1410.6418.86
", + "bbox": [ + 81, + 359, + 467, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and previous methods. Our method is significantly faster than HyperStyle (0.125 sec vs. 0.439 sec). That is because HyperStyle refines its predicted weight offsets gradually via multiple iterations. Our method is also faster than HFGI (0.125 sec vs. 0.130 sec), in addition to achieving better scores. We run a single stage network inference, whereas HFGI first generates an image via e4e and StyleGAN and provides the error map to a second architecture. The table for runtimes is provided in Supplementary.", + "bbox": [ + 75, + 492, + 467, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We show visuals of inversion and editing results of our method, e4e, HFGI, and HyperStyle in Fig. 5. We provide further comparisons in Supplementary with other attribute editorings and with other methods. Compared to previous methods, our method achieves significantly better fidelity to the input images and preserves the identity and details when edited. It is the only method in these comparisons that preserves the background, earings (second row), hands (second-fourth rows), and hats (fourth row). Our method also achieves facial detail reconstruction; for example, in the fifth row, the person has a mole at the corner of her mouth. Among the inversions, our method is the only one preserving that. Furthermore, during the pose edit, it is transformed correctly. On the sixth row, our method is the only one that achieves the correct inversion and edit. In car examples, we again achieve high fidelity to the input images both in inversion and editing. e4e and HyperStyle do not reconstruct the image faithfully. On the", + "bbox": [ + 75, + 628, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/323b7dea6ed040bf91e1d3464a5b225612ebe14f8e41b0d67629629342d9445f.jpg", + "image_caption": [ + "Figure 6. Comparison of our method with e4e, HFGI, and HyperStyle with StyleClip edits [26]. The first column shows eyeglasses addition, the second and third columns show bangs addition, and the last column shows bob cut hairstyle results." + ], + "image_footnote": [], + "bbox": [ + 516, + 99, + 880, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "other hand, HFGI outputs artefacts during edits. We additionally show results with edits explored by StyleClip [26] and GradCtrl [7] methods in Figs. 6 and 7, respectively.", + "bbox": [ + 496, + 551, + 890, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We run extensive experiments to validate the role of each proposed design choices. We first experiment with an architecture that does not have $E_{1}$ and $E_{2}$ modules. This experiment refers to the case where we learn a network that does not take input from StyleGAN Part 1 features, neither the original $(G_{W+})$ nor the edited features $(G_{\\alpha})$ . The network directly takes higher layer features $(F_{0})$ from the encoder and outputs them to the generator. As shown in Fig. 8, the network still tries to learn residual features to achieve reconstructions; however, it achieves poor edits. Without $E_{2}$ refers to the experiment with $E_{1}$ directly outputting features to Generator Part 2 (as shown in Fig. 3). Without $E_{1}$ refers to the experiment where $E_{2}$ directly takes input from the encoder $(F_{0})$ and $(G_{\\alpha})$ as defined in Sec. 3.2. None of the networks are able to learn the residual features and how to transform them as well as our final architecture since they are not designed to take the original and edited features separately. We also experiment without cycle consistency losses. Fig. 9 shows visual results of methods trained with and without cycle consistency constrain. We", + "bbox": [ + 496, + 599, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1834", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60fcc17268c68bc0e499e97a161b701fcfd8ac9c00e101ceda5e5bb5910b8bde.jpg", + "image_caption": [ + "Figure 7. Additional results of our method and e4e with GradCtrl edits [7]. The first three rows show blue sky edit, and the others show tree background edits." + ], + "image_footnote": [], + "bbox": [ + 84, + 99, + 467, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "observe that with cycle consistency constrain, our network achieves preserving fine image details even when images are edited.", + "bbox": [ + 75, + 633, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 693, + 194, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We show state-of-the-art results on a wide range of edits explored with different methods both quantitatively and qualitatively while achieving faster run-time than competing methods.", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9b2b4eb28f870563d8a3cdf00dd08fcade7106beee3157f8693e5cbf708d223.jpg", + "image_caption": [ + "Figure 8. Ablation Study. Pose edit outputs of our final architecture and architecture with missing modules. w/o $E_{1}$ or $E_{2}$ , networks struggle to transform features correctly." + ], + "image_footnote": [], + "bbox": [ + 511, + 99, + 885, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07bec594a38d21401644b96d3794980a1c29c83f59a7c3a3f20843ee26b28bc2.jpg", + "image_caption": [ + "Figure 9. Ablation Study. Pose edit outputs of models trained with and without cycle consistency loss. The model trained with cycle consistency is better at preserving the details as shown in enlarged boxes." + ], + "image_footnote": [], + "bbox": [ + 513, + 523, + 880, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 797, + 658, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has been funded by The Scientific and Technological Research Council of Turkey (TUBITAK), 3501 Research Project under Grant No 121E097. A. Dundar was supported by Marie Skłodowska-Curie Individual Fellowship.", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1835", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2", + "[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2, 3", + "[4] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 6, 7", + "[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 1, 2, 6, 7", + "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2", + "[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In European Conference on Computer Vision, pages 104-119. Springer, 2022. 5, 7, 8", + "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[9] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. IEEE transactions on neural networks and learning systems, 30(7):1967-1974, 2018. 2", + "[10] Yusuf Dalva, Said Fahri Altindis, and Aysegul Dundar. Vecgan: Image-to-image translation with interpretable latent directions. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 2", + "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5", + "[12] Aysegul Dundar, Karan Sapra, Guilin Liu, Andrew Tao, and Bryan Catanzaro. Panoptic-based image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8070-8079, 2020. 2", + "[13] Yue Gao, Fangyun Wei, Jianmin Bao, Shuyang Gu, Dong Chen, Fang Wen, and Zhouhui Lian. High-fidelity and arbitrary face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16115-16124, 2021. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33, 2020. 1, 3, 5, 6", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[17] Xianxu Hou, Xiaokang Zhang, Hanbang Liang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. Neural Networks, 145:209-220, 2022. 2", + "[18] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11337-11346, June 2022. 6, 7", + "[19] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. Eur. Conf. Comput. Vis., 2018. 2", + "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 5", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2, 5", + "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 3", + "[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5", + "[24] Xinyang Li, Shengchuan Zhang, Jie Hu, Liujuan Cao, Xiaopeng Hong, Xudong Mao, Feiyue Huang, Yongjian Wu, and Rongrong Ji. Image-to-image translation via hierarchical style disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8639-8648, 2021. 2", + "[25] Guilin Liu, Aysegul Dundar, Kevin J Shih, Ting-Chun Wang, Fitsum A Reda, Karan Sapra, Zhiding Yu, Xiaodong Yang, Andrew Tao, and Bryan Catanzaro. Partial convolution for padding, inpainting, and image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1836", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 1, 2, 6, 7", + "[28] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 2", + "[29] Wei Shen and Rujie Liu. Learning residual images for face attribute manipulation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4030-4038, 2017. 2", + "[30] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 3, 5, 6", + "[31] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 3", + "[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 1, 2, 3, 5, 6, 7, 8", + "[33] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International Conference on Machine Learning, pages 9786-9796. PMLR, 2020. 1, 3", + "[34] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 1, 2, 5, 6, 7", + "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2", + "[36] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5914–5922, 2019. 2", + "[37] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 3", + "[38] Taihong Xiao, Jiapeng Hong, and Jinwen Ma. Elegant: Exchanging latent encodings with gan for transferring multiple face attributes. In Proceedings of the European conference on computer vision (ECCV), pages 168-184, 2018. 2", + "[39] Guoxing Yang, Nanyi Fei, Mingyu Ding, Guangzhen Liu, Zhiwu Lu, and Tao Xiang. L2m-gan: Learning to manipulate latent space semantics for facial attribute editing. In" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2951-2960, 2021. 2", + "[40] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 6, 7", + "[41] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 1", + "[42] Gang Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Generative adversarial network with spatial attention for face attribute editing. In Proceedings of the European conference on computer vision (ECCV), pages 417-432, 2018. 2", + "[43] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR, 2019. 1", + "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[45] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2", + "[46] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A Efros, Oliver Wang, and Eli Shechtman. Multimodal image-to-image translation by enforcing bi-cycle consistency. In Advances in neural information processing systems, pages 465-476, 2017. 2" + ], + "bbox": [ + 501, + 92, + 893, + 558 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1837", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_model.json b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2df1b3fb6cbeb204ff32eee49e7a071703f31ef1 --- /dev/null +++ b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "StyleRes: Transforming the Residuals for Real Image Editing with StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.181, + 0.709, + 0.218 + ], + "angle": 0, + "content": "Hamza Pehlivan Yusuf Dalva Aysegul Dundar Bilkent University" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.22, + 0.674, + 0.234 + ], + "angle": 0, + "content": "{hamza.pehlivan,yusuf.dalva}@bilkent.edu.tr" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.238, + 0.597, + 0.251 + ], + "angle": 0, + "content": "adundar@cs.bilkent.edu.tr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.636 + ], + "angle": 0, + "content": "We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. Inverting real images into StyleGAN's latent space is an extensively studied problem, yet the trade-off between the image reconstruction fidelity and image editing quality remains an open challenge. The low-rate latent spaces are limited in their expressiveness power for high-fidelity reconstruction. On the other hand, high-rate latent spaces result in degradation in editing quality. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We train the framework to extract residual features and transform them via a novel architecture pipeline and cycle consistency losses. We run extensive experiments and compare our method with state-of-the-art inversion methods. Qualitative metrics and visual comparisons show significant improvements. Code: https://github.com/hamzapehlivan/StyleRes" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Generative Adversarial Networks (GANs) achieve high quality synthesis of various objects that are hard to distinguish from real images [14,21,22,41,43]. These networks also have an important property that they organize their latent space in a semantically meaningful way; as such, via latent editing, one can manipulate an attribute of a generated image. This property makes GANs a promising technology for image attribute editing and not only for generated images but also for real images. However, for real images, one also needs to find the corresponding latent code that will generate the particular real image. For this purpose, different GAN inversion methods are proposed, aiming to project real images to pretrained GAN latent space [15,30,31,33,37]. Even though this is an extensively stud" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.296, + 0.907, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.667, + 0.895, + 0.723 + ], + "angle": 0, + "content": "Figure 1. Comparison of our method with e4e, HFGI, and HyperStyle for the pose, bob cut hairstyle, smile removal, and color change edits. Our method achieves high fidelity to the input and high quality edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.795 + ], + "angle": 0, + "content": "ied problem with significant progress, the trade-off between image reconstruction fidelity and image editing quality remains an open challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The trade-off between image reconstruction fidelity and image editing quality is referred to as the distortion-editability trade-off [32]. Both are essential for real image editing. However, it is shown that the low-rate latent spaces are limited in their expressiveness power, and not every image can be inverted with high fidelity reconstruction [1, 27, 32, 34]. For that reason, higher bit encodings" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "and more expressive style spaces are explored for image inversion [1, 2]. Although with these techniques, images can be reconstructed with better fidelity, the editing quality decreases since there is no guarantee that projected codes will naturally lie in the generator's latent manifold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.473, + 0.575 + ], + "angle": 0, + "content": "In this work, we propose a framework that achieves high fidelity input reconstruction and significantly improved editability compared to the state-of-the-art. We learn residual features in higher-rate latent codes that are missing in the reconstruction of encoded features. This enables us to reconstruct image details and background information which are difficult to reconstruct via low rate latent encodings. Our architecture is single stage and learns the residuals based on the encoded features from the encoder and generated features from the pretrained GANs. We also learn a module to transform the higher-latent codes if needed based on the generated features (e.g. when the low-rate latent codes are manipulated). This way, when low-rate latent codes are edited for attribute manipulation, the decoded features can adapt to the edits to reconstruct details. While the attributes are not edited, the encoder can be trained with image reconstruction and adversarial losses. On the other hand, when the image is edited, we cannot use image reconstruction loss to regularize the network to preserve the details. To guide the network to learn correct transformations based on the generated features, we train the model with adversarial loss and cycle consistency constraint; that is, after we edit the latent code and generate an image, we reverse the edit and aim at reconstructing the original image. Since we do not want our method to be limited to predefined edits, during training, we simulate edits by randomly interpolating them with sampled latent codes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.698 + ], + "angle": 0, + "content": "The closest to our approach is HFGI [34], which also learns higher-rate encodings. Our framework is different as we learn a single-stage architecture designed to learn features that are missing from low-rate encodings and we learn how to transform them based on edits. As shown in Fig. 1, our framework achieves significantly better results than HFGI and other methods in editing quality. In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.709, + 0.47, + 0.754 + ], + "angle": 0, + "content": "- We propose a single-stage framework that achieves high-fidelity input embedding and editing. Our framework achieves that with a novel encoder architecture." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.767, + 0.47, + 0.859 + ], + "angle": 0, + "content": "- We propose to guide the image projection with cycle consistency and adversarial losses. We edit encoded images by taking a step toward a randomly sampled latent code. We expect to reconstruct the original image when the edit is reversed. This way, edited images preserve the details of the input image, and edits become high quality." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We conduct extensive experiments to show the effectiveness of our framework and achieve significant improve" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.709, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "ments over state-of-the-art for both reconstruction and real image attribute manipulations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.136, + 0.642, + 0.151 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.895, + 0.705 + ], + "angle": 0, + "content": "GAN Inversion. GAN inversion methods aim at projecting a real image into GANs embedding so that from the embedding, GAN can generate the given image. Currently, the biggest motivation of the inversion is the ability to edit the image via semantically rich disentangled features of GANs; therefore models aim for high reconstruction and editing quality. Inversion methods can be categorized into two; 1) methods that directly optimize the latent vector to minimize the reconstruction error between the output and target image [1,2,9,22,28], 2) methods that learn encoders to reconstruct images over training images [32,34,45]. Optimization based methods require per-image optimization and iterative passes on GANs, which take several minutes per image. Additionally, overfitting on a single image results in latent codes that do not lie in GAN's natural latent distribution, leading to poor editing quality. Among these methods, PTI [28] takes a different approach, and instead of searching for a latent code that will reconstruct the image most accurately, PTI fine-tunes the generator in order to insert encoded latent code into well-behaved regions of the latent space. It shows better editability; however, the method still suffers from long run-time optimizations and tunings. Encoder based methods leverage the knowledge of training sets while projecting images. They output results with less fidelity to the input, but edits on the projected latents are better quality. They also operate mostly in real-time. They either project the latent code with a single feed-forward pass [27,32,45], or two stage feed-forward passes where the first encoder learns to embed an image, the second learns to reconstruct the missing details [34], or iterative forward passes where the network tries to minimize the reconstruction loss at each pass by taking the original image and reconstructed output as inputs [4,5]. In this work, we propose a novel encoder architecture that achieves significantly better results than state-of-the-art via a single feed-forward pass on the input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image Translation. There is quite an interest in image translation algorithms that can change an attribute of an image while preserving a given content [12, 25, 35], especially for faces editing [3, 8, 10, 13, 17, 24, 29, 36, 38, 42]. These algorithms set an encoder-decoder architecture and train the models with reconstruction and GAN losses [8, 19, 24, 39, 46]. Even though they are successful at image manipulation, they are trained for a given task and rely on predefined translation datasets. On the other hand, it is shown that GANs that are trained to synthesize objects in an unconditional way can be used for attribute manipulation [6, 21, 22] via their semantically rich disentangled features. This makes pretrained GANs promising technol-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.189, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.157, + 0.158, + 0.165 + ], + "angle": 0, + "content": "(a1) Input" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.089, + 0.28, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.158, + 0.277, + 0.173 + ], + "angle": 0, + "content": "(b1) Low-rate Inversion (e4e)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.089, + 0.371, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.158, + 0.357, + 0.174 + ], + "angle": 0, + "content": "(c1) Low-rate Edit (e4e)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.462, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.158, + 0.461, + 0.174 + ], + "angle": 0, + "content": "(d1) Low-rate Inversion High-rate Inversion" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.174, + 0.28, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.243, + 0.262, + 0.259 + ], + "angle": 0, + "content": "(1) Low-rate Edit High-rate Inversion" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.174, + 0.371, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.243, + 0.372, + 0.26 + ], + "angle": 0, + "content": "(1) Low-rate Edit High-rate Transformation" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.205, + 0.465, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.595, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.157, + 0.561, + 0.164 + ], + "angle": 0, + "content": "(a2) Input" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.09, + 0.685, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.158, + 0.68, + 0.174 + ], + "angle": 0, + "content": "(b3) Low-rate Inversion (e4e)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.09, + 0.776, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.158, + 0.759, + 0.174 + ], + "angle": 0, + "content": "(c2) Low-rate Edit (e4e)" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.089, + 0.868, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.158, + 0.867, + 0.174 + ], + "angle": 0, + "content": "(d2) Low-rate Inversion High-rate Inversion" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.174, + 0.584, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.174, + 0.685, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.243, + 0.665, + 0.259 + ], + "angle": 0, + "content": "62)Low-rate Edit High-rate Inversion" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.174, + 0.777, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.243, + 0.776, + 0.26 + ], + "angle": 0, + "content": "(f2) Low-rate Edit High-rate Transformation" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.176, + 0.845, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.894, + 0.34 + ], + "angle": 0, + "content": "Figure 2. When images are encoded to \\(\\mathrm{W + }\\) space, as shown in (b), the reconstructions miss many image details. However, the edits are in good quality (c). Additional features can be learned in higher-rate latent codes. For example, skip connections from encoder to generator in higher resolution features can enable high-fidelity reconstruction to input images as shown in (d). However, if the high-rate features do not transform with the edits, they result in ghosting effects as shown in (e). In this work, we propose high-rate encoding and transformation for successful inversions and edits (f)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.535 + ], + "angle": 0, + "content": "ogy for image editing, given that via their ability of image generation, they can achieve many downstream tasks simultaneously. Many methods have been proposed for finding latent directions to edit images [15, 30, 31, 33, 37]. These directions are explored both in supervised [3, 30] and unsupervised ways [15, 31, 33, 37] resulting in exploration of attribute manipulations beyond the predefined attributes of labeled datasets. However, finding these directions is only one part of the problem; to achieve real image editing, there is a need for a successful image inversion method which is the topic of this paper." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.55, + 0.169, + 0.566 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.623 + ], + "angle": 0, + "content": "In Section 3.1, we describe the motivation for our approach. The model architecture and the training procedure are presented in Sections 3.2 and 3.3, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.199, + 0.649 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Current inversion methods aim at learning an encoder to project an input image into StyleGAN's natural latent space so that the inverted image is editable. However, it is observed by previous works that when images are encoded into GAN's natural space (low-rate W or W+ space), their reconstructions suffer from low fidelity to input images, as shown in Fig. 2(b). On the other hand, if the image is encoded to higher feature maps of pretrained GANs, they will not be editable since they are not inverted to the semantic space of GANs. On the other hand, low-rate inversion is editable, as shown in Fig. 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "StyleGAN, during training and inference, relies on direct noise inputs to generate stochastic details. These additional noise inputs provide stochastic image details such as for face images, the exact placement of hairs, stubble, freckles, or skin pores which are not modeled by \\(\\mathrm{W}+\\) space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.473 + ], + "angle": 0, + "content": "During inversion to \\(\\mathrm{W + }\\) space alone, that mechanism is discarded. One can tune the noise maps as well via the reconstruction loss; however, then there will be no mechanism to adapt those stochastic details to attribute manipulation. For example, the freckles should move as the pose is manipulated, but with noise optimization alone, it will not be possible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.685 + ], + "angle": 0, + "content": "We propose to embed images to both low-rate and high-rate embeddings. Consistent with the design of StyleGAN, we aim at encoding the overall composition and high-level aspects of the image to \\(\\mathrm{W + }\\) space (low-rate encoding). Our goal is to embed the stochastic image details and the diverse background details, which are difficult to reconstruct from \\(\\mathrm{W + }\\) space to higher latent codes. However, this setting also requires a mechanism for encoded image details to adopt manipulations in the image. Otherwise, it will cause a ghosting effect, as shown in Fig 2(e). For example, in Fig 2(e1), the smile is removed by \\(\\mathrm{W + }\\) code edit; however, the higher-rate encodings stay the same and cause artifacts around the mouth area. In Fig. 2(e2), the pose is edited but higher detail encodings did not move and caused blur." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.747 + ], + "angle": 0, + "content": "In this work, we design an encoder architecture and training pipeline to achieve both learning residual features (the ones \\(\\mathrm{W}+\\) space could not reconstruct) and how to transform them to be consistent with the manipulations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.704, + 0.772 + ], + "angle": 0, + "content": "3.2. StyleRes Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our method utilizes an encoder \\( E_0 \\) that can embed images into \\( W+ \\) latent space and a StyleGAN generator \\( G \\). In our setup, we utilize a pretrained encoder for \\( E_0 \\) [32] and StyleGAN2 generator [22] and fix them in our training. Because it is difficult to preserve image details only from low-rate latent codes, we also extract high level features from the encoder. First, the high and low rate features are extracted as \\( F_0 \\), \\( W^+ = E_0(x) \\) using the encoder \\( E_0 \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1830" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.093, + 0.895, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Figure 3. StyleRes encodes missing features for high-fidelity reconstruction of given input via the first encoder, \\( E_{1} \\). Those encoded features are the ones which could not be encoded to low-rate W+ space via \\( E_{0} \\) due to the information bottleneck. Through the second encoder, \\( E_{2} \\), StyleRes learns to transform features based on the manipulated features. During training, latent codes are edited by interpolating encoded W+’s with randomly generated ones by StyleGAN's mapping network. During inference, they are edited with semantically meaningful directions discovered by methods such as InterfaceGAN and GANSpace. Note that StyleGAN generator is shown as two parts just for the ease of visualizing the diagram. First part includes the layers that generate features to \\( 64 \\times 64 \\) and the second part generates the higher resolution features and final image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.473, + 0.595 + ], + "angle": 0, + "content": "from the input image \\( x \\) as shown in Fig. 3. \\( F_{0} \\) has the spatial dimension of \\( 64 \\times 64 \\) and provides us with more image details. Next, from \\( F_{0} \\), our aim is to encode residuals that are missing from the image reconstruction. For this goal, we set an encoder, \\( E_{1} \\), which takes \\( F_{0} \\) as input. Since the goal of \\( E_{1} \\) is to encode residual features, we also feed the generated features with \\( W^{+} \\) from the StyleGAN generator, \\( G_{W} = G_{0\\rightarrow n}(W) \\), where the arrow operator indicates the indices of convolutional layers used from \\( G \\)." + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.612, + 0.47, + 0.629 + ], + "angle": 0, + "content": "\\[\nF _ {a} = E _ {1} \\left(F _ {0}, G _ {W +}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.473, + 0.821 + ], + "angle": 0, + "content": "With the inputs of \\( F_{0} \\) and \\( G_{W+} \\), \\( E_{1} \\) can learn what the missing features are by comparing the encoded features \\( F_{0} \\) and generated ones, \\( G_{W+} \\), so far. While \\( E_{1} \\) can learn the residual features, it is not guided on how to transform them if images are edited. For that purpose, we train \\( E_{2} \\), which takes \\( F_{a} \\) and edited features from the generator. Since we do not target predefined edits (e.g., smile, pose, age), we simulate the edits by taking random directions. Specifically, we sample a \\( z \\) vector from the normal distribution and obtain \\( W_{r}^{+} \\) by StyleGAN's mapping network, \\( M \\); \\( W_{r}^{+} = M(z) \\). Next, we take a step towards \\( W_{r}^{+} \\) to obtain a mixed style code \\( W_{\\alpha}^{+} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.832, + 0.47, + 0.865 + ], + "angle": 0, + "content": "\\[\nW _ {\\alpha} ^ {+} = W ^ {+} + \\alpha \\frac {W _ {r} ^ {+} - W ^ {+}}{1 0} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\alpha\\) controls the degree of the edit. During the training, we assign \\(\\alpha\\) to 0 (no edit) or a value in the range of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.505 + ], + "angle": 0, + "content": "(4,5) with \\(50\\%\\) chance. We adjust \\(F_{a}\\) to the altered generator features \\(G_{\\alpha} = G_{0 \\to n}(W_{\\alpha})\\) using a second encoder \\(E_{2}\\) and obtain the final feature map F:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.518, + 0.892, + 0.535 + ], + "angle": 0, + "content": "\\[\nF = E _ {2} \\left(F _ {a}, G _ {\\alpha}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.894, + 0.585 + ], + "angle": 0, + "content": "Finally, \\( F \\) and \\( G_{\\alpha} \\) are summed, and given as an input to next convolutional layers in the generator. Architectural details of \\( E_{1} \\) and \\( E_{2} \\) are given in Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.662, + 0.607 + ], + "angle": 0, + "content": "3.3. Training Phases" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.659 + ], + "angle": 0, + "content": "To train our model with the capabilities of high-fidelity inversion and high-quality editing, we use two training phases, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.765 + ], + "angle": 0, + "content": "No Editing Path. In this path, we reconstruct images with no editing on encoded \\( W^{+} \\). This refers to the case where \\( \\alpha = 0 \\) in Eq. 2. This is also the training path other inversion methods solely rely on. While this path can teach the network to reconstruct images faithfully, it does not include any edits and cannot guide the network to high-quality editability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Cycle Translation Path. In this path, we edit images by setting \\(\\alpha\\) a value in the range of (4, 5), which we found to work best in our ablation studies. Via the edit, the generator outputs image \\(x_{i}^{\\prime}\\). Next, we feed this intermediate output image to the encoder and reverse the edit by inverting the addition presented in Eq. 2. The generator reconstructs \\(x''\\), which is supposed to match the input image \\(x\\). The cycle translation path is important because there is no ground-truth output for the edited image, \\(x_{i}^{\\prime}\\). Adversarial loss can" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1831" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.357, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.218, + 0.327, + 0.231 + ], + "angle": 0, + "content": "(a) No Editing: Reconstruction Loss" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.091, + 0.891, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.219, + 0.798, + 0.232 + ], + "angle": 0, + "content": "(b) Cycle-consistency: Reconstruction Loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Figure 4. We train StyleRes with (a) no editing based reconstruction and (b) cycle consistency based reconstruction losses. Many details from the inference pipeline are omitted for brevity. With reconstruction losses, the model learns to preserve the image details. We additionally apply adversarial losses on \\( x_{i}^{\\prime} \\); this way, when the image is edited, the transformation network learn to output realistic images. With cycle consistency based reconstruction, the network is regularized to keep the input details also during edits. We use additional losses as explained in Section 3.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.371 + ], + "angle": 0, + "content": "guide \\( x_{i}^{\\prime} \\) to look realistic but will not guide the network to keep the input image details if edited." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.378, + 0.267, + 0.395 + ], + "angle": 0, + "content": "3.4. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.507 + ], + "angle": 0, + "content": "Reconstruction Losses. For both no editing and cycle consistency path outputs, our goal is to reconstruct the input image. To supervise this behavior, we use \\( L_{2} \\) loss, perceptual loss, and identity loss between the input and output images. The first reconstruction loss is as follows, where \\( x \\) is the input image, \\( x' \\) and \\( x'' \\) are output images, as given in Fig. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.534, + 0.47, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c - l 2} = \\left| \\left| x ^ {\\prime} - x \\right| \\right| _ {2} + \\left| \\left| x ^ {\\prime \\prime} - x \\right| \\right| _ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.47, + 0.61 + ], + "angle": 0, + "content": "We use perceptual losses from VGG \\((\\Phi)\\) at different feature layers \\((j)\\) between these images from the loss objective as given in Eq. 5." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.632, + 0.47, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c - p} = \\left| \\left| \\Phi_ {j} \\left(x ^ {\\prime}\\right) - \\Phi_ {j} (x) \\right| \\right| _ {2} + \\left| \\left| \\Phi_ {j} \\left(x ^ {\\prime \\prime}\\right) - \\Phi_ {j} (x) \\right| \\right| _ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Identity loss is calculated with a pre-trained network \\( A \\). \\( A \\) is an ArcFace model [11] when training on the face domain and a domain specific ResNet-50 model [32] for our training on car class." + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.728, + 0.47, + 0.746 + ], + "angle": 0, + "content": "\\[\nL _ {r e c - i d} = (1 - \\langle A (x), A \\left(x ^ {\\prime}\\right) \\rangle) + (1 - \\langle A (x), A \\left(x ^ {\\prime \\prime}\\right) \\rangle) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.755, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Adversarial Losses. We also use adversarial losses to guide the network to output images that match the real image distribution. This objective improves realistic image inversions and edits. We load the pretrained discriminator from StyleGAN training, \\( D \\), and train the discriminator together with the encoders." + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.868, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {a d v} = 2 \\log D (x) + \\log \\left(1 - D \\left(x ^ {\\prime}\\right)\\right) \\tag {7} \\\\ + \\log \\left(1 - D \\left(x _ {i} ^ {\\prime}\\right)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.339, + 0.893, + 0.384 + ], + "angle": 0, + "content": "Feature Regularizer. To prevent our encoder from deviating much from the original StyleGAN space, we regularize the residual features to be small:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.394, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\nL _ {F} = \\sum_ {F \\in \\phi} \\| F \\| _ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Full Objective. We use the overall objectives given below. The hyper parameters are provided in Supplementary." + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.492, + 0.891, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {E _ {1}, E _ {2}} \\max _ {D} \\lambda_ {a} \\mathcal {L} _ {a d v} + \\lambda_ {r 1} \\mathcal {L} _ {r e c - l 2} + \\lambda_ {r 2} \\mathcal {L} _ {r e c - p} \\tag {9} \\\\ + \\lambda_ {r 3} \\mathcal {L} _ {r e c - i d} + \\lambda_ {f} \\mathcal {L} _ {F} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.634, + 0.576 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Set-up. For datasets and attribute editing, we follow the previous work [34]. For the human face domain, we train the model on the FFHQ [21] dataset and evaluate it on the CelebA-HQ [20] dataset. For the car domain, we use Stanford Cars [23] for training and evaluation. We run extensive experiments with directions explored with InterfaceGAN [30], GANSpace [15], StyleClip [26], and GradCtrl [7] methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Evaluation. We report metrics for reconstruction and editing qualities. For reconstruction, we report Frechet Inception Distance (FID) metric [16], which looks at the realism by comparing the target image distribution and reconstructed images, Learned Perceptual Image Patch Similarity (LPIPS) [44] and Structural Similarity Index Measure (SSIM), which compares the target and output pairs at the feature level of a pretrained deep network and pixel level, respectively. Additionally, FIDs are calculated for adding and removing the smile attributes on the CelebAHQ dataset. By using the ground-truth attributes, we add smile to images that do not have smile attribute and calculate FIDs between smile addition edited images and smiling" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.099, + 0.895, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.655, + 0.895, + 0.684 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of inversion and editing. For each method, the first column shows inversion, and the second shows InterfaceGAN [30] and GANSpace [15] edits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.471, + 0.787 + ], + "angle": 0, + "content": "ground-truth images. The same set-up is used for smile removal. On the Stanford cars dataset, we change grass and color attribute of images, and since there is no ground-truth attribute, we calculate the FIDs between the edited and original images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare our method with state-of-the-art image inversion methods pSp [27], e4e [32], ReStyle [4], HyperStyle [5], HFGI [34], StyleTransformer [18], and FeatureStyle [40]. We use the author's released models. Hence, for the Stanford car dataset, some methods are omitted from comparisons if the models are not released. Among those, we only train HFGI for car model with au" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.892, + 0.74 + ], + "angle": 0, + "content": "thor's released code since we base our main comparisons with it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. In Table 1, we provide reconstruction and editing scores on the CelebA-HQ dataset. Our method achieves better results than all competing methods on all metrics. Most significantly, we achieve better FID scores on both reconstruction and editing qualities. While FeatureStyle achieves comparable SSIM and LPIPS scores on reconstruction, the editing FIDs are worse than our model. As shown in Table 2, we achieve significantly better results than previous methods on the Stanford Car dataset as well. We also compare the runtime of our method" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1833" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.471, + 0.145 + ], + "angle": 0, + "content": "Table 1. Quantitative results of reconstruction and editing on CelebA-HQ dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for smile addition (+) and removal (-)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.157, + 0.468, + 0.278 + ], + "angle": 0, + "content": "
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSSmile(+)Smile(-)
pSp [27]23.860.750.1732.4734.0
e4e [32]30.220.710.2138.5839.68
ReStyle [4]24.820.730.2030.3533.69
HyperStyle [5]16.080.830.1126.4325.26
HFGI [34]12.170.850.1325.2227.10
StyleTransformer [18]21.820.750.1734.3234.61
FeatureStyle [40]11.330.900.1027.2026.15
StyleRes (Ours)7.040.900.0923.5221.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.293, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Table 2. Quantitative results of reconstruction and editing on the Stanford Cars Dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for grass addition and color change." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.36, + 0.468, + 0.466 + ], + "angle": 0, + "content": "
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSGrassColor
e4e [32]14.040.500.3218.0229.79
ReStyle [4]13.380.570.3016.0121.34
HyperStyle [5]11.640.630.2817.1326.30
HFGI [34]9.410.830.1614.8426.65
StyleTransformer [18]14.010.570.2819.4719.94
StyleRes (Ours)7.600.830.1410.6418.86
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.629 + ], + "angle": 0, + "content": "and previous methods. Our method is significantly faster than HyperStyle (0.125 sec vs. 0.439 sec). That is because HyperStyle refines its predicted weight offsets gradually via multiple iterations. Our method is also faster than HFGI (0.125 sec vs. 0.130 sec), in addition to achieving better scores. We run a single stage network inference, whereas HFGI first generates an image via e4e and StyleGAN and provides the error map to a second architecture. The table for runtimes is provided in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. We show visuals of inversion and editing results of our method, e4e, HFGI, and HyperStyle in Fig. 5. We provide further comparisons in Supplementary with other attribute editorings and with other methods. Compared to previous methods, our method achieves significantly better fidelity to the input images and preserves the identity and details when edited. It is the only method in these comparisons that preserves the background, earings (second row), hands (second-fourth rows), and hats (fourth row). Our method also achieves facial detail reconstruction; for example, in the fifth row, the person has a mole at the corner of her mouth. Among the inversions, our method is the only one preserving that. Furthermore, during the pose edit, it is transformed correctly. On the sixth row, our method is the only one that achieves the correct inversion and edit. In car examples, we again achieve high fidelity to the input images both in inversion and editing. e4e and HyperStyle do not reconstruct the image faithfully. On the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.1, + 0.882, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Figure 6. Comparison of our method with e4e, HFGI, and HyperStyle with StyleClip edits [26]. The first column shows eyeglasses addition, the second and third columns show bangs addition, and the last column shows bob cut hairstyle results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.891, + 0.599 + ], + "angle": 0, + "content": "other hand, HFGI outputs artefacts during edits. We additionally show results with edits explored by StyleClip [26] and GradCtrl [7] methods in Figs. 6 and 7, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation Study. We run extensive experiments to validate the role of each proposed design choices. We first experiment with an architecture that does not have \\( E_{1} \\) and \\( E_{2} \\) modules. This experiment refers to the case where we learn a network that does not take input from StyleGAN Part 1 features, neither the original \\( (G_{W+}) \\) nor the edited features \\( (G_{\\alpha}) \\). The network directly takes higher layer features \\( (F_{0}) \\) from the encoder and outputs them to the generator. As shown in Fig. 8, the network still tries to learn residual features to achieve reconstructions; however, it achieves poor edits. Without \\( E_{2} \\) refers to the experiment with \\( E_{1} \\) directly outputting features to Generator Part 2 (as shown in Fig. 3). Without \\( E_{1} \\) refers to the experiment where \\( E_{2} \\) directly takes input from the encoder \\( (F_{0}) \\) and \\( (G_{\\alpha}) \\) as defined in Sec. 3.2. None of the networks are able to learn the residual features and how to transform them as well as our final architecture since they are not designed to take the original and edited features separately. We also experiment without cycle consistency losses. Fig. 9 shows visual results of methods trained with and without cycle consistency constrain. We" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1834" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.1, + 0.468, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.565, + 0.472, + 0.608 + ], + "angle": 0, + "content": "Figure 7. Additional results of our method and e4e with GradCtrl edits [7]. The first three rows show blue sky edit, and the others show tree background edits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.68 + ], + "angle": 0, + "content": "observe that with cycle consistency constrain, our network achieves preserving fine image details even when images are edited." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.196, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We show state-of-the-art results on a wide range of edits explored with different methods both quantitatively and qualitatively while achieving faster run-time than competing methods." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.1, + 0.886, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.464, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 8. Ablation Study. Pose edit outputs of our final architecture and architecture with missing modules. w/o \\( E_{1} \\) or \\( E_{2} \\), networks struggle to transform features correctly." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.525, + 0.882, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.767 + ], + "angle": 0, + "content": "Figure 9. Ablation Study. Pose edit outputs of models trained with and without cycle consistency loss. The model trained with cycle consistency is better at preserving the details as shown in enlarged boxes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.798, + 0.66, + 0.815 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This work has been funded by The Scientific and Technological Research Council of Turkey (TUBITAK), 3501 Research Project under Grant No 121E097. A. Dundar was supported by Marie Skłodowska-Curie Individual Fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1835" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.299 + ], + "angle": 0, + "content": "[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In European Conference on Computer Vision, pages 104-119. Springer, 2022. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[9] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. IEEE transactions on neural networks and learning systems, 30(7):1967-1974, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[10] Yusuf Dalva, Said Fahri Altindis, and Aysegul Dundar. Vecgan: Image-to-image translation with interpretable latent directions. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[12] Aysegul Dundar, Karan Sapra, Guilin Liu, Andrew Tao, and Bryan Catanzaro. Panoptic-based image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8070-8079, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Yue Gao, Fangyun Wei, Jianmin Bao, Shuyang Gu, Dong Chen, Fang Wen, and Zhouhui Lian. High-fidelity and arbitrary face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16115-16124, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Xianxu Hou, Xiaokang Zhang, Hanbang Liang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. Neural Networks, 145:209-220, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11337-11346, June 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[19] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. Eur. Conf. Comput. Vis., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[24] Xinyang Li, Shengchuan Zhang, Jie Hu, Liujuan Cao, Xiaopeng Hong, Xudong Mao, Feiyue Huang, Yongjian Wu, and Rongrong Ji. Image-to-image translation via hierarchical style disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8639-8648, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Guilin Liu, Aysegul Dundar, Kevin J Shih, Ting-Chun Wang, Fitsum A Reda, Karan Sapra, Zhiding Yu, Xiaodong Yang, Andrew Tao, and Bryan Catanzaro. Partial convolution for padding, inpainting, and image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1836" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[27] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.176, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[28] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[29] Wei Shen and Rujie Liu. Learning residual images for face attribute manipulation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4030-4038, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[30] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[31] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[33] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International Conference on Machine Learning, pages 9786-9796. PMLR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[34] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.664 + ], + "angle": 0, + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[36] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5914–5922, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[37] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[38] Taihong Xiao, Jiapeng Hong, and Jinwen Ma. Elegant: Exchanging latent encodings with gan for transferring multiple face attributes. In Proceedings of the European conference on computer vision (ECCV), pages 168-184, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[39] Guoxing Yang, Nanyi Fei, Mingyu Ding, Guangzhen Liu, Zhiwu Lu, and Tao Xiang. L2m-gan: Learning to manipulate latent space semantics for facial attribute editing. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2951-2960, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[40] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[41] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[42] Gang Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Generative adversarial network with spatial attention for face attribute editing. In Proceedings of the European conference on computer vision (ECCV), pages 417-432, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[43] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[45] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[46] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A Efros, Oliver Wang, and Eli Shechtman. Multimodal image-to-image translation by enforcing bi-cycle consistency. In Advances in neural information processing systems, pages 465-476, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1837" + } + ] +] \ No newline at end of file diff --git a/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_origin.pdf b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5fa0c54a2a315c6556a48808b755d5941d7557cb --- /dev/null +++ b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/25e05c83-4dbb-4c83-a9f4-55eb76b40913_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ce1b49cf95ee838065195369cb1e347065c197acb3e6708975aa8125def58c0 +size 2242331 diff --git a/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/full.md b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/full.md new file mode 100644 index 0000000000000000000000000000000000000000..580ec3928df2aa2144cdcb43a64f658fe11f0726 --- /dev/null +++ b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/full.md @@ -0,0 +1,293 @@ +# StyleRes: Transforming the Residuals for Real Image Editing with StyleGAN + +Hamza Pehlivan Yusuf Dalva Aysegul Dundar Bilkent University + +{hamza.pehlivan,yusuf.dalva}@bilkent.edu.tr + +adundar@cs.bilkent.edu.tr + +# Abstract + +We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. Inverting real images into StyleGAN's latent space is an extensively studied problem, yet the trade-off between the image reconstruction fidelity and image editing quality remains an open challenge. The low-rate latent spaces are limited in their expressiveness power for high-fidelity reconstruction. On the other hand, high-rate latent spaces result in degradation in editing quality. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We train the framework to extract residual features and transform them via a novel architecture pipeline and cycle consistency losses. We run extensive experiments and compare our method with state-of-the-art inversion methods. Qualitative metrics and visual comparisons show significant improvements. Code: https://github.com/hamzapehlivan/StyleRes + +# 1. Introduction + +Generative Adversarial Networks (GANs) achieve high quality synthesis of various objects that are hard to distinguish from real images [14,21,22,41,43]. These networks also have an important property that they organize their latent space in a semantically meaningful way; as such, via latent editing, one can manipulate an attribute of a generated image. This property makes GANs a promising technology for image attribute editing and not only for generated images but also for real images. However, for real images, one also needs to find the corresponding latent code that will generate the particular real image. For this purpose, different GAN inversion methods are proposed, aiming to project real images to pretrained GAN latent space [15,30,31,33,37]. Even though this is an extensively stud + +![](images/b6a110052faad7df5ba714c0095b57d515f516b8c8916a12ecbf6a57ea694544.jpg) +Figure 1. Comparison of our method with e4e, HFGI, and HyperStyle for the pose, bob cut hairstyle, smile removal, and color change edits. Our method achieves high fidelity to the input and high quality edits. + +ied problem with significant progress, the trade-off between image reconstruction fidelity and image editing quality remains an open challenge. + +The trade-off between image reconstruction fidelity and image editing quality is referred to as the distortion-editability trade-off [32]. Both are essential for real image editing. However, it is shown that the low-rate latent spaces are limited in their expressiveness power, and not every image can be inverted with high fidelity reconstruction [1, 27, 32, 34]. For that reason, higher bit encodings + +and more expressive style spaces are explored for image inversion [1, 2]. Although with these techniques, images can be reconstructed with better fidelity, the editing quality decreases since there is no guarantee that projected codes will naturally lie in the generator's latent manifold. + +In this work, we propose a framework that achieves high fidelity input reconstruction and significantly improved editability compared to the state-of-the-art. We learn residual features in higher-rate latent codes that are missing in the reconstruction of encoded features. This enables us to reconstruct image details and background information which are difficult to reconstruct via low rate latent encodings. Our architecture is single stage and learns the residuals based on the encoded features from the encoder and generated features from the pretrained GANs. We also learn a module to transform the higher-latent codes if needed based on the generated features (e.g. when the low-rate latent codes are manipulated). This way, when low-rate latent codes are edited for attribute manipulation, the decoded features can adapt to the edits to reconstruct details. While the attributes are not edited, the encoder can be trained with image reconstruction and adversarial losses. On the other hand, when the image is edited, we cannot use image reconstruction loss to regularize the network to preserve the details. To guide the network to learn correct transformations based on the generated features, we train the model with adversarial loss and cycle consistency constraint; that is, after we edit the latent code and generate an image, we reverse the edit and aim at reconstructing the original image. Since we do not want our method to be limited to predefined edits, during training, we simulate edits by randomly interpolating them with sampled latent codes. + +The closest to our approach is HFGI [34], which also learns higher-rate encodings. Our framework is different as we learn a single-stage architecture designed to learn features that are missing from low-rate encodings and we learn how to transform them based on edits. As shown in Fig. 1, our framework achieves significantly better results than HFGI and other methods in editing quality. In summary, our main contributions are: + +- We propose a single-stage framework that achieves high-fidelity input embedding and editing. Our framework achieves that with a novel encoder architecture. +- We propose to guide the image projection with cycle consistency and adversarial losses. We edit encoded images by taking a step toward a randomly sampled latent code. We expect to reconstruct the original image when the edit is reversed. This way, edited images preserve the details of the input image, and edits become high quality. +- We conduct extensive experiments to show the effectiveness of our framework and achieve significant improve + +ments over state-of-the-art for both reconstruction and real image attribute manipulations. + +# 2. Related Work + +GAN Inversion. GAN inversion methods aim at projecting a real image into GANs embedding so that from the embedding, GAN can generate the given image. Currently, the biggest motivation of the inversion is the ability to edit the image via semantically rich disentangled features of GANs; therefore models aim for high reconstruction and editing quality. Inversion methods can be categorized into two; 1) methods that directly optimize the latent vector to minimize the reconstruction error between the output and target image [1,2,9,22,28], 2) methods that learn encoders to reconstruct images over training images [32,34,45]. Optimization based methods require per-image optimization and iterative passes on GANs, which take several minutes per image. Additionally, overfitting on a single image results in latent codes that do not lie in GAN's natural latent distribution, leading to poor editing quality. Among these methods, PTI [28] takes a different approach, and instead of searching for a latent code that will reconstruct the image most accurately, PTI fine-tunes the generator in order to insert encoded latent code into well-behaved regions of the latent space. It shows better editability; however, the method still suffers from long run-time optimizations and tunings. Encoder based methods leverage the knowledge of training sets while projecting images. They output results with less fidelity to the input, but edits on the projected latents are better quality. They also operate mostly in real-time. They either project the latent code with a single feed-forward pass [27,32,45], or two stage feed-forward passes where the first encoder learns to embed an image, the second learns to reconstruct the missing details [34], or iterative forward passes where the network tries to minimize the reconstruction loss at each pass by taking the original image and reconstructed output as inputs [4,5]. In this work, we propose a novel encoder architecture that achieves significantly better results than state-of-the-art via a single feed-forward pass on the input image. + +Image Translation. There is quite an interest in image translation algorithms that can change an attribute of an image while preserving a given content [12, 25, 35], especially for faces editing [3, 8, 10, 13, 17, 24, 29, 36, 38, 42]. These algorithms set an encoder-decoder architecture and train the models with reconstruction and GAN losses [8, 19, 24, 39, 46]. Even though they are successful at image manipulation, they are trained for a given task and rely on predefined translation datasets. On the other hand, it is shown that GANs that are trained to synthesize objects in an unconditional way can be used for attribute manipulation [6, 21, 22] via their semantically rich disentangled features. This makes pretrained GANs promising technol- + +![](images/7d12b3a79f2e1d372a601a2a9b5437bbdf542e752b780479378e718f2567b974.jpg) +(a1) Input + +![](images/e90e53b430e865f52e850830a6770d5c91149fb4a94c272f2a2f3fba976c720f.jpg) +(b1) Low-rate Inversion (e4e) + +![](images/64d5da3d2c3d32ae117b8b30ff23edfbacb02d053761e744b20d9ab7ff6ba50d.jpg) +(c1) Low-rate Edit (e4e) + +![](images/5f7508805408e03985480a1ccdd29648729d411517e28c06bc28396b48289a1b.jpg) +(d1) Low-rate Inversion High-rate Inversion + +![](images/75eb5812a2381097f15a8345007b2d4bb12e17377e6e23ff6a6babff13e0e523.jpg) +(1) Low-rate Edit High-rate Inversion + +![](images/0415f2b0cfb09985f92a6cbe2afabbd9542faef8ab6a3e374623e051ead97ce5.jpg) +(1) Low-rate Edit High-rate Transformation + +![](images/49ef69017312bf2abdbf04d01381c1c2cb8cc7c74b4513e5631e1a03b726315f.jpg) + +![](images/85685a9346cdd43ea4940b70bc302c9da019708b51777694bf8db148a0e12af1.jpg) +(a2) Input + +![](images/3ba937bcb007e3cbd1586cb384e5ac6bde53fb5cba5df00935f2d6bbc62743aa.jpg) +(b3) Low-rate Inversion (e4e) + +![](images/14fea19cdefd9de980c839b807c52745dc3fcaabd3b9d6dfd63903c8239fa957.jpg) +(c2) Low-rate Edit (e4e) + +![](images/d2eb7cd88e13b3cd2610f5cca8db0676538c2d3e341dcdae278c5738662a3562.jpg) +(d2) Low-rate Inversion High-rate Inversion + +![](images/e95e929c5586ac277dae561c506543f2693c56bb4d7c89b24602bad2721e8140.jpg) +Figure 2. When images are encoded to $\mathrm{W + }$ space, as shown in (b), the reconstructions miss many image details. However, the edits are in good quality (c). Additional features can be learned in higher-rate latent codes. For example, skip connections from encoder to generator in higher resolution features can enable high-fidelity reconstruction to input images as shown in (d). However, if the high-rate features do not transform with the edits, they result in ghosting effects as shown in (e). In this work, we propose high-rate encoding and transformation for successful inversions and edits (f). + +![](images/fd886fff90b9f737e70ab6a673353f738bf2bfd857bf94ecd707d80dddff59ac.jpg) +62)Low-rate Edit High-rate Inversion +(f2) Low-rate Edit High-rate Transformation + +![](images/7df925b8e6d83d187cd5a93187f4c93968d11d74ff1df7d42317835a3eb69acb.jpg) + +![](images/2f424ffd6f3e78a0d3f6725991688d26ff06ef70235ad99bf01e59f965cef20e.jpg) + +ogy for image editing, given that via their ability of image generation, they can achieve many downstream tasks simultaneously. Many methods have been proposed for finding latent directions to edit images [15, 30, 31, 33, 37]. These directions are explored both in supervised [3, 30] and unsupervised ways [15, 31, 33, 37] resulting in exploration of attribute manipulations beyond the predefined attributes of labeled datasets. However, finding these directions is only one part of the problem; to achieve real image editing, there is a need for a successful image inversion method which is the topic of this paper. + +# 3. Method + +In Section 3.1, we describe the motivation for our approach. The model architecture and the training procedure are presented in Sections 3.2 and 3.3, respectively. + +# 3.1. Motivation + +Current inversion methods aim at learning an encoder to project an input image into StyleGAN's natural latent space so that the inverted image is editable. However, it is observed by previous works that when images are encoded into GAN's natural space (low-rate W or W+ space), their reconstructions suffer from low fidelity to input images, as shown in Fig. 2(b). On the other hand, if the image is encoded to higher feature maps of pretrained GANs, they will not be editable since they are not inverted to the semantic space of GANs. On the other hand, low-rate inversion is editable, as shown in Fig. 2(c). + +StyleGAN, during training and inference, relies on direct noise inputs to generate stochastic details. These additional noise inputs provide stochastic image details such as for face images, the exact placement of hairs, stubble, freckles, or skin pores which are not modeled by $\mathrm{W}+$ space. + +During inversion to $\mathrm{W + }$ space alone, that mechanism is discarded. One can tune the noise maps as well via the reconstruction loss; however, then there will be no mechanism to adapt those stochastic details to attribute manipulation. For example, the freckles should move as the pose is manipulated, but with noise optimization alone, it will not be possible. + +We propose to embed images to both low-rate and high-rate embeddings. Consistent with the design of StyleGAN, we aim at encoding the overall composition and high-level aspects of the image to $\mathrm{W + }$ space (low-rate encoding). Our goal is to embed the stochastic image details and the diverse background details, which are difficult to reconstruct from $\mathrm{W + }$ space to higher latent codes. However, this setting also requires a mechanism for encoded image details to adopt manipulations in the image. Otherwise, it will cause a ghosting effect, as shown in Fig 2(e). For example, in Fig 2(e1), the smile is removed by $\mathrm{W + }$ code edit; however, the higher-rate encodings stay the same and cause artifacts around the mouth area. In Fig. 2(e2), the pose is edited but higher detail encodings did not move and caused blur. + +In this work, we design an encoder architecture and training pipeline to achieve both learning residual features (the ones $\mathrm{W}+$ space could not reconstruct) and how to transform them to be consistent with the manipulations. + +# 3.2. StyleRes Architecture + +Our method utilizes an encoder $E_0$ that can embed images into $W+$ latent space and a StyleGAN generator $G$ . In our setup, we utilize a pretrained encoder for $E_0$ [32] and StyleGAN2 generator [22] and fix them in our training. Because it is difficult to preserve image details only from low-rate latent codes, we also extract high level features from the encoder. First, the high and low rate features are extracted as $F_0$ , $W^+ = E_0(x)$ using the encoder $E_0$ + +![](images/b3de877285e98036d1c3895e5a4e5cdf66bebd1f3469d5800068931c14ee4be6.jpg) +Figure 3. StyleRes encodes missing features for high-fidelity reconstruction of given input via the first encoder, $E_{1}$ . Those encoded features are the ones which could not be encoded to low-rate W+ space via $E_{0}$ due to the information bottleneck. Through the second encoder, $E_{2}$ , StyleRes learns to transform features based on the manipulated features. During training, latent codes are edited by interpolating encoded W+’s with randomly generated ones by StyleGAN's mapping network. During inference, they are edited with semantically meaningful directions discovered by methods such as InterfaceGAN and GANSpace. Note that StyleGAN generator is shown as two parts just for the ease of visualizing the diagram. First part includes the layers that generate features to $64 \times 64$ and the second part generates the higher resolution features and final image. + +from the input image $x$ as shown in Fig. 3. $F_{0}$ has the spatial dimension of $64 \times 64$ and provides us with more image details. Next, from $F_{0}$ , our aim is to encode residuals that are missing from the image reconstruction. For this goal, we set an encoder, $E_{1}$ , which takes $F_{0}$ as input. Since the goal of $E_{1}$ is to encode residual features, we also feed the generated features with $W^{+}$ from the StyleGAN generator, $G_{W} = G_{0\rightarrow n}(W)$ , where the arrow operator indicates the indices of convolutional layers used from $G$ . + +$$ +F _ {a} = E _ {1} \left(F _ {0}, G _ {W +}\right) \tag {1} +$$ + +With the inputs of $F_{0}$ and $G_{W+}$ , $E_{1}$ can learn what the missing features are by comparing the encoded features $F_{0}$ and generated ones, $G_{W+}$ , so far. While $E_{1}$ can learn the residual features, it is not guided on how to transform them if images are edited. For that purpose, we train $E_{2}$ , which takes $F_{a}$ and edited features from the generator. Since we do not target predefined edits (e.g., smile, pose, age), we simulate the edits by taking random directions. Specifically, we sample a $z$ vector from the normal distribution and obtain $W_{r}^{+}$ by StyleGAN's mapping network, $M$ ; $W_{r}^{+} = M(z)$ . Next, we take a step towards $W_{r}^{+}$ to obtain a mixed style code $W_{\alpha}^{+}$ . + +$$ +W _ {\alpha} ^ {+} = W ^ {+} + \alpha \frac {W _ {r} ^ {+} - W ^ {+}}{1 0} \tag {2} +$$ + +where $\alpha$ controls the degree of the edit. During the training, we assign $\alpha$ to 0 (no edit) or a value in the range of + +(4,5) with $50\%$ chance. We adjust $F_{a}$ to the altered generator features $G_{\alpha} = G_{0 \to n}(W_{\alpha})$ using a second encoder $E_{2}$ and obtain the final feature map F: + +$$ +F = E _ {2} \left(F _ {a}, G _ {\alpha}\right) \tag {3} +$$ + +Finally, $F$ and $G_{\alpha}$ are summed, and given as an input to next convolutional layers in the generator. Architectural details of $E_{1}$ and $E_{2}$ are given in Supplementary. + +# 3.3. Training Phases + +To train our model with the capabilities of high-fidelity inversion and high-quality editing, we use two training phases, as shown in Fig. 4. + +No Editing Path. In this path, we reconstruct images with no editing on encoded $W^{+}$ . This refers to the case where $\alpha = 0$ in Eq. 2. This is also the training path other inversion methods solely rely on. While this path can teach the network to reconstruct images faithfully, it does not include any edits and cannot guide the network to high-quality editability. + +Cycle Translation Path. In this path, we edit images by setting $\alpha$ a value in the range of (4, 5), which we found to work best in our ablation studies. Via the edit, the generator outputs image $x_{i}^{\prime}$ . Next, we feed this intermediate output image to the encoder and reverse the edit by inverting the addition presented in Eq. 2. The generator reconstructs $x''$ , which is supposed to match the input image $x$ . The cycle translation path is important because there is no ground-truth output for the edited image, $x_{i}^{\prime}$ . Adversarial loss can + +![](images/5332370ca3472db546b2e80c75b86e8a03b5a141fbbce49fe66d376fa4a163d1.jpg) +(a) No Editing: Reconstruction Loss + +![](images/edc70d8352bb6ec52392320332ed22397dca7b3fa788a2e951500470de8545a1.jpg) +(b) Cycle-consistency: Reconstruction Loss +Figure 4. We train StyleRes with (a) no editing based reconstruction and (b) cycle consistency based reconstruction losses. Many details from the inference pipeline are omitted for brevity. With reconstruction losses, the model learns to preserve the image details. We additionally apply adversarial losses on $x_{i}^{\prime}$ ; this way, when the image is edited, the transformation network learn to output realistic images. With cycle consistency based reconstruction, the network is regularized to keep the input details also during edits. We use additional losses as explained in Section 3.4. + +guide $x_{i}^{\prime}$ to look realistic but will not guide the network to keep the input image details if edited. + +# 3.4. Training Objectives + +Reconstruction Losses. For both no editing and cycle consistency path outputs, our goal is to reconstruct the input image. To supervise this behavior, we use $L_{2}$ loss, perceptual loss, and identity loss between the input and output images. The first reconstruction loss is as follows, where $x$ is the input image, $x'$ and $x''$ are output images, as given in Fig. 4: + +$$ +\mathcal {L} _ {r e c - l 2} = \left| \left| x ^ {\prime} - x \right| \right| _ {2} + \left| \left| x ^ {\prime \prime} - x \right| \right| _ {2} \tag {4} +$$ + +We use perceptual losses from VGG $(\Phi)$ at different feature layers $(j)$ between these images from the loss objective as given in Eq. 5. + +$$ +\mathcal {L} _ {r e c - p} = \left| \left| \Phi_ {j} \left(x ^ {\prime}\right) - \Phi_ {j} (x) \right| \right| _ {2} + \left| \left| \Phi_ {j} \left(x ^ {\prime \prime}\right) - \Phi_ {j} (x) \right| \right| _ {2} \tag {5} +$$ + +Identity loss is calculated with a pre-trained network $A$ . $A$ is an ArcFace model [11] when training on the face domain and a domain specific ResNet-50 model [32] for our training on car class. + +$$ +L _ {r e c - i d} = (1 - \langle A (x), A \left(x ^ {\prime}\right) \rangle) + (1 - \langle A (x), A \left(x ^ {\prime \prime}\right) \rangle) \tag {6} +$$ + +Adversarial Losses. We also use adversarial losses to guide the network to output images that match the real image distribution. This objective improves realistic image inversions and edits. We load the pretrained discriminator from StyleGAN training, $D$ , and train the discriminator together with the encoders. + +$$ +\begin{array}{l} L _ {a d v} = 2 \log D (x) + \log \left(1 - D \left(x ^ {\prime}\right)\right) \tag {7} \\ + \log \left(1 - D \left(x _ {i} ^ {\prime}\right)\right) \\ \end{array} +$$ + +Feature Regularizer. To prevent our encoder from deviating much from the original StyleGAN space, we regularize the residual features to be small: + +$$ +L _ {F} = \sum_ {F \in \phi} \| F \| _ {2} \tag {8} +$$ + +Full Objective. We use the overall objectives given below. The hyper parameters are provided in Supplementary. + +$$ +\begin{array}{l} \min _ {E _ {1}, E _ {2}} \max _ {D} \lambda_ {a} \mathcal {L} _ {a d v} + \lambda_ {r 1} \mathcal {L} _ {r e c - l 2} + \lambda_ {r 2} \mathcal {L} _ {r e c - p} \tag {9} \\ + \lambda_ {r 3} \mathcal {L} _ {r e c - i d} + \lambda_ {f} \mathcal {L} _ {F} \\ \end{array} +$$ + +# 4. Experiments + +Set-up. For datasets and attribute editing, we follow the previous work [34]. For the human face domain, we train the model on the FFHQ [21] dataset and evaluate it on the CelebA-HQ [20] dataset. For the car domain, we use Stanford Cars [23] for training and evaluation. We run extensive experiments with directions explored with InterfaceGAN [30], GANSpace [15], StyleClip [26], and GradCtrl [7] methods. + +Evaluation. We report metrics for reconstruction and editing qualities. For reconstruction, we report Frechet Inception Distance (FID) metric [16], which looks at the realism by comparing the target image distribution and reconstructed images, Learned Perceptual Image Patch Similarity (LPIPS) [44] and Structural Similarity Index Measure (SSIM), which compares the target and output pairs at the feature level of a pretrained deep network and pixel level, respectively. Additionally, FIDs are calculated for adding and removing the smile attributes on the CelebAHQ dataset. By using the ground-truth attributes, we add smile to images that do not have smile attribute and calculate FIDs between smile addition edited images and smiling + +![](images/97bf85cbdf4c310efa39a4b361107c0245a5f88bc739cec5115373316276cc71.jpg) +Figure 5. Qualitative results of inversion and editing. For each method, the first column shows inversion, and the second shows InterfaceGAN [30] and GANSpace [15] edits. + +ground-truth images. The same set-up is used for smile removal. On the Stanford cars dataset, we change grass and color attribute of images, and since there is no ground-truth attribute, we calculate the FIDs between the edited and original images. + +Baselines. We compare our method with state-of-the-art image inversion methods pSp [27], e4e [32], ReStyle [4], HyperStyle [5], HFGI [34], StyleTransformer [18], and FeatureStyle [40]. We use the author's released models. Hence, for the Stanford car dataset, some methods are omitted from comparisons if the models are not released. Among those, we only train HFGI for car model with au + +thor's released code since we base our main comparisons with it. + +Quantitative Results. In Table 1, we provide reconstruction and editing scores on the CelebA-HQ dataset. Our method achieves better results than all competing methods on all metrics. Most significantly, we achieve better FID scores on both reconstruction and editing qualities. While FeatureStyle achieves comparable SSIM and LPIPS scores on reconstruction, the editing FIDs are worse than our model. As shown in Table 2, we achieve significantly better results than previous methods on the Stanford Car dataset as well. We also compare the runtime of our method + +Table 1. Quantitative results of reconstruction and editing on CelebA-HQ dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for smile addition (+) and removal (-). + +
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSSmile(+)Smile(-)
pSp [27]23.860.750.1732.4734.0
e4e [32]30.220.710.2138.5839.68
ReStyle [4]24.820.730.2030.3533.69
HyperStyle [5]16.080.830.1126.4325.26
HFGI [34]12.170.850.1325.2227.10
StyleTransformer [18]21.820.750.1734.3234.61
FeatureStyle [40]11.330.900.1027.2026.15
StyleRes (Ours)7.040.900.0923.5221.80
+ +Table 2. Quantitative results of reconstruction and editing on the Stanford Cars Dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for grass addition and color change. + +
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSGrassColor
e4e [32]14.040.500.3218.0229.79
ReStyle [4]13.380.570.3016.0121.34
HyperStyle [5]11.640.630.2817.1326.30
HFGI [34]9.410.830.1614.8426.65
StyleTransformer [18]14.010.570.2819.4719.94
StyleRes (Ours)7.600.830.1410.6418.86
+ +and previous methods. Our method is significantly faster than HyperStyle (0.125 sec vs. 0.439 sec). That is because HyperStyle refines its predicted weight offsets gradually via multiple iterations. Our method is also faster than HFGI (0.125 sec vs. 0.130 sec), in addition to achieving better scores. We run a single stage network inference, whereas HFGI first generates an image via e4e and StyleGAN and provides the error map to a second architecture. The table for runtimes is provided in Supplementary. + +Qualitative Results. We show visuals of inversion and editing results of our method, e4e, HFGI, and HyperStyle in Fig. 5. We provide further comparisons in Supplementary with other attribute editorings and with other methods. Compared to previous methods, our method achieves significantly better fidelity to the input images and preserves the identity and details when edited. It is the only method in these comparisons that preserves the background, earings (second row), hands (second-fourth rows), and hats (fourth row). Our method also achieves facial detail reconstruction; for example, in the fifth row, the person has a mole at the corner of her mouth. Among the inversions, our method is the only one preserving that. Furthermore, during the pose edit, it is transformed correctly. On the sixth row, our method is the only one that achieves the correct inversion and edit. In car examples, we again achieve high fidelity to the input images both in inversion and editing. e4e and HyperStyle do not reconstruct the image faithfully. On the + +![](images/323b7dea6ed040bf91e1d3464a5b225612ebe14f8e41b0d67629629342d9445f.jpg) +Figure 6. Comparison of our method with e4e, HFGI, and HyperStyle with StyleClip edits [26]. The first column shows eyeglasses addition, the second and third columns show bangs addition, and the last column shows bob cut hairstyle results. + +other hand, HFGI outputs artefacts during edits. We additionally show results with edits explored by StyleClip [26] and GradCtrl [7] methods in Figs. 6 and 7, respectively. + +Ablation Study. We run extensive experiments to validate the role of each proposed design choices. We first experiment with an architecture that does not have $E_{1}$ and $E_{2}$ modules. This experiment refers to the case where we learn a network that does not take input from StyleGAN Part 1 features, neither the original $(G_{W+})$ nor the edited features $(G_{\alpha})$ . The network directly takes higher layer features $(F_{0})$ from the encoder and outputs them to the generator. As shown in Fig. 8, the network still tries to learn residual features to achieve reconstructions; however, it achieves poor edits. Without $E_{2}$ refers to the experiment with $E_{1}$ directly outputting features to Generator Part 2 (as shown in Fig. 3). Without $E_{1}$ refers to the experiment where $E_{2}$ directly takes input from the encoder $(F_{0})$ and $(G_{\alpha})$ as defined in Sec. 3.2. None of the networks are able to learn the residual features and how to transform them as well as our final architecture since they are not designed to take the original and edited features separately. We also experiment without cycle consistency losses. Fig. 9 shows visual results of methods trained with and without cycle consistency constrain. We + +![](images/60fcc17268c68bc0e499e97a161b701fcfd8ac9c00e101ceda5e5bb5910b8bde.jpg) +Figure 7. Additional results of our method and e4e with GradCtrl edits [7]. The first three rows show blue sky edit, and the others show tree background edits. + +observe that with cycle consistency constrain, our network achieves preserving fine image details even when images are edited. + +# 5. Conclusion + +We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We show state-of-the-art results on a wide range of edits explored with different methods both quantitatively and qualitatively while achieving faster run-time than competing methods. + +![](images/c9b2b4eb28f870563d8a3cdf00dd08fcade7106beee3157f8693e5cbf708d223.jpg) +Figure 8. Ablation Study. Pose edit outputs of our final architecture and architecture with missing modules. w/o $E_{1}$ or $E_{2}$ , networks struggle to transform features correctly. + +![](images/07bec594a38d21401644b96d3794980a1c29c83f59a7c3a3f20843ee26b28bc2.jpg) +Figure 9. Ablation Study. Pose edit outputs of models trained with and without cycle consistency loss. The model trained with cycle consistency is better at preserving the details as shown in enlarged boxes. + +# Acknowledgement + +This work has been funded by The Scientific and Technological Research Council of Turkey (TUBITAK), 3501 Research Project under Grant No 121E097. A. Dundar was supported by Marie Skłodowska-Curie Individual Fellowship. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2 +[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2, 3 +[4] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 6, 7 +[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 1, 2, 6, 7 +[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2 +[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In European Conference on Computer Vision, pages 104-119. Springer, 2022. 5, 7, 8 +[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[9] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. IEEE transactions on neural networks and learning systems, 30(7):1967-1974, 2018. 2 +[10] Yusuf Dalva, Said Fahri Altindis, and Aysegul Dundar. Vecgan: Image-to-image translation with interpretable latent directions. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 2 +[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5 +[12] Aysegul Dundar, Karan Sapra, Guilin Liu, Andrew Tao, and Bryan Catanzaro. Panoptic-based image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8070-8079, 2020. 2 +[13] Yue Gao, Fangyun Wei, Jianmin Bao, Shuyang Gu, Dong Chen, Fang Wen, and Zhouhui Lian. High-fidelity and arbitrary face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16115-16124, 2021. 2 + +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33, 2020. 1, 3, 5, 6 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[17] Xianxu Hou, Xiaokang Zhang, Hanbang Liang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. Neural Networks, 145:209-220, 2022. 2 +[18] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11337-11346, June 2022. 6, 7 +[19] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. Eur. Conf. Comput. Vis., 2018. 2 +[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 5 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2, 5 +[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 3 +[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5 +[24] Xinyang Li, Shengchuan Zhang, Jie Hu, Liujuan Cao, Xiaopeng Hong, Xudong Mao, Feiyue Huang, Yongjian Wu, and Rongrong Ji. Image-to-image translation via hierarchical style disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8639-8648, 2021. 2 +[25] Guilin Liu, Aysegul Dundar, Kevin J Shih, Ting-Chun Wang, Fitsum A Reda, Karan Sapra, Zhiding Yu, Xiaodong Yang, Andrew Tao, and Bryan Catanzaro. Partial convolution for padding, inpainting, and image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 5, 7 + +[27] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 1, 2, 6, 7 +[28] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 2 +[29] Wei Shen and Rujie Liu. Learning residual images for face attribute manipulation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4030-4038, 2017. 2 +[30] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 3, 5, 6 +[31] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 3 +[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 1, 2, 3, 5, 6, 7, 8 +[33] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International Conference on Machine Learning, pages 9786-9796. PMLR, 2020. 1, 3 +[34] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 1, 2, 5, 6, 7 +[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2 +[36] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5914–5922, 2019. 2 +[37] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 3 +[38] Taihong Xiao, Jiapeng Hong, and Jinwen Ma. Elegant: Exchanging latent encodings with gan for transferring multiple face attributes. In Proceedings of the European conference on computer vision (ECCV), pages 168-184, 2018. 2 +[39] Guoxing Yang, Nanyi Fei, Mingyu Ding, Guangzhen Liu, Zhiwu Lu, and Tao Xiang. L2m-gan: Learning to manipulate latent space semantics for facial attribute editing. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2951-2960, 2021. 2 +[40] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 6, 7 +[41] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 1 +[42] Gang Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Generative adversarial network with spatial attention for face attribute editing. In Proceedings of the European conference on computer vision (ECCV), pages 417-432, 2018. 2 +[43] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR, 2019. 1 +[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[45] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2 +[46] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A Efros, Oliver Wang, and Eli Shechtman. Multimodal image-to-image translation by enforcing bi-cycle consistency. In Advances in neural information processing systems, pages 465-476, 2017. 2 \ No newline at end of file diff --git a/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/images.zip b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f010e14ea3e7a55423b79edbc104e495e888ccf9 --- /dev/null +++ b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49ac23ad857c1b276789051b86a63e46a9d6a5d5bd99cd5610a057e5715347d0 +size 1011683 diff --git a/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/layout.json b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5872ae8dbdb350b9663c29d296fcf7bc361f2805 --- /dev/null +++ b/2023/StyleRes_ Transforming the Residuals for Real Image Editing With StyleGAN/layout.json @@ -0,0 +1,8231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 536, + 122 + ], + "type": "text", + "content": "StyleRes: Transforming the Residuals for Real Image Editing with StyleGAN" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 143, + 433, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 143, + 433, + 172 + ], + "spans": [ + { + "bbox": [ + 160, + 143, + 433, + 172 + ], + "type": "text", + "content": "Hamza Pehlivan Yusuf Dalva Aysegul Dundar Bilkent University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 174, + 412, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 174, + 412, + 185 + ], + "spans": [ + { + "bbox": [ + 181, + 174, + 412, + 185 + ], + "type": "text", + "content": "{hamza.pehlivan,yusuf.dalva}@bilkent.edu.tr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 228, + 188, + 365, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 188, + 365, + 198 + ], + "spans": [ + { + "bbox": [ + 228, + 188, + 365, + 198 + ], + "type": "text", + "content": "adundar@cs.bilkent.edu.tr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 251, + 290, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 503 + ], + "type": "text", + "content": "We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. Inverting real images into StyleGAN's latent space is an extensively studied problem, yet the trade-off between the image reconstruction fidelity and image editing quality remains an open challenge. The low-rate latent spaces are limited in their expressiveness power for high-fidelity reconstruction. On the other hand, high-rate latent spaces result in degradation in editing quality. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We train the framework to extract residual features and transform them via a novel architecture pipeline and cycle consistency losses. We run extensive experiments and compare our method with state-of-the-art inversion methods. Qualitative metrics and visual comparisons show significant improvements. Code: https://github.com/hamzapehlivan/StyleRes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Generative Adversarial Networks (GANs) achieve high quality synthesis of various objects that are hard to distinguish from real images [14,21,22,41,43]. These networks also have an important property that they organize their latent space in a semantically meaningful way; as such, via latent editing, one can manipulate an attribute of a generated image. This property makes GANs a promising technology for image attribute editing and not only for generated images but also for real images. However, for real images, one also needs to find the corresponding latent code that will generate the particular real image. For this purpose, different GAN inversion methods are proposed, aiming to project real images to pretrained GAN latent space [15,30,31,33,37]. Even though this is an extensively stud" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 234, + 555, + 518 + ], + "blocks": [ + { + "bbox": [ + 310, + 234, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 310, + 234, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 310, + 234, + 555, + 518 + ], + "type": "image", + "image_path": "b6a110052faad7df5ba714c0095b57d515f516b8c8916a12ecbf6a57ea694544.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 528, + 547, + 572 + ], + "lines": [ + { + "bbox": [ + 305, + 528, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 547, + 572 + ], + "type": "text", + "content": "Figure 1. Comparison of our method with e4e, HFGI, and HyperStyle for the pose, bob cut hairstyle, smile removal, and color change edits. Our method achieves high fidelity to the input and high quality edits." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 546, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 629 + ], + "type": "text", + "content": "ied problem with significant progress, the trade-off between image reconstruction fidelity and image editing quality remains an open challenge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "The trade-off between image reconstruction fidelity and image editing quality is referred to as the distortion-editability trade-off [32]. Both are essential for real image editing. However, it is shown that the low-rate latent spaces are limited in their expressiveness power, and not every image can be inverted with high fidelity reconstruction [1, 27, 32, 34]. For that reason, higher bit encodings" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1828" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "and more expressive style spaces are explored for image inversion [1, 2]. Although with these techniques, images can be reconstructed with better fidelity, the editing quality decreases since there is no guarantee that projected codes will naturally lie in the generator's latent manifold." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 455 + ], + "type": "text", + "content": "In this work, we propose a framework that achieves high fidelity input reconstruction and significantly improved editability compared to the state-of-the-art. We learn residual features in higher-rate latent codes that are missing in the reconstruction of encoded features. This enables us to reconstruct image details and background information which are difficult to reconstruct via low rate latent encodings. Our architecture is single stage and learns the residuals based on the encoded features from the encoder and generated features from the pretrained GANs. We also learn a module to transform the higher-latent codes if needed based on the generated features (e.g. when the low-rate latent codes are manipulated). This way, when low-rate latent codes are edited for attribute manipulation, the decoded features can adapt to the edits to reconstruct details. While the attributes are not edited, the encoder can be trained with image reconstruction and adversarial losses. On the other hand, when the image is edited, we cannot use image reconstruction loss to regularize the network to preserve the details. To guide the network to learn correct transformations based on the generated features, we train the model with adversarial loss and cycle consistency constraint; that is, after we edit the latent code and generate an image, we reverse the edit and aim at reconstructing the original image. Since we do not want our method to be limited to predefined edits, during training, we simulate edits by randomly interpolating them with sampled latent codes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 552 + ], + "type": "text", + "content": "The closest to our approach is HFGI [34], which also learns higher-rate encodings. Our framework is different as we learn a single-stage architecture designed to learn features that are missing from low-rate encodings and we learn how to transform them based on edits. As shown in Fig. 1, our framework achieves significantly better results than HFGI and other methods in editing quality. In summary, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 561, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 561, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 287, + 597 + ], + "type": "text", + "content": "- We propose a single-stage framework that achieves high-fidelity input embedding and editing. Our framework achieves that with a novel encoder architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 607, + 287, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 287, + 680 + ], + "type": "text", + "content": "- We propose to guide the image projection with cycle consistency and adversarial losses. We edit encoded images by taking a step toward a randomly sampled latent code. We expect to reconstruct the original image when the edit is reversed. This way, edited images preserve the details of the input image, and edits become high quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We conduct extensive experiments to show the effectiveness of our framework and achieve significant improve" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "type": "text", + "content": "ments over state-of-the-art for both reconstruction and real image attribute manipulations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 107, + 392, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 107, + 392, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 107, + 392, + 119 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 127, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 547, + 558 + ], + "type": "text", + "content": "GAN Inversion. GAN inversion methods aim at projecting a real image into GANs embedding so that from the embedding, GAN can generate the given image. Currently, the biggest motivation of the inversion is the ability to edit the image via semantically rich disentangled features of GANs; therefore models aim for high reconstruction and editing quality. Inversion methods can be categorized into two; 1) methods that directly optimize the latent vector to minimize the reconstruction error between the output and target image [1,2,9,22,28], 2) methods that learn encoders to reconstruct images over training images [32,34,45]. Optimization based methods require per-image optimization and iterative passes on GANs, which take several minutes per image. Additionally, overfitting on a single image results in latent codes that do not lie in GAN's natural latent distribution, leading to poor editing quality. Among these methods, PTI [28] takes a different approach, and instead of searching for a latent code that will reconstruct the image most accurately, PTI fine-tunes the generator in order to insert encoded latent code into well-behaved regions of the latent space. It shows better editability; however, the method still suffers from long run-time optimizations and tunings. Encoder based methods leverage the knowledge of training sets while projecting images. They output results with less fidelity to the input, but edits on the projected latents are better quality. They also operate mostly in real-time. They either project the latent code with a single feed-forward pass [27,32,45], or two stage feed-forward passes where the first encoder learns to embed an image, the second learns to reconstruct the missing details [34], or iterative forward passes where the network tries to minimize the reconstruction loss at each pass by taking the original image and reconstructed output as inputs [4,5]. In this work, we propose a novel encoder architecture that achieves significantly better results than state-of-the-art via a single feed-forward pass on the input image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Image Translation. There is quite an interest in image translation algorithms that can change an attribute of an image while preserving a given content [12, 25, 35], especially for faces editing [3, 8, 10, 13, 17, 24, 29, 36, 38, 42]. These algorithms set an encoder-decoder architecture and train the models with reconstruction and GAN losses [8, 19, 24, 39, 46]. Even though they are successful at image manipulation, they are trained for a given task and rely on predefined translation datasets. On the other hand, it is shown that GANs that are trained to synthesize objects in an unconditional way can be used for attribute manipulation [6, 21, 22] via their semantically rich disentangled features. This makes pretrained GANs promising technol-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1829" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 115, + 125 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 115, + 125 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 115, + 125 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 115, + 125 + ], + "type": "image", + "image_path": "7d12b3a79f2e1d372a601a2a9b5437bbdf542e752b780479378e718f2567b974.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 124, + 96, + 130 + ], + "lines": [ + { + "bbox": [ + 74, + 124, + 96, + 130 + ], + "spans": [ + { + "bbox": [ + 74, + 124, + 96, + 130 + ], + "type": "text", + "content": "(a1) Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 117, + 70, + 171, + 125 + ], + "blocks": [ + { + "bbox": [ + 117, + 70, + 171, + 125 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 171, + 125 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 171, + 125 + ], + "type": "image", + "image_path": "e90e53b430e865f52e850830a6770d5c91149fb4a94c272f2a2f3fba976c720f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 125, + 169, + 137 + ], + "lines": [ + { + "bbox": [ + 118, + 125, + 169, + 137 + ], + "spans": [ + { + "bbox": [ + 118, + 125, + 169, + 137 + ], + "type": "text", + "content": "(b1) Low-rate Inversion (e4e)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 172, + 70, + 227, + 125 + ], + "blocks": [ + { + "bbox": [ + 172, + 70, + 227, + 125 + ], + "lines": [ + { + "bbox": [ + 172, + 70, + 227, + 125 + ], + "spans": [ + { + "bbox": [ + 172, + 70, + 227, + 125 + ], + "type": "image", + "image_path": "64d5da3d2c3d32ae117b8b30ff23edfbacb02d053761e744b20d9ab7ff6ba50d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 125, + 218, + 137 + ], + "lines": [ + { + "bbox": [ + 179, + 125, + 218, + 137 + ], + "spans": [ + { + "bbox": [ + 179, + 125, + 218, + 137 + ], + "type": "text", + "content": "(c1) Low-rate Edit (e4e)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 282, + 125 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 282, + 125 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 282, + 125 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 282, + 125 + ], + "type": "image", + "image_path": "5f7508805408e03985480a1ccdd29648729d411517e28c06bc28396b48289a1b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 125, + 282, + 137 + ], + "lines": [ + { + "bbox": [ + 230, + 125, + 282, + 137 + ], + "spans": [ + { + "bbox": [ + 230, + 125, + 282, + 137 + ], + "type": "text", + "content": "(d1) Low-rate Inversion High-rate Inversion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 61, + 137, + 171, + 191 + ], + "blocks": [ + { + "bbox": [ + 61, + 137, + 171, + 191 + ], + "lines": [ + { + "bbox": [ + 61, + 137, + 171, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 171, + 191 + ], + "type": "image", + "image_path": "75eb5812a2381097f15a8345007b2d4bb12e17377e6e23ff6a6babff13e0e523.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 192, + 160, + 205 + ], + "lines": [ + { + "bbox": [ + 118, + 192, + 160, + 205 + ], + "spans": [ + { + "bbox": [ + 118, + 192, + 160, + 205 + ], + "type": "text", + "content": "(1) Low-rate Edit High-rate Inversion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 171, + 137, + 227, + 192 + ], + "blocks": [ + { + "bbox": [ + 171, + 137, + 227, + 192 + ], + "lines": [ + { + "bbox": [ + 171, + 137, + 227, + 192 + ], + "spans": [ + { + "bbox": [ + 171, + 137, + 227, + 192 + ], + "type": "image", + "image_path": "0415f2b0cfb09985f92a6cbe2afabbd9542faef8ab6a3e374623e051ead97ce5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 192, + 227, + 205 + ], + "lines": [ + { + "bbox": [ + 172, + 192, + 227, + 205 + ], + "spans": [ + { + "bbox": [ + 172, + 192, + 227, + 205 + ], + "type": "text", + "content": "(1) Low-rate Edit High-rate Transformation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 228, + 162, + 284, + 191 + ], + "blocks": [ + { + "bbox": [ + 228, + 162, + 284, + 191 + ], + "lines": [ + { + "bbox": [ + 228, + 162, + 284, + 191 + ], + "spans": [ + { + "bbox": [ + 228, + 162, + 284, + 191 + ], + "type": "image", + "image_path": "49ef69017312bf2abdbf04d01381c1c2cb8cc7c74b4513e5631e1a03b726315f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 364, + 125 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 364, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 364, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 364, + 125 + ], + "type": "image", + "image_path": "85685a9346cdd43ea4940b70bc302c9da019708b51777694bf8db148a0e12af1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 124, + 343, + 129 + ], + "lines": [ + { + "bbox": [ + 321, + 124, + 343, + 129 + ], + "spans": [ + { + "bbox": [ + 321, + 124, + 343, + 129 + ], + "type": "text", + "content": "(a2) Input" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 71, + 419, + 125 + ], + "blocks": [ + { + "bbox": [ + 365, + 71, + 419, + 125 + ], + "lines": [ + { + "bbox": [ + 365, + 71, + 419, + 125 + ], + "spans": [ + { + "bbox": [ + 365, + 71, + 419, + 125 + ], + "type": "image", + "image_path": "3ba937bcb007e3cbd1586cb384e5ac6bde53fb5cba5df00935f2d6bbc62743aa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 125, + 416, + 137 + ], + "lines": [ + { + "bbox": [ + 365, + 125, + 416, + 137 + ], + "spans": [ + { + "bbox": [ + 365, + 125, + 416, + 137 + ], + "type": "text", + "content": "(b3) Low-rate Inversion (e4e)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 420, + 71, + 474, + 125 + ], + "blocks": [ + { + "bbox": [ + 420, + 71, + 474, + 125 + ], + "lines": [ + { + "bbox": [ + 420, + 71, + 474, + 125 + ], + "spans": [ + { + "bbox": [ + 420, + 71, + 474, + 125 + ], + "type": "image", + "image_path": "14fea19cdefd9de980c839b807c52745dc3fcaabd3b9d6dfd63903c8239fa957.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 125, + 464, + 137 + ], + "lines": [ + { + "bbox": [ + 426, + 125, + 464, + 137 + ], + "spans": [ + { + "bbox": [ + 426, + 125, + 464, + 137 + ], + "type": "text", + "content": "(c2) Low-rate Edit (e4e)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 476, + 70, + 531, + 125 + ], + "blocks": [ + { + "bbox": [ + 476, + 70, + 531, + 125 + ], + "lines": [ + { + "bbox": [ + 476, + 70, + 531, + 125 + ], + "spans": [ + { + "bbox": [ + 476, + 70, + 531, + 125 + ], + "type": "image", + "image_path": "d2eb7cd88e13b3cd2610f5cca8db0676538c2d3e341dcdae278c5738662a3562.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 125, + 530, + 137 + ], + "lines": [ + { + "bbox": [ + 478, + 125, + 530, + 137 + ], + "spans": [ + { + "bbox": [ + 478, + 125, + 530, + 137 + ], + "type": "text", + "content": "(d2) Low-rate Inversion High-rate Inversion" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 323, + 137, + 357, + 193 + ], + "blocks": [ + { + "bbox": [ + 323, + 137, + 357, + 193 + ], + "lines": [ + { + "bbox": [ + 323, + 137, + 357, + 193 + ], + "spans": [ + { + "bbox": [ + 323, + 137, + 357, + 193 + ], + "type": "image", + "image_path": "e95e929c5586ac277dae561c506543f2693c56bb4d7c89b24602bad2721e8140.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 547, + 269 + ], + "type": "text", + "content": "Figure 2. When images are encoded to " + }, + { + "bbox": [ + 46, + 214, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 214, + 547, + 269 + ], + "type": "text", + "content": " space, as shown in (b), the reconstructions miss many image details. However, the edits are in good quality (c). Additional features can be learned in higher-rate latent codes. For example, skip connections from encoder to generator in higher resolution features can enable high-fidelity reconstruction to input images as shown in (d). However, if the high-rate features do not transform with the edits, they result in ghosting effects as shown in (e). In this work, we propose high-rate encoding and transformation for successful inversions and edits (f)." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 365, + 137, + 419, + 192 + ], + "blocks": [ + { + "bbox": [ + 365, + 137, + 419, + 192 + ], + "lines": [ + { + "bbox": [ + 365, + 137, + 419, + 192 + ], + "spans": [ + { + "bbox": [ + 365, + 137, + 419, + 192 + ], + "type": "image", + "image_path": "fd886fff90b9f737e70ab6a673353f738bf2bfd857bf94ecd707d80dddff59ac.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 192, + 406, + 205 + ], + "lines": [ + { + "bbox": [ + 365, + 192, + 406, + 205 + ], + "spans": [ + { + "bbox": [ + 365, + 192, + 406, + 205 + ], + "type": "text", + "content": "62)Low-rate Edit High-rate Inversion" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 192, + 474, + 205 + ], + "lines": [ + { + "bbox": [ + 419, + 192, + 474, + 205 + ], + "spans": [ + { + "bbox": [ + 419, + 192, + 474, + 205 + ], + "type": "text", + "content": "(f2) Low-rate Edit High-rate Transformation" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 420, + 137, + 475, + 192 + ], + "blocks": [ + { + "bbox": [ + 420, + 137, + 475, + 192 + ], + "lines": [ + { + "bbox": [ + 420, + 137, + 475, + 192 + ], + "spans": [ + { + "bbox": [ + 420, + 137, + 475, + 192 + ], + "type": "image", + "image_path": "7df925b8e6d83d187cd5a93187f4c93968d11d74ff1df7d42317835a3eb69acb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 482, + 139, + 517, + 193 + ], + "blocks": [ + { + "bbox": [ + 482, + 139, + 517, + 193 + ], + "lines": [ + { + "bbox": [ + 482, + 139, + 517, + 193 + ], + "spans": [ + { + "bbox": [ + 482, + 139, + 517, + 193 + ], + "type": "image", + "image_path": "2f424ffd6f3e78a0d3f6725991688d26ff06ef70235ad99bf01e59f965cef20e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 291, + 289, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 423 + ], + "type": "text", + "content": "ogy for image editing, given that via their ability of image generation, they can achieve many downstream tasks simultaneously. Many methods have been proposed for finding latent directions to edit images [15, 30, 31, 33, 37]. These directions are explored both in supervised [3, 30] and unsupervised ways [15, 31, 33, 37] resulting in exploration of attribute manipulations beyond the predefined attributes of labeled datasets. However, finding these directions is only one part of the problem; to achieve real image editing, there is a need for a successful image inversion method which is the topic of this paper." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 435, + 103, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 103, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 103, + 448 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 456, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 493 + ], + "type": "text", + "content": "In Section 3.1, we describe the motivation for our approach. The model architecture and the training procedure are presented in Sections 3.2 and 3.3, respectively." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 47, + 502, + 121, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 121, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 121, + 514 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": "Current inversion methods aim at learning an encoder to project an input image into StyleGAN's natural latent space so that the inverted image is editable. However, it is observed by previous works that when images are encoded into GAN's natural space (low-rate W or W+ space), their reconstructions suffer from low fidelity to input images, as shown in Fig. 2(b). On the other hand, if the image is encoded to higher feature maps of pretrained GANs, they will not be editable since they are not inverted to the semantic space of GANs. On the other hand, low-rate inversion is editable, as shown in Fig. 2(c)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "StyleGAN, during training and inference, relies on direct noise inputs to generate stochastic details. These additional noise inputs provide stochastic image details such as for face images, the exact placement of hairs, stubble, freckles, or skin pores which are not modeled by " + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{W}+" + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": " space." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "content": "During inversion to " + }, + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 291, + 545, + 374 + ], + "type": "text", + "content": " space alone, that mechanism is discarded. One can tune the noise maps as well via the reconstruction loss; however, then there will be no mechanism to adapt those stochastic details to attribute manipulation. For example, the freckles should move as the pose is manipulated, but with noise optimization alone, it will not be possible." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "content": "We propose to embed images to both low-rate and high-rate embeddings. Consistent with the design of StyleGAN, we aim at encoding the overall composition and high-level aspects of the image to " + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "content": " space (low-rate encoding). Our goal is to embed the stochastic image details and the diverse background details, which are difficult to reconstruct from " + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "content": " space to higher latent codes. However, this setting also requires a mechanism for encoded image details to adopt manipulations in the image. Otherwise, it will cause a ghosting effect, as shown in Fig 2(e). For example, in Fig 2(e1), the smile is removed by " + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 304, + 375, + 546, + 542 + ], + "type": "text", + "content": " code edit; however, the higher-rate encodings stay the same and cause artifacts around the mouth area. In Fig. 2(e2), the pose is edited but higher detail encodings did not move and caused blur." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 543, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 591 + ], + "type": "text", + "content": "In this work, we design an encoder architecture and training pipeline to achieve both learning residual features (the ones " + }, + { + "bbox": [ + 304, + 543, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\mathrm{W}+" + }, + { + "bbox": [ + 304, + 543, + 545, + 591 + ], + "type": "text", + "content": " space could not reconstruct) and how to transform them to be consistent with the manipulations." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 599, + 430, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 430, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 430, + 611 + ], + "type": "text", + "content": "3.2. StyleRes Architecture" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Our method utilizes an encoder " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " that can embed images into " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "W+" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " latent space and a StyleGAN generator " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ". In our setup, we utilize a pretrained encoder for " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " [32] and StyleGAN2 generator [22] and fix them in our training. Because it is difficult to preserve image details only from low-rate latent codes, we also extract high level features from the encoder. First, the high and low rate features are extracted as " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "W^+ = E_0(x)" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " using the encoder " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "E_0" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1830" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 73, + 547, + 257 + ], + "blocks": [ + { + "bbox": [ + 47, + 73, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 73, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 547, + 257 + ], + "type": "image", + "image_path": "b3de877285e98036d1c3895e5a4e5cdf66bebd1f3469d5800068931c14ee4be6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "text", + "content": "Figure 3. StyleRes encodes missing features for high-fidelity reconstruction of given input via the first encoder, " + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "text", + "content": ". Those encoded features are the ones which could not be encoded to low-rate W+ space via " + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "inline_equation", + "content": "E_{0}" + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "text", + "content": " due to the information bottleneck. Through the second encoder, " + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "text", + "content": ", StyleRes learns to transform features based on the manipulated features. During training, latent codes are edited by interpolating encoded W+’s with randomly generated ones by StyleGAN's mapping network. During inference, they are edited with semantically meaningful directions discovered by methods such as InterfaceGAN and GANSpace. Note that StyleGAN generator is shown as two parts just for the ease of visualizing the diagram. First part includes the layers that generate features to " + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 264, + 547, + 344 + ], + "type": "text", + "content": " and the second part generates the higher resolution features and final image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": "from the input image " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": " as shown in Fig. 3. " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": " has the spatial dimension of " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": " and provides us with more image details. Next, from " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": ", our aim is to encode residuals that are missing from the image reconstruction. For this goal, we set an encoder, " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": ", which takes " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": " as input. Since the goal of " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": " is to encode residual features, we also feed the generated features with " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": " from the StyleGAN generator, " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "G_{W} = G_{0\\rightarrow n}(W)" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": ", where the arrow operator indicates the indices of convolutional layers used from " + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 362, + 289, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 484, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 484, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 124, + 484, + 287, + 498 + ], + "type": "interline_equation", + "content": "F _ {a} = E _ {1} \\left(F _ {0}, G _ {W +}\\right) \\tag {1}", + "image_path": "fa864db759bdaf9f6d9dbf2fa1aaeee39c42b5f3822b12db713675bf593fabd5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": "With the inputs of " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "G_{W+}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " can learn what the missing features are by comparing the encoded features " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " and generated ones, " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "G_{W+}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": ", so far. While " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " can learn the residual features, it is not guided on how to transform them if images are edited. For that purpose, we train " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": ", which takes " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " and edited features from the generator. Since we do not target predefined edits (e.g., smile, pose, age), we simulate the edits by taking random directions. Specifically, we sample a " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " vector from the normal distribution and obtain " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "W_{r}^{+}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " by StyleGAN's mapping network, " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "W_{r}^{+} = M(z)" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": ". Next, we take a step towards " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "W_{r}^{+}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": " to obtain a mixed style code " + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "inline_equation", + "content": "W_{\\alpha}^{+}" + }, + { + "bbox": [ + 46, + 505, + 289, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 658, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 658, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 107, + 658, + 287, + 685 + ], + "type": "interline_equation", + "content": "W _ {\\alpha} ^ {+} = W ^ {+} + \\alpha \\frac {W _ {r} ^ {+} - W ^ {+}}{1 0} \\tag {2}", + "image_path": "9dc10e660daeac59e279c79a20cbed192a0a7333980ac4569b0d27b6aeee9584.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": " controls the degree of the edit. During the training, we assign " + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": " to 0 (no edit) or a value in the range of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "text", + "content": "(4,5) with " + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "text", + "content": " chance. We adjust " + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "text", + "content": " to the altered generator features " + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "inline_equation", + "content": "G_{\\alpha} = G_{0 \\to n}(W_{\\alpha})" + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "text", + "content": " using a second encoder " + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 305, + 362, + 545, + 399 + ], + "type": "text", + "content": " and obtain the final feature map F:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 389, + 410, + 545, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 410, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 389, + 410, + 545, + 423 + ], + "type": "interline_equation", + "content": "F = E _ {2} \\left(F _ {a}, G _ {\\alpha}\\right) \\tag {3}", + "image_path": "59ab5c2c8bf0df838a16b39a4b2a58cba59da5cbcf3807f5ee55a8594159ba65.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "inline_equation", + "content": "G_{\\alpha}" + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "text", + "content": " are summed, and given as an input to next convolutional layers in the generator. Architectural details of " + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 304, + 426, + 547, + 463 + ], + "type": "text", + "content": " are given in Supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 467, + 405, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 405, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 405, + 480 + ], + "type": "text", + "content": "3.3. Training Phases" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "type": "text", + "content": "To train our model with the capabilities of high-fidelity inversion and high-quality editing, we use two training phases, as shown in Fig. 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "content": "No Editing Path. In this path, we reconstruct images with no editing on encoded " + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "content": ". This refers to the case where " + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 304, + 522, + 545, + 605 + ], + "type": "text", + "content": " in Eq. 2. This is also the training path other inversion methods solely rely on. While this path can teach the network to reconstruct images faithfully, it does not include any edits and cannot guide the network to high-quality editability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Cycle Translation Path. In this path, we edit images by setting " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " a value in the range of (4, 5), which we found to work best in our ablation studies. Via the edit, the generator outputs image " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ". Next, we feed this intermediate output image to the encoder and reverse the edit by inverting the addition presented in Eq. 2. The generator reconstructs " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "x''" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ", which is supposed to match the input image " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ". The cycle translation path is important because there is no ground-truth output for the edited image, " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ". Adversarial loss can" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1831" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 218, + 169 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 218, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 218, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 218, + 169 + ], + "type": "image", + "image_path": "5332370ca3472db546b2e80c75b86e8a03b5a141fbbce49fe66d376fa4a163d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 172, + 200, + 182 + ], + "lines": [ + { + "bbox": [ + 80, + 172, + 200, + 182 + ], + "spans": [ + { + "bbox": [ + 80, + 172, + 200, + 182 + ], + "type": "text", + "content": "(a) No Editing: Reconstruction Loss" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 72, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 244, + 72, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 244, + 72, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 244, + 72, + 545, + 170 + ], + "type": "image", + "image_path": "edc70d8352bb6ec52392320332ed22397dca7b3fa788a2e951500470de8545a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 173, + 488, + 183 + ], + "lines": [ + { + "bbox": [ + 345, + 173, + 488, + 183 + ], + "spans": [ + { + "bbox": [ + 345, + 173, + 488, + 183 + ], + "type": "text", + "content": "(b) Cycle-consistency: Reconstruction Loss" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 192, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 547, + 248 + ], + "type": "text", + "content": "Figure 4. We train StyleRes with (a) no editing based reconstruction and (b) cycle consistency based reconstruction losses. Many details from the inference pipeline are omitted for brevity. With reconstruction losses, the model learns to preserve the image details. We additionally apply adversarial losses on " + }, + { + "bbox": [ + 46, + 192, + 547, + 248 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 192, + 547, + 248 + ], + "type": "text", + "content": "; this way, when the image is edited, the transformation network learn to output realistic images. With cycle consistency based reconstruction, the network is regularized to keep the input details also during edits. We use additional losses as explained in Section 3.4." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 293 + ], + "type": "text", + "content": "guide " + }, + { + "bbox": [ + 46, + 268, + 287, + 293 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 46, + 268, + 287, + 293 + ], + "type": "text", + "content": " to look realistic but will not guide the network to keep the input image details if edited." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 299, + 163, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 163, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 163, + 312 + ], + "type": "text", + "content": "3.4. Training Objectives" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "text", + "content": "Reconstruction Losses. For both no editing and cycle consistency path outputs, our goal is to reconstruct the input image. To supervise this behavior, we use " + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "text", + "content": " loss, perceptual loss, and identity loss between the input and output images. The first reconstruction loss is as follows, where " + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "inline_equation", + "content": "x''" + }, + { + "bbox": [ + 46, + 317, + 288, + 401 + ], + "type": "text", + "content": " are output images, as given in Fig. 4:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 422, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 422, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 94, + 422, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c - l 2} = \\left| \\left| x ^ {\\prime} - x \\right| \\right| _ {2} + \\left| \\left| x ^ {\\prime \\prime} - x \\right| \\right| _ {2} \\tag {4}", + "image_path": "36c873de6832a224991026b2c13e3aa5cee55529a1db65ab6fbde0b17064f464.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "type": "text", + "content": "We use perceptual losses from VGG " + }, + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "type": "inline_equation", + "content": "(\\Phi)" + }, + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "type": "text", + "content": " at different feature layers " + }, + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "type": "inline_equation", + "content": "(j)" + }, + { + "bbox": [ + 47, + 446, + 287, + 483 + ], + "type": "text", + "content": " between these images from the loss objective as given in Eq. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 500, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 500, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 59, + 500, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c - p} = \\left| \\left| \\Phi_ {j} \\left(x ^ {\\prime}\\right) - \\Phi_ {j} (x) \\right| \\right| _ {2} + \\left| \\left| \\Phi_ {j} \\left(x ^ {\\prime \\prime}\\right) - \\Phi_ {j} (x) \\right| \\right| _ {2} \\tag {5}", + "image_path": "5be5f8631f6860d80e3a4f2ba05b8a838f27c6fd8361baaa63d2b668fcd0a1a1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "type": "text", + "content": "Identity loss is calculated with a pre-trained network " + }, + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 520, + 287, + 568 + ], + "type": "text", + "content": " is an ArcFace model [11] when training on the face domain and a domain specific ResNet-50 model [32] for our training on car class." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 576, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 576, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 52, + 576, + 287, + 590 + ], + "type": "interline_equation", + "content": "L _ {r e c - i d} = (1 - \\langle A (x), A \\left(x ^ {\\prime}\\right) \\rangle) + (1 - \\langle A (x), A \\left(x ^ {\\prime \\prime}\\right) \\rangle) \\tag {6}", + "image_path": "dbc67b9ba7a49e47afd546830c3b98fffeb7904100633d77d5a76215498027c2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 597, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 670 + ], + "type": "text", + "content": "Adversarial Losses. We also use adversarial losses to guide the network to output images that match the real image distribution. This objective improves realistic image inversions and edits. We load the pretrained discriminator from StyleGAN training, " + }, + { + "bbox": [ + 47, + 597, + 287, + 670 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 597, + 287, + 670 + ], + "type": "text", + "content": ", and train the discriminator together with the encoders." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 687, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 687, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 88, + 687, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {a d v} = 2 \\log D (x) + \\log \\left(1 - D \\left(x ^ {\\prime}\\right)\\right) \\tag {7} \\\\ + \\log \\left(1 - D \\left(x _ {i} ^ {\\prime}\\right)\\right) \\\\ \\end{array}", + "image_path": "77e2a557a00847677b69f229ff75ece8f926e03346db71ec5dd3bea9d307a8b7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 268, + 546, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 268, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 268, + 546, + 304 + ], + "type": "text", + "content": "Feature Regularizer. To prevent our encoder from deviating much from the original StyleGAN space, we regularize the residual features to be small:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 391, + 312, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 312, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 391, + 312, + 545, + 338 + ], + "type": "interline_equation", + "content": "L _ {F} = \\sum_ {F \\in \\phi} \\| F \\| _ {2} \\tag {8}", + "image_path": "2a2df9222f0053b8c10193322705fc5c152f0305791f8724f62ce49359863c10.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 346, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 371 + ], + "type": "text", + "content": "Full Objective. We use the overall objectives given below. The hyper parameters are provided in Supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 332, + 389, + 545, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 389, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 332, + 389, + 545, + 423 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {E _ {1}, E _ {2}} \\max _ {D} \\lambda_ {a} \\mathcal {L} _ {a d v} + \\lambda_ {r 1} \\mathcal {L} _ {r e c - l 2} + \\lambda_ {r 2} \\mathcal {L} _ {r e c - p} \\tag {9} \\\\ + \\lambda_ {r 3} \\mathcal {L} _ {r e c - i d} + \\lambda_ {f} \\mathcal {L} _ {F} \\\\ \\end{array}", + "image_path": "773f398b126300179230c98987065cfd6cd30957ba9b7493337316e1974b7184.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 442, + 388, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 388, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 388, + 456 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": "Set-up. For datasets and attribute editing, we follow the previous work [34]. For the human face domain, we train the model on the FFHQ [21] dataset and evaluate it on the CelebA-HQ [20] dataset. For the car domain, we use Stanford Cars [23] for training and evaluation. We run extensive experiments with directions explored with InterfaceGAN [30], GANSpace [15], StyleClip [26], and GradCtrl [7] methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": "Evaluation. We report metrics for reconstruction and editing qualities. For reconstruction, we report Frechet Inception Distance (FID) metric [16], which looks at the realism by comparing the target image distribution and reconstructed images, Learned Perceptual Image Patch Similarity (LPIPS) [44] and Structural Similarity Index Measure (SSIM), which compares the target and output pairs at the feature level of a pretrained deep network and pixel level, respectively. Additionally, FIDs are calculated for adding and removing the smile attributes on the CelebAHQ dataset. By using the ground-truth attributes, we add smile to images that do not have smile attribute and calculate FIDs between smile addition edited images and smiling" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1832" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 78, + 547, + 510 + ], + "blocks": [ + { + "bbox": [ + 50, + 78, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 50, + 78, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 50, + 78, + 547, + 510 + ], + "type": "image", + "image_path": "97bf85cbdf4c310efa39a4b361107c0245a5f88bc739cec5115373316276cc71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 518, + 547, + 541 + ], + "lines": [ + { + "bbox": [ + 46, + 518, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 547, + 541 + ], + "type": "text", + "content": "Figure 5. Qualitative results of inversion and editing. For each method, the first column shows inversion, and the second shows InterfaceGAN [30] and GANSpace [15] edits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 562, + 288, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 288, + 623 + ], + "type": "text", + "content": "ground-truth images. The same set-up is used for smile removal. On the Stanford cars dataset, we change grass and color attribute of images, and since there is no ground-truth attribute, we calculate the FIDs between the edited and original images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Baselines. We compare our method with state-of-the-art image inversion methods pSp [27], e4e [32], ReStyle [4], HyperStyle [5], HFGI [34], StyleTransformer [18], and FeatureStyle [40]. We use the author's released models. Hence, for the Stanford car dataset, some methods are omitted from comparisons if the models are not released. Among those, we only train HFGI for car model with au" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 562, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 586 + ], + "type": "text", + "content": "thor's released code since we base our main comparisons with it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Quantitative Results. In Table 1, we provide reconstruction and editing scores on the CelebA-HQ dataset. Our method achieves better results than all competing methods on all metrics. Most significantly, we achieve better FID scores on both reconstruction and editing qualities. While FeatureStyle achieves comparable SSIM and LPIPS scores on reconstruction, the editing FIDs are worse than our model. As shown in Table 2, we achieve significantly better results than previous methods on the Stanford Car dataset as well. We also compare the runtime of our method" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1833" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 124, + 286, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 114 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 114 + ], + "type": "text", + "content": "Table 1. Quantitative results of reconstruction and editing on CelebA-HQ dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for smile addition (+) and removal (-)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 124, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 124, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 124, + 286, + 220 + ], + "type": "table", + "html": "
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSSmile(+)Smile(-)
pSp [27]23.860.750.1732.4734.0
e4e [32]30.220.710.2138.5839.68
ReStyle [4]24.820.730.2030.3533.69
HyperStyle [5]16.080.830.1126.4325.26
HFGI [34]12.170.850.1325.2227.10
StyleTransformer [18]21.820.750.1734.3234.61
FeatureStyle [40]11.330.900.1027.2026.15
StyleRes (Ours)7.040.900.0923.5221.80
", + "image_path": "983909fc62436f9c50afef4969a58c0a7ea90929d952c72176200d258472f9af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 285, + 286, + 369 + ], + "blocks": [ + { + "bbox": [ + 47, + 232, + 287, + 276 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 287, + 276 + ], + "type": "text", + "content": "Table 2. Quantitative results of reconstruction and editing on the Stanford Cars Dataset. For reconstruction, we report FID, SSIM, and LPIPS scores. For editing, we report FID metrics for grass addition and color change." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 285, + 286, + 369 + ], + "lines": [ + { + "bbox": [ + 50, + 285, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 50, + 285, + 286, + 369 + ], + "type": "table", + "html": "
ReconstructionEditing - FIDs
MethodFIDSSIMLPIPSGrassColor
e4e [32]14.040.500.3218.0229.79
ReStyle [4]13.380.570.3016.0121.34
HyperStyle [5]11.640.630.2817.1326.30
HFGI [34]9.410.830.1614.8426.65
StyleTransformer [18]14.010.570.2819.4719.94
StyleRes (Ours)7.600.830.1410.6418.86
", + "image_path": "8ceb05a457aead6a7bd5033b9792fe04a720b404ac97d75de9b6b9e17d0a38b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 498 + ], + "type": "text", + "content": "and previous methods. Our method is significantly faster than HyperStyle (0.125 sec vs. 0.439 sec). That is because HyperStyle refines its predicted weight offsets gradually via multiple iterations. Our method is also faster than HFGI (0.125 sec vs. 0.130 sec), in addition to achieving better scores. We run a single stage network inference, whereas HFGI first generates an image via e4e and StyleGAN and provides the error map to a second architecture. The table for runtimes is provided in Supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": "Qualitative Results. We show visuals of inversion and editing results of our method, e4e, HFGI, and HyperStyle in Fig. 5. We provide further comparisons in Supplementary with other attribute editorings and with other methods. Compared to previous methods, our method achieves significantly better fidelity to the input images and preserves the identity and details when edited. It is the only method in these comparisons that preserves the background, earings (second row), hands (second-fourth rows), and hats (fourth row). Our method also achieves facial detail reconstruction; for example, in the fifth row, the person has a mole at the corner of her mouth. Among the inversions, our method is the only one preserving that. Furthermore, during the pose edit, it is transformed correctly. On the sixth row, our method is the only one that achieves the correct inversion and edit. In car examples, we again achieve high fidelity to the input images both in inversion and editing. e4e and HyperStyle do not reconstruct the image faithfully. On the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 79, + 539, + 361 + ], + "blocks": [ + { + "bbox": [ + 316, + 79, + 539, + 361 + ], + "lines": [ + { + "bbox": [ + 316, + 79, + 539, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 79, + 539, + 361 + ], + "type": "image", + "image_path": "323b7dea6ed040bf91e1d3464a5b225612ebe14f8e41b0d67629629342d9445f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 372, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 416 + ], + "type": "text", + "content": "Figure 6. Comparison of our method with e4e, HFGI, and HyperStyle with StyleClip edits [26]. The first column shows eyeglasses addition, the second and third columns show bangs addition, and the last column shows bob cut hairstyle results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "text", + "content": "other hand, HFGI outputs artefacts during edits. We additionally show results with edits explored by StyleClip [26] and GradCtrl [7] methods in Figs. 6 and 7, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": "Ablation Study. We run extensive experiments to validate the role of each proposed design choices. We first experiment with an architecture that does not have " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " modules. This experiment refers to the case where we learn a network that does not take input from StyleGAN Part 1 features, neither the original " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "(G_{W+})" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " nor the edited features " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "(G_{\\alpha})" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": ". The network directly takes higher layer features " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "(F_{0})" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " from the encoder and outputs them to the generator. As shown in Fig. 8, the network still tries to learn residual features to achieve reconstructions; however, it achieves poor edits. Without " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " refers to the experiment with " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " directly outputting features to Generator Part 2 (as shown in Fig. 3). Without " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " refers to the experiment where " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " directly takes input from the encoder " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "(F_{0})" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "(G_{\\alpha})" + }, + { + "bbox": [ + 304, + 475, + 545, + 713 + ], + "type": "text", + "content": " as defined in Sec. 3.2. None of the networks are able to learn the residual features and how to transform them as well as our final architecture since they are not designed to take the original and edited features separately. We also experiment without cycle consistency losses. Fig. 9 shows visual results of methods trained with and without cycle consistency constrain. We" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1834" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 79, + 286, + 439 + ], + "blocks": [ + { + "bbox": [ + 52, + 79, + 286, + 439 + ], + "lines": [ + { + "bbox": [ + 52, + 79, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 52, + 79, + 286, + 439 + ], + "type": "image", + "image_path": "60fcc17268c68bc0e499e97a161b701fcfd8ac9c00e101ceda5e5bb5910b8bde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 447, + 288, + 481 + ], + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 481 + ], + "type": "text", + "content": "Figure 7. Additional results of our method and e4e with GradCtrl edits [7]. The first three rows show blue sky edit, and the others show tree background edits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 502, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 538 + ], + "type": "text", + "content": "observe that with cycle consistency constrain, our network achieves preserving fine image details even when images are edited." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 549, + 119, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 119, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 119, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "We present a novel image inversion framework and a training pipeline to achieve high-fidelity image inversion with high-quality attribute editing. In this work, to achieve high-fidelity inversion, we learn residual features in higher latent codes that lower latent codes were not able to encode. This enables preserving image details in reconstruction. To achieve high-quality editing, we learn how to transform the residual features for adapting to manipulations in latent codes. We show state-of-the-art results on a wide range of edits explored with different methods both quantitatively and qualitatively while achieving faster run-time than competing methods." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 313, + 79, + 542, + 358 + ], + "blocks": [ + { + "bbox": [ + 313, + 79, + 542, + 358 + ], + "lines": [ + { + "bbox": [ + 313, + 79, + 542, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 79, + 542, + 358 + ], + "type": "image", + "image_path": "c9b2b4eb28f870563d8a3cdf00dd08fcade7106beee3157f8693e5cbf708d223.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "type": "text", + "content": "Figure 8. Ablation Study. Pose edit outputs of our final architecture and architecture with missing modules. w/o " + }, + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 305, + 367, + 545, + 401 + ], + "type": "text", + "content": ", networks struggle to transform features correctly." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 415, + 539, + 555 + ], + "blocks": [ + { + "bbox": [ + 314, + 415, + 539, + 555 + ], + "lines": [ + { + "bbox": [ + 314, + 415, + 539, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 539, + 555 + ], + "type": "image", + "image_path": "07bec594a38d21401644b96d3794980a1c29c83f59a7c3a3f20843ee26b28bc2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 563, + 545, + 607 + ], + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 607 + ], + "type": "text", + "content": "Figure 9. Ablation Study. Pose edit outputs of models trained with and without cycle consistency loss. The model trained with cycle consistency is better at preserving the details as shown in enlarged boxes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 632, + 403, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 403, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 403, + 645 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "This work has been funded by The Scientific and Technological Research Council of Turkey (TUBITAK), 3501 Research Project under Grant No 121E097. A. Dundar was supported by Marie Skłodowska-Curie Individual Fellowship." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1835" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 236 + ], + "type": "text", + "content": "[3] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "type": "text", + "content": "[4] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 339 + ], + "type": "text", + "content": "[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 373 + ], + "type": "text", + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 418 + ], + "type": "text", + "content": "[7]Zikun Chen, Ruowei Jiang, Brendan Duke, Han Zhao, and Parham Aarabi. Exploring gradient-based multi-directional controls in gans. In European Conference on Computer Vision, pages 104-119. Springer, 2022. 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "text", + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 509 + ], + "type": "text", + "content": "[9] Antonia Creswell and Anil Anthony Bharath. Inverting the generator of a generative adversarial network. IEEE transactions on neural networks and learning systems, 30(7):1967-1974, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 555 + ], + "type": "text", + "content": "[10] Yusuf Dalva, Said Fahri Altindis, and Aysegul Dundar. Vecgan: Image-to-image translation with interpretable latent directions. In European Conference on Computer Vision, pages 153-169. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[12] Aysegul Dundar, Karan Sapra, Guilin Liu, Andrew Tao, and Bryan Catanzaro. Panoptic-based image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8070-8079, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] Yue Gao, Fangyun Wei, Jianmin Bao, Shuyang Gu, Dong Chen, Fang Wen, and Zhouhui Lian. High-fidelity and arbitrary face editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16115-16124, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[15] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[17] Xianxu Hou, Xiaokang Zhang, Hanbang Liang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. Neural Networks, 145:209-220, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "text", + "content": "[18] Xueqi Hu, Qiusheng Huang, Zhengyi Shi, Siyuan Li, Changxin Gao, Li Sun, and Qingli Li. Style transformer for image inversion and editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11337-11346, June 2022. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[19] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. Eur. Conf. Comput. Vis., 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 492 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[23] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 602 + ], + "type": "text", + "content": "[24] Xinyang Li, Shengchuan Zhang, Jie Hu, Liujuan Cao, Xiaopeng Hong, Xudong Mao, Feiyue Huang, Yongjian Wu, and Rongrong Ji. Image-to-image translation via hierarchical style disentanglement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8639-8648, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 658 + ], + "type": "text", + "content": "[25] Guilin Liu, Aysegul Dundar, Kevin J Shih, Ting-Chun Wang, Fitsum A Reda, Karan Sapra, Zhiding Yu, Xiaodong Yang, Andrew Tao, and Bryan Catanzaro. Partial convolution for padding, inpainting, and image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 5, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1836" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[27] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 139, + 287, + 183 + ], + "type": "text", + "content": "[28] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "text", + "content": "[29] Wei Shen and Rujie Liu. Learning residual images for face attribute manipulation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4030-4038, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 282 + ], + "type": "text", + "content": "[30] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 327 + ], + "type": "text", + "content": "[31] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[32] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[33] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In International Conference on Machine Learning, pages 9786-9796. PMLR, 2020. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[34] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 581 + ], + "type": "text", + "content": "[36] Po-Wei Wu, Yu-Jing Lin, Che-Han Chang, Edward Y Chang, and Shih-Wei Liao. Relgan: Multi-domain image-to-image translation via relative attributes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5914–5922, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[37] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 680 + ], + "type": "text", + "content": "[38] Taihong Xiao, Jiapeng Hong, and Jinwen Ma. Elegant: Exchanging latent encodings with gan for transferring multiple face attributes. In Proceedings of the European conference on computer vision (ECCV), pages 168-184, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[39] Guoxing Yang, Nanyi Fei, Mingyu Ding, Guangzhen Liu, Zhiwu Lu, and Tao Xiang. L2m-gan: Learning to manipulate latent space semantics for facial attribute editing. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 442 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2951-2960, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[40] Xu Yao, Alasdair Newson, Yann Gousseau, and Pierre Hellier. A style-based gan encoder for high fidelity reconstruction of images and videos. European conference on computer vision, 2022. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 195 + ], + "type": "text", + "content": "[41] Ning Yu, Guilin Liu, Aysegul Dundar, Andrew Tao, Bryan Catanzaro, Larry S Davis, and Mario Fritz. Dual contrastive loss and attention for gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6731-6742, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 241 + ], + "type": "text", + "content": "[42] Gang Zhang, Meina Kan, Shiguang Shan, and Xilin Chen. Generative adversarial network with spatial attention for face attribute editing. In Proceedings of the European conference on computer vision (ECCV), pages 417-432, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 242, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 547, + 285 + ], + "type": "text", + "content": "[43] Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena. Self-attention generative adversarial networks. In International conference on machine learning, pages 7354–7363. PMLR, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 342 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 385 + ], + "type": "text", + "content": "[45] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 547, + 442 + ], + "type": "text", + "content": "[46] Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A Efros, Oliver Wang, and Eli Shechtman. Multimodal image-to-image translation by enforcing bi-cycle consistency. In Advances in neural information processing systems, pages 465-476, 2017. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1837" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_content_list.json b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..789d355bb5a5d9e8896c7b696f64daad0132e99d --- /dev/null +++ b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_content_list.json @@ -0,0 +1,1554 @@ +[ + { + "type": "text", + "text": "StyleSync: High-Fidelity Generalized and Personalized Lip Sync in Style-based Generator", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiazhi Guan $^{1,2*}$ Zhanwang Zhang $^{1*}$ Hang Zhou $^{1\\dagger}$ Tianshu Hu $^{1\\dagger}$ Kaisiyuan Wang $^{3}$ Dongliang He $^{1}$ Haocheng Feng $^{1}$ Jingtuo Liu $^{1}$ Errui Ding $^{1}$ Ziwei Liu $^{4}$ Jingdong Wang $^{1}$ $^{1}$ Department of Computer Vision Technology (VIS), Baidu Inc. $^{2}$ Tsinghua University $^{3}$ The University of Sydney $^{4}$ S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn {zhangzhanwang,zhouhang09,hutianshu01}@baidu.com", + "bbox": [ + 94, + 202, + 872, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/220009c78f4595caa6657435d0e5f7944517bc63bd4fa6a9a05966715de9482e.jpg", + "image_caption": [ + "Figure 1. Personalized lip-sync results generated by our StyleSync framework. Our method not only supports high-fidelity modification to any target template video according to conditional audio but can further adapt to specific styles with personalized optimization. In this figure, our lip-sync results should have the same mouth shapes as the lip-synced video of the conditional audio." + ], + "image_footnote": [], + "bbox": [ + 94, + 301, + 870, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 690, + 313, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advances in syncing lip movements with any audio waves, current methods still struggle to balance generation quality and the model's generalization ability. Previous studies either require long-term data for training or produce a similar movement pattern on all subjects with low quality. In this paper, we propose StyleSync, an effective framework that enables high-fidelity lip synchronization. We identify that a style-based generator would sufficiently enable such a charming property on both one-shot", + "bbox": [ + 75, + 723, + 470, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and few-shot scenarios. Specifically, we design a mask-guided spatial information encoding module that preserves the details of the given face. The mouth shapes are accurately modified by audio through modulated convolutions. Moreover, our design also enables personalized lip-sync by introducing style space and generator refinement on only limited frames. Thus the identity and talking style of a target person could be accurately preserved. Extensive experiments demonstrate the effectiveness of our method in producing high-fidelity results on a variety of scenes. Resources can be found at https://hangz-nju-cuhk.github.io/projects/StyleSync.", + "bbox": [ + 500, + 691, + 890, + 873 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 204, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1505", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The problem of generating lip-synced videos according to conditional audio is of great importance to the field of digital human creation, audio dubbing, film-making, and entertainment. While the rapid development of this area has been witnessed within recent years, most methods [6,8,9,16,19,24,27,28,38,39,46,50,54,57-60] focus on generating a whole dynamic talking head. Results created under such settings can hardly be blended into an existing scene. Under real-world scenarios like audio dubbing, one crucial need is to seamlessly alter the mouth or facial area while preserving other parts of the scene unchanged, making these methods non-feasible.", + "bbox": [ + 80, + 114, + 467, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous methods take two different paths for achieving seamless mouth modification. A number of studies [38, 44] pursue realistic results on person-specific settings, which require long-term clips for target modeling. Moreover, they rely on prior 3D facial structural information. The uncertainty and errors accumulated in the 3D fitting procedure would greatly influence their performances. On the other hand, it is desired to build models that break the data limitation on more generalized scenes. As a result, a few methods [31, 32, 41] design person-agnostic models without relying on 3D or 2D structural priors. Nevertheless, such a setting is extremely challenging.", + "bbox": [ + 80, + 296, + 467, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to produce high-fidelity lip-synced results on any-length videos, two essential challenges need to be addressed. 1) How to efficiently design a powerful generative backbone network that supports both accurate audio information expression and seamless local alternation. Intuitively, the lip-sync quality naturally contradicts the preservation of the original target frame information [32, 59]. 2) How to effectively leverage as much provided information as possible and involve the personalized properties to a generalized model. Though few-shot meta-learning has been proven effective in generating talking heads [5, 7, 56, 60], how to involve such ability into a lip-syncing pipeline has not been explored.", + "bbox": [ + 80, + 478, + 467, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a highly concise and comprehensive framework named StyleSync, which produces high-fidelity lip-sync results on both generalized and personalized scenarios. The key is our simple but lip-sync-oriented modifications to style-based generator. Though style-based generators [22, 23] have been leveraged in various talking head generation methods [5, 55, 59], their successes are only partially instructive. They aim at producing the whole head, which leads to unstable background and distortions which are non-acceptable in our scenarios.", + "bbox": [ + 80, + 674, + 467, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By revisiting the details of style-based generators, we identify a few simple but essential modifications that make our framework suitable for lip-syncing. Different from the above methods, we adopt a masked mouth modeling protocol [32, 41] and delicately design a Mask-based Spa", + "bbox": [ + 80, + 824, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tial Information Encoding strategy, where both the target and reference frames' information is encoded into a noise space [52, 53] of the generator according to different masking schemes. While the information on audio dynamics and high-level reference frame is injected into the style-modulated convolution in a similar manner as [25, 59]. In this way, our method can be benefited from the strong generative power of style-based generators and also keeps the advantage of easy implementation and fast training.", + "bbox": [ + 503, + 90, + 888, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, our network modification enables personalized information preserving (e.g., speaking styles and details of the mouth and jaw). We take inspiration from the recent studies of inverting StyleGAN priors [1, 2, 35, 45] and propose a Personalized Optimization scheme. As audio dubbing is normally performed on speaking videos, our model can make use of only a few seconds of the person's information and optimize additional person-specific parameters including the $W^{+}$ and the generator. Extensive experiments show that our framework clearly outperforms previous state of the arts on the one-shot setting by a large margin, and the target-specific optimization further enhances the fidelity of our results.", + "bbox": [ + 503, + 228, + 888, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: 1) We present the StyleSync framework, which adopts simple but effective modifications including the Mask-based Spatial Information Encoding to a style-based generator. 2) We propose the Personalized Optimization procedure which involves few-shot person-specific optimization into our framework. 3) Extensive experiments demonstrate that our framework can directly produce accurate and high-fidelity one-shot lip-sync results. Moreover, our proposed personalized optimization further improves the generation quality. Our method outperforms previous methods by a clear margin.", + "bbox": [ + 503, + 425, + 888, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 622, + 638, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Audio-Driven Facial Animation", + "text_level": 1, + "bbox": [ + 503, + 648, + 774, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The topic of audio-driven facial animation has long been a research interest in both the computer vision and graphics community. Studies have been carried out on both digital 3D human faces [15, 20, 33, 61] and realistic human heads [14, 17, 42]. We focus on human heads in the real world.", + "bbox": [ + 503, + 672, + 888, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Talking Head Generation. Most studies on lip-syncing target generating the whole head of a talking person [6, 8, 9, 16-18, 25, 26, 40, 42, 54, 58-60]. Specifically, a number of studies leverage structural information such as 2D [9] landmarks, 3D landmarks [60] and 3D meshes [6]. The uncertainty and inaccuracy of such representations would lead to error accumulation during the talking head generation procedure. Most person-specific methods [19, 27, 42] rely on them and produce results with a poor generalization or lip", + "bbox": [ + 503, + 763, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1506", + "bbox": [ + 485, + 946, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sync quality. Recently, NeRF is also used in person-specific modeling [16, 26, 43], but they also perform poorly when driven by audio from a different person.", + "bbox": [ + 75, + 90, + 467, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lip-Syncing on Faces. The other type of study focuses on lip-syncing the mouth part while keeping other information untouched in videos [31, 32, 39, 44]. Our work lies within the same scope. While Thies et al. [44] and Song et al. [38] produce realistic results, they rely on person-specific training on more than 2 minutes of data.", + "bbox": [ + 75, + 136, + 467, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, Wav2Lip [32] generates person-agnostic lip-sync results that are highly accurate. However, they build the generative model on low-resolution images, leading to blurry results. We identify that it would be easier for the model to learn the mouth motions if less information about the image quality needs to be recovered. Moreover, their methods cannot capture the personalized patterns given the target template video.", + "bbox": [ + 75, + 227, + 467, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Style-based Generator for Faces", + "text_level": 1, + "bbox": [ + 76, + 356, + 359, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN models [21-23] have shown great success on image generation tasks, particularly on facial image generation and editing [1, 2, 34, 36, 37, 45].", + "bbox": [ + 75, + 378, + 467, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN Inversion. StyleGAN inversion is the practice of recovering the latent space of a given image with a pretrained StyleGAN. Abdal et al. [1, 2] propose to invert images not only using the style space $W$ , but also expand it to the $\\mathcal{W}^+$ , which better preserves details. Recent studies propose to learn encoders [3, 34, 45] for better editability and faster optimization results. Furthermore, recent studies try to refine the generator's parameters through pivotal finetuning [35]. In our work, we take inspiration of previous StyleGAN inversion studies and propose a personalized optimization procedure. We encode a $\\mathcal{W}^+$ space from reference frames following [45] which enables personalized lip movements learning.", + "bbox": [ + 75, + 425, + 467, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Applications. StyleGAN architecture has also been leveraged in face restorations [48,53] and face swapping [51,52]. These tasks require preserving the original facial emotion and expressions. On the other hand, style-based generators have also been applied to create facial animations [5,25,59], which are highly related to our task. However, they rely on the style vectors in the $W$ or $W^{+}$ space for controlling both the appearances and motion dynamics. One major drawback of this setting is that $W^{+}$ space cannot easily accounts for the spatial consistency of backgrounds, leading to nonrealistic results or visible artifacts.", + "bbox": [ + 75, + 621, + 467, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 800, + 210, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the details of our StyleSync framework as depicted in Fig. 2. It is simply designed by leveraging several successes of previous style-based generators. We introduce our modifications to the style-based generator that make a successful backbone in Sec. 3.1", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the training objectives of our generalized model in Sec. 3.2. Specifically, we illustrate our Personalized Optimization procedure which further strengthens our results with person-specific optimization in Sec. 3.3.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Formulation. Our goal is to sync the lip motion of a target person with any provided audio and seamlessly blend it into the original target video. We formulate our training in a typical self-reconstruction manner. The training setting is similar to Wav2Lip [32]. Given a target video $\\mathbf{V} = \\{I_1, \\dots, I_T\\}$ with $T$ frames, we mask out the lower half of the face including the jaw and cheeks with a mask $M$ . The goal is to recover the whole face with its corresponding audio $\\mathbf{a} = \\{a_1, \\dots, a_T\\}$ ( $\\mathbf{a}$ is processed to spectrograms). As no information about the mouth and jaw shape is provided on the masked target frame $I_t^m = (\\mathbf{1} - M)*I_t$ , we leverage a random reference frame $I_{ref} \\in \\mathbf{V}$ during training to provide the desired context.", + "bbox": [ + 496, + 152, + 890, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Modifying Style-based Generator for Lip Sync", + "text_level": 1, + "bbox": [ + 498, + 359, + 890, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Style-based Generator Overview. We perform lip-sync-oriented modifications to build a successful backbone for lip sync with the StyleGAN2 [23] architecture. The original StyleGAN2 takes a style vector $w$ as input and uses it to modulate convolution operations in a total of $L$ generative layers during training. During inference, different $w$ s can be used at different layers, formulating the $W^{+}$ space $\\mathbf{W} = \\{w_{1},\\dots ,w_{L}\\}$ . The basic elements of the original StyleGAN2, including the $W^{+}$ space and the StyleGAN generator G's layers are depicted in blue on Fig. 2. For simplification, certain detailed structures are omitted.", + "bbox": [ + 496, + 383, + 890, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mask-based Spatial Information Encoding. Our goal is to seamlessly blend a lip-synced mouth into the target frame with the assistance of a reference image. As discussed in Sec. 2.2, previous methods leveraging the style-based generator are not directly applicable to our scene, we seek a different way to encode the spatial information on faces.", + "bbox": [ + 496, + 551, + 890, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is proven by recent studies on face restoration [53] and swapping [52] that content-guided feature maps can be attached to StyleGAN layers in a similar status as noise without affecting the expressive power of the generative model. These contextual-rich noises $\\mathbf{N} = \\{N_1,\\dots ,N_L\\}$ could well preserve the original spatial structure and attributes of the encoded visual information. This property provides an option for leveraging the attributes information lying in the reference image. Thus following Wav2Lip [32], we concatenate the masked and reference input together as the visual input of the framework and encode the corresponding feature maps $\\mathbf{F}_f = \\{F_1,\\ldots ,F_L\\}$ . The simplest setting is to regard $\\mathbf{N} = \\mathbf{F}$ .", + "bbox": [ + 496, + 643, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the goal of generating lip motions needs to create dynamics that are different from the original facial structure but keeps facial identity, which is substantially different from previous methods that aim to keep the structure", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1507", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70c1b84618322593295d202b149ded560bde30d5279920e2a606abd7ec617b7c.jpg", + "image_caption": [ + "Figure 2. Our StyleSync framework. The building blocks in Blue indicate the style-based generator. The masked target frame $I_{t}^{m}$ is concatenated with a reference $I_{ref}$ and encoded to $f_{f}$ by $\\mathrm{E}_{\\text{face}}(\\mathbf{Red})$ . The audio information is encoded to an audio feature $f_{a}$ (Yellow). The features are concatenated together to form the style $(W)$ space. Specifically, we devise a Personalized Optimization procedure (Green) including the learning of the $\\Delta \\mathbf{W}$ and the $\\Delta \\mathbf{P}$ . This part is not trained during the initial backbone learning." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fixed. The mouth shapes would be greatly influenced by the reference image $I_{ref}$ when adopting the same protocol. We identify that more than desired facial structure information is infused in the low-level information of $\\mathbf{F}$ (layers with higher resolutions). As a result, we intuitively mask out the low-level part of the information in the face features. Here we simply define", + "bbox": [ + 75, + 398, + 470, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l} ^ {\\prime} = (\\mathbf {1} - M) * F _ {l}, \\text {f o r} l > \\left\\lfloor \\frac {L}{2} \\right\\rfloor . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 513, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The facial information is colored in red on Fig. 2. The 1 is an all-one matrix that has the same size as $M$ .", + "bbox": [ + 75, + 551, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style Information Encoding. We follow previous studies [5, 40, 59] to encode both audio dynamics and facial information into the style space ( $W$ space) of the style-based generator. The audio feature $f_{a}$ is encoded from the encoder $E_{a}$ and the face feature $f_{f}$ is the bottleneck of the face encoder $E_{face}, w = \\mathrm{concat}(f_{a}, f_{f})$ . As the information in the reference frames is already fused to the generator through spatial noise encoding, the fusing of the face feature is less necessary. Experiments show that the difference is subtle with or without $f_{f}$ , but we still keep this design in accordance with common practice.", + "bbox": [ + 75, + 583, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ingredients for Personalization. Our above designs are not only simple but also possess external potentials of personalized lip-sync modeling. We leave the details to Sec. 3.3 and briefly discuss the additional ingredients that make our model comprehensive.", + "bbox": [ + 75, + 750, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, our extension is inspired by the advances in StyleGAN inversion [1, 2, 35]. It has pointed out that extending $w$ to the $W^{+}$ space which contains $\\mathbf{W} = \\{w + \\Delta w_{1},\\dots ,w + \\Delta w_{L}\\}$ is essential to recovering specific images. Moreover, recent studies [4, 35] also explore tuning", + "bbox": [ + 76, + 824, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the parameters of the generator to improve the network's fitting ability on a specific target. Thus our generator has the potential of learning limited parameter shifts $\\Delta \\mathbf{P}$ in order to fit a specific person.", + "bbox": [ + 496, + 398, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These personalized ingredients are depicted in green on Fig. 2 and are not optimized during the generalized training procedure. In order to keep the pipeline consistent across both settings, we set $\\Delta \\mathbf{W} = \\mathbf{0}$ and $\\Delta \\mathbf{P} = \\mathbf{0}$ in our generalized model.", + "bbox": [ + 496, + 458, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Backbone Training Objectives", + "text_level": 1, + "bbox": [ + 500, + 542, + 767, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, the whole backbone networks take the 6-channel concatenation of $[I_t^m, I_{ref}]$ and the audio clip $a_t$ at the same time step as input. It predicts $I_t' = \\mathrm{G}(\\mathbf{N}, \\mathbf{W})$ , which aims at recovering the unmasked target image $I_t$ . In order to keep our design simple, the training objectives are mostly aligned with StyleGAN2 [23] and Wav2Lip [32].", + "bbox": [ + 496, + 566, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Loss. The pixel-level reconstruction loss $\\mathcal{L}_{rec}$ is fundamental for training our task. Here we leverage the commonly used $L_{1}$ loss and perceptual loss [47].", + "bbox": [ + 498, + 656, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left\\| I _ {t} ^ {\\prime} - I _ {t} \\right\\| _ {1} + \\sum_ {m = 1} ^ {N _ {v g g}} \\left\\| \\mathrm {V G G} _ {m} \\left(I _ {t} ^ {\\prime}\\right) - \\mathrm {V G G} _ {m} \\left(I _ {t}\\right) \\right\\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 709, + 890, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{VGG}_m$ is the $m$ th layer's output of a pre-trained VGG19 network.", + "bbox": [ + 498, + 773, + 890, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial Loss. We directly adopt the same discriminator D from StyleGAN2 [23] for the adversarial training:", + "bbox": [ + 498, + 805, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} = \\min _ {\\mathrm {G}} \\max _ {\\mathrm {D}} \\left(\\mathbb {E} _ {I _ {t}} \\left[ \\log \\mathrm {D} \\left(I _ {t}\\right) \\right] + \\mathbb {E} _ {I _ {t} ^ {\\prime}} \\left[ \\log \\left(1 - \\mathrm {D} \\left(I _ {t} ^ {\\prime}\\right)\\right) \\right]\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 858, + 890, + 896 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1508", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Particularly, we initialize the discriminator's weights with the pre-trained version in StyleGAN2 [23]. This practice produces visibly sharper results.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lip-Sync Loss. We additionally train a SyncNet [12] which consists of a visual encoder $S_v$ and an audio encoder $S_a$ . It identifies whether visual and audio clips are timely aligned with contrastive loss. When training the generator, we predict 5 consecutive frames $I_{t:t+4}'$ within one batch at each inference step and supervise the training with SyncNet's assistance. The objective is:", + "bbox": [ + 75, + 136, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s y n c}} = - \\frac {\\mathrm {S} _ {v} \\left(I _ {t : t + 4} ^ {\\prime}\\right) ^ {\\mathrm {T}} \\cdot \\mathrm {S} _ {a} \\left(a _ {t : t + 4}\\right)}{\\| \\mathrm {S} _ {v} \\left(I _ {t : t + 4} ^ {\\prime}\\right) \\| _ {2} \\| \\mathrm {S} _ {a} \\left(a _ {t : t + 4}\\right) \\| _ {2}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 252, + 468, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss functions across the generalized training can be written as:", + "bbox": [ + 76, + 297, + 468, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g} = \\mathcal {L} _ {a d v} + \\lambda_ {r} \\mathcal {L} _ {r e c} + \\lambda_ {s} \\mathcal {L} _ {\\text {s y n c}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 356, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Personalized Optimization", + "text_level": 1, + "bbox": [ + 76, + 393, + 316, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the training of the backbone networks, our method can readily generate high-fidelity lip sync results for arbitrary subject. However, as the model is trained in a generalized manner, the generated mouth motion patterns across different people are basically the same. It has been verified that different identities possess different talking styles [50, 54]. Here we pursue to capture such personalized property with the original template video. Notably, we focus on the few-shot setting that only less than one minute of the original video is given, which cannot be handled by previous person-specific models [16, 19, 38, 44]. Below, we illustrate how we successfully design our personalized optimization module based on the above discussions.", + "bbox": [ + 75, + 417, + 468, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Basic Learning Settings. The basic learning procedure of the personalized optimization is similar to backbone training. On a template video $\\bar{\\mathbf{V}} = \\{\\bar{I}_1,\\dots ,\\bar{I}_T\\}$ with audio $\\bar{\\mathbf{a}} = \\{\\bar{a}_1\\dots ,\\bar{a}_T\\}$ , we randomly select two frames as target $\\bar{I}_t$ and reference $\\bar{I}_{ref}$ for training as one sample. During training, the generalized loss $\\mathcal{L}_g$ continuously supervises the whole procedure.", + "bbox": [ + 75, + 613, + 468, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the encoder $\\mathrm{E}_a$ encodes person-agnostic speech content information, it is fixed during the personalized optimization step. While the face encoder $\\mathrm{E}_{face}$ encodes the attribute information of the target person, we do not optimize its parameters on a single person to avoid overfitting.", + "bbox": [ + 75, + 719, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Style Space Optimization. According to our formulation, the $W^{+}$ space contains both the lip motion and the high-level facial style information. Thus optimizing it would ideally lead to person-specific lip motion pattern. E4E [45] verifies that learning the small displacements around the original $W$ space would both extend the inversion quality of the image and enables strong editing ability. Thus we propose", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to learn a set of $\\Delta \\mathbf{W} = MLP_{\\Delta w}(f_f)$ in a similar way according to the encoded face feature $f_{f}$ with a few layers of MLPs $MLP_{\\Delta w}$ .", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generator Tuning. The generator most accounts for the synthesizing and the blending quality. We allow the parameters $\\mathbf{P}$ of our generator $\\mathbf{G}$ to shift a little margin to $\\Delta \\mathbf{P}$ with the personalized data we use.", + "bbox": [ + 496, + 137, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning Objectives. The final learning objectives of the personalized optimization procedure can be summarized as:", + "bbox": [ + 496, + 199, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\mathcal {L} _ {g} + \\lambda_ {p} \\left(\\sum_ {i, j} \\left| \\Delta \\mathbf {W} _ {i} ^ {j} \\right| _ {2} ^ {2} + \\sum_ {m, n} \\left| \\Delta \\mathbf {P} _ {n} ^ {m} \\right| _ {2} ^ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 255, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_i^j$ denote the $j$ th parameter of the $i$ th element in $\\mathbf{W}$ . This stands the same for $\\mathbf{P}_n^m$ . We restrict the displacements in a limited step.", + "bbox": [ + 496, + 303, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 366, + 632, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two commonly used audio-visual datasets, LRW [11] and VoxCeleb2 [10]. We follow the datasets' original train/test split and train our model on a mixture of these two datasets.", + "bbox": [ + 496, + 392, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- LRW [11] is an audio-visual dataset collected from BBC news for lip reading. It is one of the earliest large-scale audio-visual datasets with high quality. It consists of 1,000 one-second utterances in 500 words.", + "- VoxCeleb2 [10] VoxCeleb [30] and VoxCeleb2 are large-scale audio-visual datasets collected for the speaker verification task. We re-download around one-fifth of VoxCeleb2 for training, which are of high quality. The official test set is processed with GPEN [53] for evaluation." + ], + "bbox": [ + 517, + 465, + 890, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We process all videos at 25 fps and align all faces according to pre-detected landmarks at the eyes. All faces are cropped to the size of $256 \\times 256$ . A same U-shape mask is adopted as shown in Fig. 2 to erase the mouth, cheeks and jaws at the target frame. All audiois processed in the same way as Wav2Lip [32]. In order to keep our design simple and avoid extra hyperparameter tuning, we adopt most settings from the previous studies [23, 32, 53]. Specifically, the generator has a total of $L = 14$ style-convolution layers. $\\lambda_{r}$ is empirically set to 10 and all other $\\lambda s$ are selected to 1. For personalized optimization, we train the person's data for 5 epochs. Long training would not lead to better results. Particularly, we find that training the personalized model together with the general training data leads to more stable results, the portion of personal data and general data is set as 1:1 in our experiments.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1509", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8e4fea59578dd2f24612c242704a827769897df1dd075646a2915ef935cf337.jpg", + "image_caption": [ + "Figure 3. Qualitative Results. The top row shows the lip-synced videos of the conditional driving audio. MakeitTalk [60] fails to generate accurate mouth shape and lacks head dynamics. PC-AVS [59] and Wav2Lip [32] generate lip-synced results aligned well with audios. But they produce visibly blurry results. Particularly, our reproduced Wav2Lip-H generates high-quality results with good lip-sync. However, they still produce certain artifacts. While our generalized model directly produces high-fidelity results that have the same lip motion as the original syncsed video." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 480, + 440 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ebf1340171f28ebeef39bba2ffb50bbcd782769173a93cf2f61fa754bd66a8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 88, + 890, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing Methods. As our model is built based on the person-agnostic setting, we compare our StyleSync framework with three state-of-the-art methods including, MakeiTalk [60], PC-AVS [59] and Wav2Lip [32]. MakeiTalk [60] synthesizes talking head videos with natural head pose. PC-AVS [59] a pose source video to achieve pose-controllable talking face generation. While MakeiTalk and PC-AVS produce a whole talking head, Wav2Lip which focuses on the mouth part is our main competing method. As Wav2Lip is originally trained on a low-resolution with small networks, we carefully reproduce it on the same data and even loss functions as our methods. We denote this model as Wav2Lip-H (High-Quality). For fair comparison, all basic comparisons are carried out on the generalized setting Ours-G.", + "bbox": [ + 75, + 545, + 472, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also perform personalized optimization on the test set, where each frame lasts less than 10 seconds for both datasets. The results are denoted as Ours-P. We equally perform the generator finetuning on the above methods, and all of them perform worse than their original models, thus the results are not listed below. Additionally, we compare our personalized setting with ADNeRF [16], a recent talking head advance using neural radiance fields [29].", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 498, + 544, + 720, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As quantitative evaluations can only be performed on the self-construction setting, we avoid directly leveraging all information within the frame for reconstruction. We uniformly select a same random frame as the reference frame for PC-AVS, Wav2Lip, Wav2Lip-H and ours.", + "bbox": [ + 496, + 568, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow previous studies [8, 9, 58, 59] to adopt the popularly used SSIM [49], PSNR metrics for evaluating the generation quality, and the landmark distances around the mouth (LMD) and SyncNet's [12] confidence score to show the lip-sync quality. Please be noted that the SyncNet metric is computed on the officially released version, which is different from our implementation in the loss functions.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Particularly, as the details of the target person cannot be fully recovered, we leverage the ArcFace [13] network and compute the frame-by-frame feature cosine distances $\\mathcal{D}_{ID}$ for evaluating whether the generated results preserve the identity well.", + "bbox": [ + 496, + 763, + 890, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Results. The quantitative experiments are carried out on the test set of LRW [11] and VoxCeleb2 [10] datasets. Please refer to Table 1 for the results. It can be seen that on the LRW dataset, the generalized version", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1510", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/abc1b9eb4fb7ab1c36c04fff71d8e5c279a3c183b200443b2f2d76797120544c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLRWVoxCeleb2
SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID↑SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID ↑
MakeiTalk0.6929.832.753.880.790.6328.386.942.150.71
PC-AVS0.7930.261.847.190.820.7129.532.758.160.74
Wav2Lip0.7930.541.287.390.900.8030.531.928.900.90
Wav2Lip-H0.8031.381.207.190.880.8130.531.878.350.90
GT1.00100.000.007.651.001.00100.000.007.711.00
Ours-G0.8531.781.187.250.890.7931.001.478.250.90
Ours-P0.8832.660.866.350.930.8231.541.157.260.93
", + "bbox": [ + 81, + 88, + 890, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on LRW and VoxCeleb. For LMD the lower the better, and the higher the better for other metrics.", + "bbox": [ + 101, + 252, + 866, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of our StyleSync already outperforms previous methods by a large margin for generation quality. While on the VoxCeleb2 dataset, the gap is less obvious. The reason might be that most training data of our methods are basically frontal view faces selected from LRW, while VoxCeleb2 contains more complicated scenes. On the other hand, after our personalized optimization, the performance of our model advances again. This shows the generated results are clearly more similar to the targets at this stage, which can also be verified by looking into the identity distances.", + "bbox": [ + 75, + 290, + 468, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meanwhile, our method also achieves comparable performance on the lip-sync metrics on both datasets. Our LMD score is slightly better than the competing methods. As for the SyncNet score, we achieve comparable results on LRW that are closer to the ground truth's SyncNet score. We argue it is meaningless to refer to the SyncNet scores for verifying the lip-sync quality once the metric has outperformed the ground truth. The $\\text{Sync}_{\\text{conf}}$ only reflects how well an audio-visual pair fits the learned SyncNet model rather than the true perceptual quality. Thus though generated results might outperform ground truth on the metric, it does not mean better sync quality.", + "bbox": [ + 75, + 441, + 468, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After the personalization, the lip-sync score degrades. We assume that the specificity of the speaking style reduces the mouth opening level. This is also shown in Fig. 5. The woman tends not to open her mouth wide, leading to smaller mouth movements", + "bbox": [ + 76, + 623, + 468, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 709, + 287, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subjective evaluation is crucial in identifying the ability of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video at https://hangz-nju-cuhk.github.io/projects/StyleSync.", + "bbox": [ + 75, + 733, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two cross-driven examples and their comparisons with the SOTA methods are shown in Fig. 3. We use an conditional audio from an arbitrary person selected from the test set to drive the template. It can be seen that MakeiTalk [60] cannot produce accurate lip movements. Moreover, both PC-AVS and Wav2Lip produce visible artifacts or blurry", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b43d453f1557b4c49c6416fe4aeebf4bcea100aa2bd18d7de9df88488b251f3e.jpg", + "image_caption": [ + "Figure 4. Self-driven results compared with personalized method." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 890, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results. Our reproduced Wav2Lip-H adopts the same training setting as ours, thus generates plausible results for most cases. Nevertheless, our generalized model still produces the most accurate mouth shapes with the highest fidelity.", + "bbox": [ + 496, + 454, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we show a case of personalized optimization on a 50-second video clip in Fig. 1. It can be seen that our method preserves the speaking style of the target person with accurate lip sync. Comparing our personalized results (Ours-P) with AD-NeRF [16] (Fig. 4), it is evident that our method maintains more person-specific details with higher fidelity, even though we use significantly less personal data (10s for tuning v.s. 5min for training in [16]).", + "bbox": [ + 496, + 517, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We also invite 15 participants to conduct a user study for further subjective evaluation and the results are reported in Table 2. Specifically, we randomly select 63 videos from LRW and VoxCeleb2 datasets as test sets and generate videos by using our generalized StyleSync and the comparison methods accordingly. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on three aspects for each generated video: (1) Lip-Sync quality; (2) Generation quality; (3) Video realness.", + "bbox": [ + 496, + 641, + 890, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, MakeitTalk [60] achieves the lowest scores in all the aspects due to lacking reasonable lip movements and head pose. PC-AVS [59], Wav2Lip [32] and Wav2Lip-H achieve relatively higher lip-sync score, but only Wav2Lip-H is able to synthesize videos with less blurry textures around the mouth. Overall, our StyleSync", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1511", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fa00b56397ce052857d8bdd60f9945d88222cfe47a65a70ba9faffce6c9ed0b8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MOS on \\ ApproachMakeitTalkPCALPSWav2LipWav2Lip-HOurs-G
Lip-Sync Quality2.063.003.493.674.24
Generation Quality2.632.161.873.424.52
Video Realness1.892.162.172.984.06
", + "bbox": [ + 81, + 88, + 467, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ec2312e4a9f7b05b3fa7bdf7b4791c5dfe0e37c357b2a0f080ad895e67ca9e3.jpg", + "table_caption": [ + "Table 2. User study measured by Mean Opinion Scores. The scores are ranged from 1 (worst) to 5 (best)." + ], + "table_footnote": [], + "table_body": "
MethodSSIM ↑PSNR ↑LMD ↓Syncconf ↑LID ↑
w/o mask0.8131.201.807.890.90
w/o sync0.8031.341.557.410.90
Ours-G0.7931.001.478.250.90
P w/o ΔW0.8231.511.327.310.92
Ours-P0.8231.541.157.260.93
", + "bbox": [ + 81, + 232, + 465, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative ablations results on VoxCeleb2.", + "bbox": [ + 112, + 333, + 429, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outperforms its counterparts in all the three aspects by a large margin, indicating the effectiveness of our approach.", + "bbox": [ + 75, + 366, + 468, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 407, + 228, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further demonstrate the contributions of our novel designs, we perform an ablation study on VoxCeleb2 dataset under both generalized setting and personalized optimization setting (denoted as \"Ours-G\" and \"Ours-P\", respectively). Concretely, we construct two variants for \"Ours-G\" (denoted as \"w/o mask\" and \"w/o sync\") by removing the masking operation in Eq. 1 and the lip-sync loss during training, respectively. While for \"Ours-P\", we form one variant (denoted as \"Ours-P w/o $\\Delta \\mathbf{W}$ ) by setting $\\Delta \\mathbf{W} = \\mathbf{0}$ . We have also experimented personalized optimization with fixed generated, however, this would lead to blurry results. This ablation is omitted here.", + "bbox": [ + 75, + 431, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative and qualitative results are shown in Table 3 and Fig. 5. As shown in Table 3, we observe that both \"w/o mask\" and \"w/o sync\" achieve comparable performance on the image quality metrics but leads to a performance drop on the lip-sync metrics. \"w/o mask\" achieves worse scores when compared with \"Ours-G\" in terms of LMD and Sync, which indicates that the proposed masking strategy can obviously alleviate the influence from the reference frames. While \"w/o sync\" reasonably suffers from lip-sync degradation and the results in Fig. 5 illustrate inconsistent lip movements when compared with the lip-synced video. The results demonstrate the effectiveness of the additional supervision from SyncNet.", + "bbox": [ + 75, + 613, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It can be seen from Fig. 5 that with the personalized optimization, both the identity and the pattern of the mouth opening on the generated frames become more similar to the original template video. It is also clear that after the personalized training, the mouth opening is less obvious than in the generalized model. We analyze that this is part of the", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1435fc58b50a1e4bab9eae4c4538d0fe2746fae5050c185db0d51b432bf3a94c.jpg", + "image_caption": [ + "Figure 5. Ablation study with visual results. Zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "talking style of this target person. This also leads to poorer metric values on SyncNet.", + "bbox": [ + 498, + 363, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In terms of the comparisons between our personalized model and \"Ours-P w/o $\\Delta \\mathbf{W}$ \", it achieves similar scores on the image quality as well as lip-sync quality. However, the identity score without $\\Delta \\mathbf{W}$ is slightly degraded, which indicates that it is beneficial to involve $\\Delta \\mathbf{W}$ into the training procedure for better preserving identity.", + "bbox": [ + 498, + 393, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussions", + "text_level": 1, + "bbox": [ + 500, + 497, + 756, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we propose StyleSync, which produces high-fidelity lip sync results for both the one-shot and the few-shot settings. We highlight the unique proprieties of our method: 1) Video results produced by our generalized model clearly outperform previous state-of-the-arts. 2) Our model is built upon the success of recent style-based generators with simple modification. It is easy to implement and friendly to train. 3) By involving StyleGAN-inspired personalized optimization procedure, our model can further be improved on a specific person given only a few clips.", + "bbox": [ + 496, + 523, + 890, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. As our method blends a lip-synced face into an existing video with a fixed mask, the head pose and expressions of the target person cannot be changed. Additionally, under certain extreme cases where the target's jaw is extremely large, it might get out of our masked area.", + "bbox": [ + 496, + 674, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical Considerations. Our method could be used to create non-existing talks and speeches, which might be maliciously used. We will issue our core models strictly to research institutions.", + "bbox": [ + 496, + 750, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1512", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2, 3, 4", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 3, 4", + "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3", + "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 4", + "[5] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2, 3, 4", + "[6] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation?: A survey and benchmark. arXiv preprint arXiv:2005.03201, 2020. 2", + "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2", + "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2, 6", + "[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 2, 6", + "[10] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 5, 6", + "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016. 5, 6", + "[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016. 5, 6", + "[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 6", + "[14] Bo Fan, Lijuan Wang, Frank K Soong, and Lei Xie. Photoreal talking head with deep bidirectional LSTM. In 2015 IEEE" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4884-4888. IEEE, 2015. 2", + "[15] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2", + "[16] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 5, 6, 7", + "[17] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127(11):1767-1779, 2019. 2", + "[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH, 2022. 2", + "[19] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14080-14089, 2021. 2, 5", + "[20] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2", + "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3", + "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 3", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2, 3, 4, 5", + "[24] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2755-2764, 2021. 2", + "[25] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2, 3", + "[26] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. ECCV, 2022. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1513", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2", + "[28] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. AAAI, 2023. 2", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 6", + "[30] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 5", + "[31] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence, 2022. 2, 3", + "[32] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 4, 5, 6, 7", + "[33] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 3", + "[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 4", + "[36] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 3", + "[37] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3", + "[38] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. arXiv preprint arXiv:2001.05201, 2020. 2, 3, 5", + "[39] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018. 2, 3", + "[40] Yasheng Sun, Hang Zhou, Ziwei Liu, and Hideki Koike. Speech2talking-face: Inferring and driving a face with synchronized audio-visual representation. In *IJCAI*, 2021. 2, 4", + "[41] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2", + "[42] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2", + "[43] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3", + "[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, pages 716-731. Springer, 2020. 2, 3, 5", + "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 2, 3, 5", + "[46] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2", + "[47] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 4", + "[48] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[50] Haozhe Wu, Jia Jia, Haoyu Wang, Yishun Dou, Chao Duan, and Qingshan Deng. Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1478-1486, 2021. 2, 5", + "[51] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 3", + "[52] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677. Springer, 2022. 2, 3", + "[53] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 2, 3, 5", + "[54] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1514", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning-based personalized head pose. arXiv e-prints, pages arXiv-2002, 2020. 2, 5", + "[55] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. ECCV, 2022. 2", + "[56] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2", + "[57] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3867-3876, 2021. 2", + "[58] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 2, 6", + "[59] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 3, 4, 6, 7", + "[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 6, 7", + "[61] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + ], + "bbox": [ + 78, + 90, + 468, + 612 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1515", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_model.json b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c626a73037a0670a9d66467da1ebc2b6cf439897 --- /dev/null +++ b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.812, + 0.175 + ], + "angle": 0, + "content": "StyleSync: High-Fidelity Generalized and Personalized Lip Sync in Style-based Generator" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.203, + 0.874, + 0.293 + ], + "angle": 0, + "content": "Jiazhi Guan\\(^{1,2*}\\) Zhanwang Zhang\\(^{1*}\\) Hang Zhou\\(^{1\\dagger}\\) Tianshu Hu\\(^{1\\dagger}\\) Kaisiyuan Wang\\(^{3}\\) Dongliang He\\(^{1}\\) Haocheng Feng\\(^{1}\\) Jingtuo Liu\\(^{1}\\) Errui Ding\\(^{1}\\) Ziwei Liu\\(^{4}\\) Jingdong Wang\\(^{1}\\) \\(^{1}\\)Department of Computer Vision Technology (VIS), Baidu Inc. \\(^{2}\\)Tsinghua University \\(^{3}\\)The University of Sydney \\(^{4}\\)S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn {zhangzhanwang,zhouhang09,hutianshu01}@baidu.com" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.302, + 0.872, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.636, + 0.895, + 0.681 + ], + "angle": 0, + "content": "Figure 1. Personalized lip-sync results generated by our StyleSync framework. Our method not only supports high-fidelity modification to any target template video according to conditional audio but can further adapt to specific styles with personalized optimization. In this figure, our lip-sync results should have the same mouth shapes as the lip-synced video of the conditional audio." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.691, + 0.314, + 0.707 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.471, + 0.86 + ], + "angle": 0, + "content": "Despite recent advances in syncing lip movements with any audio waves, current methods still struggle to balance generation quality and the model's generalization ability. Previous studies either require long-term data for training or produce a similar movement pattern on all subjects with low quality. In this paper, we propose StyleSync, an effective framework that enables high-fidelity lip synchronization. We identify that a style-based generator would sufficiently enable such a charming property on both one-shot" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.692, + 0.892, + 0.874 + ], + "angle": 0, + "content": "and few-shot scenarios. Specifically, we design a mask-guided spatial information encoding module that preserves the details of the given face. The mouth shapes are accurately modified by audio through modulated convolutions. Moreover, our design also enables personalized lip-sync by introducing style space and generator refinement on only limited frames. Thus the identity and talking style of a target person could be accurately preserved. Extensive experiments demonstrate the effectiveness of our method in producing high-fidelity results on a variety of scenes. Resources can be found at https://hangz-nju-cuhk.github.io/projects/StyleSync." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.205, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.227, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1505" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.115, + 0.468, + 0.295 + ], + "angle": 0, + "content": "The problem of generating lip-synced videos according to conditional audio is of great importance to the field of digital human creation, audio dubbing, film-making, and entertainment. While the rapid development of this area has been witnessed within recent years, most methods [6,8,9,16,19,24,27,28,38,39,46,50,54,57-60] focus on generating a whole dynamic talking head. Results created under such settings can hardly be blended into an existing scene. Under real-world scenarios like audio dubbing, one crucial need is to seamlessly alter the mouth or facial area while preserving other parts of the scene unchanged, making these methods non-feasible." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.297, + 0.468, + 0.477 + ], + "angle": 0, + "content": "Previous methods take two different paths for achieving seamless mouth modification. A number of studies [38, 44] pursue realistic results on person-specific settings, which require long-term clips for target modeling. Moreover, they rely on prior 3D facial structural information. The uncertainty and errors accumulated in the 3D fitting procedure would greatly influence their performances. On the other hand, it is desired to build models that break the data limitation on more generalized scenes. As a result, a few methods [31, 32, 41] design person-agnostic models without relying on 3D or 2D structural priors. Nevertheless, such a setting is extremely challenging." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.479, + 0.468, + 0.673 + ], + "angle": 0, + "content": "In order to produce high-fidelity lip-synced results on any-length videos, two essential challenges need to be addressed. 1) How to efficiently design a powerful generative backbone network that supports both accurate audio information expression and seamless local alternation. Intuitively, the lip-sync quality naturally contradicts the preservation of the original target frame information [32, 59]. 2) How to effectively leverage as much provided information as possible and involve the personalized properties to a generalized model. Though few-shot meta-learning has been proven effective in generating talking heads [5, 7, 56, 60], how to involve such ability into a lip-syncing pipeline has not been explored." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.675, + 0.468, + 0.825 + ], + "angle": 0, + "content": "In this paper, we propose a highly concise and comprehensive framework named StyleSync, which produces high-fidelity lip-sync results on both generalized and personalized scenarios. The key is our simple but lip-sync-oriented modifications to style-based generator. Though style-based generators [22, 23] have been leveraged in various talking head generation methods [5, 55, 59], their successes are only partially instructive. They aim at producing the whole head, which leads to unstable background and distortions which are non-acceptable in our scenarios." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "By revisiting the details of style-based generators, we identify a few simple but essential modifications that make our framework suitable for lip-syncing. Different from the above methods, we adopt a masked mouth modeling protocol [32, 41] and delicately design a Mask-based Spa" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.227 + ], + "angle": 0, + "content": "tial Information Encoding strategy, where both the target and reference frames' information is encoded into a noise space [52, 53] of the generator according to different masking schemes. While the information on audio dynamics and high-level reference frame is injected into the style-modulated convolution in a similar manner as [25, 59]. In this way, our method can be benefited from the strong generative power of style-based generators and also keeps the advantage of easy implementation and fast training." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.229, + 0.89, + 0.424 + ], + "angle": 0, + "content": "Moreover, our network modification enables personalized information preserving (e.g., speaking styles and details of the mouth and jaw). We take inspiration from the recent studies of inverting StyleGAN priors [1, 2, 35, 45] and propose a Personalized Optimization scheme. As audio dubbing is normally performed on speaking videos, our model can make use of only a few seconds of the person's information and optimize additional person-specific parameters including the \\( W^{+} \\) and the generator. Extensive experiments show that our framework clearly outperforms previous state of the arts on the one-shot setting by a large margin, and the target-specific optimization further enhances the fidelity of our results." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.426, + 0.89, + 0.606 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: 1) We present the StyleSync framework, which adopts simple but effective modifications including the Mask-based Spatial Information Encoding to a style-based generator. 2) We propose the Personalized Optimization procedure which involves few-shot person-specific optimization into our framework. 3) Extensive experiments demonstrate that our framework can directly produce accurate and high-fidelity one-shot lip-sync results. Moreover, our proposed personalized optimization further improves the generation quality. Our method outperforms previous methods by a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.623, + 0.64, + 0.639 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.649, + 0.776, + 0.665 + ], + "angle": 0, + "content": "2.1. Audio-Driven Facial Animation" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.673, + 0.89, + 0.763 + ], + "angle": 0, + "content": "The topic of audio-driven facial animation has long been a research interest in both the computer vision and graphics community. Studies have been carried out on both digital 3D human faces [15, 20, 33, 61] and realistic human heads [14, 17, 42]. We focus on human heads in the real world." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.765, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Talking Head Generation. Most studies on lip-syncing target generating the whole head of a talking person [6, 8, 9, 16-18, 25, 26, 40, 42, 54, 58-60]. Specifically, a number of studies leverage structural information such as 2D [9] landmarks, 3D landmarks [60] and 3D meshes [6]. The uncertainty and inaccuracy of such representations would lead to error accumulation during the talking head generation procedure. Most person-specific methods [19, 27, 42] rely on them and produce results with a poor generalization or lip" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "1506" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "sync quality. Recently, NeRF is also used in person-specific modeling [16, 26, 43], but they also perform poorly when driven by audio from a different person." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.226 + ], + "angle": 0, + "content": "Lip-Syncing on Faces. The other type of study focuses on lip-syncing the mouth part while keeping other information untouched in videos [31, 32, 39, 44]. Our work lies within the same scope. While Thies et al. [44] and Song et al. [38] produce realistic results, they rely on person-specific training on more than 2 minutes of data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.349 + ], + "angle": 0, + "content": "Specifically, Wav2Lip [32] generates person-agnostic lip-sync results that are highly accurate. However, they build the generative model on low-resolution images, leading to blurry results. We identify that it would be easier for the model to learn the mouth motions if less information about the image quality needs to be recovered. Moreover, their methods cannot capture the personalized patterns given the target template video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.357, + 0.36, + 0.373 + ], + "angle": 0, + "content": "2.2. Style-based Generator for Faces" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.468, + 0.424 + ], + "angle": 0, + "content": "StyleGAN models [21-23] have shown great success on image generation tasks, particularly on facial image generation and editing [1, 2, 34, 36, 37, 45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.468, + 0.621 + ], + "angle": 0, + "content": "StyleGAN Inversion. StyleGAN inversion is the practice of recovering the latent space of a given image with a pretrained StyleGAN. Abdal et al. [1, 2] propose to invert images not only using the style space \\( W \\), but also expand it to the \\( \\mathcal{W}^+ \\), which better preserves details. Recent studies propose to learn encoders [3, 34, 45] for better editability and faster optimization results. Furthermore, recent studies try to refine the generator's parameters through pivotal finetuning [35]. In our work, we take inspiration of previous StyleGAN inversion studies and propose a personalized optimization procedure. We encode a \\( \\mathcal{W}^+ \\) space from reference frames following [45] which enables personalized lip movements learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.468, + 0.786 + ], + "angle": 0, + "content": "Applications. StyleGAN architecture has also been leveraged in face restorations [48,53] and face swapping [51,52]. These tasks require preserving the original facial emotion and expressions. On the other hand, style-based generators have also been applied to create facial animations [5,25,59], which are highly related to our task. However, they rely on the style vectors in the \\( W \\) or \\( W^{+} \\) space for controlling both the appearances and motion dynamics. One major drawback of this setting is that \\( W^{+} \\) space cannot easily accounts for the spatial consistency of backgrounds, leading to nonrealistic results or visible artifacts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.212, + 0.817 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the details of our StyleSync framework as depicted in Fig. 2. It is simply designed by leveraging several successes of previous style-based generators. We introduce our modifications to the style-based generator that make a successful backbone in Sec. 3.1" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "and the training objectives of our generalized model in Sec. 3.2. Specifically, we illustrate our Personalized Optimization procedure which further strengthens our results with person-specific optimization in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Task Formulation. Our goal is to sync the lip motion of a target person with any provided audio and seamlessly blend it into the original target video. We formulate our training in a typical self-reconstruction manner. The training setting is similar to Wav2Lip [32]. Given a target video \\(\\mathbf{V} = \\{I_1, \\dots, I_T\\}\\) with \\(T\\) frames, we mask out the lower half of the face including the jaw and cheeks with a mask \\(M\\). The goal is to recover the whole face with its corresponding audio \\(\\mathbf{a} = \\{a_1, \\dots, a_T\\}\\) (\\(\\mathbf{a}\\) is processed to spectrograms). As no information about the mouth and jaw shape is provided on the masked target frame \\(I_t^m = (\\mathbf{1} - M)*I_t\\), we leverage a random reference frame \\(I_{ref} \\in \\mathbf{V}\\) during training to provide the desired context." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.377 + ], + "angle": 0, + "content": "3.1. Modifying Style-based Generator for Lip Sync" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Style-based Generator Overview. We perform lip-sync-oriented modifications to build a successful backbone for lip sync with the StyleGAN2 [23] architecture. The original StyleGAN2 takes a style vector \\( w \\) as input and uses it to modulate convolution operations in a total of \\( L \\) generative layers during training. During inference, different \\( w \\)s can be used at different layers, formulating the \\( W^{+} \\) space \\( \\mathbf{W} = \\{w_{1},\\dots ,w_{L}\\} \\). The basic elements of the original StyleGAN2, including the \\( W^{+} \\) space and the StyleGAN generator G's layers are depicted in blue on Fig. 2. For simplification, certain detailed structures are omitted." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Mask-based Spatial Information Encoding. Our goal is to seamlessly blend a lip-synced mouth into the target frame with the assistance of a reference image. As discussed in Sec. 2.2, previous methods leveraging the style-based generator are not directly applicable to our scene, we seek a different way to encode the spatial information on faces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.839 + ], + "angle": 0, + "content": "It is proven by recent studies on face restoration [53] and swapping [52] that content-guided feature maps can be attached to StyleGAN layers in a similar status as noise without affecting the expressive power of the generative model. These contextual-rich noises \\(\\mathbf{N} = \\{N_1,\\dots ,N_L\\}\\) could well preserve the original spatial structure and attributes of the encoded visual information. This property provides an option for leveraging the attributes information lying in the reference image. Thus following Wav2Lip [32], we concatenate the masked and reference input together as the visual input of the framework and encode the corresponding feature maps \\(\\mathbf{F}_f = \\{F_1,\\ldots ,F_L\\}\\). The simplest setting is to regard \\(\\mathbf{N} = \\mathbf{F}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, the goal of generating lip motions needs to create dynamics that are different from the original facial structure but keeps facial identity, which is substantially different from previous methods that aim to keep the structure" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Our StyleSync framework. The building blocks in Blue indicate the style-based generator. The masked target frame \\( I_{t}^{m} \\) is concatenated with a reference \\( I_{ref} \\) and encoded to \\( f_{f} \\) by \\( \\mathrm{E}_{\\text{face}}(\\mathbf{Red}) \\). The audio information is encoded to an audio feature \\( f_{a} \\) (Yellow). The features are concatenated together to form the style \\( (W) \\) space. Specifically, we devise a Personalized Optimization procedure (Green) including the learning of the \\( \\Delta \\mathbf{W} \\) and the \\( \\Delta \\mathbf{P} \\). This part is not trained during the initial backbone learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.506 + ], + "angle": 0, + "content": "fixed. The mouth shapes would be greatly influenced by the reference image \\( I_{ref} \\) when adopting the same protocol. We identify that more than desired facial structure information is infused in the low-level information of \\( \\mathbf{F} \\) (layers with higher resolutions). As a result, we intuitively mask out the low-level part of the information in the face features. Here we simply define" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.515, + 0.469, + 0.545 + ], + "angle": 0, + "content": "\\[\nF _ {l} ^ {\\prime} = (\\mathbf {1} - M) * F _ {l}, \\text {f o r} l > \\left\\lfloor \\frac {L}{2} \\right\\rfloor . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.584 + ], + "angle": 0, + "content": "The facial information is colored in red on Fig. 2. The 1 is an all-one matrix that has the same size as \\( M \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Style Information Encoding. We follow previous studies [5, 40, 59] to encode both audio dynamics and facial information into the style space (\\(W\\) space) of the style-based generator. The audio feature \\(f_{a}\\) is encoded from the encoder \\(E_{a}\\) and the face feature \\(f_{f}\\) is the bottleneck of the face encoder \\(E_{face}, w = \\mathrm{concat}(f_{a}, f_{f})\\). As the information in the reference frames is already fused to the generator through spatial noise encoding, the fusing of the face feature is less necessary. Experiments show that the difference is subtle with or without \\(f_{f}\\), but we still keep this design in accordance with common practice." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Ingredients for Personalization. Our above designs are not only simple but also possess external potentials of personalized lip-sync modeling. We leave the details to Sec. 3.3 and briefly discuss the additional ingredients that make our model comprehensive." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Specifically, our extension is inspired by the advances in StyleGAN inversion [1, 2, 35]. It has pointed out that extending \\( w \\) to the \\( W^{+} \\) space which contains \\( \\mathbf{W} = \\{w + \\Delta w_{1},\\dots ,w + \\Delta w_{L}\\} \\) is essential to recovering specific images. Moreover, recent studies [4, 35] also explore tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.459 + ], + "angle": 0, + "content": "the parameters of the generator to improve the network's fitting ability on a specific target. Thus our generator has the potential of learning limited parameter shifts \\(\\Delta \\mathbf{P}\\) in order to fit a specific person." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.534 + ], + "angle": 0, + "content": "These personalized ingredients are depicted in green on Fig. 2 and are not optimized during the generalized training procedure. In order to keep the pipeline consistent across both settings, we set \\(\\Delta \\mathbf{W} = \\mathbf{0}\\) and \\(\\Delta \\mathbf{P} = \\mathbf{0}\\) in our generalized model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.768, + 0.56 + ], + "angle": 0, + "content": "3.2. Backbone Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.657 + ], + "angle": 0, + "content": "During training, the whole backbone networks take the 6-channel concatenation of \\([I_t^m, I_{ref}]\\) and the audio clip \\(a_t\\) at the same time step as input. It predicts \\(I_t' = \\mathrm{G}(\\mathbf{N}, \\mathbf{W})\\), which aims at recovering the unmasked target image \\(I_t\\). In order to keep our design simple, the training objectives are mostly aligned with StyleGAN2 [23] and Wav2Lip [32]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.657, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Reconstruction Loss. The pixel-level reconstruction loss \\(\\mathcal{L}_{rec}\\) is fundamental for training our task. Here we leverage the commonly used \\(L_{1}\\) loss and perceptual loss [47]." + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.71, + 0.891, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left\\| I _ {t} ^ {\\prime} - I _ {t} \\right\\| _ {1} + \\sum_ {m = 1} ^ {N _ {v g g}} \\left\\| \\mathrm {V G G} _ {m} \\left(I _ {t} ^ {\\prime}\\right) - \\mathrm {V G G} _ {m} \\left(I _ {t}\\right) \\right\\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.775, + 0.891, + 0.804 + ], + "angle": 0, + "content": "where \\(\\mathrm{VGG}_m\\) is the \\(m\\)th layer's output of a pre-trained VGG19 network." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.806, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Adversarial Loss. We directly adopt the same discriminator D from StyleGAN2 [23] for the adversarial training:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.859, + 0.891, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} = \\min _ {\\mathrm {G}} \\max _ {\\mathrm {D}} \\left(\\mathbb {E} _ {I _ {t}} \\left[ \\log \\mathrm {D} \\left(I _ {t}\\right) \\right] + \\mathbb {E} _ {I _ {t} ^ {\\prime}} \\left[ \\log \\left(1 - \\mathrm {D} \\left(I _ {t} ^ {\\prime}\\right)\\right) \\right]\\right). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1508" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "Particularly, we initialize the discriminator's weights with the pre-trained version in StyleGAN2 [23]. This practice produces visibly sharper results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Lip-Sync Loss. We additionally train a SyncNet [12] which consists of a visual encoder \\( S_v \\) and an audio encoder \\( S_a \\). It identifies whether visual and audio clips are timely aligned with contrastive loss. When training the generator, we predict 5 consecutive frames \\( I_{t:t+4}' \\) within one batch at each inference step and supervise the training with SyncNet's assistance. The objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.253, + 0.47, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s y n c}} = - \\frac {\\mathrm {S} _ {v} \\left(I _ {t : t + 4} ^ {\\prime}\\right) ^ {\\mathrm {T}} \\cdot \\mathrm {S} _ {a} \\left(a _ {t : t + 4}\\right)}{\\| \\mathrm {S} _ {v} \\left(I _ {t : t + 4} ^ {\\prime}\\right) \\| _ {2} \\| \\mathrm {S} _ {a} \\left(a _ {t : t + 4}\\right) \\| _ {2}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.299, + 0.469, + 0.329 + ], + "angle": 0, + "content": "The overall loss functions across the generalized training can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.357, + 0.469, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g} = \\mathcal {L} _ {a d v} + \\lambda_ {r} \\mathcal {L} _ {r e c} + \\lambda_ {s} \\mathcal {L} _ {\\text {s y n c}}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.318, + 0.411 + ], + "angle": 0, + "content": "3.3. Personalized Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.612 + ], + "angle": 0, + "content": "After the training of the backbone networks, our method can readily generate high-fidelity lip sync results for arbitrary subject. However, as the model is trained in a generalized manner, the generated mouth motion patterns across different people are basically the same. It has been verified that different identities possess different talking styles [50, 54]. Here we pursue to capture such personalized property with the original template video. Notably, we focus on the few-shot setting that only less than one minute of the original video is given, which cannot be handled by previous person-specific models [16, 19, 38, 44]. Below, we illustrate how we successfully design our personalized optimization module based on the above discussions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Basic Learning Settings. The basic learning procedure of the personalized optimization is similar to backbone training. On a template video \\(\\bar{\\mathbf{V}} = \\{\\bar{I}_1,\\dots ,\\bar{I}_T\\}\\) with audio \\(\\bar{\\mathbf{a}} = \\{\\bar{a}_1\\dots ,\\bar{a}_T\\}\\), we randomly select two frames as target \\(\\bar{I}_t\\) and reference \\(\\bar{I}_{ref}\\) for training as one sample. During training, the generalized loss \\(\\mathcal{L}_g\\) continuously supervises the whole procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.795 + ], + "angle": 0, + "content": "As the encoder \\( \\mathrm{E}_a \\) encodes person-agnostic speech content information, it is fixed during the personalized optimization step. While the face encoder \\( \\mathrm{E}_{face} \\) encodes the attribute information of the target person, we do not optimize its parameters on a single person to avoid overfitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Style Space Optimization. According to our formulation, the \\( W^{+} \\) space contains both the lip motion and the high-level facial style information. Thus optimizing it would ideally lead to person-specific lip motion pattern. E4E [45] verifies that learning the small displacements around the original \\( W \\) space would both extend the inversion quality of the image and enables strong editing ability. Thus we propose" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "to learn a set of \\(\\Delta \\mathbf{W} = MLP_{\\Delta w}(f_f)\\) in a similar way according to the encoded face feature \\(f_{f}\\) with a few layers of MLPs \\(MLP_{\\Delta w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Generator Tuning. The generator most accounts for the synthesizing and the blending quality. We allow the parameters \\(\\mathbf{P}\\) of our generator \\(\\mathbf{G}\\) to shift a little margin to \\(\\Delta \\mathbf{P}\\) with the personalized data we use." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Learning Objectives. The final learning objectives of the personalized optimization procedure can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.256, + 0.892, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\mathcal {L} _ {g} + \\lambda_ {p} \\left(\\sum_ {i, j} \\left| \\Delta \\mathbf {W} _ {i} ^ {j} \\right| _ {2} ^ {2} + \\sum_ {m, n} \\left| \\Delta \\mathbf {P} _ {n} ^ {m} \\right| _ {2} ^ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.351 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_i^j\\) denote the \\(j\\)th parameter of the \\(i\\)th element in \\(\\mathbf{W}\\). This stands the same for \\(\\mathbf{P}_n^m\\). We restrict the displacements in a limited step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.367, + 0.633, + 0.384 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two commonly used audio-visual datasets, LRW [11] and VoxCeleb2 [10]. We follow the datasets' original train/test split and train our model on a mixture of these two datasets." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.466, + 0.892, + 0.527 + ], + "angle": 0, + "content": "- LRW [11] is an audio-visual dataset collected from BBC news for lip reading. It is one of the earliest large-scale audio-visual datasets with high quality. It consists of 1,000 one-second utterances in 500 words." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.541, + 0.892, + 0.631 + ], + "angle": 0, + "content": "- VoxCeleb2 [10] VoxCeleb [30] and VoxCeleb2 are large-scale audio-visual datasets collected for the speaker verification task. We re-download around one-fifth of VoxCeleb2 for training, which are of high quality. The official test set is processed with GPEN [53] for evaluation." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.466, + 0.892, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We process all videos at 25 fps and align all faces according to pre-detected landmarks at the eyes. All faces are cropped to the size of \\(256 \\times 256\\). A same U-shape mask is adopted as shown in Fig. 2 to erase the mouth, cheeks and jaws at the target frame. All audiois processed in the same way as Wav2Lip [32]. In order to keep our design simple and avoid extra hyperparameter tuning, we adopt most settings from the previous studies [23, 32, 53]. Specifically, the generator has a total of \\(L = 14\\) style-convolution layers. \\(\\lambda_{r}\\) is empirically set to 10 and all other \\(\\lambda s\\) are selected to 1. For personalized optimization, we train the person's data for 5 epochs. Long training would not lead to better results. Particularly, we find that training the personalized model together with the general training data leads to more stable results, the portion of personal data and general data is set as 1:1 in our experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1509" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.482, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.089, + 0.892, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.449, + 0.895, + 0.52 + ], + "angle": 0, + "content": "Figure 3. Qualitative Results. The top row shows the lip-synced videos of the conditional driving audio. MakeitTalk [60] fails to generate accurate mouth shape and lacks head dynamics. PC-AVS [59] and Wav2Lip [32] generate lip-synced results aligned well with audios. But they produce visibly blurry results. Particularly, our reproduced Wav2Lip-H generates high-quality results with good lip-sync. However, they still produce certain artifacts. While our generalized model directly produces high-fidelity results that have the same lip motion as the original syncsed video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.473, + 0.775 + ], + "angle": 0, + "content": "Comparing Methods. As our model is built based on the person-agnostic setting, we compare our StyleSync framework with three state-of-the-art methods including, MakeiTalk [60], PC-AVS [59] and Wav2Lip [32]. MakeiTalk [60] synthesizes talking head videos with natural head pose. PC-AVS [59] a pose source video to achieve pose-controllable talking face generation. While MakeiTalk and PC-AVS produce a whole talking head, Wav2Lip which focuses on the mouth part is our main competing method. As Wav2Lip is originally trained on a low-resolution with small networks, we carefully reproduce it on the same data and even loss functions as our methods. We denote this model as Wav2Lip-H (High-Quality). For fair comparison, all basic comparisons are carried out on the generalized setting Ours-G." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We also perform personalized optimization on the test set, where each frame lasts less than 10 seconds for both datasets. The results are denoted as Ours-P. We equally perform the generator finetuning on the above methods, and all of them perform worse than their original models, thus the results are not listed below. Additionally, we compare our personalized setting with ADNeRF [16], a recent talking head advance using neural radiance fields [29]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.722, + 0.562 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.645 + ], + "angle": 0, + "content": "As quantitative evaluations can only be performed on the self-construction setting, we avoid directly leveraging all information within the frame for reconstruction. We uniformly select a same random frame as the reference frame for PC-AVS, Wav2Lip, Wav2Lip-H and ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow previous studies [8, 9, 58, 59] to adopt the popularly used SSIM [49], PSNR metrics for evaluating the generation quality, and the landmark distances around the mouth (LMD) and SyncNet's [12] confidence score to show the lip-sync quality. Please be noted that the SyncNet metric is computed on the officially released version, which is different from our implementation in the loss functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Particularly, as the details of the target person cannot be fully recovered, we leverage the ArcFace [13] network and compute the frame-by-frame feature cosine distances \\(\\mathcal{D}_{ID}\\) for evaluating whether the generated results preserve the identity well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Results. The quantitative experiments are carried out on the test set of LRW [11] and VoxCeleb2 [10] datasets. Please refer to Table 1 for the results. It can be seen that on the LRW dataset, the generalized version" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1510" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.244 + ], + "angle": 0, + "content": "
MethodLRWVoxCeleb2
SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID↑SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID ↑
MakeiTalk0.6929.832.753.880.790.6328.386.942.150.71
PC-AVS0.7930.261.847.190.820.7129.532.758.160.74
Wav2Lip0.7930.541.287.390.900.8030.531.928.900.90
Wav2Lip-H0.8031.381.207.190.880.8130.531.878.350.90
GT1.00100.000.007.651.001.00100.000.007.711.00
Ours-G0.8531.781.187.250.890.7931.001.478.250.90
Ours-P0.8832.660.866.350.930.8231.541.157.260.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.253, + 0.867, + 0.268 + ], + "angle": 0, + "content": "Table 1. Quantitative results on LRW and VoxCeleb. For LMD the lower the better, and the higher the better for other metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.442 + ], + "angle": 0, + "content": "of our StyleSync already outperforms previous methods by a large margin for generation quality. While on the VoxCeleb2 dataset, the gap is less obvious. The reason might be that most training data of our methods are basically frontal view faces selected from LRW, while VoxCeleb2 contains more complicated scenes. On the other hand, after our personalized optimization, the performance of our model advances again. This shows the generated results are clearly more similar to the targets at this stage, which can also be verified by looking into the identity distances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.624 + ], + "angle": 0, + "content": "Meanwhile, our method also achieves comparable performance on the lip-sync metrics on both datasets. Our LMD score is slightly better than the competing methods. As for the SyncNet score, we achieve comparable results on LRW that are closer to the ground truth's SyncNet score. We argue it is meaningless to refer to the SyncNet scores for verifying the lip-sync quality once the metric has outperformed the ground truth. The \\( \\text{Sync}_{\\text{conf}} \\) only reflects how well an audio-visual pair fits the learned SyncNet model rather than the true perceptual quality. Thus though generated results might outperform ground truth on the metric, it does not mean better sync quality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.624, + 0.47, + 0.699 + ], + "angle": 0, + "content": "After the personalization, the lip-sync score degrades. We assume that the specificity of the speaking style reduces the mouth opening level. This is also shown in Fig. 5. The woman tends not to open her mouth wide, leading to smaller mouth movements" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.288, + 0.727 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Subjective evaluation is crucial in identifying the ability of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video at https://hangz-nju-cuhk.github.io/projects/StyleSync." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Two cross-driven examples and their comparisons with the SOTA methods are shown in Fig. 3. We use an conditional audio from an arbitrary person selected from the test set to drive the template. It can be seen that MakeiTalk [60] cannot produce accurate lip movements. Moreover, both PC-AVS and Wav2Lip produce visible artifacts or blurry" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.288, + 0.892, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.418, + 0.892, + 0.433 + ], + "angle": 0, + "content": "Figure 4. Self-driven results compared with personalized method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.515 + ], + "angle": 0, + "content": "results. Our reproduced Wav2Lip-H adopts the same training setting as ours, thus generates plausible results for most cases. Nevertheless, our generalized model still produces the most accurate mouth shapes with the highest fidelity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Additionally, we show a case of personalized optimization on a 50-second video clip in Fig. 1. It can be seen that our method preserves the speaking style of the target person with accurate lip sync. Comparing our personalized results (Ours-P) with AD-NeRF [16] (Fig. 4), it is evident that our method maintains more person-specific details with higher fidelity, even though we use significantly less personal data (10s for tuning v.s. 5min for training in [16])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.807 + ], + "angle": 0, + "content": "User Study. We also invite 15 participants to conduct a user study for further subjective evaluation and the results are reported in Table 2. Specifically, we randomly select 63 videos from LRW and VoxCeleb2 datasets as test sets and generate videos by using our generalized StyleSync and the comparison methods accordingly. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on three aspects for each generated video: (1) Lip-Sync quality; (2) Generation quality; (3) Video realness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 2, MakeitTalk [60] achieves the lowest scores in all the aspects due to lacking reasonable lip movements and head pose. PC-AVS [59], Wav2Lip [32] and Wav2Lip-H achieve relatively higher lip-sync score, but only Wav2Lip-H is able to synthesize videos with less blurry textures around the mouth. Overall, our StyleSync" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1511" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.187 + ], + "angle": 0, + "content": "
MOS on \\ ApproachMakeitTalkPCALPSWav2LipWav2Lip-HOurs-G
Lip-Sync Quality2.063.003.493.674.24
Generation Quality2.632.161.873.424.52
Video Realness1.892.162.172.984.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.197, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Table 2. User study measured by Mean Opinion Scores. The scores are ranged from 1 (worst) to 5 (best)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.233, + 0.466, + 0.324 + ], + "angle": 0, + "content": "
MethodSSIM ↑PSNR ↑LMD ↓Syncconf ↑LID ↑
w/o mask0.8131.201.807.890.90
w/o sync0.8031.341.557.410.90
Ours-G0.7931.001.478.250.90
P w/o ΔW0.8231.511.327.310.92
Ours-P0.8231.541.157.260.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.334, + 0.431, + 0.348 + ], + "angle": 0, + "content": "Table 3. Quantitative ablations results on VoxCeleb2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.398 + ], + "angle": 0, + "content": "outperforms its counterparts in all the three aspects by a large margin, indicating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.23, + 0.424 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.612 + ], + "angle": 0, + "content": "To further demonstrate the contributions of our novel designs, we perform an ablation study on VoxCeleb2 dataset under both generalized setting and personalized optimization setting (denoted as \"Ours-G\" and \"Ours-P\", respectively). Concretely, we construct two variants for \"Ours-G\" (denoted as \"w/o mask\" and \"w/o sync\") by removing the masking operation in Eq. 1 and the lip-sync loss during training, respectively. While for \"Ours-P\", we form one variant (denoted as \"Ours-P w/o \\(\\Delta \\mathbf{W}\\)) by setting \\(\\Delta \\mathbf{W} = \\mathbf{0}\\). We have also experimented personalized optimization with fixed generated, however, this would lead to blurry results. This ablation is omitted here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.81 + ], + "angle": 0, + "content": "The quantitative and qualitative results are shown in Table 3 and Fig. 5. As shown in Table 3, we observe that both \"w/o mask\" and \"w/o sync\" achieve comparable performance on the image quality metrics but leads to a performance drop on the lip-sync metrics. \"w/o mask\" achieves worse scores when compared with \"Ours-G\" in terms of LMD and Sync, which indicates that the proposed masking strategy can obviously alleviate the influence from the reference frames. While \"w/o sync\" reasonably suffers from lip-sync degradation and the results in Fig. 5 illustrate inconsistent lip movements when compared with the lip-synced video. The results demonstrate the effectiveness of the additional supervision from SyncNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "It can be seen from Fig. 5 that with the personalized optimization, both the identity and the pattern of the mouth opening on the generated frames become more similar to the original template video. It is also clear that after the personalized training, the mouth opening is less obvious than in the generalized model. We analyze that this is part of the" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.891, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.334, + 0.888, + 0.348 + ], + "angle": 0, + "content": "Figure 5. Ablation study with visual results. Zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.394 + ], + "angle": 0, + "content": "talking style of this target person. This also leads to poorer metric values on SyncNet." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.892, + 0.485 + ], + "angle": 0, + "content": "In terms of the comparisons between our personalized model and \"Ours-P w/o \\(\\Delta \\mathbf{W}\\)\", it achieves similar scores on the image quality as well as lip-sync quality. However, the identity score without \\(\\Delta \\mathbf{W}\\) is slightly degraded, which indicates that it is beneficial to involve \\(\\Delta \\mathbf{W}\\) into the training procedure for better preserving identity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.757, + 0.514 + ], + "angle": 0, + "content": "5. Conclusion and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Conclusion. In this paper, we propose StyleSync, which produces high-fidelity lip sync results for both the one-shot and the few-shot settings. We highlight the unique proprieties of our method: 1) Video results produced by our generalized model clearly outperform previous state-of-the-arts. 2) Our model is built upon the success of recent style-based generators with simple modification. It is easy to implement and friendly to train. 3) By involving StyleGAN-inspired personalized optimization procedure, our model can further be improved on a specific person given only a few clips." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Limitations. As our method blends a lip-synced face into an existing video with a fixed mask, the head pose and expressions of the target person cannot be changed. Additionally, under certain extreme cases where the target's jaw is extremely large, it might get out of our masked area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Ethical Considerations. Our method could be used to create non-existing talks and speeches, which might be maliciously used. We will issue our core models strictly to research institutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1512" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.471, + 0.441 + ], + "angle": 0, + "content": "[5] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.471, + 0.498 + ], + "angle": 0, + "content": "[6] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation?: A survey and benchmark. arXiv preprint arXiv:2005.03201, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.615, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[10] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Bo Fan, Lijuan Wang, Frank K Soong, and Lei Xie. Photoreal talking head with deep bidirectional LSTM. In 2015 IEEE" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4884-4888. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.124, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[15] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[16] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[17] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127(11):1767-1779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14080-14089, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[24] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2755-2764, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. ECCV, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1513" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[27] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[28] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[30] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[31] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[32] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[33] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.518, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.588, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.631, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[36] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.674, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[37] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[38] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. arXiv preprint arXiv:2001.05201, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[39] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[40] Yasheng Sun, Hang Zhou, Ziwei Liu, and Hideki Koike. Speech2talking-face: Inferring and driving a face with synchronized audio-visual representation. In *IJCAI*, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[42] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[43] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, pages 716-731. Springer, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[46] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[47] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[48] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[50] Haozhe Wu, Jia Jia, Haoyu Wang, Yishun Dou, Chao Duan, and Qingshan Deng. Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1478-1486, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[51] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[52] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[53] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1514" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "learning-based personalized head pose. arXiv e-prints, pages arXiv-2002, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[55] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[56] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[57] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3867-3876, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[58] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[59] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[61] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1515" + } + ] +] \ No newline at end of file diff --git a/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_origin.pdf b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32c1a3d501f1288109d058c0c69a8e15c350239e --- /dev/null +++ b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/16a4d87e-841b-4c93-8d78-c26bcd8dff23_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2aeddc5ebbb0e6bbb7f8df3bd8264a25c9c32dce303f3ebcbb611f7c465e552 +size 5402436 diff --git a/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/full.md b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ba0b4c4e09d857880cbc7555ccbe6792e683364 --- /dev/null +++ b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/full.md @@ -0,0 +1,306 @@ +# StyleSync: High-Fidelity Generalized and Personalized Lip Sync in Style-based Generator + +Jiazhi Guan $^{1,2*}$ Zhanwang Zhang $^{1*}$ Hang Zhou $^{1\dagger}$ Tianshu Hu $^{1\dagger}$ Kaisiyuan Wang $^{3}$ Dongliang He $^{1}$ Haocheng Feng $^{1}$ Jingtuo Liu $^{1}$ Errui Ding $^{1}$ Ziwei Liu $^{4}$ Jingdong Wang $^{1}$ $^{1}$ Department of Computer Vision Technology (VIS), Baidu Inc. $^{2}$ Tsinghua University $^{3}$ The University of Sydney $^{4}$ S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn {zhangzhanwang,zhouhang09,hutianshu01}@baidu.com + +![](images/220009c78f4595caa6657435d0e5f7944517bc63bd4fa6a9a05966715de9482e.jpg) +Figure 1. Personalized lip-sync results generated by our StyleSync framework. Our method not only supports high-fidelity modification to any target template video according to conditional audio but can further adapt to specific styles with personalized optimization. In this figure, our lip-sync results should have the same mouth shapes as the lip-synced video of the conditional audio. + +# Abstract + +Despite recent advances in syncing lip movements with any audio waves, current methods still struggle to balance generation quality and the model's generalization ability. Previous studies either require long-term data for training or produce a similar movement pattern on all subjects with low quality. In this paper, we propose StyleSync, an effective framework that enables high-fidelity lip synchronization. We identify that a style-based generator would sufficiently enable such a charming property on both one-shot + +and few-shot scenarios. Specifically, we design a mask-guided spatial information encoding module that preserves the details of the given face. The mouth shapes are accurately modified by audio through modulated convolutions. Moreover, our design also enables personalized lip-sync by introducing style space and generator refinement on only limited frames. Thus the identity and talking style of a target person could be accurately preserved. Extensive experiments demonstrate the effectiveness of our method in producing high-fidelity results on a variety of scenes. Resources can be found at https://hangz-nju-cuhk.github.io/projects/StyleSync. + +# 1. Introduction + +The problem of generating lip-synced videos according to conditional audio is of great importance to the field of digital human creation, audio dubbing, film-making, and entertainment. While the rapid development of this area has been witnessed within recent years, most methods [6,8,9,16,19,24,27,28,38,39,46,50,54,57-60] focus on generating a whole dynamic talking head. Results created under such settings can hardly be blended into an existing scene. Under real-world scenarios like audio dubbing, one crucial need is to seamlessly alter the mouth or facial area while preserving other parts of the scene unchanged, making these methods non-feasible. + +Previous methods take two different paths for achieving seamless mouth modification. A number of studies [38, 44] pursue realistic results on person-specific settings, which require long-term clips for target modeling. Moreover, they rely on prior 3D facial structural information. The uncertainty and errors accumulated in the 3D fitting procedure would greatly influence their performances. On the other hand, it is desired to build models that break the data limitation on more generalized scenes. As a result, a few methods [31, 32, 41] design person-agnostic models without relying on 3D or 2D structural priors. Nevertheless, such a setting is extremely challenging. + +In order to produce high-fidelity lip-synced results on any-length videos, two essential challenges need to be addressed. 1) How to efficiently design a powerful generative backbone network that supports both accurate audio information expression and seamless local alternation. Intuitively, the lip-sync quality naturally contradicts the preservation of the original target frame information [32, 59]. 2) How to effectively leverage as much provided information as possible and involve the personalized properties to a generalized model. Though few-shot meta-learning has been proven effective in generating talking heads [5, 7, 56, 60], how to involve such ability into a lip-syncing pipeline has not been explored. + +In this paper, we propose a highly concise and comprehensive framework named StyleSync, which produces high-fidelity lip-sync results on both generalized and personalized scenarios. The key is our simple but lip-sync-oriented modifications to style-based generator. Though style-based generators [22, 23] have been leveraged in various talking head generation methods [5, 55, 59], their successes are only partially instructive. They aim at producing the whole head, which leads to unstable background and distortions which are non-acceptable in our scenarios. + +By revisiting the details of style-based generators, we identify a few simple but essential modifications that make our framework suitable for lip-syncing. Different from the above methods, we adopt a masked mouth modeling protocol [32, 41] and delicately design a Mask-based Spa + +tial Information Encoding strategy, where both the target and reference frames' information is encoded into a noise space [52, 53] of the generator according to different masking schemes. While the information on audio dynamics and high-level reference frame is injected into the style-modulated convolution in a similar manner as [25, 59]. In this way, our method can be benefited from the strong generative power of style-based generators and also keeps the advantage of easy implementation and fast training. + +Moreover, our network modification enables personalized information preserving (e.g., speaking styles and details of the mouth and jaw). We take inspiration from the recent studies of inverting StyleGAN priors [1, 2, 35, 45] and propose a Personalized Optimization scheme. As audio dubbing is normally performed on speaking videos, our model can make use of only a few seconds of the person's information and optimize additional person-specific parameters including the $W^{+}$ and the generator. Extensive experiments show that our framework clearly outperforms previous state of the arts on the one-shot setting by a large margin, and the target-specific optimization further enhances the fidelity of our results. + +Our contributions can be summarized as follows: 1) We present the StyleSync framework, which adopts simple but effective modifications including the Mask-based Spatial Information Encoding to a style-based generator. 2) We propose the Personalized Optimization procedure which involves few-shot person-specific optimization into our framework. 3) Extensive experiments demonstrate that our framework can directly produce accurate and high-fidelity one-shot lip-sync results. Moreover, our proposed personalized optimization further improves the generation quality. Our method outperforms previous methods by a clear margin. + +# 2. Related Work + +# 2.1. Audio-Driven Facial Animation + +The topic of audio-driven facial animation has long been a research interest in both the computer vision and graphics community. Studies have been carried out on both digital 3D human faces [15, 20, 33, 61] and realistic human heads [14, 17, 42]. We focus on human heads in the real world. + +Talking Head Generation. Most studies on lip-syncing target generating the whole head of a talking person [6, 8, 9, 16-18, 25, 26, 40, 42, 54, 58-60]. Specifically, a number of studies leverage structural information such as 2D [9] landmarks, 3D landmarks [60] and 3D meshes [6]. The uncertainty and inaccuracy of such representations would lead to error accumulation during the talking head generation procedure. Most person-specific methods [19, 27, 42] rely on them and produce results with a poor generalization or lip + +sync quality. Recently, NeRF is also used in person-specific modeling [16, 26, 43], but they also perform poorly when driven by audio from a different person. + +Lip-Syncing on Faces. The other type of study focuses on lip-syncing the mouth part while keeping other information untouched in videos [31, 32, 39, 44]. Our work lies within the same scope. While Thies et al. [44] and Song et al. [38] produce realistic results, they rely on person-specific training on more than 2 minutes of data. + +Specifically, Wav2Lip [32] generates person-agnostic lip-sync results that are highly accurate. However, they build the generative model on low-resolution images, leading to blurry results. We identify that it would be easier for the model to learn the mouth motions if less information about the image quality needs to be recovered. Moreover, their methods cannot capture the personalized patterns given the target template video. + +# 2.2. Style-based Generator for Faces + +StyleGAN models [21-23] have shown great success on image generation tasks, particularly on facial image generation and editing [1, 2, 34, 36, 37, 45]. + +StyleGAN Inversion. StyleGAN inversion is the practice of recovering the latent space of a given image with a pretrained StyleGAN. Abdal et al. [1, 2] propose to invert images not only using the style space $W$ , but also expand it to the $\mathcal{W}^+$ , which better preserves details. Recent studies propose to learn encoders [3, 34, 45] for better editability and faster optimization results. Furthermore, recent studies try to refine the generator's parameters through pivotal finetuning [35]. In our work, we take inspiration of previous StyleGAN inversion studies and propose a personalized optimization procedure. We encode a $\mathcal{W}^+$ space from reference frames following [45] which enables personalized lip movements learning. + +Applications. StyleGAN architecture has also been leveraged in face restorations [48,53] and face swapping [51,52]. These tasks require preserving the original facial emotion and expressions. On the other hand, style-based generators have also been applied to create facial animations [5,25,59], which are highly related to our task. However, they rely on the style vectors in the $W$ or $W^{+}$ space for controlling both the appearances and motion dynamics. One major drawback of this setting is that $W^{+}$ space cannot easily accounts for the spatial consistency of backgrounds, leading to nonrealistic results or visible artifacts. + +# 3. Methodology + +In this section, we introduce the details of our StyleSync framework as depicted in Fig. 2. It is simply designed by leveraging several successes of previous style-based generators. We introduce our modifications to the style-based generator that make a successful backbone in Sec. 3.1 + +and the training objectives of our generalized model in Sec. 3.2. Specifically, we illustrate our Personalized Optimization procedure which further strengthens our results with person-specific optimization in Sec. 3.3. + +Task Formulation. Our goal is to sync the lip motion of a target person with any provided audio and seamlessly blend it into the original target video. We formulate our training in a typical self-reconstruction manner. The training setting is similar to Wav2Lip [32]. Given a target video $\mathbf{V} = \{I_1, \dots, I_T\}$ with $T$ frames, we mask out the lower half of the face including the jaw and cheeks with a mask $M$ . The goal is to recover the whole face with its corresponding audio $\mathbf{a} = \{a_1, \dots, a_T\}$ ( $\mathbf{a}$ is processed to spectrograms). As no information about the mouth and jaw shape is provided on the masked target frame $I_t^m = (\mathbf{1} - M)*I_t$ , we leverage a random reference frame $I_{ref} \in \mathbf{V}$ during training to provide the desired context. + +# 3.1. Modifying Style-based Generator for Lip Sync + +Style-based Generator Overview. We perform lip-sync-oriented modifications to build a successful backbone for lip sync with the StyleGAN2 [23] architecture. The original StyleGAN2 takes a style vector $w$ as input and uses it to modulate convolution operations in a total of $L$ generative layers during training. During inference, different $w$ s can be used at different layers, formulating the $W^{+}$ space $\mathbf{W} = \{w_{1},\dots ,w_{L}\}$ . The basic elements of the original StyleGAN2, including the $W^{+}$ space and the StyleGAN generator G's layers are depicted in blue on Fig. 2. For simplification, certain detailed structures are omitted. + +Mask-based Spatial Information Encoding. Our goal is to seamlessly blend a lip-synced mouth into the target frame with the assistance of a reference image. As discussed in Sec. 2.2, previous methods leveraging the style-based generator are not directly applicable to our scene, we seek a different way to encode the spatial information on faces. + +It is proven by recent studies on face restoration [53] and swapping [52] that content-guided feature maps can be attached to StyleGAN layers in a similar status as noise without affecting the expressive power of the generative model. These contextual-rich noises $\mathbf{N} = \{N_1,\dots ,N_L\}$ could well preserve the original spatial structure and attributes of the encoded visual information. This property provides an option for leveraging the attributes information lying in the reference image. Thus following Wav2Lip [32], we concatenate the masked and reference input together as the visual input of the framework and encode the corresponding feature maps $\mathbf{F}_f = \{F_1,\ldots ,F_L\}$ . The simplest setting is to regard $\mathbf{N} = \mathbf{F}$ . + +However, the goal of generating lip motions needs to create dynamics that are different from the original facial structure but keeps facial identity, which is substantially different from previous methods that aim to keep the structure + +![](images/70c1b84618322593295d202b149ded560bde30d5279920e2a606abd7ec617b7c.jpg) +Figure 2. Our StyleSync framework. The building blocks in Blue indicate the style-based generator. The masked target frame $I_{t}^{m}$ is concatenated with a reference $I_{ref}$ and encoded to $f_{f}$ by $\mathrm{E}_{\text{face}}(\mathbf{Red})$ . The audio information is encoded to an audio feature $f_{a}$ (Yellow). The features are concatenated together to form the style $(W)$ space. Specifically, we devise a Personalized Optimization procedure (Green) including the learning of the $\Delta \mathbf{W}$ and the $\Delta \mathbf{P}$ . This part is not trained during the initial backbone learning. + +fixed. The mouth shapes would be greatly influenced by the reference image $I_{ref}$ when adopting the same protocol. We identify that more than desired facial structure information is infused in the low-level information of $\mathbf{F}$ (layers with higher resolutions). As a result, we intuitively mask out the low-level part of the information in the face features. Here we simply define + +$$ +F _ {l} ^ {\prime} = (\mathbf {1} - M) * F _ {l}, \text {f o r} l > \left\lfloor \frac {L}{2} \right\rfloor . \tag {1} +$$ + +The facial information is colored in red on Fig. 2. The 1 is an all-one matrix that has the same size as $M$ . + +Style Information Encoding. We follow previous studies [5, 40, 59] to encode both audio dynamics and facial information into the style space ( $W$ space) of the style-based generator. The audio feature $f_{a}$ is encoded from the encoder $E_{a}$ and the face feature $f_{f}$ is the bottleneck of the face encoder $E_{face}, w = \mathrm{concat}(f_{a}, f_{f})$ . As the information in the reference frames is already fused to the generator through spatial noise encoding, the fusing of the face feature is less necessary. Experiments show that the difference is subtle with or without $f_{f}$ , but we still keep this design in accordance with common practice. + +Ingredients for Personalization. Our above designs are not only simple but also possess external potentials of personalized lip-sync modeling. We leave the details to Sec. 3.3 and briefly discuss the additional ingredients that make our model comprehensive. + +Specifically, our extension is inspired by the advances in StyleGAN inversion [1, 2, 35]. It has pointed out that extending $w$ to the $W^{+}$ space which contains $\mathbf{W} = \{w + \Delta w_{1},\dots ,w + \Delta w_{L}\}$ is essential to recovering specific images. Moreover, recent studies [4, 35] also explore tuning + +the parameters of the generator to improve the network's fitting ability on a specific target. Thus our generator has the potential of learning limited parameter shifts $\Delta \mathbf{P}$ in order to fit a specific person. + +These personalized ingredients are depicted in green on Fig. 2 and are not optimized during the generalized training procedure. In order to keep the pipeline consistent across both settings, we set $\Delta \mathbf{W} = \mathbf{0}$ and $\Delta \mathbf{P} = \mathbf{0}$ in our generalized model. + +# 3.2. Backbone Training Objectives + +During training, the whole backbone networks take the 6-channel concatenation of $[I_t^m, I_{ref}]$ and the audio clip $a_t$ at the same time step as input. It predicts $I_t' = \mathrm{G}(\mathbf{N}, \mathbf{W})$ , which aims at recovering the unmasked target image $I_t$ . In order to keep our design simple, the training objectives are mostly aligned with StyleGAN2 [23] and Wav2Lip [32]. + +Reconstruction Loss. The pixel-level reconstruction loss $\mathcal{L}_{rec}$ is fundamental for training our task. Here we leverage the commonly used $L_{1}$ loss and perceptual loss [47]. + +$$ +\mathcal {L} _ {r e c} = \left\| I _ {t} ^ {\prime} - I _ {t} \right\| _ {1} + \sum_ {m = 1} ^ {N _ {v g g}} \left\| \mathrm {V G G} _ {m} \left(I _ {t} ^ {\prime}\right) - \mathrm {V G G} _ {m} \left(I _ {t}\right) \right\| _ {1}, \tag {2} +$$ + +where $\mathrm{VGG}_m$ is the $m$ th layer's output of a pre-trained VGG19 network. + +Adversarial Loss. We directly adopt the same discriminator D from StyleGAN2 [23] for the adversarial training: + +$$ +\mathcal {L} _ {a d v} = \min _ {\mathrm {G}} \max _ {\mathrm {D}} \left(\mathbb {E} _ {I _ {t}} \left[ \log \mathrm {D} \left(I _ {t}\right) \right] + \mathbb {E} _ {I _ {t} ^ {\prime}} \left[ \log \left(1 - \mathrm {D} \left(I _ {t} ^ {\prime}\right)\right) \right]\right). \tag {3} +$$ + +Particularly, we initialize the discriminator's weights with the pre-trained version in StyleGAN2 [23]. This practice produces visibly sharper results. + +Lip-Sync Loss. We additionally train a SyncNet [12] which consists of a visual encoder $S_v$ and an audio encoder $S_a$ . It identifies whether visual and audio clips are timely aligned with contrastive loss. When training the generator, we predict 5 consecutive frames $I_{t:t+4}'$ within one batch at each inference step and supervise the training with SyncNet's assistance. The objective is: + +$$ +\mathcal {L} _ {\text {s y n c}} = - \frac {\mathrm {S} _ {v} \left(I _ {t : t + 4} ^ {\prime}\right) ^ {\mathrm {T}} \cdot \mathrm {S} _ {a} \left(a _ {t : t + 4}\right)}{\| \mathrm {S} _ {v} \left(I _ {t : t + 4} ^ {\prime}\right) \| _ {2} \| \mathrm {S} _ {a} \left(a _ {t : t + 4}\right) \| _ {2}} \tag {4} +$$ + +The overall loss functions across the generalized training can be written as: + +$$ +\mathcal {L} _ {g} = \mathcal {L} _ {a d v} + \lambda_ {r} \mathcal {L} _ {r e c} + \lambda_ {s} \mathcal {L} _ {\text {s y n c}}. \tag {5} +$$ + +# 3.3. Personalized Optimization + +After the training of the backbone networks, our method can readily generate high-fidelity lip sync results for arbitrary subject. However, as the model is trained in a generalized manner, the generated mouth motion patterns across different people are basically the same. It has been verified that different identities possess different talking styles [50, 54]. Here we pursue to capture such personalized property with the original template video. Notably, we focus on the few-shot setting that only less than one minute of the original video is given, which cannot be handled by previous person-specific models [16, 19, 38, 44]. Below, we illustrate how we successfully design our personalized optimization module based on the above discussions. + +Basic Learning Settings. The basic learning procedure of the personalized optimization is similar to backbone training. On a template video $\bar{\mathbf{V}} = \{\bar{I}_1,\dots ,\bar{I}_T\}$ with audio $\bar{\mathbf{a}} = \{\bar{a}_1\dots ,\bar{a}_T\}$ , we randomly select two frames as target $\bar{I}_t$ and reference $\bar{I}_{ref}$ for training as one sample. During training, the generalized loss $\mathcal{L}_g$ continuously supervises the whole procedure. + +As the encoder $\mathrm{E}_a$ encodes person-agnostic speech content information, it is fixed during the personalized optimization step. While the face encoder $\mathrm{E}_{face}$ encodes the attribute information of the target person, we do not optimize its parameters on a single person to avoid overfitting. + +Style Space Optimization. According to our formulation, the $W^{+}$ space contains both the lip motion and the high-level facial style information. Thus optimizing it would ideally lead to person-specific lip motion pattern. E4E [45] verifies that learning the small displacements around the original $W$ space would both extend the inversion quality of the image and enables strong editing ability. Thus we propose + +to learn a set of $\Delta \mathbf{W} = MLP_{\Delta w}(f_f)$ in a similar way according to the encoded face feature $f_{f}$ with a few layers of MLPs $MLP_{\Delta w}$ . + +Generator Tuning. The generator most accounts for the synthesizing and the blending quality. We allow the parameters $\mathbf{P}$ of our generator $\mathbf{G}$ to shift a little margin to $\Delta \mathbf{P}$ with the personalized data we use. + +Learning Objectives. The final learning objectives of the personalized optimization procedure can be summarized as: + +$$ +\mathcal {L} _ {p} = \mathcal {L} _ {g} + \lambda_ {p} \left(\sum_ {i, j} \left| \Delta \mathbf {W} _ {i} ^ {j} \right| _ {2} ^ {2} + \sum_ {m, n} \left| \Delta \mathbf {P} _ {n} ^ {m} \right| _ {2} ^ {2}\right), \tag {6} +$$ + +where $\mathbf{W}_i^j$ denote the $j$ th parameter of the $i$ th element in $\mathbf{W}$ . This stands the same for $\mathbf{P}_n^m$ . We restrict the displacements in a limited step. + +# 4. Experiments + +Datasets. We conduct experiments on two commonly used audio-visual datasets, LRW [11] and VoxCeleb2 [10]. We follow the datasets' original train/test split and train our model on a mixture of these two datasets. + +- LRW [11] is an audio-visual dataset collected from BBC news for lip reading. It is one of the earliest large-scale audio-visual datasets with high quality. It consists of 1,000 one-second utterances in 500 words. +- VoxCeleb2 [10] VoxCeleb [30] and VoxCeleb2 are large-scale audio-visual datasets collected for the speaker verification task. We re-download around one-fifth of VoxCeleb2 for training, which are of high quality. The official test set is processed with GPEN [53] for evaluation. + +Implementation Details. We process all videos at 25 fps and align all faces according to pre-detected landmarks at the eyes. All faces are cropped to the size of $256 \times 256$ . A same U-shape mask is adopted as shown in Fig. 2 to erase the mouth, cheeks and jaws at the target frame. All audiois processed in the same way as Wav2Lip [32]. In order to keep our design simple and avoid extra hyperparameter tuning, we adopt most settings from the previous studies [23, 32, 53]. Specifically, the generator has a total of $L = 14$ style-convolution layers. $\lambda_{r}$ is empirically set to 10 and all other $\lambda s$ are selected to 1. For personalized optimization, we train the person's data for 5 epochs. Long training would not lead to better results. Particularly, we find that training the personalized model together with the general training data leads to more stable results, the portion of personal data and general data is set as 1:1 in our experiments. + +![](images/d8e4fea59578dd2f24612c242704a827769897df1dd075646a2915ef935cf337.jpg) +Figure 3. Qualitative Results. The top row shows the lip-synced videos of the conditional driving audio. MakeitTalk [60] fails to generate accurate mouth shape and lacks head dynamics. PC-AVS [59] and Wav2Lip [32] generate lip-synced results aligned well with audios. But they produce visibly blurry results. Particularly, our reproduced Wav2Lip-H generates high-quality results with good lip-sync. However, they still produce certain artifacts. While our generalized model directly produces high-fidelity results that have the same lip motion as the original syncsed video. + +![](images/1ebf1340171f28ebeef39bba2ffb50bbcd782769173a93cf2f61fa754bd66a8e.jpg) + +Comparing Methods. As our model is built based on the person-agnostic setting, we compare our StyleSync framework with three state-of-the-art methods including, MakeiTalk [60], PC-AVS [59] and Wav2Lip [32]. MakeiTalk [60] synthesizes talking head videos with natural head pose. PC-AVS [59] a pose source video to achieve pose-controllable talking face generation. While MakeiTalk and PC-AVS produce a whole talking head, Wav2Lip which focuses on the mouth part is our main competing method. As Wav2Lip is originally trained on a low-resolution with small networks, we carefully reproduce it on the same data and even loss functions as our methods. We denote this model as Wav2Lip-H (High-Quality). For fair comparison, all basic comparisons are carried out on the generalized setting Ours-G. + +We also perform personalized optimization on the test set, where each frame lasts less than 10 seconds for both datasets. The results are denoted as Ours-P. We equally perform the generator finetuning on the above methods, and all of them perform worse than their original models, thus the results are not listed below. Additionally, we compare our personalized setting with ADNeRF [16], a recent talking head advance using neural radiance fields [29]. + +# 4.1.Quantitative Evaluation + +As quantitative evaluations can only be performed on the self-construction setting, we avoid directly leveraging all information within the frame for reconstruction. We uniformly select a same random frame as the reference frame for PC-AVS, Wav2Lip, Wav2Lip-H and ours. + +Evaluation Metrics. We follow previous studies [8, 9, 58, 59] to adopt the popularly used SSIM [49], PSNR metrics for evaluating the generation quality, and the landmark distances around the mouth (LMD) and SyncNet's [12] confidence score to show the lip-sync quality. Please be noted that the SyncNet metric is computed on the officially released version, which is different from our implementation in the loss functions. + +Particularly, as the details of the target person cannot be fully recovered, we leverage the ArcFace [13] network and compute the frame-by-frame feature cosine distances $\mathcal{D}_{ID}$ for evaluating whether the generated results preserve the identity well. + +Evaluation Results. The quantitative experiments are carried out on the test set of LRW [11] and VoxCeleb2 [10] datasets. Please refer to Table 1 for the results. It can be seen that on the LRW dataset, the generalized version + +
MethodLRWVoxCeleb2
SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID↑SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID ↑
MakeiTalk0.6929.832.753.880.790.6328.386.942.150.71
PC-AVS0.7930.261.847.190.820.7129.532.758.160.74
Wav2Lip0.7930.541.287.390.900.8030.531.928.900.90
Wav2Lip-H0.8031.381.207.190.880.8130.531.878.350.90
GT1.00100.000.007.651.001.00100.000.007.711.00
Ours-G0.8531.781.187.250.890.7931.001.478.250.90
Ours-P0.8832.660.866.350.930.8231.541.157.260.93
+ +Table 1. Quantitative results on LRW and VoxCeleb. For LMD the lower the better, and the higher the better for other metrics. + +of our StyleSync already outperforms previous methods by a large margin for generation quality. While on the VoxCeleb2 dataset, the gap is less obvious. The reason might be that most training data of our methods are basically frontal view faces selected from LRW, while VoxCeleb2 contains more complicated scenes. On the other hand, after our personalized optimization, the performance of our model advances again. This shows the generated results are clearly more similar to the targets at this stage, which can also be verified by looking into the identity distances. + +Meanwhile, our method also achieves comparable performance on the lip-sync metrics on both datasets. Our LMD score is slightly better than the competing methods. As for the SyncNet score, we achieve comparable results on LRW that are closer to the ground truth's SyncNet score. We argue it is meaningless to refer to the SyncNet scores for verifying the lip-sync quality once the metric has outperformed the ground truth. The $\text{Sync}_{\text{conf}}$ only reflects how well an audio-visual pair fits the learned SyncNet model rather than the true perceptual quality. Thus though generated results might outperform ground truth on the metric, it does not mean better sync quality. + +After the personalization, the lip-sync score degrades. We assume that the specificity of the speaking style reduces the mouth opening level. This is also shown in Fig. 5. The woman tends not to open her mouth wide, leading to smaller mouth movements + +# 4.2. Qualitative Evaluation + +Subjective evaluation is crucial in identifying the ability of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video at https://hangz-nju-cuhk.github.io/projects/StyleSync. + +Two cross-driven examples and their comparisons with the SOTA methods are shown in Fig. 3. We use an conditional audio from an arbitrary person selected from the test set to drive the template. It can be seen that MakeiTalk [60] cannot produce accurate lip movements. Moreover, both PC-AVS and Wav2Lip produce visible artifacts or blurry + +![](images/b43d453f1557b4c49c6416fe4aeebf4bcea100aa2bd18d7de9df88488b251f3e.jpg) +Figure 4. Self-driven results compared with personalized method. + +results. Our reproduced Wav2Lip-H adopts the same training setting as ours, thus generates plausible results for most cases. Nevertheless, our generalized model still produces the most accurate mouth shapes with the highest fidelity. + +Additionally, we show a case of personalized optimization on a 50-second video clip in Fig. 1. It can be seen that our method preserves the speaking style of the target person with accurate lip sync. Comparing our personalized results (Ours-P) with AD-NeRF [16] (Fig. 4), it is evident that our method maintains more person-specific details with higher fidelity, even though we use significantly less personal data (10s for tuning v.s. 5min for training in [16]). + +User Study. We also invite 15 participants to conduct a user study for further subjective evaluation and the results are reported in Table 2. Specifically, we randomly select 63 videos from LRW and VoxCeleb2 datasets as test sets and generate videos by using our generalized StyleSync and the comparison methods accordingly. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on three aspects for each generated video: (1) Lip-Sync quality; (2) Generation quality; (3) Video realness. + +As shown in Table 2, MakeitTalk [60] achieves the lowest scores in all the aspects due to lacking reasonable lip movements and head pose. PC-AVS [59], Wav2Lip [32] and Wav2Lip-H achieve relatively higher lip-sync score, but only Wav2Lip-H is able to synthesize videos with less blurry textures around the mouth. Overall, our StyleSync + +
MOS on \ ApproachMakeitTalkPCALPSWav2LipWav2Lip-HOurs-G
Lip-Sync Quality2.063.003.493.674.24
Generation Quality2.632.161.873.424.52
Video Realness1.892.162.172.984.06
+ +Table 2. User study measured by Mean Opinion Scores. The scores are ranged from 1 (worst) to 5 (best). + +
MethodSSIM ↑PSNR ↑LMD ↓Syncconf ↑LID ↑
w/o mask0.8131.201.807.890.90
w/o sync0.8031.341.557.410.90
Ours-G0.7931.001.478.250.90
P w/o ΔW0.8231.511.327.310.92
Ours-P0.8231.541.157.260.93
+ +Table 3. Quantitative ablations results on VoxCeleb2. + +outperforms its counterparts in all the three aspects by a large margin, indicating the effectiveness of our approach. + +# 4.3. Ablation Study + +To further demonstrate the contributions of our novel designs, we perform an ablation study on VoxCeleb2 dataset under both generalized setting and personalized optimization setting (denoted as "Ours-G" and "Ours-P", respectively). Concretely, we construct two variants for "Ours-G" (denoted as "w/o mask" and "w/o sync") by removing the masking operation in Eq. 1 and the lip-sync loss during training, respectively. While for "Ours-P", we form one variant (denoted as "Ours-P w/o $\Delta \mathbf{W}$ ) by setting $\Delta \mathbf{W} = \mathbf{0}$ . We have also experimented personalized optimization with fixed generated, however, this would lead to blurry results. This ablation is omitted here. + +The quantitative and qualitative results are shown in Table 3 and Fig. 5. As shown in Table 3, we observe that both "w/o mask" and "w/o sync" achieve comparable performance on the image quality metrics but leads to a performance drop on the lip-sync metrics. "w/o mask" achieves worse scores when compared with "Ours-G" in terms of LMD and Sync, which indicates that the proposed masking strategy can obviously alleviate the influence from the reference frames. While "w/o sync" reasonably suffers from lip-sync degradation and the results in Fig. 5 illustrate inconsistent lip movements when compared with the lip-synced video. The results demonstrate the effectiveness of the additional supervision from SyncNet. + +It can be seen from Fig. 5 that with the personalized optimization, both the identity and the pattern of the mouth opening on the generated frames become more similar to the original template video. It is also clear that after the personalized training, the mouth opening is less obvious than in the generalized model. We analyze that this is part of the + +![](images/1435fc58b50a1e4bab9eae4c4538d0fe2746fae5050c185db0d51b432bf3a94c.jpg) +Figure 5. Ablation study with visual results. Zoom in for details. + +talking style of this target person. This also leads to poorer metric values on SyncNet. + +In terms of the comparisons between our personalized model and "Ours-P w/o $\Delta \mathbf{W}$ ", it achieves similar scores on the image quality as well as lip-sync quality. However, the identity score without $\Delta \mathbf{W}$ is slightly degraded, which indicates that it is beneficial to involve $\Delta \mathbf{W}$ into the training procedure for better preserving identity. + +# 5. Conclusion and Discussions + +Conclusion. In this paper, we propose StyleSync, which produces high-fidelity lip sync results for both the one-shot and the few-shot settings. We highlight the unique proprieties of our method: 1) Video results produced by our generalized model clearly outperform previous state-of-the-arts. 2) Our model is built upon the success of recent style-based generators with simple modification. It is easy to implement and friendly to train. 3) By involving StyleGAN-inspired personalized optimization procedure, our model can further be improved on a specific person given only a few clips. + +Limitations. As our method blends a lip-synced face into an existing video with a fixed mask, the head pose and expressions of the target person cannot be changed. Additionally, under certain extreme cases where the target's jaw is extremely large, it might get out of our masked area. + +Ethical Considerations. Our method could be used to create non-existing talks and speeches, which might be maliciously used. We will issue our core models strictly to research institutions. + +Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2, 3, 4 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 3, 4 +[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3 +[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 4 +[5] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2, 3, 4 +[6] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation?: A survey and benchmark. arXiv preprint arXiv:2005.03201, 2020. 2 +[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2 +[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2, 6 +[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 2, 6 +[10] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 5, 6 +[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016. 5, 6 +[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016. 5, 6 +[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 6 +[14] Bo Fan, Lijuan Wang, Frank K Soong, and Lei Xie. Photoreal talking head with deep bidirectional LSTM. In 2015 IEEE + +International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4884-4888. IEEE, 2015. 2 +[15] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2 +[16] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 5, 6, 7 +[17] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127(11):1767-1779, 2019. 2 +[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH, 2022. 2 +[19] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14080-14089, 2021. 2, 5 +[20] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2 +[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3 +[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 3 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2, 3, 4, 5 +[24] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2755-2764, 2021. 2 +[25] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2, 3 +[26] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. ECCV, 2022. 2, 3 + +[27] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2 +[28] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. AAAI, 2023. 2 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 6 +[30] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 5 +[31] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence, 2022. 2, 3 +[32] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 4, 5, 6, 7 +[33] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 3 +[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 4 +[36] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 3 +[37] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3 +[38] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. arXiv preprint arXiv:2001.05201, 2020. 2, 3, 5 +[39] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018. 2, 3 +[40] Yasheng Sun, Hang Zhou, Ziwei Liu, and Hideki Koike. Speech2talking-face: Inferring and driving a face with synchronized audio-visual representation. In *IJCAI*, 2021. 2, 4 +[41] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by + +audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[42] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2 +[43] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3 +[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, pages 716-731. Springer, 2020. 2, 3, 5 +[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 2, 3, 5 +[46] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2 +[47] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 4 +[48] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[50] Haozhe Wu, Jia Jia, Haoyu Wang, Yishun Dou, Chao Duan, and Qingshan Deng. Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1478-1486, 2021. 2, 5 +[51] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 3 +[52] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677. Springer, 2022. 2, 3 +[53] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 2, 3, 5 +[54] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with + +learning-based personalized head pose. arXiv e-prints, pages arXiv-2002, 2020. 2, 5 +[55] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. ECCV, 2022. 2 +[56] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2 +[57] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3867-3876, 2021. 2 +[58] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 2, 6 +[59] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 3, 4, 6, 7 +[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 6, 7 +[61] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2 \ No newline at end of file diff --git a/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/images.zip b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db30ff105f224fa0345200a68dd0b44e82fc4331 --- /dev/null +++ b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44706bdaa8a8aa0b34851f130e504c057bebf48f95f388236699c0a4db2316c7 +size 681321 diff --git a/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/layout.json b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..501c03430fb00b9a0927d9fe95faba618811d5a9 --- /dev/null +++ b/2023/StyleSync_ High-Fidelity Generalized and Personalized Lip Sync in Style-Based Generator/layout.json @@ -0,0 +1,8813 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 138 + ], + "type": "text", + "content": "StyleSync: High-Fidelity Generalized and Personalized Lip Sync in Style-based Generator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": "Jiazhi Guan" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Zhanwang Zhang" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Hang Zhou" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Tianshu Hu" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Kaisiyuan Wang" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Dongliang He" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Haocheng Feng" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Jingtuo Liu" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Errui Ding" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": " Jingdong Wang" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": "Department of Computer Vision Technology (VIS), Baidu Inc. " + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": "The University of Sydney " + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 58, + 160, + 534, + 232 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University guanjz20@mails.tsinghua.edu.cn {zhangzhanwang,zhouhang09,hutianshu01}@baidu.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 239, + 533, + 497 + ], + "blocks": [ + { + "bbox": [ + 58, + 239, + 533, + 497 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 533, + 497 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 533, + 497 + ], + "type": "image", + "image_path": "220009c78f4595caa6657435d0e5f7944517bc63bd4fa6a9a05966715de9482e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 503, + 547, + 539 + ], + "lines": [ + { + "bbox": [ + 46, + 503, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 547, + 539 + ], + "type": "text", + "content": "Figure 1. Personalized lip-sync results generated by our StyleSync framework. Our method not only supports high-fidelity modification to any target template video according to conditional audio but can further adapt to specific styles with personalized optimization. In this figure, our lip-sync results should have the same mouth shapes as the lip-synced video of the conditional audio." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 547, + 192, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 547, + 192, + 559 + ], + "spans": [ + { + "bbox": [ + 143, + 547, + 192, + 559 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 573, + 288, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 288, + 681 + ], + "type": "text", + "content": "Despite recent advances in syncing lip movements with any audio waves, current methods still struggle to balance generation quality and the model's generalization ability. Previous studies either require long-term data for training or produce a similar movement pattern on all subjects with low quality. In this paper, we propose StyleSync, an effective framework that enables high-fidelity lip synchronization. We identify that a style-based generator would sufficiently enable such a charming property on both one-shot" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 548, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 545, + 692 + ], + "type": "text", + "content": "and few-shot scenarios. Specifically, we design a mask-guided spatial information encoding module that preserves the details of the given face. The mouth shapes are accurately modified by audio through modulated convolutions. Moreover, our design also enables personalized lip-sync by introducing style space and generator refinement on only limited frames. Thus the identity and talking style of a target person could be accurately preserved. Extensive experiments demonstrate the effectiveness of our method in producing high-fidelity results on a variety of scenes. Resources can be found at https://hangz-nju-cuhk.github.io/projects/StyleSync." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 125, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 125, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 125, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1505" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 286, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 233 + ], + "type": "text", + "content": "The problem of generating lip-synced videos according to conditional audio is of great importance to the field of digital human creation, audio dubbing, film-making, and entertainment. While the rapid development of this area has been witnessed within recent years, most methods [6,8,9,16,19,24,27,28,38,39,46,50,54,57-60] focus on generating a whole dynamic talking head. Results created under such settings can hardly be blended into an existing scene. Under real-world scenarios like audio dubbing, one crucial need is to seamlessly alter the mouth or facial area while preserving other parts of the scene unchanged, making these methods non-feasible." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 235, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 235, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 49, + 235, + 286, + 377 + ], + "type": "text", + "content": "Previous methods take two different paths for achieving seamless mouth modification. A number of studies [38, 44] pursue realistic results on person-specific settings, which require long-term clips for target modeling. Moreover, they rely on prior 3D facial structural information. The uncertainty and errors accumulated in the 3D fitting procedure would greatly influence their performances. On the other hand, it is desired to build models that break the data limitation on more generalized scenes. As a result, a few methods [31, 32, 41] design person-agnostic models without relying on 3D or 2D structural priors. Nevertheless, such a setting is extremely challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 379, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 379, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 49, + 379, + 286, + 533 + ], + "type": "text", + "content": "In order to produce high-fidelity lip-synced results on any-length videos, two essential challenges need to be addressed. 1) How to efficiently design a powerful generative backbone network that supports both accurate audio information expression and seamless local alternation. Intuitively, the lip-sync quality naturally contradicts the preservation of the original target frame information [32, 59]. 2) How to effectively leverage as much provided information as possible and involve the personalized properties to a generalized model. Though few-shot meta-learning has been proven effective in generating talking heads [5, 7, 56, 60], how to involve such ability into a lip-syncing pipeline has not been explored." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 534, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 286, + 653 + ], + "type": "text", + "content": "In this paper, we propose a highly concise and comprehensive framework named StyleSync, which produces high-fidelity lip-sync results on both generalized and personalized scenarios. The key is our simple but lip-sync-oriented modifications to style-based generator. Though style-based generators [22, 23] have been leveraged in various talking head generation methods [5, 55, 59], their successes are only partially instructive. They aim at producing the whole head, which leads to unstable background and distortions which are non-acceptable in our scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "type": "text", + "content": "By revisiting the details of style-based generators, we identify a few simple but essential modifications that make our framework suitable for lip-syncing. Different from the above methods, we adopt a masked mouth modeling protocol [32, 41] and delicately design a Mask-based Spa" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "type": "text", + "content": "tial Information Encoding strategy, where both the target and reference frames' information is encoded into a noise space [52, 53] of the generator according to different masking schemes. While the information on audio dynamics and high-level reference frame is injected into the style-modulated convolution in a similar manner as [25, 59]. In this way, our method can be benefited from the strong generative power of style-based generators and also keeps the advantage of easy implementation and fast training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 181, + 544, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 181, + 544, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 544, + 335 + ], + "type": "text", + "content": "Moreover, our network modification enables personalized information preserving (e.g., speaking styles and details of the mouth and jaw). We take inspiration from the recent studies of inverting StyleGAN priors [1, 2, 35, 45] and propose a Personalized Optimization scheme. As audio dubbing is normally performed on speaking videos, our model can make use of only a few seconds of the person's information and optimize additional person-specific parameters including the " + }, + { + "bbox": [ + 308, + 181, + 544, + 335 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 308, + 181, + 544, + 335 + ], + "type": "text", + "content": " and the generator. Extensive experiments show that our framework clearly outperforms previous state of the arts on the one-shot setting by a large margin, and the target-specific optimization further enhances the fidelity of our results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 337, + 544, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 337, + 544, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 544, + 479 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: 1) We present the StyleSync framework, which adopts simple but effective modifications including the Mask-based Spatial Information Encoding to a style-based generator. 2) We propose the Personalized Optimization procedure which involves few-shot person-specific optimization into our framework. 3) Extensive experiments demonstrate that our framework can directly produce accurate and high-fidelity one-shot lip-sync results. Moreover, our proposed personalized optimization further improves the generation quality. Our method outperforms previous methods by a clear margin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 493, + 391, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 391, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 391, + 506 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 514, + 474, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 474, + 526 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 474, + 526 + ], + "type": "text", + "content": "2.1. Audio-Driven Facial Animation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 533, + 544, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 544, + 604 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 544, + 604 + ], + "type": "text", + "content": "The topic of audio-driven facial animation has long been a research interest in both the computer vision and graphics community. Studies have been carried out on both digital 3D human faces [15, 20, 33, 61] and realistic human heads [14, 17, 42]. We focus on human heads in the real world." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 605, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 605, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 605, + 544, + 712 + ], + "type": "text", + "content": "Talking Head Generation. Most studies on lip-syncing target generating the whole head of a talking person [6, 8, 9, 16-18, 25, 26, 40, 42, 54, 58-60]. Specifically, a number of studies leverage structural information such as 2D [9] landmarks, 3D landmarks [60] and 3D meshes [6]. The uncertainty and inaccuracy of such representations would lead to error accumulation during the talking head generation procedure. Most person-specific methods [19, 27, 42] rely on them and produce results with a poor generalization or lip" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 750, + 315, + 757 + ], + "type": "text", + "content": "1506" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": "sync quality. Recently, NeRF is also used in person-specific modeling [16, 26, 43], but they also perform poorly when driven by audio from a different person." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 178 + ], + "type": "text", + "content": "Lip-Syncing on Faces. The other type of study focuses on lip-syncing the mouth part while keeping other information untouched in videos [31, 32, 39, 44]. Our work lies within the same scope. While Thies et al. [44] and Song et al. [38] produce realistic results, they rely on person-specific training on more than 2 minutes of data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 286, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 276 + ], + "type": "text", + "content": "Specifically, Wav2Lip [32] generates person-agnostic lip-sync results that are highly accurate. However, they build the generative model on low-resolution images, leading to blurry results. We identify that it would be easier for the model to learn the mouth motions if less information about the image quality needs to be recovered. Moreover, their methods cannot capture the personalized patterns given the target template video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 282, + 220, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 220, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 220, + 295 + ], + "type": "text", + "content": "2.2. Style-based Generator for Faces" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 335 + ], + "type": "text", + "content": "StyleGAN models [21-23] have shown great success on image generation tasks, particularly on facial image generation and editing [1, 2, 34, 36, 37, 45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "text", + "content": "StyleGAN Inversion. StyleGAN inversion is the practice of recovering the latent space of a given image with a pretrained StyleGAN. Abdal et al. [1, 2] propose to invert images not only using the style space " + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "text", + "content": ", but also expand it to the " + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "text", + "content": ", which better preserves details. Recent studies propose to learn encoders [3, 34, 45] for better editability and faster optimization results. Furthermore, recent studies try to refine the generator's parameters through pivotal finetuning [35]. In our work, we take inspiration of previous StyleGAN inversion studies and propose a personalized optimization procedure. We encode a " + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 337, + 286, + 491 + ], + "type": "text", + "content": " space from reference frames following [45] which enables personalized lip movements learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "text", + "content": "Applications. StyleGAN architecture has also been leveraged in face restorations [48,53] and face swapping [51,52]. These tasks require preserving the original facial emotion and expressions. On the other hand, style-based generators have also been applied to create facial animations [5,25,59], which are highly related to our task. However, they rely on the style vectors in the " + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "text", + "content": " space for controlling both the appearances and motion dynamics. One major drawback of this setting is that " + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 46, + 492, + 286, + 622 + ], + "type": "text", + "content": " space cannot easily accounts for the spatial consistency of backgrounds, leading to nonrealistic results or visible artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 129, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 129, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 129, + 647 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "In this section, we introduce the details of our StyleSync framework as depicted in Fig. 2. It is simply designed by leveraging several successes of previous style-based generators. We introduce our modifications to the style-based generator that make a successful backbone in Sec. 3.1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "and the training objectives of our generalized model in Sec. 3.2. Specifically, we illustrate our Personalized Optimization procedure which further strengthens our results with person-specific optimization in Sec. 3.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": "Task Formulation. Our goal is to sync the lip motion of a target person with any provided audio and seamlessly blend it into the original target video. We formulate our training in a typical self-reconstruction manner. The training setting is similar to Wav2Lip [32]. Given a target video " + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{I_1, \\dots, I_T\\}" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": " frames, we mask out the lower half of the face including the jaw and cheeks with a mask " + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": ". The goal is to recover the whole face with its corresponding audio " + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = \\{a_1, \\dots, a_T\\}" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": " is processed to spectrograms). As no information about the mouth and jaw shape is provided on the masked target frame " + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "I_t^m = (\\mathbf{1} - M)*I_t" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": ", we leverage a random reference frame " + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "inline_equation", + "content": "I_{ref} \\in \\mathbf{V}" + }, + { + "bbox": [ + 304, + 121, + 545, + 276 + ], + "type": "text", + "content": " during training to provide the desired context." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 285, + 545, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 298 + ], + "type": "text", + "content": "3.1. Modifying Style-based Generator for Lip Sync" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": "Style-based Generator Overview. We perform lip-sync-oriented modifications to build a successful backbone for lip sync with the StyleGAN2 [23] architecture. The original StyleGAN2 takes a style vector " + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": " as input and uses it to modulate convolution operations in a total of " + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": " generative layers during training. During inference, different " + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": "s can be used at different layers, formulating the " + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": " space " + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\{w_{1},\\dots ,w_{L}\\}" + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": ". The basic elements of the original StyleGAN2, including the " + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 304, + 304, + 545, + 436 + ], + "type": "text", + "content": " space and the StyleGAN generator G's layers are depicted in blue on Fig. 2. For simplification, certain detailed structures are omitted." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 509 + ], + "type": "text", + "content": "Mask-based Spatial Information Encoding. Our goal is to seamlessly blend a lip-synced mouth into the target frame with the assistance of a reference image. As discussed in Sec. 2.2, previous methods leveraging the style-based generator are not directly applicable to our scene, we seek a different way to encode the spatial information on faces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "text", + "content": "It is proven by recent studies on face restoration [53] and swapping [52] that content-guided feature maps can be attached to StyleGAN layers in a similar status as noise without affecting the expressive power of the generative model. These contextual-rich noises " + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{N} = \\{N_1,\\dots ,N_L\\}" + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "text", + "content": " could well preserve the original spatial structure and attributes of the encoded visual information. This property provides an option for leveraging the attributes information lying in the reference image. Thus following Wav2Lip [32], we concatenate the masked and reference input together as the visual input of the framework and encode the corresponding feature maps " + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f = \\{F_1,\\ldots ,F_L\\}" + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "text", + "content": ". The simplest setting is to regard " + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{N} = \\mathbf{F}" + }, + { + "bbox": [ + 304, + 510, + 545, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "However, the goal of generating lip motions needs to create dynamics that are different from the original facial structure but keeps facial identity, which is substantially different from previous methods that aim to keep the structure" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1507" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 240 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 240 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 240 + ], + "type": "image", + "image_path": "70c1b84618322593295d202b149ded560bde30d5279920e2a606abd7ec617b7c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": "Figure 2. Our StyleSync framework. The building blocks in Blue indicate the style-based generator. The masked target frame " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "I_{t}^{m}" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": " is concatenated with a reference " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": " and encoded to " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\text{face}}(\\mathbf{Red})" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": ". The audio information is encoded to an audio feature " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": " (Yellow). The features are concatenated together to form the style " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": " space. Specifically, we devise a Personalized Optimization procedure (Green) including the learning of the " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P}" + }, + { + "bbox": [ + 46, + 251, + 546, + 297 + ], + "type": "text", + "content": ". This part is not trained during the initial backbone learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "type": "text", + "content": "fixed. The mouth shapes would be greatly influenced by the reference image " + }, + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "type": "text", + "content": " when adopting the same protocol. We identify that more than desired facial structure information is infused in the low-level information of " + }, + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 316, + 288, + 400 + ], + "type": "text", + "content": " (layers with higher resolutions). As a result, we intuitively mask out the low-level part of the information in the face features. Here we simply define" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 407, + 287, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 407, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 97, + 407, + 287, + 431 + ], + "type": "interline_equation", + "content": "F _ {l} ^ {\\prime} = (\\mathbf {1} - M) * F _ {l}, \\text {f o r} l > \\left\\lfloor \\frac {L}{2} \\right\\rfloor . \\tag {1}", + "image_path": "a36e37b49555bbb0371f05f8c43d542d60edfe984632d1b007cb20da1f66e359.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 462 + ], + "type": "text", + "content": "The facial information is colored in red on Fig. 2. The 1 is an all-one matrix that has the same size as " + }, + { + "bbox": [ + 46, + 437, + 287, + 462 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 437, + 287, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": "Style Information Encoding. We follow previous studies [5, 40, 59] to encode both audio dynamics and facial information into the style space (" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": " space) of the style-based generator. The audio feature " + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": " is encoded from the encoder " + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": " and the face feature " + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": " is the bottleneck of the face encoder " + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "inline_equation", + "content": "E_{face}, w = \\mathrm{concat}(f_{a}, f_{f})" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": ". As the information in the reference frames is already fused to the generator through spatial noise encoding, the fusing of the face feature is less necessary. Experiments show that the difference is subtle with or without " + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 46, + 462, + 287, + 594 + ], + "type": "text", + "content": ", but we still keep this design in accordance with common practice." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "content": "Ingredients for Personalization. Our above designs are not only simple but also possess external potentials of personalized lip-sync modeling. We leave the details to Sec. 3.3 and briefly discuss the additional ingredients that make our model comprehensive." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": "Specifically, our extension is inspired by the advances in StyleGAN inversion [1, 2, 35]. It has pointed out that extending " + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": " space which contains " + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\{w + \\Delta w_{1},\\dots ,w + \\Delta w_{L}\\}" + }, + { + "bbox": [ + 47, + 653, + 287, + 714 + ], + "type": "text", + "content": " is essential to recovering specific images. Moreover, recent studies [4, 35] also explore tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 363 + ], + "type": "text", + "content": "the parameters of the generator to improve the network's fitting ability on a specific target. Thus our generator has the potential of learning limited parameter shifts " + }, + { + "bbox": [ + 304, + 316, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P}" + }, + { + "bbox": [ + 304, + 316, + 545, + 363 + ], + "type": "text", + "content": " in order to fit a specific person." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "type": "text", + "content": "These personalized ingredients are depicted in green on Fig. 2 and are not optimized during the generalized training procedure. In order to keep the pipeline consistent across both settings, we set " + }, + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W} = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P} = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 363, + 545, + 422 + ], + "type": "text", + "content": " in our generalized model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 430, + 470, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 470, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 470, + 443 + ], + "type": "text", + "content": "3.2. Backbone Training Objectives" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": "During training, the whole backbone networks take the 6-channel concatenation of " + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "inline_equation", + "content": "[I_t^m, I_{ref}]" + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": " and the audio clip " + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": " at the same time step as input. It predicts " + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "inline_equation", + "content": "I_t' = \\mathrm{G}(\\mathbf{N}, \\mathbf{W})" + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": ", which aims at recovering the unmasked target image " + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": ". In order to keep our design simple, the training objectives are mostly aligned with StyleGAN2 [23] and Wav2Lip [32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "text", + "content": "Reconstruction Loss. The pixel-level reconstruction loss " + }, + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "text", + "content": " is fundamental for training our task. Here we leverage the commonly used " + }, + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 520, + 545, + 556 + ], + "type": "text", + "content": " loss and perceptual loss [47]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 562, + 545, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 562, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 311, + 562, + 545, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left\\| I _ {t} ^ {\\prime} - I _ {t} \\right\\| _ {1} + \\sum_ {m = 1} ^ {N _ {v g g}} \\left\\| \\mathrm {V G G} _ {m} \\left(I _ {t} ^ {\\prime}\\right) - \\mathrm {V G G} _ {m} \\left(I _ {t}\\right) \\right\\| _ {1}, \\tag {2}", + "image_path": "5a77a96f5e283a1ab928ab462dc278e681daffb10bc7d2a05f408787655c1a0c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{VGG}_m" + }, + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 613, + 545, + 636 + ], + "type": "text", + "content": "th layer's output of a pre-trained VGG19 network." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 638, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 638, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 638, + 545, + 662 + ], + "type": "text", + "content": "Adversarial Loss. We directly adopt the same discriminator D from StyleGAN2 [23] for the adversarial training:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 680, + 545, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 680, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 311, + 680, + 545, + 710 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} = \\min _ {\\mathrm {G}} \\max _ {\\mathrm {D}} \\left(\\mathbb {E} _ {I _ {t}} \\left[ \\log \\mathrm {D} \\left(I _ {t}\\right) \\right] + \\mathbb {E} _ {I _ {t} ^ {\\prime}} \\left[ \\log \\left(1 - \\mathrm {D} \\left(I _ {t} ^ {\\prime}\\right)\\right) \\right]\\right). \\tag {3}", + "image_path": "24e5458a913e2da2cb2f9493d309a06bf5d56e2ce8f2bdf7523382899629eb3d.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1508" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "Particularly, we initialize the discriminator's weights with the pre-trained version in StyleGAN2 [23]. This practice produces visibly sharper results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "text", + "content": "Lip-Sync Loss. We additionally train a SyncNet [12] which consists of a visual encoder " + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "inline_equation", + "content": "S_v" + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "text", + "content": " and an audio encoder " + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "inline_equation", + "content": "S_a" + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "text", + "content": ". It identifies whether visual and audio clips are timely aligned with contrastive loss. When training the generator, we predict 5 consecutive frames " + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "inline_equation", + "content": "I_{t:t+4}'" + }, + { + "bbox": [ + 46, + 108, + 287, + 192 + ], + "type": "text", + "content": " within one batch at each inference step and supervise the training with SyncNet's assistance. The objective is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 85, + 200, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 200, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 85, + 200, + 287, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s y n c}} = - \\frac {\\mathrm {S} _ {v} \\left(I _ {t : t + 4} ^ {\\prime}\\right) ^ {\\mathrm {T}} \\cdot \\mathrm {S} _ {a} \\left(a _ {t : t + 4}\\right)}{\\| \\mathrm {S} _ {v} \\left(I _ {t : t + 4} ^ {\\prime}\\right) \\| _ {2} \\| \\mathrm {S} _ {a} \\left(a _ {t : t + 4}\\right) \\| _ {2}} \\tag {4}", + "image_path": "f84bdd3122d1e3b978f6a81fc9916535d1d64d86b7e9481f49456c4174fb1daf.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 236, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 260 + ], + "type": "text", + "content": "The overall loss functions across the generalized training can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 282, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 282, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 97, + 282, + 287, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g} = \\mathcal {L} _ {a d v} + \\lambda_ {r} \\mathcal {L} _ {r e c} + \\lambda_ {s} \\mathcal {L} _ {\\text {s y n c}}. \\tag {5}", + "image_path": "eb5e8b621f16c0ab4417f52ce347a93d16abf614cc58ab5ff7e23f3b6b9c1df2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 312, + 194, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 194, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 194, + 325 + ], + "type": "text", + "content": "3.3. Personalized Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 331, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 484 + ], + "type": "text", + "content": "After the training of the backbone networks, our method can readily generate high-fidelity lip sync results for arbitrary subject. However, as the model is trained in a generalized manner, the generated mouth motion patterns across different people are basically the same. It has been verified that different identities possess different talking styles [50, 54]. Here we pursue to capture such personalized property with the original template video. Notably, we focus on the few-shot setting that only less than one minute of the original video is given, which cannot be handled by previous person-specific models [16, 19, 38, 44]. Below, we illustrate how we successfully design our personalized optimization module based on the above discussions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": "Basic Learning Settings. The basic learning procedure of the personalized optimization is similar to backbone training. On a template video " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{V}} = \\{\\bar{I}_1,\\dots ,\\bar{I}_T\\}" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " with audio " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}} = \\{\\bar{a}_1\\dots ,\\bar{a}_T\\}" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": ", we randomly select two frames as target " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\bar{I}_t" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " and reference " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\bar{I}_{ref}" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " for training as one sample. During training, the generalized loss " + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_g" + }, + { + "bbox": [ + 46, + 486, + 287, + 569 + ], + "type": "text", + "content": " continuously supervises the whole procedure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": "As the encoder " + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_a" + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": " encodes person-agnostic speech content information, it is fixed during the personalized optimization step. While the face encoder " + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{face}" + }, + { + "bbox": [ + 46, + 570, + 287, + 629 + ], + "type": "text", + "content": " encodes the attribute information of the target person, we do not optimize its parameters on a single person to avoid overfitting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Style Space Optimization. According to our formulation, the " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " space contains both the lip motion and the high-level facial style information. Thus optimizing it would ideally lead to person-specific lip motion pattern. E4E [45] verifies that learning the small displacements around the original " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " space would both extend the inversion quality of the image and enables strong editing ability. Thus we propose" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "to learn a set of " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W} = MLP_{\\Delta w}(f_f)" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " in a similar way according to the encoded face feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "f_{f}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " with a few layers of MLPs " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "MLP_{\\Delta w}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "text", + "content": "Generator Tuning. The generator most accounts for the synthesizing and the blending quality. We allow the parameters " + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "text", + "content": " of our generator " + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "text", + "content": " to shift a little margin to " + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P}" + }, + { + "bbox": [ + 304, + 109, + 545, + 156 + ], + "type": "text", + "content": " with the personalized data we use." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 158, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 182 + ], + "type": "text", + "content": "Learning Objectives. The final learning objectives of the personalized optimization procedure can be summarized as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 202, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 202, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 335, + 202, + 545, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\mathcal {L} _ {g} + \\lambda_ {p} \\left(\\sum_ {i, j} \\left| \\Delta \\mathbf {W} _ {i} ^ {j} \\right| _ {2} ^ {2} + \\sum_ {m, n} \\left| \\Delta \\mathbf {P} _ {n} ^ {m} \\right| _ {2} ^ {2}\\right), \\tag {6}", + "image_path": "30c43f202c6c442b290f5fff7280ea325a29676123ddf8fba280f451feb100a2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^j" + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "content": "th parameter of the " + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "content": "th element in " + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "content": ". This stands the same for " + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_n^m" + }, + { + "bbox": [ + 304, + 240, + 545, + 277 + ], + "type": "text", + "content": ". We restrict the displacements in a limited step." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 290, + 387, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 387, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 387, + 304 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 311, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 358 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two commonly used audio-visual datasets, LRW [11] and VoxCeleb2 [10]. We follow the datasets' original train/test split and train our model on a mixture of these two datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 369, + 545, + 499 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 317, + 369, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 369, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 545, + 417 + ], + "type": "text", + "content": "- LRW [11] is an audio-visual dataset collected from BBC news for lip reading. It is one of the earliest large-scale audio-visual datasets with high quality. It consists of 1,000 one-second utterances in 500 words." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 428, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 428, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 428, + 545, + 499 + ], + "type": "text", + "content": "- VoxCeleb2 [10] VoxCeleb [30] and VoxCeleb2 are large-scale audio-visual datasets collected for the speaker verification task. We re-download around one-fifth of VoxCeleb2 for training, which are of high quality. The official test set is processed with GPEN [53] for evaluation." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We process all videos at 25 fps and align all faces according to pre-detected landmarks at the eyes. All faces are cropped to the size of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". A same U-shape mask is adopted as shown in Fig. 2 to erase the mouth, cheeks and jaws at the target frame. All audiois processed in the same way as Wav2Lip [32]. In order to keep our design simple and avoid extra hyperparameter tuning, we adopt most settings from the previous studies [23, 32, 53]. Specifically, the generator has a total of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "L = 14" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " style-convolution layers. " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{r}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is empirically set to 10 and all other " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda s" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " are selected to 1. For personalized optimization, we train the person's data for 5 epochs. Long training would not lead to better results. Particularly, we find that training the personalized model together with the general training data leads to more stable results, the portion of personal data and general data is set as 1:1 in our experiments." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1509" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 294, + 349 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 294, + 349 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 294, + 349 + ], + "type": "image", + "image_path": "d8e4fea59578dd2f24612c242704a827769897df1dd075646a2915ef935cf337.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 355, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 547, + 411 + ], + "type": "text", + "content": "Figure 3. Qualitative Results. The top row shows the lip-synced videos of the conditional driving audio. MakeitTalk [60] fails to generate accurate mouth shape and lacks head dynamics. PC-AVS [59] and Wav2Lip [32] generate lip-synced results aligned well with audios. But they produce visibly blurry results. Particularly, our reproduced Wav2Lip-H generates high-quality results with good lip-sync. However, they still produce certain artifacts. While our generalized model directly produces high-fidelity results that have the same lip motion as the original syncsed video." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 70, + 545, + 348 + ], + "blocks": [ + { + "bbox": [ + 300, + 70, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 300, + 70, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 300, + 70, + 545, + 348 + ], + "type": "image", + "image_path": "1ebf1340171f28ebeef39bba2ffb50bbcd782769173a93cf2f61fa754bd66a8e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 432, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 289, + 613 + ], + "type": "text", + "content": "Comparing Methods. As our model is built based on the person-agnostic setting, we compare our StyleSync framework with three state-of-the-art methods including, MakeiTalk [60], PC-AVS [59] and Wav2Lip [32]. MakeiTalk [60] synthesizes talking head videos with natural head pose. PC-AVS [59] a pose source video to achieve pose-controllable talking face generation. While MakeiTalk and PC-AVS produce a whole talking head, Wav2Lip which focuses on the mouth part is our main competing method. As Wav2Lip is originally trained on a low-resolution with small networks, we carefully reproduce it on the same data and even loss functions as our methods. We denote this model as Wav2Lip-H (High-Quality). For fair comparison, all basic comparisons are carried out on the generalized setting Ours-G." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "We also perform personalized optimization on the test set, where each frame lasts less than 10 seconds for both datasets. The results are denoted as Ours-P. We equally perform the generator finetuning on the above methods, and all of them perform worse than their original models, thus the results are not listed below. Additionally, we compare our personalized setting with ADNeRF [16], a recent talking head advance using neural radiance fields [29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 431, + 441, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 441, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 441, + 445 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 510 + ], + "type": "text", + "content": "As quantitative evaluations can only be performed on the self-construction setting, we avoid directly leveraging all information within the frame for reconstruction. We uniformly select a same random frame as the reference frame for PC-AVS, Wav2Lip, Wav2Lip-H and ours." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "Evaluation Metrics. We follow previous studies [8, 9, 58, 59] to adopt the popularly used SSIM [49], PSNR metrics for evaluating the generation quality, and the landmark distances around the mouth (LMD) and SyncNet's [12] confidence score to show the lip-sync quality. Please be noted that the SyncNet metric is computed on the officially released version, which is different from our implementation in the loss functions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "type": "text", + "content": "Particularly, as the details of the target person cannot be fully recovered, we leverage the ArcFace [13] network and compute the frame-by-frame feature cosine distances " + }, + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{ID}" + }, + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "type": "text", + "content": " for evaluating whether the generated results preserve the identity well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Results. The quantitative experiments are carried out on the test set of LRW [11] and VoxCeleb2 [10] datasets. Please refer to Table 1 for the results. It can be seen that on the LRW dataset, the generalized version" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1510" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 193 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 193 + ], + "type": "table", + "html": "
MethodLRWVoxCeleb2
SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID↑SSIM ↑PSNR ↑LMD ↓Syncconf ↑ID ↑
MakeiTalk0.6929.832.753.880.790.6328.386.942.150.71
PC-AVS0.7930.261.847.190.820.7129.532.758.160.74
Wav2Lip0.7930.541.287.390.900.8030.531.928.900.90
Wav2Lip-H0.8031.381.207.190.880.8130.531.878.350.90
GT1.00100.000.007.651.001.00100.000.007.711.00
Ours-G0.8531.781.187.250.890.7931.001.478.250.90
Ours-P0.8832.660.866.350.930.8231.541.157.260.93
", + "image_path": "abc1b9eb4fb7ab1c36c04fff71d8e5c279a3c183b200443b2f2d76797120544c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 200, + 530, + 212 + ], + "lines": [ + { + "bbox": [ + 62, + 200, + 530, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 200, + 530, + 212 + ], + "type": "text", + "content": "Table 1. Quantitative results on LRW and VoxCeleb. For LMD the lower the better, and the higher the better for other metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "text", + "content": "of our StyleSync already outperforms previous methods by a large margin for generation quality. While on the VoxCeleb2 dataset, the gap is less obvious. The reason might be that most training data of our methods are basically frontal view faces selected from LRW, while VoxCeleb2 contains more complicated scenes. On the other hand, after our personalized optimization, the performance of our model advances again. This shows the generated results are clearly more similar to the targets at this stage, which can also be verified by looking into the identity distances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 494 + ], + "type": "text", + "content": "Meanwhile, our method also achieves comparable performance on the lip-sync metrics on both datasets. Our LMD score is slightly better than the competing methods. As for the SyncNet score, we achieve comparable results on LRW that are closer to the ground truth's SyncNet score. We argue it is meaningless to refer to the SyncNet scores for verifying the lip-sync quality once the metric has outperformed the ground truth. The " + }, + { + "bbox": [ + 46, + 350, + 287, + 494 + ], + "type": "inline_equation", + "content": "\\text{Sync}_{\\text{conf}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 494 + ], + "type": "text", + "content": " only reflects how well an audio-visual pair fits the learned SyncNet model rather than the true perceptual quality. Thus though generated results might outperform ground truth on the metric, it does not mean better sync quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 494, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 287, + 553 + ], + "type": "text", + "content": "After the personalization, the lip-sync score degrades. We assume that the specificity of the speaking style reduces the mouth opening level. This is also shown in Fig. 5. The woman tends not to open her mouth wide, leading to smaller mouth movements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 562, + 176, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 176, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 176, + 575 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 581, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 641 + ], + "type": "text", + "content": "Subjective evaluation is crucial in identifying the ability of generative models, particularly on videos. We strongly recommend readers to watch our supplementary video at https://hangz-nju-cuhk.github.io/projects/StyleSync." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Two cross-driven examples and their comparisons with the SOTA methods are shown in Fig. 3. We use an conditional audio from an arbitrary person selected from the test set to drive the template. It can be seen that MakeiTalk [60] cannot produce accurate lip movements. Moreover, both PC-AVS and Wav2Lip produce visible artifacts or blurry" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 323 + ], + "type": "image", + "image_path": "b43d453f1557b4c49c6416fe4aeebf4bcea100aa2bd18d7de9df88488b251f3e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 331, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 545, + 342 + ], + "type": "text", + "content": "Figure 4. Self-driven results compared with personalized method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 360, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 407 + ], + "type": "text", + "content": "results. Our reproduced Wav2Lip-H adopts the same training setting as ours, thus generates plausible results for most cases. Nevertheless, our generalized model still produces the most accurate mouth shapes with the highest fidelity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 410, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 506 + ], + "type": "text", + "content": "Additionally, we show a case of personalized optimization on a 50-second video clip in Fig. 1. It can be seen that our method preserves the speaking style of the target person with accurate lip sync. Comparing our personalized results (Ours-P) with AD-NeRF [16] (Fig. 4), it is evident that our method maintains more person-specific details with higher fidelity, even though we use significantly less personal data (10s for tuning v.s. 5min for training in [16])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 639 + ], + "type": "text", + "content": "User Study. We also invite 15 participants to conduct a user study for further subjective evaluation and the results are reported in Table 2. Specifically, we randomly select 63 videos from LRW and VoxCeleb2 datasets as test sets and generate videos by using our generalized StyleSync and the comparison methods accordingly. By adopting the commonly used Mean Opinion Scores (MOS) rating protocol, we request all the participants to provide their ratings (from 1 to 5, the higher the better) on three aspects for each generated video: (1) Lip-Sync quality; (2) Generation quality; (3) Video realness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "As shown in Table 2, MakeitTalk [60] achieves the lowest scores in all the aspects due to lacking reasonable lip movements and head pose. PC-AVS [59], Wav2Lip [32] and Wav2Lip-H achieve relatively higher lip-sync score, but only Wav2Lip-H is able to synthesize videos with less blurry textures around the mouth. Overall, our StyleSync" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1511" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 148 + ], + "type": "table", + "html": "
MOS on \\ ApproachMakeitTalkPCALPSWav2LipWav2Lip-HOurs-G
Lip-Sync Quality2.063.003.493.674.24
Generation Quality2.632.161.873.424.52
Video Realness1.892.162.172.984.06
", + "image_path": "fa00b56397ce052857d8bdd60f9945d88222cfe47a65a70ba9faffce6c9ed0b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 184, + 285, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 177 + ], + "type": "text", + "content": "Table 2. User study measured by Mean Opinion Scores. The scores are ranged from 1 (worst) to 5 (best)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 184, + 285, + 256 + ], + "lines": [ + { + "bbox": [ + 50, + 184, + 285, + 256 + ], + "spans": [ + { + "bbox": [ + 50, + 184, + 285, + 256 + ], + "type": "table", + "html": "
MethodSSIM ↑PSNR ↑LMD ↓Syncconf ↑LID ↑
w/o mask0.8131.201.807.890.90
w/o sync0.8031.341.557.410.90
Ours-G0.7931.001.478.250.90
P w/o ΔW0.8231.511.327.310.92
Ours-P0.8231.541.157.260.93
", + "image_path": "9ec2312e4a9f7b05b3fa7bdf7b4791c5dfe0e37c357b2a0f080ad895e67ca9e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 264, + 263, + 275 + ], + "lines": [ + { + "bbox": [ + 69, + 264, + 263, + 275 + ], + "spans": [ + { + "bbox": [ + 69, + 264, + 263, + 275 + ], + "type": "text", + "content": "Table 3. Quantitative ablations results on VoxCeleb2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 290, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 315 + ], + "type": "text", + "content": "outperforms its counterparts in all the three aspects by a large margin, indicating the effectiveness of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 323, + 140, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 140, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 140, + 335 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "type": "text", + "content": "To further demonstrate the contributions of our novel designs, we perform an ablation study on VoxCeleb2 dataset under both generalized setting and personalized optimization setting (denoted as \"Ours-G\" and \"Ours-P\", respectively). Concretely, we construct two variants for \"Ours-G\" (denoted as \"w/o mask\" and \"w/o sync\") by removing the masking operation in Eq. 1 and the lip-sync loss during training, respectively. While for \"Ours-P\", we form one variant (denoted as \"Ours-P w/o " + }, + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "type": "text", + "content": ") by setting " + }, + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W} = \\mathbf{0}" + }, + { + "bbox": [ + 46, + 342, + 287, + 484 + ], + "type": "text", + "content": ". We have also experimented personalized optimization with fixed generated, however, this would lead to blurry results. This ablation is omitted here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 641 + ], + "type": "text", + "content": "The quantitative and qualitative results are shown in Table 3 and Fig. 5. As shown in Table 3, we observe that both \"w/o mask\" and \"w/o sync\" achieve comparable performance on the image quality metrics but leads to a performance drop on the lip-sync metrics. \"w/o mask\" achieves worse scores when compared with \"Ours-G\" in terms of LMD and Sync, which indicates that the proposed masking strategy can obviously alleviate the influence from the reference frames. While \"w/o sync\" reasonably suffers from lip-sync degradation and the results in Fig. 5 illustrate inconsistent lip movements when compared with the lip-synced video. The results demonstrate the effectiveness of the additional supervision from SyncNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "It can be seen from Fig. 5 that with the personalized optimization, both the identity and the pattern of the mouth opening on the generated frames become more similar to the original template video. It is also clear that after the personalized training, the mouth opening is less obvious than in the generalized model. We analyze that this is part of the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 255 + ], + "type": "image", + "image_path": "1435fc58b50a1e4bab9eae4c4538d0fe2746fae5050c185db0d51b432bf3a94c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 264, + 543, + 275 + ], + "lines": [ + { + "bbox": [ + 307, + 264, + 543, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 543, + 275 + ], + "type": "text", + "content": "Figure 5. Ablation study with visual results. Zoom in for details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 312 + ], + "type": "text", + "content": "talking style of this target person. This also leads to poorer metric values on SyncNet." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": "In terms of the comparisons between our personalized model and \"Ours-P w/o " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": "\", it achieves similar scores on the image quality as well as lip-sync quality. However, the identity score without " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": " is slightly degraded, which indicates that it is beneficial to involve " + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 305, + 312, + 545, + 384 + ], + "type": "text", + "content": " into the training procedure for better preserving identity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 394, + 463, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 463, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 463, + 407 + ], + "type": "text", + "content": "5. Conclusion and Discussions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 415, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 534 + ], + "type": "text", + "content": "Conclusion. In this paper, we propose StyleSync, which produces high-fidelity lip sync results for both the one-shot and the few-shot settings. We highlight the unique proprieties of our method: 1) Video results produced by our generalized model clearly outperform previous state-of-the-arts. 2) Our model is built upon the success of recent style-based generators with simple modification. It is easy to implement and friendly to train. 3) By involving StyleGAN-inspired personalized optimization procedure, our model can further be improved on a specific person given only a few clips." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 594 + ], + "type": "text", + "content": "Limitations. As our method blends a lip-synced face into an existing video with a fixed mask, the head pose and expressions of the target person cannot be changed. Additionally, under certain extreme cases where the target's jaw is extremely large, it might get out of our masked area." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": "Ethical Considerations. Our method could be used to create non-existing talks and speeches, which might be maliciously used. We will issue our core models strictly to research institutions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1512" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 293 + ], + "type": "text", + "content": "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 349 + ], + "type": "text", + "content": "[5] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13786-13795, 2020. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 288, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 394 + ], + "type": "text", + "content": "[6] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation?: A survey and benchmark. arXiv preprint arXiv:2005.03201, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 441 + ], + "type": "text", + "content": "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 441, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 288, + 485 + ], + "type": "text", + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "text", + "content": "[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 565 + ], + "type": "text", + "content": "[10] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 623 + ], + "type": "text", + "content": "[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Bo Fan, Lijuan Wang, Frank K Soong, and Lei Xie. Photoreal talking head with deep bidirectional LSTM. In 2015 IEEE" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4884-4888. IEEE, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 98, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 98, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 547, + 152 + ], + "type": "text", + "content": "[15] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 154, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 208 + ], + "type": "text", + "content": "[16] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 545, + 255 + ], + "type": "text", + "content": "[17] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. International Journal of Computer Vision, 127(11):1767-1779, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 258, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 300 + ], + "type": "text", + "content": "[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 303, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 358 + ], + "type": "text", + "content": "[19] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14080-14089, 2021. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 406, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 440 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 441, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 495 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 541 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 609 + ], + "type": "text", + "content": "[24] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2755-2764, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 612, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 678 + ], + "type": "text", + "content": "[25] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1513" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[27] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[28] Yifeng Ma, Suzhen Wang, Zhipeng Hu, Changjie Fan, Tangjie Lv, Yu Ding, Zhidong Deng, and Xin Yu. Styletalk: One-shot talking head generation with controllable speaking styles. AAAI, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 239 + ], + "type": "text", + "content": "[30] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 297 + ], + "type": "text", + "content": "[31] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 352 + ], + "type": "text", + "content": "[32] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 287, + 408 + ], + "type": "text", + "content": "[33] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 410, + 287, + 464 + ], + "type": "text", + "content": "[34] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 465, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 465, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 465, + 287, + 498 + ], + "type": "text", + "content": "[35] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 499, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 499, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 499, + 287, + 532 + ], + "type": "text", + "content": "[36] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. TPAMI, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 533, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 287, + 555 + ], + "type": "text", + "content": "[37] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 556, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 589 + ], + "type": "text", + "content": "[38] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. arXiv preprint arXiv:2001.05201, 2020. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 633 + ], + "type": "text", + "content": "[39] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 678 + ], + "type": "text", + "content": "[40] Yasheng Sun, Hang Zhou, Ziwei Liu, and Hideki Koike. Speech2talking-face: Inferring and driving a face with synchronized audio-visual representation. In *IJCAI*, 2021. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "text", + "content": "[41] Yasheng Sun, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Zhibin Hong, Jingtuo Liu, Errui Ding, Jingdong Wang, Zwei Liu, and Koike Hideki. Masked lip-sync prediction by" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "audio-visual contextual exploitation in transformers. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "text", + "content": "[42] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 196 + ], + "type": "text", + "content": "[43] Jiaxiang Tang, Kaisiyuan Wang, Hang Zhou, Xiaokang Chen, Dongliang He, Tianshu Hu, Jingtuo Liu, Gang Zeng, and Jingdong Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[44] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, pages 716-731. Springer, 2020. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[45] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[46] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "type": "text", + "content": "[47] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[48] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "type": "text", + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 533 + ], + "type": "text", + "content": "[50] Haozhe Wu, Jia Jia, Haoyu Wang, Yishun Dou, Chao Duan, and Qingshan Deng. Imitating arbitrary talking style for realistic audio-driven talking face synthesis. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1478-1486, 2021. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "text", + "content": "[51] Chao Xu, Jiangning Zhang, Miao Hua, Qian He, Zili Yi, and Yong Liu. Region-aware face swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7632-7641, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 580, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 634 + ], + "type": "text", + "content": "[52] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[53] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 2, 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[54] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1514" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 485 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "learning-based personalized head pose. arXiv e-prints, pages arXiv-2002, 2020. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[55] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "text", + "content": "[56] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9459-9468, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "text", + "content": "[57] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3867-3876, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 328 + ], + "type": "text", + "content": "[58] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 395 + ], + "type": "text", + "content": "[59] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 287, + 441 + ], + "type": "text", + "content": "[60] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 485 + ], + "type": "text", + "content": "[61] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1515" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_content_list.json b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..58c1841ff604649107ab57991d5fa6dba11dd176 --- /dev/null +++ b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_content_list.json @@ -0,0 +1,2262 @@ +[ + { + "type": "text", + "text": "SunStage: Portrait Reconstruction and Relighting using the Sun as a Light Stage", + "text_level": 1, + "bbox": [ + 76, + 157, + 890, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Wang1 Aleksander Holynski1 Xiuming Zhang2 Xuaner Zhang2 \n1University of Washington 2Adobe Inc. sunstage.cs.washington.edu", + "bbox": [ + 181, + 207, + 787, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72a12d15f29c5906340f76c801ccc4c15a96e5877a8f469d131cc56faa4009a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 284, + 243, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1647eeccb2d79f41272998fc706007083d921392df0450417e5c22ca93e0c29.jpg", + "image_caption": [ + "(a) Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 86, + 407, + 243, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48fcdcab75a29a0909e9910f52cc24f3b29967214d99a9a5c9b46f441cd61f58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 284, + 401, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c070d1dc26fcb9105ab4de5ebce71ebaa7d1d35fa9e87583a3f0d6023ba520bf.jpg", + "image_caption": [ + "(b) Edit lighting" + ], + "image_footnote": [], + "bbox": [ + 246, + 407, + 400, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe54843eb01bf19c9fcc6df87800725b06c232909981659bb78d65e4737af92a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 284, + 560, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d94d69297e90d8ffe7512bbcbe3e7e0cdc582872dfacb0af8f04be6a3798a7f6.jpg", + "image_caption": [ + "(c) Soften shadows", + "Figure 1. Given a selfie video rotating under the sun, SunStage reconstructs geometry, material, camera pose, and lighting information. This recovered information can be used to (a) realistically re-render the input images, (b) modify the lighting conditions by adding / removing lights, (c) soften harsh shadows by changing the size of the reconstructed light sources (d) render the person in an entirely new environment, and (e) edit the albedo or material properties to add freckles, makeup, or stickers that realistically interact with scene lighting." + ], + "image_footnote": [], + "bbox": [ + 403, + 407, + 558, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dea3acc30753508cf83b136bde64244904c67a4a3436c4b4eb6638c914885923.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 284, + 717, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/251f48050b9b0dfa4cb01dcfb64672ba2facd76b6241852f1822a5184c846272.jpg", + "image_caption": [ + "(d) Swap lighting" + ], + "image_footnote": [], + "bbox": [ + 560, + 407, + 715, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/277e23ea4b2eba9a410a8677de4158bc29ea2e68c64353837eefcc92899a625a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 284, + 875, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6aefb98a22232c890df094f89cd1b64eeb7491df40ec494e3281c082c1137599.jpg", + "image_caption": [ + "(e) Edit texture" + ], + "image_footnote": [], + "bbox": [ + 718, + 407, + 872, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 616, + 313, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A light stage uses a series of calibrated cameras and lights to capture a subject's facial appearance under varying illumination and viewpoint. This captured information is crucial for facial reconstruction and relighting. Unfortunately, light stages are often inaccessible: they are expensive and require significant technical expertise for construction and operation. In this paper, we present SunStage: a lightweight alternative to a light stage that captures comparable data using only a smartphone camera and the sun. Our method only requires the user to capture a selfie video outdoors, rotating in place, and uses the varying angles between the sun and the face as guidance in joint reconstruction of facial geometry, reflectance, camera pose, and lighting parameters. Despite the in-the-wild un-calibrated setting, our approach is able to reconstruct detailed facial appearance and geometry, enabling compelling effects such as relighting, novel view synthesis, and reflectance editing.", + "bbox": [ + 75, + 636, + 472, + 893 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 616, + 630, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A light stage [11] acquires the shape and material properties of a face in high detail using a series of images captured under synchronized cameras and lights. This captured information can be used to synthesize novel images of the subject under arbitrary lighting conditions or from arbitrary viewpoints. This process enables a number of visual effects, such as creating digital replicas of actors that can be used in movies [1] or high-quality postproduction relighting [46].", + "bbox": [ + 496, + 642, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In many cases, however, it is often infeasible to get access to a light stage for capturing a particular subject, because light stages are not easy to find: they are expensive and require significant technical expertise (often teams of people) to build and operate. In these cases, hope is not lost — one can turn to methods that are trained on light stage data, with the intention of generalizing to new subjects. These methods do not require the subject to be captured by a light stage but instead use a machine learning", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20792", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model trained on a collection of previously acquired light stage captures to enable the same applications as a light stage, but from only one or several images of a new subject [6, 25, 30, 38, 40, 50, 52]. Unfortunately, these methods have difficulty faithfully reproducing and editing the appearance of new subjects, as they lack much of the signal necessary to resolve the ambiguities of single-view reconstruction, i.e., a single image of a face can be reasonably explained by different combinations of geometry, illumination, and reflectance.", + "bbox": [ + 75, + 90, + 468, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an intermediate solution — one that allows for personalized, high-quality capture of a given subject, but without the need for expensive, calibrated capture equipment. Our method, which we dub SunStage, uses only a handheld smartphone camera and the sun to simulate a minimalist light stage, enabling the reconstruction of individually-tailored geometry and reflectance without specialized equipment. Our capture setup only requires the user to hold the camera at arm's length and rotate in place, allowing the face to be observed under varying angles of incident sunlight, which causes specular highlights to move and shadows to swing across the face. This provides strong signals for the reconstruction of facial geometry and spatially-varying reflectance properties. The reconstructed face and scene parameters estimated by our system can be used to realistically render the subject in new, unseen lighting conditions — even with complex details like self-occluding cast shadows, which are typically missing in purely image-based relighting techniques, i.e., those that do not explicitly model geometry. In addition to relighting, we also show applications in view synthesis, correcting facial perspective distortion, and editing skin reflectance.", + "bbox": [ + 75, + 243, + 470, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions include: (1) a novel capture technique for personalized facial scanning without custom equipment, (2) a system for optimization and disentanglement of scene parameters (geometry, materials, lighting, and camera poses) from an unaligned, handheld video, and (3) multiple portrait editing applications that produce photorealistic results, using as input only a single selfie video.", + "bbox": [ + 75, + 575, + 468, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 76, + 696, + 222, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face modeling. Extensive research has been devoted to the modeling of human faces, leading to various 3D morphable models (3DMMs) [2, 3, 5-7, 9, 10, 25, 31-33, 43]. These models are parametric (maybe in the form of neural networks [33]), allowing one to express variations compactly with a vector. They also encode strong priors learned from real scans. The groundbreaking face 3DMM is that of Blanz and Vetter [3] containing models for shape, expression, and appearance (the Phong model). Also influential is the FLAME model [25] that uses vertex-based Linear Blend Skinning (LBS). FLAME is described by a mapping from", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shape, pose, and expression vectors to a list of vertices. We refer the reader to the survey by Egger et al. [13] for different face morphable models.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such parametric face models provide a low-dimensional space for optimization or learning algorithms. DECA [14] uses the FLAME model to estimate detailed facial geometry (and albedo) from single images, by predicting additional displacement maps and adding them to the estimated FLAME models. More recently, NextFace [12] employs the 3DMM geometry and albedo priors to learn an albedo residual that captures more facial details.", + "bbox": [ + 496, + 136, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Without modeling 3D face geometry, researchers have also achieved photorealistic synthesis of portrait images using generative models and large-scale high-quality image datasets [22, 23].", + "bbox": [ + 496, + 257, + 890, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light stage capture. The light stage as described in Debevec et al., achieves impressive portrait reconstruction and relighting by capturing a series of images of the face under varying illumination [11]. Subsequent work made this process faster, more efficient, and explored different types of illuminants [15, 16, 28].", + "bbox": [ + 496, + 330, + 890, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given that a light stage is not always accessible, a number of methods have been proposed to achieve similar outputs from a single (or few) input portrait images [18,19,29,30,40,41,48-50]. These method rely on a dataset of light stage captures or synthetic examples as training data.", + "bbox": [ + 496, + 422, + 890, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our setup can be thought of as a \"minimalist light stage\" formed by just the sun and a rotating camera, without requiring the high construction and maintenance costs of building a light stage. This parameterization of a sun and skylight model has been shown to be effective in photometric stereo [17,21] and scene factorization [27,42]. In a similar spirit, Calian et al. [8] focus on lighting estimation using faces as \"light probes\". Sengupta et al. propose to circumvent the need for a complicated light stage by recording the facial appearance responses to varying contents displayed on a desk monitor, and then perform portrait relighting [37]. Sevastopolsky et al. also attempt to simplify the capture setup from a light stage to a mobile phone camera with a co-located flash [39]. Unlike our work, which is physically-based, their approaches use neural rendering, and therefore have less direct control over lighting, material, and scene parameters.", + "bbox": [ + 496, + 512, + 890, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Overview", + "text_level": 1, + "bbox": [ + 500, + 784, + 604, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method targets accurate reconstruction of scene lighting, subject geometry, and material properties from a handheld video sequence of a person rotating in place under the sun. Given a selfie video, we take a test-time optimization approach that uses the information from all frames of the video to solve for a physical model of the", + "bbox": [ + 496, + 810, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20793", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c538ae1aca93b227dc391b75f76279fb153275fca6f86a09cbc7ad677a06251.jpg", + "image_caption": [ + "Figure 2. Overview. Our method jointly reconstructs geometry, skin reflectance, lighting, and camera pose from a selfie video sequence of a person rotating under the sun. Our system begins by extracting supervisory information from the video sequence: facial landmarks, foreground alpha mattes, and camera orientations. These are used to supervise the optimization of a collection of scene parameters (full list in Sec. 5.2) used in a physically-based renderer. The rendered output is an image consisting of diffuse, specular, and ambient light contributions. After optimization, the solved scene parameters can be used for a number of editing applications, shown in Sec. 7." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 336, + 358 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/74643e55364fe4de4bc3d155f6e3eae252ad6fc92e38065d2e9321af88206377.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 87, + 591, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1070b43376eb454b7e7766e07c1e5befa01db106d541ebed13ad6e99df8ab4cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 88, + 890, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scene: the geometry and material properties of the face, scene lighting, and camera parameters (Fig. 2). This physical model consists of a base face shape parameterized by a low-dimensional deformable model $X^b$ , a displacement map $\\Delta X$ , a reflectance model with diffuse $R^d$ and specular components $R^s$ , scene lighting $L_i$ , and a perspective camera $C$ . These components are explained in detail in Sec. 4.", + "bbox": [ + 75, + 474, + 468, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After this model has been recovered, we can modify the scene and the subject parameters to re-render images. We show several editing applications in Sec. 7: editing skin reflectance, relighting with arbitrary environment map, improving harsh lighting conditions (by softening shadows and adding fill lights), and adjusting camera parameters to change viewpoint or manipulate perspective effects.", + "bbox": [ + 75, + 580, + 470, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Formulation", + "text_level": 1, + "bbox": [ + 76, + 705, + 207, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input video, our system reconstructs the parameters of a physical model: i.e., geometry and reflectance of the subject, the scene lighting parameters, and the camera parameters. In this section, we detail all of these parameters and describe the rendering process that turns these parameters into an image.", + "bbox": [ + 75, + 733, + 470, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometry. We denote $X_{j}$ as the full mesh of the subject for frame $j$ , composed of a per-frame coarse mesh $X_{j}^{b}(\\beta, \\theta_{j}, \\psi_{j})$ and a global displacement map $\\Delta X$ . The coarse mesh $X_{j}^{b}$ is a FLAME deformable face model [25]", + "bbox": [ + 75, + 837, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defined by global shape code $\\beta$ , per-frame pose code $\\theta_{j}$ , and per-frame expression code $\\psi_{j}$ . $X_{j}^{b}$ also contains per-vertex UV coordinates, which maintain correspondence across variations in $\\theta_{j}$ and $\\psi_{j}$ . As such, we model all our global (per-subject) spatially varying parameters in UV space, and sample values per-fragment when rasterizing.", + "bbox": [ + 496, + 474, + 892, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The displacement map $\\Delta X$ is used to model fine details like wrinkles that cannot be represented by $X_{j}^{b}$ . We displace the coarse geometry by $\\Delta X$ at rasterization time, by sampling a displacement value per-fragment and displacing each fragment along the surface normal $N_{j}$ of the coarse mesh $X_{j}^{b}$ . After displacement, the updated fragment positions are used to compute a new surface normal $N_{j}^{\\prime}$ .", + "bbox": [ + 496, + 566, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$X_{j}^{b}$ is optimized per-frame, since it accounts for subtle (and unavoidable) variations in expression and head pose during the capture, which are modeled by $\\theta_{j}$ and $\\psi_{j}$ . $\\Delta X$ , on the other hand, is optimized in UV-space (i.e., globally per-subject), since the deformations it encodes are invariant to the changes in expression or pose. Formally, the final geometry $X_{j}$ is given by:", + "bbox": [ + 496, + 674, + 890, + 781 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX _ {j} = X _ {j} ^ {b} \\left(\\beta , \\theta_ {j}, \\psi_ {j}\\right) + \\Delta X \\odot N _ {j} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 794, + 890, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard product.", + "bbox": [ + 500, + 825, + 730, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reflectance. We model the skin reflectance, denoted as $R(x,\\omega_i,\\omega_o) \\in \\mathbb{R}^3$ , where $x$ is a 3D point on the face geometry $X$ , $\\omega_i$ is the incoming light direction, and $\\omega_o$ is the", + "bbox": [ + 498, + 854, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20794", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "outgoing direction, using a diffuse and a specular component: $R = R^{d} + R^{s}$ .", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diffuse component $R^d (x,\\omega_i)\\in \\mathbb{R}^3$ is a Lambertian reflectance model consisting of an albedo map, $a$ , which we optimize as a per-subject UV-space image. For the skin's specular component, we use the Blinn-Phong model [4].", + "bbox": [ + 75, + 119, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right) = k _ {s} \\frac {s + 2}{2 \\pi} \\left(h \\left(\\omega_ {i}, \\omega_ {o}\\right) \\cdot n (x)\\right) ^ {s} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 200, + 468, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h(\\omega_i, \\omega_o) = \\text{normalize}(\\omega_i + \\omega_o)$ is the half vector, $k_s$ is the specular intensity, $s$ is the specular exponent, and $(s + 2) / (2\\pi)$ is the normalization term for the reflection lobe to integrate to 1. Following [47], we segment the UV-space map into 10 segmented specular reflectance clusters. We then optimize for a spatially-varying pair of values $(s, k_s)$ per-cluster, enabling varying shininess across the face.", + "bbox": [ + 75, + 236, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While Blinn-Phong does not model many complex effects such as subsurface scattering, our experiments with other models for facial reflectance, such as microfacet models [44], show no significant quality improvements, and often introduce unstable training. More analysis is provided in the supplementary material.", + "bbox": [ + 75, + 342, + 468, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lighting. We use a sun-sky model to represent lighting as the sum of an \"ambient\" environment map and the sun: $L_{i}(x,\\omega_{i}) = L_{i}^{\\mathrm{amb}}(\\omega_{i}) + L_{i}^{\\mathrm{sun}}(\\omega_{i})$ . Note neither $L_{i}^{\\mathrm{amb}}(\\omega_{i})$ nor $L_{i}^{\\mathrm{sun}}(\\omega_{i})$ depends on the 3D point $x$ , since we model both as directional lights. Optimization-wise, our lighting parameters consist of a $16 \\times 32 \\times 3$ environment map for ambient lighting, the sun direction $p^{\\mathrm{sun}} \\in S^{3}$ , and the scalar sun intensity $k^{\\mathrm{sun}}$ . We fix the sun color to white [1, 1, 1] in our lighting model to resolve the albedo-illumination ambiguity.", + "bbox": [ + 75, + 444, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Rendering", + "text_level": 1, + "bbox": [ + 76, + 604, + 194, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We calculate the outgoing radiance $L_{o}$ at 3D location $x$ as viewed from viewing direction $\\omega_{o}$ as:", + "bbox": [ + 75, + 627, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {o} (x, \\omega_ {o}) \\\\ = \\int_ {\\Omega} V \\left(x, \\omega_ {i}\\right) L _ {i} \\left(x, \\omega_ {i}\\right) \\odot R \\left(x, \\omega_ {i}, \\omega_ {o}\\right) \\left(\\omega_ {i} \\cdot n (x)\\right) d \\omega_ {i} (3) \\\\ = \\sum_ {\\omega_ {i}} V (x, \\omega_ {i}) \\left(L _ {i} ^ {\\mathrm {a m b}} (\\omega_ {i}) \\odot R ^ {d} (x, \\omega_ {i}) \\right. (4) \\\\ + L _ {i} ^ {\\text {s u n}} \\left(\\omega_ {i}\\right) \\odot R ^ {d} \\left(x, \\omega_ {i}\\right) + L _ {i} ^ {\\text {a m b}} \\left(\\omega_ {i}\\right) \\odot R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right) \\\\ \\left. + L _ {i} ^ {\\operatorname {s u n}} \\left(\\omega_ {i}\\right) \\odot R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right)\\right) \\left(\\omega_ {i} \\cdot n (x)\\right) \\triangle \\omega_ {i} (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 680, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V(x,\\omega_{i})$ is the light visibility at $x$ from $\\omega_{i}$ , and $L_{i}(x,\\omega_{i})$ is the incoming radiance reaching $x$ from $\\omega_{i}$ . We ignore the specular reflection caused by the ambient lightning, i.e., $L_{i}^{\\mathrm{amb}}(\\omega_{i})\\odot R^{s}(x,\\omega_{i},\\omega_{o})$ , since it is much weaker", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "than the specular reflection of the sun. In the next subsections, we will group the terms into a diffuse contribution $L_{o}^{d}$ and a specular contribution $L_{o}^{s}$ : $L_{o} = L_{o}^{d} + L_{o}^{s}$ . For the final rendered color value, we apply the Reinhard operator [35] and a gamma correction of $\\gamma = 2.2$ to $L_{o}$ to convert from linear to sRGB space.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffuse contribution. The diffuse contribution $L_{o}^{d}$ is then given by only the diffuse terms of Equation 5:", + "bbox": [ + 496, + 193, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {o} ^ {d} (x) = \\sum_ {\\omega_ {i}} L _ {i} ^ {\\mathrm {a m b}} (\\omega_ {i}) \\odot \\frac {a (x)}{\\pi} (\\omega_ {i} \\cdot n (x)) \\Delta \\omega_ {i} \\\\ + V (x, p ^ {\\text {s u n}}) k ^ {\\text {s u n}} [ 1, 1, 1 ] \\odot \\frac {a (x)}{\\pi} \\left(p ^ {\\text {s u n}} \\cdot n (x)\\right) \\Delta p ^ {\\text {s u n}} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 236, + 890, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a(x)$ is the albedo at point $x$ , $k^{\\mathrm{sun}}$ is the (optimized) sun intensity, and $p^{\\mathrm{sun}}$ is the (optimized) sun direction. The sun is modeled as a directional light source, so the second summation can be simplified to a single term, only in the direction of $p^{\\mathrm{sun}}$ . We additionally optimize for a high-dynamic-range (HDR) environment map $E \\in \\mathbb{R}^{16 \\times 32 \\times 3}$ , from which values of $L_{i}^{\\mathrm{amb}}$ are sampled.", + "bbox": [ + 496, + 330, + 890, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specular contribution. The specular contribution $L_{o}^{s}$ at each pixel is given by only the specular term due to the sun in Equation 5 (recall that we ignore the specular ambient term due to its weak contribution):", + "bbox": [ + 496, + 449, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {o} ^ {s} (x, \\omega_ {o}) = V (x, p ^ {\\mathrm {s u n}}) k ^ {\\mathrm {s u n}} [ 1, 1, 1 ] k _ {s} \\frac {s + 2}{2 \\pi} \\\\ \\left(h \\left(p ^ {\\text {s u n}}, \\omega_ {o}\\right) \\cdot n (x)\\right) ^ {s} \\left(p ^ {\\text {s u n}} \\cdot n (x)\\right) \\Delta p ^ {\\text {s u n}} \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 513, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we have substituted Equation 2 and reduced the summation to just one term at $p^{\\mathrm{sun}}$ (since $L_{i}^{\\mathrm{sun}}$ is 0 elsewhere).", + "bbox": [ + 496, + 566, + 890, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shadow map. In order to generate a map of self-occluded shadows, we perform two passes of rasterization: first, we render a $z$ -buffer from a virtual orthographic camera aligned with the sun direction, $p^{\\mathrm{sun}}$ , and then, when rasterizing a given camera viewpoint, compare all fragment positions $d_{\\mathrm{hit}}$ to the light's $z$ -buffer $d_{\\mathrm{shadow}}$ . To avoid precision issues and ensure smooth gradients for back-propagation, we implement a soft comparison as follows in generating shadow/visibility maps:", + "bbox": [ + 496, + 609, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV \\left(x, \\omega_ {i}\\right) = 1 - \\text {s i g m o i d} \\left(k \\left(d _ {\\text {h i t}} - d _ {\\text {s h a d o w}} \\times b\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 758, + 890, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k$ is the falloff slope, and $b$ the tolerance. We use $k = 800$ and $b = 1.0015$ .", + "bbox": [ + 496, + 785, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Optimization", + "text_level": 1, + "bbox": [ + 500, + 829, + 635, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The described physical model contains a large number of parameters to be optimized, controlling scene elements like lighting, geometry, pose, and texture. Unfortunately,", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20795", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "naively optimizing all these parameters from scratch does not result in an optimal solution, since the final observed appearance of the face can often be explained variously through changes to geometry, material properties, lighting or camera parameters, making optimization severely under-constrained and ambiguous. Therefore, we adopt a two-stage optimization approach, through which parameters are gradually enabled. In this section, we describe this process and the relevant losses that guide optimization.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Coarse alignment", + "text_level": 1, + "bbox": [ + 76, + 239, + 250, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our system begins by using an off-the-shelf network (DECA [14]) to generate, for each input image, a set of shape parameters $\\beta$ , pose parameters $\\theta_{j}$ , and expression parameters $\\psi_{j}$ of a FLAME face model [25], as well as the relative pose parameters of the virtual camera observing the 3D face. Unfortunately, as with many other single-image facial geometry estimators, DECA assumes an orthographic projection model and therefore cannot accurately recover geometry for our selfie capture sequences, which contain heavy perspective effects (Figure 6). Without a good initialization for geometry, optimization of lighting and material properties seldom converges to an optimal solution due to the heavily ambiguous nature of our optimization problem.", + "bbox": [ + 75, + 263, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To circumvent this issue, we employ a first stage of optimization where we only optimize for the parameters of a perspective camera (with a known focal length, extracted from input metadata) and the face geometry parameters $(\\beta, \\theta_j, \\psi_j)$ . As initialization for this optimization process, we use the predicted DECA values for each frame's pose $\\theta_j$ and expression $\\psi_j$ , but set all frames to the average predicted shape $\\beta_{\\mathrm{avg}} = \\frac{1}{N} \\sum_j \\beta_j$ , since the identity remains constant across all frames. To convert DECA's orthographic camera to a perspective camera, we additionally optimize for an unknown scale $S$ and translation $T_j$ , which are initialized to empirically chosen values $S = 2.6e4$ , $T_j = (0, 0, 1.5e5)$ . During optimization, the face shape $\\beta$ and scale $S$ are shared across all frames, while camera pose $T_j$ , expression $\\psi_j$ , and pose $\\theta_j$ are optimized per-frame. Note that DECA controls the relative orientation of the camera and the face by varying the pose code $\\theta$ instead of the camera rotation. We adopt this formulation and keep the camera orientation fixed relative to the face. The global orientation of the camera at each frame (and therefore the face) is extracted from the capture video, either through a structure-from-motion system or IMU measurements commonly available on a smartphone.", + "bbox": [ + 75, + 462, + 472, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use two losses to guide this optimization: a mask loss $L_{\\mathrm{mask}}$ and a landmark loss $L_{\\mathrm{lmk}}$ . The FLAME model includes 3D facial landmark points, corresponding to the standard 68-point facial landmarks set [36] used in facial tracking. Our landmark loss minimizes the L1 distance between the 2D projection of these 3D landmarks (into the", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "input camera viewpoint) and 2D landmarks estimated from the input frame by a 2D landmark detector HRNets [45].", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The facial landmarks provide a stong constraint on facial feature alignment, but are sparse, and therefore cannot constrain the overall shape or boundary of the mesh. To supplement it, we include a silhouette loss $L_{\\mathrm{mask}}$ , which penalizes the L2 difference between the rasterized mask of the mesh $I_{\\mathrm{sil}}$ and the semantic segmentation mask $I_{\\mathrm{mask}}$ of the input image, using an off-the-shelf semantic segmentation network [26] trained to segment humans in portrait photographs.", + "bbox": [ + 496, + 121, + 893, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final pose loss is then: $L_{\\mathrm{pose}} = L_{\\mathrm{mask}} + L_{\\mathrm{lmk}}$ , optimized using an ADAM optimizer [24]. See supplemental for optimization parameters.", + "bbox": [ + 496, + 257, + 893, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Photometric optimization", + "text_level": 1, + "bbox": [ + 500, + 310, + 732, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the 3D model and camera parameters are approximately aligned, we proceed to the second stage of optimization, in which we optimize the precise facial geometry, lighting, and reflectance properties. All the parameters optimized in the first stage (Section 5.1) remain as free variables. In total, the parameters optimized during this stage include: Lighting parameters: (1) $p^{\\mathrm{sun}}$ , the global sun direction, (2) $E$ , the global environment map, (3) $k^{\\mathrm{sun}}$ , the global sun intensity, Facial geometry parameters: (4) $\\beta$ , the global FLAME shape code, (5) $\\psi_j$ , the per-frame expression code, (6) $\\theta_j$ , the per-frame pose code, (7) $\\Delta X$ , the global deformation map, Material properties: (8) $k_s$ , the global, spatially-varying specular intensity, (9) $s$ , the global, spatially-varying specular roughness, (10) $a$ , the global, spatially-varying surface albedo, Camera pose parameters: (11) $T_j$ , the per-frame perspective camera translation, and (12) $S$ , the global scene scale.", + "bbox": [ + 496, + 333, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During optimization, we randomly select a frame $j$ , render the face using a differentiable rasterizer [34] and the equations described in Section 4 to get the rendered image $\\hat{I}$ . In addition to the previously defined landmark and mask losses, we include L2 and VGG [20] photometric losses, comparing the original and reconstructed images:", + "bbox": [ + 496, + 590, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p h o t o}} = \\left| \\left| \\hat {I} _ {j} \\cdot I _ {\\text {s i l}} - I _ {j} \\cdot I _ {\\text {m a s k}} \\right| \\right| _ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 688, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also include an L2 regularization $L_{E}$ and L2-smoothness regularization $L_{Es}$ on the reconstructed environment map, to encourage the majority of the lighting to be explained by direct sunlight and to aid in disentangle-ment of the sun and ambient lighting. The total optimized loss becomes:", + "bbox": [ + 496, + 712, + 890, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\lambda_ {\\text {m a s k}} L _ {\\text {m a s k}} + \\lambda_ {\\text {l m k}} L _ {\\text {l m k}} + \\lambda_ {E} L _ {E} + \\lambda_ {E \\mathrm {s}} L _ {E \\mathrm {s}} \\\\ + \\lambda_ {\\mathrm {V G G}} L _ {\\mathrm {V G G}} + \\lambda_ {\\text {p h o t o}} L _ {\\text {p h o t o}} \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 811, + 890, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\lambda_{\\mathrm{mask}}, \\lambda_{\\mathrm{lmk}} = 0.05, \\lambda_{\\mathrm{VGG}} = 0.005, \\lambda_{E} = 0.01, \\lambda_{Es}, \\lambda_{photo} = 1$ . Additional optimization details are provided in the supplemental materials.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20796", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/52f5d3506df2da247be34ce6af111df2b650be54aec27ea643eda1ba2a6e61d0.jpg", + "image_caption": [ + "(a) DECA [14]" + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 181, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b88f041e2d99bc9832219eb121f46731d55fb38da3a7582a7c6c43d89a550965.jpg", + "image_caption": [ + "(b) DPR [52]" + ], + "image_footnote": [], + "bbox": [ + 181, + 88, + 282, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b20f9c935a049ac92da3097ee7dc4c6e57419465fb34578d57cd5f3a10deeea.jpg", + "image_caption": [ + "(c) GCFR [18]" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 383, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31ca6ac7b4b2f6315f2ab8c53c295c821ec0d9980c4f34f83059d409ca250085.jpg", + "image_caption": [ + "(d) TR [30]" + ], + "image_footnote": [], + "bbox": [ + 385, + 88, + 483, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d089802bbdb35ea642b96694c0e65a03343638319389c7936a9d5debda5d5f4d.jpg", + "image_caption": [ + "(e) NextFace [12]" + ], + "image_footnote": [], + "bbox": [ + 485, + 88, + 586, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/389ed353915131fb1aa231f1c2d610b58e8e3feb2c7080222657736a3d4a0dc4.jpg", + "image_caption": [ + "(f) NLT [51]" + ], + "image_footnote": [], + "bbox": [ + 586, + 88, + 686, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96d06ce9832da61fc03bc2768eb050b47a1bc29c66894022180b1ba67e22a1a8.jpg", + "image_caption": [ + "(g) Ours" + ], + "image_footnote": [], + "bbox": [ + 687, + 88, + 787, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0481f3271eaac39ce508bacd29ddae5422c60f9fbf655c4ecb1d0583ce0f336d.jpg", + "image_caption": [ + "(h) GT" + ], + "image_footnote": [], + "bbox": [ + 789, + 88, + 888, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63f889f69e65178df554d6b6108e518b25678feffe0ccc03b4168367f1687eea.jpg", + "image_caption": [ + "Figure 3. Qualitative: Relighting. A comparison of our method at rendering a new (unseen) lighting environment (h). Our method is able to realistically synthesize the novel lighting condition, including cast shadows and specularity, and nearly matches the (unseen) target reference image. See supplement for additional details on experimental setup and analysis of results." + ], + "image_footnote": [], + "bbox": [ + 84, + 337, + 215, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70e7609460fb80139be943c0d1a9362611ce45a7b84481341cfbffc5f0136549.jpg", + "image_caption": [ + "(a) Nearest view" + ], + "image_footnote": [], + "bbox": [ + 84, + 439, + 215, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d77e6ffd4dac729bcfd27b8c6405737fd3a4b7884c8323dae15b42d6bf79f8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 337, + 349, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82e79896fb021f99081dd578fd2625fafc214ce25057acfc1b4723ba93a479ac.jpg", + "image_caption": [ + "(b) DECA [14]" + ], + "image_footnote": [], + "bbox": [ + 220, + 439, + 349, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f28d4fe4a2a58d714b7a518425b1dbcec9bbd209968c5734c4a31b160408397.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 337, + 483, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e987664c92a9483dd4062b5b292a9ec18f9bc0d72d4f23098f2a144c31c6720.jpg", + "image_caption": [ + "(c) NextFace [12]" + ], + "image_footnote": [], + "bbox": [ + 352, + 439, + 483, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca138ccb57cb9ba789ab336cce4533901e3033782cf28b11ab3f33ee005bfbd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 337, + 617, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b3bc4291722f72be538d4a74369afc093b808eff9488db8778dc3184ae9b98c.jpg", + "image_caption": [ + "(d) NLT [51]" + ], + "image_footnote": [], + "bbox": [ + 486, + 439, + 617, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44c665776405e58528ae23d3d3f1772ce139be37341cdf658ff3c2313fa970dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 337, + 751, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62114643901ca4b5168b3d1a5f302c9b661457ad37214e278b78d936931780bf.jpg", + "image_caption": [ + "(e) Ours", + "Figure 4. Qualitative: view synthesis. A comparison of our method at the task of generating an image from an unseen viewpoint (f), having only seen a limited collection of input viewpoints. See supplement for more details on experimental setup and analysis of results." + ], + "image_footnote": [], + "bbox": [ + 619, + 439, + 750, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d45431b527777126b725e35b985f6217c22f1f8fee6f15cbf90992fcab6b65c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 337, + 883, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9410ea937733907d2ef32f96d3cac579537b70d2644566c9c5c992a047668017.jpg", + "image_caption": [ + "(f) GT" + ], + "image_footnote": [], + "bbox": [ + 754, + 439, + 883, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 633, + 192, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we detail quantitative and qualitative experiments comparing our approach with state-of-the-art methods and ablated variants of our method.", + "bbox": [ + 75, + 659, + 468, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline comparisons. We evaluate our method on the tasks of novel-view synthesis and relighting. For novel-view synthesis, we compare our method with DECA [14], Neural Light Transport (NLT) [51], and NextFace [12]. For relighting, we compare with DECA, NLT, NextFace, GCFR [18], image-based methods Deep Single Image Portrait Relighting (DPR) [52], Total Relighting (TR) [30] and NVPR [49]. Additional comparisons and details on the experimental setups are provided in the supplemental materials.", + "bbox": [ + 75, + 718, + 467, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present qualitative comparisons for relighting in Figure 3 and novel view synthesis in Figure 4. Quantitative", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comparisons on these images are provided in Table 1. These testing images consist of (1) a multi-view capture of the face, in which the subject remains still and the camera is moved to novel viewpoints in the same environment as the original capture, and (2) front-facing sequences in novel environment lighting and unseen sun positions. All testing images are not seen during training of our method, NLT or NextFace. The results shown in Figures 3 and 4 as well as Table 1 clearly demonstrate that our method outperforms all the baselines at both relighting and view synthesis. Single-image methods (DECA, GCFR, DPR, TR) can generalize to other subjects, but fail to recover more faithful and physically accurate facial details. Comparison with multi-image methods (NLT, NextFace) demonstrates that SunStage is a better reconstruction system. Additional analysis of the comparisons is provided in the supplemental materials.", + "bbox": [ + 496, + 635, + 890, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20797", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32d160419810920fd0346182a686c69c47e9250e107026d7fdd6949f3a73e12e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 196, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9c33fb6b2de3789d68733d03be92948914101c742df20584b1d85dca464b8d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 88, + 310, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/630c6c2a58f07c1a036f57b37fcda6ab6d1e63dff504fca8a40e17bf34b0e098.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 88, + 424, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/baa9a6d1b3736c55f5b66c91b53bf2f5ec88c8939fef3fa5e537798ebb65ec8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 88, + 539, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae520b4158da99aaab95b336a3cd11a13efdd3785beba4a352a232fad5d6fd4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 88, + 653, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c97590b23c88a48e4b282520d1b12ce392391de9ff745e51becfcc4529d98033.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 88, + 769, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f75e2f17afc0f648eeb2b37139e7832cbcea571a55da79424c9c225d3dd2707.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 88, + 887, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6645e13675557b3ed95647b24a162194f1325aa3bdfc2284aec35b0bb80e8b15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 181, + 196, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c603800cd4d7a07ba485a548889b9690205252c37ff0b9a6688f76970e5fd16b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 181, + 310, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52fc257f218259558efcf5c9436cd8d81749b66ce813448d5c1c6fd185e09655.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 181, + 424, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e430f445978ebbb89f326ef48b1830637e61341664463b1c0d335232870dce32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 181, + 539, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4482c11b62f0c41a2fb5db699013d87b007d426a0ff496208a3120573f8b1565.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 181, + 653, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67af2371cabb1c1e86ca548744ee11cbaaad7045eb5e9059977aa271f00e8ac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 181, + 769, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71e3128a28d7dbfbcfe8cd2eda9293d94b6c2d259eb7fdd84eed344a3ebf91d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 181, + 887, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/710bd2c7abdd4e57051c963654097daf5125ca9f115239f71b1ec7324e6dbf99.jpg", + "image_caption": [ + "Normal" + ], + "image_footnote": [], + "bbox": [ + 81, + 273, + 196, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb4bd385f83cf045e620d1bd44b404475b62a9e3cf0c513ca46a0caf49fe7517.jpg", + "image_caption": [ + "Albedo" + ], + "image_footnote": [], + "bbox": [ + 197, + 273, + 310, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/300c7b77082e13e199cd371d843c25a2833c3500874c1da3f2314b869970b8e9.jpg", + "image_caption": [ + "Specular" + ], + "image_footnote": [], + "bbox": [ + 310, + 273, + 424, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40b0cb98907b7d1f5800d9e60b2f79245ac342a1d9c4d2108b0e4027a53435e7.jpg", + "image_caption": [ + "Ambient" + ], + "image_footnote": [], + "bbox": [ + 424, + 273, + 539, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d4b93c6d8dbe5e719ed206d278823135cdb553fe1fa470a3de5b30acdbba09d.jpg", + "image_caption": [ + "Diffuse" + ], + "image_footnote": [], + "bbox": [ + 540, + 273, + 653, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/122249523a9bed97ec54557dcfc649ec8362bbe4251358b2ec65c4d27f4eca41.jpg", + "image_caption": [ + "Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 655, + 273, + 769, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41828a35e4cf3fffd69f21bf2bf3f0fbfde939fa18d24f0856c5d75a2eb496f2.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 771, + 273, + 887, + 363 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/017ee2a711f1e8edee4def443bed49e6b2fbd136554ab234ef65fdf854686861.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RelightingNovel view synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DECA [14]16.410.690.2516.640.660.29
GCFR [18]16.970.700.20---
DPR [52]19.030.720.19---
NLT [51]20.150.750.1822.270.790.15
Total Relighting [30]20.240.790.16---
NextFace [12]22.980.760.1522.550.750.15
Ours23.640.830.1025.280.840.09
Ours w/o coarse17.830.660.2319.650.700.17
Ours w/o SV ks, s21.310.770.1321.940.770.12
Ours w/o Lmask, Llmk16.460.610.3018.830.680.20
Ours w/o Lmask20.130.750.1520.540.740.15
Ours w/o opt. (β, θi, φi)18.670.690.1918.280.660.19
Ours w/o soft shadow21.460.770.1322.050.770.12
Ours w/o ΔX21.160.750.1521.800.750.14
", + "bbox": [ + 76, + 459, + 470, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison. Comparison of our method on the tasks of novel view synthesis and relighting. See Section 6 for a description of the ablated variants.", + "bbox": [ + 75, + 651, + 468, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Disentanglement. In Figure 5, we demonstrate how Sun-Stage decomposes the appearance of a portrait photograph into different components: specular, diffuse, and ambient. We also visualize the surface normal, albedo, and other intermediate representations to show that our method is able to effectively recover a physically plausible reconstruction of the real world and disentangle the different components that contribute to the final appearance. We further validate the quality of the reconstructed geometry and materials in the supplementary material.", + "bbox": [ + 75, + 720, + 468, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies. In addition to our comparisons with", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c7469f6efaaed23fba602c343523ebed64979bcf417288432441edd0a5a4c02.jpg", + "image_caption": [ + "Figure 5. Decomposition. We show all the components which comprise our final rendered image to demonstrate that our method not only closely recreates the ground truth image (reproducing realistic highlights and shadows), but also performs a meaningful decomposition of lighting components and facial geometry. Note that our reconstructed surface normals include high frequency details specific to each subject, like wrinkles and birthmarks, which are used in computation of the shadows and specular reflections.", + "(a) Input image", + "Figure 6. Perspective. DECA's assumption of an orthographic camera is broken by the strong perspective effects in selfies, causing poor alignment (b) with input images (a). Our first stage of optimization (c) (Sec. 5.1) improves alignment by solving for the parameters of a perspective camera and refined shape parameters. In the second stage we additionally optimize for a displacement map $\\Delta X$ to produce our final shape with finer geometric details like wrinkles (d). Red line added to highlight alignment with (a)." + ], + "image_footnote": [], + "bbox": [ + 503, + 459, + 599, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/befc0e0b5d14c0df09622bb32efca79c3a5ff72feb3f75a6dcfb51c51faea23d.jpg", + "image_caption": [ + "(b) DECA [14]" + ], + "image_footnote": [], + "bbox": [ + 599, + 459, + 694, + 532 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f5bfc03898bc74c2796b4542eeecf4c8492ee6d1fc0f35ca042892d7146b116.jpg", + "image_caption": [ + "(c) Stage 1" + ], + "image_footnote": [], + "bbox": [ + 697, + 459, + 792, + 532 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/644d3fc7d0c0a59ca6ef58f75db49803ef155778a8726cfc755d6e12590d0ecc.jpg", + "image_caption": [ + "(d) Final" + ], + "image_footnote": [], + "bbox": [ + 794, + 459, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the state-of-the-art baselines, we also compare with ablated variants of our own method. In particular, we include seven such experiments in Table 1: our method (1) without the initial first stage of coarse geometric alignment, i.e., directly optimizing both geometric and photometric parameters from the start, (2) without the spatially varying specular parameters, instead using a single global scalar $s$ and $k_{s}$ , (3) without the geometric alignment losses $L_{\\mathrm{mask}}$ and $L_{\\mathrm{lmk}}$ , (4) without just $L_{\\mathrm{mask}}$ , (5) without shape optimization, i.e., keeping the initial shape code predicted by DECA, (6) without soft shadow computation, i.e., using a hard z-buffer comparison to compute a shadow map instead of our soft", + "bbox": [ + 498, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20798", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85fba16c5fbf0821758be6dd5c407c10ffcb4c91ae636bd503da3c288b4c0d3b.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 176, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab834c038346f3deae18213cc163346f51647dbbddf4f6b5f2d122140f5031c3.jpg", + "image_caption": [ + "(b) Soft shadows" + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 272, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ca8f30c33e86b4b60a3039b729a3d85ad3353e57db7f33f571f3ca0b8094206.jpg", + "image_caption": [ + "(c) (b) + fill light" + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 369, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71e20c4ee897b13f12516cc0833afe8536be8551fa1446a62ace94dd31049b55.jpg", + "image_caption": [ + "(d) Swap scene", + "Figure 7. Adjusting lighting parameters. We can adjust the recovered scene parameters to improve the lighting conditions in an input image (a) by softening the harsh shadows cast by the nose (b), adding a fill light to brighten the shaded region (b), or replacing the environment altogether (d)." + ], + "image_footnote": [], + "bbox": [ + 372, + 88, + 467, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparison operator in Equation 8, and (7) without the displacement map $\\Delta X$ . Visual results for each of these variants are provided in the supplemental material.", + "bbox": [ + 76, + 356, + 468, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Applications", + "text_level": 1, + "bbox": [ + 76, + 414, + 207, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relighting. We demonstrate two types of relighting applications: (1) lighting modification and (2) lighting replacement. Practical lighting modification is common in portrait photography when the lighting conditions are not ideal, e.g., when direct sunlight casts undesirable harsh shadows with high contrast. A common practice is to make the light source larger and more diffuse by using a scrim or bounce card. In Fig. 7b, we show that by virtually increasing the size and spreading the energy of our reconstructed lighting source (i.e., the sun), we are able to soften the shadows and re-render a more visually pleasing face. Another approach to reducing the effects of harsh shadows is adding local fill lights, which reduces the contrast between the lit and shaded regions (Fig. 7c). Alternatively, fill lights can also be used for artistic purposes, to create dramatic lighting effects (Fig. 1b). Finally, replacing the scene lighting with that of a novel environment (Fig. 7d) is a necessary step in realistically inserting a captured subject into a virtual scene, which is useful for visual effects and VR applications.", + "bbox": [ + 75, + 452, + 468, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "View Synthesis. In Figure 8, we show that our reconstructed 3D model of the face can be used to synthesize new views by manipulating the viewpoint of the camera. We can also change other camera parameters, such as the focal length, to reduce the perspective effects on the face, which is often desirable for selfie images that contain significant facial distortion due to perspective.", + "bbox": [ + 76, + 752, + 468, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Skin Reflectance Editing. We are also able to edit the reflectance components of the subject. As shown in Figure 1e,", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66dae6f19b9f6d50ceae33285e157f6ccd096bab7dd701c10c305dee0453db39.jpg", + "image_caption": [ + "(a) Short focal" + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 598, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47b3313a562875c48a0277e625984bf02508ff9e849aca500d864a7cf8b7d358.jpg", + "image_caption": [ + "(b) Long focal" + ], + "image_footnote": [], + "bbox": [ + 601, + 88, + 694, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef8f4c0cf1ec2cb121e170c4c7d70c6832dfc8a2870d61cc84814ea46d4145f4.jpg", + "image_caption": [ + "(c) Novel view" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 792, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e42e876402a4bdb2cd81e76b8c1ee78d50de1c3df27ad855407c9d136cf4fbba.jpg", + "image_caption": [ + "(d) Novel view 2" + ], + "image_footnote": [], + "bbox": [ + 794, + 88, + 888, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32a964bbde0e982d6f0a192d8b8fe7a5c30b9df1962a21b0f2f699776cdc8fb5.jpg", + "image_caption": [ + "Figure 8. Changing camera parameters. We can change the recovered camera parameters to render novel views (c,d) or change the focal length (a,b).", + "(a) Less shinier", + "Figure 9. Adjusting the specularity. We can change the the specular properties of the face, making the face less shinier (a, c) or more shinier (b, d)." + ], + "image_footnote": [], + "bbox": [ + 503, + 311, + 599, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/371ce804727e1555a587e8cfd5407b473d8c005e19b8616e728a4864d328780f.jpg", + "image_caption": [ + "(b) More shinier" + ], + "image_footnote": [], + "bbox": [ + 601, + 311, + 696, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1aaa84b6d71119d71f3d7fa59871cff6d9f12c3385a996aef39ac16683409d02.jpg", + "image_caption": [ + "(c) Less shinier" + ], + "image_footnote": [], + "bbox": [ + 697, + 311, + 792, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/066a1cbed01d26fdf80fdf8dd23baed3abadc50ca6826d4a48bfa1f7a3f6481f.jpg", + "image_caption": [ + "(d) More shinier" + ], + "image_footnote": [], + "bbox": [ + 794, + 311, + 888, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we can adjust the optimized albedo to add freckles, stickers, or other textures that realistically interact with reflections, shadows, and other elements of scene lighting, or we can adjust the specular properties of the face, making the face more or less shinier, as shown in Figure 9.", + "bbox": [ + 498, + 470, + 890, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion and Discussions", + "text_level": 1, + "bbox": [ + 500, + 560, + 754, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SunStage, a lightweight and practical facial capture, rendering, and editing system that can serve as a minimalist light stage. With a video of an individual rotating in-place under the sun, our system reconstructs a physical model of the subject and the scene lighting, which enables us to relight the subject with realistic reflections and cast shadows. Our system allows arbitrary lighting and reflectance control in the reconstructed physical space, which can be rendered to produce photo-realistic results. We demonstrate several applications such as editing skin reflectance, relighting, and view synthesis.", + "bbox": [ + 496, + 585, + 890, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our system inherits the limitations of morphable face models and is unable to model hair, teeth, or clothing geometry, beyond slight deformations. Additionally, certain regions which are seen under constant shadow or specular reflection (and therefore have no cues on reflectance or albedo) are sometimes unable to be decomposed accurately into separate reflectance and lighting components. Visualization and further discussions of the system's limitations are provided in the supplemental material.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20799", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Oleg Alexander, Mike Rogers, William Lambeth, Jen-Yuan Chiang, Wan-Chun Ma, Chuan-Chang Wang, and Paul Debevec. The digital emily project: Achieving a photorealistic digital actor. IEEE Computer Graphics and Applications, 30(4):20-31, 2010. 1", + "[2] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4), 2021. 2", + "[3] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 2", + "[4] James F Blinn. Models of light reflection for computer synthesized pictures. In Proceedings of the 4th annual conference on Computer graphics and interactive techniques, pages 192-198, 1977. 4", + "[5] Timo Bolkart and Stefanie Wuhrer. A groupwise multilinear correspondence optimization for 3d faces. In Proceedings of the IEEE international conference on computer vision, pages 3604-3612, 2015. 2", + "[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2", + "[7] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In European Conference on Computer Vision, pages 297-312. Springer, 2014. 2", + "[8] Dan A Calian, Jean-François Lalonde, Paulo Gotardo, Tomas Simon, Iain Matthews, and Kenny Mitchell. From faces to outdoor light probes. In Computer Graphics Forum, volume 37, pages 51-61. Wiley Online Library, 2018. 2", + "[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 2", + "[10] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3085-3093, 2017. 2", + "[11] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 2", + "[12] Abdallah Dib, Cedric Thebault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 2, 6, 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2", + "[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2, 5, 6, 7", + "[15] Graham Fyffe and Paul Debevec. Single-shot reflectance measurement from polarized color gradient illumination. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2", + "[16] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. In Proceedings of the 2011 SIGGRAPH Asia Conference, pages 1-10, 2011. 2", + "[17] Yannick Hold-Geoffroy, Jinsong Zhang, Paulo F U Gotardo, and Jean-François Lalonde. $x$ -hour outdoor photometric stereo. In International Conference on 3D Vision, 2015. 2", + "[18] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2, 6, 7", + "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2", + "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 5", + "[21] Jiyoung Jung, Joon-Young Lee, and In So Kweon. One-day outdoor photometric stereo via skylight estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4521–4529, 2015. 2", + "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[25] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 3, 5", + "[26] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with tempo-" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "footer", + "text": "20800", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5", + "[27] Andrew Liu, Shiry Ginosar, Tinghui Zhou, Alexei A Efros, and Noah Snavely. Learning to factorize and relight a city. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, pages 544–561. Springer, 2020. 2", + "[28] Abhinitra Meka, Christian Haene, Rohit Pandey, Michael Zollhöfer, Sean Fanello, Graham Fyffe, Adarsh Kowdle, Xueming Yu, Jay Busch, Jason Dourgarian, et al. Deep reflectance fields: high-quality facial reflectance field inference from color gradient illumination. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2", + "[29] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2", + "[30] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 6, 7", + "[31] Frédéric Pighin, Jamie Hecker, Dani Lischinski, Richard Szeliski, and David H Salesin. Synthesizing realistic facial expressions from photographs. In ACM SIGGRAPH 2006 Courses, pages 19-es. 2006. 2", + "[32] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2", + "[33] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 704-720, 2018. 2", + "[34] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 5", + "[35] Erik Reinhard. Parameter estimation for photographic tone reproduction. Journal of graphics tools, 7(1):45-51, 2002. 4", + "[36] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 5", + "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2", + "[38] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2", + "[39] Artem Sevastopolsky, Savva Ignatiev, Gonzalo Ferrer, Evgeny Burnaev, and Victor Lempitsky. Relightable 3d head portraits from a smartphone video. arXiv preprint arXiv:2012.09963, 2020. 2", + "[40] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul E Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):79-1, 2019. 2", + "[41] Tiancheng Sun, Zexiang Xu, Xiuming Zhang, Sean Fanello, Christoph Rhemann, Paul Debevec, Yun-Ta Tsai, Jonathan T Barron, and Ravi Ramamoorthi. Light stage superresolution: Continuous high-frequency relighting. ACM Transactions on Graphics (TOG), 39(6):1-12, 2020. 2", + "[42] Kalyan Sunkavalli, Wojciech Matusik, Hanspeter Pfister, and Szymon Rusinkiewicz. Factored time-lapse video. ACM Transactions on Graphics (Proc. SIGGRAPH), 26(3), Aug. 2007. 2", + "[43] Luan Tran and Xiaoming Liu. Nonlinear 3d face morphable model. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7346-7355, 2018. 2", + "[44] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4", + "[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 2020. 5", + "[46] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 1", + "[47] Tim Weyrich, Wojciech Matusik, Hanspeter Pfister, Bernd Bickel, Craig Donner, Chien Tu, Janet McAndless, Jinho Lee, Addy Ngan, Henrik Wann Jensen, et al. Analysis of human faces using a measurement-based skin reflectance model. ACM Transactions on Graphics (ToG), 25(3):1013-1024, 2006. 4", + "[48] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 2022. 2", + "[49] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 2, 6", + "[50] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20801", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. ACM Transactions on Graphics (TOG), 40(1):1-17, 2021. 6, 7", + "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, pages 7194-7202, 2019. 2, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 218 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20802", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_model.json b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_model.json new file mode 100644 index 0000000000000000000000000000000000000000..991a05ebc83588229a9d3be5d868da0490f53a9c --- /dev/null +++ b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_model.json @@ -0,0 +1,3115 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.159, + 0.892, + 0.182 + ], + "angle": 0, + "content": "SunStage: Portrait Reconstruction and Relighting using the Sun as a Light Stage" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.208, + 0.788, + 0.27 + ], + "angle": 0, + "content": "Yifan Wang1 Aleksander Holynski1 Xiuming Zhang2 Xuaner Zhang2 \n1University of Washington 2Adobe Inc. sunstage.cs.washington.edu" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.285, + 0.245, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.409, + 0.244, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.532, + 0.216, + 0.543 + ], + "angle": 0, + "content": "(a) Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.285, + 0.403, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.409, + 0.401, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.532, + 0.367, + 0.544 + ], + "angle": 0, + "content": "(b) Edit lighting" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.285, + 0.561, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.409, + 0.56, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.532, + 0.532, + 0.543 + ], + "angle": 0, + "content": "(c) Soften shadows" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.285, + 0.718, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.409, + 0.717, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.532, + 0.686, + 0.544 + ], + "angle": 0, + "content": "(d) Swap lighting" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.285, + 0.876, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.409, + 0.874, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.532, + 0.837, + 0.543 + ], + "angle": 0, + "content": "(e) Edit texture" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.548, + 0.894, + 0.604 + ], + "angle": 0, + "content": "Figure 1. Given a selfie video rotating under the sun, SunStage reconstructs geometry, material, camera pose, and lighting information. This recovered information can be used to (a) realistically re-render the input images, (b) modify the lighting conditions by adding / removing lights, (c) soften harsh shadows by changing the size of the reconstructed light sources (d) render the person in an entirely new environment, and (e) edit the albedo or material properties to add freckles, makeup, or stickers that realistically interact with scene lighting." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.617, + 0.314, + 0.632 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.473, + 0.894 + ], + "angle": 0, + "content": "A light stage uses a series of calibrated cameras and lights to capture a subject's facial appearance under varying illumination and viewpoint. This captured information is crucial for facial reconstruction and relighting. Unfortunately, light stages are often inaccessible: they are expensive and require significant technical expertise for construction and operation. In this paper, we present SunStage: a lightweight alternative to a light stage that captures comparable data using only a smartphone camera and the sun. Our method only requires the user to capture a selfie video outdoors, rotating in place, and uses the varying angles between the sun and the face as guidance in joint reconstruction of facial geometry, reflectance, camera pose, and lighting parameters. Despite the in-the-wild un-calibrated setting, our approach is able to reconstruct detailed facial appearance and geometry, enabling compelling effects such as relighting, novel view synthesis, and reflectance editing." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.617, + 0.631, + 0.632 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.764 + ], + "angle": 0, + "content": "A light stage [11] acquires the shape and material properties of a face in high detail using a series of images captured under synchronized cameras and lights. This captured information can be used to synthesize novel images of the subject under arbitrary lighting conditions or from arbitrary viewpoints. This process enables a number of visual effects, such as creating digital replicas of actors that can be used in movies [1] or high-quality postproduction relighting [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In many cases, however, it is often infeasible to get access to a light stage for capturing a particular subject, because light stages are not easy to find: they are expensive and require significant technical expertise (often teams of people) to build and operate. In these cases, hope is not lost — one can turn to methods that are trained on light stage data, with the intention of generalizing to new subjects. These methods do not require the subject to be captured by a light stage but instead use a machine learning" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20792" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.242 + ], + "angle": 0, + "content": "model trained on a collection of previously acquired light stage captures to enable the same applications as a light stage, but from only one or several images of a new subject [6, 25, 30, 38, 40, 50, 52]. Unfortunately, these methods have difficulty faithfully reproducing and editing the appearance of new subjects, as they lack much of the signal necessary to resolve the ambiguities of single-view reconstruction, i.e., a single image of a face can be reasonably explained by different combinations of geometry, illumination, and reflectance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.471, + 0.575 + ], + "angle": 0, + "content": "In this paper, we propose an intermediate solution — one that allows for personalized, high-quality capture of a given subject, but without the need for expensive, calibrated capture equipment. Our method, which we dub SunStage, uses only a handheld smartphone camera and the sun to simulate a minimalist light stage, enabling the reconstruction of individually-tailored geometry and reflectance without specialized equipment. Our capture setup only requires the user to hold the camera at arm's length and rotate in place, allowing the face to be observed under varying angles of incident sunlight, which causes specular highlights to move and shadows to swing across the face. This provides strong signals for the reconstruction of facial geometry and spatially-varying reflectance properties. The reconstructed face and scene parameters estimated by our system can be used to realistically render the subject in new, unseen lighting conditions — even with complex details like self-occluding cast shadows, which are typically missing in purely image-based relighting techniques, i.e., those that do not explicitly model geometry. In addition to relighting, we also show applications in view synthesis, correcting facial perspective distortion, and editing skin reflectance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Our contributions include: (1) a novel capture technique for personalized facial scanning without custom equipment, (2) a system for optimization and disentanglement of scene parameters (geometry, materials, lighting, and camera poses) from an unaligned, handheld video, and (3) multiple portrait editing applications that produce photorealistic results, using as input only a single selfie video." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.223, + 0.713 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Face modeling. Extensive research has been devoted to the modeling of human faces, leading to various 3D morphable models (3DMMs) [2, 3, 5-7, 9, 10, 25, 31-33, 43]. These models are parametric (maybe in the form of neural networks [33]), allowing one to express variations compactly with a vector. They also encode strong priors learned from real scans. The groundbreaking face 3DMM is that of Blanz and Vetter [3] containing models for shape, expression, and appearance (the Phong model). Also influential is the FLAME model [25] that uses vertex-based Linear Blend Skinning (LBS). FLAME is described by a mapping from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "shape, pose, and expression vectors to a list of vertices. We refer the reader to the survey by Egger et al. [13] for different face morphable models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Such parametric face models provide a low-dimensional space for optimization or learning algorithms. DECA [14] uses the FLAME model to estimate detailed facial geometry (and albedo) from single images, by predicting additional displacement maps and adding them to the estimated FLAME models. More recently, NextFace [12] employs the 3DMM geometry and albedo priors to learn an albedo residual that captures more facial details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Without modeling 3D face geometry, researchers have also achieved photorealistic synthesis of portrait images using generative models and large-scale high-quality image datasets [22, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Light stage capture. The light stage as described in Debevec et al., achieves impressive portrait reconstruction and relighting by capturing a series of images of the face under varying illumination [11]. Subsequent work made this process faster, more efficient, and explored different types of illuminants [15, 16, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Given that a light stage is not always accessible, a number of methods have been proposed to achieve similar outputs from a single (or few) input portrait images [18,19,29,30,40,41,48-50]. These method rely on a dataset of light stage captures or synthetic examples as training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Our setup can be thought of as a \"minimalist light stage\" formed by just the sun and a rotating camera, without requiring the high construction and maintenance costs of building a light stage. This parameterization of a sun and skylight model has been shown to be effective in photometric stereo [17,21] and scene factorization [27,42]. In a similar spirit, Calian et al. [8] focus on lighting estimation using faces as \"light probes\". Sengupta et al. propose to circumvent the need for a complicated light stage by recording the facial appearance responses to varying contents displayed on a desk monitor, and then perform portrait relighting [37]. Sevastopolsky et al. also attempt to simplify the capture setup from a light stage to a mobile phone camera with a co-located flash [39]. Unlike our work, which is physically-based, their approaches use neural rendering, and therefore have less direct control over lighting, material, and scene parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.606, + 0.801 + ], + "angle": 0, + "content": "3. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Our method targets accurate reconstruction of scene lighting, subject geometry, and material properties from a handheld video sequence of a person rotating in place under the sun. Given a selfie video, we take a test-time optimization approach that uses the information from all frames of the video to solve for a physical model of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.337, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.088, + 0.593, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.089, + 0.891, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.386, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Overview. Our method jointly reconstructs geometry, skin reflectance, lighting, and camera pose from a selfie video sequence of a person rotating under the sun. Our system begins by extracting supervisory information from the video sequence: facial landmarks, foreground alpha mattes, and camera orientations. These are used to supervise the optimization of a collection of scene parameters (full list in Sec. 5.2) used in a physically-based renderer. The rendered output is an image consisting of diffuse, specular, and ambient light contributions. After optimization, the solved scene parameters can be used for a number of editing applications, shown in Sec. 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.581 + ], + "angle": 0, + "content": "scene: the geometry and material properties of the face, scene lighting, and camera parameters (Fig. 2). This physical model consists of a base face shape parameterized by a low-dimensional deformable model \\( X^b \\), a displacement map \\( \\Delta X \\), a reflectance model with diffuse \\( R^d \\) and specular components \\( R^s \\), scene lighting \\( L_i \\), and a perspective camera \\( C \\). These components are explained in detail in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.471, + 0.69 + ], + "angle": 0, + "content": "After this model has been recovered, we can modify the scene and the subject parameters to re-render images. We show several editing applications in Sec. 7: editing skin reflectance, relighting with arbitrary environment map, improving harsh lighting conditions (by softening shadows and adding fill lights), and adjusting camera parameters to change viewpoint or manipulate perspective effects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.208, + 0.723 + ], + "angle": 0, + "content": "4. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.471, + 0.826 + ], + "angle": 0, + "content": "Given an input video, our system reconstructs the parameters of a physical model: i.e., geometry and reflectance of the subject, the scene lighting parameters, and the camera parameters. In this section, we detail all of these parameters and describe the rendering process that turns these parameters into an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.838, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Geometry. We denote \\( X_{j} \\) as the full mesh of the subject for frame \\( j \\), composed of a per-frame coarse mesh \\( X_{j}^{b}(\\beta, \\theta_{j}, \\psi_{j}) \\) and a global displacement map \\( \\Delta X \\). The coarse mesh \\( X_{j}^{b} \\) is a FLAME deformable face model [25]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.567 + ], + "angle": 0, + "content": "defined by global shape code \\(\\beta\\), per-frame pose code \\(\\theta_{j}\\), and per-frame expression code \\(\\psi_{j}\\). \\(X_{j}^{b}\\) also contains per-vertex UV coordinates, which maintain correspondence across variations in \\(\\theta_{j}\\) and \\(\\psi_{j}\\). As such, we model all our global (per-subject) spatially varying parameters in UV space, and sample values per-fragment when rasterizing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.674 + ], + "angle": 0, + "content": "The displacement map \\(\\Delta X\\) is used to model fine details like wrinkles that cannot be represented by \\(X_{j}^{b}\\). We displace the coarse geometry by \\(\\Delta X\\) at rasterization time, by sampling a displacement value per-fragment and displacing each fragment along the surface normal \\(N_{j}\\) of the coarse mesh \\(X_{j}^{b}\\). After displacement, the updated fragment positions are used to compute a new surface normal \\(N_{j}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.782 + ], + "angle": 0, + "content": "\\(X_{j}^{b}\\) is optimized per-frame, since it accounts for subtle (and unavoidable) variations in expression and head pose during the capture, which are modeled by \\(\\theta_{j}\\) and \\(\\psi_{j}\\). \\(\\Delta X\\), on the other hand, is optimized in UV-space (i.e., globally per-subject), since the deformations it encodes are invariant to the changes in expression or pose. Formally, the final geometry \\(X_{j}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.795, + 0.892, + 0.814 + ], + "angle": 0, + "content": "\\[\nX _ {j} = X _ {j} ^ {b} \\left(\\beta , \\theta_ {j}, \\psi_ {j}\\right) + \\Delta X \\odot N _ {j} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.827, + 0.731, + 0.843 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Reflectance. We model the skin reflectance, denoted as \\( R(x,\\omega_i,\\omega_o) \\in \\mathbb{R}^3 \\), where \\( x \\) is a 3D point on the face geometry \\( X \\), \\( \\omega_i \\) is the incoming light direction, and \\( \\omega_o \\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "outgoing direction, using a diffuse and a specular component: \\( R = R^{d} + R^{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.121, + 0.47, + 0.183 + ], + "angle": 0, + "content": "The diffuse component \\(R^d (x,\\omega_i)\\in \\mathbb{R}^3\\) is a Lambertian reflectance model consisting of an albedo map, \\(a\\), which we optimize as a per-subject UV-space image. For the skin's specular component, we use the Blinn-Phong model [4]." + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.201, + 0.47, + 0.23 + ], + "angle": 0, + "content": "\\[\nR ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right) = k _ {s} \\frac {s + 2}{2 \\pi} \\left(h \\left(\\omega_ {i}, \\omega_ {o}\\right) \\cdot n (x)\\right) ^ {s} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.469, + 0.342 + ], + "angle": 0, + "content": "where \\( h(\\omega_i, \\omega_o) = \\text{normalize}(\\omega_i + \\omega_o) \\) is the half vector, \\( k_s \\) is the specular intensity, \\( s \\) is the specular exponent, and \\( (s + 2) / (2\\pi) \\) is the normalization term for the reflection lobe to integrate to 1. Following [47], we segment the UV-space map into 10 segmented specular reflectance clusters. We then optimize for a spatially-varying pair of values \\( (s, k_s) \\) per-cluster, enabling varying shininess across the face." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.434 + ], + "angle": 0, + "content": "While Blinn-Phong does not model many complex effects such as subsurface scattering, our experiments with other models for facial reflectance, such as microfacet models [44], show no significant quality improvements, and often introduce unstable training. More analysis is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.598 + ], + "angle": 0, + "content": "Lighting. We use a sun-sky model to represent lighting as the sum of an \"ambient\" environment map and the sun: \\( L_{i}(x,\\omega_{i}) = L_{i}^{\\mathrm{amb}}(\\omega_{i}) + L_{i}^{\\mathrm{sun}}(\\omega_{i}) \\). Note neither \\( L_{i}^{\\mathrm{amb}}(\\omega_{i}) \\) nor \\( L_{i}^{\\mathrm{sun}}(\\omega_{i}) \\) depends on the 3D point \\( x \\), since we model both as directional lights. Optimization-wise, our lighting parameters consist of a \\( 16 \\times 32 \\times 3 \\) environment map for ambient lighting, the sun direction \\( p^{\\mathrm{sun}} \\in S^{3} \\), and the scalar sun intensity \\( k^{\\mathrm{sun}} \\). We fix the sun color to white [1, 1, 1] in our lighting model to resolve the albedo-illumination ambiguity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.195, + 0.621 + ], + "angle": 0, + "content": "4.1. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.658 + ], + "angle": 0, + "content": "We calculate the outgoing radiance \\( L_{o} \\) at 3D location \\( x \\) as viewed from viewing direction \\( \\omega_{o} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.681, + 0.469, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {o} (x, \\omega_ {o}) \\\\ = \\int_ {\\Omega} V \\left(x, \\omega_ {i}\\right) L _ {i} \\left(x, \\omega_ {i}\\right) \\odot R \\left(x, \\omega_ {i}, \\omega_ {o}\\right) \\left(\\omega_ {i} \\cdot n (x)\\right) d \\omega_ {i} (3) \\\\ = \\sum_ {\\omega_ {i}} V (x, \\omega_ {i}) \\left(L _ {i} ^ {\\mathrm {a m b}} (\\omega_ {i}) \\odot R ^ {d} (x, \\omega_ {i}) \\right. (4) \\\\ + L _ {i} ^ {\\text {s u n}} \\left(\\omega_ {i}\\right) \\odot R ^ {d} \\left(x, \\omega_ {i}\\right) + L _ {i} ^ {\\text {a m b}} \\left(\\omega_ {i}\\right) \\odot R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right) \\\\ \\left. + L _ {i} ^ {\\operatorname {s u n}} \\left(\\omega_ {i}\\right) \\odot R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right)\\right) \\left(\\omega_ {i} \\cdot n (x)\\right) \\triangle \\omega_ {i} (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(V(x,\\omega_{i})\\) is the light visibility at \\(x\\) from \\(\\omega_{i}\\), and \\(L_{i}(x,\\omega_{i})\\) is the incoming radiance reaching \\(x\\) from \\(\\omega_{i}\\). We ignore the specular reflection caused by the ambient lightning, i.e., \\(L_{i}^{\\mathrm{amb}}(\\omega_{i})\\odot R^{s}(x,\\omega_{i},\\omega_{o})\\), since it is much weaker" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "than the specular reflection of the sun. In the next subsections, we will group the terms into a diffuse contribution \\( L_{o}^{d} \\) and a specular contribution \\( L_{o}^{s} \\): \\( L_{o} = L_{o}^{d} + L_{o}^{s} \\). For the final rendered color value, we apply the Reinhard operator [35] and a gamma correction of \\( \\gamma = 2.2 \\) to \\( L_{o} \\) to convert from linear to sRGB space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.892, + 0.226 + ], + "angle": 0, + "content": "Diffuse contribution. The diffuse contribution \\( L_{o}^{d} \\) is then given by only the diffuse terms of Equation 5:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.237, + 0.891, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {o} ^ {d} (x) = \\sum_ {\\omega_ {i}} L _ {i} ^ {\\mathrm {a m b}} (\\omega_ {i}) \\odot \\frac {a (x)}{\\pi} (\\omega_ {i} \\cdot n (x)) \\Delta \\omega_ {i} \\\\ + V (x, p ^ {\\text {s u n}}) k ^ {\\text {s u n}} [ 1, 1, 1 ] \\odot \\frac {a (x)}{\\pi} \\left(p ^ {\\text {s u n}} \\cdot n (x)\\right) \\Delta p ^ {\\text {s u n}} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.439 + ], + "angle": 0, + "content": "where \\(a(x)\\) is the albedo at point \\(x\\), \\(k^{\\mathrm{sun}}\\) is the (optimized) sun intensity, and \\(p^{\\mathrm{sun}}\\) is the (optimized) sun direction. The sun is modeled as a directional light source, so the second summation can be simplified to a single term, only in the direction of \\(p^{\\mathrm{sun}}\\). We additionally optimize for a high-dynamic-range (HDR) environment map \\(E \\in \\mathbb{R}^{16 \\times 32 \\times 3}\\), from which values of \\(L_{i}^{\\mathrm{amb}}\\) are sampled." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Specular contribution. The specular contribution \\( L_{o}^{s} \\) at each pixel is given by only the specular term due to the sun in Equation 5 (recall that we ignore the specular ambient term due to its weak contribution):" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.515, + 0.892, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {o} ^ {s} (x, \\omega_ {o}) = V (x, p ^ {\\mathrm {s u n}}) k ^ {\\mathrm {s u n}} [ 1, 1, 1 ] k _ {s} \\frac {s + 2}{2 \\pi} \\\\ \\left(h \\left(p ^ {\\text {s u n}}, \\omega_ {o}\\right) \\cdot n (x)\\right) ^ {s} \\left(p ^ {\\text {s u n}} \\cdot n (x)\\right) \\Delta p ^ {\\text {s u n}} \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.599 + ], + "angle": 0, + "content": "where we have substituted Equation 2 and reduced the summation to just one term at \\( p^{\\mathrm{sun}} \\) (since \\( L_{i}^{\\mathrm{sun}} \\) is 0 elsewhere)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Shadow map. In order to generate a map of self-occluded shadows, we perform two passes of rasterization: first, we render a \\( z \\)-buffer from a virtual orthographic camera aligned with the sun direction, \\( p^{\\mathrm{sun}} \\), and then, when rasterizing a given camera viewpoint, compare all fragment positions \\( d_{\\mathrm{hit}} \\) to the light's \\( z \\)-buffer \\( d_{\\mathrm{shadow}} \\). To avoid precision issues and ensure smooth gradients for back-propagation, we implement a soft comparison as follows in generating shadow/visibility maps:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.759, + 0.892, + 0.776 + ], + "angle": 0, + "content": "\\[\nV \\left(x, \\omega_ {i}\\right) = 1 - \\text {s i g m o i d} \\left(k \\left(d _ {\\text {h i t}} - d _ {\\text {s h a d o w}} \\times b\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.892, + 0.816 + ], + "angle": 0, + "content": "where \\( k \\) is the falloff slope, and \\( b \\) the tolerance. We use \\( k = 800 \\) and \\( b = 1.0015 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.637, + 0.847 + ], + "angle": 0, + "content": "5. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The described physical model contains a large number of parameters to be optimized, controlling scene elements like lighting, geometry, pose, and texture. Unfortunately," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20795" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "naively optimizing all these parameters from scratch does not result in an optimal solution, since the final observed appearance of the face can often be explained variously through changes to geometry, material properties, lighting or camera parameters, making optimization severely under-constrained and ambiguous. Therefore, we adopt a two-stage optimization approach, through which parameters are gradually enabled. In this section, we describe this process and the relevant losses that guide optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.24, + 0.251, + 0.258 + ], + "angle": 0, + "content": "5.1. Coarse alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Our system begins by using an off-the-shelf network (DECA [14]) to generate, for each input image, a set of shape parameters \\(\\beta\\), pose parameters \\(\\theta_{j}\\), and expression parameters \\(\\psi_{j}\\) of a FLAME face model [25], as well as the relative pose parameters of the virtual camera observing the 3D face. Unfortunately, as with many other single-image facial geometry estimators, DECA assumes an orthographic projection model and therefore cannot accurately recover geometry for our selfie capture sequences, which contain heavy perspective effects (Figure 6). Without a good initialization for geometry, optimization of lighting and material properties seldom converges to an optimal solution due to the heavily ambiguous nature of our optimization problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.809 + ], + "angle": 0, + "content": "To circumvent this issue, we employ a first stage of optimization where we only optimize for the parameters of a perspective camera (with a known focal length, extracted from input metadata) and the face geometry parameters \\((\\beta, \\theta_j, \\psi_j)\\). As initialization for this optimization process, we use the predicted DECA values for each frame's pose \\(\\theta_j\\) and expression \\(\\psi_j\\), but set all frames to the average predicted shape \\(\\beta_{\\mathrm{avg}} = \\frac{1}{N} \\sum_j \\beta_j\\), since the identity remains constant across all frames. To convert DECA's orthographic camera to a perspective camera, we additionally optimize for an unknown scale \\(S\\) and translation \\(T_j\\), which are initialized to empirically chosen values \\(S = 2.6e4\\), \\(T_j = (0, 0, 1.5e5)\\). During optimization, the face shape \\(\\beta\\) and scale \\(S\\) are shared across all frames, while camera pose \\(T_j\\), expression \\(\\psi_j\\), and pose \\(\\theta_j\\) are optimized per-frame. Note that DECA controls the relative orientation of the camera and the face by varying the pose code \\(\\theta\\) instead of the camera rotation. We adopt this formulation and keep the camera orientation fixed relative to the face. The global orientation of the camera at each frame (and therefore the face) is extracted from the capture video, either through a structure-from-motion system or IMU measurements commonly available on a smartphone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We use two losses to guide this optimization: a mask loss \\( L_{\\mathrm{mask}} \\) and a landmark loss \\( L_{\\mathrm{lmk}} \\). The FLAME model includes 3D facial landmark points, corresponding to the standard 68-point facial landmarks set [36] used in facial tracking. Our landmark loss minimizes the L1 distance between the 2D projection of these 3D landmarks (into the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "input camera viewpoint) and 2D landmarks estimated from the input frame by a 2D landmark detector HRNets [45]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.257 + ], + "angle": 0, + "content": "The facial landmarks provide a stong constraint on facial feature alignment, but are sparse, and therefore cannot constrain the overall shape or boundary of the mesh. To supplement it, we include a silhouette loss \\( L_{\\mathrm{mask}} \\), which penalizes the L2 difference between the rasterized mask of the mesh \\( I_{\\mathrm{sil}} \\) and the semantic segmentation mask \\( I_{\\mathrm{mask}} \\) of the input image, using an off-the-shelf semantic segmentation network [26] trained to segment humans in portrait photographs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.894, + 0.303 + ], + "angle": 0, + "content": "The final pose loss is then: \\( L_{\\mathrm{pose}} = L_{\\mathrm{mask}} + L_{\\mathrm{lmk}} \\), optimized using an ADAM optimizer [24]. See supplemental for optimization parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.311, + 0.733, + 0.327 + ], + "angle": 0, + "content": "5.2. Photometric optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Once the 3D model and camera parameters are approximately aligned, we proceed to the second stage of optimization, in which we optimize the precise facial geometry, lighting, and reflectance properties. All the parameters optimized in the first stage (Section 5.1) remain as free variables. In total, the parameters optimized during this stage include: Lighting parameters: (1) \\( p^{\\mathrm{sun}} \\), the global sun direction, (2) \\( E \\), the global environment map, (3) \\( k^{\\mathrm{sun}} \\), the global sun intensity, Facial geometry parameters: (4) \\( \\beta \\), the global FLAME shape code, (5) \\( \\psi_j \\), the per-frame expression code, (6) \\( \\theta_j \\), the per-frame pose code, (7) \\( \\Delta X \\), the global deformation map, Material properties: (8) \\( k_s \\), the global, spatially-varying specular intensity, (9) \\( s \\), the global, spatially-varying specular roughness, (10) \\( a \\), the global, spatially-varying surface albedo, Camera pose parameters: (11) \\( T_j \\), the per-frame perspective camera translation, and (12) \\( S \\), the global scene scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.681 + ], + "angle": 0, + "content": "During optimization, we randomly select a frame \\( j \\), render the face using a differentiable rasterizer [34] and the equations described in Section 4 to get the rendered image \\( \\hat{I} \\). In addition to the previously defined landmark and mask losses, we include L2 and VGG [20] photometric losses, comparing the original and reconstructed images:" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.689, + 0.891, + 0.707 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p h o t o}} = \\left| \\left| \\hat {I} _ {j} \\cdot I _ {\\text {s i l}} - I _ {j} \\cdot I _ {\\text {m a s k}} \\right| \\right| _ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.803 + ], + "angle": 0, + "content": "We also include an L2 regularization \\(L_{E}\\) and L2-smoothness regularization \\(L_{Es}\\) on the reconstructed environment map, to encourage the majority of the lighting to be explained by direct sunlight and to aid in disentangle-ment of the sun and ambient lighting. The total optimized loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.813, + 0.892, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\lambda_ {\\text {m a s k}} L _ {\\text {m a s k}} + \\lambda_ {\\text {l m k}} L _ {\\text {l m k}} + \\lambda_ {E} L _ {E} + \\lambda_ {E \\mathrm {s}} L _ {E \\mathrm {s}} \\\\ + \\lambda_ {\\mathrm {V G G}} L _ {\\mathrm {V G G}} + \\lambda_ {\\text {p h o t o}} L _ {\\text {p h o t o}} \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "with \\(\\lambda_{\\mathrm{mask}}, \\lambda_{\\mathrm{lmk}} = 0.05, \\lambda_{\\mathrm{VGG}} = 0.005, \\lambda_{E} = 0.01, \\lambda_{Es}, \\lambda_{photo} = 1\\). Additional optimization details are provided in the supplemental materials." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.182, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.244, + 0.17, + 0.257 + ], + "angle": 0, + "content": "(a) DECA [14]" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.283, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.244, + 0.268, + 0.257 + ], + "angle": 0, + "content": "(b) DPR [52]" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.385, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.244, + 0.373, + 0.257 + ], + "angle": 0, + "content": "(c) GCFR [18]" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.089, + 0.485, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.244, + 0.466, + 0.257 + ], + "angle": 0, + "content": "(d) TR [30]" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.089, + 0.587, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.244, + 0.584, + 0.257 + ], + "angle": 0, + "content": "(e) NextFace [12]" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.089, + 0.687, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.244, + 0.671, + 0.257 + ], + "angle": 0, + "content": "(f) NLT [51]" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.089, + 0.788, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.244, + 0.762, + 0.257 + ], + "angle": 0, + "content": "(g) Ours" + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.089, + 0.889, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.244, + 0.858, + 0.256 + ], + "angle": 0, + "content": "(h) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Qualitative: Relighting. A comparison of our method at rendering a new (unseen) lighting environment (h). Our method is able to realistically synthesize the novel lighting condition, including cast shadows and specularity, and nearly matches the (unseen) target reference image. See supplement for additional details on experimental setup and analysis of results." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.338, + 0.217, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.44, + 0.217, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.544, + 0.195, + 0.555 + ], + "angle": 0, + "content": "(a) Nearest view" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.338, + 0.351, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.44, + 0.351, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.544, + 0.326, + 0.556 + ], + "angle": 0, + "content": "(b) DECA [14]" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.338, + 0.484, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.44, + 0.484, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.544, + 0.466, + 0.556 + ], + "angle": 0, + "content": "(c) NextFace [12]" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.338, + 0.618, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.44, + 0.618, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.544, + 0.587, + 0.556 + ], + "angle": 0, + "content": "(d) NLT [51]" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.338, + 0.752, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.44, + 0.751, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.544, + 0.708, + 0.555 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.338, + 0.885, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.44, + 0.885, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.544, + 0.837, + 0.555 + ], + "angle": 0, + "content": "(f) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.568, + 0.89, + 0.597 + ], + "angle": 0, + "content": "Figure 4. Qualitative: view synthesis. A comparison of our method at the task of generating an image from an unseen viewpoint (f), having only seen a limited collection of input viewpoints. See supplement for more details on experimental setup and analysis of results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.194, + 0.649 + ], + "angle": 0, + "content": "6. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.706 + ], + "angle": 0, + "content": "In this section, we detail quantitative and qualitative experiments comparing our approach with state-of-the-art methods and ablated variants of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.468, + 0.868 + ], + "angle": 0, + "content": "Baseline comparisons. We evaluate our method on the tasks of novel-view synthesis and relighting. For novel-view synthesis, we compare our method with DECA [14], Neural Light Transport (NLT) [51], and NextFace [12]. For relighting, we compare with DECA, NLT, NextFace, GCFR [18], image-based methods Deep Single Image Portrait Relighting (DPR) [52], Total Relighting (TR) [30] and NVPR [49]. Additional comparisons and details on the experimental setups are provided in the supplemental materials." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We present qualitative comparisons for relighting in Figure 3 and novel view synthesis in Figure 4. Quantitative" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.877 + ], + "angle": 0, + "content": "comparisons on these images are provided in Table 1. These testing images consist of (1) a multi-view capture of the face, in which the subject remains still and the camera is moved to novel viewpoints in the same environment as the original capture, and (2) front-facing sequences in novel environment lighting and unseen sun positions. All testing images are not seen during training of our method, NLT or NextFace. The results shown in Figures 3 and 4 as well as Table 1 clearly demonstrate that our method outperforms all the baselines at both relighting and view synthesis. Single-image methods (DECA, GCFR, DPR, TR) can generalize to other subjects, but fail to recover more faithful and physically accurate facial details. Comparison with multi-image methods (NLT, NextFace) demonstrates that SunStage is a better reconstruction system. Additional analysis of the comparisons is provided in the supplemental materials." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.197, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.089, + 0.312, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.089, + 0.425, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.089, + 0.54, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.655, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.089, + 0.771, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.089, + 0.888, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.182, + 0.197, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.182, + 0.311, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.182, + 0.425, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.182, + 0.54, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.182, + 0.655, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.182, + 0.771, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.182, + 0.888, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.274, + 0.197, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.366, + 0.166, + 0.378 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.275, + 0.311, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.366, + 0.279, + 0.378 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.275, + 0.425, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.366, + 0.398, + 0.38 + ], + "angle": 0, + "content": "Specular" + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.275, + 0.54, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.366, + 0.511, + 0.378 + ], + "angle": 0, + "content": "Ambient" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.275, + 0.655, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.366, + 0.623, + 0.378 + ], + "angle": 0, + "content": "Diffuse" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.275, + 0.771, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.366, + 0.762, + 0.378 + ], + "angle": 0, + "content": "Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.275, + 0.888, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.366, + 0.874, + 0.378 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Figure 5. Decomposition. We show all the components which comprise our final rendered image to demonstrate that our method not only closely recreates the ground truth image (reproducing realistic highlights and shadows), but also performs a meaningful decomposition of lighting components and facial geometry. Note that our reconstructed surface normals include high frequency details specific to each subject, like wrinkles and birthmarks, which are used in computation of the shadows and specular reflections." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.46, + 0.472, + 0.643 + ], + "angle": 0, + "content": "
RelightingNovel view synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DECA [14]16.410.690.2516.640.660.29
GCFR [18]16.970.700.20---
DPR [52]19.030.720.19---
NLT [51]20.150.750.1822.270.790.15
Total Relighting [30]20.240.790.16---
NextFace [12]22.980.760.1522.550.750.15
Ours23.640.830.1025.280.840.09
Ours w/o coarse17.830.660.2319.650.700.17
Ours w/o SV ks, s21.310.770.1321.940.770.12
Ours w/o Lmask, Llmk16.460.610.3018.830.680.20
Ours w/o Lmask20.130.750.1520.540.740.15
Ours w/o opt. (β, θi, φi)18.670.690.1918.280.660.19
Ours w/o soft shadow21.460.770.1322.050.770.12
Ours w/o ΔX21.160.750.1521.800.750.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.694 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison. Comparison of our method on the tasks of novel view synthesis and relighting. See Section 6 for a description of the ablated variants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.47, + 0.873 + ], + "angle": 0, + "content": "Disentanglement. In Figure 5, we demonstrate how Sun-Stage decomposes the appearance of a portrait photograph into different components: specular, diffuse, and ambient. We also visualize the surface normal, albedo, and other intermediate representations to show that our method is able to effectively recover a physically plausible reconstruction of the real world and disentangle the different components that contribute to the final appearance. We further validate the quality of the reconstructed geometry and materials in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Ablation studies. In addition to our comparisons with" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.46, + 0.6, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.536, + 0.594, + 0.548 + ], + "angle": 0, + "content": "(a) Input image" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.46, + 0.696, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.536, + 0.69, + 0.548 + ], + "angle": 0, + "content": "(b) DECA [14]" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.46, + 0.794, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.536, + 0.773, + 0.548 + ], + "angle": 0, + "content": "(c) Stage 1" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.46, + 0.891, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.536, + 0.866, + 0.548 + ], + "angle": 0, + "content": "(d) Final" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.561, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Figure 6. Perspective. DECA's assumption of an orthographic camera is broken by the strong perspective effects in selfies, causing poor alignment (b) with input images (a). Our first stage of optimization (c) (Sec. 5.1) improves alignment by solving for the parameters of a perspective camera and refined shape parameters. In the second stage we additionally optimize for a displacement map \\(\\Delta X\\) to produce our final shape with finer geometric details like wrinkles (d). Red line added to highlight alignment with (a)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "the state-of-the-art baselines, we also compare with ablated variants of our own method. In particular, we include seven such experiments in Table 1: our method (1) without the initial first stage of coarse geometric alignment, i.e., directly optimizing both geometric and photometric parameters from the start, (2) without the spatially varying specular parameters, instead using a single global scalar \\( s \\) and \\( k_{s} \\), (3) without the geometric alignment losses \\( L_{\\mathrm{mask}} \\) and \\( L_{\\mathrm{lmk}} \\), (4) without just \\( L_{\\mathrm{mask}} \\), (5) without shape optimization, i.e., keeping the initial shape code predicted by DECA, (6) without soft shadow computation, i.e., using a hard z-buffer comparison to compute a shadow map instead of our soft" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.177, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.238, + 0.153, + 0.25 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.273, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.238, + 0.27, + 0.249 + ], + "angle": 0, + "content": "(b) Soft shadows" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.089, + 0.37, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.238, + 0.368, + 0.25 + ], + "angle": 0, + "content": "(c) (b) + fill light" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.468, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.238, + 0.461, + 0.25 + ], + "angle": 0, + "content": "(d) Swap scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.262, + 0.47, + 0.332 + ], + "angle": 0, + "content": "Figure 7. Adjusting lighting parameters. We can adjust the recovered scene parameters to improve the lighting conditions in an input image (a) by softening the harsh shadows cast by the nose (b), adding a fill light to brighten the shaded region (b), or replacing the environment altogether (d)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.469, + 0.403 + ], + "angle": 0, + "content": "comparison operator in Equation 8, and (7) without the displacement map \\(\\Delta X\\). Visual results for each of these variants are provided in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.415, + 0.208, + 0.433 + ], + "angle": 0, + "content": "7. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Relighting. We demonstrate two types of relighting applications: (1) lighting modification and (2) lighting replacement. Practical lighting modification is common in portrait photography when the lighting conditions are not ideal, e.g., when direct sunlight casts undesirable harsh shadows with high contrast. A common practice is to make the light source larger and more diffuse by using a scrim or bounce card. In Fig. 7b, we show that by virtually increasing the size and spreading the energy of our reconstructed lighting source (i.e., the sun), we are able to soften the shadows and re-render a more visually pleasing face. Another approach to reducing the effects of harsh shadows is adding local fill lights, which reduces the contrast between the lit and shaded regions (Fig. 7c). Alternatively, fill lights can also be used for artistic purposes, to create dramatic lighting effects (Fig. 1b). Finally, replacing the scene lighting with that of a novel environment (Fig. 7d) is a necessary step in realistically inserting a captured subject into a virtual scene, which is useful for visual effects and VR applications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.469, + 0.858 + ], + "angle": 0, + "content": "View Synthesis. In Figure 8, we show that our reconstructed 3D model of the face can be used to synthesize new views by manipulating the viewpoint of the camera. We can also change other camera parameters, such as the focal length, to reduce the perspective effects on the face, which is often desirable for selfie images that contain significant facial distortion due to perspective." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Skin Reflectance Editing. We are also able to edit the reflectance components of the subject. As shown in Figure 1e," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.599, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.238, + 0.591, + 0.25 + ], + "angle": 0, + "content": "(a) Short focal" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.089, + 0.696, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.238, + 0.688, + 0.25 + ], + "angle": 0, + "content": "(b) Long focal" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.793, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.238, + 0.782, + 0.25 + ], + "angle": 0, + "content": "(c) Novel view" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.089, + 0.89, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.238, + 0.888, + 0.25 + ], + "angle": 0, + "content": "(d) Novel view 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.262, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 8. Changing camera parameters. We can change the recovered camera parameters to render novel views (c,d) or change the focal length (a,b)." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.312, + 0.6, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.389, + 0.593, + 0.4 + ], + "angle": 0, + "content": "(a) Less shinier" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.312, + 0.697, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.389, + 0.693, + 0.4 + ], + "angle": 0, + "content": "(b) More shinier" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.312, + 0.793, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.389, + 0.787, + 0.399 + ], + "angle": 0, + "content": "(c) Less shinier" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.312, + 0.89, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.389, + 0.886, + 0.399 + ], + "angle": 0, + "content": "(d) More shinier" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.413, + 0.892, + 0.454 + ], + "angle": 0, + "content": "Figure 9. Adjusting the specularity. We can change the the specular properties of the face, making the face less shinier (a, c) or more shinier (b, d)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.547 + ], + "angle": 0, + "content": "we can adjust the optimized albedo to add freckles, stickers, or other textures that realistically interact with reflections, shadows, and other elements of scene lighting, or we can adjust the specular properties of the face, making the face more or less shinier, as shown in Figure 9." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.755, + 0.577 + ], + "angle": 0, + "content": "8. Conclusion and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.753 + ], + "angle": 0, + "content": "In this paper, we propose SunStage, a lightweight and practical facial capture, rendering, and editing system that can serve as a minimalist light stage. With a video of an individual rotating in-place under the sun, our system reconstructs a physical model of the subject and the scene lighting, which enables us to relight the subject with realistic reflections and cast shadows. Our system allows arbitrary lighting and reflectance control in the reconstructed physical space, which can be rendered to produce photo-realistic results. We demonstrate several applications such as editing skin reflectance, relighting, and view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations. Our system inherits the limitations of morphable face models and is unable to model hair, teeth, or clothing geometry, beyond slight deformations. Additionally, certain regions which are seen under constant shadow or specular reflection (and therefore have no cues on reflectance or albedo) are sometimes unable to be decomposed accurately into separate reflectance and lighting components. Visualization and further discussions of the system's limitations are provided in the supplemental material." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20799" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Oleg Alexander, Mike Rogers, William Lambeth, Jen-Yuan Chiang, Wan-Chun Ma, Chuan-Chang Wang, and Paul Debevec. The digital emily project: Achieving a photorealistic digital actor. IEEE Computer Graphics and Applications, 30(4):20-31, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[3] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[4] James F Blinn. Models of light reflection for computer synthesized pictures. In Proceedings of the 4th annual conference on Computer graphics and interactive techniques, pages 192-198, 1977. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[5] Timo Bolkart and Stefanie Wuhrer. A groupwise multilinear correspondence optimization for 3d faces. In Proceedings of the IEEE international conference on computer vision, pages 3604-3612, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[7] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In European Conference on Computer Vision, pages 297-312. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[8] Dan A Calian, Jean-François Lalonde, Paulo Gotardo, Tomas Simon, Iain Matthews, and Kenny Mitchell. From faces to outdoor light probes. In Computer Graphics Forum, volume 37, pages 51-61. Wiley Online Library, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.676, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[10] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3085-3093, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[12] Abdallah Dib, Cedric Thebault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[13] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[15] Graham Fyffe and Paul Debevec. Single-shot reflectance measurement from polarized color gradient illumination. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[16] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. In Proceedings of the 2011 SIGGRAPH Asia Conference, pages 1-10, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[17] Yannick Hold-Geoffroy, Jinsong Zhang, Paulo F U Gotardo, and Jean-François Lalonde. \\(x\\)-hour outdoor photometric stereo. In International Conference on 3D Vision, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[18] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[21] Jiyoung Jung, Joon-Young Lee, and In So Kweon. One-day outdoor photometric stereo via skylight estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4521–4529, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with tempo-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20800" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "ral guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[27] Andrew Liu, Shiry Ginosar, Tinghui Zhou, Alexei A Efros, and Noah Snavely. Learning to factorize and relight a city. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, pages 544–561. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[28] Abhinitra Meka, Christian Haene, Rohit Pandey, Michael Zollhöfer, Sean Fanello, Graham Fyffe, Adarsh Kowdle, Xueming Yu, Jay Busch, Jason Dourgarian, et al. Deep reflectance fields: high-quality facial reflectance field inference from color gradient illumination. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[29] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[30] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[31] Frédéric Pighin, Jamie Hecker, Dani Lischinski, Richard Szeliski, and David H Salesin. Synthesizing realistic facial expressions from photographs. In ACM SIGGRAPH 2006 Courses, pages 19-es. 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[32] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.643 + ], + "angle": 0, + "content": "[33] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 704-720, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[34] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[35] Erik Reinhard. Parameter estimation for photographic tone reproduction. Journal of graphics tools, 7(1):45-51, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[36] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[38] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[39] Artem Sevastopolsky, Savva Ignatiev, Gonzalo Ferrer, Evgeny Burnaev, and Victor Lempitsky. Relightable 3d head portraits from a smartphone video. arXiv preprint arXiv:2012.09963, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[40] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul E Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):79-1, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[41] Tiancheng Sun, Zexiang Xu, Xiuming Zhang, Sean Fanello, Christoph Rhemann, Paul Debevec, Yun-Ta Tsai, Jonathan T Barron, and Ravi Ramamoorthi. Light stage superresolution: Continuous high-frequency relighting. ACM Transactions on Graphics (TOG), 39(6):1-12, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[42] Kalyan Sunkavalli, Wojciech Matusik, Hanspeter Pfister, and Szymon Rusinkiewicz. Factored time-lapse video. ACM Transactions on Graphics (Proc. SIGGRAPH), 26(3), Aug. 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.894, + 0.42 + ], + "angle": 0, + "content": "[43] Luan Tran and Xiaoming Liu. Nonlinear 3d face morphable model. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7346-7355, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.894, + 0.489 + ], + "angle": 0, + "content": "[44] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.894, + 0.561 + ], + "angle": 0, + "content": "[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.894, + 0.631 + ], + "angle": 0, + "content": "[46] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[47] Tim Weyrich, Wojciech Matusik, Hanspeter Pfister, Bernd Bickel, Craig Donner, Chien Tu, Janet McAndless, Jinho Lee, Addy Ngan, Henrik Wann Jensen, et al. Analysis of human faces using a measurement-based skin reflectance model. ACM Transactions on Graphics (ToG), 25(3):1013-1024, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[48] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[49] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[50] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20801" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[51] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. ACM Transactions on Graphics (TOG), 40(1):1-17, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, pages 7194-7202, 2019. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20802" + } + ] +] \ No newline at end of file diff --git a/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_origin.pdf b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3bc8ed82f0f4da18dc86d40602ef8aa47ed1967a --- /dev/null +++ b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/0c3e205e-2bfe-4ba0-8856-0ef79f6d0027_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3b60f7bc6f87d24a5019b3e930347638d2db2884b0c0681d438f22ad5d66fb0 +size 3968552 diff --git a/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/full.md b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/full.md new file mode 100644 index 0000000000000000000000000000000000000000..04e6bf16c6304e530ca2c7f15ded0d88bf988110 --- /dev/null +++ b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/full.md @@ -0,0 +1,442 @@ +# SunStage: Portrait Reconstruction and Relighting using the Sun as a Light Stage + +Yifan Wang1 Aleksander Holynski1 Xiuming Zhang2 Xuaner Zhang2 +1University of Washington 2Adobe Inc. sunstage.cs.washington.edu + +![](images/72a12d15f29c5906340f76c801ccc4c15a96e5877a8f469d131cc56faa4009a3.jpg) + +![](images/b1647eeccb2d79f41272998fc706007083d921392df0450417e5c22ca93e0c29.jpg) +(a) Reconstruction + +![](images/48fcdcab75a29a0909e9910f52cc24f3b29967214d99a9a5c9b46f441cd61f58.jpg) + +![](images/c070d1dc26fcb9105ab4de5ebce71ebaa7d1d35fa9e87583a3f0d6023ba520bf.jpg) +(b) Edit lighting + +![](images/fe54843eb01bf19c9fcc6df87800725b06c232909981659bb78d65e4737af92a.jpg) + +![](images/d94d69297e90d8ffe7512bbcbe3e7e0cdc582872dfacb0af8f04be6a3798a7f6.jpg) +(c) Soften shadows +Figure 1. Given a selfie video rotating under the sun, SunStage reconstructs geometry, material, camera pose, and lighting information. This recovered information can be used to (a) realistically re-render the input images, (b) modify the lighting conditions by adding / removing lights, (c) soften harsh shadows by changing the size of the reconstructed light sources (d) render the person in an entirely new environment, and (e) edit the albedo or material properties to add freckles, makeup, or stickers that realistically interact with scene lighting. + +![](images/dea3acc30753508cf83b136bde64244904c67a4a3436c4b4eb6638c914885923.jpg) + +![](images/251f48050b9b0dfa4cb01dcfb64672ba2facd76b6241852f1822a5184c846272.jpg) +(d) Swap lighting + +![](images/277e23ea4b2eba9a410a8677de4158bc29ea2e68c64353837eefcc92899a625a.jpg) + +![](images/6aefb98a22232c890df094f89cd1b64eeb7491df40ec494e3281c082c1137599.jpg) +(e) Edit texture + +# Abstract + +A light stage uses a series of calibrated cameras and lights to capture a subject's facial appearance under varying illumination and viewpoint. This captured information is crucial for facial reconstruction and relighting. Unfortunately, light stages are often inaccessible: they are expensive and require significant technical expertise for construction and operation. In this paper, we present SunStage: a lightweight alternative to a light stage that captures comparable data using only a smartphone camera and the sun. Our method only requires the user to capture a selfie video outdoors, rotating in place, and uses the varying angles between the sun and the face as guidance in joint reconstruction of facial geometry, reflectance, camera pose, and lighting parameters. Despite the in-the-wild un-calibrated setting, our approach is able to reconstruct detailed facial appearance and geometry, enabling compelling effects such as relighting, novel view synthesis, and reflectance editing. + +# 1. Introduction + +A light stage [11] acquires the shape and material properties of a face in high detail using a series of images captured under synchronized cameras and lights. This captured information can be used to synthesize novel images of the subject under arbitrary lighting conditions or from arbitrary viewpoints. This process enables a number of visual effects, such as creating digital replicas of actors that can be used in movies [1] or high-quality postproduction relighting [46]. + +In many cases, however, it is often infeasible to get access to a light stage for capturing a particular subject, because light stages are not easy to find: they are expensive and require significant technical expertise (often teams of people) to build and operate. In these cases, hope is not lost — one can turn to methods that are trained on light stage data, with the intention of generalizing to new subjects. These methods do not require the subject to be captured by a light stage but instead use a machine learning + +model trained on a collection of previously acquired light stage captures to enable the same applications as a light stage, but from only one or several images of a new subject [6, 25, 30, 38, 40, 50, 52]. Unfortunately, these methods have difficulty faithfully reproducing and editing the appearance of new subjects, as they lack much of the signal necessary to resolve the ambiguities of single-view reconstruction, i.e., a single image of a face can be reasonably explained by different combinations of geometry, illumination, and reflectance. + +In this paper, we propose an intermediate solution — one that allows for personalized, high-quality capture of a given subject, but without the need for expensive, calibrated capture equipment. Our method, which we dub SunStage, uses only a handheld smartphone camera and the sun to simulate a minimalist light stage, enabling the reconstruction of individually-tailored geometry and reflectance without specialized equipment. Our capture setup only requires the user to hold the camera at arm's length and rotate in place, allowing the face to be observed under varying angles of incident sunlight, which causes specular highlights to move and shadows to swing across the face. This provides strong signals for the reconstruction of facial geometry and spatially-varying reflectance properties. The reconstructed face and scene parameters estimated by our system can be used to realistically render the subject in new, unseen lighting conditions — even with complex details like self-occluding cast shadows, which are typically missing in purely image-based relighting techniques, i.e., those that do not explicitly model geometry. In addition to relighting, we also show applications in view synthesis, correcting facial perspective distortion, and editing skin reflectance. + +Our contributions include: (1) a novel capture technique for personalized facial scanning without custom equipment, (2) a system for optimization and disentanglement of scene parameters (geometry, materials, lighting, and camera poses) from an unaligned, handheld video, and (3) multiple portrait editing applications that produce photorealistic results, using as input only a single selfie video. + +# 2. Related works + +Face modeling. Extensive research has been devoted to the modeling of human faces, leading to various 3D morphable models (3DMMs) [2, 3, 5-7, 9, 10, 25, 31-33, 43]. These models are parametric (maybe in the form of neural networks [33]), allowing one to express variations compactly with a vector. They also encode strong priors learned from real scans. The groundbreaking face 3DMM is that of Blanz and Vetter [3] containing models for shape, expression, and appearance (the Phong model). Also influential is the FLAME model [25] that uses vertex-based Linear Blend Skinning (LBS). FLAME is described by a mapping from + +shape, pose, and expression vectors to a list of vertices. We refer the reader to the survey by Egger et al. [13] for different face morphable models. + +Such parametric face models provide a low-dimensional space for optimization or learning algorithms. DECA [14] uses the FLAME model to estimate detailed facial geometry (and albedo) from single images, by predicting additional displacement maps and adding them to the estimated FLAME models. More recently, NextFace [12] employs the 3DMM geometry and albedo priors to learn an albedo residual that captures more facial details. + +Without modeling 3D face geometry, researchers have also achieved photorealistic synthesis of portrait images using generative models and large-scale high-quality image datasets [22, 23]. + +Light stage capture. The light stage as described in Debevec et al., achieves impressive portrait reconstruction and relighting by capturing a series of images of the face under varying illumination [11]. Subsequent work made this process faster, more efficient, and explored different types of illuminants [15, 16, 28]. + +Given that a light stage is not always accessible, a number of methods have been proposed to achieve similar outputs from a single (or few) input portrait images [18,19,29,30,40,41,48-50]. These method rely on a dataset of light stage captures or synthetic examples as training data. + +Our setup can be thought of as a "minimalist light stage" formed by just the sun and a rotating camera, without requiring the high construction and maintenance costs of building a light stage. This parameterization of a sun and skylight model has been shown to be effective in photometric stereo [17,21] and scene factorization [27,42]. In a similar spirit, Calian et al. [8] focus on lighting estimation using faces as "light probes". Sengupta et al. propose to circumvent the need for a complicated light stage by recording the facial appearance responses to varying contents displayed on a desk monitor, and then perform portrait relighting [37]. Sevastopolsky et al. also attempt to simplify the capture setup from a light stage to a mobile phone camera with a co-located flash [39]. Unlike our work, which is physically-based, their approaches use neural rendering, and therefore have less direct control over lighting, material, and scene parameters. + +# 3. Overview + +Our method targets accurate reconstruction of scene lighting, subject geometry, and material properties from a handheld video sequence of a person rotating in place under the sun. Given a selfie video, we take a test-time optimization approach that uses the information from all frames of the video to solve for a physical model of the + +![](images/3c538ae1aca93b227dc391b75f76279fb153275fca6f86a09cbc7ad677a06251.jpg) +Figure 2. Overview. Our method jointly reconstructs geometry, skin reflectance, lighting, and camera pose from a selfie video sequence of a person rotating under the sun. Our system begins by extracting supervisory information from the video sequence: facial landmarks, foreground alpha mattes, and camera orientations. These are used to supervise the optimization of a collection of scene parameters (full list in Sec. 5.2) used in a physically-based renderer. The rendered output is an image consisting of diffuse, specular, and ambient light contributions. After optimization, the solved scene parameters can be used for a number of editing applications, shown in Sec. 7. + +![](images/74643e55364fe4de4bc3d155f6e3eae252ad6fc92e38065d2e9321af88206377.jpg) + +![](images/1070b43376eb454b7e7766e07c1e5befa01db106d541ebed13ad6e99df8ab4cf.jpg) + +scene: the geometry and material properties of the face, scene lighting, and camera parameters (Fig. 2). This physical model consists of a base face shape parameterized by a low-dimensional deformable model $X^b$ , a displacement map $\Delta X$ , a reflectance model with diffuse $R^d$ and specular components $R^s$ , scene lighting $L_i$ , and a perspective camera $C$ . These components are explained in detail in Sec. 4. + +After this model has been recovered, we can modify the scene and the subject parameters to re-render images. We show several editing applications in Sec. 7: editing skin reflectance, relighting with arbitrary environment map, improving harsh lighting conditions (by softening shadows and adding fill lights), and adjusting camera parameters to change viewpoint or manipulate perspective effects. + +# 4. Formulation + +Given an input video, our system reconstructs the parameters of a physical model: i.e., geometry and reflectance of the subject, the scene lighting parameters, and the camera parameters. In this section, we detail all of these parameters and describe the rendering process that turns these parameters into an image. + +Geometry. We denote $X_{j}$ as the full mesh of the subject for frame $j$ , composed of a per-frame coarse mesh $X_{j}^{b}(\beta, \theta_{j}, \psi_{j})$ and a global displacement map $\Delta X$ . The coarse mesh $X_{j}^{b}$ is a FLAME deformable face model [25] + +defined by global shape code $\beta$ , per-frame pose code $\theta_{j}$ , and per-frame expression code $\psi_{j}$ . $X_{j}^{b}$ also contains per-vertex UV coordinates, which maintain correspondence across variations in $\theta_{j}$ and $\psi_{j}$ . As such, we model all our global (per-subject) spatially varying parameters in UV space, and sample values per-fragment when rasterizing. + +The displacement map $\Delta X$ is used to model fine details like wrinkles that cannot be represented by $X_{j}^{b}$ . We displace the coarse geometry by $\Delta X$ at rasterization time, by sampling a displacement value per-fragment and displacing each fragment along the surface normal $N_{j}$ of the coarse mesh $X_{j}^{b}$ . After displacement, the updated fragment positions are used to compute a new surface normal $N_{j}^{\prime}$ . + +$X_{j}^{b}$ is optimized per-frame, since it accounts for subtle (and unavoidable) variations in expression and head pose during the capture, which are modeled by $\theta_{j}$ and $\psi_{j}$ . $\Delta X$ , on the other hand, is optimized in UV-space (i.e., globally per-subject), since the deformations it encodes are invariant to the changes in expression or pose. Formally, the final geometry $X_{j}$ is given by: + +$$ +X _ {j} = X _ {j} ^ {b} \left(\beta , \theta_ {j}, \psi_ {j}\right) + \Delta X \odot N _ {j} \tag {1} +$$ + +where $\odot$ is the Hadamard product. + +Reflectance. We model the skin reflectance, denoted as $R(x,\omega_i,\omega_o) \in \mathbb{R}^3$ , where $x$ is a 3D point on the face geometry $X$ , $\omega_i$ is the incoming light direction, and $\omega_o$ is the + +outgoing direction, using a diffuse and a specular component: $R = R^{d} + R^{s}$ . + +The diffuse component $R^d (x,\omega_i)\in \mathbb{R}^3$ is a Lambertian reflectance model consisting of an albedo map, $a$ , which we optimize as a per-subject UV-space image. For the skin's specular component, we use the Blinn-Phong model [4]. + +$$ +R ^ {s} \left(x, \omega_ {i}, \omega_ {o}\right) = k _ {s} \frac {s + 2}{2 \pi} \left(h \left(\omega_ {i}, \omega_ {o}\right) \cdot n (x)\right) ^ {s} \tag {2} +$$ + +where $h(\omega_i, \omega_o) = \text{normalize}(\omega_i + \omega_o)$ is the half vector, $k_s$ is the specular intensity, $s$ is the specular exponent, and $(s + 2) / (2\pi)$ is the normalization term for the reflection lobe to integrate to 1. Following [47], we segment the UV-space map into 10 segmented specular reflectance clusters. We then optimize for a spatially-varying pair of values $(s, k_s)$ per-cluster, enabling varying shininess across the face. + +While Blinn-Phong does not model many complex effects such as subsurface scattering, our experiments with other models for facial reflectance, such as microfacet models [44], show no significant quality improvements, and often introduce unstable training. More analysis is provided in the supplementary material. + +Lighting. We use a sun-sky model to represent lighting as the sum of an "ambient" environment map and the sun: $L_{i}(x,\omega_{i}) = L_{i}^{\mathrm{amb}}(\omega_{i}) + L_{i}^{\mathrm{sun}}(\omega_{i})$ . Note neither $L_{i}^{\mathrm{amb}}(\omega_{i})$ nor $L_{i}^{\mathrm{sun}}(\omega_{i})$ depends on the 3D point $x$ , since we model both as directional lights. Optimization-wise, our lighting parameters consist of a $16 \times 32 \times 3$ environment map for ambient lighting, the sun direction $p^{\mathrm{sun}} \in S^{3}$ , and the scalar sun intensity $k^{\mathrm{sun}}$ . We fix the sun color to white [1, 1, 1] in our lighting model to resolve the albedo-illumination ambiguity. + +# 4.1. Rendering + +We calculate the outgoing radiance $L_{o}$ at 3D location $x$ as viewed from viewing direction $\omega_{o}$ as: + +$$ +\begin{array}{l} L _ {o} (x, \omega_ {o}) \\ = \int_ {\Omega} V \left(x, \omega_ {i}\right) L _ {i} \left(x, \omega_ {i}\right) \odot R \left(x, \omega_ {i}, \omega_ {o}\right) \left(\omega_ {i} \cdot n (x)\right) d \omega_ {i} (3) \\ = \sum_ {\omega_ {i}} V (x, \omega_ {i}) \left(L _ {i} ^ {\mathrm {a m b}} (\omega_ {i}) \odot R ^ {d} (x, \omega_ {i}) \right. (4) \\ + L _ {i} ^ {\text {s u n}} \left(\omega_ {i}\right) \odot R ^ {d} \left(x, \omega_ {i}\right) + L _ {i} ^ {\text {a m b}} \left(\omega_ {i}\right) \odot R ^ {s} \left(x, \omega_ {i}, \omega_ {o}\right) \\ \left. + L _ {i} ^ {\operatorname {s u n}} \left(\omega_ {i}\right) \odot R ^ {s} \left(x, \omega_ {i}, \omega_ {o}\right)\right) \left(\omega_ {i} \cdot n (x)\right) \triangle \omega_ {i} (5) \\ \end{array} +$$ + +where $V(x,\omega_{i})$ is the light visibility at $x$ from $\omega_{i}$ , and $L_{i}(x,\omega_{i})$ is the incoming radiance reaching $x$ from $\omega_{i}$ . We ignore the specular reflection caused by the ambient lightning, i.e., $L_{i}^{\mathrm{amb}}(\omega_{i})\odot R^{s}(x,\omega_{i},\omega_{o})$ , since it is much weaker + +than the specular reflection of the sun. In the next subsections, we will group the terms into a diffuse contribution $L_{o}^{d}$ and a specular contribution $L_{o}^{s}$ : $L_{o} = L_{o}^{d} + L_{o}^{s}$ . For the final rendered color value, we apply the Reinhard operator [35] and a gamma correction of $\gamma = 2.2$ to $L_{o}$ to convert from linear to sRGB space. + +Diffuse contribution. The diffuse contribution $L_{o}^{d}$ is then given by only the diffuse terms of Equation 5: + +$$ +\begin{array}{l} L _ {o} ^ {d} (x) = \sum_ {\omega_ {i}} L _ {i} ^ {\mathrm {a m b}} (\omega_ {i}) \odot \frac {a (x)}{\pi} (\omega_ {i} \cdot n (x)) \Delta \omega_ {i} \\ + V (x, p ^ {\text {s u n}}) k ^ {\text {s u n}} [ 1, 1, 1 ] \odot \frac {a (x)}{\pi} \left(p ^ {\text {s u n}} \cdot n (x)\right) \Delta p ^ {\text {s u n}} \tag {6} \\ \end{array} +$$ + +where $a(x)$ is the albedo at point $x$ , $k^{\mathrm{sun}}$ is the (optimized) sun intensity, and $p^{\mathrm{sun}}$ is the (optimized) sun direction. The sun is modeled as a directional light source, so the second summation can be simplified to a single term, only in the direction of $p^{\mathrm{sun}}$ . We additionally optimize for a high-dynamic-range (HDR) environment map $E \in \mathbb{R}^{16 \times 32 \times 3}$ , from which values of $L_{i}^{\mathrm{amb}}$ are sampled. + +Specular contribution. The specular contribution $L_{o}^{s}$ at each pixel is given by only the specular term due to the sun in Equation 5 (recall that we ignore the specular ambient term due to its weak contribution): + +$$ +\begin{array}{l} L _ {o} ^ {s} (x, \omega_ {o}) = V (x, p ^ {\mathrm {s u n}}) k ^ {\mathrm {s u n}} [ 1, 1, 1 ] k _ {s} \frac {s + 2}{2 \pi} \\ \left(h \left(p ^ {\text {s u n}}, \omega_ {o}\right) \cdot n (x)\right) ^ {s} \left(p ^ {\text {s u n}} \cdot n (x)\right) \Delta p ^ {\text {s u n}} \tag {7} \\ \end{array} +$$ + +where we have substituted Equation 2 and reduced the summation to just one term at $p^{\mathrm{sun}}$ (since $L_{i}^{\mathrm{sun}}$ is 0 elsewhere). + +Shadow map. In order to generate a map of self-occluded shadows, we perform two passes of rasterization: first, we render a $z$ -buffer from a virtual orthographic camera aligned with the sun direction, $p^{\mathrm{sun}}$ , and then, when rasterizing a given camera viewpoint, compare all fragment positions $d_{\mathrm{hit}}$ to the light's $z$ -buffer $d_{\mathrm{shadow}}$ . To avoid precision issues and ensure smooth gradients for back-propagation, we implement a soft comparison as follows in generating shadow/visibility maps: + +$$ +V \left(x, \omega_ {i}\right) = 1 - \text {s i g m o i d} \left(k \left(d _ {\text {h i t}} - d _ {\text {s h a d o w}} \times b\right)\right) \tag {8} +$$ + +where $k$ is the falloff slope, and $b$ the tolerance. We use $k = 800$ and $b = 1.0015$ . + +# 5. Optimization + +The described physical model contains a large number of parameters to be optimized, controlling scene elements like lighting, geometry, pose, and texture. Unfortunately, + +naively optimizing all these parameters from scratch does not result in an optimal solution, since the final observed appearance of the face can often be explained variously through changes to geometry, material properties, lighting or camera parameters, making optimization severely under-constrained and ambiguous. Therefore, we adopt a two-stage optimization approach, through which parameters are gradually enabled. In this section, we describe this process and the relevant losses that guide optimization. + +# 5.1. Coarse alignment + +Our system begins by using an off-the-shelf network (DECA [14]) to generate, for each input image, a set of shape parameters $\beta$ , pose parameters $\theta_{j}$ , and expression parameters $\psi_{j}$ of a FLAME face model [25], as well as the relative pose parameters of the virtual camera observing the 3D face. Unfortunately, as with many other single-image facial geometry estimators, DECA assumes an orthographic projection model and therefore cannot accurately recover geometry for our selfie capture sequences, which contain heavy perspective effects (Figure 6). Without a good initialization for geometry, optimization of lighting and material properties seldom converges to an optimal solution due to the heavily ambiguous nature of our optimization problem. + +To circumvent this issue, we employ a first stage of optimization where we only optimize for the parameters of a perspective camera (with a known focal length, extracted from input metadata) and the face geometry parameters $(\beta, \theta_j, \psi_j)$ . As initialization for this optimization process, we use the predicted DECA values for each frame's pose $\theta_j$ and expression $\psi_j$ , but set all frames to the average predicted shape $\beta_{\mathrm{avg}} = \frac{1}{N} \sum_j \beta_j$ , since the identity remains constant across all frames. To convert DECA's orthographic camera to a perspective camera, we additionally optimize for an unknown scale $S$ and translation $T_j$ , which are initialized to empirically chosen values $S = 2.6e4$ , $T_j = (0, 0, 1.5e5)$ . During optimization, the face shape $\beta$ and scale $S$ are shared across all frames, while camera pose $T_j$ , expression $\psi_j$ , and pose $\theta_j$ are optimized per-frame. Note that DECA controls the relative orientation of the camera and the face by varying the pose code $\theta$ instead of the camera rotation. We adopt this formulation and keep the camera orientation fixed relative to the face. The global orientation of the camera at each frame (and therefore the face) is extracted from the capture video, either through a structure-from-motion system or IMU measurements commonly available on a smartphone. + +We use two losses to guide this optimization: a mask loss $L_{\mathrm{mask}}$ and a landmark loss $L_{\mathrm{lmk}}$ . The FLAME model includes 3D facial landmark points, corresponding to the standard 68-point facial landmarks set [36] used in facial tracking. Our landmark loss minimizes the L1 distance between the 2D projection of these 3D landmarks (into the + +input camera viewpoint) and 2D landmarks estimated from the input frame by a 2D landmark detector HRNets [45]. + +The facial landmarks provide a stong constraint on facial feature alignment, but are sparse, and therefore cannot constrain the overall shape or boundary of the mesh. To supplement it, we include a silhouette loss $L_{\mathrm{mask}}$ , which penalizes the L2 difference between the rasterized mask of the mesh $I_{\mathrm{sil}}$ and the semantic segmentation mask $I_{\mathrm{mask}}$ of the input image, using an off-the-shelf semantic segmentation network [26] trained to segment humans in portrait photographs. + +The final pose loss is then: $L_{\mathrm{pose}} = L_{\mathrm{mask}} + L_{\mathrm{lmk}}$ , optimized using an ADAM optimizer [24]. See supplemental for optimization parameters. + +# 5.2. Photometric optimization + +Once the 3D model and camera parameters are approximately aligned, we proceed to the second stage of optimization, in which we optimize the precise facial geometry, lighting, and reflectance properties. All the parameters optimized in the first stage (Section 5.1) remain as free variables. In total, the parameters optimized during this stage include: Lighting parameters: (1) $p^{\mathrm{sun}}$ , the global sun direction, (2) $E$ , the global environment map, (3) $k^{\mathrm{sun}}$ , the global sun intensity, Facial geometry parameters: (4) $\beta$ , the global FLAME shape code, (5) $\psi_j$ , the per-frame expression code, (6) $\theta_j$ , the per-frame pose code, (7) $\Delta X$ , the global deformation map, Material properties: (8) $k_s$ , the global, spatially-varying specular intensity, (9) $s$ , the global, spatially-varying specular roughness, (10) $a$ , the global, spatially-varying surface albedo, Camera pose parameters: (11) $T_j$ , the per-frame perspective camera translation, and (12) $S$ , the global scene scale. + +During optimization, we randomly select a frame $j$ , render the face using a differentiable rasterizer [34] and the equations described in Section 4 to get the rendered image $\hat{I}$ . In addition to the previously defined landmark and mask losses, we include L2 and VGG [20] photometric losses, comparing the original and reconstructed images: + +$$ +L _ {\text {p h o t o}} = \left| \left| \hat {I} _ {j} \cdot I _ {\text {s i l}} - I _ {j} \cdot I _ {\text {m a s k}} \right| \right| _ {2} \tag {9} +$$ + +We also include an L2 regularization $L_{E}$ and L2-smoothness regularization $L_{Es}$ on the reconstructed environment map, to encourage the majority of the lighting to be explained by direct sunlight and to aid in disentangle-ment of the sun and ambient lighting. The total optimized loss becomes: + +$$ +\begin{array}{l} L = \lambda_ {\text {m a s k}} L _ {\text {m a s k}} + \lambda_ {\text {l m k}} L _ {\text {l m k}} + \lambda_ {E} L _ {E} + \lambda_ {E \mathrm {s}} L _ {E \mathrm {s}} \\ + \lambda_ {\mathrm {V G G}} L _ {\mathrm {V G G}} + \lambda_ {\text {p h o t o}} L _ {\text {p h o t o}} \tag {10} \\ \end{array} +$$ + +with $\lambda_{\mathrm{mask}}, \lambda_{\mathrm{lmk}} = 0.05, \lambda_{\mathrm{VGG}} = 0.005, \lambda_{E} = 0.01, \lambda_{Es}, \lambda_{photo} = 1$ . Additional optimization details are provided in the supplemental materials. + +![](images/52f5d3506df2da247be34ce6af111df2b650be54aec27ea643eda1ba2a6e61d0.jpg) +(a) DECA [14] + +![](images/b88f041e2d99bc9832219eb121f46731d55fb38da3a7582a7c6c43d89a550965.jpg) +(b) DPR [52] + +![](images/0b20f9c935a049ac92da3097ee7dc4c6e57419465fb34578d57cd5f3a10deeea.jpg) +(c) GCFR [18] + +![](images/31ca6ac7b4b2f6315f2ab8c53c295c821ec0d9980c4f34f83059d409ca250085.jpg) +(d) TR [30] + +![](images/d089802bbdb35ea642b96694c0e65a03343638319389c7936a9d5debda5d5f4d.jpg) +(e) NextFace [12] + +![](images/389ed353915131fb1aa231f1c2d610b58e8e3feb2c7080222657736a3d4a0dc4.jpg) +(f) NLT [51] + +![](images/96d06ce9832da61fc03bc2768eb050b47a1bc29c66894022180b1ba67e22a1a8.jpg) +(g) Ours + +![](images/0481f3271eaac39ce508bacd29ddae5422c60f9fbf655c4ecb1d0583ce0f336d.jpg) +(h) GT + +![](images/63f889f69e65178df554d6b6108e518b25678feffe0ccc03b4168367f1687eea.jpg) +Figure 3. Qualitative: Relighting. A comparison of our method at rendering a new (unseen) lighting environment (h). Our method is able to realistically synthesize the novel lighting condition, including cast shadows and specularity, and nearly matches the (unseen) target reference image. See supplement for additional details on experimental setup and analysis of results. + +![](images/70e7609460fb80139be943c0d1a9362611ce45a7b84481341cfbffc5f0136549.jpg) +(a) Nearest view + +![](images/6d77e6ffd4dac729bcfd27b8c6405737fd3a4b7884c8323dae15b42d6bf79f8a.jpg) + +![](images/82e79896fb021f99081dd578fd2625fafc214ce25057acfc1b4723ba93a479ac.jpg) +(b) DECA [14] + +![](images/2f28d4fe4a2a58d714b7a518425b1dbcec9bbd209968c5734c4a31b160408397.jpg) + +![](images/6e987664c92a9483dd4062b5b292a9ec18f9bc0d72d4f23098f2a144c31c6720.jpg) +(c) NextFace [12] + +![](images/ca138ccb57cb9ba789ab336cce4533901e3033782cf28b11ab3f33ee005bfbd9.jpg) + +![](images/4b3bc4291722f72be538d4a74369afc093b808eff9488db8778dc3184ae9b98c.jpg) +(d) NLT [51] + +![](images/44c665776405e58528ae23d3d3f1772ce139be37341cdf658ff3c2313fa970dc.jpg) + +![](images/62114643901ca4b5168b3d1a5f302c9b661457ad37214e278b78d936931780bf.jpg) +(e) Ours +Figure 4. Qualitative: view synthesis. A comparison of our method at the task of generating an image from an unseen viewpoint (f), having only seen a limited collection of input viewpoints. See supplement for more details on experimental setup and analysis of results. + +![](images/2d45431b527777126b725e35b985f6217c22f1f8fee6f15cbf90992fcab6b65c.jpg) + +![](images/9410ea937733907d2ef32f96d3cac579537b70d2644566c9c5c992a047668017.jpg) +(f) GT + +# 6. Evaluation + +In this section, we detail quantitative and qualitative experiments comparing our approach with state-of-the-art methods and ablated variants of our method. + +Baseline comparisons. We evaluate our method on the tasks of novel-view synthesis and relighting. For novel-view synthesis, we compare our method with DECA [14], Neural Light Transport (NLT) [51], and NextFace [12]. For relighting, we compare with DECA, NLT, NextFace, GCFR [18], image-based methods Deep Single Image Portrait Relighting (DPR) [52], Total Relighting (TR) [30] and NVPR [49]. Additional comparisons and details on the experimental setups are provided in the supplemental materials. + +We present qualitative comparisons for relighting in Figure 3 and novel view synthesis in Figure 4. Quantitative + +comparisons on these images are provided in Table 1. These testing images consist of (1) a multi-view capture of the face, in which the subject remains still and the camera is moved to novel viewpoints in the same environment as the original capture, and (2) front-facing sequences in novel environment lighting and unseen sun positions. All testing images are not seen during training of our method, NLT or NextFace. The results shown in Figures 3 and 4 as well as Table 1 clearly demonstrate that our method outperforms all the baselines at both relighting and view synthesis. Single-image methods (DECA, GCFR, DPR, TR) can generalize to other subjects, but fail to recover more faithful and physically accurate facial details. Comparison with multi-image methods (NLT, NextFace) demonstrates that SunStage is a better reconstruction system. Additional analysis of the comparisons is provided in the supplemental materials. + +![](images/32d160419810920fd0346182a686c69c47e9250e107026d7fdd6949f3a73e12e.jpg) + +![](images/c9c33fb6b2de3789d68733d03be92948914101c742df20584b1d85dca464b8d3.jpg) + +![](images/630c6c2a58f07c1a036f57b37fcda6ab6d1e63dff504fca8a40e17bf34b0e098.jpg) + +![](images/baa9a6d1b3736c55f5b66c91b53bf2f5ec88c8939fef3fa5e537798ebb65ec8e.jpg) + +![](images/ae520b4158da99aaab95b336a3cd11a13efdd3785beba4a352a232fad5d6fd4d.jpg) + +![](images/c97590b23c88a48e4b282520d1b12ce392391de9ff745e51becfcc4529d98033.jpg) + +![](images/2f75e2f17afc0f648eeb2b37139e7832cbcea571a55da79424c9c225d3dd2707.jpg) + +![](images/6645e13675557b3ed95647b24a162194f1325aa3bdfc2284aec35b0bb80e8b15.jpg) + +![](images/c603800cd4d7a07ba485a548889b9690205252c37ff0b9a6688f76970e5fd16b.jpg) + +![](images/52fc257f218259558efcf5c9436cd8d81749b66ce813448d5c1c6fd185e09655.jpg) + +![](images/e430f445978ebbb89f326ef48b1830637e61341664463b1c0d335232870dce32.jpg) + +![](images/4482c11b62f0c41a2fb5db699013d87b007d426a0ff496208a3120573f8b1565.jpg) + +![](images/67af2371cabb1c1e86ca548744ee11cbaaad7045eb5e9059977aa271f00e8ac4.jpg) + +![](images/71e3128a28d7dbfbcfe8cd2eda9293d94b6c2d259eb7fdd84eed344a3ebf91d8.jpg) + +![](images/710bd2c7abdd4e57051c963654097daf5125ca9f115239f71b1ec7324e6dbf99.jpg) +Normal + +![](images/eb4bd385f83cf045e620d1bd44b404475b62a9e3cf0c513ca46a0caf49fe7517.jpg) +Albedo + +![](images/300c7b77082e13e199cd371d843c25a2833c3500874c1da3f2314b869970b8e9.jpg) +Specular + +![](images/40b0cb98907b7d1f5800d9e60b2f79245ac342a1d9c4d2108b0e4027a53435e7.jpg) +Ambient + +![](images/9d4b93c6d8dbe5e719ed206d278823135cdb553fe1fa470a3de5b30acdbba09d.jpg) +Diffuse + +![](images/122249523a9bed97ec54557dcfc649ec8362bbe4251358b2ec65c4d27f4eca41.jpg) +Reconstruction + +![](images/41828a35e4cf3fffd69f21bf2bf3f0fbfde939fa18d24f0856c5d75a2eb496f2.jpg) +Ground Truth + +
RelightingNovel view synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DECA [14]16.410.690.2516.640.660.29
GCFR [18]16.970.700.20---
DPR [52]19.030.720.19---
NLT [51]20.150.750.1822.270.790.15
Total Relighting [30]20.240.790.16---
NextFace [12]22.980.760.1522.550.750.15
Ours23.640.830.1025.280.840.09
Ours w/o coarse17.830.660.2319.650.700.17
Ours w/o SV ks, s21.310.770.1321.940.770.12
Ours w/o Lmask, Llmk16.460.610.3018.830.680.20
Ours w/o Lmask20.130.750.1520.540.740.15
Ours w/o opt. (β, θi, φi)18.670.690.1918.280.660.19
Ours w/o soft shadow21.460.770.1322.050.770.12
Ours w/o ΔX21.160.750.1521.800.750.14
+ +Table 1. Quantitative comparison. Comparison of our method on the tasks of novel view synthesis and relighting. See Section 6 for a description of the ablated variants. + +Disentanglement. In Figure 5, we demonstrate how Sun-Stage decomposes the appearance of a portrait photograph into different components: specular, diffuse, and ambient. We also visualize the surface normal, albedo, and other intermediate representations to show that our method is able to effectively recover a physically plausible reconstruction of the real world and disentangle the different components that contribute to the final appearance. We further validate the quality of the reconstructed geometry and materials in the supplementary material. + +Ablation studies. In addition to our comparisons with + +![](images/3c7469f6efaaed23fba602c343523ebed64979bcf417288432441edd0a5a4c02.jpg) +Figure 5. Decomposition. We show all the components which comprise our final rendered image to demonstrate that our method not only closely recreates the ground truth image (reproducing realistic highlights and shadows), but also performs a meaningful decomposition of lighting components and facial geometry. Note that our reconstructed surface normals include high frequency details specific to each subject, like wrinkles and birthmarks, which are used in computation of the shadows and specular reflections. +(a) Input image +Figure 6. Perspective. DECA's assumption of an orthographic camera is broken by the strong perspective effects in selfies, causing poor alignment (b) with input images (a). Our first stage of optimization (c) (Sec. 5.1) improves alignment by solving for the parameters of a perspective camera and refined shape parameters. In the second stage we additionally optimize for a displacement map $\Delta X$ to produce our final shape with finer geometric details like wrinkles (d). Red line added to highlight alignment with (a). + +![](images/befc0e0b5d14c0df09622bb32efca79c3a5ff72feb3f75a6dcfb51c51faea23d.jpg) +(b) DECA [14] + +![](images/1f5bfc03898bc74c2796b4542eeecf4c8492ee6d1fc0f35ca042892d7146b116.jpg) +(c) Stage 1 + +![](images/644d3fc7d0c0a59ca6ef58f75db49803ef155778a8726cfc755d6e12590d0ecc.jpg) +(d) Final + +the state-of-the-art baselines, we also compare with ablated variants of our own method. In particular, we include seven such experiments in Table 1: our method (1) without the initial first stage of coarse geometric alignment, i.e., directly optimizing both geometric and photometric parameters from the start, (2) without the spatially varying specular parameters, instead using a single global scalar $s$ and $k_{s}$ , (3) without the geometric alignment losses $L_{\mathrm{mask}}$ and $L_{\mathrm{lmk}}$ , (4) without just $L_{\mathrm{mask}}$ , (5) without shape optimization, i.e., keeping the initial shape code predicted by DECA, (6) without soft shadow computation, i.e., using a hard z-buffer comparison to compute a shadow map instead of our soft + +![](images/85fba16c5fbf0821758be6dd5c407c10ffcb4c91ae636bd503da3c288b4c0d3b.jpg) +(a) Input + +![](images/ab834c038346f3deae18213cc163346f51647dbbddf4f6b5f2d122140f5031c3.jpg) +(b) Soft shadows + +![](images/7ca8f30c33e86b4b60a3039b729a3d85ad3353e57db7f33f571f3ca0b8094206.jpg) +(c) (b) + fill light + +![](images/71e20c4ee897b13f12516cc0833afe8536be8551fa1446a62ace94dd31049b55.jpg) +(d) Swap scene +Figure 7. Adjusting lighting parameters. We can adjust the recovered scene parameters to improve the lighting conditions in an input image (a) by softening the harsh shadows cast by the nose (b), adding a fill light to brighten the shaded region (b), or replacing the environment altogether (d). + +comparison operator in Equation 8, and (7) without the displacement map $\Delta X$ . Visual results for each of these variants are provided in the supplemental material. + +# 7. Applications + +Relighting. We demonstrate two types of relighting applications: (1) lighting modification and (2) lighting replacement. Practical lighting modification is common in portrait photography when the lighting conditions are not ideal, e.g., when direct sunlight casts undesirable harsh shadows with high contrast. A common practice is to make the light source larger and more diffuse by using a scrim or bounce card. In Fig. 7b, we show that by virtually increasing the size and spreading the energy of our reconstructed lighting source (i.e., the sun), we are able to soften the shadows and re-render a more visually pleasing face. Another approach to reducing the effects of harsh shadows is adding local fill lights, which reduces the contrast between the lit and shaded regions (Fig. 7c). Alternatively, fill lights can also be used for artistic purposes, to create dramatic lighting effects (Fig. 1b). Finally, replacing the scene lighting with that of a novel environment (Fig. 7d) is a necessary step in realistically inserting a captured subject into a virtual scene, which is useful for visual effects and VR applications. + +View Synthesis. In Figure 8, we show that our reconstructed 3D model of the face can be used to synthesize new views by manipulating the viewpoint of the camera. We can also change other camera parameters, such as the focal length, to reduce the perspective effects on the face, which is often desirable for selfie images that contain significant facial distortion due to perspective. + +Skin Reflectance Editing. We are also able to edit the reflectance components of the subject. As shown in Figure 1e, + +![](images/66dae6f19b9f6d50ceae33285e157f6ccd096bab7dd701c10c305dee0453db39.jpg) +(a) Short focal + +![](images/47b3313a562875c48a0277e625984bf02508ff9e849aca500d864a7cf8b7d358.jpg) +(b) Long focal + +![](images/ef8f4c0cf1ec2cb121e170c4c7d70c6832dfc8a2870d61cc84814ea46d4145f4.jpg) +(c) Novel view + +![](images/e42e876402a4bdb2cd81e76b8c1ee78d50de1c3df27ad855407c9d136cf4fbba.jpg) +(d) Novel view 2 + +![](images/32a964bbde0e982d6f0a192d8b8fe7a5c30b9df1962a21b0f2f699776cdc8fb5.jpg) +Figure 8. Changing camera parameters. We can change the recovered camera parameters to render novel views (c,d) or change the focal length (a,b). +(a) Less shinier +Figure 9. Adjusting the specularity. We can change the the specular properties of the face, making the face less shinier (a, c) or more shinier (b, d). + +![](images/371ce804727e1555a587e8cfd5407b473d8c005e19b8616e728a4864d328780f.jpg) +(b) More shinier + +![](images/1aaa84b6d71119d71f3d7fa59871cff6d9f12c3385a996aef39ac16683409d02.jpg) +(c) Less shinier + +![](images/066a1cbed01d26fdf80fdf8dd23baed3abadc50ca6826d4a48bfa1f7a3f6481f.jpg) +(d) More shinier + +we can adjust the optimized albedo to add freckles, stickers, or other textures that realistically interact with reflections, shadows, and other elements of scene lighting, or we can adjust the specular properties of the face, making the face more or less shinier, as shown in Figure 9. + +# 8. Conclusion and Discussions + +In this paper, we propose SunStage, a lightweight and practical facial capture, rendering, and editing system that can serve as a minimalist light stage. With a video of an individual rotating in-place under the sun, our system reconstructs a physical model of the subject and the scene lighting, which enables us to relight the subject with realistic reflections and cast shadows. Our system allows arbitrary lighting and reflectance control in the reconstructed physical space, which can be rendered to produce photo-realistic results. We demonstrate several applications such as editing skin reflectance, relighting, and view synthesis. + +Limitations. Our system inherits the limitations of morphable face models and is unable to model hair, teeth, or clothing geometry, beyond slight deformations. Additionally, certain regions which are seen under constant shadow or specular reflection (and therefore have no cues on reflectance or albedo) are sometimes unable to be decomposed accurately into separate reflectance and lighting components. Visualization and further discussions of the system's limitations are provided in the supplemental material. + +# References + +[1] Oleg Alexander, Mike Rogers, William Lambeth, Jen-Yuan Chiang, Wan-Chun Ma, Chuan-Chang Wang, and Paul Debevec. The digital emily project: Achieving a photorealistic digital actor. IEEE Computer Graphics and Applications, 30(4):20-31, 2010. 1 +[2] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4), 2021. 2 +[3] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 2 +[4] James F Blinn. Models of light reflection for computer synthesized pictures. In Proceedings of the 4th annual conference on Computer graphics and interactive techniques, pages 192-198, 1977. 4 +[5] Timo Bolkart and Stefanie Wuhrer. A groupwise multilinear correspondence optimization for 3d faces. In Proceedings of the IEEE international conference on computer vision, pages 3604-3612, 2015. 2 +[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2 +[7] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In European Conference on Computer Vision, pages 297-312. Springer, 2014. 2 +[8] Dan A Calian, Jean-François Lalonde, Paulo Gotardo, Tomas Simon, Iain Matthews, and Kenny Mitchell. From faces to outdoor light probes. In Computer Graphics Forum, volume 37, pages 51-61. Wiley Online Library, 2018. 2 +[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 2 +[10] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3085-3093, 2017. 2 +[11] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 2 +[12] Abdallah Dib, Cedric Thebault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 2, 6, 7 + +[13] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2 +[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2, 5, 6, 7 +[15] Graham Fyffe and Paul Debevec. Single-shot reflectance measurement from polarized color gradient illumination. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2 +[16] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. In Proceedings of the 2011 SIGGRAPH Asia Conference, pages 1-10, 2011. 2 +[17] Yannick Hold-Geoffroy, Jinsong Zhang, Paulo F U Gotardo, and Jean-François Lalonde. $x$ -hour outdoor photometric stereo. In International Conference on 3D Vision, 2015. 2 +[18] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2, 6, 7 +[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2 +[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 5 +[21] Jiyoung Jung, Joon-Young Lee, and In So Kweon. One-day outdoor photometric stereo via skylight estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4521–4529, 2015. 2 +[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[25] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 3, 5 +[26] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with tempo- + +ral guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5 +[27] Andrew Liu, Shiry Ginosar, Tinghui Zhou, Alexei A Efros, and Noah Snavely. Learning to factorize and relight a city. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, pages 544–561. Springer, 2020. 2 +[28] Abhinitra Meka, Christian Haene, Rohit Pandey, Michael Zollhöfer, Sean Fanello, Graham Fyffe, Adarsh Kowdle, Xueming Yu, Jay Busch, Jason Dourgarian, et al. Deep reflectance fields: high-quality facial reflectance field inference from color gradient illumination. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2 +[29] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2 +[30] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 6, 7 +[31] Frédéric Pighin, Jamie Hecker, Dani Lischinski, Richard Szeliski, and David H Salesin. Synthesizing realistic facial expressions from photographs. In ACM SIGGRAPH 2006 Courses, pages 19-es. 2006. 2 +[32] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2 +[33] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 704-720, 2018. 2 +[34] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 5 +[35] Erik Reinhard. Parameter estimation for photographic tone reproduction. Journal of graphics tools, 7(1):45-51, 2002. 4 +[36] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 5 +[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2 +[38] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In Proceedings of the + +IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2 +[39] Artem Sevastopolsky, Savva Ignatiev, Gonzalo Ferrer, Evgeny Burnaev, and Victor Lempitsky. Relightable 3d head portraits from a smartphone video. arXiv preprint arXiv:2012.09963, 2020. 2 +[40] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul E Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):79-1, 2019. 2 +[41] Tiancheng Sun, Zexiang Xu, Xiuming Zhang, Sean Fanello, Christoph Rhemann, Paul Debevec, Yun-Ta Tsai, Jonathan T Barron, and Ravi Ramamoorthi. Light stage superresolution: Continuous high-frequency relighting. ACM Transactions on Graphics (TOG), 39(6):1-12, 2020. 2 +[42] Kalyan Sunkavalli, Wojciech Matusik, Hanspeter Pfister, and Szymon Rusinkiewicz. Factored time-lapse video. ACM Transactions on Graphics (Proc. SIGGRAPH), 26(3), Aug. 2007. 2 +[43] Luan Tran and Xiaoming Liu. Nonlinear 3d face morphable model. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7346-7355, 2018. 2 +[44] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4 +[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 2020. 5 +[46] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 1 +[47] Tim Weyrich, Wojciech Matusik, Hanspeter Pfister, Bernd Bickel, Craig Donner, Chien Tu, Janet McAndless, Jinho Lee, Addy Ngan, Henrik Wann Jensen, et al. Analysis of human faces using a measurement-based skin reflectance model. ACM Transactions on Graphics (ToG), 25(3):1013-1024, 2006. 4 +[48] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 2022. 2 +[49] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 2, 6 +[50] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2 + +[51] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. ACM Transactions on Graphics (TOG), 40(1):1-17, 2021. 6, 7 +[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, pages 7194-7202, 2019. 2, 6, 7 \ No newline at end of file diff --git a/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/images.zip b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf6519c6ee13d9ba0ea58649cfa193042ab99ab3 --- /dev/null +++ b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:044bb7504639d37c3e030a8e73c98dd98d468659c3cb98de2b6853f87877db6c +size 751874 diff --git a/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/layout.json b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fde8e97d02841c719d5d9f44a9c839e03c3acb2 --- /dev/null +++ b/2023/SunStage_ Portrait Reconstruction and Relighting Using the Sun as a Light Stage/layout.json @@ -0,0 +1,12796 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 125, + 545, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 125, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 545, + 144 + ], + "type": "text", + "content": "SunStage: Portrait Reconstruction and Relighting using the Sun as a Light Stage" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 164, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 164, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 164, + 482, + 213 + ], + "type": "text", + "content": "Yifan Wang1 Aleksander Holynski1 Xiuming Zhang2 Xuaner Zhang2 \n1University of Washington 2Adobe Inc. sunstage.cs.washington.edu" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 225, + 149, + 322 + ], + "blocks": [ + { + "bbox": [ + 53, + 225, + 149, + 322 + ], + "lines": [ + { + "bbox": [ + 53, + 225, + 149, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 149, + 322 + ], + "type": "image", + "image_path": "72a12d15f29c5906340f76c801ccc4c15a96e5877a8f469d131cc56faa4009a3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 323, + 149, + 419 + ], + "blocks": [ + { + "bbox": [ + 53, + 323, + 149, + 419 + ], + "lines": [ + { + "bbox": [ + 53, + 323, + 149, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 149, + 419 + ], + "type": "image", + "image_path": "b1647eeccb2d79f41272998fc706007083d921392df0450417e5c22ca93e0c29.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 421, + 132, + 430 + ], + "lines": [ + { + "bbox": [ + 70, + 421, + 132, + 430 + ], + "spans": [ + { + "bbox": [ + 70, + 421, + 132, + 430 + ], + "type": "text", + "content": "(a) Reconstruction" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 225, + 246, + 322 + ], + "blocks": [ + { + "bbox": [ + 151, + 225, + 246, + 322 + ], + "lines": [ + { + "bbox": [ + 151, + 225, + 246, + 322 + ], + "spans": [ + { + "bbox": [ + 151, + 225, + 246, + 322 + ], + "type": "image", + "image_path": "48fcdcab75a29a0909e9910f52cc24f3b29967214d99a9a5c9b46f441cd61f58.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 151, + 323, + 245, + 419 + ], + "blocks": [ + { + "bbox": [ + 151, + 323, + 245, + 419 + ], + "lines": [ + { + "bbox": [ + 151, + 323, + 245, + 419 + ], + "spans": [ + { + "bbox": [ + 151, + 323, + 245, + 419 + ], + "type": "image", + "image_path": "c070d1dc26fcb9105ab4de5ebce71ebaa7d1d35fa9e87583a3f0d6023ba520bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 421, + 224, + 430 + ], + "lines": [ + { + "bbox": [ + 171, + 421, + 224, + 430 + ], + "spans": [ + { + "bbox": [ + 171, + 421, + 224, + 430 + ], + "type": "text", + "content": "(b) Edit lighting" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 247, + 225, + 343, + 322 + ], + "blocks": [ + { + "bbox": [ + 247, + 225, + 343, + 322 + ], + "lines": [ + { + "bbox": [ + 247, + 225, + 343, + 322 + ], + "spans": [ + { + "bbox": [ + 247, + 225, + 343, + 322 + ], + "type": "image", + "image_path": "fe54843eb01bf19c9fcc6df87800725b06c232909981659bb78d65e4737af92a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 247, + 323, + 342, + 419 + ], + "blocks": [ + { + "bbox": [ + 247, + 323, + 342, + 419 + ], + "lines": [ + { + "bbox": [ + 247, + 323, + 342, + 419 + ], + "spans": [ + { + "bbox": [ + 247, + 323, + 342, + 419 + ], + "type": "image", + "image_path": "d94d69297e90d8ffe7512bbcbe3e7e0cdc582872dfacb0af8f04be6a3798a7f6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 421, + 325, + 430 + ], + "lines": [ + { + "bbox": [ + 262, + 421, + 325, + 430 + ], + "spans": [ + { + "bbox": [ + 262, + 421, + 325, + 430 + ], + "type": "text", + "content": "(c) Soften shadows" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 434, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 45, + 434, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 45, + 434, + 547, + 478 + ], + "type": "text", + "content": "Figure 1. Given a selfie video rotating under the sun, SunStage reconstructs geometry, material, camera pose, and lighting information. This recovered information can be used to (a) realistically re-render the input images, (b) modify the lighting conditions by adding / removing lights, (c) soften harsh shadows by changing the size of the reconstructed light sources (d) render the person in an entirely new environment, and (e) edit the albedo or material properties to add freckles, makeup, or stickers that realistically interact with scene lighting." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 343, + 225, + 439, + 322 + ], + "blocks": [ + { + "bbox": [ + 343, + 225, + 439, + 322 + ], + "lines": [ + { + "bbox": [ + 343, + 225, + 439, + 322 + ], + "spans": [ + { + "bbox": [ + 343, + 225, + 439, + 322 + ], + "type": "image", + "image_path": "dea3acc30753508cf83b136bde64244904c67a4a3436c4b4eb6638c914885923.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 343, + 323, + 438, + 419 + ], + "blocks": [ + { + "bbox": [ + 343, + 323, + 438, + 419 + ], + "lines": [ + { + "bbox": [ + 343, + 323, + 438, + 419 + ], + "spans": [ + { + "bbox": [ + 343, + 323, + 438, + 419 + ], + "type": "image", + "image_path": "251f48050b9b0dfa4cb01dcfb64672ba2facd76b6241852f1822a5184c846272.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 421, + 419, + 430 + ], + "lines": [ + { + "bbox": [ + 362, + 421, + 419, + 430 + ], + "spans": [ + { + "bbox": [ + 362, + 421, + 419, + 430 + ], + "type": "text", + "content": "(d) Swap lighting" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 441, + 225, + 536, + 322 + ], + "blocks": [ + { + "bbox": [ + 441, + 225, + 536, + 322 + ], + "lines": [ + { + "bbox": [ + 441, + 225, + 536, + 322 + ], + "spans": [ + { + "bbox": [ + 441, + 225, + 536, + 322 + ], + "type": "image", + "image_path": "277e23ea4b2eba9a410a8677de4158bc29ea2e68c64353837eefcc92899a625a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 440, + 323, + 534, + 419 + ], + "blocks": [ + { + "bbox": [ + 440, + 323, + 534, + 419 + ], + "lines": [ + { + "bbox": [ + 440, + 323, + 534, + 419 + ], + "spans": [ + { + "bbox": [ + 440, + 323, + 534, + 419 + ], + "type": "image", + "image_path": "6aefb98a22232c890df094f89cd1b64eeb7491df40ec494e3281c082c1137599.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 421, + 512, + 430 + ], + "lines": [ + { + "bbox": [ + 462, + 421, + 512, + 430 + ], + "spans": [ + { + "bbox": [ + 462, + 421, + 512, + 430 + ], + "type": "text", + "content": "(e) Edit texture" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 488, + 192, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 488, + 192, + 500 + ], + "spans": [ + { + "bbox": [ + 143, + 488, + 192, + 500 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 504, + 289, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 289, + 708 + ], + "type": "text", + "content": "A light stage uses a series of calibrated cameras and lights to capture a subject's facial appearance under varying illumination and viewpoint. This captured information is crucial for facial reconstruction and relighting. Unfortunately, light stages are often inaccessible: they are expensive and require significant technical expertise for construction and operation. In this paper, we present SunStage: a lightweight alternative to a light stage that captures comparable data using only a smartphone camera and the sun. Our method only requires the user to capture a selfie video outdoors, rotating in place, and uses the varying angles between the sun and the face as guidance in joint reconstruction of facial geometry, reflectance, camera pose, and lighting parameters. Despite the in-the-wild un-calibrated setting, our approach is able to reconstruct detailed facial appearance and geometry, enabling compelling effects such as relighting, novel view synthesis, and reflectance editing." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 488, + 386, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 386, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 386, + 500 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 509, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 605 + ], + "type": "text", + "content": "A light stage [11] acquires the shape and material properties of a face in high detail using a series of images captured under synchronized cameras and lights. This captured information can be used to synthesize novel images of the subject under arbitrary lighting conditions or from arbitrary viewpoints. This process enables a number of visual effects, such as creating digital replicas of actors that can be used in movies [1] or high-quality postproduction relighting [46]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "In many cases, however, it is often infeasible to get access to a light stage for capturing a particular subject, because light stages are not easy to find: they are expensive and require significant technical expertise (often teams of people) to build and operate. In these cases, hope is not lost — one can turn to methods that are trained on light stage data, with the intention of generalizing to new subjects. These methods do not require the subject to be captured by a light stage but instead use a machine learning" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20792" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 191 + ], + "type": "text", + "content": "model trained on a collection of previously acquired light stage captures to enable the same applications as a light stage, but from only one or several images of a new subject [6, 25, 30, 38, 40, 50, 52]. Unfortunately, these methods have difficulty faithfully reproducing and editing the appearance of new subjects, as they lack much of the signal necessary to resolve the ambiguities of single-view reconstruction, i.e., a single image of a face can be reasonably explained by different combinations of geometry, illumination, and reflectance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 288, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 288, + 455 + ], + "type": "text", + "content": "In this paper, we propose an intermediate solution — one that allows for personalized, high-quality capture of a given subject, but without the need for expensive, calibrated capture equipment. Our method, which we dub SunStage, uses only a handheld smartphone camera and the sun to simulate a minimalist light stage, enabling the reconstruction of individually-tailored geometry and reflectance without specialized equipment. Our capture setup only requires the user to hold the camera at arm's length and rotate in place, allowing the face to be observed under varying angles of incident sunlight, which causes specular highlights to move and shadows to swing across the face. This provides strong signals for the reconstruction of facial geometry and spatially-varying reflectance properties. The reconstructed face and scene parameters estimated by our system can be used to realistically render the subject in new, unseen lighting conditions — even with complex details like self-occluding cast shadows, which are typically missing in purely image-based relighting techniques, i.e., those that do not explicitly model geometry. In addition to relighting, we also show applications in view synthesis, correcting facial perspective distortion, and editing skin reflectance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 456, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 541 + ], + "type": "text", + "content": "Our contributions include: (1) a novel capture technique for personalized facial scanning without custom equipment, (2) a system for optimization and disentanglement of scene parameters (geometry, materials, lighting, and camera poses) from an unaligned, handheld video, and (3) multiple portrait editing applications that produce photorealistic results, using as input only a single selfie video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 552, + 136, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 136, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 136, + 564 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Face modeling. Extensive research has been devoted to the modeling of human faces, leading to various 3D morphable models (3DMMs) [2, 3, 5-7, 9, 10, 25, 31-33, 43]. These models are parametric (maybe in the form of neural networks [33]), allowing one to express variations compactly with a vector. They also encode strong priors learned from real scans. The groundbreaking face 3DMM is that of Blanz and Vetter [3] containing models for shape, expression, and appearance (the Phong model). Also influential is the FLAME model [25] that uses vertex-based Linear Blend Skinning (LBS). FLAME is described by a mapping from" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "shape, pose, and expression vectors to a list of vertices. We refer the reader to the survey by Egger et al. [13] for different face morphable models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 108, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 204 + ], + "type": "text", + "content": "Such parametric face models provide a low-dimensional space for optimization or learning algorithms. DECA [14] uses the FLAME model to estimate detailed facial geometry (and albedo) from single images, by predicting additional displacement maps and adding them to the estimated FLAME models. More recently, NextFace [12] employs the 3DMM geometry and albedo priors to learn an albedo residual that captures more facial details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 204, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 252 + ], + "type": "text", + "content": "Without modeling 3D face geometry, researchers have also achieved photorealistic synthesis of portrait images using generative models and large-scale high-quality image datasets [22, 23]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "content": "Light stage capture. The light stage as described in Debevec et al., achieves impressive portrait reconstruction and relighting by capturing a series of images of the face under varying illumination [11]. Subsequent work made this process faster, more efficient, and explored different types of illuminants [15, 16, 28]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 405 + ], + "type": "text", + "content": "Given that a light stage is not always accessible, a number of methods have been proposed to achieve similar outputs from a single (or few) input portrait images [18,19,29,30,40,41,48-50]. These method rely on a dataset of light stage captures or synthetic examples as training data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 406, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 610 + ], + "type": "text", + "content": "Our setup can be thought of as a \"minimalist light stage\" formed by just the sun and a rotating camera, without requiring the high construction and maintenance costs of building a light stage. This parameterization of a sun and skylight model has been shown to be effective in photometric stereo [17,21] and scene factorization [27,42]. In a similar spirit, Calian et al. [8] focus on lighting estimation using faces as \"light probes\". Sengupta et al. propose to circumvent the need for a complicated light stage by recording the facial appearance responses to varying contents displayed on a desk monitor, and then perform portrait relighting [37]. Sevastopolsky et al. also attempt to simplify the capture setup from a light stage to a mobile phone camera with a co-located flash [39]. Unlike our work, which is physically-based, their approaches use neural rendering, and therefore have less direct control over lighting, material, and scene parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 621, + 370, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 370, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 370, + 634 + ], + "type": "text", + "content": "3. Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "content": "Our method targets accurate reconstruction of scene lighting, subject geometry, and material properties from a handheld video sequence of a person rotating in place under the sun. Given a selfie video, we take a test-time optimization approach that uses the information from all frames of the video to solve for a physical model of the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20793" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 206, + 284 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 206, + 284 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 206, + 284 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 206, + 284 + ], + "type": "image", + "image_path": "3c538ae1aca93b227dc391b75f76279fb153275fca6f86a09cbc7ad677a06251.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 305, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 547, + 361 + ], + "type": "text", + "content": "Figure 2. Overview. Our method jointly reconstructs geometry, skin reflectance, lighting, and camera pose from a selfie video sequence of a person rotating under the sun. Our system begins by extracting supervisory information from the video sequence: facial landmarks, foreground alpha mattes, and camera orientations. These are used to supervise the optimization of a collection of scene parameters (full list in Sec. 5.2) used in a physically-based renderer. The rendered output is an image consisting of diffuse, specular, and ambient light contributions. After optimization, the solved scene parameters can be used for a number of editing applications, shown in Sec. 7." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 69, + 362, + 274 + ], + "blocks": [ + { + "bbox": [ + 212, + 69, + 362, + 274 + ], + "lines": [ + { + "bbox": [ + 212, + 69, + 362, + 274 + ], + "spans": [ + { + "bbox": [ + 212, + 69, + 362, + 274 + ], + "type": "image", + "image_path": "74643e55364fe4de4bc3d155f6e3eae252ad6fc92e38065d2e9321af88206377.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 366, + 70, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 366, + 70, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 366, + 70, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 366, + 70, + 545, + 297 + ], + "type": "image", + "image_path": "1070b43376eb454b7e7766e07c1e5befa01db106d541ebed13ad6e99df8ab4cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": "scene: the geometry and material properties of the face, scene lighting, and camera parameters (Fig. 2). This physical model consists of a base face shape parameterized by a low-dimensional deformable model " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "X^b" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": ", a displacement map " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": ", a reflectance model with diffuse " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "R^d" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": " and specular components " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "R^s" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": ", scene lighting " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "L_i" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": ", and a perspective camera " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": ". These components are explained in detail in Sec. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 288, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 546 + ], + "type": "text", + "content": "After this model has been recovered, we can modify the scene and the subject parameters to re-render images. We show several editing applications in Sec. 7: editing skin reflectance, relighting with arbitrary environment map, improving harsh lighting conditions (by softening shadows and adding fill lights), and adjusting camera parameters to change viewpoint or manipulate perspective effects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 559, + 127, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 127, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 127, + 572 + ], + "type": "text", + "content": "4. Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 288, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 288, + 654 + ], + "type": "text", + "content": "Given an input video, our system reconstructs the parameters of a physical model: i.e., geometry and reflectance of the subject, the scene lighting parameters, and the camera parameters. In this section, we detail all of these parameters and describe the rendering process that turns these parameters into an image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": "Geometry. We denote " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": " as the full mesh of the subject for frame " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": ", composed of a per-frame coarse mesh " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "X_{j}^{b}(\\beta, \\theta_{j}, \\psi_{j})" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": " and a global displacement map " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": ". The coarse mesh " + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "inline_equation", + "content": "X_{j}^{b}" + }, + { + "bbox": [ + 46, + 663, + 288, + 715 + ], + "type": "text", + "content": " is a FLAME deformable face model [25]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": "defined by global shape code " + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": ", per-frame pose code " + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": ", and per-frame expression code " + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "inline_equation", + "content": "X_{j}^{b}" + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": " also contains per-vertex UV coordinates, which maintain correspondence across variations in " + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 304, + 376, + 546, + 449 + ], + "type": "text", + "content": ". As such, we model all our global (per-subject) spatially varying parameters in UV space, and sample values per-fragment when rasterizing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": "The displacement map " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " is used to model fine details like wrinkles that cannot be represented by " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "X_{j}^{b}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": ". We displace the coarse geometry by " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " at rasterization time, by sampling a displacement value per-fragment and displacing each fragment along the surface normal " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " of the coarse mesh " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "X_{j}^{b}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": ". After displacement, the updated fragment positions are used to compute a new surface normal " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "N_{j}^{\\prime}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "inline_equation", + "content": "X_{j}^{b}" + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "text", + "content": " is optimized per-frame, since it accounts for subtle (and unavoidable) variations in expression and head pose during the capture, which are modeled by " + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "text", + "content": ", on the other hand, is optimized in UV-space (i.e., globally per-subject), since the deformations it encodes are invariant to the changes in expression or pose. Formally, the final geometry " + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 304, + 534, + 545, + 619 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 629, + 545, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 629, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 356, + 629, + 545, + 644 + ], + "type": "interline_equation", + "content": "X _ {j} = X _ {j} ^ {b} \\left(\\beta , \\theta_ {j}, \\psi_ {j}\\right) + \\Delta X \\odot N _ {j} \\tag {1}", + "image_path": "75946ff78f5c735bfd9c43f14b107f710064f8217407e04b0c6890ed35faf93f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 654, + 447, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 447, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 447, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 654, + 447, + 667 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 654, + 447, + 667 + ], + "type": "text", + "content": " is the Hadamard product." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": "Reflectance. We model the skin reflectance, denoted as " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "R(x,\\omega_i,\\omega_o) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " is a 3D point on the face geometry " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " is the incoming light direction, and " + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\omega_o" + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20794" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "outgoing direction, using a diffuse and a specular component: " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "R = R^{d} + R^{s}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "type": "text", + "content": "The diffuse component " + }, + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "type": "inline_equation", + "content": "R^d (x,\\omega_i)\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "type": "text", + "content": " is a Lambertian reflectance model consisting of an albedo map, " + }, + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 95, + 287, + 144 + ], + "type": "text", + "content": ", which we optimize as a per-subject UV-space image. For the skin's specular component, we use the Blinn-Phong model [4]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 159, + 287, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 159, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 75, + 159, + 287, + 182 + ], + "type": "interline_equation", + "content": "R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right) = k _ {s} \\frac {s + 2}{2 \\pi} \\left(h \\left(\\omega_ {i}, \\omega_ {o}\\right) \\cdot n (x)\\right) ^ {s} \\tag {2}", + "image_path": "c3e47a2e1f7327ff5307f831e3fe40eb8bc0da1b8e1f74618add2d1a3f0490ca.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "inline_equation", + "content": "h(\\omega_i, \\omega_o) = \\text{normalize}(\\omega_i + \\omega_o)" + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "content": " is the half vector, " + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "inline_equation", + "content": "k_s" + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "content": " is the specular intensity, " + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "content": " is the specular exponent, and " + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "inline_equation", + "content": "(s + 2) / (2\\pi)" + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "content": " is the normalization term for the reflection lobe to integrate to 1. Following [47], we segment the UV-space map into 10 segmented specular reflectance clusters. We then optimize for a spatially-varying pair of values " + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "inline_equation", + "content": "(s, k_s)" + }, + { + "bbox": [ + 46, + 187, + 287, + 270 + ], + "type": "text", + "content": " per-cluster, enabling varying shininess across the face." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 271, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 343 + ], + "type": "text", + "content": "While Blinn-Phong does not model many complex effects such as subsurface scattering, our experiments with other models for facial reflectance, such as microfacet models [44], show no significant quality improvements, and often introduce unstable training. More analysis is provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": "Lighting. We use a sun-sky model to represent lighting as the sum of an \"ambient\" environment map and the sun: " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "L_{i}(x,\\omega_{i}) = L_{i}^{\\mathrm{amb}}(\\omega_{i}) + L_{i}^{\\mathrm{sun}}(\\omega_{i})" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": ". Note neither " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "L_{i}^{\\mathrm{amb}}(\\omega_{i})" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": " nor " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "L_{i}^{\\mathrm{sun}}(\\omega_{i})" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": " depends on the 3D point " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": ", since we model both as directional lights. Optimization-wise, our lighting parameters consist of a " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "16 \\times 32 \\times 3" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": " environment map for ambient lighting, the sun direction " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sun}} \\in S^{3}" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": ", and the scalar sun intensity " + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{sun}}" + }, + { + "bbox": [ + 46, + 352, + 287, + 473 + ], + "type": "text", + "content": ". We fix the sun color to white [1, 1, 1] in our lighting model to resolve the albedo-illumination ambiguity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 119, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 119, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 119, + 491 + ], + "type": "text", + "content": "4.1. Rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "text", + "content": "We calculate the outgoing radiance " + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "inline_equation", + "content": "L_{o}" + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "text", + "content": " at 3D location " + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "text", + "content": " as viewed from viewing direction " + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 46, + 497, + 287, + 521 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 539, + 287, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 539, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 54, + 539, + 287, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {o} (x, \\omega_ {o}) \\\\ = \\int_ {\\Omega} V \\left(x, \\omega_ {i}\\right) L _ {i} \\left(x, \\omega_ {i}\\right) \\odot R \\left(x, \\omega_ {i}, \\omega_ {o}\\right) \\left(\\omega_ {i} \\cdot n (x)\\right) d \\omega_ {i} (3) \\\\ = \\sum_ {\\omega_ {i}} V (x, \\omega_ {i}) \\left(L _ {i} ^ {\\mathrm {a m b}} (\\omega_ {i}) \\odot R ^ {d} (x, \\omega_ {i}) \\right. (4) \\\\ + L _ {i} ^ {\\text {s u n}} \\left(\\omega_ {i}\\right) \\odot R ^ {d} \\left(x, \\omega_ {i}\\right) + L _ {i} ^ {\\text {a m b}} \\left(\\omega_ {i}\\right) \\odot R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right) \\\\ \\left. + L _ {i} ^ {\\operatorname {s u n}} \\left(\\omega_ {i}\\right) \\odot R ^ {s} \\left(x, \\omega_ {i}, \\omega_ {o}\\right)\\right) \\left(\\omega_ {i} \\cdot n (x)\\right) \\triangle \\omega_ {i} (5) \\\\ \\end{array}", + "image_path": "20a8219dd0bbe314c9ef30e0e1cbc911fb43d2f92e2bd1eaf8bc7888a1fc1140.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "V(x,\\omega_{i})" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is the light visibility at " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{i}(x,\\omega_{i})" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is the incoming radiance reaching " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ". We ignore the specular reflection caused by the ambient lightning, i.e., " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{i}^{\\mathrm{amb}}(\\omega_{i})\\odot R^{s}(x,\\omega_{i},\\omega_{o})" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", since it is much weaker" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "than the specular reflection of the sun. In the next subsections, we will group the terms into a diffuse contribution " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{o}^{d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and a specular contribution " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{o}^{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{o} = L_{o}^{d} + L_{o}^{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". For the final rendered color value, we apply the Reinhard operator [35] and a gamma correction of " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\gamma = 2.2" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{o}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to convert from linear to sRGB space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 153, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 545, + 178 + ], + "type": "text", + "content": "Diffuse contribution. The diffuse contribution " + }, + { + "bbox": [ + 304, + 153, + 545, + 178 + ], + "type": "inline_equation", + "content": "L_{o}^{d}" + }, + { + "bbox": [ + 304, + 153, + 545, + 178 + ], + "type": "text", + "content": " is then given by only the diffuse terms of Equation 5:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 187, + 545, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 187, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 187, + 545, + 253 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {o} ^ {d} (x) = \\sum_ {\\omega_ {i}} L _ {i} ^ {\\mathrm {a m b}} (\\omega_ {i}) \\odot \\frac {a (x)}{\\pi} (\\omega_ {i} \\cdot n (x)) \\Delta \\omega_ {i} \\\\ + V (x, p ^ {\\text {s u n}}) k ^ {\\text {s u n}} [ 1, 1, 1 ] \\odot \\frac {a (x)}{\\pi} \\left(p ^ {\\text {s u n}} \\cdot n (x)\\right) \\Delta p ^ {\\text {s u n}} \\tag {6} \\\\ \\end{array}", + "image_path": "32bb9c997849eaeb40ef8cf5ae13434ba54f96fa8aebf4d616cf497b84ba0a17.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "a(x)" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": " is the albedo at point " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": " is the (optimized) sun intensity, and " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": " is the (optimized) sun direction. The sun is modeled as a directional light source, so the second summation can be simplified to a single term, only in the direction of " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": ". We additionally optimize for a high-dynamic-range (HDR) environment map " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{16 \\times 32 \\times 3}" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": ", from which values of " + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "inline_equation", + "content": "L_{i}^{\\mathrm{amb}}" + }, + { + "bbox": [ + 304, + 262, + 545, + 347 + ], + "type": "text", + "content": " are sampled." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 356, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 399 + ], + "type": "text", + "content": "Specular contribution. The specular contribution " + }, + { + "bbox": [ + 304, + 356, + 545, + 399 + ], + "type": "inline_equation", + "content": "L_{o}^{s}" + }, + { + "bbox": [ + 304, + 356, + 545, + 399 + ], + "type": "text", + "content": " at each pixel is given by only the specular term due to the sun in Equation 5 (recall that we ignore the specular ambient term due to its weak contribution):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 407, + 545, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 407, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 331, + 407, + 545, + 442 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {o} ^ {s} (x, \\omega_ {o}) = V (x, p ^ {\\mathrm {s u n}}) k ^ {\\mathrm {s u n}} [ 1, 1, 1 ] k _ {s} \\frac {s + 2}{2 \\pi} \\\\ \\left(h \\left(p ^ {\\text {s u n}}, \\omega_ {o}\\right) \\cdot n (x)\\right) ^ {s} \\left(p ^ {\\text {s u n}} \\cdot n (x)\\right) \\Delta p ^ {\\text {s u n}} \\tag {7} \\\\ \\end{array}", + "image_path": "76d0b1ffeaaeb9d0dc54b17ac27d6b7c17a17027a3322aad195a280089a85183.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "content": "where we have substituted Equation 2 and reduced the summation to just one term at " + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "content": " (since " + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "inline_equation", + "content": "L_{i}^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 449, + 545, + 474 + ], + "type": "text", + "content": " is 0 elsewhere)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "content": "Shadow map. In order to generate a map of self-occluded shadows, we perform two passes of rasterization: first, we render a " + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "content": "-buffer from a virtual orthographic camera aligned with the sun direction, " + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "content": ", and then, when rasterizing a given camera viewpoint, compare all fragment positions " + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{hit}}" + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "content": " to the light's " + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "content": "-buffer " + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{shadow}}" + }, + { + "bbox": [ + 304, + 483, + 545, + 591 + ], + "type": "text", + "content": ". To avoid precision issues and ensure smooth gradients for back-propagation, we implement a soft comparison as follows in generating shadow/visibility maps:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 601, + 545, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 601, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 323, + 601, + 545, + 614 + ], + "type": "interline_equation", + "content": "V \\left(x, \\omega_ {i}\\right) = 1 - \\text {s i g m o i d} \\left(k \\left(d _ {\\text {h i t}} - d _ {\\text {s h a d o w}} \\times b\\right)\\right) \\tag {8}", + "image_path": "016ec7e1e86edeb391e1d7276e34cd87241626c43bbe09cae6aa2447ead3613c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": " is the falloff slope, and " + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": " the tolerance. We use " + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "inline_equation", + "content": "k = 800" + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "inline_equation", + "content": "b = 1.0015" + }, + { + "bbox": [ + 304, + 622, + 545, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 657, + 389, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 389, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 389, + 670 + ], + "type": "text", + "content": "5. Optimization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "The described physical model contains a large number of parameters to be optimized, controlling scene elements like lighting, geometry, pose, and texture. Unfortunately," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20795" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "naively optimizing all these parameters from scratch does not result in an optimal solution, since the final observed appearance of the face can often be explained variously through changes to geometry, material properties, lighting or camera parameters, making optimization severely under-constrained and ambiguous. Therefore, we adopt a two-stage optimization approach, through which parameters are gradually enabled. In this section, we describe this process and the relevant losses that guide optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 190, + 153, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 153, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 153, + 204 + ], + "type": "text", + "content": "5.1. Coarse alignment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "text", + "content": "Our system begins by using an off-the-shelf network (DECA [14]) to generate, for each input image, a set of shape parameters " + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "text", + "content": ", pose parameters " + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "text", + "content": ", and expression parameters " + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 46, + 209, + 287, + 365 + ], + "type": "text", + "content": " of a FLAME face model [25], as well as the relative pose parameters of the virtual camera observing the 3D face. Unfortunately, as with many other single-image facial geometry estimators, DECA assumes an orthographic projection model and therefore cannot accurately recover geometry for our selfie capture sequences, which contain heavy perspective effects (Figure 6). Without a good initialization for geometry, optimization of lighting and material properties seldom converges to an optimal solution due to the heavily ambiguous nature of our optimization problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": "To circumvent this issue, we employ a first stage of optimization where we only optimize for the parameters of a perspective camera (with a known focal length, extracted from input metadata) and the face geometry parameters " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta_j, \\psi_j)" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ". As initialization for this optimization process, we use the predicted DECA values for each frame's pose " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\theta_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": " and expression " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\psi_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ", but set all frames to the average predicted shape " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathrm{avg}} = \\frac{1}{N} \\sum_j \\beta_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ", since the identity remains constant across all frames. To convert DECA's orthographic camera to a perspective camera, we additionally optimize for an unknown scale " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "T_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ", which are initialized to empirically chosen values " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "S = 2.6e4" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "T_j = (0, 0, 1.5e5)" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ". During optimization, the face shape " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": " are shared across all frames, while camera pose " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "T_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\psi_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": ", and pose " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\theta_j" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": " are optimized per-frame. Note that DECA controls the relative orientation of the camera and the face by varying the pose code " + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 366, + 289, + 640 + ], + "type": "text", + "content": " instead of the camera rotation. We adopt this formulation and keep the camera orientation fixed relative to the face. The global orientation of the camera at each frame (and therefore the face) is extracted from the capture video, either through a structure-from-motion system or IMU measurements commonly available on a smartphone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "We use two losses to guide this optimization: a mask loss " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{mask}}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and a landmark loss " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{lmk}}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ". The FLAME model includes 3D facial landmark points, corresponding to the standard 68-point facial landmarks set [36] used in facial tracking. Our landmark loss minimizes the L1 distance between the 2D projection of these 3D landmarks (into the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "input camera viewpoint) and 2D landmarks estimated from the input frame by a 2D landmark detector HRNets [45]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "text", + "content": "The facial landmarks provide a stong constraint on facial feature alignment, but are sparse, and therefore cannot constrain the overall shape or boundary of the mesh. To supplement it, we include a silhouette loss " + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "text", + "content": ", which penalizes the L2 difference between the rasterized mask of the mesh " + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{sil}}" + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "text", + "content": " and the semantic segmentation mask " + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 96, + 547, + 203 + ], + "type": "text", + "content": " of the input image, using an off-the-shelf semantic segmentation network [26] trained to segment humans in portrait photographs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 204, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 239 + ], + "type": "text", + "content": "The final pose loss is then: " + }, + { + "bbox": [ + 304, + 204, + 547, + 239 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pose}} = L_{\\mathrm{mask}} + L_{\\mathrm{lmk}}" + }, + { + "bbox": [ + 304, + 204, + 547, + 239 + ], + "type": "text", + "content": ", optimized using an ADAM optimizer [24]. See supplemental for optimization parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 246, + 448, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 448, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 448, + 258 + ], + "type": "text", + "content": "5.2. Photometric optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": "Once the 3D model and camera parameters are approximately aligned, we proceed to the second stage of optimization, in which we optimize the precise facial geometry, lighting, and reflectance properties. All the parameters optimized in the first stage (Section 5.1) remain as free variables. In total, the parameters optimized during this stage include: Lighting parameters: (1) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global sun direction, (2) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global environment map, (3) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{sun}}" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global sun intensity, Facial geometry parameters: (4) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global FLAME shape code, (5) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\psi_j" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the per-frame expression code, (6) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\theta_j" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the per-frame pose code, (7) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global deformation map, Material properties: (8) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "k_s" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global, spatially-varying specular intensity, (9) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global, spatially-varying specular roughness, (10) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global, spatially-varying surface albedo, Camera pose parameters: (11) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "T_j" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the per-frame perspective camera translation, and (12) " + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 264, + 545, + 467 + ], + "type": "text", + "content": ", the global scene scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "type": "text", + "content": "During optimization, we randomly select a frame " + }, + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "type": "text", + "content": ", render the face using a differentiable rasterizer [34] and the equations described in Section 4 to get the rendered image " + }, + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 468, + 545, + 539 + ], + "type": "text", + "content": ". In addition to the previously defined landmark and mask losses, we include L2 and VGG [20] photometric losses, comparing the original and reconstructed images:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 360, + 545, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 545, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 360, + 545, + 545, + 559 + ], + "type": "interline_equation", + "content": "L _ {\\text {p h o t o}} = \\left| \\left| \\hat {I} _ {j} \\cdot I _ {\\text {s i l}} - I _ {j} \\cdot I _ {\\text {m a s k}} \\right| \\right| _ {2} \\tag {9}", + "image_path": "b8b4adf22e169888c5d696804842343d38af8f27e95fbce94b251826fae4632a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "type": "text", + "content": "We also include an L2 regularization " + }, + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "type": "inline_equation", + "content": "L_{E}" + }, + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "type": "text", + "content": " and L2-smoothness regularization " + }, + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "type": "inline_equation", + "content": "L_{Es}" + }, + { + "bbox": [ + 304, + 564, + 545, + 635 + ], + "type": "text", + "content": " on the reconstructed environment map, to encourage the majority of the lighting to be explained by direct sunlight and to aid in disentangle-ment of the sun and ambient lighting. The total optimized loss becomes:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 643, + 545, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 643, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 328, + 643, + 545, + 672 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\lambda_ {\\text {m a s k}} L _ {\\text {m a s k}} + \\lambda_ {\\text {l m k}} L _ {\\text {l m k}} + \\lambda_ {E} L _ {E} + \\lambda_ {E \\mathrm {s}} L _ {E \\mathrm {s}} \\\\ + \\lambda_ {\\mathrm {V G G}} L _ {\\mathrm {V G G}} + \\lambda_ {\\text {p h o t o}} L _ {\\text {p h o t o}} \\tag {10} \\\\ \\end{array}", + "image_path": "caba4598d90ae32faf96d7e59437abedca11c10851dd71f13ac2489f02a17e0c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mask}}, \\lambda_{\\mathrm{lmk}} = 0.05, \\lambda_{\\mathrm{VGG}} = 0.005, \\lambda_{E} = 0.01, \\lambda_{Es}, \\lambda_{photo} = 1" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". Additional optimization details are provided in the supplemental materials." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20796" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 111, + 191 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 111, + 191 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 111, + 191 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 111, + 191 + ], + "type": "image", + "image_path": "52f5d3506df2da247be34ce6af111df2b650be54aec27ea643eda1ba2a6e61d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 193, + 104, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 193, + 104, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 104, + 203 + ], + "type": "text", + "content": "(a) DECA [14]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 70, + 173, + 191 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 173, + 191 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 173, + 191 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 173, + 191 + ], + "type": "image", + "image_path": "b88f041e2d99bc9832219eb121f46731d55fb38da3a7582a7c6c43d89a550965.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 193, + 164, + 203 + ], + "lines": [ + { + "bbox": [ + 121, + 193, + 164, + 203 + ], + "spans": [ + { + "bbox": [ + 121, + 193, + 164, + 203 + ], + "type": "text", + "content": "(b) DPR [52]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 235, + 191 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 235, + 191 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 235, + 191 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 235, + 191 + ], + "type": "image", + "image_path": "0b20f9c935a049ac92da3097ee7dc4c6e57419465fb34578d57cd5f3a10deeea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 193, + 228, + 203 + ], + "lines": [ + { + "bbox": [ + 180, + 193, + 228, + 203 + ], + "spans": [ + { + "bbox": [ + 180, + 193, + 228, + 203 + ], + "type": "text", + "content": "(c) GCFR [18]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 236, + 70, + 296, + 191 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 296, + 191 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 296, + 191 + ], + "type": "image", + "image_path": "31ca6ac7b4b2f6315f2ab8c53c295c821ec0d9980c4f34f83059d409ca250085.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 193, + 285, + 203 + ], + "lines": [ + { + "bbox": [ + 247, + 193, + 285, + 203 + ], + "spans": [ + { + "bbox": [ + 247, + 193, + 285, + 203 + ], + "type": "text", + "content": "(d) TR [30]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 297, + 70, + 359, + 191 + ], + "blocks": [ + { + "bbox": [ + 297, + 70, + 359, + 191 + ], + "lines": [ + { + "bbox": [ + 297, + 70, + 359, + 191 + ], + "spans": [ + { + "bbox": [ + 297, + 70, + 359, + 191 + ], + "type": "image", + "image_path": "d089802bbdb35ea642b96694c0e65a03343638319389c7936a9d5debda5d5f4d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 193, + 357, + 203 + ], + "lines": [ + { + "bbox": [ + 299, + 193, + 357, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 193, + 357, + 203 + ], + "type": "text", + "content": "(e) NextFace [12]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 420, + 191 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 420, + 191 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 420, + 191 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 420, + 191 + ], + "type": "image", + "image_path": "389ed353915131fb1aa231f1c2d610b58e8e3feb2c7080222657736a3d4a0dc4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 193, + 410, + 203 + ], + "lines": [ + { + "bbox": [ + 369, + 193, + 410, + 203 + ], + "spans": [ + { + "bbox": [ + 369, + 193, + 410, + 203 + ], + "type": "text", + "content": "(f) NLT [51]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 421, + 70, + 482, + 191 + ], + "blocks": [ + { + "bbox": [ + 421, + 70, + 482, + 191 + ], + "lines": [ + { + "bbox": [ + 421, + 70, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 421, + 70, + 482, + 191 + ], + "type": "image", + "image_path": "96d06ce9832da61fc03bc2768eb050b47a1bc29c66894022180b1ba67e22a1a8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 193, + 466, + 203 + ], + "lines": [ + { + "bbox": [ + 438, + 193, + 466, + 203 + ], + "spans": [ + { + "bbox": [ + 438, + 193, + 466, + 203 + ], + "type": "text", + "content": "(g) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 483, + 70, + 544, + 191 + ], + "blocks": [ + { + "bbox": [ + 483, + 70, + 544, + 191 + ], + "lines": [ + { + "bbox": [ + 483, + 70, + 544, + 191 + ], + "spans": [ + { + "bbox": [ + 483, + 70, + 544, + 191 + ], + "type": "image", + "image_path": "0481f3271eaac39ce508bacd29ddae5422c60f9fbf655c4ecb1d0583ce0f336d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 193, + 525, + 202 + ], + "lines": [ + { + "bbox": [ + 502, + 193, + 525, + 202 + ], + "spans": [ + { + "bbox": [ + 502, + 193, + 525, + 202 + ], + "type": "text", + "content": "(h) GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 52, + 267, + 132, + 347 + ], + "blocks": [ + { + "bbox": [ + 46, + 213, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 246 + ], + "type": "text", + "content": "Figure 3. Qualitative: Relighting. A comparison of our method at rendering a new (unseen) lighting environment (h). Our method is able to realistically synthesize the novel lighting condition, including cast shadows and specularity, and nearly matches the (unseen) target reference image. See supplement for additional details on experimental setup and analysis of results." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 267, + 132, + 347 + ], + "lines": [ + { + "bbox": [ + 52, + 267, + 132, + 347 + ], + "spans": [ + { + "bbox": [ + 52, + 267, + 132, + 347 + ], + "type": "image", + "image_path": "63f889f69e65178df554d6b6108e518b25678feffe0ccc03b4168367f1687eea.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 52, + 348, + 132, + 429 + ], + "blocks": [ + { + "bbox": [ + 52, + 348, + 132, + 429 + ], + "lines": [ + { + "bbox": [ + 52, + 348, + 132, + 429 + ], + "spans": [ + { + "bbox": [ + 52, + 348, + 132, + 429 + ], + "type": "image", + "image_path": "70e7609460fb80139be943c0d1a9362611ce45a7b84481341cfbffc5f0136549.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 430, + 119, + 439 + ], + "lines": [ + { + "bbox": [ + 66, + 430, + 119, + 439 + ], + "spans": [ + { + "bbox": [ + 66, + 430, + 119, + 439 + ], + "type": "text", + "content": "(a) Nearest view" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 134, + 267, + 214, + 347 + ], + "blocks": [ + { + "bbox": [ + 134, + 267, + 214, + 347 + ], + "lines": [ + { + "bbox": [ + 134, + 267, + 214, + 347 + ], + "spans": [ + { + "bbox": [ + 134, + 267, + 214, + 347 + ], + "type": "image", + "image_path": "6d77e6ffd4dac729bcfd27b8c6405737fd3a4b7884c8323dae15b42d6bf79f8a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 135, + 348, + 214, + 429 + ], + "blocks": [ + { + "bbox": [ + 135, + 348, + 214, + 429 + ], + "lines": [ + { + "bbox": [ + 135, + 348, + 214, + 429 + ], + "spans": [ + { + "bbox": [ + 135, + 348, + 214, + 429 + ], + "type": "image", + "image_path": "82e79896fb021f99081dd578fd2625fafc214ce25057acfc1b4723ba93a479ac.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 430, + 199, + 440 + ], + "lines": [ + { + "bbox": [ + 149, + 430, + 199, + 440 + ], + "spans": [ + { + "bbox": [ + 149, + 430, + 199, + 440 + ], + "type": "text", + "content": "(b) DECA [14]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 216, + 267, + 296, + 347 + ], + "blocks": [ + { + "bbox": [ + 216, + 267, + 296, + 347 + ], + "lines": [ + { + "bbox": [ + 216, + 267, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 216, + 267, + 296, + 347 + ], + "type": "image", + "image_path": "2f28d4fe4a2a58d714b7a518425b1dbcec9bbd209968c5734c4a31b160408397.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 216, + 348, + 296, + 429 + ], + "blocks": [ + { + "bbox": [ + 216, + 348, + 296, + 429 + ], + "lines": [ + { + "bbox": [ + 216, + 348, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 216, + 348, + 296, + 429 + ], + "type": "image", + "image_path": "6e987664c92a9483dd4062b5b292a9ec18f9bc0d72d4f23098f2a144c31c6720.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 430, + 285, + 440 + ], + "lines": [ + { + "bbox": [ + 227, + 430, + 285, + 440 + ], + "spans": [ + { + "bbox": [ + 227, + 430, + 285, + 440 + ], + "type": "text", + "content": "(c) NextFace [12]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 298, + 267, + 378, + 347 + ], + "blocks": [ + { + "bbox": [ + 298, + 267, + 378, + 347 + ], + "lines": [ + { + "bbox": [ + 298, + 267, + 378, + 347 + ], + "spans": [ + { + "bbox": [ + 298, + 267, + 378, + 347 + ], + "type": "image", + "image_path": "ca138ccb57cb9ba789ab336cce4533901e3033782cf28b11ab3f33ee005bfbd9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 298, + 348, + 378, + 429 + ], + "blocks": [ + { + "bbox": [ + 298, + 348, + 378, + 429 + ], + "lines": [ + { + "bbox": [ + 298, + 348, + 378, + 429 + ], + "spans": [ + { + "bbox": [ + 298, + 348, + 378, + 429 + ], + "type": "image", + "image_path": "4b3bc4291722f72be538d4a74369afc093b808eff9488db8778dc3184ae9b98c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 430, + 359, + 440 + ], + "lines": [ + { + "bbox": [ + 316, + 430, + 359, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 359, + 440 + ], + "type": "text", + "content": "(d) NLT [51]" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 379, + 267, + 460, + 347 + ], + "blocks": [ + { + "bbox": [ + 379, + 267, + 460, + 347 + ], + "lines": [ + { + "bbox": [ + 379, + 267, + 460, + 347 + ], + "spans": [ + { + "bbox": [ + 379, + 267, + 460, + 347 + ], + "type": "image", + "image_path": "44c665776405e58528ae23d3d3f1772ce139be37341cdf658ff3c2313fa970dc.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 379, + 348, + 459, + 429 + ], + "blocks": [ + { + "bbox": [ + 379, + 348, + 459, + 429 + ], + "lines": [ + { + "bbox": [ + 379, + 348, + 459, + 429 + ], + "spans": [ + { + "bbox": [ + 379, + 348, + 459, + 429 + ], + "type": "image", + "image_path": "62114643901ca4b5168b3d1a5f302c9b661457ad37214e278b78d936931780bf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 430, + 433, + 439 + ], + "lines": [ + { + "bbox": [ + 406, + 430, + 433, + 439 + ], + "spans": [ + { + "bbox": [ + 406, + 430, + 433, + 439 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 449, + 544, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 544, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 544, + 472 + ], + "type": "text", + "content": "Figure 4. Qualitative: view synthesis. A comparison of our method at the task of generating an image from an unseen viewpoint (f), having only seen a limited collection of input viewpoints. See supplement for more details on experimental setup and analysis of results." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 462, + 267, + 541, + 347 + ], + "blocks": [ + { + "bbox": [ + 462, + 267, + 541, + 347 + ], + "lines": [ + { + "bbox": [ + 462, + 267, + 541, + 347 + ], + "spans": [ + { + "bbox": [ + 462, + 267, + 541, + 347 + ], + "type": "image", + "image_path": "2d45431b527777126b725e35b985f6217c22f1f8fee6f15cbf90992fcab6b65c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 462, + 348, + 541, + 429 + ], + "blocks": [ + { + "bbox": [ + 462, + 348, + 541, + 429 + ], + "lines": [ + { + "bbox": [ + 462, + 348, + 541, + 429 + ], + "spans": [ + { + "bbox": [ + 462, + 348, + 541, + 429 + ], + "type": "image", + "image_path": "9410ea937733907d2ef32f96d3cac579537b70d2644566c9c5c992a047668017.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 430, + 512, + 439 + ], + "lines": [ + { + "bbox": [ + 490, + 430, + 512, + 439 + ], + "spans": [ + { + "bbox": [ + 490, + 430, + 512, + 439 + ], + "type": "text", + "content": "(f) GT" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 502, + 118, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 118, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 118, + 514 + ], + "type": "text", + "content": "6. Evaluation" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 522, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 559 + ], + "type": "text", + "content": "In this section, we detail quantitative and qualitative experiments comparing our approach with state-of-the-art methods and ablated variants of our method." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 569, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 286, + 687 + ], + "type": "text", + "content": "Baseline comparisons. We evaluate our method on the tasks of novel-view synthesis and relighting. For novel-view synthesis, we compare our method with DECA [14], Neural Light Transport (NLT) [51], and NextFace [12]. For relighting, we compare with DECA, NLT, NextFace, GCFR [18], image-based methods Deep Single Image Portrait Relighting (DPR) [52], Total Relighting (TR) [30] and NVPR [49]. Additional comparisons and details on the experimental setups are provided in the supplemental materials." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "We present qualitative comparisons for relighting in Figure 3 and novel view synthesis in Figure 4. Quantitative" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 503, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 694 + ], + "type": "text", + "content": "comparisons on these images are provided in Table 1. These testing images consist of (1) a multi-view capture of the face, in which the subject remains still and the camera is moved to novel viewpoints in the same environment as the original capture, and (2) front-facing sequences in novel environment lighting and unseen sun positions. All testing images are not seen during training of our method, NLT or NextFace. The results shown in Figures 3 and 4 as well as Table 1 clearly demonstrate that our method outperforms all the baselines at both relighting and view synthesis. Single-image methods (DECA, GCFR, DPR, TR) can generalize to other subjects, but fail to recover more faithful and physically accurate facial details. Comparison with multi-image methods (NLT, NextFace) demonstrates that SunStage is a better reconstruction system. Additional analysis of the comparisons is provided in the supplemental materials." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20797" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 120, + 141 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 120, + 141 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 120, + 141 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 120, + 141 + ], + "type": "image", + "image_path": "32d160419810920fd0346182a686c69c47e9250e107026d7fdd6949f3a73e12e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 70, + 190, + 141 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 190, + 141 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 190, + 141 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 190, + 141 + ], + "type": "image", + "image_path": "c9c33fb6b2de3789d68733d03be92948914101c742df20584b1d85dca464b8d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 191, + 70, + 260, + 141 + ], + "blocks": [ + { + "bbox": [ + 191, + 70, + 260, + 141 + ], + "lines": [ + { + "bbox": [ + 191, + 70, + 260, + 141 + ], + "spans": [ + { + "bbox": [ + 191, + 70, + 260, + 141 + ], + "type": "image", + "image_path": "630c6c2a58f07c1a036f57b37fcda6ab6d1e63dff504fca8a40e17bf34b0e098.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 70, + 330, + 141 + ], + "blocks": [ + { + "bbox": [ + 260, + 70, + 330, + 141 + ], + "lines": [ + { + "bbox": [ + 260, + 70, + 330, + 141 + ], + "spans": [ + { + "bbox": [ + 260, + 70, + 330, + 141 + ], + "type": "image", + "image_path": "baa9a6d1b3736c55f5b66c91b53bf2f5ec88c8939fef3fa5e537798ebb65ec8e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 400, + 141 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 400, + 141 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 400, + 141 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 400, + 141 + ], + "type": "image", + "image_path": "ae520b4158da99aaab95b336a3cd11a13efdd3785beba4a352a232fad5d6fd4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 401, + 70, + 471, + 141 + ], + "blocks": [ + { + "bbox": [ + 401, + 70, + 471, + 141 + ], + "lines": [ + { + "bbox": [ + 401, + 70, + 471, + 141 + ], + "spans": [ + { + "bbox": [ + 401, + 70, + 471, + 141 + ], + "type": "image", + "image_path": "c97590b23c88a48e4b282520d1b12ce392391de9ff745e51becfcc4529d98033.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 472, + 70, + 543, + 141 + ], + "blocks": [ + { + "bbox": [ + 472, + 70, + 543, + 141 + ], + "lines": [ + { + "bbox": [ + 472, + 70, + 543, + 141 + ], + "spans": [ + { + "bbox": [ + 472, + 70, + 543, + 141 + ], + "type": "image", + "image_path": "2f75e2f17afc0f648eeb2b37139e7832cbcea571a55da79424c9c225d3dd2707.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 144, + 120, + 215 + ], + "blocks": [ + { + "bbox": [ + 49, + 144, + 120, + 215 + ], + "lines": [ + { + "bbox": [ + 49, + 144, + 120, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 120, + 215 + ], + "type": "image", + "image_path": "6645e13675557b3ed95647b24a162194f1325aa3bdfc2284aec35b0bb80e8b15.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 121, + 144, + 190, + 215 + ], + "blocks": [ + { + "bbox": [ + 121, + 144, + 190, + 215 + ], + "lines": [ + { + "bbox": [ + 121, + 144, + 190, + 215 + ], + "spans": [ + { + "bbox": [ + 121, + 144, + 190, + 215 + ], + "type": "image", + "image_path": "c603800cd4d7a07ba485a548889b9690205252c37ff0b9a6688f76970e5fd16b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 144, + 260, + 215 + ], + "blocks": [ + { + "bbox": [ + 190, + 144, + 260, + 215 + ], + "lines": [ + { + "bbox": [ + 190, + 144, + 260, + 215 + ], + "spans": [ + { + "bbox": [ + 190, + 144, + 260, + 215 + ], + "type": "image", + "image_path": "52fc257f218259558efcf5c9436cd8d81749b66ce813448d5c1c6fd185e09655.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 260, + 144, + 330, + 215 + ], + "blocks": [ + { + "bbox": [ + 260, + 144, + 330, + 215 + ], + "lines": [ + { + "bbox": [ + 260, + 144, + 330, + 215 + ], + "spans": [ + { + "bbox": [ + 260, + 144, + 330, + 215 + ], + "type": "image", + "image_path": "e430f445978ebbb89f326ef48b1830637e61341664463b1c0d335232870dce32.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 144, + 400, + 215 + ], + "blocks": [ + { + "bbox": [ + 331, + 144, + 400, + 215 + ], + "lines": [ + { + "bbox": [ + 331, + 144, + 400, + 215 + ], + "spans": [ + { + "bbox": [ + 331, + 144, + 400, + 215 + ], + "type": "image", + "image_path": "4482c11b62f0c41a2fb5db699013d87b007d426a0ff496208a3120573f8b1565.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 401, + 144, + 471, + 215 + ], + "blocks": [ + { + "bbox": [ + 401, + 144, + 471, + 215 + ], + "lines": [ + { + "bbox": [ + 401, + 144, + 471, + 215 + ], + "spans": [ + { + "bbox": [ + 401, + 144, + 471, + 215 + ], + "type": "image", + "image_path": "67af2371cabb1c1e86ca548744ee11cbaaad7045eb5e9059977aa271f00e8ac4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 472, + 144, + 543, + 215 + ], + "blocks": [ + { + "bbox": [ + 472, + 144, + 543, + 215 + ], + "lines": [ + { + "bbox": [ + 472, + 144, + 543, + 215 + ], + "spans": [ + { + "bbox": [ + 472, + 144, + 543, + 215 + ], + "type": "image", + "image_path": "71e3128a28d7dbfbcfe8cd2eda9293d94b6c2d259eb7fdd84eed344a3ebf91d8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 50, + 217, + 120, + 288 + ], + "blocks": [ + { + "bbox": [ + 50, + 217, + 120, + 288 + ], + "lines": [ + { + "bbox": [ + 50, + 217, + 120, + 288 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 120, + 288 + ], + "type": "image", + "image_path": "710bd2c7abdd4e57051c963654097daf5125ca9f115239f71b1ec7324e6dbf99.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 289, + 101, + 299 + ], + "lines": [ + { + "bbox": [ + 69, + 289, + 101, + 299 + ], + "spans": [ + { + "bbox": [ + 69, + 289, + 101, + 299 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 121, + 217, + 190, + 288 + ], + "blocks": [ + { + "bbox": [ + 121, + 217, + 190, + 288 + ], + "lines": [ + { + "bbox": [ + 121, + 217, + 190, + 288 + ], + "spans": [ + { + "bbox": [ + 121, + 217, + 190, + 288 + ], + "type": "image", + "image_path": "eb4bd385f83cf045e620d1bd44b404475b62a9e3cf0c513ca46a0caf49fe7517.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 289, + 170, + 299 + ], + "lines": [ + { + "bbox": [ + 141, + 289, + 170, + 299 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 170, + 299 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 190, + 217, + 260, + 288 + ], + "blocks": [ + { + "bbox": [ + 190, + 217, + 260, + 288 + ], + "lines": [ + { + "bbox": [ + 190, + 217, + 260, + 288 + ], + "spans": [ + { + "bbox": [ + 190, + 217, + 260, + 288 + ], + "type": "image", + "image_path": "300c7b77082e13e199cd371d843c25a2833c3500874c1da3f2314b869970b8e9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 289, + 243, + 300 + ], + "lines": [ + { + "bbox": [ + 207, + 289, + 243, + 300 + ], + "spans": [ + { + "bbox": [ + 207, + 289, + 243, + 300 + ], + "type": "text", + "content": "Specular" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 260, + 217, + 330, + 288 + ], + "blocks": [ + { + "bbox": [ + 260, + 217, + 330, + 288 + ], + "lines": [ + { + "bbox": [ + 260, + 217, + 330, + 288 + ], + "spans": [ + { + "bbox": [ + 260, + 217, + 330, + 288 + ], + "type": "image", + "image_path": "40b0cb98907b7d1f5800d9e60b2f79245ac342a1d9c4d2108b0e4027a53435e7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 289, + 312, + 299 + ], + "lines": [ + { + "bbox": [ + 277, + 289, + 312, + 299 + ], + "spans": [ + { + "bbox": [ + 277, + 289, + 312, + 299 + ], + "type": "text", + "content": "Ambient" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 331, + 217, + 400, + 288 + ], + "blocks": [ + { + "bbox": [ + 331, + 217, + 400, + 288 + ], + "lines": [ + { + "bbox": [ + 331, + 217, + 400, + 288 + ], + "spans": [ + { + "bbox": [ + 331, + 217, + 400, + 288 + ], + "type": "image", + "image_path": "9d4b93c6d8dbe5e719ed206d278823135cdb553fe1fa470a3de5b30acdbba09d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 289, + 381, + 299 + ], + "lines": [ + { + "bbox": [ + 350, + 289, + 381, + 299 + ], + "spans": [ + { + "bbox": [ + 350, + 289, + 381, + 299 + ], + "type": "text", + "content": "Diffuse" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 401, + 217, + 471, + 288 + ], + "blocks": [ + { + "bbox": [ + 401, + 217, + 471, + 288 + ], + "lines": [ + { + "bbox": [ + 401, + 217, + 471, + 288 + ], + "spans": [ + { + "bbox": [ + 401, + 217, + 471, + 288 + ], + "type": "image", + "image_path": "122249523a9bed97ec54557dcfc649ec8362bbe4251358b2ec65c4d27f4eca41.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 289, + 466, + 299 + ], + "lines": [ + { + "bbox": [ + 406, + 289, + 466, + 299 + ], + "spans": [ + { + "bbox": [ + 406, + 289, + 466, + 299 + ], + "type": "text", + "content": "Reconstruction" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 472, + 217, + 543, + 288 + ], + "blocks": [ + { + "bbox": [ + 472, + 217, + 543, + 288 + ], + "lines": [ + { + "bbox": [ + 472, + 217, + 543, + 288 + ], + "spans": [ + { + "bbox": [ + 472, + 217, + 543, + 288 + ], + "type": "image", + "image_path": "41828a35e4cf3fffd69f21bf2bf3f0fbfde939fa18d24f0856c5d75a2eb496f2.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 289, + 534, + 299 + ], + "lines": [ + { + "bbox": [ + 480, + 289, + 534, + 299 + ], + "spans": [ + { + "bbox": [ + 480, + 289, + 534, + 299 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 47, + 364, + 288, + 509 + ], + "blocks": [ + { + "bbox": [ + 47, + 364, + 288, + 509 + ], + "lines": [ + { + "bbox": [ + 47, + 364, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 288, + 509 + ], + "type": "table", + "html": "
RelightingNovel view synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DECA [14]16.410.690.2516.640.660.29
GCFR [18]16.970.700.20---
DPR [52]19.030.720.19---
NLT [51]20.150.750.1822.270.790.15
Total Relighting [30]20.240.790.16---
NextFace [12]22.980.760.1522.550.750.15
Ours23.640.830.1025.280.840.09
Ours w/o coarse17.830.660.2319.650.700.17
Ours w/o SV ks, s21.310.770.1321.940.770.12
Ours w/o Lmask, Llmk16.460.610.3018.830.680.20
Ours w/o Lmask20.130.750.1520.540.740.15
Ours w/o opt. (β, θi, φi)18.670.690.1918.280.660.19
Ours w/o soft shadow21.460.770.1322.050.770.12
Ours w/o ΔX21.160.750.1521.800.750.14
", + "image_path": "017ee2a711f1e8edee4def443bed49e6b2fbd136554ab234ef65fdf854686861.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 549 + ], + "type": "text", + "content": "Table 1. Quantitative comparison. Comparison of our method on the tasks of novel view synthesis and relighting. See Section 6 for a description of the ablated variants." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 691 + ], + "type": "text", + "content": "Disentanglement. In Figure 5, we demonstrate how Sun-Stage decomposes the appearance of a portrait photograph into different components: specular, diffuse, and ambient. We also visualize the surface normal, albedo, and other intermediate representations to show that our method is able to effectively recover a physically plausible reconstruction of the real world and disentangle the different components that contribute to the final appearance. We further validate the quality of the reconstructed geometry and materials in the supplementary material." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Ablation studies. In addition to our comparisons with" + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 308, + 364, + 367, + 423 + ], + "blocks": [ + { + "bbox": [ + 46, + 311, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 547, + 355 + ], + "type": "text", + "content": "Figure 5. Decomposition. We show all the components which comprise our final rendered image to demonstrate that our method not only closely recreates the ground truth image (reproducing realistic highlights and shadows), but also performs a meaningful decomposition of lighting components and facial geometry. Note that our reconstructed surface normals include high frequency details specific to each subject, like wrinkles and birthmarks, which are used in computation of the shadows and specular reflections." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 364, + 367, + 423 + ], + "lines": [ + { + "bbox": [ + 308, + 364, + 367, + 423 + ], + "spans": [ + { + "bbox": [ + 308, + 364, + 367, + 423 + ], + "type": "image", + "image_path": "3c7469f6efaaed23fba602c343523ebed64979bcf417288432441edd0a5a4c02.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 424, + 363, + 434 + ], + "lines": [ + { + "bbox": [ + 312, + 424, + 363, + 434 + ], + "spans": [ + { + "bbox": [ + 312, + 424, + 363, + 434 + ], + "type": "text", + "content": "(a) Input image" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 444, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 305, + 444, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 545, + 533 + ], + "type": "text", + "content": "Figure 6. Perspective. DECA's assumption of an orthographic camera is broken by the strong perspective effects in selfies, causing poor alignment (b) with input images (a). Our first stage of optimization (c) (Sec. 5.1) improves alignment by solving for the parameters of a perspective camera and refined shape parameters. In the second stage we additionally optimize for a displacement map " + }, + { + "bbox": [ + 305, + 444, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 305, + 444, + 545, + 533 + ], + "type": "text", + "content": " to produce our final shape with finer geometric details like wrinkles (d). Red line added to highlight alignment with (a)." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 367, + 364, + 425, + 422 + ], + "blocks": [ + { + "bbox": [ + 367, + 364, + 425, + 422 + ], + "lines": [ + { + "bbox": [ + 367, + 364, + 425, + 422 + ], + "spans": [ + { + "bbox": [ + 367, + 364, + 425, + 422 + ], + "type": "image", + "image_path": "befc0e0b5d14c0df09622bb32efca79c3a5ff72feb3f75a6dcfb51c51faea23d.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 424, + 422, + 434 + ], + "lines": [ + { + "bbox": [ + 372, + 424, + 422, + 434 + ], + "spans": [ + { + "bbox": [ + 372, + 424, + 422, + 434 + ], + "type": "text", + "content": "(b) DECA [14]" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 427, + 364, + 485, + 422 + ], + "blocks": [ + { + "bbox": [ + 427, + 364, + 485, + 422 + ], + "lines": [ + { + "bbox": [ + 427, + 364, + 485, + 422 + ], + "spans": [ + { + "bbox": [ + 427, + 364, + 485, + 422 + ], + "type": "image", + "image_path": "1f5bfc03898bc74c2796b4542eeecf4c8492ee6d1fc0f35ca042892d7146b116.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 424, + 473, + 434 + ], + "lines": [ + { + "bbox": [ + 438, + 424, + 473, + 434 + ], + "spans": [ + { + "bbox": [ + 438, + 424, + 473, + 434 + ], + "type": "text", + "content": "(c) Stage 1" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 486, + 364, + 545, + 422 + ], + "blocks": [ + { + "bbox": [ + 486, + 364, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 486, + 364, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 486, + 364, + 545, + 422 + ], + "type": "image", + "image_path": "644d3fc7d0c0a59ca6ef58f75db49803ef155778a8726cfc755d6e12590d0ecc.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 424, + 529, + 434 + ], + "lines": [ + { + "bbox": [ + 501, + 424, + 529, + 434 + ], + "spans": [ + { + "bbox": [ + 501, + 424, + 529, + 434 + ], + "type": "text", + "content": "(d) Final" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "content": "the state-of-the-art baselines, we also compare with ablated variants of our own method. In particular, we include seven such experiments in Table 1: our method (1) without the initial first stage of coarse geometric alignment, i.e., directly optimizing both geometric and photometric parameters from the start, (2) without the spatially varying specular parameters, instead using a single global scalar " + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "k_{s}" + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "content": ", (3) without the geometric alignment losses " + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{mask}}" + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{lmk}}" + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "content": ", (4) without just " + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{mask}}" + }, + { + "bbox": [ + 305, + 570, + 546, + 713 + ], + "type": "text", + "content": ", (5) without shape optimization, i.e., keeping the initial shape code predicted by DECA, (6) without soft shadow computation, i.e., using a hard z-buffer comparison to compute a shadow map instead of our soft" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20798" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 108, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 108, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 108, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 108, + 186 + ], + "type": "image", + "image_path": "85fba16c5fbf0821758be6dd5c407c10ffcb4c91ae636bd503da3c288b4c0d3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 188, + 93, + 198 + ], + "lines": [ + { + "bbox": [ + 64, + 188, + 93, + 198 + ], + "spans": [ + { + "bbox": [ + 64, + 188, + 93, + 198 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 70, + 167, + 186 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 167, + 186 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 167, + 186 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 167, + 186 + ], + "type": "image", + "image_path": "ab834c038346f3deae18213cc163346f51647dbbddf4f6b5f2d122140f5031c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 188, + 165, + 197 + ], + "lines": [ + { + "bbox": [ + 111, + 188, + 165, + 197 + ], + "spans": [ + { + "bbox": [ + 111, + 188, + 165, + 197 + ], + "type": "text", + "content": "(b) Soft shadows" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 226, + 186 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 226, + 186 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 226, + 186 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 226, + 186 + ], + "type": "image", + "image_path": "7ca8f30c33e86b4b60a3039b729a3d85ad3353e57db7f33f571f3ca0b8094206.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 188, + 225, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 188, + 225, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 188, + 225, + 198 + ], + "type": "text", + "content": "(c) (b) + fill light" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 286, + 186 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 286, + 186 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 286, + 186 + ], + "type": "image", + "image_path": "71e20c4ee897b13f12516cc0833afe8536be8551fa1446a62ace94dd31049b55.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 188, + 282, + 198 + ], + "lines": [ + { + "bbox": [ + 232, + 188, + 282, + 198 + ], + "spans": [ + { + "bbox": [ + 232, + 188, + 282, + 198 + ], + "type": "text", + "content": "(d) Swap scene" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 207, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 287, + 262 + ], + "type": "text", + "content": "Figure 7. Adjusting lighting parameters. We can adjust the recovered scene parameters to improve the lighting conditions in an input image (a) by softening the harsh shadows cast by the nose (b), adding a fill light to brighten the shaded region (b), or replacing the environment altogether (d)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 282, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 287, + 319 + ], + "type": "text", + "content": "comparison operator in Equation 8, and (7) without the displacement map " + }, + { + "bbox": [ + 47, + 282, + 287, + 319 + ], + "type": "inline_equation", + "content": "\\Delta X" + }, + { + "bbox": [ + 47, + 282, + 287, + 319 + ], + "type": "text", + "content": ". Visual results for each of these variants are provided in the supplemental material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 328, + 127, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 127, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 127, + 342 + ], + "type": "text", + "content": "7. Applications" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 358, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 586 + ], + "type": "text", + "content": "Relighting. We demonstrate two types of relighting applications: (1) lighting modification and (2) lighting replacement. Practical lighting modification is common in portrait photography when the lighting conditions are not ideal, e.g., when direct sunlight casts undesirable harsh shadows with high contrast. A common practice is to make the light source larger and more diffuse by using a scrim or bounce card. In Fig. 7b, we show that by virtually increasing the size and spreading the energy of our reconstructed lighting source (i.e., the sun), we are able to soften the shadows and re-render a more visually pleasing face. Another approach to reducing the effects of harsh shadows is adding local fill lights, which reduces the contrast between the lit and shaded regions (Fig. 7c). Alternatively, fill lights can also be used for artistic purposes, to create dramatic lighting effects (Fig. 1b). Finally, replacing the scene lighting with that of a novel environment (Fig. 7d) is a necessary step in realistically inserting a captured subject into a virtual scene, which is useful for visual effects and VR applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 596, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 287, + 679 + ], + "type": "text", + "content": "View Synthesis. In Figure 8, we show that our reconstructed 3D model of the face can be used to synthesize new views by manipulating the viewpoint of the camera. We can also change other camera parameters, such as the focal length, to reduce the perspective effects on the face, which is often desirable for selfie images that contain significant facial distortion due to perspective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Skin Reflectance Editing. We are also able to edit the reflectance components of the subject. As shown in Figure 1e," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 366, + 186 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 366, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 366, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 366, + 186 + ], + "type": "image", + "image_path": "66dae6f19b9f6d50ceae33285e157f6ccd096bab7dd701c10c305dee0453db39.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 188, + 361, + 198 + ], + "lines": [ + { + "bbox": [ + 314, + 188, + 361, + 198 + ], + "spans": [ + { + "bbox": [ + 314, + 188, + 361, + 198 + ], + "type": "text", + "content": "(a) Short focal" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 368, + 70, + 425, + 186 + ], + "blocks": [ + { + "bbox": [ + 368, + 70, + 425, + 186 + ], + "lines": [ + { + "bbox": [ + 368, + 70, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 368, + 70, + 425, + 186 + ], + "type": "image", + "image_path": "47b3313a562875c48a0277e625984bf02508ff9e849aca500d864a7cf8b7d358.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 188, + 421, + 198 + ], + "lines": [ + { + "bbox": [ + 373, + 188, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 373, + 188, + 421, + 198 + ], + "type": "text", + "content": "(b) Long focal" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 485, + 186 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 485, + 186 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 485, + 186 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 485, + 186 + ], + "type": "image", + "image_path": "ef8f4c0cf1ec2cb121e170c4c7d70c6832dfc8a2870d61cc84814ea46d4145f4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 188, + 478, + 198 + ], + "lines": [ + { + "bbox": [ + 429, + 188, + 478, + 198 + ], + "spans": [ + { + "bbox": [ + 429, + 188, + 478, + 198 + ], + "type": "text", + "content": "(c) Novel view" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 486, + 70, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 486, + 70, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 486, + 70, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 486, + 70, + 544, + 186 + ], + "type": "image", + "image_path": "e42e876402a4bdb2cd81e76b8c1ee78d50de1c3df27ad855407c9d136cf4fbba.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 188, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 488, + 188, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 488, + 188, + 543, + 198 + ], + "type": "text", + "content": "(d) Novel view 2" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 308, + 247, + 367, + 306 + ], + "blocks": [ + { + "bbox": [ + 306, + 207, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 306, + 207, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 545, + 240 + ], + "type": "text", + "content": "Figure 8. Changing camera parameters. We can change the recovered camera parameters to render novel views (c,d) or change the focal length (a,b)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 247, + 367, + 306 + ], + "lines": [ + { + "bbox": [ + 308, + 247, + 367, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 367, + 306 + ], + "type": "image", + "image_path": "32a964bbde0e982d6f0a192d8b8fe7a5c30b9df1962a21b0f2f699776cdc8fb5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 308, + 362, + 316 + ], + "lines": [ + { + "bbox": [ + 312, + 308, + 362, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 362, + 316 + ], + "type": "text", + "content": "(a) Less shinier" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 327, + 545, + 359 + ], + "lines": [ + { + "bbox": [ + 306, + 327, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 545, + 359 + ], + "type": "text", + "content": "Figure 9. Adjusting the specularity. We can change the the specular properties of the face, making the face less shinier (a, c) or more shinier (b, d)." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 368, + 247, + 426, + 306 + ], + "blocks": [ + { + "bbox": [ + 368, + 247, + 426, + 306 + ], + "lines": [ + { + "bbox": [ + 368, + 247, + 426, + 306 + ], + "spans": [ + { + "bbox": [ + 368, + 247, + 426, + 306 + ], + "type": "image", + "image_path": "371ce804727e1555a587e8cfd5407b473d8c005e19b8616e728a4864d328780f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 308, + 424, + 316 + ], + "lines": [ + { + "bbox": [ + 370, + 308, + 424, + 316 + ], + "spans": [ + { + "bbox": [ + 370, + 308, + 424, + 316 + ], + "type": "text", + "content": "(b) More shinier" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 427, + 247, + 485, + 306 + ], + "blocks": [ + { + "bbox": [ + 427, + 247, + 485, + 306 + ], + "lines": [ + { + "bbox": [ + 427, + 247, + 485, + 306 + ], + "spans": [ + { + "bbox": [ + 427, + 247, + 485, + 306 + ], + "type": "image", + "image_path": "1aaa84b6d71119d71f3d7fa59871cff6d9f12c3385a996aef39ac16683409d02.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 308, + 481, + 316 + ], + "lines": [ + { + "bbox": [ + 430, + 308, + 481, + 316 + ], + "spans": [ + { + "bbox": [ + 430, + 308, + 481, + 316 + ], + "type": "text", + "content": "(c) Less shinier" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 486, + 247, + 544, + 306 + ], + "blocks": [ + { + "bbox": [ + 486, + 247, + 544, + 306 + ], + "lines": [ + { + "bbox": [ + 486, + 247, + 544, + 306 + ], + "spans": [ + { + "bbox": [ + 486, + 247, + 544, + 306 + ], + "type": "image", + "image_path": "066a1cbed01d26fdf80fdf8dd23baed3abadc50ca6826d4a48bfa1f7a3f6481f.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 308, + 542, + 316 + ], + "lines": [ + { + "bbox": [ + 488, + 308, + 542, + 316 + ], + "spans": [ + { + "bbox": [ + 488, + 308, + 542, + 316 + ], + "type": "text", + "content": "(d) More shinier" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 373, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 433 + ], + "type": "text", + "content": "we can adjust the optimized albedo to add freckles, stickers, or other textures that realistically interact with reflections, shadows, and other elements of scene lighting, or we can adjust the specular properties of the face, making the face more or less shinier, as shown in Figure 9." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 444, + 462, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 462, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 462, + 456 + ], + "type": "text", + "content": "8. Conclusion and Discussions" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "type": "text", + "content": "In this paper, we propose SunStage, a lightweight and practical facial capture, rendering, and editing system that can serve as a minimalist light stage. With a video of an individual rotating in-place under the sun, our system reconstructs a physical model of the subject and the scene lighting, which enables us to relight the subject with realistic reflections and cast shadows. Our system allows arbitrary lighting and reflectance control in the reconstructed physical space, which can be rendered to produce photo-realistic results. We demonstrate several applications such as editing skin reflectance, relighting, and view synthesis." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Limitations. Our system inherits the limitations of morphable face models and is unable to model hair, teeth, or clothing geometry, beyond slight deformations. Additionally, certain regions which are seen under constant shadow or specular reflection (and therefore have no cues on reflectance or albedo) are sometimes unable to be decomposed accurately into separate reflectance and lighting components. Visualization and further discussions of the system's limitations are provided in the supplemental material." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20799" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Oleg Alexander, Mike Rogers, William Lambeth, Jen-Yuan Chiang, Wan-Chun Ma, Chuan-Chang Wang, and Paul Debevec. The digital emily project: Achieving a photorealistic digital actor. IEEE Computer Graphics and Applications, 30(4):20-31, 2010. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Sai Bi, Stephen Lombardi, Shunsuke Saito, Tomas Simon, Shih-En Wei, Kevyn McPhail, Ravi Ramamoorthi, Yaser Sheikh, and Jason Saragih. Deep relightable appearance models for animatable faces. ACM Trans. Graph. (Proc. SIGGRAPH), 40(4), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "text", + "content": "[3] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "text", + "content": "[4] James F Blinn. Models of light reflection for computer synthesized pictures. In Proceedings of the 4th annual conference on Computer graphics and interactive techniques, pages 192-198, 1977. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "text", + "content": "[5] Timo Bolkart and Stefanie Wuhrer. A groupwise multilinear correspondence optimization for 3d faces. In Proceedings of the IEEE international conference on computer vision, pages 3604-3612, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "text", + "content": "[6] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5543-5552, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 442 + ], + "type": "text", + "content": "[7] Alan Brunton, Timo Bolkart, and Stefanie Wuhrer. Multilinear wavelets: A statistical shape space for human faces. In European Conference on Computer Vision, pages 297-312. Springer, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 487 + ], + "type": "text", + "content": "[8] Dan A Calian, Jean-François Lalonde, Paulo Gotardo, Tomas Simon, Iain Matthews, and Kenny Mitchell. From faces to outdoor light probes. In Computer Graphics Forum, volume 37, pages 51-61. Wiley Online Library, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 287, + 533 + ], + "type": "text", + "content": "[9] Chen Cao, Yanlin Weng, Shun Zhou, Yiying Tong, and Kun Zhou. Facewarehouse: A 3d facial expression database for visual computing. IEEE Transactions on Visualization and Computer Graphics, 20(3):413-425, 2013. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[10] Hang Dai, Nick Pears, William AP Smith, and Christian Duncan. A 3d morphable model of craniofacial shape and texture variation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3085-3093, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "text", + "content": "[11] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "text", + "content": "[12] Abdallah Dib, Cedric Thebault, Junghyun Ahn, Philippe-Henri Gosselin, Christian Theobalt, and Louis Chevallier. Towards high fidelity monocular face reconstruction with rich reflectance using self-supervised learning and ray tracing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12819-12829, 2021. 2, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[13] Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39(5):1-38, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[14] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[15] Graham Fyffe and Paul Debevec. Single-shot reflectance measurement from polarized color gradient illumination. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[16] Abhijeet Ghosh, Graham Fyffe, Borom Tunwattanapong, Jay Busch, Xueming Yu, and Paul Debevec. Multiview face capture using polarized spherical gradient illumination. In Proceedings of the 2011 SIGGRAPH Asia Conference, pages 1-10, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "type": "text", + "content": "[17] Yannick Hold-Geoffroy, Jinsong Zhang, Paulo F U Gotardo, and Jean-François Lalonde. " + }, + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 307, + 275, + 545, + 308 + ], + "type": "text", + "content": "-hour outdoor photometric stereo. In International Conference on 3D Vision, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "text", + "content": "[18] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4217-4226, 2022. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14719-14728, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 421, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 465 + ], + "type": "text", + "content": "[20] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694-711. Springer, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "text", + "content": "[21] Jiyoung Jung, Joon-Young Lee, and In So Kweon. One-day outdoor photometric stereo via skylight estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4521–4529, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[25] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[26] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with tempo-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20800" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "ral guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "text", + "content": "[27] Andrew Liu, Shiry Ginosar, Tinghui Zhou, Alexei A Efros, and Noah Snavely. Learning to factorize and relight a city. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16, pages 544–561. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 229 + ], + "type": "text", + "content": "[28] Abhinitra Meka, Christian Haene, Rohit Pandey, Michael Zollhöfer, Sean Fanello, Graham Fyffe, Adarsh Kowdle, Xueming Yu, Jay Busch, Jason Dourgarian, et al. Deep reflectance fields: high-quality facial reflectance field inference from color gradient illumination. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "text", + "content": "[29] Thomas Nestmeyer, Jean-François Lalonde, Iain Matthews, and Andreas Lehrmann. Learning physics-guided face relighting under directional light. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5124-5133, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 342 + ], + "type": "text", + "content": "[30] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 387 + ], + "type": "text", + "content": "[31] Frédéric Pighin, Jamie Hecker, Dani Lischinski, Richard Szeliski, and David H Salesin. Synthesizing realistic facial expressions from photographs. In ACM SIGGRAPH 2006 Courses, pages 19-es. 2006. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "text", + "content": "[32] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 509 + ], + "type": "text", + "content": "[33] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 704-720, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 510, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 555 + ], + "type": "text", + "content": "[34] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv preprint arXiv:2007.08501, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 556, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 578 + ], + "type": "text", + "content": "[35] Erik Reinhard. Parameter estimation for photographic tone reproduction. Journal of graphics tools, 7(1):45-51, 2002. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 635 + ], + "type": "text", + "content": "[36] Christos Sagonas, Georgios Tzimiropoulos, Stefanos Zafeiriou, and Maja Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 397-403, 2013. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[37] Soumyadip Sengupta, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M Seitz. A light stage on every desk. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2420-2429, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[38] Soumyadip Sengupta, Angjoo Kanazawa, Carlos D Castillo, and David W Jacobs. Sfsnet: Learning shape, reflectance and illuminance of faces in the wild'. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE conference on computer vision and pattern recognition, pages 6296-6305, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[39] Artem Sevastopolsky, Savva Ignatiev, Gonzalo Ferrer, Evgeny Burnaev, and Victor Lempitsky. Relightable 3d head portraits from a smartphone video. arXiv preprint arXiv:2012.09963, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 196 + ], + "type": "text", + "content": "[40] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul E Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):79-1, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 252 + ], + "type": "text", + "content": "[41] Tiancheng Sun, Zexiang Xu, Xiuming Zhang, Sean Fanello, Christoph Rhemann, Paul Debevec, Yun-Ta Tsai, Jonathan T Barron, and Ravi Ramamoorthi. Light stage superresolution: Continuous high-frequency relighting. ACM Transactions on Graphics (TOG), 39(6):1-12, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 297 + ], + "type": "text", + "content": "[42] Kalyan Sunkavalli, Wojciech Matusik, Hanspeter Pfister, and Szymon Rusinkiewicz. Factored time-lapse video. ACM Transactions on Graphics (Proc. SIGGRAPH), 26(3), Aug. 2007. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 547, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 332 + ], + "type": "text", + "content": "[43] Luan Tran and Xiaoming Liu. Nonlinear 3d face morphable model. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7346-7355, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 333, + 547, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 547, + 387 + ], + "type": "text", + "content": "[44] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. In Proceedings of the 18th Eurographics conference on Rendering Techniques, pages 195-206, 2007. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 547, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 444 + ], + "type": "text", + "content": "[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "type": "text", + "content": "[46] Andreas Wenger, Andrew Gardner, Chris Tchou, Jonas Unger, Tim Hawkins, and Paul Debevec. Performance relighting and reflectance transformation with timemultiplexed illumination. ACM Transactions on Graphics (TOG), 24(3):756-764, 2005. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 501, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 547, + 567 + ], + "type": "text", + "content": "[47] Tim Weyrich, Wojciech Matusik, Hanspeter Pfister, Bernd Bickel, Craig Donner, Chien Tu, Janet McAndless, Jinho Lee, Addy Ngan, Henrik Wann Jensen, et al. Analysis of human faces using a measurement-based skin reflectance model. ACM Transactions on Graphics (ToG), 25(3):1013-1024, 2006. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 568, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 547, + 612 + ], + "type": "text", + "content": "[48] Yu-Ying Yeh, Koki Nagano, Sameh Khamis, Jan Kautz, Ming-Yu Liu, and Ting-Chun Wang. Learning to relight portrait images via a virtual light stage and synthetic-to-real adaptation. ACM Transactions on Graphics (TOG), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 667 + ], + "type": "text", + "content": "[49] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "text", + "content": "[50] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20801" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[51] Xiuming Zhang, Sean Fanello, Yun-Ta Tsai, Tiancheng Sun, Tianfan Xue, Rohit Pandey, Sergio Orts-Escalano, Philip Davidson, Christoph Rhemann, Paul Debevec, et al. Neural light transport for relighting and view synthesis. ACM Transactions on Graphics (TOG), 40(1):1-17, 2021. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE International Conference on Computer Vision, pages 7194-7202, 2019. 2, 6, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20802" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_content_list.json b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e53837960b462de07c89b3cd753502d52a7392a1 --- /dev/null +++ b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_content_list.json @@ -0,0 +1,1549 @@ +[ + { + "type": "text", + "text": "Super-CLEVR: A Virtual Benchmark to Diagnose Domain Robustness in Visual Reasoning", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuowan Li $^{1}$ Xingrui Wang $^{2}$ Elias Stengel-Eskin $^{1}$ Adam Kortylewski $^{3,4}$ Wufei Ma $^{1}$ Benjamin Van Durme $^{1}$ Alan Yuille $^{1}$ $^{1}$ Johns Hopkins University $^{2}$ University of Southern California \n $^{3}$ Max Planck Institute for Informatics $^{4}$ University of Freiburg", + "bbox": [ + 155, + 202, + 823, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Question Answering (VQA) models often perform poorly on out-of-distribution data and struggle on domain generalization. Due to the multi-modal nature of this task, multiple factors of variation are intertwined, making generalization difficult to analyze. This motivates us to introduce a virtual benchmark, Super-CLEVR, where different factors in VQA domain shifts can be isolated in order that their effects can be studied independently. Four factors are considered: visual complexity, question redundancy, concept distribution and concept compositionality. With controllably generated data, Super-CLEVR enables us to test VQA methods in situations where the test data differs from the training data along each of these axes. We study four existing methods, including two neural symbolic methods NSCL [45] and NSVQA [59], and two non-symbolic methods FiLM [50] and mDETR [29]; and our proposed method, probabilistic NSVQA (P-NSVQA), which extends NSVQA with uncertainty reasoning. P-NSVQA outperforms other methods on three of the four domain shift factors. Our results suggest that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. The dataset and code are released at https://github.com/Lizw14/Super-CLEVR.", + "bbox": [ + 75, + 342, + 472, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual question answering (VQA) is a challenging task that assesses the reasoning ability of models to answer questions based on both visual and linguistic inputs. Current VQA methods are typically developed on standard benchmarks like VQAv2 [16] or GQA [25], with the implicit assumption that testing data comes from the same underlying distribution as training data. However, as has been widely studied in computer vision [15,36,51], algorithms trained on one domain often fail to generalize to other domains.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, having learned the distributional prior of training data, models often struggle on out-of-distribution tests. This has been studied in VQA from the perspective of domain transfer [8,57,62], dataset bias [2,11,48], counter-factual diagnosis [9,47], and out-of-distribution benchmarking [30].", + "bbox": [ + 496, + 310, + 890, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The multi-modal nature of VQA gives rise to multiple intertwined factors of variation, making domain shift an especially difficult problem to study. For example, [8] suggests that VQA domain shifts are a combination of differences in images, questions or answers; and [39] reveals a gap between synthetic and real VQA datasets by differences in the over-specification of questions and the underlying distribution of concepts. However, despite a wealth of research on domain generalization in VQA [3,26,57,62], there is no systematic analysis of the contributing factors in domain shifts.", + "bbox": [ + 496, + 386, + 892, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we introduce a virtual benchmark, SuperCLEVR, which enables us to test VQA algorithms in situations where the test data differs from the training data. We decompose the domain shift into a set of isolated contributing factors, so that their effects can be diagnosed independently. We study four factors: visual complexity, question redundancy, concept distribution, and concept compositionality. These are illustrated in Fig. 1 and described in Sec. 3.1. With controllable data generation using our SuperCLEVR virtual benchmark, we are able to isolate the different factors in VQA domain shifts so that their effects can be studied independently. Compared with the original CLEVR dataset [27], Super-CLEVR contains more complicated visual components and has better controllability over the domain shift factors. As shown in Fig. 1, the Super-CLEVR dataset contains images rendered from 3D graphical vehicle models in the UDA-Part dataset [40], paired with questions and answers automatically generated from templates. The objects and questions are sampled based on the specified underlying probability distribution, which can be controlled to produce distribution shifts in different factors.", + "bbox": [ + 496, + 537, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With Super-CLEVR, we diagnose the domain robustness of current VQA models. Four representative models are studied: for the classic two-stream feature fusing", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14963", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f0dce66fed1595f30b56d6ba88aa8155a34bdbd2d3fca4679910dd5f06bb79b.jpg", + "image_caption": [ + "Figure 1. We decompose VQA domain shifts into four contributing factors: visual complexity, question redundancy, concept distribution and concept compositionality. The domain shifts along each factor can be independently studied with the proposed Super-CLEVR dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 64, + 872, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "architecture, we choose FiLM [50]; for a large-scale pretrained model we take mDETR [29]; we use NSCL [45] and NSVQA [59] as representative neuro-symbolic methods. We observe that all these models suffer from domain shifts to varying degrees of sensitivity. We analyze each factor separately to examine the influence of different model designs. Specifically, we find that the step-by-step design of neural modular methods enhances their robustness to changes in question redundancy compared with non-modular ones; however, the non-modular models are more robust to visual complexity. Furthermore, thanks to its decomposed reasoning and perception, NSVQA is more robust to concept distribution shifts.", + "bbox": [ + 75, + 378, + 472, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While existing models suffer from domain shifts with different characteristics, we make a technical improvement over NSVQA which enables it to significantly outperform existing models on three of the four factors. In particular, we inject probabilities into the deterministic symbolic executor of NSVQA, empowering it to take into account the uncertainty of scene understanding. We name our model probabilistic NSVQA (P-NSVQA), and show that its performance improvement in both the in-domain and out-of-domain settings. With superior results of P-NSVQA, we suggest that disentangling reasoning from vision and language understanding, together with probabilistic uncertainty, gives a strong model that is robust to domain shifts.", + "bbox": [ + 75, + 579, + 468, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows. (1) We introduce the Super-CLEVR benchmark to diagnose VQA robustness along four different factors independently. This benchmark can also be used for part-based reasoning. (2) We enhance a neural-symbolic method by taking the uncertainty of visual understanding into account in reasoning. (3) We conduct detailed analysis of four existing methods, as well as our novel approach to study the influence of model designs on", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distinct robustness factors. We conclude that disentangled reasoning and perception plus explicit modeling of uncertainty leads to a more robust VQA model.", + "bbox": [ + 496, + 378, + 893, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 498, + 436, + 638, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual question answering (VQA). Popular VQA methods fall into three categories. Two-stream methods extract features for image and questions using CNN and LSTM respectively, then enable interaction between the two modalities with different feature fusing methods [4, 13, 24, 32, 37, 50, 60]. Neural symbolic methods, on the other hand, use a parse-then-execute pipeline where the question is parsed into a functional program, which is then executed on the image using neural modules [45, 59]. Recently, transformers-based models have achieved impressive performance on various vision-and-language tasks by pretraining on large scale dataset then finetuning for downstream tasks [29, 38, 44, 55, 63]. We choose FiLM [50], NSCL [45] and mDETR [29] as category representatives.", + "bbox": [ + 496, + 462, + 890, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VQA datasets. Datasets containing real images and human-written questions have been widely used to benchmark VQA models, e.g. VQA [5], VQAv2 [16], Visual 7w [65], VizWiz [18], Visual Genome [35], COCO QA [52], etc. However, subsequent work has revealed the strong prior and bias in those datasets which might be exploited by models to correctly predict the answers without reasoning [1, 16, 17, 30, 31, 47, 48]. Attempts to address this problem include better balancing datasets [2] and creating counterfactual examples [9, 11]. To assess a model's true reasoning ability, the CLEVR dataset [27] proposes to generate complex multi-step questions on synthetic images, which is then extended to various vision-and-language tasks [6,23,34,41,53,58,61]. The GQA dataset [25] extends CLEVR-style questions to real images. Our benchmark is", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14964", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distinct from existing ones because we introduce more complex visual scenes into CLEVR and provide controllability to study domain robustness on isolated factors.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain shift in VQA. Domain shift is a long-standing challenge in computer vision, explore in prior works in domain adaptation [14, 15, 22, 43] and domain generalization [36, 51]. Recent works have focused on domain shifts in VQA. [8, 57] improves model adaptation between datasets by feature learning. [62] analyze domain shifts between nine popular VQA datasets and proposes an unsupervised method to bridge the gaps. [39] generalize symbolic reasoning from synthetic to real dataset. [3] introduce a question-answer generation module that simulates the domain shifts. [26] propose a training scheme X-GGM to improve out-of-distribution generalization. [6] assess models generalization on the CLOSURE of linguistic components. In contrast to prior works we study each of the different domain shift factors independently with our virtual benchmark.", + "bbox": [ + 75, + 137, + 472, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Super-CLEVR", + "text_level": 1, + "bbox": [ + 76, + 380, + 225, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation: domain shift factors", + "text_level": 1, + "bbox": [ + 76, + 406, + 362, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual complexity. A major difference between different VQA datasets is visual complexity. For example, in the CLEVR dataset, objects are simple, atomic shapes while in real-world data, objects are more complex and have hierarchical parts. While hard to quantify, visual complexity is related to various factors, such as object variety, object size, background, texture, lighting, occlusion, view point, etc. In our work, we control visual complexity by introducing more challenging objects that can have distinct attributes associated with their parts, and by optionally pasting various textures onto objects. Examples of generated images with different complexity levels are shown in Fig. 1.", + "bbox": [ + 75, + 430, + 468, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Question redundancy. Question redundancy refers to the amount of over-specified or redundant information in the question, which can be in the form of either attributes or relationships. For example, in Fig. 1, \"what color is the large bus behind the cyan car\", large (attributes) and behind the cyan car (relationship) are redundant because there is only one bus in the image. As observed in linguistics and cognitive science [12, 33, 49, 54], human speakers may include over-specified information when identifying a target object, which has also been studied in referring expression generation [46]. For VQA, as analyzed in [39], a significant difference between synthetic and real datasets is that real questions contain some redundant information, which sometimes is a distraction leading to model prediction errors. Therefore, in this work, we generate questions with different redundancy levels and study the effect of question redundancy on model behaviors.", + "bbox": [ + 75, + 613, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept distribution. The distributions of concept, i.e. objects (e.g. car) and attributes (e.g. large), are distinct", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object with parts:", + "bbox": [ + 519, + 87, + 638, + 101 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d554702fdd1b286096780825b6cbc3e072b38b459d927da85bb575bbc36bf38.jpg", + "image_caption": [ + "car" + ], + "image_footnote": [], + "bbox": [ + 540, + 103, + 642, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3d881699209d1c12c4778b0f3a10a1c295f69af38f17380a51c815890ff0f426.jpg", + "image_caption": [ + "bus" + ], + "image_footnote": [], + "bbox": [ + 650, + 113, + 756, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc8dcd72b3ede28d4de876a2e2bbb94111f416073a9345cb1c8b56392cac2f5b.jpg", + "image_caption": [ + "airplane", + "Figure 2. Super-CLEVR contains 21 vehicle models belonging to 5 categories, with controllable attributes." + ], + "image_footnote": [], + "bbox": [ + 774, + 107, + 870, + 143 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f26aca1c5ab30f524deea69bf0a41e0175f46a4b775f5087d1f776d8a9a72b4a.jpg", + "image_caption": [ + "bike" + ], + "image_footnote": [], + "bbox": [ + 537, + 166, + 642, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/374244f3d17e2095ac2c499c015ac3be9dd4c5d41f8d1ea33660efebdb3ab56d.jpg", + "image_caption": [ + "motorcycle" + ], + "image_footnote": [], + "bbox": [ + 656, + 160, + 746, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Texture: dotted, checkered, stripped, none", + "bbox": [ + 517, + 229, + 787, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Color: green, gray, brown, yellow, red, purple, cyan, blue", + "bbox": [ + 519, + 243, + 870, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Size: large, small", + "bbox": [ + 519, + 257, + 625, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material: rubber, metal", + "bbox": [ + 519, + 271, + 668, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "across different VQA datasets. For example, while colors are well-balanced in CLEVR dataset, in the GQA dataset, the color distribution is long-tailed where \"white\" appears $>50$ times more frequently than \"gold\". Long-tailed distributions have been a challenge in many computer vision tasks [20, 28, 42, 56, 64]. In VQA, the long-tailed concept distribution not only hinders the learning of infrequent concepts due to few training samples, but also introduces strong biases and priors in the dataset that may mislead the models. For example, \"tennis\" is the correct answer to most questions with \"what sport is ...\" [2]. With strong priors in data, it is hard to assess the true reasoning capacity of current models. While previous works address this problem by carefully re-balancing datasets [2], in our work, we controllably vary the concept distribution in our dataset and study model robustness to concept distribution shifts.", + "bbox": [ + 496, + 332, + 890, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept compositionality. Concept compositionality refers to how different concepts (shapes, attributes) compose and co-occur with each other, e.g. roses are usually red while violets are usually blue [30]. Concept compositionality can be viewed as a conditional concept distribution in the context of other concepts. Shifts in concept compositionality impede the generalization of VQA models. For example, the model may fail to recognize a green banana because most bananas are yellow in the training data [7]. Previous works evaluate the out-of-distribution performance by collecting counterfactual testing examples [47]. In our work, we control the compositionality of shapes and colors with an intuitive motivation: if, for example, in the training data, bicycles are red and cars are blue, will the models be able to recognize blue bicycles and red cars in testing?", + "bbox": [ + 496, + 575, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dataset generation", + "text_level": 1, + "bbox": [ + 500, + 816, + 679, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Super-CLEVR follows a similar data generation pipeline as CLEVR, but with more complex visual components and better control of domain gap factors. We describe the generation procedure below.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14965", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objects with parts. To improve the visual complexity of CLEVR scenes, we replace the simple shapes (e.g., cube, sphere) in CLEVR dataset with vehicles from UDA-Part dataset [40]. There are 21 vehicle models, belonging to 5 categories: car, motorbike, aeroplane, bus, and bicycle. Each 3D model comes with part annotations, e.g., left front wheel or left right door for car. Examples for the vehicle models are shown in Fig. 2. We remove or merge small parts from the original annotations to avoid severe difficulty in visual understanding. The full object and parts list is in the supplementary material.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attributes. Besides the attributes in the original CLEVR dataset, i.e. color, material, size, we optionally add texture as an additional attribute to increase visual complexity. Note that in order to enable part-based questions, the attributes (color or material) of object parts can be different from that of the object. For example, a blue car can have a red wheel or a green door. In this case, the attribute of the holistic object refers to the attribute of its main body (e.g. the blue car has blue frame).", + "bbox": [ + 75, + 257, + 472, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene rendering. Following CLEVR, each scene contains 3 to 10 objects. The objects are placed onto the ground plane with random position and orientation. When placing the objects, we ensure that the objects do not overlap with each other and we avoid severe occlusion by thresholding the number of visible pixels for each object. Random jitters are added to lamp and camera positions. When rendering, we also save the ground-truth bounding boxes and segmentation masks for each of the objects and their parts, which are required when training some of the models.", + "bbox": [ + 75, + 395, + 472, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question generation. Super-CLEVR follows similar question generation pipeline in CLEVR, which instantiates question templates using the underlying reasoning program that can be operated on the scene graph. For example, the program select_shape truck) $\\rightarrow$ query_color (·) can be instantiated as question \"what is the color of the truck\". Therefore, redundancy level of questions can be controlled by removing or adding redundant reasoning steps in the underlying reasoning program.", + "bbox": [ + 75, + 547, + 472, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Controlling the dataset", + "text_level": 1, + "bbox": [ + 76, + 694, + 292, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To study domain generalization, we generate several variants of the dataset for each of the domain shift factors. The variants of the datasets serve as different data domains to test the model robustness. Here we describe the method for controllably generating the dataset variants.", + "bbox": [ + 75, + 718, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual complexity. We generate three variants of the dataset with different levels of visual complexity: easy, mid (middle) and hard. The only difference between the 3 versions is visual complexity: for the easy version, objects with different sizes, colors and materials are placed into the scene; for the middle version, we choose 3 parts on each object that are visible and randomly change their attributes;", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the hard version, we further add random textures to the objects and parts. An example of the 3 dataset versions can be found in Fig. 1. Note that the scene layout and the questions are shared, so that the influence of visual complexity can be isolated and studied independently.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question redundancy. Three variants of the dataset with different redundancy levels are generated: $rd-$ , $rd$ (default), $rd+$ . By default ( $rd$ ), as in original CLEVR dataset, the questions contain some redundant attributes resulting from random sampling, while all redundant relationships are removed. In $rd-$ , we also remove all redundant attributes from the questions, leading to no redundancy in the questions. In $rd+$ , we add all possible attributes and relationships into the question, so that questions contain a high level of redundancy. For all the variants, the questions are ensured to be valid.", + "bbox": [ + 496, + 169, + 892, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept distribution. We generate three dataset variants with different concept distributions: bal (balanced), slt (slightly unbalanced) and long (long-tail distributed). More specifically, we change the distribution of shapes, colors and materials while the distribution of size is kept fixed in order to keep visual complexity consistent, since objects with smaller sizes are visually harder to recognize. By default (bal), the shapes and attributes are randomly sampled, leading to a balanced distribution. For slt and long, the concept distribution d is generated by $d_{i} = a^{-i}$ , where $i$ is the index of the concept. $a$ is a hyper-parameter controlling the length of the tail. A larger $a$ leads to more imbalanced distribution and $a = 1$ leads to flat distribution (cf. Fig. 1). For slt, $a = 1.3$ ; for long, $a = 2.0$ . In addition, to better analyze the performance on the frequent and rare concepts, we generate three variants for testing purpose only: head (frequent concepts in the long-tail distribution), tail (infrequent/rare concepts), and oppo (opposite to the long-tail distribution). We test each model on those three variants to analyze the performance on concepts with different degrees of frequency.", + "bbox": [ + 496, + 338, + 893, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept compositionality. We generate 3 versions of the dataset, co-0, co-1 and co-2, with different compositions of the 21 shapes (from 5 categories) and the 8 colors. The compositionality of the dataset is controlled with the codistribution matrix $M \\in \\mathbb{R}^{21 \\times 8}$ , where each entry $M_{ij}$ is the probability an object of the $i$ -th shape has the $j$ -th color. Entries in each row of $M$ sum up to 1. In the version co-0, $M$ is a flat matrix so that the shapes and colors are randomly composed. In co-1, each shape in one category has a different color distribution, e.g. truck and sedan, while shapes from different categories may share the same color distribution e.g. sedan and airliner. Oppositely, in co-2, we make the shapes in same category have the same color distribution, while shapes from different categories have different distributions. The motivation is that since shapes from the same category are visually similar, the difference in co-1", + "bbox": [ + 496, + 659, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14966", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and co-2 will help analyze the difference in model predictions on visually similar objects and dissimilar objects when composed with different color distributions.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Statistics. Every dataset variant contains 30k images, including 20k for training, 5k for validation and 5k for testing. Each image is paired with 10 object-based and 10 part-based questions. By default, the dataset refers to the version with mid visual complexity level (i.e. objects are untextured and has up to 3 parts with distinct attributes), rd redundancy level, balanced (bal) concept distribution and random (co-0) compositionality. More dataset statistics are in supplementary materials.", + "bbox": [ + 75, + 136, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluated methods", + "text_level": 1, + "bbox": [ + 76, + 287, + 261, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Simple baselines", + "text_level": 1, + "bbox": [ + 76, + 311, + 238, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random, Majority. These simple baselines pick a random or the most frequent answer for each question type in the training set as the predicted answer.", + "bbox": [ + 75, + 335, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LSTM. This question-only baseline encodes the question word embeddings with LSTM [21] and predicts answer with an MLP on top of the final hidden states of the LSTM.", + "bbox": [ + 75, + 380, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CNN+LSTM. The image is represented with features extracted by CNN and question is encoded by LSTM. An MLP predicts answer scores based on the concatenation of image and question features.", + "bbox": [ + 75, + 422, + 468, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Existing models", + "text_level": 1, + "bbox": [ + 76, + 493, + 235, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FiLM. We choose Feature-wise Linear Modulation [50] as a representative of classic two-stream feature merging methods. The question features extracted with GRU [10] and image features extracted with CNN are fused with the proposed FiLM module.", + "bbox": [ + 75, + 513, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mDETR. mDETR [29] is a transformer-based detector trained to detects objects in an image conditioned on a text query. The model is pretrained with 1.3M image and text pairs and can be finetuned for various downstream tasks like referring expression understanding or VQA.", + "bbox": [ + 75, + 588, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NSCL. The Neuro-Symbolic Concept Learner [45] is a representative neural symbolic method. NSCL executes neural modules on the scene representation based on the reasoning program, during which the modules learn embeddings of each concept with the answer supervision.", + "bbox": [ + 75, + 662, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NSVQA. Neural-Symbolic VQA [59] is a neural symbolic method composed of three components: A scene parser (Mask-RCNN [19]) that segments an input image and recovers a structural scene representation, a question parser that converts a question from natural language into a program and a program executor that runs the program on the structural scene representation to obtain the answer. Notably, compared to NSCL, the individual components of NSVQA can be learned separately, hence, for example the scene parser can be learned from data that does not necessarily have Visual-Question annotations.", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Probabilistic NSVQA (P-NSVQA)", + "text_level": 1, + "bbox": [ + 498, + 90, + 797, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the program executor in NSVQA is a collection of deterministic, generic functional modules, it can be augmented with a probabilistic reasoning process that takes into account the confidence of the predictions of the scene parser. This allows the model to execute the program that has the largest joint likelihood, instead of only taking the maximal likelihood execution at each step of the program. The experiment results demonstrate a significant performance improvement of this probabilistic approach over the deterministic NSVQA model proposed in [59].", + "bbox": [ + 496, + 113, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, we interpret the confidence of the Mask-RCNN output as a likelihood function for all detected object classes $p_{\\text{object}}$ and their attributes $p_{\\text{att}}$ . Moreover, we define a likelihood $p_{\\text{spatial}}$ for the spatial relations between objects (behind, in front, left, right) that is proportional to the distance between the centers of two bounding boxes. Given a reasoning program containing multiple reasoning steps, we execute each step based on the scene parsing likelihood and produce an step-wise output with confidence. Finally, we use a factorized model, multiplying the output for all the steps to get the final answer prediction. We refer readers to the Appendix for more details.", + "bbox": [ + 496, + 265, + 890, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 454, + 714, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training mDETR requires ground-truth grounding of question tokens to image regions, which is available in Super-CLEVR. NSCL requires bounding box of objects, which can be predicted using a trained Faster RCNN, and the reasoning program, which can be parsed using a trained parser. Similarly, ground-truth programs are used for training NSVQA and P-NSVQA. Note that we empirically find that the question-to-program parsing is a relatively easy task ( $>99\\%$ accuracy using a simple LSTM), so we focus more on models' reasoning ability in our analysis.", + "bbox": [ + 496, + 477, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unless specified, the models are trained with default setting as in the official implementation. FiLM is trained for 100k iterations with batch size 256. mDETR is trained for 30 epochs with batch size 64 using 2 GPUs for both the grounding stage and the answer classification stage. NSCL is trained for 80 epochs with batch size 32. For NSVQA and P-NSVQA, we first train the object parser (Mask RCNN [19]) for 30k iterations with batch size 16, then train the attribute extraction model (using the Res50 backbone) for 100 epochs with batch size 64. For P-NSVQA, when counting the objects or determining whether objects exist in the scene, we use a threshold (0.7) to obtain the final selected objects. Early stopping is used based on validation accuracy. All the models are trained with 200k questions. We repeat experiments on default split for 3 times with different random seeds and get std of $\\pm 0.10$ (P-NSVQA) and $\\pm 0.40$ (NSVQA), showing the statistical significance of our results, then we only run other experiments once.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14967", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results and analysis", + "text_level": 1, + "bbox": [ + 76, + 89, + 272, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first show evaluation results for indomain setting, then provide results and analysis on out-of-domain evaluation. Finally, we describe additional studies and future works.", + "bbox": [ + 76, + 114, + 468, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. In-domain results", + "text_level": 1, + "bbox": [ + 76, + 184, + 250, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-domain evaluation refers to the setting where the training and testing data come from the same domain (the default dataset variant in this case). We compare the indomain results on Super-CLEVR and CLEVR. The results are shown in Fig. 3.", + "bbox": [ + 76, + 207, + 467, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aaa9891acc31869c8baa353199fd99d14b1ff7b1e2b49763bd49e710dfe1370b.jpg", + "image_caption": [ + "Figure 3. Comparison of models' accuracy on Super-CLEVR and the original CLEVR dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 291, + 460, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all the models, the performance is lower on SuperCLEVR than CLEVR, suggesting that Super-CLEVR is a more challenging benchmark. The scenes with vehicles are much harder visually for the models to understand compared with the simpler shapes from CLEVR. Note that the performance gap on two datasets for simple baselines (Random, Majority, LSTM, CNN+LSTM) is smaller than for the other better models. This is because Super-CLEVR contains more object types, and therefore the performance of simply guessing is lower than on CLEVR.", + "bbox": [ + 75, + 537, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When comparing the performance of different models, we find that the neural modular methods, i.e. NSCL, NSVQA, P-NSVQA, perform much better than non-modular ones. This is not surprising given their nearly perfect performance on the original CLEVR dataset, which shows its strong ability to model synthetic images. The large-scale pretrained grounding model mDETR, which is a leading model on both real and synthetic images, also achieves good performance $(82.7\\%)$ on Super-CLEVR. The two-stream method FiLM does not achieve very strong performance $(53.2\\%)$ , but is still much better than the other simple baselines.", + "bbox": [ + 75, + 688, + 467, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed P-NSVQA outperforms all the other models. In particular, on Super-CLEVR, it outperforms its de", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "terministic counterpart, NSVQA, by $2.96\\%$ . This shows the advantage of taking into account the probabilities when the scenes are challenging thus the model's uncertainty of predictions can be utilized.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Out-of-domain results", + "text_level": 1, + "bbox": [ + 500, + 162, + 705, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we train and test the five models (FiLM, mDETR, NSCL, NSVQA and P-NSVQA) on different dataset variants, and diagnose their domain robustness on each of the four domain shift factors. Please refer to Sec. 3.3 for a description of different variants. The validation accuracy is used for analysis here and the results are shown in Tab. 1.", + "bbox": [ + 496, + 185, + 890, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All the methods suffer from domain shifts. The results show that the best performance mostly occurs in situations where the model is tested on the same dataset variant as it is trained on, i.e. the bold or underlined numbers fall mostly on the diagonals in Tab. 1.", + "bbox": [ + 496, + 292, + 890, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the domain robustness of the five models by measuring the relative performance decrease when the testing data differs from the training data, i.e. smaller performance drop on different testing domains means better robustness. Based on this intuition, for easier understanding of Tab. 1, we propose a measurement metric for domain robustness named Relative Degrade (RD) to better analyze the results. We define Relative Degrade as the percentage of accuracy decrease when the model is tested under a domain shift, i.e. the accuracy drop divided by the in-domain accuracy. Specifically, if a model gets accuracy $a$ under in-domain testing (i.e. testing with the same dataset variant as training) and accuracy $b$ under out-of-domain testing (i.e. testing with a different dataset variant from training), then $RD = (a - b) / a$ . Since we train each model on three data variants, the RD's for the three models are averaged to measure its domain robustness.", + "bbox": [ + 496, + 368, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 shows the Relative Degrade of the five models on the four factors. We see that P-NSVQA outperforms other models by a significant margin on three of the four factors, indicating that it has better overall domain robustness. In the following, we take a closer look at the results on each of the factors separately, to diagnose the influence of different model designs.", + "bbox": [ + 496, + 625, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Question redundancy. Neural modular methods are much more robust to question redundancy shifts than non-modular ones. The relative degrades for modular methods are less than $2\\%$ , while one-modular ones degrade for around $20\\%$ . Due to the step-by-step design of the reasoning in modular methods, each reasoning step is independent of the others so that the models are less likely to learn the spurious correlation between question and answers. There-", + "bbox": [ + 496, + 731, + 890, + 852 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1For concept distributions, we compute relative degrade with a slight change: we compute the accuracy drop from head to tail and the drop from long to oppo, take their average, and divide by the accuracy on bal.", + "bbox": [ + 496, + 862, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14968", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a2b92c7b83207a35c6bd6b4acd76b8ef8839d8224ee93b02685e3a0970b352b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FiLMmDETRNSCLNSVQAProb NSVQA
Visual Complexity
easymidhardeasymidhardeasymidhardeasymidhardeasymidhard
easy59.9653.9550.6693.3684.3082.9795.1392.3190.8195.1994.1994.0996.7695.9896.37
mid57.4153.2850.1883.3482.3681.2784.589.1086.3381.9992.8093.7886.2595.7695.11
hard55.9553.1150.4779.7179.9480.7176.8578.6685.0873.1179.7192.6579.8186.4795.36
Question Redundancy
rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+
rd-51.4252.5453.5183.9480.3766.2888.6488.8290.3392.9592.9492.6795.6695.7295.43
rd-50.3953.2854.7882.7782.3670.3688.4589.1091.4591.1992.7892.1494.8795.7295.43
rd+46.1452.3071.4778.4884.0590.4287.9488.3491.1691.3891.9692.8094.8895.4795.72
Concept Distribution
balsltlongbalsltlongbalsltlongbalsltlongbalsltlong
bal50.4753.0454.3580.7175.7974.5485.0883.7975.1092.6590.8283.7495.3694.8989.88
long49.4354.7562.9679.0680.2990.6685.3389.4291.1092.7393.3892.5396.3196.3295.25
head48.6058.0661.6080.7579.6087.4684.5888.3990.1993.8794.8292.4896.4296.8095.92
tail51.8048.7050.0881.5070.8860.9486.1080.2760.5590.2689.2075.3294.0893.2082.68
oppo49.0648.9346.6879.1368.3756.9885.0777.8655.1491.2288.6571.3295.7694.0979.74
Concept Compositionality
co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2
co-053.2857.0056.183.3677.0382.4389.182.5283.7792.8090.1191.5995.7694.0295.12
co-152.4160.5756.6779.4682.4583.9378.8987.1884.278.7489.9990.6787.1294.5394.78
co-252.9657.3760.5380.0377.4187.2478.4081.5588.8477.8589.2892.2387.1993.4995.61
", + "bbox": [ + 81, + 77, + 890, + 483 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a0ec7897ee949d17dd17af71317509a10d136586514c9a79a0164ab789f1ccfa.jpg", + "table_caption": [ + "Table 1. Accuracy of models trained and tested on different domains. Column headings indicate training settings, while rows indicate the dataset variant for testing. The best performance in each row (i.e. the best training setting) is marked in bold and best performance in each column (i.e. the best testing setting) is underlined. Description for different splits is in Sec. 3.3 and analysis is in Sec. 5.2." + ], + "table_footnote": [], + "table_body": "
VisualRedund.Dist.Comp.
FiLM4.0321.3328.469.04
mDETR9.8119.0536.349.45
NSCL15.570.9237.4415.40
NSVQA17.481.7220.9211.44
Prob NSVQA12.880.8413.727.00
", + "bbox": [ + 81, + 549, + 465, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Relative Degrade under domain shifts, i.e. the percentage of accuracy decrease when the model is tested with domain that differs with training. Lower $RD$ means better robustness.", + "bbox": [ + 75, + 667, + 468, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fore the modular methods are less vulnerable to change in question/program length.", + "bbox": [ + 75, + 729, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual complexity. Different from our findings on question redundancy, for domain shifts in visual complexity, non-modular methods are more robust compared to modular ones. As shown in Tab. 2, while FiLM and mDETR gets less than $10\\%$ degrade, NSCL and (P-)NSVQA degrade for more than $12\\%$ . The reason might be that the simple reasoning modules in modular methods can not process the visual signals as well as the dense non-modular models.", + "bbox": [ + 75, + 761, + 468, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing P-NSVQA with NSVQA, we find that in", + "bbox": [ + 96, + 883, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "jecting probability into deterministic symbolic reasoning greatly improves the robustness on visual complexity (4.04% decrease in $RD$ ). This suggests that some errors in visual understanding can be corrected and recovered by taking into account the uncertainty of visual parsing and combining the results of each reasoning step with probability.", + "bbox": [ + 496, + 551, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Concept distribution. While all the four existing models suffer a lot (larger than $20\\%$ RD) on domain shifts in concept distribution, we see that the symbolic method NSVQA is better than the other three (by more than $7.5\\%$ ). With the disentangled reasoning and visual understanding components in NSVQA, the distribution priors in the images and the programs/answers cannot intertwine with each other, which prevent the model heavily relying on the priors. With uncertainty, we can further boost the robustness of NSVQA with a large margin (from $21\\%$ to $14\\%$ RD).", + "bbox": [ + 496, + 643, + 892, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, the head-tail results suggests that the overall accuracy, which is commonly used to measure VQA performance, should be taken with cautious. When the testing split is imbalanced, the seemingly high accuracy is misleading because the head concepts dominate the testing while the tail ones are not well-reflected. For example, for NSCL, although it gets high accuracy (91%) on the long-tailed", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14969", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data, its performance is only $60.6\\%$ on the tail concepts. In real-world datasets, the data are usually not well-balanced, which suggests the value of synthetic testing.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concept compositionality. Comparing the existing methods, we find that the non-modular methods seems to be more robust than modular methods NSCL or NSVQA. However, with uncertainty, P-NSVQA improves the result of NSVQA, which even outperforms the non-modular methods. This suggest the large potential of better robustness of modular methods by improving current models.", + "bbox": [ + 75, + 137, + 467, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, while non-modular methods are more robust to visual complexity shifts, the modular symbolic methods (improved with uncertainty) are more robust on the other three factors. By disentangling reasoning with visual understanding, separately executing every each reasoning step then merging the results of the steps using probabilities based on uncertainty, our P-NSVQA outperforms all the existing models in question redundancy, concept distribution and compositionality. Therefore, we suggest that symbolic reasoning with uncertainty leads to strong VQA models that are robust to domain shifts.", + "bbox": [ + 75, + 243, + 467, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. More analysis and future work", + "text_level": 1, + "bbox": [ + 76, + 421, + 349, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic-to-real transfer. We provide an additional proof-of-concept study to show that the findings drawn from Super-CLEVR dataset can transfer to real datasets. In the following experiments, we show our finding that neuro-symbolic methods (NSCL, NSVQA, P-NSVQA) are more robust than mDETR on question redundancy also holds true on the real GQA dataset [25]. More precisely, we progressively removed the redundant operations from the reasoning program in GQA testdev split, and then regenerated questions using a program-to-question generator. Using the change of models' testing accuracy as the redundant operations are removed, we can evaluate the models' robustness towards question redundancy. The results are show in Tab. 3.2 We observe that the performance drop of mDETR is much larger than neuro-symbolic methods as the redundant information is progressively removed, which indicates that symbolic methods have better question redundancy than mDETR on GQA dataset. This is consistent with our findings on Super-CLEVR.", + "bbox": [ + 75, + 444, + 467, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reasoning with part-object hierarchies. In addition to evaluating domain generalization, Super-CLEVR can be extended for broader purposes, e.g. part-based reasoning. We can ask questions like \"what is the color of the front wheel of the bike?\", \"what is the color of the vehicle that has a yellow wheel\", etc. Those questions require the model to correctly understanding the part-object hierarchy, which", + "bbox": [ + 75, + 733, + 467, + 838 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e85fcd9c2c01ecd6cfb12f2c632c2dd41d1451667785fc437e4eee5dbd38774a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
0%14%32%70%91%100%
mDETR0-4.82-8.46-13.16-13.88-14.56
NSCL0-0.14-0.34-1.09-1.71-2.59
NSVQA0-3.47-4.80-7.01-7.02-7.02
P-NSVQA0-1.93-3.15-5.73-5.91-5.78
", + "bbox": [ + 504, + 89, + 890, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Accuracy drop on the GQA dataset when redundant information is progressively removed.", + "bbox": [ + 500, + 172, + 890, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is an ability that current VQA models lack.", + "bbox": [ + 500, + 215, + 784, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The main limitations of our work lie in the synthetic nature of our dataset. Future efforts can be made in collecting better controlled and balanced real datasets for model diagnosis. We emphasize that the purpose of the dataset is for model diagnosis and that models should also be tested on real data.", + "bbox": [ + 498, + 229, + 890, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 334, + 617, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We diagnose domain shifts in visual reasoning using a proposed virtual benchmark, Super-CLEVR, where distinct factors can be independently studied with controlled data generation. We evaluate four existing methods and show that all of them struggle with domain shifts, highlighting the importance of out-of-domain testing. Among the evaluated methods, neural modular methods are more robust towards question redundancy. In particular, NSVQA with disentangled perception and reasoning shows better robustness towards distribution and compositionality shifts. We further propose P-NSVQA, which improves NSVQA with uncertainty in the reasoning modules. We show that P-NSVQA outperforms all the existing methods in both in-domain testing and out-of-domain testing. With detailed analysis, our study suggests that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. We hope our analysis may facilitate better understanding of strengths and weaknesses of VQA models and, more broadly, future work might explore using the Super-CLEVR benchmark for other tasks like part-based reasoning.", + "bbox": [ + 498, + 359, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 686, + 653, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would like to thank Nils Holzenberger, Kate Sanders, Chenxi Liu, Zihao Xiao, Qing Liu, Reno Kriz, and David Etter for their helpful comments and suggestions, as well as the anonymous reviewers. Zhuowan Li is supported by ONR N00014-21-1-2812 and grant IAA80052272 from the Institute for Assured Autonomy at JHU. Elias Stengel-Eskin is supported by an NSF GRFP. A. Kortylewski acknowledges support via his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No.468670075.", + "bbox": [ + 498, + 709, + 890, + 859 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2For implementation of NSCL on a real-word dataset, we use the model in [39] (the version without calibration). The model accuracies on the original not-perturbed GQA testdev split are as following: mDETR $(61.67\\%)$ NSCL $(56.13\\%)$ NSVQA $(39.58\\%)$ P-NSVQA $(39.66\\%)$", + "bbox": [ + 75, + 849, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14970", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2", + "[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4971-4980, 2018. 1, 2, 3", + "[3] Arjun Akula, Soravit Changpinyo, Boqing Gong, Piyush Sharma, Song-Chun Zhu, and Radu Soricut. Crossvqa: scalably generating benchmarks for systematically testing vqa generalization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2148-2166, 2021. 1, 3", + "[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2", + "[5] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2", + "[6] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019. 2, 3", + "[7] Remi Cadene, Corentin Dancette, Matthieu Cord, Devi Parikh, et al. Rubi: Reducing unimodal biases for visual question answering. Advances in neural information processing systems, 32, 2019. 3", + "[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 1, 3", + "[9] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 1, 2", + "[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.5", + "[11] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1554–1563, 2021. 1, 2", + "[12] William Ford and David Olson. The elaboration of the noun phrase in children's description of objects. Journal of Experimental Child Psychology, 19(3):371-382, 1975. 3" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2", + "[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3", + "[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 3", + "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2", + "[17] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. arXiv preprint arXiv:2204.02285, 2022. 2", + "[18] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5", + "[20] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6930-6940, 2021. 3", + "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5", + "[22] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. PMLR, 2018. 3", + "[23] Yining Hong, Li Yi, Josh Tenenbaum, Antonio Torralba, and Chuang Gan. *Ptr: A benchmark for part-based conceptual, relational, and physical reasoning*. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[24] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. 2018. 2", + "[25] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 2, 8", + "[26] Jingjing Jiang, Ziyi Liu, Yifan Liu, Zhixiong Nan, and Nanning Zheng. X-ggm: Graph generative modeling for out-of-distribution generalization in visual question answering. In" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14971", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the 29th ACM International Conference on Multimedia, pages 199-208, 2021. 1, 3", + "[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 1, 2", + "[28] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 3-12. Springer, 2021. 3", + "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. arXiv preprint arXiv:2104.12763, 2021. 1, 2, 5", + "[30] Corentin Kervadec, Grigory Antipov, Moez Baccouche, and Christian Wolf. Roses are red, violets are blue... but should vqa expect them to? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2776-2785, 2021. 1, 2, 3", + "[31] Corentin Kervadec, Theo Jaunet, Grigory Antipov, Moez Baccouche, Romain Vuillemot, and Christian Wolf. How transferable are reasoning patterns in vqa? 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4205-4214, 2021. 2", + "[32] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[33] Ruud Koolen, Martijn Goudbeek, and Emiel Krahmer. Effects of scene variation on referential overspecification. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 33, 2011. 3", + "[34] Satwik Kottur, José MF Moura, Devi Parikh, Dhruv Batra, and Marcus Rohrbach. Clevr-dialog: A diagnostic dataset for multi-round reasoning in visual dialog. arXiv preprint arXiv:1903.03166, 2019. 2", + "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2", + "[36] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1, 3", + "[37] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10313-10322, 2019. 2", + "[38] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 2", + "[39] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14910-14919, October 2021. 1, 3, 8", + "[40] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning part segmentation through unsupervised domain adaptation from synthetic vehicles. In CVPR, 2022. 1, 4", + "[41] Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194, 2019. 2", + "[42] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3", + "[43] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Advances in neural information processing systems, 31, 2018. 3", + "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2", + "[45] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision. In International Conference on Learning Representations, 2019. 1, 2, 5", + "[46] Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184, 2013. 3", + "[47] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12695–12705, 2021. 1, 2, 3", + "[48] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12700-12710, June 2021. 1, 2", + "[49] Thomas Pechmann. Incremental speech production and referential overspecification. 1989. 3", + "[50] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14972", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 3", + "[52] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 2", + "[53] Leonard Salewski, A Koepke, Hendrik Lensch, and Zeynep Akata. Clevr-x: A visual reasoning dataset for natural language explanations. In International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers, pages 69-88. Springer, 2022. 2", + "[54] Susan Sonnenschein. The development of referential communication skills: Some situations in which speakers give redundant messages. Journal of Psycholinguistic Research, 14(5):489-508, 1985. 3", + "[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2", + "[56] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 3", + "[57] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 3", + "[58] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2", + "[59] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Joshua B Tenenbaum. Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding. In Advances in Neural Information Processing Systems (NIPS), 2018. 1, 2, 5", + "[60] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6281–6290, 2019. 2", + "[61] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5317-5327, 2019. 2", + "[62] Mingda Zhang, Tristan Maidment, Ahmad Diab, Adriana Kovashka, and Rebecca Hwa. Domain-robust vqa with diverse datasets and methods but no target labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7046–7056, 2021. 1, 3", + "[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 3", + "[65] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 2" + ], + "bbox": [ + 501, + 90, + 892, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14973", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_model.json b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6acd377d4982c2a9e589db7bbfd839e5a72d5732 --- /dev/null +++ b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.131, + 0.738, + 0.177 + ], + "angle": 0, + "content": "Super-CLEVR: A Virtual Benchmark to Diagnose Domain Robustness in Visual Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.203, + 0.825, + 0.276 + ], + "angle": 0, + "content": "Zhuowan Li\\(^{1}\\) Xingrui Wang\\(^{2}\\) Elias Stengel-Eskin\\(^{1}\\) Adam Kortylewski\\(^{3,4}\\) Wufei Ma\\(^{1}\\) Benjamin Van Durme\\(^{1}\\) Alan Yuille\\(^{1}\\) \n\\(^{1}\\) Johns Hopkins University \\(^{2}\\) University of Southern California \n\\(^{3}\\) Max Planck Institute for Informatics \\(^{4}\\) University of Freiburg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.706 + ], + "angle": 0, + "content": "Visual Question Answering (VQA) models often perform poorly on out-of-distribution data and struggle on domain generalization. Due to the multi-modal nature of this task, multiple factors of variation are intertwined, making generalization difficult to analyze. This motivates us to introduce a virtual benchmark, Super-CLEVR, where different factors in VQA domain shifts can be isolated in order that their effects can be studied independently. Four factors are considered: visual complexity, question redundancy, concept distribution and concept compositionality. With controllably generated data, Super-CLEVR enables us to test VQA methods in situations where the test data differs from the training data along each of these axes. We study four existing methods, including two neural symbolic methods NSCL [45] and NSVQA [59], and two non-symbolic methods FiLM [50] and mDETR [29]; and our proposed method, probabilistic NSVQA (P-NSVQA), which extends NSVQA with uncertainty reasoning. P-NSVQA outperforms other methods on three of the four domain shift factors. Our results suggest that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. The dataset and code are released at https://github.com/Lizw14/Super-CLEVR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Visual question answering (VQA) is a challenging task that assesses the reasoning ability of models to answer questions based on both visual and linguistic inputs. Current VQA methods are typically developed on standard benchmarks like VQAv2 [16] or GQA [25], with the implicit assumption that testing data comes from the same underlying distribution as training data. However, as has been widely studied in computer vision [15,36,51], algorithms trained on one domain often fail to generalize to other domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Moreover, having learned the distributional prior of training data, models often struggle on out-of-distribution tests. This has been studied in VQA from the perspective of domain transfer [8,57,62], dataset bias [2,11,48], counter-factual diagnosis [9,47], and out-of-distribution benchmarking [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.893, + 0.538 + ], + "angle": 0, + "content": "The multi-modal nature of VQA gives rise to multiple intertwined factors of variation, making domain shift an especially difficult problem to study. For example, [8] suggests that VQA domain shifts are a combination of differences in images, questions or answers; and [39] reveals a gap between synthetic and real VQA datasets by differences in the over-specification of questions and the underlying distribution of concepts. However, despite a wealth of research on domain generalization in VQA [3,26,57,62], there is no systematic analysis of the contributing factors in domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.854 + ], + "angle": 0, + "content": "To this end, we introduce a virtual benchmark, SuperCLEVR, which enables us to test VQA algorithms in situations where the test data differs from the training data. We decompose the domain shift into a set of isolated contributing factors, so that their effects can be diagnosed independently. We study four factors: visual complexity, question redundancy, concept distribution, and concept compositionality. These are illustrated in Fig. 1 and described in Sec. 3.1. With controllable data generation using our SuperCLEVR virtual benchmark, we are able to isolate the different factors in VQA domain shifts so that their effects can be studied independently. Compared with the original CLEVR dataset [27], Super-CLEVR contains more complicated visual components and has better controllability over the domain shift factors. As shown in Fig. 1, the Super-CLEVR dataset contains images rendered from 3D graphical vehicle models in the UDA-Part dataset [40], paired with questions and answers automatically generated from templates. The objects and questions are sampled based on the specified underlying probability distribution, which can be controlled to produce distribution shifts in different factors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With Super-CLEVR, we diagnose the domain robustness of current VQA models. Four representative models are studied: for the classic two-stream feature fusing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14963" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.065, + 0.873, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.338, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 1. We decompose VQA domain shifts into four contributing factors: visual complexity, question redundancy, concept distribution and concept compositionality. The domain shifts along each factor can be independently studied with the proposed Super-CLEVR dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.576 + ], + "angle": 0, + "content": "architecture, we choose FiLM [50]; for a large-scale pretrained model we take mDETR [29]; we use NSCL [45] and NSVQA [59] as representative neuro-symbolic methods. We observe that all these models suffer from domain shifts to varying degrees of sensitivity. We analyze each factor separately to examine the influence of different model designs. Specifically, we find that the step-by-step design of neural modular methods enhances their robustness to changes in question redundancy compared with non-modular ones; however, the non-modular models are more robust to visual complexity. Furthermore, thanks to its decomposed reasoning and perception, NSVQA is more robust to concept distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.777 + ], + "angle": 0, + "content": "While existing models suffer from domain shifts with different characteristics, we make a technical improvement over NSVQA which enables it to significantly outperform existing models on three of the four factors. In particular, we inject probabilities into the deterministic symbolic executor of NSVQA, empowering it to take into account the uncertainty of scene understanding. We name our model probabilistic NSVQA (P-NSVQA), and show that its performance improvement in both the in-domain and out-of-domain settings. With superior results of P-NSVQA, we suggest that disentangling reasoning from vision and language understanding, together with probabilistic uncertainty, gives a strong model that is robust to domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our contributions are as follows. (1) We introduce the Super-CLEVR benchmark to diagnose VQA robustness along four different factors independently. This benchmark can also be used for part-based reasoning. (2) We enhance a neural-symbolic method by taking the uncertainty of visual understanding into account in reasoning. (3) We conduct detailed analysis of four existing methods, as well as our novel approach to study the influence of model designs on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.426 + ], + "angle": 0, + "content": "distinct robustness factors. We conclude that disentangled reasoning and perception plus explicit modeling of uncertainty leads to a more robust VQA model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.438, + 0.639, + 0.453 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Visual question answering (VQA). Popular VQA methods fall into three categories. Two-stream methods extract features for image and questions using CNN and LSTM respectively, then enable interaction between the two modalities with different feature fusing methods [4, 13, 24, 32, 37, 50, 60]. Neural symbolic methods, on the other hand, use a parse-then-execute pipeline where the question is parsed into a functional program, which is then executed on the image using neural modules [45, 59]. Recently, transformers-based models have achieved impressive performance on various vision-and-language tasks by pretraining on large scale dataset then finetuning for downstream tasks [29, 38, 44, 55, 63]. We choose FiLM [50], NSCL [45] and mDETR [29] as category representatives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "VQA datasets. Datasets containing real images and human-written questions have been widely used to benchmark VQA models, e.g. VQA [5], VQAv2 [16], Visual 7w [65], VizWiz [18], Visual Genome [35], COCO QA [52], etc. However, subsequent work has revealed the strong prior and bias in those datasets which might be exploited by models to correctly predict the answers without reasoning [1, 16, 17, 30, 31, 47, 48]. Attempts to address this problem include better balancing datasets [2] and creating counterfactual examples [9, 11]. To assess a model's true reasoning ability, the CLEVR dataset [27] proposes to generate complex multi-step questions on synthetic images, which is then extended to various vision-and-language tasks [6,23,34,41,53,58,61]. The GQA dataset [25] extends CLEVR-style questions to real images. Our benchmark is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14964" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "distinct from existing ones because we introduce more complex visual scenes into CLEVR and provide controllability to study domain robustness on isolated factors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.365 + ], + "angle": 0, + "content": "Domain shift in VQA. Domain shift is a long-standing challenge in computer vision, explore in prior works in domain adaptation [14, 15, 22, 43] and domain generalization [36, 51]. Recent works have focused on domain shifts in VQA. [8, 57] improves model adaptation between datasets by feature learning. [62] analyze domain shifts between nine popular VQA datasets and proposes an unsupervised method to bridge the gaps. [39] generalize symbolic reasoning from synthetic to real dataset. [3] introduce a question-answer generation module that simulates the domain shifts. [26] propose a training scheme X-GGM to improve out-of-distribution generalization. [6] assess models generalization on the CLOSURE of linguistic components. In contrast to prior works we study each of the different domain shift factors independently with our virtual benchmark." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.227, + 0.398 + ], + "angle": 0, + "content": "3. Super-CLEVR" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.407, + 0.364, + 0.421 + ], + "angle": 0, + "content": "3.1. Motivation: domain shift factors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Visual complexity. A major difference between different VQA datasets is visual complexity. For example, in the CLEVR dataset, objects are simple, atomic shapes while in real-world data, objects are more complex and have hierarchical parts. While hard to quantify, visual complexity is related to various factors, such as object variety, object size, background, texture, lighting, occlusion, view point, etc. In our work, we control visual complexity by introducing more challenging objects that can have distinct attributes associated with their parts, and by optionally pasting various textures onto objects. Examples of generated images with different complexity levels are shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Question redundancy. Question redundancy refers to the amount of over-specified or redundant information in the question, which can be in the form of either attributes or relationships. For example, in Fig. 1, \"what color is the large bus behind the cyan car\", large (attributes) and behind the cyan car (relationship) are redundant because there is only one bus in the image. As observed in linguistics and cognitive science [12, 33, 49, 54], human speakers may include over-specified information when identifying a target object, which has also been studied in referring expression generation [46]. For VQA, as analyzed in [39], a significant difference between synthetic and real datasets is that real questions contain some redundant information, which sometimes is a distraction leading to model prediction errors. Therefore, in this work, we generate questions with different redundancy levels and study the effect of question redundancy on model behaviors." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Concept distribution. The distributions of concept, i.e. objects (e.g. car) and attributes (e.g. large), are distinct" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.088, + 0.639, + 0.102 + ], + "angle": 0, + "content": "Object with parts:" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.104, + 0.643, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.145, + 0.601, + 0.153 + ], + "angle": 0, + "content": "car" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.114, + 0.758, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.145, + 0.709, + 0.155 + ], + "angle": 0, + "content": "bus" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.108, + 0.872, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.145, + 0.845, + 0.157 + ], + "angle": 0, + "content": "airplane" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.167, + 0.643, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.214, + 0.605, + 0.224 + ], + "angle": 0, + "content": "bike" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.161, + 0.748, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.213, + 0.739, + 0.224 + ], + "angle": 0, + "content": "motorcycle" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.23, + 0.789, + 0.243 + ], + "angle": 0, + "content": "Texture: dotted, checkered, stripped, none" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.244, + 0.872, + 0.257 + ], + "angle": 0, + "content": "Color: green, gray, brown, yellow, red, purple, cyan, blue" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.258, + 0.626, + 0.271 + ], + "angle": 0, + "content": "Size: large, small" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.272, + 0.669, + 0.285 + ], + "angle": 0, + "content": "Material: rubber, metal" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.295, + 0.892, + 0.322 + ], + "angle": 0, + "content": "Figure 2. Super-CLEVR contains 21 vehicle models belonging to 5 categories, with controllable attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.574 + ], + "angle": 0, + "content": "across different VQA datasets. For example, while colors are well-balanced in CLEVR dataset, in the GQA dataset, the color distribution is long-tailed where \"white\" appears \\(>50\\) times more frequently than \"gold\". Long-tailed distributions have been a challenge in many computer vision tasks [20, 28, 42, 56, 64]. In VQA, the long-tailed concept distribution not only hinders the learning of infrequent concepts due to few training samples, but also introduces strong biases and priors in the dataset that may mislead the models. For example, \"tennis\" is the correct answer to most questions with \"what sport is ...\" [2]. With strong priors in data, it is hard to assess the true reasoning capacity of current models. While previous works address this problem by carefully re-balancing datasets [2], in our work, we controllably vary the concept distribution in our dataset and study model robustness to concept distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Concept compositionality. Concept compositionality refers to how different concepts (shapes, attributes) compose and co-occur with each other, e.g. roses are usually red while violets are usually blue [30]. Concept compositionality can be viewed as a conditional concept distribution in the context of other concepts. Shifts in concept compositionality impede the generalization of VQA models. For example, the model may fail to recognize a green banana because most bananas are yellow in the training data [7]. Previous works evaluate the out-of-distribution performance by collecting counterfactual testing examples [47]. In our work, we control the compositionality of shapes and colors with an intuitive motivation: if, for example, in the training data, bicycles are red and cars are blue, will the models be able to recognize blue bicycles and red cars in testing?" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.681, + 0.832 + ], + "angle": 0, + "content": "3.2. Dataset generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Super-CLEVR follows a similar data generation pipeline as CLEVR, but with more complex visual components and better control of domain gap factors. We describe the generation procedure below." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14965" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Objects with parts. To improve the visual complexity of CLEVR scenes, we replace the simple shapes (e.g., cube, sphere) in CLEVR dataset with vehicles from UDA-Part dataset [40]. There are 21 vehicle models, belonging to 5 categories: car, motorbike, aeroplane, bus, and bicycle. Each 3D model comes with part annotations, e.g., left front wheel or left right door for car. Examples for the vehicle models are shown in Fig. 2. We remove or merge small parts from the original annotations to avoid severe difficulty in visual understanding. The full object and parts list is in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Attributes. Besides the attributes in the original CLEVR dataset, i.e. color, material, size, we optionally add texture as an additional attribute to increase visual complexity. Note that in order to enable part-based questions, the attributes (color or material) of object parts can be different from that of the object. For example, a blue car can have a red wheel or a green door. In this case, the attribute of the holistic object refers to the attribute of its main body (e.g. the blue car has blue frame)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.473, + 0.546 + ], + "angle": 0, + "content": "Scene rendering. Following CLEVR, each scene contains 3 to 10 objects. The objects are placed onto the ground plane with random position and orientation. When placing the objects, we ensure that the objects do not overlap with each other and we avoid severe occlusion by thresholding the number of visible pixels for each object. Random jitters are added to lamp and camera positions. When rendering, we also save the ground-truth bounding boxes and segmentation masks for each of the objects and their parts, which are required when training some of the models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.473, + 0.684 + ], + "angle": 0, + "content": "Question generation. Super-CLEVR follows similar question generation pipeline in CLEVR, which instantiates question templates using the underlying reasoning program that can be operated on the scene graph. For example, the program select_shape truck) \\(\\rightarrow\\) query_color (·) can be instantiated as question \"what is the color of the truck\". Therefore, redundancy level of questions can be controlled by removing or adding redundant reasoning steps in the underlying reasoning program." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.293, + 0.71 + ], + "angle": 0, + "content": "3.3. Controlling the dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.795 + ], + "angle": 0, + "content": "To study domain generalization, we generate several variants of the dataset for each of the domain shift factors. The variants of the datasets serve as different data domains to test the model robustness. Here we describe the method for controllably generating the dataset variants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Visual complexity. We generate three variants of the dataset with different levels of visual complexity: easy, mid (middle) and hard. The only difference between the 3 versions is visual complexity: for the easy version, objects with different sizes, colors and materials are placed into the scene; for the middle version, we choose 3 parts on each object that are visible and randomly change their attributes;" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "for the hard version, we further add random textures to the objects and parts. An example of the 3 dataset versions can be found in Fig. 1. Note that the scene layout and the questions are shared, so that the influence of visual complexity can be isolated and studied independently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.893, + 0.336 + ], + "angle": 0, + "content": "Question redundancy. Three variants of the dataset with different redundancy levels are generated: \\( rd- \\), \\( rd \\) (default), \\( rd+ \\). By default (\\( rd \\)), as in original CLEVR dataset, the questions contain some redundant attributes resulting from random sampling, while all redundant relationships are removed. In \\( rd- \\), we also remove all redundant attributes from the questions, leading to no redundancy in the questions. In \\( rd+ \\), we add all possible attributes and relationships into the question, so that questions contain a high level of redundancy. For all the variants, the questions are ensured to be valid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.657 + ], + "angle": 0, + "content": "Concept distribution. We generate three dataset variants with different concept distributions: bal (balanced), slt (slightly unbalanced) and long (long-tail distributed). More specifically, we change the distribution of shapes, colors and materials while the distribution of size is kept fixed in order to keep visual complexity consistent, since objects with smaller sizes are visually harder to recognize. By default (bal), the shapes and attributes are randomly sampled, leading to a balanced distribution. For slt and long, the concept distribution d is generated by \\(d_{i} = a^{-i}\\), where \\(i\\) is the index of the concept. \\(a\\) is a hyper-parameter controlling the length of the tail. A larger \\(a\\) leads to more imbalanced distribution and \\(a = 1\\) leads to flat distribution (cf. Fig. 1). For slt, \\(a = 1.3\\); for long, \\(a = 2.0\\). In addition, to better analyze the performance on the frequent and rare concepts, we generate three variants for testing purpose only: head (frequent concepts in the long-tail distribution), tail (infrequent/rare concepts), and oppo (opposite to the long-tail distribution). We test each model on those three variants to analyze the performance on concepts with different degrees of frequency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Concept compositionality. We generate 3 versions of the dataset, co-0, co-1 and co-2, with different compositions of the 21 shapes (from 5 categories) and the 8 colors. The compositionality of the dataset is controlled with the codistribution matrix \\( M \\in \\mathbb{R}^{21 \\times 8} \\), where each entry \\( M_{ij} \\) is the probability an object of the \\( i \\)-th shape has the \\( j \\)-th color. Entries in each row of \\( M \\) sum up to 1. In the version co-0, \\( M \\) is a flat matrix so that the shapes and colors are randomly composed. In co-1, each shape in one category has a different color distribution, e.g. truck and sedan, while shapes from different categories may share the same color distribution e.g. sedan and airliner. Oppositely, in co-2, we make the shapes in same category have the same color distribution, while shapes from different categories have different distributions. The motivation is that since shapes from the same category are visually similar, the difference in co-1" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "14966" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "and co-2 will help analyze the difference in model predictions on visually similar objects and dissimilar objects when composed with different color distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.274 + ], + "angle": 0, + "content": "Dataset Statistics. Every dataset variant contains 30k images, including 20k for training, 5k for validation and 5k for testing. Each image is paired with 10 object-based and 10 part-based questions. By default, the dataset refers to the version with mid visual complexity level (i.e. objects are untextured and has up to 3 parts with distinct attributes), rd redundancy level, balanced (bal) concept distribution and random (co-0) compositionality. More dataset statistics are in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.288, + 0.263, + 0.304 + ], + "angle": 0, + "content": "4. Evaluated methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.313, + 0.24, + 0.329 + ], + "angle": 0, + "content": "4.1. Simple baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.381 + ], + "angle": 0, + "content": "Random, Majority. These simple baselines pick a random or the most frequent answer for each question type in the training set as the predicted answer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.424 + ], + "angle": 0, + "content": "LSTM. This question-only baseline encodes the question word embeddings with LSTM [21] and predicts answer with an MLP on top of the final hidden states of the LSTM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.469, + 0.483 + ], + "angle": 0, + "content": "CNN+LSTM. The image is represented with features extracted by CNN and question is encoded by LSTM. An MLP predicts answer scores based on the concatenation of image and question features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.236, + 0.51 + ], + "angle": 0, + "content": "4.2. Existing models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.469, + 0.589 + ], + "angle": 0, + "content": "FiLM. We choose Feature-wise Linear Modulation [50] as a representative of classic two-stream feature merging methods. The question features extracted with GRU [10] and image features extracted with CNN are fused with the proposed FiLM module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.663 + ], + "angle": 0, + "content": "mDETR. mDETR [29] is a transformer-based detector trained to detects objects in an image conditioned on a text query. The model is pretrained with 1.3M image and text pairs and can be finetuned for various downstream tasks like referring expression understanding or VQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.469, + 0.736 + ], + "angle": 0, + "content": "NSCL. The Neuro-Symbolic Concept Learner [45] is a representative neural symbolic method. NSCL executes neural modules on the scene representation based on the reasoning program, during which the modules learn embeddings of each concept with the answer supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.469, + 0.901 + ], + "angle": 0, + "content": "NSVQA. Neural-Symbolic VQA [59] is a neural symbolic method composed of three components: A scene parser (Mask-RCNN [19]) that segments an input image and recovers a structural scene representation, a question parser that converts a question from natural language into a program and a program executor that runs the program on the structural scene representation to obtain the answer. Notably, compared to NSCL, the individual components of NSVQA can be learned separately, hence, for example the scene parser can be learned from data that does not necessarily have Visual-Question annotations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.798, + 0.108 + ], + "angle": 0, + "content": "4.3. Probabilistic NSVQA (P-NSVQA)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Since the program executor in NSVQA is a collection of deterministic, generic functional modules, it can be augmented with a probabilistic reasoning process that takes into account the confidence of the predictions of the scene parser. This allows the model to execute the program that has the largest joint likelihood, instead of only taking the maximal likelihood execution at each step of the program. The experiment results demonstrate a significant performance improvement of this probabilistic approach over the deterministic NSVQA model proposed in [59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.446 + ], + "angle": 0, + "content": "In particular, we interpret the confidence of the Mask-RCNN output as a likelihood function for all detected object classes \\( p_{\\text{object}} \\) and their attributes \\( p_{\\text{att}} \\). Moreover, we define a likelihood \\( p_{\\text{spatial}} \\) for the spatial relations between objects (behind, in front, left, right) that is proportional to the distance between the centers of two bounding boxes. Given a reasoning program containing multiple reasoning steps, we execute each step based on the scene parsing likelihood and produce an step-wise output with confidence. Finally, we use a factorized model, multiplying the output for all the steps to get the final answer prediction. We refer readers to the Appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.715, + 0.471 + ], + "angle": 0, + "content": "4.4. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Training mDETR requires ground-truth grounding of question tokens to image regions, which is available in Super-CLEVR. NSCL requires bounding box of objects, which can be predicted using a trained Faster RCNN, and the reasoning program, which can be parsed using a trained parser. Similarly, ground-truth programs are used for training NSVQA and P-NSVQA. Note that we empirically find that the question-to-program parsing is a relatively easy task (\\(>99\\%\\) accuracy using a simple LSTM), so we focus more on models' reasoning ability in our analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unless specified, the models are trained with default setting as in the official implementation. FiLM is trained for 100k iterations with batch size 256. mDETR is trained for 30 epochs with batch size 64 using 2 GPUs for both the grounding stage and the answer classification stage. NSCL is trained for 80 epochs with batch size 32. For NSVQA and P-NSVQA, we first train the object parser (Mask RCNN [19]) for 30k iterations with batch size 16, then train the attribute extraction model (using the Res50 backbone) for 100 epochs with batch size 64. For P-NSVQA, when counting the objects or determining whether objects exist in the scene, we use a threshold (0.7) to obtain the final selected objects. Early stopping is used based on validation accuracy. All the models are trained with 200k questions. We repeat experiments on default split for 3 times with different random seeds and get std of \\(\\pm 0.10\\) (P-NSVQA) and \\(\\pm 0.40\\) (NSVQA), showing the statistical significance of our results, then we only run other experiments once." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14967" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.273, + 0.108 + ], + "angle": 0, + "content": "5. Results and analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.176 + ], + "angle": 0, + "content": "In this section, we first show evaluation results for indomain setting, then provide results and analysis on out-of-domain evaluation. Finally, we describe additional studies and future works." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.185, + 0.251, + 0.2 + ], + "angle": 0, + "content": "5.1. In-domain results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.208, + 0.468, + 0.284 + ], + "angle": 0, + "content": "In-domain evaluation refers to the setting where the training and testing data come from the same domain (the default dataset variant in this case). We compare the indomain results on Super-CLEVR and CLEVR. The results are shown in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.292, + 0.462, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.529 + ], + "angle": 0, + "content": "Figure 3. Comparison of models' accuracy on Super-CLEVR and the original CLEVR dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.689 + ], + "angle": 0, + "content": "For all the models, the performance is lower on SuperCLEVR than CLEVR, suggesting that Super-CLEVR is a more challenging benchmark. The scenes with vehicles are much harder visually for the models to understand compared with the simpler shapes from CLEVR. Note that the performance gap on two datasets for simple baselines (Random, Majority, LSTM, CNN+LSTM) is smaller than for the other better models. This is because Super-CLEVR contains more object types, and therefore the performance of simply guessing is lower than on CLEVR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.87 + ], + "angle": 0, + "content": "When comparing the performance of different models, we find that the neural modular methods, i.e. NSCL, NSVQA, P-NSVQA, perform much better than non-modular ones. This is not surprising given their nearly perfect performance on the original CLEVR dataset, which shows its strong ability to model synthetic images. The large-scale pretrained grounding model mDETR, which is a leading model on both real and synthetic images, also achieves good performance \\((82.7\\%)\\) on Super-CLEVR. The two-stream method FiLM does not achieve very strong performance \\((53.2\\%)\\), but is still much better than the other simple baselines." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our proposed P-NSVQA outperforms all the other models. In particular, on Super-CLEVR, it outperforms its de" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "terministic counterpart, NSVQA, by \\(2.96\\%\\). This shows the advantage of taking into account the probabilities when the scenes are challenging thus the model's uncertainty of predictions can be utilized." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.163, + 0.707, + 0.178 + ], + "angle": 0, + "content": "5.2. Out-of-domain results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.291 + ], + "angle": 0, + "content": "In this section, we train and test the five models (FiLM, mDETR, NSCL, NSVQA and P-NSVQA) on different dataset variants, and diagnose their domain robustness on each of the four domain shift factors. Please refer to Sec. 3.3 for a description of different variants. The validation accuracy is used for analysis here and the results are shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.368 + ], + "angle": 0, + "content": "All the methods suffer from domain shifts. The results show that the best performance mostly occurs in situations where the model is tested on the same dataset variant as it is trained on, i.e. the bold or underlined numbers fall mostly on the diagonals in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.625 + ], + "angle": 0, + "content": "We compare the domain robustness of the five models by measuring the relative performance decrease when the testing data differs from the training data, i.e. smaller performance drop on different testing domains means better robustness. Based on this intuition, for easier understanding of Tab. 1, we propose a measurement metric for domain robustness named Relative Degrade (RD) to better analyze the results. We define Relative Degrade as the percentage of accuracy decrease when the model is tested under a domain shift, i.e. the accuracy drop divided by the in-domain accuracy. Specifically, if a model gets accuracy \\(a\\) under in-domain testing (i.e. testing with the same dataset variant as training) and accuracy \\(b\\) under out-of-domain testing (i.e. testing with a different dataset variant from training), then \\(RD = (a - b) / a\\). Since we train each model on three data variants, the RD's for the three models are averaged to measure its domain robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Tab. 2 shows the Relative Degrade of the five models on the four factors. We see that P-NSVQA outperforms other models by a significant margin on three of the four factors, indicating that it has better overall domain robustness. In the following, we take a closer look at the results on each of the factors separately, to diagnose the influence of different model designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Question redundancy. Neural modular methods are much more robust to question redundancy shifts than non-modular ones. The relative degrades for modular methods are less than \\(2\\%\\), while one-modular ones degrade for around \\(20\\%\\). Due to the step-by-step design of the reasoning in modular methods, each reasoning step is independent of the others so that the models are less likely to learn the spurious correlation between question and answers. There-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1For concept distributions, we compute relative degrade with a slight change: we compute the accuracy drop from head to tail and the drop from long to oppo, take their average, and divide by the accuracy on bal." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14968" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.078, + 0.892, + 0.484 + ], + "angle": 0, + "content": "
FiLMmDETRNSCLNSVQAProb NSVQA
Visual Complexity
easymidhardeasymidhardeasymidhardeasymidhardeasymidhard
easy59.9653.9550.6693.3684.3082.9795.1392.3190.8195.1994.1994.0996.7695.9896.37
mid57.4153.2850.1883.3482.3681.2784.589.1086.3381.9992.8093.7886.2595.7695.11
hard55.9553.1150.4779.7179.9480.7176.8578.6685.0873.1179.7192.6579.8186.4795.36
Question Redundancy
rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+
rd-51.4252.5453.5183.9480.3766.2888.6488.8290.3392.9592.9492.6795.6695.7295.43
rd-50.3953.2854.7882.7782.3670.3688.4589.1091.4591.1992.7892.1494.8795.7295.43
rd+46.1452.3071.4778.4884.0590.4287.9488.3491.1691.3891.9692.8094.8895.4795.72
Concept Distribution
balsltlongbalsltlongbalsltlongbalsltlongbalsltlong
bal50.4753.0454.3580.7175.7974.5485.0883.7975.1092.6590.8283.7495.3694.8989.88
long49.4354.7562.9679.0680.2990.6685.3389.4291.1092.7393.3892.5396.3196.3295.25
head48.6058.0661.6080.7579.6087.4684.5888.3990.1993.8794.8292.4896.4296.8095.92
tail51.8048.7050.0881.5070.8860.9486.1080.2760.5590.2689.2075.3294.0893.2082.68
oppo49.0648.9346.6879.1368.3756.9885.0777.8655.1491.2288.6571.3295.7694.0979.74
Concept Compositionality
co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2
co-053.2857.0056.183.3677.0382.4389.182.5283.7792.8090.1191.5995.7694.0295.12
co-152.4160.5756.6779.4682.4583.9378.8987.1884.278.7489.9990.6787.1294.5394.78
co-252.9657.3760.5380.0377.4187.2478.4081.5588.8477.8589.2892.2387.1993.4995.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.49, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Table 1. Accuracy of models trained and tested on different domains. Column headings indicate training settings, while rows indicate the dataset variant for testing. The best performance in each row (i.e. the best training setting) is marked in bold and best performance in each column (i.e. the best testing setting) is underlined. Description for different splits is in Sec. 3.3 and analysis is in Sec. 5.2." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.55, + 0.466, + 0.664 + ], + "angle": 0, + "content": "
VisualRedund.Dist.Comp.
FiLM4.0321.3328.469.04
mDETR9.8119.0536.349.45
NSCL15.570.9237.4415.40
NSVQA17.481.7220.9211.44
Prob NSVQA12.880.8413.727.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Table 2. Relative Degrade under domain shifts, i.e. the percentage of accuracy decrease when the model is tested with domain that differs with training. Lower \\( RD \\) means better robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.76 + ], + "angle": 0, + "content": "fore the modular methods are less vulnerable to change in question/program length." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Visual complexity. Different from our findings on question redundancy, for domain shifts in visual complexity, non-modular methods are more robust compared to modular ones. As shown in Tab. 2, while FiLM and mDETR gets less than \\(10\\%\\) degrade, NSCL and (P-)NSVQA degrade for more than \\(12\\%\\). The reason might be that the simple reasoning modules in modular methods can not process the visual signals as well as the dense non-modular models." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.885, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparing P-NSVQA with NSVQA, we find that in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.644 + ], + "angle": 0, + "content": "jecting probability into deterministic symbolic reasoning greatly improves the robustness on visual complexity (4.04% decrease in \\( RD \\)). This suggests that some errors in visual understanding can be corrected and recovered by taking into account the uncertainty of visual parsing and combining the results of each reasoning step with probability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.795 + ], + "angle": 0, + "content": "Concept distribution. While all the four existing models suffer a lot (larger than \\(20\\%\\) RD) on domain shifts in concept distribution, we see that the symbolic method NSVQA is better than the other three (by more than \\(7.5\\%\\)). With the disentangled reasoning and visual understanding components in NSVQA, the distribution priors in the images and the programs/answers cannot intertwine with each other, which prevent the model heavily relying on the priors. With uncertainty, we can further boost the robustness of NSVQA with a large margin (from \\(21\\%\\) to \\(14\\%\\) RD)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Moreover, the head-tail results suggests that the overall accuracy, which is commonly used to measure VQA performance, should be taken with cautious. When the testing split is imbalanced, the seemingly high accuracy is misleading because the head concepts dominate the testing while the tail ones are not well-reflected. For example, for NSCL, although it gets high accuracy (91%) on the long-tailed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14969" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "data, its performance is only \\(60.6\\%\\) on the tail concepts. In real-world datasets, the data are usually not well-balanced, which suggests the value of synthetic testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.243 + ], + "angle": 0, + "content": "Concept compositionality. Comparing the existing methods, we find that the non-modular methods seems to be more robust than modular methods NSCL or NSVQA. However, with uncertainty, P-NSVQA improves the result of NSVQA, which even outperforms the non-modular methods. This suggest the large potential of better robustness of modular methods by improving current models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.468, + 0.409 + ], + "angle": 0, + "content": "In summary, while non-modular methods are more robust to visual complexity shifts, the modular symbolic methods (improved with uncertainty) are more robust on the other three factors. By disentangling reasoning with visual understanding, separately executing every each reasoning step then merging the results of the steps using probabilities based on uncertainty, our P-NSVQA outperforms all the existing models in question redundancy, concept distribution and compositionality. Therefore, we suggest that symbolic reasoning with uncertainty leads to strong VQA models that are robust to domain shifts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.35, + 0.437 + ], + "angle": 0, + "content": "5.3. More analysis and future work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.468, + 0.732 + ], + "angle": 0, + "content": "Synthetic-to-real transfer. We provide an additional proof-of-concept study to show that the findings drawn from Super-CLEVR dataset can transfer to real datasets. In the following experiments, we show our finding that neuro-symbolic methods (NSCL, NSVQA, P-NSVQA) are more robust than mDETR on question redundancy also holds true on the real GQA dataset [25]. More precisely, we progressively removed the redundant operations from the reasoning program in GQA testdev split, and then regenerated questions using a program-to-question generator. Using the change of models' testing accuracy as the redundant operations are removed, we can evaluate the models' robustness towards question redundancy. The results are show in Tab. 3.2 We observe that the performance drop of mDETR is much larger than neuro-symbolic methods as the redundant information is progressively removed, which indicates that symbolic methods have better question redundancy than mDETR on GQA dataset. This is consistent with our findings on Super-CLEVR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Reasoning with part-object hierarchies. In addition to evaluating domain generalization, Super-CLEVR can be extended for broader purposes, e.g. part-based reasoning. We can ask questions like \"what is the color of the front wheel of the bike?\", \"what is the color of the vehicle that has a yellow wheel\", etc. Those questions require the model to correctly understanding the part-object hierarchy, which" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.172 + ], + "angle": 0, + "content": "
0%14%32%70%91%100%
mDETR0-4.82-8.46-13.16-13.88-14.56
NSCL0-0.14-0.34-1.09-1.71-2.59
NSVQA0-3.47-4.80-7.01-7.02-7.02
P-NSVQA0-1.93-3.15-5.73-5.91-5.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.174, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Table 3. Accuracy drop on the GQA dataset when redundant information is progressively removed." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.216, + 0.785, + 0.231 + ], + "angle": 0, + "content": "is an ability that current VQA models lack." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.231, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Limitations. The main limitations of our work lie in the synthetic nature of our dataset. Future efforts can be made in collecting better controlled and balanced real datasets for model diagnosis. We emphasize that the purpose of the dataset is for model diagnosis and that models should also be tested on real data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.335, + 0.618, + 0.35 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.36, + 0.892, + 0.678 + ], + "angle": 0, + "content": "We diagnose domain shifts in visual reasoning using a proposed virtual benchmark, Super-CLEVR, where distinct factors can be independently studied with controlled data generation. We evaluate four existing methods and show that all of them struggle with domain shifts, highlighting the importance of out-of-domain testing. Among the evaluated methods, neural modular methods are more robust towards question redundancy. In particular, NSVQA with disentangled perception and reasoning shows better robustness towards distribution and compositionality shifts. We further propose P-NSVQA, which improves NSVQA with uncertainty in the reasoning modules. We show that P-NSVQA outperforms all the existing methods in both in-domain testing and out-of-domain testing. With detailed analysis, our study suggests that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. We hope our analysis may facilitate better understanding of strengths and weaknesses of VQA models and, more broadly, future work might explore using the Super-CLEVR benchmark for other tasks like part-based reasoning." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.687, + 0.655, + 0.703 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.892, + 0.86 + ], + "angle": 0, + "content": "We would like to thank Nils Holzenberger, Kate Sanders, Chenxi Liu, Zihao Xiao, Qing Liu, Reno Kriz, and David Etter for their helpful comments and suggestions, as well as the anonymous reviewers. Zhuowan Li is supported by ONR N00014-21-1-2812 and grant IAA80052272 from the Institute for Assured Autonomy at JHU. Elias Stengel-Eskin is supported by an NSF GRFP. A. Kortylewski acknowledges support via his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No.468670075." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.468, + 0.9 + ], + "angle": 0, + "content": "2For implementation of NSCL on a real-word dataset, we use the model in [39] (the version without calibration). The model accuracies on the original not-perturbed GQA testdev split are as following: mDETR \\((61.67\\%)\\) NSCL \\((56.13\\%)\\) NSVQA \\((39.58\\%)\\) P-NSVQA \\((39.66\\%)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14970" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4971-4980, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[3] Arjun Akula, Soravit Changpinyo, Boqing Gong, Piyush Sharma, Song-Chun Zhu, and Radu Soricut. Crossvqa: scalably generating benchmarks for systematically testing vqa generalization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2148-2166, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[5] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.535, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[7] Remi Cadene, Corentin Dancette, Matthieu Cord, Devi Parikh, et al. Rubi: Reducing unimodal biases for visual question answering. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[9] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[11] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1554–1563, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] William Ford and David Olson. The elaboration of the noun phrase in children's description of objects. Journal of Experimental Child Psychology, 19(3):371-382, 1975. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. arXiv preprint arXiv:2204.02285, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[18] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[20] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6930-6940, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[22] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[23] Yining Hong, Li Yi, Josh Tenenbaum, Antonio Torralba, and Chuang Gan. *Ptr: A benchmark for part-based conceptual, relational, and physical reasoning*. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[24] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Jingjing Jiang, Ziyi Liu, Yifan Liu, Zhixiong Nan, and Nanning Zheng. X-ggm: Graph generative modeling for out-of-distribution generalization in visual question answering. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14971" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the 29th ACM International Conference on Multimedia, pages 199-208, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[28] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 3-12. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. arXiv preprint arXiv:2104.12763, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[30] Corentin Kervadec, Grigory Antipov, Moez Baccouche, and Christian Wolf. Roses are red, violets are blue... but should vqa expect them to? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2776-2785, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[31] Corentin Kervadec, Theo Jaunet, Grigory Antipov, Moez Baccouche, Romain Vuillemot, and Christian Wolf. How transferable are reasoning patterns in vqa? 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4205-4214, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[32] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[33] Ruud Koolen, Martijn Goudbeek, and Emiel Krahmer. Effects of scene variation on referential overspecification. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 33, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[34] Satwik Kottur, José MF Moura, Devi Parikh, Dhruv Batra, and Marcus Rohrbach. Clevr-dialog: A diagnostic dataset for multi-round reasoning in visual dialog. arXiv preprint arXiv:1903.03166, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[36] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[37] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10313-10322, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[39] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14910-14919, October 2021. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[40] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning part segmentation through unsupervised domain adaptation from synthetic vehicles. In CVPR, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[41] Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[42] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[43] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[45] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision. In International Conference on Learning Representations, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[46] Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[47] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12695–12705, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[48] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12700-12710, June 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[49] Thomas Pechmann. Incremental speech production and referential overspecification. 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14972" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[51] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[52] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[53] Leonard Salewski, A Koepke, Hendrik Lensch, and Zeynep Akata. Clevr-x: A visual reasoning dataset for natural language explanations. In International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers, pages 69-88. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[54] Susan Sonnenschein. The development of referential communication skills: Some situations in which speakers give redundant messages. Journal of Psycholinguistic Research, 14(5):489-508, 1985. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.363 + ], + "angle": 0, + "content": "[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[56] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[57] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.561 + ], + "angle": 0, + "content": "[58] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[59] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Joshua B Tenenbaum. Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding. In Advances in Neural Information Processing Systems (NIPS), 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[60] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6281–6290, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[61] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5317-5327, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[62] Mingda Zhang, Tristan Maidment, Ahmad Diab, Adriana Kovashka, and Rebecca Hwa. Domain-robust vqa with diverse datasets and methods but no target labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7046–7056, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.149 + ], + "angle": 0, + "content": "[64] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[65] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14973" + } + ] +] \ No newline at end of file diff --git a/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_origin.pdf b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a6b124efa827d16b602f6f5c882f412a8678f8c --- /dev/null +++ b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/e3182ce7-71fa-4a83-aeb4-c095eaebb6c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d216c267c808046f5d4a6a11bde3033fb14c55fb47996dfc79954d0eafb8fce6 +size 1623545 diff --git a/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/full.md b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f62c12db296465ebdc095976f970d769e0d6bb36 --- /dev/null +++ b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/full.md @@ -0,0 +1,297 @@ +# Super-CLEVR: A Virtual Benchmark to Diagnose Domain Robustness in Visual Reasoning + +Zhuowan Li $^{1}$ Xingrui Wang $^{2}$ Elias Stengel-Eskin $^{1}$ Adam Kortylewski $^{3,4}$ Wufei Ma $^{1}$ Benjamin Van Durme $^{1}$ Alan Yuille $^{1}$ $^{1}$ Johns Hopkins University $^{2}$ University of Southern California + $^{3}$ Max Planck Institute for Informatics $^{4}$ University of Freiburg + +# Abstract + +Visual Question Answering (VQA) models often perform poorly on out-of-distribution data and struggle on domain generalization. Due to the multi-modal nature of this task, multiple factors of variation are intertwined, making generalization difficult to analyze. This motivates us to introduce a virtual benchmark, Super-CLEVR, where different factors in VQA domain shifts can be isolated in order that their effects can be studied independently. Four factors are considered: visual complexity, question redundancy, concept distribution and concept compositionality. With controllably generated data, Super-CLEVR enables us to test VQA methods in situations where the test data differs from the training data along each of these axes. We study four existing methods, including two neural symbolic methods NSCL [45] and NSVQA [59], and two non-symbolic methods FiLM [50] and mDETR [29]; and our proposed method, probabilistic NSVQA (P-NSVQA), which extends NSVQA with uncertainty reasoning. P-NSVQA outperforms other methods on three of the four domain shift factors. Our results suggest that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. The dataset and code are released at https://github.com/Lizw14/Super-CLEVR. + +# 1. Introduction + +Visual question answering (VQA) is a challenging task that assesses the reasoning ability of models to answer questions based on both visual and linguistic inputs. Current VQA methods are typically developed on standard benchmarks like VQAv2 [16] or GQA [25], with the implicit assumption that testing data comes from the same underlying distribution as training data. However, as has been widely studied in computer vision [15,36,51], algorithms trained on one domain often fail to generalize to other domains. + +Moreover, having learned the distributional prior of training data, models often struggle on out-of-distribution tests. This has been studied in VQA from the perspective of domain transfer [8,57,62], dataset bias [2,11,48], counter-factual diagnosis [9,47], and out-of-distribution benchmarking [30]. + +The multi-modal nature of VQA gives rise to multiple intertwined factors of variation, making domain shift an especially difficult problem to study. For example, [8] suggests that VQA domain shifts are a combination of differences in images, questions or answers; and [39] reveals a gap between synthetic and real VQA datasets by differences in the over-specification of questions and the underlying distribution of concepts. However, despite a wealth of research on domain generalization in VQA [3,26,57,62], there is no systematic analysis of the contributing factors in domain shifts. + +To this end, we introduce a virtual benchmark, SuperCLEVR, which enables us to test VQA algorithms in situations where the test data differs from the training data. We decompose the domain shift into a set of isolated contributing factors, so that their effects can be diagnosed independently. We study four factors: visual complexity, question redundancy, concept distribution, and concept compositionality. These are illustrated in Fig. 1 and described in Sec. 3.1. With controllable data generation using our SuperCLEVR virtual benchmark, we are able to isolate the different factors in VQA domain shifts so that their effects can be studied independently. Compared with the original CLEVR dataset [27], Super-CLEVR contains more complicated visual components and has better controllability over the domain shift factors. As shown in Fig. 1, the Super-CLEVR dataset contains images rendered from 3D graphical vehicle models in the UDA-Part dataset [40], paired with questions and answers automatically generated from templates. The objects and questions are sampled based on the specified underlying probability distribution, which can be controlled to produce distribution shifts in different factors. + +With Super-CLEVR, we diagnose the domain robustness of current VQA models. Four representative models are studied: for the classic two-stream feature fusing + +![](images/1f0dce66fed1595f30b56d6ba88aa8155a34bdbd2d3fca4679910dd5f06bb79b.jpg) +Figure 1. We decompose VQA domain shifts into four contributing factors: visual complexity, question redundancy, concept distribution and concept compositionality. The domain shifts along each factor can be independently studied with the proposed Super-CLEVR dataset. + +architecture, we choose FiLM [50]; for a large-scale pretrained model we take mDETR [29]; we use NSCL [45] and NSVQA [59] as representative neuro-symbolic methods. We observe that all these models suffer from domain shifts to varying degrees of sensitivity. We analyze each factor separately to examine the influence of different model designs. Specifically, we find that the step-by-step design of neural modular methods enhances their robustness to changes in question redundancy compared with non-modular ones; however, the non-modular models are more robust to visual complexity. Furthermore, thanks to its decomposed reasoning and perception, NSVQA is more robust to concept distribution shifts. + +While existing models suffer from domain shifts with different characteristics, we make a technical improvement over NSVQA which enables it to significantly outperform existing models on three of the four factors. In particular, we inject probabilities into the deterministic symbolic executor of NSVQA, empowering it to take into account the uncertainty of scene understanding. We name our model probabilistic NSVQA (P-NSVQA), and show that its performance improvement in both the in-domain and out-of-domain settings. With superior results of P-NSVQA, we suggest that disentangling reasoning from vision and language understanding, together with probabilistic uncertainty, gives a strong model that is robust to domain shifts. + +Our contributions are as follows. (1) We introduce the Super-CLEVR benchmark to diagnose VQA robustness along four different factors independently. This benchmark can also be used for part-based reasoning. (2) We enhance a neural-symbolic method by taking the uncertainty of visual understanding into account in reasoning. (3) We conduct detailed analysis of four existing methods, as well as our novel approach to study the influence of model designs on + +distinct robustness factors. We conclude that disentangled reasoning and perception plus explicit modeling of uncertainty leads to a more robust VQA model. + +# 2. Related work + +Visual question answering (VQA). Popular VQA methods fall into three categories. Two-stream methods extract features for image and questions using CNN and LSTM respectively, then enable interaction between the two modalities with different feature fusing methods [4, 13, 24, 32, 37, 50, 60]. Neural symbolic methods, on the other hand, use a parse-then-execute pipeline where the question is parsed into a functional program, which is then executed on the image using neural modules [45, 59]. Recently, transformers-based models have achieved impressive performance on various vision-and-language tasks by pretraining on large scale dataset then finetuning for downstream tasks [29, 38, 44, 55, 63]. We choose FiLM [50], NSCL [45] and mDETR [29] as category representatives. + +VQA datasets. Datasets containing real images and human-written questions have been widely used to benchmark VQA models, e.g. VQA [5], VQAv2 [16], Visual 7w [65], VizWiz [18], Visual Genome [35], COCO QA [52], etc. However, subsequent work has revealed the strong prior and bias in those datasets which might be exploited by models to correctly predict the answers without reasoning [1, 16, 17, 30, 31, 47, 48]. Attempts to address this problem include better balancing datasets [2] and creating counterfactual examples [9, 11]. To assess a model's true reasoning ability, the CLEVR dataset [27] proposes to generate complex multi-step questions on synthetic images, which is then extended to various vision-and-language tasks [6,23,34,41,53,58,61]. The GQA dataset [25] extends CLEVR-style questions to real images. Our benchmark is + +distinct from existing ones because we introduce more complex visual scenes into CLEVR and provide controllability to study domain robustness on isolated factors. + +Domain shift in VQA. Domain shift is a long-standing challenge in computer vision, explore in prior works in domain adaptation [14, 15, 22, 43] and domain generalization [36, 51]. Recent works have focused on domain shifts in VQA. [8, 57] improves model adaptation between datasets by feature learning. [62] analyze domain shifts between nine popular VQA datasets and proposes an unsupervised method to bridge the gaps. [39] generalize symbolic reasoning from synthetic to real dataset. [3] introduce a question-answer generation module that simulates the domain shifts. [26] propose a training scheme X-GGM to improve out-of-distribution generalization. [6] assess models generalization on the CLOSURE of linguistic components. In contrast to prior works we study each of the different domain shift factors independently with our virtual benchmark. + +# 3. Super-CLEVR + +# 3.1. Motivation: domain shift factors + +Visual complexity. A major difference between different VQA datasets is visual complexity. For example, in the CLEVR dataset, objects are simple, atomic shapes while in real-world data, objects are more complex and have hierarchical parts. While hard to quantify, visual complexity is related to various factors, such as object variety, object size, background, texture, lighting, occlusion, view point, etc. In our work, we control visual complexity by introducing more challenging objects that can have distinct attributes associated with their parts, and by optionally pasting various textures onto objects. Examples of generated images with different complexity levels are shown in Fig. 1. + +Question redundancy. Question redundancy refers to the amount of over-specified or redundant information in the question, which can be in the form of either attributes or relationships. For example, in Fig. 1, "what color is the large bus behind the cyan car", large (attributes) and behind the cyan car (relationship) are redundant because there is only one bus in the image. As observed in linguistics and cognitive science [12, 33, 49, 54], human speakers may include over-specified information when identifying a target object, which has also been studied in referring expression generation [46]. For VQA, as analyzed in [39], a significant difference between synthetic and real datasets is that real questions contain some redundant information, which sometimes is a distraction leading to model prediction errors. Therefore, in this work, we generate questions with different redundancy levels and study the effect of question redundancy on model behaviors. + +Concept distribution. The distributions of concept, i.e. objects (e.g. car) and attributes (e.g. large), are distinct + +Object with parts: + +![](images/0d554702fdd1b286096780825b6cbc3e072b38b459d927da85bb575bbc36bf38.jpg) +car + +![](images/3d881699209d1c12c4778b0f3a10a1c295f69af38f17380a51c815890ff0f426.jpg) +bus + +![](images/fc8dcd72b3ede28d4de876a2e2bbb94111f416073a9345cb1c8b56392cac2f5b.jpg) +airplane +Figure 2. Super-CLEVR contains 21 vehicle models belonging to 5 categories, with controllable attributes. + +![](images/f26aca1c5ab30f524deea69bf0a41e0175f46a4b775f5087d1f776d8a9a72b4a.jpg) +bike + +![](images/374244f3d17e2095ac2c499c015ac3be9dd4c5d41f8d1ea33660efebdb3ab56d.jpg) +motorcycle + +Texture: dotted, checkered, stripped, none + +Color: green, gray, brown, yellow, red, purple, cyan, blue + +Size: large, small + +Material: rubber, metal + +across different VQA datasets. For example, while colors are well-balanced in CLEVR dataset, in the GQA dataset, the color distribution is long-tailed where "white" appears $>50$ times more frequently than "gold". Long-tailed distributions have been a challenge in many computer vision tasks [20, 28, 42, 56, 64]. In VQA, the long-tailed concept distribution not only hinders the learning of infrequent concepts due to few training samples, but also introduces strong biases and priors in the dataset that may mislead the models. For example, "tennis" is the correct answer to most questions with "what sport is ..." [2]. With strong priors in data, it is hard to assess the true reasoning capacity of current models. While previous works address this problem by carefully re-balancing datasets [2], in our work, we controllably vary the concept distribution in our dataset and study model robustness to concept distribution shifts. + +Concept compositionality. Concept compositionality refers to how different concepts (shapes, attributes) compose and co-occur with each other, e.g. roses are usually red while violets are usually blue [30]. Concept compositionality can be viewed as a conditional concept distribution in the context of other concepts. Shifts in concept compositionality impede the generalization of VQA models. For example, the model may fail to recognize a green banana because most bananas are yellow in the training data [7]. Previous works evaluate the out-of-distribution performance by collecting counterfactual testing examples [47]. In our work, we control the compositionality of shapes and colors with an intuitive motivation: if, for example, in the training data, bicycles are red and cars are blue, will the models be able to recognize blue bicycles and red cars in testing? + +# 3.2. Dataset generation + +Super-CLEVR follows a similar data generation pipeline as CLEVR, but with more complex visual components and better control of domain gap factors. We describe the generation procedure below. + +Objects with parts. To improve the visual complexity of CLEVR scenes, we replace the simple shapes (e.g., cube, sphere) in CLEVR dataset with vehicles from UDA-Part dataset [40]. There are 21 vehicle models, belonging to 5 categories: car, motorbike, aeroplane, bus, and bicycle. Each 3D model comes with part annotations, e.g., left front wheel or left right door for car. Examples for the vehicle models are shown in Fig. 2. We remove or merge small parts from the original annotations to avoid severe difficulty in visual understanding. The full object and parts list is in the supplementary material. + +Attributes. Besides the attributes in the original CLEVR dataset, i.e. color, material, size, we optionally add texture as an additional attribute to increase visual complexity. Note that in order to enable part-based questions, the attributes (color or material) of object parts can be different from that of the object. For example, a blue car can have a red wheel or a green door. In this case, the attribute of the holistic object refers to the attribute of its main body (e.g. the blue car has blue frame). + +Scene rendering. Following CLEVR, each scene contains 3 to 10 objects. The objects are placed onto the ground plane with random position and orientation. When placing the objects, we ensure that the objects do not overlap with each other and we avoid severe occlusion by thresholding the number of visible pixels for each object. Random jitters are added to lamp and camera positions. When rendering, we also save the ground-truth bounding boxes and segmentation masks for each of the objects and their parts, which are required when training some of the models. + +Question generation. Super-CLEVR follows similar question generation pipeline in CLEVR, which instantiates question templates using the underlying reasoning program that can be operated on the scene graph. For example, the program select_shape truck) $\rightarrow$ query_color (·) can be instantiated as question "what is the color of the truck". Therefore, redundancy level of questions can be controlled by removing or adding redundant reasoning steps in the underlying reasoning program. + +# 3.3. Controlling the dataset + +To study domain generalization, we generate several variants of the dataset for each of the domain shift factors. The variants of the datasets serve as different data domains to test the model robustness. Here we describe the method for controllably generating the dataset variants. + +Visual complexity. We generate three variants of the dataset with different levels of visual complexity: easy, mid (middle) and hard. The only difference between the 3 versions is visual complexity: for the easy version, objects with different sizes, colors and materials are placed into the scene; for the middle version, we choose 3 parts on each object that are visible and randomly change their attributes; + +for the hard version, we further add random textures to the objects and parts. An example of the 3 dataset versions can be found in Fig. 1. Note that the scene layout and the questions are shared, so that the influence of visual complexity can be isolated and studied independently. + +Question redundancy. Three variants of the dataset with different redundancy levels are generated: $rd-$ , $rd$ (default), $rd+$ . By default ( $rd$ ), as in original CLEVR dataset, the questions contain some redundant attributes resulting from random sampling, while all redundant relationships are removed. In $rd-$ , we also remove all redundant attributes from the questions, leading to no redundancy in the questions. In $rd+$ , we add all possible attributes and relationships into the question, so that questions contain a high level of redundancy. For all the variants, the questions are ensured to be valid. + +Concept distribution. We generate three dataset variants with different concept distributions: bal (balanced), slt (slightly unbalanced) and long (long-tail distributed). More specifically, we change the distribution of shapes, colors and materials while the distribution of size is kept fixed in order to keep visual complexity consistent, since objects with smaller sizes are visually harder to recognize. By default (bal), the shapes and attributes are randomly sampled, leading to a balanced distribution. For slt and long, the concept distribution d is generated by $d_{i} = a^{-i}$ , where $i$ is the index of the concept. $a$ is a hyper-parameter controlling the length of the tail. A larger $a$ leads to more imbalanced distribution and $a = 1$ leads to flat distribution (cf. Fig. 1). For slt, $a = 1.3$ ; for long, $a = 2.0$ . In addition, to better analyze the performance on the frequent and rare concepts, we generate three variants for testing purpose only: head (frequent concepts in the long-tail distribution), tail (infrequent/rare concepts), and oppo (opposite to the long-tail distribution). We test each model on those three variants to analyze the performance on concepts with different degrees of frequency. + +Concept compositionality. We generate 3 versions of the dataset, co-0, co-1 and co-2, with different compositions of the 21 shapes (from 5 categories) and the 8 colors. The compositionality of the dataset is controlled with the codistribution matrix $M \in \mathbb{R}^{21 \times 8}$ , where each entry $M_{ij}$ is the probability an object of the $i$ -th shape has the $j$ -th color. Entries in each row of $M$ sum up to 1. In the version co-0, $M$ is a flat matrix so that the shapes and colors are randomly composed. In co-1, each shape in one category has a different color distribution, e.g. truck and sedan, while shapes from different categories may share the same color distribution e.g. sedan and airliner. Oppositely, in co-2, we make the shapes in same category have the same color distribution, while shapes from different categories have different distributions. The motivation is that since shapes from the same category are visually similar, the difference in co-1 + +and co-2 will help analyze the difference in model predictions on visually similar objects and dissimilar objects when composed with different color distributions. + +Dataset Statistics. Every dataset variant contains 30k images, including 20k for training, 5k for validation and 5k for testing. Each image is paired with 10 object-based and 10 part-based questions. By default, the dataset refers to the version with mid visual complexity level (i.e. objects are untextured and has up to 3 parts with distinct attributes), rd redundancy level, balanced (bal) concept distribution and random (co-0) compositionality. More dataset statistics are in supplementary materials. + +# 4. Evaluated methods + +# 4.1. Simple baselines + +Random, Majority. These simple baselines pick a random or the most frequent answer for each question type in the training set as the predicted answer. + +LSTM. This question-only baseline encodes the question word embeddings with LSTM [21] and predicts answer with an MLP on top of the final hidden states of the LSTM. + +CNN+LSTM. The image is represented with features extracted by CNN and question is encoded by LSTM. An MLP predicts answer scores based on the concatenation of image and question features. + +# 4.2. Existing models + +FiLM. We choose Feature-wise Linear Modulation [50] as a representative of classic two-stream feature merging methods. The question features extracted with GRU [10] and image features extracted with CNN are fused with the proposed FiLM module. + +mDETR. mDETR [29] is a transformer-based detector trained to detects objects in an image conditioned on a text query. The model is pretrained with 1.3M image and text pairs and can be finetuned for various downstream tasks like referring expression understanding or VQA. + +NSCL. The Neuro-Symbolic Concept Learner [45] is a representative neural symbolic method. NSCL executes neural modules on the scene representation based on the reasoning program, during which the modules learn embeddings of each concept with the answer supervision. + +NSVQA. Neural-Symbolic VQA [59] is a neural symbolic method composed of three components: A scene parser (Mask-RCNN [19]) that segments an input image and recovers a structural scene representation, a question parser that converts a question from natural language into a program and a program executor that runs the program on the structural scene representation to obtain the answer. Notably, compared to NSCL, the individual components of NSVQA can be learned separately, hence, for example the scene parser can be learned from data that does not necessarily have Visual-Question annotations. + +# 4.3. Probabilistic NSVQA (P-NSVQA) + +Since the program executor in NSVQA is a collection of deterministic, generic functional modules, it can be augmented with a probabilistic reasoning process that takes into account the confidence of the predictions of the scene parser. This allows the model to execute the program that has the largest joint likelihood, instead of only taking the maximal likelihood execution at each step of the program. The experiment results demonstrate a significant performance improvement of this probabilistic approach over the deterministic NSVQA model proposed in [59]. + +In particular, we interpret the confidence of the Mask-RCNN output as a likelihood function for all detected object classes $p_{\text{object}}$ and their attributes $p_{\text{att}}$ . Moreover, we define a likelihood $p_{\text{spatial}}$ for the spatial relations between objects (behind, in front, left, right) that is proportional to the distance between the centers of two bounding boxes. Given a reasoning program containing multiple reasoning steps, we execute each step based on the scene parsing likelihood and produce an step-wise output with confidence. Finally, we use a factorized model, multiplying the output for all the steps to get the final answer prediction. We refer readers to the Appendix for more details. + +# 4.4. Implementation details + +Training mDETR requires ground-truth grounding of question tokens to image regions, which is available in Super-CLEVR. NSCL requires bounding box of objects, which can be predicted using a trained Faster RCNN, and the reasoning program, which can be parsed using a trained parser. Similarly, ground-truth programs are used for training NSVQA and P-NSVQA. Note that we empirically find that the question-to-program parsing is a relatively easy task ( $>99\%$ accuracy using a simple LSTM), so we focus more on models' reasoning ability in our analysis. + +Unless specified, the models are trained with default setting as in the official implementation. FiLM is trained for 100k iterations with batch size 256. mDETR is trained for 30 epochs with batch size 64 using 2 GPUs for both the grounding stage and the answer classification stage. NSCL is trained for 80 epochs with batch size 32. For NSVQA and P-NSVQA, we first train the object parser (Mask RCNN [19]) for 30k iterations with batch size 16, then train the attribute extraction model (using the Res50 backbone) for 100 epochs with batch size 64. For P-NSVQA, when counting the objects or determining whether objects exist in the scene, we use a threshold (0.7) to obtain the final selected objects. Early stopping is used based on validation accuracy. All the models are trained with 200k questions. We repeat experiments on default split for 3 times with different random seeds and get std of $\pm 0.10$ (P-NSVQA) and $\pm 0.40$ (NSVQA), showing the statistical significance of our results, then we only run other experiments once. + +# 5. Results and analysis + +In this section, we first show evaluation results for indomain setting, then provide results and analysis on out-of-domain evaluation. Finally, we describe additional studies and future works. + +# 5.1. In-domain results + +In-domain evaluation refers to the setting where the training and testing data come from the same domain (the default dataset variant in this case). We compare the indomain results on Super-CLEVR and CLEVR. The results are shown in Fig. 3. + +![](images/aaa9891acc31869c8baa353199fd99d14b1ff7b1e2b49763bd49e710dfe1370b.jpg) +Figure 3. Comparison of models' accuracy on Super-CLEVR and the original CLEVR dataset. + +For all the models, the performance is lower on SuperCLEVR than CLEVR, suggesting that Super-CLEVR is a more challenging benchmark. The scenes with vehicles are much harder visually for the models to understand compared with the simpler shapes from CLEVR. Note that the performance gap on two datasets for simple baselines (Random, Majority, LSTM, CNN+LSTM) is smaller than for the other better models. This is because Super-CLEVR contains more object types, and therefore the performance of simply guessing is lower than on CLEVR. + +When comparing the performance of different models, we find that the neural modular methods, i.e. NSCL, NSVQA, P-NSVQA, perform much better than non-modular ones. This is not surprising given their nearly perfect performance on the original CLEVR dataset, which shows its strong ability to model synthetic images. The large-scale pretrained grounding model mDETR, which is a leading model on both real and synthetic images, also achieves good performance $(82.7\%)$ on Super-CLEVR. The two-stream method FiLM does not achieve very strong performance $(53.2\%)$ , but is still much better than the other simple baselines. + +Our proposed P-NSVQA outperforms all the other models. In particular, on Super-CLEVR, it outperforms its de + +terministic counterpart, NSVQA, by $2.96\%$ . This shows the advantage of taking into account the probabilities when the scenes are challenging thus the model's uncertainty of predictions can be utilized. + +# 5.2. Out-of-domain results + +In this section, we train and test the five models (FiLM, mDETR, NSCL, NSVQA and P-NSVQA) on different dataset variants, and diagnose their domain robustness on each of the four domain shift factors. Please refer to Sec. 3.3 for a description of different variants. The validation accuracy is used for analysis here and the results are shown in Tab. 1. + +All the methods suffer from domain shifts. The results show that the best performance mostly occurs in situations where the model is tested on the same dataset variant as it is trained on, i.e. the bold or underlined numbers fall mostly on the diagonals in Tab. 1. + +We compare the domain robustness of the five models by measuring the relative performance decrease when the testing data differs from the training data, i.e. smaller performance drop on different testing domains means better robustness. Based on this intuition, for easier understanding of Tab. 1, we propose a measurement metric for domain robustness named Relative Degrade (RD) to better analyze the results. We define Relative Degrade as the percentage of accuracy decrease when the model is tested under a domain shift, i.e. the accuracy drop divided by the in-domain accuracy. Specifically, if a model gets accuracy $a$ under in-domain testing (i.e. testing with the same dataset variant as training) and accuracy $b$ under out-of-domain testing (i.e. testing with a different dataset variant from training), then $RD = (a - b) / a$ . Since we train each model on three data variants, the RD's for the three models are averaged to measure its domain robustness. + +Tab. 2 shows the Relative Degrade of the five models on the four factors. We see that P-NSVQA outperforms other models by a significant margin on three of the four factors, indicating that it has better overall domain robustness. In the following, we take a closer look at the results on each of the factors separately, to diagnose the influence of different model designs. + +Question redundancy. Neural modular methods are much more robust to question redundancy shifts than non-modular ones. The relative degrades for modular methods are less than $2\%$ , while one-modular ones degrade for around $20\%$ . Due to the step-by-step design of the reasoning in modular methods, each reasoning step is independent of the others so that the models are less likely to learn the spurious correlation between question and answers. There- + +
FiLMmDETRNSCLNSVQAProb NSVQA
Visual Complexity
easymidhardeasymidhardeasymidhardeasymidhardeasymidhard
easy59.9653.9550.6693.3684.3082.9795.1392.3190.8195.1994.1994.0996.7695.9896.37
mid57.4153.2850.1883.3482.3681.2784.589.1086.3381.9992.8093.7886.2595.7695.11
hard55.9553.1150.4779.7179.9480.7176.8578.6685.0873.1179.7192.6579.8186.4795.36
Question Redundancy
rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+
rd-51.4252.5453.5183.9480.3766.2888.6488.8290.3392.9592.9492.6795.6695.7295.43
rd-50.3953.2854.7882.7782.3670.3688.4589.1091.4591.1992.7892.1494.8795.7295.43
rd+46.1452.3071.4778.4884.0590.4287.9488.3491.1691.3891.9692.8094.8895.4795.72
Concept Distribution
balsltlongbalsltlongbalsltlongbalsltlongbalsltlong
bal50.4753.0454.3580.7175.7974.5485.0883.7975.1092.6590.8283.7495.3694.8989.88
long49.4354.7562.9679.0680.2990.6685.3389.4291.1092.7393.3892.5396.3196.3295.25
head48.6058.0661.6080.7579.6087.4684.5888.3990.1993.8794.8292.4896.4296.8095.92
tail51.8048.7050.0881.5070.8860.9486.1080.2760.5590.2689.2075.3294.0893.2082.68
oppo49.0648.9346.6879.1368.3756.9885.0777.8655.1491.2288.6571.3295.7694.0979.74
Concept Compositionality
co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2
co-053.2857.0056.183.3677.0382.4389.182.5283.7792.8090.1191.5995.7694.0295.12
co-152.4160.5756.6779.4682.4583.9378.8987.1884.278.7489.9990.6787.1294.5394.78
co-252.9657.3760.5380.0377.4187.2478.4081.5588.8477.8589.2892.2387.1993.4995.61
+ +Table 1. Accuracy of models trained and tested on different domains. Column headings indicate training settings, while rows indicate the dataset variant for testing. The best performance in each row (i.e. the best training setting) is marked in bold and best performance in each column (i.e. the best testing setting) is underlined. Description for different splits is in Sec. 3.3 and analysis is in Sec. 5.2. + +
VisualRedund.Dist.Comp.
FiLM4.0321.3328.469.04
mDETR9.8119.0536.349.45
NSCL15.570.9237.4415.40
NSVQA17.481.7220.9211.44
Prob NSVQA12.880.8413.727.00
+ +Table 2. Relative Degrade under domain shifts, i.e. the percentage of accuracy decrease when the model is tested with domain that differs with training. Lower $RD$ means better robustness. + +fore the modular methods are less vulnerable to change in question/program length. + +Visual complexity. Different from our findings on question redundancy, for domain shifts in visual complexity, non-modular methods are more robust compared to modular ones. As shown in Tab. 2, while FiLM and mDETR gets less than $10\%$ degrade, NSCL and (P-)NSVQA degrade for more than $12\%$ . The reason might be that the simple reasoning modules in modular methods can not process the visual signals as well as the dense non-modular models. + +Comparing P-NSVQA with NSVQA, we find that in + +jecting probability into deterministic symbolic reasoning greatly improves the robustness on visual complexity (4.04% decrease in $RD$ ). This suggests that some errors in visual understanding can be corrected and recovered by taking into account the uncertainty of visual parsing and combining the results of each reasoning step with probability. + +Concept distribution. While all the four existing models suffer a lot (larger than $20\%$ RD) on domain shifts in concept distribution, we see that the symbolic method NSVQA is better than the other three (by more than $7.5\%$ ). With the disentangled reasoning and visual understanding components in NSVQA, the distribution priors in the images and the programs/answers cannot intertwine with each other, which prevent the model heavily relying on the priors. With uncertainty, we can further boost the robustness of NSVQA with a large margin (from $21\%$ to $14\%$ RD). + +Moreover, the head-tail results suggests that the overall accuracy, which is commonly used to measure VQA performance, should be taken with cautious. When the testing split is imbalanced, the seemingly high accuracy is misleading because the head concepts dominate the testing while the tail ones are not well-reflected. For example, for NSCL, although it gets high accuracy (91%) on the long-tailed + +data, its performance is only $60.6\%$ on the tail concepts. In real-world datasets, the data are usually not well-balanced, which suggests the value of synthetic testing. + +Concept compositionality. Comparing the existing methods, we find that the non-modular methods seems to be more robust than modular methods NSCL or NSVQA. However, with uncertainty, P-NSVQA improves the result of NSVQA, which even outperforms the non-modular methods. This suggest the large potential of better robustness of modular methods by improving current models. + +In summary, while non-modular methods are more robust to visual complexity shifts, the modular symbolic methods (improved with uncertainty) are more robust on the other three factors. By disentangling reasoning with visual understanding, separately executing every each reasoning step then merging the results of the steps using probabilities based on uncertainty, our P-NSVQA outperforms all the existing models in question redundancy, concept distribution and compositionality. Therefore, we suggest that symbolic reasoning with uncertainty leads to strong VQA models that are robust to domain shifts. + +# 5.3. More analysis and future work + +Synthetic-to-real transfer. We provide an additional proof-of-concept study to show that the findings drawn from Super-CLEVR dataset can transfer to real datasets. In the following experiments, we show our finding that neuro-symbolic methods (NSCL, NSVQA, P-NSVQA) are more robust than mDETR on question redundancy also holds true on the real GQA dataset [25]. More precisely, we progressively removed the redundant operations from the reasoning program in GQA testdev split, and then regenerated questions using a program-to-question generator. Using the change of models' testing accuracy as the redundant operations are removed, we can evaluate the models' robustness towards question redundancy. The results are show in Tab. 3.2 We observe that the performance drop of mDETR is much larger than neuro-symbolic methods as the redundant information is progressively removed, which indicates that symbolic methods have better question redundancy than mDETR on GQA dataset. This is consistent with our findings on Super-CLEVR. + +Reasoning with part-object hierarchies. In addition to evaluating domain generalization, Super-CLEVR can be extended for broader purposes, e.g. part-based reasoning. We can ask questions like "what is the color of the front wheel of the bike?", "what is the color of the vehicle that has a yellow wheel", etc. Those questions require the model to correctly understanding the part-object hierarchy, which + +
0%14%32%70%91%100%
mDETR0-4.82-8.46-13.16-13.88-14.56
NSCL0-0.14-0.34-1.09-1.71-2.59
NSVQA0-3.47-4.80-7.01-7.02-7.02
P-NSVQA0-1.93-3.15-5.73-5.91-5.78
+ +Table 3. Accuracy drop on the GQA dataset when redundant information is progressively removed. + +is an ability that current VQA models lack. + +Limitations. The main limitations of our work lie in the synthetic nature of our dataset. Future efforts can be made in collecting better controlled and balanced real datasets for model diagnosis. We emphasize that the purpose of the dataset is for model diagnosis and that models should also be tested on real data. + +# 6. Conclusion + +We diagnose domain shifts in visual reasoning using a proposed virtual benchmark, Super-CLEVR, where distinct factors can be independently studied with controlled data generation. We evaluate four existing methods and show that all of them struggle with domain shifts, highlighting the importance of out-of-domain testing. Among the evaluated methods, neural modular methods are more robust towards question redundancy. In particular, NSVQA with disentangled perception and reasoning shows better robustness towards distribution and compositionality shifts. We further propose P-NSVQA, which improves NSVQA with uncertainty in the reasoning modules. We show that P-NSVQA outperforms all the existing methods in both in-domain testing and out-of-domain testing. With detailed analysis, our study suggests that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. We hope our analysis may facilitate better understanding of strengths and weaknesses of VQA models and, more broadly, future work might explore using the Super-CLEVR benchmark for other tasks like part-based reasoning. + +# Acknowledgements + +We would like to thank Nils Holzenberger, Kate Sanders, Chenxi Liu, Zihao Xiao, Qing Liu, Reno Kriz, and David Etter for their helpful comments and suggestions, as well as the anonymous reviewers. Zhuowan Li is supported by ONR N00014-21-1-2812 and grant IAA80052272 from the Institute for Assured Autonomy at JHU. Elias Stengel-Eskin is supported by an NSF GRFP. A. Kortylewski acknowledges support via his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No.468670075. + +# References + +[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2 +[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4971-4980, 2018. 1, 2, 3 +[3] Arjun Akula, Soravit Changpinyo, Boqing Gong, Piyush Sharma, Song-Chun Zhu, and Radu Soricut. Crossvqa: scalably generating benchmarks for systematically testing vqa generalization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2148-2166, 2021. 1, 3 +[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2 +[5] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2 +[6] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019. 2, 3 +[7] Remi Cadene, Corentin Dancette, Matthieu Cord, Devi Parikh, et al. Rubi: Reducing unimodal biases for visual question answering. Advances in neural information processing systems, 32, 2019. 3 +[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 1, 3 +[9] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 1, 2 +[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.5 +[11] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1554–1563, 2021. 1, 2 +[12] William Ford and David Olson. The elaboration of the noun phrase in children's description of objects. Journal of Experimental Child Psychology, 19(3):371-382, 1975. 3 + +[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2 +[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3 +[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 3 +[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2 +[17] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. arXiv preprint arXiv:2204.02285, 2022. 2 +[18] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2 +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5 +[20] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6930-6940, 2021. 3 +[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5 +[22] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. PMLR, 2018. 3 +[23] Yining Hong, Li Yi, Josh Tenenbaum, Antonio Torralba, and Chuang Gan. *Ptr: A benchmark for part-based conceptual, relational, and physical reasoning*. Advances in Neural Information Processing Systems, 34, 2021. 2 +[24] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. 2018. 2 +[25] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 2, 8 +[26] Jingjing Jiang, Ziyi Liu, Yifan Liu, Zhixiong Nan, and Nanning Zheng. X-ggm: Graph generative modeling for out-of-distribution generalization in visual question answering. In + +Proceedings of the 29th ACM International Conference on Multimedia, pages 199-208, 2021. 1, 3 +[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 1, 2 +[28] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 3-12. Springer, 2021. 3 +[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. arXiv preprint arXiv:2104.12763, 2021. 1, 2, 5 +[30] Corentin Kervadec, Grigory Antipov, Moez Baccouche, and Christian Wolf. Roses are red, violets are blue... but should vqa expect them to? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2776-2785, 2021. 1, 2, 3 +[31] Corentin Kervadec, Theo Jaunet, Grigory Antipov, Moez Baccouche, Romain Vuillemot, and Christian Wolf. How transferable are reasoning patterns in vqa? 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4205-4214, 2021. 2 +[32] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. Advances in Neural Information Processing Systems, 31, 2018. 2 +[33] Ruud Koolen, Martijn Goudbeek, and Emiel Krahmer. Effects of scene variation on referential overspecification. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 33, 2011. 3 +[34] Satwik Kottur, José MF Moura, Devi Parikh, Dhruv Batra, and Marcus Rohrbach. Clevr-dialog: A diagnostic dataset for multi-round reasoning in visual dialog. arXiv preprint arXiv:1903.03166, 2019. 2 +[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2 +[36] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1, 3 +[37] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10313-10322, 2019. 2 +[38] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object- + +semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 2 +[39] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14910-14919, October 2021. 1, 3, 8 +[40] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning part segmentation through unsupervised domain adaptation from synthetic vehicles. In CVPR, 2022. 1, 4 +[41] Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194, 2019. 2 +[42] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3 +[43] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Advances in neural information processing systems, 31, 2018. 3 +[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2 +[45] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision. In International Conference on Learning Representations, 2019. 1, 2, 5 +[46] Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184, 2013. 3 +[47] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12695–12705, 2021. 1, 2, 3 +[48] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12700-12710, June 2021. 1, 2 +[49] Thomas Pechmann. Incremental speech production and referential overspecification. 1989. 3 +[50] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5 + +[51] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 3 +[52] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 2 +[53] Leonard Salewski, A Koepke, Hendrik Lensch, and Zeynep Akata. Clevr-x: A visual reasoning dataset for natural language explanations. In International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers, pages 69-88. Springer, 2022. 2 +[54] Susan Sonnenschein. The development of referential communication skills: Some situations in which speakers give redundant messages. Journal of Psycholinguistic Research, 14(5):489-508, 1985. 3 +[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2 +[56] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 3 +[57] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 3 +[58] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2 +[59] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Joshua B Tenenbaum. Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding. In Advances in Neural Information Processing Systems (NIPS), 2018. 1, 2, 5 +[60] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6281–6290, 2019. 2 +[61] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5317-5327, 2019. 2 +[62] Mingda Zhang, Tristan Maidment, Ahmad Diab, Adriana Kovashka, and Rebecca Hwa. Domain-robust vqa with diverse datasets and methods but no target labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7046–7056, 2021. 1, 3 +[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 2 + +[64] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 3 +[65] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 2 \ No newline at end of file diff --git a/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/images.zip b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18f85750f0f4d92f701118eb8d627da6248ef5ec --- /dev/null +++ b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:281fcbdf716d1c061b233a90a8d76966e1429c28bd7611392dfc38575bd0c2c4 +size 358625 diff --git a/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/layout.json b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc30394261400d51ef6b553f015cbe136cc9f913 --- /dev/null +++ b/2023/Super-CLEVR_ A Virtual Benchmark To Diagnose Domain Robustness in Visual Reasoning/layout.json @@ -0,0 +1,8028 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 140 + ], + "type": "text", + "content": "Super-CLEVR: A Virtual Benchmark to Diagnose Domain Robustness in Visual Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "spans": [ + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": "Zhuowan Li" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Xingrui Wang" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Elias Stengel-Eskin" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Adam Kortylewski" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Wufei Ma" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Benjamin Van Durme" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Alan Yuille" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Johns Hopkins University " + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " University of Southern California \n" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " Max Planck Institute for Informatics " + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 95, + 160, + 504, + 218 + ], + "type": "text", + "content": " University of Freiburg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 271, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 559 + ], + "type": "text", + "content": "Visual Question Answering (VQA) models often perform poorly on out-of-distribution data and struggle on domain generalization. Due to the multi-modal nature of this task, multiple factors of variation are intertwined, making generalization difficult to analyze. This motivates us to introduce a virtual benchmark, Super-CLEVR, where different factors in VQA domain shifts can be isolated in order that their effects can be studied independently. Four factors are considered: visual complexity, question redundancy, concept distribution and concept compositionality. With controllably generated data, Super-CLEVR enables us to test VQA methods in situations where the test data differs from the training data along each of these axes. We study four existing methods, including two neural symbolic methods NSCL [45] and NSVQA [59], and two non-symbolic methods FiLM [50] and mDETR [29]; and our proposed method, probabilistic NSVQA (P-NSVQA), which extends NSVQA with uncertainty reasoning. P-NSVQA outperforms other methods on three of the four domain shift factors. Our results suggest that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. The dataset and code are released at https://github.com/Lizw14/Super-CLEVR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 584, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Visual question answering (VQA) is a challenging task that assesses the reasoning ability of models to answer questions based on both visual and linguistic inputs. Current VQA methods are typically developed on standard benchmarks like VQAv2 [16] or GQA [25], with the implicit assumption that testing data comes from the same underlying distribution as training data. However, as has been widely studied in computer vision [15,36,51], algorithms trained on one domain often fail to generalize to other domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 246, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 305 + ], + "type": "text", + "content": "Moreover, having learned the distributional prior of training data, models often struggle on out-of-distribution tests. This has been studied in VQA from the perspective of domain transfer [8,57,62], dataset bias [2,11,48], counter-factual diagnosis [9,47], and out-of-distribution benchmarking [30]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 306, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 546, + 426 + ], + "type": "text", + "content": "The multi-modal nature of VQA gives rise to multiple intertwined factors of variation, making domain shift an especially difficult problem to study. For example, [8] suggests that VQA domain shifts are a combination of differences in images, questions or answers; and [39] reveals a gap between synthetic and real VQA datasets by differences in the over-specification of questions and the underlying distribution of concepts. However, despite a wealth of research on domain generalization in VQA [3,26,57,62], there is no systematic analysis of the contributing factors in domain shifts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 676 + ], + "type": "text", + "content": "To this end, we introduce a virtual benchmark, SuperCLEVR, which enables us to test VQA algorithms in situations where the test data differs from the training data. We decompose the domain shift into a set of isolated contributing factors, so that their effects can be diagnosed independently. We study four factors: visual complexity, question redundancy, concept distribution, and concept compositionality. These are illustrated in Fig. 1 and described in Sec. 3.1. With controllable data generation using our SuperCLEVR virtual benchmark, we are able to isolate the different factors in VQA domain shifts so that their effects can be studied independently. Compared with the original CLEVR dataset [27], Super-CLEVR contains more complicated visual components and has better controllability over the domain shift factors. As shown in Fig. 1, the Super-CLEVR dataset contains images rendered from 3D graphical vehicle models in the UDA-Part dataset [40], paired with questions and answers automatically generated from templates. The objects and questions are sampled based on the specified underlying probability distribution, which can be controlled to produce distribution shifts in different factors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "With Super-CLEVR, we diagnose the domain robustness of current VQA models. Four representative models are studied: for the classic two-stream feature fusing" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14963" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 51, + 534, + 262 + ], + "blocks": [ + { + "bbox": [ + 56, + 51, + 534, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 51, + 534, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 51, + 534, + 262 + ], + "type": "image", + "image_path": "1f0dce66fed1595f30b56d6ba88aa8155a34bdbd2d3fca4679910dd5f06bb79b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 267, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 547, + 291 + ], + "type": "text", + "content": "Figure 1. We decompose VQA domain shifts into four contributing factors: visual complexity, question redundancy, concept distribution and concept compositionality. The domain shifts along each factor can be independently studied with the proposed Super-CLEVR dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 456 + ], + "type": "text", + "content": "architecture, we choose FiLM [50]; for a large-scale pretrained model we take mDETR [29]; we use NSCL [45] and NSVQA [59] as representative neuro-symbolic methods. We observe that all these models suffer from domain shifts to varying degrees of sensitivity. We analyze each factor separately to examine the influence of different model designs. Specifically, we find that the step-by-step design of neural modular methods enhances their robustness to changes in question redundancy compared with non-modular ones; however, the non-modular models are more robust to visual complexity. Furthermore, thanks to its decomposed reasoning and perception, NSVQA is more robust to concept distribution shifts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 459, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 615 + ], + "type": "text", + "content": "While existing models suffer from domain shifts with different characteristics, we make a technical improvement over NSVQA which enables it to significantly outperform existing models on three of the four factors. In particular, we inject probabilities into the deterministic symbolic executor of NSVQA, empowering it to take into account the uncertainty of scene understanding. We name our model probabilistic NSVQA (P-NSVQA), and show that its performance improvement in both the in-domain and out-of-domain settings. With superior results of P-NSVQA, we suggest that disentangling reasoning from vision and language understanding, together with probabilistic uncertainty, gives a strong model that is robust to domain shifts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Our contributions are as follows. (1) We introduce the Super-CLEVR benchmark to diagnose VQA robustness along four different factors independently. This benchmark can also be used for part-based reasoning. (2) We enhance a neural-symbolic method by taking the uncertainty of visual understanding into account in reasoning. (3) We conduct detailed analysis of four existing methods, as well as our novel approach to study the influence of model designs on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 300, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 337 + ], + "type": "text", + "content": "distinct robustness factors. We conclude that disentangled reasoning and perception plus explicit modeling of uncertainty leads to a more robust VQA model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 346, + 391, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 391, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 391, + 358 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 366, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 534 + ], + "type": "text", + "content": "Visual question answering (VQA). Popular VQA methods fall into three categories. Two-stream methods extract features for image and questions using CNN and LSTM respectively, then enable interaction between the two modalities with different feature fusing methods [4, 13, 24, 32, 37, 50, 60]. Neural symbolic methods, on the other hand, use a parse-then-execute pipeline where the question is parsed into a functional program, which is then executed on the image using neural modules [45, 59]. Recently, transformers-based models have achieved impressive performance on various vision-and-language tasks by pretraining on large scale dataset then finetuning for downstream tasks [29, 38, 44, 55, 63]. We choose FiLM [50], NSCL [45] and mDETR [29] as category representatives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "VQA datasets. Datasets containing real images and human-written questions have been widely used to benchmark VQA models, e.g. VQA [5], VQAv2 [16], Visual 7w [65], VizWiz [18], Visual Genome [35], COCO QA [52], etc. However, subsequent work has revealed the strong prior and bias in those datasets which might be exploited by models to correctly predict the answers without reasoning [1, 16, 17, 30, 31, 47, 48]. Attempts to address this problem include better balancing datasets [2] and creating counterfactual examples [9, 11]. To assess a model's true reasoning ability, the CLEVR dataset [27] proposes to generate complex multi-step questions on synthetic images, which is then extended to various vision-and-language tasks [6,23,34,41,53,58,61]. The GQA dataset [25] extends CLEVR-style questions to real images. Our benchmark is" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14964" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "distinct from existing ones because we introduce more complex visual scenes into CLEVR and provide controllability to study domain robustness on isolated factors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 289 + ], + "type": "text", + "content": "Domain shift in VQA. Domain shift is a long-standing challenge in computer vision, explore in prior works in domain adaptation [14, 15, 22, 43] and domain generalization [36, 51]. Recent works have focused on domain shifts in VQA. [8, 57] improves model adaptation between datasets by feature learning. [62] analyze domain shifts between nine popular VQA datasets and proposes an unsupervised method to bridge the gaps. [39] generalize symbolic reasoning from synthetic to real dataset. [3] introduce a question-answer generation module that simulates the domain shifts. [26] propose a training scheme X-GGM to improve out-of-distribution generalization. [6] assess models generalization on the CLOSURE of linguistic components. In contrast to prior works we study each of the different domain shift factors independently with our virtual benchmark." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 301, + 138, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 138, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 138, + 315 + ], + "type": "text", + "content": "3. Super-CLEVR" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 322, + 222, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 222, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 222, + 333 + ], + "type": "text", + "content": "3.1. Motivation: domain shift factors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 484 + ], + "type": "text", + "content": "Visual complexity. A major difference between different VQA datasets is visual complexity. For example, in the CLEVR dataset, objects are simple, atomic shapes while in real-world data, objects are more complex and have hierarchical parts. While hard to quantify, visual complexity is related to various factors, such as object variety, object size, background, texture, lighting, occlusion, view point, etc. In our work, we control visual complexity by introducing more challenging objects that can have distinct attributes associated with their parts, and by optionally pasting various textures onto objects. Examples of generated images with different complexity levels are shown in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": "Question redundancy. Question redundancy refers to the amount of over-specified or redundant information in the question, which can be in the form of either attributes or relationships. For example, in Fig. 1, \"what color is the large bus behind the cyan car\", large (attributes) and behind the cyan car (relationship) are redundant because there is only one bus in the image. As observed in linguistics and cognitive science [12, 33, 49, 54], human speakers may include over-specified information when identifying a target object, which has also been studied in referring expression generation [46]. For VQA, as analyzed in [39], a significant difference between synthetic and real datasets is that real questions contain some redundant information, which sometimes is a distraction leading to model prediction errors. Therefore, in this work, we generate questions with different redundancy levels and study the effect of question redundancy on model behaviors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Concept distribution. The distributions of concept, i.e. objects (e.g. car) and attributes (e.g. large), are distinct" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 69, + 391, + 80 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 69, + 391, + 80 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 391, + 80 + ], + "type": "text", + "content": "Object with parts:" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 82, + 393, + 113 + ], + "blocks": [ + { + "bbox": [ + 331, + 82, + 393, + 113 + ], + "lines": [ + { + "bbox": [ + 331, + 82, + 393, + 113 + ], + "spans": [ + { + "bbox": [ + 331, + 82, + 393, + 113 + ], + "type": "image", + "image_path": "0d554702fdd1b286096780825b6cbc3e072b38b459d927da85bb575bbc36bf38.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 114, + 367, + 121 + ], + "lines": [ + { + "bbox": [ + 354, + 114, + 367, + 121 + ], + "spans": [ + { + "bbox": [ + 354, + 114, + 367, + 121 + ], + "type": "text", + "content": "car" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 398, + 90, + 463, + 113 + ], + "blocks": [ + { + "bbox": [ + 398, + 90, + 463, + 113 + ], + "lines": [ + { + "bbox": [ + 398, + 90, + 463, + 113 + ], + "spans": [ + { + "bbox": [ + 398, + 90, + 463, + 113 + ], + "type": "image", + "image_path": "3d881699209d1c12c4778b0f3a10a1c295f69af38f17380a51c815890ff0f426.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 114, + 433, + 122 + ], + "lines": [ + { + "bbox": [ + 419, + 114, + 433, + 122 + ], + "spans": [ + { + "bbox": [ + 419, + 114, + 433, + 122 + ], + "type": "text", + "content": "bus" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 85, + 533, + 114 + ], + "blocks": [ + { + "bbox": [ + 474, + 85, + 533, + 114 + ], + "lines": [ + { + "bbox": [ + 474, + 85, + 533, + 114 + ], + "spans": [ + { + "bbox": [ + 474, + 85, + 533, + 114 + ], + "type": "image", + "image_path": "fc8dcd72b3ede28d4de876a2e2bbb94111f416073a9345cb1c8b56392cac2f5b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 114, + 517, + 124 + ], + "lines": [ + { + "bbox": [ + 484, + 114, + 517, + 124 + ], + "spans": [ + { + "bbox": [ + 484, + 114, + 517, + 124 + ], + "type": "text", + "content": "airplane" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "type": "text", + "content": "Figure 2. Super-CLEVR contains 21 vehicle models belonging to 5 categories, with controllable attributes." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 329, + 132, + 393, + 168 + ], + "blocks": [ + { + "bbox": [ + 329, + 132, + 393, + 168 + ], + "lines": [ + { + "bbox": [ + 329, + 132, + 393, + 168 + ], + "spans": [ + { + "bbox": [ + 329, + 132, + 393, + 168 + ], + "type": "image", + "image_path": "f26aca1c5ab30f524deea69bf0a41e0175f46a4b775f5087d1f776d8a9a72b4a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 169, + 370, + 177 + ], + "lines": [ + { + "bbox": [ + 353, + 169, + 370, + 177 + ], + "spans": [ + { + "bbox": [ + 353, + 169, + 370, + 177 + ], + "type": "text", + "content": "bike" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 402, + 127, + 457, + 167 + ], + "blocks": [ + { + "bbox": [ + 402, + 127, + 457, + 167 + ], + "lines": [ + { + "bbox": [ + 402, + 127, + 457, + 167 + ], + "spans": [ + { + "bbox": [ + 402, + 127, + 457, + 167 + ], + "type": "image", + "image_path": "374244f3d17e2095ac2c499c015ac3be9dd4c5d41f8d1ea33660efebdb3ab56d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 168, + 452, + 177 + ], + "lines": [ + { + "bbox": [ + 408, + 168, + 452, + 177 + ], + "spans": [ + { + "bbox": [ + 408, + 168, + 452, + 177 + ], + "type": "text", + "content": "motorcycle" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 182, + 482, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 182, + 482, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 482, + 192 + ], + "type": "text", + "content": "Texture: dotted, checkered, stripped, none" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 193, + 533, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 193, + 533, + 203 + ], + "spans": [ + { + "bbox": [ + 318, + 193, + 533, + 203 + ], + "type": "text", + "content": "Color: green, gray, brown, yellow, red, purple, cyan, blue" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 204, + 383, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 204, + 383, + 214 + ], + "spans": [ + { + "bbox": [ + 318, + 204, + 383, + 214 + ], + "type": "text", + "content": "Size: large, small" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 215, + 409, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 215, + 409, + 225 + ], + "spans": [ + { + "bbox": [ + 318, + 215, + 409, + 225 + ], + "type": "text", + "content": "Material: rubber, metal" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 263, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 454 + ], + "type": "text", + "content": "across different VQA datasets. For example, while colors are well-balanced in CLEVR dataset, in the GQA dataset, the color distribution is long-tailed where \"white\" appears " + }, + { + "bbox": [ + 304, + 263, + 545, + 454 + ], + "type": "inline_equation", + "content": ">50" + }, + { + "bbox": [ + 304, + 263, + 545, + 454 + ], + "type": "text", + "content": " times more frequently than \"gold\". Long-tailed distributions have been a challenge in many computer vision tasks [20, 28, 42, 56, 64]. In VQA, the long-tailed concept distribution not only hinders the learning of infrequent concepts due to few training samples, but also introduces strong biases and priors in the dataset that may mislead the models. For example, \"tennis\" is the correct answer to most questions with \"what sport is ...\" [2]. With strong priors in data, it is hard to assess the true reasoning capacity of current models. While previous works address this problem by carefully re-balancing datasets [2], in our work, we controllably vary the concept distribution in our dataset and study model robustness to concept distribution shifts." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "content": "Concept compositionality. Concept compositionality refers to how different concepts (shapes, attributes) compose and co-occur with each other, e.g. roses are usually red while violets are usually blue [30]. Concept compositionality can be viewed as a conditional concept distribution in the context of other concepts. Shifts in concept compositionality impede the generalization of VQA models. For example, the model may fail to recognize a green banana because most bananas are yellow in the training data [7]. Previous works evaluate the out-of-distribution performance by collecting counterfactual testing examples [47]. In our work, we control the compositionality of shapes and colors with an intuitive motivation: if, for example, in the training data, bicycles are red and cars are blue, will the models be able to recognize blue bicycles and red cars in testing?" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 647, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 416, + 658 + ], + "type": "text", + "content": "3.2. Dataset generation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Super-CLEVR follows a similar data generation pipeline as CLEVR, but with more complex visual components and better control of domain gap factors. We describe the generation procedure below." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14965" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Objects with parts. To improve the visual complexity of CLEVR scenes, we replace the simple shapes (e.g., cube, sphere) in CLEVR dataset with vehicles from UDA-Part dataset [40]. There are 21 vehicle models, belonging to 5 categories: car, motorbike, aeroplane, bus, and bicycle. Each 3D model comes with part annotations, e.g., left front wheel or left right door for car. Examples for the vehicle models are shown in Fig. 2. We remove or merge small parts from the original annotations to avoid severe difficulty in visual understanding. The full object and parts list is in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 312 + ], + "type": "text", + "content": "Attributes. Besides the attributes in the original CLEVR dataset, i.e. color, material, size, we optionally add texture as an additional attribute to increase visual complexity. Note that in order to enable part-based questions, the attributes (color or material) of object parts can be different from that of the object. For example, a blue car can have a red wheel or a green door. In this case, the attribute of the holistic object refers to the attribute of its main body (e.g. the blue car has blue frame)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 313, + 289, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 289, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 289, + 432 + ], + "type": "text", + "content": "Scene rendering. Following CLEVR, each scene contains 3 to 10 objects. The objects are placed onto the ground plane with random position and orientation. When placing the objects, we ensure that the objects do not overlap with each other and we avoid severe occlusion by thresholding the number of visible pixels for each object. Random jitters are added to lamp and camera positions. When rendering, we also save the ground-truth bounding boxes and segmentation masks for each of the objects and their parts, which are required when training some of the models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "type": "text", + "content": "Question generation. Super-CLEVR follows similar question generation pipeline in CLEVR, which instantiates question templates using the underlying reasoning program that can be operated on the scene graph. For example, the program select_shape truck) " + }, + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 434, + 289, + 541 + ], + "type": "text", + "content": " query_color (·) can be instantiated as question \"what is the color of the truck\". Therefore, redundancy level of questions can be controlled by removing or adding redundant reasoning steps in the underlying reasoning program." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 550, + 179, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 179, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 179, + 562 + ], + "type": "text", + "content": "3.3. Controlling the dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 569, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 629 + ], + "type": "text", + "content": "To study domain generalization, we generate several variants of the dataset for each of the domain shift factors. The variants of the datasets serve as different data domains to test the model robustness. Here we describe the method for controllably generating the dataset variants." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Visual complexity. We generate three variants of the dataset with different levels of visual complexity: easy, mid (middle) and hard. The only difference between the 3 versions is visual complexity: for the easy version, objects with different sizes, colors and materials are placed into the scene; for the middle version, we choose 3 parts on each object that are visible and randomly change their attributes;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "for the hard version, we further add random textures to the objects and parts. An example of the 3 dataset versions can be found in Fig. 1. Note that the scene layout and the questions are shared, so that the influence of visual complexity can be isolated and studied independently." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": "Question redundancy. Three variants of the dataset with different redundancy levels are generated: " + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "inline_equation", + "content": "rd-" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "inline_equation", + "content": "rd" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": " (default), " + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "inline_equation", + "content": "rd+" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": ". By default (" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "inline_equation", + "content": "rd" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": "), as in original CLEVR dataset, the questions contain some redundant attributes resulting from random sampling, while all redundant relationships are removed. In " + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "inline_equation", + "content": "rd-" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": ", we also remove all redundant attributes from the questions, leading to no redundancy in the questions. In " + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "inline_equation", + "content": "rd+" + }, + { + "bbox": [ + 304, + 134, + 546, + 266 + ], + "type": "text", + "content": ", we add all possible attributes and relationships into the question, so that questions contain a high level of redundancy. For all the variants, the questions are ensured to be valid." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": "Concept distribution. We generate three dataset variants with different concept distributions: bal (balanced), slt (slightly unbalanced) and long (long-tail distributed). More specifically, we change the distribution of shapes, colors and materials while the distribution of size is kept fixed in order to keep visual complexity consistent, since objects with smaller sizes are visually harder to recognize. By default (bal), the shapes and attributes are randomly sampled, leading to a balanced distribution. For slt and long, the concept distribution d is generated by " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "d_{i} = a^{-i}" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": " is the index of the concept. " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": " is a hyper-parameter controlling the length of the tail. A larger " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": " leads to more imbalanced distribution and " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "a = 1" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": " leads to flat distribution (cf. Fig. 1). For slt, " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "a = 1.3" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": "; for long, " + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "inline_equation", + "content": "a = 2.0" + }, + { + "bbox": [ + 304, + 268, + 547, + 520 + ], + "type": "text", + "content": ". In addition, to better analyze the performance on the frequent and rare concepts, we generate three variants for testing purpose only: head (frequent concepts in the long-tail distribution), tail (infrequent/rare concepts), and oppo (opposite to the long-tail distribution). We test each model on those three variants to analyze the performance on concepts with different degrees of frequency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "Concept compositionality. We generate 3 versions of the dataset, co-0, co-1 and co-2, with different compositions of the 21 shapes (from 5 categories) and the 8 colors. The compositionality of the dataset is controlled with the codistribution matrix " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{21 \\times 8}" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": ", where each entry " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "M_{ij}" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " is the probability an object of the " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "-th shape has the " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": "-th color. Entries in each row of " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " sum up to 1. In the version co-0, " + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 522, + 547, + 714 + ], + "type": "text", + "content": " is a flat matrix so that the shapes and colors are randomly composed. In co-1, each shape in one category has a different color distribution, e.g. truck and sedan, while shapes from different categories may share the same color distribution e.g. sedan and airliner. Oppositely, in co-2, we make the shapes in same category have the same color distribution, while shapes from different categories have different distributions. The motivation is that since shapes from the same category are visually similar, the difference in co-1" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14966" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "and co-2 will help analyze the difference in model predictions on visually similar objects and dissimilar objects when composed with different color distributions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 217 + ], + "type": "text", + "content": "Dataset Statistics. Every dataset variant contains 30k images, including 20k for training, 5k for validation and 5k for testing. Each image is paired with 10 object-based and 10 part-based questions. By default, the dataset refers to the version with mid visual complexity level (i.e. objects are untextured and has up to 3 parts with distinct attributes), rd redundancy level, balanced (bal) concept distribution and random (co-0) compositionality. More dataset statistics are in supplementary materials." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 160, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 160, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 160, + 240 + ], + "type": "text", + "content": "4. Evaluated methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 146, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 146, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 146, + 260 + ], + "type": "text", + "content": "4.1. Simple baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "content": "Random, Majority. These simple baselines pick a random or the most frequent answer for each question type in the training set as the predicted answer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 301, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 335 + ], + "type": "text", + "content": "LSTM. This question-only baseline encodes the question word embeddings with LSTM [21] and predicts answer with an MLP on top of the final hidden states of the LSTM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 382 + ], + "type": "text", + "content": "CNN+LSTM. The image is represented with features extracted by CNN and question is encoded by LSTM. An MLP predicts answer scores based on the concatenation of image and question features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 391, + 144, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 144, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 144, + 403 + ], + "type": "text", + "content": "4.2. Existing models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "text", + "content": "FiLM. We choose Feature-wise Linear Modulation [50] as a representative of classic two-stream feature merging methods. The question features extracted with GRU [10] and image features extracted with CNN are fused with the proposed FiLM module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 466, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 525 + ], + "type": "text", + "content": "mDETR. mDETR [29] is a transformer-based detector trained to detects objects in an image conditioned on a text query. The model is pretrained with 1.3M image and text pairs and can be finetuned for various downstream tasks like referring expression understanding or VQA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 525, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 582 + ], + "type": "text", + "content": "NSCL. The Neuro-Symbolic Concept Learner [45] is a representative neural symbolic method. NSCL executes neural modules on the scene representation based on the reasoning program, during which the modules learn embeddings of each concept with the answer supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "NSVQA. Neural-Symbolic VQA [59] is a neural symbolic method composed of three components: A scene parser (Mask-RCNN [19]) that segments an input image and recovers a structural scene representation, a question parser that converts a question from natural language into a program and a program executor that runs the program on the structural scene representation to obtain the answer. Notably, compared to NSCL, the individual components of NSVQA can be learned separately, hence, for example the scene parser can be learned from data that does not necessarily have Visual-Question annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 488, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 488, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 488, + 85 + ], + "type": "text", + "content": "4.3. Probabilistic NSVQA (P-NSVQA)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": "Since the program executor in NSVQA is a collection of deterministic, generic functional modules, it can be augmented with a probabilistic reasoning process that takes into account the confidence of the predictions of the scene parser. This allows the model to execute the program that has the largest joint likelihood, instead of only taking the maximal likelihood execution at each step of the program. The experiment results demonstrate a significant performance improvement of this probabilistic approach over the deterministic NSVQA model proposed in [59]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "content": "In particular, we interpret the confidence of the Mask-RCNN output as a likelihood function for all detected object classes " + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "inline_equation", + "content": "p_{\\text{object}}" + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "content": " and their attributes " + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "inline_equation", + "content": "p_{\\text{att}}" + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "content": ". Moreover, we define a likelihood " + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "inline_equation", + "content": "p_{\\text{spatial}}" + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "content": " for the spatial relations between objects (behind, in front, left, right) that is proportional to the distance between the centers of two bounding boxes. Given a reasoning program containing multiple reasoning steps, we execute each step based on the scene parsing likelihood and produce an step-wise output with confidence. Finally, we use a factorized model, multiplying the output for all the steps to get the final answer prediction. We refer readers to the Appendix for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 360, + 437, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 437, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 437, + 373 + ], + "type": "text", + "content": "4.4. Implementation details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 378, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 498 + ], + "type": "text", + "content": "Training mDETR requires ground-truth grounding of question tokens to image regions, which is available in Super-CLEVR. NSCL requires bounding box of objects, which can be predicted using a trained Faster RCNN, and the reasoning program, which can be parsed using a trained parser. Similarly, ground-truth programs are used for training NSVQA and P-NSVQA. Note that we empirically find that the question-to-program parsing is a relatively easy task (" + }, + { + "bbox": [ + 304, + 378, + 545, + 498 + ], + "type": "inline_equation", + "content": ">99\\%" + }, + { + "bbox": [ + 304, + 378, + 545, + 498 + ], + "type": "text", + "content": " accuracy using a simple LSTM), so we focus more on models' reasoning ability in our analysis." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Unless specified, the models are trained with default setting as in the official implementation. FiLM is trained for 100k iterations with batch size 256. mDETR is trained for 30 epochs with batch size 64 using 2 GPUs for both the grounding stage and the answer classification stage. NSCL is trained for 80 epochs with batch size 32. For NSVQA and P-NSVQA, we first train the object parser (Mask RCNN [19]) for 30k iterations with batch size 16, then train the attribute extraction model (using the Res50 backbone) for 100 epochs with batch size 64. For P-NSVQA, when counting the objects or determining whether objects exist in the scene, we use a threshold (0.7) to obtain the final selected objects. Early stopping is used based on validation accuracy. All the models are trained with 200k questions. We repeat experiments on default split for 3 times with different random seeds and get std of " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pm 0.10" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " (P-NSVQA) and " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pm 0.40" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " (NSVQA), showing the statistical significance of our results, then we only run other experiments once." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14967" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 167, + 85 + ], + "type": "text", + "content": "5. Results and analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 139 + ], + "type": "text", + "content": "In this section, we first show evaluation results for indomain setting, then provide results and analysis on out-of-domain evaluation. Finally, we describe additional studies and future works." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 146, + 153, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 146, + 153, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 153, + 158 + ], + "type": "text", + "content": "5.1. In-domain results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 164, + 286, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 286, + 224 + ], + "type": "text", + "content": "In-domain evaluation refers to the setting where the training and testing data come from the same domain (the default dataset variant in this case). We compare the indomain results on Super-CLEVR and CLEVR. The results are shown in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 231, + 282, + 391 + ], + "blocks": [ + { + "bbox": [ + 50, + 231, + 282, + 391 + ], + "lines": [ + { + "bbox": [ + 50, + 231, + 282, + 391 + ], + "spans": [ + { + "bbox": [ + 50, + 231, + 282, + 391 + ], + "type": "image", + "image_path": "aaa9891acc31869c8baa353199fd99d14b1ff7b1e2b49763bd49e710dfe1370b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 287, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 418 + ], + "type": "text", + "content": "Figure 3. Comparison of models' accuracy on Super-CLEVR and the original CLEVR dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": "For all the models, the performance is lower on SuperCLEVR than CLEVR, suggesting that Super-CLEVR is a more challenging benchmark. The scenes with vehicles are much harder visually for the models to understand compared with the simpler shapes from CLEVR. Note that the performance gap on two datasets for simple baselines (Random, Majority, LSTM, CNN+LSTM) is smaller than for the other better models. This is because Super-CLEVR contains more object types, and therefore the performance of simply guessing is lower than on CLEVR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "type": "text", + "content": "When comparing the performance of different models, we find that the neural modular methods, i.e. NSCL, NSVQA, P-NSVQA, perform much better than non-modular ones. This is not surprising given their nearly perfect performance on the original CLEVR dataset, which shows its strong ability to model synthetic images. The large-scale pretrained grounding model mDETR, which is a leading model on both real and synthetic images, also achieves good performance " + }, + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "type": "inline_equation", + "content": "(82.7\\%)" + }, + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "type": "text", + "content": " on Super-CLEVR. The two-stream method FiLM does not achieve very strong performance " + }, + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "type": "inline_equation", + "content": "(53.2\\%)" + }, + { + "bbox": [ + 46, + 545, + 286, + 689 + ], + "type": "text", + "content": ", but is still much better than the other simple baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Our proposed P-NSVQA outperforms all the other models. In particular, on Super-CLEVR, it outperforms its de" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "terministic counterpart, NSVQA, by " + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "2.96\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": ". This shows the advantage of taking into account the probabilities when the scenes are challenging thus the model's uncertainty of predictions can be utilized." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 129, + 432, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 129, + 432, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 432, + 140 + ], + "type": "text", + "content": "5.2. Out-of-domain results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 147, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 230 + ], + "type": "text", + "content": "In this section, we train and test the five models (FiLM, mDETR, NSCL, NSVQA and P-NSVQA) on different dataset variants, and diagnose their domain robustness on each of the four domain shift factors. Please refer to Sec. 3.3 for a description of different variants. The validation accuracy is used for analysis here and the results are shown in Tab. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 232, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 291 + ], + "type": "text", + "content": "All the methods suffer from domain shifts. The results show that the best performance mostly occurs in situations where the model is tested on the same dataset variant as it is trained on, i.e. the bold or underlined numbers fall mostly on the diagonals in Tab. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "text", + "content": "We compare the domain robustness of the five models by measuring the relative performance decrease when the testing data differs from the training data, i.e. smaller performance drop on different testing domains means better robustness. Based on this intuition, for easier understanding of Tab. 1, we propose a measurement metric for domain robustness named Relative Degrade (RD) to better analyze the results. We define Relative Degrade as the percentage of accuracy decrease when the model is tested under a domain shift, i.e. the accuracy drop divided by the in-domain accuracy. Specifically, if a model gets accuracy " + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "text", + "content": " under in-domain testing (i.e. testing with the same dataset variant as training) and accuracy " + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "text", + "content": " under out-of-domain testing (i.e. testing with a different dataset variant from training), then " + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "inline_equation", + "content": "RD = (a - b) / a" + }, + { + "bbox": [ + 304, + 292, + 545, + 495 + ], + "type": "text", + "content": ". Since we train each model on three data variants, the RD's for the three models are averaged to measure its domain robustness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 495, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 579 + ], + "type": "text", + "content": "Tab. 2 shows the Relative Degrade of the five models on the four factors. We see that P-NSVQA outperforms other models by a significant margin on three of the four factors, indicating that it has better overall domain robustness. In the following, we take a closer look at the results on each of the factors separately, to diagnose the influence of different model designs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "text", + "content": "Question redundancy. Neural modular methods are much more robust to question redundancy shifts than non-modular ones. The relative degrades for modular methods are less than " + }, + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "text", + "content": ", while one-modular ones degrade for around " + }, + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 579, + 545, + 675 + ], + "type": "text", + "content": ". Due to the step-by-step design of the reasoning in modular methods, each reasoning step is independent of the others so that the models are less likely to learn the spurious correlation between question and answers. There-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 683, + 545, + 713 + ], + "type": "text", + "content": "1For concept distributions, we compute relative degrade with a slight change: we compute the accuracy drop from head to tail and the drop from long to oppo, take their average, and divide by the accuracy on bal." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14968" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 61, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 50, + 61, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 50, + 61, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 50, + 61, + 545, + 383 + ], + "type": "table", + "html": "
FiLMmDETRNSCLNSVQAProb NSVQA
Visual Complexity
easymidhardeasymidhardeasymidhardeasymidhardeasymidhard
easy59.9653.9550.6693.3684.3082.9795.1392.3190.8195.1994.1994.0996.7695.9896.37
mid57.4153.2850.1883.3482.3681.2784.589.1086.3381.9992.8093.7886.2595.7695.11
hard55.9553.1150.4779.7179.9480.7176.8578.6685.0873.1179.7192.6579.8186.4795.36
Question Redundancy
rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+rd-rdrd+
rd-51.4252.5453.5183.9480.3766.2888.6488.8290.3392.9592.9492.6795.6695.7295.43
rd-50.3953.2854.7882.7782.3670.3688.4589.1091.4591.1992.7892.1494.8795.7295.43
rd+46.1452.3071.4778.4884.0590.4287.9488.3491.1691.3891.9692.8094.8895.4795.72
Concept Distribution
balsltlongbalsltlongbalsltlongbalsltlongbalsltlong
bal50.4753.0454.3580.7175.7974.5485.0883.7975.1092.6590.8283.7495.3694.8989.88
long49.4354.7562.9679.0680.2990.6685.3389.4291.1092.7393.3892.5396.3196.3295.25
head48.6058.0661.6080.7579.6087.4684.5888.3990.1993.8794.8292.4896.4296.8095.92
tail51.8048.7050.0881.5070.8860.9486.1080.2760.5590.2689.2075.3294.0893.2082.68
oppo49.0648.9346.6879.1368.3756.9885.0777.8655.1491.2288.6571.3295.7694.0979.74
Concept Compositionality
co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2co-0co-1co-2
co-053.2857.0056.183.3677.0382.4389.182.5283.7792.8090.1191.5995.7694.0295.12
co-152.4160.5756.6779.4682.4583.9378.8987.1884.278.7489.9990.6787.1294.5394.78
co-252.9657.3760.5380.0377.4187.2478.4081.5588.8477.8589.2892.2387.1993.4995.61
", + "image_path": "7a2b92c7b83207a35c6bd6b4acd76b8ef8839d8224ee93b02685e3a0970b352b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 435, + 285, + 525 + ], + "blocks": [ + { + "bbox": [ + 46, + 388, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 546, + 422 + ], + "type": "text", + "content": "Table 1. Accuracy of models trained and tested on different domains. Column headings indicate training settings, while rows indicate the dataset variant for testing. The best performance in each row (i.e. the best training setting) is marked in bold and best performance in each column (i.e. the best testing setting) is underlined. Description for different splits is in Sec. 3.3 and analysis is in Sec. 5.2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 435, + 285, + 525 + ], + "lines": [ + { + "bbox": [ + 50, + 435, + 285, + 525 + ], + "spans": [ + { + "bbox": [ + 50, + 435, + 285, + 525 + ], + "type": "table", + "html": "
VisualRedund.Dist.Comp.
FiLM4.0321.3328.469.04
mDETR9.8119.0536.349.45
NSCL15.570.9237.4415.40
NSVQA17.481.7220.9211.44
Prob NSVQA12.880.8413.727.00
", + "image_path": "a0ec7897ee949d17dd17af71317509a10d136586514c9a79a0164ab789f1ccfa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 529, + 287, + 562 + ], + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 562 + ], + "type": "text", + "content": "Table 2. Relative Degrade under domain shifts, i.e. the percentage of accuracy decrease when the model is tested with domain that differs with training. Lower " + }, + { + "bbox": [ + 46, + 529, + 287, + 562 + ], + "type": "inline_equation", + "content": "RD" + }, + { + "bbox": [ + 46, + 529, + 287, + 562 + ], + "type": "text", + "content": " means better robustness." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 578, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 601 + ], + "type": "text", + "content": "fore the modular methods are less vulnerable to change in question/program length." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "type": "text", + "content": "Visual complexity. Different from our findings on question redundancy, for domain shifts in visual complexity, non-modular methods are more robust compared to modular ones. As shown in Tab. 2, while FiLM and mDETR gets less than " + }, + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "type": "text", + "content": " degrade, NSCL and (P-)NSVQA degrade for more than " + }, + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 603, + 287, + 699 + ], + "type": "text", + "content": ". The reason might be that the simple reasoning modules in modular methods can not process the visual signals as well as the dense non-modular models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "content": "Comparing P-NSVQA with NSVQA, we find that in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "content": "jecting probability into deterministic symbolic reasoning greatly improves the robustness on visual complexity (4.04% decrease in " + }, + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "inline_equation", + "content": "RD" + }, + { + "bbox": [ + 304, + 437, + 545, + 510 + ], + "type": "text", + "content": "). This suggests that some errors in visual understanding can be corrected and recovered by taking into account the uncertainty of visual parsing and combining the results of each reasoning step with probability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "text", + "content": "Concept distribution. While all the four existing models suffer a lot (larger than " + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "text", + "content": " RD) on domain shifts in concept distribution, we see that the symbolic method NSVQA is better than the other three (by more than " + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "text", + "content": "). With the disentangled reasoning and visual understanding components in NSVQA, the distribution priors in the images and the programs/answers cannot intertwine with each other, which prevent the model heavily relying on the priors. With uncertainty, we can further boost the robustness of NSVQA with a large margin (from " + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 304, + 510, + 546, + 629 + ], + "type": "text", + "content": " RD)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "Moreover, the head-tail results suggests that the overall accuracy, which is commonly used to measure VQA performance, should be taken with cautious. When the testing split is imbalanced, the seemingly high accuracy is misleading because the head concepts dominate the testing while the tail ones are not well-reflected. For example, for NSCL, although it gets high accuracy (91%) on the long-tailed" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14969" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "data, its performance is only " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " on the tail concepts. In real-world datasets, the data are usually not well-balanced, which suggests the value of synthetic testing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 192 + ], + "type": "text", + "content": "Concept compositionality. Comparing the existing methods, we find that the non-modular methods seems to be more robust than modular methods NSCL or NSVQA. However, with uncertainty, P-NSVQA improves the result of NSVQA, which even outperforms the non-modular methods. This suggest the large potential of better robustness of modular methods by improving current models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 286, + 323 + ], + "type": "text", + "content": "In summary, while non-modular methods are more robust to visual complexity shifts, the modular symbolic methods (improved with uncertainty) are more robust on the other three factors. By disentangling reasoning with visual understanding, separately executing every each reasoning step then merging the results of the steps using probabilities based on uncertainty, our P-NSVQA outperforms all the existing models in question redundancy, concept distribution and compositionality. Therefore, we suggest that symbolic reasoning with uncertainty leads to strong VQA models that are robust to domain shifts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 334, + 214, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 214, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 214, + 346 + ], + "type": "text", + "content": "5.3. More analysis and future work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 352, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 286, + 579 + ], + "type": "text", + "content": "Synthetic-to-real transfer. We provide an additional proof-of-concept study to show that the findings drawn from Super-CLEVR dataset can transfer to real datasets. In the following experiments, we show our finding that neuro-symbolic methods (NSCL, NSVQA, P-NSVQA) are more robust than mDETR on question redundancy also holds true on the real GQA dataset [25]. More precisely, we progressively removed the redundant operations from the reasoning program in GQA testdev split, and then regenerated questions using a program-to-question generator. Using the change of models' testing accuracy as the redundant operations are removed, we can evaluate the models' robustness towards question redundancy. The results are show in Tab. 3.2 We observe that the performance drop of mDETR is much larger than neuro-symbolic methods as the redundant information is progressively removed, which indicates that symbolic methods have better question redundancy than mDETR on GQA dataset. This is consistent with our findings on Super-CLEVR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 581, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 664 + ], + "type": "text", + "content": "Reasoning with part-object hierarchies. In addition to evaluating domain generalization, Super-CLEVR can be extended for broader purposes, e.g. part-based reasoning. We can ask questions like \"what is the color of the front wheel of the bike?\", \"what is the color of the vehicle that has a yellow wheel\", etc. Those questions require the model to correctly understanding the part-object hierarchy, which" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 136 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 136 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 136 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 136 + ], + "type": "table", + "html": "
0%14%32%70%91%100%
mDETR0-4.82-8.46-13.16-13.88-14.56
NSCL0-0.14-0.34-1.09-1.71-2.59
NSVQA0-3.47-4.80-7.01-7.02-7.02
P-NSVQA0-1.93-3.15-5.73-5.91-5.78
", + "image_path": "e85fcd9c2c01ecd6cfb12f2c632c2dd41d1451667785fc437e4eee5dbd38774a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 137, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 306, + 137, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 545, + 159 + ], + "type": "text", + "content": "Table 3. Accuracy drop on the GQA dataset when redundant information is progressively removed." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 171, + 480, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 480, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 480, + 182 + ], + "type": "text", + "content": "is an ability that current VQA models lack." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 182, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 253 + ], + "type": "text", + "content": "Limitations. The main limitations of our work lie in the synthetic nature of our dataset. Future efforts can be made in collecting better controlled and balanced real datasets for model diagnosis. We emphasize that the purpose of the dataset is for model diagnosis and that models should also be tested on real data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 265, + 378, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 378, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 378, + 277 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 285, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 536 + ], + "type": "text", + "content": "We diagnose domain shifts in visual reasoning using a proposed virtual benchmark, Super-CLEVR, where distinct factors can be independently studied with controlled data generation. We evaluate four existing methods and show that all of them struggle with domain shifts, highlighting the importance of out-of-domain testing. Among the evaluated methods, neural modular methods are more robust towards question redundancy. In particular, NSVQA with disentangled perception and reasoning shows better robustness towards distribution and compositionality shifts. We further propose P-NSVQA, which improves NSVQA with uncertainty in the reasoning modules. We show that P-NSVQA outperforms all the existing methods in both in-domain testing and out-of-domain testing. With detailed analysis, our study suggests that disentangling reasoning and perception, combined with probabilistic uncertainty, form a strong VQA model that is more robust to domain shifts. We hope our analysis may facilitate better understanding of strengths and weaknesses of VQA models and, more broadly, future work might explore using the Super-CLEVR benchmark for other tasks like part-based reasoning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 544, + 400, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 400, + 556 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 400, + 556 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 562, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 681 + ], + "type": "text", + "content": "We would like to thank Nils Holzenberger, Kate Sanders, Chenxi Liu, Zihao Xiao, Qing Liu, Reno Kriz, and David Etter for their helpful comments and suggestions, as well as the anonymous reviewers. Zhuowan Li is supported by ONR N00014-21-1-2812 and grant IAA80052272 from the Institute for Assured Autonomy at JHU. Elias Stengel-Eskin is supported by an NSF GRFP. A. Kortylewski acknowledges support via his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No.468670075." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "text", + "content": "2For implementation of NSCL on a real-word dataset, we use the model in [39] (the version without calibration). The model accuracies on the original not-perturbed GQA testdev split are as following: mDETR " + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "inline_equation", + "content": "(61.67\\%)" + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "text", + "content": " NSCL " + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "inline_equation", + "content": "(56.13\\%)" + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "text", + "content": " NSVQA " + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "inline_equation", + "content": "(39.58\\%)" + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "text", + "content": " P-NSVQA " + }, + { + "bbox": [ + 46, + 673, + 286, + 712 + ], + "type": "inline_equation", + "content": "(39.66\\%)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14970" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 179 + ], + "type": "text", + "content": "[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4971-4980, 2018. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 287, + 246 + ], + "type": "text", + "content": "[3] Arjun Akula, Soravit Changpinyo, Boqing Gong, Piyush Sharma, Song-Chun Zhu, and Radu Soricut. Crossvqa: scalably generating benchmarks for systematically testing vqa generalization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2148-2166, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 312 + ], + "type": "text", + "content": "[4] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 313, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 368 + ], + "type": "text", + "content": "[5] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 423 + ], + "type": "text", + "content": "[6] Dzmitry Bahdanau, Harm de Vries, Timothy J O'Donnell, Shikhar Murty, Philippe Beaudoin, Yoshua Bengio, and Aaron Courville. Closure: Assessing systematic generalization of clevr models. arXiv preprint arXiv:1912.05783, 2019. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 423, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 287, + 468 + ], + "type": "text", + "content": "[7] Remi Cadene, Corentin Dancette, Matthieu Cord, Devi Parikh, et al. Rubi: Reducing unimodal biases for visual question answering. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 514 + ], + "type": "text", + "content": "[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "text", + "content": "[9] Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800-10809, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "text", + "content": "[10] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 680 + ], + "type": "text", + "content": "[11] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1554–1563, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] William Ford and David Olson. The elaboration of the noun phrase in children's description of objects. Journal of Experimental Child Psychology, 19(3):371-382, 1975. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "text", + "content": "[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "text", + "content": "[14] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 263 + ], + "type": "text", + "content": "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 319 + ], + "type": "text", + "content": "[17] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. arXiv preprint arXiv:2204.02285, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 386 + ], + "type": "text", + "content": "[18] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 387, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 421 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 421, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 476 + ], + "type": "text", + "content": "[20] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6930-6940, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 476, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 499 + ], + "type": "text", + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 499, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 545, + 555 + ], + "type": "text", + "content": "[22] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. PMLR, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 601 + ], + "type": "text", + "content": "[23] Yining Hong, Li Yi, Josh Tenenbaum, Antonio Torralba, and Chuang Gan. *Ptr: A benchmark for part-based conceptual, relational, and physical reasoning*. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 624 + ], + "type": "text", + "content": "[24] Drew A Hudson and Christopher D Manning. Compositional attention networks for machine reasoning. 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 679 + ], + "type": "text", + "content": "[25] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "text", + "content": "[26] Jingjing Jiang, Ziyi Liu, Yifan Liu, Zhixiong Nan, and Nanning Zheng. X-ggm: Graph generative modeling for out-of-distribution generalization in visual question answering. In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14971" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Proceedings of the 29th ACM International Conference on Multimedia, pages 199-208, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 229 + ], + "type": "text", + "content": "[28] Lie Ju, Xin Wang, Lin Wang, Tongliang Liu, Xin Zhao, Tom Drummond, Dwarikanath Mahapatra, and Zongyuan Ge. Relational subsets knowledge distillation for long-tailed retinal diseases recognition. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 3-12. Springer, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "text", + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. arXiv preprint arXiv:2104.12763, 2021. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 330 + ], + "type": "text", + "content": "[30] Corentin Kervadec, Grigory Antipov, Moez Baccouche, and Christian Wolf. Roses are red, violets are blue... but should vqa expect them to? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2776-2785, 2021. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "type": "text", + "content": "[31] Corentin Kervadec, Theo Jaunet, Grigory Antipov, Moez Baccouche, Romain Vuillemot, and Christian Wolf. How transferable are reasoning patterns in vqa? 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4205-4214, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "text", + "content": "[32] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[33] Ruud Koolen, Martijn Goudbeek, and Emiel Krahmer. Effects of scene variation on referential overspecification. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 33, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 510 + ], + "type": "text", + "content": "[34] Satwik Kottur, José MF Moura, Devi Parikh, Dhruv Batra, and Marcus Rohrbach. Clevr-dialog: A diagnostic dataset for multi-round reasoning in visual dialog. arXiv preprint arXiv:1903.03166, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 577 + ], + "type": "text", + "content": "[35] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[36] Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1446-1455, 2019. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[37] Linjie Li, Zhe Gan, Yu Cheng, and Jingjing Liu. Relation-aware graph attention network for visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10313-10322, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[38] Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "semantics aligned pre-training for vision-language tasks. ECCV 2020, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "text", + "content": "[39] Zhuowan Li, Elias Stengel-Eskin, Yixiao Zhang, Cihang Xie, Quan Hung Tran, Benjamin Van Durme, and Alan Yuille. Calibrating concepts and operations: Towards symbolic reasoning on real images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14910-14919, October 2021. 1, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[40] Qing Liu, Adam Kortylewski, Zhishuai Zhang, Zizhang Li, Mengqi Guo, Qihao Liu, Xiaoding Yuan, Jiteng Mu, Weichao Qiu, and Alan Yuille. Learning part segmentation through unsupervised domain adaptation from synthetic vehicles. In CVPR, 2022. 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[41] Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[42] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[43] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "text", + "content": "[45] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, and Jiajun Wu. The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision. In International Conference on Learning Representations, 2019. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[46] Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184, 2013. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 588 + ], + "type": "text", + "content": "[47] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xiansheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12695–12705, 2021. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 644 + ], + "type": "text", + "content": "[48] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12700-12710, June 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 667 + ], + "type": "text", + "content": "[49] Thomas Pechmann. Incremental speech production and referential overspecification. 1989. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[50] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14972" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[51] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 152 + ], + "type": "text", + "content": "[52] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 208 + ], + "type": "text", + "content": "[53] Leonard Salewski, A Koepke, Hendrik Lensch, and Zeynep Akata. Clevr-x: A visual reasoning dataset for natural language explanations. In International Workshop on Extending Explainable AI Beyond Deep Models and Classifiers, pages 69-88. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[54] Susan Sonnenschein. The development of referential communication skills: Some situations in which speakers give redundant messages. Journal of Psycholinguistic Research, 14(5):489-508, 1985. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 287 + ], + "type": "text", + "content": "[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 288, + 353 + ], + "type": "text", + "content": "[56] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "text", + "content": "[57] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 444 + ], + "type": "text", + "content": "[58] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 499 + ], + "type": "text", + "content": "[59] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Joshua B Tenenbaum. Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding. In Advances in Neural Information Processing Systems (NIPS), 2018. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[60] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6281–6290, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[61] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5317-5327, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 669 + ], + "type": "text", + "content": "[62] Mingda Zhang, Tristan Maidment, Ahmad Diab, Adriana Kovashka, and Rebecca Hwa. Domain-robust vqa with diverse datasets and methods but no target labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7046–7056, 2021. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[63] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. CVPR 2021, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 162 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 118 + ], + "type": "text", + "content": "[64] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In Proceedings of the IEEE International Conference on Computer Vision, pages 5409-5418, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "text", + "content": "[65] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4995–5004, 2016. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14973" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_content_list.json b/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..41b36cd5480d5ef758cc8fd3b5430a91b40f3fd4 --- /dev/null +++ b/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_content_list.json @@ -0,0 +1,1651 @@ +[ + { + "type": "text", + "text": "Super-Resolution Neural Operator", + "text_level": 1, + "bbox": [ + 307, + 130, + 661, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Min Wei* Xuesong Zhang*†", + "bbox": [ + 361, + 180, + 612, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing University of Posts and Telecommunications, China", + "bbox": [ + 246, + 199, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mw，xuesong_zhang}@bupt.edu.cn", + "bbox": [ + 344, + 219, + 619, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Super-resolution Neural Operator (SRNO), a deep operator learning framework that can resolve high-resolution (HR) images at arbitrary scales from the low-resolution (LR) counterparts. Treating the LR-HR image pairs as continuous functions approximated with different grid sizes, SRNO learns the mapping between the corresponding function spaces. From the perspective of approximation theory, SRNO first embeds the LR input into a higher-dimensional latent representation space, trying to capture sufficient basis functions, and then iteratively approximates the implicit image function with a kernel integral mechanism, followed by a final dimensionality reduction step to generate the RGB representation at the target coordinates. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial domain and therefore benefits the grid-free continuum; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to \"hallucinate\" high-frequency information from the LR image. Experiments show that SRNO outperforms existing continuous SR methods in terms of both accuracy and running time. Our code is at https://github.com/2y7c3/Super-Resolution-Neural-Operator.", + "bbox": [ + 75, + 300, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 707, + 209, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SR) addresses the inverse problem of reconstructing high-resolution (HR) images from their low-resolution (LR) counterparts. In a data-driven way, deep neural networks (DNNs) learn the inversion map from many LR-HR sample pairs and have demonstrated appealing performances [4, 20, 21, 24, 29, 40, 41]. Nevertheless, most DNNs are developed in the configuration of single scaling factors, which cannot be used in sce", + "bbox": [ + 75, + 732, + 468, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1c8a91dd468da760dfaac890b183515eb452f8ec13b0121f1bb055729e3a684.jpg", + "image_caption": [ + "Figure 1. Overview of Super-Resolution Neural operator (SRNO). SRNO is composed of three parts, $\\mathcal{L}$ (Lifting), $\\kappa$ (kernel integrals) and $\\mathcal{P}$ (Projection), which perform consecutively to learn mappings between approximation spaces $\\mathcal{H}(\\Omega_{h_c})$ and $\\mathcal{H}(\\Omega_{h_f})$ associated with grid sizes $h_c$ and $h_f$ , respectively. The key component, $\\kappa$ , uses test functions in the latent Hilbert space $\\mathcal{V}(\\Omega_{h_f})$ to seek instance-specific basis functions." + ], + "image_footnote": [], + "bbox": [ + 517, + 271, + 879, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "narios requiring arbitrary SR factors [37, 38]. Recently, implicit neural functions (INF) [5, 18] have been proposed to represent images in arbitrary resolution, and paving a feasible way for continuous SR. These networks, as opposed to storing discrete signals in grid-based formats, represent signals with evaluations of continuous functions at specified coordinates, where the functions are generally parameterized by a multi-layer perceptron (MLP). To share knowledge across instances instead of fitting individual functions for each signal, encoder-based methods [5, 15, 26] are proposed to retrieve latent codes for each signal, and then a decoding MLP is shared by all the instances to generate the required output, where both the coordinates and the corresponding latent codes are taken as input. However, the point-wise behavior of MLP in the spatial dimensions results in limited performance when decoding various objects, particularly for high-frequency components [30, 32].", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. This work is supported by the National Natural Science Foundation of China (61871055). \n†Corresponding author.", + "bbox": [ + 76, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural operator is a newly proposed neural network architecture in the field of computational physics [17, 19, 23] for numerically efficient solvers of partial differential equations (PDE). Stemming from the operator theory, nerval operators learn mappings between infinite-dimensional function spaces, which is inherently capable of continuous function evaluations and has shown promising potentials in various applications [9, 13, 27]. Typically, neural operator consists of three components: 1) lifting, 2) iterative kernel integral, and 3) projection. The kernel integrals operate in the spatial domain, and thus can explicitly capture the global relationship constraining the underlying solution function of the PDE. The attention mechanism in transformers [36] is a special case of kernel integral where linear transforms are first exerted to the feature maps prior to the inner product operations [17]. Tremendous successes of transformers in various tasks [6, 22, 34] have shown the importance of capturing global correlations, and this is also true for SR to improve performance. [8].", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the super-resolution neural operator (SRNO), a deep operator learning framework that can resolve HR images from their LR counterparts at arbitrary scales. As shown in Fig.1, SRNO learns the mapping between the corresponding function spaces by treating the LR-HR image pairs as continuous functions approximated with different grid sizes. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial dimensions and have been proved to be comparable to a Petrov-Galerkin projection [3]; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to \"hallucinate\" high-frequency information from the LR image. When employing same encoders to capture features, our method outperforms previous continuous SR methods in terms of both reconstruction accuracy and running time.", + "bbox": [ + 75, + 378, + 472, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 96, + 650, + 434, + 664 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the methodology of super-resolution neural operator that maps between finite-dimensional function spaces, allowing for continuous and zero-shot super-resolution irrespective the discretization used on the input and output spaces.", + "- We develop an architecture for SRNO that first explores the common latent basis for the whole training set and subsequently refines an instance-specific basis by the Galerkin-type attention mechanism.", + "- Numerically, we show that the proposed SRNO outperforms existing continuous SR methods with less running time, and even generates better results on the resolutions for which the fixed scale SR networks were trained." + ], + "bbox": [ + 96, + 672, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning based SR methods. [4, 20, 21, 24, 29, 40, 41] have achieved impressive performances, in multiscale scenarios one has to train and store several models for each scale factor, which is unfeasible when considering time and memory budgets. In recent years, several methods [11, 31, 37] are proposed to achieve arbitrary-scale SR with a single model, but their performances are limited when dealing with out-of-distribution scaling factors. Inspired by INF, LIIF [5] takes continuous coordinates and latent variables as inputs, and employs an MLP to achieve outstanding performances for both in-distribution and out-of-distribution factors. In contrast, LTE [18] transforms input coordinates into the Fourier domain and uses the dominant frequencies extracted from latent variables to address the spectral bias problem [30, 32]. In a nutshell, treating images as RGB-valued functions and sharing the implicit function space are the keys to the success of LIIF-like works [5, 18]. Nevertheless, a purely local decoder, like MLP, is not able to accurately approximate arbitrary images, although it is rather sensitive to the input coordinates.", + "bbox": [ + 496, + 114, + 890, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Operators. Recently, a novel neural network architecture, Neural Operator (NO), was proposed for discretization invariant solutions of PDEs via infinitedimensional operator learning [10,17,19,23]. Neural operators only need to be trained once and are capable of transferring solutions between differently discretized meshes while keeping a fixed approximation error. A valuable merit of NO is that it does not require knowledge of the underlying PDE, which allows us to introduce it by the following abstract form,", + "bbox": [ + 496, + 417, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\mathbf {L} _ {a} u\\right) (x) = f (x), \\quad x \\in D, \\tag {1} \\\\ u (x) = 0, \\quad x \\in \\partial D, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 578, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $u: D \\to \\mathbb{R}^{d_u}$ is the solution function residing in the Banach space $\\mathcal{U}$ , and $\\mathbf{L}: \\mathcal{A} \\to \\mathbf{L}(\\mathcal{U}; \\mathcal{U}^*)$ is an operator-valued functional that maps the coefficient function $a \\in \\mathcal{A}$ of the PDE to $f \\in \\mathcal{U}^*$ , the dual space of $\\mathcal{U}$ . As in many cases the inverse operator of $\\mathbf{L}$ even does not exist, NO seeks a feasible operator $\\mathcal{G}: \\mathcal{A} \\to \\mathcal{U}, a \\mapsto u$ , directly mapping the coefficient to the solution within an acceptable tolerance.", + "bbox": [ + 496, + 626, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The operator $\\mathcal{G}$ is numerically approximated by training a neural network $\\mathcal{G}_{\\theta}:\\mathcal{A}\\to \\mathcal{U}$ , where $\\theta$ are the trainable parameters. Suppose we have $N$ pairs of observations $\\{a_j,u_j\\}_{j = 1}^N$ where the input functions $a_{j}$ are sampled from probability measure $\\mu$ compactly supported on $\\mathcal{A}$ , and $u_{j} = \\mathcal{G}(a_{j})$ are used as the supervisory output functions. The infinitely dimensional operator learning problem $\\mathcal{G}\\gets \\mathcal{G}_{\\theta}$ thus is associated with the empirical-risk minimization problem [35]. In practice, we actually measure the approximation loss using the sampled observations $u_{o}^{(j)}$", + "bbox": [ + 496, + 747, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18248", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/298c6f66eac28aebf4816051fa45e1f9793d93e3f5dfe996743d1f3749fad5a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 88, + 854, + 175 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/67a29d95ac6ad0b143a2d521a5832267442f7108d7c536340ac503b21472a7ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 176, + 468, + 266 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b66767cb366d3ef67eb8d6eca3a6290246e2b45268dc199fa8e4a55480fc482f.jpg", + "image_caption": [ + "Figure 2. Super-resolution neural operator (SRNO) architecture for continuous SR. The input LR image $f_{h_c}$ undergoes three phases to output the HR image $f_{h_f}$ with the specified resolution: (a) Lifting the LR pixel values $a(\\mathbf{x})$ on the set of coordinates $\\mathbf{x} = \\{x_i\\}_{i=1}^{n_{h_f}}$ to a higher dimensional feature space by a CNN-based encoder $E_{\\psi}$ , constructing the latent representation $\\hat{a}(x)$ , and linearly transforming into the first layer's input $z_0(\\mathbf{x})$ . (b) kernel integrals composed of $T$ layers of Galerkin-type attention, and (c) finally project to the RGB space." + ], + "image_footnote": [], + "bbox": [ + 94, + 268, + 470, + 359 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4fc242f9cdb755364d257f3bf4e8010bbecc4bbe7a300a26b2fdf7ebbb4f2308.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 176, + 875, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $a_{o}^{(j)}$ , which are the direct results of discretization:", + "bbox": [ + 76, + 436, + 431, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\mathbb {E} _ {a \\sim \\mu} \\| \\mathcal {G} (a) - \\mathcal {G} _ {\\theta} (a) \\| _ {\\mathcal {U}} \\\\ \\approx \\min _ {\\theta} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\| u _ {o} ^ {(j)} - \\mathcal {G} _ {\\theta} \\left(a _ {o} ^ {(j)}\\right) \\| _ {\\mathcal {U}}. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 462, + 468, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to classical feedforward neural networks (FFNs), the NO is of an iterative architecture. For ease of exposition, suppose $\\mathcal{A}$ is defined on the bounded domain $D\\subset$ $\\mathbb{R}^d$ , and the inputs and outputs of the intermediate layers are all vector-valued functions, with dimension $d_{z}$ . Then $\\mathcal{G}_{\\theta}:\\mathcal{A}\\to \\mathcal{U}$ can be formulated as follows:", + "bbox": [ + 76, + 539, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {0} (x) = \\mathcal {L} (x, a (x)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 638, + 468, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} (x) = \\sigma \\left(W _ {t} z _ {t} (x) + \\left(\\mathcal {K} _ {t} \\left(z _ {t}; \\Phi\\right)\\right) (x)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 657, + 468, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu (x) = \\mathcal {P} \\left(z _ {T} (x)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 676, + 468, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}:\\mathbb{R}^{d_a + d}\\to \\mathbb{R}^{d_z}$ , and $\\mathcal{P}:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_u}$ are the local lifting and projection functions respectively, mapping the input $a$ to its first layer hidden representation $z_{0}$ and the last layer hidden representation $z_{T}$ back to the output function $u$ . $W:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_z}$ is a point-wise linear transformation, and $\\sigma :\\mathbb{R}^{d_z}\\rightarrow \\mathbb{R}^{d_z}$ is the nonlinear activation function.", + "bbox": [ + 76, + 702, + 468, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the PDE in (1) point-wisely defines the behavior of the solution function $u(x)$ , the solution operator $\\mathcal{G}$ we are seeking should exhibit the non-local property such that $\\mathcal{G}(a)$ can approximate $u$ everywhere rather than locally. For this purpose, NO may employ the kernel integral operators $\\mathcal{K}_t: \\{z_t : D_t \\to \\mathbb{R}^{d_z}\\} \\mapsto \\{z_{t+1} : D_{t+1} \\to \\mathbb{R}^{d_z}\\}$ to maintain the continuum in the spatial domain, and one of the", + "bbox": [ + 76, + 795, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "most adopted forms [19] is defined as", + "bbox": [ + 500, + 439, + 750, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathcal {K} _ {t} \\left(z _ {t}; \\Phi\\right)\\right) (x) = \\int_ {D} K _ {t} (x, y; \\Phi) z _ {t} (y) \\mathrm {d} y, \\quad \\forall x \\in D, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 465, + 890, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the kernel matrix $K_{t}:\\mathbb{R}^{d + d}\\to \\mathbb{R}^{d_{z}\\times d_{z}}$ is parameterized by $\\Phi$ .", + "bbox": [ + 498, + 510, + 890, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Super-resolution Neural Operator", + "text_level": 1, + "bbox": [ + 500, + 556, + 808, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All operations of NO are defined in function space and the training data are just samples at coordinates $x_{i}$ irrespective of the discretization sizes. This property inspires us to employ the NO's architecture for continuous SR. Different from NO that aims at mapping between infinitedimensional function spaces, we define SRNO in the sense of mappings between two approximation spaces of finitedimensional but continuous functions. This is reasonable because imaging optics inevitably limit the highest possible frequency in the radiance field under observation. With this configuration, we can design the architecture of SRNO using the well-developed theoretical tool, the Galerkin-type method that is widely used in the filed of Finite elements [7].", + "bbox": [ + 496, + 582, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Let $(\\mathcal{H},\\langle \\cdot ,\\cdot \\rangle_{\\mathcal{H}})$ denote a Hilbert space equipped with the inner-product structure, which is continuously embedded in the space of continuous functions $C^0 (\\Omega)$ , with $\\Omega \\subset \\mathbb{R}^2$ a bounded domain. An image is defined as a vector-valued function $\\mathcal{H}\\ni f:\\Omega \\to \\mathbb{R}^3$ . Assume we can access the function values of $f$ at the coordinates $\\{x_i\\}_{i = 1}^{nh}\\subset \\Omega$ with the biggest discretiza", + "bbox": [ + 496, + 795, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18249", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion size $h$ . Note that $x_{i}$ 's are not necessarily equidistant. Our goal is to learn a super-resolution neural operator between two Hilbert spaces with different resolutions: $S_{\\theta}: \\mathcal{H} \\supset \\mathcal{H}(\\Omega_{h_c}) \\to \\mathcal{H}(\\Omega_{h_f}) \\subset \\mathcal{H}$ , where $h_c, h_f$ denotes the coarse and the fine grid sizes, respectively. Given $N$ function pairs $\\{a^{(k)}, u^{(k)}\\}_{k=1}^N$ , where $a^{(k)} \\in \\mathcal{H}(\\Omega_{h_c})$ and $u^{(k)} \\in \\mathcal{H}(\\Omega_{h_f})$ . Our SRNO parameterized by $\\theta$ can be solved through the associated empirical-risk minimization problem:", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\| u _ {h _ {f}} ^ {(k)} - \\mathcal {S} _ {\\theta} \\left(a _ {h _ {c}} ^ {(k)}\\right) \\| _ {\\mathcal {H}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 229, + 468, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our data $a^{(k)}$ and $u^{(k)}$ are functions, to work with them numerically, we assume access only to their pointwise evaluations. Let $\\mathbb{A}_{h_c} \\subset \\mathcal{H}(\\Omega_{h_c}) \\subset \\mathcal{H}$ be an approximation space [3] associated with $\\{x_i\\}_{i=1}^{n_{h_c}}$ , such that for any $a \\in \\mathbb{A}_{h_c}$ , $a(\\cdot) = \\sum_{i=1}^{n_{h_c}} a(x_i) \\xi_{x_i}(\\cdot)$ , where $\\{\\xi_{x_i}(\\cdot)\\}$ form a set of nodal basis for $\\mathbb{A}_{h_c}$ in the sense that $\\xi_{x_i}(x_j) = \\delta_{ij}$ . Similarly, $\\mathbb{U}_{h_f} \\subset \\mathcal{H}(\\Omega_{h_f}) \\subset \\mathcal{H}$ is another approximation space associate with grid size $h_f$ . Note that Our observed LR-HR image pairs $\\{a, u\\}$ may have different grid size pairs $\\{h_c, h_f\\}$ , which means they belong to different approximation spaces $\\{\\mathbb{A}_{h_c}, \\mathbb{U}_{h_f}\\}$ .", + "bbox": [ + 75, + 273, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lifting. The neural operators from PDEs usually use a simple pointwise function, $\\mathcal{L}$ in (3), to expand the input channels. But for the SR problem, a deep feature encoder $E_{\\psi}$ (with trainable parameters $\\psi$ ) and spatial interpolations should be considered in the lifting operation, due to the complexity of natural images and the discretization inconsistency between LR and HR image functions. Another critical ingredient in Lifting is the incorporation scheme of coordinates. Prior study [14] demonstrates that deep CNNs can implicitly learn to encode the information about absolute positions. As shown in Fig.2 (c), the position information of the grid points $\\{a_l\\}_{l=1}^4$ have been implicitly encoded by $E_{\\psi}$ employing CNNs. Therefore, we only need to explicitly construct the coordinate features $\\hat{a}(x)$ with the fractional part of coordinate $x$ inside a grid. In order to reduce the blocky artifacts resulting from direct interpolation of the LR feature maps, we propose to concatenate the features weighted by the corresponding bilinear interpolation factors $s_l$ . Compared to the local ensemble trick in [5], our method takes shorter running time and overcomes the oversmoothing problem. We reformulate the lifting operation in (3) as:", + "bbox": [ + 75, + 441, + 472, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {0} (x) = \\mathcal {L} (\\mathbf {c}, \\left\\{s _ {l} \\cdot E _ {\\psi} \\left(a \\left(\\hat {x} _ {l}\\right)\\right), \\delta_ {l} (x) \\right\\} _ {l = 1} ^ {4}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 782, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ are coordinates of HR image functions, $\\hat{x}_l$ are the coordinates of the four neighbors of $x$ , $\\delta_l(x) = x - \\hat{x}_l$ , $\\mathbf{c} = (2 / r_x, 2 / r_y)$ represents a $r_x \\times r_y$ local area in HR images with $r_x$ and $r_y$ the scaling factors, and $\\mathcal{L}: \\mathbb{R}^{4 \\times (d_e + 2) + 2} \\to \\mathbb{R}^{d_z}$ is an local linear transformation function, with $d_e$ the number of channels after $E_{\\psi}$ .", + "bbox": [ + 76, + 809, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kernel integral. The kernel integral operator, in (4), actually tells that we can identify the hidden representation $z(x)$ of the underlying image function $f(x)$ with distributions [28], inner products in the settings of SRNO. We first define the kernel $K_{t}:\\mathbb{R}^{d_{z} + d_{z}}\\to \\mathbb{R}^{d_{z}\\times d_{z}}$ with respect to the input pair $(z(x),z(y))$ , rather than like (6) depending on the spatial variables $(x,y)$ . Furthermore, in order to more efficiently explore $z(x)$ we can use multiple sets of test functions defining the distributions, which may immediately remind one of a single-head self-attention [17]. Denote $z_{i} = z(x_{i})\\in \\mathbb{R}^{d_{z}}$ for $i = 1,\\dots ,n_{h_f}$ . The kernel integral operator can be approximated by the Monte-Carlo method (omitting the layer index):", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} (\\mathcal {K} (z)) (x) = \\int_ {\\Omega} K (z (x), z (y)) z (y) \\mathrm {d} y \\\\ n _ {h _ {f}} \\end{array} \\tag {9} \\\\ \\approx \\sum_ {i = 1} ^ {\\dots h _ {f}} K (z (x), z _ {i}) z _ {i}, \\quad \\forall x \\in \\Omega_ {h _ {f}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 290, + 893, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 368, + 544, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK (z (x), z _ {i}) = \\frac {\\exp \\left(\\frac {\\langle W _ {q} z (x) , W _ {k} z _ {i} \\rangle}{\\sqrt {d _ {z}}}\\right)}{\\sum_ {j = 1} ^ {n _ {h _ {f}}} \\exp \\left(\\frac {\\langle W _ {q} z _ {j} , W _ {k} z _ {i} \\rangle}{\\sqrt {d _ {z}}}\\right)} W _ {v}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 381, + 890, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the language of transformers, the matrices $W_{q}, W_{k}, W_{v} \\in \\mathbb{R}^{d_{z} \\times d_{z}}$ correspond to the queries, keys and values functions respectively. By allowing $K$ to be the sum of multiple functions with separate trainable parameters, the multi-head self-attention can also be rewritten like in (9). For every observation $(a, u)$ , this operation has a complexity of $O(n_{h_{f}}^{2} d_{z})$ , which is unaffordable for the SR problem where the sampling number $n_{h_{f}}$ usually reaches to 10,000 or beyond. To overcome this issue, we employ the Galerkin-type attention operator which has a linear complexity $O(n_{h_{f}} d_{z}^{2})$ . The approximation capacity of a linearized Galerkin-type attention has been proved to be comparable to a Petrov-Galerkin projection [3].", + "bbox": [ + 496, + 430, + 893, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We suppose that a image function $f$ is locally integrable, i.e., $f$ is measurable and $\\int_{\\Omega} |f| < \\infty$ for every compact $\\Omega \\subset \\mathbb{R}^2$ . The idea behind distributions [28] is to identify $f$ with $\\int f \\phi$ by suitably choosing a \"test function\" $\\phi$ , rather than by a series of separate evaluations $f(x)$ at coordinates $x \\in \\Omega$ . Let $v(x) = W_v z(x)$ , $k(x) = W_k z(x)$ , and $q(x) = W_q z(x)$ be the keys, queries and values functions respectively, which are $d_z$ dimensional vector-valued functions, with a subscript denoting the corresponding component. The kernel integral attention then can be written in a component-wise form ( $j = 1, \\dots, d_z$ ):", + "bbox": [ + 496, + 630, + 893, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\left(\\mathcal {K} (z)\\right) (x)\\right) _ {j} = \\sum_ {l = 1} ^ {d _ {z}} \\left\\langle k _ {l}, v _ {j} \\right\\rangle q _ {l} (x) \\tag {11} \\\\ \\approx \\sum_ {l = 1} ^ {d _ {z}} \\left(\\int_ {\\Omega} k _ {l} (y) v _ {j} (y) \\mathrm {d} y\\right) q _ {l} (x), \\quad \\forall x \\in \\Omega_ {h _ {f}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 816, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18250", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denote the evaluations $(z(x_{1}),\\ldots ,z(x_{n_{h_{f}}}))^{T}$ by $\\mathbf{z}\\in$ $\\mathbb{R}^{n_{h_f}\\times d_z}$ . Let $Q = \\mathbf{z}W_{q},K = \\mathbf{z}W_{k},V = \\mathbf{z}W_{v}\\in$ $\\mathbb{R}^{n_{h_f}\\times d_z}$ . The columns of $Q / K / V$ contain the vector representations of the learned basis functions, spanning certain subspaces of the latent representation Hilbert spaces respectively.", + "bbox": [ + 76, + 89, + 468, + 186 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = Q \\left(\\widetilde {K} ^ {T} \\widetilde {V}\\right) / n _ {h _ {f}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 186, + 468, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{K} = \\operatorname{Ln}(K), \\widetilde{V} = \\operatorname{Ln}(K)$ , with $\\operatorname{Ln}(\\cdot)$ the layer normalization. The $j$ -th column of $\\widetilde{K}^T \\widetilde{V}$ contains the coefficients for the linear combination of the basis vectors $\\{q_j\\}_{j=1}^{d_z}$ to form the output $\\mathbf{z}$ . Consequently, the global correlations are reflected through the components of $\\widetilde{K}^T \\widetilde{V}$ . As a result, SRNO can utilize all information effectively without the feature unfolding [5] or local convolutions [18], significantly suppressing the discontinuous patterns that appear around feature boundaries.", + "bbox": [ + 76, + 210, + 468, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to providing a global aggregation for the output $\\mathbf{z}$ at each sampling point, the linear Galerkin-type attention (12) has the ability to obtain a quasi-optimal approximation in the current approximation space spanned with the columns of $Q$ [3]. However, the expressive capability of the current bases in $Q$ , $K$ , and $V$ solely depends on the current input LR image through the latent representation $\\mathbf{z}$ , is there any chance that we can enrich the bases which contains some extra and useful information for SR reconstruction, but not included in the input $\\mathbf{z}$ ? For this purpose, the point-wise FFN $\\mathcal{O}:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_z}$ in Fig.2 (b) introduces nonlinearities on one hand, and the positions concatenated in $\\mathbf{z}$ enhance the bases on the other. In this way, the basis functions are being constantly enriched, and we reformulate the iterative process (See supplementary for a principled discussion):", + "bbox": [ + 75, + 354, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} (x) = z _ {t} (x) + \\mathcal {O} ((\\mathcal {K} _ {t} (z _ {t})) (x) + z _ {t} (x)). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 604, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network details. The Network architecture is shown in Fig.2. As to the feature encoder $E_{\\psi}$ , we employ EDSR-baseline [21], or RDN [41], both of which drop their upsampling layers, and their output channel dimensions $d_{e} = 64$ . The CNNs in $E_{\\psi}$ assist SRNO in capturing the common basis functions from the ensemble of training samples, while the Galerkin-type attention layers provide the instance-specific basis enhancement. We employ the multi-head attention scheme in [36] by dividing the queries, keys and values into $n_{\\text{heads}}$ parts with each of dimension $d_{z} / n_{\\text{heads}}$ . In our implementation, $d_{z} = 256$ , $n_{\\text{heads}} = 16$ , yielding 16-dimensional output values. We only use two iterations ( $T = 2$ ) in the kernel integral operator, which already outperforms previous works, while keeping the running time advantage. Note that we utilize $1 \\times 1$ convolutions to replace all the linear layers in SRNO, since they have a GPU-friendly data structure. The detailed network structures are listed in the supplementary.", + "bbox": [ + 75, + 628, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 630, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training", + "text_level": 1, + "bbox": [ + 500, + 114, + 604, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We used the DIV2K dataset [1] for network training, while the DIV2K validation set [1] and four benchmark datasets, including Set5 [2], Set14 [39], B100 [25] and Urban100 [12], for evaluation. Peak signal-to-noise ratio (PSNR) is used as the evaluation metric. Similar to [5, 11], we cropped image boundaries when computing PSNRs.", + "bbox": [ + 498, + 138, + 890, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Numerically, we have to explore the underlying continuous image function through access to its point-wise evaluations on the HR and LR grids. Let $B$ be the batch size and $n_{h_c}$ be the LR sampling counts. We first sample $B$ random scales $r^{(i)}$ following a uniform distribution $\\mathcal{U}(1,4)$ , and then crop $B$ patches of sizes $\\{\\sqrt{n}_{h_c}r^{(i)}\\times \\sqrt{n}_{h_c}r^{(i)}\\}_{i = 1}^B$ from the HR training images (one per each). We find $n_{h_c} = 128^2$ is sufficient for SRNO to represent most natural image patches and fix this choice in the sequel. The LR counterparts are downsampled using bicubic interpolation with the corresponding $r^{(i)}$ . In order to keep a consistent number of supervision points for the LR patches of different scales in a single batch, we randomly sample $128^2$ HR pixels and calculate the corresponding fractional coordinates on the coarse grid associated with $r^{(i)}$ , as is done in [5]. We use an L1 loss [21] and the Adam [16] optimizer with an initial learning rate $4\\times 10^{-5}$ and the maximum $4\\times 10^{-4}$ . All models are trained for 1000 epochs with batch size 64, and the learning rate decays by the cosine annealing after a warm-up phase of 50 epochs.", + "bbox": [ + 498, + 229, + 892, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 544, + 620, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative result. On the DIV2K validation set, Table 1 presents a quantitative comparison among our SRNO and three existing arbitrary-scale SR methods, MetaSR [5, 11], LIIF [5], and LTE [18]. Results when EDSR-baseline [21], and RDN [41] are used as encoders are displayed in the top and bottom rows, respectively. One can observe that SRNO provides the best results over all scale factors, irrespective of the encoder employed. Additionally, on in-distribution scales $(\\times 2, \\times 3, \\times 4)$ , our method outperforms earlier works by wide margins.", + "bbox": [ + 498, + 566, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As RDN facilitates better reconstruction accuracy than EDSR-baseline in Tab.1, we choose RDN as the encoder in the comparisons over the four benchmark datasets. The results are listed in Tab.2, where SRNO obtains all the best performances. Note that only SRNO produces substantial improvements over RDN on every in-distribution scale, although the latter trains separate models for each scale. As the experimental settings in Tab.1 and 2 only differ in the decoder part, we conclude that the Galerkin-type attention mechanism employed in SRNO does contribute to better function approximation capability.", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative result. Figure 3 illustrates the visual results", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd0426702f33bfeae6d1f2418da3ef6b46c72dc177f57296e2092f32ce5e12f2.jpg", + "image_caption": [ + "LR Image" + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 267, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e91cfeaafc5249e79c6f62f954dd4454bd038cd5dca262627d56e27388ed08e.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 269, + 88, + 390, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/825f491808d54b660b4da944a130f3232eb296bef87516314bf4726f03f9dc13.jpg", + "image_caption": [ + "LIIF [5]" + ], + "image_footnote": [], + "bbox": [ + 393, + 88, + 514, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be9609ed140f485ae62d5be21306490b3371a6cff09dc11df817e2b6132b4c25.jpg", + "image_caption": [ + "LTE [18]" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 637, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/883ea36ecbec0662783d605e54c8a509b16bfb296ddfc3b4790fda4bad9b6408.jpg", + "image_caption": [ + "SRNO (ours)" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 758, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/48526af59f86c9902ba932eaf5b32acf5ebfb6d31d017cb98bde051a1efa11d9.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 759, + 88, + 880, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34b39f7b2e18c000d0108ce4270c88e12eb1d6f4dce5ffc1f4ef1576aaa63c4a.jpg", + "image_caption": [ + "Figure 3. Visual comparison on other zero-shot SR. The boxes in the first column indicate the areas that the close-ups on the right display. All methods are trained with continuous random scales in $\\times 1 - \\times 4$ and tested on $\\times 8 / \\times 12$ to evaluate the generalization capability to unknown scaling factors. RDN is used as the encoder for all methods." + ], + "image_footnote": [], + "bbox": [ + 86, + 345, + 261, + 443 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/000d2fa49b6761eb1a8b2b734692ebcfb7dddc4ff7fd03a879895c788f53183b.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 86, + 444, + 261, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b669bb5797577475d1bf76f1041d42035ce06d54384da873c2d2f908051515c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 347, + 415, + 443 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e96ae44bd28d2c30447fed97abf27444ecb99f0ddb13f9152845ff522e685758.jpg", + "image_caption": [ + "LIIF [5]" + ], + "image_footnote": [], + "bbox": [ + 261, + 444, + 415, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/560bf66cbf7ebf40df96a061f29366d74a9b25bd02987b085b2cee4fcea7a16e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 347, + 571, + 443 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcdea1cca7989fa75a045bb33ee49a1669bd7d86e3d825a19333c7450cbfb226.jpg", + "image_caption": [ + "LTE [18]" + ], + "image_footnote": [], + "bbox": [ + 416, + 444, + 570, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df8d331c306c59f4cfad06bcc5d4b097b0eb184c01a3ae5c326fa08cd60dd322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 347, + 725, + 443 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4be934e6402d0f6565dea6dcb0865fed227593ee66ac3e6acb7539f6dfdde7f7.jpg", + "image_caption": [ + "SRNO (ours)", + "Figure 4. Visual comparison on non-integer scales. All methods use RDN as the encoder and are trained with continuous random scales in $\\times 1 - \\times 4$ ." + ], + "image_footnote": [], + "bbox": [ + 571, + 444, + 723, + 537 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21a95dc3d49abdc3f5abc335caa86fc6081b67bcbe16ae45978a9a601640615a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 347, + 880, + 443 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8ce4b02a4aa04fab8f37bbea4290623205ff0d34d3f8869774d963c207cdbe5.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 725, + 444, + 880, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtained from two images in the Urban100 and DIV2K validation datasets. Both the regions marked with a rectangle are rich in high-frequency structures. The bicubic interpolations exhibit severe aliasing artifacts, which obviously have been transferred into the latent codes of LIIF [5] as it directly relies on the interpolation of the feature maps and the bilinear ensemble. Thanks to the local Fourier frequency estimation step, LTE [18] performs somewhat better in partial regions (e.g. the top-right corner of the building in the second row), but aliasing still dominates on the whole. In contrast, SRNO successfully recovers the fine structures, which again verifies the importance of the global integral kernels in capturing the correct overall structures.", + "bbox": [ + 75, + 604, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 shows two examples of scene text images on non-integer scales. In the first row, only SRNO can clearly recover the word \"ANTIQUES\". And in the second row, only SRNO can consistently restore the two e's. Obviously, the Galerkin-type attention attributes to this global consistency.", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 604, + 653, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data sampling. The quantity of sampling points is crucial in our super-resolution operator learning architecture. Figure 5 shows that almost all scales perform better when using additional pixel samples. However, the benefit becomes less as the numbers of samples increases. $96^{2}$ or $128^{2}$ points would be good trade-offs between training time and performance. Figure 6 shows that more sampling points can more accurately represent an RGB-based function, which facilitates to learn the map between two approximation spaces. Using the DIV2K (800 images) and Flickr2K (2650 images) datasets, we also train an EDSR-baseline-MLP with position embedding, denoted by $\\mathbf{MLP}^{\\varphi}$ . We can see that our SRNO $(64^{2}/96^{2}/128^{2})$ obtains better performance for scales $\\times 3$ , $\\times 4$ , $\\times 6$ , $\\times 8$ , trained with only 800 images in DIV2K. The curve for $\\mathrm{MLP}(128^{2})$ in Fig.5 reflects the performance advantage of SRNO $(128^{2})$ , both trained with DIV2K dataset. Random and sequential sampling results are presented in Supplementary.", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18252", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/725d5f433ee0999f5ddd039dba2f18de096fda1dde0f1688213a437b8b666298.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIn-distributionOut-of-distribution
×2×3×4×6×12×18×24×30
Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [21]34.5530.9028.94-----
EDSR-baseline-MetaSR [5,11]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-LTE [18]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-SRNO (ours)34.8531.1129.1626.9023.8422.2921.2720.56
RDN-MetaSR [5,11]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-LTE [18]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-SRNO (ours)35.1631.4229.4227.1224.0322.4621.4120.68
", + "bbox": [ + 200, + 87, + 772, + 243 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a0dd12f77d9fb82989ffc5b062257d5d9dba1364d3ccf666fcff8dd742420992.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on DIV2K validation set (PSNR (dB)). The best performance are bolded. EDSR-baseline trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in $\\times 1 - \\times 4$ ." + ], + "table_footnote": [], + "table_body": "
MethodIn-distributionOut-of-distributionIn-distributionOut-of-distribution
×2×3×4×6×8×2×3×4×6×8
Set5Set14
RDN [41]38.2434.7132.47--34.0130.5728.81--
RDN-MetaSR [5, 11]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.97
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.15
RDN-LTE [18]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.16
RDN-SRNO (ours)38.3234.8432.6929.3827.2834.2730.7128.9726.7625.26
B100Urban100
RDN [41]32.3429.2627.72--32.8928.8026.61--
RDN-MetaSR [5, 11]32.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LIIF [5]32.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-LTE [18]32.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-SRNO (ours)32.4329.3727.8326.0424.9933.3329.1426.9824.4323.02
", + "bbox": [ + 156, + 292, + 820, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/509000c426838d5e84106f14ec1da6e75946067ce4a4d045df13b45c3cd3e426.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on benchmark datasets (PSNR (dB)). The best performances are in bold. RDN trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in $\\times 1 - \\times 4$" + ], + "table_footnote": [], + "table_body": "
Method#Params. (M)#FLOPs (G)
EDSR-baseline-LIIF1.685.0
EDSR-baseline-LTE1.775.3
EDSR-baseline-SRNO (ours)2.065.8
RDN-LIIF22.4765.2
RDN-LTE22.5755.5
RDN-SRNO (ours)22.8746.0
", + "bbox": [ + 98, + 532, + 452, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of model parameters and FLOPs. Metrics are measured under the setting of $128^{2}$ sample points.", + "bbox": [ + 76, + 625, + 468, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Other design choices. In Tab.4, we retrain the following models using EDSR-baseline [21]. All the tests were conducted using a single NVIDIA RTX 3090. In comparing our interpolation-free method to SRNO(-i) using nearest-neighbor interpolation, we find that it works better and faster. It should be noted that we only generate the image once, as opposed to employing an MLP four times as in LIIF [5] and LTE [18]. Figure 7 shows that SRNO, compared to SRNO(-i), alleviates the blocky artifact, and compared to the local ensemble [18], overcomes the oversmoothing problem. By comparing SRNO with SRNO(-h), it can be shown that cutting the number of heads from 16 to 8 causes an increase in time costs because GPU's parallelism is reduced. We see that reducing the number of basis functions and the iterative updating layers results in a", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47fcb728c79952a60a3eb99c9bfc229849dc434e0bd30dcd79e6ffd3dad3fad0.jpg", + "image_caption": [ + "Figure 5. Data samples vs. PSNRs. Evaluated on the Urban100. We train models using $48^{2} / 64^{2} / 96^{2} / 128^{2}$ pixel samples and calculate the PSNR increment relative to LTE. $\\varphi$ indicates training on DIV2K [1] (800 images) plus Flickr2K [33] (2650 images) dataset." + ], + "image_footnote": [], + "bbox": [ + 527, + 536, + 866, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "considerable performance drop by comparing SRNO with SRNO(-w) and SRNO(-l). The number of basis functions $(d_z)$ and the number of subspaces $(n_{\\text{heads}})$ play an important role in the approximation power of our model.", + "bbox": [ + 498, + 771, + 892, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Dynamic Basis", + "text_level": 1, + "bbox": [ + 500, + 845, + 651, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate the dynamic basis updating phenomenon across the approximation subspaces in this section. We", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2044e06426e8c62a221facb2786f11b500859df8d43c6051e656d83ecb1a59d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIn-dis.Out-of-dis.Time (ms)
×2×3×4×6×8
SRNO33.8330.5028.7926.5525.05149
SRNO (-i)33.8130.4728.7626.5525.00156
SRNO (-h)33.8730.4728.7626.5625.02153
SRNO (-w)33.8030.4528.7526.5325.00124
SRNO (-l)33.8230.4528.7526.5525.00117
LIIF [5]33.6830.3628.6426.4624.94183
LTE [18]33.7230.3728.6526.5024.99205
LTE+ [18]33.7130.4128.6726.4924.98160
", + "bbox": [ + 81, + 88, + 472, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative ablation study on design choices of SRNO. Evaluated on the Set14 (PSNR(dB)). -h/w/l refers to $n_{\\text{heads}} = 8$ , $d_z = 128$ , and $T = 1$ correspondingly. -i refers to using nearest-neighborhood interpolation and position embedding. EDSR-baseline is used in lifting. The running time is averaged over Set14 on $\\times 2/3/4/6/8$ SR. For each scale SR, we run every model three times and choose the minimum running time.", + "bbox": [ + 75, + 218, + 470, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e7a8a0b139c49ab63c9b550cc267a7c8c55951b849abfef372032801904cdf0.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 86, + 325, + 209, + 380 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a10d1b959a59eef5e1fdc24f3be3f27577b4764b4ee8fbd275603de62e49ee0d.jpg", + "image_caption": [ + "SRNO $(48^{2})$" + ], + "image_footnote": [], + "bbox": [ + 210, + 325, + 334, + 380 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a4fe9cf1714ab1205905d12439dc3714d161207350b20afeaf60e99192ded3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 325, + 460, + 380 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54b7c0d48005834ecae47f16e58e6872d750433dd29996d10a5bb52279ab6999.jpg", + "image_caption": [ + "SRNO $(96^{2})$" + ], + "image_footnote": [], + "bbox": [ + 86, + 397, + 209, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3327236be7feb94ddbe9f32b927f0bca7acc719c6a1843be3e97ee0ac1302a24.jpg", + "image_caption": [ + "SRNO $(128^{2})$" + ], + "image_footnote": [], + "bbox": [ + 210, + 397, + 334, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2e62743bd075ebccf3a343d5ea855c14df138552e87fe2bc1323f6e2f51e101.jpg", + "image_caption": [ + "SRNO $(64^{2})$", + "GT" + ], + "image_footnote": [], + "bbox": [ + 334, + 397, + 460, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db291cadddc205beffb22995687137e328b777d3069979df3562cff2578a470c.jpg", + "image_caption": [ + "Figure 6. Effects of sampling global image structures. Test on $\\times 8$ scale. EDSR-baseline is used as the encoder, while the data in parentheses indicate the pixel numbers sampled in the HR images during training." + ], + "image_footnote": [], + "bbox": [ + 91, + 539, + 272, + 599 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b22024d9b9f4740c4e9ea6b7ccbf9e902f562134ac17ee535a281bf9542fbff0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 539, + 454, + 599 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37dacdff4b4147a94590800ff0e304ae9e78377f8f456c4889da294bae421134.jpg", + "image_caption": [ + "LTE", + "SRNO", + "Figure 7. Interpolation vs. interpolation-free of SRNO. Test on $\\times 8$ scale. EDSR-baseline is used as an encoder. -i refers to using nearest-neighborhood interpolation and position embedding." + ], + "image_footnote": [], + "bbox": [ + 91, + 617, + 272, + 678 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f20ac7ba617af700e2c7d0f2e3f247d1cd2af586f71af502fee684edcaf096ba.jpg", + "image_caption": [ + "SRNO (-i)", + "GT" + ], + "image_footnote": [], + "bbox": [ + 274, + 617, + 454, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "show two exemplar basis functions in the two consecutive Galerkin attention layers in Fig.8. The basis function in the second layer appears to be more structured than in the previous layer. The dynamic update of the approximation subspaces spanned by the column vectors of $Q / K / V$ comes from two factors, the concatenated random coordinates and the nonlinearity introduced by the FNN. Without them, the Galerkin-type attention would just be a linear combination of the bases in the current approximation subspaces, and", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29aeeff4b6814001c6ea44ea010f290411846abb9a254ea79ed3cb4f151d48f8.jpg", + "image_caption": [ + "Figure 8. Dynamic basis update. The annotation $t = k$ refers to the iterative layer number. The pictures display one of the learned basis functions in $V$ and the latent representations $\\mathbf{z}$" + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 883, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f242339dc087f80730c3780d06708d78e9dca9de35a1e1be5105fc0cfa4a03c.jpg", + "image_caption": [ + "Figure 9. Ranks of the matrix $\\mathbf{z}$ . Evaluated on the Set5 dataset." + ], + "image_footnote": [], + "bbox": [ + 524, + 292, + 870, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SRNO would not be able to take any chances to enrich the basis for every individual image through optimizations. The increasing ranks of the latent representation matrix $\\mathbf{z}$ , shown in Fig.9, provide the evidence that Galerkin-type attention is becoming more complex than a straightforward linear combination of the existing bases. Due to the space limitation, we have provided a principled discussion on the dynamic basis update processes in the supplementary material.", + "bbox": [ + 496, + 462, + 890, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed the Super-Resolution Neural Operator (SRNO) for continuous super-resolution. In SRNO, each image is seen as a function, and our method learns a map between finite-dimensional function spaces, which allows SRNO can be trained and generalize on different levels of discretization. First, in the Lifting, we use CNN-based encoders to capture feature maps from LR images and design a simple but efficient interpolation-free method, addressing the discretization inconsistency problem encountered in SR. Second, to approximate our target function in the Iterative Kernel Integration, we employ a linear attention operator that has been proved to be comparable to the Petrov-Galerkin projection. Finally, we map the last hidden representation to the output function. Experimental results show that our SRNO outperforms other arbitrary-scale SR methods in performance as well as computation time, and particularly in the capability of capturing the global image structures.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 7", + "[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 5", + "[3] Shuhao Cao. Choose a transformer: Fourier or galerkin. Advances in Neural Information Processing Systems, 34:24924-24940, 2021. 2, 4, 5", + "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 2", + "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7, 8", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2", + "[7] Alexandre Ern and Jean-Luc Guermond. Theory and practice of finite elements, volume 159. Springer, 2004. 3", + "[8] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 2", + "[9] John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Efficient token mixing for transformers via adaptive fourier neural operators. In International Conference on Learning Representations, 2021. 2", + "[10] Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations. Advances in Neural Information Processing Systems, 34:24048-24062, 2021. 2", + "[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2, 5, 7", + "[12] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5", + "[13] Rakhoon Hwang, Jae Yong Lee, Jin Young Shin, and Hyung Ju Hwang. Solving pde-constrained control problems using operator learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4):4504-4512, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? In International Conference on Learning Representations, 2019. 4", + "[15] Chiyu\" Max\" Jiang, Soheil Esmaeilzadeh, Kamyar Azizzadenesheli, Karthik Kashinath, Mustafa Mustafa, Hamdi A Tchelepi, Philip Marcus, and Anima Anandkumar. Meshfreeflownet: a physics-constrained deep continuous spacetime super-resolution framework. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-15, 2020. 1", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5", + "[17] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces. arXiv preprint arXiv:2108.08481, 2021. 2, 4", + "[18] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 5, 6, 7, 8", + "[19] Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, Anima Anandkumar, et al. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2020. 2, 3", + "[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833-1844, 2021. 1, 2", + "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 7", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[23] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218-229, 2021. 2", + "[24] Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister. Dynamic high-pass filtering and multi-spectral attention for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4288-4297, 2021. 1, 2", + "[25] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18255", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 5", + "[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1", + "[27] Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214, 2022. 2", + "[28] Walter Rudin. Functional analysis, mcrawhill. Inc, New York, 45:46, 1991. 4", + "[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 1, 2", + "[30] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 1, 2", + "[31] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 2", + "[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2", + "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 7", + "[34] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2", + "[35] V Vapnik. Statistical learning theory wiley-interscience. New York, 1998. 2", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5", + "[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 1", + "[39] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5", + "[40] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2", + "[41] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 1, 2, 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18256", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_model.json b/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b99b377745662a7786e4a07cacbc7065cc399d5 --- /dev/null +++ b/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.308, + 0.131, + 0.662, + 0.154 + ], + "angle": 0, + "content": "Super-Resolution Neural Operator" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.181, + 0.614, + 0.2 + ], + "angle": 0, + "content": "Min Wei* Xuesong Zhang*†" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.2, + 0.721, + 0.217 + ], + "angle": 0, + "content": "Beijing University of Posts and Telecommunications, China" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.22, + 0.62, + 0.234 + ], + "angle": 0, + "content": "{mw，xuesong_zhang}@bupt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.681 + ], + "angle": 0, + "content": "We propose Super-resolution Neural Operator (SRNO), a deep operator learning framework that can resolve high-resolution (HR) images at arbitrary scales from the low-resolution (LR) counterparts. Treating the LR-HR image pairs as continuous functions approximated with different grid sizes, SRNO learns the mapping between the corresponding function spaces. From the perspective of approximation theory, SRNO first embeds the LR input into a higher-dimensional latent representation space, trying to capture sufficient basis functions, and then iteratively approximates the implicit image function with a kernel integral mechanism, followed by a final dimensionality reduction step to generate the RGB representation at the target coordinates. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial domain and therefore benefits the grid-free continuum; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to \"hallucinate\" high-frequency information from the LR image. Experiments show that SRNO outperforms existing continuous SR methods in terms of both accuracy and running time. Our code is at https://github.com/2y7c3/Super-Resolution-Neural-Operator." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.21, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Single image super-resolution (SR) addresses the inverse problem of reconstructing high-resolution (HR) images from their low-resolution (LR) counterparts. In a data-driven way, deep neural networks (DNNs) learn the inversion map from many LR-HR sample pairs and have demonstrated appealing performances [4, 20, 21, 24, 29, 40, 41]. Nevertheless, most DNNs are developed in the configuration of single scaling factors, which cannot be used in sce" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.272, + 0.88, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.516, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Figure 1. Overview of Super-Resolution Neural operator (SRNO). SRNO is composed of three parts, \\(\\mathcal{L}\\) (Lifting), \\(\\kappa\\) (kernel integrals) and \\(\\mathcal{P}\\) (Projection), which perform consecutively to learn mappings between approximation spaces \\(\\mathcal{H}(\\Omega_{h_c})\\) and \\(\\mathcal{H}(\\Omega_{h_f})\\) associated with grid sizes \\(h_c\\) and \\(h_f\\), respectively. The key component, \\(\\kappa\\), uses test functions in the latent Hilbert space \\(\\mathcal{V}(\\Omega_{h_f})\\) to seek instance-specific basis functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "narios requiring arbitrary SR factors [37, 38]. Recently, implicit neural functions (INF) [5, 18] have been proposed to represent images in arbitrary resolution, and paving a feasible way for continuous SR. These networks, as opposed to storing discrete signals in grid-based formats, represent signals with evaluations of continuous functions at specified coordinates, where the functions are generally parameterized by a multi-layer perceptron (MLP). To share knowledge across instances instead of fitting individual functions for each signal, encoder-based methods [5, 15, 26] are proposed to retrieve latent codes for each signal, and then a decoding MLP is shared by all the instances to generate the required output, where both the coordinates and the corresponding latent codes are taken as input. However, the point-wise behavior of MLP in the spatial dimensions results in limited performance when decoding various objects, particularly for high-frequency components [30, 32]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions. This work is supported by the National Natural Science Foundation of China (61871055). \n†Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18247" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "Neural operator is a newly proposed neural network architecture in the field of computational physics [17, 19, 23] for numerically efficient solvers of partial differential equations (PDE). Stemming from the operator theory, nerval operators learn mappings between infinite-dimensional function spaces, which is inherently capable of continuous function evaluations and has shown promising potentials in various applications [9, 13, 27]. Typically, neural operator consists of three components: 1) lifting, 2) iterative kernel integral, and 3) projection. The kernel integrals operate in the spatial domain, and thus can explicitly capture the global relationship constraining the underlying solution function of the PDE. The attention mechanism in transformers [36] is a special case of kernel integral where linear transforms are first exerted to the feature maps prior to the inner product operations [17]. Tremendous successes of transformers in various tasks [6, 22, 34] have shown the importance of capturing global correlations, and this is also true for SR to improve performance. [8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.473, + 0.651 + ], + "angle": 0, + "content": "In this paper, we propose the super-resolution neural operator (SRNO), a deep operator learning framework that can resolve HR images from their LR counterparts at arbitrary scales. As shown in Fig.1, SRNO learns the mapping between the corresponding function spaces by treating the LR-HR image pairs as continuous functions approximated with different grid sizes. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial dimensions and have been proved to be comparable to a Petrov-Galerkin projection [3]; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to \"hallucinate\" high-frequency information from the LR image. When employing same encoders to capture features, our method outperforms previous continuous SR methods in terms of both reconstruction accuracy and running time." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.651, + 0.435, + 0.665 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.673, + 0.47, + 0.749 + ], + "angle": 0, + "content": "- We propose the methodology of super-resolution neural operator that maps between finite-dimensional function spaces, allowing for continuous and zero-shot super-resolution irrespective the discretization used on the input and output spaces." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.756, + 0.47, + 0.817 + ], + "angle": 0, + "content": "- We develop an architecture for SRNO that first explores the common latent basis for the whole training set and subsequently refines an instance-specific basis by the Galerkin-type attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "- Numerically, we show that the proposed SRNO outperforms existing continuous SR methods with less running time, and even generates better results on the resolutions for which the fixed scale SR networks were trained." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.673, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Deep learning based SR methods. [4, 20, 21, 24, 29, 40, 41] have achieved impressive performances, in multiscale scenarios one has to train and store several models for each scale factor, which is unfeasible when considering time and memory budgets. In recent years, several methods [11, 31, 37] are proposed to achieve arbitrary-scale SR with a single model, but their performances are limited when dealing with out-of-distribution scaling factors. Inspired by INF, LIIF [5] takes continuous coordinates and latent variables as inputs, and employs an MLP to achieve outstanding performances for both in-distribution and out-of-distribution factors. In contrast, LTE [18] transforms input coordinates into the Fourier domain and uses the dominant frequencies extracted from latent variables to address the spectral bias problem [30, 32]. In a nutshell, treating images as RGB-valued functions and sharing the implicit function space are the keys to the success of LIIF-like works [5, 18]. Nevertheless, a purely local decoder, like MLP, is not able to accurately approximate arbitrary images, although it is rather sensitive to the input coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Neural Operators. Recently, a novel neural network architecture, Neural Operator (NO), was proposed for discretization invariant solutions of PDEs via infinitedimensional operator learning [10,17,19,23]. Neural operators only need to be trained once and are capable of transferring solutions between differently discretized meshes while keeping a fixed approximation error. A valuable merit of NO is that it does not require knowledge of the underlying PDE, which allows us to introduce it by the following abstract form," + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.579, + 0.891, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\mathbf {L} _ {a} u\\right) (x) = f (x), \\quad x \\in D, \\tag {1} \\\\ u (x) = 0, \\quad x \\in \\partial D, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.746 + ], + "angle": 0, + "content": "where \\(u: D \\to \\mathbb{R}^{d_u}\\) is the solution function residing in the Banach space \\(\\mathcal{U}\\), and \\(\\mathbf{L}: \\mathcal{A} \\to \\mathbf{L}(\\mathcal{U}; \\mathcal{U}^*)\\) is an operator-valued functional that maps the coefficient function \\(a \\in \\mathcal{A}\\) of the PDE to \\(f \\in \\mathcal{U}^*\\), the dual space of \\(\\mathcal{U}\\). As in many cases the inverse operator of \\(\\mathbf{L}\\) even does not exist, NO seeks a feasible operator \\(\\mathcal{G}: \\mathcal{A} \\to \\mathcal{U}, a \\mapsto u\\), directly mapping the coefficient to the solution within an acceptable tolerance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The operator \\(\\mathcal{G}\\) is numerically approximated by training a neural network \\(\\mathcal{G}_{\\theta}:\\mathcal{A}\\to \\mathcal{U}\\), where \\(\\theta\\) are the trainable parameters. Suppose we have \\(N\\) pairs of observations \\(\\{a_j,u_j\\}_{j = 1}^N\\) where the input functions \\(a_{j}\\) are sampled from probability measure \\(\\mu\\) compactly supported on \\(\\mathcal{A}\\), and \\(u_{j} = \\mathcal{G}(a_{j})\\) are used as the supervisory output functions. The infinitely dimensional operator learning problem \\(\\mathcal{G}\\gets \\mathcal{G}_{\\theta}\\) thus is associated with the empirical-risk minimization problem [35]. In practice, we actually measure the approximation loss using the sampled observations \\(u_{o}^{(j)}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18248" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.856, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.177, + 0.47, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.269, + 0.472, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.177, + 0.877, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.893, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Super-resolution neural operator (SRNO) architecture for continuous SR. The input LR image \\( f_{h_c} \\) undergoes three phases to output the HR image \\( f_{h_f} \\) with the specified resolution: (a) Lifting the LR pixel values \\( a(\\mathbf{x}) \\) on the set of coordinates \\( \\mathbf{x} = \\{x_i\\}_{i=1}^{n_{h_f}} \\) to a higher dimensional feature space by a CNN-based encoder \\( E_{\\psi} \\), constructing the latent representation \\( \\hat{a}(x) \\), and linearly transforming into the first layer's input \\( z_0(\\mathbf{x}) \\). (b) kernel integrals composed of \\( T \\) layers of Galerkin-type attention, and (c) finally project to the RGB space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.438, + 0.433, + 0.455 + ], + "angle": 0, + "content": "and \\(a_{o}^{(j)}\\), which are the direct results of discretization:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.463, + 0.469, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\mathbb {E} _ {a \\sim \\mu} \\| \\mathcal {G} (a) - \\mathcal {G} _ {\\theta} (a) \\| _ {\\mathcal {U}} \\\\ \\approx \\min _ {\\theta} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\| u _ {o} ^ {(j)} - \\mathcal {G} _ {\\theta} \\left(a _ {o} ^ {(j)}\\right) \\| _ {\\mathcal {U}}. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.54, + 0.47, + 0.63 + ], + "angle": 0, + "content": "Similar to classical feedforward neural networks (FFNs), the NO is of an iterative architecture. For ease of exposition, suppose \\(\\mathcal{A}\\) is defined on the bounded domain \\(D\\subset\\) \\(\\mathbb{R}^d\\) , and the inputs and outputs of the intermediate layers are all vector-valued functions, with dimension \\(d_{z}\\) . Then \\(\\mathcal{G}_{\\theta}:\\mathcal{A}\\to \\mathcal{U}\\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.64, + 0.469, + 0.657 + ], + "angle": 0, + "content": "\\[\nz _ {0} (x) = \\mathcal {L} (x, a (x)), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.659, + 0.469, + 0.675 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} (x) = \\sigma \\left(W _ {t} z _ {t} (x) + \\left(\\mathcal {K} _ {t} \\left(z _ {t}; \\Phi\\right)\\right) (x)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.678, + 0.469, + 0.695 + ], + "angle": 0, + "content": "\\[\nu (x) = \\mathcal {P} \\left(z _ {T} (x)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.703, + 0.469, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}:\\mathbb{R}^{d_a + d}\\to \\mathbb{R}^{d_z}\\), and \\(\\mathcal{P}:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_u}\\) are the local lifting and projection functions respectively, mapping the input \\(a\\) to its first layer hidden representation \\(z_{0}\\) and the last layer hidden representation \\(z_{T}\\) back to the output function \\(u\\). \\(W:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_z}\\) is a point-wise linear transformation, and \\(\\sigma :\\mathbb{R}^{d_z}\\rightarrow \\mathbb{R}^{d_z}\\) is the nonlinear activation function." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Although the PDE in (1) point-wisely defines the behavior of the solution function \\( u(x) \\), the solution operator \\( \\mathcal{G} \\) we are seeking should exhibit the non-local property such that \\( \\mathcal{G}(a) \\) can approximate \\( u \\) everywhere rather than locally. For this purpose, NO may employ the kernel integral operators \\( \\mathcal{K}_t: \\{z_t : D_t \\to \\mathbb{R}^{d_z}\\} \\mapsto \\{z_{t+1} : D_{t+1} \\to \\mathbb{R}^{d_z}\\} \\) to maintain the continuum in the spatial domain, and one of the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.44, + 0.75, + 0.455 + ], + "angle": 0, + "content": "most adopted forms [19] is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.467, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathcal {K} _ {t} \\left(z _ {t}; \\Phi\\right)\\right) (x) = \\int_ {D} K _ {t} (x, y; \\Phi) z _ {t} (y) \\mathrm {d} y, \\quad \\forall x \\in D, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.511, + 0.892, + 0.542 + ], + "angle": 0, + "content": "where the kernel matrix \\(K_{t}:\\mathbb{R}^{d + d}\\to \\mathbb{R}^{d_{z}\\times d_{z}}\\) is parameterized by \\(\\Phi\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.809, + 0.575 + ], + "angle": 0, + "content": "3. Super-resolution Neural Operator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.794 + ], + "angle": 0, + "content": "All operations of NO are defined in function space and the training data are just samples at coordinates \\( x_{i} \\) irrespective of the discretization sizes. This property inspires us to employ the NO's architecture for continuous SR. Different from NO that aims at mapping between infinitedimensional function spaces, we define SRNO in the sense of mappings between two approximation spaces of finitedimensional but continuous functions. This is reasonable because imaging optics inevitably limit the highest possible frequency in the radiance field under observation. With this configuration, we can design the architecture of SRNO using the well-developed theoretical tool, the Galerkin-type method that is widely used in the filed of Finite elements [7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Problem Formulation. Let \\((\\mathcal{H},\\langle \\cdot ,\\cdot \\rangle_{\\mathcal{H}})\\) denote a Hilbert space equipped with the inner-product structure, which is continuously embedded in the space of continuous functions \\(C^0 (\\Omega)\\), with \\(\\Omega \\subset \\mathbb{R}^2\\) a bounded domain. An image is defined as a vector-valued function \\(\\mathcal{H}\\ni f:\\Omega \\to \\mathbb{R}^3\\). Assume we can access the function values of \\(f\\) at the coordinates \\(\\{x_i\\}_{i = 1}^{nh}\\subset \\Omega\\) with the biggest discretiza" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18249" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "tion size \\( h \\). Note that \\( x_{i} \\)'s are not necessarily equidistant. Our goal is to learn a super-resolution neural operator between two Hilbert spaces with different resolutions: \\( S_{\\theta}: \\mathcal{H} \\supset \\mathcal{H}(\\Omega_{h_c}) \\to \\mathcal{H}(\\Omega_{h_f}) \\subset \\mathcal{H} \\), where \\( h_c, h_f \\) denotes the coarse and the fine grid sizes, respectively. Given \\( N \\) function pairs \\( \\{a^{(k)}, u^{(k)}\\}_{k=1}^N \\), where \\( a^{(k)} \\in \\mathcal{H}(\\Omega_{h_c}) \\) and \\( u^{(k)} \\in \\mathcal{H}(\\Omega_{h_f}) \\). Our SRNO parameterized by \\( \\theta \\) can be solved through the associated empirical-risk minimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.23, + 0.47, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\| u _ {h _ {f}} ^ {(k)} - \\mathcal {S} _ {\\theta} \\left(a _ {h _ {c}} ^ {(k)}\\right) \\| _ {\\mathcal {H}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Since our data \\(a^{(k)}\\) and \\(u^{(k)}\\) are functions, to work with them numerically, we assume access only to their pointwise evaluations. Let \\(\\mathbb{A}_{h_c} \\subset \\mathcal{H}(\\Omega_{h_c}) \\subset \\mathcal{H}\\) be an approximation space [3] associated with \\(\\{x_i\\}_{i=1}^{n_{h_c}}\\), such that for any \\(a \\in \\mathbb{A}_{h_c}\\), \\(a(\\cdot) = \\sum_{i=1}^{n_{h_c}} a(x_i) \\xi_{x_i}(\\cdot)\\), where \\(\\{\\xi_{x_i}(\\cdot)\\}\\) form a set of nodal basis for \\(\\mathbb{A}_{h_c}\\) in the sense that \\(\\xi_{x_i}(x_j) = \\delta_{ij}\\). Similarly, \\(\\mathbb{U}_{h_f} \\subset \\mathcal{H}(\\Omega_{h_f}) \\subset \\mathcal{H}\\) is another approximation space associate with grid size \\(h_f\\). Note that Our observed LR-HR image pairs \\(\\{a, u\\}\\) may have different grid size pairs \\(\\{h_c, h_f\\}\\), which means they belong to different approximation spaces \\(\\{\\mathbb{A}_{h_c}, \\mathbb{U}_{h_f}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.775 + ], + "angle": 0, + "content": "Lifting. The neural operators from PDEs usually use a simple pointwise function, \\(\\mathcal{L}\\) in (3), to expand the input channels. But for the SR problem, a deep feature encoder \\(E_{\\psi}\\) (with trainable parameters \\(\\psi\\)) and spatial interpolations should be considered in the lifting operation, due to the complexity of natural images and the discretization inconsistency between LR and HR image functions. Another critical ingredient in Lifting is the incorporation scheme of coordinates. Prior study [14] demonstrates that deep CNNs can implicitly learn to encode the information about absolute positions. As shown in Fig.2 (c), the position information of the grid points \\(\\{a_l\\}_{l=1}^4\\) have been implicitly encoded by \\(E_{\\psi}\\) employing CNNs. Therefore, we only need to explicitly construct the coordinate features \\(\\hat{a}(x)\\) with the fractional part of coordinate \\(x\\) inside a grid. In order to reduce the blocky artifacts resulting from direct interpolation of the LR feature maps, we propose to concatenate the features weighted by the corresponding bilinear interpolation factors \\(s_l\\). Compared to the local ensemble trick in [5], our method takes shorter running time and overcomes the oversmoothing problem. We reformulate the lifting operation in (3) as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.784, + 0.47, + 0.801 + ], + "angle": 0, + "content": "\\[\nz _ {0} (x) = \\mathcal {L} (\\mathbf {c}, \\left\\{s _ {l} \\cdot E _ {\\psi} \\left(a \\left(\\hat {x} _ {l}\\right)\\right), \\delta_ {l} (x) \\right\\} _ {l = 1} ^ {4}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(x\\) are coordinates of HR image functions, \\(\\hat{x}_l\\) are the coordinates of the four neighbors of \\(x\\), \\(\\delta_l(x) = x - \\hat{x}_l\\), \\(\\mathbf{c} = (2 / r_x, 2 / r_y)\\) represents a \\(r_x \\times r_y\\) local area in HR images with \\(r_x\\) and \\(r_y\\) the scaling factors, and \\(\\mathcal{L}: \\mathbb{R}^{4 \\times (d_e + 2) + 2} \\to \\mathbb{R}^{d_z}\\) is an local linear transformation function, with \\(d_e\\) the number of channels after \\(E_{\\psi}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Kernel integral. The kernel integral operator, in (4), actually tells that we can identify the hidden representation \\( z(x) \\) of the underlying image function \\( f(x) \\) with distributions [28], inner products in the settings of SRNO. We first define the kernel \\( K_{t}:\\mathbb{R}^{d_{z} + d_{z}}\\to \\mathbb{R}^{d_{z}\\times d_{z}} \\) with respect to the input pair \\( (z(x),z(y)) \\), rather than like (6) depending on the spatial variables \\( (x,y) \\). Furthermore, in order to more efficiently explore \\( z(x) \\) we can use multiple sets of test functions defining the distributions, which may immediately remind one of a single-head self-attention [17]. Denote \\( z_{i} = z(x_{i})\\in \\mathbb{R}^{d_{z}} \\) for \\( i = 1,\\dots ,n_{h_f} \\). The kernel integral operator can be approximated by the Monte-Carlo method (omitting the layer index):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.291, + 0.895, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} (\\mathcal {K} (z)) (x) = \\int_ {\\Omega} K (z (x), z (y)) z (y) \\mathrm {d} y \\\\ n _ {h _ {f}} \\end{array} \\tag {9} \\\\ \\approx \\sum_ {i = 1} ^ {\\dots h _ {f}} K (z (x), z _ {i}) z _ {i}, \\quad \\forall x \\in \\Omega_ {h _ {f}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.369, + 0.545, + 0.381 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.382, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\nK (z (x), z _ {i}) = \\frac {\\exp \\left(\\frac {\\langle W _ {q} z (x) , W _ {k} z _ {i} \\rangle}{\\sqrt {d _ {z}}}\\right)}{\\sum_ {j = 1} ^ {n _ {h _ {f}}} \\exp \\left(\\frac {\\langle W _ {q} z _ {j} , W _ {k} z _ {i} \\rangle}{\\sqrt {d _ {z}}}\\right)} W _ {v}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.627 + ], + "angle": 0, + "content": "In the language of transformers, the matrices \\( W_{q}, W_{k}, W_{v} \\in \\mathbb{R}^{d_{z} \\times d_{z}} \\) correspond to the queries, keys and values functions respectively. By allowing \\( K \\) to be the sum of multiple functions with separate trainable parameters, the multi-head self-attention can also be rewritten like in (9). For every observation \\( (a, u) \\), this operation has a complexity of \\( O(n_{h_{f}}^{2} d_{z}) \\), which is unaffordable for the SR problem where the sampling number \\( n_{h_{f}} \\) usually reaches to 10,000 or beyond. To overcome this issue, we employ the Galerkin-type attention operator which has a linear complexity \\( O(n_{h_{f}} d_{z}^{2}) \\). The approximation capacity of a linearized Galerkin-type attention has been proved to be comparable to a Petrov-Galerkin projection [3]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.895, + 0.799 + ], + "angle": 0, + "content": "We suppose that a image function \\( f \\) is locally integrable, i.e., \\( f \\) is measurable and \\( \\int_{\\Omega} |f| < \\infty \\) for every compact \\( \\Omega \\subset \\mathbb{R}^2 \\). The idea behind distributions [28] is to identify \\( f \\) with \\( \\int f \\phi \\) by suitably choosing a \"test function\" \\( \\phi \\), rather than by a series of separate evaluations \\( f(x) \\) at coordinates \\( x \\in \\Omega \\). Let \\( v(x) = W_v z(x) \\), \\( k(x) = W_k z(x) \\), and \\( q(x) = W_q z(x) \\) be the keys, queries and values functions respectively, which are \\( d_z \\) dimensional vector-valued functions, with a subscript denoting the corresponding component. The kernel integral attention then can be written in a component-wise form (\\( j = 1, \\dots, d_z \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.818, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\left(\\mathcal {K} (z)\\right) (x)\\right) _ {j} = \\sum_ {l = 1} ^ {d _ {z}} \\left\\langle k _ {l}, v _ {j} \\right\\rangle q _ {l} (x) \\tag {11} \\\\ \\approx \\sum_ {l = 1} ^ {d _ {z}} \\left(\\int_ {\\Omega} k _ {l} (y) v _ {j} (y) \\mathrm {d} y\\right) q _ {l} (x), \\quad \\forall x \\in \\Omega_ {h _ {f}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18250" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.188 + ], + "angle": 0, + "content": "Denote the evaluations \\((z(x_{1}),\\ldots ,z(x_{n_{h_{f}}}))^{T}\\) by \\(\\mathbf{z}\\in\\) \\(\\mathbb{R}^{n_{h_f}\\times d_z}\\). Let \\(Q = \\mathbf{z}W_{q},K = \\mathbf{z}W_{k},V = \\mathbf{z}W_{v}\\in\\) \\(\\mathbb{R}^{n_{h_f}\\times d_z}\\). The columns of \\(Q / K / V\\) contain the vector representations of the learned basis functions, spanning certain subspaces of the latent representation Hilbert spaces respectively." + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.187, + 0.469, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = Q \\left(\\widetilde {K} ^ {T} \\widetilde {V}\\right) / n _ {h _ {f}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.211, + 0.469, + 0.352 + ], + "angle": 0, + "content": "where \\(\\widetilde{K} = \\operatorname{Ln}(K), \\widetilde{V} = \\operatorname{Ln}(K)\\), with \\(\\operatorname{Ln}(\\cdot)\\) the layer normalization. The \\(j\\)-th column of \\(\\widetilde{K}^T \\widetilde{V}\\) contains the coefficients for the linear combination of the basis vectors \\(\\{q_j\\}_{j=1}^{d_z}\\) to form the output \\(\\mathbf{z}\\). Consequently, the global correlations are reflected through the components of \\(\\widetilde{K}^T \\widetilde{V}\\). As a result, SRNO can utilize all information effectively without the feature unfolding [5] or local convolutions [18], significantly suppressing the discontinuous patterns that appear around feature boundaries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.596 + ], + "angle": 0, + "content": "In addition to providing a global aggregation for the output \\(\\mathbf{z}\\) at each sampling point, the linear Galerkin-type attention (12) has the ability to obtain a quasi-optimal approximation in the current approximation space spanned with the columns of \\(Q\\) [3]. However, the expressive capability of the current bases in \\(Q\\), \\(K\\), and \\(V\\) solely depends on the current input LR image through the latent representation \\(\\mathbf{z}\\), is there any chance that we can enrich the bases which contains some extra and useful information for SR reconstruction, but not included in the input \\(\\mathbf{z}\\)? For this purpose, the point-wise FFN \\(\\mathcal{O}:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_z}\\) in Fig.2 (b) introduces nonlinearities on one hand, and the positions concatenated in \\(\\mathbf{z}\\) enhance the bases on the other. In this way, the basis functions are being constantly enriched, and we reformulate the iterative process (See supplementary for a principled discussion):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.605, + 0.469, + 0.621 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} (x) = z _ {t} (x) + \\mathcal {O} ((\\mathcal {K} _ {t} (z _ {t})) (x) + z _ {t} (x)). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Network details. The Network architecture is shown in Fig.2. As to the feature encoder \\( E_{\\psi} \\), we employ EDSR-baseline [21], or RDN [41], both of which drop their upsampling layers, and their output channel dimensions \\( d_{e} = 64 \\). The CNNs in \\( E_{\\psi} \\) assist SRNO in capturing the common basis functions from the ensemble of training samples, while the Galerkin-type attention layers provide the instance-specific basis enhancement. We employ the multi-head attention scheme in [36] by dividing the queries, keys and values into \\( n_{\\text{heads}} \\) parts with each of dimension \\( d_{z} / n_{\\text{heads}} \\). In our implementation, \\( d_{z} = 256 \\), \\( n_{\\text{heads}} = 16 \\), yielding 16-dimensional output values. We only use two iterations (\\( T = 2 \\)) in the kernel integral operator, which already outperforms previous works, while keeping the running time advantage. Note that we utilize \\( 1 \\times 1 \\) convolutions to replace all the linear layers in SRNO, since they have a GPU-friendly data structure. The detailed network structures are listed in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.632, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.605, + 0.132 + ], + "angle": 0, + "content": "4.1. Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.139, + 0.892, + 0.229 + ], + "angle": 0, + "content": "Dataset. We used the DIV2K dataset [1] for network training, while the DIV2K validation set [1] and four benchmark datasets, including Set5 [2], Set14 [39], B100 [25] and Urban100 [12], for evaluation. Peak signal-to-noise ratio (PSNR) is used as the evaluation metric. Similar to [5, 11], we cropped image boundaries when computing PSNRs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.23, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Implementation Details. Numerically, we have to explore the underlying continuous image function through access to its point-wise evaluations on the HR and LR grids. Let \\( B \\) be the batch size and \\( n_{h_c} \\) be the LR sampling counts. We first sample \\( B \\) random scales \\( r^{(i)} \\) following a uniform distribution \\( \\mathcal{U}(1,4) \\), and then crop \\( B \\) patches of sizes \\( \\{\\sqrt{n}_{h_c}r^{(i)}\\times \\sqrt{n}_{h_c}r^{(i)}\\}_{i = 1}^B \\) from the HR training images (one per each). We find \\( n_{h_c} = 128^2 \\) is sufficient for SRNO to represent most natural image patches and fix this choice in the sequel. The LR counterparts are downsampled using bicubic interpolation with the corresponding \\( r^{(i)} \\). In order to keep a consistent number of supervision points for the LR patches of different scales in a single batch, we randomly sample \\( 128^2 \\) HR pixels and calculate the corresponding fractional coordinates on the coarse grid associated with \\( r^{(i)} \\), as is done in [5]. We use an L1 loss [21] and the Adam [16] optimizer with an initial learning rate \\( 4\\times 10^{-5} \\) and the maximum \\( 4\\times 10^{-4} \\). All models are trained for 1000 epochs with batch size 64, and the learning rate decays by the cosine annealing after a warm-up phase of 50 epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.621, + 0.559 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Quantitative result. On the DIV2K validation set, Table 1 presents a quantitative comparison among our SRNO and three existing arbitrary-scale SR methods, MetaSR [5, 11], LIIF [5], and LTE [18]. Results when EDSR-baseline [21], and RDN [41] are used as encoders are displayed in the top and bottom rows, respectively. One can observe that SRNO provides the best results over all scale factors, irrespective of the encoder employed. Additionally, on in-distribution scales \\((\\times 2, \\times 3, \\times 4)\\), our method outperforms earlier works by wide margins." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.885 + ], + "angle": 0, + "content": "As RDN facilitates better reconstruction accuracy than EDSR-baseline in Tab.1, we choose RDN as the encoder in the comparisons over the four benchmark datasets. The results are listed in Tab.2, where SRNO obtains all the best performances. Note that only SRNO produces substantial improvements over RDN on every in-distribution scale, although the latter trains separate models for each scale. As the experimental settings in Tab.1 and 2 only differ in the decoder part, we conclude that the Galerkin-type attention mechanism employed in SRNO does contribute to better function approximation capability." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative result. Figure 3 illustrates the visual results" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18251" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.088, + 0.269, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.281, + 0.214, + 0.295 + ], + "angle": 0, + "content": "LR Image" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.089, + 0.392, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.281, + 0.358, + 0.294 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.089, + 0.515, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.281, + 0.482, + 0.295 + ], + "angle": 0, + "content": "LIIF [5]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.638, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.281, + 0.607, + 0.295 + ], + "angle": 0, + "content": "LTE [18]" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.089, + 0.759, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.281, + 0.743, + 0.295 + ], + "angle": 0, + "content": "SRNO (ours)" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.089, + 0.882, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.281, + 0.833, + 0.294 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Visual comparison on other zero-shot SR. The boxes in the first column indicate the areas that the close-ups on the right display. All methods are trained with continuous random scales in \\(\\times 1 - \\times 4\\) and tested on \\(\\times 8 / \\times 12\\) to evaluate the generalization capability to unknown scaling factors. RDN is used as the encoder for all methods." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.347, + 0.263, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.445, + 0.262, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.542, + 0.204, + 0.555 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.348, + 0.416, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.445, + 0.416, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.542, + 0.368, + 0.556 + ], + "angle": 0, + "content": "LIIF [5]" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.348, + 0.572, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.445, + 0.571, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.542, + 0.525, + 0.556 + ], + "angle": 0, + "content": "LTE [18]" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.348, + 0.726, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.445, + 0.725, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.542, + 0.694, + 0.556 + ], + "angle": 0, + "content": "SRNO (ours)" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.348, + 0.882, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.445, + 0.881, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.542, + 0.816, + 0.555 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.561, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Figure 4. Visual comparison on non-integer scales. All methods use RDN as the encoder and are trained with continuous random scales in \\(\\times 1 - \\times 4\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.803 + ], + "angle": 0, + "content": "obtained from two images in the Urban100 and DIV2K validation datasets. Both the regions marked with a rectangle are rich in high-frequency structures. The bicubic interpolations exhibit severe aliasing artifacts, which obviously have been transferred into the latent codes of LIIF [5] as it directly relies on the interpolation of the feature maps and the bilinear ensemble. Thanks to the local Fourier frequency estimation step, LTE [18] performs somewhat better in partial regions (e.g. the top-right corner of the building in the second row), but aliasing still dominates on the whole. In contrast, SRNO successfully recovers the fine structures, which again verifies the importance of the global integral kernels in capturing the correct overall structures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Figure 4 shows two examples of scene text images on non-integer scales. In the first row, only SRNO can clearly recover the word \"ANTIQUES\". And in the second row, only SRNO can consistently restore the two e's. Obviously, the Galerkin-type attention attributes to this global consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.655, + 0.623 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Data sampling. The quantity of sampling points is crucial in our super-resolution operator learning architecture. Figure 5 shows that almost all scales perform better when using additional pixel samples. However, the benefit becomes less as the numbers of samples increases. \\(96^{2}\\) or \\(128^{2}\\) points would be good trade-offs between training time and performance. Figure 6 shows that more sampling points can more accurately represent an RGB-based function, which facilitates to learn the map between two approximation spaces. Using the DIV2K (800 images) and Flickr2K (2650 images) datasets, we also train an EDSR-baseline-MLP with position embedding, denoted by \\(\\mathbf{MLP}^{\\varphi}\\). We can see that our SRNO \\((64^{2}/96^{2}/128^{2})\\) obtains better performance for scales \\(\\times 3\\), \\(\\times 4\\), \\(\\times 6\\), \\(\\times 8\\), trained with only 800 images in DIV2K. The curve for \\(\\mathrm{MLP}(128^{2})\\) in Fig.5 reflects the performance advantage of SRNO \\((128^{2})\\), both trained with DIV2K dataset. Random and sequential sampling results are presented in Supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18252" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.202, + 0.088, + 0.773, + 0.244 + ], + "angle": 0, + "content": "
MethodIn-distributionOut-of-distribution
×2×3×4×6×12×18×24×30
Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [21]34.5530.9028.94-----
EDSR-baseline-MetaSR [5,11]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-LTE [18]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-SRNO (ours)34.8531.1129.1626.9023.8422.2921.2720.56
RDN-MetaSR [5,11]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-LTE [18]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-SRNO (ours)35.1631.4229.4227.1224.0322.4621.4120.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.246, + 0.895, + 0.289 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on DIV2K validation set (PSNR (dB)). The best performance are bolded. EDSR-baseline trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in \\(\\times 1 - \\times 4\\)." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.293, + 0.821, + 0.475 + ], + "angle": 0, + "content": "
MethodIn-distributionOut-of-distributionIn-distributionOut-of-distribution
×2×3×4×6×8×2×3×4×6×8
Set5Set14
RDN [41]38.2434.7132.47--34.0130.5728.81--
RDN-MetaSR [5, 11]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.97
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.15
RDN-LTE [18]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.16
RDN-SRNO (ours)38.3234.8432.6929.3827.2834.2730.7128.9726.7625.26
B100Urban100
RDN [41]32.3429.2627.72--32.8928.8026.61--
RDN-MetaSR [5, 11]32.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LIIF [5]32.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-LTE [18]32.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-SRNO (ours)32.4329.3727.8326.0424.9933.3329.1426.9824.4323.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.476, + 0.893, + 0.518 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on benchmark datasets (PSNR (dB)). The best performances are in bold. RDN trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in \\( \\times 1 - \\times 4 \\)" + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.533, + 0.454, + 0.624 + ], + "angle": 0, + "content": "
Method#Params. (M)#FLOPs (G)
EDSR-baseline-LIIF1.685.0
EDSR-baseline-LTE1.775.3
EDSR-baseline-SRNO (ours)2.065.8
RDN-LIIF22.4765.2
RDN-LTE22.5755.5
RDN-SRNO (ours)22.8746.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.626, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Table 3. Comparison of model parameters and FLOPs. Metrics are measured under the setting of \\(128^{2}\\) sample points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Other design choices. In Tab.4, we retrain the following models using EDSR-baseline [21]. All the tests were conducted using a single NVIDIA RTX 3090. In comparing our interpolation-free method to SRNO(-i) using nearest-neighbor interpolation, we find that it works better and faster. It should be noted that we only generate the image once, as opposed to employing an MLP four times as in LIIF [5] and LTE [18]. Figure 7 shows that SRNO, compared to SRNO(-i), alleviates the blocky artifact, and compared to the local ensemble [18], overcomes the oversmoothing problem. By comparing SRNO with SRNO(-h), it can be shown that cutting the number of heads from 16 to 8 causes an increase in time costs because GPU's parallelism is reduced. We see that reducing the number of basis functions and the iterative updating layers results in a" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.537, + 0.867, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.696, + 0.892, + 0.754 + ], + "angle": 0, + "content": "Figure 5. Data samples vs. PSNRs. Evaluated on the Urban100. We train models using \\(48^{2} / 64^{2} / 96^{2} / 128^{2}\\) pixel samples and calculate the PSNR increment relative to LTE. \\(\\varphi\\) indicates training on DIV2K [1] (800 images) plus Flickr2K [33] (2650 images) dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.772, + 0.893, + 0.835 + ], + "angle": 0, + "content": "considerable performance drop by comparing SRNO with SRNO(-w) and SRNO(-l). The number of basis functions \\((d_z)\\) and the number of subspaces \\((n_{\\text{heads}})\\) play an important role in the approximation power of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.652, + 0.863 + ], + "angle": 0, + "content": "4.4. Dynamic Basis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We investigate the dynamic basis updating phenomenon across the approximation subspaces in this section. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18253" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.218 + ], + "angle": 0, + "content": "
MethodIn-dis.Out-of-dis.Time (ms)
×2×3×4×6×8
SRNO33.8330.5028.7926.5525.05149
SRNO (-i)33.8130.4728.7626.5525.00156
SRNO (-h)33.8730.4728.7626.5625.02153
SRNO (-w)33.8030.4528.7526.5325.00124
SRNO (-l)33.8230.4528.7526.5525.00117
LIIF [5]33.6830.3628.6426.4624.94183
LTE [18]33.7230.3728.6526.5024.99205
LTE+ [18]33.7130.4128.6726.4924.98160
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.219, + 0.472, + 0.318 + ], + "angle": 0, + "content": "Table 4. Quantitative ablation study on design choices of SRNO. Evaluated on the Set14 (PSNR(dB)). -h/w/l refers to \\( n_{\\text{heads}} = 8 \\), \\( d_z = 128 \\), and \\( T = 1 \\) correspondingly. -i refers to using nearest-neighborhood interpolation and position embedding. EDSR-baseline is used in lifting. The running time is averaged over Set14 on \\( \\times 2/3/4/6/8 \\) SR. For each scale SR, we run every model three times and choose the minimum running time." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.326, + 0.21, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.384, + 0.177, + 0.396 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.326, + 0.335, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.383, + 0.316, + 0.397 + ], + "angle": 0, + "content": "SRNO \\((48^{2})\\)" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.326, + 0.461, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.384, + 0.441, + 0.397 + ], + "angle": 0, + "content": "SRNO \\((64^{2})\\)" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.398, + 0.21, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.455, + 0.192, + 0.47 + ], + "angle": 0, + "content": "SRNO \\((96^{2})\\)" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.398, + 0.335, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.455, + 0.32, + 0.47 + ], + "angle": 0, + "content": "SRNO \\((128^{2})\\)" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.398, + 0.461, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.455, + 0.411, + 0.468 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Figure 6. Effects of sampling global image structures. Test on \\(\\times 8\\) scale. EDSR-baseline is used as the encoder, while the data in parentheses indicate the pixel numbers sampled in the HR images during training." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.54, + 0.274, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.604, + 0.199, + 0.616 + ], + "angle": 0, + "content": "LTE" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.54, + 0.455, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.604, + 0.4, + 0.617 + ], + "angle": 0, + "content": "SRNO (-i)" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.618, + 0.274, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.683, + 0.207, + 0.695 + ], + "angle": 0, + "content": "SRNO" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.618, + 0.455, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.683, + 0.377, + 0.695 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.7, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Figure 7. Interpolation vs. interpolation-free of SRNO. Test on \\(\\times 8\\) scale. EDSR-baseline is used as an encoder. -i refers to using nearest-neighborhood interpolation and position embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "show two exemplar basis functions in the two consecutive Galerkin attention layers in Fig.8. The basis function in the second layer appears to be more structured than in the previous layer. The dynamic update of the approximation subspaces spanned by the column vectors of \\( Q / K / V \\) comes from two factors, the concatenated random coordinates and the nonlinearity introduced by the FNN. Without them, the Galerkin-type attention would just be a linear combination of the bases in the current approximation subspaces, and" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.089, + 0.885, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.238, + 0.894, + 0.28 + ], + "angle": 0, + "content": "Figure 8. Dynamic basis update. The annotation \\( t = k \\) refers to the iterative layer number. The pictures display one of the learned basis functions in \\( V \\) and the latent representations \\( \\mathbf{z} \\)" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.293, + 0.871, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.428, + 0.886, + 0.442 + ], + "angle": 0, + "content": "Figure 9. Ranks of the matrix \\(\\mathbf{z}\\). Evaluated on the Set5 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.585 + ], + "angle": 0, + "content": "SRNO would not be able to take any chances to enrich the basis for every individual image through optimizations. The increasing ranks of the latent representation matrix \\(\\mathbf{z}\\), shown in Fig.9, provide the evidence that Galerkin-type attention is becoming more complex than a straightforward linear combination of the existing bases. Due to the space limitation, we have provided a principled discussion on the dynamic basis update processes in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed the Super-Resolution Neural Operator (SRNO) for continuous super-resolution. In SRNO, each image is seen as a function, and our method learns a map between finite-dimensional function spaces, which allows SRNO can be trained and generalize on different levels of discretization. First, in the Lifting, we use CNN-based encoders to capture feature maps from LR images and design a simple but efficient interpolation-free method, addressing the discretization inconsistency problem encountered in SR. Second, to approximate our target function in the Iterative Kernel Integration, we employ a linear attention operator that has been proved to be comparable to the Petrov-Galerkin projection. Finally, we map the last hidden representation to the output function. Experimental results show that our SRNO outperforms other arbitrary-scale SR methods in performance as well as computation time, and particularly in the capability of capturing the global image structures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18254" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.283 + ], + "angle": 0, + "content": "[3] Shuhao Cao. Choose a transformer: Fourier or galerkin. Advances in Neural Information Processing Systems, 34:24924-24940, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[7] Alexandre Ern and Jean-Luc Guermond. Theory and practice of finite elements, volume 159. Springer, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[9] John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Efficient token mixing for transformers via adaptive fourier neural operators. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[10] Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations. Advances in Neural Information Processing Systems, 34:24048-24062, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[12] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Rakhoon Hwang, Jae Yong Lee, Jin Young Shin, and Hyung Ju Hwang. Solving pde-constrained control problems using operator learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4):4504-4512, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? In International Conference on Learning Representations, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Chiyu\" Max\" Jiang, Soheil Esmaeilzadeh, Kamyar Azizzadenesheli, Karthik Kashinath, Mustafa Mustafa, Hamdi A Tchelepi, Philip Marcus, and Anima Anandkumar. Meshfreeflownet: a physics-constrained deep continuous spacetime super-resolution framework. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-15, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[17] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces. arXiv preprint arXiv:2108.08481, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.346, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[18] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[19] Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, Anima Anandkumar, et al. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833-1844, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[23] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218-229, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[24] Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister. Dynamic high-pass filtering and multi-spectral attention for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4288-4297, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.843, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18255" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.469, + 0.193 + ], + "angle": 0, + "content": "[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.195, + 0.469, + 0.278 + ], + "angle": 0, + "content": "[27] Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[28] Walter Rudin. Functional analysis, mcrawhill. Inc, New York, 45:46, 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[30] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[31] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.595 + ], + "angle": 0, + "content": "[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[34] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[35] V Vapnik. Statistical learning theory wiley-interscience. New York, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[39] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[40] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[41] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 1, 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18256" + } + ] +] \ No newline at end of file diff --git a/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_origin.pdf b/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c9dcb15fd62bd9a312a45fcac678ef9679acb16 --- /dev/null +++ b/2023/Super-Resolution Neural Operator/1174fe30-fb20-45e0-93e7-549281ad4be5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eefa07041db31beeacc2f10e8186f2aad64cf47593c6c87380f6253c891c3609 +size 4885903 diff --git a/2023/Super-Resolution Neural Operator/full.md b/2023/Super-Resolution Neural Operator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9d366b590859ead614f28a7b10c7529f99b94e6e --- /dev/null +++ b/2023/Super-Resolution Neural Operator/full.md @@ -0,0 +1,333 @@ +# Super-Resolution Neural Operator + +Min Wei* Xuesong Zhang*† + +Beijing University of Posts and Telecommunications, China + +{mw，xuesong_zhang}@bupt.edu.cn + +# Abstract + +We propose Super-resolution Neural Operator (SRNO), a deep operator learning framework that can resolve high-resolution (HR) images at arbitrary scales from the low-resolution (LR) counterparts. Treating the LR-HR image pairs as continuous functions approximated with different grid sizes, SRNO learns the mapping between the corresponding function spaces. From the perspective of approximation theory, SRNO first embeds the LR input into a higher-dimensional latent representation space, trying to capture sufficient basis functions, and then iteratively approximates the implicit image function with a kernel integral mechanism, followed by a final dimensionality reduction step to generate the RGB representation at the target coordinates. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial domain and therefore benefits the grid-free continuum; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to "hallucinate" high-frequency information from the LR image. Experiments show that SRNO outperforms existing continuous SR methods in terms of both accuracy and running time. Our code is at https://github.com/2y7c3/Super-Resolution-Neural-Operator. + +# 1. Introduction + +Single image super-resolution (SR) addresses the inverse problem of reconstructing high-resolution (HR) images from their low-resolution (LR) counterparts. In a data-driven way, deep neural networks (DNNs) learn the inversion map from many LR-HR sample pairs and have demonstrated appealing performances [4, 20, 21, 24, 29, 40, 41]. Nevertheless, most DNNs are developed in the configuration of single scaling factors, which cannot be used in sce + +![](images/c1c8a91dd468da760dfaac890b183515eb452f8ec13b0121f1bb055729e3a684.jpg) +Figure 1. Overview of Super-Resolution Neural operator (SRNO). SRNO is composed of three parts, $\mathcal{L}$ (Lifting), $\kappa$ (kernel integrals) and $\mathcal{P}$ (Projection), which perform consecutively to learn mappings between approximation spaces $\mathcal{H}(\Omega_{h_c})$ and $\mathcal{H}(\Omega_{h_f})$ associated with grid sizes $h_c$ and $h_f$ , respectively. The key component, $\kappa$ , uses test functions in the latent Hilbert space $\mathcal{V}(\Omega_{h_f})$ to seek instance-specific basis functions. + +narios requiring arbitrary SR factors [37, 38]. Recently, implicit neural functions (INF) [5, 18] have been proposed to represent images in arbitrary resolution, and paving a feasible way for continuous SR. These networks, as opposed to storing discrete signals in grid-based formats, represent signals with evaluations of continuous functions at specified coordinates, where the functions are generally parameterized by a multi-layer perceptron (MLP). To share knowledge across instances instead of fitting individual functions for each signal, encoder-based methods [5, 15, 26] are proposed to retrieve latent codes for each signal, and then a decoding MLP is shared by all the instances to generate the required output, where both the coordinates and the corresponding latent codes are taken as input. However, the point-wise behavior of MLP in the spatial dimensions results in limited performance when decoding various objects, particularly for high-frequency components [30, 32]. + +Neural operator is a newly proposed neural network architecture in the field of computational physics [17, 19, 23] for numerically efficient solvers of partial differential equations (PDE). Stemming from the operator theory, nerval operators learn mappings between infinite-dimensional function spaces, which is inherently capable of continuous function evaluations and has shown promising potentials in various applications [9, 13, 27]. Typically, neural operator consists of three components: 1) lifting, 2) iterative kernel integral, and 3) projection. The kernel integrals operate in the spatial domain, and thus can explicitly capture the global relationship constraining the underlying solution function of the PDE. The attention mechanism in transformers [36] is a special case of kernel integral where linear transforms are first exerted to the feature maps prior to the inner product operations [17]. Tremendous successes of transformers in various tasks [6, 22, 34] have shown the importance of capturing global correlations, and this is also true for SR to improve performance. [8]. + +In this paper, we propose the super-resolution neural operator (SRNO), a deep operator learning framework that can resolve HR images from their LR counterparts at arbitrary scales. As shown in Fig.1, SRNO learns the mapping between the corresponding function spaces by treating the LR-HR image pairs as continuous functions approximated with different grid sizes. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial dimensions and have been proved to be comparable to a Petrov-Galerkin projection [3]; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to "hallucinate" high-frequency information from the LR image. When employing same encoders to capture features, our method outperforms previous continuous SR methods in terms of both reconstruction accuracy and running time. + +In summary, our main contributions are as follows: + +- We propose the methodology of super-resolution neural operator that maps between finite-dimensional function spaces, allowing for continuous and zero-shot super-resolution irrespective the discretization used on the input and output spaces. +- We develop an architecture for SRNO that first explores the common latent basis for the whole training set and subsequently refines an instance-specific basis by the Galerkin-type attention mechanism. +- Numerically, we show that the proposed SRNO outperforms existing continuous SR methods with less running time, and even generates better results on the resolutions for which the fixed scale SR networks were trained. + +# 2. Related Work + +Deep learning based SR methods. [4, 20, 21, 24, 29, 40, 41] have achieved impressive performances, in multiscale scenarios one has to train and store several models for each scale factor, which is unfeasible when considering time and memory budgets. In recent years, several methods [11, 31, 37] are proposed to achieve arbitrary-scale SR with a single model, but their performances are limited when dealing with out-of-distribution scaling factors. Inspired by INF, LIIF [5] takes continuous coordinates and latent variables as inputs, and employs an MLP to achieve outstanding performances for both in-distribution and out-of-distribution factors. In contrast, LTE [18] transforms input coordinates into the Fourier domain and uses the dominant frequencies extracted from latent variables to address the spectral bias problem [30, 32]. In a nutshell, treating images as RGB-valued functions and sharing the implicit function space are the keys to the success of LIIF-like works [5, 18]. Nevertheless, a purely local decoder, like MLP, is not able to accurately approximate arbitrary images, although it is rather sensitive to the input coordinates. + +Neural Operators. Recently, a novel neural network architecture, Neural Operator (NO), was proposed for discretization invariant solutions of PDEs via infinitedimensional operator learning [10,17,19,23]. Neural operators only need to be trained once and are capable of transferring solutions between differently discretized meshes while keeping a fixed approximation error. A valuable merit of NO is that it does not require knowledge of the underlying PDE, which allows us to introduce it by the following abstract form, + +$$ +\begin{array}{l} \left(\mathbf {L} _ {a} u\right) (x) = f (x), \quad x \in D, \tag {1} \\ u (x) = 0, \quad x \in \partial D, \\ \end{array} +$$ + +where $u: D \to \mathbb{R}^{d_u}$ is the solution function residing in the Banach space $\mathcal{U}$ , and $\mathbf{L}: \mathcal{A} \to \mathbf{L}(\mathcal{U}; \mathcal{U}^*)$ is an operator-valued functional that maps the coefficient function $a \in \mathcal{A}$ of the PDE to $f \in \mathcal{U}^*$ , the dual space of $\mathcal{U}$ . As in many cases the inverse operator of $\mathbf{L}$ even does not exist, NO seeks a feasible operator $\mathcal{G}: \mathcal{A} \to \mathcal{U}, a \mapsto u$ , directly mapping the coefficient to the solution within an acceptable tolerance. + +The operator $\mathcal{G}$ is numerically approximated by training a neural network $\mathcal{G}_{\theta}:\mathcal{A}\to \mathcal{U}$ , where $\theta$ are the trainable parameters. Suppose we have $N$ pairs of observations $\{a_j,u_j\}_{j = 1}^N$ where the input functions $a_{j}$ are sampled from probability measure $\mu$ compactly supported on $\mathcal{A}$ , and $u_{j} = \mathcal{G}(a_{j})$ are used as the supervisory output functions. The infinitely dimensional operator learning problem $\mathcal{G}\gets \mathcal{G}_{\theta}$ thus is associated with the empirical-risk minimization problem [35]. In practice, we actually measure the approximation loss using the sampled observations $u_{o}^{(j)}$ + +![](images/298c6f66eac28aebf4816051fa45e1f9793d93e3f5dfe996743d1f3749fad5a1.jpg) + +![](images/67a29d95ac6ad0b143a2d521a5832267442f7108d7c536340ac503b21472a7ff.jpg) + +![](images/b66767cb366d3ef67eb8d6eca3a6290246e2b45268dc199fa8e4a55480fc482f.jpg) +Figure 2. Super-resolution neural operator (SRNO) architecture for continuous SR. The input LR image $f_{h_c}$ undergoes three phases to output the HR image $f_{h_f}$ with the specified resolution: (a) Lifting the LR pixel values $a(\mathbf{x})$ on the set of coordinates $\mathbf{x} = \{x_i\}_{i=1}^{n_{h_f}}$ to a higher dimensional feature space by a CNN-based encoder $E_{\psi}$ , constructing the latent representation $\hat{a}(x)$ , and linearly transforming into the first layer's input $z_0(\mathbf{x})$ . (b) kernel integrals composed of $T$ layers of Galerkin-type attention, and (c) finally project to the RGB space. + +![](images/4fc242f9cdb755364d257f3bf4e8010bbecc4bbe7a300a26b2fdf7ebbb4f2308.jpg) + +and $a_{o}^{(j)}$ , which are the direct results of discretization: + +$$ +\begin{array}{l} \min _ {\theta} \mathbb {E} _ {a \sim \mu} \| \mathcal {G} (a) - \mathcal {G} _ {\theta} (a) \| _ {\mathcal {U}} \\ \approx \min _ {\theta} \frac {1}{N} \sum_ {j = 1} ^ {N} \| u _ {o} ^ {(j)} - \mathcal {G} _ {\theta} \left(a _ {o} ^ {(j)}\right) \| _ {\mathcal {U}}. \tag {2} \\ \end{array} +$$ + +Similar to classical feedforward neural networks (FFNs), the NO is of an iterative architecture. For ease of exposition, suppose $\mathcal{A}$ is defined on the bounded domain $D\subset$ $\mathbb{R}^d$ , and the inputs and outputs of the intermediate layers are all vector-valued functions, with dimension $d_{z}$ . Then $\mathcal{G}_{\theta}:\mathcal{A}\to \mathcal{U}$ can be formulated as follows: + +$$ +z _ {0} (x) = \mathcal {L} (x, a (x)), \tag {3} +$$ + +$$ +z _ {t + 1} (x) = \sigma \left(W _ {t} z _ {t} (x) + \left(\mathcal {K} _ {t} \left(z _ {t}; \Phi\right)\right) (x)\right), \tag {4} +$$ + +$$ +u (x) = \mathcal {P} \left(z _ {T} (x)\right), \tag {5} +$$ + +where $\mathcal{L}:\mathbb{R}^{d_a + d}\to \mathbb{R}^{d_z}$ , and $\mathcal{P}:\mathbb{R}^{d_z}\to \mathbb{R}^{d_u}$ are the local lifting and projection functions respectively, mapping the input $a$ to its first layer hidden representation $z_{0}$ and the last layer hidden representation $z_{T}$ back to the output function $u$ . $W:\mathbb{R}^{d_z}\to \mathbb{R}^{d_z}$ is a point-wise linear transformation, and $\sigma :\mathbb{R}^{d_z}\rightarrow \mathbb{R}^{d_z}$ is the nonlinear activation function. + +Although the PDE in (1) point-wisely defines the behavior of the solution function $u(x)$ , the solution operator $\mathcal{G}$ we are seeking should exhibit the non-local property such that $\mathcal{G}(a)$ can approximate $u$ everywhere rather than locally. For this purpose, NO may employ the kernel integral operators $\mathcal{K}_t: \{z_t : D_t \to \mathbb{R}^{d_z}\} \mapsto \{z_{t+1} : D_{t+1} \to \mathbb{R}^{d_z}\}$ to maintain the continuum in the spatial domain, and one of the + +most adopted forms [19] is defined as + +$$ +\left(\mathcal {K} _ {t} \left(z _ {t}; \Phi\right)\right) (x) = \int_ {D} K _ {t} (x, y; \Phi) z _ {t} (y) \mathrm {d} y, \quad \forall x \in D, \tag {6} +$$ + +where the kernel matrix $K_{t}:\mathbb{R}^{d + d}\to \mathbb{R}^{d_{z}\times d_{z}}$ is parameterized by $\Phi$ . + +# 3. Super-resolution Neural Operator + +All operations of NO are defined in function space and the training data are just samples at coordinates $x_{i}$ irrespective of the discretization sizes. This property inspires us to employ the NO's architecture for continuous SR. Different from NO that aims at mapping between infinitedimensional function spaces, we define SRNO in the sense of mappings between two approximation spaces of finitedimensional but continuous functions. This is reasonable because imaging optics inevitably limit the highest possible frequency in the radiance field under observation. With this configuration, we can design the architecture of SRNO using the well-developed theoretical tool, the Galerkin-type method that is widely used in the filed of Finite elements [7]. + +Problem Formulation. Let $(\mathcal{H},\langle \cdot ,\cdot \rangle_{\mathcal{H}})$ denote a Hilbert space equipped with the inner-product structure, which is continuously embedded in the space of continuous functions $C^0 (\Omega)$ , with $\Omega \subset \mathbb{R}^2$ a bounded domain. An image is defined as a vector-valued function $\mathcal{H}\ni f:\Omega \to \mathbb{R}^3$ . Assume we can access the function values of $f$ at the coordinates $\{x_i\}_{i = 1}^{nh}\subset \Omega$ with the biggest discretiza + +tion size $h$ . Note that $x_{i}$ 's are not necessarily equidistant. Our goal is to learn a super-resolution neural operator between two Hilbert spaces with different resolutions: $S_{\theta}: \mathcal{H} \supset \mathcal{H}(\Omega_{h_c}) \to \mathcal{H}(\Omega_{h_f}) \subset \mathcal{H}$ , where $h_c, h_f$ denotes the coarse and the fine grid sizes, respectively. Given $N$ function pairs $\{a^{(k)}, u^{(k)}\}_{k=1}^N$ , where $a^{(k)} \in \mathcal{H}(\Omega_{h_c})$ and $u^{(k)} \in \mathcal{H}(\Omega_{h_f})$ . Our SRNO parameterized by $\theta$ can be solved through the associated empirical-risk minimization problem: + +$$ +\min _ {\theta} \frac {1}{N} \sum_ {k = 1} ^ {N} \| u _ {h _ {f}} ^ {(k)} - \mathcal {S} _ {\theta} \left(a _ {h _ {c}} ^ {(k)}\right) \| _ {\mathcal {H}}. \tag {7} +$$ + +Since our data $a^{(k)}$ and $u^{(k)}$ are functions, to work with them numerically, we assume access only to their pointwise evaluations. Let $\mathbb{A}_{h_c} \subset \mathcal{H}(\Omega_{h_c}) \subset \mathcal{H}$ be an approximation space [3] associated with $\{x_i\}_{i=1}^{n_{h_c}}$ , such that for any $a \in \mathbb{A}_{h_c}$ , $a(\cdot) = \sum_{i=1}^{n_{h_c}} a(x_i) \xi_{x_i}(\cdot)$ , where $\{\xi_{x_i}(\cdot)\}$ form a set of nodal basis for $\mathbb{A}_{h_c}$ in the sense that $\xi_{x_i}(x_j) = \delta_{ij}$ . Similarly, $\mathbb{U}_{h_f} \subset \mathcal{H}(\Omega_{h_f}) \subset \mathcal{H}$ is another approximation space associate with grid size $h_f$ . Note that Our observed LR-HR image pairs $\{a, u\}$ may have different grid size pairs $\{h_c, h_f\}$ , which means they belong to different approximation spaces $\{\mathbb{A}_{h_c}, \mathbb{U}_{h_f}\}$ . + +Lifting. The neural operators from PDEs usually use a simple pointwise function, $\mathcal{L}$ in (3), to expand the input channels. But for the SR problem, a deep feature encoder $E_{\psi}$ (with trainable parameters $\psi$ ) and spatial interpolations should be considered in the lifting operation, due to the complexity of natural images and the discretization inconsistency between LR and HR image functions. Another critical ingredient in Lifting is the incorporation scheme of coordinates. Prior study [14] demonstrates that deep CNNs can implicitly learn to encode the information about absolute positions. As shown in Fig.2 (c), the position information of the grid points $\{a_l\}_{l=1}^4$ have been implicitly encoded by $E_{\psi}$ employing CNNs. Therefore, we only need to explicitly construct the coordinate features $\hat{a}(x)$ with the fractional part of coordinate $x$ inside a grid. In order to reduce the blocky artifacts resulting from direct interpolation of the LR feature maps, we propose to concatenate the features weighted by the corresponding bilinear interpolation factors $s_l$ . Compared to the local ensemble trick in [5], our method takes shorter running time and overcomes the oversmoothing problem. We reformulate the lifting operation in (3) as: + +$$ +z _ {0} (x) = \mathcal {L} (\mathbf {c}, \left\{s _ {l} \cdot E _ {\psi} \left(a \left(\hat {x} _ {l}\right)\right), \delta_ {l} (x) \right\} _ {l = 1} ^ {4}), \tag {8} +$$ + +where $x$ are coordinates of HR image functions, $\hat{x}_l$ are the coordinates of the four neighbors of $x$ , $\delta_l(x) = x - \hat{x}_l$ , $\mathbf{c} = (2 / r_x, 2 / r_y)$ represents a $r_x \times r_y$ local area in HR images with $r_x$ and $r_y$ the scaling factors, and $\mathcal{L}: \mathbb{R}^{4 \times (d_e + 2) + 2} \to \mathbb{R}^{d_z}$ is an local linear transformation function, with $d_e$ the number of channels after $E_{\psi}$ . + +Kernel integral. The kernel integral operator, in (4), actually tells that we can identify the hidden representation $z(x)$ of the underlying image function $f(x)$ with distributions [28], inner products in the settings of SRNO. We first define the kernel $K_{t}:\mathbb{R}^{d_{z} + d_{z}}\to \mathbb{R}^{d_{z}\times d_{z}}$ with respect to the input pair $(z(x),z(y))$ , rather than like (6) depending on the spatial variables $(x,y)$ . Furthermore, in order to more efficiently explore $z(x)$ we can use multiple sets of test functions defining the distributions, which may immediately remind one of a single-head self-attention [17]. Denote $z_{i} = z(x_{i})\in \mathbb{R}^{d_{z}}$ for $i = 1,\dots ,n_{h_f}$ . The kernel integral operator can be approximated by the Monte-Carlo method (omitting the layer index): + +$$ +\begin{array}{l} \begin{array}{l} (\mathcal {K} (z)) (x) = \int_ {\Omega} K (z (x), z (y)) z (y) \mathrm {d} y \\ n _ {h _ {f}} \end{array} \tag {9} \\ \approx \sum_ {i = 1} ^ {\dots h _ {f}} K (z (x), z _ {i}) z _ {i}, \quad \forall x \in \Omega_ {h _ {f}}, \\ \end{array} +$$ + +where + +$$ +K (z (x), z _ {i}) = \frac {\exp \left(\frac {\langle W _ {q} z (x) , W _ {k} z _ {i} \rangle}{\sqrt {d _ {z}}}\right)}{\sum_ {j = 1} ^ {n _ {h _ {f}}} \exp \left(\frac {\langle W _ {q} z _ {j} , W _ {k} z _ {i} \rangle}{\sqrt {d _ {z}}}\right)} W _ {v}. \tag {10} +$$ + +In the language of transformers, the matrices $W_{q}, W_{k}, W_{v} \in \mathbb{R}^{d_{z} \times d_{z}}$ correspond to the queries, keys and values functions respectively. By allowing $K$ to be the sum of multiple functions with separate trainable parameters, the multi-head self-attention can also be rewritten like in (9). For every observation $(a, u)$ , this operation has a complexity of $O(n_{h_{f}}^{2} d_{z})$ , which is unaffordable for the SR problem where the sampling number $n_{h_{f}}$ usually reaches to 10,000 or beyond. To overcome this issue, we employ the Galerkin-type attention operator which has a linear complexity $O(n_{h_{f}} d_{z}^{2})$ . The approximation capacity of a linearized Galerkin-type attention has been proved to be comparable to a Petrov-Galerkin projection [3]. + +We suppose that a image function $f$ is locally integrable, i.e., $f$ is measurable and $\int_{\Omega} |f| < \infty$ for every compact $\Omega \subset \mathbb{R}^2$ . The idea behind distributions [28] is to identify $f$ with $\int f \phi$ by suitably choosing a "test function" $\phi$ , rather than by a series of separate evaluations $f(x)$ at coordinates $x \in \Omega$ . Let $v(x) = W_v z(x)$ , $k(x) = W_k z(x)$ , and $q(x) = W_q z(x)$ be the keys, queries and values functions respectively, which are $d_z$ dimensional vector-valued functions, with a subscript denoting the corresponding component. The kernel integral attention then can be written in a component-wise form ( $j = 1, \dots, d_z$ ): + +$$ +\begin{array}{l} \left(\left(\mathcal {K} (z)\right) (x)\right) _ {j} = \sum_ {l = 1} ^ {d _ {z}} \left\langle k _ {l}, v _ {j} \right\rangle q _ {l} (x) \tag {11} \\ \approx \sum_ {l = 1} ^ {d _ {z}} \left(\int_ {\Omega} k _ {l} (y) v _ {j} (y) \mathrm {d} y\right) q _ {l} (x), \quad \forall x \in \Omega_ {h _ {f}}. \\ \end{array} +$$ + +Denote the evaluations $(z(x_{1}),\ldots ,z(x_{n_{h_{f}}}))^{T}$ by $\mathbf{z}\in$ $\mathbb{R}^{n_{h_f}\times d_z}$ . Let $Q = \mathbf{z}W_{q},K = \mathbf{z}W_{k},V = \mathbf{z}W_{v}\in$ $\mathbb{R}^{n_{h_f}\times d_z}$ . The columns of $Q / K / V$ contain the vector representations of the learned basis functions, spanning certain subspaces of the latent representation Hilbert spaces respectively. + +$$ +\mathbf {z} = Q \left(\widetilde {K} ^ {T} \widetilde {V}\right) / n _ {h _ {f}}, \tag {12} +$$ + +where $\widetilde{K} = \operatorname{Ln}(K), \widetilde{V} = \operatorname{Ln}(K)$ , with $\operatorname{Ln}(\cdot)$ the layer normalization. The $j$ -th column of $\widetilde{K}^T \widetilde{V}$ contains the coefficients for the linear combination of the basis vectors $\{q_j\}_{j=1}^{d_z}$ to form the output $\mathbf{z}$ . Consequently, the global correlations are reflected through the components of $\widetilde{K}^T \widetilde{V}$ . As a result, SRNO can utilize all information effectively without the feature unfolding [5] or local convolutions [18], significantly suppressing the discontinuous patterns that appear around feature boundaries. + +In addition to providing a global aggregation for the output $\mathbf{z}$ at each sampling point, the linear Galerkin-type attention (12) has the ability to obtain a quasi-optimal approximation in the current approximation space spanned with the columns of $Q$ [3]. However, the expressive capability of the current bases in $Q$ , $K$ , and $V$ solely depends on the current input LR image through the latent representation $\mathbf{z}$ , is there any chance that we can enrich the bases which contains some extra and useful information for SR reconstruction, but not included in the input $\mathbf{z}$ ? For this purpose, the point-wise FFN $\mathcal{O}:\mathbb{R}^{d_z}\to \mathbb{R}^{d_z}$ in Fig.2 (b) introduces nonlinearities on one hand, and the positions concatenated in $\mathbf{z}$ enhance the bases on the other. In this way, the basis functions are being constantly enriched, and we reformulate the iterative process (See supplementary for a principled discussion): + +$$ +z _ {t + 1} (x) = z _ {t} (x) + \mathcal {O} ((\mathcal {K} _ {t} (z _ {t})) (x) + z _ {t} (x)). \tag {13} +$$ + +Network details. The Network architecture is shown in Fig.2. As to the feature encoder $E_{\psi}$ , we employ EDSR-baseline [21], or RDN [41], both of which drop their upsampling layers, and their output channel dimensions $d_{e} = 64$ . The CNNs in $E_{\psi}$ assist SRNO in capturing the common basis functions from the ensemble of training samples, while the Galerkin-type attention layers provide the instance-specific basis enhancement. We employ the multi-head attention scheme in [36] by dividing the queries, keys and values into $n_{\text{heads}}$ parts with each of dimension $d_{z} / n_{\text{heads}}$ . In our implementation, $d_{z} = 256$ , $n_{\text{heads}} = 16$ , yielding 16-dimensional output values. We only use two iterations ( $T = 2$ ) in the kernel integral operator, which already outperforms previous works, while keeping the running time advantage. Note that we utilize $1 \times 1$ convolutions to replace all the linear layers in SRNO, since they have a GPU-friendly data structure. The detailed network structures are listed in the supplementary. + +# 4. Experiments + +# 4.1. Training + +Dataset. We used the DIV2K dataset [1] for network training, while the DIV2K validation set [1] and four benchmark datasets, including Set5 [2], Set14 [39], B100 [25] and Urban100 [12], for evaluation. Peak signal-to-noise ratio (PSNR) is used as the evaluation metric. Similar to [5, 11], we cropped image boundaries when computing PSNRs. + +Implementation Details. Numerically, we have to explore the underlying continuous image function through access to its point-wise evaluations on the HR and LR grids. Let $B$ be the batch size and $n_{h_c}$ be the LR sampling counts. We first sample $B$ random scales $r^{(i)}$ following a uniform distribution $\mathcal{U}(1,4)$ , and then crop $B$ patches of sizes $\{\sqrt{n}_{h_c}r^{(i)}\times \sqrt{n}_{h_c}r^{(i)}\}_{i = 1}^B$ from the HR training images (one per each). We find $n_{h_c} = 128^2$ is sufficient for SRNO to represent most natural image patches and fix this choice in the sequel. The LR counterparts are downsampled using bicubic interpolation with the corresponding $r^{(i)}$ . In order to keep a consistent number of supervision points for the LR patches of different scales in a single batch, we randomly sample $128^2$ HR pixels and calculate the corresponding fractional coordinates on the coarse grid associated with $r^{(i)}$ , as is done in [5]. We use an L1 loss [21] and the Adam [16] optimizer with an initial learning rate $4\times 10^{-5}$ and the maximum $4\times 10^{-4}$ . All models are trained for 1000 epochs with batch size 64, and the learning rate decays by the cosine annealing after a warm-up phase of 50 epochs. + +# 4.2. Evaluation + +Quantitative result. On the DIV2K validation set, Table 1 presents a quantitative comparison among our SRNO and three existing arbitrary-scale SR methods, MetaSR [5, 11], LIIF [5], and LTE [18]. Results when EDSR-baseline [21], and RDN [41] are used as encoders are displayed in the top and bottom rows, respectively. One can observe that SRNO provides the best results over all scale factors, irrespective of the encoder employed. Additionally, on in-distribution scales $(\times 2, \times 3, \times 4)$ , our method outperforms earlier works by wide margins. + +As RDN facilitates better reconstruction accuracy than EDSR-baseline in Tab.1, we choose RDN as the encoder in the comparisons over the four benchmark datasets. The results are listed in Tab.2, where SRNO obtains all the best performances. Note that only SRNO produces substantial improvements over RDN on every in-distribution scale, although the latter trains separate models for each scale. As the experimental settings in Tab.1 and 2 only differ in the decoder part, we conclude that the Galerkin-type attention mechanism employed in SRNO does contribute to better function approximation capability. + +Qualitative result. Figure 3 illustrates the visual results + +![](images/dd0426702f33bfeae6d1f2418da3ef6b46c72dc177f57296e2092f32ce5e12f2.jpg) +LR Image + +![](images/3e91cfeaafc5249e79c6f62f954dd4454bd038cd5dca262627d56e27388ed08e.jpg) +Bicubic + +![](images/825f491808d54b660b4da944a130f3232eb296bef87516314bf4726f03f9dc13.jpg) +LIIF [5] + +![](images/be9609ed140f485ae62d5be21306490b3371a6cff09dc11df817e2b6132b4c25.jpg) +LTE [18] + +![](images/883ea36ecbec0662783d605e54c8a509b16bfb296ddfc3b4790fda4bad9b6408.jpg) +SRNO (ours) + +![](images/48526af59f86c9902ba932eaf5b32acf5ebfb6d31d017cb98bde051a1efa11d9.jpg) +GT + +![](images/34b39f7b2e18c000d0108ce4270c88e12eb1d6f4dce5ffc1f4ef1576aaa63c4a.jpg) +Figure 3. Visual comparison on other zero-shot SR. The boxes in the first column indicate the areas that the close-ups on the right display. All methods are trained with continuous random scales in $\times 1 - \times 4$ and tested on $\times 8 / \times 12$ to evaluate the generalization capability to unknown scaling factors. RDN is used as the encoder for all methods. + +![](images/000d2fa49b6761eb1a8b2b734692ebcfb7dddc4ff7fd03a879895c788f53183b.jpg) +Bicubic + +![](images/6b669bb5797577475d1bf76f1041d42035ce06d54384da873c2d2f908051515c.jpg) + +![](images/e96ae44bd28d2c30447fed97abf27444ecb99f0ddb13f9152845ff522e685758.jpg) +LIIF [5] + +![](images/560bf66cbf7ebf40df96a061f29366d74a9b25bd02987b085b2cee4fcea7a16e.jpg) + +![](images/dcdea1cca7989fa75a045bb33ee49a1669bd7d86e3d825a19333c7450cbfb226.jpg) +LTE [18] + +![](images/df8d331c306c59f4cfad06bcc5d4b097b0eb184c01a3ae5c326fa08cd60dd322.jpg) + +![](images/4be934e6402d0f6565dea6dcb0865fed227593ee66ac3e6acb7539f6dfdde7f7.jpg) +SRNO (ours) +Figure 4. Visual comparison on non-integer scales. All methods use RDN as the encoder and are trained with continuous random scales in $\times 1 - \times 4$ . + +![](images/21a95dc3d49abdc3f5abc335caa86fc6081b67bcbe16ae45978a9a601640615a.jpg) + +![](images/b8ce4b02a4aa04fab8f37bbea4290623205ff0d34d3f8869774d963c207cdbe5.jpg) +GT + +obtained from two images in the Urban100 and DIV2K validation datasets. Both the regions marked with a rectangle are rich in high-frequency structures. The bicubic interpolations exhibit severe aliasing artifacts, which obviously have been transferred into the latent codes of LIIF [5] as it directly relies on the interpolation of the feature maps and the bilinear ensemble. Thanks to the local Fourier frequency estimation step, LTE [18] performs somewhat better in partial regions (e.g. the top-right corner of the building in the second row), but aliasing still dominates on the whole. In contrast, SRNO successfully recovers the fine structures, which again verifies the importance of the global integral kernels in capturing the correct overall structures. + +Figure 4 shows two examples of scene text images on non-integer scales. In the first row, only SRNO can clearly recover the word "ANTIQUES". And in the second row, only SRNO can consistently restore the two e's. Obviously, the Galerkin-type attention attributes to this global consistency. + +# 4.3. Ablation Study + +Data sampling. The quantity of sampling points is crucial in our super-resolution operator learning architecture. Figure 5 shows that almost all scales perform better when using additional pixel samples. However, the benefit becomes less as the numbers of samples increases. $96^{2}$ or $128^{2}$ points would be good trade-offs between training time and performance. Figure 6 shows that more sampling points can more accurately represent an RGB-based function, which facilitates to learn the map between two approximation spaces. Using the DIV2K (800 images) and Flickr2K (2650 images) datasets, we also train an EDSR-baseline-MLP with position embedding, denoted by $\mathbf{MLP}^{\varphi}$ . We can see that our SRNO $(64^{2}/96^{2}/128^{2})$ obtains better performance for scales $\times 3$ , $\times 4$ , $\times 6$ , $\times 8$ , trained with only 800 images in DIV2K. The curve for $\mathrm{MLP}(128^{2})$ in Fig.5 reflects the performance advantage of SRNO $(128^{2})$ , both trained with DIV2K dataset. Random and sequential sampling results are presented in Supplementary. + +
MethodIn-distributionOut-of-distribution
×2×3×4×6×12×18×24×30
Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [21]34.5530.9028.94-----
EDSR-baseline-MetaSR [5,11]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-LTE [18]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-SRNO (ours)34.8531.1129.1626.9023.8422.2921.2720.56
RDN-MetaSR [5,11]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-LTE [18]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-SRNO (ours)35.1631.4229.4227.1224.0322.4621.4120.68
+ +Table 1. Quantitative comparison on DIV2K validation set (PSNR (dB)). The best performance are bolded. EDSR-baseline trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in $\times 1 - \times 4$ . + +
MethodIn-distributionOut-of-distributionIn-distributionOut-of-distribution
×2×3×4×6×8×2×3×4×6×8
Set5Set14
RDN [41]38.2434.7132.47--34.0130.5728.81--
RDN-MetaSR [5, 11]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.97
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.15
RDN-LTE [18]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.16
RDN-SRNO (ours)38.3234.8432.6929.3827.2834.2730.7128.9726.7625.26
B100Urban100
RDN [41]32.3429.2627.72--32.8928.8026.61--
RDN-MetaSR [5, 11]32.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LIIF [5]32.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-LTE [18]32.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-SRNO (ours)32.4329.3727.8326.0424.9933.3329.1426.9824.4323.02
+ +Table 2. Quantitative comparison on benchmark datasets (PSNR (dB)). The best performances are in bold. RDN trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in $\times 1 - \times 4$ + +
Method#Params. (M)#FLOPs (G)
EDSR-baseline-LIIF1.685.0
EDSR-baseline-LTE1.775.3
EDSR-baseline-SRNO (ours)2.065.8
RDN-LIIF22.4765.2
RDN-LTE22.5755.5
RDN-SRNO (ours)22.8746.0
+ +Table 3. Comparison of model parameters and FLOPs. Metrics are measured under the setting of $128^{2}$ sample points. + +Other design choices. In Tab.4, we retrain the following models using EDSR-baseline [21]. All the tests were conducted using a single NVIDIA RTX 3090. In comparing our interpolation-free method to SRNO(-i) using nearest-neighbor interpolation, we find that it works better and faster. It should be noted that we only generate the image once, as opposed to employing an MLP four times as in LIIF [5] and LTE [18]. Figure 7 shows that SRNO, compared to SRNO(-i), alleviates the blocky artifact, and compared to the local ensemble [18], overcomes the oversmoothing problem. By comparing SRNO with SRNO(-h), it can be shown that cutting the number of heads from 16 to 8 causes an increase in time costs because GPU's parallelism is reduced. We see that reducing the number of basis functions and the iterative updating layers results in a + +![](images/47fcb728c79952a60a3eb99c9bfc229849dc434e0bd30dcd79e6ffd3dad3fad0.jpg) +Figure 5. Data samples vs. PSNRs. Evaluated on the Urban100. We train models using $48^{2} / 64^{2} / 96^{2} / 128^{2}$ pixel samples and calculate the PSNR increment relative to LTE. $\varphi$ indicates training on DIV2K [1] (800 images) plus Flickr2K [33] (2650 images) dataset. + +considerable performance drop by comparing SRNO with SRNO(-w) and SRNO(-l). The number of basis functions $(d_z)$ and the number of subspaces $(n_{\text{heads}})$ play an important role in the approximation power of our model. + +# 4.4. Dynamic Basis + +We investigate the dynamic basis updating phenomenon across the approximation subspaces in this section. We + +
MethodIn-dis.Out-of-dis.Time (ms)
×2×3×4×6×8
SRNO33.8330.5028.7926.5525.05149
SRNO (-i)33.8130.4728.7626.5525.00156
SRNO (-h)33.8730.4728.7626.5625.02153
SRNO (-w)33.8030.4528.7526.5325.00124
SRNO (-l)33.8230.4528.7526.5525.00117
LIIF [5]33.6830.3628.6426.4624.94183
LTE [18]33.7230.3728.6526.5024.99205
LTE+ [18]33.7130.4128.6726.4924.98160
+ +Table 4. Quantitative ablation study on design choices of SRNO. Evaluated on the Set14 (PSNR(dB)). -h/w/l refers to $n_{\text{heads}} = 8$ , $d_z = 128$ , and $T = 1$ correspondingly. -i refers to using nearest-neighborhood interpolation and position embedding. EDSR-baseline is used in lifting. The running time is averaged over Set14 on $\times 2/3/4/6/8$ SR. For each scale SR, we run every model three times and choose the minimum running time. + +![](images/5e7a8a0b139c49ab63c9b550cc267a7c8c55951b849abfef372032801904cdf0.jpg) +Bicubic + +![](images/a10d1b959a59eef5e1fdc24f3be3f27577b4764b4ee8fbd275603de62e49ee0d.jpg) +SRNO $(48^{2})$ + +![](images/9a4fe9cf1714ab1205905d12439dc3714d161207350b20afeaf60e99192ded3d.jpg) + +![](images/54b7c0d48005834ecae47f16e58e6872d750433dd29996d10a5bb52279ab6999.jpg) +SRNO $(96^{2})$ + +![](images/3327236be7feb94ddbe9f32b927f0bca7acc719c6a1843be3e97ee0ac1302a24.jpg) +SRNO $(128^{2})$ + +![](images/e2e62743bd075ebccf3a343d5ea855c14df138552e87fe2bc1323f6e2f51e101.jpg) +SRNO $(64^{2})$ +GT + +![](images/db291cadddc205beffb22995687137e328b777d3069979df3562cff2578a470c.jpg) +Figure 6. Effects of sampling global image structures. Test on $\times 8$ scale. EDSR-baseline is used as the encoder, while the data in parentheses indicate the pixel numbers sampled in the HR images during training. + +![](images/b22024d9b9f4740c4e9ea6b7ccbf9e902f562134ac17ee535a281bf9542fbff0.jpg) + +![](images/37dacdff4b4147a94590800ff0e304ae9e78377f8f456c4889da294bae421134.jpg) +LTE +SRNO +Figure 7. Interpolation vs. interpolation-free of SRNO. Test on $\times 8$ scale. EDSR-baseline is used as an encoder. -i refers to using nearest-neighborhood interpolation and position embedding. + +![](images/f20ac7ba617af700e2c7d0f2e3f247d1cd2af586f71af502fee684edcaf096ba.jpg) +SRNO (-i) +GT + +show two exemplar basis functions in the two consecutive Galerkin attention layers in Fig.8. The basis function in the second layer appears to be more structured than in the previous layer. The dynamic update of the approximation subspaces spanned by the column vectors of $Q / K / V$ comes from two factors, the concatenated random coordinates and the nonlinearity introduced by the FNN. Without them, the Galerkin-type attention would just be a linear combination of the bases in the current approximation subspaces, and + +![](images/29aeeff4b6814001c6ea44ea010f290411846abb9a254ea79ed3cb4f151d48f8.jpg) +Figure 8. Dynamic basis update. The annotation $t = k$ refers to the iterative layer number. The pictures display one of the learned basis functions in $V$ and the latent representations $\mathbf{z}$ + +![](images/0f242339dc087f80730c3780d06708d78e9dca9de35a1e1be5105fc0cfa4a03c.jpg) +Figure 9. Ranks of the matrix $\mathbf{z}$ . Evaluated on the Set5 dataset. + +SRNO would not be able to take any chances to enrich the basis for every individual image through optimizations. The increasing ranks of the latent representation matrix $\mathbf{z}$ , shown in Fig.9, provide the evidence that Galerkin-type attention is becoming more complex than a straightforward linear combination of the existing bases. Due to the space limitation, we have provided a principled discussion on the dynamic basis update processes in the supplementary material. + +# 5. Conclusion + +In this paper, we proposed the Super-Resolution Neural Operator (SRNO) for continuous super-resolution. In SRNO, each image is seen as a function, and our method learns a map between finite-dimensional function spaces, which allows SRNO can be trained and generalize on different levels of discretization. First, in the Lifting, we use CNN-based encoders to capture feature maps from LR images and design a simple but efficient interpolation-free method, addressing the discretization inconsistency problem encountered in SR. Second, to approximate our target function in the Iterative Kernel Integration, we employ a linear attention operator that has been proved to be comparable to the Petrov-Galerkin projection. Finally, we map the last hidden representation to the output function. Experimental results show that our SRNO outperforms other arbitrary-scale SR methods in performance as well as computation time, and particularly in the capability of capturing the global image structures. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 7 +[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 5 +[3] Shuhao Cao. Choose a transformer: Fourier or galerkin. Advances in Neural Information Processing Systems, 34:24924-24940, 2021. 2, 4, 5 +[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 2 +[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7, 8 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2 +[7] Alexandre Ern and Jean-Luc Guermond. Theory and practice of finite elements, volume 159. Springer, 2004. 3 +[8] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 2 +[9] John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Efficient token mixing for transformers via adaptive fourier neural operators. In International Conference on Learning Representations, 2021. 2 +[10] Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations. Advances in Neural Information Processing Systems, 34:24048-24062, 2021. 2 +[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2, 5, 7 +[12] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5 +[13] Rakhoon Hwang, Jae Yong Lee, Jin Young Shin, and Hyung Ju Hwang. Solving pde-constrained control problems using operator learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4):4504-4512, 2022. 2 + +[14] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? In International Conference on Learning Representations, 2019. 4 +[15] Chiyu" Max" Jiang, Soheil Esmaeilzadeh, Kamyar Azizzadenesheli, Karthik Kashinath, Mustafa Mustafa, Hamdi A Tchelepi, Philip Marcus, and Anima Anandkumar. Meshfreeflownet: a physics-constrained deep continuous spacetime super-resolution framework. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-15, 2020. 1 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5 +[17] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces. arXiv preprint arXiv:2108.08481, 2021. 2, 4 +[18] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 5, 6, 7, 8 +[19] Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, Anima Anandkumar, et al. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2020. 2, 3 +[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833-1844, 2021. 1, 2 +[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 7 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[23] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218-229, 2021. 2 +[24] Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister. Dynamic high-pass filtering and multi-spectral attention for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4288-4297, 2021. 1, 2 +[25] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE + +International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 5 +[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1 +[27] Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214, 2022. 2 +[28] Walter Rudin. Functional analysis, mcrawhill. Inc, New York, 45:46, 1991. 4 +[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 1, 2 +[30] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 1, 2 +[31] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 2 +[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2 +[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 7 +[34] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2 +[35] V Vapnik. Statistical learning theory wiley-interscience. New York, 1998. 2 +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5 +[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 1, 2 + +[38] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 1 +[39] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5 +[40] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2 +[41] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 1, 2, 5, 7 \ No newline at end of file diff --git a/2023/Super-Resolution Neural Operator/images.zip b/2023/Super-Resolution Neural Operator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31ec02168b8c66cd7293fe03a93751cd62f644f3 --- /dev/null +++ b/2023/Super-Resolution Neural Operator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c00826552972150625a0c4b2f782e8fec844864ebffa80ab80520b4b6194ccde +size 781358 diff --git a/2023/Super-Resolution Neural Operator/layout.json b/2023/Super-Resolution Neural Operator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4cb06bc55869d78748b8cac8e131b1264d797d09 --- /dev/null +++ b/2023/Super-Resolution Neural Operator/layout.json @@ -0,0 +1,11525 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 188, + 103, + 405, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 103, + 405, + 121 + ], + "spans": [ + { + "bbox": [ + 188, + 103, + 405, + 121 + ], + "type": "text", + "content": "Super-Resolution Neural Operator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 221, + 143, + 375, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 143, + 375, + 158 + ], + "spans": [ + { + "bbox": [ + 221, + 143, + 375, + 158 + ], + "type": "text", + "content": "Min Wei* Xuesong Zhang*†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 158, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 158, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 151, + 158, + 441, + 171 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 211, + 174, + 379, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 174, + 379, + 185 + ], + "spans": [ + { + "bbox": [ + 211, + 174, + 379, + 185 + ], + "type": "text", + "content": "{mw，xuesong_zhang}@bupt.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 539 + ], + "type": "text", + "content": "We propose Super-resolution Neural Operator (SRNO), a deep operator learning framework that can resolve high-resolution (HR) images at arbitrary scales from the low-resolution (LR) counterparts. Treating the LR-HR image pairs as continuous functions approximated with different grid sizes, SRNO learns the mapping between the corresponding function spaces. From the perspective of approximation theory, SRNO first embeds the LR input into a higher-dimensional latent representation space, trying to capture sufficient basis functions, and then iteratively approximates the implicit image function with a kernel integral mechanism, followed by a final dimensionality reduction step to generate the RGB representation at the target coordinates. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial domain and therefore benefits the grid-free continuum; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to \"hallucinate\" high-frequency information from the LR image. Experiments show that SRNO outperforms existing continuous SR methods in terms of both accuracy and running time. Our code is at https://github.com/2y7c3/Super-Resolution-Neural-Operator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 128, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 676 + ], + "type": "text", + "content": "Single image super-resolution (SR) addresses the inverse problem of reconstructing high-resolution (HR) images from their low-resolution (LR) counterparts. In a data-driven way, deep neural networks (DNNs) learn the inversion map from many LR-HR sample pairs and have demonstrated appealing performances [4, 20, 21, 24, 29, 40, 41]. Nevertheless, most DNNs are developed in the configuration of single scaling factors, which cannot be used in sce" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 215, + 538, + 403 + ], + "blocks": [ + { + "bbox": [ + 317, + 215, + 538, + 403 + ], + "lines": [ + { + "bbox": [ + 317, + 215, + 538, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 215, + 538, + 403 + ], + "type": "image", + "image_path": "c1c8a91dd468da760dfaac890b183515eb452f8ec13b0121f1bb055729e3a684.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "lines": [ + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": "Figure 1. Overview of Super-Resolution Neural operator (SRNO). SRNO is composed of three parts, " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " (Lifting), " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " (kernel integrals) and " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " (Projection), which perform consecutively to learn mappings between approximation spaces " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\Omega_{h_c})" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\Omega_{h_f})" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " associated with grid sizes " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": ", respectively. The key component, " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": ", uses test functions in the latent Hilbert space " + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\Omega_{h_f})" + }, + { + "bbox": [ + 305, + 408, + 547, + 486 + ], + "type": "text", + "content": " to seek instance-specific basis functions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "narios requiring arbitrary SR factors [37, 38]. Recently, implicit neural functions (INF) [5, 18] have been proposed to represent images in arbitrary resolution, and paving a feasible way for continuous SR. These networks, as opposed to storing discrete signals in grid-based formats, represent signals with evaluations of continuous functions at specified coordinates, where the functions are generally parameterized by a multi-layer perceptron (MLP). To share knowledge across instances instead of fitting individual functions for each signal, encoder-based methods [5, 15, 26] are proposed to retrieve latent codes for each signal, and then a decoding MLP is shared by all the instances to generate the required output, where both the coordinates and the corresponding latent codes are taken as input. However, the point-wise behavior of MLP in the spatial dimensions results in limited performance when decoding various objects, particularly for high-frequency components [30, 32]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "text", + "content": "*Equal contributions. This work is supported by the National Natural Science Foundation of China (61871055). \n†Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18247" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "Neural operator is a newly proposed neural network architecture in the field of computational physics [17, 19, 23] for numerically efficient solvers of partial differential equations (PDE). Stemming from the operator theory, nerval operators learn mappings between infinite-dimensional function spaces, which is inherently capable of continuous function evaluations and has shown promising potentials in various applications [9, 13, 27]. Typically, neural operator consists of three components: 1) lifting, 2) iterative kernel integral, and 3) projection. The kernel integrals operate in the spatial domain, and thus can explicitly capture the global relationship constraining the underlying solution function of the PDE. The attention mechanism in transformers [36] is a special case of kernel integral where linear transforms are first exerted to the feature maps prior to the inner product operations [17]. Tremendous successes of transformers in various tasks [6, 22, 34] have shown the importance of capturing global correlations, and this is also true for SR to improve performance. [8]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 515 + ], + "type": "text", + "content": "In this paper, we propose the super-resolution neural operator (SRNO), a deep operator learning framework that can resolve HR images from their LR counterparts at arbitrary scales. As shown in Fig.1, SRNO learns the mapping between the corresponding function spaces by treating the LR-HR image pairs as continuous functions approximated with different grid sizes. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efficiently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial dimensions and have been proved to be comparable to a Petrov-Galerkin projection [3]; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to \"hallucinate\" high-frequency information from the LR image. When employing same encoders to capture features, our method outperforms previous continuous SR methods in terms of both reconstruction accuracy and running time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 515, + 266, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 515, + 266, + 526 + ], + "spans": [ + { + "bbox": [ + 59, + 515, + 266, + 526 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 533, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 533, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 533, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 59, + 533, + 287, + 593 + ], + "type": "text", + "content": "- We propose the methodology of super-resolution neural operator that maps between finite-dimensional function spaces, allowing for continuous and zero-shot super-resolution irrespective the discretization used on the input and output spaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 598, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 598, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 59, + 598, + 287, + 647 + ], + "type": "text", + "content": "- We develop an architecture for SRNO that first explores the common latent basis for the whole training set and subsequently refines an instance-specific basis by the Galerkin-type attention mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 653, + 287, + 712 + ], + "type": "text", + "content": "- Numerically, we show that the proposed SRNO outperforms existing continuous SR methods with less running time, and even generates better results on the resolutions for which the fixed scale SR networks were trained." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 330 + ], + "type": "text", + "content": "Deep learning based SR methods. [4, 20, 21, 24, 29, 40, 41] have achieved impressive performances, in multiscale scenarios one has to train and store several models for each scale factor, which is unfeasible when considering time and memory budgets. In recent years, several methods [11, 31, 37] are proposed to achieve arbitrary-scale SR with a single model, but their performances are limited when dealing with out-of-distribution scaling factors. Inspired by INF, LIIF [5] takes continuous coordinates and latent variables as inputs, and employs an MLP to achieve outstanding performances for both in-distribution and out-of-distribution factors. In contrast, LTE [18] transforms input coordinates into the Fourier domain and uses the dominant frequencies extracted from latent variables to address the spectral bias problem [30, 32]. In a nutshell, treating images as RGB-valued functions and sharing the implicit function space are the keys to the success of LIIF-like works [5, 18]. Nevertheless, a purely local decoder, like MLP, is not able to accurately approximate arbitrary images, although it is rather sensitive to the input coordinates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 449 + ], + "type": "text", + "content": "Neural Operators. Recently, a novel neural network architecture, Neural Operator (NO), was proposed for discretization invariant solutions of PDEs via infinitedimensional operator learning [10,17,19,23]. Neural operators only need to be trained once and are capable of transferring solutions between differently discretized meshes while keeping a fixed approximation error. A valuable merit of NO is that it does not require knowledge of the underlying PDE, which allows us to introduce it by the following abstract form," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 458, + 545, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 458, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 367, + 458, + 545, + 486 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\mathbf {L} _ {a} u\\right) (x) = f (x), \\quad x \\in D, \\tag {1} \\\\ u (x) = 0, \\quad x \\in \\partial D, \\\\ \\end{array}", + "image_path": "e9defc2d4a7437b7b53297c19548c054f7114f29cd072a5247fd505d921b21af.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "u: D \\to \\mathbb{R}^{d_u}" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": " is the solution function residing in the Banach space " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{L}: \\mathcal{A} \\to \\mathbf{L}(\\mathcal{U}; \\mathcal{U}^*)" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": " is an operator-valued functional that maps the coefficient function " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{A}" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": " of the PDE to " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "f \\in \\mathcal{U}^*" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": ", the dual space of " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": ". As in many cases the inverse operator of " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": " even does not exist, NO seeks a feasible operator " + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{G}: \\mathcal{A} \\to \\mathcal{U}, a \\mapsto u" + }, + { + "bbox": [ + 304, + 496, + 545, + 590 + ], + "type": "text", + "content": ", directly mapping the coefficient to the solution within an acceptable tolerance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": "The operator " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " is numerically approximated by training a neural network " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}:\\mathcal{A}\\to \\mathcal{U}" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " are the trainable parameters. Suppose we have " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " pairs of observations " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{a_j,u_j\\}_{j = 1}^N" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " where the input functions " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " are sampled from probability measure " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " compactly supported on " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "u_{j} = \\mathcal{G}(a_{j})" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " are used as the supervisory output functions. The infinitely dimensional operator learning problem " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}\\gets \\mathcal{G}_{\\theta}" + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "text", + "content": " thus is associated with the empirical-risk minimization problem [35]. In practice, we actually measure the approximation loss using the sampled observations " + }, + { + "bbox": [ + 304, + 592, + 545, + 714 + ], + "type": "inline_equation", + "content": "u_{o}^{(j)}" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18248" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 523, + 139 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 523, + 139 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 523, + 139 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 523, + 139 + ], + "type": "image", + "image_path": "298c6f66eac28aebf4816051fa45e1f9793d93e3f5dfe996743d1f3749fad5a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 140, + 287, + 211 + ], + "blocks": [ + { + "bbox": [ + 58, + 140, + 287, + 211 + ], + "lines": [ + { + "bbox": [ + 58, + 140, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 287, + 211 + ], + "type": "image", + "image_path": "67a29d95ac6ad0b143a2d521a5832267442f7108d7c536340ac503b21472a7ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 213, + 288, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 213, + 288, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 288, + 285 + ], + "type": "image", + "image_path": "b66767cb366d3ef67eb8d6eca3a6290246e2b45268dc199fa8e4a55480fc482f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": "Figure 2. Super-resolution neural operator (SRNO) architecture for continuous SR. The input LR image " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "f_{h_c}" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": " undergoes three phases to output the HR image " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "f_{h_f}" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": " with the specified resolution: (a) Lifting the LR pixel values " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "a(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": " on the set of coordinates " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{x_i\\}_{i=1}^{n_{h_f}}" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": " to a higher dimensional feature space by a CNN-based encoder " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "E_{\\psi}" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": ", constructing the latent representation " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "\\hat{a}(x)" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": ", and linearly transforming into the first layer's input " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "z_0(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": ". (b) kernel integrals composed of " + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 288, + 546, + 335 + ], + "type": "text", + "content": " layers of Galerkin-type attention, and (c) finally project to the RGB space." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 291, + 140, + 536, + 284 + ], + "blocks": [ + { + "bbox": [ + 291, + 140, + 536, + 284 + ], + "lines": [ + { + "bbox": [ + 291, + 140, + 536, + 284 + ], + "spans": [ + { + "bbox": [ + 291, + 140, + 536, + 284 + ], + "type": "image", + "image_path": "4fc242f9cdb755364d257f3bf4e8010bbecc4bbe7a300a26b2fdf7ebbb4f2308.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 346, + 264, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 264, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 264, + 360 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 346, + 264, + 360 + ], + "type": "inline_equation", + "content": "a_{o}^{(j)}" + }, + { + "bbox": [ + 47, + 346, + 264, + 360 + ], + "type": "text", + "content": ", which are the direct results of discretization:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 366, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 366, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 97, + 366, + 287, + 420 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\mathbb {E} _ {a \\sim \\mu} \\| \\mathcal {G} (a) - \\mathcal {G} _ {\\theta} (a) \\| _ {\\mathcal {U}} \\\\ \\approx \\min _ {\\theta} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\| u _ {o} ^ {(j)} - \\mathcal {G} _ {\\theta} \\left(a _ {o} ^ {(j)}\\right) \\| _ {\\mathcal {U}}. \\tag {2} \\\\ \\end{array}", + "image_path": "c64f2e2fd71979125ba6c655b2cdb7b7f17dbc403d9459336862dc77e720042e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "text", + "content": "Similar to classical feedforward neural networks (FFNs), the NO is of an iterative architecture. For ease of exposition, suppose " + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "text", + "content": " is defined on the bounded domain " + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "inline_equation", + "content": "D\\subset" + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "text", + "content": " , and the inputs and outputs of the intermediate layers are all vector-valued functions, with dimension " + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "inline_equation", + "content": "d_{z}" + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "text", + "content": " . Then " + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}:\\mathcal{A}\\to \\mathcal{U}" + }, + { + "bbox": [ + 47, + 427, + 287, + 498 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 506, + 287, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 506, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 91, + 506, + 287, + 520 + ], + "type": "interline_equation", + "content": "z _ {0} (x) = \\mathcal {L} (x, a (x)), \\tag {3}", + "image_path": "acfca9e6bba3f931e1b5311b1a13d42c087aa4a8a244b5a17f8cc75f36c69e2f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 521, + 287, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 521, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 83, + 521, + 287, + 534 + ], + "type": "interline_equation", + "content": "z _ {t + 1} (x) = \\sigma \\left(W _ {t} z _ {t} (x) + \\left(\\mathcal {K} _ {t} \\left(z _ {t}; \\Phi\\right)\\right) (x)\\right), \\tag {4}", + "image_path": "1edc2dbc658616ddbe75f82ddb40a81191f96a4ab70f23ef333346b5a4e94e4e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 536, + 287, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 536, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 94, + 536, + 287, + 550 + ], + "type": "interline_equation", + "content": "u (x) = \\mathcal {P} \\left(z _ {T} (x)\\right), \\tag {5}", + "image_path": "03e3ba98046004d566f23fb45853e70bb5d0c8e309108c850da114d2ac6a2c35.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}:\\mathbb{R}^{d_a + d}\\to \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{P}:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_u}" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": " are the local lifting and projection functions respectively, mapping the input " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": " to its first layer hidden representation " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": " and the last layer hidden representation " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": " back to the output function " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "W:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": " is a point-wise linear transformation, and " + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\sigma :\\mathbb{R}^{d_z}\\rightarrow \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 47, + 556, + 287, + 629 + ], + "type": "text", + "content": " is the nonlinear activation function." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": "Although the PDE in (1) point-wisely defines the behavior of the solution function " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "u(x)" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": ", the solution operator " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " we are seeking should exhibit the non-local property such that " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(a)" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " can approximate " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " everywhere rather than locally. For this purpose, NO may employ the kernel integral operators " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_t: \\{z_t : D_t \\to \\mathbb{R}^{d_z}\\} \\mapsto \\{z_{t+1} : D_{t+1} \\to \\mathbb{R}^{d_z}\\}" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " to maintain the continuum in the spatial domain, and one of the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 348, + 459, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 459, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 459, + 360 + ], + "type": "text", + "content": "most adopted forms [19] is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 369, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\left(\\mathcal {K} _ {t} \\left(z _ {t}; \\Phi\\right)\\right) (x) = \\int_ {D} K _ {t} (x, y; \\Phi) z _ {t} (y) \\mathrm {d} y, \\quad \\forall x \\in D, \\tag {6}", + "image_path": "1cbb5ee2365e44c810d4726790ab8e2d5670cda1bd6a219b470791eae0a1f40e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "type": "text", + "content": "where the kernel matrix " + }, + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "type": "inline_equation", + "content": "K_{t}:\\mathbb{R}^{d + d}\\to \\mathbb{R}^{d_{z}\\times d_{z}}" + }, + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "type": "text", + "content": " is parameterized by " + }, + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 404, + 545, + 429 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 441, + 495, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 495, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 495, + 455 + ], + "type": "text", + "content": "3. Super-resolution Neural Operator" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "content": "All operations of NO are defined in function space and the training data are just samples at coordinates " + }, + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "content": " irrespective of the discretization sizes. This property inspires us to employ the NO's architecture for continuous SR. Different from NO that aims at mapping between infinitedimensional function spaces, we define SRNO in the sense of mappings between two approximation spaces of finitedimensional but continuous functions. This is reasonable because imaging optics inevitably limit the highest possible frequency in the radiance field under observation. With this configuration, we can design the architecture of SRNO using the well-developed theoretical tool, the Galerkin-type method that is widely used in the filed of Finite elements [7]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": "Problem Formulation. Let " + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{H},\\langle \\cdot ,\\cdot \\rangle_{\\mathcal{H}})" + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": " denote a Hilbert space equipped with the inner-product structure, which is continuously embedded in the space of continuous functions " + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "inline_equation", + "content": "C^0 (\\Omega)" + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": " a bounded domain. An image is defined as a vector-valued function " + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}\\ni f:\\Omega \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": ". Assume we can access the function values of " + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": " at the coordinates " + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{x_i\\}_{i = 1}^{nh}\\subset \\Omega" + }, + { + "bbox": [ + 304, + 630, + 546, + 715 + ], + "type": "text", + "content": " with the biggest discretiza" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18249" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "tion size " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "'s are not necessarily equidistant. Our goal is to learn a super-resolution neural operator between two Hilbert spaces with different resolutions: " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "S_{\\theta}: \\mathcal{H} \\supset \\mathcal{H}(\\Omega_{h_c}) \\to \\mathcal{H}(\\Omega_{h_f}) \\subset \\mathcal{H}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "h_c, h_f" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " denotes the coarse and the fine grid sizes, respectively. Given " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " function pairs " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\{a^{(k)}, u^{(k)}\\}_{k=1}^N" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "a^{(k)} \\in \\mathcal{H}(\\Omega_{h_c})" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "u^{(k)} \\in \\mathcal{H}(\\Omega_{h_f})" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Our SRNO parameterized by " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " can be solved through the associated empirical-risk minimization problem:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 182, + 287, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 182, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 100, + 182, + 287, + 214 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\| u _ {h _ {f}} ^ {(k)} - \\mathcal {S} _ {\\theta} \\left(a _ {h _ {c}} ^ {(k)}\\right) \\| _ {\\mathcal {H}}. \\tag {7}", + "image_path": "4c5576f18799741d001ea77f9e73c00d85ad16f1017c7f3b2f519ef64e35a141.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": "Since our data " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "a^{(k)}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "u^{(k)}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " are functions, to work with them numerically, we assume access only to their pointwise evaluations. Let " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{h_c} \\subset \\mathcal{H}(\\Omega_{h_c}) \\subset \\mathcal{H}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " be an approximation space [3] associated with " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\{x_i\\}_{i=1}^{n_{h_c}}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": ", such that for any " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{A}_{h_c}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "a(\\cdot) = \\sum_{i=1}^{n_{h_c}} a(x_i) \\xi_{x_i}(\\cdot)" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\{\\xi_{x_i}(\\cdot)\\}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " form a set of nodal basis for " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{h_c}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " in the sense that " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\xi_{x_i}(x_j) = \\delta_{ij}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{U}_{h_f} \\subset \\mathcal{H}(\\Omega_{h_f}) \\subset \\mathcal{H}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " is another approximation space associate with grid size " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": ". Note that Our observed LR-HR image pairs " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\{a, u\\}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": " may have different grid size pairs " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\{h_c, h_f\\}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": ", which means they belong to different approximation spaces " + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{A}_{h_c}, \\mathbb{U}_{h_f}\\}" + }, + { + "bbox": [ + 46, + 217, + 287, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": "Lifting. The neural operators from PDEs usually use a simple pointwise function, " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": " in (3), to expand the input channels. But for the SR problem, a deep feature encoder " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "E_{\\psi}" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": " (with trainable parameters " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": ") and spatial interpolations should be considered in the lifting operation, due to the complexity of natural images and the discretization inconsistency between LR and HR image functions. Another critical ingredient in Lifting is the incorporation scheme of coordinates. Prior study [14] demonstrates that deep CNNs can implicitly learn to encode the information about absolute positions. As shown in Fig.2 (c), the position information of the grid points " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "\\{a_l\\}_{l=1}^4" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": " have been implicitly encoded by " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "E_{\\psi}" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": " employing CNNs. Therefore, we only need to explicitly construct the coordinate features " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "\\hat{a}(x)" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": " with the fractional part of coordinate " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": " inside a grid. In order to reduce the blocky artifacts resulting from direct interpolation of the LR feature maps, we propose to concatenate the features weighted by the corresponding bilinear interpolation factors " + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "inline_equation", + "content": "s_l" + }, + { + "bbox": [ + 46, + 350, + 289, + 613 + ], + "type": "text", + "content": ". Compared to the local ensemble trick in [5], our method takes shorter running time and overcomes the oversmoothing problem. We reformulate the lifting operation in (3) as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 620, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 620, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 81, + 620, + 287, + 634 + ], + "type": "interline_equation", + "content": "z _ {0} (x) = \\mathcal {L} (\\mathbf {c}, \\left\\{s _ {l} \\cdot E _ {\\psi} \\left(a \\left(\\hat {x} _ {l}\\right)\\right), \\delta_ {l} (x) \\right\\} _ {l = 1} ^ {4}), \\tag {8}", + "image_path": "7796fcd9bdd32350acb6caad9c4c5db10c2642e4e92655578d6c82dbe0c1ede9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " are coordinates of HR image functions, " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}_l" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " are the coordinates of the four neighbors of " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta_l(x) = x - \\hat{x}_l" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = (2 / r_x, 2 / r_y)" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " represents a " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "r_x \\times r_y" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " local area in HR images with " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "r_x" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "r_y" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " the scaling factors, and " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}: \\mathbb{R}^{4 \\times (d_e + 2) + 2} \\to \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " is an local linear transformation function, with " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_e" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " the number of channels after " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_{\\psi}" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "Kernel integral. The kernel integral operator, in (4), actually tells that we can identify the hidden representation " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "z(x)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " of the underlying image function " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " with distributions [28], inner products in the settings of SRNO. We first define the kernel " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "K_{t}:\\mathbb{R}^{d_{z} + d_{z}}\\to \\mathbb{R}^{d_{z}\\times d_{z}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " with respect to the input pair " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "(z(x),z(y))" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ", rather than like (6) depending on the spatial variables " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". Furthermore, in order to more efficiently explore " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "z(x)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " we can use multiple sets of test functions defining the distributions, which may immediately remind one of a single-head self-attention [17]. Denote " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "z_{i} = z(x_{i})\\in \\mathbb{R}^{d_{z}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,n_{h_f}" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": ". The kernel integral operator can be approximated by the Monte-Carlo method (omitting the layer index):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 347, + 230, + 547, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 230, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 347, + 230, + 547, + 289 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} (\\mathcal {K} (z)) (x) = \\int_ {\\Omega} K (z (x), z (y)) z (y) \\mathrm {d} y \\\\ n _ {h _ {f}} \\end{array} \\tag {9} \\\\ \\approx \\sum_ {i = 1} ^ {\\dots h _ {f}} K (z (x), z _ {i}) z _ {i}, \\quad \\forall x \\in \\Omega_ {h _ {f}}, \\\\ \\end{array}", + "image_path": "36111aa5be85d5e240b9aaae6c6a060b505a69bc6b1ea519f53ea264a42195cb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 292, + 333, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 333, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 333, + 301 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 302, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 302, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 325, + 302, + 545, + 340 + ], + "type": "interline_equation", + "content": "K (z (x), z _ {i}) = \\frac {\\exp \\left(\\frac {\\langle W _ {q} z (x) , W _ {k} z _ {i} \\rangle}{\\sqrt {d _ {z}}}\\right)}{\\sum_ {j = 1} ^ {n _ {h _ {f}}} \\exp \\left(\\frac {\\langle W _ {q} z _ {j} , W _ {k} z _ {i} \\rangle}{\\sqrt {d _ {z}}}\\right)} W _ {v}. \\tag {10}", + "image_path": "a5ec307a7b05219a491758e8181f1562659222e26130757b70b28855293d3660.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": "In the language of transformers, the matrices " + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}, W_{v} \\in \\mathbb{R}^{d_{z} \\times d_{z}}" + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": " correspond to the queries, keys and values functions respectively. By allowing " + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": " to be the sum of multiple functions with separate trainable parameters, the multi-head self-attention can also be rewritten like in (9). For every observation " + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "inline_equation", + "content": "(a, u)" + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": ", this operation has a complexity of " + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "inline_equation", + "content": "O(n_{h_{f}}^{2} d_{z})" + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": ", which is unaffordable for the SR problem where the sampling number " + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "inline_equation", + "content": "n_{h_{f}}" + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": " usually reaches to 10,000 or beyond. To overcome this issue, we employ the Galerkin-type attention operator which has a linear complexity " + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "inline_equation", + "content": "O(n_{h_{f}} d_{z}^{2})" + }, + { + "bbox": [ + 304, + 341, + 547, + 496 + ], + "type": "text", + "content": ". The approximation capacity of a linearized Galerkin-type attention has been proved to be comparable to a Petrov-Galerkin projection [3]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": "We suppose that a image function " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " is locally integrable, i.e., " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " is measurable and " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\int_{\\Omega} |f| < \\infty" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " for every compact " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": ". The idea behind distributions [28] is to identify " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\int f \\phi" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " by suitably choosing a \"test function\" " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": ", rather than by a series of separate evaluations " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " at coordinates " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "x \\in \\Omega" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "v(x) = W_v z(x)" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "k(x) = W_k z(x)" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "q(x) = W_q z(x)" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " be the keys, queries and values functions respectively, which are " + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "d_z" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": " dimensional vector-valued functions, with a subscript denoting the corresponding component. The kernel integral attention then can be written in a component-wise form (" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "inline_equation", + "content": "j = 1, \\dots, d_z" + }, + { + "bbox": [ + 304, + 499, + 547, + 632 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 647, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\left(\\mathcal {K} (z)\\right) (x)\\right) _ {j} = \\sum_ {l = 1} ^ {d _ {z}} \\left\\langle k _ {l}, v _ {j} \\right\\rangle q _ {l} (x) \\tag {11} \\\\ \\approx \\sum_ {l = 1} ^ {d _ {z}} \\left(\\int_ {\\Omega} k _ {l} (y) v _ {j} (y) \\mathrm {d} y\\right) q _ {l} (x), \\quad \\forall x \\in \\Omega_ {h _ {f}}. \\\\ \\end{array}", + "image_path": "d7e8e31520aee3d11855d85a0b39d3a24286ec8a1668df7524151d39a1f709a7.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18250" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "text", + "content": "Denote the evaluations " + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "inline_equation", + "content": "(z(x_{1}),\\ldots ,z(x_{n_{h_{f}}}))^{T}" + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\in" + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n_{h_f}\\times d_z}" + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "inline_equation", + "content": "Q = \\mathbf{z}W_{q},K = \\mathbf{z}W_{k},V = \\mathbf{z}W_{v}\\in" + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n_{h_f}\\times d_z}" + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "text", + "content": ". The columns of " + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "inline_equation", + "content": "Q / K / V" + }, + { + "bbox": [ + 47, + 71, + 287, + 148 + ], + "type": "text", + "content": " contain the vector representations of the learned basis functions, spanning certain subspaces of the latent representation Hilbert spaces respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 124, + 148, + 287, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 148, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 124, + 148, + 287, + 163 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = Q \\left(\\widetilde {K} ^ {T} \\widetilde {V}\\right) / n _ {h _ {f}}, \\tag {12}", + "image_path": "ce090d808afe5d327bc287056273a757ac1ab015043506f4ecbba74d2bfa7b4d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\widetilde{K} = \\operatorname{Ln}(K), \\widetilde{V} = \\operatorname{Ln}(K)" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\operatorname{Ln}(\\cdot)" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": " the layer normalization. The " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\widetilde{K}^T \\widetilde{V}" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": " contains the coefficients for the linear combination of the basis vectors " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\{q_j\\}_{j=1}^{d_z}" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": " to form the output " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": ". Consequently, the global correlations are reflected through the components of " + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\widetilde{K}^T \\widetilde{V}" + }, + { + "bbox": [ + 47, + 167, + 287, + 278 + ], + "type": "text", + "content": ". As a result, SRNO can utilize all information effectively without the feature unfolding [5] or local convolutions [18], significantly suppressing the discontinuous patterns that appear around feature boundaries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": "In addition to providing a global aggregation for the output " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": " at each sampling point, the linear Galerkin-type attention (12) has the ability to obtain a quasi-optimal approximation in the current approximation space spanned with the columns of " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": " [3]. However, the expressive capability of the current bases in " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": " solely depends on the current input LR image through the latent representation " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ", is there any chance that we can enrich the bases which contains some extra and useful information for SR reconstruction, but not included in the input " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": "? For this purpose, the point-wise FFN " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{O}:\\mathbb{R}^{d_z}\\to \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": " in Fig.2 (b) introduces nonlinearities on one hand, and the positions concatenated in " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": " enhance the bases on the other. In this way, the basis functions are being constantly enriched, and we reformulate the iterative process (See supplementary for a principled discussion):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 479, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 479, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 479, + 287, + 491 + ], + "type": "interline_equation", + "content": "z _ {t + 1} (x) = z _ {t} (x) + \\mathcal {O} ((\\mathcal {K} _ {t} (z _ {t})) (x) + z _ {t} (x)). \\tag {13}", + "image_path": "73b337ffc412cbe641624017359dc4e92e1ba055988e90244293c5d32f2a9040.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": "Network details. The Network architecture is shown in Fig.2. As to the feature encoder " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{\\psi}" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": ", we employ EDSR-baseline [21], or RDN [41], both of which drop their upsampling layers, and their output channel dimensions " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{e} = 64" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": ". The CNNs in " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{\\psi}" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " assist SRNO in capturing the common basis functions from the ensemble of training samples, while the Galerkin-type attention layers provide the instance-specific basis enhancement. We employ the multi-head attention scheme in [36] by dividing the queries, keys and values into " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "n_{\\text{heads}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " parts with each of dimension " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{z} / n_{\\text{heads}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": ". In our implementation, " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{z} = 256" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "n_{\\text{heads}} = 16" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": ", yielding 16-dimensional output values. We only use two iterations (" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "T = 2" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": ") in the kernel integral operator, which already outperforms previous works, while keeping the running time advantage. Note that we utilize " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " convolutions to replace all the linear layers in SRNO, since they have a GPU-friendly data structure. The detailed network structures are listed in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 386, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 386, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 386, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 91, + 370, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 370, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 370, + 104 + ], + "type": "text", + "content": "4.1. Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 110, + 545, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 110, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 305, + 110, + 545, + 181 + ], + "type": "text", + "content": "Dataset. We used the DIV2K dataset [1] for network training, while the DIV2K validation set [1] and four benchmark datasets, including Set5 [2], Set14 [39], B100 [25] and Urban100 [12], for evaluation. Peak signal-to-noise ratio (PSNR) is used as the evaluation metric. Similar to [5, 11], we cropped image boundaries when computing PSNRs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": "Implementation Details. Numerically, we have to explore the underlying continuous image function through access to its point-wise evaluations on the HR and LR grids. Let " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " be the batch size and " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "n_{h_c}" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " be the LR sampling counts. We first sample " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " random scales " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "r^{(i)}" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " following a uniform distribution " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(1,4)" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": ", and then crop " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " patches of sizes " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\{\\sqrt{n}_{h_c}r^{(i)}\\times \\sqrt{n}_{h_c}r^{(i)}\\}_{i = 1}^B" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " from the HR training images (one per each). We find " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "n_{h_c} = 128^2" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " is sufficient for SRNO to represent most natural image patches and fix this choice in the sequel. The LR counterparts are downsampled using bicubic interpolation with the corresponding " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "r^{(i)}" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": ". In order to keep a consistent number of supervision points for the LR patches of different scales in a single batch, we randomly sample " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "128^2" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " HR pixels and calculate the corresponding fractional coordinates on the coarse grid associated with " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "r^{(i)}" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": ", as is done in [5]. We use an L1 loss [21] and the Adam [16] optimizer with an initial learning rate " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "4\\times 10^{-5}" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": " and the maximum " + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "inline_equation", + "content": "4\\times 10^{-4}" + }, + { + "bbox": [ + 305, + 182, + 546, + 422 + ], + "type": "text", + "content": ". All models are trained for 1000 epochs with batch size 64, and the learning rate decays by the cosine annealing after a warm-up phase of 50 epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 431, + 380, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 380, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 380, + 442 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 449, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 568 + ], + "type": "text", + "content": "Quantitative result. On the DIV2K validation set, Table 1 presents a quantitative comparison among our SRNO and three existing arbitrary-scale SR methods, MetaSR [5, 11], LIIF [5], and LTE [18]. Results when EDSR-baseline [21], and RDN [41] are used as encoders are displayed in the top and bottom rows, respectively. One can observe that SRNO provides the best results over all scale factors, irrespective of the encoder employed. Additionally, on in-distribution scales " + }, + { + "bbox": [ + 305, + 449, + 545, + 568 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 3, \\times 4)" + }, + { + "bbox": [ + 305, + 449, + 545, + 568 + ], + "type": "text", + "content": ", our method outperforms earlier works by wide margins." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "As RDN facilitates better reconstruction accuracy than EDSR-baseline in Tab.1, we choose RDN as the encoder in the comparisons over the four benchmark datasets. The results are listed in Tab.2, where SRNO obtains all the best performances. Note that only SRNO produces substantial improvements over RDN on every in-distribution scale, although the latter trains separate models for each scale. As the experimental settings in Tab.1 and 2 only differ in the decoder part, we conclude that the Galerkin-type attention mechanism employed in SRNO does contribute to better function approximation capability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Qualitative result. Figure 3 illustrates the visual results" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18251" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 164, + 220 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 164, + 220 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 164, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 164, + 220 + ], + "type": "image", + "image_path": "dd0426702f33bfeae6d1f2418da3ef6b46c72dc177f57296e2092f32ce5e12f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 222, + 130, + 233 + ], + "lines": [ + { + "bbox": [ + 88, + 222, + 130, + 233 + ], + "spans": [ + { + "bbox": [ + 88, + 222, + 130, + 233 + ], + "type": "text", + "content": "LR Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 70, + 239, + 220 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 239, + 220 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 239, + 220 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 239, + 220 + ], + "type": "image", + "image_path": "3e91cfeaafc5249e79c6f62f954dd4454bd038cd5dca262627d56e27388ed08e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 222, + 219, + 232 + ], + "lines": [ + { + "bbox": [ + 187, + 222, + 219, + 232 + ], + "spans": [ + { + "bbox": [ + 187, + 222, + 219, + 232 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 241, + 70, + 315, + 220 + ], + "blocks": [ + { + "bbox": [ + 241, + 70, + 315, + 220 + ], + "lines": [ + { + "bbox": [ + 241, + 70, + 315, + 220 + ], + "spans": [ + { + "bbox": [ + 241, + 70, + 315, + 220 + ], + "type": "image", + "image_path": "825f491808d54b660b4da944a130f3232eb296bef87516314bf4726f03f9dc13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 222, + 294, + 233 + ], + "lines": [ + { + "bbox": [ + 260, + 222, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 260, + 222, + 294, + 233 + ], + "type": "text", + "content": "LIIF [5]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 390, + 220 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 390, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 390, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 390, + 220 + ], + "type": "image", + "image_path": "be9609ed140f485ae62d5be21306490b3371a6cff09dc11df817e2b6132b4c25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 222, + 371, + 233 + ], + "lines": [ + { + "bbox": [ + 333, + 222, + 371, + 233 + ], + "spans": [ + { + "bbox": [ + 333, + 222, + 371, + 233 + ], + "type": "text", + "content": "LTE [18]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 464, + 220 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 464, + 220 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 464, + 220 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 464, + 220 + ], + "type": "image", + "image_path": "883ea36ecbec0662783d605e54c8a509b16bfb296ddfc3b4790fda4bad9b6408.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 222, + 454, + 233 + ], + "lines": [ + { + "bbox": [ + 400, + 222, + 454, + 233 + ], + "spans": [ + { + "bbox": [ + 400, + 222, + 454, + 233 + ], + "type": "text", + "content": "SRNO (ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 539, + 220 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 539, + 220 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 539, + 220 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 539, + 220 + ], + "type": "image", + "image_path": "48526af59f86c9902ba932eaf5b32acf5ebfb6d31d017cb98bde051a1efa11d9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 222, + 509, + 232 + ], + "lines": [ + { + "bbox": [ + 495, + 222, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 495, + 222, + 509, + 232 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 53, + 274, + 160, + 351 + ], + "blocks": [ + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "text", + "content": "Figure 3. Visual comparison on other zero-shot SR. The boxes in the first column indicate the areas that the close-ups on the right display. All methods are trained with continuous random scales in " + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\times 1 - \\times 4" + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "text", + "content": " and tested on " + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "inline_equation", + "content": "\\times 8 / \\times 12" + }, + { + "bbox": [ + 46, + 237, + 547, + 270 + ], + "type": "text", + "content": " to evaluate the generalization capability to unknown scaling factors. RDN is used as the encoder for all methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 274, + 160, + 351 + ], + "lines": [ + { + "bbox": [ + 53, + 274, + 160, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 160, + 351 + ], + "type": "image", + "image_path": "34b39f7b2e18c000d0108ce4270c88e12eb1d6f4dce5ffc1f4ef1576aaa63c4a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 53, + 352, + 160, + 426 + ], + "blocks": [ + { + "bbox": [ + 53, + 352, + 160, + 426 + ], + "lines": [ + { + "bbox": [ + 53, + 352, + 160, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 160, + 426 + ], + "type": "image", + "image_path": "000d2fa49b6761eb1a8b2b734692ebcfb7dddc4ff7fd03a879895c788f53183b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 429, + 124, + 439 + ], + "lines": [ + { + "bbox": [ + 91, + 429, + 124, + 439 + ], + "spans": [ + { + "bbox": [ + 91, + 429, + 124, + 439 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 160, + 275, + 254, + 351 + ], + "blocks": [ + { + "bbox": [ + 160, + 275, + 254, + 351 + ], + "lines": [ + { + "bbox": [ + 160, + 275, + 254, + 351 + ], + "spans": [ + { + "bbox": [ + 160, + 275, + 254, + 351 + ], + "type": "image", + "image_path": "6b669bb5797577475d1bf76f1041d42035ce06d54384da873c2d2f908051515c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 160, + 352, + 254, + 426 + ], + "blocks": [ + { + "bbox": [ + 160, + 352, + 254, + 426 + ], + "lines": [ + { + "bbox": [ + 160, + 352, + 254, + 426 + ], + "spans": [ + { + "bbox": [ + 160, + 352, + 254, + 426 + ], + "type": "image", + "image_path": "e96ae44bd28d2c30447fed97abf27444ecb99f0ddb13f9152845ff522e685758.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 429, + 225, + 440 + ], + "lines": [ + { + "bbox": [ + 190, + 429, + 225, + 440 + ], + "spans": [ + { + "bbox": [ + 190, + 429, + 225, + 440 + ], + "type": "text", + "content": "LIIF [5]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 255, + 275, + 350, + 351 + ], + "blocks": [ + { + "bbox": [ + 255, + 275, + 350, + 351 + ], + "lines": [ + { + "bbox": [ + 255, + 275, + 350, + 351 + ], + "spans": [ + { + "bbox": [ + 255, + 275, + 350, + 351 + ], + "type": "image", + "image_path": "560bf66cbf7ebf40df96a061f29366d74a9b25bd02987b085b2cee4fcea7a16e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 255, + 352, + 349, + 426 + ], + "blocks": [ + { + "bbox": [ + 255, + 352, + 349, + 426 + ], + "lines": [ + { + "bbox": [ + 255, + 352, + 349, + 426 + ], + "spans": [ + { + "bbox": [ + 255, + 352, + 349, + 426 + ], + "type": "image", + "image_path": "dcdea1cca7989fa75a045bb33ee49a1669bd7d86e3d825a19333c7450cbfb226.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 429, + 321, + 440 + ], + "lines": [ + { + "bbox": [ + 282, + 429, + 321, + 440 + ], + "spans": [ + { + "bbox": [ + 282, + 429, + 321, + 440 + ], + "type": "text", + "content": "LTE [18]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 350, + 275, + 444, + 351 + ], + "blocks": [ + { + "bbox": [ + 350, + 275, + 444, + 351 + ], + "lines": [ + { + "bbox": [ + 350, + 275, + 444, + 351 + ], + "spans": [ + { + "bbox": [ + 350, + 275, + 444, + 351 + ], + "type": "image", + "image_path": "df8d331c306c59f4cfad06bcc5d4b097b0eb184c01a3ae5c326fa08cd60dd322.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 350, + 352, + 443, + 426 + ], + "blocks": [ + { + "bbox": [ + 350, + 352, + 443, + 426 + ], + "lines": [ + { + "bbox": [ + 350, + 352, + 443, + 426 + ], + "spans": [ + { + "bbox": [ + 350, + 352, + 443, + 426 + ], + "type": "image", + "image_path": "4be934e6402d0f6565dea6dcb0865fed227593ee66ac3e6acb7539f6dfdde7f7.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 429, + 424, + 440 + ], + "lines": [ + { + "bbox": [ + 369, + 429, + 424, + 440 + ], + "spans": [ + { + "bbox": [ + 369, + 429, + 424, + 440 + ], + "type": "text", + "content": "SRNO (ours)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 444, + 546, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 444, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 546, + 464 + ], + "type": "text", + "content": "Figure 4. Visual comparison on non-integer scales. All methods use RDN as the encoder and are trained with continuous random scales in " + }, + { + "bbox": [ + 46, + 444, + 546, + 464 + ], + "type": "inline_equation", + "content": "\\times 1 - \\times 4" + }, + { + "bbox": [ + 46, + 444, + 546, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 444, + 275, + 539, + 351 + ], + "blocks": [ + { + "bbox": [ + 444, + 275, + 539, + 351 + ], + "lines": [ + { + "bbox": [ + 444, + 275, + 539, + 351 + ], + "spans": [ + { + "bbox": [ + 444, + 275, + 539, + 351 + ], + "type": "image", + "image_path": "21a95dc3d49abdc3f5abc335caa86fc6081b67bcbe16ae45978a9a601640615a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 444, + 352, + 539, + 426 + ], + "blocks": [ + { + "bbox": [ + 444, + 352, + 539, + 426 + ], + "lines": [ + { + "bbox": [ + 444, + 352, + 539, + 426 + ], + "spans": [ + { + "bbox": [ + 444, + 352, + 539, + 426 + ], + "type": "image", + "image_path": "b8ce4b02a4aa04fab8f37bbea4290623205ff0d34d3f8869774d963c207cdbe5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 429, + 499, + 439 + ], + "lines": [ + { + "bbox": [ + 484, + 429, + 499, + 439 + ], + "spans": [ + { + "bbox": [ + 484, + 429, + 499, + 439 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 479, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 635 + ], + "type": "text", + "content": "obtained from two images in the Urban100 and DIV2K validation datasets. Both the regions marked with a rectangle are rich in high-frequency structures. The bicubic interpolations exhibit severe aliasing artifacts, which obviously have been transferred into the latent codes of LIIF [5] as it directly relies on the interpolation of the feature maps and the bilinear ensemble. Thanks to the local Fourier frequency estimation step, LTE [18] performs somewhat better in partial regions (e.g. the top-right corner of the building in the second row), but aliasing still dominates on the whole. In contrast, SRNO successfully recovers the fine structures, which again verifies the importance of the global integral kernels in capturing the correct overall structures." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Figure 4 shows two examples of scene text images on non-integer scales. In the first row, only SRNO can clearly recover the word \"ANTIQUES\". And in the second row, only SRNO can consistently restore the two e's. Obviously, the Galerkin-type attention attributes to this global consistency." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 479, + 400, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 400, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 400, + 493 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "Data sampling. The quantity of sampling points is crucial in our super-resolution operator learning architecture. Figure 5 shows that almost all scales perform better when using additional pixel samples. However, the benefit becomes less as the numbers of samples increases. " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "96^{2}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " points would be good trade-offs between training time and performance. Figure 6 shows that more sampling points can more accurately represent an RGB-based function, which facilitates to learn the map between two approximation spaces. Using the DIV2K (800 images) and Flickr2K (2650 images) datasets, we also train an EDSR-baseline-MLP with position embedding, denoted by " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{MLP}^{\\varphi}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ". We can see that our SRNO " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "(64^{2}/96^{2}/128^{2})" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " obtains better performance for scales " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 6" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ", trained with only 800 images in DIV2K. The curve for " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(128^{2})" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " in Fig.5 reflects the performance advantage of SRNO " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "(128^{2})" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ", both trained with DIV2K dataset. Random and sequential sampling results are presented in Supplementary." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18252" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 69, + 473, + 193 + ], + "blocks": [ + { + "bbox": [ + 123, + 69, + 473, + 193 + ], + "lines": [ + { + "bbox": [ + 123, + 69, + 473, + 193 + ], + "spans": [ + { + "bbox": [ + 123, + 69, + 473, + 193 + ], + "type": "table", + "html": "
MethodIn-distributionOut-of-distribution
×2×3×4×6×12×18×24×30
Bicubic31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [21]34.5530.9028.94-----
EDSR-baseline-MetaSR [5,11]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.6730.9629.0026.7523.7122.1721.1820.48
EDSR-baseline-LTE [18]34.7231.0229.0426.8123.7822.2321.2420.53
EDSR-baseline-SRNO (ours)34.8531.1129.1626.9023.8422.2921.2720.56
RDN-MetaSR [5,11]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.9931.2629.2726.9923.8922.3421.3120.59
RDN-LTE [18]35.0431.3229.3327.0423.9522.4021.3620.64
RDN-SRNO (ours)35.1631.4229.4227.1224.0322.4621.4120.68
", + "image_path": "725d5f433ee0999f5ddd039dba2f18de096fda1dde0f1688213a437b8b666298.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 232, + 502, + 376 + ], + "blocks": [ + { + "bbox": [ + 46, + 194, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 228 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on DIV2K validation set (PSNR (dB)). The best performance are bolded. EDSR-baseline trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in " + }, + { + "bbox": [ + 46, + 194, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\times 1 - \\times 4" + }, + { + "bbox": [ + 46, + 194, + 547, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 232, + 502, + 376 + ], + "lines": [ + { + "bbox": [ + 96, + 232, + 502, + 376 + ], + "spans": [ + { + "bbox": [ + 96, + 232, + 502, + 376 + ], + "type": "table", + "html": "
MethodIn-distributionOut-of-distributionIn-distributionOut-of-distribution
×2×3×4×6×8×2×3×4×6×8
Set5Set14
RDN [41]38.2434.7132.47--34.0130.5728.81--
RDN-MetaSR [5, 11]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.97
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.15
RDN-LTE [18]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.16
RDN-SRNO (ours)38.3234.8432.6929.3827.2834.2730.7128.9726.7625.26
B100Urban100
RDN [41]32.3429.2627.72--32.8928.8026.61--
RDN-MetaSR [5, 11]32.3329.2627.7125.9024.8332.9228.8226.5523.9922.59
RDN-LIIF [5]32.3229.2627.7425.9824.9132.8728.8226.6824.2022.79
RDN-LTE [18]32.3629.3027.7726.0124.9533.0428.9726.8124.2822.88
RDN-SRNO (ours)32.4329.3727.8326.0424.9933.3329.1426.9824.4323.02
", + "image_path": "a0dd12f77d9fb82989ffc5b062257d5d9dba1364d3ccf666fcff8dd742420992.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 60, + 422, + 277, + 494 + ], + "blocks": [ + { + "bbox": [ + 46, + 376, + 546, + 410 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 546, + 410 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on benchmark datasets (PSNR (dB)). The best performances are in bold. RDN trains separate models for the three in-distribution scales. The rest methods use a single model for all scales, and are trained with continuous random scales uniformly sampled in " + }, + { + "bbox": [ + 46, + 376, + 546, + 410 + ], + "type": "inline_equation", + "content": "\\times 1 - \\times 4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 422, + 277, + 494 + ], + "lines": [ + { + "bbox": [ + 60, + 422, + 277, + 494 + ], + "spans": [ + { + "bbox": [ + 60, + 422, + 277, + 494 + ], + "type": "table", + "html": "
Method#Params. (M)#FLOPs (G)
EDSR-baseline-LIIF1.685.0
EDSR-baseline-LTE1.775.3
EDSR-baseline-SRNO (ours)2.065.8
RDN-LIIF22.4765.2
RDN-LTE22.5755.5
RDN-SRNO (ours)22.8746.0
", + "image_path": "509000c426838d5e84106f14ec1da6e75946067ce4a4d045df13b45c3cd3e426.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "type": "text", + "content": "Table 3. Comparison of model parameters and FLOPs. Metrics are measured under the setting of " + }, + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 47, + 495, + 287, + 518 + ], + "type": "text", + "content": " sample points." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "Other design choices. In Tab.4, we retrain the following models using EDSR-baseline [21]. All the tests were conducted using a single NVIDIA RTX 3090. In comparing our interpolation-free method to SRNO(-i) using nearest-neighbor interpolation, we find that it works better and faster. It should be noted that we only generate the image once, as opposed to employing an MLP four times as in LIIF [5] and LTE [18]. Figure 7 shows that SRNO, compared to SRNO(-i), alleviates the blocky artifact, and compared to the local ensemble [18], overcomes the oversmoothing problem. By comparing SRNO with SRNO(-h), it can be shown that cutting the number of heads from 16 to 8 causes an increase in time costs because GPU's parallelism is reduced. We see that reducing the number of basis functions and the iterative updating layers results in a" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 323, + 425, + 530, + 548 + ], + "blocks": [ + { + "bbox": [ + 323, + 425, + 530, + 548 + ], + "lines": [ + { + "bbox": [ + 323, + 425, + 530, + 548 + ], + "spans": [ + { + "bbox": [ + 323, + 425, + 530, + 548 + ], + "type": "image", + "image_path": "47fcb728c79952a60a3eb99c9bfc229849dc434e0bd30dcd79e6ffd3dad3fad0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "lines": [ + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "type": "text", + "content": "Figure 5. Data samples vs. PSNRs. Evaluated on the Urban100. We train models using " + }, + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "type": "inline_equation", + "content": "48^{2} / 64^{2} / 96^{2} / 128^{2}" + }, + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "type": "text", + "content": " pixel samples and calculate the PSNR increment relative to LTE. " + }, + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 305, + 551, + 545, + 597 + ], + "type": "text", + "content": " indicates training on DIV2K [1] (800 images) plus Flickr2K [33] (2650 images) dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "type": "text", + "content": "considerable performance drop by comparing SRNO with SRNO(-w) and SRNO(-l). The number of basis functions " + }, + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "type": "inline_equation", + "content": "(d_z)" + }, + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "type": "text", + "content": " and the number of subspaces " + }, + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "type": "inline_equation", + "content": "(n_{\\text{heads}})" + }, + { + "bbox": [ + 305, + 611, + 546, + 661 + ], + "type": "text", + "content": " play an important role in the approximation power of our model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 670, + 399, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 399, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 399, + 683 + ], + "type": "text", + "content": "4.4. Dynamic Basis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "We investigate the dynamic basis updating phenomenon across the approximation subspaces in this section. We" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18253" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 289, + 172 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 172 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 172 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 172 + ], + "type": "table", + "html": "
MethodIn-dis.Out-of-dis.Time (ms)
×2×3×4×6×8
SRNO33.8330.5028.7926.5525.05149
SRNO (-i)33.8130.4728.7626.5525.00156
SRNO (-h)33.8730.4728.7626.5625.02153
SRNO (-w)33.8030.4528.7526.5325.00124
SRNO (-l)33.8230.4528.7526.5525.00117
LIIF [5]33.6830.3628.6426.4624.94183
LTE [18]33.7230.3728.6526.5024.99205
LTE+ [18]33.7130.4128.6726.4924.98160
", + "image_path": "e2044e06426e8c62a221facb2786f11b500859df8d43c6051e656d83ecb1a59d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "text", + "content": "Table 4. Quantitative ablation study on design choices of SRNO. Evaluated on the Set14 (PSNR(dB)). -h/w/l refers to " + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "inline_equation", + "content": "n_{\\text{heads}} = 8" + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "inline_equation", + "content": "d_z = 128" + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "text", + "content": " correspondingly. -i refers to using nearest-neighborhood interpolation and position embedding. EDSR-baseline is used in lifting. The running time is averaged over Set14 on " + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\times 2/3/4/6/8" + }, + { + "bbox": [ + 46, + 173, + 288, + 251 + ], + "type": "text", + "content": " SR. For each scale SR, we run every model three times and choose the minimum running time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 53, + 258, + 128, + 301 + ], + "blocks": [ + { + "bbox": [ + 53, + 258, + 128, + 301 + ], + "lines": [ + { + "bbox": [ + 53, + 258, + 128, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 128, + 301 + ], + "type": "image", + "image_path": "5e7a8a0b139c49ab63c9b550cc267a7c8c55951b849abfef372032801904cdf0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 304, + 108, + 313 + ], + "lines": [ + { + "bbox": [ + 75, + 304, + 108, + 313 + ], + "spans": [ + { + "bbox": [ + 75, + 304, + 108, + 313 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 258, + 205, + 301 + ], + "blocks": [ + { + "bbox": [ + 129, + 258, + 205, + 301 + ], + "lines": [ + { + "bbox": [ + 129, + 258, + 205, + 301 + ], + "spans": [ + { + "bbox": [ + 129, + 258, + 205, + 301 + ], + "type": "image", + "image_path": "a10d1b959a59eef5e1fdc24f3be3f27577b4764b4ee8fbd275603de62e49ee0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 303, + 193, + 314 + ], + "lines": [ + { + "bbox": [ + 141, + 303, + 193, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 193, + 314 + ], + "type": "text", + "content": "SRNO " + }, + { + "bbox": [ + 141, + 303, + 193, + 314 + ], + "type": "inline_equation", + "content": "(48^{2})" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 205, + 258, + 282, + 301 + ], + "blocks": [ + { + "bbox": [ + 205, + 258, + 282, + 301 + ], + "lines": [ + { + "bbox": [ + 205, + 258, + 282, + 301 + ], + "spans": [ + { + "bbox": [ + 205, + 258, + 282, + 301 + ], + "type": "image", + "image_path": "9a4fe9cf1714ab1205905d12439dc3714d161207350b20afeaf60e99192ded3d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 315, + 128, + 357 + ], + "blocks": [ + { + "bbox": [ + 53, + 315, + 128, + 357 + ], + "lines": [ + { + "bbox": [ + 53, + 315, + 128, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 128, + 357 + ], + "type": "image", + "image_path": "54b7c0d48005834ecae47f16e58e6872d750433dd29996d10a5bb52279ab6999.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 360, + 117, + 372 + ], + "lines": [ + { + "bbox": [ + 65, + 360, + 117, + 372 + ], + "spans": [ + { + "bbox": [ + 65, + 360, + 117, + 372 + ], + "type": "text", + "content": "SRNO " + }, + { + "bbox": [ + 65, + 360, + 117, + 372 + ], + "type": "inline_equation", + "content": "(96^{2})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 129, + 315, + 205, + 357 + ], + "blocks": [ + { + "bbox": [ + 129, + 315, + 205, + 357 + ], + "lines": [ + { + "bbox": [ + 129, + 315, + 205, + 357 + ], + "spans": [ + { + "bbox": [ + 129, + 315, + 205, + 357 + ], + "type": "image", + "image_path": "3327236be7feb94ddbe9f32b927f0bca7acc719c6a1843be3e97ee0ac1302a24.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 360, + 195, + 372 + ], + "lines": [ + { + "bbox": [ + 139, + 360, + 195, + 372 + ], + "spans": [ + { + "bbox": [ + 139, + 360, + 195, + 372 + ], + "type": "text", + "content": "SRNO " + }, + { + "bbox": [ + 139, + 360, + 195, + 372 + ], + "type": "inline_equation", + "content": "(128^{2})" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 205, + 315, + 282, + 357 + ], + "blocks": [ + { + "bbox": [ + 218, + 304, + 269, + 314 + ], + "lines": [ + { + "bbox": [ + 218, + 304, + 269, + 314 + ], + "spans": [ + { + "bbox": [ + 218, + 304, + 269, + 314 + ], + "type": "text", + "content": "SRNO " + }, + { + "bbox": [ + 218, + 304, + 269, + 314 + ], + "type": "inline_equation", + "content": "(64^{2})" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 315, + 282, + 357 + ], + "lines": [ + { + "bbox": [ + 205, + 315, + 282, + 357 + ], + "spans": [ + { + "bbox": [ + 205, + 315, + 282, + 357 + ], + "type": "image", + "image_path": "e2e62743bd075ebccf3a343d5ea855c14df138552e87fe2bc1323f6e2f51e101.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 360, + 251, + 370 + ], + "lines": [ + { + "bbox": [ + 236, + 360, + 251, + 370 + ], + "spans": [ + { + "bbox": [ + 236, + 360, + 251, + 370 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 56, + 427, + 167, + 475 + ], + "blocks": [ + { + "bbox": [ + 46, + 375, + 287, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 419 + ], + "type": "text", + "content": "Figure 6. Effects of sampling global image structures. Test on " + }, + { + "bbox": [ + 46, + 375, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 46, + 375, + 287, + 419 + ], + "type": "text", + "content": " scale. EDSR-baseline is used as the encoder, while the data in parentheses indicate the pixel numbers sampled in the HR images during training." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 427, + 167, + 475 + ], + "lines": [ + { + "bbox": [ + 56, + 427, + 167, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 167, + 475 + ], + "type": "image", + "image_path": "db291cadddc205beffb22995687137e328b777d3069979df3562cff2578a470c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 168, + 427, + 278, + 475 + ], + "blocks": [ + { + "bbox": [ + 168, + 427, + 278, + 475 + ], + "lines": [ + { + "bbox": [ + 168, + 427, + 278, + 475 + ], + "spans": [ + { + "bbox": [ + 168, + 427, + 278, + 475 + ], + "type": "image", + "image_path": "b22024d9b9f4740c4e9ea6b7ccbf9e902f562134ac17ee535a281bf9542fbff0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 56, + 489, + 167, + 537 + ], + "blocks": [ + { + "bbox": [ + 102, + 478, + 121, + 487 + ], + "lines": [ + { + "bbox": [ + 102, + 478, + 121, + 487 + ], + "spans": [ + { + "bbox": [ + 102, + 478, + 121, + 487 + ], + "type": "text", + "content": "LTE" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 489, + 167, + 537 + ], + "lines": [ + { + "bbox": [ + 56, + 489, + 167, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 167, + 537 + ], + "type": "image", + "image_path": "37dacdff4b4147a94590800ff0e304ae9e78377f8f456c4889da294bae421134.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 540, + 126, + 550 + ], + "lines": [ + { + "bbox": [ + 97, + 540, + 126, + 550 + ], + "spans": [ + { + "bbox": [ + 97, + 540, + 126, + 550 + ], + "type": "text", + "content": "SRNO" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 554, + 287, + 588 + ], + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 588 + ], + "type": "text", + "content": "Figure 7. Interpolation vs. interpolation-free of SRNO. Test on " + }, + { + "bbox": [ + 47, + 554, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 47, + 554, + 287, + 588 + ], + "type": "text", + "content": " scale. EDSR-baseline is used as an encoder. -i refers to using nearest-neighborhood interpolation and position embedding." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 168, + 489, + 278, + 537 + ], + "blocks": [ + { + "bbox": [ + 201, + 478, + 244, + 488 + ], + "lines": [ + { + "bbox": [ + 201, + 478, + 244, + 488 + ], + "spans": [ + { + "bbox": [ + 201, + 478, + 244, + 488 + ], + "type": "text", + "content": "SRNO (-i)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 489, + 278, + 537 + ], + "lines": [ + { + "bbox": [ + 168, + 489, + 278, + 537 + ], + "spans": [ + { + "bbox": [ + 168, + 489, + 278, + 537 + ], + "type": "image", + "image_path": "f20ac7ba617af700e2c7d0f2e3f247d1cd2af586f71af502fee684edcaf096ba.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 540, + 230, + 550 + ], + "lines": [ + { + "bbox": [ + 215, + 540, + 230, + 550 + ], + "spans": [ + { + "bbox": [ + 215, + 540, + 230, + 550 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "show two exemplar basis functions in the two consecutive Galerkin attention layers in Fig.8. The basis function in the second layer appears to be more structured than in the previous layer. The dynamic update of the approximation subspaces spanned by the column vectors of " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "Q / K / V" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " comes from two factors, the concatenated random coordinates and the nonlinearity introduced by the FNN. Without them, the Galerkin-type attention would just be a linear combination of the bases in the current approximation subspaces, and" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 186 + ], + "type": "image", + "image_path": "29aeeff4b6814001c6ea44ea010f290411846abb9a254ea79ed3cb4f151d48f8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "type": "text", + "content": "Figure 8. Dynamic basis update. The annotation " + }, + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "type": "inline_equation", + "content": "t = k" + }, + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "type": "text", + "content": " refers to the iterative layer number. The pictures display one of the learned basis functions in " + }, + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "type": "text", + "content": " and the latent representations " + }, + { + "bbox": [ + 305, + 188, + 547, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 321, + 232, + 533, + 337 + ], + "blocks": [ + { + "bbox": [ + 321, + 232, + 533, + 337 + ], + "lines": [ + { + "bbox": [ + 321, + 232, + 533, + 337 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 533, + 337 + ], + "type": "image", + "image_path": "0f242339dc087f80730c3780d06708d78e9dca9de35a1e1be5105fc0cfa4a03c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 338, + 542, + 350 + ], + "lines": [ + { + "bbox": [ + 308, + 338, + 542, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 542, + 350 + ], + "type": "text", + "content": "Figure 9. Ranks of the matrix " + }, + { + "bbox": [ + 308, + 338, + 542, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 308, + 338, + 542, + 350 + ], + "type": "text", + "content": ". Evaluated on the Set5 dataset." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 366, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 463 + ], + "type": "text", + "content": "SRNO would not be able to take any chances to enrich the basis for every individual image through optimizations. The increasing ranks of the latent representation matrix " + }, + { + "bbox": [ + 304, + 366, + 545, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 366, + 545, + 463 + ], + "type": "text", + "content": ", shown in Fig.9, provide the evidence that Galerkin-type attention is becoming more complex than a straightforward linear combination of the existing bases. Due to the space limitation, we have provided a principled discussion on the dynamic basis update processes in the supplementary material." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we proposed the Super-Resolution Neural Operator (SRNO) for continuous super-resolution. In SRNO, each image is seen as a function, and our method learns a map between finite-dimensional function spaces, which allows SRNO can be trained and generalize on different levels of discretization. First, in the Lifting, we use CNN-based encoders to capture feature maps from LR images and design a simple but efficient interpolation-free method, addressing the discretization inconsistency problem encountered in SR. Second, to approximate our target function in the Iterative Kernel Integration, we employ a linear attention operator that has been proved to be comparable to the Petrov-Galerkin projection. Finally, we map the last hidden representation to the output function. Experimental results show that our SRNO outperforms other arbitrary-scale SR methods in performance as well as computation time, and particularly in the capability of capturing the global image structures." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18254" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 191 + ], + "type": "text", + "content": "[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In Proceedings of the British Machine Vision Conference, pages 135.1-135.10. BMVA Press, 2012. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "type": "text", + "content": "[3] Shuhao Cao. Choose a transformer: Fourier or galerkin. Advances in Neural Information Processing Systems, 34:24924-24940, 2021. 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "type": "text", + "content": "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 335 + ], + "type": "text", + "content": "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 401 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 424 + ], + "type": "text", + "content": "[7] Alexandre Ern and Jean-Luc Guermond. Theory and practice of finite elements, volume 159. Springer, 2004. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[8] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9199-9208, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 523 + ], + "type": "text", + "content": "[9] John Guibas, Morteza Mardani, Zongyi Li, Andrew Tao, Anima Anandkumar, and Bryan Catanzaro. Efficient token mixing for transformers via adaptive fourier neural operators. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 568 + ], + "type": "text", + "content": "[10] Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations. Advances in Neural Information Processing Systems, 34:24048-24062, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "text", + "content": "[11] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 2, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 669 + ], + "type": "text", + "content": "[12] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] Rakhoon Hwang, Jae Yong Lee, Jin Young Shin, and Hyung Ju Hwang. Solving pde-constrained control problems using operator learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(4):4504-4512, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Md Amirul Islam, Sen Jia, and Neil DB Bruce. How much position information do convolutional neural networks encode? In International Conference on Learning Representations, 2019. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 195 + ], + "type": "text", + "content": "[15] Chiyu\" Max\" Jiang, Soheil Esmaeilzadeh, Kamyar Azizzadenesheli, Karthik Kashinath, Mustafa Mustafa, Hamdi A Tchelepi, Philip Marcus, and Anima Anandkumar. Meshfreeflownet: a physics-constrained deep continuous spacetime super-resolution framework. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-15, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 219 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "type": "text", + "content": "[17] Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces. arXiv preprint arXiv:2108.08481, 2021. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 274, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 320 + ], + "type": "text", + "content": "[18] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "type": "text", + "content": "[19] Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, Anima Anandkumar, et al. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2020. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "type": "text", + "content": "[20] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833-1844, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 432, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 489 + ], + "type": "text", + "content": "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 489, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 545 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "text", + "content": "[23] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3(3):218-229, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 600, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 667 + ], + "type": "text", + "content": "[24] Salma Abdel Magid, Yulun Zhang, Donglai Wei, Won-Dong Jang, Zudi Lin, Yun Fu, and Hanspeter Pfister. Dynamic high-pass filtering and multi-spectral attention for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4288-4297, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 667, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 667, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 667, + 545, + 713 + ], + "type": "text", + "content": "[25] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18255" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 152 + ], + "type": "text", + "content": "[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 154, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 154, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 154, + 287, + 220 + ], + "type": "text", + "content": "[27] Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 243 + ], + "type": "text", + "content": "[28] Walter Rudin. Functional analysis, mcrawhill. Inc, New York, 45:46, 1991. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 246, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 312 + ], + "type": "text", + "content": "[29] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 358 + ], + "type": "text", + "content": "[30] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "text", + "content": "[31] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 471 + ], + "type": "text", + "content": "[32] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 474, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 529 + ], + "type": "text", + "content": "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 531, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 586 + ], + "type": "text", + "content": "[34] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 588, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 286, + 609 + ], + "type": "text", + "content": "[35] V Vapnik. Statistical learning theory wiley-interscience. New York, 1998. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 656 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 262 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[38] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[39] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[40] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[41] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 1, 2, 5, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18256" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_content_list.json b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec42223653d57b2d60dac1298e8fca30b769c86e --- /dev/null +++ b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_content_list.json @@ -0,0 +1,1474 @@ +[ + { + "type": "text", + "text": "SuperDisco: Super-Class Discovery Improves Visual Recognition for the Long-Tail", + "text_level": 1, + "bbox": [ + 76, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingjun Du1, Jiayi Shen1, Xiantong Zhen1,2*, Cees G. M. Snoek1 \n1AIM Lab, University of Amsterdam 2Inception Institute of Artificial Intelligence", + "bbox": [ + 161, + 179, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 312, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern image classifiers perform well on populated classes, while degrading considerably on tail classes with only a few instances. Humans, by contrast, effortlessly handle the long-tailed recognition challenge, since they can learn the tail representation based on different levels of semantic abstraction, making the learned tail features more discriminative. This phenomenon motivated us to propose SuperDisco, an algorithm that discovers super-class representations for long-tailed recognition using a graph model. We learn to construct the super-class graph to guide the representation learning to deal with long-tailed distributions. Through message passing on the super-class graph, image representations are rectified and refined by attending to the most relevant entities based on the semantic similarity among their super-classes. Moreover, we propose to meta-learn the super-class graph under the supervision of a prototype graph constructed from a small amount of imbalanced data. By doing so, we obtain a more robust super-class graph that further improves the long-tailed recognition performance. The consistent state-of-the-art experiments on the long-tailed CIFAR-100, ImageNet, Places and iNaturalist demonstrate the benefit of the discovered super-class graph for dealing with long-tailed distributions.", + "bbox": [ + 75, + 282, + 473, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 657, + 207, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper strives for long-tailed visual recognition. A computer vision challenge that has received renewed attention in the context of representation learning, as real-world deployment demands moving from balanced to imbalanced scenarios. Three active strands of work involve class re-balancing [15, 22, 32, 43, 65], information augmentation [34, 51, 54] and module improvement [29, 31, 76]. Each of these strands is intuitive and has proven empirically successful. However, all these approaches seek to improve the classification performance of the original feature space. In this paper, we instead explore a graph learning algorithm to discover the imbalanced super-class space hidden in the original feature representation.", + "bbox": [ + 75, + 681, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e749440ec9f0eb8901b26a268a20b0131ab105ebf5edf7e9cbf59627b9a42ac.jpg", + "image_caption": [ + "(a) 100 original classes" + ], + "image_footnote": [], + "bbox": [ + 506, + 260, + 658, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eee3c5a6ed26573b9ce27470e8e7f07790888c39416c22245be44f1bf5a129e2.jpg", + "image_caption": [ + "(b) 20 ground truth super-classes" + ], + "image_footnote": [], + "bbox": [ + 697, + 258, + 849, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10d07d7e9e3e83fbc64f71685a36825e3f9d556f081d90e7f11b3f7b2279bf06.jpg", + "image_caption": [ + "(c) 16 discovered super-classes" + ], + "image_footnote": [], + "bbox": [ + 506, + 369, + 658, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8cb7e4f30c411cd68eaecc5cda9fd3cc86877cd3b2fa2e2b65a67057277a5f20.jpg", + "image_caption": [ + "(d) 32 discovered super-classes", + "Figure 1. SuperDisco learns to project the original class space (a) into a relatively balanced super-class space. Different color curves indicate the different imbalance factors on the long-tailed CIFAR-100 dataset. Like the 20 super-class ground truth (b) our discovered super-classes for 16 super-classes (c) or 32 super-classes (d) provide a much better balance than the original classes." + ], + "image_footnote": [], + "bbox": [ + 700, + 368, + 849, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The fundamental problem in long-tailed recognition [18, 32, 44, 77] is that the head features and the tail features are indistinguishable. Since the head data dominate the feature distribution, they cause the tail features to fall within the head feature space. Nonetheless, humans effortlessly handle long-tailed recognition [2, 16] by leveraging semantic abstractions existing in language to gain better representations of tail objects. This intuition hints that we may discover the semantic hierarchy from the original feature space and use it for better representations of tail objects. Moreover, intermediate concepts have been shown advantageous for classification [5, 36] by allowing the transfer of shared features across classes. Nevertheless, it remains unexplored to exploit intermediate super-classes in long-tailed visual recognition that rectify and refine the original features.", + "bbox": [ + 496, + 566, + 893, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the real world, each category has a corresponding superclass, e.g., bus, taxi, and train all belong to the vehicle superclass. This observation raises the question: are super-classes of categories also distributed along a long-tail? We find empirical evidence that within the super-class space of popular datasets, the long-tailed distribution almost disappears, and each super-class has essentially the same number of samples.", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Currently with United Imaging Healthcare, Co., Ltd., China.", + "bbox": [ + 94, + 886, + 419, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19944", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In Figure 1, we show the number of training samples for each of the original classes and their corresponding super-classes in the long-tailed CIFAR-100 dataset. We observe the data imbalance of super-classes is considerably lower than those of the original classes. This reflects the fact that the original imbalanced data hardly affects the degree of imbalance of the super-classes, which means the distribution of the super-classes and original data is relatively independent. These balanced super-class features could be used to guide the original tail data away from the dominant role of the head data, thus making the tail data more discriminative. Therefore, if the super-classes on different levels of semantic abstraction over the original classes can be accurately discovered, it will help the model generalize over the tail classes. As not all datasets provide labels for super-classes, we propose to learn to discover the super-classes in this paper.", + "bbox": [ + 80, + 90, + 468, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the above observation, we make in this paper two algorithmic contributions. First, we propose in Section 3 an algorithm that learns to discover the super-class graph for long-tailed visual recognition, which we call SuperDisco. We construct a learnable graph that discovers the super-class in a hierarchy of semantic abstraction to guide feature representation learning. By message passing on the super-class graph, the original features are rectified and refined, which attend to the most relevant entities according to the similarity between the original image features and super-classes. Thus, the model is endowed with the ability to free the original tail features from the dominance of the head features using the discovered and relatively balanced super-class representations. Even when faced with the severe class imbalance challenges, e.g., iNaturalist, our SuperDisco can still refine the original features by finding a more balanced super-class space using a more complex hierarchy. As a second contribution, we propose in Section 4 a meta-learning variant of our SuperDisco algorithm to discover the super-class graph, enabling the model to achieve even more balanced image representations. To do so, we use a small amount of balanced data to construct a prototype-based relational graph, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. Last, in Section 5 we report experiments on four long-tailed benchmarks: CIFAR-100-LT, ImageNet-LT, Places-LT, and iNaturalist, and verify that our discovered super-class graph performs better for tail data in each dataset. Before detailing our contributions, we first embed our proposal in related work.", + "bbox": [ + 80, + 333, + 468, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 80, + 784, + 210, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-tailed recognition. Several strategies have been proposed to address class imbalance in recognition. We categorize them into three groups. Those in the first group are based on class re-balancing [8, 30, 44, 75], which balance the training sample numbers of different classes during model training. Class re-balancing methods also could be catego-", + "bbox": [ + 80, + 809, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rized into three different groups, i.e., re-sampling [22, 32, 43, 67], cost-sensitive learning [14, 37, 56, 78, 79, 86] and logit adjustment [27, 45, 57, 59]. Class re-balancing methods improve the performance of the tail classes at the expense of the performance of the head classes. The second group is based on information augmentation, introducing additional information into model training to improve long-tailed learning performance. We identify four kinds of methods in the information augmentation scope, i.e., transfer learning, which includes head-to-tail knowledge transfer [6, 42, 64, 73], knowledge distillation [28, 40, 71], model pre-training [9, 33, 72] and self-training [24, 68, 74]. The third group focuses on improving network modules in long-tailed learning. This group includes representation learning [13, 46, 76], classifier learning [32, 41, 42, 69, 73], decoupled training [31, 32, 82], and ensemble learning [20, 83]. These methods introduce additional computation costs for increased performance. Our method belongs to the third group as it aims to learn a better representation of unbalanced training samples by the superclass graph, which is unexplored for long-tail recognition.", + "bbox": [ + 501, + 90, + 892, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Super-class learning. Super-class learning adds superclass labels as intermediate supervision into traditional deep learning. A super-class guided network [38] integrated the high-level semantic information into the network for image classification and object detection, which took two-level class annotations that contain both super-class and finer class labels. In [11], a two-phase multi-expert architecture was proposed for still image action recognition, which includes fine-grained and coarse-grained phases. However, they leveraged the ground truth of the super-class as supervision during the coarse-grained phase. Wu et al. [70] propose a taxonomic classifier to address the long-tail recognition problem, which classified each sample to the level that the classifier is competent. Zhou et al. [85] clustered the original categories into super-classes to produce a relatively balanced distribution in the super-class space, which also leveraged the ground truth of the super-class in the training phase. In contrast with the previous super-class learning, we do not use ground truth to group the original categories into the super-class space. To the best of our knowledge, no work exists that relies on graph learning to discover the super-class for long-tailed visual recognition, thus motivating this work.", + "bbox": [ + 503, + 398, + 892, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph neural networks. Recently, several graph neural network models (GNN) have been proposed to exploit the structures underlying graphs to benefit a variety of applications. There are two main research lines of GNN methods: non-spectral methods and spectral methods. The spectral methods [4, 10, 26, 35] focus on learning graph representations in a spectral domain, in which the learned filters are based on Laplacian matrices. The non-spectral methods [21, 63] develop an aggregator to aggregate a local set of features. Note that, message passing [19] is a key mechanism that allows GCNs and other graph neural networks to capture", + "bbox": [ + 503, + 734, + 892, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19945", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c833d2254310196ad13978c626cb615ac36571b82aa1f3734e163c9f4177c52e.jpg", + "image_caption": [ + "(a) Illustration of proposed SuperDisco" + ], + "image_footnote": [], + "bbox": [ + 158, + 89, + 480, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f91be0e202c4f80185112a6318b335b79fc5fcf749edabb6995ad2031be21b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 87, + 790, + 200 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a9511f57fd543c1a448cd45cd3fbc81b8339c52d175aad5c94022071d802859.jpg", + "image_caption": [ + "(c) Visualization of discovered super-class graph $c^2$", + "Figure 2. Illustration of proposed SuperDisco (a) and visualization of the similarity between the classes and discovered super-class at different levels (b), (c). In (a), we show two levels of super-class graphs $\\mathcal{C}^1$ and $\\mathcal{C}^2$ . The colour in each graph represents the discovered super-class. SuperDisco discovers the potential super-class at different levels hidden in each category from (b) and (c). $\\mathcal{C}^1$ roughly categorizes the original classes into three relatively balanced super-classes, and $\\mathcal{C}^2$ then finely categorize them into four more balanced super-classes." + ], + "image_footnote": [], + "bbox": [ + 535, + 208, + 767, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complex relationships and dependencies between nodes in a graph, and is a major reason why they have been successful in a variety of tasks involving graph-structured data. Our method belongs to the non-spectral methods, which leverage a GNN as the base architecture to discover the super-class representation. Our proposed super-class graph would refine and rectify the original imbalanced feature to a relatively balanced feature space, which has not been explored for long-tail recognition either.", + "bbox": [ + 75, + 421, + 470, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta-learning for the long-tail. Meta-learning or learning to learn [3, 53, 58, 80, 81], is a learning paradigm where a model is trained on the distribution of tasks so as to enable rapid learning on new tasks. Ren et al. [51] first proposed meta-learning for the long-tailed problem by reweighting training examples. Shu et al. [54] proposed Meta-weight-Net to adaptively extract sample weights to guarantee robust recognition in the presence of training data bias. Li et al. [39] introduced meta-semantic augmentation for long-tailed recognition, which produces diversified augmented samples by translating features along many semantically meaningful directions by meta-learning. Our uniqueness is that our model aims to discover an improved super-class representation by meta-learning, which enables the original feature representation to adjust its corresponding higher-level super-class space.", + "bbox": [ + 75, + 558, + 470, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Learning to discover the super-class graph", + "text_level": 1, + "bbox": [ + 76, + 814, + 455, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we discuss how to learn to discover the super-class graph from the training samples and then expand on how to leverage such a graph to benefit the unbalanced data by refining the feature representations of samples. The", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "overall illustration of SuperDisco and a visualization of the discovered super-class hierarchy are shown in Figure 2.", + "bbox": [ + 498, + 421, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. For long-tailed visual recognition, the goal is to learn an image classification model from an imbalanced training set and to evaluate the model on a balanced test set. We first define the notation for long-tailed recognition used throughout our paper. We define a training input $x_{k} \\in \\mathbb{R}, i \\in \\{1, \\dots, n\\}$ , where $n$ is the number of training samples in the dataset. The corresponding labels are $y_{k} \\in \\{1, 2, \\dots, C\\}$ where $C$ is the number of classes. Let $n_{j}$ denote the number of training samples for the class $j$ . Here, we assume that $n_{i} \\geq n_{j}$ when $i < j$ shows the long-tailed problem simply. In this work, we typically consider a deep network model with three main components: a feature extractor $f(\\cdot)$ , a proposed graph model $g(\\cdot)$ and a classifier $h(\\cdot)$ . The feature extractor $f(\\cdot)$ first extracts an image representation as $\\mathbf{z} = f(x; \\theta)$ , which is then fed into the proposed graph model to refine a new representation as $g(\\mathbf{z}; \\phi) = \\tilde{\\mathbf{z}}$ . The final class prediction $\\tilde{y}$ is given by a classifier function $h(\\cdot)$ , i.e., $\\tilde{y} = \\arg \\max h(\\tilde{\\mathbf{z}}; \\psi)$ . Before detailing our approach, we add Table 1 to detail the meaning of each symbol for easy lookup.", + "bbox": [ + 496, + 455, + 892, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SuperDisco. We construct the super-class graph to organize and distill knowledge from the training process. The vertices represent different types of super-classes (e.g., the common contour between birds and airplanes) and the edges are automatically constructed to reflect the relationship between different super-classes. Our super-class graph contains multiple levels, which is closer to the relationship between various objects in the real world. Before detailing the structure, we first explicate why the multi-level super-class graphs are preferred over a flat super-class graph: a single level of", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19946", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "super-class groups is likely insufficient to model complex task relationships in real-world applications; for example, the similarities among different bird species are high, but there are also similarities between birds and mammals, e.g., they are both animals.", + "bbox": [ + 75, + 90, + 470, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assume the vertex representation $g$ as $\\mathbf{h}^g\\in \\mathbb{R}^d$ and define the super-class graph as $\\mathcal{C}^l = (\\mathbf{H}_c^l,\\mathbf{A}_c^l)$ , where $\\mathbf{H}_c^l = \\{\\mathbf{h}^{i^l}|\\forall i^l\\in [1,C^l ]\\} \\in \\mathbb{R}^{C^l\\times d}$ is the vertex feature matrix of the $l$ -th super-class level and $\\mathbf{A}_c^l = \\{A_c^l (\\mathbf{h}^{i^l},\\mathbf{h}^{j^l})|\\forall i^l,j^l\\in [1,C^l ]\\} \\in \\mathbb{R}^{C^l\\times C^l}$ is the vertex adjacency matrix in the $l$ -th super-class level, $C^l$ denotes the number of vertices in the $l$ -th super-class level. Our vertex representation $\\mathbf{H}_c^l$ of the super-class graph is defined to get parameterized and learned during training. The initial vertex representations of each super-class level are randomly initialized, which encourages diversity of the discovered super-classes.", + "bbox": [ + 75, + 166, + 468, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we introduce how to compute the edge weight $A_{\\mathcal{C}}^{l}$ in the super-class graph. The edge weight $A_{\\mathcal{C}}^{l}(\\mathbf{h}^{i^{l}}, \\mathbf{h}^{j^{l}})$ between a pair of vertices $i$ and $j$ is gauged by the similarity between them. Formally:", + "bbox": [ + 75, + 335, + 468, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\mathcal {C}} ^ {l} \\left(\\mathbf {h} ^ {i l}, \\mathbf {h} ^ {j l}\\right) = \\sigma \\left(\\mathbf {W} _ {c} ^ {l} \\left(| \\mathbf {h} ^ {i l} - \\mathbf {h} ^ {j l} | / \\gamma_ {c} ^ {l}\\right) + \\mathbf {b} _ {c} ^ {l}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 402, + 468, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_c^l$ and $\\mathbf{b}_c^l$ indicate learnable parameters of the $l$ -th super-class level, $\\gamma_c^l$ of $l$ -th super-class level is a scalar and $\\sigma$ indicates the Sigmoid function, which normalizes the weight between 0 and 1. To adjust the representation of training samples by the involvement of super-classes, we first query the training samples in the super-class graph to obtain the relevant super-class. In light of this, we construct a new graph $\\mathcal{R}$ , which adds the original sample feature $\\mathbf{z}$ to the super-class graph. We define $\\mathbf{z}^l$ as the refined feature after the $l$ -th super-class graph. Here we define graph $\\mathcal{R}^l = (\\mathbf{H}_{\\mathcal{R}}^l, \\mathbf{A}_{\\mathcal{R}}^l)$ , where $\\mathbf{H}_{\\mathcal{R}}^l = \\{[\\mathbf{z}^l, \\mathbf{h}^{i^l}] | \\forall i^l \\in [1, C^l]\\} \\in \\mathbb{R}^{(C^l + 1) \\times d}$ denotes the vertex feature matrix of the $l$ -th super-class level, and $\\mathbf{A}_{\\mathcal{R}}^l = \\{[A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{z}^1), A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{h}^{j^l})] | \\forall i^l, j^l \\in [1, C^l]\\} \\in \\mathbb{R}^{C^l + 1 \\times C^{l + 1}}$ denotes the vertex adjacency matrix in the $l$ -th super-class level. The link between $\\mathbf{z}^l$ and vertex $\\mathbf{h}^i$ in the hierarchical graph is constructed by their similarity. In particular, analogous to the definition of weight in the super-class graph in Eq. (1), the weight $A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{z}^1)$ is constructed as:", + "bbox": [ + 75, + 424, + 470, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\mathcal {R}} ^ {l} \\left(\\mathbf {h} ^ {i ^ {l}}, \\mathbf {z} ^ {1}\\right) = \\sigma \\left(\\mathbf {W} _ {r} ^ {l} \\left(\\left| \\mathbf {h} ^ {i ^ {l}} - \\mathbf {z} ^ {1} \\right| / \\gamma_ {r} ^ {l}\\right) + \\mathbf {b} _ {r} ^ {l}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 715, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_r^l$ and $\\mathbf{b}_c^r$ indicate learnable parameters of the $l$ -th super-class level, $\\gamma_r^l$ of the $l$ -th super-class level is a scalar.", + "bbox": [ + 75, + 738, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After constructing the new graph $\\mathcal{R}$ , we propagate the most relevant super-class by message passing [19] from the discovered super-classes $\\mathcal{C}$ to the features $\\mathbf{z}^l$ by introducing a Graph Neural Network (GNN). The message passing operation over the graph is formulated as:", + "bbox": [ + 75, + 768, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {\\mathcal {R}} ^ {(m + 1)} = \\mathrm {M P} \\left(\\mathbf {A} _ {\\mathcal {R}} ^ {l}, \\mathbf {H} _ {\\mathcal {R}} ^ {(m)}; \\mathbf {W} ^ {(m)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 848, + 468, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{M}\\mathbb{P}(\\cdot)$ is the message passing function, $\\mathbf{H}^{(m)}$ is the vertex embedding after $m$ layers of GNN and $\\mathbf{W}^{(m)}$ is a", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2b60d4353fb0b6097f8a3ea65021dcc6cb81cbe9d234e61c4374e8f462601afe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NotationDescription
hVertex representation
zOriginal sample feature
CSuper-class graph
HcVertex feature matrix of Cl
AcVertex adjacency matrix of Cl
RGraph which adds z to graph Cl
PPrototype graph
CpVertex feature matrix of P
ApVertex adjacency matrix of P
SSuper graph which connecting P and C
AVertex feature matrix of S
MVertex adjacency matrix of S
", + "bbox": [ + 553, + 88, + 836, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Summary of the core notation used for SuperDisco.", + "bbox": [ + 514, + 250, + 872, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learnable weight matrix of layer $m$ . After stacking $M$ GNN layers, we get the information-propagated feature representation $\\tilde{\\mathbf{z}}^L$ for each level of the super-class graph $\\mathcal{C}$ . Once we obtain the refined representation $\\tilde{\\mathbf{z}}^L$ for a training sample by the super-class graph, we feed them into the classifier to make the predictions and compute the corresponding loss, i.e., Cross-entropy loss for optimization. Using gradient descent, we then update the super-class graph $\\mathcal{C}$ . To be able to discover a more accurate super-class graph in the face of severe imbalance problems, we propose meta-learning super-class graph discovery in the next section.", + "bbox": [ + 496, + 268, + 893, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Meta-learning super-class graph discovery", + "text_level": 1, + "bbox": [ + 498, + 450, + 879, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explore and exploit a more accurate and richer superclass graph, we propose the Meta-SuperDisco to discover the super-class graph using meta-learning, making the model more robust. In the traditional meta-learning setting [17, 50], it includes meta-training tasks and meta-test task. Each task contains a support set $S$ and a query set $Q$ . Each task is first trained by $S$ to get the task-specific learner and $Q$ optimizes this learner to update the meta-learner. For long-tailed recognition with meta-learning, previous works [51, 54] randomly sample a small amount of balanced data denoted as $\\mathcal{M}$ . The imbalanced data and the small balanced data can be seen as $S$ and $Q$ in the training phase. The goal of meta-learning for long-tailed recognition is to use a small set of balanced data to optimize the model obtained from unbalanced data. We follow [51, 54] by randomly selecting the same number of samples (e.g., ten) as $\\mathcal{M}$ per class from the training set.", + "bbox": [ + 496, + 474, + 892, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meta-SuperDisco. To meta-learn the super-class graph, we construct a prototype graph $\\mathcal{P}$ from $\\mathcal{M}$ , since $\\mathcal{M}$ is a balanced dataset. The prototype graph extracts the sample-level relation information, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. For the prototype graph, we need to compute the prototype of each category [55], which is defined as: $\\mathbf{c}^k = \\frac{1}{N^k}\\sum_{i = 1}^{N^k}\\mathbf{z}_j$ , where $N^k$ denotes the number of samples in class $k$ , $\\mathbf{z}_j$ is the feature from $\\mathcal{M}$ .", + "bbox": [ + 496, + 717, + 893, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After calculating all prototype representations $\\{\\mathbf{c}^k |\\forall k\\in [1,K]\\}$ , which serve as the vertices in the prototype graph $\\mathcal{P}_i$ , we further need to define the edges and the corresponding", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edge weights. The edge weight $A_{\\mathcal{P}}(\\mathbf{c}^i,\\mathbf{c}^j)$ between two prototypes $\\mathbf{c}^i$ and $\\mathbf{c}^j$ is gauged by the similarity between them. The edge weight is calculated as follows:", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\mathcal {P}} \\left(\\mathbf {c} ^ {i}, \\mathbf {c} ^ {j}\\right) = \\sigma \\left(\\mathbf {W} _ {p} \\left(\\left| \\mathbf {c} ^ {i} - \\mathbf {c} ^ {j} \\right| / \\gamma_ {p}\\right) + \\mathbf {b} _ {p}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 140, + 470, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_p$ and $\\mathbf{b}_p$ are the learnable parameters, $\\gamma_p$ is a scalar. Thus, we denote the prototype graph as $\\mathcal{P} = (\\mathbf{C}_{\\mathcal{P}}, \\mathbf{A}_{\\mathcal{P}})$ , where $\\mathbf{C}_{\\mathcal{P}_i} = \\{\\mathbf{c}^i | \\forall i \\in [1, K]\\} \\in \\mathbb{R}^{K \\times d}$ represent a set of vertices, with each one corresponding to the prototype from a class, while $\\mathbf{A}_{\\mathcal{P}} = \\{|A_{\\mathcal{P}}(\\mathbf{c}^i, \\mathbf{c}^j)| \\forall i, j \\in [1, K]\\} \\in \\mathbb{R}^{K \\times K}$ gives the adjacency matrix, which indicates the proximity between prototypes. We then use the prototype graph to guide the learning of the meta super-class graph. We construct a super graph $\\mathcal{S}$ by connecting prototype graph $\\mathcal{P}$ to super-class graph $\\mathcal{C}$ . In the super graph $\\mathcal{S}$ , the vertices are $\\mathbf{M} = (\\mathbf{C}_{\\mathcal{P}^l}^l; \\mathbf{H}_{\\mathcal{C}^l}^l)$ , where $\\mathcal{P}^l$ denotes the refined prototype graph vertex after the $l$ -th level super-class graph. Then, we calculate the link weight $A_{\\mathcal{S}}^l(c^i, \\{\\mathbf{h}^j\\})$ of the super graph as:", + "bbox": [ + 75, + 162, + 472, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {S} ^ {l} \\left(\\mathbf {c} ^ {i}, \\mathbf {h} ^ {j ^ {l}}\\right) = \\frac {\\exp \\left(- \\| \\left(\\mathbf {c} ^ {i} - \\mathbf {h} ^ {j ^ {l}}\\right) / \\gamma_ {s} ^ {l} \\| _ {2} ^ {2} / 2\\right)}{\\sum_ {j ^ {l ^ {\\prime}} = 1} ^ {J} \\exp \\left(- \\| \\left(\\mathbf {c} ^ {i} - \\mathbf {h} ^ {j ^ {l ^ {\\prime}}}\\right) / \\gamma_ {s} ^ {l} \\| _ {2} ^ {2} / 2\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 364, + 468, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_s^l$ is a scaling factor. Note that, here we use softmax to ensure that the total weight of edges between the prototype graph $\\mathcal{P}$ and the super-class graph $\\mathcal{C}$ is equal to 1, giving the prototype graph a unique influence on the expression of each super-class. Thus, the adjacent matrix and feature matrix of the super graph $S^l = (\\mathbf{A}^l,\\mathbf{M}^l)$ is defined as $\\mathbf{A}^l = (\\mathbf{A}_{\\mathcal{P}},\\mathbf{A}_S^l;\\mathbf{A}_S^{l^\\mathrm{T}},\\mathbf{A}_\\mathcal{C}^l)$ and $\\mathbf{M}^l = (\\mathbf{C}_{\\mathcal{P}^l}^l;\\mathbf{H}_{\\mathcal{C}^l}^l)$ . Once we constructed the super graph $S$ , we use message-passing again to propagate the most relevant knowledge from the prototype graph $\\mathcal{P}$ to the super-class graph $\\mathcal{C}$ . Similar to eq. (3):", + "bbox": [ + 75, + 410, + 470, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {(m + 1)} = \\mathbb {M P} \\left(\\mathbf {A} ^ {l}, \\mathbf {M} ^ {(m)}; \\mathbf {W} ^ {(m)}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 566, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We leverage the graph $S$ to refine the super-class graph. Finally, we feed the original feature $\\mathbf{z}$ into the super-class graph to get the information-propagated feature representation $\\tilde{\\mathbf{z}}^L$ , which refines the original feature by its corresponding discovered super-classes. We provide the complete SuperDisco and Meta-SuperDisco algorithm specifications in the supplemental material.", + "bbox": [ + 75, + 589, + 470, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 709, + 209, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We apply our method to four commonly used long-tail recognition benchmarks. Sample images and the number of categories for all datasets are provided in the supplement material. CIFAR-100-LT reduces the number of training samples per class according to an exponential function $n = n_i \\mu^i$ , where $i$ is the class index, $n_i$ is the original number of training samples, and $\\mu \\in (0, 1)$ . The imbalance factor of a dataset is defined as the number of training samples in the most populated class divided by the minority class. We consider imbalance factors {10, 50, 100}. ImageNet-LT [44] is a subset of ImageNet [12] consisting of", + "bbox": [ + 75, + 734, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a5fbe34349fdfe1a8d5eab768f3c46c286c64ccb8a17c534908fa27e94119147.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
SuperDisco65.960.757.250.945.2
Meta-SuperDisco68.563.158.353.847.5
", + "bbox": [ + 542, + 88, + 846, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Benefit of SuperDisco and Meta-SuperDisco. SuperDisco achieves better performance compared to a baseline fin-tuning on all imbalance factors, while Meta-SuperDisco is even better for long-tailed recognition.", + "bbox": [ + 496, + 176, + 893, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "115.8K images from 1000 categories, with maximally 1,280 images per class and minimally 5 images per class, and a balanced test set. Places-LT [44] has an imbalanced training set with 62,500 images for 365 classes from Places [84]. It contains images from 365 classes and the number of images per class ranges from 4980 to 5. The test sets are balanced and contain 100 images per class. iNaturalist [61] is a real-world long-tailed dataset with 675,170 training images for 5,089 classes, where the top $1\\%$ most populated classes contain more than $16\\%$ of the training images. Additionally, there is also a severe imbalance among the super-classes of iNaturalist. The 13 ground truth super-classes images range from 158,407 to 308.", + "bbox": [ + 496, + 243, + 893, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We follow [32] by first training a feature extractor with instance-balanced sampling, and then training our graph model and classifier based on the trained features. For CIFAR-100-LT, we follow [54] and use a ResNet-32 backbone. For ImageNet-LT, we use ResNeXt-50 [23] as our backbone, following [32]. For Places-LT, we report results with ResNet-152 following [44]. For iNaturalist, we use a ResNet-50 backbone. We train each dataset for 200 epochs with batch size 512. We use random left-right flipping and cropping as our training augmentation. For all experiments, we use an SGD optimizer with a momentum of 0.9 and a batch size of 512. We randomly selected 10 images per class from the training set for all datasets as $\\mathcal{M}$ . Code available at: https://github.com/Yingjun-Du/SuperDisco.", + "bbox": [ + 496, + 444, + 895, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benefit of SuperDisco and Meta-SuperDisco. To show the benefit of SuperDisco, we compare it with a fine-tuning baseline, which retrans the classifier only. Table 2 shows SuperDisco improves over fine-tuning on CIFAR-100-LT, and the results for the other long-tailed datasets are provided in the supplemental materials Table 1. In the most challenging setting with the largest imbalance factor of 200, our SuperDisco delivers $45.2\\%$ , surpassing the baseline by $5.9\\%$ . We attribute improvement to our model's ability to refine original features, allowing the discovered super-class graph to guide the tail features away from the dominant role of head features, thus leading to improvements over the original features. We also investigate the benefit of meta-learning with Meta-SuperDisco. The Meta-SuperDisco consistently surpasses the SuperDisco for all imbalance factors. The con", + "bbox": [ + 496, + 674, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19948", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c5cc9c860ce870541f92e9265c2dd556fce6faa4b4060297e47137541a15a4bc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
(20)61.260.149.947.341.9
(2, 4, 8)65.362.753.149.843.2
(4, 8, 16)69.164.255.252.345.9
(4, 8, 16, 32)68.563.158.353.847.5
(4, 8, 16, 32, 64)66.962.758.952.946.3
Oracle super-classes66.963.254.751.443.2
", + "bbox": [ + 89, + 88, + 450, + 241 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d5025da6338e812cbd109bd740ca354b1fa8113cee9a3788244dfb9f7c6c365.jpg", + "table_caption": [ + "Table 3. Effect of number of super-class levels on CIFAR-100-LT. Compared to a baseline [32] and an oracle setting, Meta-SuperDisco provides higher performance gains with more complex hierarchies." + ], + "table_footnote": [], + "table_body": "
ManyMediumFewAll
Baseline65.066.365.565.9
(13)71.870.266.170.8
(2,4,8)70.569.365.969.4
(4,8,16)72.270.966.470.3
(4,8,16,32)73.670.267.370.9
(4,8,16,32,64)73.472.968.372.3
(4,8,16,32,64,128)72.171.366.270.9
Oracle super-classes70.770.565.970.2
", + "bbox": [ + 112, + 292, + 426, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Effect of number of super-class levels on iNaturalist. Meta-SuperDisco achieves consistent performance gains with more complex hierarchies.", + "bbox": [ + 76, + 426, + 470, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sistent improvements confirm that Meta-SuperDisco learns even more robust super-class graphs, leading to a discriminative representation of the tail data.", + "bbox": [ + 76, + 474, + 468, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of the number of super-class levels. A significant challenge with any structure-aware learning algorithm is determining the appropriate complexity for the knowledge structure. So, we further analyze the effect of the super-class hierarchies, including the level (number of depths L) or the number of super-classes in each level. The results are shown in Table 3 and Table 4. The super-class number from the bottom layer to the top layer is saved in a tuple. For example, (2, 4, 8) represents three depth, with two super-classes in the top layer. The baseline is Decouple-LWS [32], which only inputs the original feature to learn a new classifier. The oracle super-classes are first trained on two long-tailed datasets using the ground truth super-class labels for super-class classifications. Once the training is completed, each oracle super-class is obtained by averaging the samples of each super-class. We constructed a one-layer super-class graph using these super-classes, where the vertices of the graph are for each super-class, and the edges of the graph are computed according to Eq. (1). Then, we use the message passing by Eq. (3) to refine the original features and input them into the classifier to get the final predictions. From Table 3, we observe that using oracle super-classes achieves better performance compared to the learned super-class (20) since it uses the ground truth super-classes as supervision. We also conclude that too few levels may not be enough", + "bbox": [ + 76, + 523, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad75c3bb258313c39fce850e20694d6365887bcf6881cc9730db255e0f253392.jpg", + "image_caption": [ + "(a) SuperDisco" + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 669, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40d1be7db1d59e8e1c62c8fe9d07dabbf7f39b6a9a571b7be13daaabdb088a17.jpg", + "image_caption": [ + "(b) Meta-SuperDisco", + "Figure 3. Similarity between discovered super-classes and classes. SuperDisco discovers super-classes hidden in each class, while Meta-SuperDisco discovers more accurate super-classes." + ], + "image_footnote": [], + "bbox": [ + 669, + 88, + 880, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to learn the precise super-classes (e.g., tuple (20) or (2, 4, 8)). In this dataset, increasing levels (e.g., tuple (4, 8, 16, 32)) achieves better performance on the smaller imbalance factor (e.g., 10), and similar performance compared with (4, 8, 16). For the real-world long-tailed dataset iNaturalist [62] in Table 4, we find no significant improvement for the few-shot classes in the performance of the oracle super-class compared to the baseline, and the same is true for the performance of the discovered super-class structure (13). This is because the super-class of iNaturalist also have serious long-tailed problems, resulting in the refined features of tail classes remaining indistinguishable from the refined features of head classes. However, with a more complex graph structure (4, 8, 16, 32, 64), the few-shot performance improves by a good margin compared with the baseline, and even the oracle super-classes. We attribute this to our model's ability to explore relatively balanced super-class spaces, thus making the refined tail category features discriminative. By comparing Table 3 and Table 4, we conclude that deeper as well as wider graphs are needed to discover the super-classes in the case of severe class imbalance.", + "bbox": [ + 496, + 219, + 893, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization of SuperDisco. To understand the meaning of the discovered super-classes more clearly, we present a visualization in Figure. 3. We selected 12 different categories from the CIFAR-100 test dataset. We calculate the similarity of each of these 12 categories to the different vertices in the graph we explore. Here we show the similarity with the second layer of graph vertices (C1, C2, C3, C4). We can see different categories mainly activate different vertices, e.g., bus $\\rightarrow$ C3 and road $\\rightarrow$ C2. As shown in this heatmap, we find that C1 reflects the super-class of flowers, C2 reflects the super-class of buildings, C3 reflects the super-class of vehicles, C4 reflects the super-class of fish. Another observation is that the second-largest activated super-class is also meaningful, promoting knowledge transfer between super-classes. For example, road and bridge are related to the C3 superclass, since some vehicles may be on the road and bridge. This visualization reflects that we can use graph models to discover the super-classes and the relationships between each super-class. We also visualize the discovered meta-learning super-classes in Figure 3 (b). The discovered super-classes are even more accurate, e.g., roses have high similarity to C2, which mainly reflects the buildings super-class, while it has high similarity to C1, which is the flowers super-class. This once again validates the benefit of Meta-SuperDisco.", + "bbox": [ + 496, + 537, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19949", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86d685881b3449474338255a4a8358b62b9a0b0cc79def00f46464450880b6b9.jpg", + "image_caption": [ + "(a) Original features" + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 297, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af99ecc2e2d54628f761f3c66477bea79fcd1fef79a2ee50ae03a11ea59bd622.jpg", + "image_caption": [ + "(c) Refined features by SuperDisco" + ], + "image_footnote": [], + "bbox": [ + 315, + 89, + 473, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6020ca1af763ca7a9554ee400ba03cfdc5304c11cb1dbd37ac8e6f616fc3beb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 90, + 651, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/198d2ca7405cd4c2eb306a90c989b2a274cebb29f0a75a7f2c26cc031467bf12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 90, + 828, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e90b07d166fe11b6d1af84e15ae336f85caad00ce4b7b45b5f995dafdea392f.jpg", + "image_caption": [ + "(b) Training distribution" + ], + "image_footnote": [], + "bbox": [ + 145, + 209, + 292, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8ae94c57d2f09d1dab25a8d230eb70e3f03f8467354aec06464f7efca0f094f.jpg", + "image_caption": [ + "(d) Refined features by Meta-SuperDisco" + ], + "image_footnote": [], + "bbox": [ + 308, + 209, + 467, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8afc546fa9283e628ad771f991a32aee87eaa912a683b333d8fe5816e712f0c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 209, + 650, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b06e08889990f3daf85c823d38d35ef5d74c93d14d47808e3fcce43d97b6976.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 209, + 828, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0a0492f9a18cb8a7a45dd84b0102d813c0e4e1a9757b0ee80f79f430448c340.jpg", + "image_caption": [ + "Figure 4. Visualization of refined features on CIFAR-100-LT, with the original features (a) and their corresponding training distribution (b). Colours indicate categories. SuperDisco (c) guides the original features on being clustered into the corresponding super-class space at different levels, while Meta-SuperDisco (d) obtains even more discriminative intra-class features.", + "Figure 5. Effect of refined features. Accuracy increases along with the increased super-class levels, revealing that more accurate and richer super-classes facilitate better long-tailed recognition." + ], + "image_footnote": [], + "bbox": [ + 156, + 387, + 390, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, in the (c) learned the hierarchical concept of each class, we can see that bus and bridge have the same concept C2 in the last concept level, which may be due to the possible presence of cars on the bridge.", + "bbox": [ + 75, + 582, + 468, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of refined features. To understand the empirical benefit of SuperDisco, we visualize in Figure 4 the original features and refined features with super-class graphs of the different levels using t-SNE [60]. We choose the vertices numbers as (2, 4, 6), meaning the super-class graph has three different levels, each with a different number of vertices. The original features of the category with a small sample size will overlap with the (original) features of the category with a large sample size. Super-class graphs discovered by our model project the original features into a high-level super-class space, where the imbalance is relatively small. Hence, its corresponding subcategory can be predicted more accurately. It is worth noting that when comparing the two different super-class graphs on top and below, the features obtained by Meta-SuperDisco are even more distinctive and distant from each other. To better measure the goodness of the refined features obtained at different levels,", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2cdd2765b8154a2e03d71f9bec1892ca06ad79d5e97fcacfae9064fffbed972.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ManyMediumFewAll
Baseline58.449.334.852.7
Multi-layer perceptron63.551.835.955.0
Graph convolution network66.153.337.157.1
", + "bbox": [ + 519, + 378, + 864, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Analysis of super-class mechanism on ImageNet-LT. The super-class mechanism contributes most, the graph convolution network improves results further.", + "bbox": [ + 500, + 446, + 890, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we show in Figure 5 the prediction accuracy using different refined features. We find that the accuracy increases along with the increased super-class levels, which shows that using more accurate and richer super-classes facilitates better performance. This again demonstrates that Meta-SuperDisco is most suitable for long-tailed visual recognition.", + "bbox": [ + 498, + 503, + 893, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of super-class mechanism. To demonstrate that the improved performance of our SuperDisco cannot solely be attributed to the graph convolutional network module, we conducted an experiment where we replaced it with a multi-layer perceptron to obtain the representation per sample. In Table 5, the performance gains of our method are primarily due to the super-classes rather than the graph convolution network. The results suggest that incorporating the super-classes mechanism plays a crucial role in improving the performance of long-tailed problems. Furthermore, the results improve further when we replace the multi-layer perceptron with our graph convolution network module.", + "bbox": [ + 496, + 597, + 893, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with the state-of-the-art. We evaluate our method on the four long-tailed datasets under different imbalance factors in Table 6 and 7. Our model achieves state-of-the-art performance on the tail data of all datasets. For ImageNet-LT, our model achieves state-of-the-art performance on both few-shot and all data. In the most challenging Places-LT, our model delivers $40.3\\%$ on all classes, surpassing the second-best Parisot et al. [47] by $1.1\\%$ . On the real-", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19950", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8b18dd2c1f55b54bb837eb29ea136002029a4f0010d93ef3e1eb7289acc4078.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VenueImageNet-LTPlaces-LTiNaturalist
ManyMediumFewAllManyMediumFewAllManyMediumFewAll
Kang et al. [32]ICLR 1960.247.230.349.940.639.128.637.665.066.365.565.9
Kang et al. [31]ICLR 2161.849.430.951.5-------68.6
He et al. [25]ICCV 2164.150.431.553.1----70.670.167.669.1
Li et al. [40]CVPR 2166.851.135.456.0-------69.3
Samuel et al. [52]ICCV 2164.049.833.153.5-------69.7
Alshammari et al. [1]CVPR 2262.550.441.553.9----71.270.469.770.2
Zhang et al. [75]CVPR 2161.352.231.452.940.442.430.139.369.071.170.270.6
Parisot et al. [47]CVPR 2263.252.136.954.139.741.034.939.2----
Park et al. [48]CVPR 2266.453.935.656.2----73.172.668.772.8
This paper66.153.337.157.145.342.835.340.372.372.971.373.6
", + "bbox": [ + 135, + 88, + 834, + 242 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/55c923aaba8eae561a17b9f199dc021010334bb22be6445abf07b9b81cead53b.jpg", + "table_caption": [ + "Table 6. Comparison with the state-of-the-art on ImageNet-LT, Places-LT and iNaturalist. Best and second best results are highlighted in bold and italic bold. Our Meta-SuperDisco achieves either better or comparable performance than state-of-the-art methods under the tail and all data for long-tailed visual recognition." + ], + "table_footnote": [], + "table_body": "
VenueImbalance ratio
1050100
Park et al. [49]ICCV 2159.547.442.0
Li et al. [40]CVPR 2162.350.546.0
Zhong et al. [82]CVPR 2162.551.546.8
Samuel et al. [52]ICCV 2163.457.647.3
Wang et al. [66]ICLR 2161.851.748.0
Zhu et al. [87]CVPR 2264.956.651.9
Cui et al. [7]ICCV 2164.256.052.0
Alshammari et al. [1]CVPR 2268.857.753.3
This paper69.358.353.8
", + "bbox": [ + 125, + 301, + 419, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Comparison with the state-of-the-art on CIFAR-100-LT. Our model achieves best performance.", + "bbox": [ + 76, + 459, + 470, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "world long-tailed dataset iNaturalist, our model achieves the three best performances under four different shots. On the long-tailed synthetic dataset CIFAR-100-LT, our model achieves the best performance under each imbalance factor. The consistent improvements on all benchmarks under various configurations confirm that our Meta-SuperDisco is effective for long-tailed visual recognition.", + "bbox": [ + 75, + 512, + 470, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We show that SuperDisco and Meta-SuperDisco achieve good performance on tail data while being less successful on the head data. Based on this result, we also perform an experiment on balanced CIFAR-100 in Figure 6. With SuperDisco and Meta-SuperDisco, there is only a slight change in performance at the expense of an increased inference time. This reveals that our SuperDisco does not change the original features much through message passing on a balanced dataset. This may be because the obtained super-classes are still the original class itself. In addition, as the computation of graphs involves many matrix operations, our model also requires a relatively long computational speed. Due to introducing a prototype graph and more data, Meta-SuperDisco takes longer to compute. In addition, the training time of SuperDisco and its meta variants is also 1.5 times higher than the baseline. Future work could investigate how to use the discovered super-class graph in balanced datasets and how to reduce the computation time.", + "bbox": [ + 75, + 628, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb1ad1ce828239038659c4f2b6829fc7fb3aaff56e41c134c834b74c4ab4b136.jpg", + "image_caption": [ + "Figure 6. Limitation. Accuracy $(\\%)$ vs. speed (ms) comparison with different methods on balanced CIFAR-100. SuperDisco has little impact on the performance of balanced datasets at the expense of increased inference time." + ], + "image_footnote": [], + "bbox": [ + 560, + 315, + 823, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 537, + 627, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes learning to discover a super-class graph for long-tailed visual recognition. The proposed superclass graph could rectify and refine the original features by message passing, which results in attending to the most relevant entities based on their semantic similarity between concepts for more accurate predictions. To obtain a more informative super-class graph and more balanced image representations, we further propose to meta-learn the super-class graph based on the prototype graph from a small amount of imbalanced data. We conduct thorough ablation studies to demonstrate the effectiveness of the proposed SuperDisco and Meta-SuperDisco algorithms. The state-of-the-art performance on the long-tailed version of four datasets further substantiates the benefit of our proposal.", + "bbox": [ + 496, + 565, + 893, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This work is financially supported by the Inception Institute of Artificial Intelligence, the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy.", + "bbox": [ + 496, + 809, + 893, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, pages 6897-6907, 2022. 8", + "[2] Chris Anderson. The long tail: How endless choice is creating unlimited demand. Random House, 2007. 1", + "[3] Yoshua Bengio, Samy Bengio, and Jocelyn Cloutier. Learning a synaptic learning rule. In IJCNN, 1991. 3", + "[4] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. 2", + "[5] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. In Nature Machine Intelligence, 2020. 1", + "[6] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 2", + "[7] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In ICCV, 2021. 8", + "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 2", + "[9] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, 2018. 2", + "[10] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 2", + "[11] Hojat Asgarian Dehkordi, Ali Soltani Nezhad, Hossein Kashiani, Shahriar Baradaran Shokouhi, and Ahmad Aytollahi. Multi-expert human action recognition with hierarchical super-class learning. Knowledge-Based Systems, page 109091, 2022. 2", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5", + "[13] Qi Dong, Shaogang Gong, and Xiatian Zhu. Class rectification hard mining for imbalanced deep learning. In ICCV, 2017. 2", + "[14] Charles Elkan. The foundations of cost-sensitive learning. In IJCAI, 2001. 2", + "[15] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Computational Intelligence, 20(1):18-36, 2004. 1", + "[16] Adam R Ferguson, Jessica L Nielson, Melissa H Cragin, Anita E Bandrowski, and Maryann E Martone. Big data from small data: data-sharing in the 'long tail' of neuroscience. Nature Neuroscience, 17(11):1442-1447, 2014. 1", + "[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICLR, pages 1126–1135, 2017. 4", + "[18] Camille Garcin, Alexis Joly, Pierre Bonnet, Jean-Christophe Lombardo, Antoine Affouard, Mathias Chouet, Maximilien Servajean, Joseph Salmon, and Titouan Lorieul. P1@ ntnet-300k: a plant image dataset with high label ambiguity and a long-tailed distribution. In NeurIPS 2021-35th Conference on Neural Information Processing Systems, 2021. 1" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICLR, pages 1263-1272, 2017. 2, 4", + "[20] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In CVPR, pages 15089-15098, 2021. 2", + "[21] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, 2017. 2", + "[22] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderlinesmote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878-887, 2005. 1, 2", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[24] Ruifei He, Jihan Yang, and Xiaojun Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In ICCV, 2021. 2", + "[25] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In ICCV, 2021. 8", + "[26] Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163, 2015. 2", + "[27] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In CVPR, 2021. 2", + "[28] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In CVPR, 2020. 2", + "[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, 2016. 1", + "[30] Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In CVPR, pages 7610-7619, 2020. 2", + "[31] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In ICLR, 2021. 1, 2, 8", + "[32] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 5, 6, 8", + "[33] Shyamgopal Karthik, Jérôme Revaud, and Chidrovskii Boris. Learning from long-tailed data with noisy labels. arXiv:2108.11096, 2021. 2", + "[34] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In CVPR, 2020. 1", + "[35] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2", + "[36] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICLR, 2020. 1" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19952", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Bo Li, Yongqiang Yao, Jingru Tan, Gang Zhang, Fengwei Yu, Jianwei Lu, and Ye Luo. Equalized focal loss for dense long-tailed object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6990-6999, 2022. 2", + "[38] Kaidong Li, Nina Y Wang, Yiju Yang, and Guanghui Wang. Sgnet: A super-class guided network for image classification and object detection. In Conference on Robots and Vision, pages 127-134, 2021. 2", + "[39] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In CVPR, pages 5212-5221, 2021. 3", + "[40] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In CVPR, pages 630-639, 2021. 2, 8", + "[41] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In ICCV, 2021. 2", + "[42] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, 2020. 2", + "[43] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, 39(2):539-550, 2008. 1, 2", + "[44] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, pages 2537-2546, 2019. 1, 2, 5", + "[45] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR, 2021. 2", + "[46] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2", + "[47] Sarah Parisot, Pedro M Esperança, Steven McDonagh, Tamas J Madarasz, Yongxin Yang, and Zhenguo Li. Long-tail recognition via compositional knowledge transfer. In CVPR, pages 6939-6948, 2022. 7, 8", + "[48] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 8", + "[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In ICCV, pages 735-744, 2021. 8", + "[50] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 4", + "[51] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 1, 3, 4", + "[52] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In ICCV, 2021. 8" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3", + "[54] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, 2019. 1, 3, 4, 5", + "[55] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 4", + "[56] Yanmin Sun, Mohamed S Kamel, Andrew KC Wong, and Yang Wang. Cost-sensitive boosting for classification of imbalanced data. Pattern Recognition, 40(12):3358-3378, 2007. 2", + "[57] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In NeurlPS, 2020. 2", + "[58] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, USA, 1998. 3", + "[59] Junjiao Tian, Yen-Cheng Liu, Nathan Glaser, Yen-Chang Hsu, and Zsolt Kira. Posterior re-calibration for imbalanced datasets. In NeurlPS, 2020. 2", + "[60] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(11), 2008. 7", + "[61] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5", + "[62] Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450, 2017. 6", + "[63] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 2", + "[64] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3784-3793, 2021. 2", + "[65] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is in classification: A simple framework for long-tail instance segmentation. In ECCV, 2020. 1", + "[66] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2021. 8", + "[67] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2", + "[68] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In CVPR, 2021. 2", + "[69] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In CVPR, pages 8659-8668, 2021. 2", + "[70] Tz-Ying Wu, Pedro Morgado, Pei Wang, Chih-Hui Ho, and Nuno Vasconcelos. Solving long-tailed recognition with deep" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19953", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "realistic taxonomic classifier. In ECCV, pages 171-189, 2020. 2", + "[71] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, pages 247-263, 2020. 2", + "[72] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In NeurIPS, 2020. 2", + "[73] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5704-5713, 2019. 2", + "[74] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2", + "[75] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In CVPR, pages 2361-2370, 2021. 2, 8", + "[76] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In ICCV, 2017. 1, 2", + "[77] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1", + "[78] Yifan Zhang, Peilin Zhao, Jiezhang Cao, Wenye Ma, Junzhou Huang, Qingyao Wu, and Mingkui Tan. Online adaptive asymmetric active learning for budgeted imbalanced data. In KDD, pages 2768-2777, 2018. 2", + "[79] Peilin Zhao, Yifan Zhang, Min Wu, Steven CH Hoi, Mingkui Tan, and Junzhou Huang. Adaptive cost-sensitive online classification. IEEE Transactions on Knowledge and Data Engineering, 31(2):214-228, 2018. 2", + "[80] Xiantong Zhen, Yingjun Du, Huan Xiong, Qiang Qiu, Cees Snoek, and Ling Shao. Learning to learn variational semantic memory. In NeurIPS, 2020. 3", + "[81] Xiantong Zhen, Haoliang Sun, Yingjun Du, Jun Xu, Yilong Yin, Ling Shao, and Cees Snoek. Learning to learn kernels with variational random features. In ICML, 2020. 3", + "[82] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 8", + "[83] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In CVPR, pages 9719–9728, 2020. 2", + "[84] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5", + "[85] Yucan Zhou, Qinghua Hu, and Yu Wang. Deep super-class learning for long-tail distributed image classification. Pattern Recognition, 80:118-128, 2018. 2", + "[86] ZhiHua Zhou and XuYing Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on Knowledge and Data Engineering, 18(1):63-77, 2005. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[87] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In CVPR, pages 6908-6917, 2022. 8", + "bbox": [ + 501, + 90, + 893, + 146 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19954", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_model.json b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e0ab3c5686213f180a4cd2fbfd1029ceac3bde55 --- /dev/null +++ b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "SuperDisco: Super-Class Discovery Improves Visual Recognition for the Long-Tail" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.18, + 0.806, + 0.217 + ], + "angle": 0, + "content": "Yingjun Du1, Jiayi Shen1, Xiantong Zhen1,2*, Cees G. M. Snoek1 \n1AIM Lab, University of Amsterdam 2Inception Institute of Artificial Intelligence" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.313, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.474, + 0.631 + ], + "angle": 0, + "content": "Modern image classifiers perform well on populated classes, while degrading considerably on tail classes with only a few instances. Humans, by contrast, effortlessly handle the long-tailed recognition challenge, since they can learn the tail representation based on different levels of semantic abstraction, making the learned tail features more discriminative. This phenomenon motivated us to propose SuperDisco, an algorithm that discovers super-class representations for long-tailed recognition using a graph model. We learn to construct the super-class graph to guide the representation learning to deal with long-tailed distributions. Through message passing on the super-class graph, image representations are rectified and refined by attending to the most relevant entities based on the semantic similarity among their super-classes. Moreover, we propose to meta-learn the super-class graph under the supervision of a prototype graph constructed from a small amount of imbalanced data. By doing so, we obtain a more robust super-class graph that further improves the long-tailed recognition performance. The consistent state-of-the-art experiments on the long-tailed CIFAR-100, ImageNet, Places and iNaturalist demonstrate the benefit of the discovered super-class graph for dealing with long-tailed distributions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.658, + 0.208, + 0.673 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.471, + 0.88 + ], + "angle": 0, + "content": "This paper strives for long-tailed visual recognition. A computer vision challenge that has received renewed attention in the context of representation learning, as real-world deployment demands moving from balanced to imbalanced scenarios. Three active strands of work involve class re-balancing [15, 22, 32, 43, 65], information augmentation [34, 51, 54] and module improvement [29, 31, 76]. Each of these strands is intuitive and has proven empirically successful. However, all these approaches seek to improve the classification performance of the original feature space. In this paper, we instead explore a graph learning algorithm to discover the imbalanced super-class space hidden in the original feature representation." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.261, + 0.659, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.348, + 0.649, + 0.361 + ], + "angle": 0, + "content": "(a) 100 original classes" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.259, + 0.85, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.348, + 0.865, + 0.361 + ], + "angle": 0, + "content": "(b) 20 ground truth super-classes" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.37, + 0.659, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.458, + 0.67, + 0.47 + ], + "angle": 0, + "content": "(c) 16 discovered super-classes" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.369, + 0.851, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.458, + 0.864, + 0.47 + ], + "angle": 0, + "content": "(d) 32 discovered super-classes" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Figure 1. SuperDisco learns to project the original class space (a) into a relatively balanced super-class space. Different color curves indicate the different imbalance factors on the long-tailed CIFAR-100 dataset. Like the 20 super-class ground truth (b) our discovered super-classes for 16 super-classes (c) or 32 super-classes (d) provide a much better balance than the original classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.795 + ], + "angle": 0, + "content": "The fundamental problem in long-tailed recognition [18, 32, 44, 77] is that the head features and the tail features are indistinguishable. Since the head data dominate the feature distribution, they cause the tail features to fall within the head feature space. Nonetheless, humans effortlessly handle long-tailed recognition [2, 16] by leveraging semantic abstractions existing in language to gain better representations of tail objects. This intuition hints that we may discover the semantic hierarchy from the original feature space and use it for better representations of tail objects. Moreover, intermediate concepts have been shown advantageous for classification [5, 36] by allowing the transfer of shared features across classes. Nevertheless, it remains unexplored to exploit intermediate super-classes in long-tailed visual recognition that rectify and refine the original features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In the real world, each category has a corresponding superclass, e.g., bus, taxi, and train all belong to the vehicle superclass. This observation raises the question: are super-classes of categories also distributed along a long-tail? We find empirical evidence that within the super-class space of popular datasets, the long-tailed distribution almost disappears, and each super-class has essentially the same number of samples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.42, + 0.9 + ], + "angle": 0, + "content": "*Currently with United Imaging Healthcare, Co., Ltd., China." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.47, + 0.333 + ], + "angle": 0, + "content": "In Figure 1, we show the number of training samples for each of the original classes and their corresponding super-classes in the long-tailed CIFAR-100 dataset. We observe the data imbalance of super-classes is considerably lower than those of the original classes. This reflects the fact that the original imbalanced data hardly affects the degree of imbalance of the super-classes, which means the distribution of the super-classes and original data is relatively independent. These balanced super-class features could be used to guide the original tail data away from the dominant role of the head data, thus making the tail data more discriminative. Therefore, if the super-classes on different levels of semantic abstraction over the original classes can be accurately discovered, it will help the model generalize over the tail classes. As not all datasets provide labels for super-classes, we propose to learn to discover the super-classes in this paper." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.334, + 0.47, + 0.77 + ], + "angle": 0, + "content": "Inspired by the above observation, we make in this paper two algorithmic contributions. First, we propose in Section 3 an algorithm that learns to discover the super-class graph for long-tailed visual recognition, which we call SuperDisco. We construct a learnable graph that discovers the super-class in a hierarchy of semantic abstraction to guide feature representation learning. By message passing on the super-class graph, the original features are rectified and refined, which attend to the most relevant entities according to the similarity between the original image features and super-classes. Thus, the model is endowed with the ability to free the original tail features from the dominance of the head features using the discovered and relatively balanced super-class representations. Even when faced with the severe class imbalance challenges, e.g., iNaturalist, our SuperDisco can still refine the original features by finding a more balanced super-class space using a more complex hierarchy. As a second contribution, we propose in Section 4 a meta-learning variant of our SuperDisco algorithm to discover the super-class graph, enabling the model to achieve even more balanced image representations. To do so, we use a small amount of balanced data to construct a prototype-based relational graph, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. Last, in Section 5 we report experiments on four long-tailed benchmarks: CIFAR-100-LT, ImageNet-LT, Places-LT, and iNaturalist, and verify that our discovered super-class graph performs better for tail data in each dataset. Before detailing our contributions, we first embed our proposal in related work." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.785, + 0.212, + 0.801 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.81, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Long-tailed recognition. Several strategies have been proposed to address class imbalance in recognition. We categorize them into three groups. Those in the first group are based on class re-balancing [8, 30, 44, 75], which balance the training sample numbers of different classes during model training. Class re-balancing methods also could be catego-" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.394 + ], + "angle": 0, + "content": "rized into three different groups, i.e., re-sampling [22, 32, 43, 67], cost-sensitive learning [14, 37, 56, 78, 79, 86] and logit adjustment [27, 45, 57, 59]. Class re-balancing methods improve the performance of the tail classes at the expense of the performance of the head classes. The second group is based on information augmentation, introducing additional information into model training to improve long-tailed learning performance. We identify four kinds of methods in the information augmentation scope, i.e., transfer learning, which includes head-to-tail knowledge transfer [6, 42, 64, 73], knowledge distillation [28, 40, 71], model pre-training [9, 33, 72] and self-training [24, 68, 74]. The third group focuses on improving network modules in long-tailed learning. This group includes representation learning [13, 46, 76], classifier learning [32, 41, 42, 69, 73], decoupled training [31, 32, 82], and ensemble learning [20, 83]. These methods introduce additional computation costs for increased performance. Our method belongs to the third group as it aims to learn a better representation of unbalanced training samples by the superclass graph, which is unexplored for long-tail recognition." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.399, + 0.893, + 0.73 + ], + "angle": 0, + "content": "Super-class learning. Super-class learning adds superclass labels as intermediate supervision into traditional deep learning. A super-class guided network [38] integrated the high-level semantic information into the network for image classification and object detection, which took two-level class annotations that contain both super-class and finer class labels. In [11], a two-phase multi-expert architecture was proposed for still image action recognition, which includes fine-grained and coarse-grained phases. However, they leveraged the ground truth of the super-class as supervision during the coarse-grained phase. Wu et al. [70] propose a taxonomic classifier to address the long-tail recognition problem, which classified each sample to the level that the classifier is competent. Zhou et al. [85] clustered the original categories into super-classes to produce a relatively balanced distribution in the super-class space, which also leveraged the ground truth of the super-class in the training phase. In contrast with the previous super-class learning, we do not use ground truth to group the original categories into the super-class space. To the best of our knowledge, no work exists that relies on graph learning to discover the super-class for long-tailed visual recognition, thus motivating this work." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.735, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Graph neural networks. Recently, several graph neural network models (GNN) have been proposed to exploit the structures underlying graphs to benefit a variety of applications. There are two main research lines of GNN methods: non-spectral methods and spectral methods. The spectral methods [4, 10, 26, 35] focus on learning graph representations in a spectral domain, in which the learned filters are based on Laplacian matrices. The non-spectral methods [21, 63] develop an aggregator to aggregate a local set of features. Note that, message passing [19] is a key mechanism that allows GCNs and other graph neural networks to capture" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.09, + 0.482, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.343, + 0.398, + 0.355 + ], + "angle": 0, + "content": "(a) Illustration of proposed SuperDisco" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.088, + 0.791, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.209, + 0.768, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.343, + 0.774, + 0.355 + ], + "angle": 0, + "content": "(c) Visualization of discovered super-class graph \\( c^2 \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Figure 2. Illustration of proposed SuperDisco (a) and visualization of the similarity between the classes and discovered super-class at different levels (b), (c). In (a), we show two levels of super-class graphs \\(\\mathcal{C}^1\\) and \\(\\mathcal{C}^2\\). The colour in each graph represents the discovered super-class. SuperDisco discovers the potential super-class at different levels hidden in each category from (b) and (c). \\(\\mathcal{C}^1\\) roughly categorizes the original classes into three relatively balanced super-classes, and \\(\\mathcal{C}^2\\) then finely categorize them into four more balanced super-classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.471, + 0.557 + ], + "angle": 0, + "content": "complex relationships and dependencies between nodes in a graph, and is a major reason why they have been successful in a variety of tasks involving graph-structured data. Our method belongs to the non-spectral methods, which leverage a GNN as the base architecture to discover the super-class representation. Our proposed super-class graph would refine and rectify the original imbalanced feature to a relatively balanced feature space, which has not been explored for long-tail recognition either." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.471, + 0.801 + ], + "angle": 0, + "content": "Meta-learning for the long-tail. Meta-learning or learning to learn [3, 53, 58, 80, 81], is a learning paradigm where a model is trained on the distribution of tasks so as to enable rapid learning on new tasks. Ren et al. [51] first proposed meta-learning for the long-tailed problem by reweighting training examples. Shu et al. [54] proposed Meta-weight-Net to adaptively extract sample weights to guarantee robust recognition in the presence of training data bias. Li et al. [39] introduced meta-semantic augmentation for long-tailed recognition, which produces diversified augmented samples by translating features along many semantically meaningful directions by meta-learning. Our uniqueness is that our model aims to discover an improved super-class representation by meta-learning, which enables the original feature representation to adjust its corresponding higher-level super-class space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.456, + 0.833 + ], + "angle": 0, + "content": "3. Learning to discover the super-class graph" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this section, we discuss how to learn to discover the super-class graph from the training samples and then expand on how to leverage such a graph to benefit the unbalanced data by refining the feature representations of samples. The" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.453 + ], + "angle": 0, + "content": "overall illustration of SuperDisco and a visualization of the discovered super-class hierarchy are shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Preliminary. For long-tailed visual recognition, the goal is to learn an image classification model from an imbalanced training set and to evaluate the model on a balanced test set. We first define the notation for long-tailed recognition used throughout our paper. We define a training input \\( x_{k} \\in \\mathbb{R}, i \\in \\{1, \\dots, n\\} \\), where \\( n \\) is the number of training samples in the dataset. The corresponding labels are \\( y_{k} \\in \\{1, 2, \\dots, C\\} \\) where \\( C \\) is the number of classes. Let \\( n_{j} \\) denote the number of training samples for the class \\( j \\). Here, we assume that \\( n_{i} \\geq n_{j} \\) when \\( i < j \\) shows the long-tailed problem simply. In this work, we typically consider a deep network model with three main components: a feature extractor \\( f(\\cdot) \\), a proposed graph model \\( g(\\cdot) \\) and a classifier \\( h(\\cdot) \\). The feature extractor \\( f(\\cdot) \\) first extracts an image representation as \\( \\mathbf{z} = f(x; \\theta) \\), which is then fed into the proposed graph model to refine a new representation as \\( g(\\mathbf{z}; \\phi) = \\tilde{\\mathbf{z}} \\). The final class prediction \\( \\tilde{y} \\) is given by a classifier function \\( h(\\cdot) \\), i.e., \\( \\tilde{y} = \\arg \\max h(\\tilde{\\mathbf{z}}; \\psi) \\). Before detailing our approach, we add Table 1 to detail the meaning of each symbol for easy lookup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "SuperDisco. We construct the super-class graph to organize and distill knowledge from the training process. The vertices represent different types of super-classes (e.g., the common contour between birds and airplanes) and the edges are automatically constructed to reflect the relationship between different super-classes. Our super-class graph contains multiple levels, which is closer to the relationship between various objects in the real world. Before detailing the structure, we first explicate why the multi-level super-class graphs are preferred over a flat super-class graph: a single level of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "super-class groups is likely insufficient to model complex task relationships in real-world applications; for example, the similarities among different bird species are high, but there are also similarities between birds and mammals, e.g., they are both animals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.47, + 0.337 + ], + "angle": 0, + "content": "We assume the vertex representation \\(g\\) as \\(\\mathbf{h}^g\\in \\mathbb{R}^d\\) and define the super-class graph as \\(\\mathcal{C}^l = (\\mathbf{H}_c^l,\\mathbf{A}_c^l)\\), where \\(\\mathbf{H}_c^l = \\{\\mathbf{h}^{i^l}|\\forall i^l\\in [1,C^l ]\\} \\in \\mathbb{R}^{C^l\\times d}\\) is the vertex feature matrix of the \\(l\\)-th super-class level and \\(\\mathbf{A}_c^l = \\{A_c^l (\\mathbf{h}^{i^l},\\mathbf{h}^{j^l})|\\forall i^l,j^l\\in [1,C^l ]\\} \\in \\mathbb{R}^{C^l\\times C^l}\\) is the vertex adjacency matrix in the \\(l\\)-th super-class level, \\(C^l\\) denotes the number of vertices in the \\(l\\)-th super-class level. Our vertex representation \\(\\mathbf{H}_c^l\\) of the super-class graph is defined to get parameterized and learned during training. The initial vertex representations of each super-class level are randomly initialized, which encourages diversity of the discovered super-classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.4 + ], + "angle": 0, + "content": "Next, we introduce how to compute the edge weight \\( A_{\\mathcal{C}}^{l} \\) in the super-class graph. The edge weight \\( A_{\\mathcal{C}}^{l}(\\mathbf{h}^{i^{l}}, \\mathbf{h}^{j^{l}}) \\) between a pair of vertices \\( i \\) and \\( j \\) is gauged by the similarity between them. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.404, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\nA _ {\\mathcal {C}} ^ {l} \\left(\\mathbf {h} ^ {i l}, \\mathbf {h} ^ {j l}\\right) = \\sigma \\left(\\mathbf {W} _ {c} ^ {l} \\left(| \\mathbf {h} ^ {i l} - \\mathbf {h} ^ {j l} | / \\gamma_ {c} ^ {l}\\right) + \\mathbf {b} _ {c} ^ {l}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.716 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_c^l\\) and \\(\\mathbf{b}_c^l\\) indicate learnable parameters of the \\(l\\)-th super-class level, \\(\\gamma_c^l\\) of \\(l\\)-th super-class level is a scalar and \\(\\sigma\\) indicates the Sigmoid function, which normalizes the weight between 0 and 1. To adjust the representation of training samples by the involvement of super-classes, we first query the training samples in the super-class graph to obtain the relevant super-class. In light of this, we construct a new graph \\(\\mathcal{R}\\), which adds the original sample feature \\(\\mathbf{z}\\) to the super-class graph. We define \\(\\mathbf{z}^l\\) as the refined feature after the \\(l\\)-th super-class graph. Here we define graph \\(\\mathcal{R}^l = (\\mathbf{H}_{\\mathcal{R}}^l, \\mathbf{A}_{\\mathcal{R}}^l)\\), where \\(\\mathbf{H}_{\\mathcal{R}}^l = \\{[\\mathbf{z}^l, \\mathbf{h}^{i^l}] | \\forall i^l \\in [1, C^l]\\} \\in \\mathbb{R}^{(C^l + 1) \\times d}\\) denotes the vertex feature matrix of the \\(l\\)-th super-class level, and \\(\\mathbf{A}_{\\mathcal{R}}^l = \\{[A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{z}^1), A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{h}^{j^l})] | \\forall i^l, j^l \\in [1, C^l]\\} \\in \\mathbb{R}^{C^l + 1 \\times C^{l + 1}}\\) denotes the vertex adjacency matrix in the \\(l\\)-th super-class level. The link between \\(\\mathbf{z}^l\\) and vertex \\(\\mathbf{h}^i\\) in the hierarchical graph is constructed by their similarity. In particular, analogous to the definition of weight in the super-class graph in Eq. (1), the weight \\(A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{z}^1)\\) is constructed as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.717, + 0.47, + 0.736 + ], + "angle": 0, + "content": "\\[\nA _ {\\mathcal {R}} ^ {l} \\left(\\mathbf {h} ^ {i ^ {l}}, \\mathbf {z} ^ {1}\\right) = \\sigma \\left(\\mathbf {W} _ {r} ^ {l} \\left(\\left| \\mathbf {h} ^ {i ^ {l}} - \\mathbf {z} ^ {1} \\right| / \\gamma_ {r} ^ {l}\\right) + \\mathbf {b} _ {r} ^ {l}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.769 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_r^l\\) and \\(\\mathbf{b}_c^r\\) indicate learnable parameters of the \\(l\\)-th super-class level, \\(\\gamma_r^l\\) of the \\(l\\)-th super-class level is a scalar." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.846 + ], + "angle": 0, + "content": "After constructing the new graph \\(\\mathcal{R}\\), we propagate the most relevant super-class by message passing [19] from the discovered super-classes \\(\\mathcal{C}\\) to the features \\(\\mathbf{z}^l\\) by introducing a Graph Neural Network (GNN). The message passing operation over the graph is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.849, + 0.47, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {\\mathcal {R}} ^ {(m + 1)} = \\mathrm {M P} \\left(\\mathbf {A} _ {\\mathcal {R}} ^ {l}, \\mathbf {H} _ {\\mathcal {R}} ^ {(m)}; \\mathbf {W} ^ {(m)}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbb{M}\\mathbb{P}(\\cdot)\\) is the message passing function, \\(\\mathbf{H}^{(m)}\\) is the vertex embedding after \\(m\\) layers of GNN and \\(\\mathbf{W}^{(m)}\\) is a" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.089, + 0.837, + 0.249 + ], + "angle": 0, + "content": "
NotationDescription
hVertex representation
zOriginal sample feature
CSuper-class graph
HcVertex feature matrix of Cl
AcVertex adjacency matrix of Cl
RGraph which adds z to graph Cl
PPrototype graph
CpVertex feature matrix of P
ApVertex adjacency matrix of P
SSuper graph which connecting P and C
AVertex feature matrix of S
MVertex adjacency matrix of S
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.251, + 0.874, + 0.264 + ], + "angle": 0, + "content": "Table 1. Summary of the core notation used for SuperDisco." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.894, + 0.436 + ], + "angle": 0, + "content": "learnable weight matrix of layer \\( m \\). After stacking \\( M \\) GNN layers, we get the information-propagated feature representation \\( \\tilde{\\mathbf{z}}^L \\) for each level of the super-class graph \\( \\mathcal{C} \\). Once we obtain the refined representation \\( \\tilde{\\mathbf{z}}^L \\) for a training sample by the super-class graph, we feed them into the classifier to make the predictions and compute the corresponding loss, i.e., Cross-entropy loss for optimization. Using gradient descent, we then update the super-class graph \\( \\mathcal{C} \\). To be able to discover a more accurate super-class graph in the face of severe imbalance problems, we propose meta-learning super-class graph discovery in the next section." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.451, + 0.88, + 0.468 + ], + "angle": 0, + "content": "4. Meta-learning super-class graph discovery" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.717 + ], + "angle": 0, + "content": "To explore and exploit a more accurate and richer superclass graph, we propose the Meta-SuperDisco to discover the super-class graph using meta-learning, making the model more robust. In the traditional meta-learning setting [17, 50], it includes meta-training tasks and meta-test task. Each task contains a support set \\( S \\) and a query set \\( Q \\). Each task is first trained by \\( S \\) to get the task-specific learner and \\( Q \\) optimizes this learner to update the meta-learner. For long-tailed recognition with meta-learning, previous works [51, 54] randomly sample a small amount of balanced data denoted as \\( \\mathcal{M} \\). The imbalanced data and the small balanced data can be seen as \\( S \\) and \\( Q \\) in the training phase. The goal of meta-learning for long-tailed recognition is to use a small set of balanced data to optimize the model obtained from unbalanced data. We follow [51, 54] by randomly selecting the same number of samples (e.g., ten) as \\( \\mathcal{M} \\) per class from the training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.855 + ], + "angle": 0, + "content": "Meta-SuperDisco. To meta-learn the super-class graph, we construct a prototype graph \\(\\mathcal{P}\\) from \\(\\mathcal{M}\\), since \\(\\mathcal{M}\\) is a balanced dataset. The prototype graph extracts the sample-level relation information, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. For the prototype graph, we need to compute the prototype of each category [55], which is defined as: \\(\\mathbf{c}^k = \\frac{1}{N^k}\\sum_{i = 1}^{N^k}\\mathbf{z}_j\\), where \\(N^k\\) denotes the number of samples in class \\(k\\), \\(\\mathbf{z}_j\\) is the feature from \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "After calculating all prototype representations \\(\\{\\mathbf{c}^k |\\forall k\\in [1,K]\\}\\), which serve as the vertices in the prototype graph \\(\\mathcal{P}_i\\), we further need to define the edges and the corresponding" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.137 + ], + "angle": 0, + "content": "edge weights. The edge weight \\( A_{\\mathcal{P}}(\\mathbf{c}^i,\\mathbf{c}^j) \\) between two prototypes \\( \\mathbf{c}^i \\) and \\( \\mathbf{c}^j \\) is gauged by the similarity between them. The edge weight is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.141, + 0.471, + 0.157 + ], + "angle": 0, + "content": "\\[\nA _ {\\mathcal {P}} \\left(\\mathbf {c} ^ {i}, \\mathbf {c} ^ {j}\\right) = \\sigma \\left(\\mathbf {W} _ {p} \\left(\\left| \\mathbf {c} ^ {i} - \\mathbf {c} ^ {j} \\right| / \\gamma_ {p}\\right) + \\mathbf {b} _ {p}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.163, + 0.473, + 0.36 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_p\\) and \\(\\mathbf{b}_p\\) are the learnable parameters, \\(\\gamma_p\\) is a scalar. Thus, we denote the prototype graph as \\(\\mathcal{P} = (\\mathbf{C}_{\\mathcal{P}}, \\mathbf{A}_{\\mathcal{P}})\\), where \\(\\mathbf{C}_{\\mathcal{P}_i} = \\{\\mathbf{c}^i | \\forall i \\in [1, K]\\} \\in \\mathbb{R}^{K \\times d}\\) represent a set of vertices, with each one corresponding to the prototype from a class, while \\(\\mathbf{A}_{\\mathcal{P}} = \\{|A_{\\mathcal{P}}(\\mathbf{c}^i, \\mathbf{c}^j)| \\forall i, j \\in [1, K]\\} \\in \\mathbb{R}^{K \\times K}\\) gives the adjacency matrix, which indicates the proximity between prototypes. We then use the prototype graph to guide the learning of the meta super-class graph. We construct a super graph \\(\\mathcal{S}\\) by connecting prototype graph \\(\\mathcal{P}\\) to super-class graph \\(\\mathcal{C}\\). In the super graph \\(\\mathcal{S}\\), the vertices are \\(\\mathbf{M} = (\\mathbf{C}_{\\mathcal{P}^l}^l; \\mathbf{H}_{\\mathcal{C}^l}^l)\\), where \\(\\mathcal{P}^l\\) denotes the refined prototype graph vertex after the \\(l\\)-th level super-class graph. Then, we calculate the link weight \\(A_{\\mathcal{S}}^l(c^i, \\{\\mathbf{h}^j\\})\\) of the super graph as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.365, + 0.47, + 0.406 + ], + "angle": 0, + "content": "\\[\nA _ {S} ^ {l} \\left(\\mathbf {c} ^ {i}, \\mathbf {h} ^ {j ^ {l}}\\right) = \\frac {\\exp \\left(- \\| \\left(\\mathbf {c} ^ {i} - \\mathbf {h} ^ {j ^ {l}}\\right) / \\gamma_ {s} ^ {l} \\| _ {2} ^ {2} / 2\\right)}{\\sum_ {j ^ {l ^ {\\prime}} = 1} ^ {J} \\exp \\left(- \\| \\left(\\mathbf {c} ^ {i} - \\mathbf {h} ^ {j ^ {l ^ {\\prime}}}\\right) / \\gamma_ {s} ^ {l} \\| _ {2} ^ {2} / 2\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.472, + 0.563 + ], + "angle": 0, + "content": "where \\(\\gamma_s^l\\) is a scaling factor. Note that, here we use softmax to ensure that the total weight of edges between the prototype graph \\(\\mathcal{P}\\) and the super-class graph \\(\\mathcal{C}\\) is equal to 1, giving the prototype graph a unique influence on the expression of each super-class. Thus, the adjacent matrix and feature matrix of the super graph \\(S^l = (\\mathbf{A}^l,\\mathbf{M}^l)\\) is defined as \\(\\mathbf{A}^l = (\\mathbf{A}_{\\mathcal{P}},\\mathbf{A}_S^l;\\mathbf{A}_S^{l^\\mathrm{T}},\\mathbf{A}_\\mathcal{C}^l)\\) and \\(\\mathbf{M}^l = (\\mathbf{C}_{\\mathcal{P}^l}^l;\\mathbf{H}_{\\mathcal{C}^l}^l)\\). Once we constructed the super graph \\(S\\), we use message-passing again to propagate the most relevant knowledge from the prototype graph \\(\\mathcal{P}\\) to the super-class graph \\(\\mathcal{C}\\). Similar to eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.567, + 0.47, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {(m + 1)} = \\mathbb {M P} \\left(\\mathbf {A} ^ {l}, \\mathbf {M} ^ {(m)}; \\mathbf {W} ^ {(m)}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.472, + 0.696 + ], + "angle": 0, + "content": "We leverage the graph \\( S \\) to refine the super-class graph. Finally, we feed the original feature \\( \\mathbf{z} \\) into the super-class graph to get the information-propagated feature representation \\( \\tilde{\\mathbf{z}}^L \\), which refines the original feature by its corresponding discovered super-classes. We provide the complete SuperDisco and Meta-SuperDisco algorithm specifications in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.21, + 0.726 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Datasets. We apply our method to four commonly used long-tail recognition benchmarks. Sample images and the number of categories for all datasets are provided in the supplement material. CIFAR-100-LT reduces the number of training samples per class according to an exponential function \\( n = n_i \\mu^i \\), where \\( i \\) is the class index, \\( n_i \\) is the original number of training samples, and \\( \\mu \\in (0, 1) \\). The imbalance factor of a dataset is defined as the number of training samples in the most populated class divided by the minority class. We consider imbalance factors {10, 50, 100}. ImageNet-LT [44] is a subset of ImageNet [12] consisting of" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.089, + 0.848, + 0.174 + ], + "angle": 0, + "content": "
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
SuperDisco65.960.757.250.945.2
Meta-SuperDisco68.563.158.353.847.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.177, + 0.895, + 0.233 + ], + "angle": 0, + "content": "Table 2. Benefit of SuperDisco and Meta-SuperDisco. SuperDisco achieves better performance compared to a baseline fin-tuning on all imbalance factors, while Meta-SuperDisco is even better for long-tailed recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.895, + 0.44 + ], + "angle": 0, + "content": "115.8K images from 1000 categories, with maximally 1,280 images per class and minimally 5 images per class, and a balanced test set. Places-LT [44] has an imbalanced training set with 62,500 images for 365 classes from Places [84]. It contains images from 365 classes and the number of images per class ranges from 4980 to 5. The test sets are balanced and contain 100 images per class. iNaturalist [61] is a real-world long-tailed dataset with 675,170 training images for 5,089 classes, where the top \\(1\\%\\) most populated classes contain more than \\(16\\%\\) of the training images. Additionally, there is also a severe imbalance among the super-classes of iNaturalist. The 13 ground truth super-classes images range from 158,407 to 308." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.897, + 0.672 + ], + "angle": 0, + "content": "Implementation details. We follow [32] by first training a feature extractor with instance-balanced sampling, and then training our graph model and classifier based on the trained features. For CIFAR-100-LT, we follow [54] and use a ResNet-32 backbone. For ImageNet-LT, we use ResNeXt-50 [23] as our backbone, following [32]. For Places-LT, we report results with ResNet-152 following [44]. For iNaturalist, we use a ResNet-50 backbone. We train each dataset for 200 epochs with batch size 512. We use random left-right flipping and cropping as our training augmentation. For all experiments, we use an SGD optimizer with a momentum of 0.9 and a batch size of 512. We randomly selected 10 images per class from the training set for all datasets as \\(\\mathcal{M}\\). Code available at: https://github.com/Yingjun-Du/SuperDisco." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Benefit of SuperDisco and Meta-SuperDisco. To show the benefit of SuperDisco, we compare it with a fine-tuning baseline, which retrans the classifier only. Table 2 shows SuperDisco improves over fine-tuning on CIFAR-100-LT, and the results for the other long-tailed datasets are provided in the supplemental materials Table 1. In the most challenging setting with the largest imbalance factor of 200, our SuperDisco delivers \\(45.2\\%\\), surpassing the baseline by \\(5.9\\%\\). We attribute improvement to our model's ability to refine original features, allowing the discovered super-class graph to guide the tail features away from the dominant role of head features, thus leading to improvements over the original features. We also investigate the benefit of meta-learning with Meta-SuperDisco. The Meta-SuperDisco consistently surpasses the SuperDisco for all imbalance factors. The con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19948" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.451, + 0.242 + ], + "angle": 0, + "content": "
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
(20)61.260.149.947.341.9
(2, 4, 8)65.362.753.149.843.2
(4, 8, 16)69.164.255.252.345.9
(4, 8, 16, 32)68.563.158.353.847.5
(4, 8, 16, 32, 64)66.962.758.952.946.3
Oracle super-classes66.963.254.751.443.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.244, + 0.47, + 0.287 + ], + "angle": 0, + "content": "Table 3. Effect of number of super-class levels on CIFAR-100-LT. Compared to a baseline [32] and an oracle setting, Meta-SuperDisco provides higher performance gains with more complex hierarchies." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.294, + 0.428, + 0.426 + ], + "angle": 0, + "content": "
ManyMediumFewAll
Baseline65.066.365.565.9
(13)71.870.266.170.8
(2,4,8)70.569.365.969.4
(4,8,16)72.270.966.470.3
(4,8,16,32)73.670.267.370.9
(4,8,16,32,64)73.472.968.372.3
(4,8,16,32,64,128)72.171.366.270.9
Oracle super-classes70.770.565.970.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.428, + 0.472, + 0.469 + ], + "angle": 0, + "content": "Table 4. Effect of number of super-class levels on iNaturalist. Meta-SuperDisco achieves consistent performance gains with more complex hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.47, + 0.521 + ], + "angle": 0, + "content": "sistent improvements confirm that Meta-SuperDisco learns even more robust super-class graphs, leading to a discriminative representation of the tail data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effect of the number of super-class levels. A significant challenge with any structure-aware learning algorithm is determining the appropriate complexity for the knowledge structure. So, we further analyze the effect of the super-class hierarchies, including the level (number of depths L) or the number of super-classes in each level. The results are shown in Table 3 and Table 4. The super-class number from the bottom layer to the top layer is saved in a tuple. For example, (2, 4, 8) represents three depth, with two super-classes in the top layer. The baseline is Decouple-LWS [32], which only inputs the original feature to learn a new classifier. The oracle super-classes are first trained on two long-tailed datasets using the ground truth super-class labels for super-class classifications. Once the training is completed, each oracle super-class is obtained by averaging the samples of each super-class. We constructed a one-layer super-class graph using these super-classes, where the vertices of the graph are for each super-class, and the edges of the graph are computed according to Eq. (1). Then, we use the message passing by Eq. (3) to refine the original features and input them into the classifier to get the final predictions. From Table 3, we observe that using oracle super-classes achieves better performance compared to the learned super-class (20) since it uses the ground truth super-classes as supervision. We also conclude that too few levels may not be enough" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.088, + 0.67, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.161, + 0.638, + 0.173 + ], + "angle": 0, + "content": "(a) SuperDisco" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.089, + 0.882, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.161, + 0.842, + 0.173 + ], + "angle": 0, + "content": "(b) Meta-SuperDisco" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.175, + 0.894, + 0.217 + ], + "angle": 0, + "content": "Figure 3. Similarity between discovered super-classes and classes. SuperDisco discovers super-classes hidden in each class, while Meta-SuperDisco discovers more accurate super-classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.895, + 0.536 + ], + "angle": 0, + "content": "to learn the precise super-classes (e.g., tuple (20) or (2, 4, 8)). In this dataset, increasing levels (e.g., tuple (4, 8, 16, 32)) achieves better performance on the smaller imbalance factor (e.g., 10), and similar performance compared with (4, 8, 16). For the real-world long-tailed dataset iNaturalist [62] in Table 4, we find no significant improvement for the few-shot classes in the performance of the oracle super-class compared to the baseline, and the same is true for the performance of the discovered super-class structure (13). This is because the super-class of iNaturalist also have serious long-tailed problems, resulting in the refined features of tail classes remaining indistinguishable from the refined features of head classes. However, with a more complex graph structure (4, 8, 16, 32, 64), the few-shot performance improves by a good margin compared with the baseline, and even the oracle super-classes. We attribute this to our model's ability to explore relatively balanced super-class spaces, thus making the refined tail category features discriminative. By comparing Table 3 and Table 4, we conclude that deeper as well as wider graphs are needed to discover the super-classes in the case of severe class imbalance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Visualization of SuperDisco. To understand the meaning of the discovered super-classes more clearly, we present a visualization in Figure. 3. We selected 12 different categories from the CIFAR-100 test dataset. We calculate the similarity of each of these 12 categories to the different vertices in the graph we explore. Here we show the similarity with the second layer of graph vertices (C1, C2, C3, C4). We can see different categories mainly activate different vertices, e.g., bus \\(\\rightarrow\\) C3 and road \\(\\rightarrow\\) C2. As shown in this heatmap, we find that C1 reflects the super-class of flowers, C2 reflects the super-class of buildings, C3 reflects the super-class of vehicles, C4 reflects the super-class of fish. Another observation is that the second-largest activated super-class is also meaningful, promoting knowledge transfer between super-classes. For example, road and bridge are related to the C3 superclass, since some vehicles may be on the road and bridge. This visualization reflects that we can use graph models to discover the super-classes and the relationships between each super-class. We also visualize the discovered meta-learning super-classes in Figure 3 (b). The discovered super-classes are even more accurate, e.g., roses have high similarity to C2, which mainly reflects the buildings super-class, while it has high similarity to C1, which is the flowers super-class. This once again validates the benefit of Meta-SuperDisco." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.089, + 0.298, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.19, + 0.27, + 0.203 + ], + "angle": 0, + "content": "(a) Original features" + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.09, + 0.475, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.091, + 0.653, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.091, + 0.829, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.19, + 0.676, + 0.203 + ], + "angle": 0, + "content": "(c) Refined features by SuperDisco" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.21, + 0.294, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.309, + 0.287, + 0.322 + ], + "angle": 0, + "content": "(b) Training distribution" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.21, + 0.468, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.21, + 0.651, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.21, + 0.829, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.309, + 0.692, + 0.322 + ], + "angle": 0, + "content": "(d) Refined features by Meta-SuperDisco" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Figure 4. Visualization of refined features on CIFAR-100-LT, with the original features (a) and their corresponding training distribution (b). Colours indicate categories. SuperDisco (c) guides the original features on being clustered into the corresponding super-class space at different levels, while Meta-SuperDisco (d) obtains even more discriminative intra-class features." + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.388, + 0.391, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Figure 5. Effect of refined features. Accuracy increases along with the increased super-class levels, revealing that more accurate and richer super-classes facilitate better long-tailed recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Furthermore, in the (c) learned the hierarchical concept of each class, we can see that bus and bridge have the same concept C2 in the last concept level, which may be due to the possible presence of cars on the bridge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Visualization of refined features. To understand the empirical benefit of SuperDisco, we visualize in Figure 4 the original features and refined features with super-class graphs of the different levels using t-SNE [60]. We choose the vertices numbers as (2, 4, 6), meaning the super-class graph has three different levels, each with a different number of vertices. The original features of the category with a small sample size will overlap with the (original) features of the category with a large sample size. Super-class graphs discovered by our model project the original features into a high-level super-class space, where the imbalance is relatively small. Hence, its corresponding subcategory can be predicted more accurately. It is worth noting that when comparing the two different super-class graphs on top and below, the features obtained by Meta-SuperDisco are even more distinctive and distant from each other. To better measure the goodness of the refined features obtained at different levels," + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.38, + 0.866, + 0.442 + ], + "angle": 0, + "content": "
ManyMediumFewAll
Baseline58.449.334.852.7
Multi-layer perceptron63.551.835.955.0
Graph convolution network66.153.337.157.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Table 5. Analysis of super-class mechanism on ImageNet-LT. The super-class mechanism contributes most, the graph convolution network improves results further." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.894, + 0.595 + ], + "angle": 0, + "content": "we show in Figure 5 the prediction accuracy using different refined features. We find that the accuracy increases along with the increased super-class levels, which shows that using more accurate and richer super-classes facilitates better performance. This again demonstrates that Meta-SuperDisco is most suitable for long-tailed visual recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.778 + ], + "angle": 0, + "content": "Analysis of super-class mechanism. To demonstrate that the improved performance of our SuperDisco cannot solely be attributed to the graph convolutional network module, we conducted an experiment where we replaced it with a multi-layer perceptron to obtain the representation per sample. In Table 5, the performance gains of our method are primarily due to the super-classes rather than the graph convolution network. The results suggest that incorporating the super-classes mechanism plays a crucial role in improving the performance of long-tailed problems. Furthermore, the results improve further when we replace the multi-layer perceptron with our graph convolution network module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Comparison with the state-of-the-art. We evaluate our method on the four long-tailed datasets under different imbalance factors in Table 6 and 7. Our model achieves state-of-the-art performance on the tail data of all datasets. For ImageNet-LT, our model achieves state-of-the-art performance on both few-shot and all data. In the most challenging Places-LT, our model delivers \\(40.3\\%\\) on all classes, surpassing the second-best Parisot et al. [47] by \\(1.1\\%\\). On the real-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19950" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.835, + 0.243 + ], + "angle": 0, + "content": "
VenueImageNet-LTPlaces-LTiNaturalist
ManyMediumFewAllManyMediumFewAllManyMediumFewAll
Kang et al. [32]ICLR 1960.247.230.349.940.639.128.637.665.066.365.565.9
Kang et al. [31]ICLR 2161.849.430.951.5-------68.6
He et al. [25]ICCV 2164.150.431.553.1----70.670.167.669.1
Li et al. [40]CVPR 2166.851.135.456.0-------69.3
Samuel et al. [52]ICCV 2164.049.833.153.5-------69.7
Alshammari et al. [1]CVPR 2262.550.441.553.9----71.270.469.770.2
Zhang et al. [75]CVPR 2161.352.231.452.940.442.430.139.369.071.170.270.6
Parisot et al. [47]CVPR 2263.252.136.954.139.741.034.939.2----
Park et al. [48]CVPR 2266.453.935.656.2----73.172.668.772.8
This paper66.153.337.157.145.342.835.340.372.372.971.373.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.254, + 0.895, + 0.297 + ], + "angle": 0, + "content": "Table 6. Comparison with the state-of-the-art on ImageNet-LT, Places-LT and iNaturalist. Best and second best results are highlighted in bold and italic bold. Our Meta-SuperDisco achieves either better or comparable performance than state-of-the-art methods under the tail and all data for long-tailed visual recognition." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.302, + 0.421, + 0.449 + ], + "angle": 0, + "content": "
VenueImbalance ratio
1050100
Park et al. [49]ICCV 2159.547.442.0
Li et al. [40]CVPR 2162.350.546.0
Zhong et al. [82]CVPR 2162.551.546.8
Samuel et al. [52]ICCV 2163.457.647.3
Wang et al. [66]ICLR 2161.851.748.0
Zhu et al. [87]CVPR 2264.956.651.9
Cui et al. [7]ICCV 2164.256.052.0
Alshammari et al. [1]CVPR 2268.857.753.3
This paper69.358.353.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.46, + 0.472, + 0.489 + ], + "angle": 0, + "content": "Table 7. Comparison with the state-of-the-art on CIFAR-100-LT. Our model achieves best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.471, + 0.62 + ], + "angle": 0, + "content": "world long-tailed dataset iNaturalist, our model achieves the three best performances under four different shots. On the long-tailed synthetic dataset CIFAR-100-LT, our model achieves the best performance under each imbalance factor. The consistent improvements on all benchmarks under various configurations confirm that our Meta-SuperDisco is effective for long-tailed visual recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Limitations. We show that SuperDisco and Meta-SuperDisco achieve good performance on tail data while being less successful on the head data. Based on this result, we also perform an experiment on balanced CIFAR-100 in Figure 6. With SuperDisco and Meta-SuperDisco, there is only a slight change in performance at the expense of an increased inference time. This reveals that our SuperDisco does not change the original features much through message passing on a balanced dataset. This may be because the obtained super-classes are still the original class itself. In addition, as the computation of graphs involves many matrix operations, our model also requires a relatively long computational speed. Due to introducing a prototype graph and more data, Meta-SuperDisco takes longer to compute. In addition, the training time of SuperDisco and its meta variants is also 1.5 times higher than the baseline. Future work could investigate how to use the discovered super-class graph in balanced datasets and how to reduce the computation time." + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.316, + 0.825, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.529 + ], + "angle": 0, + "content": "Figure 6. Limitation. Accuracy \\((\\%)\\) vs. speed (ms) comparison with different methods on balanced CIFAR-100. SuperDisco has little impact on the performance of balanced datasets at the expense of increased inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.628, + 0.554 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.895, + 0.779 + ], + "angle": 0, + "content": "This paper proposes learning to discover a super-class graph for long-tailed visual recognition. The proposed superclass graph could rectify and refine the original features by message passing, which results in attending to the most relevant entities based on their semantic similarity between concepts for more accurate predictions. To obtain a more informative super-class graph and more balanced image representations, we further propose to meta-learn the super-class graph based on the prototype graph from a small amount of imbalanced data. We conduct thorough ablation studies to demonstrate the effectiveness of the proposed SuperDisco and Meta-SuperDisco algorithms. The state-of-the-art performance on the long-tailed version of four datasets further substantiates the benefit of our proposal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.887 + ], + "angle": 0, + "content": "Acknowledgment This work is financially supported by the Inception Institute of Artificial Intelligence, the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19951" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, pages 6897-6907, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[2] Chris Anderson. The long tail: How endless choice is creating unlimited demand. Random House, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Samy Bengio, and Jocelyn Cloutier. Learning a synaptic learning rule. In IJCNN, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[4] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[5] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. In Nature Machine Intelligence, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[6] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.469, + 0.368 + ], + "angle": 0, + "content": "[7] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.469, + 0.41 + ], + "angle": 0, + "content": "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[9] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.494 + ], + "angle": 0, + "content": "[10] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[11] Hojat Asgarian Dehkordi, Ali Soltani Nezhad, Hossein Kashiani, Shahriar Baradaran Shokouhi, and Ahmad Aytollahi. Multi-expert human action recognition with hierarchical super-class learning. Knowledge-Based Systems, page 109091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[13] Qi Dong, Shaogang Gong, and Xiatian Zhu. Class rectification hard mining for imbalanced deep learning. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[14] Charles Elkan. The foundations of cost-sensitive learning. In IJCAI, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[15] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Computational Intelligence, 20(1):18-36, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[16] Adam R Ferguson, Jessica L Nielson, Melissa H Cragin, Anita E Bandrowski, and Maryann E Martone. Big data from small data: data-sharing in the 'long tail' of neuroscience. Nature Neuroscience, 17(11):1442-1447, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICLR, pages 1126–1135, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[18] Camille Garcin, Alexis Joly, Pierre Bonnet, Jean-Christophe Lombardo, Antoine Affouard, Mathias Chouet, Maximilien Servajean, Joseph Salmon, and Titouan Lorieul. P1@ ntnet-300k: a plant image dataset with high label ambiguity and a long-tailed distribution. In NeurIPS 2021-35th Conference on Neural Information Processing Systems, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICLR, pages 1263-1272, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.187 + ], + "angle": 0, + "content": "[20] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In CVPR, pages 15089-15098, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.189, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[21] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[22] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderlinesmote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878-887, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.315, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[24] Ruifei He, Jihan Yang, and Xiaojun Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[25] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[26] Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[27] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[28] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[30] Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In CVPR, pages 7610-7619, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[31] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In ICLR, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[32] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[33] Shyamgopal Karthik, Jérôme Revaud, and Chidrovskii Boris. Learning from long-tailed data with noisy labels. arXiv:2108.11096, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[34] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[35] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[36] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICLR, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[37] Bo Li, Yongqiang Yao, Jingru Tan, Gang Zhang, Fengwei Yu, Jianwei Lu, and Ye Luo. Equalized focal loss for dense long-tailed object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6990-6999, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[38] Kaidong Li, Nina Y Wang, Yiju Yang, and Guanghui Wang. Sgnet: A super-class guided network for image classification and object detection. In Conference on Robots and Vision, pages 127-134, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[39] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In CVPR, pages 5212-5221, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[40] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In CVPR, pages 630-639, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.472, + 0.361 + ], + "angle": 0, + "content": "[41] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.416 + ], + "angle": 0, + "content": "[42] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[43] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, 39(2):539-550, 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.53 + ], + "angle": 0, + "content": "[44] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, pages 2537-2546, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[45] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[46] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[47] Sarah Parisot, Pedro M Esperança, Steven McDonagh, Tamas J Madarasz, Yongxin Yang, and Zhenguo Li. Long-tail recognition via compositional knowledge transfer. In CVPR, pages 6939-6948, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.756 + ], + "angle": 0, + "content": "[48] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In ICCV, pages 735-744, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[50] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[51] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[52] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In ICCV, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.146 + ], + "angle": 0, + "content": "[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[54] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, 2019. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[55] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[56] Yanmin Sun, Mohamed S Kamel, Andrew KC Wong, and Yang Wang. Cost-sensitive boosting for classification of imbalanced data. Pattern Recognition, 40(12):3358-3378, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[57] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In NeurlPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.894, + 0.361 + ], + "angle": 0, + "content": "[58] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, USA, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[59] Junjiao Tian, Yen-Cheng Liu, Nathan Glaser, Yen-Chang Hsu, and Zsolt Kira. Posterior re-calibration for imbalanced datasets. In NeurlPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[60] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[61] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[62] Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[63] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[64] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3784-3793, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[65] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is in classification: A simple framework for long-tail instance segmentation. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[66] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[67] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[68] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[69] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In CVPR, pages 8659-8668, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[70] Tz-Ying Wu, Pedro Morgado, Pei Wang, Chih-Hui Ho, and Nuno Vasconcelos. Solving long-tailed recognition with deep" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19953" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "realistic taxonomic classifier. In ECCV, pages 171-189, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[71] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, pages 247-263, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[72] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[73] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5704-5713, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[74] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[75] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In CVPR, pages 2361-2370, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[76] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.47, + 0.447 + ], + "angle": 0, + "content": "[77] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[78] Yifan Zhang, Peilin Zhao, Jiezhang Cao, Wenye Ma, Junzhou Huang, Qingyao Wu, and Mingkui Tan. Online adaptive asymmetric active learning for budgeted imbalanced data. In KDD, pages 2768-2777, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[79] Peilin Zhao, Yifan Zhang, Min Wu, Steven CH Hoi, Mingkui Tan, and Junzhou Huang. Adaptive cost-sensitive online classification. IEEE Transactions on Knowledge and Data Engineering, 31(2):214-228, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[80] Xiantong Zhen, Yingjun Du, Huan Xiong, Qiang Qiu, Cees Snoek, and Ling Shao. Learning to learn variational semantic memory. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[81] Xiantong Zhen, Haoliang Sun, Yingjun Du, Jun Xu, Yilong Yin, Ling Shao, and Cees Snoek. Learning to learn kernels with variational random features. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[82] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[83] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In CVPR, pages 9719–9728, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[84] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[85] Yucan Zhou, Qinghua Hu, and Yu Wang. Deep super-class learning for long-tail distributed image classification. Pattern Recognition, 80:118-128, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[86] ZhiHua Zhou and XuYing Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on Knowledge and Data Engineering, 18(1):63-77, 2005. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[87] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In CVPR, pages 6908-6917, 2022. 8" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19954" + } + ] +] \ No newline at end of file diff --git a/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_origin.pdf b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a0ef2c91927bacc02df7d2187ae8b13844890d3 --- /dev/null +++ b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/c1ca0beb-07ce-4086-8c07-7b5c73f59b64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf6401b0458760858c8e202373e17b7ff73ac9e9e203e8278f35cd3fbe5d8093 +size 1616592 diff --git a/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/full.md b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df693f06334dfe5f4cc029b494cde21fbeee31e0 --- /dev/null +++ b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/full.md @@ -0,0 +1,318 @@ +# SuperDisco: Super-Class Discovery Improves Visual Recognition for the Long-Tail + +Yingjun Du1, Jiayi Shen1, Xiantong Zhen1,2*, Cees G. M. Snoek1 +1AIM Lab, University of Amsterdam 2Inception Institute of Artificial Intelligence + +# Abstract + +Modern image classifiers perform well on populated classes, while degrading considerably on tail classes with only a few instances. Humans, by contrast, effortlessly handle the long-tailed recognition challenge, since they can learn the tail representation based on different levels of semantic abstraction, making the learned tail features more discriminative. This phenomenon motivated us to propose SuperDisco, an algorithm that discovers super-class representations for long-tailed recognition using a graph model. We learn to construct the super-class graph to guide the representation learning to deal with long-tailed distributions. Through message passing on the super-class graph, image representations are rectified and refined by attending to the most relevant entities based on the semantic similarity among their super-classes. Moreover, we propose to meta-learn the super-class graph under the supervision of a prototype graph constructed from a small amount of imbalanced data. By doing so, we obtain a more robust super-class graph that further improves the long-tailed recognition performance. The consistent state-of-the-art experiments on the long-tailed CIFAR-100, ImageNet, Places and iNaturalist demonstrate the benefit of the discovered super-class graph for dealing with long-tailed distributions. + +# 1. Introduction + +This paper strives for long-tailed visual recognition. A computer vision challenge that has received renewed attention in the context of representation learning, as real-world deployment demands moving from balanced to imbalanced scenarios. Three active strands of work involve class re-balancing [15, 22, 32, 43, 65], information augmentation [34, 51, 54] and module improvement [29, 31, 76]. Each of these strands is intuitive and has proven empirically successful. However, all these approaches seek to improve the classification performance of the original feature space. In this paper, we instead explore a graph learning algorithm to discover the imbalanced super-class space hidden in the original feature representation. + +![](images/3e749440ec9f0eb8901b26a268a20b0131ab105ebf5edf7e9cbf59627b9a42ac.jpg) +(a) 100 original classes + +![](images/eee3c5a6ed26573b9ce27470e8e7f07790888c39416c22245be44f1bf5a129e2.jpg) +(b) 20 ground truth super-classes + +![](images/10d07d7e9e3e83fbc64f71685a36825e3f9d556f081d90e7f11b3f7b2279bf06.jpg) +(c) 16 discovered super-classes + +![](images/8cb7e4f30c411cd68eaecc5cda9fd3cc86877cd3b2fa2e2b65a67057277a5f20.jpg) +(d) 32 discovered super-classes +Figure 1. SuperDisco learns to project the original class space (a) into a relatively balanced super-class space. Different color curves indicate the different imbalance factors on the long-tailed CIFAR-100 dataset. Like the 20 super-class ground truth (b) our discovered super-classes for 16 super-classes (c) or 32 super-classes (d) provide a much better balance than the original classes. + +The fundamental problem in long-tailed recognition [18, 32, 44, 77] is that the head features and the tail features are indistinguishable. Since the head data dominate the feature distribution, they cause the tail features to fall within the head feature space. Nonetheless, humans effortlessly handle long-tailed recognition [2, 16] by leveraging semantic abstractions existing in language to gain better representations of tail objects. This intuition hints that we may discover the semantic hierarchy from the original feature space and use it for better representations of tail objects. Moreover, intermediate concepts have been shown advantageous for classification [5, 36] by allowing the transfer of shared features across classes. Nevertheless, it remains unexplored to exploit intermediate super-classes in long-tailed visual recognition that rectify and refine the original features. + +In the real world, each category has a corresponding superclass, e.g., bus, taxi, and train all belong to the vehicle superclass. This observation raises the question: are super-classes of categories also distributed along a long-tail? We find empirical evidence that within the super-class space of popular datasets, the long-tailed distribution almost disappears, and each super-class has essentially the same number of samples. + +In Figure 1, we show the number of training samples for each of the original classes and their corresponding super-classes in the long-tailed CIFAR-100 dataset. We observe the data imbalance of super-classes is considerably lower than those of the original classes. This reflects the fact that the original imbalanced data hardly affects the degree of imbalance of the super-classes, which means the distribution of the super-classes and original data is relatively independent. These balanced super-class features could be used to guide the original tail data away from the dominant role of the head data, thus making the tail data more discriminative. Therefore, if the super-classes on different levels of semantic abstraction over the original classes can be accurately discovered, it will help the model generalize over the tail classes. As not all datasets provide labels for super-classes, we propose to learn to discover the super-classes in this paper. + +Inspired by the above observation, we make in this paper two algorithmic contributions. First, we propose in Section 3 an algorithm that learns to discover the super-class graph for long-tailed visual recognition, which we call SuperDisco. We construct a learnable graph that discovers the super-class in a hierarchy of semantic abstraction to guide feature representation learning. By message passing on the super-class graph, the original features are rectified and refined, which attend to the most relevant entities according to the similarity between the original image features and super-classes. Thus, the model is endowed with the ability to free the original tail features from the dominance of the head features using the discovered and relatively balanced super-class representations. Even when faced with the severe class imbalance challenges, e.g., iNaturalist, our SuperDisco can still refine the original features by finding a more balanced super-class space using a more complex hierarchy. As a second contribution, we propose in Section 4 a meta-learning variant of our SuperDisco algorithm to discover the super-class graph, enabling the model to achieve even more balanced image representations. To do so, we use a small amount of balanced data to construct a prototype-based relational graph, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. Last, in Section 5 we report experiments on four long-tailed benchmarks: CIFAR-100-LT, ImageNet-LT, Places-LT, and iNaturalist, and verify that our discovered super-class graph performs better for tail data in each dataset. Before detailing our contributions, we first embed our proposal in related work. + +# 2. Related work + +Long-tailed recognition. Several strategies have been proposed to address class imbalance in recognition. We categorize them into three groups. Those in the first group are based on class re-balancing [8, 30, 44, 75], which balance the training sample numbers of different classes during model training. Class re-balancing methods also could be catego- + +rized into three different groups, i.e., re-sampling [22, 32, 43, 67], cost-sensitive learning [14, 37, 56, 78, 79, 86] and logit adjustment [27, 45, 57, 59]. Class re-balancing methods improve the performance of the tail classes at the expense of the performance of the head classes. The second group is based on information augmentation, introducing additional information into model training to improve long-tailed learning performance. We identify four kinds of methods in the information augmentation scope, i.e., transfer learning, which includes head-to-tail knowledge transfer [6, 42, 64, 73], knowledge distillation [28, 40, 71], model pre-training [9, 33, 72] and self-training [24, 68, 74]. The third group focuses on improving network modules in long-tailed learning. This group includes representation learning [13, 46, 76], classifier learning [32, 41, 42, 69, 73], decoupled training [31, 32, 82], and ensemble learning [20, 83]. These methods introduce additional computation costs for increased performance. Our method belongs to the third group as it aims to learn a better representation of unbalanced training samples by the superclass graph, which is unexplored for long-tail recognition. + +Super-class learning. Super-class learning adds superclass labels as intermediate supervision into traditional deep learning. A super-class guided network [38] integrated the high-level semantic information into the network for image classification and object detection, which took two-level class annotations that contain both super-class and finer class labels. In [11], a two-phase multi-expert architecture was proposed for still image action recognition, which includes fine-grained and coarse-grained phases. However, they leveraged the ground truth of the super-class as supervision during the coarse-grained phase. Wu et al. [70] propose a taxonomic classifier to address the long-tail recognition problem, which classified each sample to the level that the classifier is competent. Zhou et al. [85] clustered the original categories into super-classes to produce a relatively balanced distribution in the super-class space, which also leveraged the ground truth of the super-class in the training phase. In contrast with the previous super-class learning, we do not use ground truth to group the original categories into the super-class space. To the best of our knowledge, no work exists that relies on graph learning to discover the super-class for long-tailed visual recognition, thus motivating this work. + +Graph neural networks. Recently, several graph neural network models (GNN) have been proposed to exploit the structures underlying graphs to benefit a variety of applications. There are two main research lines of GNN methods: non-spectral methods and spectral methods. The spectral methods [4, 10, 26, 35] focus on learning graph representations in a spectral domain, in which the learned filters are based on Laplacian matrices. The non-spectral methods [21, 63] develop an aggregator to aggregate a local set of features. Note that, message passing [19] is a key mechanism that allows GCNs and other graph neural networks to capture + +![](images/c833d2254310196ad13978c626cb615ac36571b82aa1f3734e163c9f4177c52e.jpg) +(a) Illustration of proposed SuperDisco + +![](images/8f91be0e202c4f80185112a6318b335b79fc5fcf749edabb6995ad2031be21b8.jpg) + +![](images/1a9511f57fd543c1a448cd45cd3fbc81b8339c52d175aad5c94022071d802859.jpg) +(c) Visualization of discovered super-class graph $c^2$ +Figure 2. Illustration of proposed SuperDisco (a) and visualization of the similarity between the classes and discovered super-class at different levels (b), (c). In (a), we show two levels of super-class graphs $\mathcal{C}^1$ and $\mathcal{C}^2$ . The colour in each graph represents the discovered super-class. SuperDisco discovers the potential super-class at different levels hidden in each category from (b) and (c). $\mathcal{C}^1$ roughly categorizes the original classes into three relatively balanced super-classes, and $\mathcal{C}^2$ then finely categorize them into four more balanced super-classes. + +complex relationships and dependencies between nodes in a graph, and is a major reason why they have been successful in a variety of tasks involving graph-structured data. Our method belongs to the non-spectral methods, which leverage a GNN as the base architecture to discover the super-class representation. Our proposed super-class graph would refine and rectify the original imbalanced feature to a relatively balanced feature space, which has not been explored for long-tail recognition either. + +Meta-learning for the long-tail. Meta-learning or learning to learn [3, 53, 58, 80, 81], is a learning paradigm where a model is trained on the distribution of tasks so as to enable rapid learning on new tasks. Ren et al. [51] first proposed meta-learning for the long-tailed problem by reweighting training examples. Shu et al. [54] proposed Meta-weight-Net to adaptively extract sample weights to guarantee robust recognition in the presence of training data bias. Li et al. [39] introduced meta-semantic augmentation for long-tailed recognition, which produces diversified augmented samples by translating features along many semantically meaningful directions by meta-learning. Our uniqueness is that our model aims to discover an improved super-class representation by meta-learning, which enables the original feature representation to adjust its corresponding higher-level super-class space. + +# 3. Learning to discover the super-class graph + +In this section, we discuss how to learn to discover the super-class graph from the training samples and then expand on how to leverage such a graph to benefit the unbalanced data by refining the feature representations of samples. The + +overall illustration of SuperDisco and a visualization of the discovered super-class hierarchy are shown in Figure 2. + +Preliminary. For long-tailed visual recognition, the goal is to learn an image classification model from an imbalanced training set and to evaluate the model on a balanced test set. We first define the notation for long-tailed recognition used throughout our paper. We define a training input $x_{k} \in \mathbb{R}, i \in \{1, \dots, n\}$ , where $n$ is the number of training samples in the dataset. The corresponding labels are $y_{k} \in \{1, 2, \dots, C\}$ where $C$ is the number of classes. Let $n_{j}$ denote the number of training samples for the class $j$ . Here, we assume that $n_{i} \geq n_{j}$ when $i < j$ shows the long-tailed problem simply. In this work, we typically consider a deep network model with three main components: a feature extractor $f(\cdot)$ , a proposed graph model $g(\cdot)$ and a classifier $h(\cdot)$ . The feature extractor $f(\cdot)$ first extracts an image representation as $\mathbf{z} = f(x; \theta)$ , which is then fed into the proposed graph model to refine a new representation as $g(\mathbf{z}; \phi) = \tilde{\mathbf{z}}$ . The final class prediction $\tilde{y}$ is given by a classifier function $h(\cdot)$ , i.e., $\tilde{y} = \arg \max h(\tilde{\mathbf{z}}; \psi)$ . Before detailing our approach, we add Table 1 to detail the meaning of each symbol for easy lookup. + +SuperDisco. We construct the super-class graph to organize and distill knowledge from the training process. The vertices represent different types of super-classes (e.g., the common contour between birds and airplanes) and the edges are automatically constructed to reflect the relationship between different super-classes. Our super-class graph contains multiple levels, which is closer to the relationship between various objects in the real world. Before detailing the structure, we first explicate why the multi-level super-class graphs are preferred over a flat super-class graph: a single level of + +super-class groups is likely insufficient to model complex task relationships in real-world applications; for example, the similarities among different bird species are high, but there are also similarities between birds and mammals, e.g., they are both animals. + +We assume the vertex representation $g$ as $\mathbf{h}^g\in \mathbb{R}^d$ and define the super-class graph as $\mathcal{C}^l = (\mathbf{H}_c^l,\mathbf{A}_c^l)$ , where $\mathbf{H}_c^l = \{\mathbf{h}^{i^l}|\forall i^l\in [1,C^l ]\} \in \mathbb{R}^{C^l\times d}$ is the vertex feature matrix of the $l$ -th super-class level and $\mathbf{A}_c^l = \{A_c^l (\mathbf{h}^{i^l},\mathbf{h}^{j^l})|\forall i^l,j^l\in [1,C^l ]\} \in \mathbb{R}^{C^l\times C^l}$ is the vertex adjacency matrix in the $l$ -th super-class level, $C^l$ denotes the number of vertices in the $l$ -th super-class level. Our vertex representation $\mathbf{H}_c^l$ of the super-class graph is defined to get parameterized and learned during training. The initial vertex representations of each super-class level are randomly initialized, which encourages diversity of the discovered super-classes. + +Next, we introduce how to compute the edge weight $A_{\mathcal{C}}^{l}$ in the super-class graph. The edge weight $A_{\mathcal{C}}^{l}(\mathbf{h}^{i^{l}}, \mathbf{h}^{j^{l}})$ between a pair of vertices $i$ and $j$ is gauged by the similarity between them. Formally: + +$$ +A _ {\mathcal {C}} ^ {l} \left(\mathbf {h} ^ {i l}, \mathbf {h} ^ {j l}\right) = \sigma \left(\mathbf {W} _ {c} ^ {l} \left(| \mathbf {h} ^ {i l} - \mathbf {h} ^ {j l} | / \gamma_ {c} ^ {l}\right) + \mathbf {b} _ {c} ^ {l}\right), \tag {1} +$$ + +where $\mathbf{W}_c^l$ and $\mathbf{b}_c^l$ indicate learnable parameters of the $l$ -th super-class level, $\gamma_c^l$ of $l$ -th super-class level is a scalar and $\sigma$ indicates the Sigmoid function, which normalizes the weight between 0 and 1. To adjust the representation of training samples by the involvement of super-classes, we first query the training samples in the super-class graph to obtain the relevant super-class. In light of this, we construct a new graph $\mathcal{R}$ , which adds the original sample feature $\mathbf{z}$ to the super-class graph. We define $\mathbf{z}^l$ as the refined feature after the $l$ -th super-class graph. Here we define graph $\mathcal{R}^l = (\mathbf{H}_{\mathcal{R}}^l, \mathbf{A}_{\mathcal{R}}^l)$ , where $\mathbf{H}_{\mathcal{R}}^l = \{[\mathbf{z}^l, \mathbf{h}^{i^l}] | \forall i^l \in [1, C^l]\} \in \mathbb{R}^{(C^l + 1) \times d}$ denotes the vertex feature matrix of the $l$ -th super-class level, and $\mathbf{A}_{\mathcal{R}}^l = \{[A_{\mathcal{R}}^l (\mathbf{h}^{i^l}, \mathbf{z}^1), A_{\mathcal{R}}^l (\mathbf{h}^{i^l}, \mathbf{h}^{j^l})] | \forall i^l, j^l \in [1, C^l]\} \in \mathbb{R}^{C^l + 1 \times C^{l + 1}}$ denotes the vertex adjacency matrix in the $l$ -th super-class level. The link between $\mathbf{z}^l$ and vertex $\mathbf{h}^i$ in the hierarchical graph is constructed by their similarity. In particular, analogous to the definition of weight in the super-class graph in Eq. (1), the weight $A_{\mathcal{R}}^l (\mathbf{h}^{i^l}, \mathbf{z}^1)$ is constructed as: + +$$ +A _ {\mathcal {R}} ^ {l} \left(\mathbf {h} ^ {i ^ {l}}, \mathbf {z} ^ {1}\right) = \sigma \left(\mathbf {W} _ {r} ^ {l} \left(\left| \mathbf {h} ^ {i ^ {l}} - \mathbf {z} ^ {1} \right| / \gamma_ {r} ^ {l}\right) + \mathbf {b} _ {r} ^ {l}\right), \tag {2} +$$ + +where $\mathbf{W}_r^l$ and $\mathbf{b}_c^r$ indicate learnable parameters of the $l$ -th super-class level, $\gamma_r^l$ of the $l$ -th super-class level is a scalar. + +After constructing the new graph $\mathcal{R}$ , we propagate the most relevant super-class by message passing [19] from the discovered super-classes $\mathcal{C}$ to the features $\mathbf{z}^l$ by introducing a Graph Neural Network (GNN). The message passing operation over the graph is formulated as: + +$$ +\mathbf {H} _ {\mathcal {R}} ^ {(m + 1)} = \mathrm {M P} \left(\mathbf {A} _ {\mathcal {R}} ^ {l}, \mathbf {H} _ {\mathcal {R}} ^ {(m)}; \mathbf {W} ^ {(m)}\right), \tag {3} +$$ + +where $\mathbb{M}\mathbb{P}(\cdot)$ is the message passing function, $\mathbf{H}^{(m)}$ is the vertex embedding after $m$ layers of GNN and $\mathbf{W}^{(m)}$ is a + +
NotationDescription
hVertex representation
zOriginal sample feature
CSuper-class graph
HcVertex feature matrix of Cl
AcVertex adjacency matrix of Cl
RGraph which adds z to graph Cl
PPrototype graph
CpVertex feature matrix of P
ApVertex adjacency matrix of P
SSuper graph which connecting P and C
AVertex feature matrix of S
MVertex adjacency matrix of S
+ +Table 1. Summary of the core notation used for SuperDisco. + +learnable weight matrix of layer $m$ . After stacking $M$ GNN layers, we get the information-propagated feature representation $\tilde{\mathbf{z}}^L$ for each level of the super-class graph $\mathcal{C}$ . Once we obtain the refined representation $\tilde{\mathbf{z}}^L$ for a training sample by the super-class graph, we feed them into the classifier to make the predictions and compute the corresponding loss, i.e., Cross-entropy loss for optimization. Using gradient descent, we then update the super-class graph $\mathcal{C}$ . To be able to discover a more accurate super-class graph in the face of severe imbalance problems, we propose meta-learning super-class graph discovery in the next section. + +# 4. Meta-learning super-class graph discovery + +To explore and exploit a more accurate and richer superclass graph, we propose the Meta-SuperDisco to discover the super-class graph using meta-learning, making the model more robust. In the traditional meta-learning setting [17, 50], it includes meta-training tasks and meta-test task. Each task contains a support set $S$ and a query set $Q$ . Each task is first trained by $S$ to get the task-specific learner and $Q$ optimizes this learner to update the meta-learner. For long-tailed recognition with meta-learning, previous works [51, 54] randomly sample a small amount of balanced data denoted as $\mathcal{M}$ . The imbalanced data and the small balanced data can be seen as $S$ and $Q$ in the training phase. The goal of meta-learning for long-tailed recognition is to use a small set of balanced data to optimize the model obtained from unbalanced data. We follow [51, 54] by randomly selecting the same number of samples (e.g., ten) as $\mathcal{M}$ per class from the training set. + +Meta-SuperDisco. To meta-learn the super-class graph, we construct a prototype graph $\mathcal{P}$ from $\mathcal{M}$ , since $\mathcal{M}$ is a balanced dataset. The prototype graph extracts the sample-level relation information, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. For the prototype graph, we need to compute the prototype of each category [55], which is defined as: $\mathbf{c}^k = \frac{1}{N^k}\sum_{i = 1}^{N^k}\mathbf{z}_j$ , where $N^k$ denotes the number of samples in class $k$ , $\mathbf{z}_j$ is the feature from $\mathcal{M}$ . + +After calculating all prototype representations $\{\mathbf{c}^k |\forall k\in [1,K]\}$ , which serve as the vertices in the prototype graph $\mathcal{P}_i$ , we further need to define the edges and the corresponding + +edge weights. The edge weight $A_{\mathcal{P}}(\mathbf{c}^i,\mathbf{c}^j)$ between two prototypes $\mathbf{c}^i$ and $\mathbf{c}^j$ is gauged by the similarity between them. The edge weight is calculated as follows: + +$$ +A _ {\mathcal {P}} \left(\mathbf {c} ^ {i}, \mathbf {c} ^ {j}\right) = \sigma \left(\mathbf {W} _ {p} \left(\left| \mathbf {c} ^ {i} - \mathbf {c} ^ {j} \right| / \gamma_ {p}\right) + \mathbf {b} _ {p}\right), \tag {4} +$$ + +where $\mathbf{W}_p$ and $\mathbf{b}_p$ are the learnable parameters, $\gamma_p$ is a scalar. Thus, we denote the prototype graph as $\mathcal{P} = (\mathbf{C}_{\mathcal{P}}, \mathbf{A}_{\mathcal{P}})$ , where $\mathbf{C}_{\mathcal{P}_i} = \{\mathbf{c}^i | \forall i \in [1, K]\} \in \mathbb{R}^{K \times d}$ represent a set of vertices, with each one corresponding to the prototype from a class, while $\mathbf{A}_{\mathcal{P}} = \{|A_{\mathcal{P}}(\mathbf{c}^i, \mathbf{c}^j)| \forall i, j \in [1, K]\} \in \mathbb{R}^{K \times K}$ gives the adjacency matrix, which indicates the proximity between prototypes. We then use the prototype graph to guide the learning of the meta super-class graph. We construct a super graph $\mathcal{S}$ by connecting prototype graph $\mathcal{P}$ to super-class graph $\mathcal{C}$ . In the super graph $\mathcal{S}$ , the vertices are $\mathbf{M} = (\mathbf{C}_{\mathcal{P}^l}^l; \mathbf{H}_{\mathcal{C}^l}^l)$ , where $\mathcal{P}^l$ denotes the refined prototype graph vertex after the $l$ -th level super-class graph. Then, we calculate the link weight $A_{\mathcal{S}}^l(c^i, \{\mathbf{h}^j\})$ of the super graph as: + +$$ +A _ {S} ^ {l} \left(\mathbf {c} ^ {i}, \mathbf {h} ^ {j ^ {l}}\right) = \frac {\exp \left(- \| \left(\mathbf {c} ^ {i} - \mathbf {h} ^ {j ^ {l}}\right) / \gamma_ {s} ^ {l} \| _ {2} ^ {2} / 2\right)}{\sum_ {j ^ {l ^ {\prime}} = 1} ^ {J} \exp \left(- \| \left(\mathbf {c} ^ {i} - \mathbf {h} ^ {j ^ {l ^ {\prime}}}\right) / \gamma_ {s} ^ {l} \| _ {2} ^ {2} / 2\right)}, \tag {5} +$$ + +where $\gamma_s^l$ is a scaling factor. Note that, here we use softmax to ensure that the total weight of edges between the prototype graph $\mathcal{P}$ and the super-class graph $\mathcal{C}$ is equal to 1, giving the prototype graph a unique influence on the expression of each super-class. Thus, the adjacent matrix and feature matrix of the super graph $S^l = (\mathbf{A}^l,\mathbf{M}^l)$ is defined as $\mathbf{A}^l = (\mathbf{A}_{\mathcal{P}},\mathbf{A}_S^l;\mathbf{A}_S^{l^\mathrm{T}},\mathbf{A}_\mathcal{C}^l)$ and $\mathbf{M}^l = (\mathbf{C}_{\mathcal{P}^l}^l;\mathbf{H}_{\mathcal{C}^l}^l)$ . Once we constructed the super graph $S$ , we use message-passing again to propagate the most relevant knowledge from the prototype graph $\mathcal{P}$ to the super-class graph $\mathcal{C}$ . Similar to eq. (3): + +$$ +\mathbf {M} ^ {(m + 1)} = \mathbb {M P} \left(\mathbf {A} ^ {l}, \mathbf {M} ^ {(m)}; \mathbf {W} ^ {(m)}\right). \tag {6} +$$ + +We leverage the graph $S$ to refine the super-class graph. Finally, we feed the original feature $\mathbf{z}$ into the super-class graph to get the information-propagated feature representation $\tilde{\mathbf{z}}^L$ , which refines the original feature by its corresponding discovered super-classes. We provide the complete SuperDisco and Meta-SuperDisco algorithm specifications in the supplemental material. + +# 5. Experiments + +Datasets. We apply our method to four commonly used long-tail recognition benchmarks. Sample images and the number of categories for all datasets are provided in the supplement material. CIFAR-100-LT reduces the number of training samples per class according to an exponential function $n = n_i \mu^i$ , where $i$ is the class index, $n_i$ is the original number of training samples, and $\mu \in (0, 1)$ . The imbalance factor of a dataset is defined as the number of training samples in the most populated class divided by the minority class. We consider imbalance factors {10, 50, 100}. ImageNet-LT [44] is a subset of ImageNet [12] consisting of + +
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
SuperDisco65.960.757.250.945.2
Meta-SuperDisco68.563.158.353.847.5
+ +Table 2. Benefit of SuperDisco and Meta-SuperDisco. SuperDisco achieves better performance compared to a baseline fin-tuning on all imbalance factors, while Meta-SuperDisco is even better for long-tailed recognition. + +115.8K images from 1000 categories, with maximally 1,280 images per class and minimally 5 images per class, and a balanced test set. Places-LT [44] has an imbalanced training set with 62,500 images for 365 classes from Places [84]. It contains images from 365 classes and the number of images per class ranges from 4980 to 5. The test sets are balanced and contain 100 images per class. iNaturalist [61] is a real-world long-tailed dataset with 675,170 training images for 5,089 classes, where the top $1\%$ most populated classes contain more than $16\%$ of the training images. Additionally, there is also a severe imbalance among the super-classes of iNaturalist. The 13 ground truth super-classes images range from 158,407 to 308. + +Implementation details. We follow [32] by first training a feature extractor with instance-balanced sampling, and then training our graph model and classifier based on the trained features. For CIFAR-100-LT, we follow [54] and use a ResNet-32 backbone. For ImageNet-LT, we use ResNeXt-50 [23] as our backbone, following [32]. For Places-LT, we report results with ResNet-152 following [44]. For iNaturalist, we use a ResNet-50 backbone. We train each dataset for 200 epochs with batch size 512. We use random left-right flipping and cropping as our training augmentation. For all experiments, we use an SGD optimizer with a momentum of 0.9 and a batch size of 512. We randomly selected 10 images per class from the training set for all datasets as $\mathcal{M}$ . Code available at: https://github.com/Yingjun-Du/SuperDisco. + +Benefit of SuperDisco and Meta-SuperDisco. To show the benefit of SuperDisco, we compare it with a fine-tuning baseline, which retrans the classifier only. Table 2 shows SuperDisco improves over fine-tuning on CIFAR-100-LT, and the results for the other long-tailed datasets are provided in the supplemental materials Table 1. In the most challenging setting with the largest imbalance factor of 200, our SuperDisco delivers $45.2\%$ , surpassing the baseline by $5.9\%$ . We attribute improvement to our model's ability to refine original features, allowing the discovered super-class graph to guide the tail features away from the dominant role of head features, thus leading to improvements over the original features. We also investigate the benefit of meta-learning with Meta-SuperDisco. The Meta-SuperDisco consistently surpasses the SuperDisco for all imbalance factors. The con + +
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
(20)61.260.149.947.341.9
(2, 4, 8)65.362.753.149.843.2
(4, 8, 16)69.164.255.252.345.9
(4, 8, 16, 32)68.563.158.353.847.5
(4, 8, 16, 32, 64)66.962.758.952.946.3
Oracle super-classes66.963.254.751.443.2
+ +Table 3. Effect of number of super-class levels on CIFAR-100-LT. Compared to a baseline [32] and an oracle setting, Meta-SuperDisco provides higher performance gains with more complex hierarchies. + +
ManyMediumFewAll
Baseline65.066.365.565.9
(13)71.870.266.170.8
(2,4,8)70.569.365.969.4
(4,8,16)72.270.966.470.3
(4,8,16,32)73.670.267.370.9
(4,8,16,32,64)73.472.968.372.3
(4,8,16,32,64,128)72.171.366.270.9
Oracle super-classes70.770.565.970.2
+ +Table 4. Effect of number of super-class levels on iNaturalist. Meta-SuperDisco achieves consistent performance gains with more complex hierarchies. + +sistent improvements confirm that Meta-SuperDisco learns even more robust super-class graphs, leading to a discriminative representation of the tail data. + +Effect of the number of super-class levels. A significant challenge with any structure-aware learning algorithm is determining the appropriate complexity for the knowledge structure. So, we further analyze the effect of the super-class hierarchies, including the level (number of depths L) or the number of super-classes in each level. The results are shown in Table 3 and Table 4. The super-class number from the bottom layer to the top layer is saved in a tuple. For example, (2, 4, 8) represents three depth, with two super-classes in the top layer. The baseline is Decouple-LWS [32], which only inputs the original feature to learn a new classifier. The oracle super-classes are first trained on two long-tailed datasets using the ground truth super-class labels for super-class classifications. Once the training is completed, each oracle super-class is obtained by averaging the samples of each super-class. We constructed a one-layer super-class graph using these super-classes, where the vertices of the graph are for each super-class, and the edges of the graph are computed according to Eq. (1). Then, we use the message passing by Eq. (3) to refine the original features and input them into the classifier to get the final predictions. From Table 3, we observe that using oracle super-classes achieves better performance compared to the learned super-class (20) since it uses the ground truth super-classes as supervision. We also conclude that too few levels may not be enough + +![](images/ad75c3bb258313c39fce850e20694d6365887bcf6881cc9730db255e0f253392.jpg) +(a) SuperDisco + +![](images/40d1be7db1d59e8e1c62c8fe9d07dabbf7f39b6a9a571b7be13daaabdb088a17.jpg) +(b) Meta-SuperDisco +Figure 3. Similarity between discovered super-classes and classes. SuperDisco discovers super-classes hidden in each class, while Meta-SuperDisco discovers more accurate super-classes. + +to learn the precise super-classes (e.g., tuple (20) or (2, 4, 8)). In this dataset, increasing levels (e.g., tuple (4, 8, 16, 32)) achieves better performance on the smaller imbalance factor (e.g., 10), and similar performance compared with (4, 8, 16). For the real-world long-tailed dataset iNaturalist [62] in Table 4, we find no significant improvement for the few-shot classes in the performance of the oracle super-class compared to the baseline, and the same is true for the performance of the discovered super-class structure (13). This is because the super-class of iNaturalist also have serious long-tailed problems, resulting in the refined features of tail classes remaining indistinguishable from the refined features of head classes. However, with a more complex graph structure (4, 8, 16, 32, 64), the few-shot performance improves by a good margin compared with the baseline, and even the oracle super-classes. We attribute this to our model's ability to explore relatively balanced super-class spaces, thus making the refined tail category features discriminative. By comparing Table 3 and Table 4, we conclude that deeper as well as wider graphs are needed to discover the super-classes in the case of severe class imbalance. + +Visualization of SuperDisco. To understand the meaning of the discovered super-classes more clearly, we present a visualization in Figure. 3. We selected 12 different categories from the CIFAR-100 test dataset. We calculate the similarity of each of these 12 categories to the different vertices in the graph we explore. Here we show the similarity with the second layer of graph vertices (C1, C2, C3, C4). We can see different categories mainly activate different vertices, e.g., bus $\rightarrow$ C3 and road $\rightarrow$ C2. As shown in this heatmap, we find that C1 reflects the super-class of flowers, C2 reflects the super-class of buildings, C3 reflects the super-class of vehicles, C4 reflects the super-class of fish. Another observation is that the second-largest activated super-class is also meaningful, promoting knowledge transfer between super-classes. For example, road and bridge are related to the C3 superclass, since some vehicles may be on the road and bridge. This visualization reflects that we can use graph models to discover the super-classes and the relationships between each super-class. We also visualize the discovered meta-learning super-classes in Figure 3 (b). The discovered super-classes are even more accurate, e.g., roses have high similarity to C2, which mainly reflects the buildings super-class, while it has high similarity to C1, which is the flowers super-class. This once again validates the benefit of Meta-SuperDisco. + +![](images/86d685881b3449474338255a4a8358b62b9a0b0cc79def00f46464450880b6b9.jpg) +(a) Original features + +![](images/af99ecc2e2d54628f761f3c66477bea79fcd1fef79a2ee50ae03a11ea59bd622.jpg) +(c) Refined features by SuperDisco + +![](images/f6020ca1af763ca7a9554ee400ba03cfdc5304c11cb1dbd37ac8e6f616fc3beb.jpg) + +![](images/198d2ca7405cd4c2eb306a90c989b2a274cebb29f0a75a7f2c26cc031467bf12.jpg) + +![](images/1e90b07d166fe11b6d1af84e15ae336f85caad00ce4b7b45b5f995dafdea392f.jpg) +(b) Training distribution + +![](images/d8ae94c57d2f09d1dab25a8d230eb70e3f03f8467354aec06464f7efca0f094f.jpg) +(d) Refined features by Meta-SuperDisco + +![](images/8afc546fa9283e628ad771f991a32aee87eaa912a683b333d8fe5816e712f0c1.jpg) + +![](images/9b06e08889990f3daf85c823d38d35ef5d74c93d14d47808e3fcce43d97b6976.jpg) + +![](images/e0a0492f9a18cb8a7a45dd84b0102d813c0e4e1a9757b0ee80f79f430448c340.jpg) +Figure 4. Visualization of refined features on CIFAR-100-LT, with the original features (a) and their corresponding training distribution (b). Colours indicate categories. SuperDisco (c) guides the original features on being clustered into the corresponding super-class space at different levels, while Meta-SuperDisco (d) obtains even more discriminative intra-class features. +Figure 5. Effect of refined features. Accuracy increases along with the increased super-class levels, revealing that more accurate and richer super-classes facilitate better long-tailed recognition. + +Furthermore, in the (c) learned the hierarchical concept of each class, we can see that bus and bridge have the same concept C2 in the last concept level, which may be due to the possible presence of cars on the bridge. + +Visualization of refined features. To understand the empirical benefit of SuperDisco, we visualize in Figure 4 the original features and refined features with super-class graphs of the different levels using t-SNE [60]. We choose the vertices numbers as (2, 4, 6), meaning the super-class graph has three different levels, each with a different number of vertices. The original features of the category with a small sample size will overlap with the (original) features of the category with a large sample size. Super-class graphs discovered by our model project the original features into a high-level super-class space, where the imbalance is relatively small. Hence, its corresponding subcategory can be predicted more accurately. It is worth noting that when comparing the two different super-class graphs on top and below, the features obtained by Meta-SuperDisco are even more distinctive and distant from each other. To better measure the goodness of the refined features obtained at different levels, + +
ManyMediumFewAll
Baseline58.449.334.852.7
Multi-layer perceptron63.551.835.955.0
Graph convolution network66.153.337.157.1
+ +Table 5. Analysis of super-class mechanism on ImageNet-LT. The super-class mechanism contributes most, the graph convolution network improves results further. + +we show in Figure 5 the prediction accuracy using different refined features. We find that the accuracy increases along with the increased super-class levels, which shows that using more accurate and richer super-classes facilitates better performance. This again demonstrates that Meta-SuperDisco is most suitable for long-tailed visual recognition. + +Analysis of super-class mechanism. To demonstrate that the improved performance of our SuperDisco cannot solely be attributed to the graph convolutional network module, we conducted an experiment where we replaced it with a multi-layer perceptron to obtain the representation per sample. In Table 5, the performance gains of our method are primarily due to the super-classes rather than the graph convolution network. The results suggest that incorporating the super-classes mechanism plays a crucial role in improving the performance of long-tailed problems. Furthermore, the results improve further when we replace the multi-layer perceptron with our graph convolution network module. + +Comparison with the state-of-the-art. We evaluate our method on the four long-tailed datasets under different imbalance factors in Table 6 and 7. Our model achieves state-of-the-art performance on the tail data of all datasets. For ImageNet-LT, our model achieves state-of-the-art performance on both few-shot and all data. In the most challenging Places-LT, our model delivers $40.3\%$ on all classes, surpassing the second-best Parisot et al. [47] by $1.1\%$ . On the real- + +
VenueImageNet-LTPlaces-LTiNaturalist
ManyMediumFewAllManyMediumFewAllManyMediumFewAll
Kang et al. [32]ICLR 1960.247.230.349.940.639.128.637.665.066.365.565.9
Kang et al. [31]ICLR 2161.849.430.951.5-------68.6
He et al. [25]ICCV 2164.150.431.553.1----70.670.167.669.1
Li et al. [40]CVPR 2166.851.135.456.0-------69.3
Samuel et al. [52]ICCV 2164.049.833.153.5-------69.7
Alshammari et al. [1]CVPR 2262.550.441.553.9----71.270.469.770.2
Zhang et al. [75]CVPR 2161.352.231.452.940.442.430.139.369.071.170.270.6
Parisot et al. [47]CVPR 2263.252.136.954.139.741.034.939.2----
Park et al. [48]CVPR 2266.453.935.656.2----73.172.668.772.8
This paper66.153.337.157.145.342.835.340.372.372.971.373.6
+ +Table 6. Comparison with the state-of-the-art on ImageNet-LT, Places-LT and iNaturalist. Best and second best results are highlighted in bold and italic bold. Our Meta-SuperDisco achieves either better or comparable performance than state-of-the-art methods under the tail and all data for long-tailed visual recognition. + +
VenueImbalance ratio
1050100
Park et al. [49]ICCV 2159.547.442.0
Li et al. [40]CVPR 2162.350.546.0
Zhong et al. [82]CVPR 2162.551.546.8
Samuel et al. [52]ICCV 2163.457.647.3
Wang et al. [66]ICLR 2161.851.748.0
Zhu et al. [87]CVPR 2264.956.651.9
Cui et al. [7]ICCV 2164.256.052.0
Alshammari et al. [1]CVPR 2268.857.753.3
This paper69.358.353.8
+ +Table 7. Comparison with the state-of-the-art on CIFAR-100-LT. Our model achieves best performance. + +world long-tailed dataset iNaturalist, our model achieves the three best performances under four different shots. On the long-tailed synthetic dataset CIFAR-100-LT, our model achieves the best performance under each imbalance factor. The consistent improvements on all benchmarks under various configurations confirm that our Meta-SuperDisco is effective for long-tailed visual recognition. + +Limitations. We show that SuperDisco and Meta-SuperDisco achieve good performance on tail data while being less successful on the head data. Based on this result, we also perform an experiment on balanced CIFAR-100 in Figure 6. With SuperDisco and Meta-SuperDisco, there is only a slight change in performance at the expense of an increased inference time. This reveals that our SuperDisco does not change the original features much through message passing on a balanced dataset. This may be because the obtained super-classes are still the original class itself. In addition, as the computation of graphs involves many matrix operations, our model also requires a relatively long computational speed. Due to introducing a prototype graph and more data, Meta-SuperDisco takes longer to compute. In addition, the training time of SuperDisco and its meta variants is also 1.5 times higher than the baseline. Future work could investigate how to use the discovered super-class graph in balanced datasets and how to reduce the computation time. + +![](images/bb1ad1ce828239038659c4f2b6829fc7fb3aaff56e41c134c834b74c4ab4b136.jpg) +Figure 6. Limitation. Accuracy $(\%)$ vs. speed (ms) comparison with different methods on balanced CIFAR-100. SuperDisco has little impact on the performance of balanced datasets at the expense of increased inference time. + +# 6. Conclusions + +This paper proposes learning to discover a super-class graph for long-tailed visual recognition. The proposed superclass graph could rectify and refine the original features by message passing, which results in attending to the most relevant entities based on their semantic similarity between concepts for more accurate predictions. To obtain a more informative super-class graph and more balanced image representations, we further propose to meta-learn the super-class graph based on the prototype graph from a small amount of imbalanced data. We conduct thorough ablation studies to demonstrate the effectiveness of the proposed SuperDisco and Meta-SuperDisco algorithms. The state-of-the-art performance on the long-tailed version of four datasets further substantiates the benefit of our proposal. + +Acknowledgment This work is financially supported by the Inception Institute of Artificial Intelligence, the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy. + +# References + +[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, pages 6897-6907, 2022. 8 +[2] Chris Anderson. The long tail: How endless choice is creating unlimited demand. Random House, 2007. 1 +[3] Yoshua Bengio, Samy Bengio, and Jocelyn Cloutier. Learning a synaptic learning rule. In IJCNN, 1991. 3 +[4] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. 2 +[5] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. In Nature Machine Intelligence, 2020. 1 +[6] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 2 +[7] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In ICCV, 2021. 8 +[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 2 +[9] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, 2018. 2 +[10] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 2 +[11] Hojat Asgarian Dehkordi, Ali Soltani Nezhad, Hossein Kashiani, Shahriar Baradaran Shokouhi, and Ahmad Aytollahi. Multi-expert human action recognition with hierarchical super-class learning. Knowledge-Based Systems, page 109091, 2022. 2 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5 +[13] Qi Dong, Shaogang Gong, and Xiatian Zhu. Class rectification hard mining for imbalanced deep learning. In ICCV, 2017. 2 +[14] Charles Elkan. The foundations of cost-sensitive learning. In IJCAI, 2001. 2 +[15] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Computational Intelligence, 20(1):18-36, 2004. 1 +[16] Adam R Ferguson, Jessica L Nielson, Melissa H Cragin, Anita E Bandrowski, and Maryann E Martone. Big data from small data: data-sharing in the 'long tail' of neuroscience. Nature Neuroscience, 17(11):1442-1447, 2014. 1 +[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICLR, pages 1126–1135, 2017. 4 +[18] Camille Garcin, Alexis Joly, Pierre Bonnet, Jean-Christophe Lombardo, Antoine Affouard, Mathias Chouet, Maximilien Servajean, Joseph Salmon, and Titouan Lorieul. P1@ ntnet-300k: a plant image dataset with high label ambiguity and a long-tailed distribution. In NeurIPS 2021-35th Conference on Neural Information Processing Systems, 2021. 1 + +[19] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICLR, pages 1263-1272, 2017. 2, 4 +[20] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In CVPR, pages 15089-15098, 2021. 2 +[21] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, 2017. 2 +[22] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderlinesmote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878-887, 2005. 1, 2 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[24] Ruifei He, Jihan Yang, and Xiaojun Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In ICCV, 2021. 2 +[25] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In ICCV, 2021. 8 +[26] Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163, 2015. 2 +[27] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In CVPR, 2021. 2 +[28] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In CVPR, 2020. 2 +[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, 2016. 1 +[30] Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In CVPR, pages 7610-7619, 2020. 2 +[31] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In ICLR, 2021. 1, 2, 8 +[32] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 5, 6, 8 +[33] Shyamgopal Karthik, Jérôme Revaud, and Chidrovskii Boris. Learning from long-tailed data with noisy labels. arXiv:2108.11096, 2021. 2 +[34] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In CVPR, 2020. 1 +[35] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2 +[36] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICLR, 2020. 1 + +[37] Bo Li, Yongqiang Yao, Jingru Tan, Gang Zhang, Fengwei Yu, Jianwei Lu, and Ye Luo. Equalized focal loss for dense long-tailed object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6990-6999, 2022. 2 +[38] Kaidong Li, Nina Y Wang, Yiju Yang, and Guanghui Wang. Sgnet: A super-class guided network for image classification and object detection. In Conference on Robots and Vision, pages 127-134, 2021. 2 +[39] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In CVPR, pages 5212-5221, 2021. 3 +[40] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In CVPR, pages 630-639, 2021. 2, 8 +[41] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In ICCV, 2021. 2 +[42] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, 2020. 2 +[43] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, 39(2):539-550, 2008. 1, 2 +[44] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, pages 2537-2546, 2019. 1, 2, 5 +[45] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR, 2021. 2 +[46] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2 +[47] Sarah Parisot, Pedro M Esperança, Steven McDonagh, Tamas J Madarasz, Yongxin Yang, and Zhenguo Li. Long-tail recognition via compositional knowledge transfer. In CVPR, pages 6939-6948, 2022. 7, 8 +[48] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 8 +[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In ICCV, pages 735-744, 2021. 8 +[50] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 4 +[51] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 1, 3, 4 +[52] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In ICCV, 2021. 8 + +[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3 +[54] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, 2019. 1, 3, 4, 5 +[55] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 4 +[56] Yanmin Sun, Mohamed S Kamel, Andrew KC Wong, and Yang Wang. Cost-sensitive boosting for classification of imbalanced data. Pattern Recognition, 40(12):3358-3378, 2007. 2 +[57] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In NeurlPS, 2020. 2 +[58] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, USA, 1998. 3 +[59] Junjiao Tian, Yen-Cheng Liu, Nathan Glaser, Yen-Chang Hsu, and Zsolt Kira. Posterior re-calibration for imbalanced datasets. In NeurlPS, 2020. 2 +[60] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(11), 2008. 7 +[61] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5 +[62] Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450, 2017. 6 +[63] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 2 +[64] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3784-3793, 2021. 2 +[65] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is in classification: A simple framework for long-tail instance segmentation. In ECCV, 2020. 1 +[66] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2021. 8 +[67] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2 +[68] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In CVPR, 2021. 2 +[69] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In CVPR, pages 8659-8668, 2021. 2 +[70] Tz-Ying Wu, Pedro Morgado, Pei Wang, Chih-Hui Ho, and Nuno Vasconcelos. Solving long-tailed recognition with deep + +realistic taxonomic classifier. In ECCV, pages 171-189, 2020. 2 +[71] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, pages 247-263, 2020. 2 +[72] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In NeurIPS, 2020. 2 +[73] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5704-5713, 2019. 2 +[74] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2 +[75] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In CVPR, pages 2361-2370, 2021. 2, 8 +[76] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In ICCV, 2017. 1, 2 +[77] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1 +[78] Yifan Zhang, Peilin Zhao, Jiezhang Cao, Wenye Ma, Junzhou Huang, Qingyao Wu, and Mingkui Tan. Online adaptive asymmetric active learning for budgeted imbalanced data. In KDD, pages 2768-2777, 2018. 2 +[79] Peilin Zhao, Yifan Zhang, Min Wu, Steven CH Hoi, Mingkui Tan, and Junzhou Huang. Adaptive cost-sensitive online classification. IEEE Transactions on Knowledge and Data Engineering, 31(2):214-228, 2018. 2 +[80] Xiantong Zhen, Yingjun Du, Huan Xiong, Qiang Qiu, Cees Snoek, and Ling Shao. Learning to learn variational semantic memory. In NeurIPS, 2020. 3 +[81] Xiantong Zhen, Haoliang Sun, Yingjun Du, Jun Xu, Yilong Yin, Ling Shao, and Cees Snoek. Learning to learn kernels with variational random features. In ICML, 2020. 3 +[82] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 8 +[83] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In CVPR, pages 9719–9728, 2020. 2 +[84] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5 +[85] Yucan Zhou, Qinghua Hu, and Yu Wang. Deep super-class learning for long-tail distributed image classification. Pattern Recognition, 80:118-128, 2018. 2 +[86] ZhiHua Zhou and XuYing Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on Knowledge and Data Engineering, 18(1):63-77, 2005. 2 + +[87] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In CVPR, pages 6908-6917, 2022. 8 \ No newline at end of file diff --git a/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/images.zip b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e2a9800c4b05e08192ecc2907cb96f6fed8df058 --- /dev/null +++ b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75497c281f1242b113f4c086045cb8e56908716727c06d1e2c2dcd9440fd6f9e +size 467932 diff --git a/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/layout.json b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ffabbb4e70602fc13eaad69353afbdf13832e211 --- /dev/null +++ b/2023/SuperDisco_ Super-Class Discovery Improves Visual Recognition for the Long-Tail/layout.json @@ -0,0 +1,10333 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 545, + 120 + ], + "type": "text", + "content": "SuperDisco: Super-Class Discovery Improves Visual Recognition for the Long-Tail" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 493, + 171 + ], + "type": "text", + "content": "Yingjun Du1, Jiayi Shen1, Xiantong Zhen1,2*, Cees G. M. Snoek1 \n1AIM Lab, University of Amsterdam 2Inception Institute of Artificial Intelligence" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 191, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 290, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 499 + ], + "type": "text", + "content": "Modern image classifiers perform well on populated classes, while degrading considerably on tail classes with only a few instances. Humans, by contrast, effortlessly handle the long-tailed recognition challenge, since they can learn the tail representation based on different levels of semantic abstraction, making the learned tail features more discriminative. This phenomenon motivated us to propose SuperDisco, an algorithm that discovers super-class representations for long-tailed recognition using a graph model. We learn to construct the super-class graph to guide the representation learning to deal with long-tailed distributions. Through message passing on the super-class graph, image representations are rectified and refined by attending to the most relevant entities based on the semantic similarity among their super-classes. Moreover, we propose to meta-learn the super-class graph under the supervision of a prototype graph constructed from a small amount of imbalanced data. By doing so, we obtain a more robust super-class graph that further improves the long-tailed recognition performance. The consistent state-of-the-art experiments on the long-tailed CIFAR-100, ImageNet, Places and iNaturalist demonstrate the benefit of the discovered super-class graph for dealing with long-tailed distributions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 521, + 127, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 127, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 127, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 540, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 696 + ], + "type": "text", + "content": "This paper strives for long-tailed visual recognition. A computer vision challenge that has received renewed attention in the context of representation learning, as real-world deployment demands moving from balanced to imbalanced scenarios. Three active strands of work involve class re-balancing [15, 22, 32, 43, 65], information augmentation [34, 51, 54] and module improvement [29, 31, 76]. Each of these strands is intuitive and has proven empirically successful. However, all these approaches seek to improve the classification performance of the original feature space. In this paper, we instead explore a graph learning algorithm to discover the imbalanced super-class space hidden in the original feature representation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 206, + 403, + 274 + ], + "blocks": [ + { + "bbox": [ + 310, + 206, + 403, + 274 + ], + "lines": [ + { + "bbox": [ + 310, + 206, + 403, + 274 + ], + "spans": [ + { + "bbox": [ + 310, + 206, + 403, + 274 + ], + "type": "image", + "image_path": "3e749440ec9f0eb8901b26a268a20b0131ab105ebf5edf7e9cbf59627b9a42ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 275, + 397, + 285 + ], + "lines": [ + { + "bbox": [ + 321, + 275, + 397, + 285 + ], + "spans": [ + { + "bbox": [ + 321, + 275, + 397, + 285 + ], + "type": "text", + "content": "(a) 100 original classes" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 205, + 520, + 274 + ], + "blocks": [ + { + "bbox": [ + 427, + 205, + 520, + 274 + ], + "lines": [ + { + "bbox": [ + 427, + 205, + 520, + 274 + ], + "spans": [ + { + "bbox": [ + 427, + 205, + 520, + 274 + ], + "type": "image", + "image_path": "eee3c5a6ed26573b9ce27470e8e7f07790888c39416c22245be44f1bf5a129e2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 275, + 529, + 285 + ], + "lines": [ + { + "bbox": [ + 426, + 275, + 529, + 285 + ], + "spans": [ + { + "bbox": [ + 426, + 275, + 529, + 285 + ], + "type": "text", + "content": "(b) 20 ground truth super-classes" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 293, + 403, + 361 + ], + "blocks": [ + { + "bbox": [ + 310, + 293, + 403, + 361 + ], + "lines": [ + { + "bbox": [ + 310, + 293, + 403, + 361 + ], + "spans": [ + { + "bbox": [ + 310, + 293, + 403, + 361 + ], + "type": "image", + "image_path": "10d07d7e9e3e83fbc64f71685a36825e3f9d556f081d90e7f11b3f7b2279bf06.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 362, + 410, + 372 + ], + "lines": [ + { + "bbox": [ + 309, + 362, + 410, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 362, + 410, + 372 + ], + "type": "text", + "content": "(c) 16 discovered super-classes" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 429, + 292, + 520, + 361 + ], + "blocks": [ + { + "bbox": [ + 429, + 292, + 520, + 361 + ], + "lines": [ + { + "bbox": [ + 429, + 292, + 520, + 361 + ], + "spans": [ + { + "bbox": [ + 429, + 292, + 520, + 361 + ], + "type": "image", + "image_path": "8cb7e4f30c411cd68eaecc5cda9fd3cc86877cd3b2fa2e2b65a67057277a5f20.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 362, + 528, + 372 + ], + "lines": [ + { + "bbox": [ + 427, + 362, + 528, + 372 + ], + "spans": [ + { + "bbox": [ + 427, + 362, + 528, + 372 + ], + "type": "text", + "content": "(d) 32 discovered super-classes" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 382, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 448 + ], + "type": "text", + "content": "Figure 1. SuperDisco learns to project the original class space (a) into a relatively balanced super-class space. Different color curves indicate the different imbalance factors on the long-tailed CIFAR-100 dataset. Like the 20 super-class ground truth (b) our discovered super-classes for 16 super-classes (c) or 32 super-classes (d) provide a much better balance than the original classes." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 629 + ], + "type": "text", + "content": "The fundamental problem in long-tailed recognition [18, 32, 44, 77] is that the head features and the tail features are indistinguishable. Since the head data dominate the feature distribution, they cause the tail features to fall within the head feature space. Nonetheless, humans effortlessly handle long-tailed recognition [2, 16] by leveraging semantic abstractions existing in language to gain better representations of tail objects. This intuition hints that we may discover the semantic hierarchy from the original feature space and use it for better representations of tail objects. Moreover, intermediate concepts have been shown advantageous for classification [5, 36] by allowing the transfer of shared features across classes. Nevertheless, it remains unexplored to exploit intermediate super-classes in long-tailed visual recognition that rectify and refine the original features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "In the real world, each category has a corresponding superclass, e.g., bus, taxi, and train all belong to the vehicle superclass. This observation raises the question: are super-classes of categories also distributed along a long-tail? We find empirical evidence that within the super-class space of popular datasets, the long-tailed distribution almost disappears, and each super-class has essentially the same number of samples." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 257, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 257, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 257, + 712 + ], + "type": "text", + "content": "*Currently with United Imaging Healthcare, Co., Ltd., China." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19944" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 287, + 263 + ], + "type": "text", + "content": "In Figure 1, we show the number of training samples for each of the original classes and their corresponding super-classes in the long-tailed CIFAR-100 dataset. We observe the data imbalance of super-classes is considerably lower than those of the original classes. This reflects the fact that the original imbalanced data hardly affects the degree of imbalance of the super-classes, which means the distribution of the super-classes and original data is relatively independent. These balanced super-class features could be used to guide the original tail data away from the dominant role of the head data, thus making the tail data more discriminative. Therefore, if the super-classes on different levels of semantic abstraction over the original classes can be accurately discovered, it will help the model generalize over the tail classes. As not all datasets provide labels for super-classes, we propose to learn to discover the super-classes in this paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 264, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 609 + ], + "type": "text", + "content": "Inspired by the above observation, we make in this paper two algorithmic contributions. First, we propose in Section 3 an algorithm that learns to discover the super-class graph for long-tailed visual recognition, which we call SuperDisco. We construct a learnable graph that discovers the super-class in a hierarchy of semantic abstraction to guide feature representation learning. By message passing on the super-class graph, the original features are rectified and refined, which attend to the most relevant entities according to the similarity between the original image features and super-classes. Thus, the model is endowed with the ability to free the original tail features from the dominance of the head features using the discovered and relatively balanced super-class representations. Even when faced with the severe class imbalance challenges, e.g., iNaturalist, our SuperDisco can still refine the original features by finding a more balanced super-class space using a more complex hierarchy. As a second contribution, we propose in Section 4 a meta-learning variant of our SuperDisco algorithm to discover the super-class graph, enabling the model to achieve even more balanced image representations. To do so, we use a small amount of balanced data to construct a prototype-based relational graph, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. Last, in Section 5 we report experiments on four long-tailed benchmarks: CIFAR-100-LT, ImageNet-LT, Places-LT, and iNaturalist, and verify that our discovered super-class graph performs better for tail data in each dataset. Before detailing our contributions, we first embed our proposal in related work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 621, + 129, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 621, + 129, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 621, + 129, + 634 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 641, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 641, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 641, + 287, + 712 + ], + "type": "text", + "content": "Long-tailed recognition. Several strategies have been proposed to address class imbalance in recognition. We categorize them into three groups. Those in the first group are based on class re-balancing [8, 30, 44, 75], which balance the training sample numbers of different classes during model training. Class re-balancing methods also could be catego-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 307, + 72, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 312 + ], + "type": "text", + "content": "rized into three different groups, i.e., re-sampling [22, 32, 43, 67], cost-sensitive learning [14, 37, 56, 78, 79, 86] and logit adjustment [27, 45, 57, 59]. Class re-balancing methods improve the performance of the tail classes at the expense of the performance of the head classes. The second group is based on information augmentation, introducing additional information into model training to improve long-tailed learning performance. We identify four kinds of methods in the information augmentation scope, i.e., transfer learning, which includes head-to-tail knowledge transfer [6, 42, 64, 73], knowledge distillation [28, 40, 71], model pre-training [9, 33, 72] and self-training [24, 68, 74]. The third group focuses on improving network modules in long-tailed learning. This group includes representation learning [13, 46, 76], classifier learning [32, 41, 42, 69, 73], decoupled training [31, 32, 82], and ensemble learning [20, 83]. These methods introduce additional computation costs for increased performance. Our method belongs to the third group as it aims to learn a better representation of unbalanced training samples by the superclass graph, which is unexplored for long-tail recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 316, + 546, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 546, + 578 + ], + "type": "text", + "content": "Super-class learning. Super-class learning adds superclass labels as intermediate supervision into traditional deep learning. A super-class guided network [38] integrated the high-level semantic information into the network for image classification and object detection, which took two-level class annotations that contain both super-class and finer class labels. In [11], a two-phase multi-expert architecture was proposed for still image action recognition, which includes fine-grained and coarse-grained phases. However, they leveraged the ground truth of the super-class as supervision during the coarse-grained phase. Wu et al. [70] propose a taxonomic classifier to address the long-tail recognition problem, which classified each sample to the level that the classifier is competent. Zhou et al. [85] clustered the original categories into super-classes to produce a relatively balanced distribution in the super-class space, which also leveraged the ground truth of the super-class in the training phase. In contrast with the previous super-class learning, we do not use ground truth to group the original categories into the super-class space. To the best of our knowledge, no work exists that relies on graph learning to discover the super-class for long-tailed visual recognition, thus motivating this work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 582, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 546, + 712 + ], + "type": "text", + "content": "Graph neural networks. Recently, several graph neural network models (GNN) have been proposed to exploit the structures underlying graphs to benefit a variety of applications. There are two main research lines of GNN methods: non-spectral methods and spectral methods. The spectral methods [4, 10, 26, 35] focus on learning graph representations in a spectral domain, in which the learned filters are based on Laplacian matrices. The non-spectral methods [21, 63] develop an aggregator to aggregate a local set of features. Note that, message passing [19] is a key mechanism that allows GCNs and other graph neural networks to capture" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "19945" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 71, + 294, + 266 + ], + "blocks": [ + { + "bbox": [ + 97, + 71, + 294, + 266 + ], + "lines": [ + { + "bbox": [ + 97, + 71, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 294, + 266 + ], + "type": "image", + "image_path": "c833d2254310196ad13978c626cb615ac36571b82aa1f3734e163c9f4177c52e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 271, + 243, + 281 + ], + "lines": [ + { + "bbox": [ + 136, + 271, + 243, + 281 + ], + "spans": [ + { + "bbox": [ + 136, + 271, + 243, + 281 + ], + "type": "text", + "content": "(a) Illustration of proposed SuperDisco" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 327, + 69, + 484, + 159 + ], + "blocks": [ + { + "bbox": [ + 327, + 69, + 484, + 159 + ], + "lines": [ + { + "bbox": [ + 327, + 69, + 484, + 159 + ], + "spans": [ + { + "bbox": [ + 327, + 69, + 484, + 159 + ], + "type": "image", + "image_path": "8f91be0e202c4f80185112a6318b335b79fc5fcf749edabb6995ad2031be21b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 328, + 165, + 470, + 266 + ], + "blocks": [ + { + "bbox": [ + 328, + 165, + 470, + 266 + ], + "lines": [ + { + "bbox": [ + 328, + 165, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 328, + 165, + 470, + 266 + ], + "type": "image", + "image_path": "1a9511f57fd543c1a448cd45cd3fbc81b8339c52d175aad5c94022071d802859.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 271, + 473, + 281 + ], + "lines": [ + { + "bbox": [ + 329, + 271, + 473, + 281 + ], + "spans": [ + { + "bbox": [ + 329, + 271, + 473, + 281 + ], + "type": "text", + "content": "(c) Visualization of discovered super-class graph " + }, + { + "bbox": [ + 329, + 271, + 473, + 281 + ], + "type": "inline_equation", + "content": "c^2" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "text", + "content": "Figure 2. Illustration of proposed SuperDisco (a) and visualization of the similarity between the classes and discovered super-class at different levels (b), (c). In (a), we show two levels of super-class graphs " + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^1" + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^2" + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "text", + "content": ". The colour in each graph represents the discovered super-class. SuperDisco discovers the potential super-class at different levels hidden in each category from (b) and (c). " + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^1" + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "text", + "content": " roughly categorizes the original classes into three relatively balanced super-classes, and " + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^2" + }, + { + "bbox": [ + 46, + 285, + 545, + 330 + ], + "type": "text", + "content": " then finely categorize them into four more balanced super-classes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 288, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 288, + 441 + ], + "type": "text", + "content": "complex relationships and dependencies between nodes in a graph, and is a major reason why they have been successful in a variety of tasks involving graph-structured data. Our method belongs to the non-spectral methods, which leverage a GNN as the base architecture to discover the super-class representation. Our proposed super-class graph would refine and rectify the original imbalanced feature to a relatively balanced feature space, which has not been explored for long-tail recognition either." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 442, + 288, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 288, + 634 + ], + "type": "text", + "content": "Meta-learning for the long-tail. Meta-learning or learning to learn [3, 53, 58, 80, 81], is a learning paradigm where a model is trained on the distribution of tasks so as to enable rapid learning on new tasks. Ren et al. [51] first proposed meta-learning for the long-tailed problem by reweighting training examples. Shu et al. [54] proposed Meta-weight-Net to adaptively extract sample weights to guarantee robust recognition in the presence of training data bias. Li et al. [39] introduced meta-semantic augmentation for long-tailed recognition, which produces diversified augmented samples by translating features along many semantically meaningful directions by meta-learning. Our uniqueness is that our model aims to discover an improved super-class representation by meta-learning, which enables the original feature representation to adjust its corresponding higher-level super-class space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 645, + 279, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 279, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 279, + 659 + ], + "type": "text", + "content": "3. Learning to discover the super-class graph" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "In this section, we discuss how to learn to discover the super-class graph from the training samples and then expand on how to leverage such a graph to benefit the unbalanced data by refining the feature representations of samples. The" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 334, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 358 + ], + "type": "text", + "content": "overall illustration of SuperDisco and a visualization of the discovered super-class hierarchy are shown in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": "Preliminary. For long-tailed visual recognition, the goal is to learn an image classification model from an imbalanced training set and to evaluate the model on a balanced test set. We first define the notation for long-tailed recognition used throughout our paper. We define a training input " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "x_{k} \\in \\mathbb{R}, i \\in \\{1, \\dots, n\\}" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " is the number of training samples in the dataset. The corresponding labels are " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "y_{k} \\in \\{1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " is the number of classes. Let " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "n_{j}" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " denote the number of training samples for the class " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ". Here, we assume that " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "n_{i} \\geq n_{j}" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "i < j" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " shows the long-tailed problem simply. In this work, we typically consider a deep network model with three main components: a feature extractor " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ", a proposed graph model " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ". The feature extractor " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " first extracts an image representation as " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = f(x; \\theta)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ", which is then fed into the proposed graph model to refine a new representation as " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "g(\\mathbf{z}; \\phi) = \\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ". The final class prediction " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": " is given by a classifier function " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\tilde{y} = \\arg \\max h(\\tilde{\\mathbf{z}}; \\psi)" + }, + { + "bbox": [ + 304, + 361, + 546, + 589 + ], + "type": "text", + "content": ". Before detailing our approach, we add Table 1 to detail the meaning of each symbol for easy lookup." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "SuperDisco. We construct the super-class graph to organize and distill knowledge from the training process. The vertices represent different types of super-classes (e.g., the common contour between birds and airplanes) and the edges are automatically constructed to reflect the relationship between different super-classes. Our super-class graph contains multiple levels, which is closer to the relationship between various objects in the real world. Before detailing the structure, we first explicate why the multi-level super-class graphs are preferred over a flat super-class graph: a single level of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19946" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "content": "super-class groups is likely insufficient to model complex task relationships in real-world applications; for example, the similarities among different bird species are high, but there are also similarities between birds and mammals, e.g., they are both animals." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": "We assume the vertex representation " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^g\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": " and define the super-class graph as " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^l = (\\mathbf{H}_c^l,\\mathbf{A}_c^l)" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c^l = \\{\\mathbf{h}^{i^l}|\\forall i^l\\in [1,C^l ]\\} \\in \\mathbb{R}^{C^l\\times d}" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": " is the vertex feature matrix of the " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": "-th super-class level and " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_c^l = \\{A_c^l (\\mathbf{h}^{i^l},\\mathbf{h}^{j^l})|\\forall i^l,j^l\\in [1,C^l ]\\} \\in \\mathbb{R}^{C^l\\times C^l}" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": " is the vertex adjacency matrix in the " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": "-th super-class level, " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "C^l" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": " denotes the number of vertices in the " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": "-th super-class level. Our vertex representation " + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c^l" + }, + { + "bbox": [ + 46, + 132, + 287, + 266 + ], + "type": "text", + "content": " of the super-class graph is defined to get parameterized and learned during training. The initial vertex representations of each super-class level are randomly initialized, which encourages diversity of the discovered super-classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "text", + "content": "Next, we introduce how to compute the edge weight " + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "inline_equation", + "content": "A_{\\mathcal{C}}^{l}" + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "text", + "content": " in the super-class graph. The edge weight " + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "inline_equation", + "content": "A_{\\mathcal{C}}^{l}(\\mathbf{h}^{i^{l}}, \\mathbf{h}^{j^{l}})" + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "text", + "content": " between a pair of vertices " + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 266, + 287, + 316 + ], + "type": "text", + "content": " is gauged by the similarity between them. Formally:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 319, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 319, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 86, + 319, + 287, + 334 + ], + "type": "interline_equation", + "content": "A _ {\\mathcal {C}} ^ {l} \\left(\\mathbf {h} ^ {i l}, \\mathbf {h} ^ {j l}\\right) = \\sigma \\left(\\mathbf {W} _ {c} ^ {l} \\left(| \\mathbf {h} ^ {i l} - \\mathbf {h} ^ {j l} | / \\gamma_ {c} ^ {l}\\right) + \\mathbf {b} _ {c} ^ {l}\\right), \\tag {1}", + "image_path": "1768c804de68f08b186a9c21e5ce88fc36c1fed2b3b4cf88f1a02f73d673764f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_c^l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_c^l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " indicate learnable parameters of the " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": "-th super-class level, " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\gamma_c^l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": "-th super-class level is a scalar and " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " indicates the Sigmoid function, which normalizes the weight between 0 and 1. To adjust the representation of training samples by the involvement of super-classes, we first query the training samples in the super-class graph to obtain the relevant super-class. In light of this, we construct a new graph " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": ", which adds the original sample feature " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " to the super-class graph. We define " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " as the refined feature after the " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": "-th super-class graph. Here we define graph " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^l = (\\mathbf{H}_{\\mathcal{R}}^l, \\mathbf{A}_{\\mathcal{R}}^l)" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathcal{R}}^l = \\{[\\mathbf{z}^l, \\mathbf{h}^{i^l}] | \\forall i^l \\in [1, C^l]\\} \\in \\mathbb{R}^{(C^l + 1) \\times d}" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " denotes the vertex feature matrix of the " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": "-th super-class level, and " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathcal{R}}^l = \\{[A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{z}^1), A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{h}^{j^l})] | \\forall i^l, j^l \\in [1, C^l]\\} \\in \\mathbb{R}^{C^l + 1 \\times C^{l + 1}}" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " denotes the vertex adjacency matrix in the " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": "-th super-class level. The link between " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^l" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " and vertex " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^i" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " in the hierarchical graph is constructed by their similarity. In particular, analogous to the definition of weight in the super-class graph in Eq. (1), the weight " + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "inline_equation", + "content": "A_{\\mathcal{R}}^l (\\mathbf{h}^{i^l}, \\mathbf{z}^1)" + }, + { + "bbox": [ + 46, + 336, + 288, + 567 + ], + "type": "text", + "content": " is constructed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 567, + 287, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 567, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 91, + 567, + 287, + 582 + ], + "type": "interline_equation", + "content": "A _ {\\mathcal {R}} ^ {l} \\left(\\mathbf {h} ^ {i ^ {l}}, \\mathbf {z} ^ {1}\\right) = \\sigma \\left(\\mathbf {W} _ {r} ^ {l} \\left(\\left| \\mathbf {h} ^ {i ^ {l}} - \\mathbf {z} ^ {1} \\right| / \\gamma_ {r} ^ {l}\\right) + \\mathbf {b} _ {r} ^ {l}\\right), \\tag {2}", + "image_path": "6ee05b3150880ffa185d29b6af5d30d28bd570f54290a59cd26d0fd4bfa60af4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_r^l" + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_c^r" + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "content": " indicate learnable parameters of the " + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "content": "-th super-class level, " + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\gamma_r^l" + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 585, + 287, + 609 + ], + "type": "text", + "content": "-th super-class level is a scalar." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "text", + "content": "After constructing the new graph " + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "text", + "content": ", we propagate the most relevant super-class by message passing [19] from the discovered super-classes " + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "text", + "content": " to the features " + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^l" + }, + { + "bbox": [ + 46, + 609, + 287, + 670 + ], + "type": "text", + "content": " by introducing a Graph Neural Network (GNN). The message passing operation over the graph is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 672, + 287, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 672, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 287, + 687 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {\\mathcal {R}} ^ {(m + 1)} = \\mathrm {M P} \\left(\\mathbf {A} _ {\\mathcal {R}} ^ {l}, \\mathbf {H} _ {\\mathcal {R}} ^ {(m)}; \\mathbf {W} ^ {(m)}\\right), \\tag {3}", + "image_path": "1d3261d8c9021bc12355f9aba8d5df50c941ec382ac7593d2665a2928bcdfdc1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{M}\\mathbb{P}(\\cdot)" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the message passing function, " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{(m)}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the vertex embedding after " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " layers of GNN and " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{(m)}" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " is a" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 339, + 70, + 512, + 197 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 512, + 197 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 512, + 197 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 512, + 197 + ], + "type": "table", + "html": "
NotationDescription
hVertex representation
zOriginal sample feature
CSuper-class graph
HcVertex feature matrix of Cl
AcVertex adjacency matrix of Cl
RGraph which adds z to graph Cl
PPrototype graph
CpVertex feature matrix of P
ApVertex adjacency matrix of P
SSuper graph which connecting P and C
AVertex feature matrix of S
MVertex adjacency matrix of S
", + "image_path": "2b60d4353fb0b6097f8a3ea65021dcc6cb81cbe9d234e61c4374e8f462601afe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 198, + 534, + 209 + ], + "lines": [ + { + "bbox": [ + 315, + 198, + 534, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 534, + 209 + ], + "type": "text", + "content": "Table 1. Summary of the core notation used for SuperDisco." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": "learnable weight matrix of layer " + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": ". After stacking " + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": " GNN layers, we get the information-propagated feature representation " + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}^L" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": " for each level of the super-class graph " + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": ". Once we obtain the refined representation " + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}^L" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": " for a training sample by the super-class graph, we feed them into the classifier to make the predictions and compute the corresponding loss, i.e., Cross-entropy loss for optimization. Using gradient descent, we then update the super-class graph " + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 213, + 547, + 345 + ], + "type": "text", + "content": ". To be able to discover a more accurate super-class graph in the face of severe imbalance problems, we propose meta-learning super-class graph discovery in the next section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 357, + 538, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 538, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 538, + 370 + ], + "type": "text", + "content": "4. Meta-learning super-class graph discovery" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": "To explore and exploit a more accurate and richer superclass graph, we propose the Meta-SuperDisco to discover the super-class graph using meta-learning, making the model more robust. In the traditional meta-learning setting [17, 50], it includes meta-training tasks and meta-test task. Each task contains a support set " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": " and a query set " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": ". Each task is first trained by " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": " to get the task-specific learner and " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": " optimizes this learner to update the meta-learner. For long-tailed recognition with meta-learning, previous works [51, 54] randomly sample a small amount of balanced data denoted as " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": ". The imbalanced data and the small balanced data can be seen as " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": " in the training phase. The goal of meta-learning for long-tailed recognition is to use a small set of balanced data to optimize the model obtained from unbalanced data. We follow [51, 54] by randomly selecting the same number of samples (e.g., ten) as " + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 376, + 546, + 567 + ], + "type": "text", + "content": " per class from the training set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": "Meta-SuperDisco. To meta-learn the super-class graph, we construct a prototype graph " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": " is a balanced dataset. The prototype graph extracts the sample-level relation information, which captures the underlying relationship behind samples and alleviates the potential effects of abnormal samples. For the prototype graph, we need to compute the prototype of each category [55], which is defined as: " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^k = \\frac{1}{N^k}\\sum_{i = 1}^{N^k}\\mathbf{z}_j" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "N^k" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": " denotes the number of samples in class " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": " is the feature from " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "After calculating all prototype representations " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}^k |\\forall k\\in [1,K]\\}" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": ", which serve as the vertices in the prototype graph " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": ", we further need to define the edges and the corresponding" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19947" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "edge weights. The edge weight " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "A_{\\mathcal{P}}(\\mathbf{c}^i,\\mathbf{c}^j)" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " between two prototypes " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^i" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^j" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " is gauged by the similarity between them. The edge weight is calculated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 111, + 288, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 111, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 91, + 111, + 288, + 124 + ], + "type": "interline_equation", + "content": "A _ {\\mathcal {P}} \\left(\\mathbf {c} ^ {i}, \\mathbf {c} ^ {j}\\right) = \\sigma \\left(\\mathbf {W} _ {p} \\left(\\left| \\mathbf {c} ^ {i} - \\mathbf {c} ^ {j} \\right| / \\gamma_ {p}\\right) + \\mathbf {b} _ {p}\\right), \\tag {4}", + "image_path": "9c80be5ded78e56f000c0f0e6624acdc11f69ca76783000ea0e2f0ca4fbf0260.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_p" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_p" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " are the learnable parameters, " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\gamma_p" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " is a scalar. Thus, we denote the prototype graph as " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = (\\mathbf{C}_{\\mathcal{P}}, \\mathbf{A}_{\\mathcal{P}})" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathcal{P}_i} = \\{\\mathbf{c}^i | \\forall i \\in [1, K]\\} \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " represent a set of vertices, with each one corresponding to the prototype from a class, while " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathcal{P}} = \\{|A_{\\mathcal{P}}(\\mathbf{c}^i, \\mathbf{c}^j)| \\forall i, j \\in [1, K]\\} \\in \\mathbb{R}^{K \\times K}" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " gives the adjacency matrix, which indicates the proximity between prototypes. We then use the prototype graph to guide the learning of the meta super-class graph. We construct a super graph " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " by connecting prototype graph " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " to super-class graph " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": ". In the super graph " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": ", the vertices are " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = (\\mathbf{C}_{\\mathcal{P}^l}^l; \\mathbf{H}_{\\mathcal{C}^l}^l)" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^l" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " denotes the refined prototype graph vertex after the " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": "-th level super-class graph. Then, we calculate the link weight " + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "inline_equation", + "content": "A_{\\mathcal{S}}^l(c^i, \\{\\mathbf{h}^j\\})" + }, + { + "bbox": [ + 46, + 129, + 289, + 285 + ], + "type": "text", + "content": " of the super graph as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 289, + 287, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 289, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 71, + 289, + 287, + 321 + ], + "type": "interline_equation", + "content": "A _ {S} ^ {l} \\left(\\mathbf {c} ^ {i}, \\mathbf {h} ^ {j ^ {l}}\\right) = \\frac {\\exp \\left(- \\| \\left(\\mathbf {c} ^ {i} - \\mathbf {h} ^ {j ^ {l}}\\right) / \\gamma_ {s} ^ {l} \\| _ {2} ^ {2} / 2\\right)}{\\sum_ {j ^ {l ^ {\\prime}} = 1} ^ {J} \\exp \\left(- \\| \\left(\\mathbf {c} ^ {i} - \\mathbf {h} ^ {j ^ {l ^ {\\prime}}}\\right) / \\gamma_ {s} ^ {l} \\| _ {2} ^ {2} / 2\\right)}, \\tag {5}", + "image_path": "1d1a08b7320a301ad307b7a2819e5d71a5754dcf870b91a0c18606824c6f99a5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\gamma_s^l" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": " is a scaling factor. Note that, here we use softmax to ensure that the total weight of edges between the prototype graph " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": " and the super-class graph " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": " is equal to 1, giving the prototype graph a unique influence on the expression of each super-class. Thus, the adjacent matrix and feature matrix of the super graph " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "S^l = (\\mathbf{A}^l,\\mathbf{M}^l)" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^l = (\\mathbf{A}_{\\mathcal{P}},\\mathbf{A}_S^l;\\mathbf{A}_S^{l^\\mathrm{T}},\\mathbf{A}_\\mathcal{C}^l)" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^l = (\\mathbf{C}_{\\mathcal{P}^l}^l;\\mathbf{H}_{\\mathcal{C}^l}^l)" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": ". Once we constructed the super graph " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": ", we use message-passing again to propagate the most relevant knowledge from the prototype graph " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": " to the super-class graph " + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 325, + 288, + 445 + ], + "type": "text", + "content": ". Similar to eq. (3):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 449, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 449, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 103, + 449, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {(m + 1)} = \\mathbb {M P} \\left(\\mathbf {A} ^ {l}, \\mathbf {M} ^ {(m)}; \\mathbf {W} ^ {(m)}\\right). \\tag {6}", + "image_path": "bce17b638c2890afa1d33b1c7a6540c2a3b2903e0a529d8c6fcd75afb04da409.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "text", + "content": "We leverage the graph " + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "text", + "content": " to refine the super-class graph. Finally, we feed the original feature " + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "text", + "content": " into the super-class graph to get the information-propagated feature representation " + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}^L" + }, + { + "bbox": [ + 46, + 467, + 288, + 551 + ], + "type": "text", + "content": ", which refines the original feature by its corresponding discovered super-classes. We provide the complete SuperDisco and Meta-SuperDisco algorithm specifications in the supplemental material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 128, + 574 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": "Datasets. We apply our method to four commonly used long-tail recognition benchmarks. Sample images and the number of categories for all datasets are provided in the supplement material. CIFAR-100-LT reduces the number of training samples per class according to an exponential function " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "n = n_i \\mu^i" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": " is the class index, " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": " is the original number of training samples, and " + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mu \\in (0, 1)" + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": ". The imbalance factor of a dataset is defined as the number of training samples in the most populated class divided by the minority class. We consider imbalance factors {10, 50, 100}. ImageNet-LT [44] is a subset of ImageNet [12] consisting of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 70, + 518, + 137 + ], + "blocks": [ + { + "bbox": [ + 332, + 70, + 518, + 137 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 518, + 137 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 518, + 137 + ], + "type": "table", + "html": "
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
SuperDisco65.960.757.250.945.2
Meta-SuperDisco68.563.158.353.847.5
", + "image_path": "a5fbe34349fdfe1a8d5eab768f3c46c286c64ccb8a17c534908fa27e94119147.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 140, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 304, + 140, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 547, + 184 + ], + "type": "text", + "content": "Table 2. Benefit of SuperDisco and Meta-SuperDisco. SuperDisco achieves better performance compared to a baseline fin-tuning on all imbalance factors, while Meta-SuperDisco is even better for long-tailed recognition." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "type": "text", + "content": "115.8K images from 1000 categories, with maximally 1,280 images per class and minimally 5 images per class, and a balanced test set. Places-LT [44] has an imbalanced training set with 62,500 images for 365 classes from Places [84]. It contains images from 365 classes and the number of images per class ranges from 4980 to 5. The test sets are balanced and contain 100 images per class. iNaturalist [61] is a real-world long-tailed dataset with 675,170 training images for 5,089 classes, where the top " + }, + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "type": "text", + "content": " most populated classes contain more than " + }, + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 304, + 193, + 547, + 348 + ], + "type": "text", + "content": " of the training images. Additionally, there is also a severe imbalance among the super-classes of iNaturalist. The 13 ground truth super-classes images range from 158,407 to 308." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 352, + 548, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 548, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 548, + 532 + ], + "type": "text", + "content": "Implementation details. We follow [32] by first training a feature extractor with instance-balanced sampling, and then training our graph model and classifier based on the trained features. For CIFAR-100-LT, we follow [54] and use a ResNet-32 backbone. For ImageNet-LT, we use ResNeXt-50 [23] as our backbone, following [32]. For Places-LT, we report results with ResNet-152 following [44]. For iNaturalist, we use a ResNet-50 backbone. We train each dataset for 200 epochs with batch size 512. We use random left-right flipping and cropping as our training augmentation. For all experiments, we use an SGD optimizer with a momentum of 0.9 and a batch size of 512. We randomly selected 10 images per class from the training set for all datasets as " + }, + { + "bbox": [ + 304, + 352, + 548, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 352, + 548, + 532 + ], + "type": "text", + "content": ". Code available at: https://github.com/Yingjun-Du/SuperDisco." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "type": "text", + "content": "Benefit of SuperDisco and Meta-SuperDisco. To show the benefit of SuperDisco, we compare it with a fine-tuning baseline, which retrans the classifier only. Table 2 shows SuperDisco improves over fine-tuning on CIFAR-100-LT, and the results for the other long-tailed datasets are provided in the supplemental materials Table 1. In the most challenging setting with the largest imbalance factor of 200, our SuperDisco delivers " + }, + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "type": "inline_equation", + "content": "45.2\\%" + }, + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "type": "text", + "content": ", surpassing the baseline by " + }, + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 304, + 534, + 548, + 713 + ], + "type": "text", + "content": ". We attribute improvement to our model's ability to refine original features, allowing the discovered super-class graph to guide the tail features away from the dominant role of head features, thus leading to improvements over the original features. We also investigate the benefit of meta-learning with Meta-SuperDisco. The Meta-SuperDisco consistently surpasses the SuperDisco for all imbalance factors. The con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19948" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 276, + 191 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 276, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 276, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 276, + 191 + ], + "type": "table", + "html": "
Imbalance ratio
102050100200
Baseline60.357.347.544.939.3
(20)61.260.149.947.341.9
(2, 4, 8)65.362.753.149.843.2
(4, 8, 16)69.164.255.252.345.9
(4, 8, 16, 32)68.563.158.353.847.5
(4, 8, 16, 32, 64)66.962.758.952.946.3
Oracle super-classes66.963.254.751.443.2
", + "image_path": "c5cc9c860ce870541f92e9265c2dd556fce6faa4b4060297e47137541a15a4bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 232, + 261, + 337 + ], + "blocks": [ + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "type": "text", + "content": "Table 3. Effect of number of super-class levels on CIFAR-100-LT. Compared to a baseline [32] and an oracle setting, Meta-SuperDisco provides higher performance gains with more complex hierarchies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 232, + 261, + 337 + ], + "lines": [ + { + "bbox": [ + 69, + 232, + 261, + 337 + ], + "spans": [ + { + "bbox": [ + 69, + 232, + 261, + 337 + ], + "type": "table", + "html": "
ManyMediumFewAll
Baseline65.066.365.565.9
(13)71.870.266.170.8
(2,4,8)70.569.365.969.4
(4,8,16)72.270.966.470.3
(4,8,16,32)73.670.267.370.9
(4,8,16,32,64)73.472.968.372.3
(4,8,16,32,64,128)72.171.366.270.9
Oracle super-classes70.770.565.970.2
", + "image_path": "8d5025da6338e812cbd109bd740ca354b1fa8113cee9a3788244dfb9f7c6c365.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 338, + 288, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 288, + 371 + ], + "type": "text", + "content": "Table 4. Effect of number of super-class levels on iNaturalist. Meta-SuperDisco achieves consistent performance gains with more complex hierarchies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 376, + 287, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 412 + ], + "type": "text", + "content": "sistent improvements confirm that Meta-SuperDisco learns even more robust super-class graphs, leading to a discriminative representation of the tail data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 415, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 288, + 714 + ], + "type": "text", + "content": "Effect of the number of super-class levels. A significant challenge with any structure-aware learning algorithm is determining the appropriate complexity for the knowledge structure. So, we further analyze the effect of the super-class hierarchies, including the level (number of depths L) or the number of super-classes in each level. The results are shown in Table 3 and Table 4. The super-class number from the bottom layer to the top layer is saved in a tuple. For example, (2, 4, 8) represents three depth, with two super-classes in the top layer. The baseline is Decouple-LWS [32], which only inputs the original feature to learn a new classifier. The oracle super-classes are first trained on two long-tailed datasets using the ground truth super-class labels for super-class classifications. Once the training is completed, each oracle super-class is obtained by averaging the samples of each super-class. We constructed a one-layer super-class graph using these super-classes, where the vertices of the graph are for each super-class, and the edges of the graph are computed according to Eq. (1). Then, we use the message passing by Eq. (3) to refine the original features and input them into the classifier to get the final predictions. From Table 3, we observe that using oracle super-classes achieves better performance compared to the learned super-class (20) since it uses the ground truth super-classes as supervision. We also conclude that too few levels may not be enough" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 410, + 123 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 410, + 123 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 410, + 123 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 410, + 123 + ], + "type": "image", + "image_path": "ad75c3bb258313c39fce850e20694d6365887bcf6881cc9730db255e0f253392.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 127, + 390, + 137 + ], + "lines": [ + { + "bbox": [ + 341, + 127, + 390, + 137 + ], + "spans": [ + { + "bbox": [ + 341, + 127, + 390, + 137 + ], + "type": "text", + "content": "(a) SuperDisco" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 410, + 70, + 539, + 122 + ], + "blocks": [ + { + "bbox": [ + 410, + 70, + 539, + 122 + ], + "lines": [ + { + "bbox": [ + 410, + 70, + 539, + 122 + ], + "spans": [ + { + "bbox": [ + 410, + 70, + 539, + 122 + ], + "type": "image", + "image_path": "40d1be7db1d59e8e1c62c8fe9d07dabbf7f39b6a9a571b7be13daaabdb088a17.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 127, + 515, + 137 + ], + "lines": [ + { + "bbox": [ + 445, + 127, + 515, + 137 + ], + "spans": [ + { + "bbox": [ + 445, + 127, + 515, + 137 + ], + "type": "text", + "content": "(b) Meta-SuperDisco" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 138, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 304, + 138, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 547, + 171 + ], + "type": "text", + "content": "Figure 3. Similarity between discovered super-classes and classes. SuperDisco discovers super-classes hidden in each class, while Meta-SuperDisco discovers more accurate super-classes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 174, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 547, + 424 + ], + "type": "text", + "content": "to learn the precise super-classes (e.g., tuple (20) or (2, 4, 8)). In this dataset, increasing levels (e.g., tuple (4, 8, 16, 32)) achieves better performance on the smaller imbalance factor (e.g., 10), and similar performance compared with (4, 8, 16). For the real-world long-tailed dataset iNaturalist [62] in Table 4, we find no significant improvement for the few-shot classes in the performance of the oracle super-class compared to the baseline, and the same is true for the performance of the discovered super-class structure (13). This is because the super-class of iNaturalist also have serious long-tailed problems, resulting in the refined features of tail classes remaining indistinguishable from the refined features of head classes. However, with a more complex graph structure (4, 8, 16, 32, 64), the few-shot performance improves by a good margin compared with the baseline, and even the oracle super-classes. We attribute this to our model's ability to explore relatively balanced super-class spaces, thus making the refined tail category features discriminative. By comparing Table 3 and Table 4, we conclude that deeper as well as wider graphs are needed to discover the super-classes in the case of severe class imbalance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "type": "text", + "content": "Visualization of SuperDisco. To understand the meaning of the discovered super-classes more clearly, we present a visualization in Figure. 3. We selected 12 different categories from the CIFAR-100 test dataset. We calculate the similarity of each of these 12 categories to the different vertices in the graph we explore. Here we show the similarity with the second layer of graph vertices (C1, C2, C3, C4). We can see different categories mainly activate different vertices, e.g., bus " + }, + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "type": "text", + "content": " C3 and road " + }, + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 426, + 547, + 714 + ], + "type": "text", + "content": " C2. As shown in this heatmap, we find that C1 reflects the super-class of flowers, C2 reflects the super-class of buildings, C3 reflects the super-class of vehicles, C4 reflects the super-class of fish. Another observation is that the second-largest activated super-class is also meaningful, promoting knowledge transfer between super-classes. For example, road and bridge are related to the C3 superclass, since some vehicles may be on the road and bridge. This visualization reflects that we can use graph models to discover the super-classes and the relationships between each super-class. We also visualize the discovered meta-learning super-classes in Figure 3 (b). The discovered super-classes are even more accurate, e.g., roses have high similarity to C2, which mainly reflects the buildings super-class, while it has high similarity to C1, which is the flowers super-class. This once again validates the benefit of Meta-SuperDisco." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19949" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 182, + 145 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 182, + 145 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 182, + 145 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 182, + 145 + ], + "type": "image", + "image_path": "86d685881b3449474338255a4a8358b62b9a0b0cc79def00f46464450880b6b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 150, + 165, + 160 + ], + "lines": [ + { + "bbox": [ + 94, + 150, + 165, + 160 + ], + "spans": [ + { + "bbox": [ + 94, + 150, + 165, + 160 + ], + "type": "text", + "content": "(a) Original features" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 193, + 71, + 290, + 144 + ], + "blocks": [ + { + "bbox": [ + 193, + 71, + 290, + 144 + ], + "lines": [ + { + "bbox": [ + 193, + 71, + 290, + 144 + ], + "spans": [ + { + "bbox": [ + 193, + 71, + 290, + 144 + ], + "type": "image", + "image_path": "af99ecc2e2d54628f761f3c66477bea79fcd1fef79a2ee50ae03a11ea59bd622.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 150, + 413, + 160 + ], + "lines": [ + { + "bbox": [ + 289, + 150, + 413, + 160 + ], + "spans": [ + { + "bbox": [ + 289, + 150, + 413, + 160 + ], + "type": "text", + "content": "(c) Refined features by SuperDisco" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 72, + 399, + 144 + ], + "blocks": [ + { + "bbox": [ + 302, + 72, + 399, + 144 + ], + "lines": [ + { + "bbox": [ + 302, + 72, + 399, + 144 + ], + "spans": [ + { + "bbox": [ + 302, + 72, + 399, + 144 + ], + "type": "image", + "image_path": "f6020ca1af763ca7a9554ee400ba03cfdc5304c11cb1dbd37ac8e6f616fc3beb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 410, + 72, + 507, + 144 + ], + "blocks": [ + { + "bbox": [ + 410, + 72, + 507, + 144 + ], + "lines": [ + { + "bbox": [ + 410, + 72, + 507, + 144 + ], + "spans": [ + { + "bbox": [ + 410, + 72, + 507, + 144 + ], + "type": "image", + "image_path": "198d2ca7405cd4c2eb306a90c989b2a274cebb29f0a75a7f2c26cc031467bf12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 89, + 166, + 179, + 237 + ], + "blocks": [ + { + "bbox": [ + 89, + 166, + 179, + 237 + ], + "lines": [ + { + "bbox": [ + 89, + 166, + 179, + 237 + ], + "spans": [ + { + "bbox": [ + 89, + 166, + 179, + 237 + ], + "type": "image", + "image_path": "1e90b07d166fe11b6d1af84e15ae336f85caad00ce4b7b45b5f995dafdea392f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 244, + 175, + 255 + ], + "lines": [ + { + "bbox": [ + 94, + 244, + 175, + 255 + ], + "spans": [ + { + "bbox": [ + 94, + 244, + 175, + 255 + ], + "type": "text", + "content": "(b) Training distribution" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 189, + 166, + 286, + 238 + ], + "blocks": [ + { + "bbox": [ + 189, + 166, + 286, + 238 + ], + "lines": [ + { + "bbox": [ + 189, + 166, + 286, + 238 + ], + "spans": [ + { + "bbox": [ + 189, + 166, + 286, + 238 + ], + "type": "image", + "image_path": "d8ae94c57d2f09d1dab25a8d230eb70e3f03f8467354aec06464f7efca0f094f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 244, + 423, + 255 + ], + "lines": [ + { + "bbox": [ + 279, + 244, + 423, + 255 + ], + "spans": [ + { + "bbox": [ + 279, + 244, + 423, + 255 + ], + "type": "text", + "content": "(d) Refined features by Meta-SuperDisco" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 295, + 166, + 398, + 238 + ], + "blocks": [ + { + "bbox": [ + 295, + 166, + 398, + 238 + ], + "lines": [ + { + "bbox": [ + 295, + 166, + 398, + 238 + ], + "spans": [ + { + "bbox": [ + 295, + 166, + 398, + 238 + ], + "type": "image", + "image_path": "8afc546fa9283e628ad771f991a32aee87eaa912a683b333d8fe5816e712f0c1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 410, + 166, + 507, + 238 + ], + "blocks": [ + { + "bbox": [ + 410, + 166, + 507, + 238 + ], + "lines": [ + { + "bbox": [ + 410, + 166, + 507, + 238 + ], + "spans": [ + { + "bbox": [ + 410, + 166, + 507, + 238 + ], + "type": "image", + "image_path": "9b06e08889990f3daf85c823d38d35ef5d74c93d14d47808e3fcce43d97b6976.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 96, + 307, + 239, + 414 + ], + "blocks": [ + { + "bbox": [ + 46, + 266, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 545, + 299 + ], + "type": "text", + "content": "Figure 4. Visualization of refined features on CIFAR-100-LT, with the original features (a) and their corresponding training distribution (b). Colours indicate categories. SuperDisco (c) guides the original features on being clustered into the corresponding super-class space at different levels, while Meta-SuperDisco (d) obtains even more discriminative intra-class features." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 307, + 239, + 414 + ], + "lines": [ + { + "bbox": [ + 96, + 307, + 239, + 414 + ], + "spans": [ + { + "bbox": [ + 96, + 307, + 239, + 414 + ], + "type": "image", + "image_path": "e0a0492f9a18cb8a7a45dd84b0102d813c0e4e1a9757b0ee80f79f430448c340.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 423, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 456 + ], + "type": "text", + "content": "Figure 5. Effect of refined features. Accuracy increases along with the increased super-class levels, revealing that more accurate and richer super-classes facilitate better long-tailed recognition." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 461, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 509 + ], + "type": "text", + "content": "Furthermore, in the (c) learned the hierarchical concept of each class, we can see that bus and bridge have the same concept C2 in the last concept level, which may be due to the possible presence of cars on the bridge." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "Visualization of refined features. To understand the empirical benefit of SuperDisco, we visualize in Figure 4 the original features and refined features with super-class graphs of the different levels using t-SNE [60]. We choose the vertices numbers as (2, 4, 6), meaning the super-class graph has three different levels, each with a different number of vertices. The original features of the category with a small sample size will overlap with the (original) features of the category with a large sample size. Super-class graphs discovered by our model project the original features into a high-level super-class space, where the imbalance is relatively small. Hence, its corresponding subcategory can be predicted more accurately. It is worth noting that when comparing the two different super-class graphs on top and below, the features obtained by Meta-SuperDisco are even more distinctive and distant from each other. To better measure the goodness of the refined features obtained at different levels," + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 318, + 300, + 529, + 350 + ], + "blocks": [ + { + "bbox": [ + 318, + 300, + 529, + 350 + ], + "lines": [ + { + "bbox": [ + 318, + 300, + 529, + 350 + ], + "spans": [ + { + "bbox": [ + 318, + 300, + 529, + 350 + ], + "type": "table", + "html": "
ManyMediumFewAll
Baseline58.449.334.852.7
Multi-layer perceptron63.551.835.955.0
Graph convolution network66.153.337.157.1
", + "image_path": "f2cdd2765b8154a2e03d71f9bec1892ca06ad79d5e97fcacfae9064fffbed972.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 354, + 545, + 386 + ], + "lines": [ + { + "bbox": [ + 306, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 545, + 386 + ], + "type": "text", + "content": "Table 5. Analysis of super-class mechanism on ImageNet-LT. The super-class mechanism contributes most, the graph convolution network improves results further." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 399, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 471 + ], + "type": "text", + "content": "we show in Figure 5 the prediction accuracy using different refined features. We find that the accuracy increases along with the increased super-class levels, which shows that using more accurate and richer super-classes facilitates better performance. This again demonstrates that Meta-SuperDisco is most suitable for long-tailed visual recognition." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 473, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 616 + ], + "type": "text", + "content": "Analysis of super-class mechanism. To demonstrate that the improved performance of our SuperDisco cannot solely be attributed to the graph convolutional network module, we conducted an experiment where we replaced it with a multi-layer perceptron to obtain the representation per sample. In Table 5, the performance gains of our method are primarily due to the super-classes rather than the graph convolution network. The results suggest that incorporating the super-classes mechanism plays a crucial role in improving the performance of long-tailed problems. Furthermore, the results improve further when we replace the multi-layer perceptron with our graph convolution network module." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Comparison with the state-of-the-art. We evaluate our method on the four long-tailed datasets under different imbalance factors in Table 6 and 7. Our model achieves state-of-the-art performance on the tail data of all datasets. For ImageNet-LT, our model achieves state-of-the-art performance on both few-shot and all data. In the most challenging Places-LT, our model delivers " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "40.3\\%" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " on all classes, surpassing the second-best Parisot et al. [47] by " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": ". On the real-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19950" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 511, + 192 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 511, + 192 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 511, + 192 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 511, + 192 + ], + "type": "table", + "html": "
VenueImageNet-LTPlaces-LTiNaturalist
ManyMediumFewAllManyMediumFewAllManyMediumFewAll
Kang et al. [32]ICLR 1960.247.230.349.940.639.128.637.665.066.365.565.9
Kang et al. [31]ICLR 2161.849.430.951.5-------68.6
He et al. [25]ICCV 2164.150.431.553.1----70.670.167.669.1
Li et al. [40]CVPR 2166.851.135.456.0-------69.3
Samuel et al. [52]ICCV 2164.049.833.153.5-------69.7
Alshammari et al. [1]CVPR 2262.550.441.553.9----71.270.469.770.2
Zhang et al. [75]CVPR 2161.352.231.452.940.442.430.139.369.071.170.270.6
Parisot et al. [47]CVPR 2263.252.136.954.139.741.034.939.2----
Park et al. [48]CVPR 2266.453.935.656.2----73.172.668.772.8
This paper66.153.337.157.145.342.835.340.372.372.971.373.6
", + "image_path": "d8b18dd2c1f55b54bb837eb29ea136002029a4f0010d93ef3e1eb7289acc4078.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 239, + 257, + 355 + ], + "blocks": [ + { + "bbox": [ + 46, + 201, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 235 + ], + "type": "text", + "content": "Table 6. Comparison with the state-of-the-art on ImageNet-LT, Places-LT and iNaturalist. Best and second best results are highlighted in bold and italic bold. Our Meta-SuperDisco achieves either better or comparable performance than state-of-the-art methods under the tail and all data for long-tailed visual recognition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 239, + 257, + 355 + ], + "lines": [ + { + "bbox": [ + 77, + 239, + 257, + 355 + ], + "spans": [ + { + "bbox": [ + 77, + 239, + 257, + 355 + ], + "type": "table", + "html": "
VenueImbalance ratio
1050100
Park et al. [49]ICCV 2159.547.442.0
Li et al. [40]CVPR 2162.350.546.0
Zhong et al. [82]CVPR 2162.551.546.8
Samuel et al. [52]ICCV 2163.457.647.3
Wang et al. [66]ICLR 2161.851.748.0
Zhu et al. [87]CVPR 2264.956.651.9
Cui et al. [7]ICCV 2164.256.052.0
Alshammari et al. [1]CVPR 2268.857.753.3
This paper69.358.353.8
", + "image_path": "55c923aaba8eae561a17b9f199dc021010334bb22be6445abf07b9b81cead53b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 364, + 288, + 387 + ], + "lines": [ + { + "bbox": [ + 47, + 364, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 288, + 387 + ], + "type": "text", + "content": "Table 7. Comparison with the state-of-the-art on CIFAR-100-LT. Our model achieves best performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 406, + 288, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 288, + 491 + ], + "type": "text", + "content": "world long-tailed dataset iNaturalist, our model achieves the three best performances under four different shots. On the long-tailed synthetic dataset CIFAR-100-LT, our model achieves the best performance under each imbalance factor. The consistent improvements on all benchmarks under various configurations confirm that our Meta-SuperDisco is effective for long-tailed visual recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": "Limitations. We show that SuperDisco and Meta-SuperDisco achieve good performance on tail data while being less successful on the head data. Based on this result, we also perform an experiment on balanced CIFAR-100 in Figure 6. With SuperDisco and Meta-SuperDisco, there is only a slight change in performance at the expense of an increased inference time. This reveals that our SuperDisco does not change the original features much through message passing on a balanced dataset. This may be because the obtained super-classes are still the original class itself. In addition, as the computation of graphs involves many matrix operations, our model also requires a relatively long computational speed. Due to introducing a prototype graph and more data, Meta-SuperDisco takes longer to compute. In addition, the training time of SuperDisco and its meta variants is also 1.5 times higher than the baseline. Future work could investigate how to use the discovered super-class graph in balanced datasets and how to reduce the computation time." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 343, + 250, + 504, + 365 + ], + "blocks": [ + { + "bbox": [ + 343, + 250, + 504, + 365 + ], + "lines": [ + { + "bbox": [ + 343, + 250, + 504, + 365 + ], + "spans": [ + { + "bbox": [ + 343, + 250, + 504, + 365 + ], + "type": "image", + "image_path": "bb1ad1ce828239038659c4f2b6829fc7fb3aaff56e41c134c834b74c4ab4b136.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 374, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 418 + ], + "type": "text", + "content": "Figure 6. Limitation. Accuracy " + }, + { + "bbox": [ + 304, + 374, + 547, + 418 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 374, + 547, + 418 + ], + "type": "text", + "content": " vs. speed (ms) comparison with different methods on balanced CIFAR-100. SuperDisco has little impact on the performance of balanced datasets at the expense of increased inference time." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 426, + 384, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 384, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 384, + 438 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 448, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 616 + ], + "type": "text", + "content": "This paper proposes learning to discover a super-class graph for long-tailed visual recognition. The proposed superclass graph could rectify and refine the original features by message passing, which results in attending to the most relevant entities based on their semantic similarity between concepts for more accurate predictions. To obtain a more informative super-class graph and more balanced image representations, we further propose to meta-learn the super-class graph based on the prototype graph from a small amount of imbalanced data. We conduct thorough ablation studies to demonstrate the effectiveness of the proposed SuperDisco and Meta-SuperDisco algorithms. The state-of-the-art performance on the long-tailed version of four datasets further substantiates the benefit of our proposal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 547, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 702 + ], + "type": "text", + "content": "Acknowledgment This work is financially supported by the Inception Institute of Artificial Intelligence, the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19951" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In CVPR, pages 6897-6907, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 146 + ], + "type": "text", + "content": "[2] Chris Anderson. The long tail: How endless choice is creating unlimited demand. Random House, 2007. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 147, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 287, + 168 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Samy Bengio, and Jocelyn Cloutier. Learning a synaptic learning rule. In IJCNN, 1991. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 201 + ], + "type": "text", + "content": "[4] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 203, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 287, + 235 + ], + "type": "text", + "content": "[5] Zhi Chen, Yijie Bei, and Cynthia Rudin. Concept whitening for interpretable image recognition. In Nature Machine Intelligence, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "type": "text", + "content": "[6] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 269, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 291 + ], + "type": "text", + "content": "[7] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In ICCV, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 292, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 287, + 324 + ], + "type": "text", + "content": "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 357 + ], + "type": "text", + "content": "[9] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In CVPR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 358, + 287, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 391 + ], + "type": "text", + "content": "[10] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 392, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 287, + 445 + ], + "type": "text", + "content": "[11] Hojat Asgarian Dehkordi, Ali Soltani Nezhad, Hossein Kashiani, Shahriar Baradaran Shokouhi, and Ahmad Aytollahi. Multi-expert human action recognition with hierarchical super-class learning. Knowledge-Based Systems, page 109091, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "text", + "content": "[13] Qi Dong, Shaogang Gong, and Xiatian Zhu. Class rectification hard mining for imbalanced deep learning. In ICCV, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 514, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 534 + ], + "type": "text", + "content": "[14] Charles Elkan. The foundations of cost-sensitive learning. In IJCAI, 2001. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[15] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. A multiple resampling method for learning from imbalanced data sets. Computational Intelligence, 20(1):18-36, 2004. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[16] Adam R Ferguson, Jessica L Nielson, Melissa H Cragin, Anita E Bandrowski, and Maryann E Martone. Big data from small data: data-sharing in the 'long tail' of neuroscience. Nature Neuroscience, 17(11):1442-1447, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[17] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICLR, pages 1126–1135, 2017. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[18] Camille Garcin, Alexis Joly, Pierre Bonnet, Jean-Christophe Lombardo, Antoine Affouard, Mathias Chouet, Maximilien Servajean, Joseph Salmon, and Titouan Lorieul. P1@ ntnet-300k: a plant image dataset with high label ambiguity and a long-tailed distribution. In NeurIPS 2021-35th Conference on Neural Information Processing Systems, 2021. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICLR, pages 1263-1272, 2017. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 148 + ], + "type": "text", + "content": "[20] Hao Guo and Song Wang. Long-tailed multi-label visual recognition by collaborative training on uniform and rebalanced samplings. In CVPR, pages 15089-15098, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 149, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 149, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 545, + 172 + ], + "type": "text", + "content": "[21] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "text", + "content": "[22] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderlinesmote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878-887, 2005. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 217, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 247 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 249, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 249, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 249, + 545, + 281 + ], + "type": "text", + "content": "[24] Ruifei He, Jihan Yang, and Xiaojun Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "type": "text", + "content": "[25] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In ICCV, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 305, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 337 + ], + "type": "text", + "content": "[26] Mikael Henaff, Joan Bruna, and Yann LeCun. Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "text", + "content": "[27] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 381, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 414 + ], + "type": "text", + "content": "[28] Xinting Hu, Yi Jiang, Kaihua Tang, Jingyuan Chen, Chunyan Miao, and Hanwang Zhang. Learning to segment the tail. In CVPR, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 415, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 447 + ], + "type": "text", + "content": "[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In CVPR, 2016. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "text", + "content": "[30] Muhammad Abdullah Jamal, Matthew Brown, Ming-Hsuan Yang, Liqiang Wang, and Boqing Gong. Rethinking class-balanced methods for long-tailed visual recognition from a domain adaptation perspective. In CVPR, pages 7610-7619, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 503, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 535 + ], + "type": "text", + "content": "[31] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In ICLR, 2021. 1, 2, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 536, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 580 + ], + "type": "text", + "content": "[32] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 582, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 612 + ], + "type": "text", + "content": "[33] Shyamgopal Karthik, Jérôme Revaud, and Chidrovskii Boris. Learning from long-tailed data with noisy labels. arXiv:2108.11096, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 613, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 646 + ], + "type": "text", + "content": "[34] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In CVPR, 2020. 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[35] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[36] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In ICLR, 2020. 1" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19952" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[37] Bo Li, Yongqiang Yao, Jingru Tan, Gang Zhang, Fengwei Yu, Jianwei Lu, and Ye Luo. Equalized focal loss for dense long-tailed object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6990-6999, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[38] Kaidong Li, Nina Y Wang, Yiju Yang, and Guanghui Wang. Sgnet: A super-class guided network for image classification and object detection. In Conference on Robots and Vision, pages 127-134, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 217 + ], + "type": "text", + "content": "[39] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In CVPR, pages 5212-5221, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 251 + ], + "type": "text", + "content": "[40] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In CVPR, pages 630-639, 2021. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "type": "text", + "content": "[41] Bo Liu, Haoxiang Li, Hao Kang, Gang Hua, and Nuno Vasconcelos. Gistnet: a geometric structure transfer network for long-tailed recognition. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 329 + ], + "type": "text", + "content": "[42] Jialun Liu, Yifan Sun, Chuchu Han, Zhaopeng Dou, and Wenhui Li. Deep representation learning on long-tailed data: A learnable embedding augmentation perspective. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 374 + ], + "type": "text", + "content": "[43] Xu-Ying Liu, Jianxin Wu, and Zhi-Hua Zhou. Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, 39(2):539-550, 2008. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "type": "text", + "content": "[44] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, pages 2537-2546, 2019. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 454 + ], + "type": "text", + "content": "[45] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In ICLR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 487 + ], + "type": "text", + "content": "[46] Wanli Ouyang, Xiaogang Wang, Cong Zhang, and Xiaokang Yang. Factors in finetuning deep model for object detection with long-tail distribution. In CVPR, pages 864-873, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[47] Sarah Parisot, Pedro M Esperança, Steven McDonagh, Tamas J Madarasz, Yongxin Yang, and Zhenguo Li. Long-tail recognition via compositional knowledge transfer. In CVPR, pages 6939-6948, 2022. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 598 + ], + "type": "text", + "content": "[48] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 633 + ], + "type": "text", + "content": "[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In ICCV, pages 735-744, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 656 + ], + "type": "text", + "content": "[50] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[51] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 1, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[52] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In ICCV, 2021. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "text", + "content": "[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[54] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. In NeurIPS, 2019. 1, 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "text", + "content": "[55] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 186, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 228 + ], + "type": "text", + "content": "[56] Yanmin Sun, Mohamed S Kamel, Andrew KC Wong, and Yang Wang. Cost-sensitive boosting for classification of imbalanced data. Pattern Recognition, 40(12):3358-3378, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 230, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 547, + 262 + ], + "type": "text", + "content": "[57] Kaihua Tang, Jianqiang Huang, and Hanwang Zhang. Long-tailed classification by keeping the good and removing the bad momentum causal effect. In NeurlPS, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 264, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 547, + 285 + ], + "type": "text", + "content": "[58] Sebastian Thrun and Lorien Pratt, editors. Learning to Learn. Kluwer Academic Publishers, USA, 1998. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 319 + ], + "type": "text", + "content": "[59] Junjiao Tian, Yen-Cheng Liu, Nathan Glaser, Yen-Chang Hsu, and Zsolt Kira. Posterior re-calibration for imbalanced datasets. In NeurlPS, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 321, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 353 + ], + "type": "text", + "content": "[60] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(11), 2008. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[61] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, pages 8769-8778, 2018. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 431 + ], + "type": "text", + "content": "[62] Grant Van Horn and Pietro Perona. The devil is in the tails: Fine-grained classification in the wild. arXiv preprint arXiv:1709.01450, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "type": "text", + "content": "[63] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 467, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 509 + ], + "type": "text", + "content": "[64] Jianfeng Wang, Thomas Lukasiewicz, Xiaolin Hu, Jianfei Cai, and Zhenghua Xu. Rsg: A simple but effective module for learning imbalanced datasets. In CVPR, pages 3784-3793, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": "[65] Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is in classification: A simple framework for long-tail instance segmentation. In ECCV, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 555, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 589 + ], + "type": "text", + "content": "[66] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. In ICLR, 2021. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "text", + "content": "[67] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "text", + "content": "[68] Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[69] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In CVPR, pages 8659-8668, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[70] Tz-Ying Wu, Pedro Morgado, Pei Wang, Chih-Hui Ho, and Nuno Vasconcelos. Solving long-tailed recognition with deep" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19953" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "realistic taxonomic classifier. In ECCV, pages 171-189, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[71] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In ECCV, pages 247-263, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 152 + ], + "type": "text", + "content": "[72] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 153, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 288, + 195 + ], + "type": "text", + "content": "[73] Xi Yin, Xiang Yu, Kihyuk Sohn, Xiaoming Liu, and Manmohan Chandraker. Feature transfer learning for face recognition with under-represented data. In CVPR, pages 5704-5713, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[74] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. Mosaicos: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "type": "text", + "content": "[75] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In CVPR, pages 2361-2370, 2021. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[76] Xiao Zhang, Zhiyuan Fang, Yandong Wen, Zhifeng Li, and Yu Qiao. Range loss for deep face recognition with long-tailed training data. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "text", + "content": "[77] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. arXiv preprint arXiv:2110.04596, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "text", + "content": "[78] Yifan Zhang, Peilin Zhao, Jiezhang Cao, Wenye Ma, Junzhou Huang, Qingyao Wu, and Mingkui Tan. Online adaptive asymmetric active learning for budgeted imbalanced data. In KDD, pages 2768-2777, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[79] Peilin Zhao, Yifan Zhang, Min Wu, Steven CH Hoi, Mingkui Tan, and Junzhou Huang. Adaptive cost-sensitive online classification. IEEE Transactions on Knowledge and Data Engineering, 31(2):214-228, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 477 + ], + "type": "text", + "content": "[80] Xiantong Zhen, Yingjun Du, Huan Xiong, Qiang Qiu, Cees Snoek, and Ling Shao. Learning to learn variational semantic memory. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 510 + ], + "type": "text", + "content": "[81] Xiantong Zhen, Haoliang Sun, Yingjun Du, Jun Xu, Yilong Yin, Ling Shao, and Cees Snoek. Learning to learn kernels with variational random features. In ICML, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 513, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 544 + ], + "type": "text", + "content": "[82] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In CVPR, 2021. 2, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 546, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 588 + ], + "type": "text", + "content": "[83] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In CVPR, pages 9719–9728, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[84] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "text", + "content": "[85] Yucan Zhou, Qinghua Hu, and Yu Wang. Deep super-class learning for long-tail distributed image classification. Pattern Recognition, 80:118-128, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[86] ZhiHua Zhou and XuYing Liu. Training cost-sensitive neural networks with methods addressing the class imbalance problem. IEEE Transactions on Knowledge and Data Engineering, 18(1):63-77, 2005. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 116 + ], + "type": "text", + "content": "[87] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In CVPR, pages 6908-6917, 2022. 8" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19954" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_content_list.json b/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3710eee362550a7ca8d02943585f2c5181bfce27 --- /dev/null +++ b/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_content_list.json @@ -0,0 +1,1927 @@ +[ + { + "type": "text", + "text": "Superclass Learning with Representation Enhancement", + "text_level": 1, + "bbox": [ + 202, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyu Gan $^{1,2}$", + "bbox": [ + 187, + 179, + 290, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suyun Zhao $^{1,2,*}$", + "bbox": [ + 330, + 180, + 462, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hong Chen $^{1,2}$", + "bbox": [ + 352, + 199, + 464, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinlong Kang", + "bbox": [ + 496, + 180, + 612, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liyuan Shang", + "bbox": [ + 651, + 180, + 772, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cuiping $\\mathbf{L}\\mathbf{i}^{1,2}$", + "bbox": [ + 504, + 199, + 611, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Key Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China \n2Renmin University of China, Beijing, China", + "bbox": [ + 101, + 215, + 867, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zygan, zhaosuyun, kangjinlong, shangliyuan4032, chong, LICUping}@ruc.edu.cn", + "bbox": [ + 140, + 253, + 821, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In many real scenarios, data are often divided into a handful of artificial super categories in terms of expert knowledge rather than the representations of images. Concretely, a superclass may contain massive and various raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classification techniques are intractable to recognize superclass without raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhancement(SCLRE), to recognize super categories by leveraging enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE collapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision boundary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Experimentally, extensive results demonstrate that SCLRE outperforms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets.", + "bbox": [ + 75, + 335, + 473, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 713, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent decades, basic-level raw categorization (e.g. cats vs dogs, apples vs bananas) has greatly developed [9, 27] while high-level or super-coarse-grained visual categorization (e.g., recyclable waste vs kitchen waste, creatures vs non-creatures) has received little attention. In many real scenarios, there often exist a handful of high-level categories, wherein numerous images from diverse basic-level categories share one common label. We tend to define this kind of super-coarse-grained class as Superclass.", + "bbox": [ + 75, + 739, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a24cc70987046ff79367574bad04da73da7a217a5ed361c57c8f9b6bc12b72d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 301, + 696, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f6e157739725d90d4c41884b9037e7881d6442447e6adc792b2b231c3273962.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 303, + 875, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bdc5511a9e598233af0f47ca10706d2cd00ffe6cbb2a2a406e17ccb3a1050de6.jpg", + "image_caption": [ + "(a) Image Feature Domain" + ], + "image_footnote": [], + "bbox": [ + 503, + 422, + 614, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18f9b99a87527fa81959da251f8889ee2b0573c7f12f19a752546e126de18b77.jpg", + "image_caption": [ + "(b) Separate Fine Domain" + ], + "image_footnote": [], + "bbox": [ + 635, + 422, + 743, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3dc8ab9df2cc866ad3e76b08bf80995eda9a018993739b909a7859875970070.jpg", + "image_caption": [ + "(c) Superclass Concept Domain", + "Figure 1. Illustration of superclass learning. The samples from a same superclass will be scatteredly distributed in the embedding space. The process of superclass learning is to break old domains and construct new domains. Red indicates the superclass of kitchen waste, and blue indicates the superclass of recyclable waste." + ], + "image_footnote": [], + "bbox": [ + 764, + 422, + 875, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Refuse sorting, as an example, is such a recognition problem with four superclasses, i.e., kitchen waste, recyclable waste, hazardous waste, and others waste. One task of refuse sorting is to accurately collect various items, such as rotten fruits, bones, raw vegetables, and eggshells, into kitchen waste. Traditional recognition needs to identify what exact basic-level categories they are, then sort them out. Obviously, it is wasteful and unrealistic for superclass identification.", + "bbox": [ + 496, + 654, + 890, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Essentially, high-level superclasses contain two characteristics, remarkably distinct from basic-level classes. First, the basic-level classes contained in superclass problems are usually scattered and share few common features. As depicted in the top-left corner subgraph of Fig. 1, the fruit apple, bone, and eggs are remote from each other in feature spaces, though all of them belong to kitchen waste.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24060", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, the instances from two distinct superclasses may share common features. Just as illustrated in Fig. 1, fruit apple, from kitchen waste, and toy apple, from recyclable waste, are close to each other as they share common semantic features. Obviously, the above-mentioned characteristics indicate that the smoothness assumption [27] under basic-level classification (nearby images tend to have the same label) does not hold in the superclass scenarios. Thus, the existing classification techniques based on smoothness assumption are not practically deployable and scalable and they may suffer severe performance damage in superclass settings. Accordingly, it is valuable and promising to investigate superclass identification.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the superclass problems, we have to address two main challenges. First, we need to break the original decision boundaries of basic-level classes and disclose a superclass-aware boundary at the basic class level. As depicted in bottom subgraph (a) of Fig. 1, the boundary of the apple domain is original, however, it is useless and even harmful for the refuse sorting as both fruit apples and toy apples belong to this domain. To get the required domain boundary, the apple domain needs to be separated into the fruit domain and toy domain by leveraging their individual local features, as depicted in the bottom subgraph (b) of Fig. 1. In superclass scenarios, it is not enough to investigate the boundary at a basic class level. Consequently, the second challenge is to reconstruct a decision boundary at a superclass level. To achieve this end, it is necessary to merge the domain of classes, such as fruit apples, eggs, and bones into a new rotten superclass domain, as depicted in the bottom subgraph (c) of Fig. 1.", + "bbox": [ + 75, + 287, + 472, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a SuperClass Learning framework with Representation Enhancement. Considering that the semantic representation at the basic class level is not workable for superclass recognition, we propose one cross-instance attention module which could seize the representation across the instances with the same superclass label. By leveraging contrastive adjustment loss, the attention mechanism enhances this representation. Moreover, to overcome the imbalance distribution of superclasses, we adopt target adjustment loss to reconstruct a superclass-aware decision boundary on the enhanced representation space.", + "bbox": [ + 75, + 560, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this paper makes the following contributions:", + "bbox": [ + 76, + 726, + 470, + 755 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an under-study but realistic problem, superclass identification, that has notably distinct distribution from basic-level categorization.", + "- We propose a novel representation enhancement method by leveraging cross-instance attention and then exploit it in superclass identification. And by theoretical analyses, we verify that this self-attention technique can bound the generalization error of superclass" + ], + "bbox": [ + 94, + 768, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recognition.", + "bbox": [ + 532, + 92, + 614, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments demonstrate that SCLRE outperforms the SOTA classification techniques on one artificial superclass dataset and three real datasets.", + "bbox": [ + 517, + 114, + 890, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remainder of the paper is organized as follows: in Sec. 2 we briefly review related work and in Sec. 3 we describe our method for superclass recognition. Then, extensive experiments and generalization error analysis are conducted in Secs. 4 and 5. Finally, Sec. 6 draws a brief conclusion.", + "bbox": [ + 496, + 167, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 271, + 640, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Learning. The contrastive framework has achieved great success on instance-level problems. Chen et al [3] designed a simple framework to get visual representations in a contrastive way. He et al [14] developed a dynamic dictionary with queues and moving averages to improve contrastive encoder quality. To learn basic-level features and discover more semantic structures for the embedding space, Li et al [22] bridge contrastive learning with clustering. More recently, Caron et al [2] presented a contrastive framework without the need for pairwise comparison computation. To get rid of the limitation of negative samples, Grill et al [12] use two neural networks and predict the representation of the same image under a different augmented view. Chen et al [4] further designed a stop-gradient operation to prevent collapsing. Contrastive learning has also been applied to various research domains and evolved quickly [11, 33, 34]. Instead of focusing on the basic-level boundary, we construct a high-level superclass-aware boundary.", + "bbox": [ + 496, + 296, + 893, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Contrastive Loss. Traditional contrastive frameworks encode data either unsupervised or self-supervised and apply supervised information in downstream tasks. Khosla et al [19] extend the contrastive mode to a fully-supervised setting and use label information more effectively. To address the impact of long-tailed distribution data, Li et al [23] propose a targeted contrastive loss and guide the encoder to obtain representations based on several preset anchors. Recently, supervised contrastive learning is used in broader research domains both in computer vision and natural language processing [16,25, 35]. We are inspired by the function of pulling close or away features in these supervised contrastive losses, and therefore propose a new loss, which will pull the features by the superclass labels for better performance in classification.", + "bbox": [ + 496, + 584, + 893, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Attention Module. Self-attention module has been widely used in deep learning. It was first mentioned by Vaswani et al [29] as a part of the transformer model and is widely used in the natural language processing field [5, 21, 24, 26]. Recently, the attention module and transformer also obtained great success in the computer vision", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24061", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/81486a35630af085eae0ba8a9fbd1854d1bb712364438f0ccbb74a45bd629405.jpg", + "image_caption": [ + "Figure 2. Illustration of SCLRE. The images first generate their representations through a convolutional neural network, then mix with each other in a trainable cross-instance attention module for enhancement. After enhancement, the representations are then adjusted according to their superclass labels and the target anchors." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 455, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "field. Dosovitskiy et al [6] organize image patches as a sequence and propose a vision transformer. Furthermore, He et al [13] propose a masked auto-encoder based on vision transformers. Self-attention modules can also be integrated into convolutional neural networks to improve performance [1, 18, 31]. Fu et al [10] address the fine-grained image recognition problem by using an attention network to find the unique and pivotal features in samples (for example, one kind of bird's tail color). Different from focusing the attention relationship in one sample, we propose a cross-instance attention(CIA) module that cares more about the relationship between instances.", + "bbox": [ + 75, + 463, + 472, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 679, + 168, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Superclass Learning with Representation Enhancement(SCLRE) framework is proposed for superclass image recognition. It exploits the self-attention technique across instances to perform a representation enhancement, thus it achieves the goal to break the basic-level boundary in representations. Then SCLRE reconstructs a new superclass-aware boundary in the enhanced representation space by a series of adjustment losses.", + "bbox": [ + 75, + 712, + 468, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the overview of SCLRE. Sec. 3.1 presents the process of breaking basic-level boundary, while Sec. 3.2 describes the details of constructing the new high-level superclass-aware boundary.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4a9f5c0094da18574f9cf87e87c98adbfecaff13cb10171d816766958c13456.jpg", + "image_caption": [ + "Figure 3. Different segmentation strategies of cross-instance attention module. We experiment on three kinds of segmentation strategies and choose the best for the experiment." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 893, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Representation Enhancement", + "text_level": 1, + "bbox": [ + 500, + 239, + 764, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To break the basic-level boundary of representations and obtain valuable enhancements, we mix the representations by leveraging a cross-instance attention module. Unlike the existing attention techniques [32], the cross-instance attention module cares more about the relationship between representations of the instances, rather than the feature representation inside the instance.", + "bbox": [ + 496, + 263, + 890, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Enhanced Representation. Let $\\mathcal{X}$ be the data space, $\\mathcal{R}^D$ be the $D$ -dimensional embedding space, an encoder $f: \\mathcal{X} \\to \\mathcal{R}^D$ be a mapping from the data space to the embedding space, and an enhancement process $\\mathrm{EnH}: \\mathcal{R}^D \\to \\mathcal{R}^D$ be a mapping from the embedding space to the embedding space. For each $x_j \\in \\mathcal{X}$ , the representation $z_j = f(x_j)$ , and $Z \\in \\mathcal{R}^{n \\times D}$ is the representation matrix for the batch with size $n$ . The enhanced representation $v_j$ is an output of $\\mathrm{EnH}$ . More specifically, $v_j$ is defined as follows:", + "bbox": [ + 498, + 368, + 892, + 505 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv _ {j} = \\operatorname {E n H} \\left(z _ {j}; Z\\right) = a _ {j} \\times Z, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 518, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $a_{j}$ is a trainable vector for representation $z_{j}$ computed by cross-instance attention module. Actually, the vector $a_{j}$ is the $j$ -th row of the attention distribution matrix(ADM) computed in the following Q-K-V mode multi-head attention module [29]:", + "bbox": [ + 496, + 542, + 890, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A D M} (Z) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 627, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The matrix $Q = Z \\times W_{Q}$ , where $W_{Q} \\in \\mathcal{R}^{D \\times d_{q}}$ . Matrix $K = Z \\times W_{K}$ , where $W_{K} \\in \\mathcal{R}^{D \\times d_{k}}$ and $d_{q} = d_{k}$ .", + "bbox": [ + 500, + 672, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The enhancement process breaks the basic-level boundary by merging the semantic features of instances. Cross-instance attention module automatically explores the key instances for representation enhancement and integrates them to obtain a superclass-aware enhanced representation. Moreover, we conduct strict theoretical analysis on the generalization ability of our model in Sec. 5, and discover that the cross-instance module can tightly bound the generalization error of SCLRE.", + "bbox": [ + 496, + 704, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Segmentation Strategy. In the cross-instance attention module, the segmentation strategy of input data is important for representation enhancement. For each instance, the segmentation strategy determines whose repre", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24062", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sentations will be received as an enhancement. We considered three kinds of segmentation strategies for the input data in the cross-instance attention module. Fig. 3 illustrates the details of the strategies. In (a), images are gathered in terms of their superclass label, we put the same label images into one batch. In (b), we randomly and evenly select images from the different superclasses and gather them into the batch. In (c), we keep the distribution of the dataset and randomly select the samples to form the batch. The result shows that (c) selecting the samples according to the distribution of the dataset benefits the model most, so we fixed this strategy in the following part of this work.", + "bbox": [ + 75, + 90, + 472, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SCLRE Loss", + "text_level": 1, + "bbox": [ + 76, + 279, + 215, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reconstruct the high-level superclass-aware boundary of representations, we design a SCLRE loss to adjust the representations after the attention-weighted enhancement. We keep using the letter $v$ in Eq. (1) to demonstrate representations being enhanced by the cross-instance attention module, and we call it enhanced representation. After that, we adjust the representations by taking three aspects into consideration.", + "bbox": [ + 75, + 303, + 468, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category Classification Loss. We first adopt the cross-entropy loss as the base to obtain a category-level adjustment and prevent the model from collapsing:", + "bbox": [ + 75, + 422, + 468, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i, k} \\log M (v _ {i}) _ {k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 474, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ demonstrates batch size, $y_{i}$ demonstrates the superclass label of the $i$ -th sample in one-hot embedding, $M(v_{i})$ demonstrates the output prediction of $v_{i}$ , and the footprint $k$ demonstrates the $k$ -th component of a vector.", + "bbox": [ + 76, + 521, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Adjustment Loss. Besides classification loss, the samples belonging to the same superclass may share few common features. To obtain a representation-level adjustment, we design a contrastive adjustment loss to pull close the samples in the same superclass and pull away samples in distinct subclasses. We directly regard samples from the same superclass as positive pairs and samples from other subclasses as negative pairs. For one sample $v_{i}$ and one of its positive sample $v^{+}$ (i.e. $v_{i}$ and $v^{+}$ belongs to the same superclass), the contrastive adjustment loss is:", + "bbox": [ + 75, + 582, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell \\left(v _ {i}, v ^ {+}\\right) = - \\log \\frac {\\exp \\left(s \\left(v _ {i} , v ^ {+}\\right) / \\tau\\right)}{\\exp \\left(s \\left(v _ {i} , v ^ {+}\\right) / \\tau\\right) + \\sum_ {v ^ {-} \\notin P \\left(v _ {i}\\right)} \\exp \\left(s \\left(v _ {i} , v ^ {-}\\right) / \\tau\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 739, + 467, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By summing up all the losses of positive pairs and for a batch with size $N$ , we can get the total loss as:", + "bbox": [ + 76, + 784, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c a} = \\sum_ {i = 1} ^ {N} \\frac {1}{| | P (v _ {i}) | | - 1} \\sum_ {v ^ {+} \\in P (v _ {i}) \\backslash \\{v _ {i} \\}} \\ell \\left(v _ {i}, v ^ {+}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 820, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P(v_{i})$ stands for all the other data that share the same superclass label with $v_{i}$ in the batch, i.e. $P(v_{i}) =$", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{v_{k}|p(v_{k}) = p(v_{i})\\}$ $p$ means the superclass label. $s(\\cdot ,\\cdot)$ stands for the similarity between two vectors, we use cosine similarity for measurement, i.e. $s(v_{s},v_{t}) = \\frac{v_{s}^{T}v_{t}}{\\|v_{s}\\|\\cdot\\|v_{t}\\|}$ $\\tau$ is the temperature hyperparameter. $||P(v_i)||$ stands for the size of set $P(v_{i})$ , and it aims to normalize the total loss.", + "bbox": [ + 498, + 90, + 890, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Targeted Adjustment Loss. Since superclass learning naturally faces the problem of lack of classification centers due to the scattered distribution of representations, then follows damage to the constructing process when the contrastive adjustment loss works. To address this problem, we design a targeted adjustment loss to give each superclass a pre-defined category center (i.e. target anchors). We first generate the target anchors and make them far away from each other inspired by [23]. Then we pull every sample close to its pre-defined target anchor by minimizing the following targeted adjustment loss:", + "bbox": [ + 496, + 174, + 892, + 339 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t a} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (s \\left(v _ {i} , t \\left(v _ {i}\\right)\\right) / \\tau)}{\\sum_ {t \\in T} \\exp (s \\left(v _ {i} , t\\right) / \\tau)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 354, + 890, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ stands for batch size, $s(\\cdot ,\\cdot)$ stands for cosine similarity function, $T$ is the set of all the target anchors, and $t(v_{i})$ demonstrates the target anchor allocated to $v_{i}$ .", + "bbox": [ + 498, + 410, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SCLRE Adjustment Loss. To construct the high-level superclass-aware boundary with both category-level and instance-level adjustments and overcome the problem of lack of classification center, the total loss of SCLRE is a weighted summation of the above three losses(i.e. Eqs. (3), (5) and (6)):", + "bbox": [ + 498, + 458, + 892, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S C L R E} = (1 - \\alpha) \\mathcal {L} _ {c e} + \\alpha \\mathcal {L} _ {c a} + \\lambda \\mathcal {L} _ {t a}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 564, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\lambda$ range from 0 to 1. $\\mathcal{L}_{ce}$ and $\\mathcal{L}_{ca}$ are designed as a combination to help the encoder explore the concept of superclasses, a larger $\\alpha$ means a stronger adjustment strategy and weaker classification ability. $\\mathcal{L}_{ta}$ is designed to help the model overcome the imbalanced distribution, and a larger $\\lambda$ means stronger guidance towards the destination of adjustment.", + "bbox": [ + 496, + 595, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SCLRE adjustment loss can help the model learn more about the concept of the superclass and the high-level categorization boundary. By balancing the influence of cross-entropy loss and contrastive adjustment loss, the model can adjust representations properly. By controlling the utilization of targeted adjustment loss, the model can run smoothly and converge steadily.", + "bbox": [ + 496, + 703, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 828, + 624, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments on several artificially constructed superclass datasets to demonstrate the effectiveness of our proposed approach, SCLRE.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24063", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/609df7a44d6c32fd6a63754200b04bdceb25bd2698b0a0861e9a4a1924098e1c.jpg", + "image_caption": [ + "Figure 4. Illustration of superclass construction. We integrate the initial class labels of the original dataset to superclass labels, according to field knowledge. The constructed dataset contains a handful of subclasses." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 472, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 362, + 253, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In this study, superclass datasets are artificially reorganized from three benchmarks, CIFAR-100, mini-ImageNet [30] and VOC [8], and two real-world datasets, FMoW [20] and Adience [7]. CIFAR-100 contains 50000 training images and 10000 test images from 100 categories with a size of $32 \\times 32 \\times 3$ . mini-ImageNet [30] contains 60000 images in total with a size of $84 \\times 84 \\times 3$ , which are evenly distributed among 100 classes. VOC [8] contains 2501 training images, 2510 validation images and 4952 test images from 20 categories. The real-world dataset FMoW [20] is about a hybrid domain generalization and subpopulation shift problem, which consists of 76863 training images, 11483 validation images and 11327 test images with a size of $224 \\times 224 \\times 3$ . And the real-world dataset Adience [7] contains 26,580 images from 2,284 people.", + "bbox": [ + 75, + 385, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, superclass datasets are artificially constructed by integrating the initial classes, just as illustrated in Fig. 4. To be concrete, We reorganize the 100 classes of CIFAR-100 into 3 subclasses, 4 subclasses, and 7 subclasses, respectively, forming three distinct superclass datasets, CIFAR100-3, CIFAR100-4, and CIFAR100-7. The subclasses in the three data sets are based on movement mode, life form, and are more complex than the original 20 coarse classes in CIFAR-100. For miniImagenet and VOC, we reorganize their subclasses into 2 subclasses based on life form. For FMoW, we drop the original labels and turn to predict the geographical location of the images. For Adience, we utilize the age groups as the superclass labels. Thus, those 7 superclass datasets can well simulate the distribution of real-world superclass problems. More reorganization details are listed in the Appendix.", + "bbox": [ + 75, + 613, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared Methods. We compare SCLRE with the baseline model, i.e., ResNet [15], and two SOTA contrastive techniques SupCon [19], and SimCLR [3]. As contrastive", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/56a1fc5e498a00aa071ea332bdcd0a086acfeb1b67484df9281a2fb8c3f7fe84.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodAccuracy(%)
CIFAR100-3Baseline72.8
SupCon [19]78.1
SimCLR [3]79.0
SCLRE80.1
CIFAR100-4Baseline76.0
SupCon [19]80.1
SimCLR [3]80.6
SCLRE84.0
CIFAR100-7Baseline68.9
SupCon [19]72.7
SimCLR [3]73.9
SCLRE78.1
", + "bbox": [ + 535, + 87, + 859, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Classification accuracy on low pixels dataset: CIFAR-100. We compared the classification accuracy on CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets.", + "bbox": [ + 498, + 300, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "techniques perform excellently in representation learning. We adopt the default optimal settings in the training details.", + "bbox": [ + 498, + 366, + 890, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SupCon: In our experiments, SupCon adopts ResNet-50 as backbone, and SupCon loss as the super-classification loss.", + "bbox": [ + 498, + 396, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SimCLR: It is a known self-supervised contrastive learning framework. In superclass recognition, we keep using ResNet-50 as the backbone and use a supervised downstream task to keep the instance-level features pure, wherein the upstream instance-level results are a kind of learning of the basic level class labels.", + "bbox": [ + 496, + 441, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We adopt ResNet-50 [15] as our default backbone and we reduce the feature dimension from 2048 to 128 with an extra Multi-Layer Perceptron(MLP) projection head. We train the network using the SGD with a momentum of 0.9 and a batch size of 64. We set the learning rate as 0.001 and the training stage as 200 epochs. Additionally, we resize all the images to $32 \\times 32$ pixels. For our SCLRE method, we set the key $\\mathrm{K} = 256$ , value $\\mathrm{V} = 128$ and 8 multi-heads are adopted in the attention module.", + "bbox": [ + 496, + 532, + 890, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Classification Evaluation", + "text_level": 1, + "bbox": [ + 500, + 695, + 725, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accuracy on CIFAR-100 Datasets. Tab. 1 shows the classification accuracies of SCLRE and the compared methods on CIFAR100-3, CIFAR100-4, and CIFAR100-7, respectively. And we have the following observations: 1) It is observed that SCLRE has a great improvement compared with the baseline. This shows that the cross-instance attention module can effectively activate the contrastive adjustment loss. 2) SCLRE has higher classification accuracies than SimCLR on superclass datasets. This shows that the instance-level semantic information learned by SimCLR may fail in the superclass problem because the images from the same superclass may have quite different, even con", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24064", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06f53139251811f5874f0e64e7e48052768f42b461ad91ea8db539a69374ee69.jpg", + "image_caption": [ + "CIFAR100-7" + ], + "image_footnote": [], + "bbox": [ + 184, + 99, + 308, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0479c9dc9936b3fd319e48b6ee109247249ab34b61c6405043e8042fa442f74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 101, + 450, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b3f8ff954d68f6ffa179375ea2da723a0b11d3f966aee4c16c2f3d2838da5528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 101, + 594, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58df38e3ae129d775a41f5cdecb19c03590bfb1820c03511d3cf8a0312365046.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 101, + 735, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/715d8d83c2103e1968a11ab02845731529af4dd50125808b9d0243b72cd2baed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 101, + 879, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e7b5f7ae7b935b8cdece9d7650035165186783fccfc94faaefed562971db8db.jpg", + "image_caption": [ + "CIFAR100-4" + ], + "image_footnote": [], + "bbox": [ + 186, + 200, + 307, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cec2af3f87cdfaae066591103d22423c1a36d8c7d16b4e55e834a5175cdca01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 200, + 450, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/99547b98d18370f43647e53c6b43ef07b154a240f8de9b343abec3531d0ec552.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 472, + 200, + 594, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2cff40e1c84402a1573929e3762c70364ced9b4d68c338cdf987e5114154357.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 200, + 733, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f619e0ee88af89763c3d4dc261feed3fb6533d4f55557339888b2a0881ead613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 202, + 877, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c03d9cfa75d7f5094416e3d3f7f73c20f630119605d0ee8b5137cfe29295cc66.jpg", + "image_caption": [ + "CIFAR100-3", + "(a) Baseline", + "Figure 5. Visualization results by t-SNE. We sampled part of the test set and encode them into representations, then visualize them in the embedding space by t-SNE. (a) We directly generate the representations using the backbone as a baseline. (b) We use the upstream encoder of SimCLR. (c) We use the encoder of SupCon. (d) We use part of SCLRE by removing the targeted adjustment loss. (e) We use the entire SCLRE." + ], + "image_footnote": [], + "bbox": [ + 186, + 301, + 307, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9380385bbebbd3e11920fda1ea224b778aedc3c4af528af16ec8a6a8ca88ea98.jpg", + "image_caption": [ + "(b) SimCLR" + ], + "image_footnote": [], + "bbox": [ + 328, + 301, + 450, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/655630c26a3a012c46fbf843d8f0ccdd7a9edcf1af41b11cf0d31f3abbd3e994.jpg", + "image_caption": [ + "(c) SupCon" + ], + "image_footnote": [], + "bbox": [ + 472, + 301, + 594, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b431494ab043b92ca92389694a2536816674a73d69413df42d422792a14225f2.jpg", + "image_caption": [ + "(d) SCLRE w\\o targets" + ], + "image_footnote": [], + "bbox": [ + 614, + 301, + 735, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0069a6cb1947ee08f4d16d96fed3bf4db794432fc8dfe782c67430572faca62.jpg", + "image_caption": [ + "(e) SCLRE" + ], + "image_footnote": [], + "bbox": [ + 756, + 301, + 877, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tradictory semantic information. Comparatively, SCLRE works well by the superclass-aware decision boundary.", + "bbox": [ + 75, + 526, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accuracy on High-Resolution Datasets. Tab. 2 shows the evaluation results on three high-resolution datasets, from which we observe the following facts. 1) we observe that, on the real-world dataset FMoW, SCLRE outperforms other methods by a large margin. On FMoW, the geographic location of building images is adopted as the superclass label, rather than their type. The methods without representation enhancement may perform worse due to the lack of superclass-aware common features. This further confirms the importance of representation enhancement in the superclass problems. 2) SCLRE also outperforms other methods on real-world dataset Adience, which contains 26580 face images from 2284 people. In case every single person is a raw class, the results confirm that SCLRE is also workable in large-scale situations. 3) Moreover, we observe that on FMoW and Adience, SimCLR performs even worse than the baseline. Because the instance-level representation by SimCLR is conflicted with the given superclass labels. 4) The performance on mini-[ImageNet appears to be less attractive. This may be because the mini-[ImageNet dataset is divided into 2 distinct subclasses: animals and non-animals. For such kind of distribution, basic-level methods", + "bbox": [ + 75, + 568, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb51c939d8ae35367f7a268126eb1ce8cc10a075ecd66c5fe743ca007f651d1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mini-ImageNet [30]FMoW [20]Adience [7]VOC [8]
Baseline87.756.565.778.3
SupCon87.058.366.278.2
SimCLR90.151.552.980.0
SCLRE89.364.868.781.5
", + "bbox": [ + 503, + 523, + 888, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Classification accuracy on high-resolution datasets. We compared the accuracy on datasets with more complex and informative content.", + "bbox": [ + 498, + 594, + 890, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "can also easily sort them out by exploring the obvious common features inside each superclass(e.g. whether the image contains eyes), which makes it more like a traditional instance-level problem, and SimCLR may perform comparably or even slightly better than SCLRE.", + "bbox": [ + 498, + 652, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Visualization", + "text_level": 1, + "bbox": [ + 500, + 739, + 637, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we generate the visualization results on representation space by the t-SNE technique [28]. The CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets are adopted for the visualization experiments.", + "bbox": [ + 498, + 763, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe the following facts from Fig. 5. 1) It is observed that in columns (a) and (b) the images are scattered in their respective representation space, although the images (b) cluster slightly closer than the ones in (a). This further shows that the instance-level representation of SimCLR is", + "bbox": [ + 496, + 824, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24065", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "little related to superclass labels. Consequently, in superclass scenarios, SimCLR performs a little better than baseline but worse than SCLRE, just as illustrated in Tab. 1. 2) We observed that the images in column (c) have a more obvious cluster structure than in columns (a) and (b). This shows that the use of superclass labels is important for constructing clear boundaries. But due to the diverse and complex superclass contents, the representations are mostly inaccurate and thus result in less favorable results as reported in Tab. 1. 3) It is also observed that the images in columns (d) and (e) are notably closer than the ones in (a), (b), and (c). This shows that the enhanced representation of SCLRE is superclass-aware. 4) Furtherly, we observe that the images in column (e) are obviously closer than the ones in column (d). This shows that the targeted adjustment loss, one important component of SCLRE, can furtherly cluster images in the same subclasses, accordingly, it benefits reconstructing a superclass-aware decision boundary.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 76, + 371, + 264, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We discuss the hyperparameters $\\alpha$ and $\\lambda$ of SCLRE by evaluating their sensitivity on the CIFAR100-7 dataset. When analyzing $\\alpha$ , to avoid the influence of target anchors, we removed targeted adjustment loss by fixing $\\lambda$ to 0. We test the sensitivity of $\\alpha$ by ranging it from 0 to 1. When analyzing $\\lambda$ , we fix $\\alpha$ to 0.5 to avoid its influence.", + "bbox": [ + 75, + 393, + 468, + 484 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/418e5259c0adf52acfbebe0d4b6a24a4cc92e6615bd8fc333a43ab1ee802159d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
α00.10.30.50.71
Accuracy69.272.173.775.173.6-
", + "bbox": [ + 96, + 494, + 450, + 530 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/39a5a9d5c68ca14b1d993419c3ac2ff38928ca8e25715302152d0f3cc61844b0.jpg", + "table_caption": [ + "Table 3. Sensitivity analysis of $\\alpha$ . We range $\\alpha$ from 0 to 1 and calculate the accuracy of SCLRE respectively. When $\\alpha = 1$ , the cross-entropy loss is invalid and the model will lose its classification ability." + ], + "table_footnote": [], + "table_body": "
λ00.10.20.40.81
Accuracy75.176.777.976.978.176.0
", + "bbox": [ + 86, + 612, + 460, + 647 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df104f321c4edddcd7d6c00dc504608951b77f5869f8d1a411312e8511fa13f5.jpg", + "table_caption": [ + "Table 4. Sensitivity analysis of $\\lambda$ . We range $\\lambda$ from 0 to 1 and calculate the accuracy of SCLRE respectively." + ], + "table_footnote": [], + "table_body": "
Batch Size64128256512
Accuracy76.776.977.377.9
", + "bbox": [ + 130, + 702, + 416, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sensitivity of $\\alpha$ . Tab. 3 shows the sensitivity of $\\alpha$ . Besides extreme situations, we find that $\\alpha$ has a little influence on the accuracy. The optimal value of $\\alpha$ falls around 0.5 and we fix it in our experiments.", + "bbox": [ + 75, + 780, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the case of $\\alpha = 0$ , the contrastive adjustment loss is invalid and the model degenerates into the baseline. In the case of $\\alpha = 1$ , the cross-entropy loss is invalid and then the model is not workable for classification.", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc01d5d462bc24164d625ba959f00e28ec6a87209dc8a011e3bb95411e6d246c.jpg", + "table_caption": [ + "Table 5. Sensitivity analysis of batch size. We vary batch size from 64 to 512 and calculate the accuracy of SCLRE respectively." + ], + "table_footnote": [], + "table_body": "
ArchitectureParams.(M)FLOPs(G)Acc.(%)
Baseline23.511.3068.9
+CIA37.211.3169.2
+Lca--72.7
+CIA, Lca--75.1
SCLRE (+CIA, Lca, Lta)37.211.3277.9
", + "bbox": [ + 517, + 87, + 870, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on SCLRE. We calculated the model size (Params.), computational complexity (FLOPs) and accuracy contributions of four components of SCLRE.", + "bbox": [ + 498, + 209, + 890, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sensitivity of $\\lambda$ . Tab. 4 demonstrates the sensitivity analysis of $\\lambda$ . $\\lambda$ is a hyperparameter that controls the influence of targeted adjustment loss.", + "bbox": [ + 498, + 276, + 890, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike $\\alpha$ , there are two most proper $\\lambda$ values which fall around 0.2 and 0.8. When $\\lambda$ is too small or too large, the performance is both less satisfactory. As a too large $\\lambda$ may be harmful to the accuracy of the model, a too small $\\lambda$ may be short of guidance and has limited performance improvement.", + "bbox": [ + 498, + 325, + 890, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sensitivity of Batch Size. Intuitively, a larger batch size can bring more abundant samples of superclasses, thus making it easier to construct superclass-aware representations. The experimental results are conducted to verify this statement. Tab. 5 illustrated the results of the sensitivity study of the batch size. The results show that a larger batch size can improve the performance of SCLRE, though the improvement is not significant.", + "bbox": [ + 498, + 420, + 890, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 561, + 653, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here ablation studies are conducted to demonstrate the contribution of each component in our proposed SCLRE. Just as designed in Sec. 3, SCLRE is composed of the baseline model, contrastive adjustment loss $(\\mathcal{L}_{ca})$ , cross-instance attention module(CIA), and targeted adjustment loss $(\\mathcal{L}_{ta})$ . The baseline model is the same as our SCLRE without $\\mathcal{L}_{ca}$ , $\\mathcal{L}_{ta}$ and CIA.", + "bbox": [ + 498, + 590, + 890, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to the results reported in Tab. 6, we have the following observations:", + "bbox": [ + 498, + 700, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1) It is observed that only CIA itself has an observable contribution to the performance improvements, but CIA joint with $\\mathcal{L}_{ca}$ brings notable improvements. This indicates that the CIA module is workable to break the basic class boundary with the great assistance of $\\mathcal{L}_{ca}$ . They enhance the representation of the superclass by cross-instance attention mechanism. 2) With CIA, $\\mathcal{L}_{ca}$ and $\\mathcal{L}_{ta}$ , SCLRE achieves the optimal improvements. It shows that the targeted adjustment loss does function on reconstructing superclass-aware boundary on the enhanced representation space.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24066", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f0eb61f22e22ac0e99f43c1275b8669fc8824434f20bc60d29f512c58ab6b653.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ResNet-18ResNet-34ResNet-50ResNet-101ResNet-152
Baseline83.983.778.283.482.6
SCLRE86.887.682.987.787.2
Improve.(%)+2.9+2.9+4.7+4.3+4.6
", + "bbox": [ + 192, + 87, + 781, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Robustness research of SCLRE on different backbones. We calculated the accuracy on a CIFAR100-3 superclass dataset of baseline and SCLRE. We also calculated the improvement SCLRE gets in the different backbones.", + "bbox": [ + 75, + 162, + 893, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Robustness for Different Backbones", + "text_level": 1, + "bbox": [ + 76, + 213, + 390, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the robustness of SCLRE by changing the backbone to other convolutional neural networks. Tab. 7 shows the result of the experiments on robustness. The experiments are performed on a CIFAR100-3 superclass dataset and we vary the backbone from smaller networks to larger networks. The result shows that SCLRE can make stable improvements on different backbones.", + "bbox": [ + 75, + 237, + 470, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Analysis of Generalization Ability", + "text_level": 1, + "bbox": [ + 76, + 354, + 385, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an analysis of the generalization ability of SCLRE and prove that the cross-instance attention module can further compress the upper bound of the generalization error. The details of the proof are in the Appendix.", + "bbox": [ + 75, + 378, + 470, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the contrastive adjustment loss in Sec. 3.2, and with ignoring the constant items, we can rewrite the contrastive loss into an expectation form:", + "bbox": [ + 75, + 439, + 470, + 484 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal{L}_{ca}\\propto -\\mathbb{E}_{v,v^{\\prime}}\\mathbb{E}_{\\substack{v_{1},v_{2}\\in P(v)\\\\ v^{-}\\in P(v^{\\prime})}}\\log \\frac{\\exp(v_{1}^{T}v_{2})}{\\exp(v_{1}^{T}v_{2}) + \\exp(v_{1}^{T}v^{-})} \\\\ = \\underbrace {- \\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2}} _ {L _ {1}} \\\\ + \\underbrace {\\mathbb {E} _ {v , v ^ {\\prime}} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} \\log \\left(\\exp (v _ {1} ^ {T} v _ {2}) + \\exp (v _ {1} ^ {T} v ^ {-})\\right)} _ {L _ {2}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 489, + 455, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $L_{1}$ measures the alignment between two transformed features and $L_{2}$ is the regularizer preventing the collapse of representation.", + "bbox": [ + 75, + 616, + 468, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lemma 1. As a conclusion in [17], if $L_{2}$ is trained to be small than the threshold, which is easily satisfied, the generalization error of downstream classifier $G_{f}$ has an upper bound:", + "bbox": [ + 75, + 661, + 468, + 720 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E r r (G _ {f}) \\leqslant (1 - \\sigma) + \\eta (\\epsilon) \\sqrt {2 - 2 \\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2}} \\\\ = (1 - \\sigma) + \\eta (\\epsilon) \\sqrt {2 + 2 L _ {1}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 724, + 457, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lemma 2. Sample pairs with the same coarse label consist of two parts: those share the same fine label $(O_1)$ and those have different fine labels $(O_2)$ . The former can be closely aligned after contrastive learning. We focus on the latter and prove that:", + "bbox": [ + 75, + 792, + 468, + 868 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1}, v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2} \\geqslant C _ {\\varphi} + \\frac {(1 - \\rho) K}{M _ {1} M _ {2}} s (a _ {1}, a _ {2})\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 872, + 426, + 905 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $C_{\\varphi}$ is a constant determined by the data distribution. $M_1, M_2, K$ are constants. $\\rho$ is the ratio of sample pairs with same fine label in all positive pairs.", + "bbox": [ + 500, + 214, + 890, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 1. The generalization error has another upper bound based on attention vectors according to lemma 1 and lemma 2:", + "bbox": [ + 498, + 263, + 893, + 308 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E r r (G _ {f}) \\leqslant (1 - \\sigma) + \\\\ \\sqrt {2} \\eta (\\epsilon) \\sqrt {1 - C _ {\\varphi} - \\frac {(1 - \\rho) K}{M _ {1} M _ {2}} \\mathbb {E} _ {v _ {1} , v _ {2} \\in O _ {2}} s (a _ {1} , a _ {2}))} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 325, + 867, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The attention vector is a representation of the importance between samples. When we define a superclass, the representation vector $z$ with the same coarse label will be pulled together by the contrastive loss. During this process, samples in the superclass will focus on the same important enhanced representations, thus leading to similar vectors and declining the upper bound of generalization error.", + "bbox": [ + 498, + 409, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 541, + 619, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we explore the high-level categories recognition problem, i.e., superclass categorization. And a superclass learning framework, that exploits self-attention techniques cross the instances to enhance the representation, is proposed. Thus, the distribution of superclass is modified in the enhanced representation space and a new superclass-aware decision boundary is then reconstructed. In theory, the generalization error of SCLRE can be bounded by attention constraints. In extensive experiments, SCLRE outperforms the SOTA classification methods.", + "bbox": [ + 496, + 570, + 890, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the near future, we would like to tackle the superclass problems under the scenarios without full annotation by leveraging semi-supervised learning and active learning.", + "bbox": [ + 496, + 724, + 890, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Key Research & Develop Plan(2018YFB1004401); National Natural Science Foundation of China (62276270,62072460, 62172424); Beijing Natural Science Foundation (4212022). it is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence.", + "bbox": [ + 496, + 787, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24067", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amjad Almahairi, Nicolas Ballas, Tim Coolijmans, Yin Zheng, H. Larochelle, and Aaron C. Courville. Dynamic capacity networks. In ICML, 2016. 3", + "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2", + "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. ArXiv, abs/2002.05709, 2020. 2, 5", + "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 2", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 3", + "[7] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Transactions on information forensics and security, 9(12):2170-2179, 2014. 5, 6", + "[8] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2009. 5, 6", + "[9] Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2010. 1", + "[10] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4476-4484, 2017. 3", + "[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2", + "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3", + "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2", + "[15] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5", + "[16] Qiushi Huang, Tom Ko, H Lilian Tang, Xubo Liu, and Boyong Wu. Token-level supervised contrastive learning for punctuation restoration. In Interspeech, 2021. 2", + "[17] Weiran Huang, Mingyang Yi, and Xuyang Zhao. Towards the generalization of contrastive self-supervised learning. arXiv preprint arXiv:2111.00743, 2021. 8", + "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In NIPS, 2015. 3", + "[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. ArXiv, abs/2004.11362, 2020. 2, 5", + "[20] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton A. Earnshaw, Imran S. Haque, Sara Beery, Jure Leskovec, Anshul Kundaje, Emma Pierson, Sergey Levine, Chelsea Finn, and Percy Liang. WILDS: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning (ICML), 2021. 5, 6", + "[21] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 2", + "[22] Junnan Li, Pan Zhou, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Prototypical contrastive learning of unsupervised representations. ArXiv, abs/2005.04966, 2021. 2", + "[23] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogério Schmidt Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6908-6918, 2022. 2, 4", + "[24] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2", + "[25] Zheda Mai, Ruiwen Li, Hyunwoo J. Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3584-3594, 2021. 2", + "[26] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24068", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Lars Schmarje, Monty Santarossa, Simon-Martin Schroder, and Reinhard Koch. A survey on semi-, self- and unsupervised learning for image classification. IEEE Access, 9:82146-82168, 2021. 1, 2", + "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3", + "[30] Oriol Vinyals, Charles Blundell, Timothy P. Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 5, 6", + "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In-So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3", + "[32] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2020. 3", + "[33] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. ArXiv, abs/2010.13902, 2020. 2", + "[34] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-842, 2021. 2", + "[35] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10022-10031, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 571 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24069", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_model.json b/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40ffebd0689632f3b10f8f42c78a241b5e8cec70 --- /dev/null +++ b/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "Superclass Learning with Representation Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.18, + 0.292, + 0.199 + ], + "angle": 0, + "content": "Zeyu Gan\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.181, + 0.464, + 0.199 + ], + "angle": 0, + "content": "Suyun Zhao\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.2, + 0.465, + 0.217 + ], + "angle": 0, + "content": "Hong Chen\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.181, + 0.613, + 0.199 + ], + "angle": 0, + "content": "Jinlong Kang" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.181, + 0.774, + 0.199 + ], + "angle": 0, + "content": "Liyuan Shang" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.2, + 0.612, + 0.217 + ], + "angle": 0, + "content": "Cuiping \\(\\mathbf{L}\\mathbf{i}^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.217, + 0.868, + 0.253 + ], + "angle": 0, + "content": "1Key Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China \n2Renmin University of China, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.255, + 0.822, + 0.27 + ], + "angle": 0, + "content": "{zygan, zhaosuyun, kangjinlong, shangliyuan4032, chong, LICUping}@ruc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.685 + ], + "angle": 0, + "content": "In many real scenarios, data are often divided into a handful of artificial super categories in terms of expert knowledge rather than the representations of images. Concretely, a superclass may contain massive and various raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classification techniques are intractable to recognize superclass without raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhancement(SCLRE), to recognize super categories by leveraging enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE collapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision boundary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Experimentally, extensive results demonstrate that SCLRE outperforms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.714, + 0.21, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.47, + 0.877 + ], + "angle": 0, + "content": "In recent decades, basic-level raw categorization (e.g. cats vs dogs, apples vs bananas) has greatly developed [9, 27] while high-level or super-coarse-grained visual categorization (e.g., recyclable waste vs kitchen waste, creatures vs non-creatures) has received little attention. In many real scenarios, there often exist a handful of high-level categories, wherein numerous images from diverse basic-level categories share one common label. We tend to define this kind of super-coarse-grained class as Superclass." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.303, + 0.697, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.304, + 0.877, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.423, + 0.615, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.51, + 0.61, + 0.52 + ], + "angle": 0, + "content": "(a) Image Feature Domain" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.423, + 0.745, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.51, + 0.74, + 0.52 + ], + "angle": 0, + "content": "(b) Separate Fine Domain" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.423, + 0.877, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.51, + 0.885, + 0.52 + ], + "angle": 0, + "content": "(c) Superclass Concept Domain" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Illustration of superclass learning. The samples from a same superclass will be scatteredly distributed in the embedding space. The process of superclass learning is to break old domains and construct new domains. Red indicates the superclass of kitchen waste, and blue indicates the superclass of recyclable waste." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Refuse sorting, as an example, is such a recognition problem with four superclasses, i.e., kitchen waste, recyclable waste, hazardous waste, and others waste. One task of refuse sorting is to accurately collect various items, such as rotten fruits, bones, raw vegetables, and eggshells, into kitchen waste. Traditional recognition needs to identify what exact basic-level categories they are, then sort them out. Obviously, it is wasteful and unrealistic for superclass identification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Essentially, high-level superclasses contain two characteristics, remarkably distinct from basic-level classes. First, the basic-level classes contained in superclass problems are usually scattered and share few common features. As depicted in the top-left corner subgraph of Fig. 1, the fruit apple, bone, and eggs are remote from each other in feature spaces, though all of them belong to kitchen waste." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24060" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "Second, the instances from two distinct superclasses may share common features. Just as illustrated in Fig. 1, fruit apple, from kitchen waste, and toy apple, from recyclable waste, are close to each other as they share common semantic features. Obviously, the above-mentioned characteristics indicate that the smoothness assumption [27] under basic-level classification (nearby images tend to have the same label) does not hold in the superclass scenarios. Thus, the existing classification techniques based on smoothness assumption are not practically deployable and scalable and they may suffer severe performance damage in superclass settings. Accordingly, it is valuable and promising to investigate superclass identification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.56 + ], + "angle": 0, + "content": "To tackle the superclass problems, we have to address two main challenges. First, we need to break the original decision boundaries of basic-level classes and disclose a superclass-aware boundary at the basic class level. As depicted in bottom subgraph (a) of Fig. 1, the boundary of the apple domain is original, however, it is useless and even harmful for the refuse sorting as both fruit apples and toy apples belong to this domain. To get the required domain boundary, the apple domain needs to be separated into the fruit domain and toy domain by leveraging their individual local features, as depicted in the bottom subgraph (b) of Fig. 1. In superclass scenarios, it is not enough to investigate the boundary at a basic class level. Consequently, the second challenge is to reconstruct a decision boundary at a superclass level. To achieve this end, it is necessary to merge the domain of classes, such as fruit apples, eggs, and bones into a new rotten superclass domain, as depicted in the bottom subgraph (c) of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.473, + 0.727 + ], + "angle": 0, + "content": "In this paper, we propose a SuperClass Learning framework with Representation Enhancement. Considering that the semantic representation at the basic class level is not workable for superclass recognition, we propose one cross-instance attention module which could seize the representation across the instances with the same superclass label. By leveraging contrastive adjustment loss, the attention mechanism enhances this representation. Moreover, to overcome the imbalance distribution of superclasses, we adopt target adjustment loss to reconstruct a superclass-aware decision boundary on the enhanced representation space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.472, + 0.756 + ], + "angle": 0, + "content": "In summary, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.769, + 0.47, + 0.814 + ], + "angle": 0, + "content": "- We propose an under-study but realistic problem, superclass identification, that has notably distinct distribution from basic-level categorization." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel representation enhancement method by leveraging cross-instance attention and then exploit it in superclass identification. And by theoretical analyses, we verify that this self-attention technique can bound the generalization error of superclass" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.769, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.093, + 0.616, + 0.107 + ], + "angle": 0, + "content": "recognition." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.116, + 0.892, + 0.161 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that SCLRE outperforms the SOTA classification techniques on one artificial superclass dataset and three real datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.895, + 0.259 + ], + "angle": 0, + "content": "The remainder of the paper is organized as follows: in Sec. 2 we briefly review related work and in Sec. 3 we describe our method for superclass recognition. Then, extensive experiments and generalization error analysis are conducted in Secs. 4 and 5. Finally, Sec. 6 draws a brief conclusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.272, + 0.642, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Contrastive Learning. The contrastive framework has achieved great success on instance-level problems. Chen et al [3] designed a simple framework to get visual representations in a contrastive way. He et al [14] developed a dynamic dictionary with queues and moving averages to improve contrastive encoder quality. To learn basic-level features and discover more semantic structures for the embedding space, Li et al [22] bridge contrastive learning with clustering. More recently, Caron et al [2] presented a contrastive framework without the need for pairwise comparison computation. To get rid of the limitation of negative samples, Grill et al [12] use two neural networks and predict the representation of the same image under a different augmented view. Chen et al [4] further designed a stop-gradient operation to prevent collapsing. Contrastive learning has also been applied to various research domains and evolved quickly [11, 33, 34]. Instead of focusing on the basic-level boundary, we construct a high-level superclass-aware boundary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Supervised Contrastive Loss. Traditional contrastive frameworks encode data either unsupervised or self-supervised and apply supervised information in downstream tasks. Khosla et al [19] extend the contrastive mode to a fully-supervised setting and use label information more effectively. To address the impact of long-tailed distribution data, Li et al [23] propose a targeted contrastive loss and guide the encoder to obtain representations based on several preset anchors. Recently, supervised contrastive learning is used in broader research domains both in computer vision and natural language processing [16,25, 35]. We are inspired by the function of pulling close or away features in these supervised contrastive losses, and therefore propose a new loss, which will pull the features by the superclass labels for better performance in classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Self-Attention Module. Self-attention module has been widely used in deep learning. It was first mentioned by Vaswani et al [29] as a part of the transformer model and is widely used in the natural language processing field [5, 21, 24, 26]. Recently, the attention module and transformer also obtained great success in the computer vision" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24061" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.457, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Illustration of SCLRE. The images first generate their representations through a convolutional neural network, then mix with each other in a trainable cross-instance attention module for enhancement. After enhancement, the representations are then adjusted according to their superclass labels and the target anchors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.473, + 0.645 + ], + "angle": 0, + "content": "field. Dosovitskiy et al [6] organize image patches as a sequence and propose a vision transformer. Furthermore, He et al [13] propose a masked auto-encoder based on vision transformers. Self-attention modules can also be integrated into convolutional neural networks to improve performance [1, 18, 31]. Fu et al [10] address the fine-grained image recognition problem by using an attention network to find the unique and pivotal features in samples (for example, one kind of bird's tail color). Different from focusing the attention relationship in one sample, we propose a cross-instance attention(CIA) module that cares more about the relationship between instances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.169, + 0.696 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Superclass Learning with Representation Enhancement(SCLRE) framework is proposed for superclass image recognition. It exploits the self-attention technique across instances to perform a representation enhancement, thus it achieves the goal to break the basic-level boundary in representations. Then SCLRE reconstructs a new superclass-aware boundary in the enhanced representation space by a series of adjustment losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Fig. 2 shows the overview of SCLRE. Sec. 3.1 presents the process of breaking basic-level boundary, while Sec. 3.2 describes the details of constructing the new high-level superclass-aware boundary." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.895, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.183, + 0.894, + 0.225 + ], + "angle": 0, + "content": "Figure 3. Different segmentation strategies of cross-instance attention module. We experiment on three kinds of segmentation strategies and choose the best for the experiment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.24, + 0.766, + 0.258 + ], + "angle": 0, + "content": "3.1. Representation Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.367 + ], + "angle": 0, + "content": "To break the basic-level boundary of representations and obtain valuable enhancements, we mix the representations by leveraging a cross-instance attention module. Unlike the existing attention techniques [32], the cross-instance attention module cares more about the relationship between representations of the instances, rather than the feature representation inside the instance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.893, + 0.506 + ], + "angle": 0, + "content": "Enhanced Representation. Let \\(\\mathcal{X}\\) be the data space, \\(\\mathcal{R}^D\\) be the \\(D\\)-dimensional embedding space, an encoder \\(f: \\mathcal{X} \\to \\mathcal{R}^D\\) be a mapping from the data space to the embedding space, and an enhancement process \\(\\mathrm{EnH}: \\mathcal{R}^D \\to \\mathcal{R}^D\\) be a mapping from the embedding space to the embedding space. For each \\(x_j \\in \\mathcal{X}\\), the representation \\(z_j = f(x_j)\\), and \\(Z \\in \\mathcal{R}^{n \\times D}\\) is the representation matrix for the batch with size \\(n\\). The enhanced representation \\(v_j\\) is an output of \\(\\mathrm{EnH}\\). More specifically, \\(v_j\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.52, + 0.892, + 0.537 + ], + "angle": 0, + "content": "\\[\nv _ {j} = \\operatorname {E n H} \\left(z _ {j}; Z\\right) = a _ {j} \\times Z, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.618 + ], + "angle": 0, + "content": "where \\(a_{j}\\) is a trainable vector for representation \\(z_{j}\\) computed by cross-instance attention module. Actually, the vector \\(a_{j}\\) is the \\(j\\)-th row of the attention distribution matrix(ADM) computed in the following Q-K-V mode multi-head attention module [29]:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.628, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A D M} (Z) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.674, + 0.891, + 0.704 + ], + "angle": 0, + "content": "The matrix \\(Q = Z \\times W_{Q}\\), where \\(W_{Q} \\in \\mathcal{R}^{D \\times d_{q}}\\). Matrix \\(K = Z \\times W_{K}\\), where \\(W_{K} \\in \\mathcal{R}^{D \\times d_{k}}\\) and \\(d_{q} = d_{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The enhancement process breaks the basic-level boundary by merging the semantic features of instances. Cross-instance attention module automatically explores the key instances for representation enhancement and integrates them to obtain a superclass-aware enhanced representation. Moreover, we conduct strict theoretical analysis on the generalization ability of our model in Sec. 5, and discover that the cross-instance module can tightly bound the generalization error of SCLRE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Data Segmentation Strategy. In the cross-instance attention module, the segmentation strategy of input data is important for representation enhancement. For each instance, the segmentation strategy determines whose repre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24062" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.274 + ], + "angle": 0, + "content": "sentations will be received as an enhancement. We considered three kinds of segmentation strategies for the input data in the cross-instance attention module. Fig. 3 illustrates the details of the strategies. In (a), images are gathered in terms of their superclass label, we put the same label images into one batch. In (b), we randomly and evenly select images from the different superclasses and gather them into the batch. In (c), we keep the distribution of the dataset and randomly select the samples to form the batch. The result shows that (c) selecting the samples according to the distribution of the dataset benefits the model most, so we fixed this strategy in the following part of this work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.28, + 0.216, + 0.295 + ], + "angle": 0, + "content": "3.2. SCLRE Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.422 + ], + "angle": 0, + "content": "To reconstruct the high-level superclass-aware boundary of representations, we design a SCLRE loss to adjust the representations after the attention-weighted enhancement. We keep using the letter \\( v \\) in Eq. (1) to demonstrate representations being enhanced by the cross-instance attention module, and we call it enhanced representation. After that, we adjust the representations by taking three aspects into consideration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.47 + ], + "angle": 0, + "content": "Category Classification Loss. We first adopt the cross-entropy loss as the base to obtain a category-level adjustment and prevent the model from collapsing:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.476, + 0.47, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i, k} \\log M (v _ {i}) _ {k}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.47, + 0.582 + ], + "angle": 0, + "content": "where \\(N\\) demonstrates batch size, \\(y_{i}\\) demonstrates the superclass label of the \\(i\\)-th sample in one-hot embedding, \\(M(v_{i})\\) demonstrates the output prediction of \\(v_{i}\\), and the footprint \\(k\\) demonstrates the \\(k\\)-th component of a vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Contrastive Adjustment Loss. Besides classification loss, the samples belonging to the same superclass may share few common features. To obtain a representation-level adjustment, we design a contrastive adjustment loss to pull close the samples in the same superclass and pull away samples in distinct subclasses. We directly regard samples from the same superclass as positive pairs and samples from other subclasses as negative pairs. For one sample \\( v_{i} \\) and one of its positive sample \\( v^{+} \\) (i.e. \\( v_{i} \\) and \\( v^{+} \\) belongs to the same superclass), the contrastive adjustment loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.74, + 0.468, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\ell \\left(v _ {i}, v ^ {+}\\right) = - \\log \\frac {\\exp \\left(s \\left(v _ {i} , v ^ {+}\\right) / \\tau\\right)}{\\exp \\left(s \\left(v _ {i} , v ^ {+}\\right) / \\tau\\right) + \\sum_ {v ^ {-} \\notin P \\left(v _ {i}\\right)} \\exp \\left(s \\left(v _ {i} , v ^ {-}\\right) / \\tau\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.785, + 0.47, + 0.814 + ], + "angle": 0, + "content": "By summing up all the losses of positive pairs and for a batch with size \\(N\\), we can get the total loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.821, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c a} = \\sum_ {i = 1} ^ {N} \\frac {1}{| | P (v _ {i}) | | - 1} \\sum_ {v ^ {+} \\in P (v _ {i}) \\backslash \\{v _ {i} \\}} \\ell \\left(v _ {i}, v ^ {+}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( P(v_{i}) \\) stands for all the other data that share the same superclass label with \\( v_{i} \\) in the batch, i.e. \\( P(v_{i}) = \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.173 + ], + "angle": 0, + "content": "\\(\\{v_{k}|p(v_{k}) = p(v_{i})\\}\\) \\(p\\) means the superclass label. \\(s(\\cdot ,\\cdot)\\) stands for the similarity between two vectors, we use cosine similarity for measurement, i.e. \\(s(v_{s},v_{t}) = \\frac{v_{s}^{T}v_{t}}{\\|v_{s}\\|\\cdot\\|v_{t}\\|}\\) \\(\\tau\\) is the temperature hyperparameter. \\(||P(v_i)||\\) stands for the size of set \\(P(v_{i})\\) , and it aims to normalize the total loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.175, + 0.893, + 0.34 + ], + "angle": 0, + "content": "Targeted Adjustment Loss. Since superclass learning naturally faces the problem of lack of classification centers due to the scattered distribution of representations, then follows damage to the constructing process when the contrastive adjustment loss works. To address this problem, we design a targeted adjustment loss to give each superclass a pre-defined category center (i.e. target anchors). We first generate the target anchors and make them far away from each other inspired by [23]. Then we pull every sample close to its pre-defined target anchor by minimizing the following targeted adjustment loss:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.355, + 0.892, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t a} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (s \\left(v _ {i} , t \\left(v _ {i}\\right)\\right) / \\tau)}{\\sum_ {t \\in T} \\exp (s \\left(v _ {i} , t\\right) / \\tau)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.457 + ], + "angle": 0, + "content": "where \\(N\\) stands for batch size, \\(s(\\cdot ,\\cdot)\\) stands for cosine similarity function, \\(T\\) is the set of all the target anchors, and \\(t(v_{i})\\) demonstrates the target anchor allocated to \\(v_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.893, + 0.549 + ], + "angle": 0, + "content": "SCLRE Adjustment Loss. To construct the high-level superclass-aware boundary with both category-level and instance-level adjustments and overcome the problem of lack of classification center, the total loss of SCLRE is a weighted summation of the above three losses(i.e. Eqs. (3), (5) and (6)):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.565, + 0.892, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S C L R E} = (1 - \\alpha) \\mathcal {L} _ {c e} + \\alpha \\mathcal {L} _ {c a} + \\lambda \\mathcal {L} _ {t a}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.702 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\lambda\\) range from 0 to 1. \\(\\mathcal{L}_{ce}\\) and \\(\\mathcal{L}_{ca}\\) are designed as a combination to help the encoder explore the concept of superclasses, a larger \\(\\alpha\\) means a stronger adjustment strategy and weaker classification ability. \\(\\mathcal{L}_{ta}\\) is designed to help the model overcome the imbalanced distribution, and a larger \\(\\lambda\\) means stronger guidance towards the destination of adjustment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.81 + ], + "angle": 0, + "content": "SCLRE adjustment loss can help the model learn more about the concept of the superclass and the high-level categorization boundary. By balancing the influence of cross-entropy loss and contrastive adjustment loss, the model can adjust representations properly. By controlling the utilization of targeted adjustment loss, the model can run smoothly and converge steadily." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.625, + 0.846 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments on several artificially constructed superclass datasets to demonstrate the effectiveness of our proposed approach, SCLRE." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24063" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.473, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.471, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Illustration of superclass construction. We integrate the initial class labels of the original dataset to superclass labels, according to field knowledge. The constructed dataset contains a handful of subclasses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.363, + 0.254, + 0.379 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Datasets. In this study, superclass datasets are artificially reorganized from three benchmarks, CIFAR-100, mini-ImageNet [30] and VOC [8], and two real-world datasets, FMoW [20] and Adience [7]. CIFAR-100 contains 50000 training images and 10000 test images from 100 categories with a size of \\(32 \\times 32 \\times 3\\). mini-ImageNet [30] contains 60000 images in total with a size of \\(84 \\times 84 \\times 3\\), which are evenly distributed among 100 classes. VOC [8] contains 2501 training images, 2510 validation images and 4952 test images from 20 categories. The real-world dataset FMoW [20] is about a hybrid domain generalization and subpopulation shift problem, which consists of 76863 training images, 11483 validation images and 11327 test images with a size of \\(224 \\times 224 \\times 3\\). And the real-world dataset Adience [7] contains 26,580 images from 2,284 people." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.856 + ], + "angle": 0, + "content": "In our experiments, superclass datasets are artificially constructed by integrating the initial classes, just as illustrated in Fig. 4. To be concrete, We reorganize the 100 classes of CIFAR-100 into 3 subclasses, 4 subclasses, and 7 subclasses, respectively, forming three distinct superclass datasets, CIFAR100-3, CIFAR100-4, and CIFAR100-7. The subclasses in the three data sets are based on movement mode, life form, and are more complex than the original 20 coarse classes in CIFAR-100. For miniImagenet and VOC, we reorganize their subclasses into 2 subclasses based on life form. For FMoW, we drop the original labels and turn to predict the geographical location of the images. For Adience, we utilize the age groups as the superclass labels. Thus, those 7 superclass datasets can well simulate the distribution of real-world superclass problems. More reorganization details are listed in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Compared Methods. We compare SCLRE with the baseline model, i.e., ResNet [15], and two SOTA contrastive techniques SupCon [19], and SimCLR [3]. As contrastive" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.088, + 0.86, + 0.291 + ], + "angle": 0, + "content": "
DatasetMethodAccuracy(%)
CIFAR100-3Baseline72.8
SupCon [19]78.1
SimCLR [3]79.0
SCLRE80.1
CIFAR100-4Baseline76.0
SupCon [19]80.1
SimCLR [3]80.6
SCLRE84.0
CIFAR100-7Baseline68.9
SupCon [19]72.7
SimCLR [3]73.9
SCLRE78.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Table 1. Classification accuracy on low pixels dataset: CIFAR-100. We compared the classification accuracy on CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.891, + 0.397 + ], + "angle": 0, + "content": "techniques perform excellently in representation learning. We adopt the default optimal settings in the training details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.442 + ], + "angle": 0, + "content": "SupCon: In our experiments, SupCon adopts ResNet-50 as backbone, and SupCon loss as the super-classification loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.533 + ], + "angle": 0, + "content": "SimCLR: It is a known self-supervised contrastive learning framework. In superclass recognition, we keep using ResNet-50 as the backbone and use a supervised downstream task to keep the instance-level features pure, wherein the upstream instance-level results are a kind of learning of the basic level class labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Implementation details. We adopt ResNet-50 [15] as our default backbone and we reduce the feature dimension from 2048 to 128 with an extra Multi-Layer Perceptron(MLP) projection head. We train the network using the SGD with a momentum of 0.9 and a batch size of 64. We set the learning rate as 0.001 and the training stage as 200 epochs. Additionally, we resize all the images to \\(32 \\times 32\\) pixels. For our SCLRE method, we set the key \\(\\mathrm{K} = 256\\), value \\(\\mathrm{V} = 128\\) and 8 multi-heads are adopted in the attention module." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.727, + 0.71 + ], + "angle": 0, + "content": "4.2. Classification Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Accuracy on CIFAR-100 Datasets. Tab. 1 shows the classification accuracies of SCLRE and the compared methods on CIFAR100-3, CIFAR100-4, and CIFAR100-7, respectively. And we have the following observations: 1) It is observed that SCLRE has a great improvement compared with the baseline. This shows that the cross-instance attention module can effectively activate the contrastive adjustment loss. 2) SCLRE has higher classification accuracies than SimCLR on superclass datasets. This shows that the instance-level semantic information learned by SimCLR may fail in the superclass problem because the images from the same superclass may have quite different, even con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24064" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.143, + 0.164, + 0.156 + ], + "angle": 0, + "content": "CIFAR100-7" + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.101, + 0.31, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.102, + 0.451, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.102, + 0.596, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.102, + 0.736, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.102, + 0.88, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.242, + 0.162, + 0.254 + ], + "angle": 0, + "content": "CIFAR100-4" + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.202, + 0.308, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.202, + 0.451, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.202, + 0.596, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.202, + 0.735, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.203, + 0.879, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.338, + 0.162, + 0.349 + ], + "angle": 0, + "content": "CIFAR100-3" + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.302, + 0.308, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.302, + 0.451, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.302, + 0.596, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.303, + 0.736, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.303, + 0.879, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.414, + 0.277, + 0.427 + ], + "angle": 0, + "content": "(a) Baseline" + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.414, + 0.421, + 0.427 + ], + "angle": 0, + "content": "(b) SimCLR" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.414, + 0.562, + 0.427 + ], + "angle": 0, + "content": "(c) SupCon" + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.414, + 0.739, + 0.427 + ], + "angle": 0, + "content": "(d) SCLRE w\\o targets" + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.414, + 0.843, + 0.427 + ], + "angle": 0, + "content": "(e) SCLRE" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.894, + 0.499 + ], + "angle": 0, + "content": "Figure 5. Visualization results by t-SNE. We sampled part of the test set and encode them into representations, then visualize them in the embedding space by t-SNE. (a) We directly generate the representations using the backbone as a baseline. (b) We use the upstream encoder of SimCLR. (c) We use the encoder of SupCon. (d) We use part of SCLRE by removing the targeted adjustment loss. (e) We use the entire SCLRE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.558 + ], + "angle": 0, + "content": "tradictory semantic information. Comparatively, SCLRE works well by the superclass-aware decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Accuracy on High-Resolution Datasets. Tab. 2 shows the evaluation results on three high-resolution datasets, from which we observe the following facts. 1) we observe that, on the real-world dataset FMoW, SCLRE outperforms other methods by a large margin. On FMoW, the geographic location of building images is adopted as the superclass label, rather than their type. The methods without representation enhancement may perform worse due to the lack of superclass-aware common features. This further confirms the importance of representation enhancement in the superclass problems. 2) SCLRE also outperforms other methods on real-world dataset Adience, which contains 26580 face images from 2284 people. In case every single person is a raw class, the results confirm that SCLRE is also workable in large-scale situations. 3) Moreover, we observe that on FMoW and Adience, SimCLR performs even worse than the baseline. Because the instance-level representation by SimCLR is conflicted with the given superclass labels. 4) The performance on mini-[ImageNet appears to be less attractive. This may be because the mini-[ImageNet dataset is divided into 2 distinct subclasses: animals and non-animals. For such kind of distribution, basic-level methods" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.524, + 0.89, + 0.584 + ], + "angle": 0, + "content": "
mini-ImageNet [30]FMoW [20]Adience [7]VOC [8]
Baseline87.756.565.778.3
SupCon87.058.366.278.2
SimCLR90.151.552.980.0
SCLRE89.364.868.781.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Table 2. Classification accuracy on high-resolution datasets. We compared the accuracy on datasets with more complex and informative content." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.892, + 0.73 + ], + "angle": 0, + "content": "can also easily sort them out by exploring the obvious common features inside each superclass(e.g. whether the image contains eyes), which makes it more like a traditional instance-level problem, and SimCLR may perform comparably or even slightly better than SCLRE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.638, + 0.755 + ], + "angle": 0, + "content": "4.3. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Here we generate the visualization results on representation space by the t-SNE technique [28]. The CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets are adopted for the visualization experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.9 + ], + "angle": 0, + "content": "We observe the following facts from Fig. 5. 1) It is observed that in columns (a) and (b) the images are scattered in their respective representation space, although the images (b) cluster slightly closer than the ones in (a). This further shows that the instance-level representation of SimCLR is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "little related to superclass labels. Consequently, in superclass scenarios, SimCLR performs a little better than baseline but worse than SCLRE, just as illustrated in Tab. 1. 2) We observed that the images in column (c) have a more obvious cluster structure than in columns (a) and (b). This shows that the use of superclass labels is important for constructing clear boundaries. But due to the diverse and complex superclass contents, the representations are mostly inaccurate and thus result in less favorable results as reported in Tab. 1. 3) It is also observed that the images in columns (d) and (e) are notably closer than the ones in (a), (b), and (c). This shows that the enhanced representation of SCLRE is superclass-aware. 4) Furtherly, we observe that the images in column (e) are obviously closer than the ones in column (d). This shows that the targeted adjustment loss, one important component of SCLRE, can furtherly cluster images in the same subclasses, accordingly, it benefits reconstructing a superclass-aware decision boundary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.372, + 0.265, + 0.387 + ], + "angle": 0, + "content": "4.4. Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.486 + ], + "angle": 0, + "content": "We discuss the hyperparameters \\(\\alpha\\) and \\(\\lambda\\) of SCLRE by evaluating their sensitivity on the CIFAR100-7 dataset. When analyzing \\(\\alpha\\), to avoid the influence of target anchors, we removed targeted adjustment loss by fixing \\(\\lambda\\) to 0. We test the sensitivity of \\(\\alpha\\) by ranging it from 0 to 1. When analyzing \\(\\lambda\\), we fix \\(\\alpha\\) to 0.5 to avoid its influence." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.496, + 0.451, + 0.531 + ], + "angle": 0, + "content": "
α00.10.30.50.71
Accuracy69.272.173.775.173.6-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.598 + ], + "angle": 0, + "content": "Table 3. Sensitivity analysis of \\(\\alpha\\). We range \\(\\alpha\\) from 0 to 1 and calculate the accuracy of SCLRE respectively. When \\(\\alpha = 1\\), the cross-entropy loss is invalid and the model will lose its classification ability." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.613, + 0.462, + 0.648 + ], + "angle": 0, + "content": "
λ00.10.20.40.81
Accuracy75.176.777.976.978.176.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Table 4. Sensitivity analysis of \\( \\lambda \\). We range \\( \\lambda \\) from 0 to 1 and calculate the accuracy of SCLRE respectively." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.703, + 0.418, + 0.737 + ], + "angle": 0, + "content": "
Batch Size64128256512
Accuracy76.776.977.377.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.776 + ], + "angle": 0, + "content": "Table 5. Sensitivity analysis of batch size. We vary batch size from 64 to 512 and calculate the accuracy of SCLRE respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Sensitivity of \\(\\alpha\\). Tab. 3 shows the sensitivity of \\(\\alpha\\). Besides extreme situations, we find that \\(\\alpha\\) has a little influence on the accuracy. The optimal value of \\(\\alpha\\) falls around 0.5 and we fix it in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": "In the case of \\(\\alpha = 0\\), the contrastive adjustment loss is invalid and the model degenerates into the baseline. In the case of \\(\\alpha = 1\\), the cross-entropy loss is invalid and then the model is not workable for classification." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.088, + 0.872, + 0.201 + ], + "angle": 0, + "content": "
ArchitectureParams.(M)FLOPs(G)Acc.(%)
Baseline23.511.3068.9
+CIA37.211.3169.2
+Lca--72.7
+CIA, Lca--75.1
SCLRE (+CIA, Lca, Lta)37.211.3277.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.252 + ], + "angle": 0, + "content": "Table 6. Ablation study on SCLRE. We calculated the model size (Params.), computational complexity (FLOPs) and accuracy contributions of four components of SCLRE." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.322 + ], + "angle": 0, + "content": "Sensitivity of \\(\\lambda\\). Tab. 4 demonstrates the sensitivity analysis of \\(\\lambda\\). \\(\\lambda\\) is a hyperparameter that controls the influence of targeted adjustment loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.326, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Unlike \\(\\alpha\\), there are two most proper \\(\\lambda\\) values which fall around 0.2 and 0.8. When \\(\\lambda\\) is too small or too large, the performance is both less satisfactory. As a too large \\(\\lambda\\) may be harmful to the accuracy of the model, a too small \\(\\lambda\\) may be short of guidance and has limited performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.421, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Sensitivity of Batch Size. Intuitively, a larger batch size can bring more abundant samples of superclasses, thus making it easier to construct superclass-aware representations. The experimental results are conducted to verify this statement. Tab. 5 illustrated the results of the sensitivity study of the batch size. The results show that a larger batch size can improve the performance of SCLRE, though the improvement is not significant." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.654, + 0.58 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.591, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Here ablation studies are conducted to demonstrate the contribution of each component in our proposed SCLRE. Just as designed in Sec. 3, SCLRE is composed of the baseline model, contrastive adjustment loss \\((\\mathcal{L}_{ca})\\), cross-instance attention module(CIA), and targeted adjustment loss \\((\\mathcal{L}_{ta})\\). The baseline model is the same as our SCLRE without \\(\\mathcal{L}_{ca}\\), \\(\\mathcal{L}_{ta}\\) and CIA." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.701, + 0.892, + 0.731 + ], + "angle": 0, + "content": "According to the results reported in Tab. 6, we have the following observations:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1) It is observed that only CIA itself has an observable contribution to the performance improvements, but CIA joint with \\(\\mathcal{L}_{ca}\\) brings notable improvements. This indicates that the CIA module is workable to break the basic class boundary with the great assistance of \\(\\mathcal{L}_{ca}\\). They enhance the representation of the superclass by cross-instance attention mechanism. 2) With CIA, \\(\\mathcal{L}_{ca}\\) and \\(\\mathcal{L}_{ta}\\), SCLRE achieves the optimal improvements. It shows that the targeted adjustment loss does function on reconstructing superclass-aware boundary on the enhanced representation space." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24066" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.193, + 0.088, + 0.782, + 0.155 + ], + "angle": 0, + "content": "
ResNet-18ResNet-34ResNet-50ResNet-101ResNet-152
Baseline83.983.778.283.482.6
SCLRE86.887.682.987.787.2
Improve.(%)+2.9+2.9+4.7+4.3+4.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.164, + 0.895, + 0.194 + ], + "angle": 0, + "content": "Table 7. Robustness research of SCLRE on different backbones. We calculated the accuracy on a CIFAR100-3 superclass dataset of baseline and SCLRE. We also calculated the improvement SCLRE gets in the different backbones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.214, + 0.391, + 0.23 + ], + "angle": 0, + "content": "4.6. Robustness for Different Backbones" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.471, + 0.345 + ], + "angle": 0, + "content": "We investigate the robustness of SCLRE by changing the backbone to other convolutional neural networks. Tab. 7 shows the result of the experiments on robustness. The experiments are performed on a CIFAR100-3 superclass dataset and we vary the backbone from smaller networks to larger networks. The result shows that SCLRE can make stable improvements on different backbones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.386, + 0.373 + ], + "angle": 0, + "content": "5. Analysis of Generalization Ability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.471, + 0.44 + ], + "angle": 0, + "content": "We conduct an analysis of the generalization ability of SCLRE and prove that the cross-instance attention module can further compress the upper bound of the generalization error. The details of the proof are in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.486 + ], + "angle": 0, + "content": "Based on the contrastive adjustment loss in Sec. 3.2, and with ignoring the constant items, we can rewrite the contrastive loss into an expectation form:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.49, + 0.457, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal{L}_{ca}\\propto -\\mathbb{E}_{v,v^{\\prime}}\\mathbb{E}_{\\substack{v_{1},v_{2}\\in P(v)\\\\ v^{-}\\in P(v^{\\prime})}}\\log \\frac{\\exp(v_{1}^{T}v_{2})}{\\exp(v_{1}^{T}v_{2}) + \\exp(v_{1}^{T}v^{-})} \\\\ = \\underbrace {- \\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2}} _ {L _ {1}} \\\\ + \\underbrace {\\mathbb {E} _ {v , v ^ {\\prime}} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} \\log \\left(\\exp (v _ {1} ^ {T} v _ {2}) + \\exp (v _ {1} ^ {T} v ^ {-})\\right)} _ {L _ {2}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.661 + ], + "angle": 0, + "content": "where \\( L_{1} \\) measures the alignment between two transformed features and \\( L_{2} \\) is the regularizer preventing the collapse of representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Lemma 1. As a conclusion in [17], if \\( L_{2} \\) is trained to be small than the threshold, which is easily satisfied, the generalization error of downstream classifier \\( G_{f} \\) has an upper bound:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.725, + 0.459, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E r r (G _ {f}) \\leqslant (1 - \\sigma) + \\eta (\\epsilon) \\sqrt {2 - 2 \\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2}} \\\\ = (1 - \\sigma) + \\eta (\\epsilon) \\sqrt {2 + 2 L _ {1}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Lemma 2. Sample pairs with the same coarse label consist of two parts: those share the same fine label \\((O_1)\\) and those have different fine labels \\((O_2)\\). The former can be closely aligned after contrastive learning. We focus on the latter and prove that:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.873, + 0.427, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1}, v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2} \\geqslant C _ {\\varphi} + \\frac {(1 - \\rho) K}{M _ {1} M _ {2}} s (a _ {1}, a _ {2})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.215, + 0.892, + 0.261 + ], + "angle": 0, + "content": "where \\( C_{\\varphi} \\) is a constant determined by the data distribution. \\( M_1, M_2, K \\) are constants. \\( \\rho \\) is the ratio of sample pairs with same fine label in all positive pairs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.894, + 0.309 + ], + "angle": 0, + "content": "Theorem 1. The generalization error has another upper bound based on attention vectors according to lemma 1 and lemma 2:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.326, + 0.868, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E r r (G _ {f}) \\leqslant (1 - \\sigma) + \\\\ \\sqrt {2} \\eta (\\epsilon) \\sqrt {1 - C _ {\\varphi} - \\frac {(1 - \\rho) K}{M _ {1} M _ {2}} \\mathbb {E} _ {v _ {1} , v _ {2} \\in O _ {2}} s (a _ {1} , a _ {2}))} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.516 + ], + "angle": 0, + "content": "The attention vector is a representation of the importance between samples. When we define a superclass, the representation vector \\( z \\) with the same coarse label will be pulled together by the contrastive loss. During this process, samples in the superclass will focus on the same important enhanced representations, thus leading to similar vectors and declining the upper bound of generalization error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.62, + 0.558 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.722 + ], + "angle": 0, + "content": "In this study, we explore the high-level categories recognition problem, i.e., superclass categorization. And a superclass learning framework, that exploits self-attention techniques cross the instances to enhance the representation, is proposed. Thus, the distribution of superclass is modified in the enhanced representation space and a new superclass-aware decision boundary is then reconstructed. In theory, the generalization error of SCLRE can be bounded by attention constraints. In extensive experiments, SCLRE outperforms the SOTA classification methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.772 + ], + "angle": 0, + "content": "In the near future, we would like to tackle the superclass problems under the scenarios without full annotation by leveraging semi-supervised learning and active learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Key Research & Develop Plan(2018YFB1004401); National Natural Science Foundation of China (62276270,62072460, 62172424); Beijing Natural Science Foundation (4212022). it is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24067" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Amjad Almahairi, Nicolas Ballas, Tim Coolijmans, Yin Zheng, H. Larochelle, and Aaron C. Courville. Dynamic capacity networks. In ICML, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.228 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. ArXiv, abs/2002.05709, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.528 + ], + "angle": 0, + "content": "[7] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Transactions on information forensics and security, 9(12):2170-2179, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.471, + 0.587 + ], + "angle": 0, + "content": "[8] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2009. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.471, + 0.656 + ], + "angle": 0, + "content": "[9] Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[10] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4476-4484, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Qiushi Huang, Tom Ko, H Lilian Tang, Xubo Liu, and Boyong Wu. Token-level supervised contrastive learning for punctuation restoration. In Interspeech, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[17] Weiran Huang, Mingyang Yi, and Xuyang Zhao. Towards the generalization of contrastive self-supervised learning. arXiv preprint arXiv:2111.00743, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In NIPS, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. ArXiv, abs/2004.11362, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton A. Earnshaw, Imran S. Haque, Sara Beery, Jure Leskovec, Anshul Kundaje, Emma Pierson, Sergey Levine, Chelsea Finn, and Percy Liang. WILDS: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning (ICML), 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Junnan Li, Pan Zhou, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Prototypical contrastive learning of unsupervised representations. ArXiv, abs/2005.04966, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogério Schmidt Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6908-6918, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[24] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Zheda Mai, Ruiwen Li, Hyunwoo J. Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3584-3594, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24068" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Lars Schmarje, Monty Santarossa, Simon-Martin Schroder, and Reinhard Koch. A survey on semi-, self- and unsupervised learning for image classification. IEEE Access, 9:82146-82168, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[30] Oriol Vinyals, Charles Blundell, Timothy P. Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In-So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[32] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[33] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. ArXiv, abs/2010.13902, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.501 + ], + "angle": 0, + "content": "[34] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-842, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[35] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10022-10031, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24069" + } + ] +] \ No newline at end of file diff --git a/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_origin.pdf b/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef339f3162d39140458b012798005c574b9b17a4 --- /dev/null +++ b/2023/Superclass Learning With Representation Enhancement/2c1b722e-766f-4d50-b4f7-13afac67bce5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:615760cc321a044eb8983fd30d576ba8660df852d7558b8239c57375d98f9004 +size 4849171 diff --git a/2023/Superclass Learning With Representation Enhancement/full.md b/2023/Superclass Learning With Representation Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..97766d769a1f86eb530ee8a1fee79de990292675 --- /dev/null +++ b/2023/Superclass Learning With Representation Enhancement/full.md @@ -0,0 +1,370 @@ +# Superclass Learning with Representation Enhancement + +Zeyu Gan $^{1,2}$ + +Suyun Zhao $^{1,2,*}$ + +Hong Chen $^{1,2}$ + +Jinlong Kang + +Liyuan Shang + +Cuiping $\mathbf{L}\mathbf{i}^{1,2}$ + +1Key Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China +2Renmin University of China, Beijing, China + +{zygan, zhaosuyun, kangjinlong, shangliyuan4032, chong, LICUping}@ruc.edu.cn + +# Abstract + +In many real scenarios, data are often divided into a handful of artificial super categories in terms of expert knowledge rather than the representations of images. Concretely, a superclass may contain massive and various raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classification techniques are intractable to recognize superclass without raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhancement(SCLRE), to recognize super categories by leveraging enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE collapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision boundary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Experimentally, extensive results demonstrate that SCLRE outperforms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets. + +# 1. Introduction + +In recent decades, basic-level raw categorization (e.g. cats vs dogs, apples vs bananas) has greatly developed [9, 27] while high-level or super-coarse-grained visual categorization (e.g., recyclable waste vs kitchen waste, creatures vs non-creatures) has received little attention. In many real scenarios, there often exist a handful of high-level categories, wherein numerous images from diverse basic-level categories share one common label. We tend to define this kind of super-coarse-grained class as Superclass. + +![](images/a24cc70987046ff79367574bad04da73da7a217a5ed361c57c8f9b6bc12b72d1.jpg) + +![](images/0f6e157739725d90d4c41884b9037e7881d6442447e6adc792b2b231c3273962.jpg) + +![](images/bdc5511a9e598233af0f47ca10706d2cd00ffe6cbb2a2a406e17ccb3a1050de6.jpg) +(a) Image Feature Domain + +![](images/18f9b99a87527fa81959da251f8889ee2b0573c7f12f19a752546e126de18b77.jpg) +(b) Separate Fine Domain + +![](images/d3dc8ab9df2cc866ad3e76b08bf80995eda9a018993739b909a7859875970070.jpg) +(c) Superclass Concept Domain +Figure 1. Illustration of superclass learning. The samples from a same superclass will be scatteredly distributed in the embedding space. The process of superclass learning is to break old domains and construct new domains. Red indicates the superclass of kitchen waste, and blue indicates the superclass of recyclable waste. + +Refuse sorting, as an example, is such a recognition problem with four superclasses, i.e., kitchen waste, recyclable waste, hazardous waste, and others waste. One task of refuse sorting is to accurately collect various items, such as rotten fruits, bones, raw vegetables, and eggshells, into kitchen waste. Traditional recognition needs to identify what exact basic-level categories they are, then sort them out. Obviously, it is wasteful and unrealistic for superclass identification. + +Essentially, high-level superclasses contain two characteristics, remarkably distinct from basic-level classes. First, the basic-level classes contained in superclass problems are usually scattered and share few common features. As depicted in the top-left corner subgraph of Fig. 1, the fruit apple, bone, and eggs are remote from each other in feature spaces, though all of them belong to kitchen waste. + +Second, the instances from two distinct superclasses may share common features. Just as illustrated in Fig. 1, fruit apple, from kitchen waste, and toy apple, from recyclable waste, are close to each other as they share common semantic features. Obviously, the above-mentioned characteristics indicate that the smoothness assumption [27] under basic-level classification (nearby images tend to have the same label) does not hold in the superclass scenarios. Thus, the existing classification techniques based on smoothness assumption are not practically deployable and scalable and they may suffer severe performance damage in superclass settings. Accordingly, it is valuable and promising to investigate superclass identification. + +To tackle the superclass problems, we have to address two main challenges. First, we need to break the original decision boundaries of basic-level classes and disclose a superclass-aware boundary at the basic class level. As depicted in bottom subgraph (a) of Fig. 1, the boundary of the apple domain is original, however, it is useless and even harmful for the refuse sorting as both fruit apples and toy apples belong to this domain. To get the required domain boundary, the apple domain needs to be separated into the fruit domain and toy domain by leveraging their individual local features, as depicted in the bottom subgraph (b) of Fig. 1. In superclass scenarios, it is not enough to investigate the boundary at a basic class level. Consequently, the second challenge is to reconstruct a decision boundary at a superclass level. To achieve this end, it is necessary to merge the domain of classes, such as fruit apples, eggs, and bones into a new rotten superclass domain, as depicted in the bottom subgraph (c) of Fig. 1. + +In this paper, we propose a SuperClass Learning framework with Representation Enhancement. Considering that the semantic representation at the basic class level is not workable for superclass recognition, we propose one cross-instance attention module which could seize the representation across the instances with the same superclass label. By leveraging contrastive adjustment loss, the attention mechanism enhances this representation. Moreover, to overcome the imbalance distribution of superclasses, we adopt target adjustment loss to reconstruct a superclass-aware decision boundary on the enhanced representation space. + +In summary, this paper makes the following contributions: + +- We propose an under-study but realistic problem, superclass identification, that has notably distinct distribution from basic-level categorization. +- We propose a novel representation enhancement method by leveraging cross-instance attention and then exploit it in superclass identification. And by theoretical analyses, we verify that this self-attention technique can bound the generalization error of superclass + +recognition. + +- Extensive experiments demonstrate that SCLRE outperforms the SOTA classification techniques on one artificial superclass dataset and three real datasets. + +The remainder of the paper is organized as follows: in Sec. 2 we briefly review related work and in Sec. 3 we describe our method for superclass recognition. Then, extensive experiments and generalization error analysis are conducted in Secs. 4 and 5. Finally, Sec. 6 draws a brief conclusion. + +# 2. Related Work + +Contrastive Learning. The contrastive framework has achieved great success on instance-level problems. Chen et al [3] designed a simple framework to get visual representations in a contrastive way. He et al [14] developed a dynamic dictionary with queues and moving averages to improve contrastive encoder quality. To learn basic-level features and discover more semantic structures for the embedding space, Li et al [22] bridge contrastive learning with clustering. More recently, Caron et al [2] presented a contrastive framework without the need for pairwise comparison computation. To get rid of the limitation of negative samples, Grill et al [12] use two neural networks and predict the representation of the same image under a different augmented view. Chen et al [4] further designed a stop-gradient operation to prevent collapsing. Contrastive learning has also been applied to various research domains and evolved quickly [11, 33, 34]. Instead of focusing on the basic-level boundary, we construct a high-level superclass-aware boundary. + +Supervised Contrastive Loss. Traditional contrastive frameworks encode data either unsupervised or self-supervised and apply supervised information in downstream tasks. Khosla et al [19] extend the contrastive mode to a fully-supervised setting and use label information more effectively. To address the impact of long-tailed distribution data, Li et al [23] propose a targeted contrastive loss and guide the encoder to obtain representations based on several preset anchors. Recently, supervised contrastive learning is used in broader research domains both in computer vision and natural language processing [16,25, 35]. We are inspired by the function of pulling close or away features in these supervised contrastive losses, and therefore propose a new loss, which will pull the features by the superclass labels for better performance in classification. + +Self-Attention Module. Self-attention module has been widely used in deep learning. It was first mentioned by Vaswani et al [29] as a part of the transformer model and is widely used in the natural language processing field [5, 21, 24, 26]. Recently, the attention module and transformer also obtained great success in the computer vision + +![](images/81486a35630af085eae0ba8a9fbd1854d1bb712364438f0ccbb74a45bd629405.jpg) +Figure 2. Illustration of SCLRE. The images first generate their representations through a convolutional neural network, then mix with each other in a trainable cross-instance attention module for enhancement. After enhancement, the representations are then adjusted according to their superclass labels and the target anchors. + +field. Dosovitskiy et al [6] organize image patches as a sequence and propose a vision transformer. Furthermore, He et al [13] propose a masked auto-encoder based on vision transformers. Self-attention modules can also be integrated into convolutional neural networks to improve performance [1, 18, 31]. Fu et al [10] address the fine-grained image recognition problem by using an attention network to find the unique and pivotal features in samples (for example, one kind of bird's tail color). Different from focusing the attention relationship in one sample, we propose a cross-instance attention(CIA) module that cares more about the relationship between instances. + +# 3. Method + +Superclass Learning with Representation Enhancement(SCLRE) framework is proposed for superclass image recognition. It exploits the self-attention technique across instances to perform a representation enhancement, thus it achieves the goal to break the basic-level boundary in representations. Then SCLRE reconstructs a new superclass-aware boundary in the enhanced representation space by a series of adjustment losses. + +Fig. 2 shows the overview of SCLRE. Sec. 3.1 presents the process of breaking basic-level boundary, while Sec. 3.2 describes the details of constructing the new high-level superclass-aware boundary. + +![](images/a4a9f5c0094da18574f9cf87e87c98adbfecaff13cb10171d816766958c13456.jpg) +Figure 3. Different segmentation strategies of cross-instance attention module. We experiment on three kinds of segmentation strategies and choose the best for the experiment. + +# 3.1. Representation Enhancement + +To break the basic-level boundary of representations and obtain valuable enhancements, we mix the representations by leveraging a cross-instance attention module. Unlike the existing attention techniques [32], the cross-instance attention module cares more about the relationship between representations of the instances, rather than the feature representation inside the instance. + +Enhanced Representation. Let $\mathcal{X}$ be the data space, $\mathcal{R}^D$ be the $D$ -dimensional embedding space, an encoder $f: \mathcal{X} \to \mathcal{R}^D$ be a mapping from the data space to the embedding space, and an enhancement process $\mathrm{EnH}: \mathcal{R}^D \to \mathcal{R}^D$ be a mapping from the embedding space to the embedding space. For each $x_j \in \mathcal{X}$ , the representation $z_j = f(x_j)$ , and $Z \in \mathcal{R}^{n \times D}$ is the representation matrix for the batch with size $n$ . The enhanced representation $v_j$ is an output of $\mathrm{EnH}$ . More specifically, $v_j$ is defined as follows: + +$$ +v _ {j} = \operatorname {E n H} \left(z _ {j}; Z\right) = a _ {j} \times Z, \tag {1} +$$ + +where $a_{j}$ is a trainable vector for representation $z_{j}$ computed by cross-instance attention module. Actually, the vector $a_{j}$ is the $j$ -th row of the attention distribution matrix(ADM) computed in the following Q-K-V mode multi-head attention module [29]: + +$$ +\operatorname {A D M} (Z) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right). \tag {2} +$$ + +The matrix $Q = Z \times W_{Q}$ , where $W_{Q} \in \mathcal{R}^{D \times d_{q}}$ . Matrix $K = Z \times W_{K}$ , where $W_{K} \in \mathcal{R}^{D \times d_{k}}$ and $d_{q} = d_{k}$ . + +The enhancement process breaks the basic-level boundary by merging the semantic features of instances. Cross-instance attention module automatically explores the key instances for representation enhancement and integrates them to obtain a superclass-aware enhanced representation. Moreover, we conduct strict theoretical analysis on the generalization ability of our model in Sec. 5, and discover that the cross-instance module can tightly bound the generalization error of SCLRE. + +Data Segmentation Strategy. In the cross-instance attention module, the segmentation strategy of input data is important for representation enhancement. For each instance, the segmentation strategy determines whose repre + +sentations will be received as an enhancement. We considered three kinds of segmentation strategies for the input data in the cross-instance attention module. Fig. 3 illustrates the details of the strategies. In (a), images are gathered in terms of their superclass label, we put the same label images into one batch. In (b), we randomly and evenly select images from the different superclasses and gather them into the batch. In (c), we keep the distribution of the dataset and randomly select the samples to form the batch. The result shows that (c) selecting the samples according to the distribution of the dataset benefits the model most, so we fixed this strategy in the following part of this work. + +# 3.2. SCLRE Loss + +To reconstruct the high-level superclass-aware boundary of representations, we design a SCLRE loss to adjust the representations after the attention-weighted enhancement. We keep using the letter $v$ in Eq. (1) to demonstrate representations being enhanced by the cross-instance attention module, and we call it enhanced representation. After that, we adjust the representations by taking three aspects into consideration. + +Category Classification Loss. We first adopt the cross-entropy loss as the base to obtain a category-level adjustment and prevent the model from collapsing: + +$$ +\mathcal {L} _ {c e} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {k = 1} ^ {K} y _ {i, k} \log M (v _ {i}) _ {k}, \tag {3} +$$ + +where $N$ demonstrates batch size, $y_{i}$ demonstrates the superclass label of the $i$ -th sample in one-hot embedding, $M(v_{i})$ demonstrates the output prediction of $v_{i}$ , and the footprint $k$ demonstrates the $k$ -th component of a vector. + +Contrastive Adjustment Loss. Besides classification loss, the samples belonging to the same superclass may share few common features. To obtain a representation-level adjustment, we design a contrastive adjustment loss to pull close the samples in the same superclass and pull away samples in distinct subclasses. We directly regard samples from the same superclass as positive pairs and samples from other subclasses as negative pairs. For one sample $v_{i}$ and one of its positive sample $v^{+}$ (i.e. $v_{i}$ and $v^{+}$ belongs to the same superclass), the contrastive adjustment loss is: + +$$ +\ell \left(v _ {i}, v ^ {+}\right) = - \log \frac {\exp \left(s \left(v _ {i} , v ^ {+}\right) / \tau\right)}{\exp \left(s \left(v _ {i} , v ^ {+}\right) / \tau\right) + \sum_ {v ^ {-} \notin P \left(v _ {i}\right)} \exp \left(s \left(v _ {i} , v ^ {-}\right) / \tau\right)}. \tag {4} +$$ + +By summing up all the losses of positive pairs and for a batch with size $N$ , we can get the total loss as: + +$$ +\mathcal {L} _ {c a} = \sum_ {i = 1} ^ {N} \frac {1}{| | P (v _ {i}) | | - 1} \sum_ {v ^ {+} \in P (v _ {i}) \backslash \{v _ {i} \}} \ell \left(v _ {i}, v ^ {+}\right), \tag {5} +$$ + +where $P(v_{i})$ stands for all the other data that share the same superclass label with $v_{i}$ in the batch, i.e. $P(v_{i}) =$ + +$\{v_{k}|p(v_{k}) = p(v_{i})\}$ $p$ means the superclass label. $s(\cdot ,\cdot)$ stands for the similarity between two vectors, we use cosine similarity for measurement, i.e. $s(v_{s},v_{t}) = \frac{v_{s}^{T}v_{t}}{\|v_{s}\|\cdot\|v_{t}\|}$ $\tau$ is the temperature hyperparameter. $||P(v_i)||$ stands for the size of set $P(v_{i})$ , and it aims to normalize the total loss. + +Targeted Adjustment Loss. Since superclass learning naturally faces the problem of lack of classification centers due to the scattered distribution of representations, then follows damage to the constructing process when the contrastive adjustment loss works. To address this problem, we design a targeted adjustment loss to give each superclass a pre-defined category center (i.e. target anchors). We first generate the target anchors and make them far away from each other inspired by [23]. Then we pull every sample close to its pre-defined target anchor by minimizing the following targeted adjustment loss: + +$$ +\mathcal {L} _ {t a} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp (s \left(v _ {i} , t \left(v _ {i}\right)\right) / \tau)}{\sum_ {t \in T} \exp (s \left(v _ {i} , t\right) / \tau)}, \tag {6} +$$ + +where $N$ stands for batch size, $s(\cdot ,\cdot)$ stands for cosine similarity function, $T$ is the set of all the target anchors, and $t(v_{i})$ demonstrates the target anchor allocated to $v_{i}$ . + +SCLRE Adjustment Loss. To construct the high-level superclass-aware boundary with both category-level and instance-level adjustments and overcome the problem of lack of classification center, the total loss of SCLRE is a weighted summation of the above three losses(i.e. Eqs. (3), (5) and (6)): + +$$ +\mathcal {L} _ {S C L R E} = (1 - \alpha) \mathcal {L} _ {c e} + \alpha \mathcal {L} _ {c a} + \lambda \mathcal {L} _ {t a}, \tag {7} +$$ + +where $\alpha$ and $\lambda$ range from 0 to 1. $\mathcal{L}_{ce}$ and $\mathcal{L}_{ca}$ are designed as a combination to help the encoder explore the concept of superclasses, a larger $\alpha$ means a stronger adjustment strategy and weaker classification ability. $\mathcal{L}_{ta}$ is designed to help the model overcome the imbalanced distribution, and a larger $\lambda$ means stronger guidance towards the destination of adjustment. + +SCLRE adjustment loss can help the model learn more about the concept of the superclass and the high-level categorization boundary. By balancing the influence of cross-entropy loss and contrastive adjustment loss, the model can adjust representations properly. By controlling the utilization of targeted adjustment loss, the model can run smoothly and converge steadily. + +# 4. Experiment + +In this section, we conduct extensive experiments on several artificially constructed superclass datasets to demonstrate the effectiveness of our proposed approach, SCLRE. + +![](images/609df7a44d6c32fd6a63754200b04bdceb25bd2698b0a0861e9a4a1924098e1c.jpg) +Figure 4. Illustration of superclass construction. We integrate the initial class labels of the original dataset to superclass labels, according to field knowledge. The constructed dataset contains a handful of subclasses. + +# 4.1. Experiment Setup + +Datasets. In this study, superclass datasets are artificially reorganized from three benchmarks, CIFAR-100, mini-ImageNet [30] and VOC [8], and two real-world datasets, FMoW [20] and Adience [7]. CIFAR-100 contains 50000 training images and 10000 test images from 100 categories with a size of $32 \times 32 \times 3$ . mini-ImageNet [30] contains 60000 images in total with a size of $84 \times 84 \times 3$ , which are evenly distributed among 100 classes. VOC [8] contains 2501 training images, 2510 validation images and 4952 test images from 20 categories. The real-world dataset FMoW [20] is about a hybrid domain generalization and subpopulation shift problem, which consists of 76863 training images, 11483 validation images and 11327 test images with a size of $224 \times 224 \times 3$ . And the real-world dataset Adience [7] contains 26,580 images from 2,284 people. + +In our experiments, superclass datasets are artificially constructed by integrating the initial classes, just as illustrated in Fig. 4. To be concrete, We reorganize the 100 classes of CIFAR-100 into 3 subclasses, 4 subclasses, and 7 subclasses, respectively, forming three distinct superclass datasets, CIFAR100-3, CIFAR100-4, and CIFAR100-7. The subclasses in the three data sets are based on movement mode, life form, and are more complex than the original 20 coarse classes in CIFAR-100. For miniImagenet and VOC, we reorganize their subclasses into 2 subclasses based on life form. For FMoW, we drop the original labels and turn to predict the geographical location of the images. For Adience, we utilize the age groups as the superclass labels. Thus, those 7 superclass datasets can well simulate the distribution of real-world superclass problems. More reorganization details are listed in the Appendix. + +Compared Methods. We compare SCLRE with the baseline model, i.e., ResNet [15], and two SOTA contrastive techniques SupCon [19], and SimCLR [3]. As contrastive + +
DatasetMethodAccuracy(%)
CIFAR100-3Baseline72.8
SupCon [19]78.1
SimCLR [3]79.0
SCLRE80.1
CIFAR100-4Baseline76.0
SupCon [19]80.1
SimCLR [3]80.6
SCLRE84.0
CIFAR100-7Baseline68.9
SupCon [19]72.7
SimCLR [3]73.9
SCLRE78.1
+ +Table 1. Classification accuracy on low pixels dataset: CIFAR-100. We compared the classification accuracy on CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets. + +techniques perform excellently in representation learning. We adopt the default optimal settings in the training details. + +SupCon: In our experiments, SupCon adopts ResNet-50 as backbone, and SupCon loss as the super-classification loss. + +SimCLR: It is a known self-supervised contrastive learning framework. In superclass recognition, we keep using ResNet-50 as the backbone and use a supervised downstream task to keep the instance-level features pure, wherein the upstream instance-level results are a kind of learning of the basic level class labels. + +Implementation details. We adopt ResNet-50 [15] as our default backbone and we reduce the feature dimension from 2048 to 128 with an extra Multi-Layer Perceptron(MLP) projection head. We train the network using the SGD with a momentum of 0.9 and a batch size of 64. We set the learning rate as 0.001 and the training stage as 200 epochs. Additionally, we resize all the images to $32 \times 32$ pixels. For our SCLRE method, we set the key $\mathrm{K} = 256$ , value $\mathrm{V} = 128$ and 8 multi-heads are adopted in the attention module. + +# 4.2. Classification Evaluation + +Accuracy on CIFAR-100 Datasets. Tab. 1 shows the classification accuracies of SCLRE and the compared methods on CIFAR100-3, CIFAR100-4, and CIFAR100-7, respectively. And we have the following observations: 1) It is observed that SCLRE has a great improvement compared with the baseline. This shows that the cross-instance attention module can effectively activate the contrastive adjustment loss. 2) SCLRE has higher classification accuracies than SimCLR on superclass datasets. This shows that the instance-level semantic information learned by SimCLR may fail in the superclass problem because the images from the same superclass may have quite different, even con + +![](images/06f53139251811f5874f0e64e7e48052768f42b461ad91ea8db539a69374ee69.jpg) +CIFAR100-7 + +![](images/e0479c9dc9936b3fd319e48b6ee109247249ab34b61c6405043e8042fa442f74.jpg) + +![](images/b3f8ff954d68f6ffa179375ea2da723a0b11d3f966aee4c16c2f3d2838da5528.jpg) + +![](images/58df38e3ae129d775a41f5cdecb19c03590bfb1820c03511d3cf8a0312365046.jpg) + +![](images/715d8d83c2103e1968a11ab02845731529af4dd50125808b9d0243b72cd2baed.jpg) + +![](images/0e7b5f7ae7b935b8cdece9d7650035165186783fccfc94faaefed562971db8db.jpg) +CIFAR100-4 + +![](images/7cec2af3f87cdfaae066591103d22423c1a36d8c7d16b4e55e834a5175cdca01.jpg) + +![](images/99547b98d18370f43647e53c6b43ef07b154a240f8de9b343abec3531d0ec552.jpg) + +![](images/b2cff40e1c84402a1573929e3762c70364ced9b4d68c338cdf987e5114154357.jpg) + +![](images/f619e0ee88af89763c3d4dc261feed3fb6533d4f55557339888b2a0881ead613.jpg) + +![](images/c03d9cfa75d7f5094416e3d3f7f73c20f630119605d0ee8b5137cfe29295cc66.jpg) +CIFAR100-3 +(a) Baseline +Figure 5. Visualization results by t-SNE. We sampled part of the test set and encode them into representations, then visualize them in the embedding space by t-SNE. (a) We directly generate the representations using the backbone as a baseline. (b) We use the upstream encoder of SimCLR. (c) We use the encoder of SupCon. (d) We use part of SCLRE by removing the targeted adjustment loss. (e) We use the entire SCLRE. + +![](images/9380385bbebbd3e11920fda1ea224b778aedc3c4af528af16ec8a6a8ca88ea98.jpg) +(b) SimCLR + +![](images/655630c26a3a012c46fbf843d8f0ccdd7a9edcf1af41b11cf0d31f3abbd3e994.jpg) +(c) SupCon + +![](images/b431494ab043b92ca92389694a2536816674a73d69413df42d422792a14225f2.jpg) +(d) SCLRE w\o targets + +![](images/f0069a6cb1947ee08f4d16d96fed3bf4db794432fc8dfe782c67430572faca62.jpg) +(e) SCLRE + +tradictory semantic information. Comparatively, SCLRE works well by the superclass-aware decision boundary. + +Accuracy on High-Resolution Datasets. Tab. 2 shows the evaluation results on three high-resolution datasets, from which we observe the following facts. 1) we observe that, on the real-world dataset FMoW, SCLRE outperforms other methods by a large margin. On FMoW, the geographic location of building images is adopted as the superclass label, rather than their type. The methods without representation enhancement may perform worse due to the lack of superclass-aware common features. This further confirms the importance of representation enhancement in the superclass problems. 2) SCLRE also outperforms other methods on real-world dataset Adience, which contains 26580 face images from 2284 people. In case every single person is a raw class, the results confirm that SCLRE is also workable in large-scale situations. 3) Moreover, we observe that on FMoW and Adience, SimCLR performs even worse than the baseline. Because the instance-level representation by SimCLR is conflicted with the given superclass labels. 4) The performance on mini-[ImageNet appears to be less attractive. This may be because the mini-[ImageNet dataset is divided into 2 distinct subclasses: animals and non-animals. For such kind of distribution, basic-level methods + +
mini-ImageNet [30]FMoW [20]Adience [7]VOC [8]
Baseline87.756.565.778.3
SupCon87.058.366.278.2
SimCLR90.151.552.980.0
SCLRE89.364.868.781.5
+ +Table 2. Classification accuracy on high-resolution datasets. We compared the accuracy on datasets with more complex and informative content. + +can also easily sort them out by exploring the obvious common features inside each superclass(e.g. whether the image contains eyes), which makes it more like a traditional instance-level problem, and SimCLR may perform comparably or even slightly better than SCLRE. + +# 4.3. Visualization + +Here we generate the visualization results on representation space by the t-SNE technique [28]. The CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets are adopted for the visualization experiments. + +We observe the following facts from Fig. 5. 1) It is observed that in columns (a) and (b) the images are scattered in their respective representation space, although the images (b) cluster slightly closer than the ones in (a). This further shows that the instance-level representation of SimCLR is + +little related to superclass labels. Consequently, in superclass scenarios, SimCLR performs a little better than baseline but worse than SCLRE, just as illustrated in Tab. 1. 2) We observed that the images in column (c) have a more obvious cluster structure than in columns (a) and (b). This shows that the use of superclass labels is important for constructing clear boundaries. But due to the diverse and complex superclass contents, the representations are mostly inaccurate and thus result in less favorable results as reported in Tab. 1. 3) It is also observed that the images in columns (d) and (e) are notably closer than the ones in (a), (b), and (c). This shows that the enhanced representation of SCLRE is superclass-aware. 4) Furtherly, we observe that the images in column (e) are obviously closer than the ones in column (d). This shows that the targeted adjustment loss, one important component of SCLRE, can furtherly cluster images in the same subclasses, accordingly, it benefits reconstructing a superclass-aware decision boundary. + +# 4.4. Sensitivity Analysis + +We discuss the hyperparameters $\alpha$ and $\lambda$ of SCLRE by evaluating their sensitivity on the CIFAR100-7 dataset. When analyzing $\alpha$ , to avoid the influence of target anchors, we removed targeted adjustment loss by fixing $\lambda$ to 0. We test the sensitivity of $\alpha$ by ranging it from 0 to 1. When analyzing $\lambda$ , we fix $\alpha$ to 0.5 to avoid its influence. + +
α00.10.30.50.71
Accuracy69.272.173.775.173.6-
+ +Table 3. Sensitivity analysis of $\alpha$ . We range $\alpha$ from 0 to 1 and calculate the accuracy of SCLRE respectively. When $\alpha = 1$ , the cross-entropy loss is invalid and the model will lose its classification ability. + +
λ00.10.20.40.81
Accuracy75.176.777.976.978.176.0
+ +Table 4. Sensitivity analysis of $\lambda$ . We range $\lambda$ from 0 to 1 and calculate the accuracy of SCLRE respectively. + +
Batch Size64128256512
Accuracy76.776.977.377.9
+ +Sensitivity of $\alpha$ . Tab. 3 shows the sensitivity of $\alpha$ . Besides extreme situations, we find that $\alpha$ has a little influence on the accuracy. The optimal value of $\alpha$ falls around 0.5 and we fix it in our experiments. + +In the case of $\alpha = 0$ , the contrastive adjustment loss is invalid and the model degenerates into the baseline. In the case of $\alpha = 1$ , the cross-entropy loss is invalid and then the model is not workable for classification. + +Table 5. Sensitivity analysis of batch size. We vary batch size from 64 to 512 and calculate the accuracy of SCLRE respectively. + +
ArchitectureParams.(M)FLOPs(G)Acc.(%)
Baseline23.511.3068.9
+CIA37.211.3169.2
+Lca--72.7
+CIA, Lca--75.1
SCLRE (+CIA, Lca, Lta)37.211.3277.9
+ +Table 6. Ablation study on SCLRE. We calculated the model size (Params.), computational complexity (FLOPs) and accuracy contributions of four components of SCLRE. + +Sensitivity of $\lambda$ . Tab. 4 demonstrates the sensitivity analysis of $\lambda$ . $\lambda$ is a hyperparameter that controls the influence of targeted adjustment loss. + +Unlike $\alpha$ , there are two most proper $\lambda$ values which fall around 0.2 and 0.8. When $\lambda$ is too small or too large, the performance is both less satisfactory. As a too large $\lambda$ may be harmful to the accuracy of the model, a too small $\lambda$ may be short of guidance and has limited performance improvement. + +Sensitivity of Batch Size. Intuitively, a larger batch size can bring more abundant samples of superclasses, thus making it easier to construct superclass-aware representations. The experimental results are conducted to verify this statement. Tab. 5 illustrated the results of the sensitivity study of the batch size. The results show that a larger batch size can improve the performance of SCLRE, though the improvement is not significant. + +# 4.5. Ablation Study + +Here ablation studies are conducted to demonstrate the contribution of each component in our proposed SCLRE. Just as designed in Sec. 3, SCLRE is composed of the baseline model, contrastive adjustment loss $(\mathcal{L}_{ca})$ , cross-instance attention module(CIA), and targeted adjustment loss $(\mathcal{L}_{ta})$ . The baseline model is the same as our SCLRE without $\mathcal{L}_{ca}$ , $\mathcal{L}_{ta}$ and CIA. + +According to the results reported in Tab. 6, we have the following observations: + +1) It is observed that only CIA itself has an observable contribution to the performance improvements, but CIA joint with $\mathcal{L}_{ca}$ brings notable improvements. This indicates that the CIA module is workable to break the basic class boundary with the great assistance of $\mathcal{L}_{ca}$ . They enhance the representation of the superclass by cross-instance attention mechanism. 2) With CIA, $\mathcal{L}_{ca}$ and $\mathcal{L}_{ta}$ , SCLRE achieves the optimal improvements. It shows that the targeted adjustment loss does function on reconstructing superclass-aware boundary on the enhanced representation space. + +
ResNet-18ResNet-34ResNet-50ResNet-101ResNet-152
Baseline83.983.778.283.482.6
SCLRE86.887.682.987.787.2
Improve.(%)+2.9+2.9+4.7+4.3+4.6
+ +Table 7. Robustness research of SCLRE on different backbones. We calculated the accuracy on a CIFAR100-3 superclass dataset of baseline and SCLRE. We also calculated the improvement SCLRE gets in the different backbones. + +# 4.6. Robustness for Different Backbones + +We investigate the robustness of SCLRE by changing the backbone to other convolutional neural networks. Tab. 7 shows the result of the experiments on robustness. The experiments are performed on a CIFAR100-3 superclass dataset and we vary the backbone from smaller networks to larger networks. The result shows that SCLRE can make stable improvements on different backbones. + +# 5. Analysis of Generalization Ability + +We conduct an analysis of the generalization ability of SCLRE and prove that the cross-instance attention module can further compress the upper bound of the generalization error. The details of the proof are in the Appendix. + +Based on the contrastive adjustment loss in Sec. 3.2, and with ignoring the constant items, we can rewrite the contrastive loss into an expectation form: + +$$ +\begin{array}{l} \mathcal{L}_{ca}\propto -\mathbb{E}_{v,v^{\prime}}\mathbb{E}_{\substack{v_{1},v_{2}\in P(v)\\ v^{-}\in P(v^{\prime})}}\log \frac{\exp(v_{1}^{T}v_{2})}{\exp(v_{1}^{T}v_{2}) + \exp(v_{1}^{T}v^{-})} \\ = \underbrace {- \mathbb {E} _ {v} \mathbb {E} _ {v _ {1} , v _ {2} \in P (v)} v _ {1} ^ {T} v _ {2}} _ {L _ {1}} \\ + \underbrace {\mathbb {E} _ {v , v ^ {\prime}} \mathbb {E} _ {v _ {1} , v _ {2} \in P (v)} \log \left(\exp (v _ {1} ^ {T} v _ {2}) + \exp (v _ {1} ^ {T} v ^ {-})\right)} _ {L _ {2}} \\ \end{array} +$$ + +where $L_{1}$ measures the alignment between two transformed features and $L_{2}$ is the regularizer preventing the collapse of representation. + +Lemma 1. As a conclusion in [17], if $L_{2}$ is trained to be small than the threshold, which is easily satisfied, the generalization error of downstream classifier $G_{f}$ has an upper bound: + +$$ +\begin{array}{l} E r r (G _ {f}) \leqslant (1 - \sigma) + \eta (\epsilon) \sqrt {2 - 2 \mathbb {E} _ {v} \mathbb {E} _ {v _ {1} , v _ {2} \in P (v)} v _ {1} ^ {T} v _ {2}} \\ = (1 - \sigma) + \eta (\epsilon) \sqrt {2 + 2 L _ {1}} \\ \end{array} +$$ + +Lemma 2. Sample pairs with the same coarse label consist of two parts: those share the same fine label $(O_1)$ and those have different fine labels $(O_2)$ . The former can be closely aligned after contrastive learning. We focus on the latter and prove that: + +$$ +\mathbb {E} _ {v} \mathbb {E} _ {v _ {1}, v _ {2} \in P (v)} v _ {1} ^ {T} v _ {2} \geqslant C _ {\varphi} + \frac {(1 - \rho) K}{M _ {1} M _ {2}} s (a _ {1}, a _ {2}) +$$ + +where $C_{\varphi}$ is a constant determined by the data distribution. $M_1, M_2, K$ are constants. $\rho$ is the ratio of sample pairs with same fine label in all positive pairs. + +Theorem 1. The generalization error has another upper bound based on attention vectors according to lemma 1 and lemma 2: + +$$ +\begin{array}{l} E r r (G _ {f}) \leqslant (1 - \sigma) + \\ \sqrt {2} \eta (\epsilon) \sqrt {1 - C _ {\varphi} - \frac {(1 - \rho) K}{M _ {1} M _ {2}} \mathbb {E} _ {v _ {1} , v _ {2} \in O _ {2}} s (a _ {1} , a _ {2}))} \\ \end{array} +$$ + +The attention vector is a representation of the importance between samples. When we define a superclass, the representation vector $z$ with the same coarse label will be pulled together by the contrastive loss. During this process, samples in the superclass will focus on the same important enhanced representations, thus leading to similar vectors and declining the upper bound of generalization error. + +# 6. Conclusion + +In this study, we explore the high-level categories recognition problem, i.e., superclass categorization. And a superclass learning framework, that exploits self-attention techniques cross the instances to enhance the representation, is proposed. Thus, the distribution of superclass is modified in the enhanced representation space and a new superclass-aware decision boundary is then reconstructed. In theory, the generalization error of SCLRE can be bounded by attention constraints. In extensive experiments, SCLRE outperforms the SOTA classification methods. + +In the near future, we would like to tackle the superclass problems under the scenarios without full annotation by leveraging semi-supervised learning and active learning. + +Acknowledgements. This work was supported by the National Key Research & Develop Plan(2018YFB1004401); National Natural Science Foundation of China (62276270,62072460, 62172424); Beijing Natural Science Foundation (4212022). it is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence. + +# References + +[1] Amjad Almahairi, Nicolas Ballas, Tim Coolijmans, Yin Zheng, H. Larochelle, and Aaron C. Courville. Dynamic capacity networks. In ICML, 2016. 3 +[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2 +[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. ArXiv, abs/2002.05709, 2020. 2, 5 +[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 2 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 3 +[7] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Transactions on information forensics and security, 9(12):2170-2179, 2014. 5, 6 +[8] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2009. 5, 6 +[9] Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2010. 1 +[10] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4476-4484, 2017. 3 +[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2 +[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3 +[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2 +[15] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5 +[16] Qiushi Huang, Tom Ko, H Lilian Tang, Xubo Liu, and Boyong Wu. Token-level supervised contrastive learning for punctuation restoration. In Interspeech, 2021. 2 +[17] Weiran Huang, Mingyang Yi, and Xuyang Zhao. Towards the generalization of contrastive self-supervised learning. arXiv preprint arXiv:2111.00743, 2021. 8 +[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In NIPS, 2015. 3 +[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. ArXiv, abs/2004.11362, 2020. 2, 5 +[20] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton A. Earnshaw, Imran S. Haque, Sara Beery, Jure Leskovec, Anshul Kundaje, Emma Pierson, Sergey Levine, Chelsea Finn, and Percy Liang. WILDS: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning (ICML), 2021. 5, 6 +[21] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 2 +[22] Junnan Li, Pan Zhou, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Prototypical contrastive learning of unsupervised representations. ArXiv, abs/2005.04966, 2021. 2 +[23] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogério Schmidt Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6908-6918, 2022. 2, 4 +[24] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2 +[25] Zheda Mai, Ruiwen Li, Hyunwoo J. Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3584-3594, 2021. 2 +[26] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2 + +[27] Lars Schmarje, Monty Santarossa, Simon-Martin Schroder, and Reinhard Koch. A survey on semi-, self- and unsupervised learning for image classification. IEEE Access, 9:82146-82168, 2021. 1, 2 +[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3 +[30] Oriol Vinyals, Charles Blundell, Timothy P. Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 5, 6 +[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In-So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3 +[32] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2020. 3 +[33] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. ArXiv, abs/2010.13902, 2020. 2 +[34] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-842, 2021. 2 +[35] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10022-10031, 2021. 2 \ No newline at end of file diff --git a/2023/Superclass Learning With Representation Enhancement/images.zip b/2023/Superclass Learning With Representation Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a54732096ccff89608ee392f589ece0d27c254d3 --- /dev/null +++ b/2023/Superclass Learning With Representation Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba7c353cc810322a6f84aa8aca1f78148b860bf64200923c881397d30d05442f +size 441331 diff --git a/2023/Superclass Learning With Representation Enhancement/layout.json b/2023/Superclass Learning With Representation Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3374dc617e74c353257805f178b13f55ee185b7 --- /dev/null +++ b/2023/Superclass Learning With Representation Enhancement/layout.json @@ -0,0 +1,9666 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "text", + "content": "Superclass Learning with Representation Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 142, + 178, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 142, + 178, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 178, + 157 + ], + "type": "text", + "content": "Zeyu Gan" + }, + { + "bbox": [ + 115, + 142, + 178, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 143, + 283, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 143, + 283, + 157 + ], + "spans": [ + { + "bbox": [ + 202, + 143, + 283, + 157 + ], + "type": "text", + "content": "Suyun Zhao" + }, + { + "bbox": [ + 202, + 143, + 283, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 158, + 284, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 158, + 284, + 171 + ], + "spans": [ + { + "bbox": [ + 216, + 158, + 284, + 171 + ], + "type": "text", + "content": "Hong Chen" + }, + { + "bbox": [ + 216, + 158, + 284, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 143, + 375, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 143, + 375, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 375, + 157 + ], + "type": "text", + "content": "Jinlong Kang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 399, + 143, + 473, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 143, + 473, + 157 + ], + "spans": [ + { + "bbox": [ + 399, + 143, + 473, + 157 + ], + "type": "text", + "content": "Liyuan Shang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 158, + 374, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 158, + 374, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 158, + 374, + 171 + ], + "type": "text", + "content": "Cuiping " + }, + { + "bbox": [ + 309, + 158, + 374, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{L}\\mathbf{i}^{1,2}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 171, + 531, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 531, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 531, + 200 + ], + "type": "text", + "content": "1Key Lab of Data Engineering and Knowledge Engineering of MOE Renmin University of China \n2Renmin University of China, Beijing, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 201, + 503, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 201, + 503, + 213 + ], + "spans": [ + { + "bbox": [ + 86, + 201, + 503, + 213 + ], + "type": "text", + "content": "{zygan, zhaosuyun, kangjinlong, shangliyuan4032, chong, LICUping}@ruc.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 266, + 290, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 542 + ], + "type": "text", + "content": "In many real scenarios, data are often divided into a handful of artificial super categories in terms of expert knowledge rather than the representations of images. Concretely, a superclass may contain massive and various raw categories, such as refuse sorting. Due to the lack of common semantic features, the existing classification techniques are intractable to recognize superclass without raw class labels, thus they suffer severe performance damage or require huge annotation costs. To narrow this gap, this paper proposes a superclass learning framework, called SuperClass Learning with Representation Enhancement(SCLRE), to recognize super categories by leveraging enhanced representation. Specifically, by exploiting the self-attention technique across the batch, SCLRE collapses the boundaries of those raw categories and enhances the representation of each superclass. On the enhanced representation space, a superclass-aware decision boundary is then reconstructed. Theoretically, we prove that by leveraging attention techniques the generalization error of SCLRE can be bounded under superclass scenarios. Experimentally, extensive results demonstrate that SCLRE outperforms the baseline and other contrastive-based methods on CIFAR-100 datasets and four high-resolution datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 694 + ], + "type": "text", + "content": "In recent decades, basic-level raw categorization (e.g. cats vs dogs, apples vs bananas) has greatly developed [9, 27] while high-level or super-coarse-grained visual categorization (e.g., recyclable waste vs kitchen waste, creatures vs non-creatures) has received little attention. In many real scenarios, there often exist a handful of high-level categories, wherein numerous images from diverse basic-level categories share one common label. We tend to define this kind of super-coarse-grained class as Superclass." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 239, + 426, + 332 + ], + "blocks": [ + { + "bbox": [ + 308, + 239, + 426, + 332 + ], + "lines": [ + { + "bbox": [ + 308, + 239, + 426, + 332 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 426, + 332 + ], + "type": "image", + "image_path": "a24cc70987046ff79367574bad04da73da7a217a5ed361c57c8f9b6bc12b72d1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 434, + 240, + 536, + 285 + ], + "blocks": [ + { + "bbox": [ + 434, + 240, + 536, + 285 + ], + "lines": [ + { + "bbox": [ + 434, + 240, + 536, + 285 + ], + "spans": [ + { + "bbox": [ + 434, + 240, + 536, + 285 + ], + "type": "image", + "image_path": "0f6e157739725d90d4c41884b9037e7881d6442447e6adc792b2b231c3273962.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 335, + 376, + 401 + ], + "blocks": [ + { + "bbox": [ + 308, + 335, + 376, + 401 + ], + "lines": [ + { + "bbox": [ + 308, + 335, + 376, + 401 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 376, + 401 + ], + "type": "image", + "image_path": "bdc5511a9e598233af0f47ca10706d2cd00ffe6cbb2a2a406e17ccb3a1050de6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 403, + 373, + 411 + ], + "lines": [ + { + "bbox": [ + 310, + 403, + 373, + 411 + ], + "spans": [ + { + "bbox": [ + 310, + 403, + 373, + 411 + ], + "type": "text", + "content": "(a) Image Feature Domain" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 389, + 335, + 455, + 401 + ], + "blocks": [ + { + "bbox": [ + 389, + 335, + 455, + 401 + ], + "lines": [ + { + "bbox": [ + 389, + 335, + 455, + 401 + ], + "spans": [ + { + "bbox": [ + 389, + 335, + 455, + 401 + ], + "type": "image", + "image_path": "18f9b99a87527fa81959da251f8889ee2b0573c7f12f19a752546e126de18b77.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 403, + 452, + 411 + ], + "lines": [ + { + "bbox": [ + 391, + 403, + 452, + 411 + ], + "spans": [ + { + "bbox": [ + 391, + 403, + 452, + 411 + ], + "type": "text", + "content": "(b) Separate Fine Domain" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 468, + 335, + 536, + 401 + ], + "blocks": [ + { + "bbox": [ + 468, + 335, + 536, + 401 + ], + "lines": [ + { + "bbox": [ + 468, + 335, + 536, + 401 + ], + "spans": [ + { + "bbox": [ + 468, + 335, + 536, + 401 + ], + "type": "image", + "image_path": "d3dc8ab9df2cc866ad3e76b08bf80995eda9a018993739b909a7859875970070.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 403, + 541, + 411 + ], + "lines": [ + { + "bbox": [ + 466, + 403, + 541, + 411 + ], + "spans": [ + { + "bbox": [ + 466, + 403, + 541, + 411 + ], + "type": "text", + "content": "(c) Superclass Concept Domain" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 424, + 545, + 489 + ], + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 489 + ], + "type": "text", + "content": "Figure 1. Illustration of superclass learning. The samples from a same superclass will be scatteredly distributed in the embedding space. The process of superclass learning is to break old domains and construct new domains. Red indicates the superclass of kitchen waste, and blue indicates the superclass of recyclable waste." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 518, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 625 + ], + "type": "text", + "content": "Refuse sorting, as an example, is such a recognition problem with four superclasses, i.e., kitchen waste, recyclable waste, hazardous waste, and others waste. One task of refuse sorting is to accurately collect various items, such as rotten fruits, bones, raw vegetables, and eggshells, into kitchen waste. Traditional recognition needs to identify what exact basic-level categories they are, then sort them out. Obviously, it is wasteful and unrealistic for superclass identification." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Essentially, high-level superclasses contain two characteristics, remarkably distinct from basic-level classes. First, the basic-level classes contained in superclass problems are usually scattered and share few common features. As depicted in the top-left corner subgraph of Fig. 1, the fruit apple, bone, and eggs are remote from each other in feature spaces, though all of them belong to kitchen waste." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24060" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "Second, the instances from two distinct superclasses may share common features. Just as illustrated in Fig. 1, fruit apple, from kitchen waste, and toy apple, from recyclable waste, are close to each other as they share common semantic features. Obviously, the above-mentioned characteristics indicate that the smoothness assumption [27] under basic-level classification (nearby images tend to have the same label) does not hold in the superclass scenarios. Thus, the existing classification techniques based on smoothness assumption are not practically deployable and scalable and they may suffer severe performance damage in superclass settings. Accordingly, it is valuable and promising to investigate superclass identification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 443 + ], + "type": "text", + "content": "To tackle the superclass problems, we have to address two main challenges. First, we need to break the original decision boundaries of basic-level classes and disclose a superclass-aware boundary at the basic class level. As depicted in bottom subgraph (a) of Fig. 1, the boundary of the apple domain is original, however, it is useless and even harmful for the refuse sorting as both fruit apples and toy apples belong to this domain. To get the required domain boundary, the apple domain needs to be separated into the fruit domain and toy domain by leveraging their individual local features, as depicted in the bottom subgraph (b) of Fig. 1. In superclass scenarios, it is not enough to investigate the boundary at a basic class level. Consequently, the second challenge is to reconstruct a decision boundary at a superclass level. To achieve this end, it is necessary to merge the domain of classes, such as fruit apples, eggs, and bones into a new rotten superclass domain, as depicted in the bottom subgraph (c) of Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 444, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 289, + 575 + ], + "type": "text", + "content": "In this paper, we propose a SuperClass Learning framework with Representation Enhancement. Considering that the semantic representation at the basic class level is not workable for superclass recognition, we propose one cross-instance attention module which could seize the representation across the instances with the same superclass label. By leveraging contrastive adjustment loss, the attention mechanism enhances this representation. Moreover, to overcome the imbalance distribution of superclasses, we adopt target adjustment loss to reconstruct a superclass-aware decision boundary on the enhanced representation space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 575, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 288, + 598 + ], + "type": "text", + "content": "In summary, this paper makes the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 609, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 609, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 609, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 59, + 609, + 287, + 644 + ], + "type": "text", + "content": "- We propose an under-study but realistic problem, superclass identification, that has notably distinct distribution from basic-level categorization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 654, + 288, + 714 + ], + "type": "text", + "content": "- We propose a novel representation enhancement method by leveraging cross-instance attention and then exploit it in superclass identification. And by theoretical analyses, we verify that this self-attention technique can bound the generalization error of superclass" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 73, + 376, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 376, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 376, + 84 + ], + "type": "text", + "content": "recognition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 91, + 545, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 545, + 127 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that SCLRE outperforms the SOTA classification techniques on one artificial superclass dataset and three real datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "content": "The remainder of the paper is organized as follows: in Sec. 2 we briefly review related work and in Sec. 3 we describe our method for superclass recognition. Then, extensive experiments and generalization error analysis are conducted in Secs. 4 and 5. Finally, Sec. 6 draws a brief conclusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 235, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 547, + 462 + ], + "type": "text", + "content": "Contrastive Learning. The contrastive framework has achieved great success on instance-level problems. Chen et al [3] designed a simple framework to get visual representations in a contrastive way. He et al [14] developed a dynamic dictionary with queues and moving averages to improve contrastive encoder quality. To learn basic-level features and discover more semantic structures for the embedding space, Li et al [22] bridge contrastive learning with clustering. More recently, Caron et al [2] presented a contrastive framework without the need for pairwise comparison computation. To get rid of the limitation of negative samples, Grill et al [12] use two neural networks and predict the representation of the same image under a different augmented view. Chen et al [4] further designed a stop-gradient operation to prevent collapsing. Contrastive learning has also been applied to various research domains and evolved quickly [11, 33, 34]. Instead of focusing on the basic-level boundary, we construct a high-level superclass-aware boundary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 463, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 547, + 641 + ], + "type": "text", + "content": "Supervised Contrastive Loss. Traditional contrastive frameworks encode data either unsupervised or self-supervised and apply supervised information in downstream tasks. Khosla et al [19] extend the contrastive mode to a fully-supervised setting and use label information more effectively. To address the impact of long-tailed distribution data, Li et al [23] propose a targeted contrastive loss and guide the encoder to obtain representations based on several preset anchors. Recently, supervised contrastive learning is used in broader research domains both in computer vision and natural language processing [16,25, 35]. We are inspired by the function of pulling close or away features in these supervised contrastive losses, and therefore propose a new loss, which will pull the features by the superclass labels for better performance in classification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Self-Attention Module. Self-attention module has been widely used in deep learning. It was first mentioned by Vaswani et al [29] as a part of the transformer model and is widely used in the natural language processing field [5, 21, 24, 26]. Recently, the attention module and transformer also obtained great success in the computer vision" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24061" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 279, + 272 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 279, + 272 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 279, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 279, + 272 + ], + "type": "image", + "image_path": "81486a35630af085eae0ba8a9fbd1854d1bb712364438f0ccbb74a45bd629405.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 289, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 335 + ], + "type": "text", + "content": "Figure 2. Illustration of SCLRE. The images first generate their representations through a convolutional neural network, then mix with each other in a trainable cross-instance attention module for enhancement. After enhancement, the representations are then adjusted according to their superclass labels and the target anchors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 367, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 289, + 510 + ], + "type": "text", + "content": "field. Dosovitskiy et al [6] organize image patches as a sequence and propose a vision transformer. Furthermore, He et al [13] propose a masked auto-encoder based on vision transformers. Self-attention modules can also be integrated into convolutional neural networks to improve performance [1, 18, 31]. Fu et al [10] address the fine-grained image recognition problem by using an attention network to find the unique and pivotal features in samples (for example, one kind of bird's tail color). Different from focusing the attention relationship in one sample, we propose a cross-instance attention(CIA) module that cares more about the relationship between instances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 538, + 103, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 103, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 103, + 551 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 564, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 660 + ], + "type": "text", + "content": "Superclass Learning with Representation Enhancement(SCLRE) framework is proposed for superclass image recognition. It exploits the self-attention technique across instances to perform a representation enhancement, thus it achieves the goal to break the basic-level boundary in representations. Then SCLRE reconstructs a new superclass-aware boundary in the enhanced representation space by a series of adjustment losses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Fig. 2 shows the overview of SCLRE. Sec. 3.1 presents the process of breaking basic-level boundary, while Sec. 3.2 describes the details of constructing the new high-level superclass-aware boundary." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 547, + 135 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 547, + 135 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 135 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 135 + ], + "type": "image", + "image_path": "a4a9f5c0094da18574f9cf87e87c98adbfecaff13cb10171d816766958c13456.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 144, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 547, + 178 + ], + "type": "text", + "content": "Figure 3. Different segmentation strategies of cross-instance attention module. We experiment on three kinds of segmentation strategies and choose the best for the experiment." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 190, + 468, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 468, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 468, + 204 + ], + "type": "text", + "content": "3.1. Representation Enhancement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 209, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 290 + ], + "type": "text", + "content": "To break the basic-level boundary of representations and obtain valuable enhancements, we mix the representations by leveraging a cross-instance attention module. Unlike the existing attention techniques [32], the cross-instance attention module cares more about the relationship between representations of the instances, rather than the feature representation inside the instance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": "Enhanced Representation. Let " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " be the data space, " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^D" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": "-dimensional embedding space, an encoder " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{R}^D" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " be a mapping from the data space to the embedding space, and an enhancement process " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{EnH}: \\mathcal{R}^D \\to \\mathcal{R}^D" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " be a mapping from the embedding space to the embedding space. For each " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "x_j \\in \\mathcal{X}" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": ", the representation " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "z_j = f(x_j)" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "Z \\in \\mathcal{R}^{n \\times D}" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " is the representation matrix for the batch with size " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": ". The enhanced representation " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " is an output of " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{EnH}" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": ". More specifically, " + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 305, + 292, + 546, + 400 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 411, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 411, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 365, + 411, + 545, + 425 + ], + "type": "interline_equation", + "content": "v _ {j} = \\operatorname {E n H} \\left(z _ {j}; Z\\right) = a _ {j} \\times Z, \\tag {1}", + "image_path": "8d3a87153985363af1fd1292cc08f5c14af8cbf6d2462750901043eca7f54c20.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " is a trainable vector for representation " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " computed by cross-instance attention module. Actually, the vector " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "a_{j}" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": "-th row of the attention distribution matrix(ADM) computed in the following Q-K-V mode multi-head attention module [29]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 497, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 497, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 357, + 497, + 545, + 525 + ], + "type": "interline_equation", + "content": "\\operatorname {A D M} (Z) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right). \\tag {2}", + "image_path": "a7c8e40f9b51c9fd1e3d6f20bdedf360d2efab82b9c72d97e98fe109d534a11b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "Q = Z \\times W_{Q}" + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "W_{Q} \\in \\mathcal{R}^{D \\times d_{q}}" + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "content": ". Matrix " + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "K = Z \\times W_{K}" + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "W_{K} \\in \\mathcal{R}^{D \\times d_{k}}" + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "d_{q} = d_{k}" + }, + { + "bbox": [ + 306, + 533, + 545, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": "The enhancement process breaks the basic-level boundary by merging the semantic features of instances. Cross-instance attention module automatically explores the key instances for representation enhancement and integrates them to obtain a superclass-aware enhanced representation. Moreover, we conduct strict theoretical analysis on the generalization ability of our model in Sec. 5, and discover that the cross-instance module can tightly bound the generalization error of SCLRE." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Data Segmentation Strategy. In the cross-instance attention module, the segmentation strategy of input data is important for representation enhancement. For each instance, the segmentation strategy determines whose repre" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24062" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": "sentations will be received as an enhancement. We considered three kinds of segmentation strategies for the input data in the cross-instance attention module. Fig. 3 illustrates the details of the strategies. In (a), images are gathered in terms of their superclass label, we put the same label images into one batch. In (b), we randomly and evenly select images from the different superclasses and gather them into the batch. In (c), we keep the distribution of the dataset and randomly select the samples to form the batch. The result shows that (c) selecting the samples according to the distribution of the dataset benefits the model most, so we fixed this strategy in the following part of this work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 221, + 132, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 132, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 132, + 233 + ], + "type": "text", + "content": "3.2. SCLRE Loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 334 + ], + "type": "text", + "content": "To reconstruct the high-level superclass-aware boundary of representations, we design a SCLRE loss to adjust the representations after the attention-weighted enhancement. We keep using the letter " + }, + { + "bbox": [ + 46, + 240, + 287, + 334 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 240, + 287, + 334 + ], + "type": "text", + "content": " in Eq. (1) to demonstrate representations being enhanced by the cross-instance attention module, and we call it enhanced representation. After that, we adjust the representations by taking three aspects into consideration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 372 + ], + "type": "text", + "content": "Category Classification Loss. We first adopt the cross-entropy loss as the base to obtain a category-level adjustment and prevent the model from collapsing:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 376, + 287, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 376, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 91, + 376, + 287, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i, k} \\log M (v _ {i}) _ {k}, \\tag {3}", + "image_path": "fdf641b983713722bbf64f5cfe9dc883f7327a7761b2307efb9b932de4f613be.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": " demonstrates batch size, " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": " demonstrates the superclass label of the " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": "-th sample in one-hot embedding, " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "M(v_{i})" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": " demonstrates the output prediction of " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": ", and the footprint " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": " demonstrates the " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": "-th component of a vector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": "Contrastive Adjustment Loss. Besides classification loss, the samples belonging to the same superclass may share few common features. To obtain a representation-level adjustment, we design a contrastive adjustment loss to pull close the samples in the same superclass and pull away samples in distinct subclasses. We directly regard samples from the same superclass as positive pairs and samples from other subclasses as negative pairs. For one sample " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " and one of its positive sample " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "v^{+}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "v^{+}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " belongs to the same superclass), the contrastive adjustment loss is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 586, + 286, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 586, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 58, + 586, + 286, + 620 + ], + "type": "interline_equation", + "content": "\\ell \\left(v _ {i}, v ^ {+}\\right) = - \\log \\frac {\\exp \\left(s \\left(v _ {i} , v ^ {+}\\right) / \\tau\\right)}{\\exp \\left(s \\left(v _ {i} , v ^ {+}\\right) / \\tau\\right) + \\sum_ {v ^ {-} \\notin P \\left(v _ {i}\\right)} \\exp \\left(s \\left(v _ {i} , v ^ {-}\\right) / \\tau\\right)}. \\tag {4}", + "image_path": "d5c2b0648d30f634e373dd41eb79087c007421c736d64fb6a93c76b05d5c2b7d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 621, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 287, + 644 + ], + "type": "text", + "content": "By summing up all the losses of positive pairs and for a batch with size " + }, + { + "bbox": [ + 47, + 621, + 287, + 644 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 621, + 287, + 644 + ], + "type": "text", + "content": ", we can get the total loss as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 650, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 650, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 61, + 650, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c a} = \\sum_ {i = 1} ^ {N} \\frac {1}{| | P (v _ {i}) | | - 1} \\sum_ {v ^ {+} \\in P (v _ {i}) \\backslash \\{v _ {i} \\}} \\ell \\left(v _ {i}, v ^ {+}\\right), \\tag {5}", + "image_path": "b6c0069ad74c980f6d7f795c1716ca9751309f51ad3896276e07147d50d130e5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "P(v_{i})" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " stands for all the other data that share the same superclass label with " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " in the batch, i.e. " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "P(v_{i}) =" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\{v_{k}|p(v_{k}) = p(v_{i})\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "text", + "content": " means the superclass label. " + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "s(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "text", + "content": " stands for the similarity between two vectors, we use cosine similarity for measurement, i.e. " + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "s(v_{s},v_{t}) = \\frac{v_{s}^{T}v_{t}}{\\|v_{s}\\|\\cdot\\|v_{t}\\|}" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "text", + "content": " is the temperature hyperparameter. " + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "||P(v_i)||" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "text", + "content": " stands for the size of set " + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "inline_equation", + "content": "P(v_{i})" + }, + { + "bbox": [ + 305, + 72, + 545, + 137 + ], + "type": "text", + "content": " , and it aims to normalize the total loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 138, + 546, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 138, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 546, + 269 + ], + "type": "text", + "content": "Targeted Adjustment Loss. Since superclass learning naturally faces the problem of lack of classification centers due to the scattered distribution of representations, then follows damage to the constructing process when the contrastive adjustment loss works. To address this problem, we design a targeted adjustment loss to give each superclass a pre-defined category center (i.e. target anchors). We first generate the target anchors and make them far away from each other inspired by [23]. Then we pull every sample close to its pre-defined target anchor by minimizing the following targeted adjustment loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 337, + 281, + 545, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 281, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 337, + 281, + 545, + 314 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t a} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp (s \\left(v _ {i} , t \\left(v _ {i}\\right)\\right) / \\tau)}{\\sum_ {t \\in T} \\exp (s \\left(v _ {i} , t\\right) / \\tau)}, \\tag {6}", + "image_path": "2a1d097c96d435cd84a618e3343cf73f6c8c9322caf9408f26bb517d579e90f5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": " stands for batch size, " + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "inline_equation", + "content": "s(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": " stands for cosine similarity function, " + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": " is the set of all the target anchors, and " + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "inline_equation", + "content": "t(v_{i})" + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": " demonstrates the target anchor allocated to " + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 305, + 325, + 545, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 363, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 546, + 434 + ], + "type": "text", + "content": "SCLRE Adjustment Loss. To construct the high-level superclass-aware boundary with both category-level and instance-level adjustments and overcome the problem of lack of classification center, the total loss of SCLRE is a weighted summation of the above three losses(i.e. Eqs. (3), (5) and (6)):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 447, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 447, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 341, + 447, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S C L R E} = (1 - \\alpha) \\mathcal {L} _ {c e} + \\alpha \\mathcal {L} _ {c a} + \\lambda \\mathcal {L} _ {t a}, \\tag {7}", + "image_path": "c08c6cb9da3b6b512dbb5e968ac9fcf398664a8497025cdd8a048cc545d3bd75.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " range from 0 to 1. " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ca}" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " are designed as a combination to help the encoder explore the concept of superclasses, a larger " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " means a stronger adjustment strategy and weaker classification ability. " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ta}" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " is designed to help the model overcome the imbalanced distribution, and a larger " + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": " means stronger guidance towards the destination of adjustment." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "type": "text", + "content": "SCLRE adjustment loss can help the model learn more about the concept of the superclass and the high-level categorization boundary. By balancing the influence of cross-entropy loss and contrastive adjustment loss, the model can adjust representations properly. By controlling the utilization of targeted adjustment loss, the model can run smoothly and converge steadily." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 656, + 382, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 382, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 382, + 670 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments on several artificially constructed superclass datasets to demonstrate the effectiveness of our proposed approach, SCLRE." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24063" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 289, + 210 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 210 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 210 + ], + "type": "image", + "image_path": "609df7a44d6c32fd6a63754200b04bdceb25bd2698b0a0861e9a4a1924098e1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 288, + 266 + ], + "type": "text", + "content": "Figure 4. Illustration of superclass construction. We integrate the initial class labels of the original dataset to superclass labels, according to field knowledge. The constructed dataset contains a handful of subclasses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 287, + 155, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 155, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 155, + 300 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "text", + "content": "Datasets. In this study, superclass datasets are artificially reorganized from three benchmarks, CIFAR-100, mini-ImageNet [30] and VOC [8], and two real-world datasets, FMoW [20] and Adience [7]. CIFAR-100 contains 50000 training images and 10000 test images from 100 categories with a size of " + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 3" + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "text", + "content": ". mini-ImageNet [30] contains 60000 images in total with a size of " + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "inline_equation", + "content": "84 \\times 84 \\times 3" + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "text", + "content": ", which are evenly distributed among 100 classes. VOC [8] contains 2501 training images, 2510 validation images and 4952 test images from 20 categories. The real-world dataset FMoW [20] is about a hybrid domain generalization and subpopulation shift problem, which consists of 76863 training images, 11483 validation images and 11327 test images with a size of " + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 46, + 305, + 287, + 486 + ], + "type": "text", + "content": ". And the real-world dataset Adience [7] contains 26,580 images from 2,284 people." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 677 + ], + "type": "text", + "content": "In our experiments, superclass datasets are artificially constructed by integrating the initial classes, just as illustrated in Fig. 4. To be concrete, We reorganize the 100 classes of CIFAR-100 into 3 subclasses, 4 subclasses, and 7 subclasses, respectively, forming three distinct superclass datasets, CIFAR100-3, CIFAR100-4, and CIFAR100-7. The subclasses in the three data sets are based on movement mode, life form, and are more complex than the original 20 coarse classes in CIFAR-100. For miniImagenet and VOC, we reorganize their subclasses into 2 subclasses based on life form. For FMoW, we drop the original labels and turn to predict the geographical location of the images. For Adience, we utilize the age groups as the superclass labels. Thus, those 7 superclass datasets can well simulate the distribution of real-world superclass problems. More reorganization details are listed in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Compared Methods. We compare SCLRE with the baseline model, i.e., ResNet [15], and two SOTA contrastive techniques SupCon [19], and SimCLR [3]. As contrastive" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 69, + 526, + 230 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 526, + 230 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 526, + 230 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 526, + 230 + ], + "type": "table", + "html": "
DatasetMethodAccuracy(%)
CIFAR100-3Baseline72.8
SupCon [19]78.1
SimCLR [3]79.0
SCLRE80.1
CIFAR100-4Baseline76.0
SupCon [19]80.1
SimCLR [3]80.6
SCLRE84.0
CIFAR100-7Baseline68.9
SupCon [19]72.7
SimCLR [3]73.9
SCLRE78.1
", + "image_path": "56a1fc5e498a00aa071ea332bdcd0a086acfeb1b67484df9281a2fb8c3f7fe84.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 238, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 270 + ], + "type": "text", + "content": "Table 1. Classification accuracy on low pixels dataset: CIFAR-100. We compared the classification accuracy on CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 290, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 314 + ], + "type": "text", + "content": "techniques perform excellently in representation learning. We adopt the default optimal settings in the training details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 350 + ], + "type": "text", + "content": "SupCon: In our experiments, SupCon adopts ResNet-50 as backbone, and SupCon loss as the super-classification loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": "SimCLR: It is a known self-supervised contrastive learning framework. In superclass recognition, we keep using ResNet-50 as the backbone and use a supervised downstream task to keep the instance-level features pure, wherein the upstream instance-level results are a kind of learning of the basic level class labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "text", + "content": "Implementation details. We adopt ResNet-50 [15] as our default backbone and we reduce the feature dimension from 2048 to 128 with an extra Multi-Layer Perceptron(MLP) projection head. We train the network using the SGD with a momentum of 0.9 and a batch size of 64. We set the learning rate as 0.001 and the training stage as 200 epochs. Additionally, we resize all the images to " + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "text", + "content": " pixels. For our SCLRE method, we set the key " + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 256" + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "text", + "content": ", value " + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{V} = 128" + }, + { + "bbox": [ + 304, + 422, + 545, + 542 + ], + "type": "text", + "content": " and 8 multi-heads are adopted in the attention module." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 551, + 444, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 444, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 444, + 562 + ], + "type": "text", + "content": "4.2. Classification Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Accuracy on CIFAR-100 Datasets. Tab. 1 shows the classification accuracies of SCLRE and the compared methods on CIFAR100-3, CIFAR100-4, and CIFAR100-7, respectively. And we have the following observations: 1) It is observed that SCLRE has a great improvement compared with the baseline. This shows that the cross-instance attention module can effectively activate the contrastive adjustment loss. 2) SCLRE has higher classification accuracies than SimCLR on superclass datasets. This shows that the instance-level semantic information learned by SimCLR may fail in the superclass problem because the images from the same superclass may have quite different, even con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24064" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 113, + 79, + 189, + 152 + ], + "blocks": [ + { + "bbox": [ + 52, + 113, + 100, + 123 + ], + "lines": [ + { + "bbox": [ + 52, + 113, + 100, + 123 + ], + "spans": [ + { + "bbox": [ + 52, + 113, + 100, + 123 + ], + "type": "text", + "content": "CIFAR100-7" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 113, + 79, + 189, + 152 + ], + "lines": [ + { + "bbox": [ + 113, + 79, + 189, + 152 + ], + "spans": [ + { + "bbox": [ + 113, + 79, + 189, + 152 + ], + "type": "image", + "image_path": "06f53139251811f5874f0e64e7e48052768f42b461ad91ea8db539a69374ee69.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 200, + 80, + 276, + 152 + ], + "blocks": [ + { + "bbox": [ + 200, + 80, + 276, + 152 + ], + "lines": [ + { + "bbox": [ + 200, + 80, + 276, + 152 + ], + "spans": [ + { + "bbox": [ + 200, + 80, + 276, + 152 + ], + "type": "image", + "image_path": "e0479c9dc9936b3fd319e48b6ee109247249ab34b61c6405043e8042fa442f74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 289, + 80, + 364, + 152 + ], + "blocks": [ + { + "bbox": [ + 289, + 80, + 364, + 152 + ], + "lines": [ + { + "bbox": [ + 289, + 80, + 364, + 152 + ], + "spans": [ + { + "bbox": [ + 289, + 80, + 364, + 152 + ], + "type": "image", + "image_path": "b3f8ff954d68f6ffa179375ea2da723a0b11d3f966aee4c16c2f3d2838da5528.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 375, + 80, + 450, + 152 + ], + "blocks": [ + { + "bbox": [ + 375, + 80, + 450, + 152 + ], + "lines": [ + { + "bbox": [ + 375, + 80, + 450, + 152 + ], + "spans": [ + { + "bbox": [ + 375, + 80, + 450, + 152 + ], + "type": "image", + "image_path": "58df38e3ae129d775a41f5cdecb19c03590bfb1820c03511d3cf8a0312365046.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 462, + 80, + 538, + 153 + ], + "blocks": [ + { + "bbox": [ + 462, + 80, + 538, + 153 + ], + "lines": [ + { + "bbox": [ + 462, + 80, + 538, + 153 + ], + "spans": [ + { + "bbox": [ + 462, + 80, + 538, + 153 + ], + "type": "image", + "image_path": "715d8d83c2103e1968a11ab02845731529af4dd50125808b9d0243b72cd2baed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 114, + 159, + 188, + 232 + ], + "blocks": [ + { + "bbox": [ + 53, + 191, + 99, + 201 + ], + "lines": [ + { + "bbox": [ + 53, + 191, + 99, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 99, + 201 + ], + "type": "text", + "content": "CIFAR100-4" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 114, + 159, + 188, + 232 + ], + "lines": [ + { + "bbox": [ + 114, + 159, + 188, + 232 + ], + "spans": [ + { + "bbox": [ + 114, + 159, + 188, + 232 + ], + "type": "image", + "image_path": "0e7b5f7ae7b935b8cdece9d7650035165186783fccfc94faaefed562971db8db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 201, + 159, + 276, + 232 + ], + "blocks": [ + { + "bbox": [ + 201, + 159, + 276, + 232 + ], + "lines": [ + { + "bbox": [ + 201, + 159, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 201, + 159, + 276, + 232 + ], + "type": "image", + "image_path": "7cec2af3f87cdfaae066591103d22423c1a36d8c7d16b4e55e834a5175cdca01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 289, + 159, + 364, + 232 + ], + "blocks": [ + { + "bbox": [ + 289, + 159, + 364, + 232 + ], + "lines": [ + { + "bbox": [ + 289, + 159, + 364, + 232 + ], + "spans": [ + { + "bbox": [ + 289, + 159, + 364, + 232 + ], + "type": "image", + "image_path": "99547b98d18370f43647e53c6b43ef07b154a240f8de9b343abec3531d0ec552.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 159, + 449, + 232 + ], + "blocks": [ + { + "bbox": [ + 376, + 159, + 449, + 232 + ], + "lines": [ + { + "bbox": [ + 376, + 159, + 449, + 232 + ], + "spans": [ + { + "bbox": [ + 376, + 159, + 449, + 232 + ], + "type": "image", + "image_path": "b2cff40e1c84402a1573929e3762c70364ced9b4d68c338cdf987e5114154357.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 463, + 160, + 537, + 232 + ], + "blocks": [ + { + "bbox": [ + 463, + 160, + 537, + 232 + ], + "lines": [ + { + "bbox": [ + 463, + 160, + 537, + 232 + ], + "spans": [ + { + "bbox": [ + 463, + 160, + 537, + 232 + ], + "type": "image", + "image_path": "f619e0ee88af89763c3d4dc261feed3fb6533d4f55557339888b2a0881ead613.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 114, + 239, + 188, + 312 + ], + "blocks": [ + { + "bbox": [ + 53, + 267, + 99, + 276 + ], + "lines": [ + { + "bbox": [ + 53, + 267, + 99, + 276 + ], + "spans": [ + { + "bbox": [ + 53, + 267, + 99, + 276 + ], + "type": "text", + "content": "CIFAR100-3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 114, + 239, + 188, + 312 + ], + "lines": [ + { + "bbox": [ + 114, + 239, + 188, + 312 + ], + "spans": [ + { + "bbox": [ + 114, + 239, + 188, + 312 + ], + "type": "image", + "image_path": "c03d9cfa75d7f5094416e3d3f7f73c20f630119605d0ee8b5137cfe29295cc66.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 327, + 169, + 338 + ], + "lines": [ + { + "bbox": [ + 128, + 327, + 169, + 338 + ], + "spans": [ + { + "bbox": [ + 128, + 327, + 169, + 338 + ], + "type": "text", + "content": "(a) Baseline" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 352, + 547, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 547, + 395 + ], + "type": "text", + "content": "Figure 5. Visualization results by t-SNE. We sampled part of the test set and encode them into representations, then visualize them in the embedding space by t-SNE. (a) We directly generate the representations using the backbone as a baseline. (b) We use the upstream encoder of SimCLR. (c) We use the encoder of SupCon. (d) We use part of SCLRE by removing the targeted adjustment loss. (e) We use the entire SCLRE." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 201, + 239, + 276, + 312 + ], + "blocks": [ + { + "bbox": [ + 201, + 239, + 276, + 312 + ], + "lines": [ + { + "bbox": [ + 201, + 239, + 276, + 312 + ], + "spans": [ + { + "bbox": [ + 201, + 239, + 276, + 312 + ], + "type": "image", + "image_path": "9380385bbebbd3e11920fda1ea224b778aedc3c4af528af16ec8a6a8ca88ea98.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 327, + 257, + 338 + ], + "lines": [ + { + "bbox": [ + 215, + 327, + 257, + 338 + ], + "spans": [ + { + "bbox": [ + 215, + 327, + 257, + 338 + ], + "type": "text", + "content": "(b) SimCLR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 289, + 239, + 364, + 311 + ], + "blocks": [ + { + "bbox": [ + 289, + 239, + 364, + 311 + ], + "lines": [ + { + "bbox": [ + 289, + 239, + 364, + 311 + ], + "spans": [ + { + "bbox": [ + 289, + 239, + 364, + 311 + ], + "type": "image", + "image_path": "655630c26a3a012c46fbf843d8f0ccdd7a9edcf1af41b11cf0d31f3abbd3e994.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 327, + 343, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 327, + 343, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 343, + 338 + ], + "type": "text", + "content": "(c) SupCon" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 376, + 239, + 450, + 312 + ], + "blocks": [ + { + "bbox": [ + 376, + 239, + 450, + 312 + ], + "lines": [ + { + "bbox": [ + 376, + 239, + 450, + 312 + ], + "spans": [ + { + "bbox": [ + 376, + 239, + 450, + 312 + ], + "type": "image", + "image_path": "b431494ab043b92ca92389694a2536816674a73d69413df42d422792a14225f2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 327, + 452, + 338 + ], + "lines": [ + { + "bbox": [ + 376, + 327, + 452, + 338 + ], + "spans": [ + { + "bbox": [ + 376, + 327, + 452, + 338 + ], + "type": "text", + "content": "(d) SCLRE w\\o targets" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 463, + 239, + 537, + 312 + ], + "blocks": [ + { + "bbox": [ + 463, + 239, + 537, + 312 + ], + "lines": [ + { + "bbox": [ + 463, + 239, + 537, + 312 + ], + "spans": [ + { + "bbox": [ + 463, + 239, + 537, + 312 + ], + "type": "image", + "image_path": "f0069a6cb1947ee08f4d16d96fed3bf4db794432fc8dfe782c67430572faca62.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 327, + 515, + 338 + ], + "lines": [ + { + "bbox": [ + 478, + 327, + 515, + 338 + ], + "spans": [ + { + "bbox": [ + 478, + 327, + 515, + 338 + ], + "type": "text", + "content": "(e) SCLRE" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 417, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 441 + ], + "type": "text", + "content": "tradictory semantic information. Comparatively, SCLRE works well by the superclass-aware decision boundary." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 450, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 713 + ], + "type": "text", + "content": "Accuracy on High-Resolution Datasets. Tab. 2 shows the evaluation results on three high-resolution datasets, from which we observe the following facts. 1) we observe that, on the real-world dataset FMoW, SCLRE outperforms other methods by a large margin. On FMoW, the geographic location of building images is adopted as the superclass label, rather than their type. The methods without representation enhancement may perform worse due to the lack of superclass-aware common features. This further confirms the importance of representation enhancement in the superclass problems. 2) SCLRE also outperforms other methods on real-world dataset Adience, which contains 26580 face images from 2284 people. In case every single person is a raw class, the results confirm that SCLRE is also workable in large-scale situations. 3) Moreover, we observe that on FMoW and Adience, SimCLR performs even worse than the baseline. Because the instance-level representation by SimCLR is conflicted with the given superclass labels. 4) The performance on mini-[ImageNet appears to be less attractive. This may be because the mini-[ImageNet dataset is divided into 2 distinct subclasses: animals and non-animals. For such kind of distribution, basic-level methods" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 308, + 415, + 544, + 462 + ], + "blocks": [ + { + "bbox": [ + 308, + 415, + 544, + 462 + ], + "lines": [ + { + "bbox": [ + 308, + 415, + 544, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 544, + 462 + ], + "type": "table", + "html": "
mini-ImageNet [30]FMoW [20]Adience [7]VOC [8]
Baseline87.756.565.778.3
SupCon87.058.366.278.2
SimCLR90.151.552.980.0
SCLRE89.364.868.781.5
", + "image_path": "eb51c939d8ae35367f7a268126eb1ce8cc10a075ecd66c5fe743ca007f651d1e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 471, + 545, + 503 + ], + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 503 + ], + "type": "text", + "content": "Table 2. Classification accuracy on high-resolution datasets. We compared the accuracy on datasets with more complex and informative content." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 517, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 578 + ], + "type": "text", + "content": "can also easily sort them out by exploring the obvious common features inside each superclass(e.g. whether the image contains eyes), which makes it more like a traditional instance-level problem, and SimCLR may perform comparably or even slightly better than SCLRE." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 586, + 390, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 390, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 390, + 597 + ], + "type": "text", + "content": "4.3. Visualization" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 653 + ], + "type": "text", + "content": "Here we generate the visualization results on representation space by the t-SNE technique [28]. The CIFAR100-3, CIFAR100-4, and CIFAR100-7 datasets are adopted for the visualization experiments." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "We observe the following facts from Fig. 5. 1) It is observed that in columns (a) and (b) the images are scattered in their respective representation space, although the images (b) cluster slightly closer than the ones in (a). This further shows that the instance-level representation of SimCLR is" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24065" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "little related to superclass labels. Consequently, in superclass scenarios, SimCLR performs a little better than baseline but worse than SCLRE, just as illustrated in Tab. 1. 2) We observed that the images in column (c) have a more obvious cluster structure than in columns (a) and (b). This shows that the use of superclass labels is important for constructing clear boundaries. But due to the diverse and complex superclass contents, the representations are mostly inaccurate and thus result in less favorable results as reported in Tab. 1. 3) It is also observed that the images in columns (d) and (e) are notably closer than the ones in (a), (b), and (c). This shows that the enhanced representation of SCLRE is superclass-aware. 4) Furtherly, we observe that the images in column (e) are obviously closer than the ones in column (d). This shows that the targeted adjustment loss, one important component of SCLRE, can furtherly cluster images in the same subclasses, accordingly, it benefits reconstructing a superclass-aware decision boundary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 294, + 162, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 162, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 162, + 306 + ], + "type": "text", + "content": "4.4. Sensitivity Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": "We discuss the hyperparameters " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": " of SCLRE by evaluating their sensitivity on the CIFAR100-7 dataset. When analyzing " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": ", to avoid the influence of target anchors, we removed targeted adjustment loss by fixing " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": " to 0. We test the sensitivity of " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": " by ranging it from 0 to 1. When analyzing " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": ", we fix " + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 312, + 287, + 384 + ], + "type": "text", + "content": " to 0.5 to avoid its influence." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 392, + 276, + 420 + ], + "blocks": [ + { + "bbox": [ + 59, + 392, + 276, + 420 + ], + "lines": [ + { + "bbox": [ + 59, + 392, + 276, + 420 + ], + "spans": [ + { + "bbox": [ + 59, + 392, + 276, + 420 + ], + "type": "table", + "html": "
α00.10.30.50.71
Accuracy69.272.173.775.173.6-
", + "image_path": "418e5259c0adf52acfbebe0d4b6a24a4cc92e6615bd8fc333a43ab1ee802159d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 53, + 485, + 282, + 513 + ], + "blocks": [ + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "text", + "content": "Table 3. Sensitivity analysis of " + }, + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "text", + "content": ". We range " + }, + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "text", + "content": " from 0 to 1 and calculate the accuracy of SCLRE respectively. When " + }, + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 429, + 287, + 473 + ], + "type": "text", + "content": ", the cross-entropy loss is invalid and the model will lose its classification ability." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 485, + 282, + 513 + ], + "lines": [ + { + "bbox": [ + 53, + 485, + 282, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 485, + 282, + 513 + ], + "type": "table", + "html": "
λ00.10.20.40.81
Accuracy75.176.777.976.978.176.0
", + "image_path": "39a5a9d5c68ca14b1d993419c3ac2ff38928ca8e25715302152d0f3cc61844b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 80, + 556, + 255, + 583 + ], + "blocks": [ + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "type": "text", + "content": "Table 4. Sensitivity analysis of " + }, + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "type": "text", + "content": ". We range " + }, + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 521, + 287, + 544 + ], + "type": "text", + "content": " from 0 to 1 and calculate the accuracy of SCLRE respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 556, + 255, + 583 + ], + "lines": [ + { + "bbox": [ + 80, + 556, + 255, + 583 + ], + "spans": [ + { + "bbox": [ + 80, + 556, + 255, + 583 + ], + "type": "table", + "html": "
Batch Size64128256512
Accuracy76.776.977.377.9
", + "image_path": "df104f321c4edddcd7d6c00dc504608951b77f5869f8d1a411312e8511fa13f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "text", + "content": "Sensitivity of " + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "text", + "content": ". Tab. 3 shows the sensitivity of " + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "text", + "content": ". Besides extreme situations, we find that " + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "text", + "content": " has a little influence on the accuracy. The optimal value of " + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 618, + 287, + 665 + ], + "type": "text", + "content": " falls around 0.5 and we fix it in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "In the case of " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": ", the contrastive adjustment loss is invalid and the model degenerates into the baseline. In the case of " + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": ", the cross-entropy loss is invalid and then the model is not workable for classification." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 69, + 533, + 159 + ], + "blocks": [ + { + "bbox": [ + 46, + 591, + 287, + 614 + ], + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 614 + ], + "type": "text", + "content": "Table 5. Sensitivity analysis of batch size. We vary batch size from 64 to 512 and calculate the accuracy of SCLRE respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 69, + 533, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 533, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 533, + 159 + ], + "type": "table", + "html": "
ArchitectureParams.(M)FLOPs(G)Acc.(%)
Baseline23.511.3068.9
+CIA37.211.3169.2
+Lca--72.7
+CIA, Lca--75.1
SCLRE (+CIA, Lca, Lta)37.211.3277.9
", + "image_path": "cc01d5d462bc24164d625ba959f00e28ec6a87209dc8a011e3bb95411e6d246c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 166, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 199 + ], + "type": "text", + "content": "Table 6. Ablation study on SCLRE. We calculated the model size (Params.), computational complexity (FLOPs) and accuracy contributions of four components of SCLRE." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "text", + "content": "Sensitivity of " + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "text", + "content": ". Tab. 4 demonstrates the sensitivity analysis of " + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 219, + 545, + 255 + ], + "type": "text", + "content": " is a hyperparameter that controls the influence of targeted adjustment loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "content": "Unlike " + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "content": ", there are two most proper " + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "content": " values which fall around 0.2 and 0.8. When " + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "content": " is too small or too large, the performance is both less satisfactory. As a too large " + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "content": " may be harmful to the accuracy of the model, a too small " + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 258, + 545, + 328 + ], + "type": "text", + "content": " may be short of guidance and has limited performance improvement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 333, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 333, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 545, + 429 + ], + "type": "text", + "content": "Sensitivity of Batch Size. Intuitively, a larger batch size can bring more abundant samples of superclasses, thus making it easier to construct superclass-aware representations. The experimental results are conducted to verify this statement. Tab. 5 illustrated the results of the sensitivity study of the batch size. The results show that a larger batch size can improve the performance of SCLRE, though the improvement is not significant." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 445, + 400, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 400, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 400, + 459 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "text", + "content": "Here ablation studies are conducted to demonstrate the contribution of each component in our proposed SCLRE. Just as designed in Sec. 3, SCLRE is composed of the baseline model, contrastive adjustment loss " + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ca})" + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "text", + "content": ", cross-instance attention module(CIA), and targeted adjustment loss " + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{ta})" + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "text", + "content": ". The baseline model is the same as our SCLRE without " + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ca}" + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ta}" + }, + { + "bbox": [ + 305, + 468, + 545, + 552 + ], + "type": "text", + "content": " and CIA." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 555, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 578 + ], + "type": "text", + "content": "According to the results reported in Tab. 6, we have the following observations:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "1) It is observed that only CIA itself has an observable contribution to the performance improvements, but CIA joint with " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ca}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " brings notable improvements. This indicates that the CIA module is workable to break the basic class boundary with the great assistance of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ca}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". They enhance the representation of the superclass by cross-instance attention mechanism. 2) With CIA, " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ca}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ta}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", SCLRE achieves the optimal improvements. It shows that the targeted adjustment loss does function on reconstructing superclass-aware boundary on the enhanced representation space." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24066" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 69, + 478, + 122 + ], + "blocks": [ + { + "bbox": [ + 118, + 69, + 478, + 122 + ], + "lines": [ + { + "bbox": [ + 118, + 69, + 478, + 122 + ], + "spans": [ + { + "bbox": [ + 118, + 69, + 478, + 122 + ], + "type": "table", + "html": "
ResNet-18ResNet-34ResNet-50ResNet-101ResNet-152
Baseline83.983.778.283.482.6
SCLRE86.887.682.987.787.2
Improve.(%)+2.9+2.9+4.7+4.3+4.6
", + "image_path": "f0eb61f22e22ac0e99f43c1275b8669fc8824434f20bc60d29f512c58ab6b653.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 129, + 547, + 153 + ], + "lines": [ + { + "bbox": [ + 46, + 129, + 547, + 153 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 547, + 153 + ], + "type": "text", + "content": "Table 7. Robustness research of SCLRE on different backbones. We calculated the accuracy on a CIFAR100-3 superclass dataset of baseline and SCLRE. We also calculated the improvement SCLRE gets in the different backbones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 169, + 239, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 169, + 239, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 239, + 182 + ], + "type": "text", + "content": "4.6. Robustness for Different Backbones" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 188, + 288, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 288, + 273 + ], + "type": "text", + "content": "We investigate the robustness of SCLRE by changing the backbone to other convolutional neural networks. Tab. 7 shows the result of the experiments on robustness. The experiments are performed on a CIFAR100-3 superclass dataset and we vary the backbone from smaller networks to larger networks. The result shows that SCLRE can make stable improvements on different backbones." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 281, + 236, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 236, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 236, + 295 + ], + "type": "text", + "content": "5. Analysis of Generalization Ability" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 300, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 348 + ], + "type": "text", + "content": "We conduct an analysis of the generalization ability of SCLRE and prove that the cross-instance attention module can further compress the upper bound of the generalization error. The details of the proof are in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 384 + ], + "type": "text", + "content": "Based on the contrastive adjustment loss in Sec. 3.2, and with ignoring the constant items, we can rewrite the contrastive loss into an expectation form:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 388, + 279, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 388, + 279, + 485 + ], + "spans": [ + { + "bbox": [ + 52, + 388, + 279, + 485 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal{L}_{ca}\\propto -\\mathbb{E}_{v,v^{\\prime}}\\mathbb{E}_{\\substack{v_{1},v_{2}\\in P(v)\\\\ v^{-}\\in P(v^{\\prime})}}\\log \\frac{\\exp(v_{1}^{T}v_{2})}{\\exp(v_{1}^{T}v_{2}) + \\exp(v_{1}^{T}v^{-})} \\\\ = \\underbrace {- \\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2}} _ {L _ {1}} \\\\ + \\underbrace {\\mathbb {E} _ {v , v ^ {\\prime}} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} \\log \\left(\\exp (v _ {1} ^ {T} v _ {2}) + \\exp (v _ {1} ^ {T} v ^ {-})\\right)} _ {L _ {2}} \\\\ \\end{array}", + "image_path": "1d8a869f3c5d30cb52a5086060d5c135b3b791b27da708c6b2e5b2828c0b881c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "type": "text", + "content": " measures the alignment between two transformed features and " + }, + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 488, + 287, + 523 + ], + "type": "text", + "content": " is the regularizer preventing the collapse of representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "text", + "content": "Lemma 1. As a conclusion in [17], if " + }, + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "text", + "content": " is trained to be small than the threshold, which is easily satisfied, the generalization error of downstream classifier " + }, + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "inline_equation", + "content": "G_{f}" + }, + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "text", + "content": " has an upper bound:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 574, + 280, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 574, + 280, + 611 + ], + "spans": [ + { + "bbox": [ + 53, + 574, + 280, + 611 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E r r (G _ {f}) \\leqslant (1 - \\sigma) + \\eta (\\epsilon) \\sqrt {2 - 2 \\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1} , v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2}} \\\\ = (1 - \\sigma) + \\eta (\\epsilon) \\sqrt {2 + 2 L _ {1}} \\\\ \\end{array}", + "image_path": "73c3fb1bfecbc44896d3c3094e13c954132ca7321620d691c56ee0bcd4687fec.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "type": "text", + "content": "Lemma 2. Sample pairs with the same coarse label consist of two parts: those share the same fine label " + }, + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "type": "inline_equation", + "content": "(O_1)" + }, + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "type": "text", + "content": " and those have different fine labels " + }, + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "type": "inline_equation", + "content": "(O_2)" + }, + { + "bbox": [ + 46, + 628, + 287, + 688 + ], + "type": "text", + "content": ". The former can be closely aligned after contrastive learning. We focus on the latter and prove that:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 691, + 261, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 261, + 717 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 261, + 717 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {v} \\mathbb {E} _ {v _ {1}, v _ {2} \\in P (v)} v _ {1} ^ {T} v _ {2} \\geqslant C _ {\\varphi} + \\frac {(1 - \\rho) K}{M _ {1} M _ {2}} s (a _ {1}, a _ {2})", + "image_path": "81071e2d5dbb2b9d452b79dd9195dbb09021e01679842b1276e2a97346f09689.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "inline_equation", + "content": "C_{\\varphi}" + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "text", + "content": " is a constant determined by the data distribution. " + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "inline_equation", + "content": "M_1, M_2, K" + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "text", + "content": " are constants. " + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 306, + 170, + 545, + 206 + ], + "type": "text", + "content": " is the ratio of sample pairs with same fine label in all positive pairs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 209, + 547, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 547, + 244 + ], + "type": "text", + "content": "Theorem 1. The generalization error has another upper bound based on attention vectors according to lemma 1 and lemma 2:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 258, + 531, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 258, + 531, + 306 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 531, + 306 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E r r (G _ {f}) \\leqslant (1 - \\sigma) + \\\\ \\sqrt {2} \\eta (\\epsilon) \\sqrt {1 - C _ {\\varphi} - \\frac {(1 - \\rho) K}{M _ {1} M _ {2}} \\mathbb {E} _ {v _ {1} , v _ {2} \\in O _ {2}} s (a _ {1} , a _ {2}))} \\\\ \\end{array}", + "image_path": "a10031f98f3aa97d652a12b624dd92f23ce2f1ddc472cffab2122533f45345f9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "type": "text", + "content": "The attention vector is a representation of the importance between samples. When we define a superclass, the representation vector " + }, + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 324, + 545, + 408 + ], + "type": "text", + "content": " with the same coarse label will be pulled together by the contrastive loss. During this process, samples in the superclass will focus on the same important enhanced representations, thus leading to similar vectors and declining the upper bound of generalization error." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 429, + 379, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 379, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 379, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 452, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 571 + ], + "type": "text", + "content": "In this study, we explore the high-level categories recognition problem, i.e., superclass categorization. And a superclass learning framework, that exploits self-attention techniques cross the instances to enhance the representation, is proposed. Thus, the distribution of superclass is modified in the enhanced representation space and a new superclass-aware decision boundary is then reconstructed. In theory, the generalization error of SCLRE can be bounded by attention constraints. In extensive experiments, SCLRE outperforms the SOTA classification methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 574, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 611 + ], + "type": "text", + "content": "In the near future, we would like to tackle the superclass problems under the scenarios without full annotation by leveraging semi-supervised learning and active learning." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 624, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Key Research & Develop Plan(2018YFB1004401); National Natural Science Foundation of China (62276270,62072460, 62172424); Beijing Natural Science Foundation (4212022). it is also partially supported by the Opening Fund of Hebei Key Laboratory of Machine Learning and Computational Intelligence." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24067" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Amjad Almahairi, Nicolas Ballas, Tim Coolijmans, Yin Zheng, H. Larochelle, and Aaron C. Courville. Dynamic capacity networks. In ICML, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "type": "text", + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "text", + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. ArXiv, abs/2002.05709, 2020. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 306 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 308, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 373 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "text", + "content": "[7] Eran Eidinger, Roee Enbar, and Tal Hassner. Age and gender estimation of unfiltered faces. IEEE Transactions on information forensics and security, 9(12):2170-2179, 2014. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "text", + "content": "[8] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88:303-338, 2009. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 519 + ], + "type": "text", + "content": "[9] Pedro F Felzenszwalb, Ross B Girshick, David McAllester, and Deva Ramanan. Object detection with discriminatively trained part-based models. IEEE transactions on pattern analysis and machine intelligence, 32(9):1627-1645, 2010. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 576 + ], + "type": "text", + "content": "[10] Jianlong Fu, Heliang Zheng, and Tao Mei. Look closer to see better: Recurrent attention convolutional neural network for fine-grained image recognition. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4476-4484, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 610 + ], + "type": "text", + "content": "[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[12] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[15] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "text", + "content": "[16] Qiushi Huang, Tom Ko, H Lilian Tang, Xubo Liu, and Boyong Wu. Token-level supervised contrastive learning for punctuation restoration. In Interspeech, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 260 + ], + "type": "text", + "content": "[17] Weiran Huang, Mingyang Yi, and Xuyang Zhao. Towards the generalization of contrastive self-supervised learning. arXiv preprint arXiv:2111.00743, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 262, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 293 + ], + "type": "text", + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In NIPS, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 295, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 337 + ], + "type": "text", + "content": "[19] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. ArXiv, abs/2004.11362, 2020. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 339, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 437 + ], + "type": "text", + "content": "[20] Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton A. Earnshaw, Imran S. Haque, Sara Beery, Jure Leskovec, Anshul Kundaje, Emma Pierson, Sergey Levine, Chelsea Finn, and Percy Liang. WILDS: A benchmark of in-the-wild distribution shifts. In International Conference on Machine Learning (ICML), 2021. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "text", + "content": "[21] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[22] Junnan Li, Pan Zhou, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Prototypical contrastive learning of unsupervised representations. ArXiv, abs/2005.04966, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 516, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 516, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 545, + 569 + ], + "type": "text", + "content": "[23] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogério Schmidt Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6908-6918, 2022. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[24] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 615, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 678 + ], + "type": "text", + "content": "[25] Zheda Mai, Ruiwen Li, Hyunwoo J. Kim, and Scott Sanner. Supervised contrastive replay: Revisiting the nearest class mean classifier in online class-incremental continual learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3584-3594, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24068" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 453 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Lars Schmarje, Monty Santarossa, Simon-Martin Schroder, and Reinhard Koch. A survey on semi-, self- and unsupervised learning for image classification. IEEE Access, 9:82146-82168, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "type": "text", + "content": "[30] Oriol Vinyals, Charles Blundell, Timothy P. Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NIPS, 2016. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 262 + ], + "type": "text", + "content": "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In-So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "text", + "content": "[32] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8805-8814, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "text", + "content": "[33] Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. ArXiv, abs/2010.13902, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 396 + ], + "type": "text", + "content": "[34] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 833-842, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "text", + "content": "[35] Mingkai Zheng, Fei Wang, Shan You, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. Weakly supervised contrastive learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10022-10031, 2021. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24069" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_content_list.json b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c4efc8af75d32ce07b6bd30192ef59bdc0749cd --- /dev/null +++ b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_content_list.json @@ -0,0 +1,1461 @@ +[ + { + "type": "text", + "text": "Supervised Masked Knowledge Distillation for Few-Shot Transformers", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Lin*, Guangxing Han\\*†, Jiawei Ma, Shiyuan Huang, Xudong Lin, Shih-Fu Chang Columbia University", + "bbox": [ + 130, + 180, + 833, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{h13199, gh2561, jiawei.m, sh3813, x12798, sc250}@columbia.edu", + "bbox": [ + 210, + 227, + 759, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 277, + 313, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) emerge to achieve impressive performance on many data-abundant computer vision tasks by capturing long-range dependencies among local features. However, under few-shot learning (FSL) settings on small datasets with only a few labeled data, ViT tends to overfit and suffers from severe performance degradation due to its absence of CNN-alike inductive bias. Previous works in FSL avoid such problem either through the help of self-supervised auxiliary losses, or through the dextile uses of label information under supervised settings. But the gap between self-supervised and supervised few-shot Transformers is still unfilled. Inspired by recent advances in self-supervised knowledge distillation and masked image modeling (MIM), we propose a novel Supervised Masked Knowledge Distillation model (SMKD) for few-shot Transformers which incorporates label information into self-distillation frameworks. Compared with previous self-supervised methods, we allow intra-class knowledge distillation on both class and patch tokens, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. Experimental results on four few-shot classification benchmark datasets show that our method with simple design outperforms previous methods by a large margin and achieves a new start-of-the-art. Detailed ablation studies confirm the effectiveness of each component of our model. Code for this paper is available here: https://github.com/HL-hanlin/SMKD.", + "bbox": [ + 73, + 310, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) [16] have emerged as a competitive alternative to Convolutional Neural Networks (CNNs) [31] in recent years, and have achieved impressive performance in many vision tasks including image classification [16, 38, 59, 66], object detection [3, 10, 26-28, 79], and object segmentation [50, 55]. Compared with CNNs, which introduce inductive bias through convolutional kernels with fixed receptive fields [35], the attention layers in", + "bbox": [ + 75, + 755, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a569f99be3d3fbeffe35c16e1a351519717b86e1318889b45f343c6814f476a6.jpg", + "image_caption": [ + "(a) Few-Shot Transformers with Soft Label Regularization" + ], + "image_footnote": [], + "bbox": [ + 537, + 290, + 862, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7639a8994228dc8bd335e0f4d910b602ff0d7cb61cb44573f51d684dffe1efa0.jpg", + "image_caption": [ + "(b) Few-Shot Transformers with Self-Supervision Auxiliary Regularization" + ], + "image_footnote": [], + "bbox": [ + 527, + 368, + 879, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a8d7ae397840fc6de041cf50d80ee656ca1e14a0d3bdbbb02aa83c5bbd202f1.jpg", + "image_caption": [ + "(c) Few-Shot Transformers with Our Supervised Masked Knowledge Distillation", + "Figure 1. Comparison of the proposed idea and other existing methods for few-shot Transformers. Our model mitigates the overfitting issue of few-shot Transformers, by extending the masked knowledge distillation framework into the supervised setting, and enforcing the alignment of [cls] and corresponding [patch] tokens for intra-class images." + ], + "image_footnote": [], + "bbox": [ + 504, + 455, + 887, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ViT allow it to model global token relations to capture long-range token dependencies. However, such flexibility also comes at a cost: ViT is data-hungry and it needs to learn token dependencies purely from data. This property often makes it easy to overfit to datasets with small training set and suffer from severe generalization performance degradation [36, 37]. Therefore, we are motivated to study how to make ViTs generalize well on these small datasets, especially under the few-shot learning (FSL) setting [19, 39, 62] which aims to recognize unseen new instances at test time just from only a few (e.g. one or five) labeled samples from each new categories.", + "bbox": [ + 496, + 628, + 892, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of the existing methods mitigate the overfitting issue of few-shot Transformers [14] using various regularizations. For instance, some works utilize label information in a weaker [70], softer [42] way, or use label information efficiently through patch-level supervision [14]. However, these models usually design sophisticated learning targets.", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 94, + 886, + 333, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19649", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, self-distillation techniques [4, 8], and particularly, the recent masked self-distillation [29, 45, 77], which distills knowledge learned from an uncorrupted image to the knowledge predicted from a masked image, have lead an emerging trend in self-supervised Transformers in various fields [15, 68]. Inspired by such success, recent works in FSL attempt to incorporate self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46], or to adopt a self-supervised pretraining, supervised training two-stage framework to train few-shot Transformers [18, 32]. Compared with traditional supervised methods, self-supervision can learn less biased representations towards base class, which usually leads to better generalization ability for novel classes [40]. However, the two learning objectives of self-supervision and supervision are conflicting and it is hard to balance them during training. Therefore, how to efficiently leverage the strengths of self-supervised learning to alleviate the overfitting issue of supervised training remains a challenge.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel supervised masked knowledge distillation framework (SMKD) for few-shot Transformers, which handles the aforementioned challenge through a natural extension of the self-supervised masked knowledge distillation framework into the supervised setting (shown in Fig. 1). Different from supervised contrastive learning [33] which only utilizes global image features for training, we leverage multi-scale information from the images (both global [cls] token and local [patch] tokens) to formulate the learning objectives, which has been demonstrated to be effective in the recent self-supervised Transformer methods [29, 77]. For global [cls] tokens, we can simply maximize the similarity for intra-class images. However, it is non-trivial and challenging to formulate the learning objectives for local [patch] tokens because we do not have ground-truth patch-level annotations. To address this problem, we propose to estimate the similarity between [patch] tokens across intra-class images using cross-attention, and enforce the alignment of corresponding [patch] tokens. Particularly, reconstructing masked [patch] tokens across intra-class images increases the difficulty of model learning, thus encouraging learning generalizable few-shot Transformer models by jointly exploiting the holistic knowledge of the image and the similarity of intra-class images.", + "bbox": [ + 75, + 382, + 472, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we compare our model with the existing self-supervised/supervised learning methods. Our model is a natural extension of the supervised contrastive learning method [33] and self-supervised knowledge distillation methods [4, 77]. Thus our model inherits both the advantage of method [33] for effectively leveraging label information, and the advantages of methods [4, 77] for not needing large batch size and negative samples. Meanwhile, the newly-introduced challenging task of masked", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5e95c3374f45762922ae8483735494ef364b93840e6e69b793cb0578d89dadb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 89, + 679, + 183 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/904899dc3403ebb8f1bbaaff312a687d96a9424860c65ae3953ef0ed91008a7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 88, + 875, + 183 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad243b514655fef5979eb6795ed347e98214cf1b0b10e98ca2d0f490ef8fc638.jpg", + "image_caption": [ + "Figure 2. Comparison of other self-supervised/supervised frameworks. Our method (d) is a natural extension of (b) and (c), with the newly-introduced challenging task of masked [patch] tokens reconstruction across intra-class images." + ], + "image_footnote": [], + "bbox": [ + 501, + 189, + 679, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/101bdef3dc637e02f74da15925811687b9c3e559329269caf6a547509e1845e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 189, + 885, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[patch] tokens reconstruction across intra-class images makes our method more powerful for learning generalizable few-shot Transformer models.", + "bbox": [ + 496, + 361, + 890, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared with contemporary works on few-shot Transformers [14, 32, 70], our framework enjoys several good properties from a practical point of view. (1) Our method does not introduce any additional learnable parameters besides the ViT backbone and projection head, which makes it easy to be combined with other methods [32, 70, 74]. (2) Our method is both effective and training-efficient, with stronger performance and less training time on four few-shot classification benchmarks, compared with [32, 70]. In a nutshell, our main contributions can be summarized as follows:", + "bbox": [ + 496, + 406, + 892, + 556 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new supervised knowledge distillation framework (SMKD) that incorporates class label information into self-distillation, thus filling the gap between self-supervised knowledge distillation and traditional supervised learning.", + "- Within the proposed framework, we design two supervised-contrastive losses on both class and patch levels, and introduce the challenging task of masked patch tokens reconstruction across intra-class images.", + "- Given its simple design, we test our SMKD on four few-shot datasets, and show that it achieves a new SOTA on CIFAR-FS and FC100 by a large margin, as well as competitive performance on mini-ImageNet and tiered-ImageNet using the simple prototype classification method for few-shot evaluation." + ], + "bbox": [ + 514, + 560, + 890, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 799, + 640, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot Learning. Few-shot learning aims at fast knowledge transfer from seen base classes to unseen novel classes given only a few labeled samples from each novel class. The meta-learning paradigm, which simulates few-shot tasks episodically to mimic the human-learning process in the", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19650", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "real world, once dominated FSL [22-25, 44, 53, 54, 56, 62]. Surprisingly, recent works [6, 21, 58, 72] have shown that the state-of-the-art meta-learning methods can be outperformed by simple baseline methods using just a distance-based classifier without the complicated design of meta-learning strategies. Therefore, recent methods [17, 42, 70] in FSL start to focus less on meta-learning, and more on learning embeddings with good generalization ability. Our paper follows this trend and proposes a knowledge distillation framework to learn generalizable embeddings.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformers in FSL. Variants of Transformer [61] have achieved great success in NLP [12, 49], computer vision [16, 59], and multimodal learning [48]. However, the lack of inductive bias makes Transformer infamous for its data-hungry property, which is especially significant in few-shot settings when the amount of data is limited. A line of works focuses on introducing inductive bias back to the Transformer architecture, including methods using pyramid structure [63, 64], shifted windows [38], and explicit convolutional token embeddings [66]. With this said, some recent works [13, 24, 32, 71] still show that few-shot Transformers have the potential of fast adaptation to novel classes. Our work also studies few-shot Transformers, and we show that our method can work well even on the vanilla ViT structure without explicit incorporation of inductive bias.", + "bbox": [ + 75, + 244, + 472, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervision for FSL. Self-supervised learning (SSL) has shown great potential in FSL due to its good generalization ability to novel classes. Previous methods incorporate SSL into FSL in various ways. Some works propose to include self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46]. For example, [37] proposed a regularization loss that extracts additional information from images by predicting the geometric distance between patch tokens. [46] designed their loss by discouraging spatially disordered attention maps based on the idea that objects usually occupy connected regions, and [43] derived their auxiliary loss from self-supervision tasks of rotation and exemplar. Some other works [18, 32], instead, adopt a two-stage procedure by pretraining a model via self-supervision before supervised training. [32] takes advantage of self-supervised training with iBOT [77] as a pretext task, and then uses inner loop token importance reweighting for supervised fine-tuning. [18] initializes its model with a pre-trained self-supervised DINO [4], then trained with supervised cross-entropy loss. Our work follows the second branch of work, but the difference is obvious from previous works. Instead of designing complicated training pipeline or leveraging additional learnable modules at inference time, we use supervised training on self-supervised pre-trained model, and focus mainly on bridging the gap between self-supervised and supervised knowledge distillation with minimum extra design.", + "bbox": [ + 75, + 474, + 472, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another emerging trend in SSL is Masked Image Mod", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "eling (MIM) [1, 29, 45, 77], which aims at recovering the patch-level target (e.g. image pixels, patch features) of the masked content in a corrupted input image. In iBOT [77], the class and patch tokens share the same projection head, thus helping the ViT backbone to learn semantically meaningful patch embeddings, which can be used as a supplement to image-level global self-supervision.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 498, + 215, + 648, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present our Supervised Masked Knowledge Distillation (SMKD) framework in this section. Fig. 3 shows an overview of our model. Task definition for few-shot classification is presented in Sec. 3.1. We explain the formulation of SSL knowledge distillation in Sec. 3.2, and how we extend it into our supervised knowledge distillation framework in Sec. 3.3. Training pipeline is given in Sec. 3.4.", + "bbox": [ + 496, + 244, + 893, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learning Task Definition", + "text_level": 1, + "bbox": [ + 498, + 364, + 728, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In few-shot classification, we are given two datasets $\\mathcal{D}_{base}$ and $\\mathcal{D}_{novel}$ with disjoint class labels $\\mathcal{C}_{base}$ and $\\mathcal{C}_{novel}$ ( $\\mathcal{C}_{base} \\cap \\mathcal{C}_{novel} = \\emptyset$ ). The base dataset $\\mathcal{D}_{base}$ , which contains abundant labeled samples, is used to train the feature extractor. And the novel dataset $\\mathcal{D}_{novel}$ is then used to sample episodes for prototype estimation and few-shot evaluation. In a standard N-way K-shot task, each episode $\\mathcal{D}_{epi} = (\\mathcal{D}_S, \\mathcal{D}_Q)$ covers N classes from $\\mathcal{C}_{novel}$ . The support dataset $\\mathcal{D}_S$ , which contains K samples from each class, is used for class prototype estimation, and the query dataset $\\mathcal{D}_Q$ is then used for evaluation. This task aims at correctly classifying $\\mathcal{D}_Q$ into N classes from sampled episodes. And the main focus of this paper is to train a feature extractor with good generalization ability from $\\mathcal{D}_{base}$ .", + "bbox": [ + 496, + 391, + 893, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Preliminary: SSL with Knowledge Distillation", + "text_level": 1, + "bbox": [ + 498, + 618, + 890, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work is inspired by the self-supervised knowledge distillation frameworks proposed recently [4]. Specifically, given an input image $\\pmb{x}$ uniformly sampled from the training set $\\mathcal{I}$ , random data augmentations are applied to generate two augmented views $\\pmb{x}^1$ and $\\pmb{x}^2$ (we represent image view indices as superscripts and patch indices as subscripts), which are then fed into the teacher and student networks. The student network, parameterized by $\\theta_{s}$ , consists of an encoder with ViT backbone and a projection head with a 3-layer multi-layer perceptron (MLP) followed by $l_{2}$ -normalized bottleneck. The ViT backbone first generates a [cls] token, which is then entered into the projection head and outputs a probability distribution $P_{s}^{[\\mathrm{cls}]}$ over K dimensions. The teacher network, parameterized by $\\theta_{t}$ , is Exponentially Moving Averaged (EMA) updated by the student network $\\theta_{s}$ , which distills its knowledge to the student by minimizing the cross-entropy loss over the outputs of the", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19651", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2d4f90643870ab2108a670f8c23a0de7f96160d43079a1eaf2d2306d8148aea0.jpg", + "image_caption": [ + "Figure 3. Overview of our SMKD framework. Two views are generated from a pair of images sampled (with replacement) from the same class. The first view, applied with random blockwise masking, is passed to the student network, while the second unmasked view is passed to the teacher network. Both networks consist of a ViT backbone and a projection head. The parameters of the teacher network are Exponentially Moving Averaged (EMA) updated by the student network. SMKD distills knowledge between intra-class cross-views on both class and patch levels. $\\mathcal{L}_{[cls]}$ distills knowledge from [cls] tokens, while $\\mathcal{L}_{[patch]}$ distills knowledge from [patch] tokens by finding dense correspondence of matched token pairs (connected by red dashed lines) with highest similarities." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 883, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "categorical distributions from their projection heads:", + "bbox": [ + 76, + 363, + 424, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {[ \\mathrm {c l s} ]} = \\mathcal {H} (\\boldsymbol {P} _ {t} ^ {[ \\mathrm {c l s} ]} (\\boldsymbol {x} ^ {1}), \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\mathrm {c l s} ]} (\\boldsymbol {x} ^ {2})) \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 383, + 468, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{H}(x,y) = -x\\log y$", + "bbox": [ + 76, + 411, + 261, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked Image Modeling (MIM) [1, 57] can be performed via self-distillation as follows [77]. Given a randomly sampled mask sequence $m \\in \\{0,1\\}^N$ over an image with $N$ [patch] tokens $\\pmb{x} = \\{\\pmb{x}_i\\}_{i=1}^N$ , $\\pmb{x}_i$ 's with $m_i = 1$ are then replaced by a learnable token embedding $e_{[\\text{MASK}]}$ , which results in a corrupted image $\\widehat{\\pmb{x}} = \\{\\widehat{\\pmb{x}}_i\\}_{i=1}^N = \\{(1 - m_i)\\pmb{x}_i + m_i e_{[\\text{MASK}]}\\}_{i=1}^N$ . This corrupted image and the original uncorrupted image are fed into the student and teacher networks respectively. The objective of MIM is to recover the masked tokens from the corrupted image, which is equivalent to minimizing the cross-entropy loss between the outputs of the categorical distributions of the student and teacher networks on masked patches:", + "bbox": [ + 75, + 426, + 470, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M I M}} = \\sum_ {i = 1} ^ {N} m _ {i} \\cdot \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\text {p a t c h} ]} (\\boldsymbol {x}) _ {i}, \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\text {p a t c h} ]} (\\widehat {\\boldsymbol {x}}) _ {i}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 630, + 468, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Supervised Masked Knowledge Distillation", + "text_level": 1, + "bbox": [ + 76, + 681, + 442, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distill the Class Tokens. Recent self-distillation frameworks [4, 20, 77] distill knowledge on [cls] tokens from cross-view images via Eq.(1). To incorporate label information into such self-supervised frameworks, we further allow knowledge on [cls] tokens to be distilled from intra-class cross-views. This can be achieved by a small extension of the way we sample images. Rather than sampling a single image $\\boldsymbol{x} \\sim \\mathcal{I}$ and generating two views, now we sample two images $\\boldsymbol{x}, \\boldsymbol{y} \\sim \\mathcal{I}^c$ (with replacement) and generate one view for each of them1. $\\mathcal{I}^c \\subseteq \\mathcal{I}$ here denotes the set", + "bbox": [ + 75, + 705, + 468, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of images with the same class label $c$ in the training set $\\mathcal{I}$ . Specifically, we denote $\\pmb{x}^{\\prime}$ and $\\pmb{y}^{\\prime}$ as the augmented views generated from image $\\pmb{x}$ and $\\pmb{y}$ respectively. We apply additional random blockwise masking on $\\pmb{x}^{\\prime}$ and denote the resulting corrupted view as $\\widehat{\\pmb{x}}^{\\prime}$ . Then the corrupted view $\\widehat{\\pmb{x}}^{\\prime}$ and uncorrupted view $\\pmb{y}^{\\prime}$ are sent to the student and teacher networks separately. Now our supervised-contrastive loss on [cls] tokens becomes:", + "bbox": [ + 496, + 362, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {[ \\mathrm {c l s} ]} = \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\mathrm {c l s} ]} \\left(\\boldsymbol {y} ^ {\\prime}\\right), \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\mathrm {c l s} ]} \\left(\\hat {\\boldsymbol {x}} ^ {\\prime}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 493, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $\\pmb{x}, \\pmb{y}$ are sampled as the same image $(\\pmb{x} = \\pmb{y})$ , our loss performs masked self-distillation across two views of the same image similar to Eq.(1). In the other situation when $\\pmb{x}$ and $\\pmb{y}$ represent different images $(\\pmb{x} \\neq \\pmb{y})$ , our loss then minimizes the cross-entropy loss of projected [c1s] tokens among all cross-view pairs between images $\\pmb{x}$ and $\\pmb{y}$ .", + "bbox": [ + 496, + 523, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Such design has two major advantages. (1) It can be implemented efficiently. Instead of intentionally sampling image pairs from the same class, we just need to look through the images in a mini-batch, find image pairs belonging to the same class, and then apply our loss in Eq.(3). (2) Unlike previous works that use either supervised [42,76] or self-supervised [5, 30] contrastive losses, our method follows the recent trend in SSL works [4, 7, 20, 73] and avoids the need for negative examples.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distill the Patch Tokens. Beyond the knowledge distillation on the global [cls] tokens, we introduce the challenging task of masked patch tokens reconstruction across intra-class images, to fully exploit the local details of the image for training. Our main intuition here is based on the following hypothesis: for intra-class images, even though their semantic information can be vastly different at the patch level, there should at least exist some patches that share similar semantic meanings.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1For simplicity of illustration, we show our method with only one augmented view from each image. We sample two views from an image in our implementation. Loss is then averaged over all cross-view pairs.", + "bbox": [ + 75, + 862, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19652", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each patch $k$ from an input view $\\pmb{y}^{\\prime}$ sent to the teacher network (with its corresponding token embedding defined as $\\pmb{f}_{k}^{t}$ ), we need to first find the most similar patch $k^{+}$ from the masked view $\\widehat{\\pmb{x}}^{\\prime}$ of the student network (with its corresponding token embedding defined as $\\pmb{f}_{k^{+}}^{s}$ ), and then perform knowledge distillation between these two matched tokens. The token embedding $\\pmb{f}_{k^{+}}^{s}$ represents either the $k^{+th}$ [patch] token if patch $k^{+}$ is unmasked, or the reconstructed patch token if it is masked.", + "bbox": [ + 76, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we do not have any patch-level annotations, we use cosine similarity to find the best-matched patch of $k$ among all the [patch] tokens in the student network:", + "bbox": [ + 76, + 229, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nk ^ {+} = \\arg \\max _ {l \\in [ N ]} \\frac {\\boldsymbol {f} _ {k} ^ {t} {} ^ {\\top} \\boldsymbol {f} _ {l} ^ {s}}{\\| \\boldsymbol {f} _ {k} ^ {t} \\| \\| \\boldsymbol {f} _ {l} ^ {s} \\|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 281, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our patch-level knowledge distillation loss now becomes:", + "bbox": [ + 76, + 330, + 459, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {[ \\text {p a t c h} ]} = \\sum_ {k = 1} ^ {N} \\omega_ {k +} \\cdot \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\text {p a t c h} ]} \\left(\\boldsymbol {y} ^ {\\prime}\\right) _ {k}, \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\text {p a t c h} ]} \\left(\\hat {\\boldsymbol {x}} ^ {\\prime}\\right) _ {k +}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 351, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{k+}$ is a scalar representing the weight we give to each loss term. We find that taking it as a constant value is effective enough. The ablation study of more complex designs of $\\omega_{k+}$ is given in the appendix.", + "bbox": [ + 76, + 412, + 468, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our loss shares some similarities with DenseCL [65]. However, the differences are also obvious: (1) Our loss serves as an extension of their self-supervised contrastive loss into a supervised variant. (2) We further incorporate MIM into our design and allow masked patches to be matched, which makes our task harder and leads to more semantically meaningful patch embeddings.", + "bbox": [ + 75, + 473, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Pipeline", + "text_level": 1, + "bbox": [ + 76, + 588, + 246, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model in two stages: self-supervised pretraining and supervised training. In the first stage, we use the recently proposed MIM framework [77] for self-supervised pretraining. The self-supervised loss is the summation of $\\mathcal{L}_{[\\mathrm{cls}]}$ and $\\mathcal{L}_{\\mathrm{MIM}}$ in Eq.(1) and Eq.(2) without scaling. In the second stage, we continue training the pre-trained model using our supervised-contrastive losses $\\mathcal{L}_{[\\mathrm{cls}]}$ and $\\mathcal{L}_{[\\mathrm{patch}]}$ from Eq.(3) and Eq.(5). We define our training loss as: $\\mathcal{L} = \\mathcal{L}_{[\\mathrm{cls}]} + \\lambda \\mathcal{L}_{[\\mathrm{patch}]}$ , where $\\lambda$ controls the relative scale of these two components. A relatively large $\\lambda$ will make our model focus more on localization and less on high-level semantics.", + "bbox": [ + 75, + 611, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 806, + 200, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 832, + 179, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is evaluated on four widely used and publicly available few-shot classification datasets: mini-ImageNet [62], tiered-ImageNet [51], CIFAR-FS [2], and FC100", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[44]. mini-ImageNet and tiered-ImageNet are derived from ImageNet [11], while CIFAR-FS and FC100 are derived from CIFAR100 [34]. mini-ImageNet contains 100 classes, which are randomly split into 64 base classes for training, 16 classes for few-shot validation, and 20 classes for few-shot evaluation. There are 600 images in each class. tiered-ImageNet contains 609 classes with 779165 images in total. The class split for training, few-shot validation and few-shot evaluation are 351, 97, and 160. CIFAR-FS contains 100 classes with class split as 64, 16, and 20. FC100 contains 100 classes with class split as 60, 20, and 20. For both of these two datasets, each class has 600 images with smaller resolutions $(32 \\times 32)$ compared with ImageNet.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 300, + 718, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised pretraining. We pre-train our Vision Transformer backbone and projection head following the same pipeline in iBOT [77]. Most of the hyper-parameter settings are kept unchanged without tuning. We use a batch size of 640 and a learning rate of 0.0005 decayed with the cosine schedule. mini-[ImageNet and tiered-[ImageNet are pre-trained for 1200 epochs, and CIFAR-FS and FC100 are pre-trained for 900 epochs. All models are trained on 8 Nvidia RTX 3090 GPUs. Detailed training parameter settings are included in the appendix.", + "bbox": [ + 496, + 324, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised knowledge distillation After getting the pre-trained model, we train it with our supervised-contrastive losses. We find that the model with the best performance can converge within 60 epochs in the validation set. We use the same batch size and learning rate as in the pretraining stage. Compared with the first stage, the only extra hyper-parameter is the scaling parameter $\\lambda$ . We set it to make the ratio between $\\mathcal{L}_{[\\mathrm{patch}]}$ and $\\mathcal{L}_{[\\mathrm{cls}]}$ roughly around 2. Ablation over different $\\lambda$ is given in the appendix.", + "bbox": [ + 496, + 477, + 892, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Few-shot Evaluation We use the simple prototype classification method (Prototype) [44, 54] and the linear classifier (Classifier) used in S2M2 [43] as our default methods for few-shot evaluation. For each sampled episode data in an N-way K-shot task, Prototype first estimates each class prototype using the averaged feature over K support samples. Then a new sampled query image is classified into one of the N classes which has the highest cosine similarity between its feature vector and the class prototype. Instead, Classifier trains a linear classifier from the $\\mathrm{N}\\times \\mathrm{K}$ support samples, which is then used to classify new query samples. More complex evaluation methods (e.g. DeepEMD) are also compatible with our framework. The feature we used for evaluation is the concatenation of the [cls] token with the weighted average [patch] token (weighted avg pool). The weights of weighted avg pool is the average of the self-attention values of the [cls] token with all heads of the last attention layer. Ablation over different choices of features for evaluation is studied in Sec. 4.4.", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19653", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b8e7b2d7097df8ab7b8d4a500df1ac669e54f5f07521daa70b6de30f4ca4b43b.jpg", + "table_caption": [ + "Table 1. Results on mini-ImageNet and tiered-ImageNet. Top three methods are colored in red, blue and green respectively. A more comprehensive version of this table is shown in the appendix." + ], + "table_footnote": [], + "table_body": "
MethodBackbone#ParamsminiImageNet,5-waytieredImageNet,5-way
1-shot5-shot1-shot5-shot
DeepEMD [74]ResNet-1212.4M65.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
IE [52]ResNet-1212.4M67.28 ± 0.8084.78 ± 0.5272.21 ± 0.9087.08 ± 0.58
COSOC [41]ResNet-1212.4M69.28 ± 0.4985.16 ± 0.4273.57 ± 0.4387.57 ± 0.10
Meta-QDA [75]WRN-28-1036.5M67.38 ± 0.5584.27 ± 0.7574.29 ± 0.6689.41 ± 0.77
OM [47]WRN-28-1036.5M66.78 ± 0.3085.29 ± 0.4171.54 ± 0.2987.79 ± 0.46
SUN [14]ViT12.5M67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [32]Swin-Tiny29.0M72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
HCT Transformers [70]3× ViT-S63.0M74.74 ± 0.1789.19 ± 0.1379.67 ± 0.2091.72 ± 0.11
Ours (Prototype)ViT-S21M74.28 ± 0.1888.82 ± 0.0978.83 ± 0.2091.02 ± 0.12
Ours (Classifier)ViT-S21M74.10 ± 0.1788.89 ± 0.0978.81 ± 0.2191.21 ± 0.11
Ours + HCT [70]3× ViT-S63M75.32 ± 0.1889.57 ± 0.0979.74 ± 0.2091.68 ± 0.11
", + "bbox": [ + 107, + 119, + 867, + 316 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a35d486531feb7b13dcfa73dfce881ca5346d001545e5bf3c7d2b219c6a0ad4f.jpg", + "table_caption": [ + "Table 2. Results on CIFAR-FS and FC100. A more comprehensive version of this table is shown in the appendix." + ], + "table_footnote": [], + "table_body": "
MethodBackbone#ParamsCIFAR-FS,5-wayFC100,5-way
1-shot5-shot1-shot5-shot
BML [78]ResNet-1212.4M73.45 ± 0.4788.04 ± 0.3345.00 ± 0.4163.03 ± 0.41
IE [52]ResNet-1212.4M77.87 ± 0.8589.74 ± 0.5747.76 ± 0.7765.30 ± 0.76
TPMN [67]ResNet-1212.4M75.50 ± 0.9087.20 ± 0.6046.93 ± 0.7163.26 ± 0.74
PSST [9]WRN-28-1036.5M77.02 ± 0.3888.45 ± 0.35--
Meta-QDA [75]WRN-28-1036.5M75.95 ± 0.5988.72 ± 0.79--
SUN [14]ViT12.5M78.37 ± 0.4688.84 ± 0.32--
FewTURE [32]Swin-Tiny29.0M77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
HCT Transformers [70]3× ViT-S63.0M78.89 ± 0.1890.50 ± 0.0948.27 ± 0.1566.42 ± 0.16
Ours (Prototype)ViT-S21M80.08 ± 0.1890.63 ± 0.1350.38 ± 0.1668.37 ± 0.16
Ours (Classifier)ViT-S21M79.82 ± 0.1890.91 ± 0.1350.28 ± 0.1668.50 ± 0.16
", + "bbox": [ + 107, + 342, + 867, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison With the State-of-the-arts (SOTAs)", + "text_level": 1, + "bbox": [ + 76, + 539, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed SMKD with the above-mentioned evaluation methods on four few-shot classification datasets. Our goal in this section is to prove the effectiveness of our method given its simple training pipeline and evaluation procedure. This being said, the performance of our method can be further boosted by adopting strategies from contemporary works (see the last row \"Ours+HCT [70]\" in Table 1 and detailed results in Table 3).", + "bbox": [ + 75, + 566, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared with traditional convolutional backbones, Transformer-based models are still underdeveloped in few-shot classification. Recent methods with Transformer backbones [14, 32, 70] differ a lot in their training pipelines and evaluation procedures. SUN [14] first pre-trains a teacher network with supervised loss, then uses it to generate patch-level supervision for the student network as a supplement to class-level supervision. HCTransformers [70] converts class-level prediction into a latent prototype learning problem, and introduces spectral tokens pooling to merge neighboring tokens with similar semantic meanings adaptively. FewTURE [32], which adopts a similar two-stage training pipeline as ours, shows the effectiveness of inner loop token importance reweighting to avoid supervision collapse.", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite the success of these works, our SMKD still shows competitive performance given its simple design. Table 1 contains the results of mini-ImageNet and tiered-ImageNet. With Prototype and Classifier as the evaluation method, our proposed SMKD outperforms all methods with ResNet and WRN backbones, and ranked second among methods with Transformer backbones. By adopting the same strategies (patch size of 8 and spectral tokens pooling) as in HCTransformers [70], our method achieves a new SOTA on mini-ImageNet. The effect of each of these two strategies is shown in Table 3. Furthermore, as displayed in Table 2, our method performs the best on the two small-resolution datasets (CIFAR-FS and FC100), and outperforms all previous results by great margins (0.93% for 1-shot and 0.41% for 5-shot on CIFAR-FS, and over 2% for both 1 and 5-shot on FC100). In conclusion, our method grows more effective on datasets with smaller resolutions and fewer training images.", + "bbox": [ + 496, + 540, + 890, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The top competitor to our model is HCTransformer [70]. However, their method is more complex compared with ours in the following two aspects: (1) Their method uses a smaller patch size of 8 rather than 16 (which is used in most other methods as well as ours). As shown in Table", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19654", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ddc14a91c6e48db86f840bda0caac3c886b2b9bab4c0338c266907aa9c686147.jpg", + "table_caption": [ + "Table 3. Results for small patch size and spectral tokens pooling. Results are evaluated on mini-ImageNet with Prototype and Classifier methods, and the best is reported." + ], + "table_footnote": [], + "table_body": "
MethodPatch 8Spec. Pool1-shot5-shot
HCT [70]✓71.27 ± 0.1784.68 ± 0.10
HCT [70]✓✓74.62 ± 0.2089.19 ± 0.13
Ours74.28 ± 0.1888.82 ± 0.09
Ours✓74.00 ± 0.1789.57 ± 0.09
Ours✓✓75.32 ± 0.1889.23 ± 0.18
", + "bbox": [ + 81, + 133, + 472, + 228 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4, this makes their training procedure $4 \\times$ slower than ours in each epoch. (2) Their method contains three sets of cascaded transformers, each corresponding to a set of teacher-student network with ViT-S backbone. This triples the total number of learnable parameters (63M versus 21M in ours) and incurs much longer inference time.", + "bbox": [ + 75, + 246, + 468, + 335 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2eb5dab5addc20ac8c7aa5585bd534cc1b21206842152f3d87363fbb4032530.jpg", + "table_caption": [ + "Table 4. Training clock time comparison on mini-ImageNet. For HCT, we report the result provided in their paper. All experiments are conducted on 8 Nvidia RTX 3090 GPUs." + ], + "table_footnote": [], + "table_body": "
HCT [70]Stage1Stage2Stage3
21.1h (400 epoch)0.58h (2 epoch)0.21h (2 epoch)
OursSelf-Supervised Pre-trainSupervised Training
15.0h (1200 epoch)1.08h (60 epoch)
", + "bbox": [ + 81, + 393, + 468, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a fair comparison, we compare few-shot classification results of methods with the same ViT-S backbone (21M parameters) on mini-ImageNet in Table 5. The results for HCT [70] and FewTURE [32] are copy-pasted from their paper. Our model outperforms them by a large margin when the number of trainable parameters is comparable.", + "bbox": [ + 75, + 468, + 468, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualizations of the multi-head self-attention and patch token correspondence are shown in Fig. 4 and Fig. 5.", + "bbox": [ + 76, + 559, + 468, + 589 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2e8b772251b01846a0b5c5e1ab0e04ce9a46c7867954224b040dc3a69c96f50.jpg", + "table_caption": [ + "Table 5. Comparison results with the same ViT-S backbone on mini-ImageNet. For HCT, we report their result of the first student transformer (training time shown in Stage 1 of Table 4)." + ], + "table_footnote": [], + "table_body": "
MethodBackbone1-shot5-shot
HCT [70]ViT-S71.27 ± 0.1784.68 ± 0.10
FewTURE [32]ViT-S68.02 ± 0.8884.51 ± 0.53
OursViT-S74.28 ± 0.1888.82 ± 0.09
", + "bbox": [ + 84, + 646, + 465, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 741, + 228, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-Supervised MIM Pretraining Helps. Our training pipeline contains two stages. A natural question is whether the first self-supervised pretraining stage is necessary. We can see from Table 6 that the classification accuracy of our method drops a lot without pretraining. If we train in a single stage, with the combination of the self-supervised learning losses and supervised contrastive learning losses, the model is still hard to converge without a good initialization (1-shot: 43.71 vs. our 74.28. 5-shot: 60.27 vs. our", + "bbox": [ + 75, + 763, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c15894a23af3177d1e84c737d22eec888e62334b442a236058cad160f8a756f.jpg", + "image_caption": [ + "Figure 4. Visualization of multi-head self-attention maps. The self-attention of the [c1s] tokens with different heads in the last attention layer of ViT are visualized in different colors. iBOT+CE represents the model first pre-trained with iBOT, then trained with CE loss. Our SMKD pays more attention to the foreground objects, especially the most discriminative parts." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 883, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f784f34866ff047b8c767b7590e38ffe6fc32699d24df9e2c222566df7e32ec.jpg", + "image_caption": [ + "Figure 5. Visualization of dense correspondence. We use the patches with the highest self-attention of the [cls] token on each attention head (6 in total) of the last layer of ViT-S as queries. Best matched patches with highest similarities are connected with lines." + ], + "image_footnote": [], + "bbox": [ + 514, + 382, + 879, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "88.82 on mini-[ImageNet]. An explanation is that it will be hard for our patch-level knowledge distillation to find matched patches if the ViT backbone is initialized with random weights. What we observe is that our model finds some shortcut solution that prohibits it from proper training. Additionally, we also test the traditional CE loss with and without self-supervised pretraining. The model pre-trained with self-supervision and then trained with supervised CE loss works surprisingly well and even outperforms most of the methods in Table 1 on mini-[ImageNet]. This further supports the usefulness of our two-stage training procedure. For the masking strategy in the first pretraining stage, our ablation study over three masking types (blockwise, random, and no mask) in the appendix shows that blockwise masking balances the best between 1 and 5-shot classification accuracy.", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19655", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/48ae272fe4ddecde22882161463b91966c7cbafcf8e2c64d54106be1d9590171.jpg", + "table_caption": [ + "Table 6. Few-shot evaluation w/wo self-supervised pretraining. For models with pretraining stage, we pre-train with iBOT for 1200 epochs, then train with CE or our loss for 60 epochs. For models without pretraining stage, we train with supervised losses for 60 epochs directly. Results are evaluated by Prototype classification method on mini-ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodPre-train1-shot5-shot
CE48.98 ± 0.1667.62 ± 0.14
CE✓71.02 ± 0.1887.25 ± 0.10
Ours33.14 ± 0.1345.02 ± 0.14
Ours✓74.28 ± 0.1888.82 ± 0.09
", + "bbox": [ + 102, + 174, + 447, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52d1fbf0dd58ac661d4ecb1670dd2f22fb152e0bfd2cf8d4845af6eff859c17d.jpg", + "table_caption": [ + "Table 7. Ablation of different losses in the supervised training stage. All losses are trained for 60 epochs initialized with the pretrained iBOT model. The first row represents the pre-trained iBOT without additional supervised training. $\\mathcal{L}_{CE} + \\mathcal{L}_{[\\text{patch}]}$ stands for the combination of CE loss and our patch-level loss. Results are evaluated by Prototype method on mini-ImageNet." + ], + "table_footnote": [], + "table_body": "
Loss1-shot5-shot
-60.93 ± 0.1780.38 ± 0.12
LCE71.02 ± 0.1887.25 ± 0.10
L[cls]70.21 ± 0.1787.03 ± 0.10
L[patch]70.84 ± 0.1885.90 ± 0.11
LCE + L[patch]70.70 ± 0.1886.77 ± 0.10
L[cls] + L[patch]74.28 ± 0.1888.82 ± 0.09
", + "bbox": [ + 107, + 356, + 444, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class and Patch-Level Distillation Complement Each Other. Our SMKD uses the combination of class-level and patch-level supervised-contrastive losses as the final objective. Here we study several different combinations of supervised losses to confirm our proposed method performs the best. Results in Table 7 show that combining $\\mathcal{L}_{[\\mathrm{cls}]}$ with $\\mathcal{L}_{[\\mathrm{patch}]}$ achieves $13.35\\%$ improvement over baseline for 1-shot and $8.44\\%$ for 5-shot classification, which outperforms each of its components. Sharing the parameters of projection heads for class and patch tokens allows the semantics obtained from their distillation to complement each other, which leads to much better performance. This can also be seen from the comparison between $\\mathcal{L}_{CE} + \\mathcal{L}_{[\\mathrm{patch}]}$ and $\\mathcal{L}_{[\\mathrm{cls}]} + \\mathcal{L}_{[\\mathrm{patch}]}$ . $\\mathcal{L}_{CE}$ alone achieves stronger performance than $\\mathcal{L}_{[\\mathrm{cls}]}$ , but combining it with $\\mathcal{L}_{[\\mathrm{patch}]}$ does not provide additional benefits with separate classification heads.", + "bbox": [ + 75, + 489, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weighted Average Pooling Boosts Performance. Some recent works [60,69] find that using average pooling instead of the [cls] token during evaluation can encourage token-level tasks (e.g. localization, segmentation). Inspired by such findings, we evaluate our model with different tokens as well as their combinations. Results from Table 8 show that the performance of the [cls] token can be boosted by $3.02\\%$ for 1-shot and $0.79\\%$ for 5-shot by just concatenating with the weighted average pooling token. This shows", + "bbox": [ + 75, + 763, + 467, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that our proposed method can learn meaningful representations for both [cls] and [patch] tokens. Moreover, weighted average pooling and average pooling tokens share similar information, but the former performs better because it gives less weight to the background.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b345ba9ea13f92a23076607638ccfa6648b12792e4b82d5e1971d3d47ec99c08.jpg", + "table_caption": [ + "Table 8. Ablation of different tokens for few-shot evaluation. \"cls\" stands for the [cls] token, \"avg pool\" stands for using average pooling as token, \"weighted avg pool\" stands for weighted average [patch] token with weights from the self-attention module from the last block of ViT. The last four rows represent concatenation of these three tokens. All tokens are normalized to unit length. Results are evaluated by Prototype classification method." + ], + "table_footnote": [], + "table_body": "
Embedding1-shot5-shot
①cls71.26 ± 0.1788.03 ± 0.10
②avg pool71.65 ± 0.1986.40 ± 0.11
③weighted avg pool71.83 ± 0.1986.38 ± 0.11
① + ②73.83 ± 0.1888.58 ± 0.10
① + ③74.28 ± 0.1888.82 ± 0.09
② + ③71.86 ± 0.1986.47 ± 0.11
① + ② + ③73.80 ± 0.1888.26 ± 0.10
", + "bbox": [ + 517, + 273, + 879, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 415, + 617, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel supervised knowledge distillation framework (SMKD) for few-shot Transformers, which extends the self-supervised masked knowledge distillation framework into the supervised setting. With our design of supervised-contrastive losses, we incorporate supervision into both class and patch-level knowledge distillation while still enjoy the benefits of not needing large batch size and negative samples. Evaluation results together with ablation studies demonstrate the superiority of our method given its simple design compared with contemporary works. Our two-stage training is a special case of curriculum learning from the easy samples to the hard ones. We can unify the learning objectives of self-supervised learning and supervised contrastive learning, using a carefully designed curriculum learning strategy for future works. We hope our work can bridge the gap between self-supervised and supervised knowledge distillation, and inspire more works on supervised few-shot learning methods.", + "bbox": [ + 496, + 441, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 727, + 686, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is sponsored by Air Force Research Laboratory (AFRL) under agreement number FA8750-19-1-1000. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation therein. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of Air Force Laboratory, DARPA or the U.S. Government.", + "bbox": [ + 496, + 751, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19656", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3,4", + "[2] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4", + "[6] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 3", + "[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4", + "[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2", + "[9] Zhengyu Chen, Jixie Ge, Heshen Zhan, Siteng Huang, and Donglin Wang. Pareto self-supervised training for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13663-13672, 2021. 6", + "[10] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 1", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[13] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. Advances in Neural Information Processing Systems, 33:21981-21993, 2020. 3", + "[14] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3", + "[17] Yutong Feng, Jianwen Jiang, Mingqian Tang, Rong Jin, and Yue Gao. Rethinking supervised pre-training for better downstream transferring. arXiv preprint arXiv:2110.06014, 2021.3", + "[18] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? arXiv preprint arXiv:2210.07240, 2022. 2, 3", + "[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8059-8068, 2019. 1", + "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4", + "[21] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 3", + "[22] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3263-3272, October 2021. 3", + "[23] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 780-789, 2022. 3", + "[24] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, June 2022. 3", + "[25] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, Rama Chellappa, and Shih-Fu Chang. Multimodal few-shot object detection with meta-learning based cross-modal prompting. arXiv preprint arXiv:2204.07841, 2022. 3", + "[26] Guangxing Han, Xuan Zhang, and Chongrong Li. Revisiting faster r-cnn: A deeper look at region proposal network. In International Conference on Neural Information Processing, pages 14-24, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19657", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Guangxing Han, Xuan Zhang, and Chongrong Li. Single shot object detection with top-down refinement. In 2017 IEEE International Conference on Image Processing (ICIP), pages 3360-3364. IEEE, 2017. 1", + "[28] Guangxing Han, Xuan Zhang, and Chongrong Li. Semi-supervised diff: Decoupling detection and feature flow for video object detectors. In Proceedings of the 26th ACM international conference on Multimedia, pages 1811-1819, 2018. 1", + "[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3", + "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4", + "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[32] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. arXiv preprint arXiv:2206.07267, 2022. 2, 3, 6, 7", + "[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2", + "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[35] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1", + "[36] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 1", + "[37] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34:23818-23830, 2021. 1, 2, 3", + "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 3", + "[39] Jiang Lu, Pinghua Gong, Jieping Ye, and Changshui Zhang. Learning from very few samples: A survey. arXiv preprint arXiv:2009.02653, 2020. 1", + "[40] Yuning Lu, Liangjian Wen, Jianzhuang Liu, Yajing Liu, and Xinmei Tian. Self-supervision can be a good few-shot learner. arXiv preprint arXiv:2207.09176, 2022. 2", + "[41] Xu Luo, Longhui Wei, Liangjian Wen, Jinrong Yang, Lingxi Xie, Zenglin Xu, and Qi Tian. Rectifying the shortcut learning of background: Shared object concentration for few-shot" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image recognition. arXiv preprint arXiv:2107.07746, 2021. 6", + "[42] Jiawei Ma, Hanchen Xie, Guangxing Han, Shih-Fu Chang, Aram Galstyan, and Wael Abd-Almageed. Partner-assisted learning for few-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10573-10582, 2021. 1, 3, 4", + "[43] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2218-2227, 2020. 2, 3, 5", + "[44] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5", + "[45] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. arXiv preprint arXiv:2210.10615, 2022. 2, 3", + "[46] Elia Peruzzo, Enver Sangineto, Yahui Liu, Marco De Nadai, Wei Bi, Bruno Lepri, and Nicu Sebe. Spatial entropy regularization for vision transformers. arXiv preprint arXiv:2206.04636, 2022. 2, 3", + "[47] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8412-8422, 2021. 6", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3", + "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1", + "[51] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5", + "[52] Mamshad Nayeem Rizve, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Exploring complementary strengths of invariant and equivariant representations for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10836-10846, June 2021. 6", + "[53] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 3" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19658", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5", + "[55] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1", + "[56] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 3", + "[57] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 4", + "[58] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282. Springer, 2020. 3", + "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 3", + "[60] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 8", + "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[62] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 1, 3, 5", + "[63] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 3", + "[64] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3", + "[65] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 5", + "[66] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Jiamin Wu, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Task-aware part mining network for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8433-8442, 2021. 6", + "[68] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2", + "[69] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 8", + "[70] Dongyang Zhao Hong-Yu Zhou Weifeng Ge Yizhou Yu Wenqiang Zhang Yangji He, Weihan Liang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning, 2022. 1, 2, 3, 6, 7", + "[71] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8808-8817, 2020. 3", + "[72] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The met dataset: Instance-level recognition for artworks. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 3", + "[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 4", + "[74] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 6", + "[75] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 6", + "[76] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? arXiv preprint arXiv:2006.06606, 2020. 4", + "[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2, 3, 4, 5", + "[78] Ziqi Zhou, Xi Qiu, Jiangtao Xie, Jianan Wu, and Chi Zhang. Binocular mutual learning for improving few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8402-8411, 2021. 6", + "[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1" + ], + "bbox": [ + 501, + 92, + 890, + 880 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19659", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_model.json b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c9f1a731ffde623641f2ff6305e540438d37efee --- /dev/null +++ b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.843, + 0.154 + ], + "angle": 0, + "content": "Supervised Masked Knowledge Distillation for Few-Shot Transformers" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.181, + 0.834, + 0.224 + ], + "angle": 0, + "content": "Han Lin*, Guangxing Han\\*†, Jiawei Ma, Shiyuan Huang, Xudong Lin, Shih-Fu Chang Columbia University" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.228, + 0.761, + 0.243 + ], + "angle": 0, + "content": "{h13199, gh2561, jiawei.m, sh3813, x12798, sc250}@columbia.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.279, + 0.314, + 0.294 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.311, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) emerge to achieve impressive performance on many data-abundant computer vision tasks by capturing long-range dependencies among local features. However, under few-shot learning (FSL) settings on small datasets with only a few labeled data, ViT tends to overfit and suffers from severe performance degradation due to its absence of CNN-alike inductive bias. Previous works in FSL avoid such problem either through the help of self-supervised auxiliary losses, or through the dextile uses of label information under supervised settings. But the gap between self-supervised and supervised few-shot Transformers is still unfilled. Inspired by recent advances in self-supervised knowledge distillation and masked image modeling (MIM), we propose a novel Supervised Masked Knowledge Distillation model (SMKD) for few-shot Transformers which incorporates label information into self-distillation frameworks. Compared with previous self-supervised methods, we allow intra-class knowledge distillation on both class and patch tokens, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. Experimental results on four few-shot classification benchmark datasets show that our method with simple design outperforms previous methods by a large margin and achieves a new start-of-the-art. Detailed ablation studies confirm the effectiveness of each component of our model. Code for this paper is available here: https://github.com/HL-hanlin/SMKD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.73, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) [16] have emerged as a competitive alternative to Convolutional Neural Networks (CNNs) [31] in recent years, and have achieved impressive performance in many vision tasks including image classification [16, 38, 59, 66], object detection [3, 10, 26-28, 79], and object segmentation [50, 55]. Compared with CNNs, which introduce inductive bias through convolutional kernels with fixed receptive fields [35], the attention layers in" + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.277, + 0.78, + 0.288 + ], + "angle": 0, + "content": "(a) Few-Shot Transformers with Soft Label Regularization" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.291, + 0.864, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.356, + 0.856, + 0.368 + ], + "angle": 0, + "content": "(b) Few-Shot Transformers with Self-Supervision Auxiliary Regularization" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.369, + 0.88, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.44, + 0.885, + 0.451 + ], + "angle": 0, + "content": "(c) Few-Shot Transformers with Our Supervised Masked Knowledge Distillation" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.456, + 0.888, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.616 + ], + "angle": 0, + "content": "Figure 1. Comparison of the proposed idea and other existing methods for few-shot Transformers. Our model mitigates the overfitting issue of few-shot Transformers, by extending the masked knowledge distillation framework into the supervised setting, and enforcing the alignment of [cls] and corresponding [patch] tokens for intra-class images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.809 + ], + "angle": 0, + "content": "ViT allow it to model global token relations to capture long-range token dependencies. However, such flexibility also comes at a cost: ViT is data-hungry and it needs to learn token dependencies purely from data. This property often makes it easy to overfit to datasets with small training set and suffer from severe generalization performance degradation [36, 37]. Therefore, we are motivated to study how to make ViTs generalize well on these small datasets, especially under the few-shot learning (FSL) setting [19, 39, 62] which aims to recognize unseen new instances at test time just from only a few (e.g. one or five) labeled samples from each new categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Most of the existing methods mitigate the overfitting issue of few-shot Transformers [14] using various regularizations. For instance, some works utilize label information in a weaker [70], softer [42] way, or use label information efficiently through patch-level supervision [14]. However, these models usually design sophisticated learning targets." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.334, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.38 + ], + "angle": 0, + "content": "On the other hand, self-distillation techniques [4, 8], and particularly, the recent masked self-distillation [29, 45, 77], which distills knowledge learned from an uncorrupted image to the knowledge predicted from a masked image, have lead an emerging trend in self-supervised Transformers in various fields [15, 68]. Inspired by such success, recent works in FSL attempt to incorporate self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46], or to adopt a self-supervised pretraining, supervised training two-stage framework to train few-shot Transformers [18, 32]. Compared with traditional supervised methods, self-supervision can learn less biased representations towards base class, which usually leads to better generalization ability for novel classes [40]. However, the two learning objectives of self-supervision and supervision are conflicting and it is hard to balance them during training. Therefore, how to efficiently leverage the strengths of self-supervised learning to alleviate the overfitting issue of supervised training remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.761 + ], + "angle": 0, + "content": "In this work, we propose a novel supervised masked knowledge distillation framework (SMKD) for few-shot Transformers, which handles the aforementioned challenge through a natural extension of the self-supervised masked knowledge distillation framework into the supervised setting (shown in Fig. 1). Different from supervised contrastive learning [33] which only utilizes global image features for training, we leverage multi-scale information from the images (both global [cls] token and local [patch] tokens) to formulate the learning objectives, which has been demonstrated to be effective in the recent self-supervised Transformer methods [29, 77]. For global [cls] tokens, we can simply maximize the similarity for intra-class images. However, it is non-trivial and challenging to formulate the learning objectives for local [patch] tokens because we do not have ground-truth patch-level annotations. To address this problem, we propose to estimate the similarity between [patch] tokens across intra-class images using cross-attention, and enforce the alignment of corresponding [patch] tokens. Particularly, reconstructing masked [patch] tokens across intra-class images increases the difficulty of model learning, thus encouraging learning generalizable few-shot Transformer models by jointly exploiting the holistic knowledge of the image and the similarity of intra-class images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 2, we compare our model with the existing self-supervised/supervised learning methods. Our model is a natural extension of the supervised contrastive learning method [33] and self-supervised knowledge distillation methods [4, 77]. Thus our model inherits both the advantage of method [33] for effectively leveraging label information, and the advantages of methods [4, 77] for not needing large batch size and negative samples. Meanwhile, the newly-introduced challenging task of masked" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.681, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.089, + 0.877, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.19, + 0.681, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.19, + 0.887, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Comparison of other self-supervised/supervised frameworks. Our method (d) is a natural extension of (b) and (c), with the newly-introduced challenging task of masked [patch] tokens reconstruction across intra-class images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[patch] tokens reconstruction across intra-class images makes our method more powerful for learning generalizable few-shot Transformer models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.557 + ], + "angle": 0, + "content": "Compared with contemporary works on few-shot Transformers [14, 32, 70], our framework enjoys several good properties from a practical point of view. (1) Our method does not introduce any additional learnable parameters besides the ViT backbone and projection head, which makes it easy to be combined with other methods [32, 70, 74]. (2) Our method is both effective and training-efficient, with stronger performance and less training time on four few-shot classification benchmarks, compared with [32, 70]. In a nutshell, our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.561, + 0.892, + 0.637 + ], + "angle": 0, + "content": "- We propose a new supervised knowledge distillation framework (SMKD) that incorporates class label information into self-distillation, thus filling the gap between self-supervised knowledge distillation and traditional supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.637, + 0.892, + 0.697 + ], + "angle": 0, + "content": "- Within the proposed framework, we design two supervised-contrastive losses on both class and patch levels, and introduce the challenging task of masked patch tokens reconstruction across intra-class images." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.697, + 0.892, + 0.787 + ], + "angle": 0, + "content": "- Given its simple design, we test our SMKD on four few-shot datasets, and show that it achieves a new SOTA on CIFAR-FS and FC100 by a large margin, as well as competitive performance on mini-ImageNet and tiered-ImageNet using the simple prototype classification method for few-shot evaluation." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.561, + 0.892, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.642, + 0.816 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Few-shot Learning. Few-shot learning aims at fast knowledge transfer from seen base classes to unseen novel classes given only a few labeled samples from each novel class. The meta-learning paradigm, which simulates few-shot tasks episodically to mimic the human-learning process in the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19650" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "real world, once dominated FSL [22-25, 44, 53, 54, 56, 62]. Surprisingly, recent works [6, 21, 58, 72] have shown that the state-of-the-art meta-learning methods can be outperformed by simple baseline methods using just a distance-based classifier without the complicated design of meta-learning strategies. Therefore, recent methods [17, 42, 70] in FSL start to focus less on meta-learning, and more on learning embeddings with good generalization ability. Our paper follows this trend and proposes a knowledge distillation framework to learn generalizable embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.472 + ], + "angle": 0, + "content": "Vision Transformers in FSL. Variants of Transformer [61] have achieved great success in NLP [12, 49], computer vision [16, 59], and multimodal learning [48]. However, the lack of inductive bias makes Transformer infamous for its data-hungry property, which is especially significant in few-shot settings when the amount of data is limited. A line of works focuses on introducing inductive bias back to the Transformer architecture, including methods using pyramid structure [63, 64], shifted windows [38], and explicit convolutional token embeddings [66]. With this said, some recent works [13, 24, 32, 71] still show that few-shot Transformers have the potential of fast adaptation to novel classes. Our work also studies few-shot Transformers, and we show that our method can work well even on the vanilla ViT structure without explicit incorporation of inductive bias." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Self-Supervision for FSL. Self-supervised learning (SSL) has shown great potential in FSL due to its good generalization ability to novel classes. Previous methods incorporate SSL into FSL in various ways. Some works propose to include self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46]. For example, [37] proposed a regularization loss that extracts additional information from images by predicting the geometric distance between patch tokens. [46] designed their loss by discouraging spatially disordered attention maps based on the idea that objects usually occupy connected regions, and [43] derived their auxiliary loss from self-supervision tasks of rotation and exemplar. Some other works [18, 32], instead, adopt a two-stage procedure by pretraining a model via self-supervision before supervised training. [32] takes advantage of self-supervised training with iBOT [77] as a pretext task, and then uses inner loop token importance reweighting for supervised fine-tuning. [18] initializes its model with a pre-trained self-supervised DINO [4], then trained with supervised cross-entropy loss. Our work follows the second branch of work, but the difference is obvious from previous works. Instead of designing complicated training pipeline or leveraging additional learnable modules at inference time, we use supervised training on self-supervised pre-trained model, and focus mainly on bridging the gap between self-supervised and supervised knowledge distillation with minimum extra design." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Another emerging trend in SSL is Masked Image Mod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "eling (MIM) [1, 29, 45, 77], which aims at recovering the patch-level target (e.g. image pixels, patch features) of the masked content in a corrupted input image. In iBOT [77], the class and patch tokens share the same projection head, thus helping the ViT backbone to learn semantically meaningful patch embeddings, which can be used as a supplement to image-level global self-supervision." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.217, + 0.65, + 0.234 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.895, + 0.352 + ], + "angle": 0, + "content": "We present our Supervised Masked Knowledge Distillation (SMKD) framework in this section. Fig. 3 shows an overview of our model. Task definition for few-shot classification is presented in Sec. 3.1. We explain the formulation of SSL knowledge distillation in Sec. 3.2, and how we extend it into our supervised knowledge distillation framework in Sec. 3.3. Training pipeline is given in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.366, + 0.73, + 0.383 + ], + "angle": 0, + "content": "3.1. Learning Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.604 + ], + "angle": 0, + "content": "In few-shot classification, we are given two datasets \\(\\mathcal{D}_{base}\\) and \\(\\mathcal{D}_{novel}\\) with disjoint class labels \\(\\mathcal{C}_{base}\\) and \\(\\mathcal{C}_{novel}\\) (\\(\\mathcal{C}_{base} \\cap \\mathcal{C}_{novel} = \\emptyset\\)). The base dataset \\(\\mathcal{D}_{base}\\), which contains abundant labeled samples, is used to train the feature extractor. And the novel dataset \\(\\mathcal{D}_{novel}\\) is then used to sample episodes for prototype estimation and few-shot evaluation. In a standard N-way K-shot task, each episode \\(\\mathcal{D}_{epi} = (\\mathcal{D}_S, \\mathcal{D}_Q)\\) covers N classes from \\(\\mathcal{C}_{novel}\\). The support dataset \\(\\mathcal{D}_S\\), which contains K samples from each class, is used for class prototype estimation, and the query dataset \\(\\mathcal{D}_Q\\) is then used for evaluation. This task aims at correctly classifying \\(\\mathcal{D}_Q\\) into N classes from sampled episodes. And the main focus of this paper is to train a feature extractor with good generalization ability from \\(\\mathcal{D}_{base}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.619, + 0.892, + 0.636 + ], + "angle": 0, + "content": "3.2. Preliminary: SSL with Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our work is inspired by the self-supervised knowledge distillation frameworks proposed recently [4]. Specifically, given an input image \\(\\pmb{x}\\) uniformly sampled from the training set \\(\\mathcal{I}\\), random data augmentations are applied to generate two augmented views \\(\\pmb{x}^1\\) and \\(\\pmb{x}^2\\) (we represent image view indices as superscripts and patch indices as subscripts), which are then fed into the teacher and student networks. The student network, parameterized by \\(\\theta_{s}\\), consists of an encoder with ViT backbone and a projection head with a 3-layer multi-layer perceptron (MLP) followed by \\(l_{2}\\)-normalized bottleneck. The ViT backbone first generates a [cls] token, which is then entered into the projection head and outputs a probability distribution \\(P_{s}^{[\\mathrm{cls}]}\\) over K dimensions. The teacher network, parameterized by \\(\\theta_{t}\\), is Exponentially Moving Averaged (EMA) updated by the student network \\(\\theta_{s}\\), which distills its knowledge to the student by minimizing the cross-entropy loss over the outputs of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19651" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.885, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 3. Overview of our SMKD framework. Two views are generated from a pair of images sampled (with replacement) from the same class. The first view, applied with random blockwise masking, is passed to the student network, while the second unmasked view is passed to the teacher network. Both networks consist of a ViT backbone and a projection head. The parameters of the teacher network are Exponentially Moving Averaged (EMA) updated by the student network. SMKD distills knowledge between intra-class cross-views on both class and patch levels. \\(\\mathcal{L}_{[cls]}\\) distills knowledge from [cls] tokens, while \\(\\mathcal{L}_{[patch]}\\) distills knowledge from [patch] tokens by finding dense correspondence of matched token pairs (connected by red dashed lines) with highest similarities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.364, + 0.426, + 0.38 + ], + "angle": 0, + "content": "categorical distributions from their projection heads:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.385, + 0.47, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {[ \\mathrm {c l s} ]} = \\mathcal {H} (\\boldsymbol {P} _ {t} ^ {[ \\mathrm {c l s} ]} (\\boldsymbol {x} ^ {1}), \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\mathrm {c l s} ]} (\\boldsymbol {x} ^ {2})) \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.262, + 0.427 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}(x,y) = -x\\log y\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.471, + 0.624 + ], + "angle": 0, + "content": "Masked Image Modeling (MIM) [1, 57] can be performed via self-distillation as follows [77]. Given a randomly sampled mask sequence \\( m \\in \\{0,1\\}^N \\) over an image with \\( N \\) [patch] tokens \\( \\pmb{x} = \\{\\pmb{x}_i\\}_{i=1}^N \\), \\( \\pmb{x}_i \\)'s with \\( m_i = 1 \\) are then replaced by a learnable token embedding \\( e_{[\\text{MASK}]} \\), which results in a corrupted image \\( \\widehat{\\pmb{x}} = \\{\\widehat{\\pmb{x}}_i\\}_{i=1}^N = \\{(1 - m_i)\\pmb{x}_i + m_i e_{[\\text{MASK}]}\\}_{i=1}^N \\). This corrupted image and the original uncorrupted image are fed into the student and teacher networks respectively. The objective of MIM is to recover the masked tokens from the corrupted image, which is equivalent to minimizing the cross-entropy loss between the outputs of the categorical distributions of the student and teacher networks on masked patches:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.631, + 0.47, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M I M}} = \\sum_ {i = 1} ^ {N} m _ {i} \\cdot \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\text {p a t c h} ]} (\\boldsymbol {x}) _ {i}, \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\text {p a t c h} ]} (\\widehat {\\boldsymbol {x}}) _ {i}\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.444, + 0.699 + ], + "angle": 0, + "content": "3.3. Supervised Masked Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.47, + 0.857 + ], + "angle": 0, + "content": "Distill the Class Tokens. Recent self-distillation frameworks [4, 20, 77] distill knowledge on [cls] tokens from cross-view images via Eq.(1). To incorporate label information into such self-supervised frameworks, we further allow knowledge on [cls] tokens to be distilled from intra-class cross-views. This can be achieved by a small extension of the way we sample images. Rather than sampling a single image \\( \\boldsymbol{x} \\sim \\mathcal{I} \\) and generating two views, now we sample two images \\( \\boldsymbol{x}, \\boldsymbol{y} \\sim \\mathcal{I}^c \\) (with replacement) and generate one view for each of them1. \\( \\mathcal{I}^c \\subseteq \\mathcal{I} \\) here denotes the set" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1For simplicity of illustration, we show our method with only one augmented view from each image. We sample two views from an image in our implementation. Loss is then averaged over all cross-view pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.485 + ], + "angle": 0, + "content": "of images with the same class label \\(c\\) in the training set \\(\\mathcal{I}\\). Specifically, we denote \\(\\pmb{x}^{\\prime}\\) and \\(\\pmb{y}^{\\prime}\\) as the augmented views generated from image \\(\\pmb{x}\\) and \\(\\pmb{y}\\) respectively. We apply additional random blockwise masking on \\(\\pmb{x}^{\\prime}\\) and denote the resulting corrupted view as \\(\\widehat{\\pmb{x}}^{\\prime}\\). Then the corrupted view \\(\\widehat{\\pmb{x}}^{\\prime}\\) and uncorrupted view \\(\\pmb{y}^{\\prime}\\) are sent to the student and teacher networks separately. Now our supervised-contrastive loss on [cls] tokens becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.494, + 0.892, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {[ \\mathrm {c l s} ]} = \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\mathrm {c l s} ]} \\left(\\boldsymbol {y} ^ {\\prime}\\right), \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\mathrm {c l s} ]} \\left(\\hat {\\boldsymbol {x}} ^ {\\prime}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.614 + ], + "angle": 0, + "content": "When \\( \\pmb{x}, \\pmb{y} \\) are sampled as the same image \\( (\\pmb{x} = \\pmb{y}) \\), our loss performs masked self-distillation across two views of the same image similar to Eq.(1). In the other situation when \\( \\pmb{x} \\) and \\( \\pmb{y} \\) represent different images \\( (\\pmb{x} \\neq \\pmb{y}) \\), our loss then minimizes the cross-entropy loss of projected [c1s] tokens among all cross-view pairs between images \\( \\pmb{x} \\) and \\( \\pmb{y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Such design has two major advantages. (1) It can be implemented efficiently. Instead of intentionally sampling image pairs from the same class, we just need to look through the images in a mini-batch, find image pairs belonging to the same class, and then apply our loss in Eq.(3). (2) Unlike previous works that use either supervised [42,76] or self-supervised [5, 30] contrastive losses, our method follows the recent trend in SSL works [4, 7, 20, 73] and avoids the need for negative examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Distill the Patch Tokens. Beyond the knowledge distillation on the global [cls] tokens, we introduce the challenging task of masked patch tokens reconstruction across intra-class images, to fully exploit the local details of the image for training. Our main intuition here is based on the following hypothesis: for intra-class images, even though their semantic information can be vastly different at the patch level, there should at least exist some patches that share similar semantic meanings." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "For each patch \\(k\\) from an input view \\(\\pmb{y}^{\\prime}\\) sent to the teacher network (with its corresponding token embedding defined as \\(\\pmb{f}_{k}^{t}\\)), we need to first find the most similar patch \\(k^{+}\\) from the masked view \\(\\widehat{\\pmb{x}}^{\\prime}\\) of the student network (with its corresponding token embedding defined as \\(\\pmb{f}_{k^{+}}^{s}\\)), and then perform knowledge distillation between these two matched tokens. The token embedding \\(\\pmb{f}_{k^{+}}^{s}\\) represents either the \\(k^{+th}\\) [patch] token if patch \\(k^{+}\\) is unmasked, or the reconstructed patch token if it is masked." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.23, + 0.47, + 0.275 + ], + "angle": 0, + "content": "As we do not have any patch-level annotations, we use cosine similarity to find the best-matched patch of \\( k \\) among all the [patch] tokens in the student network:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.282, + 0.47, + 0.322 + ], + "angle": 0, + "content": "\\[\nk ^ {+} = \\arg \\max _ {l \\in [ N ]} \\frac {\\boldsymbol {f} _ {k} ^ {t} {} ^ {\\top} \\boldsymbol {f} _ {l} ^ {s}}{\\| \\boldsymbol {f} _ {k} ^ {t} \\| \\| \\boldsymbol {f} _ {l} ^ {s} \\|} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.331, + 0.46, + 0.347 + ], + "angle": 0, + "content": "Our patch-level knowledge distillation loss now becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.352, + 0.469, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {[ \\text {p a t c h} ]} = \\sum_ {k = 1} ^ {N} \\omega_ {k +} \\cdot \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\text {p a t c h} ]} \\left(\\boldsymbol {y} ^ {\\prime}\\right) _ {k}, \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\text {p a t c h} ]} \\left(\\hat {\\boldsymbol {x}} ^ {\\prime}\\right) _ {k +}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.469, + 0.473 + ], + "angle": 0, + "content": "where \\(\\omega_{k+}\\) is a scalar representing the weight we give to each loss term. We find that taking it as a constant value is effective enough. The ablation study of more complex designs of \\(\\omega_{k+}\\) is given in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.58 + ], + "angle": 0, + "content": "Our loss shares some similarities with DenseCL [65]. However, the differences are also obvious: (1) Our loss serves as an extension of their self-supervised contrastive loss into a supervised variant. (2) We further incorporate MIM into our design and allow masked patches to be matched, which makes our task harder and leads to more semantically meaningful patch embeddings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.248, + 0.606 + ], + "angle": 0, + "content": "3.4. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We train our model in two stages: self-supervised pretraining and supervised training. In the first stage, we use the recently proposed MIM framework [77] for self-supervised pretraining. The self-supervised loss is the summation of \\(\\mathcal{L}_{[\\mathrm{cls}]}\\) and \\(\\mathcal{L}_{\\mathrm{MIM}}\\) in Eq.(1) and Eq.(2) without scaling. In the second stage, we continue training the pre-trained model using our supervised-contrastive losses \\(\\mathcal{L}_{[\\mathrm{cls}]}\\) and \\(\\mathcal{L}_{[\\mathrm{patch}]}\\) from Eq.(3) and Eq.(5). We define our training loss as: \\(\\mathcal{L} = \\mathcal{L}_{[\\mathrm{cls}]} + \\lambda \\mathcal{L}_{[\\mathrm{patch}]}\\), where \\(\\lambda\\) controls the relative scale of these two components. A relatively large \\(\\lambda\\) will make our model focus more on localization and less on high-level semantics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.808, + 0.202, + 0.825 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.18, + 0.847 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our model is evaluated on four widely used and publicly available few-shot classification datasets: mini-ImageNet [62], tiered-ImageNet [51], CIFAR-FS [2], and FC100" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[44]. mini-ImageNet and tiered-ImageNet are derived from ImageNet [11], while CIFAR-FS and FC100 are derived from CIFAR100 [34]. mini-ImageNet contains 100 classes, which are randomly split into 64 base classes for training, 16 classes for few-shot validation, and 20 classes for few-shot evaluation. There are 600 images in each class. tiered-ImageNet contains 609 classes with 779165 images in total. The class split for training, few-shot validation and few-shot evaluation are 351, 97, and 160. CIFAR-FS contains 100 classes with class split as 64, 16, and 20. FC100 contains 100 classes with class split as 60, 20, and 20. For both of these two datasets, each class has 600 images with smaller resolutions \\((32 \\times 32)\\) compared with ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.719, + 0.317 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Self-supervised pretraining. We pre-train our Vision Transformer backbone and projection head following the same pipeline in iBOT [77]. Most of the hyper-parameter settings are kept unchanged without tuning. We use a batch size of 640 and a learning rate of 0.0005 decayed with the cosine schedule. mini-[ImageNet and tiered-[ImageNet are pre-trained for 1200 epochs, and CIFAR-FS and FC100 are pre-trained for 900 epochs. All models are trained on 8 Nvidia RTX 3090 GPUs. Detailed training parameter settings are included in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.613 + ], + "angle": 0, + "content": "Supervised knowledge distillation After getting the pre-trained model, we train it with our supervised-contrastive losses. We find that the model with the best performance can converge within 60 epochs in the validation set. We use the same batch size and learning rate as in the pretraining stage. Compared with the first stage, the only extra hyper-parameter is the scaling parameter \\(\\lambda\\). We set it to make the ratio between \\(\\mathcal{L}_{[\\mathrm{patch}]}\\) and \\(\\mathcal{L}_{[\\mathrm{cls}]}\\) roughly around 2. Ablation over different \\(\\lambda\\) is given in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Few-shot Evaluation We use the simple prototype classification method (Prototype) [44, 54] and the linear classifier (Classifier) used in S2M2 [43] as our default methods for few-shot evaluation. For each sampled episode data in an N-way K-shot task, Prototype first estimates each class prototype using the averaged feature over K support samples. Then a new sampled query image is classified into one of the N classes which has the highest cosine similarity between its feature vector and the class prototype. Instead, Classifier trains a linear classifier from the \\(\\mathrm{N}\\times \\mathrm{K}\\) support samples, which is then used to classify new query samples. More complex evaluation methods (e.g. DeepEMD) are also compatible with our framework. The feature we used for evaluation is the concatenation of the [cls] token with the weighted average [patch] token (weighted avg pool). The weights of weighted avg pool is the average of the self-attention values of the [cls] token with all heads of the last attention layer. Ablation over different choices of features for evaluation is studied in Sec. 4.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19653" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Results on mini-ImageNet and tiered-ImageNet. Top three methods are colored in red, blue and green respectively. A more comprehensive version of this table is shown in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.12, + 0.869, + 0.317 + ], + "angle": 0, + "content": "
MethodBackbone#ParamsminiImageNet,5-waytieredImageNet,5-way
1-shot5-shot1-shot5-shot
DeepEMD [74]ResNet-1212.4M65.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
IE [52]ResNet-1212.4M67.28 ± 0.8084.78 ± 0.5272.21 ± 0.9087.08 ± 0.58
COSOC [41]ResNet-1212.4M69.28 ± 0.4985.16 ± 0.4273.57 ± 0.4387.57 ± 0.10
Meta-QDA [75]WRN-28-1036.5M67.38 ± 0.5584.27 ± 0.7574.29 ± 0.6689.41 ± 0.77
OM [47]WRN-28-1036.5M66.78 ± 0.3085.29 ± 0.4171.54 ± 0.2987.79 ± 0.46
SUN [14]ViT12.5M67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [32]Swin-Tiny29.0M72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
HCT Transformers [70]3× ViT-S63.0M74.74 ± 0.1789.19 ± 0.1379.67 ± 0.2091.72 ± 0.11
Ours (Prototype)ViT-S21M74.28 ± 0.1888.82 ± 0.0978.83 ± 0.2091.02 ± 0.12
Ours (Classifier)ViT-S21M74.10 ± 0.1788.89 ± 0.0978.81 ± 0.2191.21 ± 0.11
Ours + HCT [70]3× ViT-S63M75.32 ± 0.1889.57 ± 0.0979.74 ± 0.2091.68 ± 0.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.327, + 0.825, + 0.342 + ], + "angle": 0, + "content": "Table 2. Results on CIFAR-FS and FC100. A more comprehensive version of this table is shown in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.343, + 0.868, + 0.525 + ], + "angle": 0, + "content": "
MethodBackbone#ParamsCIFAR-FS,5-wayFC100,5-way
1-shot5-shot1-shot5-shot
BML [78]ResNet-1212.4M73.45 ± 0.4788.04 ± 0.3345.00 ± 0.4163.03 ± 0.41
IE [52]ResNet-1212.4M77.87 ± 0.8589.74 ± 0.5747.76 ± 0.7765.30 ± 0.76
TPMN [67]ResNet-1212.4M75.50 ± 0.9087.20 ± 0.6046.93 ± 0.7163.26 ± 0.74
PSST [9]WRN-28-1036.5M77.02 ± 0.3888.45 ± 0.35--
Meta-QDA [75]WRN-28-1036.5M75.95 ± 0.5988.72 ± 0.79--
SUN [14]ViT12.5M78.37 ± 0.4688.84 ± 0.32--
FewTURE [32]Swin-Tiny29.0M77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
HCT Transformers [70]3× ViT-S63.0M78.89 ± 0.1890.50 ± 0.0948.27 ± 0.1566.42 ± 0.16
Ours (Prototype)ViT-S21M80.08 ± 0.1890.63 ± 0.1350.38 ± 0.1668.37 ± 0.16
Ours (Classifier)ViT-S21M79.82 ± 0.1890.91 ± 0.1350.28 ± 0.1668.50 ± 0.16
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.47, + 0.558 + ], + "angle": 0, + "content": "4.3. Comparison With the State-of-the-arts (SOTAs)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.687 + ], + "angle": 0, + "content": "We evaluate our proposed SMKD with the above-mentioned evaluation methods on four few-shot classification datasets. Our goal in this section is to prove the effectiveness of our method given its simple training pipeline and evaluation procedure. This being said, the performance of our method can be further boosted by adopting strategies from contemporary works (see the last row \"Ours+HCT [70]\" in Table 1 and detailed results in Table 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Compared with traditional convolutional backbones, Transformer-based models are still underdeveloped in few-shot classification. Recent methods with Transformer backbones [14, 32, 70] differ a lot in their training pipelines and evaluation procedures. SUN [14] first pre-trains a teacher network with supervised loss, then uses it to generate patch-level supervision for the student network as a supplement to class-level supervision. HCTransformers [70] converts class-level prediction into a latent prototype learning problem, and introduces spectral tokens pooling to merge neighboring tokens with similar semantic meanings adaptively. FewTURE [32], which adopts a similar two-stage training pipeline as ours, shows the effectiveness of inner loop token importance reweighting to avoid supervision collapse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Despite the success of these works, our SMKD still shows competitive performance given its simple design. Table 1 contains the results of mini-ImageNet and tiered-ImageNet. With Prototype and Classifier as the evaluation method, our proposed SMKD outperforms all methods with ResNet and WRN backbones, and ranked second among methods with Transformer backbones. By adopting the same strategies (patch size of 8 and spectral tokens pooling) as in HCTransformers [70], our method achieves a new SOTA on mini-ImageNet. The effect of each of these two strategies is shown in Table 3. Furthermore, as displayed in Table 2, our method performs the best on the two small-resolution datasets (CIFAR-FS and FC100), and outperforms all previous results by great margins (0.93% for 1-shot and 0.41% for 5-shot on CIFAR-FS, and over 2% for both 1 and 5-shot on FC100). In conclusion, our method grows more effective on datasets with smaller resolutions and fewer training images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The top competitor to our model is HCTransformer [70]. However, their method is more complex compared with ours in the following two aspects: (1) Their method uses a smaller patch size of 8 rather than 16 (which is used in most other methods as well as ours). As shown in Table" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19654" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.131 + ], + "angle": 0, + "content": "Table 3. Results for small patch size and spectral tokens pooling. Results are evaluated on mini-ImageNet with Prototype and Classifier methods, and the best is reported." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.134, + 0.473, + 0.229 + ], + "angle": 0, + "content": "
MethodPatch 8Spec. Pool1-shot5-shot
HCT [70]✓71.27 ± 0.1784.68 ± 0.10
HCT [70]✓✓74.62 ± 0.2089.19 ± 0.13
Ours74.28 ± 0.1888.82 ± 0.09
Ours✓74.00 ± 0.1789.57 ± 0.09
Ours✓✓75.32 ± 0.1889.23 ± 0.18
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.469, + 0.336 + ], + "angle": 0, + "content": "4, this makes their training procedure \\( 4 \\times \\) slower than ours in each epoch. (2) Their method contains three sets of cascaded transformers, each corresponding to a set of teacher-student network with ViT-S backbone. This triples the total number of learnable parameters (63M versus 21M in ours) and incurs much longer inference time." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.351, + 0.468, + 0.391 + ], + "angle": 0, + "content": "Table 4. Training clock time comparison on mini-ImageNet. For HCT, we report the result provided in their paper. All experiments are conducted on 8 Nvidia RTX 3090 GPUs." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.394, + 0.469, + 0.455 + ], + "angle": 0, + "content": "
HCT [70]Stage1Stage2Stage3
21.1h (400 epoch)0.58h (2 epoch)0.21h (2 epoch)
OursSelf-Supervised Pre-trainSupervised Training
15.0h (1200 epoch)1.08h (60 epoch)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.469, + 0.559 + ], + "angle": 0, + "content": "For a fair comparison, we compare few-shot classification results of methods with the same ViT-S backbone (21M parameters) on mini-ImageNet in Table 5. The results for HCT [70] and FewTURE [32] are copy-pasted from their paper. Our model outperforms them by a large margin when the number of trainable parameters is comparable." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.56, + 0.469, + 0.59 + ], + "angle": 0, + "content": "Visualizations of the multi-head self-attention and patch token correspondence are shown in Fig. 4 and Fig. 5." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.604, + 0.469, + 0.645 + ], + "angle": 0, + "content": "Table 5. Comparison results with the same ViT-S backbone on mini-ImageNet. For HCT, we report their result of the first student transformer (training time shown in Stage 1 of Table 4)." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.647, + 0.466, + 0.717 + ], + "angle": 0, + "content": "
MethodBackbone1-shot5-shot
HCT [70]ViT-S71.27 ± 0.1784.68 ± 0.10
FewTURE [32]ViT-S68.02 ± 0.8884.51 ± 0.53
OursViT-S74.28 ± 0.1888.82 ± 0.09
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.229, + 0.758 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Self-Supervised MIM Pretraining Helps. Our training pipeline contains two stages. A natural question is whether the first self-supervised pretraining stage is necessary. We can see from Table 6 that the classification accuracy of our method drops a lot without pretraining. If we train in a single stage, with the combination of the self-supervised learning losses and supervised contrastive learning losses, the model is still hard to converge without a good initialization (1-shot: 43.71 vs. our 74.28. 5-shot: 60.27 vs. our" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.885, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Figure 4. Visualization of multi-head self-attention maps. The self-attention of the [c1s] tokens with different heads in the last attention layer of ViT are visualized in different colors. iBOT+CE represents the model first pre-trained with iBOT, then trained with CE loss. Our SMKD pays more attention to the foreground objects, especially the most discriminative parts." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.383, + 0.88, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Figure 5. Visualization of dense correspondence. We use the patches with the highest self-attention of the [cls] token on each attention head (6 in total) of the last layer of ViT-S as queries. Best matched patches with highest similarities are connected with lines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "88.82 on mini-[ImageNet]. An explanation is that it will be hard for our patch-level knowledge distillation to find matched patches if the ViT backbone is initialized with random weights. What we observe is that our model finds some shortcut solution that prohibits it from proper training. Additionally, we also test the traditional CE loss with and without self-supervised pretraining. The model pre-trained with self-supervision and then trained with supervised CE loss works surprisingly well and even outperforms most of the methods in Table 1 on mini-[ImageNet]. This further supports the usefulness of our two-stage training procedure. For the masking strategy in the first pretraining stage, our ablation study over three masking types (blockwise, random, and no mask) in the appendix shows that blockwise masking balances the best between 1 and 5-shot classification accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19655" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.173 + ], + "angle": 0, + "content": "Table 6. Few-shot evaluation w/wo self-supervised pretraining. For models with pretraining stage, we pre-train with iBOT for 1200 epochs, then train with CE or our loss for 60 epochs. For models without pretraining stage, we train with supervised losses for 60 epochs directly. Results are evaluated by Prototype classification method on mini-ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.175, + 0.449, + 0.261 + ], + "angle": 0, + "content": "
MethodPre-train1-shot5-shot
CE48.98 ± 0.1667.62 ± 0.14
CE✓71.02 ± 0.1887.25 ± 0.10
Ours33.14 ± 0.1345.02 ± 0.14
Ours✓74.28 ± 0.1888.82 ± 0.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.272, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Table 7. Ablation of different losses in the supervised training stage. All losses are trained for 60 epochs initialized with the pretrained iBOT model. The first row represents the pre-trained iBOT without additional supervised training. \\(\\mathcal{L}_{CE} + \\mathcal{L}_{[\\text{patch}]}\\) stands for the combination of CE loss and our patch-level loss. Results are evaluated by Prototype method on mini-ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.357, + 0.445, + 0.471 + ], + "angle": 0, + "content": "
Loss1-shot5-shot
-60.93 ± 0.1780.38 ± 0.12
LCE71.02 ± 0.1887.25 ± 0.10
L[cls]70.21 ± 0.1787.03 ± 0.10
L[patch]70.84 ± 0.1885.90 ± 0.11
LCE + L[patch]70.70 ± 0.1886.77 ± 0.10
L[cls] + L[patch]74.28 ± 0.1888.82 ± 0.09
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Class and Patch-Level Distillation Complement Each Other. Our SMKD uses the combination of class-level and patch-level supervised-contrastive losses as the final objective. Here we study several different combinations of supervised losses to confirm our proposed method performs the best. Results in Table 7 show that combining \\(\\mathcal{L}_{[\\mathrm{cls}]}\\) with \\(\\mathcal{L}_{[\\mathrm{patch}]}\\) achieves \\(13.35\\%\\) improvement over baseline for 1-shot and \\(8.44\\%\\) for 5-shot classification, which outperforms each of its components. Sharing the parameters of projection heads for class and patch tokens allows the semantics obtained from their distillation to complement each other, which leads to much better performance. This can also be seen from the comparison between \\(\\mathcal{L}_{CE} + \\mathcal{L}_{[\\mathrm{patch}]}\\) and \\(\\mathcal{L}_{[\\mathrm{cls}]} + \\mathcal{L}_{[\\mathrm{patch}]}\\). \\(\\mathcal{L}_{CE}\\) alone achieves stronger performance than \\(\\mathcal{L}_{[\\mathrm{cls}]}\\), but combining it with \\(\\mathcal{L}_{[\\mathrm{patch}]}\\) does not provide additional benefits with separate classification heads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.903 + ], + "angle": 0, + "content": "Weighted Average Pooling Boosts Performance. Some recent works [60,69] find that using average pooling instead of the [cls] token during evaluation can encourage token-level tasks (e.g. localization, segmentation). Inspired by such findings, we evaluate our model with different tokens as well as their combinations. Results from Table 8 show that the performance of the [cls] token can be boosted by \\(3.02\\%\\) for 1-shot and \\(0.79\\%\\) for 5-shot by just concatenating with the weighted average pooling token. This shows" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "that our proposed method can learn meaningful representations for both [cls] and [patch] tokens. Moreover, weighted average pooling and average pooling tokens share similar information, but the former performs better because it gives less weight to the background." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Table 8. Ablation of different tokens for few-shot evaluation. \"cls\" stands for the [cls] token, \"avg pool\" stands for using average pooling as token, \"weighted avg pool\" stands for weighted average [patch] token with weights from the self-attention module from the last block of ViT. The last four rows represent concatenation of these three tokens. All tokens are normalized to unit length. Results are evaluated by Prototype classification method." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.275, + 0.88, + 0.405 + ], + "angle": 0, + "content": "
Embedding1-shot5-shot
①cls71.26 ± 0.1788.03 ± 0.10
②avg pool71.65 ± 0.1986.40 ± 0.11
③weighted avg pool71.83 ± 0.1986.38 ± 0.11
① + ②73.83 ± 0.1888.58 ± 0.10
① + ③74.28 ± 0.1888.82 ± 0.09
② + ③71.86 ± 0.1986.47 ± 0.11
① + ② + ③73.80 ± 0.1888.26 ± 0.10
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.619, + 0.432 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.713 + ], + "angle": 0, + "content": "In this work, we propose a novel supervised knowledge distillation framework (SMKD) for few-shot Transformers, which extends the self-supervised masked knowledge distillation framework into the supervised setting. With our design of supervised-contrastive losses, we incorporate supervision into both class and patch-level knowledge distillation while still enjoy the benefits of not needing large batch size and negative samples. Evaluation results together with ablation studies demonstrate the superiority of our method given its simple design compared with contemporary works. Our two-stage training is a special case of curriculum learning from the easy samples to the hard ones. We can unify the learning objectives of self-supervised learning and supervised contrastive learning, using a carefully designed curriculum learning strategy for future works. We hope our work can bridge the gap between self-supervised and supervised knowledge distillation, and inspire more works on supervised few-shot learning methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.728, + 0.687, + 0.744 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.889 + ], + "angle": 0, + "content": "This research is sponsored by Air Force Research Laboratory (AFRL) under agreement number FA8750-19-1-1000. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation therein. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of Air Force Laboratory, DARPA or the U.S. Government." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19656" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[2] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.201, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.257, + 0.468, + 0.325 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.327, + 0.468, + 0.381 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.383, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[6] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.426, + 0.468, + 0.478 + ], + "angle": 0, + "content": "[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.481, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.538, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[9] Zhengyu Chen, Jixie Ge, Heshen Zhan, Siteng Huang, and Donglin Wang. Pareto self-supervised training for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13663-13672, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[10] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.732 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.787 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[13] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. Advances in Neural Information Processing Systems, 33:21981-21993, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[14] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[15] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[17] Yutong Feng, Jianwen Jiang, Mingqian Tang, Rong Jin, and Yue Gao. Rethinking supervised pre-training for better downstream transferring. arXiv preprint arXiv:2110.06014, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[18] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? arXiv preprint arXiv:2210.07240, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8059-8068, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[22] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3263-3272, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 780-789, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[24] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[25] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, Rama Chellappa, and Shih-Fu Chang. Multimodal few-shot object detection with meta-learning based cross-modal prompting. arXiv preprint arXiv:2204.07841, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Guangxing Han, Xuan Zhang, and Chongrong Li. Revisiting faster r-cnn: A deeper look at region proposal network. In International Conference on Neural Information Processing, pages 14-24, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19657" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Guangxing Han, Xuan Zhang, and Chongrong Li. Single shot object detection with top-down refinement. In 2017 IEEE International Conference on Image Processing (ICIP), pages 3360-3364. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.216 + ], + "angle": 0, + "content": "[28] Guangxing Han, Xuan Zhang, and Chongrong Li. Semi-supervised diff: Decoupling detection and feature flow for video object detectors. In Proceedings of the 26th ACM international conference on Multimedia, pages 1811-1819, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[32] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. arXiv preprint arXiv:2206.07267, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[35] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[36] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[37] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34:23818-23830, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[39] Jiang Lu, Pinghua Gong, Jieping Ye, and Changshui Zhang. Learning from very few samples: A survey. arXiv preprint arXiv:2009.02653, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[40] Yuning Lu, Liangjian Wen, Jianzhuang Liu, Yajing Liu, and Xinmei Tian. Self-supervision can be a good few-shot learner. arXiv preprint arXiv:2207.09176, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Xu Luo, Longhui Wei, Liangjian Wen, Jinrong Yang, Lingxi Xie, Zenglin Xu, and Qi Tian. Rectifying the shortcut learning of background: Shared object concentration for few-shot" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.119 + ], + "angle": 0, + "content": "image recognition. arXiv preprint arXiv:2107.07746, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[42] Jiawei Ma, Hanchen Xie, Guangxing Han, Shih-Fu Chang, Aram Galstyan, and Wael Abd-Almageed. Partner-assisted learning for few-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10573-10582, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[43] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2218-2227, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[44] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[45] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. arXiv preprint arXiv:2210.10615, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[46] Elia Peruzzo, Enver Sangineto, Yahui Liu, Marco De Nadai, Wei Bi, Bruno Lepri, and Nicu Sebe. Spatial entropy regularization for vision transformers. arXiv preprint arXiv:2206.04636, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[47] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8412-8422, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.5, + 0.892, + 0.581 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.893, + 0.685 + ], + "angle": 0, + "content": "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[51] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[52] Mamshad Nayeem Rizve, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Exploring complementary strengths of invariant and equivariant representations for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10836-10846, June 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19658" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[55] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[56] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[57] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[58] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.461 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[62] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[63] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[64] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[65] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[66] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[67] Jiamin Wu, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Task-aware part mining network for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8433-8442, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[68] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[69] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[70] Dongyang Zhao Hong-Yu Zhou Weifeng Ge Yizhou Yu Wenqiang Zhang Yangji He, Weihan Liang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[71] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8808-8817, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[72] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The met dataset: Instance-level recognition for artworks. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[74] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[75] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[76] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? arXiv preprint arXiv:2006.06606, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[78] Ziqi Zhou, Xi Qiu, Jiangtao Xie, Jianan Wu, and Chi Zhang. Binocular mutual learning for improving few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8402-8411, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.828, + 0.892, + 0.881 + ], + "angle": 0, + "content": "[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.881 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19659" + } + ] +] \ No newline at end of file diff --git a/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_origin.pdf b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dba36503df01c011fdd06e3eb6c199af814ad4be --- /dev/null +++ b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/19caecc7-ab68-426b-8fbc-ffe20dc02942_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efc868ab5cf5a03f071beb2e3e1067ed99f0b5d2897358910f557d388db88642 +size 1684070 diff --git a/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/full.md b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3cbd67428626bbbe2dd3981872de7759e8473587 --- /dev/null +++ b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/full.md @@ -0,0 +1,309 @@ +# Supervised Masked Knowledge Distillation for Few-Shot Transformers + +Han Lin*, Guangxing Han\*†, Jiawei Ma, Shiyuan Huang, Xudong Lin, Shih-Fu Chang Columbia University + +{h13199, gh2561, jiawei.m, sh3813, x12798, sc250}@columbia.edu + +# Abstract + +Vision Transformers (ViTs) emerge to achieve impressive performance on many data-abundant computer vision tasks by capturing long-range dependencies among local features. However, under few-shot learning (FSL) settings on small datasets with only a few labeled data, ViT tends to overfit and suffers from severe performance degradation due to its absence of CNN-alike inductive bias. Previous works in FSL avoid such problem either through the help of self-supervised auxiliary losses, or through the dextile uses of label information under supervised settings. But the gap between self-supervised and supervised few-shot Transformers is still unfilled. Inspired by recent advances in self-supervised knowledge distillation and masked image modeling (MIM), we propose a novel Supervised Masked Knowledge Distillation model (SMKD) for few-shot Transformers which incorporates label information into self-distillation frameworks. Compared with previous self-supervised methods, we allow intra-class knowledge distillation on both class and patch tokens, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. Experimental results on four few-shot classification benchmark datasets show that our method with simple design outperforms previous methods by a large margin and achieves a new start-of-the-art. Detailed ablation studies confirm the effectiveness of each component of our model. Code for this paper is available here: https://github.com/HL-hanlin/SMKD. + +# 1. Introduction + +Vision Transformers (ViTs) [16] have emerged as a competitive alternative to Convolutional Neural Networks (CNNs) [31] in recent years, and have achieved impressive performance in many vision tasks including image classification [16, 38, 59, 66], object detection [3, 10, 26-28, 79], and object segmentation [50, 55]. Compared with CNNs, which introduce inductive bias through convolutional kernels with fixed receptive fields [35], the attention layers in + +![](images/a569f99be3d3fbeffe35c16e1a351519717b86e1318889b45f343c6814f476a6.jpg) +(a) Few-Shot Transformers with Soft Label Regularization + +![](images/7639a8994228dc8bd335e0f4d910b602ff0d7cb61cb44573f51d684dffe1efa0.jpg) +(b) Few-Shot Transformers with Self-Supervision Auxiliary Regularization + +![](images/7a8d7ae397840fc6de041cf50d80ee656ca1e14a0d3bdbbb02aa83c5bbd202f1.jpg) +(c) Few-Shot Transformers with Our Supervised Masked Knowledge Distillation +Figure 1. Comparison of the proposed idea and other existing methods for few-shot Transformers. Our model mitigates the overfitting issue of few-shot Transformers, by extending the masked knowledge distillation framework into the supervised setting, and enforcing the alignment of [cls] and corresponding [patch] tokens for intra-class images. + +ViT allow it to model global token relations to capture long-range token dependencies. However, such flexibility also comes at a cost: ViT is data-hungry and it needs to learn token dependencies purely from data. This property often makes it easy to overfit to datasets with small training set and suffer from severe generalization performance degradation [36, 37]. Therefore, we are motivated to study how to make ViTs generalize well on these small datasets, especially under the few-shot learning (FSL) setting [19, 39, 62] which aims to recognize unseen new instances at test time just from only a few (e.g. one or five) labeled samples from each new categories. + +Most of the existing methods mitigate the overfitting issue of few-shot Transformers [14] using various regularizations. For instance, some works utilize label information in a weaker [70], softer [42] way, or use label information efficiently through patch-level supervision [14]. However, these models usually design sophisticated learning targets. + +On the other hand, self-distillation techniques [4, 8], and particularly, the recent masked self-distillation [29, 45, 77], which distills knowledge learned from an uncorrupted image to the knowledge predicted from a masked image, have lead an emerging trend in self-supervised Transformers in various fields [15, 68]. Inspired by such success, recent works in FSL attempt to incorporate self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46], or to adopt a self-supervised pretraining, supervised training two-stage framework to train few-shot Transformers [18, 32]. Compared with traditional supervised methods, self-supervision can learn less biased representations towards base class, which usually leads to better generalization ability for novel classes [40]. However, the two learning objectives of self-supervision and supervision are conflicting and it is hard to balance them during training. Therefore, how to efficiently leverage the strengths of self-supervised learning to alleviate the overfitting issue of supervised training remains a challenge. + +In this work, we propose a novel supervised masked knowledge distillation framework (SMKD) for few-shot Transformers, which handles the aforementioned challenge through a natural extension of the self-supervised masked knowledge distillation framework into the supervised setting (shown in Fig. 1). Different from supervised contrastive learning [33] which only utilizes global image features for training, we leverage multi-scale information from the images (both global [cls] token and local [patch] tokens) to formulate the learning objectives, which has been demonstrated to be effective in the recent self-supervised Transformer methods [29, 77]. For global [cls] tokens, we can simply maximize the similarity for intra-class images. However, it is non-trivial and challenging to formulate the learning objectives for local [patch] tokens because we do not have ground-truth patch-level annotations. To address this problem, we propose to estimate the similarity between [patch] tokens across intra-class images using cross-attention, and enforce the alignment of corresponding [patch] tokens. Particularly, reconstructing masked [patch] tokens across intra-class images increases the difficulty of model learning, thus encouraging learning generalizable few-shot Transformer models by jointly exploiting the holistic knowledge of the image and the similarity of intra-class images. + +As shown in Fig. 2, we compare our model with the existing self-supervised/supervised learning methods. Our model is a natural extension of the supervised contrastive learning method [33] and self-supervised knowledge distillation methods [4, 77]. Thus our model inherits both the advantage of method [33] for effectively leveraging label information, and the advantages of methods [4, 77] for not needing large batch size and negative samples. Meanwhile, the newly-introduced challenging task of masked + +![](images/5e95c3374f45762922ae8483735494ef364b93840e6e69b793cb0578d89dadb5.jpg) + +![](images/904899dc3403ebb8f1bbaaff312a687d96a9424860c65ae3953ef0ed91008a7b.jpg) + +![](images/ad243b514655fef5979eb6795ed347e98214cf1b0b10e98ca2d0f490ef8fc638.jpg) +Figure 2. Comparison of other self-supervised/supervised frameworks. Our method (d) is a natural extension of (b) and (c), with the newly-introduced challenging task of masked [patch] tokens reconstruction across intra-class images. + +![](images/101bdef3dc637e02f74da15925811687b9c3e559329269caf6a547509e1845e4.jpg) + +[patch] tokens reconstruction across intra-class images makes our method more powerful for learning generalizable few-shot Transformer models. + +Compared with contemporary works on few-shot Transformers [14, 32, 70], our framework enjoys several good properties from a practical point of view. (1) Our method does not introduce any additional learnable parameters besides the ViT backbone and projection head, which makes it easy to be combined with other methods [32, 70, 74]. (2) Our method is both effective and training-efficient, with stronger performance and less training time on four few-shot classification benchmarks, compared with [32, 70]. In a nutshell, our main contributions can be summarized as follows: + +- We propose a new supervised knowledge distillation framework (SMKD) that incorporates class label information into self-distillation, thus filling the gap between self-supervised knowledge distillation and traditional supervised learning. +- Within the proposed framework, we design two supervised-contrastive losses on both class and patch levels, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. +- Given its simple design, we test our SMKD on four few-shot datasets, and show that it achieves a new SOTA on CIFAR-FS and FC100 by a large margin, as well as competitive performance on mini-ImageNet and tiered-ImageNet using the simple prototype classification method for few-shot evaluation. + +# 2. Related Work + +Few-shot Learning. Few-shot learning aims at fast knowledge transfer from seen base classes to unseen novel classes given only a few labeled samples from each novel class. The meta-learning paradigm, which simulates few-shot tasks episodically to mimic the human-learning process in the + +real world, once dominated FSL [22-25, 44, 53, 54, 56, 62]. Surprisingly, recent works [6, 21, 58, 72] have shown that the state-of-the-art meta-learning methods can be outperformed by simple baseline methods using just a distance-based classifier without the complicated design of meta-learning strategies. Therefore, recent methods [17, 42, 70] in FSL start to focus less on meta-learning, and more on learning embeddings with good generalization ability. Our paper follows this trend and proposes a knowledge distillation framework to learn generalizable embeddings. + +Vision Transformers in FSL. Variants of Transformer [61] have achieved great success in NLP [12, 49], computer vision [16, 59], and multimodal learning [48]. However, the lack of inductive bias makes Transformer infamous for its data-hungry property, which is especially significant in few-shot settings when the amount of data is limited. A line of works focuses on introducing inductive bias back to the Transformer architecture, including methods using pyramid structure [63, 64], shifted windows [38], and explicit convolutional token embeddings [66]. With this said, some recent works [13, 24, 32, 71] still show that few-shot Transformers have the potential of fast adaptation to novel classes. Our work also studies few-shot Transformers, and we show that our method can work well even on the vanilla ViT structure without explicit incorporation of inductive bias. + +Self-Supervision for FSL. Self-supervised learning (SSL) has shown great potential in FSL due to its good generalization ability to novel classes. Previous methods incorporate SSL into FSL in various ways. Some works propose to include self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46]. For example, [37] proposed a regularization loss that extracts additional information from images by predicting the geometric distance between patch tokens. [46] designed their loss by discouraging spatially disordered attention maps based on the idea that objects usually occupy connected regions, and [43] derived their auxiliary loss from self-supervision tasks of rotation and exemplar. Some other works [18, 32], instead, adopt a two-stage procedure by pretraining a model via self-supervision before supervised training. [32] takes advantage of self-supervised training with iBOT [77] as a pretext task, and then uses inner loop token importance reweighting for supervised fine-tuning. [18] initializes its model with a pre-trained self-supervised DINO [4], then trained with supervised cross-entropy loss. Our work follows the second branch of work, but the difference is obvious from previous works. Instead of designing complicated training pipeline or leveraging additional learnable modules at inference time, we use supervised training on self-supervised pre-trained model, and focus mainly on bridging the gap between self-supervised and supervised knowledge distillation with minimum extra design. + +Another emerging trend in SSL is Masked Image Mod + +eling (MIM) [1, 29, 45, 77], which aims at recovering the patch-level target (e.g. image pixels, patch features) of the masked content in a corrupted input image. In iBOT [77], the class and patch tokens share the same projection head, thus helping the ViT backbone to learn semantically meaningful patch embeddings, which can be used as a supplement to image-level global self-supervision. + +# 3. Our Approach + +We present our Supervised Masked Knowledge Distillation (SMKD) framework in this section. Fig. 3 shows an overview of our model. Task definition for few-shot classification is presented in Sec. 3.1. We explain the formulation of SSL knowledge distillation in Sec. 3.2, and how we extend it into our supervised knowledge distillation framework in Sec. 3.3. Training pipeline is given in Sec. 3.4. + +# 3.1. Learning Task Definition + +In few-shot classification, we are given two datasets $\mathcal{D}_{base}$ and $\mathcal{D}_{novel}$ with disjoint class labels $\mathcal{C}_{base}$ and $\mathcal{C}_{novel}$ ( $\mathcal{C}_{base} \cap \mathcal{C}_{novel} = \emptyset$ ). The base dataset $\mathcal{D}_{base}$ , which contains abundant labeled samples, is used to train the feature extractor. And the novel dataset $\mathcal{D}_{novel}$ is then used to sample episodes for prototype estimation and few-shot evaluation. In a standard N-way K-shot task, each episode $\mathcal{D}_{epi} = (\mathcal{D}_S, \mathcal{D}_Q)$ covers N classes from $\mathcal{C}_{novel}$ . The support dataset $\mathcal{D}_S$ , which contains K samples from each class, is used for class prototype estimation, and the query dataset $\mathcal{D}_Q$ is then used for evaluation. This task aims at correctly classifying $\mathcal{D}_Q$ into N classes from sampled episodes. And the main focus of this paper is to train a feature extractor with good generalization ability from $\mathcal{D}_{base}$ . + +# 3.2. Preliminary: SSL with Knowledge Distillation + +Our work is inspired by the self-supervised knowledge distillation frameworks proposed recently [4]. Specifically, given an input image $\pmb{x}$ uniformly sampled from the training set $\mathcal{I}$ , random data augmentations are applied to generate two augmented views $\pmb{x}^1$ and $\pmb{x}^2$ (we represent image view indices as superscripts and patch indices as subscripts), which are then fed into the teacher and student networks. The student network, parameterized by $\theta_{s}$ , consists of an encoder with ViT backbone and a projection head with a 3-layer multi-layer perceptron (MLP) followed by $l_{2}$ -normalized bottleneck. The ViT backbone first generates a [cls] token, which is then entered into the projection head and outputs a probability distribution $P_{s}^{[\mathrm{cls}]}$ over K dimensions. The teacher network, parameterized by $\theta_{t}$ , is Exponentially Moving Averaged (EMA) updated by the student network $\theta_{s}$ , which distills its knowledge to the student by minimizing the cross-entropy loss over the outputs of the + +![](images/2d4f90643870ab2108a670f8c23a0de7f96160d43079a1eaf2d2306d8148aea0.jpg) +Figure 3. Overview of our SMKD framework. Two views are generated from a pair of images sampled (with replacement) from the same class. The first view, applied with random blockwise masking, is passed to the student network, while the second unmasked view is passed to the teacher network. Both networks consist of a ViT backbone and a projection head. The parameters of the teacher network are Exponentially Moving Averaged (EMA) updated by the student network. SMKD distills knowledge between intra-class cross-views on both class and patch levels. $\mathcal{L}_{[cls]}$ distills knowledge from [cls] tokens, while $\mathcal{L}_{[patch]}$ distills knowledge from [patch] tokens by finding dense correspondence of matched token pairs (connected by red dashed lines) with highest similarities. + +categorical distributions from their projection heads: + +$$ +\mathcal {L} _ {[ \mathrm {c l s} ]} = \mathcal {H} (\boldsymbol {P} _ {t} ^ {[ \mathrm {c l s} ]} (\boldsymbol {x} ^ {1}), \boldsymbol {P} _ {\boldsymbol {s}} ^ {[ \mathrm {c l s} ]} (\boldsymbol {x} ^ {2})) \qquad (1) +$$ + +where $\mathcal{H}(x,y) = -x\log y$ + +Masked Image Modeling (MIM) [1, 57] can be performed via self-distillation as follows [77]. Given a randomly sampled mask sequence $m \in \{0,1\}^N$ over an image with $N$ [patch] tokens $\pmb{x} = \{\pmb{x}_i\}_{i=1}^N$ , $\pmb{x}_i$ 's with $m_i = 1$ are then replaced by a learnable token embedding $e_{[\text{MASK}]}$ , which results in a corrupted image $\widehat{\pmb{x}} = \{\widehat{\pmb{x}}_i\}_{i=1}^N = \{(1 - m_i)\pmb{x}_i + m_i e_{[\text{MASK}]}\}_{i=1}^N$ . This corrupted image and the original uncorrupted image are fed into the student and teacher networks respectively. The objective of MIM is to recover the masked tokens from the corrupted image, which is equivalent to minimizing the cross-entropy loss between the outputs of the categorical distributions of the student and teacher networks on masked patches: + +$$ +\mathcal {L} _ {\mathrm {M I M}} = \sum_ {i = 1} ^ {N} m _ {i} \cdot \mathcal {H} \left(\boldsymbol {P} _ {\boldsymbol {t}} ^ {[ \text {p a t c h} ]} (\boldsymbol {x}) _ {i}, \boldsymbol {P} _ {\boldsymbol {s}} ^ {[ \text {p a t c h} ]} (\widehat {\boldsymbol {x}}) _ {i}\right) \tag {2} +$$ + +# 3.3. Supervised Masked Knowledge Distillation + +Distill the Class Tokens. Recent self-distillation frameworks [4, 20, 77] distill knowledge on [cls] tokens from cross-view images via Eq.(1). To incorporate label information into such self-supervised frameworks, we further allow knowledge on [cls] tokens to be distilled from intra-class cross-views. This can be achieved by a small extension of the way we sample images. Rather than sampling a single image $\boldsymbol{x} \sim \mathcal{I}$ and generating two views, now we sample two images $\boldsymbol{x}, \boldsymbol{y} \sim \mathcal{I}^c$ (with replacement) and generate one view for each of them1. $\mathcal{I}^c \subseteq \mathcal{I}$ here denotes the set + +of images with the same class label $c$ in the training set $\mathcal{I}$ . Specifically, we denote $\pmb{x}^{\prime}$ and $\pmb{y}^{\prime}$ as the augmented views generated from image $\pmb{x}$ and $\pmb{y}$ respectively. We apply additional random blockwise masking on $\pmb{x}^{\prime}$ and denote the resulting corrupted view as $\widehat{\pmb{x}}^{\prime}$ . Then the corrupted view $\widehat{\pmb{x}}^{\prime}$ and uncorrupted view $\pmb{y}^{\prime}$ are sent to the student and teacher networks separately. Now our supervised-contrastive loss on [cls] tokens becomes: + +$$ +\mathcal {L} _ {[ \mathrm {c l s} ]} = \mathcal {H} \left(\boldsymbol {P} _ {\boldsymbol {t}} ^ {[ \mathrm {c l s} ]} \left(\boldsymbol {y} ^ {\prime}\right), \boldsymbol {P} _ {\boldsymbol {s}} ^ {[ \mathrm {c l s} ]} \left(\hat {\boldsymbol {x}} ^ {\prime}\right)\right) \tag {3} +$$ + +When $\pmb{x}, \pmb{y}$ are sampled as the same image $(\pmb{x} = \pmb{y})$ , our loss performs masked self-distillation across two views of the same image similar to Eq.(1). In the other situation when $\pmb{x}$ and $\pmb{y}$ represent different images $(\pmb{x} \neq \pmb{y})$ , our loss then minimizes the cross-entropy loss of projected [c1s] tokens among all cross-view pairs between images $\pmb{x}$ and $\pmb{y}$ . + +Such design has two major advantages. (1) It can be implemented efficiently. Instead of intentionally sampling image pairs from the same class, we just need to look through the images in a mini-batch, find image pairs belonging to the same class, and then apply our loss in Eq.(3). (2) Unlike previous works that use either supervised [42,76] or self-supervised [5, 30] contrastive losses, our method follows the recent trend in SSL works [4, 7, 20, 73] and avoids the need for negative examples. + +Distill the Patch Tokens. Beyond the knowledge distillation on the global [cls] tokens, we introduce the challenging task of masked patch tokens reconstruction across intra-class images, to fully exploit the local details of the image for training. Our main intuition here is based on the following hypothesis: for intra-class images, even though their semantic information can be vastly different at the patch level, there should at least exist some patches that share similar semantic meanings. + +For each patch $k$ from an input view $\pmb{y}^{\prime}$ sent to the teacher network (with its corresponding token embedding defined as $\pmb{f}_{k}^{t}$ ), we need to first find the most similar patch $k^{+}$ from the masked view $\widehat{\pmb{x}}^{\prime}$ of the student network (with its corresponding token embedding defined as $\pmb{f}_{k^{+}}^{s}$ ), and then perform knowledge distillation between these two matched tokens. The token embedding $\pmb{f}_{k^{+}}^{s}$ represents either the $k^{+th}$ [patch] token if patch $k^{+}$ is unmasked, or the reconstructed patch token if it is masked. + +As we do not have any patch-level annotations, we use cosine similarity to find the best-matched patch of $k$ among all the [patch] tokens in the student network: + +$$ +k ^ {+} = \arg \max _ {l \in [ N ]} \frac {\boldsymbol {f} _ {k} ^ {t} {} ^ {\top} \boldsymbol {f} _ {l} ^ {s}}{\| \boldsymbol {f} _ {k} ^ {t} \| \| \boldsymbol {f} _ {l} ^ {s} \|} \tag {4} +$$ + +Our patch-level knowledge distillation loss now becomes: + +$$ +\mathcal {L} _ {[ \text {p a t c h} ]} = \sum_ {k = 1} ^ {N} \omega_ {k +} \cdot \mathcal {H} \left(\boldsymbol {P} _ {\boldsymbol {t}} ^ {[ \text {p a t c h} ]} \left(\boldsymbol {y} ^ {\prime}\right) _ {k}, \boldsymbol {P} _ {\boldsymbol {s}} ^ {[ \text {p a t c h} ]} \left(\hat {\boldsymbol {x}} ^ {\prime}\right) _ {k +}\right) \tag {5} +$$ + +where $\omega_{k+}$ is a scalar representing the weight we give to each loss term. We find that taking it as a constant value is effective enough. The ablation study of more complex designs of $\omega_{k+}$ is given in the appendix. + +Our loss shares some similarities with DenseCL [65]. However, the differences are also obvious: (1) Our loss serves as an extension of their self-supervised contrastive loss into a supervised variant. (2) We further incorporate MIM into our design and allow masked patches to be matched, which makes our task harder and leads to more semantically meaningful patch embeddings. + +# 3.4. Training Pipeline + +We train our model in two stages: self-supervised pretraining and supervised training. In the first stage, we use the recently proposed MIM framework [77] for self-supervised pretraining. The self-supervised loss is the summation of $\mathcal{L}_{[\mathrm{cls}]}$ and $\mathcal{L}_{\mathrm{MIM}}$ in Eq.(1) and Eq.(2) without scaling. In the second stage, we continue training the pre-trained model using our supervised-contrastive losses $\mathcal{L}_{[\mathrm{cls}]}$ and $\mathcal{L}_{[\mathrm{patch}]}$ from Eq.(3) and Eq.(5). We define our training loss as: $\mathcal{L} = \mathcal{L}_{[\mathrm{cls}]} + \lambda \mathcal{L}_{[\mathrm{patch}]}$ , where $\lambda$ controls the relative scale of these two components. A relatively large $\lambda$ will make our model focus more on localization and less on high-level semantics. + +# 4. Experiment + +# 4.1. Datasets + +Our model is evaluated on four widely used and publicly available few-shot classification datasets: mini-ImageNet [62], tiered-ImageNet [51], CIFAR-FS [2], and FC100 + +[44]. mini-ImageNet and tiered-ImageNet are derived from ImageNet [11], while CIFAR-FS and FC100 are derived from CIFAR100 [34]. mini-ImageNet contains 100 classes, which are randomly split into 64 base classes for training, 16 classes for few-shot validation, and 20 classes for few-shot evaluation. There are 600 images in each class. tiered-ImageNet contains 609 classes with 779165 images in total. The class split for training, few-shot validation and few-shot evaluation are 351, 97, and 160. CIFAR-FS contains 100 classes with class split as 64, 16, and 20. FC100 contains 100 classes with class split as 60, 20, and 20. For both of these two datasets, each class has 600 images with smaller resolutions $(32 \times 32)$ compared with ImageNet. + +# 4.2. Implementation Details + +Self-supervised pretraining. We pre-train our Vision Transformer backbone and projection head following the same pipeline in iBOT [77]. Most of the hyper-parameter settings are kept unchanged without tuning. We use a batch size of 640 and a learning rate of 0.0005 decayed with the cosine schedule. mini-[ImageNet and tiered-[ImageNet are pre-trained for 1200 epochs, and CIFAR-FS and FC100 are pre-trained for 900 epochs. All models are trained on 8 Nvidia RTX 3090 GPUs. Detailed training parameter settings are included in the appendix. + +Supervised knowledge distillation After getting the pre-trained model, we train it with our supervised-contrastive losses. We find that the model with the best performance can converge within 60 epochs in the validation set. We use the same batch size and learning rate as in the pretraining stage. Compared with the first stage, the only extra hyper-parameter is the scaling parameter $\lambda$ . We set it to make the ratio between $\mathcal{L}_{[\mathrm{patch}]}$ and $\mathcal{L}_{[\mathrm{cls}]}$ roughly around 2. Ablation over different $\lambda$ is given in the appendix. + +Few-shot Evaluation We use the simple prototype classification method (Prototype) [44, 54] and the linear classifier (Classifier) used in S2M2 [43] as our default methods for few-shot evaluation. For each sampled episode data in an N-way K-shot task, Prototype first estimates each class prototype using the averaged feature over K support samples. Then a new sampled query image is classified into one of the N classes which has the highest cosine similarity between its feature vector and the class prototype. Instead, Classifier trains a linear classifier from the $\mathrm{N}\times \mathrm{K}$ support samples, which is then used to classify new query samples. More complex evaluation methods (e.g. DeepEMD) are also compatible with our framework. The feature we used for evaluation is the concatenation of the [cls] token with the weighted average [patch] token (weighted avg pool). The weights of weighted avg pool is the average of the self-attention values of the [cls] token with all heads of the last attention layer. Ablation over different choices of features for evaluation is studied in Sec. 4.4. + +Table 1. Results on mini-ImageNet and tiered-ImageNet. Top three methods are colored in red, blue and green respectively. A more comprehensive version of this table is shown in the appendix. + +
MethodBackbone#ParamsminiImageNet,5-waytieredImageNet,5-way
1-shot5-shot1-shot5-shot
DeepEMD [74]ResNet-1212.4M65.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
IE [52]ResNet-1212.4M67.28 ± 0.8084.78 ± 0.5272.21 ± 0.9087.08 ± 0.58
COSOC [41]ResNet-1212.4M69.28 ± 0.4985.16 ± 0.4273.57 ± 0.4387.57 ± 0.10
Meta-QDA [75]WRN-28-1036.5M67.38 ± 0.5584.27 ± 0.7574.29 ± 0.6689.41 ± 0.77
OM [47]WRN-28-1036.5M66.78 ± 0.3085.29 ± 0.4171.54 ± 0.2987.79 ± 0.46
SUN [14]ViT12.5M67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [32]Swin-Tiny29.0M72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
HCT Transformers [70]3× ViT-S63.0M74.74 ± 0.1789.19 ± 0.1379.67 ± 0.2091.72 ± 0.11
Ours (Prototype)ViT-S21M74.28 ± 0.1888.82 ± 0.0978.83 ± 0.2091.02 ± 0.12
Ours (Classifier)ViT-S21M74.10 ± 0.1788.89 ± 0.0978.81 ± 0.2191.21 ± 0.11
Ours + HCT [70]3× ViT-S63M75.32 ± 0.1889.57 ± 0.0979.74 ± 0.2091.68 ± 0.11
+ +Table 2. Results on CIFAR-FS and FC100. A more comprehensive version of this table is shown in the appendix. + +
MethodBackbone#ParamsCIFAR-FS,5-wayFC100,5-way
1-shot5-shot1-shot5-shot
BML [78]ResNet-1212.4M73.45 ± 0.4788.04 ± 0.3345.00 ± 0.4163.03 ± 0.41
IE [52]ResNet-1212.4M77.87 ± 0.8589.74 ± 0.5747.76 ± 0.7765.30 ± 0.76
TPMN [67]ResNet-1212.4M75.50 ± 0.9087.20 ± 0.6046.93 ± 0.7163.26 ± 0.74
PSST [9]WRN-28-1036.5M77.02 ± 0.3888.45 ± 0.35--
Meta-QDA [75]WRN-28-1036.5M75.95 ± 0.5988.72 ± 0.79--
SUN [14]ViT12.5M78.37 ± 0.4688.84 ± 0.32--
FewTURE [32]Swin-Tiny29.0M77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
HCT Transformers [70]3× ViT-S63.0M78.89 ± 0.1890.50 ± 0.0948.27 ± 0.1566.42 ± 0.16
Ours (Prototype)ViT-S21M80.08 ± 0.1890.63 ± 0.1350.38 ± 0.1668.37 ± 0.16
Ours (Classifier)ViT-S21M79.82 ± 0.1890.91 ± 0.1350.28 ± 0.1668.50 ± 0.16
+ +# 4.3. Comparison With the State-of-the-arts (SOTAs) + +We evaluate our proposed SMKD with the above-mentioned evaluation methods on four few-shot classification datasets. Our goal in this section is to prove the effectiveness of our method given its simple training pipeline and evaluation procedure. This being said, the performance of our method can be further boosted by adopting strategies from contemporary works (see the last row "Ours+HCT [70]" in Table 1 and detailed results in Table 3). + +Compared with traditional convolutional backbones, Transformer-based models are still underdeveloped in few-shot classification. Recent methods with Transformer backbones [14, 32, 70] differ a lot in their training pipelines and evaluation procedures. SUN [14] first pre-trains a teacher network with supervised loss, then uses it to generate patch-level supervision for the student network as a supplement to class-level supervision. HCTransformers [70] converts class-level prediction into a latent prototype learning problem, and introduces spectral tokens pooling to merge neighboring tokens with similar semantic meanings adaptively. FewTURE [32], which adopts a similar two-stage training pipeline as ours, shows the effectiveness of inner loop token importance reweighting to avoid supervision collapse. + +Despite the success of these works, our SMKD still shows competitive performance given its simple design. Table 1 contains the results of mini-ImageNet and tiered-ImageNet. With Prototype and Classifier as the evaluation method, our proposed SMKD outperforms all methods with ResNet and WRN backbones, and ranked second among methods with Transformer backbones. By adopting the same strategies (patch size of 8 and spectral tokens pooling) as in HCTransformers [70], our method achieves a new SOTA on mini-ImageNet. The effect of each of these two strategies is shown in Table 3. Furthermore, as displayed in Table 2, our method performs the best on the two small-resolution datasets (CIFAR-FS and FC100), and outperforms all previous results by great margins (0.93% for 1-shot and 0.41% for 5-shot on CIFAR-FS, and over 2% for both 1 and 5-shot on FC100). In conclusion, our method grows more effective on datasets with smaller resolutions and fewer training images. + +The top competitor to our model is HCTransformer [70]. However, their method is more complex compared with ours in the following two aspects: (1) Their method uses a smaller patch size of 8 rather than 16 (which is used in most other methods as well as ours). As shown in Table + +Table 3. Results for small patch size and spectral tokens pooling. Results are evaluated on mini-ImageNet with Prototype and Classifier methods, and the best is reported. + +
MethodPatch 8Spec. Pool1-shot5-shot
HCT [70]✓71.27 ± 0.1784.68 ± 0.10
HCT [70]✓✓74.62 ± 0.2089.19 ± 0.13
Ours74.28 ± 0.1888.82 ± 0.09
Ours✓74.00 ± 0.1789.57 ± 0.09
Ours✓✓75.32 ± 0.1889.23 ± 0.18
+ +4, this makes their training procedure $4 \times$ slower than ours in each epoch. (2) Their method contains three sets of cascaded transformers, each corresponding to a set of teacher-student network with ViT-S backbone. This triples the total number of learnable parameters (63M versus 21M in ours) and incurs much longer inference time. + +Table 4. Training clock time comparison on mini-ImageNet. For HCT, we report the result provided in their paper. All experiments are conducted on 8 Nvidia RTX 3090 GPUs. + +
HCT [70]Stage1Stage2Stage3
21.1h (400 epoch)0.58h (2 epoch)0.21h (2 epoch)
OursSelf-Supervised Pre-trainSupervised Training
15.0h (1200 epoch)1.08h (60 epoch)
+ +For a fair comparison, we compare few-shot classification results of methods with the same ViT-S backbone (21M parameters) on mini-ImageNet in Table 5. The results for HCT [70] and FewTURE [32] are copy-pasted from their paper. Our model outperforms them by a large margin when the number of trainable parameters is comparable. + +Visualizations of the multi-head self-attention and patch token correspondence are shown in Fig. 4 and Fig. 5. + +Table 5. Comparison results with the same ViT-S backbone on mini-ImageNet. For HCT, we report their result of the first student transformer (training time shown in Stage 1 of Table 4). + +
MethodBackbone1-shot5-shot
HCT [70]ViT-S71.27 ± 0.1784.68 ± 0.10
FewTURE [32]ViT-S68.02 ± 0.8884.51 ± 0.53
OursViT-S74.28 ± 0.1888.82 ± 0.09
+ +# 4.4. Ablation Study + +Self-Supervised MIM Pretraining Helps. Our training pipeline contains two stages. A natural question is whether the first self-supervised pretraining stage is necessary. We can see from Table 6 that the classification accuracy of our method drops a lot without pretraining. If we train in a single stage, with the combination of the self-supervised learning losses and supervised contrastive learning losses, the model is still hard to converge without a good initialization (1-shot: 43.71 vs. our 74.28. 5-shot: 60.27 vs. our + +![](images/9c15894a23af3177d1e84c737d22eec888e62334b442a236058cad160f8a756f.jpg) +Figure 4. Visualization of multi-head self-attention maps. The self-attention of the [c1s] tokens with different heads in the last attention layer of ViT are visualized in different colors. iBOT+CE represents the model first pre-trained with iBOT, then trained with CE loss. Our SMKD pays more attention to the foreground objects, especially the most discriminative parts. + +![](images/3f784f34866ff047b8c767b7590e38ffe6fc32699d24df9e2c222566df7e32ec.jpg) +Figure 5. Visualization of dense correspondence. We use the patches with the highest self-attention of the [cls] token on each attention head (6 in total) of the last layer of ViT-S as queries. Best matched patches with highest similarities are connected with lines. + +88.82 on mini-[ImageNet]. An explanation is that it will be hard for our patch-level knowledge distillation to find matched patches if the ViT backbone is initialized with random weights. What we observe is that our model finds some shortcut solution that prohibits it from proper training. Additionally, we also test the traditional CE loss with and without self-supervised pretraining. The model pre-trained with self-supervision and then trained with supervised CE loss works surprisingly well and even outperforms most of the methods in Table 1 on mini-[ImageNet]. This further supports the usefulness of our two-stage training procedure. For the masking strategy in the first pretraining stage, our ablation study over three masking types (blockwise, random, and no mask) in the appendix shows that blockwise masking balances the best between 1 and 5-shot classification accuracy. + +Table 6. Few-shot evaluation w/wo self-supervised pretraining. For models with pretraining stage, we pre-train with iBOT for 1200 epochs, then train with CE or our loss for 60 epochs. For models without pretraining stage, we train with supervised losses for 60 epochs directly. Results are evaluated by Prototype classification method on mini-ImageNet. + +
MethodPre-train1-shot5-shot
CE48.98 ± 0.1667.62 ± 0.14
CE✓71.02 ± 0.1887.25 ± 0.10
Ours33.14 ± 0.1345.02 ± 0.14
Ours✓74.28 ± 0.1888.82 ± 0.09
+ +Table 7. Ablation of different losses in the supervised training stage. All losses are trained for 60 epochs initialized with the pretrained iBOT model. The first row represents the pre-trained iBOT without additional supervised training. $\mathcal{L}_{CE} + \mathcal{L}_{[\text{patch}]}$ stands for the combination of CE loss and our patch-level loss. Results are evaluated by Prototype method on mini-ImageNet. + +
Loss1-shot5-shot
-60.93 ± 0.1780.38 ± 0.12
LCE71.02 ± 0.1887.25 ± 0.10
L[cls]70.21 ± 0.1787.03 ± 0.10
L[patch]70.84 ± 0.1885.90 ± 0.11
LCE + L[patch]70.70 ± 0.1886.77 ± 0.10
L[cls] + L[patch]74.28 ± 0.1888.82 ± 0.09
+ +Class and Patch-Level Distillation Complement Each Other. Our SMKD uses the combination of class-level and patch-level supervised-contrastive losses as the final objective. Here we study several different combinations of supervised losses to confirm our proposed method performs the best. Results in Table 7 show that combining $\mathcal{L}_{[\mathrm{cls}]}$ with $\mathcal{L}_{[\mathrm{patch}]}$ achieves $13.35\%$ improvement over baseline for 1-shot and $8.44\%$ for 5-shot classification, which outperforms each of its components. Sharing the parameters of projection heads for class and patch tokens allows the semantics obtained from their distillation to complement each other, which leads to much better performance. This can also be seen from the comparison between $\mathcal{L}_{CE} + \mathcal{L}_{[\mathrm{patch}]}$ and $\mathcal{L}_{[\mathrm{cls}]} + \mathcal{L}_{[\mathrm{patch}]}$ . $\mathcal{L}_{CE}$ alone achieves stronger performance than $\mathcal{L}_{[\mathrm{cls}]}$ , but combining it with $\mathcal{L}_{[\mathrm{patch}]}$ does not provide additional benefits with separate classification heads. + +Weighted Average Pooling Boosts Performance. Some recent works [60,69] find that using average pooling instead of the [cls] token during evaluation can encourage token-level tasks (e.g. localization, segmentation). Inspired by such findings, we evaluate our model with different tokens as well as their combinations. Results from Table 8 show that the performance of the [cls] token can be boosted by $3.02\%$ for 1-shot and $0.79\%$ for 5-shot by just concatenating with the weighted average pooling token. This shows + +that our proposed method can learn meaningful representations for both [cls] and [patch] tokens. Moreover, weighted average pooling and average pooling tokens share similar information, but the former performs better because it gives less weight to the background. + +Table 8. Ablation of different tokens for few-shot evaluation. "cls" stands for the [cls] token, "avg pool" stands for using average pooling as token, "weighted avg pool" stands for weighted average [patch] token with weights from the self-attention module from the last block of ViT. The last four rows represent concatenation of these three tokens. All tokens are normalized to unit length. Results are evaluated by Prototype classification method. + +
Embedding1-shot5-shot
①cls71.26 ± 0.1788.03 ± 0.10
②avg pool71.65 ± 0.1986.40 ± 0.11
③weighted avg pool71.83 ± 0.1986.38 ± 0.11
① + ②73.83 ± 0.1888.58 ± 0.10
① + ③74.28 ± 0.1888.82 ± 0.09
② + ③71.86 ± 0.1986.47 ± 0.11
① + ② + ③73.80 ± 0.1888.26 ± 0.10
+ +# 5. Conclusion + +In this work, we propose a novel supervised knowledge distillation framework (SMKD) for few-shot Transformers, which extends the self-supervised masked knowledge distillation framework into the supervised setting. With our design of supervised-contrastive losses, we incorporate supervision into both class and patch-level knowledge distillation while still enjoy the benefits of not needing large batch size and negative samples. Evaluation results together with ablation studies demonstrate the superiority of our method given its simple design compared with contemporary works. Our two-stage training is a special case of curriculum learning from the easy samples to the hard ones. We can unify the learning objectives of self-supervised learning and supervised contrastive learning, using a carefully designed curriculum learning strategy for future works. We hope our work can bridge the gap between self-supervised and supervised knowledge distillation, and inspire more works on supervised few-shot learning methods. + +# 6. Acknowledgements + +This research is sponsored by Air Force Research Laboratory (AFRL) under agreement number FA8750-19-1-1000. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation therein. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of Air Force Laboratory, DARPA or the U.S. Government. + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3,4 +[2] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4 +[6] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 3 +[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4 +[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2 +[9] Zhengyu Chen, Jixie Ge, Heshen Zhan, Siteng Huang, and Donglin Wang. Pareto self-supervised training for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13663-13672, 2021. 6 +[10] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 1 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[13] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. Advances in Neural Information Processing Systems, 33:21981-21993, 2020. 3 +[14] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6 + +[15] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3 +[17] Yutong Feng, Jianwen Jiang, Mingqian Tang, Rong Jin, and Yue Gao. Rethinking supervised pre-training for better downstream transferring. arXiv preprint arXiv:2110.06014, 2021.3 +[18] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? arXiv preprint arXiv:2210.07240, 2022. 2, 3 +[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8059-8068, 2019. 1 +[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4 +[21] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 3 +[22] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3263-3272, October 2021. 3 +[23] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 780-789, 2022. 3 +[24] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, June 2022. 3 +[25] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, Rama Chellappa, and Shih-Fu Chang. Multimodal few-shot object detection with meta-learning based cross-modal prompting. arXiv preprint arXiv:2204.07841, 2022. 3 +[26] Guangxing Han, Xuan Zhang, and Chongrong Li. Revisiting faster r-cnn: A deeper look at region proposal network. In International Conference on Neural Information Processing, pages 14-24, 2017. 1 + +[27] Guangxing Han, Xuan Zhang, and Chongrong Li. Single shot object detection with top-down refinement. In 2017 IEEE International Conference on Image Processing (ICIP), pages 3360-3364. IEEE, 2017. 1 +[28] Guangxing Han, Xuan Zhang, and Chongrong Li. Semi-supervised diff: Decoupling detection and feature flow for video object detectors. In Proceedings of the 26th ACM international conference on Multimedia, pages 1811-1819, 2018. 1 +[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3 +[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4 +[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[32] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. arXiv preprint arXiv:2206.07267, 2022. 2, 3, 6, 7 +[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2 +[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[35] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1 +[36] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 1 +[37] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34:23818-23830, 2021. 1, 2, 3 +[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 3 +[39] Jiang Lu, Pinghua Gong, Jieping Ye, and Changshui Zhang. Learning from very few samples: A survey. arXiv preprint arXiv:2009.02653, 2020. 1 +[40] Yuning Lu, Liangjian Wen, Jianzhuang Liu, Yajing Liu, and Xinmei Tian. Self-supervision can be a good few-shot learner. arXiv preprint arXiv:2207.09176, 2022. 2 +[41] Xu Luo, Longhui Wei, Liangjian Wen, Jinrong Yang, Lingxi Xie, Zenglin Xu, and Qi Tian. Rectifying the shortcut learning of background: Shared object concentration for few-shot + +image recognition. arXiv preprint arXiv:2107.07746, 2021. 6 +[42] Jiawei Ma, Hanchen Xie, Guangxing Han, Shih-Fu Chang, Aram Galstyan, and Wael Abd-Almageed. Partner-assisted learning for few-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10573-10582, 2021. 1, 3, 4 +[43] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2218-2227, 2020. 2, 3, 5 +[44] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5 +[45] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. arXiv preprint arXiv:2210.10615, 2022. 2, 3 +[46] Elia Peruzzo, Enver Sangineto, Yahui Liu, Marco De Nadai, Wei Bi, Bruno Lepri, and Nicu Sebe. Spatial entropy regularization for vision transformers. arXiv preprint arXiv:2206.04636, 2022. 2, 3 +[47] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8412-8422, 2021. 6 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1 +[51] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5 +[52] Mamshad Nayeem Rizve, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Exploring complementary strengths of invariant and equivariant representations for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10836-10846, June 2021. 6 +[53] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 3 + +[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5 +[55] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1 +[56] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 3 +[57] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 4 +[58] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282. Springer, 2020. 3 +[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 3 +[60] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 8 +[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[62] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 1, 3, 5 +[63] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 3 +[64] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3 +[65] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 5 +[66] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 3 + +[67] Jiamin Wu, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Task-aware part mining network for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8433-8442, 2021. 6 +[68] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2 +[69] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 8 +[70] Dongyang Zhao Hong-Yu Zhou Weifeng Ge Yizhou Yu Wenqiang Zhang Yangji He, Weihan Liang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning, 2022. 1, 2, 3, 6, 7 +[71] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8808-8817, 2020. 3 +[72] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The met dataset: Instance-level recognition for artworks. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 3 +[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 4 +[74] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 6 +[75] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 6 +[76] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? arXiv preprint arXiv:2006.06606, 2020. 4 +[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2, 3, 4, 5 +[78] Ziqi Zhou, Xi Qiu, Jiangtao Xie, Jianan Wu, and Chi Zhang. Binocular mutual learning for improving few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8402-8411, 2021. 6 +[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1 \ No newline at end of file diff --git a/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/images.zip b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..76c7a9dc7b1bcca19aa259c99903ca693875d925 --- /dev/null +++ b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3232d948c0b22d3546de1b6b63f08eec3dc79215d1a7e22d398753cc3362b729 +size 635582 diff --git a/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/layout.json b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9feca7e125b2ebe43f152ac1226af4423ddcebfc --- /dev/null +++ b/2023/Supervised Masked Knowledge Distillation for Few-Shot Transformers/layout.json @@ -0,0 +1,9075 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "Supervised Masked Knowledge Distillation for Few-Shot Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 143, + 510, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 143, + 510, + 177 + ], + "spans": [ + { + "bbox": [ + 80, + 143, + 510, + 177 + ], + "type": "text", + "content": "Han Lin*, Guangxing Han\\*†, Jiawei Ma, Shiyuan Huang, Xudong Lin, Shih-Fu Chang Columbia University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 180, + 465, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 180, + 465, + 192 + ], + "spans": [ + { + "bbox": [ + 129, + 180, + 465, + 192 + ], + "type": "text", + "content": "{h13199, gh2561, jiawei.m, sh3813, x12798, sc250}@columbia.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 220, + 192, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 220, + 192, + 232 + ], + "spans": [ + { + "bbox": [ + 143, + 220, + 192, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 246, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 246, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 45, + 246, + 290, + 568 + ], + "type": "text", + "content": "Vision Transformers (ViTs) emerge to achieve impressive performance on many data-abundant computer vision tasks by capturing long-range dependencies among local features. However, under few-shot learning (FSL) settings on small datasets with only a few labeled data, ViT tends to overfit and suffers from severe performance degradation due to its absence of CNN-alike inductive bias. Previous works in FSL avoid such problem either through the help of self-supervised auxiliary losses, or through the dextile uses of label information under supervised settings. But the gap between self-supervised and supervised few-shot Transformers is still unfilled. Inspired by recent advances in self-supervised knowledge distillation and masked image modeling (MIM), we propose a novel Supervised Masked Knowledge Distillation model (SMKD) for few-shot Transformers which incorporates label information into self-distillation frameworks. Compared with previous self-supervised methods, we allow intra-class knowledge distillation on both class and patch tokens, and introduce the challenging task of masked patch tokens reconstruction across intra-class images. Experimental results on four few-shot classification benchmark datasets show that our method with simple design outperforms previous methods by a large margin and achieves a new start-of-the-art. Detailed ablation studies confirm the effectiveness of each component of our model. Code for this paper is available here: https://github.com/HL-hanlin/SMKD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 598, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 694 + ], + "type": "text", + "content": "Vision Transformers (ViTs) [16] have emerged as a competitive alternative to Convolutional Neural Networks (CNNs) [31] in recent years, and have achieved impressive performance in many vision tasks including image classification [16, 38, 59, 66], object detection [3, 10, 26-28, 79], and object segmentation [50, 55]. Compared with CNNs, which introduce inductive bias through convolutional kernels with fixed receptive fields [35], the attention layers in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 230, + 528, + 278 + ], + "blocks": [ + { + "bbox": [ + 308, + 219, + 477, + 228 + ], + "lines": [ + { + "bbox": [ + 308, + 219, + 477, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 477, + 228 + ], + "type": "text", + "content": "(a) Few-Shot Transformers with Soft Label Regularization" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 230, + 528, + 278 + ], + "lines": [ + { + "bbox": [ + 329, + 230, + 528, + 278 + ], + "spans": [ + { + "bbox": [ + 329, + 230, + 528, + 278 + ], + "type": "image", + "image_path": "a569f99be3d3fbeffe35c16e1a351519717b86e1318889b45f343c6814f476a6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 292, + 538, + 342 + ], + "blocks": [ + { + "bbox": [ + 308, + 281, + 523, + 291 + ], + "lines": [ + { + "bbox": [ + 308, + 281, + 523, + 291 + ], + "spans": [ + { + "bbox": [ + 308, + 281, + 523, + 291 + ], + "type": "text", + "content": "(b) Few-Shot Transformers with Self-Supervision Auxiliary Regularization" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 292, + 538, + 342 + ], + "lines": [ + { + "bbox": [ + 323, + 292, + 538, + 342 + ], + "spans": [ + { + "bbox": [ + 323, + 292, + 538, + 342 + ], + "type": "image", + "image_path": "7639a8994228dc8bd335e0f4d910b602ff0d7cb61cb44573f51d684dffe1efa0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 361, + 543, + 418 + ], + "blocks": [ + { + "bbox": [ + 308, + 348, + 541, + 357 + ], + "lines": [ + { + "bbox": [ + 308, + 348, + 541, + 357 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 541, + 357 + ], + "type": "text", + "content": "(c) Few-Shot Transformers with Our Supervised Masked Knowledge Distillation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 361, + 543, + 418 + ], + "lines": [ + { + "bbox": [ + 309, + 361, + 543, + 418 + ], + "spans": [ + { + "bbox": [ + 309, + 361, + 543, + 418 + ], + "type": "image", + "image_path": "7a8d7ae397840fc6de041cf50d80ee656ca1e14a0d3bdbbb02aa83c5bbd202f1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 422, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 487 + ], + "type": "text", + "content": "Figure 1. Comparison of the proposed idea and other existing methods for few-shot Transformers. Our model mitigates the overfitting issue of few-shot Transformers, by extending the masked knowledge distillation framework into the supervised setting, and enforcing the alignment of [cls] and corresponding [patch] tokens for intra-class images." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 640 + ], + "type": "text", + "content": "ViT allow it to model global token relations to capture long-range token dependencies. However, such flexibility also comes at a cost: ViT is data-hungry and it needs to learn token dependencies purely from data. This property often makes it easy to overfit to datasets with small training set and suffer from severe generalization performance degradation [36, 37]. Therefore, we are motivated to study how to make ViTs generalize well on these small datasets, especially under the few-shot learning (FSL) setting [19, 39, 62] which aims to recognize unseen new instances at test time just from only a few (e.g. one or five) labeled samples from each new categories." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Most of the existing methods mitigate the overfitting issue of few-shot Transformers [14] using various regularizations. For instance, some works utilize label information in a weaker [70], softer [42] way, or use label information efficiently through patch-level supervision [14]. However, these models usually design sophisticated learning targets." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 204, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 204, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 204, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19649" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "On the other hand, self-distillation techniques [4, 8], and particularly, the recent masked self-distillation [29, 45, 77], which distills knowledge learned from an uncorrupted image to the knowledge predicted from a masked image, have lead an emerging trend in self-supervised Transformers in various fields [15, 68]. Inspired by such success, recent works in FSL attempt to incorporate self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46], or to adopt a self-supervised pretraining, supervised training two-stage framework to train few-shot Transformers [18, 32]. Compared with traditional supervised methods, self-supervision can learn less biased representations towards base class, which usually leads to better generalization ability for novel classes [40]. However, the two learning objectives of self-supervision and supervision are conflicting and it is hard to balance them during training. Therefore, how to efficiently leverage the strengths of self-supervised learning to alleviate the overfitting issue of supervised training remains a challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 602 + ], + "type": "text", + "content": "In this work, we propose a novel supervised masked knowledge distillation framework (SMKD) for few-shot Transformers, which handles the aforementioned challenge through a natural extension of the self-supervised masked knowledge distillation framework into the supervised setting (shown in Fig. 1). Different from supervised contrastive learning [33] which only utilizes global image features for training, we leverage multi-scale information from the images (both global [cls] token and local [patch] tokens) to formulate the learning objectives, which has been demonstrated to be effective in the recent self-supervised Transformer methods [29, 77]. For global [cls] tokens, we can simply maximize the similarity for intra-class images. However, it is non-trivial and challenging to formulate the learning objectives for local [patch] tokens because we do not have ground-truth patch-level annotations. To address this problem, we propose to estimate the similarity between [patch] tokens across intra-class images using cross-attention, and enforce the alignment of corresponding [patch] tokens. Particularly, reconstructing masked [patch] tokens across intra-class images increases the difficulty of model learning, thus encouraging learning generalizable few-shot Transformer models by jointly exploiting the holistic knowledge of the image and the similarity of intra-class images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "As shown in Fig. 2, we compare our model with the existing self-supervised/supervised learning methods. Our model is a natural extension of the supervised contrastive learning method [33] and self-supervised knowledge distillation methods [4, 77]. Thus our model inherits both the advantage of method [33] for effectively leveraging label information, and the advantages of methods [4, 77] for not needing large batch size and negative samples. Meanwhile, the newly-introduced challenging task of masked" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 416, + 145 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 416, + 145 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 416, + 145 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 416, + 145 + ], + "type": "image", + "image_path": "5e95c3374f45762922ae8483735494ef364b93840e6e69b793cb0578d89dadb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 419, + 70, + 536, + 145 + ], + "blocks": [ + { + "bbox": [ + 419, + 70, + 536, + 145 + ], + "lines": [ + { + "bbox": [ + 419, + 70, + 536, + 145 + ], + "spans": [ + { + "bbox": [ + 419, + 70, + 536, + 145 + ], + "type": "image", + "image_path": "904899dc3403ebb8f1bbaaff312a687d96a9424860c65ae3953ef0ed91008a7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 150, + 416, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 150, + 416, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 150, + 416, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 416, + 225 + ], + "type": "image", + "image_path": "ad243b514655fef5979eb6795ed347e98214cf1b0b10e98ca2d0f490ef8fc638.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 273 + ], + "type": "text", + "content": "Figure 2. Comparison of other self-supervised/supervised frameworks. Our method (d) is a natural extension of (b) and (c), with the newly-introduced challenging task of masked [patch] tokens reconstruction across intra-class images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 419, + 150, + 542, + 224 + ], + "blocks": [ + { + "bbox": [ + 419, + 150, + 542, + 224 + ], + "lines": [ + { + "bbox": [ + 419, + 150, + 542, + 224 + ], + "spans": [ + { + "bbox": [ + 419, + 150, + 542, + 224 + ], + "type": "image", + "image_path": "101bdef3dc637e02f74da15925811687b9c3e559329269caf6a547509e1845e4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 286, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 320 + ], + "type": "text", + "content": "[patch] tokens reconstruction across intra-class images makes our method more powerful for learning generalizable few-shot Transformer models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 322, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 441 + ], + "type": "text", + "content": "Compared with contemporary works on few-shot Transformers [14, 32, 70], our framework enjoys several good properties from a practical point of view. (1) Our method does not introduce any additional learnable parameters besides the ViT backbone and projection head, which makes it easy to be combined with other methods [32, 70, 74]. (2) Our method is both effective and training-efficient, with stronger performance and less training time on four few-shot classification benchmarks, compared with [32, 70]. In a nutshell, our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 444, + 545, + 623 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 315, + 444, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 444, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 315, + 444, + 545, + 504 + ], + "type": "text", + "content": "- We propose a new supervised knowledge distillation framework (SMKD) that incorporates class label information into self-distillation, thus filling the gap between self-supervised knowledge distillation and traditional supervised learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 504, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 504, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 315, + 504, + 545, + 552 + ], + "type": "text", + "content": "- Within the proposed framework, we design two supervised-contrastive losses on both class and patch levels, and introduce the challenging task of masked patch tokens reconstruction across intra-class images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 552, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 552, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 315, + 552, + 545, + 623 + ], + "type": "text", + "content": "- Given its simple design, we test our SMKD on four few-shot datasets, and show that it achieves a new SOTA on CIFAR-FS and FC100 by a large margin, as well as competitive performance on mini-ImageNet and tiered-ImageNet using the simple prototype classification method for few-shot evaluation." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 633, + 392, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 392, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 392, + 646 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Few-shot Learning. Few-shot learning aims at fast knowledge transfer from seen base classes to unseen novel classes given only a few labeled samples from each novel class. The meta-learning paradigm, which simulates few-shot tasks episodically to mimic the human-learning process in the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19650" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "real world, once dominated FSL [22-25, 44, 53, 54, 56, 62]. Surprisingly, recent works [6, 21, 58, 72] have shown that the state-of-the-art meta-learning methods can be outperformed by simple baseline methods using just a distance-based classifier without the complicated design of meta-learning strategies. Therefore, recent methods [17, 42, 70] in FSL start to focus less on meta-learning, and more on learning embeddings with good generalization ability. Our paper follows this trend and proposes a knowledge distillation framework to learn generalizable embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 373 + ], + "type": "text", + "content": "Vision Transformers in FSL. Variants of Transformer [61] have achieved great success in NLP [12, 49], computer vision [16, 59], and multimodal learning [48]. However, the lack of inductive bias makes Transformer infamous for its data-hungry property, which is especially significant in few-shot settings when the amount of data is limited. A line of works focuses on introducing inductive bias back to the Transformer architecture, including methods using pyramid structure [63, 64], shifted windows [38], and explicit convolutional token embeddings [66]. With this said, some recent works [13, 24, 32, 71] still show that few-shot Transformers have the potential of fast adaptation to novel classes. Our work also studies few-shot Transformers, and we show that our method can work well even on the vanilla ViT structure without explicit incorporation of inductive bias." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 700 + ], + "type": "text", + "content": "Self-Supervision for FSL. Self-supervised learning (SSL) has shown great potential in FSL due to its good generalization ability to novel classes. Previous methods incorporate SSL into FSL in various ways. Some works propose to include self-supervised pretext tasks into the standard supervised learning through auxiliary losses [37, 43, 46]. For example, [37] proposed a regularization loss that extracts additional information from images by predicting the geometric distance between patch tokens. [46] designed their loss by discouraging spatially disordered attention maps based on the idea that objects usually occupy connected regions, and [43] derived their auxiliary loss from self-supervision tasks of rotation and exemplar. Some other works [18, 32], instead, adopt a two-stage procedure by pretraining a model via self-supervision before supervised training. [32] takes advantage of self-supervised training with iBOT [77] as a pretext task, and then uses inner loop token importance reweighting for supervised fine-tuning. [18] initializes its model with a pre-trained self-supervised DINO [4], then trained with supervised cross-entropy loss. Our work follows the second branch of work, but the difference is obvious from previous works. Instead of designing complicated training pipeline or leveraging additional learnable modules at inference time, we use supervised training on self-supervised pre-trained model, and focus mainly on bridging the gap between self-supervised and supervised knowledge distillation with minimum extra design." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Another emerging trend in SSL is Masked Image Mod" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "eling (MIM) [1, 29, 45, 77], which aims at recovering the patch-level target (e.g. image pixels, patch features) of the masked content in a corrupted input image. In iBOT [77], the class and patch tokens share the same projection head, thus helping the ViT backbone to learn semantically meaningful patch embeddings, which can be used as a supplement to image-level global self-supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 171, + 397, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 171, + 397, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 397, + 185 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 194, + 547, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 278 + ], + "type": "text", + "content": "We present our Supervised Masked Knowledge Distillation (SMKD) framework in this section. Fig. 3 shows an overview of our model. Task definition for few-shot classification is presented in Sec. 3.1. We explain the formulation of SSL knowledge distillation in Sec. 3.2, and how we extend it into our supervised knowledge distillation framework in Sec. 3.3. Training pipeline is given in Sec. 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 289, + 446, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 446, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 446, + 303 + ], + "type": "text", + "content": "3.1. Learning Task Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": "In few-shot classification, we are given two datasets " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{novel}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " with disjoint class labels " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{base}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{novel}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{base} \\cap \\mathcal{C}_{novel} = \\emptyset" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": "). The base dataset " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": ", which contains abundant labeled samples, is used to train the feature extractor. And the novel dataset " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{novel}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " is then used to sample episodes for prototype estimation and few-shot evaluation. In a standard N-way K-shot task, each episode " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{epi} = (\\mathcal{D}_S, \\mathcal{D}_Q)" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " covers N classes from " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{novel}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": ". The support dataset " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": ", which contains K samples from each class, is used for class prototype estimation, and the query dataset " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " is then used for evaluation. This task aims at correctly classifying " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_Q" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": " into N classes from sampled episodes. And the main focus of this paper is to train a feature extractor with good generalization ability from " + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{base}" + }, + { + "bbox": [ + 304, + 310, + 547, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 490, + 545, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 490, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 545, + 503 + ], + "type": "text", + "content": "3.2. Preliminary: SSL with Knowledge Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "Our work is inspired by the self-supervised knowledge distillation frameworks proposed recently [4]. Specifically, given an input image " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " uniformly sampled from the training set " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": ", random data augmentations are applied to generate two augmented views " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}^1" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}^2" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " (we represent image view indices as superscripts and patch indices as subscripts), which are then fed into the teacher and student networks. The student network, parameterized by " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": ", consists of an encoder with ViT backbone and a projection head with a 3-layer multi-layer perceptron (MLP) followed by " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "-normalized bottleneck. The ViT backbone first generates a [cls] token, which is then entered into the projection head and outputs a probability distribution " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "P_{s}^{[\\mathrm{cls}]}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " over K dimensions. The teacher network, parameterized by " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": ", is Exponentially Moving Averaged (EMA) updated by the student network " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": ", which distills its knowledge to the student by minimizing the cross-entropy loss over the outputs of the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19651" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 541, + 210 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 541, + 210 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 541, + 210 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 541, + 210 + ], + "type": "image", + "image_path": "2d4f90643870ab2108a670f8c23a0de7f96160d43079a1eaf2d2306d8148aea0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "type": "text", + "content": "Figure 3. Overview of our SMKD framework. Two views are generated from a pair of images sampled (with replacement) from the same class. The first view, applied with random blockwise masking, is passed to the student network, while the second unmasked view is passed to the teacher network. Both networks consist of a ViT backbone and a projection head. The parameters of the teacher network are Exponentially Moving Averaged (EMA) updated by the student network. SMKD distills knowledge between intra-class cross-views on both class and patch levels. " + }, + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[cls]}" + }, + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "type": "text", + "content": " distills knowledge from [cls] tokens, while " + }, + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[patch]}" + }, + { + "bbox": [ + 46, + 212, + 547, + 278 + ], + "type": "text", + "content": " distills knowledge from [patch] tokens by finding dense correspondence of matched token pairs (connected by red dashed lines) with highest similarities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 288, + 260, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 288, + 260, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 260, + 300 + ], + "type": "text", + "content": "categorical distributions from their projection heads:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 304, + 287, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 304, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 82, + 304, + 287, + 321 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {[ \\mathrm {c l s} ]} = \\mathcal {H} (\\boldsymbol {P} _ {t} ^ {[ \\mathrm {c l s} ]} (\\boldsymbol {x} ^ {1}), \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\mathrm {c l s} ]} (\\boldsymbol {x} ^ {2})) \\qquad (1)", + "image_path": "bba51d8877ebf93cf4f24dba8f7adab03b7b88eccafd66b5f591c557634d4ca0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 326, + 160, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 160, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 160, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 326, + 160, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(x,y) = -x\\log y" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": "Masked Image Modeling (MIM) [1, 57] can be performed via self-distillation as follows [77]. Given a randomly sampled mask sequence " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": " over an image with " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": " [patch] tokens " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "\\pmb{x} = \\{\\pmb{x}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": "'s with " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "m_i = 1" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": " are then replaced by a learnable token embedding " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "e_{[\\text{MASK}]}" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": ", which results in a corrupted image " + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}} = \\{\\widehat{\\pmb{x}}_i\\}_{i=1}^N = \\{(1 - m_i)\\pmb{x}_i + m_i e_{[\\text{MASK}]}\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 338, + 288, + 494 + ], + "type": "text", + "content": ". This corrupted image and the original uncorrupted image are fed into the student and teacher networks respectively. The objective of MIM is to recover the masked tokens from the corrupted image, which is equivalent to minimizing the cross-entropy loss between the outputs of the categorical distributions of the student and teacher networks on masked patches:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 499, + 287, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 287, + 532 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M I M}} = \\sum_ {i = 1} ^ {N} m _ {i} \\cdot \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\text {p a t c h} ]} (\\boldsymbol {x}) _ {i}, \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\text {p a t c h} ]} (\\widehat {\\boldsymbol {x}}) _ {i}\\right) \\tag {2}", + "image_path": "547b4361bf10d96bd9df133b4d1a4f0f8add3b1eac328be1442ffc633a471bd3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 540, + 271, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 271, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 271, + 553 + ], + "type": "text", + "content": "3.3. Supervised Masked Knowledge Distillation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "text", + "content": "Distill the Class Tokens. Recent self-distillation frameworks [4, 20, 77] distill knowledge on [cls] tokens from cross-view images via Eq.(1). To incorporate label information into such self-supervised frameworks, we further allow knowledge on [cls] tokens to be distilled from intra-class cross-views. This can be achieved by a small extension of the way we sample images. Rather than sampling a single image " + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\sim \\mathcal{I}" + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "text", + "content": " and generating two views, now we sample two images " + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}, \\boldsymbol{y} \\sim \\mathcal{I}^c" + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "text", + "content": " (with replacement) and generate one view for each of them1. " + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^c \\subseteq \\mathcal{I}" + }, + { + "bbox": [ + 46, + 559, + 287, + 678 + ], + "type": "text", + "content": " here denotes the set" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": "of images with the same class label " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " in the training set " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": ". Specifically, we denote " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\prime}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " as the augmented views generated from image " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " respectively. We apply additional random blockwise masking on " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " and denote the resulting corrupted view as " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}^{\\prime}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": ". Then the corrupted view " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}^{\\prime}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " and uncorrupted view " + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\prime}" + }, + { + "bbox": [ + 304, + 287, + 545, + 384 + ], + "type": "text", + "content": " are sent to the student and teacher networks separately. Now our supervised-contrastive loss on [cls] tokens becomes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 391, + 545, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 391, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 342, + 391, + 545, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {[ \\mathrm {c l s} ]} = \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\mathrm {c l s} ]} \\left(\\boldsymbol {y} ^ {\\prime}\\right), \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\mathrm {c l s} ]} \\left(\\hat {\\boldsymbol {x}} ^ {\\prime}\\right)\\right) \\tag {3}", + "image_path": "13ce48cdaa602b922297be2e24a376966c1ca3b21e95db92c7738bcf588b38d8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{x}, \\pmb{y}" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": " are sampled as the same image " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "(\\pmb{x} = \\pmb{y})" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": ", our loss performs masked self-distillation across two views of the same image similar to Eq.(1). In the other situation when " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": " represent different images " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "(\\pmb{x} \\neq \\pmb{y})" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": ", our loss then minimizes the cross-entropy loss of projected [c1s] tokens among all cross-view pairs between images " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "Such design has two major advantages. (1) It can be implemented efficiently. Instead of intentionally sampling image pairs from the same class, we just need to look through the images in a mini-batch, find image pairs belonging to the same class, and then apply our loss in Eq.(3). (2) Unlike previous works that use either supervised [42,76] or self-supervised [5, 30] contrastive losses, our method follows the recent trend in SSL works [4, 7, 20, 73] and avoids the need for negative examples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Distill the Patch Tokens. Beyond the knowledge distillation on the global [cls] tokens, we introduce the challenging task of masked patch tokens reconstruction across intra-class images, to fully exploit the local details of the image for training. Our main intuition here is based on the following hypothesis: for intra-class images, even though their semantic information can be vastly different at the patch level, there should at least exist some patches that share similar semantic meanings." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": "1For simplicity of illustration, we show our method with only one augmented view from each image. We sample two views from an image in our implementation. Loss is then averaged over all cross-view pairs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19652" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "For each patch " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " from an input view " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{\\prime}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " sent to the teacher network (with its corresponding token embedding defined as " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{k}^{t}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "), we need to first find the most similar patch " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "k^{+}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " from the masked view " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\widehat{\\pmb{x}}^{\\prime}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " of the student network (with its corresponding token embedding defined as " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{k^{+}}^{s}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "), and then perform knowledge distillation between these two matched tokens. The token embedding " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{k^{+}}^{s}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " represents either the " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "k^{+th}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " [patch] token if patch " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "k^{+}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " is unmasked, or the reconstructed patch token if it is masked." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 182, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 287, + 217 + ], + "type": "text", + "content": "As we do not have any patch-level annotations, we use cosine similarity to find the best-matched patch of " + }, + { + "bbox": [ + 47, + 182, + 287, + 217 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 182, + 287, + 217 + ], + "type": "text", + "content": " among all the [patch] tokens in the student network:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 223, + 287, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 223, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 111, + 223, + 287, + 255 + ], + "type": "interline_equation", + "content": "k ^ {+} = \\arg \\max _ {l \\in [ N ]} \\frac {\\boldsymbol {f} _ {k} ^ {t} {} ^ {\\top} \\boldsymbol {f} _ {l} ^ {s}}{\\| \\boldsymbol {f} _ {k} ^ {t} \\| \\| \\boldsymbol {f} _ {l} ^ {s} \\|} \\tag {4}", + "image_path": "ca9d6ce46ff4704dc7a0eac259cd895500bfe892302929583fab2dd4a9296b23.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 262, + 281, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 281, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 281, + 274 + ], + "type": "text", + "content": "Our patch-level knowledge distillation loss now becomes:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 278, + 287, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 287, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {[ \\text {p a t c h} ]} = \\sum_ {k = 1} ^ {N} \\omega_ {k +} \\cdot \\mathcal {H} \\left(\\boldsymbol {P} _ {\\boldsymbol {t}} ^ {[ \\text {p a t c h} ]} \\left(\\boldsymbol {y} ^ {\\prime}\\right) _ {k}, \\boldsymbol {P} _ {\\boldsymbol {s}} ^ {[ \\text {p a t c h} ]} \\left(\\hat {\\boldsymbol {x}} ^ {\\prime}\\right) _ {k +}\\right) \\tag {5}", + "image_path": "6cc86952f62269c0edc00134545c4b796dcb77059a5415437746d6998fd56c7b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\omega_{k+}" + }, + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "text", + "content": " is a scalar representing the weight we give to each loss term. We find that taking it as a constant value is effective enough. The ablation study of more complex designs of " + }, + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\omega_{k+}" + }, + { + "bbox": [ + 47, + 327, + 287, + 374 + ], + "type": "text", + "content": " is given in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 459 + ], + "type": "text", + "content": "Our loss shares some similarities with DenseCL [65]. However, the differences are also obvious: (1) Our loss serves as an extension of their self-supervised contrastive loss into a supervised variant. (2) We further incorporate MIM into our design and allow masked patches to be matched, which makes our task harder and leads to more semantically meaningful patch embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 466, + 151, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 151, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 151, + 479 + ], + "type": "text", + "content": "3.4. Training Pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": "We train our model in two stages: self-supervised pretraining and supervised training. In the first stage, we use the recently proposed MIM framework [77] for self-supervised pretraining. The self-supervised loss is the summation of " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{cls}]}" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MIM}}" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": " in Eq.(1) and Eq.(2) without scaling. In the second stage, we continue training the pre-trained model using our supervised-contrastive losses " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{cls}]}" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": " from Eq.(3) and Eq.(5). We define our training loss as: " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{[\\mathrm{cls}]} + \\lambda \\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": " controls the relative scale of these two components. A relatively large " + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 484, + 287, + 628 + ], + "type": "text", + "content": " will make our model focus more on localization and less on high-level semantics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 639, + 123, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 123, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 123, + 653 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 659, + 110, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 110, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 110, + 670 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Our model is evaluated on four widely used and publicly available few-shot classification datasets: mini-ImageNet [62], tiered-ImageNet [51], CIFAR-FS [2], and FC100" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "[44]. mini-ImageNet and tiered-ImageNet are derived from ImageNet [11], while CIFAR-FS and FC100 are derived from CIFAR100 [34]. mini-ImageNet contains 100 classes, which are randomly split into 64 base classes for training, 16 classes for few-shot validation, and 20 classes for few-shot evaluation. There are 600 images in each class. tiered-ImageNet contains 609 classes with 779165 images in total. The class split for training, few-shot validation and few-shot evaluation are 351, 97, and 160. CIFAR-FS contains 100 classes with class split as 64, 16, and 20. FC100 contains 100 classes with class split as 60, 20, and 20. For both of these two datasets, each class has 600 images with smaller resolutions " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " compared with ImageNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 238, + 440, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 440, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 440, + 251 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 257, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 376 + ], + "type": "text", + "content": "Self-supervised pretraining. We pre-train our Vision Transformer backbone and projection head following the same pipeline in iBOT [77]. Most of the hyper-parameter settings are kept unchanged without tuning. We use a batch size of 640 and a learning rate of 0.0005 decayed with the cosine schedule. mini-[ImageNet and tiered-[ImageNet are pre-trained for 1200 epochs, and CIFAR-FS and FC100 are pre-trained for 900 epochs. All models are trained on 8 Nvidia RTX 3090 GPUs. Detailed training parameter settings are included in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "content": "Supervised knowledge distillation After getting the pre-trained model, we train it with our supervised-contrastive losses. We find that the model with the best performance can converge within 60 epochs in the validation set. We use the same batch size and learning rate as in the pretraining stage. Compared with the first stage, the only extra hyper-parameter is the scaling parameter " + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "content": ". We set it to make the ratio between " + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{cls}]}" + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "content": " roughly around 2. Ablation over different " + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 378, + 546, + 485 + ], + "type": "text", + "content": " is given in the appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Few-shot Evaluation We use the simple prototype classification method (Prototype) [44, 54] and the linear classifier (Classifier) used in S2M2 [43] as our default methods for few-shot evaluation. For each sampled episode data in an N-way K-shot task, Prototype first estimates each class prototype using the averaged feature over K support samples. Then a new sampled query image is classified into one of the N classes which has the highest cosine similarity between its feature vector and the class prototype. Instead, Classifier trains a linear classifier from the " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{N}\\times \\mathrm{K}" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": " support samples, which is then used to classify new query samples. More complex evaluation methods (e.g. DeepEMD) are also compatible with our framework. The feature we used for evaluation is the concatenation of the [cls] token with the weighted average [patch] token (weighted avg pool). The weights of weighted avg pool is the average of the self-attention values of the [cls] token with all heads of the last attention layer. Ablation over different choices of features for evaluation is studied in Sec. 4.4." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19653" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 95, + 531, + 251 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Results on mini-ImageNet and tiered-ImageNet. Top three methods are colored in red, blue and green respectively. A more comprehensive version of this table is shown in the appendix." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 95, + 531, + 251 + ], + "lines": [ + { + "bbox": [ + 66, + 95, + 531, + 251 + ], + "spans": [ + { + "bbox": [ + 66, + 95, + 531, + 251 + ], + "type": "table", + "html": "
MethodBackbone#ParamsminiImageNet,5-waytieredImageNet,5-way
1-shot5-shot1-shot5-shot
DeepEMD [74]ResNet-1212.4M65.91 ± 0.8282.41 ± 0.5671.16 ± 0.8786.03 ± 0.58
IE [52]ResNet-1212.4M67.28 ± 0.8084.78 ± 0.5272.21 ± 0.9087.08 ± 0.58
COSOC [41]ResNet-1212.4M69.28 ± 0.4985.16 ± 0.4273.57 ± 0.4387.57 ± 0.10
Meta-QDA [75]WRN-28-1036.5M67.38 ± 0.5584.27 ± 0.7574.29 ± 0.6689.41 ± 0.77
OM [47]WRN-28-1036.5M66.78 ± 0.3085.29 ± 0.4171.54 ± 0.2987.79 ± 0.46
SUN [14]ViT12.5M67.80 ± 0.4583.25 ± 0.3072.99 ± 0.5086.74 ± 0.33
FewTURE [32]Swin-Tiny29.0M72.40 ± 0.7886.38 ± 0.4976.32 ± 0.8789.96 ± 0.55
HCT Transformers [70]3× ViT-S63.0M74.74 ± 0.1789.19 ± 0.1379.67 ± 0.2091.72 ± 0.11
Ours (Prototype)ViT-S21M74.28 ± 0.1888.82 ± 0.0978.83 ± 0.2091.02 ± 0.12
Ours (Classifier)ViT-S21M74.10 ± 0.1788.89 ± 0.0978.81 ± 0.2191.21 ± 0.11
Ours + HCT [70]3× ViT-S63M75.32 ± 0.1889.57 ± 0.0979.74 ± 0.2091.68 ± 0.11
", + "image_path": "b8e7b2d7097df8ab7b8d4a500df1ac669e54f5f07521daa70b6de30f4ca4b43b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 271, + 531, + 415 + ], + "blocks": [ + { + "bbox": [ + 88, + 258, + 504, + 270 + ], + "lines": [ + { + "bbox": [ + 88, + 258, + 504, + 270 + ], + "spans": [ + { + "bbox": [ + 88, + 258, + 504, + 270 + ], + "type": "text", + "content": "Table 2. Results on CIFAR-FS and FC100. A more comprehensive version of this table is shown in the appendix." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 271, + 531, + 415 + ], + "lines": [ + { + "bbox": [ + 66, + 271, + 531, + 415 + ], + "spans": [ + { + "bbox": [ + 66, + 271, + 531, + 415 + ], + "type": "table", + "html": "
MethodBackbone#ParamsCIFAR-FS,5-wayFC100,5-way
1-shot5-shot1-shot5-shot
BML [78]ResNet-1212.4M73.45 ± 0.4788.04 ± 0.3345.00 ± 0.4163.03 ± 0.41
IE [52]ResNet-1212.4M77.87 ± 0.8589.74 ± 0.5747.76 ± 0.7765.30 ± 0.76
TPMN [67]ResNet-1212.4M75.50 ± 0.9087.20 ± 0.6046.93 ± 0.7163.26 ± 0.74
PSST [9]WRN-28-1036.5M77.02 ± 0.3888.45 ± 0.35--
Meta-QDA [75]WRN-28-1036.5M75.95 ± 0.5988.72 ± 0.79--
SUN [14]ViT12.5M78.37 ± 0.4688.84 ± 0.32--
FewTURE [32]Swin-Tiny29.0M77.76 ± 0.8188.90 ± 0.5947.68 ± 0.7863.81 ± 0.75
HCT Transformers [70]3× ViT-S63.0M78.89 ± 0.1890.50 ± 0.0948.27 ± 0.1566.42 ± 0.16
Ours (Prototype)ViT-S21M80.08 ± 0.1890.63 ± 0.1350.38 ± 0.1668.37 ± 0.16
Ours (Classifier)ViT-S21M79.82 ± 0.1890.91 ± 0.1350.28 ± 0.1668.50 ± 0.16
", + "image_path": "a35d486531feb7b13dcfa73dfce881ca5346d001545e5bf3c7d2b219c6a0ad4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 427, + 287, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 287, + 441 + ], + "type": "text", + "content": "4.3. Comparison With the State-of-the-arts (SOTAs)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 544 + ], + "type": "text", + "content": "We evaluate our proposed SMKD with the above-mentioned evaluation methods on four few-shot classification datasets. Our goal in this section is to prove the effectiveness of our method given its simple training pipeline and evaluation procedure. This being said, the performance of our method can be further boosted by adopting strategies from contemporary works (see the last row \"Ours+HCT [70]\" in Table 1 and detailed results in Table 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Compared with traditional convolutional backbones, Transformer-based models are still underdeveloped in few-shot classification. Recent methods with Transformer backbones [14, 32, 70] differ a lot in their training pipelines and evaluation procedures. SUN [14] first pre-trains a teacher network with supervised loss, then uses it to generate patch-level supervision for the student network as a supplement to class-level supervision. HCTransformers [70] converts class-level prediction into a latent prototype learning problem, and introduces spectral tokens pooling to merge neighboring tokens with similar semantic meanings adaptively. FewTURE [32], which adopts a similar two-stage training pipeline as ours, shows the effectiveness of inner loop token importance reweighting to avoid supervision collapse." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 428, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 644 + ], + "type": "text", + "content": "Despite the success of these works, our SMKD still shows competitive performance given its simple design. Table 1 contains the results of mini-ImageNet and tiered-ImageNet. With Prototype and Classifier as the evaluation method, our proposed SMKD outperforms all methods with ResNet and WRN backbones, and ranked second among methods with Transformer backbones. By adopting the same strategies (patch size of 8 and spectral tokens pooling) as in HCTransformers [70], our method achieves a new SOTA on mini-ImageNet. The effect of each of these two strategies is shown in Table 3. Furthermore, as displayed in Table 2, our method performs the best on the two small-resolution datasets (CIFAR-FS and FC100), and outperforms all previous results by great margins (0.93% for 1-shot and 0.41% for 5-shot on CIFAR-FS, and over 2% for both 1 and 5-shot on FC100). In conclusion, our method grows more effective on datasets with smaller resolutions and fewer training images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "The top competitor to our model is HCTransformer [70]. However, their method is more complex compared with ours in the following two aspects: (1) Their method uses a smaller patch size of 8 rather than 16 (which is used in most other methods as well as ours). As shown in Table" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19654" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 106, + 289, + 181 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 103 + ], + "type": "text", + "content": "Table 3. Results for small patch size and spectral tokens pooling. Results are evaluated on mini-ImageNet with Prototype and Classifier methods, and the best is reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 106, + 289, + 181 + ], + "lines": [ + { + "bbox": [ + 50, + 106, + 289, + 181 + ], + "spans": [ + { + "bbox": [ + 50, + 106, + 289, + 181 + ], + "type": "table", + "html": "
MethodPatch 8Spec. Pool1-shot5-shot
HCT [70]✓71.27 ± 0.1784.68 ± 0.10
HCT [70]✓✓74.62 ± 0.2089.19 ± 0.13
Ours74.28 ± 0.1888.82 ± 0.09
Ours✓74.00 ± 0.1789.57 ± 0.09
Ours✓✓75.32 ± 0.1889.23 ± 0.18
", + "image_path": "ddc14a91c6e48db86f840bda0caac3c886b2b9bab4c0338c266907aa9c686147.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": "4, this makes their training procedure " + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 195, + 287, + 266 + ], + "type": "text", + "content": " slower than ours in each epoch. (2) Their method contains three sets of cascaded transformers, each corresponding to a set of teacher-student network with ViT-S backbone. This triples the total number of learnable parameters (63M versus 21M in ours) and incurs much longer inference time." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 312, + 287, + 360 + ], + "blocks": [ + { + "bbox": [ + 47, + 277, + 286, + 309 + ], + "lines": [ + { + "bbox": [ + 47, + 277, + 286, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 286, + 309 + ], + "type": "text", + "content": "Table 4. Training clock time comparison on mini-ImageNet. For HCT, we report the result provided in their paper. All experiments are conducted on 8 Nvidia RTX 3090 GPUs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 312, + 287, + 360 + ], + "lines": [ + { + "bbox": [ + 50, + 312, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 50, + 312, + 287, + 360 + ], + "type": "table", + "html": "
HCT [70]Stage1Stage2Stage3
21.1h (400 epoch)0.58h (2 epoch)0.21h (2 epoch)
OursSelf-Supervised Pre-trainSupervised Training
15.0h (1200 epoch)1.08h (60 epoch)
", + "image_path": "c2eb5dab5addc20ac8c7aa5585bd534cc1b21206842152f3d87363fbb4032530.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 371, + 287, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 442 + ], + "type": "text", + "content": "For a fair comparison, we compare few-shot classification results of methods with the same ViT-S backbone (21M parameters) on mini-ImageNet in Table 5. The results for HCT [70] and FewTURE [32] are copy-pasted from their paper. Our model outperforms them by a large margin when the number of trainable parameters is comparable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 443, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 467 + ], + "type": "text", + "content": "Visualizations of the multi-head self-attention and patch token correspondence are shown in Fig. 4 and Fig. 5." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 52, + 512, + 285, + 567 + ], + "blocks": [ + { + "bbox": [ + 47, + 478, + 287, + 510 + ], + "lines": [ + { + "bbox": [ + 47, + 478, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 287, + 510 + ], + "type": "text", + "content": "Table 5. Comparison results with the same ViT-S backbone on mini-ImageNet. For HCT, we report their result of the first student transformer (training time shown in Stage 1 of Table 4)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 512, + 285, + 567 + ], + "lines": [ + { + "bbox": [ + 52, + 512, + 285, + 567 + ], + "spans": [ + { + "bbox": [ + 52, + 512, + 285, + 567 + ], + "type": "table", + "html": "
MethodBackbone1-shot5-shot
HCT [70]ViT-S71.27 ± 0.1784.68 ± 0.10
FewTURE [32]ViT-S68.02 ± 0.8884.51 ± 0.53
OursViT-S74.28 ± 0.1888.82 ± 0.09
", + "image_path": "e2e8b772251b01846a0b5c5e1ab0e04ce9a46c7867954224b040dc3a69c96f50.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 587, + 140, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 140, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 140, + 600 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 712 + ], + "type": "text", + "content": "Self-Supervised MIM Pretraining Helps. Our training pipeline contains two stages. A natural question is whether the first self-supervised pretraining stage is necessary. We can see from Table 6 that the classification accuracy of our method drops a lot without pretraining. If we train in a single stage, with the combination of the self-supervised learning losses and supervised contrastive learning losses, the model is still hard to converge without a good initialization (1-shot: 43.71 vs. our 74.28. 5-shot: 60.27 vs. our" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 541, + 222 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 541, + 222 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 541, + 222 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 541, + 222 + ], + "type": "image", + "image_path": "9c15894a23af3177d1e84c737d22eec888e62334b442a236058cad160f8a756f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 224, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 291 + ], + "type": "text", + "content": "Figure 4. Visualization of multi-head self-attention maps. The self-attention of the [c1s] tokens with different heads in the last attention layer of ViT are visualized in different colors. iBOT+CE represents the model first pre-trained with iBOT, then trained with CE loss. Our SMKD pays more attention to the foreground objects, especially the most discriminative parts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 303, + 538, + 474 + ], + "blocks": [ + { + "bbox": [ + 315, + 303, + 538, + 474 + ], + "lines": [ + { + "bbox": [ + 315, + 303, + 538, + 474 + ], + "spans": [ + { + "bbox": [ + 315, + 303, + 538, + 474 + ], + "type": "image", + "image_path": "3f784f34866ff047b8c767b7590e38ffe6fc32699d24df9e2c222566df7e32ec.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 479, + 545, + 522 + ], + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 522 + ], + "type": "text", + "content": "Figure 5. Visualization of dense correspondence. We use the patches with the highest self-attention of the [cls] token on each attention head (6 in total) of the last layer of ViT-S as queries. Best matched patches with highest similarities are connected with lines." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "88.82 on mini-[ImageNet]. An explanation is that it will be hard for our patch-level knowledge distillation to find matched patches if the ViT backbone is initialized with random weights. What we observe is that our model finds some shortcut solution that prohibits it from proper training. Additionally, we also test the traditional CE loss with and without self-supervised pretraining. The model pre-trained with self-supervision and then trained with supervised CE loss works surprisingly well and even outperforms most of the methods in Table 1 on mini-[ImageNet]. This further supports the usefulness of our two-stage training procedure. For the masking strategy in the first pretraining stage, our ablation study over three masking types (blockwise, random, and no mask) in the appendix shows that blockwise masking balances the best between 1 and 5-shot classification accuracy." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19655" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 138, + 274, + 206 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 137 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 137 + ], + "type": "text", + "content": "Table 6. Few-shot evaluation w/wo self-supervised pretraining. For models with pretraining stage, we pre-train with iBOT for 1200 epochs, then train with CE or our loss for 60 epochs. For models without pretraining stage, we train with supervised losses for 60 epochs directly. Results are evaluated by Prototype classification method on mini-ImageNet." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 138, + 274, + 206 + ], + "lines": [ + { + "bbox": [ + 63, + 138, + 274, + 206 + ], + "spans": [ + { + "bbox": [ + 63, + 138, + 274, + 206 + ], + "type": "table", + "html": "
MethodPre-train1-shot5-shot
CE48.98 ± 0.1667.62 ± 0.14
CE✓71.02 ± 0.1887.25 ± 0.10
Ours33.14 ± 0.1345.02 ± 0.14
Ours✓74.28 ± 0.1888.82 ± 0.09
", + "image_path": "48ae272fe4ddecde22882161463b91966c7cbafcf8e2c64d54106be1d9590171.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 282, + 272, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 215, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 281 + ], + "type": "text", + "content": "Table 7. Ablation of different losses in the supervised training stage. All losses are trained for 60 epochs initialized with the pretrained iBOT model. The first row represents the pre-trained iBOT without additional supervised training. " + }, + { + "bbox": [ + 46, + 215, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE} + \\mathcal{L}_{[\\text{patch}]}" + }, + { + "bbox": [ + 46, + 215, + 287, + 281 + ], + "type": "text", + "content": " stands for the combination of CE loss and our patch-level loss. Results are evaluated by Prototype method on mini-ImageNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 282, + 272, + 373 + ], + "lines": [ + { + "bbox": [ + 66, + 282, + 272, + 373 + ], + "spans": [ + { + "bbox": [ + 66, + 282, + 272, + 373 + ], + "type": "table", + "html": "
Loss1-shot5-shot
-60.93 ± 0.1780.38 ± 0.12
LCE71.02 ± 0.1887.25 ± 0.10
L[cls]70.21 ± 0.1787.03 ± 0.10
L[patch]70.84 ± 0.1885.90 ± 0.11
LCE + L[patch]70.70 ± 0.1886.77 ± 0.10
L[cls] + L[patch]74.28 ± 0.1888.82 ± 0.09
", + "image_path": "52d1fbf0dd58ac661d4ecb1670dd2f22fb152e0bfd2cf8d4845af6eff859c17d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": "Class and Patch-Level Distillation Complement Each Other. Our SMKD uses the combination of class-level and patch-level supervised-contrastive losses as the final objective. Here we study several different combinations of supervised losses to confirm our proposed method performs the best. Results in Table 7 show that combining " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{cls}]}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "13.35\\%" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " improvement over baseline for 1-shot and " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "8.44\\%" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " for 5-shot classification, which outperforms each of its components. Sharing the parameters of projection heads for class and patch tokens allows the semantics obtained from their distillation to complement each other, which leads to much better performance. This can also be seen from the comparison between " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE} + \\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{cls}]} + \\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " alone achieves stronger performance than " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{cls}]}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": ", but combining it with " + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{[\\mathrm{patch}]}" + }, + { + "bbox": [ + 46, + 388, + 287, + 593 + ], + "type": "text", + "content": " does not provide additional benefits with separate classification heads." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "text", + "content": "Weighted Average Pooling Boosts Performance. Some recent works [60,69] find that using average pooling instead of the [cls] token during evaluation can encourage token-level tasks (e.g. localization, segmentation). Inspired by such findings, we evaluate our model with different tokens as well as their combinations. Results from Table 8 show that the performance of the [cls] token can be boosted by " + }, + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "inline_equation", + "content": "3.02\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "text", + "content": " for 1-shot and " + }, + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "text", + "content": " for 5-shot by just concatenating with the weighted average pooling token. This shows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "that our proposed method can learn meaningful representations for both [cls] and [patch] tokens. Moreover, weighted average pooling and average pooling tokens share similar information, but the former performs better because it gives less weight to the background." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 217, + 538, + 320 + ], + "blocks": [ + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "type": "text", + "content": "Table 8. Ablation of different tokens for few-shot evaluation. \"cls\" stands for the [cls] token, \"avg pool\" stands for using average pooling as token, \"weighted avg pool\" stands for weighted average [patch] token with weights from the self-attention module from the last block of ViT. The last four rows represent concatenation of these three tokens. All tokens are normalized to unit length. Results are evaluated by Prototype classification method." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 217, + 538, + 320 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 538, + 320 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 538, + 320 + ], + "type": "table", + "html": "
Embedding1-shot5-shot
①cls71.26 ± 0.1788.03 ± 0.10
②avg pool71.65 ± 0.1986.40 ± 0.11
③weighted avg pool71.83 ± 0.1986.38 ± 0.11
① + ②73.83 ± 0.1888.58 ± 0.10
① + ③74.28 ± 0.1888.82 ± 0.09
② + ③71.86 ± 0.1986.47 ± 0.11
① + ② + ③73.80 ± 0.1888.26 ± 0.10
", + "image_path": "b345ba9ea13f92a23076607638ccfa6648b12792e4b82d5e1971d3d47ec99c08.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 329, + 378, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 378, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 378, + 342 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 350, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 564 + ], + "type": "text", + "content": "In this work, we propose a novel supervised knowledge distillation framework (SMKD) for few-shot Transformers, which extends the self-supervised masked knowledge distillation framework into the supervised setting. With our design of supervised-contrastive losses, we incorporate supervision into both class and patch-level knowledge distillation while still enjoy the benefits of not needing large batch size and negative samples. Evaluation results together with ablation studies demonstrate the superiority of our method given its simple design compared with contemporary works. Our two-stage training is a special case of curriculum learning from the easy samples to the hard ones. We can unify the learning objectives of self-supervised learning and supervised contrastive learning, using a carefully designed curriculum learning strategy for future works. We hope our work can bridge the gap between self-supervised and supervised knowledge distillation, and inspire more works on supervised few-shot learning methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 576, + 420, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 576, + 420, + 589 + ], + "spans": [ + { + "bbox": [ + 306, + 576, + 420, + 589 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 595, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 704 + ], + "type": "text", + "content": "This research is sponsored by Air Force Research Laboratory (AFRL) under agreement number FA8750-19-1-1000. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation therein. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of Air Force Laboratory, DARPA or the U.S. Government." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19656" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3,4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 157 + ], + "type": "text", + "content": "[2] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. arXiv preprint arXiv:1805.08136, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 201 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 203, + 286, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 203, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 286, + 257 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 286, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 301 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 286, + 335 + ], + "type": "text", + "content": "[6] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 286, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 286, + 378 + ], + "type": "text", + "content": "[7] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 380, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 286, + 423 + ], + "type": "text", + "content": "[8] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 286, + 479 + ], + "type": "text", + "content": "[9] Zhengyu Chen, Jixie Ge, Heshen Zhan, Siteng Huang, and Donglin Wang. Pareto self-supervised training for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13663-13672, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 286, + 534 + ], + "type": "text", + "content": "[10] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1601-1610, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 286, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 623 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 667 + ], + "type": "text", + "content": "[13] Carl Doersch, Ankush Gupta, and Andrew Zisserman. Crosstransformers: spatially-aware few-shot transfer. Advances in Neural Information Processing Systems, 33:21981-21993, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng Zuo. Self-promoted supervision for few-shot transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[15] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 193 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 236 + ], + "type": "text", + "content": "[17] Yutong Feng, Jianwen Jiang, Mingqian Tang, Rong Jin, and Yue Gao. Rethinking supervised pre-training for better downstream transferring. arXiv preprint arXiv:2110.06014, 2021.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "type": "text", + "content": "[18] Hanan Gani, Muzammal Naseer, and Mohammad Yaqub. How to train vision transformer on small-scale datasets? arXiv preprint arXiv:2210.07240, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 325 + ], + "type": "text", + "content": "[19] Spyros Gidaris, Andrei Bursuc, Nikos Komodakis, Patrick Pérez, and Matthieu Cord. Boosting few-shot visual learning with self-supervision. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8059-8068, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "type": "text", + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[21] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "type": "text", + "content": "[22] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3263-3272, October 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "text", + "content": "[23] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 780-789, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[24] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5321-5330, June 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[25] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, Rama Chellappa, and Shih-Fu Chang. Multimodal few-shot object detection with meta-learning based cross-modal prompting. arXiv preprint arXiv:2204.07841, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[26] Guangxing Han, Xuan Zhang, and Chongrong Li. Revisiting faster r-cnn: A deeper look at region proposal network. In International Conference on Neural Information Processing, pages 14-24, 2017. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19657" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Guangxing Han, Xuan Zhang, and Chongrong Li. Single shot object detection with top-down refinement. In 2017 IEEE International Conference on Image Processing (ICIP), pages 3360-3364. IEEE, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[28] Guangxing Han, Xuan Zhang, and Chongrong Li. Semi-supervised diff: Decoupling detection and feature flow for video object detectors. In Proceedings of the 26th ACM international conference on Multimedia, pages 1811-1819, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 226 + ], + "type": "text", + "content": "[29] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 281 + ], + "type": "text", + "content": "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "text", + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 369 + ], + "type": "text", + "content": "[32] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. arXiv preprint arXiv:2206.07267, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 424 + ], + "type": "text", + "content": "[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 448 + ], + "type": "text", + "content": "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[35] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "type": "text", + "content": "[36] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 557 + ], + "type": "text", + "content": "[37] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. Advances in Neural Information Processing Systems, 34:23818-23830, 2021. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[39] Jiang Lu, Pinghua Gong, Jieping Ye, and Changshui Zhang. Learning from very few samples: A survey. arXiv preprint arXiv:2009.02653, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[40] Yuning Lu, Liangjian Wen, Jianzhuang Liu, Yajing Liu, and Xinmei Tian. Self-supervision can be a good few-shot learner. arXiv preprint arXiv:2207.09176, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[41] Xu Luo, Longhui Wei, Liangjian Wen, Jinrong Yang, Lingxi Xie, Zenglin Xu, and Qi Tian. Rectifying the shortcut learning of background: Shared object concentration for few-shot" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "image recognition. arXiv preprint arXiv:2107.07746, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 97, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 152 + ], + "type": "text", + "content": "[42] Jiawei Ma, Hanchen Xie, Guangxing Han, Shih-Fu Chang, Aram Galstyan, and Wael Abd-Almageed. Partner-assisted learning for few-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10573-10582, 2021. 1, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 154, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 219 + ], + "type": "text", + "content": "[43] Puneet Mangla, Nupur Kumari, Abhishek Sinha, Mayank Singh, Balaji Krishnamurthy, and Vineeth N Balasubramanian. Charting the right manifold: Manifold mixup for few-shot learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2218-2227, 2020. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "text", + "content": "[44] Boris Oreshkin, Pau Rodríguez López, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few-shot learning. Advances in neural information processing systems, 31, 2018. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "type": "text", + "content": "[45] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. A unified view of masked image modeling. arXiv preprint arXiv:2210.10615, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 303, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 346 + ], + "type": "text", + "content": "[46] Elia Peruzzo, Enver Sangineto, Yahui Liu, Marco De Nadai, Wei Bi, Bruno Lepri, and Nicu Sebe. Spatial entropy regularization for vision transformers. arXiv preprint arXiv:2206.04636, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[47] Guodong Qi, Huimin Yu, Zhaohui Lu, and Shuzhao Li. Transductive few-shot classification on the oblique manifold. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8412-8422, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 396, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 460 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "text", + "content": "[49] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 498, + 546, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 546, + 542 + ], + "type": "text", + "content": "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 597 + ], + "type": "text", + "content": "[51] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenenbaum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 666 + ], + "type": "text", + "content": "[52] Mamshad Nayeem Rizve, Salman Khan, Fahad Shahbaz Khan, and Mubarak Shah. Exploring complementary strengths of invariant and equivariant representations for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10836-10846, June 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[53] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19658" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[55] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[56] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199-1208, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 251 + ], + "type": "text", + "content": "[57] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "text", + "content": "[58] Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? In European Conference on Computer Vision, pages 266-282. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 365 + ], + "type": "text", + "content": "[59] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 398 + ], + "type": "text", + "content": "[60] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "text", + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 487 + ], + "type": "text", + "content": "[62] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 1, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 555 + ], + "type": "text", + "content": "[63] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[64] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "type": "text", + "content": "[65] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[66] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22-31, 2021. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 697 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[67] Jiamin Wu, Tianzhu Zhang, Yongdong Zhang, and Feng Wu. Task-aware part mining network for few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8433-8442, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[68] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[69] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[70] Dongyang Zhao Hong-Yu Zhou Weifeng Ge Yizhou Yu Wenqiang Zhang Yangji He, Weihan Liang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[71] Han-Jia Ye, Hexiang Hu, De-Chuan Zhan, and Fei Sha. Few-shot learning via embedding adaptation with set-to-set functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8808-8817, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "type": "text", + "content": "[72] Nikolaos-Antonios Ypsilantis, Noa Garcia, Guangxing Han, Sarah Ibrahimi, Nanne Van Noord, and Giorgos Tolias. The met dataset: Instance-level recognition for artworks. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 374, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 545, + 419 + ], + "type": "text", + "content": "[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 420, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 474 + ], + "type": "text", + "content": "[74] Chi Zhang, Yujun Cai, Guosheng Lin, and Chunhua Shen. Deepemd: Few-shot image classification with differentiable earth mover's distance and structured classifiers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "text", + "content": "[75] Xueting Zhang, Debin Meng, Henry Gouk, and Timothy M Hospedales. Shallow bayesian meta learning for real-world few-shot recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 651-660, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 531, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 564 + ], + "type": "text", + "content": "[76] Nanxuan Zhao, Zhirong Wu, Rynson WH Lau, and Stephen Lin. What makes instance discrimination good for transfer learning? arXiv preprint arXiv:2006.06606, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 609 + ], + "type": "text", + "content": "[77] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 654 + ], + "type": "text", + "content": "[78] Ziqi Zhou, Xi Qiu, Jiangtao Xie, Jianan Wu, and Chi Zhang. Binocular mutual learning for improving few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8402-8411, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 655, + 545, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 655, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 307, + 655, + 545, + 697 + ], + "type": "text", + "content": "[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19659" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_content_list.json b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..639e7507ee5d6581d3b943839f67fb65bda116b7 --- /dev/null +++ b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "SurfelNeRF: Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes", + "text_level": 1, + "bbox": [ + 138, + 128, + 831, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiming Gao Yan-Pei Cao Ying Shan", + "bbox": [ + 300, + 203, + 669, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ARC Lab, Tencent PCG", + "bbox": [ + 387, + 222, + 581, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gaoym9@mail3.sysu.edu.cn caoyanpei@gmail.com yingsshan@tencent.com", + "bbox": [ + 168, + 242, + 795, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online reconstructing and rendering of large-scale indoor scenes is a long-standing challenge. SLAM-based methods can reconstruct 3D scene geometry progressively in real time but can not render photorealistic results. While NeRF-based methods produce promising novel view synthesis results, their long offline optimization time and lack of geometric constraints pose challenges to efficiently handling online input. Inspired by the complementary advantages of classical 3D reconstruction and NeRF, we thus investigate marrying explicit geometric representation with NeRF rendering to achieve efficient online reconstruction and high-quality rendering. We introduce SurfelNeRF, a variant of neural radiance field which employs a flexible and scalable neural surfel representation to store geometric attributes and extracted appearance features from input images. We further extend the conventional surfel-based fusion scheme to progressively integrate incoming input frames into the reconstructed global neural scene representation. In addition, we propose a highly-efficient differentiable rasterization scheme for rendering neural surfel radiance fields, which helps SurfelNeRF achieve $10 \\times$ speedups in training and inference time, respectively. Experimental results show that our method achieves the state-of-the-art 23.82 PSNR and 29.58 PSNR on ScanNet in feedforward inference and per-scene optimization settings, respectively. $^{1}$", + "bbox": [ + 75, + 323, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale scene reconstruction and rendering is a crucial but challenging task in computer vision and graphics with many applications. Classical visual simultaneous localization and mapping (SLAM) systems [6, 12, 16, 24, 39, 41] can perform real-time 3D scene reconstruction. However, they usually represent the scene geometry as solid surfaces and appearance as vertex color or texture maps; thus,", + "bbox": [ + 75, + 757, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c1222365b0a2fe61ffe68b3ebee31b0cf87c99fb6fd47f9946ae667013c53d4.jpg", + "image_caption": [ + "Figure 1. Examples to illustrate the task of online photorealistic reconstruction of an indoor scene. The online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views." + ], + "image_footnote": [], + "bbox": [ + 506, + 292, + 893, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the reconstructed results fail to fully capture the scene content and cannot be used for photorealistic rendering. Recently, neural radiance fields (NeRF) and its variants [2, 3, 22,27,35,42] have achieved unprecedented novel view synthesis quality on both object-centric and large-scale scenes. However, NeRFs suffer from long per-scene optimization time and slow rendering speed, especially for large-scale scenes. Although recent advances [7, 14, 19, 23, 33] achieve faster optimization and rendering via incorporating explicit representations, they still require gathering all input images in an offline fashion before optimizing each scene.", + "bbox": [ + 496, + 597, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we target the challenging task of online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views. The online setting can unlock a variety of real-time interactive applications, providing crucial immediate feedback to users during 3D capture. However, this task brings multiple extra requirements, including the scalability", + "bbox": [ + 496, + 765, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project website: https://gymat.github.io/SurfelNeRF-web", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "108", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/668b0c804464494a7c6f0c16257db4ccc6ccc9f509a648dcad3a535c035cad51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRepresentationGeneralizationReal-time RenderingOnline FusionScalability
DeepSurfel [21]SurfelsX✓XX
Instant-NGP [23]Hash GridsX✓XX
PointNeRF [43]Point Clouds✓XX✓
VBA [10]B+ TreesXX✓✓
NeRFusion [50]Voxel Grids✓X✓X
OursSurfels✓✓✓✓
", + "bbox": [ + 161, + 88, + 810, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of representation and features with existing methods.", + "bbox": [ + 261, + 205, + 705, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity of the underlying scene representation, the ability to perform on-the-fly updates to the scene representation, and optimizing and rendering at interactive framereates. Recently, NeRFusion [50] followed NeuralRecon [34] to unproject input images into local sparse feature volumes, fusing them to a global volume via Gated Recurrent Units (GRUs), and then generating photorealistic results from the global feature volume via volume rendering. However, updating the sparse volumetric feature involves computationally heavy operations; the volume rendering is also very slow since it requires hundreds of MLP evaluations to render a pixel. Thus, although NeRFusion achieves efficient online scene reconstruction, it still needs dozens of seconds to render a frame. VBA [10] is another recent approach to online photorealistic scene capture, but it only applies to object-centric scenes. We compare with representation and key features used in online photorealistic rendering with existing methods, which is shown in Tab. 1.", + "bbox": [ + 75, + 233, + 472, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose surfel-based neural radiance fields, SurfelNeRF, for online photorealistic reconstruction and rendering of large-scale indoor scenes. Surfels (surface elements) [25] are point primitives containing geometric attributes, e.g., position, color, normal, and radius. We extend this representation to neural surfels, storing extra neural features that encode the neural radiance field of the target scene. Compared with volumetric representations, neural surfels are more compact and flexible and can easily scale to large scenes. Besides, we further employ a fast and differentiable rasterization process to render neural surfel radiance fields, which produces a pixel with only a few MLP evaluations based on the rasterized neural surfels.", + "bbox": [ + 75, + 506, + 468, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by classical real-time surfel-based geometric reconstruction methods [6, 15, 41], we propose an efficient neural radiance field fusion method to progressively build the scene representation by integrating neighboring neural surfels. Unlike point-based representations [1, 26, 30, 43] that are computationally heavy when finding neighboring points, it is easier to locate overlapping surfels and then merge neural features from multiview observations. By coupling the SurfelNeRF representation, the efficient neural surfel fusion approach, and the fast neural surfel rasterization algorithm, we achieve high-quality, photorealistic 3D scene reconstruction in an online manner.", + "bbox": [ + 75, + 703, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on the large-scale indoor scene", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset ScanNet [11], which contains complex scene structures and a large variety of scene appearances. We train the SurfelNeRF end-to-end across the scenes on the ScanNet, obtaining a generalizable model that enables both feedforward inference on unseen data and per-scene fine-tuning. We demonstrate in experiments that the proposed SurfelNeRF achieves favorably better rendering quality than the state-of-the-art approaches in both feedforward and finetuning settings while maintaining high training and rendering efficiency. We believe the proposed online photorealistic reconstruction framework has great potential in practical applications.", + "bbox": [ + 496, + 233, + 893, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 433, + 640, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis. Novel view synthesis is a challenging task for computer vision and computer graphics. Previous methods utilize geometry-based view synthesis. Zhou et al. [51] proposes to use the multiplane image (MPI) representation for novel view synthesis. Recently, neural radiance fields [22] and neural-based geometry representation have attracted more interest due to high-quality rendering. Recent methods propose to associate neural features on geometry structure and then synthesize novel views via rendering rasterized features, such as point cloud [1, 30], texture [36] and surfel [21]. DeepSurfel performs online appearance fusion only on a fixed given geometry during online fusion, while our SurfelNeRF performs online appearance and geometry integration. To render outputs, DeepSurfel rasterizes surfels and then passes them through 2D CNN filters, while our SurfelNeRF performs neural volume rendering which could achieve higher-quality rendering. Some recent works [9, 13, 18, 45] focus on real-time neural rendering methods. However, these methods all require per-scene optimization, which makes them impractical to handle novel scenes. In contrast, our SurfelNeRF is generalizable, which can reconstruct new scenes via direct feedforward inference.", + "bbox": [ + 496, + 460, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields. Recently, neural radiance fields (NeRF) methods [2, 22, 42] achieve promising results on novel view synthesis, but it requires expensive per-scene optimization. Some methods proposed to reduce the long per-scene optimization time via deferred neural rendering [13, 47], hash table [23] or geometry prior [20]. Recent,", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "109", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generalization NeRF methods [5,8,37,43,46] are proposed to produce novel view without per-scene optimization by learning a radiance fields representation from offline given images of novel scenes. PixelNeRF [46] and IBRNet [37] render novel view images using volume rendering based on corresponding warped features from nearby given images. MVSNeRF [8] reconstructs local neural volumes from local nearby views, and then conducts volume rendering based on volumes. FWD [5] renders novel views by warping all input images to the target view, and fusing and refining features of these warped inputs. However, these methods are generalizable on local small-scale scenes, and require taking all input images before rendering novel views in an offline fashion. PointNeRF [43] learns to reconstruct a point-based neural radiance field for the novel scene. However, point-based representation is redundant, especially on large-scale scenes, and the point clouds cannot be updated based on multi-view input. Unlike these works, our method can represent not only large-scale scenes but also update compact surfel representation in an online fashion.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online scene reconstruction and rendering. Most online scene reconstruction methods [4, 34, 41, 52, 53] focus on 3D reconstruction only via TSDF-based fusion [4, 34], surfel-based fusion [41] and SLAM-based reconstruction [52]. However, they cannot render photorealistic novel view synthesis. Recently, some methods [10, 32, 50] employ radiance fields to tackle online scene representation for high-quality rendering. VBA [10] builds a neural dynamic B+Tree to represent scenes and apply volume rendering on it, but it still requires per-scene optimization. SRT [32] represents scenes with set-latent with a transformer encoder and decoder, but it cannot synthesize high-resolution images, which makes it not ready for real-world applications. NeRFusion [50] represents large-scale scenes with neural voxel grids, update them via GRU in an online fashion, and renders results via slow volume rendering. However, grid-based representation is wasteful for large-scale scenes, and is inefficient during features update due to features in empty areas are also updated. In contrast, our surfel representation is compact for large-scale scenes, and surfel-based fusion and corresponding rendering are more efficient.", + "bbox": [ + 76, + 393, + 472, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 724, + 168, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike most existing methods [22, 23, 43] that build neural radiance fields in an offline fashion, we aim for the task of online photorealistic scene reconstruction. Given a sequence of images $\\{I_n\\}$ , the goal is to progressively build the neural scene representation on the fly as the input images arrive.", + "bbox": [ + 76, + 750, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this goal, we present SurfelNeRF, an online neural reconstruction framework based on the neural surfel radiance fields representation. The insight of our method is that the classical explicit surfel representation is efficient", + "bbox": [ + 76, + 839, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for both reconstruction and rendering, has compact memory footprints, has the scalability to large and complex scenes, and can be naturally extended to equip neural attributes. Meanwhile, one can integrate multi-frame information and perform incremental updates on neural surfels in a geometrically meaningful way, and the operations involved can be streamlined.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 gives an overview of our proposed method. The framework first learns to generate per-view local neural surfel radiance fields $S_{t}$ for each input frame $I_{t}$ . We then adopt the conventional surfel fusion process [6,15,41] to the neural setting, parameterizing the neural surfel fusion procedural as a neural network $N_{f}$ . Through $N_{f}$ , per-view representations $\\{S_t\\}$ can be integrated into a global neural surfel radiance field $S_{g}$ on the fly, aggregating geometric and neural attributes. The neural surfel radiance fields can be efficiently rendered to photorealistic images using a differentiable rasterizer without extra optimization. The framework can be trained end-to-end on large-scale indoor scene datasets. Once trained, it can generalize to previously unseen data and perform online photorealistic reconstruction; if needed, the rendering quality can be further improved by a short period time of fine-tuning.", + "bbox": [ + 496, + 196, + 892, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Neural Surfel Radiance Field Construction", + "text_level": 1, + "bbox": [ + 498, + 450, + 864, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Surfel Representation. Adopted from classical 3D reconstruction research, each surfel $s^n$ is equipped with the following geometric attributes: a position $\\mathbf{p}^n \\in \\mathbb{R}^3$ , normal $\\mathbf{n}^n \\in \\mathbb{R}^3$ , weight $w^n \\in \\mathbb{R}$ and radius $r^n \\in \\mathbb{R}$ . The weight $w^n$ measures the confidence of a surfel based on the distance of the current depth measurement from the camera center. The radius $r^n$ of each surfel is designed to represent the local support area around the projected 3D position while minimizing visible holes. The weight and radius of each surfel are initialized following [41]. Besides geometric properties, we further associate each surfel with neural features $\\mathbf{f}^n \\in \\mathbb{R}^N$ from input images for modeling semi-transparent structures and view-dependent effects. Although a similar representation has been explored in DeepSurfels [21], we construct neural surfels in a feedforward manner and incorporate them in progressive scene reconstruction.", + "bbox": [ + 496, + 476, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local Neural Surfel Radiance Field. Given an input frame $I_{t}$ , camera parameters $\\mathbf{c}_t$ , and estimated depth $D_{t}$ , the corresponding local neural surfel radiance field $\\mathcal{S}_t$ can be constructed by unprojecting pixels in the input image to 3D space using camera parameters, then estimating corresponding normal, weight, radius, and neural features. The depth $D_{t}$ can be captured from sensors or estimated via off-the-shelf MVS-based methods [44], which might be incomplete and noisy. Thus, we employ a U-Net structure [29] to refine the depth following [5].", + "bbox": [ + 496, + 733, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Features on Surfels. We estimate neural features", + "bbox": [ + 500, + 885, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "110", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b7b4e20fde4ff81e9924c8dcbac88033f882241b5f9eef8d8e50279d428fa70.jpg", + "image_caption": [ + "Figure 2. Overview of our SurfelNeRF. Given an online input stream of image sequences, we first reconstruct a surfel representation associated with neural features to build a local neural surfel radiance field for each input keyframe. Then the neural surfel radiance field integration is used to fuse the input local neural surfel radiance field into the global neural surfel radiance field by updating both surfel position and features. More specifically, input local neural surfels associated with global surfels are fused to global surfels with corresponding global ones, and remaining local surfels without corresponding ones are added to the global model. Furthermore, the novel views can be rendered from updated global surfels via our efficient rasterization-guided render. Our proposed rasterization-guided renderer renders color only on the intersection points of ray and surfels, which is faster than volume rendering." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 851, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{f}^n$ of surfels by extracting corresponding pixel features in the input image. We employ a 2D CNN to extract features of input images. Specifically, we use a MasNet following [34, 50] to extract multi-scale view-dependent features.", + "bbox": [ + 75, + 439, + 468, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Online Integration of Neural Surfels", + "text_level": 1, + "bbox": [ + 76, + 512, + 392, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the neural surfel representation, we adopt a geometry-guided fusion scheme to efficiently and progressively integrate local neural surfel radiance fields into a global model. Compared with previous volumetric fusion methods [10,50], our surfel-based solution offers the advantages of 1) efficiency - the fusion operation only needs to be performed at surfel locations and 2) scalability - the memory footprint only grows linearly to the number of neural surfels.", + "bbox": [ + 75, + 536, + 468, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, our fusion scheme can be divided into two steps. Firstly, for each local neural surfel, we determine whether it should be added as a global neural surfel or should be merged with corresponding global surfels. Secondly, we insert it into the global model or perform updates to corresponding global surfels with local features.", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Surfel Association. Given an input image frame $I_{t}$ and corresponding camera parameter, each surfel in global surfels can be projected as an ellipse onto the image plane of the current input camera view. For each pixel in the input image, we obtain top- $K$ global surfels that are overlapped with it after rasterization. We discard irrelevant surfels with a normal filter. Then we determine the incremental surfel by comparing the depth distance between respective top-M global surfels from near to far step by step. If the min", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "imal distance between a local surfel and its global surfel candidates is smaller than a threshold $\\delta_{depth}$ , then we determine that this input surfel needs to be merged; otherwise, it should be directly added to global surfels. The overall workflow follows [41], and $K$ is set to 8.", + "bbox": [ + 496, + 439, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Surfel Fusion. After the surfel association, there are a set of input surfels that need to be added directly and another set of surfels that will be merged with corresponding global surfels. For the former set of surfels, we directly construct neural surfels (Sec. 3.1) and then add them to global surfels. For the latter surfels, we use a GRU network to update neural features between the features $\\mathbf{f}_t^{\\mathrm{merge}}$ of input surfels and features $\\mathbf{f}_{t - 1}^{\\mathrm{corr}}$ of corresponding global surfels, which is given as", + "bbox": [ + 496, + 515, + 892, + 651 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t} ^ {\\text {c o r r s}} = \\operatorname {G R U} \\left(\\mathbf {f} _ {t} ^ {\\text {m e r g e}}, \\mathbf {f} _ {t - 1} ^ {\\text {c o r r s}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 656, + 890, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_t^{\\mathrm{merge}}$ and $\\mathbf{f}_{t - 1}^{\\mathrm{corrs}}$ are input and hidden state of GRU respectively. We then update the geometric attributes of these global surfels via a weighted sum. The update rule of each respective pair of local surfel $s_t^j$ and global surfel $s_{t - 1}^{j}$ can be expressed by", + "bbox": [ + 496, + 678, + 890, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {t} ^ {j} = \\frac {w _ {t} ^ {i} \\mathbf {p} _ {t} ^ {i} + w _ {t - 1} ^ {j} \\mathbf {p} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}},\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 761, + 782, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} _ {t} ^ {j} = \\frac {w _ {t} ^ {i} \\mathbf {n} _ {t} ^ {i} + w _ {t - 1} ^ {j} \\mathbf {n} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 803, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} ^ {j} = \\frac {w _ {t} ^ {i} r _ {t} ^ {i} + w _ {t - 1} ^ {j} r _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}},\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 844, + 776, + 882 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} ^ {j} = w _ {t} ^ {i} + w _ {t - 1} ^ {j},\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 885, + 728, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "111", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e9622e389164ead116c083dea46c41fdcf3a20c957f3b2d1a8ae900f2abb4dd.jpg", + "image_caption": [ + "Figure 3. Our efficient rasterization-guided surfel rendering for neural surfel radiance fields is visualized using a toy model as an example. We begin by rasterizing neural surfels into the image plane, and then render the ray using shading points on the intersections between the ray and corresponding surfels. This method uses a combination of rasterization and ray tracing to render view-dependency results with surfel representation and 3D rendering." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_t^j$ , $\\mathbf{n}_t^j$ , $r_t^j$ and $w_t^j$ are the updated position, normal, radius, and weight of the global surfels, respectively.", + "bbox": [ + 76, + 340, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Rendering Neural Surfel Radiance Fields", + "text_level": 1, + "bbox": [ + 76, + 381, + 429, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a fast rasterization-based rendering algorithm for neural surfel radiance fields (see Figure. 3). Given a neural surfel radiance field $S$ and the camera parameter $\\mathbf{c}_{tar}$ of a target view, we first conduct a differentiable rasterization process on the neural surfels to project them on the target image plane. For a pixel $p_x$ in the target image, we gather the neural surfels $\\{s_i\\}_{i=1}^M$ that cover it and render its color $C(p_x)$ based on a few shading points $\\{\\mathbf{x}_i\\}_{i=1}^M$ that are the intersections between $\\{s_i\\}_{i=1}^M$ and pixel ray, using the following volume rendering formula [22]:", + "bbox": [ + 75, + 405, + 472, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} C \\left(p _ {x}\\right) = \\sum_ {i - 1} ^ {M} \\tau_ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\Delta_ {i}\\right)\\right) c _ {i}, \\tag {3} \\\\ \\tau_ {i} = \\exp (- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\Delta_ {j}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 565, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ indicates the accumulated transmittance; $\\sigma_{i}$ and $c_{i}$ indicate the volume density and radiance on shading point $\\mathbf{x}_i$ ; $\\Delta_j$ is the distance between adjacent shading points. For each shading point $x_{i}$ , we regress its volume density $\\sigma_{i}$ and radiance $c_{i}$ based on its view direction $\\mathbf{d}$ and shading point features $\\mathbf{f}^i (\\mathbf{x}_i)$ associated from surfel $s_i$ , which can be formulated as:", + "bbox": [ + 75, + 651, + 468, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i}, c _ {i} = \\operatorname {R e n d e r} \\left(x _ {i}, \\mathbf {f} ^ {i} \\left(\\mathbf {x} _ {i}\\right), \\mathbf {d}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 766, + 468, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use an MLP to parameterize the Render module. The \"interpolated\" surfel feature $\\mathbf{f}^i (\\mathbf{x}_i)$ is given as:", + "bbox": [ + 76, + 792, + 467, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {i} \\left(\\mathbf {x} _ {i}\\right) = \\frac {\\| \\mathbf {x} ^ {i} - \\mathbf {p} ^ {i} \\|}{r ^ {i}} \\mathrm {F} \\left(\\mathbf {f} ^ {i}, \\mathbf {d}, \\mathbf {n} ^ {i}, w ^ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 832, + 468, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ is also parameterized as an MLP-like network. Note that a recent concurrent work [48] also explores the", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hybrid rendering process by combining rasterization and volume rendering, but using a different neural representation.", + "bbox": [ + 498, + 90, + 892, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Optimization", + "text_level": 1, + "bbox": [ + 500, + 145, + 743, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All modules in SurfelNeRF are differentiable, so they can be trained end-to-end on large-scale scene datasets with sequential images as input. We utilize L2 pixel-wise rendering loss to train our network supervised by the target ground truth image:", + "bbox": [ + 498, + 167, + 893, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n d e r}} = \\left\\| C _ {g t} - C _ {r} \\right\\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 253, + 892, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_{gt}$ is the ground truth pixel color and $C_r$ denotes the rendered pixel color output from SurfelNeRF. In addition, if we can acquire sensor depth $D_i^{\\mathrm{sen}}$ senor for image $I_i$ , we further employ an L1 loss to train the depth refinement network:", + "bbox": [ + 498, + 281, + 892, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\left| \\left| D _ {i} ^ {\\text {s e n}} \\odot M _ {i} - D ^ {i} \\odot M _ {i} \\right| \\right| _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 367, + 890, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D^{i}$ is the output from depth refinement network, and $M_{i}$ is a binary mask indicating valid sensor depths. This supervision can make the depth refinement network to learn to match the sensor depths.", + "bbox": [ + 498, + 393, + 892, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To sum up, the total loss to train SurfelNeRF end-to-end is given as", + "bbox": [ + 498, + 455, + 893, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda \\cdot \\mathbb {1} \\left(\\mathcal {L} _ {d}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 484, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{1}$ denotes the indicator function whose value is set to 1 if sensor depth is available; $\\lambda$ is the balancing weight.", + "bbox": [ + 500, + 507, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Following [50], we uniformly sample keyframes from each image sequence as input for training SurfelNeRF. To train SurfelNeRF, we sample about $5\\%$ of the full sequence as keyframes, and other frames are used as supervision.", + "bbox": [ + 498, + 537, + 893, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Per-scene Fine-tuning. After being trained on large-scale datasets, our SurfelNeRF is capable of constructing neural surfel radiance fields and rendering photorealistic novel views on unseen scenes in a feedforward manner. To further improve the rendering performance on test scenes, one can also perform a short period of per-scene optimization via fine-tuning the surfel representations and the rendering module. To make a fair comparison, we do not use the additional depths in the fine-tuning process and keep the depth refinement module fixed. Thanks to our efficient rendering scheme, SurfelNeRF takes less time when fine-tuning compared with existing methods [43, 50].", + "bbox": [ + 496, + 613, + 893, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 633, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 830, + 707, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. Our training data consists of large-scale indoor scenes from ScanNet [11] which provides the online sequence input of images with known camera parameters,", + "bbox": [ + 498, + 854, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "112", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the captured incomplete sensor depths. We focus on the online input of large-scale scenes in the main paper. Following [37, 50], we randomly sample 100 scenes from ScanNet for training.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We train our SurfelNeRF on the large-scale training dataset mentioned above using the Adam [17] optimizer with an initial learning rate of 0.001 for total 200k iterations on 4 NVIDIA Tesla V100 32GB GPUs. For per-scene optimization settings, we fine-tune our method also using the Adam optimizer with an initial learning rate of 0.0002. All input images are resized to $640 \\times 480$ . The $\\lambda$ in Eq. 8 is set to 0.1. The threshold of $\\delta_{depth}$ is set to $0.1\\mathrm{m}$ in surfel fusion (Sec. 3.2). The maximum number of candidates M in Eq. 3 for rendering is set to 80. We implement our method using Pytorch and Pytorch3D.", + "bbox": [ + 75, + 152, + 470, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following previous works [50], we report the PSNR, SSIM [38] and LPIPS $_{vgg}$ [49] to evaluate and compare the results. For fair comparisons with previous methods, we also evaluate our method on the same test set on the ScanNet dataset. The time cost of rendering and perscene fine-tuning is measured on the same platform using one NVIDIA Tesla V100 32GB GPU.", + "bbox": [ + 75, + 319, + 468, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with state-of-the-arts", + "text_level": 1, + "bbox": [ + 76, + 436, + 375, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our method with other state-of-the-arts on the large-scale dataset, ScanNet. To achieve fair comparisons, we evaluate our methods strictly following the same training and evaluation scheme as NeRFusion [50] and NerfingMVS [40]. There are 8 selected testing scenes in ScanNet where 1/8 of images are held out for testing. Following [50], we report no per-scene optimization results via direct inference network and per-scene optimization results.", + "bbox": [ + 75, + 460, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. In addition to comparing SurfelNeRF with NeRF-Fusion [50], we also compare it with other state-of-the-art (SOTA) methods, such as fast per-scene optimization NeRF and neural point cloud representation, generalizable NeRF, scalable MLP-based representation NeRF, NeRF with additional depth supervision, and classical image-based rendering methods, including Instant-NGP [23], ADOP [31], PointNeRF [43], Mip-NeRF-360 [3], Dense Depth Prior NeRF/DDP-NeRF) [28], and IBRNet [37]. Among these methods, Mip-NeRF-360 is one of the state-of-the-art NeRF methods for large-scale scenes. DDP-NeRF employs an offline NeRF with additional depth supervision. PointNeRF achieves state-of-the-art performance thanks to point-based representation. ADOP achieves promising results on offline large-scale rendering. These two methods employ a similar representation to our method; thus we also compare with them. We run generalizable PointNeRF with the same No per-scene optimization and per-scene optimization settings using its open-source code.", + "bbox": [ + 75, + 597, + 470, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The result is shown in Tab. 2 (a). For no per-scene", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f952aa366259a02d4825ed6e4cdfdcbbf9a619b8ab649e81de72a78097b5e02e.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Results of No per-scene optimization setting on ScanNet. And we report the average time to render an image. $\\sharp$ indicates that the time of NeR-Fusion is reported by authors using a 2080Ti since they have not released executable code before submission." + ], + "table_body": "
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
IBRNet [37]21.190.7860.358-
NeRFusion [50]22.990.8380.33538s#
PointNeRF [43]20.470.6420.54430s
SurfelNeRF23.820.8450.3270.2s
", + "bbox": [ + 517, + 89, + 874, + 180 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/474f874671318d063e92faa8a14da31ad1499f60cbeeaf468b2fb0dd0256fa5d.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Per-scene optimization. Time↓ indicates the average time to render an image.", + "Table 2. Quantitative comparisons with SOTAs on the ScanNet dataset with No per-scene and per-scene optimization." + ], + "table_body": "
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
Instant-NGP [23]23.230.7140.4590.03
ADOP [31]25.010.8070.2721s
NeRFingMVS [40]26.370.9030.245-
IBRNet [37]25.140.8710.266-
DDP-NeRF [28]26.610.800.26-
Mip-NeRF-360 [3]27.850.860.25-
NeRFusion [50]26.490.9150.20938s
PointNeRF [43]28.990.8290.32430s
SurfelNeRF29.580.9190.2150.2s
", + "bbox": [ + 506, + 244, + 883, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimization settings, NeRFusion achieves PSNR of 22.99 thanks to the fusion between input features to global grid representations. Our SurfelNeRF outperforms NeRFusion since SurfelNeRF benefits from compact and precise feature fusion and the geometric guidance of surfels. For the per-scene optimization task, the result is shown in Tab. 2 (b). PointNeRF gains a large improvement after fine-tuning, which illustrates it achieves sub-optimal performance in no per-scene optimization settings due to its inability to fuse and update point features. Compared with DDP-NeRF, which employs additional depth supervision, our SurfelNeRF still outperforms it by taking advantage of surfel representation and its integration. Our method achieves state-of-the-art results again, which shows that our neural surfel representation can provide optimal initialization for fine-tuning.", + "bbox": [ + 496, + 489, + 892, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative comparison is shown in Figure. 4. We compare our method with PointNeRF without per-scene optimization, and compare with per-scene optimization methods of PoingNeRF and ADOP. Since NeRFusion has not release the executable code, we cannot generate qualitative results. As shown in Figure. 4, our results are visually much better than state-of-the-art methods, PointNeRF and ADOP.", + "bbox": [ + 496, + 732, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 847, + 653, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We investigate the effectiveness of designs in our SurfelNeRF by conducting the following ablation studies on", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "113", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eeacfc929f7b6e21c64a68edd2dcc04db4273566b1e45b31fa8aa36614d98bce.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons of rendering quality. The 1-2 column are the results of no per-scene optimization. And the left three columns are per-scene optimization results. It is obvious that with or without per-scene optimization, our method can still achieve the most photorealistic rendering results. Since PointNeRF without a depth refinement network relies on captured depths only, it cannot handle the scenario that depth is incomplete. However, our method can tackle this problem with a lightweight depth refinement network. ADOP can produce geometrically correct results, but its color is less photorealistic." + ], + "image_footnote": [], + "bbox": [ + 127, + 90, + 836, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ScanNet.", + "text_level": 1, + "bbox": [ + 76, + 700, + 140, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fusion Schemes. Our surfel-based fusion is inspired by classical Surfel-based fusion in conventional 3D reconstruction [41]. The classical surfel-based fusion update features via the simple weighted sum of two overlapped surfels. We conduct an experiment that updates surfel features by weighted sum based on surfel weights, which is a similar operation of updating geometry attributes in Eq. 2. The results are shown in Tab. 3. Since the GRU module can learn to adaptively update features based on the global surfels and new incremental surfels, it can improve performance with negligible computation overhead.", + "bbox": [ + 75, + 715, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth Refinement Module. To investigate the effective-", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0dd2e28d516e4fce62ce15ddc3dfd26c6395441cbefb083066c101f1da85d173.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Fusion SchemeSettingPSNR↑SSIM↑LPIPS ↓
Weighted SumNo per-scene23.090.8330.353
GRUoptimization23.820.8450.327
Weighted SumPer-scene28.540.8840.293
GRUoptimization29.580.9190.215
", + "bbox": [ + 500, + 696, + 906, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies about fusion schemes in our SurfelNeRF.", + "bbox": [ + 500, + 800, + 890, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ness of our depth refinement, we conduct experiments that reconstruct scenes with incomplete depth captured by sensors. The quantitative results are shown in Tab. 4, which show that PSNR performance drops about $4\\%$ without the", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "114", + "bbox": [ + 486, + 944, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6a1ad81b652212160046247df99fb73ecbff989a9268b3129ce8458e1ba848c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Depth RefinementPSNR↑SSIM↑LPIPS ↓
w/o19.510.7130.454
w23.090.8330.353
", + "bbox": [ + 111, + 88, + 436, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c4c9c89e7c58a961d2e64936b2d193ffb63fb472059a17472578fa0826c7c40c.jpg", + "table_caption": [ + "Table 4. Ablation studies about the effectiveness of depth refinement in scene0000_01 in the ScanNet with the no per-scene optimization setting. The depth captured by the sensor in this scene is incomplete. If using captured depths only, some pixels cannot be projected to the 3D space, as the visualization results of PointNeRF shown in the first row in the qualitative results in Fig. 4." + ], + "table_footnote": [], + "table_body": "
Rendering SchemeSettingPSNR↑Time ↓
Volume RenderingNo per-scene23.8030s
Oursoptimization23.820.2s
Volume RenderingPer-scene29.4530s
Oursoptimization29.580.2s
", + "bbox": [ + 94, + 255, + 452, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparing with two rendering schemes, volume rendering [22] and our proposed rasterization-guided rendering, in the ScanNet dataset.", + "bbox": [ + 75, + 358, + 468, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "depth refinement. Depth quality affects surfel reconstruction geometry quality, but captured depth is always incomplete (visualized in the first row in Fig. 4). Thus, a depth refinement network is necessarily required to refine the input depth for better surfel geometry quality. The experimental result shows that the depth refinement network is effective in alleviating the noisy or incomplete captured or estimated depth. Moreover, we also conduct an experiment with worse depth quality that resizes input depth resolution to half size. Our performance just slightly drops $0.5\\%$ PSNR. These experiments show the effectiveness of our depth refinement module.", + "bbox": [ + 75, + 416, + 468, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Rendering Schemes.** Our proposed renderer renders a ray with a few shading points, which are the intersection points of ray and surfels, unlike volume rendering in NeRF [22] that requires hundreds of shading points each with one MLP evaluation. To validate the effectiveness of our differentiable rasterization scheme for rendering, we conduct an experiment that replaced our rendering scheme with NeRFs' volume rendering. The result is shown in Tab. 5. Since online integration of surfels can limit the number of intersecting surfels, our rendering speed is persistently faster than volume rendering with fewer shading points. Specifically, The average and maximal number of rasterized surfels per pixel at novel views on ScanNet are 2.92 and 10, respectively.", + "bbox": [ + 75, + 598, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Cost and Memory Usage. To evaluate the computational cost and memory usage on a large-scale scene, we plot components of GPU processing time, surfel number, and memory usage of our method on a large three rooms scene measured on a V100 GPU implemented by Pytorch3D. During inference, our method renders $640 \\times 480$", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32897e5ca832b5c88f757b61d307af1a2c813fcd90c1c5bae0801efe25beb265.jpg", + "image_caption": [ + "(a) GPU processing time.", + "Figure 5. The computation cost, memory footprint, and the number of surfels grow according to input frames measured on a three rooms scene." + ], + "image_footnote": [], + "bbox": [ + 501, + 98, + 689, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0b92ee29a1294a28051cec34bc7d4885cee3bb8b248ef2c2ee223733347edcf.jpg", + "image_caption": [ + "(b) GPU memory usage." + ], + "image_footnote": [], + "bbox": [ + 700, + 98, + 893, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images at 5.2FPS on average. As shown in Fig. 5a and Fig. 5b, the number of surfels has a sub-linear growth thanks to the proposed surfel integration. The fluctuation of Time (Rasterization) is caused by the different numbers of corresponding surfels at novel views. Although the computational cost of rasterization depends on the number of rasterized surfels, but it is still faster than volume rendering [22]. These results show that our surfel representation is compact and efficient. The major computation cost comes from rasterization, and the cost of surfel construction as well as integration could be neglected.", + "bbox": [ + 496, + 266, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 444, + 622, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method requires high-quality surfel geometry as neural surfel representation since SurfelNeRF relies on it to conduct surfel integration and rasterization-guided render. To obtain better geometry quality, we may require more accurate depth quality to construct surfel geometry, which necessitates extra depth supervision or the MVS depth estimator. Since captured or estimated depth quality is not exact, there may be geometry discontinuities or missing parts in the surfel representation. Consequently, there may be some missing parts or incorrect geometry in the rendered output (examples of missing parts are shown in the first row in Fig. 4 or video). As discussed in ablation studies of depth refinement above, inaccurate and noisy estimated depth as well as depth discontinuity could affect the construction of neural surfel representation. To alleviate this issue, we propose a depth refinement module (Tab. 4) to refine depth information.", + "bbox": [ + 496, + 469, + 890, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SurfelNeRF for online photorealistic reconstruction and rendering of large-scale indoor scenes. We employ neural surfels to store a neural radiance field, adopt an efficient neural radiance field fusion method to progressively build the scene representation, and propose a fast and differentiable rasterization-based rendering algorithm to tackle the challenging task. Experiments on the ScanNet dataset demonstrate that our method achieving favorably better results than the state-of-the-art methods.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "115", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. 2020. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 1, 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 6", + "[4] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 3", + "[5] Ang Cao, Chris Rockwell, and Justin Johnson. Fwd: Real-time novel view synthesis with forward warping and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15713-15724, 2022. 3", + "[6] Yan-Pei Cao, Leif Kobbelt, and Shi-Min Hu. Real-time high-accuracy three-dimensional reconstruction with consumer RGB-D cameras. ACM Transactions on Graphics, 37(5):171:1-171:16, 2018. 1, 2, 3", + "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1", + "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 3", + "[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.2", + "[10] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6124-6132, 2022. 2, 3, 4", + "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE CVPR, pages 2432-2443, 2017. 2, 5", + "[12] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(3):24:1-24:18, 2017. 1", + "[13] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2", + "[14] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 1", + "[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2, 3", + "[16] Jargon Kim, Hyomin Kim, Hyeonseo Nam, Jaesik Park, and Seungyong Lee. Textureme: High-quality textured scene reconstruction in real time. ACM Trans. Graph., 41(3), mar 2022. 1", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[18] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2", + "[19] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. 1", + "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2", + "[21] Marko Mihajlovic, Silvan Weder, Marc Pollefeys, and Martin R Oswald. Deepsurfels: Learning online appearance fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14524-14535, 2021. 2, 3", + "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 3, 5, 8", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 6", + "[24] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, pages 127-136, 2011. 1", + "[25] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 2", + "[26] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "116", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, June 2022. 2", + "[27] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. arXiv preprint arXiv:2111.14643, 2021. 1", + "[28] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 6", + "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3", + "[30] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. arXiv preprint arXiv:2110.06635, 2021. 2", + "[31] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 6", + "[32] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022. 3", + "[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021. 1", + "[34] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In IEEE CVPR, pages 15598-15607, 2021. 2, 3, 4", + "[35] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. arXiv preprint arXiv:2202.05263, 2022.1", + "[36] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2", + "[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3, 6", + "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[39] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 1", + "[40] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 6", + "[41] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2, 3, 4, 7", + "[42] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 1, 2", + "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 3, 5, 6", + "[44] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. European Conference on Computer Vision (ECCV), 2018. 3", + "[45] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2", + "[46] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3", + "[47] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2", + "[48] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 5", + "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[50] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 2, 3, 4, 5, 6", + "[51] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "117", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In IEEE CVPR, pages 12786-12796, 2022. 3", + "[53] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "118", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_model.json b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ae8e59d528d2096f2660957d51596d68eae7caad --- /dev/null +++ b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.13, + 0.833, + 0.175 + ], + "angle": 0, + "content": "SurfelNeRF: Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.204, + 0.671, + 0.222 + ], + "angle": 0, + "content": "Yiming Gao Yan-Pei Cao Ying Shan" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.223, + 0.583, + 0.239 + ], + "angle": 0, + "content": "ARC Lab, Tencent PCG" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.243, + 0.796, + 0.257 + ], + "angle": 0, + "content": "gaoym9@mail3.sysu.edu.cn caoyanpei@gmail.com yingsshan@tencent.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Online reconstructing and rendering of large-scale indoor scenes is a long-standing challenge. SLAM-based methods can reconstruct 3D scene geometry progressively in real time but can not render photorealistic results. While NeRF-based methods produce promising novel view synthesis results, their long offline optimization time and lack of geometric constraints pose challenges to efficiently handling online input. Inspired by the complementary advantages of classical 3D reconstruction and NeRF, we thus investigate marrying explicit geometric representation with NeRF rendering to achieve efficient online reconstruction and high-quality rendering. We introduce SurfelNeRF, a variant of neural radiance field which employs a flexible and scalable neural surfel representation to store geometric attributes and extracted appearance features from input images. We further extend the conventional surfel-based fusion scheme to progressively integrate incoming input frames into the reconstructed global neural scene representation. In addition, we propose a highly-efficient differentiable rasterization scheme for rendering neural surfel radiance fields, which helps SurfelNeRF achieve \\(10 \\times\\) speedups in training and inference time, respectively. Experimental results show that our method achieves the state-of-the-art 23.82 PSNR and 29.58 PSNR on ScanNet in feedforward inference and per-scene optimization settings, respectively.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.732, + 0.21, + 0.748 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Large-scale scene reconstruction and rendering is a crucial but challenging task in computer vision and graphics with many applications. Classical visual simultaneous localization and mapping (SLAM) systems [6, 12, 16, 24, 39, 41] can perform real-time 3D scene reconstruction. However, they usually represent the scene geometry as solid surfaces and appearance as vertex color or texture maps; thus," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.293, + 0.895, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.488, + 0.895, + 0.572 + ], + "angle": 0, + "content": "Figure 1. Examples to illustrate the task of online photorealistic reconstruction of an indoor scene. The online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.765 + ], + "angle": 0, + "content": "the reconstructed results fail to fully capture the scene content and cannot be used for photorealistic rendering. Recently, neural radiance fields (NeRF) and its variants [2, 3, 22,27,35,42] have achieved unprecedented novel view synthesis quality on both object-centric and large-scale scenes. However, NeRFs suffer from long per-scene optimization time and slow rendering speed, especially for large-scale scenes. Although recent advances [7, 14, 19, 23, 33] achieve faster optimization and rendering via incorporating explicit representations, they still require gathering all input images in an offline fashion before optimizing each scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.896, + 0.903 + ], + "angle": 0, + "content": "In this paper, we target the challenging task of online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views. The online setting can unlock a variety of real-time interactive applications, providing crucial immediate feedback to users during 3D capture. However, this task brings multiple extra requirements, including the scalability" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "1Project website: https://gymat.github.io/SurfelNeRF-web" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "108" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.812, + 0.198 + ], + "angle": 0, + "content": "
MethodsRepresentationGeneralizationReal-time RenderingOnline FusionScalability
DeepSurfel [21]SurfelsX✓XX
Instant-NGP [23]Hash GridsX✓XX
PointNeRF [43]Point Clouds✓XX✓
VBA [10]B+ TreesXX✓✓
NeRFusion [50]Voxel Grids✓X✓X
OursSurfels✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.263, + 0.207, + 0.706, + 0.221 + ], + "angle": 0, + "content": "Table 1. Comparison of representation and features with existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.473, + 0.504 + ], + "angle": 0, + "content": "ity of the underlying scene representation, the ability to perform on-the-fly updates to the scene representation, and optimizing and rendering at interactive framereates. Recently, NeRFusion [50] followed NeuralRecon [34] to unproject input images into local sparse feature volumes, fusing them to a global volume via Gated Recurrent Units (GRUs), and then generating photorealistic results from the global feature volume via volume rendering. However, updating the sparse volumetric feature involves computationally heavy operations; the volume rendering is also very slow since it requires hundreds of MLP evaluations to render a pixel. Thus, although NeRFusion achieves efficient online scene reconstruction, it still needs dozens of seconds to render a frame. VBA [10] is another recent approach to online photorealistic scene capture, but it only applies to object-centric scenes. We compare with representation and key features used in online photorealistic rendering with existing methods, which is shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.702 + ], + "angle": 0, + "content": "We propose surfel-based neural radiance fields, SurfelNeRF, for online photorealistic reconstruction and rendering of large-scale indoor scenes. Surfels (surface elements) [25] are point primitives containing geometric attributes, e.g., position, color, normal, and radius. We extend this representation to neural surfels, storing extra neural features that encode the neural radiance field of the target scene. Compared with volumetric representations, neural surfels are more compact and flexible and can easily scale to large scenes. Besides, we further employ a fast and differentiable rasterization process to render neural surfel radiance fields, which produces a pixel with only a few MLP evaluations based on the rasterized neural surfels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Inspired by classical real-time surfel-based geometric reconstruction methods [6, 15, 41], we propose an efficient neural radiance field fusion method to progressively build the scene representation by integrating neighboring neural surfels. Unlike point-based representations [1, 26, 30, 43] that are computationally heavy when finding neighboring points, it is easier to locate overlapping surfels and then merge neural features from multiview observations. By coupling the SurfelNeRF representation, the efficient neural surfel fusion approach, and the fast neural surfel rasterization algorithm, we achieve high-quality, photorealistic 3D scene reconstruction in an online manner." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on the large-scale indoor scene" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.895, + 0.416 + ], + "angle": 0, + "content": "dataset ScanNet [11], which contains complex scene structures and a large variety of scene appearances. We train the SurfelNeRF end-to-end across the scenes on the ScanNet, obtaining a generalizable model that enables both feedforward inference on unseen data and per-scene fine-tuning. We demonstrate in experiments that the proposed SurfelNeRF achieves favorably better rendering quality than the state-of-the-art approaches in both feedforward and finetuning settings while maintaining high training and rendering efficiency. We believe the proposed online photorealistic reconstruction framework has great potential in practical applications." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.434, + 0.642, + 0.451 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Novel view synthesis. Novel view synthesis is a challenging task for computer vision and computer graphics. Previous methods utilize geometry-based view synthesis. Zhou et al. [51] proposes to use the multiplane image (MPI) representation for novel view synthesis. Recently, neural radiance fields [22] and neural-based geometry representation have attracted more interest due to high-quality rendering. Recent methods propose to associate neural features on geometry structure and then synthesize novel views via rendering rasterized features, such as point cloud [1, 30], texture [36] and surfel [21]. DeepSurfel performs online appearance fusion only on a fixed given geometry during online fusion, while our SurfelNeRF performs online appearance and geometry integration. To render outputs, DeepSurfel rasterizes surfels and then passes them through 2D CNN filters, while our SurfelNeRF performs neural volume rendering which could achieve higher-quality rendering. Some recent works [9, 13, 18, 45] focus on real-time neural rendering methods. However, these methods all require per-scene optimization, which makes them impractical to handle novel scenes. In contrast, our SurfelNeRF is generalizable, which can reconstruct new scenes via direct feedforward inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Fields. Recently, neural radiance fields (NeRF) methods [2, 22, 42] achieve promising results on novel view synthesis, but it requires expensive per-scene optimization. Some methods proposed to reduce the long per-scene optimization time via deferred neural rendering [13, 47], hash table [23] or geometry prior [20]. Recent," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "109" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "generalization NeRF methods [5,8,37,43,46] are proposed to produce novel view without per-scene optimization by learning a radiance fields representation from offline given images of novel scenes. PixelNeRF [46] and IBRNet [37] render novel view images using volume rendering based on corresponding warped features from nearby given images. MVSNeRF [8] reconstructs local neural volumes from local nearby views, and then conducts volume rendering based on volumes. FWD [5] renders novel views by warping all input images to the target view, and fusing and refining features of these warped inputs. However, these methods are generalizable on local small-scale scenes, and require taking all input images before rendering novel views in an offline fashion. PointNeRF [43] learns to reconstruct a point-based neural radiance field for the novel scene. However, point-based representation is redundant, especially on large-scale scenes, and the point clouds cannot be updated based on multi-view input. Unlike these works, our method can represent not only large-scale scenes but also update compact surfel representation in an online fashion." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.394, + 0.473, + 0.711 + ], + "angle": 0, + "content": "Online scene reconstruction and rendering. Most online scene reconstruction methods [4, 34, 41, 52, 53] focus on 3D reconstruction only via TSDF-based fusion [4, 34], surfel-based fusion [41] and SLAM-based reconstruction [52]. However, they cannot render photorealistic novel view synthesis. Recently, some methods [10, 32, 50] employ radiance fields to tackle online scene representation for high-quality rendering. VBA [10] builds a neural dynamic B+Tree to represent scenes and apply volume rendering on it, but it still requires per-scene optimization. SRT [32] represents scenes with set-latent with a transformer encoder and decoder, but it cannot synthesize high-resolution images, which makes it not ready for real-world applications. NeRFusion [50] represents large-scale scenes with neural voxel grids, update them via GRU in an online fashion, and renders results via slow volume rendering. However, grid-based representation is wasteful for large-scale scenes, and is inefficient during features update due to features in empty areas are also updated. In contrast, our surfel representation is compact for large-scale scenes, and surfel-based fusion and corresponding rendering are more efficient." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.169, + 0.74 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Unlike most existing methods [22, 23, 43] that build neural radiance fields in an offline fashion, we aim for the task of online photorealistic scene reconstruction. Given a sequence of images \\(\\{I_n\\}\\), the goal is to progressively build the neural scene representation on the fly as the input images arrive." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To achieve this goal, we present SurfelNeRF, an online neural reconstruction framework based on the neural surfel radiance fields representation. The insight of our method is that the classical explicit surfel representation is efficient" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "for both reconstruction and rendering, has compact memory footprints, has the scalability to large and complex scenes, and can be naturally extended to equip neural attributes. Meanwhile, one can integrate multi-frame information and perform incremental updates on neural surfels in a geometrically meaningful way, and the operations involved can be streamlined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Figure 2 gives an overview of our proposed method. The framework first learns to generate per-view local neural surfel radiance fields \\( S_{t} \\) for each input frame \\( I_{t} \\). We then adopt the conventional surfel fusion process [6,15,41] to the neural setting, parameterizing the neural surfel fusion procedural as a neural network \\( N_{f} \\). Through \\( N_{f} \\), per-view representations \\( \\{S_t\\} \\) can be integrated into a global neural surfel radiance field \\( S_{g} \\) on the fly, aggregating geometric and neural attributes. The neural surfel radiance fields can be efficiently rendered to photorealistic images using a differentiable rasterizer without extra optimization. The framework can be trained end-to-end on large-scale indoor scene datasets. Once trained, it can generalize to previously unseen data and perform online photorealistic reconstruction; if needed, the rendering quality can be further improved by a short period time of fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.452, + 0.866, + 0.467 + ], + "angle": 0, + "content": "3.1. Neural Surfel Radiance Field Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Surfel Representation. Adopted from classical 3D reconstruction research, each surfel \\( s^n \\) is equipped with the following geometric attributes: a position \\( \\mathbf{p}^n \\in \\mathbb{R}^3 \\), normal \\( \\mathbf{n}^n \\in \\mathbb{R}^3 \\), weight \\( w^n \\in \\mathbb{R} \\) and radius \\( r^n \\in \\mathbb{R} \\). The weight \\( w^n \\) measures the confidence of a surfel based on the distance of the current depth measurement from the camera center. The radius \\( r^n \\) of each surfel is designed to represent the local support area around the projected 3D position while minimizing visible holes. The weight and radius of each surfel are initialized following [41]. Besides geometric properties, we further associate each surfel with neural features \\( \\mathbf{f}^n \\in \\mathbb{R}^N \\) from input images for modeling semi-transparent structures and view-dependent effects. Although a similar representation has been explored in DeepSurfels [21], we construct neural surfels in a feedforward manner and incorporate them in progressive scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Local Neural Surfel Radiance Field. Given an input frame \\(I_{t}\\), camera parameters \\(\\mathbf{c}_t\\), and estimated depth \\(D_{t}\\), the corresponding local neural surfel radiance field \\(\\mathcal{S}_t\\) can be constructed by unprojecting pixels in the input image to 3D space using camera parameters, then estimating corresponding normal, weight, radius, and neural features. The depth \\(D_{t}\\) can be captured from sensors or estimated via off-the-shelf MVS-based methods [44], which might be incomplete and noisy. Thus, we employ a U-Net structure [29] to refine the depth following [5]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Neural Features on Surfels. We estimate neural features" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.852, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Overview of our SurfelNeRF. Given an online input stream of image sequences, we first reconstruct a surfel representation associated with neural features to build a local neural surfel radiance field for each input keyframe. Then the neural surfel radiance field integration is used to fuse the input local neural surfel radiance field into the global neural surfel radiance field by updating both surfel position and features. More specifically, input local neural surfels associated with global surfels are fused to global surfels with corresponding global ones, and remaining local surfels without corresponding ones are added to the global model. Furthermore, the novel views can be rendered from updated global surfels via our efficient rasterization-guided render. Our proposed rasterization-guided renderer renders color only on the intersection points of ray and surfels, which is faster than volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.502 + ], + "angle": 0, + "content": "\\(\\mathbf{f}^n\\) of surfels by extracting corresponding pixel features in the input image. We employ a 2D CNN to extract features of input images. Specifically, we use a MasNet following [34, 50] to extract multi-scale view-dependent features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.393, + 0.529 + ], + "angle": 0, + "content": "3.2. Online Integration of Neural Surfels" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.672 + ], + "angle": 0, + "content": "Based on the neural surfel representation, we adopt a geometry-guided fusion scheme to efficiently and progressively integrate local neural surfel radiance fields into a global model. Compared with previous volumetric fusion methods [10,50], our surfel-based solution offers the advantages of 1) efficiency - the fusion operation only needs to be performed at surfel locations and 2) scalability - the memory footprint only grows linearly to the number of neural surfels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.764 + ], + "angle": 0, + "content": "In particular, our fusion scheme can be divided into two steps. Firstly, for each local neural surfel, we determine whether it should be added as a global neural surfel or should be merged with corresponding global surfels. Secondly, we insert it into the global model or perform updates to corresponding global surfels with local features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Neural Surfel Association. Given an input image frame \\( I_{t} \\) and corresponding camera parameter, each surfel in global surfels can be projected as an ellipse onto the image plane of the current input camera view. For each pixel in the input image, we obtain top- \\( K \\) global surfels that are overlapped with it after rasterization. We discard irrelevant surfels with a normal filter. Then we determine the incremental surfel by comparing the depth distance between respective top-M global surfels from near to far step by step. If the min" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.516 + ], + "angle": 0, + "content": "imal distance between a local surfel and its global surfel candidates is smaller than a threshold \\(\\delta_{depth}\\), then we determine that this input surfel needs to be merged; otherwise, it should be directly added to global surfels. The overall workflow follows [41], and \\(K\\) is set to 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.893, + 0.652 + ], + "angle": 0, + "content": "Neural Surfel Fusion. After the surfel association, there are a set of input surfels that need to be added directly and another set of surfels that will be merged with corresponding global surfels. For the former set of surfels, we directly construct neural surfels (Sec. 3.1) and then add them to global surfels. For the latter surfels, we use a GRU network to update neural features between the features \\(\\mathbf{f}_t^{\\mathrm{merge}}\\) of input surfels and features \\(\\mathbf{f}_{t - 1}^{\\mathrm{corr}}\\) of corresponding global surfels, which is given as" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.657, + 0.891, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t} ^ {\\text {c o r r s}} = \\operatorname {G R U} \\left(\\mathbf {f} _ {t} ^ {\\text {m e r g e}}, \\mathbf {f} _ {t - 1} ^ {\\text {c o r r s}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.756 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_t^{\\mathrm{merge}}\\) and \\(\\mathbf{f}_{t - 1}^{\\mathrm{corrs}}\\) are input and hidden state of GRU respectively. We then update the geometric attributes of these global surfels via a weighted sum. The update rule of each respective pair of local surfel \\(s_t^j\\) and global surfel \\(s_{t - 1}^{j}\\) can be expressed by" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.762, + 0.783, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {t} ^ {j} = \\frac {w _ {t} ^ {i} \\mathbf {p} _ {t} ^ {i} + w _ {t - 1} ^ {j} \\mathbf {p} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.804, + 0.892, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} _ {t} ^ {j} = \\frac {w _ {t} ^ {i} \\mathbf {n} _ {t} ^ {i} + w _ {t - 1} ^ {j} \\mathbf {n} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.845, + 0.777, + 0.883 + ], + "angle": 0, + "content": "\\[\nr _ {t} ^ {j} = \\frac {w _ {t} ^ {i} r _ {t} ^ {i} + w _ {t - 1} ^ {j} r _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.886, + 0.73, + 0.904 + ], + "angle": 0, + "content": "\\[\nw _ {t} ^ {j} = w _ {t} ^ {i} + w _ {t - 1} ^ {j},\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "111" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.473, + 0.32 + ], + "angle": 0, + "content": "Figure 3. Our efficient rasterization-guided surfel rendering for neural surfel radiance fields is visualized using a toy model as an example. We begin by rasterizing neural surfels into the image plane, and then render the ray using shading points on the intersections between the ray and corresponding surfels. This method uses a combination of rasterization and ray tracing to render view-dependency results with surfel representation and 3D rendering." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.47, + 0.375 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_t^j\\), \\(\\mathbf{n}_t^j\\), \\(r_t^j\\) and \\(w_t^j\\) are the updated position, normal, radius, and weight of the global surfels, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.382, + 0.431, + 0.399 + ], + "angle": 0, + "content": "3.3. Rendering Neural Surfel Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.473, + 0.558 + ], + "angle": 0, + "content": "We propose a fast rasterization-based rendering algorithm for neural surfel radiance fields (see Figure. 3). Given a neural surfel radiance field \\( S \\) and the camera parameter \\( \\mathbf{c}_{tar} \\) of a target view, we first conduct a differentiable rasterization process on the neural surfels to project them on the target image plane. For a pixel \\( p_x \\) in the target image, we gather the neural surfels \\( \\{s_i\\}_{i=1}^M \\) that cover it and render its color \\( C(p_x) \\) based on a few shading points \\( \\{\\mathbf{x}_i\\}_{i=1}^M \\) that are the intersections between \\( \\{s_i\\}_{i=1}^M \\) and pixel ray, using the following volume rendering formula [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.566, + 0.47, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} C \\left(p _ {x}\\right) = \\sum_ {i - 1} ^ {M} \\tau_ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\Delta_ {i}\\right)\\right) c _ {i}, \\tag {3} \\\\ \\tau_ {i} = \\exp (- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\Delta_ {j}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.47, + 0.756 + ], + "angle": 0, + "content": "where \\(\\tau\\) indicates the accumulated transmittance; \\(\\sigma_{i}\\) and \\(c_{i}\\) indicate the volume density and radiance on shading point \\(\\mathbf{x}_i\\); \\(\\Delta_j\\) is the distance between adjacent shading points. For each shading point \\(x_{i}\\), we regress its volume density \\(\\sigma_{i}\\) and radiance \\(c_{i}\\) based on its view direction \\(\\mathbf{d}\\) and shading point features \\(\\mathbf{f}^i (\\mathbf{x}_i)\\) associated from surfel \\(s_i\\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.767, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i}, c _ {i} = \\operatorname {R e n d e r} \\left(x _ {i}, \\mathbf {f} ^ {i} \\left(\\mathbf {x} _ {i}\\right), \\mathbf {d}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.793, + 0.468, + 0.824 + ], + "angle": 0, + "content": "where we use an MLP to parameterize the Render module. The \"interpolated\" surfel feature \\(\\mathbf{f}^i (\\mathbf{x}_i)\\) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.833, + 0.47, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {i} \\left(\\mathbf {x} _ {i}\\right) = \\frac {\\| \\mathbf {x} ^ {i} - \\mathbf {p} ^ {i} \\|}{r ^ {i}} \\mathrm {F} \\left(\\mathbf {f} ^ {i}, \\mathbf {d}, \\mathbf {n} ^ {i}, w ^ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( \\mathbf{F} \\) is also parameterized as an MLP-like network. Note that a recent concurrent work [48] also explores the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.893, + 0.136 + ], + "angle": 0, + "content": "hybrid rendering process by combining rasterization and volume rendering, but using a different neural representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.146, + 0.744, + 0.163 + ], + "angle": 0, + "content": "3.4. Training and Optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.169, + 0.894, + 0.245 + ], + "angle": 0, + "content": "All modules in SurfelNeRF are differentiable, so they can be trained end-to-end on large-scale scene datasets with sequential images as input. We utilize L2 pixel-wise rendering loss to train our network supervised by the target ground truth image:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.255, + 0.893, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n d e r}} = \\left\\| C _ {g t} - C _ {r} \\right\\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.282, + 0.893, + 0.356 + ], + "angle": 0, + "content": "where \\( C_{gt} \\) is the ground truth pixel color and \\( C_r \\) denotes the rendered pixel color output from SurfelNeRF. In addition, if we can acquire sensor depth \\( D_i^{\\mathrm{sen}} \\) senor for image \\( I_i \\), we further employ an L1 loss to train the depth refinement network:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.368, + 0.892, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\left| \\left| D _ {i} ^ {\\text {s e n}} \\odot M _ {i} - D ^ {i} \\odot M _ {i} \\right| \\right| _ {1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.893, + 0.456 + ], + "angle": 0, + "content": "where \\(D^{i}\\) is the output from depth refinement network, and \\(M_{i}\\) is a binary mask indicating valid sensor depths. This supervision can make the depth refinement network to learn to match the sensor depths." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.456, + 0.894, + 0.485 + ], + "angle": 0, + "content": "To sum up, the total loss to train SurfelNeRF end-to-end is given as" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.486, + 0.892, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda \\cdot \\mathbb {1} \\left(\\mathcal {L} _ {d}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.508, + 0.892, + 0.538 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\) denotes the indicator function whose value is set to 1 if sensor depth is available; \\(\\lambda\\) is the balancing weight." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.894, + 0.613 + ], + "angle": 0, + "content": "Training Details. Following [50], we uniformly sample keyframes from each image sequence as input for training SurfelNeRF. To train SurfelNeRF, we sample about \\(5\\%\\) of the full sequence as keyframes, and other frames are used as supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.894, + 0.795 + ], + "angle": 0, + "content": "Per-scene Fine-tuning. After being trained on large-scale datasets, our SurfelNeRF is capable of constructing neural surfel radiance fields and rendering photorealistic novel views on unseen scenes in a feedforward manner. To further improve the rendering performance on test scenes, one can also perform a short period of per-scene optimization via fine-tuning the surfel representations and the rendering module. To make a fair comparison, we do not use the additional depths in the fine-tuning process and keep the depth refinement module fixed. Thanks to our efficient rendering scheme, SurfelNeRF takes less time when fine-tuning compared with existing methods [43, 50]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.634, + 0.825 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.708, + 0.849 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Training Datasets. Our training data consists of large-scale indoor scenes from ScanNet [11] which provides the online sequence input of images with known camera parameters," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "112" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "and the captured incomplete sensor depths. We focus on the online input of large-scale scenes in the main paper. Following [37, 50], we randomly sample 100 scenes from ScanNet for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.319 + ], + "angle": 0, + "content": "Implementation Details. We train our SurfelNeRF on the large-scale training dataset mentioned above using the Adam [17] optimizer with an initial learning rate of 0.001 for total 200k iterations on 4 NVIDIA Tesla V100 32GB GPUs. For per-scene optimization settings, we fine-tune our method also using the Adam optimizer with an initial learning rate of 0.0002. All input images are resized to \\(640 \\times 480\\). The \\(\\lambda\\) in Eq. 8 is set to 0.1. The threshold of \\(\\delta_{depth}\\) is set to \\(0.1\\mathrm{m}\\) in surfel fusion (Sec. 3.2). The maximum number of candidates M in Eq. 3 for rendering is set to 80. We implement our method using Pytorch and Pytorch3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.47, + 0.425 + ], + "angle": 0, + "content": "Metrics. Following previous works [50], we report the PSNR, SSIM [38] and LPIPS\\(_{vgg}\\) [49] to evaluate and compare the results. For fair comparisons with previous methods, we also evaluate our method on the same test set on the ScanNet dataset. The time cost of rendering and perscene fine-tuning is measured on the same platform using one NVIDIA Tesla V100 32GB GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.377, + 0.455 + ], + "angle": 0, + "content": "4.2. Comparison with state-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.596 + ], + "angle": 0, + "content": "In this section, we evaluate our method with other state-of-the-arts on the large-scale dataset, ScanNet. To achieve fair comparisons, we evaluate our methods strictly following the same training and evaluation scheme as NeRFusion [50] and NerfingMVS [40]. There are 8 selected testing scenes in ScanNet where 1/8 of images are held out for testing. Following [50], we report no per-scene optimization results via direct inference network and per-scene optimization results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Baselines. In addition to comparing SurfelNeRF with NeRF-Fusion [50], we also compare it with other state-of-the-art (SOTA) methods, such as fast per-scene optimization NeRF and neural point cloud representation, generalizable NeRF, scalable MLP-based representation NeRF, NeRF with additional depth supervision, and classical image-based rendering methods, including Instant-NGP [23], ADOP [31], PointNeRF [43], Mip-NeRF-360 [3], Dense Depth Prior NeRF/DDP-NeRF) [28], and IBRNet [37]. Among these methods, Mip-NeRF-360 is one of the state-of-the-art NeRF methods for large-scale scenes. DDP-NeRF employs an offline NeRF with additional depth supervision. PointNeRF achieves state-of-the-art performance thanks to point-based representation. ADOP achieves promising results on offline large-scale rendering. These two methods employ a similar representation to our method; thus we also compare with them. We run generalizable PointNeRF with the same No per-scene optimization and per-scene optimization settings using its open-source code." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The result is shown in Tab. 2 (a). For no per-scene" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.875, + 0.181 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
IBRNet [37]21.190.7860.358-
NeRFusion [50]22.990.8380.33538s#
PointNeRF [43]20.470.6420.54430s
SurfelNeRF23.820.8450.3270.2s
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.231 + ], + "angle": 0, + "content": "(a) Results of No per-scene optimization setting on ScanNet. And we report the average time to render an image. \\(\\sharp\\) indicates that the time of NeR-Fusion is reported by authors using a 2080Ti since they have not released executable code before submission." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.245, + 0.885, + 0.413 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
Instant-NGP [23]23.230.7140.4590.03
ADOP [31]25.010.8070.2721s
NeRFingMVS [40]26.370.9030.245-
IBRNet [37]25.140.8710.266-
DDP-NeRF [28]26.610.800.26-
Mip-NeRF-360 [3]27.850.860.25-
NeRFusion [50]26.490.9150.20938s
PointNeRF [43]28.990.8290.32430s
SurfelNeRF29.580.9190.2150.2s
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.415, + 0.892, + 0.44 + ], + "angle": 0, + "content": "(b) Per-scene optimization. Time↓ indicates the average time to render an image." + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.452, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons with SOTAs on the ScanNet dataset with No per-scene and per-scene optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.732 + ], + "angle": 0, + "content": "optimization settings, NeRFusion achieves PSNR of 22.99 thanks to the fusion between input features to global grid representations. Our SurfelNeRF outperforms NeRFusion since SurfelNeRF benefits from compact and precise feature fusion and the geometric guidance of surfels. For the per-scene optimization task, the result is shown in Tab. 2 (b). PointNeRF gains a large improvement after fine-tuning, which illustrates it achieves sub-optimal performance in no per-scene optimization settings due to its inability to fuse and update point features. Compared with DDP-NeRF, which employs additional depth supervision, our SurfelNeRF still outperforms it by taking advantage of surfel representation and its integration. Our method achieves state-of-the-art results again, which shows that our neural surfel representation can provide optimal initialization for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.839 + ], + "angle": 0, + "content": "The qualitative comparison is shown in Figure. 4. We compare our method with PointNeRF without per-scene optimization, and compare with per-scene optimization methods of PoingNeRF and ADOP. Since NeRFusion has not release the executable code, we cannot generate qualitative results. As shown in Figure. 4, our results are visually much better than state-of-the-art methods, PointNeRF and ADOP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.655, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We investigate the effectiveness of designs in our SurfelNeRF by conducting the following ablation studies on" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "113" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.092, + 0.837, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.614, + 0.895, + 0.687 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of rendering quality. The 1-2 column are the results of no per-scene optimization. And the left three columns are per-scene optimization results. It is obvious that with or without per-scene optimization, our method can still achieve the most photorealistic rendering results. Since PointNeRF without a depth refinement network relies on captured depths only, it cannot handle the scenario that depth is incomplete. However, our method can tackle this problem with a lightweight depth refinement network. ADOP can produce geometrically correct results, but its color is less photorealistic." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.142, + 0.714 + ], + "angle": 0, + "content": "ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Fusion Schemes. Our surfel-based fusion is inspired by classical Surfel-based fusion in conventional 3D reconstruction [41]. The classical surfel-based fusion update features via the simple weighted sum of two overlapped surfels. We conduct an experiment that updates surfel features by weighted sum based on surfel weights, which is a similar operation of updating geometry attributes in Eq. 2. The results are shown in Tab. 3. Since the GRU module can learn to adaptively update features based on the global surfels and new incremental surfels, it can improve performance with negligible computation overhead." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Depth Refinement Module. To investigate the effective-" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.698, + 0.908, + 0.791 + ], + "angle": 0, + "content": "
Fusion SchemeSettingPSNR↑SSIM↑LPIPS ↓
Weighted SumNo per-scene23.090.8330.353
GRUoptimization23.820.8450.327
Weighted SumPer-scene28.540.8840.293
GRUoptimization29.580.9190.215
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.801, + 0.892, + 0.815 + ], + "angle": 0, + "content": "Table 3. Ablation studies about fusion schemes in our SurfelNeRF." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ness of our depth refinement, we conduct experiments that reconstruct scenes with incomplete depth captured by sensors. The quantitative results are shown in Tab. 4, which show that PSNR performance drops about \\(4\\%\\) without the" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "114" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.112, + 0.089, + 0.437, + 0.148 + ], + "angle": 0, + "content": "
Depth RefinementPSNR↑SSIM↑LPIPS ↓
w/o19.510.7130.454
w23.090.8330.353
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.158, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Table 4. Ablation studies about the effectiveness of depth refinement in scene0000_01 in the ScanNet with the no per-scene optimization setting. The depth captured by the sensor in this scene is incomplete. If using captured depths only, some pixels cannot be projected to the 3D space, as the visualization results of PointNeRF shown in the first row in the qualitative results in Fig. 4." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.256, + 0.454, + 0.349 + ], + "angle": 0, + "content": "
Rendering SchemeSettingPSNR↑Time ↓
Volume RenderingNo per-scene23.8030s
Oursoptimization23.820.2s
Volume RenderingPer-scene29.4530s
Oursoptimization29.580.2s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.4 + ], + "angle": 0, + "content": "Table 5. Comparing with two rendering schemes, volume rendering [22] and our proposed rasterization-guided rendering, in the ScanNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.598 + ], + "angle": 0, + "content": "depth refinement. Depth quality affects surfel reconstruction geometry quality, but captured depth is always incomplete (visualized in the first row in Fig. 4). Thus, a depth refinement network is necessarily required to refine the input depth for better surfel geometry quality. The experimental result shows that the depth refinement network is effective in alleviating the noisy or incomplete captured or estimated depth. Moreover, we also conduct an experiment with worse depth quality that resizes input depth resolution to half size. Our performance just slightly drops \\(0.5\\%\\) PSNR. These experiments show the effectiveness of our depth refinement module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.809 + ], + "angle": 0, + "content": "**Rendering Schemes.** Our proposed renderer renders a ray with a few shading points, which are the intersection points of ray and surfels, unlike volume rendering in NeRF [22] that requires hundreds of shading points each with one MLP evaluation. To validate the effectiveness of our differentiable rasterization scheme for rendering, we conduct an experiment that replaced our rendering scheme with NeRFs' volume rendering. The result is shown in Tab. 5. Since online integration of surfels can limit the number of intersecting surfels, our rendering speed is persistently faster than volume rendering with fewer shading points. Specifically, The average and maximal number of rasterized surfels per pixel at novel views on ScanNet are 2.92 and 10, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Computational Cost and Memory Usage. To evaluate the computational cost and memory usage on a large-scale scene, we plot components of GPU processing time, surfel number, and memory usage of our method on a large three rooms scene measured on a V100 GPU implemented by Pytorch3D. During inference, our method renders \\(640 \\times 480\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.099, + 0.691, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.188, + 0.663, + 0.201 + ], + "angle": 0, + "content": "(a) GPU processing time." + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.099, + 0.895, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.188, + 0.862, + 0.201 + ], + "angle": 0, + "content": "(b) GPU memory usage." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.212, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Figure 5. The computation cost, memory footprint, and the number of surfels grow according to input frames measured on a three rooms scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.432 + ], + "angle": 0, + "content": "images at 5.2FPS on average. As shown in Fig. 5a and Fig. 5b, the number of surfels has a sub-linear growth thanks to the proposed surfel integration. The fluctuation of Time (Rasterization) is caused by the different numbers of corresponding surfels at novel views. Although the computational cost of rasterization depends on the number of rasterized surfels, but it is still faster than volume rendering [22]. These results show that our surfel representation is compact and efficient. The major computation cost comes from rasterization, and the cost of surfel construction as well as integration could be neglected." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.445, + 0.623, + 0.461 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Our method requires high-quality surfel geometry as neural surfel representation since SurfelNeRF relies on it to conduct surfel integration and rasterization-guided render. To obtain better geometry quality, we may require more accurate depth quality to construct surfel geometry, which necessitates extra depth supervision or the MVS depth estimator. Since captured or estimated depth quality is not exact, there may be geometry discontinuities or missing parts in the surfel representation. Consequently, there may be some missing parts or incorrect geometry in the rendered output (examples of missing parts are shown in the first row in Fig. 4 or video). As discussed in ablation studies of depth refinement above, inaccurate and noisy estimated depth as well as depth discontinuity could affect the construction of neural surfel representation. To alleviate this issue, we propose a depth refinement module (Tab. 4) to refine depth information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.619, + 0.755 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose SurfelNeRF for online photorealistic reconstruction and rendering of large-scale indoor scenes. We employ neural surfels to store a neural radiance field, adopt an efficient neural radiance field fusion method to progressively build the scene representation, and propose a fast and differentiable rasterization-based rendering algorithm to tackle the challenging task. Experiments on the ScanNet dataset demonstrate that our method achieving favorably better results than the state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "115" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.155 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.468, + 0.239 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.31 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[4] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.435 + ], + "angle": 0, + "content": "[5] Ang Cao, Chris Rockwell, and Justin Johnson. Fwd: Real-time novel view synthesis with forward warping and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15713-15724, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.468, + 0.492 + ], + "angle": 0, + "content": "[6] Yan-Pei Cao, Leif Kobbelt, and Shi-Min Hu. Real-time high-accuracy three-dimensional reconstruction with consumer RGB-D cameras. ACM Transactions on Graphics, 37(5):171:1-171:16, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.732 + ], + "angle": 0, + "content": "[10] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6124-6132, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE CVPR, pages 2432-2443, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[12] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(3):24:1-24:18, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[14] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[16] Jargon Kim, Hyomin Kim, Hyeonseo Nam, Jaesik Park, and Seungyong Lee. Textureme: High-quality textured scene reconstruction in real time. ACM Trans. Graph., 41(3), mar 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[18] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[19] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[21] Marko Mihajlovic, Silvan Weder, Marc Pollefeys, and Martin R Oswald. Deepsurfels: Learning online appearance fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14524-14535, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[24] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, pages 127-136, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[25] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "116" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[27] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. arXiv preprint arXiv:2111.14643, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[28] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[30] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. arXiv preprint arXiv:2110.06635, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[31] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[32] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[34] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In IEEE CVPR, pages 15598-15607, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[35] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. arXiv preprint arXiv:2202.05263, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[36] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[41] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[42] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[44] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[45] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[46] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[47] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[48] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[50] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "117" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In IEEE CVPR, pages 12786-12796, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[53] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "118" + } + ] +] \ No newline at end of file diff --git a/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_origin.pdf b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dde9dff7da55108c053bb5986427c6f037cbc444 --- /dev/null +++ b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/ca098c26-5851-4e50-aa91-b22763ee66ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d6c9304a45d537b13ad334372767ad2c802aaa42e0e784cbdf3eaceebf7a4f9 +size 2028091 diff --git a/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/full.md b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dcbcef21b7c8a2cce36c6c71017157ff33ec7584 --- /dev/null +++ b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/full.md @@ -0,0 +1,306 @@ +# SurfelNeRF: Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes + +Yiming Gao Yan-Pei Cao Ying Shan + +ARC Lab, Tencent PCG + +gaoym9@mail3.sysu.edu.cn caoyanpei@gmail.com yingsshan@tencent.com + +# Abstract + +Online reconstructing and rendering of large-scale indoor scenes is a long-standing challenge. SLAM-based methods can reconstruct 3D scene geometry progressively in real time but can not render photorealistic results. While NeRF-based methods produce promising novel view synthesis results, their long offline optimization time and lack of geometric constraints pose challenges to efficiently handling online input. Inspired by the complementary advantages of classical 3D reconstruction and NeRF, we thus investigate marrying explicit geometric representation with NeRF rendering to achieve efficient online reconstruction and high-quality rendering. We introduce SurfelNeRF, a variant of neural radiance field which employs a flexible and scalable neural surfel representation to store geometric attributes and extracted appearance features from input images. We further extend the conventional surfel-based fusion scheme to progressively integrate incoming input frames into the reconstructed global neural scene representation. In addition, we propose a highly-efficient differentiable rasterization scheme for rendering neural surfel radiance fields, which helps SurfelNeRF achieve $10 \times$ speedups in training and inference time, respectively. Experimental results show that our method achieves the state-of-the-art 23.82 PSNR and 29.58 PSNR on ScanNet in feedforward inference and per-scene optimization settings, respectively. $^{1}$ + +# 1. Introduction + +Large-scale scene reconstruction and rendering is a crucial but challenging task in computer vision and graphics with many applications. Classical visual simultaneous localization and mapping (SLAM) systems [6, 12, 16, 24, 39, 41] can perform real-time 3D scene reconstruction. However, they usually represent the scene geometry as solid surfaces and appearance as vertex color or texture maps; thus, + +![](images/3c1222365b0a2fe61ffe68b3ebee31b0cf87c99fb6fd47f9946ae667013c53d4.jpg) +Figure 1. Examples to illustrate the task of online photorealistic reconstruction of an indoor scene. The online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views. + +the reconstructed results fail to fully capture the scene content and cannot be used for photorealistic rendering. Recently, neural radiance fields (NeRF) and its variants [2, 3, 22,27,35,42] have achieved unprecedented novel view synthesis quality on both object-centric and large-scale scenes. However, NeRFs suffer from long per-scene optimization time and slow rendering speed, especially for large-scale scenes. Although recent advances [7, 14, 19, 23, 33] achieve faster optimization and rendering via incorporating explicit representations, they still require gathering all input images in an offline fashion before optimizing each scene. + +In this paper, we target the challenging task of online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views. The online setting can unlock a variety of real-time interactive applications, providing crucial immediate feedback to users during 3D capture. However, this task brings multiple extra requirements, including the scalability + +
MethodsRepresentationGeneralizationReal-time RenderingOnline FusionScalability
DeepSurfel [21]SurfelsX✓XX
Instant-NGP [23]Hash GridsX✓XX
PointNeRF [43]Point Clouds✓XX✓
VBA [10]B+ TreesXX✓✓
NeRFusion [50]Voxel Grids✓X✓X
OursSurfels✓✓✓✓
+ +Table 1. Comparison of representation and features with existing methods. + +ity of the underlying scene representation, the ability to perform on-the-fly updates to the scene representation, and optimizing and rendering at interactive framereates. Recently, NeRFusion [50] followed NeuralRecon [34] to unproject input images into local sparse feature volumes, fusing them to a global volume via Gated Recurrent Units (GRUs), and then generating photorealistic results from the global feature volume via volume rendering. However, updating the sparse volumetric feature involves computationally heavy operations; the volume rendering is also very slow since it requires hundreds of MLP evaluations to render a pixel. Thus, although NeRFusion achieves efficient online scene reconstruction, it still needs dozens of seconds to render a frame. VBA [10] is another recent approach to online photorealistic scene capture, but it only applies to object-centric scenes. We compare with representation and key features used in online photorealistic rendering with existing methods, which is shown in Tab. 1. + +We propose surfel-based neural radiance fields, SurfelNeRF, for online photorealistic reconstruction and rendering of large-scale indoor scenes. Surfels (surface elements) [25] are point primitives containing geometric attributes, e.g., position, color, normal, and radius. We extend this representation to neural surfels, storing extra neural features that encode the neural radiance field of the target scene. Compared with volumetric representations, neural surfels are more compact and flexible and can easily scale to large scenes. Besides, we further employ a fast and differentiable rasterization process to render neural surfel radiance fields, which produces a pixel with only a few MLP evaluations based on the rasterized neural surfels. + +Inspired by classical real-time surfel-based geometric reconstruction methods [6, 15, 41], we propose an efficient neural radiance field fusion method to progressively build the scene representation by integrating neighboring neural surfels. Unlike point-based representations [1, 26, 30, 43] that are computationally heavy when finding neighboring points, it is easier to locate overlapping surfels and then merge neural features from multiview observations. By coupling the SurfelNeRF representation, the efficient neural surfel fusion approach, and the fast neural surfel rasterization algorithm, we achieve high-quality, photorealistic 3D scene reconstruction in an online manner. + +We conduct experiments on the large-scale indoor scene + +dataset ScanNet [11], which contains complex scene structures and a large variety of scene appearances. We train the SurfelNeRF end-to-end across the scenes on the ScanNet, obtaining a generalizable model that enables both feedforward inference on unseen data and per-scene fine-tuning. We demonstrate in experiments that the proposed SurfelNeRF achieves favorably better rendering quality than the state-of-the-art approaches in both feedforward and finetuning settings while maintaining high training and rendering efficiency. We believe the proposed online photorealistic reconstruction framework has great potential in practical applications. + +# 2. Related Work + +Novel view synthesis. Novel view synthesis is a challenging task for computer vision and computer graphics. Previous methods utilize geometry-based view synthesis. Zhou et al. [51] proposes to use the multiplane image (MPI) representation for novel view synthesis. Recently, neural radiance fields [22] and neural-based geometry representation have attracted more interest due to high-quality rendering. Recent methods propose to associate neural features on geometry structure and then synthesize novel views via rendering rasterized features, such as point cloud [1, 30], texture [36] and surfel [21]. DeepSurfel performs online appearance fusion only on a fixed given geometry during online fusion, while our SurfelNeRF performs online appearance and geometry integration. To render outputs, DeepSurfel rasterizes surfels and then passes them through 2D CNN filters, while our SurfelNeRF performs neural volume rendering which could achieve higher-quality rendering. Some recent works [9, 13, 18, 45] focus on real-time neural rendering methods. However, these methods all require per-scene optimization, which makes them impractical to handle novel scenes. In contrast, our SurfelNeRF is generalizable, which can reconstruct new scenes via direct feedforward inference. + +Neural Radiance Fields. Recently, neural radiance fields (NeRF) methods [2, 22, 42] achieve promising results on novel view synthesis, but it requires expensive per-scene optimization. Some methods proposed to reduce the long per-scene optimization time via deferred neural rendering [13, 47], hash table [23] or geometry prior [20]. Recent, + +generalization NeRF methods [5,8,37,43,46] are proposed to produce novel view without per-scene optimization by learning a radiance fields representation from offline given images of novel scenes. PixelNeRF [46] and IBRNet [37] render novel view images using volume rendering based on corresponding warped features from nearby given images. MVSNeRF [8] reconstructs local neural volumes from local nearby views, and then conducts volume rendering based on volumes. FWD [5] renders novel views by warping all input images to the target view, and fusing and refining features of these warped inputs. However, these methods are generalizable on local small-scale scenes, and require taking all input images before rendering novel views in an offline fashion. PointNeRF [43] learns to reconstruct a point-based neural radiance field for the novel scene. However, point-based representation is redundant, especially on large-scale scenes, and the point clouds cannot be updated based on multi-view input. Unlike these works, our method can represent not only large-scale scenes but also update compact surfel representation in an online fashion. + +Online scene reconstruction and rendering. Most online scene reconstruction methods [4, 34, 41, 52, 53] focus on 3D reconstruction only via TSDF-based fusion [4, 34], surfel-based fusion [41] and SLAM-based reconstruction [52]. However, they cannot render photorealistic novel view synthesis. Recently, some methods [10, 32, 50] employ radiance fields to tackle online scene representation for high-quality rendering. VBA [10] builds a neural dynamic B+Tree to represent scenes and apply volume rendering on it, but it still requires per-scene optimization. SRT [32] represents scenes with set-latent with a transformer encoder and decoder, but it cannot synthesize high-resolution images, which makes it not ready for real-world applications. NeRFusion [50] represents large-scale scenes with neural voxel grids, update them via GRU in an online fashion, and renders results via slow volume rendering. However, grid-based representation is wasteful for large-scale scenes, and is inefficient during features update due to features in empty areas are also updated. In contrast, our surfel representation is compact for large-scale scenes, and surfel-based fusion and corresponding rendering are more efficient. + +# 3. Method + +Unlike most existing methods [22, 23, 43] that build neural radiance fields in an offline fashion, we aim for the task of online photorealistic scene reconstruction. Given a sequence of images $\{I_n\}$ , the goal is to progressively build the neural scene representation on the fly as the input images arrive. + +To achieve this goal, we present SurfelNeRF, an online neural reconstruction framework based on the neural surfel radiance fields representation. The insight of our method is that the classical explicit surfel representation is efficient + +for both reconstruction and rendering, has compact memory footprints, has the scalability to large and complex scenes, and can be naturally extended to equip neural attributes. Meanwhile, one can integrate multi-frame information and perform incremental updates on neural surfels in a geometrically meaningful way, and the operations involved can be streamlined. + +Figure 2 gives an overview of our proposed method. The framework first learns to generate per-view local neural surfel radiance fields $S_{t}$ for each input frame $I_{t}$ . We then adopt the conventional surfel fusion process [6,15,41] to the neural setting, parameterizing the neural surfel fusion procedural as a neural network $N_{f}$ . Through $N_{f}$ , per-view representations $\{S_t\}$ can be integrated into a global neural surfel radiance field $S_{g}$ on the fly, aggregating geometric and neural attributes. The neural surfel radiance fields can be efficiently rendered to photorealistic images using a differentiable rasterizer without extra optimization. The framework can be trained end-to-end on large-scale indoor scene datasets. Once trained, it can generalize to previously unseen data and perform online photorealistic reconstruction; if needed, the rendering quality can be further improved by a short period time of fine-tuning. + +# 3.1. Neural Surfel Radiance Field Construction + +Surfel Representation. Adopted from classical 3D reconstruction research, each surfel $s^n$ is equipped with the following geometric attributes: a position $\mathbf{p}^n \in \mathbb{R}^3$ , normal $\mathbf{n}^n \in \mathbb{R}^3$ , weight $w^n \in \mathbb{R}$ and radius $r^n \in \mathbb{R}$ . The weight $w^n$ measures the confidence of a surfel based on the distance of the current depth measurement from the camera center. The radius $r^n$ of each surfel is designed to represent the local support area around the projected 3D position while minimizing visible holes. The weight and radius of each surfel are initialized following [41]. Besides geometric properties, we further associate each surfel with neural features $\mathbf{f}^n \in \mathbb{R}^N$ from input images for modeling semi-transparent structures and view-dependent effects. Although a similar representation has been explored in DeepSurfels [21], we construct neural surfels in a feedforward manner and incorporate them in progressive scene reconstruction. + +Local Neural Surfel Radiance Field. Given an input frame $I_{t}$ , camera parameters $\mathbf{c}_t$ , and estimated depth $D_{t}$ , the corresponding local neural surfel radiance field $\mathcal{S}_t$ can be constructed by unprojecting pixels in the input image to 3D space using camera parameters, then estimating corresponding normal, weight, radius, and neural features. The depth $D_{t}$ can be captured from sensors or estimated via off-the-shelf MVS-based methods [44], which might be incomplete and noisy. Thus, we employ a U-Net structure [29] to refine the depth following [5]. + +Neural Features on Surfels. We estimate neural features + +![](images/2b7b4e20fde4ff81e9924c8dcbac88033f882241b5f9eef8d8e50279d428fa70.jpg) +Figure 2. Overview of our SurfelNeRF. Given an online input stream of image sequences, we first reconstruct a surfel representation associated with neural features to build a local neural surfel radiance field for each input keyframe. Then the neural surfel radiance field integration is used to fuse the input local neural surfel radiance field into the global neural surfel radiance field by updating both surfel position and features. More specifically, input local neural surfels associated with global surfels are fused to global surfels with corresponding global ones, and remaining local surfels without corresponding ones are added to the global model. Furthermore, the novel views can be rendered from updated global surfels via our efficient rasterization-guided render. Our proposed rasterization-guided renderer renders color only on the intersection points of ray and surfels, which is faster than volume rendering. + +$\mathbf{f}^n$ of surfels by extracting corresponding pixel features in the input image. We employ a 2D CNN to extract features of input images. Specifically, we use a MasNet following [34, 50] to extract multi-scale view-dependent features. + +# 3.2. Online Integration of Neural Surfels + +Based on the neural surfel representation, we adopt a geometry-guided fusion scheme to efficiently and progressively integrate local neural surfel radiance fields into a global model. Compared with previous volumetric fusion methods [10,50], our surfel-based solution offers the advantages of 1) efficiency - the fusion operation only needs to be performed at surfel locations and 2) scalability - the memory footprint only grows linearly to the number of neural surfels. + +In particular, our fusion scheme can be divided into two steps. Firstly, for each local neural surfel, we determine whether it should be added as a global neural surfel or should be merged with corresponding global surfels. Secondly, we insert it into the global model or perform updates to corresponding global surfels with local features. + +Neural Surfel Association. Given an input image frame $I_{t}$ and corresponding camera parameter, each surfel in global surfels can be projected as an ellipse onto the image plane of the current input camera view. For each pixel in the input image, we obtain top- $K$ global surfels that are overlapped with it after rasterization. We discard irrelevant surfels with a normal filter. Then we determine the incremental surfel by comparing the depth distance between respective top-M global surfels from near to far step by step. If the min + +imal distance between a local surfel and its global surfel candidates is smaller than a threshold $\delta_{depth}$ , then we determine that this input surfel needs to be merged; otherwise, it should be directly added to global surfels. The overall workflow follows [41], and $K$ is set to 8. + +Neural Surfel Fusion. After the surfel association, there are a set of input surfels that need to be added directly and another set of surfels that will be merged with corresponding global surfels. For the former set of surfels, we directly construct neural surfels (Sec. 3.1) and then add them to global surfels. For the latter surfels, we use a GRU network to update neural features between the features $\mathbf{f}_t^{\mathrm{merge}}$ of input surfels and features $\mathbf{f}_{t - 1}^{\mathrm{corr}}$ of corresponding global surfels, which is given as + +$$ +\mathbf {f} _ {t} ^ {\text {c o r r s}} = \operatorname {G R U} \left(\mathbf {f} _ {t} ^ {\text {m e r g e}}, \mathbf {f} _ {t - 1} ^ {\text {c o r r s}}\right), \tag {1} +$$ + +where $\mathbf{f}_t^{\mathrm{merge}}$ and $\mathbf{f}_{t - 1}^{\mathrm{corrs}}$ are input and hidden state of GRU respectively. We then update the geometric attributes of these global surfels via a weighted sum. The update rule of each respective pair of local surfel $s_t^j$ and global surfel $s_{t - 1}^{j}$ can be expressed by + +$$ +\mathbf {p} _ {t} ^ {j} = \frac {w _ {t} ^ {i} \mathbf {p} _ {t} ^ {i} + w _ {t - 1} ^ {j} \mathbf {p} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}}, +$$ + +$$ +\mathbf {n} _ {t} ^ {j} = \frac {w _ {t} ^ {i} \mathbf {n} _ {t} ^ {i} + w _ {t - 1} ^ {j} \mathbf {n} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}}, \tag {2} +$$ + +$$ +r _ {t} ^ {j} = \frac {w _ {t} ^ {i} r _ {t} ^ {i} + w _ {t - 1} ^ {j} r _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}}, +$$ + +$$ +w _ {t} ^ {j} = w _ {t} ^ {i} + w _ {t - 1} ^ {j}, +$$ + +![](images/5e9622e389164ead116c083dea46c41fdcf3a20c957f3b2d1a8ae900f2abb4dd.jpg) +Figure 3. Our efficient rasterization-guided surfel rendering for neural surfel radiance fields is visualized using a toy model as an example. We begin by rasterizing neural surfels into the image plane, and then render the ray using shading points on the intersections between the ray and corresponding surfels. This method uses a combination of rasterization and ray tracing to render view-dependency results with surfel representation and 3D rendering. + +where $\mathbf{p}_t^j$ , $\mathbf{n}_t^j$ , $r_t^j$ and $w_t^j$ are the updated position, normal, radius, and weight of the global surfels, respectively. + +# 3.3. Rendering Neural Surfel Radiance Fields + +We propose a fast rasterization-based rendering algorithm for neural surfel radiance fields (see Figure. 3). Given a neural surfel radiance field $S$ and the camera parameter $\mathbf{c}_{tar}$ of a target view, we first conduct a differentiable rasterization process on the neural surfels to project them on the target image plane. For a pixel $p_x$ in the target image, we gather the neural surfels $\{s_i\}_{i=1}^M$ that cover it and render its color $C(p_x)$ based on a few shading points $\{\mathbf{x}_i\}_{i=1}^M$ that are the intersections between $\{s_i\}_{i=1}^M$ and pixel ray, using the following volume rendering formula [22]: + +$$ +\begin{array}{l} C \left(p _ {x}\right) = \sum_ {i - 1} ^ {M} \tau_ {i} \left(1 - \exp \left(- \sigma_ {i} \Delta_ {i}\right)\right) c _ {i}, \tag {3} \\ \tau_ {i} = \exp (- \sum_ {j = 1} ^ {i - 1} \sigma_ {j} \Delta_ {j}), \\ \end{array} +$$ + +where $\tau$ indicates the accumulated transmittance; $\sigma_{i}$ and $c_{i}$ indicate the volume density and radiance on shading point $\mathbf{x}_i$ ; $\Delta_j$ is the distance between adjacent shading points. For each shading point $x_{i}$ , we regress its volume density $\sigma_{i}$ and radiance $c_{i}$ based on its view direction $\mathbf{d}$ and shading point features $\mathbf{f}^i (\mathbf{x}_i)$ associated from surfel $s_i$ , which can be formulated as: + +$$ +\sigma_ {i}, c _ {i} = \operatorname {R e n d e r} \left(x _ {i}, \mathbf {f} ^ {i} \left(\mathbf {x} _ {i}\right), \mathbf {d}\right), \tag {4} +$$ + +where we use an MLP to parameterize the Render module. The "interpolated" surfel feature $\mathbf{f}^i (\mathbf{x}_i)$ is given as: + +$$ +\mathbf {f} ^ {i} \left(\mathbf {x} _ {i}\right) = \frac {\| \mathbf {x} ^ {i} - \mathbf {p} ^ {i} \|}{r ^ {i}} \mathrm {F} \left(\mathbf {f} ^ {i}, \mathbf {d}, \mathbf {n} ^ {i}, w ^ {i}\right), \tag {5} +$$ + +where $\mathbf{F}$ is also parameterized as an MLP-like network. Note that a recent concurrent work [48] also explores the + +hybrid rendering process by combining rasterization and volume rendering, but using a different neural representation. + +# 3.4. Training and Optimization + +All modules in SurfelNeRF are differentiable, so they can be trained end-to-end on large-scale scene datasets with sequential images as input. We utilize L2 pixel-wise rendering loss to train our network supervised by the target ground truth image: + +$$ +\mathcal {L} _ {\text {r e n d e r}} = \left\| C _ {g t} - C _ {r} \right\| _ {2} ^ {2}, \tag {6} +$$ + +where $C_{gt}$ is the ground truth pixel color and $C_r$ denotes the rendered pixel color output from SurfelNeRF. In addition, if we can acquire sensor depth $D_i^{\mathrm{sen}}$ senor for image $I_i$ , we further employ an L1 loss to train the depth refinement network: + +$$ +\mathcal {L} _ {d} = \left| \left| D _ {i} ^ {\text {s e n}} \odot M _ {i} - D ^ {i} \odot M _ {i} \right| \right| _ {1}, \tag {7} +$$ + +where $D^{i}$ is the output from depth refinement network, and $M_{i}$ is a binary mask indicating valid sensor depths. This supervision can make the depth refinement network to learn to match the sensor depths. + +To sum up, the total loss to train SurfelNeRF end-to-end is given as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e n d e r}} + \lambda \cdot \mathbb {1} \left(\mathcal {L} _ {d}\right), \tag {8} +$$ + +where $\mathbb{1}$ denotes the indicator function whose value is set to 1 if sensor depth is available; $\lambda$ is the balancing weight. + +Training Details. Following [50], we uniformly sample keyframes from each image sequence as input for training SurfelNeRF. To train SurfelNeRF, we sample about $5\%$ of the full sequence as keyframes, and other frames are used as supervision. + +Per-scene Fine-tuning. After being trained on large-scale datasets, our SurfelNeRF is capable of constructing neural surfel radiance fields and rendering photorealistic novel views on unseen scenes in a feedforward manner. To further improve the rendering performance on test scenes, one can also perform a short period of per-scene optimization via fine-tuning the surfel representations and the rendering module. To make a fair comparison, we do not use the additional depths in the fine-tuning process and keep the depth refinement module fixed. Thanks to our efficient rendering scheme, SurfelNeRF takes less time when fine-tuning compared with existing methods [43, 50]. + +# 4. Experiments + +# 4.1. Experimental Settings + +Training Datasets. Our training data consists of large-scale indoor scenes from ScanNet [11] which provides the online sequence input of images with known camera parameters, + +and the captured incomplete sensor depths. We focus on the online input of large-scale scenes in the main paper. Following [37, 50], we randomly sample 100 scenes from ScanNet for training. + +Implementation Details. We train our SurfelNeRF on the large-scale training dataset mentioned above using the Adam [17] optimizer with an initial learning rate of 0.001 for total 200k iterations on 4 NVIDIA Tesla V100 32GB GPUs. For per-scene optimization settings, we fine-tune our method also using the Adam optimizer with an initial learning rate of 0.0002. All input images are resized to $640 \times 480$ . The $\lambda$ in Eq. 8 is set to 0.1. The threshold of $\delta_{depth}$ is set to $0.1\mathrm{m}$ in surfel fusion (Sec. 3.2). The maximum number of candidates M in Eq. 3 for rendering is set to 80. We implement our method using Pytorch and Pytorch3D. + +Metrics. Following previous works [50], we report the PSNR, SSIM [38] and LPIPS $_{vgg}$ [49] to evaluate and compare the results. For fair comparisons with previous methods, we also evaluate our method on the same test set on the ScanNet dataset. The time cost of rendering and perscene fine-tuning is measured on the same platform using one NVIDIA Tesla V100 32GB GPU. + +# 4.2. Comparison with state-of-the-arts + +In this section, we evaluate our method with other state-of-the-arts on the large-scale dataset, ScanNet. To achieve fair comparisons, we evaluate our methods strictly following the same training and evaluation scheme as NeRFusion [50] and NerfingMVS [40]. There are 8 selected testing scenes in ScanNet where 1/8 of images are held out for testing. Following [50], we report no per-scene optimization results via direct inference network and per-scene optimization results. + +Baselines. In addition to comparing SurfelNeRF with NeRF-Fusion [50], we also compare it with other state-of-the-art (SOTA) methods, such as fast per-scene optimization NeRF and neural point cloud representation, generalizable NeRF, scalable MLP-based representation NeRF, NeRF with additional depth supervision, and classical image-based rendering methods, including Instant-NGP [23], ADOP [31], PointNeRF [43], Mip-NeRF-360 [3], Dense Depth Prior NeRF/DDP-NeRF) [28], and IBRNet [37]. Among these methods, Mip-NeRF-360 is one of the state-of-the-art NeRF methods for large-scale scenes. DDP-NeRF employs an offline NeRF with additional depth supervision. PointNeRF achieves state-of-the-art performance thanks to point-based representation. ADOP achieves promising results on offline large-scale rendering. These two methods employ a similar representation to our method; thus we also compare with them. We run generalizable PointNeRF with the same No per-scene optimization and per-scene optimization settings using its open-source code. + +The result is shown in Tab. 2 (a). For no per-scene + +
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
IBRNet [37]21.190.7860.358-
NeRFusion [50]22.990.8380.33538s#
PointNeRF [43]20.470.6420.54430s
SurfelNeRF23.820.8450.3270.2s
+ +(a) Results of No per-scene optimization setting on ScanNet. And we report the average time to render an image. $\sharp$ indicates that the time of NeR-Fusion is reported by authors using a 2080Ti since they have not released executable code before submission. + +
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
Instant-NGP [23]23.230.7140.4590.03
ADOP [31]25.010.8070.2721s
NeRFingMVS [40]26.370.9030.245-
IBRNet [37]25.140.8710.266-
DDP-NeRF [28]26.610.800.26-
Mip-NeRF-360 [3]27.850.860.25-
NeRFusion [50]26.490.9150.20938s
PointNeRF [43]28.990.8290.32430s
SurfelNeRF29.580.9190.2150.2s
+ +(b) Per-scene optimization. Time↓ indicates the average time to render an image. +Table 2. Quantitative comparisons with SOTAs on the ScanNet dataset with No per-scene and per-scene optimization. + +optimization settings, NeRFusion achieves PSNR of 22.99 thanks to the fusion between input features to global grid representations. Our SurfelNeRF outperforms NeRFusion since SurfelNeRF benefits from compact and precise feature fusion and the geometric guidance of surfels. For the per-scene optimization task, the result is shown in Tab. 2 (b). PointNeRF gains a large improvement after fine-tuning, which illustrates it achieves sub-optimal performance in no per-scene optimization settings due to its inability to fuse and update point features. Compared with DDP-NeRF, which employs additional depth supervision, our SurfelNeRF still outperforms it by taking advantage of surfel representation and its integration. Our method achieves state-of-the-art results again, which shows that our neural surfel representation can provide optimal initialization for fine-tuning. + +The qualitative comparison is shown in Figure. 4. We compare our method with PointNeRF without per-scene optimization, and compare with per-scene optimization methods of PoingNeRF and ADOP. Since NeRFusion has not release the executable code, we cannot generate qualitative results. As shown in Figure. 4, our results are visually much better than state-of-the-art methods, PointNeRF and ADOP. + +# 4.3. Ablation Study + +We investigate the effectiveness of designs in our SurfelNeRF by conducting the following ablation studies on + +![](images/eeacfc929f7b6e21c64a68edd2dcc04db4273566b1e45b31fa8aa36614d98bce.jpg) +Figure 4. Qualitative comparisons of rendering quality. The 1-2 column are the results of no per-scene optimization. And the left three columns are per-scene optimization results. It is obvious that with or without per-scene optimization, our method can still achieve the most photorealistic rendering results. Since PointNeRF without a depth refinement network relies on captured depths only, it cannot handle the scenario that depth is incomplete. However, our method can tackle this problem with a lightweight depth refinement network. ADOP can produce geometrically correct results, but its color is less photorealistic. + +# ScanNet. + +Fusion Schemes. Our surfel-based fusion is inspired by classical Surfel-based fusion in conventional 3D reconstruction [41]. The classical surfel-based fusion update features via the simple weighted sum of two overlapped surfels. We conduct an experiment that updates surfel features by weighted sum based on surfel weights, which is a similar operation of updating geometry attributes in Eq. 2. The results are shown in Tab. 3. Since the GRU module can learn to adaptively update features based on the global surfels and new incremental surfels, it can improve performance with negligible computation overhead. + +Depth Refinement Module. To investigate the effective- + +
Fusion SchemeSettingPSNR↑SSIM↑LPIPS ↓
Weighted SumNo per-scene23.090.8330.353
GRUoptimization23.820.8450.327
Weighted SumPer-scene28.540.8840.293
GRUoptimization29.580.9190.215
+ +Table 3. Ablation studies about fusion schemes in our SurfelNeRF. + +ness of our depth refinement, we conduct experiments that reconstruct scenes with incomplete depth captured by sensors. The quantitative results are shown in Tab. 4, which show that PSNR performance drops about $4\%$ without the + +
Depth RefinementPSNR↑SSIM↑LPIPS ↓
w/o19.510.7130.454
w23.090.8330.353
+ +Table 4. Ablation studies about the effectiveness of depth refinement in scene0000_01 in the ScanNet with the no per-scene optimization setting. The depth captured by the sensor in this scene is incomplete. If using captured depths only, some pixels cannot be projected to the 3D space, as the visualization results of PointNeRF shown in the first row in the qualitative results in Fig. 4. + +
Rendering SchemeSettingPSNR↑Time ↓
Volume RenderingNo per-scene23.8030s
Oursoptimization23.820.2s
Volume RenderingPer-scene29.4530s
Oursoptimization29.580.2s
+ +Table 5. Comparing with two rendering schemes, volume rendering [22] and our proposed rasterization-guided rendering, in the ScanNet dataset. + +depth refinement. Depth quality affects surfel reconstruction geometry quality, but captured depth is always incomplete (visualized in the first row in Fig. 4). Thus, a depth refinement network is necessarily required to refine the input depth for better surfel geometry quality. The experimental result shows that the depth refinement network is effective in alleviating the noisy or incomplete captured or estimated depth. Moreover, we also conduct an experiment with worse depth quality that resizes input depth resolution to half size. Our performance just slightly drops $0.5\%$ PSNR. These experiments show the effectiveness of our depth refinement module. + +**Rendering Schemes.** Our proposed renderer renders a ray with a few shading points, which are the intersection points of ray and surfels, unlike volume rendering in NeRF [22] that requires hundreds of shading points each with one MLP evaluation. To validate the effectiveness of our differentiable rasterization scheme for rendering, we conduct an experiment that replaced our rendering scheme with NeRFs' volume rendering. The result is shown in Tab. 5. Since online integration of surfels can limit the number of intersecting surfels, our rendering speed is persistently faster than volume rendering with fewer shading points. Specifically, The average and maximal number of rasterized surfels per pixel at novel views on ScanNet are 2.92 and 10, respectively. + +Computational Cost and Memory Usage. To evaluate the computational cost and memory usage on a large-scale scene, we plot components of GPU processing time, surfel number, and memory usage of our method on a large three rooms scene measured on a V100 GPU implemented by Pytorch3D. During inference, our method renders $640 \times 480$ + +![](images/32897e5ca832b5c88f757b61d307af1a2c813fcd90c1c5bae0801efe25beb265.jpg) +(a) GPU processing time. +Figure 5. The computation cost, memory footprint, and the number of surfels grow according to input frames measured on a three rooms scene. + +![](images/f0b92ee29a1294a28051cec34bc7d4885cee3bb8b248ef2c2ee223733347edcf.jpg) +(b) GPU memory usage. + +images at 5.2FPS on average. As shown in Fig. 5a and Fig. 5b, the number of surfels has a sub-linear growth thanks to the proposed surfel integration. The fluctuation of Time (Rasterization) is caused by the different numbers of corresponding surfels at novel views. Although the computational cost of rasterization depends on the number of rasterized surfels, but it is still faster than volume rendering [22]. These results show that our surfel representation is compact and efficient. The major computation cost comes from rasterization, and the cost of surfel construction as well as integration could be neglected. + +# 5. Limitations + +Our method requires high-quality surfel geometry as neural surfel representation since SurfelNeRF relies on it to conduct surfel integration and rasterization-guided render. To obtain better geometry quality, we may require more accurate depth quality to construct surfel geometry, which necessitates extra depth supervision or the MVS depth estimator. Since captured or estimated depth quality is not exact, there may be geometry discontinuities or missing parts in the surfel representation. Consequently, there may be some missing parts or incorrect geometry in the rendered output (examples of missing parts are shown in the first row in Fig. 4 or video). As discussed in ablation studies of depth refinement above, inaccurate and noisy estimated depth as well as depth discontinuity could affect the construction of neural surfel representation. To alleviate this issue, we propose a depth refinement module (Tab. 4) to refine depth information. + +# 6. Conclusion + +In this paper, we propose SurfelNeRF for online photorealistic reconstruction and rendering of large-scale indoor scenes. We employ neural surfels to store a neural radiance field, adopt an efficient neural radiance field fusion method to progressively build the scene representation, and propose a fast and differentiable rasterization-based rendering algorithm to tackle the challenging task. Experiments on the ScanNet dataset demonstrate that our method achieving favorably better results than the state-of-the-art methods. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. 2020. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 1, 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 6 +[4] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 3 +[5] Ang Cao, Chris Rockwell, and Justin Johnson. Fwd: Real-time novel view synthesis with forward warping and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15713-15724, 2022. 3 +[6] Yan-Pei Cao, Leif Kobbelt, and Shi-Min Hu. Real-time high-accuracy three-dimensional reconstruction with consumer RGB-D cameras. ACM Transactions on Graphics, 37(5):171:1-171:16, 2018. 1, 2, 3 +[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1 +[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 3 +[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.2 +[10] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6124-6132, 2022. 2, 3, 4 +[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE CVPR, pages 2432-2443, 2017. 2, 5 +[12] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(3):24:1-24:18, 2017. 1 +[13] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings + +of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2 +[14] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 1 +[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2, 3 +[16] Jargon Kim, Hyomin Kim, Hyeonseo Nam, Jaesik Park, and Seungyong Lee. Textureme: High-quality textured scene reconstruction in real time. ACM Trans. Graph., 41(3), mar 2022. 1 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[18] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2 +[19] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. 1 +[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2 +[21] Marko Mihajlovic, Silvan Weder, Marc Pollefeys, and Martin R Oswald. Deepsurfels: Learning online appearance fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14524-14535, 2021. 2, 3 +[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 3, 5, 8 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 6 +[24] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, pages 127-136, 2011. 1 +[25] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 2 +[26] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, June 2022. 2 +[27] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. arXiv preprint arXiv:2111.14643, 2021. 1 +[28] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 6 +[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3 +[30] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. arXiv preprint arXiv:2110.06635, 2021. 2 +[31] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 6 +[32] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022. 3 +[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021. 1 +[34] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In IEEE CVPR, pages 15598-15607, 2021. 2, 3, 4 +[35] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. arXiv preprint arXiv:2202.05263, 2022.1 +[36] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2 +[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3, 6 +[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[39] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth + +map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 1 +[40] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 6 +[41] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2, 3, 4, 7 +[42] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 1, 2 +[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 3, 5, 6 +[44] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. European Conference on Computer Vision (ECCV), 2018. 3 +[45] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2 +[46] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3 +[47] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2 +[48] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 5 +[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[50] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 2, 3, 4, 5, 6 +[51] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2 + +[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In IEEE CVPR, pages 12786-12796, 2022. 3 +[53] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 3 \ No newline at end of file diff --git a/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/images.zip b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8fc79d095acea50b3b1f4c08637efc57e54a5676 --- /dev/null +++ b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7261a75f857f61d68010dbd42062ddbea4b2a798b1674053833383f4663b5fab +size 550373 diff --git a/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/layout.json b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a4ca54e8e92478250b27c3dbfa0a5e4b6e42d0b0 --- /dev/null +++ b/2023/SurfelNeRF_ Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes/layout.json @@ -0,0 +1,7966 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 102, + 509, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 102, + 509, + 138 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 509, + 138 + ], + "type": "text", + "content": "SurfelNeRF: Neural Surfel Radiance Fields for Online Photorealistic Reconstruction of Indoor Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 161, + 410, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 161, + 410, + 175 + ], + "spans": [ + { + "bbox": [ + 184, + 161, + 410, + 175 + ], + "type": "text", + "content": "Yiming Gao Yan-Pei Cao Ying Shan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 237, + 176, + 356, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 176, + 356, + 189 + ], + "spans": [ + { + "bbox": [ + 237, + 176, + 356, + 189 + ], + "type": "text", + "content": "ARC Lab, Tencent PCG" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 192, + 487, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 192, + 487, + 203 + ], + "spans": [ + { + "bbox": [ + 103, + 192, + 487, + 203 + ], + "type": "text", + "content": "gaoym9@mail3.sysu.edu.cn caoyanpei@gmail.com yingsshan@tencent.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "content": "Online reconstructing and rendering of large-scale indoor scenes is a long-standing challenge. SLAM-based methods can reconstruct 3D scene geometry progressively in real time but can not render photorealistic results. While NeRF-based methods produce promising novel view synthesis results, their long offline optimization time and lack of geometric constraints pose challenges to efficiently handling online input. Inspired by the complementary advantages of classical 3D reconstruction and NeRF, we thus investigate marrying explicit geometric representation with NeRF rendering to achieve efficient online reconstruction and high-quality rendering. We introduce SurfelNeRF, a variant of neural radiance field which employs a flexible and scalable neural surfel representation to store geometric attributes and extracted appearance features from input images. We further extend the conventional surfel-based fusion scheme to progressively integrate incoming input frames into the reconstructed global neural scene representation. In addition, we propose a highly-efficient differentiable rasterization scheme for rendering neural surfel radiance fields, which helps SurfelNeRF achieve " + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "text", + "content": " speedups in training and inference time, respectively. Experimental results show that our method achieves the state-of-the-art 23.82 PSNR and 29.58 PSNR on ScanNet in feedforward inference and per-scene optimization settings, respectively." + }, + { + "bbox": [ + 46, + 256, + 290, + 558 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 579, + 128, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 592 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 600, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 685 + ], + "type": "text", + "content": "Large-scale scene reconstruction and rendering is a crucial but challenging task in computer vision and graphics with many applications. Classical visual simultaneous localization and mapping (SLAM) systems [6, 12, 16, 24, 39, 41] can perform real-time 3D scene reconstruction. However, they usually represent the scene geometry as solid surfaces and appearance as vertex color or texture maps; thus," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 232, + 547, + 375 + ], + "blocks": [ + { + "bbox": [ + 310, + 232, + 547, + 375 + ], + "lines": [ + { + "bbox": [ + 310, + 232, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 310, + 232, + 547, + 375 + ], + "type": "image", + "image_path": "3c1222365b0a2fe61ffe68b3ebee31b0cf87c99fb6fd47f9946ae667013c53d4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 386, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 304, + 386, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 547, + 453 + ], + "type": "text", + "content": "Figure 1. Examples to illustrate the task of online photorealistic reconstruction of an indoor scene. The online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 473, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 605 + ], + "type": "text", + "content": "the reconstructed results fail to fully capture the scene content and cannot be used for photorealistic rendering. Recently, neural radiance fields (NeRF) and its variants [2, 3, 22,27,35,42] have achieved unprecedented novel view synthesis quality on both object-centric and large-scale scenes. However, NeRFs suffer from long per-scene optimization time and slow rendering speed, especially for large-scale scenes. Although recent advances [7, 14, 19, 23, 33] achieve faster optimization and rendering via incorporating explicit representations, they still require gathering all input images in an offline fashion before optimizing each scene." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 606, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 548, + 715 + ], + "type": "text", + "content": "In this paper, we target the challenging task of online photorealistic reconstruction of large-scale indoor scenes: given an online input image stream of a previously unseen scene, the goal is to progressively build and update a scene representation that allows for high-quality rendering from novel views. The online setting can unlock a variety of real-time interactive applications, providing crucial immediate feedback to users during 3D capture. However, this task brings multiple extra requirements, including the scalability" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "1Project website: https://gymat.github.io/SurfelNeRF-web" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 496, + 156 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 496, + 156 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 496, + 156 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 496, + 156 + ], + "type": "table", + "html": "
MethodsRepresentationGeneralizationReal-time RenderingOnline FusionScalability
DeepSurfel [21]SurfelsX✓XX
Instant-NGP [23]Hash GridsX✓XX
PointNeRF [43]Point Clouds✓XX✓
VBA [10]B+ TreesXX✓✓
NeRFusion [50]Voxel Grids✓X✓X
OursSurfels✓✓✓✓
", + "image_path": "668b0c804464494a7c6f0c16257db4ccc6ccc9f509a648dcad3a535c035cad51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 160, + 163, + 432, + 175 + ], + "lines": [ + { + "bbox": [ + 160, + 163, + 432, + 175 + ], + "spans": [ + { + "bbox": [ + 160, + 163, + 432, + 175 + ], + "type": "text", + "content": "Table 1. Comparison of representation and features with existing methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 185, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 289, + 399 + ], + "type": "text", + "content": "ity of the underlying scene representation, the ability to perform on-the-fly updates to the scene representation, and optimizing and rendering at interactive framereates. Recently, NeRFusion [50] followed NeuralRecon [34] to unproject input images into local sparse feature volumes, fusing them to a global volume via Gated Recurrent Units (GRUs), and then generating photorealistic results from the global feature volume via volume rendering. However, updating the sparse volumetric feature involves computationally heavy operations; the volume rendering is also very slow since it requires hundreds of MLP evaluations to render a pixel. Thus, although NeRFusion achieves efficient online scene reconstruction, it still needs dozens of seconds to render a frame. VBA [10] is another recent approach to online photorealistic scene capture, but it only applies to object-centric scenes. We compare with representation and key features used in online photorealistic rendering with existing methods, which is shown in Tab. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": "We propose surfel-based neural radiance fields, SurfelNeRF, for online photorealistic reconstruction and rendering of large-scale indoor scenes. Surfels (surface elements) [25] are point primitives containing geometric attributes, e.g., position, color, normal, and radius. We extend this representation to neural surfels, storing extra neural features that encode the neural radiance field of the target scene. Compared with volumetric representations, neural surfels are more compact and flexible and can easily scale to large scenes. Besides, we further employ a fast and differentiable rasterization process to render neural surfel radiance fields, which produces a pixel with only a few MLP evaluations based on the rasterized neural surfels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 700 + ], + "type": "text", + "content": "Inspired by classical real-time surfel-based geometric reconstruction methods [6, 15, 41], we propose an efficient neural radiance field fusion method to progressively build the scene representation by integrating neighboring neural surfels. Unlike point-based representations [1, 26, 30, 43] that are computationally heavy when finding neighboring points, it is easier to locate overlapping surfels and then merge neural features from multiview observations. By coupling the SurfelNeRF representation, the efficient neural surfel fusion approach, and the fast neural surfel rasterization algorithm, we achieve high-quality, photorealistic 3D scene reconstruction in an online manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We conduct experiments on the large-scale indoor scene" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 185, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 329 + ], + "type": "text", + "content": "dataset ScanNet [11], which contains complex scene structures and a large variety of scene appearances. We train the SurfelNeRF end-to-end across the scenes on the ScanNet, obtaining a generalizable model that enables both feedforward inference on unseen data and per-scene fine-tuning. We demonstrate in experiments that the proposed SurfelNeRF achieves favorably better rendering quality than the state-of-the-art approaches in both feedforward and finetuning settings while maintaining high training and rendering efficiency. We believe the proposed online photorealistic reconstruction framework has great potential in practical applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 343, + 392, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 392, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 392, + 357 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 639 + ], + "type": "text", + "content": "Novel view synthesis. Novel view synthesis is a challenging task for computer vision and computer graphics. Previous methods utilize geometry-based view synthesis. Zhou et al. [51] proposes to use the multiplane image (MPI) representation for novel view synthesis. Recently, neural radiance fields [22] and neural-based geometry representation have attracted more interest due to high-quality rendering. Recent methods propose to associate neural features on geometry structure and then synthesize novel views via rendering rasterized features, such as point cloud [1, 30], texture [36] and surfel [21]. DeepSurfel performs online appearance fusion only on a fixed given geometry during online fusion, while our SurfelNeRF performs online appearance and geometry integration. To render outputs, DeepSurfel rasterizes surfels and then passes them through 2D CNN filters, while our SurfelNeRF performs neural volume rendering which could achieve higher-quality rendering. Some recent works [9, 13, 18, 45] focus on real-time neural rendering methods. However, these methods all require per-scene optimization, which makes them impractical to handle novel scenes. In contrast, our SurfelNeRF is generalizable, which can reconstruct new scenes via direct feedforward inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Neural Radiance Fields. Recently, neural radiance fields (NeRF) methods [2, 22, 42] achieve promising results on novel view synthesis, but it requires expensive per-scene optimization. Some methods proposed to reduce the long per-scene optimization time via deferred neural rendering [13, 47], hash table [23] or geometry prior [20]. Recent," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "109" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "generalization NeRF methods [5,8,37,43,46] are proposed to produce novel view without per-scene optimization by learning a radiance fields representation from offline given images of novel scenes. PixelNeRF [46] and IBRNet [37] render novel view images using volume rendering based on corresponding warped features from nearby given images. MVSNeRF [8] reconstructs local neural volumes from local nearby views, and then conducts volume rendering based on volumes. FWD [5] renders novel views by warping all input images to the target view, and fusing and refining features of these warped inputs. However, these methods are generalizable on local small-scale scenes, and require taking all input images before rendering novel views in an offline fashion. PointNeRF [43] learns to reconstruct a point-based neural radiance field for the novel scene. However, point-based representation is redundant, especially on large-scale scenes, and the point clouds cannot be updated based on multi-view input. Unlike these works, our method can represent not only large-scale scenes but also update compact surfel representation in an online fashion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 312, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 289, + 563 + ], + "type": "text", + "content": "Online scene reconstruction and rendering. Most online scene reconstruction methods [4, 34, 41, 52, 53] focus on 3D reconstruction only via TSDF-based fusion [4, 34], surfel-based fusion [41] and SLAM-based reconstruction [52]. However, they cannot render photorealistic novel view synthesis. Recently, some methods [10, 32, 50] employ radiance fields to tackle online scene representation for high-quality rendering. VBA [10] builds a neural dynamic B+Tree to represent scenes and apply volume rendering on it, but it still requires per-scene optimization. SRT [32] represents scenes with set-latent with a transformer encoder and decoder, but it cannot synthesize high-resolution images, which makes it not ready for real-world applications. NeRFusion [50] represents large-scale scenes with neural voxel grids, update them via GRU in an online fashion, and renders results via slow volume rendering. However, grid-based representation is wasteful for large-scale scenes, and is inefficient during features update due to features in empty areas are also updated. In contrast, our surfel representation is compact for large-scale scenes, and surfel-based fusion and corresponding rendering are more efficient." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 574, + 103, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 103, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 103, + 586 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "content": "Unlike most existing methods [22, 23, 43] that build neural radiance fields in an offline fashion, we aim for the task of online photorealistic scene reconstruction. Given a sequence of images " + }, + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\{I_n\\}" + }, + { + "bbox": [ + 47, + 594, + 287, + 665 + ], + "type": "text", + "content": ", the goal is to progressively build the neural scene representation on the fly as the input images arrive." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": "To achieve this goal, we present SurfelNeRF, an online neural reconstruction framework based on the neural surfel radiance fields representation. The insight of our method is that the classical explicit surfel representation is efficient" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "for both reconstruction and rendering, has compact memory footprints, has the scalability to large and complex scenes, and can be naturally extended to equip neural attributes. Meanwhile, one can integrate multi-frame information and perform incremental updates on neural surfels in a geometrically meaningful way, and the operations involved can be streamlined." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": "Figure 2 gives an overview of our proposed method. The framework first learns to generate per-view local neural surfel radiance fields " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": " for each input frame " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": ". We then adopt the conventional surfel fusion process [6,15,41] to the neural setting, parameterizing the neural surfel fusion procedural as a neural network " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": ". Through " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": ", per-view representations " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "\\{S_t\\}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": " can be integrated into a global neural surfel radiance field " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "S_{g}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": " on the fly, aggregating geometric and neural attributes. The neural surfel radiance fields can be efficiently rendered to photorealistic images using a differentiable rasterizer without extra optimization. The framework can be trained end-to-end on large-scale indoor scene datasets. Once trained, it can generalize to previously unseen data and perform online photorealistic reconstruction; if needed, the rendering quality can be further improved by a short period time of fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 357, + 529, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 529, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 529, + 369 + ], + "type": "text", + "content": "3.1. Neural Surfel Radiance Field Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": "Surfel Representation. Adopted from classical 3D reconstruction research, each surfel " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "s^n" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " is equipped with the following geometric attributes: a position " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^n \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": ", normal " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^n \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": ", weight " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "w^n \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " and radius " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "r^n \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": ". The weight " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "w^n" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " measures the confidence of a surfel based on the distance of the current depth measurement from the camera center. The radius " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "r^n" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " of each surfel is designed to represent the local support area around the projected 3D position while minimizing visible holes. The weight and radius of each surfel are initialized following [41]. Besides geometric properties, we further associate each surfel with neural features " + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^n \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 377, + 545, + 579 + ], + "type": "text", + "content": " from input images for modeling semi-transparent structures and view-dependent effects. Although a similar representation has been explored in DeepSurfels [21], we construct neural surfels in a feedforward manner and incorporate them in progressive scene reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "content": "Local Neural Surfel Radiance Field. Given an input frame " + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "content": ", camera parameters " + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t" + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "content": ", and estimated depth " + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "content": ", the corresponding local neural surfel radiance field " + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_t" + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "content": " can be constructed by unprojecting pixels in the input image to 3D space using camera parameters, then estimating corresponding normal, weight, radius, and neural features. The depth " + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 304, + 581, + 545, + 700 + ], + "type": "text", + "content": " can be captured from sensors or estimated via off-the-shelf MVS-based methods [44], which might be incomplete and noisy. Thus, we employ a U-Net structure [29] to refine the depth following [5]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": "Neural Features on Surfels. We estimate neural features" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "110" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 521, + 242 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 521, + 242 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 521, + 242 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 521, + 242 + ], + "type": "image", + "image_path": "2b7b4e20fde4ff81e9924c8dcbac88033f882241b5f9eef8d8e50279d428fa70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 545, + 327 + ], + "type": "text", + "content": "Figure 2. Overview of our SurfelNeRF. Given an online input stream of image sequences, we first reconstruct a surfel representation associated with neural features to build a local neural surfel radiance field for each input keyframe. Then the neural surfel radiance field integration is used to fuse the input local neural surfel radiance field into the global neural surfel radiance field by updating both surfel position and features. More specifically, input local neural surfels associated with global surfels are fused to global surfels with corresponding global ones, and remaining local surfels without corresponding ones are added to the global model. Furthermore, the novel views can be rendered from updated global surfels via our efficient rasterization-guided render. Our proposed rasterization-guided renderer renders color only on the intersection points of ray and surfels, which is faster than volume rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 348, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^n" + }, + { + "bbox": [ + 46, + 348, + 287, + 397 + ], + "type": "text", + "content": " of surfels by extracting corresponding pixel features in the input image. We employ a 2D CNN to extract features of input images. Specifically, we use a MasNet following [34, 50] to extract multi-scale view-dependent features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 406, + 240, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 240, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 240, + 418 + ], + "type": "text", + "content": "3.2. Online Integration of Neural Surfels" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 532 + ], + "type": "text", + "content": "Based on the neural surfel representation, we adopt a geometry-guided fusion scheme to efficiently and progressively integrate local neural surfel radiance fields into a global model. Compared with previous volumetric fusion methods [10,50], our surfel-based solution offers the advantages of 1) efficiency - the fusion operation only needs to be performed at surfel locations and 2) scalability - the memory footprint only grows linearly to the number of neural surfels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "In particular, our fusion scheme can be divided into two steps. Firstly, for each local neural surfel, we determine whether it should be added as a global neural surfel or should be merged with corresponding global surfels. Secondly, we insert it into the global model or perform updates to corresponding global surfels with local features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Neural Surfel Association. Given an input image frame " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and corresponding camera parameter, each surfel in global surfels can be projected as an ellipse onto the image plane of the current input camera view. For each pixel in the input image, we obtain top- " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " global surfels that are overlapped with it after rasterization. We discard irrelevant surfels with a normal filter. Then we determine the incremental surfel by comparing the depth distance between respective top-M global surfels from near to far step by step. If the min" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "content": "imal distance between a local surfel and its global surfel candidates is smaller than a threshold " + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\delta_{depth}" + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "content": ", then we determine that this input surfel needs to be merged; otherwise, it should be directly added to global surfels. The overall workflow follows [41], and " + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 348, + 545, + 408 + ], + "type": "text", + "content": " is set to 8." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "type": "text", + "content": "Neural Surfel Fusion. After the surfel association, there are a set of input surfels that need to be added directly and another set of surfels that will be merged with corresponding global surfels. For the former set of surfels, we directly construct neural surfels (Sec. 3.1) and then add them to global surfels. For the latter surfels, we use a GRU network to update neural features between the features " + }, + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\mathrm{merge}}" + }, + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "type": "text", + "content": " of input surfels and features " + }, + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t - 1}^{\\mathrm{corr}}" + }, + { + "bbox": [ + 304, + 408, + 546, + 516 + ], + "type": "text", + "content": " of corresponding global surfels, which is given as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 366, + 520, + 545, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 520, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 366, + 520, + 545, + 534 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t} ^ {\\text {c o r r s}} = \\operatorname {G R U} \\left(\\mathbf {f} _ {t} ^ {\\text {m e r g e}}, \\mathbf {f} _ {t - 1} ^ {\\text {c o r r s}}\\right), \\tag {1}", + "image_path": "2e34a8a0067f07c8c064186596e820703abd514e27688d54ea6f5c3f03b9add8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t^{\\mathrm{merge}}" + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t - 1}^{\\mathrm{corrs}}" + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "text", + "content": " are input and hidden state of GRU respectively. We then update the geometric attributes of these global surfels via a weighted sum. The update rule of each respective pair of local surfel " + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "inline_equation", + "content": "s_t^j" + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "text", + "content": " and global surfel " + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "inline_equation", + "content": "s_{t - 1}^{j}" + }, + { + "bbox": [ + 304, + 537, + 545, + 598 + ], + "type": "text", + "content": " can be expressed by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 603, + 479, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 603, + 479, + 634 + ], + "spans": [ + { + "bbox": [ + 372, + 603, + 479, + 634 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {t} ^ {j} = \\frac {w _ {t} ^ {i} \\mathbf {p} _ {t} ^ {i} + w _ {t - 1} ^ {j} \\mathbf {p} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}},", + "image_path": "fd62510fe59fd3468ec881698330990298616387d2b6f59e1f55f6a988913b42.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 636, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 636, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 373, + 636, + 545, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {n} _ {t} ^ {j} = \\frac {w _ {t} ^ {i} \\mathbf {n} _ {t} ^ {i} + w _ {t - 1} ^ {j} \\mathbf {n} _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}}, \\tag {2}", + "image_path": "6361b4ab382940e8c43d8fd7ff456bf7f8fa7dd79181ea4ab2108916f4559263.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 376, + 669, + 475, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 669, + 475, + 699 + ], + "spans": [ + { + "bbox": [ + 376, + 669, + 475, + 699 + ], + "type": "interline_equation", + "content": "r _ {t} ^ {j} = \\frac {w _ {t} ^ {i} r _ {t} ^ {i} + w _ {t - 1} ^ {j} r _ {t - 1} ^ {j}}{w _ {t} ^ {i} + w _ {t - 1} ^ {j}},", + "image_path": "7258a51d4fccc00246c7ff2a01e1248b6225db5fece1cc0c81ce7e2776ca1fd4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 701, + 446, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 701, + 446, + 715 + ], + "spans": [ + { + "bbox": [ + 373, + 701, + 446, + 715 + ], + "type": "interline_equation", + "content": "w _ {t} ^ {j} = w _ {t} ^ {i} + w _ {t - 1} ^ {j},", + "image_path": "98275473fe5a2c77e24ea4d023f7f6cee3a45e1df6469b1e0d45ceaf70e7e2f9.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "111" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 167 + ], + "type": "image", + "image_path": "5e9622e389164ead116c083dea46c41fdcf3a20c957f3b2d1a8ae900f2abb4dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 289, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 253 + ], + "type": "text", + "content": "Figure 3. Our efficient rasterization-guided surfel rendering for neural surfel radiance fields is visualized using a toy model as an example. We begin by rasterizing neural surfels into the image plane, and then render the ray using shading points on the intersections between the ray and corresponding surfels. This method uses a combination of rasterization and ray tracing to render view-dependency results with surfel representation and 3D rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t^j" + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_t^j" + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "inline_equation", + "content": "r_t^j" + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "inline_equation", + "content": "w_t^j" + }, + { + "bbox": [ + 47, + 270, + 287, + 297 + ], + "type": "text", + "content": " are the updated position, normal, radius, and weight of the global surfels, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 302, + 263, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 263, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 263, + 316 + ], + "type": "text", + "content": "3.3. Rendering Neural Surfel Radiance Fields" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": "We propose a fast rasterization-based rendering algorithm for neural surfel radiance fields (see Figure. 3). Given a neural surfel radiance field " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " and the camera parameter " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{tar}" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " of a target view, we first conduct a differentiable rasterization process on the neural surfels to project them on the target image plane. For a pixel " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "p_x" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " in the target image, we gather the neural surfels " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " that cover it and render its color " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "C(p_x)" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " based on a few shading points " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " that are the intersections between " + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 321, + 289, + 441 + ], + "type": "text", + "content": " and pixel ray, using the following volume rendering formula [22]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 448, + 287, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 448, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 90, + 448, + 287, + 509 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} C \\left(p _ {x}\\right) = \\sum_ {i - 1} ^ {M} \\tau_ {i} \\left(1 - \\exp \\left(- \\sigma_ {i} \\Delta_ {i}\\right)\\right) c _ {i}, \\tag {3} \\\\ \\tau_ {i} = \\exp (- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\Delta_ {j}), \\\\ \\end{array}", + "image_path": "faa792cba855e6e9cac726ee658571b9cd0ca814f926a85788d1dd9ea63a2eba.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " indicates the accumulated transmittance; " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " indicate the volume density and radiance on shading point " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\Delta_j" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " is the distance between adjacent shading points. For each shading point " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": ", we regress its volume density " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " and radiance " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " based on its view direction " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " and shading point features " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^i (\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": " associated from surfel " + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 516, + 287, + 598 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 607, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 607, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 102, + 607, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\sigma_ {i}, c _ {i} = \\operatorname {R e n d e r} \\left(x _ {i}, \\mathbf {f} ^ {i} \\left(\\mathbf {x} _ {i}\\right), \\mathbf {d}\\right), \\tag {4}", + "image_path": "80cd459cda47e7e80551ef8982e2999c4974e4aa4c3a1aa02b26d7948c82ec2d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "content": "where we use an MLP to parameterize the Render module. The \"interpolated\" surfel feature " + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^i (\\mathbf{x}_i)" + }, + { + "bbox": [ + 47, + 628, + 286, + 652 + ], + "type": "text", + "content": " is given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 659, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 659, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 91, + 659, + 287, + 684 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {i} \\left(\\mathbf {x} _ {i}\\right) = \\frac {\\| \\mathbf {x} ^ {i} - \\mathbf {p} ^ {i} \\|}{r ^ {i}} \\mathrm {F} \\left(\\mathbf {f} ^ {i}, \\mathbf {d}, \\mathbf {n} ^ {i}, w ^ {i}\\right), \\tag {5}", + "image_path": "81a0ef4c1aff5dd7f5a42d451734b2eb385f5a6945478a63cfb608afdfedca79.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is also parameterized as an MLP-like network. Note that a recent concurrent work [48] also explores the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 546, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 546, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 546, + 107 + ], + "type": "text", + "content": "hybrid rendering process by combining rasterization and volume rendering, but using a different neural representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 115, + 455, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 115, + 455, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 455, + 129 + ], + "type": "text", + "content": "3.4. Training and Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 133, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 194 + ], + "type": "text", + "content": "All modules in SurfelNeRF are differentiable, so they can be trained end-to-end on large-scale scene datasets with sequential images as input. We utilize L2 pixel-wise rendering loss to train our network supervised by the target ground truth image:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 201, + 546, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 201, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 373, + 201, + 546, + 216 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n d e r}} = \\left\\| C _ {g t} - C _ {r} \\right\\| _ {2} ^ {2}, \\tag {6}", + "image_path": "8d474248f719065b6b197ee5a098d9a20672297684a20efe4ee818d995d8a2c7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "inline_equation", + "content": "C_{gt}" + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "text", + "content": " is the ground truth pixel color and " + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "inline_equation", + "content": "C_r" + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "text", + "content": " denotes the rendered pixel color output from SurfelNeRF. In addition, if we can acquire sensor depth " + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "inline_equation", + "content": "D_i^{\\mathrm{sen}}" + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "text", + "content": " senor for image " + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 305, + 223, + 546, + 281 + ], + "type": "text", + "content": ", we further employ an L1 loss to train the depth refinement network:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 352, + 291, + 545, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 291, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 352, + 291, + 545, + 305 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\left| \\left| D _ {i} ^ {\\text {s e n}} \\odot M _ {i} - D ^ {i} \\odot M _ {i} \\right| \\right| _ {1}, \\tag {7}", + "image_path": "6a49e1b39bffbb52fd14b3c18c7e993dad5dc3632431be66fed9ab3ae34b984c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "D^{i}" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": " is the output from depth refinement network, and " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": " is a binary mask indicating valid sensor depths. This supervision can make the depth refinement network to learn to match the sensor depths." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 361, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 547, + 384 + ], + "type": "text", + "content": "To sum up, the total loss to train SurfelNeRF end-to-end is given as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 372, + 384, + 545, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 384, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 372, + 384, + 545, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda \\cdot \\mathbb {1} \\left(\\mathcal {L} _ {d}\\right), \\tag {8}", + "image_path": "f4a6d0936a57fbcbfadda46134aeb75637dc3d01894840a782d58aa3f59dcd78.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "type": "text", + "content": " denotes the indicator function whose value is set to 1 if sensor depth is available; " + }, + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 402, + 545, + 426 + ], + "type": "text", + "content": " is the balancing weight." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 426, + 547, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 485 + ], + "type": "text", + "content": "Training Details. Following [50], we uniformly sample keyframes from each image sequence as input for training SurfelNeRF. To train SurfelNeRF, we sample about " + }, + { + "bbox": [ + 305, + 426, + 547, + 485 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 305, + 426, + 547, + 485 + ], + "type": "text", + "content": " of the full sequence as keyframes, and other frames are used as supervision." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 486, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 629 + ], + "type": "text", + "content": "Per-scene Fine-tuning. After being trained on large-scale datasets, our SurfelNeRF is capable of constructing neural surfel radiance fields and rendering photorealistic novel views on unseen scenes in a feedforward manner. To further improve the rendering performance on test scenes, one can also perform a short period of per-scene optimization via fine-tuning the surfel representations and the rendering module. To make a fair comparison, we do not use the additional depths in the fine-tuning process and keep the depth refinement module fixed. Thanks to our efficient rendering scheme, SurfelNeRF takes less time when fine-tuning compared with existing methods [43, 50]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 658, + 433, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 433, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 433, + 672 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 714 + ], + "type": "text", + "content": "Training Datasets. Our training data consists of large-scale indoor scenes from ScanNet [11] which provides the online sequence input of images with known camera parameters," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "112" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "and the captured incomplete sensor depths. We focus on the online input of large-scale scenes in the main paper. Following [37, 50], we randomly sample 100 scenes from ScanNet for training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "content": "Implementation Details. We train our SurfelNeRF on the large-scale training dataset mentioned above using the Adam [17] optimizer with an initial learning rate of 0.001 for total 200k iterations on 4 NVIDIA Tesla V100 32GB GPUs. For per-scene optimization settings, we fine-tune our method also using the Adam optimizer with an initial learning rate of 0.0002. All input images are resized to " + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "content": " in Eq. 8 is set to 0.1. The threshold of " + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "inline_equation", + "content": "\\delta_{depth}" + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "content": " in surfel fusion (Sec. 3.2). The maximum number of candidates M in Eq. 3 for rendering is set to 80. We implement our method using Pytorch and Pytorch3D." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 336 + ], + "type": "text", + "content": "Metrics. Following previous works [50], we report the PSNR, SSIM [38] and LPIPS" + }, + { + "bbox": [ + 46, + 253, + 287, + 336 + ], + "type": "inline_equation", + "content": "_{vgg}" + }, + { + "bbox": [ + 46, + 253, + 287, + 336 + ], + "type": "text", + "content": " [49] to evaluate and compare the results. For fair comparisons with previous methods, we also evaluate our method on the same test set on the ScanNet dataset. The time cost of rendering and perscene fine-tuning is measured on the same platform using one NVIDIA Tesla V100 32GB GPU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 230, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 230, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 230, + 360 + ], + "type": "text", + "content": "4.2. Comparison with state-of-the-arts" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 472 + ], + "type": "text", + "content": "In this section, we evaluate our method with other state-of-the-arts on the large-scale dataset, ScanNet. To achieve fair comparisons, we evaluate our methods strictly following the same training and evaluation scheme as NeRFusion [50] and NerfingMVS [40]. There are 8 selected testing scenes in ScanNet where 1/8 of images are held out for testing. Following [50], we report no per-scene optimization results via direct inference network and per-scene optimization results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 700 + ], + "type": "text", + "content": "Baselines. In addition to comparing SurfelNeRF with NeRF-Fusion [50], we also compare it with other state-of-the-art (SOTA) methods, such as fast per-scene optimization NeRF and neural point cloud representation, generalizable NeRF, scalable MLP-based representation NeRF, NeRF with additional depth supervision, and classical image-based rendering methods, including Instant-NGP [23], ADOP [31], PointNeRF [43], Mip-NeRF-360 [3], Dense Depth Prior NeRF/DDP-NeRF) [28], and IBRNet [37]. Among these methods, Mip-NeRF-360 is one of the state-of-the-art NeRF methods for large-scale scenes. DDP-NeRF employs an offline NeRF with additional depth supervision. PointNeRF achieves state-of-the-art performance thanks to point-based representation. ADOP achieves promising results on offline large-scale rendering. These two methods employ a similar representation to our method; thus we also compare with them. We run generalizable PointNeRF with the same No per-scene optimization and per-scene optimization settings using its open-source code." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "The result is shown in Tab. 2 (a). For no per-scene" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 535, + 143 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 535, + 143 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 535, + 143 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 535, + 143 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
IBRNet [37]21.190.7860.358-
NeRFusion [50]22.990.8380.33538s#
PointNeRF [43]20.470.6420.54430s
SurfelNeRF23.820.8450.3270.2s
", + "image_path": "f952aa366259a02d4825ed6e4cdfdcbbf9a619b8ab649e81de72a78097b5e02e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 144, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 182 + ], + "type": "text", + "content": "(a) Results of No per-scene optimization setting on ScanNet. And we report the average time to render an image. " + }, + { + "bbox": [ + 305, + 144, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 305, + 144, + 545, + 182 + ], + "type": "text", + "content": " indicates that the time of NeR-Fusion is reported by authors using a 2080Ti since they have not released executable code before submission." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 194, + 541, + 327 + ], + "blocks": [ + { + "bbox": [ + 310, + 194, + 541, + 327 + ], + "lines": [ + { + "bbox": [ + 310, + 194, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 310, + 194, + 541, + 327 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑LPIPS ↓Time ↓
Instant-NGP [23]23.230.7140.4590.03
ADOP [31]25.010.8070.2721s
NeRFingMVS [40]26.370.9030.245-
IBRNet [37]25.140.8710.266-
DDP-NeRF [28]26.610.800.26-
Mip-NeRF-360 [3]27.850.860.25-
NeRFusion [50]26.490.9150.20938s
PointNeRF [43]28.990.8290.32430s
SurfelNeRF29.580.9190.2150.2s
", + "image_path": "474f874671318d063e92faa8a14da31ad1499f60cbeeaf468b2fb0dd0256fa5d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 328, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 306, + 328, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 545, + 348 + ], + "type": "text", + "content": "(b) Per-scene optimization. Time↓ indicates the average time to render an image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 305, + 357, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 357, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 545, + 380 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons with SOTAs on the ScanNet dataset with No per-scene and per-scene optimization." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 579 + ], + "type": "text", + "content": "optimization settings, NeRFusion achieves PSNR of 22.99 thanks to the fusion between input features to global grid representations. Our SurfelNeRF outperforms NeRFusion since SurfelNeRF benefits from compact and precise feature fusion and the geometric guidance of surfels. For the per-scene optimization task, the result is shown in Tab. 2 (b). PointNeRF gains a large improvement after fine-tuning, which illustrates it achieves sub-optimal performance in no per-scene optimization settings due to its inability to fuse and update point features. Compared with DDP-NeRF, which employs additional depth supervision, our SurfelNeRF still outperforms it by taking advantage of surfel representation and its integration. Our method achieves state-of-the-art results again, which shows that our neural surfel representation can provide optimal initialization for fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": "The qualitative comparison is shown in Figure. 4. We compare our method with PointNeRF without per-scene optimization, and compare with per-scene optimization methods of PoingNeRF and ADOP. Since NeRFusion has not release the executable code, we cannot generate qualitative results. As shown in Figure. 4, our results are visually much better than state-of-the-art methods, PointNeRF and ADOP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 400, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 400, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 400, + 684 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We investigate the effectiveness of designs in our SurfelNeRF by conducting the following ablation studies on" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "113" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 72, + 512, + 475 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 512, + 475 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 512, + 475 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 512, + 475 + ], + "type": "image", + "image_path": "eeacfc929f7b6e21c64a68edd2dcc04db4273566b1e45b31fa8aa36614d98bce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 486, + 547, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 486, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 547, + 544 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of rendering quality. The 1-2 column are the results of no per-scene optimization. And the left three columns are per-scene optimization results. It is obvious that with or without per-scene optimization, our method can still achieve the most photorealistic rendering results. Since PointNeRF without a depth refinement network relies on captured depths only, it cannot handle the scenario that depth is incomplete. However, our method can tackle this problem with a lightweight depth refinement network. ADOP can produce geometrically correct results, but its color is less photorealistic." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 555, + 86, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 86, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 86, + 565 + ], + "type": "text", + "content": "ScanNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 567, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 288, + 700 + ], + "type": "text", + "content": "Fusion Schemes. Our surfel-based fusion is inspired by classical Surfel-based fusion in conventional 3D reconstruction [41]. The classical surfel-based fusion update features via the simple weighted sum of two overlapped surfels. We conduct an experiment that updates surfel features by weighted sum based on surfel weights, which is a similar operation of updating geometry attributes in Eq. 2. The results are shown in Tab. 3. Since the GRU module can learn to adaptively update features based on the global surfels and new incremental surfels, it can improve performance with negligible computation overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Depth Refinement Module. To investigate the effective-" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 552, + 555, + 626 + ], + "blocks": [ + { + "bbox": [ + 306, + 552, + 555, + 626 + ], + "lines": [ + { + "bbox": [ + 306, + 552, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 555, + 626 + ], + "type": "table", + "html": "
Fusion SchemeSettingPSNR↑SSIM↑LPIPS ↓
Weighted SumNo per-scene23.090.8330.353
GRUoptimization23.820.8450.327
Weighted SumPer-scene28.540.8840.293
GRUoptimization29.580.9190.215
", + "image_path": "0dd2e28d516e4fce62ce15ddc3dfd26c6395441cbefb083066c101f1da85d173.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 634, + 545, + 645 + ], + "lines": [ + { + "bbox": [ + 306, + 634, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 545, + 645 + ], + "type": "text", + "content": "Table 3. Ablation studies about fusion schemes in our SurfelNeRF." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "ness of our depth refinement, we conduct experiments that reconstruct scenes with incomplete depth captured by sensors. The quantitative results are shown in Tab. 4, which show that PSNR performance drops about " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " without the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "text", + "content": "114" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 70, + 267, + 117 + ], + "blocks": [ + { + "bbox": [ + 68, + 70, + 267, + 117 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 267, + 117 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 267, + 117 + ], + "type": "table", + "html": "
Depth RefinementPSNR↑SSIM↑LPIPS ↓
w/o19.510.7130.454
w23.090.8330.353
", + "image_path": "e6a1ad81b652212160046247df99fb73ecbff989a9268b3129ce8458e1ba848c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 202, + 277, + 276 + ], + "blocks": [ + { + "bbox": [ + 46, + 125, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 125, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 287, + 191 + ], + "type": "text", + "content": "Table 4. Ablation studies about the effectiveness of depth refinement in scene0000_01 in the ScanNet with the no per-scene optimization setting. The depth captured by the sensor in this scene is incomplete. If using captured depths only, some pixels cannot be projected to the 3D space, as the visualization results of PointNeRF shown in the first row in the qualitative results in Fig. 4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 202, + 277, + 276 + ], + "lines": [ + { + "bbox": [ + 58, + 202, + 277, + 276 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 277, + 276 + ], + "type": "table", + "html": "
Rendering SchemeSettingPSNR↑Time ↓
Volume RenderingNo per-scene23.8030s
Oursoptimization23.820.2s
Volume RenderingPer-scene29.4530s
Oursoptimization29.580.2s
", + "image_path": "c4c9c89e7c58a961d2e64936b2d193ffb63fb472059a17472578fa0826c7c40c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 284, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 316 + ], + "type": "text", + "content": "Table 5. Comparing with two rendering schemes, volume rendering [22] and our proposed rasterization-guided rendering, in the ScanNet dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 330, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 473 + ], + "type": "text", + "content": "depth refinement. Depth quality affects surfel reconstruction geometry quality, but captured depth is always incomplete (visualized in the first row in Fig. 4). Thus, a depth refinement network is necessarily required to refine the input depth for better surfel geometry quality. The experimental result shows that the depth refinement network is effective in alleviating the noisy or incomplete captured or estimated depth. Moreover, we also conduct an experiment with worse depth quality that resizes input depth resolution to half size. Our performance just slightly drops " + }, + { + "bbox": [ + 46, + 330, + 287, + 473 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 330, + 287, + 473 + ], + "type": "text", + "content": " PSNR. These experiments show the effectiveness of our depth refinement module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 640 + ], + "type": "text", + "content": "**Rendering Schemes.** Our proposed renderer renders a ray with a few shading points, which are the intersection points of ray and surfels, unlike volume rendering in NeRF [22] that requires hundreds of shading points each with one MLP evaluation. To validate the effectiveness of our differentiable rasterization scheme for rendering, we conduct an experiment that replaced our rendering scheme with NeRFs' volume rendering. The result is shown in Tab. 5. Since online integration of surfels can limit the number of intersecting surfels, our rendering speed is persistently faster than volume rendering with fewer shading points. Specifically, The average and maximal number of rasterized surfels per pixel at novel views on ScanNet are 2.92 and 10, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Computational Cost and Memory Usage. To evaluate the computational cost and memory usage on a large-scale scene, we plot components of GPU processing time, surfel number, and memory usage of our method on a large three rooms scene measured on a V100 GPU implemented by Pytorch3D. During inference, our method renders " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "640 \\times 480" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 78, + 422, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 78, + 422, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 78, + 422, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 78, + 422, + 148 + ], + "type": "image", + "image_path": "32897e5ca832b5c88f757b61d307af1a2c813fcd90c1c5bae0801efe25beb265.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 148, + 405, + 159 + ], + "lines": [ + { + "bbox": [ + 323, + 148, + 405, + 159 + ], + "spans": [ + { + "bbox": [ + 323, + 148, + 405, + 159 + ], + "type": "text", + "content": "(a) GPU processing time." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 167, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 200 + ], + "type": "text", + "content": "Figure 5. The computation cost, memory footprint, and the number of surfels grow according to input frames measured on a three rooms scene." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 429, + 78, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 429, + 78, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 429, + 78, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 429, + 78, + 547, + 148 + ], + "type": "image", + "image_path": "f0b92ee29a1294a28051cec34bc7d4885cee3bb8b248ef2c2ee223733347edcf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 148, + 527, + 159 + ], + "lines": [ + { + "bbox": [ + 449, + 148, + 527, + 159 + ], + "spans": [ + { + "bbox": [ + 449, + 148, + 527, + 159 + ], + "type": "text", + "content": "(b) GPU memory usage." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 211, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 342 + ], + "type": "text", + "content": "images at 5.2FPS on average. As shown in Fig. 5a and Fig. 5b, the number of surfels has a sub-linear growth thanks to the proposed surfel integration. The fluctuation of Time (Rasterization) is caused by the different numbers of corresponding surfels at novel views. Although the computational cost of rasterization depends on the number of rasterized surfels, but it is still faster than volume rendering [22]. These results show that our surfel representation is compact and efficient. The major computation cost comes from rasterization, and the cost of surfel construction as well as integration could be neglected." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 352, + 381, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 381, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 381, + 365 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 575 + ], + "type": "text", + "content": "Our method requires high-quality surfel geometry as neural surfel representation since SurfelNeRF relies on it to conduct surfel integration and rasterization-guided render. To obtain better geometry quality, we may require more accurate depth quality to construct surfel geometry, which necessitates extra depth supervision or the MVS depth estimator. Since captured or estimated depth quality is not exact, there may be geometry discontinuities or missing parts in the surfel representation. Consequently, there may be some missing parts or incorrect geometry in the rendered output (examples of missing parts are shown in the first row in Fig. 4 or video). As discussed in ablation studies of depth refinement above, inaccurate and noisy estimated depth as well as depth discontinuity could affect the construction of neural surfel representation. To alleviate this issue, we propose a depth refinement module (Tab. 4) to refine depth information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose SurfelNeRF for online photorealistic reconstruction and rendering of large-scale indoor scenes. We employ neural surfels to store a neural radiance field, adopt an efficient neural radiance field fusion method to progressively build the scene representation, and propose a fast and differentiable rasterization-based rendering algorithm to tackle the challenging task. Experiments on the ScanNet dataset demonstrate that our method achieving favorably better results than the state-of-the-art methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "115" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 189 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855-5864, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "text", + "content": "[4] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 286, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 344 + ], + "type": "text", + "content": "[5] Ang Cao, Chris Rockwell, and Justin Johnson. Fwd: Real-time novel view synthesis with forward warping and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15713-15724, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 286, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 286, + 389 + ], + "type": "text", + "content": "[6] Yan-Pei Cao, Leif Kobbelt, and Shi-Min Hu. Real-time high-accuracy three-dimensional reconstruction with consumer RGB-D cameras. ACM Transactions on Graphics, 37(5):171:1-171:16, 2018. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 286, + 479 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 286, + 533 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. arXiv preprint arXiv:2208.00277, 2022.2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "type": "text", + "content": "[10] Ronald Clark. Volumetric bundle adjustment for online photorealistic scene capture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6124-6132, 2022. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE CVPR, pages 2432-2443, 2017. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 678 + ], + "type": "text", + "content": "[12] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(3):24:1-24:18, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[13] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 152 + ], + "type": "text", + "content": "[14] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy Mitra. Relu fields: The little non-linearity that could. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "text", + "content": "[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[16] Jargon Kim, Hyomin Kim, Hyeonseo Nam, Jaesik Park, and Seungyong Lee. Textureme: High-quality textured scene reconstruction in real time. ACM Trans. Graph., 41(3), mar 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 319 + ], + "type": "text", + "content": "[18] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "text", + "content": "[19] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. arXiv preprint arXiv:2205.15723, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 408 + ], + "type": "text", + "content": "[20] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[21] Marko Mihajlovic, Silvan Weder, Marc Pollefeys, and Martin R Oswald. Deepsurfels: Learning online appearance fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14524-14535, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "text", + "content": "[24] Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Pushmeet Kohli, Jamie Shotton, Steve Hodges, and Andrew W. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, pages 127-136, 2011. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "text", + "content": "[25] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "116" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, June 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[27] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. arXiv preprint arXiv:2111.14643, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[28] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 252 + ], + "type": "text", + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 286 + ], + "type": "text", + "content": "[30] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. arXiv preprint arXiv:2110.06635, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 319 + ], + "type": "text", + "content": "[31] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (TOG), 41(4):1-14, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 398 + ], + "type": "text", + "content": "[32] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "text", + "content": "[33] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. arXiv preprint arXiv:2111.11215, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "text", + "content": "[34] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In IEEE CVPR, pages 15598-15607, 2021. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 532 + ], + "type": "text", + "content": "[35] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. arXiv preprint arXiv:2202.05263, 2022.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[36] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 645 + ], + "type": "text", + "content": "[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[39] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 162 + ], + "type": "text", + "content": "[40] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5610-5619, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[41] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 251 + ], + "type": "text", + "content": "[42] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Citynerf: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 310, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 342 + ], + "type": "text", + "content": "[44] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 398 + ], + "type": "text", + "content": "[45] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "type": "text", + "content": "[46] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "text", + "content": "[47] Jian Zhang, Jinchi Huang, Bowen Cai, Huan Fu, Mingming Gong, Chaohui Wang, Jiaming Wang, Hongchen Luo, Rongfei Jia, Binqiang Zhao, et al. Digging into radiance grid for real-time view synthesis with detail preservation. In European Conference on Computer Vision, pages 724-740. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "text", + "content": "[48] Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. arXiv preprint arXiv:2205.14330, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 668 + ], + "type": "text", + "content": "[50] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449-5458, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[51] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "117" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[52] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In IEEE CVPR, pages 12786-12796, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[53] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, and Hongbo Fu. Mononeuralfusion: Online monocular neural 3d reconstruction with geometric priors. arXiv preprint arXiv:2209.15153, 2022. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "118" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_content_list.json b/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..05ce305eed62303f125e498e18b61b5ff3878aef --- /dev/null +++ b/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_content_list.json @@ -0,0 +1,1800 @@ +[ + { + "type": "text", + "text": "Swept-Angle Synthetic Wavelength Interferometry", + "text_level": 1, + "bbox": [ + 228, + 130, + 738, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alankar Kotwal1, Anat Levin2, and Ioannis Gkioulekas1", + "bbox": [ + 261, + 156, + 704, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University, $^{2}$ Technion", + "bbox": [ + 323, + 175, + 640, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 227, + 313, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a new imaging technique, swept-angle synthetic wavelength interferometry, for full-field micron-scale 3D sensing. As in conventional synthetic wavelength interferometry, our technique uses light consisting of two narrowly-separated optical wavelengths, resulting in per-pixel interferometric measurements whose phase encodes scene depth. Our technique additionally uses a new type of light source that, by emulating spatially-incoherent illumination, makes interferometric measurements insensitive to aberrations and (sub)surface scattering, effects that corrupt phase measurements. The resulting technique combines the robustness to such corruptions of scanning interferometric setups, with the speed of full-field interferometric setups. Overall, our technique can recover full-frame depth at a lateral and axial resolution of $5\\mu \\mathrm{m}$ , at frame rates of $5\\mathrm{Hz}$ , even under strong ambient light. We build an experimental prototype, and use it to demonstrate these capabilities by scanning a variety of objects, including objects representative of applications in inspection and fabrication, and objects that contain challenging light scattering effects.", + "bbox": [ + 75, + 247, + 473, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 564, + 207, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth sensing is among the core problems of computer vision and computational imaging, with widespread applications in medicine, industry, and robotics. An array of techniques is available for acquiring depth maps of three-dimensional (3D) scenes at different scales. In particular, micrometer-resolution depth sensing, our focus in this paper, is important in biomedical imaging because biological features are often micron-scale, industrial fabrication and inspection of critical parts that must conform to their specifications (Figure 1), and robotics to handle fine objects.", + "bbox": [ + 75, + 583, + 470, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active illumination depth sensing techniques such as lidar, structured light, and correlation time-of-flight (ToF) cannot provide micrometer axial resolution. Instead, we focus on interferometric techniques that can achieve such resolutions. The operational principles and characteristics of interferometric techniques vary, depending on the type of active illumination and optical configurations they use.", + "bbox": [ + 75, + 734, + 470, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The choice of illumination spectrum leads to techniques such as optical coherence tomography (OCT), which uses broadband illumination, and phase-shifting interferometry (PSI), which uses monochromatic illumination. We consider", + "bbox": [ + 75, + 839, + 470, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b4d699b630e04c539ae5b1f410ca5570181d27d53a7a94b715398926d13c97d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 227, + 696, + 304 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5be579f24c1237fcad3248b622c95333103d0944a395421fa5ec0f1280134dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 305, + 709, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6d086e19473fb4a9da562c40f8035887fe782bb75bfea229237d6c912f7e597.jpg", + "image_caption": [ + "photo of image of scanned object scanned scene" + ], + "image_footnote": [], + "bbox": [ + 522, + 358, + 709, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c1056f30b12d514f52ae5ea614b628718903c6adaa28822ec1adb484a28627d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 227, + 867, + 303 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26bd931f927f7d2b6a1b351aed415f6638152ac1ab1ffae1917c23954fdfa609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 305, + 867, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21190cf1af0d97804549dd211292cad8f9c61719e91463d9f98215a904903aa0.jpg", + "image_caption": [ + "ept-angle (ours)", + "without swept-angle", + "Figure 1. Applications of swept-angle SWI in industrial inspection and fabrication. We show depth reconstructions for two scenes representative of these applications: millimeter-scale dents on an aircraft fuselage section, and a 3D-printed coin." + ], + "image_footnote": [], + "bbox": [ + 709, + 358, + 867, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "synthetic wavelength interferometry (SWI), which operates between these two extremes: By using illumination consisting of two narrowly-separated optical wavelengths, SWI provides a controllable trade-off between the large unambiguous depth range of OCT, and the large axial resolution of PSI. SWI can achieve micrometer resolution at depth ranges in the order of hundreds of micrometers.", + "bbox": [ + 496, + 492, + 893, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The choice of optical configuration results in full-field versus scanning implementations, which offer different trade-offs. Full-field implementations acquire entire 2D depth maps, offering simultaneously fast operation and pixel-level lateral resolutions. However, full-field implementations are very sensitive to effects that corrupt depth estimation, such as imperfections in free-space optics (e.g., lens aberrations) and indirect illumination (e.g., subsurface scattering). By contrast, scanning implementations use beam steering to sequentially scan points in a scene and produce a 2D depth map. Scanning implementations offer robustness to depth corruption effects, through the use of fiber optics to reduce aberrations, and co-axial illumination and sensing to eliminate most indirect illumination. However, scanning makes acquiring depth maps with pixel-level lateral resolution and megapixel sizes impractically slow.", + "bbox": [ + 496, + 598, + 895, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We develop a 3D sensing technique, swept-angle synthetic wavelength interferometry, that combines the complementary advantages of full-field and scanning implementations. We draw inspiration from previous work showing", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8233", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cafa133a664e0212e031a64c142df4a37ceac0e099ab31ed85fc41d0d377b1f4.jpg", + "image_caption": [ + "bill" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 207, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/55e57b66b4c1dc777942244a0845adc0c73ce87365a819e75e2f92c12c7ba357.jpg", + "image_caption": [ + "scene" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 362, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a7296b1c72c1e129ce8d69b7364b3b4906b44bf053bcdf5d6bc1974a715918e6.jpg", + "image_caption": [ + "input image", + "Figure 2. Reconstructing the eagle embossed on a $20 bill. The features on the eagle are raised $10\\mu \\mathrm{m}$ off the surface of the bill. The recovered depth shows fine details such as the gaps between the wings reconstructed with high lateral and axial resolution." + ], + "image_footnote": [], + "bbox": [ + 364, + 88, + 516, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ab74e865e7a658ed19f3de3ccc2776691427057fd72dad222da505fc34465c8.jpg", + "image_caption": [ + "depth surface" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 671, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66bdf49815e1cd2fa80dfe60d47b0e3d048b38a80eb3932340e5622e7f22350b.jpg", + "image_caption": [ + "depth map" + ], + "image_footnote": [], + "bbox": [ + 673, + 88, + 864, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that the use of spatially-incoherent illumination in full-field interferometry mitigates aberration and indirect illumination effects. We design a light source that emulates spatial incoherence, by scanning within exposure a dichromatic point source at the focal plane of the illumination lens, effectively sweeping the illumination angle—hence swept-angle. We combine this light source with full-field SWI, resulting in a 3D sensing technique with the following performance characteristics: $5\\mathrm{Hz}$ full-frame 2 Mpixel acquisition; $5\\mu \\mathrm{m}$ lateral and axial resolution; range and resolution tunability; and robustness to aberrations, ambient illumination, and indirect illumination. We build an experimental prototype, and use it to demonstrate these capabilities, as in Figure 2. We provide setup details, reconstruction code, and data in the supplement and project website. $^1$", + "bbox": [ + 75, + 228, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Potential impact. Swept-angle SWI is relevant for critical applications, including industrial fabrication and inspection. In industrial fabrication, swept-angle SWI can be used to provide feedback during additive and subtractive manufacturing processes [67]. In industrial inspection, swept-angle SWI can be used to examine newly-fabricated or used in-place critical parts and ensure they comply with operational specifications. Swept-angle SWI, uniquely among 3D scanning techniques, offers a combination of features that are critical for both applications: First, high acquisition speed, which is necessary to avoid slowing down the manufacturing process, and to perform inspection efficiently. Second, micrometer lateral and axial resolution, which is necessary to detect critical defects. Third, robustness to aberrations and indirect illumination, which is necessary because of the materials used for fabrication, which often have strong subsurface scattering. Figure 1 showcases results representative of these applications: We scan a fuselage section from a Boeing aircraft, to detect critical defects such as scratches and bumps from collisions, at axial and lateral scales of a couple dozen micrometers. We also scan a coin pattern 3D-printed by a commercial material printer on a translucent material.", + "bbox": [ + 75, + 455, + 472, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 791, + 215, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth sensing. There are many technologies for acquiring depth in computer vision. Passive techniques rely on scene appearance under ambient light, and use cues such as stereo [4,30,50], (de)focus [21,31,69], or shading [28,34].", + "bbox": [ + 75, + 810, + 470, + 871 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7b6369174956027eea477a09f471a29942c86557e02f146f4e430eec2804d028.jpg", + "table_caption": [ + "Table 1. Comparison of interferometric depth sensing techniques for millimeter-scale scenes. ('f.f': full-field; 'scan': scanning)" + ], + "table_footnote": [], + "table_body": "
methodaxial reslateral resdepth rangeacq timerobust
f.f. TD-OCT✓✓✓✗✓
scan. FD-OCT✗✓✓✗✓
scan. SS-OCT✗✓✓✗✓
f.f. PSI✓✓✗✓✗
scan. PSI✗✓✗✗✓
f.f. SWI✓✓✓✓✗
scan. SWI✗✓✓✗✓
f.f. SA-SWI✓✓✓✓✓
", + "bbox": [ + 519, + 253, + 874, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Active techniques inject controlled light in the scene, to overcome issues such as lack of texture and limited resolution. These include structured light [6, 24, 56, 65], impulse ToF [2, 19, 22, 23, 32, 39, 46, 52, 54, 62, 70, 71], and correlation ToF [3, 15, 16, 26, 27, 33, 42, 43, 58, 60, 66]. These techniques cannot easily achieve axial resolutions below hundreds of micrometers, placing them out of scope for applications requiring micrometer resolutions.", + "bbox": [ + 496, + 395, + 893, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical interferometry. Interferometry is a classic wave-optics technique that measures the correlation, or interference, between two or more light beams that have traveled along different paths [29]. Most relevant to our work are three broad classes of interferometric techniques that provide depth sensing capabilities with different advantages and disadvantages, which we summarize in Table 1.", + "bbox": [ + 496, + 515, + 893, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Phase-shifting interferometry (PSI) techniques [9, 37] use single-frequency illumination for nanometer-scale depth sensing. Similar to correlation ToF, they estimate depth by measuring the phase of a sinusoidal waveform. However, whereas correlation ToF uses external modulation to generate megahertz-frequency waveforms, PSI directly uses the terahertz-frequency light waveform. Unfortunately, the unambiguous depth range of PSI is limited to the illumination wavelength, typically around a micrometer.", + "bbox": [ + 496, + 621, + 893, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic wavelength interferometry (SWI) (or heterodyne interferometry) techniques [7, 8, 10, 14, 44, 45] use illumination comprising multiple narrow spectral bands to synthesize waveforms at frequencies between the megahertz rates of correlation ToF and terahertz rates of PSI. These techniques allow control of the trade off between unambiguous depth range and resolution, by tuning the frequency of the synthetic wavelength. As our technique builds upon SWI, we detail its operational principles in Section 3.", + "bbox": [ + 496, + 757, + 895, + 893 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://imaging.cs.cmu.edu/swept_angle_swi", + "bbox": [ + 94, + 886, + 433, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8234", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lastly, optical coherence tomography (OCT) [20, 35, 41] uses broadband illumination to perform depth sensing analogously to impulse time-of-flight, by processing temporally-resolved (transient) responses. OCT decouples range and resolution, achieving unambiguous depth ranges up to centimeters, at micrometer axial resolutions. This flexibility comes at the cost of increased acquisition time, because of the need for axial (time-domain OCT) or lateral (Fourier-domain and swept-source OCT) scanning [13].", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mitigating indirect illumination effects. In their basic forms, active depth sensing techniques assume the presence of only direct illumination in the scene. Indirect illumination effects, such as interreflections and subsurface scattering, confound depth information. For example, in correlation ToF techniques (including PSI and SWI), indirect illumination results in incorrect phase, and thus depth estimation, an effect also known as multi-path interference. Several techniques exist for computationally mitigating indirect illumination effects, using models of multi-bounce transport [18, 36, 49], sparse reconstruction [17, 38], multiwavelength approaches [5], and neural approaches [48, 68].", + "bbox": [ + 75, + 227, + 472, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other techniques optically remove indirect illumination by probing light transport [57]. Such techniques include epipolar imaging [53, 56], high-spatial-frequency illumination [51, 61], and spatio-temporal coded illumination [1, 25, 55]. Similar probing capabilities are possible in interferometric systems, by exploiting the spatio-temporal coherence properties of the illumination [20, 41]. We adapt these techniques for robust micrometer-scale depth sensing.", + "bbox": [ + 75, + 407, + 472, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background on interferometry", + "text_level": 1, + "bbox": [ + 76, + 539, + 361, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Michelson interferometer. Our optical setup is based on the classical Michelson interferometer (Figure 3(c)). The interferometer uses a beam splitter to divide collimated input illumination into two beams: one propagates toward the scene arm, and another propagates toward the reference arm—typically a planar mirror mounted on a translation stage that can vary the mirror's distance from the beam splitter. After reflection, the two light beams recombine at the beam splitter and propagate toward the sensor.", + "bbox": [ + 75, + 558, + 470, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote by $l$ and $d(x)$ the distance from the beamsplitter of the reference mirror and the scene point that pixel $x$ images, respectively. As $l$ is a controllable parameter, we denote it explicitly. We denote by $\\pmb{u}_r(x,l)$ and $\\pmb{u}_s(x)$ the complex fields arriving at sensor pixel $x$ from the reference and scene arms respectively. Then, the sensor measures,", + "bbox": [ + 75, + 694, + 470, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, l) = \\underbrace {\\left| \\boldsymbol {u} _ {s} (x) \\right| ^ {2} + \\left| \\boldsymbol {u} _ {r} (x , l) \\right| ^ {2}} _ {\\equiv \\mathcal {I} (x, l)} + 2 \\operatorname {R e} \\left\\{\\underbrace {\\boldsymbol {u} _ {s} (x) \\boldsymbol {u} _ {r} ^ {*} (x , l)} _ {\\equiv \\mathcal {C} (x, l)} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 785, + 470, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first two terms in Equation (1) are the intensities the sensor would measure if it were observing each of the two arms separately. We call their sum the interference-free image $\\mathcal{I}$ . The third term, which we call interference, is the real part of the complex correlation $\\mathcal{C}$ between the reflected", + "bbox": [ + 75, + 825, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f70f35dd8704b3a044561a1dfdf58727b672d1199c01218665a1ddc36d65a995.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 87, + 870, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6c2ec06fa13cc80151c7cc387ce31e7d01a93729ff53d291217c80a17fe14c18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 256, + 864, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09707a6187575de1a56c8f2e2c73f78a4af31d3f81d12e8f0dea70c8cce139d7.jpg", + "image_caption": [ + "Figure 3. Synthetic wavelength interferometry. (a-b) Collimated illumination from a point source emitting at two narrowly-separated wavelengths is injected into a Michelson interferometer. (c) As the reference mirror position $l$ is scanned, each wavelength contributes to the interference a sinusoid with period equal to its wavelength. (d) The sum of these sinusoids has an envelope that is another sinusoid at a synthetic wavelength peaked at $l = d(x)$ ." + ], + "image_footnote": [], + "bbox": [ + 524, + 334, + 864, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scene and reference fields. We elaborate on how to isolate the interference term from Equation (1) in Section 4.", + "bbox": [ + 496, + 511, + 890, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic wavelength interferometry. SWI uses illumination comprising two distinct, but narrowly-separated, wavelengths that are incoherent with each other. We denote these wavelengths as $\\lambda$ and $\\lambda /1 + \\epsilon$ , corresponding to wavenumbers $\\kappa \\equiv 2\\pi /\\lambda$ and $(1 + \\epsilon)\\kappa$ , respectively. We assume that this illumination is injected into the interferometer as a collimated beam—for example, created by placing the outputs of two fiber-coupled single-frequency lasers at the focal plane of a lens, as in Figures 3 and 5. Then, we show in the supplement that the correlation $\\mathcal{C}(x,l)$ equals", + "bbox": [ + 496, + 541, + 893, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {C} (x, l) = \\exp (- 2 i \\kappa (d (x) - l)) \\\\ [ 1 + \\exp (- 2 i \\kappa \\epsilon (d (x) - l)) ]. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 696, + 893, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The interference component of the camera intensity measurements in Equation (1) equals,", + "bbox": [ + 496, + 734, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {R e} (\\mathcal {C} (x, l)) = 2 \\sin (\\kappa (2 + \\epsilon) (d (x) - l)) \\sin (\\kappa \\epsilon (d (x) - l)) (3) \\\\ \\approx 2 \\sin (2 \\kappa (d (x) - l)) \\sin (\\kappa \\epsilon (d (x) - l)). (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 767, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The approximation is accurate when $\\epsilon \\ll 1$ , i.e., when the two wavelengths are close. We observe that, as a function of $d(x) - l$ , the interference is the product of two sinusoids: first, a carrier sinusoid with carrier wavelength $\\lambda_{c} \\equiv \\lambda / 2$ and corresponding carrier wavenumber $\\kappa_{c} \\equiv 2\\lambda$ ; second, an envelope sinusoid $\\mathcal{E}$ with synthetic wavelength $\\lambda_{s} \\equiv \\lambda / \\epsilon$", + "bbox": [ + 496, + 806, + 893, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8235", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and corresponding synthetic wavenumber $\\kappa_{s} \\equiv \\epsilon \\kappa$ :", + "bbox": [ + 76, + 90, + 415, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {E}} (x, l) \\equiv \\sin (\\kappa_ {s} (d (x) - l)). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 109, + 468, + 125 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3(d) visualizes $\\operatorname{Re}\\{\\mathcal{C}(x, l)\\}$ and $\\mathcal{E}(x, l)$ . In practice, we measure only the squared amplitude of the envelope,", + "bbox": [ + 75, + 128, + 470, + 159 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\boldsymbol {\\mathcal {E}} (x, l) \\right| ^ {2} = \\sin^ {2} \\left(\\kappa_ {s} (d (x) - l)\\right) = \\frac {1 - \\cos \\left(2 \\kappa_ {s} (d (x) - l)\\right)}{2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 161, + 468, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Equation (6), we see that SWI encodes scene depth $d(x)$ in the phase $\\phi(d(x)) \\equiv 2\\kappa_s(d(x) - l)$ of the envelope sinusoid. We defer details on how to measure the envelope and estimate this phase until Section 4. We make two observations: First, SWI provides depth measurements at intervals of $[0, \\lambda_s / 2]$ , and cannot disambiguate between depths differing by an integer multiple of $\\lambda_s / 2$ . Second, the use of two wavelengths makes it possible to control the unambiguous depth range: by decreasing the separation $\\epsilon$ between the two wavelengths, we increase the unambiguous depth range, at the cost of decreasing depth resolution.", + "bbox": [ + 75, + 191, + 468, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Full-field and scanning interferometry. Figure 5 shows two types of Michelson interferometer setups that implement SWI: (a) a full-field interferometer; and (b) a scanning interferometer. We discuss their relative merits, which will motivate our proposed swept-angle interferometer setup.", + "bbox": [ + 75, + 358, + 468, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Full-field interferometers (Figure 5(a)) use free-space optics to illuminate and image the entire field of view in the scene and reference arms. They also use a two-dimensional sensor to measure interference at all locations $x$ . This enables fast interference measurements for all scene points at once, at lateral resolutions equal to the sensor pixel pitch.", + "bbox": [ + 75, + 434, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unfortunately, full-field interferometers are susceptible to phase corruption effects, as we visualize in Figure 4. Equation (5) assumes that the scene field is due to only the direct light path (solid line in Figure 4(a)), which produces a sinusoidal envelope with phase delay $d(x)$ (solid orange curve in Figure 4(b)). In practice, the scene field will include contributions from additional paths: First, indirect paths due to subsurface scattering (dashed lines in Figure 4(a)). Second, stray paths due to optical aberrations (Figure 4(c)). These paths have different lengths, and thus contribute to the envelope sinusoidal terms of different phases (dashed curves in Figure 4(b)). Their summation produces an overall sinusoidal envelope (black) with phase $d' \\neq d(x)$ , resulting in incorrect depth estimation.", + "bbox": [ + 75, + 523, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scanning interferometers (Figure 5(b)) use fiber optics (couplers, circulators, collimators) to generate a focused beam that illuminates only one point in the scene and reference arms. Additionally, they use a single-pixel sensor, focused at the same point. They also use steering optics (e.g., MEMS mirrors) to scan the focus point and capture interference measurements for the entire scene.", + "bbox": [ + 75, + 734, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scanning interferometers effectively mitigate the phase corruption effects in Figure 4: Because, at any given time, they only illuminate and image one point in the scene, they", + "bbox": [ + 75, + 840, + 468, + 886 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f56d1833565b35556c613a6a6692b69feac8aec3701d65d4eb5d2127662d677a.jpg", + "image_caption": [ + "(a) light paths in scattering slab" + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 689, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eeb47b533cb97067ac74da62988a98c468c9f8a22062ff4f1a8e945501fc0a4d.jpg", + "image_caption": [ + "(b) envelope contributions" + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 870, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5291baa79d9de5872302809f755a9860b36bfa84bcc17f5c7be27cdb2b00ce93.jpg", + "image_caption": [ + "(c) stray aberrated light paths" + ], + "image_footnote": [], + "bbox": [ + 519, + 175, + 689, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/89cb66496337ccc51bb74f2df4ab79f9c5b19db67bdd33b7ab1157f0bf0dc588.jpg", + "image_caption": [ + "(d) reference arm images", + "Figure 4. Phase corruption effects. (a) A typical scene contains both direct (solid line) and indirect (dashed lines) light paths. (b) The direct path contributes a sinusoid with the correct phase (solid orange line). The indirect paths contribute sinusoids with incorrect phases (dashed lines). Their summation results in erroneous phase estimation (dark solid line). (c) Aberrations in free-space optics result in stray light paths that also contribute incorrect phases. (d) Images of the reference arm (mirror) visualize the effects of aberrations, and their mitigation using swept-angle illumination." + ], + "image_footnote": [], + "bbox": [ + 692, + 176, + 869, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eliminate contributions from indirect paths such as those in Figure 4(a). Additionally, because the use of fiber optics minimizes aberrations, they eliminate stray paths such as those in Figure 4(c). Unfortunately, this robustness comes at the cost of having to use beam steering to scan the entire scene. This translates into long acquisition times, especially when it is necessary to measure depth at pixel-level lateral resolutions and at a sensor-equivalent field of view.", + "bbox": [ + 496, + 383, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the next section, we introduce an interferometer design that combines the fast acquisition and high lateral resolution of full-field interferometers, with the robustness to phase corruption effects of scanning interferometers.", + "bbox": [ + 496, + 505, + 890, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Swept-angle illumination", + "text_level": 1, + "bbox": [ + 500, + 579, + 733, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we design a new light source for use with full-field SWI, to mitigate the phase corruption effects shown in Figure 4—indirect illumination, aberrations. Figure 5(c) shows our final swept-angle optical design.", + "bbox": [ + 496, + 599, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interferometry with spatially-incoherent illumination. Our starting point is previous findings on the use of spatially-incoherent illumination in full-field interferometric implementations. Specifically, Gkioulekas et al. [20] showed that using spatially incoherent illumination in a Michelson interferometer is equivalent to direct-only probing: this optically rejects indirect paths and makes the correlation term $\\mathcal{C}$ depend predominantly on contributions from direct paths. Additionally, Xiao et al. [72] showed that using spatially-incoherent illumination makes the correlation term $\\mathcal{C}$ insensitive to aberrations in the free-space optics.", + "bbox": [ + 496, + 659, + 893, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both Gkioulekas et al. [20] and Xiao et al. [72] realize spatially-incoherent illumination by replacing the point emitter in Figure 3 with an area emitter, e.g., LED or halogen lamp. Unfortunately, fundamental physics dictate that extending emission area of a light source is coupled with", + "bbox": [ + 496, + 825, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8236", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9016a9ad2e7c2d5e86dd621001ffacf1ccadbcfe455ac1521b341c15642873b.jpg", + "image_caption": [ + "(a) full-field SWI" + ], + "image_footnote": [], + "bbox": [ + 116, + 85, + 349, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c2607d50d04cb10c8d0fb8d8a043db74318fde433609c539f8cae86c24588d4.jpg", + "image_caption": [ + "(b) scanning SWI" + ], + "image_footnote": [], + "bbox": [ + 354, + 87, + 560, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b197703bb937d68effbdace7500b9c0b5329be89e906caed6358b25e87a9c41e.jpg", + "image_caption": [ + "(c) full-field swept-angle SWI" + ], + "image_footnote": [], + "bbox": [ + 563, + 85, + 849, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1fd8eeeea364b497c19637e65277053565e0d62b92dc96f4b01a6148012a178.jpg", + "image_caption": [ + "Figure 5. Synthetic wavelength interferometry setups. (a) A full-field interferometer efficiently acquires full-frame depth, but is susceptible to phase corruptions due to aberrations and indirect illumination. (b) A scanning interferometer is robust to such corruptions, but requires slow lateral scanning. (c) A swept-angle full-field interferometer achieves both efficiency and robustness.", + "Figure 6. The $\\{4,4\\}$ -shift phase retrieval pipeline. (a) We measure intensity at 16 reference positions, corresponding to 4 synthetic by 4 carrier subwavelength shifts. (b) and (c) For each synthetic sub-wavelength shift, we estimate interference-free and envelope images. (d) We denoise the envelope image using joint bilateral filtering. (e) We use 4-shift phase retrieval to estimate envelope phase and depth." + ], + "image_footnote": [], + "bbox": [ + 96, + 333, + 869, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "broadening its emission spectrum. This coupling was not an issue for prior work, which focused on OCT applications that require broadband illumination; but it makes spatially-incoherent light sources incompatible with SWI, which requires narrow-linewidth dichromatic illumination.", + "bbox": [ + 75, + 478, + 470, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Emulating spatial coherence. To resolve this conundrum, we consider the following: Replacing the point emitter in Figure 3 with an area emitter changes the illumination, from a single collimated beam parallel to the optical axis, to the superposition of beams traveling along directions offset from the optical axis by angles $\\theta \\in [-\\Theta /2,\\Theta /2]$ , where $\\Theta$ depends on the emission area and lens focal length. We denote by $\\pmb{u}_s^\\theta (x)$ and $\\pmb{u}_r^\\theta (x,l)$ the complex fields resulting by each such beam reflecting on the scene and reference arms. Then, because different points on the area emitter are incoherent with each other, we can update Equation (1) as", + "bbox": [ + 75, + 553, + 470, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (x, l) = \\int_ {\\theta} I ^ {\\theta} (x, l) \\mathrm {d} \\theta ,\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 720, + 251, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {I}} (x, l) = \\int_ {\\theta} \\boldsymbol {\\mathcal {I}} ^ {\\theta} (x, l) \\mathrm {d} \\theta , \\quad \\boldsymbol {\\mathcal {C}} (x, l) = \\int_ {\\theta} \\boldsymbol {\\mathcal {C}} ^ {\\theta} (x, l) \\mathrm {d} \\theta , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 755, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI^{\\theta}(x,l)\\equiv \\underbrace{\\left|\\boldsymbol{u}_{s}^{\\theta}(x)\\right|^{2} + \\left|\\boldsymbol{u}_{r}^{\\theta}(x,l)\\right|^{2}}_{\\equiv \\boldsymbol{\\mathcal{I}}^{\\theta}(x,l)} + 2\\operatorname {Re}\\bigl\\{\\underbrace{\\boldsymbol{u}_{s}^{\\theta}(x)\\boldsymbol{u}_{r}^{\\theta*}(x,l)}_{\\equiv \\boldsymbol{\\mathcal{C}}^{\\theta}(x,l)}\\bigr \\} .\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 789, + 452, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can interpret Equation (7) as follows: the image $I$ measured using an area emitter equals the sum of the images that would be measured using independent point emitters spanning the emission area, each producing illumination at", + "bbox": [ + 75, + 832, + 470, + 892 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an angle $\\theta$ (and respectively for correlation $\\mathcal{C}$ ).", + "bbox": [ + 500, + 478, + 808, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this interpretation at hand, we design the setup of Figure 5(c) to emulate a spatially-incoherent source suitable for SWI through time-division multiplexing. We use a galvo mirror to steer a narrow collimated beam of narrow-linewidth dichromatic illumination. We focus this beam through a relay lens at a point on the focal plane of the illumination lens. As we steer the beam direction, the focus point scans an area on the focal plane. In turn, each point location results in the injection of illumination at an angle $\\theta$ in the interferometer. By scanning the focus point over a target effective emission area within a single exposure, we acquire image and correlation measurements that equal those of Equation (7). As this design sweeps the illumination angle, we term it swept-angle illumination, and its combination with full-field SWI swept-angle synthetic wavelength interferometry.", + "bbox": [ + 496, + 493, + 893, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison to Kotwal et al. [41]. The illumination module of the proposed setup in Figure 5(c) is an adaptation of the Fourier-domain redistributive projector of Kotwal et al. [41]. Compared to their setup, we do not need to use an amplitude electro-optic modulator, as we are only interested in emulating spatial incoherence (equivalent, directly probing). Additionally, their setup was designed for use with monochromatic illumination, to enable different light transport probing capabilities. By contrast, we use our setup with narrow-linewidth dichromatic illumination, to enable synthetic wavelength interferometry. More broadly,", + "bbox": [ + 496, + 719, + 893, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8237", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our key insight is that swept-angle illumination can emulate extended-area emitters with arbitrary spectral profiles.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. The $\\{M,N\\}$ -shift phase retrieval algorithm", + "text_level": 1, + "bbox": [ + 76, + 128, + 464, + 146 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We discuss our pipeline for acquiring and processing measurements to estimate the envelope phase, and thus depth. For this, in Figure 5(c), we use a nanometer-accuracy stage to vary the position $l$ of the reference mirror.", + "bbox": [ + 75, + 148, + 468, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Phase retrieval. We use the $N$ -shift phase retrieval algorithm [9, 63, 64] to compute the envelope phase: We assume we have estimates of the envelope's squared amplitude $|\\pmb{\\mathcal{E}}(x,l_n)|^2$ at reference positions $l_{n}$ corresponding to shifts by $N$ -th fractions of the synthetic half-wavelength $\\lambda_s / 2$ , $l_{n} = l + n\\lambda_{s} / 2N$ , $n\\in \\{0,\\dots ,N - 1\\}$ . Then, from Equation (6), we estimate the envelope phase as", + "bbox": [ + 76, + 208, + 470, + 314 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (d (x)) = \\arctan \\left[ \\frac {\\sum_ {n = 0} ^ {N - 1} \\left| \\boldsymbol {\\mathcal {E}} (x , l _ {n}) \\right| ^ {2} \\sin \\left(2 \\pi n / N\\right)}{\\sum_ {n = 0} ^ {N - 1} \\left| \\boldsymbol {\\mathcal {E}} (x , l _ {n}) \\right| ^ {2} \\cos \\left(2 \\pi n / N\\right)} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 315, + 470, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the depth (up to an integer multiple of $\\lambda_{s} / 2$ ) as", + "bbox": [ + 76, + 359, + 410, + 375 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd (x) = l + \\frac {\\phi (d (x))}{2 \\kappa_ {s}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 377, + 468, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Envelope estimation. Equation (8) requires estimates of $|\\mathcal{E}(x,l_n)|^2$ . For each $l_{n}$ , we capture intensity measurements $I(x,l_n^m)$ at $M$ reference positions corresponding to shifts by $M$ -th fractions of the carrier wavelength, $l_{n}^{m} = l_{n} + m\\lambda_{c} / M$ , $m \\in \\{0,\\dots,M - 1\\}$ . As the carrier wavelength $\\lambda_{c}$ is half the optical wavelength $\\lambda$ and orders of magnitude smaller than the synthetic wavelength $\\lambda_{s}$ , the envelope and interference-free image remain approximately constant across shifts, $\\mathcal{E}(x,l_n^m) \\approx \\mathcal{E}(x,l_n)$ , $\\mathcal{I}(x,l_n^m) \\approx \\mathcal{I}(x,l_n)$ . Then, from Equations (1) and (4),", + "bbox": [ + 75, + 397, + 470, + 547 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI (x, l _ {n} ^ {m}) = \\boldsymbol {\\mathcal {T}} (x, l _ {n}) + 2 \\sin \\left(\\kappa_ {c} \\left(d (x) - l _ {n} ^ {m}\\right)\\right) \\boldsymbol {\\mathcal {E}} (x, l _ {n}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 550, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We estimate the interference-free image and envelope as", + "bbox": [ + 76, + 570, + 449, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} (x, l _ {n}) = ^ {1} / M \\sum_ {m = 0} ^ {M - 1} I (x, l _ {n} ^ {m}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 585, + 468, + 614 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\boldsymbol {\\mathcal {E}} \\left(x, l _ {n}\\right) \\right| ^ {2} = 1 / _ {2 M} \\sum_ {m = 0} ^ {M - 1} \\left(I \\left(x, l _ {n} ^ {m}\\right) - \\boldsymbol {\\mathcal {I}} \\left(x, l _ {n}\\right)\\right) ^ {2}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 617, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Equations (11), (12), (8), and (9) complete our phase estimation pipeline, which we summarize in Figure 6. We note that the runtime of this pipeline is negligible relative to acquisition time. We provide algorithmic details and pseudocode in the supplement. We call this pipeline the $\\{M,N\\}$ -shift phase retrieval algorithm. The parameters $M$ and $N$ are design parameters that we can fine-tune to control a trade-off between acquisition time and depth accuracy: As we increase $M$ and $N$ , the number $M\\cdot N$ of images we capture also increases, and depth estimates become more robust to noise. The theoretical minimum number of images is achieved using $\\{3,3\\}$ shifts (9 images). In practice, we found $\\{4,4\\}$ shifts to be robust. We evaluate different $\\{M,N\\}$ combinations in the supplement.", + "bbox": [ + 75, + 645, + 470, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dealing with speckle. Interference in non-specular scenes takes the form of speckle, a high-frequency pseudo-random pattern (Figure 6(a)). This can result in noisy envelope,", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d12483adad0e5981445b2adc67e5aa2fc03c1b62909219ff2d29932cb30e341.jpg", + "image_caption": [ + "(a) with swept-angle source" + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 602, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a06672c511b5a7b2d25f3dd120bc7af3e7ed40e57f9b96c5010985314f008ec.jpg", + "image_caption": [ + "(b) without swept-angle source" + ], + "image_footnote": [], + "bbox": [ + 602, + 89, + 692, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e4cf0d27b7f3dbb7fe55e84c4354910dd7d8f07ec5ac4f6b2d0fc8e0d72734b.jpg", + "image_caption": [ + "Figure 7. Effect of swept-angle scanning. Swept-angle illumination helps mitigate speckle noise and subsurface scattering effects." + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 785, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d97392e45807ebb0e3ccc33cdc0edc06eb63038680deb24ae2d1852775a77c9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 88, + 879, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and thus phase and depth, estimates. As Figures 4(d) and 7 show, swept-angle illumination greatly mitigates speckle effects. We can further reduce the impact of speckle by denoising estimated quantities with a low-pass filter (e.g., Gaussian) [20, 41]. Alternatively, to avoid blurring image details, we can use joint bilateral filtering [12, 59] with a guide image of the scene under ambient light. Empirically, we found it better to blur the envelope estimates of Equation (12) before using them in Equation (8) (Figure 6(d)), compared to blurring the phase or depth estimates.", + "bbox": [ + 496, + 208, + 893, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 368, + 632, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental prototype. For all our experiments, we use the experimental prototype in Figure 5(c). Our prototype comprises: two distributed-Bragg-reflector lasers (wavelengths $780\\mathrm{nm}$ and $781\\mathrm{nm}$ , power $45\\mathrm{mW}$ , linewidth $1\\mathrm{MHz}$ ); a galvo mirror pair; a translation stage (resolution $10\\mathrm{nm}$ ); two compound macro lenses (focal length $200\\mathrm{mm}$ ); a CCD sensor (pixel pitch $3.7\\mu \\mathrm{m}$ , $3400\\times 2700$ pixels); and other free-space optics and mounts. We provide a detailed component list and specifications, calibration instructions, and other information in the supplement.", + "bbox": [ + 496, + 387, + 893, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the following experimental specifications: The reproduction ratio is 1:1, the field of view is $12.5\\mathrm{mm}\\times 10\\mathrm{mm}$ and the working distance is $400\\mathrm{mm}$ . The unambiguous depth range is approximately $500\\mu \\mathrm{m}$ . We use $\\{4,4\\}$ -shifts (16 images), and a minimum per-image exposure time of $10\\mathrm{ms}$ , resulting at a frame rate of $5\\mathrm{Hz}$ .", + "bbox": [ + 496, + 537, + 893, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth recovery on challenging scenes. Figures 8 and 2 show scans of many challenging scenes. We scan materials ranging from rough metallic (coins), to diffuse (music box, pill), to highly-scattering (soap and chocolate). Most scenes include fine features requiring high lateral and axial resolution (music box, business card, US quarter, dollar bill). We compare scan results using swept-angle SWI (with bilateral or Gaussian filtering), versus conventional full-field SWI (Figure 5(a), emulated by deactivating the angle-sweep galvo). In all scenes, using swept-angle illumination greatly improves reconstruction quality. The difference is more pronounced in scenes with strong subsurface scattering (music box, chocolate, soap). Even in metallic scenes where there is no subsurface scattering (coins), the use of swept-angle illumination still improves reconstruction quality, by helping mitigate aberration artifacts. Bilateral filtering further improves lateral detail compared to Gaussian filtering.", + "bbox": [ + 496, + 628, + 893, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with scanning SWI. A direct comparison of", + "bbox": [ + 500, + 885, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8238", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/130d8b8ce6a1503c46e48a29387c13cd729fecada546014e99fd33934182c5d7.jpg", + "image_caption": [ + "Figure 8. Depth reconstruction. Depth maps (left) and surface renderings (right) acquired using full-field SWI with: (c) swept-angle scanning and bilateral filtering; (d) swept-angle scanning and Gaussian filtering; (e) no swept-angle scanning and with Gaussian filtering." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 869, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "swept-angle SWI with scanning SWI is challenging, because of the differences between scanning and full-field setups (Figure 5, (b) versus (a), (c)). To qualitatively assess their relative performance, in Figure 9 we use the following experimental protocol: We downsample the depth map from our technique, by a factor of 35 in each dimension. This is approximately equal to the number of depths points a scanning SWI system operating in point-to-point mode would acquire at the same total exposure time as ours, when equipped with beam-steering optics that operate at a $30\\mathrm{kHz}$ scan rate—we detail the exact calculation in the supplement. We then use joint bilateral upsampling [40] with the same guide image as in our technique, to upsample the downsampled depth map back to its original resolution. We observe that, due to the sparse set of points scanning SWI can acquire, the reconstructions miss fine features such as the hair on the quarter and letters on the business card.", + "bbox": [ + 75, + 613, + 472, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scanning SWI systems can also operate in resonant mode, typically using Lissajous scanning to raster scan the field of", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "view [47]. Resonant mode enables much faster operation than the point-to-point mode we compare against. However, Lissajous scanning results in non-rectilinear sampling of the image plane, as scan lines smear across multiple image rows (when the fast axis is horizontal). At high magnifications, this produces strong spatial distortions and artifacts, making resonant mode operation unsuitable for applications that require micrometer lateral resolution. Our technique also uses resonant-mode Lissajous scanning, but to raster scan the focal plane of the illumination lens (Figure 5(d)), and not the image plane. Instead, our technique uses a two-dimensional sensor to sample the image plane, thus completely avoiding the artifacts above and achieving micrometer lateral resolution. The supplement discusses other challenges in achieving high lateral resolution with scanning SWI.", + "bbox": [ + 496, + 613, + 893, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, our technique is orders of magnitude faster than resonant-mode scanning SWI at the same lateral resolution. This is because, whereas scanning SWI must scan the image plane at the target (pixel-level) lateral resolution, our", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8239", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/149abe0ca9835118ada176022492b1bf15a4a621b8badd2d710de1ae592b4cc8.jpg", + "image_caption": [ + "(a) full-field SWI (ours) (b) scanning+upsampling", + "Figure 9. Comparison with upsampled scanning SWI. We emulate scanning SWI by downsampling our swept-angle SWI depth. We then bilaterally upsample it to the original resolution." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 455, + 294 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/95f960a67bca7b3c90fa037b965991f0f6030fd8b7eb5e0d066d5327e6ea59df.jpg", + "table_caption": [ + "Table 2. Depth accuracy. MedAE is the median absolute error between ground truth and estimated depth. Kernel width is the lateral size of the speckle blur filter. All quantities are in $\\mu$ m." + ], + "table_footnote": [], + "table_body": "
kernel widthwith swept-anglew/o swept-angle
RMSEMedAERMSEMedAE
78.24.818.913.2
155.13.611.29.5
212.01.610.57.3
301.61.011.16.7
", + "bbox": [ + 98, + 398, + 449, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "technique remains effective even when scanning the focal plane at much lower resolutions. Figure 10 shows images of the focal plane during scanning, taken with the measurement camera. The dark regions (insets) show that scanning resolution is much lower than pixel-level resolution.", + "bbox": [ + 75, + 493, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Axial resolution. To quantify the axial resolution of our technique, we perform the following experiment. We use a second nanometer-accurate translation stage to place the chocolate scene from Figure 8 at different depths from the camera, at increments of $1\\mu \\mathrm{m}$ . We choose this scene because it is has strong sub-surface scattering. We then compare how well full-field SWI with and without swept-angle illumination can track the scene depth. We perform this experiment using Gaussian filtering with different kernel sizes, to additionally quantify lateral resolution. The results in Table 2 show that we can achieve an axial resolution of approximately $5\\mu \\mathrm{m}$ and $1\\mu \\mathrm{m}$ , for kernel sizes $7\\mu \\mathrm{m}$ and $30\\mu \\mathrm{m}$ , respectively. We can trade off lateral for axial resolution, by increasing kernel size. Figure 8 shows that this trade-off can become more favorable with bilateral filtering.", + "bbox": [ + 75, + 568, + 470, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional experiments. In the supplement: 1) Comparisons with full-field OCT. We show that swept-angle SWI can approximate the reconstruction quality of OCT, despite being $50 \\times$ faster. 2) Tunable depth range. We show that, by adjusting the wavelength separation between the two lasers, swept-angle SWI can scan scenes with depth range $16\\mathrm{mm}$ at resolution $50\\mu \\mathrm{m}$ , at significantly better quality than con", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbd3fffb0ba0eb3e5b642bcf93fde51434095b975a64d6e67461378a4169e091.jpg", + "image_caption": [ + "Figure 10. Lissajous scanning. Image and insets of the illumination lens' focal plane, showing the Lissajous scanning pattern." + ], + "image_footnote": [], + "bbox": [ + 527, + 88, + 859, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ventional full-field SWI. 3) Robustness to ambient lighting. We show that the use of near-monochromatic illumination makes swept-angle SWI robust to strong ambient light, even at $10\\%$ signal-to-background ratios. 4) Effect of scanning resolution and $\\{M,N\\}$ settings. We show that, by adjusting the focal plane scanning resolution and $\\{M,N\\}$ parameters for phase retrieval, swept-angle SWI can trade off between acquisition time versus reconstruction quality.", + "bbox": [ + 498, + 208, + 892, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations and conclusion", + "text_level": 1, + "bbox": [ + 500, + 338, + 751, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Phase wrapping. SWI determines depth only up to an integer multiple of the synthetic half-wavelength $\\lambda_{s} / 2$ . Equivalently, all depth values are wrapped to $[0, \\lambda_{s} / 2]$ , even if the depth range is greater. To mitigate phase wrapping and extend unambiguous depth range, it is common to capture measurements at multiple synthetic wavelengths, and use them to unwrap the phase estimate [3,7,8,11]. Theoretically two synthetic wavelengths are enough to uniquely determine depth, but in practice phase unwrapping techniques are very sensitive to noise [25]. Combining phase unwrapping with our technique is an interesting future direction.", + "bbox": [ + 498, + 357, + 893, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Toward real-time operation. Our prototype acquires measurements at a frame rate of $5\\mathrm{Hz}$ , due to the need for $10\\mathrm{ms}$ per-frame exposure time. If we use a stronger laser to reduce this time, the main speed bottleneck will be the need to perform phase shifts by physically translating the reference mirror. We can mitigate this bottleneck using faster translation stages (e.g., fast microscopy stages).", + "bbox": [ + 498, + 523, + 893, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We presented a technique for fast depth sensing at micron-scale lateral and axial resolutions. Our technique, swept-angle synthetic wavelength interferometry, combines the complementary advantages of full-field interferometry (speed, pixel-level lateral resolution) and scanning interferometry (robustness to aberrations and indirect illumination). We demonstrated these advantages by scanning multiple scenes with fine geometry and strong subsurface scattering. We expect our results to motivate applications of swept-angle SWI in areas such as biomedical imaging, and industrial fabrication and inspection (Figure 1).", + "bbox": [ + 498, + 628, + 893, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank Sudershan Boovaraghavan, Yuvraj Agrawal, Arpit Agarwal, Wenzhen Yuan from CMU, and Veniamin V. Stryzheus, Brian T. Miller from The Boeing Company, who provided the samples for the experiments in Figure 1. This work was supported by NSF awards 1730147, 2047341, 2008123 (NSF-BSF 2019758), and a Sloan Research Fellowship for Ioannis Gkioulekas.", + "bbox": [ + 498, + 795, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8240", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Supreeth Achar, Joseph R Bartels, William L Whittaker, Kiriakos N Kutulakos, and Srinivasa G Narasimhan. Epipolar time-of-flight imaging. ACM TOG, 2017. 3", + "[2] Brian Aull. 3d imaging with geiger-mode avalanche photodiodes. Opt. Photon. News, 2005. 2", + "[3] Seung-Hwan Baek, Noah Walsh, Ilya Chugunov, Zheng Shi, and Felix Heide. Centimeter-wave free-space neural time-of-flight imaging. ACM Transactions on Graphics (TOG), 2022. 2, 8", + "[4] Stephen T. Barnard and William B. Thompson. Disparity analysis of images. IEEE TPAMI, 1980. 2", + "[5] Ayush Bhandari, Achuta Kadambi, Refael Whyte, Christopher Barsi, Micha Feigin, Adrian Dorrington, and Ramesh Raskar. Resolving multipath interference in time-of-flight imaging via modulation frequency diversity and sparse regularization. Optics Letters, 2014. 3", + "[6] Tongbo Chen, Hans-Peter Seidel, and Hendrik P. A. Lensch. Modulated phase-shifting for 3d scanning. IEEE/CVF CVPR, 2008. 2", + "[7] Yeou-Yen Cheng and James C. Wyatt. Two-wavelength phase shifting interferometry. Applied Optics, 1984. 2, 8", + "[8] Yeou-Yen Cheng and James C. Wyant. Multiple-wavelength phase-shifting interferometry. Applied Optics, 1985. 2, 8", + "[9] Peter de Groot. Phase Shifting Interferometry. 2011. 2, 6", + "[10] Peter de Groot and John McGarvey. Chirped synthetic-wavelength interferometry. Optics Letters, 1992. 2", + "[11] David Droeschel, Dirk Holz, and Sven Behnke. Multifrequency phase unwrapping for time-of-flight cameras. IEEE/RSJ IROS, 2010. 8", + "[12] Elmar Eisemann and Frédo Durand. Flash photography enhancement via intrinsic relighting. ACM transactions on graphics (TOG), 2004. 6", + "[13] Adolf F Fercher, Wolfgang Drexler, Christoph K Hitzenberger, and Theo Lasser. Optical coherence tomography-principles and applications. Reports on progress in physics, 2003. 3", + "[14] Adolf F. Fercher, H Z. Hu, and U. Vry. Rough surface interferometry with a two-wavelength heterodyne speckle interferometer. Applied Optics, 1985. 2", + "[15] Richard Ferriere, Johann Cussey, and John M. Dudley. Time-of-flight range detection using low-frequency intensity modulation of a cw laser diode: application to fiber length measurement. Optical Engineering, 2008. 2", + "[16] Angel Flores, Craig Robin, Ann Lanari, and Iyad Dajani. Pseudo-random binary sequence phase modulation for narrow linewidth, kilowatt, monolithic fiber amplifiers. Optics Express, 2014. 2", + "[17] Daniel Freedman, Eyal Krupka, Yoni Smolin, Ido Leichter, and Mirko Schmidt. Sra: Fast removal of general multipath for tof sensors, 2014. 3", + "[18] Stefan Fuchs. Multipath interference compensation in time-of-flight camera images. ICPR, 2010. 3", + "[19] Genevieve Gariepy, Nikola Krstajic, Robert Henderson, Chunyong Li, Robert R Thomson, Gerald S Buller, Barmak Heshmat, Ramesh Raskar, Jonathan Leach, and Daniele Faccio." + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Single-photon sensitive light-in-fight imaging. Nature Comm., 2015. 2", + "[20] Ioannis Gkioulekas, Anat Levin, Frédo Durand, and Todd Zickler. Micron-scale light transport decomposition using interferometry. ACM TOG, 2015. 3, 4, 6", + "[21] Pavel Grossmann. Depth from focus. PRL, 1987. 2", + "[22] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In ICCV, 2019. 2", + "[23] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. IEEE/CVF CVPR, 2019. 2", + "[24] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. IEEE/CVF CVPR, 2011. 2", + "[25] Mohit Gupta, Shree K Nayar, Matthias B Hullin, and Jaime Martin. Phasor imaging: A generalization of correlation-based time-of-flight imaging. ACM TOG, 2015. 3, 8", + "[26] Mohit Gupta, Andreas Velten, Shree K. Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM TOG, 2018. 2", + "[27] Felipe Gutierrez-Barragan, Syed Azer Reza, Andreas Velten, and Mohit Gupta. Practical coding function design for time-of-flight imaging. IEEE/CVF CVPR, 2019. 2", + "[28] Yudeog Han, Joon-Young Lee, and In So Kweon. High quality shape from a single rgb-d image under uncalibrated natural illumination. IEEE ICCV, 2013. 2", + "[29] Parameswaran Hariharan. Optical interferometry. Elsevier, 2003. 2", + "[30] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. 2004. 2", + "[31] Caner Hazirbas, Sebastian Georg Soyer, Maximilian Christian Staab, Laura Leal-Taixe, and Daniel Cremers. Deep depth from focus, 2018. 2", + "[32] Felix Heide, Steven Diamond, David B. Lindell, and Gordon Wetzstein. Sub-picosecond photon-efficient 3d imaging using single-photon sensors. Scientific Reports, 2018. 2", + "[33] Felix Heide, Matthias B Hullin, James Gregson, and Wolfgang Heidrich. Low-budget transient imaging using photonic mixer devices. ACM TOG, 2013. 2", + "[34] Berthold Klaus Paul Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. Technical report, 1970. 2", + "[35] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, and James G Fujimoto. Optical coherence tomography. Science, 1991. 3", + "[36] David Jimeneza, Daniel Pizarrob, Manuel Mazoa, and Sira Palazuelos. Modeling and correction of multipath interference in time of flight cameras. Image and Vision Computing, 2014. 3", + "[37] Jon L. Johnson, Timothy D. Dorney, and Daniel M. Mittleman. Enhanced depth resolution in terahertz imaging using phase-shift interferometry. Applied Physics Letters, 2001. 2", + "[38] Achuta Kadambi, Refael Whyte, Ayush Bhandari, Lee Streeter, Christopher Barsi, Adrian Dorrington, and Ramesh" + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8241", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raskar. Coded Time of Flight Cameras: Sparse Deconvolution to Address Multipath Interference and Recover Time Profiles. ACM TOG, 2013. 3", + "[39] Ahmed Kirmani, Dheera Venkatraman, Dongeek Shin, Andrea Colaço, Franco N. C. Wong, Jeffrey H. Shapiro, and Vivek K Goyal. First-photon imaging. Science, 2014. 2", + "[40] Johannes Kopf, Michael F. Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Trans. Graph., 26(3):96-es, jul 2007. 7", + "[41] Alankar Kotwal, Anat Levin, and Ioannis Gkioulekas. Interferometric transmission probing with coded mutual intensity. ACM TOG, 2020. 3, 5, 6", + "[42] Robert Lange and Peter Seitz. Solid-state time-of-flight range camera. IEEE JQE, 2001. 2", + "[43] Robert Lange, Peter Seitz, Alice Biber, and Stefan Lauxtermann. Demodulation pixels in ccd and cmos technologies. SPIE, 2000. 2", + "[44] Fengqiang Li, Florian Willomitzer, Prasanna Rangarajan, Mohit Gupta, Andreas Velten, and Oliver Cossairt. Sh-tof: Micro resolution time-of-flight imaging with superheterodyne interferometry. IEEE ICCP, 2018. 2", + "[45] Fengqiang Li, Joshua Yablon, Andreas Velten, Mohit Gupta, and Oliver Cossairt. High-depth-resolution range imaging with multiple-wavelength superheterodyne interferometry using 1550-nm lasers. Applied Optics, 2017. 2", + "[46] David B Lindell, Matthew O'Toole, and Gordon Wetzstein. Single-photon 3d imaging with deep sensor fusion. ACM TOG, 2018. 2", + "[47] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 2021. 7", + "[48] Julio Marco, Quercus Hernandez, Adolfo Muñoz, Yue Dong, Adrian Jarabo, Min H. Kim, Xin Tong, and Diego Gutierrez. Deeptof: Off-the-shelf real-time correction of multipath interference in time-of-flight imaging. ACM TOG, 2017. 3", + "[49] Nikhil Naik, Achuta Kadambi, Christoph Rhemann, Shahram Izadi, Ramesh Raskar, and Sing Bing Kang. A light transport model for mitigating multipath interference in time-of-flight sensors. IEEE/CVF CVPR, 2015. 3", + "[50] Lazaros Nalpantidis, Georgios Sirakoulis, and Antonios Gasteratos. Review of stereo vision algorithms: From software to hardware. Int. J. Optomechatronics, 2008. 2", + "[51] Shree Nayar, Gurunandan Krishnan, Michael Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. ACM TOG, 2006. 3", + "[52] Cristiano Niclass, Alexis Rochas, Pierre. Besse, and Edoardo Charbon. Design and characterization of a cmos 3-d image sensor based on single photon avalanche diodes. IEEE JSSC, 2005. 2", + "[53] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM TOG, 2015. 3", + "[54] Matthew O'Toole, Felix Heide, David B Lindell, Kai Zang, Steven Diamond, and Gordon Wetzstein. Reconstructing transient images from single-photon sensors. IEEE/CVF CVPR, 2017. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Matthew O'Toole, Felix Heide, Lei Xiao, Matthias B Hullin, Wolfgang Heidrich, and Kiriakos N Kutulakos. Temporal Frequency Probing for 5D Transient Analysis of Global Light Transport. ACM TOG, 2014. 3", + "[56] Matthew O'Toole, John Mather, and Kiriakos N. Kutulakos. 3d shape and indirect appearance by structured light transport. IEEE TPAMI, 2016. 2, 3", + "[57] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual Coding to Probe Light Transport. ACM TOG, 2012. 3", + "[58] Andrew Payne, A Dorrington, and Michael Cree. Illumination waveform optimization for time-of-flight range imaging cameras. SPIE, 2011. 2", + "[59] Georg Petschnigg, Richard Szeliski, Maneesh Agrawala, Michael Cohen, Hugues Hoppe, and Kentaro Toyama. Digital photography with flash and no-flash image pairs. ACM transactions on graphics (TOG), 2004. 6", + "[60] Dario Piatti, Fabio Remondino, and David Stoppa. State-of-the-Art of TOF Range-Imaging Sensors. 2013. 2", + "[61] Dikpal Reddy, Ravi Ramamoorthi, and Brian Curless. Frequency-space Decomposition and Acquisition of Light Transport Under Spatially Varying Illumination. ECCV, 2012. 3", + "[62] Alexis Rochas, Michael Gosch, Alexandre Serov, Pierre Besse, Rade S. Popovic, T. Lasser, and Rudolph Rigler. First fully integrated 2-d array of single-photon detectors in standard cmos technology. IEEE PTL, 2003. 2", + "[63] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. Applied optics, 36(34):8871–8876, 1997. 6", + "[64] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. ii. a generalized solution. Applied optics, 37(31):7253-7258, 1998. 6", + "[65] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. IEEE/CVF CVPR, 2003. 2", + "[66] Rudolf Schwarte, Zhanping Xu, Horst-Guenther Heinol, Joachim Olk, Ruediger Klein, Bernd Buxbaum, Helmut Fischer, and Juergen Schulte. New electro-optical mixing and correlating sensor: facilities and applications of the photonic mixer device (PMD). In Sensors, Sensor Systems, and Sensor Data Processing, 1997. 2", + "[67] Pitchaya Sitthi-Amorn, Javier E Ramos, Yuwang Wangy, Joyce Kwan, Justin Lan, Wenshou Wang, and Wojciech Matusik. Multifab: a machine vision assisted platform for multi-material 3d printing. Acm Transactions on Graphics (Tog), 2015. 2", + "[68] Shuochen Su, Felix Heide, Gordon Wetzstein, and Wolfgang Heidrich. Deep end-to-end time-of-flight imaging. CVPR, 2018. 3", + "[69] Murali Subbarao and Gopal Surya. Depth from defocus: A spatial domain approach. IJCV, 1994. 2", + "[70] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G. Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature Comm., 2012. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8242", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Federica Villa, Rudi Lussana, Danilo Bronzi, Simone Tisa, Alberto Tosi, Franco Zappa, Alberto Dalla Mora, Davide Contini, Daniel Durini, Sasha Weyers, and Werner Brockherde. Cmos imager with 1024 spads and tdcs for single-photon timing and 3-d time-of-flight. IEEE JSTQE, 2014. 2", + "[72] Peng Xiao, Mathias Fink, and A Claude Boccara. Full-field spatially incoherent illumination interferometry: a spatial resolution almost insensitive to aberrations. Optics letters, 2016. 4" + ], + "bbox": [ + 78, + 90, + 470, + 215 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8243", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_model.json b/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1efbacd7fec574c72529aaba56e2c756ece012c3 --- /dev/null +++ b/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_model.json @@ -0,0 +1,2741 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.131, + 0.74, + 0.152 + ], + "angle": 0, + "content": "Swept-Angle Synthetic Wavelength Interferometry" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.157, + 0.705, + 0.175 + ], + "angle": 0, + "content": "Alankar Kotwal1, Anat Levin2, and Ioannis Gkioulekas1" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.176, + 0.642, + 0.193 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University, \\(^{2}\\)Technion" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.228, + 0.314, + 0.244 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.474, + 0.55 + ], + "angle": 0, + "content": "We present a new imaging technique, swept-angle synthetic wavelength interferometry, for full-field micron-scale 3D sensing. As in conventional synthetic wavelength interferometry, our technique uses light consisting of two narrowly-separated optical wavelengths, resulting in per-pixel interferometric measurements whose phase encodes scene depth. Our technique additionally uses a new type of light source that, by emulating spatially-incoherent illumination, makes interferometric measurements insensitive to aberrations and (sub)surface scattering, effects that corrupt phase measurements. The resulting technique combines the robustness to such corruptions of scanning interferometric setups, with the speed of full-field interferometric setups. Overall, our technique can recover full-frame depth at a lateral and axial resolution of \\(5\\mu \\mathrm{m}\\), at frame rates of \\(5\\mathrm{Hz}\\), even under strong ambient light. We build an experimental prototype, and use it to demonstrate these capabilities by scanning a variety of objects, including objects representative of applications in inspection and fabrication, and objects that contain challenging light scattering effects." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.565, + 0.208, + 0.58 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.734 + ], + "angle": 0, + "content": "Depth sensing is among the core problems of computer vision and computational imaging, with widespread applications in medicine, industry, and robotics. An array of techniques is available for acquiring depth maps of three-dimensional (3D) scenes at different scales. In particular, micrometer-resolution depth sensing, our focus in this paper, is important in biomedical imaging because biological features are often micron-scale, industrial fabrication and inspection of critical parts that must conform to their specifications (Figure 1), and robotics to handle fine objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Active illumination depth sensing techniques such as lidar, structured light, and correlation time-of-flight (ToF) cannot provide micrometer axial resolution. Instead, we focus on interferometric techniques that can achieve such resolutions. The operational principles and characteristics of interferometric techniques vary, depending on the type of active illumination and optical configurations they use." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "The choice of illumination spectrum leads to techniques such as optical coherence tomography (OCT), which uses broadband illumination, and phase-shifting interferometry (PSI), which uses monochromatic illumination. We consider" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.228, + 0.697, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.306, + 0.71, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.359, + 0.71, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.412, + 0.706, + 0.433 + ], + "angle": 0, + "content": "photo of image of scanned object scanned scene" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.228, + 0.868, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.306, + 0.868, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.359, + 0.868, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.412, + 0.787, + 0.432 + ], + "angle": 0, + "content": "ept-angle (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.412, + 0.857, + 0.433 + ], + "angle": 0, + "content": "without swept-angle" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.434, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Figure 1. Applications of swept-angle SWI in industrial inspection and fabrication. We show depth reconstructions for two scenes representative of these applications: millimeter-scale dents on an aircraft fuselage section, and a 3D-printed coin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.598 + ], + "angle": 0, + "content": "synthetic wavelength interferometry (SWI), which operates between these two extremes: By using illumination consisting of two narrowly-separated optical wavelengths, SWI provides a controllable trade-off between the large unambiguous depth range of OCT, and the large axial resolution of PSI. SWI can achieve micrometer resolution at depth ranges in the order of hundreds of micrometers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.896, + 0.84 + ], + "angle": 0, + "content": "The choice of optical configuration results in full-field versus scanning implementations, which offer different trade-offs. Full-field implementations acquire entire 2D depth maps, offering simultaneously fast operation and pixel-level lateral resolutions. However, full-field implementations are very sensitive to effects that corrupt depth estimation, such as imperfections in free-space optics (e.g., lens aberrations) and indirect illumination (e.g., subsurface scattering). By contrast, scanning implementations use beam steering to sequentially scan points in a scene and produce a 2D depth map. Scanning implementations offer robustness to depth corruption effects, through the use of fiber optics to reduce aberrations, and co-axial illumination and sensing to eliminate most indirect illumination. However, scanning makes acquiring depth maps with pixel-level lateral resolution and megapixel sizes impractically slow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We develop a 3D sensing technique, swept-angle synthetic wavelength interferometry, that combines the complementary advantages of full-field and scanning implementations. We draw inspiration from previous work showing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8233" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.208, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.184, + 0.163, + 0.195 + ], + "angle": 0, + "content": "bill" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.363, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.185, + 0.303, + 0.196 + ], + "angle": 0, + "content": "scene" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.517, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.185, + 0.476, + 0.196 + ], + "angle": 0, + "content": "input image" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.089, + 0.672, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.185, + 0.636, + 0.196 + ], + "angle": 0, + "content": "depth surface" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.089, + 0.866, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.185, + 0.781, + 0.197 + ], + "angle": 0, + "content": "depth map" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.196, + 0.892, + 0.224 + ], + "angle": 0, + "content": "Figure 2. Reconstructing the eagle embossed on a $20 bill. The features on the eagle are raised \\(10\\mu \\mathrm{m}\\) off the surface of the bill. The recovered depth shows fine details such as the gaps between the wings reconstructed with high lateral and axial resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.473, + 0.456 + ], + "angle": 0, + "content": "that the use of spatially-incoherent illumination in full-field interferometry mitigates aberration and indirect illumination effects. We design a light source that emulates spatial incoherence, by scanning within exposure a dichromatic point source at the focal plane of the illumination lens, effectively sweeping the illumination angle—hence swept-angle. We combine this light source with full-field SWI, resulting in a 3D sensing technique with the following performance characteristics: \\(5\\mathrm{Hz}\\) full-frame 2 Mpixel acquisition; \\(5\\mu \\mathrm{m}\\) lateral and axial resolution; range and resolution tunability; and robustness to aberrations, ambient illumination, and indirect illumination. We build an experimental prototype, and use it to demonstrate these capabilities, as in Figure 2. We provide setup details, reconstruction code, and data in the supplement and project website.\\(^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.788 + ], + "angle": 0, + "content": "Potential impact. Swept-angle SWI is relevant for critical applications, including industrial fabrication and inspection. In industrial fabrication, swept-angle SWI can be used to provide feedback during additive and subtractive manufacturing processes [67]. In industrial inspection, swept-angle SWI can be used to examine newly-fabricated or used in-place critical parts and ensure they comply with operational specifications. Swept-angle SWI, uniquely among 3D scanning techniques, offers a combination of features that are critical for both applications: First, high acquisition speed, which is necessary to avoid slowing down the manufacturing process, and to perform inspection efficiently. Second, micrometer lateral and axial resolution, which is necessary to detect critical defects. Third, robustness to aberrations and indirect illumination, which is necessary because of the materials used for fabrication, which often have strong subsurface scattering. Figure 1 showcases results representative of these applications: We scan a fuselage section from a Boeing aircraft, to detect critical defects such as scratches and bumps from collisions, at axial and lateral scales of a couple dozen micrometers. We also scan a coin pattern 3D-printed by a commercial material printer on a translucent material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.216, + 0.807 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.872 + ], + "angle": 0, + "content": "Depth sensing. There are many technologies for acquiring depth in computer vision. Passive techniques rely on scene appearance under ambient light, and use cues such as stereo [4,30,50], (de)focus [21,31,69], or shading [28,34]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Table 1. Comparison of interferometric depth sensing techniques for millimeter-scale scenes. ('f.f': full-field; 'scan': scanning)" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.255, + 0.875, + 0.392 + ], + "angle": 0, + "content": "
methodaxial reslateral resdepth rangeacq timerobust
f.f. TD-OCT✓✓✓✗✓
scan. FD-OCT✗✓✓✗✓
scan. SS-OCT✗✓✓✗✓
f.f. PSI✓✓✗✓✗
scan. PSI✗✓✗✗✓
f.f. SWI✓✓✓✓✗
scan. SWI✗✓✓✗✓
f.f. SA-SWI✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.895, + 0.517 + ], + "angle": 0, + "content": "Active techniques inject controlled light in the scene, to overcome issues such as lack of texture and limited resolution. These include structured light [6, 24, 56, 65], impulse ToF [2, 19, 22, 23, 32, 39, 46, 52, 54, 62, 70, 71], and correlation ToF [3, 15, 16, 26, 27, 33, 42, 43, 58, 60, 66]. These techniques cannot easily achieve axial resolutions below hundreds of micrometers, placing them out of scope for applications requiring micrometer resolutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Optical interferometry. Interferometry is a classic wave-optics technique that measures the correlation, or interference, between two or more light beams that have traveled along different paths [29]. Most relevant to our work are three broad classes of interferometric techniques that provide depth sensing capabilities with different advantages and disadvantages, which we summarize in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.895, + 0.759 + ], + "angle": 0, + "content": "Phase-shifting interferometry (PSI) techniques [9, 37] use single-frequency illumination for nanometer-scale depth sensing. Similar to correlation ToF, they estimate depth by measuring the phase of a sinusoidal waveform. However, whereas correlation ToF uses external modulation to generate megahertz-frequency waveforms, PSI directly uses the terahertz-frequency light waveform. Unfortunately, the unambiguous depth range of PSI is limited to the illumination wavelength, typically around a micrometer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.897, + 0.894 + ], + "angle": 0, + "content": "Synthetic wavelength interferometry (SWI) (or heterodyne interferometry) techniques [7, 8, 10, 14, 44, 45] use illumination comprising multiple narrow spectral bands to synthesize waveforms at frequencies between the megahertz rates of correlation ToF and terahertz rates of PSI. These techniques allow control of the trade off between unambiguous depth range and resolution, by tuning the frequency of the synthetic wavelength. As our technique builds upon SWI, we detail its operational principles in Section 3." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.434, + 0.902 + ], + "angle": 0, + "content": "1https://imaging.cs.cmu.edu/swept_angle_swi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8234" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "Lastly, optical coherence tomography (OCT) [20, 35, 41] uses broadband illumination to perform depth sensing analogously to impulse time-of-flight, by processing temporally-resolved (transient) responses. OCT decouples range and resolution, achieving unambiguous depth ranges up to centimeters, at micrometer axial resolutions. This flexibility comes at the cost of increased acquisition time, because of the need for axial (time-domain OCT) or lateral (Fourier-domain and swept-source OCT) scanning [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.409 + ], + "angle": 0, + "content": "Mitigating indirect illumination effects. In their basic forms, active depth sensing techniques assume the presence of only direct illumination in the scene. Indirect illumination effects, such as interreflections and subsurface scattering, confound depth information. For example, in correlation ToF techniques (including PSI and SWI), indirect illumination results in incorrect phase, and thus depth estimation, an effect also known as multi-path interference. Several techniques exist for computationally mitigating indirect illumination effects, using models of multi-bounce transport [18, 36, 49], sparse reconstruction [17, 38], multiwavelength approaches [5], and neural approaches [48, 68]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Other techniques optically remove indirect illumination by probing light transport [57]. Such techniques include epipolar imaging [53, 56], high-spatial-frequency illumination [51, 61], and spatio-temporal coded illumination [1, 25, 55]. Similar probing capabilities are possible in interferometric systems, by exploiting the spatio-temporal coherence properties of the illumination [20, 41]. We adapt these techniques for robust micrometer-scale depth sensing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.362, + 0.557 + ], + "angle": 0, + "content": "3. Background on interferometry" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.472, + 0.693 + ], + "angle": 0, + "content": "The Michelson interferometer. Our optical setup is based on the classical Michelson interferometer (Figure 3(c)). The interferometer uses a beam splitter to divide collimated input illumination into two beams: one propagates toward the scene arm, and another propagates toward the reference arm—typically a planar mirror mounted on a translation stage that can vary the mirror's distance from the beam splitter. After reflection, the two light beams recombine at the beam splitter and propagate toward the sensor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.472, + 0.785 + ], + "angle": 0, + "content": "We denote by \\(l\\) and \\(d(x)\\) the distance from the beamsplitter of the reference mirror and the scene point that pixel \\(x\\) images, respectively. As \\(l\\) is a controllable parameter, we denote it explicitly. We denote by \\(\\pmb{u}_r(x,l)\\) and \\(\\pmb{u}_s(x)\\) the complex fields arriving at sensor pixel \\(x\\) from the reference and scene arms respectively. Then, the sensor measures," + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.786, + 0.472, + 0.824 + ], + "angle": 0, + "content": "\\[\nI (x, l) = \\underbrace {\\left| \\boldsymbol {u} _ {s} (x) \\right| ^ {2} + \\left| \\boldsymbol {u} _ {r} (x , l) \\right| ^ {2}} _ {\\equiv \\mathcal {I} (x, l)} + 2 \\operatorname {R e} \\left\\{\\underbrace {\\boldsymbol {u} _ {s} (x) \\boldsymbol {u} _ {r} ^ {*} (x , l)} _ {\\equiv \\mathcal {C} (x, l)} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.901 + ], + "angle": 0, + "content": "The first two terms in Equation (1) are the intensities the sensor would measure if it were observing each of the two arms separately. We call their sum the interference-free image \\(\\mathcal{I}\\). The third term, which we call interference, is the real part of the complex correlation \\(\\mathcal{C}\\) between the reflected" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.872, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.257, + 0.866, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.335, + 0.865, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.41, + 0.895, + 0.508 + ], + "angle": 0, + "content": "Figure 3. Synthetic wavelength interferometry. (a-b) Collimated illumination from a point source emitting at two narrowly-separated wavelengths is injected into a Michelson interferometer. (c) As the reference mirror position \\( l \\) is scanned, each wavelength contributes to the interference a sinusoid with period equal to its wavelength. (d) The sum of these sinusoids has an envelope that is another sinusoid at a synthetic wavelength peaked at \\( l = d(x) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.892, + 0.542 + ], + "angle": 0, + "content": "scene and reference fields. We elaborate on how to isolate the interference term from Equation (1) in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.895, + 0.693 + ], + "angle": 0, + "content": "Synthetic wavelength interferometry. SWI uses illumination comprising two distinct, but narrowly-separated, wavelengths that are incoherent with each other. We denote these wavelengths as \\(\\lambda\\) and \\(\\lambda /1 + \\epsilon\\), corresponding to wavenumbers \\(\\kappa \\equiv 2\\pi /\\lambda\\) and \\((1 + \\epsilon)\\kappa\\), respectively. We assume that this illumination is injected into the interferometer as a collimated beam—for example, created by placing the outputs of two fiber-coupled single-frequency lasers at the focal plane of a lens, as in Figures 3 and 5. Then, we show in the supplement that the correlation \\(\\mathcal{C}(x,l)\\) equals" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.697, + 0.894, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {C} (x, l) = \\exp (- 2 i \\kappa (d (x) - l)) \\\\ [ 1 + \\exp (- 2 i \\kappa \\epsilon (d (x) - l)) ]. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.765 + ], + "angle": 0, + "content": "The interference component of the camera intensity measurements in Equation (1) equals," + }, + { + "type": "equation", + "bbox": [ + 0.497, + 0.768, + 0.892, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {R e} (\\mathcal {C} (x, l)) = 2 \\sin (\\kappa (2 + \\epsilon) (d (x) - l)) \\sin (\\kappa \\epsilon (d (x) - l)) (3) \\\\ \\approx 2 \\sin (2 \\kappa (d (x) - l)) \\sin (\\kappa \\epsilon (d (x) - l)). (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.895, + 0.898 + ], + "angle": 0, + "content": "The approximation is accurate when \\(\\epsilon \\ll 1\\), i.e., when the two wavelengths are close. We observe that, as a function of \\(d(x) - l\\), the interference is the product of two sinusoids: first, a carrier sinusoid with carrier wavelength \\(\\lambda_{c} \\equiv \\lambda / 2\\) and corresponding carrier wavenumber \\(\\kappa_{c} \\equiv 2\\lambda\\); second, an envelope sinusoid \\(\\mathcal{E}\\) with synthetic wavelength \\(\\lambda_{s} \\equiv \\lambda / \\epsilon\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8235" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.416, + 0.106 + ], + "angle": 0, + "content": "and corresponding synthetic wavenumber \\(\\kappa_{s} \\equiv \\epsilon \\kappa\\):" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.11, + 0.47, + 0.126 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {E}} (x, l) \\equiv \\sin (\\kappa_ {s} (d (x) - l)). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.129, + 0.471, + 0.16 + ], + "angle": 0, + "content": "Figure 3(d) visualizes \\(\\operatorname{Re}\\{\\mathcal{C}(x, l)\\}\\) and \\(\\mathcal{E}(x, l)\\). In practice, we measure only the squared amplitude of the envelope," + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.162, + 0.47, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\left| \\boldsymbol {\\mathcal {E}} (x, l) \\right| ^ {2} = \\sin^ {2} \\left(\\kappa_ {s} (d (x) - l)\\right) = \\frac {1 - \\cos \\left(2 \\kappa_ {s} (d (x) - l)\\right)}{2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.193, + 0.469, + 0.357 + ], + "angle": 0, + "content": "From Equation (6), we see that SWI encodes scene depth \\( d(x) \\) in the phase \\( \\phi(d(x)) \\equiv 2\\kappa_s(d(x) - l) \\) of the envelope sinusoid. We defer details on how to measure the envelope and estimate this phase until Section 4. We make two observations: First, SWI provides depth measurements at intervals of \\( [0, \\lambda_s / 2] \\), and cannot disambiguate between depths differing by an integer multiple of \\( \\lambda_s / 2 \\). Second, the use of two wavelengths makes it possible to control the unambiguous depth range: by decreasing the separation \\( \\epsilon \\) between the two wavelengths, we increase the unambiguous depth range, at the cost of decreasing depth resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.469, + 0.434 + ], + "angle": 0, + "content": "Full-field and scanning interferometry. Figure 5 shows two types of Michelson interferometer setups that implement SWI: (a) a full-field interferometer; and (b) a scanning interferometer. We discuss their relative merits, which will motivate our proposed swept-angle interferometer setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.525 + ], + "angle": 0, + "content": "Full-field interferometers (Figure 5(a)) use free-space optics to illuminate and image the entire field of view in the scene and reference arms. They also use a two-dimensional sensor to measure interference at all locations \\( x \\). This enables fast interference measurements for all scene points at once, at lateral resolutions equal to the sensor pixel pitch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.736 + ], + "angle": 0, + "content": "Unfortunately, full-field interferometers are susceptible to phase corruption effects, as we visualize in Figure 4. Equation (5) assumes that the scene field is due to only the direct light path (solid line in Figure 4(a)), which produces a sinusoidal envelope with phase delay \\( d(x) \\) (solid orange curve in Figure 4(b)). In practice, the scene field will include contributions from additional paths: First, indirect paths due to subsurface scattering (dashed lines in Figure 4(a)). Second, stray paths due to optical aberrations (Figure 4(c)). These paths have different lengths, and thus contribute to the envelope sinusoidal terms of different phases (dashed curves in Figure 4(b)). Their summation produces an overall sinusoidal envelope (black) with phase \\( d' \\neq d(x) \\), resulting in incorrect depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Scanning interferometers (Figure 5(b)) use fiber optics (couplers, circulators, collimators) to generate a focused beam that illuminates only one point in the scene and reference arms. Additionally, they use a single-pixel sensor, focused at the same point. They also use steering optics (e.g., MEMS mirrors) to scan the focus point and capture interference measurements for the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.887 + ], + "angle": 0, + "content": "Scanning interferometers effectively mitigate the phase corruption effects in Figure 4: Because, at any given time, they only illuminate and image one point in the scene, they" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.691, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.163, + 0.694, + 0.176 + ], + "angle": 0, + "content": "(a) light paths in scattering slab" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.871, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.164, + 0.859, + 0.175 + ], + "angle": 0, + "content": "(b) envelope contributions" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.176, + 0.691, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.244, + 0.689, + 0.257 + ], + "angle": 0, + "content": "(c) stray aberrated light paths" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.178, + 0.87, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.244, + 0.853, + 0.257 + ], + "angle": 0, + "content": "(d) reference arm images" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Figure 4. Phase corruption effects. (a) A typical scene contains both direct (solid line) and indirect (dashed lines) light paths. (b) The direct path contributes a sinusoid with the correct phase (solid orange line). The indirect paths contribute sinusoids with incorrect phases (dashed lines). Their summation results in erroneous phase estimation (dark solid line). (c) Aberrations in free-space optics result in stray light paths that also contribute incorrect phases. (d) Images of the reference arm (mirror) visualize the effects of aberrations, and their mitigation using swept-angle illumination." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.891, + 0.505 + ], + "angle": 0, + "content": "eliminate contributions from indirect paths such as those in Figure 4(a). Additionally, because the use of fiber optics minimizes aberrations, they eliminate stray paths such as those in Figure 4(c). Unfortunately, this robustness comes at the cost of having to use beam steering to scan the entire scene. This translates into long acquisition times, especially when it is necessary to measure depth at pixel-level lateral resolutions and at a sensor-equivalent field of view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.891, + 0.566 + ], + "angle": 0, + "content": "In the next section, we introduce an interferometer design that combines the fast acquisition and high lateral resolution of full-field interferometers, with the robustness to phase corruption effects of scanning interferometers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.734, + 0.596 + ], + "angle": 0, + "content": "4. Swept-angle illumination" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.659 + ], + "angle": 0, + "content": "In this section, we design a new light source for use with full-field SWI, to mitigate the phase corruption effects shown in Figure 4—indirect illumination, aberrations. Figure 5(c) shows our final swept-angle optical design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.825 + ], + "angle": 0, + "content": "Interferometry with spatially-incoherent illumination. Our starting point is previous findings on the use of spatially-incoherent illumination in full-field interferometric implementations. Specifically, Gkioulekas et al. [20] showed that using spatially incoherent illumination in a Michelson interferometer is equivalent to direct-only probing: this optically rejects indirect paths and makes the correlation term \\(\\mathcal{C}\\) depend predominantly on contributions from direct paths. Additionally, Xiao et al. [72] showed that using spatially-incoherent illumination makes the correlation term \\(\\mathcal{C}\\) insensitive to aberrations in the free-space optics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Both Gkioulekas et al. [20] and Xiao et al. [72] realize spatially-incoherent illumination by replacing the point emitter in Figure 3 with an area emitter, e.g., LED or halogen lamp. Unfortunately, fundamental physics dictate that extending emission area of a light source is coupled with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8236" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.087, + 0.351, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.278, + 0.284, + 0.289 + ], + "angle": 0, + "content": "(a) full-field SWI" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.088, + 0.562, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.278, + 0.506, + 0.29 + ], + "angle": 0, + "content": "(b) scanning SWI" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.087, + 0.85, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.278, + 0.788, + 0.29 + ], + "angle": 0, + "content": "(c) full-field swept-angle SWI" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 5. Synthetic wavelength interferometry setups. (a) A full-field interferometer efficiently acquires full-frame depth, but is susceptible to phase corruptions due to aberrations and indirect illumination. (b) A scanning interferometer is robust to such corruptions, but requires slow lateral scanning. (c) A swept-angle full-field interferometer achieves both efficiency and robustness." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.334, + 0.87, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.43, + 0.893, + 0.472 + ], + "angle": 0, + "content": "Figure 6. The \\(\\{4,4\\}\\)-shift phase retrieval pipeline. (a) We measure intensity at 16 reference positions, corresponding to 4 synthetic by 4 carrier subwavelength shifts. (b) and (c) For each synthetic sub-wavelength shift, we estimate interference-free and envelope images. (d) We denoise the envelope image using joint bilateral filtering. (e) We use 4-shift phase retrieval to estimate envelope phase and depth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.471, + 0.553 + ], + "angle": 0, + "content": "broadening its emission spectrum. This coupling was not an issue for prior work, which focused on OCT applications that require broadband illumination; but it makes spatially-incoherent light sources incompatible with SWI, which requires narrow-linewidth dichromatic illumination." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.72 + ], + "angle": 0, + "content": "Emulating spatial coherence. To resolve this conundrum, we consider the following: Replacing the point emitter in Figure 3 with an area emitter changes the illumination, from a single collimated beam parallel to the optical axis, to the superposition of beams traveling along directions offset from the optical axis by angles \\(\\theta \\in [-\\Theta /2,\\Theta /2]\\), where \\(\\Theta\\) depends on the emission area and lens focal length. We denote by \\(\\pmb{u}_s^\\theta (x)\\) and \\(\\pmb{u}_r^\\theta (x,l)\\) the complex fields resulting by each such beam reflecting on the scene and reference arms. Then, because different points on the area emitter are incoherent with each other, we can update Equation (1) as" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.722, + 0.253, + 0.754 + ], + "angle": 0, + "content": "\\[\nI (x, l) = \\int_ {\\theta} I ^ {\\theta} (x, l) \\mathrm {d} \\theta ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.756, + 0.47, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {I}} (x, l) = \\int_ {\\theta} \\boldsymbol {\\mathcal {I}} ^ {\\theta} (x, l) \\mathrm {d} \\theta , \\quad \\boldsymbol {\\mathcal {C}} (x, l) = \\int_ {\\theta} \\boldsymbol {\\mathcal {C}} ^ {\\theta} (x, l) \\mathrm {d} \\theta , \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.79, + 0.454, + 0.831 + ], + "angle": 0, + "content": "\\[\nI^{\\theta}(x,l)\\equiv \\underbrace{\\left|\\boldsymbol{u}_{s}^{\\theta}(x)\\right|^{2} + \\left|\\boldsymbol{u}_{r}^{\\theta}(x,l)\\right|^{2}}_{\\equiv \\boldsymbol{\\mathcal{I}}^{\\theta}(x,l)} + 2\\operatorname {Re}\\bigl\\{\\underbrace{\\boldsymbol{u}_{s}^{\\theta}(x)\\boldsymbol{u}_{r}^{\\theta*}(x,l)}_{\\equiv \\boldsymbol{\\mathcal{C}}^{\\theta}(x,l)}\\bigr \\} .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.471, + 0.893 + ], + "angle": 0, + "content": "We can interpret Equation (7) as follows: the image \\(I\\) measured using an area emitter equals the sum of the images that would be measured using independent point emitters spanning the emission area, each producing illumination at" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.479, + 0.809, + 0.493 + ], + "angle": 0, + "content": "an angle \\(\\theta\\) (and respectively for correlation \\(\\mathcal{C}\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.72 + ], + "angle": 0, + "content": "With this interpretation at hand, we design the setup of Figure 5(c) to emulate a spatially-incoherent source suitable for SWI through time-division multiplexing. We use a galvo mirror to steer a narrow collimated beam of narrow-linewidth dichromatic illumination. We focus this beam through a relay lens at a point on the focal plane of the illumination lens. As we steer the beam direction, the focus point scans an area on the focal plane. In turn, each point location results in the injection of illumination at an angle \\(\\theta\\) in the interferometer. By scanning the focus point over a target effective emission area within a single exposure, we acquire image and correlation measurements that equal those of Equation (7). As this design sweeps the illumination angle, we term it swept-angle illumination, and its combination with full-field SWI swept-angle synthetic wavelength interferometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.887 + ], + "angle": 0, + "content": "Comparison to Kotwal et al. [41]. The illumination module of the proposed setup in Figure 5(c) is an adaptation of the Fourier-domain redistributive projector of Kotwal et al. [41]. Compared to their setup, we do not need to use an amplitude electro-optic modulator, as we are only interested in emulating spatial incoherence (equivalent, directly probing). Additionally, their setup was designed for use with monochromatic illumination, to enable different light transport probing capabilities. By contrast, we use our setup with narrow-linewidth dichromatic illumination, to enable synthetic wavelength interferometry. More broadly," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8237" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "our key insight is that swept-angle illumination can emulate extended-area emitters with arbitrary spectral profiles." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.129, + 0.465, + 0.147 + ], + "angle": 0, + "content": "5. The \\(\\{M,N\\}\\)-shift phase retrieval algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.149, + 0.47, + 0.208 + ], + "angle": 0, + "content": "We discuss our pipeline for acquiring and processing measurements to estimate the envelope phase, and thus depth. For this, in Figure 5(c), we use a nanometer-accuracy stage to vary the position \\(l\\) of the reference mirror." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.209, + 0.471, + 0.315 + ], + "angle": 0, + "content": "Phase retrieval. We use the \\(N\\)-shift phase retrieval algorithm [9, 63, 64] to compute the envelope phase: We assume we have estimates of the envelope's squared amplitude \\(|\\pmb{\\mathcal{E}}(x,l_n)|^2\\) at reference positions \\(l_{n}\\) corresponding to shifts by \\(N\\)-th fractions of the synthetic half-wavelength \\(\\lambda_s / 2\\), \\(l_{n} = l + n\\lambda_{s} / 2N\\), \\(n\\in \\{0,\\dots ,N - 1\\}\\). Then, from Equation (6), we estimate the envelope phase as" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.316, + 0.471, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\phi (d (x)) = \\arctan \\left[ \\frac {\\sum_ {n = 0} ^ {N - 1} \\left| \\boldsymbol {\\mathcal {E}} (x , l _ {n}) \\right| ^ {2} \\sin \\left(2 \\pi n / N\\right)}{\\sum_ {n = 0} ^ {N - 1} \\left| \\boldsymbol {\\mathcal {E}} (x , l _ {n}) \\right| ^ {2} \\cos \\left(2 \\pi n / N\\right)} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.411, + 0.375 + ], + "angle": 0, + "content": "and the depth (up to an integer multiple of \\(\\lambda_{s} / 2\\)) as" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.378, + 0.47, + 0.394 + ], + "angle": 0, + "content": "\\[\nd (x) = l + \\frac {\\phi (d (x))}{2 \\kappa_ {s}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.472, + 0.548 + ], + "angle": 0, + "content": "Envelope estimation. Equation (8) requires estimates of \\(|\\mathcal{E}(x,l_n)|^2\\). For each \\(l_{n}\\), we capture intensity measurements \\(I(x,l_n^m)\\) at \\(M\\) reference positions corresponding to shifts by \\(M\\)-th fractions of the carrier wavelength, \\(l_{n}^{m} = l_{n} + m\\lambda_{c} / M\\), \\(m \\in \\{0,\\dots,M - 1\\}\\). As the carrier wavelength \\(\\lambda_{c}\\) is half the optical wavelength \\(\\lambda\\) and orders of magnitude smaller than the synthetic wavelength \\(\\lambda_{s}\\), the envelope and interference-free image remain approximately constant across shifts, \\(\\mathcal{E}(x,l_n^m) \\approx \\mathcal{E}(x,l_n)\\), \\(\\mathcal{I}(x,l_n^m) \\approx \\mathcal{I}(x,l_n)\\). Then, from Equations (1) and (4)," + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.551, + 0.47, + 0.568 + ], + "angle": 0, + "content": "\\[\nI (x, l _ {n} ^ {m}) = \\boldsymbol {\\mathcal {T}} (x, l _ {n}) + 2 \\sin \\left(\\kappa_ {c} \\left(d (x) - l _ {n} ^ {m}\\right)\\right) \\boldsymbol {\\mathcal {E}} (x, l _ {n}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.571, + 0.45, + 0.586 + ], + "angle": 0, + "content": "We estimate the interference-free image and envelope as" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.587, + 0.47, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} (x, l _ {n}) = ^ {1} / M \\sum_ {m = 0} ^ {M - 1} I (x, l _ {n} ^ {m}), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.618, + 0.47, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\left| \\boldsymbol {\\mathcal {E}} \\left(x, l _ {n}\\right) \\right| ^ {2} = 1 / _ {2 M} \\sum_ {m = 0} ^ {M - 1} \\left(I \\left(x, l _ {n} ^ {m}\\right) - \\boldsymbol {\\mathcal {I}} \\left(x, l _ {n}\\right)\\right) ^ {2}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Equations (11), (12), (8), and (9) complete our phase estimation pipeline, which we summarize in Figure 6. We note that the runtime of this pipeline is negligible relative to acquisition time. We provide algorithmic details and pseudocode in the supplement. We call this pipeline the \\(\\{M,N\\}\\)-shift phase retrieval algorithm. The parameters \\(M\\) and \\(N\\) are design parameters that we can fine-tune to control a trade-off between acquisition time and depth accuracy: As we increase \\(M\\) and \\(N\\), the number \\(M\\cdot N\\) of images we capture also increases, and depth estimates become more robust to noise. The theoretical minimum number of images is achieved using \\(\\{3,3\\}\\) shifts (9 images). In practice, we found \\(\\{4,4\\}\\) shifts to be robust. We evaluate different \\(\\{M,N\\}\\) combinations in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Dealing with speckle. Interference in non-specular scenes takes the form of speckle, a high-frequency pseudo-random pattern (Figure 6(a)). This can result in noisy envelope," + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.603, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.09, + 0.693, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.786, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.089, + 0.88, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.163, + 0.683, + 0.177 + ], + "angle": 0, + "content": "(a) with swept-angle source" + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.163, + 0.877, + 0.177 + ], + "angle": 0, + "content": "(b) without swept-angle source" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.178, + 0.894, + 0.205 + ], + "angle": 0, + "content": "Figure 7. Effect of swept-angle scanning. Swept-angle illumination helps mitigate speckle noise and subsurface scattering effects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.894, + 0.361 + ], + "angle": 0, + "content": "and thus phase and depth, estimates. As Figures 4(d) and 7 show, swept-angle illumination greatly mitigates speckle effects. We can further reduce the impact of speckle by denoising estimated quantities with a low-pass filter (e.g., Gaussian) [20, 41]. Alternatively, to avoid blurring image details, we can use joint bilateral filtering [12, 59] with a guide image of the scene under ambient light. Empirically, we found it better to blur the envelope estimates of Equation (12) before using them in Equation (8) (Figure 6(d)), compared to blurring the phase or depth estimates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.369, + 0.633, + 0.386 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.894, + 0.538 + ], + "angle": 0, + "content": "Experimental prototype. For all our experiments, we use the experimental prototype in Figure 5(c). Our prototype comprises: two distributed-Bragg-reflector lasers (wavelengths \\(780\\mathrm{nm}\\) and \\(781\\mathrm{nm}\\), power \\(45\\mathrm{mW}\\), linewidth \\(1\\mathrm{MHz}\\)); a galvo mirror pair; a translation stage (resolution \\(10\\mathrm{nm}\\)); two compound macro lenses (focal length \\(200\\mathrm{mm}\\)); a CCD sensor (pixel pitch \\(3.7\\mu \\mathrm{m}\\), \\(3400\\times 2700\\) pixels); and other free-space optics and mounts. We provide a detailed component list and specifications, calibration instructions, and other information in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.894, + 0.628 + ], + "angle": 0, + "content": "We use the following experimental specifications: The reproduction ratio is 1:1, the field of view is \\(12.5\\mathrm{mm}\\times 10\\mathrm{mm}\\) and the working distance is \\(400\\mathrm{mm}\\). The unambiguous depth range is approximately \\(500\\mu \\mathrm{m}\\). We use \\(\\{4,4\\}\\)-shifts (16 images), and a minimum per-image exposure time of \\(10\\mathrm{ms}\\), resulting at a frame rate of \\(5\\mathrm{Hz}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Depth recovery on challenging scenes. Figures 8 and 2 show scans of many challenging scenes. We scan materials ranging from rough metallic (coins), to diffuse (music box, pill), to highly-scattering (soap and chocolate). Most scenes include fine features requiring high lateral and axial resolution (music box, business card, US quarter, dollar bill). We compare scan results using swept-angle SWI (with bilateral or Gaussian filtering), versus conventional full-field SWI (Figure 5(a), emulated by deactivating the angle-sweep galvo). In all scenes, using swept-angle illumination greatly improves reconstruction quality. The difference is more pronounced in scenes with strong subsurface scattering (music box, chocolate, soap). Even in metallic scenes where there is no subsurface scattering (coins), the use of swept-angle illumination still improves reconstruction quality, by helping mitigate aberration artifacts. Bilateral filtering further improves lateral detail compared to Gaussian filtering." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Comparison with scanning SWI. A direct comparison of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8238" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.87, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.891, + 0.611 + ], + "angle": 0, + "content": "Figure 8. Depth reconstruction. Depth maps (left) and surface renderings (right) acquired using full-field SWI with: (c) swept-angle scanning and bilateral filtering; (d) swept-angle scanning and Gaussian filtering; (e) no swept-angle scanning and with Gaussian filtering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.869 + ], + "angle": 0, + "content": "swept-angle SWI with scanning SWI is challenging, because of the differences between scanning and full-field setups (Figure 5, (b) versus (a), (c)). To qualitatively assess their relative performance, in Figure 9 we use the following experimental protocol: We downsample the depth map from our technique, by a factor of 35 in each dimension. This is approximately equal to the number of depths points a scanning SWI system operating in point-to-point mode would acquire at the same total exposure time as ours, when equipped with beam-steering optics that operate at a \\(30\\mathrm{kHz}\\) scan rate—we detail the exact calculation in the supplement. We then use joint bilateral upsampling [40] with the same guide image as in our technique, to upsample the downsampled depth map back to its original resolution. We observe that, due to the sparse set of points scanning SWI can acquire, the reconstructions miss fine features such as the hair on the quarter and letters on the business card." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Scanning SWI systems can also operate in resonant mode, typically using Lissajous scanning to raster scan the field of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.84 + ], + "angle": 0, + "content": "view [47]. Resonant mode enables much faster operation than the point-to-point mode we compare against. However, Lissajous scanning results in non-rectilinear sampling of the image plane, as scan lines smear across multiple image rows (when the fast axis is horizontal). At high magnifications, this produces strong spatial distortions and artifacts, making resonant mode operation unsuitable for applications that require micrometer lateral resolution. Our technique also uses resonant-mode Lissajous scanning, but to raster scan the focal plane of the illumination lens (Figure 5(d)), and not the image plane. Instead, our technique uses a two-dimensional sensor to sample the image plane, thus completely avoiding the artifacts above and achieving micrometer lateral resolution. The supplement discusses other challenges in achieving high lateral resolution with scanning SWI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Additionally, our technique is orders of magnitude faster than resonant-mode scanning SWI at the same lateral resolution. This is because, whereas scanning SWI must scan the image plane at the target (pixel-level) lateral resolution, our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8239" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.457, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.297, + 0.417, + 0.31 + ], + "angle": 0, + "content": "(a) full-field SWI (ours) (b) scanning+upsampling" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.31, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Figure 9. Comparison with upsampled scanning SWI. We emulate scanning SWI by downsampling our swept-angle SWI depth. We then bilaterally upsample it to the original resolution." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.354, + 0.47, + 0.396 + ], + "angle": 0, + "content": "Table 2. Depth accuracy. MedAE is the median absolute error between ground truth and estimated depth. Kernel width is the lateral size of the speckle blur filter. All quantities are in \\(\\mu\\)m." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.4, + 0.45, + 0.49 + ], + "angle": 0, + "content": "
kernel widthwith swept-anglew/o swept-angle
RMSEMedAERMSEMedAE
78.24.818.913.2
155.13.611.29.5
212.01.610.57.3
301.61.011.16.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.569 + ], + "angle": 0, + "content": "technique remains effective even when scanning the focal plane at much lower resolutions. Figure 10 shows images of the focal plane during scanning, taken with the measurement camera. The dark regions (insets) show that scanning resolution is much lower than pixel-level resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.795 + ], + "angle": 0, + "content": "Axial resolution. To quantify the axial resolution of our technique, we perform the following experiment. We use a second nanometer-accurate translation stage to place the chocolate scene from Figure 8 at different depths from the camera, at increments of \\(1\\mu \\mathrm{m}\\). We choose this scene because it is has strong sub-surface scattering. We then compare how well full-field SWI with and without swept-angle illumination can track the scene depth. We perform this experiment using Gaussian filtering with different kernel sizes, to additionally quantify lateral resolution. The results in Table 2 show that we can achieve an axial resolution of approximately \\(5\\mu \\mathrm{m}\\) and \\(1\\mu \\mathrm{m}\\), for kernel sizes \\(7\\mu \\mathrm{m}\\) and \\(30\\mu \\mathrm{m}\\), respectively. We can trade off lateral for axial resolution, by increasing kernel size. Figure 8 shows that this trade-off can become more favorable with bilateral filtering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Additional experiments. In the supplement: 1) Comparisons with full-field OCT. We show that swept-angle SWI can approximate the reconstruction quality of OCT, despite being \\(50 \\times\\) faster. 2) Tunable depth range. We show that, by adjusting the wavelength separation between the two lasers, swept-angle SWI can scan scenes with depth range \\(16\\mathrm{mm}\\) at resolution \\(50\\mu \\mathrm{m}\\), at significantly better quality than con" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.089, + 0.861, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.178, + 0.894, + 0.204 + ], + "angle": 0, + "content": "Figure 10. Lissajous scanning. Image and insets of the illumination lens' focal plane, showing the Lissajous scanning pattern." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.209, + 0.893, + 0.33 + ], + "angle": 0, + "content": "ventional full-field SWI. 3) Robustness to ambient lighting. We show that the use of near-monochromatic illumination makes swept-angle SWI robust to strong ambient light, even at \\(10\\%\\) signal-to-background ratios. 4) Effect of scanning resolution and \\(\\{M,N\\}\\) settings. We show that, by adjusting the focal plane scanning resolution and \\(\\{M,N\\}\\) parameters for phase retrieval, swept-angle SWI can trade off between acquisition time versus reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.339, + 0.753, + 0.354 + ], + "angle": 0, + "content": "7. Limitations and conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.358, + 0.895, + 0.523 + ], + "angle": 0, + "content": "Phase wrapping. SWI determines depth only up to an integer multiple of the synthetic half-wavelength \\(\\lambda_{s} / 2\\). Equivalently, all depth values are wrapped to \\([0, \\lambda_{s} / 2]\\), even if the depth range is greater. To mitigate phase wrapping and extend unambiguous depth range, it is common to capture measurements at multiple synthetic wavelengths, and use them to unwrap the phase estimate [3,7,8,11]. Theoretically two synthetic wavelengths are enough to uniquely determine depth, but in practice phase unwrapping techniques are very sensitive to noise [25]. Combining phase unwrapping with our technique is an interesting future direction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.895, + 0.63 + ], + "angle": 0, + "content": "Toward real-time operation. Our prototype acquires measurements at a frame rate of \\(5\\mathrm{Hz}\\), due to the need for \\(10\\mathrm{ms}\\) per-frame exposure time. If we use a stronger laser to reduce this time, the main speed bottleneck will be the need to perform phase shifts by physically translating the reference mirror. We can mitigate this bottleneck using faster translation stages (e.g., fast microscopy stages)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.63, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Conclusion. We presented a technique for fast depth sensing at micron-scale lateral and axial resolutions. Our technique, swept-angle synthetic wavelength interferometry, combines the complementary advantages of full-field interferometry (speed, pixel-level lateral resolution) and scanning interferometry (robustness to aberrations and indirect illumination). We demonstrated these advantages by scanning multiple scenes with fine geometry and strong subsurface scattering. We expect our results to motivate applications of swept-angle SWI in areas such as biomedical imaging, and industrial fabrication and inspection (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. We thank Sudershan Boovaraghavan, Yuvraj Agrawal, Arpit Agarwal, Wenzhen Yuan from CMU, and Veniamin V. Stryzheus, Brian T. Miller from The Boeing Company, who provided the samples for the experiments in Figure 1. This work was supported by NSF awards 1730147, 2047341, 2008123 (NSF-BSF 2019758), and a Sloan Research Fellowship for Ioannis Gkioulekas." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8240" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.157 + ], + "angle": 0, + "content": "[1] Supreeth Achar, Joseph R Bartels, William L Whittaker, Kiriakos N Kutulakos, and Srinivasa G Narasimhan. Epipolar time-of-flight imaging. ACM TOG, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.186 + ], + "angle": 0, + "content": "[2] Brian Aull. 3d imaging with geiger-mode avalanche photodiodes. Opt. Photon. News, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[3] Seung-Hwan Baek, Noah Walsh, Ilya Chugunov, Zheng Shi, and Felix Heide. Centimeter-wave free-space neural time-of-flight imaging. ACM Transactions on Graphics (TOG), 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[4] Stephen T. Barnard and William B. Thompson. Disparity analysis of images. IEEE TPAMI, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[5] Ayush Bhandari, Achuta Kadambi, Refael Whyte, Christopher Barsi, Micha Feigin, Adrian Dorrington, and Ramesh Raskar. Resolving multipath interference in time-of-flight imaging via modulation frequency diversity and sparse regularization. Optics Letters, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.343, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[6] Tongbo Chen, Hans-Peter Seidel, and Hendrik P. A. Lensch. Modulated phase-shifting for 3d scanning. IEEE/CVF CVPR, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.414 + ], + "angle": 0, + "content": "[7] Yeou-Yen Cheng and James C. Wyatt. Two-wavelength phase shifting interferometry. Applied Optics, 1984. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[8] Yeou-Yen Cheng and James C. Wyant. Multiple-wavelength phase-shifting interferometry. Applied Optics, 1985. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[9] Peter de Groot. Phase Shifting Interferometry. 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[10] Peter de Groot and John McGarvey. Chirped synthetic-wavelength interferometry. Optics Letters, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[11] David Droeschel, Dirk Holz, and Sven Behnke. Multifrequency phase unwrapping for time-of-flight cameras. IEEE/RSJ IROS, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[12] Elmar Eisemann and Frédo Durand. Flash photography enhancement via intrinsic relighting. ACM transactions on graphics (TOG), 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.628 + ], + "angle": 0, + "content": "[13] Adolf F Fercher, Wolfgang Drexler, Christoph K Hitzenberger, and Theo Lasser. Optical coherence tomography-principles and applications. Reports on progress in physics, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[14] Adolf F. Fercher, H Z. Hu, and U. Vry. Rough surface interferometry with a two-wavelength heterodyne speckle interferometer. Applied Optics, 1985. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[15] Richard Ferriere, Johann Cussey, and John M. Dudley. Time-of-flight range detection using low-frequency intensity modulation of a cw laser diode: application to fiber length measurement. Optical Engineering, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[16] Angel Flores, Craig Robin, Ann Lanari, and Iyad Dajani. Pseudo-random binary sequence phase modulation for narrow linewidth, kilowatt, monolithic fiber amplifiers. Optics Express, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.828 + ], + "angle": 0, + "content": "[17] Daniel Freedman, Eyal Krupka, Yoni Smolin, Ido Leichter, and Mirko Schmidt. Sra: Fast removal of general multipath for tof sensors, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[18] Stefan Fuchs. Multipath interference compensation in time-of-flight camera images. ICPR, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[19] Genevieve Gariepy, Nikola Krstajic, Robert Henderson, Chunyong Li, Robert R Thomson, Gerald S Buller, Barmak Heshmat, Ramesh Raskar, Jonathan Leach, and Daniele Faccio." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "Single-photon sensitive light-in-fight imaging. Nature Comm., 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.893, + 0.164 + ], + "angle": 0, + "content": "[20] Ioannis Gkioulekas, Anat Levin, Frédo Durand, and Todd Zickler. Micron-scale light transport decomposition using interferometry. ACM TOG, 2015. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.842, + 0.18 + ], + "angle": 0, + "content": "[21] Pavel Grossmann. Depth from focus. PRL, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.182, + 0.892, + 0.209 + ], + "angle": 0, + "content": "[22] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.211, + 0.893, + 0.25 + ], + "angle": 0, + "content": "[23] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. IEEE/CVF CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.254, + 0.893, + 0.295 + ], + "angle": 0, + "content": "[24] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. IEEE/CVF CVPR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.893, + 0.339 + ], + "angle": 0, + "content": "[25] Mohit Gupta, Shree K Nayar, Matthias B Hullin, and Jaime Martin. Phasor imaging: A generalization of correlation-based time-of-flight imaging. ACM TOG, 2015. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.341, + 0.893, + 0.381 + ], + "angle": 0, + "content": "[26] Mohit Gupta, Andreas Velten, Shree K. Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM TOG, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.893, + 0.425 + ], + "angle": 0, + "content": "[27] Felipe Gutierrez-Barragan, Syed Azer Reza, Andreas Velten, and Mohit Gupta. Practical coding function design for time-of-flight imaging. IEEE/CVF CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.893, + 0.468 + ], + "angle": 0, + "content": "[28] Yudeog Han, Joon-Young Lee, and In So Kweon. High quality shape from a single rgb-d image under uncalibrated natural illumination. IEEE ICCV, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[29] Parameswaran Hariharan. Optical interferometry. Elsevier, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.893, + 0.528 + ], + "angle": 0, + "content": "[30] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.893, + 0.57 + ], + "angle": 0, + "content": "[31] Caner Hazirbas, Sebastian Georg Soyer, Maximilian Christian Staab, Laura Leal-Taixe, and Daniel Cremers. Deep depth from focus, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.572, + 0.893, + 0.613 + ], + "angle": 0, + "content": "[32] Felix Heide, Steven Diamond, David B. Lindell, and Gordon Wetzstein. Sub-picosecond photon-efficient 3d imaging using single-photon sensors. Scientific Reports, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.893, + 0.656 + ], + "angle": 0, + "content": "[33] Felix Heide, Matthias B Hullin, James Gregson, and Wolfgang Heidrich. Low-budget transient imaging using photonic mixer devices. ACM TOG, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.893, + 0.7 + ], + "angle": 0, + "content": "[34] Berthold Klaus Paul Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. Technical report, 1970. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.893, + 0.77 + ], + "angle": 0, + "content": "[35] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, and James G Fujimoto. Optical coherence tomography. Science, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.828 + ], + "angle": 0, + "content": "[36] David Jimeneza, Daniel Pizarrob, Manuel Mazoa, and Sira Palazuelos. Modeling and correction of multipath interference in time of flight cameras. Image and Vision Computing, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[37] Jon L. Johnson, Timothy D. Dorney, and Daniel M. Mittleman. Enhanced depth resolution in terahertz imaging using phase-shift interferometry. Applied Physics Letters, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[38] Achuta Kadambi, Refael Whyte, Ayush Bhandari, Lee Streeter, Christopher Barsi, Adrian Dorrington, and Ramesh" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8241" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "Raskar. Coded Time of Flight Cameras: Sparse Deconvolution to Address Multipath Interference and Recover Time Profiles. ACM TOG, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[39] Ahmed Kirmani, Dheera Venkatraman, Dongeek Shin, Andrea Colaço, Franco N. C. Wong, Jeffrey H. Shapiro, and Vivek K Goyal. First-photon imaging. Science, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[40] Johannes Kopf, Michael F. Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Trans. Graph., 26(3):96-es, jul 2007. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[41] Alankar Kotwal, Anat Levin, and Ioannis Gkioulekas. Interferometric transmission probing with coded mutual intensity. ACM TOG, 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[42] Robert Lange and Peter Seitz. Solid-state time-of-flight range camera. IEEE JQE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[43] Robert Lange, Peter Seitz, Alice Biber, and Stefan Lauxtermann. Demodulation pixels in ccd and cmos technologies. SPIE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.385 + ], + "angle": 0, + "content": "[44] Fengqiang Li, Florian Willomitzer, Prasanna Rangarajan, Mohit Gupta, Andreas Velten, and Oliver Cossairt. Sh-tof: Micro resolution time-of-flight imaging with superheterodyne interferometry. IEEE ICCP, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.472, + 0.442 + ], + "angle": 0, + "content": "[45] Fengqiang Li, Joshua Yablon, Andreas Velten, Mohit Gupta, and Oliver Cossairt. High-depth-resolution range imaging with multiple-wavelength superheterodyne interferometry using 1550-nm lasers. Applied Optics, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[46] David B Lindell, Matthew O'Toole, and Gordon Wetzstein. Single-photon 3d imaging with deep sensor fusion. ACM TOG, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.537 + ], + "angle": 0, + "content": "[47] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[48] Julio Marco, Quercus Hernandez, Adolfo Muñoz, Yue Dong, Adrian Jarabo, Min H. Kim, Xin Tong, and Diego Gutierrez. Deeptof: Off-the-shelf real-time correction of multipath interference in time-of-flight imaging. ACM TOG, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[49] Nikhil Naik, Achuta Kadambi, Christoph Rhemann, Shahram Izadi, Ramesh Raskar, and Sing Bing Kang. A light transport model for mitigating multipath interference in time-of-flight sensors. IEEE/CVF CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.692 + ], + "angle": 0, + "content": "[50] Lazaros Nalpantidis, Georgios Sirakoulis, and Antonios Gasteratos. Review of stereo vision algorithms: From software to hardware. Int. J. Optomechatronics, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.747 + ], + "angle": 0, + "content": "[51] Shree Nayar, Gurunandan Krishnan, Michael Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. ACM TOG, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[52] Cristiano Niclass, Alexis Rochas, Pierre. Besse, and Edoardo Charbon. Design and characterization of a cmos 3-d image sensor based on single photon avalanche diodes. IEEE JSSC, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[53] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM TOG, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[54] Matthew O'Toole, Felix Heide, David B Lindell, Kai Zang, Steven Diamond, and Gordon Wetzstein. Reconstructing transient images from single-photon sensors. IEEE/CVF CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[55] Matthew O'Toole, Felix Heide, Lei Xiao, Matthias B Hullin, Wolfgang Heidrich, and Kiriakos N Kutulakos. Temporal Frequency Probing for 5D Transient Analysis of Global Light Transport. ACM TOG, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[56] Matthew O'Toole, John Mather, and Kiriakos N. Kutulakos. 3d shape and indirect appearance by structured light transport. IEEE TPAMI, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[57] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual Coding to Probe Light Transport. ACM TOG, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.279 + ], + "angle": 0, + "content": "[58] Andrew Payne, A Dorrington, and Michael Cree. Illumination waveform optimization for time-of-flight range imaging cameras. SPIE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.894, + 0.337 + ], + "angle": 0, + "content": "[59] Georg Petschnigg, Richard Szeliski, Maneesh Agrawala, Michael Cohen, Hugues Hoppe, and Kentaro Toyama. Digital photography with flash and no-flash image pairs. ACM transactions on graphics (TOG), 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.894, + 0.368 + ], + "angle": 0, + "content": "[60] Dario Piatti, Fabio Remondino, and David Stoppa. State-of-the-Art of TOF Range-Imaging Sensors. 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.894, + 0.423 + ], + "angle": 0, + "content": "[61] Dikpal Reddy, Ravi Ramamoorthi, and Brian Curless. Frequency-space Decomposition and Acquisition of Light Transport Under Spatially Varying Illumination. ECCV, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.893, + 0.482 + ], + "angle": 0, + "content": "[62] Alexis Rochas, Michael Gosch, Alexandre Serov, Pierre Besse, Rade S. Popovic, T. Lasser, and Rudolph Rigler. First fully integrated 2-d array of single-photon detectors in standard cmos technology. IEEE PTL, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[63] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. Applied optics, 36(34):8871–8876, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.893, + 0.582 + ], + "angle": 0, + "content": "[64] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. ii. a generalized solution. Applied optics, 37(31):7253-7258, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.893, + 0.613 + ], + "angle": 0, + "content": "[65] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. IEEE/CVF CVPR, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.893, + 0.698 + ], + "angle": 0, + "content": "[66] Rudolf Schwarte, Zhanping Xu, Horst-Guenther Heinol, Joachim Olk, Ruediger Klein, Bernd Buxbaum, Helmut Fischer, and Juergen Schulte. New electro-optical mixing and correlating sensor: facilities and applications of the photonic mixer device (PMD). In Sensors, Sensor Systems, and Sensor Data Processing, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.893, + 0.768 + ], + "angle": 0, + "content": "[67] Pitchaya Sitthi-Amorn, Javier E Ramos, Yuwang Wangy, Joyce Kwan, Justin Lan, Wenshou Wang, and Wojciech Matusik. Multifab: a machine vision assisted platform for multi-material 3d printing. Acm Transactions on Graphics (Tog), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.771, + 0.893, + 0.812 + ], + "angle": 0, + "content": "[68] Shuochen Su, Felix Heide, Gordon Wetzstein, and Wolfgang Heidrich. Deep end-to-end time-of-flight imaging. CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.814, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[69] Murali Subbarao and Gopal Surya. Depth from defocus: A spatial domain approach. IJCV, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[70] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G. Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature Comm., 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8242" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[71] Federica Villa, Rudi Lussana, Danilo Bronzi, Simone Tisa, Alberto Tosi, Franco Zappa, Alberto Dalla Mora, Davide Contini, Daniel Durini, Sasha Weyers, and Werner Brockherde. Cmos imager with 1024 spads and tdcs for single-photon timing and 3-d time-of-flight. IEEE JSTQE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[72] Peng Xiao, Mathias Fink, and A Claude Boccara. Full-field spatially incoherent illumination interferometry: a spatial resolution almost insensitive to aberrations. Optics letters, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8243" + } + ] +] \ No newline at end of file diff --git a/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_origin.pdf b/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4e14fb348b5e004aaef72b83735458dcef53996 --- /dev/null +++ b/2023/Swept-Angle Synthetic Wavelength Interferometry/3690b80e-cefe-4aee-937f-0f4b7af2250e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec1893096f24404ddbe716158fe4f3a13388b7db3304fe252846e8fc67b281d8 +size 5018678 diff --git a/2023/Swept-Angle Synthetic Wavelength Interferometry/full.md b/2023/Swept-Angle Synthetic Wavelength Interferometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb75684882b74597665c529f9cad528aa3b1e7ff --- /dev/null +++ b/2023/Swept-Angle Synthetic Wavelength Interferometry/full.md @@ -0,0 +1,384 @@ +# Swept-Angle Synthetic Wavelength Interferometry + +Alankar Kotwal1, Anat Levin2, and Ioannis Gkioulekas1 + +$^{1}$ Carnegie Mellon University, $^{2}$ Technion + +# Abstract + +We present a new imaging technique, swept-angle synthetic wavelength interferometry, for full-field micron-scale 3D sensing. As in conventional synthetic wavelength interferometry, our technique uses light consisting of two narrowly-separated optical wavelengths, resulting in per-pixel interferometric measurements whose phase encodes scene depth. Our technique additionally uses a new type of light source that, by emulating spatially-incoherent illumination, makes interferometric measurements insensitive to aberrations and (sub)surface scattering, effects that corrupt phase measurements. The resulting technique combines the robustness to such corruptions of scanning interferometric setups, with the speed of full-field interferometric setups. Overall, our technique can recover full-frame depth at a lateral and axial resolution of $5\mu \mathrm{m}$ , at frame rates of $5\mathrm{Hz}$ , even under strong ambient light. We build an experimental prototype, and use it to demonstrate these capabilities by scanning a variety of objects, including objects representative of applications in inspection and fabrication, and objects that contain challenging light scattering effects. + +# 1. Introduction + +Depth sensing is among the core problems of computer vision and computational imaging, with widespread applications in medicine, industry, and robotics. An array of techniques is available for acquiring depth maps of three-dimensional (3D) scenes at different scales. In particular, micrometer-resolution depth sensing, our focus in this paper, is important in biomedical imaging because biological features are often micron-scale, industrial fabrication and inspection of critical parts that must conform to their specifications (Figure 1), and robotics to handle fine objects. + +Active illumination depth sensing techniques such as lidar, structured light, and correlation time-of-flight (ToF) cannot provide micrometer axial resolution. Instead, we focus on interferometric techniques that can achieve such resolutions. The operational principles and characteristics of interferometric techniques vary, depending on the type of active illumination and optical configurations they use. + +The choice of illumination spectrum leads to techniques such as optical coherence tomography (OCT), which uses broadband illumination, and phase-shifting interferometry (PSI), which uses monochromatic illumination. We consider + +![](images/5b4d699b630e04c539ae5b1f410ca5570181d27d53a7a94b715398926d13c97d.jpg) + +![](images/c5be579f24c1237fcad3248b622c95333103d0944a395421fa5ec0f1280134dd.jpg) + +![](images/d6d086e19473fb4a9da562c40f8035887fe782bb75bfea229237d6c912f7e597.jpg) +photo of image of scanned object scanned scene + +![](images/2c1056f30b12d514f52ae5ea614b628718903c6adaa28822ec1adb484a28627d.jpg) + +![](images/26bd931f927f7d2b6a1b351aed415f6638152ac1ab1ffae1917c23954fdfa609.jpg) + +![](images/21190cf1af0d97804549dd211292cad8f9c61719e91463d9f98215a904903aa0.jpg) +ept-angle (ours) +without swept-angle +Figure 1. Applications of swept-angle SWI in industrial inspection and fabrication. We show depth reconstructions for two scenes representative of these applications: millimeter-scale dents on an aircraft fuselage section, and a 3D-printed coin. + +synthetic wavelength interferometry (SWI), which operates between these two extremes: By using illumination consisting of two narrowly-separated optical wavelengths, SWI provides a controllable trade-off between the large unambiguous depth range of OCT, and the large axial resolution of PSI. SWI can achieve micrometer resolution at depth ranges in the order of hundreds of micrometers. + +The choice of optical configuration results in full-field versus scanning implementations, which offer different trade-offs. Full-field implementations acquire entire 2D depth maps, offering simultaneously fast operation and pixel-level lateral resolutions. However, full-field implementations are very sensitive to effects that corrupt depth estimation, such as imperfections in free-space optics (e.g., lens aberrations) and indirect illumination (e.g., subsurface scattering). By contrast, scanning implementations use beam steering to sequentially scan points in a scene and produce a 2D depth map. Scanning implementations offer robustness to depth corruption effects, through the use of fiber optics to reduce aberrations, and co-axial illumination and sensing to eliminate most indirect illumination. However, scanning makes acquiring depth maps with pixel-level lateral resolution and megapixel sizes impractically slow. + +We develop a 3D sensing technique, swept-angle synthetic wavelength interferometry, that combines the complementary advantages of full-field and scanning implementations. We draw inspiration from previous work showing + +![](images/cafa133a664e0212e031a64c142df4a37ceac0e099ab31ed85fc41d0d377b1f4.jpg) +bill + +![](images/55e57b66b4c1dc777942244a0845adc0c73ce87365a819e75e2f92c12c7ba357.jpg) +scene + +![](images/a7296b1c72c1e129ce8d69b7364b3b4906b44bf053bcdf5d6bc1974a715918e6.jpg) +input image +Figure 2. Reconstructing the eagle embossed on a $20 bill. The features on the eagle are raised $10\mu \mathrm{m}$ off the surface of the bill. The recovered depth shows fine details such as the gaps between the wings reconstructed with high lateral and axial resolution. + +![](images/2ab74e865e7a658ed19f3de3ccc2776691427057fd72dad222da505fc34465c8.jpg) +depth surface + +![](images/66bdf49815e1cd2fa80dfe60d47b0e3d048b38a80eb3932340e5622e7f22350b.jpg) +depth map + +that the use of spatially-incoherent illumination in full-field interferometry mitigates aberration and indirect illumination effects. We design a light source that emulates spatial incoherence, by scanning within exposure a dichromatic point source at the focal plane of the illumination lens, effectively sweeping the illumination angle—hence swept-angle. We combine this light source with full-field SWI, resulting in a 3D sensing technique with the following performance characteristics: $5\mathrm{Hz}$ full-frame 2 Mpixel acquisition; $5\mu \mathrm{m}$ lateral and axial resolution; range and resolution tunability; and robustness to aberrations, ambient illumination, and indirect illumination. We build an experimental prototype, and use it to demonstrate these capabilities, as in Figure 2. We provide setup details, reconstruction code, and data in the supplement and project website. $^1$ + +Potential impact. Swept-angle SWI is relevant for critical applications, including industrial fabrication and inspection. In industrial fabrication, swept-angle SWI can be used to provide feedback during additive and subtractive manufacturing processes [67]. In industrial inspection, swept-angle SWI can be used to examine newly-fabricated or used in-place critical parts and ensure they comply with operational specifications. Swept-angle SWI, uniquely among 3D scanning techniques, offers a combination of features that are critical for both applications: First, high acquisition speed, which is necessary to avoid slowing down the manufacturing process, and to perform inspection efficiently. Second, micrometer lateral and axial resolution, which is necessary to detect critical defects. Third, robustness to aberrations and indirect illumination, which is necessary because of the materials used for fabrication, which often have strong subsurface scattering. Figure 1 showcases results representative of these applications: We scan a fuselage section from a Boeing aircraft, to detect critical defects such as scratches and bumps from collisions, at axial and lateral scales of a couple dozen micrometers. We also scan a coin pattern 3D-printed by a commercial material printer on a translucent material. + +# 2. Related work + +Depth sensing. There are many technologies for acquiring depth in computer vision. Passive techniques rely on scene appearance under ambient light, and use cues such as stereo [4,30,50], (de)focus [21,31,69], or shading [28,34]. + +Table 1. Comparison of interferometric depth sensing techniques for millimeter-scale scenes. ('f.f': full-field; 'scan': scanning) + +
methodaxial reslateral resdepth rangeacq timerobust
f.f. TD-OCT✓✓✓✗✓
scan. FD-OCT✗✓✓✗✓
scan. SS-OCT✗✓✓✗✓
f.f. PSI✓✓✗✓✗
scan. PSI✗✓✗✗✓
f.f. SWI✓✓✓✓✗
scan. SWI✗✓✓✗✓
f.f. SA-SWI✓✓✓✓✓
+ +Active techniques inject controlled light in the scene, to overcome issues such as lack of texture and limited resolution. These include structured light [6, 24, 56, 65], impulse ToF [2, 19, 22, 23, 32, 39, 46, 52, 54, 62, 70, 71], and correlation ToF [3, 15, 16, 26, 27, 33, 42, 43, 58, 60, 66]. These techniques cannot easily achieve axial resolutions below hundreds of micrometers, placing them out of scope for applications requiring micrometer resolutions. + +Optical interferometry. Interferometry is a classic wave-optics technique that measures the correlation, or interference, between two or more light beams that have traveled along different paths [29]. Most relevant to our work are three broad classes of interferometric techniques that provide depth sensing capabilities with different advantages and disadvantages, which we summarize in Table 1. + +Phase-shifting interferometry (PSI) techniques [9, 37] use single-frequency illumination for nanometer-scale depth sensing. Similar to correlation ToF, they estimate depth by measuring the phase of a sinusoidal waveform. However, whereas correlation ToF uses external modulation to generate megahertz-frequency waveforms, PSI directly uses the terahertz-frequency light waveform. Unfortunately, the unambiguous depth range of PSI is limited to the illumination wavelength, typically around a micrometer. + +Synthetic wavelength interferometry (SWI) (or heterodyne interferometry) techniques [7, 8, 10, 14, 44, 45] use illumination comprising multiple narrow spectral bands to synthesize waveforms at frequencies between the megahertz rates of correlation ToF and terahertz rates of PSI. These techniques allow control of the trade off between unambiguous depth range and resolution, by tuning the frequency of the synthetic wavelength. As our technique builds upon SWI, we detail its operational principles in Section 3. + +Lastly, optical coherence tomography (OCT) [20, 35, 41] uses broadband illumination to perform depth sensing analogously to impulse time-of-flight, by processing temporally-resolved (transient) responses. OCT decouples range and resolution, achieving unambiguous depth ranges up to centimeters, at micrometer axial resolutions. This flexibility comes at the cost of increased acquisition time, because of the need for axial (time-domain OCT) or lateral (Fourier-domain and swept-source OCT) scanning [13]. + +Mitigating indirect illumination effects. In their basic forms, active depth sensing techniques assume the presence of only direct illumination in the scene. Indirect illumination effects, such as interreflections and subsurface scattering, confound depth information. For example, in correlation ToF techniques (including PSI and SWI), indirect illumination results in incorrect phase, and thus depth estimation, an effect also known as multi-path interference. Several techniques exist for computationally mitigating indirect illumination effects, using models of multi-bounce transport [18, 36, 49], sparse reconstruction [17, 38], multiwavelength approaches [5], and neural approaches [48, 68]. + +Other techniques optically remove indirect illumination by probing light transport [57]. Such techniques include epipolar imaging [53, 56], high-spatial-frequency illumination [51, 61], and spatio-temporal coded illumination [1, 25, 55]. Similar probing capabilities are possible in interferometric systems, by exploiting the spatio-temporal coherence properties of the illumination [20, 41]. We adapt these techniques for robust micrometer-scale depth sensing. + +# 3. Background on interferometry + +The Michelson interferometer. Our optical setup is based on the classical Michelson interferometer (Figure 3(c)). The interferometer uses a beam splitter to divide collimated input illumination into two beams: one propagates toward the scene arm, and another propagates toward the reference arm—typically a planar mirror mounted on a translation stage that can vary the mirror's distance from the beam splitter. After reflection, the two light beams recombine at the beam splitter and propagate toward the sensor. + +We denote by $l$ and $d(x)$ the distance from the beamsplitter of the reference mirror and the scene point that pixel $x$ images, respectively. As $l$ is a controllable parameter, we denote it explicitly. We denote by $\pmb{u}_r(x,l)$ and $\pmb{u}_s(x)$ the complex fields arriving at sensor pixel $x$ from the reference and scene arms respectively. Then, the sensor measures, + +$$ +I (x, l) = \underbrace {\left| \boldsymbol {u} _ {s} (x) \right| ^ {2} + \left| \boldsymbol {u} _ {r} (x , l) \right| ^ {2}} _ {\equiv \mathcal {I} (x, l)} + 2 \operatorname {R e} \left\{\underbrace {\boldsymbol {u} _ {s} (x) \boldsymbol {u} _ {r} ^ {*} (x , l)} _ {\equiv \mathcal {C} (x, l)} \right\}. \tag {1} +$$ + +The first two terms in Equation (1) are the intensities the sensor would measure if it were observing each of the two arms separately. We call their sum the interference-free image $\mathcal{I}$ . The third term, which we call interference, is the real part of the complex correlation $\mathcal{C}$ between the reflected + +![](images/f70f35dd8704b3a044561a1dfdf58727b672d1199c01218665a1ddc36d65a995.jpg) + +![](images/6c2ec06fa13cc80151c7cc387ce31e7d01a93729ff53d291217c80a17fe14c18.jpg) + +![](images/09707a6187575de1a56c8f2e2c73f78a4af31d3f81d12e8f0dea70c8cce139d7.jpg) +Figure 3. Synthetic wavelength interferometry. (a-b) Collimated illumination from a point source emitting at two narrowly-separated wavelengths is injected into a Michelson interferometer. (c) As the reference mirror position $l$ is scanned, each wavelength contributes to the interference a sinusoid with period equal to its wavelength. (d) The sum of these sinusoids has an envelope that is another sinusoid at a synthetic wavelength peaked at $l = d(x)$ . + +scene and reference fields. We elaborate on how to isolate the interference term from Equation (1) in Section 4. + +Synthetic wavelength interferometry. SWI uses illumination comprising two distinct, but narrowly-separated, wavelengths that are incoherent with each other. We denote these wavelengths as $\lambda$ and $\lambda /1 + \epsilon$ , corresponding to wavenumbers $\kappa \equiv 2\pi /\lambda$ and $(1 + \epsilon)\kappa$ , respectively. We assume that this illumination is injected into the interferometer as a collimated beam—for example, created by placing the outputs of two fiber-coupled single-frequency lasers at the focal plane of a lens, as in Figures 3 and 5. Then, we show in the supplement that the correlation $\mathcal{C}(x,l)$ equals + +$$ +\begin{array}{l} \boldsymbol {C} (x, l) = \exp (- 2 i \kappa (d (x) - l)) \\ [ 1 + \exp (- 2 i \kappa \epsilon (d (x) - l)) ]. \tag {2} \\ \end{array} +$$ + +The interference component of the camera intensity measurements in Equation (1) equals, + +$$ +\begin{array}{l} \operatorname {R e} (\mathcal {C} (x, l)) = 2 \sin (\kappa (2 + \epsilon) (d (x) - l)) \sin (\kappa \epsilon (d (x) - l)) (3) \\ \approx 2 \sin (2 \kappa (d (x) - l)) \sin (\kappa \epsilon (d (x) - l)). (4) \\ \end{array} +$$ + +The approximation is accurate when $\epsilon \ll 1$ , i.e., when the two wavelengths are close. We observe that, as a function of $d(x) - l$ , the interference is the product of two sinusoids: first, a carrier sinusoid with carrier wavelength $\lambda_{c} \equiv \lambda / 2$ and corresponding carrier wavenumber $\kappa_{c} \equiv 2\lambda$ ; second, an envelope sinusoid $\mathcal{E}$ with synthetic wavelength $\lambda_{s} \equiv \lambda / \epsilon$ + +and corresponding synthetic wavenumber $\kappa_{s} \equiv \epsilon \kappa$ : + +$$ +\boldsymbol {\mathcal {E}} (x, l) \equiv \sin (\kappa_ {s} (d (x) - l)). \tag {5} +$$ + +Figure 3(d) visualizes $\operatorname{Re}\{\mathcal{C}(x, l)\}$ and $\mathcal{E}(x, l)$ . In practice, we measure only the squared amplitude of the envelope, + +$$ +\left| \boldsymbol {\mathcal {E}} (x, l) \right| ^ {2} = \sin^ {2} \left(\kappa_ {s} (d (x) - l)\right) = \frac {1 - \cos \left(2 \kappa_ {s} (d (x) - l)\right)}{2}. \tag {6} +$$ + +From Equation (6), we see that SWI encodes scene depth $d(x)$ in the phase $\phi(d(x)) \equiv 2\kappa_s(d(x) - l)$ of the envelope sinusoid. We defer details on how to measure the envelope and estimate this phase until Section 4. We make two observations: First, SWI provides depth measurements at intervals of $[0, \lambda_s / 2]$ , and cannot disambiguate between depths differing by an integer multiple of $\lambda_s / 2$ . Second, the use of two wavelengths makes it possible to control the unambiguous depth range: by decreasing the separation $\epsilon$ between the two wavelengths, we increase the unambiguous depth range, at the cost of decreasing depth resolution. + +Full-field and scanning interferometry. Figure 5 shows two types of Michelson interferometer setups that implement SWI: (a) a full-field interferometer; and (b) a scanning interferometer. We discuss their relative merits, which will motivate our proposed swept-angle interferometer setup. + +Full-field interferometers (Figure 5(a)) use free-space optics to illuminate and image the entire field of view in the scene and reference arms. They also use a two-dimensional sensor to measure interference at all locations $x$ . This enables fast interference measurements for all scene points at once, at lateral resolutions equal to the sensor pixel pitch. + +Unfortunately, full-field interferometers are susceptible to phase corruption effects, as we visualize in Figure 4. Equation (5) assumes that the scene field is due to only the direct light path (solid line in Figure 4(a)), which produces a sinusoidal envelope with phase delay $d(x)$ (solid orange curve in Figure 4(b)). In practice, the scene field will include contributions from additional paths: First, indirect paths due to subsurface scattering (dashed lines in Figure 4(a)). Second, stray paths due to optical aberrations (Figure 4(c)). These paths have different lengths, and thus contribute to the envelope sinusoidal terms of different phases (dashed curves in Figure 4(b)). Their summation produces an overall sinusoidal envelope (black) with phase $d' \neq d(x)$ , resulting in incorrect depth estimation. + +Scanning interferometers (Figure 5(b)) use fiber optics (couplers, circulators, collimators) to generate a focused beam that illuminates only one point in the scene and reference arms. Additionally, they use a single-pixel sensor, focused at the same point. They also use steering optics (e.g., MEMS mirrors) to scan the focus point and capture interference measurements for the entire scene. + +Scanning interferometers effectively mitigate the phase corruption effects in Figure 4: Because, at any given time, they only illuminate and image one point in the scene, they + +![](images/f56d1833565b35556c613a6a6692b69feac8aec3701d65d4eb5d2127662d677a.jpg) +(a) light paths in scattering slab + +![](images/eeb47b533cb97067ac74da62988a98c468c9f8a22062ff4f1a8e945501fc0a4d.jpg) +(b) envelope contributions + +![](images/5291baa79d9de5872302809f755a9860b36bfa84bcc17f5c7be27cdb2b00ce93.jpg) +(c) stray aberrated light paths + +![](images/89cb66496337ccc51bb74f2df4ab79f9c5b19db67bdd33b7ab1157f0bf0dc588.jpg) +(d) reference arm images +Figure 4. Phase corruption effects. (a) A typical scene contains both direct (solid line) and indirect (dashed lines) light paths. (b) The direct path contributes a sinusoid with the correct phase (solid orange line). The indirect paths contribute sinusoids with incorrect phases (dashed lines). Their summation results in erroneous phase estimation (dark solid line). (c) Aberrations in free-space optics result in stray light paths that also contribute incorrect phases. (d) Images of the reference arm (mirror) visualize the effects of aberrations, and their mitigation using swept-angle illumination. + +eliminate contributions from indirect paths such as those in Figure 4(a). Additionally, because the use of fiber optics minimizes aberrations, they eliminate stray paths such as those in Figure 4(c). Unfortunately, this robustness comes at the cost of having to use beam steering to scan the entire scene. This translates into long acquisition times, especially when it is necessary to measure depth at pixel-level lateral resolutions and at a sensor-equivalent field of view. + +In the next section, we introduce an interferometer design that combines the fast acquisition and high lateral resolution of full-field interferometers, with the robustness to phase corruption effects of scanning interferometers. + +# 4. Swept-angle illumination + +In this section, we design a new light source for use with full-field SWI, to mitigate the phase corruption effects shown in Figure 4—indirect illumination, aberrations. Figure 5(c) shows our final swept-angle optical design. + +Interferometry with spatially-incoherent illumination. Our starting point is previous findings on the use of spatially-incoherent illumination in full-field interferometric implementations. Specifically, Gkioulekas et al. [20] showed that using spatially incoherent illumination in a Michelson interferometer is equivalent to direct-only probing: this optically rejects indirect paths and makes the correlation term $\mathcal{C}$ depend predominantly on contributions from direct paths. Additionally, Xiao et al. [72] showed that using spatially-incoherent illumination makes the correlation term $\mathcal{C}$ insensitive to aberrations in the free-space optics. + +Both Gkioulekas et al. [20] and Xiao et al. [72] realize spatially-incoherent illumination by replacing the point emitter in Figure 3 with an area emitter, e.g., LED or halogen lamp. Unfortunately, fundamental physics dictate that extending emission area of a light source is coupled with + +![](images/e9016a9ad2e7c2d5e86dd621001ffacf1ccadbcfe455ac1521b341c15642873b.jpg) +(a) full-field SWI + +![](images/9c2607d50d04cb10c8d0fb8d8a043db74318fde433609c539f8cae86c24588d4.jpg) +(b) scanning SWI + +![](images/b197703bb937d68effbdace7500b9c0b5329be89e906caed6358b25e87a9c41e.jpg) +(c) full-field swept-angle SWI + +![](images/d1fd8eeeea364b497c19637e65277053565e0d62b92dc96f4b01a6148012a178.jpg) +Figure 5. Synthetic wavelength interferometry setups. (a) A full-field interferometer efficiently acquires full-frame depth, but is susceptible to phase corruptions due to aberrations and indirect illumination. (b) A scanning interferometer is robust to such corruptions, but requires slow lateral scanning. (c) A swept-angle full-field interferometer achieves both efficiency and robustness. +Figure 6. The $\{4,4\}$ -shift phase retrieval pipeline. (a) We measure intensity at 16 reference positions, corresponding to 4 synthetic by 4 carrier subwavelength shifts. (b) and (c) For each synthetic sub-wavelength shift, we estimate interference-free and envelope images. (d) We denoise the envelope image using joint bilateral filtering. (e) We use 4-shift phase retrieval to estimate envelope phase and depth. + +broadening its emission spectrum. This coupling was not an issue for prior work, which focused on OCT applications that require broadband illumination; but it makes spatially-incoherent light sources incompatible with SWI, which requires narrow-linewidth dichromatic illumination. + +Emulating spatial coherence. To resolve this conundrum, we consider the following: Replacing the point emitter in Figure 3 with an area emitter changes the illumination, from a single collimated beam parallel to the optical axis, to the superposition of beams traveling along directions offset from the optical axis by angles $\theta \in [-\Theta /2,\Theta /2]$ , where $\Theta$ depends on the emission area and lens focal length. We denote by $\pmb{u}_s^\theta (x)$ and $\pmb{u}_r^\theta (x,l)$ the complex fields resulting by each such beam reflecting on the scene and reference arms. Then, because different points on the area emitter are incoherent with each other, we can update Equation (1) as + +$$ +I (x, l) = \int_ {\theta} I ^ {\theta} (x, l) \mathrm {d} \theta , +$$ + +$$ +\boldsymbol {\mathcal {I}} (x, l) = \int_ {\theta} \boldsymbol {\mathcal {I}} ^ {\theta} (x, l) \mathrm {d} \theta , \quad \boldsymbol {\mathcal {C}} (x, l) = \int_ {\theta} \boldsymbol {\mathcal {C}} ^ {\theta} (x, l) \mathrm {d} \theta , \tag {7} +$$ + +$$ +I^{\theta}(x,l)\equiv \underbrace{\left|\boldsymbol{u}_{s}^{\theta}(x)\right|^{2} + \left|\boldsymbol{u}_{r}^{\theta}(x,l)\right|^{2}}_{\equiv \boldsymbol{\mathcal{I}}^{\theta}(x,l)} + 2\operatorname {Re}\bigl\{\underbrace{\boldsymbol{u}_{s}^{\theta}(x)\boldsymbol{u}_{r}^{\theta*}(x,l)}_{\equiv \boldsymbol{\mathcal{C}}^{\theta}(x,l)}\bigr \} . +$$ + +We can interpret Equation (7) as follows: the image $I$ measured using an area emitter equals the sum of the images that would be measured using independent point emitters spanning the emission area, each producing illumination at + +an angle $\theta$ (and respectively for correlation $\mathcal{C}$ ). + +With this interpretation at hand, we design the setup of Figure 5(c) to emulate a spatially-incoherent source suitable for SWI through time-division multiplexing. We use a galvo mirror to steer a narrow collimated beam of narrow-linewidth dichromatic illumination. We focus this beam through a relay lens at a point on the focal plane of the illumination lens. As we steer the beam direction, the focus point scans an area on the focal plane. In turn, each point location results in the injection of illumination at an angle $\theta$ in the interferometer. By scanning the focus point over a target effective emission area within a single exposure, we acquire image and correlation measurements that equal those of Equation (7). As this design sweeps the illumination angle, we term it swept-angle illumination, and its combination with full-field SWI swept-angle synthetic wavelength interferometry. + +Comparison to Kotwal et al. [41]. The illumination module of the proposed setup in Figure 5(c) is an adaptation of the Fourier-domain redistributive projector of Kotwal et al. [41]. Compared to their setup, we do not need to use an amplitude electro-optic modulator, as we are only interested in emulating spatial incoherence (equivalent, directly probing). Additionally, their setup was designed for use with monochromatic illumination, to enable different light transport probing capabilities. By contrast, we use our setup with narrow-linewidth dichromatic illumination, to enable synthetic wavelength interferometry. More broadly, + +our key insight is that swept-angle illumination can emulate extended-area emitters with arbitrary spectral profiles. + +# 5. The $\{M,N\}$ -shift phase retrieval algorithm + +We discuss our pipeline for acquiring and processing measurements to estimate the envelope phase, and thus depth. For this, in Figure 5(c), we use a nanometer-accuracy stage to vary the position $l$ of the reference mirror. + +Phase retrieval. We use the $N$ -shift phase retrieval algorithm [9, 63, 64] to compute the envelope phase: We assume we have estimates of the envelope's squared amplitude $|\pmb{\mathcal{E}}(x,l_n)|^2$ at reference positions $l_{n}$ corresponding to shifts by $N$ -th fractions of the synthetic half-wavelength $\lambda_s / 2$ , $l_{n} = l + n\lambda_{s} / 2N$ , $n\in \{0,\dots ,N - 1\}$ . Then, from Equation (6), we estimate the envelope phase as + +$$ +\phi (d (x)) = \arctan \left[ \frac {\sum_ {n = 0} ^ {N - 1} \left| \boldsymbol {\mathcal {E}} (x , l _ {n}) \right| ^ {2} \sin \left(2 \pi n / N\right)}{\sum_ {n = 0} ^ {N - 1} \left| \boldsymbol {\mathcal {E}} (x , l _ {n}) \right| ^ {2} \cos \left(2 \pi n / N\right)} \right], \tag {8} +$$ + +and the depth (up to an integer multiple of $\lambda_{s} / 2$ ) as + +$$ +d (x) = l + \frac {\phi (d (x))}{2 \kappa_ {s}}. \tag {9} +$$ + +Envelope estimation. Equation (8) requires estimates of $|\mathcal{E}(x,l_n)|^2$ . For each $l_{n}$ , we capture intensity measurements $I(x,l_n^m)$ at $M$ reference positions corresponding to shifts by $M$ -th fractions of the carrier wavelength, $l_{n}^{m} = l_{n} + m\lambda_{c} / M$ , $m \in \{0,\dots,M - 1\}$ . As the carrier wavelength $\lambda_{c}$ is half the optical wavelength $\lambda$ and orders of magnitude smaller than the synthetic wavelength $\lambda_{s}$ , the envelope and interference-free image remain approximately constant across shifts, $\mathcal{E}(x,l_n^m) \approx \mathcal{E}(x,l_n)$ , $\mathcal{I}(x,l_n^m) \approx \mathcal{I}(x,l_n)$ . Then, from Equations (1) and (4), + +$$ +I (x, l _ {n} ^ {m}) = \boldsymbol {\mathcal {T}} (x, l _ {n}) + 2 \sin \left(\kappa_ {c} \left(d (x) - l _ {n} ^ {m}\right)\right) \boldsymbol {\mathcal {E}} (x, l _ {n}). \tag {10} +$$ + +We estimate the interference-free image and envelope as + +$$ +\mathcal {I} (x, l _ {n}) = ^ {1} / M \sum_ {m = 0} ^ {M - 1} I (x, l _ {n} ^ {m}), \tag {11} +$$ + +$$ +\left| \boldsymbol {\mathcal {E}} \left(x, l _ {n}\right) \right| ^ {2} = 1 / _ {2 M} \sum_ {m = 0} ^ {M - 1} \left(I \left(x, l _ {n} ^ {m}\right) - \boldsymbol {\mathcal {I}} \left(x, l _ {n}\right)\right) ^ {2}. \tag {12} +$$ + +Equations (11), (12), (8), and (9) complete our phase estimation pipeline, which we summarize in Figure 6. We note that the runtime of this pipeline is negligible relative to acquisition time. We provide algorithmic details and pseudocode in the supplement. We call this pipeline the $\{M,N\}$ -shift phase retrieval algorithm. The parameters $M$ and $N$ are design parameters that we can fine-tune to control a trade-off between acquisition time and depth accuracy: As we increase $M$ and $N$ , the number $M\cdot N$ of images we capture also increases, and depth estimates become more robust to noise. The theoretical minimum number of images is achieved using $\{3,3\}$ shifts (9 images). In practice, we found $\{4,4\}$ shifts to be robust. We evaluate different $\{M,N\}$ combinations in the supplement. + +Dealing with speckle. Interference in non-specular scenes takes the form of speckle, a high-frequency pseudo-random pattern (Figure 6(a)). This can result in noisy envelope, + +![](images/0d12483adad0e5981445b2adc67e5aa2fc03c1b62909219ff2d29932cb30e341.jpg) +(a) with swept-angle source + +![](images/6a06672c511b5a7b2d25f3dd120bc7af3e7ed40e57f9b96c5010985314f008ec.jpg) +(b) without swept-angle source + +![](images/1e4cf0d27b7f3dbb7fe55e84c4354910dd7d8f07ec5ac4f6b2d0fc8e0d72734b.jpg) +Figure 7. Effect of swept-angle scanning. Swept-angle illumination helps mitigate speckle noise and subsurface scattering effects. + +![](images/d97392e45807ebb0e3ccc33cdc0edc06eb63038680deb24ae2d1852775a77c9e.jpg) + +and thus phase and depth, estimates. As Figures 4(d) and 7 show, swept-angle illumination greatly mitigates speckle effects. We can further reduce the impact of speckle by denoising estimated quantities with a low-pass filter (e.g., Gaussian) [20, 41]. Alternatively, to avoid blurring image details, we can use joint bilateral filtering [12, 59] with a guide image of the scene under ambient light. Empirically, we found it better to blur the envelope estimates of Equation (12) before using them in Equation (8) (Figure 6(d)), compared to blurring the phase or depth estimates. + +# 6. Experiments + +Experimental prototype. For all our experiments, we use the experimental prototype in Figure 5(c). Our prototype comprises: two distributed-Bragg-reflector lasers (wavelengths $780\mathrm{nm}$ and $781\mathrm{nm}$ , power $45\mathrm{mW}$ , linewidth $1\mathrm{MHz}$ ); a galvo mirror pair; a translation stage (resolution $10\mathrm{nm}$ ); two compound macro lenses (focal length $200\mathrm{mm}$ ); a CCD sensor (pixel pitch $3.7\mu \mathrm{m}$ , $3400\times 2700$ pixels); and other free-space optics and mounts. We provide a detailed component list and specifications, calibration instructions, and other information in the supplement. + +We use the following experimental specifications: The reproduction ratio is 1:1, the field of view is $12.5\mathrm{mm}\times 10\mathrm{mm}$ and the working distance is $400\mathrm{mm}$ . The unambiguous depth range is approximately $500\mu \mathrm{m}$ . We use $\{4,4\}$ -shifts (16 images), and a minimum per-image exposure time of $10\mathrm{ms}$ , resulting at a frame rate of $5\mathrm{Hz}$ . + +Depth recovery on challenging scenes. Figures 8 and 2 show scans of many challenging scenes. We scan materials ranging from rough metallic (coins), to diffuse (music box, pill), to highly-scattering (soap and chocolate). Most scenes include fine features requiring high lateral and axial resolution (music box, business card, US quarter, dollar bill). We compare scan results using swept-angle SWI (with bilateral or Gaussian filtering), versus conventional full-field SWI (Figure 5(a), emulated by deactivating the angle-sweep galvo). In all scenes, using swept-angle illumination greatly improves reconstruction quality. The difference is more pronounced in scenes with strong subsurface scattering (music box, chocolate, soap). Even in metallic scenes where there is no subsurface scattering (coins), the use of swept-angle illumination still improves reconstruction quality, by helping mitigate aberration artifacts. Bilateral filtering further improves lateral detail compared to Gaussian filtering. + +Comparison with scanning SWI. A direct comparison of + +![](images/130d8b8ce6a1503c46e48a29387c13cd729fecada546014e99fd33934182c5d7.jpg) +Figure 8. Depth reconstruction. Depth maps (left) and surface renderings (right) acquired using full-field SWI with: (c) swept-angle scanning and bilateral filtering; (d) swept-angle scanning and Gaussian filtering; (e) no swept-angle scanning and with Gaussian filtering. + +swept-angle SWI with scanning SWI is challenging, because of the differences between scanning and full-field setups (Figure 5, (b) versus (a), (c)). To qualitatively assess their relative performance, in Figure 9 we use the following experimental protocol: We downsample the depth map from our technique, by a factor of 35 in each dimension. This is approximately equal to the number of depths points a scanning SWI system operating in point-to-point mode would acquire at the same total exposure time as ours, when equipped with beam-steering optics that operate at a $30\mathrm{kHz}$ scan rate—we detail the exact calculation in the supplement. We then use joint bilateral upsampling [40] with the same guide image as in our technique, to upsample the downsampled depth map back to its original resolution. We observe that, due to the sparse set of points scanning SWI can acquire, the reconstructions miss fine features such as the hair on the quarter and letters on the business card. + +Scanning SWI systems can also operate in resonant mode, typically using Lissajous scanning to raster scan the field of + +view [47]. Resonant mode enables much faster operation than the point-to-point mode we compare against. However, Lissajous scanning results in non-rectilinear sampling of the image plane, as scan lines smear across multiple image rows (when the fast axis is horizontal). At high magnifications, this produces strong spatial distortions and artifacts, making resonant mode operation unsuitable for applications that require micrometer lateral resolution. Our technique also uses resonant-mode Lissajous scanning, but to raster scan the focal plane of the illumination lens (Figure 5(d)), and not the image plane. Instead, our technique uses a two-dimensional sensor to sample the image plane, thus completely avoiding the artifacts above and achieving micrometer lateral resolution. The supplement discusses other challenges in achieving high lateral resolution with scanning SWI. + +Additionally, our technique is orders of magnitude faster than resonant-mode scanning SWI at the same lateral resolution. This is because, whereas scanning SWI must scan the image plane at the target (pixel-level) lateral resolution, our + +![](images/149abe0ca9835118ada176022492b1bf15a4a621b8badd2d710de1ae592b4cc8.jpg) +(a) full-field SWI (ours) (b) scanning+upsampling +Figure 9. Comparison with upsampled scanning SWI. We emulate scanning SWI by downsampling our swept-angle SWI depth. We then bilaterally upsample it to the original resolution. + +Table 2. Depth accuracy. MedAE is the median absolute error between ground truth and estimated depth. Kernel width is the lateral size of the speckle blur filter. All quantities are in $\mu$ m. + +
kernel widthwith swept-anglew/o swept-angle
RMSEMedAERMSEMedAE
78.24.818.913.2
155.13.611.29.5
212.01.610.57.3
301.61.011.16.7
+ +technique remains effective even when scanning the focal plane at much lower resolutions. Figure 10 shows images of the focal plane during scanning, taken with the measurement camera. The dark regions (insets) show that scanning resolution is much lower than pixel-level resolution. + +Axial resolution. To quantify the axial resolution of our technique, we perform the following experiment. We use a second nanometer-accurate translation stage to place the chocolate scene from Figure 8 at different depths from the camera, at increments of $1\mu \mathrm{m}$ . We choose this scene because it is has strong sub-surface scattering. We then compare how well full-field SWI with and without swept-angle illumination can track the scene depth. We perform this experiment using Gaussian filtering with different kernel sizes, to additionally quantify lateral resolution. The results in Table 2 show that we can achieve an axial resolution of approximately $5\mu \mathrm{m}$ and $1\mu \mathrm{m}$ , for kernel sizes $7\mu \mathrm{m}$ and $30\mu \mathrm{m}$ , respectively. We can trade off lateral for axial resolution, by increasing kernel size. Figure 8 shows that this trade-off can become more favorable with bilateral filtering. + +Additional experiments. In the supplement: 1) Comparisons with full-field OCT. We show that swept-angle SWI can approximate the reconstruction quality of OCT, despite being $50 \times$ faster. 2) Tunable depth range. We show that, by adjusting the wavelength separation between the two lasers, swept-angle SWI can scan scenes with depth range $16\mathrm{mm}$ at resolution $50\mu \mathrm{m}$ , at significantly better quality than con + +![](images/cbd3fffb0ba0eb3e5b642bcf93fde51434095b975a64d6e67461378a4169e091.jpg) +Figure 10. Lissajous scanning. Image and insets of the illumination lens' focal plane, showing the Lissajous scanning pattern. + +ventional full-field SWI. 3) Robustness to ambient lighting. We show that the use of near-monochromatic illumination makes swept-angle SWI robust to strong ambient light, even at $10\%$ signal-to-background ratios. 4) Effect of scanning resolution and $\{M,N\}$ settings. We show that, by adjusting the focal plane scanning resolution and $\{M,N\}$ parameters for phase retrieval, swept-angle SWI can trade off between acquisition time versus reconstruction quality. + +# 7. Limitations and conclusion + +Phase wrapping. SWI determines depth only up to an integer multiple of the synthetic half-wavelength $\lambda_{s} / 2$ . Equivalently, all depth values are wrapped to $[0, \lambda_{s} / 2]$ , even if the depth range is greater. To mitigate phase wrapping and extend unambiguous depth range, it is common to capture measurements at multiple synthetic wavelengths, and use them to unwrap the phase estimate [3,7,8,11]. Theoretically two synthetic wavelengths are enough to uniquely determine depth, but in practice phase unwrapping techniques are very sensitive to noise [25]. Combining phase unwrapping with our technique is an interesting future direction. + +Toward real-time operation. Our prototype acquires measurements at a frame rate of $5\mathrm{Hz}$ , due to the need for $10\mathrm{ms}$ per-frame exposure time. If we use a stronger laser to reduce this time, the main speed bottleneck will be the need to perform phase shifts by physically translating the reference mirror. We can mitigate this bottleneck using faster translation stages (e.g., fast microscopy stages). + +Conclusion. We presented a technique for fast depth sensing at micron-scale lateral and axial resolutions. Our technique, swept-angle synthetic wavelength interferometry, combines the complementary advantages of full-field interferometry (speed, pixel-level lateral resolution) and scanning interferometry (robustness to aberrations and indirect illumination). We demonstrated these advantages by scanning multiple scenes with fine geometry and strong subsurface scattering. We expect our results to motivate applications of swept-angle SWI in areas such as biomedical imaging, and industrial fabrication and inspection (Figure 1). + +Acknowledgments. We thank Sudershan Boovaraghavan, Yuvraj Agrawal, Arpit Agarwal, Wenzhen Yuan from CMU, and Veniamin V. Stryzheus, Brian T. Miller from The Boeing Company, who provided the samples for the experiments in Figure 1. This work was supported by NSF awards 1730147, 2047341, 2008123 (NSF-BSF 2019758), and a Sloan Research Fellowship for Ioannis Gkioulekas. + +# References + +[1] Supreeth Achar, Joseph R Bartels, William L Whittaker, Kiriakos N Kutulakos, and Srinivasa G Narasimhan. Epipolar time-of-flight imaging. ACM TOG, 2017. 3 +[2] Brian Aull. 3d imaging with geiger-mode avalanche photodiodes. Opt. Photon. News, 2005. 2 +[3] Seung-Hwan Baek, Noah Walsh, Ilya Chugunov, Zheng Shi, and Felix Heide. Centimeter-wave free-space neural time-of-flight imaging. ACM Transactions on Graphics (TOG), 2022. 2, 8 +[4] Stephen T. Barnard and William B. Thompson. Disparity analysis of images. IEEE TPAMI, 1980. 2 +[5] Ayush Bhandari, Achuta Kadambi, Refael Whyte, Christopher Barsi, Micha Feigin, Adrian Dorrington, and Ramesh Raskar. Resolving multipath interference in time-of-flight imaging via modulation frequency diversity and sparse regularization. Optics Letters, 2014. 3 +[6] Tongbo Chen, Hans-Peter Seidel, and Hendrik P. A. Lensch. Modulated phase-shifting for 3d scanning. IEEE/CVF CVPR, 2008. 2 +[7] Yeou-Yen Cheng and James C. Wyatt. Two-wavelength phase shifting interferometry. Applied Optics, 1984. 2, 8 +[8] Yeou-Yen Cheng and James C. Wyant. Multiple-wavelength phase-shifting interferometry. Applied Optics, 1985. 2, 8 +[9] Peter de Groot. Phase Shifting Interferometry. 2011. 2, 6 +[10] Peter de Groot and John McGarvey. Chirped synthetic-wavelength interferometry. Optics Letters, 1992. 2 +[11] David Droeschel, Dirk Holz, and Sven Behnke. Multifrequency phase unwrapping for time-of-flight cameras. IEEE/RSJ IROS, 2010. 8 +[12] Elmar Eisemann and Frédo Durand. Flash photography enhancement via intrinsic relighting. ACM transactions on graphics (TOG), 2004. 6 +[13] Adolf F Fercher, Wolfgang Drexler, Christoph K Hitzenberger, and Theo Lasser. Optical coherence tomography-principles and applications. Reports on progress in physics, 2003. 3 +[14] Adolf F. Fercher, H Z. Hu, and U. Vry. Rough surface interferometry with a two-wavelength heterodyne speckle interferometer. Applied Optics, 1985. 2 +[15] Richard Ferriere, Johann Cussey, and John M. Dudley. Time-of-flight range detection using low-frequency intensity modulation of a cw laser diode: application to fiber length measurement. Optical Engineering, 2008. 2 +[16] Angel Flores, Craig Robin, Ann Lanari, and Iyad Dajani. Pseudo-random binary sequence phase modulation for narrow linewidth, kilowatt, monolithic fiber amplifiers. Optics Express, 2014. 2 +[17] Daniel Freedman, Eyal Krupka, Yoni Smolin, Ido Leichter, and Mirko Schmidt. Sra: Fast removal of general multipath for tof sensors, 2014. 3 +[18] Stefan Fuchs. Multipath interference compensation in time-of-flight camera images. ICPR, 2010. 3 +[19] Genevieve Gariepy, Nikola Krstajic, Robert Henderson, Chunyong Li, Robert R Thomson, Gerald S Buller, Barmak Heshmat, Ramesh Raskar, Jonathan Leach, and Daniele Faccio. + +Single-photon sensitive light-in-fight imaging. Nature Comm., 2015. 2 +[20] Ioannis Gkioulekas, Anat Levin, Frédo Durand, and Todd Zickler. Micron-scale light transport decomposition using interferometry. ACM TOG, 2015. 3, 4, 6 +[21] Pavel Grossmann. Depth from focus. PRL, 1987. 2 +[22] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In ICCV, 2019. 2 +[23] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. IEEE/CVF CVPR, 2019. 2 +[24] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. IEEE/CVF CVPR, 2011. 2 +[25] Mohit Gupta, Shree K Nayar, Matthias B Hullin, and Jaime Martin. Phasor imaging: A generalization of correlation-based time-of-flight imaging. ACM TOG, 2015. 3, 8 +[26] Mohit Gupta, Andreas Velten, Shree K. Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM TOG, 2018. 2 +[27] Felipe Gutierrez-Barragan, Syed Azer Reza, Andreas Velten, and Mohit Gupta. Practical coding function design for time-of-flight imaging. IEEE/CVF CVPR, 2019. 2 +[28] Yudeog Han, Joon-Young Lee, and In So Kweon. High quality shape from a single rgb-d image under uncalibrated natural illumination. IEEE ICCV, 2013. 2 +[29] Parameswaran Hariharan. Optical interferometry. Elsevier, 2003. 2 +[30] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. 2004. 2 +[31] Caner Hazirbas, Sebastian Georg Soyer, Maximilian Christian Staab, Laura Leal-Taixe, and Daniel Cremers. Deep depth from focus, 2018. 2 +[32] Felix Heide, Steven Diamond, David B. Lindell, and Gordon Wetzstein. Sub-picosecond photon-efficient 3d imaging using single-photon sensors. Scientific Reports, 2018. 2 +[33] Felix Heide, Matthias B Hullin, James Gregson, and Wolfgang Heidrich. Low-budget transient imaging using photonic mixer devices. ACM TOG, 2013. 2 +[34] Berthold Klaus Paul Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. Technical report, 1970. 2 +[35] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, and James G Fujimoto. Optical coherence tomography. Science, 1991. 3 +[36] David Jimeneza, Daniel Pizarrob, Manuel Mazoa, and Sira Palazuelos. Modeling and correction of multipath interference in time of flight cameras. Image and Vision Computing, 2014. 3 +[37] Jon L. Johnson, Timothy D. Dorney, and Daniel M. Mittleman. Enhanced depth resolution in terahertz imaging using phase-shift interferometry. Applied Physics Letters, 2001. 2 +[38] Achuta Kadambi, Refael Whyte, Ayush Bhandari, Lee Streeter, Christopher Barsi, Adrian Dorrington, and Ramesh + +Raskar. Coded Time of Flight Cameras: Sparse Deconvolution to Address Multipath Interference and Recover Time Profiles. ACM TOG, 2013. 3 +[39] Ahmed Kirmani, Dheera Venkatraman, Dongeek Shin, Andrea Colaço, Franco N. C. Wong, Jeffrey H. Shapiro, and Vivek K Goyal. First-photon imaging. Science, 2014. 2 +[40] Johannes Kopf, Michael F. Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Trans. Graph., 26(3):96-es, jul 2007. 7 +[41] Alankar Kotwal, Anat Levin, and Ioannis Gkioulekas. Interferometric transmission probing with coded mutual intensity. ACM TOG, 2020. 3, 5, 6 +[42] Robert Lange and Peter Seitz. Solid-state time-of-flight range camera. IEEE JQE, 2001. 2 +[43] Robert Lange, Peter Seitz, Alice Biber, and Stefan Lauxtermann. Demodulation pixels in ccd and cmos technologies. SPIE, 2000. 2 +[44] Fengqiang Li, Florian Willomitzer, Prasanna Rangarajan, Mohit Gupta, Andreas Velten, and Oliver Cossairt. Sh-tof: Micro resolution time-of-flight imaging with superheterodyne interferometry. IEEE ICCP, 2018. 2 +[45] Fengqiang Li, Joshua Yablon, Andreas Velten, Mohit Gupta, and Oliver Cossairt. High-depth-resolution range imaging with multiple-wavelength superheterodyne interferometry using 1550-nm lasers. Applied Optics, 2017. 2 +[46] David B Lindell, Matthew O'Toole, and Gordon Wetzstein. Single-photon 3d imaging with deep sensor fusion. ACM TOG, 2018. 2 +[47] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 2021. 7 +[48] Julio Marco, Quercus Hernandez, Adolfo Muñoz, Yue Dong, Adrian Jarabo, Min H. Kim, Xin Tong, and Diego Gutierrez. Deeptof: Off-the-shelf real-time correction of multipath interference in time-of-flight imaging. ACM TOG, 2017. 3 +[49] Nikhil Naik, Achuta Kadambi, Christoph Rhemann, Shahram Izadi, Ramesh Raskar, and Sing Bing Kang. A light transport model for mitigating multipath interference in time-of-flight sensors. IEEE/CVF CVPR, 2015. 3 +[50] Lazaros Nalpantidis, Georgios Sirakoulis, and Antonios Gasteratos. Review of stereo vision algorithms: From software to hardware. Int. J. Optomechatronics, 2008. 2 +[51] Shree Nayar, Gurunandan Krishnan, Michael Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. ACM TOG, 2006. 3 +[52] Cristiano Niclass, Alexis Rochas, Pierre. Besse, and Edoardo Charbon. Design and characterization of a cmos 3-d image sensor based on single photon avalanche diodes. IEEE JSSC, 2005. 2 +[53] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM TOG, 2015. 3 +[54] Matthew O'Toole, Felix Heide, David B Lindell, Kai Zang, Steven Diamond, and Gordon Wetzstein. Reconstructing transient images from single-photon sensors. IEEE/CVF CVPR, 2017. 2 + +[55] Matthew O'Toole, Felix Heide, Lei Xiao, Matthias B Hullin, Wolfgang Heidrich, and Kiriakos N Kutulakos. Temporal Frequency Probing for 5D Transient Analysis of Global Light Transport. ACM TOG, 2014. 3 +[56] Matthew O'Toole, John Mather, and Kiriakos N. Kutulakos. 3d shape and indirect appearance by structured light transport. IEEE TPAMI, 2016. 2, 3 +[57] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual Coding to Probe Light Transport. ACM TOG, 2012. 3 +[58] Andrew Payne, A Dorrington, and Michael Cree. Illumination waveform optimization for time-of-flight range imaging cameras. SPIE, 2011. 2 +[59] Georg Petschnigg, Richard Szeliski, Maneesh Agrawala, Michael Cohen, Hugues Hoppe, and Kentaro Toyama. Digital photography with flash and no-flash image pairs. ACM transactions on graphics (TOG), 2004. 6 +[60] Dario Piatti, Fabio Remondino, and David Stoppa. State-of-the-Art of TOF Range-Imaging Sensors. 2013. 2 +[61] Dikpal Reddy, Ravi Ramamoorthi, and Brian Curless. Frequency-space Decomposition and Acquisition of Light Transport Under Spatially Varying Illumination. ECCV, 2012. 3 +[62] Alexis Rochas, Michael Gosch, Alexandre Serov, Pierre Besse, Rade S. Popovic, T. Lasser, and Rudolph Rigler. First fully integrated 2-d array of single-photon detectors in standard cmos technology. IEEE PTL, 2003. 2 +[63] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. Applied optics, 36(34):8871–8876, 1997. 6 +[64] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. ii. a generalized solution. Applied optics, 37(31):7253-7258, 1998. 6 +[65] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. IEEE/CVF CVPR, 2003. 2 +[66] Rudolf Schwarte, Zhanping Xu, Horst-Guenther Heinol, Joachim Olk, Ruediger Klein, Bernd Buxbaum, Helmut Fischer, and Juergen Schulte. New electro-optical mixing and correlating sensor: facilities and applications of the photonic mixer device (PMD). In Sensors, Sensor Systems, and Sensor Data Processing, 1997. 2 +[67] Pitchaya Sitthi-Amorn, Javier E Ramos, Yuwang Wangy, Joyce Kwan, Justin Lan, Wenshou Wang, and Wojciech Matusik. Multifab: a machine vision assisted platform for multi-material 3d printing. Acm Transactions on Graphics (Tog), 2015. 2 +[68] Shuochen Su, Felix Heide, Gordon Wetzstein, and Wolfgang Heidrich. Deep end-to-end time-of-flight imaging. CVPR, 2018. 3 +[69] Murali Subbarao and Gopal Surya. Depth from defocus: A spatial domain approach. IJCV, 1994. 2 +[70] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G. Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature Comm., 2012. 2 + +[71] Federica Villa, Rudi Lussana, Danilo Bronzi, Simone Tisa, Alberto Tosi, Franco Zappa, Alberto Dalla Mora, Davide Contini, Daniel Durini, Sasha Weyers, and Werner Brockherde. Cmos imager with 1024 spads and tdcs for single-photon timing and 3-d time-of-flight. IEEE JSTQE, 2014. 2 +[72] Peng Xiao, Mathias Fink, and A Claude Boccara. Full-field spatially incoherent illumination interferometry: a spatial resolution almost insensitive to aberrations. Optics letters, 2016. 4 \ No newline at end of file diff --git a/2023/Swept-Angle Synthetic Wavelength Interferometry/images.zip b/2023/Swept-Angle Synthetic Wavelength Interferometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a794e8e903e809b9c8e38696ee9ac3b62475f127 --- /dev/null +++ b/2023/Swept-Angle Synthetic Wavelength Interferometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0546144193e4c2bccf25861c4f325549005ff2e6d6c380e900fc36cecfe710f +size 855605 diff --git a/2023/Swept-Angle Synthetic Wavelength Interferometry/layout.json b/2023/Swept-Angle Synthetic Wavelength Interferometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b586b7bf011b4b42de4dcc933977b2fdb69abf3e --- /dev/null +++ b/2023/Swept-Angle Synthetic Wavelength Interferometry/layout.json @@ -0,0 +1,10820 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 452, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 452, + 120 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 452, + 120 + ], + "type": "text", + "content": "Swept-Angle Synthetic Wavelength Interferometry" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 124, + 431, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 124, + 431, + 138 + ], + "spans": [ + { + "bbox": [ + 160, + 124, + 431, + 138 + ], + "type": "text", + "content": "Alankar Kotwal1, Anat Levin2, and Ioannis Gkioulekas1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 139, + 392, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 139, + 392, + 152 + ], + "spans": [ + { + "bbox": [ + 198, + 139, + 392, + 152 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 139, + 392, + 152 + ], + "type": "text", + "content": "Carnegie Mellon University, " + }, + { + "bbox": [ + 198, + 139, + 392, + 152 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 139, + 392, + 152 + ], + "type": "text", + "content": "Technion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 180, + 192, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 180, + 192, + 193 + ], + "spans": [ + { + "bbox": [ + 143, + 180, + 192, + 193 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "type": "text", + "content": "We present a new imaging technique, swept-angle synthetic wavelength interferometry, for full-field micron-scale 3D sensing. As in conventional synthetic wavelength interferometry, our technique uses light consisting of two narrowly-separated optical wavelengths, resulting in per-pixel interferometric measurements whose phase encodes scene depth. Our technique additionally uses a new type of light source that, by emulating spatially-incoherent illumination, makes interferometric measurements insensitive to aberrations and (sub)surface scattering, effects that corrupt phase measurements. The resulting technique combines the robustness to such corruptions of scanning interferometric setups, with the speed of full-field interferometric setups. Overall, our technique can recover full-frame depth at a lateral and axial resolution of " + }, + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "type": "inline_equation", + "content": "5\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "type": "text", + "content": ", at frame rates of " + }, + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 196, + 290, + 435 + ], + "type": "text", + "content": ", even under strong ambient light. We build an experimental prototype, and use it to demonstrate these capabilities by scanning a variety of objects, including objects representative of applications in inspection and fabrication, and objects that contain challenging light scattering effects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 447, + 127, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 127, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 127, + 459 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 581 + ], + "type": "text", + "content": "Depth sensing is among the core problems of computer vision and computational imaging, with widespread applications in medicine, industry, and robotics. An array of techniques is available for acquiring depth maps of three-dimensional (3D) scenes at different scales. In particular, micrometer-resolution depth sensing, our focus in this paper, is important in biomedical imaging because biological features are often micron-scale, industrial fabrication and inspection of critical parts that must conform to their specifications (Figure 1), and robotics to handle fine objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "type": "text", + "content": "Active illumination depth sensing techniques such as lidar, structured light, and correlation time-of-flight (ToF) cannot provide micrometer axial resolution. Instead, we focus on interferometric techniques that can achieve such resolutions. The operational principles and characteristics of interferometric techniques vary, depending on the type of active illumination and optical configurations they use." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": "The choice of illumination spectrum leads to techniques such as optical coherence tomography (OCT), which uses broadband illumination, and phase-shifting interferometry (PSI), which uses monochromatic illumination. We consider" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 180, + 426, + 241 + ], + "blocks": [ + { + "bbox": [ + 319, + 180, + 426, + 241 + ], + "lines": [ + { + "bbox": [ + 319, + 180, + 426, + 241 + ], + "spans": [ + { + "bbox": [ + 319, + 180, + 426, + 241 + ], + "type": "image", + "image_path": "5b4d699b630e04c539ae5b1f410ca5570181d27d53a7a94b715398926d13c97d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 242, + 434, + 283 + ], + "blocks": [ + { + "bbox": [ + 320, + 242, + 434, + 283 + ], + "lines": [ + { + "bbox": [ + 320, + 242, + 434, + 283 + ], + "spans": [ + { + "bbox": [ + 320, + 242, + 434, + 283 + ], + "type": "image", + "image_path": "c5be579f24c1237fcad3248b622c95333103d0944a395421fa5ec0f1280134dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 284, + 434, + 325 + ], + "blocks": [ + { + "bbox": [ + 320, + 284, + 434, + 325 + ], + "lines": [ + { + "bbox": [ + 320, + 284, + 434, + 325 + ], + "spans": [ + { + "bbox": [ + 320, + 284, + 434, + 325 + ], + "type": "image", + "image_path": "d6d086e19473fb4a9da562c40f8035887fe782bb75bfea229237d6c912f7e597.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 326, + 432, + 342 + ], + "lines": [ + { + "bbox": [ + 340, + 326, + 432, + 342 + ], + "spans": [ + { + "bbox": [ + 340, + 326, + 432, + 342 + ], + "type": "text", + "content": "photo of image of scanned object scanned scene" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 180, + 531, + 240 + ], + "blocks": [ + { + "bbox": [ + 434, + 180, + 531, + 240 + ], + "lines": [ + { + "bbox": [ + 434, + 180, + 531, + 240 + ], + "spans": [ + { + "bbox": [ + 434, + 180, + 531, + 240 + ], + "type": "image", + "image_path": "2c1056f30b12d514f52ae5ea614b628718903c6adaa28822ec1adb484a28627d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 434, + 242, + 531, + 283 + ], + "blocks": [ + { + "bbox": [ + 434, + 242, + 531, + 283 + ], + "lines": [ + { + "bbox": [ + 434, + 242, + 531, + 283 + ], + "spans": [ + { + "bbox": [ + 434, + 242, + 531, + 283 + ], + "type": "image", + "image_path": "26bd931f927f7d2b6a1b351aed415f6638152ac1ab1ffae1917c23954fdfa609.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 434, + 284, + 531, + 325 + ], + "blocks": [ + { + "bbox": [ + 434, + 284, + 531, + 325 + ], + "lines": [ + { + "bbox": [ + 434, + 284, + 531, + 325 + ], + "spans": [ + { + "bbox": [ + 434, + 284, + 531, + 325 + ], + "type": "image", + "image_path": "21190cf1af0d97804549dd211292cad8f9c61719e91463d9f98215a904903aa0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 326, + 481, + 342 + ], + "lines": [ + { + "bbox": [ + 449, + 326, + 481, + 342 + ], + "spans": [ + { + "bbox": [ + 449, + 326, + 481, + 342 + ], + "type": "text", + "content": "ept-angle (ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 326, + 524, + 342 + ], + "lines": [ + { + "bbox": [ + 487, + 326, + 524, + 342 + ], + "spans": [ + { + "bbox": [ + 487, + 326, + 524, + 342 + ], + "type": "text", + "content": "without swept-angle" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 343, + 547, + 386 + ], + "lines": [ + { + "bbox": [ + 305, + 343, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 547, + 386 + ], + "type": "text", + "content": "Figure 1. Applications of swept-angle SWI in industrial inspection and fabrication. We show depth reconstructions for two scenes representative of these applications: millimeter-scale dents on an aircraft fuselage section, and a 3D-printed coin." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 390, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 473 + ], + "type": "text", + "content": "synthetic wavelength interferometry (SWI), which operates between these two extremes: By using illumination consisting of two narrowly-separated optical wavelengths, SWI provides a controllable trade-off between the large unambiguous depth range of OCT, and the large axial resolution of PSI. SWI can achieve micrometer resolution at depth ranges in the order of hundreds of micrometers." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 474, + 548, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 665 + ], + "type": "text", + "content": "The choice of optical configuration results in full-field versus scanning implementations, which offer different trade-offs. Full-field implementations acquire entire 2D depth maps, offering simultaneously fast operation and pixel-level lateral resolutions. However, full-field implementations are very sensitive to effects that corrupt depth estimation, such as imperfections in free-space optics (e.g., lens aberrations) and indirect illumination (e.g., subsurface scattering). By contrast, scanning implementations use beam steering to sequentially scan points in a scene and produce a 2D depth map. Scanning implementations offer robustness to depth corruption effects, through the use of fiber optics to reduce aberrations, and co-axial illumination and sensing to eliminate most indirect illumination. However, scanning makes acquiring depth maps with pixel-level lateral resolution and megapixel sizes impractically slow." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We develop a 3D sensing technique, swept-angle synthetic wavelength interferometry, that combines the complementary advantages of full-field and scanning implementations. We draw inspiration from previous work showing" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8233" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 127, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 127, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 127, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 127, + 144 + ], + "type": "image", + "image_path": "cafa133a664e0212e031a64c142df4a37ceac0e099ab31ed85fc41d0d377b1f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 145, + 99, + 154 + ], + "lines": [ + { + "bbox": [ + 86, + 145, + 99, + 154 + ], + "spans": [ + { + "bbox": [ + 86, + 145, + 99, + 154 + ], + "type": "text", + "content": "bill" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 222, + 144 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 222, + 144 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 222, + 144 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 222, + 144 + ], + "type": "image", + "image_path": "55e57b66b4c1dc777942244a0845adc0c73ce87365a819e75e2f92c12c7ba357.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 146, + 185, + 155 + ], + "lines": [ + { + "bbox": [ + 164, + 146, + 185, + 155 + ], + "spans": [ + { + "bbox": [ + 164, + 146, + 185, + 155 + ], + "type": "text", + "content": "scene" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 316, + 144 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 316, + 144 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 316, + 144 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 316, + 144 + ], + "type": "image", + "image_path": "a7296b1c72c1e129ce8d69b7364b3b4906b44bf053bcdf5d6bc1974a715918e6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 146, + 291, + 155 + ], + "lines": [ + { + "bbox": [ + 248, + 146, + 291, + 155 + ], + "spans": [ + { + "bbox": [ + 248, + 146, + 291, + 155 + ], + "type": "text", + "content": "input image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 155, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 46, + 155, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 545, + 177 + ], + "type": "text", + "content": "Figure 2. Reconstructing the eagle embossed on a $20 bill. The features on the eagle are raised " + }, + { + "bbox": [ + 46, + 155, + 545, + 177 + ], + "type": "inline_equation", + "content": "10\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 155, + 545, + 177 + ], + "type": "text", + "content": " off the surface of the bill. The recovered depth shows fine details such as the gaps between the wings reconstructed with high lateral and axial resolution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 411, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 411, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 411, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 411, + 144 + ], + "type": "image", + "image_path": "2ab74e865e7a658ed19f3de3ccc2776691427057fd72dad222da505fc34465c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 146, + 389, + 155 + ], + "lines": [ + { + "bbox": [ + 341, + 146, + 389, + 155 + ], + "spans": [ + { + "bbox": [ + 341, + 146, + 389, + 155 + ], + "type": "text", + "content": "depth surface" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 412, + 70, + 529, + 144 + ], + "blocks": [ + { + "bbox": [ + 412, + 70, + 529, + 144 + ], + "lines": [ + { + "bbox": [ + 412, + 70, + 529, + 144 + ], + "spans": [ + { + "bbox": [ + 412, + 70, + 529, + 144 + ], + "type": "image", + "image_path": "66bdf49815e1cd2fa80dfe60d47b0e3d048b38a80eb3932340e5622e7f22350b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 146, + 477, + 156 + ], + "lines": [ + { + "bbox": [ + 441, + 146, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 441, + 146, + 477, + 156 + ], + "type": "text", + "content": "depth map" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "text", + "content": "that the use of spatially-incoherent illumination in full-field interferometry mitigates aberration and indirect illumination effects. We design a light source that emulates spatial incoherence, by scanning within exposure a dichromatic point source at the focal plane of the illumination lens, effectively sweeping the illumination angle—hence swept-angle. We combine this light source with full-field SWI, resulting in a 3D sensing technique with the following performance characteristics: " + }, + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "text", + "content": " full-frame 2 Mpixel acquisition; " + }, + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "inline_equation", + "content": "5\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "text", + "content": " lateral and axial resolution; range and resolution tunability; and robustness to aberrations, ambient illumination, and indirect illumination. We build an experimental prototype, and use it to demonstrate these capabilities, as in Figure 2. We provide setup details, reconstruction code, and data in the supplement and project website." + }, + { + "bbox": [ + 46, + 181, + 289, + 361 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 361, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 624 + ], + "type": "text", + "content": "Potential impact. Swept-angle SWI is relevant for critical applications, including industrial fabrication and inspection. In industrial fabrication, swept-angle SWI can be used to provide feedback during additive and subtractive manufacturing processes [67]. In industrial inspection, swept-angle SWI can be used to examine newly-fabricated or used in-place critical parts and ensure they comply with operational specifications. Swept-angle SWI, uniquely among 3D scanning techniques, offers a combination of features that are critical for both applications: First, high acquisition speed, which is necessary to avoid slowing down the manufacturing process, and to perform inspection efficiently. Second, micrometer lateral and axial resolution, which is necessary to detect critical defects. Third, robustness to aberrations and indirect illumination, which is necessary because of the materials used for fabrication, which often have strong subsurface scattering. Figure 1 showcases results representative of these applications: We scan a fuselage section from a Boeing aircraft, to detect critical defects such as scratches and bumps from collisions, at axial and lateral scales of a couple dozen micrometers. We also scan a coin pattern 3D-printed by a commercial material printer on a translucent material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 627, + 132, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 132, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 132, + 639 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 642, + 288, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 690 + ], + "type": "text", + "content": "Depth sensing. There are many technologies for acquiring depth in computer vision. Passive techniques rely on scene appearance under ambient light, and use cues such as stereo [4,30,50], (de)focus [21,31,69], or shading [28,34]." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 201, + 535, + 310 + ], + "blocks": [ + { + "bbox": [ + 305, + 179, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 179, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 545, + 201 + ], + "type": "text", + "content": "Table 1. Comparison of interferometric depth sensing techniques for millimeter-scale scenes. ('f.f': full-field; 'scan': scanning)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 201, + 535, + 310 + ], + "lines": [ + { + "bbox": [ + 318, + 201, + 535, + 310 + ], + "spans": [ + { + "bbox": [ + 318, + 201, + 535, + 310 + ], + "type": "table", + "html": "
methodaxial reslateral resdepth rangeacq timerobust
f.f. TD-OCT✓✓✓✗✓
scan. FD-OCT✗✓✓✗✓
scan. SS-OCT✗✓✓✗✓
f.f. PSI✓✓✗✓✗
scan. PSI✗✓✗✗✓
f.f. SWI✓✓✓✓✗
scan. SWI✗✓✓✗✓
f.f. SA-SWI✓✓✓✓✓
", + "image_path": "7b6369174956027eea477a09f471a29942c86557e02f146f4e430eec2804d028.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 313, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 547, + 409 + ], + "type": "text", + "content": "Active techniques inject controlled light in the scene, to overcome issues such as lack of texture and limited resolution. These include structured light [6, 24, 56, 65], impulse ToF [2, 19, 22, 23, 32, 39, 46, 52, 54, 62, 70, 71], and correlation ToF [3, 15, 16, 26, 27, 33, 42, 43, 58, 60, 66]. These techniques cannot easily achieve axial resolutions below hundreds of micrometers, placing them out of scope for applications requiring micrometer resolutions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 408, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 492 + ], + "type": "text", + "content": "Optical interferometry. Interferometry is a classic wave-optics technique that measures the correlation, or interference, between two or more light beams that have traveled along different paths [29]. Most relevant to our work are three broad classes of interferometric techniques that provide depth sensing capabilities with different advantages and disadvantages, which we summarize in Table 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 492, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 547, + 601 + ], + "type": "text", + "content": "Phase-shifting interferometry (PSI) techniques [9, 37] use single-frequency illumination for nanometer-scale depth sensing. Similar to correlation ToF, they estimate depth by measuring the phase of a sinusoidal waveform. However, whereas correlation ToF uses external modulation to generate megahertz-frequency waveforms, PSI directly uses the terahertz-frequency light waveform. Unfortunately, the unambiguous depth range of PSI is limited to the illumination wavelength, typically around a micrometer." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 600, + 548, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 548, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 548, + 708 + ], + "type": "text", + "content": "Synthetic wavelength interferometry (SWI) (or heterodyne interferometry) techniques [7, 8, 10, 14, 44, 45] use illumination comprising multiple narrow spectral bands to synthesize waveforms at frequencies between the megahertz rates of correlation ToF and terahertz rates of PSI. These techniques allow control of the trade off between unambiguous depth range and resolution, by tuning the frequency of the synthetic wavelength. As our technique builds upon SWI, we detail its operational principles in Section 3." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 265, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 265, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 265, + 714 + ], + "type": "text", + "content": "1https://imaging.cs.cmu.edu/swept_angle_swi" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8234" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "Lastly, optical coherence tomography (OCT) [20, 35, 41] uses broadband illumination to perform depth sensing analogously to impulse time-of-flight, by processing temporally-resolved (transient) responses. OCT decouples range and resolution, achieving unambiguous depth ranges up to centimeters, at micrometer axial resolutions. This flexibility comes at the cost of increased acquisition time, because of the need for axial (time-domain OCT) or lateral (Fourier-domain and swept-source OCT) scanning [13]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 323 + ], + "type": "text", + "content": "Mitigating indirect illumination effects. In their basic forms, active depth sensing techniques assume the presence of only direct illumination in the scene. Indirect illumination effects, such as interreflections and subsurface scattering, confound depth information. For example, in correlation ToF techniques (including PSI and SWI), indirect illumination results in incorrect phase, and thus depth estimation, an effect also known as multi-path interference. Several techniques exist for computationally mitigating indirect illumination effects, using models of multi-bounce transport [18, 36, 49], sparse reconstruction [17, 38], multiwavelength approaches [5], and neural approaches [48, 68]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 420 + ], + "type": "text", + "content": "Other techniques optically remove indirect illumination by probing light transport [57]. Such techniques include epipolar imaging [53, 56], high-spatial-frequency illumination [51, 61], and spatio-temporal coded illumination [1, 25, 55]. Similar probing capabilities are possible in interferometric systems, by exploiting the spatio-temporal coherence properties of the illumination [20, 41]. We adapt these techniques for robust micrometer-scale depth sensing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 427, + 221, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 221, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 221, + 441 + ], + "type": "text", + "content": "3. Background on interferometry" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 442, + 288, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 288, + 548 + ], + "type": "text", + "content": "The Michelson interferometer. Our optical setup is based on the classical Michelson interferometer (Figure 3(c)). The interferometer uses a beam splitter to divide collimated input illumination into two beams: one propagates toward the scene arm, and another propagates toward the reference arm—typically a planar mirror mounted on a translation stage that can vary the mirror's distance from the beam splitter. After reflection, the two light beams recombine at the beam splitter and propagate toward the sensor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": "We denote by " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " the distance from the beamsplitter of the reference mirror and the scene point that pixel " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " images, respectively. As " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " is a controllable parameter, we denote it explicitly. We denote by " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "\\pmb{u}_r(x,l)" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "\\pmb{u}_s(x)" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " the complex fields arriving at sensor pixel " + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 550, + 288, + 621 + ], + "type": "text", + "content": " from the reference and scene arms respectively. Then, the sensor measures," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 622, + 288, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 288, + 652 + ], + "type": "interline_equation", + "content": "I (x, l) = \\underbrace {\\left| \\boldsymbol {u} _ {s} (x) \\right| ^ {2} + \\left| \\boldsymbol {u} _ {r} (x , l) \\right| ^ {2}} _ {\\equiv \\mathcal {I} (x, l)} + 2 \\operatorname {R e} \\left\\{\\underbrace {\\boldsymbol {u} _ {s} (x) \\boldsymbol {u} _ {r} ^ {*} (x , l)} _ {\\equiv \\mathcal {C} (x, l)} \\right\\}. \\tag {1}", + "image_path": "4cf499f7c985f19a762c8a3f4bc047f43c75e180ffbcb281f464dce069319c2b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "type": "text", + "content": "The first two terms in Equation (1) are the intensities the sensor would measure if it were observing each of the two arms separately. We call their sum the interference-free image " + }, + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "type": "text", + "content": ". The third term, which we call interference, is the real part of the complex correlation " + }, + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 654, + 289, + 713 + ], + "type": "text", + "content": " between the reflected" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 533, + 201 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 533, + 201 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 533, + 201 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 533, + 201 + ], + "type": "image", + "image_path": "f70f35dd8704b3a044561a1dfdf58727b672d1199c01218665a1ddc36d65a995.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 203, + 529, + 264 + ], + "blocks": [ + { + "bbox": [ + 320, + 203, + 529, + 264 + ], + "lines": [ + { + "bbox": [ + 320, + 203, + 529, + 264 + ], + "spans": [ + { + "bbox": [ + 320, + 203, + 529, + 264 + ], + "type": "image", + "image_path": "6c2ec06fa13cc80151c7cc387ce31e7d01a93729ff53d291217c80a17fe14c18.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 265, + 529, + 324 + ], + "blocks": [ + { + "bbox": [ + 321, + 265, + 529, + 324 + ], + "lines": [ + { + "bbox": [ + 321, + 265, + 529, + 324 + ], + "spans": [ + { + "bbox": [ + 321, + 265, + 529, + 324 + ], + "type": "image", + "image_path": "09707a6187575de1a56c8f2e2c73f78a4af31d3f81d12e8f0dea70c8cce139d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "lines": [ + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "type": "text", + "content": "Figure 3. Synthetic wavelength interferometry. (a-b) Collimated illumination from a point source emitting at two narrowly-separated wavelengths is injected into a Michelson interferometer. (c) As the reference mirror position " + }, + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "type": "text", + "content": " is scanned, each wavelength contributes to the interference a sinusoid with period equal to its wavelength. (d) The sum of these sinusoids has an envelope that is another sinusoid at a synthetic wavelength peaked at " + }, + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "type": "inline_equation", + "content": "l = d(x)" + }, + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 405, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 545, + 429 + ], + "type": "text", + "content": "scene and reference fields. We elaborate on how to isolate the interference term from Equation (1) in Section 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "content": "Synthetic wavelength interferometry. SWI uses illumination comprising two distinct, but narrowly-separated, wavelengths that are incoherent with each other. We denote these wavelengths as " + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "inline_equation", + "content": "\\lambda /1 + \\epsilon" + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "content": ", corresponding to wavenumbers " + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "inline_equation", + "content": "\\kappa \\equiv 2\\pi /\\lambda" + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "inline_equation", + "content": "(1 + \\epsilon)\\kappa" + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "content": ", respectively. We assume that this illumination is injected into the interferometer as a collimated beam—for example, created by placing the outputs of two fiber-coupled single-frequency lasers at the focal plane of a lens, as in Figures 3 and 5. Then, we show in the supplement that the correlation " + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(x,l)" + }, + { + "bbox": [ + 304, + 429, + 547, + 548 + ], + "type": "text", + "content": " equals" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 552, + 547, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 552, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 341, + 552, + 547, + 579 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {C} (x, l) = \\exp (- 2 i \\kappa (d (x) - l)) \\\\ [ 1 + \\exp (- 2 i \\kappa \\epsilon (d (x) - l)) ]. \\tag {2} \\\\ \\end{array}", + "image_path": "d8606d68d26e52a9d754315366193ddbec9ceb40a4d06f193f91844783e067b8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 605 + ], + "type": "text", + "content": "The interference component of the camera intensity measurements in Equation (1) equals," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 608, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 545, + 635 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {R e} (\\mathcal {C} (x, l)) = 2 \\sin (\\kappa (2 + \\epsilon) (d (x) - l)) \\sin (\\kappa \\epsilon (d (x) - l)) (3) \\\\ \\approx 2 \\sin (2 \\kappa (d (x) - l)) \\sin (\\kappa \\epsilon (d (x) - l)). (4) \\\\ \\end{array}", + "image_path": "8c892e3071c1d56833b00e4e88d892d0e7e964097ad8594e3c3e014ff1473130.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "content": "The approximation is accurate when " + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "inline_equation", + "content": "\\epsilon \\ll 1" + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "content": ", i.e., when the two wavelengths are close. We observe that, as a function of " + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "inline_equation", + "content": "d(x) - l" + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "content": ", the interference is the product of two sinusoids: first, a carrier sinusoid with carrier wavelength " + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "inline_equation", + "content": "\\lambda_{c} \\equiv \\lambda / 2" + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "content": " and corresponding carrier wavenumber " + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "inline_equation", + "content": "\\kappa_{c} \\equiv 2\\lambda" + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "content": "; second, an envelope sinusoid " + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "text", + "content": " with synthetic wavelength " + }, + { + "bbox": [ + 304, + 639, + 547, + 711 + ], + "type": "inline_equation", + "content": "\\lambda_{s} \\equiv \\lambda / \\epsilon" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8235" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 254, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 254, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 254, + 83 + ], + "type": "text", + "content": "and corresponding synthetic wavenumber " + }, + { + "bbox": [ + 47, + 72, + 254, + 83 + ], + "type": "inline_equation", + "content": "\\kappa_{s} \\equiv \\epsilon \\kappa" + }, + { + "bbox": [ + 47, + 72, + 254, + 83 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 87, + 287, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 87, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 103, + 87, + 287, + 99 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {E}} (x, l) \\equiv \\sin (\\kappa_ {s} (d (x) - l)). \\tag {5}", + "image_path": "b4fd81d3ee1aeb98a9000790dc68764e310b133552422a2abb24277b4b05ad91.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "type": "text", + "content": "Figure 3(d) visualizes " + }, + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "type": "inline_equation", + "content": "\\operatorname{Re}\\{\\mathcal{C}(x, l)\\}" + }, + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(x, l)" + }, + { + "bbox": [ + 46, + 102, + 288, + 126 + ], + "type": "text", + "content": ". In practice, we measure only the squared amplitude of the envelope," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 128, + 287, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 287, + 152 + ], + "type": "interline_equation", + "content": "\\left| \\boldsymbol {\\mathcal {E}} (x, l) \\right| ^ {2} = \\sin^ {2} \\left(\\kappa_ {s} (d (x) - l)\\right) = \\frac {1 - \\cos \\left(2 \\kappa_ {s} (d (x) - l)\\right)}{2}. \\tag {6}", + "image_path": "d98d0daf76321b1e18a31f4bd220994778fd42070f00da0003b5e59a42445dd4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": "From Equation (6), we see that SWI encodes scene depth " + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": " in the phase " + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\phi(d(x)) \\equiv 2\\kappa_s(d(x) - l)" + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": " of the envelope sinusoid. We defer details on how to measure the envelope and estimate this phase until Section 4. We make two observations: First, SWI provides depth measurements at intervals of " + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "inline_equation", + "content": "[0, \\lambda_s / 2]" + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": ", and cannot disambiguate between depths differing by an integer multiple of " + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\lambda_s / 2" + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": ". Second, the use of two wavelengths makes it possible to control the unambiguous depth range: by decreasing the separation " + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 152, + 287, + 282 + ], + "type": "text", + "content": " between the two wavelengths, we increase the unambiguous depth range, at the cost of decreasing depth resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 284, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 343 + ], + "type": "text", + "content": "Full-field and scanning interferometry. Figure 5 shows two types of Michelson interferometer setups that implement SWI: (a) a full-field interferometer; and (b) a scanning interferometer. We discuss their relative merits, which will motivate our proposed swept-angle interferometer setup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": "Full-field interferometers (Figure 5(a)) use free-space optics to illuminate and image the entire field of view in the scene and reference arms. They also use a two-dimensional sensor to measure interference at all locations " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": ". This enables fast interference measurements for all scene points at once, at lateral resolutions equal to the sensor pixel pitch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "text", + "content": "Unfortunately, full-field interferometers are susceptible to phase corruption effects, as we visualize in Figure 4. Equation (5) assumes that the scene field is due to only the direct light path (solid line in Figure 4(a)), which produces a sinusoidal envelope with phase delay " + }, + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "inline_equation", + "content": "d(x)" + }, + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "text", + "content": " (solid orange curve in Figure 4(b)). In practice, the scene field will include contributions from additional paths: First, indirect paths due to subsurface scattering (dashed lines in Figure 4(a)). Second, stray paths due to optical aberrations (Figure 4(c)). These paths have different lengths, and thus contribute to the envelope sinusoidal terms of different phases (dashed curves in Figure 4(b)). Their summation produces an overall sinusoidal envelope (black) with phase " + }, + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "inline_equation", + "content": "d' \\neq d(x)" + }, + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "text", + "content": ", resulting in incorrect depth estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 665 + ], + "type": "text", + "content": "Scanning interferometers (Figure 5(b)) use fiber optics (couplers, circulators, collimators) to generate a focused beam that illuminates only one point in the scene and reference arms. Additionally, they use a single-pixel sensor, focused at the same point. They also use steering optics (e.g., MEMS mirrors) to scan the focus point and capture interference measurements for the entire scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 287, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 702 + ], + "type": "text", + "content": "Scanning interferometers effectively mitigate the phase corruption effects in Figure 4: Because, at any given time, they only illuminate and image one point in the scene, they" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 422, + 129 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 422, + 129 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 422, + 129 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 422, + 129 + ], + "type": "image", + "image_path": "f56d1833565b35556c613a6a6692b69feac8aec3701d65d4eb5d2127662d677a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 129, + 424, + 139 + ], + "lines": [ + { + "bbox": [ + 320, + 129, + 424, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 424, + 139 + ], + "type": "text", + "content": "(a) light paths in scattering slab" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 533, + 129 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 533, + 129 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 533, + 129 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 533, + 129 + ], + "type": "image", + "image_path": "eeb47b533cb97067ac74da62988a98c468c9f8a22062ff4f1a8e945501fc0a4d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 129, + 525, + 138 + ], + "lines": [ + { + "bbox": [ + 437, + 129, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 437, + 129, + 525, + 138 + ], + "type": "text", + "content": "(b) envelope contributions" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 139, + 422, + 192 + ], + "blocks": [ + { + "bbox": [ + 318, + 139, + 422, + 192 + ], + "lines": [ + { + "bbox": [ + 318, + 139, + 422, + 192 + ], + "spans": [ + { + "bbox": [ + 318, + 139, + 422, + 192 + ], + "type": "image", + "image_path": "5291baa79d9de5872302809f755a9860b36bfa84bcc17f5c7be27cdb2b00ce93.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 193, + 421, + 203 + ], + "lines": [ + { + "bbox": [ + 323, + 193, + 421, + 203 + ], + "spans": [ + { + "bbox": [ + 323, + 193, + 421, + 203 + ], + "type": "text", + "content": "(c) stray aberrated light paths" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 424, + 140, + 532, + 193 + ], + "blocks": [ + { + "bbox": [ + 424, + 140, + 532, + 193 + ], + "lines": [ + { + "bbox": [ + 424, + 140, + 532, + 193 + ], + "spans": [ + { + "bbox": [ + 424, + 140, + 532, + 193 + ], + "type": "image", + "image_path": "89cb66496337ccc51bb74f2df4ab79f9c5b19db67bdd33b7ab1157f0bf0dc588.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 193, + 522, + 203 + ], + "lines": [ + { + "bbox": [ + 438, + 193, + 522, + 203 + ], + "spans": [ + { + "bbox": [ + 438, + 193, + 522, + 203 + ], + "type": "text", + "content": "(d) reference arm images" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 203, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 300 + ], + "type": "text", + "content": "Figure 4. Phase corruption effects. (a) A typical scene contains both direct (solid line) and indirect (dashed lines) light paths. (b) The direct path contributes a sinusoid with the correct phase (solid orange line). The indirect paths contribute sinusoids with incorrect phases (dashed lines). Their summation results in erroneous phase estimation (dark solid line). (c) Aberrations in free-space optics result in stray light paths that also contribute incorrect phases. (d) Images of the reference arm (mirror) visualize the effects of aberrations, and their mitigation using swept-angle illumination." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 399 + ], + "type": "text", + "content": "eliminate contributions from indirect paths such as those in Figure 4(a). Additionally, because the use of fiber optics minimizes aberrations, they eliminate stray paths such as those in Figure 4(c). Unfortunately, this robustness comes at the cost of having to use beam steering to scan the entire scene. This translates into long acquisition times, especially when it is necessary to measure depth at pixel-level lateral resolutions and at a sensor-equivalent field of view." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 400, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 448 + ], + "type": "text", + "content": "In the next section, we introduce an interferometer design that combines the fast acquisition and high lateral resolution of full-field interferometers, with the robustness to phase corruption effects of scanning interferometers." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 459, + 449, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 449, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 449, + 472 + ], + "type": "text", + "content": "4. Swept-angle illumination" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 475, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 521 + ], + "type": "text", + "content": "In this section, we design a new light source for use with full-field SWI, to mitigate the phase corruption effects shown in Figure 4—indirect illumination, aberrations. Figure 5(c) shows our final swept-angle optical design." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": "Interferometry with spatially-incoherent illumination. Our starting point is previous findings on the use of spatially-incoherent illumination in full-field interferometric implementations. Specifically, Gkioulekas et al. [20] showed that using spatially incoherent illumination in a Michelson interferometer is equivalent to direct-only probing: this optically rejects indirect paths and makes the correlation term " + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": " depend predominantly on contributions from direct paths. Additionally, Xiao et al. [72] showed that using spatially-incoherent illumination makes the correlation term " + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": " insensitive to aberrations in the free-space optics." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 712 + ], + "type": "text", + "content": "Both Gkioulekas et al. [20] and Xiao et al. [72] realize spatially-incoherent illumination by replacing the point emitter in Figure 3 with an area emitter, e.g., LED or halogen lamp. Unfortunately, fundamental physics dictate that extending emission area of a light source is coupled with" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8236" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 68, + 214, + 217 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 214, + 217 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 214, + 217 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 214, + 217 + ], + "type": "image", + "image_path": "e9016a9ad2e7c2d5e86dd621001ffacf1ccadbcfe455ac1521b341c15642873b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 220, + 173, + 228 + ], + "lines": [ + { + "bbox": [ + 115, + 220, + 173, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 220, + 173, + 228 + ], + "type": "text", + "content": "(a) full-field SWI" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 69, + 343, + 217 + ], + "blocks": [ + { + "bbox": [ + 217, + 69, + 343, + 217 + ], + "lines": [ + { + "bbox": [ + 217, + 69, + 343, + 217 + ], + "spans": [ + { + "bbox": [ + 217, + 69, + 343, + 217 + ], + "type": "image", + "image_path": "9c2607d50d04cb10c8d0fb8d8a043db74318fde433609c539f8cae86c24588d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 220, + 309, + 229 + ], + "lines": [ + { + "bbox": [ + 251, + 220, + 309, + 229 + ], + "spans": [ + { + "bbox": [ + 251, + 220, + 309, + 229 + ], + "type": "text", + "content": "(b) scanning SWI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 68, + 520, + 217 + ], + "blocks": [ + { + "bbox": [ + 345, + 68, + 520, + 217 + ], + "lines": [ + { + "bbox": [ + 345, + 68, + 520, + 217 + ], + "spans": [ + { + "bbox": [ + 345, + 68, + 520, + 217 + ], + "type": "image", + "image_path": "b197703bb937d68effbdace7500b9c0b5329be89e906caed6358b25e87a9c41e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 220, + 482, + 229 + ], + "lines": [ + { + "bbox": [ + 382, + 220, + 482, + 229 + ], + "spans": [ + { + "bbox": [ + 382, + 220, + 482, + 229 + ], + "type": "text", + "content": "(c) full-field swept-angle SWI" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 264, + 532, + 339 + ], + "blocks": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "type": "text", + "content": "Figure 5. Synthetic wavelength interferometry setups. (a) A full-field interferometer efficiently acquires full-frame depth, but is susceptible to phase corruptions due to aberrations and indirect illumination. (b) A scanning interferometer is robust to such corruptions, but requires slow lateral scanning. (c) A swept-angle full-field interferometer achieves both efficiency and robustness." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 264, + 532, + 339 + ], + "lines": [ + { + "bbox": [ + 59, + 264, + 532, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 264, + 532, + 339 + ], + "type": "image", + "image_path": "d1fd8eeeea364b497c19637e65277053565e0d62b92dc96f4b01a6148012a178.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 340, + 546, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 340, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 546, + 373 + ], + "type": "text", + "content": "Figure 6. The " + }, + { + "bbox": [ + 46, + 340, + 546, + 373 + ], + "type": "inline_equation", + "content": "\\{4,4\\}" + }, + { + "bbox": [ + 46, + 340, + 546, + 373 + ], + "type": "text", + "content": "-shift phase retrieval pipeline. (a) We measure intensity at 16 reference positions, corresponding to 4 synthetic by 4 carrier subwavelength shifts. (b) and (c) For each synthetic sub-wavelength shift, we estimate interference-free and envelope images. (d) We denoise the envelope image using joint bilateral filtering. (e) We use 4-shift phase retrieval to estimate envelope phase and depth." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 379, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 437 + ], + "type": "text", + "content": "broadening its emission spectrum. This coupling was not an issue for prior work, which focused on OCT applications that require broadband illumination; but it makes spatially-incoherent light sources incompatible with SWI, which requires narrow-linewidth dichromatic illumination." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": "Emulating spatial coherence. To resolve this conundrum, we consider the following: Replacing the point emitter in Figure 3 with an area emitter changes the illumination, from a single collimated beam parallel to the optical axis, to the superposition of beams traveling along directions offset from the optical axis by angles " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\theta \\in [-\\Theta /2,\\Theta /2]" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " depends on the emission area and lens focal length. We denote by " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{u}_s^\\theta (x)" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "inline_equation", + "content": "\\pmb{u}_r^\\theta (x,l)" + }, + { + "bbox": [ + 46, + 438, + 288, + 570 + ], + "type": "text", + "content": " the complex fields resulting by each such beam reflecting on the scene and reference arms. Then, because different points on the area emitter are incoherent with each other, we can update Equation (1) as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 571, + 154, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 571, + 154, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 571, + 154, + 597 + ], + "type": "interline_equation", + "content": "I (x, l) = \\int_ {\\theta} I ^ {\\theta} (x, l) \\mathrm {d} \\theta ,", + "image_path": "7fe9dbdc844eb3926e86f38ad0338be6a542bce8e15bf5d98a4f0fb3cc187f68.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 598, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 598, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 598, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {I}} (x, l) = \\int_ {\\theta} \\boldsymbol {\\mathcal {I}} ^ {\\theta} (x, l) \\mathrm {d} \\theta , \\quad \\boldsymbol {\\mathcal {C}} (x, l) = \\int_ {\\theta} \\boldsymbol {\\mathcal {C}} ^ {\\theta} (x, l) \\mathrm {d} \\theta , \\tag {7}", + "image_path": "1385f9fa96a1a981504ec097a447e1527c9c5432dd5ae21865338f2fe396b2a9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 625, + 277, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 277, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 277, + 658 + ], + "type": "interline_equation", + "content": "I^{\\theta}(x,l)\\equiv \\underbrace{\\left|\\boldsymbol{u}_{s}^{\\theta}(x)\\right|^{2} + \\left|\\boldsymbol{u}_{r}^{\\theta}(x,l)\\right|^{2}}_{\\equiv \\boldsymbol{\\mathcal{I}}^{\\theta}(x,l)} + 2\\operatorname {Re}\\bigl\\{\\underbrace{\\boldsymbol{u}_{s}^{\\theta}(x)\\boldsymbol{u}_{r}^{\\theta*}(x,l)}_{\\equiv \\boldsymbol{\\mathcal{C}}^{\\theta}(x,l)}\\bigr \\} .", + "image_path": "2baebce16d644e8b772919350f97c0cef1688280f3aadcbf0e49ca955ff36f67.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 659, + 288, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 288, + 707 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 288, + 707 + ], + "type": "text", + "content": "We can interpret Equation (7) as follows: the image " + }, + { + "bbox": [ + 46, + 659, + 288, + 707 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 659, + 288, + 707 + ], + "type": "text", + "content": " measured using an area emitter equals the sum of the images that would be measured using independent point emitters spanning the emission area, each producing illumination at" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "type": "text", + "content": "an angle " + }, + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "type": "text", + "content": " (and respectively for correlation " + }, + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 306, + 379, + 495, + 390 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 391, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 570 + ], + "type": "text", + "content": "With this interpretation at hand, we design the setup of Figure 5(c) to emulate a spatially-incoherent source suitable for SWI through time-division multiplexing. We use a galvo mirror to steer a narrow collimated beam of narrow-linewidth dichromatic illumination. We focus this beam through a relay lens at a point on the focal plane of the illumination lens. As we steer the beam direction, the focus point scans an area on the focal plane. In turn, each point location results in the injection of illumination at an angle " + }, + { + "bbox": [ + 304, + 391, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 391, + 547, + 570 + ], + "type": "text", + "content": " in the interferometer. By scanning the focus point over a target effective emission area within a single exposure, we acquire image and correlation measurements that equal those of Equation (7). As this design sweeps the illumination angle, we term it swept-angle illumination, and its combination with full-field SWI swept-angle synthetic wavelength interferometry." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 547, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 702 + ], + "type": "text", + "content": "Comparison to Kotwal et al. [41]. The illumination module of the proposed setup in Figure 5(c) is an adaptation of the Fourier-domain redistributive projector of Kotwal et al. [41]. Compared to their setup, we do not need to use an amplitude electro-optic modulator, as we are only interested in emulating spatial incoherence (equivalent, directly probing). Additionally, their setup was designed for use with monochromatic illumination, to enable different light transport probing capabilities. By contrast, we use our setup with narrow-linewidth dichromatic illumination, to enable synthetic wavelength interferometry. More broadly," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8237" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "our key insight is that swept-angle illumination can emulate extended-area emitters with arbitrary spectral profiles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 284, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 284, + 116 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 284, + 116 + ], + "type": "text", + "content": "5. The " + }, + { + "bbox": [ + 47, + 102, + 284, + 116 + ], + "type": "inline_equation", + "content": "\\{M,N\\}" + }, + { + "bbox": [ + 47, + 102, + 284, + 116 + ], + "type": "text", + "content": "-shift phase retrieval algorithm" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 118, + 287, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 118, + 287, + 164 + ], + "spans": [ + { + "bbox": [ + 46, + 118, + 287, + 164 + ], + "type": "text", + "content": "We discuss our pipeline for acquiring and processing measurements to estimate the envelope phase, and thus depth. For this, in Figure 5(c), we use a nanometer-accuracy stage to vary the position " + }, + { + "bbox": [ + 46, + 118, + 287, + 164 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 118, + 287, + 164 + ], + "type": "text", + "content": " of the reference mirror." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": "Phase retrieval. We use the " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": "-shift phase retrieval algorithm [9, 63, 64] to compute the envelope phase: We assume we have estimates of the envelope's squared amplitude " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "|\\pmb{\\mathcal{E}}(x,l_n)|^2" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": " at reference positions " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "l_{n}" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": " corresponding to shifts by " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": "-th fractions of the synthetic half-wavelength " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "\\lambda_s / 2" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "l_{n} = l + n\\lambda_{s} / 2N" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "inline_equation", + "content": "n\\in \\{0,\\dots ,N - 1\\}" + }, + { + "bbox": [ + 47, + 165, + 288, + 249 + ], + "type": "text", + "content": ". Then, from Equation (6), we estimate the envelope phase as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 250, + 288, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 288, + 283 + ], + "type": "interline_equation", + "content": "\\phi (d (x)) = \\arctan \\left[ \\frac {\\sum_ {n = 0} ^ {N - 1} \\left| \\boldsymbol {\\mathcal {E}} (x , l _ {n}) \\right| ^ {2} \\sin \\left(2 \\pi n / N\\right)}{\\sum_ {n = 0} ^ {N - 1} \\left| \\boldsymbol {\\mathcal {E}} (x , l _ {n}) \\right| ^ {2} \\cos \\left(2 \\pi n / N\\right)} \\right], \\tag {8}", + "image_path": "6c7ae42d73274f1e1b1df9faf9494c19a8b9cf034f2e48277cbd94e28068cb37.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 285, + 251, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 251, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 251, + 297 + ], + "type": "text", + "content": "and the depth (up to an integer multiple of " + }, + { + "bbox": [ + 47, + 285, + 251, + 297 + ], + "type": "inline_equation", + "content": "\\lambda_{s} / 2" + }, + { + "bbox": [ + 47, + 285, + 251, + 297 + ], + "type": "text", + "content": ") as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 299, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 299, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 119, + 299, + 287, + 312 + ], + "type": "interline_equation", + "content": "d (x) = l + \\frac {\\phi (d (x))}{2 \\kappa_ {s}}. \\tag {9}", + "image_path": "780061eda99b66f782c951792c09c89be1220e93562bd9324e1595e0fc4e142f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": "Envelope estimation. Equation (8) requires estimates of " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "|\\mathcal{E}(x,l_n)|^2" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "l_{n}" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ", we capture intensity measurements " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "I(x,l_n^m)" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": " reference positions corresponding to shifts by " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": "-th fractions of the carrier wavelength, " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "l_{n}^{m} = l_{n} + m\\lambda_{c} / M" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "m \\in \\{0,\\dots,M - 1\\}" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ". As the carrier wavelength " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": " is half the optical wavelength " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": " and orders of magnitude smaller than the synthetic wavelength " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ", the envelope and interference-free image remain approximately constant across shifts, " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(x,l_n^m) \\approx \\mathcal{E}(x,l_n)" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(x,l_n^m) \\approx \\mathcal{I}(x,l_n)" + }, + { + "bbox": [ + 46, + 315, + 288, + 434 + ], + "type": "text", + "content": ". Then, from Equations (1) and (4)," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 50, + 436, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 436, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 50, + 436, + 287, + 449 + ], + "type": "interline_equation", + "content": "I (x, l _ {n} ^ {m}) = \\boldsymbol {\\mathcal {T}} (x, l _ {n}) + 2 \\sin \\left(\\kappa_ {c} \\left(d (x) - l _ {n} ^ {m}\\right)\\right) \\boldsymbol {\\mathcal {E}} (x, l _ {n}). \\tag {10}", + "image_path": "8da718c07636f27e166a1e66db0a8a16eb265834ec59c155fd2e682933841783.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 452, + 275, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 275, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 275, + 464 + ], + "type": "text", + "content": "We estimate the interference-free image and envelope as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 464, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 464, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 64, + 464, + 287, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {I} (x, l _ {n}) = ^ {1} / M \\sum_ {m = 0} ^ {M - 1} I (x, l _ {n} ^ {m}), \\tag {11}", + "image_path": "eabfbb4bc9e6fd021183dd5126a7066a533bc6f489d82e9f689e815452f118cf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 489, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\left| \\boldsymbol {\\mathcal {E}} \\left(x, l _ {n}\\right) \\right| ^ {2} = 1 / _ {2 M} \\sum_ {m = 0} ^ {M - 1} \\left(I \\left(x, l _ {n} ^ {m}\\right) - \\boldsymbol {\\mathcal {I}} \\left(x, l _ {n}\\right)\\right) ^ {2}. \\tag {12}", + "image_path": "b535b2203a4ce63e474eee7904fd244b771dbcae33c7baec3ccd61ee2658b2bd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": "Equations (11), (12), (8), and (9) complete our phase estimation pipeline, which we summarize in Figure 6. We note that the runtime of this pipeline is negligible relative to acquisition time. We provide algorithmic details and pseudocode in the supplement. We call this pipeline the " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\{M,N\\}" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": "-shift phase retrieval algorithm. The parameters " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " are design parameters that we can fine-tune to control a trade-off between acquisition time and depth accuracy: As we increase " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": ", the number " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "M\\cdot N" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " of images we capture also increases, and depth estimates become more robust to noise. The theoretical minimum number of images is achieved using " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\{3,3\\}" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " shifts (9 images). In practice, we found " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\{4,4\\}" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " shifts to be robust. We evaluate different " + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\{M,N\\}" + }, + { + "bbox": [ + 46, + 511, + 288, + 677 + ], + "type": "text", + "content": " combinations in the supplement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Dealing with speckle. Interference in non-specular scenes takes the form of speckle, a high-frequency pseudo-random pattern (Figure 6(a)). This can result in noisy envelope," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 369, + 127 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 369, + 127 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 369, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 369, + 127 + ], + "type": "image", + "image_path": "0d12483adad0e5981445b2adc67e5aa2fc03c1b62909219ff2d29932cb30e341.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 129, + 417, + 140 + ], + "lines": [ + { + "bbox": [ + 318, + 129, + 417, + 140 + ], + "spans": [ + { + "bbox": [ + 318, + 129, + 417, + 140 + ], + "type": "text", + "content": "(a) with swept-angle source" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 369, + 71, + 424, + 127 + ], + "blocks": [ + { + "bbox": [ + 369, + 71, + 424, + 127 + ], + "lines": [ + { + "bbox": [ + 369, + 71, + 424, + 127 + ], + "spans": [ + { + "bbox": [ + 369, + 71, + 424, + 127 + ], + "type": "image", + "image_path": "6a06672c511b5a7b2d25f3dd120bc7af3e7ed40e57f9b96c5010985314f008ec.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 129, + 536, + 140 + ], + "lines": [ + { + "bbox": [ + 427, + 129, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 427, + 129, + 536, + 140 + ], + "type": "text", + "content": "(b) without swept-angle source" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 481, + 127 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 481, + 127 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 481, + 127 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 481, + 127 + ], + "type": "image", + "image_path": "1e4cf0d27b7f3dbb7fe55e84c4354910dd7d8f07ec5ac4f6b2d0fc8e0d72734b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 140, + 547, + 162 + ], + "lines": [ + { + "bbox": [ + 305, + 140, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 547, + 162 + ], + "type": "text", + "content": "Figure 7. Effect of swept-angle scanning. Swept-angle illumination helps mitigate speckle noise and subsurface scattering effects." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 482, + 70, + 538, + 127 + ], + "blocks": [ + { + "bbox": [ + 482, + 70, + 538, + 127 + ], + "lines": [ + { + "bbox": [ + 482, + 70, + 538, + 127 + ], + "spans": [ + { + "bbox": [ + 482, + 70, + 538, + 127 + ], + "type": "image", + "image_path": "d97392e45807ebb0e3ccc33cdc0edc06eb63038680deb24ae2d1852775a77c9e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 165, + 547, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 547, + 285 + ], + "type": "text", + "content": "and thus phase and depth, estimates. As Figures 4(d) and 7 show, swept-angle illumination greatly mitigates speckle effects. We can further reduce the impact of speckle by denoising estimated quantities with a low-pass filter (e.g., Gaussian) [20, 41]. Alternatively, to avoid blurring image details, we can use joint bilateral filtering [12, 59] with a guide image of the scene under ambient light. Empirically, we found it better to blur the envelope estimates of Equation (12) before using them in Equation (8) (Figure 6(d)), compared to blurring the phase or depth estimates." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 292, + 387, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 387, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 387, + 305 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": "Experimental prototype. For all our experiments, we use the experimental prototype in Figure 5(c). Our prototype comprises: two distributed-Bragg-reflector lasers (wavelengths " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "780\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "781\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": ", power " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "45\\mathrm{mW}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": ", linewidth " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "1\\mathrm{MHz}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": "); a galvo mirror pair; a translation stage (resolution " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "10\\mathrm{nm}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": "); two compound macro lenses (focal length " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": "); a CCD sensor (pixel pitch " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "3.7\\mu \\mathrm{m}" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "inline_equation", + "content": "3400\\times 2700" + }, + { + "bbox": [ + 304, + 307, + 547, + 426 + ], + "type": "text", + "content": " pixels); and other free-space optics and mounts. We provide a detailed component list and specifications, calibration instructions, and other information in the supplement." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": "We use the following experimental specifications: The reproduction ratio is 1:1, the field of view is " + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "inline_equation", + "content": "12.5\\mathrm{mm}\\times 10\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": " and the working distance is " + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "inline_equation", + "content": "400\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": ". The unambiguous depth range is approximately " + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "inline_equation", + "content": "500\\mu \\mathrm{m}" + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "inline_equation", + "content": "\\{4,4\\}" + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": "-shifts (16 images), and a minimum per-image exposure time of " + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "inline_equation", + "content": "10\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": ", resulting at a frame rate of " + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 426, + 547, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 498, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 700 + ], + "type": "text", + "content": "Depth recovery on challenging scenes. Figures 8 and 2 show scans of many challenging scenes. We scan materials ranging from rough metallic (coins), to diffuse (music box, pill), to highly-scattering (soap and chocolate). Most scenes include fine features requiring high lateral and axial resolution (music box, business card, US quarter, dollar bill). We compare scan results using swept-angle SWI (with bilateral or Gaussian filtering), versus conventional full-field SWI (Figure 5(a), emulated by deactivating the angle-sweep galvo). In all scenes, using swept-angle illumination greatly improves reconstruction quality. The difference is more pronounced in scenes with strong subsurface scattering (music box, chocolate, soap). Even in metallic scenes where there is no subsurface scattering (coins), the use of swept-angle illumination still improves reconstruction quality, by helping mitigate aberration artifacts. Bilateral filtering further improves lateral detail compared to Gaussian filtering." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 713 + ], + "type": "text", + "content": "Comparison with scanning SWI. A direct comparison of" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8238" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 532, + 459 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 532, + 459 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 532, + 459 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 532, + 459 + ], + "type": "image", + "image_path": "130d8b8ce6a1503c46e48a29387c13cd729fecada546014e99fd33934182c5d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 460, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 545, + 483 + ], + "type": "text", + "content": "Figure 8. Depth reconstruction. Depth maps (left) and surface renderings (right) acquired using full-field SWI with: (c) swept-angle scanning and bilateral filtering; (d) swept-angle scanning and Gaussian filtering; (e) no swept-angle scanning and with Gaussian filtering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 486, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 688 + ], + "type": "text", + "content": "swept-angle SWI with scanning SWI is challenging, because of the differences between scanning and full-field setups (Figure 5, (b) versus (a), (c)). To qualitatively assess their relative performance, in Figure 9 we use the following experimental protocol: We downsample the depth map from our technique, by a factor of 35 in each dimension. This is approximately equal to the number of depths points a scanning SWI system operating in point-to-point mode would acquire at the same total exposure time as ours, when equipped with beam-steering optics that operate at a " + }, + { + "bbox": [ + 46, + 486, + 289, + 688 + ], + "type": "inline_equation", + "content": "30\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 486, + 289, + 688 + ], + "type": "text", + "content": " scan rate—we detail the exact calculation in the supplement. We then use joint bilateral upsampling [40] with the same guide image as in our technique, to upsample the downsampled depth map back to its original resolution. We observe that, due to the sparse set of points scanning SWI can acquire, the reconstructions miss fine features such as the hair on the quarter and letters on the business card." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Scanning SWI systems can also operate in resonant mode, typically using Lissajous scanning to raster scan the field of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 486, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 665 + ], + "type": "text", + "content": "view [47]. Resonant mode enables much faster operation than the point-to-point mode we compare against. However, Lissajous scanning results in non-rectilinear sampling of the image plane, as scan lines smear across multiple image rows (when the fast axis is horizontal). At high magnifications, this produces strong spatial distortions and artifacts, making resonant mode operation unsuitable for applications that require micrometer lateral resolution. Our technique also uses resonant-mode Lissajous scanning, but to raster scan the focal plane of the illumination lens (Figure 5(d)), and not the image plane. Instead, our technique uses a two-dimensional sensor to sample the image plane, thus completely avoiding the artifacts above and achieving micrometer lateral resolution. The supplement discusses other challenges in achieving high lateral resolution with scanning SWI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Additionally, our technique is orders of magnitude faster than resonant-mode scanning SWI at the same lateral resolution. This is because, whereas scanning SWI must scan the image plane at the target (pixel-level) lateral resolution, our" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8239" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 279, + 233 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 279, + 233 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 279, + 233 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 279, + 233 + ], + "type": "image", + "image_path": "149abe0ca9835118ada176022492b1bf15a4a621b8badd2d710de1ae592b4cc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 235, + 255, + 245 + ], + "lines": [ + { + "bbox": [ + 75, + 235, + 255, + 245 + ], + "spans": [ + { + "bbox": [ + 75, + 235, + 255, + 245 + ], + "type": "text", + "content": "(a) full-field SWI (ours) (b) scanning+upsampling" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 245, + 288, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 245, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 288, + 277 + ], + "type": "text", + "content": "Figure 9. Comparison with upsampled scanning SWI. We emulate scanning SWI by downsampling our swept-angle SWI depth. We then bilaterally upsample it to the original resolution." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 316, + 275, + 388 + ], + "blocks": [ + { + "bbox": [ + 47, + 280, + 287, + 313 + ], + "lines": [ + { + "bbox": [ + 47, + 280, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 287, + 313 + ], + "type": "text", + "content": "Table 2. Depth accuracy. MedAE is the median absolute error between ground truth and estimated depth. Kernel width is the lateral size of the speckle blur filter. All quantities are in " + }, + { + "bbox": [ + 47, + 280, + 287, + 313 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 280, + 287, + 313 + ], + "type": "text", + "content": "m." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 316, + 275, + 388 + ], + "lines": [ + { + "bbox": [ + 60, + 316, + 275, + 388 + ], + "spans": [ + { + "bbox": [ + 60, + 316, + 275, + 388 + ], + "type": "table", + "html": "
kernel widthwith swept-anglew/o swept-angle
RMSEMedAERMSEMedAE
78.24.818.913.2
155.13.611.29.5
212.01.610.57.3
301.61.011.16.7
", + "image_path": "95f960a67bca7b3c90fa037b965991f0f6030fd8b7eb5e0d066d5327e6ea59df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 450 + ], + "type": "text", + "content": "technique remains effective even when scanning the focal plane at much lower resolutions. Figure 10 shows images of the focal plane during scanning, taken with the measurement camera. The dark regions (insets) show that scanning resolution is much lower than pixel-level resolution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "content": "Axial resolution. To quantify the axial resolution of our technique, we perform the following experiment. We use a second nanometer-accurate translation stage to place the chocolate scene from Figure 8 at different depths from the camera, at increments of " + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "inline_equation", + "content": "1\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "content": ". We choose this scene because it is has strong sub-surface scattering. We then compare how well full-field SWI with and without swept-angle illumination can track the scene depth. We perform this experiment using Gaussian filtering with different kernel sizes, to additionally quantify lateral resolution. The results in Table 2 show that we can achieve an axial resolution of approximately " + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "inline_equation", + "content": "5\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "inline_equation", + "content": "1\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "content": ", for kernel sizes " + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "inline_equation", + "content": "7\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "inline_equation", + "content": "30\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 450, + 288, + 629 + ], + "type": "text", + "content": ", respectively. We can trade off lateral for axial resolution, by increasing kernel size. Figure 8 shows that this trade-off can become more favorable with bilateral filtering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Additional experiments. In the supplement: 1) Comparisons with full-field OCT. We show that swept-angle SWI can approximate the reconstruction quality of OCT, despite being " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " faster. 2) Tunable depth range. We show that, by adjusting the wavelength separation between the two lasers, swept-angle SWI can scan scenes with depth range " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "16\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " at resolution " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "50\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", at significantly better quality than con" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 526, + 140 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 526, + 140 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 526, + 140 + ], + "type": "image", + "image_path": "cbd3fffb0ba0eb3e5b642bcf93fde51434095b975a64d6e67461378a4169e091.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 140, + 547, + 161 + ], + "lines": [ + { + "bbox": [ + 306, + 140, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 547, + 161 + ], + "type": "text", + "content": "Figure 10. Lissajous scanning. Image and insets of the illumination lens' focal plane, showing the Lissajous scanning pattern." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "text", + "content": "ventional full-field SWI. 3) Robustness to ambient lighting. We show that the use of near-monochromatic illumination makes swept-angle SWI robust to strong ambient light, even at " + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "text", + "content": " signal-to-background ratios. 4) Effect of scanning resolution and " + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "inline_equation", + "content": "\\{M,N\\}" + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "text", + "content": " settings. We show that, by adjusting the focal plane scanning resolution and " + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "inline_equation", + "content": "\\{M,N\\}" + }, + { + "bbox": [ + 305, + 165, + 546, + 261 + ], + "type": "text", + "content": " parameters for phase retrieval, swept-angle SWI can trade off between acquisition time versus reconstruction quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 268, + 460, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 460, + 280 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 460, + 280 + ], + "type": "text", + "content": "7. Limitations and conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "type": "text", + "content": "Phase wrapping. SWI determines depth only up to an integer multiple of the synthetic half-wavelength " + }, + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "type": "inline_equation", + "content": "\\lambda_{s} / 2" + }, + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "type": "text", + "content": ". Equivalently, all depth values are wrapped to " + }, + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "type": "inline_equation", + "content": "[0, \\lambda_{s} / 2]" + }, + { + "bbox": [ + 305, + 283, + 547, + 414 + ], + "type": "text", + "content": ", even if the depth range is greater. To mitigate phase wrapping and extend unambiguous depth range, it is common to capture measurements at multiple synthetic wavelengths, and use them to unwrap the phase estimate [3,7,8,11]. Theoretically two synthetic wavelengths are enough to uniquely determine depth, but in practice phase unwrapping techniques are very sensitive to noise [25]. Combining phase unwrapping with our technique is an interesting future direction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "type": "text", + "content": "Toward real-time operation. Our prototype acquires measurements at a frame rate of " + }, + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "type": "text", + "content": ", due to the need for " + }, + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "type": "inline_equation", + "content": "10\\mathrm{ms}" + }, + { + "bbox": [ + 305, + 415, + 547, + 498 + ], + "type": "text", + "content": " per-frame exposure time. If we use a stronger laser to reduce this time, the main speed bottleneck will be the need to perform phase shifts by physically translating the reference mirror. We can mitigate this bottleneck using faster translation stages (e.g., fast microscopy stages)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 498, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 547, + 629 + ], + "type": "text", + "content": "Conclusion. We presented a technique for fast depth sensing at micron-scale lateral and axial resolutions. Our technique, swept-angle synthetic wavelength interferometry, combines the complementary advantages of full-field interferometry (speed, pixel-level lateral resolution) and scanning interferometry (robustness to aberrations and indirect illumination). We demonstrated these advantages by scanning multiple scenes with fine geometry and strong subsurface scattering. We expect our results to motivate applications of swept-angle SWI in areas such as biomedical imaging, and industrial fabrication and inspection (Figure 1)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments. We thank Sudershan Boovaraghavan, Yuvraj Agrawal, Arpit Agarwal, Wenzhen Yuan from CMU, and Veniamin V. Stryzheus, Brian T. Miller from The Boeing Company, who provided the samples for the experiments in Figure 1. This work was supported by NSF awards 1730147, 2047341, 2008123 (NSF-BSF 2019758), and a Sloan Research Fellowship for Ioannis Gkioulekas." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8240" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "type": "text", + "content": "[1] Supreeth Achar, Joseph R Bartels, William L Whittaker, Kiriakos N Kutulakos, and Srinivasa G Narasimhan. Epipolar time-of-flight imaging. ACM TOG, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 147 + ], + "type": "text", + "content": "[2] Brian Aull. 3d imaging with geiger-mode avalanche photodiodes. Opt. Photon. News, 2005. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "text", + "content": "[3] Seung-Hwan Baek, Noah Walsh, Ilya Chugunov, Zheng Shi, and Felix Heide. Centimeter-wave free-space neural time-of-flight imaging. ACM Transactions on Graphics (TOG), 2022. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 193, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 215 + ], + "type": "text", + "content": "[4] Stephen T. Barnard and William B. Thompson. Disparity analysis of images. IEEE TPAMI, 1980. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "text", + "content": "[5] Ayush Bhandari, Achuta Kadambi, Refael Whyte, Christopher Barsi, Micha Feigin, Adrian Dorrington, and Ramesh Raskar. Resolving multipath interference in time-of-flight imaging via modulation frequency diversity and sparse regularization. Optics Letters, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 271, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 271, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 288, + 303 + ], + "type": "text", + "content": "[6] Tongbo Chen, Hans-Peter Seidel, and Hendrik P. A. Lensch. Modulated phase-shifting for 3d scanning. IEEE/CVF CVPR, 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 306, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 327 + ], + "type": "text", + "content": "[7] Yeou-Yen Cheng and James C. Wyatt. Two-wavelength phase shifting interferometry. Applied Optics, 1984. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 329, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 351 + ], + "type": "text", + "content": "[8] Yeou-Yen Cheng and James C. Wyant. Multiple-wavelength phase-shifting interferometry. Applied Optics, 1985. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 352, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 363 + ], + "type": "text", + "content": "[9] Peter de Groot. Phase Shifting Interferometry. 2011. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 364, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 288, + 385 + ], + "type": "text", + "content": "[10] Peter de Groot and John McGarvey. Chirped synthetic-wavelength interferometry. Optics Letters, 1992. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 387, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 288, + 419 + ], + "type": "text", + "content": "[11] David Droeschel, Dirk Holz, and Sven Behnke. Multifrequency phase unwrapping for time-of-flight cameras. IEEE/RSJ IROS, 2010. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 421, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 453 + ], + "type": "text", + "content": "[12] Elmar Eisemann and Frédo Durand. Flash photography enhancement via intrinsic relighting. ACM transactions on graphics (TOG), 2004. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 455, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 497 + ], + "type": "text", + "content": "[13] Adolf F Fercher, Wolfgang Drexler, Christoph K Hitzenberger, and Theo Lasser. Optical coherence tomography-principles and applications. Reports on progress in physics, 2003. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 499, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 288, + 533 + ], + "type": "text", + "content": "[14] Adolf F. Fercher, H Z. Hu, and U. Vry. Rough surface interferometry with a two-wavelength heterodyne speckle interferometer. Applied Optics, 1985. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 533, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 578 + ], + "type": "text", + "content": "[15] Richard Ferriere, Johann Cussey, and John M. Dudley. Time-of-flight range detection using low-frequency intensity modulation of a cw laser diode: application to fiber length measurement. Optical Engineering, 2008. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "text", + "content": "[16] Angel Flores, Craig Robin, Ann Lanari, and Iyad Dajani. Pseudo-random binary sequence phase modulation for narrow linewidth, kilowatt, monolithic fiber amplifiers. Optics Express, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 624, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 655 + ], + "type": "text", + "content": "[17] Daniel Freedman, Eyal Krupka, Yoni Smolin, Ido Leichter, and Mirko Schmidt. Sra: Fast removal of general multipath for tof sensors, 2014. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 657, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 680 + ], + "type": "text", + "content": "[18] Stefan Fuchs. Multipath interference compensation in time-of-flight camera images. ICPR, 2010. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 681, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 712 + ], + "type": "text", + "content": "[19] Genevieve Gariepy, Nikola Krstajic, Robert Henderson, Chunyong Li, Robert R Thomson, Gerald S Buller, Barmak Heshmat, Ramesh Raskar, Jonathan Leach, and Daniele Faccio." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 41, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "Single-photon sensitive light-in-fight imaging. Nature Comm., 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 96, + 546, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 546, + 129 + ], + "type": "text", + "content": "[20] Ioannis Gkioulekas, Anat Levin, Frédo Durand, and Todd Zickler. Micron-scale light transport decomposition using interferometry. ACM TOG, 2015. 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 131, + 515, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 515, + 142 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 515, + 142 + ], + "type": "text", + "content": "[21] Pavel Grossmann. Depth from focus. PRL, 1987. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 144, + 545, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 144, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 545, + 165 + ], + "type": "text", + "content": "[22] Anant Gupta, Atul Ingle, and Mohit Gupta. Asynchronous single-photon 3d imaging. In ICCV, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 167, + 546, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 546, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 546, + 198 + ], + "type": "text", + "content": "[23] Anant Gupta, Atul Ingle, Andreas Velten, and Mohit Gupta. Photon-flooded single-photon 3d cameras. IEEE/CVF CVPR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 201, + 546, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 546, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 546, + 233 + ], + "type": "text", + "content": "[24] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G Narasimhan. Structured light 3d scanning in the presence of global illumination. IEEE/CVF CVPR, 2011. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 235, + 546, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 546, + 268 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 546, + 268 + ], + "type": "text", + "content": "[25] Mohit Gupta, Shree K Nayar, Matthias B Hullin, and Jaime Martin. Phasor imaging: A generalization of correlation-based time-of-flight imaging. ACM TOG, 2015. 3, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 270, + 546, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 270, + 546, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 546, + 301 + ], + "type": "text", + "content": "[26] Mohit Gupta, Andreas Velten, Shree K. Nayar, and Eric Breitbach. What are optimal coding functions for time-of-flight imaging? ACM TOG, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 303, + 546, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 546, + 336 + ], + "type": "text", + "content": "[27] Felipe Gutierrez-Barragan, Syed Azer Reza, Andreas Velten, and Mohit Gupta. Practical coding function design for time-of-flight imaging. IEEE/CVF CVPR, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 338, + 546, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 546, + 370 + ], + "type": "text", + "content": "[28] Yudeog Han, Joon-Young Lee, and In So Kweon. High quality shape from a single rgb-d image under uncalibrated natural illumination. IEEE ICCV, 2013. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 372, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 546, + 392 + ], + "type": "text", + "content": "[29] Parameswaran Hariharan. Optical interferometry. Elsevier, 2003. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 395, + 546, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 546, + 418 + ], + "type": "text", + "content": "[30] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. 2004. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 418, + 546, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 546, + 451 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 546, + 451 + ], + "type": "text", + "content": "[31] Caner Hazirbas, Sebastian Georg Soyer, Maximilian Christian Staab, Laura Leal-Taixe, and Daniel Cremers. Deep depth from focus, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 453, + 546, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 546, + 485 + ], + "type": "text", + "content": "[32] Felix Heide, Steven Diamond, David B. Lindell, and Gordon Wetzstein. Sub-picosecond photon-efficient 3d imaging using single-photon sensors. Scientific Reports, 2018. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 487, + 546, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 546, + 519 + ], + "type": "text", + "content": "[33] Felix Heide, Matthias B Hullin, James Gregson, and Wolfgang Heidrich. Low-budget transient imaging using photonic mixer devices. ACM TOG, 2013. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 521, + 546, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 546, + 554 + ], + "type": "text", + "content": "[34] Berthold Klaus Paul Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. Technical report, 1970. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "type": "text", + "content": "[35] David Huang, Eric A Swanson, Charles P Lin, Joel S Schuman, William G Stinson, Warren Chang, Michael R Hee, Thomas Flotte, Kenton Gregory, Carmen A Puliafito, and James G Fujimoto. Optical coherence tomography. Science, 1991. 3" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 612, + 546, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 546, + 655 + ], + "type": "text", + "content": "[36] David Jimeneza, Daniel Pizarrob, Manuel Mazoa, and Sira Palazuelos. Modeling and correction of multipath interference in time of flight cameras. Image and Vision Computing, 2014. 3" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 307, + 657, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 546, + 689 + ], + "type": "text", + "content": "[37] Jon L. Johnson, Timothy D. Dorney, and Daniel M. Mittleman. Enhanced depth resolution in terahertz imaging using phase-shift interferometry. Applied Physics Letters, 2001. 2" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 307, + 691, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 712 + ], + "type": "text", + "content": "[38] Achuta Kadambi, Refael Whyte, Ayush Bhandari, Lee Streeter, Christopher Barsi, Adrian Dorrington, and Ramesh" + } + ] + } + ], + "index": 40 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8241" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "Raskar. Coded Time of Flight Cameras: Sparse Deconvolution to Address Multipath Interference and Recover Time Profiles. ACM TOG, 2013. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 139 + ], + "type": "text", + "content": "[39] Ahmed Kirmani, Dheera Venkatraman, Dongeek Shin, Andrea Colaço, Franco N. C. Wong, Jeffrey H. Shapiro, and Vivek K Goyal. First-photon imaging. Science, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 172 + ], + "type": "text", + "content": "[40] Johannes Kopf, Michael F. Cohen, Dani Lischinski, and Matt Uytendaele. Joint bilateral upsampling. ACM Trans. Graph., 26(3):96-es, jul 2007. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 205 + ], + "type": "text", + "content": "[41] Alankar Kotwal, Anat Levin, and Ioannis Gkioulekas. Interferometric transmission probing with coded mutual intensity. ACM TOG, 2020. 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 205, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 227 + ], + "type": "text", + "content": "[42] Robert Lange and Peter Seitz. Solid-state time-of-flight range camera. IEEE JQE, 2001. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 259 + ], + "type": "text", + "content": "[43] Robert Lange, Peter Seitz, Alice Biber, and Stefan Lauxtermann. Demodulation pixels in ccd and cmos technologies. SPIE, 2000. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "text", + "content": "[44] Fengqiang Li, Florian Willomitzer, Prasanna Rangarajan, Mohit Gupta, Andreas Velten, and Oliver Cossairt. Sh-tof: Micro resolution time-of-flight imaging with superheterodyne interferometry. IEEE ICCP, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "type": "text", + "content": "[45] Fengqiang Li, Joshua Yablon, Andreas Velten, Mohit Gupta, and Oliver Cossairt. High-depth-resolution range imaging with multiple-wavelength superheterodyne interferometry using 1550-nm lasers. Applied Optics, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 381 + ], + "type": "text", + "content": "[46] David B Lindell, Matthew O'Toole, and Gordon Wetzstein. Single-photon 3d imaging with deep sensor fusion. ACM TOG, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 383, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 425 + ], + "type": "text", + "content": "[47] Xiaomeng Liu, Kristofer Henderson, Joshua Rego, Suren Jayasuriya, and Sanjeev Koppal. Dense lissajous sampling and interpolation for dynamic light-transport. Optics Express, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[48] Julio Marco, Quercus Hernandez, Adolfo Muñoz, Yue Dong, Adrian Jarabo, Min H. Kim, Xin Tong, and Diego Gutierrez. Deeptof: Off-the-shelf real-time correction of multipath interference in time-of-flight imaging. ACM TOG, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[49] Nikhil Naik, Achuta Kadambi, Christoph Rhemann, Shahram Izadi, Ramesh Raskar, and Sing Bing Kang. A light transport model for mitigating multipath interference in time-of-flight sensors. IEEE/CVF CVPR, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 548 + ], + "type": "text", + "content": "[50] Lazaros Nalpantidis, Georgios Sirakoulis, and Antonios Gasteratos. Review of stereo vision algorithms: From software to hardware. Int. J. Optomechatronics, 2008. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "text", + "content": "[51] Shree Nayar, Gurunandan Krishnan, Michael Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. ACM TOG, 2006. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[52] Cristiano Niclass, Alexis Rochas, Pierre. Besse, and Edoardo Charbon. Design and characterization of a cmos 3-d image sensor based on single photon avalanche diodes. IEEE JSSC, 2005. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 669 + ], + "type": "text", + "content": "[53] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM TOG, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[54] Matthew O'Toole, Felix Heide, David B Lindell, Kai Zang, Steven Diamond, and Gordon Wetzstein. Reconstructing transient images from single-photon sensors. IEEE/CVF CVPR, 2017. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[55] Matthew O'Toole, Felix Heide, Lei Xiao, Matthias B Hullin, Wolfgang Heidrich, and Kiriakos N Kutulakos. Temporal Frequency Probing for 5D Transient Analysis of Global Light Transport. ACM TOG, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 151 + ], + "type": "text", + "content": "[56] Matthew O'Toole, John Mather, and Kiriakos N. Kutulakos. 3d shape and indirect appearance by structured light transport. IEEE TPAMI, 2016. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 185 + ], + "type": "text", + "content": "[57] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual Coding to Probe Light Transport. ACM TOG, 2012. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 547, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 220 + ], + "type": "text", + "content": "[58] Andrew Payne, A Dorrington, and Michael Cree. Illumination waveform optimization for time-of-flight range imaging cameras. SPIE, 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 222, + 547, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 547, + 266 + ], + "type": "text", + "content": "[59] Georg Petschnigg, Richard Szeliski, Maneesh Agrawala, Michael Cohen, Hugues Hoppe, and Kentaro Toyama. Digital photography with flash and no-flash image pairs. ACM transactions on graphics (TOG), 2004. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 268, + 547, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 547, + 291 + ], + "type": "text", + "content": "[60] Dario Piatti, Fabio Remondino, and David Stoppa. State-of-the-Art of TOF Range-Imaging Sensors. 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 292, + 547, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 547, + 335 + ], + "type": "text", + "content": "[61] Dikpal Reddy, Ravi Ramamoorthi, and Brian Curless. Frequency-space Decomposition and Acquisition of Light Transport Under Spatially Varying Illumination. ECCV, 2012. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 337, + 546, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 546, + 381 + ], + "type": "text", + "content": "[62] Alexis Rochas, Michael Gosch, Alexandre Serov, Pierre Besse, Rade S. Popovic, T. Lasser, and Rudolph Rigler. First fully integrated 2-d array of single-photon detectors in standard cmos technology. IEEE PTL, 2003. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 383, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 546, + 415 + ], + "type": "text", + "content": "[63] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. Applied optics, 36(34):8871–8876, 1997. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 417, + 546, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 546, + 460 + ], + "type": "text", + "content": "[64] Eric W Rogala and Harrison H Barrett. Phase-shifting interferometry and maximum-likelihood estimation theory. ii. a generalized solution. Applied optics, 37(31):7253-7258, 1998. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 462, + 546, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 546, + 485 + ], + "type": "text", + "content": "[65] Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured light. IEEE/CVF CVPR, 2003. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 487, + 546, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 546, + 552 + ], + "type": "text", + "content": "[66] Rudolf Schwarte, Zhanping Xu, Horst-Guenther Heinol, Joachim Olk, Ruediger Klein, Bernd Buxbaum, Helmut Fischer, and Juergen Schulte. New electro-optical mixing and correlating sensor: facilities and applications of the photonic mixer device (PMD). In Sensors, Sensor Systems, and Sensor Data Processing, 1997. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 554, + 546, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 546, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 546, + 608 + ], + "type": "text", + "content": "[67] Pitchaya Sitthi-Amorn, Javier E Ramos, Yuwang Wangy, Joyce Kwan, Justin Lan, Wenshou Wang, and Wojciech Matusik. Multifab: a machine vision assisted platform for multi-material 3d printing. Acm Transactions on Graphics (Tog), 2015. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 610, + 546, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 610, + 546, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 610, + 546, + 643 + ], + "type": "text", + "content": "[68] Shuochen Su, Felix Heide, Gordon Wetzstein, and Wolfgang Heidrich. Deep end-to-end time-of-flight imaging. CVPR, 2018. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 644, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 644, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 644, + 546, + 667 + ], + "type": "text", + "content": "[69] Murali Subbarao and Gopal Surya. Depth from defocus: A spatial domain approach. IJCV, 1994. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "text", + "content": "[70] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G. Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature Comm., 2012. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8242" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 171 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[71] Federica Villa, Rudi Lussana, Danilo Bronzi, Simone Tisa, Alberto Tosi, Franco Zappa, Alberto Dalla Mora, Davide Contini, Daniel Durini, Sasha Weyers, and Werner Brockherde. Cmos imager with 1024 spads and tdcs for single-photon timing and 3-d time-of-flight. IEEE JSTQE, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 171 + ], + "type": "text", + "content": "[72] Peng Xiao, Mathias Fink, and A Claude Boccara. Full-field spatially incoherent illumination interferometry: a spatial resolution almost insensitive to aberrations. Optics letters, 2016. 4" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8243" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_content_list.json b/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1abf2d8f81ce696771da8f7ddcd43a4b19d50c26 --- /dev/null +++ b/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_content_list.json @@ -0,0 +1,1652 @@ +[ + { + "type": "text", + "text": "Switchable Representation Learning Framework with Self-compatibility", + "text_level": 1, + "bbox": [ + 119, + 130, + 848, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengsen Wu $^{1,2}$ , Yan Bai $^{2}$ , Yihang Lou $^{3}$ , Xiongkun Linghu $^{4}$ , Jianzhong He $^{3}$ , Ling-Yu Duan $^{1,2,*}$ $^{1}$ Peng Cheng Laboratory, China; $^{2}$ Peking University, China; $^{3}$ Huawei Technologies; $^{4}$ Tsinghua University, China", + "bbox": [ + 109, + 179, + 864, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sswu, yanbai, lingyu}@pku.edu.cn, {louyihang1, jianzhong.he}@huawei.com, lhxk20@mails.tsinghua.edu.cn", + "bbox": [ + 194, + 243, + 769, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 311, + 313, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-world visual search systems involve deployments on multiple platforms with different computing and storage resources. Deploying a unified model that suits the minimal-constrain platforms leads to limited accuracy. It is expected to deploy models with different capacities adapting to the resource constraints, which requires features extracted by these models to be aligned in the metric space. The method to achieve feature alignments is called \"compatible learning\". Existing research mainly focuses on the one-to-one compatible paradigm, which is limited in learning compatibility among multiple models. We propose a Switchable representation learning Framework with Self-Compatibility (SFSC). SFSC generates a series of compatible sub-models with different capacities through one training process. The optimization of sub-models faces gradients conflict, and we mitigate this problem from the perspective of the magnitude and direction. We adjust the priorities of sub-models dynamically through uncertainty estimation to co-optimize sub-models properly. Besides, the gradients with conflicting directions are projected to avoid mutual interference. SFSC achieves state-of-the-art performance on the evaluated datasets.", + "bbox": [ + 76, + 343, + 472, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 703, + 209, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual search systems are widely deployed, which recall the nearest neighbors in gallery features according to their distances to the query feature. Real-world visual search systems consist of multiple models deployed on different platforms [17, 38, 45] (e.g. clouds, mobiles, smart cameras), where different platforms interact with each other by visual features. Typically, as for person re-identification systems, images are captured and processed into features on edge sides. And such features are sent to the cloud side to compare with the database features for identifying", + "bbox": [ + 75, + 728, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a specific person by feature similarities. As diverse platforms meet different computing and storage resource limitations, deploying the unified model which suits the minimal-constrain platforms leads to a waste of resources on other platforms and limited accuracy. To make better use of resources and achieve higher accuracy, it is expected to deploy models with different capacities which adapt to the resource limitations. Such a solution requires compatibility among the models to satisfy their interaction requirements, which means that the similar data processed by different models are close to each other in the feature space while dissimilar data are far apart.", + "bbox": [ + 500, + 311, + 890, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve compatibility, compatible learning methods have been proposed. Existing research focuses on the one-to-one compatible paradigm [6,49,62], which constrains the learned features of a latter (learnable) model to be compatible with its previous (fixed) version. They are limited in achieving many-to-many compatibilities, which means any two in a series of models are compatible with each other.", + "bbox": [ + 496, + 500, + 890, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a Switchable representation learning Framework with Self-Compatibility (SFSC) to achieve many-to-many compatibilities. As shown in Figure 1, SFSC can deploy different sub-models to suit the diverse computing and storage resource limitation of different platforms. The compatibility between any two sub-models can be achieved, termed as \"self-compatibility\". However, there are gradients conflicts between sub-models as they are co-optimized during the training process. We summarize such conflicts into the gradient magnitude and gradient direction. Specifically, a gradient of a large magnitude will dominate the optimization process, which may lead to the overfitting of its corresponding sub-model and impair the improvements of other sub-models. To solve this problem, we estimate the uncertainty of different sub-models to weight the gradient, denoting the optimization priorities. Besides, as the sub-models may produce gradients in various directions, there is mutual interference between them. Thus the improvement in different sub-models may be overestimated", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15943", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70e765d57ccf0a68bef5a3b047029b61a69e6e0bb8e330762e2af7676e2b3878.jpg", + "image_caption": [ + "Figure 1. An example of SFSC in multi-platform deployments. The switchable neural network generates a series of sub-networks $\\{\\phi_1, \\phi_2, \\dots, \\phi_m\\}$ with different capacities for different computing resources. $\\forall i, j \\in \\{1, 2, \\dots, m\\}$ , $|\\phi_i(a) - \\phi_j(p)| < |\\phi_i(a) - \\phi_j(n)|$ where $(a, p)$ is a pair of samples with the same class ID, while $(a, n)$ is a pair of samples with different class IDs." + ], + "image_footnote": [], + "bbox": [ + 202, + 90, + 777, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or underestimated. To tackle such conflict, the gradients are projected onto planes that are orthogonal to each other.", + "bbox": [ + 75, + 313, + 468, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are summarized as follows:", + "bbox": [ + 75, + 343, + 468, + 375 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a switchable representation learning framework with self-compatibility (SFSC). SFSC generates a series of feature-compatible sub-models with different capacities for visual search systems, which can be deployed on different platforms flexibly.", + "- We resolve the conflicts between sub-models from the aspect of gradient magnitude and gradient direction. A compatible loss based on uncertainty estimation is proposed to guide optimization priorities and alleviate the imbalance of gradient magnitude between sub-models. An aggregation method based on the gradient projection is proposed to avoid mutual interference and find a generic optimal direction for all sub-models.", + "- SFSC achieves state-of-the-art performance on the evaluated benchmark datasets. Compared to deploying a unified model, adopting SFSC to obtain different sub-models can achieve $6\\%$ to $8\\%$ performance improvements on three different datasets." + ], + "bbox": [ + 94, + 382, + 468, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 686, + 220, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. One-to-one Compatible Learning Paradigm", + "text_level": 1, + "bbox": [ + 76, + 710, + 449, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a fixed old model, the one-to-one compatible learning methods aim at aligning the learnable models with the fixed old one in metric space, termed as \"backward compatible learning\". Existing methods focus on the optimization difficulties caused by the old model. Yan et al. [63] propose Focal Distillation to reduce the \"negative flips\" which means a test sample is correctly classified by the old model but incorrectly classified by the new one. Wu et al. [62] start from neighborhood structures and classification entropy to eliminate the effect of outliers in old features. Shen et al. [48] propose a pseudo classifier and enhance it with a", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "random walk algorithm to get rid of the old classifier. Different from these one-to-one compatible learning methods, this paper studies a new compatible learning paradigm that aims to learn many-to-many compatibility among multiple learnable models.", + "bbox": [ + 496, + 313, + 893, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Dynamic Neural Networks", + "text_level": 1, + "bbox": [ + 498, + 401, + 741, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic neural networks are able to adapt their structures or parameters to the input during inference [19]. In recent years, research in this field has mainly focused on dynamic depth [5, 18, 26, 54, 55, 58], dynamic width [3, 4, 11, 27, 29, 36], dynamic routing [22, 25, 28, 64], dynamic parameters [12, 15, 21, 51, 69], etc. In addition to hand-crafted methods, Neural Architecture Search (NAS) method is also adopted to train and optimize multiple sub-nets with various capacities in one-time training [7, 8, 44, 53]. Dynamic neural networks can allocate computations on demand for different platforms by selectively activating model components. They have been used in classification [41, 42, 59], segmentation [31, 34, 43], and detection tasks [12, 56, 69]. In general, current dynamic neural networks output with definite semantic information (e.g. class id, detection box) which are naturally interoperable. However, for retrieval tasks, the networks output with features, which requires feature compatibility.", + "bbox": [ + 496, + 425, + 893, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Gradient-based Multi-task Optimization", + "text_level": 1, + "bbox": [ + 498, + 710, + 846, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the gradient perspective, multi-task optimization is to aggregate gradients corresponding to different tasks and find a descent direction that decreases all objectives. Existing methods treat it as a nonlinear programming problem and take the multi-objective Karush-Kuhn-Tucker [33] (KKT) conditions to perform gradient projection and rescaling under different assumptions on tasks, e.g. multiple tasks which are equally important [10, 46, 66], main tasks with some auxiliary tasks [13, 57], a series of the tasks which are learned sequentially [14, 35, 37]. This paper treats optimizations of multiple sub-models as multiple tasks. As", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15944", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/87150a85c8cbcecf8d2b8dd20aa49ecb3141df1c0b332e505b983f8a2f8ccc58.jpg", + "image_caption": [ + "Figure 2. Overview of SFSC. We first transform the full modelcalculate the compatible learning loss $\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}$ on different sub-models with various capacitie. $\\phi$ into the sub-models. Then we calculate the compatible learning loss $\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}$ on different sub-models with various capacities. Finally, we optimize the sub-models by aggregating losses." + ], + "image_footnote": [], + "bbox": [ + 178, + 87, + 794, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the sub-models vary greatly in their capacity, the optimization of different models converges at different speeds. The challenge is to dynamically adjust the priorities of different sub-models to avoid the overfitting of someones as well as underfitting of others.", + "bbox": [ + 75, + 299, + 470, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Evidential Deep Learning", + "text_level": 1, + "bbox": [ + 76, + 387, + 312, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evidential deep learning [47] is used to model the uncertainty of deterministic neural networks. Compared with the Bayesian neural networks [32, 40], evidential deep learning methods have the advantage of low computational complexity, as there is no need for complex posterior probability estimates and multiple Mento Carlo samplings. Based on the DST [16] (Dempster-Shafer Theory of Evidence), evidential deep learning treats the output of the neural networks as subjective opinions and formalizes them as Dirichlet distribution. It has been used in classification tasks to overcome the overconfidence problem of softmax cross entropy [30], action recognition task to identify unknown samples [2], and multi-view learning task [20] to dynamically integrate different views. This paper treats the output features of different sub-models as evidence to access uncertainty and relates the priority of different sub-models with the obtained uncertainty.", + "bbox": [ + 75, + 412, + 473, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 685, + 174, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will describe the proposed SFSC method. We first introduce the construction of switchable neural networks. Then, we analyze the difficulties for optimizations and propose our compatible loss as well as the gradient aggregation method.", + "bbox": [ + 75, + 712, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 800, + 192, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SFSC involves the training process of multiple sub-models whose features are compatible. Sub-models with different capacities are transformed from the full model $\\phi$ according to the crop ratio list $W$ . We termed the crop ratio list of each sub-model as $W = \\{\\gamma_{1},\\gamma_{2},\\dots,\\gamma_{m}\\}$ , where $\\gamma_{j}$", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "denotes the ratio of the width of the sub-model $\\phi_j$ to the full model $\\phi$ .", + "bbox": [ + 496, + 299, + 890, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, for each batch input, SFSC first calculates the original loss $L_{ori}$ . The compatible loss $\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}$ is calculated on different sub-models. By aggregating losses $\\{L_{ori}, L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}$ , SFSC computes the general gradient and updates the full model $\\phi$ by gradient descent methods.", + "bbox": [ + 496, + 330, + 890, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Switchable Neural Network for Sub-models Construction", + "text_level": 1, + "bbox": [ + 498, + 433, + 890, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first convert the traditional convolutional neural network into its corresponding switchable neural network. As the platforms generally perform model inference layer by layer, the resource limitation is more related to the widths of layers than the number of layers. Here we perform channel pruning on $\\phi$ with crop ratio $\\gamma_{j}$ to obtain sub-model $\\phi_{j}$ . Specifically, the classic convolutional neural network is mainly composed of some basic modules, such as convolution, fully connected layer, and batch normalization. We modify these basic modules as follows:", + "bbox": [ + 496, + 473, + 892, + 625 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) For convolution (Conv) and fully connected (FC) layers, if their output is not the final feature for retrieval, we perform pruning in both input and output channels. Otherwise, pruning is only performed in input channels.", + "2) For batch normalization (BN), we assign an independent BN to the sub-model $\\phi_j$ . As the intermediate features extracted by different sub-models vary greatly in distributions [65], channel pruning can not be performed on BN. Fortunately, the parameters of BN only occupy a small part of all parameters of the entire network (shown in Table 1). Such an operation only results in a slight increase in storage resources in training progress but does not incur additional deployment costs." + ], + "bbox": [ + 496, + 626, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through such design, we obtain the sub-model $\\phi_j$ with the following characteristics: 1) The features extracted by different sub-models have the same dimensions. 2) The ratio of layer widths in the sub-model $\\phi_j$ to that in the full model $\\phi$ is around $\\gamma_j$ . 3) The sub-model $\\phi_j$ reuses the pa", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4300379dd771e7b9aae344c9a8fafd844b62fce87b21c233798285035fa6fc2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NetWorksResNet-18 [23]ResNet-50 [23]MobileNet V2 [45]ShuffleNet [67]
Conv and FC11,679,91225,503,9123,470,8705,402,826
99.918%99.792%99.027%99.102%
BN9,60053,12034,11248,960
0.0820%0.208%0.973%0.898%
", + "bbox": [ + 78, + 88, + 470, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Number and proportion of different modules' parameters in different network structures", + "bbox": [ + 76, + 136, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rameters from the full model $\\phi$ except for BN.", + "bbox": [ + 76, + 191, + 382, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Aggregation Method based on the Gradients Projection", + "text_level": 1, + "bbox": [ + 76, + 215, + 468, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the gradients corresponding to different sub-models may have different directions, simply aggregating the gradients by summation may cause mutual interference, which means the improvements of sub-models are overestimated or underestimated, as shown in Figure 3.", + "bbox": [ + 75, + 253, + 468, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by multi-task learning [1,46,66], we tackle such a problem by projecting conflicting gradients and keeping the components perpendicular to each other. Given a pair of gradients $(g_i, g_j)$ corresponding to sub-model $\\phi_i$ and $\\phi_j$ , if the cosine of their direction $\\cos(g_i, g_j) < 0$ , they are conflict with each other. For a pair of conflicting gradients $(g_i, g_j)$ , we project them to the orthogonal plane of each other, respectively, as shown in Equation 1.", + "bbox": [ + 75, + 330, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {j} ^ {\\prime} = g _ {i} - \\frac {g _ {j} * g _ {i}}{\\left\\| g _ {i} \\right\\| ^ {2}} * g _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 450, + 468, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Such projection results in a direction that improves both the $i$ -th and $j$ -th sub-model simultaneously.", + "bbox": [ + 76, + 491, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c2bb56d278f8395edd91472d8d1c17a4c4037834b3b88552cef4823db39d1e4.jpg", + "image_caption": [ + "Figure 3. An example of gradient conflict between pair $(g_i, g_j)$ . Simply taking the summation aggregation results in the elimination of the improvement of the $i$ -th sub-model. Our aggregation module ensures the generic direction to improve both the $i$ -th and $j$ -th sub-model." + ], + "image_footnote": [], + "bbox": [ + 86, + 532, + 459, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Compatible Loss based on the Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 76, + 696, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The aggregation method aims to optimize all sub-models coordinately. However, it may not work as expected when facing the imbalance of gradient magnitude between sub-models, as shown in Figure 4. Therefore, it is critical to design a compatible loss.", + "bbox": [ + 75, + 734, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As sub-models are co-optimized, constraining them with existing compatible loss [6, 49] faces conflicts in the magnitude of gradients. As the sub-models vary in capacities, their optimization process converges at different speeds. The small-capacity sub-models will converge faster. Therefore, the small-capacity sub-model may have reached its", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/43ead0d0647de42c59033410ee596870fe0267f53c8438e8b673842a265dda17.jpg", + "image_caption": [ + "Figure 4. An example of gradient conflict between pair $(g_i, g_j)$ when there is an imbalance of gradient magnitude. The $i$ -th sub-model contributes little to the optimization even using gradient projection." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 879, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "performance ceiling while still producing losses of large magnitude, resulting in the overfitting of such sub-model and weakening of other sub-models. Intuitively, a sub-model with low uncertainty tends to reach its performance ceiling. Therefore, it is expected to adaptively adjust the optimization priorities of each sub-models through their uncertainty dynamically during the training process.", + "bbox": [ + 496, + 238, + 890, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we propose a compatible loss based on the uncertainty estimation. Inspired by Dempster-Shafer Theory of evidence [16] (DST), we measure the uncertainty through belief, where high uncertainty corresponds to the model with small capacities. Typically, the belief is measured by a set of mass $B = \\{b_{1}, b_{2}, \\ldots, b_{C}\\}$ and an overall uncertainty mass $u$ , where $b_{c}$ denotes the belief degree belonging to the class $c$ . $B$ and $u$ are related to a Dirichlet distribution, which satisfy Equation 2.", + "bbox": [ + 496, + 344, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu + \\sum_ {c = 1} ^ {C} b _ {c} = 1, u \\geq 0, b _ {c} \\geq 0. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 482, + 890, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The belief is estimated by the features extracted by sub-models. Given sample $i$ , its features extracted by the $j$ -th sub-model termed $\\phi_j(i)$ . The evidence $E_i^j = \\{e_{i1}^j, e_{i2}^j, \\dots, e_{iC}^j\\}$ is transformed by $\\phi_j(i)$ through linear projection and activation function (i.e., $\\exp(\\cdot)$ ), where the linear projection layers share parameters with the classifier of the $1.0 \\times$ model $\\phi$ . The belief of sample $i$ on the $j$ -th sub-model is computed by Equation 3, where $\\alpha_i^j = \\{\\alpha_{i1}^j, \\alpha_{i2}^j, \\dots, \\alpha_{iC}^j\\}$ is the Dirichlet intensity on all classes.", + "bbox": [ + 498, + 526, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i c} ^ {j} = e _ {i c} ^ {j} + 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 667, + 750, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} ^ {j} = \\sum_ {c = 1} ^ {C} \\alpha_ {i c} ^ {j}, b _ {i c} ^ {j} = \\frac {\\alpha_ {i c} ^ {j} - 1}{S _ {i} ^ {j}}, u _ {i} ^ {j} = \\frac {C}{S _ {i} ^ {j}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 689, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With $\\alpha_{i}^{j}$ , the Dirichlet distribution $\\mathrm{D}(p_i|\\alpha_i^j)$ is obtained, where $p_i$ is the class assignment probabilities on a simplex. Inspired by the Cross-Entropy loss, we integrate the probability $p_i$ on $\\mathrm{D}(p_i|\\alpha_i^j)$ , as shown in Equation 4.", + "bbox": [ + 496, + 732, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\alpha_ {i} ^ {j}) = \\int [ \\sum_ {c = 1} ^ {C} - y _ {i c} ^ {j} \\log (p _ {i c} ^ {j}) ] \\frac {1}{\\beta (\\alpha_ {i} ^ {j})} \\Pi_ {c = 1} ^ {C} p _ {i c} ^ {j} ^ {(\\alpha_ {i c} - 1)} \\mathrm {d} p _ {i} ^ {j} \\\\ = \\sum_ {c = 1} ^ {C} y _ {i c} ^ {j} \\left(\\psi \\left(S _ {i} ^ {j}\\right) - \\psi \\left(\\alpha_ {i c} ^ {j}\\right)\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 796, + 890, + 896 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15946", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot)$ is the digamma function. Since Equation 4 may converge to a trivial solution $\\mathrm{D}(p_i|1)$ during optimization, we penalize the distance between $\\mathrm{D}(p_i|1)$ and $\\mathrm{D}(p_i|\\alpha_i^j)$ as the regularization. The complete compatible loss $L_{c_j}$ is shown in Equation 5. $L_{c_j}$ adaptively adjusts the magnitude of gradients through their corresponding uncertainty.", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K L \\left[ \\mathrm {D} \\left(p _ {i} \\mid \\alpha_ {i} ^ {j}\\right) | | \\mathrm {D} \\left(p _ {i} \\mid \\mathbf {1}\\right) \\right] = \\log \\left(\\frac {\\sum_ {c = 1} ^ {C} \\Gamma \\left(\\alpha_ {i c} ^ {j}\\right)}{\\Gamma (C) \\prod_ {c = 1} ^ {C} \\left(\\alpha_ {i c} ^ {j}\\right)}\\right) \\\\ + \\sum_ {c = 1} ^ {C} \\left(\\alpha_ {i c} ^ {j} - 1\\right) \\left[ \\psi \\left(\\alpha_ {i c} ^ {j} - \\psi \\left(\\sum_ {c = 1} ^ {C} \\alpha_ {i c} ^ {j}\\right)\\right) \\right], \\tag {5} \\\\ L _ {c _ {j}} (i) = \\mathcal {L} \\left(\\alpha_ {i} ^ {j}\\right) + \\lambda * K L \\left[ \\mathrm {D} \\left(p _ {i} \\mid \\alpha_ {i} ^ {j}\\right) | | \\mathrm {D} \\left(p _ {i} \\mid \\mathbf {1}\\right) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 203, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further conduct a gradient analysis for the proposed compatible loss. The Cross-Entropy loss is written as:", + "bbox": [ + 76, + 318, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i c} = \\frac {e ^ {z _ {i c}}}{\\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}}, L _ {c e} = \\sum_ {c = 1} ^ {C} - y _ {i c} \\log \\left(p _ {i c}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 351, + 468, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{i}$ is the classification vector of sample $i$ outputted by the model. The corresponding gradient of Equation 6 can be computed as:", + "bbox": [ + 76, + 400, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L _ {c e}}{\\partial z _ {i c}} = \\frac {e ^ {z _ {i c}}}{\\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}} - y _ {i c}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 455, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The corresponding gradient of the proposed compatible loss in 5 can be computed as:", + "bbox": [ + 76, + 500, + 468, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial L _ {c _ {j}} \\left(\\alpha_ {i} ^ {j}\\right)}{\\partial \\alpha_ {i c} ^ {j}} = \\psi^ {\\prime} \\left(\\alpha_ {i c} ^ {j}\\right) \\left[ \\lambda \\left(\\alpha_ {i c} ^ {j} - 1\\right) - \\mathbf {y} _ {i c} \\right] \\tag {8} \\\\ + \\psi^ {'} (S _ {i} ^ {j}) [ - \\lambda (S _ {i} ^ {j} - C) + 1 ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 531, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We take $\\frac{1}{\\alpha_{ic}^j}$ as the approximation of $\\psi'(\\alpha_{ic}^j)$ and obtain Equation 9.", + "bbox": [ + 76, + 598, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial L _ {c _ {j}} (\\alpha_ {i} ^ {j})}{\\partial \\alpha_ {i c} ^ {j}} \\approx \\frac {1}{\\alpha_ {i c} ^ {j}} [ \\lambda (\\alpha_ {i c} ^ {j} - 1) - \\mathbf {y} _ {i c} ] \\\\ + \\frac {1}{S _ {i} ^ {j}} [ - \\lambda (S _ {i} ^ {j} - C) + 1 ] \\tag {9} \\\\ = - (\\frac {\\lambda + y _ {i c}}{\\alpha_ {i c} ^ {j}} + \\frac {\\lambda C + 1}{S _ {i} ^ {j}}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 635, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term corresponds to the Cross-Entropy loss in Equation 7, and the second term corresponds to the uncertainty estimation. As the uncertainty $u_{i}^{j} = \\frac{C}{S_{i}^{j}} \\propto \\frac{1}{S_{i}^{j}}$ , thus higher $u_{i}^{j}$ would result in a gradient with larger magnitude through $\\frac{\\lambda C + 1}{S_{i}^{j}}$ . Therefore, the magnitude of gradients corresponding to different sub-models is dynamically adjusted to avoid someone dominating the optimization process.", + "bbox": [ + 76, + 753, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Such compatible loss can work well with gradient projection, as shown in Algorithm 1.", + "bbox": [ + 73, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Training Process of SFSC" + ], + "code_body": "Require: batch input $\\mathcal{B}$ the crop list $W = \\{\\gamma_0,\\gamma_1,\\dots ,\\gamma_m\\}$ ,the full model $\\phi$ \n1: $g_{ori}\\gets \\frac{\\delta L_{ori}(\\phi,\\mathcal{B})}{\\delta\\phi}$ \n2: for $\\gamma_{j}\\in W$ do \n3: Switch to the sub-model $\\phi_j$ \n4: $g_{j}\\leftarrow \\frac{\\delta L_{c_{j}}(\\phi_{j},\\mathcal{B})}{\\delta\\phi}$ \n5: end for \n6: $G,G^{p}\\gets \\{g_{ori},g_{0},\\ldots g_{m}\\}$ \n7: for $g_{a}\\in G^{p}$ do \n8: Shuffle(G) \n9: for $g_{b}\\in G$ do \n10: $d = g_{a}*g_{b}$ \n11: if $d < 0$ then //have conflicts \n12: $g_{a}\\gets g_{a} - \\frac{g_{b}*d}{||g_{b}||^{2}}$ \n13: end if \n14: end for \n15: end for \n16: return Update $\\phi$ by $\\Delta \\theta = -\\sum_{g_a\\in GP}g_a$", + "bbox": [ + 500, + 109, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In general, SFSC aims to achieve the Pareto optimality [9] among the optimization of sub-models. The proposed compatible loss adaptively adjusts the magnitude of gradients through uncertainty to control the optimization priorities of sub-models dynamically. Besides, the aggregation method performs the gradient projection to find a generic direction to improve all sub-models.", + "bbox": [ + 498, + 364, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 486, + 632, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Configuration", + "text_level": 1, + "bbox": [ + 500, + 512, + 753, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed SFSC on person ReID datasets, i.e., Market-1501 [68], MSMT17 [61], and one vehicle ReID dataset VeRi-776 [60]. Market1501 consists of 32,668 annotated images of 1,501 identities shot from 6 cameras. MSMT17 consists of 126,441 bounding boxes of 4,101 identities taken by 15 cameras. VeRi-776 contains over 50,000 images of 776 vehicles captured by 20 cameras.", + "bbox": [ + 496, + 536, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison methods. For the one-to-one compatible learning paradigm, BCT [49] and Asymmetric [6] are adopted, as they are representatives of compatible learning through classification losses and embedding losses. BCT and Asymmetric can only obtain sub-models in a one-to-one compatible way. Specifically, given the fixed largest capacity model which is trained without compatible regularization, we train a series of small-capacity models to be compatible with it, respectively. Besides, we also incorporate BCT/Asymmetric loss in our proposed self-compatible learning paradigm for a fair comparison. We take the compatible loss in BCT/Asymmetric to constrain each sub-model compatible with the full model, and use summation to aggregate losses from sub-models, termed as \"BCT-S\" and \"Asymmetric-S\".", + "bbox": [ + 496, + 657, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Mean average precision (mAP)", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/774c9c30a758a1d2a80fe73b9fbf057c3e30a2c8fa2c996809d710721eb61475.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M(φq, φg) φq φqφ1/16φ1/4φ9/16φφ1/16φ1/4φ9/16φ
Unified ModelOurs
φ1/1655.25---58.1962.8863.4869.43
φ1/4-67.48--61.2470.7471.3776.37
φ9/16--71.25-62.4371.2572.0677.26
φ---80.9168.0374.7675.6781.43
BCT-SAsymmetric-S
φ1/1654.4058.1060.2261.2948.7951.9654.0956.31
φ1/455.7166.3868.2969.7152.2462.4564.9867.23
φ9/1656.5068.3068.4271.9954.4464.7467.9771.28
φ61.7969.7371.9373.6156.5467.3071.9178.38
BCTAsymmetric
φ1/1655.5556.6459.4060.3555.8350.8553.7155.39
φ1/455.8365.7467.1068.5452.6667.0462.0866.59
φ9/1655.6668.0267.4970.9354.4862.3366.3370.14
φ58.6969.4870.04-56.0866.6669.71-
", + "bbox": [ + 181, + 77, + 785, + 347 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7fe1f7ded5631968cf2e5afbbee07ed2dc5b90bc4ea92c2509f51cf247632489.jpg", + "table_caption": [ + "Table 2. Baseline performance comparison on Market1501 (mAP). \"Unified Model\" are models without any compatible regularization." + ], + "table_footnote": [], + "table_body": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
MSMT17
Unified Model28.8211.6948.6522.8657.5630.0670.2243.06
BCT-S46.4920.1156.2228.1358.1130.1458.7630.99
Asymmetric-S32.6413.3449.6224.0457.1830.0466.0840.39
Ours56.2228.1663.2935.3265.1337.7470.3843.89
VERI-776
Unified Model81.0744.7985.3053.6388.6955.4292.2066.50
BCT-S86.2549.3887.0857.0390.4258.2086.7358.46
Asymmetric-S68.7542.1684.2956.5588.0461.3789.8265.35
Ours79.7655.0489.9462.2890.4862.7292.3266.55
", + "bbox": [ + 166, + 378, + 802, + 559 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/62fa9aa18a85316a9a075c33bceb5295a03f2885a771210fb392d684f672ef39.jpg", + "table_caption": [ + "Table 3. Performance comparison on MSMT17 and VeRi-776 datasets." + ], + "table_footnote": [], + "table_body": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
Circle Loss
Unified Model70.2550.7972.6660.4378.2164.6687.3874.06
BCT-S63.1842.4572.7452.0477.3859.4987.5374.58
Asymmetric-S69.6648.2177.8557.0280.5261.8186.9073.08
Ours73.9954.6180.5863.4682.3967.0288.9577.05
Softmax + Triplet
Unified Model77.5555.8283.9765.3087.8970.6990.9478.30
BCT-S81.6259.5887.8669.0588.7571.3890.1174.71
Asymmetric-S75.5652.6085.3664.2186.9168.8390.4577.34
Ours87.3666.2190.9476.6491.2177.9392.6481.46
", + "bbox": [ + 166, + 589, + 800, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Performance comparison between different loss functions on Market1501.", + "bbox": [ + 238, + 771, + 730, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and top-1 accuracy (R1) are adopted to evaluate the performances of retrieval. Define $M(\\phi_{a},\\phi_{b})$ as the performances of retrieval where the query features are extracted by model $\\phi_{a}$ and gallery features are extracted by model $\\phi_{b}$ . We set $\\phi_{a}$ and $\\phi_{b}$ as different sub-models to evaluate the perfor", + "bbox": [ + 76, + 811, + 470, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mance of compatible learning.", + "bbox": [ + 500, + 811, + 702, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The implementations of all compared methods and SFSC are based on the FastReLU [24]. The default configuration named \"bag of tricks\" [39] is adopted. $\\lambda$ is set to 0.2 for all experiments.", + "bbox": [ + 498, + 839, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15948", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bb8932007960713cf71bf46c470153f192a06aca1a6c9ddeb944fc02350a30aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
ResNet-50
Unified Model85.2765.4488.8472.0790.876.4092.2282.13
BCT-S89.3172.0991.0077.2291.3678.2791.8679.40
Asymmetric-S85.0467.3990.2676.2891.6078.5391.9582.79
Ours90.4176.4091.0378.2091.6678.7693.4483.66
MobileNet V2
Unified Model69.0041.1489.5872.8591.5477.9992.2879.50
BCT-S80.4655.1788.4870.6889.8873.7289.9673.87
Asymmetric-S67.2843.6486.9768.9589.3173.8290.7478.45
Ours83.1163.8290.5376.5291.1878.3092.6181.34
ShuffleNet
Unified Model82.5460.1289.8564.4591.9572.9684.3879.13
BCT-S85.5165.7688.4266.8589.7370.5986.6774.41
Asymmetric-S82.3062.5084.4157.4388.4266.4378.9274.92
Ours89.2573.0989.9374.8689.9375.1492.8181.97
", + "bbox": [ + 165, + 87, + 802, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Baseline Comparisons", + "text_level": 1, + "bbox": [ + 76, + 381, + 284, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct baseline experiments on Market1501. We take \"Softmax + Cross Entropy\" as the loss function, ResNet-18 [23] as the backbone, and $W = \\{0.25 \\times, 0.50 \\times, 0.75 \\times\\}$ as the crop ratio list of the sub-models. Notes that for a model with crop ratio $\\gamma_i$ , the ratio of its parameters to the full model is $\\gamma_i^2$ . Therefore, the ratio of the number of parameters in the $0.25 \\times$ , $0.50 \\times$ and $0.75 \\times$ sub-model to that in the full model are $\\frac{1}{16}, \\frac{1}{4}, \\frac{9}{16}$ , respectively.", + "bbox": [ + 75, + 406, + 470, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison methods. As for one-to-one compatible learning methods BCT and Asymmetric, the compatibility does not outperform that achieved by their corresponding self-compatible methods especially when the models are not constrained to be compatible directly. For self-compatible learning methods, BCT-S and Asymmetric-S, the conflicts among sub-models lead to a limited average accuracy.", + "bbox": [ + 75, + 542, + 468, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our Method (SFSC). As we resolve the conflicts among different sub-models, our methods can achieve many-to-many compatibility among all the sub-models, which significantly outperforms all comparison methods.", + "bbox": [ + 75, + 666, + 468, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Performances under Different Settings", + "text_level": 1, + "bbox": [ + 76, + 738, + 410, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different datasets. We evaluate each self-compatible learning method on the other person ReID datasets MSMT17, and the vehicle ReID dataset VeRi-776 under baseline configurations, as shown in Table 3. The results on the two datasets with larger scales are consistent with that on the Market1501. Our methods outperform others significantly for the small capacity sub-model (e.g. the $0.25 \\times$ sub-model $\\phi_{\\frac{1}{4}}$ ).", + "bbox": [ + 75, + 763, + 468, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different loss functions. We replace the original loss", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "function $L_{ori}$ from \"Softmax\" to \"Softmax + Triplet\" and \"Circle Loss\" [52] to evaluate the effectiveness of each method, while Triplet loss is a kind embedding loss and circle loss is a kind of classification loss. The results in Table 4 show that the accuracy of BCT-S and Asymmetric-S drops greatly when training with circle loss, as Our method still stably achieves the best performance.", + "bbox": [ + 496, + 382, + 890, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different model architectures. We also study whether each method can be stably applied to various model architectures, i.e., ResNet-50 [23], MobileNet V2 [45] and ShuffleNet [67]. ShuffleNet and MobileNet are both designed especially for edge devices, which meet the application scenarios of self-compatible learning. It is shown in Table 5 that BCT-S and Asymmetric-S do not achieve higher accuracy than deploying a unified model with ShuffleNet and MobileNet (e.g. for the $0.75 \\times$ sub-model $\\phi_{\\frac{9}{16}}$ ). And our methods always improve the accuracy in all cases.", + "bbox": [ + 496, + 489, + 892, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 657, + 653, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We replace the proposed compatible loss and aggregation module with existing compatible loss BCT [50] and summation aggregation to test the proposed two methods.", + "bbox": [ + 496, + 681, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e56a94025003d2aa1243aedc0abd9773b137c4af0573f18da42cde2c7d570d1.jpg", + "table_caption": [ + "Table 5. Performance comparison between different model architectures on Market1501." + ], + "table_footnote": [], + "table_body": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
Unified Model55.2567.4871.2580.91
w/o Ours Loss66.4174.5876.8677.87
w/o Ours Aggregation68.0873.7674.8475.38
SFSC69.4376.3777.2681.43
", + "bbox": [ + 500, + 743, + 887, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation Study on Market1501.", + "bbox": [ + 573, + 804, + 815, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 shows that the proposed compatible loss mainly improves the retrieval accuracy of the small capacity sub-models while the proposed aggregation method can ensure the performance of the full model.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15949", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0d939f33c064f2251efe58c9745ee250bf0ecf1f8edcc940cf211ac8366b115.jpg", + "image_caption": [ + "Figure 5. Gradient visualization on Market1501-R18 baseline. As for the original cross-entropy loss (BCT [49]), the magnitude of the $0.25 \\times$ sub-model is much larger than that of the $0.75 \\times$ sub-model, especially at the early stage of the training process ( $0 \\sim 1000$ iterations). Therefore, the $0.25 \\times$ sub-model dominates the training progress and the improvements of the $0.75 \\times$ sub-models are eliminated. While our method ensures co-optimization through uncertainty estimation." + ], + "image_footnote": [], + "bbox": [ + 187, + 88, + 478, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0418f3bd8988d888bfa2d350dd9a8267a43fc75606e329ca551bcc3c85dc09e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 88, + 781, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Hyperparameter Analysis", + "text_level": 1, + "bbox": [ + 76, + 351, + 313, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we verify the impact of hyperparameter settings on the performance of the proposed method. The key hyperparameter is the crop ratio list for sub-models $W$ . Intuitively, if the number of $W$ increases or the minimal crop ratio decreases, the compatibility constraints on model $\\phi$ would be more strict, thereby increasing the difficulty of compatible learning. Therefore, we take the baseline configuration and set $W$ with different values.", + "bbox": [ + 75, + 375, + 470, + 494 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5bcc03e90ae30d84a52a2a8b9de23054d2265897722e2aa15277f45d0dce696a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. Performance comparison with $W = \\{0.25\\times ,0.45\\times ,0.65\\times ,0.85\\times \\}$ on Market1501 (mAP)." + ], + "table_body": "
M(φ1/16, φ)M(φ81/100, φ)M(φ169/400, φ)M(φ283/400, φ)M(φ, φ)
Unified Model55.2562.5469.5472.5780.27
BCT-S61.4568.6373.4272.8075.39
Asymmetric-S62.8769.4575.2073.1474.60
Ours69.6375.8177.1077.3680.59
", + "bbox": [ + 78, + 500, + 470, + 556 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/539ae55da7a03421439d23783534022ff69753f5c946fb8771684273492d339f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 8. Performance comparison with $W = \\{0.10\\times ,0.25\\times ,0.75\\times \\}$ on Market1501 (mAP)." + ], + "table_body": "
M(φ1/100, φ)M(φ1/16, φ)M(φ9/16, φ)M(φ, φ)
Unified Model40.4055.2571.2580.27
BCT-S42.3459.7571.7376.39
Asymmetric-S42.5858.2070.5075.42
Ours45.4468.1878.0081.46
", + "bbox": [ + 78, + 595, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7 shows the performances with the number of $W$ increases, while Table 8 shows the performances with the minimum among $W$ decreases. The result of the comparison methods has fluctuated to varying degrees. SFSC still achieves the best retrieval accuracy.", + "bbox": [ + 75, + 686, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Visualization", + "text_level": 1, + "bbox": [ + 76, + 771, + 215, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the norm of the gradient during the training progress on Market1501-R18 baseline, as shown in Figure 5. We also visualize the cumulative number of gradient pairs which are conflict with each other during the training process. As shown in Figure 6, our method has found a generic direction to improve all sub-models. Therefore, there are almost no conflict pairs after 5000 iterations. As", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9bdcb1363751d8acb4ec965101591a89329bf776554b19f20be4e5544a4311b1.jpg", + "image_caption": [ + "Figure 6. The cumulative number of conflict pairs during the training process." + ], + "image_footnote": [], + "bbox": [ + 535, + 352, + 856, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for the summation aggregation, it faces fierce conflicts especially after 10000 iterations, resulting in the improvement of different sub-models overestimated or underestimated.", + "bbox": [ + 498, + 566, + 892, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 625, + 619, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a Switchable Representation Learning Framework with Self Compatibility (SFSC) for multi-platform model collaboration. SFSC enables us to obtain models with various capacities to fit different computing and storage resource constraints on diverse platforms. It finds a generic direction to improve all sub-models by uncertainty estimation and gradient projection. SFSC achieves state-of-the-art performances and shows convincing robustness in different case studies.", + "bbox": [ + 498, + 651, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 800, + 679, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, the Basic and Frontier Research Project of PCL, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation.", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15950", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lu Bai, Yew-Soon Ong, Tiantian He, and Abhishek Gupta. Multi-task gradient descent for multi-task learning. Memetic Computing, 12(4):355-369, 2020. 4", + "[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13349-13358, 2021. 3", + "[3] Emmanuel Bengio, Pierre-Luc Bacon, Joelle Pineau, and Doina Precup. Conditional computation in neural networks for faster models. arXiv preprint arXiv:1511.06297, 2015. 2", + "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 2", + "[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2", + "[6] Mateusz Budnik and Yannis Avrithis. Asymmetric metric learning for knowledge transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2021. 1, 4, 5", + "[7] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. 2", + "[8] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 2", + "[9] Yair Censor. Pareto optimality in multiobjective problems. Applied Mathematics and Optimization, 4(1):41-59, 1977. 5", + "[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International Conference on Machine Learning, pages 794-803. PMLR, 2018. 2", + "[11] Kyunghyun Cho and Yoshua Bengio. Exponentially increasing the capacity-to-computation ratio for conditional computation in deep learning. arXiv preprint arXiv:1406.7362, 2014. 2", + "[12] Jifeng Dai, Haozhi Qi, Yuwen Xiong, and Yi Li. Guodong, zhang, han hu, and yichen wei. Deformable convolutional networks. In, ICCV, 2(3):7, 2017. 2", + "[13] Yunshu Du, Wojciech M Czarnecki, Siddhant M Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224, 2018. 2", + "[14] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, pages 3762-3773. PMLR, 2020. 2", + "[15] Hang Gao, Xizhou Zhu, Steve Lin, and Jifeng Dai. Deformable kernels: Adapting effective receptive fields for object deformation. arXiv preprint arXiv:1910.02940, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jean Gordon and Edward H Shortliffe. The dampster-shafer theory of evidence. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, 3:832-838, 1984. 3, 4", + "[17] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12259-12269, 2021. 1", + "[18] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2", + "[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2", + "[20] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 3", + "[21] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2", + "[22] Hussein Hazimeh, Natalia Ponomareva, Petros Mol, Zhenyu Tan, and Rahul Mazumder. The tree ensemble layer: Differentiability meets conditional computation. In International Conference on Machine Learning, pages 4138-4148. PMLR, 2020. 2", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 4, 7", + "[24] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631, 6(7):8, 2020. 6", + "[25] Thomas M Hehn, Julian FP Kooij, and Fred A Hamprecht. End-to-end learning of decision trees and forests. International Journal of Computer Vision, 128(4):997-1011, 2020. 2", + "[26] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2", + "[27] Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2761, 2018. 2", + "[28] Yani Ioannou, Duncan Robertson, Darko Zikic, Peter Kontschieder, Jamie Shotton, Matthew Brown, and Antonio Criminisi. Decision forests, convolutional networks and the models in-between. arXiv preprint arXiv:1603.01250, 2016. 2", + "[29] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Taejong Joo, Uijung Chung, and Min-Gwan Seo. Being bayesian about categorical probability. In International Conference on Machine Learning, pages 4950-4961. PMLR, 2020. 3", + "[31] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 2", + "[32] Igor Kononenko. Bayesian neural networks. Biological Cybernetics, 61(5):361-370, 1989. 3", + "[33] Harold W Kuhn and Albert W Tucker. Nonlinear programming. In Traces and emergence of nonlinear programming, pages 247-258. Springer, 2014. 2", + "[34] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2", + "[35] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Trgp: Trust region gradient projection for continual learning. arXiv preprint arXiv:2202.02931, 2022. 2", + "[36] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In Proceedings of the IEEE international conference on computer vision, pages 2736-2744, 2017. 2", + "[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2", + "[38] Yihang Lou, Ling-Yu Duan, Shiqi Wang, Ziqian Chen, Yan Bai, Changwen Chen, and Wen Gao. Front-end smart visual sensing and back-end intelligent analysis: A unified infrastructure for economizing the visual system of city brain. IEEE Journal on Selected Areas in Communications, 37(7):1489-1503, 2019. 1", + "[39] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6", + "[40] David JC MacKay. Bayesian neural networks and density networks. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 354(1):73-80, 1995. 3", + "[41] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. Advances in neural information processing systems, 27, 2014. 2", + "[42] Amir Rosenfeld and Shimon Ullman. Visual concept recognition and localization via iterative introspection. In *Asian Conference on Computer Vision*, pages 264-279. Springer, 2016. 2", + "[43] Abhijit Guha Roy, Nassir Navab, and Christian Wachinger. Concurrent spatial and channel 'squeeze & excitation' in fully convolutional networks. In International conference on medical image computing and computer-assisted intervention, pages 421-429. Springer, 2018. 2", + "[44] Manas Sahni, Shreya Varshini, Alind Khare, and Alexey Tumanov. Compofa: Compound once-for-all networks" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for faster multi-platform deployment. arXiv preprint arXiv:2104.12642, 2021. 2", + "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 4, 7", + "[46] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2, 4", + "[47] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in Neural Information Processing Systems, 31, 2018. 3", + "[48] Yantao Shen, Fanzi Wu, and Ying Shan. Non-inherent feature compatible learning. 2020. 2", + "[49] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 1, 4, 5, 8", + "[50] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 7", + "[51] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 2", + "[52] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 7", + "[53] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 2", + "[54] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2", + "[55] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2", + "[56] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 2", + "[57] Xinyi Wang, Ankur Bapna, Melvin Johnson, and Orhan Firat. Gradient-guided loss masking for neural machine translation. arXiv preprint arXiv:2102.13549, 2021. 2", + "[58] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15952", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2", + "[59] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. Advances in Neural Information Processing Systems, 33:2432-2444, 2020. 2", + "[60] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. Proceedings of the IEEE International Conference on Computer Vision, pages 379–387, 2017. 5", + "[61] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 5", + "[62] Shengsen Wu, Liang Chen, Yihang Lou, Yan Bai, Tao Bai, Minghua Deng, and Lingyu Duan. Neighborhood consensus contrastive learning for backward-compatible representation. arXiv preprint arXiv:2108.03372, 2021. 1, 2", + "[63] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021. 2", + "[64] Zhicheng Yan, Hao Zhang, Robinson Piramuthu, Vignesh Jagadeesh, Dennis DeCoste, Wei Di, and Yizhou Yu. Hd-cnn: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision, pages 2740-2748, 2015. 2", + "[65] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018. 3", + "[66] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 4", + "[67] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4, 7", + "[68] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5", + "[69] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9308-9316, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 853 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15953", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_model.json b/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1897bc8b37c0e92e9e5af0a0ecbf069c89baf92 --- /dev/null +++ b/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.849, + 0.154 + ], + "angle": 0, + "content": "Switchable Representation Learning Framework with Self-compatibility" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.18, + 0.865, + 0.235 + ], + "angle": 0, + "content": "Shengsen Wu\\(^{1,2}\\), Yan Bai\\(^{2}\\), Yihang Lou\\(^{3}\\), Xiongkun Linghu\\(^{4}\\), Jianzhong He\\(^{3}\\), Ling-Yu Duan\\(^{1,2,*}\\) \n\\(^{1}\\) Peng Cheng Laboratory, China; \\(^{2}\\) Peking University, China; \\(^{3}\\) Huawei Technologies; \\(^{4}\\) Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.244, + 0.77, + 0.278 + ], + "angle": 0, + "content": "{sswu, yanbai, lingyu}@pku.edu.cn, {louyihang1, jianzhong.he}@huawei.com, lhxk20@mails.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.312, + 0.314, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.473, + 0.676 + ], + "angle": 0, + "content": "Real-world visual search systems involve deployments on multiple platforms with different computing and storage resources. Deploying a unified model that suits the minimal-constrain platforms leads to limited accuracy. It is expected to deploy models with different capacities adapting to the resource constraints, which requires features extracted by these models to be aligned in the metric space. The method to achieve feature alignments is called \"compatible learning\". Existing research mainly focuses on the one-to-one compatible paradigm, which is limited in learning compatibility among multiple models. We propose a Switchable representation learning Framework with Self-Compatibility (SFSC). SFSC generates a series of compatible sub-models with different capacities through one training process. The optimization of sub-models faces gradients conflict, and we mitigate this problem from the perspective of the magnitude and direction. We adjust the priorities of sub-models dynamically through uncertainty estimation to co-optimize sub-models properly. Besides, the gradients with conflicting directions are projected to avoid mutual interference. SFSC achieves state-of-the-art performance on the evaluated datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.704, + 0.21, + 0.719 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Visual search systems are widely deployed, which recall the nearest neighbors in gallery features according to their distances to the query feature. Real-world visual search systems consist of multiple models deployed on different platforms [17, 38, 45] (e.g. clouds, mobiles, smart cameras), where different platforms interact with each other by visual features. Typically, as for person re-identification systems, images are captured and processed into features on edge sides. And such features are sent to the cloud side to compare with the database features for identifying" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.313, + 0.892, + 0.496 + ], + "angle": 0, + "content": "a specific person by feature similarities. As diverse platforms meet different computing and storage resource limitations, deploying the unified model which suits the minimal-constrain platforms leads to a waste of resources on other platforms and limited accuracy. To make better use of resources and achieve higher accuracy, it is expected to deploy models with different capacities which adapt to the resource limitations. Such a solution requires compatibility among the models to satisfy their interaction requirements, which means that the similar data processed by different models are close to each other in the feature space while dissimilar data are far apart." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.608 + ], + "angle": 0, + "content": "To achieve compatibility, compatible learning methods have been proposed. Existing research focuses on the one-to-one compatible paradigm [6,49,62], which constrains the learned features of a latter (learnable) model to be compatible with its previous (fixed) version. They are limited in achieving many-to-many compatibilities, which means any two in a series of models are compatible with each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a Switchable representation learning Framework with Self-Compatibility (SFSC) to achieve many-to-many compatibilities. As shown in Figure 1, SFSC can deploy different sub-models to suit the diverse computing and storage resource limitation of different platforms. The compatibility between any two sub-models can be achieved, termed as \"self-compatibility\". However, there are gradients conflicts between sub-models as they are co-optimized during the training process. We summarize such conflicts into the gradient magnitude and gradient direction. Specifically, a gradient of a large magnitude will dominate the optimization process, which may lead to the overfitting of its corresponding sub-model and impair the improvements of other sub-models. To solve this problem, we estimate the uncertainty of different sub-models to weight the gradient, denoting the optimization priorities. Besides, as the sub-models may produce gradients in various directions, there is mutual interference between them. Thus the improvement in different sub-models may be overestimated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.203, + 0.092, + 0.778, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 1. An example of SFSC in multi-platform deployments. The switchable neural network generates a series of sub-networks \\(\\{\\phi_1, \\phi_2, \\dots, \\phi_m\\}\\) with different capacities for different computing resources. \\(\\forall i, j \\in \\{1, 2, \\dots, m\\}\\), \\(|\\phi_i(a) - \\phi_j(p)| < |\\phi_i(a) - \\phi_j(n)|\\) where \\((a, p)\\) is a pair of samples with the same class ID, while \\((a, n)\\) is a pair of samples with different class IDs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.344 + ], + "angle": 0, + "content": "or underestimated. To tackle such conflict, the gradients are projected onto planes that are orthogonal to each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.375 + ], + "angle": 0, + "content": "The contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.383, + 0.47, + 0.459 + ], + "angle": 0, + "content": "- We propose a switchable representation learning framework with self-compatibility (SFSC). SFSC generates a series of feature-compatible sub-models with different capacities for visual search systems, which can be deployed on different platforms flexibly." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.468, + 0.47, + 0.591 + ], + "angle": 0, + "content": "- We resolve the conflicts between sub-models from the aspect of gradient magnitude and gradient direction. A compatible loss based on uncertainty estimation is proposed to guide optimization priorities and alleviate the imbalance of gradient magnitude between sub-models. An aggregation method based on the gradient projection is proposed to avoid mutual interference and find a generic optimal direction for all sub-models." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.599, + 0.47, + 0.675 + ], + "angle": 0, + "content": "- SFSC achieves state-of-the-art performance on the evaluated benchmark datasets. Compared to deploying a unified model, adopting SFSC to obtain different sub-models can achieve \\(6\\%\\) to \\(8\\%\\) performance improvements on three different datasets." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.383, + 0.47, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.687, + 0.221, + 0.703 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.45, + 0.729 + ], + "angle": 0, + "content": "2.1. One-to-one Compatible Learning Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given a fixed old model, the one-to-one compatible learning methods aim at aligning the learnable models with the fixed old one in metric space, termed as \"backward compatible learning\". Existing methods focus on the optimization difficulties caused by the old model. Yan et al. [63] propose Focal Distillation to reduce the \"negative flips\" which means a test sample is correctly classified by the old model but incorrectly classified by the new one. Wu et al. [62] start from neighborhood structures and classification entropy to eliminate the effect of outliers in old features. Shen et al. [48] propose a pseudo classifier and enhance it with a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.895, + 0.39 + ], + "angle": 0, + "content": "random walk algorithm to get rid of the old classifier. Different from these one-to-one compatible learning methods, this paper studies a new compatible learning paradigm that aims to learn many-to-many compatibility among multiple learnable models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.402, + 0.742, + 0.42 + ], + "angle": 0, + "content": "2.2. Dynamic Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.699 + ], + "angle": 0, + "content": "Dynamic neural networks are able to adapt their structures or parameters to the input during inference [19]. In recent years, research in this field has mainly focused on dynamic depth [5, 18, 26, 54, 55, 58], dynamic width [3, 4, 11, 27, 29, 36], dynamic routing [22, 25, 28, 64], dynamic parameters [12, 15, 21, 51, 69], etc. In addition to hand-crafted methods, Neural Architecture Search (NAS) method is also adopted to train and optimize multiple sub-nets with various capacities in one-time training [7, 8, 44, 53]. Dynamic neural networks can allocate computations on demand for different platforms by selectively activating model components. They have been used in classification [41, 42, 59], segmentation [31, 34, 43], and detection tasks [12, 56, 69]. In general, current dynamic neural networks output with definite semantic information (e.g. class id, detection box) which are naturally interoperable. However, for retrieval tasks, the networks output with features, which requires feature compatibility." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.711, + 0.848, + 0.728 + ], + "angle": 0, + "content": "2.3. Gradient-based Multi-task Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "From the gradient perspective, multi-task optimization is to aggregate gradients corresponding to different tasks and find a descent direction that decreases all objectives. Existing methods treat it as a nonlinear programming problem and take the multi-objective Karush-Kuhn-Tucker [33] (KKT) conditions to perform gradient projection and rescaling under different assumptions on tasks, e.g. multiple tasks which are equally important [10, 46, 66], main tasks with some auxiliary tasks [13, 57], a series of the tasks which are learned sequentially [14, 35, 37]. This paper treats optimizations of multiple sub-models as multiple tasks. As" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.088, + 0.795, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.895, + 0.274 + ], + "angle": 0, + "content": "Figure 2. Overview of SFSC. We first transform the full modelcalculate the compatible learning loss \\(\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}\\) on different sub-models with various capacitie. \\(\\phi\\) into the sub-models. Then we calculate the compatible learning loss \\(\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}\\) on different sub-models with various capacities. Finally, we optimize the sub-models by aggregating losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.377 + ], + "angle": 0, + "content": "the sub-models vary greatly in their capacity, the optimization of different models converges at different speeds. The challenge is to dynamically adjust the priorities of different sub-models to avoid the overfitting of someones as well as underfitting of others." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.388, + 0.313, + 0.405 + ], + "angle": 0, + "content": "2.4. Evidential Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.474, + 0.671 + ], + "angle": 0, + "content": "Evidential deep learning [47] is used to model the uncertainty of deterministic neural networks. Compared with the Bayesian neural networks [32, 40], evidential deep learning methods have the advantage of low computational complexity, as there is no need for complex posterior probability estimates and multiple Mento Carlo samplings. Based on the DST [16] (Dempster-Shafer Theory of Evidence), evidential deep learning treats the output of the neural networks as subjective opinions and formalizes them as Dirichlet distribution. It has been used in classification tasks to overcome the overconfidence problem of softmax cross entropy [30], action recognition task to identify unknown samples [2], and multi-view learning task [20] to dynamically integrate different views. This paper treats the output features of different sub-models as evidence to access uncertainty and relates the priority of different sub-models with the obtained uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.686, + 0.176, + 0.701 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.789 + ], + "angle": 0, + "content": "In this section, we will describe the proposed SFSC method. We first introduce the construction of switchable neural networks. Then, we analyze the difficulties for optimizations and propose our compatible loss as well as the gradient aggregation method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.194, + 0.815 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "SFSC involves the training process of multiple sub-models whose features are compatible. Sub-models with different capacities are transformed from the full model \\(\\phi\\) according to the crop ratio list \\(W\\). We termed the crop ratio list of each sub-model as \\(W = \\{\\gamma_{1},\\gamma_{2},\\dots,\\gamma_{m}\\}\\), where \\(\\gamma_{j}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.329 + ], + "angle": 0, + "content": "denotes the ratio of the width of the sub-model \\(\\phi_j\\) to the full model \\(\\phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.892, + 0.422 + ], + "angle": 0, + "content": "As shown in Figure 2, for each batch input, SFSC first calculates the original loss \\( L_{ori} \\). The compatible loss \\( \\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\} \\) is calculated on different sub-models. By aggregating losses \\( \\{L_{ori}, L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\} \\), SFSC computes the general gradient and updates the full model \\( \\phi \\) by gradient descent methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.434, + 0.892, + 0.465 + ], + "angle": 0, + "content": "3.2. Switchable Neural Network for Sub-models Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.625 + ], + "angle": 0, + "content": "We first convert the traditional convolutional neural network into its corresponding switchable neural network. As the platforms generally perform model inference layer by layer, the resource limitation is more related to the widths of layers than the number of layers. Here we perform channel pruning on \\(\\phi\\) with crop ratio \\(\\gamma_{j}\\) to obtain sub-model \\(\\phi_{j}\\). Specifically, the classic convolutional neural network is mainly composed of some basic modules, such as convolution, fully connected layer, and batch normalization. We modify these basic modules as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.687 + ], + "angle": 0, + "content": "1) For convolution (Conv) and fully connected (FC) layers, if their output is not the final feature for retrieval, we perform pruning in both input and output channels. Otherwise, pruning is only performed in input channels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.824 + ], + "angle": 0, + "content": "2) For batch normalization (BN), we assign an independent BN to the sub-model \\(\\phi_j\\). As the intermediate features extracted by different sub-models vary greatly in distributions [65], channel pruning can not be performed on BN. Fortunately, the parameters of BN only occupy a small part of all parameters of the entire network (shown in Table 1). Such an operation only results in a slight increase in storage resources in training progress but does not incur additional deployment costs." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Through such design, we obtain the sub-model \\(\\phi_j\\) with the following characteristics: 1) The features extracted by different sub-models have the same dimensions. 2) The ratio of layer widths in the sub-model \\(\\phi_j\\) to that in the full model \\(\\phi\\) is around \\(\\gamma_j\\). 3) The sub-model \\(\\phi_j\\) reuses the pa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15945" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.133 + ], + "angle": 0, + "content": "
NetWorksResNet-18 [23]ResNet-50 [23]MobileNet V2 [45]ShuffleNet [67]
Conv and FC11,679,91225,503,9123,470,8705,402,826
99.918%99.792%99.027%99.102%
BN9,60053,12034,11248,960
0.0820%0.208%0.973%0.898%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.137, + 0.47, + 0.166 + ], + "angle": 0, + "content": "Table 1. Number and proportion of different modules' parameters in different network structures" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.193, + 0.383, + 0.208 + ], + "angle": 0, + "content": "rameters from the full model \\(\\phi\\) except for BN." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.216, + 0.469, + 0.248 + ], + "angle": 0, + "content": "3.3. Aggregation Method based on the Gradients Projection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.469, + 0.331 + ], + "angle": 0, + "content": "As the gradients corresponding to different sub-models may have different directions, simply aggregating the gradients by summation may cause mutual interference, which means the improvements of sub-models are overestimated or underestimated, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.469, + 0.451 + ], + "angle": 0, + "content": "Inspired by multi-task learning [1,46,66], we tackle such a problem by projecting conflicting gradients and keeping the components perpendicular to each other. Given a pair of gradients \\((g_i, g_j)\\) corresponding to sub-model \\(\\phi_i\\) and \\(\\phi_j\\), if the cosine of their direction \\(\\cos(g_i, g_j) < 0\\), they are conflict with each other. For a pair of conflicting gradients \\((g_i, g_j)\\), we project them to the orthogonal plane of each other, respectively, as shown in Equation 1." + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.452, + 0.469, + 0.482 + ], + "angle": 0, + "content": "\\[\ng _ {j} ^ {\\prime} = g _ {i} - \\frac {g _ {j} * g _ {i}}{\\left\\| g _ {i} \\right\\| ^ {2}} * g _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Such projection results in a direction that improves both the \\(i\\)-th and \\(j\\)-th sub-model simultaneously." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.534, + 0.46, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.619, + 0.469, + 0.688 + ], + "angle": 0, + "content": "Figure 3. An example of gradient conflict between pair \\((g_i, g_j)\\). Simply taking the summation aggregation results in the elimination of the improvement of the \\(i\\)-th sub-model. Our aggregation module ensures the generic direction to improve both the \\(i\\)-th and \\(j\\)-th sub-model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.469, + 0.726 + ], + "angle": 0, + "content": "3.4. Compatible Loss based on the Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.81 + ], + "angle": 0, + "content": "The aggregation method aims to optimize all sub-models coordinately. However, it may not work as expected when facing the imbalance of gradient magnitude between sub-models, as shown in Figure 4. Therefore, it is critical to design a compatible loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As sub-models are co-optimized, constraining them with existing compatible loss [6, 49] faces conflicts in the magnitude of gradients. As the sub-models vary in capacities, their optimization process converges at different speeds. The small-capacity sub-models will converge faster. Therefore, the small-capacity sub-model may have reached its" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.092, + 0.88, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.173, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Figure 4. An example of gradient conflict between pair \\((g_i, g_j)\\) when there is an imbalance of gradient magnitude. The \\(i\\)-th sub-model contributes little to the optimization even using gradient projection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.892, + 0.344 + ], + "angle": 0, + "content": "performance ceiling while still producing losses of large magnitude, resulting in the overfitting of such sub-model and weakening of other sub-models. Intuitively, a sub-model with low uncertainty tends to reach its performance ceiling. Therefore, it is expected to adaptively adjust the optimization priorities of each sub-models through their uncertainty dynamically during the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Here we propose a compatible loss based on the uncertainty estimation. Inspired by Dempster-Shafer Theory of evidence [16] (DST), we measure the uncertainty through belief, where high uncertainty corresponds to the model with small capacities. Typically, the belief is measured by a set of mass \\( B = \\{b_{1}, b_{2}, \\ldots, b_{C}\\} \\) and an overall uncertainty mass \\( u \\), where \\( b_{c} \\) denotes the belief degree belonging to the class \\( c \\). \\( B \\) and \\( u \\) are related to a Dirichlet distribution, which satisfy Equation 2." + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.483, + 0.891, + 0.523 + ], + "angle": 0, + "content": "\\[\nu + \\sum_ {c = 1} ^ {C} b _ {c} = 1, u \\geq 0, b _ {c} \\geq 0. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.892, + 0.667 + ], + "angle": 0, + "content": "The belief is estimated by the features extracted by sub-models. Given sample \\(i\\), its features extracted by the \\(j\\)-th sub-model termed \\(\\phi_j(i)\\). The evidence \\(E_i^j = \\{e_{i1}^j, e_{i2}^j, \\dots, e_{iC}^j\\}\\) is transformed by \\(\\phi_j(i)\\) through linear projection and activation function (i.e., \\(\\exp(\\cdot)\\)), where the linear projection layers share parameters with the classifier of the \\(1.0 \\times\\) model \\(\\phi\\). The belief of sample \\(i\\) on the \\(j\\)-th sub-model is computed by Equation 3, where \\(\\alpha_i^j = \\{\\alpha_{i1}^j, \\alpha_{i2}^j, \\dots, \\alpha_{iC}^j\\}\\) is the Dirichlet intensity on all classes." + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.668, + 0.751, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i c} ^ {j} = e _ {i c} ^ {j} + 1,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.69, + 0.891, + 0.731 + ], + "angle": 0, + "content": "\\[\nS _ {i} ^ {j} = \\sum_ {c = 1} ^ {C} \\alpha_ {i c} ^ {j}, b _ {i c} ^ {j} = \\frac {\\alpha_ {i c} ^ {j} - 1}{S _ {i} ^ {j}}, u _ {i} ^ {j} = \\frac {C}{S _ {i} ^ {j}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.891, + 0.795 + ], + "angle": 0, + "content": "With \\(\\alpha_{i}^{j}\\), the Dirichlet distribution \\(\\mathrm{D}(p_i|\\alpha_i^j)\\) is obtained, where \\(p_i\\) is the class assignment probabilities on a simplex. Inspired by the Cross-Entropy loss, we integrate the probability \\(p_i\\) on \\(\\mathrm{D}(p_i|\\alpha_i^j)\\), as shown in Equation 4." + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.797, + 0.891, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\alpha_ {i} ^ {j}) = \\int [ \\sum_ {c = 1} ^ {C} - y _ {i c} ^ {j} \\log (p _ {i c} ^ {j}) ] \\frac {1}{\\beta (\\alpha_ {i} ^ {j})} \\Pi_ {c = 1} ^ {C} p _ {i c} ^ {j} ^ {(\\alpha_ {i c} - 1)} \\mathrm {d} p _ {i} ^ {j} \\\\ = \\sum_ {c = 1} ^ {C} y _ {i c} ^ {j} \\left(\\psi \\left(S _ {i} ^ {j}\\right) - \\psi \\left(\\alpha_ {i c} ^ {j}\\right)\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot)\\) is the digamma function. Since Equation 4 may converge to a trivial solution \\(\\mathrm{D}(p_i|1)\\) during optimization, we penalize the distance between \\(\\mathrm{D}(p_i|1)\\) and \\(\\mathrm{D}(p_i|\\alpha_i^j)\\) as the regularization. The complete compatible loss \\(L_{c_j}\\) is shown in Equation 5. \\(L_{c_j}\\) adaptively adjusts the magnitude of gradients through their corresponding uncertainty." + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.204, + 0.469, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K L \\left[ \\mathrm {D} \\left(p _ {i} \\mid \\alpha_ {i} ^ {j}\\right) | | \\mathrm {D} \\left(p _ {i} \\mid \\mathbf {1}\\right) \\right] = \\log \\left(\\frac {\\sum_ {c = 1} ^ {C} \\Gamma \\left(\\alpha_ {i c} ^ {j}\\right)}{\\Gamma (C) \\prod_ {c = 1} ^ {C} \\left(\\alpha_ {i c} ^ {j}\\right)}\\right) \\\\ + \\sum_ {c = 1} ^ {C} \\left(\\alpha_ {i c} ^ {j} - 1\\right) \\left[ \\psi \\left(\\alpha_ {i c} ^ {j} - \\psi \\left(\\sum_ {c = 1} ^ {C} \\alpha_ {i c} ^ {j}\\right)\\right) \\right], \\tag {5} \\\\ L _ {c _ {j}} (i) = \\mathcal {L} \\left(\\alpha_ {i} ^ {j}\\right) + \\lambda * K L \\left[ \\mathrm {D} \\left(p _ {i} \\mid \\alpha_ {i} ^ {j}\\right) | | \\mathrm {D} \\left(p _ {i} \\mid \\mathbf {1}\\right) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.319, + 0.47, + 0.35 + ], + "angle": 0, + "content": "We further conduct a gradient analysis for the proposed compatible loss. The Cross-Entropy loss is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.352, + 0.469, + 0.392 + ], + "angle": 0, + "content": "\\[\np _ {i c} = \\frac {e ^ {z _ {i c}}}{\\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}}, L _ {c e} = \\sum_ {c = 1} ^ {C} - y _ {i c} \\log \\left(p _ {i c}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.401, + 0.47, + 0.448 + ], + "angle": 0, + "content": "where \\( z_{i} \\) is the classification vector of sample \\( i \\) outputted by the model. The corresponding gradient of Equation 6 can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.456, + 0.469, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L _ {c e}}{\\partial z _ {i c}} = \\frac {e ^ {z _ {i c}}}{\\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}} - y _ {i c}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.5, + 0.469, + 0.53 + ], + "angle": 0, + "content": "The corresponding gradient of the proposed compatible loss in 5 can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.532, + 0.469, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial L _ {c _ {j}} \\left(\\alpha_ {i} ^ {j}\\right)}{\\partial \\alpha_ {i c} ^ {j}} = \\psi^ {\\prime} \\left(\\alpha_ {i c} ^ {j}\\right) \\left[ \\lambda \\left(\\alpha_ {i c} ^ {j} - 1\\right) - \\mathbf {y} _ {i c} \\right] \\tag {8} \\\\ + \\psi^ {'} (S _ {i} ^ {j}) [ - \\lambda (S _ {i} ^ {j} - C) + 1 ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.469, + 0.634 + ], + "angle": 0, + "content": "We take \\(\\frac{1}{\\alpha_{ic}^j}\\) as the approximation of \\(\\psi'(\\alpha_{ic}^j)\\) and obtain Equation 9." + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.636, + 0.469, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial L _ {c _ {j}} (\\alpha_ {i} ^ {j})}{\\partial \\alpha_ {i c} ^ {j}} \\approx \\frac {1}{\\alpha_ {i c} ^ {j}} [ \\lambda (\\alpha_ {i c} ^ {j} - 1) - \\mathbf {y} _ {i c} ] \\\\ + \\frac {1}{S _ {i} ^ {j}} [ - \\lambda (S _ {i} ^ {j} - C) + 1 ] \\tag {9} \\\\ = - (\\frac {\\lambda + y _ {i c}}{\\alpha_ {i c} ^ {j}} + \\frac {\\lambda C + 1}{S _ {i} ^ {j}}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.755, + 0.469, + 0.871 + ], + "angle": 0, + "content": "where the first term corresponds to the Cross-Entropy loss in Equation 7, and the second term corresponds to the uncertainty estimation. As the uncertainty \\( u_{i}^{j} = \\frac{C}{S_{i}^{j}} \\propto \\frac{1}{S_{i}^{j}} \\), thus higher \\( u_{i}^{j} \\) would result in a gradient with larger magnitude through \\( \\frac{\\lambda C + 1}{S_{i}^{j}} \\). Therefore, the magnitude of gradients corresponding to different sub-models is dynamically adjusted to avoid someone dominating the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Such compatible loss can work well with gradient projection, as shown in Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.764, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Training Process of SFSC" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.111, + 0.892, + 0.343 + ], + "angle": 0, + "content": "Require: batch input \\(\\mathcal{B}\\) the crop list \\(W = \\{\\gamma_0,\\gamma_1,\\dots ,\\gamma_m\\}\\) ,the full model \\(\\phi\\) \n1: \\(g_{ori}\\gets \\frac{\\delta L_{ori}(\\phi,\\mathcal{B})}{\\delta\\phi}\\) \n2: for \\(\\gamma_{j}\\in W\\) do \n3: Switch to the sub-model \\(\\phi_j\\) \n4: \\(g_{j}\\leftarrow \\frac{\\delta L_{c_{j}}(\\phi_{j},\\mathcal{B})}{\\delta\\phi}\\) \n5: end for \n6: \\(G,G^{p}\\gets \\{g_{ori},g_{0},\\ldots g_{m}\\}\\) \n7: for \\(g_{a}\\in G^{p}\\) do \n8: Shuffle(G) \n9: for \\(g_{b}\\in G\\) do \n10: \\(d = g_{a}*g_{b}\\) \n11: if \\(d < 0\\) then //have conflicts \n12: \\(g_{a}\\gets g_{a} - \\frac{g_{b}*d}{||g_{b}||^{2}}\\) \n13: end if \n14: end for \n15: end for \n16: return Update \\(\\phi\\) by \\(\\Delta \\theta = -\\sum_{g_a\\in GP}g_a\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.366, + 0.892, + 0.471 + ], + "angle": 0, + "content": "In general, SFSC aims to achieve the Pareto optimality [9] among the optimization of sub-models. The proposed compatible loss adaptively adjusts the magnitude of gradients through uncertainty to control the optimization priorities of sub-models dynamically. Besides, the aggregation method performs the gradient projection to find a generic direction to improve all sub-models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.487, + 0.633, + 0.505 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.754, + 0.53 + ], + "angle": 0, + "content": "4.1. Experimental Configuration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed SFSC on person ReID datasets, i.e., Market-1501 [68], MSMT17 [61], and one vehicle ReID dataset VeRi-776 [60]. Market1501 consists of 32,668 annotated images of 1,501 identities shot from 6 cameras. MSMT17 consists of 126,441 bounding boxes of 4,101 identities taken by 15 cameras. VeRi-776 contains over 50,000 images of 776 vehicles captured by 20 cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Comparison methods. For the one-to-one compatible learning paradigm, BCT [49] and Asymmetric [6] are adopted, as they are representatives of compatible learning through classification losses and embedding losses. BCT and Asymmetric can only obtain sub-models in a one-to-one compatible way. Specifically, given the fixed largest capacity model which is trained without compatible regularization, we train a series of small-capacity models to be compatible with it, respectively. Besides, we also incorporate BCT/Asymmetric loss in our proposed self-compatible learning paradigm for a fair comparison. We take the compatible loss in BCT/Asymmetric to constrain each sub-model compatible with the full model, and use summation to aggregate losses from sub-models, termed as \"BCT-S\" and \"Asymmetric-S\"." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. Mean average precision (mAP)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15947" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.078, + 0.787, + 0.348 + ], + "angle": 0, + "content": "
M(φq, φg) φq φqφ1/16φ1/4φ9/16φφ1/16φ1/4φ9/16φ
Unified ModelOurs
φ1/1655.25---58.1962.8863.4869.43
φ1/4-67.48--61.2470.7471.3776.37
φ9/16--71.25-62.4371.2572.0677.26
φ---80.9168.0374.7675.6781.43
BCT-SAsymmetric-S
φ1/1654.4058.1060.2261.2948.7951.9654.0956.31
φ1/455.7166.3868.2969.7152.2462.4564.9867.23
φ9/1656.5068.3068.4271.9954.4464.7467.9771.28
φ61.7969.7371.9373.6156.5467.3071.9178.38
BCTAsymmetric
φ1/1655.5556.6459.4060.3555.8350.8553.7155.39
φ1/455.8365.7467.1068.5452.6667.0462.0866.59
φ9/1655.6668.0267.4970.9354.4862.3366.3370.14
φ58.6969.4870.04-56.0866.6669.71-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.353, + 0.882, + 0.368 + ], + "angle": 0, + "content": "Table 2. Baseline performance comparison on Market1501 (mAP). \"Unified Model\" are models without any compatible regularization." + }, + { + "type": "table", + "bbox": [ + 0.167, + 0.38, + 0.803, + 0.56 + ], + "angle": 0, + "content": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
MSMT17
Unified Model28.8211.6948.6522.8657.5630.0670.2243.06
BCT-S46.4920.1156.2228.1358.1130.1458.7630.99
Asymmetric-S32.6413.3449.6224.0457.1830.0466.0840.39
Ours56.2228.1663.2935.3265.1337.7470.3843.89
VERI-776
Unified Model81.0744.7985.3053.6388.6955.4292.2066.50
BCT-S86.2549.3887.0857.0390.4258.2086.7358.46
Asymmetric-S68.7542.1684.2956.5588.0461.3789.8265.35
Ours79.7655.0489.9462.2890.4862.7292.3266.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.563, + 0.696, + 0.576 + ], + "angle": 0, + "content": "Table 3. Performance comparison on MSMT17 and VeRi-776 datasets." + }, + { + "type": "table", + "bbox": [ + 0.167, + 0.59, + 0.802, + 0.77 + ], + "angle": 0, + "content": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
Circle Loss
Unified Model70.2550.7972.6660.4378.2164.6687.3874.06
BCT-S63.1842.4572.7452.0477.3859.4987.5374.58
Asymmetric-S69.6648.2177.8557.0280.5261.8186.9073.08
Ours73.9954.6180.5863.4682.3967.0288.9577.05
Softmax + Triplet
Unified Model77.5555.8283.9765.3087.8970.6990.9478.30
BCT-S81.6259.5887.8669.0588.7571.3890.1174.71
Asymmetric-S75.5652.6085.3664.2186.9168.8390.4577.34
Ours87.3666.2190.9476.6491.2177.9392.6481.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.772, + 0.731, + 0.785 + ], + "angle": 0, + "content": "Table 4. Performance comparison between different loss functions on Market1501." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.813, + 0.471, + 0.889 + ], + "angle": 0, + "content": "and top-1 accuracy (R1) are adopted to evaluate the performances of retrieval. Define \\( M(\\phi_{a},\\phi_{b}) \\) as the performances of retrieval where the query features are extracted by model \\( \\phi_{a} \\) and gallery features are extracted by model \\( \\phi_{b} \\). We set \\( \\phi_{a} \\) and \\( \\phi_{b} \\) as different sub-models to evaluate the perfor" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.813, + 0.704, + 0.828 + ], + "angle": 0, + "content": "mance of compatible learning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Implementation details. The implementations of all compared methods and SFSC are based on the FastReLU [24]. The default configuration named \"bag of tricks\" [39] is adopted. \\(\\lambda\\) is set to 0.2 for all experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15948" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.166, + 0.088, + 0.803, + 0.34 + ], + "angle": 0, + "content": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
ResNet-50
Unified Model85.2765.4488.8472.0790.876.4092.2282.13
BCT-S89.3172.0991.0077.2291.3678.2791.8679.40
Asymmetric-S85.0467.3990.2676.2891.6078.5391.9582.79
Ours90.4176.4091.0378.2091.6678.7693.4483.66
MobileNet V2
Unified Model69.0041.1489.5872.8591.5477.9992.2879.50
BCT-S80.4655.1788.4870.6889.8873.7289.9673.87
Asymmetric-S67.2843.6486.9768.9589.3173.8290.7478.45
Ours83.1163.8290.5376.5291.1878.3092.6181.34
ShuffleNet
Unified Model82.5460.1289.8564.4591.9572.9684.3879.13
BCT-S85.5165.7688.4266.8589.7370.5986.6774.41
Asymmetric-S82.3062.5084.4157.4388.4266.4378.9274.92
Ours89.2573.0989.9374.8689.9375.1492.8181.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.221, + 0.343, + 0.747, + 0.356 + ], + "angle": 0, + "content": "Table 5. Performance comparison between different model architectures on Market1501." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.382, + 0.285, + 0.399 + ], + "angle": 0, + "content": "4.2. Baseline Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.543 + ], + "angle": 0, + "content": "We conduct baseline experiments on Market1501. We take \"Softmax + Cross Entropy\" as the loss function, ResNet-18 [23] as the backbone, and \\( W = \\{0.25 \\times, 0.50 \\times, 0.75 \\times\\} \\) as the crop ratio list of the sub-models. Notes that for a model with crop ratio \\( \\gamma_i \\), the ratio of its parameters to the full model is \\( \\gamma_i^2 \\). Therefore, the ratio of the number of parameters in the \\( 0.25 \\times \\), \\( 0.50 \\times \\) and \\( 0.75 \\times \\) sub-model to that in the full model are \\( \\frac{1}{16}, \\frac{1}{4}, \\frac{9}{16} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Comparison methods. As for one-to-one compatible learning methods BCT and Asymmetric, the compatibility does not outperform that achieved by their corresponding self-compatible methods especially when the models are not constrained to be compatible directly. For self-compatible learning methods, BCT-S and Asymmetric-S, the conflicts among sub-models lead to a limited average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Our Method (SFSC). As we resolve the conflicts among different sub-models, our methods can achieve many-to-many compatibility among all the sub-models, which significantly outperforms all comparison methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.411, + 0.756 + ], + "angle": 0, + "content": "4.3. Performances under Different Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Different datasets. We evaluate each self-compatible learning method on the other person ReID datasets MSMT17, and the vehicle ReID dataset VeRi-776 under baseline configurations, as shown in Table 3. The results on the two datasets with larger scales are consistent with that on the Market1501. Our methods outperform others significantly for the small capacity sub-model (e.g. the \\(0.25 \\times\\) sub-model \\(\\phi_{\\frac{1}{4}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Different loss functions. We replace the original loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.489 + ], + "angle": 0, + "content": "function \\(L_{ori}\\) from \"Softmax\" to \"Softmax + Triplet\" and \"Circle Loss\" [52] to evaluate the effectiveness of each method, while Triplet loss is a kind embedding loss and circle loss is a kind of classification loss. The results in Table 4 show that the accuracy of BCT-S and Asymmetric-S drops greatly when training with circle loss, as Our method still stably achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.893, + 0.642 + ], + "angle": 0, + "content": "Different model architectures. We also study whether each method can be stably applied to various model architectures, i.e., ResNet-50 [23], MobileNet V2 [45] and ShuffleNet [67]. ShuffleNet and MobileNet are both designed especially for edge devices, which meet the application scenarios of self-compatible learning. It is shown in Table 5 that BCT-S and Asymmetric-S do not achieve higher accuracy than deploying a unified model with ShuffleNet and MobileNet (e.g. for the \\(0.75 \\times\\) sub-model \\(\\phi_{\\frac{9}{16}}\\)). And our methods always improve the accuracy in all cases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.658, + 0.655, + 0.674 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.73 + ], + "angle": 0, + "content": "We replace the proposed compatible loss and aggregation module with existing compatible loss BCT [50] and summation aggregation to test the proposed two methods." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.744, + 0.888, + 0.804 + ], + "angle": 0, + "content": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
Unified Model55.2567.4871.2580.91
w/o Ours Loss66.4174.5876.8677.87
w/o Ours Aggregation68.0873.7674.8475.38
SFSC69.4376.3777.2681.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.805, + 0.816, + 0.817 + ], + "angle": 0, + "content": "Table 6. Ablation Study on Market1501." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 6 shows that the proposed compatible loss mainly improves the retrieval accuracy of the small capacity sub-models while the proposed aggregation method can ensure the performance of the full model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.188, + 0.089, + 0.48, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.089, + 0.782, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Figure 5. Gradient visualization on Market1501-R18 baseline. As for the original cross-entropy loss (BCT [49]), the magnitude of the \\(0.25 \\times\\) sub-model is much larger than that of the \\(0.75 \\times\\) sub-model, especially at the early stage of the training process (\\(0 \\sim 1000\\) iterations). Therefore, the \\(0.25 \\times\\) sub-model dominates the training progress and the improvements of the \\(0.75 \\times\\) sub-models are eliminated. While our method ensures co-optimization through uncertainty estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.315, + 0.37 + ], + "angle": 0, + "content": "4.5. Hyperparameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.496 + ], + "angle": 0, + "content": "Here we verify the impact of hyperparameter settings on the performance of the proposed method. The key hyperparameter is the crop ratio list for sub-models \\( W \\). Intuitively, if the number of \\( W \\) increases or the minimal crop ratio decreases, the compatibility constraints on model \\( \\phi \\) would be more strict, thereby increasing the difficulty of compatible learning. Therefore, we take the baseline configuration and set \\( W \\) with different values." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.501, + 0.472, + 0.558 + ], + "angle": 0, + "content": "
M(φ1/16, φ)M(φ81/100, φ)M(φ169/400, φ)M(φ283/400, φ)M(φ, φ)
Unified Model55.2562.5469.5472.5780.27
BCT-S61.4568.6373.4272.8075.39
Asymmetric-S62.8769.4575.2073.1474.60
Ours69.6375.8177.1077.3680.59
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.559, + 0.47, + 0.587 + ], + "angle": 0, + "content": "Table 7. Performance comparison with \\(W = \\{0.25\\times ,0.45\\times ,0.65\\times ,0.85\\times \\}\\) on Market1501 (mAP)." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.596, + 0.469, + 0.655 + ], + "angle": 0, + "content": "
M(φ1/100, φ)M(φ1/16, φ)M(φ9/16, φ)M(φ, φ)
Unified Model40.4055.2571.2580.27
BCT-S42.3459.7571.7376.39
Asymmetric-S42.5858.2070.5075.42
Ours45.4468.1878.0081.46
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.079, + 0.656, + 0.469, + 0.684 + ], + "angle": 0, + "content": "Table 8. Performance comparison with \\(W = \\{0.10\\times ,0.25\\times ,0.75\\times \\}\\) on Market1501 (mAP)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Table 7 shows the performances with the number of \\( W \\) increases, while Table 8 shows the performances with the minimum among \\( W \\) decreases. The result of the comparison methods has fluctuated to varying degrees. SFSC still achieves the best retrieval accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.216, + 0.787 + ], + "angle": 0, + "content": "4.6. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We visualize the norm of the gradient during the training progress on Market1501-R18 baseline, as shown in Figure 5. We also visualize the cumulative number of gradient pairs which are conflict with each other during the training process. As shown in Figure 6, our method has found a generic direction to improve all sub-models. Therefore, there are almost no conflict pairs after 5000 iterations. As" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.353, + 0.857, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.532, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Figure 6. The cumulative number of conflict pairs during the training process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.893, + 0.613 + ], + "angle": 0, + "content": "for the summation aggregation, it faces fierce conflicts especially after 10000 iterations, resulting in the improvement of different sub-models overestimated or underestimated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.62, + 0.642 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.652, + 0.892, + 0.786 + ], + "angle": 0, + "content": "This paper proposes a Switchable Representation Learning Framework with Self Compatibility (SFSC) for multi-platform model collaboration. SFSC enables us to obtain models with various capacities to fit different computing and storage resource constraints on diverse platforms. It finds a generic direction to improve all sub-models by uncertainty estimation and gradient projection. SFSC achieves state-of-the-art performances and shows convincing robustness in different case studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.681, + 0.818 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, the Basic and Frontier Research Project of PCL, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15950" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Lu Bai, Yew-Soon Ong, Tiantian He, and Abhishek Gupta. Multi-task gradient descent for multi-task learning. Memetic Computing, 12(4):355-369, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.214 + ], + "angle": 0, + "content": "[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13349-13358, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[3] Emmanuel Bengio, Pierre-Luc Bacon, Joelle Pineau, and Doina Precup. Conditional computation in neural networks for faster models. arXiv preprint arXiv:1511.06297, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.371 + ], + "angle": 0, + "content": "[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[6] Mateusz Budnik and Yannis Avrithis. Asymmetric metric learning for knowledge transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2021. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[7] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[8] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[9] Yair Censor. Pareto optimality in multiobjective problems. Applied Mathematics and Optimization, 4(1):41-59, 1977. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International Conference on Machine Learning, pages 794-803. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[11] Kyunghyun Cho and Yoshua Bengio. Exponentially increasing the capacity-to-computation ratio for conditional computation in deep learning. arXiv preprint arXiv:1406.7362, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[12] Jifeng Dai, Haozhi Qi, Yuwen Xiong, and Yi Li. Guodong, zhang, han hu, and yichen wei. Deformable convolutional networks. In, ICCV, 2(3):7, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[13] Yunshu Du, Wojciech M Czarnecki, Siddhant M Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[14] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, pages 3762-3773. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[15] Hang Gao, Xizhou Zhu, Steve Lin, and Jifeng Dai. Deformable kernels: Adapting effective receptive fields for object deformation. arXiv preprint arXiv:1910.02940, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[16] Jean Gordon and Edward H Shortliffe. The dampster-shafer theory of evidence. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, 3:832-838, 1984. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[17] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12259-12269, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.893, + 0.259 + ], + "angle": 0, + "content": "[18] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[20] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.893, + 0.425 + ], + "angle": 0, + "content": "[21] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.893, + 0.495 + ], + "angle": 0, + "content": "[22] Hussein Hazimeh, Natalia Ponomareva, Petros Mol, Zhenyu Tan, and Rahul Mazumder. The tree ensemble layer: Differentiability meets conditional computation. In International Conference on Machine Learning, pages 4138-4148. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[24] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631, 6(7):8, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.893, + 0.662 + ], + "angle": 0, + "content": "[25] Thomas M Hehn, Julian FP Kooij, and Fred A Hamprecht. End-to-end learning of decision trees and forests. International Journal of Computer Vision, 128(4):997-1011, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.893, + 0.72 + ], + "angle": 0, + "content": "[26] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[27] Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2761, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[28] Yani Ioannou, Duncan Robertson, Darko Zikic, Peter Kontschieder, Jamie Shotton, Matthew Brown, and Antonio Criminisi. Decision forests, convolutional networks and the models in-between. arXiv preprint arXiv:1603.01250, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[29] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15951" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[30] Taejong Joo, Uijung Chung, and Min-Gwan Seo. Being bayesian about categorical probability. In International Conference on Machine Learning, pages 4950-4961. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[31] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.231 + ], + "angle": 0, + "content": "[32] Igor Kononenko. Bayesian neural networks. Biological Cybernetics, 61(5):361-370, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[33] Harold W Kuhn and Albert W Tucker. Nonlinear programming. In Traces and emergence of nonlinear programming, pages 247-258. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[34] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[35] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Trgp: Trust region gradient projection for continual learning. arXiv preprint arXiv:2202.02931, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[36] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In Proceedings of the IEEE international conference on computer vision, pages 2736-2744, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[38] Yihang Lou, Ling-Yu Duan, Shiqi Wang, Ziqian Chen, Yan Bai, Changwen Chen, and Wen Gao. Front-end smart visual sensing and back-end intelligent analysis: A unified infrastructure for economizing the visual system of city brain. IEEE Journal on Selected Areas in Communications, 37(7):1489-1503, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[39] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[40] David JC MacKay. Bayesian neural networks and density networks. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 354(1):73-80, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[41] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[42] Amir Rosenfeld and Shimon Ullman. Visual concept recognition and localization via iterative introspection. In *Asian Conference on Computer Vision*, pages 264-279. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[43] Abhijit Guha Roy, Nassir Navab, and Christian Wachinger. Concurrent spatial and channel 'squeeze & excitation' in fully convolutional networks. In International conference on medical image computing and computer-assisted intervention, pages 421-429. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[44] Manas Sahni, Shreya Varshini, Alind Khare, and Alexey Tumanov. Compofa: Compound once-for-all networks" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "for faster multi-platform deployment. arXiv preprint arXiv:2104.12642, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[46] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[47] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in Neural Information Processing Systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[48] Yantao Shen, Fanzi Wu, and Ying Shan. Non-inherent feature compatible learning. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[49] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 1, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[50] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[51] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[52] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[53] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[54] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[55] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[56] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[57] Xinyi Wang, Ankur Bapna, Melvin Johnson, and Orhan Firat. Gradient-guided loss masking for neural machine translation. arXiv preprint arXiv:2102.13549, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[58] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[59] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. Advances in Neural Information Processing Systems, 33:2432-2444, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[60] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. Proceedings of the IEEE International Conference on Computer Vision, pages 379–387, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[61] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[62] Shengsen Wu, Liang Chen, Yihang Lou, Yan Bai, Tao Bai, Minghua Deng, and Lingyu Duan. Neighborhood consensus contrastive learning for backward-compatible representation. arXiv preprint arXiv:2108.03372, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[63] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.488, + 0.469, + 0.569 + ], + "angle": 0, + "content": "[64] Zhicheng Yan, Hao Zhang, Robinson Piramuthu, Vignesh Jagadeesh, Dennis DeCoste, Wei Di, and Yizhou Yu. Hd-cnn: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision, pages 2740-2748, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.572, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[65] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.616, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[66] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.672, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[67] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.743, + 0.469, + 0.797 + ], + "angle": 0, + "content": "[68] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.799, + 0.469, + 0.854 + ], + "angle": 0, + "content": "[69] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9308-9316, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15953" + } + ] +] \ No newline at end of file diff --git a/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_origin.pdf b/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff381b8b8f03423da7478f138c8e4503f8343724 --- /dev/null +++ b/2023/Switchable Representation Learning Framework With Self-Compatibility/879611fd-27b7-4de2-b6e0-22124b67ae58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b7c075ddb5d8138712d0001c7c2194ec499947f4ddedad3924965b0a1c765bb +size 997468 diff --git a/2023/Switchable Representation Learning Framework With Self-Compatibility/full.md b/2023/Switchable Representation Learning Framework With Self-Compatibility/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63fb4127806d20ff0eae07b49ab9e18b26b5b642 --- /dev/null +++ b/2023/Switchable Representation Learning Framework With Self-Compatibility/full.md @@ -0,0 +1,359 @@ +# Switchable Representation Learning Framework with Self-compatibility + +Shengsen Wu $^{1,2}$ , Yan Bai $^{2}$ , Yihang Lou $^{3}$ , Xiongkun Linghu $^{4}$ , Jianzhong He $^{3}$ , Ling-Yu Duan $^{1,2,*}$ $^{1}$ Peng Cheng Laboratory, China; $^{2}$ Peking University, China; $^{3}$ Huawei Technologies; $^{4}$ Tsinghua University, China + +{sswu, yanbai, lingyu}@pku.edu.cn, {louyihang1, jianzhong.he}@huawei.com, lhxk20@mails.tsinghua.edu.cn + +# Abstract + +Real-world visual search systems involve deployments on multiple platforms with different computing and storage resources. Deploying a unified model that suits the minimal-constrain platforms leads to limited accuracy. It is expected to deploy models with different capacities adapting to the resource constraints, which requires features extracted by these models to be aligned in the metric space. The method to achieve feature alignments is called "compatible learning". Existing research mainly focuses on the one-to-one compatible paradigm, which is limited in learning compatibility among multiple models. We propose a Switchable representation learning Framework with Self-Compatibility (SFSC). SFSC generates a series of compatible sub-models with different capacities through one training process. The optimization of sub-models faces gradients conflict, and we mitigate this problem from the perspective of the magnitude and direction. We adjust the priorities of sub-models dynamically through uncertainty estimation to co-optimize sub-models properly. Besides, the gradients with conflicting directions are projected to avoid mutual interference. SFSC achieves state-of-the-art performance on the evaluated datasets. + +# 1. Introduction + +Visual search systems are widely deployed, which recall the nearest neighbors in gallery features according to their distances to the query feature. Real-world visual search systems consist of multiple models deployed on different platforms [17, 38, 45] (e.g. clouds, mobiles, smart cameras), where different platforms interact with each other by visual features. Typically, as for person re-identification systems, images are captured and processed into features on edge sides. And such features are sent to the cloud side to compare with the database features for identifying + +a specific person by feature similarities. As diverse platforms meet different computing and storage resource limitations, deploying the unified model which suits the minimal-constrain platforms leads to a waste of resources on other platforms and limited accuracy. To make better use of resources and achieve higher accuracy, it is expected to deploy models with different capacities which adapt to the resource limitations. Such a solution requires compatibility among the models to satisfy their interaction requirements, which means that the similar data processed by different models are close to each other in the feature space while dissimilar data are far apart. + +To achieve compatibility, compatible learning methods have been proposed. Existing research focuses on the one-to-one compatible paradigm [6,49,62], which constrains the learned features of a latter (learnable) model to be compatible with its previous (fixed) version. They are limited in achieving many-to-many compatibilities, which means any two in a series of models are compatible with each other. + +In this work, we propose a Switchable representation learning Framework with Self-Compatibility (SFSC) to achieve many-to-many compatibilities. As shown in Figure 1, SFSC can deploy different sub-models to suit the diverse computing and storage resource limitation of different platforms. The compatibility between any two sub-models can be achieved, termed as "self-compatibility". However, there are gradients conflicts between sub-models as they are co-optimized during the training process. We summarize such conflicts into the gradient magnitude and gradient direction. Specifically, a gradient of a large magnitude will dominate the optimization process, which may lead to the overfitting of its corresponding sub-model and impair the improvements of other sub-models. To solve this problem, we estimate the uncertainty of different sub-models to weight the gradient, denoting the optimization priorities. Besides, as the sub-models may produce gradients in various directions, there is mutual interference between them. Thus the improvement in different sub-models may be overestimated + +![](images/70e765d57ccf0a68bef5a3b047029b61a69e6e0bb8e330762e2af7676e2b3878.jpg) +Figure 1. An example of SFSC in multi-platform deployments. The switchable neural network generates a series of sub-networks $\{\phi_1, \phi_2, \dots, \phi_m\}$ with different capacities for different computing resources. $\forall i, j \in \{1, 2, \dots, m\}$ , $|\phi_i(a) - \phi_j(p)| < |\phi_i(a) - \phi_j(n)|$ where $(a, p)$ is a pair of samples with the same class ID, while $(a, n)$ is a pair of samples with different class IDs. + +or underestimated. To tackle such conflict, the gradients are projected onto planes that are orthogonal to each other. + +The contributions of this paper are summarized as follows: + +- We propose a switchable representation learning framework with self-compatibility (SFSC). SFSC generates a series of feature-compatible sub-models with different capacities for visual search systems, which can be deployed on different platforms flexibly. +- We resolve the conflicts between sub-models from the aspect of gradient magnitude and gradient direction. A compatible loss based on uncertainty estimation is proposed to guide optimization priorities and alleviate the imbalance of gradient magnitude between sub-models. An aggregation method based on the gradient projection is proposed to avoid mutual interference and find a generic optimal direction for all sub-models. +- SFSC achieves state-of-the-art performance on the evaluated benchmark datasets. Compared to deploying a unified model, adopting SFSC to obtain different sub-models can achieve $6\%$ to $8\%$ performance improvements on three different datasets. + +# 2. Related Work + +# 2.1. One-to-one Compatible Learning Paradigm + +Given a fixed old model, the one-to-one compatible learning methods aim at aligning the learnable models with the fixed old one in metric space, termed as "backward compatible learning". Existing methods focus on the optimization difficulties caused by the old model. Yan et al. [63] propose Focal Distillation to reduce the "negative flips" which means a test sample is correctly classified by the old model but incorrectly classified by the new one. Wu et al. [62] start from neighborhood structures and classification entropy to eliminate the effect of outliers in old features. Shen et al. [48] propose a pseudo classifier and enhance it with a + +random walk algorithm to get rid of the old classifier. Different from these one-to-one compatible learning methods, this paper studies a new compatible learning paradigm that aims to learn many-to-many compatibility among multiple learnable models. + +# 2.2. Dynamic Neural Networks + +Dynamic neural networks are able to adapt their structures or parameters to the input during inference [19]. In recent years, research in this field has mainly focused on dynamic depth [5, 18, 26, 54, 55, 58], dynamic width [3, 4, 11, 27, 29, 36], dynamic routing [22, 25, 28, 64], dynamic parameters [12, 15, 21, 51, 69], etc. In addition to hand-crafted methods, Neural Architecture Search (NAS) method is also adopted to train and optimize multiple sub-nets with various capacities in one-time training [7, 8, 44, 53]. Dynamic neural networks can allocate computations on demand for different platforms by selectively activating model components. They have been used in classification [41, 42, 59], segmentation [31, 34, 43], and detection tasks [12, 56, 69]. In general, current dynamic neural networks output with definite semantic information (e.g. class id, detection box) which are naturally interoperable. However, for retrieval tasks, the networks output with features, which requires feature compatibility. + +# 2.3. Gradient-based Multi-task Optimization + +From the gradient perspective, multi-task optimization is to aggregate gradients corresponding to different tasks and find a descent direction that decreases all objectives. Existing methods treat it as a nonlinear programming problem and take the multi-objective Karush-Kuhn-Tucker [33] (KKT) conditions to perform gradient projection and rescaling under different assumptions on tasks, e.g. multiple tasks which are equally important [10, 46, 66], main tasks with some auxiliary tasks [13, 57], a series of the tasks which are learned sequentially [14, 35, 37]. This paper treats optimizations of multiple sub-models as multiple tasks. As + +![](images/87150a85c8cbcecf8d2b8dd20aa49ecb3141df1c0b332e505b983f8a2f8ccc58.jpg) +Figure 2. Overview of SFSC. We first transform the full modelcalculate the compatible learning loss $\{L_{c_1}, L_{c_2}, \dots, L_{c_m}\}$ on different sub-models with various capacitie. $\phi$ into the sub-models. Then we calculate the compatible learning loss $\{L_{c_1}, L_{c_2}, \dots, L_{c_m}\}$ on different sub-models with various capacities. Finally, we optimize the sub-models by aggregating losses. + +the sub-models vary greatly in their capacity, the optimization of different models converges at different speeds. The challenge is to dynamically adjust the priorities of different sub-models to avoid the overfitting of someones as well as underfitting of others. + +# 2.4. Evidential Deep Learning + +Evidential deep learning [47] is used to model the uncertainty of deterministic neural networks. Compared with the Bayesian neural networks [32, 40], evidential deep learning methods have the advantage of low computational complexity, as there is no need for complex posterior probability estimates and multiple Mento Carlo samplings. Based on the DST [16] (Dempster-Shafer Theory of Evidence), evidential deep learning treats the output of the neural networks as subjective opinions and formalizes them as Dirichlet distribution. It has been used in classification tasks to overcome the overconfidence problem of softmax cross entropy [30], action recognition task to identify unknown samples [2], and multi-view learning task [20] to dynamically integrate different views. This paper treats the output features of different sub-models as evidence to access uncertainty and relates the priority of different sub-models with the obtained uncertainty. + +# 3. Methods + +In this section, we will describe the proposed SFSC method. We first introduce the construction of switchable neural networks. Then, we analyze the difficulties for optimizations and propose our compatible loss as well as the gradient aggregation method. + +# 3.1. Overview + +SFSC involves the training process of multiple sub-models whose features are compatible. Sub-models with different capacities are transformed from the full model $\phi$ according to the crop ratio list $W$ . We termed the crop ratio list of each sub-model as $W = \{\gamma_{1},\gamma_{2},\dots,\gamma_{m}\}$ , where $\gamma_{j}$ + +denotes the ratio of the width of the sub-model $\phi_j$ to the full model $\phi$ . + +As shown in Figure 2, for each batch input, SFSC first calculates the original loss $L_{ori}$ . The compatible loss $\{L_{c_1}, L_{c_2}, \dots, L_{c_m}\}$ is calculated on different sub-models. By aggregating losses $\{L_{ori}, L_{c_1}, L_{c_2}, \dots, L_{c_m}\}$ , SFSC computes the general gradient and updates the full model $\phi$ by gradient descent methods. + +# 3.2. Switchable Neural Network for Sub-models Construction + +We first convert the traditional convolutional neural network into its corresponding switchable neural network. As the platforms generally perform model inference layer by layer, the resource limitation is more related to the widths of layers than the number of layers. Here we perform channel pruning on $\phi$ with crop ratio $\gamma_{j}$ to obtain sub-model $\phi_{j}$ . Specifically, the classic convolutional neural network is mainly composed of some basic modules, such as convolution, fully connected layer, and batch normalization. We modify these basic modules as follows: + +1) For convolution (Conv) and fully connected (FC) layers, if their output is not the final feature for retrieval, we perform pruning in both input and output channels. Otherwise, pruning is only performed in input channels. +2) For batch normalization (BN), we assign an independent BN to the sub-model $\phi_j$ . As the intermediate features extracted by different sub-models vary greatly in distributions [65], channel pruning can not be performed on BN. Fortunately, the parameters of BN only occupy a small part of all parameters of the entire network (shown in Table 1). Such an operation only results in a slight increase in storage resources in training progress but does not incur additional deployment costs. + +Through such design, we obtain the sub-model $\phi_j$ with the following characteristics: 1) The features extracted by different sub-models have the same dimensions. 2) The ratio of layer widths in the sub-model $\phi_j$ to that in the full model $\phi$ is around $\gamma_j$ . 3) The sub-model $\phi_j$ reuses the pa + +
NetWorksResNet-18 [23]ResNet-50 [23]MobileNet V2 [45]ShuffleNet [67]
Conv and FC11,679,91225,503,9123,470,8705,402,826
99.918%99.792%99.027%99.102%
BN9,60053,12034,11248,960
0.0820%0.208%0.973%0.898%
+ +Table 1. Number and proportion of different modules' parameters in different network structures + +rameters from the full model $\phi$ except for BN. + +# 3.3. Aggregation Method based on the Gradients Projection + +As the gradients corresponding to different sub-models may have different directions, simply aggregating the gradients by summation may cause mutual interference, which means the improvements of sub-models are overestimated or underestimated, as shown in Figure 3. + +Inspired by multi-task learning [1,46,66], we tackle such a problem by projecting conflicting gradients and keeping the components perpendicular to each other. Given a pair of gradients $(g_i, g_j)$ corresponding to sub-model $\phi_i$ and $\phi_j$ , if the cosine of their direction $\cos(g_i, g_j) < 0$ , they are conflict with each other. For a pair of conflicting gradients $(g_i, g_j)$ , we project them to the orthogonal plane of each other, respectively, as shown in Equation 1. + +$$ +g _ {j} ^ {\prime} = g _ {i} - \frac {g _ {j} * g _ {i}}{\left\| g _ {i} \right\| ^ {2}} * g _ {i}. \tag {1} +$$ + +Such projection results in a direction that improves both the $i$ -th and $j$ -th sub-model simultaneously. + +![](images/8c2bb56d278f8395edd91472d8d1c17a4c4037834b3b88552cef4823db39d1e4.jpg) +Figure 3. An example of gradient conflict between pair $(g_i, g_j)$ . Simply taking the summation aggregation results in the elimination of the improvement of the $i$ -th sub-model. Our aggregation module ensures the generic direction to improve both the $i$ -th and $j$ -th sub-model. + +# 3.4. Compatible Loss based on the Uncertainty Estimation + +The aggregation method aims to optimize all sub-models coordinately. However, it may not work as expected when facing the imbalance of gradient magnitude between sub-models, as shown in Figure 4. Therefore, it is critical to design a compatible loss. + +As sub-models are co-optimized, constraining them with existing compatible loss [6, 49] faces conflicts in the magnitude of gradients. As the sub-models vary in capacities, their optimization process converges at different speeds. The small-capacity sub-models will converge faster. Therefore, the small-capacity sub-model may have reached its + +![](images/43ead0d0647de42c59033410ee596870fe0267f53c8438e8b673842a265dda17.jpg) +Figure 4. An example of gradient conflict between pair $(g_i, g_j)$ when there is an imbalance of gradient magnitude. The $i$ -th sub-model contributes little to the optimization even using gradient projection. + +performance ceiling while still producing losses of large magnitude, resulting in the overfitting of such sub-model and weakening of other sub-models. Intuitively, a sub-model with low uncertainty tends to reach its performance ceiling. Therefore, it is expected to adaptively adjust the optimization priorities of each sub-models through their uncertainty dynamically during the training process. + +Here we propose a compatible loss based on the uncertainty estimation. Inspired by Dempster-Shafer Theory of evidence [16] (DST), we measure the uncertainty through belief, where high uncertainty corresponds to the model with small capacities. Typically, the belief is measured by a set of mass $B = \{b_{1}, b_{2}, \ldots, b_{C}\}$ and an overall uncertainty mass $u$ , where $b_{c}$ denotes the belief degree belonging to the class $c$ . $B$ and $u$ are related to a Dirichlet distribution, which satisfy Equation 2. + +$$ +u + \sum_ {c = 1} ^ {C} b _ {c} = 1, u \geq 0, b _ {c} \geq 0. \tag {2} +$$ + +The belief is estimated by the features extracted by sub-models. Given sample $i$ , its features extracted by the $j$ -th sub-model termed $\phi_j(i)$ . The evidence $E_i^j = \{e_{i1}^j, e_{i2}^j, \dots, e_{iC}^j\}$ is transformed by $\phi_j(i)$ through linear projection and activation function (i.e., $\exp(\cdot)$ ), where the linear projection layers share parameters with the classifier of the $1.0 \times$ model $\phi$ . The belief of sample $i$ on the $j$ -th sub-model is computed by Equation 3, where $\alpha_i^j = \{\alpha_{i1}^j, \alpha_{i2}^j, \dots, \alpha_{iC}^j\}$ is the Dirichlet intensity on all classes. + +$$ +\alpha_ {i c} ^ {j} = e _ {i c} ^ {j} + 1, +$$ + +$$ +S _ {i} ^ {j} = \sum_ {c = 1} ^ {C} \alpha_ {i c} ^ {j}, b _ {i c} ^ {j} = \frac {\alpha_ {i c} ^ {j} - 1}{S _ {i} ^ {j}}, u _ {i} ^ {j} = \frac {C}{S _ {i} ^ {j}}. \tag {3} +$$ + +With $\alpha_{i}^{j}$ , the Dirichlet distribution $\mathrm{D}(p_i|\alpha_i^j)$ is obtained, where $p_i$ is the class assignment probabilities on a simplex. Inspired by the Cross-Entropy loss, we integrate the probability $p_i$ on $\mathrm{D}(p_i|\alpha_i^j)$ , as shown in Equation 4. + +$$ +\begin{array}{l} \mathcal {L} (\alpha_ {i} ^ {j}) = \int [ \sum_ {c = 1} ^ {C} - y _ {i c} ^ {j} \log (p _ {i c} ^ {j}) ] \frac {1}{\beta (\alpha_ {i} ^ {j})} \Pi_ {c = 1} ^ {C} p _ {i c} ^ {j} ^ {(\alpha_ {i c} - 1)} \mathrm {d} p _ {i} ^ {j} \\ = \sum_ {c = 1} ^ {C} y _ {i c} ^ {j} \left(\psi \left(S _ {i} ^ {j}\right) - \psi \left(\alpha_ {i c} ^ {j}\right)\right), \tag {4} \\ \end{array} +$$ + +where $\psi (\cdot)$ is the digamma function. Since Equation 4 may converge to a trivial solution $\mathrm{D}(p_i|1)$ during optimization, we penalize the distance between $\mathrm{D}(p_i|1)$ and $\mathrm{D}(p_i|\alpha_i^j)$ as the regularization. The complete compatible loss $L_{c_j}$ is shown in Equation 5. $L_{c_j}$ adaptively adjusts the magnitude of gradients through their corresponding uncertainty. + +$$ +\begin{array}{l} K L \left[ \mathrm {D} \left(p _ {i} \mid \alpha_ {i} ^ {j}\right) | | \mathrm {D} \left(p _ {i} \mid \mathbf {1}\right) \right] = \log \left(\frac {\sum_ {c = 1} ^ {C} \Gamma \left(\alpha_ {i c} ^ {j}\right)}{\Gamma (C) \prod_ {c = 1} ^ {C} \left(\alpha_ {i c} ^ {j}\right)}\right) \\ + \sum_ {c = 1} ^ {C} \left(\alpha_ {i c} ^ {j} - 1\right) \left[ \psi \left(\alpha_ {i c} ^ {j} - \psi \left(\sum_ {c = 1} ^ {C} \alpha_ {i c} ^ {j}\right)\right) \right], \tag {5} \\ L _ {c _ {j}} (i) = \mathcal {L} \left(\alpha_ {i} ^ {j}\right) + \lambda * K L \left[ \mathrm {D} \left(p _ {i} \mid \alpha_ {i} ^ {j}\right) | | \mathrm {D} \left(p _ {i} \mid \mathbf {1}\right) \right]. \\ \end{array} +$$ + +We further conduct a gradient analysis for the proposed compatible loss. The Cross-Entropy loss is written as: + +$$ +p _ {i c} = \frac {e ^ {z _ {i c}}}{\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}}, L _ {c e} = \sum_ {c = 1} ^ {C} - y _ {i c} \log \left(p _ {i c}\right), \tag {6} +$$ + +where $z_{i}$ is the classification vector of sample $i$ outputted by the model. The corresponding gradient of Equation 6 can be computed as: + +$$ +\frac {\partial L _ {c e}}{\partial z _ {i c}} = \frac {e ^ {z _ {i c}}}{\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}} - y _ {i c}. \tag {7} +$$ + +The corresponding gradient of the proposed compatible loss in 5 can be computed as: + +$$ +\begin{array}{l} \frac {\partial L _ {c _ {j}} \left(\alpha_ {i} ^ {j}\right)}{\partial \alpha_ {i c} ^ {j}} = \psi^ {\prime} \left(\alpha_ {i c} ^ {j}\right) \left[ \lambda \left(\alpha_ {i c} ^ {j} - 1\right) - \mathbf {y} _ {i c} \right] \tag {8} \\ + \psi^ {'} (S _ {i} ^ {j}) [ - \lambda (S _ {i} ^ {j} - C) + 1 ]. \\ \end{array} +$$ + +We take $\frac{1}{\alpha_{ic}^j}$ as the approximation of $\psi'(\alpha_{ic}^j)$ and obtain Equation 9. + +$$ +\begin{array}{l} \frac {\partial L _ {c _ {j}} (\alpha_ {i} ^ {j})}{\partial \alpha_ {i c} ^ {j}} \approx \frac {1}{\alpha_ {i c} ^ {j}} [ \lambda (\alpha_ {i c} ^ {j} - 1) - \mathbf {y} _ {i c} ] \\ + \frac {1}{S _ {i} ^ {j}} [ - \lambda (S _ {i} ^ {j} - C) + 1 ] \tag {9} \\ = - (\frac {\lambda + y _ {i c}}{\alpha_ {i c} ^ {j}} + \frac {\lambda C + 1}{S _ {i} ^ {j}}), \\ \end{array} +$$ + +where the first term corresponds to the Cross-Entropy loss in Equation 7, and the second term corresponds to the uncertainty estimation. As the uncertainty $u_{i}^{j} = \frac{C}{S_{i}^{j}} \propto \frac{1}{S_{i}^{j}}$ , thus higher $u_{i}^{j}$ would result in a gradient with larger magnitude through $\frac{\lambda C + 1}{S_{i}^{j}}$ . Therefore, the magnitude of gradients corresponding to different sub-models is dynamically adjusted to avoid someone dominating the optimization process. + +Such compatible loss can work well with gradient projection, as shown in Algorithm 1. + +Algorithm 1 Training Process of SFSC +Require: batch input $\mathcal{B}$ the crop list $W = \{\gamma_0,\gamma_1,\dots ,\gamma_m\}$ ,the full model $\phi$ +1: $g_{ori}\gets \frac{\delta L_{ori}(\phi,\mathcal{B})}{\delta\phi}$ +2: for $\gamma_{j}\in W$ do +3: Switch to the sub-model $\phi_j$ +4: $g_{j}\leftarrow \frac{\delta L_{c_{j}}(\phi_{j},\mathcal{B})}{\delta\phi}$ +5: end for +6: $G,G^{p}\gets \{g_{ori},g_{0},\ldots g_{m}\}$ +7: for $g_{a}\in G^{p}$ do +8: Shuffle(G) +9: for $g_{b}\in G$ do +10: $d = g_{a}*g_{b}$ +11: if $d < 0$ then //have conflicts +12: $g_{a}\gets g_{a} - \frac{g_{b}*d}{||g_{b}||^{2}}$ +13: end if +14: end for +15: end for +16: return Update $\phi$ by $\Delta \theta = -\sum_{g_a\in GP}g_a$ + +In general, SFSC aims to achieve the Pareto optimality [9] among the optimization of sub-models. The proposed compatible loss adaptively adjusts the magnitude of gradients through uncertainty to control the optimization priorities of sub-models dynamically. Besides, the aggregation method performs the gradient projection to find a generic direction to improve all sub-models. + +# 4. Experiments + +# 4.1. Experimental Configuration + +Datasets. We evaluate the proposed SFSC on person ReID datasets, i.e., Market-1501 [68], MSMT17 [61], and one vehicle ReID dataset VeRi-776 [60]. Market1501 consists of 32,668 annotated images of 1,501 identities shot from 6 cameras. MSMT17 consists of 126,441 bounding boxes of 4,101 identities taken by 15 cameras. VeRi-776 contains over 50,000 images of 776 vehicles captured by 20 cameras. + +Comparison methods. For the one-to-one compatible learning paradigm, BCT [49] and Asymmetric [6] are adopted, as they are representatives of compatible learning through classification losses and embedding losses. BCT and Asymmetric can only obtain sub-models in a one-to-one compatible way. Specifically, given the fixed largest capacity model which is trained without compatible regularization, we train a series of small-capacity models to be compatible with it, respectively. Besides, we also incorporate BCT/Asymmetric loss in our proposed self-compatible learning paradigm for a fair comparison. We take the compatible loss in BCT/Asymmetric to constrain each sub-model compatible with the full model, and use summation to aggregate losses from sub-models, termed as "BCT-S" and "Asymmetric-S". + +Evaluation metrics. Mean average precision (mAP) + +
M(φq, φg) φq φqφ1/16φ1/4φ9/16φφ1/16φ1/4φ9/16φ
Unified ModelOurs
φ1/1655.25---58.1962.8863.4869.43
φ1/4-67.48--61.2470.7471.3776.37
φ9/16--71.25-62.4371.2572.0677.26
φ---80.9168.0374.7675.6781.43
BCT-SAsymmetric-S
φ1/1654.4058.1060.2261.2948.7951.9654.0956.31
φ1/455.7166.3868.2969.7152.2462.4564.9867.23
φ9/1656.5068.3068.4271.9954.4464.7467.9771.28
φ61.7969.7371.9373.6156.5467.3071.9178.38
BCTAsymmetric
φ1/1655.5556.6459.4060.3555.8350.8553.7155.39
φ1/455.8365.7467.1068.5452.6667.0462.0866.59
φ9/1655.6668.0267.4970.9354.4862.3366.3370.14
φ58.6969.4870.04-56.0866.6669.71-
+ +Table 2. Baseline performance comparison on Market1501 (mAP). "Unified Model" are models without any compatible regularization. + +
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
MSMT17
Unified Model28.8211.6948.6522.8657.5630.0670.2243.06
BCT-S46.4920.1156.2228.1358.1130.1458.7630.99
Asymmetric-S32.6413.3449.6224.0457.1830.0466.0840.39
Ours56.2228.1663.2935.3265.1337.7470.3843.89
VERI-776
Unified Model81.0744.7985.3053.6388.6955.4292.2066.50
BCT-S86.2549.3887.0857.0390.4258.2086.7358.46
Asymmetric-S68.7542.1684.2956.5588.0461.3789.8265.35
Ours79.7655.0489.9462.2890.4862.7292.3266.55
+ +Table 3. Performance comparison on MSMT17 and VeRi-776 datasets. + +
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
Circle Loss
Unified Model70.2550.7972.6660.4378.2164.6687.3874.06
BCT-S63.1842.4572.7452.0477.3859.4987.5374.58
Asymmetric-S69.6648.2177.8557.0280.5261.8186.9073.08
Ours73.9954.6180.5863.4682.3967.0288.9577.05
Softmax + Triplet
Unified Model77.5555.8283.9765.3087.8970.6990.9478.30
BCT-S81.6259.5887.8669.0588.7571.3890.1174.71
Asymmetric-S75.5652.6085.3664.2186.9168.8390.4577.34
Ours87.3666.2190.9476.6491.2177.9392.6481.46
+ +Table 4. Performance comparison between different loss functions on Market1501. + +and top-1 accuracy (R1) are adopted to evaluate the performances of retrieval. Define $M(\phi_{a},\phi_{b})$ as the performances of retrieval where the query features are extracted by model $\phi_{a}$ and gallery features are extracted by model $\phi_{b}$ . We set $\phi_{a}$ and $\phi_{b}$ as different sub-models to evaluate the perfor + +mance of compatible learning. + +Implementation details. The implementations of all compared methods and SFSC are based on the FastReLU [24]. The default configuration named "bag of tricks" [39] is adopted. $\lambda$ is set to 0.2 for all experiments. + +
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
ResNet-50
Unified Model85.2765.4488.8472.0790.876.4092.2282.13
BCT-S89.3172.0991.0077.2291.3678.2791.8679.40
Asymmetric-S85.0467.3990.2676.2891.6078.5391.9582.79
Ours90.4176.4091.0378.2091.6678.7693.4483.66
MobileNet V2
Unified Model69.0041.1489.5872.8591.5477.9992.2879.50
BCT-S80.4655.1788.4870.6889.8873.7289.9673.87
Asymmetric-S67.2843.6486.9768.9589.3173.8290.7478.45
Ours83.1163.8290.5376.5291.1878.3092.6181.34
ShuffleNet
Unified Model82.5460.1289.8564.4591.9572.9684.3879.13
BCT-S85.5165.7688.4266.8589.7370.5986.6774.41
Asymmetric-S82.3062.5084.4157.4388.4266.4378.9274.92
Ours89.2573.0989.9374.8689.9375.1492.8181.97
+ +# 4.2. Baseline Comparisons + +We conduct baseline experiments on Market1501. We take "Softmax + Cross Entropy" as the loss function, ResNet-18 [23] as the backbone, and $W = \{0.25 \times, 0.50 \times, 0.75 \times\}$ as the crop ratio list of the sub-models. Notes that for a model with crop ratio $\gamma_i$ , the ratio of its parameters to the full model is $\gamma_i^2$ . Therefore, the ratio of the number of parameters in the $0.25 \times$ , $0.50 \times$ and $0.75 \times$ sub-model to that in the full model are $\frac{1}{16}, \frac{1}{4}, \frac{9}{16}$ , respectively. + +Comparison methods. As for one-to-one compatible learning methods BCT and Asymmetric, the compatibility does not outperform that achieved by their corresponding self-compatible methods especially when the models are not constrained to be compatible directly. For self-compatible learning methods, BCT-S and Asymmetric-S, the conflicts among sub-models lead to a limited average accuracy. + +Our Method (SFSC). As we resolve the conflicts among different sub-models, our methods can achieve many-to-many compatibility among all the sub-models, which significantly outperforms all comparison methods. + +# 4.3. Performances under Different Settings + +Different datasets. We evaluate each self-compatible learning method on the other person ReID datasets MSMT17, and the vehicle ReID dataset VeRi-776 under baseline configurations, as shown in Table 3. The results on the two datasets with larger scales are consistent with that on the Market1501. Our methods outperform others significantly for the small capacity sub-model (e.g. the $0.25 \times$ sub-model $\phi_{\frac{1}{4}}$ ). + +Different loss functions. We replace the original loss + +function $L_{ori}$ from "Softmax" to "Softmax + Triplet" and "Circle Loss" [52] to evaluate the effectiveness of each method, while Triplet loss is a kind embedding loss and circle loss is a kind of classification loss. The results in Table 4 show that the accuracy of BCT-S and Asymmetric-S drops greatly when training with circle loss, as Our method still stably achieves the best performance. + +Different model architectures. We also study whether each method can be stably applied to various model architectures, i.e., ResNet-50 [23], MobileNet V2 [45] and ShuffleNet [67]. ShuffleNet and MobileNet are both designed especially for edge devices, which meet the application scenarios of self-compatible learning. It is shown in Table 5 that BCT-S and Asymmetric-S do not achieve higher accuracy than deploying a unified model with ShuffleNet and MobileNet (e.g. for the $0.75 \times$ sub-model $\phi_{\frac{9}{16}}$ ). And our methods always improve the accuracy in all cases. + +# 4.4. Ablation Study + +We replace the proposed compatible loss and aggregation module with existing compatible loss BCT [50] and summation aggregation to test the proposed two methods. + +Table 5. Performance comparison between different model architectures on Market1501. + +
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
Unified Model55.2567.4871.2580.91
w/o Ours Loss66.4174.5876.8677.87
w/o Ours Aggregation68.0873.7674.8475.38
SFSC69.4376.3777.2681.43
+ +Table 6. Ablation Study on Market1501. + +Table 6 shows that the proposed compatible loss mainly improves the retrieval accuracy of the small capacity sub-models while the proposed aggregation method can ensure the performance of the full model. + +![](images/a0d939f33c064f2251efe58c9745ee250bf0ecf1f8edcc940cf211ac8366b115.jpg) +Figure 5. Gradient visualization on Market1501-R18 baseline. As for the original cross-entropy loss (BCT [49]), the magnitude of the $0.25 \times$ sub-model is much larger than that of the $0.75 \times$ sub-model, especially at the early stage of the training process ( $0 \sim 1000$ iterations). Therefore, the $0.25 \times$ sub-model dominates the training progress and the improvements of the $0.75 \times$ sub-models are eliminated. While our method ensures co-optimization through uncertainty estimation. + +![](images/0418f3bd8988d888bfa2d350dd9a8267a43fc75606e329ca551bcc3c85dc09e8.jpg) + +# 4.5. Hyperparameter Analysis + +Here we verify the impact of hyperparameter settings on the performance of the proposed method. The key hyperparameter is the crop ratio list for sub-models $W$ . Intuitively, if the number of $W$ increases or the minimal crop ratio decreases, the compatibility constraints on model $\phi$ would be more strict, thereby increasing the difficulty of compatible learning. Therefore, we take the baseline configuration and set $W$ with different values. + +
M(φ1/16, φ)M(φ81/100, φ)M(φ169/400, φ)M(φ283/400, φ)M(φ, φ)
Unified Model55.2562.5469.5472.5780.27
BCT-S61.4568.6373.4272.8075.39
Asymmetric-S62.8769.4575.2073.1474.60
Ours69.6375.8177.1077.3680.59
+ +Table 7. Performance comparison with $W = \{0.25\times ,0.45\times ,0.65\times ,0.85\times \}$ on Market1501 (mAP). + +
M(φ1/100, φ)M(φ1/16, φ)M(φ9/16, φ)M(φ, φ)
Unified Model40.4055.2571.2580.27
BCT-S42.3459.7571.7376.39
Asymmetric-S42.5858.2070.5075.42
Ours45.4468.1878.0081.46
+ +Table 8. Performance comparison with $W = \{0.10\times ,0.25\times ,0.75\times \}$ on Market1501 (mAP). + +Table 7 shows the performances with the number of $W$ increases, while Table 8 shows the performances with the minimum among $W$ decreases. The result of the comparison methods has fluctuated to varying degrees. SFSC still achieves the best retrieval accuracy. + +# 4.6. Visualization + +We visualize the norm of the gradient during the training progress on Market1501-R18 baseline, as shown in Figure 5. We also visualize the cumulative number of gradient pairs which are conflict with each other during the training process. As shown in Figure 6, our method has found a generic direction to improve all sub-models. Therefore, there are almost no conflict pairs after 5000 iterations. As + +![](images/9bdcb1363751d8acb4ec965101591a89329bf776554b19f20be4e5544a4311b1.jpg) +Figure 6. The cumulative number of conflict pairs during the training process. + +for the summation aggregation, it faces fierce conflicts especially after 10000 iterations, resulting in the improvement of different sub-models overestimated or underestimated. + +# 5. Conclusion + +This paper proposes a Switchable Representation Learning Framework with Self Compatibility (SFSC) for multi-platform model collaboration. SFSC enables us to obtain models with various capacities to fit different computing and storage resource constraints on diverse platforms. It finds a generic direction to improve all sub-models by uncertainty estimation and gradient projection. SFSC achieves state-of-the-art performances and shows convincing robustness in different case studies. + +# 6. Acknowledgments + +This work was supported by the National Natural Science Foundation of China under Grant 62088102, the Basic and Frontier Research Project of PCL, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation. + +# References + +[1] Lu Bai, Yew-Soon Ong, Tiantian He, and Abhishek Gupta. Multi-task gradient descent for multi-task learning. Memetic Computing, 12(4):355-369, 2020. 4 +[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13349-13358, 2021. 3 +[3] Emmanuel Bengio, Pierre-Luc Bacon, Joelle Pineau, and Doina Precup. Conditional computation in neural networks for faster models. arXiv preprint arXiv:1511.06297, 2015. 2 +[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 2 +[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2 +[6] Mateusz Budnik and Yannis Avrithis. Asymmetric metric learning for knowledge transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2021. 1, 4, 5 +[7] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. 2 +[8] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 2 +[9] Yair Censor. Pareto optimality in multiobjective problems. Applied Mathematics and Optimization, 4(1):41-59, 1977. 5 +[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International Conference on Machine Learning, pages 794-803. PMLR, 2018. 2 +[11] Kyunghyun Cho and Yoshua Bengio. Exponentially increasing the capacity-to-computation ratio for conditional computation in deep learning. arXiv preprint arXiv:1406.7362, 2014. 2 +[12] Jifeng Dai, Haozhi Qi, Yuwen Xiong, and Yi Li. Guodong, zhang, han hu, and yichen wei. Deformable convolutional networks. In, ICCV, 2(3):7, 2017. 2 +[13] Yunshu Du, Wojciech M Czarnecki, Siddhant M Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224, 2018. 2 +[14] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, pages 3762-3773. PMLR, 2020. 2 +[15] Hang Gao, Xizhou Zhu, Steve Lin, and Jifeng Dai. Deformable kernels: Adapting effective receptive fields for object deformation. arXiv preprint arXiv:1910.02940, 2019. 2 + +[16] Jean Gordon and Edward H Shortliffe. The dampster-shafer theory of evidence. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, 3:832-838, 1984. 3, 4 +[17] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12259-12269, 2021. 1 +[18] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2 +[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2 +[20] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 3 +[21] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2 +[22] Hussein Hazimeh, Natalia Ponomareva, Petros Mol, Zhenyu Tan, and Rahul Mazumder. The tree ensemble layer: Differentiability meets conditional computation. In International Conference on Machine Learning, pages 4138-4148. PMLR, 2020. 2 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 4, 7 +[24] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631, 6(7):8, 2020. 6 +[25] Thomas M Hehn, Julian FP Kooij, and Fred A Hamprecht. End-to-end learning of decision trees and forests. International Journal of Computer Vision, 128(4):997-1011, 2020. 2 +[26] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2 +[27] Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2761, 2018. 2 +[28] Yani Ioannou, Duncan Robertson, Darko Zikic, Peter Kontschieder, Jamie Shotton, Matthew Brown, and Antonio Criminisi. Decision forests, convolutional networks and the models in-between. arXiv preprint arXiv:1603.01250, 2016. 2 +[29] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2 + +[30] Taejong Joo, Uijung Chung, and Min-Gwan Seo. Being bayesian about categorical probability. In International Conference on Machine Learning, pages 4950-4961. PMLR, 2020. 3 +[31] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 2 +[32] Igor Kononenko. Bayesian neural networks. Biological Cybernetics, 61(5):361-370, 1989. 3 +[33] Harold W Kuhn and Albert W Tucker. Nonlinear programming. In Traces and emergence of nonlinear programming, pages 247-258. Springer, 2014. 2 +[34] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2 +[35] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Trgp: Trust region gradient projection for continual learning. arXiv preprint arXiv:2202.02931, 2022. 2 +[36] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In Proceedings of the IEEE international conference on computer vision, pages 2736-2744, 2017. 2 +[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2 +[38] Yihang Lou, Ling-Yu Duan, Shiqi Wang, Ziqian Chen, Yan Bai, Changwen Chen, and Wen Gao. Front-end smart visual sensing and back-end intelligent analysis: A unified infrastructure for economizing the visual system of city brain. IEEE Journal on Selected Areas in Communications, 37(7):1489-1503, 2019. 1 +[39] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6 +[40] David JC MacKay. Bayesian neural networks and density networks. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 354(1):73-80, 1995. 3 +[41] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. Advances in neural information processing systems, 27, 2014. 2 +[42] Amir Rosenfeld and Shimon Ullman. Visual concept recognition and localization via iterative introspection. In *Asian Conference on Computer Vision*, pages 264-279. Springer, 2016. 2 +[43] Abhijit Guha Roy, Nassir Navab, and Christian Wachinger. Concurrent spatial and channel 'squeeze & excitation' in fully convolutional networks. In International conference on medical image computing and computer-assisted intervention, pages 421-429. Springer, 2018. 2 +[44] Manas Sahni, Shreya Varshini, Alind Khare, and Alexey Tumanov. Compofa: Compound once-for-all networks + +for faster multi-platform deployment. arXiv preprint arXiv:2104.12642, 2021. 2 +[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 4, 7 +[46] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2, 4 +[47] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in Neural Information Processing Systems, 31, 2018. 3 +[48] Yantao Shen, Fanzi Wu, and Ying Shan. Non-inherent feature compatible learning. 2020. 2 +[49] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 1, 4, 5, 8 +[50] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 7 +[51] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 2 +[52] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 7 +[53] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 2 +[54] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2 +[55] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2 +[56] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 2 +[57] Xinyi Wang, Ankur Bapna, Melvin Johnson, and Orhan Firat. Gradient-guided loss masking for neural machine translation. arXiv preprint arXiv:2102.13549, 2021. 2 +[58] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in + +convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2 +[59] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. Advances in Neural Information Processing Systems, 33:2432-2444, 2020. 2 +[60] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. Proceedings of the IEEE International Conference on Computer Vision, pages 379–387, 2017. 5 +[61] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 5 +[62] Shengsen Wu, Liang Chen, Yihang Lou, Yan Bai, Tao Bai, Minghua Deng, and Lingyu Duan. Neighborhood consensus contrastive learning for backward-compatible representation. arXiv preprint arXiv:2108.03372, 2021. 1, 2 +[63] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021. 2 +[64] Zhicheng Yan, Hao Zhang, Robinson Piramuthu, Vignesh Jagadeesh, Dennis DeCoste, Wei Di, and Yizhou Yu. Hd-cnn: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision, pages 2740-2748, 2015. 2 +[65] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018. 3 +[66] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 4 +[67] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4, 7 +[68] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5 +[69] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9308-9316, 2019. 2 \ No newline at end of file diff --git a/2023/Switchable Representation Learning Framework With Self-Compatibility/images.zip b/2023/Switchable Representation Learning Framework With Self-Compatibility/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..09dbaf2ceb86df78900d91d31acd0032f49adb8a --- /dev/null +++ b/2023/Switchable Representation Learning Framework With Self-Compatibility/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afef20c87dca4407b5c9be89c390850993b5ee55e7f12f9bcc9ae68ce0dafc8a +size 649280 diff --git a/2023/Switchable Representation Learning Framework With Self-Compatibility/layout.json b/2023/Switchable Representation Learning Framework With Self-Compatibility/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e4c5a01e4398d835ddd7e89f6080ec0b3621b5e5 --- /dev/null +++ b/2023/Switchable Representation Learning Framework With Self-Compatibility/layout.json @@ -0,0 +1,10148 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 519, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 519, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 519, + 121 + ], + "type": "text", + "content": "Switchable Representation Learning Framework with Self-compatibility" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": "Shengsen Wu" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": ", Yan Bai" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": ", Yihang Lou" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": ", Xiongkun Linghu" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": ", Jianzhong He" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": ", Ling-Yu Duan" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": " Peng Cheng Laboratory, China; " + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": " Peking University, China; " + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": " Huawei Technologies; " + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 142, + 529, + 186 + ], + "type": "text", + "content": " Tsinghua University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 193, + 471, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 193, + 471, + 220 + ], + "spans": [ + { + "bbox": [ + 119, + 193, + 471, + 220 + ], + "type": "text", + "content": "{sswu, yanbai, lingyu}@pku.edu.cn, {louyihang1, jianzhong.he}@huawei.com, lhxk20@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 272, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 289, + 535 + ], + "type": "text", + "content": "Real-world visual search systems involve deployments on multiple platforms with different computing and storage resources. Deploying a unified model that suits the minimal-constrain platforms leads to limited accuracy. It is expected to deploy models with different capacities adapting to the resource constraints, which requires features extracted by these models to be aligned in the metric space. The method to achieve feature alignments is called \"compatible learning\". Existing research mainly focuses on the one-to-one compatible paradigm, which is limited in learning compatibility among multiple models. We propose a Switchable representation learning Framework with Self-Compatibility (SFSC). SFSC generates a series of compatible sub-models with different capacities through one training process. The optimization of sub-models faces gradients conflict, and we mitigate this problem from the perspective of the magnitude and direction. We adjust the priorities of sub-models dynamically through uncertainty estimation to co-optimize sub-models properly. Besides, the gradients with conflicting directions are projected to avoid mutual interference. SFSC achieves state-of-the-art performance on the evaluated datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 557, + 128, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 128, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 128, + 569 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 577, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 696 + ], + "type": "text", + "content": "Visual search systems are widely deployed, which recall the nearest neighbors in gallery features according to their distances to the query feature. Real-world visual search systems consist of multiple models deployed on different platforms [17, 38, 45] (e.g. clouds, mobiles, smart cameras), where different platforms interact with each other by visual features. Typically, as for person re-identification systems, images are captured and processed into features on edge sides. And such features are sent to the cloud side to compare with the database features for identifying" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 247, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 545, + 392 + ], + "type": "text", + "content": "a specific person by feature similarities. As diverse platforms meet different computing and storage resource limitations, deploying the unified model which suits the minimal-constrain platforms leads to a waste of resources on other platforms and limited accuracy. To make better use of resources and achieve higher accuracy, it is expected to deploy models with different capacities which adapt to the resource limitations. Such a solution requires compatibility among the models to satisfy their interaction requirements, which means that the similar data processed by different models are close to each other in the feature space while dissimilar data are far apart." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 481 + ], + "type": "text", + "content": "To achieve compatibility, compatible learning methods have been proposed. Existing research focuses on the one-to-one compatible paradigm [6,49,62], which constrains the learned features of a latter (learnable) model to be compatible with its previous (fixed) version. They are limited in achieving many-to-many compatibilities, which means any two in a series of models are compatible with each other." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "In this work, we propose a Switchable representation learning Framework with Self-Compatibility (SFSC) to achieve many-to-many compatibilities. As shown in Figure 1, SFSC can deploy different sub-models to suit the diverse computing and storage resource limitation of different platforms. The compatibility between any two sub-models can be achieved, termed as \"self-compatibility\". However, there are gradients conflicts between sub-models as they are co-optimized during the training process. We summarize such conflicts into the gradient magnitude and gradient direction. Specifically, a gradient of a large magnitude will dominate the optimization process, which may lead to the overfitting of its corresponding sub-model and impair the improvements of other sub-models. To solve this problem, we estimate the uncertainty of different sub-models to weight the gradient, denoting the optimization priorities. Besides, as the sub-models may produce gradients in various directions, there is mutual interference between them. Thus the improvement in different sub-models may be overestimated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15943" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 72, + 476, + 187 + ], + "blocks": [ + { + "bbox": [ + 124, + 72, + 476, + 187 + ], + "lines": [ + { + "bbox": [ + 124, + 72, + 476, + 187 + ], + "spans": [ + { + "bbox": [ + 124, + 72, + 476, + 187 + ], + "type": "image", + "image_path": "70e765d57ccf0a68bef5a3b047029b61a69e6e0bb8e330762e2af7676e2b3878.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": "Figure 1. An example of SFSC in multi-platform deployments. The switchable neural network generates a series of sub-networks " + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "inline_equation", + "content": "\\{\\phi_1, \\phi_2, \\dots, \\phi_m\\}" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": " with different capacities for different computing resources. " + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "inline_equation", + "content": "\\forall i, j \\in \\{1, 2, \\dots, m\\}" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "inline_equation", + "content": "|\\phi_i(a) - \\phi_j(p)| < |\\phi_i(a) - \\phi_j(n)|" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "inline_equation", + "content": "(a, p)" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": " is a pair of samples with the same class ID, while " + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "inline_equation", + "content": "(a, n)" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": " is a pair of samples with different class IDs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "type": "text", + "content": "or underestimated. To tackle such conflict, the gradients are projected onto planes that are orthogonal to each other." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 272, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 297 + ], + "type": "text", + "content": "The contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 303, + 287, + 534 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 303, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 303, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 287, + 363 + ], + "type": "text", + "content": "- We propose a switchable representation learning framework with self-compatibility (SFSC). SFSC generates a series of feature-compatible sub-models with different capacities for visual search systems, which can be deployed on different platforms flexibly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 370, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 370, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 287, + 468 + ], + "type": "text", + "content": "- We resolve the conflicts between sub-models from the aspect of gradient magnitude and gradient direction. A compatible loss based on uncertainty estimation is proposed to guide optimization priorities and alleviate the imbalance of gradient magnitude between sub-models. An aggregation method based on the gradient projection is proposed to avoid mutual interference and find a generic optimal direction for all sub-models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "type": "text", + "content": "- SFSC achieves state-of-the-art performance on the evaluated benchmark datasets. Compared to deploying a unified model, adopting SFSC to obtain different sub-models can achieve " + }, + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 58, + 474, + 287, + 534 + ], + "type": "text", + "content": " performance improvements on three different datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 544, + 135, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 135, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 135, + 556 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 275, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 275, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 275, + 577 + ], + "type": "text", + "content": "2.1. One-to-one Compatible Learning Paradigm" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Given a fixed old model, the one-to-one compatible learning methods aim at aligning the learnable models with the fixed old one in metric space, termed as \"backward compatible learning\". Existing methods focus on the optimization difficulties caused by the old model. Yan et al. [63] propose Focal Distillation to reduce the \"negative flips\" which means a test sample is correctly classified by the old model but incorrectly classified by the new one. Wu et al. [62] start from neighborhood structures and classification entropy to eliminate the effect of outliers in old features. Shen et al. [48] propose a pseudo classifier and enhance it with a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 248, + 547, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 547, + 308 + ], + "type": "text", + "content": "random walk algorithm to get rid of the old classifier. Different from these one-to-one compatible learning methods, this paper studies a new compatible learning paradigm that aims to learn many-to-many compatibility among multiple learnable models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 318, + 454, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 454, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 454, + 332 + ], + "type": "text", + "content": "2.2. Dynamic Neural Networks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 337, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 553 + ], + "type": "text", + "content": "Dynamic neural networks are able to adapt their structures or parameters to the input during inference [19]. In recent years, research in this field has mainly focused on dynamic depth [5, 18, 26, 54, 55, 58], dynamic width [3, 4, 11, 27, 29, 36], dynamic routing [22, 25, 28, 64], dynamic parameters [12, 15, 21, 51, 69], etc. In addition to hand-crafted methods, Neural Architecture Search (NAS) method is also adopted to train and optimize multiple sub-nets with various capacities in one-time training [7, 8, 44, 53]. Dynamic neural networks can allocate computations on demand for different platforms by selectively activating model components. They have been used in classification [41, 42, 59], segmentation [31, 34, 43], and detection tasks [12, 56, 69]. In general, current dynamic neural networks output with definite semantic information (e.g. class id, detection box) which are naturally interoperable. However, for retrieval tasks, the networks output with features, which requires feature compatibility." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 563, + 518, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 518, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 518, + 576 + ], + "type": "text", + "content": "2.3. Gradient-based Multi-task Optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "From the gradient perspective, multi-task optimization is to aggregate gradients corresponding to different tasks and find a descent direction that decreases all objectives. Existing methods treat it as a nonlinear programming problem and take the multi-objective Karush-Kuhn-Tucker [33] (KKT) conditions to perform gradient projection and rescaling under different assumptions on tasks, e.g. multiple tasks which are equally important [10, 46, 66], main tasks with some auxiliary tasks [13, 57], a series of the tasks which are learned sequentially [14, 35, 37]. This paper treats optimizations of multiple sub-models as multiple tasks. As" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15944" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 69, + 486, + 174 + ], + "blocks": [ + { + "bbox": [ + 109, + 69, + 486, + 174 + ], + "lines": [ + { + "bbox": [ + 109, + 69, + 486, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 69, + 486, + 174 + ], + "type": "image", + "image_path": "87150a85c8cbcecf8d2b8dd20aa49ecb3141df1c0b332e505b983f8a2f8ccc58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "text", + "content": "Figure 2. Overview of SFSC. We first transform the full modelcalculate the compatible learning loss " + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "inline_equation", + "content": "\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}" + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "text", + "content": " on different sub-models with various capacitie. " + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "text", + "content": " into the sub-models. Then we calculate the compatible learning loss " + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "inline_equation", + "content": "\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}" + }, + { + "bbox": [ + 46, + 182, + 547, + 217 + ], + "type": "text", + "content": " on different sub-models with various capacities. Finally, we optimize the sub-models by aggregating losses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 288, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 298 + ], + "type": "text", + "content": "the sub-models vary greatly in their capacity, the optimization of different models converges at different speeds. The challenge is to dynamically adjust the priorities of different sub-models to avoid the overfitting of someones as well as underfitting of others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 307, + 191, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 191, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 191, + 320 + ], + "type": "text", + "content": "2.4. Evidential Deep Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 290, + 531 + ], + "type": "text", + "content": "Evidential deep learning [47] is used to model the uncertainty of deterministic neural networks. Compared with the Bayesian neural networks [32, 40], evidential deep learning methods have the advantage of low computational complexity, as there is no need for complex posterior probability estimates and multiple Mento Carlo samplings. Based on the DST [16] (Dempster-Shafer Theory of Evidence), evidential deep learning treats the output of the neural networks as subjective opinions and formalizes them as Dirichlet distribution. It has been used in classification tasks to overcome the overconfidence problem of softmax cross entropy [30], action recognition task to identify unknown samples [2], and multi-view learning task [20] to dynamically integrate different views. This paper treats the output features of different sub-models as evidence to access uncertainty and relates the priority of different sub-models with the obtained uncertainty." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 543, + 107, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 107, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 107, + 555 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 564, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 624 + ], + "type": "text", + "content": "In this section, we will describe the proposed SFSC method. We first introduce the construction of switchable neural networks. Then, we analyze the difficulties for optimizations and propose our compatible loss as well as the gradient aggregation method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 118, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 118, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 118, + 645 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "SFSC involves the training process of multiple sub-models whose features are compatible. Sub-models with different capacities are transformed from the full model " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " according to the crop ratio list " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". We termed the crop ratio list of each sub-model as " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "W = \\{\\gamma_{1},\\gamma_{2},\\dots,\\gamma_{m}\\}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "type": "text", + "content": "denotes the ratio of the width of the sub-model " + }, + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "type": "text", + "content": " to the full model " + }, + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 237, + 545, + 260 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "content": "As shown in Figure 2, for each batch input, SFSC first calculates the original loss " + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "L_{ori}" + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "content": ". The compatible loss " + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\{L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}" + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "content": " is calculated on different sub-models. By aggregating losses " + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\{L_{ori}, L_{c_1}, L_{c_2}, \\dots, L_{c_m}\\}" + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "content": ", SFSC computes the general gradient and updates the full model " + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 262, + 545, + 334 + ], + "type": "text", + "content": " by gradient descent methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 343, + 545, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 343, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 343, + 545, + 368 + ], + "type": "text", + "content": "3.2. Switchable Neural Network for Sub-models Construction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "text", + "content": "We first convert the traditional convolutional neural network into its corresponding switchable neural network. As the platforms generally perform model inference layer by layer, the resource limitation is more related to the widths of layers than the number of layers. Here we perform channel pruning on " + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "text", + "content": " with crop ratio " + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "text", + "content": " to obtain sub-model " + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "inline_equation", + "content": "\\phi_{j}" + }, + { + "bbox": [ + 304, + 375, + 546, + 495 + ], + "type": "text", + "content": ". Specifically, the classic convolutional neural network is mainly composed of some basic modules, such as convolution, fully connected layer, and batch normalization. We modify these basic modules as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 496, + 545, + 652 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "text", + "content": "1) For convolution (Conv) and fully connected (FC) layers, if their output is not the final feature for retrieval, we perform pruning in both input and output channels. Otherwise, pruning is only performed in input channels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": "2) For batch normalization (BN), we assign an independent BN to the sub-model " + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 304, + 545, + 545, + 652 + ], + "type": "text", + "content": ". As the intermediate features extracted by different sub-models vary greatly in distributions [65], channel pruning can not be performed on BN. Fortunately, the parameters of BN only occupy a small part of all parameters of the entire network (shown in Table 1). Such an operation only results in a slight increase in storage resources in training progress but does not incur additional deployment costs." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Through such design, we obtain the sub-model " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " with the following characteristics: 1) The features extracted by different sub-models have the same dimensions. 2) The ratio of layer widths in the sub-model " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " to that in the full model " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " is around " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_j" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": ". 3) The sub-model " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " reuses the pa" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15945" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 105 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 105 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 105 + ], + "type": "table", + "html": "
NetWorksResNet-18 [23]ResNet-50 [23]MobileNet V2 [45]ShuffleNet [67]
Conv and FC11,679,91225,503,9123,470,8705,402,826
99.918%99.792%99.027%99.102%
BN9,60053,12034,11248,960
0.0820%0.208%0.973%0.898%
", + "image_path": "4300379dd771e7b9aae344c9a8fafd844b62fce87b21c233798285035fa6fc2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 287, + 131 + ], + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 131 + ], + "type": "text", + "content": "Table 1. Number and proportion of different modules' parameters in different network structures" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 152, + 234, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 234, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 234, + 164 + ], + "type": "text", + "content": "rameters from the full model " + }, + { + "bbox": [ + 47, + 152, + 234, + 164 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 152, + 234, + 164 + ], + "type": "text", + "content": " except for BN." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 171, + 287, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 287, + 196 + ], + "type": "text", + "content": "3.3. Aggregation Method based on the Gradients Projection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 201, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 262 + ], + "type": "text", + "content": "As the gradients corresponding to different sub-models may have different directions, simply aggregating the gradients by summation may cause mutual interference, which means the improvements of sub-models are overestimated or underestimated, as shown in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": "Inspired by multi-task learning [1,46,66], we tackle such a problem by projecting conflicting gradients and keeping the components perpendicular to each other. Given a pair of gradients " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "(g_i, g_j)" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": " corresponding to sub-model " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": ", if the cosine of their direction " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\cos(g_i, g_j) < 0" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": ", they are conflict with each other. For a pair of conflicting gradients " + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "inline_equation", + "content": "(g_i, g_j)" + }, + { + "bbox": [ + 46, + 262, + 287, + 357 + ], + "type": "text", + "content": ", we project them to the orthogonal plane of each other, respectively, as shown in Equation 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 357, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 357, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 119, + 357, + 287, + 381 + ], + "type": "interline_equation", + "content": "g _ {j} ^ {\\prime} = g _ {i} - \\frac {g _ {j} * g _ {i}}{\\left\\| g _ {i} \\right\\| ^ {2}} * g _ {i}. \\tag {1}", + "image_path": "4d4dafe35b5bc682672c592167d5c964c48ea195faa5c755119262063b209e63.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "text", + "content": "Such projection results in a direction that improves both the " + }, + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "text", + "content": "-th sub-model simultaneously." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 53, + 422, + 281, + 482 + ], + "blocks": [ + { + "bbox": [ + 53, + 422, + 281, + 482 + ], + "lines": [ + { + "bbox": [ + 53, + 422, + 281, + 482 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 281, + 482 + ], + "type": "image", + "image_path": "8c2bb56d278f8395edd91472d8d1c17a4c4037834b3b88552cef4823db39d1e4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "text", + "content": "Figure 3. An example of gradient conflict between pair " + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "inline_equation", + "content": "(g_i, g_j)" + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "text", + "content": ". Simply taking the summation aggregation results in the elimination of the improvement of the " + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "text", + "content": "-th sub-model. Our aggregation module ensures the generic direction to improve both the " + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 490, + 287, + 544 + ], + "type": "text", + "content": "-th sub-model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "content": "3.4. Compatible Loss based on the Uncertainty Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "The aggregation method aims to optimize all sub-models coordinately. However, it may not work as expected when facing the imbalance of gradient magnitude between sub-models, as shown in Figure 4. Therefore, it is critical to design a compatible loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "As sub-models are co-optimized, constraining them with existing compatible loss [6, 49] faces conflicts in the magnitude of gradients. As the sub-models vary in capacities, their optimization process converges at different speeds. The small-capacity sub-models will converge faster. Therefore, the small-capacity sub-model may have reached its" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 538, + 128 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 538, + 128 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 538, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 538, + 128 + ], + "type": "image", + "image_path": "43ead0d0647de42c59033410ee596870fe0267f53c8438e8b673842a265dda17.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "type": "text", + "content": "Figure 4. An example of gradient conflict between pair " + }, + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "type": "inline_equation", + "content": "(g_i, g_j)" + }, + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "type": "text", + "content": " when there is an imbalance of gradient magnitude. The " + }, + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 137, + 545, + 182 + ], + "type": "text", + "content": "-th sub-model contributes little to the optimization even using gradient projection." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 189, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 545, + 272 + ], + "type": "text", + "content": "performance ceiling while still producing losses of large magnitude, resulting in the overfitting of such sub-model and weakening of other sub-models. Intuitively, a sub-model with low uncertainty tends to reach its performance ceiling. Therefore, it is expected to adaptively adjust the optimization priorities of each sub-models through their uncertainty dynamically during the training process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": "Here we propose a compatible loss based on the uncertainty estimation. Inspired by Dempster-Shafer Theory of evidence [16] (DST), we measure the uncertainty through belief, where high uncertainty corresponds to the model with small capacities. Typically, the belief is measured by a set of mass " + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "inline_equation", + "content": "B = \\{b_{1}, b_{2}, \\ldots, b_{C}\\}" + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": " and an overall uncertainty mass " + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "inline_equation", + "content": "b_{c}" + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": " denotes the belief degree belonging to the class " + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 273, + 545, + 380 + ], + "type": "text", + "content": " are related to a Dirichlet distribution, which satisfy Equation 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 382, + 545, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 382, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 362, + 382, + 545, + 414 + ], + "type": "interline_equation", + "content": "u + \\sum_ {c = 1} ^ {C} b _ {c} = 1, u \\geq 0, b _ {c} \\geq 0. \\tag {2}", + "image_path": "ea3b510b75ce7005401d3471a7e336108d0594a9d7a77d5cabdafe81ba24b40a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": "The belief is estimated by the features extracted by sub-models. Given sample " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": ", its features extracted by the " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": "-th sub-model termed " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\phi_j(i)" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": ". The evidence " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "E_i^j = \\{e_{i1}^j, e_{i2}^j, \\dots, e_{iC}^j\\}" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": " is transformed by " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\phi_j(i)" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": " through linear projection and activation function (i.e., " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\exp(\\cdot)" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": "), where the linear projection layers share parameters with the classifier of the " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "1.0 \\times" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": " model " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": ". The belief of sample " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": "-th sub-model is computed by Equation 3, where " + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\alpha_i^j = \\{\\alpha_{i1}^j, \\alpha_{i2}^j, \\dots, \\alpha_{iC}^j\\}" + }, + { + "bbox": [ + 305, + 417, + 545, + 528 + ], + "type": "text", + "content": " is the Dirichlet intensity on all classes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 399, + 529, + 459, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 529, + 459, + 544 + ], + "spans": [ + { + "bbox": [ + 399, + 529, + 459, + 544 + ], + "type": "interline_equation", + "content": "\\alpha_ {i c} ^ {j} = e _ {i c} ^ {j} + 1,", + "image_path": "a18b6da2da3eb71114db8b05569562ea63882490c28445febb1f50b6dc0188b5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 546, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 546, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 345, + 546, + 545, + 578 + ], + "type": "interline_equation", + "content": "S _ {i} ^ {j} = \\sum_ {c = 1} ^ {C} \\alpha_ {i c} ^ {j}, b _ {i c} ^ {j} = \\frac {\\alpha_ {i c} ^ {j} - 1}{S _ {i} ^ {j}}, u _ {i} ^ {j} = \\frac {C}{S _ {i} ^ {j}}. \\tag {3}", + "image_path": "eb48d0e1b9a3594e16481cc29bb1d0f53aeb9f6752c155c2d22e471f11a3fe74.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{j}" + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "content": ", the Dirichlet distribution " + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(p_i|\\alpha_i^j)" + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "content": " is obtained, where " + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "content": " is the class assignment probabilities on a simplex. Inspired by the Cross-Entropy loss, we integrate the probability " + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(p_i|\\alpha_i^j)" + }, + { + "bbox": [ + 304, + 580, + 545, + 629 + ], + "type": "text", + "content": ", as shown in Equation 4." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 631, + 545, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 631, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 310, + 631, + 545, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\alpha_ {i} ^ {j}) = \\int [ \\sum_ {c = 1} ^ {C} - y _ {i c} ^ {j} \\log (p _ {i c} ^ {j}) ] \\frac {1}{\\beta (\\alpha_ {i} ^ {j})} \\Pi_ {c = 1} ^ {C} p _ {i c} ^ {j} ^ {(\\alpha_ {i c} - 1)} \\mathrm {d} p _ {i} ^ {j} \\\\ = \\sum_ {c = 1} ^ {C} y _ {i c} ^ {j} \\left(\\psi \\left(S _ {i} ^ {j}\\right) - \\psi \\left(\\alpha_ {i c} ^ {j}\\right)\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "085bac1bb1b65904ed96c598750bad1f696fbc8d9b7854588f744f9cad22827f.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "15946" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " is the digamma function. Since Equation 4 may converge to a trivial solution " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(p_i|1)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " during optimization, we penalize the distance between " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(p_i|1)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathrm{D}(p_i|\\alpha_i^j)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " as the regularization. The complete compatible loss " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "L_{c_j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " is shown in Equation 5. " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "L_{c_j}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " adaptively adjusts the magnitude of gradients through their corresponding uncertainty." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 161, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 161, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 52, + 161, + 287, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K L \\left[ \\mathrm {D} \\left(p _ {i} \\mid \\alpha_ {i} ^ {j}\\right) | | \\mathrm {D} \\left(p _ {i} \\mid \\mathbf {1}\\right) \\right] = \\log \\left(\\frac {\\sum_ {c = 1} ^ {C} \\Gamma \\left(\\alpha_ {i c} ^ {j}\\right)}{\\Gamma (C) \\prod_ {c = 1} ^ {C} \\left(\\alpha_ {i c} ^ {j}\\right)}\\right) \\\\ + \\sum_ {c = 1} ^ {C} \\left(\\alpha_ {i c} ^ {j} - 1\\right) \\left[ \\psi \\left(\\alpha_ {i c} ^ {j} - \\psi \\left(\\sum_ {c = 1} ^ {C} \\alpha_ {i c} ^ {j}\\right)\\right) \\right], \\tag {5} \\\\ L _ {c _ {j}} (i) = \\mathcal {L} \\left(\\alpha_ {i} ^ {j}\\right) + \\lambda * K L \\left[ \\mathrm {D} \\left(p _ {i} \\mid \\alpha_ {i} ^ {j}\\right) | | \\mathrm {D} \\left(p _ {i} \\mid \\mathbf {1}\\right) \\right]. \\\\ \\end{array}", + "image_path": "9b8702920177c1329201891ce65fb70e86b29dbfcd0f80df905bfbbc29d58d99.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 252, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 287, + 277 + ], + "type": "text", + "content": "We further conduct a gradient analysis for the proposed compatible loss. The Cross-Entropy loss is written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 278, + 287, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 278, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 77, + 278, + 287, + 310 + ], + "type": "interline_equation", + "content": "p _ {i c} = \\frac {e ^ {z _ {i c}}}{\\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}}, L _ {c e} = \\sum_ {c = 1} ^ {C} - y _ {i c} \\log \\left(p _ {i c}\\right), \\tag {6}", + "image_path": "8ff2c589eed6b338004f098b9101b8491d65eae5ee948698ec2ffc086e4d5b12.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "content": " is the classification vector of sample " + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 317, + 287, + 354 + ], + "type": "text", + "content": " outputted by the model. The corresponding gradient of Equation 6 can be computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 361, + 287, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 361, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 111, + 361, + 287, + 390 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L _ {c e}}{\\partial z _ {i c}} = \\frac {e ^ {z _ {i c}}}{\\sum_ {k = 1} ^ {C} e ^ {z _ {i k}}} - y _ {i c}. \\tag {7}", + "image_path": "e65d223d7e14cdb3c68b42114fbee35889e7521d76eb723a78c6ca86f744c108.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 419 + ], + "type": "text", + "content": "The corresponding gradient of the proposed compatible loss in 5 can be computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 421, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 421, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 85, + 421, + 287, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial L _ {c _ {j}} \\left(\\alpha_ {i} ^ {j}\\right)}{\\partial \\alpha_ {i c} ^ {j}} = \\psi^ {\\prime} \\left(\\alpha_ {i c} ^ {j}\\right) \\left[ \\lambda \\left(\\alpha_ {i c} ^ {j} - 1\\right) - \\mathbf {y} _ {i c} \\right] \\tag {8} \\\\ + \\psi^ {'} (S _ {i} ^ {j}) [ - \\lambda (S _ {i} ^ {j} - C) + 1 ]. \\\\ \\end{array}", + "image_path": "8e9d8c59f17e9df75ad92ef9c14781c62adae7b8a0563d82522279b2d458aa8d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "type": "text", + "content": "We take " + }, + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\alpha_{ic}^j}" + }, + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "type": "text", + "content": " as the approximation of " + }, + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\psi'(\\alpha_{ic}^j)" + }, + { + "bbox": [ + 47, + 474, + 287, + 502 + ], + "type": "text", + "content": " and obtain Equation 9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 503, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 503, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 93, + 503, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial L _ {c _ {j}} (\\alpha_ {i} ^ {j})}{\\partial \\alpha_ {i c} ^ {j}} \\approx \\frac {1}{\\alpha_ {i c} ^ {j}} [ \\lambda (\\alpha_ {i c} ^ {j} - 1) - \\mathbf {y} _ {i c} ] \\\\ + \\frac {1}{S _ {i} ^ {j}} [ - \\lambda (S _ {i} ^ {j} - C) + 1 ] \\tag {9} \\\\ = - (\\frac {\\lambda + y _ {i c}}{\\alpha_ {i c} ^ {j}} + \\frac {\\lambda C + 1}{S _ {i} ^ {j}}), \\\\ \\end{array}", + "image_path": "060780f73b3ac35aca58082c183ad3ae4bd467fa9cd91632c711e8e6517d8a75.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "text", + "content": "where the first term corresponds to the Cross-Entropy loss in Equation 7, and the second term corresponds to the uncertainty estimation. As the uncertainty " + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "inline_equation", + "content": "u_{i}^{j} = \\frac{C}{S_{i}^{j}} \\propto \\frac{1}{S_{i}^{j}}" + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "text", + "content": ", thus higher " + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "inline_equation", + "content": "u_{i}^{j}" + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "text", + "content": " would result in a gradient with larger magnitude through " + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\frac{\\lambda C + 1}{S_{i}^{j}}" + }, + { + "bbox": [ + 47, + 597, + 287, + 689 + ], + "type": "text", + "content": ". Therefore, the magnitude of gradients corresponding to different sub-models is dynamically adjusted to avoid someone dominating the optimization process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 689, + 287, + 714 + ], + "type": "text", + "content": "Such compatible loss can work well with gradient projection, as shown in Algorithm 1." + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 306, + 87, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 467, + 84 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 467, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 467, + 84 + ], + "type": "text", + "content": "Algorithm 1 Training Process of SFSC" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": "Require: batch input " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " the crop list " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "W = \\{\\gamma_0,\\gamma_1,\\dots ,\\gamma_m\\}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " ,the full model " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "g_{ori}\\gets \\frac{\\delta L_{ori}(\\phi,\\mathcal{B})}{\\delta\\phi}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\gamma_{j}\\in W" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " do \n3: Switch to the sub-model " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "g_{j}\\leftarrow \\frac{\\delta L_{c_{j}}(\\phi_{j},\\mathcal{B})}{\\delta\\phi}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n5: end for \n6: " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "G,G^{p}\\gets \\{g_{ori},g_{0},\\ldots g_{m}\\}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n7: for " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "g_{a}\\in G^{p}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " do \n8: Shuffle(G) \n9: for " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "g_{b}\\in G" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " do \n10: " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "d = g_{a}*g_{b}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n11: if " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "d < 0" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " then //have conflicts \n12: " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "g_{a}\\gets g_{a} - \\frac{g_{b}*d}{||g_{b}||^{2}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " \n13: end if \n14: end for \n15: end for \n16: return Update " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 306, + 87, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\Delta \\theta = -\\sum_{g_a\\in GP}g_a" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 289, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 373 + ], + "type": "text", + "content": "In general, SFSC aims to achieve the Pareto optimality [9] among the optimization of sub-models. The proposed compatible loss adaptively adjusts the magnitude of gradients through uncertainty to control the optimization priorities of sub-models dynamically. Besides, the aggregation method performs the gradient projection to find a generic direction to improve all sub-models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 385, + 387, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 385, + 387, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 387, + 399 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 406, + 461, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 461, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 461, + 419 + ], + "type": "text", + "content": "4.1. Experimental Configuration" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 425, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 519 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed SFSC on person ReID datasets, i.e., Market-1501 [68], MSMT17 [61], and one vehicle ReID dataset VeRi-776 [60]. Market1501 consists of 32,668 annotated images of 1,501 identities shot from 6 cameras. MSMT17 consists of 126,441 bounding boxes of 4,101 identities taken by 15 cameras. VeRi-776 contains over 50,000 images of 776 vehicles captured by 20 cameras." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": "Comparison methods. For the one-to-one compatible learning paradigm, BCT [49] and Asymmetric [6] are adopted, as they are representatives of compatible learning through classification losses and embedding losses. BCT and Asymmetric can only obtain sub-models in a one-to-one compatible way. Specifically, given the fixed largest capacity model which is trained without compatible regularization, we train a series of small-capacity models to be compatible with it, respectively. Besides, we also incorporate BCT/Asymmetric loss in our proposed self-compatible learning paradigm for a fair comparison. We take the compatible loss in BCT/Asymmetric to constrain each sub-model compatible with the full model, and use summation to aggregate losses from sub-models, termed as \"BCT-S\" and \"Asymmetric-S\"." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Evaluation metrics. Mean average precision (mAP)" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15947" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 61, + 481, + 275 + ], + "blocks": [ + { + "bbox": [ + 111, + 61, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 111, + 61, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 111, + 61, + 481, + 275 + ], + "type": "table", + "html": "
M(φq, φg) φq φqφ1/16φ1/4φ9/16φφ1/16φ1/4φ9/16φ
Unified ModelOurs
φ1/1655.25---58.1962.8863.4869.43
φ1/4-67.48--61.2470.7471.3776.37
φ9/16--71.25-62.4371.2572.0677.26
φ---80.9168.0374.7675.6781.43
BCT-SAsymmetric-S
φ1/1654.4058.1060.2261.2948.7951.9654.0956.31
φ1/455.7166.3868.2969.7152.2462.4564.9867.23
φ9/1656.5068.3068.4271.9954.4464.7467.9771.28
φ61.7969.7371.9373.6156.5467.3071.9178.38
BCTAsymmetric
φ1/1655.5556.6459.4060.3555.8350.8553.7155.39
φ1/455.8365.7467.1068.5452.6667.0462.0866.59
φ9/1655.6668.0267.4970.9354.4862.3366.3370.14
φ58.6969.4870.04-56.0866.6669.71-
", + "image_path": "774c9c30a758a1d2a80fe73b9fbf057c3e30a2c8fa2c996809d710721eb61475.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 102, + 300, + 491, + 443 + ], + "blocks": [ + { + "bbox": [ + 52, + 279, + 539, + 291 + ], + "lines": [ + { + "bbox": [ + 52, + 279, + 539, + 291 + ], + "spans": [ + { + "bbox": [ + 52, + 279, + 539, + 291 + ], + "type": "text", + "content": "Table 2. Baseline performance comparison on Market1501 (mAP). \"Unified Model\" are models without any compatible regularization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 300, + 491, + 443 + ], + "lines": [ + { + "bbox": [ + 102, + 300, + 491, + 443 + ], + "spans": [ + { + "bbox": [ + 102, + 300, + 491, + 443 + ], + "type": "table", + "html": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
MSMT17
Unified Model28.8211.6948.6522.8657.5630.0670.2243.06
BCT-S46.4920.1156.2228.1358.1130.1458.7630.99
Asymmetric-S32.6413.3449.6224.0457.1830.0466.0840.39
Ours56.2228.1663.2935.3265.1337.7470.3843.89
VERI-776
Unified Model81.0744.7985.3053.6388.6955.4292.2066.50
BCT-S86.2549.3887.0857.0390.4258.2086.7358.46
Asymmetric-S68.7542.1684.2956.5588.0461.3789.8265.35
Ours79.7655.0489.9462.2890.4862.7292.3266.55
", + "image_path": "7fe1f7ded5631968cf2e5afbbee07ed2dc5b90bc4ea92c2509f51cf247632489.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 102, + 467, + 490, + 609 + ], + "blocks": [ + { + "bbox": [ + 168, + 445, + 425, + 456 + ], + "lines": [ + { + "bbox": [ + 168, + 445, + 425, + 456 + ], + "spans": [ + { + "bbox": [ + 168, + 445, + 425, + 456 + ], + "type": "text", + "content": "Table 3. Performance comparison on MSMT17 and VeRi-776 datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 467, + 490, + 609 + ], + "lines": [ + { + "bbox": [ + 102, + 467, + 490, + 609 + ], + "spans": [ + { + "bbox": [ + 102, + 467, + 490, + 609 + ], + "type": "table", + "html": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
Circle Loss
Unified Model70.2550.7972.6660.4378.2164.6687.3874.06
BCT-S63.1842.4572.7452.0477.3859.4987.5374.58
Asymmetric-S69.6648.2177.8557.0280.5261.8186.9073.08
Ours73.9954.6180.5863.4682.3967.0288.9577.05
Softmax + Triplet
Unified Model77.5555.8283.9765.3087.8970.6990.9478.30
BCT-S81.6259.5887.8669.0588.7571.3890.1174.71
Asymmetric-S75.5652.6085.3664.2186.9168.8390.4577.34
Ours87.3666.2190.9476.6491.2177.9392.6481.46
", + "image_path": "62fa9aa18a85316a9a075c33bceb5295a03f2885a771210fb392d684f672ef39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 611, + 447, + 621 + ], + "lines": [ + { + "bbox": [ + 146, + 611, + 447, + 621 + ], + "spans": [ + { + "bbox": [ + 146, + 611, + 447, + 621 + ], + "type": "text", + "content": "Table 4. Performance comparison between different loss functions on Market1501." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "content": "and top-1 accuracy (R1) are adopted to evaluate the performances of retrieval. Define " + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "inline_equation", + "content": "M(\\phi_{a},\\phi_{b})" + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "content": " as the performances of retrieval where the query features are extracted by model " + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "inline_equation", + "content": "\\phi_{a}" + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "content": " and gallery features are extracted by model " + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "inline_equation", + "content": "\\phi_{b}" + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "inline_equation", + "content": "\\phi_{a}" + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "inline_equation", + "content": "\\phi_{b}" + }, + { + "bbox": [ + 47, + 643, + 288, + 704 + ], + "type": "text", + "content": " as different sub-models to evaluate the perfor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 643, + 430, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 430, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 430, + 655 + ], + "type": "text", + "content": "mance of compatible learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": "Implementation details. The implementations of all compared methods and SFSC are based on the FastReLU [24]. The default configuration named \"bag of tricks\" [39] is adopted. " + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 665, + 545, + 715 + ], + "type": "text", + "content": " is set to 0.2 for all experiments." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15948" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 101, + 69, + 491, + 269 + ], + "blocks": [ + { + "bbox": [ + 101, + 69, + 491, + 269 + ], + "lines": [ + { + "bbox": [ + 101, + 69, + 491, + 269 + ], + "spans": [ + { + "bbox": [ + 101, + 69, + 491, + 269 + ], + "type": "table", + "html": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
R1mAPR1mAPR1mAPR1mAP
ResNet-50
Unified Model85.2765.4488.8472.0790.876.4092.2282.13
BCT-S89.3172.0991.0077.2291.3678.2791.8679.40
Asymmetric-S85.0467.3990.2676.2891.6078.5391.9582.79
Ours90.4176.4091.0378.2091.6678.7693.4483.66
MobileNet V2
Unified Model69.0041.1489.5872.8591.5477.9992.2879.50
BCT-S80.4655.1788.4870.6889.8873.7289.9673.87
Asymmetric-S67.2843.6486.9768.9589.3173.8290.7478.45
Ours83.1163.8290.5376.5291.1878.3092.6181.34
ShuffleNet
Unified Model82.5460.1289.8564.4591.9572.9684.3879.13
BCT-S85.5165.7688.4266.8589.7370.5986.6774.41
Asymmetric-S82.3062.5084.4157.4388.4266.4378.9274.92
Ours89.2573.0989.9374.8689.9375.1492.8181.97
", + "image_path": "bb8932007960713cf71bf46c470153f192a06aca1a6c9ddeb944fc02350a30aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 302, + 174, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 174, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 174, + 316 + ], + "type": "text", + "content": "4.2. Baseline Comparisons" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": "We conduct baseline experiments on Market1501. We take \"Softmax + Cross Entropy\" as the loss function, ResNet-18 [23] as the backbone, and " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "W = \\{0.25 \\times, 0.50 \\times, 0.75 \\times\\}" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " as the crop ratio list of the sub-models. Notes that for a model with crop ratio " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\gamma_i" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": ", the ratio of its parameters to the full model is " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\gamma_i^2" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": ". Therefore, the ratio of the number of parameters in the " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "0.25 \\times" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "0.50 \\times" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "0.75 \\times" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " sub-model to that in the full model are " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}, \\frac{1}{4}, \\frac{9}{16}" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 527 + ], + "type": "text", + "content": "Comparison methods. As for one-to-one compatible learning methods BCT and Asymmetric, the compatibility does not outperform that achieved by their corresponding self-compatible methods especially when the models are not constrained to be compatible directly. For self-compatible learning methods, BCT-S and Asymmetric-S, the conflicts among sub-models lead to a limited average accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 528, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 575 + ], + "type": "text", + "content": "Our Method (SFSC). As we resolve the conflicts among different sub-models, our methods can achieve many-to-many compatibility among all the sub-models, which significantly outperforms all comparison methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 585, + 251, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 251, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 251, + 598 + ], + "type": "text", + "content": "4.3. Performances under Different Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "type": "text", + "content": "Different datasets. We evaluate each self-compatible learning method on the other person ReID datasets MSMT17, and the vehicle ReID dataset VeRi-776 under baseline configurations, as shown in Table 3. The results on the two datasets with larger scales are consistent with that on the Market1501. Our methods outperform others significantly for the small capacity sub-model (e.g. the " + }, + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "type": "inline_equation", + "content": "0.25 \\times" + }, + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "type": "text", + "content": " sub-model " + }, + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "type": "inline_equation", + "content": "\\phi_{\\frac{1}{4}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 701 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Different loss functions. We replace the original loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "type": "text", + "content": "function " + }, + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "type": "inline_equation", + "content": "L_{ori}" + }, + { + "bbox": [ + 304, + 303, + 545, + 387 + ], + "type": "text", + "content": " from \"Softmax\" to \"Softmax + Triplet\" and \"Circle Loss\" [52] to evaluate the effectiveness of each method, while Triplet loss is a kind embedding loss and circle loss is a kind of classification loss. The results in Table 4 show that the accuracy of BCT-S and Asymmetric-S drops greatly when training with circle loss, as Our method still stably achieves the best performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "type": "text", + "content": "Different model architectures. We also study whether each method can be stably applied to various model architectures, i.e., ResNet-50 [23], MobileNet V2 [45] and ShuffleNet [67]. ShuffleNet and MobileNet are both designed especially for edge devices, which meet the application scenarios of self-compatible learning. It is shown in Table 5 that BCT-S and Asymmetric-S do not achieve higher accuracy than deploying a unified model with ShuffleNet and MobileNet (e.g. for the " + }, + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "type": "inline_equation", + "content": "0.75 \\times" + }, + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "type": "text", + "content": " sub-model " + }, + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "type": "inline_equation", + "content": "\\phi_{\\frac{9}{16}}" + }, + { + "bbox": [ + 304, + 388, + 546, + 508 + ], + "type": "text", + "content": "). And our methods always improve the accuracy in all cases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 521, + 400, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 400, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 400, + 533 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 540, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 578 + ], + "type": "text", + "content": "We replace the proposed compatible loss and aggregation module with existing compatible loss BCT [50] and summation aggregation to test the proposed two methods." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 306, + 589, + 543, + 636 + ], + "blocks": [ + { + "bbox": [ + 135, + 271, + 457, + 281 + ], + "lines": [ + { + "bbox": [ + 135, + 271, + 457, + 281 + ], + "spans": [ + { + "bbox": [ + 135, + 271, + 457, + 281 + ], + "type": "text", + "content": "Table 5. Performance comparison between different model architectures on Market1501." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 589, + 543, + 636 + ], + "lines": [ + { + "bbox": [ + 306, + 589, + 543, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 543, + 636 + ], + "type": "table", + "html": "
M(φ1/16, φ)M(φ1/4, φ)M(φ9/16, φ)M(φ, φ)
Unified Model55.2567.4871.2580.91
w/o Ours Loss66.4174.5876.8677.87
w/o Ours Aggregation68.0873.7674.8475.38
SFSC69.4376.3777.2681.43
", + "image_path": "9e56a94025003d2aa1243aedc0abd9773b137c4af0573f18da42cde2c7d570d1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 637, + 499, + 647 + ], + "lines": [ + { + "bbox": [ + 351, + 637, + 499, + 647 + ], + "spans": [ + { + "bbox": [ + 351, + 637, + 499, + 647 + ], + "type": "text", + "content": "Table 6. Ablation Study on Market1501." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Table 6 shows that the proposed compatible loss mainly improves the retrieval accuracy of the small capacity sub-models while the proposed aggregation method can ensure the performance of the full model." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15949" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 70, + 293, + 213 + ], + "blocks": [ + { + "bbox": [ + 115, + 70, + 293, + 213 + ], + "lines": [ + { + "bbox": [ + 115, + 70, + 293, + 213 + ], + "spans": [ + { + "bbox": [ + 115, + 70, + 293, + 213 + ], + "type": "image", + "image_path": "a0d939f33c064f2251efe58c9745ee250bf0ecf1f8edcc940cf211ac8366b115.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "text", + "content": "Figure 5. Gradient visualization on Market1501-R18 baseline. As for the original cross-entropy loss (BCT [49]), the magnitude of the " + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "inline_equation", + "content": "0.25 \\times" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "text", + "content": " sub-model is much larger than that of the " + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "inline_equation", + "content": "0.75 \\times" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "text", + "content": " sub-model, especially at the early stage of the training process (" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "inline_equation", + "content": "0 \\sim 1000" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "text", + "content": " iterations). Therefore, the " + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "inline_equation", + "content": "0.25 \\times" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "text", + "content": " sub-model dominates the training progress and the improvements of the " + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "inline_equation", + "content": "0.75 \\times" + }, + { + "bbox": [ + 46, + 222, + 547, + 268 + ], + "type": "text", + "content": " sub-models are eliminated. While our method ensures co-optimization through uncertainty estimation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 70, + 478, + 213 + ], + "blocks": [ + { + "bbox": [ + 300, + 70, + 478, + 213 + ], + "lines": [ + { + "bbox": [ + 300, + 70, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 300, + 70, + 478, + 213 + ], + "type": "image", + "image_path": "0418f3bd8988d888bfa2d350dd9a8267a43fc75606e329ca551bcc3c85dc09e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 278, + 192, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 192, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 192, + 293 + ], + "type": "text", + "content": "4.5. Hyperparameter Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "text", + "content": "Here we verify the impact of hyperparameter settings on the performance of the proposed method. The key hyperparameter is the crop ratio list for sub-models " + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "text", + "content": ". Intuitively, if the number of " + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "text", + "content": " increases or the minimal crop ratio decreases, the compatibility constraints on model " + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "text", + "content": " would be more strict, thereby increasing the difficulty of compatible learning. Therefore, we take the baseline configuration and set " + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 297, + 288, + 392 + ], + "type": "text", + "content": " with different values." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 396, + 288, + 441 + ], + "blocks": [ + { + "bbox": [ + 48, + 396, + 288, + 441 + ], + "lines": [ + { + "bbox": [ + 48, + 396, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 288, + 441 + ], + "type": "table", + "html": "
M(φ1/16, φ)M(φ81/100, φ)M(φ169/400, φ)M(φ283/400, φ)M(φ, φ)
Unified Model55.2562.5469.5472.5780.27
BCT-S61.4568.6373.4272.8075.39
Asymmetric-S62.8769.4575.2073.1474.60
Ours69.6375.8177.1077.3680.59
", + "image_path": "5bcc03e90ae30d84a52a2a8b9de23054d2265897722e2aa15277f45d0dce696a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 442, + 287, + 464 + ], + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 464 + ], + "type": "text", + "content": "Table 7. Performance comparison with " + }, + { + "bbox": [ + 48, + 442, + 287, + 464 + ], + "type": "inline_equation", + "content": "W = \\{0.25\\times ,0.45\\times ,0.65\\times ,0.85\\times \\}" + }, + { + "bbox": [ + 48, + 442, + 287, + 464 + ], + "type": "text", + "content": " on Market1501 (mAP)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 472, + 287, + 518 + ], + "blocks": [ + { + "bbox": [ + 48, + 472, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 518 + ], + "type": "table", + "html": "
M(φ1/100, φ)M(φ1/16, φ)M(φ9/16, φ)M(φ, φ)
Unified Model40.4055.2571.2580.27
BCT-S42.3459.7571.7376.39
Asymmetric-S42.5858.2070.5075.42
Ours45.4468.1878.0081.46
", + "image_path": "539ae55da7a03421439d23783534022ff69753f5c946fb8771684273492d339f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "type": "text", + "content": "Table 8. Performance comparison with " + }, + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "type": "inline_equation", + "content": "W = \\{0.10\\times ,0.25\\times ,0.75\\times \\}" + }, + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "type": "text", + "content": " on Market1501 (mAP)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "content": "Table 7 shows the performances with the number of " + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "content": " increases, while Table 8 shows the performances with the minimum among " + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 544, + 287, + 605 + ], + "type": "text", + "content": " decreases. The result of the comparison methods has fluctuated to varying degrees. SFSC still achieves the best retrieval accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 611, + 132, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 132, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 132, + 623 + ], + "type": "text", + "content": "4.6. Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "We visualize the norm of the gradient during the training progress on Market1501-R18 baseline, as shown in Figure 5. We also visualize the cumulative number of gradient pairs which are conflict with each other during the training process. As shown in Figure 6, our method has found a generic direction to improve all sub-models. Therefore, there are almost no conflict pairs after 5000 iterations. As" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 279, + 524, + 419 + ], + "blocks": [ + { + "bbox": [ + 328, + 279, + 524, + 419 + ], + "lines": [ + { + "bbox": [ + 328, + 279, + 524, + 419 + ], + "spans": [ + { + "bbox": [ + 328, + 279, + 524, + 419 + ], + "type": "image", + "image_path": "9bdcb1363751d8acb4ec965101591a89329bf776554b19f20be4e5544a4311b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 421, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 306, + 421, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 545, + 444 + ], + "type": "text", + "content": "Figure 6. The cumulative number of conflict pairs during the training process." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 449, + 546, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 546, + 485 + ], + "type": "text", + "content": "for the summation aggregation, it faces fierce conflicts especially after 10000 iterations, resulting in the improvement of different sub-models overestimated or underestimated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 495, + 379, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 379, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 379, + 508 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 516, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 516, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 516, + 545, + 622 + ], + "type": "text", + "content": "This paper proposes a Switchable Representation Learning Framework with Self Compatibility (SFSC) for multi-platform model collaboration. SFSC enables us to obtain models with various capacities to fit different computing and storage resource constraints on diverse platforms. It finds a generic direction to improve all sub-models by uncertainty estimation and gradient projection. SFSC achieves state-of-the-art performances and shows convincing robustness in different case studies." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 634, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 416, + 647 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, the Basic and Frontier Research Project of PCL, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15950" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Lu Bai, Yew-Soon Ong, Tiantian He, and Abhishek Gupta. Multi-task gradient descent for multi-task learning. Memetic Computing, 12(4):355-369, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 169 + ], + "type": "text", + "content": "[2] Wentao Bao, Qi Yu, and Yu Kong. Evidential deep learning for open set action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13349-13358, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 204 + ], + "type": "text", + "content": "[3] Emmanuel Bengio, Pierre-Luc Bacon, Joelle Pineau, and Doina Precup. Conditional computation in neural networks for faster models. arXiv preprint arXiv:1511.06297, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "text", + "content": "[5] Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In International Conference on Machine Learning, pages 527-536. PMLR, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "text", + "content": "[6] Mateusz Budnik and Yannis Avrithis. Asymmetric metric learning for knowledge transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2021. 1, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "text", + "content": "[7] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "type": "text", + "content": "[8] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 442 + ], + "type": "text", + "content": "[9] Yair Censor. Pareto optimality in multiobjective problems. Applied Mathematics and Optimization, 4(1):41-59, 1977. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "text", + "content": "[10] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International Conference on Machine Learning, pages 794-803. PMLR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 542 + ], + "type": "text", + "content": "[11] Kyunghyun Cho and Yoshua Bengio. Exponentially increasing the capacity-to-computation ratio for conditional computation in deep learning. arXiv preprint arXiv:1406.7362, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[12] Jifeng Dai, Haozhi Qi, Yuwen Xiong, and Yi Li. Guodong, zhang, han hu, and yichen wei. Deformable convolutional networks. In, ICCV, 2(3):7, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 623 + ], + "type": "text", + "content": "[13] Yunshu Du, Wojciech M Czarnecki, Siddhant M Jayakumar, Mehrdad Farajtabar, Razvan Pascanu, and Balaji Lakshminarayanan. Adapting auxiliary losses using gradient similarity. arXiv preprint arXiv:1812.02224, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[14] Mehrdad Farajtabar, Navig Azizan, Alex Mott, and Ang Li. Orthogonal gradient descent for continual learning. In International Conference on Artificial Intelligence and Statistics, pages 3762-3773. PMLR, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[15] Hang Gao, Xizhou Zhu, Steve Lin, and Jifeng Dai. Deformable kernels: Adapting effective receptive fields for object deformation. arXiv preprint arXiv:1910.02940, 2019. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[16] Jean Gordon and Edward H Shortliffe. The dampster-shafer theory of evidence. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, 3:832-838, 1984. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 182 + ], + "type": "text", + "content": "[17] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12259-12269, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 184, + 546, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 546, + 205 + ], + "type": "text", + "content": "[18] Alex Graves. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 206, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 546, + 249 + ], + "type": "text", + "content": "[19] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 546, + 282 + ], + "type": "text", + "content": "[20] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. arXiv preprint arXiv:2102.02051, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 284, + 546, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 546, + 336 + ], + "type": "text", + "content": "[21] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 546, + 392 + ], + "type": "text", + "content": "[22] Hussein Hazimeh, Natalia Ponomareva, Petros Mol, Zhenyu Tan, and Rahul Mazumder. The tree ensemble layer: Differentiability meets conditional computation. In International Conference on Machine Learning, pages 4138-4148. PMLR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 394, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 546, + 436 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 438, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 546, + 480 + ], + "type": "text", + "content": "[24] Lingxiao He, Xingyu Liao, Wu Liu, Xinchen Liu, Peng Cheng, and Tao Mei. Fastreid: A pytorch toolbox for general instance re-identification. arXiv preprint arXiv:2006.02631, 6(7):8, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 482, + 546, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 546, + 524 + ], + "type": "text", + "content": "[25] Thomas M Hehn, Julian FP Kooij, and Fred A Hamprecht. End-to-end learning of decision trees and forests. International Journal of Computer Vision, 128(4):997-1011, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 526, + 546, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 546, + 570 + ], + "type": "text", + "content": "[26] Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens Van Der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 571, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 546, + 624 + ], + "type": "text", + "content": "[27] Gao Huang, Shichen Liu, Laurens Van der Maaten, and Kilian Q Weinberger. Condensenet: An efficient densenet using learned group convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2761, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 546, + 678 + ], + "type": "text", + "content": "[28] Yani Ioannou, Duncan Robertson, Darko Zikic, Peter Kontschieder, Jamie Shotton, Matthew Brown, and Antonio Criminisi. Decision forests, convolutional networks and the models in-between. arXiv preprint arXiv:1603.01250, 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "text", + "content": "[29] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15951" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[30] Taejong Joo, Uijung Chung, and Min-Gwan Seo. Being bayesian about categorical probability. In International Conference on Machine Learning, pages 4950-4961. PMLR, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[31] Shu Kong and Charless Fowlkes. Pixel-wise attentional gating for scene parsing. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1024-1033. IEEE, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 182 + ], + "type": "text", + "content": "[32] Igor Kononenko. Bayesian neural networks. Biological Cybernetics, 61(5):361-370, 1989. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[33] Harold W Kuhn and Albert W Tucker. Nonlinear programming. In Traces and emergence of nonlinear programming, pages 247-258. Springer, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[34] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3202, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[35] Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. Trgp: Trust region gradient projection for continual learning. arXiv preprint arXiv:2202.02931, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "text", + "content": "[36] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In Proceedings of the IEEE international conference on computer vision, pages 2736-2744, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "type": "text", + "content": "[37] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "text", + "content": "[38] Yihang Lou, Ling-Yu Duan, Shiqi Wang, Ziqian Chen, Yan Bai, Changwen Chen, and Wen Gao. Front-end smart visual sensing and back-end intelligent analysis: A unified infrastructure for economizing the visual system of city brain. IEEE Journal on Selected Areas in Communications, 37(7):1489-1503, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[39] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "text", + "content": "[40] David JC MacKay. Bayesian neural networks and density networks. *Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment*, 354(1):73-80, 1995. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "text", + "content": "[41] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 634 + ], + "type": "text", + "content": "[42] Amir Rosenfeld and Shimon Ullman. Visual concept recognition and localization via iterative introspection. In *Asian Conference on Computer Vision*, pages 264-279. Springer, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[43] Abhijit Guha Roy, Nassir Navab, and Christian Wachinger. Concurrent spatial and channel 'squeeze & excitation' in fully convolutional networks. In International conference on medical image computing and computer-assisted intervention, pages 421-429. Springer, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[44] Manas Sahni, Shreya Varshini, Alind Khare, and Alexey Tumanov. Compofa: Compound once-for-all networks" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "for faster multi-platform deployment. arXiv preprint arXiv:2104.12642, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 149 + ], + "type": "text", + "content": "[45] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "type": "text", + "content": "[46] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 184, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 226 + ], + "type": "text", + "content": "[47] Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in Neural Information Processing Systems, 31, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 249 + ], + "type": "text", + "content": "[48] Yantao Shen, Fanzi Wu, and Ying Shan. Non-inherent feature compatible learning. 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 303 + ], + "type": "text", + "content": "[49] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 1, 4, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "type": "text", + "content": "[50] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6368-6377, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "text", + "content": "[51] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[52] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "text", + "content": "[53] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2820-2828, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 514, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 568 + ], + "type": "text", + "content": "[54] Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pages 2464-2469. IEEE, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "type": "text", + "content": "[55] Andreas Veit and Serge Belongie. Convolutional networks with adaptive inference graphs. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-18, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 658 + ], + "type": "text", + "content": "[56] Thomas Verelst and Tinne Tuytelaars. Dynamic convolutions: Exploiting spatial sparsity for faster inference. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2320-2329, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[57] Xinyi Wang, Ankur Bapna, Melvin Johnson, and Orhan Firat. Gradient-guided loss masking for neural machine translation. arXiv preprint arXiv:2102.13549, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[58] Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15952" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 676 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "convolutional networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 409-424, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[59] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. Advances in Neural Information Processing Systems, 33:2432-2444, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 228 + ], + "type": "text", + "content": "[60] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. Proceedings of the IEEE International Conference on Computer Vision, pages 379–387, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "text", + "content": "[61] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 275, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 319 + ], + "type": "text", + "content": "[62] Shengsen Wu, Liang Chen, Yihang Lou, Yan Bai, Tao Bai, Minghua Deng, and Lingyu Duan. Neighborhood consensus contrastive learning for backward-compatible representation. arXiv preprint arXiv:2108.03372, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 320, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 384 + ], + "type": "text", + "content": "[63] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 386, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 386, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 49, + 386, + 287, + 450 + ], + "type": "text", + "content": "[64] Zhicheng Yan, Hao Zhang, Robinson Piramuthu, Vignesh Jagadeesh, Dennis DeCoste, Wei Di, and Yizhou Yu. Hd-cnn: hierarchical deep convolutional neural networks for large scale visual recognition. In Proceedings of the IEEE international conference on computer vision, pages 2740-2748, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 453, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 453, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 453, + 287, + 485 + ], + "type": "text", + "content": "[65] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 487, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 287, + 529 + ], + "type": "text", + "content": "[66] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 532, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 532, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 49, + 532, + 287, + 585 + ], + "type": "text", + "content": "[67] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 588, + 287, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 588, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 49, + 588, + 287, + 631 + ], + "type": "text", + "content": "[68] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 632, + 287, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 632, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 49, + 632, + 287, + 676 + ], + "type": "text", + "content": "[69] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9308-9316, 2019. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15953" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_content_list.json b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c65e3e28f401b1f3c9fde8952bf0e5de1f780d74 --- /dev/null +++ b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_content_list.json @@ -0,0 +1,1528 @@ +[ + { + "type": "text", + "text": "Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changfeng Ma1, Yinuo Chen1, Pengxiao Guo1, Jie Guo1, Chongjun Wang1, Yanwen Guo1* \n1Nanjing University, Nanjing, China \n{changfengma, yinuochen, pxguo}@smail.nju.edu.cn \n{guojie, chjwang, ywguo}@nju.edu.cn", + "bbox": [ + 127, + 202, + 846, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised completion of real scene objects is of vital importance but still remains extremely challenging in preserving input shapes, predicting accurate results, and adapting to multi-category data. To solve these problems, we propose in this paper an Unsupervised Symmetric Shape-Preserving Autoencoding Network, termed USSPA, to predict complete point clouds of objects from real scenes. One of our main observations is that many natural and man-made objects exhibit significant symmetries. To accommodate this, we devise a symmetry learning module to learn from those objects and to preserve structural symmetries. Starting from an initial coarse predictor, our autoencoder refines the complete shape with a carefully designed upsampling refinement module. Besides the discriminative process on the latent space, the discriminators of our USSPA also take predicted point clouds as direct guidance, enabling more detailed shape prediction. Clearly different from previous methods which train each category separately, our USSPA can be adapted to the training of multi-category data in one pass through a classifier-guided discriminator, with consistent performance on single category. For more accurate evaluation, we contribute to the community a real scene dataset with paired CAD models as ground truth. Extensive experiments and comparisons demonstrate our superiority and generalization and show that our method achieves state-of-the-art performance on unsupervised completion of real scene objects.", + "bbox": [ + 75, + 324, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 736, + 209, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the standard outputs of 3D scanners [12, 32], point clouds are becoming more and more popular [9] which are also the basic data structure in 3D geometry processing [4, 5, 13]. Complete point clouds are hard to obtain due to the nature of the scanning process and object occlusion [35]. Due to the defects of incomplete point clouds on downstream applications such as reconstruction [10], recent works [17, 19, 22, 23, 26, 30, 33, 35] pay more attention", + "bbox": [ + 75, + 753, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af7482530a4e9f27d70c7ce72e1e41ee8a46c6298f17a80089df7905dd3a3d6b.jpg", + "image_caption": [ + "Figure 1. Visual comparison of predicted results on real scene data by our USSPA and other works (top) and our complete result on a whole real scene (bottom). (a) shows an example of a real scene partial point cloud of a chair and the complete predictions by Disp3D [23], ShapeInv [36], Unpaired [32] and our method. As shown, our prediction result is more accurate and uniform according to the input, which contains complete arms and legs. (b) and (c) show the original point cloud of a real scene and the complete results of all the objects in this scene." + ], + "image_footnote": [], + "bbox": [ + 511, + 306, + 879, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to point cloud completion which relies on paired artificial complete point clouds for supervised training to complete partial point clouds. However, these supervised works are difficult to apply in practice because of the great gap between artificial data and real scene data and the inaccessibility to the ground truth of real scene data. Therefore, it is important to complete partial point clouds from real scene in an unsupervised way.", + "bbox": [ + 496, + 761, + 892, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent unsupervised works [24, 32, 36] only require real", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13560", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene partial point clouds and artificial CAD models for unpaired completion utilizing GANs [8] as their fundamental frameworks, most of which need pre-training on artificial data. The main ideas are to transform latent codes from the space of real scene partial data to the space of artificial complete data and then employ the decoder trained on artificial data to predict the complete point cloud. Essentially, these methods make the predictions whose distributions are consistent with the artificial models. Most of them, however, just extract a global feature from the partial input without fully exploiting its geometry information, leading to the prediction severely deviated from the input. And such information actually provides vital clues and constraints for completion. Furthermore, prediction results by these methods usually lack enough geometric details due to the absence of an explicit discriminative process on point clouds. These domain transforming methods are also hard to adapt to multi-category data or other datasets.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present an unsupervised symmetric shape-preserving autoencoding network, termed USSPA, for the completion of real scene objects, as shown in Figure 2, which is a GAN-based end-to-end network without the requirement of pre-trained weights. Different from previous domain transforming methods which cannot fully leverage existing incomplete models, we argue that the existing partial scanning, which also provides vital clues and constraints for the prediction of the missing part, should be preserved to some extent. To this end, we exploit the symmetries shown in many natural or man-made objects and devise a novel symmetry learning module to generate symmetrical point clouds of existing parts by predicting the symmetric planes. This enables our network to preserve the shapes of input symmetrically, intrinsically facilitating structure completion, as shown in Figure 1. For those parts that can not be directly inferred from inputs, we employ an initial coarse module for an initial prediction first. Starting from the initial guess, we specifically design a refinement autoencoder with an upsampling refinement module for detailed refinement and the local feature grouping for extracting local information, to learn detailed structures of artificial data through the autoencoding process. Benefiting from this, our final prediction is accurate, uniform, and symmetric shape-preserving. Besides the indirect guidance of the feature discriminator on latent space, our point discriminator takes predicted point clouds as direct guidance for generating more accurate shapes. Compared with previous methods which train each category separately, our method can classify the partial point clouds simultaneously through a classifier-guided discriminator when adapted to multi-category data, with consistent performance on the single category.", + "bbox": [ + 75, + 367, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To measure the performance of unsupervised completion quantitatively, we build a dataset from ScanNet [5] and", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ShapeNet [2] utilizing the annotations of Scan2CAD [1]. Our dataset contains real scene partial point clouds and paired ground truths that are only used for evaluation in our experiments. Extensive comparisons against previous works on this dataset and the public PCN Dataset [35] show the superiority and generalization of our method which achieves state-of-the-art performance on unsupervised completion of real scene objects.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows.", + "bbox": [ + 517, + 214, + 772, + 227 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel USSPA for unsupervised real scene point cloud completion whose prediction is accurate, uniform and symmetric shape-preserving. Clearly different from previous works training each category separately, our USSPA can be adapted to the training of multi-category data in one pass by classifying the input simultaneously.", + "- We propose a novel symmetry learning module and a novel refinement autoencoder. The symmetry learning module preserves input shapes by generating symmetrical point clouds, and the refinement autoencoder learns the detailed information from artificial data to refine the initial guess by an autoencoding process.", + "- We propose a new evaluation method for obtaining paired ground truths and partial data from artificial and real scene datasets using alignment information, which can be used to more accurately evaluate unsupervised completion of real scene objects." + ], + "bbox": [ + 517, + 234, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 527, + 640, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Supervised Point Cloud Completion", + "text_level": 1, + "bbox": [ + 500, + 550, + 812, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To infer the complete point cloud from an incomplete input, many deep learning methods were introduced in recent years, most of which are supervised. At first, the 3D data were voxelized into occupancy grids or distance fields before fed into convolutional networks [15, 16, 20]. To avoid loss of details and huge memory cost brought by these data representations, Yuan et al. [35] propose PCN which directly operates on raw point clouds. Giancola et al. [7] also leverage a new method for 3D shape completion which directly works on LiDAR point clouds. Later works [11, 19, 22] pay much attention to details refinement and denoise. Disp3D [23] investigates grouping local features to bring improvement to point cloud completion. Taking structures and topology information into consideration, SA-Net [25], GRNet [31], PoinTr [34] and LAKe-Net [17] succeed to utilize the detailed geometry information from input to generate more reasonable complete point clouds. Other works like designing new metrics [28] and combining information of single-view images [37] also make contributions to the point cloud completion task. However, these works are all supervised that are not suitable for practice since ground truths are in lack.", + "bbox": [ + 496, + 571, + 893, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2fe9346696598d15132b477864fb0ce92de4d77bc56ce7b3fd53267d9f81b370.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our USSPA. Our network has a generator composed of an initial coarse predictor (blue) and a refinement autoencoder (green), and the point and feature discriminators (yellow). The red lines and boxes represent the employed losses. The blue and green lines (bottom) denote the forward processes having the same network architectures with sharing weights." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 537, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29ab211dd5f971895a3b505959b9a7e101daaec8c276d433151726698b027b70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 88, + 870, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Unsupervised Point Cloud Completion", + "text_level": 1, + "bbox": [ + 76, + 395, + 410, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the absence of suitable paired training data for supervised models, some unsupervised models for point cloud completion have emerged in recent years. Unpaired [32] first introduces the GAN to learn the transformation from the latent space of partial point cloud to the latent space of complete point cloud, and then decode the transformed latent code to predict. Zhang et al. [36] employ optimization-based GAN inversion [29] and directly find the complete prediction by optimizing the latent code such that the degradation of the complete point cloud matches the input partial point cloud, which requires fine-tuning the network for each input data. Cycle4 [24] proposes two simultaneous cycle transformations between the latent spaces of complete point cloud and partial ones, facilitating the network's understanding of 3D shapes by building a bidirectional mapping. Wu et al. [27] propose the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. However, the application of pre-trained GANs makes their predictions lack details and difficult to preserve the shape features of the input. For evaluation, Unpaired [32] employs classification accuracy of the complete results which is hard to measure detailed structures. Thus, a more accurate evaluation method is required.", + "bbox": [ + 75, + 414, + 473, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 782, + 169, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 805, + 282, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the partial point cloud $\\mathcal{P}_R\\in \\mathbb{M}_{n_0\\times 3}$ from real scene data with distribution $p_r$ and the complete point cloud $\\mathcal{P}_A\\in \\mathbb{M}_{n_0\\times 3}$ from artificial data with distribution $p_a$ , the goal of unsupervised real scene point cloud completion is to predict the complete point cloud $Com(\\mathcal{P}_R)$ of $\\mathcal{P}_R$ such", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that $\\operatorname{Com}(\\mathcal{P}_R) \\sim p_a$ and $\\mathcal{P}_R \\subset \\operatorname{Com}(\\mathcal{P}_R)$ . This means that the prediction must be similar to the artificial data and meanwhile retain the shapes of the input partial point cloud.", + "bbox": [ + 496, + 395, + 890, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We carefully design several modules of our generator $\\mathcal{G}$ and discriminators, $\\mathcal{D}_p$ and $\\mathcal{D}_f$ , for unsupervised completion on real scene data through generative adversarial learning [8] as shown in Figures 2 and 3. As we mentioned above, many natural and man-made objects exhibit significant symmetries. To fully leverage these symmetries and preserve the shape of existing partial input $\\mathcal{P}_R$ , we devise a novel symmetry learning module that predicts the symmetric plane and the symmetrical point cloud of $\\mathcal{P}_R$ for further refinement. Our refinement autoencoder learns the detailed information of artificial data and refines the details of the initial, but coarse prediction $P_{R_0}$ for uniform and accurate prediction. With the local feature grouping operation, our elaborate upsampling refinement module can utilize the local features for more detailed results. Different from previous works focusing on the discriminative process of latent space, our point discriminator $\\mathcal{D}_p$ also takes point clouds as input for direct guidance.", + "bbox": [ + 496, + 443, + 892, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initial Coarse Predictor. We first employ an initial coarse predictor for shape-preserving and preliminary prediction. The encoder $E_{R}$ takes the real scene partial point cloud $\\mathcal{P}_R$ as input and encodes the global feature $f_{R}^{\\prime} = E_{R}(\\mathcal{P}_{R})\\in \\mathbb{M}_{1\\times c_{0}}$ . $f_{R}^{\\prime}$ is then taken by the decoder $D_{R}$ for the prediction of initial, but coarse complete result $P_R' = D_R(f_R')\\in \\mathbb{M}_{n_1\\times 3}$ . Encoder-decoder architecture is hard to retain shape features on the completion task since there is no explicit complete ground truth. To this end, we devise the symmetry learning module to accommodate symmetrical shape features widely existed in man-made and natural objects, as shown in Figure 3. The decoder $D_{sl}$ pre", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13562", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dicts symmetric arguments $\\mathbf{A} = D_{sl}(f_R') = [x_A, y_A, z_A]^T$ and then we introduce symmetry operation to calculate the symmetrical point cloud $P_{sym}$ of $\\mathcal{P}_R$ by:", + "bbox": [ + 75, + 89, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s y m} = \\boldsymbol {p} - 2 \\frac {\\boldsymbol {A} \\cdot \\boldsymbol {p}}{\\| \\boldsymbol {A} \\| ^ {2}} \\boldsymbol {A}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 147, + 468, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{p} = [x,y,z]^T\\in \\mathcal{P}_R$ $\\cdot$ represents the inner product and $||\\pmb {x}||$ denotes the module of the vector $\\pmb{x}$ . Without the loss of generality, we can safely assume that all the objects are axis-aligned. Under such an assumption, the symmetric plane is perpendicular to the $xz$ -plane and zero-crossing. Thus we set $y_{A} = 0$ . Then $P_{sym}$ $\\mathcal{P}_R$ and $P_R^{\\prime}$ are fused and resampled through FPS (farthest point sampling) to get $P_{R_0}$ which is the complete and shape-preserving result.", + "bbox": [ + 75, + 189, + 468, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Refinement Autoencoder. Noisy $\\mathcal{P}_R$ and the fusion operation may cause noisy $P_{R_0}$ with uneven point distribution. The initial coarse predictor or the point discriminator only predicts or judges through the global feature, which makes $P_{R_0}$ lack details. Thus, $P_{R_0}$ requires denoising and further refinement on details and point distribution. To this end, we employ the refinement autoencoder with a carefully designed upsampling refinement module. As shown in Figure 2, the blue and green lines of the refinement autoencoder represent the same forward process having the same network architectures with sharing weights. In this way, our autoencoder extracts the detailed information of artificial data and refines the details of $P_{R_0}$ for more accurate and uniform prediction.", + "bbox": [ + 75, + 311, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The encoder and decoder $E_{A},D_{A}$ generate the skeleton point cloud $P_{R_1 / A_1}\\in \\mathbb{M}_{n_2\\times 3}$ of input point cloud. We set $n_2$ to a small number for easier prediction of $P_{R_1 / A_1}$ . And the density loss is faster to calculate with fewer points. We then employ upsampling refinement module to upsample $P_{R_1 / A_1}$ and add details to the skeleton for final prediction $P_{R_2 / A_2}$ . As shown in Figure 3, we utilize wight-shared MLP [3] to obtain the point features and max-pooling operation on point features to obtain the global feature of the point cloud. We also concatenate the point and global features to fuse the information. Since wight-shared MLP only extracts the feature of one point, lacking the local information, we also employ the local feature grouping to get the local information around each point. For point $p_0$ , $f_{group}\\in \\mathbb{M}_{n_2\\times (c_1\\cdot m)}$ are the features grouped from $f_{input}\\in \\mathbb{M}_{n_2\\times c_1}$ corresponding to the points $p$ that $\\| p - p_0\\| \\leq r$ , where $m$ is the feature number of each group. After applying max-pooling operation on $m$ features of each group, $f_{local}\\in \\mathbb{M}_{n_2\\times c_1}$ are concatenated with $f_{input}$ to generate the output $f_{output}\\in \\mathbb{M}_{n_2\\times 2c_1}$ . We utilize a weight-shared MLP for upsampling that takes fused features composed of $P_{R_1 / A_1}$ , local and global features, and predict $k$ shifts for each point. Then, each point splits into $k$ points by adding the predicted shifts to its coordinate whose result is $P_{R_2 / A_2}\\in \\mathbb{M}_{kn_2\\times 3}$ .", + "bbox": [ + 75, + 523, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6df62d8a68bf4a85c11c6c19a6e2f60514742af0ef23556f6afefaee378b946d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 90, + 879, + 175 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/90846916c9692d15b10200c537e9d1d31275caf446df0ff45d2474a0723554ac.jpg", + "image_caption": [ + "Upsampling Refinement Module", + "Local Feature Grouping" + ], + "image_footnote": [], + "bbox": [ + 516, + 191, + 879, + 289 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cae283e51f0ca5bc5d18d4521552a9c672ce864e735014bf914f3a886644a2b.jpg", + "image_caption": [ + "Symmetry Learning Module" + ], + "image_footnote": [], + "bbox": [ + 514, + 305, + 756, + 404 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efe0b02cc3fd5d45a1089578c90f3d8cff7a594ce3c219542db67ecf9e35bacf.jpg", + "image_caption": [ + "Concatenate" + ], + "image_footnote": [], + "bbox": [ + 766, + 310, + 792, + 330 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/923708c274ca2ebb767d9b0811e99451a6afa3dda58de7e1f890b7a4d3481246.jpg", + "image_caption": [ + "Max-Pooling" + ], + "image_footnote": [], + "bbox": [ + 767, + 335, + 792, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a938c514f554c0d23359bd5d2c0ad9391f993c3e4b86512f87f6675443a67458.jpg", + "image_caption": [ + "Local Feature Grouping" + ], + "image_footnote": [], + "bbox": [ + 767, + 359, + 792, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0877c0040db32685acef03a3bef9f5f2338a32a16dc7d1875b1e391dd32c479.jpg", + "image_caption": [ + "Element Addition", + "Figure 3. Detailed structures of our symmetry learning module (red), upsampling refinement module (yellow), and local feature sampling (blue)." + ], + "image_footnote": [], + "bbox": [ + 767, + 392, + 792, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point and Feature Discriminator. We employ the point discriminator $\\mathcal{D}_p$ to directly guide the generation of $P_{R_0}$ and $P_{R_2}$ comparing with the artificial complete data $\\mathcal{P}_A$ for unsupervised completion. We also guide the generation of latent code of autoencoder through the feature discriminator $\\mathcal{D}_f$ such that $f_R$ and $f_A$ have the same distribution for more accurate prediction. Our discriminators predict the probability that input point cloud $P_i$ or feature $f_i$ is real data. For multi-category data, we replace the probability-guided discriminator with a classifier-guided discriminator which outputs the category label of input rather than probability. We utilize label 0 to represent fake data and labels $l > 0$ to denote real data of different categories such as the chair, table, etc. For fake data, the discriminator tends to output label 0, while the generator tends to make the discriminator output label $l > 0$ . And for real data with label $l$ , the discriminator tends to output $l$ correctly. In this way, our USSPA can unsupervisedly classify the real scene partial point cloud when completing multi-category data.", + "bbox": [ + 498, + 465, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Optimization", + "text_level": 1, + "bbox": [ + 500, + 757, + 638, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The abbreviate optimization goal of our USSPA for unsupervised learning is:", + "bbox": [ + 500, + 775, + 888, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G}} \\max _ {\\mathcal {D} _ {p}, \\mathcal {D} _ {f}} \\mathbb {E} _ {P _ {i} \\sim p _ {a}} \\log p + \\mathbb {E} _ {P _ {i} \\sim p _ {r}} \\log (1 - p), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 819, + 888, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ is the probability of real data. By playing the maxmin game with $\\mathcal{D}_p$ and $\\mathcal{D}_f$ , our generator $\\mathcal{G}$ tends to predict point clouds similar to artificial data. We employ", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8446cea51927e05b3827bb2793b11c453971f6f7b1faf68ac74ed3f865d2a945.jpg", + "table_caption": [ + "Table 1. Point cloud completion comparison on our dataset in terms of L1 Chamfer Distance $cd^{l1} \\times 10^2$ (lower is better). All the methods are trained with our dataset and the unsupervised methods including ours are trained on single category data. Boldface denotes the best among unsupervised methods in Tables 1, 2 and 3." + ], + "table_footnote": [], + "table_body": "
MethodAVGchairtabletrash binTVcabinetbookshelfsofalampbedtub
Sup.PoinTr [34]14.3713.6512.5215.2612.6917.3213.9912.3617.0515.1313.77
Disp3D [23]7.786.248.207.127.1210.366.945.6014.036.905.32
TopNet [19]7.076.395.797.406.268.377.025.948.507.817.25
Unsup. OptBasedShapeInv [36]21.3917.9717.2833.5115.6926.2625.5114.2816.6932.3314.43
Unsup.Unpaired [32]10.478.417.5212.086.7217.459.956.9219.3610.046.22
Cycle4 [24]11.539.1111.3511.938.4015.4712.5110.6312.2515.737.92
Ours8.568.227.6810.367.6610.777.846.1411.938.206.75
", + "bbox": [ + 86, + 132, + 883, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Chamfer Distance(CD) [6] to measure the difference between point clouds. The $cd_{P_1 \\to P_2}^l$ indicates the single direction Chamfer Distance with L2-norm from point cloud $P_1$ to $P_2$ :", + "bbox": [ + 76, + 281, + 468, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc d _ {P _ {1} \\rightarrow P _ {2}} ^ {l 2} = \\frac {1}{| P _ {1} |} \\sum_ {\\boldsymbol {x} \\in P _ {1}} \\min _ {\\boldsymbol {y} \\in P _ {2}} \\| \\boldsymbol {x} - \\boldsymbol {y} \\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 348, + 468, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which guides $P_{1}$ to be a part of $P_{2}$ . The $cd_{P_1\\leftrightarrow P_2}^{l^2}$ between $P_{1}$ and $P_{2}$ is:", + "bbox": [ + 76, + 393, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc d _ {P _ {1} \\leftrightarrow P _ {2}} ^ {l 2} = c d _ {P _ {1} \\rightarrow P _ {2}} ^ {l 2} + c d _ {P _ {2} \\rightarrow P _ {1}} ^ {l 2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 431, + 468, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which guides $P_{1}$ and $P_{2}$ to be identical. As shown in Figure 2, to restrict the generation of $P_{R}^{\\prime}$ , we employ single direction Chamfer Distance from $\\mathcal{P}_R$ to $P_{R}^{\\prime}$ . We also utilize Chamfer Distance between $P_{R_1 / A_1}, P_{R_2 / A_2}$ and $P_{R_0}, \\mathcal{P}_A$ for supervision of autoencoder. For uniform prediction, we employ density loss by first calculating the local density of $\\boldsymbol{x} \\in P$ :", + "bbox": [ + 76, + 458, + 472, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl d (\\boldsymbol {x}, P) = \\frac {1}{k _ {d}} \\sum_ {k = 1} ^ {k _ {d}} \\min _ {\\boldsymbol {y} \\in P} ^ {k} | | \\boldsymbol {x} - \\boldsymbol {y} | |, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 571, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and then calculating the standard deviation for all $x$ :", + "bbox": [ + 76, + 619, + 423, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd l (P) = \\sqrt {\\frac {1}{| P |} \\sum_ {\\boldsymbol {x} \\in P} \\left(l d (\\boldsymbol {x} , P) - \\overline {{l d (\\boldsymbol {x} , P)}}\\right) ^ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 643, + 468, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\min$ indicates the $k$ -th-min element. The smaller the $ld(\\pmb{x}, P)$ is, the bigger the local density of $\\pmb{x}$ is.", + "bbox": [ + 76, + 693, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we alternately optimize $\\mathcal{G}$ with loss $\\mathcal{L}_{\\mathcal{G}}$ :", + "bbox": [ + 96, + 731, + 437, + 746 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} = \\alpha_ {1} \\mathcal {L} _ {F \\rightarrow R} + \\alpha_ {2} \\mathcal {L} _ {P} + \\alpha_ {3} \\mathcal {L} _ {d}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 756, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 782, + 122, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {F \\rightarrow R} = - \\left(\\log \\mathcal {D} _ {p} \\left(P _ {R _ {0}}\\right) + \\log \\mathcal {D} _ {p} \\left(P _ {R _ {2}}\\right)\\right. \\tag {8} \\\\ + \\log \\mathcal {D} _ {f} (f _ {R})) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 803, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {P} = c d _ {P _ {R _ {1}} \\leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {P _ {R _ {2}} \\leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {\\mathcal {P} _ {R} \\rightarrow P _ {R} ^ {\\prime}} ^ {l 2} \\tag {9} \\\\ + c d _ {P _ {A _ {1}} \\leftrightarrow \\mathcal {P} _ {A}} ^ {l ^ {2}} + c d _ {P _ {A _ {2}} \\leftrightarrow \\mathcal {P} _ {A}} ^ {l ^ {2}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 859, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = d l \\left(P _ {R _ {1}}\\right) + d l \\left(P _ {A _ {1}}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 310, + 890, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mathcal{D}$ with loss $\\mathcal{L}_{\\mathcal{D}}$", + "bbox": [ + 500, + 340, + 640, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {D}} = \\alpha_ {4} \\mathcal {L} _ {F} + \\alpha_ {5} \\mathcal {L} _ {R}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 369, + 890, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 398, + 545, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {F} = - \\left(\\log \\left[ 1 - \\mathcal {D} _ {p} \\left(P _ {R _ {0}}\\right) \\right] + \\log \\left[ 1 - \\mathcal {D} _ {p} \\left(P _ {R _ {2}}\\right) \\right] \\right. \\tag {12} \\\\ \\left. + \\log \\left[ 1 - \\mathcal {D} _ {f} \\left(f _ {R}\\right) \\right]\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 421, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = - \\left(\\log \\mathcal {D} _ {p} \\left(\\mathcal {P} _ {A}\\right) + \\log \\mathcal {D} _ {f} \\left(f _ {A}\\right)\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 489, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and here $\\alpha_{1\\sim 5}$ are the weights used for balancing the influences between each term. And for multi-category data, we use cross-entropy to calculate the loss between predicted labels and ground truth labels.", + "bbox": [ + 496, + 521, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 589, + 633, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 609, + 812, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from the Unpaired [32] which utilizes a classifier to evaluate the completion results of real scene data, we build a new dataset from ScanNet [5], ShapeNet [2] and Scan2CAD [1], containing paired partial and complete point clouds for more accurate evaluation. Scan2CAD contains annotations that align the CAD models of ShapeNet to the real scenes of ScanNet. We first extract the CAD models from ShapeNet and partial point clouds from ScanNet which are paired according to the correspondences offered by Scan2CAD, and then uniformly sample the CAD models to get complete point clouds. Finally, we transform the paired partial and complete point clouds with the transform matrix given by Scan2CAD to normalize their face direction, scale, and position. We select ten categories including the chair, table, etc., to establish the dataset and randomly split the data into training, validation, and testing sets. In our experiments, these ground truths are only used for the evaluation of the complete results.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13564", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50ebc953567d54ff4c32e2be7a363499d0e024e27777648ddb9aa67791a13ff5.jpg", + "image_caption": [ + "Figure 4. Visualization of completion results by supervised works, including TopNet [19], Disp3D [23], unsupervised works Cycle4 [24], Unpaired [32], optimization-based work ShapeInv [36] and our method. From top to bottom: table, bed, cabinet, and TV." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 869, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize PyTorch to implement our USSPA. In Section 3.1, the point numbers $n_{0\\sim 2}$ are set to 2048, 512 and 512. The channel numbers $c_{0\\sim 1}$ are set to 512 and 256. The feature number $m$ , the radius $r$ of local feature grouping, and the split number of upsampling $k$ are set to 32, 0.5, and 4. $k_d$ of density loss in Section 3.2 is set to 16. The weights $\\alpha_{1\\sim 5}$ of losses are set to 1, 100, 7.5, 0.5 and 0.5. We employ Adam optimizer whose learning rate is $1.0 \\times 10^{-5}$ . The batch size is set to 4, and the maximum epoch of training is 240 to 960 according to the size of the dataset. Training our USSPA takes about 20 hours with a GTX 2080Ti GPU. We will make our dataset and codes public in the future.", + "bbox": [ + 75, + 443, + 472, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison", + "text_level": 1, + "bbox": [ + 76, + 636, + 209, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our USSPA against existing representative point cloud completion methods including supervised [19, 23, 34], unsupervised [24, 32] and optimization-based [36] methods on our dataset and the PCN Dataset [35]. The metrics used for comparison include pre-point L1 Chamfer Distance $cd^{l1}$ [6] and F-Scores $F_{score}^{1\\%}$ , $F_{score}^{0.1\\%}$ [18].", + "bbox": [ + 75, + 656, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, we compare these methods on our dataset with 10 categories where all methods are trained on the training set and evaluated on the testing set. All the unsupervised methods are trained with single category data for each class. The average $cd^{l1}$ of our USSPA is $18.2\\%$ lower than Unpaired and the average $F_{score}^{0.1\\%}$ is $30\\%$ higher than Unpaired as shown in Table 2, indicating the superior performance of our method on completion of real scene objects. We also mix up all 10 categories as the multi-category data to train the methods for evaluation of the generalization", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison on our dataset trained with single category and multi-category data in terms of L1 Chamfer Distance $cd^{l1} \\times 10^2$ (lower is better) and F-scores $F_{score}^{0.1\\%} \\times 10^2$ , $F_{score}^{1\\%} \\times 10^2$ (higher is better). \"Ours\" and \"Ours(classifier)\" denote our method with probability-guided and classifier-guided discriminators, respectively.", + "bbox": [ + 498, + 441, + 892, + 525 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/41260711794fb64449907b46a49018ba3574405f354d2d9a2dd7253fb1153f31.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsingle categorymulti-category
↑F0.1% score↑F1% score↓cdl1↑F0.1% score↑F1% score
PoinTr [34]--14.3718.3580.41
Disp3D [23]--7.7830.2978.26
TopNet [19]--7.0712.3380.37
ShapeInv [36]15.5866.5319.3516.9869.66
Unpaired [32]12.2064.3310.1210.8666.68
Cycle4 [24]9.9860.1412.008.6156.57
Ours17.4973.418.9616.8872.31
Ours(classifier)--8.7617.1273.75
", + "bbox": [ + 501, + 529, + 887, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ability. As shown in Table 2, our USSPA keeps the same performance by replacing the probability-guided discriminators with classifier-guided discriminators and has better generalization ability on multi-category than other unsupervised networks. Further comparisons show the generalization capability of our USSPA on different datasets. We train unsupervised models with ScanNet and then test them on the PCN Dataset which is a point cloud completion dataset built upon artificial data. As shown in Table 3, our method also performs better than other unsupervised methods, the average $cd^{l1}$ of which is $15\\%$ lower than the SOTA method. Table 4 shows the classification accuracy of PN2 [14] and", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5db3ff3690db2481351a8d59a5efa06691b717af27790384ae178a0bb546337e.jpg", + "table_caption": [ + "Table 3. Point cloud completion comparison on PCN Dataset in terms of L1 Chamfer Distance $cd^{l1} \\times 10^2$ (lower is better). All unsupervised models are trained with ScanNet." + ], + "table_footnote": [], + "table_body": "
MethodAVGchairtablecabinetsofalamp
PoinTr [34]5.495.615.686.085.674.44
Disp3D [23]2.512.422.302.382.443.00
TopNet [19]5.926.345.456.065.805.95
ShapeInv [36]19.0523.1815.6617.1422.8516.40
Unpaired [32]14.8712.878.1414.3018.2320.82
Cycle4 [24]17.6014.2515.7321.0621.5415.40
Ours12.6313.529.668.8915.5115.57
", + "bbox": [ + 83, + 142, + 464, + 273 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d83dcb74061f492c34a807dfbcda873a5a344c9148792f5bb07e244fcefd07c.jpg", + "table_caption": [ + "Table 4. Point cloud classification in terms of accuracy. \"AVG\" denotes the average accuracy on 10 categories of our dataset. Only four categories are shown due to width limitation." + ], + "table_footnote": [], + "table_body": "
MethodAVGchairtabletrash binsofa
PN2 [14]58.6%91.0%94.0%17.2%48.9%
Ours69.8%91.0%91.0%51.6%93.3%
", + "bbox": [ + 86, + 342, + 457, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method. Our method can also unsupervisedly classify the real scene partial point cloud while completing. The accuracy of our USSPA is $11.2\\%$ higher than PN2 on average.", + "bbox": [ + 75, + 409, + 468, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows the qualitative results on 4 categories. The complete point clouds predicted by our USSPA exhibit more accurate shapes, such as the leg of the table and the division plate of the bookshelf, which are hard to generate even for supervised methods. Our method can successfully generate the circular top of the table. By comparison, other unsupervised works generate a square one. Our predictions are also more uniform with fewer noises. Benefiting from our shape-preserving method, our predicted TV retains the support part from the input point cloud. Our symmetry learning module also plays an important role in completing the pillars of the bed by copying the pillars from the input symmetrically. Visualization of our predicted point clouds shows the superiority of our methods in the generation of detailed completion on real scene data.", + "bbox": [ + 75, + 455, + 468, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 684, + 230, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on network architecture. We conduct experiments on ablated models for evaluation of the necessity of important structures of our USSPA, including the symmetry learning module, the refinement autoencoder, the local feature grouping, the point and feature discriminators, the single direction Chamfer Distance, and the upsampling refinement module. As shown in Table 5, these modules are removed from our full network separately. We downsample the final prediction of our USSPA to 512 points for a fair comparison to the ablated model without the upsampling refinement module. We train and utilize $F_{score}^{0.1\\%}$ to evaluate these ablated models with our dataset on the chair category, and the comparison indicates the necessity of these modules", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/84a266a8b271bb3de1ada255cea998db0e46a0b091648c78fd67cf4a609ce1e5.jpg", + "table_caption": [ + "Table 5. Ablation study in terms of $F_{score}^{0.1\\%}$ , where “-” represents removing a module from the full network respectively and “512” denotes the point number." + ], + "table_footnote": [], + "table_body": "
ModelF0.1% score
- Symmetry Learning Module16.43
- Refinement Autoencoder17.08
- Local Feature Grouping15.72
- Feature Discriminator17.11
- Point Discriminator15.68
- Single Direction Chamfer Distance17.84
Ours (full)18.43
- Upsampling Refinement (512)5.60
Ours (full) (512)6.46
", + "bbox": [ + 529, + 142, + 862, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ad04e4de23b29a7956529e38d60b2ba392bfe4dfde3913da6fab02cdf01b98b.jpg", + "image_caption": [ + "Figure 5. Visualization of the symmetric planes and predictions by USSPA without the symmetry learning module and full network. Blue and orange denote original and mirrored points, separately." + ], + "image_footnote": [], + "bbox": [ + 501, + 318, + 890, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which play significant roles in shape-preserving, extracting, and utilizing detailed features, upsampling and refining for more accurate point cloud completion.", + "bbox": [ + 498, + 474, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Symmetry learning module. Our symmetry learning module learns different symmetric planes according to the input point clouds as shown in Figures 5(a) and 5(b). The predicted symmetric plane is xy-plane for the chair and yz-plane for the table. Through different symmetric planes, our network can fully leverage the symmetrical structures of input point clouds, such as the arm of the chair and the legs of the table, which exist on only one side. Figure 5(c) shows the prediction $P_{R_2}$ of our network without the symmetry learning module, and Figure 5(d) shows the result of our full network. Figure 5(e) shows the input point cloud whose left arm is missing while the right arm is complete. Without the symmetry learning module, the left arm of the prediction is incomplete as shown in the red box. And benefiting from such a module, our full network can learn the symmetric structures, and generate a complete left arm.", + "bbox": [ + 496, + 522, + 892, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Density loss. We visualize the local density of points as shown in Figure 6. The warmer the color of the point is, the bigger its local density (described in equation 5) is. (a) is the prediction without the supervision of the density loss, and (b) is the prediction supervised by the density loss. Without the supervision of density loss, the network prefers generating many points in the same location as shown in the blue circle to minimize the Chamfer Distance, which makes the predicted point clouds contain fewer points describing", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13566", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60e4cb56b0df4722a39c54d55d9d235c74ce3002e273332eabea170d5a73a084.jpg", + "table_caption": [ + "Table 6. Comparison of Unpaired and our USSPA in terms of density loss, where \"ours (w/o $dl$ )\" indicates our USSPA without the supervision of the density loss. The results show that our predicted point clouds are more uniform." + ], + "table_footnote": [], + "table_body": "
Unpairedours (w/o dl)ours
Density Loss0.09570.12380.0149
", + "bbox": [ + 106, + 152, + 437, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc48897284b2cd60fdb4dea05a44266c4b191d525270a63d1f40eec626af6d9f.jpg", + "image_caption": [ + "Figure 6. Visualization of the local density of point clouds, where the warmth of the color indicates the values of local densities. The warmer the color is, the bigger the local density is. (a) and (b) are the comparison of the ablation experiment on the density loss. (c) and (d) are the examples of $P_{R_0}$ and $P_{R_2}$ . The density loss and the point number of each point cloud are shown at its bottom." + ], + "image_footnote": [], + "bbox": [ + 80, + 203, + 173, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71ec2aea586c7429d4c70b60810b95212742fa774dc044fcba4a7fe544000308.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 204, + 271, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/014284f2971eb8e5dc8dae2871dbfc862107835bfd3946161fe642da7ac99679.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 209, + 372, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7134bc6d3af70df8f1d07f6cc0dc1dd3db7d449322a6644677d1f34c5d9721e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 207, + 470, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "details such as chair arms and legs. Table 6 compares Unpaired and our USSPA on density loss indicating that our predicted point clouds are more uniform.", + "bbox": [ + 75, + 426, + 468, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Autoencoding refiner. As shown in Figure 6, (c) and (d) are the examples of $P_{R_0}$ and $P_{R_2}$ which is the refinement result of the autoencoding refiner on $P_{R_0}$ . The refiner removes the noises from input shown in blue boxes and generates more points on legs as shown in green boxes, which makes the whole point cloud more uniform and accurate with less density loss.", + "bbox": [ + 75, + 474, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real scene data. We also test our USSPA on real objects, as shown in Figure 7. We obtain the point clouds of real objects by the 3D scanner and then sample them to 2048 points as the input to our network. Our network can preserve the right arm of the chair and complete the left arm. The predictions are more accurate and uniform as shown in red boxes. This indicates the generalization ability of our USSPA on real objects.", + "bbox": [ + 75, + 580, + 468, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of latent codes on multi-category data. We utilize t-SNE [21] to visualize the latent codes $f_{R}$ and $f_{A}$ of Unpaired and our USSPA on multi-category data. As shown in Figure 8, ten different colors indicate ten categories where dark colors indicate real scene data and light colors indicate artificial data. Figure 8 (c) and (d) show the t-SNE results of latent codes on chair category only. Benefiting from our feature discriminator and autoencoder, the latent codes of real scene data are similar to artificial data which can be shown from the distribution of dark and light points. The dark and light points of Unpaired are separated while ours are fixed together for each category. Benefiting from our classifier-guided discriminator, the distributions of", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cf0cecbbab6595a26b95c0353320bfd95fadf19f47261ad6018f23eb772c395.jpg", + "image_caption": [ + "Figure 7. Completion on real objects by Unpaired [32], ShapeInv [36] and our USSPA." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34b72b68879acc150a93e2ced5a68c5095a513b1861ec8a7c726c5ea4bd38f8c.jpg", + "image_caption": [ + "Figure 8. Visualization of t-SNE results of the latent codes by Unpaired [32] and our method. (a) and (b) are the results on multi-category, where different colors denote different categories. Dark and light denote real scene data and artificial data respectively. (c) and (d) are the results on the chair category only." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 694, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48e7821b250ad1cd936a453b8c8bc9080ae71a957b7c5e1d1387e8666091f951.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 290, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our latent codes of different categories are separated and clean. For example, as Figure 8 (b) shows, the distances between pink, orange and blue clusters are bigger and the clusters are tighter.", + "bbox": [ + 498, + 513, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Limitation and Future Work", + "text_level": 1, + "bbox": [ + 500, + 578, + 867, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented USSPA, an end-to-end unsupervised network for completion of real scene point cloud objects. Benefiting from our carefully designed symmetry learning module and refinement autoencoder, our prediction preserves the symmetrical shape of input with more accurate and uniform points. We also contribute a real scene dataset for accurate evaluation, and extensive experiments and comparisons show the superiority and generalization of our method on different categories of different datasets and real objects, which achieves state-of-the-art performance on unsupervised completion of real scene objects. For those categories without artificial data as assistance, previous works and our method are hard to apply in this situation. This is our limitation. Unsupervised point cloud completion without the assistance of artificial data is our future work.", + "bbox": [ + 496, + 599, + 890, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 835, + 658, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant numbers 62032011, 61972194.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X. Chang, and Matthias Niessner. Scan2cad: Learning cad model alignment in rgb-d scans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2, 5", + "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5", + "[3] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jul 2017. 4", + "[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1", + "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 1, 2, 5", + "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5, 6", + "[7] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1359-1368, 2019. 2", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 3", + "[9] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 2020. 1", + "[10] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh. ACM Transactions on Graphics (TOG), 39:126:1 - 126:12, 2020. 1", + "[11] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7662-7670, 2020. 2", + "[12] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 1", + "[13] Bo Li. 3d fully convolutional network for vehicle detection in point cloud. In 2017 IEEE/RSJ International Conference" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Intelligent Robots and Systems (IROS), pages 1513-1518. IEEE, 2017. 1", + "[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, 2017 Advances in Neural Information Processing Systems (NIPS), volume 30. Curran Associates, Inc., 2017. 6, 7", + "[15] Abhishek Sharma, Oliver Grau, and Mario Fritz. Vconv-dae: Deep volumetric shape learning without object labels. In European Conference on Computer Vision, pages 236-250. Springer, 2016. 2", + "[16] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1955-1964, 2018. 2", + "[17] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1716-1725, 2022. 1, 2", + "[18] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 6", + "[19] Lyne P Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese. Topnet: Structural point cloud decoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 383-392, 2019. 1, 2, 5, 6, 7", + "[20] Duc Thanh Nguyen, Binh-Son Hua, Khoi Tran, Quang-Hieu Pham, and Sai-Kit Yeung. A field model for repairing 3d shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5676-5684, 2016. 2", + "[21] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 8", + "[22] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 790-799, 2020. 1, 2", + "[23] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Learning local displacements for point cloud completion. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022. 1, 2, 5, 6, 7", + "[24] Xin Wen, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Cycle4completion: Unpaired point cloud completion using cycle transformation with missing region coding. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13075-13084, 2021. 1, 3, 5, 6, 7", + "[25] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF Con" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 1939-1948, 2020. 2", + "[26] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net: Point cloud completion by learning multi-step point moving paths. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[27] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In The European Conference on Computer Vision (ECCV), August 2020. 3", + "[28] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Density-aware chamfer distance as a comprehensive metric for point cloud completion. 11 2021. 2", + "[29] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2022. 3", + "[30] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1", + "[31] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2", + "[32] Niloy J. Mitra Xuelin Chen, Baoquan Chen. Unpaired point cloud completion on real scans using adversarial training. In 2020 International Conference on Learning Representations 2020 (ICLR), 2020. 1, 3, 5, 6, 7, 8", + "[33] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 206–215, 2018. 1", + "[34] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 2, 5, 6, 7", + "[35] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 International Conference on 3D Vision (3DV), pages 728-737. IEEE, 2018. 1, 2, 6", + "[36] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1768-1777, 2021. 1, 3, 5, 6, 7, 8", + "[37] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15890-15899, June 2021. 2" + ], + "bbox": [ + 78, + 92, + 467, + 882 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13569", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_model.json b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..37b07b0c84fd36cda5931e71b5f859f4516377a1 --- /dev/null +++ b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.848, + 0.177 + ], + "angle": 0, + "content": "Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.203, + 0.848, + 0.275 + ], + "angle": 0, + "content": "Changfeng Ma1, Yinuo Chen1, Pengxiao Guo1, Jie Guo1, Chongjun Wang1, Yanwen Guo1* \n1Nanjing University, Nanjing, China \n{changfengma, yinuochen, pxguo}@smail.nju.edu.cn \n{guojie, chjwang, ywguo}@nju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.324 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Unsupervised completion of real scene objects is of vital importance but still remains extremely challenging in preserving input shapes, predicting accurate results, and adapting to multi-category data. To solve these problems, we propose in this paper an Unsupervised Symmetric Shape-Preserving Autoencoding Network, termed USSPA, to predict complete point clouds of objects from real scenes. One of our main observations is that many natural and man-made objects exhibit significant symmetries. To accommodate this, we devise a symmetry learning module to learn from those objects and to preserve structural symmetries. Starting from an initial coarse predictor, our autoencoder refines the complete shape with a carefully designed upsampling refinement module. Besides the discriminative process on the latent space, the discriminators of our USSPA also take predicted point clouds as direct guidance, enabling more detailed shape prediction. Clearly different from previous methods which train each category separately, our USSPA can be adapted to the training of multi-category data in one pass through a classifier-guided discriminator, with consistent performance on single category. For more accurate evaluation, we contribute to the community a real scene dataset with paired CAD models as ground truth. Extensive experiments and comparisons demonstrate our superiority and generalization and show that our method achieves state-of-the-art performance on unsupervised completion of real scene objects." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.737, + 0.21, + 0.752 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.47, + 0.877 + ], + "angle": 0, + "content": "As the standard outputs of 3D scanners [12, 32], point clouds are becoming more and more popular [9] which are also the basic data structure in 3D geometry processing [4, 5, 13]. Complete point clouds are hard to obtain due to the nature of the scanning process and object occlusion [35]. Due to the defects of incomplete point clouds on downstream applications such as reconstruction [10], recent works [17, 19, 22, 23, 26, 30, 33, 35] pay more attention" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.308, + 0.88, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.632, + 0.894, + 0.756 + ], + "angle": 0, + "content": "Figure 1. Visual comparison of predicted results on real scene data by our USSPA and other works (top) and our complete result on a whole real scene (bottom). (a) shows an example of a real scene partial point cloud of a chair and the complete predictions by Disp3D [23], ShapeInv [36], Unpaired [32] and our method. As shown, our prediction result is more accurate and uniform according to the input, which contains complete arms and legs. (b) and (c) show the original point cloud of a real scene and the complete results of all the objects in this scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.893, + 0.883 + ], + "angle": 0, + "content": "to point cloud completion which relies on paired artificial complete point clouds for supervised training to complete partial point clouds. However, these supervised works are difficult to apply in practice because of the great gap between artificial data and real scene data and the inaccessibility to the ground truth of real scene data. Therefore, it is important to complete partial point clouds from real scene in an unsupervised way." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Recent unsupervised works [24, 32, 36] only require real" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "scene partial point clouds and artificial CAD models for unpaired completion utilizing GANs [8] as their fundamental frameworks, most of which need pre-training on artificial data. The main ideas are to transform latent codes from the space of real scene partial data to the space of artificial complete data and then employ the decoder trained on artificial data to predict the complete point cloud. Essentially, these methods make the predictions whose distributions are consistent with the artificial models. Most of them, however, just extract a global feature from the partial input without fully exploiting its geometry information, leading to the prediction severely deviated from the input. And such information actually provides vital clues and constraints for completion. Furthermore, prediction results by these methods usually lack enough geometric details due to the absence of an explicit discriminative process on point clouds. These domain transforming methods are also hard to adapt to multi-category data or other datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.868 + ], + "angle": 0, + "content": "In this paper, we present an unsupervised symmetric shape-preserving autoencoding network, termed USSPA, for the completion of real scene objects, as shown in Figure 2, which is a GAN-based end-to-end network without the requirement of pre-trained weights. Different from previous domain transforming methods which cannot fully leverage existing incomplete models, we argue that the existing partial scanning, which also provides vital clues and constraints for the prediction of the missing part, should be preserved to some extent. To this end, we exploit the symmetries shown in many natural or man-made objects and devise a novel symmetry learning module to generate symmetrical point clouds of existing parts by predicting the symmetric planes. This enables our network to preserve the shapes of input symmetrically, intrinsically facilitating structure completion, as shown in Figure 1. For those parts that can not be directly inferred from inputs, we employ an initial coarse module for an initial prediction first. Starting from the initial guess, we specifically design a refinement autoencoder with an upsampling refinement module for detailed refinement and the local feature grouping for extracting local information, to learn detailed structures of artificial data through the autoencoding process. Benefiting from this, our final prediction is accurate, uniform, and symmetric shape-preserving. Besides the indirect guidance of the feature discriminator on latent space, our point discriminator takes predicted point clouds as direct guidance for generating more accurate shapes. Compared with previous methods which train each category separately, our method can classify the partial point clouds simultaneously through a classifier-guided discriminator when adapted to multi-category data, with consistent performance on the single category." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To measure the performance of unsupervised completion quantitatively, we build a dataset from ScanNet [5] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "ShapeNet [2] utilizing the annotations of Scan2CAD [1]. Our dataset contains real scene partial point clouds and paired ground truths that are only used for evaluation in our experiments. Extensive comparisons against previous works on this dataset and the public PCN Dataset [35] show the superiority and generalization of our method which achieves state-of-the-art performance on unsupervised completion of real scene objects." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.215, + 0.774, + 0.228 + ], + "angle": 0, + "content": "Our main contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.236, + 0.892, + 0.344 + ], + "angle": 0, + "content": "- We propose a novel USSPA for unsupervised real scene point cloud completion whose prediction is accurate, uniform and symmetric shape-preserving. Clearly different from previous works training each category separately, our USSPA can be adapted to the training of multi-category data in one pass by classifying the input simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.345, + 0.892, + 0.435 + ], + "angle": 0, + "content": "- We propose a novel symmetry learning module and a novel refinement autoencoder. The symmetry learning module preserves input shapes by generating symmetrical point clouds, and the refinement autoencoder learns the detailed information from artificial data to refine the initial guess by an autoencoding process." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.439, + 0.892, + 0.514 + ], + "angle": 0, + "content": "- We propose a new evaluation method for obtaining paired ground truths and partial data from artificial and real scene datasets using alignment information, which can be used to more accurately evaluate unsupervised completion of real scene objects." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.236, + 0.892, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.642, + 0.543 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.813, + 0.567 + ], + "angle": 0, + "content": "2.1. Supervised Point Cloud Completion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.895, + 0.905 + ], + "angle": 0, + "content": "To infer the complete point cloud from an incomplete input, many deep learning methods were introduced in recent years, most of which are supervised. At first, the 3D data were voxelized into occupancy grids or distance fields before fed into convolutional networks [15, 16, 20]. To avoid loss of details and huge memory cost brought by these data representations, Yuan et al. [35] propose PCN which directly operates on raw point clouds. Giancola et al. [7] also leverage a new method for 3D shape completion which directly works on LiDAR point clouds. Later works [11, 19, 22] pay much attention to details refinement and denoise. Disp3D [23] investigates grouping local features to bring improvement to point cloud completion. Taking structures and topology information into consideration, SA-Net [25], GRNet [31], PoinTr [34] and LAKe-Net [17] succeed to utilize the detailed geometry information from input to generate more reasonable complete point clouds. Other works like designing new metrics [28] and combining information of single-view images [37] also make contributions to the point cloud completion task. However, these works are all supervised that are not suitable for practice since ground truths are in lack." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13561" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.538, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.871, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.895, + 0.392 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our USSPA. Our network has a generator composed of an initial coarse predictor (blue) and a refinement autoencoder (green), and the point and feature discriminators (yellow). The red lines and boxes represent the employed losses. The blue and green lines (bottom) denote the forward processes having the same network architectures with sharing weights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.396, + 0.411, + 0.412 + ], + "angle": 0, + "content": "2.2. Unsupervised Point Cloud Completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.474, + 0.779 + ], + "angle": 0, + "content": "Due to the absence of suitable paired training data for supervised models, some unsupervised models for point cloud completion have emerged in recent years. Unpaired [32] first introduces the GAN to learn the transformation from the latent space of partial point cloud to the latent space of complete point cloud, and then decode the transformed latent code to predict. Zhang et al. [36] employ optimization-based GAN inversion [29] and directly find the complete prediction by optimizing the latent code such that the degradation of the complete point cloud matches the input partial point cloud, which requires fine-tuning the network for each input data. Cycle4 [24] proposes two simultaneous cycle transformations between the latent spaces of complete point cloud and partial ones, facilitating the network's understanding of 3D shapes by building a bidirectional mapping. Wu et al. [27] propose the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. However, the application of pre-trained GANs makes their predictions lack details and difficult to preserve the shape features of the input. For evaluation, Unpaired [32] employs classification accuracy of the complete results which is hard to measure detailed structures. Thus, a more accurate evaluation method is required." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.17, + 0.799 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.283, + 0.82 + ], + "angle": 0, + "content": "3.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Given the partial point cloud \\(\\mathcal{P}_R\\in \\mathbb{M}_{n_0\\times 3}\\) from real scene data with distribution \\(p_r\\) and the complete point cloud \\(\\mathcal{P}_A\\in \\mathbb{M}_{n_0\\times 3}\\) from artificial data with distribution \\(p_a\\), the goal of unsupervised real scene point cloud completion is to predict the complete point cloud \\(Com(\\mathcal{P}_R)\\) of \\(\\mathcal{P}_R\\) such" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.892, + 0.442 + ], + "angle": 0, + "content": "that \\( \\operatorname{Com}(\\mathcal{P}_R) \\sim p_a \\) and \\( \\mathcal{P}_R \\subset \\operatorname{Com}(\\mathcal{P}_R) \\). This means that the prediction must be similar to the artificial data and meanwhile retain the shapes of the input partial point cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.893, + 0.718 + ], + "angle": 0, + "content": "We carefully design several modules of our generator \\(\\mathcal{G}\\) and discriminators, \\(\\mathcal{D}_p\\) and \\(\\mathcal{D}_f\\), for unsupervised completion on real scene data through generative adversarial learning [8] as shown in Figures 2 and 3. As we mentioned above, many natural and man-made objects exhibit significant symmetries. To fully leverage these symmetries and preserve the shape of existing partial input \\(\\mathcal{P}_R\\), we devise a novel symmetry learning module that predicts the symmetric plane and the symmetrical point cloud of \\(\\mathcal{P}_R\\) for further refinement. Our refinement autoencoder learns the detailed information of artificial data and refines the details of the initial, but coarse prediction \\(P_{R_0}\\) for uniform and accurate prediction. With the local feature grouping operation, our elaborate upsampling refinement module can utilize the local features for more detailed results. Different from previous works focusing on the discriminative process of latent space, our point discriminator \\(\\mathcal{D}_p\\) also takes point clouds as input for direct guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Initial Coarse Predictor. We first employ an initial coarse predictor for shape-preserving and preliminary prediction. The encoder \\( E_{R} \\) takes the real scene partial point cloud \\( \\mathcal{P}_R \\) as input and encodes the global feature \\( f_{R}^{\\prime} = E_{R}(\\mathcal{P}_{R})\\in \\mathbb{M}_{1\\times c_{0}} \\). \\( f_{R}^{\\prime} \\) is then taken by the decoder \\( D_{R} \\) for the prediction of initial, but coarse complete result \\( P_R' = D_R(f_R')\\in \\mathbb{M}_{n_1\\times 3} \\). Encoder-decoder architecture is hard to retain shape features on the completion task since there is no explicit complete ground truth. To this end, we devise the symmetry learning module to accommodate symmetrical shape features widely existed in man-made and natural objects, as shown in Figure 3. The decoder \\( D_{sl} \\) pre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13562" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.138 + ], + "angle": 0, + "content": "dicts symmetric arguments \\( \\mathbf{A} = D_{sl}(f_R') = [x_A, y_A, z_A]^T \\) and then we introduce symmetry operation to calculate the symmetrical point cloud \\( P_{sym} \\) of \\( \\mathcal{P}_R \\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.148, + 0.469, + 0.18 + ], + "angle": 0, + "content": "\\[\nP _ {s y m} = \\boldsymbol {p} - 2 \\frac {\\boldsymbol {A} \\cdot \\boldsymbol {p}}{\\| \\boldsymbol {A} \\| ^ {2}} \\boldsymbol {A}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.19, + 0.469, + 0.312 + ], + "angle": 0, + "content": "where \\(\\pmb{p} = [x,y,z]^T\\in \\mathcal{P}_R\\) \\(\\cdot\\) represents the inner product and \\(||\\pmb {x}||\\) denotes the module of the vector \\(\\pmb{x}\\) . Without the loss of generality, we can safely assume that all the objects are axis-aligned. Under such an assumption, the symmetric plane is perpendicular to the \\(xz\\) -plane and zero-crossing. Thus we set \\(y_{A} = 0\\) . Then \\(P_{sym}\\) \\(\\mathcal{P}_R\\) and \\(P_R^{\\prime}\\) are fused and resampled through FPS (farthest point sampling) to get \\(P_{R_0}\\) which is the complete and shape-preserving result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Refinement Autoencoder. Noisy \\(\\mathcal{P}_R\\) and the fusion operation may cause noisy \\(P_{R_0}\\) with uneven point distribution. The initial coarse predictor or the point discriminator only predicts or judges through the global feature, which makes \\(P_{R_0}\\) lack details. Thus, \\(P_{R_0}\\) requires denoising and further refinement on details and point distribution. To this end, we employ the refinement autoencoder with a carefully designed upsampling refinement module. As shown in Figure 2, the blue and green lines of the refinement autoencoder represent the same forward process having the same network architectures with sharing weights. In this way, our autoencoder extracts the detailed information of artificial data and refines the details of \\(P_{R_0}\\) for more accurate and uniform prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.904 + ], + "angle": 0, + "content": "The encoder and decoder \\(E_{A},D_{A}\\) generate the skeleton point cloud \\(P_{R_1 / A_1}\\in \\mathbb{M}_{n_2\\times 3}\\) of input point cloud. We set \\(n_2\\) to a small number for easier prediction of \\(P_{R_1 / A_1}\\). And the density loss is faster to calculate with fewer points. We then employ upsampling refinement module to upsample \\(P_{R_1 / A_1}\\) and add details to the skeleton for final prediction \\(P_{R_2 / A_2}\\). As shown in Figure 3, we utilize wight-shared MLP [3] to obtain the point features and max-pooling operation on point features to obtain the global feature of the point cloud. We also concatenate the point and global features to fuse the information. Since wight-shared MLP only extracts the feature of one point, lacking the local information, we also employ the local feature grouping to get the local information around each point. For point \\(p_0\\), \\(f_{group}\\in \\mathbb{M}_{n_2\\times (c_1\\cdot m)}\\) are the features grouped from \\(f_{input}\\in \\mathbb{M}_{n_2\\times c_1}\\) corresponding to the points \\(p\\) that \\(\\| p - p_0\\| \\leq r\\), where \\(m\\) is the feature number of each group. After applying max-pooling operation on \\(m\\) features of each group, \\(f_{local}\\in \\mathbb{M}_{n_2\\times c_1}\\) are concatenated with \\(f_{input}\\) to generate the output \\(f_{output}\\in \\mathbb{M}_{n_2\\times 2c_1}\\). We utilize a weight-shared MLP for upsampling that takes fused features composed of \\(P_{R_1 / A_1}\\), local and global features, and predict \\(k\\) shifts for each point. Then, each point splits into \\(k\\) points by adding the predicted shifts to its coordinate whose result is \\(P_{R_2 / A_2}\\in \\mathbb{M}_{kn_2\\times 3}\\)." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.88, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.178, + 0.731, + 0.192 + ], + "angle": 0, + "content": "Upsampling Refinement Module" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.193, + 0.88, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.291, + 0.682, + 0.305 + ], + "angle": 0, + "content": "Local Feature Grouping" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.306, + 0.758, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.405, + 0.7, + 0.419 + ], + "angle": 0, + "content": "Symmetry Learning Module" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.311, + 0.794, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.336, + 0.794, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.361, + 0.794, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.393, + 0.794, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.316, + 0.874, + 0.326 + ], + "angle": 0, + "content": "Concatenate" + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.34, + 0.874, + 0.352 + ], + "angle": 0, + "content": "Max-Pooling" + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.361, + 0.876, + 0.385 + ], + "angle": 0, + "content": "Local Feature Grouping" + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.393, + 0.86, + 0.415 + ], + "angle": 0, + "content": "Element Addition" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.421, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Detailed structures of our symmetry learning module (red), upsampling refinement module (yellow), and local feature sampling (blue)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Point and Feature Discriminator. We employ the point discriminator \\(\\mathcal{D}_p\\) to directly guide the generation of \\(P_{R_0}\\) and \\(P_{R_2}\\) comparing with the artificial complete data \\(\\mathcal{P}_A\\) for unsupervised completion. We also guide the generation of latent code of autoencoder through the feature discriminator \\(\\mathcal{D}_f\\) such that \\(f_R\\) and \\(f_A\\) have the same distribution for more accurate prediction. Our discriminators predict the probability that input point cloud \\(P_i\\) or feature \\(f_i\\) is real data. For multi-category data, we replace the probability-guided discriminator with a classifier-guided discriminator which outputs the category label of input rather than probability. We utilize label 0 to represent fake data and labels \\(l > 0\\) to denote real data of different categories such as the chair, table, etc. For fake data, the discriminator tends to output label 0, while the generator tends to make the discriminator output label \\(l > 0\\). And for real data with label \\(l\\), the discriminator tends to output \\(l\\) correctly. In this way, our USSPA can unsupervisedly classify the real scene partial point cloud when completing multi-category data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.758, + 0.639, + 0.773 + ], + "angle": 0, + "content": "3.2. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.776, + 0.89, + 0.807 + ], + "angle": 0, + "content": "The abbreviate optimization goal of our USSPA for unsupervised learning is:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.82, + 0.89, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G}} \\max _ {\\mathcal {D} _ {p}, \\mathcal {D} _ {f}} \\mathbb {E} _ {P _ {i} \\sim p _ {a}} \\log p + \\mathbb {E} _ {P _ {i} \\sim p _ {r}} \\log (1 - p), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.891, + 0.902 + ], + "angle": 0, + "content": "where \\( p \\) is the probability of real data. By playing the maxmin game with \\( \\mathcal{D}_p \\) and \\( \\mathcal{D}_f \\), our generator \\( \\mathcal{G} \\) tends to predict point clouds similar to artificial data. We employ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13563" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. Point cloud completion comparison on our dataset in terms of L1 Chamfer Distance \\(cd^{l1} \\times 10^2\\) (lower is better). All the methods are trained with our dataset and the unsupervised methods including ours are trained on single category data. Boldface denotes the best among unsupervised methods in Tables 1, 2 and 3." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.133, + 0.885, + 0.279 + ], + "angle": 0, + "content": "
MethodAVGchairtabletrash binTVcabinetbookshelfsofalampbedtub
Sup.PoinTr [34]14.3713.6512.5215.2612.6917.3213.9912.3617.0515.1313.77
Disp3D [23]7.786.248.207.127.1210.366.945.6014.036.905.32
TopNet [19]7.076.395.797.406.268.377.025.948.507.817.25
Unsup. OptBasedShapeInv [36]21.3917.9717.2833.5115.6926.2625.5114.2816.6932.3314.43
Unsup.Unpaired [32]10.478.417.5212.086.7217.459.956.9219.3610.046.22
Cycle4 [24]11.539.1111.3511.938.4015.4712.5110.6312.2515.737.92
Ours8.568.227.6810.367.6610.777.846.1411.938.206.75
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.47, + 0.343 + ], + "angle": 0, + "content": "Chamfer Distance(CD) [6] to measure the difference between point clouds. The \\(cd_{P_1 \\to P_2}^l\\) indicates the single direction Chamfer Distance with L2-norm from point cloud \\(P_1\\) to \\(P_2\\):" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.349, + 0.47, + 0.387 + ], + "angle": 0, + "content": "\\[\nc d _ {P _ {1} \\rightarrow P _ {2}} ^ {l 2} = \\frac {1}{| P _ {1} |} \\sum_ {\\boldsymbol {x} \\in P _ {1}} \\min _ {\\boldsymbol {y} \\in P _ {2}} \\| \\boldsymbol {x} - \\boldsymbol {y} \\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.395, + 0.47, + 0.426 + ], + "angle": 0, + "content": "which guides \\( P_{1} \\) to be a part of \\( P_{2} \\). The \\( cd_{P_1\\leftrightarrow P_2}^{l^2} \\) between \\( P_{1} \\) and \\( P_{2} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.433, + 0.469, + 0.453 + ], + "angle": 0, + "content": "\\[\nc d _ {P _ {1} \\leftrightarrow P _ {2}} ^ {l 2} = c d _ {P _ {1} \\rightarrow P _ {2}} ^ {l 2} + c d _ {P _ {2} \\rightarrow P _ {1}} ^ {l 2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.473, + 0.564 + ], + "angle": 0, + "content": "which guides \\(P_{1}\\) and \\(P_{2}\\) to be identical. As shown in Figure 2, to restrict the generation of \\(P_{R}^{\\prime}\\), we employ single direction Chamfer Distance from \\(\\mathcal{P}_R\\) to \\(P_{R}^{\\prime}\\). We also utilize Chamfer Distance between \\(P_{R_1 / A_1}, P_{R_2 / A_2}\\) and \\(P_{R_0}, \\mathcal{P}_A\\) for supervision of autoencoder. For uniform prediction, we employ density loss by first calculating the local density of \\(\\boldsymbol{x} \\in P\\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.572, + 0.47, + 0.613 + ], + "angle": 0, + "content": "\\[\nl d (\\boldsymbol {x}, P) = \\frac {1}{k _ {d}} \\sum_ {k = 1} ^ {k _ {d}} \\min _ {\\boldsymbol {y} \\in P} ^ {k} | | \\boldsymbol {x} - \\boldsymbol {y} | |, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.62, + 0.424, + 0.635 + ], + "angle": 0, + "content": "and then calculating the standard deviation for all \\(x\\):" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.644, + 0.47, + 0.684 + ], + "angle": 0, + "content": "\\[\nd l (P) = \\sqrt {\\frac {1}{| P |} \\sum_ {\\boldsymbol {x} \\in P} \\left(l d (\\boldsymbol {x} , P) - \\overline {{l d (\\boldsymbol {x} , P)}}\\right) ^ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.694, + 0.469, + 0.731 + ], + "angle": 0, + "content": "where \\(\\min\\) indicates the \\(k\\)-th-min element. The smaller the \\(ld(\\pmb{x}, P)\\) is, the bigger the local density of \\(\\pmb{x}\\) is." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.732, + 0.438, + 0.747 + ], + "angle": 0, + "content": "In practice, we alternately optimize \\(\\mathcal{G}\\) with loss \\(\\mathcal{L}_{\\mathcal{G}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.757, + 0.469, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} = \\alpha_ {1} \\mathcal {L} _ {F \\rightarrow R} + \\alpha_ {2} \\mathcal {L} _ {P} + \\alpha_ {3} \\mathcal {L} _ {d}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.784, + 0.124, + 0.797 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.804, + 0.47, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {F \\rightarrow R} = - \\left(\\log \\mathcal {D} _ {p} \\left(P _ {R _ {0}}\\right) + \\log \\mathcal {D} _ {p} \\left(P _ {R _ {2}}\\right)\\right. \\tag {8} \\\\ + \\log \\mathcal {D} _ {f} (f _ {R})) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.861, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {P} = c d _ {P _ {R _ {1}} \\leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {P _ {R _ {2}} \\leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {\\mathcal {P} _ {R} \\rightarrow P _ {R} ^ {\\prime}} ^ {l 2} \\tag {9} \\\\ + c d _ {P _ {A _ {1}} \\leftrightarrow \\mathcal {P} _ {A}} ^ {l ^ {2}} + c d _ {P _ {A _ {2}} \\leftrightarrow \\mathcal {P} _ {A}} ^ {l ^ {2}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.311, + 0.892, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = d l \\left(P _ {R _ {1}}\\right) + d l \\left(P _ {A _ {1}}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.641, + 0.356 + ], + "angle": 0, + "content": "and \\(\\mathcal{D}\\) with loss \\(\\mathcal{L}_{\\mathcal{D}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.37, + 0.892, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {D}} = \\alpha_ {4} \\mathcal {L} _ {F} + \\alpha_ {5} \\mathcal {L} _ {R}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.4, + 0.547, + 0.413 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.422, + 0.892, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {F} = - \\left(\\log \\left[ 1 - \\mathcal {D} _ {p} \\left(P _ {R _ {0}}\\right) \\right] + \\log \\left[ 1 - \\mathcal {D} _ {p} \\left(P _ {R _ {2}}\\right) \\right] \\right. \\tag {12} \\\\ \\left. + \\log \\left[ 1 - \\mathcal {D} _ {f} \\left(f _ {R}\\right) \\right]\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.491, + 0.892, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = - \\left(\\log \\mathcal {D} _ {p} \\left(\\mathcal {P} _ {A}\\right) + \\log \\mathcal {D} _ {f} \\left(f _ {A}\\right)\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.582 + ], + "angle": 0, + "content": "and here \\(\\alpha_{1\\sim 5}\\) are the weights used for balancing the influences between each term. And for multi-category data, we use cross-entropy to calculate the loss between predicted labels and ground truth labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.634, + 0.607 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.813, + 0.626 + ], + "angle": 0, + "content": "4.1. Dataset and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Different from the Unpaired [32] which utilizes a classifier to evaluate the completion results of real scene data, we build a new dataset from ScanNet [5], ShapeNet [2] and Scan2CAD [1], containing paired partial and complete point clouds for more accurate evaluation. Scan2CAD contains annotations that align the CAD models of ShapeNet to the real scenes of ScanNet. We first extract the CAD models from ShapeNet and partial point clouds from ScanNet which are paired according to the correspondences offered by Scan2CAD, and then uniformly sample the CAD models to get complete point clouds. Finally, we transform the paired partial and complete point clouds with the transform matrix given by Scan2CAD to normalize their face direction, scale, and position. We select ten categories including the chair, table, etc., to establish the dataset and randomly split the data into training, validation, and testing sets. In our experiments, these ground truths are only used for the evaluation of the complete results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13564" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.87, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Figure 4. Visualization of completion results by supervised works, including TopNet [19], Disp3D [23], unsupervised works Cycle4 [24], Unpaired [32], optimization-based work ShapeInv [36] and our method. From top to bottom: table, bed, cabinet, and TV." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.473, + 0.626 + ], + "angle": 0, + "content": "We utilize PyTorch to implement our USSPA. In Section 3.1, the point numbers \\( n_{0\\sim 2} \\) are set to 2048, 512 and 512. The channel numbers \\( c_{0\\sim 1} \\) are set to 512 and 256. The feature number \\( m \\), the radius \\( r \\) of local feature grouping, and the split number of upsampling \\( k \\) are set to 32, 0.5, and 4. \\( k_d \\) of density loss in Section 3.2 is set to 16. The weights \\( \\alpha_{1\\sim 5} \\) of losses are set to 1, 100, 7.5, 0.5 and 0.5. We employ Adam optimizer whose learning rate is \\( 1.0 \\times 10^{-5} \\). The batch size is set to 4, and the maximum epoch of training is 240 to 960 according to the size of the dataset. Training our USSPA takes about 20 hours with a GTX 2080Ti GPU. We will make our dataset and codes public in the future." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.637, + 0.21, + 0.654 + ], + "angle": 0, + "content": "4.2. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We compare our USSPA against existing representative point cloud completion methods including supervised [19, 23, 34], unsupervised [24, 32] and optimization-based [36] methods on our dataset and the PCN Dataset [35]. The metrics used for comparison include pre-point L1 Chamfer Distance \\(cd^{l1}\\) [6] and F-Scores \\(F_{score}^{1\\%}\\), \\(F_{score}^{0.1\\%}\\) [18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1, we compare these methods on our dataset with 10 categories where all methods are trained on the training set and evaluated on the testing set. All the unsupervised methods are trained with single category data for each class. The average \\(cd^{l1}\\) of our USSPA is \\(18.2\\%\\) lower than Unpaired and the average \\(F_{score}^{0.1\\%}\\) is \\(30\\%\\) higher than Unpaired as shown in Table 2, indicating the superior performance of our method on completion of real scene objects. We also mix up all 10 categories as the multi-category data to train the methods for evaluation of the generalization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.442, + 0.893, + 0.526 + ], + "angle": 0, + "content": "Table 2. Comparison on our dataset trained with single category and multi-category data in terms of L1 Chamfer Distance \\(cd^{l1} \\times 10^2\\) (lower is better) and F-scores \\(F_{score}^{0.1\\%} \\times 10^2\\), \\(F_{score}^{1\\%} \\times 10^2\\) (higher is better). \"Ours\" and \"Ours(classifier)\" denote our method with probability-guided and classifier-guided discriminators, respectively." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.53, + 0.888, + 0.711 + ], + "angle": 0, + "content": "
Methodsingle categorymulti-category
↑F0.1% score↑F1% score↓cdl1↑F0.1% score↑F1% score
PoinTr [34]--14.3718.3580.41
Disp3D [23]--7.7830.2978.26
TopNet [19]--7.0712.3380.37
ShapeInv [36]15.5866.5319.3516.9869.66
Unpaired [32]12.2064.3310.1210.8666.68
Cycle4 [24]9.9860.1412.008.6156.57
Ours17.4973.418.9616.8872.31
Ours(classifier)--8.7617.1273.75
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "ability. As shown in Table 2, our USSPA keeps the same performance by replacing the probability-guided discriminators with classifier-guided discriminators and has better generalization ability on multi-category than other unsupervised networks. Further comparisons show the generalization capability of our USSPA on different datasets. We train unsupervised models with ScanNet and then test them on the PCN Dataset which is a point cloud completion dataset built upon artificial data. As shown in Table 3, our method also performs better than other unsupervised methods, the average \\(cd^{l1}\\) of which is \\(15\\%\\) lower than the SOTA method. Table 4 shows the classification accuracy of PN2 [14] and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13565" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.471, + 0.132 + ], + "angle": 0, + "content": "Table 3. Point cloud completion comparison on PCN Dataset in terms of L1 Chamfer Distance \\( cd^{l1} \\times 10^2 \\) (lower is better). All unsupervised models are trained with ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.143, + 0.465, + 0.274 + ], + "angle": 0, + "content": "
MethodAVGchairtablecabinetsofalamp
PoinTr [34]5.495.615.686.085.674.44
Disp3D [23]2.512.422.302.382.443.00
TopNet [19]5.926.345.456.065.805.95
ShapeInv [36]19.0523.1815.6617.1422.8516.40
Unpaired [32]14.8712.878.1414.3018.2320.82
Cycle4 [24]17.6014.2515.7321.0621.5415.40
Ours12.6313.529.668.8915.5115.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.289, + 0.47, + 0.332 + ], + "angle": 0, + "content": "Table 4. Point cloud classification in terms of accuracy. \"AVG\" denotes the average accuracy on 10 categories of our dataset. Only four categories are shown due to width limitation." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.343, + 0.458, + 0.402 + ], + "angle": 0, + "content": "
MethodAVGchairtabletrash binsofa
PN2 [14]58.6%91.0%94.0%17.2%48.9%
Ours69.8%91.0%91.0%51.6%93.3%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.469, + 0.455 + ], + "angle": 0, + "content": "our method. Our method can also unsupervisedly classify the real scene partial point cloud while completing. The accuracy of our USSPA is \\(11.2\\%\\) higher than PN2 on average." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.683 + ], + "angle": 0, + "content": "Figure 4 shows the qualitative results on 4 categories. The complete point clouds predicted by our USSPA exhibit more accurate shapes, such as the leg of the table and the division plate of the bookshelf, which are hard to generate even for supervised methods. Our method can successfully generate the circular top of the table. By comparison, other unsupervised works generate a square one. Our predictions are also more uniform with fewer noises. Benefiting from our shape-preserving method, our predicted TV retains the support part from the input point cloud. Our symmetry learning module also plays an important role in completing the pillars of the bed by copying the pillars from the input symmetrically. Visualization of our predicted point clouds shows the superiority of our methods in the generation of detailed completion on real scene data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.231, + 0.7 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Ablation study on network architecture. We conduct experiments on ablated models for evaluation of the necessity of important structures of our USSPA, including the symmetry learning module, the refinement autoencoder, the local feature grouping, the point and feature discriminators, the single direction Chamfer Distance, and the upsampling refinement module. As shown in Table 5, these modules are removed from our full network separately. We downsample the final prediction of our USSPA to 512 points for a fair comparison to the ablated model without the upsampling refinement module. We train and utilize \\( F_{score}^{0.1\\%} \\) to evaluate these ablated models with our dataset on the chair category, and the comparison indicates the necessity of these modules" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.133 + ], + "angle": 0, + "content": "Table 5. Ablation study in terms of \\( F_{score}^{0.1\\%} \\), where “-” represents removing a module from the full network respectively and “512” denotes the point number." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.143, + 0.864, + 0.307 + ], + "angle": 0, + "content": "
ModelF0.1% score
- Symmetry Learning Module16.43
- Refinement Autoencoder17.08
- Local Feature Grouping15.72
- Feature Discriminator17.11
- Point Discriminator15.68
- Single Direction Chamfer Distance17.84
Ours (full)18.43
- Upsampling Refinement (512)5.60
Ours (full) (512)6.46
" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.319, + 0.892, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Figure 5. Visualization of the symmetric planes and predictions by USSPA without the symmetry learning module and full network. Blue and orange denote original and mirrored points, separately." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.892, + 0.522 + ], + "angle": 0, + "content": "which play significant roles in shape-preserving, extracting, and utilizing detailed features, upsampling and refining for more accurate point cloud completion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.893, + 0.764 + ], + "angle": 0, + "content": "Symmetry learning module. Our symmetry learning module learns different symmetric planes according to the input point clouds as shown in Figures 5(a) and 5(b). The predicted symmetric plane is xy-plane for the chair and yz-plane for the table. Through different symmetric planes, our network can fully leverage the symmetrical structures of input point clouds, such as the arm of the chair and the legs of the table, which exist on only one side. Figure 5(c) shows the prediction \\( P_{R_2} \\) of our network without the symmetry learning module, and Figure 5(d) shows the result of our full network. Figure 5(e) shows the input point cloud whose left arm is missing while the right arm is complete. Without the symmetry learning module, the left arm of the prediction is incomplete as shown in the red box. And benefiting from such a module, our full network can learn the symmetric structures, and generate a complete left arm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Density loss. We visualize the local density of points as shown in Figure 6. The warmer the color of the point is, the bigger its local density (described in equation 5) is. (a) is the prediction without the supervision of the density loss, and (b) is the prediction supervised by the density loss. Without the supervision of density loss, the network prefers generating many points in the same location as shown in the blue circle to minimize the Chamfer Distance, which makes the predicted point clouds contain fewer points describing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13566" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.146 + ], + "angle": 0, + "content": "Table 6. Comparison of Unpaired and our USSPA in terms of density loss, where \"ours (w/o \\( dl \\))\" indicates our USSPA without the supervision of the density loss. The results show that our predicted point clouds are more uniform." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.154, + 0.439, + 0.197 + ], + "angle": 0, + "content": "
Unpairedours (w/o dl)ours
Density Loss0.09570.12380.0149
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.204, + 0.174, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.205, + 0.272, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.21, + 0.374, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.208, + 0.472, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.331, + 0.47, + 0.415 + ], + "angle": 0, + "content": "Figure 6. Visualization of the local density of point clouds, where the warmth of the color indicates the values of local densities. The warmer the color is, the bigger the local density is. (a) and (b) are the comparison of the ablation experiment on the density loss. (c) and (d) are the examples of \\( P_{R_0} \\) and \\( P_{R_2} \\). The density loss and the point number of each point cloud are shown at its bottom." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.472 + ], + "angle": 0, + "content": "details such as chair arms and legs. Table 6 compares Unpaired and our USSPA on density loss indicating that our predicted point clouds are more uniform." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Autoencoding refiner. As shown in Figure 6, (c) and (d) are the examples of \\( P_{R_0} \\) and \\( P_{R_2} \\) which is the refinement result of the autoencoding refiner on \\( P_{R_0} \\). The refiner removes the noises from input shown in blue boxes and generates more points on legs as shown in green boxes, which makes the whole point cloud more uniform and accurate with less density loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Real scene data. We also test our USSPA on real objects, as shown in Figure 7. We obtain the point clouds of real objects by the 3D scanner and then sample them to 2048 points as the input to our network. Our network can preserve the right arm of the chair and complete the left arm. The predictions are more accurate and uniform as shown in red boxes. This indicates the generalization ability of our USSPA on real objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Visualization of latent codes on multi-category data. We utilize t-SNE [21] to visualize the latent codes \\( f_{R} \\) and \\( f_{A} \\) of Unpaired and our USSPA on multi-category data. As shown in Figure 8, ten different colors indicate ten categories where dark colors indicate real scene data and light colors indicate artificial data. Figure 8 (c) and (d) show the t-SNE results of latent codes on chair category only. Benefiting from our feature discriminator and autoencoder, the latent codes of real scene data are similar to artificial data which can be shown from the distribution of dark and light points. The dark and light points of Unpaired are separated while ours are fixed together for each category. Benefiting from our classifier-guided discriminator, the distributions of" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Figure 7. Completion on real objects by Unpaired [32], ShapeInv [36] and our USSPA." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.288, + 0.695, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.291, + 0.891, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.436, + 0.892, + 0.506 + ], + "angle": 0, + "content": "Figure 8. Visualization of t-SNE results of the latent codes by Unpaired [32] and our method. (a) and (b) are the results on multi-category, where different colors denote different categories. Dark and light denote real scene data and artificial data respectively. (c) and (d) are the results on the chair category only." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.576 + ], + "angle": 0, + "content": "our latent codes of different categories are separated and clean. For example, as Figure 8 (b) shows, the distances between pink, orange and blue clusters are bigger and the clusters are tighter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.579, + 0.869, + 0.595 + ], + "angle": 0, + "content": "5. Conclusion, Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.827 + ], + "angle": 0, + "content": "We have presented USSPA, an end-to-end unsupervised network for completion of real scene point cloud objects. Benefiting from our carefully designed symmetry learning module and refinement autoencoder, our prediction preserves the symmetrical shape of input with more accurate and uniform points. We also contribute a real scene dataset for accurate evaluation, and extensive experiments and comparisons show the superiority and generalization of our method on different categories of different datasets and real objects, which achieves state-of-the-art performance on unsupervised completion of real scene objects. For those categories without artificial data as assistance, previous works and our method are hard to apply in this situation. This is our limitation. Unsupervised point cloud completion without the assistance of artificial data is our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.836, + 0.66, + 0.853 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant numbers 62032011, 61972194." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13567" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X. Chang, and Matthias Niessner. Scan2cad: Learning cad model alignment in rgb-d scans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jul 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.379 + ], + "angle": 0, + "content": "[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[7] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1359-1368, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[9] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[10] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh. ACM Transactions on Graphics (TOG), 39:126:1 - 126:12, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[11] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7662-7670, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[12] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Bo Li. 3d fully convolutional network for vehicle detection in point cloud. In 2017 IEEE/RSJ International Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "on Intelligent Robots and Systems (IROS), pages 1513-1518. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, 2017 Advances in Neural Information Processing Systems (NIPS), volume 30. Curran Associates, Inc., 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[15] Abhishek Sharma, Oliver Grau, and Mario Fritz. Vconv-dae: Deep volumetric shape learning without object labels. In European Conference on Computer Vision, pages 236-250. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1955-1964, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1716-1725, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[18] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[19] Lyne P Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese. Topnet: Structural point cloud decoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 383-392, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[20] Duc Thanh Nguyen, Binh-Son Hua, Khoi Tran, Quang-Hieu Pham, and Sai-Kit Yeung. A field model for repairing 3d shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5676-5684, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[22] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 790-799, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Learning local displacements for point cloud completion. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Xin Wen, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Cycle4completion: Unpaired point cloud completion using cycle transformation with missing region coding. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13075-13084, 2021. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF Con" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13568" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 1939-1948, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[26] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net: Point cloud completion by learning multi-step point moving paths. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[27] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In The European Conference on Computer Vision (ECCV), August 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[28] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Density-aware chamfer distance as a comprehensive metric for point cloud completion. 11 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[29] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[30] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[31] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[32] Niloy J. Mitra Xuelin Chen, Baoquan Chen. Unpaired point cloud completion on real scans using adversarial training. In 2020 International Conference on Learning Representations 2020 (ICLR), 2020. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.468, + 0.601 + ], + "angle": 0, + "content": "[33] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 206–215, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[34] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.468, + 0.727 + ], + "angle": 0, + "content": "[35] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 International Conference on 3D Vision (3DV), pages 728-737. IEEE, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.468, + 0.811 + ], + "angle": 0, + "content": "[36] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1768-1777, 2021. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.468, + 0.883 + ], + "angle": 0, + "content": "[37] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15890-15899, June 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13569" + } + ] +] \ No newline at end of file diff --git a/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_origin.pdf b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edea86f79fbc03c3e5c51775722af74b014b5f65 --- /dev/null +++ b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/324b2114-0ac8-491b-ac58-edad9a307e8f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff7afc7a3fb6953e4774cb606fee9a320ee879e92fc8963d0de37c934711518b +size 7575785 diff --git a/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/full.md b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..07f374f122ab7979b0ec476f16d1d894743ffb6f --- /dev/null +++ b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/full.md @@ -0,0 +1,311 @@ +# Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion + +Changfeng Ma1, Yinuo Chen1, Pengxiao Guo1, Jie Guo1, Chongjun Wang1, Yanwen Guo1* +1Nanjing University, Nanjing, China +{changfengma, yinuochen, pxguo}@smail.nju.edu.cn +{guojie, chjwang, ywguo}@nju.edu.cn + +# Abstract + +Unsupervised completion of real scene objects is of vital importance but still remains extremely challenging in preserving input shapes, predicting accurate results, and adapting to multi-category data. To solve these problems, we propose in this paper an Unsupervised Symmetric Shape-Preserving Autoencoding Network, termed USSPA, to predict complete point clouds of objects from real scenes. One of our main observations is that many natural and man-made objects exhibit significant symmetries. To accommodate this, we devise a symmetry learning module to learn from those objects and to preserve structural symmetries. Starting from an initial coarse predictor, our autoencoder refines the complete shape with a carefully designed upsampling refinement module. Besides the discriminative process on the latent space, the discriminators of our USSPA also take predicted point clouds as direct guidance, enabling more detailed shape prediction. Clearly different from previous methods which train each category separately, our USSPA can be adapted to the training of multi-category data in one pass through a classifier-guided discriminator, with consistent performance on single category. For more accurate evaluation, we contribute to the community a real scene dataset with paired CAD models as ground truth. Extensive experiments and comparisons demonstrate our superiority and generalization and show that our method achieves state-of-the-art performance on unsupervised completion of real scene objects. + +# 1. Introduction + +As the standard outputs of 3D scanners [12, 32], point clouds are becoming more and more popular [9] which are also the basic data structure in 3D geometry processing [4, 5, 13]. Complete point clouds are hard to obtain due to the nature of the scanning process and object occlusion [35]. Due to the defects of incomplete point clouds on downstream applications such as reconstruction [10], recent works [17, 19, 22, 23, 26, 30, 33, 35] pay more attention + +![](images/af7482530a4e9f27d70c7ce72e1e41ee8a46c6298f17a80089df7905dd3a3d6b.jpg) +Figure 1. Visual comparison of predicted results on real scene data by our USSPA and other works (top) and our complete result on a whole real scene (bottom). (a) shows an example of a real scene partial point cloud of a chair and the complete predictions by Disp3D [23], ShapeInv [36], Unpaired [32] and our method. As shown, our prediction result is more accurate and uniform according to the input, which contains complete arms and legs. (b) and (c) show the original point cloud of a real scene and the complete results of all the objects in this scene. + +to point cloud completion which relies on paired artificial complete point clouds for supervised training to complete partial point clouds. However, these supervised works are difficult to apply in practice because of the great gap between artificial data and real scene data and the inaccessibility to the ground truth of real scene data. Therefore, it is important to complete partial point clouds from real scene in an unsupervised way. + +Recent unsupervised works [24, 32, 36] only require real + +scene partial point clouds and artificial CAD models for unpaired completion utilizing GANs [8] as their fundamental frameworks, most of which need pre-training on artificial data. The main ideas are to transform latent codes from the space of real scene partial data to the space of artificial complete data and then employ the decoder trained on artificial data to predict the complete point cloud. Essentially, these methods make the predictions whose distributions are consistent with the artificial models. Most of them, however, just extract a global feature from the partial input without fully exploiting its geometry information, leading to the prediction severely deviated from the input. And such information actually provides vital clues and constraints for completion. Furthermore, prediction results by these methods usually lack enough geometric details due to the absence of an explicit discriminative process on point clouds. These domain transforming methods are also hard to adapt to multi-category data or other datasets. + +In this paper, we present an unsupervised symmetric shape-preserving autoencoding network, termed USSPA, for the completion of real scene objects, as shown in Figure 2, which is a GAN-based end-to-end network without the requirement of pre-trained weights. Different from previous domain transforming methods which cannot fully leverage existing incomplete models, we argue that the existing partial scanning, which also provides vital clues and constraints for the prediction of the missing part, should be preserved to some extent. To this end, we exploit the symmetries shown in many natural or man-made objects and devise a novel symmetry learning module to generate symmetrical point clouds of existing parts by predicting the symmetric planes. This enables our network to preserve the shapes of input symmetrically, intrinsically facilitating structure completion, as shown in Figure 1. For those parts that can not be directly inferred from inputs, we employ an initial coarse module for an initial prediction first. Starting from the initial guess, we specifically design a refinement autoencoder with an upsampling refinement module for detailed refinement and the local feature grouping for extracting local information, to learn detailed structures of artificial data through the autoencoding process. Benefiting from this, our final prediction is accurate, uniform, and symmetric shape-preserving. Besides the indirect guidance of the feature discriminator on latent space, our point discriminator takes predicted point clouds as direct guidance for generating more accurate shapes. Compared with previous methods which train each category separately, our method can classify the partial point clouds simultaneously through a classifier-guided discriminator when adapted to multi-category data, with consistent performance on the single category. + +To measure the performance of unsupervised completion quantitatively, we build a dataset from ScanNet [5] and + +ShapeNet [2] utilizing the annotations of Scan2CAD [1]. Our dataset contains real scene partial point clouds and paired ground truths that are only used for evaluation in our experiments. Extensive comparisons against previous works on this dataset and the public PCN Dataset [35] show the superiority and generalization of our method which achieves state-of-the-art performance on unsupervised completion of real scene objects. + +Our main contributions are as follows. + +- We propose a novel USSPA for unsupervised real scene point cloud completion whose prediction is accurate, uniform and symmetric shape-preserving. Clearly different from previous works training each category separately, our USSPA can be adapted to the training of multi-category data in one pass by classifying the input simultaneously. +- We propose a novel symmetry learning module and a novel refinement autoencoder. The symmetry learning module preserves input shapes by generating symmetrical point clouds, and the refinement autoencoder learns the detailed information from artificial data to refine the initial guess by an autoencoding process. +- We propose a new evaluation method for obtaining paired ground truths and partial data from artificial and real scene datasets using alignment information, which can be used to more accurately evaluate unsupervised completion of real scene objects. + +# 2. Related Work + +# 2.1. Supervised Point Cloud Completion + +To infer the complete point cloud from an incomplete input, many deep learning methods were introduced in recent years, most of which are supervised. At first, the 3D data were voxelized into occupancy grids or distance fields before fed into convolutional networks [15, 16, 20]. To avoid loss of details and huge memory cost brought by these data representations, Yuan et al. [35] propose PCN which directly operates on raw point clouds. Giancola et al. [7] also leverage a new method for 3D shape completion which directly works on LiDAR point clouds. Later works [11, 19, 22] pay much attention to details refinement and denoise. Disp3D [23] investigates grouping local features to bring improvement to point cloud completion. Taking structures and topology information into consideration, SA-Net [25], GRNet [31], PoinTr [34] and LAKe-Net [17] succeed to utilize the detailed geometry information from input to generate more reasonable complete point clouds. Other works like designing new metrics [28] and combining information of single-view images [37] also make contributions to the point cloud completion task. However, these works are all supervised that are not suitable for practice since ground truths are in lack. + +![](images/2fe9346696598d15132b477864fb0ce92de4d77bc56ce7b3fd53267d9f81b370.jpg) +Figure 2. The overall architecture of our USSPA. Our network has a generator composed of an initial coarse predictor (blue) and a refinement autoencoder (green), and the point and feature discriminators (yellow). The red lines and boxes represent the employed losses. The blue and green lines (bottom) denote the forward processes having the same network architectures with sharing weights. + +![](images/29ab211dd5f971895a3b505959b9a7e101daaec8c276d433151726698b027b70.jpg) + +# 2.2. Unsupervised Point Cloud Completion + +Due to the absence of suitable paired training data for supervised models, some unsupervised models for point cloud completion have emerged in recent years. Unpaired [32] first introduces the GAN to learn the transformation from the latent space of partial point cloud to the latent space of complete point cloud, and then decode the transformed latent code to predict. Zhang et al. [36] employ optimization-based GAN inversion [29] and directly find the complete prediction by optimizing the latent code such that the degradation of the complete point cloud matches the input partial point cloud, which requires fine-tuning the network for each input data. Cycle4 [24] proposes two simultaneous cycle transformations between the latent spaces of complete point cloud and partial ones, facilitating the network's understanding of 3D shapes by building a bidirectional mapping. Wu et al. [27] propose the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. However, the application of pre-trained GANs makes their predictions lack details and difficult to preserve the shape features of the input. For evaluation, Unpaired [32] employs classification accuracy of the complete results which is hard to measure detailed structures. Thus, a more accurate evaluation method is required. + +# 3. Method + +# 3.1. Network Architecture + +Given the partial point cloud $\mathcal{P}_R\in \mathbb{M}_{n_0\times 3}$ from real scene data with distribution $p_r$ and the complete point cloud $\mathcal{P}_A\in \mathbb{M}_{n_0\times 3}$ from artificial data with distribution $p_a$ , the goal of unsupervised real scene point cloud completion is to predict the complete point cloud $Com(\mathcal{P}_R)$ of $\mathcal{P}_R$ such + +that $\operatorname{Com}(\mathcal{P}_R) \sim p_a$ and $\mathcal{P}_R \subset \operatorname{Com}(\mathcal{P}_R)$ . This means that the prediction must be similar to the artificial data and meanwhile retain the shapes of the input partial point cloud. + +We carefully design several modules of our generator $\mathcal{G}$ and discriminators, $\mathcal{D}_p$ and $\mathcal{D}_f$ , for unsupervised completion on real scene data through generative adversarial learning [8] as shown in Figures 2 and 3. As we mentioned above, many natural and man-made objects exhibit significant symmetries. To fully leverage these symmetries and preserve the shape of existing partial input $\mathcal{P}_R$ , we devise a novel symmetry learning module that predicts the symmetric plane and the symmetrical point cloud of $\mathcal{P}_R$ for further refinement. Our refinement autoencoder learns the detailed information of artificial data and refines the details of the initial, but coarse prediction $P_{R_0}$ for uniform and accurate prediction. With the local feature grouping operation, our elaborate upsampling refinement module can utilize the local features for more detailed results. Different from previous works focusing on the discriminative process of latent space, our point discriminator $\mathcal{D}_p$ also takes point clouds as input for direct guidance. + +Initial Coarse Predictor. We first employ an initial coarse predictor for shape-preserving and preliminary prediction. The encoder $E_{R}$ takes the real scene partial point cloud $\mathcal{P}_R$ as input and encodes the global feature $f_{R}^{\prime} = E_{R}(\mathcal{P}_{R})\in \mathbb{M}_{1\times c_{0}}$ . $f_{R}^{\prime}$ is then taken by the decoder $D_{R}$ for the prediction of initial, but coarse complete result $P_R' = D_R(f_R')\in \mathbb{M}_{n_1\times 3}$ . Encoder-decoder architecture is hard to retain shape features on the completion task since there is no explicit complete ground truth. To this end, we devise the symmetry learning module to accommodate symmetrical shape features widely existed in man-made and natural objects, as shown in Figure 3. The decoder $D_{sl}$ pre + +dicts symmetric arguments $\mathbf{A} = D_{sl}(f_R') = [x_A, y_A, z_A]^T$ and then we introduce symmetry operation to calculate the symmetrical point cloud $P_{sym}$ of $\mathcal{P}_R$ by: + +$$ +P _ {s y m} = \boldsymbol {p} - 2 \frac {\boldsymbol {A} \cdot \boldsymbol {p}}{\| \boldsymbol {A} \| ^ {2}} \boldsymbol {A}, \tag {1} +$$ + +where $\pmb{p} = [x,y,z]^T\in \mathcal{P}_R$ $\cdot$ represents the inner product and $||\pmb {x}||$ denotes the module of the vector $\pmb{x}$ . Without the loss of generality, we can safely assume that all the objects are axis-aligned. Under such an assumption, the symmetric plane is perpendicular to the $xz$ -plane and zero-crossing. Thus we set $y_{A} = 0$ . Then $P_{sym}$ $\mathcal{P}_R$ and $P_R^{\prime}$ are fused and resampled through FPS (farthest point sampling) to get $P_{R_0}$ which is the complete and shape-preserving result. + +Refinement Autoencoder. Noisy $\mathcal{P}_R$ and the fusion operation may cause noisy $P_{R_0}$ with uneven point distribution. The initial coarse predictor or the point discriminator only predicts or judges through the global feature, which makes $P_{R_0}$ lack details. Thus, $P_{R_0}$ requires denoising and further refinement on details and point distribution. To this end, we employ the refinement autoencoder with a carefully designed upsampling refinement module. As shown in Figure 2, the blue and green lines of the refinement autoencoder represent the same forward process having the same network architectures with sharing weights. In this way, our autoencoder extracts the detailed information of artificial data and refines the details of $P_{R_0}$ for more accurate and uniform prediction. + +The encoder and decoder $E_{A},D_{A}$ generate the skeleton point cloud $P_{R_1 / A_1}\in \mathbb{M}_{n_2\times 3}$ of input point cloud. We set $n_2$ to a small number for easier prediction of $P_{R_1 / A_1}$ . And the density loss is faster to calculate with fewer points. We then employ upsampling refinement module to upsample $P_{R_1 / A_1}$ and add details to the skeleton for final prediction $P_{R_2 / A_2}$ . As shown in Figure 3, we utilize wight-shared MLP [3] to obtain the point features and max-pooling operation on point features to obtain the global feature of the point cloud. We also concatenate the point and global features to fuse the information. Since wight-shared MLP only extracts the feature of one point, lacking the local information, we also employ the local feature grouping to get the local information around each point. For point $p_0$ , $f_{group}\in \mathbb{M}_{n_2\times (c_1\cdot m)}$ are the features grouped from $f_{input}\in \mathbb{M}_{n_2\times c_1}$ corresponding to the points $p$ that $\| p - p_0\| \leq r$ , where $m$ is the feature number of each group. After applying max-pooling operation on $m$ features of each group, $f_{local}\in \mathbb{M}_{n_2\times c_1}$ are concatenated with $f_{input}$ to generate the output $f_{output}\in \mathbb{M}_{n_2\times 2c_1}$ . We utilize a weight-shared MLP for upsampling that takes fused features composed of $P_{R_1 / A_1}$ , local and global features, and predict $k$ shifts for each point. Then, each point splits into $k$ points by adding the predicted shifts to its coordinate whose result is $P_{R_2 / A_2}\in \mathbb{M}_{kn_2\times 3}$ . + +![](images/6df62d8a68bf4a85c11c6c19a6e2f60514742af0ef23556f6afefaee378b946d.jpg) + +![](images/90846916c9692d15b10200c537e9d1d31275caf446df0ff45d2474a0723554ac.jpg) +Upsampling Refinement Module +Local Feature Grouping + +![](images/8cae283e51f0ca5bc5d18d4521552a9c672ce864e735014bf914f3a886644a2b.jpg) +Symmetry Learning Module + +![](images/efe0b02cc3fd5d45a1089578c90f3d8cff7a594ce3c219542db67ecf9e35bacf.jpg) +Concatenate + +![](images/923708c274ca2ebb767d9b0811e99451a6afa3dda58de7e1f890b7a4d3481246.jpg) +Max-Pooling + +![](images/a938c514f554c0d23359bd5d2c0ad9391f993c3e4b86512f87f6675443a67458.jpg) +Local Feature Grouping + +![](images/f0877c0040db32685acef03a3bef9f5f2338a32a16dc7d1875b1e391dd32c479.jpg) +Element Addition +Figure 3. Detailed structures of our symmetry learning module (red), upsampling refinement module (yellow), and local feature sampling (blue). + +Point and Feature Discriminator. We employ the point discriminator $\mathcal{D}_p$ to directly guide the generation of $P_{R_0}$ and $P_{R_2}$ comparing with the artificial complete data $\mathcal{P}_A$ for unsupervised completion. We also guide the generation of latent code of autoencoder through the feature discriminator $\mathcal{D}_f$ such that $f_R$ and $f_A$ have the same distribution for more accurate prediction. Our discriminators predict the probability that input point cloud $P_i$ or feature $f_i$ is real data. For multi-category data, we replace the probability-guided discriminator with a classifier-guided discriminator which outputs the category label of input rather than probability. We utilize label 0 to represent fake data and labels $l > 0$ to denote real data of different categories such as the chair, table, etc. For fake data, the discriminator tends to output label 0, while the generator tends to make the discriminator output label $l > 0$ . And for real data with label $l$ , the discriminator tends to output $l$ correctly. In this way, our USSPA can unsupervisedly classify the real scene partial point cloud when completing multi-category data. + +# 3.2. Optimization + +The abbreviate optimization goal of our USSPA for unsupervised learning is: + +$$ +\min _ {\mathcal {G}} \max _ {\mathcal {D} _ {p}, \mathcal {D} _ {f}} \mathbb {E} _ {P _ {i} \sim p _ {a}} \log p + \mathbb {E} _ {P _ {i} \sim p _ {r}} \log (1 - p), \tag {2} +$$ + +where $p$ is the probability of real data. By playing the maxmin game with $\mathcal{D}_p$ and $\mathcal{D}_f$ , our generator $\mathcal{G}$ tends to predict point clouds similar to artificial data. We employ + +Table 1. Point cloud completion comparison on our dataset in terms of L1 Chamfer Distance $cd^{l1} \times 10^2$ (lower is better). All the methods are trained with our dataset and the unsupervised methods including ours are trained on single category data. Boldface denotes the best among unsupervised methods in Tables 1, 2 and 3. + +
MethodAVGchairtabletrash binTVcabinetbookshelfsofalampbedtub
Sup.PoinTr [34]14.3713.6512.5215.2612.6917.3213.9912.3617.0515.1313.77
Disp3D [23]7.786.248.207.127.1210.366.945.6014.036.905.32
TopNet [19]7.076.395.797.406.268.377.025.948.507.817.25
Unsup. OptBasedShapeInv [36]21.3917.9717.2833.5115.6926.2625.5114.2816.6932.3314.43
Unsup.Unpaired [32]10.478.417.5212.086.7217.459.956.9219.3610.046.22
Cycle4 [24]11.539.1111.3511.938.4015.4712.5110.6312.2515.737.92
Ours8.568.227.6810.367.6610.777.846.1411.938.206.75
+ +Chamfer Distance(CD) [6] to measure the difference between point clouds. The $cd_{P_1 \to P_2}^l$ indicates the single direction Chamfer Distance with L2-norm from point cloud $P_1$ to $P_2$ : + +$$ +c d _ {P _ {1} \rightarrow P _ {2}} ^ {l 2} = \frac {1}{| P _ {1} |} \sum_ {\boldsymbol {x} \in P _ {1}} \min _ {\boldsymbol {y} \in P _ {2}} \| \boldsymbol {x} - \boldsymbol {y} \| ^ {2}, \tag {3} +$$ + +which guides $P_{1}$ to be a part of $P_{2}$ . The $cd_{P_1\leftrightarrow P_2}^{l^2}$ between $P_{1}$ and $P_{2}$ is: + +$$ +c d _ {P _ {1} \leftrightarrow P _ {2}} ^ {l 2} = c d _ {P _ {1} \rightarrow P _ {2}} ^ {l 2} + c d _ {P _ {2} \rightarrow P _ {1}} ^ {l 2}, \tag {4} +$$ + +which guides $P_{1}$ and $P_{2}$ to be identical. As shown in Figure 2, to restrict the generation of $P_{R}^{\prime}$ , we employ single direction Chamfer Distance from $\mathcal{P}_R$ to $P_{R}^{\prime}$ . We also utilize Chamfer Distance between $P_{R_1 / A_1}, P_{R_2 / A_2}$ and $P_{R_0}, \mathcal{P}_A$ for supervision of autoencoder. For uniform prediction, we employ density loss by first calculating the local density of $\boldsymbol{x} \in P$ : + +$$ +l d (\boldsymbol {x}, P) = \frac {1}{k _ {d}} \sum_ {k = 1} ^ {k _ {d}} \min _ {\boldsymbol {y} \in P} ^ {k} | | \boldsymbol {x} - \boldsymbol {y} | |, \tag {5} +$$ + +and then calculating the standard deviation for all $x$ : + +$$ +d l (P) = \sqrt {\frac {1}{| P |} \sum_ {\boldsymbol {x} \in P} \left(l d (\boldsymbol {x} , P) - \overline {{l d (\boldsymbol {x} , P)}}\right) ^ {2}}, \tag {6} +$$ + +where $\min$ indicates the $k$ -th-min element. The smaller the $ld(\pmb{x}, P)$ is, the bigger the local density of $\pmb{x}$ is. + +In practice, we alternately optimize $\mathcal{G}$ with loss $\mathcal{L}_{\mathcal{G}}$ : + +$$ +\mathcal {L} _ {\mathcal {G}} = \alpha_ {1} \mathcal {L} _ {F \rightarrow R} + \alpha_ {2} \mathcal {L} _ {P} + \alpha_ {3} \mathcal {L} _ {d}, \tag {7} +$$ + +where + +$$ +\begin{array}{l} \mathcal {L} _ {F \rightarrow R} = - \left(\log \mathcal {D} _ {p} \left(P _ {R _ {0}}\right) + \log \mathcal {D} _ {p} \left(P _ {R _ {2}}\right)\right. \tag {8} \\ + \log \mathcal {D} _ {f} (f _ {R})) \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {P} = c d _ {P _ {R _ {1}} \leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {P _ {R _ {2}} \leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {\mathcal {P} _ {R} \rightarrow P _ {R} ^ {\prime}} ^ {l 2} \tag {9} \\ + c d _ {P _ {A _ {1}} \leftrightarrow \mathcal {P} _ {A}} ^ {l ^ {2}} + c d _ {P _ {A _ {2}} \leftrightarrow \mathcal {P} _ {A}} ^ {l ^ {2}}, \\ \end{array} +$$ + +$$ +\mathcal {L} _ {d} = d l \left(P _ {R _ {1}}\right) + d l \left(P _ {A _ {1}}\right) \tag {10} +$$ + +and $\mathcal{D}$ with loss $\mathcal{L}_{\mathcal{D}}$ + +$$ +\mathcal {L} _ {\mathcal {D}} = \alpha_ {4} \mathcal {L} _ {F} + \alpha_ {5} \mathcal {L} _ {R}, \tag {11} +$$ + +where + +$$ +\begin{array}{l} \mathcal {L} _ {F} = - \left(\log \left[ 1 - \mathcal {D} _ {p} \left(P _ {R _ {0}}\right) \right] + \log \left[ 1 - \mathcal {D} _ {p} \left(P _ {R _ {2}}\right) \right] \right. \tag {12} \\ \left. + \log \left[ 1 - \mathcal {D} _ {f} \left(f _ {R}\right) \right]\right), \\ \end{array} +$$ + +$$ +\mathcal {L} _ {R} = - \left(\log \mathcal {D} _ {p} \left(\mathcal {P} _ {A}\right) + \log \mathcal {D} _ {f} \left(f _ {A}\right)\right), \tag {13} +$$ + +and here $\alpha_{1\sim 5}$ are the weights used for balancing the influences between each term. And for multi-category data, we use cross-entropy to calculate the loss between predicted labels and ground truth labels. + +# 4. Experiments + +# 4.1. Dataset and Implementation Details + +Different from the Unpaired [32] which utilizes a classifier to evaluate the completion results of real scene data, we build a new dataset from ScanNet [5], ShapeNet [2] and Scan2CAD [1], containing paired partial and complete point clouds for more accurate evaluation. Scan2CAD contains annotations that align the CAD models of ShapeNet to the real scenes of ScanNet. We first extract the CAD models from ShapeNet and partial point clouds from ScanNet which are paired according to the correspondences offered by Scan2CAD, and then uniformly sample the CAD models to get complete point clouds. Finally, we transform the paired partial and complete point clouds with the transform matrix given by Scan2CAD to normalize their face direction, scale, and position. We select ten categories including the chair, table, etc., to establish the dataset and randomly split the data into training, validation, and testing sets. In our experiments, these ground truths are only used for the evaluation of the complete results. + +![](images/50ebc953567d54ff4c32e2be7a363499d0e024e27777648ddb9aa67791a13ff5.jpg) +Figure 4. Visualization of completion results by supervised works, including TopNet [19], Disp3D [23], unsupervised works Cycle4 [24], Unpaired [32], optimization-based work ShapeInv [36] and our method. From top to bottom: table, bed, cabinet, and TV. + +We utilize PyTorch to implement our USSPA. In Section 3.1, the point numbers $n_{0\sim 2}$ are set to 2048, 512 and 512. The channel numbers $c_{0\sim 1}$ are set to 512 and 256. The feature number $m$ , the radius $r$ of local feature grouping, and the split number of upsampling $k$ are set to 32, 0.5, and 4. $k_d$ of density loss in Section 3.2 is set to 16. The weights $\alpha_{1\sim 5}$ of losses are set to 1, 100, 7.5, 0.5 and 0.5. We employ Adam optimizer whose learning rate is $1.0 \times 10^{-5}$ . The batch size is set to 4, and the maximum epoch of training is 240 to 960 according to the size of the dataset. Training our USSPA takes about 20 hours with a GTX 2080Ti GPU. We will make our dataset and codes public in the future. + +# 4.2. Comparison + +We compare our USSPA against existing representative point cloud completion methods including supervised [19, 23, 34], unsupervised [24, 32] and optimization-based [36] methods on our dataset and the PCN Dataset [35]. The metrics used for comparison include pre-point L1 Chamfer Distance $cd^{l1}$ [6] and F-Scores $F_{score}^{1\%}$ , $F_{score}^{0.1\%}$ [18]. + +As shown in Table 1, we compare these methods on our dataset with 10 categories where all methods are trained on the training set and evaluated on the testing set. All the unsupervised methods are trained with single category data for each class. The average $cd^{l1}$ of our USSPA is $18.2\%$ lower than Unpaired and the average $F_{score}^{0.1\%}$ is $30\%$ higher than Unpaired as shown in Table 2, indicating the superior performance of our method on completion of real scene objects. We also mix up all 10 categories as the multi-category data to train the methods for evaluation of the generalization + +Table 2. Comparison on our dataset trained with single category and multi-category data in terms of L1 Chamfer Distance $cd^{l1} \times 10^2$ (lower is better) and F-scores $F_{score}^{0.1\%} \times 10^2$ , $F_{score}^{1\%} \times 10^2$ (higher is better). "Ours" and "Ours(classifier)" denote our method with probability-guided and classifier-guided discriminators, respectively. + +
Methodsingle categorymulti-category
↑F0.1% score↑F1% score↓cdl1↑F0.1% score↑F1% score
PoinTr [34]--14.3718.3580.41
Disp3D [23]--7.7830.2978.26
TopNet [19]--7.0712.3380.37
ShapeInv [36]15.5866.5319.3516.9869.66
Unpaired [32]12.2064.3310.1210.8666.68
Cycle4 [24]9.9860.1412.008.6156.57
Ours17.4973.418.9616.8872.31
Ours(classifier)--8.7617.1273.75
+ +ability. As shown in Table 2, our USSPA keeps the same performance by replacing the probability-guided discriminators with classifier-guided discriminators and has better generalization ability on multi-category than other unsupervised networks. Further comparisons show the generalization capability of our USSPA on different datasets. We train unsupervised models with ScanNet and then test them on the PCN Dataset which is a point cloud completion dataset built upon artificial data. As shown in Table 3, our method also performs better than other unsupervised methods, the average $cd^{l1}$ of which is $15\%$ lower than the SOTA method. Table 4 shows the classification accuracy of PN2 [14] and + +Table 3. Point cloud completion comparison on PCN Dataset in terms of L1 Chamfer Distance $cd^{l1} \times 10^2$ (lower is better). All unsupervised models are trained with ScanNet. + +
MethodAVGchairtablecabinetsofalamp
PoinTr [34]5.495.615.686.085.674.44
Disp3D [23]2.512.422.302.382.443.00
TopNet [19]5.926.345.456.065.805.95
ShapeInv [36]19.0523.1815.6617.1422.8516.40
Unpaired [32]14.8712.878.1414.3018.2320.82
Cycle4 [24]17.6014.2515.7321.0621.5415.40
Ours12.6313.529.668.8915.5115.57
+ +Table 4. Point cloud classification in terms of accuracy. "AVG" denotes the average accuracy on 10 categories of our dataset. Only four categories are shown due to width limitation. + +
MethodAVGchairtabletrash binsofa
PN2 [14]58.6%91.0%94.0%17.2%48.9%
Ours69.8%91.0%91.0%51.6%93.3%
+ +our method. Our method can also unsupervisedly classify the real scene partial point cloud while completing. The accuracy of our USSPA is $11.2\%$ higher than PN2 on average. + +Figure 4 shows the qualitative results on 4 categories. The complete point clouds predicted by our USSPA exhibit more accurate shapes, such as the leg of the table and the division plate of the bookshelf, which are hard to generate even for supervised methods. Our method can successfully generate the circular top of the table. By comparison, other unsupervised works generate a square one. Our predictions are also more uniform with fewer noises. Benefiting from our shape-preserving method, our predicted TV retains the support part from the input point cloud. Our symmetry learning module also plays an important role in completing the pillars of the bed by copying the pillars from the input symmetrically. Visualization of our predicted point clouds shows the superiority of our methods in the generation of detailed completion on real scene data. + +# 4.3. Ablation Study + +Ablation study on network architecture. We conduct experiments on ablated models for evaluation of the necessity of important structures of our USSPA, including the symmetry learning module, the refinement autoencoder, the local feature grouping, the point and feature discriminators, the single direction Chamfer Distance, and the upsampling refinement module. As shown in Table 5, these modules are removed from our full network separately. We downsample the final prediction of our USSPA to 512 points for a fair comparison to the ablated model without the upsampling refinement module. We train and utilize $F_{score}^{0.1\%}$ to evaluate these ablated models with our dataset on the chair category, and the comparison indicates the necessity of these modules + +Table 5. Ablation study in terms of $F_{score}^{0.1\%}$ , where “-” represents removing a module from the full network respectively and “512” denotes the point number. + +
ModelF0.1% score
- Symmetry Learning Module16.43
- Refinement Autoencoder17.08
- Local Feature Grouping15.72
- Feature Discriminator17.11
- Point Discriminator15.68
- Single Direction Chamfer Distance17.84
Ours (full)18.43
- Upsampling Refinement (512)5.60
Ours (full) (512)6.46
+ +![](images/3ad04e4de23b29a7956529e38d60b2ba392bfe4dfde3913da6fab02cdf01b98b.jpg) +Figure 5. Visualization of the symmetric planes and predictions by USSPA without the symmetry learning module and full network. Blue and orange denote original and mirrored points, separately. + +which play significant roles in shape-preserving, extracting, and utilizing detailed features, upsampling and refining for more accurate point cloud completion. + +Symmetry learning module. Our symmetry learning module learns different symmetric planes according to the input point clouds as shown in Figures 5(a) and 5(b). The predicted symmetric plane is xy-plane for the chair and yz-plane for the table. Through different symmetric planes, our network can fully leverage the symmetrical structures of input point clouds, such as the arm of the chair and the legs of the table, which exist on only one side. Figure 5(c) shows the prediction $P_{R_2}$ of our network without the symmetry learning module, and Figure 5(d) shows the result of our full network. Figure 5(e) shows the input point cloud whose left arm is missing while the right arm is complete. Without the symmetry learning module, the left arm of the prediction is incomplete as shown in the red box. And benefiting from such a module, our full network can learn the symmetric structures, and generate a complete left arm. + +Density loss. We visualize the local density of points as shown in Figure 6. The warmer the color of the point is, the bigger its local density (described in equation 5) is. (a) is the prediction without the supervision of the density loss, and (b) is the prediction supervised by the density loss. Without the supervision of density loss, the network prefers generating many points in the same location as shown in the blue circle to minimize the Chamfer Distance, which makes the predicted point clouds contain fewer points describing + +Table 6. Comparison of Unpaired and our USSPA in terms of density loss, where "ours (w/o $dl$ )" indicates our USSPA without the supervision of the density loss. The results show that our predicted point clouds are more uniform. + +
Unpairedours (w/o dl)ours
Density Loss0.09570.12380.0149
+ +![](images/dc48897284b2cd60fdb4dea05a44266c4b191d525270a63d1f40eec626af6d9f.jpg) +Figure 6. Visualization of the local density of point clouds, where the warmth of the color indicates the values of local densities. The warmer the color is, the bigger the local density is. (a) and (b) are the comparison of the ablation experiment on the density loss. (c) and (d) are the examples of $P_{R_0}$ and $P_{R_2}$ . The density loss and the point number of each point cloud are shown at its bottom. + +![](images/71ec2aea586c7429d4c70b60810b95212742fa774dc044fcba4a7fe544000308.jpg) + +![](images/014284f2971eb8e5dc8dae2871dbfc862107835bfd3946161fe642da7ac99679.jpg) + +![](images/7134bc6d3af70df8f1d07f6cc0dc1dd3db7d449322a6644677d1f34c5d9721e9.jpg) + +details such as chair arms and legs. Table 6 compares Unpaired and our USSPA on density loss indicating that our predicted point clouds are more uniform. + +Autoencoding refiner. As shown in Figure 6, (c) and (d) are the examples of $P_{R_0}$ and $P_{R_2}$ which is the refinement result of the autoencoding refiner on $P_{R_0}$ . The refiner removes the noises from input shown in blue boxes and generates more points on legs as shown in green boxes, which makes the whole point cloud more uniform and accurate with less density loss. + +Real scene data. We also test our USSPA on real objects, as shown in Figure 7. We obtain the point clouds of real objects by the 3D scanner and then sample them to 2048 points as the input to our network. Our network can preserve the right arm of the chair and complete the left arm. The predictions are more accurate and uniform as shown in red boxes. This indicates the generalization ability of our USSPA on real objects. + +Visualization of latent codes on multi-category data. We utilize t-SNE [21] to visualize the latent codes $f_{R}$ and $f_{A}$ of Unpaired and our USSPA on multi-category data. As shown in Figure 8, ten different colors indicate ten categories where dark colors indicate real scene data and light colors indicate artificial data. Figure 8 (c) and (d) show the t-SNE results of latent codes on chair category only. Benefiting from our feature discriminator and autoencoder, the latent codes of real scene data are similar to artificial data which can be shown from the distribution of dark and light points. The dark and light points of Unpaired are separated while ours are fixed together for each category. Benefiting from our classifier-guided discriminator, the distributions of + +![](images/7cf0cecbbab6595a26b95c0353320bfd95fadf19f47261ad6018f23eb772c395.jpg) +Figure 7. Completion on real objects by Unpaired [32], ShapeInv [36] and our USSPA. + +![](images/34b72b68879acc150a93e2ced5a68c5095a513b1861ec8a7c726c5ea4bd38f8c.jpg) +Figure 8. Visualization of t-SNE results of the latent codes by Unpaired [32] and our method. (a) and (b) are the results on multi-category, where different colors denote different categories. Dark and light denote real scene data and artificial data respectively. (c) and (d) are the results on the chair category only. + +![](images/48e7821b250ad1cd936a453b8c8bc9080ae71a957b7c5e1d1387e8666091f951.jpg) + +our latent codes of different categories are separated and clean. For example, as Figure 8 (b) shows, the distances between pink, orange and blue clusters are bigger and the clusters are tighter. + +# 5. Conclusion, Limitation and Future Work + +We have presented USSPA, an end-to-end unsupervised network for completion of real scene point cloud objects. Benefiting from our carefully designed symmetry learning module and refinement autoencoder, our prediction preserves the symmetrical shape of input with more accurate and uniform points. We also contribute a real scene dataset for accurate evaluation, and extensive experiments and comparisons show the superiority and generalization of our method on different categories of different datasets and real objects, which achieves state-of-the-art performance on unsupervised completion of real scene objects. For those categories without artificial data as assistance, previous works and our method are hard to apply in this situation. This is our limitation. Unsupervised point cloud completion without the assistance of artificial data is our future work. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grant numbers 62032011, 61972194. + +# References + +[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X. Chang, and Matthias Niessner. Scan2cad: Learning cad model alignment in rgb-d scans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2, 5 +[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5 +[3] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jul 2017. 4 +[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 1, 2, 5 +[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5, 6 +[7] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1359-1368, 2019. 2 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 3 +[9] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 2020. 1 +[10] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh. ACM Transactions on Graphics (TOG), 39:126:1 - 126:12, 2020. 1 +[11] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7662-7670, 2020. 2 +[12] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 1 +[13] Bo Li. 3d fully convolutional network for vehicle detection in point cloud. In 2017 IEEE/RSJ International Conference + +on Intelligent Robots and Systems (IROS), pages 1513-1518. IEEE, 2017. 1 +[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, 2017 Advances in Neural Information Processing Systems (NIPS), volume 30. Curran Associates, Inc., 2017. 6, 7 +[15] Abhishek Sharma, Oliver Grau, and Mario Fritz. Vconv-dae: Deep volumetric shape learning without object labels. In European Conference on Computer Vision, pages 236-250. Springer, 2016. 2 +[16] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1955-1964, 2018. 2 +[17] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1716-1725, 2022. 1, 2 +[18] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 6 +[19] Lyne P Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese. Topnet: Structural point cloud decoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 383-392, 2019. 1, 2, 5, 6, 7 +[20] Duc Thanh Nguyen, Binh-Son Hua, Khoi Tran, Quang-Hieu Pham, and Sai-Kit Yeung. A field model for repairing 3d shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5676-5684, 2016. 2 +[21] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 8 +[22] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 790-799, 2020. 1, 2 +[23] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Learning local displacements for point cloud completion. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022. 1, 2, 5, 6, 7 +[24] Xin Wen, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Cycle4completion: Unpaired point cloud completion using cycle transformation with missing region coding. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13075-13084, 2021. 1, 3, 5, 6, 7 +[25] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF Con + +ference on Computer Vision and Pattern Recognition, pages 1939-1948, 2020. 2 +[26] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net: Point cloud completion by learning multi-step point moving paths. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[27] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In The European Conference on Computer Vision (ECCV), August 2020. 3 +[28] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Density-aware chamfer distance as a comprehensive metric for point cloud completion. 11 2021. 2 +[29] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2022. 3 +[30] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1 +[31] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[32] Niloy J. Mitra Xuelin Chen, Baoquan Chen. Unpaired point cloud completion on real scans using adversarial training. In 2020 International Conference on Learning Representations 2020 (ICLR), 2020. 1, 3, 5, 6, 7, 8 +[33] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 206–215, 2018. 1 +[34] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 2, 5, 6, 7 +[35] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 International Conference on 3D Vision (3DV), pages 728-737. IEEE, 2018. 1, 2, 6 +[36] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1768-1777, 2021. 1, 3, 5, 6, 7, 8 +[37] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15890-15899, June 2021. 2 \ No newline at end of file diff --git a/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/images.zip b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..91a9eb5fb6bcbaeeec22f81c1a9ecebea9e7dde4 --- /dev/null +++ b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb11453d2aba2ccc1072a0c7897645372ebb103c54d26f37a5c878ade1c89928 +size 882343 diff --git a/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/layout.json b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a74640a3bb2bc1700fde924e13e41b58afc2e07b --- /dev/null +++ b/2023/Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion/layout.json @@ -0,0 +1,9174 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 140 + ], + "type": "text", + "content": "Symmetric Shape-Preserving Autoencoder for Unsupervised Real Scene Point Cloud Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 518, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 518, + 217 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 518, + 217 + ], + "type": "text", + "content": "Changfeng Ma1, Yinuo Chen1, Pengxiao Guo1, Jie Guo1, Chongjun Wang1, Yanwen Guo1* \n1Nanjing University, Nanjing, China \n{changfengma, yinuochen, pxguo}@smail.nju.edu.cn \n{guojie, chjwang, ywguo}@nju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 245, + 192, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 256 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 256 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 257, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 579 + ], + "type": "text", + "content": "Unsupervised completion of real scene objects is of vital importance but still remains extremely challenging in preserving input shapes, predicting accurate results, and adapting to multi-category data. To solve these problems, we propose in this paper an Unsupervised Symmetric Shape-Preserving Autoencoding Network, termed USSPA, to predict complete point clouds of objects from real scenes. One of our main observations is that many natural and man-made objects exhibit significant symmetries. To accommodate this, we devise a symmetry learning module to learn from those objects and to preserve structural symmetries. Starting from an initial coarse predictor, our autoencoder refines the complete shape with a carefully designed upsampling refinement module. Besides the discriminative process on the latent space, the discriminators of our USSPA also take predicted point clouds as direct guidance, enabling more detailed shape prediction. Clearly different from previous methods which train each category separately, our USSPA can be adapted to the training of multi-category data in one pass through a classifier-guided discriminator, with consistent performance on single category. For more accurate evaluation, we contribute to the community a real scene dataset with paired CAD models as ground truth. Extensive experiments and comparisons demonstrate our superiority and generalization and show that our method achieves state-of-the-art performance on unsupervised completion of real scene objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 583, + 128, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 128, + 595 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 128, + 595 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 597, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 694 + ], + "type": "text", + "content": "As the standard outputs of 3D scanners [12, 32], point clouds are becoming more and more popular [9] which are also the basic data structure in 3D geometry processing [4, 5, 13]. Complete point clouds are hard to obtain due to the nature of the scanning process and object occlusion [35]. Due to the defects of incomplete point clouds on downstream applications such as reconstruction [10], recent works [17, 19, 22, 23, 26, 30, 33, 35] pay more attention" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 243, + 538, + 498 + ], + "blocks": [ + { + "bbox": [ + 313, + 243, + 538, + 498 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 538, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 538, + 498 + ], + "type": "image", + "image_path": "af7482530a4e9f27d70c7ce72e1e41ee8a46c6298f17a80089df7905dd3a3d6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 500, + 547, + 598 + ], + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 598 + ], + "type": "text", + "content": "Figure 1. Visual comparison of predicted results on real scene data by our USSPA and other works (top) and our complete result on a whole real scene (bottom). (a) shows an example of a real scene partial point cloud of a chair and the complete predictions by Disp3D [23], ShapeInv [36], Unpaired [32] and our method. As shown, our prediction result is more accurate and uniform according to the input, which contains complete arms and legs. (b) and (c) show the original point cloud of a real scene and the complete results of all the objects in this scene." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 603, + 546, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 546, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 546, + 699 + ], + "type": "text", + "content": "to point cloud completion which relies on paired artificial complete point clouds for supervised training to complete partial point clouds. However, these supervised works are difficult to apply in practice because of the great gap between artificial data and real scene data and the inaccessibility to the ground truth of real scene data. Therefore, it is important to complete partial point clouds from real scene in an unsupervised way." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "Recent unsupervised works [24, 32, 36] only require real" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13560" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "scene partial point clouds and artificial CAD models for unpaired completion utilizing GANs [8] as their fundamental frameworks, most of which need pre-training on artificial data. The main ideas are to transform latent codes from the space of real scene partial data to the space of artificial complete data and then employ the decoder trained on artificial data to predict the complete point cloud. Essentially, these methods make the predictions whose distributions are consistent with the artificial models. Most of them, however, just extract a global feature from the partial input without fully exploiting its geometry information, leading to the prediction severely deviated from the input. And such information actually provides vital clues and constraints for completion. Furthermore, prediction results by these methods usually lack enough geometric details due to the absence of an explicit discriminative process on point clouds. These domain transforming methods are also hard to adapt to multi-category data or other datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 687 + ], + "type": "text", + "content": "In this paper, we present an unsupervised symmetric shape-preserving autoencoding network, termed USSPA, for the completion of real scene objects, as shown in Figure 2, which is a GAN-based end-to-end network without the requirement of pre-trained weights. Different from previous domain transforming methods which cannot fully leverage existing incomplete models, we argue that the existing partial scanning, which also provides vital clues and constraints for the prediction of the missing part, should be preserved to some extent. To this end, we exploit the symmetries shown in many natural or man-made objects and devise a novel symmetry learning module to generate symmetrical point clouds of existing parts by predicting the symmetric planes. This enables our network to preserve the shapes of input symmetrically, intrinsically facilitating structure completion, as shown in Figure 1. For those parts that can not be directly inferred from inputs, we employ an initial coarse module for an initial prediction first. Starting from the initial guess, we specifically design a refinement autoencoder with an upsampling refinement module for detailed refinement and the local feature grouping for extracting local information, to learn detailed structures of artificial data through the autoencoding process. Benefiting from this, our final prediction is accurate, uniform, and symmetric shape-preserving. Besides the indirect guidance of the feature discriminator on latent space, our point discriminator takes predicted point clouds as direct guidance for generating more accurate shapes. Compared with previous methods which train each category separately, our method can classify the partial point clouds simultaneously through a classifier-guided discriminator when adapted to multi-category data, with consistent performance on the single category." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "To measure the performance of unsupervised completion quantitatively, we build a dataset from ScanNet [5] and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "ShapeNet [2] utilizing the annotations of Scan2CAD [1]. Our dataset contains real scene partial point clouds and paired ground truths that are only used for evaluation in our experiments. Extensive comparisons against previous works on this dataset and the public PCN Dataset [35] show the superiority and generalization of our method which achieves state-of-the-art performance on unsupervised completion of real scene objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 170, + 473, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 170, + 473, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 473, + 180 + ], + "type": "text", + "content": "Our main contributions are as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 186, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 186, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 545, + 272 + ], + "type": "text", + "content": "- We propose a novel USSPA for unsupervised real scene point cloud completion whose prediction is accurate, uniform and symmetric shape-preserving. Clearly different from previous works training each category separately, our USSPA can be adapted to the training of multi-category data in one pass by classifying the input simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "type": "text", + "content": "- We propose a novel symmetry learning module and a novel refinement autoencoder. The symmetry learning module preserves input shapes by generating symmetrical point clouds, and the refinement autoencoder learns the detailed information from artificial data to refine the initial guess by an autoencoding process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 347, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 347, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 347, + 545, + 407 + ], + "type": "text", + "content": "- We propose a new evaluation method for obtaining paired ground truths and partial data from artificial and real scene datasets using alignment information, which can be used to more accurately evaluate unsupervised completion of real scene objects." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 436, + 497, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 497, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 497, + 449 + ], + "type": "text", + "content": "2.1. Supervised Point Cloud Completion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 453, + 547, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 716 + ], + "type": "text", + "content": "To infer the complete point cloud from an incomplete input, many deep learning methods were introduced in recent years, most of which are supervised. At first, the 3D data were voxelized into occupancy grids or distance fields before fed into convolutional networks [15, 16, 20]. To avoid loss of details and huge memory cost brought by these data representations, Yuan et al. [35] propose PCN which directly operates on raw point clouds. Giancola et al. [7] also leverage a new method for 3D shape completion which directly works on LiDAR point clouds. Later works [11, 19, 22] pay much attention to details refinement and denoise. Disp3D [23] investigates grouping local features to bring improvement to point cloud completion. Taking structures and topology information into consideration, SA-Net [25], GRNet [31], PoinTr [34] and LAKe-Net [17] succeed to utilize the detailed geometry information from input to generate more reasonable complete point clouds. Other works like designing new metrics [28] and combining information of single-view images [37] also make contributions to the point cloud completion task. However, these works are all supervised that are not suitable for practice since ground truths are in lack." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13561" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 329, + 274 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 329, + 274 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 329, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 329, + 274 + ], + "type": "image", + "image_path": "2fe9346696598d15132b477864fb0ce92de4d77bc56ce7b3fd53267d9f81b370.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 547, + 310 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our USSPA. Our network has a generator composed of an initial coarse predictor (blue) and a refinement autoencoder (green), and the point and feature discriminators (yellow). The red lines and boxes represent the employed losses. The blue and green lines (bottom) denote the forward processes having the same network architectures with sharing weights." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 533, + 274 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 533, + 274 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 533, + 274 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 533, + 274 + ], + "type": "image", + "image_path": "29ab211dd5f971895a3b505959b9a7e101daaec8c276d433151726698b027b70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 313, + 251, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 251, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 251, + 326 + ], + "type": "text", + "content": "2.2. Unsupervised Point Cloud Completion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 328, + 290, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 290, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 290, + 616 + ], + "type": "text", + "content": "Due to the absence of suitable paired training data for supervised models, some unsupervised models for point cloud completion have emerged in recent years. Unpaired [32] first introduces the GAN to learn the transformation from the latent space of partial point cloud to the latent space of complete point cloud, and then decode the transformed latent code to predict. Zhang et al. [36] employ optimization-based GAN inversion [29] and directly find the complete prediction by optimizing the latent code such that the degradation of the complete point cloud matches the input partial point cloud, which requires fine-tuning the network for each input data. Cycle4 [24] proposes two simultaneous cycle transformations between the latent spaces of complete point cloud and partial ones, facilitating the network's understanding of 3D shapes by building a bidirectional mapping. Wu et al. [27] propose the first multimodal shape completion method that completes the partial shape via conditional generative modeling, without requiring paired training data. However, the application of pre-trained GANs makes their predictions lack details and difficult to preserve the shape features of the input. For evaluation, Unpaired [32] employs classification accuracy of the complete results which is hard to measure detailed structures. Thus, a more accurate evaluation method is required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 620, + 104, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 104, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 104, + 632 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 638, + 173, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 173, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 173, + 649 + ], + "type": "text", + "content": "3.1. Network Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Given the partial point cloud " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R\\in \\mathbb{M}_{n_0\\times 3}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " from real scene data with distribution " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_r" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " and the complete point cloud " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A\\in \\mathbb{M}_{n_0\\times 3}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " from artificial data with distribution " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", the goal of unsupervised real scene point cloud completion is to predict the complete point cloud " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "Com(\\mathcal{P}_R)" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " such" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "type": "text", + "content": "that " + }, + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\operatorname{Com}(\\mathcal{P}_R) \\sim p_a" + }, + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R \\subset \\operatorname{Com}(\\mathcal{P}_R)" + }, + { + "bbox": [ + 304, + 313, + 545, + 350 + ], + "type": "text", + "content": ". This means that the prediction must be similar to the artificial data and meanwhile retain the shapes of the input partial point cloud." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": "We carefully design several modules of our generator " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": " and discriminators, " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": ", for unsupervised completion on real scene data through generative adversarial learning [8] as shown in Figures 2 and 3. As we mentioned above, many natural and man-made objects exhibit significant symmetries. To fully leverage these symmetries and preserve the shape of existing partial input " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": ", we devise a novel symmetry learning module that predicts the symmetric plane and the symmetrical point cloud of " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": " for further refinement. Our refinement autoencoder learns the detailed information of artificial data and refines the details of the initial, but coarse prediction " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": " for uniform and accurate prediction. With the local feature grouping operation, our elaborate upsampling refinement module can utilize the local features for more detailed results. Different from previous works focusing on the discriminative process of latent space, our point discriminator " + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 304, + 351, + 546, + 568 + ], + "type": "text", + "content": " also takes point clouds as input for direct guidance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Initial Coarse Predictor. We first employ an initial coarse predictor for shape-preserving and preliminary prediction. The encoder " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "E_{R}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " takes the real scene partial point cloud " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " as input and encodes the global feature " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_{R}^{\\prime} = E_{R}(\\mathcal{P}_{R})\\in \\mathbb{M}_{1\\times c_{0}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_{R}^{\\prime}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " is then taken by the decoder " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{R}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " for the prediction of initial, but coarse complete result " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "P_R' = D_R(f_R')\\in \\mathbb{M}_{n_1\\times 3}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": ". Encoder-decoder architecture is hard to retain shape features on the completion task since there is no explicit complete ground truth. To this end, we devise the symmetry learning module to accommodate symmetrical shape features widely existed in man-made and natural objects, as shown in Figure 3. The decoder " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{sl}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " pre" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13562" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "text", + "content": "dicts symmetric arguments " + }, + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = D_{sl}(f_R') = [x_A, y_A, z_A]^T" + }, + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "text", + "content": " and then we introduce symmetry operation to calculate the symmetrical point cloud " + }, + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "P_{sym}" + }, + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 46, + 71, + 287, + 109 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 117, + 287, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 117, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 117, + 117, + 287, + 142 + ], + "type": "interline_equation", + "content": "P _ {s y m} = \\boldsymbol {p} - 2 \\frac {\\boldsymbol {A} \\cdot \\boldsymbol {p}}{\\| \\boldsymbol {A} \\| ^ {2}} \\boldsymbol {A}, \\tag {1}", + "image_path": "080ba1f2bf4ce0b859fee372bb68346f395751295e0dcc7938eebb23f8387898.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{p} = [x,y,z]^T\\in \\mathcal{P}_R" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " represents the inner product and " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "||\\pmb {x}||" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " denotes the module of the vector " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " . Without the loss of generality, we can safely assume that all the objects are axis-aligned. Under such an assumption, the symmetric plane is perpendicular to the " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "xz" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " -plane and zero-crossing. Thus we set " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "y_{A} = 0" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " . Then " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "P_{sym}" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "P_R^{\\prime}" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " are fused and resampled through FPS (farthest point sampling) to get " + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 150, + 287, + 247 + ], + "type": "text", + "content": " which is the complete and shape-preserving result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "content": "Refinement Autoencoder. Noisy " + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "content": " and the fusion operation may cause noisy " + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "content": " with uneven point distribution. The initial coarse predictor or the point discriminator only predicts or judges through the global feature, which makes " + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "content": " lack details. Thus, " + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "content": " requires denoising and further refinement on details and point distribution. To this end, we employ the refinement autoencoder with a carefully designed upsampling refinement module. As shown in Figure 2, the blue and green lines of the refinement autoencoder represent the same forward process having the same network architectures with sharing weights. In this way, our autoencoder extracts the detailed information of artificial data and refines the details of " + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 247, + 287, + 413 + ], + "type": "text", + "content": " for more accurate and uniform prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": "The encoder and decoder " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "E_{A},D_{A}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " generate the skeleton point cloud " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "P_{R_1 / A_1}\\in \\mathbb{M}_{n_2\\times 3}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " of input point cloud. We set " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " to a small number for easier prediction of " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "P_{R_1 / A_1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": ". And the density loss is faster to calculate with fewer points. We then employ upsampling refinement module to upsample " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "P_{R_1 / A_1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " and add details to the skeleton for final prediction " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "P_{R_2 / A_2}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": ". As shown in Figure 3, we utilize wight-shared MLP [3] to obtain the point features and max-pooling operation on point features to obtain the global feature of the point cloud. We also concatenate the point and global features to fuse the information. Since wight-shared MLP only extracts the feature of one point, lacking the local information, we also employ the local feature grouping to get the local information around each point. For point " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "f_{group}\\in \\mathbb{M}_{n_2\\times (c_1\\cdot m)}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " are the features grouped from " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "f_{input}\\in \\mathbb{M}_{n_2\\times c_1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " corresponding to the points " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " that " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\| p - p_0\\| \\leq r" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " is the feature number of each group. After applying max-pooling operation on " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " features of each group, " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "f_{local}\\in \\mathbb{M}_{n_2\\times c_1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " are concatenated with " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "f_{input}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " to generate the output " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "f_{output}\\in \\mathbb{M}_{n_2\\times 2c_1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": ". We utilize a weight-shared MLP for upsampling that takes fused features composed of " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "P_{R_1 / A_1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": ", local and global features, and predict " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " shifts for each point. Then, each point splits into " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": " points by adding the predicted shifts to its coordinate whose result is " + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "inline_equation", + "content": "P_{R_2 / A_2}\\in \\mathbb{M}_{kn_2\\times 3}" + }, + { + "bbox": [ + 46, + 415, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 538, + 139 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 538, + 139 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 538, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 538, + 139 + ], + "type": "image", + "image_path": "6df62d8a68bf4a85c11c6c19a6e2f60514742af0ef23556f6afefaee378b946d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 152, + 538, + 229 + ], + "blocks": [ + { + "bbox": [ + 317, + 140, + 447, + 152 + ], + "lines": [ + { + "bbox": [ + 317, + 140, + 447, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 447, + 152 + ], + "type": "text", + "content": "Upsampling Refinement Module" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 152, + 538, + 229 + ], + "lines": [ + { + "bbox": [ + 316, + 152, + 538, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 538, + 229 + ], + "type": "image", + "image_path": "90846916c9692d15b10200c537e9d1d31275caf446df0ff45d2474a0723554ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 230, + 417, + 241 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 417, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 417, + 241 + ], + "type": "text", + "content": "Local Feature Grouping" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 242, + 463, + 320 + ], + "blocks": [ + { + "bbox": [ + 315, + 242, + 463, + 320 + ], + "lines": [ + { + "bbox": [ + 315, + 242, + 463, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 242, + 463, + 320 + ], + "type": "image", + "image_path": "8cae283e51f0ca5bc5d18d4521552a9c672ce864e735014bf914f3a886644a2b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 320, + 428, + 331 + ], + "lines": [ + { + "bbox": [ + 315, + 320, + 428, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 428, + 331 + ], + "type": "text", + "content": "Symmetry Learning Module" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 469, + 246, + 485, + 262 + ], + "blocks": [ + { + "bbox": [ + 469, + 246, + 485, + 262 + ], + "lines": [ + { + "bbox": [ + 469, + 246, + 485, + 262 + ], + "spans": [ + { + "bbox": [ + 469, + 246, + 485, + 262 + ], + "type": "image", + "image_path": "efe0b02cc3fd5d45a1089578c90f3d8cff7a594ce3c219542db67ecf9e35bacf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 250, + 534, + 258 + ], + "lines": [ + { + "bbox": [ + 489, + 250, + 534, + 258 + ], + "spans": [ + { + "bbox": [ + 489, + 250, + 534, + 258 + ], + "type": "text", + "content": "Concatenate" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 470, + 266, + 485, + 281 + ], + "blocks": [ + { + "bbox": [ + 470, + 266, + 485, + 281 + ], + "lines": [ + { + "bbox": [ + 470, + 266, + 485, + 281 + ], + "spans": [ + { + "bbox": [ + 470, + 266, + 485, + 281 + ], + "type": "image", + "image_path": "923708c274ca2ebb767d9b0811e99451a6afa3dda58de7e1f890b7a4d3481246.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 269, + 534, + 278 + ], + "lines": [ + { + "bbox": [ + 489, + 269, + 534, + 278 + ], + "spans": [ + { + "bbox": [ + 489, + 269, + 534, + 278 + ], + "type": "text", + "content": "Max-Pooling" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 470, + 285, + 485, + 304 + ], + "blocks": [ + { + "bbox": [ + 470, + 285, + 485, + 304 + ], + "lines": [ + { + "bbox": [ + 470, + 285, + 485, + 304 + ], + "spans": [ + { + "bbox": [ + 470, + 285, + 485, + 304 + ], + "type": "image", + "image_path": "a938c514f554c0d23359bd5d2c0ad9391f993c3e4b86512f87f6675443a67458.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 285, + 536, + 304 + ], + "lines": [ + { + "bbox": [ + 488, + 285, + 536, + 304 + ], + "spans": [ + { + "bbox": [ + 488, + 285, + 536, + 304 + ], + "type": "text", + "content": "Local Feature Grouping" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 311, + 485, + 327 + ], + "blocks": [ + { + "bbox": [ + 470, + 311, + 485, + 327 + ], + "lines": [ + { + "bbox": [ + 470, + 311, + 485, + 327 + ], + "spans": [ + { + "bbox": [ + 470, + 311, + 485, + 327 + ], + "type": "image", + "image_path": "f0877c0040db32685acef03a3bef9f5f2338a32a16dc7d1875b1e391dd32c479.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 311, + 526, + 328 + ], + "lines": [ + { + "bbox": [ + 495, + 311, + 526, + 328 + ], + "spans": [ + { + "bbox": [ + 495, + 311, + 526, + 328 + ], + "type": "text", + "content": "Element Addition" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 333, + 545, + 366 + ], + "lines": [ + { + "bbox": [ + 305, + 333, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 545, + 366 + ], + "type": "text", + "content": "Figure 3. Detailed structures of our symmetry learning module (red), upsampling refinement module (yellow), and local feature sampling (blue)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": "Point and Feature Discriminator. We employ the point discriminator " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " to directly guide the generation of " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "P_{R_2}" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " comparing with the artificial complete data " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " for unsupervised completion. We also guide the generation of latent code of autoencoder through the feature discriminator " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "f_R" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "f_A" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " have the same distribution for more accurate prediction. Our discriminators predict the probability that input point cloud " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " or feature " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " is real data. For multi-category data, we replace the probability-guided discriminator with a classifier-guided discriminator which outputs the category label of input rather than probability. We utilize label 0 to represent fake data and labels " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "l > 0" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " to denote real data of different categories such as the chair, table, etc. For fake data, the discriminator tends to output label 0, while the generator tends to make the discriminator output label " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "l > 0" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": ". And for real data with label " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": ", the discriminator tends to output " + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 369, + 545, + 596 + ], + "type": "text", + "content": " correctly. In this way, our USSPA can unsupervisedly classify the real scene partial point cloud when completing multi-category data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 600, + 391, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 600, + 391, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 600, + 391, + 612 + ], + "type": "text", + "content": "3.2. Optimization" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 614, + 544, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 544, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 544, + 639 + ], + "type": "text", + "content": "The abbreviate optimization goal of our USSPA for unsupervised learning is:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 339, + 649, + 544, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 649, + 544, + 668 + ], + "spans": [ + { + "bbox": [ + 339, + 649, + 544, + 668 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G}} \\max _ {\\mathcal {D} _ {p}, \\mathcal {D} _ {f}} \\mathbb {E} _ {P _ {i} \\sim p _ {a}} \\log p + \\mathbb {E} _ {P _ {i} \\sim p _ {r}} \\log (1 - p), \\tag {2}", + "image_path": "66d4ea0be5002fdda450337bae76fb5a958c341caad7e9802e3eea5229c0cbf6.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " is the probability of real data. By playing the maxmin game with " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": ", our generator " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " tends to predict point clouds similar to artificial data. We employ" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13563" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 105, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Point cloud completion comparison on our dataset in terms of L1 Chamfer Distance " + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "cd^{l1} \\times 10^2" + }, + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": " (lower is better). All the methods are trained with our dataset and the unsupervised methods including ours are trained on single category data. Boldface denotes the best among unsupervised methods in Tables 1, 2 and 3." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 105, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 53, + 105, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 105, + 541, + 220 + ], + "type": "table", + "html": "
MethodAVGchairtabletrash binTVcabinetbookshelfsofalampbedtub
Sup.PoinTr [34]14.3713.6512.5215.2612.6917.3213.9912.3617.0515.1313.77
Disp3D [23]7.786.248.207.127.1210.366.945.6014.036.905.32
TopNet [19]7.076.395.797.406.268.377.025.948.507.817.25
Unsup. OptBasedShapeInv [36]21.3917.9717.2833.5115.6926.2625.5114.2816.6932.3314.43
Unsup.Unpaired [32]10.478.417.5212.086.7217.459.956.9219.3610.046.22
Cycle4 [24]11.539.1111.3511.938.4015.4712.5110.6312.2515.737.92
Ours8.568.227.6810.367.6610.777.846.1411.938.206.75
", + "image_path": "8446cea51927e05b3827bb2793b11c453971f6f7b1faf68ac74ed3f865d2a945.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "text", + "content": "Chamfer Distance(CD) [6] to measure the difference between point clouds. The " + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "inline_equation", + "content": "cd_{P_1 \\to P_2}^l" + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "text", + "content": " indicates the single direction Chamfer Distance with L2-norm from point cloud " + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "inline_equation", + "content": "P_2" + }, + { + "bbox": [ + 47, + 223, + 287, + 271 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 276, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 276, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 87, + 276, + 287, + 306 + ], + "type": "interline_equation", + "content": "c d _ {P _ {1} \\rightarrow P _ {2}} ^ {l 2} = \\frac {1}{| P _ {1} |} \\sum_ {\\boldsymbol {x} \\in P _ {1}} \\min _ {\\boldsymbol {y} \\in P _ {2}} \\| \\boldsymbol {x} - \\boldsymbol {y} \\| ^ {2}, \\tag {3}", + "image_path": "71c4e7d5d41c6829f242fc24ec235b1520fad987a64d17d11428f5bbbeaeaf31.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "content": "which guides " + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "content": " to be a part of " + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "inline_equation", + "content": "cd_{P_1\\leftrightarrow P_2}^{l^2}" + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 47, + 312, + 287, + 337 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 342, + 287, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 342, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 97, + 342, + 287, + 358 + ], + "type": "interline_equation", + "content": "c d _ {P _ {1} \\leftrightarrow P _ {2}} ^ {l 2} = c d _ {P _ {1} \\rightarrow P _ {2}} ^ {l 2} + c d _ {P _ {2} \\rightarrow P _ {1}} ^ {l 2}, \\tag {4}", + "image_path": "e35da40750e6ca75d038235ef40ad1dc765c4af3277e98ecef5a81430d4802a0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": "which guides " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": " to be identical. As shown in Figure 2, to restrict the generation of " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "P_{R}^{\\prime}" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": ", we employ single direction Chamfer Distance from " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_R" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "P_{R}^{\\prime}" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": ". We also utilize Chamfer Distance between " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "P_{R_1 / A_1}, P_{R_2 / A_2}" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "P_{R_0}, \\mathcal{P}_A" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": " for supervision of autoencoder. For uniform prediction, we employ density loss by first calculating the local density of " + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\in P" + }, + { + "bbox": [ + 47, + 363, + 289, + 446 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 453, + 287, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 453, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 97, + 453, + 287, + 485 + ], + "type": "interline_equation", + "content": "l d (\\boldsymbol {x}, P) = \\frac {1}{k _ {d}} \\sum_ {k = 1} ^ {k _ {d}} \\min _ {\\boldsymbol {y} \\in P} ^ {k} | | \\boldsymbol {x} - \\boldsymbol {y} | |, \\tag {5}", + "image_path": "1db1f25df760062eb1e8cee2833a59ad1e3264ad5aab695ad0562700fbbf72bd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 491, + 259, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 259, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 259, + 502 + ], + "type": "text", + "content": "and then calculating the standard deviation for all " + }, + { + "bbox": [ + 47, + 491, + 259, + 502 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 491, + 259, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 510, + 287, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 510, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 74, + 510, + 287, + 541 + ], + "type": "interline_equation", + "content": "d l (P) = \\sqrt {\\frac {1}{| P |} \\sum_ {\\boldsymbol {x} \\in P} \\left(l d (\\boldsymbol {x} , P) - \\overline {{l d (\\boldsymbol {x} , P)}}\\right) ^ {2}}, \\tag {6}", + "image_path": "8bc22e82093b0397c4908a645250474384b1ed3a8400f72bd81bbfcca49d57ac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\min" + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "text", + "content": "-th-min element. The smaller the " + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "inline_equation", + "content": "ld(\\pmb{x}, P)" + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "text", + "content": " is, the bigger the local density of " + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 549, + 287, + 578 + ], + "type": "text", + "content": " is." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "spans": [ + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "type": "text", + "content": "In practice, we alternately optimize " + }, + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "type": "text", + "content": " with loss " + }, + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{G}}" + }, + { + "bbox": [ + 59, + 579, + 268, + 591 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 96, + 599, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 599, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 96, + 599, + 287, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} = \\alpha_ {1} \\mathcal {L} _ {F \\rightarrow R} + \\alpha_ {2} \\mathcal {L} _ {P} + \\alpha_ {3} \\mathcal {L} _ {d}, \\tag {7}", + "image_path": "d57c8a97fd959308dc653ee783e2214b525812384e914b420923f10dcbc126be.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 620, + 75, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 75, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 75, + 631 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 80, + 636, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 636, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 80, + 636, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {F \\rightarrow R} = - \\left(\\log \\mathcal {D} _ {p} \\left(P _ {R _ {0}}\\right) + \\log \\mathcal {D} _ {p} \\left(P _ {R _ {2}}\\right)\\right. \\tag {8} \\\\ + \\log \\mathcal {D} _ {f} (f _ {R})) \\\\ \\end{array}", + "image_path": "3cfbef2a4956ee9ee0bdd6e6a87a595be0f761edd300b805731690820bdaf70a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 74, + 681, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 681, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 74, + 681, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {P} = c d _ {P _ {R _ {1}} \\leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {P _ {R _ {2}} \\leftrightarrow P _ {R _ {0}}} ^ {l 2} + c d _ {\\mathcal {P} _ {R} \\rightarrow P _ {R} ^ {\\prime}} ^ {l 2} \\tag {9} \\\\ + c d _ {P _ {A _ {1}} \\leftrightarrow \\mathcal {P} _ {A}} ^ {l ^ {2}} + c d _ {P _ {A _ {2}} \\leftrightarrow \\mathcal {P} _ {A}} ^ {l ^ {2}}, \\\\ \\end{array}", + "image_path": "3caab1cf172eaaedd754fa77d4a610e80f61f9ffa67492bc7398638e30f678e4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 373, + 246, + 545, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 246, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 373, + 246, + 545, + 259 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = d l \\left(P _ {R _ {1}}\\right) + d l \\left(P _ {A _ {1}}\\right) \\tag {10}", + "image_path": "31314f2832dc140c7248d703d2e9401506e48444feb4ea1ed164abb567cc6762.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 270, + 392, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 392, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 392, + 281 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 306, + 270, + 392, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 306, + 270, + 392, + 281 + ], + "type": "text", + "content": " with loss " + }, + { + "bbox": [ + 306, + 270, + 392, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{D}}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 293, + 545, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 293, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 378, + 293, + 545, + 306 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {D}} = \\alpha_ {4} \\mathcal {L} _ {F} + \\alpha_ {5} \\mathcal {L} _ {R}, \\tag {11}", + "image_path": "ae87a40433a0cfa43cee1956d2709c2276bc2436983e72cc3bcb8ad0438836d5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 316, + 334, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 334, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 334, + 327 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 334, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 334, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 334, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {F} = - \\left(\\log \\left[ 1 - \\mathcal {D} _ {p} \\left(P _ {R _ {0}}\\right) \\right] + \\log \\left[ 1 - \\mathcal {D} _ {p} \\left(P _ {R _ {2}}\\right) \\right] \\right. \\tag {12} \\\\ \\left. + \\log \\left[ 1 - \\mathcal {D} _ {f} \\left(f _ {R}\\right) \\right]\\right), \\\\ \\end{array}", + "image_path": "56a754259ea213f8e8005be3630aa966ec30204447d92d8ee7ffa7cce74625c6.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 388, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 388, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 345, + 388, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = - \\left(\\log \\mathcal {D} _ {p} \\left(\\mathcal {P} _ {A}\\right) + \\log \\mathcal {D} _ {f} \\left(f _ {A}\\right)\\right), \\tag {13}", + "image_path": "95da6176f9c02fa0d0715d7d2244c37de4b7d192797e675a2103d0b8453a8b0b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 413, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 460 + ], + "type": "text", + "content": "and here " + }, + { + "bbox": [ + 304, + 413, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\alpha_{1\\sim 5}" + }, + { + "bbox": [ + 304, + 413, + 545, + 460 + ], + "type": "text", + "content": " are the weights used for balancing the influences between each term. And for multi-category data, we use cross-entropy to calculate the loss between predicted labels and ground truth labels." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 467, + 388, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 388, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 388, + 480 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 483, + 497, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 497, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 497, + 495 + ], + "type": "text", + "content": "4.1. Dataset and Implementation Details" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Different from the Unpaired [32] which utilizes a classifier to evaluate the completion results of real scene data, we build a new dataset from ScanNet [5], ShapeNet [2] and Scan2CAD [1], containing paired partial and complete point clouds for more accurate evaluation. Scan2CAD contains annotations that align the CAD models of ShapeNet to the real scenes of ScanNet. We first extract the CAD models from ShapeNet and partial point clouds from ScanNet which are paired according to the correspondences offered by Scan2CAD, and then uniformly sample the CAD models to get complete point clouds. Finally, we transform the paired partial and complete point clouds with the transform matrix given by Scan2CAD to normalize their face direction, scale, and position. We select ten categories including the chair, table, etc., to establish the dataset and randomly split the data into training, validation, and testing sets. In our experiments, these ground truths are only used for the evaluation of the complete results." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13564" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 532, + 320 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 532, + 320 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 532, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 532, + 320 + ], + "type": "image", + "image_path": "50ebc953567d54ff4c32e2be7a363499d0e024e27777648ddb9aa67791a13ff5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 546, + 346 + ], + "type": "text", + "content": "Figure 4. Visualization of completion results by supervised works, including TopNet [19], Disp3D [23], unsupervised works Cycle4 [24], Unpaired [32], optimization-based work ShapeInv [36] and our method. From top to bottom: table, bed, cabinet, and TV." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": "We utilize PyTorch to implement our USSPA. In Section 3.1, the point numbers " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "n_{0\\sim 2}" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": " are set to 2048, 512 and 512. The channel numbers " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "c_{0\\sim 1}" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": " are set to 512 and 256. The feature number " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": ", the radius " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": " of local feature grouping, and the split number of upsampling " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": " are set to 32, 0.5, and 4. " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "k_d" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": " of density loss in Section 3.2 is set to 16. The weights " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\alpha_{1\\sim 5}" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": " of losses are set to 1, 100, 7.5, 0.5 and 0.5. We employ Adam optimizer whose learning rate is " + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "inline_equation", + "content": "1.0 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": ". The batch size is set to 4, and the maximum epoch of training is 240 to 960 according to the size of the dataset. Training our USSPA takes about 20 hours with a GTX 2080Ti GPU. We will make our dataset and codes public in the future." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 504, + 128, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 128, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 128, + 517 + ], + "type": "text", + "content": "4.2. Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "text", + "content": "We compare our USSPA against existing representative point cloud completion methods including supervised [19, 23, 34], unsupervised [24, 32] and optimization-based [36] methods on our dataset and the PCN Dataset [35]. The metrics used for comparison include pre-point L1 Chamfer Distance " + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "inline_equation", + "content": "cd^{l1}" + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "text", + "content": " [6] and F-Scores " + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "inline_equation", + "content": "F_{score}^{1\\%}" + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "inline_equation", + "content": "F_{score}^{0.1\\%}" + }, + { + "bbox": [ + 46, + 520, + 287, + 593 + ], + "type": "text", + "content": " [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "As shown in Table 1, we compare these methods on our dataset with 10 categories where all methods are trained on the training set and evaluated on the testing set. All the unsupervised methods are trained with single category data for each class. The average " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "cd^{l1}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " of our USSPA is " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "18.2\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " lower than Unpaired and the average " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "F_{score}^{0.1\\%}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " higher than Unpaired as shown in Table 2, indicating the superior performance of our method on completion of real scene objects. We also mix up all 10 categories as the multi-category data to train the methods for evaluation of the generalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "text", + "content": "Table 2. Comparison on our dataset trained with single category and multi-category data in terms of L1 Chamfer Distance " + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "inline_equation", + "content": "cd^{l1} \\times 10^2" + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "text", + "content": " (lower is better) and F-scores " + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "inline_equation", + "content": "F_{score}^{0.1\\%} \\times 10^2" + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "inline_equation", + "content": "F_{score}^{1\\%} \\times 10^2" + }, + { + "bbox": [ + 305, + 350, + 546, + 416 + ], + "type": "text", + "content": " (higher is better). \"Ours\" and \"Ours(classifier)\" denote our method with probability-guided and classifier-guided discriminators, respectively." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 419, + 543, + 563 + ], + "blocks": [ + { + "bbox": [ + 307, + 419, + 543, + 563 + ], + "lines": [ + { + "bbox": [ + 307, + 419, + 543, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 543, + 563 + ], + "type": "table", + "html": "
Methodsingle categorymulti-category
↑F0.1% score↑F1% score↓cdl1↑F0.1% score↑F1% score
PoinTr [34]--14.3718.3580.41
Disp3D [23]--7.7830.2978.26
TopNet [19]--7.0712.3380.37
ShapeInv [36]15.5866.5319.3516.9869.66
Unpaired [32]12.2064.3310.1210.8666.68
Cycle4 [24]9.9860.1412.008.6156.57
Ours17.4973.418.9616.8872.31
Ours(classifier)--8.7617.1273.75
", + "image_path": "41260711794fb64449907b46a49018ba3574405f354d2d9a2dd7253fb1153f31.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "ability. As shown in Table 2, our USSPA keeps the same performance by replacing the probability-guided discriminators with classifier-guided discriminators and has better generalization ability on multi-category than other unsupervised networks. Further comparisons show the generalization capability of our USSPA on different datasets. We train unsupervised models with ScanNet and then test them on the PCN Dataset which is a point cloud completion dataset built upon artificial data. As shown in Table 3, our method also performs better than other unsupervised methods, the average " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "cd^{l1}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " of which is " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " lower than the SOTA method. Table 4 shows the classification accuracy of PN2 [14] and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13565" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 113, + 284, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 3. Point cloud completion comparison on PCN Dataset in terms of L1 Chamfer Distance " + }, + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "inline_equation", + "content": "cd^{l1} \\times 10^2" + }, + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": " (lower is better). All unsupervised models are trained with ScanNet." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 113, + 284, + 217 + ], + "lines": [ + { + "bbox": [ + 51, + 113, + 284, + 217 + ], + "spans": [ + { + "bbox": [ + 51, + 113, + 284, + 217 + ], + "type": "table", + "html": "
MethodAVGchairtablecabinetsofalamp
PoinTr [34]5.495.615.686.085.674.44
Disp3D [23]2.512.422.302.382.443.00
TopNet [19]5.926.345.456.065.805.95
ShapeInv [36]19.0523.1815.6617.1422.8516.40
Unpaired [32]14.8712.878.1414.3018.2320.82
Cycle4 [24]17.6014.2515.7321.0621.5415.40
Ours12.6313.529.668.8915.5115.57
", + "image_path": "5db3ff3690db2481351a8d59a5efa06691b717af27790384ae178a0bb546337e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 271, + 280, + 318 + ], + "blocks": [ + { + "bbox": [ + 47, + 228, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 262 + ], + "type": "text", + "content": "Table 4. Point cloud classification in terms of accuracy. \"AVG\" denotes the average accuracy on 10 categories of our dataset. Only four categories are shown due to width limitation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 271, + 280, + 318 + ], + "lines": [ + { + "bbox": [ + 53, + 271, + 280, + 318 + ], + "spans": [ + { + "bbox": [ + 53, + 271, + 280, + 318 + ], + "type": "table", + "html": "
MethodAVGchairtabletrash binsofa
PN2 [14]58.6%91.0%94.0%17.2%48.9%
Ours69.8%91.0%91.0%51.6%93.3%
", + "image_path": "3d83dcb74061f492c34a807dfbcda873a5a344c9148792f5bb07e244fcefd07c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 360 + ], + "type": "text", + "content": "our method. Our method can also unsupervisedly classify the real scene partial point cloud while completing. The accuracy of our USSPA is " + }, + { + "bbox": [ + 46, + 324, + 287, + 360 + ], + "type": "inline_equation", + "content": "11.2\\%" + }, + { + "bbox": [ + 46, + 324, + 287, + 360 + ], + "type": "text", + "content": " higher than PN2 on average." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 540 + ], + "type": "text", + "content": "Figure 4 shows the qualitative results on 4 categories. The complete point clouds predicted by our USSPA exhibit more accurate shapes, such as the leg of the table and the division plate of the bookshelf, which are hard to generate even for supervised methods. Our method can successfully generate the circular top of the table. By comparison, other unsupervised works generate a square one. Our predictions are also more uniform with fewer noises. Benefiting from our shape-preserving method, our predicted TV retains the support part from the input point cloud. Our symmetry learning module also plays an important role in completing the pillars of the bed by copying the pillars from the input symmetrically. Visualization of our predicted point clouds shows the superiority of our methods in the generation of detailed completion on real scene data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 542, + 141, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 141, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 141, + 554 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Ablation study on network architecture. We conduct experiments on ablated models for evaluation of the necessity of important structures of our USSPA, including the symmetry learning module, the refinement autoencoder, the local feature grouping, the point and feature discriminators, the single direction Chamfer Distance, and the upsampling refinement module. As shown in Table 5, these modules are removed from our full network separately. We downsample the final prediction of our USSPA to 512 points for a fair comparison to the ablated model without the upsampling refinement module. We train and utilize " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "F_{score}^{0.1\\%}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " to evaluate these ablated models with our dataset on the chair category, and the comparison indicates the necessity of these modules" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 324, + 113, + 528, + 243 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 5. Ablation study in terms of " + }, + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "F_{score}^{0.1\\%}" + }, + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "text", + "content": ", where “-” represents removing a module from the full network respectively and “512” denotes the point number." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 113, + 528, + 243 + ], + "lines": [ + { + "bbox": [ + 324, + 113, + 528, + 243 + ], + "spans": [ + { + "bbox": [ + 324, + 113, + 528, + 243 + ], + "type": "table", + "html": "
ModelF0.1% score
- Symmetry Learning Module16.43
- Refinement Autoencoder17.08
- Local Feature Grouping15.72
- Feature Discriminator17.11
- Point Discriminator15.68
- Single Direction Chamfer Distance17.84
Ours (full)18.43
- Upsampling Refinement (512)5.60
Ours (full) (512)6.46
", + "image_path": "84a266a8b271bb3de1ada255cea998db0e46a0b091648c78fd67cf4a609ce1e5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 252, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 307, + 252, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 334 + ], + "type": "image", + "image_path": "3ad04e4de23b29a7956529e38d60b2ba392bfe4dfde3913da6fab02cdf01b98b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 372 + ], + "type": "text", + "content": "Figure 5. Visualization of the symmetric planes and predictions by USSPA without the symmetry learning module and full network. Blue and orange denote original and mirrored points, separately." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "text", + "content": "which play significant roles in shape-preserving, extracting, and utilizing detailed features, upsampling and refining for more accurate point cloud completion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 414, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 546, + 605 + ], + "type": "text", + "content": "Symmetry learning module. Our symmetry learning module learns different symmetric planes according to the input point clouds as shown in Figures 5(a) and 5(b). The predicted symmetric plane is xy-plane for the chair and yz-plane for the table. Through different symmetric planes, our network can fully leverage the symmetrical structures of input point clouds, such as the arm of the chair and the legs of the table, which exist on only one side. Figure 5(c) shows the prediction " + }, + { + "bbox": [ + 304, + 414, + 546, + 605 + ], + "type": "inline_equation", + "content": "P_{R_2}" + }, + { + "bbox": [ + 304, + 414, + 546, + 605 + ], + "type": "text", + "content": " of our network without the symmetry learning module, and Figure 5(d) shows the result of our full network. Figure 5(e) shows the input point cloud whose left arm is missing while the right arm is complete. Without the symmetry learning module, the left arm of the prediction is incomplete as shown in the red box. And benefiting from such a module, our full network can learn the symmetric structures, and generate a complete left arm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Density loss. We visualize the local density of points as shown in Figure 6. The warmer the color of the point is, the bigger its local density (described in equation 5) is. (a) is the prediction without the supervision of the density loss, and (b) is the prediction supervised by the density loss. Without the supervision of density loss, the network prefers generating many points in the same location as shown in the blue circle to minimize the Chamfer Distance, which makes the predicted point clouds contain fewer points describing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13566" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 121, + 268, + 156 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "type": "text", + "content": "Table 6. Comparison of Unpaired and our USSPA in terms of density loss, where \"ours (w/o " + }, + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "type": "inline_equation", + "content": "dl" + }, + { + "bbox": [ + 47, + 70, + 289, + 115 + ], + "type": "text", + "content": ")\" indicates our USSPA without the supervision of the density loss. The results show that our predicted point clouds are more uniform." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 121, + 268, + 156 + ], + "lines": [ + { + "bbox": [ + 65, + 121, + 268, + 156 + ], + "spans": [ + { + "bbox": [ + 65, + 121, + 268, + 156 + ], + "type": "table", + "html": "
Unpairedours (w/o dl)ours
Density Loss0.09570.12380.0149
", + "image_path": "60e4cb56b0df4722a39c54d55d9d235c74ce3002e273332eabea170d5a73a084.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 161, + 106, + 249 + ], + "blocks": [ + { + "bbox": [ + 49, + 161, + 106, + 249 + ], + "lines": [ + { + "bbox": [ + 49, + 161, + 106, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 161, + 106, + 249 + ], + "type": "image", + "image_path": "dc48897284b2cd60fdb4dea05a44266c4b191d525270a63d1f40eec626af6d9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "type": "text", + "content": "Figure 6. Visualization of the local density of point clouds, where the warmth of the color indicates the values of local densities. The warmer the color is, the bigger the local density is. (a) and (b) are the comparison of the ablation experiment on the density loss. (c) and (d) are the examples of " + }, + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "type": "inline_equation", + "content": "P_{R_2}" + }, + { + "bbox": [ + 46, + 262, + 287, + 328 + ], + "type": "text", + "content": ". The density loss and the point number of each point cloud are shown at its bottom." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 108, + 162, + 166, + 249 + ], + "blocks": [ + { + "bbox": [ + 108, + 162, + 166, + 249 + ], + "lines": [ + { + "bbox": [ + 108, + 162, + 166, + 249 + ], + "spans": [ + { + "bbox": [ + 108, + 162, + 166, + 249 + ], + "type": "image", + "image_path": "71ec2aea586c7429d4c70b60810b95212742fa774dc044fcba4a7fe544000308.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 169, + 166, + 228, + 247 + ], + "blocks": [ + { + "bbox": [ + 169, + 166, + 228, + 247 + ], + "lines": [ + { + "bbox": [ + 169, + 166, + 228, + 247 + ], + "spans": [ + { + "bbox": [ + 169, + 166, + 228, + 247 + ], + "type": "image", + "image_path": "014284f2971eb8e5dc8dae2871dbfc862107835bfd3946161fe642da7ac99679.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 229, + 164, + 288, + 247 + ], + "blocks": [ + { + "bbox": [ + 229, + 164, + 288, + 247 + ], + "lines": [ + { + "bbox": [ + 229, + 164, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 229, + 164, + 288, + 247 + ], + "type": "image", + "image_path": "7134bc6d3af70df8f1d07f6cc0dc1dd3db7d449322a6644677d1f34c5d9721e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "text", + "content": "details such as chair arms and legs. Table 6 compares Unpaired and our USSPA on density loss indicating that our predicted point clouds are more uniform." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": "Autoencoding refiner. As shown in Figure 6, (c) and (d) are the examples of " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "P_{R_2}" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": " which is the refinement result of the autoencoding refiner on " + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "inline_equation", + "content": "P_{R_0}" + }, + { + "bbox": [ + 46, + 376, + 287, + 460 + ], + "type": "text", + "content": ". The refiner removes the noises from input shown in blue boxes and generates more points on legs as shown in green boxes, which makes the whole point cloud more uniform and accurate with less density loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "Real scene data. We also test our USSPA on real objects, as shown in Figure 7. We obtain the point clouds of real objects by the 3D scanner and then sample them to 2048 points as the input to our network. Our network can preserve the right arm of the chair and complete the left arm. The predictions are more accurate and uniform as shown in red boxes. This indicates the generalization ability of our USSPA on real objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Visualization of latent codes on multi-category data. We utilize t-SNE [21] to visualize the latent codes " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "f_{R}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " of Unpaired and our USSPA on multi-category data. As shown in Figure 8, ten different colors indicate ten categories where dark colors indicate real scene data and light colors indicate artificial data. Figure 8 (c) and (d) show the t-SNE results of latent codes on chair category only. Benefiting from our feature discriminator and autoencoder, the latent codes of real scene data are similar to artificial data which can be shown from the distribution of dark and light points. The dark and light points of Unpaired are separated while ours are fixed together for each category. Benefiting from our classifier-guided discriminator, the distributions of" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 189 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 189 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 189 + ], + "type": "image", + "image_path": "7cf0cecbbab6595a26b95c0353320bfd95fadf19f47261ad6018f23eb772c395.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 194, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 216 + ], + "type": "text", + "content": "Figure 7. Completion on real objects by Unpaired [32], ShapeInv [36] and our USSPA." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 425, + 340 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 425, + 340 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 425, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 425, + 340 + ], + "type": "image", + "image_path": "34b72b68879acc150a93e2ced5a68c5095a513b1861ec8a7c726c5ea4bd38f8c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 345, + 545, + 400 + ], + "lines": [ + { + "bbox": [ + 305, + 345, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 545, + 400 + ], + "type": "text", + "content": "Figure 8. Visualization of t-SNE results of the latent codes by Unpaired [32] and our method. (a) and (b) are the results on multi-category, where different colors denote different categories. Dark and light denote real scene data and artificial data respectively. (c) and (d) are the results on the chair category only." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 426, + 230, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 426, + 230, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 426, + 230, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 426, + 230, + 545, + 340 + ], + "type": "image", + "image_path": "48e7821b250ad1cd936a453b8c8bc9080ae71a957b7c5e1d1387e8666091f951.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 407, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 456 + ], + "type": "text", + "content": "our latent codes of different categories are separated and clean. For example, as Figure 8 (b) shows, the distances between pink, orange and blue clusters are bigger and the clusters are tighter." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 458, + 531, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 458, + 531, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 458, + 531, + 471 + ], + "type": "text", + "content": "5. Conclusion, Limitation and Future Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 475, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 654 + ], + "type": "text", + "content": "We have presented USSPA, an end-to-end unsupervised network for completion of real scene point cloud objects. Benefiting from our carefully designed symmetry learning module and refinement autoencoder, our prediction preserves the symmetrical shape of input with more accurate and uniform points. We also contribute a real scene dataset for accurate evaluation, and extensive experiments and comparisons show the superiority and generalization of our method on different categories of different datasets and real objects, which achieves state-of-the-art performance on unsupervised completion of real scene objects. For those categories without artificial data as assistance, previous works and our method are hard to apply in this situation. This is our limitation. Unsupervised point cloud completion without the assistance of artificial data is our future work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 662, + 403, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 662, + 403, + 675 + ], + "spans": [ + { + "bbox": [ + 306, + 662, + 403, + 675 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant numbers 62032011, 61972194." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13567" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X. Chang, and Matthias Niessner. Scan2cad: Learning cad model alignment in rgb-d scans. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "text", + "content": "[3] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jul 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 300 + ], + "type": "text", + "content": "[4] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[7] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1359-1368, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 512 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "text", + "content": "[9] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[10] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh. ACM Transactions on Graphics (TOG), 39:126:1 - 126:12, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7662-7670, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[12] Giseop Kim and Ayoung Kim. Scan context: Egocentric spatial descriptor for place recognition within 3d point cloud map. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4802-4809. IEEE, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Bo Li. 3d fully convolutional network for vehicle detection in point cloud. In 2017 IEEE/RSJ International Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "on Intelligent Robots and Systems (IROS), pages 1513-1518. IEEE, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 172 + ], + "type": "text", + "content": "[14] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, 2017 Advances in Neural Information Processing Systems (NIPS), volume 30. Curran Associates, Inc., 2017. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "text", + "content": "[15] Abhishek Sharma, Oliver Grau, and Mario Fritz. Vconv-dae: Deep volumetric shape learning without object labels. In European Conference on Computer Vision, pages 236-250. Springer, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[16] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1955-1964, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[17] Junshu Tang, Zhijun Gong, Ran Yi, Yuan Xie, and Lizhuang Ma. Lake-net: Topology-aware point cloud completion by localizing aligned keypoints. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1716-1725, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 375 + ], + "type": "text", + "content": "[18] Maxim Tatarchenko, Stephan R Richter, René Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405-3414, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 431 + ], + "type": "text", + "content": "[19] Lyne P Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese. Topnet: Structural point cloud decoder. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 383-392, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "text", + "content": "[20] Duc Thanh Nguyen, Binh-Son Hua, Khoi Tran, Quang-Hieu Pham, and Sai-Kit Yeung. A field model for repairing 3d shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5676-5684, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[21] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579-2605, 2008. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[22] Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cascaded refinement network for point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 790-799, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[23] Yida Wang, David Joseph Tan, Nassir Navab, and Federico Tombari. Learning local displacements for point cloud completion. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "text", + "content": "[24] Xin Wen, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Cycle4completion: Unpaired point cloud completion using cycle transformation with missing region coding. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13075-13084, 2021. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[25] Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hierarchical folding. In Proceedings of the IEEE/CVF Con" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13568" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 699 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 1939-1948, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "text", + "content": "[26] Xin Wen, Peng Xiang, Zhizhong Han, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Yu-Shen Liu. Pmp-net: Point cloud completion by learning multi-step point moving paths. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 196 + ], + "type": "text", + "content": "[27] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In The European Conference on Computer Vision (ECCV), August 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 286, + 229 + ], + "type": "text", + "content": "[28] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Density-aware chamfer distance as a comprehensive metric for point cloud completion. 11 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 286, + 274 + ], + "type": "text", + "content": "[29] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. Gan inversion: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-17, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 286, + 330 + ], + "type": "text", + "content": "[30] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. SnowflakeNet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 286, + 385 + ], + "type": "text", + "content": "[31] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 286, + 430 + ], + "type": "text", + "content": "[32] Niloy J. Mitra Xuelin Chen, Baoquan Chen. Unpaired point cloud completion on real scans using adversarial training. In 2020 International Conference on Learning Representations 2020 (ICLR), 2020. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 286, + 475 + ], + "type": "text", + "content": "[33] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 206–215, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 286, + 531 + ], + "type": "text", + "content": "[34] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12498-12507, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 286, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 286, + 575 + ], + "type": "text", + "content": "[35] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 International Conference on 3D Vision (3DV), pages 728-737. IEEE, 2018. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 286, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 642 + ], + "type": "text", + "content": "[36] Junzhe Zhang, Xinyi Chen, Zhongang Cai, Liang Pan, Haiyu Zhao, Shuai Yi, Chai Kiat Yeo, Bo Dai, and Chen Change Loy. Unsupervised 3d shape completion through gan inversion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1768-1777, 2021. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 644, + 286, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 286, + 699 + ], + "type": "text", + "content": "[37] Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15890-15899, June 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13569" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_content_list.json b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b370d22c9a4d4f0789c590d778f278d68a340233 --- /dev/null +++ b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_content_list.json @@ -0,0 +1,1499 @@ +[ + { + "type": "text", + "text": "SynthVSR: Scaling Up Visual Speech Recognition With Synthetic Supervision", + "text_level": 1, + "bbox": [ + 91, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xubo Liu $^{1*}$ , Egor Lakomkin $^{2}$ , Konstantinos Vougioukas $^{2}$ , Pingchuan Ma $^{2}$ , Honglie Chen $^{2}$ , Ruiming Xie $^{2}$ , Morrie Doulaty $^{2}$ , Niko Moritz $^{2}$ , Jachym Kolar $^{2}$ , Stavros Petridis $^{2}$ , Maja Pantic $^{2}$ , Christian Fuegen $^{2}$", + "bbox": [ + 84, + 179, + 921, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Surrey $^{2}$ Meta AI", + "bbox": [ + 320, + 223, + 643, + 242 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c83aa001db9bf9b18f74e94f33c8cbbf5848c6198b7eb2f947af666b6faf19d.jpg", + "image_caption": [ + "Figure 1. Scaling up visual speech recognition with synthetic supervision (SynthVSR): we propose SynthVSR, a semi-supervised framework that can substantially improve the performance of VSR models by using synthetic lip movements. Firstly, we introduce a speech-driven lip animation model that generates lip movement videos conditioned on input lip images and speech utterances (left). Secondly, we generate large-scale synthetic videos using transcribed speech datasets and lip images. The combination of synthetic videos and their corresponding speech transcriptions constitutes the synthetic dataset (centre). Finally, we conduct semi-supervised VSR training with synthetic and real datasets. Our method substantially improves the performance of VSR models with large-scale synthetic data (right)." + ], + "image_footnote": [], + "bbox": [ + 78, + 257, + 341, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2204cfd53a62e8188cab6ef04555d56ed95e5a2eaa6a468886cc70f85aabeea3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 257, + 635, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c690f05f11549ce55d61c18a411de52117a0004767962da6952c7f5cec4a9da4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 257, + 888, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 542, + 313, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently reported state-of-the-art results in visual speech recognition (VSR) often rely on increasingly large amounts of video data, while the publicly available transcribed video datasets are limited in size. In this paper, for the first time, we study the potential of leveraging synthetic visual data for VSR. Our method, termed SynthVSR, substantially improves the performance of VSR systems with synthetic lip movements. The key idea behind SynthVSR is to leverage a speech-driven lip animation model that generates lip movements conditioned on the input speech. The speech-driven lip animation model is trained on an unlabeled audio-visual dataset and could be further optimized towards a pre-trained VSR model when labeled videos are available. As plenty of transcribed acoustic data and face images are available, we are able to generate large-scale synthetic data using the proposed lip animation model for semi-supervised VSR training. We evaluate the performance of our approach on the largest public VSR benchmark - Lip Reading Sentences 3 (LRS3). SynthVSR achieves a WER of $43.3\\%$ with only 30 hours of real labeled data,", + "bbox": [ + 75, + 568, + 472, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "outperforming off-the-shelf approaches using thousands of hours of video. The WER is further reduced to $27.9\\%$ when using all 438 hours of labeled data from LRS3, which is on par with the state-of-the-art self-supervised AV-HuBERT method. Furthermore, when combined with large-scale pseudo-labeled audio-visual data SynthVSR yields a new state-of-the-art VSR WER of $16.9\\%$ using publicly available data only, surpassing the recent state-of-the-art approaches trained with 29 times more non-public machine-transcribed video data (90,000 hours). Finally, we perform extensive ablation studies to understand the effect of each component in our proposed method.", + "bbox": [ + 500, + 544, + 893, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 739, + 630, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual speech recognition (VSR), also known as lip reading, is the task of recognizing speech content based on visual lip movements. VSR has a wide range of applications in real-world scenarios such as helping the hearing-impaired perceive human speech and improving automatic speech recognition (ASR) in noisy environments.", + "bbox": [ + 496, + 763, + 890, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VSR is a challenging task, as it requires capturing speech from high-dimensional spatio-temporal videos, while multiple words are visually ambiguous (e.g., \"world\" and", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Meta AI.", + "bbox": [ + 99, + 886, + 339, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18806", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"word\") in the visual streams. Recently, with the release of large-scale transcribed audio-visual datasets such as LRS2 [1] and LRS3 [2], deep neural networks have become the mainstream approach for VSR. However, even the largest public dataset for English VSR, LRS3, does not exceed 500 hours of transcribed video data. The lack of large-scale transcribed audio-visual datasets potentially results in VSR models which could only work in a laboratory environment i.e. limited word vocabulary and lip sources diversity [28].", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A common solution to this issue is to collect and annotate large-scale audio-visual datasets. For example, [41, 42] collected 90,000 hours of YouTube videos with user-uploaded transcriptions to achieve state-of-the-art performance on standard benchmarks. However, such a procedure is expensive and time-consuming, especially for most of the world's 7,000 languages [43]. If annotations are missing, the ASR can be used to generate the transcriptions automatically and this has been shown to be an effective approach to significantly improve VSR performance [28]. The other promising direction is to learn audio-visual speech representations from large amounts of parallel unlabeled audio-visual data in a self-supervised approach, and then fine-tune them on the limited labeled video dataset [43]. Nevertheless, publicly available video datasets are also limited and their usage may raise license-related concerns, barring their use in commercial applications.", + "bbox": [ + 75, + 227, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human perception of speech is inherently multimodal, involving both audition and vision [43]. ASR, which is a complementary task to VSR, has achieved impressive performance in recent years, with tens of thousands of hours of annotated speech datasets [4, 20, 33] available for large-scale training. It is intuitive to ask: Can we improve VSR with large amounts of transcribed acoustic-only ASR training data? The key to this question is to take advantage of recent advances in speech-driven visual generative models [49, 50]. By leveraging visual generative models, we can produce parallel synthetic videos for large-scale labeled audio datasets. Synthetic videos provide advantages such as having control over the target text and lip image as well as the duration of a generated utterance. To the best of our knowledge, the potential of leveraging synthetic visual data for improving VSR has never been studied in the literature.", + "bbox": [ + 75, + 484, + 472, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present SynthVSR, a novel semi-supervised framework for VSR. In particular, we first propose a speech-driven lip animation model that can generate synthetic lip movements video conditioned on the speech content. Next, the proposed lip animation model is used to generate synthetic video clips from transcribed speech datasets (e.g., Librispeech [33]) and human face datasets (e.g., CelebA [23]). Then, the synthetic videos together with the corresponding transcriptions are used in combi", + "bbox": [ + 75, + 727, + 472, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nation with the real video-text pairs (e.g., LRS3 [2]) for large-scale semi-supervised VSR training. The pipeline of SynthVSR is illustrated in Fig. 1. Unlike existing studies in exploiting unlabeled video data for VSR using methods such as pseudo-labeling [28] and self-supervised learning [43], we use the unlabeled audio-visual data to train a cross-modal generative model in order to bridge ASR training data and VSR. Furthermore, we propose to optimize the lip animation model towards a pre-trained VSR model when labeled videos are available. We empirically demonstrate that the semantically high level, spatio-temporal supervision signal from the pre-trained VSR model offers the lip animation model more accurate lip movements.", + "bbox": [ + 500, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SynthVSR achieves remarkable performance gains with labeled video data of different scales. We evaluate the performance of SynthVSR on the largest public VSR benchmark LRS3 with a Conformer-Transformer encoder-decoder VSR model [28]. In the low-resource setup using only 30 hours of labeled video data from LRS3 [2], our approach achieves a VSR WER of $43.3\\%$ , substantially outperforming the former VSR methods using hundreds or thousands of hours of video data for supervised training [1,27,37,44,52] and self-supervised learning [3,26,43]. Notably, we demonstrate the first successful attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. Using the complete 438 hours from LRS3 further improves WER to $27.9\\%$ which is on par with the state-of-the-art self-supervised method AV-HuBERT-LARGE [43] that uses external 1,759 hours of unlabeled audio-visual data, but with fewer model parameters. Furthermore, following a recent high-resource setup [25] which uses additional 2,630 hours of ASR pseudo-labeled publicly available audio-visual data, our proposed method yields a new state-of-the-art VSR WER of $16.9\\%$ , surpassing the former state-of-the-art approaches [41, 42] trained on 90,000 hours of non-public machine-transcribed data.", + "bbox": [ + 496, + 287, + 893, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we present extensive ablation studies to analyze where the improvement of SynthVSR comes from (e.g., the diversity of lip sources, the scale of ASR data). We also show considerable VSR improvement using synthetic video data derived from Text-To-Speech (TTS)-generated speech, indicating the great potential of our method for VSR.", + "bbox": [ + 496, + 635, + 893, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 739, + 640, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Speech Recognition. VSR has achieved remarkable progress with the success of deep learning and the availability of audio-visual datasets such as LRS2 [1] and LRS3 [2]. Progress was driven by adapting ASR approaches such as sequence-to-sequence models [11] and the Connectionist Temporal Classification (CTC) [5, 17] training objective. Recent studies have achieved noticeable results in VSR by using Transformer-based architecture [1], a convolutional variant [53], Conformer-based models [18], and", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Such as LRW [11] and LRS2 [1] datasets which are only permitted for the purpose of academic research.", + "bbox": [ + 75, + 875, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18807", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an attention-based feature pooling method [35]. The recent advances in VSR models are mainly dependent on leveraging increasingly large-scale transcribed non-public video datasets such as LSVSR (3,800 hours) [44] and YT31k (31,000 hours) [28]. Some recent works [41, 42] use 90,000 hours of non-publicly machine-transcribed video data to achieve state-of-the-art performance on standard benchmarks. Another popular trend is to leverage unlabeled audio-visual datasets such as AVSpeech [14] and VoxCeleb2 [10] with semi- or self-supervised learning. Using pre-trained ASR models to transcribe audio streams in unlabeled audio-visual datasets i.e., pseudo-labeling has demonstrated great potential to improve VSR performance. Such pseudo-labeled video data can be used for supervised training [28] or knowledge distillation [3]. Self-supervised approaches such as contrastive learning [12, 29, 32] and cross-modal feature prediction [26] have achieved significant improvement by exploiting unlabeled audio-visual data. AVHuBERT [43] is the state-of-the-art self-supervised method that predicts iteratively refined cluster assignments from masked audio-visual streams and achieves impressive VSR performance, especially when only limited labeled video data is available (e.g., 30 hours).", + "bbox": [ + 76, + 90, + 472, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech-Driven Facial Animation. Speech-driven facial animation aims to produce a talking face video conditioned on speech content. With the advances in deep visual generative models, speech-driven facial animation has significantly improved over the past years. Speech2Vid [9] is the first end-to-end neural network-based method for speech-driven facial animation. This method is optimized towards a pixel-level $L_{1}$ reconstruction loss which results in blurry generated frames. Recently, Generative Adversarial Networks (GANs) have made impressive progress [49,50] in facial animation, as they are able to generate sharp and highly textured facial frames. Initially, most of these methods synthesized frontal faces and focused on improving lip-sync [8, 50], preserving the identity of the target speaker [54]. As the generation of synthetic frontal faces improved, researchers began to focus on generating talking heads that could handle natural head movements [7, 55].", + "bbox": [ + 76, + 439, + 472, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning from Synthetic Supervision. Training with synthetic data has attracted increasing research attention in recent years. Synthetic data could offer improved data diversity [13], which is useful in many tasks where labeled data is limited, such as image classification [16], objection detection [34, 36], semantic segmentation [40, 51], human pose estimation [13], and ASR [15, 31]. Unlike previous work, we focus on a new multimodal problem: how to produce and leverage synthetic video data for improving VSR.", + "bbox": [ + 76, + 696, + 472, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SynthVSR", + "text_level": 1, + "bbox": [ + 76, + 844, + 191, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce SynthVSR, a novel semi-supervised VSR framework to improve the performance", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of VSR models using synthetic data, as shown in Fig. 1. We propose a speech-driven lip animation model trained on an audio-visual dataset, which is used to produce synthetic videos with large-scale transcribed speech and face datasets for scaling up VSR training. We will introduce the VSR model, the speech-driven lip animation model, and the semi-supervised VSR training next.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. VSR Model", + "text_level": 1, + "bbox": [ + 500, + 207, + 627, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The VSR model we use in this work is based on [27,28], which has achieved the state-of-the-art VSR performance on LRS3 [2] without the use of external data. The baseline VSR model is an encoder-decoder architecture. In particular, the encoder is comprised of two components, the visual front-end (a 3D convolutional layer followed by a ResNet-18 model [19,45]) and a Conformer [18] encoder. The decoder is based on the transformer architecture [47]. The baseline VSR model is trained end-to-end using a combination of the CTC loss [5,44] with an attention-based Cross-Entropy (CE) loss. Model details are described in the supplementary material.", + "bbox": [ + 496, + 229, + 893, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Speech-Driven Lip Animation", + "text_level": 1, + "bbox": [ + 500, + 422, + 767, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the recent advances in speech-driven facial animation [49, 50, 54, 55], we propose an approach for speech-driven lip animation that generates videos of talking mouth regions conditioned on speech utterances. The output space of the lip animation model is the same as the VSR input space. The proposed lip animation model is based on a temporal GAN [49, 50] with two discriminators. We further propose a VSR perceptual loss when labeled video data is available. The architecture of the speech-driven lip animation model is illustrated in the left part of Fig. 2. We will introduce each component in the next sections.", + "bbox": [ + 496, + 446, + 893, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generator. The generator $G$ is an encoder-decoder structure, as shown in the right part of the Fig. 2. The generator uses the first frame of a video clip, a speech clip, and a head rotation sequence as inputs. The head rotation is used as the additional condition which helps better model the lip movements as most training videos are not static. The speech clip is divided into overlapping $200\\mathrm{ms}$ chunks with a stride of $40~\\mathrm{ms}$ . The generator produces the corresponding video frame for each speech chunk. In the generator, an image encoder $E_{i}$ and a speech encoder $E_{s}$ are used to capture the visual information and speech context into latent embeddings $z_{i}$ and $z_{s}$ , respectively. The image encoder $E_{i}$ uses a stack of 2D convolutional layers to extract the visual embedding. The speech encoder $E_{s}$ comprises a stack of 1D convolutional layers followed by a stack of GRU layers. The head rotations are provided in the form of sequences of 3D rotation matrices [55] $z_{r}\\in \\mathbb{R}^{3\\times 3}$ with respect to the first frame. The three embeddings $z_{i},z_{s}$ and $z_{r}$ are concatenated and used to modulate the convolutional lay", + "bbox": [ + 496, + 613, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18808", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2af223e0bfa6526d1fa8498868f7d7c02a96e420565afa1521436be15bf89821.jpg", + "image_caption": [ + "Figure 2. Architecture of proposed speech-driven lip animation model. Left: GAN-based speech-driven lip animation model generating lip movements given a lip image, a speech utterance, and a rotation sequence; Right: structure of the generator in the lip animation model." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 529, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4ba1df8a8020a716329db45853c1667c4e5f63903a24aa036615d0d0bd59e65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 93, + 890, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ers in the frame decoder $D_{frame}$ , which is similar to StyleGAN2 [48]. Different from StyleGAN2, which generates frames from a learned constant input, we use the penultimate layer features of the image encoder $E_{i}$ as the input.", + "bbox": [ + 75, + 327, + 467, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discriminators. The speech-driven lip animation system has two discriminators: frame discriminator $D_{img}$ and sequence discriminator $D_{seq}$ . The frame discriminator is a stack of CNN layers that operates on the image frame level. The original first frame is concatenated channel-wise to the target frame to form the input of the frame discriminator. This helps enforce visual consistency. The sequence discriminator operates on the sequence level to ensure the temporal consistency of synthetic lip movements. The sequence discriminator uses spatio-temporal convolutions to encode the image sequence, followed by GRU layers to determine if the sequence is real or not. Specifically, the frame discriminator $D_{img}$ is trained on frames that are uniformly sampled from a video $v$ using a sampling function $S(v)$ . The first frame $v_1$ is fed to the $D_{img}$ as the condition. The input speech signal is $s$ . The adversarial loss of the $D_{img}$ is defined as:", + "bbox": [ + 75, + 387, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D i s c} ^ {i m g} = \\mathbb {E} _ {v} \\left[ \\log D _ {i m g} (S (v), v _ {1}) \\right] \\\\ + \\mathbb {E} _ {v, s} [ \\log (1 - D _ {i m g} (S (G (s, v _ {1})), v _ {1}) ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 651, + 403, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$D_{seq}$ operates on the entire sequence video $v$ . The adversarial loss of the $D_{seq}$ is defined as follows:", + "bbox": [ + 76, + 696, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {D i s c}} ^ {\\text {s e q}} = \\mathbb {E} _ {v} [ \\log D _ {\\text {s e q}} (v) ] + \\\\ \\mathbb {E} _ {v, s} [ \\log (1 - D _ {\\text {s e q}} (G (s, v _ {1})) ]. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 734, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VSR Perceptual Loss. We further propose to optimize the lip animation model towards a VSR perceptual loss if labeled video data is available. We first pre-train a VSR model as introduced in Sec. 3.1. The proposed VSR perceptual loss corresponds to a weighted sum of feature distances computed from the visual front-end and the Transformer decoder of the pre-trained VSR model for real and generated samples. We use $L_{1}$ norm to measure the visual embedding", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distance and Kullback-Leibler (KL) divergence to measure the logits distribution distance, respectively. The VSR perceptual loss is obtained by:", + "bbox": [ + 496, + 327, + 890, + 372 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V S R} = \\lambda_ {\\text {v i s u a l}} \\left\\| z _ {f} ^ {r} - z _ {f} ^ {s} \\right\\| _ {1} + \\lambda_ {\\text {l o g i t s}} \\operatorname {K L} \\left(\\hat {y} ^ {r}, \\hat {y} ^ {s}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 377, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{f}^{r}$ and $z_{f}^{s}$ are the VSR front-end visual features of the real and synthetic video, respectively, $\\hat{y}^{r}$ and $\\hat{y}^{s}$ is the VSR predicted logits distribution of real and synthetic video, respectively, $\\lambda_{\\text{visual}}$ and $\\lambda_{\\text{logits}}$ control the weights of these two perceptual losses. The VSR model is frozen during the lip animation model training.", + "bbox": [ + 496, + 407, + 890, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objectives. The speech-driven lip animation model is trained using a combination of a reconstruction loss, adversarial losses, and a VSR perceptual loss. The reconstruction loss is computed based on the $L_{1}$ distance between the generated video $\\hat{v}$ and ground truth video $v$ :", + "bbox": [ + 496, + 498, + 890, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\| v - \\hat {v} \\| _ {1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 585, + 890, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall training loss for the lip animation model is:", + "bbox": [ + 500, + 612, + 864, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {A n i m a t i o n}} = \\lambda_ {\\text {d i s c}} ^ {\\text {i m g}} \\mathcal {L} _ {\\text {d i s c}} ^ {\\text {i m g}} + \\lambda_ {\\text {d i s c}} ^ {\\text {s e q}} \\mathcal {L} _ {\\text {d i s c}} ^ {\\text {s e q}} \\tag {5} \\\\ + \\lambda_ {r e c} \\mathcal {L} _ {r e c} + \\mathcal {L} _ {V S R}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 637, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{disc}^{img},\\lambda_{disc}^{seq}$ , and $\\lambda_{rec}$ represent the coefficient of the adversarial loss of frame discriminator $D_{img}$ , the adversarial loss of sequence discriminator $D_{seq}$ , and the reconstruction loss, respectively.", + "bbox": [ + 496, + 686, + 890, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semi-Supervised VSR with Synthetic Data", + "text_level": 1, + "bbox": [ + 500, + 756, + 861, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a typical supervised setting, a VSR model is trained from the labeled dataset $\\mathcal{D}_{real} = \\{(v_i^r,y_i^r)\\}_{i = 1}^{n_r}$ , where $n_r$ is the number of paired video clips $v_{i}^{r}$ and its transcriptions $y_{i}^{r}$ . In SynthVSR, we first train a speech-driven lip animation model on an unlabeled audio-visual dataset $\\mathcal{D}_{av} = \\{(v_i^{av},s_i^{av})\\}_{i = 1}^{n_{av}}$ , where $n_{av}$ is the number of paired video clips $v_{i}^{av}$ and speech utterances $s_i^{av}$ . Furthermore, we can use the trained generator $G$ to generate synthetic video", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18809", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from a labeled speech dataset $\\mathcal{D}_s = \\{(s_i^s,y_i^s)\\}_{i = 1}^{n_s}$ and a cropped lips dataset $\\mathcal{D}_f = \\{c_i^f\\}_{i = 1}^{n_f}$ , where $n_{s}$ is the number of paired speech clips $s_i^s$ and its transcriptions $y_{i}^{s}$ and $n_f$ is the number of lip images $c_{i}^{f}$ . For each speech clip $s_i^s$ , the generator $G$ generates its parallel synthetic video $\\hat{v}_i^s$ :", + "bbox": [ + 76, + 90, + 468, + 170 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG \\left(s _ {i} ^ {s}, S \\left(\\mathcal {D} _ {f}\\right)\\right) \\mapsto \\hat {v} _ {i} ^ {s}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 179, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S(\\cdot)$ is a uniform sampling function. Then, the synthetic dataset $\\mathcal{D}_{\\text{synth}} = \\{\\hat{v}_i^s, y_i^s\\}_{i=1}^{n_s}$ is obtained, where in general $n_s >> n_r$ as there are far more annotated speech datasets (e.g., Librispeech [33]) than labeled video datasets. Finally, the VSR model is trained from $\\mathcal{D}_{\\text{real}} \\cup \\mathcal{D}_{\\text{synth}}$ , where the rich visual and text label information in $\\mathcal{D}_{\\text{synth}}$ offers significant performance improvement for VSR.", + "bbox": [ + 76, + 205, + 468, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 324, + 209, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 348, + 179, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use several public datasets in this work. (1) Audiovisual datasets: LRS3 [2], AVSpeech [14] and VoxCeleb2 [10]; (2) Speech datasets: Librispeech [33], TED-LIUM 3 [20], Common Voice [4]; (3) Facial dataset: CelebA [23].", + "bbox": [ + 76, + 371, + 468, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS3. We conduct experiments on the LRS3 [2] dataset, which is the largest public benchmark for English VSR containing 438.9 hours of video clips from TED talks (408, 30, and 0.9 hours in the pre-training, training-validation, and test set, respectively).", + "bbox": [ + 76, + 433, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVSpeech & VoxCeleb2. A recent work [25] uses public language classifier [46] and ASR models [6] to filter for English content and obtain pseudo-transcriptions for two multilingual audio-visual datasets: AVSpeech [14] and VoxCeleb2 [10], containing 1,323 hours from AVSpeech and 1,307 hours from VoxCeleb2. We follow the same data setting for our high-resource setting. Furthermore, we filter out AVSpeech videos with large jitter resulting in 933 hours of videos overall for speech-driven lip animation training.", + "bbox": [ + 76, + 508, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets for Speech-Driven Lip Animation Training. The lip animation model is trained on a combination of LRS3 (pre-training and training-validation splits) and the English subset (933 hours) of AVSpeech datasets.", + "bbox": [ + 76, + 643, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets for Synthetic Data Generation. We use Librispeech [33], TED-LIUM 3 [20], Common Voice (English split) [4] datasets as the speech sources. We split Librispeech audio clips into segments of less than 6 seconds based on the provided silence annotations. For TED-LIUM 3 and Common Voice, we filter out the audio clips which are longer than 20s. As a result, 944 hours, 465 hours, and 2,243 hours of speech data are obtained for Librispeech, TED-LIUM 3, and Common Voice, respectively. We use the CelebA [23] dataset as a source of lip images, which has 202,599 face images and 10,177 identities. For each speech clip, we randomly sample one image from CelebA to generate one synthetic video. We use a static rotation matrix", + "bbox": [ + 76, + 704, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in the generation process. In total, 3,652 hours of synthetic video clips are generated for scaling up VSR training.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Data Processing", + "text_level": 1, + "bbox": [ + 500, + 132, + 660, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train a mouth detection module using the Faster R-CNN object detection architecture [38] on the CelebA [23] dataset. We leverage the mouth corner positions provided in the dataset to identify the area around the mouth in each video frame. With this network, we locate and extract a $96 \\times 96$ bounding box around the mouth, which we subsequently convert to grayscale. We use the same video pre-processing methods for lip animation and VSR models. For the text vocabulary, we use SentencePiece [22] subword units with a vocabulary size of 5,000.", + "bbox": [ + 496, + 155, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 316, + 718, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VSR Model. We consider two model configurations: (1) Conformer-BASE (250M) with 12-layer Conformer encoder, 6-layer Transformer decoder, 768 input dimensions, 3,072 feed-forward dimensions, and 16 attention heads; (2) Conformer-LARGE (783M) with 24-layers Conformer encoder, 9-layer Transformer decoder, 1024 input dimensions, 4,096 feed-forward dimensions, and 16 attention heads. For each configuration, the encoder and decoder have the same dimensions and attention heads.", + "bbox": [ + 496, + 340, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use horizontal flipping, random cropping, and adaptive time masking as the VSR data augmentation methods [28]. We train the BASE and LARGE models for 75 epochs on 64 A100-GPUs with the AdamW [24] optimizer, a cosine learning rate scheduler, and a warm-up of 5 epochs. The peak learning rate is $1 \\times 10^{-3}$ and $8 \\times 10^{-4}$ for the BASE and LARGE models respectively. The number of frames in each batch is limited to 2,400 and 1,600 frames for the BASE and LARGE models, respectively. Following [25, 28], a pre-trained transformer-based language model is used in the VSR decoding stage. The training configuration is the same as that used in the previous work [25].", + "bbox": [ + 496, + 477, + 892, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Speech-Driven Lip Animation. For each training video clip, we sample a sequence of 75 frames that corresponds to 3 seconds. The first frame of the sampled video sequence is used to drive the lip animation. We train the lip animation model for 70 epochs on 32 A100-GPUs with a batch size of 3 per GPU. The details of the module structure in the lip animation model are described in the supplementary material. Adam [21] optimizer is used to train the lip animation model with the learning rate $1 \\times 10^{-4}$ for the generator and the frame discriminator. Sequence discriminator uses a smaller learning rate of $1 \\times 10^{-5}$ . The weights for each loss term are $\\lambda_{disc}^{img} = 1$ , $\\lambda_{disc}^{seq} = 0.2$ , $\\lambda_{rec} = 300$ in Eq. (5).", + "bbox": [ + 496, + 657, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider five lip animation model configurations. The lip animation model without VSR perceptual loss is referred to LAM-Baseline. The lip animation model with the BASE VSR model trained on LRS3 is referred to as LAM-", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18810", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/aa9303fe78c811b39c57d990b61559233b1f6b2ffc5a2540916510c56dc37e02.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
Afouras et al. [3]CNN✓595‡334-59.8
Ren et al. [37]TransformerX818‡--59.0
Afouras et al. [1]Transformer✓1,519†‡--58.9
Xu et al. [52]RNNX595‡--57.8
Shillingford et al. [44]RNN✓3,886†--55.1
Ma et al. [26]TransformerX4331,759-49.6*
Ma et al. [27]Conformer✓438--46.9
AV-HuBERT-BASE [43]TransformerX301,759-46.1
SynthVSRConformer-BASEX30--104.0
X--3,652100.3
X30-3,65244.7
✓30-3,65243.3
", + "bbox": [ + 81, + 87, + 893, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Experimental results of low-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. †Includes non-publicly available data. ‡Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours. *Result taken from [43].", + "bbox": [ + 75, + 287, + 892, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LRS3-VSR-VL, with the $\\lambda_{\\text{visual}} = 250$ and $\\lambda_{\\text{logits}} = 10$ in Eq. (3). Two variants LAM-LRS3-VSR-L and LAM-LRS3-VSR-V are further designed with the $\\lambda_{\\text{visual}} = 0$ and $\\lambda_{\\text{logits}} = 0$ , respectively. These two model configurations are used in ablation studies. Last, the lip animation model with the BASE VSR model trained on LRS3 and 2,630 hours of pseudo-labeled AVSpeech and VoxCeleb2 is referred to as LAM-LRS3-AVoX-VSR, with the $\\lambda_{\\text{visual}} = 500$ and $\\lambda_{\\text{logits}} = 10$ .", + "bbox": [ + 75, + 343, + 472, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Low-Resource Labeled Data Setting", + "text_level": 1, + "bbox": [ + 76, + 494, + 390, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The LAM-Baseline model is used to generate 3,652 hours of synthetic data. We use 30 hours of LRS3 (training-validation) to evaluate the VSR performance when labeled data is scarce. We conduct experiments on the BASE VSR model. The results are reported in Tab. 1. As supervised VSR training from scratch with long sequences often poses optimization problems [1, 11, 28], we first use 30 hours of LRS3 and the 944 hours of Librispeech synthetic data to pre-train a VSR model with the same architecture as [28], then we take the weights of the pre-trained visual front-end to train the BASE VSR model. The pre-training details are described in the supplementary material.", + "bbox": [ + 75, + 520, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training with synthetic data leads to dramatic performance improvement when only 30 hours of labeled data is available. Using 3,652 hours of synthetic data and 30 hours of LRS3 labeled data for training, the BASE model achieves the WER $43.3\\%$ (44.7% w/o. language model). Training from 30 hours of LRS3 results in a poor WER of $104\\%$ . Our proposed method significantly outperforms the former approaches using hundreds or thousands of data for supervised training [1,27,37,44,52] and self-supervised learning [3, 26]. Furthermore, our method performs better than the AV-HuBERT-BASE [43] model (46.1%) that uses 30 hours of LRS3 and 1,759 hours of unlabeled audio-visual data for self-supervised learning. Notably, we show", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the first attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. In addition, we observe that training only from synthetic data performs poorly (100.3%), which may be caused by the domain mismatch between the real and synthetic data.", + "bbox": [ + 496, + 343, + 890, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. LRS3 Labeled Data Setting", + "text_level": 1, + "bbox": [ + 498, + 431, + 746, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this setting, we report the results when using the full 438 hours of LRS3. Experiments are conducted on the BASE VSR model. To avoid the optimization problem, we initialize the weights of the visual front-end from a publicly available Conformer-based model [28] pre-trained on LRS3 using curriculum learning. The results are shown in Tab. 2.", + "bbox": [ + 496, + 455, + 890, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our BASE model achieves WER $36.7\\%$ when using 438 hours of LRS3 data for training. We generate 3,652 hours of synthetic data using the LAM-LRS3-VSR-VL model. Using 3,652 hours of synthetic data and 438 hours of LRS3 labeled data, the BASE VSR model achieves the WER $27.9\\%$ $(28.4\\% \\mathrm{w / o}$ language model, corresponding to a WER reduction of $8.3\\%$ . Our method outperforms three recent approaches using 31,000 $(33.6\\%)$ , 1,459 $(31.5\\%)$ , and 2,679 $(30.7\\%)$ hours of labeled data, respectively. When compared with the state-of-the-art self-supervised method AV-HuBERT [43] that uses additional 1,759 hours of unlabeled audio-visual data, our method outperforms the AV-HuBERT-BASE model $(34.8\\%)$ by a large margin. Our method slightly performs better than the AV-HuBERT-LARGE model $(28.6\\%)$ , but with fewer model parameters (our BASE model 250M vs AV-HuBERT-LARGE 390M). Note that we compare with the AV-HuBERT results without self-training as we do not use the pseudo-labeled 933 hours of AVSpeech subset for VSR training.", + "bbox": [ + 496, + 547, + 892, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. High-Resource Labeled Data Setting", + "text_level": 1, + "bbox": [ + 500, + 845, + 816, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate the scalability of SynthVSR: when using machine-transcribed AVSpeech [14] and VoxCeleb2", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18811", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d6ba7342a6712704b02a71e61c6bf671f3dc3a31523440d47a6e6790c75a5082.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
AV-HuBERT-BASE [43]TransformerX4331,759-34.8
Makino et al. [30]TransformerX31,000†--33.6
Ma et al. [28]Conformer✓1,459‡--31.5
Prajwal et al. [35]Transformer✓2,676†--30.7
AV-HuBERT-LARGE [43]TransformerX4331,759-28.6
AV-HuBERT-LARGE w. Self-Training [43]TransformerX4331,759-26.9
Auto-AVSR [25]Conformer✓3,448‡--19.1
Serdyuk et al. [42]TransformerX90,000†--25.9
Serdyuk et al. [41]TransformerX90,000†--17.0
SynthVSRConformer-BASEX438--36.7
X438-3,65228.4
✓438-3,65227.9
X3,068--21.2
X3,068-3,65219.4
✓3,068-3,65218.7
SynthVSRConformer-LARGEX3,068-3,65218.2
✓3,068-3,65216.9
", + "bbox": [ + 81, + 87, + 893, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[10] as additional training data. We generate 3,652 hours of synthetic data with LAM-LRS3-AVoX-VSR and conduct experiments on BASE and LARGE models. The visual front-end is initialized from a pre-trained model which is the same as that used in the LRS3 labeled data setting.", + "bbox": [ + 75, + 378, + 468, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first train the BASE model with 438 hours of labeled LRS3 and 2,630 hours of pseudo-labeled data, resulting in a strong VSR baseline with the WER $21.2\\%$ . By using additional 3,652 hours of synthetic data, the WER of the BASE model improves to $18.7\\%$ (19.4% w./o. language model), which outperforms [25] that uses additional labeled dataset LRS2 (223 hours) and LRW (157 hours) for training. Although the VSR model has seen a large amount of labeled data, and the speech-driven lip animation model is trained from part of the VSR training data, synthetic data can still lead to considerable performance gains. Furthermore, increasing the model size from BASE to LARGE results in better VSR performance with the WER of $16.9\\%$ (18.2% w./o. language model), which is the current state-of-the-art performance on LRS3, with publicly available data only.", + "bbox": [ + 75, + 455, + 470, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Why SynthVSR Improves VSR?", + "text_level": 1, + "bbox": [ + 76, + 694, + 362, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive ablation studies to understand the impact of VSR perceptual loss, the diversity of lip sources, and the scale of ASR data for SynthVSR. In addition, we assess the domain mismatch between the real and synthetic data using VSR models. We perform experiments on the LRS3 labeled data setting with the Conformer-BASE VSR model.", + "bbox": [ + 75, + 718, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of VSR Perceptual Loss. We experiment with four lip animation model variants: LAM-Baseline, LAM-LRS3-VSR-V, LAM-LRS3-VSR-L, LAM-LRS3-VSR-VL to generate 944 hours of synthetic data from Librispeech, respectively. The WER of the BASE model trained with only 438", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hours of LRS3 data is $36.7\\%$ (see in Tab. 2), the results are shown in Tab. 3. Using 944 hours of synthetic data generated from the LAM-Baseline model, the WER improves to $32.2\\%$ . The individual visual VSR loss and linguistic VSR loss improve the WER to $31.7\\%$ and $31.4\\%$ , respectively. The linguistic VSR loss performs slightly better than the visual one. Further, the combination of visual and linguistic VSR loss further improves the WER to $30.8\\%$ , which indicates that the spatio-temporal visual and high-level semantic knowledge from a pre-trained VSR model could offer the lip animation model more accurate lip movements.", + "bbox": [ + 496, + 378, + 893, + 547 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3476941763988f754577569515e39142be0ac2c1448de6d1f54f6e0f60f656ff.jpg", + "table_caption": [ + "Table 2. Experimental results of LRS3 & high-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. ${}^{ \\dagger }$ Includes non-publicly available data. ${}^{§}$ Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours." + ], + "table_footnote": [], + "table_body": "
Lip animation modelWER (%)
LAM-Baseline32.2
LAM-LRS3-VSR-V31.7
LAM-LRS3-VSR-L31.4
LAM-LRS3-VSR-VL30.8
", + "bbox": [ + 580, + 558, + 818, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on the proposed VSR perceptual loss.", + "bbox": [ + 506, + 642, + 883, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Increasing Lip Sources Diversity. To analyze the impact of the diversity of lip sources we generate 944 hours of synthetic data generated from Librispeech using two target face sources: CelebA and LRS3 (pre-training). Images from LRS3 (pre-training) do not introduce any new lip images for VSR training. The results are reported in Tab. 4. Using the synthetic data with LRS3 lip images, the BASE model achieves the WER $32.1\\%$ which is $1.3\\%$ worse than that with CelebA lip images, which means the external lip images from CelebA offer better WER improvement.", + "bbox": [ + 496, + 672, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the Scale of Speech Corpus. Using synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice with LAM-LRS3-VSR-VL model synthetic data, the LRS3 (test) WER improves from $36.7\\%$ to $30.9\\%$ , $30.8\\%$ , and $30.1\\%$ , respectively (see in Tab. 5). When combining", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18812", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a61f5cd9242931ed45566d0bf63cfdb9f28852a45e61ffab143baf844f4cc28a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Lip sourceUnique imagesWER (%)
LRS3-pretrain5,09032.1
CelebA202,59930.8
", + "bbox": [ + 112, + 88, + 437, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all synthetic data together (3,652 hours) for training, the WER improves to $28.4\\%$ . We observe that richer speech corpora can offer larger performance improvements. Furthermore, we find that the improvement from scaling speech corpora is more pronounced than when adding VSR perceptual loss and lip sources diversity, suggesting that our proposed method benefits greatly from large speech datasets.", + "bbox": [ + 75, + 202, + 467, + 308 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/63fb2aa6d18145ba3e3aa94f9ed41448d81b5bcfb062f3e7b898451168d25c5f.jpg", + "table_caption": [ + "Table 4. Ablation studies on the impact of lip sources diversity. LRS3-pretrain indicates using the LRS3 (pre-training) images." + ], + "table_footnote": [], + "table_body": "
Training dataHoursWER (%)
LRS343836.7
LRS3 + TED-Synth90330.9
LRS3 + LBS-Synth1,38230.8
LRS3 + CV-Synth2,68130.1
LRS3 + [TED + LBS + CV]-Synth4,09028.4
", + "bbox": [ + 93, + 319, + 457, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assessment of the Domain Mismatch Using VSR. We use 944 hours of Librispeech synthetic data generated from LAM-Baseline and LAM-LRS3-VSR-VL for VSR training. We conduct an assessment on the LRS3 real and synthetic test sets, the synthetic test set is obtained by passing the speech and the first frame in the real test set to the lip animation model. The results are illustrated in Fig. 3. By introducing the VSR perceptual loss, the WER on the synthetic test set is reduced from $106.8\\%$ to $49.7\\%$ , which indicates the VSR loss minimizes the domain gap between real and synthetic data. After training on synthetic data, the WER reduces for both real and synthetic data. Specifically, the WER is improved substantially from $106.8\\%$ to $36.0\\%$ and from $49.7\\%$ to $16.8\\%$ on synthetic test sets generated using LAM-Baseline and LAM-LRS3-VSR-VL, respectively. We observe that domain mismatch can be further reduced by combining real and synthetic data for VSR training.", + "bbox": [ + 75, + 474, + 467, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8. SynthVSR with TTS-Generated Speech", + "text_level": 1, + "bbox": [ + 76, + 741, + 415, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the real world, text corpora (e.g., Wikipedia) are more accessible than transcribed speech corpora. To explore the potential of using TTS-generated speech for SynthVSR, we use the off-the-shelf TTS model FastSpeech 2 [39] to generate 944 hours of synthetic speech from Librispeech transcriptions, which is further used to generate synthetic video clips using LAM-LRS3-VSR-VL model. Using these 944 hours of synthetic data to train the BASE model under the LRS3 labeled data setting, the WER is improved", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38b8a320d1f90e152d48dd0fd74848a89bf715be9b4abfdbd87789a0310d57c6.jpg", + "image_caption": [ + "Figure 3. Assessment of the domain mismatch between real and synthetic video data using VSR models trained with only real data (VSR-Real, black) and real data together with synthetic data (VSR-Real-Synth, red), respectively. Synth-Test and Real-Test refer to the synthetic and real LRS3 (test) sets." + ], + "image_footnote": [], + "bbox": [ + 506, + 80, + 846, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from $36.7\\%$ to $32.9\\%$ , as shown in Tab. 6. Although TTS-generated speech is not as natural as the original LibriSpeech data, the improvement of using synthetic data is still decent $(3.8\\%)$ , indicating the further potential of SynthVSR in real-world applications where labeled video data is sparse, such as the healthcare industry.", + "bbox": [ + 496, + 325, + 890, + 416 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d5a5b2164de4b7444917cbfc3d35bdb10934126f1bcb9a59f5a51a32be875583.jpg", + "table_caption": [ + "Table 5. Ablation studies on different speech corpus. TED-Synth, LBS-Synth, and CV-Synth indicates the synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice, respectively." + ], + "table_footnote": [], + "table_body": "
Training dataHoursWER (%)
LRS343836.7
LRS3 + LBS-Synth1,38230.8
LRS3 + TTS-LBS-Synth1,38232.9
", + "bbox": [ + 535, + 429, + 859, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Experiments results of using the synthetic speech generated synthetic video (TTS-LBS-Synth) for VSR training.", + "bbox": [ + 496, + 503, + 890, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations & Societal Impact", + "text_level": 1, + "bbox": [ + 500, + 561, + 774, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The excellent performance improvement of SynthVSR comes at the cost of high computational demands during training, as we use additional large-scale synthetic video data. Also, although SynthVSR has achieved state-of-the-art VSR performance, the LRS3 test set is relatively limited (0.9 hours), which is from TED talks, while real-world videos are more challenging (e.g., egocentric videos). SynthVSR has many positive real-world applications, such as helping the hearing-impaired or people with aphonia with everyday communication.", + "bbox": [ + 496, + 588, + 890, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 753, + 617, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a semi-supervised method for VSR enhanced with synthetic lip movements. The speech-driven lip animation model is proposed to generate synthetic video data from labeled speech datasets and face images for scaling up VSR. Our method achieves state-of-the-art results on LRS3, outperforming prior work trained on more labeled or unlabeled real video data. Our work fosters future research on generating and exploiting synthetic visual data for VSR.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18813", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2, 6", + "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: A large-scale dataset for visual speech recognition. arXiv:1809.00496, 2018. 2, 3, 5", + "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2143-2147, 2020. 2, 3, 6", + "[4] Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Conference on Language Resources and Evaluation, 2020. 2, 5", + "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. LipNet: Sentence-level lipreading. arXiv:1611.01599, 2016. 2, 3", + "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 5", + "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 3", + "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision, pages 520-535, 2018. 3", + "[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? In British Machine Vision Conference, 2017. 3", + "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 3, 5, 7", + "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, pages 87-103. Springer, 2016. 2, 6", + "[12] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: Learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 3", + "[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3D human pose estimation: Motion to the rescue. In Advances in Neural Information Processing Systems, volume 32, 2019. 3", + "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv:1804.03619, 2018. 3, 5, 6" + ], + "bbox": [ + 78, + 119, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Amin Fazel, Wei Yang, Yulan Liu, Roberto Barra-Chicote, Yixiong Meng, Roland Maas, and Jasha Droppo. SynthASR: Unlocking synthetic data for speech recognition. arXiv:2106.07803, 2021. 3", + "[16] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv:2007.04954, 2020. 3", + "[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine learning, pages 369-376, 2006. 2", + "[18] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. In INTER-SPEECH, 2020. 2, 3", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3", + "[20] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In International Conference on Speech and Computer, pages 198-208, 2018. 2, 5", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5", + "[22] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018. 5", + "[23] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision, December 2015. 2, 5", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 5", + "[25] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-visual speech recognition with automatic labels. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2023. 2, 5, 7", + "[26] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In IN-TERSPEECH, 2021. 2, 3, 6", + "[27] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 7613-7617, 2021. 2, 3, 6", + "[28] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nature Machine Intelligence, 2022. 2, 3, 5, 6, 7", + "[29] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song." + ], + "bbox": [ + 501, + 92, + 890, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18814", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Contrastive learning of global and local audio-visual representations. arXiv:2104.05418, 2021.3", + "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE Automatic Speech Recognition and Understanding Workshop, pages 905–912, 2019. 7", + "[31] Sepand Mavandadi, Tara N Sainath, Kevin Hu, and Zelin Wu. A deliberation-based joint acoustic and text decoder. In INTERSPEECH, 2021. 3", + "[32] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging uni-modal self-supervised learning for multimodal audio-visual speech recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 3", + "[33] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 5206-5210, 2015. 2, 5", + "[34] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3D models. In Proceedings of the IEEE International Conference on Computer Vision, pages 1278-1286, 2015. 3", + "[35] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 3, 7", + "[36] Aayush Prakash, Shaad Boochoon, Mark Brophy, David Acuna, Eric Cameracci, Gavriel State, Omer Shapira, and Stan Birchfield. Structured domain randomization: Bridging the reality gap by context-aware synthetic data. In 2019 International Conference on Robotics and Automation, pages 7249-7255, 2019. 3", + "[37] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6", + "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5", + "[39] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv:2006.04558, 2020. 8", + "[40] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3234-3243, 2016. 3", + "[41] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition. In INTERSPEECH, 2022. 2, 3, 7", + "[42] Dmitriy Serdyuk, Olivier Siohan, and Otavio de Pinho Forin Braga. Audio-visual speech recognition is worth $32 \\times 32 \\times 8$ voxels. In IEEE Automatic Speech Recognition and Under" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "standing Workshop, 2021. 2, 3, 7", + "[43] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In International Conference on Learning Representations, 2022. 2, 3, 6, 7", + "[44] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. In INTERSPEECH, 2019. 2, 3, 6", + "[45] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. arXiv:1703.04105, 2017. 3", + "[46] Jörgen Valk and Tanel Alumäe. Voxlingua107: A dataset for spoken language recognition. In 2021 IEEE Spoken Language Technology Workshop, pages 652-658, 2021. 5", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3", + "[48] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. StyleGAN2 distillation for feed-forward image manipulation. In European Conference on Computer Vision, pages 170-186. Springer, 2020. 4", + "[49] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven facial animation with temporal GANs. In British Machine Vision Conference, 2018. 2, 3", + "[50] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with GANs. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2, 3", + "[51] Zhonghao Wang, Mo Yu, Yunchao Wei, Rogerio Feris, Jinjun Xiong, Wen-mei Hwu, Thomas S Huang, and Honghui Shi. Differential treatment for stuff and things: A simple unsupervised domain adaptation method for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12635-12644, 2020. 3", + "[52] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14433-14442, 2020. 2, 6", + "[53] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 2", + "[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 3", + "[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18815", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_model.json b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8fe380fc5cd5565ab3538f5e25e9c8c2a907bc1b --- /dev/null +++ b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "SynthVSR: Scaling Up Visual Speech Recognition With Synthetic Supervision" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.18, + 0.923, + 0.217 + ], + "angle": 0, + "content": "Xubo Liu\\(^{1*}\\), Egor Lakomkin\\(^{2}\\), Konstantinos Vougioukas\\(^{2}\\), Pingchuan Ma\\(^{2}\\), Honglie Chen\\(^{2}\\), Ruiming Xie\\(^{2}\\), Morrie Doulaty\\(^{2}\\), Niko Moritz\\(^{2}\\), Jachym Kolar\\(^{2}\\), Stavros Petridis\\(^{2}\\), Maja Pantic\\(^{2}\\), Christian Fuegen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.224, + 0.645, + 0.243 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Surrey \\(^{2}\\)Meta AI" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.258, + 0.342, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.258, + 0.637, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.258, + 0.89, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.447, + 0.893, + 0.531 + ], + "angle": 0, + "content": "Figure 1. Scaling up visual speech recognition with synthetic supervision (SynthVSR): we propose SynthVSR, a semi-supervised framework that can substantially improve the performance of VSR models by using synthetic lip movements. Firstly, we introduce a speech-driven lip animation model that generates lip movement videos conditioned on input lip images and speech utterances (left). Secondly, we generate large-scale synthetic videos using transcribed speech datasets and lip images. The combination of synthetic videos and their corresponding speech transcriptions constitutes the synthetic dataset (centre). Finally, we conduct semi-supervised VSR training with synthetic and real datasets. Our method substantially improves the performance of VSR models with large-scale synthetic data (right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.543, + 0.314, + 0.559 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.871 + ], + "angle": 0, + "content": "Recently reported state-of-the-art results in visual speech recognition (VSR) often rely on increasingly large amounts of video data, while the publicly available transcribed video datasets are limited in size. In this paper, for the first time, we study the potential of leveraging synthetic visual data for VSR. Our method, termed SynthVSR, substantially improves the performance of VSR systems with synthetic lip movements. The key idea behind SynthVSR is to leverage a speech-driven lip animation model that generates lip movements conditioned on the input speech. The speech-driven lip animation model is trained on an unlabeled audio-visual dataset and could be further optimized towards a pre-trained VSR model when labeled videos are available. As plenty of transcribed acoustic data and face images are available, we are able to generate large-scale synthetic data using the proposed lip animation model for semi-supervised VSR training. We evaluate the performance of our approach on the largest public VSR benchmark - Lip Reading Sentences 3 (LRS3). SynthVSR achieves a WER of \\(43.3\\%\\) with only 30 hours of real labeled data," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.545, + 0.895, + 0.726 + ], + "angle": 0, + "content": "outperforming off-the-shelf approaches using thousands of hours of video. The WER is further reduced to \\(27.9\\%\\) when using all 438 hours of labeled data from LRS3, which is on par with the state-of-the-art self-supervised AV-HuBERT method. Furthermore, when combined with large-scale pseudo-labeled audio-visual data SynthVSR yields a new state-of-the-art VSR WER of \\(16.9\\%\\) using publicly available data only, surpassing the recent state-of-the-art approaches trained with 29 times more non-public machine-transcribed video data (90,000 hours). Finally, we perform extensive ablation studies to understand the effect of each component in our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.74, + 0.631, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Visual speech recognition (VSR), also known as lip reading, is the task of recognizing speech content based on visual lip movements. VSR has a wide range of applications in real-world scenarios such as helping the hearing-impaired perceive human speech and improving automatic speech recognition (ASR) in noisy environments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "VSR is a challenging task, as it requires capturing speech from high-dimensional spatio-temporal videos, while multiple words are visually ambiguous (e.g., \"world\" and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.341, + 0.9 + ], + "angle": 0, + "content": "*Work done during an internship at Meta AI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "\"word\") in the visual streams. Recently, with the release of large-scale transcribed audio-visual datasets such as LRS2 [1] and LRS3 [2], deep neural networks have become the mainstream approach for VSR. However, even the largest public dataset for English VSR, LRS3, does not exceed 500 hours of transcribed video data. The lack of large-scale transcribed audio-visual datasets potentially results in VSR models which could only work in a laboratory environment i.e. limited word vocabulary and lip sources diversity [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.485 + ], + "angle": 0, + "content": "A common solution to this issue is to collect and annotate large-scale audio-visual datasets. For example, [41, 42] collected 90,000 hours of YouTube videos with user-uploaded transcriptions to achieve state-of-the-art performance on standard benchmarks. However, such a procedure is expensive and time-consuming, especially for most of the world's 7,000 languages [43]. If annotations are missing, the ASR can be used to generate the transcriptions automatically and this has been shown to be an effective approach to significantly improve VSR performance [28]. The other promising direction is to learn audio-visual speech representations from large amounts of parallel unlabeled audio-visual data in a self-supervised approach, and then fine-tune them on the limited labeled video dataset [43]. Nevertheless, publicly available video datasets are also limited and their usage may raise license-related concerns, barring their use in commercial applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.473, + 0.728 + ], + "angle": 0, + "content": "Human perception of speech is inherently multimodal, involving both audition and vision [43]. ASR, which is a complementary task to VSR, has achieved impressive performance in recent years, with tens of thousands of hours of annotated speech datasets [4, 20, 33] available for large-scale training. It is intuitive to ask: Can we improve VSR with large amounts of transcribed acoustic-only ASR training data? The key to this question is to take advantage of recent advances in speech-driven visual generative models [49, 50]. By leveraging visual generative models, we can produce parallel synthetic videos for large-scale labeled audio datasets. Synthetic videos provide advantages such as having control over the target text and lip image as well as the duration of a generated utterance. To the best of our knowledge, the potential of leveraging synthetic visual data for improving VSR has never been studied in the literature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.473, + 0.865 + ], + "angle": 0, + "content": "In this work, we present SynthVSR, a novel semi-supervised framework for VSR. In particular, we first propose a speech-driven lip animation model that can generate synthetic lip movements video conditioned on the speech content. Next, the proposed lip animation model is used to generate synthetic video clips from transcribed speech datasets (e.g., Librispeech [33]) and human face datasets (e.g., CelebA [23]). Then, the synthetic videos together with the corresponding transcriptions are used in combi" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "nation with the real video-text pairs (e.g., LRS3 [2]) for large-scale semi-supervised VSR training. The pipeline of SynthVSR is illustrated in Fig. 1. Unlike existing studies in exploiting unlabeled video data for VSR using methods such as pseudo-labeling [28] and self-supervised learning [43], we use the unlabeled audio-visual data to train a cross-modal generative model in order to bridge ASR training data and VSR. Furthermore, we propose to optimize the lip animation model towards a pre-trained VSR model when labeled videos are available. We empirically demonstrate that the semantically high level, spatio-temporal supervision signal from the pre-trained VSR model offers the lip animation model more accurate lip movements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.635 + ], + "angle": 0, + "content": "SynthVSR achieves remarkable performance gains with labeled video data of different scales. We evaluate the performance of SynthVSR on the largest public VSR benchmark LRS3 with a Conformer-Transformer encoder-decoder VSR model [28]. In the low-resource setup using only 30 hours of labeled video data from LRS3 [2], our approach achieves a VSR WER of \\(43.3\\%\\), substantially outperforming the former VSR methods using hundreds or thousands of hours of video data for supervised training [1,27,37,44,52] and self-supervised learning [3,26,43]. Notably, we demonstrate the first successful attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. Using the complete 438 hours from LRS3 further improves WER to \\(27.9\\%\\) which is on par with the state-of-the-art self-supervised method AV-HuBERT-LARGE [43] that uses external 1,759 hours of unlabeled audio-visual data, but with fewer model parameters. Furthermore, following a recent high-resource setup [25] which uses additional 2,630 hours of ASR pseudo-labeled publicly available audio-visual data, our proposed method yields a new state-of-the-art VSR WER of \\(16.9\\%\\), surpassing the former state-of-the-art approaches [41, 42] trained on 90,000 hours of non-public machine-transcribed data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.895, + 0.728 + ], + "angle": 0, + "content": "Finally, we present extensive ablation studies to analyze where the improvement of SynthVSR comes from (e.g., the diversity of lip sources, the scale of ASR data). We also show considerable VSR improvement using synthetic video data derived from Text-To-Speech (TTS)-generated speech, indicating the great potential of our method for VSR." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.74, + 0.642, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Visual Speech Recognition. VSR has achieved remarkable progress with the success of deep learning and the availability of audio-visual datasets such as LRS2 [1] and LRS3 [2]. Progress was driven by adapting ASR approaches such as sequence-to-sequence models [11] and the Connectionist Temporal Classification (CTC) [5, 17] training objective. Recent studies have achieved noticeable results in VSR by using Transformer-based architecture [1], a convolutional variant [53], Conformer-based models [18], and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.472, + 0.903 + ], + "angle": 0, + "content": "1Such as LRW [11] and LRS2 [1] datasets which are only permitted for the purpose of academic research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18807" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.438 + ], + "angle": 0, + "content": "an attention-based feature pooling method [35]. The recent advances in VSR models are mainly dependent on leveraging increasingly large-scale transcribed non-public video datasets such as LSVSR (3,800 hours) [44] and YT31k (31,000 hours) [28]. Some recent works [41, 42] use 90,000 hours of non-publicly machine-transcribed video data to achieve state-of-the-art performance on standard benchmarks. Another popular trend is to leverage unlabeled audio-visual datasets such as AVSpeech [14] and VoxCeleb2 [10] with semi- or self-supervised learning. Using pre-trained ASR models to transcribe audio streams in unlabeled audio-visual datasets i.e., pseudo-labeling has demonstrated great potential to improve VSR performance. Such pseudo-labeled video data can be used for supervised training [28] or knowledge distillation [3]. Self-supervised approaches such as contrastive learning [12, 29, 32] and cross-modal feature prediction [26] have achieved significant improvement by exploiting unlabeled audio-visual data. AVHuBERT [43] is the state-of-the-art self-supervised method that predicts iteratively refined cluster assignments from masked audio-visual streams and achieves impressive VSR performance, especially when only limited labeled video data is available (e.g., 30 hours)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.473, + 0.697 + ], + "angle": 0, + "content": "Speech-Driven Facial Animation. Speech-driven facial animation aims to produce a talking face video conditioned on speech content. With the advances in deep visual generative models, speech-driven facial animation has significantly improved over the past years. Speech2Vid [9] is the first end-to-end neural network-based method for speech-driven facial animation. This method is optimized towards a pixel-level \\( L_{1} \\) reconstruction loss which results in blurry generated frames. Recently, Generative Adversarial Networks (GANs) have made impressive progress [49,50] in facial animation, as they are able to generate sharp and highly textured facial frames. Initially, most of these methods synthesized frontal faces and focused on improving lip-sync [8, 50], preserving the identity of the target speaker [54]. As the generation of synthetic frontal faces improved, researchers began to focus on generating talking heads that could handle natural head movements [7, 55]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.473, + 0.834 + ], + "angle": 0, + "content": "Learning from Synthetic Supervision. Training with synthetic data has attracted increasing research attention in recent years. Synthetic data could offer improved data diversity [13], which is useful in many tasks where labeled data is limited, such as image classification [16], objection detection [34, 36], semantic segmentation [40, 51], human pose estimation [13], and ASR [15, 31]. Unlike previous work, we focus on a new multimodal problem: how to produce and leverage synthetic video data for improving VSR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.192, + 0.864 + ], + "angle": 0, + "content": "3. SynthVSR" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this section, we introduce SynthVSR, a novel semi-supervised VSR framework to improve the performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "of VSR models using synthetic data, as shown in Fig. 1. We propose a speech-driven lip animation model trained on an audio-visual dataset, which is used to produce synthetic videos with large-scale transcribed speech and face datasets for scaling up VSR training. We will introduce the VSR model, the speech-driven lip animation model, and the semi-supervised VSR training next." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.208, + 0.628, + 0.223 + ], + "angle": 0, + "content": "3.1. VSR Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.415 + ], + "angle": 0, + "content": "The VSR model we use in this work is based on [27,28], which has achieved the state-of-the-art VSR performance on LRS3 [2] without the use of external data. The baseline VSR model is an encoder-decoder architecture. In particular, the encoder is comprised of two components, the visual front-end (a 3D convolutional layer followed by a ResNet-18 model [19,45]) and a Conformer [18] encoder. The decoder is based on the transformer architecture [47]. The baseline VSR model is trained end-to-end using a combination of the CTC loss [5,44] with an attention-based Cross-Entropy (CE) loss. Model details are described in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.424, + 0.768, + 0.44 + ], + "angle": 0, + "content": "3.2. Speech-Driven Lip Animation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Inspired by the recent advances in speech-driven facial animation [49, 50, 54, 55], we propose an approach for speech-driven lip animation that generates videos of talking mouth regions conditioned on speech utterances. The output space of the lip animation model is the same as the VSR input space. The proposed lip animation model is based on a temporal GAN [49, 50] with two discriminators. We further propose a VSR perceptual loss when labeled video data is available. The architecture of the speech-driven lip animation model is illustrated in the left part of Fig. 2. We will introduce each component in the next sections." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Generator. The generator \\( G \\) is an encoder-decoder structure, as shown in the right part of the Fig. 2. The generator uses the first frame of a video clip, a speech clip, and a head rotation sequence as inputs. The head rotation is used as the additional condition which helps better model the lip movements as most training videos are not static. The speech clip is divided into overlapping \\( 200\\mathrm{ms} \\) chunks with a stride of \\( 40~\\mathrm{ms} \\). The generator produces the corresponding video frame for each speech chunk. In the generator, an image encoder \\( E_{i} \\) and a speech encoder \\( E_{s} \\) are used to capture the visual information and speech context into latent embeddings \\( z_{i} \\) and \\( z_{s} \\), respectively. The image encoder \\( E_{i} \\) uses a stack of 2D convolutional layers to extract the visual embedding. The speech encoder \\( E_{s} \\) comprises a stack of 1D convolutional layers followed by a stack of GRU layers. The head rotations are provided in the form of sequences of 3D rotation matrices [55] \\( z_{r}\\in \\mathbb{R}^{3\\times 3} \\) with respect to the first frame. The three embeddings \\( z_{i},z_{s} \\) and \\( z_{r} \\) are concatenated and used to modulate the convolutional lay" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.53, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.094, + 0.891, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Architecture of proposed speech-driven lip animation model. Left: GAN-based speech-driven lip animation model generating lip movements given a lip image, a speech utterance, and a rotation sequence; Right: structure of the generator in the lip animation model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.468, + 0.388 + ], + "angle": 0, + "content": "ers in the frame decoder \\(D_{frame}\\), which is similar to StyleGAN2 [48]. Different from StyleGAN2, which generates frames from a learned constant input, we use the penultimate layer features of the image encoder \\(E_{i}\\) as the input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Discriminators. The speech-driven lip animation system has two discriminators: frame discriminator \\( D_{img} \\) and sequence discriminator \\( D_{seq} \\). The frame discriminator is a stack of CNN layers that operates on the image frame level. The original first frame is concatenated channel-wise to the target frame to form the input of the frame discriminator. This helps enforce visual consistency. The sequence discriminator operates on the sequence level to ensure the temporal consistency of synthetic lip movements. The sequence discriminator uses spatio-temporal convolutions to encode the image sequence, followed by GRU layers to determine if the sequence is real or not. Specifically, the frame discriminator \\( D_{img} \\) is trained on frames that are uniformly sampled from a video \\( v \\) using a sampling function \\( S(v) \\). The first frame \\( v_1 \\) is fed to the \\( D_{img} \\) as the condition. The input speech signal is \\( s \\). The adversarial loss of the \\( D_{img} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.652, + 0.405, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D i s c} ^ {i m g} = \\mathbb {E} _ {v} \\left[ \\log D _ {i m g} (S (v), v _ {1}) \\right] \\\\ + \\mathbb {E} _ {v, s} [ \\log (1 - D _ {i m g} (S (G (s, v _ {1})), v _ {1}) ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.697, + 0.469, + 0.726 + ], + "angle": 0, + "content": "\\(D_{seq}\\) operates on the entire sequence video \\(v\\). The adversarial loss of the \\(D_{seq}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.736, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {D i s c}} ^ {\\text {s e q}} = \\mathbb {E} _ {v} [ \\log D _ {\\text {s e q}} (v) ] + \\\\ \\mathbb {E} _ {v, s} [ \\log (1 - D _ {\\text {s e q}} (G (s, v _ {1})) ]. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "VSR Perceptual Loss. We further propose to optimize the lip animation model towards a VSR perceptual loss if labeled video data is available. We first pre-train a VSR model as introduced in Sec. 3.1. The proposed VSR perceptual loss corresponds to a weighted sum of feature distances computed from the visual front-end and the Transformer decoder of the pre-trained VSR model for real and generated samples. We use \\( L_{1} \\) norm to measure the visual embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.373 + ], + "angle": 0, + "content": "distance and Kullback-Leibler (KL) divergence to measure the logits distribution distance, respectively. The VSR perceptual loss is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.378, + 0.892, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V S R} = \\lambda_ {\\text {v i s u a l}} \\left\\| z _ {f} ^ {r} - z _ {f} ^ {s} \\right\\| _ {1} + \\lambda_ {\\text {l o g i t s}} \\operatorname {K L} \\left(\\hat {y} ^ {r}, \\hat {y} ^ {s}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.499 + ], + "angle": 0, + "content": "where \\( z_{f}^{r} \\) and \\( z_{f}^{s} \\) are the VSR front-end visual features of the real and synthetic video, respectively, \\( \\hat{y}^{r} \\) and \\( \\hat{y}^{s} \\) is the VSR predicted logits distribution of real and synthetic video, respectively, \\( \\lambda_{\\text{visual}} \\) and \\( \\lambda_{\\text{logits}} \\) control the weights of these two perceptual losses. The VSR model is frozen during the lip animation model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Training Objectives. The speech-driven lip animation model is trained using a combination of a reconstruction loss, adversarial losses, and a VSR perceptual loss. The reconstruction loss is computed based on the \\( L_{1} \\) distance between the generated video \\( \\hat{v} \\) and ground truth video \\( v \\):" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.586, + 0.891, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\| v - \\hat {v} \\| _ {1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.613, + 0.866, + 0.628 + ], + "angle": 0, + "content": "The overall training loss for the lip animation model is:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.638, + 0.892, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {A n i m a t i o n}} = \\lambda_ {\\text {d i s c}} ^ {\\text {i m g}} \\mathcal {L} _ {\\text {d i s c}} ^ {\\text {i m g}} + \\lambda_ {\\text {d i s c}} ^ {\\text {s e q}} \\mathcal {L} _ {\\text {d i s c}} ^ {\\text {s e q}} \\tag {5} \\\\ + \\lambda_ {r e c} \\mathcal {L} _ {r e c} + \\mathcal {L} _ {V S R}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.748 + ], + "angle": 0, + "content": "where \\(\\lambda_{disc}^{img},\\lambda_{disc}^{seq}\\), and \\(\\lambda_{rec}\\) represent the coefficient of the adversarial loss of frame discriminator \\(D_{img}\\), the adversarial loss of sequence discriminator \\(D_{seq}\\), and the reconstruction loss, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.862, + 0.773 + ], + "angle": 0, + "content": "3.3. Semi-Supervised VSR with Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In a typical supervised setting, a VSR model is trained from the labeled dataset \\(\\mathcal{D}_{real} = \\{(v_i^r,y_i^r)\\}_{i = 1}^{n_r}\\), where \\(n_r\\) is the number of paired video clips \\(v_{i}^{r}\\) and its transcriptions \\(y_{i}^{r}\\). In SynthVSR, we first train a speech-driven lip animation model on an unlabeled audio-visual dataset \\(\\mathcal{D}_{av} = \\{(v_i^{av},s_i^{av})\\}_{i = 1}^{n_{av}}\\), where \\(n_{av}\\) is the number of paired video clips \\(v_{i}^{av}\\) and speech utterances \\(s_i^{av}\\). Furthermore, we can use the trained generator \\(G\\) to generate synthetic video" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.171 + ], + "angle": 0, + "content": "from a labeled speech dataset \\(\\mathcal{D}_s = \\{(s_i^s,y_i^s)\\}_{i = 1}^{n_s}\\) and a cropped lips dataset \\(\\mathcal{D}_f = \\{c_i^f\\}_{i = 1}^{n_f}\\), where \\(n_{s}\\) is the number of paired speech clips \\(s_i^s\\) and its transcriptions \\(y_{i}^{s}\\) and \\(n_f\\) is the number of lip images \\(c_{i}^{f}\\). For each speech clip \\(s_i^s\\), the generator \\(G\\) generates its parallel synthetic video \\(\\hat{v}_i^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.18, + 0.47, + 0.198 + ], + "angle": 0, + "content": "\\[\nG \\left(s _ {i} ^ {s}, S \\left(\\mathcal {D} _ {f}\\right)\\right) \\mapsto \\hat {v} _ {i} ^ {s}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.469, + 0.312 + ], + "angle": 0, + "content": "where \\( S(\\cdot) \\) is a uniform sampling function. Then, the synthetic dataset \\( \\mathcal{D}_{\\text{synth}} = \\{\\hat{v}_i^s, y_i^s\\}_{i=1}^{n_s} \\) is obtained, where in general \\( n_s >> n_r \\) as there are far more annotated speech datasets (e.g., Librispeech [33]) than labeled video datasets. Finally, the VSR model is trained from \\( \\mathcal{D}_{\\text{real}} \\cup \\mathcal{D}_{\\text{synth}} \\), where the rich visual and text label information in \\( \\mathcal{D}_{\\text{synth}} \\) offers significant performance improvement for VSR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.21, + 0.342 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.349, + 0.18, + 0.364 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.372, + 0.469, + 0.434 + ], + "angle": 0, + "content": "We use several public datasets in this work. (1) Audiovisual datasets: LRS3 [2], AVSpeech [14] and VoxCeleb2 [10]; (2) Speech datasets: Librispeech [33], TED-LIUM 3 [20], Common Voice [4]; (3) Facial dataset: CelebA [23]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.434, + 0.469, + 0.509 + ], + "angle": 0, + "content": "LRS3. We conduct experiments on the LRS3 [2] dataset, which is the largest public benchmark for English VSR containing 438.9 hours of video clips from TED talks (408, 30, and 0.9 hours in the pre-training, training-validation, and test set, respectively)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.469, + 0.644 + ], + "angle": 0, + "content": "AVSpeech & VoxCeleb2. A recent work [25] uses public language classifier [46] and ASR models [6] to filter for English content and obtain pseudo-transcriptions for two multilingual audio-visual datasets: AVSpeech [14] and VoxCeleb2 [10], containing 1,323 hours from AVSpeech and 1,307 hours from VoxCeleb2. We follow the same data setting for our high-resource setting. Furthermore, we filter out AVSpeech videos with large jitter resulting in 933 hours of videos overall for speech-driven lip animation training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.469, + 0.704 + ], + "angle": 0, + "content": "Datasets for Speech-Driven Lip Animation Training. The lip animation model is trained on a combination of LRS3 (pre-training and training-validation splits) and the English subset (933 hours) of AVSpeech datasets." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets for Synthetic Data Generation. We use Librispeech [33], TED-LIUM 3 [20], Common Voice (English split) [4] datasets as the speech sources. We split Librispeech audio clips into segments of less than 6 seconds based on the provided silence annotations. For TED-LIUM 3 and Common Voice, we filter out the audio clips which are longer than 20s. As a result, 944 hours, 465 hours, and 2,243 hours of speech data are obtained for Librispeech, TED-LIUM 3, and Common Voice, respectively. We use the CelebA [23] dataset as a source of lip images, which has 202,599 face images and 10,177 identities. For each speech clip, we randomly sample one image from CelebA to generate one synthetic video. We use a static rotation matrix" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "in the generation process. In total, 3,652 hours of synthetic video clips are generated for scaling up VSR training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.133, + 0.661, + 0.149 + ], + "angle": 0, + "content": "4.2. Data Processing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.307 + ], + "angle": 0, + "content": "We train a mouth detection module using the Faster R-CNN object detection architecture [38] on the CelebA [23] dataset. We leverage the mouth corner positions provided in the dataset to identify the area around the mouth in each video frame. With this network, we locate and extract a \\(96 \\times 96\\) bounding box around the mouth, which we subsequently convert to grayscale. We use the same video pre-processing methods for lip animation and VSR models. For the text vocabulary, we use SentencePiece [22] subword units with a vocabulary size of 5,000." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.719, + 0.334 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.476 + ], + "angle": 0, + "content": "VSR Model. We consider two model configurations: (1) Conformer-BASE (250M) with 12-layer Conformer encoder, 6-layer Transformer decoder, 768 input dimensions, 3,072 feed-forward dimensions, and 16 attention heads; (2) Conformer-LARGE (783M) with 24-layers Conformer encoder, 9-layer Transformer decoder, 1024 input dimensions, 4,096 feed-forward dimensions, and 16 attention heads. For each configuration, the encoder and decoder have the same dimensions and attention heads." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.658 + ], + "angle": 0, + "content": "We use horizontal flipping, random cropping, and adaptive time masking as the VSR data augmentation methods [28]. We train the BASE and LARGE models for 75 epochs on 64 A100-GPUs with the AdamW [24] optimizer, a cosine learning rate scheduler, and a warm-up of 5 epochs. The peak learning rate is \\(1 \\times 10^{-3}\\) and \\(8 \\times 10^{-4}\\) for the BASE and LARGE models respectively. The number of frames in each batch is limited to 2,400 and 1,600 frames for the BASE and LARGE models, respectively. Following [25, 28], a pre-trained transformer-based language model is used in the VSR decoding stage. The training configuration is the same as that used in the previous work [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Speech-Driven Lip Animation. For each training video clip, we sample a sequence of 75 frames that corresponds to 3 seconds. The first frame of the sampled video sequence is used to drive the lip animation. We train the lip animation model for 70 epochs on 32 A100-GPUs with a batch size of 3 per GPU. The details of the module structure in the lip animation model are described in the supplementary material. Adam [21] optimizer is used to train the lip animation model with the learning rate \\(1 \\times 10^{-4}\\) for the generator and the frame discriminator. Sequence discriminator uses a smaller learning rate of \\(1 \\times 10^{-5}\\). The weights for each loss term are \\(\\lambda_{disc}^{img} = 1\\), \\(\\lambda_{disc}^{seq} = 0.2\\), \\(\\lambda_{rec} = 300\\) in Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We consider five lip animation model configurations. The lip animation model without VSR perceptual loss is referred to LAM-Baseline. The lip animation model with the BASE VSR model trained on LRS3 is referred to as LAM-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.894, + 0.279 + ], + "angle": 0, + "content": "
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
Afouras et al. [3]CNN✓595‡334-59.8
Ren et al. [37]TransformerX818‡--59.0
Afouras et al. [1]Transformer✓1,519†‡--58.9
Xu et al. [52]RNNX595‡--57.8
Shillingford et al. [44]RNN✓3,886†--55.1
Ma et al. [26]TransformerX4331,759-49.6*
Ma et al. [27]Conformer✓438--46.9
AV-HuBERT-BASE [43]TransformerX301,759-46.1
SynthVSRConformer-BASEX30--104.0
X--3,652100.3
X30-3,65244.7
✓30-3,65243.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.288, + 0.893, + 0.331 + ], + "angle": 0, + "content": "Table 1. Experimental results of low-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. †Includes non-publicly available data. ‡Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours. *Result taken from [43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.473, + 0.482 + ], + "angle": 0, + "content": "LRS3-VSR-VL, with the \\(\\lambda_{\\text{visual}} = 250\\) and \\(\\lambda_{\\text{logits}} = 10\\) in Eq. (3). Two variants LAM-LRS3-VSR-L and LAM-LRS3-VSR-V are further designed with the \\(\\lambda_{\\text{visual}} = 0\\) and \\(\\lambda_{\\text{logits}} = 0\\), respectively. These two model configurations are used in ablation studies. Last, the lip animation model with the BASE VSR model trained on LRS3 and 2,630 hours of pseudo-labeled AVSpeech and VoxCeleb2 is referred to as LAM-LRS3-AVoX-VSR, with the \\(\\lambda_{\\text{visual}} = 500\\) and \\(\\lambda_{\\text{logits}} = 10\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.391, + 0.513 + ], + "angle": 0, + "content": "4.4. Low-Resource Labeled Data Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.703 + ], + "angle": 0, + "content": "The LAM-Baseline model is used to generate 3,652 hours of synthetic data. We use 30 hours of LRS3 (training-validation) to evaluate the VSR performance when labeled data is scarce. We conduct experiments on the BASE VSR model. The results are reported in Tab. 1. As supervised VSR training from scratch with long sequences often poses optimization problems [1, 11, 28], we first use 30 hours of LRS3 and the 944 hours of Librispeech synthetic data to pre-train a VSR model with the same architecture as [28], then we take the weights of the pre-trained visual front-end to train the BASE VSR model. The pre-training details are described in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training with synthetic data leads to dramatic performance improvement when only 30 hours of labeled data is available. Using 3,652 hours of synthetic data and 30 hours of LRS3 labeled data for training, the BASE model achieves the WER \\(43.3\\%\\) (44.7% w/o. language model). Training from 30 hours of LRS3 results in a poor WER of \\(104\\%\\). Our proposed method significantly outperforms the former approaches using hundreds or thousands of data for supervised training [1,27,37,44,52] and self-supervised learning [3, 26]. Furthermore, our method performs better than the AV-HuBERT-BASE [43] model (46.1%) that uses 30 hours of LRS3 and 1,759 hours of unlabeled audio-visual data for self-supervised learning. Notably, we show" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.421 + ], + "angle": 0, + "content": "the first attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. In addition, we observe that training only from synthetic data performs poorly (100.3%), which may be caused by the domain mismatch between the real and synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.432, + 0.747, + 0.449 + ], + "angle": 0, + "content": "4.5. LRS3 Labeled Data Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.547 + ], + "angle": 0, + "content": "In this setting, we report the results when using the full 438 hours of LRS3. Experiments are conducted on the BASE VSR model. To avoid the optimization problem, we initialize the weights of the visual front-end from a publicly available Conformer-based model [28] pre-trained on LRS3 using curriculum learning. The results are shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.893, + 0.835 + ], + "angle": 0, + "content": "Our BASE model achieves WER \\(36.7\\%\\) when using 438 hours of LRS3 data for training. We generate 3,652 hours of synthetic data using the LAM-LRS3-VSR-VL model. Using 3,652 hours of synthetic data and 438 hours of LRS3 labeled data, the BASE VSR model achieves the WER \\(27.9\\%\\) \\((28.4\\% \\mathrm{w / o}\\) language model, corresponding to a WER reduction of \\(8.3\\%\\). Our method outperforms three recent approaches using 31,000 \\((33.6\\%)\\), 1,459 \\((31.5\\%)\\), and 2,679 \\((30.7\\%)\\) hours of labeled data, respectively. When compared with the state-of-the-art self-supervised method AV-HuBERT [43] that uses additional 1,759 hours of unlabeled audio-visual data, our method outperforms the AV-HuBERT-BASE model \\((34.8\\%)\\) by a large margin. Our method slightly performs better than the AV-HuBERT-LARGE model \\((28.6\\%)\\), but with fewer model parameters (our BASE model 250M vs AV-HuBERT-LARGE 390M). Note that we compare with the AV-HuBERT results without self-training as we do not use the pseudo-labeled 933 hours of AVSpeech subset for VSR training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.818, + 0.864 + ], + "angle": 0, + "content": "4.6. High-Resource Labeled Data Setting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further evaluate the scalability of SynthVSR: when using machine-transcribed AVSpeech [14] and VoxCeleb2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18811" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.088, + 0.894, + 0.314 + ], + "angle": 0, + "content": "
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
AV-HuBERT-BASE [43]TransformerX4331,759-34.8
Makino et al. [30]TransformerX31,000†--33.6
Ma et al. [28]Conformer✓1,459‡--31.5
Prajwal et al. [35]Transformer✓2,676†--30.7
AV-HuBERT-LARGE [43]TransformerX4331,759-28.6
AV-HuBERT-LARGE w. Self-Training [43]TransformerX4331,759-26.9
Auto-AVSR [25]Conformer✓3,448‡--19.1
Serdyuk et al. [42]TransformerX90,000†--25.9
Serdyuk et al. [41]TransformerX90,000†--17.0
SynthVSRConformer-BASEX438--36.7
X438-3,65228.4
✓438-3,65227.9
X3,068--21.2
X3,068-3,65219.4
✓3,068-3,65218.7
SynthVSRConformer-LARGEX3,068-3,65218.2
✓3,068-3,65216.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.324, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Table 2. Experimental results of LRS3 & high-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. \\( {}^{ \\dagger } \\) Includes non-publicly available data. \\( {}^{§} \\) Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.456 + ], + "angle": 0, + "content": "[10] as additional training data. We generate 3,652 hours of synthetic data with LAM-LRS3-AVoX-VSR and conduct experiments on BASE and LARGE models. The visual front-end is initialized from a pre-trained model which is the same as that used in the LRS3 labeled data setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.471, + 0.684 + ], + "angle": 0, + "content": "We first train the BASE model with 438 hours of labeled LRS3 and 2,630 hours of pseudo-labeled data, resulting in a strong VSR baseline with the WER \\(21.2\\%\\). By using additional 3,652 hours of synthetic data, the WER of the BASE model improves to \\(18.7\\%\\) (19.4% w./o. language model), which outperforms [25] that uses additional labeled dataset LRS2 (223 hours) and LRW (157 hours) for training. Although the VSR model has seen a large amount of labeled data, and the speech-driven lip animation model is trained from part of the VSR training data, synthetic data can still lead to considerable performance gains. Furthermore, increasing the model size from BASE to LARGE results in better VSR performance with the WER of \\(16.9\\%\\) (18.2% w./o. language model), which is the current state-of-the-art performance on LRS3, with publicly available data only." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.363, + 0.712 + ], + "angle": 0, + "content": "4.7. Why SynthVSR Improves VSR?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.824 + ], + "angle": 0, + "content": "We conduct extensive ablation studies to understand the impact of VSR perceptual loss, the diversity of lip sources, and the scale of ASR data for SynthVSR. In addition, we assess the domain mismatch between the real and synthetic data using VSR models. We perform experiments on the LRS3 labeled data setting with the Conformer-BASE VSR model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of VSR Perceptual Loss. We experiment with four lip animation model variants: LAM-Baseline, LAM-LRS3-VSR-V, LAM-LRS3-VSR-L, LAM-LRS3-VSR-VL to generate 944 hours of synthetic data from Librispeech, respectively. The WER of the BASE model trained with only 438" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.895, + 0.548 + ], + "angle": 0, + "content": "hours of LRS3 data is \\(36.7\\%\\) (see in Tab. 2), the results are shown in Tab. 3. Using 944 hours of synthetic data generated from the LAM-Baseline model, the WER improves to \\(32.2\\%\\). The individual visual VSR loss and linguistic VSR loss improve the WER to \\(31.7\\%\\) and \\(31.4\\%\\), respectively. The linguistic VSR loss performs slightly better than the visual one. Further, the combination of visual and linguistic VSR loss further improves the WER to \\(30.8\\%\\), which indicates that the spatio-temporal visual and high-level semantic knowledge from a pre-trained VSR model could offer the lip animation model more accurate lip movements." + }, + { + "type": "table", + "bbox": [ + 0.581, + 0.559, + 0.819, + 0.634 + ], + "angle": 0, + "content": "
Lip animation modelWER (%)
LAM-Baseline32.2
LAM-LRS3-VSR-V31.7
LAM-LRS3-VSR-L31.4
LAM-LRS3-VSR-VL30.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.643, + 0.884, + 0.658 + ], + "angle": 0, + "content": "Table 3. Ablation studies on the proposed VSR perceptual loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Effect of Increasing Lip Sources Diversity. To analyze the impact of the diversity of lip sources we generate 944 hours of synthetic data generated from Librispeech using two target face sources: CelebA and LRS3 (pre-training). Images from LRS3 (pre-training) do not introduce any new lip images for VSR training. The results are reported in Tab. 4. Using the synthetic data with LRS3 lip images, the BASE model achieves the WER \\(32.1\\%\\) which is \\(1.3\\%\\) worse than that with CelebA lip images, which means the external lip images from CelebA offer better WER improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Effect of the Scale of Speech Corpus. Using synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice with LAM-LRS3-VSR-VL model synthetic data, the LRS3 (test) WER improves from \\(36.7\\%\\) to \\(30.9\\%\\), \\(30.8\\%\\), and \\(30.1\\%\\), respectively (see in Tab. 5). When combining" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18812" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.438, + 0.139 + ], + "angle": 0, + "content": "
Lip sourceUnique imagesWER (%)
LRS3-pretrain5,09032.1
CelebA202,59930.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.149, + 0.468, + 0.178 + ], + "angle": 0, + "content": "Table 4. Ablation studies on the impact of lip sources diversity. LRS3-pretrain indicates using the LRS3 (pre-training) images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.203, + 0.468, + 0.309 + ], + "angle": 0, + "content": "all synthetic data together (3,652 hours) for training, the WER improves to \\(28.4\\%\\). We observe that richer speech corpora can offer larger performance improvements. Furthermore, we find that the improvement from scaling speech corpora is more pronounced than when adding VSR perceptual loss and lip sources diversity, suggesting that our proposed method benefits greatly from large speech datasets." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.32, + 0.459, + 0.407 + ], + "angle": 0, + "content": "
Training dataHoursWER (%)
LRS343836.7
LRS3 + TED-Synth90330.9
LRS3 + LBS-Synth1,38230.8
LRS3 + CV-Synth2,68130.1
LRS3 + [TED + LBS + CV]-Synth4,09028.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.417, + 0.468, + 0.473 + ], + "angle": 0, + "content": "Table 5. Ablation studies on different speech corpus. TED-Synth, LBS-Synth, and CV-Synth indicates the synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Assessment of the Domain Mismatch Using VSR. We use 944 hours of Librispeech synthetic data generated from LAM-Baseline and LAM-LRS3-VSR-VL for VSR training. We conduct an assessment on the LRS3 real and synthetic test sets, the synthetic test set is obtained by passing the speech and the first frame in the real test set to the lip animation model. The results are illustrated in Fig. 3. By introducing the VSR perceptual loss, the WER on the synthetic test set is reduced from \\(106.8\\%\\) to \\(49.7\\%\\), which indicates the VSR loss minimizes the domain gap between real and synthetic data. After training on synthetic data, the WER reduces for both real and synthetic data. Specifically, the WER is improved substantially from \\(106.8\\%\\) to \\(36.0\\%\\) and from \\(49.7\\%\\) to \\(16.8\\%\\) on synthetic test sets generated using LAM-Baseline and LAM-LRS3-VSR-VL, respectively. We observe that domain mismatch can be further reduced by combining real and synthetic data for VSR training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.416, + 0.759 + ], + "angle": 0, + "content": "4.8. SynthVSR with TTS-Generated Speech" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In the real world, text corpora (e.g., Wikipedia) are more accessible than transcribed speech corpora. To explore the potential of using TTS-generated speech for SynthVSR, we use the off-the-shelf TTS model FastSpeech 2 [39] to generate 944 hours of synthetic speech from Librispeech transcriptions, which is further used to generate synthetic video clips using LAM-LRS3-VSR-VL model. Using these 944 hours of synthetic data to train the BASE model under the LRS3 labeled data setting, the WER is improved" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.081, + 0.848, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Assessment of the domain mismatch between real and synthetic video data using VSR models trained with only real data (VSR-Real, black) and real data together with synthetic data (VSR-Real-Synth, red), respectively. Synth-Test and Real-Test refer to the synthetic and real LRS3 (test) sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.417 + ], + "angle": 0, + "content": "from \\(36.7\\%\\) to \\(32.9\\%\\), as shown in Tab. 6. Although TTS-generated speech is not as natural as the original LibriSpeech data, the improvement of using synthetic data is still decent \\((3.8\\%)\\), indicating the further potential of SynthVSR in real-world applications where labeled video data is sparse, such as the healthcare industry." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.43, + 0.86, + 0.494 + ], + "angle": 0, + "content": "
Training dataHoursWER (%)
LRS343836.7
LRS3 + LBS-Synth1,38230.8
LRS3 + TTS-LBS-Synth1,38232.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.504, + 0.891, + 0.532 + ], + "angle": 0, + "content": "Table 6. Experiments results of using the synthetic speech generated synthetic video (TTS-LBS-Synth) for VSR training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.776, + 0.581 + ], + "angle": 0, + "content": "5. Limitations & Societal Impact" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.74 + ], + "angle": 0, + "content": "The excellent performance improvement of SynthVSR comes at the cost of high computational demands during training, as we use additional large-scale synthetic video data. Also, although SynthVSR has achieved state-of-the-art VSR performance, the LRS3 test set is relatively limited (0.9 hours), which is from TED talks, while real-world videos are more challenging (e.g., egocentric videos). SynthVSR has many positive real-world applications, such as helping the hearing-impaired or people with aphonia with everyday communication." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.618, + 0.77 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We have presented a semi-supervised method for VSR enhanced with synthetic lip movements. The speech-driven lip animation model is proposed to generate synthetic video data from labeled speech datasets and face images for scaling up VSR. Our method achieves state-of-the-art results on LRS3, outperforming prior work trained on more labeled or unlabeled real video data. Our work fosters future research on generating and exploiting synthetic visual data for VSR." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.177, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: A large-scale dataset for visual speech recognition. arXiv:1809.00496, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2143-2147, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[4] Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Conference on Language Resources and Evaluation, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. LipNet: Sentence-level lipreading. arXiv:1611.01599, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.539, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision, pages 520-535, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? In British Machine Vision Conference, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.72 + ], + "angle": 0, + "content": "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, pages 87-103. Springer, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[12] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: Learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3D human pose estimation: Motion to the rescue. In Advances in Neural Information Processing Systems, volume 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv:1804.03619, 2018. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.121, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Amin Fazel, Wei Yang, Yulan Liu, Roberto Barra-Chicote, Yixiong Meng, Roland Maas, and Jasha Droppo. SynthASR: Unlocking synthetic data for speech recognition. arXiv:2106.07803, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[16] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv:2007.04954, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine learning, pages 369-376, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[18] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. In INTER-SPEECH, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[20] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In International Conference on Speech and Computer, pages 198-208, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[22] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[23] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision, December 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[25] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-visual speech recognition with automatic labels. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[26] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In IN-TERSPEECH, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.832 + ], + "angle": 0, + "content": "[27] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 7613-7617, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[28] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nature Machine Intelligence, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.888 + ], + "angle": 0, + "content": "[29] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Contrastive learning of global and local audio-visual representations. arXiv:2104.05418, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE Automatic Speech Recognition and Understanding Workshop, pages 905–912, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.231 + ], + "angle": 0, + "content": "[31] Sepand Mavandadi, Tara N Sainath, Kevin Hu, and Zelin Wu. A deliberation-based joint acoustic and text decoder. In INTERSPEECH, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[32] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging uni-modal self-supervised learning for multimodal audio-visual speech recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[33] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 5206-5210, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[34] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3D models. In Proceedings of the IEEE International Conference on Computer Vision, pages 1278-1286, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[35] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[36] Aayush Prakash, Shaad Boochoon, Mark Brophy, David Acuna, Eric Cameracci, Gavriel State, Omer Shapira, and Stan Birchfield. Structured domain randomization: Bridging the reality gap by context-aware synthetic data. In 2019 International Conference on Robotics and Automation, pages 7249-7255, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[37] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[39] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv:2006.04558, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[40] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3234-3243, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[41] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition. In INTERSPEECH, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Dmitriy Serdyuk, Olivier Siohan, and Otavio de Pinho Forin Braga. Audio-visual speech recognition is worth \\(32 \\times 32 \\times 8\\) voxels. In IEEE Automatic Speech Recognition and Under" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.733, + 0.106 + ], + "angle": 0, + "content": "standing Workshop, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.108, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[43] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In International Conference on Learning Representations, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[44] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. In INTERSPEECH, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[45] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. arXiv:1703.04105, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[46] Jörgen Valk and Tanel Alumäe. Voxlingua107: A dataset for spoken language recognition. In 2021 IEEE Spoken Language Technology Workshop, pages 652-658, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[48] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. StyleGAN2 distillation for feed-forward image manipulation. In European Conference on Computer Vision, pages 170-186. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[49] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven facial animation with temporal GANs. In British Machine Vision Conference, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[50] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with GANs. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[51] Zhonghao Wang, Mo Yu, Yunchao Wei, Rogerio Feris, Jinjun Xiong, Wen-mei Hwu, Thomas S Huang, and Honghui Shi. Differential treatment for stuff and things: A simple unsupervised domain adaptation method for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12635-12644, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[52] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14433-14442, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[53] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18815" + } + ] +] \ No newline at end of file diff --git a/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_origin.pdf b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2862ffa03110cde40b86ea9ae846ca02e4a69f58 --- /dev/null +++ b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/6d7987b1-44f5-4f2e-8f41-c2dafd6e1f4e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe025602346c2f67008f8751443f414d5a83165df6d983f6459690471319fcb7 +size 784070 diff --git a/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/full.md b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7edd7841df5619403302bb366be8c415bbc70dd --- /dev/null +++ b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/full.md @@ -0,0 +1,293 @@ +# SynthVSR: Scaling Up Visual Speech Recognition With Synthetic Supervision + +Xubo Liu $^{1*}$ , Egor Lakomkin $^{2}$ , Konstantinos Vougioukas $^{2}$ , Pingchuan Ma $^{2}$ , Honglie Chen $^{2}$ , Ruiming Xie $^{2}$ , Morrie Doulaty $^{2}$ , Niko Moritz $^{2}$ , Jachym Kolar $^{2}$ , Stavros Petridis $^{2}$ , Maja Pantic $^{2}$ , Christian Fuegen $^{2}$ + +$^{1}$ University of Surrey $^{2}$ Meta AI + +![](images/9c83aa001db9bf9b18f74e94f33c8cbbf5848c6198b7eb2f947af666b6faf19d.jpg) +Figure 1. Scaling up visual speech recognition with synthetic supervision (SynthVSR): we propose SynthVSR, a semi-supervised framework that can substantially improve the performance of VSR models by using synthetic lip movements. Firstly, we introduce a speech-driven lip animation model that generates lip movement videos conditioned on input lip images and speech utterances (left). Secondly, we generate large-scale synthetic videos using transcribed speech datasets and lip images. The combination of synthetic videos and their corresponding speech transcriptions constitutes the synthetic dataset (centre). Finally, we conduct semi-supervised VSR training with synthetic and real datasets. Our method substantially improves the performance of VSR models with large-scale synthetic data (right). + +![](images/2204cfd53a62e8188cab6ef04555d56ed95e5a2eaa6a468886cc70f85aabeea3.jpg) + +![](images/c690f05f11549ce55d61c18a411de52117a0004767962da6952c7f5cec4a9da4.jpg) + +# Abstract + +Recently reported state-of-the-art results in visual speech recognition (VSR) often rely on increasingly large amounts of video data, while the publicly available transcribed video datasets are limited in size. In this paper, for the first time, we study the potential of leveraging synthetic visual data for VSR. Our method, termed SynthVSR, substantially improves the performance of VSR systems with synthetic lip movements. The key idea behind SynthVSR is to leverage a speech-driven lip animation model that generates lip movements conditioned on the input speech. The speech-driven lip animation model is trained on an unlabeled audio-visual dataset and could be further optimized towards a pre-trained VSR model when labeled videos are available. As plenty of transcribed acoustic data and face images are available, we are able to generate large-scale synthetic data using the proposed lip animation model for semi-supervised VSR training. We evaluate the performance of our approach on the largest public VSR benchmark - Lip Reading Sentences 3 (LRS3). SynthVSR achieves a WER of $43.3\%$ with only 30 hours of real labeled data, + +outperforming off-the-shelf approaches using thousands of hours of video. The WER is further reduced to $27.9\%$ when using all 438 hours of labeled data from LRS3, which is on par with the state-of-the-art self-supervised AV-HuBERT method. Furthermore, when combined with large-scale pseudo-labeled audio-visual data SynthVSR yields a new state-of-the-art VSR WER of $16.9\%$ using publicly available data only, surpassing the recent state-of-the-art approaches trained with 29 times more non-public machine-transcribed video data (90,000 hours). Finally, we perform extensive ablation studies to understand the effect of each component in our proposed method. + +# 1. Introduction + +Visual speech recognition (VSR), also known as lip reading, is the task of recognizing speech content based on visual lip movements. VSR has a wide range of applications in real-world scenarios such as helping the hearing-impaired perceive human speech and improving automatic speech recognition (ASR) in noisy environments. + +VSR is a challenging task, as it requires capturing speech from high-dimensional spatio-temporal videos, while multiple words are visually ambiguous (e.g., "world" and + +"word") in the visual streams. Recently, with the release of large-scale transcribed audio-visual datasets such as LRS2 [1] and LRS3 [2], deep neural networks have become the mainstream approach for VSR. However, even the largest public dataset for English VSR, LRS3, does not exceed 500 hours of transcribed video data. The lack of large-scale transcribed audio-visual datasets potentially results in VSR models which could only work in a laboratory environment i.e. limited word vocabulary and lip sources diversity [28]. + +A common solution to this issue is to collect and annotate large-scale audio-visual datasets. For example, [41, 42] collected 90,000 hours of YouTube videos with user-uploaded transcriptions to achieve state-of-the-art performance on standard benchmarks. However, such a procedure is expensive and time-consuming, especially for most of the world's 7,000 languages [43]. If annotations are missing, the ASR can be used to generate the transcriptions automatically and this has been shown to be an effective approach to significantly improve VSR performance [28]. The other promising direction is to learn audio-visual speech representations from large amounts of parallel unlabeled audio-visual data in a self-supervised approach, and then fine-tune them on the limited labeled video dataset [43]. Nevertheless, publicly available video datasets are also limited and their usage may raise license-related concerns, barring their use in commercial applications. + +Human perception of speech is inherently multimodal, involving both audition and vision [43]. ASR, which is a complementary task to VSR, has achieved impressive performance in recent years, with tens of thousands of hours of annotated speech datasets [4, 20, 33] available for large-scale training. It is intuitive to ask: Can we improve VSR with large amounts of transcribed acoustic-only ASR training data? The key to this question is to take advantage of recent advances in speech-driven visual generative models [49, 50]. By leveraging visual generative models, we can produce parallel synthetic videos for large-scale labeled audio datasets. Synthetic videos provide advantages such as having control over the target text and lip image as well as the duration of a generated utterance. To the best of our knowledge, the potential of leveraging synthetic visual data for improving VSR has never been studied in the literature. + +In this work, we present SynthVSR, a novel semi-supervised framework for VSR. In particular, we first propose a speech-driven lip animation model that can generate synthetic lip movements video conditioned on the speech content. Next, the proposed lip animation model is used to generate synthetic video clips from transcribed speech datasets (e.g., Librispeech [33]) and human face datasets (e.g., CelebA [23]). Then, the synthetic videos together with the corresponding transcriptions are used in combi + +nation with the real video-text pairs (e.g., LRS3 [2]) for large-scale semi-supervised VSR training. The pipeline of SynthVSR is illustrated in Fig. 1. Unlike existing studies in exploiting unlabeled video data for VSR using methods such as pseudo-labeling [28] and self-supervised learning [43], we use the unlabeled audio-visual data to train a cross-modal generative model in order to bridge ASR training data and VSR. Furthermore, we propose to optimize the lip animation model towards a pre-trained VSR model when labeled videos are available. We empirically demonstrate that the semantically high level, spatio-temporal supervision signal from the pre-trained VSR model offers the lip animation model more accurate lip movements. + +SynthVSR achieves remarkable performance gains with labeled video data of different scales. We evaluate the performance of SynthVSR on the largest public VSR benchmark LRS3 with a Conformer-Transformer encoder-decoder VSR model [28]. In the low-resource setup using only 30 hours of labeled video data from LRS3 [2], our approach achieves a VSR WER of $43.3\%$ , substantially outperforming the former VSR methods using hundreds or thousands of hours of video data for supervised training [1,27,37,44,52] and self-supervised learning [3,26,43]. Notably, we demonstrate the first successful attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. Using the complete 438 hours from LRS3 further improves WER to $27.9\%$ which is on par with the state-of-the-art self-supervised method AV-HuBERT-LARGE [43] that uses external 1,759 hours of unlabeled audio-visual data, but with fewer model parameters. Furthermore, following a recent high-resource setup [25] which uses additional 2,630 hours of ASR pseudo-labeled publicly available audio-visual data, our proposed method yields a new state-of-the-art VSR WER of $16.9\%$ , surpassing the former state-of-the-art approaches [41, 42] trained on 90,000 hours of non-public machine-transcribed data. + +Finally, we present extensive ablation studies to analyze where the improvement of SynthVSR comes from (e.g., the diversity of lip sources, the scale of ASR data). We also show considerable VSR improvement using synthetic video data derived from Text-To-Speech (TTS)-generated speech, indicating the great potential of our method for VSR. + +# 2. Related Work + +Visual Speech Recognition. VSR has achieved remarkable progress with the success of deep learning and the availability of audio-visual datasets such as LRS2 [1] and LRS3 [2]. Progress was driven by adapting ASR approaches such as sequence-to-sequence models [11] and the Connectionist Temporal Classification (CTC) [5, 17] training objective. Recent studies have achieved noticeable results in VSR by using Transformer-based architecture [1], a convolutional variant [53], Conformer-based models [18], and + +an attention-based feature pooling method [35]. The recent advances in VSR models are mainly dependent on leveraging increasingly large-scale transcribed non-public video datasets such as LSVSR (3,800 hours) [44] and YT31k (31,000 hours) [28]. Some recent works [41, 42] use 90,000 hours of non-publicly machine-transcribed video data to achieve state-of-the-art performance on standard benchmarks. Another popular trend is to leverage unlabeled audio-visual datasets such as AVSpeech [14] and VoxCeleb2 [10] with semi- or self-supervised learning. Using pre-trained ASR models to transcribe audio streams in unlabeled audio-visual datasets i.e., pseudo-labeling has demonstrated great potential to improve VSR performance. Such pseudo-labeled video data can be used for supervised training [28] or knowledge distillation [3]. Self-supervised approaches such as contrastive learning [12, 29, 32] and cross-modal feature prediction [26] have achieved significant improvement by exploiting unlabeled audio-visual data. AVHuBERT [43] is the state-of-the-art self-supervised method that predicts iteratively refined cluster assignments from masked audio-visual streams and achieves impressive VSR performance, especially when only limited labeled video data is available (e.g., 30 hours). + +Speech-Driven Facial Animation. Speech-driven facial animation aims to produce a talking face video conditioned on speech content. With the advances in deep visual generative models, speech-driven facial animation has significantly improved over the past years. Speech2Vid [9] is the first end-to-end neural network-based method for speech-driven facial animation. This method is optimized towards a pixel-level $L_{1}$ reconstruction loss which results in blurry generated frames. Recently, Generative Adversarial Networks (GANs) have made impressive progress [49,50] in facial animation, as they are able to generate sharp and highly textured facial frames. Initially, most of these methods synthesized frontal faces and focused on improving lip-sync [8, 50], preserving the identity of the target speaker [54]. As the generation of synthetic frontal faces improved, researchers began to focus on generating talking heads that could handle natural head movements [7, 55]. + +Learning from Synthetic Supervision. Training with synthetic data has attracted increasing research attention in recent years. Synthetic data could offer improved data diversity [13], which is useful in many tasks where labeled data is limited, such as image classification [16], objection detection [34, 36], semantic segmentation [40, 51], human pose estimation [13], and ASR [15, 31]. Unlike previous work, we focus on a new multimodal problem: how to produce and leverage synthetic video data for improving VSR. + +# 3. SynthVSR + +In this section, we introduce SynthVSR, a novel semi-supervised VSR framework to improve the performance + +of VSR models using synthetic data, as shown in Fig. 1. We propose a speech-driven lip animation model trained on an audio-visual dataset, which is used to produce synthetic videos with large-scale transcribed speech and face datasets for scaling up VSR training. We will introduce the VSR model, the speech-driven lip animation model, and the semi-supervised VSR training next. + +# 3.1. VSR Model + +The VSR model we use in this work is based on [27,28], which has achieved the state-of-the-art VSR performance on LRS3 [2] without the use of external data. The baseline VSR model is an encoder-decoder architecture. In particular, the encoder is comprised of two components, the visual front-end (a 3D convolutional layer followed by a ResNet-18 model [19,45]) and a Conformer [18] encoder. The decoder is based on the transformer architecture [47]. The baseline VSR model is trained end-to-end using a combination of the CTC loss [5,44] with an attention-based Cross-Entropy (CE) loss. Model details are described in the supplementary material. + +# 3.2. Speech-Driven Lip Animation + +Inspired by the recent advances in speech-driven facial animation [49, 50, 54, 55], we propose an approach for speech-driven lip animation that generates videos of talking mouth regions conditioned on speech utterances. The output space of the lip animation model is the same as the VSR input space. The proposed lip animation model is based on a temporal GAN [49, 50] with two discriminators. We further propose a VSR perceptual loss when labeled video data is available. The architecture of the speech-driven lip animation model is illustrated in the left part of Fig. 2. We will introduce each component in the next sections. + +Generator. The generator $G$ is an encoder-decoder structure, as shown in the right part of the Fig. 2. The generator uses the first frame of a video clip, a speech clip, and a head rotation sequence as inputs. The head rotation is used as the additional condition which helps better model the lip movements as most training videos are not static. The speech clip is divided into overlapping $200\mathrm{ms}$ chunks with a stride of $40~\mathrm{ms}$ . The generator produces the corresponding video frame for each speech chunk. In the generator, an image encoder $E_{i}$ and a speech encoder $E_{s}$ are used to capture the visual information and speech context into latent embeddings $z_{i}$ and $z_{s}$ , respectively. The image encoder $E_{i}$ uses a stack of 2D convolutional layers to extract the visual embedding. The speech encoder $E_{s}$ comprises a stack of 1D convolutional layers followed by a stack of GRU layers. The head rotations are provided in the form of sequences of 3D rotation matrices [55] $z_{r}\in \mathbb{R}^{3\times 3}$ with respect to the first frame. The three embeddings $z_{i},z_{s}$ and $z_{r}$ are concatenated and used to modulate the convolutional lay + +![](images/2af223e0bfa6526d1fa8498868f7d7c02a96e420565afa1521436be15bf89821.jpg) +Figure 2. Architecture of proposed speech-driven lip animation model. Left: GAN-based speech-driven lip animation model generating lip movements given a lip image, a speech utterance, and a rotation sequence; Right: structure of the generator in the lip animation model. + +![](images/a4ba1df8a8020a716329db45853c1667c4e5f63903a24aa036615d0d0bd59e65.jpg) + +ers in the frame decoder $D_{frame}$ , which is similar to StyleGAN2 [48]. Different from StyleGAN2, which generates frames from a learned constant input, we use the penultimate layer features of the image encoder $E_{i}$ as the input. + +Discriminators. The speech-driven lip animation system has two discriminators: frame discriminator $D_{img}$ and sequence discriminator $D_{seq}$ . The frame discriminator is a stack of CNN layers that operates on the image frame level. The original first frame is concatenated channel-wise to the target frame to form the input of the frame discriminator. This helps enforce visual consistency. The sequence discriminator operates on the sequence level to ensure the temporal consistency of synthetic lip movements. The sequence discriminator uses spatio-temporal convolutions to encode the image sequence, followed by GRU layers to determine if the sequence is real or not. Specifically, the frame discriminator $D_{img}$ is trained on frames that are uniformly sampled from a video $v$ using a sampling function $S(v)$ . The first frame $v_1$ is fed to the $D_{img}$ as the condition. The input speech signal is $s$ . The adversarial loss of the $D_{img}$ is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {D i s c} ^ {i m g} = \mathbb {E} _ {v} \left[ \log D _ {i m g} (S (v), v _ {1}) \right] \\ + \mathbb {E} _ {v, s} [ \log (1 - D _ {i m g} (S (G (s, v _ {1})), v _ {1}) ]. \\ \end{array} +$$ + +$D_{seq}$ operates on the entire sequence video $v$ . The adversarial loss of the $D_{seq}$ is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {D i s c}} ^ {\text {s e q}} = \mathbb {E} _ {v} [ \log D _ {\text {s e q}} (v) ] + \\ \mathbb {E} _ {v, s} [ \log (1 - D _ {\text {s e q}} (G (s, v _ {1})) ]. \tag {2} \\ \end{array} +$$ + +VSR Perceptual Loss. We further propose to optimize the lip animation model towards a VSR perceptual loss if labeled video data is available. We first pre-train a VSR model as introduced in Sec. 3.1. The proposed VSR perceptual loss corresponds to a weighted sum of feature distances computed from the visual front-end and the Transformer decoder of the pre-trained VSR model for real and generated samples. We use $L_{1}$ norm to measure the visual embedding + +distance and Kullback-Leibler (KL) divergence to measure the logits distribution distance, respectively. The VSR perceptual loss is obtained by: + +$$ +\mathcal {L} _ {V S R} = \lambda_ {\text {v i s u a l}} \left\| z _ {f} ^ {r} - z _ {f} ^ {s} \right\| _ {1} + \lambda_ {\text {l o g i t s}} \operatorname {K L} \left(\hat {y} ^ {r}, \hat {y} ^ {s}\right), \tag {3} +$$ + +where $z_{f}^{r}$ and $z_{f}^{s}$ are the VSR front-end visual features of the real and synthetic video, respectively, $\hat{y}^{r}$ and $\hat{y}^{s}$ is the VSR predicted logits distribution of real and synthetic video, respectively, $\lambda_{\text{visual}}$ and $\lambda_{\text{logits}}$ control the weights of these two perceptual losses. The VSR model is frozen during the lip animation model training. + +Training Objectives. The speech-driven lip animation model is trained using a combination of a reconstruction loss, adversarial losses, and a VSR perceptual loss. The reconstruction loss is computed based on the $L_{1}$ distance between the generated video $\hat{v}$ and ground truth video $v$ : + +$$ +\mathcal {L} _ {\text {r e c}} = \| v - \hat {v} \| _ {1}. \tag {4} +$$ + +The overall training loss for the lip animation model is: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {A n i m a t i o n}} = \lambda_ {\text {d i s c}} ^ {\text {i m g}} \mathcal {L} _ {\text {d i s c}} ^ {\text {i m g}} + \lambda_ {\text {d i s c}} ^ {\text {s e q}} \mathcal {L} _ {\text {d i s c}} ^ {\text {s e q}} \tag {5} \\ + \lambda_ {r e c} \mathcal {L} _ {r e c} + \mathcal {L} _ {V S R}, \\ \end{array} +$$ + +where $\lambda_{disc}^{img},\lambda_{disc}^{seq}$ , and $\lambda_{rec}$ represent the coefficient of the adversarial loss of frame discriminator $D_{img}$ , the adversarial loss of sequence discriminator $D_{seq}$ , and the reconstruction loss, respectively. + +# 3.3. Semi-Supervised VSR with Synthetic Data + +In a typical supervised setting, a VSR model is trained from the labeled dataset $\mathcal{D}_{real} = \{(v_i^r,y_i^r)\}_{i = 1}^{n_r}$ , where $n_r$ is the number of paired video clips $v_{i}^{r}$ and its transcriptions $y_{i}^{r}$ . In SynthVSR, we first train a speech-driven lip animation model on an unlabeled audio-visual dataset $\mathcal{D}_{av} = \{(v_i^{av},s_i^{av})\}_{i = 1}^{n_{av}}$ , where $n_{av}$ is the number of paired video clips $v_{i}^{av}$ and speech utterances $s_i^{av}$ . Furthermore, we can use the trained generator $G$ to generate synthetic video + +from a labeled speech dataset $\mathcal{D}_s = \{(s_i^s,y_i^s)\}_{i = 1}^{n_s}$ and a cropped lips dataset $\mathcal{D}_f = \{c_i^f\}_{i = 1}^{n_f}$ , where $n_{s}$ is the number of paired speech clips $s_i^s$ and its transcriptions $y_{i}^{s}$ and $n_f$ is the number of lip images $c_{i}^{f}$ . For each speech clip $s_i^s$ , the generator $G$ generates its parallel synthetic video $\hat{v}_i^s$ : + +$$ +G \left(s _ {i} ^ {s}, S \left(\mathcal {D} _ {f}\right)\right) \mapsto \hat {v} _ {i} ^ {s}, \tag {6} +$$ + +where $S(\cdot)$ is a uniform sampling function. Then, the synthetic dataset $\mathcal{D}_{\text{synth}} = \{\hat{v}_i^s, y_i^s\}_{i=1}^{n_s}$ is obtained, where in general $n_s >> n_r$ as there are far more annotated speech datasets (e.g., Librispeech [33]) than labeled video datasets. Finally, the VSR model is trained from $\mathcal{D}_{\text{real}} \cup \mathcal{D}_{\text{synth}}$ , where the rich visual and text label information in $\mathcal{D}_{\text{synth}}$ offers significant performance improvement for VSR. + +# 4. Experiments + +# 4.1. Datasets + +We use several public datasets in this work. (1) Audiovisual datasets: LRS3 [2], AVSpeech [14] and VoxCeleb2 [10]; (2) Speech datasets: Librispeech [33], TED-LIUM 3 [20], Common Voice [4]; (3) Facial dataset: CelebA [23]. + +LRS3. We conduct experiments on the LRS3 [2] dataset, which is the largest public benchmark for English VSR containing 438.9 hours of video clips from TED talks (408, 30, and 0.9 hours in the pre-training, training-validation, and test set, respectively). + +AVSpeech & VoxCeleb2. A recent work [25] uses public language classifier [46] and ASR models [6] to filter for English content and obtain pseudo-transcriptions for two multilingual audio-visual datasets: AVSpeech [14] and VoxCeleb2 [10], containing 1,323 hours from AVSpeech and 1,307 hours from VoxCeleb2. We follow the same data setting for our high-resource setting. Furthermore, we filter out AVSpeech videos with large jitter resulting in 933 hours of videos overall for speech-driven lip animation training. + +Datasets for Speech-Driven Lip Animation Training. The lip animation model is trained on a combination of LRS3 (pre-training and training-validation splits) and the English subset (933 hours) of AVSpeech datasets. + +Datasets for Synthetic Data Generation. We use Librispeech [33], TED-LIUM 3 [20], Common Voice (English split) [4] datasets as the speech sources. We split Librispeech audio clips into segments of less than 6 seconds based on the provided silence annotations. For TED-LIUM 3 and Common Voice, we filter out the audio clips which are longer than 20s. As a result, 944 hours, 465 hours, and 2,243 hours of speech data are obtained for Librispeech, TED-LIUM 3, and Common Voice, respectively. We use the CelebA [23] dataset as a source of lip images, which has 202,599 face images and 10,177 identities. For each speech clip, we randomly sample one image from CelebA to generate one synthetic video. We use a static rotation matrix + +in the generation process. In total, 3,652 hours of synthetic video clips are generated for scaling up VSR training. + +# 4.2. Data Processing + +We train a mouth detection module using the Faster R-CNN object detection architecture [38] on the CelebA [23] dataset. We leverage the mouth corner positions provided in the dataset to identify the area around the mouth in each video frame. With this network, we locate and extract a $96 \times 96$ bounding box around the mouth, which we subsequently convert to grayscale. We use the same video pre-processing methods for lip animation and VSR models. For the text vocabulary, we use SentencePiece [22] subword units with a vocabulary size of 5,000. + +# 4.3. Implementation Details + +VSR Model. We consider two model configurations: (1) Conformer-BASE (250M) with 12-layer Conformer encoder, 6-layer Transformer decoder, 768 input dimensions, 3,072 feed-forward dimensions, and 16 attention heads; (2) Conformer-LARGE (783M) with 24-layers Conformer encoder, 9-layer Transformer decoder, 1024 input dimensions, 4,096 feed-forward dimensions, and 16 attention heads. For each configuration, the encoder and decoder have the same dimensions and attention heads. + +We use horizontal flipping, random cropping, and adaptive time masking as the VSR data augmentation methods [28]. We train the BASE and LARGE models for 75 epochs on 64 A100-GPUs with the AdamW [24] optimizer, a cosine learning rate scheduler, and a warm-up of 5 epochs. The peak learning rate is $1 \times 10^{-3}$ and $8 \times 10^{-4}$ for the BASE and LARGE models respectively. The number of frames in each batch is limited to 2,400 and 1,600 frames for the BASE and LARGE models, respectively. Following [25, 28], a pre-trained transformer-based language model is used in the VSR decoding stage. The training configuration is the same as that used in the previous work [25]. + +Speech-Driven Lip Animation. For each training video clip, we sample a sequence of 75 frames that corresponds to 3 seconds. The first frame of the sampled video sequence is used to drive the lip animation. We train the lip animation model for 70 epochs on 32 A100-GPUs with a batch size of 3 per GPU. The details of the module structure in the lip animation model are described in the supplementary material. Adam [21] optimizer is used to train the lip animation model with the learning rate $1 \times 10^{-4}$ for the generator and the frame discriminator. Sequence discriminator uses a smaller learning rate of $1 \times 10^{-5}$ . The weights for each loss term are $\lambda_{disc}^{img} = 1$ , $\lambda_{disc}^{seq} = 0.2$ , $\lambda_{rec} = 300$ in Eq. (5). + +We consider five lip animation model configurations. The lip animation model without VSR perceptual loss is referred to LAM-Baseline. The lip animation model with the BASE VSR model trained on LRS3 is referred to as LAM- + +
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
Afouras et al. [3]CNN✓595‡334-59.8
Ren et al. [37]TransformerX818‡--59.0
Afouras et al. [1]Transformer✓1,519†‡--58.9
Xu et al. [52]RNNX595‡--57.8
Shillingford et al. [44]RNN✓3,886†--55.1
Ma et al. [26]TransformerX4331,759-49.6*
Ma et al. [27]Conformer✓438--46.9
AV-HuBERT-BASE [43]TransformerX301,759-46.1
SynthVSRConformer-BASEX30--104.0
X--3,652100.3
X30-3,65244.7
✓30-3,65243.3
+ +Table 1. Experimental results of low-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. †Includes non-publicly available data. ‡Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours. *Result taken from [43]. + +LRS3-VSR-VL, with the $\lambda_{\text{visual}} = 250$ and $\lambda_{\text{logits}} = 10$ in Eq. (3). Two variants LAM-LRS3-VSR-L and LAM-LRS3-VSR-V are further designed with the $\lambda_{\text{visual}} = 0$ and $\lambda_{\text{logits}} = 0$ , respectively. These two model configurations are used in ablation studies. Last, the lip animation model with the BASE VSR model trained on LRS3 and 2,630 hours of pseudo-labeled AVSpeech and VoxCeleb2 is referred to as LAM-LRS3-AVoX-VSR, with the $\lambda_{\text{visual}} = 500$ and $\lambda_{\text{logits}} = 10$ . + +# 4.4. Low-Resource Labeled Data Setting + +The LAM-Baseline model is used to generate 3,652 hours of synthetic data. We use 30 hours of LRS3 (training-validation) to evaluate the VSR performance when labeled data is scarce. We conduct experiments on the BASE VSR model. The results are reported in Tab. 1. As supervised VSR training from scratch with long sequences often poses optimization problems [1, 11, 28], we first use 30 hours of LRS3 and the 944 hours of Librispeech synthetic data to pre-train a VSR model with the same architecture as [28], then we take the weights of the pre-trained visual front-end to train the BASE VSR model. The pre-training details are described in the supplementary material. + +Training with synthetic data leads to dramatic performance improvement when only 30 hours of labeled data is available. Using 3,652 hours of synthetic data and 30 hours of LRS3 labeled data for training, the BASE model achieves the WER $43.3\%$ (44.7% w/o. language model). Training from 30 hours of LRS3 results in a poor WER of $104\%$ . Our proposed method significantly outperforms the former approaches using hundreds or thousands of data for supervised training [1,27,37,44,52] and self-supervised learning [3, 26]. Furthermore, our method performs better than the AV-HuBERT-BASE [43] model (46.1%) that uses 30 hours of LRS3 and 1,759 hours of unlabeled audio-visual data for self-supervised learning. Notably, we show + +the first attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. In addition, we observe that training only from synthetic data performs poorly (100.3%), which may be caused by the domain mismatch between the real and synthetic data. + +# 4.5. LRS3 Labeled Data Setting + +In this setting, we report the results when using the full 438 hours of LRS3. Experiments are conducted on the BASE VSR model. To avoid the optimization problem, we initialize the weights of the visual front-end from a publicly available Conformer-based model [28] pre-trained on LRS3 using curriculum learning. The results are shown in Tab. 2. + +Our BASE model achieves WER $36.7\%$ when using 438 hours of LRS3 data for training. We generate 3,652 hours of synthetic data using the LAM-LRS3-VSR-VL model. Using 3,652 hours of synthetic data and 438 hours of LRS3 labeled data, the BASE VSR model achieves the WER $27.9\%$ $(28.4\% \mathrm{w / o}$ language model, corresponding to a WER reduction of $8.3\%$ . Our method outperforms three recent approaches using 31,000 $(33.6\%)$ , 1,459 $(31.5\%)$ , and 2,679 $(30.7\%)$ hours of labeled data, respectively. When compared with the state-of-the-art self-supervised method AV-HuBERT [43] that uses additional 1,759 hours of unlabeled audio-visual data, our method outperforms the AV-HuBERT-BASE model $(34.8\%)$ by a large margin. Our method slightly performs better than the AV-HuBERT-LARGE model $(28.6\%)$ , but with fewer model parameters (our BASE model 250M vs AV-HuBERT-LARGE 390M). Note that we compare with the AV-HuBERT results without self-training as we do not use the pseudo-labeled 933 hours of AVSpeech subset for VSR training. + +# 4.6. High-Resource Labeled Data Setting + +We further evaluate the scalability of SynthVSR: when using machine-transcribed AVSpeech [14] and VoxCeleb2 + +
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
AV-HuBERT-BASE [43]TransformerX4331,759-34.8
Makino et al. [30]TransformerX31,000†--33.6
Ma et al. [28]Conformer✓1,459‡--31.5
Prajwal et al. [35]Transformer✓2,676†--30.7
AV-HuBERT-LARGE [43]TransformerX4331,759-28.6
AV-HuBERT-LARGE w. Self-Training [43]TransformerX4331,759-26.9
Auto-AVSR [25]Conformer✓3,448‡--19.1
Serdyuk et al. [42]TransformerX90,000†--25.9
Serdyuk et al. [41]TransformerX90,000†--17.0
SynthVSRConformer-BASEX438--36.7
X438-3,65228.4
✓438-3,65227.9
X3,068--21.2
X3,068-3,65219.4
✓3,068-3,65218.7
SynthVSRConformer-LARGEX3,068-3,65218.2
✓3,068-3,65216.9
+ +[10] as additional training data. We generate 3,652 hours of synthetic data with LAM-LRS3-AVoX-VSR and conduct experiments on BASE and LARGE models. The visual front-end is initialized from a pre-trained model which is the same as that used in the LRS3 labeled data setting. + +We first train the BASE model with 438 hours of labeled LRS3 and 2,630 hours of pseudo-labeled data, resulting in a strong VSR baseline with the WER $21.2\%$ . By using additional 3,652 hours of synthetic data, the WER of the BASE model improves to $18.7\%$ (19.4% w./o. language model), which outperforms [25] that uses additional labeled dataset LRS2 (223 hours) and LRW (157 hours) for training. Although the VSR model has seen a large amount of labeled data, and the speech-driven lip animation model is trained from part of the VSR training data, synthetic data can still lead to considerable performance gains. Furthermore, increasing the model size from BASE to LARGE results in better VSR performance with the WER of $16.9\%$ (18.2% w./o. language model), which is the current state-of-the-art performance on LRS3, with publicly available data only. + +# 4.7. Why SynthVSR Improves VSR? + +We conduct extensive ablation studies to understand the impact of VSR perceptual loss, the diversity of lip sources, and the scale of ASR data for SynthVSR. In addition, we assess the domain mismatch between the real and synthetic data using VSR models. We perform experiments on the LRS3 labeled data setting with the Conformer-BASE VSR model. + +Effect of VSR Perceptual Loss. We experiment with four lip animation model variants: LAM-Baseline, LAM-LRS3-VSR-V, LAM-LRS3-VSR-L, LAM-LRS3-VSR-VL to generate 944 hours of synthetic data from Librispeech, respectively. The WER of the BASE model trained with only 438 + +hours of LRS3 data is $36.7\%$ (see in Tab. 2), the results are shown in Tab. 3. Using 944 hours of synthetic data generated from the LAM-Baseline model, the WER improves to $32.2\%$ . The individual visual VSR loss and linguistic VSR loss improve the WER to $31.7\%$ and $31.4\%$ , respectively. The linguistic VSR loss performs slightly better than the visual one. Further, the combination of visual and linguistic VSR loss further improves the WER to $30.8\%$ , which indicates that the spatio-temporal visual and high-level semantic knowledge from a pre-trained VSR model could offer the lip animation model more accurate lip movements. + +Table 2. Experimental results of LRS3 & high-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. ${}^{ \dagger }$ Includes non-publicly available data. ${}^{§}$ Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours. + +
Lip animation modelWER (%)
LAM-Baseline32.2
LAM-LRS3-VSR-V31.7
LAM-LRS3-VSR-L31.4
LAM-LRS3-VSR-VL30.8
+ +Table 3. Ablation studies on the proposed VSR perceptual loss. + +Effect of Increasing Lip Sources Diversity. To analyze the impact of the diversity of lip sources we generate 944 hours of synthetic data generated from Librispeech using two target face sources: CelebA and LRS3 (pre-training). Images from LRS3 (pre-training) do not introduce any new lip images for VSR training. The results are reported in Tab. 4. Using the synthetic data with LRS3 lip images, the BASE model achieves the WER $32.1\%$ which is $1.3\%$ worse than that with CelebA lip images, which means the external lip images from CelebA offer better WER improvement. + +Effect of the Scale of Speech Corpus. Using synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice with LAM-LRS3-VSR-VL model synthetic data, the LRS3 (test) WER improves from $36.7\%$ to $30.9\%$ , $30.8\%$ , and $30.1\%$ , respectively (see in Tab. 5). When combining + +
Lip sourceUnique imagesWER (%)
LRS3-pretrain5,09032.1
CelebA202,59930.8
+ +all synthetic data together (3,652 hours) for training, the WER improves to $28.4\%$ . We observe that richer speech corpora can offer larger performance improvements. Furthermore, we find that the improvement from scaling speech corpora is more pronounced than when adding VSR perceptual loss and lip sources diversity, suggesting that our proposed method benefits greatly from large speech datasets. + +Table 4. Ablation studies on the impact of lip sources diversity. LRS3-pretrain indicates using the LRS3 (pre-training) images. + +
Training dataHoursWER (%)
LRS343836.7
LRS3 + TED-Synth90330.9
LRS3 + LBS-Synth1,38230.8
LRS3 + CV-Synth2,68130.1
LRS3 + [TED + LBS + CV]-Synth4,09028.4
+ +Assessment of the Domain Mismatch Using VSR. We use 944 hours of Librispeech synthetic data generated from LAM-Baseline and LAM-LRS3-VSR-VL for VSR training. We conduct an assessment on the LRS3 real and synthetic test sets, the synthetic test set is obtained by passing the speech and the first frame in the real test set to the lip animation model. The results are illustrated in Fig. 3. By introducing the VSR perceptual loss, the WER on the synthetic test set is reduced from $106.8\%$ to $49.7\%$ , which indicates the VSR loss minimizes the domain gap between real and synthetic data. After training on synthetic data, the WER reduces for both real and synthetic data. Specifically, the WER is improved substantially from $106.8\%$ to $36.0\%$ and from $49.7\%$ to $16.8\%$ on synthetic test sets generated using LAM-Baseline and LAM-LRS3-VSR-VL, respectively. We observe that domain mismatch can be further reduced by combining real and synthetic data for VSR training. + +# 4.8. SynthVSR with TTS-Generated Speech + +In the real world, text corpora (e.g., Wikipedia) are more accessible than transcribed speech corpora. To explore the potential of using TTS-generated speech for SynthVSR, we use the off-the-shelf TTS model FastSpeech 2 [39] to generate 944 hours of synthetic speech from Librispeech transcriptions, which is further used to generate synthetic video clips using LAM-LRS3-VSR-VL model. Using these 944 hours of synthetic data to train the BASE model under the LRS3 labeled data setting, the WER is improved + +![](images/38b8a320d1f90e152d48dd0fd74848a89bf715be9b4abfdbd87789a0310d57c6.jpg) +Figure 3. Assessment of the domain mismatch between real and synthetic video data using VSR models trained with only real data (VSR-Real, black) and real data together with synthetic data (VSR-Real-Synth, red), respectively. Synth-Test and Real-Test refer to the synthetic and real LRS3 (test) sets. + +from $36.7\%$ to $32.9\%$ , as shown in Tab. 6. Although TTS-generated speech is not as natural as the original LibriSpeech data, the improvement of using synthetic data is still decent $(3.8\%)$ , indicating the further potential of SynthVSR in real-world applications where labeled video data is sparse, such as the healthcare industry. + +Table 5. Ablation studies on different speech corpus. TED-Synth, LBS-Synth, and CV-Synth indicates the synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice, respectively. + +
Training dataHoursWER (%)
LRS343836.7
LRS3 + LBS-Synth1,38230.8
LRS3 + TTS-LBS-Synth1,38232.9
+ +Table 6. Experiments results of using the synthetic speech generated synthetic video (TTS-LBS-Synth) for VSR training. + +# 5. Limitations & Societal Impact + +The excellent performance improvement of SynthVSR comes at the cost of high computational demands during training, as we use additional large-scale synthetic video data. Also, although SynthVSR has achieved state-of-the-art VSR performance, the LRS3 test set is relatively limited (0.9 hours), which is from TED talks, while real-world videos are more challenging (e.g., egocentric videos). SynthVSR has many positive real-world applications, such as helping the hearing-impaired or people with aphonia with everyday communication. + +# 6. Conclusion + +We have presented a semi-supervised method for VSR enhanced with synthetic lip movements. The speech-driven lip animation model is proposed to generate synthetic video data from labeled speech datasets and face images for scaling up VSR. Our method achieves state-of-the-art results on LRS3, outperforming prior work trained on more labeled or unlabeled real video data. Our work fosters future research on generating and exploiting synthetic visual data for VSR. + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2, 6 +[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: A large-scale dataset for visual speech recognition. arXiv:1809.00496, 2018. 2, 3, 5 +[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2143-2147, 2020. 2, 3, 6 +[4] Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Conference on Language Resources and Evaluation, 2020. 2, 5 +[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. LipNet: Sentence-level lipreading. arXiv:1611.01599, 2016. 2, 3 +[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 5 +[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 3 +[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision, pages 520-535, 2018. 3 +[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? In British Machine Vision Conference, 2017. 3 +[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 3, 5, 7 +[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, pages 87-103. Springer, 2016. 2, 6 +[12] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: Learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 3 +[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3D human pose estimation: Motion to the rescue. In Advances in Neural Information Processing Systems, volume 32, 2019. 3 +[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv:1804.03619, 2018. 3, 5, 6 + +[15] Amin Fazel, Wei Yang, Yulan Liu, Roberto Barra-Chicote, Yixiong Meng, Roland Maas, and Jasha Droppo. SynthASR: Unlocking synthetic data for speech recognition. arXiv:2106.07803, 2021. 3 +[16] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv:2007.04954, 2020. 3 +[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine learning, pages 369-376, 2006. 2 +[18] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. In INTER-SPEECH, 2020. 2, 3 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3 +[20] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In International Conference on Speech and Computer, pages 198-208, 2018. 2, 5 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5 +[22] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018. 5 +[23] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision, December 2015. 2, 5 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 5 +[25] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-visual speech recognition with automatic labels. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2023. 2, 5, 7 +[26] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In IN-TERSPEECH, 2021. 2, 3, 6 +[27] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 7613-7617, 2021. 2, 3, 6 +[28] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nature Machine Intelligence, 2022. 2, 3, 5, 6, 7 +[29] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song. + +Contrastive learning of global and local audio-visual representations. arXiv:2104.05418, 2021.3 +[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE Automatic Speech Recognition and Understanding Workshop, pages 905–912, 2019. 7 +[31] Sepand Mavandadi, Tara N Sainath, Kevin Hu, and Zelin Wu. A deliberation-based joint acoustic and text decoder. In INTERSPEECH, 2021. 3 +[32] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging uni-modal self-supervised learning for multimodal audio-visual speech recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 3 +[33] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 5206-5210, 2015. 2, 5 +[34] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3D models. In Proceedings of the IEEE International Conference on Computer Vision, pages 1278-1286, 2015. 3 +[35] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 3, 7 +[36] Aayush Prakash, Shaad Boochoon, Mark Brophy, David Acuna, Eric Cameracci, Gavriel State, Omer Shapira, and Stan Birchfield. Structured domain randomization: Bridging the reality gap by context-aware synthetic data. In 2019 International Conference on Robotics and Automation, pages 7249-7255, 2019. 3 +[37] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6 +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5 +[39] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv:2006.04558, 2020. 8 +[40] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3234-3243, 2016. 3 +[41] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition. In INTERSPEECH, 2022. 2, 3, 7 +[42] Dmitriy Serdyuk, Olivier Siohan, and Otavio de Pinho Forin Braga. Audio-visual speech recognition is worth $32 \times 32 \times 8$ voxels. In IEEE Automatic Speech Recognition and Under + +standing Workshop, 2021. 2, 3, 7 +[43] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In International Conference on Learning Representations, 2022. 2, 3, 6, 7 +[44] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. In INTERSPEECH, 2019. 2, 3, 6 +[45] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. arXiv:1703.04105, 2017. 3 +[46] Jörgen Valk and Tanel Alumäe. Voxlingua107: A dataset for spoken language recognition. In 2021 IEEE Spoken Language Technology Workshop, pages 652-658, 2021. 5 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3 +[48] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. StyleGAN2 distillation for feed-forward image manipulation. In European Conference on Computer Vision, pages 170-186. Springer, 2020. 4 +[49] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven facial animation with temporal GANs. In British Machine Vision Conference, 2018. 2, 3 +[50] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with GANs. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2, 3 +[51] Zhonghao Wang, Mo Yu, Yunchao Wei, Rogerio Feris, Jinjun Xiong, Wen-mei Hwu, Thomas S Huang, and Honghui Shi. Differential treatment for stuff and things: A simple unsupervised domain adaptation method for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12635-12644, 2020. 3 +[52] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14433-14442, 2020. 2, 6 +[53] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 2 +[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 3 +[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 3 \ No newline at end of file diff --git a/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/images.zip b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0b3fc01165dc506eb96a99603e5ef2085e21404f --- /dev/null +++ b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6b47ee9772585d2f7ebc0b4af1276efe44dd5ee4b25e1e047070382f464def8 +size 345929 diff --git a/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/layout.json b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..53b91af04c480e868b8ad62797b511b8e1494e05 --- /dev/null +++ b/2023/SynthVSR_ Scaling Up Visual Speech Recognition With Synthetic Supervision/layout.json @@ -0,0 +1,9067 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 536, + 121 + ], + "type": "text", + "content": "SynthVSR: Scaling Up Visual Speech Recognition With Synthetic Supervision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "spans": [ + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": "Xubo Liu" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Egor Lakomkin" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Konstantinos Vougioukas" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Pingchuan Ma" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Honglie Chen" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Ruiming Xie" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Morrie Doulaty" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Niko Moritz" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Jachym Kolar" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Stavros Petridis" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Maja Pantic" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "text", + "content": ", Christian Fuegen" + }, + { + "bbox": [ + 52, + 142, + 564, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 177, + 394, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 177, + 394, + 192 + ], + "spans": [ + { + "bbox": [ + 196, + 177, + 394, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 177, + 394, + 192 + ], + "type": "text", + "content": "University of Surrey " + }, + { + "bbox": [ + 196, + 177, + 394, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 177, + 394, + 192 + ], + "type": "text", + "content": "Meta AI" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 204, + 209, + 338 + ], + "blocks": [ + { + "bbox": [ + 48, + 204, + 209, + 338 + ], + "lines": [ + { + "bbox": [ + 48, + 204, + 209, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 204, + 209, + 338 + ], + "type": "image", + "image_path": "9c83aa001db9bf9b18f74e94f33c8cbbf5848c6198b7eb2f947af666b6faf19d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 354, + 546, + 420 + ], + "lines": [ + { + "bbox": [ + 45, + 354, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 45, + 354, + 546, + 420 + ], + "type": "text", + "content": "Figure 1. Scaling up visual speech recognition with synthetic supervision (SynthVSR): we propose SynthVSR, a semi-supervised framework that can substantially improve the performance of VSR models by using synthetic lip movements. Firstly, we introduce a speech-driven lip animation model that generates lip movement videos conditioned on input lip images and speech utterances (left). Secondly, we generate large-scale synthetic videos using transcribed speech datasets and lip images. The combination of synthetic videos and their corresponding speech transcriptions constitutes the synthetic dataset (centre). Finally, we conduct semi-supervised VSR training with synthetic and real datasets. Our method substantially improves the performance of VSR models with large-scale synthetic data (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 213, + 204, + 389, + 338 + ], + "blocks": [ + { + "bbox": [ + 213, + 204, + 389, + 338 + ], + "lines": [ + { + "bbox": [ + 213, + 204, + 389, + 338 + ], + "spans": [ + { + "bbox": [ + 213, + 204, + 389, + 338 + ], + "type": "image", + "image_path": "2204cfd53a62e8188cab6ef04555d56ed95e5a2eaa6a468886cc70f85aabeea3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 204, + 544, + 338 + ], + "blocks": [ + { + "bbox": [ + 389, + 204, + 544, + 338 + ], + "lines": [ + { + "bbox": [ + 389, + 204, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 389, + 204, + 544, + 338 + ], + "type": "image", + "image_path": "c690f05f11549ce55d61c18a411de52117a0004767962da6952c7f5cec4a9da4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 430, + 192, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 430, + 192, + 442 + ], + "spans": [ + { + "bbox": [ + 143, + 430, + 192, + 442 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 450, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 689 + ], + "type": "text", + "content": "Recently reported state-of-the-art results in visual speech recognition (VSR) often rely on increasingly large amounts of video data, while the publicly available transcribed video datasets are limited in size. In this paper, for the first time, we study the potential of leveraging synthetic visual data for VSR. Our method, termed SynthVSR, substantially improves the performance of VSR systems with synthetic lip movements. The key idea behind SynthVSR is to leverage a speech-driven lip animation model that generates lip movements conditioned on the input speech. The speech-driven lip animation model is trained on an unlabeled audio-visual dataset and could be further optimized towards a pre-trained VSR model when labeled videos are available. As plenty of transcribed acoustic data and face images are available, we are able to generate large-scale synthetic data using the proposed lip animation model for semi-supervised VSR training. We evaluate the performance of our approach on the largest public VSR benchmark - Lip Reading Sentences 3 (LRS3). SynthVSR achieves a WER of " + }, + { + "bbox": [ + 46, + 450, + 289, + 689 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 46, + 450, + 289, + 689 + ], + "type": "text", + "content": " with only 30 hours of real labeled data," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "type": "text", + "content": "outperforming off-the-shelf approaches using thousands of hours of video. The WER is further reduced to " + }, + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "type": "inline_equation", + "content": "27.9\\%" + }, + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "type": "text", + "content": " when using all 438 hours of labeled data from LRS3, which is on par with the state-of-the-art self-supervised AV-HuBERT method. Furthermore, when combined with large-scale pseudo-labeled audio-visual data SynthVSR yields a new state-of-the-art VSR WER of " + }, + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 306, + 431, + 547, + 574 + ], + "type": "text", + "content": " using publicly available data only, surpassing the recent state-of-the-art approaches trained with 29 times more non-public machine-transcribed video data (90,000 hours). Finally, we perform extensive ablation studies to understand the effect of each component in our proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 386, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 386, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 386, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "Visual speech recognition (VSR), also known as lip reading, is the task of recognizing speech content based on visual lip movements. VSR has a wide range of applications in real-world scenarios such as helping the hearing-impaired perceive human speech and improving automatic speech recognition (ASR) in noisy environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "VSR is a challenging task, as it requires capturing speech from high-dimensional spatio-temporal videos, while multiple words are visually ambiguous (e.g., \"world\" and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 208, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 208, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 208, + 712 + ], + "type": "text", + "content": "*Work done during an internship at Meta AI." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18806" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "\"word\") in the visual streams. Recently, with the release of large-scale transcribed audio-visual datasets such as LRS2 [1] and LRS3 [2], deep neural networks have become the mainstream approach for VSR. However, even the largest public dataset for English VSR, LRS3, does not exceed 500 hours of transcribed video data. The lack of large-scale transcribed audio-visual datasets potentially results in VSR models which could only work in a laboratory environment i.e. limited word vocabulary and lip sources diversity [28]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "type": "text", + "content": "A common solution to this issue is to collect and annotate large-scale audio-visual datasets. For example, [41, 42] collected 90,000 hours of YouTube videos with user-uploaded transcriptions to achieve state-of-the-art performance on standard benchmarks. However, such a procedure is expensive and time-consuming, especially for most of the world's 7,000 languages [43]. If annotations are missing, the ASR can be used to generate the transcriptions automatically and this has been shown to be an effective approach to significantly improve VSR performance [28]. The other promising direction is to learn audio-visual speech representations from large amounts of parallel unlabeled audio-visual data in a self-supervised approach, and then fine-tune them on the limited labeled video dataset [43]. Nevertheless, publicly available video datasets are also limited and their usage may raise license-related concerns, barring their use in commercial applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 576 + ], + "type": "text", + "content": "Human perception of speech is inherently multimodal, involving both audition and vision [43]. ASR, which is a complementary task to VSR, has achieved impressive performance in recent years, with tens of thousands of hours of annotated speech datasets [4, 20, 33] available for large-scale training. It is intuitive to ask: Can we improve VSR with large amounts of transcribed acoustic-only ASR training data? The key to this question is to take advantage of recent advances in speech-driven visual generative models [49, 50]. By leveraging visual generative models, we can produce parallel synthetic videos for large-scale labeled audio datasets. Synthetic videos provide advantages such as having control over the target text and lip image as well as the duration of a generated utterance. To the best of our knowledge, the potential of leveraging synthetic visual data for improving VSR has never been studied in the literature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 576, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 289, + 685 + ], + "type": "text", + "content": "In this work, we present SynthVSR, a novel semi-supervised framework for VSR. In particular, we first propose a speech-driven lip animation model that can generate synthetic lip movements video conditioned on the speech content. Next, the proposed lip animation model is used to generate synthetic video clips from transcribed speech datasets (e.g., Librispeech [33]) and human face datasets (e.g., CelebA [23]). Then, the synthetic videos together with the corresponding transcriptions are used in combi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 228 + ], + "type": "text", + "content": "nation with the real video-text pairs (e.g., LRS3 [2]) for large-scale semi-supervised VSR training. The pipeline of SynthVSR is illustrated in Fig. 1. Unlike existing studies in exploiting unlabeled video data for VSR using methods such as pseudo-labeling [28] and self-supervised learning [43], we use the unlabeled audio-visual data to train a cross-modal generative model in order to bridge ASR training data and VSR. Furthermore, we propose to optimize the lip animation model towards a pre-trained VSR model when labeled videos are available. We empirically demonstrate that the semantically high level, spatio-temporal supervision signal from the pre-trained VSR model offers the lip animation model more accurate lip movements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "text", + "content": "SynthVSR achieves remarkable performance gains with labeled video data of different scales. We evaluate the performance of SynthVSR on the largest public VSR benchmark LRS3 with a Conformer-Transformer encoder-decoder VSR model [28]. In the low-resource setup using only 30 hours of labeled video data from LRS3 [2], our approach achieves a VSR WER of " + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "text", + "content": ", substantially outperforming the former VSR methods using hundreds or thousands of hours of video data for supervised training [1,27,37,44,52] and self-supervised learning [3,26,43]. Notably, we demonstrate the first successful attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. Using the complete 438 hours from LRS3 further improves WER to " + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "inline_equation", + "content": "27.9\\%" + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "text", + "content": " which is on par with the state-of-the-art self-supervised method AV-HuBERT-LARGE [43] that uses external 1,759 hours of unlabeled audio-visual data, but with fewer model parameters. Furthermore, following a recent high-resource setup [25] which uses additional 2,630 hours of ASR pseudo-labeled publicly available audio-visual data, our proposed method yields a new state-of-the-art VSR WER of " + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 304, + 228, + 547, + 502 + ], + "type": "text", + "content": ", surpassing the former state-of-the-art approaches [41, 42] trained on 90,000 hours of non-public machine-transcribed data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 503, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 547, + 576 + ], + "type": "text", + "content": "Finally, we present extensive ablation studies to analyze where the improvement of SynthVSR comes from (e.g., the diversity of lip sources, the scale of ASR data). We also show considerable VSR improvement using synthetic video data derived from Text-To-Speech (TTS)-generated speech, indicating the great potential of our method for VSR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 586, + 392, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 392, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 392, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Visual Speech Recognition. VSR has achieved remarkable progress with the success of deep learning and the availability of audio-visual datasets such as LRS2 [1] and LRS3 [2]. Progress was driven by adapting ASR approaches such as sequence-to-sequence models [11] and the Connectionist Temporal Classification (CTC) [5, 17] training objective. Recent studies have achieved noticeable results in VSR by using Transformer-based architecture [1], a convolutional variant [53], Conformer-based models [18], and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 715 + ], + "type": "text", + "content": "1Such as LRW [11] and LRS2 [1] datasets which are only permitted for the purpose of academic research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18807" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 346 + ], + "type": "text", + "content": "an attention-based feature pooling method [35]. The recent advances in VSR models are mainly dependent on leveraging increasingly large-scale transcribed non-public video datasets such as LSVSR (3,800 hours) [44] and YT31k (31,000 hours) [28]. Some recent works [41, 42] use 90,000 hours of non-publicly machine-transcribed video data to achieve state-of-the-art performance on standard benchmarks. Another popular trend is to leverage unlabeled audio-visual datasets such as AVSpeech [14] and VoxCeleb2 [10] with semi- or self-supervised learning. Using pre-trained ASR models to transcribe audio streams in unlabeled audio-visual datasets i.e., pseudo-labeling has demonstrated great potential to improve VSR performance. Such pseudo-labeled video data can be used for supervised training [28] or knowledge distillation [3]. Self-supervised approaches such as contrastive learning [12, 29, 32] and cross-modal feature prediction [26] have achieved significant improvement by exploiting unlabeled audio-visual data. AVHuBERT [43] is the state-of-the-art self-supervised method that predicts iteratively refined cluster assignments from masked audio-visual streams and achieves impressive VSR performance, especially when only limited labeled video data is available (e.g., 30 hours)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 348, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 289, + 552 + ], + "type": "text", + "content": "Speech-Driven Facial Animation. Speech-driven facial animation aims to produce a talking face video conditioned on speech content. With the advances in deep visual generative models, speech-driven facial animation has significantly improved over the past years. Speech2Vid [9] is the first end-to-end neural network-based method for speech-driven facial animation. This method is optimized towards a pixel-level " + }, + { + "bbox": [ + 47, + 348, + 289, + 552 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 348, + 289, + 552 + ], + "type": "text", + "content": " reconstruction loss which results in blurry generated frames. Recently, Generative Adversarial Networks (GANs) have made impressive progress [49,50] in facial animation, as they are able to generate sharp and highly textured facial frames. Initially, most of these methods synthesized frontal faces and focused on improving lip-sync [8, 50], preserving the identity of the target speaker [54]. As the generation of synthetic frontal faces improved, researchers began to focus on generating talking heads that could handle natural head movements [7, 55]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 552, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 289, + 660 + ], + "type": "text", + "content": "Learning from Synthetic Supervision. Training with synthetic data has attracted increasing research attention in recent years. Synthetic data could offer improved data diversity [13], which is useful in many tasks where labeled data is limited, such as image classification [16], objection detection [34, 36], semantic segmentation [40, 51], human pose estimation [13], and ASR [15, 31]. Unlike previous work, we focus on a new multimodal problem: how to produce and leverage synthetic video data for improving VSR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 669, + 117, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 117, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 117, + 684 + ], + "type": "text", + "content": "3. SynthVSR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "In this section, we introduce SynthVSR, a novel semi-supervised VSR framework to improve the performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "of VSR models using synthetic data, as shown in Fig. 1. We propose a speech-driven lip animation model trained on an audio-visual dataset, which is used to produce synthetic videos with large-scale transcribed speech and face datasets for scaling up VSR training. We will introduce the VSR model, the speech-driven lip animation model, and the semi-supervised VSR training next." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 164, + 384, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 384, + 176 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 384, + 176 + ], + "type": "text", + "content": "3.1. VSR Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 182, + 547, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 328 + ], + "type": "text", + "content": "The VSR model we use in this work is based on [27,28], which has achieved the state-of-the-art VSR performance on LRS3 [2] without the use of external data. The baseline VSR model is an encoder-decoder architecture. In particular, the encoder is comprised of two components, the visual front-end (a 3D convolutional layer followed by a ResNet-18 model [19,45]) and a Conformer [18] encoder. The decoder is based on the transformer architecture [47]. The baseline VSR model is trained end-to-end using a combination of the CTC loss [5,44] with an attention-based Cross-Entropy (CE) loss. Model details are described in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 335, + 470, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 470, + 348 + ], + "type": "text", + "content": "3.2. Speech-Driven Lip Animation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 486 + ], + "type": "text", + "content": "Inspired by the recent advances in speech-driven facial animation [49, 50, 54, 55], we propose an approach for speech-driven lip animation that generates videos of talking mouth regions conditioned on speech utterances. The output space of the lip animation model is the same as the VSR input space. The proposed lip animation model is based on a temporal GAN [49, 50] with two discriminators. We further propose a VSR perceptual loss when labeled video data is available. The architecture of the speech-driven lip animation model is illustrated in the left part of Fig. 2. We will introduce each component in the next sections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": "Generator. The generator " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " is an encoder-decoder structure, as shown in the right part of the Fig. 2. The generator uses the first frame of a video clip, a speech clip, and a head rotation sequence as inputs. The head rotation is used as the additional condition which helps better model the lip movements as most training videos are not static. The speech clip is divided into overlapping " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "200\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " chunks with a stride of " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "40~\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": ". The generator produces the corresponding video frame for each speech chunk. In the generator, an image encoder " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " and a speech encoder " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " are used to capture the visual information and speech context into latent embeddings " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": ", respectively. The image encoder " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " uses a stack of 2D convolutional layers to extract the visual embedding. The speech encoder " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " comprises a stack of 1D convolutional layers followed by a stack of GRU layers. The head rotations are provided in the form of sequences of 3D rotation matrices [55] " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "z_{r}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " with respect to the first frame. The three embeddings " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "z_{i},z_{s}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "inline_equation", + "content": "z_{r}" + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": " are concatenated and used to modulate the convolutional lay" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18808" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 324, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 324, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 324, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 324, + 220 + ], + "type": "image", + "image_path": "2af223e0bfa6526d1fa8498868f7d7c02a96e420565afa1521436be15bf89821.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 254 + ], + "type": "text", + "content": "Figure 2. Architecture of proposed speech-driven lip animation model. Left: GAN-based speech-driven lip animation model generating lip movements given a lip image, a speech utterance, and a rotation sequence; Right: structure of the generator in the lip animation model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 335, + 74, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 335, + 74, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 335, + 74, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 545, + 213 + ], + "type": "image", + "image_path": "a4ba1df8a8020a716329db45853c1667c4e5f63903a24aa036615d0d0bd59e65.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "type": "text", + "content": "ers in the frame decoder " + }, + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "type": "inline_equation", + "content": "D_{frame}" + }, + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "type": "text", + "content": ", which is similar to StyleGAN2 [48]. Different from StyleGAN2, which generates frames from a learned constant input, we use the penultimate layer features of the image encoder " + }, + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 46, + 259, + 286, + 307 + ], + "type": "text", + "content": " as the input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": "Discriminators. The speech-driven lip animation system has two discriminators: frame discriminator " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": " and sequence discriminator " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "D_{seq}" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": ". The frame discriminator is a stack of CNN layers that operates on the image frame level. The original first frame is concatenated channel-wise to the target frame to form the input of the frame discriminator. This helps enforce visual consistency. The sequence discriminator operates on the sequence level to ensure the temporal consistency of synthetic lip movements. The sequence discriminator uses spatio-temporal convolutions to encode the image sequence, followed by GRU layers to determine if the sequence is real or not. Specifically, the frame discriminator " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": " is trained on frames that are uniformly sampled from a video " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": " using a sampling function " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "S(v)" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": ". The first frame " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "v_1" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": " is fed to the " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": " as the condition. The input speech signal is " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": ". The adversarial loss of the " + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 46, + 307, + 287, + 509 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 516, + 247, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 516, + 247, + 544 + ], + "spans": [ + { + "bbox": [ + 87, + 516, + 247, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D i s c} ^ {i m g} = \\mathbb {E} _ {v} \\left[ \\log D _ {i m g} (S (v), v _ {1}) \\right] \\\\ + \\mathbb {E} _ {v, s} [ \\log (1 - D _ {i m g} (S (G (s, v _ {1})), v _ {1}) ]. \\\\ \\end{array}", + "image_path": "8e80d19a8628b16740dc0ed8aaca446b49cb766c8fa952d3e3b47319f48e74aa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "inline_equation", + "content": "D_{seq}" + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "content": " operates on the entire sequence video " + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "content": ". The adversarial loss of the " + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "inline_equation", + "content": "D_{seq}" + }, + { + "bbox": [ + 47, + 552, + 287, + 574 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 582, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 582, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 287, + 611 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {D i s c}} ^ {\\text {s e q}} = \\mathbb {E} _ {v} [ \\log D _ {\\text {s e q}} (v) ] + \\\\ \\mathbb {E} _ {v, s} [ \\log (1 - D _ {\\text {s e q}} (G (s, v _ {1})) ]. \\tag {2} \\\\ \\end{array}", + "image_path": "6807c9581912fd581fe09a91f75480863e6b2dd573508ae6062c04b14e40d81c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "VSR Perceptual Loss. We further propose to optimize the lip animation model towards a VSR perceptual loss if labeled video data is available. We first pre-train a VSR model as introduced in Sec. 3.1. The proposed VSR perceptual loss corresponds to a weighted sum of feature distances computed from the visual front-end and the Transformer decoder of the pre-trained VSR model for real and generated samples. We use " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " norm to measure the visual embedding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 259, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 295 + ], + "type": "text", + "content": "distance and Kullback-Leibler (KL) divergence to measure the logits distribution distance, respectively. The VSR perceptual loss is obtained by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 299, + 545, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 299, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 299, + 545, + 315 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V S R} = \\lambda_ {\\text {v i s u a l}} \\left\\| z _ {f} ^ {r} - z _ {f} ^ {s} \\right\\| _ {1} + \\lambda_ {\\text {l o g i t s}} \\operatorname {K L} \\left(\\hat {y} ^ {r}, \\hat {y} ^ {s}\\right), \\tag {3}", + "image_path": "6153e72fdeb31169bda01a98e64ba917d2316c1d1413118e2ab16bc66ed3a455.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "z_{f}^{r}" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "z_{f}^{s}" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " are the VSR front-end visual features of the real and synthetic video, respectively, " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\hat{y}^{r}" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\hat{y}^{s}" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " is the VSR predicted logits distribution of real and synthetic video, respectively, " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{visual}}" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{logits}}" + }, + { + "bbox": [ + 304, + 323, + 545, + 395 + ], + "type": "text", + "content": " control the weights of these two perceptual losses. The VSR model is frozen during the lip animation model training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "content": "Training Objectives. The speech-driven lip animation model is trained using a combination of a reconstruction loss, adversarial losses, and a VSR perceptual loss. The reconstruction loss is computed based on the " + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "content": " distance between the generated video " + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\hat{v}" + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "content": " and ground truth video " + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 395, + 545, + 455 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 388, + 464, + 545, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 464, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 388, + 464, + 545, + 477 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\| v - \\hat {v} \\| _ {1}. \\tag {4}", + "image_path": "09ba0f7a24e9d57a0c6ce91e0e16cfe31ed49f8ab277e5dcfdd958b4a2778754.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 485, + 529, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 529, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 529, + 497 + ], + "type": "text", + "content": "The overall training loss for the lip animation model is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 505, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 505, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 344, + 505, + 545, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {A n i m a t i o n}} = \\lambda_ {\\text {d i s c}} ^ {\\text {i m g}} \\mathcal {L} _ {\\text {d i s c}} ^ {\\text {i m g}} + \\lambda_ {\\text {d i s c}} ^ {\\text {s e q}} \\mathcal {L} _ {\\text {d i s c}} ^ {\\text {s e q}} \\tag {5} \\\\ + \\lambda_ {r e c} \\mathcal {L} _ {r e c} + \\mathcal {L} _ {V S R}, \\\\ \\end{array}", + "image_path": "61c8ac2b11be6dd79bbf0fd6f6b293cc0752a223c843722088b44f309ffc12e1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{disc}^{img},\\lambda_{disc}^{seq}" + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{rec}" + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "text", + "content": " represent the coefficient of the adversarial loss of frame discriminator " + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "text", + "content": ", the adversarial loss of sequence discriminator " + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "inline_equation", + "content": "D_{seq}" + }, + { + "bbox": [ + 304, + 544, + 545, + 592 + ], + "type": "text", + "content": ", and the reconstruction loss, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 599, + 527, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 527, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 527, + 612 + ], + "type": "text", + "content": "3.3. Semi-Supervised VSR with Synthetic Data" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "In a typical supervised setting, a VSR model is trained from the labeled dataset " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{real} = \\{(v_i^r,y_i^r)\\}_{i = 1}^{n_r}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "n_r" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " is the number of paired video clips " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{i}^{r}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and its transcriptions " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "y_{i}^{r}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". In SynthVSR, we first train a speech-driven lip animation model on an unlabeled audio-visual dataset " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{av} = \\{(v_i^{av},s_i^{av})\\}_{i = 1}^{n_{av}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "n_{av}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " is the number of paired video clips " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{i}^{av}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and speech utterances " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "s_i^{av}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". Furthermore, we can use the trained generator " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " to generate synthetic video" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18809" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": "from a labeled speech dataset " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{(s_i^s,y_i^s)\\}_{i = 1}^{n_s}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " and a cropped lips dataset " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f = \\{c_i^f\\}_{i = 1}^{n_f}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "n_{s}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " is the number of paired speech clips " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "s_i^s" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " and its transcriptions " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "y_{i}^{s}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " is the number of lip images " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "c_{i}^{f}" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": ". For each speech clip " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "s_i^s" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": ", the generator " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": " generates its parallel synthetic video " + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "inline_equation", + "content": "\\hat{v}_i^s" + }, + { + "bbox": [ + 47, + 72, + 287, + 135 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 142, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 142, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 287, + 156 + ], + "type": "interline_equation", + "content": "G \\left(s _ {i} ^ {s}, S \\left(\\mathcal {D} _ {f}\\right)\\right) \\mapsto \\hat {v} _ {i} ^ {s}, \\tag {6}", + "image_path": "2570f3cf4d0830e0679ebced374e270f9ec9a44d700c74feee512f7a2db9ffd9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "content": " is a uniform sampling function. Then, the synthetic dataset " + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{synth}} = \\{\\hat{v}_i^s, y_i^s\\}_{i=1}^{n_s}" + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "content": " is obtained, where in general " + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "n_s >> n_r" + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "content": " as there are far more annotated speech datasets (e.g., Librispeech [33]) than labeled video datasets. Finally, the VSR model is trained from " + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{real}} \\cup \\mathcal{D}_{\\text{synth}}" + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "content": ", where the rich visual and text label information in " + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{synth}}" + }, + { + "bbox": [ + 47, + 163, + 287, + 247 + ], + "type": "text", + "content": " offers significant performance improvement for VSR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 128, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 128, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 128, + 270 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 276, + 110, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 110, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 110, + 288 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 294, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 287, + 343 + ], + "type": "text", + "content": "We use several public datasets in this work. (1) Audiovisual datasets: LRS3 [2], AVSpeech [14] and VoxCeleb2 [10]; (2) Speech datasets: Librispeech [33], TED-LIUM 3 [20], Common Voice [4]; (3) Facial dataset: CelebA [23]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 343, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 287, + 403 + ], + "type": "text", + "content": "LRS3. We conduct experiments on the LRS3 [2] dataset, which is the largest public benchmark for English VSR containing 438.9 hours of video clips from TED talks (408, 30, and 0.9 hours in the pre-training, training-validation, and test set, respectively)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 403, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 510 + ], + "type": "text", + "content": "AVSpeech & VoxCeleb2. A recent work [25] uses public language classifier [46] and ASR models [6] to filter for English content and obtain pseudo-transcriptions for two multilingual audio-visual datasets: AVSpeech [14] and VoxCeleb2 [10], containing 1,323 hours from AVSpeech and 1,307 hours from VoxCeleb2. We follow the same data setting for our high-resource setting. Furthermore, we filter out AVSpeech videos with large jitter resulting in 933 hours of videos overall for speech-driven lip animation training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 510, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 557 + ], + "type": "text", + "content": "Datasets for Speech-Driven Lip Animation Training. The lip animation model is trained on a combination of LRS3 (pre-training and training-validation splits) and the English subset (933 hours) of AVSpeech datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "type": "text", + "content": "Datasets for Synthetic Data Generation. We use Librispeech [33], TED-LIUM 3 [20], Common Voice (English split) [4] datasets as the speech sources. We split Librispeech audio clips into segments of less than 6 seconds based on the provided silence annotations. For TED-LIUM 3 and Common Voice, we filter out the audio clips which are longer than 20s. As a result, 944 hours, 465 hours, and 2,243 hours of speech data are obtained for Librispeech, TED-LIUM 3, and Common Voice, respectively. We use the CelebA [23] dataset as a source of lip images, which has 202,599 face images and 10,177 identities. For each speech clip, we randomly sample one image from CelebA to generate one synthetic video. We use a static rotation matrix" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "in the generation process. In total, 3,652 hours of synthetic video clips are generated for scaling up VSR training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 105, + 404, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 105, + 404, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 105, + 404, + 118 + ], + "type": "text", + "content": "4.2. Data Processing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 123, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 243 + ], + "type": "text", + "content": "We train a mouth detection module using the Faster R-CNN object detection architecture [38] on the CelebA [23] dataset. We leverage the mouth corner positions provided in the dataset to identify the area around the mouth in each video frame. With this network, we locate and extract a " + }, + { + "bbox": [ + 304, + 123, + 545, + 243 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 304, + 123, + 545, + 243 + ], + "type": "text", + "content": " bounding box around the mouth, which we subsequently convert to grayscale. We use the same video pre-processing methods for lip animation and VSR models. For the text vocabulary, we use SentencePiece [22] subword units with a vocabulary size of 5,000." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 251, + 440, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 440, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 440, + 264 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 270, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 376 + ], + "type": "text", + "content": "VSR Model. We consider two model configurations: (1) Conformer-BASE (250M) with 12-layer Conformer encoder, 6-layer Transformer decoder, 768 input dimensions, 3,072 feed-forward dimensions, and 16 attention heads; (2) Conformer-LARGE (783M) with 24-layers Conformer encoder, 9-layer Transformer decoder, 1024 input dimensions, 4,096 feed-forward dimensions, and 16 attention heads. For each configuration, the encoder and decoder have the same dimensions and attention heads." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": "We use horizontal flipping, random cropping, and adaptive time masking as the VSR data augmentation methods [28]. We train the BASE and LARGE models for 75 epochs on 64 A100-GPUs with the AdamW [24] optimizer, a cosine learning rate scheduler, and a warm-up of 5 epochs. The peak learning rate is " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "inline_equation", + "content": "8 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 378, + 546, + 521 + ], + "type": "text", + "content": " for the BASE and LARGE models respectively. The number of frames in each batch is limited to 2,400 and 1,600 frames for the BASE and LARGE models, respectively. Following [25, 28], a pre-trained transformer-based language model is used in the VSR decoding stage. The training configuration is the same as that used in the previous work [25]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "content": "Speech-Driven Lip Animation. For each training video clip, we sample a sequence of 75 frames that corresponds to 3 seconds. The first frame of the sampled video sequence is used to drive the lip animation. We train the lip animation model for 70 epochs on 32 A100-GPUs with a batch size of 3 per GPU. The details of the module structure in the lip animation model are described in the supplementary material. Adam [21] optimizer is used to train the lip animation model with the learning rate " + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "content": " for the generator and the frame discriminator. Sequence discriminator uses a smaller learning rate of " + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "content": ". The weights for each loss term are " + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{disc}^{img} = 1" + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{disc}^{seq} = 0.2" + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{rec} = 300" + }, + { + "bbox": [ + 304, + 521, + 545, + 666 + ], + "type": "text", + "content": " in Eq. (5)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "We consider five lip animation model configurations. The lip animation model without VSR perceptual loss is referred to LAM-Baseline. The lip animation model with the BASE VSR model trained on LRS3 is referred to as LAM-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18810" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 220 + ], + "type": "table", + "html": "
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
Afouras et al. [3]CNN✓595‡334-59.8
Ren et al. [37]TransformerX818‡--59.0
Afouras et al. [1]Transformer✓1,519†‡--58.9
Xu et al. [52]RNNX595‡--57.8
Shillingford et al. [44]RNN✓3,886†--55.1
Ma et al. [26]TransformerX4331,759-49.6*
Ma et al. [27]Conformer✓438--46.9
AV-HuBERT-BASE [43]TransformerX301,759-46.1
SynthVSRConformer-BASEX30--104.0
X--3,652100.3
X30-3,65244.7
✓30-3,65243.3
", + "image_path": "aa9303fe78c811b39c57d990b61559233b1f6b2ffc5a2540916510c56dc37e02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 546, + 262 + ], + "type": "text", + "content": "Table 1. Experimental results of low-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. †Includes non-publicly available data. ‡Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours. *Result taken from [43]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": "LRS3-VSR-VL, with the " + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{visual}} = 250" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{logits}} = 10" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": " in Eq. (3). Two variants LAM-LRS3-VSR-L and LAM-LRS3-VSR-V are further designed with the " + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{visual}} = 0" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{logits}} = 0" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": ", respectively. These two model configurations are used in ablation studies. Last, the lip animation model with the BASE VSR model trained on LRS3 and 2,630 hours of pseudo-labeled AVSpeech and VoxCeleb2 is referred to as LAM-LRS3-AVoX-VSR, with the " + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{visual}} = 500" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{logits}} = 10" + }, + { + "bbox": [ + 46, + 272, + 289, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 239, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 239, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 239, + 406 + ], + "type": "text", + "content": "4.4. Low-Resource Labeled Data Setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 556 + ], + "type": "text", + "content": "The LAM-Baseline model is used to generate 3,652 hours of synthetic data. We use 30 hours of LRS3 (training-validation) to evaluate the VSR performance when labeled data is scarce. We conduct experiments on the BASE VSR model. The results are reported in Tab. 1. As supervised VSR training from scratch with long sequences often poses optimization problems [1, 11, 28], we first use 30 hours of LRS3 and the 944 hours of Librispeech synthetic data to pre-train a VSR model with the same architecture as [28], then we take the weights of the pre-trained visual front-end to train the BASE VSR model. The pre-training details are described in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Training with synthetic data leads to dramatic performance improvement when only 30 hours of labeled data is available. Using 3,652 hours of synthetic data and 30 hours of LRS3 labeled data for training, the BASE model achieves the WER " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " (44.7% w/o. language model). Training from 30 hours of LRS3 results in a poor WER of " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "104\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ". Our proposed method significantly outperforms the former approaches using hundreds or thousands of data for supervised training [1,27,37,44,52] and self-supervised learning [3, 26]. Furthermore, our method performs better than the AV-HuBERT-BASE [43] model (46.1%) that uses 30 hours of LRS3 and 1,759 hours of unlabeled audio-visual data for self-supervised learning. Notably, we show" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 272, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 333 + ], + "type": "text", + "content": "the first attempt that trains a VSR model with considerable WER performance using only 30 hours of real video data. In addition, we observe that training only from synthetic data performs poorly (100.3%), which may be caused by the domain mismatch between the real and synthetic data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 342, + 457, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 457, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 457, + 355 + ], + "type": "text", + "content": "4.5. LRS3 Labeled Data Setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 433 + ], + "type": "text", + "content": "In this setting, we report the results when using the full 438 hours of LRS3. Experiments are conducted on the BASE VSR model. To avoid the optimization problem, we initialize the weights of the visual front-end from a publicly available Conformer-based model [28] pre-trained on LRS3 using curriculum learning. The results are shown in Tab. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": "Our BASE model achieves WER " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": " when using 438 hours of LRS3 data for training. We generate 3,652 hours of synthetic data using the LAM-LRS3-VSR-VL model. Using 3,652 hours of synthetic data and 438 hours of LRS3 labeled data, the BASE VSR model achieves the WER " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "27.9\\%" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "(28.4\\% \\mathrm{w / o}" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": " language model, corresponding to a WER reduction of " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": ". Our method outperforms three recent approaches using 31,000 " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "(33.6\\%)" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": ", 1,459 " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "(31.5\\%)" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": ", and 2,679 " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "(30.7\\%)" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": " hours of labeled data, respectively. When compared with the state-of-the-art self-supervised method AV-HuBERT [43] that uses additional 1,759 hours of unlabeled audio-visual data, our method outperforms the AV-HuBERT-BASE model " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "(34.8\\%)" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": " by a large margin. Our method slightly performs better than the AV-HuBERT-LARGE model " + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "inline_equation", + "content": "(28.6\\%)" + }, + { + "bbox": [ + 304, + 434, + 546, + 661 + ], + "type": "text", + "content": ", but with fewer model parameters (our BASE model 250M vs AV-HuBERT-LARGE 390M). Note that we compare with the AV-HuBERT results without self-training as we do not use the pseudo-labeled 933 hours of AVSpeech subset for VSR training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 670, + 500, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 500, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 500, + 684 + ], + "type": "text", + "content": "4.6. High-Resource Labeled Data Setting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We further evaluate the scalability of SynthVSR: when using machine-transcribed AVSpeech [14] and VoxCeleb2" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18811" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 69, + 547, + 248 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 248 + ], + "type": "table", + "html": "
MethodBackboneLMLabeled data (hrs)Unlabeled data (hrs)Synthetic data (hrs)WER (%)
AV-HuBERT-BASE [43]TransformerX4331,759-34.8
Makino et al. [30]TransformerX31,000†--33.6
Ma et al. [28]Conformer✓1,459‡--31.5
Prajwal et al. [35]Transformer✓2,676†--30.7
AV-HuBERT-LARGE [43]TransformerX4331,759-28.6
AV-HuBERT-LARGE w. Self-Training [43]TransformerX4331,759-26.9
Auto-AVSR [25]Conformer✓3,448‡--19.1
Serdyuk et al. [42]TransformerX90,000†--25.9
Serdyuk et al. [41]TransformerX90,000†--17.0
SynthVSRConformer-BASEX438--36.7
X438-3,65228.4
✓438-3,65227.9
X3,068--21.2
X3,068-3,65219.4
✓3,068-3,65218.7
SynthVSRConformer-LARGEX3,068-3,65218.2
✓3,068-3,65216.9
", + "image_path": "d6ba7342a6712704b02a71e61c6bf671f3dc3a31523440d47a6e6790c75a5082.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 361 + ], + "type": "text", + "content": "[10] as additional training data. We generate 3,652 hours of synthetic data with LAM-LRS3-AVoX-VSR and conduct experiments on BASE and LARGE models. The visual front-end is initialized from a pre-trained model which is the same as that used in the LRS3 labeled data setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "text", + "content": "We first train the BASE model with 438 hours of labeled LRS3 and 2,630 hours of pseudo-labeled data, resulting in a strong VSR baseline with the WER " + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "inline_equation", + "content": "21.2\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "text", + "content": ". By using additional 3,652 hours of synthetic data, the WER of the BASE model improves to " + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "text", + "content": " (19.4% w./o. language model), which outperforms [25] that uses additional labeled dataset LRS2 (223 hours) and LRW (157 hours) for training. Although the VSR model has seen a large amount of labeled data, and the speech-driven lip animation model is trained from part of the VSR training data, synthetic data can still lead to considerable performance gains. Furthermore, increasing the model size from BASE to LARGE results in better VSR performance with the WER of " + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 541 + ], + "type": "text", + "content": " (18.2% w./o. language model), which is the current state-of-the-art performance on LRS3, with publicly available data only." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 550, + 222, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 222, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 222, + 563 + ], + "type": "text", + "content": "4.7. Why SynthVSR Improves VSR?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "content": "We conduct extensive ablation studies to understand the impact of VSR perceptual loss, the diversity of lip sources, and the scale of ASR data for SynthVSR. In addition, we assess the domain mismatch between the real and synthetic data using VSR models. We perform experiments on the LRS3 labeled data setting with the Conformer-BASE VSR model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Effect of VSR Perceptual Loss. We experiment with four lip animation model variants: LAM-Baseline, LAM-LRS3-VSR-V, LAM-LRS3-VSR-L, LAM-LRS3-VSR-VL to generate 944 hours of synthetic data from Librispeech, respectively. The WER of the BASE model trained with only 438" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "content": "hours of LRS3 data is " + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "content": " (see in Tab. 2), the results are shown in Tab. 3. Using 944 hours of synthetic data generated from the LAM-Baseline model, the WER improves to " + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "inline_equation", + "content": "32.2\\%" + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "content": ". The individual visual VSR loss and linguistic VSR loss improve the WER to " + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "inline_equation", + "content": "31.7\\%" + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "inline_equation", + "content": "31.4\\%" + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "content": ", respectively. The linguistic VSR loss performs slightly better than the visual one. Further, the combination of visual and linguistic VSR loss further improves the WER to " + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "inline_equation", + "content": "30.8\\%" + }, + { + "bbox": [ + 304, + 300, + 547, + 434 + ], + "type": "text", + "content": ", which indicates that the spatio-temporal visual and high-level semantic knowledge from a pre-trained VSR model could offer the lip animation model more accurate lip movements." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 355, + 442, + 501, + 502 + ], + "blocks": [ + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "type": "text", + "content": "Table 2. Experimental results of LRS3 & high-resource labeled data setting on LRS3 (test). LM denotes whether or not a language model is used in the decoding. " + }, + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "type": "text", + "content": " Includes non-publicly available data. " + }, + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "type": "inline_equation", + "content": "{}^{§}" + }, + { + "bbox": [ + 46, + 256, + 546, + 289 + ], + "type": "text", + "content": " Includes datasets that are only permitted for the purpose of academic research. hrs is an abbreviation for hours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 355, + 442, + 501, + 502 + ], + "lines": [ + { + "bbox": [ + 355, + 442, + 501, + 502 + ], + "spans": [ + { + "bbox": [ + 355, + 442, + 501, + 502 + ], + "type": "table", + "html": "
Lip animation modelWER (%)
LAM-Baseline32.2
LAM-LRS3-VSR-V31.7
LAM-LRS3-VSR-L31.4
LAM-LRS3-VSR-VL30.8
", + "image_path": "3476941763988f754577569515e39142be0ac2c1448de6d1f54f6e0f60f656ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 509, + 541, + 521 + ], + "lines": [ + { + "bbox": [ + 310, + 509, + 541, + 521 + ], + "spans": [ + { + "bbox": [ + 310, + 509, + 541, + 521 + ], + "type": "text", + "content": "Table 3. Ablation studies on the proposed VSR perceptual loss." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "type": "text", + "content": "Effect of Increasing Lip Sources Diversity. To analyze the impact of the diversity of lip sources we generate 944 hours of synthetic data generated from Librispeech using two target face sources: CelebA and LRS3 (pre-training). Images from LRS3 (pre-training) do not introduce any new lip images for VSR training. The results are reported in Tab. 4. Using the synthetic data with LRS3 lip images, the BASE model achieves the WER " + }, + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "type": "inline_equation", + "content": "32.1\\%" + }, + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "type": "text", + "content": " which is " + }, + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 533, + 545, + 653 + ], + "type": "text", + "content": " worse than that with CelebA lip images, which means the external lip images from CelebA offer better WER improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Effect of the Scale of Speech Corpus. Using synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice with LAM-LRS3-VSR-VL model synthetic data, the LRS3 (test) WER improves from " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "30.9\\%" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "30.8\\%" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "30.1\\%" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", respectively (see in Tab. 5). When combining" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18812" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 268, + 110 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 268, + 110 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 268, + 110 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 268, + 110 + ], + "type": "table", + "html": "
Lip sourceUnique imagesWER (%)
LRS3-pretrain5,09032.1
CelebA202,59930.8
", + "image_path": "a61f5cd9242931ed45566d0bf63cfdb9f28852a45e61ffab143baf844f4cc28a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 160, + 286, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 160, + 286, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 286, + 244 + ], + "type": "text", + "content": "all synthetic data together (3,652 hours) for training, the WER improves to " + }, + { + "bbox": [ + 46, + 160, + 286, + 244 + ], + "type": "inline_equation", + "content": "28.4\\%" + }, + { + "bbox": [ + 46, + 160, + 286, + 244 + ], + "type": "text", + "content": ". We observe that richer speech corpora can offer larger performance improvements. Furthermore, we find that the improvement from scaling speech corpora is more pronounced than when adding VSR perceptual loss and lip sources diversity, suggesting that our proposed method benefits greatly from large speech datasets." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 253, + 280, + 322 + ], + "blocks": [ + { + "bbox": [ + 47, + 118, + 286, + 140 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 286, + 140 + ], + "type": "text", + "content": "Table 4. Ablation studies on the impact of lip sources diversity. LRS3-pretrain indicates using the LRS3 (pre-training) images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 253, + 280, + 322 + ], + "lines": [ + { + "bbox": [ + 57, + 253, + 280, + 322 + ], + "spans": [ + { + "bbox": [ + 57, + 253, + 280, + 322 + ], + "type": "table", + "html": "
Training dataHoursWER (%)
LRS343836.7
LRS3 + TED-Synth90330.9
LRS3 + LBS-Synth1,38230.8
LRS3 + CV-Synth2,68130.1
LRS3 + [TED + LBS + CV]-Synth4,09028.4
", + "image_path": "63fb2aa6d18145ba3e3aa94f9ed41448d81b5bcfb062f3e7b898451168d25c5f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": "Assessment of the Domain Mismatch Using VSR. We use 944 hours of Librispeech synthetic data generated from LAM-Baseline and LAM-LRS3-VSR-VL for VSR training. We conduct an assessment on the LRS3 real and synthetic test sets, the synthetic test set is obtained by passing the speech and the first frame in the real test set to the lip animation model. The results are illustrated in Fig. 3. By introducing the VSR perceptual loss, the WER on the synthetic test set is reduced from " + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "inline_equation", + "content": "106.8\\%" + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "inline_equation", + "content": "49.7\\%" + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": ", which indicates the VSR loss minimizes the domain gap between real and synthetic data. After training on synthetic data, the WER reduces for both real and synthetic data. Specifically, the WER is improved substantially from " + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "inline_equation", + "content": "106.8\\%" + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "inline_equation", + "content": "36.0\\%" + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": " and from " + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "inline_equation", + "content": "49.7\\%" + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 46, + 376, + 286, + 581 + ], + "type": "text", + "content": " on synthetic test sets generated using LAM-Baseline and LAM-LRS3-VSR-VL, respectively. We observe that domain mismatch can be further reduced by combining real and synthetic data for VSR training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 587, + 254, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 254, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 254, + 601 + ], + "type": "text", + "content": "4.8. SynthVSR with TTS-Generated Speech" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "In the real world, text corpora (e.g., Wikipedia) are more accessible than transcribed speech corpora. To explore the potential of using TTS-generated speech for SynthVSR, we use the off-the-shelf TTS model FastSpeech 2 [39] to generate 944 hours of synthetic speech from Librispeech transcriptions, which is further used to generate synthetic video clips using LAM-LRS3-VSR-VL model. Using these 944 hours of synthetic data to train the BASE model under the LRS3 labeled data setting, the WER is improved" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 64, + 518, + 178 + ], + "blocks": [ + { + "bbox": [ + 310, + 64, + 518, + 178 + ], + "lines": [ + { + "bbox": [ + 310, + 64, + 518, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 64, + 518, + 178 + ], + "type": "image", + "image_path": "38b8a320d1f90e152d48dd0fd74848a89bf715be9b4abfdbd87789a0310d57c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 191, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 246 + ], + "type": "text", + "content": "Figure 3. Assessment of the domain mismatch between real and synthetic video data using VSR models trained with only real data (VSR-Real, black) and real data together with synthetic data (VSR-Real-Synth, red), respectively. Synth-Test and Real-Test refer to the synthetic and real LRS3 (test) sets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "36.7\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "32.9\\%" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": ", as shown in Tab. 6. Although TTS-generated speech is not as natural as the original LibriSpeech data, the improvement of using synthetic data is still decent " + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "inline_equation", + "content": "(3.8\\%)" + }, + { + "bbox": [ + 304, + 258, + 545, + 330 + ], + "type": "text", + "content": ", indicating the further potential of SynthVSR in real-world applications where labeled video data is sparse, such as the healthcare industry." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 328, + 340, + 526, + 391 + ], + "blocks": [ + { + "bbox": [ + 46, + 330, + 286, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 286, + 374 + ], + "type": "text", + "content": "Table 5. Ablation studies on different speech corpus. TED-Synth, LBS-Synth, and CV-Synth indicates the synthetic data generated from TED-LIUM 3, Librispeech, and Common Voice, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 340, + 526, + 391 + ], + "lines": [ + { + "bbox": [ + 328, + 340, + 526, + 391 + ], + "spans": [ + { + "bbox": [ + 328, + 340, + 526, + 391 + ], + "type": "table", + "html": "
Training dataHoursWER (%)
LRS343836.7
LRS3 + LBS-Synth1,38230.8
LRS3 + TTS-LBS-Synth1,38232.9
", + "image_path": "d5a5b2164de4b7444917cbfc3d35bdb10934126f1bcb9a59f5a51a32be875583.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 421 + ], + "type": "text", + "content": "Table 6. Experiments results of using the synthetic speech generated synthetic video (TTS-LBS-Synth) for VSR training." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 445, + 474, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 474, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 474, + 460 + ], + "type": "text", + "content": "5. Limitations & Societal Impact" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 586 + ], + "type": "text", + "content": "The excellent performance improvement of SynthVSR comes at the cost of high computational demands during training, as we use additional large-scale synthetic video data. Also, although SynthVSR has achieved state-of-the-art VSR performance, the LRS3 test set is relatively limited (0.9 hours), which is from TED talks, while real-world videos are more challenging (e.g., egocentric videos). SynthVSR has many positive real-world applications, such as helping the hearing-impaired or people with aphonia with everyday communication." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 597, + 378, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 378, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 378, + 609 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We have presented a semi-supervised method for VSR enhanced with synthetic lip movements. The speech-driven lip animation model is proposed to generate synthetic video data from labeled speech datasets and face images for scaling up VSR. Our method achieves state-of-the-art results on LRS3, outperforming prior work trained on more labeled or unlabeled real video data. Our work fosters future research on generating and exploiting synthetic visual data for VSR." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18813" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 139 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 172 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. LRS3-TED: A large-scale dataset for visual speech recognition. arXiv:1809.00496, 2018. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 228 + ], + "type": "text", + "content": "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. ASR is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2143-2147, 2020. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 293 + ], + "type": "text", + "content": "[4] Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Conference on Language Resources and Evaluation, 2020. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 293, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 288, + 326 + ], + "type": "text", + "content": "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. LipNet: Sentence-level lipreading. arXiv:1611.01599, 2016. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 381 + ], + "type": "text", + "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 382, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 288, + 426 + ], + "type": "text", + "content": "[7] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 470 + ], + "type": "text", + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision, pages 520-535, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 471, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 288, + 502 + ], + "type": "text", + "content": "[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? In British Machine Vision Conference, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 3, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 570 + ], + "type": "text", + "content": "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, pages 87-103. Springer, 2016. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "text", + "content": "[12] Soo-Whan Chung, Hong Goo Kang, and Joon Son Chung. Seeing voices and hearing voices: Learning discriminative embeddings using cross-modal self-supervision. In INTER-SPEECH, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3D human pose estimation: Motion to the rescue. In Advances in Neural Information Processing Systems, volume 32, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv:1804.03619, 2018. 3, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 703 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Amin Fazel, Wei Yang, Yulan Liu, Roberto Barra-Chicote, Yixiong Meng, Roland Maas, and Jasha Droppo. SynthASR: Unlocking synthetic data for speech recognition. arXiv:2106.07803, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[16] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv:2007.04954, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 228 + ], + "type": "text", + "content": "[17] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine learning, pages 369-376, 2006. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "text", + "content": "[18] Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. In INTER-SPEECH, 2020. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 327 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 382 + ], + "type": "text", + "content": "[20] François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia Tomashenko, and Yannick Esteve. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In International Conference on Speech and Computer, pages 198-208, 2018. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 403 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 449 + ], + "type": "text", + "content": "[22] Taku Kudo. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[23] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision, December 2015. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 492, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 514 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 515, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 570 + ], + "type": "text", + "content": "[25] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-AVSR: Audio-visual speech recognition with automatic labels. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2023. 2, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "text", + "content": "[26] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. LiRA: Learning visual speech representations from audio through self-supervision. In IN-TERSPEECH, 2021. 2, 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "text", + "content": "[27] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with Conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 7613-7617, 2021. 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 691 + ], + "type": "text", + "content": "[28] Pingchuan Ma, Stavros Petridis, and Maja Pantic. Visual speech recognition for multiple languages in the wild. Nature Machine Intelligence, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 703 + ], + "type": "text", + "content": "[29] Shuang Ma, Zhaoyang Zeng, Daniel McDuff, and Yale Song." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18814" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "text", + "content": "Contrastive learning of global and local audio-visual representations. arXiv:2104.05418, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE Automatic Speech Recognition and Understanding Workshop, pages 905–912, 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 182 + ], + "type": "text", + "content": "[31] Sepand Mavandadi, Tara N Sainath, Kevin Hu, and Zelin Wu. A deliberation-based joint acoustic and text decoder. In INTERSPEECH, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 239 + ], + "type": "text", + "content": "[32] Xichen Pan, Peiyu Chen, Yichen Gong, Helong Zhou, Xinbing Wang, and Zhouhan Lin. Leveraging uni-modal self-supervised learning for multimodal audio-visual speech recognition. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[33] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 5206-5210, 2015. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "text", + "content": "[34] Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning deep object detectors from 3D models. In Proceedings of the IEEE International Conference on Computer Vision, pages 1278-1286, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "text", + "content": "[35] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 447 + ], + "type": "text", + "content": "[36] Aayush Prakash, Shaad Boochoon, Mark Brophy, David Acuna, Eric Cameracci, Gavriel State, Omer Shapira, and Stan Birchfield. Structured domain randomization: Bridging the reality gap by context-aware synthetic data. In 2019 International Conference on Robotics and Automation, pages 7249-7255, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[37] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[39] Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech 2: Fast and high-quality end-to-end text to speech. arXiv:2006.04558, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "text", + "content": "[40] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3234-3243, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[41] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition. In INTERSPEECH, 2022. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[42] Dmitriy Serdyuk, Olivier Siohan, and Otavio de Pinho Forin Braga. Audio-visual speech recognition is worth " + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 8" + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": " voxels. In IEEE Automatic Speech Recognition and Under" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 448, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 448, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 448, + 83 + ], + "type": "text", + "content": "standing Workshop, 2021. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 85, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 85, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 85, + 545, + 128 + ], + "type": "text", + "content": "[43] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In International Conference on Learning Representations, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "text", + "content": "[44] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. In INTERSPEECH, 2019. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 216 + ], + "type": "text", + "content": "[45] Themos Stafylakis and Georgios Tzimiropoulos. Combining residual networks with LSTMs for lipreading. arXiv:1703.04105, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "type": "text", + "content": "[46] Jörgen Valk and Tanel Alumäe. Voxlingua107: A dataset for spoken language recognition. In 2021 IEEE Spoken Language Technology Workshop, pages 652-658, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 251, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 294 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[48] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. StyleGAN2 distillation for feed-forward image manipulation. In European Conference on Computer Vision, pages 170-186. Springer, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "type": "text", + "content": "[49] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven facial animation with temporal GANs. In British Machine Vision Conference, 2018. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[50] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with GANs. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 502 + ], + "type": "text", + "content": "[51] Zhonghao Wang, Mo Yu, Yunchao Wei, Rogerio Feris, Jinjun Xiong, Wen-mei Hwu, Thomas S Huang, and Honghui Shi. Differential treatment for stuff and things: A simple unsupervised domain adaptation method for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12635-12644, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[52] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14433-14442, 2020. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[53] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "text", + "content": "[54] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9299–9306, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[55] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18815" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_content_list.json b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..08aba7d2f20ca738cc70d8df81b6287954199669 --- /dev/null +++ b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_content_list.json @@ -0,0 +1,1356 @@ +[ + { + "type": "text", + "text": "Synthesizing Photorealistic Virtual Humans Through Cross-modal Disentanglement", + "text_level": 1, + "bbox": [ + 148, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siddarth Ravichandran, Ondrej Texler, Dimitar Dinev, Hyun Jae Kang NEON, Samsung Research America", + "bbox": [ + 191, + 203, + 776, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{siddarth.r, o.texler, dimitar.d, hyunjae.k}@samsung.com", + "bbox": [ + 233, + 241, + 730, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9455b5d041dda44847ef46f66803753111631caaaf7347704e5ea1bae1acd808.jpg", + "image_caption": [ + "Figure 1. Three sequences of talking faces generated using the proposed framework. Our method produces high-texture quality, sharp images, and correct lips shapes that are inline with the spoken audio. See the zoom-in patches, our method faithfully reproduces fine textural and content details such as teeth." + ], + "image_footnote": [], + "bbox": [ + 81, + 292, + 890, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 497, + 313, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the last few decades, many aspects of human life have been enhanced with virtual domains, from the advent of digital assistants such as Amazon's Alexa and Apple's Siri to the latest metaverse efforts of the rebranded Meta. These trends underscore the importance of generating photorealistic visual depictions of humans. This has led to the rapid growth of so-called deepfake and talking-head generation methods in recent years. Despite their impressive results and popularity, they usually lack certain qualitative aspects such as texture quality, lips synchronization, or resolution, and practical aspects such as the ability to run in real-time. To allow for virtual human avatars to be used in practical scenarios, we propose an end-to-end framework for synthesizing high-quality virtual human faces capable of speaking with accurate lip motion with a special emphasis on performance. We introduce a novel network utilizing visemes as an intermediate audio representation and a novel data augmentation strategy employing a hierarchical image synthesis approach that allows disentanglement of the different modalities used to control the global head motion. Our method runs in real-time, and is able to deliver superior results compared to the current state-of-the-art.", + "bbox": [ + 73, + 550, + 473, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 497, + 630, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the metaverse gaining traction in the media and the rapid virtualization of many aspects of life, virtual humans are becoming increasingly popular. Propped up by technological advances in hardware, GPU computing, and deep learning, digitization of humans has become a very active research topic.", + "bbox": [ + 496, + 523, + 890, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One challenging problem is creating virtual avatars that attempt to faithfully recreate a specific human indistinguishable from their real counterparts. This entails a perfect visual recreation, with high quality textures that recreate the minutia that make a human visually realistic, e.g., hair that appears to be composed of individual strands, clothes that appear made out of fabric, and pores or vellus hair on the skin. Recent advances in neural rendering have achieved considerable success in creating very convincing images that are able to visually recreate such fine details. Additionally, many use-cases involving virtual humans require them to appear on large screens, e.g., human-size displays in airport halls, hotels, and billboards. For such use-cases, it is vital for the virtual humans to be rendered in a relatively high resolution.", + "bbox": [ + 496, + 613, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, rendering quality alone is not sufficient to recreate a convincing virtual human that can be interacted with. Just as important as the visual fidelity is the fidelity of the motion. Specifically, for speech-driven avatars, the", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4585", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "way the avatar talks, opens its mouth, and moves its lips is crucial. The human perception system is very sensitive to even the slightest mistake when it comes to the face, specifically the appearance of mouth and lips during speech. In order for a virtual avatar to be believable, its mouth and lip motion must match the speech a human listener hears.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a novel training regime that utilizes 1-dimensional audio features such as visemes [9] or wav2vec 2.0 [1]. These are higher-level features compared to traditional features such as MFCC [23] which require more complex encoders. This enables us to efficiently synthesize talking heads with high-quality rendering and lip motion. However, this alone is insufficient for creating an avatar suitable for the aforementioned use-cases. In order to have more control over the avatar, it needs to be conditioned on multiple constraints in addition to speech, such as head pose or body shape. With speech represented as a 1-dimensional vector of visemes, head pose represented as rotation angles, and body shape represented as a 2-dimensional outline, we quickly run into a problem with different modalities among our data representation.", + "bbox": [ + 75, + 184, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training deep neural networks on multimodal data is a difficult problem, especially when the different modalities are innately correlated. The network then tends to overfit on the modality that is easier to learn from, and this kind of overfitting leads to very poor performance at testing time. While there exist many techniques to mitigate traditional overfitting such as reducing number of parameters, regularization, etc., they are less effective when the overfitting is caused by multiple modalities. To break the correlations between these modalities, we introduce a novel data augmentation strategy: for every given lip shape, we use a generative oracle network to synthesize photorealistic images of the same lip shapes in a variety of head poses. This requires a generative network capable of producing images that are as high-resolution as our training data to preserve sharpness in important areas such as teeth.", + "bbox": [ + 75, + 412, + 470, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even with modern hardware and the latest research, generative neural networks still struggle with generating high-resolution images, and the resolution usually comes at the cost of quality. When increasing the resolution and the complexity of what the network is asked to generate, the generative network often loses the semantic quality or other qualitative aspects of the result images. Increasing the network's capacity alone by adding more learnable parameters does not necessarily help and can introduce noise. To alleviate this, we propose a hierarchical approach to generating a high-quality synthetic image of a talking face that preserves the quality in the mouth region via high-resolution supervision.", + "bbox": [ + 75, + 656, + 468, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to reproduce a believable interaction with a virtual avatar, inference speed is paramount. Our paper presents an efficient framework for creating high-quality", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "virtual artificial humans in real-time where each of the aforementioned concerns is addressed. The contribution of our works is:", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A data augmentation method to disentangle audio and visual modalities so that the whole framework can be trained end-to-end.", + "- A hierarchical outpainting approach which allows for generation of high-resolution synthetic data.", + "- An end-to-end framework that utilizes 2-encoder-2-decoder neural network architecture and leverages synthetic data." + ], + "bbox": [ + 517, + 151, + 890, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 315, + 640, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Talking faces. In the last several decades, many methods to generate videos of talking faces have been developed, and most of the recent ones heavily leverage generative adversarial networks-GANs [12]. Having the photorealism as a main objective, various methods, colloquially known as \"deepfakes\" [2,7,20,32,34,36,45], have been proposed and successfully used to synthesize images of human faces. Talking face animation can be categorized into face re-enactment approaches [15, 24, 25, 40, 42, 47] and speech driven approaches [3, 26, 27, 31, 46, 49, 51]. Speech driven models have used a variety of intermediate representations such as 2D keypoints [3,32], meshes [6,27,33], which themselves can be derived from various audio features such as MFCC, spectral coefficients, or DeepSpeech features [13]. Some methods directly produce talking faces from audio features without additional geometric representations [46] and some [26, 49] use an additional identity image or video as guiding signal. A common issue these methods face is the reduced high-frequency motion, such as lip motion during speech. Unrealistic lip movement is easy to notice, and disrupts the overall experience. This is partially addressed in the work of Prajwal et al. [26], they utilize SyncNet [5] to supervise the audio and lips synchronization, which results in a dynamic mouth movement. However, they do not consider the photorealistic appearance. As the mandate of our work is to create virtual humans that are indistinguishable from real people, we need to secure both realistic lip motion and photorealistic textures.", + "bbox": [ + 496, + 340, + 890, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hierarchical image generation. Synthesizing high-resolution images while retaining various fine details and overall sharpness is a difficult task. Researchers have been approaching this problem from various angles, and one of widely used solutions is to tackle this challenge in a hierarchical fashion. InsetGAN [11] propose to combine multiple pre-trained GANs, where one network generates the coarse body image, and other networks refine certain regions. The work of Li et al. [22] proposes to have several", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4586", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30b393907b45f629de66a669f799a04aebcb3d9f164c0a4de3dfa2ccfdb73f71.jpg", + "image_caption": [ + "Figure 2. Overview of our multi-modal neural rendering pipeline. We generate $k$ audio features of the provided audio, and sample a sequence of keypoint drawings and contours which are then mapped to a shared latent space using Audio Encoder (a) and the Landmark Encoder (b) respectively; which is then passed through the mouth (c) and face (d) decoders to produce the mouth and face images respectively. The mouth feature maps are pasted into the head feature map as denoted by the red arrow. During inference, only the output of the head decoder is used." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 888, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "class-specific generators that are used to refine certain semantic regions of the previously synthesized image. TileGAN [10] presents a way to combine small-resolution synthesized images in order to produce a single high-resolution texture. Xu and Xu [43] hierarchically compose the image from smaller patches, and Iizuka et al. [17] solve inpainting task using two discriminators, one for the whole image, and the other for a small area centered at the completed region. Although those approaches are not directly applicable in our talking virtual human scenario, our pipeline also leverages the hierarchical composition.", + "bbox": [ + 75, + 462, + 472, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal training. To puppet a virtual human, multiple different inputs are usually needed, e.g., speech to control the lips and head-pose to specify head position and rotation. Those inputs-modalities are usually strongly correlated, which translates in the need of a multi-modal training [4, 37, 48, 50]. NeuralVoicePuppetry [33] trains a generalized expression network and an identity targeted neural renderer method decoupling the modalities using an intermediate mesh representation. MakeItTalk [51] propose to disentangle the modalities by learning displacements from a reference facial landmark using a separate module and perform rendering from the intermediate 2-dimensional landmark representations. Similarly, L. Chen et al. [3] use a learned high-level hierarchical structure to bridge audio signal with the pixel image. MeshTalk [27] deals with the modality disentanglement in a lower dimensional space by employing the cross-modality loss. Many of these methods require an intermediate representation for the face in the", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "form of drawings or mesh, hindering their ability to produce diverse lip shapes. We address these limitations in our work by using a mix of 1-dimensional and 2-dimensional representations, requiring modality disentanglement in order to synthesize speech-driven talking faces.", + "bbox": [ + 496, + 462, + 893, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 498, + 551, + 648, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a framework to generate photorealistic virtual human avatars that can be streamed in real-time to enable interactive applications. We train our system on a single identity and during the inference we drive it using audio features and partial face keypoints, as shown in Fig. 2. Due to bias in the way humans speak, there is a strong correlation between head motion and speech which causes the network to model undesirable relations. To combat that, we propose a novel data augmentation strategy employing keypoint mashing and a high-resolution generative oracle network to achieve disentanglement of lip motion from upper face motion. This allows synthesizing videos with arbitrary upper face keypoints and audio.", + "bbox": [ + 496, + 579, + 893, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Input Features", + "text_level": 1, + "bbox": [ + 500, + 786, + 651, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio features. One of the crucial requirements for our method is the ability to faithfully reproduce motion and shapes of mouth and lips. Thus, it is vital to have an intermediate audio representation that is capable of capturing those geometric characteristics. Visemes [9] are lip shape categories that describe the basic lip shapes used to make", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4587", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sounds common in human speech. They can be thought of as the geometric analog to phonemes, and have been utilized to create a geometric basis for head models used in commercial 3D graphics products to parameterize and animate speech [52]. Despite the aforementioned theoretical advantages, visemes have certain practical difficulties, often requiring commercial tools such as JALI [8]. Our framework can also handle different frame-level audio feature representations such as wav2vec 2.0 [1].", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Commonly used intermediate representations, such as keypoints [39] can be inaccurate, especially for difficult-to-detect regions such as lips. We propose to forego the use of intermediate geometric representation and instead render the image directly given visemes.", + "bbox": [ + 75, + 227, + 468, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Facial features. To render the rest of the head, landmarks and contours are strong guiding signals for the network to produce semantically correct images. We use these features of the non-mouth areas of the face as additional inputs for the network, enabling control of non-speech-related behavior such as head motion.", + "bbox": [ + 75, + 304, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-Modal Neural Rendering", + "text_level": 1, + "bbox": [ + 76, + 406, + 354, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start by processing the audio to obtain $k$ features per frame $\\mathbf{v} \\in \\mathbb{R}^{k \\times 1}$ . We take a window of 6 consecutive frames centered around the corresponding image frame. This results in a vector of size $\\mathbf{v} \\in \\mathbb{R}^{k \\times 6}$ , shown in Fig. 2a. For visemes, $k = 34$ .", + "bbox": [ + 75, + 430, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second modality is the upper face keypoints and the contours drawn on a canvas of size $512 \\times 512$ px, in Fig. 2b. Rather than passing the keypoints as vectors, we draw them as a 2D image which helps the model to learn spatial correspondences, resulting in better generalization to translations and motion in the keypoints.", + "bbox": [ + 75, + 507, + 468, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model architecture details. Our model consists of two encoders, one for each modality. The audio encoder consists of a 1D convolutional neural network followed by a 2-layer perceptron network. The resulting latent vector is reshaped to a 2D representation and concatenated with the latent feature map of the keypoint encoder which consist of 4-layer strided convolution blocks with a residual block [14] in between consecutive strided convolutions. Taking the inspiration from Pix2PixHD [39], we adopted a two image decoder scheme. One decoder is trained to generate a full face (Fig. 2d), and the second decoder (Fig. 2c) is trained to generate a fixed crop region of the mouth at a larger resolution. Both decoders produce different images from the same latent space. The model is trained end to end in a GAN setup, using two patch discriminators [18]: one for the head and the other for the mouth. We use the VGG [29] feature matching loss and smooth L1 loss alongside the GAN loss to train the generators. To obtain sharper mouth images, especially for the teeth, we resize and place the mouth region feature map into the head feature map at the appro", + "bbox": [ + 75, + 599, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c15ed914154b4d4131b5cacc2e2842bc76275dea903db4d674981999a259b7eb.jpg", + "image_caption": [ + "Figure 3. Training a full-head model at $512\\mathrm{px}$ resolution (a) produces lower details in the $192\\mathrm{px}$ mouth region (b) than training a mouth-only model on $512\\mathrm{px}$ , then resampled to $192\\mathrm{px}$ (c)." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 629, + 188 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb0aadf133720a3e9a8e245943d79dc83552932432e67c2682a236047f3e4ee2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 89, + 759, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9018ec539fde69748b1e492f1a626f9b5cf8e37f2ca5ca0127ac3e9d19b254e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 89, + 890, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "priate crop position in the penultimate layer of the head decoder, as shown by the red arrow in Fig. 2. We do this in the penultimate layer rather than the final image to avoid any blending artifacts that could arise.", + "bbox": [ + 496, + 270, + 890, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Disentanglement of Audio and Head-Pose", + "text_level": 1, + "bbox": [ + 500, + 344, + 854, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned in Section 3.2, we use two inputs to the network: a $k \\times 6$ vector of audio features, and a $512 \\times 512$ line drawing of the face. Since our training dataset only contains a single subject speaking one language, this leads to considerable entanglement between audio features and the 2D contour drawing, making the model learn unwanted relationships between the line drawing and lip motion. This manifests in, for example, lip motion being generated from just head motion in the contour drawing, even without providing any audio.", + "bbox": [ + 496, + 369, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To remedy this, we propose a novel data augmentation strategy for disentangling the audio features from the head pose. One way to enforce the correlation between audio features and lip shape is to show the network samples of the subject uttering the same phrases with different head poses. While this can be done during data capture, this significantly lengthens and complicates the capture process. To remedy this, we propose training an oracle network that converts 2D contour drawings into photoreal renders. Then, we propose a method for transferring the mouth shape from one frame to the head pose of another. Combining these two methods allows us to augment the capture dataset with synthetic images containing a greater variety of lip shapes and head poses, forcing the network to break the correlation between head pose and lip shape.", + "bbox": [ + 496, + 521, + 890, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Oracle Network. The oracle generates photorealistic images of the subject given the 2D line drawings from in Fig. 2. This can be accomplished by employing a network based on Pix2Pix [18,39]. While the images produced by this network appear photorealistic, Pix2Pix can struggle with blurriness, especially in regions where there is significant motion and occlusions, such as the teeth. This blurriness pollutes our training data, degrading the visual quality of our method. One observation is that if we train a higher-resolution network that focuses on the mouth then the teeth", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4588", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e23d522c7194e2d8bb23383367379a3abe07a1378cd7f252add95c9af107669.jpg", + "image_caption": [ + "Figure 4. Our oracle network works in a hierarchical fashion, starting with the network to generate mouth (a) in a high-resolution. The result is then fed into the subsequent network (b) to generate the rest of the head. The head image and the high resolution mouth image are used for supervision of our multi-modal renderer." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "will come out clearer, even if the result is then downsampled as shown in Fig. 3. Here, we trained two versions of the same network: one on the full-head $512 \\times 512$ px data Fig. 3a and the other on high-resolution $512 \\times 512$ px mouth crop, then downscaled to $192 \\times 192$ px Fig. 3c. If we extract the same $192 \\times 192$ region from the full head model (Fig. 3b), we can see that there is a degradation in quality in the teeth region.", + "bbox": [ + 75, + 338, + 468, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To produce high-resolution renderings, we propose to render the image in a hierarchical manner starting with the region where the details and quality is most important for the given task – the mouth region in our case. Our solution can be viewed as a chaining of neural network models where the next network is conditioned by the result of the previous network, shown in Fig. 4. First, we train a network that generates a high-quality mouth image( Fig. 4a). This image is downsampled and placed in a line drawing of the full head, similar to the process outlined in Section 3.2. Then, it is used as input to the full-head network( Fig. 4b), which learns to render the rest of the head. We refer to this hierarchical approach as outpaint since it resembles the task of outpainting.", + "bbox": [ + 75, + 459, + 470, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Keypoint mashing. The next missing piece for generating novel head pose and mouth shape combinations is the ability to combine the mouth keypoints from one frame with the head keypoints in another frame. However, the two frames will have different head poses, meaning we need to remove the global pose information before the mouth keypoints can be replaced.", + "bbox": [ + 75, + 670, + 468, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For any given frame, we extracted $n$ 2D facial keypoints $\\mathbf{K} \\in \\mathbb{R}^{n \\times 2}$ and a pose matrix $\\mathbf{P} \\in \\mathbb{R}^{4 \\times 4}$ based on a 3D canonical space for the keypoints. The position of the keypoints can be expressed as:", + "bbox": [ + 76, + 776, + 470, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} = p r o j (\\mathbf {P K ^ {f}}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 843, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{K}^{\\mathbf{f}}\\in \\mathbb{R}^{n\\times 4}$ are the un-posed keypoints with depth in homogeneous coordinates, and $proj()$ is a projection func", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d0b2eedaceec469be29edda1acbd6219f379d75da292da63605e18a62fb3a50.jpg", + "image_caption": [ + "Figure 5. Using keypoint mashing and the proposed oracle network, we are able to generate a synthetic image (c) by combining the mouth keypoints of (a) with the head keypoints of (b)." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion (e.g. perspective projection) that converts keypoints from 3D space to 2D screen space. Using Eq. 1, we can compute the un-posed keypoints:", + "bbox": [ + 498, + 342, + 890, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} ^ {\\mathbf {f}} = \\mathbf {P} ^ {- 1} p r o j ^ {- 1} (\\mathbf {K}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 398, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $proj^{-1}$ is the inverse projection function. For an orthographic projection, this would just be an inverse of the projection matrix, provided the keypoint data is 3D. However, many keypoints detectors only provide 2D information. In order to fill in the missing depth, we obtain depth values by solving the forward problem Eq. 1 for the canonical keypoints and concatenating the resulting depth value to the extracted 2D keypoints. This same process can be used to recover the perspective divide if $proj$ is a perspective transformation. While this is only an approximation for the true depth values, most of the depth variation in the face comes from the underlying facial bone structure and geometry captured in the canonical keypoints, making this a viable approximation.", + "bbox": [ + 496, + 429, + 890, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating synthetic data. With keypoint mashing and a generative oracle in place, we now define a method for data augmentation. We select two arbitrary frames $i$ and $j$ , with corresponding keypoints and poses $\\mathbf{K}_i, \\mathbf{P}_i$ (Fig. 5a, top) and $\\mathbf{K}_j, \\mathbf{P}_j$ (Fig. 5b, top), we compute the un-posed keypoints $\\mathbf{K}_i^f$ . Then, we apply $\\mathbf{P}_j$ resulting in a new keypoints set $\\mathbf{K}_{i \\rightarrow j}$ : the keypoints from frame $i$ in the head pose of frame $j$ . Since the two keypoints sets $\\mathbf{K}_{i \\rightarrow j}$ and $\\mathbf{K}_j$ are now in the same space, we can trivially replace the mouth keypoints: $\\mathbf{K}_j^{mouth} = \\mathbf{K}_{i \\rightarrow j}^{mouth}$ resulting in the mouth position from $i$ transferred over to frame $j$ , $\\mathbf{K}_j'$ . Now we have audio features for frame $i$ , but a new keypoint set $\\mathbf{K}_j'$ (Fig. 5c, top).", + "bbox": [ + 496, + 641, + 892, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using our oracle network, we can now generate synthetic images (Fig. 5c, bottom) from $\\mathbf{K}_j^\\prime$ to be used as the \"ground truth\" for this new head pose and mouth shape combination. This makes the network to see different head poses", + "bbox": [ + 498, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4589", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/733e20b67313ffd28c78fc285613011eba50df50103b93ad03605d26b66d9c1d.jpg", + "image_caption": [ + "(a) MakeItTalk", + "(b) Wav2lip", + "(c) NVP+Pix2Pix", + "(d) Ours", + "(e) Ground Truth", + "Figure 6. Comparison with state-of-the-art techniques. MakeItTalk [51] (a), Wav2Lip [26] (b), NVP [33] + Pix2pix [38] (c), our result (d), and ground truth image (e). The avatar says 'ee'. Our results have significantly better lip shapes as compared to (a, c), and better texture quality as compared to (a, b)." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 890, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for every audio feature frame, preventing it from learning an erroneous correlation between the head motion and lip shape.", + "bbox": [ + 75, + 529, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 587, + 161, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First we present comparisons with the current state-of-the-art techniques. Then, we elaborate on data collection and implementation details, discuss hyper-parameters, followed by further analysis and discussion. Our method is single identity, and each subject shown has their own model trained on their own dataset. For more results, please, see our supplementary pdf and video.", + "bbox": [ + 75, + 612, + 468, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparisons", + "text_level": 1, + "bbox": [ + 76, + 726, + 215, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. We compare our method with two recent multi-identity approaches, Fig. 6a-b: MakeItTalk [51] and Wav2Lip [26]. These methods are trained on a large corpus of data containing many identities, and extract the identity of the subject at inference time. While it is indeed practical that MakeItTalk only requires a single image to synthesize a new identity, the resulting visual and lip sync quality is markedly worse than our method. Wav2Lip, which infers identity from a source video, arguably produces better lip sync than our method,", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "but at the expense of visual quality: Wav2Lip only supports $96 \\times 96$ , of which only $48 \\times 96$ are used for the mouth, and attempting higher resolutions fails. Notice that both of these above methods produce black blobs in the mouth region without any teeth textures. We then compared our method to TalkingFace [30], which is also trained on a single identity using data which is similar to our approach making the comparison straightforward. Our method produces superior lip motion: Fig. 7 shows how the two methods compare when producing the word \"we\". Our method is able to correctly capture the lip transition from the more rounded \"w\" sound to the \"e\" sound. Another single-identity method for generating talking heads is outlined in Neural Voice Puppetry [33]. We implemented a variation using Pix2Pix [38] and trained it on our data; our method produces superior lip motion Fig. 6c-d. Overall, our method produces higher quality videos with sufficiently accurate lip motion than the competing methods. Please refer to the supplementary video for more detailed comparisons.", + "bbox": [ + 496, + 529, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons. In Table 1, we provide quantitative comparisons between our method and the above competing methods. To evaluate image quality, we compute the PSNR and SSIM [41] scores for 500 inferred images against their ground truth counterparts, and show", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4590", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0070b703b6372ff6b1b24cca126de070c8ff86aacfad619b6541cbc9fdde6133.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 88, + 243, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6759508db93345edd05fb68b192211ce33e24c37b13bdd5eea1fcad9f7a408ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 88, + 439, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e014f936663919ac45c9501b5fce67866b26bec00f2be6b5f5a61f35052204da.jpg", + "image_caption": [ + "Figure 7. TalkingFace [30] (top) and our model (bottom) were both asked to produce the word \"we\". Our model captures the transition from \"w\" (a) to \"e\" (b) more accurately." + ], + "image_footnote": [], + "bbox": [ + 80, + 160, + 272, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1acb79375dc082857b34467af717c68cab35b96b13995414b451cde77380c15.jpg", + "image_caption": [ + "Figure 8. The proposed augmentation is essential for the model to correctly respect different inputs. Without the augmentation (a), the virtual human erroneously opens her mouth even when the audio is silence; while (b) classical data augmentation methods might slightly mitigate this issue the model still produces wrong lip shapes; this is fixed by our augmentation technique (c)." + ], + "image_footnote": [], + "bbox": [ + 274, + 160, + 467, + 231 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca1f22a41de4902f7b0341c6dcf0afb2d972bda445e03364d709fa7fc319c7bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFPS↑SSIM↑PSNR↑Sync↑D_lip↓
Wav2Lip490.8833.169.518.65
TalkingFace63nana6.45na
NVP+p2p470.8227.194.8318.34
Ours1100.9435.717.4112.81
", + "bbox": [ + 81, + 301, + 464, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons with Wav2Lip [26], Talking-Face [30] and NVP [33] +Pix2Pix [38]. Sync confidence score Sync and Lip Distance metric $\\mathbf{D}_{lip}$ are used to determine accuracy of lip synchronization. Structure similarity score SSIM and peak signal to noise ration PSNR are used to judge the image quality. Our methods outperform other methods by a large margin on image quality and trails close behind Wav2Lip in lip-sync accuracy while achieving the highest inference speed FPS by a significant margin.", + "bbox": [ + 75, + 416, + 468, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that we produce the best results. We also provide two metrics to evaluate the quality of the lip sync: the score provided by SyncNet [5] (Sync) and Lip Marker Distance $(\\mathrm{D}_{lip})$ in Table 1. In terms of lip-sync, our results are superior to all of the other comparisons except Wav2Lip. This is because Wav2Lip was trained to maximize Sync, which also translates into better $\\mathrm{D}_{lip}$ . Our method comes close to Wav2Lip in terms of lip sync quality, but with a higher visual quality. Since we do not have the ground truth videos for Talking Face, we are unable to compute SSIM, PSNR and $\\mathrm{D}_{lip}$ metrics, which rely on ground truth frames.", + "bbox": [ + 75, + 582, + 467, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance. All results shown in the Table 1 were run on a single RTX 3080 Ti with a Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz. Our method is close to twice as fast as the second fastest method, TalkingFace. This is because by using a vector of visemes as input, we are able to greatly simplify our network. Additionally, since our architecture does not use any custom layers, our models can be easily converted to TensorRT. This further increases our inference speeds up to 200 frames per second, making it very suitable for production-level interactive applications.", + "bbox": [ + 75, + 750, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56e31e96f053338fd832375cbe81f4b33daa7fe3c0e2fb1e443b44d1a22224d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 629, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bef5df2419c8a10008c1e34f1b09cb9245964c252962e8b742fcaa69519899fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 88, + 759, + 188 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03344dea644abf1f0426649c4890ed18d2abe2981ba40b0c8ec88f4c1d8628f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 88, + 890, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Discussion and Analysis", + "text_level": 1, + "bbox": [ + 500, + 313, + 718, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Collection and Capture Setup. We recorded subjects using cameras capable of capturing 6K footage at 30 frames per second. From the full-sized video, we crop out the head region and a high-resolution mouth image that is used to supervise the training. We then run facial landmark detection, pose estimation, and contour extraction on the head crops, which provides us with the information we need to produce line-drawings such as the ones in Fig. 2. For audio data, we extract visemes by providing the audio waveforms and a transcript to JALI [8].", + "bbox": [ + 496, + 338, + 890, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We detect facial keypoints and head pose using a commercial product AlgoFace (www.algoface.ai). To obtain contours, we first generate a head segmentation using a BiSeNetV2 [44] and then we draw a line on the background-foreground boundary. Finally, we use JALI [8] to extract visemes. The viseme window size is a hyper-parameter; we empirically determined that using a window of size 6 is optimal. Larger window sizes lead to muffled mouth motion and very small window sizes produces unstable results. For the generator and both of the discriminators, we use Adam optimizer with a learning rate of 0.0002 and learning rate scheduling. We train the system on a 20 minute video of the subject speaking and evaluate on out of domain TTS audio paired with arbitrary sequences of keypoints. The oracle and multi-modal renderer each take roughly 6-8 hours on two 3080 GPUs with a batch size of 1.", + "bbox": [ + 496, + 489, + 890, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Disentanglement. Our network takes two types of input: speech data represented as visemes and head motion data represented as keypoint drawings. These two modalities are inherently connected: humans move their heads rhythmically while talking in a way that matches their speech, e.g., nodding up and down subtly when saying \"yes\". Without any special handling, our network learns a strong correlation between the head motion and word being spoken. This results in the model producing mouth openings just from head motion, depicted in Fig. 8a. Generating synthetic data", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4591", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "using classical data augmentation techniques such as rotation, scale, translation, cropping, or adding noise would not be effective in our case. While it is true that it might slightly help with the overfitting as it introduces randomness to the data and it might have effect similar to regularization, this alone does not break the correlation between head motion and audio, see Fig. 8b and our supplementary video, the lips still move and the mouth is not in the neutral resting pose. Our data augmentation successfully breaks the correlation by introducing a variety of head pose and viseme combinations. Fig. 8c shows that our model properly keeps its mouth closed during silence, regardless of head motion. We empirically observed that at least $80\\%$ of the training data have to be synthetic data for the disentanglement to have desired effect.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cbf5ca958f004e9b8f06e2e83f03ee18d7505d35db8a3a9b131aed28d867c132.jpg", + "image_caption": [ + "Figure 9. Even though the oracle (a) can sometimes produce artifacts in our synthetic training data, our multimodal renderer (b) is robust to them." + ], + "image_footnote": [], + "bbox": [ + 80, + 349, + 274, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6568c5a263eb2d1f746db715620454c6325932ada3e6f4431f7bdb961c1a473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 349, + 467, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the keypoint mashing Section 3.3 can produce unnatural combinations that are substantially different from the oracle's training data, the synthetic data produced can contain noticeable artifacts (Fig. 9a). However, such artifacts only occur when the two frames have vastly different head poses; this is rare and results in only a small percentage of the training data being polluted. Our multi-modal network is robust to such artifacts, shown in Fig. 9b.", + "bbox": [ + 75, + 556, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different backbones and performance optimization. MobileNetV2 [28] is usually used as compressed architectures primarily for image recognition tasks. We tried to directly apply the MobileNetV2 backbone for our image synthesis tasks and observed that both large and small variants of the MobileNetV2 were slower than our current approach. Using hard swish activation as proposed in MobileNetV2 led to rendering artifacts across multiple runs, and we chose to use the original leaky relu activation functions. Finally, replacing all convolutions in our network with depth wise separable convolutions did improve the performance, but also introduced some quality degradation in the teeth and produced grid like artifacts near high contrast texture regions like the hair.", + "bbox": [ + 75, + 689, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 500, + 90, + 772, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our work significantly advances the state-of-the-art in generating realistic virtual human avatars, there are still certain limitations and a fair amount of potential future work that needs to be done before the virtual humans are truly indistinguishable from real people.", + "bbox": [ + 496, + 114, + 890, + 190 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One of the major limitations of our framework is that it does not perform well on large motions, head rotations, and extreme head poses. To alleviate this in a future work, we envision involving 3D geometry and other aspects from the 3D rendering domain, in particular, using a mesh as an intermediate representation instead of 2D images, which would allow for better occlusion and collision handling and for using 3D neural rendering techniques, such as deferred neural rendering [35] that have been proven very successful.", + "bbox": [ + 496, + 191, + 892, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we also observed certain texture-sticking artifacts between frames when the motion is large, this is due to the fully convolutional nature of our network; this is a known problem discussed in, for example, StyleGAN3 [19], and might be mitigated by adopting vision transformers [21]. Also, as a part of further work, we encourage exploration into how multiple modalities can be used to target the same part of the face, enabling us to selectively learn from each of the modalities [16]. In the context of this work, this could mean, for example, having visemes as one modality and a signal to controls smiling as another modality; both modalities used to synthesize lips.", + "bbox": [ + 496, + 328, + 893, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 523, + 617, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While the problem of creating virtual humans that perfectly mimic the appearance and behavior of real people is still far from being solved, our work considerably pushes the state-of-the-art boundaries. We presented a robust and efficient framework for generating photorealistic talking face animations from audio in real-time. Thanks to the proposed alternative data representation, a training system to prevent modality entanglement, and supervision from high resolution around the mouth area, we are able to produce superior face rendering quality with better lip synchronization compared to recent approaches, all while maintaining real-time inference.", + "bbox": [ + 496, + 549, + 890, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 746, + 658, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would like to thank Sajid Sadi, Ankur Gupta, Anthony Liot, Anil Unnikrishnan, Janvi Palan, and the rest of the team at neonlife.ai for their extensive engineering, hardware, and design support.", + "bbox": [ + 496, + 771, + 890, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 847, + 596, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised", + "bbox": [ + 506, + 872, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4592", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 2, 4", + "[2] Christoph Bregler, Michele Covell, and Malcolm Slaney. Video rewrite: Driving visual speech with audio. ACM Transactions on Graphics, 31:353-360, 01 1997. 2", + "[3] Lele Chen, Ross K. Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2019. 2, 3", + "[4] Sen Chen, Zhilei Liu, Jiaxing Liu, Zhengxiang Yan, and Longbiao Wang. Talking head generation with audio and speech related facial action units. arXiv preprint arxiv:2110.09951, 2021.3", + "[5] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In ACCV Workshops, 2016. 2, 7", + "[6] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 2", + "[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2", + "[8] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: An animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics, 35:1-11, 07 2016. 4, 7", + "[9] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2, 3", + "[10] Anna Frühstück, Ibraheem Alhashim, and Peter Wonka. TitleGAN: Synthesis of large-scale non-homogeneous textures. ACM Transactions on Graphics, 38(4):58:1-58:11, 2019. 3", + "[11] Anna Fruhstuck, Krishna Kumar Singh, Eli Shechtman, Niloy J Mitra, Peter Wonka, and Jingwan Lu. Insetgan for full-body image generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 7723-7732, 2022. 2", + "[12] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 2", + "[13] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 2", + "[16] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts GANs. In ECCV, 2022. 8", + "[17] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and Locally Consistent Image Completion. ACM Transactions on Graphics, 36(4):107:1-107:14, 2017. 3", + "[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 5967-5976, 2017. 4", + "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems, volume 34, pages 852-863, 2021. 8", + "[20] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of European Conference on Computer Vision, 2022. 2", + "[21] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. ViTGAN: Training GANs with vision transformers. In International Conference on Learning Representations, 2022. 8", + "[22] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Collaging class-specific gans for semantic image synthesis. In Proceedings of IEEE International Conference on Computer Vision, pages 14398-14407, 2021. 2", + "[23] Paul Mermelstein. Distance measures for speech recognition, psychological and instrumental. Pattern recognition and artificial intelligence, 116:374-388, 1976. 2", + "[24] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2", + "[25] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 6, 7", + "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of IEEE International Conference on Computer Vision, pages 1173-1182, October 2021. 2, 3", + "[28] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4593", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "residuals and linear bottlenecks. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 8", + "[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 4", + "[30] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 21425–21430, June 2022. 6, 7", + "[31] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018.2", + "[32] Supasorn Suwajanakorn, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: Learning lip sync from audio. ACM Transactions on Graphics, 36(4), 2017. 2", + "[33] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Proceedings of European Conference on Computer Vision, pages 716-731, 2020. 2, 3, 6, 7", + "[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 2387-2395, 2016. 2", + "[35] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics, 38:1-12, 07 2019. 8", + "[36] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven realistic facial animation with temporal gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Workshops, June 2019. 2", + "[37] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 2020. 3", + "[38] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems, pages 1144-1156, 2018. 6, 7", + "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 4", + "[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[41] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to struc" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6", + "[42] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In Proceedings of European Conference on Computer Vision, 2018. 2", + "[43] Xiaogang Xu and Ning Xu. Hierarchical image generation via transformer-based sequential patch selection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2938-2945, 2022. 3", + "[44] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:1-18, 11 2021. 7", + "[45] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor S. Lempitsky. Few-shot adversarial learning of realistic neural talking head models. Proceedings of IEEE International Conference on Computer Vision, pages 9458-9467, 2019. 2", + "[46] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2659-2663, 2022. 2", + "[47] Yunxuan Zhang, Siwei Zhang, Yue He, Cheng Li, Chen Change Loy, and Ziwei Liu. One-shot face reenactment. In British Machine Vision Conference, 2019. 2", + "[48] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3661–3670, June 2021. 3", + "[49] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence, 2019. 2", + "[50] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[51] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makeittalk: Speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6), 2020. 2, 3, 6", + "[52] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics, 37(4):1-10, 2018. 4" + ], + "bbox": [ + 501, + 92, + 890, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4594", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_model.json b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1931d51025d83e35017854fbf1dff27d149a3210 --- /dev/null +++ b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.82, + 0.175 + ], + "angle": 0, + "content": "Synthesizing Photorealistic Virtual Humans Through Cross-modal Disentanglement" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.204, + 0.777, + 0.24 + ], + "angle": 0, + "content": "Siddarth Ravichandran, Ondrej Texler, Dimitar Dinev, Hyun Jae Kang NEON, Samsung Research America" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.242, + 0.731, + 0.256 + ], + "angle": 0, + "content": "{siddarth.r, o.texler, dimitar.d, hyunjae.k}@samsung.com" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.293, + 0.891, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.443, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Figure 1. Three sequences of talking faces generated using the proposed framework. Our method produces high-texture quality, sharp images, and correct lips shapes that are inline with the spoken audio. See the zoom-in patches, our method faithfully reproduces fine textural and content details such as teeth." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.498, + 0.314, + 0.515 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.551, + 0.474, + 0.884 + ], + "angle": 0, + "content": "Over the last few decades, many aspects of human life have been enhanced with virtual domains, from the advent of digital assistants such as Amazon's Alexa and Apple's Siri to the latest metaverse efforts of the rebranded Meta. These trends underscore the importance of generating photorealistic visual depictions of humans. This has led to the rapid growth of so-called deepfake and talking-head generation methods in recent years. Despite their impressive results and popularity, they usually lack certain qualitative aspects such as texture quality, lips synchronization, or resolution, and practical aspects such as the ability to run in real-time. To allow for virtual human avatars to be used in practical scenarios, we propose an end-to-end framework for synthesizing high-quality virtual human faces capable of speaking with accurate lip motion with a special emphasis on performance. We introduce a novel network utilizing visemes as an intermediate audio representation and a novel data augmentation strategy employing a hierarchical image synthesis approach that allows disentanglement of the different modalities used to control the global head motion. Our method runs in real-time, and is able to deliver superior results compared to the current state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.498, + 0.631, + 0.513 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.614 + ], + "angle": 0, + "content": "With the metaverse gaining traction in the media and the rapid virtualization of many aspects of life, virtual humans are becoming increasingly popular. Propped up by technological advances in hardware, GPU computing, and deep learning, digitization of humans has become a very active research topic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.84 + ], + "angle": 0, + "content": "One challenging problem is creating virtual avatars that attempt to faithfully recreate a specific human indistinguishable from their real counterparts. This entails a perfect visual recreation, with high quality textures that recreate the minutia that make a human visually realistic, e.g., hair that appears to be composed of individual strands, clothes that appear made out of fabric, and pores or vellus hair on the skin. Recent advances in neural rendering have achieved considerable success in creating very convincing images that are able to visually recreate such fine details. Additionally, many use-cases involving virtual humans require them to appear on large screens, e.g., human-size displays in airport halls, hotels, and billboards. For such use-cases, it is vital for the virtual humans to be rendered in a relatively high resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unfortunately, rendering quality alone is not sufficient to recreate a convincing virtual human that can be interacted with. Just as important as the visual fidelity is the fidelity of the motion. Specifically, for speech-driven avatars, the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "way the avatar talks, opens its mouth, and moves its lips is crucial. The human perception system is very sensitive to even the slightest mistake when it comes to the face, specifically the appearance of mouth and lips during speech. In order for a virtual avatar to be believable, its mouth and lip motion must match the speech a human listener hears." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.471, + 0.412 + ], + "angle": 0, + "content": "In this work, we introduce a novel training regime that utilizes 1-dimensional audio features such as visemes [9] or wav2vec 2.0 [1]. These are higher-level features compared to traditional features such as MFCC [23] which require more complex encoders. This enables us to efficiently synthesize talking heads with high-quality rendering and lip motion. However, this alone is insufficient for creating an avatar suitable for the aforementioned use-cases. In order to have more control over the avatar, it needs to be conditioned on multiple constraints in addition to speech, such as head pose or body shape. With speech represented as a 1-dimensional vector of visemes, head pose represented as rotation angles, and body shape represented as a 2-dimensional outline, we quickly run into a problem with different modalities among our data representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.471, + 0.655 + ], + "angle": 0, + "content": "Training deep neural networks on multimodal data is a difficult problem, especially when the different modalities are innately correlated. The network then tends to overfit on the modality that is easier to learn from, and this kind of overfitting leads to very poor performance at testing time. While there exist many techniques to mitigate traditional overfitting such as reducing number of parameters, regularization, etc., they are less effective when the overfitting is caused by multiple modalities. To break the correlations between these modalities, we introduce a novel data augmentation strategy: for every given lip shape, we use a generative oracle network to synthesize photorealistic images of the same lip shapes in a variety of head poses. This requires a generative network capable of producing images that are as high-resolution as our training data to preserve sharpness in important areas such as teeth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Even with modern hardware and the latest research, generative neural networks still struggle with generating high-resolution images, and the resolution usually comes at the cost of quality. When increasing the resolution and the complexity of what the network is asked to generate, the generative network often loses the semantic quality or other qualitative aspects of the result images. Increasing the network's capacity alone by adding more learnable parameters does not necessarily help and can introduce noise. To alleviate this, we propose a hierarchical approach to generating a high-quality synthetic image of a talking face that preserves the quality in the mouth region via high-resolution supervision." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In order to reproduce a believable interaction with a virtual avatar, inference speed is paramount. Our paper presents an efficient framework for creating high-quality" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "virtual artificial humans in real-time where each of the aforementioned concerns is addressed. The contribution of our works is:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.152, + 0.892, + 0.198 + ], + "angle": 0, + "content": "- A data augmentation method to disentangle audio and visual modalities so that the whole framework can be trained end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.211, + 0.892, + 0.242 + ], + "angle": 0, + "content": "- A hierarchical outpainting approach which allows for generation of high-resolution synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.254, + 0.892, + 0.298 + ], + "angle": 0, + "content": "- An end-to-end framework that utilizes 2-encoder-2-decoder neural network architecture and leverages synthetic data." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.152, + 0.892, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.316, + 0.642, + 0.332 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Talking faces. In the last several decades, many methods to generate videos of talking faces have been developed, and most of the recent ones heavily leverage generative adversarial networks-GANs [12]. Having the photorealism as a main objective, various methods, colloquially known as \"deepfakes\" [2,7,20,32,34,36,45], have been proposed and successfully used to synthesize images of human faces. Talking face animation can be categorized into face re-enactment approaches [15, 24, 25, 40, 42, 47] and speech driven approaches [3, 26, 27, 31, 46, 49, 51]. Speech driven models have used a variety of intermediate representations such as 2D keypoints [3,32], meshes [6,27,33], which themselves can be derived from various audio features such as MFCC, spectral coefficients, or DeepSpeech features [13]. Some methods directly produce talking faces from audio features without additional geometric representations [46] and some [26, 49] use an additional identity image or video as guiding signal. A common issue these methods face is the reduced high-frequency motion, such as lip motion during speech. Unrealistic lip movement is easy to notice, and disrupts the overall experience. This is partially addressed in the work of Prajwal et al. [26], they utilize SyncNet [5] to supervise the audio and lips synchronization, which results in a dynamic mouth movement. However, they do not consider the photorealistic appearance. As the mandate of our work is to create virtual humans that are indistinguishable from real people, we need to secure both realistic lip motion and photorealistic textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Hierarchical image generation. Synthesizing high-resolution images while retaining various fine details and overall sharpness is a difficult task. Researchers have been approaching this problem from various angles, and one of widely used solutions is to tackle this challenge in a hierarchical fashion. InsetGAN [11] propose to combine multiple pre-trained GANs, where one network generates the coarse body image, and other networks refine certain regions. The work of Li et al. [22] proposes to have several" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4586" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.889, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.376, + 0.896, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Overview of our multi-modal neural rendering pipeline. We generate \\( k \\) audio features of the provided audio, and sample a sequence of keypoint drawings and contours which are then mapped to a shared latent space using Audio Encoder (a) and the Landmark Encoder (b) respectively; which is then passed through the mouth (c) and face (d) decoders to produce the mouth and face images respectively. The mouth feature maps are pasted into the head feature map as denoted by the red arrow. During inference, only the output of the head decoder is used." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.629 + ], + "angle": 0, + "content": "class-specific generators that are used to refine certain semantic regions of the previously synthesized image. TileGAN [10] presents a way to combine small-resolution synthesized images in order to produce a single high-resolution texture. Xu and Xu [43] hierarchically compose the image from smaller patches, and Iizuka et al. [17] solve inpainting task using two discriminators, one for the whole image, and the other for a small area centered at the completed region. Although those approaches are not directly applicable in our talking virtual human scenario, our pipeline also leverages the hierarchical composition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Multi-modal training. To puppet a virtual human, multiple different inputs are usually needed, e.g., speech to control the lips and head-pose to specify head position and rotation. Those inputs-modalities are usually strongly correlated, which translates in the need of a multi-modal training [4, 37, 48, 50]. NeuralVoicePuppetry [33] trains a generalized expression network and an identity targeted neural renderer method decoupling the modalities using an intermediate mesh representation. MakeItTalk [51] propose to disentangle the modalities by learning displacements from a reference facial landmark using a separate module and perform rendering from the intermediate 2-dimensional landmark representations. Similarly, L. Chen et al. [3] use a learned high-level hierarchical structure to bridge audio signal with the pixel image. MeshTalk [27] deals with the modality disentanglement in a lower dimensional space by employing the cross-modality loss. Many of these methods require an intermediate representation for the face in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.541 + ], + "angle": 0, + "content": "form of drawings or mesh, hindering their ability to produce diverse lip shapes. We address these limitations in our work by using a mix of 1-dimensional and 2-dimensional representations, requiring modality disentanglement in order to synthesize speech-driven talking faces." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.553, + 0.65, + 0.571 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.777 + ], + "angle": 0, + "content": "We present a framework to generate photorealistic virtual human avatars that can be streamed in real-time to enable interactive applications. We train our system on a single identity and during the inference we drive it using audio features and partial face keypoints, as shown in Fig. 2. Due to bias in the way humans speak, there is a strong correlation between head motion and speech which causes the network to model undesirable relations. To combat that, we propose a novel data augmentation strategy employing keypoint mashing and a high-resolution generative oracle network to achieve disentanglement of lip motion from upper face motion. This allows synthesizing videos with arbitrary upper face keypoints and audio." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.652, + 0.803 + ], + "angle": 0, + "content": "3.1. Input Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Audio features. One of the crucial requirements for our method is the ability to faithfully reproduce motion and shapes of mouth and lips. Thus, it is vital to have an intermediate audio representation that is capable of capturing those geometric characteristics. Visemes [9] are lip shape categories that describe the basic lip shapes used to make" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "4587" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "sounds common in human speech. They can be thought of as the geometric analog to phonemes, and have been utilized to create a geometric basis for head models used in commercial 3D graphics products to parameterize and animate speech [52]. Despite the aforementioned theoretical advantages, visemes have certain practical difficulties, often requiring commercial tools such as JALI [8]. Our framework can also handle different frame-level audio feature representations such as wav2vec 2.0 [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Commonly used intermediate representations, such as keypoints [39] can be inaccurate, especially for difficult-to-detect regions such as lips. We propose to forego the use of intermediate geometric representation and instead render the image directly given visemes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.395 + ], + "angle": 0, + "content": "Facial features. To render the rest of the head, landmarks and contours are strong guiding signals for the network to produce semantically correct images. We use these features of the non-mouth areas of the face as additional inputs for the network, enabling control of non-speech-related behavior such as head motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.407, + 0.356, + 0.423 + ], + "angle": 0, + "content": "3.2. Multi-Modal Neural Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.506 + ], + "angle": 0, + "content": "We start by processing the audio to obtain \\( k \\) features per frame \\( \\mathbf{v} \\in \\mathbb{R}^{k \\times 1} \\). We take a window of 6 consecutive frames centered around the corresponding image frame. This results in a vector of size \\( \\mathbf{v} \\in \\mathbb{R}^{k \\times 6} \\), shown in Fig. 2a. For visemes, \\( k = 34 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.599 + ], + "angle": 0, + "content": "The second modality is the upper face keypoints and the contours drawn on a canvas of size \\(512 \\times 512\\) px, in Fig. 2b. Rather than passing the keypoints as vectors, we draw them as a 2D image which helps the model to learn spatial correspondences, resulting in better generalization to translations and motion in the keypoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Model architecture details. Our model consists of two encoders, one for each modality. The audio encoder consists of a 1D convolutional neural network followed by a 2-layer perceptron network. The resulting latent vector is reshaped to a 2D representation and concatenated with the latent feature map of the keypoint encoder which consist of 4-layer strided convolution blocks with a residual block [14] in between consecutive strided convolutions. Taking the inspiration from Pix2PixHD [39], we adopted a two image decoder scheme. One decoder is trained to generate a full face (Fig. 2d), and the second decoder (Fig. 2c) is trained to generate a fixed crop region of the mouth at a larger resolution. Both decoders produce different images from the same latent space. The model is trained end to end in a GAN setup, using two patch discriminators [18]: one for the head and the other for the mouth. We use the VGG [29] feature matching loss and smooth L1 loss alongside the GAN loss to train the generators. To obtain sharper mouth images, especially for the teeth, we resize and place the mouth region feature map into the head feature map at the appro" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.63, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.09, + 0.76, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.09, + 0.892, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Figure 3. Training a full-head model at \\(512\\mathrm{px}\\) resolution (a) produces lower details in the \\(192\\mathrm{px}\\) mouth region (b) than training a mouth-only model on \\(512\\mathrm{px}\\), then resampled to \\(192\\mathrm{px}\\) (c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.332 + ], + "angle": 0, + "content": "priate crop position in the penultimate layer of the head decoder, as shown by the red arrow in Fig. 2. We do this in the penultimate layer rather than the final image to avoid any blending artifacts that could arise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.855, + 0.361 + ], + "angle": 0, + "content": "3.3. Disentanglement of Audio and Head-Pose" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.521 + ], + "angle": 0, + "content": "As mentioned in Section 3.2, we use two inputs to the network: a \\( k \\times 6 \\) vector of audio features, and a \\( 512 \\times 512 \\) line drawing of the face. Since our training dataset only contains a single subject speaking one language, this leads to considerable entanglement between audio features and the 2D contour drawing, making the model learn unwanted relationships between the line drawing and lip motion. This manifests in, for example, lip motion being generated from just head motion in the contour drawing, even without providing any audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.749 + ], + "angle": 0, + "content": "To remedy this, we propose a novel data augmentation strategy for disentangling the audio features from the head pose. One way to enforce the correlation between audio features and lip shape is to show the network samples of the subject uttering the same phrases with different head poses. While this can be done during data capture, this significantly lengthens and complicates the capture process. To remedy this, we propose training an oracle network that converts 2D contour drawings into photoreal renders. Then, we propose a method for transferring the mouth shape from one frame to the head pose of another. Combining these two methods allows us to augment the capture dataset with synthetic images containing a greater variety of lip shapes and head poses, forcing the network to break the correlation between head pose and lip shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Oracle Network. The oracle generates photorealistic images of the subject given the 2D line drawings from in Fig. 2. This can be accomplished by employing a network based on Pix2Pix [18,39]. While the images produced by this network appear photorealistic, Pix2Pix can struggle with blurriness, especially in regions where there is significant motion and occlusions, such as the teeth. This blurriness pollutes our training data, degrading the visual quality of our method. One observation is that if we train a higher-resolution network that focuses on the mouth then the teeth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Our oracle network works in a hierarchical fashion, starting with the network to generate mouth (a) in a high-resolution. The result is then fed into the subsequent network (b) to generate the rest of the head. The head image and the high resolution mouth image are used for supervision of our multi-modal renderer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.459 + ], + "angle": 0, + "content": "will come out clearer, even if the result is then downsampled as shown in Fig. 3. Here, we trained two versions of the same network: one on the full-head \\(512 \\times 512\\) px data Fig. 3a and the other on high-resolution \\(512 \\times 512\\) px mouth crop, then downscaled to \\(192 \\times 192\\) px Fig. 3c. If we extract the same \\(192 \\times 192\\) region from the full head model (Fig. 3b), we can see that there is a degradation in quality in the teeth region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.471, + 0.67 + ], + "angle": 0, + "content": "To produce high-resolution renderings, we propose to render the image in a hierarchical manner starting with the region where the details and quality is most important for the given task – the mouth region in our case. Our solution can be viewed as a chaining of neural network models where the next network is conditioned by the result of the previous network, shown in Fig. 4. First, we train a network that generates a high-quality mouth image( Fig. 4a). This image is downsampled and placed in a line drawing of the full head, similar to the process outlined in Section 3.2. Then, it is used as input to the full-head network( Fig. 4b), which learns to render the rest of the head. We refer to this hierarchical approach as outpaint since it resembles the task of outpainting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.776 + ], + "angle": 0, + "content": "Keypoint mashing. The next missing piece for generating novel head pose and mouth shape combinations is the ability to combine the mouth keypoints from one frame with the head keypoints in another frame. However, the two frames will have different head poses, meaning we need to remove the global pose information before the mouth keypoints can be replaced." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.777, + 0.471, + 0.838 + ], + "angle": 0, + "content": "For any given frame, we extracted \\(n\\) 2D facial keypoints \\(\\mathbf{K} \\in \\mathbb{R}^{n \\times 2}\\) and a pose matrix \\(\\mathbf{P} \\in \\mathbb{R}^{4 \\times 4}\\) based on a 3D canonical space for the keypoints. The position of the keypoints can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.844, + 0.47, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} = p r o j (\\mathbf {P K ^ {f}}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}^{\\mathbf{f}}\\in \\mathbb{R}^{n\\times 4}\\) are the un-posed keypoints with depth in homogeneous coordinates, and \\(proj()\\) is a projection func" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Figure 5. Using keypoint mashing and the proposed oracle network, we are able to generate a synthetic image (c) by combining the mouth keypoints of (a) with the head keypoints of (b)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.388 + ], + "angle": 0, + "content": "tion (e.g. perspective projection) that converts keypoints from 3D space to 2D screen space. Using Eq. 1, we can compute the un-posed keypoints:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.399, + 0.891, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} ^ {\\mathbf {f}} = \\mathbf {P} ^ {- 1} p r o j ^ {- 1} (\\mathbf {K}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.642 + ], + "angle": 0, + "content": "Where \\( proj^{-1} \\) is the inverse projection function. For an orthographic projection, this would just be an inverse of the projection matrix, provided the keypoint data is 3D. However, many keypoints detectors only provide 2D information. In order to fill in the missing depth, we obtain depth values by solving the forward problem Eq. 1 for the canonical keypoints and concatenating the resulting depth value to the extracted 2D keypoints. This same process can be used to recover the perspective divide if \\( proj \\) is a perspective transformation. While this is only an approximation for the true depth values, most of the depth variation in the face comes from the underlying facial bone structure and geometry captured in the canonical keypoints, making this a viable approximation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.893, + 0.841 + ], + "angle": 0, + "content": "Generating synthetic data. With keypoint mashing and a generative oracle in place, we now define a method for data augmentation. We select two arbitrary frames \\(i\\) and \\(j\\), with corresponding keypoints and poses \\(\\mathbf{K}_i, \\mathbf{P}_i\\) (Fig. 5a, top) and \\(\\mathbf{K}_j, \\mathbf{P}_j\\) (Fig. 5b, top), we compute the un-posed keypoints \\(\\mathbf{K}_i^f\\). Then, we apply \\(\\mathbf{P}_j\\) resulting in a new keypoints set \\(\\mathbf{K}_{i \\rightarrow j}\\): the keypoints from frame \\(i\\) in the head pose of frame \\(j\\). Since the two keypoints sets \\(\\mathbf{K}_{i \\rightarrow j}\\) and \\(\\mathbf{K}_j\\) are now in the same space, we can trivially replace the mouth keypoints: \\(\\mathbf{K}_j^{mouth} = \\mathbf{K}_{i \\rightarrow j}^{mouth}\\) resulting in the mouth position from \\(i\\) transferred over to frame \\(j\\), \\(\\mathbf{K}_j'\\). Now we have audio features for frame \\(i\\), but a new keypoint set \\(\\mathbf{K}_j'\\) (Fig. 5c, top)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Using our oracle network, we can now generate synthetic images (Fig. 5c, bottom) from \\(\\mathbf{K}_j^\\prime\\) to be used as the \"ground truth\" for this new head pose and mouth shape combination. This makes the network to see different head poses" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4589" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.891, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.436, + 0.202, + 0.45 + ], + "angle": 0, + "content": "(a) MakeItTalk" + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.437, + 0.356, + 0.451 + ], + "angle": 0, + "content": "(b) Wav2lip" + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.437, + 0.535, + 0.45 + ], + "angle": 0, + "content": "(c) NVP+Pix2Pix" + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.437, + 0.667, + 0.45 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.437, + 0.853, + 0.45 + ], + "angle": 0, + "content": "(e) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Figure 6. Comparison with state-of-the-art techniques. MakeItTalk [51] (a), Wav2Lip [26] (b), NVP [33] + Pix2pix [38] (c), our result (d), and ground truth image (e). The avatar says 'ee'. Our results have significantly better lip shapes as compared to (a, c), and better texture quality as compared to (a, b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.576 + ], + "angle": 0, + "content": "for every audio feature frame, preventing it from learning an erroneous correlation between the head motion and lip shape." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.163, + 0.603 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.718 + ], + "angle": 0, + "content": "First we present comparisons with the current state-of-the-art techniques. Then, we elaborate on data collection and implementation details, discuss hyper-parameters, followed by further analysis and discussion. Our method is single identity, and each subject shown has their own model trained on their own dataset. For more results, please, see our supplementary pdf and video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.216, + 0.743 + ], + "angle": 0, + "content": "4.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparisons. We compare our method with two recent multi-identity approaches, Fig. 6a-b: MakeItTalk [51] and Wav2Lip [26]. These methods are trained on a large corpus of data containing many identities, and extract the identity of the subject at inference time. While it is indeed practical that MakeItTalk only requires a single image to synthesize a new identity, the resulting visual and lip sync quality is markedly worse than our method. Wav2Lip, which infers identity from a source video, arguably produces better lip sync than our method," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.817 + ], + "angle": 0, + "content": "but at the expense of visual quality: Wav2Lip only supports \\(96 \\times 96\\), of which only \\(48 \\times 96\\) are used for the mouth, and attempting higher resolutions fails. Notice that both of these above methods produce black blobs in the mouth region without any teeth textures. We then compared our method to TalkingFace [30], which is also trained on a single identity using data which is similar to our approach making the comparison straightforward. Our method produces superior lip motion: Fig. 7 shows how the two methods compare when producing the word \"we\". Our method is able to correctly capture the lip transition from the more rounded \"w\" sound to the \"e\" sound. Another single-identity method for generating talking heads is outlined in Neural Voice Puppetry [33]. We implemented a variation using Pix2Pix [38] and trained it on our data; our method produces superior lip motion Fig. 6c-d. Overall, our method produces higher quality videos with sufficiently accurate lip motion than the competing methods. Please refer to the supplementary video for more detailed comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparisons. In Table 1, we provide quantitative comparisons between our method and the above competing methods. To evaluate image quality, we compute the PSNR and SSIM [41] scores for 500 inferred images against their ground truth counterparts, and show" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.245, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.089, + 0.44, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.161, + 0.273, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.161, + 0.468, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Figure 7. TalkingFace [30] (top) and our model (bottom) were both asked to produce the word \"we\". Our model captures the transition from \"w\" (a) to \"e\" (b) more accurately." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.302, + 0.465, + 0.405 + ], + "angle": 0, + "content": "
MethodFPS↑SSIM↑PSNR↑Sync↑D_lip↓
Wav2Lip490.8833.169.518.65
TalkingFace63nana6.45na
NVP+p2p470.8227.194.8318.34
Ours1100.9435.717.4112.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.542 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with Wav2Lip [26], Talking-Face [30] and NVP [33] +Pix2Pix [38]. Sync confidence score Sync and Lip Distance metric \\(\\mathbf{D}_{lip}\\) are used to determine accuracy of lip synchronization. Structure similarity score SSIM and peak signal to noise ration PSNR are used to judge the image quality. Our methods outperform other methods by a large margin on image quality and trails close behind Wav2Lip in lip-sync accuracy while achieving the highest inference speed FPS by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.749 + ], + "angle": 0, + "content": "that we produce the best results. We also provide two metrics to evaluate the quality of the lip sync: the score provided by SyncNet [5] (Sync) and Lip Marker Distance \\((\\mathrm{D}_{lip})\\) in Table 1. In terms of lip-sync, our results are superior to all of the other comparisons except Wav2Lip. This is because Wav2Lip was trained to maximize Sync, which also translates into better \\(\\mathrm{D}_{lip}\\). Our method comes close to Wav2Lip in terms of lip sync quality, but with a higher visual quality. Since we do not have the ground truth videos for Talking Face, we are unable to compute SSIM, PSNR and \\(\\mathrm{D}_{lip}\\) metrics, which rely on ground truth frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Performance. All results shown in the Table 1 were run on a single RTX 3080 Ti with a Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz. Our method is close to twice as fast as the second fastest method, TalkingFace. This is because by using a vector of visemes as input, we are able to greatly simplify our network. Additionally, since our architecture does not use any custom layers, our models can be easily converted to TensorRT. This further increases our inference speeds up to 200 frames per second, making it very suitable for production-level interactive applications." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.63, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.089, + 0.761, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.089, + 0.892, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.284 + ], + "angle": 0, + "content": "Figure 8. The proposed augmentation is essential for the model to correctly respect different inputs. Without the augmentation (a), the virtual human erroneously opens her mouth even when the audio is silence; while (b) classical data augmentation methods might slightly mitigate this issue the model still produces wrong lip shapes; this is fixed by our augmentation technique (c)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.72, + 0.33 + ], + "angle": 0, + "content": "4.2. Discussion and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Data Collection and Capture Setup. We recorded subjects using cameras capable of capturing 6K footage at 30 frames per second. From the full-sized video, we crop out the head region and a high-resolution mouth image that is used to supervise the training. We then run facial landmark detection, pose estimation, and contour extraction on the head crops, which provides us with the information we need to produce line-drawings such as the ones in Fig. 2. For audio data, we extract visemes by providing the audio waveforms and a transcript to JALI [8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Implementation Details. We detect facial keypoints and head pose using a commercial product AlgoFace (www.algoface.ai). To obtain contours, we first generate a head segmentation using a BiSeNetV2 [44] and then we draw a line on the background-foreground boundary. Finally, we use JALI [8] to extract visemes. The viseme window size is a hyper-parameter; we empirically determined that using a window of size 6 is optimal. Larger window sizes lead to muffled mouth motion and very small window sizes produces unstable results. For the generator and both of the discriminators, we use Adam optimizer with a learning rate of 0.0002 and learning rate scheduling. We train the system on a 20 minute video of the subject speaking and evaluate on out of domain TTS audio paired with arbitrary sequences of keypoints. The oracle and multi-modal renderer each take roughly 6-8 hours on two 3080 GPUs with a batch size of 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Disentanglement. Our network takes two types of input: speech data represented as visemes and head motion data represented as keypoint drawings. These two modalities are inherently connected: humans move their heads rhythmically while talking in a way that matches their speech, e.g., nodding up and down subtly when saying \"yes\". Without any special handling, our network learns a strong correlation between the head motion and word being spoken. This results in the model producing mouth openings just from head motion, depicted in Fig. 8a. Generating synthetic data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4591" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "using classical data augmentation techniques such as rotation, scale, translation, cropping, or adding noise would not be effective in our case. While it is true that it might slightly help with the overfitting as it introduces randomness to the data and it might have effect similar to regularization, this alone does not break the correlation between head motion and audio, see Fig. 8b and our supplementary video, the lips still move and the mouth is not in the neutral resting pose. Our data augmentation successfully breaks the correlation by introducing a variety of head pose and viseme combinations. Fig. 8c shows that our model properly keeps its mouth closed during silence, regardless of head motion. We empirically observed that at least \\(80\\%\\) of the training data have to be synthetic data for the disentanglement to have desired effect." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.351, + 0.275, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.351, + 0.468, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.515 + ], + "angle": 0, + "content": "Figure 9. Even though the oracle (a) can sometimes produce artifacts in our synthetic training data, our multimodal renderer (b) is robust to them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Since the keypoint mashing Section 3.3 can produce unnatural combinations that are substantially different from the oracle's training data, the synthetic data produced can contain noticeable artifacts (Fig. 9a). However, such artifacts only occur when the two frames have vastly different head poses; this is rare and results in only a small percentage of the training data being polluted. Our multi-modal network is robust to such artifacts, shown in Fig. 9b." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Different backbones and performance optimization. MobileNetV2 [28] is usually used as compressed architectures primarily for image recognition tasks. We tried to directly apply the MobileNetV2 backbone for our image synthesis tasks and observed that both large and small variants of the MobileNetV2 were slower than our current approach. Using hard swish activation as proposed in MobileNetV2 led to rendering artifacts across multiple runs, and we chose to use the original leaky relu activation functions. Finally, replacing all convolutions in our network with depth wise separable convolutions did improve the performance, but also introduced some quality degradation in the teeth and produced grid like artifacts near high contrast texture regions like the hair." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.773, + 0.107 + ], + "angle": 0, + "content": "5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.191 + ], + "angle": 0, + "content": "While our work significantly advances the state-of-the-art in generating realistic virtual human avatars, there are still certain limitations and a fair amount of potential future work that needs to be done before the virtual humans are truly indistinguishable from real people." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.893, + 0.328 + ], + "angle": 0, + "content": "One of the major limitations of our framework is that it does not perform well on large motions, head rotations, and extreme head poses. To alleviate this in a future work, we envision involving 3D geometry and other aspects from the 3D rendering domain, in particular, using a mesh as an intermediate representation instead of 2D images, which would allow for better occlusion and collision handling and for using 3D neural rendering techniques, such as deferred neural rendering [35] that have been proven very successful." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.895, + 0.51 + ], + "angle": 0, + "content": "Furthermore, we also observed certain texture-sticking artifacts between frames when the motion is large, this is due to the fully convolutional nature of our network; this is a known problem discussed in, for example, StyleGAN3 [19], and might be mitigated by adopting vision transformers [21]. Also, as a part of further work, we encourage exploration into how multiple modalities can be used to target the same part of the face, enabling us to selectively learn from each of the modalities [16]. In the context of this work, this could mean, for example, having visemes as one modality and a signal to controls smiling as another modality; both modalities used to synthesize lips." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.619, + 0.54 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.732 + ], + "angle": 0, + "content": "While the problem of creating virtual humans that perfectly mimic the appearance and behavior of real people is still far from being solved, our work considerably pushes the state-of-the-art boundaries. We presented a robust and efficient framework for generating photorealistic talking face animations from audio in real-time. Thanks to the proposed alternative data representation, a training system to prevent modality entanglement, and supervision from high resolution around the mouth area, we are able to produce superior face rendering quality with better lip synchronization compared to recent approaches, all while maintaining real-time inference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.747, + 0.66, + 0.763 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.833 + ], + "angle": 0, + "content": "We would like to thank Sajid Sadi, Ankur Gupta, Anthony Liot, Anil Unnikrishnan, Janvi Palan, and the rest of the team at neonlife.ai for their extensive engineering, hardware, and design support." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.848, + 0.597, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4592" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.471, + 0.177 + ], + "angle": 0, + "content": "[2] Christoph Bregler, Michele Covell, and Malcolm Slaney. Video rewrite: Driving visual speech with audio. ACM Transactions on Graphics, 31:353-360, 01 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[3] Lele Chen, Ross K. Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.25, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[4] Sen Chen, Zhilei Liu, Jiaxing Liu, Zhengxiang Yan, and Longbiao Wang. Talking head generation with audio and speech related facial action units. arXiv preprint arxiv:2110.09951, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.307, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[5] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In ACCV Workshops, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.349, + 0.471, + 0.419 + ], + "angle": 0, + "content": "[6] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.421, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.491, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[8] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: An animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics, 35:1-11, 07 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.548, + 0.471, + 0.589 + ], + "angle": 0, + "content": "[9] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.632 + ], + "angle": 0, + "content": "[10] Anna Frühstück, Ibraheem Alhashim, and Peter Wonka. TitleGAN: Synthesis of large-scale non-homogeneous textures. ACM Transactions on Graphics, 38(4):58:1-58:11, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[11] Anna Fruhstuck, Krishna Kumar Singh, Eli Shechtman, Niloy J Mitra, Peter Wonka, and Jingwan Lu. Insetgan for full-body image generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 7723-7732, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[12] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[13] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[16] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts GANs. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[17] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and Locally Consistent Image Completion. ACM Transactions on Graphics, 36(4):107:1-107:14, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.893, + 0.317 + ], + "angle": 0, + "content": "[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 5967-5976, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems, volume 34, pages 852-863, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.893, + 0.445 + ], + "angle": 0, + "content": "[20] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[21] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. ViTGAN: Training GANs with vision transformers. In International Conference on Learning Representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.893, + 0.573 + ], + "angle": 0, + "content": "[22] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Collaging class-specific gans for semantic image synthesis. In Proceedings of IEEE International Conference on Computer Vision, pages 14398-14407, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.893, + 0.617 + ], + "angle": 0, + "content": "[23] Paul Mermelstein. Distance measures for speech recognition, psychological and instrumental. Pattern recognition and artificial intelligence, 116:374-388, 1976. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.893, + 0.673 + ], + "angle": 0, + "content": "[24] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.893, + 0.731 + ], + "angle": 0, + "content": "[25] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.893, + 0.8 + ], + "angle": 0, + "content": "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of IEEE International Conference on Computer Vision, pages 1173-1182, October 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[28] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4593" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "residuals and linear bottlenecks. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[30] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 21425–21430, June 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[31] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[32] Supasorn Suwajanakorn, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: Learning lip sync from audio. ACM Transactions on Graphics, 36(4), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[33] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Proceedings of European Conference on Computer Vision, pages 716-731, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 2387-2395, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[35] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics, 38:1-12, 07 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[36] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven realistic facial animation with temporal gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Workshops, June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[37] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[38] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems, pages 1144-1156, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to struc" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.12 + ], + "angle": 0, + "content": "tural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[42] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In Proceedings of European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[43] Xiaogang Xu and Ning Xu. Hierarchical image generation via transformer-based sequential patch selection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2938-2945, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[44] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:1-18, 11 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[45] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor S. Lempitsky. Few-shot adversarial learning of realistic neural talking head models. Proceedings of IEEE International Conference on Computer Vision, pages 9458-9467, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[46] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2659-2663, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[47] Yunxuan Zhang, Siwei Zhang, Yue He, Cheng Li, Chen Change Loy, and Ziwei Liu. One-shot face reenactment. In British Machine Vision Conference, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[48] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3661–3670, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.56, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[49] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[50] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[51] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makeittalk: Speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6), 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[52] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics, 37(4):1-10, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4594" + } + ] +] \ No newline at end of file diff --git a/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_origin.pdf b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aa87db2fa5ca145a9433d206495bec805f483873 --- /dev/null +++ b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/f1ee982b-9321-4c5b-bbc8-a11edd225f8f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2913c96a96a47bce411310de8f287134e82c5198184653194cd611597481395 +size 7201545 diff --git a/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/full.md b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a663037be1fb5c086d36968306baee0d76f1f452 --- /dev/null +++ b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/full.md @@ -0,0 +1,265 @@ +# Synthesizing Photorealistic Virtual Humans Through Cross-modal Disentanglement + +Siddarth Ravichandran, Ondrej Texler, Dimitar Dinev, Hyun Jae Kang NEON, Samsung Research America + +{siddarth.r, o.texler, dimitar.d, hyunjae.k}@samsung.com + +![](images/9455b5d041dda44847ef46f66803753111631caaaf7347704e5ea1bae1acd808.jpg) +Figure 1. Three sequences of talking faces generated using the proposed framework. Our method produces high-texture quality, sharp images, and correct lips shapes that are inline with the spoken audio. See the zoom-in patches, our method faithfully reproduces fine textural and content details such as teeth. + +# Abstract + +Over the last few decades, many aspects of human life have been enhanced with virtual domains, from the advent of digital assistants such as Amazon's Alexa and Apple's Siri to the latest metaverse efforts of the rebranded Meta. These trends underscore the importance of generating photorealistic visual depictions of humans. This has led to the rapid growth of so-called deepfake and talking-head generation methods in recent years. Despite their impressive results and popularity, they usually lack certain qualitative aspects such as texture quality, lips synchronization, or resolution, and practical aspects such as the ability to run in real-time. To allow for virtual human avatars to be used in practical scenarios, we propose an end-to-end framework for synthesizing high-quality virtual human faces capable of speaking with accurate lip motion with a special emphasis on performance. We introduce a novel network utilizing visemes as an intermediate audio representation and a novel data augmentation strategy employing a hierarchical image synthesis approach that allows disentanglement of the different modalities used to control the global head motion. Our method runs in real-time, and is able to deliver superior results compared to the current state-of-the-art. + +# 1. Introduction + +With the metaverse gaining traction in the media and the rapid virtualization of many aspects of life, virtual humans are becoming increasingly popular. Propped up by technological advances in hardware, GPU computing, and deep learning, digitization of humans has become a very active research topic. + +One challenging problem is creating virtual avatars that attempt to faithfully recreate a specific human indistinguishable from their real counterparts. This entails a perfect visual recreation, with high quality textures that recreate the minutia that make a human visually realistic, e.g., hair that appears to be composed of individual strands, clothes that appear made out of fabric, and pores or vellus hair on the skin. Recent advances in neural rendering have achieved considerable success in creating very convincing images that are able to visually recreate such fine details. Additionally, many use-cases involving virtual humans require them to appear on large screens, e.g., human-size displays in airport halls, hotels, and billboards. For such use-cases, it is vital for the virtual humans to be rendered in a relatively high resolution. + +Unfortunately, rendering quality alone is not sufficient to recreate a convincing virtual human that can be interacted with. Just as important as the visual fidelity is the fidelity of the motion. Specifically, for speech-driven avatars, the + +way the avatar talks, opens its mouth, and moves its lips is crucial. The human perception system is very sensitive to even the slightest mistake when it comes to the face, specifically the appearance of mouth and lips during speech. In order for a virtual avatar to be believable, its mouth and lip motion must match the speech a human listener hears. + +In this work, we introduce a novel training regime that utilizes 1-dimensional audio features such as visemes [9] or wav2vec 2.0 [1]. These are higher-level features compared to traditional features such as MFCC [23] which require more complex encoders. This enables us to efficiently synthesize talking heads with high-quality rendering and lip motion. However, this alone is insufficient for creating an avatar suitable for the aforementioned use-cases. In order to have more control over the avatar, it needs to be conditioned on multiple constraints in addition to speech, such as head pose or body shape. With speech represented as a 1-dimensional vector of visemes, head pose represented as rotation angles, and body shape represented as a 2-dimensional outline, we quickly run into a problem with different modalities among our data representation. + +Training deep neural networks on multimodal data is a difficult problem, especially when the different modalities are innately correlated. The network then tends to overfit on the modality that is easier to learn from, and this kind of overfitting leads to very poor performance at testing time. While there exist many techniques to mitigate traditional overfitting such as reducing number of parameters, regularization, etc., they are less effective when the overfitting is caused by multiple modalities. To break the correlations between these modalities, we introduce a novel data augmentation strategy: for every given lip shape, we use a generative oracle network to synthesize photorealistic images of the same lip shapes in a variety of head poses. This requires a generative network capable of producing images that are as high-resolution as our training data to preserve sharpness in important areas such as teeth. + +Even with modern hardware and the latest research, generative neural networks still struggle with generating high-resolution images, and the resolution usually comes at the cost of quality. When increasing the resolution and the complexity of what the network is asked to generate, the generative network often loses the semantic quality or other qualitative aspects of the result images. Increasing the network's capacity alone by adding more learnable parameters does not necessarily help and can introduce noise. To alleviate this, we propose a hierarchical approach to generating a high-quality synthetic image of a talking face that preserves the quality in the mouth region via high-resolution supervision. + +In order to reproduce a believable interaction with a virtual avatar, inference speed is paramount. Our paper presents an efficient framework for creating high-quality + +virtual artificial humans in real-time where each of the aforementioned concerns is addressed. The contribution of our works is: + +- A data augmentation method to disentangle audio and visual modalities so that the whole framework can be trained end-to-end. +- A hierarchical outpainting approach which allows for generation of high-resolution synthetic data. +- An end-to-end framework that utilizes 2-encoder-2-decoder neural network architecture and leverages synthetic data. + +# 2. Related Work + +Talking faces. In the last several decades, many methods to generate videos of talking faces have been developed, and most of the recent ones heavily leverage generative adversarial networks-GANs [12]. Having the photorealism as a main objective, various methods, colloquially known as "deepfakes" [2,7,20,32,34,36,45], have been proposed and successfully used to synthesize images of human faces. Talking face animation can be categorized into face re-enactment approaches [15, 24, 25, 40, 42, 47] and speech driven approaches [3, 26, 27, 31, 46, 49, 51]. Speech driven models have used a variety of intermediate representations such as 2D keypoints [3,32], meshes [6,27,33], which themselves can be derived from various audio features such as MFCC, spectral coefficients, or DeepSpeech features [13]. Some methods directly produce talking faces from audio features without additional geometric representations [46] and some [26, 49] use an additional identity image or video as guiding signal. A common issue these methods face is the reduced high-frequency motion, such as lip motion during speech. Unrealistic lip movement is easy to notice, and disrupts the overall experience. This is partially addressed in the work of Prajwal et al. [26], they utilize SyncNet [5] to supervise the audio and lips synchronization, which results in a dynamic mouth movement. However, they do not consider the photorealistic appearance. As the mandate of our work is to create virtual humans that are indistinguishable from real people, we need to secure both realistic lip motion and photorealistic textures. + +Hierarchical image generation. Synthesizing high-resolution images while retaining various fine details and overall sharpness is a difficult task. Researchers have been approaching this problem from various angles, and one of widely used solutions is to tackle this challenge in a hierarchical fashion. InsetGAN [11] propose to combine multiple pre-trained GANs, where one network generates the coarse body image, and other networks refine certain regions. The work of Li et al. [22] proposes to have several + +![](images/30b393907b45f629de66a669f799a04aebcb3d9f164c0a4de3dfa2ccfdb73f71.jpg) +Figure 2. Overview of our multi-modal neural rendering pipeline. We generate $k$ audio features of the provided audio, and sample a sequence of keypoint drawings and contours which are then mapped to a shared latent space using Audio Encoder (a) and the Landmark Encoder (b) respectively; which is then passed through the mouth (c) and face (d) decoders to produce the mouth and face images respectively. The mouth feature maps are pasted into the head feature map as denoted by the red arrow. During inference, only the output of the head decoder is used. + +class-specific generators that are used to refine certain semantic regions of the previously synthesized image. TileGAN [10] presents a way to combine small-resolution synthesized images in order to produce a single high-resolution texture. Xu and Xu [43] hierarchically compose the image from smaller patches, and Iizuka et al. [17] solve inpainting task using two discriminators, one for the whole image, and the other for a small area centered at the completed region. Although those approaches are not directly applicable in our talking virtual human scenario, our pipeline also leverages the hierarchical composition. + +Multi-modal training. To puppet a virtual human, multiple different inputs are usually needed, e.g., speech to control the lips and head-pose to specify head position and rotation. Those inputs-modalities are usually strongly correlated, which translates in the need of a multi-modal training [4, 37, 48, 50]. NeuralVoicePuppetry [33] trains a generalized expression network and an identity targeted neural renderer method decoupling the modalities using an intermediate mesh representation. MakeItTalk [51] propose to disentangle the modalities by learning displacements from a reference facial landmark using a separate module and perform rendering from the intermediate 2-dimensional landmark representations. Similarly, L. Chen et al. [3] use a learned high-level hierarchical structure to bridge audio signal with the pixel image. MeshTalk [27] deals with the modality disentanglement in a lower dimensional space by employing the cross-modality loss. Many of these methods require an intermediate representation for the face in the + +form of drawings or mesh, hindering their ability to produce diverse lip shapes. We address these limitations in our work by using a mix of 1-dimensional and 2-dimensional representations, requiring modality disentanglement in order to synthesize speech-driven talking faces. + +# 3. Our Approach + +We present a framework to generate photorealistic virtual human avatars that can be streamed in real-time to enable interactive applications. We train our system on a single identity and during the inference we drive it using audio features and partial face keypoints, as shown in Fig. 2. Due to bias in the way humans speak, there is a strong correlation between head motion and speech which causes the network to model undesirable relations. To combat that, we propose a novel data augmentation strategy employing keypoint mashing and a high-resolution generative oracle network to achieve disentanglement of lip motion from upper face motion. This allows synthesizing videos with arbitrary upper face keypoints and audio. + +# 3.1. Input Features + +Audio features. One of the crucial requirements for our method is the ability to faithfully reproduce motion and shapes of mouth and lips. Thus, it is vital to have an intermediate audio representation that is capable of capturing those geometric characteristics. Visemes [9] are lip shape categories that describe the basic lip shapes used to make + +sounds common in human speech. They can be thought of as the geometric analog to phonemes, and have been utilized to create a geometric basis for head models used in commercial 3D graphics products to parameterize and animate speech [52]. Despite the aforementioned theoretical advantages, visemes have certain practical difficulties, often requiring commercial tools such as JALI [8]. Our framework can also handle different frame-level audio feature representations such as wav2vec 2.0 [1]. + +Commonly used intermediate representations, such as keypoints [39] can be inaccurate, especially for difficult-to-detect regions such as lips. We propose to forego the use of intermediate geometric representation and instead render the image directly given visemes. + +Facial features. To render the rest of the head, landmarks and contours are strong guiding signals for the network to produce semantically correct images. We use these features of the non-mouth areas of the face as additional inputs for the network, enabling control of non-speech-related behavior such as head motion. + +# 3.2. Multi-Modal Neural Rendering + +We start by processing the audio to obtain $k$ features per frame $\mathbf{v} \in \mathbb{R}^{k \times 1}$ . We take a window of 6 consecutive frames centered around the corresponding image frame. This results in a vector of size $\mathbf{v} \in \mathbb{R}^{k \times 6}$ , shown in Fig. 2a. For visemes, $k = 34$ . + +The second modality is the upper face keypoints and the contours drawn on a canvas of size $512 \times 512$ px, in Fig. 2b. Rather than passing the keypoints as vectors, we draw them as a 2D image which helps the model to learn spatial correspondences, resulting in better generalization to translations and motion in the keypoints. + +Model architecture details. Our model consists of two encoders, one for each modality. The audio encoder consists of a 1D convolutional neural network followed by a 2-layer perceptron network. The resulting latent vector is reshaped to a 2D representation and concatenated with the latent feature map of the keypoint encoder which consist of 4-layer strided convolution blocks with a residual block [14] in between consecutive strided convolutions. Taking the inspiration from Pix2PixHD [39], we adopted a two image decoder scheme. One decoder is trained to generate a full face (Fig. 2d), and the second decoder (Fig. 2c) is trained to generate a fixed crop region of the mouth at a larger resolution. Both decoders produce different images from the same latent space. The model is trained end to end in a GAN setup, using two patch discriminators [18]: one for the head and the other for the mouth. We use the VGG [29] feature matching loss and smooth L1 loss alongside the GAN loss to train the generators. To obtain sharper mouth images, especially for the teeth, we resize and place the mouth region feature map into the head feature map at the appro + +![](images/c15ed914154b4d4131b5cacc2e2842bc76275dea903db4d674981999a259b7eb.jpg) +Figure 3. Training a full-head model at $512\mathrm{px}$ resolution (a) produces lower details in the $192\mathrm{px}$ mouth region (b) than training a mouth-only model on $512\mathrm{px}$ , then resampled to $192\mathrm{px}$ (c). + +![](images/fb0aadf133720a3e9a8e245943d79dc83552932432e67c2682a236047f3e4ee2.jpg) + +![](images/9018ec539fde69748b1e492f1a626f9b5cf8e37f2ca5ca0127ac3e9d19b254e9.jpg) + +priate crop position in the penultimate layer of the head decoder, as shown by the red arrow in Fig. 2. We do this in the penultimate layer rather than the final image to avoid any blending artifacts that could arise. + +# 3.3. Disentanglement of Audio and Head-Pose + +As mentioned in Section 3.2, we use two inputs to the network: a $k \times 6$ vector of audio features, and a $512 \times 512$ line drawing of the face. Since our training dataset only contains a single subject speaking one language, this leads to considerable entanglement between audio features and the 2D contour drawing, making the model learn unwanted relationships between the line drawing and lip motion. This manifests in, for example, lip motion being generated from just head motion in the contour drawing, even without providing any audio. + +To remedy this, we propose a novel data augmentation strategy for disentangling the audio features from the head pose. One way to enforce the correlation between audio features and lip shape is to show the network samples of the subject uttering the same phrases with different head poses. While this can be done during data capture, this significantly lengthens and complicates the capture process. To remedy this, we propose training an oracle network that converts 2D contour drawings into photoreal renders. Then, we propose a method for transferring the mouth shape from one frame to the head pose of another. Combining these two methods allows us to augment the capture dataset with synthetic images containing a greater variety of lip shapes and head poses, forcing the network to break the correlation between head pose and lip shape. + +Oracle Network. The oracle generates photorealistic images of the subject given the 2D line drawings from in Fig. 2. This can be accomplished by employing a network based on Pix2Pix [18,39]. While the images produced by this network appear photorealistic, Pix2Pix can struggle with blurriness, especially in regions where there is significant motion and occlusions, such as the teeth. This blurriness pollutes our training data, degrading the visual quality of our method. One observation is that if we train a higher-resolution network that focuses on the mouth then the teeth + +![](images/7e23d522c7194e2d8bb23383367379a3abe07a1378cd7f252add95c9af107669.jpg) +Figure 4. Our oracle network works in a hierarchical fashion, starting with the network to generate mouth (a) in a high-resolution. The result is then fed into the subsequent network (b) to generate the rest of the head. The head image and the high resolution mouth image are used for supervision of our multi-modal renderer. + +will come out clearer, even if the result is then downsampled as shown in Fig. 3. Here, we trained two versions of the same network: one on the full-head $512 \times 512$ px data Fig. 3a and the other on high-resolution $512 \times 512$ px mouth crop, then downscaled to $192 \times 192$ px Fig. 3c. If we extract the same $192 \times 192$ region from the full head model (Fig. 3b), we can see that there is a degradation in quality in the teeth region. + +To produce high-resolution renderings, we propose to render the image in a hierarchical manner starting with the region where the details and quality is most important for the given task – the mouth region in our case. Our solution can be viewed as a chaining of neural network models where the next network is conditioned by the result of the previous network, shown in Fig. 4. First, we train a network that generates a high-quality mouth image( Fig. 4a). This image is downsampled and placed in a line drawing of the full head, similar to the process outlined in Section 3.2. Then, it is used as input to the full-head network( Fig. 4b), which learns to render the rest of the head. We refer to this hierarchical approach as outpaint since it resembles the task of outpainting. + +Keypoint mashing. The next missing piece for generating novel head pose and mouth shape combinations is the ability to combine the mouth keypoints from one frame with the head keypoints in another frame. However, the two frames will have different head poses, meaning we need to remove the global pose information before the mouth keypoints can be replaced. + +For any given frame, we extracted $n$ 2D facial keypoints $\mathbf{K} \in \mathbb{R}^{n \times 2}$ and a pose matrix $\mathbf{P} \in \mathbb{R}^{4 \times 4}$ based on a 3D canonical space for the keypoints. The position of the keypoints can be expressed as: + +$$ +\mathbf {K} = p r o j (\mathbf {P K ^ {f}}) \tag {1} +$$ + +where $\mathbf{K}^{\mathbf{f}}\in \mathbb{R}^{n\times 4}$ are the un-posed keypoints with depth in homogeneous coordinates, and $proj()$ is a projection func + +![](images/7d0b2eedaceec469be29edda1acbd6219f379d75da292da63605e18a62fb3a50.jpg) +Figure 5. Using keypoint mashing and the proposed oracle network, we are able to generate a synthetic image (c) by combining the mouth keypoints of (a) with the head keypoints of (b). + +tion (e.g. perspective projection) that converts keypoints from 3D space to 2D screen space. Using Eq. 1, we can compute the un-posed keypoints: + +$$ +\mathbf {K} ^ {\mathbf {f}} = \mathbf {P} ^ {- 1} p r o j ^ {- 1} (\mathbf {K}) \tag {2} +$$ + +Where $proj^{-1}$ is the inverse projection function. For an orthographic projection, this would just be an inverse of the projection matrix, provided the keypoint data is 3D. However, many keypoints detectors only provide 2D information. In order to fill in the missing depth, we obtain depth values by solving the forward problem Eq. 1 for the canonical keypoints and concatenating the resulting depth value to the extracted 2D keypoints. This same process can be used to recover the perspective divide if $proj$ is a perspective transformation. While this is only an approximation for the true depth values, most of the depth variation in the face comes from the underlying facial bone structure and geometry captured in the canonical keypoints, making this a viable approximation. + +Generating synthetic data. With keypoint mashing and a generative oracle in place, we now define a method for data augmentation. We select two arbitrary frames $i$ and $j$ , with corresponding keypoints and poses $\mathbf{K}_i, \mathbf{P}_i$ (Fig. 5a, top) and $\mathbf{K}_j, \mathbf{P}_j$ (Fig. 5b, top), we compute the un-posed keypoints $\mathbf{K}_i^f$ . Then, we apply $\mathbf{P}_j$ resulting in a new keypoints set $\mathbf{K}_{i \rightarrow j}$ : the keypoints from frame $i$ in the head pose of frame $j$ . Since the two keypoints sets $\mathbf{K}_{i \rightarrow j}$ and $\mathbf{K}_j$ are now in the same space, we can trivially replace the mouth keypoints: $\mathbf{K}_j^{mouth} = \mathbf{K}_{i \rightarrow j}^{mouth}$ resulting in the mouth position from $i$ transferred over to frame $j$ , $\mathbf{K}_j'$ . Now we have audio features for frame $i$ , but a new keypoint set $\mathbf{K}_j'$ (Fig. 5c, top). + +Using our oracle network, we can now generate synthetic images (Fig. 5c, bottom) from $\mathbf{K}_j^\prime$ to be used as the "ground truth" for this new head pose and mouth shape combination. This makes the network to see different head poses + +![](images/733e20b67313ffd28c78fc285613011eba50df50103b93ad03605d26b66d9c1d.jpg) +(a) MakeItTalk +(b) Wav2lip +(c) NVP+Pix2Pix +(d) Ours +(e) Ground Truth +Figure 6. Comparison with state-of-the-art techniques. MakeItTalk [51] (a), Wav2Lip [26] (b), NVP [33] + Pix2pix [38] (c), our result (d), and ground truth image (e). The avatar says 'ee'. Our results have significantly better lip shapes as compared to (a, c), and better texture quality as compared to (a, b). + +for every audio feature frame, preventing it from learning an erroneous correlation between the head motion and lip shape. + +# 4. Results + +First we present comparisons with the current state-of-the-art techniques. Then, we elaborate on data collection and implementation details, discuss hyper-parameters, followed by further analysis and discussion. Our method is single identity, and each subject shown has their own model trained on their own dataset. For more results, please, see our supplementary pdf and video. + +# 4.1. Comparisons + +Qualitative comparisons. We compare our method with two recent multi-identity approaches, Fig. 6a-b: MakeItTalk [51] and Wav2Lip [26]. These methods are trained on a large corpus of data containing many identities, and extract the identity of the subject at inference time. While it is indeed practical that MakeItTalk only requires a single image to synthesize a new identity, the resulting visual and lip sync quality is markedly worse than our method. Wav2Lip, which infers identity from a source video, arguably produces better lip sync than our method, + +but at the expense of visual quality: Wav2Lip only supports $96 \times 96$ , of which only $48 \times 96$ are used for the mouth, and attempting higher resolutions fails. Notice that both of these above methods produce black blobs in the mouth region without any teeth textures. We then compared our method to TalkingFace [30], which is also trained on a single identity using data which is similar to our approach making the comparison straightforward. Our method produces superior lip motion: Fig. 7 shows how the two methods compare when producing the word "we". Our method is able to correctly capture the lip transition from the more rounded "w" sound to the "e" sound. Another single-identity method for generating talking heads is outlined in Neural Voice Puppetry [33]. We implemented a variation using Pix2Pix [38] and trained it on our data; our method produces superior lip motion Fig. 6c-d. Overall, our method produces higher quality videos with sufficiently accurate lip motion than the competing methods. Please refer to the supplementary video for more detailed comparisons. + +Quantitative comparisons. In Table 1, we provide quantitative comparisons between our method and the above competing methods. To evaluate image quality, we compute the PSNR and SSIM [41] scores for 500 inferred images against their ground truth counterparts, and show + +![](images/0070b703b6372ff6b1b24cca126de070c8ff86aacfad619b6541cbc9fdde6133.jpg) + +![](images/6759508db93345edd05fb68b192211ce33e24c37b13bdd5eea1fcad9f7a408ac.jpg) + +![](images/e014f936663919ac45c9501b5fce67866b26bec00f2be6b5f5a61f35052204da.jpg) +Figure 7. TalkingFace [30] (top) and our model (bottom) were both asked to produce the word "we". Our model captures the transition from "w" (a) to "e" (b) more accurately. + +![](images/d1acb79375dc082857b34467af717c68cab35b96b13995414b451cde77380c15.jpg) +Figure 8. The proposed augmentation is essential for the model to correctly respect different inputs. Without the augmentation (a), the virtual human erroneously opens her mouth even when the audio is silence; while (b) classical data augmentation methods might slightly mitigate this issue the model still produces wrong lip shapes; this is fixed by our augmentation technique (c). + +
MethodFPS↑SSIM↑PSNR↑Sync↑D_lip↓
Wav2Lip490.8833.169.518.65
TalkingFace63nana6.45na
NVP+p2p470.8227.194.8318.34
Ours1100.9435.717.4112.81
+ +Table 1. Quantitative comparisons with Wav2Lip [26], Talking-Face [30] and NVP [33] +Pix2Pix [38]. Sync confidence score Sync and Lip Distance metric $\mathbf{D}_{lip}$ are used to determine accuracy of lip synchronization. Structure similarity score SSIM and peak signal to noise ration PSNR are used to judge the image quality. Our methods outperform other methods by a large margin on image quality and trails close behind Wav2Lip in lip-sync accuracy while achieving the highest inference speed FPS by a significant margin. + +that we produce the best results. We also provide two metrics to evaluate the quality of the lip sync: the score provided by SyncNet [5] (Sync) and Lip Marker Distance $(\mathrm{D}_{lip})$ in Table 1. In terms of lip-sync, our results are superior to all of the other comparisons except Wav2Lip. This is because Wav2Lip was trained to maximize Sync, which also translates into better $\mathrm{D}_{lip}$ . Our method comes close to Wav2Lip in terms of lip sync quality, but with a higher visual quality. Since we do not have the ground truth videos for Talking Face, we are unable to compute SSIM, PSNR and $\mathrm{D}_{lip}$ metrics, which rely on ground truth frames. + +Performance. All results shown in the Table 1 were run on a single RTX 3080 Ti with a Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz. Our method is close to twice as fast as the second fastest method, TalkingFace. This is because by using a vector of visemes as input, we are able to greatly simplify our network. Additionally, since our architecture does not use any custom layers, our models can be easily converted to TensorRT. This further increases our inference speeds up to 200 frames per second, making it very suitable for production-level interactive applications. + +![](images/56e31e96f053338fd832375cbe81f4b33daa7fe3c0e2fb1e443b44d1a22224d8.jpg) + +![](images/bef5df2419c8a10008c1e34f1b09cb9245964c252962e8b742fcaa69519899fe.jpg) + +![](images/03344dea644abf1f0426649c4890ed18d2abe2981ba40b0c8ec88f4c1d8628f3.jpg) + +# 4.2. Discussion and Analysis + +Data Collection and Capture Setup. We recorded subjects using cameras capable of capturing 6K footage at 30 frames per second. From the full-sized video, we crop out the head region and a high-resolution mouth image that is used to supervise the training. We then run facial landmark detection, pose estimation, and contour extraction on the head crops, which provides us with the information we need to produce line-drawings such as the ones in Fig. 2. For audio data, we extract visemes by providing the audio waveforms and a transcript to JALI [8]. + +Implementation Details. We detect facial keypoints and head pose using a commercial product AlgoFace (www.algoface.ai). To obtain contours, we first generate a head segmentation using a BiSeNetV2 [44] and then we draw a line on the background-foreground boundary. Finally, we use JALI [8] to extract visemes. The viseme window size is a hyper-parameter; we empirically determined that using a window of size 6 is optimal. Larger window sizes lead to muffled mouth motion and very small window sizes produces unstable results. For the generator and both of the discriminators, we use Adam optimizer with a learning rate of 0.0002 and learning rate scheduling. We train the system on a 20 minute video of the subject speaking and evaluate on out of domain TTS audio paired with arbitrary sequences of keypoints. The oracle and multi-modal renderer each take roughly 6-8 hours on two 3080 GPUs with a batch size of 1. + +Disentanglement. Our network takes two types of input: speech data represented as visemes and head motion data represented as keypoint drawings. These two modalities are inherently connected: humans move their heads rhythmically while talking in a way that matches their speech, e.g., nodding up and down subtly when saying "yes". Without any special handling, our network learns a strong correlation between the head motion and word being spoken. This results in the model producing mouth openings just from head motion, depicted in Fig. 8a. Generating synthetic data + +using classical data augmentation techniques such as rotation, scale, translation, cropping, or adding noise would not be effective in our case. While it is true that it might slightly help with the overfitting as it introduces randomness to the data and it might have effect similar to regularization, this alone does not break the correlation between head motion and audio, see Fig. 8b and our supplementary video, the lips still move and the mouth is not in the neutral resting pose. Our data augmentation successfully breaks the correlation by introducing a variety of head pose and viseme combinations. Fig. 8c shows that our model properly keeps its mouth closed during silence, regardless of head motion. We empirically observed that at least $80\%$ of the training data have to be synthetic data for the disentanglement to have desired effect. + +![](images/cbf5ca958f004e9b8f06e2e83f03ee18d7505d35db8a3a9b131aed28d867c132.jpg) +Figure 9. Even though the oracle (a) can sometimes produce artifacts in our synthetic training data, our multimodal renderer (b) is robust to them. + +![](images/f6568c5a263eb2d1f746db715620454c6325932ada3e6f4431f7bdb961c1a473.jpg) + +Since the keypoint mashing Section 3.3 can produce unnatural combinations that are substantially different from the oracle's training data, the synthetic data produced can contain noticeable artifacts (Fig. 9a). However, such artifacts only occur when the two frames have vastly different head poses; this is rare and results in only a small percentage of the training data being polluted. Our multi-modal network is robust to such artifacts, shown in Fig. 9b. + +Different backbones and performance optimization. MobileNetV2 [28] is usually used as compressed architectures primarily for image recognition tasks. We tried to directly apply the MobileNetV2 backbone for our image synthesis tasks and observed that both large and small variants of the MobileNetV2 were slower than our current approach. Using hard swish activation as proposed in MobileNetV2 led to rendering artifacts across multiple runs, and we chose to use the original leaky relu activation functions. Finally, replacing all convolutions in our network with depth wise separable convolutions did improve the performance, but also introduced some quality degradation in the teeth and produced grid like artifacts near high contrast texture regions like the hair. + +# 5. Limitations and Future Work + +While our work significantly advances the state-of-the-art in generating realistic virtual human avatars, there are still certain limitations and a fair amount of potential future work that needs to be done before the virtual humans are truly indistinguishable from real people. + +One of the major limitations of our framework is that it does not perform well on large motions, head rotations, and extreme head poses. To alleviate this in a future work, we envision involving 3D geometry and other aspects from the 3D rendering domain, in particular, using a mesh as an intermediate representation instead of 2D images, which would allow for better occlusion and collision handling and for using 3D neural rendering techniques, such as deferred neural rendering [35] that have been proven very successful. + +Furthermore, we also observed certain texture-sticking artifacts between frames when the motion is large, this is due to the fully convolutional nature of our network; this is a known problem discussed in, for example, StyleGAN3 [19], and might be mitigated by adopting vision transformers [21]. Also, as a part of further work, we encourage exploration into how multiple modalities can be used to target the same part of the face, enabling us to selectively learn from each of the modalities [16]. In the context of this work, this could mean, for example, having visemes as one modality and a signal to controls smiling as another modality; both modalities used to synthesize lips. + +# 6. Conclusion + +While the problem of creating virtual humans that perfectly mimic the appearance and behavior of real people is still far from being solved, our work considerably pushes the state-of-the-art boundaries. We presented a robust and efficient framework for generating photorealistic talking face animations from audio in real-time. Thanks to the proposed alternative data representation, a training system to prevent modality entanglement, and supervision from high resolution around the mouth area, we are able to produce superior face rendering quality with better lip synchronization compared to recent approaches, all while maintaining real-time inference. + +# Acknowledgments + +We would like to thank Sajid Sadi, Ankur Gupta, Anthony Liot, Anil Unnikrishnan, Janvi Palan, and the rest of the team at neonlife.ai for their extensive engineering, hardware, and design support. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised + +learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 2, 4 +[2] Christoph Bregler, Michele Covell, and Malcolm Slaney. Video rewrite: Driving visual speech with audio. ACM Transactions on Graphics, 31:353-360, 01 1997. 2 +[3] Lele Chen, Ross K. Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2019. 2, 3 +[4] Sen Chen, Zhilei Liu, Jiaxing Liu, Zhengxiang Yan, and Longbiao Wang. Talking head generation with audio and speech related facial action units. arXiv preprint arxiv:2110.09951, 2021.3 +[5] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In ACCV Workshops, 2016. 2, 7 +[6] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 2 +[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2 +[8] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: An animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics, 35:1-11, 07 2016. 4, 7 +[9] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2, 3 +[10] Anna Frühstück, Ibraheem Alhashim, and Peter Wonka. TitleGAN: Synthesis of large-scale non-homogeneous textures. ACM Transactions on Graphics, 38(4):58:1-58:11, 2019. 3 +[11] Anna Fruhstuck, Krishna Kumar Singh, Eli Shechtman, Niloy J Mitra, Peter Wonka, and Jingwan Lu. Insetgan for full-body image generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 7723-7732, 2022. 2 +[12] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 2 +[13] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4 + +[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 2 +[16] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts GANs. In ECCV, 2022. 8 +[17] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and Locally Consistent Image Completion. ACM Transactions on Graphics, 36(4):107:1-107:14, 2017. 3 +[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 5967-5976, 2017. 4 +[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems, volume 34, pages 852-863, 2021. 8 +[20] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of European Conference on Computer Vision, 2022. 2 +[21] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. ViTGAN: Training GANs with vision transformers. In International Conference on Learning Representations, 2022. 8 +[22] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Collaging class-specific gans for semantic image synthesis. In Proceedings of IEEE International Conference on Computer Vision, pages 14398-14407, 2021. 2 +[23] Paul Mermelstein. Distance measures for speech recognition, psychological and instrumental. Pattern recognition and artificial intelligence, 116:374-388, 1976. 2 +[24] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2 +[25] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 6, 7 +[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of IEEE International Conference on Computer Vision, pages 1173-1182, October 2021. 2, 3 +[28] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted + +residuals and linear bottlenecks. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 8 +[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 4 +[30] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 21425–21430, June 2022. 6, 7 +[31] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018.2 +[32] Supasorn Suwajanakorn, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: Learning lip sync from audio. ACM Transactions on Graphics, 36(4), 2017. 2 +[33] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Proceedings of European Conference on Computer Vision, pages 716-731, 2020. 2, 3, 6, 7 +[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 2387-2395, 2016. 2 +[35] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics, 38:1-12, 07 2019. 8 +[36] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven realistic facial animation with temporal gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Workshops, June 2019. 2 +[37] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 2020. 3 +[38] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems, pages 1144-1156, 2018. 6, 7 +[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 4 +[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[41] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to struc + +tural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6 +[42] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In Proceedings of European Conference on Computer Vision, 2018. 2 +[43] Xiaogang Xu and Ning Xu. Hierarchical image generation via transformer-based sequential patch selection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2938-2945, 2022. 3 +[44] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:1-18, 11 2021. 7 +[45] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor S. Lempitsky. Few-shot adversarial learning of realistic neural talking head models. Proceedings of IEEE International Conference on Computer Vision, pages 9458-9467, 2019. 2 +[46] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2659-2663, 2022. 2 +[47] Yunxuan Zhang, Siwei Zhang, Yue He, Cheng Li, Chen Change Loy, and Ziwei Liu. One-shot face reenactment. In British Machine Vision Conference, 2019. 2 +[48] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3661–3670, June 2021. 3 +[49] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence, 2019. 2 +[50] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3 +[51] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makeittalk: Speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6), 2020. 2, 3, 6 +[52] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics, 37(4):1-10, 2018. 4 \ No newline at end of file diff --git a/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/images.zip b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..40f8fbaa6766422788fbc6c26a64bb9b9d4df142 --- /dev/null +++ b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54986707df7b38b8bc89a9018d00068be380eadbb27d74a1f22534c1ceff3d00 +size 389110 diff --git a/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/layout.json b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a56644a13d6fb05d865251fe4227b50a301f6fb --- /dev/null +++ b/2023/Synthesizing Photorealistic Virtual Humans Through Cross-Modal Disentanglement/layout.json @@ -0,0 +1,7173 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 501, + 138 + ], + "type": "text", + "content": "Synthesizing Photorealistic Virtual Humans Through Cross-modal Disentanglement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 161, + 475, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 161, + 475, + 190 + ], + "spans": [ + { + "bbox": [ + 117, + 161, + 475, + 190 + ], + "type": "text", + "content": "Siddarth Ravichandran, Ondrej Texler, Dimitar Dinev, Hyun Jae Kang NEON, Samsung Research America" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 191, + 447, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 191, + 447, + 202 + ], + "spans": [ + { + "bbox": [ + 143, + 191, + 447, + 202 + ], + "type": "text", + "content": "{siddarth.r, o.texler, dimitar.d, hyunjae.k}@samsung.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 232, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 50, + 232, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 50, + 232, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 232, + 545, + 342 + ], + "type": "image", + "image_path": "9455b5d041dda44847ef46f66803753111631caaaf7347704e5ea1bae1acd808.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "type": "text", + "content": "Figure 1. Three sequences of talking faces generated using the proposed framework. Our method produces high-texture quality, sharp images, and correct lips shapes that are inline with the spoken audio. See the zoom-in patches, our method faithfully reproduces fine textural and content details such as teeth." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 394, + 192, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 394, + 192, + 407 + ], + "spans": [ + { + "bbox": [ + 143, + 394, + 192, + 407 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 436, + 290, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 436, + 290, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 436, + 290, + 700 + ], + "type": "text", + "content": "Over the last few decades, many aspects of human life have been enhanced with virtual domains, from the advent of digital assistants such as Amazon's Alexa and Apple's Siri to the latest metaverse efforts of the rebranded Meta. These trends underscore the importance of generating photorealistic visual depictions of humans. This has led to the rapid growth of so-called deepfake and talking-head generation methods in recent years. Despite their impressive results and popularity, they usually lack certain qualitative aspects such as texture quality, lips synchronization, or resolution, and practical aspects such as the ability to run in real-time. To allow for virtual human avatars to be used in practical scenarios, we propose an end-to-end framework for synthesizing high-quality virtual human faces capable of speaking with accurate lip motion with a special emphasis on performance. We introduce a novel network utilizing visemes as an intermediate audio representation and a novel data augmentation strategy employing a hierarchical image synthesis approach that allows disentanglement of the different modalities used to control the global head motion. Our method runs in real-time, and is able to deliver superior results compared to the current state-of-the-art." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 394, + 386, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 386, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 386, + 406 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": "With the metaverse gaining traction in the media and the rapid virtualization of many aspects of life, virtual humans are becoming increasingly popular. Propped up by technological advances in hardware, GPU computing, and deep learning, digitization of humans has become a very active research topic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 665 + ], + "type": "text", + "content": "One challenging problem is creating virtual avatars that attempt to faithfully recreate a specific human indistinguishable from their real counterparts. This entails a perfect visual recreation, with high quality textures that recreate the minutia that make a human visually realistic, e.g., hair that appears to be composed of individual strands, clothes that appear made out of fabric, and pores or vellus hair on the skin. Recent advances in neural rendering have achieved considerable success in creating very convincing images that are able to visually recreate such fine details. Additionally, many use-cases involving virtual humans require them to appear on large screens, e.g., human-size displays in airport halls, hotels, and billboards. For such use-cases, it is vital for the virtual humans to be rendered in a relatively high resolution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Unfortunately, rendering quality alone is not sufficient to recreate a convincing virtual human that can be interacted with. Just as important as the visual fidelity is the fidelity of the motion. Specifically, for speech-driven avatars, the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4585" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "way the avatar talks, opens its mouth, and moves its lips is crucial. The human perception system is very sensitive to even the slightest mistake when it comes to the face, specifically the appearance of mouth and lips during speech. In order for a virtual avatar to be believable, its mouth and lip motion must match the speech a human listener hears." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 288, + 326 + ], + "type": "text", + "content": "In this work, we introduce a novel training regime that utilizes 1-dimensional audio features such as visemes [9] or wav2vec 2.0 [1]. These are higher-level features compared to traditional features such as MFCC [23] which require more complex encoders. This enables us to efficiently synthesize talking heads with high-quality rendering and lip motion. However, this alone is insufficient for creating an avatar suitable for the aforementioned use-cases. In order to have more control over the avatar, it needs to be conditioned on multiple constraints in addition to speech, such as head pose or body shape. With speech represented as a 1-dimensional vector of visemes, head pose represented as rotation angles, and body shape represented as a 2-dimensional outline, we quickly run into a problem with different modalities among our data representation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 288, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 518 + ], + "type": "text", + "content": "Training deep neural networks on multimodal data is a difficult problem, especially when the different modalities are innately correlated. The network then tends to overfit on the modality that is easier to learn from, and this kind of overfitting leads to very poor performance at testing time. While there exist many techniques to mitigate traditional overfitting such as reducing number of parameters, regularization, etc., they are less effective when the overfitting is caused by multiple modalities. To break the correlations between these modalities, we introduce a novel data augmentation strategy: for every given lip shape, we use a generative oracle network to synthesize photorealistic images of the same lip shapes in a variety of head poses. This requires a generative network capable of producing images that are as high-resolution as our training data to preserve sharpness in important areas such as teeth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 520, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 674 + ], + "type": "text", + "content": "Even with modern hardware and the latest research, generative neural networks still struggle with generating high-resolution images, and the resolution usually comes at the cost of quality. When increasing the resolution and the complexity of what the network is asked to generate, the generative network often loses the semantic quality or other qualitative aspects of the result images. Increasing the network's capacity alone by adding more learnable parameters does not necessarily help and can introduce noise. To alleviate this, we propose a hierarchical approach to generating a high-quality synthetic image of a talking face that preserves the quality in the mouth region via high-resolution supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "In order to reproduce a believable interaction with a virtual avatar, inference speed is paramount. Our paper presents an efficient framework for creating high-quality" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "virtual artificial humans in real-time where each of the aforementioned concerns is addressed. The contribution of our works is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 120, + 545, + 236 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 120, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 545, + 156 + ], + "type": "text", + "content": "- A data augmentation method to disentangle audio and visual modalities so that the whole framework can be trained end-to-end." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 167, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 167, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 545, + 191 + ], + "type": "text", + "content": "- A hierarchical outpainting approach which allows for generation of high-resolution synthetic data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 201, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 201, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 545, + 236 + ], + "type": "text", + "content": "- An end-to-end framework that utilizes 2-encoder-2-decoder neural network architecture and leverages synthetic data." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 250, + 392, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 250, + 392, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 250, + 392, + 262 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 270, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 604 + ], + "type": "text", + "content": "Talking faces. In the last several decades, many methods to generate videos of talking faces have been developed, and most of the recent ones heavily leverage generative adversarial networks-GANs [12]. Having the photorealism as a main objective, various methods, colloquially known as \"deepfakes\" [2,7,20,32,34,36,45], have been proposed and successfully used to synthesize images of human faces. Talking face animation can be categorized into face re-enactment approaches [15, 24, 25, 40, 42, 47] and speech driven approaches [3, 26, 27, 31, 46, 49, 51]. Speech driven models have used a variety of intermediate representations such as 2D keypoints [3,32], meshes [6,27,33], which themselves can be derived from various audio features such as MFCC, spectral coefficients, or DeepSpeech features [13]. Some methods directly produce talking faces from audio features without additional geometric representations [46] and some [26, 49] use an additional identity image or video as guiding signal. A common issue these methods face is the reduced high-frequency motion, such as lip motion during speech. Unrealistic lip movement is easy to notice, and disrupts the overall experience. This is partially addressed in the work of Prajwal et al. [26], they utilize SyncNet [5] to supervise the audio and lips synchronization, which results in a dynamic mouth movement. However, they do not consider the photorealistic appearance. As the mandate of our work is to create virtual humans that are indistinguishable from real people, we need to secure both realistic lip motion and photorealistic textures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Hierarchical image generation. Synthesizing high-resolution images while retaining various fine details and overall sharpness is a difficult task. Researchers have been approaching this problem from various angles, and one of widely used solutions is to tackle this challenge in a hierarchical fashion. InsetGAN [11] propose to combine multiple pre-trained GANs, where one network generates the coarse body image, and other networks refine certain regions. The work of Li et al. [22] proposes to have several" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4586" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 544, + 289 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 544, + 289 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 544, + 289 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 544, + 289 + ], + "type": "image", + "image_path": "30b393907b45f629de66a669f799a04aebcb3d9f164c0a4de3dfa2ccfdb73f71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 548, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 548, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 548, + 352 + ], + "type": "text", + "content": "Figure 2. Overview of our multi-modal neural rendering pipeline. We generate " + }, + { + "bbox": [ + 46, + 297, + 548, + 352 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 297, + 548, + 352 + ], + "type": "text", + "content": " audio features of the provided audio, and sample a sequence of keypoint drawings and contours which are then mapped to a shared latent space using Audio Encoder (a) and the Landmark Encoder (b) respectively; which is then passed through the mouth (c) and face (d) decoders to produce the mouth and face images respectively. The mouth feature maps are pasted into the head feature map as denoted by the red arrow. During inference, only the output of the head decoder is used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 366, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 498 + ], + "type": "text", + "content": "class-specific generators that are used to refine certain semantic regions of the previously synthesized image. TileGAN [10] presents a way to combine small-resolution synthesized images in order to produce a single high-resolution texture. Xu and Xu [43] hierarchically compose the image from smaller patches, and Iizuka et al. [17] solve inpainting task using two discriminators, one for the whole image, and the other for a small area centered at the completed region. Although those approaches are not directly applicable in our talking virtual human scenario, our pipeline also leverages the hierarchical composition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "Multi-modal training. To puppet a virtual human, multiple different inputs are usually needed, e.g., speech to control the lips and head-pose to specify head position and rotation. Those inputs-modalities are usually strongly correlated, which translates in the need of a multi-modal training [4, 37, 48, 50]. NeuralVoicePuppetry [33] trains a generalized expression network and an identity targeted neural renderer method decoupling the modalities using an intermediate mesh representation. MakeItTalk [51] propose to disentangle the modalities by learning displacements from a reference facial landmark using a separate module and perform rendering from the intermediate 2-dimensional landmark representations. Similarly, L. Chen et al. [3] use a learned high-level hierarchical structure to bridge audio signal with the pixel image. MeshTalk [27] deals with the modality disentanglement in a lower dimensional space by employing the cross-modality loss. Many of these methods require an intermediate representation for the face in the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 366, + 547, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 428 + ], + "type": "text", + "content": "form of drawings or mesh, hindering their ability to produce diverse lip shapes. We address these limitations in our work by using a mix of 1-dimensional and 2-dimensional representations, requiring modality disentanglement in order to synthesize speech-driven talking faces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 437, + 397, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 397, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 397, + 452 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 459, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 615 + ], + "type": "text", + "content": "We present a framework to generate photorealistic virtual human avatars that can be streamed in real-time to enable interactive applications. We train our system on a single identity and during the inference we drive it using audio features and partial face keypoints, as shown in Fig. 2. Due to bias in the way humans speak, there is a strong correlation between head motion and speech which causes the network to model undesirable relations. To combat that, we propose a novel data augmentation strategy employing keypoint mashing and a high-resolution generative oracle network to achieve disentanglement of lip motion from upper face motion. This allows synthesizing videos with arbitrary upper face keypoints and audio." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 623, + 399, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 399, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 399, + 635 + ], + "type": "text", + "content": "3.1. Input Features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Audio features. One of the crucial requirements for our method is the ability to faithfully reproduce motion and shapes of mouth and lips. Thus, it is vital to have an intermediate audio representation that is capable of capturing those geometric characteristics. Visemes [9] are lip shape categories that describe the basic lip shapes used to make" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "4587" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "sounds common in human speech. They can be thought of as the geometric analog to phonemes, and have been utilized to create a geometric basis for head models used in commercial 3D graphics products to parameterize and animate speech [52]. Despite the aforementioned theoretical advantages, visemes have certain practical difficulties, often requiring commercial tools such as JALI [8]. Our framework can also handle different frame-level audio feature representations such as wav2vec 2.0 [1]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 240 + ], + "type": "text", + "content": "Commonly used intermediate representations, such as keypoints [39] can be inaccurate, especially for difficult-to-detect regions such as lips. We propose to forego the use of intermediate geometric representation and instead render the image directly given visemes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 312 + ], + "type": "text", + "content": "Facial features. To render the rest of the head, landmarks and contours are strong guiding signals for the network to produce semantically correct images. We use these features of the non-mouth areas of the face as additional inputs for the network, enabling control of non-speech-related behavior such as head motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 322, + 217, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 217, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 217, + 335 + ], + "type": "text", + "content": "3.2. Multi-Modal Neural Rendering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "content": "We start by processing the audio to obtain " + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "content": " features per frame " + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{k \\times 1}" + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "content": ". We take a window of 6 consecutive frames centered around the corresponding image frame. This results in a vector of size " + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{k \\times 6}" + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "content": ", shown in Fig. 2a. For visemes, " + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "inline_equation", + "content": "k = 34" + }, + { + "bbox": [ + 46, + 341, + 287, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": "The second modality is the upper face keypoints and the contours drawn on a canvas of size " + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 402, + 287, + 474 + ], + "type": "text", + "content": " px, in Fig. 2b. Rather than passing the keypoints as vectors, we draw them as a 2D image which helps the model to learn spatial correspondences, resulting in better generalization to translations and motion in the keypoints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 475, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 714 + ], + "type": "text", + "content": "Model architecture details. Our model consists of two encoders, one for each modality. The audio encoder consists of a 1D convolutional neural network followed by a 2-layer perceptron network. The resulting latent vector is reshaped to a 2D representation and concatenated with the latent feature map of the keypoint encoder which consist of 4-layer strided convolution blocks with a residual block [14] in between consecutive strided convolutions. Taking the inspiration from Pix2PixHD [39], we adopted a two image decoder scheme. One decoder is trained to generate a full face (Fig. 2d), and the second decoder (Fig. 2c) is trained to generate a fixed crop region of the mouth at a larger resolution. Both decoders produce different images from the same latent space. The model is trained end to end in a GAN setup, using two patch discriminators [18]: one for the head and the other for the mouth. We use the VGG [29] feature matching loss and smooth L1 loss alongside the GAN loss to train the generators. To obtain sharper mouth images, especially for the teeth, we resize and place the mouth region feature map into the head feature map at the appro" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 385, + 149 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 385, + 149 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 385, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 385, + 149 + ], + "type": "image", + "image_path": "c15ed914154b4d4131b5cacc2e2842bc76275dea903db4d674981999a259b7eb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": "Figure 3. Training a full-head model at " + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "inline_equation", + "content": "512\\mathrm{px}" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": " resolution (a) produces lower details in the " + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "inline_equation", + "content": "192\\mathrm{px}" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": " mouth region (b) than training a mouth-only model on " + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "inline_equation", + "content": "512\\mathrm{px}" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": ", then resampled to " + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "inline_equation", + "content": "192\\mathrm{px}" + }, + { + "bbox": [ + 305, + 158, + 545, + 191 + ], + "type": "text", + "content": " (c)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 386, + 71, + 465, + 148 + ], + "blocks": [ + { + "bbox": [ + 386, + 71, + 465, + 148 + ], + "lines": [ + { + "bbox": [ + 386, + 71, + 465, + 148 + ], + "spans": [ + { + "bbox": [ + 386, + 71, + 465, + 148 + ], + "type": "image", + "image_path": "fb0aadf133720a3e9a8e245943d79dc83552932432e67c2682a236047f3e4ee2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 466, + 71, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 466, + 71, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 466, + 71, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 466, + 71, + 545, + 148 + ], + "type": "image", + "image_path": "9018ec539fde69748b1e492f1a626f9b5cf8e37f2ca5ca0127ac3e9d19b254e9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 214, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 262 + ], + "type": "text", + "content": "priate crop position in the penultimate layer of the head decoder, as shown by the red arrow in Fig. 2. We do this in the penultimate layer rather than the final image to avoid any blending artifacts that could arise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 273, + 523, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 523, + 285 + ], + "type": "text", + "content": "3.3. Disentanglement of Audio and Head-Pose" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "type": "text", + "content": "As mentioned in Section 3.2, we use two inputs to the network: a " + }, + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "type": "inline_equation", + "content": "k \\times 6" + }, + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "type": "text", + "content": " vector of audio features, and a " + }, + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 293, + 545, + 412 + ], + "type": "text", + "content": " line drawing of the face. Since our training dataset only contains a single subject speaking one language, this leads to considerable entanglement between audio features and the 2D contour drawing, making the model learn unwanted relationships between the line drawing and lip motion. This manifests in, for example, lip motion being generated from just head motion in the contour drawing, even without providing any audio." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 593 + ], + "type": "text", + "content": "To remedy this, we propose a novel data augmentation strategy for disentangling the audio features from the head pose. One way to enforce the correlation between audio features and lip shape is to show the network samples of the subject uttering the same phrases with different head poses. While this can be done during data capture, this significantly lengthens and complicates the capture process. To remedy this, we propose training an oracle network that converts 2D contour drawings into photoreal renders. Then, we propose a method for transferring the mouth shape from one frame to the head pose of another. Combining these two methods allows us to augment the capture dataset with synthetic images containing a greater variety of lip shapes and head poses, forcing the network to break the correlation between head pose and lip shape." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Oracle Network. The oracle generates photorealistic images of the subject given the 2D line drawings from in Fig. 2. This can be accomplished by employing a network based on Pix2Pix [18,39]. While the images produced by this network appear photorealistic, Pix2Pix can struggle with blurriness, especially in regions where there is significant motion and occlusions, such as the teeth. This blurriness pollutes our training data, degrading the visual quality of our method. One observation is that if we train a higher-resolution network that focuses on the mouth then the teeth" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4588" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 186 + ], + "type": "image", + "image_path": "7e23d522c7194e2d8bb23383367379a3abe07a1378cd7f252add95c9af107669.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 250 + ], + "type": "text", + "content": "Figure 4. Our oracle network works in a hierarchical fashion, starting with the network to generate mouth (a) in a high-resolution. The result is then fed into the subsequent network (b) to generate the rest of the head. The head image and the high resolution mouth image are used for supervision of our multi-modal renderer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "text", + "content": "will come out clearer, even if the result is then downsampled as shown in Fig. 3. Here, we trained two versions of the same network: one on the full-head " + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "text", + "content": " px data Fig. 3a and the other on high-resolution " + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "text", + "content": " px mouth crop, then downscaled to " + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "text", + "content": " px Fig. 3c. If we extract the same " + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 46, + 268, + 287, + 363 + ], + "type": "text", + "content": " region from the full head model (Fig. 3b), we can see that there is a degradation in quality in the teeth region." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 364, + 288, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 288, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 288, + 530 + ], + "type": "text", + "content": "To produce high-resolution renderings, we propose to render the image in a hierarchical manner starting with the region where the details and quality is most important for the given task – the mouth region in our case. Our solution can be viewed as a chaining of neural network models where the next network is conditioned by the result of the previous network, shown in Fig. 4. First, we train a network that generates a high-quality mouth image( Fig. 4a). This image is downsampled and placed in a line drawing of the full head, similar to the process outlined in Section 3.2. Then, it is used as input to the full-head network( Fig. 4b), which learns to render the rest of the head. We refer to this hierarchical approach as outpaint since it resembles the task of outpainting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 531, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 614 + ], + "type": "text", + "content": "Keypoint mashing. The next missing piece for generating novel head pose and mouth shape combinations is the ability to combine the mouth keypoints from one frame with the head keypoints in another frame. However, the two frames will have different head poses, meaning we need to remove the global pose information before the mouth keypoints can be replaced." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "text", + "content": "For any given frame, we extracted " + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "text", + "content": " 2D facial keypoints " + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{n \\times 2}" + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "text", + "content": " and a pose matrix " + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 47, + 615, + 288, + 663 + ], + "type": "text", + "content": " based on a 3D canonical space for the keypoints. The position of the keypoints can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 668, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 668, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 130, + 668, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {K} = p r o j (\\mathbf {P K ^ {f}}) \\tag {1}", + "image_path": "7e3bb2446372cd322d9b65dccd8c889348ae178e7d226942552a27a5f363dc4b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\mathbf{f}}\\in \\mathbb{R}^{n\\times 4}" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " are the un-posed keypoints with depth in homogeneous coordinates, and " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "proj()" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " is a projection func" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 208 + ], + "type": "image", + "image_path": "7d0b2eedaceec469be29edda1acbd6219f379d75da292da63605e18a62fb3a50.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 216, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 250 + ], + "type": "text", + "content": "Figure 5. Using keypoint mashing and the proposed oracle network, we are able to generate a synthetic image (c) by combining the mouth keypoints of (a) with the head keypoints of (b)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 271, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 307 + ], + "type": "text", + "content": "tion (e.g. perspective projection) that converts keypoints from 3D space to 2D screen space. Using Eq. 1, we can compute the un-posed keypoints:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 316, + 545, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 316, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 378, + 316, + 545, + 331 + ], + "type": "interline_equation", + "content": "\\mathbf {K} ^ {\\mathbf {f}} = \\mathbf {P} ^ {- 1} p r o j ^ {- 1} (\\mathbf {K}) \\tag {2}", + "image_path": "117cf61e38cf90c00cfa5857b115234fff88912cd9d63573055aaeb2103c23ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "type": "inline_equation", + "content": "proj^{-1}" + }, + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "type": "text", + "content": " is the inverse projection function. For an orthographic projection, this would just be an inverse of the projection matrix, provided the keypoint data is 3D. However, many keypoints detectors only provide 2D information. In order to fill in the missing depth, we obtain depth values by solving the forward problem Eq. 1 for the canonical keypoints and concatenating the resulting depth value to the extracted 2D keypoints. This same process can be used to recover the perspective divide if " + }, + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "type": "inline_equation", + "content": "proj" + }, + { + "bbox": [ + 304, + 340, + 545, + 508 + ], + "type": "text", + "content": " is a perspective transformation. While this is only an approximation for the true depth values, most of the depth variation in the face comes from the underlying facial bone structure and geometry captured in the canonical keypoints, making this a viable approximation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": "Generating synthetic data. With keypoint mashing and a generative oracle in place, we now define a method for data augmentation. We select two arbitrary frames " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ", with corresponding keypoints and poses " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_i, \\mathbf{P}_i" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " (Fig. 5a, top) and " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j, \\mathbf{P}_j" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " (Fig. 5b, top), we compute the un-posed keypoints " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_i^f" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ". Then, we apply " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_j" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " resulting in a new keypoints set " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{i \\rightarrow j}" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ": the keypoints from frame " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " in the head pose of frame " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ". Since the two keypoints sets " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{i \\rightarrow j}" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " are now in the same space, we can trivially replace the mouth keypoints: " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j^{mouth} = \\mathbf{K}_{i \\rightarrow j}^{mouth}" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " resulting in the mouth position from " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " transferred over to frame " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j'" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ". Now we have audio features for frame " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": ", but a new keypoint set " + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j'" + }, + { + "bbox": [ + 304, + 508, + 546, + 666 + ], + "type": "text", + "content": " (Fig. 5c, top)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": "Using our oracle network, we can now generate synthetic images (Fig. 5c, bottom) from " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_j^\\prime" + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": " to be used as the \"ground truth\" for this new head pose and mouth shape combination. This makes the network to see different head poses" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4589" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 545, + 342 + ], + "type": "image", + "image_path": "733e20b67313ffd28c78fc285613011eba50df50103b93ad03605d26b66d9c1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 345, + 123, + 356 + ], + "lines": [ + { + "bbox": [ + 62, + 345, + 123, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 123, + 356 + ], + "type": "text", + "content": "(a) MakeItTalk" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 346, + 217, + 357 + ], + "lines": [ + { + "bbox": [ + 168, + 346, + 217, + 357 + ], + "spans": [ + { + "bbox": [ + 168, + 346, + 217, + 357 + ], + "type": "text", + "content": "(b) Wav2lip" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 346, + 327, + 356 + ], + "lines": [ + { + "bbox": [ + 255, + 346, + 327, + 356 + ], + "spans": [ + { + "bbox": [ + 255, + 346, + 327, + 356 + ], + "type": "text", + "content": "(c) NVP+Pix2Pix" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 373, + 346, + 408, + 356 + ], + "lines": [ + { + "bbox": [ + 373, + 346, + 408, + 356 + ], + "spans": [ + { + "bbox": [ + 373, + 346, + 408, + 356 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 346, + 522, + 356 + ], + "lines": [ + { + "bbox": [ + 452, + 346, + 522, + 356 + ], + "spans": [ + { + "bbox": [ + 452, + 346, + 522, + 356 + ], + "type": "text", + "content": "(e) Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 365, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 546, + 399 + ], + "type": "text", + "content": "Figure 6. Comparison with state-of-the-art techniques. MakeItTalk [51] (a), Wav2Lip [26] (b), NVP [33] + Pix2pix [38] (c), our result (d), and ground truth image (e). The avatar says 'ee'. Our results have significantly better lip shapes as compared to (a, c), and better texture quality as compared to (a, b)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 419, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 456 + ], + "type": "text", + "content": "for every audio feature frame, preventing it from learning an erroneous correlation between the head motion and lip shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 465, + 99, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 99, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 99, + 477 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 485, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 568 + ], + "type": "text", + "content": "First we present comparisons with the current state-of-the-art techniques. Then, we elaborate on data collection and implementation details, discuss hyper-parameters, followed by further analysis and discussion. Our method is single identity, and each subject shown has their own model trained on their own dataset. For more results, please, see our supplementary pdf and video." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 575, + 132, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 132, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 132, + 588 + ], + "type": "text", + "content": "4.1. Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Qualitative comparisons. We compare our method with two recent multi-identity approaches, Fig. 6a-b: MakeItTalk [51] and Wav2Lip [26]. These methods are trained on a large corpus of data containing many identities, and extract the identity of the subject at inference time. While it is indeed practical that MakeItTalk only requires a single image to synthesize a new identity, the resulting visual and lip sync quality is markedly worse than our method. Wav2Lip, which infers identity from a source video, arguably produces better lip sync than our method," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "type": "text", + "content": "but at the expense of visual quality: Wav2Lip only supports " + }, + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "type": "text", + "content": ", of which only " + }, + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "type": "inline_equation", + "content": "48 \\times 96" + }, + { + "bbox": [ + 304, + 419, + 545, + 647 + ], + "type": "text", + "content": " are used for the mouth, and attempting higher resolutions fails. Notice that both of these above methods produce black blobs in the mouth region without any teeth textures. We then compared our method to TalkingFace [30], which is also trained on a single identity using data which is similar to our approach making the comparison straightforward. Our method produces superior lip motion: Fig. 7 shows how the two methods compare when producing the word \"we\". Our method is able to correctly capture the lip transition from the more rounded \"w\" sound to the \"e\" sound. Another single-identity method for generating talking heads is outlined in Neural Voice Puppetry [33]. We implemented a variation using Pix2Pix [38] and trained it on our data; our method produces superior lip motion Fig. 6c-d. Overall, our method produces higher quality videos with sufficiently accurate lip motion than the competing methods. Please refer to the supplementary video for more detailed comparisons." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "Quantitative comparisons. In Table 1, we provide quantitative comparisons between our method and the above competing methods. To evaluate image quality, we compute the PSNR and SSIM [41] scores for 500 inferred images against their ground truth counterparts, and show" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4590" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 149, + 127 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 149, + 127 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 149, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 149, + 127 + ], + "type": "image", + "image_path": "0070b703b6372ff6b1b24cca126de070c8ff86aacfad619b6541cbc9fdde6133.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 269, + 128 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 269, + 128 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 269, + 128 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 269, + 128 + ], + "type": "image", + "image_path": "6759508db93345edd05fb68b192211ce33e24c37b13bdd5eea1fcad9f7a408ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 127, + 167, + 183 + ], + "blocks": [ + { + "bbox": [ + 49, + 127, + 167, + 183 + ], + "lines": [ + { + "bbox": [ + 49, + 127, + 167, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 127, + 167, + 183 + ], + "type": "image", + "image_path": "e014f936663919ac45c9501b5fce67866b26bec00f2be6b5f5a61f35052204da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 226 + ], + "type": "text", + "content": "Figure 7. TalkingFace [30] (top) and our model (bottom) were both asked to produce the word \"we\". Our model captures the transition from \"w\" (a) to \"e\" (b) more accurately." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 127, + 286, + 183 + ], + "blocks": [ + { + "bbox": [ + 168, + 127, + 286, + 183 + ], + "lines": [ + { + "bbox": [ + 168, + 127, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 168, + 127, + 286, + 183 + ], + "type": "image", + "image_path": "d1acb79375dc082857b34467af717c68cab35b96b13995414b451cde77380c15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 158, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 224 + ], + "type": "text", + "content": "Figure 8. The proposed augmentation is essential for the model to correctly respect different inputs. Without the augmentation (a), the virtual human erroneously opens her mouth even when the audio is silence; while (b) classical data augmentation methods might slightly mitigate this issue the model still produces wrong lip shapes; this is fixed by our augmentation technique (c)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 239, + 284, + 320 + ], + "blocks": [ + { + "bbox": [ + 50, + 239, + 284, + 320 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 284, + 320 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 284, + 320 + ], + "type": "table", + "html": "
MethodFPS↑SSIM↑PSNR↑Sync↑D_lip↓
Wav2Lip490.8833.169.518.65
TalkingFace63nana6.45na
NVP+p2p470.8227.194.8318.34
Ours1100.9435.717.4112.81
", + "image_path": "ca1f22a41de4902f7b0341c6dcf0afb2d972bda445e03364d709fa7fc319c7bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 330, + 287, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 429 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with Wav2Lip [26], Talking-Face [30] and NVP [33] +Pix2Pix [38]. Sync confidence score Sync and Lip Distance metric " + }, + { + "bbox": [ + 46, + 330, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{lip}" + }, + { + "bbox": [ + 46, + 330, + 287, + 429 + ], + "type": "text", + "content": " are used to determine accuracy of lip synchronization. Structure similarity score SSIM and peak signal to noise ration PSNR are used to judge the image quality. Our methods outperform other methods by a large margin on image quality and trails close behind Wav2Lip in lip-sync accuracy while achieving the highest inference speed FPS by a significant margin." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "text", + "content": "that we produce the best results. We also provide two metrics to evaluate the quality of the lip sync: the score provided by SyncNet [5] (Sync) and Lip Marker Distance " + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}_{lip})" + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "text", + "content": " in Table 1. In terms of lip-sync, our results are superior to all of the other comparisons except Wav2Lip. This is because Wav2Lip was trained to maximize Sync, which also translates into better " + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{D}_{lip}" + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "text", + "content": ". Our method comes close to Wav2Lip in terms of lip sync quality, but with a higher visual quality. Since we do not have the ground truth videos for Talking Face, we are unable to compute SSIM, PSNR and " + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{D}_{lip}" + }, + { + "bbox": [ + 46, + 461, + 286, + 593 + ], + "type": "text", + "content": " metrics, which rely on ground truth frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "content": "Performance. All results shown in the Table 1 were run on a single RTX 3080 Ti with a Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz. Our method is close to twice as fast as the second fastest method, TalkingFace. This is because by using a vector of visemes as input, we are able to greatly simplify our network. Additionally, since our architecture does not use any custom layers, our models can be easily converted to TensorRT. This further increases our inference speeds up to 200 frames per second, making it very suitable for production-level interactive applications." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 385, + 150 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 385, + 150 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 385, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 385, + 150 + ], + "type": "image", + "image_path": "56e31e96f053338fd832375cbe81f4b33daa7fe3c0e2fb1e443b44d1a22224d8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 70, + 465, + 149 + ], + "blocks": [ + { + "bbox": [ + 386, + 70, + 465, + 149 + ], + "lines": [ + { + "bbox": [ + 386, + 70, + 465, + 149 + ], + "spans": [ + { + "bbox": [ + 386, + 70, + 465, + 149 + ], + "type": "image", + "image_path": "bef5df2419c8a10008c1e34f1b09cb9245964c252962e8b742fcaa69519899fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 466, + 70, + 545, + 149 + ], + "blocks": [ + { + "bbox": [ + 466, + 70, + 545, + 149 + ], + "lines": [ + { + "bbox": [ + 466, + 70, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 466, + 70, + 545, + 149 + ], + "type": "image", + "image_path": "03344dea644abf1f0426649c4890ed18d2abe2981ba40b0c8ec88f4c1d8628f3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 248, + 440, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 440, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 440, + 261 + ], + "type": "text", + "content": "4.2. Discussion and Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 268, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 387 + ], + "type": "text", + "content": "Data Collection and Capture Setup. We recorded subjects using cameras capable of capturing 6K footage at 30 frames per second. From the full-sized video, we crop out the head region and a high-resolution mouth image that is used to supervise the training. We then run facial landmark detection, pose estimation, and contour extraction on the head crops, which provides us with the information we need to produce line-drawings such as the ones in Fig. 2. For audio data, we extract visemes by providing the audio waveforms and a transcript to JALI [8]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 388, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 591 + ], + "type": "text", + "content": "Implementation Details. We detect facial keypoints and head pose using a commercial product AlgoFace (www.algoface.ai). To obtain contours, we first generate a head segmentation using a BiSeNetV2 [44] and then we draw a line on the background-foreground boundary. Finally, we use JALI [8] to extract visemes. The viseme window size is a hyper-parameter; we empirically determined that using a window of size 6 is optimal. Larger window sizes lead to muffled mouth motion and very small window sizes produces unstable results. For the generator and both of the discriminators, we use Adam optimizer with a learning rate of 0.0002 and learning rate scheduling. We train the system on a 20 minute video of the subject speaking and evaluate on out of domain TTS audio paired with arbitrary sequences of keypoints. The oracle and multi-modal renderer each take roughly 6-8 hours on two 3080 GPUs with a batch size of 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Disentanglement. Our network takes two types of input: speech data represented as visemes and head motion data represented as keypoint drawings. These two modalities are inherently connected: humans move their heads rhythmically while talking in a way that matches their speech, e.g., nodding up and down subtly when saying \"yes\". Without any special handling, our network learns a strong correlation between the head motion and word being spoken. This results in the model producing mouth openings just from head motion, depicted in Fig. 8a. Generating synthetic data" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4591" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "using classical data augmentation techniques such as rotation, scale, translation, cropping, or adding noise would not be effective in our case. While it is true that it might slightly help with the overfitting as it introduces randomness to the data and it might have effect similar to regularization, this alone does not break the correlation between head motion and audio, see Fig. 8b and our supplementary video, the lips still move and the mouth is not in the neutral resting pose. Our data augmentation successfully breaks the correlation by introducing a variety of head pose and viseme combinations. Fig. 8c shows that our model properly keeps its mouth closed during silence, regardless of head motion. We empirically observed that at least " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " of the training data have to be synthetic data for the disentanglement to have desired effect." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 277, + 168, + 366 + ], + "blocks": [ + { + "bbox": [ + 49, + 277, + 168, + 366 + ], + "lines": [ + { + "bbox": [ + 49, + 277, + 168, + 366 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 168, + 366 + ], + "type": "image", + "image_path": "cbf5ca958f004e9b8f06e2e83f03ee18d7505d35db8a3a9b131aed28d867c132.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 407 + ], + "type": "text", + "content": "Figure 9. Even though the oracle (a) can sometimes produce artifacts in our synthetic training data, our multimodal renderer (b) is robust to them." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 169, + 277, + 286, + 366 + ], + "blocks": [ + { + "bbox": [ + 169, + 277, + 286, + 366 + ], + "lines": [ + { + "bbox": [ + 169, + 277, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 169, + 277, + 286, + 366 + ], + "type": "image", + "image_path": "f6568c5a263eb2d1f746db715620454c6325932ada3e6f4431f7bdb961c1a473.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 441, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 538 + ], + "type": "text", + "content": "Since the keypoint mashing Section 3.3 can produce unnatural combinations that are substantially different from the oracle's training data, the synthetic data produced can contain noticeable artifacts (Fig. 9a). However, such artifacts only occur when the two frames have vastly different head poses; this is rare and results in only a small percentage of the training data being polluted. Our multi-modal network is robust to such artifacts, shown in Fig. 9b." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": "Different backbones and performance optimization. MobileNetV2 [28] is usually used as compressed architectures primarily for image recognition tasks. We tried to directly apply the MobileNetV2 backbone for our image synthesis tasks and observed that both large and small variants of the MobileNetV2 were slower than our current approach. Using hard swish activation as proposed in MobileNetV2 led to rendering artifacts across multiple runs, and we chose to use the original leaky relu activation functions. Finally, replacing all convolutions in our network with depth wise separable convolutions did improve the performance, but also introduced some quality degradation in the teeth and produced grid like artifacts near high contrast texture regions like the hair." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 473, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 473, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 473, + 84 + ], + "type": "text", + "content": "5. Limitations and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 151 + ], + "type": "text", + "content": "While our work significantly advances the state-of-the-art in generating realistic virtual human avatars, there are still certain limitations and a fair amount of potential future work that needs to be done before the virtual humans are truly indistinguishable from real people." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 152, + 546, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 546, + 259 + ], + "type": "text", + "content": "One of the major limitations of our framework is that it does not perform well on large motions, head rotations, and extreme head poses. To alleviate this in a future work, we envision involving 3D geometry and other aspects from the 3D rendering domain, in particular, using a mesh as an intermediate representation instead of 2D images, which would allow for better occlusion and collision handling and for using 3D neural rendering techniques, such as deferred neural rendering [35] that have been proven very successful." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 260, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 547, + 403 + ], + "type": "text", + "content": "Furthermore, we also observed certain texture-sticking artifacts between frames when the motion is large, this is due to the fully convolutional nature of our network; this is a known problem discussed in, for example, StyleGAN3 [19], and might be mitigated by adopting vision transformers [21]. Also, as a part of further work, we encourage exploration into how multiple modalities can be used to target the same part of the face, enabling us to selectively learn from each of the modalities [16]. In the context of this work, this could mean, for example, having visemes as one modality and a signal to controls smiling as another modality; both modalities used to synthesize lips." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 415, + 378, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 378, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 378, + 427 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 579 + ], + "type": "text", + "content": "While the problem of creating virtual humans that perfectly mimic the appearance and behavior of real people is still far from being solved, our work considerably pushes the state-of-the-art boundaries. We presented a robust and efficient framework for generating photorealistic talking face animations from audio in real-time. Thanks to the proposed alternative data representation, a training system to prevent modality entanglement, and supervision from high resolution around the mouth area, we are able to produce superior face rendering quality with better lip synchronization compared to recent approaches, all while maintaining real-time inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 591, + 403, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 591, + 403, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 591, + 403, + 604 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 611, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 659 + ], + "type": "text", + "content": "We would like to thank Sajid Sadi, Ankur Gupta, Anthony Liot, Anil Unnikrishnan, Janvi Palan, and the rest of the team at neonlife.ai for their extensive engineering, hardware, and design support." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 546, + 713 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4592" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449-12460, 2020. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 140 + ], + "type": "text", + "content": "[2] Christoph Bregler, Michele Covell, and Malcolm Slaney. Video rewrite: Driving visual speech with audio. ACM Transactions on Graphics, 31:353-360, 01 1997. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 288, + 196 + ], + "type": "text", + "content": "[3] Lele Chen, Ross K. Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 241 + ], + "type": "text", + "content": "[4] Sen Chen, Zhilei Liu, Jiaxing Liu, Zhengxiang Yan, and Longbiao Wang. Talking head generation with audio and speech related facial action units. arXiv preprint arxiv:2110.09951, 2021.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 274 + ], + "type": "text", + "content": "[5] Joon Son Chung and Andrew Zisserman. Out of time: Automated lip sync in the wild. In ACCV Workshops, 2016. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 331 + ], + "type": "text", + "content": "[6] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 333, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 288, + 387 + ], + "type": "text", + "content": "[7] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 432 + ], + "type": "text", + "content": "[8] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: An animator-centric viseme model for expressive lip synchronization. ACM Transactions on Graphics, 35:1-11, 07 2016. 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 434, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 288, + 466 + ], + "type": "text", + "content": "[9] Cletus G Fisher. Confusions among visually perceived consonants. Journal of speech and hearing research, 11(4):796-804, 1968. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "type": "text", + "content": "[10] Anna Frühstück, Ibraheem Alhashim, and Peter Wonka. TitleGAN: Synthesis of large-scale non-homogeneous textures. ACM Transactions on Graphics, 38(4):58:1-58:11, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 555 + ], + "type": "text", + "content": "[11] Anna Fruhstuck, Krishna Kumar Singh, Eli Shechtman, Niloy J Mitra, Peter Wonka, and Jingwan Lu. Insetgan for full-body image generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 7723-7732, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[12] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[13] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[15] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 162 + ], + "type": "text", + "content": "[16] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts GANs. In ECCV, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 196 + ], + "type": "text", + "content": "[17] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and Locally Consistent Image Completion. ACM Transactions on Graphics, 36(4):107:1-107:14, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 546, + 251 + ], + "type": "text", + "content": "[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 5967-5976, 2017. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 546, + 308 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Information Processing Systems, volume 34, pages 852-863, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 546, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 546, + 352 + ], + "type": "text", + "content": "[20] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In Proceedings of European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "type": "text", + "content": "[21] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. ViTGAN: Training GANs with vision transformers. In International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 453 + ], + "type": "text", + "content": "[22] Yuheng Li, Yijun Li, Jingwan Lu, Eli Shechtman, Yong Jae Lee, and Krishna Kumar Singh. Collaging class-specific gans for semantic image synthesis. In Proceedings of IEEE International Conference on Computer Vision, pages 14398-14407, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 546, + 488 + ], + "type": "text", + "content": "[23] Paul Mermelstein. Distance measures for speech recognition, psychological and instrumental. Pattern recognition and artificial intelligence, 116:374-388, 1976. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 490, + 546, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 546, + 533 + ], + "type": "text", + "content": "[24] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In Proceedings of IEEE International Conference on Computer Vision, pages 7184-7193, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 535, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 546, + 578 + ], + "type": "text", + "content": "[25] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 580, + 546, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 546, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 546, + 633 + ], + "type": "text", + "content": "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 690 + ], + "type": "text", + "content": "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando de la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of IEEE International Conference on Computer Vision, pages 1173-1182, October 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "type": "text", + "content": "[28] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4593" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "residuals and linear bottlenecks. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "text", + "content": "[30] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 21425–21430, June 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "text", + "content": "[31] Yang Song, Jingwen Zhu, Dawei Li, Xiaolong Wang, and Hairong Qi. Talking face generation by conditional recurrent adversarial network. arXiv preprint arXiv:1804.04786, 2018.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "text", + "content": "[32] Supasorn Suwajanakorn, Steven M. Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: Learning lip sync from audio. ACM Transactions on Graphics, 36(4), 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "text", + "content": "[33] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In Proceedings of European Conference on Computer Vision, pages 716-731, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 2387-2395, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[35] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics, 38:1-12, 07 2019. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[36] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. End-to-end speech-driven realistic facial animation with temporal gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Workshops, June 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[37] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[38] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems, pages 1144-1156, 2018. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[39] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 8798-8807, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[40] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to struc" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 632 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[42] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. Reenactgan: Learning to reenact faces via boundary transfer. In Proceedings of European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[43] Xiaogang Xu and Ning Xu. Hierarchical image generation via transformer-based sequential patch selection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):2938-2945, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[44] Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, and Nong Sang. Bisenet v2: Bilateral network with guided aggregation for real-time semantic segmentation. International Journal of Computer Vision, 129:1-18, 11 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "text", + "content": "[45] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor S. Lempitsky. Few-shot adversarial learning of realistic neural talking head models. Proceedings of IEEE International Conference on Computer Vision, pages 9458-9467, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[46] Sibo Zhang, Jiahong Yuan, Miao Liao, and Liangjun Zhang. Text2video: Text-driven talking-head video synthesis with personalized phoneme-pose dictionary. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2659-2663, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 386 + ], + "type": "text", + "content": "[47] Yunxuan Zhang, Siwei Zhang, Yue He, Cheng Li, Chen Change Loy, and Ziwei Liu. One-shot face reenactment. In British Machine Vision Conference, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": "[48] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3661–3670, June 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 443, + 545, + 487 + ], + "type": "text", + "content": "[49] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 542 + ], + "type": "text", + "content": "[50] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "text", + "content": "[51] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makeittalk: Speaker-aware talking-head animation. ACM Transactions on Graphics, 39(6), 2020. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[52] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics, 37(4):1-10, 2018. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4594" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_content_list.json b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7c829739452d8dedc009f6c78a6b4e8a838195b --- /dev/null +++ b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_content_list.json @@ -0,0 +1,1415 @@ +[ + { + "type": "text", + "text": "System-status-aware Adaptive Network for Online Streaming Video Understanding", + "text_level": 1, + "bbox": [ + 143, + 128, + 826, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lin Geng Foo $^{1\\dagger}$ Jia Gong $^{1\\dagger}$ Zhipeng Fan $^{2§}$ Jun Liu $^{1‡}$", + "bbox": [ + 245, + 202, + 720, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Singapore University of Technology and Design $^{2}$ New York University", + "bbox": [ + 187, + 220, + 779, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lingeng_foo, jia_gong}@mypail.sutd.edu.sg, zf606@nyu.edu,jun_riu@sutd.edu.sg", + "bbox": [ + 145, + 241, + 813, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed great progress in deep neural networks for real-time applications. However, most existing works do not explicitly consider the general case where the device's state and the available resources fluctuate over time, and none of them investigate or address the impact of varying computational resources for online video understanding tasks. This paper proposes a System-status-aware Adaptive Network (SAN) that considers the device's real-time state to provide high-quality predictions with low delay. Usage of our agent's policy improves efficiency and robustness to fluctuations of the system status. On two widely used video understanding tasks, SAN obtains state-of-the-art performance while constantly keeping processing delays low. Moreover, training such an agent on various types of hardware configurations is not easy as the labeled training data might not be available, or can be computationally prohibitive. To address this challenging problem, we propose a Meta Self-supervised Adaptation (MSA) method that adapts the agent's policy to new hardware configurations at test-time, allowing for easy deployment of the model onto other unseen hardware platforms.", + "bbox": [ + 75, + 323, + 472, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 670, + 209, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online video understanding, where certain predictions are immediately made for each video frame by using information in the current frame and potentially past frames, is an important task right at the intersection of video-based research and practical vision applications (e.g., self-driving vehicles [11], security surveillance [4], streaming services [32], and human-computer interactions [20]). In particular, in many of these real-world video-based applications, a fast and timely response is often crucial to ensure high usability and reduce potential security risk. Therefore, in many practical online applications, it is essential to ensure that the model is working with low delay while maintaining a good", + "bbox": [ + 75, + 695, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance, which can be challenging for many existing deep neural networks.", + "bbox": [ + 500, + 292, + 890, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, much effort has been made to reduce the delay of deep neural networks, including research into efficient network design [16, 36, 44], input-aware dynamic networks [5,6,12,24], and latency-constrained neural architectures [1, 2, 21]. However, all these works do not explicitly consider the dynamic conditions of the hardware platform, and assume stable computation resources are readily available. In practical scenarios, the accessible computing resources of the host devices can be fluctuating and dynamic due to the fact that multiple computationally expensive yet important threads are running concurrently. For example, in addition to performing vision-related tasks such as object detection, human activity recognition, and pose estimation, state-of-the-art robotic systems usually need to simultaneously perform additional tasks like simultaneous localization and mapping (SLAM) to successfully interact with humans and the environment. Those tasks are also often computationally heavy and could compete with vision tasks for computing resources. As a result, at times when the host device is busy with other processes, conducting inference for each model might require significantly more time than usual, leading to extremely long delays, which could cause safety issues and lagging responses in many real-world applications. Therefore, the study and development of models providing reliable yet timely responses under various hardware devices and fluctuating computing resources is crucially important. Unfortunately, such studies are lacking in the field.", + "bbox": [ + 496, + 325, + 893, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve and maintain low delay for online video understanding tasks under a dynamic computing resource budget, we propose a novel System-status-aware Adaptive Network (SAN). Different from previous works, SAN explicitly considers the system status of its host device to make on-the-fly adjustments to its computational complexity, and is thus capable of processing video streams effectively and efficiently in a dynamic system environment. SAN comprises of two components: a) a simple yet effective dynamic main module that offers reliable predictions under various", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Equal contribution; § Currently at Meta; ‡ Corresponding author", + "bbox": [ + 99, + 886, + 460, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10514", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "network depths and input resolutions; b) a lightweight agent that learns a dynamic system-status-aware policy used to control the execution of the main module, which facilitates adaptation to the fluctuating system load. With the adaptivity of the main module and the control policy generated by the agent, our SAN can achieve good performance on the online video understanding task while maintaining a low delay under fluctuating system loads.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In various applications, we may need to deploy SAN onto different hardware platforms for online video understanding. However, it is inconvenient to train SAN for each hardware platform, and it might also be difficult to find adequate storage to load the large labeled dataset on all platforms (e.g., mobile devices). In light of these difficulties, we further propose a method for deployment-time self-supervised agent adaptation, which we call Meta Self-supervised Adaptation (MSA). With MSA, we can conveniently train a SAN model on a set of local platforms, and perform a quick deployment-time agent adaptation on a target device, without the need for the original labeled training data. Specifically, our proposed MSA introduces an auxiliary task of delay prediction together with a meta-learning procedure, that facilitates the adaptation to the target deployment device.", + "bbox": [ + 75, + 212, + 472, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this paper are:", + "bbox": [ + 96, + 454, + 447, + 469 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to explicitly consider the fluctuating system status of the hardware device at inference time for online video understanding. To address this, we propose SAN, a novel system-status-aware network that adapts its behavior according to the video stream and the real-time status of the host system.", + "- We further propose a novel Meta Self-supervised Adaptation method MSA that alleviates the training burden and allows our model to effectively adapt to new host devices with potentially unclear computation profiles at deployment time.", + "- We empirically demonstrate that our proposed method achieves promising performance on the challenging online action recognition and pose estimation tasks, where we achieve low delays under a rapidly fluctuating system load without jeopardizing the quality of the predictions." + ], + "bbox": [ + 96, + 478, + 468, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 768, + 218, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Online Video Understanding. Recently, motivated by the increasing real-world demand, a lot of works [5,6,29,40, 46] have attempted to improve the accuracy and efficiency of models for online video understanding tasks, such as online action recognition [23, 40, 46] and online pose estimation [6, 12, 29]. Several works design efficient networks to improve models' efficiency for online video tasks, including", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the efficient 3D CNN design [46], the Temporal Shift 2D CNN network [23] and the Skip-Convolution [12], while other researchers introduce adaptivity into the networks, termed as stream-aware dynamic networks [5, 25, 29] to further reduce the network's computation complexity while maintaining a good performance. For example, LiteEval [40] dynamically chooses between a coarse model and a fine model for each frame and MAPN [6] dynamically activates its encoder to save computation resources. However, all these existing methods do not explicitly consider the dynamic conditions of the hardware platform, and thus face limitations in the face of fluctuating system conditions. For example, in situations when the system is under high computational load, the model may still select to execute the branch with high complexity and therefore incur significantly longer delays, which is sub-optimal for time-sensitive applications. Compared to these methods, our proposed SAN is system-status-aware and thus more robust to computational fluctuations on the models' host devices. To the best of our knowledge, such a system-status-aware adaptive network that can efficiently tackle online video understanding tasks has not been explored before.", + "bbox": [ + 496, + 90, + 893, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Network Designs. There are several paradigms for efficient deep architectures, and we list a few here. Efficient deep models such as SqueezeNet [17] and MobileNet [16] introduce novel computation operators using fewer parameters, while Neural Network Quantization [3,45] methods effectively shrink the model size of the existing models by quantizing model parameters to isolated values. Teacher-student models [15] are also effective in distilling knowledge from larger models into smaller ones. Some other latency-constrained networks [37, 42] further conduct neural architecture search (NAS) to search for an efficient architecture. Different from all these approaches, our SAN explicitly accounts for the dynamic system status that is constantly changing on the device over time, allowing us to perform efficient online video understanding even with a fluctuating system status.", + "bbox": [ + 496, + 441, + 893, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reinforcement Learning (RL). RL methods [22,27,28] train agents to navigate environments such that the reward is maximized. To allow pre-trained RL agents to adapt to a new environment in fewer episodes, methods based on the few-shot setting [7, 18] have been developed. Differently, we aim to allow our RL agent to adapt at deployment time in a self-supervised manner (i.e., no reward signals can be provided due to the absence of labeled video data). We introduce MSA, a novel method for self-supervised deployment-time adaptation. MSA employs an auxiliary task of delay prediction, and performs meta-learning to allow the RL agent to effectively adapt its control policy to the target hardware platform via fine-tuning on the auxiliary task.", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10515", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 89, + 168, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the live video stream $V = \\{I_1, \\dots, I_t, \\dots\\}$ and a seen/unseen host device whose system status $sys_t$ varies over time, our goal is to build a model that dynamically adjusts the computational complexity of inference based on not only the streaming input, but also the system status to make high-quality predictions with a low delay at each step $t$ . To this end, our proposed model: 1) takes into account the available computational resources to make on-the-fly decisions to dynamically adjust the computation consumption of the model, therefore maintaining the delay at low values while keeping the prediction's accuracy; and 2) effectively adapts itself to unseen devices in a self-supervised manner while still keeping its performance during deployment.", + "bbox": [ + 75, + 114, + 472, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SAN: System-status-aware Adaptive Network", + "text_level": 1, + "bbox": [ + 76, + 318, + 465, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For better online performance amid the host device's varying background computational load, we propose a System-status-aware Adaptive Network (SAN). As shown in Fig. 1, our SAN consists of two main components: a lightweight agent $\\pi$ that decides how to process the incoming frame and a dynamic main network $M$ that processes the frame with the specified policy generated by $\\pi$ .", + "bbox": [ + 75, + 340, + 468, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic main network. Our dynamic main network $M$ flexibly produces frame-wise predictions using an architecture determined by the policy from the agent module $\\pi$ . To accomplish this without extra burden on the hardware system, we design a dynamic encoder that has both dynamic depth and dynamic resolution. Our dynamic depth mechanism gives the encoder the option of producing output features at shallower intermediate layers without having to wait for all layers to be executed. This reduces computational cost by executing a \"simpler\" sub-network. On the other hand, the dynamic resolution mechanism allows our encoder to selectively take in input images at a lower resolution. Crucially, this reduces the amount of pixels to be examined by the network, effectively reducing the amount of convolution operations. We implement both of the aforementioned mechanisms in a single network, where to enable dynamic execution depth of the network, we add early exits to the network which allow the execution to end at an intermediate layer, while to enable dynamic resolution, we leverage the fully convolutional network and add a dynamic global pooling operation at the end.", + "bbox": [ + 75, + 448, + 472, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 1, our dynamic encoder selects from $m$ input resolutions $\\{res^i\\}_{i=1}^m$ and $n$ model depths $\\{dep^j\\}_{j=1}^n$ , allowing for as many as $m \\times n$ different options in total. These $m \\times n$ dynamic options have different computation complexities and levels of performance. For online video understanding, we construct our dynamic encoder as a 2D CNN for handling each frame, and to model the temporal aspect, the output features from the dynamic encoder are fed into a Long Short Term Memory network", + "bbox": [ + 75, + 765, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(LSTM) at every step. The LSTM updates its hidden state $h_t$ based on the features of the incoming frame $I_t$ and its previous state $h_{t-1}$ , which is then used by the task head to determine the final prediction $\\hat{y}_t$ for the video task following: $\\hat{y}_t = M(I_t, rest_t, dept_t, h_{t-1}; \\theta)$ , where $\\theta$ refers to the parameters of the main network $M$ , while $rest_t$ and $dept_t$ are selected by the agent introduced next. More details of our main dynamic network are in Sec. 4 and Supplementary.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RL-based agent. The RL-based agent $\\pi$ controls the dynamic main network $M$ by generating a frame-level processing policy $(\\text{rest}_t, \\text{depth}_t)$ at each time $t$ , which aims to maintain the task-related accuracy $\\text{acc}_t$ at high level with low delay $d_t$ . To make decisions amid the fluctuating system loads, our lightweight agent $\\pi$ should be both system-status-aware and streaming input-aware. We model our task as a Markov Decision Process, where the agent $\\pi$ takes an action $a_t$ at every step $t$ , based on the observed state $s_t$ , and transferred to the next state $s_{t+1}$ . Crucially, state $s_t$ includes system status information $\\text{sys}_t$ and input stream information $h_{t-1}$ , making $\\pi$ both system status and stream aware. To learn an optimal policy, we propose an RL-based method to train our agent.", + "bbox": [ + 496, + 212, + 892, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we set the observed states of the agent as $s_t = [h_{t-1}, sys_t, g_t]$ , where $sys_t$ is the system status (e.g., the CPU and GPU utility) $^1$ , $h_{t-1}$ is the LSTM hidden state, and $g_t$ refers to other useful information (i.e., the previous step's action $a_{t-1}$ and delay $d_{t-1}$ ). Then, based on the available resolutions $\\{res^i\\}_{i=1}^m$ and depths $\\{dep^j\\}_{j=1}^n$ , we build the action space as $\\{a^{i,j} = (res^i, dep^j)\\}_{i=1,j=1}^{m,n}$ . As shown in Fig. 1, at time $t$ , the agent $\\pi$ receives $h_{t-1}$ , $sys_t$ and $g_t$ , and generates a probability distribution over the action space:", + "bbox": [ + 496, + 424, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {t} = \\left[ p _ {t} ^ {1, 1}, p _ {t} ^ {1, 2}, \\dots , p _ {t} ^ {m, n} \\right] = \\pi \\left(h _ {t - 1}, s y s _ {t}, g _ {t}; \\phi\\right) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 589, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_t^{i,j} \\in [0,1]$ is the probability of the action $a^{i,j}$ at time $t$ , $P_t \\in [0,1]^{m\\cdot n}$ represents the probability distribution over the action space, and $\\phi$ refers to the parameters of the agent $\\pi$ . Then, we take the highest probability entry in $P_t$ (denoted as $p_t^{i*,j*}$ ), select the corresponding action $a_t = (res^{i*}, dep^{j*})$ and execute the main network accordingly.", + "bbox": [ + 496, + 619, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training of SAN. To train our main network $M$ , we apply the loss $\\mathcal{L}_M$ based on the specified online video-processing task. For example, for online action recognition, $\\mathcal{L}_M$ can be the cross-entropy loss, while for online pose estimation, $\\mathcal{L}_M$ can be the Mean Squared Error (MSE) loss.", + "bbox": [ + 496, + 713, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To control the dynamic encoder to achieve a good accuracy at low delay, we apply an RL-based loss $\\mathcal{L}_{\\pi}$ to train our agent $\\pi$ . Specifically, we adopt a policy gradient loss [38, 39]. To maximize the accuracy $acc_{t}$ while reducing the", + "bbox": [ + 496, + 789, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We also collect other system parameters (e.g., the temperature of each processor, the memory usage and the status of I/O) via three linux utilities [9, 10, 43] to form system status syst. More details are in Supplementary.", + "bbox": [ + 500, + 862, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87cdabc49c873cf45f453eabff9f9662ea65ed252862a28e319b74006f782617.jpg", + "image_caption": [ + "Figure 1. Overview of our SAN framework, which consists of two modules: the agent module $\\pi$ and the dynamic main network $M$ . For each incoming frame $I_{t}$ , the agent module first takes in the LSTM hidden state $h_{t-1}$ , system-status $sys_{t}$ , and other useful information $g_{t}$ and then decides the input resolution $rest_{t}$ and the depth $depth_{t}$ of $M$ via the policy head. After that, the main network $M$ processes the frame according to these decisions. By controlling the behavior of the main network $M$ via the agent $\\pi$ , our SAN can output high quality and low delay predictions in the dynamic system environment. Note that the auxiliary task head of the agent is used for fine-tuning of the agent during deployment, calibrating the learned system-status-aware features to adapt to the unseen target device." + ], + "image_footnote": [], + "bbox": [ + 158, + 95, + 802, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "delay $d_{t}$ , we set the reward of the agent at time $t$ as:", + "bbox": [ + 76, + 364, + 421, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} = \\lambda_ {a c c} a c c _ {t} \\left(a _ {t}, I _ {t}\\right) - \\max \\left(d _ {t} \\left(a _ {t}, s y s _ {t}\\right) - d _ {b}, 0\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 385, + 468, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $acc_{t}(a_{t},I_{t})$ and $d_t(a_t,sys_t)$ are the accuracy and delay of our main network under the action $a_{t}$ , the current frame $I_{t}$ , and system status $sys_{t}$ . $\\lambda_{acc}$ is the coefficient of the accuracy term, and $d_b$ is a tolerance threshold for the delay (e.g., $d_b = 30$ ms for a rate of 30 fps), which is the acceptable delay. The policy gradient loss $\\mathcal{L}_{\\pi}$ is defined as:", + "bbox": [ + 75, + 422, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\pi} = \\sum_ {t = 1} ^ {T} - r _ {t} \\log \\left(p _ {t}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 520, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_t \\in [0,1]$ is the probability of selecting the action $a_t$ at time $t$ .", + "bbox": [ + 75, + 568, + 468, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. MSA: Meta Self-supervised Adaptation", + "text_level": 1, + "bbox": [ + 76, + 604, + 415, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, we want to deploy SAN onto many different hardware platforms for online video applications, requiring us to train SAN for each of those platforms, which is labor intensive, computationally prohibitive, and may also be infeasible at times due to storage concerns (e.g., mobile devices). Instead, a more practical option would be to effectively fine-tune a pre-trained SAN at deployment-time on a target device. However, during deployment, the ground-truth labels associated with the real-time video stream are often inaccessible. To address this issue, we introduce an auxiliary task [13, 35] of delay prediction, where the agent $\\pi$ is expected to additionally predict the actual processing time based on the selected action (i.e., the input resolution and execution depth of main network $M$ ).", + "bbox": [ + 75, + 628, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our intuition is that, on a new (unseen) platform, our agent $\\pi$ lacks understanding of hardware status of this platform, thus we need a way for the agent to quickly understand it. Hence, we introduce the auxiliary task of delay", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prediction, because delay times are directly related to the system status and the model's computational cost. Therefore, when the agent $\\pi$ is fine-tuned using delay prediction, it will be pushed to adopt a better understanding of the host platform's characteristics and the costs of different options of $M$ on that platform. Moreover, delay times can be directly observed, and can be a convenient supervision signal on a new platform since we do not have signals from our main task (which requires labeled video data). Therefore, on a new platform, we can fine-tune our agent to predict the delay by using the observed delay times as supervision, which will improve the agent's understanding of the device.", + "bbox": [ + 496, + 364, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the delay prediction task improves the quality of the intermediate features (which are also used to select actions), fine-tuning with it does not necessarily lead to a better policy on the host device. Hence, to better align the performance on the auxiliary task of delay prediction with policy learning, we further propose a meta-optimization-based learning framework to find a good initialization for the agent, such that updating based on the loss of delay prediction would also improve the quality of the policy. Combining the auxiliary task and the meta-learning framework allows us to train SAN on a set of local hardware platforms, and perform only a quick deployment-time adaptation on the target device at run-time, without requiring the actual labeled video dataset on that target device.", + "bbox": [ + 496, + 549, + 892, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Auxiliary task of delay prediction. When SAN is deployed to an unseen target device, the processing times of each action at a given system status will shift significantly due to the drift in computing profiles. Therefore, fine-tuning on the target device is required to adapt SAN to the new hardware characteristics. Due to the lack of ground-truth labels for the streaming videos at run-time, we introduce a self-supervised auxiliary task of delay prediction for adapting SAN to unseen hardware platforms. The delay predic", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion task calibrates the learned system-status-aware features for the agent module $\\pi$ using supervisory signals (processing times) that can be generated on-the-fly with only a simple timer on the device.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, an auxiliary task head is added to the agent $\\pi$ , as shown in Fig. 1. The auxiliary task head takes in the features from the agent as well as the chosen action $a_{t}$ , to predict the delay time $d_{t}$ . To train our model on the auxiliary task, we apply a mean-squared-error loss between our predictions $\\hat{d}_t$ and the observed delay $d_{t}$ . The auxiliary loss $\\mathcal{L}_{aux}$ at step $t$ , is thus defined as: $\\mathcal{L}_{aux,t} = (d_t - \\hat{d}_t)^2$ .", + "bbox": [ + 76, + 151, + 467, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At deployment time, we fine-tune the agent $\\pi$ on the target device using this auxiliary task only, allowing the agent to better understand the target device's characteristics and processing times. However, this does not necessarily mean that the agent has optimized its policy output – in order to achieve that, we further introduce a meta-optimization method as described next.", + "bbox": [ + 75, + 257, + 467, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meta-optimization for self-supervised agent adaptation. During deployment onto the target device, optimizing the auxiliary task of delay prediction is loosely connected to improving the actual policy. To better bind the quality of the developed policy to the delay prediction performance, we introduce a meta-optimization step to our agent $\\pi$ during training, such that fine-tuning of the auxiliary task will also lead to a better adaptation of the policy.", + "bbox": [ + 75, + 363, + 467, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We approach this problem by simulating the fine-tuning process using the delay prediction error on multiple target platforms, and (meta) optimize $\\phi$ such that the update based on the delay prediction error leads to better adaptation of the policy as well. This will effectively link the quality of the policy to the delay prediction performance. Here, we adopt a meta-learning framework to tackle the problem, treating $\\phi$ as the meta-parameters to meta-optimize. In this context, we denote the auxiliary task as the meta-train task and the agent's policy generation as the meta-test task. Our meta-optimization algorithm aims to learn a good initialization of $\\phi$ , such that when it is updated via the meta-train task, it can achieve good performance on the meta-test task.", + "bbox": [ + 75, + 484, + 467, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assume that our SAN is first trained on a source device $D_{src}$ , and then adapted to the additional $K$ target platforms $\\{D_{tgt}^k\\}_{k=1}^K$ . After first training our SAN on the source device $D_{src}$ , we freeze the main network $M$ , and keep a copy of the agent's parameters (denoted as $\\phi_{meta,0}$ ) to initialize the meta-learning process. In the $u$ -th meta-learning iteration, we first meta-train the agent for each target device $D_{tgt}^k$ using its loss on the auxiliary task and perform a pseudo-update on the meta parameters $\\phi_{meta,u}$ which can be formalized as:", + "bbox": [ + 75, + 680, + 467, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {t g t, u} ^ {k} \\leftarrow \\phi_ {\\text {m e t a}, u} - \\alpha \\nabla_ {\\phi_ {\\text {m e t a}, u}} \\mathcal {L} _ {\\text {a u x}, t} \\left(\\phi_ {\\text {m e t a}, u}, D _ {t g t} ^ {k}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 837, + 467, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{tgt}^k$ refers to the parameters of the meta-trained agent $\\pi_{tgt}^k$ which is about to be deployed onto the target device $D_{tgt}^k$ , and $\\alpha$ is the learning rate for meta-training.", + "bbox": [ + 498, + 90, + 890, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we check the performance of the pseudo-updated parameters on target device by meta-testing the performance of the pseudo-updated agent $\\pi_{tgt}^{k}(:,\\phi_{tgt,u}^{k})$ for the agent's policy generation following:", + "bbox": [ + 498, + 140, + 890, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m e t a}} ^ {k} \\left(\\phi_ {t g t, u} ^ {k}\\right) = \\mathcal {L} _ {\\pi} \\left(\\phi_ {t g t, u} ^ {k}, D _ {t g t} ^ {k}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 210, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{meta}^k$ is the meta-testing-loss from the target hardware platform $D_{tgt}^k$ . Intuitively, this loss penalizes our agent based on its performance on a target device after a short period of fine-tuning. Hence, we update our meta parameters $\\phi_{meta,u}$ in the $u$ -th meta-learning iteration as follows:", + "bbox": [ + 498, + 241, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {m e t a, u + 1} \\leftarrow \\phi_ {m e t a, u} - \\beta \\sum_ {k = 1} ^ {K} \\nabla_ {\\phi_ {m e t a, u}} \\mathcal {L} _ {m e t a} ^ {k} \\left(\\phi_ {t g t, u} ^ {k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 329, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the learning rate for optimizing the meta parameters, and $\\phi_{tgt,u}^k$ is obtained via the pseudo-update in Eq. 4. Through updating our model with this meta-learning scheme, our agent parameters will converge towards optimal parameters $\\phi_{meta,*}$ that are able to effectively adapt to each target device through fine-tuning via the auxiliary task.", + "bbox": [ + 498, + 383, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Testing", + "text_level": 1, + "bbox": [ + 500, + 484, + 696, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned above, we consider two scenarios: deploying our SAN on a seen or unseen device.", + "bbox": [ + 498, + 508, + 888, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model training and testing for seen device. In this scenario, we evaluate SAN when it is trained and tested on the same device, which evaluates the effectiveness of our SAN design described in Sec. 3.1. For model training, we first utilize the dataset's training set to initialize our main network $M$ , which is controlled by a random policy (randomly sampling the resolution-depth pair) to improve its accuracy via the loss $\\mathcal{L}_M$ . Then, we train our agent $\\pi$ using the loss $\\mathcal{L}_{\\pi}$ with the parameters of the main network $(M)$ fixed. Lastly, we fine-tune the main network $M$ and agent $\\pi$ jointly in a fully end-to-end manner with both $\\mathcal{L}_M$ and $\\mathcal{L}_{\\pi}$ , allowing us to obtain a high-performance video understanding dynamic network controlled by an optimal system-status-aware policy on the current local device. For evaluation, we test the trained SAN on the same device, but with an unseen system status trajectory, which evaluates both the accuracy and the delay of our SAN.", + "bbox": [ + 498, + 537, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model meta-optimization and deployment for unseen device. In this scenario, our SAN is pre-trained and meta-optimized on a set of seen source devices, and we evaluate its efficacy when deployed to an unseen device. This evaluates the efficacy of our MSA design described in Sec. 3.2. At the start of training, we first initialize and pre-train our SAN on a seen device, following the process", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2We empirically find that we can simulate a large set of target platforms by setting different utility limitations on a few selected devices, as described in our experiments.", + "bbox": [ + 76, + 862, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10518", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "described previously. Then, during meta-optimization, we freeze the main network $M$ and only optimize our agent $\\pi$ . During meta-optimization iterations, we randomly sample video clips from the training set for meta-training and meta-testing. For better generalization, video clips used for meta-training are not used for meta-testing. When deploying our SAN to an unseen target device, we first fine-tune the agent $\\pi$ via the auxiliary task of delay prediction. The adapted agent is then tested for average accuracy and delay over all frames in the testing set. When comparing our method's average accuracy and delay with others, we compute the global accuracy and delay (counting both the fine-tuning and final testing phase) for a fair comparison.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 301, + 209, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We describe our setup and results below. More analysis and details are provided in Supplementary.", + "bbox": [ + 75, + 325, + 468, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAN's architecture. For our dynamic encoder, we adopt the fully-convolutional backbone ResNet50 [14] to handle various input resolutions. We add three exit ports (i.e., $n = 3$ ) at the end of these specific layers of ResNet50 ( $\\{conv3_x, conv4_x, conv5_x\\}$ [14]) to achieve dynamic depth. Moreover, we specifically design a convolutional layer for each resolution-depth pair to unify the shapes of the encoder's output features (details in Supplementary). Global average pooling is performed on the unified encoder's output and send to the LSTM to update its hidden state $h \\in \\mathbb{R}^{2048}$ . The output of the LSTM is fed into the task head for online video understanding. On the other hand, our lightweight agent $\\pi$ consists of three fully connected layers. We utilize three widely used monitors [43] to collect system information, which contains the status of system's CPU, GPU, memory, disk, I/O and fan. For model training, we set $\\lambda_{acc} = 2$ , $d_b = 0.03$ , $\\alpha = 1e^{-3}$ and $\\beta = 1e^{-5}$ . More details of our SAN are presented in Supplementary.", + "bbox": [ + 75, + 357, + 472, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Online video task settings. We compare the performance of our SAN with the prior methods on two widely used online video understanding tasks: 1) Online action recognition: We follow the previous work [30] to use the 50Salads dataset [34], since it contains very long videos that are suitable for online video action recognition. Here, we consider three (i.e., $m = 3$ ) resolution candidates: [112, 168, 224] and use a fully-connected layer as the task head. We compare our method with SOTA online action recognition methods, including RA [30], DDLSTM [31], AR-Net [25], and LiteEval [40]. To test models on the same platform for a fair comparison, we use their publicly released code (AR-Net) when available and implement the rest (RA, DDLSTM and LiteEval) on our own. 2) Online pose estimation: We follow the existing video pose estimation works [6, 12] to conduct our experiments on the Sub-JHMDB dataset [19], a collection of 11,200 frames from 316 video clips, labeled with 15 body", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bfe88da5c1cc5de6cb4798a6b1d6305f41a58e1d631ad283a9601253dca6dc1b.jpg", + "table_caption": [ + "Table 1. Results on 50Salads dataset for online action recognition." + ], + "table_footnote": [], + "table_body": "
MethodAccuracyMax DelayMean DelayR\\T Frames
RA [30]42.6 %76.6 ms33.9 ms79.1 %
DDLSTM [31]41.1 %407.4 ms110.3 ms5.9 %
LiteEval [40]40.3 %110.7 ms68.9 ms30.4 %
AR-Net [25]40.9 %91.1 ms56.2 ms62.1 %
main network + random policy46.2%80.5 ms34.1 ms92.1%
+ stream-aware55.4 %80.8 ms40.3 ms82.0%
+ SAN53.8%49.4 ms29.9 ms95.4%
", + "bbox": [ + 501, + 103, + 888, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f978f910c782b02b1d984314c6dca1a72f226867e5acfc2d996196953c52f56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 215, + 699, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea8ecf390824b23a0fe233472130cdc99508ce257a346d3abe3de93c976689f5.jpg", + "image_caption": [ + "Figure 2. A trace of models' accuracy and delay for online video-based action recognition (left) and pose estimation (right). In the areas masked by the gray color, we can observe a rapid increase of the baseline model's delay (see the second row), which corresponds to the high-load status of the system." + ], + "image_footnote": [], + "bbox": [ + 500, + 340, + 699, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3657466d63a82059928f570e2e1b2d95081a753e2646908fc6a481b43ebb7d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 215, + 893, + 339 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5bf85d570bbd53c2484eb1bec05d789a25c635e9680adf339f5f62a8f7053fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 339, + 893, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "joints. Here, we consider three (i.e., $m = 3$ ) resolution candidates: [128, 192, 256] and follow the work of [41] to use 3 deconvolutional layers to build the task head for pose estimation. Moreover, we replace LSTM with a ConvLSTM [33] to accommodate for 2D feature maps. We compare our method with the existing dynamic pose estimation networks, i.e., DKD [29] and Skip-Convolution [12]. Similarly, we re-implement DKD and Skip-Convolutions as there is no publicly available code.", + "bbox": [ + 496, + 491, + 890, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment settings. To build dynamic systems with varying system status, we design multiple background processes (i.e., matrix calculators, video compressors, and large deep learning models) to occupy various amounts of computational resources. We can then generate dynamic system load trajectories using these processes to simulate dynamic system load environments. During model training, we randomly generate dynamic system trajectories in each iteration to simulate various dynamic environments to train our SAN. During testing, for fair comparisons, we first generate 3 dynamic trajectories unseen in training, and then run all the evaluation experiments on these 3 fixed trajectories and report the average result. To evaluate the models' efficiency, we measure the maximum delay and mean delay for each model. Moreover, as the videos of both datasets [19, 34] are at 30 fps, we consider delay time $< 30$ ms as real-time, and measure the percentage of such real-time predictions (R\\T Frames).", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10519", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/313b31f77a8e654d60fcc0e9ee13886f5a5a4a1944bd5b41be74c015e05dc295.jpg", + "table_caption": [ + "Table 2. Results on Sub-JHMDB dataset for online pose estimation." + ], + "table_footnote": [], + "table_body": "
MethodHeadSho.Elb.Wri.HipKneeAnk.AvgMax DelayMean DelayR\\T Frames
DKD [29]98.396.690.487.199.196.092,994.0133.8 ms47.9 ms59.9 %
SimpleBaseline [41]97.597.891.186.099.696.892.694.4129.6 ms45.5 ms63.8 %
Skip-Convolution [12]98.797.792.088.199.396.691.095.1115.7 ms51.2 ms53.7 %
main network + random policy97.595.291.493.577.786.872.486.9130.9 ms36.1 ms72.9%
main network + stream-aware99.498.697.598.193.994.290.095.3126.1 ms44.7 ms63.7%
main network + SAN99.198.797.398.290.192.585.093.441.2 ms33.1 ms81.4%
", + "bbox": [ + 78, + 103, + 888, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Evaluating SAN on seen devices", + "text_level": 1, + "bbox": [ + 76, + 212, + 357, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we compare our SAN with the SOTA methods for both tasks and report the results in Tab. 1 and Tab. 2. In both tables, we also report the performance of two baselines of our method: main network + random policy where actions are randomly selected; and main network + stream-aware where the agent $\\pi$ uses information from the input stream without considering system status. Both baselines are not system-status-aware.", + "bbox": [ + 75, + 234, + 470, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Online action recognition. Tab. 1 presents the results on the 50Salads dataset for online action recognition. We compare our method with four SOTA methods, including two static models (i.e., RA [30] and DDLSTM [31]) and two efficient dynamic models (i.e., AR-Net [25] and LiteEval [40]). Our SAN outperforms these methods on all metrics, achieving a higher accuracy while maintaining a low mean delay (29.9ms) and high percentage of real-time frames (95.4%). These improvements are due to our system-status-aware design of SAN, allowing it to make on-the-fly adaptations to process a higher percentage of frames in real-time. The two baselines are not system-status-aware and therefore perform significantly worse on the delay-related metrics, especially on max delay (80.5ms and 80.8ms vs our 49.4ms). Compared to the random policy, stream-aware policy significantly improves the accuracy at the cost of reduced real-time performance. Overall, our SAN achieves superior performance and at the same time also decreases the delay.", + "bbox": [ + 75, + 356, + 468, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Online pose estimation. Tab. 2 presents the results on the Sub-JHMDB dataset for online pose estimation. We compare our SAN with a static network SimpleBaseline [41] and two dynamic efficient networks: DKD [29] and Skip-Convolution [12]. As shown in Tab. 2, our main network + stream-aware baseline can achieve the SOTA performance but may still incur large delays occasionally. With the help of our system-status-aware design, our SAN can control the maximum and mean delay at a low level while still maintaining high accuracy.", + "bbox": [ + 75, + 628, + 468, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis. We analyze the effect of the system load dynamics and show results from test-time runs on a laptop with a NVIDIA Geforce GTX 1080 GPU (11 GB), where 10 background processes are randomly activated. Fig. 2 presents the models' frame-level delay and accuracy along with the system status. We compare our SAN with two baselines: main network + stream-aware and main network + random policy. In the first row of Fig. 2,", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we observe that the GPU load fluctuates over time, varying the available computational resources for the online models. As a result, the delay of models with random/stream-aware policy fluctuates up to around 100ms as shown in the second row. In contrast, by using the system-status information to control the model, our SAN is able to stabilize the delay and avoid large delays amid the dynamic system. Moreover, as shown in the last row, with the help of the stream-aware design, our SAN can still obtain a similar accuracy to the stream-aware model. These results show the advantage of our system-status-aware and stream-aware designs, which can simultaneously control the real-time delay and attain high accuracy for online video understanding tasks.", + "bbox": [ + 496, + 212, + 893, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluating SAN with MSA on unseen devices", + "text_level": 1, + "bbox": [ + 498, + 421, + 883, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed MSA is introduced to achieve effective adaptation to the unseen host devices. Here, we compare our MSA with three baselines: 1) Direct transfer: Directly deploying our SAN to the unseen device. 2) Feature alignment: Minimizing the domain gap between the source device and target device via a domain discriminator [8, 26]. 3) Self-supervised fine-tune: Using our auxiliary task (delay prediction) to fine tune the model without meta optimization. To further demonstrate the efficacy of MSA, we also present the upper bound, dubbed as Fully-supervised learning, where we directly train the agent on the target environment, which eliminates the need for adaptation.", + "bbox": [ + 496, + 446, + 890, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we collect three hardware platforms for training and testing: 1) Device a: a laptop with an AMD Ryzen Threadripper 2950X and an NVIDIA Geforce GTX 1080 GPU (11 GB). 2) Device b: a desktop with an Intel(R) Core(TM) i9-10900 and an NVIDIA Geforce RTX 2080 GPU (12 GB). 3) Device c: a workstation with an AMD Ryzen Threadripper 3960X and an NVIDIA Geforce RTX 3090 GPU (24 GB). We randomly sample two platforms as the seen source devices for model initialization and meta-optimization and use the remaining one as the unseen target device for model adaptation (e.g., Source: $\\{a,b\\} \\to$ Target: $\\{c\\}$ ). Furthermore, we augment the source platforms by restricting the number of available CPU and the highest frequencies of CPU and GPU to achieve various environments with different resource configurations for better training. The results of policy adaptation are reported in Tab. 3 and 4. More results on other portable devices (e.g., NVIDIA Jetson TX2) are presented in Supplementary.", + "bbox": [ + 496, + 628, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10520", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/644dfd6605fbe926d8b592474a095f9b586ca0aabc2ec5c79525b11db0a5f9e9.jpg", + "table_caption": [ + "Table 3. Results of policy adaptation across platforms for online pose estimation. More results on other platforms are in Supplementary." + ], + "table_footnote": [], + "table_body": "
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T Frames
Fully-supervised learning (Upper bound)93.0 %28.3 ms87.4 %93.1 %24.7 ms84.9 %92.8 %28.1 ms74.5 %
Direct transfer87.3 %28.4 ms82.4 %85.7 %25.0 ms79.5 %90.9 %33.5 ms20.4 %
Feature alignment90.8 %29.0 ms80.4 %91.0 %27.2 ms74.6 %91.0 %31.8 ms64.5 %
Our self-supervised fine-tune91.1 %29.8 ms79.8 %85.8 %21.9 ms84.1 %91.5 %29.3 ms61.7 %
Our MSA92.8 %29.7 ms86.1 %92.9 %25.1 ms82.2 %92.4 %30.0 ms71.9 %
", + "bbox": [ + 78, + 104, + 888, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6c7026ece52415f9a1393e45f19ac5923ab183b6b47ec9d9d5c9718fdf07581.jpg", + "table_caption": [ + "Table 4. Results of policy adaptation across platforms for online action recognition." + ], + "table_footnote": [], + "table_body": "
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T Frames
Fully-supervised learning (Upper bound)53.8 %26.3 ms95.4 %52.6 %31.3 ms93.1 %52.1 %31.9 ms90.4 %
Direct transfer41.7 %28.1 ms77.3 %47.2 %34.6 ms84.7 %52.0 %41.1 ms40.6 %
Feature alignment50.1 %29.8 ms82.4 %50.3 %32.7 ms88.1 %51.7 %37.2 ms60.5 %
Our self-supervised fine-tune49.1 %26.5 ms93.8 %51.0 %32.2 ms93.1 %50.7 %33.7 ms87.2 %
Our MSA52.6 %26.8 ms94.1 %52.2 %31.6 ms92.7 %51.9 %32.1 ms90.2 %
", + "bbox": [ + 78, + 196, + 888, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 3 and 4, the direct transfer method fails to maintain the high accuracy and low delay simultaneously due to the hardware configuration shift between devices. On the other hand, by aligning the system-status-aware features between the source and target devices, the agent finds a suboptimal policy to balance the accuracy and delay. Using the auxiliary task to fine-tune the agent without meta optimization can achieve real-time predictions on the target device, but it sometimes fails to maintain high accuracy. Finally, our MSA simultaneously maintains a high ratio of real-time predictions and accurate predictions. Furthermore, the results are close to the upper bound of actually training on the target device, showing the effectiveness of our MSA to generalize to unseen devices.", + "bbox": [ + 75, + 291, + 472, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 513, + 230, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further analyze the characteristics of our dynamic resolution and dynamic depth mechanism, and investigate the performance of our agent under various delay thresholds $d_{b}$ and accuracy coefficients $\\lambda_{acc}$ .", + "bbox": [ + 75, + 537, + 468, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we evaluate the impact of the dynamic resolution and depth mechanism and perform ablation studies by allowing one of them to be dynamic while fixing the other. In Tab. 5, we observe that, by allowing for dynamic resolutions or dynamic depths, our main network can reduce the maximum and mean delay while maintaining high accuracy. Moreover, by combining these two mechanisms, our SAN can further reduce the mean delay and suppress the maximum delay at a lower level, with even better accuracy.", + "bbox": [ + 75, + 598, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we investigate the performance of our SAN under different delay thresholds $d_{b}$ and accuracy coefficients $\\lambda_{acc}$ . As shown in Tab. 6, under different values of $d_{b}$ , our SAN can still significantly decrease the maximum delay while maintaining high accuracy. Furthermore, as $d_{b}$ increases, SAN is allowed to have higher delays before getting penalized, leading to higher accuracy while slightly degrading delay metrics, which is what we expected. We also evaluate different values of $\\lambda_{acc}$ and present the results in Tab. 6. We observe that our SAN generally achieves good accuracy and delay performance for different values of $\\lambda_{acc}$ , showing that", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5084f73a13bcdf2dd6c55aa6cfc48141a99c914ceaa3e9ac341f0e68f1af6d3d.jpg", + "table_caption": [ + "Table 5. Ablation results for dynamic mechanism." + ], + "table_footnote": [], + "table_body": "
SettingAvgMax DelayMean Delay
res = 256, dep = 394.4 %129.6 ms45.5 ms
res = 256, dep = [1,3]92.6 %71.4 ms39.3 ms
res = 256, dep = [1,2,3]93.3 %63.6 ms34.5 ms
res = [128,256], dep = 391.2 %84.0 ms43.7 ms
res = [64,128,192,256], dep = 391.9 %61.9 ms38.0 ms
Our SAN93.4 %41.2 ms33.1 ms
", + "bbox": [ + 501, + 304, + 888, + 390 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e8ad647ade62484aac66c4be4c0c5eef19a2b8861e9a7f3e06c13c77a4059b2b.jpg", + "table_caption": [ + "Table 6. Ablation results for different values of $d_b$ and $\\lambda_{acc}$ ." + ], + "table_footnote": [], + "table_body": "
SettingAvgMax DelayMean Delay
db=10 ms92.4 %39.2 ms27.7 ms
=20 ms92.8 %39.4 ms30.7 ms
=30 ms93.4 %41.2 ms33.1 ms
=40 ms94.0 %43.5 ms41.9 ms
λacc=191.5 %41.0 ms30.8 ms
=293.4 %41.2 ms33.1 ms
=494.0 %44.6 ms37.1 ms
", + "bbox": [ + 501, + 417, + 888, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "it is robust to the $\\lambda_{acc}$ hyperparameter. Also, when $\\lambda_{acc}$ increases, the accuracy improves and delay slightly degrades, as we would expect.", + "bbox": [ + 498, + 532, + 890, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 619, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we tackle the efficiency problem of deep networks on devices with computational fluctuations. To the best of our knowledge, this is the first time this problem is being explicitly tackled for online video understanding tasks. We propose SAN, a novel system-status-aware adaptive network for low-delay online action recognition and online pose estimation tasks with a fluctuating computation budget. A self-supervised meta optimization framework (MSA) is also proposed for more effective adaptation between hardware platforms. Experiments show that our proposed SAN and MSA obtains SOTA performance with low delay even under a rapidly fluctuating system load.", + "bbox": [ + 496, + 638, + 893, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06).", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10521", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 1", + "[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 1", + "[3] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 2", + "[4] Mario Emad, Michael Ishack, Mohamed Ahmed, Mohamed Osama, Mohamed Salah, and Ghada Khoriba. Early-anomaly prediction in surveillance cameras for security applications. In 2021 International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC), pages 124-128. IEEE, 2021. 1", + "[5] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 1, 2", + "[6] Zhipeng Fan, Jun Liu, and Yao Wang. Motion adaptive pose estimation from compressed videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11719-11728, 2021. 1, 2, 6", + "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 2", + "[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 7", + "[9] giampaolo. psutil. https://github.com/giampaolo/psutil. Accessed: 2022-4-25. 3", + "[10]gpuopenanalytics. pynvml. https://github.com/gpuopenanalytics/pynvml. Accessed: 2022-4-25.3", + "[11] Pratik Gujjar and Richard Vaughan. Classifying pedestrian actions in advance using predicted video of urban driving scenes. In 2019 International Conference on Robotics and Automation (ICRA), pages 2097-2103. IEEE, 2019. 1", + "[12] Amirhossein Habibian, Davide Abati, Taco S Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2695–2704, 2021. 1, 2, 6, 7", + "[13] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In International Conference on Learning Representations, 2021. 4", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1, 2", + "[17] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 2", + "[18] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11719-11727, 2019. 2", + "[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6", + "[20] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 1", + "[21] Hayeon Lee, Sewoong Lee, Song Chong, and Sung Ju Hwang. Help: Hardware-adaptive efficient latency prediction for nas via meta-learning. arXiv preprint arXiv:2106.08630, 2021. 1", + "[22] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 2", + "[23] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 2", + "[24] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 1", + "[25] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 2, 6, 7", + "[26] Yu Mitsuzumi, Go Irie, Daiki Ikami, and Takashi Shibata. Generalized domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1084-1093, June 2021. 7", + "[27] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Maria Florina Balcan and Kil" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10522", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1928-1937, New York, New York, USA, 20-22 Jun 2016. PMLR. 2", + "[28] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 2", + "[29] Xuecheng Nie, Yuncheng Li, Linjie Luo, Ning Zhang, and Jiashi Feng. Dynamic kernel distillation for efficient pose estimation in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6942-6950, 2019. 2, 6, 7", + "[30] Toby Perrett and Dima Damen. Recurrent assistance: Cross-dataset training of lstms on kitchen tasks. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1354-1362. IEEE Computer Society, 2017. 6, 7", + "[31] T. Perrett and D. Damen. Ddlstm: Dual-domain lstm for cross-dataset action recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7844-7853, 2019. 6, 7", + "[32] Steven C Seow. Designing and engineering time: The psychology of time perception in software. Addison-Wesley Professional, 2008. 1", + "[33] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 6", + "[34] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 6", + "[35] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 4", + "[36] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 1", + "[37] Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han. Hat: Hardware-aware transformers for efficient natural language processing. In ACL, 2020. 2", + "[38] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the 17th Conference in Uncertainty in Artificial Intelligence, UAI '01, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 3", + "[39] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3", + "[40] Zuxuan Wu, Caiming Xiong, Yu-Gang Jiang, and Larry S Davis. Liteeval: A coarse-to-fine framework for resource" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "efficient video recognition. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 6, 7", + "[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 6, 7", + "[42] Lumin Xu, Yingda Guan, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Vipnas: Efficient video pose estimation via neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16072-16081, 2021. 2", + "[43] XuehaiPan. nvitop. https://github.com/XuehaiPan/nvitop. Accessed: 2022-4-25. 3, 6", + "[44] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 1", + "[45] Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization: Towards lossless cnns with low-precision weights. In International Conference on Learning Representations, ICLR2017, 2017. 2", + "[46] Mohammadreza Zolfaghari, Kamaljeet Singh, and Thomas Brox. Eco: Efficient convolutional network for online video understanding. In Proceedings of the European conference on computer vision (ECCV), pages 695-712, 2018. 2" + ], + "bbox": [ + 501, + 90, + 893, + 501 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10523", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_model.json b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79dc165a4426ed981f9829da6221d8007aec6cf6 --- /dev/null +++ b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.13, + 0.828, + 0.177 + ], + "angle": 0, + "content": "System-status-aware Adaptive Network for Online Streaming Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.203, + 0.721, + 0.221 + ], + "angle": 0, + "content": "Lin Geng Foo\\(^{1\\dagger}\\) Jia Gong\\(^{1\\dagger}\\) Zhipeng Fan\\(^{2§}\\) Jun Liu\\(^{1‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.222, + 0.78, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Singapore University of Technology and Design \\(^{2}\\)New York University" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.242, + 0.815, + 0.257 + ], + "angle": 0, + "content": "{lingeng_foo, jia_gong}@mypail.sutd.edu.sg, zf606@nyu.edu,jun_riu@sutd.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.473, + 0.642 + ], + "angle": 0, + "content": "Recent years have witnessed great progress in deep neural networks for real-time applications. However, most existing works do not explicitly consider the general case where the device's state and the available resources fluctuate over time, and none of them investigate or address the impact of varying computational resources for online video understanding tasks. This paper proposes a System-status-aware Adaptive Network (SAN) that considers the device's real-time state to provide high-quality predictions with low delay. Usage of our agent's policy improves efficiency and robustness to fluctuations of the system status. On two widely used video understanding tasks, SAN obtains state-of-the-art performance while constantly keeping processing delays low. Moreover, training such an agent on various types of hardware configurations is not easy as the labeled training data might not be available, or can be computationally prohibitive. To address this challenging problem, we propose a Meta Self-supervised Adaptation (MSA) method that adapts the agent's policy to new hardware configurations at test-time, allowing for easy deployment of the model onto other unseen hardware platforms." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.671, + 0.21, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Online video understanding, where certain predictions are immediately made for each video frame by using information in the current frame and potentially past frames, is an important task right at the intersection of video-based research and practical vision applications (e.g., self-driving vehicles [11], security surveillance [4], streaming services [32], and human-computer interactions [20]). In particular, in many of these real-world video-based applications, a fast and timely response is often crucial to ensure high usability and reduce potential security risk. Therefore, in many practical online applications, it is essential to ensure that the model is working with low delay while maintaining a good" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.293, + 0.892, + 0.323 + ], + "angle": 0, + "content": "performance, which can be challenging for many existing deep neural networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.895, + 0.747 + ], + "angle": 0, + "content": "Recently, much effort has been made to reduce the delay of deep neural networks, including research into efficient network design [16, 36, 44], input-aware dynamic networks [5,6,12,24], and latency-constrained neural architectures [1, 2, 21]. However, all these works do not explicitly consider the dynamic conditions of the hardware platform, and assume stable computation resources are readily available. In practical scenarios, the accessible computing resources of the host devices can be fluctuating and dynamic due to the fact that multiple computationally expensive yet important threads are running concurrently. For example, in addition to performing vision-related tasks such as object detection, human activity recognition, and pose estimation, state-of-the-art robotic systems usually need to simultaneously perform additional tasks like simultaneous localization and mapping (SLAM) to successfully interact with humans and the environment. Those tasks are also often computationally heavy and could compete with vision tasks for computing resources. As a result, at times when the host device is busy with other processes, conducting inference for each model might require significantly more time than usual, leading to extremely long delays, which could cause safety issues and lagging responses in many real-world applications. Therefore, the study and development of models providing reliable yet timely responses under various hardware devices and fluctuating computing resources is crucially important. Unfortunately, such studies are lacking in the field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To achieve and maintain low delay for online video understanding tasks under a dynamic computing resource budget, we propose a novel System-status-aware Adaptive Network (SAN). Different from previous works, SAN explicitly considers the system status of its host device to make on-the-fly adjustments to its computational complexity, and is thus capable of processing video streams effectively and efficiently in a dynamic system environment. SAN comprises of two components: a) a simple yet effective dynamic main module that offers reliable predictions under various" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.461, + 0.902 + ], + "angle": 0, + "content": "† Equal contribution; § Currently at Meta; ‡ Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "network depths and input resolutions; b) a lightweight agent that learns a dynamic system-status-aware policy used to control the execution of the main module, which facilitates adaptation to the fluctuating system load. With the adaptivity of the main module and the control policy generated by the agent, our SAN can achieve good performance on the online video understanding task while maintaining a low delay under fluctuating system loads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.454 + ], + "angle": 0, + "content": "In various applications, we may need to deploy SAN onto different hardware platforms for online video understanding. However, it is inconvenient to train SAN for each hardware platform, and it might also be difficult to find adequate storage to load the large labeled dataset on all platforms (e.g., mobile devices). In light of these difficulties, we further propose a method for deployment-time self-supervised agent adaptation, which we call Meta Self-supervised Adaptation (MSA). With MSA, we can conveniently train a SAN model on a set of local platforms, and perform a quick deployment-time agent adaptation on a target device, without the need for the original labeled training data. Specifically, our proposed MSA introduces an auxiliary task of delay prediction together with a meta-learning procedure, that facilitates the adaptation to the target deployment device." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.455, + 0.449, + 0.47 + ], + "angle": 0, + "content": "In summary, the main contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.479, + 0.47, + 0.57 + ], + "angle": 0, + "content": "- We are the first to explicitly consider the fluctuating system status of the hardware device at inference time for online video understanding. To address this, we propose SAN, a novel system-status-aware network that adapts its behavior according to the video stream and the real-time status of the host system." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.58, + 0.47, + 0.656 + ], + "angle": 0, + "content": "- We further propose a novel Meta Self-supervised Adaptation method MSA that alleviates the training burden and allows our model to effectively adapt to new host devices with potentially unclear computation profiles at deployment time." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.665, + 0.47, + 0.756 + ], + "angle": 0, + "content": "- We empirically demonstrate that our proposed method achieves promising performance on the challenging online action recognition and pose estimation tasks, where we achieve low delays under a rapidly fluctuating system load without jeopardizing the quality of the predictions." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.479, + 0.47, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.22, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Online Video Understanding. Recently, motivated by the increasing real-world demand, a lot of works [5,6,29,40, 46] have attempted to improve the accuracy and efficiency of models for online video understanding tasks, such as online action recognition [23, 40, 46] and online pose estimation [6, 12, 29]. Several works design efficient networks to improve models' efficiency for online video tasks, including" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.426 + ], + "angle": 0, + "content": "the efficient 3D CNN design [46], the Temporal Shift 2D CNN network [23] and the Skip-Convolution [12], while other researchers introduce adaptivity into the networks, termed as stream-aware dynamic networks [5, 25, 29] to further reduce the network's computation complexity while maintaining a good performance. For example, LiteEval [40] dynamically chooses between a coarse model and a fine model for each frame and MAPN [6] dynamically activates its encoder to save computation resources. However, all these existing methods do not explicitly consider the dynamic conditions of the hardware platform, and thus face limitations in the face of fluctuating system conditions. For example, in situations when the system is under high computational load, the model may still select to execute the branch with high complexity and therefore incur significantly longer delays, which is sub-optimal for time-sensitive applications. Compared to these methods, our proposed SAN is system-status-aware and thus more robust to computational fluctuations on the models' host devices. To the best of our knowledge, such a system-status-aware adaptive network that can efficiently tackle online video understanding tasks has not been explored before." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.895, + 0.686 + ], + "angle": 0, + "content": "Efficient Network Designs. There are several paradigms for efficient deep architectures, and we list a few here. Efficient deep models such as SqueezeNet [17] and MobileNet [16] introduce novel computation operators using fewer parameters, while Neural Network Quantization [3,45] methods effectively shrink the model size of the existing models by quantizing model parameters to isolated values. Teacher-student models [15] are also effective in distilling knowledge from larger models into smaller ones. Some other latency-constrained networks [37, 42] further conduct neural architecture search (NAS) to search for an efficient architecture. Different from all these approaches, our SAN explicitly accounts for the dynamic system status that is constantly changing on the device over time, allowing us to perform efficient online video understanding even with a fluctuating system status." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Reinforcement Learning (RL). RL methods [22,27,28] train agents to navigate environments such that the reward is maximized. To allow pre-trained RL agents to adapt to a new environment in fewer episodes, methods based on the few-shot setting [7, 18] have been developed. Differently, we aim to allow our RL agent to adapt at deployment time in a self-supervised manner (i.e., no reward signals can be provided due to the absence of labeled video data). We introduce MSA, a novel method for self-supervised deployment-time adaptation. MSA employs an auxiliary task of delay prediction, and performs meta-learning to allow the RL agent to effectively adapt its control policy to the target hardware platform via fine-tuning on the auxiliary task." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10515" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.169, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.314 + ], + "angle": 0, + "content": "Given the live video stream \\( V = \\{I_1, \\dots, I_t, \\dots\\} \\) and a seen/unseen host device whose system status \\( sys_t \\) varies over time, our goal is to build a model that dynamically adjusts the computational complexity of inference based on not only the streaming input, but also the system status to make high-quality predictions with a low delay at each step \\( t \\). To this end, our proposed model: 1) takes into account the available computational resources to make on-the-fly decisions to dynamically adjust the computation consumption of the model, therefore maintaining the delay at low values while keeping the prediction's accuracy; and 2) effectively adapts itself to unseen devices in a self-supervised manner while still keeping its performance during deployment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.319, + 0.466, + 0.336 + ], + "angle": 0, + "content": "3.1. SAN: System-status-aware Adaptive Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.448 + ], + "angle": 0, + "content": "For better online performance amid the host device's varying background computational load, we propose a System-status-aware Adaptive Network (SAN). As shown in Fig. 1, our SAN consists of two main components: a lightweight agent \\(\\pi\\) that decides how to process the incoming frame and a dynamic main network \\(M\\) that processes the frame with the specified policy generated by \\(\\pi\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Dynamic main network. Our dynamic main network \\(M\\) flexibly produces frame-wise predictions using an architecture determined by the policy from the agent module \\(\\pi\\). To accomplish this without extra burden on the hardware system, we design a dynamic encoder that has both dynamic depth and dynamic resolution. Our dynamic depth mechanism gives the encoder the option of producing output features at shallower intermediate layers without having to wait for all layers to be executed. This reduces computational cost by executing a \"simpler\" sub-network. On the other hand, the dynamic resolution mechanism allows our encoder to selectively take in input images at a lower resolution. Crucially, this reduces the amount of pixels to be examined by the network, effectively reducing the amount of convolution operations. We implement both of the aforementioned mechanisms in a single network, where to enable dynamic execution depth of the network, we add early exits to the network which allow the execution to end at an intermediate layer, while to enable dynamic resolution, we leverage the fully convolutional network and add a dynamic global pooling operation at the end." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 1, our dynamic encoder selects from \\( m \\) input resolutions \\( \\{res^i\\}_{i=1}^m \\) and \\( n \\) model depths \\( \\{dep^j\\}_{j=1}^n \\), allowing for as many as \\( m \\times n \\) different options in total. These \\( m \\times n \\) dynamic options have different computation complexities and levels of performance. For online video understanding, we construct our dynamic encoder as a 2D CNN for handling each frame, and to model the temporal aspect, the output features from the dynamic encoder are fed into a Long Short Term Memory network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "(LSTM) at every step. The LSTM updates its hidden state \\( h_t \\) based on the features of the incoming frame \\( I_t \\) and its previous state \\( h_{t-1} \\), which is then used by the task head to determine the final prediction \\( \\hat{y}_t \\) for the video task following: \\( \\hat{y}_t = M(I_t, rest_t, dept_t, h_{t-1}; \\theta) \\), where \\( \\theta \\) refers to the parameters of the main network \\( M \\), while \\( rest_t \\) and \\( dept_t \\) are selected by the agent introduced next. More details of our main dynamic network are in Sec. 4 and Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.893, + 0.426 + ], + "angle": 0, + "content": "RL-based agent. The RL-based agent \\(\\pi\\) controls the dynamic main network \\(M\\) by generating a frame-level processing policy \\((\\text{rest}_t, \\text{depth}_t)\\) at each time \\(t\\), which aims to maintain the task-related accuracy \\(\\text{acc}_t\\) at high level with low delay \\(d_t\\). To make decisions amid the fluctuating system loads, our lightweight agent \\(\\pi\\) should be both system-status-aware and streaming input-aware. We model our task as a Markov Decision Process, where the agent \\(\\pi\\) takes an action \\(a_t\\) at every step \\(t\\), based on the observed state \\(s_t\\), and transferred to the next state \\(s_{t+1}\\). Crucially, state \\(s_t\\) includes system status information \\(\\text{sys}_t\\) and input stream information \\(h_{t-1}\\), making \\(\\pi\\) both system status and stream aware. To learn an optimal policy, we propose an RL-based method to train our agent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Specifically, we set the observed states of the agent as \\( s_t = [h_{t-1}, sys_t, g_t] \\), where \\( sys_t \\) is the system status (e.g., the CPU and GPU utility)\\(^1\\), \\( h_{t-1} \\) is the LSTM hidden state, and \\( g_t \\) refers to other useful information (i.e., the previous step's action \\( a_{t-1} \\) and delay \\( d_{t-1} \\)). Then, based on the available resolutions \\( \\{res^i\\}_{i=1}^m \\) and depths \\( \\{dep^j\\}_{j=1}^n \\), we build the action space as \\( \\{a^{i,j} = (res^i, dep^j)\\}_{i=1,j=1}^{m,n} \\). As shown in Fig. 1, at time \\( t \\), the agent \\( \\pi \\) receives \\( h_{t-1} \\), \\( sys_t \\) and \\( g_t \\), and generates a probability distribution over the action space:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.59, + 0.892, + 0.611 + ], + "angle": 0, + "content": "\\[\nP _ {t} = \\left[ p _ {t} ^ {1, 1}, p _ {t} ^ {1, 2}, \\dots , p _ {t} ^ {m, n} \\right] = \\pi \\left(h _ {t - 1}, s y s _ {t}, g _ {t}; \\phi\\right) \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.713 + ], + "angle": 0, + "content": "where \\( p_t^{i,j} \\in [0,1] \\) is the probability of the action \\( a^{i,j} \\) at time \\( t \\), \\( P_t \\in [0,1]^{m\\cdot n} \\) represents the probability distribution over the action space, and \\( \\phi \\) refers to the parameters of the agent \\( \\pi \\). Then, we take the highest probability entry in \\( P_t \\) (denoted as \\( p_t^{i*,j*} \\)), select the corresponding action \\( a_t = (res^{i*}, dep^{j*}) \\) and execute the main network accordingly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Training of SAN. To train our main network \\(M\\), we apply the loss \\(\\mathcal{L}_M\\) based on the specified online video-processing task. For example, for online action recognition, \\(\\mathcal{L}_M\\) can be the cross-entropy loss, while for online pose estimation, \\(\\mathcal{L}_M\\) can be the Mean Squared Error (MSE) loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.851 + ], + "angle": 0, + "content": "To control the dynamic encoder to achieve a good accuracy at low delay, we apply an RL-based loss \\(\\mathcal{L}_{\\pi}\\) to train our agent \\(\\pi\\). Specifically, we adopt a policy gradient loss [38, 39]. To maximize the accuracy \\(acc_{t}\\) while reducing the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1We also collect other system parameters (e.g., the temperature of each processor, the memory usage and the status of I/O) via three linux utilities [9, 10, 43] to form system status syst. More details are in Supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10516" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.096, + 0.803, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.354 + ], + "angle": 0, + "content": "Figure 1. Overview of our SAN framework, which consists of two modules: the agent module \\(\\pi\\) and the dynamic main network \\(M\\). For each incoming frame \\(I_{t}\\), the agent module first takes in the LSTM hidden state \\(h_{t-1}\\), system-status \\(sys_{t}\\), and other useful information \\(g_{t}\\) and then decides the input resolution \\(rest_{t}\\) and the depth \\(depth_{t}\\) of \\(M\\) via the policy head. After that, the main network \\(M\\) processes the frame according to these decisions. By controlling the behavior of the main network \\(M\\) via the agent \\(\\pi\\), our SAN can output high quality and low delay predictions in the dynamic system environment. Note that the auxiliary task head of the agent is used for fine-tuning of the agent during deployment, calibrating the learned system-status-aware features to adapt to the unseen target device." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.423, + 0.381 + ], + "angle": 0, + "content": "delay \\( d_{t} \\), we set the reward of the agent at time \\( t \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.386, + 0.47, + 0.42 + ], + "angle": 0, + "content": "\\[\nr _ {t} = \\lambda_ {a c c} a c c _ {t} \\left(a _ {t}, I _ {t}\\right) - \\max \\left(d _ {t} \\left(a _ {t}, s y s _ {t}\\right) - d _ {b}, 0\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.515 + ], + "angle": 0, + "content": "where \\(acc_{t}(a_{t},I_{t})\\) and \\(d_t(a_t,sys_t)\\) are the accuracy and delay of our main network under the action \\(a_{t}\\), the current frame \\(I_{t}\\), and system status \\(sys_{t}\\). \\(\\lambda_{acc}\\) is the coefficient of the accuracy term, and \\(d_b\\) is a tolerance threshold for the delay (e.g., \\(d_b = 30\\) ms for a rate of 30 fps), which is the acceptable delay. The policy gradient loss \\(\\mathcal{L}_{\\pi}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.521, + 0.47, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\pi} = \\sum_ {t = 1} ^ {T} - r _ {t} \\log \\left(p _ {t}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.598 + ], + "angle": 0, + "content": "where \\( p_t \\in [0,1] \\) is the probability of selecting the action \\( a_t \\) at time \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.416, + 0.623 + ], + "angle": 0, + "content": "3.2. MSA: Meta Self-supervised Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Ideally, we want to deploy SAN onto many different hardware platforms for online video applications, requiring us to train SAN for each of those platforms, which is labor intensive, computationally prohibitive, and may also be infeasible at times due to storage concerns (e.g., mobile devices). Instead, a more practical option would be to effectively fine-tune a pre-trained SAN at deployment-time on a target device. However, during deployment, the ground-truth labels associated with the real-time video stream are often inaccessible. To address this issue, we introduce an auxiliary task [13, 35] of delay prediction, where the agent \\(\\pi\\) is expected to additionally predict the actual processing time based on the selected action (i.e., the input resolution and execution depth of main network \\(M\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our intuition is that, on a new (unseen) platform, our agent \\(\\pi\\) lacks understanding of hardware status of this platform, thus we need a way for the agent to quickly understand it. Hence, we introduce the auxiliary task of delay" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.547 + ], + "angle": 0, + "content": "prediction, because delay times are directly related to the system status and the model's computational cost. Therefore, when the agent \\(\\pi\\) is fine-tuned using delay prediction, it will be pushed to adopt a better understanding of the host platform's characteristics and the costs of different options of \\(M\\) on that platform. Moreover, delay times can be directly observed, and can be a convenient supervision signal on a new platform since we do not have signals from our main task (which requires labeled video data). Therefore, on a new platform, we can fine-tune our agent to predict the delay by using the observed delay times as supervision, which will improve the agent's understanding of the device." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.762 + ], + "angle": 0, + "content": "Although the delay prediction task improves the quality of the intermediate features (which are also used to select actions), fine-tuning with it does not necessarily lead to a better policy on the host device. Hence, to better align the performance on the auxiliary task of delay prediction with policy learning, we further propose a meta-optimization-based learning framework to find a good initialization for the agent, such that updating based on the loss of delay prediction would also improve the quality of the policy. Combining the auxiliary task and the meta-learning framework allows us to train SAN on a set of local hardware platforms, and perform only a quick deployment-time adaptation on the target device at run-time, without requiring the actual labeled video dataset on that target device." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Auxiliary task of delay prediction. When SAN is deployed to an unseen target device, the processing times of each action at a given system status will shift significantly due to the drift in computing profiles. Therefore, fine-tuning on the target device is required to adapt SAN to the new hardware characteristics. Due to the lack of ground-truth labels for the streaming videos at run-time, we introduce a self-supervised auxiliary task of delay prediction for adapting SAN to unseen hardware platforms. The delay predic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "tion task calibrates the learned system-status-aware features for the agent module \\(\\pi\\) using supervisory signals (processing times) that can be generated on-the-fly with only a simple timer on the device." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.152, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Specifically, an auxiliary task head is added to the agent \\(\\pi\\), as shown in Fig. 1. The auxiliary task head takes in the features from the agent as well as the chosen action \\(a_{t}\\), to predict the delay time \\(d_{t}\\). To train our model on the auxiliary task, we apply a mean-squared-error loss between our predictions \\(\\hat{d}_t\\) and the observed delay \\(d_{t}\\). The auxiliary loss \\(\\mathcal{L}_{aux}\\) at step \\(t\\), is thus defined as: \\(\\mathcal{L}_{aux,t} = (d_t - \\hat{d}_t)^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.468, + 0.362 + ], + "angle": 0, + "content": "At deployment time, we fine-tune the agent \\(\\pi\\) on the target device using this auxiliary task only, allowing the agent to better understand the target device's characteristics and processing times. However, this does not necessarily mean that the agent has optimized its policy output – in order to achieve that, we further introduce a meta-optimization method as described next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.468, + 0.484 + ], + "angle": 0, + "content": "Meta-optimization for self-supervised agent adaptation. During deployment onto the target device, optimizing the auxiliary task of delay prediction is loosely connected to improving the actual policy. To better bind the quality of the developed policy to the delay prediction performance, we introduce a meta-optimization step to our agent \\(\\pi\\) during training, such that fine-tuning of the auxiliary task will also lead to a better adaptation of the policy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.468, + 0.68 + ], + "angle": 0, + "content": "We approach this problem by simulating the fine-tuning process using the delay prediction error on multiple target platforms, and (meta) optimize \\(\\phi\\) such that the update based on the delay prediction error leads to better adaptation of the policy as well. This will effectively link the quality of the policy to the delay prediction performance. Here, we adopt a meta-learning framework to tackle the problem, treating \\(\\phi\\) as the meta-parameters to meta-optimize. In this context, we denote the auxiliary task as the meta-train task and the agent's policy generation as the meta-test task. Our meta-optimization algorithm aims to learn a good initialization of \\(\\phi\\), such that when it is updated via the meta-train task, it can achieve good performance on the meta-test task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.468, + 0.831 + ], + "angle": 0, + "content": "We assume that our SAN is first trained on a source device \\(D_{src}\\), and then adapted to the additional \\(K\\) target platforms \\(\\{D_{tgt}^k\\}_{k=1}^K\\). After first training our SAN on the source device \\(D_{src}\\), we freeze the main network \\(M\\), and keep a copy of the agent's parameters (denoted as \\(\\phi_{meta,0}\\)) to initialize the meta-learning process. In the \\(u\\)-th meta-learning iteration, we first meta-train the agent for each target device \\(D_{tgt}^k\\) using its loss on the auxiliary task and perform a pseudo-update on the meta parameters \\(\\phi_{meta,u}\\) which can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.838, + 0.468, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\phi_ {t g t, u} ^ {k} \\leftarrow \\phi_ {\\text {m e t a}, u} - \\alpha \\nabla_ {\\phi_ {\\text {m e t a}, u}} \\mathcal {L} _ {\\text {a u x}, t} \\left(\\phi_ {\\text {m e t a}, u}, D _ {t g t} ^ {k}\\right) \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": "2We empirically find that we can simulate a large set of target platforms by setting different utility limitations on a few selected devices, as described in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.141 + ], + "angle": 0, + "content": "where \\(\\phi_{tgt}^k\\) refers to the parameters of the meta-trained agent \\(\\pi_{tgt}^k\\) which is about to be deployed onto the target device \\(D_{tgt}^k\\), and \\(\\alpha\\) is the learning rate for meta-training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.141, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Then, we check the performance of the pseudo-updated parameters on target device by meta-testing the performance of the pseudo-updated agent \\(\\pi_{tgt}^{k}(:,\\phi_{tgt,u}^{k})\\) for the agent's policy generation following:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.211, + 0.892, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m e t a}} ^ {k} \\left(\\phi_ {t g t, u} ^ {k}\\right) = \\mathcal {L} _ {\\pi} \\left(\\phi_ {t g t, u} ^ {k}, D _ {t g t} ^ {k}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.319 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{meta}^k\\) is the meta-testing-loss from the target hardware platform \\(D_{tgt}^k\\). Intuitively, this loss penalizes our agent based on its performance on a target device after a short period of fine-tuning. Hence, we update our meta parameters \\(\\phi_{meta,u}\\) in the \\(u\\)-th meta-learning iteration as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\phi_ {m e t a, u + 1} \\leftarrow \\phi_ {m e t a, u} - \\beta \\sum_ {k = 1} ^ {K} \\nabla_ {\\phi_ {m e t a, u}} \\mathcal {L} _ {m e t a} ^ {k} \\left(\\phi_ {t g t, u} ^ {k}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.476 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the learning rate for optimizing the meta parameters, and \\(\\phi_{tgt,u}^k\\) is obtained via the pseudo-update in Eq. 4. Through updating our model with this meta-learning scheme, our agent parameters will converge towards optimal parameters \\(\\phi_{meta,*}\\) that are able to effectively adapt to each target device through fine-tuning via the auxiliary task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.697, + 0.502 + ], + "angle": 0, + "content": "3.3. Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.509, + 0.89, + 0.538 + ], + "angle": 0, + "content": "As mentioned above, we consider two scenarios: deploying our SAN on a seen or unseen device." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.539, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Model training and testing for seen device. In this scenario, we evaluate SAN when it is trained and tested on the same device, which evaluates the effectiveness of our SAN design described in Sec. 3.1. For model training, we first utilize the dataset's training set to initialize our main network \\( M \\), which is controlled by a random policy (randomly sampling the resolution-depth pair) to improve its accuracy via the loss \\( \\mathcal{L}_M \\). Then, we train our agent \\( \\pi \\) using the loss \\( \\mathcal{L}_{\\pi} \\) with the parameters of the main network \\( (M) \\) fixed. Lastly, we fine-tune the main network \\( M \\) and agent \\( \\pi \\) jointly in a fully end-to-end manner with both \\( \\mathcal{L}_M \\) and \\( \\mathcal{L}_{\\pi} \\), allowing us to obtain a high-performance video understanding dynamic network controlled by an optimal system-status-aware policy on the current local device. For evaluation, we test the trained SAN on the same device, but with an unseen system status trajectory, which evaluates both the accuracy and the delay of our SAN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Model meta-optimization and deployment for unseen device. In this scenario, our SAN is pre-trained and meta-optimized on a set of seen source devices, and we evaluate its efficacy when deployed to an unseen device. This evaluates the efficacy of our MSA design described in Sec. 3.2. At the start of training, we first initialize and pre-train our SAN on a seen device, following the process" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10518" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "described previously. Then, during meta-optimization, we freeze the main network \\( M \\) and only optimize our agent \\( \\pi \\). During meta-optimization iterations, we randomly sample video clips from the training set for meta-training and meta-testing. For better generalization, video clips used for meta-training are not used for meta-testing. When deploying our SAN to an unseen target device, we first fine-tune the agent \\( \\pi \\) via the auxiliary task of delay prediction. The adapted agent is then tested for average accuracy and delay over all frames in the testing set. When comparing our method's average accuracy and delay with others, we compute the global accuracy and delay (counting both the fine-tuning and final testing phase) for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.302, + 0.21, + 0.319 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.358 + ], + "angle": 0, + "content": "We describe our setup and results below. More analysis and details are provided in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.473, + 0.629 + ], + "angle": 0, + "content": "SAN's architecture. For our dynamic encoder, we adopt the fully-convolutional backbone ResNet50 [14] to handle various input resolutions. We add three exit ports (i.e., \\( n = 3 \\)) at the end of these specific layers of ResNet50 (\\(\\{conv3_x, conv4_x, conv5_x\\}\\) [14]) to achieve dynamic depth. Moreover, we specifically design a convolutional layer for each resolution-depth pair to unify the shapes of the encoder's output features (details in Supplementary). Global average pooling is performed on the unified encoder's output and send to the LSTM to update its hidden state \\( h \\in \\mathbb{R}^{2048} \\). The output of the LSTM is fed into the task head for online video understanding. On the other hand, our lightweight agent \\( \\pi \\) consists of three fully connected layers. We utilize three widely used monitors [43] to collect system information, which contains the status of system's CPU, GPU, memory, disk, I/O and fan. For model training, we set \\( \\lambda_{acc} = 2 \\), \\( d_b = 0.03 \\), \\( \\alpha = 1e^{-3} \\) and \\( \\beta = 1e^{-5} \\). More details of our SAN are presented in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Online video task settings. We compare the performance of our SAN with the prior methods on two widely used online video understanding tasks: 1) Online action recognition: We follow the previous work [30] to use the 50Salads dataset [34], since it contains very long videos that are suitable for online video action recognition. Here, we consider three (i.e., \\( m = 3 \\)) resolution candidates: [112, 168, 224] and use a fully-connected layer as the task head. We compare our method with SOTA online action recognition methods, including RA [30], DDLSTM [31], AR-Net [25], and LiteEval [40]. To test models on the same platform for a fair comparison, we use their publicly released code (AR-Net) when available and implement the rest (RA, DDLSTM and LiteEval) on our own. 2) Online pose estimation: We follow the existing video pose estimation works [6, 12] to conduct our experiments on the Sub-JHMDB dataset [19], a collection of 11,200 frames from 316 video clips, labeled with 15 body" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.103 + ], + "angle": 0, + "content": "Table 1. Results on 50Salads dataset for online action recognition." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.104, + 0.889, + 0.213 + ], + "angle": 0, + "content": "
MethodAccuracyMax DelayMean DelayR\\T Frames
RA [30]42.6 %76.6 ms33.9 ms79.1 %
DDLSTM [31]41.1 %407.4 ms110.3 ms5.9 %
LiteEval [40]40.3 %110.7 ms68.9 ms30.4 %
AR-Net [25]40.9 %91.1 ms56.2 ms62.1 %
main network + random policy46.2%80.5 ms34.1 ms92.1%
+ stream-aware55.4 %80.8 ms40.3 ms82.0%
+ SAN53.8%49.4 ms29.9 ms95.4%
" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.216, + 0.7, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.341, + 0.7, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.217, + 0.894, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.34, + 0.894, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.41, + 0.893, + 0.48 + ], + "angle": 0, + "content": "Figure 2. A trace of models' accuracy and delay for online video-based action recognition (left) and pose estimation (right). In the areas masked by the gray color, we can observe a rapid increase of the baseline model's delay (see the second row), which corresponds to the high-load status of the system." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.629 + ], + "angle": 0, + "content": "joints. Here, we consider three (i.e., \\( m = 3 \\)) resolution candidates: [128, 192, 256] and follow the work of [41] to use 3 deconvolutional layers to build the task head for pose estimation. Moreover, we replace LSTM with a ConvLSTM [33] to accommodate for 2D feature maps. We compare our method with the existing dynamic pose estimation networks, i.e., DKD [29] and Skip-Convolution [12]. Similarly, we re-implement DKD and Skip-Convolutions as there is no publicly available code." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Experiment settings. To build dynamic systems with varying system status, we design multiple background processes (i.e., matrix calculators, video compressors, and large deep learning models) to occupy various amounts of computational resources. We can then generate dynamic system load trajectories using these processes to simulate dynamic system load environments. During model training, we randomly generate dynamic system trajectories in each iteration to simulate various dynamic environments to train our SAN. During testing, for fair comparisons, we first generate 3 dynamic trajectories unseen in training, and then run all the evaluation experiments on these 3 fixed trajectories and report the average result. To evaluate the models' efficiency, we measure the maximum delay and mean delay for each model. Moreover, as the videos of both datasets [19, 34] are at 30 fps, we consider delay time \\(< 30\\) ms as real-time, and measure the percentage of such real-time predictions (R\\T Frames)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10519" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.283, + 0.09, + 0.687, + 0.103 + ], + "angle": 0, + "content": "Table 2. Results on Sub-JHMDB dataset for online pose estimation." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.104, + 0.89, + 0.203 + ], + "angle": 0, + "content": "
MethodHeadSho.Elb.Wri.HipKneeAnk.AvgMax DelayMean DelayR\\T Frames
DKD [29]98.396.690.487.199.196.092,994.0133.8 ms47.9 ms59.9 %
SimpleBaseline [41]97.597.891.186.099.696.892.694.4129.6 ms45.5 ms63.8 %
Skip-Convolution [12]98.797.792.088.199.396.691.095.1115.7 ms51.2 ms53.7 %
main network + random policy97.595.291.493.577.786.872.486.9130.9 ms36.1 ms72.9%
main network + stream-aware99.498.697.598.193.994.290.095.3126.1 ms44.7 ms63.7%
main network + SAN99.198.797.398.290.192.585.093.441.2 ms33.1 ms81.4%
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.213, + 0.358, + 0.228 + ], + "angle": 0, + "content": "4.1. Evaluating SAN on seen devices" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.471, + 0.356 + ], + "angle": 0, + "content": "First, we compare our SAN with the SOTA methods for both tasks and report the results in Tab. 1 and Tab. 2. In both tables, we also report the performance of two baselines of our method: main network + random policy where actions are randomly selected; and main network + stream-aware where the agent \\(\\pi\\) uses information from the input stream without considering system status. Both baselines are not system-status-aware." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Online action recognition. Tab. 1 presents the results on the 50Salads dataset for online action recognition. We compare our method with four SOTA methods, including two static models (i.e., RA [30] and DDLSTM [31]) and two efficient dynamic models (i.e., AR-Net [25] and LiteEval [40]). Our SAN outperforms these methods on all metrics, achieving a higher accuracy while maintaining a low mean delay (29.9ms) and high percentage of real-time frames (95.4%). These improvements are due to our system-status-aware design of SAN, allowing it to make on-the-fly adaptations to process a higher percentage of frames in real-time. The two baselines are not system-status-aware and therefore perform significantly worse on the delay-related metrics, especially on max delay (80.5ms and 80.8ms vs our 49.4ms). Compared to the random policy, stream-aware policy significantly improves the accuracy at the cost of reduced real-time performance. Overall, our SAN achieves superior performance and at the same time also decreases the delay." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.781 + ], + "angle": 0, + "content": "Online pose estimation. Tab. 2 presents the results on the Sub-JHMDB dataset for online pose estimation. We compare our SAN with a static network SimpleBaseline [41] and two dynamic efficient networks: DKD [29] and Skip-Convolution [12]. As shown in Tab. 2, our main network + stream-aware baseline can achieve the SOTA performance but may still incur large delays occasionally. With the help of our system-status-aware design, our SAN can control the maximum and mean delay at a low level while still maintaining high accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Qualitative analysis. We analyze the effect of the system load dynamics and show results from test-time runs on a laptop with a NVIDIA Geforce GTX 1080 GPU (11 GB), where 10 background processes are randomly activated. Fig. 2 presents the models' frame-level delay and accuracy along with the system status. We compare our SAN with two baselines: main network + stream-aware and main network + random policy. In the first row of Fig. 2," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.895, + 0.41 + ], + "angle": 0, + "content": "we observe that the GPU load fluctuates over time, varying the available computational resources for the online models. As a result, the delay of models with random/stream-aware policy fluctuates up to around 100ms as shown in the second row. In contrast, by using the system-status information to control the model, our SAN is able to stabilize the delay and avoid large delays amid the dynamic system. Moreover, as shown in the last row, with the help of the stream-aware design, our SAN can still obtain a similar accuracy to the stream-aware model. These results show the advantage of our system-status-aware and stream-aware designs, which can simultaneously control the real-time delay and attain high accuracy for online video understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.884, + 0.438 + ], + "angle": 0, + "content": "4.2. Evaluating SAN with MSA on unseen devices" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Our proposed MSA is introduced to achieve effective adaptation to the unseen host devices. Here, we compare our MSA with three baselines: 1) Direct transfer: Directly deploying our SAN to the unseen device. 2) Feature alignment: Minimizing the domain gap between the source device and target device via a domain discriminator [8, 26]. 3) Self-supervised fine-tune: Using our auxiliary task (delay prediction) to fine tune the model without meta optimization. To further demonstrate the efficacy of MSA, we also present the upper bound, dubbed as Fully-supervised learning, where we directly train the agent on the target environment, which eliminates the need for adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Here, we collect three hardware platforms for training and testing: 1) Device a: a laptop with an AMD Ryzen Threadripper 2950X and an NVIDIA Geforce GTX 1080 GPU (11 GB). 2) Device b: a desktop with an Intel(R) Core(TM) i9-10900 and an NVIDIA Geforce RTX 2080 GPU (12 GB). 3) Device c: a workstation with an AMD Ryzen Threadripper 3960X and an NVIDIA Geforce RTX 3090 GPU (24 GB). We randomly sample two platforms as the seen source devices for model initialization and meta-optimization and use the remaining one as the unseen target device for model adaptation (e.g., Source: \\(\\{a,b\\} \\to\\) Target: \\(\\{c\\}\\)). Furthermore, we augment the source platforms by restricting the number of available CPU and the highest frequencies of CPU and GPU to achieve various environments with different resource configurations for better training. The results of policy adaptation are reported in Tab. 3 and 4. More results on other portable devices (e.g., NVIDIA Jetson TX2) are presented in Supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10520" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.09, + 0.885, + 0.104 + ], + "angle": 0, + "content": "Table 3. Results of policy adaptation across platforms for online pose estimation. More results on other platforms are in Supplementary." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.105, + 0.89, + 0.184 + ], + "angle": 0, + "content": "
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T Frames
Fully-supervised learning (Upper bound)93.0 %28.3 ms87.4 %93.1 %24.7 ms84.9 %92.8 %28.1 ms74.5 %
Direct transfer87.3 %28.4 ms82.4 %85.7 %25.0 ms79.5 %90.9 %33.5 ms20.4 %
Feature alignment90.8 %29.0 ms80.4 %91.0 %27.2 ms74.6 %91.0 %31.8 ms64.5 %
Our self-supervised fine-tune91.1 %29.8 ms79.8 %85.8 %21.9 ms84.1 %91.5 %29.3 ms61.7 %
Our MSA92.8 %29.7 ms86.1 %92.9 %25.1 ms82.2 %92.4 %30.0 ms71.9 %
" + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.185, + 0.731, + 0.197 + ], + "angle": 0, + "content": "Table 4. Results of policy adaptation across platforms for online action recognition." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.198, + 0.89, + 0.279 + ], + "angle": 0, + "content": "
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T Frames
Fully-supervised learning (Upper bound)53.8 %26.3 ms95.4 %52.6 %31.3 ms93.1 %52.1 %31.9 ms90.4 %
Direct transfer41.7 %28.1 ms77.3 %47.2 %34.6 ms84.7 %52.0 %41.1 ms40.6 %
Feature alignment50.1 %29.8 ms82.4 %50.3 %32.7 ms88.1 %51.7 %37.2 ms60.5 %
Our self-supervised fine-tune49.1 %26.5 ms93.8 %51.0 %32.2 ms93.1 %50.7 %33.7 ms87.2 %
Our MSA52.6 %26.8 ms94.1 %52.2 %31.6 ms92.7 %51.9 %32.1 ms90.2 %
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.504 + ], + "angle": 0, + "content": "As shown in Tab. 3 and 4, the direct transfer method fails to maintain the high accuracy and low delay simultaneously due to the hardware configuration shift between devices. On the other hand, by aligning the system-status-aware features between the source and target devices, the agent finds a suboptimal policy to balance the accuracy and delay. Using the auxiliary task to fine-tune the agent without meta optimization can achieve real-time predictions on the target device, but it sometimes fails to maintain high accuracy. Finally, our MSA simultaneously maintains a high ratio of real-time predictions and accurate predictions. Furthermore, the results are close to the upper bound of actually training on the target device, showing the effectiveness of our MSA to generalize to unseen devices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.232, + 0.531 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.599 + ], + "angle": 0, + "content": "We further analyze the characteristics of our dynamic resolution and dynamic depth mechanism, and investigate the performance of our agent under various delay thresholds \\( d_{b} \\) and accuracy coefficients \\( \\lambda_{acc} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.734 + ], + "angle": 0, + "content": "First, we evaluate the impact of the dynamic resolution and depth mechanism and perform ablation studies by allowing one of them to be dynamic while fixing the other. In Tab. 5, we observe that, by allowing for dynamic resolutions or dynamic depths, our main network can reduce the maximum and mean delay while maintaining high accuracy. Moreover, by combining these two mechanisms, our SAN can further reduce the mean delay and suppress the maximum delay at a lower level, with even better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Next, we investigate the performance of our SAN under different delay thresholds \\( d_{b} \\) and accuracy coefficients \\( \\lambda_{acc} \\). As shown in Tab. 6, under different values of \\( d_{b} \\), our SAN can still significantly decrease the maximum delay while maintaining high accuracy. Furthermore, as \\( d_{b} \\) increases, SAN is allowed to have higher delays before getting penalized, leading to higher accuracy while slightly degrading delay metrics, which is what we expected. We also evaluate different values of \\( \\lambda_{acc} \\) and present the results in Tab. 6. We observe that our SAN generally achieves good accuracy and delay performance for different values of \\( \\lambda_{acc} \\), showing that" + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.291, + 0.845, + 0.304 + ], + "angle": 0, + "content": "Table 5. Ablation results for dynamic mechanism." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.305, + 0.889, + 0.391 + ], + "angle": 0, + "content": "
SettingAvgMax DelayMean Delay
res = 256, dep = 394.4 %129.6 ms45.5 ms
res = 256, dep = [1,3]92.6 %71.4 ms39.3 ms
res = 256, dep = [1,2,3]93.3 %63.6 ms34.5 ms
res = [128,256], dep = 391.2 %84.0 ms43.7 ms
res = [64,128,192,256], dep = 391.9 %61.9 ms38.0 ms
Our SAN93.4 %41.2 ms33.1 ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.517, + 0.404, + 0.874, + 0.417 + ], + "angle": 0, + "content": "Table 6. Ablation results for different values of \\(d_b\\) and \\(\\lambda_{acc}\\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.419, + 0.889, + 0.515 + ], + "angle": 0, + "content": "
SettingAvgMax DelayMean Delay
db=10 ms92.4 %39.2 ms27.7 ms
=20 ms92.8 %39.4 ms30.7 ms
=30 ms93.4 %41.2 ms33.1 ms
=40 ms94.0 %43.5 ms41.9 ms
λacc=191.5 %41.0 ms30.8 ms
=293.4 %41.2 ms33.1 ms
=494.0 %44.6 ms37.1 ms
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.533, + 0.892, + 0.58 + ], + "angle": 0, + "content": "it is robust to the \\(\\lambda_{acc}\\) hyperparameter. Also, when \\(\\lambda_{acc}\\) increases, the accuracy improves and delay slightly degrades, as we would expect." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.62, + 0.624 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.821 + ], + "angle": 0, + "content": "In this work, we tackle the efficiency problem of deep networks on devices with computational fluctuations. To the best of our knowledge, this is the first time this problem is being explicitly tackled for online video understanding tasks. We propose SAN, a novel system-status-aware adaptive network for low-delay online action recognition and online pose estimation tasks with a fluctuating computation budget. A self-supervised meta optimization framework (MSA) is also proposed for more effective adaptation between hardware platforms. Experiments show that our proposed SAN and MSA obtains SOTA performance with low delay even under a rapidly fluctuating system load." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10521" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.214 + ], + "angle": 0, + "content": "[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[4] Mario Emad, Michael Ishack, Mohamed Ahmed, Mohamed Osama, Mohamed Salah, and Ghada Khoriba. Early-anomaly prediction in surveillance cameras for security applications. In 2021 International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC), pages 124-128. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Zhipeng Fan, Jun Liu, and Yao Wang. Motion adaptive pose estimation from compressed videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11719-11728, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.469, + 0.527 + ], + "angle": 0, + "content": "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[9] giampaolo. psutil. https://github.com/giampaolo/psutil. Accessed: 2022-4-25. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[10]gpuopenanalytics. pynvml. https://github.com/gpuopenanalytics/pynvml. Accessed: 2022-4-25.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[11] Pratik Gujjar and Richard Vaughan. Classifying pedestrian actions in advance using predicted video of urban driving scenes. In 2019 International Conference on Robotics and Automation (ICRA), pages 2097-2103. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[12] Amirhossein Habibian, Davide Abati, Taco S Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2695–2704, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[13] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[17] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[18] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11719-11727, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[20] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[21] Hayeon Lee, Sewoong Lee, Song Chong, and Sung Ju Hwang. Help: Hardware-adaptive efficient latency prediction for nas via meta-learning. arXiv preprint arXiv:2106.08630, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[22] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[23] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Yu Mitsuzumi, Go Irie, Daiki Ikami, and Takashi Shibata. Generalized domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1084-1093, June 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Maria Florina Balcan and Kil" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10522" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "ian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1928-1937, New York, New York, USA, 20-22 Jun 2016. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[28] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[29] Xuecheng Nie, Yuncheng Li, Linjie Luo, Ning Zhang, and Jiashi Feng. Dynamic kernel distillation for efficient pose estimation in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6942-6950, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[30] Toby Perrett and Dima Damen. Recurrent assistance: Cross-dataset training of lstms on kitchen tasks. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1354-1362. IEEE Computer Society, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[31] T. Perrett and D. Damen. Ddlstm: Dual-domain lstm for cross-dataset action recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7844-7853, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[32] Steven C Seow. Designing and engineering time: The psychology of time perception in software. Addison-Wesley Professional, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[33] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[34] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[35] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[36] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.76 + ], + "angle": 0, + "content": "[37] Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han. Hat: Hardware-aware transformers for efficient natural language processing. In ACL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[38] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the 17th Conference in Uncertainty in Artificial Intelligence, UAI '01, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[39] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[40] Zuxuan Wu, Caiming Xiong, Yu-Gang Jiang, and Larry S Davis. Liteeval: A coarse-to-fine framework for resource" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "efficient video recognition. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[42] Lumin Xu, Yingda Guan, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Vipnas: Efficient video pose estimation via neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16072-16081, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[43] XuehaiPan. nvitop. https://github.com/XuehaiPan/nvitop. Accessed: 2022-4-25. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[44] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[45] Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization: Towards lossless cnns with low-precision weights. In International Conference on Learning Representations, ICLR2017, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[46] Mohammadreza Zolfaghari, Kamaljeet Singh, and Thomas Brox. Eco: Efficient convolutional network for online video understanding. In Proceedings of the European conference on computer vision (ECCV), pages 695-712, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10523" + } + ] +] \ No newline at end of file diff --git a/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_origin.pdf b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..513a0dee010291c26a644933b09333dc993a6ed9 --- /dev/null +++ b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/d274131a-49c0-4a6a-97bb-9cb59d470bd6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cb04ed031826c46be292b168483b523a88e078c44e9e9c39f10a4d05b01c660 +size 1682064 diff --git a/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/full.md b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7cef3b13cf11b8a4667f4a12f73f8b0d3347e0c --- /dev/null +++ b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/full.md @@ -0,0 +1,276 @@ +# System-status-aware Adaptive Network for Online Streaming Video Understanding + +Lin Geng Foo $^{1\dagger}$ Jia Gong $^{1\dagger}$ Zhipeng Fan $^{2§}$ Jun Liu $^{1‡}$ + +$^{1}$ Singapore University of Technology and Design $^{2}$ New York University + +{lingeng_foo, jia_gong}@mypail.sutd.edu.sg, zf606@nyu.edu,jun_riu@sutd.edu.sg + +# Abstract + +Recent years have witnessed great progress in deep neural networks for real-time applications. However, most existing works do not explicitly consider the general case where the device's state and the available resources fluctuate over time, and none of them investigate or address the impact of varying computational resources for online video understanding tasks. This paper proposes a System-status-aware Adaptive Network (SAN) that considers the device's real-time state to provide high-quality predictions with low delay. Usage of our agent's policy improves efficiency and robustness to fluctuations of the system status. On two widely used video understanding tasks, SAN obtains state-of-the-art performance while constantly keeping processing delays low. Moreover, training such an agent on various types of hardware configurations is not easy as the labeled training data might not be available, or can be computationally prohibitive. To address this challenging problem, we propose a Meta Self-supervised Adaptation (MSA) method that adapts the agent's policy to new hardware configurations at test-time, allowing for easy deployment of the model onto other unseen hardware platforms. + +# 1. Introduction + +Online video understanding, where certain predictions are immediately made for each video frame by using information in the current frame and potentially past frames, is an important task right at the intersection of video-based research and practical vision applications (e.g., self-driving vehicles [11], security surveillance [4], streaming services [32], and human-computer interactions [20]). In particular, in many of these real-world video-based applications, a fast and timely response is often crucial to ensure high usability and reduce potential security risk. Therefore, in many practical online applications, it is essential to ensure that the model is working with low delay while maintaining a good + +performance, which can be challenging for many existing deep neural networks. + +Recently, much effort has been made to reduce the delay of deep neural networks, including research into efficient network design [16, 36, 44], input-aware dynamic networks [5,6,12,24], and latency-constrained neural architectures [1, 2, 21]. However, all these works do not explicitly consider the dynamic conditions of the hardware platform, and assume stable computation resources are readily available. In practical scenarios, the accessible computing resources of the host devices can be fluctuating and dynamic due to the fact that multiple computationally expensive yet important threads are running concurrently. For example, in addition to performing vision-related tasks such as object detection, human activity recognition, and pose estimation, state-of-the-art robotic systems usually need to simultaneously perform additional tasks like simultaneous localization and mapping (SLAM) to successfully interact with humans and the environment. Those tasks are also often computationally heavy and could compete with vision tasks for computing resources. As a result, at times when the host device is busy with other processes, conducting inference for each model might require significantly more time than usual, leading to extremely long delays, which could cause safety issues and lagging responses in many real-world applications. Therefore, the study and development of models providing reliable yet timely responses under various hardware devices and fluctuating computing resources is crucially important. Unfortunately, such studies are lacking in the field. + +To achieve and maintain low delay for online video understanding tasks under a dynamic computing resource budget, we propose a novel System-status-aware Adaptive Network (SAN). Different from previous works, SAN explicitly considers the system status of its host device to make on-the-fly adjustments to its computational complexity, and is thus capable of processing video streams effectively and efficiently in a dynamic system environment. SAN comprises of two components: a) a simple yet effective dynamic main module that offers reliable predictions under various + +network depths and input resolutions; b) a lightweight agent that learns a dynamic system-status-aware policy used to control the execution of the main module, which facilitates adaptation to the fluctuating system load. With the adaptivity of the main module and the control policy generated by the agent, our SAN can achieve good performance on the online video understanding task while maintaining a low delay under fluctuating system loads. + +In various applications, we may need to deploy SAN onto different hardware platforms for online video understanding. However, it is inconvenient to train SAN for each hardware platform, and it might also be difficult to find adequate storage to load the large labeled dataset on all platforms (e.g., mobile devices). In light of these difficulties, we further propose a method for deployment-time self-supervised agent adaptation, which we call Meta Self-supervised Adaptation (MSA). With MSA, we can conveniently train a SAN model on a set of local platforms, and perform a quick deployment-time agent adaptation on a target device, without the need for the original labeled training data. Specifically, our proposed MSA introduces an auxiliary task of delay prediction together with a meta-learning procedure, that facilitates the adaptation to the target deployment device. + +In summary, the main contributions of this paper are: + +- We are the first to explicitly consider the fluctuating system status of the hardware device at inference time for online video understanding. To address this, we propose SAN, a novel system-status-aware network that adapts its behavior according to the video stream and the real-time status of the host system. +- We further propose a novel Meta Self-supervised Adaptation method MSA that alleviates the training burden and allows our model to effectively adapt to new host devices with potentially unclear computation profiles at deployment time. +- We empirically demonstrate that our proposed method achieves promising performance on the challenging online action recognition and pose estimation tasks, where we achieve low delays under a rapidly fluctuating system load without jeopardizing the quality of the predictions. + +# 2. Related Work + +Online Video Understanding. Recently, motivated by the increasing real-world demand, a lot of works [5,6,29,40, 46] have attempted to improve the accuracy and efficiency of models for online video understanding tasks, such as online action recognition [23, 40, 46] and online pose estimation [6, 12, 29]. Several works design efficient networks to improve models' efficiency for online video tasks, including + +the efficient 3D CNN design [46], the Temporal Shift 2D CNN network [23] and the Skip-Convolution [12], while other researchers introduce adaptivity into the networks, termed as stream-aware dynamic networks [5, 25, 29] to further reduce the network's computation complexity while maintaining a good performance. For example, LiteEval [40] dynamically chooses between a coarse model and a fine model for each frame and MAPN [6] dynamically activates its encoder to save computation resources. However, all these existing methods do not explicitly consider the dynamic conditions of the hardware platform, and thus face limitations in the face of fluctuating system conditions. For example, in situations when the system is under high computational load, the model may still select to execute the branch with high complexity and therefore incur significantly longer delays, which is sub-optimal for time-sensitive applications. Compared to these methods, our proposed SAN is system-status-aware and thus more robust to computational fluctuations on the models' host devices. To the best of our knowledge, such a system-status-aware adaptive network that can efficiently tackle online video understanding tasks has not been explored before. + +Efficient Network Designs. There are several paradigms for efficient deep architectures, and we list a few here. Efficient deep models such as SqueezeNet [17] and MobileNet [16] introduce novel computation operators using fewer parameters, while Neural Network Quantization [3,45] methods effectively shrink the model size of the existing models by quantizing model parameters to isolated values. Teacher-student models [15] are also effective in distilling knowledge from larger models into smaller ones. Some other latency-constrained networks [37, 42] further conduct neural architecture search (NAS) to search for an efficient architecture. Different from all these approaches, our SAN explicitly accounts for the dynamic system status that is constantly changing on the device over time, allowing us to perform efficient online video understanding even with a fluctuating system status. + +Reinforcement Learning (RL). RL methods [22,27,28] train agents to navigate environments such that the reward is maximized. To allow pre-trained RL agents to adapt to a new environment in fewer episodes, methods based on the few-shot setting [7, 18] have been developed. Differently, we aim to allow our RL agent to adapt at deployment time in a self-supervised manner (i.e., no reward signals can be provided due to the absence of labeled video data). We introduce MSA, a novel method for self-supervised deployment-time adaptation. MSA employs an auxiliary task of delay prediction, and performs meta-learning to allow the RL agent to effectively adapt its control policy to the target hardware platform via fine-tuning on the auxiliary task. + +# 3. Method + +Given the live video stream $V = \{I_1, \dots, I_t, \dots\}$ and a seen/unseen host device whose system status $sys_t$ varies over time, our goal is to build a model that dynamically adjusts the computational complexity of inference based on not only the streaming input, but also the system status to make high-quality predictions with a low delay at each step $t$ . To this end, our proposed model: 1) takes into account the available computational resources to make on-the-fly decisions to dynamically adjust the computation consumption of the model, therefore maintaining the delay at low values while keeping the prediction's accuracy; and 2) effectively adapts itself to unseen devices in a self-supervised manner while still keeping its performance during deployment. + +# 3.1. SAN: System-status-aware Adaptive Network + +For better online performance amid the host device's varying background computational load, we propose a System-status-aware Adaptive Network (SAN). As shown in Fig. 1, our SAN consists of two main components: a lightweight agent $\pi$ that decides how to process the incoming frame and a dynamic main network $M$ that processes the frame with the specified policy generated by $\pi$ . + +Dynamic main network. Our dynamic main network $M$ flexibly produces frame-wise predictions using an architecture determined by the policy from the agent module $\pi$ . To accomplish this without extra burden on the hardware system, we design a dynamic encoder that has both dynamic depth and dynamic resolution. Our dynamic depth mechanism gives the encoder the option of producing output features at shallower intermediate layers without having to wait for all layers to be executed. This reduces computational cost by executing a "simpler" sub-network. On the other hand, the dynamic resolution mechanism allows our encoder to selectively take in input images at a lower resolution. Crucially, this reduces the amount of pixels to be examined by the network, effectively reducing the amount of convolution operations. We implement both of the aforementioned mechanisms in a single network, where to enable dynamic execution depth of the network, we add early exits to the network which allow the execution to end at an intermediate layer, while to enable dynamic resolution, we leverage the fully convolutional network and add a dynamic global pooling operation at the end. + +Specifically, as shown in Fig. 1, our dynamic encoder selects from $m$ input resolutions $\{res^i\}_{i=1}^m$ and $n$ model depths $\{dep^j\}_{j=1}^n$ , allowing for as many as $m \times n$ different options in total. These $m \times n$ dynamic options have different computation complexities and levels of performance. For online video understanding, we construct our dynamic encoder as a 2D CNN for handling each frame, and to model the temporal aspect, the output features from the dynamic encoder are fed into a Long Short Term Memory network + +(LSTM) at every step. The LSTM updates its hidden state $h_t$ based on the features of the incoming frame $I_t$ and its previous state $h_{t-1}$ , which is then used by the task head to determine the final prediction $\hat{y}_t$ for the video task following: $\hat{y}_t = M(I_t, rest_t, dept_t, h_{t-1}; \theta)$ , where $\theta$ refers to the parameters of the main network $M$ , while $rest_t$ and $dept_t$ are selected by the agent introduced next. More details of our main dynamic network are in Sec. 4 and Supplementary. + +RL-based agent. The RL-based agent $\pi$ controls the dynamic main network $M$ by generating a frame-level processing policy $(\text{rest}_t, \text{depth}_t)$ at each time $t$ , which aims to maintain the task-related accuracy $\text{acc}_t$ at high level with low delay $d_t$ . To make decisions amid the fluctuating system loads, our lightweight agent $\pi$ should be both system-status-aware and streaming input-aware. We model our task as a Markov Decision Process, where the agent $\pi$ takes an action $a_t$ at every step $t$ , based on the observed state $s_t$ , and transferred to the next state $s_{t+1}$ . Crucially, state $s_t$ includes system status information $\text{sys}_t$ and input stream information $h_{t-1}$ , making $\pi$ both system status and stream aware. To learn an optimal policy, we propose an RL-based method to train our agent. + +Specifically, we set the observed states of the agent as $s_t = [h_{t-1}, sys_t, g_t]$ , where $sys_t$ is the system status (e.g., the CPU and GPU utility) $^1$ , $h_{t-1}$ is the LSTM hidden state, and $g_t$ refers to other useful information (i.e., the previous step's action $a_{t-1}$ and delay $d_{t-1}$ ). Then, based on the available resolutions $\{res^i\}_{i=1}^m$ and depths $\{dep^j\}_{j=1}^n$ , we build the action space as $\{a^{i,j} = (res^i, dep^j)\}_{i=1,j=1}^{m,n}$ . As shown in Fig. 1, at time $t$ , the agent $\pi$ receives $h_{t-1}$ , $sys_t$ and $g_t$ , and generates a probability distribution over the action space: + +$$ +P _ {t} = \left[ p _ {t} ^ {1, 1}, p _ {t} ^ {1, 2}, \dots , p _ {t} ^ {m, n} \right] = \pi \left(h _ {t - 1}, s y s _ {t}, g _ {t}; \phi\right) \quad (1) +$$ + +where $p_t^{i,j} \in [0,1]$ is the probability of the action $a^{i,j}$ at time $t$ , $P_t \in [0,1]^{m\cdot n}$ represents the probability distribution over the action space, and $\phi$ refers to the parameters of the agent $\pi$ . Then, we take the highest probability entry in $P_t$ (denoted as $p_t^{i*,j*}$ ), select the corresponding action $a_t = (res^{i*}, dep^{j*})$ and execute the main network accordingly. + +Training of SAN. To train our main network $M$ , we apply the loss $\mathcal{L}_M$ based on the specified online video-processing task. For example, for online action recognition, $\mathcal{L}_M$ can be the cross-entropy loss, while for online pose estimation, $\mathcal{L}_M$ can be the Mean Squared Error (MSE) loss. + +To control the dynamic encoder to achieve a good accuracy at low delay, we apply an RL-based loss $\mathcal{L}_{\pi}$ to train our agent $\pi$ . Specifically, we adopt a policy gradient loss [38, 39]. To maximize the accuracy $acc_{t}$ while reducing the + +![](images/87cdabc49c873cf45f453eabff9f9662ea65ed252862a28e319b74006f782617.jpg) +Figure 1. Overview of our SAN framework, which consists of two modules: the agent module $\pi$ and the dynamic main network $M$ . For each incoming frame $I_{t}$ , the agent module first takes in the LSTM hidden state $h_{t-1}$ , system-status $sys_{t}$ , and other useful information $g_{t}$ and then decides the input resolution $rest_{t}$ and the depth $depth_{t}$ of $M$ via the policy head. After that, the main network $M$ processes the frame according to these decisions. By controlling the behavior of the main network $M$ via the agent $\pi$ , our SAN can output high quality and low delay predictions in the dynamic system environment. Note that the auxiliary task head of the agent is used for fine-tuning of the agent during deployment, calibrating the learned system-status-aware features to adapt to the unseen target device. + +delay $d_{t}$ , we set the reward of the agent at time $t$ as: + +$$ +r _ {t} = \lambda_ {a c c} a c c _ {t} \left(a _ {t}, I _ {t}\right) - \max \left(d _ {t} \left(a _ {t}, s y s _ {t}\right) - d _ {b}, 0\right) \tag {2} +$$ + +where $acc_{t}(a_{t},I_{t})$ and $d_t(a_t,sys_t)$ are the accuracy and delay of our main network under the action $a_{t}$ , the current frame $I_{t}$ , and system status $sys_{t}$ . $\lambda_{acc}$ is the coefficient of the accuracy term, and $d_b$ is a tolerance threshold for the delay (e.g., $d_b = 30$ ms for a rate of 30 fps), which is the acceptable delay. The policy gradient loss $\mathcal{L}_{\pi}$ is defined as: + +$$ +\mathcal {L} _ {\pi} = \sum_ {t = 1} ^ {T} - r _ {t} \log \left(p _ {t}\right) \tag {3} +$$ + +where $p_t \in [0,1]$ is the probability of selecting the action $a_t$ at time $t$ . + +# 3.2. MSA: Meta Self-supervised Adaptation + +Ideally, we want to deploy SAN onto many different hardware platforms for online video applications, requiring us to train SAN for each of those platforms, which is labor intensive, computationally prohibitive, and may also be infeasible at times due to storage concerns (e.g., mobile devices). Instead, a more practical option would be to effectively fine-tune a pre-trained SAN at deployment-time on a target device. However, during deployment, the ground-truth labels associated with the real-time video stream are often inaccessible. To address this issue, we introduce an auxiliary task [13, 35] of delay prediction, where the agent $\pi$ is expected to additionally predict the actual processing time based on the selected action (i.e., the input resolution and execution depth of main network $M$ ). + +Our intuition is that, on a new (unseen) platform, our agent $\pi$ lacks understanding of hardware status of this platform, thus we need a way for the agent to quickly understand it. Hence, we introduce the auxiliary task of delay + +prediction, because delay times are directly related to the system status and the model's computational cost. Therefore, when the agent $\pi$ is fine-tuned using delay prediction, it will be pushed to adopt a better understanding of the host platform's characteristics and the costs of different options of $M$ on that platform. Moreover, delay times can be directly observed, and can be a convenient supervision signal on a new platform since we do not have signals from our main task (which requires labeled video data). Therefore, on a new platform, we can fine-tune our agent to predict the delay by using the observed delay times as supervision, which will improve the agent's understanding of the device. + +Although the delay prediction task improves the quality of the intermediate features (which are also used to select actions), fine-tuning with it does not necessarily lead to a better policy on the host device. Hence, to better align the performance on the auxiliary task of delay prediction with policy learning, we further propose a meta-optimization-based learning framework to find a good initialization for the agent, such that updating based on the loss of delay prediction would also improve the quality of the policy. Combining the auxiliary task and the meta-learning framework allows us to train SAN on a set of local hardware platforms, and perform only a quick deployment-time adaptation on the target device at run-time, without requiring the actual labeled video dataset on that target device. + +Auxiliary task of delay prediction. When SAN is deployed to an unseen target device, the processing times of each action at a given system status will shift significantly due to the drift in computing profiles. Therefore, fine-tuning on the target device is required to adapt SAN to the new hardware characteristics. Due to the lack of ground-truth labels for the streaming videos at run-time, we introduce a self-supervised auxiliary task of delay prediction for adapting SAN to unseen hardware platforms. The delay predic + +tion task calibrates the learned system-status-aware features for the agent module $\pi$ using supervisory signals (processing times) that can be generated on-the-fly with only a simple timer on the device. + +Specifically, an auxiliary task head is added to the agent $\pi$ , as shown in Fig. 1. The auxiliary task head takes in the features from the agent as well as the chosen action $a_{t}$ , to predict the delay time $d_{t}$ . To train our model on the auxiliary task, we apply a mean-squared-error loss between our predictions $\hat{d}_t$ and the observed delay $d_{t}$ . The auxiliary loss $\mathcal{L}_{aux}$ at step $t$ , is thus defined as: $\mathcal{L}_{aux,t} = (d_t - \hat{d}_t)^2$ . + +At deployment time, we fine-tune the agent $\pi$ on the target device using this auxiliary task only, allowing the agent to better understand the target device's characteristics and processing times. However, this does not necessarily mean that the agent has optimized its policy output – in order to achieve that, we further introduce a meta-optimization method as described next. + +Meta-optimization for self-supervised agent adaptation. During deployment onto the target device, optimizing the auxiliary task of delay prediction is loosely connected to improving the actual policy. To better bind the quality of the developed policy to the delay prediction performance, we introduce a meta-optimization step to our agent $\pi$ during training, such that fine-tuning of the auxiliary task will also lead to a better adaptation of the policy. + +We approach this problem by simulating the fine-tuning process using the delay prediction error on multiple target platforms, and (meta) optimize $\phi$ such that the update based on the delay prediction error leads to better adaptation of the policy as well. This will effectively link the quality of the policy to the delay prediction performance. Here, we adopt a meta-learning framework to tackle the problem, treating $\phi$ as the meta-parameters to meta-optimize. In this context, we denote the auxiliary task as the meta-train task and the agent's policy generation as the meta-test task. Our meta-optimization algorithm aims to learn a good initialization of $\phi$ , such that when it is updated via the meta-train task, it can achieve good performance on the meta-test task. + +We assume that our SAN is first trained on a source device $D_{src}$ , and then adapted to the additional $K$ target platforms $\{D_{tgt}^k\}_{k=1}^K$ . After first training our SAN on the source device $D_{src}$ , we freeze the main network $M$ , and keep a copy of the agent's parameters (denoted as $\phi_{meta,0}$ ) to initialize the meta-learning process. In the $u$ -th meta-learning iteration, we first meta-train the agent for each target device $D_{tgt}^k$ using its loss on the auxiliary task and perform a pseudo-update on the meta parameters $\phi_{meta,u}$ which can be formalized as: + +$$ +\phi_ {t g t, u} ^ {k} \leftarrow \phi_ {\text {m e t a}, u} - \alpha \nabla_ {\phi_ {\text {m e t a}, u}} \mathcal {L} _ {\text {a u x}, t} \left(\phi_ {\text {m e t a}, u}, D _ {t g t} ^ {k}\right) \tag {4} +$$ + +where $\phi_{tgt}^k$ refers to the parameters of the meta-trained agent $\pi_{tgt}^k$ which is about to be deployed onto the target device $D_{tgt}^k$ , and $\alpha$ is the learning rate for meta-training. + +Then, we check the performance of the pseudo-updated parameters on target device by meta-testing the performance of the pseudo-updated agent $\pi_{tgt}^{k}(:,\phi_{tgt,u}^{k})$ for the agent's policy generation following: + +$$ +\mathcal {L} _ {\text {m e t a}} ^ {k} \left(\phi_ {t g t, u} ^ {k}\right) = \mathcal {L} _ {\pi} \left(\phi_ {t g t, u} ^ {k}, D _ {t g t} ^ {k}\right), \tag {5} +$$ + +where $\mathcal{L}_{meta}^k$ is the meta-testing-loss from the target hardware platform $D_{tgt}^k$ . Intuitively, this loss penalizes our agent based on its performance on a target device after a short period of fine-tuning. Hence, we update our meta parameters $\phi_{meta,u}$ in the $u$ -th meta-learning iteration as follows: + +$$ +\phi_ {m e t a, u + 1} \leftarrow \phi_ {m e t a, u} - \beta \sum_ {k = 1} ^ {K} \nabla_ {\phi_ {m e t a, u}} \mathcal {L} _ {m e t a} ^ {k} \left(\phi_ {t g t, u} ^ {k}\right), \tag {6} +$$ + +where $\beta$ is the learning rate for optimizing the meta parameters, and $\phi_{tgt,u}^k$ is obtained via the pseudo-update in Eq. 4. Through updating our model with this meta-learning scheme, our agent parameters will converge towards optimal parameters $\phi_{meta,*}$ that are able to effectively adapt to each target device through fine-tuning via the auxiliary task. + +# 3.3. Training and Testing + +As mentioned above, we consider two scenarios: deploying our SAN on a seen or unseen device. + +Model training and testing for seen device. In this scenario, we evaluate SAN when it is trained and tested on the same device, which evaluates the effectiveness of our SAN design described in Sec. 3.1. For model training, we first utilize the dataset's training set to initialize our main network $M$ , which is controlled by a random policy (randomly sampling the resolution-depth pair) to improve its accuracy via the loss $\mathcal{L}_M$ . Then, we train our agent $\pi$ using the loss $\mathcal{L}_{\pi}$ with the parameters of the main network $(M)$ fixed. Lastly, we fine-tune the main network $M$ and agent $\pi$ jointly in a fully end-to-end manner with both $\mathcal{L}_M$ and $\mathcal{L}_{\pi}$ , allowing us to obtain a high-performance video understanding dynamic network controlled by an optimal system-status-aware policy on the current local device. For evaluation, we test the trained SAN on the same device, but with an unseen system status trajectory, which evaluates both the accuracy and the delay of our SAN. + +Model meta-optimization and deployment for unseen device. In this scenario, our SAN is pre-trained and meta-optimized on a set of seen source devices, and we evaluate its efficacy when deployed to an unseen device. This evaluates the efficacy of our MSA design described in Sec. 3.2. At the start of training, we first initialize and pre-train our SAN on a seen device, following the process + +described previously. Then, during meta-optimization, we freeze the main network $M$ and only optimize our agent $\pi$ . During meta-optimization iterations, we randomly sample video clips from the training set for meta-training and meta-testing. For better generalization, video clips used for meta-training are not used for meta-testing. When deploying our SAN to an unseen target device, we first fine-tune the agent $\pi$ via the auxiliary task of delay prediction. The adapted agent is then tested for average accuracy and delay over all frames in the testing set. When comparing our method's average accuracy and delay with others, we compute the global accuracy and delay (counting both the fine-tuning and final testing phase) for a fair comparison. + +# 4. Experiments + +We describe our setup and results below. More analysis and details are provided in Supplementary. + +SAN's architecture. For our dynamic encoder, we adopt the fully-convolutional backbone ResNet50 [14] to handle various input resolutions. We add three exit ports (i.e., $n = 3$ ) at the end of these specific layers of ResNet50 ( $\{conv3_x, conv4_x, conv5_x\}$ [14]) to achieve dynamic depth. Moreover, we specifically design a convolutional layer for each resolution-depth pair to unify the shapes of the encoder's output features (details in Supplementary). Global average pooling is performed on the unified encoder's output and send to the LSTM to update its hidden state $h \in \mathbb{R}^{2048}$ . The output of the LSTM is fed into the task head for online video understanding. On the other hand, our lightweight agent $\pi$ consists of three fully connected layers. We utilize three widely used monitors [43] to collect system information, which contains the status of system's CPU, GPU, memory, disk, I/O and fan. For model training, we set $\lambda_{acc} = 2$ , $d_b = 0.03$ , $\alpha = 1e^{-3}$ and $\beta = 1e^{-5}$ . More details of our SAN are presented in Supplementary. + +Online video task settings. We compare the performance of our SAN with the prior methods on two widely used online video understanding tasks: 1) Online action recognition: We follow the previous work [30] to use the 50Salads dataset [34], since it contains very long videos that are suitable for online video action recognition. Here, we consider three (i.e., $m = 3$ ) resolution candidates: [112, 168, 224] and use a fully-connected layer as the task head. We compare our method with SOTA online action recognition methods, including RA [30], DDLSTM [31], AR-Net [25], and LiteEval [40]. To test models on the same platform for a fair comparison, we use their publicly released code (AR-Net) when available and implement the rest (RA, DDLSTM and LiteEval) on our own. 2) Online pose estimation: We follow the existing video pose estimation works [6, 12] to conduct our experiments on the Sub-JHMDB dataset [19], a collection of 11,200 frames from 316 video clips, labeled with 15 body + +Table 1. Results on 50Salads dataset for online action recognition. + +
MethodAccuracyMax DelayMean DelayR\T Frames
RA [30]42.6 %76.6 ms33.9 ms79.1 %
DDLSTM [31]41.1 %407.4 ms110.3 ms5.9 %
LiteEval [40]40.3 %110.7 ms68.9 ms30.4 %
AR-Net [25]40.9 %91.1 ms56.2 ms62.1 %
main network + random policy46.2%80.5 ms34.1 ms92.1%
+ stream-aware55.4 %80.8 ms40.3 ms82.0%
+ SAN53.8%49.4 ms29.9 ms95.4%
+ +![](images/4f978f910c782b02b1d984314c6dca1a72f226867e5acfc2d996196953c52f56.jpg) + +![](images/ea8ecf390824b23a0fe233472130cdc99508ce257a346d3abe3de93c976689f5.jpg) +Figure 2. A trace of models' accuracy and delay for online video-based action recognition (left) and pose estimation (right). In the areas masked by the gray color, we can observe a rapid increase of the baseline model's delay (see the second row), which corresponds to the high-load status of the system. + +![](images/3657466d63a82059928f570e2e1b2d95081a753e2646908fc6a481b43ebb7d98.jpg) + +![](images/a5bf85d570bbd53c2484eb1bec05d789a25c635e9680adf339f5f62a8f7053fa.jpg) + +joints. Here, we consider three (i.e., $m = 3$ ) resolution candidates: [128, 192, 256] and follow the work of [41] to use 3 deconvolutional layers to build the task head for pose estimation. Moreover, we replace LSTM with a ConvLSTM [33] to accommodate for 2D feature maps. We compare our method with the existing dynamic pose estimation networks, i.e., DKD [29] and Skip-Convolution [12]. Similarly, we re-implement DKD and Skip-Convolutions as there is no publicly available code. + +Experiment settings. To build dynamic systems with varying system status, we design multiple background processes (i.e., matrix calculators, video compressors, and large deep learning models) to occupy various amounts of computational resources. We can then generate dynamic system load trajectories using these processes to simulate dynamic system load environments. During model training, we randomly generate dynamic system trajectories in each iteration to simulate various dynamic environments to train our SAN. During testing, for fair comparisons, we first generate 3 dynamic trajectories unseen in training, and then run all the evaluation experiments on these 3 fixed trajectories and report the average result. To evaluate the models' efficiency, we measure the maximum delay and mean delay for each model. Moreover, as the videos of both datasets [19, 34] are at 30 fps, we consider delay time $< 30$ ms as real-time, and measure the percentage of such real-time predictions (R\T Frames). + +Table 2. Results on Sub-JHMDB dataset for online pose estimation. + +
MethodHeadSho.Elb.Wri.HipKneeAnk.AvgMax DelayMean DelayR\T Frames
DKD [29]98.396.690.487.199.196.092,994.0133.8 ms47.9 ms59.9 %
SimpleBaseline [41]97.597.891.186.099.696.892.694.4129.6 ms45.5 ms63.8 %
Skip-Convolution [12]98.797.792.088.199.396.691.095.1115.7 ms51.2 ms53.7 %
main network + random policy97.595.291.493.577.786.872.486.9130.9 ms36.1 ms72.9%
main network + stream-aware99.498.697.598.193.994.290.095.3126.1 ms44.7 ms63.7%
main network + SAN99.198.797.398.290.192.585.093.441.2 ms33.1 ms81.4%
+ +# 4.1. Evaluating SAN on seen devices + +First, we compare our SAN with the SOTA methods for both tasks and report the results in Tab. 1 and Tab. 2. In both tables, we also report the performance of two baselines of our method: main network + random policy where actions are randomly selected; and main network + stream-aware where the agent $\pi$ uses information from the input stream without considering system status. Both baselines are not system-status-aware. + +Online action recognition. Tab. 1 presents the results on the 50Salads dataset for online action recognition. We compare our method with four SOTA methods, including two static models (i.e., RA [30] and DDLSTM [31]) and two efficient dynamic models (i.e., AR-Net [25] and LiteEval [40]). Our SAN outperforms these methods on all metrics, achieving a higher accuracy while maintaining a low mean delay (29.9ms) and high percentage of real-time frames (95.4%). These improvements are due to our system-status-aware design of SAN, allowing it to make on-the-fly adaptations to process a higher percentage of frames in real-time. The two baselines are not system-status-aware and therefore perform significantly worse on the delay-related metrics, especially on max delay (80.5ms and 80.8ms vs our 49.4ms). Compared to the random policy, stream-aware policy significantly improves the accuracy at the cost of reduced real-time performance. Overall, our SAN achieves superior performance and at the same time also decreases the delay. + +Online pose estimation. Tab. 2 presents the results on the Sub-JHMDB dataset for online pose estimation. We compare our SAN with a static network SimpleBaseline [41] and two dynamic efficient networks: DKD [29] and Skip-Convolution [12]. As shown in Tab. 2, our main network + stream-aware baseline can achieve the SOTA performance but may still incur large delays occasionally. With the help of our system-status-aware design, our SAN can control the maximum and mean delay at a low level while still maintaining high accuracy. + +Qualitative analysis. We analyze the effect of the system load dynamics and show results from test-time runs on a laptop with a NVIDIA Geforce GTX 1080 GPU (11 GB), where 10 background processes are randomly activated. Fig. 2 presents the models' frame-level delay and accuracy along with the system status. We compare our SAN with two baselines: main network + stream-aware and main network + random policy. In the first row of Fig. 2, + +we observe that the GPU load fluctuates over time, varying the available computational resources for the online models. As a result, the delay of models with random/stream-aware policy fluctuates up to around 100ms as shown in the second row. In contrast, by using the system-status information to control the model, our SAN is able to stabilize the delay and avoid large delays amid the dynamic system. Moreover, as shown in the last row, with the help of the stream-aware design, our SAN can still obtain a similar accuracy to the stream-aware model. These results show the advantage of our system-status-aware and stream-aware designs, which can simultaneously control the real-time delay and attain high accuracy for online video understanding tasks. + +# 4.2. Evaluating SAN with MSA on unseen devices + +Our proposed MSA is introduced to achieve effective adaptation to the unseen host devices. Here, we compare our MSA with three baselines: 1) Direct transfer: Directly deploying our SAN to the unseen device. 2) Feature alignment: Minimizing the domain gap between the source device and target device via a domain discriminator [8, 26]. 3) Self-supervised fine-tune: Using our auxiliary task (delay prediction) to fine tune the model without meta optimization. To further demonstrate the efficacy of MSA, we also present the upper bound, dubbed as Fully-supervised learning, where we directly train the agent on the target environment, which eliminates the need for adaptation. + +Here, we collect three hardware platforms for training and testing: 1) Device a: a laptop with an AMD Ryzen Threadripper 2950X and an NVIDIA Geforce GTX 1080 GPU (11 GB). 2) Device b: a desktop with an Intel(R) Core(TM) i9-10900 and an NVIDIA Geforce RTX 2080 GPU (12 GB). 3) Device c: a workstation with an AMD Ryzen Threadripper 3960X and an NVIDIA Geforce RTX 3090 GPU (24 GB). We randomly sample two platforms as the seen source devices for model initialization and meta-optimization and use the remaining one as the unseen target device for model adaptation (e.g., Source: $\{a,b\} \to$ Target: $\{c\}$ ). Furthermore, we augment the source platforms by restricting the number of available CPU and the highest frequencies of CPU and GPU to achieve various environments with different resource configurations for better training. The results of policy adaptation are reported in Tab. 3 and 4. More results on other portable devices (e.g., NVIDIA Jetson TX2) are presented in Supplementary. + +Table 3. Results of policy adaptation across platforms for online pose estimation. More results on other platforms are in Supplementary. + +
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\T FramesAccuracyMean DelayR\T FramesAccuracyMean DelayR\T Frames
Fully-supervised learning (Upper bound)93.0 %28.3 ms87.4 %93.1 %24.7 ms84.9 %92.8 %28.1 ms74.5 %
Direct transfer87.3 %28.4 ms82.4 %85.7 %25.0 ms79.5 %90.9 %33.5 ms20.4 %
Feature alignment90.8 %29.0 ms80.4 %91.0 %27.2 ms74.6 %91.0 %31.8 ms64.5 %
Our self-supervised fine-tune91.1 %29.8 ms79.8 %85.8 %21.9 ms84.1 %91.5 %29.3 ms61.7 %
Our MSA92.8 %29.7 ms86.1 %92.9 %25.1 ms82.2 %92.4 %30.0 ms71.9 %
+ +Table 4. Results of policy adaptation across platforms for online action recognition. + +
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\T FramesAccuracyMean DelayR\T FramesAccuracyMean DelayR\T Frames
Fully-supervised learning (Upper bound)53.8 %26.3 ms95.4 %52.6 %31.3 ms93.1 %52.1 %31.9 ms90.4 %
Direct transfer41.7 %28.1 ms77.3 %47.2 %34.6 ms84.7 %52.0 %41.1 ms40.6 %
Feature alignment50.1 %29.8 ms82.4 %50.3 %32.7 ms88.1 %51.7 %37.2 ms60.5 %
Our self-supervised fine-tune49.1 %26.5 ms93.8 %51.0 %32.2 ms93.1 %50.7 %33.7 ms87.2 %
Our MSA52.6 %26.8 ms94.1 %52.2 %31.6 ms92.7 %51.9 %32.1 ms90.2 %
+ +As shown in Tab. 3 and 4, the direct transfer method fails to maintain the high accuracy and low delay simultaneously due to the hardware configuration shift between devices. On the other hand, by aligning the system-status-aware features between the source and target devices, the agent finds a suboptimal policy to balance the accuracy and delay. Using the auxiliary task to fine-tune the agent without meta optimization can achieve real-time predictions on the target device, but it sometimes fails to maintain high accuracy. Finally, our MSA simultaneously maintains a high ratio of real-time predictions and accurate predictions. Furthermore, the results are close to the upper bound of actually training on the target device, showing the effectiveness of our MSA to generalize to unseen devices. + +# 4.3. Ablation Study + +We further analyze the characteristics of our dynamic resolution and dynamic depth mechanism, and investigate the performance of our agent under various delay thresholds $d_{b}$ and accuracy coefficients $\lambda_{acc}$ . + +First, we evaluate the impact of the dynamic resolution and depth mechanism and perform ablation studies by allowing one of them to be dynamic while fixing the other. In Tab. 5, we observe that, by allowing for dynamic resolutions or dynamic depths, our main network can reduce the maximum and mean delay while maintaining high accuracy. Moreover, by combining these two mechanisms, our SAN can further reduce the mean delay and suppress the maximum delay at a lower level, with even better accuracy. + +Next, we investigate the performance of our SAN under different delay thresholds $d_{b}$ and accuracy coefficients $\lambda_{acc}$ . As shown in Tab. 6, under different values of $d_{b}$ , our SAN can still significantly decrease the maximum delay while maintaining high accuracy. Furthermore, as $d_{b}$ increases, SAN is allowed to have higher delays before getting penalized, leading to higher accuracy while slightly degrading delay metrics, which is what we expected. We also evaluate different values of $\lambda_{acc}$ and present the results in Tab. 6. We observe that our SAN generally achieves good accuracy and delay performance for different values of $\lambda_{acc}$ , showing that + +Table 5. Ablation results for dynamic mechanism. + +
SettingAvgMax DelayMean Delay
res = 256, dep = 394.4 %129.6 ms45.5 ms
res = 256, dep = [1,3]92.6 %71.4 ms39.3 ms
res = 256, dep = [1,2,3]93.3 %63.6 ms34.5 ms
res = [128,256], dep = 391.2 %84.0 ms43.7 ms
res = [64,128,192,256], dep = 391.9 %61.9 ms38.0 ms
Our SAN93.4 %41.2 ms33.1 ms
+ +Table 6. Ablation results for different values of $d_b$ and $\lambda_{acc}$ . + +
SettingAvgMax DelayMean Delay
db=10 ms92.4 %39.2 ms27.7 ms
=20 ms92.8 %39.4 ms30.7 ms
=30 ms93.4 %41.2 ms33.1 ms
=40 ms94.0 %43.5 ms41.9 ms
λacc=191.5 %41.0 ms30.8 ms
=293.4 %41.2 ms33.1 ms
=494.0 %44.6 ms37.1 ms
+ +it is robust to the $\lambda_{acc}$ hyperparameter. Also, when $\lambda_{acc}$ increases, the accuracy improves and delay slightly degrades, as we would expect. + +# 5. Conclusion + +In this work, we tackle the efficiency problem of deep networks on devices with computational fluctuations. To the best of our knowledge, this is the first time this problem is being explicitly tackled for online video understanding tasks. We propose SAN, a novel system-status-aware adaptive network for low-delay online action recognition and online pose estimation tasks with a fluctuating computation budget. A self-supervised meta optimization framework (MSA) is also proposed for more effective adaptation between hardware platforms. Experiments show that our proposed SAN and MSA obtains SOTA performance with low delay even under a rapidly fluctuating system load. + +Acknowledgments. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06). + +# References + +[1] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 1 +[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 1 +[3] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 2 +[4] Mario Emad, Michael Ishack, Mohamed Ahmed, Mohamed Osama, Mohamed Salah, and Ghada Khoriba. Early-anomaly prediction in surveillance cameras for security applications. In 2021 International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC), pages 124-128. IEEE, 2021. 1 +[5] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 1, 2 +[6] Zhipeng Fan, Jun Liu, and Yao Wang. Motion adaptive pose estimation from compressed videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11719-11728, 2021. 1, 2, 6 +[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 2 +[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 7 +[9] giampaolo. psutil. https://github.com/giampaolo/psutil. Accessed: 2022-4-25. 3 +[10]gpuopenanalytics. pynvml. https://github.com/gpuopenanalytics/pynvml. Accessed: 2022-4-25.3 +[11] Pratik Gujjar and Richard Vaughan. Classifying pedestrian actions in advance using predicted video of urban driving scenes. In 2019 International Conference on Robotics and Automation (ICRA), pages 2097-2103. IEEE, 2019. 1 +[12] Amirhossein Habibian, Davide Abati, Taco S Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2695–2704, 2021. 1, 2, 6, 7 +[13] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In International Conference on Learning Representations, 2021. 4 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1, 2 +[17] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 2 +[18] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11719-11727, 2019. 2 +[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6 +[20] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 1 +[21] Hayeon Lee, Sewoong Lee, Song Chong, and Sung Ju Hwang. Help: Hardware-adaptive efficient latency prediction for nas via meta-learning. arXiv preprint arXiv:2106.08630, 2021. 1 +[22] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 2 +[23] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 2 +[24] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 1 +[25] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 2, 6, 7 +[26] Yu Mitsuzumi, Go Irie, Daiki Ikami, and Takashi Shibata. Generalized domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1084-1093, June 2021. 7 +[27] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Maria Florina Balcan and Kil + +ian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1928-1937, New York, New York, USA, 20-22 Jun 2016. PMLR. 2 +[28] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 2 +[29] Xuecheng Nie, Yuncheng Li, Linjie Luo, Ning Zhang, and Jiashi Feng. Dynamic kernel distillation for efficient pose estimation in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6942-6950, 2019. 2, 6, 7 +[30] Toby Perrett and Dima Damen. Recurrent assistance: Cross-dataset training of lstms on kitchen tasks. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1354-1362. IEEE Computer Society, 2017. 6, 7 +[31] T. Perrett and D. Damen. Ddlstm: Dual-domain lstm for cross-dataset action recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7844-7853, 2019. 6, 7 +[32] Steven C Seow. Designing and engineering time: The psychology of time perception in software. Addison-Wesley Professional, 2008. 1 +[33] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 6 +[34] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 6 +[35] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 4 +[36] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 1 +[37] Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han. Hat: Hardware-aware transformers for efficient natural language processing. In ACL, 2020. 2 +[38] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the 17th Conference in Uncertainty in Artificial Intelligence, UAI '01, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 3 +[39] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3 +[40] Zuxuan Wu, Caiming Xiong, Yu-Gang Jiang, and Larry S Davis. Liteeval: A coarse-to-fine framework for resource + +efficient video recognition. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 6, 7 +[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 6, 7 +[42] Lumin Xu, Yingda Guan, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Vipnas: Efficient video pose estimation via neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16072-16081, 2021. 2 +[43] XuehaiPan. nvitop. https://github.com/XuehaiPan/nvitop. Accessed: 2022-4-25. 3, 6 +[44] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 1 +[45] Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization: Towards lossless cnns with low-precision weights. In International Conference on Learning Representations, ICLR2017, 2017. 2 +[46] Mohammadreza Zolfaghari, Kamaljeet Singh, and Thomas Brox. Eco: Efficient convolutional network for online video understanding. In Proceedings of the European conference on computer vision (ECCV), pages 695-712, 2018. 2 \ No newline at end of file diff --git a/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/images.zip b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..960d1bca8985c7b8d8dd23536d4900d54cfd6192 --- /dev/null +++ b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a89dc480d1ca28d2b44b6d30af4b8c545b25020bf153d8a2b951b70ce1b6e78 +size 367205 diff --git a/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/layout.json b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae24048624d2a972cd12671c495c5059c7b58e45 --- /dev/null +++ b/2023/System-Status-Aware Adaptive Network for Online Streaming Video Understanding/layout.json @@ -0,0 +1,9586 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 102, + 506, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 506, + 140 + ], + "type": "text", + "content": "System-status-aware Adaptive Network for Online Streaming Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "spans": [ + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "text", + "content": "Lin Geng Foo" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "text", + "content": " Jia Gong" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "text", + "content": " Zhipeng Fan" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{2§}" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "text", + "content": " Jun Liu" + }, + { + "bbox": [ + 150, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 175, + 477, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 175, + 477, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 477, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 175, + 477, + 190 + ], + "type": "text", + "content": "Singapore University of Technology and Design " + }, + { + "bbox": [ + 115, + 175, + 477, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 175, + 477, + 190 + ], + "type": "text", + "content": "New York University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 191, + 498, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 191, + 498, + 203 + ], + "spans": [ + { + "bbox": [ + 89, + 191, + 498, + 203 + ], + "type": "text", + "content": "{lingeng_foo, jia_gong}@mypail.sutd.edu.sg, zf606@nyu.edu,jun_riu@sutd.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 256, + 289, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 289, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 289, + 508 + ], + "type": "text", + "content": "Recent years have witnessed great progress in deep neural networks for real-time applications. However, most existing works do not explicitly consider the general case where the device's state and the available resources fluctuate over time, and none of them investigate or address the impact of varying computational resources for online video understanding tasks. This paper proposes a System-status-aware Adaptive Network (SAN) that considers the device's real-time state to provide high-quality predictions with low delay. Usage of our agent's policy improves efficiency and robustness to fluctuations of the system status. On two widely used video understanding tasks, SAN obtains state-of-the-art performance while constantly keeping processing delays low. Moreover, training such an agent on various types of hardware configurations is not easy as the labeled training data might not be available, or can be computationally prohibitive. To address this challenging problem, we propose a Meta Self-supervised Adaptation (MSA) method that adapts the agent's policy to new hardware configurations at test-time, allowing for easy deployment of the model onto other unseen hardware platforms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 128, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 551, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 288, + 696 + ], + "type": "text", + "content": "Online video understanding, where certain predictions are immediately made for each video frame by using information in the current frame and potentially past frames, is an important task right at the intersection of video-based research and practical vision applications (e.g., self-driving vehicles [11], security surveillance [4], streaming services [32], and human-computer interactions [20]). In particular, in many of these real-world video-based applications, a fast and timely response is often crucial to ensure high usability and reduce potential security risk. Therefore, in many practical online applications, it is essential to ensure that the model is working with low delay while maintaining a good" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 545, + 255 + ], + "type": "text", + "content": "performance, which can be challenging for many existing deep neural networks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 591 + ], + "type": "text", + "content": "Recently, much effort has been made to reduce the delay of deep neural networks, including research into efficient network design [16, 36, 44], input-aware dynamic networks [5,6,12,24], and latency-constrained neural architectures [1, 2, 21]. However, all these works do not explicitly consider the dynamic conditions of the hardware platform, and assume stable computation resources are readily available. In practical scenarios, the accessible computing resources of the host devices can be fluctuating and dynamic due to the fact that multiple computationally expensive yet important threads are running concurrently. For example, in addition to performing vision-related tasks such as object detection, human activity recognition, and pose estimation, state-of-the-art robotic systems usually need to simultaneously perform additional tasks like simultaneous localization and mapping (SLAM) to successfully interact with humans and the environment. Those tasks are also often computationally heavy and could compete with vision tasks for computing resources. As a result, at times when the host device is busy with other processes, conducting inference for each model might require significantly more time than usual, leading to extremely long delays, which could cause safety issues and lagging responses in many real-world applications. Therefore, the study and development of models providing reliable yet timely responses under various hardware devices and fluctuating computing resources is crucially important. Unfortunately, such studies are lacking in the field." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "To achieve and maintain low delay for online video understanding tasks under a dynamic computing resource budget, we propose a novel System-status-aware Adaptive Network (SAN). Different from previous works, SAN explicitly considers the system status of its host device to make on-the-fly adjustments to its computational complexity, and is thus capable of processing video streams effectively and efficiently in a dynamic system environment. SAN comprises of two components: a) a simple yet effective dynamic main module that offers reliable predictions under various" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 282, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 282, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 282, + 714 + ], + "type": "text", + "content": "† Equal contribution; § Currently at Meta; ‡ Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10514" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "network depths and input resolutions; b) a lightweight agent that learns a dynamic system-status-aware policy used to control the execution of the main module, which facilitates adaptation to the fluctuating system load. With the adaptivity of the main module and the control policy generated by the agent, our SAN can achieve good performance on the online video understanding task while maintaining a low delay under fluctuating system loads." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 359 + ], + "type": "text", + "content": "In various applications, we may need to deploy SAN onto different hardware platforms for online video understanding. However, it is inconvenient to train SAN for each hardware platform, and it might also be difficult to find adequate storage to load the large labeled dataset on all platforms (e.g., mobile devices). In light of these difficulties, we further propose a method for deployment-time self-supervised agent adaptation, which we call Meta Self-supervised Adaptation (MSA). With MSA, we can conveniently train a SAN model on a set of local platforms, and perform a quick deployment-time agent adaptation on a target device, without the need for the original labeled training data. Specifically, our proposed MSA introduces an auxiliary task of delay prediction together with a meta-learning procedure, that facilitates the adaptation to the target deployment device." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 360, + 274, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 360, + 274, + 372 + ], + "spans": [ + { + "bbox": [ + 59, + 360, + 274, + 372 + ], + "type": "text", + "content": "In summary, the main contributions of this paper are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 379, + 287, + 598 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 379, + 287, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 379, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 59, + 379, + 287, + 451 + ], + "type": "text", + "content": "- We are the first to explicitly consider the fluctuating system status of the hardware device at inference time for online video understanding. To address this, we propose SAN, a novel system-status-aware network that adapts its behavior according to the video stream and the real-time status of the host system." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 459, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 459, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 59, + 459, + 287, + 519 + ], + "type": "text", + "content": "- We further propose a novel Meta Self-supervised Adaptation method MSA that alleviates the training burden and allows our model to effectively adapt to new host devices with potentially unclear computation profiles at deployment time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 526, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 526, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 59, + 526, + 287, + 598 + ], + "type": "text", + "content": "- We empirically demonstrate that our proposed method achieves promising performance on the challenging online action recognition and pose estimation tasks, where we achieve low delays under a rapidly fluctuating system load without jeopardizing the quality of the predictions." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 609, + 134, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 134, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 134, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Online Video Understanding. Recently, motivated by the increasing real-world demand, a lot of works [5,6,29,40, 46] have attempted to improve the accuracy and efficiency of models for online video understanding tasks, such as online action recognition [23, 40, 46] and online pose estimation [6, 12, 29]. Several works design efficient networks to improve models' efficiency for online video tasks, including" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "content": "the efficient 3D CNN design [46], the Temporal Shift 2D CNN network [23] and the Skip-Convolution [12], while other researchers introduce adaptivity into the networks, termed as stream-aware dynamic networks [5, 25, 29] to further reduce the network's computation complexity while maintaining a good performance. For example, LiteEval [40] dynamically chooses between a coarse model and a fine model for each frame and MAPN [6] dynamically activates its encoder to save computation resources. However, all these existing methods do not explicitly consider the dynamic conditions of the hardware platform, and thus face limitations in the face of fluctuating system conditions. For example, in situations when the system is under high computational load, the model may still select to execute the branch with high complexity and therefore incur significantly longer delays, which is sub-optimal for time-sensitive applications. Compared to these methods, our proposed SAN is system-status-aware and thus more robust to computational fluctuations on the models' host devices. To the best of our knowledge, such a system-status-aware adaptive network that can efficiently tackle online video understanding tasks has not been explored before." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 350, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 547, + 543 + ], + "type": "text", + "content": "Efficient Network Designs. There are several paradigms for efficient deep architectures, and we list a few here. Efficient deep models such as SqueezeNet [17] and MobileNet [16] introduce novel computation operators using fewer parameters, while Neural Network Quantization [3,45] methods effectively shrink the model size of the existing models by quantizing model parameters to isolated values. Teacher-student models [15] are also effective in distilling knowledge from larger models into smaller ones. Some other latency-constrained networks [37, 42] further conduct neural architecture search (NAS) to search for an efficient architecture. Different from all these approaches, our SAN explicitly accounts for the dynamic system status that is constantly changing on the device over time, allowing us to perform efficient online video understanding even with a fluctuating system status." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Reinforcement Learning (RL). RL methods [22,27,28] train agents to navigate environments such that the reward is maximized. To allow pre-trained RL agents to adapt to a new environment in fewer episodes, methods based on the few-shot setting [7, 18] have been developed. Differently, we aim to allow our RL agent to adapt at deployment time in a self-supervised manner (i.e., no reward signals can be provided due to the absence of labeled video data). We introduce MSA, a novel method for self-supervised deployment-time adaptation. MSA employs an auxiliary task of delay prediction, and performs meta-learning to allow the RL agent to effectively adapt its control policy to the target hardware platform via fine-tuning on the auxiliary task." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10515" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 103, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "text", + "content": "Given the live video stream " + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "inline_equation", + "content": "V = \\{I_1, \\dots, I_t, \\dots\\}" + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "text", + "content": " and a seen/unseen host device whose system status " + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "inline_equation", + "content": "sys_t" + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "text", + "content": " varies over time, our goal is to build a model that dynamically adjusts the computational complexity of inference based on not only the streaming input, but also the system status to make high-quality predictions with a low delay at each step " + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 91, + 289, + 248 + ], + "type": "text", + "content": ". To this end, our proposed model: 1) takes into account the available computational resources to make on-the-fly decisions to dynamically adjust the computation consumption of the model, therefore maintaining the delay at low values while keeping the prediction's accuracy; and 2) effectively adapts itself to unseen devices in a self-supervised manner while still keeping its performance during deployment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 252, + 285, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 285, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 285, + 266 + ], + "type": "text", + "content": "3.1. SAN: System-status-aware Adaptive Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "text", + "content": "For better online performance amid the host device's varying background computational load, we propose a System-status-aware Adaptive Network (SAN). As shown in Fig. 1, our SAN consists of two main components: a lightweight agent " + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "text", + "content": " that decides how to process the incoming frame and a dynamic main network " + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "text", + "content": " that processes the frame with the specified policy generated by " + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 270, + 287, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "type": "text", + "content": "Dynamic main network. Our dynamic main network " + }, + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "type": "text", + "content": " flexibly produces frame-wise predictions using an architecture determined by the policy from the agent module " + }, + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 355, + 289, + 605 + ], + "type": "text", + "content": ". To accomplish this without extra burden on the hardware system, we design a dynamic encoder that has both dynamic depth and dynamic resolution. Our dynamic depth mechanism gives the encoder the option of producing output features at shallower intermediate layers without having to wait for all layers to be executed. This reduces computational cost by executing a \"simpler\" sub-network. On the other hand, the dynamic resolution mechanism allows our encoder to selectively take in input images at a lower resolution. Crucially, this reduces the amount of pixels to be examined by the network, effectively reducing the amount of convolution operations. We implement both of the aforementioned mechanisms in a single network, where to enable dynamic execution depth of the network, we add early exits to the network which allow the execution to end at an intermediate layer, while to enable dynamic resolution, we leverage the fully convolutional network and add a dynamic global pooling operation at the end." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 1, our dynamic encoder selects from " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " input resolutions " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{res^i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " model depths " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{dep^j\\}_{j=1}^n" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": ", allowing for as many as " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "m \\times n" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " different options in total. These " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "m \\times n" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " dynamic options have different computation complexities and levels of performance. For online video understanding, we construct our dynamic encoder as a 2D CNN for handling each frame, and to model the temporal aspect, the output features from the dynamic encoder are fed into a Long Short Term Memory network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "(LSTM) at every step. The LSTM updates its hidden state " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " based on the features of the incoming frame " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and its previous state " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", which is then used by the task head to determine the final prediction " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " for the video task following: " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\hat{y}_t = M(I_t, rest_t, dept_t, h_{t-1}; \\theta)" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " refers to the parameters of the main network " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "rest_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "dept_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " are selected by the agent introduced next. More details of our main dynamic network are in Sec. 4 and Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": "RL-based agent. The RL-based agent " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " controls the dynamic main network " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " by generating a frame-level processing policy " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "(\\text{rest}_t, \\text{depth}_t)" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " at each time " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": ", which aims to maintain the task-related accuracy " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\text{acc}_t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " at high level with low delay " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": ". To make decisions amid the fluctuating system loads, our lightweight agent " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " should be both system-status-aware and streaming input-aware. We model our task as a Markov Decision Process, where the agent " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " takes an action " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " at every step " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": ", based on the observed state " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": ", and transferred to the next state " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "s_{t+1}" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": ". Crucially, state " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " includes system status information " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\text{sys}_t" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " and input stream information " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": ", making " + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 168, + 546, + 337 + ], + "type": "text", + "content": " both system status and stream aware. To learn an optimal policy, we propose an RL-based method to train our agent." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": "Specifically, we set the observed states of the agent as " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "s_t = [h_{t-1}, sys_t, g_t]" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "sys_t" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " is the system status (e.g., the CPU and GPU utility)" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " is the LSTM hidden state, and " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " refers to other useful information (i.e., the previous step's action " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "a_{t-1}" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " and delay " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "d_{t-1}" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": "). Then, based on the available resolutions " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\{res^i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " and depths " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\{dep^j\\}_{j=1}^n" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ", we build the action space as " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\{a^{i,j} = (res^i, dep^j)\\}_{i=1,j=1}^{m,n}" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ". As shown in Fig. 1, at time " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ", the agent " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " receives " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "sys_t" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 304, + 336, + 545, + 459 + ], + "type": "text", + "content": ", and generates a probability distribution over the action space:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 467, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 467, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 319, + 467, + 545, + 483 + ], + "type": "interline_equation", + "content": "P _ {t} = \\left[ p _ {t} ^ {1, 1}, p _ {t} ^ {1, 2}, \\dots , p _ {t} ^ {m, n} \\right] = \\pi \\left(h _ {t - 1}, s y s _ {t}, g _ {t}; \\phi\\right) \\quad (1)", + "image_path": "9b309ceb21f1140ba4ccc3f5a51bd718cd3d6565c9cc12008351cbfd934fc1e6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "p_t^{i,j} \\in [0,1]" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " is the probability of the action " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "a^{i,j}" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "P_t \\in [0,1]^{m\\cdot n}" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " represents the probability distribution over the action space, and " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " refers to the parameters of the agent " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": ". Then, we take the highest probability entry in " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "P_t" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "p_t^{i*,j*}" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": "), select the corresponding action " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "a_t = (res^{i*}, dep^{j*})" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " and execute the main network accordingly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": "Training of SAN. To train our main network " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": ", we apply the loss " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " based on the specified online video-processing task. For example, for online action recognition, " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " can be the cross-entropy loss, while for online pose estimation, " + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 304, + 565, + 545, + 624 + ], + "type": "text", + "content": " can be the Mean Squared Error (MSE) loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "content": "To control the dynamic encoder to achieve a good accuracy at low delay, we apply an RL-based loss " + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\pi}" + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "content": " to train our agent " + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "content": ". Specifically, we adopt a policy gradient loss [38, 39]. To maximize the accuracy " + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "inline_equation", + "content": "acc_{t}" + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "content": " while reducing the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 713 + ], + "type": "text", + "content": "1We also collect other system parameters (e.g., the temperature of each processor, the memory usage and the status of I/O) via three linux utilities [9, 10, 43] to form system status syst. More details are in Supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10516" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 76, + 491, + 212 + ], + "blocks": [ + { + "bbox": [ + 97, + 76, + 491, + 212 + ], + "lines": [ + { + "bbox": [ + 97, + 76, + 491, + 212 + ], + "spans": [ + { + "bbox": [ + 97, + 76, + 491, + 212 + ], + "type": "image", + "image_path": "87cdabc49c873cf45f453eabff9f9662ea65ed252862a28e319b74006f782617.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": "Figure 1. Overview of our SAN framework, which consists of two modules: the agent module " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " and the dynamic main network " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": ". For each incoming frame " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": ", the agent module first takes in the LSTM hidden state " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": ", system-status " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "sys_{t}" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": ", and other useful information " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " and then decides the input resolution " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "rest_{t}" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " and the depth " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "depth_{t}" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " via the policy head. After that, the main network " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " processes the frame according to these decisions. By controlling the behavior of the main network " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": " via the agent " + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 213, + 547, + 280 + ], + "type": "text", + "content": ", our SAN can output high quality and low delay predictions in the dynamic system environment. Note that the auxiliary task head of the agent is used for fine-tuning of the agent during deployment, calibrating the learned system-status-aware features to adapt to the unseen target device." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "type": "text", + "content": "delay " + }, + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "type": "text", + "content": ", we set the reward of the agent at time " + }, + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 289, + 258, + 301 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 305, + 287, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 287, + 332 + ], + "type": "interline_equation", + "content": "r _ {t} = \\lambda_ {a c c} a c c _ {t} \\left(a _ {t}, I _ {t}\\right) - \\max \\left(d _ {t} \\left(a _ {t}, s y s _ {t}\\right) - d _ {b}, 0\\right) \\tag {2}", + "image_path": "6622d2c4c5e572d70d4c6e0984bbca7dd4274aa45f72b1d0ad27a7ddf60c4ff7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "acc_{t}(a_{t},I_{t})" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "d_t(a_t,sys_t)" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " are the accuracy and delay of our main network under the action " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": ", the current frame " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": ", and system status " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "sys_{t}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " is the coefficient of the accuracy term, and " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "d_b" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " is a tolerance threshold for the delay (e.g., " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "d_b = 30" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " ms for a rate of 30 fps), which is the acceptable delay. The policy gradient loss " + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\pi}" + }, + { + "bbox": [ + 46, + 335, + 287, + 407 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 412, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 412, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 121, + 412, + 287, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\pi} = \\sum_ {t = 1} ^ {T} - r _ {t} \\log \\left(p _ {t}\\right) \\tag {3}", + "image_path": "95107975a6e43557662b45ca39344292961661669f84286a8fbb9c068b7a9f63.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "inline_equation", + "content": "p_t \\in [0,1]" + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "text", + "content": " is the probability of selecting the action " + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 450, + 287, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 254, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 254, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 254, + 493 + ], + "type": "text", + "content": "3.2. MSA: Meta Self-supervised Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": "Ideally, we want to deploy SAN onto many different hardware platforms for online video applications, requiring us to train SAN for each of those platforms, which is labor intensive, computationally prohibitive, and may also be infeasible at times due to storage concerns (e.g., mobile devices). Instead, a more practical option would be to effectively fine-tune a pre-trained SAN at deployment-time on a target device. However, during deployment, the ground-truth labels associated with the real-time video stream are often inaccessible. To address this issue, we introduce an auxiliary task [13, 35] of delay prediction, where the agent " + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": " is expected to additionally predict the actual processing time based on the selected action (i.e., the input resolution and execution depth of main network " + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Our intuition is that, on a new (unseen) platform, our agent " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " lacks understanding of hardware status of this platform, thus we need a way for the agent to quickly understand it. Hence, we introduce the auxiliary task of delay" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "type": "text", + "content": "prediction, because delay times are directly related to the system status and the model's computational cost. Therefore, when the agent " + }, + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "type": "text", + "content": " is fine-tuned using delay prediction, it will be pushed to adopt a better understanding of the host platform's characteristics and the costs of different options of " + }, + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 289, + 545, + 433 + ], + "type": "text", + "content": " on that platform. Moreover, delay times can be directly observed, and can be a convenient supervision signal on a new platform since we do not have signals from our main task (which requires labeled video data). Therefore, on a new platform, we can fine-tune our agent to predict the delay by using the observed delay times as supervision, which will improve the agent's understanding of the device." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 546, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 603 + ], + "type": "text", + "content": "Although the delay prediction task improves the quality of the intermediate features (which are also used to select actions), fine-tuning with it does not necessarily lead to a better policy on the host device. Hence, to better align the performance on the auxiliary task of delay prediction with policy learning, we further propose a meta-optimization-based learning framework to find a good initialization for the agent, such that updating based on the loss of delay prediction would also improve the quality of the policy. Combining the auxiliary task and the meta-learning framework allows us to train SAN on a set of local hardware platforms, and perform only a quick deployment-time adaptation on the target device at run-time, without requiring the actual labeled video dataset on that target device." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Auxiliary task of delay prediction. When SAN is deployed to an unseen target device, the processing times of each action at a given system status will shift significantly due to the drift in computing profiles. Therefore, fine-tuning on the target device is required to adapt SAN to the new hardware characteristics. Due to the lack of ground-truth labels for the streaming videos at run-time, we introduce a self-supervised auxiliary task of delay prediction for adapting SAN to unseen hardware platforms. The delay predic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10517" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "tion task calibrates the learned system-status-aware features for the agent module " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " using supervisory signals (processing times) that can be generated on-the-fly with only a simple timer on the device." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": "Specifically, an auxiliary task head is added to the agent " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": ", as shown in Fig. 1. The auxiliary task head takes in the features from the agent as well as the chosen action " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": ", to predict the delay time " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": ". To train our model on the auxiliary task, we apply a mean-squared-error loss between our predictions " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\hat{d}_t" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": " and the observed delay " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": ". The auxiliary loss " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux}" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": ", is thus defined as: " + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aux,t} = (d_t - \\hat{d}_t)^2" + }, + { + "bbox": [ + 47, + 120, + 286, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 286, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 286, + 286 + ], + "type": "text", + "content": "At deployment time, we fine-tune the agent " + }, + { + "bbox": [ + 46, + 204, + 286, + 286 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 204, + 286, + 286 + ], + "type": "text", + "content": " on the target device using this auxiliary task only, allowing the agent to better understand the target device's characteristics and processing times. However, this does not necessarily mean that the agent has optimized its policy output – in order to achieve that, we further introduce a meta-optimization method as described next." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 288, + 286, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 286, + 383 + ], + "type": "text", + "content": "Meta-optimization for self-supervised agent adaptation. During deployment onto the target device, optimizing the auxiliary task of delay prediction is loosely connected to improving the actual policy. To better bind the quality of the developed policy to the delay prediction performance, we introduce a meta-optimization step to our agent " + }, + { + "bbox": [ + 46, + 288, + 286, + 383 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 288, + 286, + 383 + ], + "type": "text", + "content": " during training, such that fine-tuning of the auxiliary task will also lead to a better adaptation of the policy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "text", + "content": "We approach this problem by simulating the fine-tuning process using the delay prediction error on multiple target platforms, and (meta) optimize " + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "text", + "content": " such that the update based on the delay prediction error leads to better adaptation of the policy as well. This will effectively link the quality of the policy to the delay prediction performance. Here, we adopt a meta-learning framework to tackle the problem, treating " + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "text", + "content": " as the meta-parameters to meta-optimize. In this context, we denote the auxiliary task as the meta-train task and the agent's policy generation as the meta-test task. Our meta-optimization algorithm aims to learn a good initialization of " + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 384, + 286, + 538 + ], + "type": "text", + "content": ", such that when it is updated via the meta-train task, it can achieve good performance on the meta-test task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": "We assume that our SAN is first trained on a source device " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "D_{src}" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": ", and then adapted to the additional " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": " target platforms " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "\\{D_{tgt}^k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": ". After first training our SAN on the source device " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "D_{src}" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": ", we freeze the main network " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": ", and keep a copy of the agent's parameters (denoted as " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "\\phi_{meta,0}" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": ") to initialize the meta-learning process. In the " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": "-th meta-learning iteration, we first meta-train the agent for each target device " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "D_{tgt}^k" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": " using its loss on the auxiliary task and perform a pseudo-update on the meta parameters " + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "inline_equation", + "content": "\\phi_{meta,u}" + }, + { + "bbox": [ + 46, + 539, + 286, + 658 + ], + "type": "text", + "content": " which can be formalized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 663, + 286, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 663, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 52, + 663, + 286, + 678 + ], + "type": "interline_equation", + "content": "\\phi_ {t g t, u} ^ {k} \\leftarrow \\phi_ {\\text {m e t a}, u} - \\alpha \\nabla_ {\\phi_ {\\text {m e t a}, u}} \\mathcal {L} _ {\\text {a u x}, t} \\left(\\phi_ {\\text {m e t a}, u}, D _ {t g t} ^ {k}\\right) \\tag {4}", + "image_path": "5d2bfbf347d4178834c139c24d596b3cda000a320c1444a59574d7d3d144842f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "inline_equation", + "content": "\\phi_{tgt}^k" + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "text", + "content": " refers to the parameters of the meta-trained agent " + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "inline_equation", + "content": "\\pi_{tgt}^k" + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "text", + "content": " which is about to be deployed onto the target device " + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "inline_equation", + "content": "D_{tgt}^k" + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 72, + 545, + 111 + ], + "type": "text", + "content": " is the learning rate for meta-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 111, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 111, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 111, + 545, + 159 + ], + "type": "text", + "content": "Then, we check the performance of the pseudo-updated parameters on target device by meta-testing the performance of the pseudo-updated agent " + }, + { + "bbox": [ + 305, + 111, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\pi_{tgt}^{k}(:,\\phi_{tgt,u}^{k})" + }, + { + "bbox": [ + 305, + 111, + 545, + 159 + ], + "type": "text", + "content": " for the agent's policy generation following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 167, + 545, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 167, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 354, + 167, + 545, + 182 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m e t a}} ^ {k} \\left(\\phi_ {t g t, u} ^ {k}\\right) = \\mathcal {L} _ {\\pi} \\left(\\phi_ {t g t, u} ^ {k}, D _ {t g t} ^ {k}\\right), \\tag {5}", + "image_path": "7ab3795e8be9977f3bfa63f2bc82edc01f21c192ac8da613ff2333853d12d558.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{meta}^k" + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "text", + "content": " is the meta-testing-loss from the target hardware platform " + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "inline_equation", + "content": "D_{tgt}^k" + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "text", + "content": ". Intuitively, this loss penalizes our agent based on its performance on a target device after a short period of fine-tuning. Hence, we update our meta parameters " + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\phi_{meta,u}" + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 305, + 191, + 545, + 252 + ], + "type": "text", + "content": "-th meta-learning iteration as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 261, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 309, + 261, + 545, + 304 + ], + "type": "interline_equation", + "content": "\\phi_ {m e t a, u + 1} \\leftarrow \\phi_ {m e t a, u} - \\beta \\sum_ {k = 1} ^ {K} \\nabla_ {\\phi_ {m e t a, u}} \\mathcal {L} _ {m e t a} ^ {k} \\left(\\phi_ {t g t, u} ^ {k}\\right), \\tag {6}", + "image_path": "b5b3101cd4bc953b5ca25ce0bed69a628f509f1f8755ae98725d0dac30d663fc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "text", + "content": " is the learning rate for optimizing the meta parameters, and " + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\phi_{tgt,u}^k" + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "text", + "content": " is obtained via the pseudo-update in Eq. 4. Through updating our model with this meta-learning scheme, our agent parameters will converge towards optimal parameters " + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\phi_{meta,*}" + }, + { + "bbox": [ + 305, + 304, + 545, + 376 + ], + "type": "text", + "content": " that are able to effectively adapt to each target device through fine-tuning via the auxiliary task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 384, + 426, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 426, + 397 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 426, + 397 + ], + "type": "text", + "content": "3.3. Training and Testing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 403, + 544, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 544, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 544, + 426 + ], + "type": "text", + "content": "As mentioned above, we consider two scenarios: deploying our SAN on a seen or unseen device." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": "Model training and testing for seen device. In this scenario, we evaluate SAN when it is trained and tested on the same device, which evaluates the effectiveness of our SAN design described in Sec. 3.1. For model training, we first utilize the dataset's training set to initialize our main network " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": ", which is controlled by a random policy (randomly sampling the resolution-depth pair) to improve its accuracy via the loss " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": ". Then, we train our agent " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": " using the loss " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\pi}" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": " with the parameters of the main network " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": " fixed. Lastly, we fine-tune the main network " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": " and agent " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": " jointly in a fully end-to-end manner with both " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\pi}" + }, + { + "bbox": [ + 305, + 426, + 545, + 629 + ], + "type": "text", + "content": ", allowing us to obtain a high-performance video understanding dynamic network controlled by an optimal system-status-aware policy on the current local device. For evaluation, we test the trained SAN on the same device, but with an unseen system status trajectory, which evaluates both the accuracy and the delay of our SAN." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Model meta-optimization and deployment for unseen device. In this scenario, our SAN is pre-trained and meta-optimized on a set of seen source devices, and we evaluate its efficacy when deployed to an unseen device. This evaluates the efficacy of our MSA design described in Sec. 3.2. At the start of training, we first initialize and pre-train our SAN on a seen device, following the process" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "type": "text", + "content": "2We empirically find that we can simulate a large set of target platforms by setting different utility limitations on a few selected devices, as described in our experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10518" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "described previously. Then, during meta-optimization, we freeze the main network " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " and only optimize our agent " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ". During meta-optimization iterations, we randomly sample video clips from the training set for meta-training and meta-testing. For better generalization, video clips used for meta-training are not used for meta-testing. When deploying our SAN to an unseen target device, we first fine-tune the agent " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " via the auxiliary task of delay prediction. The adapted agent is then tested for average accuracy and delay over all frames in the testing set. When comparing our method's average accuracy and delay with others, we compute the global accuracy and delay (counting both the fine-tuning and final testing phase) for a fair comparison." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 128, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 128, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 128, + 252 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 258, + 287, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 283 + ], + "type": "text", + "content": "We describe our setup and results below. More analysis and details are provided in Supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": "SAN's architecture. For our dynamic encoder, we adopt the fully-convolutional backbone ResNet50 [14] to handle various input resolutions. We add three exit ports (i.e., " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": ") at the end of these specific layers of ResNet50 (" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\{conv3_x, conv4_x, conv5_x\\}" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": " [14]) to achieve dynamic depth. Moreover, we specifically design a convolutional layer for each resolution-depth pair to unify the shapes of the encoder's output features (details in Supplementary). Global average pooling is performed on the unified encoder's output and send to the LSTM to update its hidden state " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}^{2048}" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": ". The output of the LSTM is fed into the task head for online video understanding. On the other hand, our lightweight agent " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": " consists of three fully connected layers. We utilize three widely used monitors [43] to collect system information, which contains the status of system's CPU, GPU, memory, disk, I/O and fan. For model training, we set " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\lambda_{acc} = 2" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "d_b = 0.03" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\alpha = 1e^{-3}" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "inline_equation", + "content": "\\beta = 1e^{-5}" + }, + { + "bbox": [ + 46, + 283, + 289, + 498 + ], + "type": "text", + "content": ". More details of our SAN are presented in Supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "Online video task settings. We compare the performance of our SAN with the prior methods on two widely used online video understanding tasks: 1) Online action recognition: We follow the previous work [30] to use the 50Salads dataset [34], since it contains very long videos that are suitable for online video action recognition. Here, we consider three (i.e., " + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": ") resolution candidates: [112, 168, 224] and use a fully-connected layer as the task head. We compare our method with SOTA online action recognition methods, including RA [30], DDLSTM [31], AR-Net [25], and LiteEval [40]. To test models on the same platform for a fair comparison, we use their publicly released code (AR-Net) when available and implement the rest (RA, DDLSTM and LiteEval) on our own. 2) Online pose estimation: We follow the existing video pose estimation works [6, 12] to conduct our experiments on the Sub-JHMDB dataset [19], a collection of 11,200 frames from 316 video clips, labeled with 15 body" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 82, + 544, + 168 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 81 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 81 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 81 + ], + "type": "text", + "content": "Table 1. Results on 50Salads dataset for online action recognition." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 82, + 544, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 82, + 544, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 82, + 544, + 168 + ], + "type": "table", + "html": "
MethodAccuracyMax DelayMean DelayR\\T Frames
RA [30]42.6 %76.6 ms33.9 ms79.1 %
DDLSTM [31]41.1 %407.4 ms110.3 ms5.9 %
LiteEval [40]40.3 %110.7 ms68.9 ms30.4 %
AR-Net [25]40.9 %91.1 ms56.2 ms62.1 %
main network + random policy46.2%80.5 ms34.1 ms92.1%
+ stream-aware55.4 %80.8 ms40.3 ms82.0%
+ SAN53.8%49.4 ms29.9 ms95.4%
", + "image_path": "bfe88da5c1cc5de6cb4798a6b1d6305f41a58e1d631ad283a9601253dca6dc1b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 171, + 428, + 270 + ], + "blocks": [ + { + "bbox": [ + 306, + 171, + 428, + 270 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 428, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 428, + 270 + ], + "type": "image", + "image_path": "4f978f910c782b02b1d984314c6dca1a72f226867e5acfc2d996196953c52f56.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 270, + 428, + 323 + ], + "blocks": [ + { + "bbox": [ + 306, + 270, + 428, + 323 + ], + "lines": [ + { + "bbox": [ + 306, + 270, + 428, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 428, + 323 + ], + "type": "image", + "image_path": "ea8ecf390824b23a0fe233472130cdc99508ce257a346d3abe3de93c976689f5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 324, + 546, + 380 + ], + "lines": [ + { + "bbox": [ + 304, + 324, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 546, + 380 + ], + "type": "text", + "content": "Figure 2. A trace of models' accuracy and delay for online video-based action recognition (left) and pose estimation (right). In the areas masked by the gray color, we can observe a rapid increase of the baseline model's delay (see the second row), which corresponds to the high-load status of the system." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 171, + 547, + 269 + ], + "blocks": [ + { + "bbox": [ + 428, + 171, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 428, + 171, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 428, + 171, + 547, + 269 + ], + "type": "image", + "image_path": "3657466d63a82059928f570e2e1b2d95081a753e2646908fc6a481b43ebb7d98.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 428, + 269, + 547, + 323 + ], + "blocks": [ + { + "bbox": [ + 428, + 269, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 428, + 269, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 428, + 269, + 547, + 323 + ], + "type": "image", + "image_path": "a5bf85d570bbd53c2484eb1bec05d789a25c635e9680adf339f5f62a8f7053fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "type": "text", + "content": "joints. Here, we consider three (i.e., " + }, + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 304, + 389, + 545, + 498 + ], + "type": "text", + "content": ") resolution candidates: [128, 192, 256] and follow the work of [41] to use 3 deconvolutional layers to build the task head for pose estimation. Moreover, we replace LSTM with a ConvLSTM [33] to accommodate for 2D feature maps. We compare our method with the existing dynamic pose estimation networks, i.e., DKD [29] and Skip-Convolution [12]. Similarly, we re-implement DKD and Skip-Convolutions as there is no publicly available code." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "Experiment settings. To build dynamic systems with varying system status, we design multiple background processes (i.e., matrix calculators, video compressors, and large deep learning models) to occupy various amounts of computational resources. We can then generate dynamic system load trajectories using these processes to simulate dynamic system load environments. During model training, we randomly generate dynamic system trajectories in each iteration to simulate various dynamic environments to train our SAN. During testing, for fair comparisons, we first generate 3 dynamic trajectories unseen in training, and then run all the evaluation experiments on these 3 fixed trajectories and report the average result. To evaluate the models' efficiency, we measure the maximum delay and mean delay for each model. Moreover, as the videos of both datasets [19, 34] are at 30 fps, we consider delay time " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "< 30" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " ms as real-time, and measure the percentage of such real-time predictions (R\\T Frames)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10519" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 82, + 544, + 160 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 420, + 81 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 420, + 81 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 420, + 81 + ], + "type": "text", + "content": "Table 2. Results on Sub-JHMDB dataset for online pose estimation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 82, + 544, + 160 + ], + "lines": [ + { + "bbox": [ + 48, + 82, + 544, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 82, + 544, + 160 + ], + "type": "table", + "html": "
MethodHeadSho.Elb.Wri.HipKneeAnk.AvgMax DelayMean DelayR\\T Frames
DKD [29]98.396.690.487.199.196.092,994.0133.8 ms47.9 ms59.9 %
SimpleBaseline [41]97.597.891.186.099.696.892.694.4129.6 ms45.5 ms63.8 %
Skip-Convolution [12]98.797.792.088.199.396.691.095.1115.7 ms51.2 ms53.7 %
main network + random policy97.595.291.493.577.786.872.486.9130.9 ms36.1 ms72.9%
main network + stream-aware99.498.697.598.193.994.290.095.3126.1 ms44.7 ms63.7%
main network + SAN99.198.797.398.290.192.585.093.441.2 ms33.1 ms81.4%
", + "image_path": "313b31f77a8e654d60fcc0e9ee13886f5a5a4a1944bd5b41be74c015e05dc295.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 219, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 219, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 219, + 180 + ], + "type": "text", + "content": "4.1. Evaluating SAN on seen devices" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "content": "First, we compare our SAN with the SOTA methods for both tasks and report the results in Tab. 1 and Tab. 2. In both tables, we also report the performance of two baselines of our method: main network + random policy where actions are randomly selected; and main network + stream-aware where the agent " + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 186, + 288, + 281 + ], + "type": "text", + "content": " uses information from the input stream without considering system status. Both baselines are not system-status-aware." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 282, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 498 + ], + "type": "text", + "content": "Online action recognition. Tab. 1 presents the results on the 50Salads dataset for online action recognition. We compare our method with four SOTA methods, including two static models (i.e., RA [30] and DDLSTM [31]) and two efficient dynamic models (i.e., AR-Net [25] and LiteEval [40]). Our SAN outperforms these methods on all metrics, achieving a higher accuracy while maintaining a low mean delay (29.9ms) and high percentage of real-time frames (95.4%). These improvements are due to our system-status-aware design of SAN, allowing it to make on-the-fly adaptations to process a higher percentage of frames in real-time. The two baselines are not system-status-aware and therefore perform significantly worse on the delay-related metrics, especially on max delay (80.5ms and 80.8ms vs our 49.4ms). Compared to the random policy, stream-aware policy significantly improves the accuracy at the cost of reduced real-time performance. Overall, our SAN achieves superior performance and at the same time also decreases the delay." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 618 + ], + "type": "text", + "content": "Online pose estimation. Tab. 2 presents the results on the Sub-JHMDB dataset for online pose estimation. We compare our SAN with a static network SimpleBaseline [41] and two dynamic efficient networks: DKD [29] and Skip-Convolution [12]. As shown in Tab. 2, our main network + stream-aware baseline can achieve the SOTA performance but may still incur large delays occasionally. With the help of our system-status-aware design, our SAN can control the maximum and mean delay at a low level while still maintaining high accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Qualitative analysis. We analyze the effect of the system load dynamics and show results from test-time runs on a laptop with a NVIDIA Geforce GTX 1080 GPU (11 GB), where 10 background processes are randomly activated. Fig. 2 presents the models' frame-level delay and accuracy along with the system status. We compare our SAN with two baselines: main network + stream-aware and main network + random policy. In the first row of Fig. 2," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 168, + 547, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 324 + ], + "type": "text", + "content": "we observe that the GPU load fluctuates over time, varying the available computational resources for the online models. As a result, the delay of models with random/stream-aware policy fluctuates up to around 100ms as shown in the second row. In contrast, by using the system-status information to control the model, our SAN is able to stabilize the delay and avoid large delays amid the dynamic system. Moreover, as shown in the last row, with the help of the stream-aware design, our SAN can still obtain a similar accuracy to the stream-aware model. These results show the advantage of our system-status-aware and stream-aware designs, which can simultaneously control the real-time delay and attain high accuracy for online video understanding tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 334, + 541, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 541, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 541, + 346 + ], + "type": "text", + "content": "4.2. Evaluating SAN with MSA on unseen devices" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": "Our proposed MSA is introduced to achieve effective adaptation to the unseen host devices. Here, we compare our MSA with three baselines: 1) Direct transfer: Directly deploying our SAN to the unseen device. 2) Feature alignment: Minimizing the domain gap between the source device and target device via a domain discriminator [8, 26]. 3) Self-supervised fine-tune: Using our auxiliary task (delay prediction) to fine tune the model without meta optimization. To further demonstrate the efficacy of MSA, we also present the upper bound, dubbed as Fully-supervised learning, where we directly train the agent on the target environment, which eliminates the need for adaptation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": "Here, we collect three hardware platforms for training and testing: 1) Device a: a laptop with an AMD Ryzen Threadripper 2950X and an NVIDIA Geforce GTX 1080 GPU (11 GB). 2) Device b: a desktop with an Intel(R) Core(TM) i9-10900 and an NVIDIA Geforce RTX 2080 GPU (12 GB). 3) Device c: a workstation with an AMD Ryzen Threadripper 3960X and an NVIDIA Geforce RTX 3090 GPU (24 GB). We randomly sample two platforms as the seen source devices for model initialization and meta-optimization and use the remaining one as the unseen target device for model adaptation (e.g., Source: " + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{a,b\\} \\to" + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": " Target: " + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{c\\}" + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": "). Furthermore, we augment the source platforms by restricting the number of available CPU and the highest frequencies of CPU and GPU to achieve various environments with different resource configurations for better training. The results of policy adaptation are reported in Tab. 3 and 4. More results on other portable devices (e.g., NVIDIA Jetson TX2) are presented in Supplementary." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10520" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 83, + 544, + 145 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 541, + 82 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 541, + 82 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 541, + 82 + ], + "type": "text", + "content": "Table 3. Results of policy adaptation across platforms for online pose estimation. More results on other platforms are in Supplementary." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 83, + 544, + 145 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 544, + 145 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 544, + 145 + ], + "type": "table", + "html": "
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T Frames
Fully-supervised learning (Upper bound)93.0 %28.3 ms87.4 %93.1 %24.7 ms84.9 %92.8 %28.1 ms74.5 %
Direct transfer87.3 %28.4 ms82.4 %85.7 %25.0 ms79.5 %90.9 %33.5 ms20.4 %
Feature alignment90.8 %29.0 ms80.4 %91.0 %27.2 ms74.6 %91.0 %31.8 ms64.5 %
Our self-supervised fine-tune91.1 %29.8 ms79.8 %85.8 %21.9 ms84.1 %91.5 %29.3 ms61.7 %
Our MSA92.8 %29.7 ms86.1 %92.9 %25.1 ms82.2 %92.4 %30.0 ms71.9 %
", + "image_path": "644dfd6605fbe926d8b592474a095f9b586ca0aabc2ec5c79525b11db0a5f9e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 156, + 544, + 220 + ], + "blocks": [ + { + "bbox": [ + 145, + 146, + 447, + 156 + ], + "lines": [ + { + "bbox": [ + 145, + 146, + 447, + 156 + ], + "spans": [ + { + "bbox": [ + 145, + 146, + 447, + 156 + ], + "type": "text", + "content": "Table 4. Results of policy adaptation across platforms for online action recognition." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 156, + 544, + 220 + ], + "lines": [ + { + "bbox": [ + 48, + 156, + 544, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 156, + 544, + 220 + ], + "type": "table", + "html": "
Methoda+b → ca+c → bb+c → a
AccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T FramesAccuracyMean DelayR\\T Frames
Fully-supervised learning (Upper bound)53.8 %26.3 ms95.4 %52.6 %31.3 ms93.1 %52.1 %31.9 ms90.4 %
Direct transfer41.7 %28.1 ms77.3 %47.2 %34.6 ms84.7 %52.0 %41.1 ms40.6 %
Feature alignment50.1 %29.8 ms82.4 %50.3 %32.7 ms88.1 %51.7 %37.2 ms60.5 %
Our self-supervised fine-tune49.1 %26.5 ms93.8 %51.0 %32.2 ms93.1 %50.7 %33.7 ms87.2 %
Our MSA52.6 %26.8 ms94.1 %52.2 %31.6 ms92.7 %51.9 %32.1 ms90.2 %
", + "image_path": "c6c7026ece52415f9a1393e45f19ac5923ab183b6b47ec9d9d5c9718fdf07581.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 231, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 399 + ], + "type": "text", + "content": "As shown in Tab. 3 and 4, the direct transfer method fails to maintain the high accuracy and low delay simultaneously due to the hardware configuration shift between devices. On the other hand, by aligning the system-status-aware features between the source and target devices, the agent finds a suboptimal policy to balance the accuracy and delay. Using the auxiliary task to fine-tune the agent without meta optimization can achieve real-time predictions on the target device, but it sometimes fails to maintain high accuracy. Finally, our MSA simultaneously maintains a high ratio of real-time predictions and accurate predictions. Furthermore, the results are close to the upper bound of actually training on the target device, showing the effectiveness of our MSA to generalize to unseen devices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 141, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 141, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 141, + 420 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "type": "text", + "content": "We further analyze the characteristics of our dynamic resolution and dynamic depth mechanism, and investigate the performance of our agent under various delay thresholds " + }, + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "type": "inline_equation", + "content": "d_{b}" + }, + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "type": "text", + "content": " and accuracy coefficients " + }, + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 46, + 426, + 287, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 581 + ], + "type": "text", + "content": "First, we evaluate the impact of the dynamic resolution and depth mechanism and perform ablation studies by allowing one of them to be dynamic while fixing the other. In Tab. 5, we observe that, by allowing for dynamic resolutions or dynamic depths, our main network can reduce the maximum and mean delay while maintaining high accuracy. Moreover, by combining these two mechanisms, our SAN can further reduce the mean delay and suppress the maximum delay at a lower level, with even better accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Next, we investigate the performance of our SAN under different delay thresholds " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{b}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and accuracy coefficients " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ". As shown in Tab. 6, under different values of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{b}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", our SAN can still significantly decrease the maximum delay while maintaining high accuracy. Furthermore, as " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{b}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " increases, SAN is allowed to have higher delays before getting penalized, leading to higher accuracy while slightly degrading delay metrics, which is what we expected. We also evaluate different values of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " and present the results in Tab. 6. We observe that our SAN generally achieves good accuracy and delay performance for different values of " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": ", showing that" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 241, + 544, + 309 + ], + "blocks": [ + { + "bbox": [ + 334, + 230, + 517, + 240 + ], + "lines": [ + { + "bbox": [ + 334, + 230, + 517, + 240 + ], + "spans": [ + { + "bbox": [ + 334, + 230, + 517, + 240 + ], + "type": "text", + "content": "Table 5. Ablation results for dynamic mechanism." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 241, + 544, + 309 + ], + "lines": [ + { + "bbox": [ + 307, + 241, + 544, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 544, + 309 + ], + "type": "table", + "html": "
SettingAvgMax DelayMean Delay
res = 256, dep = 394.4 %129.6 ms45.5 ms
res = 256, dep = [1,3]92.6 %71.4 ms39.3 ms
res = 256, dep = [1,2,3]93.3 %63.6 ms34.5 ms
res = [128,256], dep = 391.2 %84.0 ms43.7 ms
res = [64,128,192,256], dep = 391.9 %61.9 ms38.0 ms
Our SAN93.4 %41.2 ms33.1 ms
", + "image_path": "5084f73a13bcdf2dd6c55aa6cfc48141a99c914ceaa3e9ac341f0e68f1af6d3d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 331, + 544, + 407 + ], + "blocks": [ + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "lines": [ + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "type": "text", + "content": "Table 6. Ablation results for different values of " + }, + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "type": "inline_equation", + "content": "d_b" + }, + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 316, + 319, + 534, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 331, + 544, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 331, + 544, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 544, + 407 + ], + "type": "table", + "html": "
SettingAvgMax DelayMean Delay
db=10 ms92.4 %39.2 ms27.7 ms
=20 ms92.8 %39.4 ms30.7 ms
=30 ms93.4 %41.2 ms33.1 ms
=40 ms94.0 %43.5 ms41.9 ms
λacc=191.5 %41.0 ms30.8 ms
=293.4 %41.2 ms33.1 ms
=494.0 %44.6 ms37.1 ms
", + "image_path": "e8ad647ade62484aac66c4be4c0c5eef19a2b8861e9a7f3e06c13c77a4059b2b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "type": "text", + "content": "it is robust to the " + }, + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "type": "text", + "content": " hyperparameter. Also, when " + }, + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\lambda_{acc}" + }, + { + "bbox": [ + 305, + 422, + 545, + 459 + ], + "type": "text", + "content": " increases, the accuracy improves and delay slightly degrades, as we would expect." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 481, + 379, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 379, + 494 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 379, + 494 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": "In this work, we tackle the efficiency problem of deep networks on devices with computational fluctuations. To the best of our knowledge, this is the first time this problem is being explicitly tackled for online video understanding tasks. We propose SAN, a novel system-status-aware adaptive network for low-delay online action recognition and online pose estimation tasks with a fluctuating computation budget. A self-supervised meta optimization framework (MSA) is also proposed for more effective adaptation between hardware platforms. Experiments show that our proposed SAN and MSA obtains SOTA performance with low delay even under a rapidly fluctuating system load." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by MOE AcRF Tier 2 (Proposal ID: T2EP20222-0035), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2020-065), and SUTD SKI Project (SKI 2021_02_06)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10521" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. In International Conference on Learning Representations, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "type": "text", + "content": "[3] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 281 + ], + "type": "text", + "content": "[4] Mario Emad, Michael Ishack, Mohamed Ahmed, Mohamed Osama, Mohamed Salah, and Ghada Khoriba. Early-anomaly prediction in surveillance cameras for security applications. In 2021 International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC), pages 124-128. IEEE, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Zhipeng Fan, Jun Liu, and Yao Wang. Adaptive computationally efficient network for monocular 3d hand pose estimation. In European Conference on Computer Vision, pages 127-144. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 373 + ], + "type": "text", + "content": "[6] Zhipeng Fan, Jun Liu, and Yao Wang. Motion adaptive pose estimation from compressed videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11719-11728, 2021. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 417 + ], + "type": "text", + "content": "[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "type": "text", + "content": "[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 475, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 496 + ], + "type": "text", + "content": "[9] giampaolo. psutil. https://github.com/giampaolo/psutil. Accessed: 2022-4-25. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 530 + ], + "type": "text", + "content": "[10]gpuopenanalytics. pynvml. https://github.com/gpuopenanalytics/pynvml. Accessed: 2022-4-25.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "text", + "content": "[11] Pratik Gujjar and Richard Vaughan. Classifying pedestrian actions in advance using predicted video of urban driving scenes. In 2019 International Conference on Robotics and Automation (ICRA), pages 2097-2103. IEEE, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "text", + "content": "[12] Amirhossein Habibian, Davide Abati, Taco S Cohen, and Babak Ehteshami Bejnordi. Skip-convolutions for efficient video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2695–2704, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[13] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A. Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[15] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 185 + ], + "type": "text", + "content": "[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 240 + ], + "type": "text", + "content": "[17] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[18] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11719-11727, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "text", + "content": "[20] Hema S Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE transactions on pattern analysis and machine intelligence, 38(1):14-29, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 420 + ], + "type": "text", + "content": "[21] Hayeon Lee, Sewoong Lee, Song Chong, and Sung Ju Hwang. Help: Hardware-adaptive efficient latency prediction for nas via meta-learning. arXiv preprint arXiv:2106.08630, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 466 + ], + "type": "text", + "content": "[22] Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[23] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7083-7093, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 567 + ], + "type": "text", + "content": "[24] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 624 + ], + "type": "text", + "content": "[25] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European Conference on Computer Vision, pages 86-104. Springer, 2020. 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[26] Yu Mitsuzumi, Go Irie, Daiki Ikami, and Takashi Shibata. Generalized domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1084-1093, June 2021. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Maria Florina Balcan and Kil" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10522" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "ian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1928-1937, New York, New York, USA, 20-22 Jun 2016. PMLR. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[28] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 226 + ], + "type": "text", + "content": "[29] Xuecheng Nie, Yuncheng Li, Linjie Luo, Ning Zhang, and Jiashi Feng. Dynamic kernel distillation for efficient pose estimation in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6942-6950, 2019. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "text", + "content": "[30] Toby Perrett and Dima Damen. Recurrent assistance: Cross-dataset training of lstms on kitchen tasks. In 2017 IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1354-1362. IEEE Computer Society, 2017. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "text", + "content": "[31] T. Perrett and D. Damen. Ddlstm: Dual-domain lstm for cross-dataset action recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7844-7853, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "text", + "content": "[32] Steven C Seow. Designing and engineering time: The psychology of time perception in software. Addison-Wesley Professional, 2008. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "text", + "content": "[33] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[34] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "text", + "content": "[35] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "text", + "content": "[36] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International Conference on Machine Learning, pages 10096-10106. PMLR, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[37] Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han. Hat: Hardware-aware transformers for efficient natural language processing. In ACL, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[38] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the 17th Conference in Uncertainty in Artificial Intelligence, UAI '01, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[39] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229-256, 1992. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[40] Zuxuan Wu, Caiming Xiong, Yu-Gang Jiang, and Larry S Davis. Liteeval: A coarse-to-fine framework for resource" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 397 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 116 + ], + "type": "text", + "content": "efficient video recognition. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[41] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), pages 466-481, 2018. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "text", + "content": "[42] Lumin Xu, Yingda Guan, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Vipnas: Efficient video pose estimation via neural architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16072-16081, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 229, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 251 + ], + "type": "text", + "content": "[43] XuehaiPan. nvitop. https://github.com/XuehaiPan/nvitop. Accessed: 2022-4-25. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[44] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[45] Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization: Towards lossless cnns with low-precision weights. In International Conference on Learning Representations, ICLR2017, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 353, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 397 + ], + "type": "text", + "content": "[46] Mohammadreza Zolfaghari, Kamaljeet Singh, and Thomas Brox. Eco: Efficient convolutional network for online video understanding. In Proceedings of the European conference on computer vision (ECCV), pages 695-712, 2018. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10523" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_content_list.json b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..edcba1bff94fc97579f4ae9e71952dc4a2398a23 --- /dev/null +++ b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_content_list.json @@ -0,0 +1,1621 @@ +[ + { + "type": "text", + "text": "T-SEA: Transfer-based Self-Ensemble Attack on Object Detection", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Huang* \nPeking University \nBeijing, China \nhuanghao@stu.pku.edu.cn", + "bbox": [ + 132, + 180, + 339, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyan Chen* \nPeking University \nBeijing, China \nchen.ziyan@outlook.com", + "bbox": [ + 383, + 181, + 581, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huanran Chen* Peking University Beijing, China huanran_chen@outlook.com", + "bbox": [ + 622, + 181, + 834, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongtao Wang† \nPeking University \nBeijing, China \nwyt@pku.edu.cn", + "bbox": [ + 289, + 258, + 433, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kevin Zhang \nPeking University \nBeijing, China \nkevinzyz@pku.edu.cn", + "bbox": [ + 503, + 258, + 673, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 364, + 312, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to query-based black-box attacks, transfer-based black-box attacks do not require any information of the attacked models, which ensures their secrecy. However, most existing transfer-based approaches rely on ensembling multiple models to boost the attack transferability, which is time- and resource-intensive, not to mention the difficulty of obtaining diverse models on the same task. To address this limitation, in this work, we focus on the single-model transfer-based black-box attack on object detection, utilizing only one model to achieve a high-transferability adversarial attack on multiple black-box detectors. Specifically, we first make observations on the patch optimization process of the existing method and propose an enhanced attack framework by slightly adjusting its training strategies. Then, we analyze patch optimization with regular model optimization, proposing a series of self-ensemble approaches on the input data, the attacked model, and the adversarial patch to efficiently make use of the limited information and prevent the patch from overfitting. The experimental results show that the proposed framework can be applied with multiple classical base attack methods (e.g., PGD and MIM) to greatly improve the black-box transferability of the well-optimized patch on multiple mainstream detectors, meanwhile boosting white-box performance. Our code is available at https://github.com/VDIGPKU/T-SEA.", + "bbox": [ + 76, + 397, + 472, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 804, + 207, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of computer vision, deep learning-based object detectors are being widely applied to", + "bbox": [ + 76, + 830, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1184f9effd3d46f673255c5a7de031f2917ff87c1b5020893f1c8be1da94e039.jpg", + "image_caption": [ + "Figure 1. Model optimization usually augments the training data and drop out neurons to increase generalization, motivating us to propose self-ensemble methods for adversarial patch optimization. Specifically, inspired by data augmentation in model optimization, we augment the data $x$ and the model $f$ via constrained data augmentation and model ShakeDrop, respectively. Meanwhile, inspired by drop out in model optimization, we cut out the training patch $\\tau$ to prevent it overfitting on specific models or images." + ], + "image_footnote": [], + "bbox": [ + 503, + 368, + 875, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "many aspects of our lives, many of which are highly related to our personal safety, including autonomous driving and intelligent security. Unfortunately, recent works [16, 19, 34, 37] have proved that adversarial examples can successfully disrupt the detectors in both digital and physical domains, posing a great threat to detector-based applications. Hence, the mechanism of adversarial examples on detectors should be further explored to help us improve the robustness of detector-based AI applications.", + "bbox": [ + 496, + 714, + 890, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real scenes, attackers usually cannot obtain the details of the attacked model, so black-box attacks naturally receive more attention from both academia and industry.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 875, + 349, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20514", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generally speaking, black-box adversarial attacks can be classified into 1) query-based and 2) transfer-based. For the former, we usually pre-train an adversarial perturbation on the white-box model and then fine-tune it via the information from the target black-box model, assuming we can access the target model for free. However, frequent queries may expose the attack intent, weakening the covertness of the attack. Contrarily, transfer-based black-box attacks utilize the adversarial examples' model-level transferability to attack the target model without querying and ensure the secrecy of the attack. Thus, how to enhance the model-level transferability is a key problem of transfer-based black-box attacks. Most existing works apply model ensemble strategies to enhance the transferability among black-box models, however, finding proper models for the same task is not easy and training adversarial patch on multiple models is laborious and costly. To address these issues, in this work, we focus on how to enhance the model-level transferability with only one accessible model instead of model assembling.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though the investigation of adversarial transferability is still in its early stage, the generalizability of neural networks has been investigated for a long time. Intuitively, the association between model optimization and patch optimization can be established by shifting of formal definitions. Given the input pair of data $x \\in \\mathcal{X}$ , and label $y \\in \\mathcal{Y}$ , the classical model learning is to find a parametric model $f_{\\theta}$ such that $f_{\\theta}(x) = y$ , while learning an adversarial patch treats the model $f_{\\theta}$ , the original optimization target, as a fixed input to find a hypothesis $h$ of parametric patch $\\tau$ to corrupt the trained model such that $h_{\\tau}(f_{\\theta}, x) \\neq y$ . Hence, it is straight forward to analogize patch optimization with regular model optimization. Motivated by the classical approaches for increasing model generalization, we propose our Transfer-based Self-Ensemble Attack (T-SEA), ensembling the input $x$ , the attacked model $f_{\\theta}$ , and the adversarial patch $\\tau$ from themselves to boost the adversarial transferability of the attack.", + "bbox": [ + 75, + 382, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we first introduce an enhanced attack baseline based on [32]. Observing from Fig. 4 that the original training strategies have some limitations, we slightly adjust its learning rate scheduler and training patch scale to revise [32] as our enhanced baseline (E-baseline). Then, as shown in Fig. 1, motivated by input augmentation in model optimization (e.g., training data augmentation), we introduce the constrained data augmentation (data self-ensemble) and model ShakeDrop (model self-ensemble), virtually expanding the inputs of patch optimization (i.e., the input data $x$ and the attacked model $f$ ) to increase the transferability of the patch against different data and models. Meanwhile, motivated by the dropout technique in model optimization, which utilizes sub-networks of the optimizing model to overcome overfitting and thus increasing model generalization, we propose patch cutout (patch self-ensemble), ran", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "domly performing cutout on the training patch $\\tau$ to overcome overfitting. Through comprehensive experiments, we prove that the proposed E-baseline and self-ensemble strategies perform very well on widely-used detectors with mainstream base attack methods (e.g., PGD [24], MIM [13]). Our contributions can be summarized as the following:", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a transfer-based black-box attack T-SEA, requiring only one attacked model to achieve a high adversarial transferability attack on object detectors.", + "- Observing the issues of the existing approach, we slightly adjust the training strategies to craft an enhanced baseline and increase its performance.", + "- Motivated by approaches increasing generalization of deep learning model, we propose a series of strategies to self-ensemble the input data, attacked model, and adversarial patch, which significantly increases the model-level adversarial transferability without introducing extra information.", + "- The experimental results demonstrate that the proposed T-SEA can greatly reduce the AP on multiple widely-used detectors on the black-box setting compared to the previous methods, while concurrently performing well with multiple base attack methods." + ], + "bbox": [ + 517, + 191, + 890, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 488, + 640, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Black-box Adversarial Attack", + "text_level": 1, + "bbox": [ + 500, + 513, + 764, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the discovery of adversarial attack by [31], black-box attacks have gradually attracted more attention, owing to the difficulty of obtaining details of the attacked model in real scenes. Black-box attack can be separated into query-based methods [5, 12, 21] and transfer-based methods [2, 20, 26]. The former utilizes the outputs of the target model to optimize the adversarial examples, which may disclose the attack behavior due to the frequent queries. Hence, in our work, we concentrate on the latter, optimizing adversarial examples in a substitute model without relying on any knowledge of the black-box models. Though the transfer-based attacks ensure the covertness of the attack, their performance is generally limited due to the lack of specialized adjustment catered to the target model. In this work, we propose the self-ensemble strategies to address the issue and greatly boost the black-box attack transferability of the existing transfer-based detector attack.", + "bbox": [ + 496, + 536, + 890, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Object Detectors", + "text_level": 1, + "bbox": [ + 500, + 801, + 666, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object detection is a fundamental computer vision technology predicting objects' categories and positions and is widely used in many perception tasks. In the past ten years, deep learning-based models have greatly improved the performance of object detectors. The mainstream methods", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20515", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2230a575dc7e538de8ed49b5c87171131ee30529dd74d51798e01b9b069e7aee.jpg", + "image_caption": [ + "Figure 2. The Overall Pipeline of T-SEA. During the training stage, we utilize the self-ensemble strategies (i.e., the constrained data augmentation, patch cutout and model ShakeDrop) to enhance the transferability of the well-optimized adversarial patch. During the attack (inference) stage, we attach the crafted patch into different images to disrupt the detection process of multiple widely-used detectors on black-box setting." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 880, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based on deep learning can be roughly divided into one-staged and two-staged methods. The former directly predicts the positions and classes of the object instances and are thus faster; the latter first use the region proposal network (RPN) to generate proposals, and then predict the labels and positions of the selected proposals. In this paper, we select eight mainstream detectors of both one-stage and two-stage, including YOLO v2 [27], YOLO v3 & YOLO v3tiny [28], YOLO v4 & YOLO v4tiny [4], YOLO v5 [17], Faster R-CNN [29] and SSD [22] to systemically verify the proposed T-SEA framework.", + "bbox": [ + 75, + 363, + 472, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Attacks on Object Detector", + "text_level": 1, + "bbox": [ + 76, + 537, + 323, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The security of deep learning-based models is receiving increasing attention [8, 9], especially with the existence of adversarial attacks. Recent works have explored adversarial robustness of object detectors. To begin, [37] apply adversarial attack on object detectors, performing iterative gradient-based method to misclassify the proposals, and the similar ideas are also carried out by [19, 34]. Ensuingly, for enhancing the attack capability in the physical world, [7, 32] propose real-world adversarial patches to attack the mainstream detectors, e.g., YOLO and Faster R-CNN. Recently, model ensemble approaches are used to improve the attack transferability among multiple detectors, such as [16, 35, 41] simultaneously attacking multiple detectors to generate cross-model adversarial examples. Different from the above, we focus on how to make the most of the limited information (i.e., a single detector) to carry out a high-transferability black-box attack on multiple black-box detectors.", + "bbox": [ + 75, + 560, + 472, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 844, + 168, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first give the problem formulation in Sec. 3.1 and describe the overall framework in Sec. 3.2.", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we respectively introduce the details of our enhanced baseline and self-ensemble strategies in Sec. 3.3 and Sec. 3.4.", + "bbox": [ + 496, + 363, + 892, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 417, + 702, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we carry out a transfer-based black-box attack with only one white-box detector to decrease the average precision (AP) of both white-box and black-box detectors. Given the target input data distribution $\\mathcal{D}(\\mathcal{X},\\mathcal{F})$ , we regard a single pre-trained detector $f_{w}\\in \\mathcal{F}$ as the whitebox attacked model, $x_{1,\\dots,N}\\in \\mathcal{X}$ as the input images, where $N$ is the number of training samples. We are committed to crafting a universal adversarial patch $\\tau$ from the adversarial distribution $\\mathcal{T}$ to disrupt the detection process,", + "bbox": [ + 496, + 440, + 893, + 578 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\tau} = \\underset {\\tau \\sim \\mathcal {T}} {\\arg \\min } \\sum_ {i} ^ {N} L \\left(\\widetilde {x} _ {i}, \\widetilde {f} _ {w}, \\widetilde {\\tau}\\right) + \\lambda J (\\tau), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 587, + 892, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L$ is the loss function to measure the corruption of detector, $\\widetilde{x}_i,\\widetilde{f}_w$ and $\\widetilde{\\tau}$ denote self-enssembled data, model and patch, respectively, and $J(\\cdot)$ is a regularization term which we employ total variation of $\\tau$ in our work.", + "bbox": [ + 496, + 638, + 893, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall Framework", + "text_level": 1, + "bbox": [ + 500, + 710, + 689, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, we divide the entire T-SEA pipeline into training stage and attack stage. During the training stage, the input images are first augmented with the constrained data augmentation. Ensuingly, the white-box detector (i.e., the attacked detector) is used to locate the target objects from the augmented images. Then, we attach adversarial patches to the center of each detected object of the target class, along with patch cutout for mitigating overfitting. After that, the images with adversarial patches go through the shake-dropped models, and we minimize the object confidence and continuously optimize the training", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20516", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adversarial patches until reaching maximum epoch number. During the attack stage, we apply our well-optimized adversarial patch on the test images to disrupt the detection process of multiple black-box detectors.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Enhanced Baseline", + "text_level": 1, + "bbox": [ + 76, + 162, + 259, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the details of the enhanced baseline, which is described in Algorithm 1. Firstly, we randomly initialize a patch $\\tau_0$ , prepare the training images and attacked model $f_w$ . During each training epoch, for every input image batch $X$ , we first obtain their detection results via the white-box detector $f_w$ . Ensuingly, we utilize the transformation function $T$ to apply the training patch into each image to generate adversarial image batch $X^{adv}$ . Next, we calculate the object confidence of $X^{adv}$ , and the attack loss of the image batch. Here we use $\\ell_1$ loss as the attack loss, unless otherwise stated. Finally, we update the $\\tau$ via base attack method $\\hbar$ and adjust the learning rate after each epoch until reaching the maximum number.", + "bbox": [ + 75, + 186, + 470, + 383 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Enhanced Baseline Based on [32]" + ], + "code_body": "Require: $x_{1,\\dots,N}$ (training images), $f_{w}$ (white-box detector), $M$ (maximum epoch), $BS$ (batch size), $\\tau_0$ (the initial patch), $T$ (patch applicer), $\\hbar$ (base attack method), scheduler (learning rate scheduler). \nEnsure: Well-optimized Adversarial Patch $\\tau$ \n1: $\\tau \\Leftarrow \\tau_0$ \n2: for each $i \\in [1, M]$ do \n3: for each $j \\in [1, \\frac{N}{BS}]$ do \n4: $X \\Leftarrow x_{(j-1)\\cdot BS+1}, \\ldots, x_{j\\cdot BS}$ \n5: bboxclean, confclean $\\Leftarrow f_w(X)$ \n6: $X^{adv} \\Leftarrow T(X, bbox^{clean}, \\tau)$ \n7: bbox $^{adv}$ , conf $^{adv} \\Leftarrow f_w(X^{adv})$ \n8: loss $\\Leftarrow Avg(conf^{adv})$ \n9: $\\tau \\Leftarrow \\hbar(\\tau, loss)$ \n10: end for \n11: update lr via scheduler \n12: end for", + "bbox": [ + 78, + 416, + 468, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to AdvPatch [32], we slightly adjust two training strategies. One is the learning rate scheduler. We observe that a saddle point during patch optimization may lead the original plateau-based scheduler to drastically drop the learning rate, causing inadequate optimization. Hence, we adjust the learning rate decline strategy with the following criterion,", + "bbox": [ + 75, + 694, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl r \\Leftarrow l r * \\mu , i f \\left(\\ell_ {t} - \\ell_ {t - 1}\\right) < \\epsilon_ {1} a n d \\frac {\\left(\\ell_ {t} - \\ell_ {t - 1}\\right)}{\\ell_ {t}} < \\epsilon_ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 811, + 480, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ is the decay factor, $\\ell_t$ denotes the mean loss at epoch $t$ , $\\epsilon_1, \\epsilon_2$ are thresholds to control learning rate updates. In our experiments, we ensure that the learning rate", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "decrease more stably via the above hyper-parameters. The other adjustment is to reduce the patch scale at training stage $s_t$ . Here, the patch scale is the length ratio between the patch and the bounding box of the specific object. As Fig. 4 shows, a proper scale helps the optimized patch to learn more global patterns, which will not be disrupted easily when the patch is scaled, e.g., a local pattern will lose more information when its corresponding patch is scaled to a very small size.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Self-Ensemble Strategies", + "text_level": 1, + "bbox": [ + 498, + 237, + 725, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1 Theoretical Analysis", + "text_level": 1, + "bbox": [ + 500, + 260, + 696, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typical model training usually attempts to find a mapping function $f$ from a finite hypothesis space $\\mathcal{F}$ that describes the relationship between the input data $x$ and label $y$ following the underlying joint distribution $\\mathcal{D}(\\mathcal{X},\\mathcal{Y})$ . Generally, since $\\mathcal{D}$ is unknown, we use training data set $S = \\{(x_i,y_i)|i = 1,\\dots,N\\}$ to train the model, which means the optimization involves the empirical risk minimization(ERM) [33]:", + "bbox": [ + 496, + 282, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {R} _ {S} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} L \\left(y _ {i}, f \\left(x _ {i}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 416, + 890, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L$ is a loss function to measure the difference between the model prediction $f(x)$ and the label $y$ . However, the empirical risk is unable to provide generalization on unseen data [40]. Fortunately, we can derive that the generalization error $R_{\\mathcal{D}}(f)$ is bounded by $\\widehat{R}_S(f)$ for a given confidence $1 - \\sigma \\in (0,1)$ [3,25],", + "bbox": [ + 496, + 467, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\forall f \\in \\mathcal {F}, R _ {\\mathcal {D}} (f) \\leq \\widehat {R} _ {S} (f) + \\sqrt {\\frac {\\log c + \\log \\frac {1}{\\sigma}}{N}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 571, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ denotes the input data size, $c$ is a complexity measure of $\\mathcal{F}$ , such as VC-dimension [25] or covering numbers [1]. This generally yields bounds of the generalization gap $R_{\\mathcal{D}}(f) - \\widehat{R}_S(f) = \\mathcal{O}(\\sqrt{\\frac{\\log c}{N}})$ , which gives us a lens to analyze model generalization: we can sample more training data or lower model complexity to help close the generalization gap. In practice, limited training data and large-scaled model parameters make the model prone to overfit [39], causing an undesirable generalization gap increase. Fortunately, past studies have widely adopted a series of regularization strategies on both input data and the model, such as data augmentation [11, 40, 42], Stochastic Depth [15], and Dropout [30].", + "bbox": [ + 496, + 621, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the aforementioned model regularization methods, we discuss how to improve attack transferability of the adversarial patch. During patch optimization, the model $f$ becomes an input instead of the optimization objective, which means the original input distribution", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20517", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{D}(\\mathcal{X}, \\mathcal{Y})$ is shifted to a joint input distribution $\\mathcal{D}(\\mathcal{X}, \\mathcal{F})$ . Naturally, we can improve generalization of the adversarial patch by increasing training data scale $N_{x,f}$ , i.e. sampling more training images $x$ and assembling more white-box models $f_w$ . However, obtaining massive images or models of the same task is generally expensive or impractical for the attacker. To address this issue, we propose to virtually expand the input by data augmentation and model ShakeDrop to help improve patch transferability. Meanwhile, although the large-scale models have achieved superior performance in most tasks, these models still benefit from reducing model complexity at training stage like stochastic depth [15] and Dropout [30]. Thus we propose the patch cutout to reduce its capacity $c_{\\tau}$ in training process as a new regularization to alleviate overfitting.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2 Data: Constrained Data Augmentation", + "text_level": 1, + "bbox": [ + 76, + 335, + 403, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivated by data augmentation in increasing model generalization, we can similarly improve patch transferability by virtually expanding the training set $S$ to more closely approximate the underlying distribution $\\mathcal{D}$ . Therefore, we employ a constrained policy to avoid unnatural augmentation which may not appear in natural scenes. In our work, we 1) mildly resize and crop the input images, 2) slightly alter the brightness, contrast, saturation and hue, and 3) randomly rotate the input images in a small range, to generate natural augmented images.", + "bbox": [ + 75, + 359, + 472, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3 Model: ShakeDrop", + "text_level": 1, + "bbox": [ + 76, + 527, + 264, + 544 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/734ba65804b6e881505ff51289a1a3a33114e5adf16772fe42f0acbe88435e97.jpg", + "image_caption": [ + "(a) Forward", + "Figure 3. We utilize the model ShakeDrop to linearly combine the outputs of the stacked layers and identity of a residual block." + ], + "image_footnote": [], + "bbox": [ + 99, + 565, + 261, + 723 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/82d0b5cc5ee8845bec1f99a1cd693c00ab4ab817f51dbb0b8f5181f19a69e5dc.jpg", + "image_caption": [ + "(b) Backward" + ], + "image_footnote": [], + "bbox": [ + 287, + 568, + 446, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed above, when training the patch, sampling more white-box models $f$ from the joint input distribution $\\mathcal{D}$ can improve the adversarial transferability of the crafted patch. However, it is difficult to obtain multiple models of the same task in reality. Inspired by stochastic depth [15], which randomly drops a subset of layers to virtually ensemble multiple model variants for improving generalization,", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we utilize ShakeDrop [38] to linearly combine the outputs of stacked layers and identity of a residual block, generating massive variants of the attacked white-box model as the Fig. 3 illustrated.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in forward propagation, we combine the identity $z_{l}$ and the output of stacked layers $F(z_{l})$ of a residual block to generate $z_{L}$ with the following formula:", + "bbox": [ + 496, + 151, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {L} = z _ {l} + (\\beta + \\alpha - \\beta \\cdot \\alpha) \\cdot F (z _ {l}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 210, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is sampled from a Bernoulli distribution with $P(\\beta = 1) = \\varphi_s$ , $\\alpha$ is sampled from a continuous uniform distribution $\\alpha \\sim U(1 - e, 1 + e)$ ( $e$ is a constant). In backward propagation, denoting the loss function as $\\mathcal{E}$ , from the chain rule of backpropagation, ShakeDrop can be formulated as", + "bbox": [ + 496, + 239, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\psi \\left(\\frac {\\partial \\mathcal {E}}{\\partial z _ {l}}\\right) = \\psi \\left(\\frac {\\partial \\mathcal {E}}{\\partial z _ {L}} \\cdot \\frac {\\partial z _ {L}}{\\partial z _ {l}}\\right) \\tag {6} \\\\ = \\frac {\\partial \\mathcal {E}}{\\partial z _ {L}} (1 + (\\gamma + \\alpha - \\gamma \\cdot \\alpha) \\cdot \\frac {\\partial F}{\\partial z _ {l}}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 337, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is another Bernoulli variable and identically distributed with $\\beta$ .", + "bbox": [ + 496, + 412, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.4 Patch: Cutout", + "text_level": 1, + "bbox": [ + 498, + 460, + 653, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dropout [30] is designed to randomly drop neural units along with their connections, effectively preventing overfitting of the model. Inspired by Dropout, we propose a patch cutout strategy, randomly masking a region of the adversarial patch to reduce the patch complexity at training stage and thus to prevent overfitting on specific model and data. From an other angle, model Dropout prevents excessive co-adapting of neural units and spread out features over multi neurons to alleviate overfitting. Similarly, patch cutout also prevents the adversarial nature of patch from relying too much on the patterns of a certain area.", + "bbox": [ + 496, + 483, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed patch cutout is similar to cutout [11] or random erasing [42] in model training. Specifically, for a normalized input image $I_{x,y} \\in [0,1]$ of size $H \\times W$ , we carry out the following process with a probability $\\varphi_c$ before attaching patches to the target objects: 1) we firstly conduct random sampling to obtain one point $p = (x_0, y_0)$ within the given patch; 2) then we cover a $\\eta H \\times \\eta W$ square area $(\\eta$ is the ratio) centered at $p$ with a specific value $\\eta \\in [0,1]$ .", + "bbox": [ + 496, + 650, + 890, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 784, + 624, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the implementation details in Sec. 4.1. Then, we present the main results of T-SEA in Sec. 4.2, reporting its performance on different detectors and attack methods. After that, we compare the proposed method with other state-of-the-art detection attack algorithms in Sec. 4.3 and perform the ablation study", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20518", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d7c54e807d8d83e013c538091f2669d7a5b471f43e7b52e16b93ebc874991239.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsYOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSDBlack-Box Avg↓
AdvPatch5.6640.2618.0748.4924.4443.3839.2741.2836.46
T-SEA(ours)1.734.482.415.687.756.9116.3820.55\\(9.16^{27.30}\\)↓
AdvPatch51.8513.8951.1757.1658.4370.4751.4659.7957.19
T-SEA(ours)31.025.7635.3842.3725.358.0237.6252.26\\(40.28^{16.91}\\)↓
AdvPatch56.0266.121.3969.6451.5672.1656.6160.7361.83
T-SEA(ours)38.8547.130.5152.7531.0161.1849.1255.44\\(47.93^{13.90}\\)↓
AdvPatch37.4137.1817.5819.6726.9146.3744.6743.0736.17
T-SEA(ours)9.346.469.494.2216.6511.6616.2429.43\\(14.18^{21.99}\\)↓
AdvPatch47.4159.5937.1466.4814.5069.5155.9555.2255.90
T-SEA(ours)37.6944.9815.8151.144.1151.3746.0649.39\\(42.35^{13.55}\\)↓
AdvPatch46.954.9329.2062.1646.1513.3947.7150.7348.25
T-SEA(ours)11.6116.7710.7332.5317.551.3717.1430.78\\(19.59^{28.66}\\)↓
AdvPatch24.5323.3714.5826.5425.630.468.6242.1526.75
T-SEA(ours)9.284.083.998.5513.5810.453.0826.46\\(10.91^{15.84}\\)↓
AdvPatch32.4362.9352.0366.2249.0751.1947.4215.1051.61
T-SEA(ours)12.0630.909.7327.339.0817.3828.545.13\\(19.29^{32.32}\\)↓
", + "bbox": [ + 91, + 88, + 877, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a7fd64c97a9da6e10fcb7f1832bfaa3dfc2d532caa01b99a796f5857b4de7f7b.jpg", + "table_caption": [ + "Table 1. Comparisons between T-SEA and AdvPatch [32]. We attack each detector separately and use the remaining seven detectors as the black-box detectors. The proposed T-SEA performs much better than [32] on both white-box setting and black-box setting, demonstrating the effectiveness of the proposed self-ensemble strategies." + ], + "table_footnote": [], + "table_body": "
MethodWhite Box ↓Black-Box Avg↓
AdamAdvPatch13.3948.25
T-SEA(ours)1.3719.59
SGDAdvPatch20.3955.97
T-SEA(ours)1.6615.85
MIMAdvPatch11.9133.51
T-SEA(ours)1.4315.43
BIMAdvPatch8.4732.93
T-SEA(ours)1.6219.15
PGDAdvPatch13.5838.70
T-SEA(ours)1.6018.06
", + "bbox": [ + 114, + 449, + 431, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons of T-SEA and AdvPatch with Different Base Attack Methods on YOLO v5 (white-box). We select five classical base attack methods, including optimization-based methods (Adam and SGD) and iterative methods (MIM, BIM and PGD). The results show that T-SEA can enhance the performance of all these methods and performs much better than AdvPatch.", + "bbox": [ + 75, + 657, + 470, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Sec. 4.4. Finally, we show the transferability of T-SEA across the datasets and scenes in Sec. 4.5 and Sec. 4.6.", + "bbox": [ + 75, + 756, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 800, + 294, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets In our experiments, we utilize the INRIA person dataset [10] to train and test our adversarial patch, whose training set and test set consist of 614 and 288 images, respectively. Meanwhile, to verify the transferability of the crafted patch, we select images containing per", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "son from COCO validation set (named COCO-person) and CCTV Footage of Humans* (named CCTV-person) as additional test data. The former contains 1684 human images in different scenes (e.g., sports playground, transportation routes, oceans and forests) and the latter contains 559 in-person images from camera footage.", + "bbox": [ + 496, + 452, + 890, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimization Details We use INRIA train set as the training set and regard person as the target attack class. The patch size is $300 \\times 300$ , the input image size is $416 \\times 416$ , the batch size $BS = 8$ , and the maximum epoch number $M = 1000$ . For the E-baseline, we adjust the training patch scale from 0.2 in AdvPatch [32] to 0.15, set $\\epsilon_{1} = 1e - 4$ , $\\epsilon_{2} = 1e - 4$ for the learning rate scheduler, and adopt the Adam [18] as the optimizer. For constrained data augmentation, we carry out constrained data augmentation via horizontal flip, slight color jitter, random resized crop, and random rotation; for model ShakeDrop, the constant $e$ is set to 1 so that $\\alpha \\sim U(0,2)$ , and we perturb the model with the probability of $\\varphi_{s} = 0.5$ ; for patch cutout, we set the fill value of the erased area $k = 0.5$ , the ratio $\\eta = 0.4$ , and the probability $\\varphi_{c} = 0.9$ .", + "bbox": [ + 496, + 546, + 892, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric Following [14, 32], we use the Average Precision (AP) to measure the attack capability of the crafted patch (the lower AP, the better attack) and regard the detector's predictions of the clean data as the ground truth (i.e., $AP = 1$ ).", + "bbox": [ + 496, + 776, + 890, + 852 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "\\*https://www.kaggle.com/datasets/constantinwerner/human-detectiondataset", + "bbox": [ + 500, + 875, + 888, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20519", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/609d701fa2a583fb4e18b238c9528d2a85c30af42b56b27a14e72f80dedf06fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodWhite Box ↓Black Box ↓Black-Box Avg↓
YOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSD
Gray67.7576.2280.6975.2276.8981.8661.7572.05-
Random Noise70.6775.882.4475.178.7481.7963.4172.9-
White68.5274.8980.274.7376.0980.0960.3569.41-
NPAP [14]38.0356.8558.0467.7467.4366.8556.656.6661.45
AdvCloak [36]33.7454.7753.4267.5756.1268.0555.1960.8259.42
AdvPatch [32]5.6640.2618.0748.4924.4443.3839.2741.2836.46
E-baseline(ours)3.6115.325.5018.5813.399.7729.7323.8216.59
T-SEA(ours)1.734.482.415.687.756.9116.3820.559.16
", + "bbox": [ + 96, + 88, + 874, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 303, + 215, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Results on Different Attacked Models", + "text_level": 1, + "bbox": [ + 76, + 327, + 392, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We systematically investigate attack performance of the proposed T-SEA on eight widely-used object detectors, and report the quantitative results in Tab. 1. Since T-SEA is improved based on AdvPatch [32], we also report the results of AdvPatch. The proposed T-SEA achieves significant improvements on both white-box and black-box performance of all eight detectors compared to [32], demonstrating the effectiveness of the proposed strategies. Meanwhile, for some white-box detectors (e.g., YOLO v2), the black-box average AP can drop to around 10, exhibiting effective black-box performance of T-SEA.", + "bbox": [ + 75, + 352, + 468, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Results of Different Attack Methods", + "text_level": 1, + "bbox": [ + 76, + 534, + 382, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "T-SEA is not designed only for a specific base attack method; it is important for T-SEA to perform well on different base attack methods. In Tab. 2, we give the detailed results of T-SEA with different base attack methods on YOLO v5, which includes the optimization-based methods (e.g., Adam and SGD) and iterative methods (e.g., MIM, BIM and PGD). The results show that the proposed T-SEA can improve both white-box and black-box attack performance on all methods above, demonstrating that the performance gain caused by T-SEA is not limited to a specific method. That is to say, T-SEA may potentially work well with future base attack methods to further increase the transferability of their crafted AdvPatch.", + "bbox": [ + 75, + 554, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 76, + 758, + 369, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare T-SEA with the SOTA detection attack approaches, which all regard the YOLO v2 as the white-box model and evaluate on seven black-box detectors, and we also report the results of gray/random noise/white patches as control group. As reported on Tab. 3, 1) all adversarial patch performs much better than the control group; 2) our E-baseline and T-SEA achieve the highest performance among all adversarial approaches with the", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cd5bbb7c05270c902bbd9595a66340692ad2f5f0f7445b035cde6801a394e23.jpg", + "table_caption": [ + "Table 3. Comparisons with Existing Detection Attack Methods. For clearer controlled observations, we list the results of gray, random noise, and white patch. Compared with existing methods, T-SEA achieves the best performance on both white-box and black-box attack." + ], + "table_footnote": [], + "table_body": "
White Box ↓Black Box ↓
E-baseline13.3948.25
+ Constrained Data Aug.18.4742.42
+ Model ShakeDrop12.5439.40
+ Patch cutout6.8032.54
+ Combined Strategies1.3719.59
", + "bbox": [ + 514, + 301, + 874, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation Study of Self-Ensemble Strategies on YOLO v5 (white-box). Each self-ensemble strategy is added to the E-baseline to verify its individual performance gain. The results show that: 1) though using data augmentation alone degrades the white-box performance, all strategies can obviously improve the black-box performance of E-baseline; 2) combining all achieves non-trivial results on both white-box and black-box setting.", + "bbox": [ + 496, + 415, + 892, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "same inference setting (i.e., we ensure the perturbed area is same for each method), showing that compared to the existing single model attack method, the training strategies adjustment and self-ensemble strategies can effectively enhance the attack capability on both white-box and black-box of the crafted adversarial patch.", + "bbox": [ + 496, + 523, + 890, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 625, + 653, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Training Strategies Adjustment", + "text_level": 1, + "bbox": [ + 500, + 650, + 772, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we explore the performance gain from the training strategies adjustments. As discussed in the Sec. 3.3, we modify the learning rate scheduler and training patch scale of [32]. As illustrated in Fig. 4, the modified training scale can reduce the AP in test set more quickly and effectively, and the crafted patch has clearer adversarial patterns. The adjusted scheduler can also lead the detection loss to decrease more, and can be further combined with the modified training scale to achieve a better result.", + "bbox": [ + 496, + 674, + 890, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Self-Ensemble Strategies", + "text_level": 1, + "bbox": [ + 500, + 833, + 725, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we are motivated by model training approaches that enhance generalization ability, and we propose self-ensemble strategies on the input data, the attacked", + "bbox": [ + 496, + 854, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20520", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96cd91f62347db130d1f0a83095630b1e06fd01cb3fb76029f5f0b720d6eb67c.jpg", + "image_caption": [ + "Figure 4. The improvements after adjusting learning rate scheduler (tagged as $\\mathrm{PlateauLR}^{\\dagger}$ ) and patch scale at training stage. A small patch scale during training will cause the test AP to drop lower and faster, and the adjusted $\\mathrm{PlateauLR}^{\\dagger}$ can also cause the detection loss to further decrease." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 297, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afbb33ffe759763f1f868375778744d1920f4fc4cdac8c02e30fad05670f12b8.jpg", + "image_caption": [ + "PlateauLR Scale=0.2 mAP:38.89" + ], + "image_footnote": [], + "bbox": [ + 308, + 90, + 370, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f769d13b97809adac618432521c48bdedd51f992f13bcfdd9df84fd582b7c19.jpg", + "image_caption": [ + "PlateauLR Scale=0.15 mAP:36.63" + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 444, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1dfc0af9db60426bbe2b98ff55d21bc49cf30025764186e336bbf7ffc9cba3c.jpg", + "image_caption": [ + "PlateauLR+ Scale=0.2 mAP:35.38" + ], + "image_footnote": [], + "bbox": [ + 308, + 165, + 370, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7faa8157c411a78bf38658b62f1d78fa73a19b54f34ab6161b126a9e9a3a6391.jpg", + "image_caption": [ + "PlateauLR+ Scale=0.15 mAP:24.77" + ], + "image_footnote": [], + "bbox": [ + 387, + 166, + 447, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9e4b399191c85f46e625fd388e8cc9dee99004fba161e8a0336150d535622f9c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodWhite Box ↓Black Box ↓
COCO-personAdvPatch [32]45.8352.54
T-SEA37.2838.87
CCTV-personAdvPatch [32]38.0734.08
T-SEA38.7119.91
", + "bbox": [ + 78, + 339, + 478, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The Cross Datasets Transferability of T-SEA on YOLO v5 (white-box). Compared to the [32], the adversarial patch crafted by T-SEA has much stronger attack adaptability on person images from different datasets.", + "bbox": [ + 75, + 446, + 468, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model, and the training patch, greatly augmenting adversarial transferability from themselves. Here we individually inspect the proposed strategies on our E-baseline (YOLO v5 as white-box detector) to explore the impact of each strategy. As reported in Tab. 4, though the constrained data augmentation will slightly decrease the white-box performance (other two strategies both improve the white-box performance), all these self-ensemble strategies increase the black-box results separately. and combining them can achieve the best performance.", + "bbox": [ + 75, + 529, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Cross-Dataset Verification", + "text_level": 1, + "bbox": [ + 76, + 708, + 313, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The capability of performing cross-dataset attack is also significant for a well-optimized patch. Here, we apply the INRIA-trained patch on two different datasets, COCO-person and CCTV-person. Compared to INRIA, COCO-person has much more person images on different scenes, while CCTV-person focus on persons from the security camera. As shown in Tab. 5, the patch crafted by T-SEA achieves comparable results on the white-box setting with the AdvPatch, but obtains a much better black-box attack capability on both COCO-person and CCTV-person, indicating its strong black-box cross-data attacking capability.", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c59c3f8efada64ba0dfe99e073bd1deb52862c7a70f3578d1bebf15c0a5efa0.jpg", + "image_caption": [ + "Figure 5. Physical Attack Demo of T-SEA. After showing the optimized patch on iPad, the YOLO v5 and SSD can not detect person." + ], + "image_footnote": [], + "bbox": [ + 550, + 87, + 839, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Physical Verification", + "text_level": 1, + "bbox": [ + 500, + 291, + 692, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although applying the patch attack in the physical world is not the main goal of our work, we believe the optimized patch that has high transferability may perform well in physical settings. Here we show a simple case in Fig. 5 that the patch shown on a iPad can successfully attack the human detection of YOLO v5 and SSD, without disrupting the detection process of other objects.", + "bbox": [ + 496, + 311, + 890, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Future Work", + "text_level": 1, + "bbox": [ + 500, + 430, + 633, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since most existing mainstream detectors are CNN-based, we focus on applying the self-ensemble strategies on these detectors in this work. However, the transformer-based detectors [6, 23] are achieving promising results, and the proposed model ShakeDrop can not directly applied on transformers (the others can), we will follow the motivation of designing the model ShakeDrop, to propose new approach and generate variants of transformer-based detector.", + "bbox": [ + 496, + 448, + 890, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 583, + 617, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel transfer-based self-ensemble black-box attack on object detectors, achieving stable and excellent performance gains with various base attack methods on multiple popular object detectors. Firstly, with only slight training strategy adjustments, we improve the existing method's performance and regard it as our enhanced baseline. Then, based on this baseline, we propose a series of self-ensemble strategies to augment the input data, the attacked model, and the training patch from itself to significantly enhance the adversarial transferability of the optimized patch on black-box detectors. The comprehensive experimental results reveal the potential risk that an attacker with only one model could still succeed in a high transferability black-box attack.", + "bbox": [ + 496, + 604, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 821, + 687, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Key R&D Program of China (Grant No. 2022ZD0160305) and National Key Laboratory Open Fund of China (Grant No. 6142113210204).", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Anthony and Peter Bartlett. Neural network learning: Theoretical foundations. Cambridge university press, 1999. 4", + "[2] Song Bai, Yingwei Li, Yuyin Zhou, Qizhu Li, and Philip HS Torr. Adversarial metric attack and defense for person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2119-2126, 2020. 2", + "[3] Mikhail Belkin, Siyuan Ma, and Soumik Mandal. To understand deep learning we need to understand kernel learning. In International Conference on Machine Learning, pages 541-549. PMLR, 2018. 4", + "[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3", + "[5] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248, 2017. 2", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 8", + "[7] Shang-Tse Chen, Cory Cornelius, Jason Martin, and Duen Horng Polo Chau. Shapeshifter: Robust physical adversarial attack on faster r-cnn object detector. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 52-68. Springer, 2018. 3", + "[8] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European Conference on Computer Vision, pages 529-548. Springer, 2022. 3", + "[9] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 3", + "[10] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 6", + "[11] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4, 5", + "[12] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2020. 2", + "[13] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2", + "[14] Yu-Chih-Tuan Hu, Bo-Han Kung, Daniel Stanley Tan, Jun-Cheng Chen, Kai-Lung Hua, and Wen-Huang Cheng. Naturalistic physical adversarial patch for object detectors. In" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7", + "[15] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646-661. Springer, 2016. 4, 5", + "[16] Hao Huang, Yongtao Wang, Zhaoyu Chen, Zhi Tang, Wenqiang Zhang, and Kai-Kuang Ma. Rpattack: Refined patch attack on general object detectors. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 3", + "[17] Glenn Jocher, Alex Stoken, Jirka Borovec, NanoCode012, ChristopherSTAN, Liu Changyu, Laughing, tkianai, Adam Hogan, lorenzomammana, yxNONG, AlexWang1900, Laurentiu Diaconu, Marc, wanghaoyang0106, ml5ah, Doug, Francisco Ingham, Frederik, Guilhien, Hatovix, Jake Poznanski, Jiacong Fang, Lijun Yu, changyu98, Mingyu Wang, Naman Gupta, Osama Akhtar, PetrDvoracek, and Prashant Rai. ultralytics/yolov5, Oct. 2020. 3", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Yuezun Li, Daniel Tian, Ming-Ching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. arXiv preprint arXiv:1809.05962, 2018. 1, 3", + "[20] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations. 2", + "[21] Siao Liu, Zhaoyu Chen, Wei Li, Jiwei Zhu, Jiafeng Wang, Wenqiang Zhang, and Zhongxue Gan. Efficient universal shuffle attack for visual object tracking. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2739-2743. IEEE, 2022. 2", + "[22] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 3", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8", + "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2", + "[25] Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine learning. MIT press, 2018. 4", + "[26] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pages 506-519, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20522", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 3", + "[28] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 3", + "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3", + "[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4, 5", + "[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2", + "[32] Simen Thys, Wiebe Van Ranst, and Toon Goedemé. Fooling automated surveillance cameras: adversarial patches to attack person detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 3, 4, 6, 7, 8", + "[33] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 4", + "[34] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 1, 3", + "[35] Shudeng Wu, Tao Dai, and Shu-Tao Xia. Dpattack: Diffused patch attacks against universal object detection. arXiv preprint arXiv:2010.11679, 2020. 3", + "[36] Zuxuan Wu, Ser-Nam Lim, Larry S Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 7", + "[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 1, 3", + "[38] Yoshihiro Yamada, Masakazu Iwamura, and Koichi Kise. Shakedrop regularization. 2018. 5", + "[39] Xue Ying. An overview of overfitting and its solutions. In Journal of physics: Conference series, volume 1168, page 022022. IOP Publishing, 2019. 4", + "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4", + "[41] Yusheng Zhao, Huanqian Yan, and Xingxing Wei. Object hider: Adversarial patch attack against object detectors. arXiv preprint arXiv:2010.14974, 2020. 3", + "[42] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 4, 5" + ], + "bbox": [ + 78, + 90, + 468, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20523", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_model.json b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..639182b2546a89f58420c99b5b16c17ae238cc4f --- /dev/null +++ b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.153, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "T-SEA: Transfer-based Self-Ensemble Attack on Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.181, + 0.341, + 0.251 + ], + "angle": 0, + "content": "Hao Huang* \nPeking University \nBeijing, China \nhuanghao@stu.pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.182, + 0.582, + 0.251 + ], + "angle": 0, + "content": "Ziyan Chen* \nPeking University \nBeijing, China \nchen.ziyan@outlook.com" + }, + { + "type": "text", + "bbox": [ + 0.624, + 0.182, + 0.835, + 0.251 + ], + "angle": 0, + "content": "Huanran Chen* Peking University Beijing, China huanran_chen@outlook.com" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.259, + 0.434, + 0.33 + ], + "angle": 0, + "content": "Yongtao Wang† \nPeking University \nBeijing, China \nwyt@pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.26, + 0.674, + 0.33 + ], + "angle": 0, + "content": "Kevin Zhang \nPeking University \nBeijing, China \nkevinzyz@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.365, + 0.313, + 0.381 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.398, + 0.473, + 0.789 + ], + "angle": 0, + "content": "Compared to query-based black-box attacks, transfer-based black-box attacks do not require any information of the attacked models, which ensures their secrecy. However, most existing transfer-based approaches rely on ensembling multiple models to boost the attack transferability, which is time- and resource-intensive, not to mention the difficulty of obtaining diverse models on the same task. To address this limitation, in this work, we focus on the single-model transfer-based black-box attack on object detection, utilizing only one model to achieve a high-transferability adversarial attack on multiple black-box detectors. Specifically, we first make observations on the patch optimization process of the existing method and propose an enhanced attack framework by slightly adjusting its training strategies. Then, we analyze patch optimization with regular model optimization, proposing a series of self-ensemble approaches on the input data, the attacked model, and the adversarial patch to efficiently make use of the limited information and prevent the patch from overfitting. The experimental results show that the proposed framework can be applied with multiple classical base attack methods (e.g., PGD and MIM) to greatly improve the black-box transferability of the well-optimized patch on multiple mainstream detectors, meanwhile boosting white-box performance. Our code is available at https://github.com/VDIGPKU/T-SEA." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.805, + 0.208, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.831, + 0.47, + 0.862 + ], + "angle": 0, + "content": "With the rapid development of computer vision, deep learning-based object detectors are being widely applied to" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.369, + 0.877, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Figure 1. Model optimization usually augments the training data and drop out neurons to increase generalization, motivating us to propose self-ensemble methods for adversarial patch optimization. Specifically, inspired by data augmentation in model optimization, we augment the data \\( x \\) and the model \\( f \\) via constrained data augmentation and model ShakeDrop, respectively. Meanwhile, inspired by drop out in model optimization, we cut out the training patch \\( \\tau \\) to prevent it overfitting on specific models or images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.852 + ], + "angle": 0, + "content": "many aspects of our lives, many of which are highly related to our personal safety, including autonomous driving and intelligent security. Unfortunately, recent works [16, 19, 34, 37] have proved that adversarial examples can successfully disrupt the detectors in both digital and physical domains, posing a great threat to detector-based applications. Hence, the mechanism of adversarial examples on detectors should be further explored to help us improve the robustness of detector-based AI applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In real scenes, attackers usually cannot obtain the details of the attacked model, so black-box attacks naturally receive more attention from both academia and industry." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.351, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.351, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.38 + ], + "angle": 0, + "content": "Generally speaking, black-box adversarial attacks can be classified into 1) query-based and 2) transfer-based. For the former, we usually pre-train an adversarial perturbation on the white-box model and then fine-tune it via the information from the target black-box model, assuming we can access the target model for free. However, frequent queries may expose the attack intent, weakening the covertness of the attack. Contrarily, transfer-based black-box attacks utilize the adversarial examples' model-level transferability to attack the target model without querying and ensure the secrecy of the attack. Thus, how to enhance the model-level transferability is a key problem of transfer-based black-box attacks. Most existing works apply model ensemble strategies to enhance the transferability among black-box models, however, finding proper models for the same task is not easy and training adversarial patch on multiple models is laborious and costly. To address these issues, in this work, we focus on how to enhance the model-level transferability with only one accessible model instead of model assembling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Though the investigation of adversarial transferability is still in its early stage, the generalizability of neural networks has been investigated for a long time. Intuitively, the association between model optimization and patch optimization can be established by shifting of formal definitions. Given the input pair of data \\( x \\in \\mathcal{X} \\), and label \\( y \\in \\mathcal{Y} \\), the classical model learning is to find a parametric model \\( f_{\\theta} \\) such that \\( f_{\\theta}(x) = y \\), while learning an adversarial patch treats the model \\( f_{\\theta} \\), the original optimization target, as a fixed input to find a hypothesis \\( h \\) of parametric patch \\( \\tau \\) to corrupt the trained model such that \\( h_{\\tau}(f_{\\theta}, x) \\neq y \\). Hence, it is straight forward to analogize patch optimization with regular model optimization. Motivated by the classical approaches for increasing model generalization, we propose our Transfer-based Self-Ensemble Attack (T-SEA), ensembling the input \\( x \\), the attacked model \\( f_{\\theta} \\), and the adversarial patch \\( \\tau \\) from themselves to boost the adversarial transferability of the attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Specifically, we first introduce an enhanced attack baseline based on [32]. Observing from Fig. 4 that the original training strategies have some limitations, we slightly adjust its learning rate scheduler and training patch scale to revise [32] as our enhanced baseline (E-baseline). Then, as shown in Fig. 1, motivated by input augmentation in model optimization (e.g., training data augmentation), we introduce the constrained data augmentation (data self-ensemble) and model ShakeDrop (model self-ensemble), virtually expanding the inputs of patch optimization (i.e., the input data \\( x \\) and the attacked model \\( f \\)) to increase the transferability of the patch against different data and models. Meanwhile, motivated by the dropout technique in model optimization, which utilizes sub-networks of the optimizing model to overcome overfitting and thus increasing model generalization, we propose patch cutout (patch self-ensemble), ran" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "domly performing cutout on the training patch \\(\\tau\\) to overcome overfitting. Through comprehensive experiments, we prove that the proposed E-baseline and self-ensemble strategies perform very well on widely-used detectors with mainstream base attack methods (e.g., PGD [24], MIM [13]). Our contributions can be summarized as the following:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.193, + 0.892, + 0.239 + ], + "angle": 0, + "content": "- We propose a transfer-based black-box attack T-SEA, requiring only one attacked model to achieve a high adversarial transferability attack on object detectors." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.248, + 0.892, + 0.293 + ], + "angle": 0, + "content": "- Observing the issues of the existing approach, we slightly adjust the training strategies to craft an enhanced baseline and increase its performance." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.302, + 0.892, + 0.392 + ], + "angle": 0, + "content": "- Motivated by approaches increasing generalization of deep learning model, we propose a series of strategies to self-ensemble the input data, attacked model, and adversarial patch, which significantly increases the model-level adversarial transferability without introducing extra information." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.402, + 0.892, + 0.478 + ], + "angle": 0, + "content": "- The experimental results demonstrate that the proposed T-SEA can greatly reduce the AP on multiple widely-used detectors on the black-box setting compared to the previous methods, while concurrently performing well with multiple base attack methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.193, + 0.892, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.642, + 0.505 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.514, + 0.766, + 0.529 + ], + "angle": 0, + "content": "2.1. Black-box Adversarial Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Since the discovery of adversarial attack by [31], black-box attacks have gradually attracted more attention, owing to the difficulty of obtaining details of the attacked model in real scenes. Black-box attack can be separated into query-based methods [5, 12, 21] and transfer-based methods [2, 20, 26]. The former utilizes the outputs of the target model to optimize the adversarial examples, which may disclose the attack behavior due to the frequent queries. Hence, in our work, we concentrate on the latter, optimizing adversarial examples in a substitute model without relying on any knowledge of the black-box models. Though the transfer-based attacks ensure the covertness of the attack, their performance is generally limited due to the lack of specialized adjustment catered to the target model. In this work, we propose the self-ensemble strategies to address the issue and greatly boost the black-box attack transferability of the existing transfer-based detector attack." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.667, + 0.818 + ], + "angle": 0, + "content": "2.2. Object Detectors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object detection is a fundamental computer vision technology predicting objects' categories and positions and is widely used in many perception tasks. In the past ten years, deep learning-based models have greatly improved the performance of object detectors. The mainstream methods" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20515" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.881, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Figure 2. The Overall Pipeline of T-SEA. During the training stage, we utilize the self-ensemble strategies (i.e., the constrained data augmentation, patch cutout and model ShakeDrop) to enhance the transferability of the well-optimized adversarial patch. During the attack (inference) stage, we attach the crafted patch into different images to disrupt the detection process of multiple widely-used detectors on black-box setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.529 + ], + "angle": 0, + "content": "based on deep learning can be roughly divided into one-staged and two-staged methods. The former directly predicts the positions and classes of the object instances and are thus faster; the latter first use the region proposal network (RPN) to generate proposals, and then predict the labels and positions of the selected proposals. In this paper, we select eight mainstream detectors of both one-stage and two-stage, including YOLO v2 [27], YOLO v3 & YOLO v3tiny [28], YOLO v4 & YOLO v4tiny [4], YOLO v5 [17], Faster R-CNN [29] and SSD [22] to systemically verify the proposed T-SEA framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.538, + 0.325, + 0.555 + ], + "angle": 0, + "content": "2.3. Attacks on Object Detector" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.473, + 0.833 + ], + "angle": 0, + "content": "The security of deep learning-based models is receiving increasing attention [8, 9], especially with the existence of adversarial attacks. Recent works have explored adversarial robustness of object detectors. To begin, [37] apply adversarial attack on object detectors, performing iterative gradient-based method to misclassify the proposals, and the similar ideas are also carried out by [19, 34]. Ensuingly, for enhancing the attack capability in the physical world, [7, 32] propose real-world adversarial patches to attack the mainstream detectors, e.g., YOLO and Faster R-CNN. Recently, model ensemble approaches are used to improve the attack transferability among multiple detectors, such as [16, 35, 41] simultaneously attacking multiple detectors to generate cross-model adversarial examples. Different from the above, we focus on how to make the most of the limited information (i.e., a single detector) to carry out a high-transferability black-box attack on multiple black-box detectors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.169, + 0.861 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this section, we first give the problem formulation in Sec. 3.1 and describe the overall framework in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.893, + 0.409 + ], + "angle": 0, + "content": "Then, we respectively introduce the details of our enhanced baseline and self-ensemble strategies in Sec. 3.3 and Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.418, + 0.704, + 0.433 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.895, + 0.579 + ], + "angle": 0, + "content": "In this work, we carry out a transfer-based black-box attack with only one white-box detector to decrease the average precision (AP) of both white-box and black-box detectors. Given the target input data distribution \\(\\mathcal{D}(\\mathcal{X},\\mathcal{F})\\), we regard a single pre-trained detector \\(f_{w}\\in \\mathcal{F}\\) as the whitebox attacked model, \\(x_{1,\\dots,N}\\in \\mathcal{X}\\) as the input images, where \\(N\\) is the number of training samples. We are committed to crafting a universal adversarial patch \\(\\tau\\) from the adversarial distribution \\(\\mathcal{T}\\) to disrupt the detection process," + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.588, + 0.893, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\hat {\\tau} = \\underset {\\tau \\sim \\mathcal {T}} {\\arg \\min } \\sum_ {i} ^ {N} L \\left(\\widetilde {x} _ {i}, \\widetilde {f} _ {w}, \\widetilde {\\tau}\\right) + \\lambda J (\\tau), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.703 + ], + "angle": 0, + "content": "where \\(L\\) is the loss function to measure the corruption of detector, \\(\\widetilde{x}_i,\\widetilde{f}_w\\) and \\(\\widetilde{\\tau}\\) denote self-enssembled data, model and patch, respectively, and \\(J(\\cdot)\\) is a regularization term which we employ total variation of \\(\\tau\\) in our work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.691, + 0.727 + ], + "angle": 0, + "content": "3.2. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As depicted in Fig. 2, we divide the entire T-SEA pipeline into training stage and attack stage. During the training stage, the input images are first augmented with the constrained data augmentation. Ensuingly, the white-box detector (i.e., the attacked detector) is used to locate the target objects from the augmented images. Then, we attach adversarial patches to the center of each detected object of the target class, along with patch cutout for mitigating overfitting. After that, the images with adversarial patches go through the shake-dropped models, and we minimize the object confidence and continuously optimize the training" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "adversarial patches until reaching maximum epoch number. During the attack stage, we apply our well-optimized adversarial patch on the test images to disrupt the detection process of multiple black-box detectors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.163, + 0.26, + 0.178 + ], + "angle": 0, + "content": "3.3. Enhanced Baseline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.471, + 0.384 + ], + "angle": 0, + "content": "In this section, we introduce the details of the enhanced baseline, which is described in Algorithm 1. Firstly, we randomly initialize a patch \\(\\tau_0\\), prepare the training images and attacked model \\(f_w\\). During each training epoch, for every input image batch \\(X\\), we first obtain their detection results via the white-box detector \\(f_w\\). Ensuingly, we utilize the transformation function \\(T\\) to apply the training patch into each image to generate adversarial image batch \\(X^{adv}\\). Next, we calculate the object confidence of \\(X^{adv}\\), and the attack loss of the image batch. Here we use \\(\\ell_1\\) loss as the attack loss, unless otherwise stated. Finally, we update the \\(\\tau\\) via base attack method \\(\\hbar\\) and adjust the learning rate after each epoch until reaching the maximum number." + }, + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.399, + 0.392, + 0.414 + ], + "angle": 0, + "content": "Algorithm 1 Enhanced Baseline Based on [32]" + }, + { + "type": "algorithm", + "bbox": [ + 0.079, + 0.417, + 0.47, + 0.676 + ], + "angle": 0, + "content": "Require: \\(x_{1,\\dots,N}\\) (training images), \\(f_{w}\\) (white-box detector), \\(M\\) (maximum epoch), \\(BS\\) (batch size), \\(\\tau_0\\) (the initial patch), \\(T\\) (patch applicer), \\(\\hbar\\) (base attack method), scheduler (learning rate scheduler). \nEnsure: Well-optimized Adversarial Patch \\(\\tau\\) \n1: \\(\\tau \\Leftarrow \\tau_0\\) \n2: for each \\(i \\in [1, M]\\) do \n3: for each \\(j \\in [1, \\frac{N}{BS}]\\) do \n4: \\(X \\Leftarrow x_{(j-1)\\cdot BS+1}, \\ldots, x_{j\\cdot BS}\\) \n5: bboxclean, confclean \\(\\Leftarrow f_w(X)\\) \n6: \\(X^{adv} \\Leftarrow T(X, bbox^{clean}, \\tau)\\) \n7: bbox\\(^{adv}\\), conf\\(^{adv} \\Leftarrow f_w(X^{adv})\\) \n8: loss \\(\\Leftarrow Avg(conf^{adv})\\) \n9: \\(\\tau \\Leftarrow \\hbar(\\tau, loss)\\) \n10: end for \n11: update lr via scheduler \n12: end for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.801 + ], + "angle": 0, + "content": "Compared to AdvPatch [32], we slightly adjust two training strategies. One is the learning rate scheduler. We observe that a saddle point during patch optimization may lead the original plateau-based scheduler to drastically drop the learning rate, causing inadequate optimization. Hence, we adjust the learning rate decline strategy with the following criterion," + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.813, + 0.482, + 0.845 + ], + "angle": 0, + "content": "\\[\nl r \\Leftarrow l r * \\mu , i f \\left(\\ell_ {t} - \\ell_ {t - 1}\\right) < \\epsilon_ {1} a n d \\frac {\\left(\\ell_ {t} - \\ell_ {t - 1}\\right)}{\\ell_ {t}} < \\epsilon_ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the decay factor, \\(\\ell_t\\) denotes the mean loss at epoch \\(t\\), \\(\\epsilon_1, \\epsilon_2\\) are thresholds to control learning rate updates. In our experiments, we ensure that the learning rate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "decrease more stably via the above hyper-parameters. The other adjustment is to reduce the patch scale at training stage \\( s_t \\). Here, the patch scale is the length ratio between the patch and the bounding box of the specific object. As Fig. 4 shows, a proper scale helps the optimized patch to learn more global patterns, which will not be disrupted easily when the patch is scaled, e.g., a local pattern will lose more information when its corresponding patch is scaled to a very small size." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.238, + 0.727, + 0.254 + ], + "angle": 0, + "content": "3.4. Self-Ensemble Strategies" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.261, + 0.697, + 0.276 + ], + "angle": 0, + "content": "3.4.1 Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.405 + ], + "angle": 0, + "content": "Typical model training usually attempts to find a mapping function \\( f \\) from a finite hypothesis space \\( \\mathcal{F} \\) that describes the relationship between the input data \\( x \\) and label \\( y \\) following the underlying joint distribution \\( \\mathcal{D}(\\mathcal{X},\\mathcal{Y}) \\). Generally, since \\( \\mathcal{D} \\) is unknown, we use training data set \\( S = \\{(x_i,y_i)|i = 1,\\dots,N\\} \\) to train the model, which means the optimization involves the empirical risk minimization(ERM) [33]:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.417, + 0.892, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\widehat {R} _ {S} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} L \\left(y _ {i}, f \\left(x _ {i}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.56 + ], + "angle": 0, + "content": "where \\(L\\) is a loss function to measure the difference between the model prediction \\(f(x)\\) and the label \\(y\\). However, the empirical risk is unable to provide generalization on unseen data [40]. Fortunately, we can derive that the generalization error \\(R_{\\mathcal{D}}(f)\\) is bounded by \\(\\widehat{R}_S(f)\\) for a given confidence \\(1 - \\sigma \\in (0,1)\\) [3,25]," + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.572, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\forall f \\in \\mathcal {F}, R _ {\\mathcal {D}} (f) \\leq \\widehat {R} _ {S} (f) + \\sqrt {\\frac {\\log c + \\log \\frac {1}{\\sigma}}{N}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.825 + ], + "angle": 0, + "content": "where \\(N\\) denotes the input data size, \\(c\\) is a complexity measure of \\(\\mathcal{F}\\), such as VC-dimension [25] or covering numbers [1]. This generally yields bounds of the generalization gap \\(R_{\\mathcal{D}}(f) - \\widehat{R}_S(f) = \\mathcal{O}(\\sqrt{\\frac{\\log c}{N}})\\), which gives us a lens to analyze model generalization: we can sample more training data or lower model complexity to help close the generalization gap. In practice, limited training data and large-scaled model parameters make the model prone to overfit [39], causing an undesirable generalization gap increase. Fortunately, past studies have widely adopted a series of regularization strategies on both input data and the model, such as data augmentation [11, 40, 42], Stochastic Depth [15], and Dropout [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motivated by the aforementioned model regularization methods, we discuss how to improve attack transferability of the adversarial patch. During patch optimization, the model \\( f \\) becomes an input instead of the optimization objective, which means the original input distribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.319 + ], + "angle": 0, + "content": "\\(\\mathcal{D}(\\mathcal{X}, \\mathcal{Y})\\) is shifted to a joint input distribution \\(\\mathcal{D}(\\mathcal{X}, \\mathcal{F})\\). Naturally, we can improve generalization of the adversarial patch by increasing training data scale \\(N_{x,f}\\), i.e. sampling more training images \\(x\\) and assembling more white-box models \\(f_w\\). However, obtaining massive images or models of the same task is generally expensive or impractical for the attacker. To address this issue, we propose to virtually expand the input by data augmentation and model ShakeDrop to help improve patch transferability. Meanwhile, although the large-scale models have achieved superior performance in most tasks, these models still benefit from reducing model complexity at training stage like stochastic depth [15] and Dropout [30]. Thus we propose the patch cutout to reduce its capacity \\(c_{\\tau}\\) in training process as a new regularization to alleviate overfitting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.336, + 0.405, + 0.352 + ], + "angle": 0, + "content": "3.4.2 Data: Constrained Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.473, + 0.512 + ], + "angle": 0, + "content": "Motivated by data augmentation in increasing model generalization, we can similarly improve patch transferability by virtually expanding the training set \\( S \\) to more closely approximate the underlying distribution \\( \\mathcal{D} \\). Therefore, we employ a constrained policy to avoid unnatural augmentation which may not appear in natural scenes. In our work, we 1) mildly resize and crop the input images, 2) slightly alter the brightness, contrast, saturation and hue, and 3) randomly rotate the input images in a small range, to generate natural augmented images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.529, + 0.266, + 0.545 + ], + "angle": 0, + "content": "3.4.3 Model: ShakeDrop" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.566, + 0.262, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.727, + 0.213, + 0.739 + ], + "angle": 0, + "content": "(a) Forward" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.569, + 0.447, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.727, + 0.405, + 0.739 + ], + "angle": 0, + "content": "(b) Backward" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.751, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Figure 3. We utilize the model ShakeDrop to linearly combine the outputs of the stacked layers and identity of a residual block." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "As discussed above, when training the patch, sampling more white-box models \\( f \\) from the joint input distribution \\( \\mathcal{D} \\) can improve the adversarial transferability of the crafted patch. However, it is difficult to obtain multiple models of the same task in reality. Inspired by stochastic depth [15], which randomly drops a subset of layers to virtually ensemble multiple model variants for improving generalization," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "we utilize ShakeDrop [38] to linearly combine the outputs of stacked layers and identity of a residual block, generating massive variants of the attacked white-box model as the Fig. 3 illustrated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Specifically, in forward propagation, we combine the identity \\( z_{l} \\) and the output of stacked layers \\( F(z_{l}) \\) of a residual block to generate \\( z_{L} \\) with the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.211, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\[\nz _ {L} = z _ {l} + (\\beta + \\alpha - \\beta \\cdot \\alpha) \\cdot F (z _ {l}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.331 + ], + "angle": 0, + "content": "where \\(\\beta\\) is sampled from a Bernoulli distribution with \\(P(\\beta = 1) = \\varphi_s\\), \\(\\alpha\\) is sampled from a continuous uniform distribution \\(\\alpha \\sim U(1 - e, 1 + e)\\) (\\(e\\) is a constant). In backward propagation, denoting the loss function as \\(\\mathcal{E}\\), from the chain rule of backpropagation, ShakeDrop can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.338, + 0.892, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\psi \\left(\\frac {\\partial \\mathcal {E}}{\\partial z _ {l}}\\right) = \\psi \\left(\\frac {\\partial \\mathcal {E}}{\\partial z _ {L}} \\cdot \\frac {\\partial z _ {L}}{\\partial z _ {l}}\\right) \\tag {6} \\\\ = \\frac {\\partial \\mathcal {E}}{\\partial z _ {L}} (1 + (\\gamma + \\alpha - \\gamma \\cdot \\alpha) \\cdot \\frac {\\partial F}{\\partial z _ {l}}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.443 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is another Bernoulli variable and identically distributed with \\(\\beta\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.461, + 0.655, + 0.475 + ], + "angle": 0, + "content": "3.4.4 Patch: Cutout" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.65 + ], + "angle": 0, + "content": "Dropout [30] is designed to randomly drop neural units along with their connections, effectively preventing overfitting of the model. Inspired by Dropout, we propose a patch cutout strategy, randomly masking a region of the adversarial patch to reduce the patch complexity at training stage and thus to prevent overfitting on specific model and data. From an other angle, model Dropout prevents excessive co-adapting of neural units and spread out features over multi neurons to alleviate overfitting. Similarly, patch cutout also prevents the adversarial nature of patch from relying too much on the patterns of a certain area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.773 + ], + "angle": 0, + "content": "The proposed patch cutout is similar to cutout [11] or random erasing [42] in model training. Specifically, for a normalized input image \\( I_{x,y} \\in [0,1] \\) of size \\( H \\times W \\), we carry out the following process with a probability \\( \\varphi_c \\) before attaching patches to the target objects: 1) we firstly conduct random sampling to obtain one point \\( p = (x_0, y_0) \\) within the given patch; 2) then we cover a \\( \\eta H \\times \\eta W \\) square area \\( (\\eta \\) is the ratio) centered at \\( p \\) with a specific value \\( \\eta \\in [0,1] \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.625, + 0.802 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this section, we first introduce the implementation details in Sec. 4.1. Then, we present the main results of T-SEA in Sec. 4.2, reporting its performance on different detectors and attack methods. After that, we compare the proposed method with other state-of-the-art detection attack algorithms in Sec. 4.3 and perform the ablation study" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20518" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.878, + 0.389 + ], + "angle": 0, + "content": "
MethodsYOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSDBlack-Box Avg↓
AdvPatch5.6640.2618.0748.4924.4443.3839.2741.2836.46
T-SEA(ours)1.734.482.415.687.756.9116.3820.55\\(9.16^{27.30}\\)↓
AdvPatch51.8513.8951.1757.1658.4370.4751.4659.7957.19
T-SEA(ours)31.025.7635.3842.3725.358.0237.6252.26\\(40.28^{16.91}\\)↓
AdvPatch56.0266.121.3969.6451.5672.1656.6160.7361.83
T-SEA(ours)38.8547.130.5152.7531.0161.1849.1255.44\\(47.93^{13.90}\\)↓
AdvPatch37.4137.1817.5819.6726.9146.3744.6743.0736.17
T-SEA(ours)9.346.469.494.2216.6511.6616.2429.43\\(14.18^{21.99}\\)↓
AdvPatch47.4159.5937.1466.4814.5069.5155.9555.2255.90
T-SEA(ours)37.6944.9815.8151.144.1151.3746.0649.39\\(42.35^{13.55}\\)↓
AdvPatch46.954.9329.2062.1646.1513.3947.7150.7348.25
T-SEA(ours)11.6116.7710.7332.5317.551.3717.1430.78\\(19.59^{28.66}\\)↓
AdvPatch24.5323.3714.5826.5425.630.468.6242.1526.75
T-SEA(ours)9.284.083.998.5513.5810.453.0826.46\\(10.91^{15.84}\\)↓
AdvPatch32.4362.9352.0366.2249.0751.1947.4215.1051.61
T-SEA(ours)12.0630.909.7327.339.0817.3828.545.13\\(19.29^{32.32}\\)↓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.402, + 0.895, + 0.445 + ], + "angle": 0, + "content": "Table 1. Comparisons between T-SEA and AdvPatch [32]. We attack each detector separately and use the remaining seven detectors as the black-box detectors. The proposed T-SEA performs much better than [32] on both white-box setting and black-box setting, demonstrating the effectiveness of the proposed self-ensemble strategies." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.45, + 0.432, + 0.645 + ], + "angle": 0, + "content": "
MethodWhite Box ↓Black-Box Avg↓
AdamAdvPatch13.3948.25
T-SEA(ours)1.3719.59
SGDAdvPatch20.3955.97
T-SEA(ours)1.6615.85
MIMAdvPatch11.9133.51
T-SEA(ours)1.4315.43
BIMAdvPatch8.4732.93
T-SEA(ours)1.6219.15
PGDAdvPatch13.5838.70
T-SEA(ours)1.6018.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.658, + 0.472, + 0.742 + ], + "angle": 0, + "content": "Table 2. Comparisons of T-SEA and AdvPatch with Different Base Attack Methods on YOLO v5 (white-box). We select five classical base attack methods, including optimization-based methods (Adam and SGD) and iterative methods (MIM, BIM and PGD). The results show that T-SEA can enhance the performance of all these methods and performs much better than AdvPatch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.788 + ], + "angle": 0, + "content": "in Sec. 4.4. Finally, we show the transferability of T-SEA across the datasets and scenes in Sec. 4.5 and Sec. 4.6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.295, + 0.817 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Datasets In our experiments, we utilize the INRIA person dataset [10] to train and test our adversarial patch, whose training set and test set consist of 614 and 288 images, respectively. Meanwhile, to verify the transferability of the crafted patch, we select images containing per" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.544 + ], + "angle": 0, + "content": "son from COCO validation set (named COCO-person) and CCTV Footage of Humans* (named CCTV-person) as additional test data. The former contains 1684 human images in different scenes (e.g., sports playground, transportation routes, oceans and forests) and the latter contains 559 in-person images from camera footage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.893, + 0.773 + ], + "angle": 0, + "content": "Optimization Details We use INRIA train set as the training set and regard person as the target attack class. The patch size is \\(300 \\times 300\\), the input image size is \\(416 \\times 416\\), the batch size \\(BS = 8\\), and the maximum epoch number \\(M = 1000\\). For the E-baseline, we adjust the training patch scale from 0.2 in AdvPatch [32] to 0.15, set \\(\\epsilon_{1} = 1e - 4\\), \\(\\epsilon_{2} = 1e - 4\\) for the learning rate scheduler, and adopt the Adam [18] as the optimizer. For constrained data augmentation, we carry out constrained data augmentation via horizontal flip, slight color jitter, random resized crop, and random rotation; for model ShakeDrop, the constant \\(e\\) is set to 1 so that \\(\\alpha \\sim U(0,2)\\), and we perturb the model with the probability of \\(\\varphi_{s} = 0.5\\); for patch cutout, we set the fill value of the erased area \\(k = 0.5\\), the ratio \\(\\eta = 0.4\\), and the probability \\(\\varphi_{c} = 0.9\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Evaluation Metric Following [14, 32], we use the Average Precision (AP) to measure the attack capability of the crafted patch (the lower AP, the better attack) and regard the detector's predictions of the clean data as the ground truth (i.e., \\(AP = 1\\))." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.89, + 0.9 + ], + "angle": 0, + "content": "\\*https://www.kaggle.com/datasets/constantinwerner/human-detectiondataset" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20519" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.875, + 0.254 + ], + "angle": 0, + "content": "
MethodWhite Box ↓Black Box ↓Black-Box Avg↓
YOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSD
Gray67.7576.2280.6975.2276.8981.8661.7572.05-
Random Noise70.6775.882.4475.178.7481.7963.4172.9-
White68.5274.8980.274.7376.0980.0960.3569.41-
NPAP [14]38.0356.8558.0467.7467.4366.8556.656.6661.45
AdvCloak [36]33.7454.7753.4267.5756.1268.0555.1960.8259.42
AdvPatch [32]5.6640.2618.0748.4924.4443.3839.2741.2836.46
E-baseline(ours)3.6115.325.5018.5813.399.7729.7323.8216.59
T-SEA(ours)1.734.482.415.687.756.9116.3820.559.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.264, + 0.893, + 0.293 + ], + "angle": 0, + "content": "Table 3. Comparisons with Existing Detection Attack Methods. For clearer controlled observations, we list the results of gray, random noise, and white patch. Compared with existing methods, T-SEA achieves the best performance on both white-box and black-box attack." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.304, + 0.216, + 0.318 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.328, + 0.393, + 0.342 + ], + "angle": 0, + "content": "4.2.1 Results on Different Attacked Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.518 + ], + "angle": 0, + "content": "We systematically investigate attack performance of the proposed T-SEA on eight widely-used object detectors, and report the quantitative results in Tab. 1. Since T-SEA is improved based on AdvPatch [32], we also report the results of AdvPatch. The proposed T-SEA achieves significant improvements on both white-box and black-box performance of all eight detectors compared to [32], demonstrating the effectiveness of the proposed strategies. Meanwhile, for some white-box detectors (e.g., YOLO v2), the black-box average AP can drop to around 10, exhibiting effective black-box performance of T-SEA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.383, + 0.548 + ], + "angle": 0, + "content": "4.2.2 Results of Different Attack Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.47, + 0.751 + ], + "angle": 0, + "content": "T-SEA is not designed only for a specific base attack method; it is important for T-SEA to perform well on different base attack methods. In Tab. 2, we give the detailed results of T-SEA with different base attack methods on YOLO v5, which includes the optimization-based methods (e.g., Adam and SGD) and iterative methods (e.g., MIM, BIM and PGD). The results show that the proposed T-SEA can improve both white-box and black-box attack performance on all methods above, demonstrating that the performance gain caused by T-SEA is not limited to a specific method. That is to say, T-SEA may potentially work well with future base attack methods to further increase the transferability of their crafted AdvPatch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.76, + 0.37, + 0.775 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this section, we compare T-SEA with the SOTA detection attack approaches, which all regard the YOLO v2 as the white-box model and evaluate on seven black-box detectors, and we also report the results of gray/random noise/white patches as control group. As reported on Tab. 3, 1) all adversarial patch performs much better than the control group; 2) our E-baseline and T-SEA achieve the highest performance among all adversarial approaches with the" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.302, + 0.875, + 0.409 + ], + "angle": 0, + "content": "
White Box ↓Black Box ↓
E-baseline13.3948.25
+ Constrained Data Aug.18.4742.42
+ Model ShakeDrop12.5439.40
+ Patch cutout6.8032.54
+ Combined Strategies1.3719.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.513 + ], + "angle": 0, + "content": "Table 4. Ablation Study of Self-Ensemble Strategies on YOLO v5 (white-box). Each self-ensemble strategy is added to the E-baseline to verify its individual performance gain. The results show that: 1) though using data augmentation alone degrades the white-box performance, all strategies can obviously improve the black-box performance of E-baseline; 2) combining all achieves non-trivial results on both white-box and black-box setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.614 + ], + "angle": 0, + "content": "same inference setting (i.e., we ensure the perturbed area is same for each method), showing that compared to the existing single model attack method, the training strategies adjustment and self-ensemble strategies can effectively enhance the attack capability on both white-box and black-box of the crafted adversarial patch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.654, + 0.642 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.774, + 0.667 + ], + "angle": 0, + "content": "4.4.1 Training Strategies Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Here we explore the performance gain from the training strategies adjustments. As discussed in the Sec. 3.3, we modify the learning rate scheduler and training patch scale of [32]. As illustrated in Fig. 4, the modified training scale can reduce the AP in test set more quickly and effectively, and the crafted patch has clearer adversarial patterns. The adjusted scheduler can also lead the detection loss to decrease more, and can be further combined with the modified training scale to achieve a better result." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.834, + 0.727, + 0.849 + ], + "angle": 0, + "content": "4.4.2 Self-Ensemble Strategies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this work, we are motivated by model training approaches that enhance generalization ability, and we propose self-ensemble strategies on the input data, the attacked" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20520" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.299, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.092, + 0.372, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.14, + 0.364, + 0.166 + ], + "angle": 0, + "content": "PlateauLR Scale=0.2 mAP:38.89" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.092, + 0.446, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.14, + 0.442, + 0.166 + ], + "angle": 0, + "content": "PlateauLR Scale=0.15 mAP:36.63" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.166, + 0.372, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.214, + 0.365, + 0.239 + ], + "angle": 0, + "content": "PlateauLR+ Scale=0.2 mAP:35.38" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.167, + 0.449, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.214, + 0.442, + 0.239 + ], + "angle": 0, + "content": "PlateauLR+ Scale=0.15 mAP:24.77" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 4. The improvements after adjusting learning rate scheduler (tagged as \\(\\mathrm{PlateauLR}^{\\dagger}\\)) and patch scale at training stage. A small patch scale during training will cause the test AP to drop lower and faster, and the adjusted \\(\\mathrm{PlateauLR}^{\\dagger}\\) can also cause the detection loss to further decrease." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.34, + 0.48, + 0.438 + ], + "angle": 0, + "content": "
DatasetMethodWhite Box ↓Black Box ↓
COCO-personAdvPatch [32]45.8352.54
T-SEA37.2838.87
CCTV-personAdvPatch [32]38.0734.08
T-SEA38.7119.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Table 5. The Cross Datasets Transferability of T-SEA on YOLO v5 (white-box). Compared to the [32], the adversarial patch crafted by T-SEA has much stronger attack adaptability on person images from different datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.469, + 0.68 + ], + "angle": 0, + "content": "model, and the training patch, greatly augmenting adversarial transferability from themselves. Here we individually inspect the proposed strategies on our E-baseline (YOLO v5 as white-box detector) to explore the impact of each strategy. As reported in Tab. 4, though the constrained data augmentation will slightly decrease the white-box performance (other two strategies both improve the white-box performance), all these self-ensemble strategies increase the black-box results separately. and combining them can achieve the best performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.314, + 0.724 + ], + "angle": 0, + "content": "4.5. Cross-Dataset Verification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The capability of performing cross-dataset attack is also significant for a well-optimized patch. Here, we apply the INRIA-trained patch on two different datasets, COCO-person and CCTV-person. Compared to INRIA, COCO-person has much more person images on different scenes, while CCTV-person focus on persons from the security camera. As shown in Tab. 5, the patch crafted by T-SEA achieves comparable results on the white-box setting with the AdvPatch, but obtains a much better black-box attack capability on both COCO-person and CCTV-person, indicating its strong black-box cross-data attacking capability." + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.088, + 0.84, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.285 + ], + "angle": 0, + "content": "Figure 5. Physical Attack Demo of T-SEA. After showing the optimized patch on iPad, the YOLO v5 and SSD can not detect person." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.292, + 0.694, + 0.307 + ], + "angle": 0, + "content": "4.6. Physical Verification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Although applying the patch attack in the physical world is not the main goal of our work, we believe the optimized patch that has high transferability may perform well in physical settings. Here we show a simple case in Fig. 5 that the patch shown on a iPad can successfully attack the human detection of YOLO v5 and SSD, without disrupting the detection process of other objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.431, + 0.635, + 0.446 + ], + "angle": 0, + "content": "5. Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Since most existing mainstream detectors are CNN-based, we focus on applying the self-ensemble strategies on these detectors in this work. However, the transformer-based detectors [6, 23] are achieving promising results, and the proposed model ShakeDrop can not directly applied on transformers (the others can), we will follow the motivation of designing the model ShakeDrop, to propose new approach and generate variants of transformer-based detector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.584, + 0.619, + 0.599 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.816 + ], + "angle": 0, + "content": "In this paper, we propose a novel transfer-based self-ensemble black-box attack on object detectors, achieving stable and excellent performance gains with various base attack methods on multiple popular object detectors. Firstly, with only slight training strategy adjustments, we improve the existing method's performance and regard it as our enhanced baseline. Then, based on this baseline, we propose a series of self-ensemble strategies to augment the input data, the attacked model, and the training patch from itself to significantly enhance the adversarial transferability of the optimized patch on black-box detectors. The comprehensive experimental results reveal the potential risk that an attacker with only one model could still succeed in a high transferability black-box attack." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.688, + 0.839 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by National Key R&D Program of China (Grant No. 2022ZD0160305) and National Key Laboratory Open Fund of China (Grant No. 6142113210204)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20521" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Martin Anthony and Peter Bartlett. Neural network learning: Theoretical foundations. Cambridge university press, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Song Bai, Yingwei Li, Yuyin Zhou, Qizhu Li, and Philip HS Torr. Adversarial metric attack and defense for person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2119-2126, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[3] Mikhail Belkin, Siyuan Ma, and Soumik Mandal. To understand deep learning we need to understand kernel learning. In International Conference on Machine Learning, pages 541-549. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[5] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Shang-Tse Chen, Cory Cornelius, Jason Martin, and Duen Horng Polo Chau. Shapeshifter: Robust physical adversarial attack on faster r-cnn object detector. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 52-68. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[8] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European Conference on Computer Vision, pages 529-548. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[10] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[11] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.72, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[12] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[13] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[14] Yu-Chih-Tuan Hu, Bo-Han Kung, Daniel Stanley Tan, Jun-Cheng Chen, Kai-Lung Hua, and Wen-Huang Cheng. Naturalistic physical adversarial patch for object detectors. In" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646-661. Springer, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Hao Huang, Yongtao Wang, Zhaoyu Chen, Zhi Tang, Wenqiang Zhang, and Kai-Kuang Ma. Rpattack: Refined patch attack on general object detectors. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[17] Glenn Jocher, Alex Stoken, Jirka Borovec, NanoCode012, ChristopherSTAN, Liu Changyu, Laughing, tkianai, Adam Hogan, lorenzomammana, yxNONG, AlexWang1900, Laurentiu Diaconu, Marc, wanghaoyang0106, ml5ah, Doug, Francisco Ingham, Frederik, Guilhien, Hatovix, Jake Poznanski, Jiacong Fang, Lijun Yu, changyu98, Mingyu Wang, Naman Gupta, Osama Akhtar, PetrDvoracek, and Prashant Rai. ultralytics/yolov5, Oct. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[19] Yuezun Li, Daniel Tian, Ming-Ching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. arXiv preprint arXiv:1809.05962, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[20] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[21] Siao Liu, Zhaoyu Chen, Wei Li, Jiwei Zhu, Jiafeng Wang, Wenqiang Zhang, and Zhongxue Gan. Efficient universal shuffle attack for visual object tracking. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2739-2743. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[25] Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine learning. MIT press, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pages 506-519, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20522" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[27] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[28] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.346, + 0.469, + 0.415 + ], + "angle": 0, + "content": "[32] Simen Thys, Wiebe Van Ranst, and Toon Goedemé. Fooling automated surveillance cameras: adversarial patches to attack person detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.417, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[33] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[34] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.515, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[35] Shudeng Wu, Tao Dai, and Shu-Tao Xia. Dpattack: Diffused patch attacks against universal object detection. arXiv preprint arXiv:2010.11679, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.558, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[36] Zuxuan Wu, Ser-Nam Lim, Larry S Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.615, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.684, + 0.468, + 0.711 + ], + "angle": 0, + "content": "[38] Yoshihiro Yamada, Masakazu Iwamura, and Koichi Kise. Shakedrop regularization. 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.713, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[39] Xue Ying. An overview of overfitting and its solutions. In Journal of physics: Conference series, volume 1168, page 022022. IOP Publishing, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.755, + 0.469, + 0.795 + ], + "angle": 0, + "content": "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.798, + 0.469, + 0.837 + ], + "angle": 0, + "content": "[41] Yusheng Zhao, Huanqian Yan, and Xingxing Wei. Object hider: Adversarial patch attack against object detectors. arXiv preprint arXiv:2010.14974, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.84, + 0.469, + 0.895 + ], + "angle": 0, + "content": "[42] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20523" + } + ] +] \ No newline at end of file diff --git a/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_origin.pdf b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f3a247ee9d50bb24cc30afb7c2c58a9114c8651 --- /dev/null +++ b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/8fc8a2c2-bc87-41bf-a7ad-e07bcd025a9d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e263750cd167ac5c014a2621841c8d65244c156c4c364b85a26aa6f3f73ed1f0 +size 3531586 diff --git a/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/full.md b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a451fd8a9a04269f73343e3bb3ec63ffa50f518 --- /dev/null +++ b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/full.md @@ -0,0 +1,331 @@ +# T-SEA: Transfer-based Self-Ensemble Attack on Object Detection + +Hao Huang* +Peking University +Beijing, China +huanghao@stu.pku.edu.cn + +Ziyan Chen* +Peking University +Beijing, China +chen.ziyan@outlook.com + +Huanran Chen* Peking University Beijing, China huanran_chen@outlook.com + +Yongtao Wang† +Peking University +Beijing, China +wyt@pku.edu.cn + +Kevin Zhang +Peking University +Beijing, China +kevinzyz@pku.edu.cn + +# Abstract + +Compared to query-based black-box attacks, transfer-based black-box attacks do not require any information of the attacked models, which ensures their secrecy. However, most existing transfer-based approaches rely on ensembling multiple models to boost the attack transferability, which is time- and resource-intensive, not to mention the difficulty of obtaining diverse models on the same task. To address this limitation, in this work, we focus on the single-model transfer-based black-box attack on object detection, utilizing only one model to achieve a high-transferability adversarial attack on multiple black-box detectors. Specifically, we first make observations on the patch optimization process of the existing method and propose an enhanced attack framework by slightly adjusting its training strategies. Then, we analyze patch optimization with regular model optimization, proposing a series of self-ensemble approaches on the input data, the attacked model, and the adversarial patch to efficiently make use of the limited information and prevent the patch from overfitting. The experimental results show that the proposed framework can be applied with multiple classical base attack methods (e.g., PGD and MIM) to greatly improve the black-box transferability of the well-optimized patch on multiple mainstream detectors, meanwhile boosting white-box performance. Our code is available at https://github.com/VDIGPKU/T-SEA. + +# 1. Introduction + +With the rapid development of computer vision, deep learning-based object detectors are being widely applied to + +![](images/1184f9effd3d46f673255c5a7de031f2917ff87c1b5020893f1c8be1da94e039.jpg) +Figure 1. Model optimization usually augments the training data and drop out neurons to increase generalization, motivating us to propose self-ensemble methods for adversarial patch optimization. Specifically, inspired by data augmentation in model optimization, we augment the data $x$ and the model $f$ via constrained data augmentation and model ShakeDrop, respectively. Meanwhile, inspired by drop out in model optimization, we cut out the training patch $\tau$ to prevent it overfitting on specific models or images. + +many aspects of our lives, many of which are highly related to our personal safety, including autonomous driving and intelligent security. Unfortunately, recent works [16, 19, 34, 37] have proved that adversarial examples can successfully disrupt the detectors in both digital and physical domains, posing a great threat to detector-based applications. Hence, the mechanism of adversarial examples on detectors should be further explored to help us improve the robustness of detector-based AI applications. + +In real scenes, attackers usually cannot obtain the details of the attacked model, so black-box attacks naturally receive more attention from both academia and industry. + +Generally speaking, black-box adversarial attacks can be classified into 1) query-based and 2) transfer-based. For the former, we usually pre-train an adversarial perturbation on the white-box model and then fine-tune it via the information from the target black-box model, assuming we can access the target model for free. However, frequent queries may expose the attack intent, weakening the covertness of the attack. Contrarily, transfer-based black-box attacks utilize the adversarial examples' model-level transferability to attack the target model without querying and ensure the secrecy of the attack. Thus, how to enhance the model-level transferability is a key problem of transfer-based black-box attacks. Most existing works apply model ensemble strategies to enhance the transferability among black-box models, however, finding proper models for the same task is not easy and training adversarial patch on multiple models is laborious and costly. To address these issues, in this work, we focus on how to enhance the model-level transferability with only one accessible model instead of model assembling. + +Though the investigation of adversarial transferability is still in its early stage, the generalizability of neural networks has been investigated for a long time. Intuitively, the association between model optimization and patch optimization can be established by shifting of formal definitions. Given the input pair of data $x \in \mathcal{X}$ , and label $y \in \mathcal{Y}$ , the classical model learning is to find a parametric model $f_{\theta}$ such that $f_{\theta}(x) = y$ , while learning an adversarial patch treats the model $f_{\theta}$ , the original optimization target, as a fixed input to find a hypothesis $h$ of parametric patch $\tau$ to corrupt the trained model such that $h_{\tau}(f_{\theta}, x) \neq y$ . Hence, it is straight forward to analogize patch optimization with regular model optimization. Motivated by the classical approaches for increasing model generalization, we propose our Transfer-based Self-Ensemble Attack (T-SEA), ensembling the input $x$ , the attacked model $f_{\theta}$ , and the adversarial patch $\tau$ from themselves to boost the adversarial transferability of the attack. + +Specifically, we first introduce an enhanced attack baseline based on [32]. Observing from Fig. 4 that the original training strategies have some limitations, we slightly adjust its learning rate scheduler and training patch scale to revise [32] as our enhanced baseline (E-baseline). Then, as shown in Fig. 1, motivated by input augmentation in model optimization (e.g., training data augmentation), we introduce the constrained data augmentation (data self-ensemble) and model ShakeDrop (model self-ensemble), virtually expanding the inputs of patch optimization (i.e., the input data $x$ and the attacked model $f$ ) to increase the transferability of the patch against different data and models. Meanwhile, motivated by the dropout technique in model optimization, which utilizes sub-networks of the optimizing model to overcome overfitting and thus increasing model generalization, we propose patch cutout (patch self-ensemble), ran + +domly performing cutout on the training patch $\tau$ to overcome overfitting. Through comprehensive experiments, we prove that the proposed E-baseline and self-ensemble strategies perform very well on widely-used detectors with mainstream base attack methods (e.g., PGD [24], MIM [13]). Our contributions can be summarized as the following: + +- We propose a transfer-based black-box attack T-SEA, requiring only one attacked model to achieve a high adversarial transferability attack on object detectors. +- Observing the issues of the existing approach, we slightly adjust the training strategies to craft an enhanced baseline and increase its performance. +- Motivated by approaches increasing generalization of deep learning model, we propose a series of strategies to self-ensemble the input data, attacked model, and adversarial patch, which significantly increases the model-level adversarial transferability without introducing extra information. +- The experimental results demonstrate that the proposed T-SEA can greatly reduce the AP on multiple widely-used detectors on the black-box setting compared to the previous methods, while concurrently performing well with multiple base attack methods. + +# 2. Related Work + +# 2.1. Black-box Adversarial Attack + +Since the discovery of adversarial attack by [31], black-box attacks have gradually attracted more attention, owing to the difficulty of obtaining details of the attacked model in real scenes. Black-box attack can be separated into query-based methods [5, 12, 21] and transfer-based methods [2, 20, 26]. The former utilizes the outputs of the target model to optimize the adversarial examples, which may disclose the attack behavior due to the frequent queries. Hence, in our work, we concentrate on the latter, optimizing adversarial examples in a substitute model without relying on any knowledge of the black-box models. Though the transfer-based attacks ensure the covertness of the attack, their performance is generally limited due to the lack of specialized adjustment catered to the target model. In this work, we propose the self-ensemble strategies to address the issue and greatly boost the black-box attack transferability of the existing transfer-based detector attack. + +# 2.2. Object Detectors + +Object detection is a fundamental computer vision technology predicting objects' categories and positions and is widely used in many perception tasks. In the past ten years, deep learning-based models have greatly improved the performance of object detectors. The mainstream methods + +![](images/2230a575dc7e538de8ed49b5c87171131ee30529dd74d51798e01b9b069e7aee.jpg) +Figure 2. The Overall Pipeline of T-SEA. During the training stage, we utilize the self-ensemble strategies (i.e., the constrained data augmentation, patch cutout and model ShakeDrop) to enhance the transferability of the well-optimized adversarial patch. During the attack (inference) stage, we attach the crafted patch into different images to disrupt the detection process of multiple widely-used detectors on black-box setting. + +based on deep learning can be roughly divided into one-staged and two-staged methods. The former directly predicts the positions and classes of the object instances and are thus faster; the latter first use the region proposal network (RPN) to generate proposals, and then predict the labels and positions of the selected proposals. In this paper, we select eight mainstream detectors of both one-stage and two-stage, including YOLO v2 [27], YOLO v3 & YOLO v3tiny [28], YOLO v4 & YOLO v4tiny [4], YOLO v5 [17], Faster R-CNN [29] and SSD [22] to systemically verify the proposed T-SEA framework. + +# 2.3. Attacks on Object Detector + +The security of deep learning-based models is receiving increasing attention [8, 9], especially with the existence of adversarial attacks. Recent works have explored adversarial robustness of object detectors. To begin, [37] apply adversarial attack on object detectors, performing iterative gradient-based method to misclassify the proposals, and the similar ideas are also carried out by [19, 34]. Ensuingly, for enhancing the attack capability in the physical world, [7, 32] propose real-world adversarial patches to attack the mainstream detectors, e.g., YOLO and Faster R-CNN. Recently, model ensemble approaches are used to improve the attack transferability among multiple detectors, such as [16, 35, 41] simultaneously attacking multiple detectors to generate cross-model adversarial examples. Different from the above, we focus on how to make the most of the limited information (i.e., a single detector) to carry out a high-transferability black-box attack on multiple black-box detectors. + +# 3. Method + +In this section, we first give the problem formulation in Sec. 3.1 and describe the overall framework in Sec. 3.2. + +Then, we respectively introduce the details of our enhanced baseline and self-ensemble strategies in Sec. 3.3 and Sec. 3.4. + +# 3.1. Problem Formulation + +In this work, we carry out a transfer-based black-box attack with only one white-box detector to decrease the average precision (AP) of both white-box and black-box detectors. Given the target input data distribution $\mathcal{D}(\mathcal{X},\mathcal{F})$ , we regard a single pre-trained detector $f_{w}\in \mathcal{F}$ as the whitebox attacked model, $x_{1,\dots,N}\in \mathcal{X}$ as the input images, where $N$ is the number of training samples. We are committed to crafting a universal adversarial patch $\tau$ from the adversarial distribution $\mathcal{T}$ to disrupt the detection process, + +$$ +\hat {\tau} = \underset {\tau \sim \mathcal {T}} {\arg \min } \sum_ {i} ^ {N} L \left(\widetilde {x} _ {i}, \widetilde {f} _ {w}, \widetilde {\tau}\right) + \lambda J (\tau), \tag {1} +$$ + +where $L$ is the loss function to measure the corruption of detector, $\widetilde{x}_i,\widetilde{f}_w$ and $\widetilde{\tau}$ denote self-enssembled data, model and patch, respectively, and $J(\cdot)$ is a regularization term which we employ total variation of $\tau$ in our work. + +# 3.2. Overall Framework + +As depicted in Fig. 2, we divide the entire T-SEA pipeline into training stage and attack stage. During the training stage, the input images are first augmented with the constrained data augmentation. Ensuingly, the white-box detector (i.e., the attacked detector) is used to locate the target objects from the augmented images. Then, we attach adversarial patches to the center of each detected object of the target class, along with patch cutout for mitigating overfitting. After that, the images with adversarial patches go through the shake-dropped models, and we minimize the object confidence and continuously optimize the training + +adversarial patches until reaching maximum epoch number. During the attack stage, we apply our well-optimized adversarial patch on the test images to disrupt the detection process of multiple black-box detectors. + +# 3.3. Enhanced Baseline + +In this section, we introduce the details of the enhanced baseline, which is described in Algorithm 1. Firstly, we randomly initialize a patch $\tau_0$ , prepare the training images and attacked model $f_w$ . During each training epoch, for every input image batch $X$ , we first obtain their detection results via the white-box detector $f_w$ . Ensuingly, we utilize the transformation function $T$ to apply the training patch into each image to generate adversarial image batch $X^{adv}$ . Next, we calculate the object confidence of $X^{adv}$ , and the attack loss of the image batch. Here we use $\ell_1$ loss as the attack loss, unless otherwise stated. Finally, we update the $\tau$ via base attack method $\hbar$ and adjust the learning rate after each epoch until reaching the maximum number. + +Algorithm 1 Enhanced Baseline Based on [32] +Require: $x_{1,\dots,N}$ (training images), $f_{w}$ (white-box detector), $M$ (maximum epoch), $BS$ (batch size), $\tau_0$ (the initial patch), $T$ (patch applicer), $\hbar$ (base attack method), scheduler (learning rate scheduler). +Ensure: Well-optimized Adversarial Patch $\tau$ +1: $\tau \Leftarrow \tau_0$ +2: for each $i \in [1, M]$ do +3: for each $j \in [1, \frac{N}{BS}]$ do +4: $X \Leftarrow x_{(j-1)\cdot BS+1}, \ldots, x_{j\cdot BS}$ +5: bboxclean, confclean $\Leftarrow f_w(X)$ +6: $X^{adv} \Leftarrow T(X, bbox^{clean}, \tau)$ +7: bbox $^{adv}$ , conf $^{adv} \Leftarrow f_w(X^{adv})$ +8: loss $\Leftarrow Avg(conf^{adv})$ +9: $\tau \Leftarrow \hbar(\tau, loss)$ +10: end for +11: update lr via scheduler +12: end for + +Compared to AdvPatch [32], we slightly adjust two training strategies. One is the learning rate scheduler. We observe that a saddle point during patch optimization may lead the original plateau-based scheduler to drastically drop the learning rate, causing inadequate optimization. Hence, we adjust the learning rate decline strategy with the following criterion, + +$$ +l r \Leftarrow l r * \mu , i f \left(\ell_ {t} - \ell_ {t - 1}\right) < \epsilon_ {1} a n d \frac {\left(\ell_ {t} - \ell_ {t - 1}\right)}{\ell_ {t}} < \epsilon_ {2}, \tag {2} +$$ + +where $\mu$ is the decay factor, $\ell_t$ denotes the mean loss at epoch $t$ , $\epsilon_1, \epsilon_2$ are thresholds to control learning rate updates. In our experiments, we ensure that the learning rate + +decrease more stably via the above hyper-parameters. The other adjustment is to reduce the patch scale at training stage $s_t$ . Here, the patch scale is the length ratio between the patch and the bounding box of the specific object. As Fig. 4 shows, a proper scale helps the optimized patch to learn more global patterns, which will not be disrupted easily when the patch is scaled, e.g., a local pattern will lose more information when its corresponding patch is scaled to a very small size. + +# 3.4. Self-Ensemble Strategies + +# 3.4.1 Theoretical Analysis + +Typical model training usually attempts to find a mapping function $f$ from a finite hypothesis space $\mathcal{F}$ that describes the relationship between the input data $x$ and label $y$ following the underlying joint distribution $\mathcal{D}(\mathcal{X},\mathcal{Y})$ . Generally, since $\mathcal{D}$ is unknown, we use training data set $S = \{(x_i,y_i)|i = 1,\dots,N\}$ to train the model, which means the optimization involves the empirical risk minimization(ERM) [33]: + +$$ +\widehat {R} _ {S} (f) = \frac {1}{N} \sum_ {i} ^ {N} L \left(y _ {i}, f \left(x _ {i}\right)\right), \tag {3} +$$ + +where $L$ is a loss function to measure the difference between the model prediction $f(x)$ and the label $y$ . However, the empirical risk is unable to provide generalization on unseen data [40]. Fortunately, we can derive that the generalization error $R_{\mathcal{D}}(f)$ is bounded by $\widehat{R}_S(f)$ for a given confidence $1 - \sigma \in (0,1)$ [3,25], + +$$ +\forall f \in \mathcal {F}, R _ {\mathcal {D}} (f) \leq \widehat {R} _ {S} (f) + \sqrt {\frac {\log c + \log \frac {1}{\sigma}}{N}}, \tag {4} +$$ + +where $N$ denotes the input data size, $c$ is a complexity measure of $\mathcal{F}$ , such as VC-dimension [25] or covering numbers [1]. This generally yields bounds of the generalization gap $R_{\mathcal{D}}(f) - \widehat{R}_S(f) = \mathcal{O}(\sqrt{\frac{\log c}{N}})$ , which gives us a lens to analyze model generalization: we can sample more training data or lower model complexity to help close the generalization gap. In practice, limited training data and large-scaled model parameters make the model prone to overfit [39], causing an undesirable generalization gap increase. Fortunately, past studies have widely adopted a series of regularization strategies on both input data and the model, such as data augmentation [11, 40, 42], Stochastic Depth [15], and Dropout [30]. + +Motivated by the aforementioned model regularization methods, we discuss how to improve attack transferability of the adversarial patch. During patch optimization, the model $f$ becomes an input instead of the optimization objective, which means the original input distribution + +$\mathcal{D}(\mathcal{X}, \mathcal{Y})$ is shifted to a joint input distribution $\mathcal{D}(\mathcal{X}, \mathcal{F})$ . Naturally, we can improve generalization of the adversarial patch by increasing training data scale $N_{x,f}$ , i.e. sampling more training images $x$ and assembling more white-box models $f_w$ . However, obtaining massive images or models of the same task is generally expensive or impractical for the attacker. To address this issue, we propose to virtually expand the input by data augmentation and model ShakeDrop to help improve patch transferability. Meanwhile, although the large-scale models have achieved superior performance in most tasks, these models still benefit from reducing model complexity at training stage like stochastic depth [15] and Dropout [30]. Thus we propose the patch cutout to reduce its capacity $c_{\tau}$ in training process as a new regularization to alleviate overfitting. + +# 3.4.2 Data: Constrained Data Augmentation + +Motivated by data augmentation in increasing model generalization, we can similarly improve patch transferability by virtually expanding the training set $S$ to more closely approximate the underlying distribution $\mathcal{D}$ . Therefore, we employ a constrained policy to avoid unnatural augmentation which may not appear in natural scenes. In our work, we 1) mildly resize and crop the input images, 2) slightly alter the brightness, contrast, saturation and hue, and 3) randomly rotate the input images in a small range, to generate natural augmented images. + +# 3.4.3 Model: ShakeDrop + +![](images/734ba65804b6e881505ff51289a1a3a33114e5adf16772fe42f0acbe88435e97.jpg) +(a) Forward +Figure 3. We utilize the model ShakeDrop to linearly combine the outputs of the stacked layers and identity of a residual block. + +![](images/82d0b5cc5ee8845bec1f99a1cd693c00ab4ab817f51dbb0b8f5181f19a69e5dc.jpg) +(b) Backward + +As discussed above, when training the patch, sampling more white-box models $f$ from the joint input distribution $\mathcal{D}$ can improve the adversarial transferability of the crafted patch. However, it is difficult to obtain multiple models of the same task in reality. Inspired by stochastic depth [15], which randomly drops a subset of layers to virtually ensemble multiple model variants for improving generalization, + +we utilize ShakeDrop [38] to linearly combine the outputs of stacked layers and identity of a residual block, generating massive variants of the attacked white-box model as the Fig. 3 illustrated. + +Specifically, in forward propagation, we combine the identity $z_{l}$ and the output of stacked layers $F(z_{l})$ of a residual block to generate $z_{L}$ with the following formula: + +$$ +z _ {L} = z _ {l} + (\beta + \alpha - \beta \cdot \alpha) \cdot F (z _ {l}), \tag {5} +$$ + +where $\beta$ is sampled from a Bernoulli distribution with $P(\beta = 1) = \varphi_s$ , $\alpha$ is sampled from a continuous uniform distribution $\alpha \sim U(1 - e, 1 + e)$ ( $e$ is a constant). In backward propagation, denoting the loss function as $\mathcal{E}$ , from the chain rule of backpropagation, ShakeDrop can be formulated as + +$$ +\begin{array}{l} \psi \left(\frac {\partial \mathcal {E}}{\partial z _ {l}}\right) = \psi \left(\frac {\partial \mathcal {E}}{\partial z _ {L}} \cdot \frac {\partial z _ {L}}{\partial z _ {l}}\right) \tag {6} \\ = \frac {\partial \mathcal {E}}{\partial z _ {L}} (1 + (\gamma + \alpha - \gamma \cdot \alpha) \cdot \frac {\partial F}{\partial z _ {l}}), \\ \end{array} +$$ + +where $\gamma$ is another Bernoulli variable and identically distributed with $\beta$ . + +# 3.4.4 Patch: Cutout + +Dropout [30] is designed to randomly drop neural units along with their connections, effectively preventing overfitting of the model. Inspired by Dropout, we propose a patch cutout strategy, randomly masking a region of the adversarial patch to reduce the patch complexity at training stage and thus to prevent overfitting on specific model and data. From an other angle, model Dropout prevents excessive co-adapting of neural units and spread out features over multi neurons to alleviate overfitting. Similarly, patch cutout also prevents the adversarial nature of patch from relying too much on the patterns of a certain area. + +The proposed patch cutout is similar to cutout [11] or random erasing [42] in model training. Specifically, for a normalized input image $I_{x,y} \in [0,1]$ of size $H \times W$ , we carry out the following process with a probability $\varphi_c$ before attaching patches to the target objects: 1) we firstly conduct random sampling to obtain one point $p = (x_0, y_0)$ within the given patch; 2) then we cover a $\eta H \times \eta W$ square area $(\eta$ is the ratio) centered at $p$ with a specific value $\eta \in [0,1]$ . + +# 4. Experiment + +In this section, we first introduce the implementation details in Sec. 4.1. Then, we present the main results of T-SEA in Sec. 4.2, reporting its performance on different detectors and attack methods. After that, we compare the proposed method with other state-of-the-art detection attack algorithms in Sec. 4.3 and perform the ablation study + +
MethodsYOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSDBlack-Box Avg↓
AdvPatch5.6640.2618.0748.4924.4443.3839.2741.2836.46
T-SEA(ours)1.734.482.415.687.756.9116.3820.55\(9.16^{27.30}\)↓
AdvPatch51.8513.8951.1757.1658.4370.4751.4659.7957.19
T-SEA(ours)31.025.7635.3842.3725.358.0237.6252.26\(40.28^{16.91}\)↓
AdvPatch56.0266.121.3969.6451.5672.1656.6160.7361.83
T-SEA(ours)38.8547.130.5152.7531.0161.1849.1255.44\(47.93^{13.90}\)↓
AdvPatch37.4137.1817.5819.6726.9146.3744.6743.0736.17
T-SEA(ours)9.346.469.494.2216.6511.6616.2429.43\(14.18^{21.99}\)↓
AdvPatch47.4159.5937.1466.4814.5069.5155.9555.2255.90
T-SEA(ours)37.6944.9815.8151.144.1151.3746.0649.39\(42.35^{13.55}\)↓
AdvPatch46.954.9329.2062.1646.1513.3947.7150.7348.25
T-SEA(ours)11.6116.7710.7332.5317.551.3717.1430.78\(19.59^{28.66}\)↓
AdvPatch24.5323.3714.5826.5425.630.468.6242.1526.75
T-SEA(ours)9.284.083.998.5513.5810.453.0826.46\(10.91^{15.84}\)↓
AdvPatch32.4362.9352.0366.2249.0751.1947.4215.1051.61
T-SEA(ours)12.0630.909.7327.339.0817.3828.545.13\(19.29^{32.32}\)↓
+ +Table 1. Comparisons between T-SEA and AdvPatch [32]. We attack each detector separately and use the remaining seven detectors as the black-box detectors. The proposed T-SEA performs much better than [32] on both white-box setting and black-box setting, demonstrating the effectiveness of the proposed self-ensemble strategies. + +
MethodWhite Box ↓Black-Box Avg↓
AdamAdvPatch13.3948.25
T-SEA(ours)1.3719.59
SGDAdvPatch20.3955.97
T-SEA(ours)1.6615.85
MIMAdvPatch11.9133.51
T-SEA(ours)1.4315.43
BIMAdvPatch8.4732.93
T-SEA(ours)1.6219.15
PGDAdvPatch13.5838.70
T-SEA(ours)1.6018.06
+ +Table 2. Comparisons of T-SEA and AdvPatch with Different Base Attack Methods on YOLO v5 (white-box). We select five classical base attack methods, including optimization-based methods (Adam and SGD) and iterative methods (MIM, BIM and PGD). The results show that T-SEA can enhance the performance of all these methods and performs much better than AdvPatch. + +in Sec. 4.4. Finally, we show the transferability of T-SEA across the datasets and scenes in Sec. 4.5 and Sec. 4.6. + +# 4.1. Implementation Details + +Datasets In our experiments, we utilize the INRIA person dataset [10] to train and test our adversarial patch, whose training set and test set consist of 614 and 288 images, respectively. Meanwhile, to verify the transferability of the crafted patch, we select images containing per + +son from COCO validation set (named COCO-person) and CCTV Footage of Humans* (named CCTV-person) as additional test data. The former contains 1684 human images in different scenes (e.g., sports playground, transportation routes, oceans and forests) and the latter contains 559 in-person images from camera footage. + +Optimization Details We use INRIA train set as the training set and regard person as the target attack class. The patch size is $300 \times 300$ , the input image size is $416 \times 416$ , the batch size $BS = 8$ , and the maximum epoch number $M = 1000$ . For the E-baseline, we adjust the training patch scale from 0.2 in AdvPatch [32] to 0.15, set $\epsilon_{1} = 1e - 4$ , $\epsilon_{2} = 1e - 4$ for the learning rate scheduler, and adopt the Adam [18] as the optimizer. For constrained data augmentation, we carry out constrained data augmentation via horizontal flip, slight color jitter, random resized crop, and random rotation; for model ShakeDrop, the constant $e$ is set to 1 so that $\alpha \sim U(0,2)$ , and we perturb the model with the probability of $\varphi_{s} = 0.5$ ; for patch cutout, we set the fill value of the erased area $k = 0.5$ , the ratio $\eta = 0.4$ , and the probability $\varphi_{c} = 0.9$ . + +Evaluation Metric Following [14, 32], we use the Average Precision (AP) to measure the attack capability of the crafted patch (the lower AP, the better attack) and regard the detector's predictions of the clean data as the ground truth (i.e., $AP = 1$ ). + +
MethodWhite Box ↓Black Box ↓Black-Box Avg↓
YOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSD
Gray67.7576.2280.6975.2276.8981.8661.7572.05-
Random Noise70.6775.882.4475.178.7481.7963.4172.9-
White68.5274.8980.274.7376.0980.0960.3569.41-
NPAP [14]38.0356.8558.0467.7467.4366.8556.656.6661.45
AdvCloak [36]33.7454.7753.4267.5756.1268.0555.1960.8259.42
AdvPatch [32]5.6640.2618.0748.4924.4443.3839.2741.2836.46
E-baseline(ours)3.6115.325.5018.5813.399.7729.7323.8216.59
T-SEA(ours)1.734.482.415.687.756.9116.3820.559.16
+ +# 4.2. Main Results + +# 4.2.1 Results on Different Attacked Models + +We systematically investigate attack performance of the proposed T-SEA on eight widely-used object detectors, and report the quantitative results in Tab. 1. Since T-SEA is improved based on AdvPatch [32], we also report the results of AdvPatch. The proposed T-SEA achieves significant improvements on both white-box and black-box performance of all eight detectors compared to [32], demonstrating the effectiveness of the proposed strategies. Meanwhile, for some white-box detectors (e.g., YOLO v2), the black-box average AP can drop to around 10, exhibiting effective black-box performance of T-SEA. + +# 4.2.2 Results of Different Attack Methods + +T-SEA is not designed only for a specific base attack method; it is important for T-SEA to perform well on different base attack methods. In Tab. 2, we give the detailed results of T-SEA with different base attack methods on YOLO v5, which includes the optimization-based methods (e.g., Adam and SGD) and iterative methods (e.g., MIM, BIM and PGD). The results show that the proposed T-SEA can improve both white-box and black-box attack performance on all methods above, demonstrating that the performance gain caused by T-SEA is not limited to a specific method. That is to say, T-SEA may potentially work well with future base attack methods to further increase the transferability of their crafted AdvPatch. + +# 4.3. Comparison with SOTA Methods + +In this section, we compare T-SEA with the SOTA detection attack approaches, which all regard the YOLO v2 as the white-box model and evaluate on seven black-box detectors, and we also report the results of gray/random noise/white patches as control group. As reported on Tab. 3, 1) all adversarial patch performs much better than the control group; 2) our E-baseline and T-SEA achieve the highest performance among all adversarial approaches with the + +Table 3. Comparisons with Existing Detection Attack Methods. For clearer controlled observations, we list the results of gray, random noise, and white patch. Compared with existing methods, T-SEA achieves the best performance on both white-box and black-box attack. + +
White Box ↓Black Box ↓
E-baseline13.3948.25
+ Constrained Data Aug.18.4742.42
+ Model ShakeDrop12.5439.40
+ Patch cutout6.8032.54
+ Combined Strategies1.3719.59
+ +Table 4. Ablation Study of Self-Ensemble Strategies on YOLO v5 (white-box). Each self-ensemble strategy is added to the E-baseline to verify its individual performance gain. The results show that: 1) though using data augmentation alone degrades the white-box performance, all strategies can obviously improve the black-box performance of E-baseline; 2) combining all achieves non-trivial results on both white-box and black-box setting. + +same inference setting (i.e., we ensure the perturbed area is same for each method), showing that compared to the existing single model attack method, the training strategies adjustment and self-ensemble strategies can effectively enhance the attack capability on both white-box and black-box of the crafted adversarial patch. + +# 4.4. Ablation Study + +# 4.4.1 Training Strategies Adjustment + +Here we explore the performance gain from the training strategies adjustments. As discussed in the Sec. 3.3, we modify the learning rate scheduler and training patch scale of [32]. As illustrated in Fig. 4, the modified training scale can reduce the AP in test set more quickly and effectively, and the crafted patch has clearer adversarial patterns. The adjusted scheduler can also lead the detection loss to decrease more, and can be further combined with the modified training scale to achieve a better result. + +# 4.4.2 Self-Ensemble Strategies + +In this work, we are motivated by model training approaches that enhance generalization ability, and we propose self-ensemble strategies on the input data, the attacked + +![](images/96cd91f62347db130d1f0a83095630b1e06fd01cb3fb76029f5f0b720d6eb67c.jpg) +Figure 4. The improvements after adjusting learning rate scheduler (tagged as $\mathrm{PlateauLR}^{\dagger}$ ) and patch scale at training stage. A small patch scale during training will cause the test AP to drop lower and faster, and the adjusted $\mathrm{PlateauLR}^{\dagger}$ can also cause the detection loss to further decrease. + +![](images/afbb33ffe759763f1f868375778744d1920f4fc4cdac8c02e30fad05670f12b8.jpg) +PlateauLR Scale=0.2 mAP:38.89 + +![](images/0f769d13b97809adac618432521c48bdedd51f992f13bcfdd9df84fd582b7c19.jpg) +PlateauLR Scale=0.15 mAP:36.63 + +![](images/e1dfc0af9db60426bbe2b98ff55d21bc49cf30025764186e336bbf7ffc9cba3c.jpg) +PlateauLR+ Scale=0.2 mAP:35.38 + +![](images/7faa8157c411a78bf38658b62f1d78fa73a19b54f34ab6161b126a9e9a3a6391.jpg) +PlateauLR+ Scale=0.15 mAP:24.77 + +
DatasetMethodWhite Box ↓Black Box ↓
COCO-personAdvPatch [32]45.8352.54
T-SEA37.2838.87
CCTV-personAdvPatch [32]38.0734.08
T-SEA38.7119.91
+ +Table 5. The Cross Datasets Transferability of T-SEA on YOLO v5 (white-box). Compared to the [32], the adversarial patch crafted by T-SEA has much stronger attack adaptability on person images from different datasets. + +model, and the training patch, greatly augmenting adversarial transferability from themselves. Here we individually inspect the proposed strategies on our E-baseline (YOLO v5 as white-box detector) to explore the impact of each strategy. As reported in Tab. 4, though the constrained data augmentation will slightly decrease the white-box performance (other two strategies both improve the white-box performance), all these self-ensemble strategies increase the black-box results separately. and combining them can achieve the best performance. + +# 4.5. Cross-Dataset Verification + +The capability of performing cross-dataset attack is also significant for a well-optimized patch. Here, we apply the INRIA-trained patch on two different datasets, COCO-person and CCTV-person. Compared to INRIA, COCO-person has much more person images on different scenes, while CCTV-person focus on persons from the security camera. As shown in Tab. 5, the patch crafted by T-SEA achieves comparable results on the white-box setting with the AdvPatch, but obtains a much better black-box attack capability on both COCO-person and CCTV-person, indicating its strong black-box cross-data attacking capability. + +![](images/4c59c3f8efada64ba0dfe99e073bd1deb52862c7a70f3578d1bebf15c0a5efa0.jpg) +Figure 5. Physical Attack Demo of T-SEA. After showing the optimized patch on iPad, the YOLO v5 and SSD can not detect person. + +# 4.6. Physical Verification + +Although applying the patch attack in the physical world is not the main goal of our work, we believe the optimized patch that has high transferability may perform well in physical settings. Here we show a simple case in Fig. 5 that the patch shown on a iPad can successfully attack the human detection of YOLO v5 and SSD, without disrupting the detection process of other objects. + +# 5. Future Work + +Since most existing mainstream detectors are CNN-based, we focus on applying the self-ensemble strategies on these detectors in this work. However, the transformer-based detectors [6, 23] are achieving promising results, and the proposed model ShakeDrop can not directly applied on transformers (the others can), we will follow the motivation of designing the model ShakeDrop, to propose new approach and generate variants of transformer-based detector. + +# 6. Conclusion + +In this paper, we propose a novel transfer-based self-ensemble black-box attack on object detectors, achieving stable and excellent performance gains with various base attack methods on multiple popular object detectors. Firstly, with only slight training strategy adjustments, we improve the existing method's performance and regard it as our enhanced baseline. Then, based on this baseline, we propose a series of self-ensemble strategies to augment the input data, the attacked model, and the training patch from itself to significantly enhance the adversarial transferability of the optimized patch on black-box detectors. The comprehensive experimental results reveal the potential risk that an attacker with only one model could still succeed in a high transferability black-box attack. + +# 7. Acknowledgements + +This work was supported by National Key R&D Program of China (Grant No. 2022ZD0160305) and National Key Laboratory Open Fund of China (Grant No. 6142113210204). + +# References + +[1] Martin Anthony and Peter Bartlett. Neural network learning: Theoretical foundations. Cambridge university press, 1999. 4 +[2] Song Bai, Yingwei Li, Yuyin Zhou, Qizhu Li, and Philip HS Torr. Adversarial metric attack and defense for person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2119-2126, 2020. 2 +[3] Mikhail Belkin, Siyuan Ma, and Soumik Mandal. To understand deep learning we need to understand kernel learning. In International Conference on Machine Learning, pages 541-549. PMLR, 2018. 4 +[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3 +[5] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248, 2017. 2 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 8 +[7] Shang-Tse Chen, Cory Cornelius, Jason Martin, and Duen Horng Polo Chau. Shapeshifter: Robust physical adversarial attack on faster r-cnn object detector. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 52-68. Springer, 2018. 3 +[8] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European Conference on Computer Vision, pages 529-548. Springer, 2022. 3 +[9] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 3 +[10] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 6 +[11] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4, 5 +[12] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2020. 2 +[13] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2 +[14] Yu-Chih-Tuan Hu, Bo-Han Kung, Daniel Stanley Tan, Jun-Cheng Chen, Kai-Lung Hua, and Wen-Huang Cheng. Naturalistic physical adversarial patch for object detectors. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7 +[15] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646-661. Springer, 2016. 4, 5 +[16] Hao Huang, Yongtao Wang, Zhaoyu Chen, Zhi Tang, Wenqiang Zhang, and Kai-Kuang Ma. Rpattack: Refined patch attack on general object detectors. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 3 +[17] Glenn Jocher, Alex Stoken, Jirka Borovec, NanoCode012, ChristopherSTAN, Liu Changyu, Laughing, tkianai, Adam Hogan, lorenzomammana, yxNONG, AlexWang1900, Laurentiu Diaconu, Marc, wanghaoyang0106, ml5ah, Doug, Francisco Ingham, Frederik, Guilhien, Hatovix, Jake Poznanski, Jiacong Fang, Lijun Yu, changyu98, Mingyu Wang, Naman Gupta, Osama Akhtar, PetrDvoracek, and Prashant Rai. ultralytics/yolov5, Oct. 2020. 3 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Yuezun Li, Daniel Tian, Ming-Ching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. arXiv preprint arXiv:1809.05962, 2018. 1, 3 +[20] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations. 2 +[21] Siao Liu, Zhaoyu Chen, Wei Li, Jiwei Zhu, Jiafeng Wang, Wenqiang Zhang, and Zhongxue Gan. Efficient universal shuffle attack for visual object tracking. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2739-2743. IEEE, 2022. 2 +[22] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 3 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8 +[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2 +[25] Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine learning. MIT press, 2018. 4 +[26] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pages 506-519, 2017. 2 + +[27] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 3 +[28] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 3 +[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4, 5 +[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2 +[32] Simen Thys, Wiebe Van Ranst, and Toon Goedemé. Fooling automated surveillance cameras: adversarial patches to attack person detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 3, 4, 6, 7, 8 +[33] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 4 +[34] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 1, 3 +[35] Shudeng Wu, Tao Dai, and Shu-Tao Xia. Dpattack: Diffused patch attacks against universal object detection. arXiv preprint arXiv:2010.11679, 2020. 3 +[36] Zuxuan Wu, Ser-Nam Lim, Larry S Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 7 +[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 1, 3 +[38] Yoshihiro Yamada, Masakazu Iwamura, and Koichi Kise. Shakedrop regularization. 2018. 5 +[39] Xue Ying. An overview of overfitting and its solutions. In Journal of physics: Conference series, volume 1168, page 022022. IOP Publishing, 2019. 4 +[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4 +[41] Yusheng Zhao, Huanqian Yan, and Xingxing Wei. Object hider: Adversarial patch attack against object detectors. arXiv preprint arXiv:2010.14974, 2020. 3 +[42] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 4, 5 \ No newline at end of file diff --git a/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/images.zip b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec619b9764f2603cdfdab22a3b2e9c1c074f380c --- /dev/null +++ b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61d44c7ba01ef60f1fdb9eacdd75490f89a94e81aa006d60804e09351e6d82d6 +size 524864 diff --git a/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/layout.json b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ceb74cf9ef66ea3f18a9e50a33853d328793ada --- /dev/null +++ b/2023/T-SEA_ Transfer-Based Self-Ensemble Attack on Object Detection/layout.json @@ -0,0 +1,9096 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 121 + ], + "type": "text", + "content": "T-SEA: Transfer-based Self-Ensemble Attack on Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 143, + 208, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 143, + 208, + 198 + ], + "spans": [ + { + "bbox": [ + 81, + 143, + 208, + 198 + ], + "type": "text", + "content": "Hao Huang* \nPeking University \nBeijing, China \nhuanghao@stu.pku.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 235, + 144, + 356, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 144, + 356, + 198 + ], + "spans": [ + { + "bbox": [ + 235, + 144, + 356, + 198 + ], + "type": "text", + "content": "Ziyan Chen* \nPeking University \nBeijing, China \nchen.ziyan@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 381, + 144, + 511, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 144, + 511, + 198 + ], + "spans": [ + { + "bbox": [ + 381, + 144, + 511, + 198 + ], + "type": "text", + "content": "Huanran Chen* Peking University Beijing, China huanran_chen@outlook.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 205, + 265, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 205, + 265, + 261 + ], + "spans": [ + { + "bbox": [ + 177, + 205, + 265, + 261 + ], + "type": "text", + "content": "Yongtao Wang† \nPeking University \nBeijing, China \nwyt@pku.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 205, + 412, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 412, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 412, + 261 + ], + "type": "text", + "content": "Kevin Zhang \nPeking University \nBeijing, China \nkevinzyz@pku.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 289, + 191, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 289, + 191, + 301 + ], + "spans": [ + { + "bbox": [ + 143, + 289, + 191, + 301 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 315, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 289, + 624 + ], + "type": "text", + "content": "Compared to query-based black-box attacks, transfer-based black-box attacks do not require any information of the attacked models, which ensures their secrecy. However, most existing transfer-based approaches rely on ensembling multiple models to boost the attack transferability, which is time- and resource-intensive, not to mention the difficulty of obtaining diverse models on the same task. To address this limitation, in this work, we focus on the single-model transfer-based black-box attack on object detection, utilizing only one model to achieve a high-transferability adversarial attack on multiple black-box detectors. Specifically, we first make observations on the patch optimization process of the existing method and propose an enhanced attack framework by slightly adjusting its training strategies. Then, we analyze patch optimization with regular model optimization, proposing a series of self-ensemble approaches on the input data, the attacked model, and the adversarial patch to efficiently make use of the limited information and prevent the patch from overfitting. The experimental results show that the proposed framework can be applied with multiple classical base attack methods (e.g., PGD and MIM) to greatly improve the black-box transferability of the well-optimized patch on multiple mainstream detectors, meanwhile boosting white-box performance. Our code is available at https://github.com/VDIGPKU/T-SEA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 637, + 127, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 127, + 649 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 127, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 658, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 682 + ], + "type": "text", + "content": "With the rapid development of computer vision, deep learning-based object detectors are being widely applied to" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 292, + 536, + 458 + ], + "blocks": [ + { + "bbox": [ + 308, + 292, + 536, + 458 + ], + "lines": [ + { + "bbox": [ + 308, + 292, + 536, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 536, + 458 + ], + "type": "image", + "image_path": "1184f9effd3d46f673255c5a7de031f2917ff87c1b5020893f1c8be1da94e039.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "text", + "content": "Figure 1. Model optimization usually augments the training data and drop out neurons to increase generalization, motivating us to propose self-ensemble methods for adversarial patch optimization. Specifically, inspired by data augmentation in model optimization, we augment the data " + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "text", + "content": " and the model " + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "text", + "content": " via constrained data augmentation and model ShakeDrop, respectively. Meanwhile, inspired by drop out in model optimization, we cut out the training patch " + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 465, + 545, + 555 + ], + "type": "text", + "content": " to prevent it overfitting on specific models or images." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 674 + ], + "type": "text", + "content": "many aspects of our lives, many of which are highly related to our personal safety, including autonomous driving and intelligent security. Unfortunately, recent works [16, 19, 34, 37] have proved that adversarial examples can successfully disrupt the detectors in both digital and physical domains, posing a great threat to detector-based applications. Hence, the mechanism of adversarial examples on detectors should be further explored to help us improve the robustness of detector-based AI applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In real scenes, attackers usually cannot obtain the details of the attacked model, so black-box attacks naturally receive more attention from both academia and industry." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 214, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 214, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 214, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20514" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "Generally speaking, black-box adversarial attacks can be classified into 1) query-based and 2) transfer-based. For the former, we usually pre-train an adversarial perturbation on the white-box model and then fine-tune it via the information from the target black-box model, assuming we can access the target model for free. However, frequent queries may expose the attack intent, weakening the covertness of the attack. Contrarily, transfer-based black-box attacks utilize the adversarial examples' model-level transferability to attack the target model without querying and ensure the secrecy of the attack. Thus, how to enhance the model-level transferability is a key problem of transfer-based black-box attacks. Most existing works apply model ensemble strategies to enhance the transferability among black-box models, however, finding proper models for the same task is not easy and training adversarial patch on multiple models is laborious and costly. To address these issues, in this work, we focus on how to enhance the model-level transferability with only one accessible model instead of model assembling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": "Though the investigation of adversarial transferability is still in its early stage, the generalizability of neural networks has been investigated for a long time. Intuitively, the association between model optimization and patch optimization can be established by shifting of formal definitions. Given the input pair of data " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ", and label " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ", the classical model learning is to find a parametric model " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x) = y" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ", while learning an adversarial patch treats the model " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ", the original optimization target, as a fixed input to find a hypothesis " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": " of parametric patch " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": " to corrupt the trained model such that " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "h_{\\tau}(f_{\\theta}, x) \\neq y" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ". Hence, it is straight forward to analogize patch optimization with regular model optimization. Motivated by the classical approaches for increasing model generalization, we propose our Transfer-based Self-Ensemble Attack (T-SEA), ensembling the input " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ", the attacked model " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": ", and the adversarial patch " + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 303, + 289, + 518 + ], + "type": "text", + "content": " from themselves to boost the adversarial transferability of the attack." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "Specifically, we first introduce an enhanced attack baseline based on [32]. Observing from Fig. 4 that the original training strategies have some limitations, we slightly adjust its learning rate scheduler and training patch scale to revise [32] as our enhanced baseline (E-baseline). Then, as shown in Fig. 1, motivated by input augmentation in model optimization (e.g., training data augmentation), we introduce the constrained data augmentation (data self-ensemble) and model ShakeDrop (model self-ensemble), virtually expanding the inputs of patch optimization (i.e., the input data " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and the attacked model " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ") to increase the transferability of the patch against different data and models. Meanwhile, motivated by the dropout technique in model optimization, which utilizes sub-networks of the optimizing model to overcome overfitting and thus increasing model generalization, we propose patch cutout (patch self-ensemble), ran" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "domly performing cutout on the training patch " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " to overcome overfitting. Through comprehensive experiments, we prove that the proposed E-baseline and self-ensemble strategies perform very well on widely-used detectors with mainstream base attack methods (e.g., PGD [24], MIM [13]). Our contributions can be summarized as the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 152, + 545, + 378 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 152, + 545, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 545, + 189 + ], + "type": "text", + "content": "- We propose a transfer-based black-box attack T-SEA, requiring only one attacked model to achieve a high adversarial transferability attack on object detectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 196, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 196, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 196, + 545, + 232 + ], + "type": "text", + "content": "- Observing the issues of the existing approach, we slightly adjust the training strategies to craft an enhanced baseline and increase its performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 239, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 545, + 310 + ], + "type": "text", + "content": "- Motivated by approaches increasing generalization of deep learning model, we propose a series of strategies to self-ensemble the input data, attacked model, and adversarial patch, which significantly increases the model-level adversarial transferability without introducing extra information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 318, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 318, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 317, + 318, + 545, + 378 + ], + "type": "text", + "content": "- The experimental results demonstrate that the proposed T-SEA can greatly reduce the AP on multiple widely-used detectors on the black-box setting compared to the previous methods, while concurrently performing well with multiple base attack methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 387, + 392, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 392, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 392, + 399 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 407, + 468, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 468, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 468, + 418 + ], + "type": "text", + "content": "2.1. Black-box Adversarial Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 628 + ], + "type": "text", + "content": "Since the discovery of adversarial attack by [31], black-box attacks have gradually attracted more attention, owing to the difficulty of obtaining details of the attacked model in real scenes. Black-box attack can be separated into query-based methods [5, 12, 21] and transfer-based methods [2, 20, 26]. The former utilizes the outputs of the target model to optimize the adversarial examples, which may disclose the attack behavior due to the frequent queries. Hence, in our work, we concentrate on the latter, optimizing adversarial examples in a substitute model without relying on any knowledge of the black-box models. Though the transfer-based attacks ensure the covertness of the attack, their performance is generally limited due to the lack of specialized adjustment catered to the target model. In this work, we propose the self-ensemble strategies to address the issue and greatly boost the black-box attack transferability of the existing transfer-based detector attack." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 635, + 408, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 408, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 408, + 647 + ], + "type": "text", + "content": "2.2. Object Detectors" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Object detection is a fundamental computer vision technology predicting objects' categories and positions and is widely used in many perception tasks. In the past ten years, deep learning-based models have greatly improved the performance of object detectors. The mainstream methods" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20515" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 539, + 234 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 539, + 234 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 539, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 539, + 234 + ], + "type": "image", + "image_path": "2230a575dc7e538de8ed49b5c87171131ee30529dd74d51798e01b9b069e7aee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": "Figure 2. The Overall Pipeline of T-SEA. During the training stage, we utilize the self-ensemble strategies (i.e., the constrained data augmentation, patch cutout and model ShakeDrop) to enhance the transferability of the well-optimized adversarial patch. During the attack (inference) stage, we attach the crafted patch into different images to disrupt the detection process of multiple widely-used detectors on black-box setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 418 + ], + "type": "text", + "content": "based on deep learning can be roughly divided into one-staged and two-staged methods. The former directly predicts the positions and classes of the object instances and are thus faster; the latter first use the region proposal network (RPN) to generate proposals, and then predict the labels and positions of the selected proposals. In this paper, we select eight mainstream detectors of both one-stage and two-stage, including YOLO v2 [27], YOLO v3 & YOLO v3tiny [28], YOLO v4 & YOLO v4tiny [4], YOLO v5 [17], Faster R-CNN [29] and SSD [22] to systemically verify the proposed T-SEA framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 426, + 198, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 198, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 198, + 439 + ], + "type": "text", + "content": "2.3. Attacks on Object Detector" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 444, + 289, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 289, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 289, + 659 + ], + "type": "text", + "content": "The security of deep learning-based models is receiving increasing attention [8, 9], especially with the existence of adversarial attacks. Recent works have explored adversarial robustness of object detectors. To begin, [37] apply adversarial attack on object detectors, performing iterative gradient-based method to misclassify the proposals, and the similar ideas are also carried out by [19, 34]. Ensuingly, for enhancing the attack capability in the physical world, [7, 32] propose real-world adversarial patches to attack the mainstream detectors, e.g., YOLO and Faster R-CNN. Recently, model ensemble approaches are used to improve the attack transferability among multiple detectors, such as [16, 35, 41] simultaneously attacking multiple detectors to generate cross-model adversarial examples. Different from the above, we focus on how to make the most of the limited information (i.e., a single detector) to carry out a high-transferability black-box attack on multiple black-box detectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "In this section, we first give the problem formulation in Sec. 3.1 and describe the overall framework in Sec. 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 288, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 546, + 323 + ], + "type": "text", + "content": "Then, we respectively introduce the details of our enhanced baseline and self-ensemble strategies in Sec. 3.3 and Sec. 3.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 331, + 430, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 430, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 430, + 342 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": "In this work, we carry out a transfer-based black-box attack with only one white-box detector to decrease the average precision (AP) of both white-box and black-box detectors. Given the target input data distribution " + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{X},\\mathcal{F})" + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": ", we regard a single pre-trained detector " + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "inline_equation", + "content": "f_{w}\\in \\mathcal{F}" + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": " as the whitebox attacked model, " + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "inline_equation", + "content": "x_{1,\\dots,N}\\in \\mathcal{X}" + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": " as the input images, where " + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": " is the number of training samples. We are committed to crafting a universal adversarial patch " + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": " from the adversarial distribution " + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 349, + 547, + 458 + ], + "type": "text", + "content": " to disrupt the detection process," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 465, + 546, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 465, + 546, + 499 + ], + "spans": [ + { + "bbox": [ + 344, + 465, + 546, + 499 + ], + "type": "interline_equation", + "content": "\\hat {\\tau} = \\underset {\\tau \\sim \\mathcal {T}} {\\arg \\min } \\sum_ {i} ^ {N} L \\left(\\widetilde {x} _ {i}, \\widetilde {f} _ {w}, \\widetilde {\\tau}\\right) + \\lambda J (\\tau), \\tag {1}", + "image_path": "6d60de0c2641479c4de5466f7e582ce71147a4121f2c5fd0605ccf45a1f78271.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "content": " is the loss function to measure the corruption of detector, " + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_i,\\widetilde{f}_w" + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "inline_equation", + "content": "\\widetilde{\\tau}" + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "content": " denote self-enssembled data, model and patch, respectively, and " + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "inline_equation", + "content": "J(\\cdot)" + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "content": " is a regularization term which we employ total variation of " + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 506, + 547, + 556 + ], + "type": "text", + "content": " in our work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 563, + 422, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 422, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 422, + 575 + ], + "type": "text", + "content": "3.2. Overall Framework" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "As depicted in Fig. 2, we divide the entire T-SEA pipeline into training stage and attack stage. During the training stage, the input images are first augmented with the constrained data augmentation. Ensuingly, the white-box detector (i.e., the attacked detector) is used to locate the target objects from the augmented images. Then, we attach adversarial patches to the center of each detected object of the target class, along with patch cutout for mitigating overfitting. After that, the images with adversarial patches go through the shake-dropped models, and we minimize the object confidence and continuously optimize the training" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20516" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "adversarial patches until reaching maximum epoch number. During the attack stage, we apply our well-optimized adversarial patch on the test images to disrupt the detection process of multiple black-box detectors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 159, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 159, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 159, + 140 + ], + "type": "text", + "content": "3.3. Enhanced Baseline" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": "In this section, we introduce the details of the enhanced baseline, which is described in Algorithm 1. Firstly, we randomly initialize a patch " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": ", prepare the training images and attacked model " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": ". During each training epoch, for every input image batch " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": ", we first obtain their detection results via the white-box detector " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": ". Ensuingly, we utilize the transformation function " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": " to apply the training patch into each image to generate adversarial image batch " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "X^{adv}" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": ". Next, we calculate the object confidence of " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "X^{adv}" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": ", and the attack loss of the image batch. Here we use " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": " loss as the attack loss, unless otherwise stated. Finally, we update the " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": " via base attack method " + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 46, + 148, + 288, + 304 + ], + "type": "text", + "content": " and adjust the learning rate after each epoch until reaching the maximum number." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 48, + 330, + 287, + 535 + ], + "blocks": [ + { + "bbox": [ + 48, + 316, + 239, + 327 + ], + "lines": [ + { + "bbox": [ + 48, + 316, + 239, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 239, + 327 + ], + "type": "text", + "content": "Algorithm 1 Enhanced Baseline Based on [32]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "lines": [ + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "x_{1,\\dots,N}" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (training images), " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (white-box detector), " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (maximum epoch), " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "BS" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (batch size), " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (the initial patch), " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (patch applicer), " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\hbar" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " (base attack method), scheduler (learning rate scheduler). \nEnsure: Well-optimized Adversarial Patch " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\tau \\Leftarrow \\tau_0" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n2: for each " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "i \\in [1, M]" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " do \n3: for each " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "j \\in [1, \\frac{N}{BS}]" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "X \\Leftarrow x_{(j-1)\\cdot BS+1}, \\ldots, x_{j\\cdot BS}" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n5: bboxclean, confclean " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\Leftarrow f_w(X)" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "X^{adv} \\Leftarrow T(X, bbox^{clean}, \\tau)" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n7: bbox" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "^{adv}" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": ", conf" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "^{adv} \\Leftarrow f_w(X^{adv})" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n8: loss " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\Leftarrow Avg(conf^{adv})" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "inline_equation", + "content": "\\tau \\Leftarrow \\hbar(\\tau, loss)" + }, + { + "bbox": [ + 48, + 330, + 287, + 535 + ], + "type": "text", + "content": " \n10: end for \n11: update lr via scheduler \n12: end for" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 550, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 634 + ], + "type": "text", + "content": "Compared to AdvPatch [32], we slightly adjust two training strategies. One is the learning rate scheduler. We observe that a saddle point during patch optimization may lead the original plateau-based scheduler to drastically drop the learning rate, causing inadequate optimization. Hence, we adjust the learning rate decline strategy with the following criterion," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 643, + 294, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 294, + 669 + ], + "type": "interline_equation", + "content": "l r \\Leftarrow l r * \\mu , i f \\left(\\ell_ {t} - \\ell_ {t - 1}\\right) < \\epsilon_ {1} a n d \\frac {\\left(\\ell_ {t} - \\ell_ {t - 1}\\right)}{\\ell_ {t}} < \\epsilon_ {2}, \\tag {2}", + "image_path": "13e28b8bc37cd1578d61fdb327c0b5cda73a3ff8d7579335e135b1052ac7e388.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the decay factor, " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\ell_t" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " denotes the mean loss at epoch " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_1, \\epsilon_2" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " are thresholds to control learning rate updates. In our experiments, we ensure that the learning rate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "decrease more stably via the above hyper-parameters. The other adjustment is to reduce the patch scale at training stage " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". Here, the patch scale is the length ratio between the patch and the bounding box of the specific object. As Fig. 4 shows, a proper scale helps the optimized patch to learn more global patterns, which will not be disrupted easily when the patch is scaled, e.g., a local pattern will lose more information when its corresponding patch is scaled to a very small size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 188, + 444, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 188, + 444, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 444, + 201 + ], + "type": "text", + "content": "3.4. Self-Ensemble Strategies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 206, + 426, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 426, + 218 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 426, + 218 + ], + "type": "text", + "content": "3.4.1 Theoretical Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": "Typical model training usually attempts to find a mapping function " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": " from a finite hypothesis space " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": " that describes the relationship between the input data " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": " and label " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": " following the underlying joint distribution " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{X},\\mathcal{Y})" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": ". Generally, since " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": " is unknown, we use training data set " + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "inline_equation", + "content": "S = \\{(x_i,y_i)|i = 1,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 224, + 545, + 320 + ], + "type": "text", + "content": " to train the model, which means the optimization involves the empirical risk minimization(ERM) [33]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 363, + 330, + 545, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 330, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 363, + 330, + 545, + 362 + ], + "type": "interline_equation", + "content": "\\widehat {R} _ {S} (f) = \\frac {1}{N} \\sum_ {i} ^ {N} L \\left(y _ {i}, f \\left(x _ {i}\\right)\\right), \\tag {3}", + "image_path": "57a65b95bfba50bbcf93f0464aae829bda773a75b2bd964a1cdcf9f624ab13b0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": " is a loss function to measure the difference between the model prediction " + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": " and the label " + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": ". However, the empirical risk is unable to provide generalization on unseen data [40]. Fortunately, we can derive that the generalization error " + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "inline_equation", + "content": "R_{\\mathcal{D}}(f)" + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": " is bounded by " + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\widehat{R}_S(f)" + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": " for a given confidence " + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "inline_equation", + "content": "1 - \\sigma \\in (0,1)" + }, + { + "bbox": [ + 304, + 370, + 545, + 443 + ], + "type": "text", + "content": " [3,25]," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 453, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 453, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 328, + 453, + 545, + 485 + ], + "type": "interline_equation", + "content": "\\forall f \\in \\mathcal {F}, R _ {\\mathcal {D}} (f) \\leq \\widehat {R} _ {S} (f) + \\sqrt {\\frac {\\log c + \\log \\frac {1}{\\sigma}}{N}}, \\tag {4}", + "image_path": "0ba0abdb9de8f79593a4799e863879d870fc5200e51fbd4d2e49d1a6d18c3885.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "text", + "content": " denotes the input data size, " + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "text", + "content": " is a complexity measure of " + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "text", + "content": ", such as VC-dimension [25] or covering numbers [1]. This generally yields bounds of the generalization gap " + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "inline_equation", + "content": "R_{\\mathcal{D}}(f) - \\widehat{R}_S(f) = \\mathcal{O}(\\sqrt{\\frac{\\log c}{N}})" + }, + { + "bbox": [ + 304, + 492, + 545, + 653 + ], + "type": "text", + "content": ", which gives us a lens to analyze model generalization: we can sample more training data or lower model complexity to help close the generalization gap. In practice, limited training data and large-scaled model parameters make the model prone to overfit [39], causing an undesirable generalization gap increase. Fortunately, past studies have widely adopted a series of regularization strategies on both input data and the model, such as data augmentation [11, 40, 42], Stochastic Depth [15], and Dropout [30]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Motivated by the aforementioned model regularization methods, we discuss how to improve attack transferability of the adversarial patch. During patch optimization, the model " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " becomes an input instead of the optimization objective, which means the original input distribution" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20517" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{X}, \\mathcal{Y})" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " is shifted to a joint input distribution " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{X}, \\mathcal{F})" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": ". Naturally, we can improve generalization of the adversarial patch by increasing training data scale " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "N_{x,f}" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": ", i.e. sampling more training images " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " and assembling more white-box models " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": ". However, obtaining massive images or models of the same task is generally expensive or impractical for the attacker. To address this issue, we propose to virtually expand the input by data augmentation and model ShakeDrop to help improve patch transferability. Meanwhile, although the large-scale models have achieved superior performance in most tasks, these models still benefit from reducing model complexity at training stage like stochastic depth [15] and Dropout [30]. Thus we propose the patch cutout to reduce its capacity " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "c_{\\tau}" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " in training process as a new regularization to alleviate overfitting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 247, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 247, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 247, + 278 + ], + "type": "text", + "content": "3.4.2 Data: Constrained Data Augmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "type": "text", + "content": "Motivated by data augmentation in increasing model generalization, we can similarly improve patch transferability by virtually expanding the training set " + }, + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "type": "text", + "content": " to more closely approximate the underlying distribution " + }, + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 285, + 289, + 405 + ], + "type": "text", + "content": ". Therefore, we employ a constrained policy to avoid unnatural augmentation which may not appear in natural scenes. In our work, we 1) mildly resize and crop the input images, 2) slightly alter the brightness, contrast, saturation and hue, and 3) randomly rotate the input images in a small range, to generate natural augmented images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 162, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 162, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 162, + 431 + ], + "type": "text", + "content": "3.4.3 Model: ShakeDrop" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 448, + 160, + 573 + ], + "blocks": [ + { + "bbox": [ + 61, + 448, + 160, + 573 + ], + "lines": [ + { + "bbox": [ + 61, + 448, + 160, + 573 + ], + "spans": [ + { + "bbox": [ + 61, + 448, + 160, + 573 + ], + "type": "image", + "image_path": "734ba65804b6e881505ff51289a1a3a33114e5adf16772fe42f0acbe88435e97.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 575, + 130, + 585 + ], + "lines": [ + { + "bbox": [ + 91, + 575, + 130, + 585 + ], + "spans": [ + { + "bbox": [ + 91, + 575, + 130, + 585 + ], + "type": "text", + "content": "(a) Forward" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": "Figure 3. We utilize the model ShakeDrop to linearly combine the outputs of the stacked layers and identity of a residual block." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 450, + 273, + 573 + ], + "blocks": [ + { + "bbox": [ + 176, + 450, + 273, + 573 + ], + "lines": [ + { + "bbox": [ + 176, + 450, + 273, + 573 + ], + "spans": [ + { + "bbox": [ + 176, + 450, + 273, + 573 + ], + "type": "image", + "image_path": "82d0b5cc5ee8845bec1f99a1cd693c00ab4ab817f51dbb0b8f5181f19a69e5dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 575, + 247, + 585 + ], + "lines": [ + { + "bbox": [ + 202, + 575, + 247, + 585 + ], + "spans": [ + { + "bbox": [ + 202, + 575, + 247, + 585 + ], + "type": "text", + "content": "(b) Backward" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "As discussed above, when training the patch, sampling more white-box models " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " from the joint input distribution " + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": " can improve the adversarial transferability of the crafted patch. However, it is difficult to obtain multiple models of the same task in reality. Inspired by stochastic depth [15], which randomly drops a subset of layers to virtually ensemble multiple model variants for improving generalization," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "we utilize ShakeDrop [38] to linearly combine the outputs of stacked layers and identity of a residual block, generating massive variants of the attacked white-box model as the Fig. 3 illustrated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "content": "Specifically, in forward propagation, we combine the identity " + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "inline_equation", + "content": "z_{l}" + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "content": " and the output of stacked layers " + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "inline_equation", + "content": "F(z_{l})" + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "content": " of a residual block to generate " + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 304, + 120, + 545, + 156 + ], + "type": "text", + "content": " with the following formula:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 167, + 545, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 167, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 353, + 167, + 545, + 180 + ], + "type": "interline_equation", + "content": "z _ {L} = z _ {l} + (\\beta + \\alpha - \\beta \\cdot \\alpha) \\cdot F (z _ {l}), \\tag {5}", + "image_path": "58f79a600a855149784cb746cca35b161773f9300f754c914e461d1ac18f5c5c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": " is sampled from a Bernoulli distribution with " + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "inline_equation", + "content": "P(\\beta = 1) = \\varphi_s" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": " is sampled from a continuous uniform distribution " + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\alpha \\sim U(1 - e, 1 + e)" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": " is a constant). In backward propagation, denoting the loss function as " + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 190, + 545, + 262 + ], + "type": "text", + "content": ", from the chain rule of backpropagation, ShakeDrop can be formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 267, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 267, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 334, + 267, + 545, + 319 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\psi \\left(\\frac {\\partial \\mathcal {E}}{\\partial z _ {l}}\\right) = \\psi \\left(\\frac {\\partial \\mathcal {E}}{\\partial z _ {L}} \\cdot \\frac {\\partial z _ {L}}{\\partial z _ {l}}\\right) \\tag {6} \\\\ = \\frac {\\partial \\mathcal {E}}{\\partial z _ {L}} (1 + (\\gamma + \\alpha - \\gamma \\cdot \\alpha) \\cdot \\frac {\\partial F}{\\partial z _ {l}}), \\\\ \\end{array}", + "image_path": "6f0fef4c8813dc7ab5063abdefad4c695194d3a77aa2d4d61d7e3f898172ac0d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "type": "text", + "content": " is another Bernoulli variable and identically distributed with " + }, + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 327, + 545, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 365, + 400, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 400, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 400, + 376 + ], + "type": "text", + "content": "3.4.4 Patch: Cutout" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 383, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 514 + ], + "type": "text", + "content": "Dropout [30] is designed to randomly drop neural units along with their connections, effectively preventing overfitting of the model. Inspired by Dropout, we propose a patch cutout strategy, randomly masking a region of the adversarial patch to reduce the patch complexity at training stage and thus to prevent overfitting on specific model and data. From an other angle, model Dropout prevents excessive co-adapting of neural units and spread out features over multi neurons to alleviate overfitting. Similarly, patch cutout also prevents the adversarial nature of patch from relying too much on the patterns of a certain area." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": "The proposed patch cutout is similar to cutout [11] or random erasing [42] in model training. Specifically, for a normalized input image " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "I_{x,y} \\in [0,1]" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": ", we carry out the following process with a probability " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\varphi_c" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": " before attaching patches to the target objects: 1) we firstly conduct random sampling to obtain one point " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "p = (x_0, y_0)" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": " within the given patch; 2) then we cover a " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\eta H \\times \\eta W" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": " square area " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "(\\eta" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": " is the ratio) centered at " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": " with a specific value " + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1]" + }, + { + "bbox": [ + 304, + 515, + 545, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 621, + 382, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 382, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 382, + 635 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "In this section, we first introduce the implementation details in Sec. 4.1. Then, we present the main results of T-SEA in Sec. 4.2, reporting its performance on different detectors and attack methods. After that, we compare the proposed method with other state-of-the-art detection attack algorithms in Sec. 4.3 and perform the ablation study" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20518" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 537, + 308 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 537, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 537, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 537, + 308 + ], + "type": "table", + "html": "
MethodsYOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSDBlack-Box Avg↓
AdvPatch5.6640.2618.0748.4924.4443.3839.2741.2836.46
T-SEA(ours)1.734.482.415.687.756.9116.3820.55\\(9.16^{27.30}\\)↓
AdvPatch51.8513.8951.1757.1658.4370.4751.4659.7957.19
T-SEA(ours)31.025.7635.3842.3725.358.0237.6252.26\\(40.28^{16.91}\\)↓
AdvPatch56.0266.121.3969.6451.5672.1656.6160.7361.83
T-SEA(ours)38.8547.130.5152.7531.0161.1849.1255.44\\(47.93^{13.90}\\)↓
AdvPatch37.4137.1817.5819.6726.9146.3744.6743.0736.17
T-SEA(ours)9.346.469.494.2216.6511.6616.2429.43\\(14.18^{21.99}\\)↓
AdvPatch47.4159.5937.1466.4814.5069.5155.9555.2255.90
T-SEA(ours)37.6944.9815.8151.144.1151.3746.0649.39\\(42.35^{13.55}\\)↓
AdvPatch46.954.9329.2062.1646.1513.3947.7150.7348.25
T-SEA(ours)11.6116.7710.7332.5317.551.3717.1430.78\\(19.59^{28.66}\\)↓
AdvPatch24.5323.3714.5826.5425.630.468.6242.1526.75
T-SEA(ours)9.284.083.998.5513.5810.453.0826.46\\(10.91^{15.84}\\)↓
AdvPatch32.4362.9352.0366.2249.0751.1947.4215.1051.61
T-SEA(ours)12.0630.909.7327.339.0817.3828.545.13\\(19.29^{32.32}\\)↓
", + "image_path": "d7c54e807d8d83e013c538091f2669d7a5b471f43e7b52e16b93ebc874991239.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 356, + 264, + 510 + ], + "blocks": [ + { + "bbox": [ + 46, + 318, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 318, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 547, + 352 + ], + "type": "text", + "content": "Table 1. Comparisons between T-SEA and AdvPatch [32]. We attack each detector separately and use the remaining seven detectors as the black-box detectors. The proposed T-SEA performs much better than [32] on both white-box setting and black-box setting, demonstrating the effectiveness of the proposed self-ensemble strategies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 356, + 264, + 510 + ], + "lines": [ + { + "bbox": [ + 70, + 356, + 264, + 510 + ], + "spans": [ + { + "bbox": [ + 70, + 356, + 264, + 510 + ], + "type": "table", + "html": "
MethodWhite Box ↓Black-Box Avg↓
AdamAdvPatch13.3948.25
T-SEA(ours)1.3719.59
SGDAdvPatch20.3955.97
T-SEA(ours)1.6615.85
MIMAdvPatch11.9133.51
T-SEA(ours)1.4315.43
BIMAdvPatch8.4732.93
T-SEA(ours)1.6219.15
PGDAdvPatch13.5838.70
T-SEA(ours)1.6018.06
", + "image_path": "a7fd64c97a9da6e10fcb7f1832bfaa3dfc2d532caa01b99a796f5857b4de7f7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 288, + 587 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 587 + ], + "type": "text", + "content": "Table 2. Comparisons of T-SEA and AdvPatch with Different Base Attack Methods on YOLO v5 (white-box). We select five classical base attack methods, including optimization-based methods (Adam and SGD) and iterative methods (MIM, BIM and PGD). The results show that T-SEA can enhance the performance of all these methods and performs much better than AdvPatch." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 599, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 624 + ], + "type": "text", + "content": "in Sec. 4.4. Finally, we show the transferability of T-SEA across the datasets and scenes in Sec. 4.5 and Sec. 4.6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 180, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 180, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 180, + 647 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Datasets In our experiments, we utilize the INRIA person dataset [10] to train and test our adversarial patch, whose training set and test set consist of 614 and 288 images, respectively. Meanwhile, to verify the transferability of the crafted patch, we select images containing per" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 358, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 430 + ], + "type": "text", + "content": "son from COCO validation set (named COCO-person) and CCTV Footage of Humans* (named CCTV-person) as additional test data. The former contains 1684 human images in different scenes (e.g., sports playground, transportation routes, oceans and forests) and the latter contains 559 in-person images from camera footage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": "Optimization Details We use INRIA train set as the training set and regard person as the target attack class. The patch size is " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "300 \\times 300" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", the input image size is " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "416 \\times 416" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", the batch size " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "BS = 8" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", and the maximum epoch number " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "M = 1000" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ". For the E-baseline, we adjust the training patch scale from 0.2 in AdvPatch [32] to 0.15, set " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\epsilon_{1} = 1e - 4" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\epsilon_{2} = 1e - 4" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": " for the learning rate scheduler, and adopt the Adam [18] as the optimizer. For constrained data augmentation, we carry out constrained data augmentation via horizontal flip, slight color jitter, random resized crop, and random rotation; for model ShakeDrop, the constant " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": " is set to 1 so that " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\alpha \\sim U(0,2)" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", and we perturb the model with the probability of " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\varphi_{s} = 0.5" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": "; for patch cutout, we set the fill value of the erased area " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "k = 0.5" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", the ratio " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\eta = 0.4" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": ", and the probability " + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "inline_equation", + "content": "\\varphi_{c} = 0.9" + }, + { + "bbox": [ + 304, + 433, + 546, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 615, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 675 + ], + "type": "text", + "content": "Evaluation Metric Following [14, 32], we use the Average Precision (AP) to measure the attack capability of the crafted patch (the lower AP, the better attack) and regard the detector's predictions of the clean data as the ground truth (i.e., " + }, + { + "bbox": [ + 304, + 615, + 545, + 675 + ], + "type": "inline_equation", + "content": "AP = 1" + }, + { + "bbox": [ + 304, + 615, + 545, + 675 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": "\\*https://www.kaggle.com/datasets/constantinwerner/human-detectiondataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20519" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 535, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 535, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 535, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 535, + 201 + ], + "type": "table", + "html": "
MethodWhite Box ↓Black Box ↓Black-Box Avg↓
YOLO v2YOLO v3YOLO v3tinyYOLO v4YOLO v4tinyYOLO v5Faster R-CNNSSD
Gray67.7576.2280.6975.2276.8981.8661.7572.05-
Random Noise70.6775.882.4475.178.7481.7963.4172.9-
White68.5274.8980.274.7376.0980.0960.3569.41-
NPAP [14]38.0356.8558.0467.7467.4366.8556.656.6661.45
AdvCloak [36]33.7454.7753.4267.5756.1268.0555.1960.8259.42
AdvPatch [32]5.6640.2618.0748.4924.4443.3839.2741.2836.46
E-baseline(ours)3.6115.325.5018.5813.399.7729.7323.8216.59
T-SEA(ours)1.734.482.415.687.756.9116.3820.559.16
", + "image_path": "609d701fa2a583fb4e18b238c9528d2a85c30af42b56b27a14e72f80dedf06fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 240, + 132, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 132, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 132, + 251 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 259, + 240, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 240, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 240, + 270 + ], + "type": "text", + "content": "4.2.1 Results on Different Attacked Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 279, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 410 + ], + "type": "text", + "content": "We systematically investigate attack performance of the proposed T-SEA on eight widely-used object detectors, and report the quantitative results in Tab. 1. Since T-SEA is improved based on AdvPatch [32], we also report the results of AdvPatch. The proposed T-SEA achieves significant improvements on both white-box and black-box performance of all eight detectors compared to [32], demonstrating the effectiveness of the proposed strategies. Meanwhile, for some white-box detectors (e.g., YOLO v2), the black-box average AP can drop to around 10, exhibiting effective black-box performance of T-SEA." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 423, + 234, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 234, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 234, + 434 + ], + "type": "text", + "content": "4.2.2 Results of Different Attack Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 439, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 594 + ], + "type": "text", + "content": "T-SEA is not designed only for a specific base attack method; it is important for T-SEA to perform well on different base attack methods. In Tab. 2, we give the detailed results of T-SEA with different base attack methods on YOLO v5, which includes the optimization-based methods (e.g., Adam and SGD) and iterative methods (e.g., MIM, BIM and PGD). The results show that the proposed T-SEA can improve both white-box and black-box attack performance on all methods above, demonstrating that the performance gain caused by T-SEA is not limited to a specific method. That is to say, T-SEA may potentially work well with future base attack methods to further increase the transferability of their crafted AdvPatch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 601, + 226, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 226, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 226, + 613 + ], + "type": "text", + "content": "4.3. Comparison with SOTA Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "In this section, we compare T-SEA with the SOTA detection attack approaches, which all regard the YOLO v2 as the white-box model and evaluate on seven black-box detectors, and we also report the results of gray/random noise/white patches as control group. As reported on Tab. 3, 1) all adversarial patch performs much better than the control group; 2) our E-baseline and T-SEA achieve the highest performance among all adversarial approaches with the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 239, + 535, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": "Table 3. Comparisons with Existing Detection Attack Methods. For clearer controlled observations, we list the results of gray, random noise, and white patch. Compared with existing methods, T-SEA achieves the best performance on both white-box and black-box attack." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 239, + 535, + 323 + ], + "lines": [ + { + "bbox": [ + 315, + 239, + 535, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 535, + 323 + ], + "type": "table", + "html": "
White Box ↓Black Box ↓
E-baseline13.3948.25
+ Constrained Data Aug.18.4742.42
+ Model ShakeDrop12.5439.40
+ Patch cutout6.8032.54
+ Combined Strategies1.3719.59
", + "image_path": "5cd5bbb7c05270c902bbd9595a66340692ad2f5f0f7445b035cde6801a394e23.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 329, + 546, + 406 + ], + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 406 + ], + "type": "text", + "content": "Table 4. Ablation Study of Self-Ensemble Strategies on YOLO v5 (white-box). Each self-ensemble strategy is added to the E-baseline to verify its individual performance gain. The results show that: 1) though using data augmentation alone degrades the white-box performance, all strategies can obviously improve the black-box performance of E-baseline; 2) combining all achieves non-trivial results on both white-box and black-box setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 486 + ], + "type": "text", + "content": "same inference setting (i.e., we ensure the perturbed area is same for each method), showing that compared to the existing single model attack method, the training strategies adjustment and self-ensemble strategies can effectively enhance the attack capability on both white-box and black-box of the crafted adversarial patch." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 495, + 400, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 400, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 400, + 508 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 515, + 473, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 473, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 473, + 528 + ], + "type": "text", + "content": "4.4.1 Training Strategies Adjustment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 643 + ], + "type": "text", + "content": "Here we explore the performance gain from the training strategies adjustments. As discussed in the Sec. 3.3, we modify the learning rate scheduler and training patch scale of [32]. As illustrated in Fig. 4, the modified training scale can reduce the AP in test set more quickly and effectively, and the crafted patch has clearer adversarial patterns. The adjusted scheduler can also lead the detection loss to decrease more, and can be further combined with the modified training scale to achieve a better result." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 660, + 444, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 444, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 444, + 672 + ], + "type": "text", + "content": "4.4.2 Self-Ensemble Strategies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 714 + ], + "type": "text", + "content": "In this work, we are motivated by model training approaches that enhance generalization ability, and we propose self-ensemble strategies on the input data, the attacked" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20520" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 182, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 182, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 182, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 182, + 190 + ], + "type": "image", + "image_path": "96cd91f62347db130d1f0a83095630b1e06fd01cb3fb76029f5f0b720d6eb67c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "type": "text", + "content": "Figure 4. The improvements after adjusting learning rate scheduler (tagged as " + }, + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{PlateauLR}^{\\dagger}" + }, + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "type": "text", + "content": ") and patch scale at training stage. A small patch scale during training will cause the test AP to drop lower and faster, and the adjusted " + }, + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{PlateauLR}^{\\dagger}" + }, + { + "bbox": [ + 46, + 198, + 287, + 251 + ], + "type": "text", + "content": " can also cause the detection loss to further decrease." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 72, + 227, + 110 + ], + "blocks": [ + { + "bbox": [ + 189, + 72, + 227, + 110 + ], + "lines": [ + { + "bbox": [ + 189, + 72, + 227, + 110 + ], + "spans": [ + { + "bbox": [ + 189, + 72, + 227, + 110 + ], + "type": "image", + "image_path": "afbb33ffe759763f1f868375778744d1920f4fc4cdac8c02e30fad05670f12b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 110, + 222, + 131 + ], + "lines": [ + { + "bbox": [ + 194, + 110, + 222, + 131 + ], + "spans": [ + { + "bbox": [ + 194, + 110, + 222, + 131 + ], + "type": "text", + "content": "PlateauLR Scale=0.2 mAP:38.89" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 272, + 110 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 272, + 110 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 272, + 110 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 272, + 110 + ], + "type": "image", + "image_path": "0f769d13b97809adac618432521c48bdedd51f992f13bcfdd9df84fd582b7c19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 110, + 270, + 131 + ], + "lines": [ + { + "bbox": [ + 238, + 110, + 270, + 131 + ], + "spans": [ + { + "bbox": [ + 238, + 110, + 270, + 131 + ], + "type": "text", + "content": "PlateauLR Scale=0.15 mAP:36.63" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 189, + 131, + 227, + 168 + ], + "blocks": [ + { + "bbox": [ + 189, + 131, + 227, + 168 + ], + "lines": [ + { + "bbox": [ + 189, + 131, + 227, + 168 + ], + "spans": [ + { + "bbox": [ + 189, + 131, + 227, + 168 + ], + "type": "image", + "image_path": "e1dfc0af9db60426bbe2b98ff55d21bc49cf30025764186e336bbf7ffc9cba3c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 169, + 223, + 189 + ], + "lines": [ + { + "bbox": [ + 194, + 169, + 223, + 189 + ], + "spans": [ + { + "bbox": [ + 194, + 169, + 223, + 189 + ], + "type": "text", + "content": "PlateauLR+ Scale=0.2 mAP:35.38" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 237, + 132, + 274, + 168 + ], + "blocks": [ + { + "bbox": [ + 237, + 132, + 274, + 168 + ], + "lines": [ + { + "bbox": [ + 237, + 132, + 274, + 168 + ], + "spans": [ + { + "bbox": [ + 237, + 132, + 274, + 168 + ], + "type": "image", + "image_path": "7faa8157c411a78bf38658b62f1d78fa73a19b54f34ab6161b126a9e9a3a6391.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 169, + 270, + 189 + ], + "lines": [ + { + "bbox": [ + 238, + 169, + 270, + 189 + ], + "spans": [ + { + "bbox": [ + 238, + 169, + 270, + 189 + ], + "type": "text", + "content": "PlateauLR+ Scale=0.15 mAP:24.77" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 48, + 269, + 293, + 346 + ], + "blocks": [ + { + "bbox": [ + 48, + 269, + 293, + 346 + ], + "lines": [ + { + "bbox": [ + 48, + 269, + 293, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 293, + 346 + ], + "type": "table", + "html": "
DatasetMethodWhite Box ↓Black Box ↓
COCO-personAdvPatch [32]45.8352.54
T-SEA37.2838.87
CCTV-personAdvPatch [32]38.0734.08
T-SEA38.7119.91
", + "image_path": "9e4b399191c85f46e625fd388e8cc9dee99004fba161e8a0336150d535622f9c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 354, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 399 + ], + "type": "text", + "content": "Table 5. The Cross Datasets Transferability of T-SEA on YOLO v5 (white-box). Compared to the [32], the adversarial patch crafted by T-SEA has much stronger attack adaptability on person images from different datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 419, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 538 + ], + "type": "text", + "content": "model, and the training patch, greatly augmenting adversarial transferability from themselves. Here we individually inspect the proposed strategies on our E-baseline (YOLO v5 as white-box detector) to explore the impact of each strategy. As reported in Tab. 4, though the constrained data augmentation will slightly decrease the white-box performance (other two strategies both improve the white-box performance), all these self-ensemble strategies increase the black-box results separately. and combining them can achieve the best performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 561, + 192, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 192, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 192, + 573 + ], + "type": "text", + "content": "4.5. Cross-Dataset Verification" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "The capability of performing cross-dataset attack is also significant for a well-optimized patch. Here, we apply the INRIA-trained patch on two different datasets, COCO-person and CCTV-person. Compared to INRIA, COCO-person has much more person images on different scenes, while CCTV-person focus on persons from the security camera. As shown in Tab. 5, the patch crafted by T-SEA achieves comparable results on the white-box setting with the AdvPatch, but obtains a much better black-box attack capability on both COCO-person and CCTV-person, indicating its strong black-box cross-data attacking capability." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 337, + 69, + 514, + 194 + ], + "blocks": [ + { + "bbox": [ + 337, + 69, + 514, + 194 + ], + "lines": [ + { + "bbox": [ + 337, + 69, + 514, + 194 + ], + "spans": [ + { + "bbox": [ + 337, + 69, + 514, + 194 + ], + "type": "image", + "image_path": "4c59c3f8efada64ba0dfe99e073bd1deb52862c7a70f3578d1bebf15c0a5efa0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 202, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 225 + ], + "type": "text", + "content": "Figure 5. Physical Attack Demo of T-SEA. After showing the optimized patch on iPad, the YOLO v5 and SSD can not detect person." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 231, + 424, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 424, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 424, + 243 + ], + "type": "text", + "content": "4.6. Physical Verification" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 247, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 330 + ], + "type": "text", + "content": "Although applying the patch attack in the physical world is not the main goal of our work, we believe the optimized patch that has high transferability may perform well in physical settings. Here we show a simple case in Fig. 5 that the patch shown on a iPad can successfully attack the human detection of YOLO v5 and SSD, without disrupting the detection process of other objects." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 341, + 388, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 388, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 388, + 353 + ], + "type": "text", + "content": "5. Future Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 355, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 451 + ], + "type": "text", + "content": "Since most existing mainstream detectors are CNN-based, we focus on applying the self-ensemble strategies on these detectors in this work. However, the transformer-based detectors [6, 23] are achieving promising results, and the proposed model ShakeDrop can not directly applied on transformers (the others can), we will follow the motivation of designing the model ShakeDrop, to propose new approach and generate variants of transformer-based detector." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 462, + 378, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 378, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 378, + 474 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 479, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 646 + ], + "type": "text", + "content": "In this paper, we propose a novel transfer-based self-ensemble black-box attack on object detectors, achieving stable and excellent performance gains with various base attack methods on multiple popular object detectors. Firstly, with only slight training strategy adjustments, we improve the existing method's performance and regard it as our enhanced baseline. Then, based on this baseline, we propose a series of self-ensemble strategies to augment the input data, the attacked model, and the training patch from itself to significantly enhance the adversarial transferability of the optimized patch on black-box detectors. The comprehensive experimental results reveal the potential risk that an attacker with only one model could still succeed in a high transferability black-box attack." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 651, + 421, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 421, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 421, + 664 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by National Key R&D Program of China (Grant No. 2022ZD0160305) and National Key Laboratory Open Fund of China (Grant No. 6142113210204)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20521" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Martin Anthony and Peter Bartlett. Neural network learning: Theoretical foundations. Cambridge university press, 1999. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Song Bai, Yingwei Li, Yuyin Zhou, Qizhu Li, and Philip HS Torr. Adversarial metric attack and defense for person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2119-2126, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 212 + ], + "type": "text", + "content": "[3] Mikhail Belkin, Siyuan Ma, and Soumik Mandal. To understand deep learning we need to understand kernel learning. In International Conference on Machine Learning, pages 541-549. PMLR, 2018. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 247 + ], + "type": "text", + "content": "[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "text", + "content": "[5] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-based adversarial attacks: Reliable attacks against black-box machine learning models. arXiv preprint arXiv:1712.04248, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "text", + "content": "[7] Shang-Tse Chen, Cory Cornelius, Jason Martin, and Duen Horng Polo Chau. Shapeshifter: Robust physical adversarial attack on faster r-cnn object detector. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 52-68. Springer, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "text", + "content": "[8] Zhaoyu Chen, Bo Li, Shuang Wu, Jianghe Xu, Shouhong Ding, and Wenqiang Zhang. Shape matters: deformable patch attack. In European Conference on Computer Vision, pages 529-548. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 491 + ], + "type": "text", + "content": "[9] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[10] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[11] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[12] Yinpeng Dong, Qi-An Fu, Xiao Yang, Tianyu Pang, Hang Su, Zihao Xiao, and Jun Zhu. Benchmarking adversarial robustness on image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 321-331, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[13] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[14] Yu-Chih-Tuan Hu, Bo-Han Kung, Daniel Stanley Tan, Jun-Cheng Chen, Kai-Lung Hua, and Wen-Huang Cheng. Naturalistic physical adversarial patch for object detectors. In" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[15] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pages 646-661. Springer, 2016. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[16] Hao Huang, Yongtao Wang, Zhaoyu Chen, Zhi Tang, Wenqiang Zhang, and Kai-Kuang Ma. Rpattack: Refined patch attack on general object detectors. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 284 + ], + "type": "text", + "content": "[17] Glenn Jocher, Alex Stoken, Jirka Borovec, NanoCode012, ChristopherSTAN, Liu Changyu, Laughing, tkianai, Adam Hogan, lorenzomammana, yxNONG, AlexWang1900, Laurentiu Diaconu, Marc, wanghaoyang0106, ml5ah, Doug, Francisco Ingham, Frederik, Guilhien, Hatovix, Jake Poznanski, Jiacong Fang, Lijun Yu, changyu98, Mingyu Wang, Naman Gupta, Osama Akhtar, PetrDvoracek, and Prashant Rai. ultralytics/yolov5, Oct. 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 318 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 363 + ], + "type": "text", + "content": "[19] Yuezun Li, Daniel Tian, Ming-Ching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. arXiv preprint arXiv:1809.05962, 2018. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[20] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "type": "text", + "content": "[21] Siao Liu, Zhaoyu Chen, Wei Li, Jiwei Zhu, Jiafeng Wang, Wenqiang Zhang, and Zhongxue Gan. Efficient universal shuffle attack for visual object tracking. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2739-2743. IEEE, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 521 + ], + "type": "text", + "content": "[22] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 579, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 622 + ], + "type": "text", + "content": "[24] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "text", + "content": "[25] Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine learning. MIT press, 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pages 506-519, 2017. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20522" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 708 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[27] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "type": "text", + "content": "[28] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 130, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 287, + 173 + ], + "type": "text", + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "type": "text", + "content": "[30] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 219, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 272 + ], + "type": "text", + "content": "[31] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 274, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 287, + 328 + ], + "type": "text", + "content": "[32] Simen Thys, Wiebe Van Ranst, and Toon Goedemé. Fooling automated surveillance cameras: adversarial patches to attack person detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 330, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 330, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 330, + 286, + 351 + ], + "type": "text", + "content": "[33] Vladimir Vapnik. The nature of statistical learning theory. Springer science & business media, 1999. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 353, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 287, + 406 + ], + "type": "text", + "content": "[34] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 407, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 407, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 49, + 407, + 287, + 440 + ], + "type": "text", + "content": "[35] Shudeng Wu, Tao Dai, and Shu-Tao Xia. Dpattack: Diffused patch attacks against universal object detection. arXiv preprint arXiv:2010.11679, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 441, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 441, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 441, + 287, + 485 + ], + "type": "text", + "content": "[36] Zuxuan Wu, Ser-Nam Lim, Larry S Davis, and Tom Goldstein. Making an invisibility cloak: Real world adversarial attacks on object detectors. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 487, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 287, + 540 + ], + "type": "text", + "content": "[37] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In Proceedings of the IEEE international conference on computer vision, pages 1369-1378, 2017. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 541, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 541, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 541, + 286, + 563 + ], + "type": "text", + "content": "[38] Yoshihiro Yamada, Masakazu Iwamura, and Koichi Kise. Shakedrop regularization. 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 564, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 564, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 49, + 564, + 287, + 596 + ], + "type": "text", + "content": "[39] Xue Ying. An overview of overfitting and its solutions. In Journal of physics: Conference series, volume 1168, page 022022. IOP Publishing, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 597, + 287, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 597, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 49, + 597, + 287, + 629 + ], + "type": "text", + "content": "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 632, + 287, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 632, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 49, + 632, + 287, + 662 + ], + "type": "text", + "content": "[41] Yusheng Zhao, Huanqian Yan, and Xingxing Wei. Object hider: Adversarial patch attack against object detectors. arXiv preprint arXiv:2010.14974, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 665, + 287, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 665, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 49, + 665, + 287, + 708 + ], + "type": "text", + "content": "[42] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 4, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20523" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_content_list.json b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9365bab06abb019115ed9690abd310356198662 --- /dev/null +++ b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_content_list.json @@ -0,0 +1,2463 @@ +[ + { + "type": "text", + "text": "TAPS3D: Text-Guided 3D Textured Shape Generation from Pseudo Supervision", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiacheng Wei $^{1*}$ Hao Wang $^{1*}$ Jiashi Feng $^{2}$ Guosheng Lin $^{1\\dagger}$ Kim-Hui Yap $^{1}$ $^{1}$ Nanyang Technological University, Singapore $^{2}$ ByteDance \n{jiacheng002@e., hao005@e., gslin@, ekhyap@}ntu.edu.sg, jshfeng@bytedance.com", + "bbox": [ + 119, + 178, + 849, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe5a857b9162dd7165cba42c3288eb459803ae27b0cd1d10664bb92d25f41204.jpg", + "image_caption": [ + "\"a red hatchback car\"" + ], + "image_footnote": [], + "bbox": [ + 88, + 257, + 290, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21f46766d115262baf96255299c52ae29b897dfbe28841a78771350ea37f33f5.jpg", + "image_caption": [ + "Figure 1. Our proposed model is capable of generating detailed textured shapes based on given text prompts. We demonstrate the effectiveness of our approach by showcasing generated results for four distinct object classes: Car, Chair, Table, and Motorbike. Both the textured meshes and geometries are presented, with visualizations rendered using ChimeraX [10]." + ], + "image_footnote": [], + "bbox": [ + 302, + 261, + 364, + 334 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d52a4bf3050eede9191cd8ed4d3f1bc0081fb8233fb32bff1575993e28c2dfdc.jpg", + "image_caption": [ + "\"agrayarmlesschair" + ], + "image_footnote": [], + "bbox": [ + 375, + 258, + 475, + 334 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fbc3400b84d242098e2dd16ac61d1a9b2fac285a4f836a490a1670cc65e91b8.jpg", + "image_caption": [ + "\"awooden office table\"" + ], + "image_footnote": [], + "bbox": [ + 491, + 273, + 588, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/665635970e9f3d95c6c1be52840257be7b7a2c58394773e5225ea36c6c1fa4eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 258, + 687, + 334 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5b238b9babd634c15d7251ce285259ebc330cdf936715f53b1a4c24a60f8f11.jpg", + "image_caption": [ + "\"ablock motorbike\"" + ], + "image_footnote": [], + "bbox": [ + 694, + 268, + 779, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e052b58908dfaf5125946af3eb2a77b3c359ec9bf9276b50f80eef9c281e2230.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 258, + 880, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 424, + 313, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we investigate an open research task of generating controllable 3D textured shapes from the given textual descriptions. Previous works either require ground truth caption labeling or extensive optimization time. To resolve these issues, we present a novel framework, TAPS3D, to train a text-guided 3D shape generator with pseudo captions. Specifically, based on rendered 2D images, we retrieve relevant words from the CLIP vocabulary and construct pseudo captions using templates. Our constructed captions provide high-level semantic supervision for generated 3D shapes. Further, in order to produce fine-grained textures and increase geometry diversity, we propose to adopt low-level image regularization to enable fake-rendered images to align with the real ones. During the inference phase, our proposed model can generate 3D textured shapes from the given text without any additional optimization. We conduct extensive experiments to analyze each of our proposed components and show the efficacy of our framework in generating high-fidelity 3D textured and text-relevant shapes. Code is available at https://github.com/plusmultiply/TAPS3D", + "bbox": [ + 75, + 457, + 473, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 805, + 209, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D objects are essential in various applications [21-24], such as video games, film special effects, and virtual real", + "bbox": [ + 76, + 830, + 468, + 862 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ity. However, realistic and detailed 3D object models are usually hand-crafted by well-trained artists and engineers slowly and tediously. To expedite this process, many research works [3,9,13,34,48,49] use deep generative models to achieve automatic 3D object generation. However, these models are primarily unconditioned, which can hardly generate objects as humans will.", + "bbox": [ + 500, + 425, + 890, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to control the generated 3D objects from text, prior text-to-3D generation works [43, 44] leverage the pretrained vision-language alignment model CLIP [37], such that they can only use 3D shape data to achieve zero-shot learning. For example, Dream Fields [13] combines the advantages of CLIP and NeRF [27], which can produce both 3D representations and renderings. However, Dream Fields costs about 70 minutes on 8 TPU cores to produce a single result. This means the optimization time during the inference phase is too slow to use in practice. Later on, GET3D [9] is proposed with faster inference time, which incorporates StyleGAN [15] and Deep Marching Tetrahedral (DMTet) [46] as the texture and geometry generators respectively. Since GET3D adopts a pretrained model to do text-guided synthesis, they can finish optimization in less time than Dream Fields. But the requirement of test-time optimization still limits its application scenarios. CLIP-NeRF [50] utilizes conditional radiance fields [45] to avoid test-time optimization, but it requires ground truth text data for the training purpose. Therefore, CLIP-NeRF is only applicable to a few object classes that have labeled text data for training, and its generation quality is restricted by the NeRF capacity.", + "bbox": [ + 496, + 534, + 892, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned limitations, we propose", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Work done during an internship at Bytedance.", + "bbox": [ + 99, + 875, + 457, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 99, + 887, + 228, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to generate pseudo captions for 3D shape data based on their rendered 2D images and construct a large amount of $\\langle 3\\mathrm{D}$ shape, pseudo captions $\\rangle$ as training data, such that the text-guided 3D generation model can be trained over them. To this end, we propose a novel framework for Text-guided 3D textured shApe generation from Pseudo Supervision (TAPS3D), in which we can generate high-quality 3D shapes without requiring annotated text training data or test-time optimization.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our proposed framework is composed of two modules, where the first generates pseudo captions for 3D shapes and feeds them into a 3D generator to conduct text-guided training within the second module. In the pseudo caption generation module, we follow the language-free text-to-image learning scheme [20, 54]. We first adopt the CLIP model to retrieve relevant words from given rendered images. Then we construct multiple candidate sentences based on the retrieved words and pick sentences having the highest CLIP similarity scores with the given images. The selected sentences are used as our pseudo captions for each 3D shape sample.", + "bbox": [ + 75, + 228, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the notable progress of text-to-image generation models [29, 38, 39, 42, 52], we use text-conditioned GAN architecture in the text-guided 3D generator training part. We adopt the pretrained GET3D [9] model as our backbone network since it has been demonstrated to generate high-fidelity 3D textured shapes across various object classes. We input the pseudo captions as the generator conditions and supervise the training process with high-level CLIP supervision in an attempt to control the generated 3D shapes. Moreover, we introduce a low-level image regularization loss to produce fine-grained textures and increase geometry diversity. We empirically train the mapping networks only of a pretrained GET3D model so that the training is stable and fast, and also, the generation quality of the pretrained model can be preserved.", + "bbox": [ + 75, + 411, + 472, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed model TAPS3D can produce high-quality 3D textured shapes with strong text control as shown in Fig. 1, without any per prompt test-time optimization. Our contribution can be summarized as:", + "bbox": [ + 75, + 638, + 472, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new 3D textured shape generative framework, which can generate high-quality and fidelity 3D shapes without requiring paired text and 3D shape training data.", + "- We propose a simple pseudo caption generation method that enables text-conditioned 3D generator training, such that the model can generate text-controlled 3D textured shapes without test time optimization, and significantly reduce the time cost.", + "- We introduce a low-level image regularization loss on top of the high-level CLIP loss in an attempt to produce fine-grained textures and increase geometry diversity." + ], + "bbox": [ + 94, + 707, + 468, + 907 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-Guided 3D Shape Generation", + "text_level": 1, + "bbox": [ + 500, + 114, + 802, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-guided 3D shape generation aims to generate 3D shapes from the textual descriptions so that the generation process can be controlled. There are mainly two categories of methods, i.e., fully-supervised and optimization-based methods. The fully-supervised method [6,8,25] uses ground truth text and the paired 3D objects with explicit 3D representations as training data. Specifically, CLIP-Forge [43] uses a two-stage training scheme, which consists of shape autoencoder training, and conditional normalizing flow training. VQ-VAE [44] performs zero-shot training with 3D voxel data by utilizing the pretrained CLIP model [37].", + "bbox": [ + 496, + 138, + 893, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the optimization-based methods [13, 18, 26, 34], Neural Radiance Fields (NeRF) are usually adopted as the 3D generator. To generate 3D shapes for each input text prompt, they use the CLIP model to supervise the semantic alignment between rendered images and text prompts. Since NeRF suffers from extensive generation time, 3D-aware image synthesis [3, 3, 4, 11, 30, 31] has become popular, which generates multi-view consistent images by integrating neural rendering in the Generative Adversarial Networks (GANs). Specifically, there are no explicit 3D shapes generated during the process, while the 3D shapes can be extracted from the implicit representations, such as the occupancy field or signed distance function (SDF), using the marching cube algorithm. These optimization-based methods provide a solution to generate 3D shapes, but their generation speed was compromised. Although [50] is free from test-time optimization, it requires text data for training, which limits its applicable 3D object classes. Our proposed method attempts to alleviate both the paired text data shortage issues and the long optimization time in the previous work. We finally produce high-quality 3D shapes to bring our method to practical applications.", + "bbox": [ + 496, + 321, + 893, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text-to-Image Synthesis", + "text_level": 1, + "bbox": [ + 500, + 664, + 720, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We may draw inspiration from text-to-image generation methods. Typically, many research works [36,40,53] adopt the conditional GAN architecture, where they directly take the text features and concatenate them with the random noise as the input. Recently, autoregressive models [7, 39] and diffusion models [29, 38, 41, 42] made great improvement on text to image synthesis while demanding huge computational resources and massive training data.", + "bbox": [ + 496, + 688, + 893, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the introduction of StyleGAN [14-16] mapping networks, the input random noise can be first mapped to another latent space that has disentangled semantics, then the model can generate images with better quality. Further, exploring the latent space of StyleGAN has been proved useful by several works [1, 17, 33] in text-driven image synthe", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16806", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08cd0e04469de448a7fa81f72149eb38221d57c37e541e9b0ea5c09fac8c2413.jpg", + "image_caption": [ + "Figure 2. Our proposed framework supports text-guided 3D shape generation training without paired text labeling. To generate captions for training, we first adopt a pseudo caption generation module to produce textual descriptions, given rendered 2D images. Then we feed the pseudo caption CLIP embeddings into the mapping networks to train the controllable text-guided 3D shape synthesis network. Our training pipeline is supervised by high-level semantic loss $\\mathcal{L}_{clip}$ and low-level image regularization $\\mathcal{L}_{img}$ . It is notable that during the training phase, only the mapping network weights are updated, and the rest components are fixed." + ], + "image_footnote": [], + "bbox": [ + 124, + 88, + 841, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sis and manipulation tasks, where they utilize the pretrained vision-language model CLIP [37] model to manipulate pretrained unconditional StyleGAN networks. In order to relieve the need for paired text data during the training phase, Lafite [55] proposes to adopt the image CLIP embeddings as training input while using text CLIP embedding during the inference phase. However, as discussed in the recent works [20, 54], there exists a modality gap between text and image embeddings in the CLIP latent space. Therefore, language-free methods that are solely trained on CLIP image embedding could introduce distribution mismatch when encountering the CLIP text embedding input during inference. In this paper, we further explore language-free training for text-guided 3D textured shape generation by proposing a pseudo caption generation module.", + "bbox": [ + 75, + 371, + 472, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 609, + 169, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 633, + 207, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GET3D [9] is a newly proposed 3D generator that can generate high-fidelity textured 3D shapes. Specifically, GET3D maps noise vectors $\\mathbf{z} \\in \\mathcal{N}(0, \\mathbf{I})$ to a textured mesh. In order to disentangle the geometry and texture information, GET3D takes two random vectors $\\mathbf{z}_1$ and $\\mathbf{z}_2$ as inputs for the geometry and texture generation branches, respectively. Following the design of StyleGAN [15], and EG3D [3], they map $\\mathbf{z}_1$ and $\\mathbf{z}_2$ to the intermediate latent codes $\\mathbf{w}_1$ and $\\mathbf{w}_2$ with two mapping networks since the semantics would be better represented in the mapped space. The geometry and texture generators are built conditioned on $\\mathbf{w}_1$ and $\\mathbf{w}_2$ .", + "bbox": [ + 75, + 657, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 76, + 847, + 187, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the GET3D [9] model only supports unconditional 3D generation, if one need to do text-guided 3D shape syn", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "thesis, additional optimization is required with the original GET3D model. In order to control the semantics of the generated results and achieve text-based textured 3D generation in one shot, we propose to add text conditions to GET3D models. In Fig. 2, we present our training framework.", + "bbox": [ + 496, + 371, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we address two main challenges: one is the lack of training textual descriptions, and another is color and geometry distortion. Regarding the first challenge, we propose a pseudo caption generation module, which gives large amounts of captions for training purposes. These captions are used as conditions and fed into the mapping networks to control the semantic meanings of the output textured 3D shapes. We apply the high-level CLIP [37] semantic loss on the paired input text and rendered 2D images to achieve text-3D alignment. To resolve the second challenge, we propose to regularize the color and geometry training with low-level image regularization. This method provides complimentary image-level guidance to our learning objective, which gives fine-grained textures during the generation phase.", + "bbox": [ + 496, + 446, + 892, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, we empirically adopt the pretrained GET3D model and train its mapping networks only since we observe this practice not only improves the training efficiency but also stabilizes the training process.", + "bbox": [ + 496, + 672, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Pseudo Caption Generation", + "text_level": 1, + "bbox": [ + 500, + 741, + 750, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most recent 3D shape generation models are data-driven. With large-scale training data, their performance improvement is remarkable. However, when we aim to train a text-guided 3D shape generation model, we can only obtain a few textual descriptions for limited samples. To resolve this issue, we follow the method of [54] and propose to generate pseudo captions for each 3D object.", + "bbox": [ + 496, + 763, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we observe the textual descriptions for a single 3D object are highly structured, e.g., \"A yellow", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7cc311faf542181fb0e5ac97ac8033034c713890872e9c45f089f39fd4023e5e.jpg", + "image_caption": [ + "Figure 3. Our proposed pseudo caption generation module can be split into four steps. Firstly, we construct vocabulary from ShapeNet-related nouns and adjectives in the CLIP vocabulary. Then we retrieve multiple words based on the 2D-rendered images. In the third step, we create candidate captions with retrieved words. Finally, we select one caption based on the computed text-image similarities with the CLIP model." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 472, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sports car\", which typically consists of adjectives and a noun. Therefore, it is critical to obtain suitable adjectives and nouns for given 3D objects to construct pseudo captions automatically. To this end, we adopt the pretrained text-image similarity model of CLIP [37] to retrieve useful nouns and adjectives and further build complete sentences.", + "bbox": [ + 75, + 329, + 468, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pseudo caption generation module can be split into four steps, shown in Fig. 3. In the first step, we construct a vocabulary to cover the related adjectives and nouns. Since we aim to generate ShapeNet [5] object classes, we collect all the class names as nouns in the vocabulary, which has 169 tokens. Then we extract all the adjectives from CLIP vocabulary with the Natural Language Toolkit [2]. The size of our collected adjectives is 1463.", + "bbox": [ + 75, + 424, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second and third steps, we aim to retrieve relevant words from 2D-rendered images and construct candidate captions. Technically, we first extract all the vocabulary word embeddings and rendered image embeddings, where we adopt the pretrained CLIP text and image encoders, respectively. Then for each object, we compute the cosine similarities between the given image embedding and all the text embeddings. It is notable that we retrieve the nouns and adjectives separately, where we select the top- $K_{1}$ retrieved nouns and top- $K_{2}$ retrieved adjectives. With the selected words, we utilize the templates to construct candidate sentences. Since these sentences are only supposed to describe a static object, we use the template \"a{Adjective} {Noun}\", where we put 1 or 2 adjectives randomly to increase the caption diversity. Due to having $K_{1}$ nouns and $K_{2}$ adjectives, we are able to combine them and produce multiple candidate pseudo captions.", + "bbox": [ + 75, + 549, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to obtain the most suitable captions for given objects, we further conduct image-to-candidate caption retrieval, where we also use CLIP encoders to extract the image and sentence embeddings. To be specific, we use the top retrieved sentences as the final pseudo captions, which are used as the input condition for the 3D generator.", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72c57836373d24beb7efbfac9109fb78bf4b5a96d827177b3aecb4158bbffdd8.jpg", + "image_caption": [ + "Figure 4. In the training phase, we not only conduct the high-level text-to-image similarity learning through the CLIP model but also use low-level image regularization to obtain fine-grained textures and diverse shapes, where we render real and fake images with the same camera poses and apply background augmentation to allow the model focus on object features only." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Text-3D Alignment Training", + "text_level": 1, + "bbox": [ + 498, + 335, + 754, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the pseudo caption generation, we feed these captions into a pretrained GET3D [9] model in an attempt to control the semantic meanings of generated 3D objects. To be specific, we first adopt the CLIP text encoder $E_{t}$ to extract the caption $t$ features, then the caption features $E_{t}(t)$ are served as conditions and fed into the mapping networks of pretrained GET3D models. The input of the original mapping networks is random noise only. To inject the text information, we concatenate the caption embeddings and the random noise together as the new input. This enables us to generate diverse 3D objects that match the text's semantic meanings.", + "bbox": [ + 496, + 359, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we adopt the pretrained 3D generator as our model backbone, which has a good ability to synthesize high-fidelity 3D textured shapes, fixing the generator weights allows our model to focus on getting semantic alignment between input textual descriptions and generated 3D objects and also alleviates the learning difficulty. Our main training objective is to maximize the image-text similarity measured by the CLIP model, which can be denoted as:", + "bbox": [ + 496, + 541, + 890, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l i p}} = 1 - \\cos \\left(E _ {i} \\left(I _ {x}\\right), E _ {t} (t)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 674, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $E_{i}$ and $I_{x}$ represent the CLIP image encoder and rendered images, respectively. These 2D images are rendered from the generated 3D textured shapes with randomly sampled camera poses.", + "bbox": [ + 496, + 700, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Towards Diverse and Fine-Grained Generation", + "text_level": 1, + "bbox": [ + 500, + 771, + 890, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is observed using $\\mathcal{L}_{clip}$ only enables the model to generate plausible 3D shapes that align with the input text. However, the diversity of geometry generation is limited, and the generated colors are unnatural. This is caused by two main reasons. One reason is the diversity of our pseudo captions is also limited, as there are only a few descriptive words for a single object class. Moreover, the CLIP model", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16808", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "may wrongly retrieve irrelevant words to the given objects, which also affects the generated shape quality. Another reason is the CLIP model only provides high-level guidance for our training supervision while failing to give fine-grained texture information, thus making the generated results unrealistic.", + "bbox": [ + 75, + 90, + 468, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to resolve this issue, we further introduce a low-level image regularization loss. Specifically, we use the same camera poses as that used in the $I_{x}$ rendering to produce ground truth rendered images $I_{x}^{gt}$ . The optimization objective is denoted as", + "bbox": [ + 75, + 181, + 467, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m g} = 1 - \\cos \\left(E _ {i} \\left(I _ {x}\\right), E _ {i} \\left(I _ {x} ^ {g t}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 265, + 468, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the rendering process returns an alpha channel, which results in a simple white or black background, if we directly take rendered images with the same backgrounds over training iterations, the background information will also be involved during optimizing Eq. (2). To alleviate this issue and allow the model to focus on the foreground objects, we follow [13] to conduct background augmentation, as shown in Fig. 4. Technically, we randomly select random Fourier textures [28], Gaussian noise, and checkerboard patterns as backgrounds. For each paired real and fake image, we apply the same augmented backgrounds.", + "bbox": [ + 75, + 291, + 468, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Overall Training Objective", + "text_level": 1, + "bbox": [ + 76, + 465, + 320, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our overall training objective can be depicted as", + "bbox": [ + 96, + 489, + 416, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c l i p}} + \\mathcal {L} _ {\\text {i m g}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 513, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training phase, we empirically experiment with updating all model parameters and the mapping networks only. We observe training all GET3D backbone parameters would severely distort the generated geometry and textures, where we show the experimental results in the supplementary material. This is because the learning objective is two-pronged: improving the generation quality and matching the input captions. While borrowing a pretrained unconditional generator ensures good generation quality, expedites the training process, and also yields stable text-guided results. Therefore, we regard training the mapping networks only of a pretrained 3D generator as a useful and practical solution for the text-guided 3D shape generation task.", + "bbox": [ + 75, + 537, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 746, + 207, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 771, + 171, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train and evaluate our method on the large-scale 3D dataset ShapeNet [5]. Our experiments are conducted on four classes with complex geometry, Car, Table, Chair, and Motorbike, containing 7497, 8436, 6778, and 337 shapes, respectively. We follow GET3D [9] to split the data into training, validation, and test sets in a 7:1:2 ratio. The training image data are rendered following the GET3D setting.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d0298588e05126143c53974f0caaadb948f5b5028ed38b290d618f44048f526.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 602, + 125 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5ea777085ae903bc12528e00f1adfd35e2bce12d7aacca7e5aa94ff4efdc5186.jpg", + "image_caption": [ + "a white automobile a low race car" + ], + "image_footnote": [], + "bbox": [ + 617, + 88, + 679, + 127 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4202d1f167d4ad5b76527e3cb0b2f603e665b049476448befe80fd14668bc3fc.jpg", + "image_caption": [ + "a stationary chair" + ], + "image_footnote": [], + "bbox": [ + 535, + 141, + 589, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f5c112a5191829020810077f7cebeef6f5fbf989993a15222f5c19fe4fdb18b.jpg", + "image_caption": [ + "a white ladderback chair" + ], + "image_footnote": [], + "bbox": [ + 632, + 141, + 678, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f6ac98f4cafae482ef8f6b63e6cb2198b95e2df6871bc81c820fc9276fac14a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 90, + 774, + 121 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b2151929695bbb231861c8350195ef66e524e09f6cb55852f1f312dc7cf2ec2.jpg", + "image_caption": [ + "a ural motorbike" + ], + "image_footnote": [], + "bbox": [ + 705, + 155, + 774, + 185 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd2166693acb868fede55b4cbabd744a30ef4e28a3ad8909030ab71d5bc685f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 90, + 856, + 125 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f09374256299dd939c36cb47a2e9bafc431f706a08435b21ebe92e4bd7ffde6.jpg", + "image_caption": [ + "a brown motorcycle", + "a structural table", + "Figure 5. The demonstration of our generated pseudo captions." + ], + "image_footnote": [], + "bbox": [ + 784, + 150, + 859, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We render 24 views randomly for Car, Chair and Table and 100 views for Motorbike since it has less available data. The pretrained models were trained on each class separately.", + "bbox": [ + 498, + 258, + 890, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 325, + 712, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the GET3D [9] model pretrained on ShapeNet [5] dataset as our backbone 3D generator. In the pseudo caption generation module, we set $K_{1}$ and $K_{2}$ as 3 and 6, respectively. We generate 20 pseudo captions for each ShapeNet object. During the training phase, each input image is randomly paired with one of the pseudo captions generated for the object which the image was rendered from. An example of the generated pseudo can be found in Fig. 5. We only update the geometry and texture mapping networks. We set the batch size to 16 and ran all the experiments on 4 Nvidia Tesla V100-32G GPUs. We empirically observe the geometry branch is more sensitive to the weight update. Hence we set learning rates for geometry and texture mapping networks as 0.004 and 0.0005, respectively. It costs around 10 hours for model training.", + "bbox": [ + 496, + 352, + 890, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 500, + 599, + 807, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1 Qualitative comparison", + "text_level": 1, + "bbox": [ + 500, + 627, + 717, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 6, we first compare with the method TITG3SG [25], which is an encoder-decoder-based framework. They only experiment with object classes of Chair and Table. Here we conduct the comparison qualitatively. Specifically, we show the generated explicit 3D shapes with the corresponding text prompts. Since TITG3SG [25] generates 3D shapes in two steps, where the coarse voxel and refined mesh are produced separately. We show them both in the top row of Fig. 6. The second row presents our generated mesh in two different views, in which we generate final 3D results in one shot. We observe our method gives better 3D textured shapes than TITG3SG, which is a fully supervised method. For example, in the third generated sample of TITG3SG, the chair legs are distorted. Moreover, the high alignment between input prompts and our generated results demonstrates our text control ability.", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16809", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a26605101e82a9003bb56173ac0c12a9e6fdf5eab9071f71e5942fea2dead3b7.jpg", + "image_caption": [ + "Figure 6. We compare the text-guided generation results of our methods with that of the fully supervised method TITG3SG [25] on two object classes, Chair and Table. On the top row, we present the results from [25], for each text prompt, the top left one is the raw voxel output in $64^{3}$ resolution, and the top right one is the refined mesh based on the voxel output. The bottom row shows our mesh output with two different views. The renderings are generated with ChimeraX [10]." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 843, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Quantitative comparison", + "text_level": 1, + "bbox": [ + 76, + 303, + 305, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate our 3D generation quality, we compare existing works [25,43,50] from three aspects: 1) rendered 2D image quality using FID, 2) text-image relevance using R-Precision, and 3) 3D geometry quality using FPD.", + "bbox": [ + 75, + 327, + 468, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we provide a quantitative comparison with CLIP-NeRF [50], which utilizes human-written captions for training. We employ the Fréchet Inception Distance (FID) [12] as our evaluation metric and present results for two object classes: Car and Chair. To ensure fair comparisons, we downsample our results to match the resolution of CLIP-NeRF [50]. CLIP-NeRF employs NeRF as the 3D generator, which imposes limitations on the generation resolution. In contrast, our backbone GET3D model boasts a larger capacity, supporting higher resolution. Experimental results demonstrate that our method surpasses CLIP-NeRF in text-guided generation quality.", + "bbox": [ + 75, + 388, + 470, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we adopt the CLIP-R-Precision metric [32] to evaluate the consistency between input text prompts and generated shapes. Given that our model is category-specific, we perform per-category CLIP-R-Precision and compare it to the fully supervised TITG3SG method [25]. According to [32], we compute the CLIP-R-Precision using a set of unseen random captions with words sampled from existing captions, thereby enhancing correlation with human evaluations and mitigating model bias inherent compared to the conventional R-Precision metric. It is observed our method outperforms TITG3SG [25] in each category.", + "bbox": [ + 75, + 569, + 468, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 3, we assess the geometry generation quality using the Fréchet Point Distance (FPD) [19, 47], which measures the feature distance of a PointNet [35] pretrained on ModelNet40 [51]. Our method outperforms existing approaches in terms of geometry quality by a large margin.", + "bbox": [ + 75, + 736, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.3 Inference speed comparison", + "text_level": 1, + "bbox": [ + 76, + 830, + 326, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 4, we compare the inference speed of our method with other works. The optimization-based methods DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18]", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a0a825d3ebb2fb2d571fb8f9a17327796943a7824c33a111a0340efbe1a2f98.jpg", + "table_caption": [ + "Table 1. Comparison with the existing work. We evaluate the rendered 2D images using Fréchet inception distance (FID). We downsample our result to the same resolution of CLIP-NeRF [50] for fair comparisons." + ], + "table_footnote": [], + "table_body": "
CarChair
ResolutionFIDResolutionFID
CLIP-NeRF [50]256267.8128248.4
Ours256220.1128243.7
Ours1024221.71024244.8
", + "bbox": [ + 511, + 357, + 874, + 441 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/92cce2d5317a09c1fc141dc0faf2a5595aaf05989377567b8b45a3e22a8690f1.jpg", + "table_caption": [ + "Table 2. Per category CLIP-R-Precision score using random generated captions on categories Chair and Table." + ], + "table_footnote": [], + "table_body": "
MethodChairTable
TITG3SG [25]9.05 ± 1.926.67 ± 3.25
Ours20.19 ± 1.8915.21 ± 3.15
", + "bbox": [ + 545, + 477, + 839, + 532 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/829ac951fe35224096c1715e9da4260edfa4eebe7228f55a474c7cd358342ea3.jpg", + "table_caption": [ + "Table 3. Comparison of 3D generation quality in FPD score." + ], + "table_footnote": [], + "table_body": "
MethodChairTable
TITG3SG [25]1566.761639.68
CLIP-Forge [43]825.963051.31
Ours342.231468.43
", + "bbox": [ + 573, + 553, + 815, + 619 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/53c1fb399f3fc519a409ec1da7e71a12819b51300af3df511567f76594068585.jpg", + "table_caption": [ + "Table 4. The comparison of inference time per text prompt for different methods. The inference time of DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18] are given in their papers. For TITG3SG [25] and ours, we generate 1000 text prompts with batch size as 1, and calculate the average inference time." + ], + "table_footnote": [], + "table_body": "
MethodDeviceOutputTime
DreamFields [13]TPU cores x8Rendering72 min
DreamFusion [34]TPUv4 machineRendering90 min
PureCLIPNeRF [18]GTX 2080tiRendering20 min
TITG3SG [25]Telsa V100-32GVoxel2.21 sec
TITG3SG [25]Telsa V100-32GMesh24.44 sec
OursTelsa V100-32GRendering0.05 sec
OursTelsa V100-32GMesh7.09 sec
", + "bbox": [ + 504, + 694, + 887, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "take tens of minutes to optimize for each input text prompt. Liu et al. [25] use an encoder-decoder architecture and thus can perform inference at 2.21 seconds. However, the raw $64^{3}$ voxel output is very coarse, as shown in Fig. 6. The", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16810", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a1b23870babc5dc897131f96736b6ca1f9857507acdcea6c3ffac4a67492bc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 108, + 173, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/214e06bdaeed14a83587464862d93025b28c2dcac8e6709ba5e6ee9ee04b19b3.jpg", + "image_caption": [ + "\"a blue SUV\"" + ], + "image_footnote": [], + "bbox": [ + 196, + 109, + 274, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edff6ab99cc5e40d1106ac6e4d0994c66561205906593d45b1f432e024e26ab2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 109, + 370, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58ccc99d37b1dca0382cd1ba4dba6faee5094f3c3994f66bea8d0125c2ad7a68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 109, + 457, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eded246d9e8aa73d422f8e0882e7586a966445a068d7c6a6d6795fa1a09e8e70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 181, + 176, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef9afded5d387628fd364d2c4cd56fae90b334f0f9c5cddf3731d9dd3f89c94d.jpg", + "image_caption": [ + "\"a red hatchback\"" + ], + "image_footnote": [], + "bbox": [ + 197, + 186, + 272, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56a667ddbf78bb4336889deaaa17edfb52b124a8705985db8b10e246c62eec8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 186, + 367, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8541ce9317e965bf6cdebc994ca1977d5bc89a6eaf0ac595ad5e6bab6a9ec569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 188, + 455, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f48239947a516ed8281391162e8b735a1991902490a7b1709a690f404f4f28d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 270, + 174, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6dba14818a767c762bc1c615c352eaa3e265e07b29184724732082e23352fd9.jpg", + "image_caption": [ + "\"a yellow racing car\"" + ], + "image_footnote": [], + "bbox": [ + 191, + 268, + 264, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e201e724c96f4162c4d70b1afad2ff6fa4621bd39a8ad34d9b08da5d095b334.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 270, + 364, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6399f71a632cac11e6e3b17283d083c0cc5dba9b825eea9f2c8edd63a28d0300.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 271, + 452, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa86196d4cb1cb3f9954adda2ae59cfb0899ab568c5c0e0896fa8adb3547eda6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 343, + 161, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ff9b9d32a0cf3392703b2bb00814600bc0b41834526fb51e8e108ae5f6c6d47.jpg", + "image_caption": [ + "\"a yellow racing motorcycle\"" + ], + "image_footnote": [], + "bbox": [ + 197, + 347, + 254, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b75b8296f04efc925635434639a9a29bf1d35a590a270ca70da5cb5eefbc0c6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 345, + 356, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3717f6429816b70c471113303b3b685dc9883adcdf85fb54d9eb4d69f58db01c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 345, + 442, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c5444d8ab1be15714f6af8f158bf8034ff217e8b199d136d3337dfa425331df.jpg", + "image_caption": [ + "Figure 7. We show the generation diversity and text control of our method. Each row takes the same text prompt with different sampled random noise as the model input." + ], + "image_footnote": [], + "bbox": [ + 102, + 421, + 163, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a504c1a2ce7714b3c31784ba414f208110df88e6c338551072bec69c55e8050.jpg", + "image_caption": [ + "\"a red motorbike\"" + ], + "image_footnote": [], + "bbox": [ + 197, + 421, + 256, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48bde7ece2ef1a61c57ff0b227a2c0ce85e743865660cd39a43aaf68080a2dff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 422, + 354, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f10bb1b7e7d35fbb88c837e625718343c0f1a3ae78764dbdd7f0318dba7f7a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 416, + 444, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7609661119dc0896a8d575b3614efcb3103d4a4281823c6f3e0df1ec03eba42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 104, + 566, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1ee4643d9dfac1ad7dc9001e9ff52e6e23d8a11391bc868bc3be330322c3d2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 104, + 653, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecdd8747f4766720a015829bea424167f28343546e9f6b9fcdab4f1950f76102.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 104, + 745, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87ce7ea8c4cb3e8e58d6e409e593cbb48d717015a4b2ef2dc0b39ae37ddfdc4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 178, + 563, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ab7114968b1bedd42ef98ceb4f7bb6f731b2b50e66e1a27c810852196f35a9a.jpg", + "image_caption": [ + "\"agrayarmlesschair\"" + ], + "image_footnote": [], + "bbox": [ + 593, + 178, + 651, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c7f1bf21c71b346cdc59a99f252e6d8da8539e2cef8a801c640d0af4b7a900f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 178, + 745, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55d0a9969116303f0c59ccd0d5ac931693d5f40020bdea9e7fb1c062ab7f5e62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 260, + 550, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61436fb6d9ae7b707e943bb5eb78cbebba944da9bd7645debe72a86d4dd1dd5f.jpg", + "image_caption": [ + "\"agreenchair\"" + ], + "image_footnote": [], + "bbox": [ + 581, + 260, + 651, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae3808d742a0176ed08b9ff46f488081d80ed60c56c7f58dbdafa9031bb7ddb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 260, + 746, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d984dd161ffa86bbe648beb7a77a35a03b5899e2c0915474ede145863aa28088.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 107, + 848, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcd2f436a5274261f889d67dc87eb795f098949fd882927f5f563f66b6415fc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 179, + 839, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a304aa388f185a7249b6020135e6e9b99b6c1787e3d42417d857084e1dc567a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 342, + 560, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f57037c3d6098c55afe9912ebc8fbf53b42b79ff56585e0ff49446b496f489f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 345, + 656, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e986e0072a32f0a06f4006085075158792a8c6c2a76a57a162d968e8dca14293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 345, + 750, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c8a27f3c506847b373da38df7aff79cd649a23c148a6fb3bcfaa887883efa43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 260, + 849, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b3cf9233e942161b2dacb19c7fa10774c7854d9ad3161f2583210b79f297b36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 345, + 849, + 392 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf0aa2c2aa13c6eb3d9ad470e1c84c1c1bb06d9d97bdeb145733ccfac5d07082.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 417, + 563, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bacb4b37a17e1ae89c902e02d476d7661e41b116482370a209d18f8b4106f6c.jpg", + "image_caption": [ + "\"a red dining table\"", + "\"a silver metallic table\"" + ], + "image_footnote": [], + "bbox": [ + 573, + 419, + 656, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6454e7e27334ed49ea3be4a0687aecd664489cbeb25263f0c29e859ba09ecb0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 419, + 754, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7573347e3ed9cedfe310f5ea8e937b3651d42520bc892685df5e6e9eca4b75a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 420, + 849, + 463 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3643b402ba20c8916ad4ad52bf02c7df9d6d6fb0d009a5451ed46e11eca1f0a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 541, + 163, + 569 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7608b1350c6c1f3a294010a047daf196b0d0c1d675438a8d9da81519d024479c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 542, + 235, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37038c41b49cd9beda1ac903ba2afeeebca34cdff84ab236a0b7235c2c008157.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 542, + 305, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad989530d6a8c339a54c11ffa8975b30c5f387fdc9392994649aa616aa87be06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 542, + 377, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eaba7cb8d453835a6cab9c519b09897f7d22bc1d9c64aa8965875ded873a711f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 542, + 450, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9dbfc68e7cc4d5d38097902f5ccb3162f8799b7147c2cd47a095e0931d9baf8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 583, + 156, + 633 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a13f994c575b76ecd3a24fdc29c64d37773f7e1050e5ec6d07bbece13647562f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 584, + 223, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fdfe74916a25d18e386930f44abc70b3a545246c1c31a18bb87ed41c3ea7274.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 584, + 290, + 631 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1188f55a4c568cf4b911b82a79698c859f963a6716f8c9fdd5548af3fa5fdadd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 583, + 367, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf100d6544b188189ffc190003af17a3f9c449f7b49bf39e28568529c31597af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 584, + 437, + 632 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64f04db05a9e3c07045b5ecee724a241779d4e14bf4b79eb78991d0935232b39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 647, + 161, + 679 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1cdc7f32f01adf592a78141e19a5170ad54eb071c03ac18de3dc1a72d7dd315d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 646, + 235, + 679 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/746b9a72f10d9bce6dcb331a629dcd22f7f4ae9a949fef4b44c10f03080a60e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 646, + 305, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff0a8a2744f862b14fb080a8d442f3fb88db7891bf89748240560faecbece891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 646, + 377, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff3fc6719ea660eee7d6d617819f13ff6b6d1bb58e7077cfee63e70e1cfbac74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 646, + 450, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c190a8d91e5e0bed32c5b35968049a471c5ad1da6ea538c721a0236a5e3df7a.jpg", + "image_caption": [ + "\"ared{class}\"", + "Figure 8. Text control of textures on four different categories. Each row shares the same sampled random noise input." + ], + "image_footnote": [], + "bbox": [ + 101, + 686, + 143, + 719 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7eb3bd917cb8776d698cdac89beebbbdc6eff12240991edc0d378b5de1610916.jpg", + "image_caption": [ + "\"a yellow {class}\"" + ], + "image_footnote": [], + "bbox": [ + 174, + 686, + 215, + 719 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/596cb04ec69713b9445ed85fb327a9d3191c463cdc924790aecc6012be5f028f.jpg", + "image_caption": [ + "\"agreen{class}" + ], + "image_footnote": [], + "bbox": [ + 246, + 686, + 290, + 719 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79c91a0412a11ea75c96a94c7fa7c318dc67f0a158f267dd5b3967c31db44673.jpg", + "image_caption": [ + "\"a blue {class}\"" + ], + "image_footnote": [], + "bbox": [ + 320, + 686, + 361, + 719 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34d26746bdb3bcde65d7beac3cde946ef5c7d531a9f2e9f404ad883a86f850f2.jpg", + "image_caption": [ + "\"a gray {class}\"" + ], + "image_footnote": [], + "bbox": [ + 393, + 686, + 436, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coarse output is later refined with interpolations and marching cubes to generate mesh data. This process takes ten times longer than the network inference time and costs 24.44 seconds per sample. Our method generates high-resolution renderings at $1024^2$ resolution with only 0.05 seconds and generates the complete explicit mesh representation in 7.09 seconds.", + "bbox": [ + 75, + 794, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 534, + 681, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 7, we present the generation diversity and text control strength of our method in four object classes. In each row, we sample different random noises as well as the given text prompt as input to generate diversified and text-controlled 3D shapes, where the input text features are simultaneously fed into both the geometry and texture mapping network. Generally, we observe the semantic meanings of the rendered 2D images, and the given prompts are consistent. Moreover, there exists large diversity in the texture and geometry of our generated 3D shapes, even from the same text input. Here we present generated results of 4 classes, where $Car$ has a relatively simple geometry, while Table and Chair are more complex. But we still have good control over the textures and geometries for the challenging classes. For example, when we input \"a gray armless chair\", the generated samples exactly match the given instructions. However, we also observe failure cases of Table generation. For instance, the generation of \"a wooden office desk\" may have distorted geometry due to its complex structure. In Fig. 8, we show the text control ability of our method on different classes. The results in each row are generated from the same random noise input but with dif", + "bbox": [ + 496, + 568, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16811", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9de2d828a5ac7eeb5f60f28323b90d2523953aef9bf27f80649d0719142d47e3.jpg", + "table_caption": [ + "Table 5. We compare the FID score of using image CLIP features (Lafite [55]) and the pseudo text CLIP features (Ours) as training input." + ], + "table_footnote": [], + "table_body": "
MethodCarChairTableMotorbike
Lafite [55]68.781.194.3146.5
Ours21.744.843.257.2
", + "bbox": [ + 112, + 130, + 431, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e2066ec54d9853789e025666645403313a5145a394d4392e0cf1a181db644aec.jpg", + "table_caption": [ + "Table 6. We compare the FID score for our method without $\\mathcal{L}_{clip}$ , $\\mathcal{L}_{img}$ and background augmentation (BGaug)." + ], + "table_footnote": [], + "table_body": "
MethodCarChairTableMotorbike
Ours21.744.843.257.2
-w/o Lclip83.7148.967.1134.9
-w/o Limg59.8144.248.4101.8
-w/o BGaug55.165.158.1110.0
", + "bbox": [ + 102, + 224, + 439, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ferent text prompts. To summarize, the results show that our generation results are highly consistent with the input text prompts across different object classes.", + "bbox": [ + 75, + 316, + 468, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 375, + 228, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo caption generation module. In Tab. 5, we compare our method with Lafite [55], which is a language-free training method for the text-to-image generation task. Lafite claims due to the alignment between CLIP text and image embeddings. The CLIP image features can be directly used as text input during training. To implement Lafite [55] onto our framework, we feed the rendered real image CLIP features into our model and train the model by minimizing the cosine similarity between the generated and input image CLIP features. The quantitative results show that our proposed pseudo caption generation module significantly improves the generation quality compared to the baseline. We also give a qualitative comparison in Fig. 9 (b), we observe that the generator trained with Lafite [55] fails to generate correct textures. This problem might come from the modality gap between the CLIP image and text embeddings, as discussed in [20, 54].", + "bbox": [ + 75, + 398, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of our loss functions. We conduct an ablation study on our used two losses $\\mathcal{L}_{clip}$ and $\\mathcal{L}_{img}$ . In Tab. 6, we quantitatively validate the usefulness of both two losses. Specifically, $\\mathcal{L}_{clip}$ supervises the text-image alignment learning, which is more critical than $\\mathcal{L}_{img}$ , and removing $\\mathcal{L}_{clip}$ gives a more considerable performance drop. From the qualitative results in Fig. 9 (c), we observe that simply using $\\mathcal{L}_{img}$ leads to the model generating distorted geometries, which may indicate the significance of high-level semantic supervision. We also observe that solely using $\\mathcal{L}_{clip}$ (Fig. 9 (d)) may cause the model generation to lose fine-grained texture details, as well as the geometry information.", + "bbox": [ + 75, + 657, + 467, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of background augmentation. Our used GET3D [9] backbone network generates RGBA images with dark backgrounds, which may bring some noise when", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05a30e36ad4296ca32e197ddacb32424fb7993c74924a548eeca56c5ae6bf2aa.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 558, + 92, + 872, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a74c124e3363581d94064041940b941d7f12c5160f11a032706385921f174bf0.jpg", + "image_caption": [ + "Figure 9. Ablation study on our methods with text prompt \"a red car\". We show results of (a) our full model, (b) Lafite [55], (c) without high-level semantic regularization $\\mathcal{L}_{clip}$ , (d) without low-level image regularization $\\mathcal{L}_{img}$ , (e) without background augmentation. All the visualizations are generated from models trained for 250k iterations.", + "Figure 10. Examples of failure cases of our method. Our model often fails to handle text input with composed descriptions for different parts of the object." + ], + "image_footnote": [ + "\"acarwithyellowlegs, redseat and blue back.\"" + ], + "bbox": [ + 702, + 382, + 867, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we conduct low-level image regularization. As illustrated from the quantitative results in Tab. 6 and the visualizations in Fig. 9 (e), training with the augmented background can help remove the texture flaws.", + "bbox": [ + 496, + 506, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 579, + 617, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce TAPS3D, a text-guided 3D textured shape generation method. We first generate pseudo captions for model training. During the inference phase, TAPS3D does not require additional optimization. Our fast generation speed enables average users to generate high-quality 3D objects with acceptable processing time.", + "bbox": [ + 496, + 604, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Limitation.** The limitation of our method mainly lies in the capacity of our model is still bound to the training image data, and we cannot produce different fine-grained details for different object parts, as shown in Fig. 10. Although we do not require paired text data, we still need diversified training images to handle complex text input.", + "bbox": [ + 496, + 696, + 890, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 800, + 650, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is partly supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-RP-2018-003), and the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and MoE AcRF Tier 1 grants (RG14/22, RG95/20).", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16812", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2", + "[2] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc., 2009. 4", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 3", + "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4, 5", + "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2", + "[7] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2", + "[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. arXiv preprint arXiv:2207.09446, 2022. 2", + "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv preprint arXiv:2209.11163, 2022. 1, 2, 3, 4, 5, 8", + "[10] Thomas D Goddard, Conrad C Huang, Elaine C Meng, Eric F Pettersen, Gregory S Couch, John H Morris, and Thomas E Ferrin. Ucsf chimerax: Meeting modern challenges in visualization and analysis. Protein Science, 27(1):14-25, 2018. 1, 6", + "[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylererf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6", + "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2", + "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3", + "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[17] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 895–904, 2022. 2", + "[18] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 2, 6", + "[19] Ruihui Li, Xianzhi Li, Ka-Hei Hui, and Chi-Wing Fu. Spagan: Sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 6", + "[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 2, 3, 8", + "[21] Kangcheng Liu. Rm3d: Robust data-efficient 3d scene parsing via traditional and learnt 3d descriptors-based semantic region merging. International Journal of Computer Vision, pages 1-30, 2022. 1", + "[22] Kangcheng Liu, Zhi Gao, Feng Lin, and Ben M Chen. Fgnet: A fast and accurate framework for large-scale lidar point cloud understanding. IEEE Transactions on Cybernetics, 53(1):553-564, 2022. 1", + "[23] Kangcheng Liu, Aoran Xiao, Xiaoqin Zhang, Shijian Lu, and Ling Shao. Fac: 3d representation learning via foreground aware feature contrast. arXiv preprint arXiv:2303.06388, 2023.1", + "[24] Kangcheng Liu, Yuzhi Zhao, Qiang Nie, Zhi Gao, and Ben M Chen. Weakly supervised 3d scene segmentation with region-level boundary awareness and instance discrimination. In European Conference on Computer Vision 2022 (ECCV 2022), pages 37-55. Springer, Cham, 2022. 1", + "[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 6", + "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16813", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[28] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 3(7):e12, 2018. 5", + "[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2", + "[31] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2", + "[32] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6", + "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2", + "[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 6", + "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 6", + "[36] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4", + "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[40] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[42] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 6", + "[44] Aditya Sanghi, Rao Fu, Vivian Liu, Karl Willis, Hooman Shayani, Amir Hosein Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Textcraft: Zero-shot generation of high-fidelity and diverse shapes from text. arXiv preprint arXiv:2211.01427, 2022. 1, 2", + "[45] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 1", + "[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 1", + "[47] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 6", + "[48] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. 3d pose transfer with correspondence learning and mesh refinement. Advances in Neural Information Processing Systems, 34:3108-3120, 2021. 1", + "[49] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. Unsupervised 3d pose transfer with cross consistency and dual reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[50] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16814", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 1, 2, 6", + "[51] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6", + "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2", + "[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2", + "[54] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2, 3, 8", + "[55] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 3, 8" + ], + "bbox": [ + 78, + 90, + 468, + 455 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16815", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_model.json b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_model.json new file mode 100644 index 0000000000000000000000000000000000000000..305d6837cde2681a6a5ec3e36df490ec3dffa111 --- /dev/null +++ b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_model.json @@ -0,0 +1,3269 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "TAPS3D: Text-Guided 3D Textured Shape Generation from Pseudo Supervision" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.179, + 0.851, + 0.235 + ], + "angle": 0, + "content": "Jiacheng Wei\\(^{1*}\\) Hao Wang\\(^{1*}\\) Jiashi Feng\\(^{2}\\) Guosheng Lin\\(^{1\\dagger}\\) Kim-Hui Yap\\(^{1}\\) \n\\(^{1}\\)Nanyang Technological University, Singapore \\(^{2}\\)ByteDance \n{jiacheng002@e., hao005@e., gslin@, ekhyap@}ntu.edu.sg, jshfeng@bytedance.com" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.258, + 0.291, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.342, + 0.273, + 0.356 + ], + "angle": 0, + "content": "\"a red hatchback car\"" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.262, + 0.365, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.342, + 0.466, + 0.357 + ], + "angle": 0, + "content": "\"agrayarmlesschair" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.26, + 0.476, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.274, + 0.589, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.26, + 0.688, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.342, + 0.675, + 0.357 + ], + "angle": 0, + "content": "\"awooden office table\"" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.27, + 0.781, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.26, + 0.882, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.342, + 0.857, + 0.356 + ], + "angle": 0, + "content": "\"ablock motorbike\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.895, + 0.412 + ], + "angle": 0, + "content": "Figure 1. Our proposed model is capable of generating detailed textured shapes based on given text prompts. We demonstrate the effectiveness of our approach by showcasing generated results for four distinct object classes: Car, Chair, Table, and Motorbike. Both the textured meshes and geometries are presented, with visualizations rendered using ChimeraX [10]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.425, + 0.314, + 0.44 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.474, + 0.777 + ], + "angle": 0, + "content": "In this paper, we investigate an open research task of generating controllable 3D textured shapes from the given textual descriptions. Previous works either require ground truth caption labeling or extensive optimization time. To resolve these issues, we present a novel framework, TAPS3D, to train a text-guided 3D shape generator with pseudo captions. Specifically, based on rendered 2D images, we retrieve relevant words from the CLIP vocabulary and construct pseudo captions using templates. Our constructed captions provide high-level semantic supervision for generated 3D shapes. Further, in order to produce fine-grained textures and increase geometry diversity, we propose to adopt low-level image regularization to enable fake-rendered images to align with the real ones. During the inference phase, our proposed model can generate 3D textured shapes from the given text without any additional optimization. We conduct extensive experiments to analyze each of our proposed components and show the efficacy of our framework in generating high-fidelity 3D textured and text-relevant shapes. Code is available at https://github.com/plusmultiply/TAPS3D" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.21, + 0.821 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.863 + ], + "angle": 0, + "content": "3D objects are essential in various applications [21-24], such as video games, film special effects, and virtual real" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.426, + 0.892, + 0.532 + ], + "angle": 0, + "content": "ity. However, realistic and detailed 3D object models are usually hand-crafted by well-trained artists and engineers slowly and tediously. To expedite this process, many research works [3,9,13,34,48,49] use deep generative models to achieve automatic 3D object generation. However, these models are primarily unconditioned, which can hardly generate objects as humans will." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.882 + ], + "angle": 0, + "content": "In order to control the generated 3D objects from text, prior text-to-3D generation works [43, 44] leverage the pretrained vision-language alignment model CLIP [37], such that they can only use 3D shape data to achieve zero-shot learning. For example, Dream Fields [13] combines the advantages of CLIP and NeRF [27], which can produce both 3D representations and renderings. However, Dream Fields costs about 70 minutes on 8 TPU cores to produce a single result. This means the optimization time during the inference phase is too slow to use in practice. Later on, GET3D [9] is proposed with faster inference time, which incorporates StyleGAN [15] and Deep Marching Tetrahedral (DMTet) [46] as the texture and geometry generators respectively. Since GET3D adopts a pretrained model to do text-guided synthesis, they can finish optimization in less time than Dream Fields. But the requirement of test-time optimization still limits its application scenarios. CLIP-NeRF [50] utilizes conditional radiance fields [45] to avoid test-time optimization, but it requires ground truth text data for the training purpose. Therefore, CLIP-NeRF is only applicable to a few object classes that have labeled text data for training, and its generation quality is restricted by the NeRF capacity." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To address the aforementioned limitations, we propose" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.458, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution. Work done during an internship at Bytedance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.229, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.458, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16805" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "to generate pseudo captions for 3D shape data based on their rendered 2D images and construct a large amount of \\(\\langle 3\\mathrm{D}\\) shape, pseudo captions \\(\\rangle\\) as training data, such that the text-guided 3D generation model can be trained over them. To this end, we propose a novel framework for Text-guided 3D textured shApe generation from Pseudo Supervision (TAPS3D), in which we can generate high-quality 3D shapes without requiring annotated text training data or test-time optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Specifically, our proposed framework is composed of two modules, where the first generates pseudo captions for 3D shapes and feeds them into a 3D generator to conduct text-guided training within the second module. In the pseudo caption generation module, we follow the language-free text-to-image learning scheme [20, 54]. We first adopt the CLIP model to retrieve relevant words from given rendered images. Then we construct multiple candidate sentences based on the retrieved words and pick sentences having the highest CLIP similarity scores with the given images. The selected sentences are used as our pseudo captions for each 3D shape sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.473, + 0.638 + ], + "angle": 0, + "content": "Following the notable progress of text-to-image generation models [29, 38, 39, 42, 52], we use text-conditioned GAN architecture in the text-guided 3D generator training part. We adopt the pretrained GET3D [9] model as our backbone network since it has been demonstrated to generate high-fidelity 3D textured shapes across various object classes. We input the pseudo captions as the generator conditions and supervise the training process with high-level CLIP supervision in an attempt to control the generated 3D shapes. Moreover, we introduce a low-level image regularization loss to produce fine-grained textures and increase geometry diversity. We empirically train the mapping networks only of a pretrained GET3D model so that the training is stable and fast, and also, the generation quality of the pretrained model can be preserved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.7 + ], + "angle": 0, + "content": "Our proposed model TAPS3D can produce high-quality 3D textured shapes with strong text control as shown in Fig. 1, without any per prompt test-time optimization. Our contribution can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.708, + 0.47, + 0.768 + ], + "angle": 0, + "content": "- We introduce a new 3D textured shape generative framework, which can generate high-quality and fidelity 3D shapes without requiring paired text and 3D shape training data." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.777, + 0.47, + 0.853 + ], + "angle": 0, + "content": "- We propose a simple pseudo caption generation method that enables text-conditioned 3D generator training, such that the model can generate text-controlled 3D textured shapes without test time optimization, and significantly reduce the time cost." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.862, + 0.47, + 0.908 + ], + "angle": 0, + "content": "- We introduce a low-level image regularization loss on top of the high-level CLIP loss in an attempt to produce fine-grained textures and increase geometry diversity." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.708, + 0.47, + 0.908 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.803, + 0.132 + ], + "angle": 0, + "content": "2.1. Text-Guided 3D Shape Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Text-guided 3D shape generation aims to generate 3D shapes from the textual descriptions so that the generation process can be controlled. There are mainly two categories of methods, i.e., fully-supervised and optimization-based methods. The fully-supervised method [6,8,25] uses ground truth text and the paired 3D objects with explicit 3D representations as training data. Specifically, CLIP-Forge [43] uses a two-stage training scheme, which consists of shape autoencoder training, and conditional normalizing flow training. VQ-VAE [44] performs zero-shot training with 3D voxel data by utilizing the pretrained CLIP model [37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.655 + ], + "angle": 0, + "content": "Regarding the optimization-based methods [13, 18, 26, 34], Neural Radiance Fields (NeRF) are usually adopted as the 3D generator. To generate 3D shapes for each input text prompt, they use the CLIP model to supervise the semantic alignment between rendered images and text prompts. Since NeRF suffers from extensive generation time, 3D-aware image synthesis [3, 3, 4, 11, 30, 31] has become popular, which generates multi-view consistent images by integrating neural rendering in the Generative Adversarial Networks (GANs). Specifically, there are no explicit 3D shapes generated during the process, while the 3D shapes can be extracted from the implicit representations, such as the occupancy field or signed distance function (SDF), using the marching cube algorithm. These optimization-based methods provide a solution to generate 3D shapes, but their generation speed was compromised. Although [50] is free from test-time optimization, it requires text data for training, which limits its applicable 3D object classes. Our proposed method attempts to alleviate both the paired text data shortage issues and the long optimization time in the previous work. We finally produce high-quality 3D shapes to bring our method to practical applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.722, + 0.682 + ], + "angle": 0, + "content": "2.2. Text-to-Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.894, + 0.81 + ], + "angle": 0, + "content": "We may draw inspiration from text-to-image generation methods. Typically, many research works [36,40,53] adopt the conditional GAN architecture, where they directly take the text features and concatenate them with the random noise as the input. Recently, autoregressive models [7, 39] and diffusion models [29, 38, 41, 42] made great improvement on text to image synthesis while demanding huge computational resources and massive training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "With the introduction of StyleGAN [14-16] mapping networks, the input random noise can be first mapped to another latent space that has disentangled semantics, then the model can generate images with better quality. Further, exploring the latent space of StyleGAN has been proved useful by several works [1, 17, 33] in text-driven image synthe" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16806" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.843, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Our proposed framework supports text-guided 3D shape generation training without paired text labeling. To generate captions for training, we first adopt a pseudo caption generation module to produce textual descriptions, given rendered 2D images. Then we feed the pseudo caption CLIP embeddings into the mapping networks to train the controllable text-guided 3D shape synthesis network. Our training pipeline is supervised by high-level semantic loss \\(\\mathcal{L}_{clip}\\) and low-level image regularization \\(\\mathcal{L}_{img}\\). It is notable that during the training phase, only the mapping network weights are updated, and the rest components are fixed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.599 + ], + "angle": 0, + "content": "sis and manipulation tasks, where they utilize the pretrained vision-language model CLIP [37] model to manipulate pretrained unconditional StyleGAN networks. In order to relieve the need for paired text data during the training phase, Lafite [55] proposes to adopt the image CLIP embeddings as training input while using text CLIP embedding during the inference phase. However, as discussed in the recent works [20, 54], there exists a modality gap between text and image embeddings in the CLIP latent space. Therefore, language-free methods that are solely trained on CLIP image embedding could introduce distribution mismatch when encountering the CLIP text embedding input during inference. In this paper, we further explore language-free training for text-guided 3D textured shape generation by proposing a pseudo caption generation module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.17, + 0.626 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.208, + 0.651 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.84 + ], + "angle": 0, + "content": "GET3D [9] is a newly proposed 3D generator that can generate high-fidelity textured 3D shapes. Specifically, GET3D maps noise vectors \\(\\mathbf{z} \\in \\mathcal{N}(0, \\mathbf{I})\\) to a textured mesh. In order to disentangle the geometry and texture information, GET3D takes two random vectors \\(\\mathbf{z}_1\\) and \\(\\mathbf{z}_2\\) as inputs for the geometry and texture generation branches, respectively. Following the design of StyleGAN [15], and EG3D [3], they map \\(\\mathbf{z}_1\\) and \\(\\mathbf{z}_2\\) to the intermediate latent codes \\(\\mathbf{w}_1\\) and \\(\\mathbf{w}_2\\) with two mapping networks since the semantics would be better represented in the mapped space. The geometry and texture generators are built conditioned on \\(\\mathbf{w}_1\\) and \\(\\mathbf{w}_2\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.188, + 0.862 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Since the GET3D [9] model only supports unconditional 3D generation, if one need to do text-guided 3D shape syn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.447 + ], + "angle": 0, + "content": "thesis, additional optimization is required with the original GET3D model. In order to control the semantics of the generated results and achieve text-based textured 3D generation in one shot, we propose to add text conditions to GET3D models. In Fig. 2, we present our training framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.674 + ], + "angle": 0, + "content": "To this end, we address two main challenges: one is the lack of training textual descriptions, and another is color and geometry distortion. Regarding the first challenge, we propose a pseudo caption generation module, which gives large amounts of captions for training purposes. These captions are used as conditions and fed into the mapping networks to control the semantic meanings of the output textured 3D shapes. We apply the high-level CLIP [37] semantic loss on the paired input text and rendered 2D images to achieve text-3D alignment. To resolve the second challenge, we propose to regularize the color and geometry training with low-level image regularization. This method provides complimentary image-level guidance to our learning objective, which gives fine-grained textures during the generation phase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Notably, we empirically adopt the pretrained GET3D model and train its mapping networks only since we observe this practice not only improves the training efficiency but also stabilizes the training process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.75, + 0.759 + ], + "angle": 0, + "content": "3.3. Pseudo Caption Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Most recent 3D shape generation models are data-driven. With large-scale training data, their performance improvement is remarkable. However, when we aim to train a text-guided 3D shape generation model, we can only obtain a few textual descriptions for limited samples. To resolve this issue, we follow the method of [54] and propose to generate pseudo captions for each 3D object." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Specifically, we observe the textual descriptions for a single 3D object are highly structured, e.g., \"A yellow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.473, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.314 + ], + "angle": 0, + "content": "Figure 3. Our proposed pseudo caption generation module can be split into four steps. Firstly, we construct vocabulary from ShapeNet-related nouns and adjectives in the CLIP vocabulary. Then we retrieve multiple words based on the 2D-rendered images. In the third step, we create candidate captions with retrieved words. Finally, we select one caption based on the computed text-image similarities with the CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.469, + 0.422 + ], + "angle": 0, + "content": "sports car\", which typically consists of adjectives and a noun. Therefore, it is critical to obtain suitable adjectives and nouns for given 3D objects to construct pseudo captions automatically. To this end, we adopt the pretrained text-image similarity model of CLIP [37] to retrieve useful nouns and adjectives and further build complete sentences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.546 + ], + "angle": 0, + "content": "The pseudo caption generation module can be split into four steps, shown in Fig. 3. In the first step, we construct a vocabulary to cover the related adjectives and nouns. Since we aim to generate ShapeNet [5] object classes, we collect all the class names as nouns in the vocabulary, which has 169 tokens. Then we extract all the adjectives from CLIP vocabulary with the Natural Language Toolkit [2]. The size of our collected adjectives is 1463." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.807 + ], + "angle": 0, + "content": "In the second and third steps, we aim to retrieve relevant words from 2D-rendered images and construct candidate captions. Technically, we first extract all the vocabulary word embeddings and rendered image embeddings, where we adopt the pretrained CLIP text and image encoders, respectively. Then for each object, we compute the cosine similarities between the given image embedding and all the text embeddings. It is notable that we retrieve the nouns and adjectives separately, where we select the top-\\(K_{1}\\) retrieved nouns and top-\\(K_{2}\\) retrieved adjectives. With the selected words, we utilize the templates to construct candidate sentences. Since these sentences are only supposed to describe a static object, we use the template \"a{Adjective} {Noun}\", where we put 1 or 2 adjectives randomly to increase the caption diversity. Due to having \\(K_{1}\\) nouns and \\(K_{2}\\) adjectives, we are able to combine them and produce multiple candidate pseudo captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In order to obtain the most suitable captions for given objects, we further conduct image-to-candidate caption retrieval, where we also use CLIP encoders to extract the image and sentence embeddings. To be specific, we use the top retrieved sentences as the final pseudo captions, which are used as the input condition for the 3D generator." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.891, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.331 + ], + "angle": 0, + "content": "Figure 4. In the training phase, we not only conduct the high-level text-to-image similarity learning through the CLIP model but also use low-level image regularization to obtain fine-grained textures and diverse shapes, where we render real and fake images with the same camera poses and apply background augmentation to allow the model focus on object features only." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.337, + 0.755, + 0.354 + ], + "angle": 0, + "content": "3.4. Text-3D Alignment Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.542 + ], + "angle": 0, + "content": "After the pseudo caption generation, we feed these captions into a pretrained GET3D [9] model in an attempt to control the semantic meanings of generated 3D objects. To be specific, we first adopt the CLIP text encoder \\( E_{t} \\) to extract the caption \\( t \\) features, then the caption features \\( E_{t}(t) \\) are served as conditions and fed into the mapping networks of pretrained GET3D models. The input of the original mapping networks is random noise only. To inject the text information, we concatenate the caption embeddings and the random noise together as the new input. This enables us to generate diverse 3D objects that match the text's semantic meanings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.662 + ], + "angle": 0, + "content": "Since we adopt the pretrained 3D generator as our model backbone, which has a good ability to synthesize high-fidelity 3D textured shapes, fixing the generator weights allows our model to focus on getting semantic alignment between input textual descriptions and generated 3D objects and also alleviates the learning difficulty. Our main training objective is to maximize the image-text similarity measured by the CLIP model, which can be denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.675, + 0.891, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l i p}} = 1 - \\cos \\left(E _ {i} \\left(I _ {x}\\right), E _ {t} (t)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.891, + 0.763 + ], + "angle": 0, + "content": "Here \\( E_{i} \\) and \\( I_{x} \\) represent the CLIP image encoder and rendered images, respectively. These 2D images are rendered from the generated 3D textured shapes with randomly sampled camera poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.891, + 0.787 + ], + "angle": 0, + "content": "3.5. Towards Diverse and Fine-Grained Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "It is observed using \\(\\mathcal{L}_{clip}\\) only enables the model to generate plausible 3D shapes that align with the input text. However, the diversity of geometry generation is limited, and the generated colors are unnatural. This is caused by two main reasons. One reason is the diversity of our pseudo captions is also limited, as there are only a few descriptive words for a single object class. Moreover, the CLIP model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16808" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.181 + ], + "angle": 0, + "content": "may wrongly retrieve irrelevant words to the given objects, which also affects the generated shape quality. Another reason is the CLIP model only provides high-level guidance for our training supervision while failing to give fine-grained texture information, thus making the generated results unrealistic." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.468, + 0.258 + ], + "angle": 0, + "content": "In order to resolve this issue, we further introduce a low-level image regularization loss. Specifically, we use the same camera poses as that used in the \\( I_{x} \\) rendering to produce ground truth rendered images \\( I_{x}^{gt} \\). The optimization objective is denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.266, + 0.469, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m g} = 1 - \\cos \\left(E _ {i} \\left(I _ {x}\\right), E _ {i} \\left(I _ {x} ^ {g t}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.458 + ], + "angle": 0, + "content": "Since the rendering process returns an alpha channel, which results in a simple white or black background, if we directly take rendered images with the same backgrounds over training iterations, the background information will also be involved during optimizing Eq. (2). To alleviate this issue and allow the model to focus on the foreground objects, we follow [13] to conduct background augmentation, as shown in Fig. 4. Technically, we randomly select random Fourier textures [28], Gaussian noise, and checkerboard patterns as backgrounds. For each paired real and fake image, we apply the same augmented backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.321, + 0.483 + ], + "angle": 0, + "content": "3.6. Overall Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.49, + 0.417, + 0.505 + ], + "angle": 0, + "content": "Our overall training objective can be depicted as" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.514, + 0.469, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c l i p}} + \\mathcal {L} _ {\\text {i m g}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.735 + ], + "angle": 0, + "content": "During the training phase, we empirically experiment with updating all model parameters and the mapping networks only. We observe training all GET3D backbone parameters would severely distort the generated geometry and textures, where we show the experimental results in the supplementary material. This is because the learning objective is two-pronged: improving the generation quality and matching the input captions. While borrowing a pretrained unconditional generator ensures good generation quality, expedites the training process, and also yields stable text-guided results. Therefore, we regard training the mapping networks only of a pretrained 3D generator as a useful and practical solution for the text-guided 3D shape generation task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.747, + 0.208, + 0.763 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.173, + 0.787 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We train and evaluate our method on the large-scale 3D dataset ShapeNet [5]. Our experiments are conducted on four classes with complex geometry, Car, Table, Chair, and Motorbike, containing 7497, 8436, 6778, and 337 shapes, respectively. We follow GET3D [9] to split the data into training, validation, and test sets in a 7:1:2 ratio. The training image data are rendered following the GET3D setting." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.603, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.089, + 0.681, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.129, + 0.694, + 0.139 + ], + "angle": 0, + "content": "a white automobile a low race car" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.142, + 0.59, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.198, + 0.598, + 0.219 + ], + "angle": 0, + "content": "a stationary chair" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.142, + 0.679, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.198, + 0.688, + 0.219 + ], + "angle": 0, + "content": "a white ladderback chair" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.092, + 0.776, + 0.122 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.13, + 0.762, + 0.15 + ], + "angle": 0, + "content": "a ural motorbike" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.156, + 0.775, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.092, + 0.857, + 0.125 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.129, + 0.855, + 0.15 + ], + "angle": 0, + "content": "a brown motorcycle" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.151, + 0.86, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.203, + 0.871, + 0.213 + ], + "angle": 0, + "content": "a structural table" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.23, + 0.882, + 0.244 + ], + "angle": 0, + "content": "Figure 5. The demonstration of our generated pseudo captions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.305 + ], + "angle": 0, + "content": "We render 24 views randomly for Car, Chair and Table and 100 views for Motorbike since it has less available data. The pretrained models were trained on each class separately." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.326, + 0.714, + 0.342 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.58 + ], + "angle": 0, + "content": "We use the GET3D [9] model pretrained on ShapeNet [5] dataset as our backbone 3D generator. In the pseudo caption generation module, we set \\( K_{1} \\) and \\( K_{2} \\) as 3 and 6, respectively. We generate 20 pseudo captions for each ShapeNet object. During the training phase, each input image is randomly paired with one of the pseudo captions generated for the object which the image was rendered from. An example of the generated pseudo can be found in Fig. 5. We only update the geometry and texture mapping networks. We set the batch size to 16 and ran all the experiments on 4 Nvidia Tesla V100-32G GPUs. We empirically observe the geometry branch is more sensitive to the weight update. Hence we set learning rates for geometry and texture mapping networks as 0.004 and 0.0005, respectively. It costs around 10 hours for model training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.601, + 0.808, + 0.617 + ], + "angle": 0, + "content": "4.3. Comparison with Existing Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.718, + 0.644 + ], + "angle": 0, + "content": "4.3.1 Qualitative comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Fig. 6, we first compare with the method TITG3SG [25], which is an encoder-decoder-based framework. They only experiment with object classes of Chair and Table. Here we conduct the comparison qualitatively. Specifically, we show the generated explicit 3D shapes with the corresponding text prompts. Since TITG3SG [25] generates 3D shapes in two steps, where the coarse voxel and refined mesh are produced separately. We show them both in the top row of Fig. 6. The second row presents our generated mesh in two different views, in which we generate final 3D results in one shot. We observe our method gives better 3D textured shapes than TITG3SG, which is a fully supervised method. For example, in the third generated sample of TITG3SG, the chair legs are distorted. Moreover, the high alignment between input prompts and our generated results demonstrates our text control ability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.844, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Figure 6. We compare the text-guided generation results of our methods with that of the fully supervised method TITG3SG [25] on two object classes, Chair and Table. On the top row, we present the results from [25], for each text prompt, the top left one is the raw voxel output in \\(64^{3}\\) resolution, and the top right one is the refined mesh based on the voxel output. The bottom row shows our mesh output with two different views. The renderings are generated with ChimeraX [10]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.304, + 0.307, + 0.32 + ], + "angle": 0, + "content": "4.3.2 Quantitative comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.388 + ], + "angle": 0, + "content": "To demonstrate our 3D generation quality, we compare existing works [25,43,50] from three aspects: 1) rendered 2D image quality using FID, 2) text-image relevance using R-Precision, and 3) 3D geometry quality using FPD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.471, + 0.569 + ], + "angle": 0, + "content": "In Tab. 1, we provide a quantitative comparison with CLIP-NeRF [50], which utilizes human-written captions for training. We employ the Fréchet Inception Distance (FID) [12] as our evaluation metric and present results for two object classes: Car and Chair. To ensure fair comparisons, we downsample our results to match the resolution of CLIP-NeRF [50]. CLIP-NeRF employs NeRF as the 3D generator, which imposes limitations on the generation resolution. In contrast, our backbone GET3D model boasts a larger capacity, supporting higher resolution. Experimental results demonstrate that our method surpasses CLIP-NeRF in text-guided generation quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.736 + ], + "angle": 0, + "content": "In Tab. 2, we adopt the CLIP-R-Precision metric [32] to evaluate the consistency between input text prompts and generated shapes. Given that our model is category-specific, we perform per-category CLIP-R-Precision and compare it to the fully supervised TITG3SG method [25]. According to [32], we compute the CLIP-R-Precision using a set of unseen random captions with words sampled from existing captions, thereby enhancing correlation with human evaluations and mitigating model bias inherent compared to the conventional R-Precision metric. It is observed our method outperforms TITG3SG [25] in each category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.47, + 0.813 + ], + "angle": 0, + "content": "In Tab. 3, we assess the geometry generation quality using the Fréchet Point Distance (FPD) [19, 47], which measures the feature distance of a PointNet [35] pretrained on ModelNet40 [51]. Our method outperforms existing approaches in terms of geometry quality by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.327, + 0.848 + ], + "angle": 0, + "content": "4.3.3 Inference speed comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Tab. 4, we compare the inference speed of our method with other works. The optimization-based methods DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18]" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.302, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Table 1. Comparison with the existing work. We evaluate the rendered 2D images using Fréchet inception distance (FID). We downsample our result to the same resolution of CLIP-NeRF [50] for fair comparisons." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.358, + 0.875, + 0.443 + ], + "angle": 0, + "content": "
CarChair
ResolutionFIDResolutionFID
CLIP-NeRF [50]256267.8128248.4
Ours256220.1128243.7
Ours1024221.71024244.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Table 2. Per category CLIP-R-Precision score using random generated captions on categories Chair and Table." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.478, + 0.841, + 0.533 + ], + "angle": 0, + "content": "
MethodChairTable
TITG3SG [25]9.05 ± 1.926.67 ± 3.25
Ours20.19 ± 1.8915.21 ± 3.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.539, + 0.875, + 0.553 + ], + "angle": 0, + "content": "Table 3. Comparison of 3D generation quality in FPD score." + }, + { + "type": "table", + "bbox": [ + 0.575, + 0.554, + 0.816, + 0.62 + ], + "angle": 0, + "content": "
MethodChairTable
TITG3SG [25]1566.761639.68
CLIP-Forge [43]825.963051.31
Ours342.231468.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Table 4. The comparison of inference time per text prompt for different methods. The inference time of DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18] are given in their papers. For TITG3SG [25] and ours, we generate 1000 text prompts with batch size as 1, and calculate the average inference time." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.695, + 0.888, + 0.815 + ], + "angle": 0, + "content": "
MethodDeviceOutputTime
DreamFields [13]TPU cores x8Rendering72 min
DreamFusion [34]TPUv4 machineRendering90 min
PureCLIPNeRF [18]GTX 2080tiRendering20 min
TITG3SG [25]Telsa V100-32GVoxel2.21 sec
TITG3SG [25]Telsa V100-32GMesh24.44 sec
OursTelsa V100-32GRendering0.05 sec
OursTelsa V100-32GMesh7.09 sec
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "take tens of minutes to optimize for each input text prompt. Liu et al. [25] use an encoder-decoder architecture and thus can perform inference at 2.21 seconds. However, the raw \\(64^{3}\\) voxel output is very coarse, as shown in Fig. 6. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.109, + 0.174, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.11, + 0.275, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.111, + 0.372, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.111, + 0.459, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.164, + 0.317, + 0.177 + ], + "angle": 0, + "content": "\"a blue SUV\"" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.182, + 0.177, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.187, + 0.274, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.188, + 0.369, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.189, + 0.456, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.241, + 0.334, + 0.254 + ], + "angle": 0, + "content": "\"a red hatchback\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.271, + 0.175, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.27, + 0.266, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.271, + 0.366, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.272, + 0.453, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.324, + 0.347, + 0.338 + ], + "angle": 0, + "content": "\"a yellow racing car\"" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.344, + 0.162, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.348, + 0.256, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.346, + 0.357, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.346, + 0.443, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.402, + 0.374, + 0.416 + ], + "angle": 0, + "content": "\"a yellow racing motorcycle\"" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.422, + 0.164, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.422, + 0.257, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.423, + 0.355, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.417, + 0.446, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.473, + 0.324, + 0.486 + ], + "angle": 0, + "content": "\"a red motorbike\"" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.105, + 0.567, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.105, + 0.655, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.106, + 0.746, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.179, + 0.564, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.179, + 0.653, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.179, + 0.746, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.242, + 0.736, + 0.256 + ], + "angle": 0, + "content": "\"agrayarmlesschair\"" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.261, + 0.551, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.261, + 0.653, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.261, + 0.747, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.108, + 0.849, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.18, + 0.841, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.327, + 0.719, + 0.34 + ], + "angle": 0, + "content": "\"agreenchair\"" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.343, + 0.562, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.346, + 0.657, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.346, + 0.751, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.261, + 0.85, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.346, + 0.851, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.401, + 0.736, + 0.415 + ], + "angle": 0, + "content": "\"a red dining table\"" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.419, + 0.564, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.42, + 0.658, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.42, + 0.755, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.421, + 0.851, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.47, + 0.748, + 0.483 + ], + "angle": 0, + "content": "\"a silver metallic table\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.893, + 0.528 + ], + "angle": 0, + "content": "Figure 7. We show the generation diversity and text control of our method. Each row takes the same text prompt with different sampled random noise as the model input." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.542, + 0.165, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.544, + 0.236, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.544, + 0.307, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.543, + 0.378, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.543, + 0.452, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.584, + 0.157, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.585, + 0.225, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.585, + 0.292, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.584, + 0.368, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.585, + 0.439, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.648, + 0.163, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.647, + 0.236, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.647, + 0.307, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.647, + 0.379, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.647, + 0.451, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.687, + 0.144, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.687, + 0.216, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.687, + 0.291, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.687, + 0.362, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.687, + 0.437, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.724, + 0.152, + 0.734 + ], + "angle": 0, + "content": "\"ared{class}\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.724, + 0.229, + 0.734 + ], + "angle": 0, + "content": "\"a yellow {class}\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.724, + 0.306, + 0.734 + ], + "angle": 0, + "content": "\"agreen{class}" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.724, + 0.382, + 0.734 + ], + "angle": 0, + "content": "\"a blue {class}\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.724, + 0.456, + 0.734 + ], + "angle": 0, + "content": "\"a gray {class}\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.742, + 0.468, + 0.77 + ], + "angle": 0, + "content": "Figure 8. Text control of textures on four different categories. Each row shares the same sampled random noise input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.9 + ], + "angle": 0, + "content": "coarse output is later refined with interpolations and marching cubes to generate mesh data. This process takes ten times longer than the network inference time and costs 24.44 seconds per sample. Our method generates high-resolution renderings at \\(1024^2\\) resolution with only 0.05 seconds and generates the complete explicit mesh representation in 7.09 seconds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.682, + 0.551 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.903 + ], + "angle": 0, + "content": "In Fig. 7, we present the generation diversity and text control strength of our method in four object classes. In each row, we sample different random noises as well as the given text prompt as input to generate diversified and text-controlled 3D shapes, where the input text features are simultaneously fed into both the geometry and texture mapping network. Generally, we observe the semantic meanings of the rendered 2D images, and the given prompts are consistent. Moreover, there exists large diversity in the texture and geometry of our generated 3D shapes, even from the same text input. Here we present generated results of 4 classes, where \\( Car \\) has a relatively simple geometry, while Table and Chair are more complex. But we still have good control over the textures and geometries for the challenging classes. For example, when we input \"a gray armless chair\", the generated samples exactly match the given instructions. However, we also observe failure cases of Table generation. For instance, the generation of \"a wooden office desk\" may have distorted geometry due to its complex structure. In Fig. 8, we show the text control ability of our method on different classes. The results in each row are generated from the same random noise input but with dif" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16811" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.13 + ], + "angle": 0, + "content": "Table 5. We compare the FID score of using image CLIP features (Lafite [55]) and the pseudo text CLIP features (Ours) as training input." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.131, + 0.432, + 0.188 + ], + "angle": 0, + "content": "
MethodCarChairTableMotorbike
Lafite [55]68.781.194.3146.5
Ours21.744.843.257.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.198, + 0.468, + 0.226 + ], + "angle": 0, + "content": "Table 6. We compare the FID score for our method without \\(\\mathcal{L}_{clip}\\), \\(\\mathcal{L}_{img}\\) and background augmentation (BGaug)." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.226, + 0.44, + 0.309 + ], + "angle": 0, + "content": "
MethodCarChairTableMotorbike
Ours21.744.843.257.2
-w/o Lclip83.7148.967.1134.9
-w/o Limg59.8144.248.4101.8
-w/o BGaug55.165.158.1110.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.364 + ], + "angle": 0, + "content": "ferent text prompts. To summarize, the results show that our generation results are highly consistent with the input text prompts across different object classes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.23, + 0.392 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Pseudo caption generation module. In Tab. 5, we compare our method with Lafite [55], which is a language-free training method for the text-to-image generation task. Lafite claims due to the alignment between CLIP text and image embeddings. The CLIP image features can be directly used as text input during training. To implement Lafite [55] onto our framework, we feed the rendered real image CLIP features into our model and train the model by minimizing the cosine similarity between the generated and input image CLIP features. The quantitative results show that our proposed pseudo caption generation module significantly improves the generation quality compared to the baseline. We also give a qualitative comparison in Fig. 9 (b), we observe that the generator trained with Lafite [55] fails to generate correct textures. This problem might come from the modality gap between the CLIP image and text embeddings, as discussed in [20, 54]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.468, + 0.852 + ], + "angle": 0, + "content": "Effects of our loss functions. We conduct an ablation study on our used two losses \\(\\mathcal{L}_{clip}\\) and \\(\\mathcal{L}_{img}\\). In Tab. 6, we quantitatively validate the usefulness of both two losses. Specifically, \\(\\mathcal{L}_{clip}\\) supervises the text-image alignment learning, which is more critical than \\(\\mathcal{L}_{img}\\), and removing \\(\\mathcal{L}_{clip}\\) gives a more considerable performance drop. From the qualitative results in Fig. 9 (c), we observe that simply using \\(\\mathcal{L}_{img}\\) leads to the model generating distorted geometries, which may indicate the significance of high-level semantic supervision. We also observe that solely using \\(\\mathcal{L}_{clip}\\) (Fig. 9 (d)) may cause the model generation to lose fine-grained texture details, as well as the geometry information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Effects of background augmentation. Our used GET3D [9] backbone network generates RGBA images with dark backgrounds, which may bring some noise when" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.105, + 0.535, + 0.117 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.093, + 0.874, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Figure 9. Ablation study on our methods with text prompt \"a red car\". We show results of (a) our full model, (b) Lafite [55], (c) without high-level semantic regularization \\(\\mathcal{L}_{clip}\\), (d) without low-level image regularization \\(\\mathcal{L}_{img}\\), (e) without background augmentation. All the visualizations are generated from models trained for 250k iterations." + }, + { + "type": "image_footnote", + "bbox": [ + 0.533, + 0.405, + 0.667, + 0.428 + ], + "angle": 0, + "content": "\"acarwithyellowlegs, redseat and blue back.\"" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.383, + 0.868, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 10. Examples of failure cases of our method. Our model often fails to handle text input with composed descriptions for different parts of the object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.567 + ], + "angle": 0, + "content": "we conduct low-level image regularization. As illustrated from the quantitative results in Tab. 6 and the visualizations in Fig. 9 (e), training with the augmented background can help remove the texture flaws." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.619, + 0.595 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.697 + ], + "angle": 0, + "content": "We introduce TAPS3D, a text-guided 3D textured shape generation method. We first generate pseudo captions for model training. During the inference phase, TAPS3D does not require additional optimization. Our fast generation speed enables average users to generate high-quality 3D objects with acceptable processing time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.787 + ], + "angle": 0, + "content": "**Limitation.** The limitation of our method mainly lies in the capacity of our model is still bound to the training image data, and we cannot produce different fine-grained details for different object parts, as shown in Fig. 10. Although we do not require paired text data, we still need diversified training images to handle complex text input." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.651, + 0.817 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research is partly supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-RP-2018-003), and the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and MoE AcRF Tier 1 grants (RG14/22, RG95/20)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16812" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.211 + ], + "angle": 0, + "content": "[2] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc., 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.632 + ], + "angle": 0, + "content": "[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. arXiv preprint arXiv:2207.09446, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv preprint arXiv:2209.11163, 2022. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Thomas D Goddard, Conrad C Huang, Elaine C Meng, Eric F Pettersen, Gregory S Couch, John H Morris, and Thomas E Ferrin. Ucsf chimerax: Meeting modern challenges in visualization and analysis. Protein Science, 27(1):14-25, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylererf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[17] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 895–904, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[18] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[19] Ruihui Li, Xianzhi Li, Ka-Hei Hui, and Chi-Wing Fu. Spagan: Sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[21] Kangcheng Liu. Rm3d: Robust data-efficient 3d scene parsing via traditional and learnt 3d descriptors-based semantic region merging. International Journal of Computer Vision, pages 1-30, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Kangcheng Liu, Zhi Gao, Feng Lin, and Ben M Chen. Fgnet: A fast and accurate framework for large-scale lidar point cloud understanding. IEEE Transactions on Cybernetics, 53(1):553-564, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[23] Kangcheng Liu, Aoran Xiao, Xiaoqin Zhang, Shijian Lu, and Ling Shao. Fac: 3d representation learning via foreground aware feature contrast. arXiv preprint arXiv:2303.06388, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[24] Kangcheng Liu, Yuzhi Zhao, Qiang Nie, Zhi Gao, and Ben M Chen. Weakly supervised 3d scene segmentation with region-level boundary awareness and instance discrimination. In European Conference on Computer Vision 2022 (ECCV 2022), pages 37-55. Springer, Cham, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[28] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 3(7):e12, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[31] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[32] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[36] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[40] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[42] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[44] Aditya Sanghi, Rao Fu, Vivian Liu, Karl Willis, Hooman Shayani, Amir Hosein Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Textcraft: Zero-shot generation of high-fidelity and diverse shapes from text. arXiv preprint arXiv:2211.01427, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.552, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[45] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[47] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[48] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. 3d pose transfer with correspondence learning and mesh refinement. Advances in Neural Information Processing Systems, 34:3108-3120, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[49] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. Unsupervised 3d pose transfer with cross consistency and dual reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[51] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[54] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[55] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16815" + } + ] +] \ No newline at end of file diff --git a/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_origin.pdf b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..91c21a86d39b99cf7ccf978e702df20958b1286b --- /dev/null +++ b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/cb0b5b28-71a3-4466-8edb-4420e1a78920_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c9e0115d6bc005e2c47be4daa09c8fee23219a76d945a7e167beb4ec9535feb +size 6921747 diff --git a/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/full.md b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e0095a70b76ea2a460869a2c042a7ae575302f3e --- /dev/null +++ b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/full.md @@ -0,0 +1,461 @@ +# TAPS3D: Text-Guided 3D Textured Shape Generation from Pseudo Supervision + +Jiacheng Wei $^{1*}$ Hao Wang $^{1*}$ Jiashi Feng $^{2}$ Guosheng Lin $^{1\dagger}$ Kim-Hui Yap $^{1}$ $^{1}$ Nanyang Technological University, Singapore $^{2}$ ByteDance +{jiacheng002@e., hao005@e., gslin@, ekhyap@}ntu.edu.sg, jshfeng@bytedance.com + +![](images/fe5a857b9162dd7165cba42c3288eb459803ae27b0cd1d10664bb92d25f41204.jpg) +"a red hatchback car" + +![](images/21f46766d115262baf96255299c52ae29b897dfbe28841a78771350ea37f33f5.jpg) +Figure 1. Our proposed model is capable of generating detailed textured shapes based on given text prompts. We demonstrate the effectiveness of our approach by showcasing generated results for four distinct object classes: Car, Chair, Table, and Motorbike. Both the textured meshes and geometries are presented, with visualizations rendered using ChimeraX [10]. + +![](images/d52a4bf3050eede9191cd8ed4d3f1bc0081fb8233fb32bff1575993e28c2dfdc.jpg) +"agrayarmlesschair + +![](images/7fbc3400b84d242098e2dd16ac61d1a9b2fac285a4f836a490a1670cc65e91b8.jpg) +"awooden office table" + +![](images/665635970e9f3d95c6c1be52840257be7b7a2c58394773e5225ea36c6c1fa4eb.jpg) + +![](images/c5b238b9babd634c15d7251ce285259ebc330cdf936715f53b1a4c24a60f8f11.jpg) +"ablock motorbike" + +![](images/e052b58908dfaf5125946af3eb2a77b3c359ec9bf9276b50f80eef9c281e2230.jpg) + +# Abstract + +In this paper, we investigate an open research task of generating controllable 3D textured shapes from the given textual descriptions. Previous works either require ground truth caption labeling or extensive optimization time. To resolve these issues, we present a novel framework, TAPS3D, to train a text-guided 3D shape generator with pseudo captions. Specifically, based on rendered 2D images, we retrieve relevant words from the CLIP vocabulary and construct pseudo captions using templates. Our constructed captions provide high-level semantic supervision for generated 3D shapes. Further, in order to produce fine-grained textures and increase geometry diversity, we propose to adopt low-level image regularization to enable fake-rendered images to align with the real ones. During the inference phase, our proposed model can generate 3D textured shapes from the given text without any additional optimization. We conduct extensive experiments to analyze each of our proposed components and show the efficacy of our framework in generating high-fidelity 3D textured and text-relevant shapes. Code is available at https://github.com/plusmultiply/TAPS3D + +# 1. Introduction + +3D objects are essential in various applications [21-24], such as video games, film special effects, and virtual real + +ity. However, realistic and detailed 3D object models are usually hand-crafted by well-trained artists and engineers slowly and tediously. To expedite this process, many research works [3,9,13,34,48,49] use deep generative models to achieve automatic 3D object generation. However, these models are primarily unconditioned, which can hardly generate objects as humans will. + +In order to control the generated 3D objects from text, prior text-to-3D generation works [43, 44] leverage the pretrained vision-language alignment model CLIP [37], such that they can only use 3D shape data to achieve zero-shot learning. For example, Dream Fields [13] combines the advantages of CLIP and NeRF [27], which can produce both 3D representations and renderings. However, Dream Fields costs about 70 minutes on 8 TPU cores to produce a single result. This means the optimization time during the inference phase is too slow to use in practice. Later on, GET3D [9] is proposed with faster inference time, which incorporates StyleGAN [15] and Deep Marching Tetrahedral (DMTet) [46] as the texture and geometry generators respectively. Since GET3D adopts a pretrained model to do text-guided synthesis, they can finish optimization in less time than Dream Fields. But the requirement of test-time optimization still limits its application scenarios. CLIP-NeRF [50] utilizes conditional radiance fields [45] to avoid test-time optimization, but it requires ground truth text data for the training purpose. Therefore, CLIP-NeRF is only applicable to a few object classes that have labeled text data for training, and its generation quality is restricted by the NeRF capacity. + +To address the aforementioned limitations, we propose + +to generate pseudo captions for 3D shape data based on their rendered 2D images and construct a large amount of $\langle 3\mathrm{D}$ shape, pseudo captions $\rangle$ as training data, such that the text-guided 3D generation model can be trained over them. To this end, we propose a novel framework for Text-guided 3D textured shApe generation from Pseudo Supervision (TAPS3D), in which we can generate high-quality 3D shapes without requiring annotated text training data or test-time optimization. + +Specifically, our proposed framework is composed of two modules, where the first generates pseudo captions for 3D shapes and feeds them into a 3D generator to conduct text-guided training within the second module. In the pseudo caption generation module, we follow the language-free text-to-image learning scheme [20, 54]. We first adopt the CLIP model to retrieve relevant words from given rendered images. Then we construct multiple candidate sentences based on the retrieved words and pick sentences having the highest CLIP similarity scores with the given images. The selected sentences are used as our pseudo captions for each 3D shape sample. + +Following the notable progress of text-to-image generation models [29, 38, 39, 42, 52], we use text-conditioned GAN architecture in the text-guided 3D generator training part. We adopt the pretrained GET3D [9] model as our backbone network since it has been demonstrated to generate high-fidelity 3D textured shapes across various object classes. We input the pseudo captions as the generator conditions and supervise the training process with high-level CLIP supervision in an attempt to control the generated 3D shapes. Moreover, we introduce a low-level image regularization loss to produce fine-grained textures and increase geometry diversity. We empirically train the mapping networks only of a pretrained GET3D model so that the training is stable and fast, and also, the generation quality of the pretrained model can be preserved. + +Our proposed model TAPS3D can produce high-quality 3D textured shapes with strong text control as shown in Fig. 1, without any per prompt test-time optimization. Our contribution can be summarized as: + +- We introduce a new 3D textured shape generative framework, which can generate high-quality and fidelity 3D shapes without requiring paired text and 3D shape training data. +- We propose a simple pseudo caption generation method that enables text-conditioned 3D generator training, such that the model can generate text-controlled 3D textured shapes without test time optimization, and significantly reduce the time cost. +- We introduce a low-level image regularization loss on top of the high-level CLIP loss in an attempt to produce fine-grained textures and increase geometry diversity. + +# 2. Related Work + +# 2.1. Text-Guided 3D Shape Generation + +Text-guided 3D shape generation aims to generate 3D shapes from the textual descriptions so that the generation process can be controlled. There are mainly two categories of methods, i.e., fully-supervised and optimization-based methods. The fully-supervised method [6,8,25] uses ground truth text and the paired 3D objects with explicit 3D representations as training data. Specifically, CLIP-Forge [43] uses a two-stage training scheme, which consists of shape autoencoder training, and conditional normalizing flow training. VQ-VAE [44] performs zero-shot training with 3D voxel data by utilizing the pretrained CLIP model [37]. + +Regarding the optimization-based methods [13, 18, 26, 34], Neural Radiance Fields (NeRF) are usually adopted as the 3D generator. To generate 3D shapes for each input text prompt, they use the CLIP model to supervise the semantic alignment between rendered images and text prompts. Since NeRF suffers from extensive generation time, 3D-aware image synthesis [3, 3, 4, 11, 30, 31] has become popular, which generates multi-view consistent images by integrating neural rendering in the Generative Adversarial Networks (GANs). Specifically, there are no explicit 3D shapes generated during the process, while the 3D shapes can be extracted from the implicit representations, such as the occupancy field or signed distance function (SDF), using the marching cube algorithm. These optimization-based methods provide a solution to generate 3D shapes, but their generation speed was compromised. Although [50] is free from test-time optimization, it requires text data for training, which limits its applicable 3D object classes. Our proposed method attempts to alleviate both the paired text data shortage issues and the long optimization time in the previous work. We finally produce high-quality 3D shapes to bring our method to practical applications. + +# 2.2. Text-to-Image Synthesis + +We may draw inspiration from text-to-image generation methods. Typically, many research works [36,40,53] adopt the conditional GAN architecture, where they directly take the text features and concatenate them with the random noise as the input. Recently, autoregressive models [7, 39] and diffusion models [29, 38, 41, 42] made great improvement on text to image synthesis while demanding huge computational resources and massive training data. + +With the introduction of StyleGAN [14-16] mapping networks, the input random noise can be first mapped to another latent space that has disentangled semantics, then the model can generate images with better quality. Further, exploring the latent space of StyleGAN has been proved useful by several works [1, 17, 33] in text-driven image synthe + +![](images/08cd0e04469de448a7fa81f72149eb38221d57c37e541e9b0ea5c09fac8c2413.jpg) +Figure 2. Our proposed framework supports text-guided 3D shape generation training without paired text labeling. To generate captions for training, we first adopt a pseudo caption generation module to produce textual descriptions, given rendered 2D images. Then we feed the pseudo caption CLIP embeddings into the mapping networks to train the controllable text-guided 3D shape synthesis network. Our training pipeline is supervised by high-level semantic loss $\mathcal{L}_{clip}$ and low-level image regularization $\mathcal{L}_{img}$ . It is notable that during the training phase, only the mapping network weights are updated, and the rest components are fixed. + +sis and manipulation tasks, where they utilize the pretrained vision-language model CLIP [37] model to manipulate pretrained unconditional StyleGAN networks. In order to relieve the need for paired text data during the training phase, Lafite [55] proposes to adopt the image CLIP embeddings as training input while using text CLIP embedding during the inference phase. However, as discussed in the recent works [20, 54], there exists a modality gap between text and image embeddings in the CLIP latent space. Therefore, language-free methods that are solely trained on CLIP image embedding could introduce distribution mismatch when encountering the CLIP text embedding input during inference. In this paper, we further explore language-free training for text-guided 3D textured shape generation by proposing a pseudo caption generation module. + +# 3. Method + +# 3.1. Preliminary + +GET3D [9] is a newly proposed 3D generator that can generate high-fidelity textured 3D shapes. Specifically, GET3D maps noise vectors $\mathbf{z} \in \mathcal{N}(0, \mathbf{I})$ to a textured mesh. In order to disentangle the geometry and texture information, GET3D takes two random vectors $\mathbf{z}_1$ and $\mathbf{z}_2$ as inputs for the geometry and texture generation branches, respectively. Following the design of StyleGAN [15], and EG3D [3], they map $\mathbf{z}_1$ and $\mathbf{z}_2$ to the intermediate latent codes $\mathbf{w}_1$ and $\mathbf{w}_2$ with two mapping networks since the semantics would be better represented in the mapped space. The geometry and texture generators are built conditioned on $\mathbf{w}_1$ and $\mathbf{w}_2$ . + +# 3.2. Overview + +Since the GET3D [9] model only supports unconditional 3D generation, if one need to do text-guided 3D shape syn + +thesis, additional optimization is required with the original GET3D model. In order to control the semantics of the generated results and achieve text-based textured 3D generation in one shot, we propose to add text conditions to GET3D models. In Fig. 2, we present our training framework. + +To this end, we address two main challenges: one is the lack of training textual descriptions, and another is color and geometry distortion. Regarding the first challenge, we propose a pseudo caption generation module, which gives large amounts of captions for training purposes. These captions are used as conditions and fed into the mapping networks to control the semantic meanings of the output textured 3D shapes. We apply the high-level CLIP [37] semantic loss on the paired input text and rendered 2D images to achieve text-3D alignment. To resolve the second challenge, we propose to regularize the color and geometry training with low-level image regularization. This method provides complimentary image-level guidance to our learning objective, which gives fine-grained textures during the generation phase. + +Notably, we empirically adopt the pretrained GET3D model and train its mapping networks only since we observe this practice not only improves the training efficiency but also stabilizes the training process. + +# 3.3. Pseudo Caption Generation + +Most recent 3D shape generation models are data-driven. With large-scale training data, their performance improvement is remarkable. However, when we aim to train a text-guided 3D shape generation model, we can only obtain a few textual descriptions for limited samples. To resolve this issue, we follow the method of [54] and propose to generate pseudo captions for each 3D object. + +Specifically, we observe the textual descriptions for a single 3D object are highly structured, e.g., "A yellow + +![](images/7cc311faf542181fb0e5ac97ac8033034c713890872e9c45f089f39fd4023e5e.jpg) +Figure 3. Our proposed pseudo caption generation module can be split into four steps. Firstly, we construct vocabulary from ShapeNet-related nouns and adjectives in the CLIP vocabulary. Then we retrieve multiple words based on the 2D-rendered images. In the third step, we create candidate captions with retrieved words. Finally, we select one caption based on the computed text-image similarities with the CLIP model. + +sports car", which typically consists of adjectives and a noun. Therefore, it is critical to obtain suitable adjectives and nouns for given 3D objects to construct pseudo captions automatically. To this end, we adopt the pretrained text-image similarity model of CLIP [37] to retrieve useful nouns and adjectives and further build complete sentences. + +The pseudo caption generation module can be split into four steps, shown in Fig. 3. In the first step, we construct a vocabulary to cover the related adjectives and nouns. Since we aim to generate ShapeNet [5] object classes, we collect all the class names as nouns in the vocabulary, which has 169 tokens. Then we extract all the adjectives from CLIP vocabulary with the Natural Language Toolkit [2]. The size of our collected adjectives is 1463. + +In the second and third steps, we aim to retrieve relevant words from 2D-rendered images and construct candidate captions. Technically, we first extract all the vocabulary word embeddings and rendered image embeddings, where we adopt the pretrained CLIP text and image encoders, respectively. Then for each object, we compute the cosine similarities between the given image embedding and all the text embeddings. It is notable that we retrieve the nouns and adjectives separately, where we select the top- $K_{1}$ retrieved nouns and top- $K_{2}$ retrieved adjectives. With the selected words, we utilize the templates to construct candidate sentences. Since these sentences are only supposed to describe a static object, we use the template "a{Adjective} {Noun}", where we put 1 or 2 adjectives randomly to increase the caption diversity. Due to having $K_{1}$ nouns and $K_{2}$ adjectives, we are able to combine them and produce multiple candidate pseudo captions. + +In order to obtain the most suitable captions for given objects, we further conduct image-to-candidate caption retrieval, where we also use CLIP encoders to extract the image and sentence embeddings. To be specific, we use the top retrieved sentences as the final pseudo captions, which are used as the input condition for the 3D generator. + +![](images/72c57836373d24beb7efbfac9109fb78bf4b5a96d827177b3aecb4158bbffdd8.jpg) +Figure 4. In the training phase, we not only conduct the high-level text-to-image similarity learning through the CLIP model but also use low-level image regularization to obtain fine-grained textures and diverse shapes, where we render real and fake images with the same camera poses and apply background augmentation to allow the model focus on object features only. + +# 3.4. Text-3D Alignment Training + +After the pseudo caption generation, we feed these captions into a pretrained GET3D [9] model in an attempt to control the semantic meanings of generated 3D objects. To be specific, we first adopt the CLIP text encoder $E_{t}$ to extract the caption $t$ features, then the caption features $E_{t}(t)$ are served as conditions and fed into the mapping networks of pretrained GET3D models. The input of the original mapping networks is random noise only. To inject the text information, we concatenate the caption embeddings and the random noise together as the new input. This enables us to generate diverse 3D objects that match the text's semantic meanings. + +Since we adopt the pretrained 3D generator as our model backbone, which has a good ability to synthesize high-fidelity 3D textured shapes, fixing the generator weights allows our model to focus on getting semantic alignment between input textual descriptions and generated 3D objects and also alleviates the learning difficulty. Our main training objective is to maximize the image-text similarity measured by the CLIP model, which can be denoted as: + +$$ +\mathcal {L} _ {\text {c l i p}} = 1 - \cos \left(E _ {i} \left(I _ {x}\right), E _ {t} (t)\right). \tag {1} +$$ + +Here $E_{i}$ and $I_{x}$ represent the CLIP image encoder and rendered images, respectively. These 2D images are rendered from the generated 3D textured shapes with randomly sampled camera poses. + +# 3.5. Towards Diverse and Fine-Grained Generation + +It is observed using $\mathcal{L}_{clip}$ only enables the model to generate plausible 3D shapes that align with the input text. However, the diversity of geometry generation is limited, and the generated colors are unnatural. This is caused by two main reasons. One reason is the diversity of our pseudo captions is also limited, as there are only a few descriptive words for a single object class. Moreover, the CLIP model + +may wrongly retrieve irrelevant words to the given objects, which also affects the generated shape quality. Another reason is the CLIP model only provides high-level guidance for our training supervision while failing to give fine-grained texture information, thus making the generated results unrealistic. + +In order to resolve this issue, we further introduce a low-level image regularization loss. Specifically, we use the same camera poses as that used in the $I_{x}$ rendering to produce ground truth rendered images $I_{x}^{gt}$ . The optimization objective is denoted as + +$$ +\mathcal {L} _ {i m g} = 1 - \cos \left(E _ {i} \left(I _ {x}\right), E _ {i} \left(I _ {x} ^ {g t}\right)\right). \tag {2} +$$ + +Since the rendering process returns an alpha channel, which results in a simple white or black background, if we directly take rendered images with the same backgrounds over training iterations, the background information will also be involved during optimizing Eq. (2). To alleviate this issue and allow the model to focus on the foreground objects, we follow [13] to conduct background augmentation, as shown in Fig. 4. Technically, we randomly select random Fourier textures [28], Gaussian noise, and checkerboard patterns as backgrounds. For each paired real and fake image, we apply the same augmented backgrounds. + +# 3.6. Overall Training Objective + +Our overall training objective can be depicted as + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c l i p}} + \mathcal {L} _ {\text {i m g}}. \tag {3} +$$ + +During the training phase, we empirically experiment with updating all model parameters and the mapping networks only. We observe training all GET3D backbone parameters would severely distort the generated geometry and textures, where we show the experimental results in the supplementary material. This is because the learning objective is two-pronged: improving the generation quality and matching the input captions. While borrowing a pretrained unconditional generator ensures good generation quality, expedites the training process, and also yields stable text-guided results. Therefore, we regard training the mapping networks only of a pretrained 3D generator as a useful and practical solution for the text-guided 3D shape generation task. + +# 4. Experiments + +# 4.1. Dataset + +We train and evaluate our method on the large-scale 3D dataset ShapeNet [5]. Our experiments are conducted on four classes with complex geometry, Car, Table, Chair, and Motorbike, containing 7497, 8436, 6778, and 337 shapes, respectively. We follow GET3D [9] to split the data into training, validation, and test sets in a 7:1:2 ratio. The training image data are rendered following the GET3D setting. + +![](images/8d0298588e05126143c53974f0caaadb948f5b5028ed38b290d618f44048f526.jpg) + +![](images/5ea777085ae903bc12528e00f1adfd35e2bce12d7aacca7e5aa94ff4efdc5186.jpg) +a white automobile a low race car + +![](images/4202d1f167d4ad5b76527e3cb0b2f603e665b049476448befe80fd14668bc3fc.jpg) +a stationary chair + +![](images/2f5c112a5191829020810077f7cebeef6f5fbf989993a15222f5c19fe4fdb18b.jpg) +a white ladderback chair + +![](images/6f6ac98f4cafae482ef8f6b63e6cb2198b95e2df6871bc81c820fc9276fac14a.jpg) + +![](images/9b2151929695bbb231861c8350195ef66e524e09f6cb55852f1f312dc7cf2ec2.jpg) +a ural motorbike + +![](images/dd2166693acb868fede55b4cbabd744a30ef4e28a3ad8909030ab71d5bc685f0.jpg) + +![](images/3f09374256299dd939c36cb47a2e9bafc431f706a08435b21ebe92e4bd7ffde6.jpg) +a brown motorcycle +a structural table +Figure 5. The demonstration of our generated pseudo captions. + +We render 24 views randomly for Car, Chair and Table and 100 views for Motorbike since it has less available data. The pretrained models were trained on each class separately. + +# 4.2. Implementation details + +We use the GET3D [9] model pretrained on ShapeNet [5] dataset as our backbone 3D generator. In the pseudo caption generation module, we set $K_{1}$ and $K_{2}$ as 3 and 6, respectively. We generate 20 pseudo captions for each ShapeNet object. During the training phase, each input image is randomly paired with one of the pseudo captions generated for the object which the image was rendered from. An example of the generated pseudo can be found in Fig. 5. We only update the geometry and texture mapping networks. We set the batch size to 16 and ran all the experiments on 4 Nvidia Tesla V100-32G GPUs. We empirically observe the geometry branch is more sensitive to the weight update. Hence we set learning rates for geometry and texture mapping networks as 0.004 and 0.0005, respectively. It costs around 10 hours for model training. + +# 4.3. Comparison with Existing Methods + +# 4.3.1 Qualitative comparison + +In Fig. 6, we first compare with the method TITG3SG [25], which is an encoder-decoder-based framework. They only experiment with object classes of Chair and Table. Here we conduct the comparison qualitatively. Specifically, we show the generated explicit 3D shapes with the corresponding text prompts. Since TITG3SG [25] generates 3D shapes in two steps, where the coarse voxel and refined mesh are produced separately. We show them both in the top row of Fig. 6. The second row presents our generated mesh in two different views, in which we generate final 3D results in one shot. We observe our method gives better 3D textured shapes than TITG3SG, which is a fully supervised method. For example, in the third generated sample of TITG3SG, the chair legs are distorted. Moreover, the high alignment between input prompts and our generated results demonstrates our text control ability. + +![](images/a26605101e82a9003bb56173ac0c12a9e6fdf5eab9071f71e5942fea2dead3b7.jpg) +Figure 6. We compare the text-guided generation results of our methods with that of the fully supervised method TITG3SG [25] on two object classes, Chair and Table. On the top row, we present the results from [25], for each text prompt, the top left one is the raw voxel output in $64^{3}$ resolution, and the top right one is the refined mesh based on the voxel output. The bottom row shows our mesh output with two different views. The renderings are generated with ChimeraX [10]. + +# 4.3.2 Quantitative comparison + +To demonstrate our 3D generation quality, we compare existing works [25,43,50] from three aspects: 1) rendered 2D image quality using FID, 2) text-image relevance using R-Precision, and 3) 3D geometry quality using FPD. + +In Tab. 1, we provide a quantitative comparison with CLIP-NeRF [50], which utilizes human-written captions for training. We employ the Fréchet Inception Distance (FID) [12] as our evaluation metric and present results for two object classes: Car and Chair. To ensure fair comparisons, we downsample our results to match the resolution of CLIP-NeRF [50]. CLIP-NeRF employs NeRF as the 3D generator, which imposes limitations on the generation resolution. In contrast, our backbone GET3D model boasts a larger capacity, supporting higher resolution. Experimental results demonstrate that our method surpasses CLIP-NeRF in text-guided generation quality. + +In Tab. 2, we adopt the CLIP-R-Precision metric [32] to evaluate the consistency between input text prompts and generated shapes. Given that our model is category-specific, we perform per-category CLIP-R-Precision and compare it to the fully supervised TITG3SG method [25]. According to [32], we compute the CLIP-R-Precision using a set of unseen random captions with words sampled from existing captions, thereby enhancing correlation with human evaluations and mitigating model bias inherent compared to the conventional R-Precision metric. It is observed our method outperforms TITG3SG [25] in each category. + +In Tab. 3, we assess the geometry generation quality using the Fréchet Point Distance (FPD) [19, 47], which measures the feature distance of a PointNet [35] pretrained on ModelNet40 [51]. Our method outperforms existing approaches in terms of geometry quality by a large margin. + +# 4.3.3 Inference speed comparison + +In Tab. 4, we compare the inference speed of our method with other works. The optimization-based methods DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18] + +Table 1. Comparison with the existing work. We evaluate the rendered 2D images using Fréchet inception distance (FID). We downsample our result to the same resolution of CLIP-NeRF [50] for fair comparisons. + +
CarChair
ResolutionFIDResolutionFID
CLIP-NeRF [50]256267.8128248.4
Ours256220.1128243.7
Ours1024221.71024244.8
+ +Table 2. Per category CLIP-R-Precision score using random generated captions on categories Chair and Table. + +
MethodChairTable
TITG3SG [25]9.05 ± 1.926.67 ± 3.25
Ours20.19 ± 1.8915.21 ± 3.15
+ +Table 3. Comparison of 3D generation quality in FPD score. + +
MethodChairTable
TITG3SG [25]1566.761639.68
CLIP-Forge [43]825.963051.31
Ours342.231468.43
+ +Table 4. The comparison of inference time per text prompt for different methods. The inference time of DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18] are given in their papers. For TITG3SG [25] and ours, we generate 1000 text prompts with batch size as 1, and calculate the average inference time. + +
MethodDeviceOutputTime
DreamFields [13]TPU cores x8Rendering72 min
DreamFusion [34]TPUv4 machineRendering90 min
PureCLIPNeRF [18]GTX 2080tiRendering20 min
TITG3SG [25]Telsa V100-32GVoxel2.21 sec
TITG3SG [25]Telsa V100-32GMesh24.44 sec
OursTelsa V100-32GRendering0.05 sec
OursTelsa V100-32GMesh7.09 sec
+ +take tens of minutes to optimize for each input text prompt. Liu et al. [25] use an encoder-decoder architecture and thus can perform inference at 2.21 seconds. However, the raw $64^{3}$ voxel output is very coarse, as shown in Fig. 6. The + +![](images/3a1b23870babc5dc897131f96736b6ca1f9857507acdcea6c3ffac4a67492bc0.jpg) + +![](images/214e06bdaeed14a83587464862d93025b28c2dcac8e6709ba5e6ee9ee04b19b3.jpg) +"a blue SUV" + +![](images/edff6ab99cc5e40d1106ac6e4d0994c66561205906593d45b1f432e024e26ab2.jpg) + +![](images/58ccc99d37b1dca0382cd1ba4dba6faee5094f3c3994f66bea8d0125c2ad7a68.jpg) + +![](images/eded246d9e8aa73d422f8e0882e7586a966445a068d7c6a6d6795fa1a09e8e70.jpg) + +![](images/ef9afded5d387628fd364d2c4cd56fae90b334f0f9c5cddf3731d9dd3f89c94d.jpg) +"a red hatchback" + +![](images/56a667ddbf78bb4336889deaaa17edfb52b124a8705985db8b10e246c62eec8b.jpg) + +![](images/8541ce9317e965bf6cdebc994ca1977d5bc89a6eaf0ac595ad5e6bab6a9ec569.jpg) + +![](images/f48239947a516ed8281391162e8b735a1991902490a7b1709a690f404f4f28d3.jpg) + +![](images/c6dba14818a767c762bc1c615c352eaa3e265e07b29184724732082e23352fd9.jpg) +"a yellow racing car" + +![](images/8e201e724c96f4162c4d70b1afad2ff6fa4621bd39a8ad34d9b08da5d095b334.jpg) + +![](images/6399f71a632cac11e6e3b17283d083c0cc5dba9b825eea9f2c8edd63a28d0300.jpg) + +![](images/fa86196d4cb1cb3f9954adda2ae59cfb0899ab568c5c0e0896fa8adb3547eda6.jpg) + +![](images/6ff9b9d32a0cf3392703b2bb00814600bc0b41834526fb51e8e108ae5f6c6d47.jpg) +"a yellow racing motorcycle" + +![](images/b75b8296f04efc925635434639a9a29bf1d35a590a270ca70da5cb5eefbc0c6d.jpg) + +![](images/3717f6429816b70c471113303b3b685dc9883adcdf85fb54d9eb4d69f58db01c.jpg) + +![](images/2c5444d8ab1be15714f6af8f158bf8034ff217e8b199d136d3337dfa425331df.jpg) +Figure 7. We show the generation diversity and text control of our method. Each row takes the same text prompt with different sampled random noise as the model input. + +![](images/0a504c1a2ce7714b3c31784ba414f208110df88e6c338551072bec69c55e8050.jpg) +"a red motorbike" + +![](images/48bde7ece2ef1a61c57ff0b227a2c0ce85e743865660cd39a43aaf68080a2dff.jpg) + +![](images/3f10bb1b7e7d35fbb88c837e625718343c0f1a3ae78764dbdd7f0318dba7f7a0.jpg) + +![](images/d7609661119dc0896a8d575b3614efcb3103d4a4281823c6f3e0df1ec03eba42.jpg) + +![](images/f1ee4643d9dfac1ad7dc9001e9ff52e6e23d8a11391bc868bc3be330322c3d2d.jpg) + +![](images/ecdd8747f4766720a015829bea424167f28343546e9f6b9fcdab4f1950f76102.jpg) + +![](images/87ce7ea8c4cb3e8e58d6e409e593cbb48d717015a4b2ef2dc0b39ae37ddfdc4b.jpg) + +![](images/6ab7114968b1bedd42ef98ceb4f7bb6f731b2b50e66e1a27c810852196f35a9a.jpg) +"agrayarmlesschair" + +![](images/0c7f1bf21c71b346cdc59a99f252e6d8da8539e2cef8a801c640d0af4b7a900f.jpg) + +![](images/55d0a9969116303f0c59ccd0d5ac931693d5f40020bdea9e7fb1c062ab7f5e62.jpg) + +![](images/61436fb6d9ae7b707e943bb5eb78cbebba944da9bd7645debe72a86d4dd1dd5f.jpg) +"agreenchair" + +![](images/ae3808d742a0176ed08b9ff46f488081d80ed60c56c7f58dbdafa9031bb7ddb4.jpg) + +![](images/d984dd161ffa86bbe648beb7a77a35a03b5899e2c0915474ede145863aa28088.jpg) + +![](images/dcd2f436a5274261f889d67dc87eb795f098949fd882927f5f563f66b6415fc7.jpg) + +![](images/0a304aa388f185a7249b6020135e6e9b99b6c1787e3d42417d857084e1dc567a.jpg) + +![](images/3f57037c3d6098c55afe9912ebc8fbf53b42b79ff56585e0ff49446b496f489f.jpg) + +![](images/e986e0072a32f0a06f4006085075158792a8c6c2a76a57a162d968e8dca14293.jpg) + +![](images/8c8a27f3c506847b373da38df7aff79cd649a23c148a6fb3bcfaa887883efa43.jpg) + +![](images/9b3cf9233e942161b2dacb19c7fa10774c7854d9ad3161f2583210b79f297b36.jpg) + +![](images/cf0aa2c2aa13c6eb3d9ad470e1c84c1c1bb06d9d97bdeb145733ccfac5d07082.jpg) + +![](images/3bacb4b37a17e1ae89c902e02d476d7661e41b116482370a209d18f8b4106f6c.jpg) +"a red dining table" +"a silver metallic table" + +![](images/6454e7e27334ed49ea3be4a0687aecd664489cbeb25263f0c29e859ba09ecb0b.jpg) + +![](images/d7573347e3ed9cedfe310f5ea8e937b3651d42520bc892685df5e6e9eca4b75a.jpg) + +![](images/3643b402ba20c8916ad4ad52bf02c7df9d6d6fb0d009a5451ed46e11eca1f0a0.jpg) + +![](images/7608b1350c6c1f3a294010a047daf196b0d0c1d675438a8d9da81519d024479c.jpg) + +![](images/37038c41b49cd9beda1ac903ba2afeeebca34cdff84ab236a0b7235c2c008157.jpg) + +![](images/ad989530d6a8c339a54c11ffa8975b30c5f387fdc9392994649aa616aa87be06.jpg) + +![](images/eaba7cb8d453835a6cab9c519b09897f7d22bc1d9c64aa8965875ded873a711f.jpg) + +![](images/9dbfc68e7cc4d5d38097902f5ccb3162f8799b7147c2cd47a095e0931d9baf8f.jpg) + +![](images/a13f994c575b76ecd3a24fdc29c64d37773f7e1050e5ec6d07bbece13647562f.jpg) + +![](images/0fdfe74916a25d18e386930f44abc70b3a545246c1c31a18bb87ed41c3ea7274.jpg) + +![](images/1188f55a4c568cf4b911b82a79698c859f963a6716f8c9fdd5548af3fa5fdadd.jpg) + +![](images/cf100d6544b188189ffc190003af17a3f9c449f7b49bf39e28568529c31597af.jpg) + +![](images/64f04db05a9e3c07045b5ecee724a241779d4e14bf4b79eb78991d0935232b39.jpg) + +![](images/1cdc7f32f01adf592a78141e19a5170ad54eb071c03ac18de3dc1a72d7dd315d.jpg) + +![](images/746b9a72f10d9bce6dcb331a629dcd22f7f4ae9a949fef4b44c10f03080a60e8.jpg) + +![](images/ff0a8a2744f862b14fb080a8d442f3fb88db7891bf89748240560faecbece891.jpg) + +![](images/ff3fc6719ea660eee7d6d617819f13ff6b6d1bb58e7077cfee63e70e1cfbac74.jpg) + +![](images/0c190a8d91e5e0bed32c5b35968049a471c5ad1da6ea538c721a0236a5e3df7a.jpg) +"ared{class}" +Figure 8. Text control of textures on four different categories. Each row shares the same sampled random noise input. + +![](images/7eb3bd917cb8776d698cdac89beebbbdc6eff12240991edc0d378b5de1610916.jpg) +"a yellow {class}" + +![](images/596cb04ec69713b9445ed85fb327a9d3191c463cdc924790aecc6012be5f028f.jpg) +"agreen{class} + +![](images/79c91a0412a11ea75c96a94c7fa7c318dc67f0a158f267dd5b3967c31db44673.jpg) +"a blue {class}" + +![](images/34d26746bdb3bcde65d7beac3cde946ef5c7d531a9f2e9f404ad883a86f850f2.jpg) +"a gray {class}" + +coarse output is later refined with interpolations and marching cubes to generate mesh data. This process takes ten times longer than the network inference time and costs 24.44 seconds per sample. Our method generates high-resolution renderings at $1024^2$ resolution with only 0.05 seconds and generates the complete explicit mesh representation in 7.09 seconds. + +# 4.4. Qualitative Results + +In Fig. 7, we present the generation diversity and text control strength of our method in four object classes. In each row, we sample different random noises as well as the given text prompt as input to generate diversified and text-controlled 3D shapes, where the input text features are simultaneously fed into both the geometry and texture mapping network. Generally, we observe the semantic meanings of the rendered 2D images, and the given prompts are consistent. Moreover, there exists large diversity in the texture and geometry of our generated 3D shapes, even from the same text input. Here we present generated results of 4 classes, where $Car$ has a relatively simple geometry, while Table and Chair are more complex. But we still have good control over the textures and geometries for the challenging classes. For example, when we input "a gray armless chair", the generated samples exactly match the given instructions. However, we also observe failure cases of Table generation. For instance, the generation of "a wooden office desk" may have distorted geometry due to its complex structure. In Fig. 8, we show the text control ability of our method on different classes. The results in each row are generated from the same random noise input but with dif + +Table 5. We compare the FID score of using image CLIP features (Lafite [55]) and the pseudo text CLIP features (Ours) as training input. + +
MethodCarChairTableMotorbike
Lafite [55]68.781.194.3146.5
Ours21.744.843.257.2
+ +Table 6. We compare the FID score for our method without $\mathcal{L}_{clip}$ , $\mathcal{L}_{img}$ and background augmentation (BGaug). + +
MethodCarChairTableMotorbike
Ours21.744.843.257.2
-w/o Lclip83.7148.967.1134.9
-w/o Limg59.8144.248.4101.8
-w/o BGaug55.165.158.1110.0
+ +ferent text prompts. To summarize, the results show that our generation results are highly consistent with the input text prompts across different object classes. + +# 4.5. Ablation Study + +Pseudo caption generation module. In Tab. 5, we compare our method with Lafite [55], which is a language-free training method for the text-to-image generation task. Lafite claims due to the alignment between CLIP text and image embeddings. The CLIP image features can be directly used as text input during training. To implement Lafite [55] onto our framework, we feed the rendered real image CLIP features into our model and train the model by minimizing the cosine similarity between the generated and input image CLIP features. The quantitative results show that our proposed pseudo caption generation module significantly improves the generation quality compared to the baseline. We also give a qualitative comparison in Fig. 9 (b), we observe that the generator trained with Lafite [55] fails to generate correct textures. This problem might come from the modality gap between the CLIP image and text embeddings, as discussed in [20, 54]. + +Effects of our loss functions. We conduct an ablation study on our used two losses $\mathcal{L}_{clip}$ and $\mathcal{L}_{img}$ . In Tab. 6, we quantitatively validate the usefulness of both two losses. Specifically, $\mathcal{L}_{clip}$ supervises the text-image alignment learning, which is more critical than $\mathcal{L}_{img}$ , and removing $\mathcal{L}_{clip}$ gives a more considerable performance drop. From the qualitative results in Fig. 9 (c), we observe that simply using $\mathcal{L}_{img}$ leads to the model generating distorted geometries, which may indicate the significance of high-level semantic supervision. We also observe that solely using $\mathcal{L}_{clip}$ (Fig. 9 (d)) may cause the model generation to lose fine-grained texture details, as well as the geometry information. + +Effects of background augmentation. Our used GET3D [9] backbone network generates RGBA images with dark backgrounds, which may bring some noise when + +![](images/05a30e36ad4296ca32e197ddacb32424fb7993c74924a548eeca56c5ae6bf2aa.jpg) +(a) + +Figure 9. Ablation study on our methods with text prompt "a red car". We show results of (a) our full model, (b) Lafite [55], (c) without high-level semantic regularization $\mathcal{L}_{clip}$ , (d) without low-level image regularization $\mathcal{L}_{img}$ , (e) without background augmentation. All the visualizations are generated from models trained for 250k iterations. +Figure 10. Examples of failure cases of our method. Our model often fails to handle text input with composed descriptions for different parts of the object. +![](images/a74c124e3363581d94064041940b941d7f12c5160f11a032706385921f174bf0.jpg) +"acarwithyellowlegs, redseat and blue back." + +we conduct low-level image regularization. As illustrated from the quantitative results in Tab. 6 and the visualizations in Fig. 9 (e), training with the augmented background can help remove the texture flaws. + +# 5. Conclusion + +We introduce TAPS3D, a text-guided 3D textured shape generation method. We first generate pseudo captions for model training. During the inference phase, TAPS3D does not require additional optimization. Our fast generation speed enables average users to generate high-quality 3D objects with acceptable processing time. + +**Limitation.** The limitation of our method mainly lies in the capacity of our model is still bound to the training image data, and we cannot produce different fine-grained details for different object parts, as shown in Fig. 10. Although we do not require paired text data, we still need diversified training images to handle complex text input. + +# Acknowledgment + +This research is partly supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-RP-2018-003), and the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and MoE AcRF Tier 1 grants (RG14/22, RG95/20). + +# References + +[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2 +[2] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc., 2009. 4 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 3 +[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4, 5 +[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2 +[7] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2 +[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. arXiv preprint arXiv:2207.09446, 2022. 2 +[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv preprint arXiv:2209.11163, 2022. 1, 2, 3, 4, 5, 8 +[10] Thomas D Goddard, Conrad C Huang, Elaine C Meng, Eric F Pettersen, Gregory S Couch, John H Morris, and Thomas E Ferrin. Ucsf chimerax: Meeting modern challenges in visualization and analysis. Protein Science, 27(1):14-25, 2018. 1, 6 +[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylererf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 + +[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6 +[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3 +[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[17] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 895–904, 2022. 2 +[18] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 2, 6 +[19] Ruihui Li, Xianzhi Li, Ka-Hei Hui, and Chi-Wing Fu. Spagan: Sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 6 +[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 2, 3, 8 +[21] Kangcheng Liu. Rm3d: Robust data-efficient 3d scene parsing via traditional and learnt 3d descriptors-based semantic region merging. International Journal of Computer Vision, pages 1-30, 2022. 1 +[22] Kangcheng Liu, Zhi Gao, Feng Lin, and Ben M Chen. Fgnet: A fast and accurate framework for large-scale lidar point cloud understanding. IEEE Transactions on Cybernetics, 53(1):553-564, 2022. 1 +[23] Kangcheng Liu, Aoran Xiao, Xiaoqin Zhang, Shijian Lu, and Ling Shao. Fac: 3d representation learning via foreground aware feature contrast. arXiv preprint arXiv:2303.06388, 2023.1 +[24] Kangcheng Liu, Yuzhi Zhao, Qiang Nie, Zhi Gao, and Ben M Chen. Weakly supervised 3d scene segmentation with region-level boundary awareness and instance discrimination. In European Conference on Computer Vision 2022 (ECCV 2022), pages 37-55. Springer, Cham, 2022. 1 +[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 6 +[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization + +for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[28] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 3(7):e12, 2018. 5 +[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2 +[31] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2 +[32] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6 +[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2 +[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 6 +[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 6 +[36] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4 +[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen + +eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[40] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[42] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 6 +[44] Aditya Sanghi, Rao Fu, Vivian Liu, Karl Willis, Hooman Shayani, Amir Hosein Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Textcraft: Zero-shot generation of high-fidelity and diverse shapes from text. arXiv preprint arXiv:2211.01427, 2022. 1, 2 +[45] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 1 +[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 1 +[47] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 6 +[48] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. 3d pose transfer with correspondence learning and mesh refinement. Advances in Neural Information Processing Systems, 34:3108-3120, 2021. 1 +[49] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. Unsupervised 3d pose transfer with cross consistency and dual reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[50] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 1, 2, 6 +[51] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6 +[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2 +[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2 +[54] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2, 3, 8 +[55] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 3, 8 \ No newline at end of file diff --git a/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/images.zip b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ee3f972abf44ad3f1648bdde878058f88d96105b --- /dev/null +++ b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c05090d6f65eb806c81239c2e3640ca297ecf3efb93d0a34bc7907276b68c58 +size 528995 diff --git a/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/layout.json b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6493c0d5da10ccddb3c4f1ccc60c78c1b88e90b2 --- /dev/null +++ b/2023/TAPS3D_ Text-Guided 3D Textured Shape Generation From Pseudo Supervision/layout.json @@ -0,0 +1,11738 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "text", + "content": "TAPS3D: Text-Guided 3D Textured Shape Generation from Pseudo Supervision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "spans": [ + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": "Jiacheng Wei" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": " Hao Wang" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": " Jiashi Feng" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": " Guosheng Lin" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": " Kim-Hui Yap" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": "Nanyang Technological University, Singapore " + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 141, + 520, + 186 + ], + "type": "text", + "content": "ByteDance \n{jiacheng002@e., hao005@e., gslin@, ekhyap@}ntu.edu.sg, jshfeng@bytedance.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 54, + 204, + 178, + 266 + ], + "blocks": [ + { + "bbox": [ + 54, + 204, + 178, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 204, + 178, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 178, + 266 + ], + "type": "image", + "image_path": "fe5a857b9162dd7165cba42c3288eb459803ae27b0cd1d10664bb92d25f41204.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 270, + 167, + 281 + ], + "lines": [ + { + "bbox": [ + 72, + 270, + 167, + 281 + ], + "spans": [ + { + "bbox": [ + 72, + 270, + 167, + 281 + ], + "type": "text", + "content": "\"a red hatchback car\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 185, + 207, + 223, + 265 + ], + "blocks": [ + { + "bbox": [ + 185, + 207, + 223, + 265 + ], + "lines": [ + { + "bbox": [ + 185, + 207, + 223, + 265 + ], + "spans": [ + { + "bbox": [ + 185, + 207, + 223, + 265 + ], + "type": "image", + "image_path": "21f46766d115262baf96255299c52ae29b897dfbe28841a78771350ea37f33f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 326 + ], + "type": "text", + "content": "Figure 1. Our proposed model is capable of generating detailed textured shapes based on given text prompts. We demonstrate the effectiveness of our approach by showcasing generated results for four distinct object classes: Car, Chair, Table, and Motorbike. Both the textured meshes and geometries are presented, with visualizations rendered using ChimeraX [10]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 230, + 205, + 291, + 265 + ], + "blocks": [ + { + "bbox": [ + 192, + 270, + 285, + 282 + ], + "lines": [ + { + "bbox": [ + 192, + 270, + 285, + 282 + ], + "spans": [ + { + "bbox": [ + 192, + 270, + 285, + 282 + ], + "type": "text", + "content": "\"agrayarmlesschair" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 205, + 291, + 265 + ], + "lines": [ + { + "bbox": [ + 230, + 205, + 291, + 265 + ], + "spans": [ + { + "bbox": [ + 230, + 205, + 291, + 265 + ], + "type": "image", + "image_path": "d52a4bf3050eede9191cd8ed4d3f1bc0081fb8233fb32bff1575993e28c2dfdc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 217, + 360, + 259 + ], + "blocks": [ + { + "bbox": [ + 301, + 217, + 360, + 259 + ], + "lines": [ + { + "bbox": [ + 301, + 217, + 360, + 259 + ], + "spans": [ + { + "bbox": [ + 301, + 217, + 360, + 259 + ], + "type": "image", + "image_path": "7fbc3400b84d242098e2dd16ac61d1a9b2fac285a4f836a490a1670cc65e91b8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 413, + 282 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 413, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 413, + 282 + ], + "type": "text", + "content": "\"awooden office table\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 361, + 205, + 421, + 265 + ], + "blocks": [ + { + "bbox": [ + 361, + 205, + 421, + 265 + ], + "lines": [ + { + "bbox": [ + 361, + 205, + 421, + 265 + ], + "spans": [ + { + "bbox": [ + 361, + 205, + 421, + 265 + ], + "type": "image", + "image_path": "665635970e9f3d95c6c1be52840257be7b7a2c58394773e5225ea36c6c1fa4eb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 213, + 477, + 258 + ], + "blocks": [ + { + "bbox": [ + 425, + 213, + 477, + 258 + ], + "lines": [ + { + "bbox": [ + 425, + 213, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 425, + 213, + 477, + 258 + ], + "type": "image", + "image_path": "c5b238b9babd634c15d7251ce285259ebc330cdf936715f53b1a4c24a60f8f11.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 270, + 524, + 281 + ], + "lines": [ + { + "bbox": [ + 440, + 270, + 524, + 281 + ], + "spans": [ + { + "bbox": [ + 440, + 270, + 524, + 281 + ], + "type": "text", + "content": "\"ablock motorbike\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 479, + 205, + 539, + 265 + ], + "blocks": [ + { + "bbox": [ + 479, + 205, + 539, + 265 + ], + "lines": [ + { + "bbox": [ + 479, + 205, + 539, + 265 + ], + "spans": [ + { + "bbox": [ + 479, + 205, + 539, + 265 + ], + "type": "image", + "image_path": "e052b58908dfaf5125946af3eb2a77b3c359ec9bf9276b50f80eef9c281e2230.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 336, + 192, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 336, + 192, + 348 + ], + "spans": [ + { + "bbox": [ + 143, + 336, + 192, + 348 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 362, + 290, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 290, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 290, + 615 + ], + "type": "text", + "content": "In this paper, we investigate an open research task of generating controllable 3D textured shapes from the given textual descriptions. Previous works either require ground truth caption labeling or extensive optimization time. To resolve these issues, we present a novel framework, TAPS3D, to train a text-guided 3D shape generator with pseudo captions. Specifically, based on rendered 2D images, we retrieve relevant words from the CLIP vocabulary and construct pseudo captions using templates. Our constructed captions provide high-level semantic supervision for generated 3D shapes. Further, in order to produce fine-grained textures and increase geometry diversity, we propose to adopt low-level image regularization to enable fake-rendered images to align with the real ones. During the inference phase, our proposed model can generate 3D textured shapes from the given text without any additional optimization. We conduct extensive experiments to analyze each of our proposed components and show the efficacy of our framework in generating high-fidelity 3D textured and text-relevant shapes. Code is available at https://github.com/plusmultiply/TAPS3D" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 638, + 128, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 128, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 128, + 650 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "content": "3D objects are essential in various applications [21-24], such as video games, film special effects, and virtual real" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 337, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 545, + 421 + ], + "type": "text", + "content": "ity. However, realistic and detailed 3D object models are usually hand-crafted by well-trained artists and engineers slowly and tediously. To expedite this process, many research works [3,9,13,34,48,49] use deep generative models to achieve automatic 3D object generation. However, these models are primarily unconditioned, which can hardly generate objects as humans will." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 423, + 546, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 698 + ], + "type": "text", + "content": "In order to control the generated 3D objects from text, prior text-to-3D generation works [43, 44] leverage the pretrained vision-language alignment model CLIP [37], such that they can only use 3D shape data to achieve zero-shot learning. For example, Dream Fields [13] combines the advantages of CLIP and NeRF [27], which can produce both 3D representations and renderings. However, Dream Fields costs about 70 minutes on 8 TPU cores to produce a single result. This means the optimization time during the inference phase is too slow to use in practice. Later on, GET3D [9] is proposed with faster inference time, which incorporates StyleGAN [15] and Deep Marching Tetrahedral (DMTet) [46] as the texture and geometry generators respectively. Since GET3D adopts a pretrained model to do text-guided synthesis, they can finish optimization in less time than Dream Fields. But the requirement of test-time optimization still limits its application scenarios. CLIP-NeRF [50] utilizes conditional radiance fields [45] to avoid test-time optimization, but it requires ground truth text data for the training purpose. Therefore, CLIP-NeRF is only applicable to a few object classes that have labeled text data for training, and its generation quality is restricted by the NeRF capacity." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To address the aforementioned limitations, we propose" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 280, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 280, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 280, + 703 + ], + "type": "text", + "content": "*Equal contribution. Work done during an internship at Bytedance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 61, + 703, + 140, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 140, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 140, + 713 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16805" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "to generate pseudo captions for 3D shape data based on their rendered 2D images and construct a large amount of " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\langle 3\\mathrm{D}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " shape, pseudo captions " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " as training data, such that the text-guided 3D generation model can be trained over them. To this end, we propose a novel framework for Text-guided 3D textured shApe generation from Pseudo Supervision (TAPS3D), in which we can generate high-quality 3D shapes without requiring annotated text training data or test-time optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 289, + 324 + ], + "type": "text", + "content": "Specifically, our proposed framework is composed of two modules, where the first generates pseudo captions for 3D shapes and feeds them into a 3D generator to conduct text-guided training within the second module. In the pseudo caption generation module, we follow the language-free text-to-image learning scheme [20, 54]. We first adopt the CLIP model to retrieve relevant words from given rendered images. Then we construct multiple candidate sentences based on the retrieved words and pick sentences having the highest CLIP similarity scores with the given images. The selected sentences are used as our pseudo captions for each 3D shape sample." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 326, + 289, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 289, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 289, + 505 + ], + "type": "text", + "content": "Following the notable progress of text-to-image generation models [29, 38, 39, 42, 52], we use text-conditioned GAN architecture in the text-guided 3D generator training part. We adopt the pretrained GET3D [9] model as our backbone network since it has been demonstrated to generate high-fidelity 3D textured shapes across various object classes. We input the pseudo captions as the generator conditions and supervise the training process with high-level CLIP supervision in an attempt to control the generated 3D shapes. Moreover, we introduce a low-level image regularization loss to produce fine-grained textures and increase geometry diversity. We empirically train the mapping networks only of a pretrained GET3D model so that the training is stable and fast, and also, the generation quality of the pretrained model can be preserved." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 506, + 289, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 554 + ], + "type": "text", + "content": "Our proposed model TAPS3D can produce high-quality 3D textured shapes with strong text control as shown in Fig. 1, without any per prompt test-time optimization. Our contribution can be summarized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 560, + 287, + 719 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 560, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 560, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 58, + 560, + 287, + 608 + ], + "type": "text", + "content": "- We introduce a new 3D textured shape generative framework, which can generate high-quality and fidelity 3D shapes without requiring paired text and 3D shape training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 615, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 615, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 58, + 615, + 287, + 675 + ], + "type": "text", + "content": "- We propose a simple pseudo caption generation method that enables text-conditioned 3D generator training, such that the model can generate text-controlled 3D textured shapes without test time optimization, and significantly reduce the time cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 682, + 287, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 682, + 287, + 719 + ], + "spans": [ + { + "bbox": [ + 58, + 682, + 287, + 719 + ], + "type": "text", + "content": "- We introduce a low-level image regularization loss on top of the high-level CLIP loss in an attempt to produce fine-grained textures and increase geometry diversity." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 91, + 491, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 491, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 491, + 104 + ], + "type": "text", + "content": "2.1. Text-Guided 3D Shape Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 110, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 253 + ], + "type": "text", + "content": "Text-guided 3D shape generation aims to generate 3D shapes from the textual descriptions so that the generation process can be controlled. There are mainly two categories of methods, i.e., fully-supervised and optimization-based methods. The fully-supervised method [6,8,25] uses ground truth text and the paired 3D objects with explicit 3D representations as training data. Specifically, CLIP-Forge [43] uses a two-stage training scheme, which consists of shape autoencoder training, and conditional normalizing flow training. VQ-VAE [44] performs zero-shot training with 3D voxel data by utilizing the pretrained CLIP model [37]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 255, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 518 + ], + "type": "text", + "content": "Regarding the optimization-based methods [13, 18, 26, 34], Neural Radiance Fields (NeRF) are usually adopted as the 3D generator. To generate 3D shapes for each input text prompt, they use the CLIP model to supervise the semantic alignment between rendered images and text prompts. Since NeRF suffers from extensive generation time, 3D-aware image synthesis [3, 3, 4, 11, 30, 31] has become popular, which generates multi-view consistent images by integrating neural rendering in the Generative Adversarial Networks (GANs). Specifically, there are no explicit 3D shapes generated during the process, while the 3D shapes can be extracted from the implicit representations, such as the occupancy field or signed distance function (SDF), using the marching cube algorithm. These optimization-based methods provide a solution to generate 3D shapes, but their generation speed was compromised. Although [50] is free from test-time optimization, it requires text data for training, which limits its applicable 3D object classes. Our proposed method attempts to alleviate both the paired text data shortage issues and the long optimization time in the previous work. We finally produce high-quality 3D shapes to bring our method to practical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 526, + 441, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 441, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 441, + 540 + ], + "type": "text", + "content": "2.2. Text-to-Image Synthesis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 641 + ], + "type": "text", + "content": "We may draw inspiration from text-to-image generation methods. Typically, many research works [36,40,53] adopt the conditional GAN architecture, where they directly take the text features and concatenate them with the random noise as the input. Recently, autoregressive models [7, 39] and diffusion models [29, 38, 41, 42] made great improvement on text to image synthesis while demanding huge computational resources and massive training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "With the introduction of StyleGAN [14-16] mapping networks, the input random noise can be first mapped to another latent space that has disentangled semantics, then the model can generate images with better quality. Further, exploring the latent space of StyleGAN has been proved useful by several works [1, 17, 33] in text-driven image synthe" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16806" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 515, + 230 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 515, + 230 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 515, + 230 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 515, + 230 + ], + "type": "image", + "image_path": "08cd0e04469de448a7fa81f72149eb38221d57c37e541e9b0ea5c09fac8c2413.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. Our proposed framework supports text-guided 3D shape generation training without paired text labeling. To generate captions for training, we first adopt a pseudo caption generation module to produce textual descriptions, given rendered 2D images. Then we feed the pseudo caption CLIP embeddings into the mapping networks to train the controllable text-guided 3D shape synthesis network. Our training pipeline is supervised by high-level semantic loss " + }, + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "type": "text", + "content": " and low-level image regularization " + }, + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "type": "text", + "content": ". It is notable that during the training phase, only the mapping network weights are updated, and the rest components are fixed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 294, + 289, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 474 + ], + "type": "text", + "content": "sis and manipulation tasks, where they utilize the pretrained vision-language model CLIP [37] model to manipulate pretrained unconditional StyleGAN networks. In order to relieve the need for paired text data during the training phase, Lafite [55] proposes to adopt the image CLIP embeddings as training input while using text CLIP embedding during the inference phase. However, as discussed in the recent works [20, 54], there exists a modality gap between text and image embeddings in the CLIP latent space. Therefore, language-free methods that are solely trained on CLIP image embedding could introduce distribution mismatch when encountering the CLIP text embedding input during inference. In this paper, we further explore language-free training for text-guided 3D textured shape generation by proposing a pseudo caption generation module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 483, + 104, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 104, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 104, + 495 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 127, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 127, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 127, + 515 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": "GET3D [9] is a newly proposed 3D generator that can generate high-fidelity textured 3D shapes. Specifically, GET3D maps noise vectors " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " to a textured mesh. In order to disentangle the geometry and texture information, GET3D takes two random vectors " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_2" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " as inputs for the geometry and texture generation branches, respectively. Following the design of StyleGAN [15], and EG3D [3], they map " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_2" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " to the intermediate latent codes " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_1" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_2" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " with two mapping networks since the semantics would be better represented in the mapped space. The geometry and texture generators are built conditioned on " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_1" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_2" + }, + { + "bbox": [ + 46, + 521, + 287, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 115, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 115, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 115, + 682 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Since the GET3D [9] model only supports unconditional 3D generation, if one need to do text-guided 3D shape syn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 294, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 354 + ], + "type": "text", + "content": "thesis, additional optimization is required with the original GET3D model. In order to control the semantics of the generated results and achieve text-based textured 3D generation in one shot, we propose to add text conditions to GET3D models. In Fig. 2, we present our training framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 533 + ], + "type": "text", + "content": "To this end, we address two main challenges: one is the lack of training textual descriptions, and another is color and geometry distortion. Regarding the first challenge, we propose a pseudo caption generation module, which gives large amounts of captions for training purposes. These captions are used as conditions and fed into the mapping networks to control the semantic meanings of the output textured 3D shapes. We apply the high-level CLIP [37] semantic loss on the paired input text and rendered 2D images to achieve text-3D alignment. To resolve the second challenge, we propose to regularize the color and geometry training with low-level image regularization. This method provides complimentary image-level guidance to our learning objective, which gives fine-grained textures during the generation phase." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": "Notably, we empirically adopt the pretrained GET3D model and train its mapping networks only since we observe this practice not only improves the training efficiency but also stabilizes the training process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 587, + 459, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 459, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 459, + 601 + ], + "type": "text", + "content": "3.3. Pseudo Caption Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "content": "Most recent 3D shape generation models are data-driven. With large-scale training data, their performance improvement is remarkable. However, when we aim to train a text-guided 3D shape generation model, we can only obtain a few textual descriptions for limited samples. To resolve this issue, we follow the method of [54] and propose to generate pseudo captions for each 3D object." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Specifically, we observe the textual descriptions for a single 3D object are highly structured, e.g., \"A yellow" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16807" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 289, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 289, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 289, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 289, + 165 + ], + "type": "image", + "image_path": "7cc311faf542181fb0e5ac97ac8033034c713890872e9c45f089f39fd4023e5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 248 + ], + "type": "text", + "content": "Figure 3. Our proposed pseudo caption generation module can be split into four steps. Firstly, we construct vocabulary from ShapeNet-related nouns and adjectives in the CLIP vocabulary. Then we retrieve multiple words based on the 2D-rendered images. In the third step, we create candidate captions with retrieved words. Finally, we select one caption based on the computed text-image similarities with the CLIP model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 287, + 334 + ], + "type": "text", + "content": "sports car\", which typically consists of adjectives and a noun. Therefore, it is critical to obtain suitable adjectives and nouns for given 3D objects to construct pseudo captions automatically. To this end, we adopt the pretrained text-image similarity model of CLIP [37] to retrieve useful nouns and adjectives and further build complete sentences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 432 + ], + "type": "text", + "content": "The pseudo caption generation module can be split into four steps, shown in Fig. 3. In the first step, we construct a vocabulary to cover the related adjectives and nouns. Since we aim to generate ShapeNet [5] object classes, we collect all the class names as nouns in the vocabulary, which has 169 tokens. Then we extract all the adjectives from CLIP vocabulary with the Natural Language Toolkit [2]. The size of our collected adjectives is 1463." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "content": "In the second and third steps, we aim to retrieve relevant words from 2D-rendered images and construct candidate captions. Technically, we first extract all the vocabulary word embeddings and rendered image embeddings, where we adopt the pretrained CLIP text and image encoders, respectively. Then for each object, we compute the cosine similarities between the given image embedding and all the text embeddings. It is notable that we retrieve the nouns and adjectives separately, where we select the top-" + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "inline_equation", + "content": "K_{1}" + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "content": " retrieved nouns and top-" + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "inline_equation", + "content": "K_{2}" + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "content": " retrieved adjectives. With the selected words, we utilize the templates to construct candidate sentences. Since these sentences are only supposed to describe a static object, we use the template \"a{Adjective} {Noun}\", where we put 1 or 2 adjectives randomly to increase the caption diversity. Due to having " + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "inline_equation", + "content": "K_{1}" + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "content": " nouns and " + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "inline_equation", + "content": "K_{2}" + }, + { + "bbox": [ + 46, + 435, + 287, + 639 + ], + "type": "text", + "content": " adjectives, we are able to combine them and produce multiple candidate pseudo captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "In order to obtain the most suitable captions for given objects, we further conduct image-to-candidate caption retrieval, where we also use CLIP encoders to extract the image and sentence embeddings. To be specific, we use the top retrieved sentences as the final pseudo captions, which are used as the input condition for the 3D generator." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 545, + 191 + ], + "type": "image", + "image_path": "72c57836373d24beb7efbfac9109fb78bf4b5a96d827177b3aecb4158bbffdd8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 195, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 262 + ], + "type": "text", + "content": "Figure 4. In the training phase, we not only conduct the high-level text-to-image similarity learning through the CLIP model but also use low-level image regularization to obtain fine-grained textures and diverse shapes, where we render real and fake images with the same camera poses and apply background augmentation to allow the model focus on object features only." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 266, + 462, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 462, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 462, + 280 + ], + "type": "text", + "content": "3.4. Text-3D Alignment Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "text", + "content": "After the pseudo caption generation, we feed these captions into a pretrained GET3D [9] model in an attempt to control the semantic meanings of generated 3D objects. To be specific, we first adopt the CLIP text encoder " + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "text", + "content": " to extract the caption " + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "text", + "content": " features, then the caption features " + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "inline_equation", + "content": "E_{t}(t)" + }, + { + "bbox": [ + 304, + 285, + 545, + 429 + ], + "type": "text", + "content": " are served as conditions and fed into the mapping networks of pretrained GET3D models. The input of the original mapping networks is random noise only. To inject the text information, we concatenate the caption embeddings and the random noise together as the new input. This enables us to generate diverse 3D objects that match the text's semantic meanings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 429, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 524 + ], + "type": "text", + "content": "Since we adopt the pretrained 3D generator as our model backbone, which has a good ability to synthesize high-fidelity 3D textured shapes, fixing the generator weights allows our model to focus on getting semantic alignment between input textual descriptions and generated 3D objects and also alleviates the learning difficulty. Our main training objective is to maximize the image-text similarity measured by the CLIP model, which can be denoted as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 534, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 534, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 359, + 534, + 545, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l i p}} = 1 - \\cos \\left(E _ {i} \\left(I _ {x}\\right), E _ {t} (t)\\right). \\tag {1}", + "image_path": "4ce2e8ef84ee23e6967c0fc4ca8411411d275c6d41ce1b0a3a305868b8b7c177.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "inline_equation", + "content": "I_{x}" + }, + { + "bbox": [ + 304, + 555, + 545, + 604 + ], + "type": "text", + "content": " represent the CLIP image encoder and rendered images, respectively. These 2D images are rendered from the generated 3D textured shapes with randomly sampled camera poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 611, + 545, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 545, + 623 + ], + "type": "text", + "content": "3.5. Towards Diverse and Fine-Grained Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "It is observed using " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " only enables the model to generate plausible 3D shapes that align with the input text. However, the diversity of geometry generation is limited, and the generated colors are unnatural. This is caused by two main reasons. One reason is the diversity of our pseudo captions is also limited, as there are only a few descriptive words for a single object class. Moreover, the CLIP model" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16808" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 143 + ], + "type": "text", + "content": "may wrongly retrieve irrelevant words to the given objects, which also affects the generated shape quality. Another reason is the CLIP model only provides high-level guidance for our training supervision while failing to give fine-grained texture information, thus making the generated results unrealistic." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "type": "text", + "content": "In order to resolve this issue, we further introduce a low-level image regularization loss. Specifically, we use the same camera poses as that used in the " + }, + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "type": "inline_equation", + "content": "I_{x}" + }, + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "type": "text", + "content": " rendering to produce ground truth rendered images " + }, + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "type": "inline_equation", + "content": "I_{x}^{gt}" + }, + { + "bbox": [ + 46, + 144, + 286, + 204 + ], + "type": "text", + "content": ". The optimization objective is denoted as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 210, + 287, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 210, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 96, + 210, + 287, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m g} = 1 - \\cos \\left(E _ {i} \\left(I _ {x}\\right), E _ {i} \\left(I _ {x} ^ {g t}\\right)\\right). \\tag {2}", + "image_path": "2290d8f10ff1dc7623c6c02a32893c65783508348f184b2fe5d440a4cd5748be.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 231, + 287, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 362 + ], + "type": "text", + "content": "Since the rendering process returns an alpha channel, which results in a simple white or black background, if we directly take rendered images with the same backgrounds over training iterations, the background information will also be involved during optimizing Eq. (2). To alleviate this issue and allow the model to focus on the foreground objects, we follow [13] to conduct background augmentation, as shown in Fig. 4. Technically, we randomly select random Fourier textures [28], Gaussian noise, and checkerboard patterns as backgrounds. For each paired real and fake image, we apply the same augmented backgrounds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 196, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 196, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 196, + 382 + ], + "type": "text", + "content": "3.6. Overall Training Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 388, + 255, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 388, + 255, + 399 + ], + "spans": [ + { + "bbox": [ + 59, + 388, + 255, + 399 + ], + "type": "text", + "content": "Our overall training objective can be depicted as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 407, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 407, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 127, + 407, + 287, + 420 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c l i p}} + \\mathcal {L} _ {\\text {i m g}}. \\tag {3}", + "image_path": "f5c1963614b204c6058dd97c714592727cc942d0bb5596b8046ed89f96ce3ee6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 582 + ], + "type": "text", + "content": "During the training phase, we empirically experiment with updating all model parameters and the mapping networks only. We observe training all GET3D backbone parameters would severely distort the generated geometry and textures, where we show the experimental results in the supplementary material. This is because the learning objective is two-pronged: improving the generation quality and matching the input captions. While borrowing a pretrained unconditional generator ensures good generation quality, expedites the training process, and also yields stable text-guided results. Therefore, we regard training the mapping networks only of a pretrained 3D generator as a useful and practical solution for the text-guided 3D shape generation task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 591, + 127, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 127, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 127, + 604 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 611, + 105, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 105, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 105, + 623 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We train and evaluate our method on the large-scale 3D dataset ShapeNet [5]. Our experiments are conducted on four classes with complex geometry, Car, Table, Chair, and Motorbike, containing 7497, 8436, 6778, and 337 shapes, respectively. We follow GET3D [9] to split the data into training, validation, and test sets in a 7:1:2 ratio. The training image data are rendered following the GET3D setting." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 369, + 99 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 369, + 99 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 369, + 99 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 369, + 99 + ], + "type": "image", + "image_path": "8d0298588e05126143c53974f0caaadb948f5b5028ed38b290d618f44048f526.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 416, + 101 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 416, + 101 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 416, + 101 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 416, + 101 + ], + "type": "image", + "image_path": "5ea777085ae903bc12528e00f1adfd35e2bce12d7aacca7e5aa94ff4efdc5186.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 102, + 424, + 110 + ], + "lines": [ + { + "bbox": [ + 318, + 102, + 424, + 110 + ], + "spans": [ + { + "bbox": [ + 318, + 102, + 424, + 110 + ], + "type": "text", + "content": "a white automobile a low race car" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 112, + 361, + 155 + ], + "blocks": [ + { + "bbox": [ + 328, + 112, + 361, + 155 + ], + "lines": [ + { + "bbox": [ + 328, + 112, + 361, + 155 + ], + "spans": [ + { + "bbox": [ + 328, + 112, + 361, + 155 + ], + "type": "image", + "image_path": "4202d1f167d4ad5b76527e3cb0b2f603e665b049476448befe80fd14668bc3fc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 156, + 365, + 173 + ], + "lines": [ + { + "bbox": [ + 327, + 156, + 365, + 173 + ], + "spans": [ + { + "bbox": [ + 327, + 156, + 365, + 173 + ], + "type": "text", + "content": "a stationary chair" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 387, + 112, + 415, + 156 + ], + "blocks": [ + { + "bbox": [ + 387, + 112, + 415, + 156 + ], + "lines": [ + { + "bbox": [ + 387, + 112, + 415, + 156 + ], + "spans": [ + { + "bbox": [ + 387, + 112, + 415, + 156 + ], + "type": "image", + "image_path": "2f5c112a5191829020810077f7cebeef6f5fbf989993a15222f5c19fe4fdb18b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 156, + 421, + 173 + ], + "lines": [ + { + "bbox": [ + 373, + 156, + 421, + 173 + ], + "spans": [ + { + "bbox": [ + 373, + 156, + 421, + 173 + ], + "type": "text", + "content": "a white ladderback chair" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 474, + 96 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 474, + 96 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 474, + 96 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 474, + 96 + ], + "type": "image", + "image_path": "6f6ac98f4cafae482ef8f6b63e6cb2198b95e2df6871bc81c820fc9276fac14a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 432, + 123, + 474, + 147 + ], + "blocks": [ + { + "bbox": [ + 433, + 102, + 466, + 118 + ], + "lines": [ + { + "bbox": [ + 433, + 102, + 466, + 118 + ], + "spans": [ + { + "bbox": [ + 433, + 102, + 466, + 118 + ], + "type": "text", + "content": "a ural motorbike" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 123, + 474, + 147 + ], + "lines": [ + { + "bbox": [ + 432, + 123, + 474, + 147 + ], + "spans": [ + { + "bbox": [ + 432, + 123, + 474, + 147 + ], + "type": "image", + "image_path": "9b2151929695bbb231861c8350195ef66e524e09f6cb55852f1f312dc7cf2ec2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 478, + 72, + 524, + 99 + ], + "blocks": [ + { + "bbox": [ + 478, + 72, + 524, + 99 + ], + "lines": [ + { + "bbox": [ + 478, + 72, + 524, + 99 + ], + "spans": [ + { + "bbox": [ + 478, + 72, + 524, + 99 + ], + "type": "image", + "image_path": "dd2166693acb868fede55b4cbabd744a30ef4e28a3ad8909030ab71d5bc685f0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 480, + 119, + 526, + 159 + ], + "blocks": [ + { + "bbox": [ + 486, + 102, + 523, + 118 + ], + "lines": [ + { + "bbox": [ + 486, + 102, + 523, + 118 + ], + "spans": [ + { + "bbox": [ + 486, + 102, + 523, + 118 + ], + "type": "text", + "content": "a brown motorcycle" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 480, + 119, + 526, + 159 + ], + "lines": [ + { + "bbox": [ + 480, + 119, + 526, + 159 + ], + "spans": [ + { + "bbox": [ + 480, + 119, + 526, + 159 + ], + "type": "image", + "image_path": "3f09374256299dd939c36cb47a2e9bafc431f706a08435b21ebe92e4bd7ffde6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 160, + 533, + 168 + ], + "lines": [ + { + "bbox": [ + 478, + 160, + 533, + 168 + ], + "spans": [ + { + "bbox": [ + 478, + 160, + 533, + 168 + ], + "type": "text", + "content": "a structural table" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 182, + 539, + 193 + ], + "lines": [ + { + "bbox": [ + 311, + 182, + 539, + 193 + ], + "spans": [ + { + "bbox": [ + 311, + 182, + 539, + 193 + ], + "type": "text", + "content": "Figure 5. The demonstration of our generated pseudo captions." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 205, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 241 + ], + "type": "text", + "content": "We render 24 views randomly for Car, Chair and Table and 100 views for Motorbike since it has less available data. The pretrained models were trained on each class separately." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 258, + 436, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 436, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 436, + 270 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "type": "text", + "content": "We use the GET3D [9] model pretrained on ShapeNet [5] dataset as our backbone 3D generator. In the pseudo caption generation module, we set " + }, + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "type": "inline_equation", + "content": "K_{1}" + }, + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "type": "inline_equation", + "content": "K_{2}" + }, + { + "bbox": [ + 304, + 279, + 545, + 459 + ], + "type": "text", + "content": " as 3 and 6, respectively. We generate 20 pseudo captions for each ShapeNet object. During the training phase, each input image is randomly paired with one of the pseudo captions generated for the object which the image was rendered from. An example of the generated pseudo can be found in Fig. 5. We only update the geometry and texture mapping networks. We set the batch size to 16 and ran all the experiments on 4 Nvidia Tesla V100-32G GPUs. We empirically observe the geometry branch is more sensitive to the weight update. Hence we set learning rates for geometry and texture mapping networks as 0.004 and 0.0005, respectively. It costs around 10 hours for model training." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 475, + 494, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 494, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 494, + 488 + ], + "type": "text", + "content": "4.3. Comparison with Existing Methods" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 497, + 439, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 439, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 439, + 510 + ], + "type": "text", + "content": "4.3.1 Qualitative comparison" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "In Fig. 6, we first compare with the method TITG3SG [25], which is an encoder-decoder-based framework. They only experiment with object classes of Chair and Table. Here we conduct the comparison qualitatively. Specifically, we show the generated explicit 3D shapes with the corresponding text prompts. Since TITG3SG [25] generates 3D shapes in two steps, where the coarse voxel and refined mesh are produced separately. We show them both in the top row of Fig. 6. The second row presents our generated mesh in two different views, in which we generate final 3D results in one shot. We observe our method gives better 3D textured shapes than TITG3SG, which is a fully supervised method. For example, in the third generated sample of TITG3SG, the chair legs are distorted. Moreover, the high alignment between input prompts and our generated results demonstrates our text control ability." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16809" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 516, + 186 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 516, + 186 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 516, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 516, + 186 + ], + "type": "image", + "image_path": "a26605101e82a9003bb56173ac0c12a9e6fdf5eab9071f71e5942fea2dead3b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "type": "text", + "content": "Figure 6. We compare the text-guided generation results of our methods with that of the fully supervised method TITG3SG [25] on two object classes, Chair and Table. On the top row, we present the results from [25], for each text prompt, the top left one is the raw voxel output in " + }, + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "type": "text", + "content": " resolution, and the top right one is the refined mesh based on the voxel output. The bottom row shows our mesh output with two different views. The renderings are generated with ChimeraX [10]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 240, + 187, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 187, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 187, + 253 + ], + "type": "text", + "content": "4.3.2 Quantitative comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 259, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 307 + ], + "type": "text", + "content": "To demonstrate our 3D generation quality, we compare existing works [25,43,50] from three aspects: 1) rendered 2D image quality using FID, 2) text-image relevance using R-Precision, and 3) 3D geometry quality using FPD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 450 + ], + "type": "text", + "content": "In Tab. 1, we provide a quantitative comparison with CLIP-NeRF [50], which utilizes human-written captions for training. We employ the Fréchet Inception Distance (FID) [12] as our evaluation metric and present results for two object classes: Car and Chair. To ensure fair comparisons, we downsample our results to match the resolution of CLIP-NeRF [50]. CLIP-NeRF employs NeRF as the 3D generator, which imposes limitations on the generation resolution. In contrast, our backbone GET3D model boasts a larger capacity, supporting higher resolution. Experimental results demonstrate that our method surpasses CLIP-NeRF in text-guided generation quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 582 + ], + "type": "text", + "content": "In Tab. 2, we adopt the CLIP-R-Precision metric [32] to evaluate the consistency between input text prompts and generated shapes. Given that our model is category-specific, we perform per-category CLIP-R-Precision and compare it to the fully supervised TITG3SG method [25]. According to [32], we compute the CLIP-R-Precision using a set of unseen random captions with words sampled from existing captions, thereby enhancing correlation with human evaluations and mitigating model bias inherent compared to the conventional R-Precision metric. It is observed our method outperforms TITG3SG [25] in each category." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 583, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 643 + ], + "type": "text", + "content": "In Tab. 3, we assess the geometry generation quality using the Fréchet Point Distance (FPD) [19, 47], which measures the feature distance of a PointNet [35] pretrained on ModelNet40 [51]. Our method outperforms existing approaches in terms of geometry quality by a large margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 200, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 200, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 200, + 671 + ], + "type": "text", + "content": "4.3.3 Inference speed comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In Tab. 4, we compare the inference speed of our method with other works. The optimization-based methods DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18]" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 283, + 535, + 350 + ], + "blocks": [ + { + "bbox": [ + 304, + 239, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 282 + ], + "type": "text", + "content": "Table 1. Comparison with the existing work. We evaluate the rendered 2D images using Fréchet inception distance (FID). We downsample our result to the same resolution of CLIP-NeRF [50] for fair comparisons." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 283, + 535, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 283, + 535, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 535, + 350 + ], + "type": "table", + "html": "
CarChair
ResolutionFIDResolutionFID
CLIP-NeRF [50]256267.8128248.4
Ours256220.1128243.7
Ours1024221.71024244.8
", + "image_path": "3a0a825d3ebb2fb2d571fb8f9a17327796943a7824c33a111a0340efbe1a2f98.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 334, + 378, + 514, + 422 + ], + "blocks": [ + { + "bbox": [ + 305, + 356, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 378 + ], + "type": "text", + "content": "Table 2. Per category CLIP-R-Precision score using random generated captions on categories Chair and Table." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 378, + 514, + 422 + ], + "lines": [ + { + "bbox": [ + 334, + 378, + 514, + 422 + ], + "spans": [ + { + "bbox": [ + 334, + 378, + 514, + 422 + ], + "type": "table", + "html": "
MethodChairTable
TITG3SG [25]9.05 ± 1.926.67 ± 3.25
Ours20.19 ± 1.8915.21 ± 3.15
", + "image_path": "92cce2d5317a09c1fc141dc0faf2a5595aaf05989377567b8b45a3e22a8690f1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 351, + 438, + 499, + 491 + ], + "blocks": [ + { + "bbox": [ + 315, + 426, + 535, + 437 + ], + "lines": [ + { + "bbox": [ + 315, + 426, + 535, + 437 + ], + "spans": [ + { + "bbox": [ + 315, + 426, + 535, + 437 + ], + "type": "text", + "content": "Table 3. Comparison of 3D generation quality in FPD score." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 438, + 499, + 491 + ], + "lines": [ + { + "bbox": [ + 351, + 438, + 499, + 491 + ], + "spans": [ + { + "bbox": [ + 351, + 438, + 499, + 491 + ], + "type": "table", + "html": "
MethodChairTable
TITG3SG [25]1566.761639.68
CLIP-Forge [43]825.963051.31
Ours342.231468.43
", + "image_path": "829ac951fe35224096c1715e9da4260edfa4eebe7228f55a474c7cd358342ea3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 309, + 550, + 543, + 645 + ], + "blocks": [ + { + "bbox": [ + 304, + 495, + 545, + 550 + ], + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 550 + ], + "type": "text", + "content": "Table 4. The comparison of inference time per text prompt for different methods. The inference time of DreamFields [13], DreamFusion [34], and PureCLIPNeRF [18] are given in their papers. For TITG3SG [25] and ours, we generate 1000 text prompts with batch size as 1, and calculate the average inference time." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 550, + 543, + 645 + ], + "lines": [ + { + "bbox": [ + 309, + 550, + 543, + 645 + ], + "spans": [ + { + "bbox": [ + 309, + 550, + 543, + 645 + ], + "type": "table", + "html": "
MethodDeviceOutputTime
DreamFields [13]TPU cores x8Rendering72 min
DreamFusion [34]TPUv4 machineRendering90 min
PureCLIPNeRF [18]GTX 2080tiRendering20 min
TITG3SG [25]Telsa V100-32GVoxel2.21 sec
TITG3SG [25]Telsa V100-32GMesh24.44 sec
OursTelsa V100-32GRendering0.05 sec
OursTelsa V100-32GMesh7.09 sec
", + "image_path": "53c1fb399f3fc519a409ec1da7e71a12819b51300af3df511567f76594068585.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "take tens of minutes to optimize for each input text prompt. Liu et al. [25] use an encoder-decoder architecture and thus can perform inference at 2.21 seconds. However, the raw " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " voxel output is very coarse, as shown in Fig. 6. The" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16810" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 86, + 106, + 125 + ], + "blocks": [ + { + "bbox": [ + 59, + 86, + 106, + 125 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 106, + 125 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 106, + 125 + ], + "type": "image", + "image_path": "3a1b23870babc5dc897131f96736b6ca1f9857507acdcea6c3ffac4a67492bc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 87, + 168, + 125 + ], + "blocks": [ + { + "bbox": [ + 120, + 87, + 168, + 125 + ], + "lines": [ + { + "bbox": [ + 120, + 87, + 168, + 125 + ], + "spans": [ + { + "bbox": [ + 120, + 87, + 168, + 125 + ], + "type": "image", + "image_path": "214e06bdaeed14a83587464862d93025b28c2dcac8e6709ba5e6ee9ee04b19b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 129, + 194, + 140 + ], + "lines": [ + { + "bbox": [ + 141, + 129, + 194, + 140 + ], + "spans": [ + { + "bbox": [ + 141, + 129, + 194, + 140 + ], + "type": "text", + "content": "\"a blue SUV\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 87, + 227, + 125 + ], + "blocks": [ + { + "bbox": [ + 182, + 87, + 227, + 125 + ], + "lines": [ + { + "bbox": [ + 182, + 87, + 227, + 125 + ], + "spans": [ + { + "bbox": [ + 182, + 87, + 227, + 125 + ], + "type": "image", + "image_path": "edff6ab99cc5e40d1106ac6e4d0994c66561205906593d45b1f432e024e26ab2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 232, + 87, + 280, + 125 + ], + "blocks": [ + { + "bbox": [ + 232, + 87, + 280, + 125 + ], + "lines": [ + { + "bbox": [ + 232, + 87, + 280, + 125 + ], + "spans": [ + { + "bbox": [ + 232, + 87, + 280, + 125 + ], + "type": "image", + "image_path": "58ccc99d37b1dca0382cd1ba4dba6faee5094f3c3994f66bea8d0125c2ad7a68.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 144, + 108, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 144, + 108, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 144, + 108, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 144, + 108, + 186 + ], + "type": "image", + "image_path": "eded246d9e8aa73d422f8e0882e7586a966445a068d7c6a6d6795fa1a09e8e70.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 121, + 148, + 167, + 183 + ], + "blocks": [ + { + "bbox": [ + 121, + 148, + 167, + 183 + ], + "lines": [ + { + "bbox": [ + 121, + 148, + 167, + 183 + ], + "spans": [ + { + "bbox": [ + 121, + 148, + 167, + 183 + ], + "type": "image", + "image_path": "ef9afded5d387628fd364d2c4cd56fae90b334f0f9c5cddf3731d9dd3f89c94d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 204, + 201 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 204, + 201 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 204, + 201 + ], + "type": "text", + "content": "\"a red hatchback\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 180, + 148, + 225, + 183 + ], + "blocks": [ + { + "bbox": [ + 180, + 148, + 225, + 183 + ], + "lines": [ + { + "bbox": [ + 180, + 148, + 225, + 183 + ], + "spans": [ + { + "bbox": [ + 180, + 148, + 225, + 183 + ], + "type": "image", + "image_path": "56a667ddbf78bb4336889deaaa17edfb52b124a8705985db8b10e246c62eec8b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 234, + 149, + 279, + 185 + ], + "blocks": [ + { + "bbox": [ + 234, + 149, + 279, + 185 + ], + "lines": [ + { + "bbox": [ + 234, + 149, + 279, + 185 + ], + "spans": [ + { + "bbox": [ + 234, + 149, + 279, + 185 + ], + "type": "image", + "image_path": "8541ce9317e965bf6cdebc994ca1977d5bc89a6eaf0ac595ad5e6bab6a9ec569.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 214, + 107, + 246 + ], + "blocks": [ + { + "bbox": [ + 59, + 214, + 107, + 246 + ], + "lines": [ + { + "bbox": [ + 59, + 214, + 107, + 246 + ], + "spans": [ + { + "bbox": [ + 59, + 214, + 107, + 246 + ], + "type": "image", + "image_path": "f48239947a516ed8281391162e8b735a1991902490a7b1709a690f404f4f28d3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 117, + 213, + 162, + 244 + ], + "blocks": [ + { + "bbox": [ + 117, + 213, + 162, + 244 + ], + "lines": [ + { + "bbox": [ + 117, + 213, + 162, + 244 + ], + "spans": [ + { + "bbox": [ + 117, + 213, + 162, + 244 + ], + "type": "image", + "image_path": "c6dba14818a767c762bc1c615c352eaa3e265e07b29184724732082e23352fd9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 256, + 212, + 267 + ], + "lines": [ + { + "bbox": [ + 127, + 256, + 212, + 267 + ], + "spans": [ + { + "bbox": [ + 127, + 256, + 212, + 267 + ], + "type": "text", + "content": "\"a yellow racing car\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 176, + 214, + 223, + 247 + ], + "blocks": [ + { + "bbox": [ + 176, + 214, + 223, + 247 + ], + "lines": [ + { + "bbox": [ + 176, + 214, + 223, + 247 + ], + "spans": [ + { + "bbox": [ + 176, + 214, + 223, + 247 + ], + "type": "image", + "image_path": "8e201e724c96f4162c4d70b1afad2ff6fa4621bd39a8ad34d9b08da5d095b334.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 231, + 215, + 277, + 247 + ], + "blocks": [ + { + "bbox": [ + 231, + 215, + 277, + 247 + ], + "lines": [ + { + "bbox": [ + 231, + 215, + 277, + 247 + ], + "spans": [ + { + "bbox": [ + 231, + 215, + 277, + 247 + ], + "type": "image", + "image_path": "6399f71a632cac11e6e3b17283d083c0cc5dba9b825eea9f2c8edd63a28d0300.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 63, + 272, + 99, + 313 + ], + "blocks": [ + { + "bbox": [ + 63, + 272, + 99, + 313 + ], + "lines": [ + { + "bbox": [ + 63, + 272, + 99, + 313 + ], + "spans": [ + { + "bbox": [ + 63, + 272, + 99, + 313 + ], + "type": "image", + "image_path": "fa86196d4cb1cb3f9954adda2ae59cfb0899ab568c5c0e0896fa8adb3547eda6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 121, + 275, + 156, + 313 + ], + "blocks": [ + { + "bbox": [ + 121, + 275, + 156, + 313 + ], + "lines": [ + { + "bbox": [ + 121, + 275, + 156, + 313 + ], + "spans": [ + { + "bbox": [ + 121, + 275, + 156, + 313 + ], + "type": "image", + "image_path": "6ff9b9d32a0cf3392703b2bb00814600bc0b41834526fb51e8e108ae5f6c6d47.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 318, + 228, + 329 + ], + "lines": [ + { + "bbox": [ + 114, + 318, + 228, + 329 + ], + "spans": [ + { + "bbox": [ + 114, + 318, + 228, + 329 + ], + "type": "text", + "content": "\"a yellow racing motorcycle\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 184, + 274, + 218, + 314 + ], + "blocks": [ + { + "bbox": [ + 184, + 274, + 218, + 314 + ], + "lines": [ + { + "bbox": [ + 184, + 274, + 218, + 314 + ], + "spans": [ + { + "bbox": [ + 184, + 274, + 218, + 314 + ], + "type": "image", + "image_path": "b75b8296f04efc925635434639a9a29bf1d35a590a270ca70da5cb5eefbc0c6d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 236, + 274, + 271, + 314 + ], + "blocks": [ + { + "bbox": [ + 236, + 274, + 271, + 314 + ], + "lines": [ + { + "bbox": [ + 236, + 274, + 271, + 314 + ], + "spans": [ + { + "bbox": [ + 236, + 274, + 271, + 314 + ], + "type": "image", + "image_path": "3717f6429816b70c471113303b3b685dc9883adcdf85fb54d9eb4d69f58db01c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 63, + 334, + 100, + 370 + ], + "blocks": [ + { + "bbox": [ + 63, + 334, + 100, + 370 + ], + "lines": [ + { + "bbox": [ + 63, + 334, + 100, + 370 + ], + "spans": [ + { + "bbox": [ + 63, + 334, + 100, + 370 + ], + "type": "image", + "image_path": "2c5444d8ab1be15714f6af8f158bf8034ff217e8b199d136d3337dfa425331df.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 395, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 546, + 418 + ], + "type": "text", + "content": "Figure 7. We show the generation diversity and text control of our method. Each row takes the same text prompt with different sampled random noise as the model input." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 121, + 334, + 157, + 370 + ], + "blocks": [ + { + "bbox": [ + 121, + 334, + 157, + 370 + ], + "lines": [ + { + "bbox": [ + 121, + 334, + 157, + 370 + ], + "spans": [ + { + "bbox": [ + 121, + 334, + 157, + 370 + ], + "type": "image", + "image_path": "0a504c1a2ce7714b3c31784ba414f208110df88e6c338551072bec69c55e8050.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 374, + 198, + 384 + ], + "lines": [ + { + "bbox": [ + 126, + 374, + 198, + 384 + ], + "spans": [ + { + "bbox": [ + 126, + 374, + 198, + 384 + ], + "type": "text", + "content": "\"a red motorbike\"" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 180, + 335, + 217, + 370 + ], + "blocks": [ + { + "bbox": [ + 180, + 335, + 217, + 370 + ], + "lines": [ + { + "bbox": [ + 180, + 335, + 217, + 370 + ], + "spans": [ + { + "bbox": [ + 180, + 335, + 217, + 370 + ], + "type": "image", + "image_path": "48bde7ece2ef1a61c57ff0b227a2c0ce85e743865660cd39a43aaf68080a2dff.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 236, + 330, + 272, + 370 + ], + "blocks": [ + { + "bbox": [ + 236, + 330, + 272, + 370 + ], + "lines": [ + { + "bbox": [ + 236, + 330, + 272, + 370 + ], + "spans": [ + { + "bbox": [ + 236, + 330, + 272, + 370 + ], + "type": "image", + "image_path": "3f10bb1b7e7d35fbb88c837e625718343c0f1a3ae78764dbdd7f0318dba7f7a0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 303, + 83, + 347, + 137 + ], + "blocks": [ + { + "bbox": [ + 303, + 83, + 347, + 137 + ], + "lines": [ + { + "bbox": [ + 303, + 83, + 347, + 137 + ], + "spans": [ + { + "bbox": [ + 303, + 83, + 347, + 137 + ], + "type": "image", + "image_path": "d7609661119dc0896a8d575b3614efcb3103d4a4281823c6f3e0df1ec03eba42.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 365, + 83, + 400, + 134 + ], + "blocks": [ + { + "bbox": [ + 365, + 83, + 400, + 134 + ], + "lines": [ + { + "bbox": [ + 365, + 83, + 400, + 134 + ], + "spans": [ + { + "bbox": [ + 365, + 83, + 400, + 134 + ], + "type": "image", + "image_path": "f1ee4643d9dfac1ad7dc9001e9ff52e6e23d8a11391bc868bc3be330322c3d2d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 422, + 83, + 456, + 134 + ], + "blocks": [ + { + "bbox": [ + 422, + 83, + 456, + 134 + ], + "lines": [ + { + "bbox": [ + 422, + 83, + 456, + 134 + ], + "spans": [ + { + "bbox": [ + 422, + 83, + 456, + 134 + ], + "type": "image", + "image_path": "ecdd8747f4766720a015829bea424167f28343546e9f6b9fcdab4f1950f76102.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 302, + 141, + 345, + 194 + ], + "blocks": [ + { + "bbox": [ + 302, + 141, + 345, + 194 + ], + "lines": [ + { + "bbox": [ + 302, + 141, + 345, + 194 + ], + "spans": [ + { + "bbox": [ + 302, + 141, + 345, + 194 + ], + "type": "image", + "image_path": "87ce7ea8c4cb3e8e58d6e409e593cbb48d717015a4b2ef2dc0b39ae37ddfdc4b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 363, + 141, + 399, + 192 + ], + "blocks": [ + { + "bbox": [ + 363, + 141, + 399, + 192 + ], + "lines": [ + { + "bbox": [ + 363, + 141, + 399, + 192 + ], + "spans": [ + { + "bbox": [ + 363, + 141, + 399, + 192 + ], + "type": "image", + "image_path": "6ab7114968b1bedd42ef98ceb4f7bb6f731b2b50e66e1a27c810852196f35a9a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 191, + 450, + 202 + ], + "lines": [ + { + "bbox": [ + 361, + 191, + 450, + 202 + ], + "spans": [ + { + "bbox": [ + 361, + 191, + 450, + 202 + ], + "type": "text", + "content": "\"agrayarmlesschair\"" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 422, + 141, + 456, + 192 + ], + "blocks": [ + { + "bbox": [ + 422, + 141, + 456, + 192 + ], + "lines": [ + { + "bbox": [ + 422, + 141, + 456, + 192 + ], + "spans": [ + { + "bbox": [ + 422, + 141, + 456, + 192 + ], + "type": "image", + "image_path": "0c7f1bf21c71b346cdc59a99f252e6d8da8539e2cef8a801c640d0af4b7a900f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 307, + 206, + 337, + 256 + ], + "blocks": [ + { + "bbox": [ + 307, + 206, + 337, + 256 + ], + "lines": [ + { + "bbox": [ + 307, + 206, + 337, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 337, + 256 + ], + "type": "image", + "image_path": "55d0a9969116303f0c59ccd0d5ac931693d5f40020bdea9e7fb1c062ab7f5e62.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 356, + 206, + 399, + 258 + ], + "blocks": [ + { + "bbox": [ + 356, + 206, + 399, + 258 + ], + "lines": [ + { + "bbox": [ + 356, + 206, + 399, + 258 + ], + "spans": [ + { + "bbox": [ + 356, + 206, + 399, + 258 + ], + "type": "image", + "image_path": "61436fb6d9ae7b707e943bb5eb78cbebba944da9bd7645debe72a86d4dd1dd5f.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 258, + 440, + 269 + ], + "lines": [ + { + "bbox": [ + 378, + 258, + 440, + 269 + ], + "spans": [ + { + "bbox": [ + 378, + 258, + 440, + 269 + ], + "type": "text", + "content": "\"agreenchair\"" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 421, + 206, + 457, + 258 + ], + "blocks": [ + { + "bbox": [ + 421, + 206, + 457, + 258 + ], + "lines": [ + { + "bbox": [ + 421, + 206, + 457, + 258 + ], + "spans": [ + { + "bbox": [ + 421, + 206, + 457, + 258 + ], + "type": "image", + "image_path": "ae3808d742a0176ed08b9ff46f488081d80ed60c56c7f58dbdafa9031bb7ddb4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 471, + 85, + 519, + 137 + ], + "blocks": [ + { + "bbox": [ + 471, + 85, + 519, + 137 + ], + "lines": [ + { + "bbox": [ + 471, + 85, + 519, + 137 + ], + "spans": [ + { + "bbox": [ + 471, + 85, + 519, + 137 + ], + "type": "image", + "image_path": "d984dd161ffa86bbe648beb7a77a35a03b5899e2c0915474ede145863aa28088.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 480, + 142, + 514, + 194 + ], + "blocks": [ + { + "bbox": [ + 480, + 142, + 514, + 194 + ], + "lines": [ + { + "bbox": [ + 480, + 142, + 514, + 194 + ], + "spans": [ + { + "bbox": [ + 480, + 142, + 514, + 194 + ], + "type": "image", + "image_path": "dcd2f436a5274261f889d67dc87eb795f098949fd882927f5f563f66b6415fc7.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 294, + 271, + 343, + 312 + ], + "blocks": [ + { + "bbox": [ + 294, + 271, + 343, + 312 + ], + "lines": [ + { + "bbox": [ + 294, + 271, + 343, + 312 + ], + "spans": [ + { + "bbox": [ + 294, + 271, + 343, + 312 + ], + "type": "image", + "image_path": "0a304aa388f185a7249b6020135e6e9b99b6c1787e3d42417d857084e1dc567a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 351, + 274, + 402, + 311 + ], + "blocks": [ + { + "bbox": [ + 351, + 274, + 402, + 311 + ], + "lines": [ + { + "bbox": [ + 351, + 274, + 402, + 311 + ], + "spans": [ + { + "bbox": [ + 351, + 274, + 402, + 311 + ], + "type": "image", + "image_path": "3f57037c3d6098c55afe9912ebc8fbf53b42b79ff56585e0ff49446b496f489f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 413, + 274, + 459, + 311 + ], + "blocks": [ + { + "bbox": [ + 413, + 274, + 459, + 311 + ], + "lines": [ + { + "bbox": [ + 413, + 274, + 459, + 311 + ], + "spans": [ + { + "bbox": [ + 413, + 274, + 459, + 311 + ], + "type": "image", + "image_path": "e986e0072a32f0a06f4006085075158792a8c6c2a76a57a162d968e8dca14293.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 470, + 206, + 520, + 258 + ], + "blocks": [ + { + "bbox": [ + 470, + 206, + 520, + 258 + ], + "lines": [ + { + "bbox": [ + 470, + 206, + 520, + 258 + ], + "spans": [ + { + "bbox": [ + 470, + 206, + 520, + 258 + ], + "type": "image", + "image_path": "8c8a27f3c506847b373da38df7aff79cd649a23c148a6fb3bcfaa887883efa43.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 467, + 274, + 520, + 311 + ], + "blocks": [ + { + "bbox": [ + 467, + 274, + 520, + 311 + ], + "lines": [ + { + "bbox": [ + 467, + 274, + 520, + 311 + ], + "spans": [ + { + "bbox": [ + 467, + 274, + 520, + 311 + ], + "type": "image", + "image_path": "9b3cf9233e942161b2dacb19c7fa10774c7854d9ad3161f2583210b79f297b36.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 294, + 331, + 345, + 371 + ], + "blocks": [ + { + "bbox": [ + 294, + 331, + 345, + 371 + ], + "lines": [ + { + "bbox": [ + 294, + 331, + 345, + 371 + ], + "spans": [ + { + "bbox": [ + 294, + 331, + 345, + 371 + ], + "type": "image", + "image_path": "cf0aa2c2aa13c6eb3d9ad470e1c84c1c1bb06d9d97bdeb145733ccfac5d07082.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 351, + 332, + 402, + 371 + ], + "blocks": [ + { + "bbox": [ + 372, + 317, + 450, + 328 + ], + "lines": [ + { + "bbox": [ + 372, + 317, + 450, + 328 + ], + "spans": [ + { + "bbox": [ + 372, + 317, + 450, + 328 + ], + "type": "text", + "content": "\"a red dining table\"" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 332, + 402, + 371 + ], + "lines": [ + { + "bbox": [ + 351, + 332, + 402, + 371 + ], + "spans": [ + { + "bbox": [ + 351, + 332, + 402, + 371 + ], + "type": "image", + "image_path": "3bacb4b37a17e1ae89c902e02d476d7661e41b116482370a209d18f8b4106f6c.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 372, + 457, + 382 + ], + "lines": [ + { + "bbox": [ + 365, + 372, + 457, + 382 + ], + "spans": [ + { + "bbox": [ + 365, + 372, + 457, + 382 + ], + "type": "text", + "content": "\"a silver metallic table\"" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 410, + 332, + 462, + 367 + ], + "blocks": [ + { + "bbox": [ + 410, + 332, + 462, + 367 + ], + "lines": [ + { + "bbox": [ + 410, + 332, + 462, + 367 + ], + "spans": [ + { + "bbox": [ + 410, + 332, + 462, + 367 + ], + "type": "image", + "image_path": "6454e7e27334ed49ea3be4a0687aecd664489cbeb25263f0c29e859ba09ecb0b.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 469, + 333, + 520, + 367 + ], + "blocks": [ + { + "bbox": [ + 469, + 333, + 520, + 367 + ], + "lines": [ + { + "bbox": [ + 469, + 333, + 520, + 367 + ], + "spans": [ + { + "bbox": [ + 469, + 333, + 520, + 367 + ], + "type": "image", + "image_path": "d7573347e3ed9cedfe310f5ea8e937b3651d42520bc892685df5e6e9eca4b75a.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 62, + 429, + 100, + 451 + ], + "blocks": [ + { + "bbox": [ + 62, + 429, + 100, + 451 + ], + "lines": [ + { + "bbox": [ + 62, + 429, + 100, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 100, + 451 + ], + "type": "image", + "image_path": "3643b402ba20c8916ad4ad52bf02c7df9d6d6fb0d009a5451ed46e11eca1f0a0.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 107, + 430, + 144, + 450 + ], + "blocks": [ + { + "bbox": [ + 107, + 430, + 144, + 450 + ], + "lines": [ + { + "bbox": [ + 107, + 430, + 144, + 450 + ], + "spans": [ + { + "bbox": [ + 107, + 430, + 144, + 450 + ], + "type": "image", + "image_path": "7608b1350c6c1f3a294010a047daf196b0d0c1d675438a8d9da81519d024479c.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 151, + 430, + 187, + 450 + ], + "blocks": [ + { + "bbox": [ + 151, + 430, + 187, + 450 + ], + "lines": [ + { + "bbox": [ + 151, + 430, + 187, + 450 + ], + "spans": [ + { + "bbox": [ + 151, + 430, + 187, + 450 + ], + "type": "image", + "image_path": "37038c41b49cd9beda1ac903ba2afeeebca34cdff84ab236a0b7235c2c008157.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 196, + 430, + 231, + 450 + ], + "blocks": [ + { + "bbox": [ + 196, + 430, + 231, + 450 + ], + "lines": [ + { + "bbox": [ + 196, + 430, + 231, + 450 + ], + "spans": [ + { + "bbox": [ + 196, + 430, + 231, + 450 + ], + "type": "image", + "image_path": "ad989530d6a8c339a54c11ffa8975b30c5f387fdc9392994649aa616aa87be06.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 239, + 430, + 276, + 450 + ], + "blocks": [ + { + "bbox": [ + 239, + 430, + 276, + 450 + ], + "lines": [ + { + "bbox": [ + 239, + 430, + 276, + 450 + ], + "spans": [ + { + "bbox": [ + 239, + 430, + 276, + 450 + ], + "type": "image", + "image_path": "eaba7cb8d453835a6cab9c519b09897f7d22bc1d9c64aa8965875ded873a711f.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 62, + 462, + 96, + 502 + ], + "blocks": [ + { + "bbox": [ + 62, + 462, + 96, + 502 + ], + "lines": [ + { + "bbox": [ + 62, + 462, + 96, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 462, + 96, + 502 + ], + "type": "image", + "image_path": "9dbfc68e7cc4d5d38097902f5ccb3162f8799b7147c2cd47a095e0931d9baf8f.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 108, + 463, + 137, + 501 + ], + "blocks": [ + { + "bbox": [ + 108, + 463, + 137, + 501 + ], + "lines": [ + { + "bbox": [ + 108, + 463, + 137, + 501 + ], + "spans": [ + { + "bbox": [ + 108, + 463, + 137, + 501 + ], + "type": "image", + "image_path": "a13f994c575b76ecd3a24fdc29c64d37773f7e1050e5ec6d07bbece13647562f.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 156, + 463, + 178, + 500 + ], + "blocks": [ + { + "bbox": [ + 156, + 463, + 178, + 500 + ], + "lines": [ + { + "bbox": [ + 156, + 463, + 178, + 500 + ], + "spans": [ + { + "bbox": [ + 156, + 463, + 178, + 500 + ], + "type": "image", + "image_path": "0fdfe74916a25d18e386930f44abc70b3a545246c1c31a18bb87ed41c3ea7274.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 196, + 462, + 225, + 501 + ], + "blocks": [ + { + "bbox": [ + 196, + 462, + 225, + 501 + ], + "lines": [ + { + "bbox": [ + 196, + 462, + 225, + 501 + ], + "spans": [ + { + "bbox": [ + 196, + 462, + 225, + 501 + ], + "type": "image", + "image_path": "1188f55a4c568cf4b911b82a79698c859f963a6716f8c9fdd5548af3fa5fdadd.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 240, + 463, + 268, + 501 + ], + "blocks": [ + { + "bbox": [ + 240, + 463, + 268, + 501 + ], + "lines": [ + { + "bbox": [ + 240, + 463, + 268, + 501 + ], + "spans": [ + { + "bbox": [ + 240, + 463, + 268, + 501 + ], + "type": "image", + "image_path": "cf100d6544b188189ffc190003af17a3f9c449f7b49bf39e28568529c31597af.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 61, + 513, + 99, + 538 + ], + "blocks": [ + { + "bbox": [ + 61, + 513, + 99, + 538 + ], + "lines": [ + { + "bbox": [ + 61, + 513, + 99, + 538 + ], + "spans": [ + { + "bbox": [ + 61, + 513, + 99, + 538 + ], + "type": "image", + "image_path": "64f04db05a9e3c07045b5ecee724a241779d4e14bf4b79eb78991d0935232b39.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 102, + 512, + 144, + 538 + ], + "blocks": [ + { + "bbox": [ + 102, + 512, + 144, + 538 + ], + "lines": [ + { + "bbox": [ + 102, + 512, + 144, + 538 + ], + "spans": [ + { + "bbox": [ + 102, + 512, + 144, + 538 + ], + "type": "image", + "image_path": "1cdc7f32f01adf592a78141e19a5170ad54eb071c03ac18de3dc1a72d7dd315d.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 149, + 512, + 187, + 540 + ], + "blocks": [ + { + "bbox": [ + 149, + 512, + 187, + 540 + ], + "lines": [ + { + "bbox": [ + 149, + 512, + 187, + 540 + ], + "spans": [ + { + "bbox": [ + 149, + 512, + 187, + 540 + ], + "type": "image", + "image_path": "746b9a72f10d9bce6dcb331a629dcd22f7f4ae9a949fef4b44c10f03080a60e8.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 192, + 512, + 231, + 540 + ], + "blocks": [ + { + "bbox": [ + 192, + 512, + 231, + 540 + ], + "lines": [ + { + "bbox": [ + 192, + 512, + 231, + 540 + ], + "spans": [ + { + "bbox": [ + 192, + 512, + 231, + 540 + ], + "type": "image", + "image_path": "ff0a8a2744f862b14fb080a8d442f3fb88db7891bf89748240560faecbece891.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 238, + 512, + 276, + 539 + ], + "blocks": [ + { + "bbox": [ + 238, + 512, + 276, + 539 + ], + "lines": [ + { + "bbox": [ + 238, + 512, + 276, + 539 + ], + "spans": [ + { + "bbox": [ + 238, + 512, + 276, + 539 + ], + "type": "image", + "image_path": "ff3fc6719ea660eee7d6d617819f13ff6b6d1bb58e7077cfee63e70e1cfbac74.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 62, + 544, + 88, + 570 + ], + "blocks": [ + { + "bbox": [ + 62, + 544, + 88, + 570 + ], + "lines": [ + { + "bbox": [ + 62, + 544, + 88, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 88, + 570 + ], + "type": "image", + "image_path": "0c190a8d91e5e0bed32c5b35968049a471c5ad1da6ea538c721a0236a5e3df7a.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 573, + 93, + 581 + ], + "lines": [ + { + "bbox": [ + 56, + 573, + 93, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 573, + 93, + 581 + ], + "type": "text", + "content": "\"ared{class}\"" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 587, + 286, + 609 + ], + "lines": [ + { + "bbox": [ + 47, + 587, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 286, + 609 + ], + "type": "text", + "content": "Figure 8. Text control of textures on four different categories. Each row shares the same sampled random noise input." + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 107, + 544, + 132, + 570 + ], + "blocks": [ + { + "bbox": [ + 107, + 544, + 132, + 570 + ], + "lines": [ + { + "bbox": [ + 107, + 544, + 132, + 570 + ], + "spans": [ + { + "bbox": [ + 107, + 544, + 132, + 570 + ], + "type": "image", + "image_path": "7eb3bd917cb8776d698cdac89beebbbdc6eff12240991edc0d378b5de1610916.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 573, + 140, + 581 + ], + "lines": [ + { + "bbox": [ + 97, + 573, + 140, + 581 + ], + "spans": [ + { + "bbox": [ + 97, + 573, + 140, + 581 + ], + "type": "text", + "content": "\"a yellow {class}\"" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 151, + 544, + 178, + 570 + ], + "blocks": [ + { + "bbox": [ + 151, + 544, + 178, + 570 + ], + "lines": [ + { + "bbox": [ + 151, + 544, + 178, + 570 + ], + "spans": [ + { + "bbox": [ + 151, + 544, + 178, + 570 + ], + "type": "image", + "image_path": "596cb04ec69713b9445ed85fb327a9d3191c463cdc924790aecc6012be5f028f.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 573, + 187, + 581 + ], + "lines": [ + { + "bbox": [ + 145, + 573, + 187, + 581 + ], + "spans": [ + { + "bbox": [ + 145, + 573, + 187, + 581 + ], + "type": "text", + "content": "\"agreen{class}" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 196, + 544, + 221, + 570 + ], + "blocks": [ + { + "bbox": [ + 196, + 544, + 221, + 570 + ], + "lines": [ + { + "bbox": [ + 196, + 544, + 221, + 570 + ], + "spans": [ + { + "bbox": [ + 196, + 544, + 221, + 570 + ], + "type": "image", + "image_path": "79c91a0412a11ea75c96a94c7fa7c318dc67f0a158f267dd5b3967c31db44673.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 573, + 233, + 581 + ], + "lines": [ + { + "bbox": [ + 195, + 573, + 233, + 581 + ], + "spans": [ + { + "bbox": [ + 195, + 573, + 233, + 581 + ], + "type": "text", + "content": "\"a blue {class}\"" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 241, + 544, + 267, + 570 + ], + "blocks": [ + { + "bbox": [ + 241, + 544, + 267, + 570 + ], + "lines": [ + { + "bbox": [ + 241, + 544, + 267, + 570 + ], + "spans": [ + { + "bbox": [ + 241, + 544, + 267, + 570 + ], + "type": "image", + "image_path": "34d26746bdb3bcde65d7beac3cde946ef5c7d531a9f2e9f404ad883a86f850f2.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 573, + 279, + 581 + ], + "lines": [ + { + "bbox": [ + 240, + 573, + 279, + 581 + ], + "spans": [ + { + "bbox": [ + 240, + 573, + 279, + 581 + ], + "type": "text", + "content": "\"a gray {class}\"" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": "coarse output is later refined with interpolations and marching cubes to generate mesh data. This process takes ten times longer than the network inference time and costs 24.44 seconds per sample. Our method generates high-resolution renderings at " + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": " resolution with only 0.05 seconds and generates the complete explicit mesh representation in 7.09 seconds." + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 306, + 423, + 417, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 417, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 417, + 436 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 77 + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": "In Fig. 7, we present the generation diversity and text control strength of our method in four object classes. In each row, we sample different random noises as well as the given text prompt as input to generate diversified and text-controlled 3D shapes, where the input text features are simultaneously fed into both the geometry and texture mapping network. Generally, we observe the semantic meanings of the rendered 2D images, and the given prompts are consistent. Moreover, there exists large diversity in the texture and geometry of our generated 3D shapes, even from the same text input. Here we present generated results of 4 classes, where " + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "inline_equation", + "content": "Car" + }, + { + "bbox": [ + 304, + 450, + 546, + 715 + ], + "type": "text", + "content": " has a relatively simple geometry, while Table and Chair are more complex. But we still have good control over the textures and geometries for the challenging classes. For example, when we input \"a gray armless chair\", the generated samples exactly match the given instructions. However, we also observe failure cases of Table generation. For instance, the generation of \"a wooden office desk\" may have distorted geometry due to its complex structure. In Fig. 8, we show the text control ability of our method on different classes. The results in each row are generated from the same random noise input but with dif" + } + ] + } + ], + "index": 78 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16811" + } + ] + } + ], + "index": 79 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 103, + 264, + 148 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 102 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 102 + ], + "type": "text", + "content": "Table 5. We compare the FID score of using image CLIP features (Lafite [55]) and the pseudo text CLIP features (Ours) as training input." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 103, + 264, + 148 + ], + "lines": [ + { + "bbox": [ + 69, + 103, + 264, + 148 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 264, + 148 + ], + "type": "table", + "html": "
MethodCarChairTableMotorbike
Lafite [55]68.781.194.3146.5
Ours21.744.843.257.2
", + "image_path": "9de2d828a5ac7eeb5f60f28323b90d2523953aef9bf27f80649d0719142d47e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 178, + 269, + 244 + ], + "blocks": [ + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "type": "text", + "content": "Table 6. We compare the FID score for our method without " + }, + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 47, + 156, + 286, + 178 + ], + "type": "text", + "content": " and background augmentation (BGaug)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 178, + 269, + 244 + ], + "lines": [ + { + "bbox": [ + 63, + 178, + 269, + 244 + ], + "spans": [ + { + "bbox": [ + 63, + 178, + 269, + 244 + ], + "type": "table", + "html": "
MethodCarChairTableMotorbike
Ours21.744.843.257.2
-w/o Lclip83.7148.967.1134.9
-w/o Limg59.8144.248.4101.8
-w/o BGaug55.165.158.1110.0
", + "image_path": "e2066ec54d9853789e025666645403313a5145a394d4392e0cf1a181db644aec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 251, + 287, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 288 + ], + "type": "text", + "content": "ferent text prompts. To summarize, the results show that our generation results are highly consistent with the input text prompts across different object classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 297, + 140, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 140, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 140, + 310 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 316, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 519 + ], + "type": "text", + "content": "Pseudo caption generation module. In Tab. 5, we compare our method with Lafite [55], which is a language-free training method for the text-to-image generation task. Lafite claims due to the alignment between CLIP text and image embeddings. The CLIP image features can be directly used as text input during training. To implement Lafite [55] onto our framework, we feed the rendered real image CLIP features into our model and train the model by minimizing the cosine similarity between the generated and input image CLIP features. The quantitative results show that our proposed pseudo caption generation module significantly improves the generation quality compared to the baseline. We also give a qualitative comparison in Fig. 9 (b), we observe that the generator trained with Lafite [55] fails to generate correct textures. This problem might come from the modality gap between the CLIP image and text embeddings, as discussed in [20, 54]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": "Effects of our loss functions. We conduct an ablation study on our used two losses " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": ". In Tab. 6, we quantitatively validate the usefulness of both two losses. Specifically, " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": " supervises the text-image alignment learning, which is more critical than " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": ", and removing " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": " gives a more considerable performance drop. From the qualitative results in Fig. 9 (c), we observe that simply using " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": " leads to the model generating distorted geometries, which may indicate the significance of high-level semantic supervision. We also observe that solely using " + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 46, + 521, + 286, + 674 + ], + "type": "text", + "content": " (Fig. 9 (d)) may cause the model generation to lose fine-grained texture details, as well as the geometry information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Effects of background augmentation. Our used GET3D [9] backbone network generates RGBA images with dark backgrounds, which may bring some noise when" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 342, + 73, + 534, + 233 + ], + "blocks": [ + { + "bbox": [ + 317, + 83, + 327, + 92 + ], + "lines": [ + { + "bbox": [ + 317, + 83, + 327, + 92 + ], + "spans": [ + { + "bbox": [ + 317, + 83, + 327, + 92 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 73, + 534, + 233 + ], + "lines": [ + { + "bbox": [ + 342, + 73, + 534, + 233 + ], + "spans": [ + { + "bbox": [ + 342, + 73, + 534, + 233 + ], + "type": "image", + "image_path": "05a30e36ad4296ca32e197ddacb32424fb7993c74924a548eeca56c5ae6bf2aa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 303, + 531, + 357 + ], + "blocks": [ + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "type": "text", + "content": "Figure 9. Ablation study on our methods with text prompt \"a red car\". We show results of (a) our full model, (b) Lafite [55], (c) without high-level semantic regularization " + }, + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{clip}" + }, + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "type": "text", + "content": ", (d) without low-level image regularization " + }, + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img}" + }, + { + "bbox": [ + 304, + 238, + 545, + 303 + ], + "type": "text", + "content": ", (e) without background augmentation. All the visualizations are generated from models trained for 250k iterations." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 320, + 408, + 338 + ], + "lines": [ + { + "bbox": [ + 326, + 320, + 408, + 338 + ], + "spans": [ + { + "bbox": [ + 326, + 320, + 408, + 338 + ], + "type": "text", + "content": "\"acarwithyellowlegs, redseat and blue back.\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 430, + 303, + 531, + 357 + ], + "lines": [ + { + "bbox": [ + 430, + 303, + 531, + 357 + ], + "spans": [ + { + "bbox": [ + 430, + 303, + 531, + 357 + ], + "type": "image", + "image_path": "a74c124e3363581d94064041940b941d7f12c5160f11a032706385921f174bf0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 363, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 396 + ], + "type": "text", + "content": "Figure 10. Examples of failure cases of our method. Our model often fails to handle text input with composed descriptions for different parts of the object." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 401, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 449 + ], + "type": "text", + "content": "we conduct low-level image regularization. As illustrated from the quantitative results in Tab. 6 and the visualizations in Fig. 9 (e), training with the augmented background can help remove the texture flaws." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 378, + 471 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": "We introduce TAPS3D, a text-guided 3D textured shape generation method. We first generate pseudo captions for model training. During the inference phase, TAPS3D does not require additional optimization. Our fast generation speed enables average users to generate high-quality 3D objects with acceptable processing time." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 552, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 623 + ], + "type": "text", + "content": "**Limitation.** The limitation of our method mainly lies in the capacity of our model is still bound to the training image data, and we cannot produce different fine-grained details for different object parts, as shown in Fig. 10. Although we do not require paired text data, we still need diversified training images to handle complex text input." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 634, + 398, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 398, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 398, + 647 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "This research is partly supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG-RP-2018-003), and the MoE AcRF Tier 2 grant (MOE-T2EP20220-0007) and MoE AcRF Tier 1 grants (RG14/22, RG95/20)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16812" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, John Femiani, Niloy Mitra, and Peter Wonka. Clip2stylegan: Unsupervised extraction of stylegan edit directions. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 167 + ], + "type": "text", + "content": "[2] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O'Reilly Media Inc., 2009. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 288, + 235 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 402 + ], + "type": "text", + "content": "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "text", + "content": "[7] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 500 + ], + "type": "text", + "content": "[8] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecrafter: A recursive text-conditioned 3d shape generation model. arXiv preprint arXiv:2207.09446, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 557 + ], + "type": "text", + "content": "[9] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv preprint arXiv:2209.11163, 2022. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[10] Thomas D Goddard, Conrad C Huang, Elaine C Meng, Eric F Pettersen, Gregory S Couch, John H Morris, and Thomas E Ferrin. Ucsf chimerax: Meeting modern challenges in visualization and analysis. Protein Science, 27(1):14-25, 2018. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[11] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylererf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 285 + ], + "type": "text", + "content": "[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[17] Umut Kocasari, Alara Dirik, Mert Tiftikci, and Pinar Yanardag. Stylemc: Multi-channel based fast text-guided image generation and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 895–904, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 373 + ], + "type": "text", + "content": "[18] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 409 + ], + "type": "text", + "content": "[19] Ruihui Li, Xianzhi Li, Ka-Hei Hui, and Chi-Wing Fu. Spagan: Sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[20] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. 2, 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[21] Kangcheng Liu. Rm3d: Robust data-efficient 3d scene parsing via traditional and learnt 3d descriptors-based semantic region merging. International Journal of Computer Vision, pages 1-30, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[22] Kangcheng Liu, Zhi Gao, Feng Lin, and Ben M Chen. Fgnet: A fast and accurate framework for large-scale lidar point cloud understanding. IEEE Transactions on Cybernetics, 53(1):553-564, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 588 + ], + "type": "text", + "content": "[23] Kangcheng Liu, Aoran Xiao, Xiaoqin Zhang, Shijian Lu, and Ling Shao. Fac: 3d representation learning via foreground aware feature contrast. arXiv preprint arXiv:2303.06388, 2023.1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 645 + ], + "type": "text", + "content": "[24] Kangcheng Liu, Yuzhi Zhao, Qiang Nie, Zhi Gao, and Ben M Chen. Weakly supervised 3d scene segmentation with region-level boundary awareness and instance discrimination. In European Conference on Computer Vision 2022 (ECCV 2022), pages 37-55. Springer, Cham, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[25] Zhengzhe Liu, Yi Wang, Xiaojuan Qi, and Chi-Wing Fu. Towards implicit text-guided 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2022. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16813" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 194 + ], + "type": "text", + "content": "[28] Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schubert, and Chris Olah. Differentiable image parameterizations. Distill, 3(7):e12, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 248 + ], + "type": "text", + "content": "[29] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 303 + ], + "type": "text", + "content": "[30] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 369 + ], + "type": "text", + "content": "[31] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 425 + ], + "type": "text", + "content": "[32] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "text", + "content": "[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[36] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "eration with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[40] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 194, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 249 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 316 + ], + "type": "text", + "content": "[42] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 380 + ], + "type": "text", + "content": "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603-18613, 2022. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 436 + ], + "type": "text", + "content": "[44] Aditya Sanghi, Rao Fu, Vivian Liu, Karl Willis, Hooman Shayani, Amir Hosein Khasahmadi, Srinath Sridhar, and Daniel Ritchie. Textcraft: Zero-shot generation of high-fidelity and diverse shapes from text. arXiv preprint arXiv:2211.01427, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[45] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 481, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 545, + 535 + ], + "type": "text", + "content": "[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 536, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 591 + ], + "type": "text", + "content": "[47] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "text", + "content": "[48] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. 3d pose transfer with correspondence learning and mesh refinement. Advances in Neural Information Processing Systems, 34:3108-3120, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 680 + ], + "type": "text", + "content": "[49] Chaoyue Song, Jiacheng Wei, Ruibo Li, Fayao Liu, and Guosheng Lin. Unsupervised 3d pose transfer with cross consistency and dual reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[50] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16814" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 361 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[51] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "text", + "content": "[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "type": "text", + "content": "[54] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[55] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021. 3, 8" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16815" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_content_list.json b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c743625256d4ff77beef21fbd2d4782f1ae4503 --- /dev/null +++ b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_content_list.json @@ -0,0 +1,1502 @@ +[ + { + "type": "text", + "text": "TBP-Former: Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving", + "text_level": 1, + "bbox": [ + 125, + 130, + 844, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaoheng Fang $^{1*}$ Zi Wang $^{1*}$ Yiqi Zhong $^{2}$ Junhao Ge $^{1}$ Siheng Chen $^{1,3\\dagger}$", + "bbox": [ + 174, + 202, + 795, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Cooperative Medianet Innovation Center, Shanghai Jiao Tong University", + "bbox": [ + 192, + 222, + 774, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, University of Southern California $^{3}$ Shanghai AI Laboratory", + "bbox": [ + 109, + 239, + 877, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1{shfang，w4ngz1，cancaries，sihengc}@sjtu.edu.cn 2{yiqizhon}@usc.edu", + "bbox": [ + 174, + 257, + 794, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-centric joint perception and prediction (PnP) has become an emerging trend in autonomous driving research. It predicts the future states of the traffic participants in the surrounding environment from raw RGB images. However, it is still a critical challenge to synchronize features obtained at multiple camera views and timestamps due to inevitable geometric distortions and further exploit those spatial-temporal features. To address this issue, we propose a temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP, which includes two novel designs. First, a pose-synchronized BEV encoder is proposed to map raw image inputs with any camera pose at any time to a shared and synchronized BEV space for better spatial-temporal synchronization. Second, a spatial-temporal pyramid transformer is introduced to comprehensively extract multi-scale BEV features and predict future BEV states with the support of spatial priors. Extensive experiments on nuScenes dataset show that our proposed framework overall outperforms all state-of-the-art vision-based prediction methods. Code is available at: https://github.com/MediaBrain-SJTU/TBP-Former", + "bbox": [ + 75, + 340, + 473, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 699, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As one of the most fascinating engineering projects, autonomous driving has been an aspiration for many researchers and engineers for decades. Although significant progress has been made, it is still an open question in designing a practical solution to achieve the goal of full self-driving. A traditional and common solution consists of a sequential stack of perception, prediction, planning, and control. Despite the idea of divide-and-conquer having achieved tremendous success in developing software sys", + "bbox": [ + 75, + 724, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e14ef5292e54d78c4693f68ae053ebc277cef5b277c65fceee5ce62d613c6f82.jpg", + "image_caption": [ + "Previous works" + ], + "image_footnote": [], + "bbox": [ + 511, + 321, + 776, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/143dbc6f80836fda5ead3f6fb745c36c14d7a4a92e17a75a79885c92e1f1e081.jpg", + "image_caption": [ + "TBP-Former" + ], + "image_footnote": [], + "bbox": [ + 785, + 321, + 885, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d43e3d772a538bacefdb6cfc5eacc582b42d4a95beb85b679b115888f99fc7e.jpg", + "image_caption": [ + "(a) Temporal BEV feature alignment", + "Figure 1. Two major challenges in vision-based perception and prediction are (a) how to avoid distortion and deficiency when aggregating features across time and camera views; and (b) how to achieve spatial-temporal feature learning for prediction. Our Pose-Synchronized BEV Encoder can precisely map the visual features into synchronized BEV space, and Spatial-Temporal Pyramid Transformer extracts feature at multiple scales." + ], + "image_footnote": [], + "bbox": [ + 511, + 462, + 776, + 588 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/002bf9e04e7761f1d780ff6b49709ad83265ad64d6e1d315907695cf48154154.jpg", + "image_caption": [ + "(b) Spatial-temporal feature modeling for prediction" + ], + "image_footnote": [], + "bbox": [ + 785, + 462, + 888, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tems, a long stack could cause cascading failures in an autonomous system. Recently, there is a trend to combine multiple parts in an autonomous system to be a joint module, cutting down the stack. For example, [25, 46] consider joint perception and prediction and [5, 43] explore joint prediction and planning. This work focuses on joint perception and prediction.", + "bbox": [ + 496, + 744, + 893, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of joint perception and prediction (PnP) aims to predict the current and future states of the surrounding environment with the input of multi-frame raw sensor", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 94, + 875, + 351, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1368", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data. The output current and future states would directly serve as the input for motion planning. Recently, many PnP methods are proposed based on diverse sensor input choices. For example, [4, 25, 34] take multi-frame LiDAR point clouds as input and achieve encouraging 3D detection and trajectory prediction performances simultaneously. Recently, the rapid development of vision-centric methods offers a new possibility to provide a cheaper and easy-to-deploy solution for PnP. For instance, [1, 16, 17] only uses RGB images collected by multiple cameras to build PnP systems. Meanwhile, without precise 3D measurements, vision-centric PnP is more technically challenging. Therefore, this work aims to advance this direction.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core of vision-centric PnP is to learn appropriate spatial-temporal feature representations from temporal image sequences. It is a crux and difficult from three aspects. First, since the input and the output of vision-centric PnP are supported in camera front-view (FV) and bird's-eye-view (BEV) respectively, one has to deal with distortion issues during geometric transformation between two views. Second, when the vehicle is moving, the view of the image input is time-varying and it is thus nontrivial to precisely map visual features across time into a shared and synchronized space. Third, since information in temporal image sequences is sufficiently rich for humans to accurately perceive the environment, we need a powerful learning model to comprehensively exploit spatial-temporal features.", + "bbox": [ + 75, + 290, + 472, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these issues, previous works on vision-centric PnP consider diverse strategies. For example, [16, 56] follows the method in [38] to map FV features to BEV features, then synchronizes BEV features across time via rigid transformation, and finally uses a recurrent network to exploit spatial-temporal features. However, due to the image discretization nature and depth estimation uncertainty, simply relying on rigid geometric transformations would cause inevitable distortion; see Fig. 1. Some other work [49] transforms the pseudo feature point cloud to current ego coordinates and then pools the pseudo-lidar to BEV features; however, this approach encounters deficiency due to the limited sensing range in perception. Meanwhile, many works [16, 17, 56] simply employ recurrent neural networks to learn the temporal features from multiple BEV representations, which is hard to comprehensively extract spatial-temporal features.", + "bbox": [ + 75, + 505, + 472, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To promote more reliable and comprehensive feature learning across views and time, we propose the temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP. The proposed TBP-Former includes two key innovations: i) pose-synchronized BEV encoder, which leverages a pose-aware cross-attention mechanism to directly map a raw image input with any camera pose at any time to the corresponding feature map in a shared and synchronized BEV space; and ii) spatial-temporal pyra", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mid transformer, which leverages a pyramid architecture with Swin-transformer [28] blocks to learn comprehensive spatial-temporal features from sequential BEV maps at multiple scales and predict future BEV states with a set of future queries equipped with spatial priors.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to previous works, the proposed TBP-Former brings benefits from two aspects. First, previous works [16, 17, 24, 56] consider FV-to-BEV transformation and temporal synchronization as two separate steps, each of which could bring distortion due to discrete depth estimation and rigid transformation; while we merge them into one step and leverage both geometric transformation and attention-based learning ability to achieve spatial-temporal synchronization. Second, previous works [16, 53] use RNNs or 3D convolutions to learn spatial-temporal features; while we leverage a powerful pyramid transformer architecture to comprehensively capture spatial-temporal features, which makes prediction more effective.", + "bbox": [ + 496, + 167, + 892, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the main contributions of our work are:", + "bbox": [ + 519, + 364, + 880, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To tackle the distortion issues in mapping temporal image sequences to a synchronized BEV space, we propose a pose-synchronized BEV encoder (PoseSync BEV Encoder) based on cross-view attention mechanism to extract quality temporal BEV features.", + "- We propose a novel Spatial-Temporal Pyramid Transformer (STPT) to extract multi-scale spatial-temporal features from sequential BEV maps and predict future BEV states according to well-elaborated future queries integrated with spatial priors.", + "- Overall, we propose TBP-Former, a vision-based joint perception and prediction framework for autonomous driving. TBP-Former achieves state-of-the-art performance on nuScenes [2] dataset for the vision-based prediction task. Extensive experiments show that both PoseSync BEV Encoder and STPT contribute greatly to the performance. Due to the decoupling property of the framework, both proposed modules can be easily utilized as alternative modules in any vision-based BEV prediction framework." + ], + "bbox": [ + 517, + 388, + 892, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 731, + 640, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Joint Perception and Prediction", + "text_level": 1, + "bbox": [ + 500, + 756, + 779, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the two core system modules of autonomous driving, how to conduct perception and prediction tasks jointly has received a lot of attention. Traditional approaches [3,4,25,34,39] formulate this joint task as a trajectory prediction problem that relies on the perception outputs of 3D object detection and tracking. The dependency on intermediate results tends to accumulate errors and lacks the capacity to perceive unknown objects [52,53]. Subsequently,", + "bbox": [ + 496, + 780, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1369", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ecf270f42ebad10f4d13bcd27c06dd85e580b8c2d4aa88f99dd595d12d02bc97.jpg", + "image_caption": [ + "Figure 2. An overview of TBP-Former architecture. Taking consecutive surrounding camera images as inputs, TBP-Former first generates image-space features and uses the PoseSync BEV Encoder to map front-view features to BEV features in a shared and synchronized BEV space. Then the BEV features from multiple frames are processed by the Spatial-Temporal Pyramid Transformer to extract BEV spatial-temporal features and predict future BEV states in order. In this process, high-level scene representations are generated from the last frame BEV feature as spatial priors to guide the prediction. Finally, the well-predicted future states are sent to decoder heads for joint perception and prediction tasks." + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 880, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "instance-free methods [9,21,33,43,45,53] that predict dense future semantic occupancy and flow have become a growing trend to simplify the understanding of dynamic scenes. Also, several recent works [16, 17, 56] explore joint perception and prediction in the form of dense occupancy and flow using only surrounding camera input.", + "bbox": [ + 75, + 429, + 468, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In many previous works [4, 25, 39, 43], raster HD (high-definition) maps play an important role as input of the frameworks. HD maps can provide strong priors to guide the predicted results to follow the traffic lanes. However, in practice, HD maps are laborious and costly to produce and require frequent maintenance. Instead of using off-the-peg HD maps, we follow the philosophy of [5, 7, 23] in predicting online HD maps but propose to learn high-level scene geometry representations from real-time sensor inputs and take these representations as priors for the prediction task.", + "bbox": [ + 75, + 520, + 470, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. BEV Representations", + "text_level": 1, + "bbox": [ + 76, + 681, + 277, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BEV representations provide a unified and physical-interpretable way to represent the rich information of road, moving objects and occlusion in a traffic scene, which can be easily utilized for downstream tasks such as motion prediction, planning and control, etc. For camera-based methods, how to solve the problem of projecting features from perspective view to BEV is a major challenge. Some learnable methods use MLP [23, 36, 42] or transformer network [37, 57] to implicitly reason the relationship between two different views. LSS [38] proposes the approach of predicting depth distribution per pixel on 2D features, then 'lifting' the 2D features according to the corresponding depth distribution to BEV space. Numerous works, aim", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing at tasks of BEV perception [19, 20, 54], motion prediction [1, 16, 17, 56], lidar-camera fusion [26], etc., follow this form to generate BEV representations. Also, some methods [6, 13, 24] explicitly establish the correspondence from BEV location to image-view pixel using homography between image and BEV plane and achieve attractive performance in diverse tasks.", + "bbox": [ + 496, + 429, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, when dealing with temporal information, most methods [1, 16, 17, 24, 56] warp history BEV representations according to the variation of ego poses. Due to the pre-defined fixed range and size of the BEV grid, rotation and translation operations may cause distortion and out-of-range problems when aligning history BEV maps to current ego coordinates. Though [40] introduces a similar operation to us to integrate historical information into the current frame, the design of their model is unable to predict future states. To alleviate these issues, we propose a PoseSync BEV Encoder module based on deformable attention to generate pose-synchronized BEV representations from temporally consecutive image-view input.", + "bbox": [ + 496, + 536, + 892, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Spatial-Temporal Modeling", + "text_level": 1, + "bbox": [ + 500, + 741, + 750, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the BEV prediction field, how to design a temporal model to aggregate spatial-temporal information is a critical problem. Existing modeling methods can be classified into three categories: RNN-based, CNN-based and transformer-based. RNN-based methods [1, 16-18, 43, 56] utilize recurrent models such as LSTM [15], GRU [8] to predict the future latent states. Though the recurrent model is powerful to model temporal relationships, it is time-consuming for constraints in the parallelization of computation. Besides,", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1370", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c7b4199d8890da50565946c771cae33c5f05eb0a2720c25da883f3f9515e474.jpg", + "image_caption": [ + "(A)", + "(B)" + ], + "image_footnote": [], + "bbox": [ + 133, + 102, + 495, + 275 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/999941865914e9fd53b96a15718d940b3b6514b1b6f1122f5345fd09a902cd3f.jpg", + "image_caption": [ + "Figure 3. The PoseSync BEV Encoder (A) takes front-view features and camera poses as input and then maps to BEV space. The core to generate BEV features in a synchronized way is the Pose-Aware Cross Attention. Its cross-view attention mechanism is depicted in (B), where front-view features from different frames of a dynamic vehicle are projected into a uniform BEV space." + ], + "image_footnote": [], + "bbox": [ + 563, + 102, + 890, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "some CNN-based methods [5, 34, 50, 53] concatenate BEV features in the time dimension and take advantage of 3D convolution to extract spatial-temporal features.", + "bbox": [ + 75, + 344, + 468, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the great power of transformer [48] in sequence modeling, it has shown promise in many temporal modeling tasks such as trajectory prediction [11, 35, 55], object tracking [22], video prediction [12, 41, 51], video interpolation [10, 32, 47], etc. For BEV perception, [24, 27] utilize self-attention to model temporal information from multiple frames to boost perception task. [22] leverage self-attention to aggregate spatial information and cross-attention to exploit affinities among sequence frames. To explore the capacity of transformer model in BEV spatial-temporal modeling, we propose a novel Spatial-Temporal Pyramid Transformer (STPT) architecture with future queries for BEV spatial-temporal features extraction and BEV future states prediction.", + "bbox": [ + 75, + 392, + 470, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 619, + 210, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overview Architecture", + "text_level": 1, + "bbox": [ + 76, + 647, + 287, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture of the proposed TBP-Former is illustrated in Fig. 2. It takes the input of multi-view images with the corresponding camera poses at consecutive $T$ timestamps. The final output includes BEV map segmentation for current scene understanding and occupancy flow for motion prediction. The whole TBP-Former can be decoupled into three parts: (i) pose-synchronized BEV encoder, which maps raw image sequences into feature maps in a spatial-temporal-synchronized BEV space; (ii) spatial-temporal pyramid transformer, which achieves comprehensive feature learning at multiple spatial and temporal scales; and (iii) a multi-head decoder, which takes the spatial and temporal features to achieve scene understanding and motion prediction. We will elaborate on each part in the following subsections.", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pose-Synchronized BEV Encoder", + "text_level": 1, + "bbox": [ + 500, + 343, + 792, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given images collected at multiple time stamps and from various camera poses, we aim to generate the corresponding feature maps in a shared and synchronized BEV space. Different from many previous works that synchronize spatial and temporal information in two separate steps, the proposed pose-synchronized BEV encoder leverage both geometric prior and learning ability to achieve one-step synchronization, alleviating distortion effects. Following the previous transformer-based method [24], this encoder adopts a transformer architecture whose core is a novel cross-view attention operation.", + "bbox": [ + 496, + 368, + 892, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Front-view feature map. Let $\\mathcal{X} = \\{\\mathbf{X}_i^{(-t)}\\}_{i=1,t=0}^{N,T}$ be the input multi-frame multi-view images, where $N$ is the number of cameras, $T$ is the number of historical timestamps and $\\mathbf{X}_i^{(-t)} \\in \\mathbb{R}^{H \\times W \\times 3}$ is the RBG image captured by the $i$ th camera at historical time stamp $t$ . Note that each front-view image is associated with a different camera pose. Let $S_i^{(-t)} = \\{(u_i^{(-t)}, v_i^{(-t)})\\}_{1,1}^{H,W}$ be the pixel indices of the $i$ th camera's front-view space, whose image size is $H \\times W$ . We feed each RBG image $\\mathbf{X}_i^{(-t)}$ into a shared backbone network (our implementation uses ResNet-101 [14]) and obtain the corresponding front-view feature map $\\mathbf{F}_i^{(-t)} \\in \\mathbb{R}^{H' \\times W' \\times C}$ with $C$ the channel number, which is also supported on the front-view space $S_i^{(-t)}$ .", + "bbox": [ + 496, + 534, + 892, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BEV queries. Let $S_{\\mathrm{BEV}} = \\{(x,y)\\}_{x = 1,y = 1}^{X,Y}$ be the BEV grid indices, reflecting the $X\\times Y$ BEV grid space based on the vehicle-ego pose at the current timestamp. Note that $S_{\\mathrm{BEV}}$ is the only BEV space we work with in this paper. Let $\\mathbf{Q}\\in \\mathbb{R}^{X\\times Y\\times C}$ be the trainable BEV queries whose element $\\mathbf{Q}_{x,y}\\in \\mathbb{R}^C$ is a $C$ -dimensional query feature at the $(x,y)$ th geo-location in the BEV space $S_{\\mathrm{BEV}}$ .We use $\\mathbf{Q}$ as the input to query from the front-view feature map $\\mathbf{F}_i^{(-t)}$ to produce the corresponding BEV feature map.", + "bbox": [ + 496, + 744, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-view attention. As the key operation in the pose", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1371", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "synchronized BEV encoder, the proposed cross-view attention constructs a feature map in the BEV space $S_{\\mathrm{BEV}}$ by absorbing information from the corresponding pixels in the front-view feature map.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{P}_i^{(-t)}:S_{\\mathrm{BEV}}\\times \\mathcal{Z}\\to S_i^{(-t)}$ be a project operation that maps a BEV index with a specific height index to a pixel index in the $i$ th camera's front view at historical timestamp $t$ ; that is,", + "bbox": [ + 75, + 152, + 468, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(u _ {i} ^ {(- t)}, v _ {i} ^ {(- t)}\\right) = \\mathcal {P} _ {i} ^ {(- t)} \\Big ((x, y, z) \\Big),\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 229, + 388, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z \\in \\mathcal{Z} = \\{1, \\dots, Z\\}$ . The project operation $\\mathcal{P}_i^{(-t)}$ builds the geometric relationship between the BEV and a front view. The implementation of $\\mathcal{P}_i^{(-t)}$ works as", + "bbox": [ + 75, + 270, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} ^ {(- t)} \\cdot \\left[ \\begin{array}{c} u _ {i} ^ {(- t)} \\\\ v _ {i} ^ {(- t)} \\\\ 1 \\end{array} \\right] = \\mathcal {T} _ {\\mathcal {S} _ {\\mathrm {B E V}} \\to \\mathcal {S} _ {i} ^ {(- t)}} \\cdot \\left[ \\begin{array}{c} x \\\\ y \\\\ z \\\\ 1 \\end{array} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 334, + 406, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}_{\\mathcal{S}_{\\mathrm{BEV}}\\to \\mathcal{S}_i^{(-t)}}\\in \\mathbb{R}^{3\\times 4}$ is a transformation matrix that can be calculated by camera's intrinsic/ extrinsic parameters and vehicle-ego pose.", + "bbox": [ + 75, + 414, + 468, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the project operation, the cross-view attention can trace visual features in the front view through a BEV index. Let $\\mathbf{B}_i^{(-t)}\\in \\mathbb{R}^{X\\times Y\\times C}$ be the BEV feature map associated with the RBG image $\\mathbf{X}_i^{(-t)}$ . The $(x,y)$ th element of the BEV feature map is obtained as", + "bbox": [ + 75, + 464, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {B} _ {i} ^ {(- t)}\\right) _ {x, y} = \\sum_ {z} f _ {\\mathrm {D A}} \\left(\\mathbf {Q} _ {x, y}, \\mathcal {P} _ {i} ^ {(- t)} (x, y, z), \\mathbf {F} _ {i} ^ {(- t)}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 559, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $f_{\\mathrm{DA}}(\\cdot)$ represents the deformable attention operation [58]. It allows BEV query $Q_{x,y}$ only to interact with the front-view feature $\\mathbf{F}_i^{(-t)}$ within its regions of interest, which is sampled around the reference point calculated by $\\mathcal{P}_i^{(-t)}$ . Since one BEV index might lead to multiple pixel indices in the front-view image because of various height possibilities in the 3D space. We thus sum over all possible heights in (1). To further aggregate BEV feature maps across all the $N$ camera views, we simply take the average; that is, the BEV feature map at historical timestamp $t$ is $\\mathbf{B}^{(-t)} = \\frac{1}{N}\\sum_{i}\\mathbf{B}_{i}^{(-t)}$ . Note that all the front-view features across time and from multiple cameras are synchronized into the same BEV space in one step (1), leading to less information distortion or deficiency issues.", + "bbox": [ + 75, + 604, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can successively apply the cross-view attention followed by feed forward networks and normalization layers for multiple times. Finally, we order BEV feature maps at multiple timestamps and obtain a temporal BEV feature map $\\mathcal{B} = [\\mathbf{B}^{(0)},\\mathbf{B}^{(-1)},\\dots,\\mathbf{B}^{(-T)}]\\in \\mathbb{R}^{(T + 1)\\times X\\times Y\\times C}$ .", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/98c466355ad7037ecd88acf0f26e1b8e44eebe2b1f9dde4fd776cd2202957f94.jpg", + "image_caption": [ + "Figure 4. The network architecture of Spatial-Temporal Pyramid Transformer (STPT). Each encoder layer consists of an optional convolutional block for downsampling and Swin Transformer Blocks, while each decoder layer contains Swin Transformer Blocks and a deconvolutional block for upsampling. In the decoding process, we pre-define a set of future queries to represent future BEV states and query the features from encoders." + ], + "image_footnote": [], + "bbox": [ + 526, + 87, + 867, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spatial-Temporal Pyramid Transformer", + "text_level": 1, + "bbox": [ + 500, + 446, + 841, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further propose a novel spatial-temporal pyramid transformer (STPT) to learn spatial-temporal features more comprehensively and produce the future BEV states. The detailed structure of STPT is depicted in Fig. 4.", + "bbox": [ + 498, + 470, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal BEV pyramid feature learning. We encode the input temporal BEV feature map $\\mathcal{B}$ with four hierarchical layers. Each encoder layer is composed of an optional convolution layer with stride 2 to downsample the features and a swim transformer encoder, which is a stack of Swin Transformer blocks [28]. In our implementation, the window size of Swin-T blocks is set as (4,4). We can then obtain multi-scale spatial-temporal features $\\mathcal{B}_s\\in \\mathbb{R}^{(T + 1)\\times \\frac{X}{2^s}\\times \\frac{Y}{2^s}\\times C}$ , $s = 0,1,2,3$ .", + "bbox": [ + 498, + 534, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Future BEV queries. Future BEV queries are defined to represent future BEV states and query the generated multiscale spatial-temporal features. There is a set of learnable future queries $\\{\\mathbf{Q}^{(t)}\\}$ , $t = 0, \\dots, T'$ , where $\\mathbf{Q}^{(t)}$ has the same spatial dimension $\\frac{X}{8} \\times \\frac{Y}{8}$ as $\\mathcal{B}_3$ . Separate learning embeddings are employed for future queries to differentiate the predicted BEV states over time. Additionally, a map feature generator is applied to generate high-dimensional features from $\\mathbf{B}^{(T)}$ in order to extract information about the scene's geometry. To be specific, the same structure and parameters of the hdmap decoder head (excluding the last linear layer) are reused. The resulting map feature is added to all future queries to provide spatial information priors.", + "bbox": [ + 498, + 671, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Future BEV state prediction. The decoding process contains corresponding four hierarchical layers as the en", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1372", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/feb83b50d7aecfe810fcef3983927ca84336e2bd9a0a61808899555b459d3a1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRGB ResolutionFuture semantic seg.Future instance seg.FPS
IoU (Short)IoU (Long)VPQ (Short)VPQ (Long)
FIERY [16]224×48059.436.750.229.91.56
StretchBEV [1]224×48055.537.146.029.01.56
ST-P3 [17]224×480-38.9-32.11.43
BEVerse [56]256×70460.338.752.233.31.96
TBP-Former224×48064.741.956.736.92.44
", + "bbox": [ + 145, + 88, + 823, + 212 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2eec6cae78fea7371b0bacc3ac6dd1b2af0284cace4232bb0b83b62f43708a67.jpg", + "table_caption": [ + "Table 1. Prediction results on nuScenes [2] validation set. Intersection-over-Union (IoU) is used for future semantic segmentation and Video Panoptic Quality (VPQ) for future instance segmentation. Results are reported under two settings: short $(30m\\times 30m)$ range and long $(100m\\times 100m)$ range. Frame Per Second (FPS) means the inverse of inference time. All methods are tested under the same settings on a single NVIDIA A100. Our TBP-Former achieves SOTA performance and is still more computationally efficient than other methods." + ], + "table_footnote": [], + "table_body": "
MethodTemp.Veh. IoUPed. IoU
VED [31]23.311.9
VPN [36]28.210.3
PON [42]27.913.9
LSS [38]34.615.0
CVT [57]36.0-
Image2Map [44]40.2-
BEVFormer [24]44.4-
IVMP [49]✓36.817.4
FIERY [16]✓38.217.2
ST-P3 [17]✓40.114.5
TBP-Former static44.817.2
TBP-Former✓46.218.6
", + "bbox": [ + 91, + 292, + 455, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Perception results on nuScenes [2] validation set. Results of vehicles and pedestrians are compared by segmentation IoU. Temp. indicates whether temporal information is involved.", + "bbox": [ + 75, + 518, + 468, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "coding process. Unlike the encoding process, $\\{\\mathbf{Q}^{(t)}\\}$ is used as the query input of Swin-T block and performs cross attention with the encoded features $\\mathbf{E}_3$ . After the first decoding layer, the output of each layer is used as the query input of the next layer. Similar to encoding layers, the deconvolution layer is optionally applied to upsample the decoded features. The simplified process can be written as", + "bbox": [ + 75, + 571, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {s} = \\left\\{ \\begin{array}{l l} \\operatorname {S w i n T} (\\mathcal {B} _ {3}, \\{\\mathbf {Q} ^ {(t)} \\}), & s = 3 \\\\ \\operatorname {S w i n T} (\\mathcal {B} _ {s}, \\operatorname {D e C o n v} (\\mathbf {D} _ {s + 1})), & s = 0, 1, 2 \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 686, + 442, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_s\\in \\mathbb{R}^{(T' + 1)\\times \\frac{X}{2^s}\\times \\frac{Y}{2^s}\\times C}, s = 0,1,2,3$ are decoded features. The future temporal BEV feature map at the 0th scale is a temporal sequence of final predicted BEV states; that is, $\\mathcal{D}_0 = [\\mathbf{B}_*^{(0)},\\mathbf{B}_*^{(1)},\\dots,\\mathbf{B}_*^{(T')}]$ , where $\\mathbf{B}_*^{(t)}\\in \\mathbb{R}^{X\\times Y\\times C}$ is the BEV state at future time stamp $t$ .", + "bbox": [ + 76, + 739, + 467, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Multi-head decoder", + "text_level": 1, + "bbox": [ + 76, + 830, + 264, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The future temporal BEV feature map is fed into the multi-task decoder heads to generate various outputs for dynamic scene understanding, see Fig. 5. We follow the output", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "setting in [16] that predicts BEV semantic segmentation, instance center, instance offset, and future flow for joint perception and prediction. Meanwhile, we set up an additional HD map decoder head to predict basic traffic scene elements including drivable areas and lanes. The map decoder head can not only provide scene information for subsequent planning and control modules but also give guidance to the prediction process, see sec. 3.3.", + "bbox": [ + 496, + 296, + 890, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 422, + 633, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and settings", + "text_level": 1, + "bbox": [ + 500, + 446, + 689, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use nuScenes [2] datasets to evaluate our approach. NuScenes contains 1000 scenes, each of which has 20 seconds annotated at $2\\mathrm{Hz}$ . In nuScenes, the images are captured by 6 cameras with a small overlap in the field of view, which guarantees the cameras cover the full $360^{\\circ}$ field of view. For model input, raw camera images with the size of $900 \\times 1600$ are resized and cropped to a resolution of $224 \\times 480$ . We follow the training and evaluating settings used in previous methods [1,16,17,56] for fair comparisons, which use 1.0 second past states and current state to predict 2.0 seconds of the future states. It corresponds to predicting 4 future frames based on 3 observed frames. The size of the generated BEV grid map is $200 \\times 200$ . Each grid has a range of $0.5m \\times 0.5m$ , which means the perception and prediction range is $100m \\times 100m$ .", + "bbox": [ + 496, + 470, + 890, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training, we use AdamW [30] with a weight decay 0.01 to optimize the models. The learning rate is initialized as $10^{-4}$ and decays with a cosine annealing scheduler [29]. All models are trained on 4 NVIDIA A100 GPUs for 10 epochs.", + "bbox": [ + 496, + 699, + 890, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 500, + 786, + 596, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous works [1, 16, 17, 56], we mainly use two metrics for evaluation. The first is Intersection over Union (IoU), which measures the quality of segmentation at each frame. The second is Video Panoptic Quality (VPQ), which is used to measure the consistency of the detected instances over time and the accuracy of the segmentation.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1373", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/92b53a1848df79cb682b19a90bad019779d318759c80ae325879ec7913eb8018.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Exp.Warp.Sync.SLQSPEFuture semantic seg.Future instance seg.
Short (IoU)Long (IoU)Short (VPQ)Long (VPQ)
1✓58.738.450.631.8
2✓✓60.838.652.433.4
3✓✓✓62.040.753.234.3
4✓63.040.854.134.3
5✓✓63.841.155.835.8
6✓✓✓64.741.956.736.9
", + "bbox": [ + 155, + 88, + 816, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/086a6438eed1fa23a41a8454d9dfc1d037ca89c79985f33fabb0cb946824031f.jpg", + "table_caption": [ + "Table 3. Ablation of our proposed architecture. Ablation results for our PoseSync BEV Encoder (Sync.), the learnable future queries, and the spatial embedding are presented. Exp. 1-3 use the traditional warping methods to align temporal BEV features. Separate learnable queries (SLQ) represent using separate learnable future queries instead of utilizing the same query with temporal positional encoding. Spatial positional embedding (SPE) represents using spatial scene representations in future prediction queries." + ], + "table_footnote": [], + "table_body": "
Temporal modelIoUVPQVRQVSQ
MotionNet†[53]35.430.643.171.1
FIERY†[16]38.332.145.470.7
BEVerse†[56]40.234.048.070.9
TBP-Former41.936.951.572.6
", + "bbox": [ + 102, + 301, + 444, + 391 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dfd61a9af9512a84aac3f9e2849a25e9a8fd31db48a14e2ffdb69cc2413e2a4f.jpg", + "table_caption": [ + "Table 4. Ablation for the prediction model. †: We use MotionNet, FIERY and BEVerse to replace our prediction model for comparison, and the BEV encoder and task heads are the same. Besides IoU and VPQ, we also use Video Recognition Quality (VRQ) and Video Segmentation quality (VSQ) for evaluation." + ], + "table_footnote": [], + "table_body": "
AugmentationPerceptionPrediction
CamBEVVeh.Ped.IoUVPQ
45.017.740.534.4
✓44.818.540.935.3
✓45.318.641.435.6
✓✓46.218.641.936.9
", + "bbox": [ + 116, + 481, + 431, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation for data augmentation strategies. Perception of Vehicles and Pedestrians with different data augmentation strategies are evaluated on segmentation IoU. Prediction results are evaluated on segmentation IoU and Video Panoptic Quality.", + "bbox": [ + 75, + 595, + 468, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The formula is shown below:", + "bbox": [ + 76, + 662, + 272, + 676 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V P Q} = \\sum_ {t = 0} ^ {H} \\frac {\\sum_ {\\left(p _ {t} , q _ {t}\\right) \\in T P _ {t}} \\operatorname {I o U} \\left(p _ {t} , q _ {t}\\right)}{\\left| T P _ {t} \\right| + \\frac {1}{2} \\left| F P _ {t} \\right| + \\frac {1}{2} \\left| F N _ {t} \\right|}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 685, + 408, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $H$ is the sequence length, $TP_{t}$ represents the set of true positives, $FP_{t}$ represents the set of false positives and $FN_{t}$ represents the set of false negatives at timestamp $t$ .", + "bbox": [ + 75, + 724, + 468, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. PnP results", + "text_level": 1, + "bbox": [ + 76, + 775, + 202, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perception and Prediction. Table 1 compares TBP-Former with other methods of perception and prediction task based on multi-view cameras. We see that i) we achieve state-of-the-art performance and exceed previous methods by a large margin. ii) Even though BEVerse has larger RGB resolutions, TBP-Former still surpasses their performance on IoU by $7.3\\% / 8.3\\%$ for short/long settings, respectively.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TBP-Former also improves the VPQ by $12.1\\% / 10.8\\%$ . iii) Apart from the performance improvement, TBF-Former also has a larger FPS compared to other methods. Its inference speed is $25\\%$ faster than BEVerse's.", + "bbox": [ + 496, + 304, + 890, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 shows the visualization results of our proposed method. We see that i) almost all the objects are detected correctly except for those occluded ones. ii) TBP-Former is capable of capturing the motion information in past frames and precisely predicting the vehicles' trajectories by occupancy and flow. Compared with FIERY [16], TBP-Former is closer to the ground truth. iii) TBP-Former does a better job than FIERY when predicting vehicles' turning.", + "bbox": [ + 496, + 364, + 892, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perception Only. Table 2 compares the results of plenty of state-of-the-art methods on perception (segmentation) task. We see that our static model, which does not contain temporal information, can achieve 44.8 and 17.2 IoU of vehicles and pedestrians. With the input of temporal sequences, the performance improves further since auxiliary information is provided for better perception. The state-of-the-art results prove the effectiveness of the novel design of our Pose-synchronized BEV encoder.", + "bbox": [ + 496, + 484, + 893, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation", + "text_level": 1, + "bbox": [ + 500, + 625, + 604, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of PoseSync View Projection. The Exp. 1&4, 2&5, 3&6 in Table 3 compare the proposed PoseSync View Projection and the existing feature warping methods. We see that the proposed method always achieves better performance when other settings remain the same. The reasons are that: i) PoseSync View Projection based on Deformable Attention can guarantee the precise correspondence between BEV grids and image features. ii) Our projection method can alleviate distortion and our-of-range issues when synchronizing sequential BEV features.", + "bbox": [ + 496, + 643, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the designed future queries. In Exp. 1&3 in Table 3, we utilize the identical query with temporal positional encoding for future queries. Exp. 2&4 in Table 3 demonstrate that using separate learnable embedding for future queries can achieve better performance. Exp. 3&6 in Table 3 validate the efficacy of the proposed spatial priors for future queries. The generated high", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1374", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05ace495861ad71b1b2312e77f5dacfff1efdf7fd70b78a3a7c7c32bc1eb1dca.jpg", + "image_caption": [ + "Figure 5. Demonstration of our results compared with FIERY and Ground Truth. Different vehicles are assigned with different colors in order to make a distinction. The darker parts represent the perception of the current frame, and the lighter parts represent the prediction of the vehicles in future frames. The visualization is based on the predicted occupancy and flow." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 880, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dimensional map features provide the prediction model with useful geographic information. The additional spatial information can aid the prediction and lead to better scene forecasting.", + "bbox": [ + 75, + 565, + 470, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of STPT. Table 4 compares STPT with popular CNN-based [53] and RNN-based [16, 56] methods. We keep all the settings the same except for temporal modeling. To be specific, the size of input images $(224 \\times 480)$ , image backbones and BEV feature extractor are the same. And then we plug their temporal models into our architecture. We see that i) STPT model performs better in all four metrics, including semantic segmentation IoU and three instance segmentation metrics from the video prediction area. ii) Our reproduced temporal models achieve higher performance than the original implements. This further validates the effectiveness and power of our BEV feature extractor.", + "bbox": [ + 75, + 627, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Augmentation. We perform both image-view and BEV augmentations. The image-view augmentations include random scaling, rotation and flip of the input images. The BEV augmentations include similar operations on both BEV representations and corresponding ground truth labels. Table 5 compares the results of different data augmentation", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "strategies. We see that i) both augmentation methods improve the performance when used separately. ii) The combination of two methods works better than any single approach. Introducing data augmentation strategies is beneficial to the model's robustness and generalization ability.", + "bbox": [ + 496, + 565, + 890, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 619, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a novel TBP-Former for vision-centric joint perception and prediction. We design a pose-synchronized BEV encoder module using a cross-view attention mechanism to solve the distortion issues in previous works. Furthermore, we propose a powerful spatial-temporal pyramid transformer for BEV feature extraction and BEV state prediction. Experiments show that i) TBP-Former improves the prediction performance over state-of-the-art methods significantly; and ii) both PoseSync BEV Encoder and STPT contribute to better performances.", + "bbox": [ + 496, + 665, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This research is partially supported by National Natural Science Foundation of China under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005.", + "bbox": [ + 496, + 816, + 893, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1375", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022. 2, 3, 6", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6", + "[3] Sergio Casas, Cole Gulino, Renjie Liao, and Raquel Urtasun. Spagnn: Spatially-aware graph neural networks for relational behavior forecasting from sensor data. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9491-9497. IEEE, 2020. 2", + "[4] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 2, 3", + "[5] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 3, 4", + "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 3", + "[7] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 3", + "[8] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597–1600. IEEE, 2017. 3", + "[9] Artem Filatov, Andrey Rykov, and Viacheslav Murashkin. Any motion detector: Learning class-agnostic scene dynamics from a sequence of lidar point clouds. In 2020 IEEE international conference on robotics and automation (ICRA), pages 9498–9504. IEEE, 2020. 3", + "[10] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17441-17451, 2022. 4", + "[11] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2021. 4", + "[12] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 4" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. A simple baseline for bev perception without lidar. arXiv e-prints, pages arXiv-2206, 2022. 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[15] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3", + "[16] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 2, 3, 6, 7, 8", + "[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2, 3, 6", + "[18] Yue Hu, Siheng Chen, Ya Zhang, and Xiao Gu. Collaborative motion predication via neural motion message passing. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3", + "[19] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), November 2022. 3", + "[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3", + "[21] Kuan-Hui Lee, Matthew Kliemann, Adrien Gaidon, Jie Li, Chao Fang, Sudeep Pillai, and Wolfram Burgard. Pillarflow: End-to-end birds-eye-view flow estimation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2007-2013. IEEE, 2020. 3", + "[22] Peixuan Li and Jieyu Jin. Time3d: End-to-end joint monocular 3d object detection and tracking for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3885-3894, 2022. 4", + "[23] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework, 2021. 3", + "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2, 3, 4, 6", + "[25] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1, 2, 3", + "[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1376", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3", + "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 4", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 5", + "[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[31] Chenyang Lu, Marinus Jacobus Gerardus van de Molengraft, and Gijs Dubbelman. Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters, 4(2):445-452, 2019. 6", + "[32] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 4", + "[33] Chenxu Luo, Xiaodong Yang, and Alan Yuille. Self-supervised pillar motion learning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3183-3192, 2021. 3", + "[34] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2, 4", + "[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified multi-task model for behavior prediction and planning. arXiv e-prints, pages arXiv-2106, 2021. 4", + "[36] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 3, 6", + "[37] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 3", + "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 2, 3, 6", + "[39] John Phillips, Julieta Martinez, Ioan Andrei Bársan, Sergio Casas, Abbas Sadat, and Raquel Urtasun. Deep multi-task learning for joint localization, perception, and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4679-4689, 2021. 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Zequn Qin, Jingyu Chen, Chao Chen, Xiaozhi Chen, and Xi Li. Uniformer: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. arXiv preprint arXiv:2207.08536, 2022. 3", + "[41] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent video transformer. arXiv preprint arXiv:2006.10704, 2020. 4", + "[42] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 3, 6", + "[43] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In European Conference on Computer Vision, pages 414-430. Springer, 2020. 1, 3", + "[44] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In 2022 International Conference on Robotics and Automation (ICRA), pages 9200-9206. IEEE, 2022. 6", + "[45] Marcel Schreiber, Vasileios Belagiannis, Claudius Gläser, and Klaus Dietmayer. Dynamic occupancy grid mapping with recurrent neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6717-6724. IEEE, 2021. 3", + "[46] Meet Shah, Zhiling Huang, Ankit Laddha, Matthew Langford, Blake Barber, Sidney Zhang, Carlos Vallespi-Gonzalez, and Raquel Urtasun. Liranet: End-to-end trajectory prediction using spatio-temporal radar fusion. arXiv preprint arXiv:2010.00731, 2020. 1", + "[47] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 4", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[49] Hengli Wang, Peide Cai, Yuxiang Sun, Lujia Wang, and Ming Liu. Learning interpretable end-to-end vision-based motion planning for autonomous driving with optical flow distillation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13731-13737. IEEE, 2021. 2, 6", + "[50] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17093–17102, 2022. 4", + "[51] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 4", + "[52] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying unknown instances for au" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1377", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tonomous driving. In Conference on Robot Learning, pages 384-393. PMLR, 2020. 2", + "[53] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 2, 3, 4, 7, 8", + "[54] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3", + "[55] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4", + "[56] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2, 3, 6, 7, 8", + "[57] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 3, 6", + "[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + ], + "bbox": [ + 78, + 92, + 467, + 527 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1378", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_model.json b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99ad4cbe088120c3b0985e5970eaf5e89c4b9a00 --- /dev/null +++ b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.845, + 0.177 + ], + "angle": 0, + "content": "TBP-Former: Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.203, + 0.797, + 0.223 + ], + "angle": 0, + "content": "Shaoheng Fang\\(^{1*}\\) Zi Wang\\(^{1*}\\) Yiqi Zhong\\(^{2}\\) Junhao Ge\\(^{1}\\) Siheng Chen\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.223, + 0.776, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Cooperative Medianet Innovation Center, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.24, + 0.879, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, University of Southern California \\(^{3}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.258, + 0.795, + 0.275 + ], + "angle": 0, + "content": "1{shfang，w4ngz1，cancaries，sihengc}@sjtu.edu.cn 2{yiqizhon}@usc.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.66 + ], + "angle": 0, + "content": "Vision-centric joint perception and prediction (PnP) has become an emerging trend in autonomous driving research. It predicts the future states of the traffic participants in the surrounding environment from raw RGB images. However, it is still a critical challenge to synchronize features obtained at multiple camera views and timestamps due to inevitable geometric distortions and further exploit those spatial-temporal features. To address this issue, we propose a temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP, which includes two novel designs. First, a pose-synchronized BEV encoder is proposed to map raw image inputs with any camera pose at any time to a shared and synchronized BEV space for better spatial-temporal synchronization. Second, a spatial-temporal pyramid transformer is introduced to comprehensively extract multi-scale BEV features and predict future BEV states with the support of spatial priors. Extensive experiments on nuScenes dataset show that our proposed framework overall outperforms all state-of-the-art vision-based prediction methods. Code is available at: https://github.com/MediaBrain-SJTU/TBP-Former" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.7, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.862 + ], + "angle": 0, + "content": "As one of the most fascinating engineering projects, autonomous driving has been an aspiration for many researchers and engineers for decades. Although significant progress has been made, it is still an open question in designing a practical solution to achieve the goal of full self-driving. A traditional and common solution consists of a sequential stack of perception, prediction, planning, and control. Despite the idea of divide-and-conquer having achieved tremendous success in developing software sys" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.31, + 0.688, + 0.321 + ], + "angle": 0, + "content": "Previous works" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.323, + 0.777, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.31, + 0.873, + 0.321 + ], + "angle": 0, + "content": "TBP-Former" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.323, + 0.887, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.45, + 0.804, + 0.462 + ], + "angle": 0, + "content": "(a) Temporal BEV feature alignment" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.463, + 0.777, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.463, + 0.889, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.593, + 0.846, + 0.606 + ], + "angle": 0, + "content": "(b) Spatial-temporal feature modeling for prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.615, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Figure 1. Two major challenges in vision-based perception and prediction are (a) how to avoid distortion and deficiency when aggregating features across time and camera views; and (b) how to achieve spatial-temporal feature learning for prediction. Our Pose-Synchronized BEV Encoder can precisely map the visual features into synchronized BEV space, and Spatial-Temporal Pyramid Transformer extracts feature at multiple scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.894, + 0.852 + ], + "angle": 0, + "content": "tems, a long stack could cause cascading failures in an autonomous system. Recently, there is a trend to combine multiple parts in an autonomous system to be a joint module, cutting down the stack. For example, [25, 46] consider joint perception and prediction and [5, 43] explore joint prediction and planning. This work focuses on joint perception and prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The task of joint perception and prediction (PnP) aims to predict the current and future states of the surrounding environment with the input of multi-frame raw sensor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.352, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.352, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1368" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "data. The output current and future states would directly serve as the input for motion planning. Recently, many PnP methods are proposed based on diverse sensor input choices. For example, [4, 25, 34] take multi-frame LiDAR point clouds as input and achieve encouraging 3D detection and trajectory prediction performances simultaneously. Recently, the rapid development of vision-centric methods offers a new possibility to provide a cheaper and easy-to-deploy solution for PnP. For instance, [1, 16, 17] only uses RGB images collected by multiple cameras to build PnP systems. Meanwhile, without precise 3D measurements, vision-centric PnP is more technically challenging. Therefore, this work aims to advance this direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.502 + ], + "angle": 0, + "content": "The core of vision-centric PnP is to learn appropriate spatial-temporal feature representations from temporal image sequences. It is a crux and difficult from three aspects. First, since the input and the output of vision-centric PnP are supported in camera front-view (FV) and bird's-eye-view (BEV) respectively, one has to deal with distortion issues during geometric transformation between two views. Second, when the vehicle is moving, the view of the image input is time-varying and it is thus nontrivial to precisely map visual features across time into a shared and synchronized space. Third, since information in temporal image sequences is sufficiently rich for humans to accurately perceive the environment, we need a powerful learning model to comprehensively exploit spatial-temporal features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.763 + ], + "angle": 0, + "content": "To tackle these issues, previous works on vision-centric PnP consider diverse strategies. For example, [16, 56] follows the method in [38] to map FV features to BEV features, then synchronizes BEV features across time via rigid transformation, and finally uses a recurrent network to exploit spatial-temporal features. However, due to the image discretization nature and depth estimation uncertainty, simply relying on rigid geometric transformations would cause inevitable distortion; see Fig. 1. Some other work [49] transforms the pseudo feature point cloud to current ego coordinates and then pools the pseudo-lidar to BEV features; however, this approach encounters deficiency due to the limited sensing range in perception. Meanwhile, many works [16, 17, 56] simply employ recurrent neural networks to learn the temporal features from multiple BEV representations, which is hard to comprehensively extract spatial-temporal features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.904 + ], + "angle": 0, + "content": "To promote more reliable and comprehensive feature learning across views and time, we propose the temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP. The proposed TBP-Former includes two key innovations: i) pose-synchronized BEV encoder, which leverages a pose-aware cross-attention mechanism to directly map a raw image input with any camera pose at any time to the corresponding feature map in a shared and synchronized BEV space; and ii) spatial-temporal pyra" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "mid transformer, which leverages a pyramid architecture with Swin-transformer [28] blocks to learn comprehensive spatial-temporal features from sequential BEV maps at multiple scales and predict future BEV states with a set of future queries equipped with spatial priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.363 + ], + "angle": 0, + "content": "Compared to previous works, the proposed TBP-Former brings benefits from two aspects. First, previous works [16, 17, 24, 56] consider FV-to-BEV transformation and temporal synchronization as two separate steps, each of which could bring distortion due to discrete depth estimation and rigid transformation; while we merge them into one step and leverage both geometric transformation and attention-based learning ability to achieve spatial-temporal synchronization. Second, previous works [16, 53] use RNNs or 3D convolutions to learn spatial-temporal features; while we leverage a powerful pyramid transformer architecture to comprehensively capture spatial-temporal features, which makes prediction more effective." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.365, + 0.882, + 0.379 + ], + "angle": 0, + "content": "To summarize, the main contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.39, + 0.892, + 0.466 + ], + "angle": 0, + "content": "- To tackle the distortion issues in mapping temporal image sequences to a synchronized BEV space, we propose a pose-synchronized BEV encoder (PoseSync BEV Encoder) based on cross-view attention mechanism to extract quality temporal BEV features." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.476, + 0.892, + 0.552 + ], + "angle": 0, + "content": "- We propose a novel Spatial-Temporal Pyramid Transformer (STPT) to extract multi-scale spatial-temporal features from sequential BEV maps and predict future BEV states according to well-elaborated future queries integrated with spatial priors." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.563, + 0.893, + 0.713 + ], + "angle": 0, + "content": "- Overall, we propose TBP-Former, a vision-based joint perception and prediction framework for autonomous driving. TBP-Former achieves state-of-the-art performance on nuScenes [2] dataset for the vision-based prediction task. Extensive experiments show that both PoseSync BEV Encoder and STPT contribute greatly to the performance. Due to the decoupling property of the framework, both proposed modules can be easily utilized as alternative modules in any vision-based BEV prediction framework." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.39, + 0.893, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.732, + 0.642, + 0.747 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.78, + 0.773 + ], + "angle": 0, + "content": "2.1. Joint Perception and Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.903 + ], + "angle": 0, + "content": "As the two core system modules of autonomous driving, how to conduct perception and prediction tasks jointly has received a lot of attention. Traditional approaches [3,4,25,34,39] formulate this joint task as a trajectory prediction problem that relies on the perception outputs of 3D object detection and tracking. The dependency on intermediate results tends to accumulate errors and lacks the capacity to perceive unknown objects [52,53]. Subsequently," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1369" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.096, + 0.882, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.896, + 0.419 + ], + "angle": 0, + "content": "Figure 2. An overview of TBP-Former architecture. Taking consecutive surrounding camera images as inputs, TBP-Former first generates image-space features and uses the PoseSync BEV Encoder to map front-view features to BEV features in a shared and synchronized BEV space. Then the BEV features from multiple frames are processed by the Spatial-Temporal Pyramid Transformer to extract BEV spatial-temporal features and predict future BEV states in order. In this process, high-level scene representations are generated from the last frame BEV feature as spatial priors to guide the prediction. Finally, the well-predicted future states are sent to decoder heads for joint perception and prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.52 + ], + "angle": 0, + "content": "instance-free methods [9,21,33,43,45,53] that predict dense future semantic occupancy and flow have become a growing trend to simplify the understanding of dynamic scenes. Also, several recent works [16, 17, 56] explore joint perception and prediction in the form of dense occupancy and flow using only surrounding camera input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.471, + 0.673 + ], + "angle": 0, + "content": "In many previous works [4, 25, 39, 43], raster HD (high-definition) maps play an important role as input of the frameworks. HD maps can provide strong priors to guide the predicted results to follow the traffic lanes. However, in practice, HD maps are laborious and costly to produce and require frequent maintenance. Instead of using off-the-peg HD maps, we follow the philosophy of [5, 7, 23] in predicting online HD maps but propose to learn high-level scene geometry representations from real-time sensor inputs and take these representations as priors for the prediction task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.279, + 0.698 + ], + "angle": 0, + "content": "2.2. BEV Representations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "BEV representations provide a unified and physical-interpretable way to represent the rich information of road, moving objects and occlusion in a traffic scene, which can be easily utilized for downstream tasks such as motion prediction, planning and control, etc. For camera-based methods, how to solve the problem of projecting features from perspective view to BEV is a major challenge. Some learnable methods use MLP [23, 36, 42] or transformer network [37, 57] to implicitly reason the relationship between two different views. LSS [38] proposes the approach of predicting depth distribution per pixel on 2D features, then 'lifting' the 2D features according to the corresponding depth distribution to BEV space. Numerous works, aim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.534 + ], + "angle": 0, + "content": "ing at tasks of BEV perception [19, 20, 54], motion prediction [1, 16, 17, 56], lidar-camera fusion [26], etc., follow this form to generate BEV representations. Also, some methods [6, 13, 24] explicitly establish the correspondence from BEV location to image-view pixel using homography between image and BEV plane and achieve attractive performance in diverse tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.734 + ], + "angle": 0, + "content": "However, when dealing with temporal information, most methods [1, 16, 17, 24, 56] warp history BEV representations according to the variation of ego poses. Due to the pre-defined fixed range and size of the BEV grid, rotation and translation operations may cause distortion and out-of-range problems when aligning history BEV maps to current ego coordinates. Though [40] introduces a similar operation to us to integrate historical information into the current frame, the design of their model is unable to predict future states. To alleviate these issues, we propose a PoseSync BEV Encoder module based on deformable attention to generate pose-synchronized BEV representations from temporally consecutive image-view input." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.75, + 0.759 + ], + "angle": 0, + "content": "2.3. Spatial-Temporal Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In the BEV prediction field, how to design a temporal model to aggregate spatial-temporal information is a critical problem. Existing modeling methods can be classified into three categories: RNN-based, CNN-based and transformer-based. RNN-based methods [1, 16-18, 43, 56] utilize recurrent models such as LSTM [15], GRU [8] to predict the future latent states. Though the recurrent model is powerful to model temporal relationships, it is time-consuming for constraints in the parallelization of computation. Besides," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1370" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.101, + 0.106, + 0.113 + ], + "angle": 0, + "content": "(A)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.103, + 0.496, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.101, + 0.541, + 0.114 + ], + "angle": 0, + "content": "(B)" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.103, + 0.891, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Figure 3. The PoseSync BEV Encoder (A) takes front-view features and camera poses as input and then maps to BEV space. The core to generate BEV features in a synchronized way is the Pose-Aware Cross Attention. Its cross-view attention mechanism is depicted in (B), where front-view features from different frames of a dynamic vehicle are projected into a uniform BEV space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.391 + ], + "angle": 0, + "content": "some CNN-based methods [5, 34, 50, 53] concatenate BEV features in the time dimension and take advantage of 3D convolution to extract spatial-temporal features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.471, + 0.606 + ], + "angle": 0, + "content": "Due to the great power of transformer [48] in sequence modeling, it has shown promise in many temporal modeling tasks such as trajectory prediction [11, 35, 55], object tracking [22], video prediction [12, 41, 51], video interpolation [10, 32, 47], etc. For BEV perception, [24, 27] utilize self-attention to model temporal information from multiple frames to boost perception task. [22] leverage self-attention to aggregate spatial information and cross-attention to exploit affinities among sequence frames. To explore the capacity of transformer model in BEV spatial-temporal modeling, we propose a novel Spatial-Temporal Pyramid Transformer (STPT) architecture with future queries for BEV spatial-temporal features extraction and BEV future states prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.212, + 0.638 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.289, + 0.663 + ], + "angle": 0, + "content": "3.1. Overview Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The overall architecture of the proposed TBP-Former is illustrated in Fig. 2. It takes the input of multi-view images with the corresponding camera poses at consecutive \\(T\\) timestamps. The final output includes BEV map segmentation for current scene understanding and occupancy flow for motion prediction. The whole TBP-Former can be decoupled into three parts: (i) pose-synchronized BEV encoder, which maps raw image sequences into feature maps in a spatial-temporal-synchronized BEV space; (ii) spatial-temporal pyramid transformer, which achieves comprehensive feature learning at multiple spatial and temporal scales; and (iii) a multi-head decoder, which takes the spatial and temporal features to achieve scene understanding and motion prediction. We will elaborate on each part in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.344, + 0.794, + 0.361 + ], + "angle": 0, + "content": "3.2. Pose-Synchronized BEV Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Given images collected at multiple time stamps and from various camera poses, we aim to generate the corresponding feature maps in a shared and synchronized BEV space. Different from many previous works that synchronize spatial and temporal information in two separate steps, the proposed pose-synchronized BEV encoder leverage both geometric prior and learning ability to achieve one-step synchronization, alleviating distortion effects. Following the previous transformer-based method [24], this encoder adopts a transformer architecture whose core is a novel cross-view attention operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Front-view feature map. Let \\(\\mathcal{X} = \\{\\mathbf{X}_i^{(-t)}\\}_{i=1,t=0}^{N,T}\\) be the input multi-frame multi-view images, where \\(N\\) is the number of cameras, \\(T\\) is the number of historical timestamps and \\(\\mathbf{X}_i^{(-t)} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is the RBG image captured by the \\(i\\)th camera at historical time stamp \\(t\\). Note that each front-view image is associated with a different camera pose. Let \\(S_i^{(-t)} = \\{(u_i^{(-t)}, v_i^{(-t)})\\}_{1,1}^{H,W}\\) be the pixel indices of the \\(i\\)th camera's front-view space, whose image size is \\(H \\times W\\). We feed each RBG image \\(\\mathbf{X}_i^{(-t)}\\) into a shared backbone network (our implementation uses ResNet-101 [14]) and obtain the corresponding front-view feature map \\(\\mathbf{F}_i^{(-t)} \\in \\mathbb{R}^{H' \\times W' \\times C}\\) with \\(C\\) the channel number, which is also supported on the front-view space \\(S_i^{(-t)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.892, + 0.885 + ], + "angle": 0, + "content": "BEV queries. Let \\(S_{\\mathrm{BEV}} = \\{(x,y)\\}_{x = 1,y = 1}^{X,Y}\\) be the BEV grid indices, reflecting the \\(X\\times Y\\) BEV grid space based on the vehicle-ego pose at the current timestamp. Note that \\(S_{\\mathrm{BEV}}\\) is the only BEV space we work with in this paper. Let \\(\\mathbf{Q}\\in \\mathbb{R}^{X\\times Y\\times C}\\) be the trainable BEV queries whose element \\(\\mathbf{Q}_{x,y}\\in \\mathbb{R}^C\\) is a \\(C\\) -dimensional query feature at the \\((x,y)\\) th geo-location in the BEV space \\(S_{\\mathrm{BEV}}\\) .We use \\(\\mathbf{Q}\\) as the input to query from the front-view feature map \\(\\mathbf{F}_i^{(-t)}\\) to produce the corresponding BEV feature map." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Cross-view attention. As the key operation in the pose" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "synchronized BEV encoder, the proposed cross-view attention constructs a feature map in the BEV space \\(S_{\\mathrm{BEV}}\\) by absorbing information from the corresponding pixels in the front-view feature map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.469, + 0.216 + ], + "angle": 0, + "content": "Let \\(\\mathcal{P}_i^{(-t)}:S_{\\mathrm{BEV}}\\times \\mathcal{Z}\\to S_i^{(-t)}\\) be a project operation that maps a BEV index with a specific height index to a pixel index in the \\(i\\)th camera's front view at historical timestamp \\(t\\); that is," + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.231, + 0.39, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\left(u _ {i} ^ {(- t)}, v _ {i} ^ {(- t)}\\right) = \\mathcal {P} _ {i} ^ {(- t)} \\Big ((x, y, z) \\Big),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.469, + 0.323 + ], + "angle": 0, + "content": "where \\(z \\in \\mathcal{Z} = \\{1, \\dots, Z\\}\\). The project operation \\(\\mathcal{P}_i^{(-t)}\\) builds the geometric relationship between the BEV and a front view. The implementation of \\(\\mathcal{P}_i^{(-t)}\\) works as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.335, + 0.407, + 0.401 + ], + "angle": 0, + "content": "\\[\nz _ {i} ^ {(- t)} \\cdot \\left[ \\begin{array}{c} u _ {i} ^ {(- t)} \\\\ v _ {i} ^ {(- t)} \\\\ 1 \\end{array} \\right] = \\mathcal {T} _ {\\mathcal {S} _ {\\mathrm {B E V}} \\to \\mathcal {S} _ {i} ^ {(- t)}} \\cdot \\left[ \\begin{array}{c} x \\\\ y \\\\ z \\\\ 1 \\end{array} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.469, + 0.463 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}_{\\mathcal{S}_{\\mathrm{BEV}}\\to \\mathcal{S}_i^{(-t)}}\\in \\mathbb{R}^{3\\times 4}\\) is a transformation matrix that can be calculated by camera's intrinsic/ extrinsic parameters and vehicle-ego pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.469, + 0.546 + ], + "angle": 0, + "content": "Based on the project operation, the cross-view attention can trace visual features in the front view through a BEV index. Let \\(\\mathbf{B}_i^{(-t)}\\in \\mathbb{R}^{X\\times Y\\times C}\\) be the BEV feature map associated with the RBG image \\(\\mathbf{X}_i^{(-t)}\\). The \\((x,y)\\)th element of the BEV feature map is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.56, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {B} _ {i} ^ {(- t)}\\right) _ {x, y} = \\sum_ {z} f _ {\\mathrm {D A}} \\left(\\mathbf {Q} _ {x, y}, \\mathcal {P} _ {i} ^ {(- t)} (x, y, z), \\mathbf {F} _ {i} ^ {(- t)}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.824 + ], + "angle": 0, + "content": "where, \\( f_{\\mathrm{DA}}(\\cdot) \\) represents the deformable attention operation [58]. It allows BEV query \\( Q_{x,y} \\) only to interact with the front-view feature \\( \\mathbf{F}_i^{(-t)} \\) within its regions of interest, which is sampled around the reference point calculated by \\( \\mathcal{P}_i^{(-t)} \\). Since one BEV index might lead to multiple pixel indices in the front-view image because of various height possibilities in the 3D space. We thus sum over all possible heights in (1). To further aggregate BEV feature maps across all the \\( N \\) camera views, we simply take the average; that is, the BEV feature map at historical timestamp \\( t \\) is \\( \\mathbf{B}^{(-t)} = \\frac{1}{N}\\sum_{i}\\mathbf{B}_{i}^{(-t)} \\). Note that all the front-view features across time and from multiple cameras are synchronized into the same BEV space in one step (1), leading to less information distortion or deficiency issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We can successively apply the cross-view attention followed by feed forward networks and normalization layers for multiple times. Finally, we order BEV feature maps at multiple timestamps and obtain a temporal BEV feature map \\(\\mathcal{B} = [\\mathbf{B}^{(0)},\\mathbf{B}^{(-1)},\\dots,\\mathbf{B}^{(-T)}]\\in \\mathbb{R}^{(T + 1)\\times X\\times Y\\times C}\\)." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.088, + 0.868, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Figure 4. The network architecture of Spatial-Temporal Pyramid Transformer (STPT). Each encoder layer consists of an optional convolutional block for downsampling and Swin Transformer Blocks, while each decoder layer contains Swin Transformer Blocks and a deconvolutional block for upsampling. In the decoding process, we pre-define a set of future queries to represent future BEV states and query the features from encoders." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.447, + 0.843, + 0.464 + ], + "angle": 0, + "content": "3.3. Spatial-Temporal Pyramid Transformer" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.533 + ], + "angle": 0, + "content": "We further propose a novel spatial-temporal pyramid transformer (STPT) to learn spatial-temporal features more comprehensively and produce the future BEV states. The detailed structure of STPT is depicted in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Temporal BEV pyramid feature learning. We encode the input temporal BEV feature map \\(\\mathcal{B}\\) with four hierarchical layers. Each encoder layer is composed of an optional convolution layer with stride 2 to downsample the features and a swim transformer encoder, which is a stack of Swin Transformer blocks [28]. In our implementation, the window size of Swin-T blocks is set as (4,4). We can then obtain multi-scale spatial-temporal features \\(\\mathcal{B}_s\\in \\mathbb{R}^{(T + 1)\\times \\frac{X}{2^s}\\times \\frac{Y}{2^s}\\times C}\\), \\(s = 0,1,2,3\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.672, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Future BEV queries. Future BEV queries are defined to represent future BEV states and query the generated multiscale spatial-temporal features. There is a set of learnable future queries \\(\\{\\mathbf{Q}^{(t)}\\}\\), \\(t = 0, \\dots, T'\\), where \\(\\mathbf{Q}^{(t)}\\) has the same spatial dimension \\(\\frac{X}{8} \\times \\frac{Y}{8}\\) as \\(\\mathcal{B}_3\\). Separate learning embeddings are employed for future queries to differentiate the predicted BEV states over time. Additionally, a map feature generator is applied to generate high-dimensional features from \\(\\mathbf{B}^{(T)}\\) in order to extract information about the scene's geometry. To be specific, the same structure and parameters of the hdmap decoder head (excluding the last linear layer) are reused. The resulting map feature is added to all future queries to provide spatial information priors." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Future BEV state prediction. The decoding process contains corresponding four hierarchical layers as the en" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1372" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.089, + 0.825, + 0.213 + ], + "angle": 0, + "content": "
MethodRGB ResolutionFuture semantic seg.Future instance seg.FPS
IoU (Short)IoU (Long)VPQ (Short)VPQ (Long)
FIERY [16]224×48059.436.750.229.91.56
StretchBEV [1]224×48055.537.146.029.01.56
ST-P3 [17]224×480-38.9-32.11.43
BEVerse [56]256×70460.338.752.233.31.96
TBP-Former224×48064.741.956.736.92.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.225, + 0.894, + 0.281 + ], + "angle": 0, + "content": "Table 1. Prediction results on nuScenes [2] validation set. Intersection-over-Union (IoU) is used for future semantic segmentation and Video Panoptic Quality (VPQ) for future instance segmentation. Results are reported under two settings: short \\((30m\\times 30m)\\) range and long \\((100m\\times 100m)\\) range. Frame Per Second (FPS) means the inverse of inference time. All methods are tested under the same settings on a single NVIDIA A100. Our TBP-Former achieves SOTA performance and is still more computationally efficient than other methods." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.294, + 0.456, + 0.508 + ], + "angle": 0, + "content": "
MethodTemp.Veh. IoUPed. IoU
VED [31]23.311.9
VPN [36]28.210.3
PON [42]27.913.9
LSS [38]34.615.0
CVT [57]36.0-
Image2Map [44]40.2-
BEVFormer [24]44.4-
IVMP [49]✓36.817.4
FIERY [16]✓38.217.2
ST-P3 [17]✓40.114.5
TBP-Former static44.817.2
TBP-Former✓46.218.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.562 + ], + "angle": 0, + "content": "Table 2. Perception results on nuScenes [2] validation set. Results of vehicles and pedestrians are compared by segmentation IoU. Temp. indicates whether temporal information is involved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.678 + ], + "angle": 0, + "content": "coding process. Unlike the encoding process, \\(\\{\\mathbf{Q}^{(t)}\\}\\) is used as the query input of Swin-T block and performs cross attention with the encoded features \\(\\mathbf{E}_3\\). After the first decoding layer, the output of each layer is used as the query input of the next layer. Similar to encoding layers, the deconvolution layer is optionally applied to upsample the decoded features. The simplified process can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.688, + 0.443, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {s} = \\left\\{ \\begin{array}{l l} \\operatorname {S w i n T} (\\mathcal {B} _ {3}, \\{\\mathbf {Q} ^ {(t)} \\}), & s = 3 \\\\ \\operatorname {S w i n T} (\\mathcal {B} _ {s}, \\operatorname {D e C o n v} (\\mathbf {D} _ {s + 1})), & s = 0, 1, 2 \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.741, + 0.468, + 0.822 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_s\\in \\mathbb{R}^{(T' + 1)\\times \\frac{X}{2^s}\\times \\frac{Y}{2^s}\\times C}, s = 0,1,2,3\\) are decoded features. The future temporal BEV feature map at the 0th scale is a temporal sequence of final predicted BEV states; that is, \\(\\mathcal{D}_0 = [\\mathbf{B}_*^{(0)},\\mathbf{B}_*^{(1)},\\dots,\\mathbf{B}_*^{(T')}]\\), where \\(\\mathbf{B}_*^{(t)}\\in \\mathbb{R}^{X\\times Y\\times C}\\) is the BEV state at future time stamp \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.832, + 0.265, + 0.847 + ], + "angle": 0, + "content": "3.4. Multi-head decoder" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The future temporal BEV feature map is fed into the multi-task decoder heads to generate various outputs for dynamic scene understanding, see Fig. 5. We follow the output" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.417 + ], + "angle": 0, + "content": "setting in [16] that predicts BEV semantic segmentation, instance center, instance offset, and future flow for joint perception and prediction. Meanwhile, we set up an additional HD map decoder head to predict basic traffic scene elements including drivable areas and lanes. The map decoder head can not only provide scene information for subsequent planning and control modules but also give guidance to the prediction process, see sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.423, + 0.634, + 0.439 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.448, + 0.691, + 0.464 + ], + "angle": 0, + "content": "4.1. Dataset and settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.699 + ], + "angle": 0, + "content": "We use nuScenes [2] datasets to evaluate our approach. NuScenes contains 1000 scenes, each of which has 20 seconds annotated at \\(2\\mathrm{Hz}\\). In nuScenes, the images are captured by 6 cameras with a small overlap in the field of view, which guarantees the cameras cover the full \\(360^{\\circ}\\) field of view. For model input, raw camera images with the size of \\(900 \\times 1600\\) are resized and cropped to a resolution of \\(224 \\times 480\\). We follow the training and evaluating settings used in previous methods [1,16,17,56] for fair comparisons, which use 1.0 second past states and current state to predict 2.0 seconds of the future states. It corresponds to predicting 4 future frames based on 3 observed frames. The size of the generated BEV grid map is \\(200 \\times 200\\). Each grid has a range of \\(0.5m \\times 0.5m\\), which means the perception and prediction range is \\(100m \\times 100m\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.775 + ], + "angle": 0, + "content": "For training, we use AdamW [30] with a weight decay 0.01 to optimize the models. The learning rate is initialized as \\(10^{-4}\\) and decays with a cosine annealing scheduler [29]. All models are trained on 4 NVIDIA A100 GPUs for 10 epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.597, + 0.801 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following previous works [1, 16, 17, 56], we mainly use two metrics for evaluation. The first is Intersection over Union (IoU), which measures the quality of segmentation at each frame. The second is Video Panoptic Quality (VPQ), which is used to measure the consistency of the detected instances over time and the accuracy of the segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1373" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.817, + 0.226 + ], + "angle": 0, + "content": "
Exp.Warp.Sync.SLQSPEFuture semantic seg.Future instance seg.
Short (IoU)Long (IoU)Short (VPQ)Long (VPQ)
1✓58.738.450.631.8
2✓✓60.838.652.433.4
3✓✓✓62.040.753.234.3
4✓63.040.854.134.3
5✓✓63.841.155.835.8
6✓✓✓64.741.956.736.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.237, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Table 3. Ablation of our proposed architecture. Ablation results for our PoseSync BEV Encoder (Sync.), the learnable future queries, and the spatial embedding are presented. Exp. 1-3 use the traditional warping methods to align temporal BEV features. Separate learnable queries (SLQ) represent using separate learnable future queries instead of utilizing the same query with temporal positional encoding. Spatial positional embedding (SPE) represents using spatial scene representations in future prediction queries." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.303, + 0.446, + 0.392 + ], + "angle": 0, + "content": "
Temporal modelIoUVPQVRQVSQ
MotionNet†[53]35.430.643.171.1
FIERY†[16]38.332.145.470.7
BEVerse†[56]40.234.048.070.9
TBP-Former41.936.951.572.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.402, + 0.472, + 0.473 + ], + "angle": 0, + "content": "Table 4. Ablation for the prediction model. †: We use MotionNet, FIERY and BEVerse to replace our prediction model for comparison, and the BEV encoder and task heads are the same. Besides IoU and VPQ, we also use Video Recognition Quality (VRQ) and Video Segmentation quality (VSQ) for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.482, + 0.432, + 0.586 + ], + "angle": 0, + "content": "
AugmentationPerceptionPrediction
CamBEVVeh.Ped.IoUVPQ
45.017.740.534.4
✓44.818.540.935.3
✓45.318.641.435.6
✓✓46.218.641.936.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.653 + ], + "angle": 0, + "content": "Table 5. Ablation for data augmentation strategies. Perception of Vehicles and Pedestrians with different data augmentation strategies are evaluated on segmentation IoU. Prediction results are evaluated on segmentation IoU and Video Panoptic Quality." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.663, + 0.273, + 0.677 + ], + "angle": 0, + "content": "The formula is shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.686, + 0.409, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V P Q} = \\sum_ {t = 0} ^ {H} \\frac {\\sum_ {\\left(p _ {t} , q _ {t}\\right) \\in T P _ {t}} \\operatorname {I o U} \\left(p _ {t} , q _ {t}\\right)}{\\left| T P _ {t} \\right| + \\frac {1}{2} \\left| F P _ {t} \\right| + \\frac {1}{2} \\left| F N _ {t} \\right|}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.47, + 0.772 + ], + "angle": 0, + "content": "where \\( H \\) is the sequence length, \\( TP_{t} \\) represents the set of true positives, \\( FP_{t} \\) represents the set of false positives and \\( FN_{t} \\) represents the set of false negatives at timestamp \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.776, + 0.203, + 0.79 + ], + "angle": 0, + "content": "4.3. PnP results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Perception and Prediction. Table 1 compares TBP-Former with other methods of perception and prediction task based on multi-view cameras. We see that i) we achieve state-of-the-art performance and exceed previous methods by a large margin. ii) Even though BEVerse has larger RGB resolutions, TBP-Former still surpasses their performance on IoU by \\(7.3\\% / 8.3\\%\\) for short/long settings, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.365 + ], + "angle": 0, + "content": "TBP-Former also improves the VPQ by \\(12.1\\% / 10.8\\%\\). iii) Apart from the performance improvement, TBF-Former also has a larger FPS compared to other methods. Its inference speed is \\(25\\%\\) faster than BEVerse's." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Fig. 5 shows the visualization results of our proposed method. We see that i) almost all the objects are detected correctly except for those occluded ones. ii) TBP-Former is capable of capturing the motion information in past frames and precisely predicting the vehicles' trajectories by occupancy and flow. Compared with FIERY [16], TBP-Former is closer to the ground truth. iii) TBP-Former does a better job than FIERY when predicting vehicles' turning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.894, + 0.622 + ], + "angle": 0, + "content": "Perception Only. Table 2 compares the results of plenty of state-of-the-art methods on perception (segmentation) task. We see that our static model, which does not contain temporal information, can achieve 44.8 and 17.2 IoU of vehicles and pedestrians. With the input of temporal sequences, the performance improves further since auxiliary information is provided for better perception. The state-of-the-art results prove the effectiveness of the novel design of our Pose-synchronized BEV encoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.625, + 0.605, + 0.639 + ], + "angle": 0, + "content": "4.4. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Effectiveness of PoseSync View Projection. The Exp. 1&4, 2&5, 3&6 in Table 3 compare the proposed PoseSync View Projection and the existing feature warping methods. We see that the proposed method always achieves better performance when other settings remain the same. The reasons are that: i) PoseSync View Projection based on Deformable Attention can guarantee the precise correspondence between BEV grids and image features. ii) Our projection method can alleviate distortion and our-of-range issues when synchronizing sequential BEV features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of the designed future queries. In Exp. 1&3 in Table 3, we utilize the identical query with temporal positional encoding for future queries. Exp. 2&4 in Table 3 demonstrate that using separate learnable embedding for future queries can achieve better performance. Exp. 3&6 in Table 3 validate the efficacy of the proposed spatial priors for future queries. The generated high" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.881, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Figure 5. Demonstration of our results compared with FIERY and Ground Truth. Different vehicles are assigned with different colors in order to make a distinction. The darker parts represent the perception of the current frame, and the lighter parts represent the prediction of the vehicles in future frames. The visualization is based on the predicted occupancy and flow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.471, + 0.626 + ], + "angle": 0, + "content": "dimensional map features provide the prediction model with useful geographic information. The additional spatial information can aid the prediction and lead to better scene forecasting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Effectiveness of STPT. Table 4 compares STPT with popular CNN-based [53] and RNN-based [16, 56] methods. We keep all the settings the same except for temporal modeling. To be specific, the size of input images \\((224 \\times 480)\\), image backbones and BEV feature extractor are the same. And then we plug their temporal models into our architecture. We see that i) STPT model performs better in all four metrics, including semantic segmentation IoU and three instance segmentation metrics from the video prediction area. ii) Our reproduced temporal models achieve higher performance than the original implements. This further validates the effectiveness and power of our BEV feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Data Augmentation. We perform both image-view and BEV augmentations. The image-view augmentations include random scaling, rotation and flip of the input images. The BEV augmentations include similar operations on both BEV representations and corresponding ground truth labels. Table 5 compares the results of different data augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.644 + ], + "angle": 0, + "content": "strategies. We see that i) both augmentation methods improve the performance when used separately. ii) The combination of two methods works better than any single approach. Introducing data augmentation strategies is beneficial to the model's robustness and generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.62, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.817 + ], + "angle": 0, + "content": "This paper proposes a novel TBP-Former for vision-centric joint perception and prediction. We design a pose-synchronized BEV encoder module using a cross-view attention mechanism to solve the distortion issues in previous works. Furthermore, we propose a powerful spatial-temporal pyramid transformer for BEV feature extraction and BEV state prediction. Experiments show that i) TBP-Former improves the prediction performance over state-of-the-art methods significantly; and ii) both PoseSync BEV Encoder and STPT contribute to better performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.894, + 0.892 + ], + "angle": 0, + "content": "Acknowledgement This research is partially supported by National Natural Science Foundation of China under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1375" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Sergio Casas, Cole Gulino, Renjie Liao, and Raquel Urtasun. Spagnn: Spatially-aware graph neural networks for relational behavior forecasting from sensor data. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9491-9497. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[8] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597–1600. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[9] Artem Filatov, Andrey Rykov, and Viacheslav Murashkin. Any motion detector: Learning class-agnostic scene dynamics from a sequence of lidar point clouds. In 2020 IEEE international conference on robotics and automation (ICRA), pages 9498–9504. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[10] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17441-17451, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[11] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[12] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[13] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. A simple baseline for bev perception without lidar. arXiv e-prints, pages arXiv-2206, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[15] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[18] Yue Hu, Siheng Chen, Ya Zhang, and Xiao Gu. Collaborative motion predication via neural motion message passing. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[19] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), November 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[21] Kuan-Hui Lee, Matthew Kliemann, Adrien Gaidon, Jie Li, Chao Fang, Sudeep Pillai, and Wolfram Burgard. Pillarflow: End-to-end birds-eye-view flow estimation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2007-2013. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[22] Peixuan Li and Jieyu Jin. Time3d: End-to-end joint monocular 3d object detection and tracking for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3885-3894, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[23] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[25] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1376" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[31] Chenyang Lu, Marinus Jacobus Gerardus van de Molengraft, and Gijs Dubbelman. Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters, 4(2):445-452, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[32] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[33] Chenxu Luo, Xiaodong Yang, and Alan Yuille. Self-supervised pillar motion learning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3183-3192, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[34] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified multi-task model for behavior prediction and planning. arXiv e-prints, pages arXiv-2106, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[36] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[37] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] John Phillips, Julieta Martinez, Ioan Andrei Bársan, Sergio Casas, Abbas Sadat, and Raquel Urtasun. Deep multi-task learning for joint localization, perception, and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4679-4689, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[40] Zequn Qin, Jingyu Chen, Chao Chen, Xiaozhi Chen, and Xi Li. Uniformer: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. arXiv preprint arXiv:2207.08536, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[41] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent video transformer. arXiv preprint arXiv:2006.10704, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[43] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In European Conference on Computer Vision, pages 414-430. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[44] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In 2022 International Conference on Robotics and Automation (ICRA), pages 9200-9206. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[45] Marcel Schreiber, Vasileios Belagiannis, Claudius Gläser, and Klaus Dietmayer. Dynamic occupancy grid mapping with recurrent neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6717-6724. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[46] Meet Shah, Zhiling Huang, Ankit Laddha, Matthew Langford, Blake Barber, Sidney Zhang, Carlos Vallespi-Gonzalez, and Raquel Urtasun. Liranet: End-to-end trajectory prediction using spatio-temporal radar fusion. arXiv preprint arXiv:2010.00731, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[47] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[49] Hengli Wang, Peide Cai, Yuxiang Sun, Lujia Wang, and Ming Liu. Learning interpretable end-to-end vision-based motion planning for autonomous driving with optical flow distillation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13731-13737. IEEE, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[50] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17093–17102, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[51] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying unknown instances for au" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1377" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "tonomous driving. In Conference on Robot Learning, pages 384-393. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.203 + ], + "angle": 0, + "content": "[53] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.274 + ], + "angle": 0, + "content": "[54] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.345 + ], + "angle": 0, + "content": "[55] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.414 + ], + "angle": 0, + "content": "[56] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.418, + 0.468, + 0.471 + ], + "angle": 0, + "content": "[57] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.474, + 0.468, + 0.528 + ], + "angle": 0, + "content": "[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1378" + } + ] +] \ No newline at end of file diff --git a/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_origin.pdf b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6fc114d8582e05d5d7fd60ab42bedec30e7a624d --- /dev/null +++ b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/f7f592e0-a2b8-46fb-a0f7-4d6e06bddd8a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ece36aa6947f676f3afe0c3b1997ecd1698bc43c62d0d1044875bfdc0fc62b19 +size 4657804 diff --git a/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/full.md b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..14cf0b6340b57f01b494353223a2b4c17b8457bf --- /dev/null +++ b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/full.md @@ -0,0 +1,297 @@ +# TBP-Former: Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving + +Shaoheng Fang $^{1*}$ Zi Wang $^{1*}$ Yiqi Zhong $^{2}$ Junhao Ge $^{1}$ Siheng Chen $^{1,3\dagger}$ + +$^{1}$ Cooperative Medianet Innovation Center, Shanghai Jiao Tong University + +$^{2}$ Department of Computer Science, University of Southern California $^{3}$ Shanghai AI Laboratory + +1{shfang，w4ngz1，cancaries，sihengc}@sjtu.edu.cn 2{yiqizhon}@usc.edu + +# Abstract + +Vision-centric joint perception and prediction (PnP) has become an emerging trend in autonomous driving research. It predicts the future states of the traffic participants in the surrounding environment from raw RGB images. However, it is still a critical challenge to synchronize features obtained at multiple camera views and timestamps due to inevitable geometric distortions and further exploit those spatial-temporal features. To address this issue, we propose a temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP, which includes two novel designs. First, a pose-synchronized BEV encoder is proposed to map raw image inputs with any camera pose at any time to a shared and synchronized BEV space for better spatial-temporal synchronization. Second, a spatial-temporal pyramid transformer is introduced to comprehensively extract multi-scale BEV features and predict future BEV states with the support of spatial priors. Extensive experiments on nuScenes dataset show that our proposed framework overall outperforms all state-of-the-art vision-based prediction methods. Code is available at: https://github.com/MediaBrain-SJTU/TBP-Former + +# 1. Introduction + +As one of the most fascinating engineering projects, autonomous driving has been an aspiration for many researchers and engineers for decades. Although significant progress has been made, it is still an open question in designing a practical solution to achieve the goal of full self-driving. A traditional and common solution consists of a sequential stack of perception, prediction, planning, and control. Despite the idea of divide-and-conquer having achieved tremendous success in developing software sys + +![](images/e14ef5292e54d78c4693f68ae053ebc277cef5b277c65fceee5ce62d613c6f82.jpg) +Previous works + +![](images/143dbc6f80836fda5ead3f6fb745c36c14d7a4a92e17a75a79885c92e1f1e081.jpg) +TBP-Former + +![](images/8d43e3d772a538bacefdb6cfc5eacc582b42d4a95beb85b679b115888f99fc7e.jpg) +(a) Temporal BEV feature alignment +Figure 1. Two major challenges in vision-based perception and prediction are (a) how to avoid distortion and deficiency when aggregating features across time and camera views; and (b) how to achieve spatial-temporal feature learning for prediction. Our Pose-Synchronized BEV Encoder can precisely map the visual features into synchronized BEV space, and Spatial-Temporal Pyramid Transformer extracts feature at multiple scales. + +![](images/002bf9e04e7761f1d780ff6b49709ad83265ad64d6e1d315907695cf48154154.jpg) +(b) Spatial-temporal feature modeling for prediction + +tems, a long stack could cause cascading failures in an autonomous system. Recently, there is a trend to combine multiple parts in an autonomous system to be a joint module, cutting down the stack. For example, [25, 46] consider joint perception and prediction and [5, 43] explore joint prediction and planning. This work focuses on joint perception and prediction. + +The task of joint perception and prediction (PnP) aims to predict the current and future states of the surrounding environment with the input of multi-frame raw sensor + +data. The output current and future states would directly serve as the input for motion planning. Recently, many PnP methods are proposed based on diverse sensor input choices. For example, [4, 25, 34] take multi-frame LiDAR point clouds as input and achieve encouraging 3D detection and trajectory prediction performances simultaneously. Recently, the rapid development of vision-centric methods offers a new possibility to provide a cheaper and easy-to-deploy solution for PnP. For instance, [1, 16, 17] only uses RGB images collected by multiple cameras to build PnP systems. Meanwhile, without precise 3D measurements, vision-centric PnP is more technically challenging. Therefore, this work aims to advance this direction. + +The core of vision-centric PnP is to learn appropriate spatial-temporal feature representations from temporal image sequences. It is a crux and difficult from three aspects. First, since the input and the output of vision-centric PnP are supported in camera front-view (FV) and bird's-eye-view (BEV) respectively, one has to deal with distortion issues during geometric transformation between two views. Second, when the vehicle is moving, the view of the image input is time-varying and it is thus nontrivial to precisely map visual features across time into a shared and synchronized space. Third, since information in temporal image sequences is sufficiently rich for humans to accurately perceive the environment, we need a powerful learning model to comprehensively exploit spatial-temporal features. + +To tackle these issues, previous works on vision-centric PnP consider diverse strategies. For example, [16, 56] follows the method in [38] to map FV features to BEV features, then synchronizes BEV features across time via rigid transformation, and finally uses a recurrent network to exploit spatial-temporal features. However, due to the image discretization nature and depth estimation uncertainty, simply relying on rigid geometric transformations would cause inevitable distortion; see Fig. 1. Some other work [49] transforms the pseudo feature point cloud to current ego coordinates and then pools the pseudo-lidar to BEV features; however, this approach encounters deficiency due to the limited sensing range in perception. Meanwhile, many works [16, 17, 56] simply employ recurrent neural networks to learn the temporal features from multiple BEV representations, which is hard to comprehensively extract spatial-temporal features. + +To promote more reliable and comprehensive feature learning across views and time, we propose the temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP. The proposed TBP-Former includes two key innovations: i) pose-synchronized BEV encoder, which leverages a pose-aware cross-attention mechanism to directly map a raw image input with any camera pose at any time to the corresponding feature map in a shared and synchronized BEV space; and ii) spatial-temporal pyra + +mid transformer, which leverages a pyramid architecture with Swin-transformer [28] blocks to learn comprehensive spatial-temporal features from sequential BEV maps at multiple scales and predict future BEV states with a set of future queries equipped with spatial priors. + +Compared to previous works, the proposed TBP-Former brings benefits from two aspects. First, previous works [16, 17, 24, 56] consider FV-to-BEV transformation and temporal synchronization as two separate steps, each of which could bring distortion due to discrete depth estimation and rigid transformation; while we merge them into one step and leverage both geometric transformation and attention-based learning ability to achieve spatial-temporal synchronization. Second, previous works [16, 53] use RNNs or 3D convolutions to learn spatial-temporal features; while we leverage a powerful pyramid transformer architecture to comprehensively capture spatial-temporal features, which makes prediction more effective. + +To summarize, the main contributions of our work are: + +- To tackle the distortion issues in mapping temporal image sequences to a synchronized BEV space, we propose a pose-synchronized BEV encoder (PoseSync BEV Encoder) based on cross-view attention mechanism to extract quality temporal BEV features. +- We propose a novel Spatial-Temporal Pyramid Transformer (STPT) to extract multi-scale spatial-temporal features from sequential BEV maps and predict future BEV states according to well-elaborated future queries integrated with spatial priors. +- Overall, we propose TBP-Former, a vision-based joint perception and prediction framework for autonomous driving. TBP-Former achieves state-of-the-art performance on nuScenes [2] dataset for the vision-based prediction task. Extensive experiments show that both PoseSync BEV Encoder and STPT contribute greatly to the performance. Due to the decoupling property of the framework, both proposed modules can be easily utilized as alternative modules in any vision-based BEV prediction framework. + +# 2. Related Work + +# 2.1. Joint Perception and Prediction + +As the two core system modules of autonomous driving, how to conduct perception and prediction tasks jointly has received a lot of attention. Traditional approaches [3,4,25,34,39] formulate this joint task as a trajectory prediction problem that relies on the perception outputs of 3D object detection and tracking. The dependency on intermediate results tends to accumulate errors and lacks the capacity to perceive unknown objects [52,53]. Subsequently, + +![](images/ecf270f42ebad10f4d13bcd27c06dd85e580b8c2d4aa88f99dd595d12d02bc97.jpg) +Figure 2. An overview of TBP-Former architecture. Taking consecutive surrounding camera images as inputs, TBP-Former first generates image-space features and uses the PoseSync BEV Encoder to map front-view features to BEV features in a shared and synchronized BEV space. Then the BEV features from multiple frames are processed by the Spatial-Temporal Pyramid Transformer to extract BEV spatial-temporal features and predict future BEV states in order. In this process, high-level scene representations are generated from the last frame BEV feature as spatial priors to guide the prediction. Finally, the well-predicted future states are sent to decoder heads for joint perception and prediction tasks. + +instance-free methods [9,21,33,43,45,53] that predict dense future semantic occupancy and flow have become a growing trend to simplify the understanding of dynamic scenes. Also, several recent works [16, 17, 56] explore joint perception and prediction in the form of dense occupancy and flow using only surrounding camera input. + +In many previous works [4, 25, 39, 43], raster HD (high-definition) maps play an important role as input of the frameworks. HD maps can provide strong priors to guide the predicted results to follow the traffic lanes. However, in practice, HD maps are laborious and costly to produce and require frequent maintenance. Instead of using off-the-peg HD maps, we follow the philosophy of [5, 7, 23] in predicting online HD maps but propose to learn high-level scene geometry representations from real-time sensor inputs and take these representations as priors for the prediction task. + +# 2.2. BEV Representations + +BEV representations provide a unified and physical-interpretable way to represent the rich information of road, moving objects and occlusion in a traffic scene, which can be easily utilized for downstream tasks such as motion prediction, planning and control, etc. For camera-based methods, how to solve the problem of projecting features from perspective view to BEV is a major challenge. Some learnable methods use MLP [23, 36, 42] or transformer network [37, 57] to implicitly reason the relationship between two different views. LSS [38] proposes the approach of predicting depth distribution per pixel on 2D features, then 'lifting' the 2D features according to the corresponding depth distribution to BEV space. Numerous works, aim + +ing at tasks of BEV perception [19, 20, 54], motion prediction [1, 16, 17, 56], lidar-camera fusion [26], etc., follow this form to generate BEV representations. Also, some methods [6, 13, 24] explicitly establish the correspondence from BEV location to image-view pixel using homography between image and BEV plane and achieve attractive performance in diverse tasks. + +However, when dealing with temporal information, most methods [1, 16, 17, 24, 56] warp history BEV representations according to the variation of ego poses. Due to the pre-defined fixed range and size of the BEV grid, rotation and translation operations may cause distortion and out-of-range problems when aligning history BEV maps to current ego coordinates. Though [40] introduces a similar operation to us to integrate historical information into the current frame, the design of their model is unable to predict future states. To alleviate these issues, we propose a PoseSync BEV Encoder module based on deformable attention to generate pose-synchronized BEV representations from temporally consecutive image-view input. + +# 2.3. Spatial-Temporal Modeling + +In the BEV prediction field, how to design a temporal model to aggregate spatial-temporal information is a critical problem. Existing modeling methods can be classified into three categories: RNN-based, CNN-based and transformer-based. RNN-based methods [1, 16-18, 43, 56] utilize recurrent models such as LSTM [15], GRU [8] to predict the future latent states. Though the recurrent model is powerful to model temporal relationships, it is time-consuming for constraints in the parallelization of computation. Besides, + +![](images/2c7b4199d8890da50565946c771cae33c5f05eb0a2720c25da883f3f9515e474.jpg) +(A) +(B) + +![](images/999941865914e9fd53b96a15718d940b3b6514b1b6f1122f5345fd09a902cd3f.jpg) +Figure 3. The PoseSync BEV Encoder (A) takes front-view features and camera poses as input and then maps to BEV space. The core to generate BEV features in a synchronized way is the Pose-Aware Cross Attention. Its cross-view attention mechanism is depicted in (B), where front-view features from different frames of a dynamic vehicle are projected into a uniform BEV space. + +some CNN-based methods [5, 34, 50, 53] concatenate BEV features in the time dimension and take advantage of 3D convolution to extract spatial-temporal features. + +Due to the great power of transformer [48] in sequence modeling, it has shown promise in many temporal modeling tasks such as trajectory prediction [11, 35, 55], object tracking [22], video prediction [12, 41, 51], video interpolation [10, 32, 47], etc. For BEV perception, [24, 27] utilize self-attention to model temporal information from multiple frames to boost perception task. [22] leverage self-attention to aggregate spatial information and cross-attention to exploit affinities among sequence frames. To explore the capacity of transformer model in BEV spatial-temporal modeling, we propose a novel Spatial-Temporal Pyramid Transformer (STPT) architecture with future queries for BEV spatial-temporal features extraction and BEV future states prediction. + +# 3. Methodology + +# 3.1. Overview Architecture + +The overall architecture of the proposed TBP-Former is illustrated in Fig. 2. It takes the input of multi-view images with the corresponding camera poses at consecutive $T$ timestamps. The final output includes BEV map segmentation for current scene understanding and occupancy flow for motion prediction. The whole TBP-Former can be decoupled into three parts: (i) pose-synchronized BEV encoder, which maps raw image sequences into feature maps in a spatial-temporal-synchronized BEV space; (ii) spatial-temporal pyramid transformer, which achieves comprehensive feature learning at multiple spatial and temporal scales; and (iii) a multi-head decoder, which takes the spatial and temporal features to achieve scene understanding and motion prediction. We will elaborate on each part in the following subsections. + +# 3.2. Pose-Synchronized BEV Encoder + +Given images collected at multiple time stamps and from various camera poses, we aim to generate the corresponding feature maps in a shared and synchronized BEV space. Different from many previous works that synchronize spatial and temporal information in two separate steps, the proposed pose-synchronized BEV encoder leverage both geometric prior and learning ability to achieve one-step synchronization, alleviating distortion effects. Following the previous transformer-based method [24], this encoder adopts a transformer architecture whose core is a novel cross-view attention operation. + +Front-view feature map. Let $\mathcal{X} = \{\mathbf{X}_i^{(-t)}\}_{i=1,t=0}^{N,T}$ be the input multi-frame multi-view images, where $N$ is the number of cameras, $T$ is the number of historical timestamps and $\mathbf{X}_i^{(-t)} \in \mathbb{R}^{H \times W \times 3}$ is the RBG image captured by the $i$ th camera at historical time stamp $t$ . Note that each front-view image is associated with a different camera pose. Let $S_i^{(-t)} = \{(u_i^{(-t)}, v_i^{(-t)})\}_{1,1}^{H,W}$ be the pixel indices of the $i$ th camera's front-view space, whose image size is $H \times W$ . We feed each RBG image $\mathbf{X}_i^{(-t)}$ into a shared backbone network (our implementation uses ResNet-101 [14]) and obtain the corresponding front-view feature map $\mathbf{F}_i^{(-t)} \in \mathbb{R}^{H' \times W' \times C}$ with $C$ the channel number, which is also supported on the front-view space $S_i^{(-t)}$ . + +BEV queries. Let $S_{\mathrm{BEV}} = \{(x,y)\}_{x = 1,y = 1}^{X,Y}$ be the BEV grid indices, reflecting the $X\times Y$ BEV grid space based on the vehicle-ego pose at the current timestamp. Note that $S_{\mathrm{BEV}}$ is the only BEV space we work with in this paper. Let $\mathbf{Q}\in \mathbb{R}^{X\times Y\times C}$ be the trainable BEV queries whose element $\mathbf{Q}_{x,y}\in \mathbb{R}^C$ is a $C$ -dimensional query feature at the $(x,y)$ th geo-location in the BEV space $S_{\mathrm{BEV}}$ .We use $\mathbf{Q}$ as the input to query from the front-view feature map $\mathbf{F}_i^{(-t)}$ to produce the corresponding BEV feature map. + +Cross-view attention. As the key operation in the pose + +synchronized BEV encoder, the proposed cross-view attention constructs a feature map in the BEV space $S_{\mathrm{BEV}}$ by absorbing information from the corresponding pixels in the front-view feature map. + +Let $\mathcal{P}_i^{(-t)}:S_{\mathrm{BEV}}\times \mathcal{Z}\to S_i^{(-t)}$ be a project operation that maps a BEV index with a specific height index to a pixel index in the $i$ th camera's front view at historical timestamp $t$ ; that is, + +$$ +\left(u _ {i} ^ {(- t)}, v _ {i} ^ {(- t)}\right) = \mathcal {P} _ {i} ^ {(- t)} \Big ((x, y, z) \Big), +$$ + +where $z \in \mathcal{Z} = \{1, \dots, Z\}$ . The project operation $\mathcal{P}_i^{(-t)}$ builds the geometric relationship between the BEV and a front view. The implementation of $\mathcal{P}_i^{(-t)}$ works as + +$$ +z _ {i} ^ {(- t)} \cdot \left[ \begin{array}{c} u _ {i} ^ {(- t)} \\ v _ {i} ^ {(- t)} \\ 1 \end{array} \right] = \mathcal {T} _ {\mathcal {S} _ {\mathrm {B E V}} \to \mathcal {S} _ {i} ^ {(- t)}} \cdot \left[ \begin{array}{c} x \\ y \\ z \\ 1 \end{array} \right], +$$ + +where $\mathcal{T}_{\mathcal{S}_{\mathrm{BEV}}\to \mathcal{S}_i^{(-t)}}\in \mathbb{R}^{3\times 4}$ is a transformation matrix that can be calculated by camera's intrinsic/ extrinsic parameters and vehicle-ego pose. + +Based on the project operation, the cross-view attention can trace visual features in the front view through a BEV index. Let $\mathbf{B}_i^{(-t)}\in \mathbb{R}^{X\times Y\times C}$ be the BEV feature map associated with the RBG image $\mathbf{X}_i^{(-t)}$ . The $(x,y)$ th element of the BEV feature map is obtained as + +$$ +\left(\mathbf {B} _ {i} ^ {(- t)}\right) _ {x, y} = \sum_ {z} f _ {\mathrm {D A}} \left(\mathbf {Q} _ {x, y}, \mathcal {P} _ {i} ^ {(- t)} (x, y, z), \mathbf {F} _ {i} ^ {(- t)}\right), \tag {1} +$$ + +where, $f_{\mathrm{DA}}(\cdot)$ represents the deformable attention operation [58]. It allows BEV query $Q_{x,y}$ only to interact with the front-view feature $\mathbf{F}_i^{(-t)}$ within its regions of interest, which is sampled around the reference point calculated by $\mathcal{P}_i^{(-t)}$ . Since one BEV index might lead to multiple pixel indices in the front-view image because of various height possibilities in the 3D space. We thus sum over all possible heights in (1). To further aggregate BEV feature maps across all the $N$ camera views, we simply take the average; that is, the BEV feature map at historical timestamp $t$ is $\mathbf{B}^{(-t)} = \frac{1}{N}\sum_{i}\mathbf{B}_{i}^{(-t)}$ . Note that all the front-view features across time and from multiple cameras are synchronized into the same BEV space in one step (1), leading to less information distortion or deficiency issues. + +We can successively apply the cross-view attention followed by feed forward networks and normalization layers for multiple times. Finally, we order BEV feature maps at multiple timestamps and obtain a temporal BEV feature map $\mathcal{B} = [\mathbf{B}^{(0)},\mathbf{B}^{(-1)},\dots,\mathbf{B}^{(-T)}]\in \mathbb{R}^{(T + 1)\times X\times Y\times C}$ . + +![](images/98c466355ad7037ecd88acf0f26e1b8e44eebe2b1f9dde4fd776cd2202957f94.jpg) +Figure 4. The network architecture of Spatial-Temporal Pyramid Transformer (STPT). Each encoder layer consists of an optional convolutional block for downsampling and Swin Transformer Blocks, while each decoder layer contains Swin Transformer Blocks and a deconvolutional block for upsampling. In the decoding process, we pre-define a set of future queries to represent future BEV states and query the features from encoders. + +# 3.3. Spatial-Temporal Pyramid Transformer + +We further propose a novel spatial-temporal pyramid transformer (STPT) to learn spatial-temporal features more comprehensively and produce the future BEV states. The detailed structure of STPT is depicted in Fig. 4. + +Temporal BEV pyramid feature learning. We encode the input temporal BEV feature map $\mathcal{B}$ with four hierarchical layers. Each encoder layer is composed of an optional convolution layer with stride 2 to downsample the features and a swim transformer encoder, which is a stack of Swin Transformer blocks [28]. In our implementation, the window size of Swin-T blocks is set as (4,4). We can then obtain multi-scale spatial-temporal features $\mathcal{B}_s\in \mathbb{R}^{(T + 1)\times \frac{X}{2^s}\times \frac{Y}{2^s}\times C}$ , $s = 0,1,2,3$ . + +Future BEV queries. Future BEV queries are defined to represent future BEV states and query the generated multiscale spatial-temporal features. There is a set of learnable future queries $\{\mathbf{Q}^{(t)}\}$ , $t = 0, \dots, T'$ , where $\mathbf{Q}^{(t)}$ has the same spatial dimension $\frac{X}{8} \times \frac{Y}{8}$ as $\mathcal{B}_3$ . Separate learning embeddings are employed for future queries to differentiate the predicted BEV states over time. Additionally, a map feature generator is applied to generate high-dimensional features from $\mathbf{B}^{(T)}$ in order to extract information about the scene's geometry. To be specific, the same structure and parameters of the hdmap decoder head (excluding the last linear layer) are reused. The resulting map feature is added to all future queries to provide spatial information priors. + +Future BEV state prediction. The decoding process contains corresponding four hierarchical layers as the en + +
MethodRGB ResolutionFuture semantic seg.Future instance seg.FPS
IoU (Short)IoU (Long)VPQ (Short)VPQ (Long)
FIERY [16]224×48059.436.750.229.91.56
StretchBEV [1]224×48055.537.146.029.01.56
ST-P3 [17]224×480-38.9-32.11.43
BEVerse [56]256×70460.338.752.233.31.96
TBP-Former224×48064.741.956.736.92.44
+ +Table 1. Prediction results on nuScenes [2] validation set. Intersection-over-Union (IoU) is used for future semantic segmentation and Video Panoptic Quality (VPQ) for future instance segmentation. Results are reported under two settings: short $(30m\times 30m)$ range and long $(100m\times 100m)$ range. Frame Per Second (FPS) means the inverse of inference time. All methods are tested under the same settings on a single NVIDIA A100. Our TBP-Former achieves SOTA performance and is still more computationally efficient than other methods. + +
MethodTemp.Veh. IoUPed. IoU
VED [31]23.311.9
VPN [36]28.210.3
PON [42]27.913.9
LSS [38]34.615.0
CVT [57]36.0-
Image2Map [44]40.2-
BEVFormer [24]44.4-
IVMP [49]✓36.817.4
FIERY [16]✓38.217.2
ST-P3 [17]✓40.114.5
TBP-Former static44.817.2
TBP-Former✓46.218.6
+ +Table 2. Perception results on nuScenes [2] validation set. Results of vehicles and pedestrians are compared by segmentation IoU. Temp. indicates whether temporal information is involved. + +coding process. Unlike the encoding process, $\{\mathbf{Q}^{(t)}\}$ is used as the query input of Swin-T block and performs cross attention with the encoded features $\mathbf{E}_3$ . After the first decoding layer, the output of each layer is used as the query input of the next layer. Similar to encoding layers, the deconvolution layer is optionally applied to upsample the decoded features. The simplified process can be written as + +$$ +\mathcal {D} _ {s} = \left\{ \begin{array}{l l} \operatorname {S w i n T} (\mathcal {B} _ {3}, \{\mathbf {Q} ^ {(t)} \}), & s = 3 \\ \operatorname {S w i n T} (\mathcal {B} _ {s}, \operatorname {D e C o n v} (\mathbf {D} _ {s + 1})), & s = 0, 1, 2 \end{array} \right. +$$ + +where $\mathcal{D}_s\in \mathbb{R}^{(T' + 1)\times \frac{X}{2^s}\times \frac{Y}{2^s}\times C}, s = 0,1,2,3$ are decoded features. The future temporal BEV feature map at the 0th scale is a temporal sequence of final predicted BEV states; that is, $\mathcal{D}_0 = [\mathbf{B}_*^{(0)},\mathbf{B}_*^{(1)},\dots,\mathbf{B}_*^{(T')}]$ , where $\mathbf{B}_*^{(t)}\in \mathbb{R}^{X\times Y\times C}$ is the BEV state at future time stamp $t$ . + +# 3.4. Multi-head decoder + +The future temporal BEV feature map is fed into the multi-task decoder heads to generate various outputs for dynamic scene understanding, see Fig. 5. We follow the output + +setting in [16] that predicts BEV semantic segmentation, instance center, instance offset, and future flow for joint perception and prediction. Meanwhile, we set up an additional HD map decoder head to predict basic traffic scene elements including drivable areas and lanes. The map decoder head can not only provide scene information for subsequent planning and control modules but also give guidance to the prediction process, see sec. 3.3. + +# 4. Experiments + +# 4.1. Dataset and settings + +We use nuScenes [2] datasets to evaluate our approach. NuScenes contains 1000 scenes, each of which has 20 seconds annotated at $2\mathrm{Hz}$ . In nuScenes, the images are captured by 6 cameras with a small overlap in the field of view, which guarantees the cameras cover the full $360^{\circ}$ field of view. For model input, raw camera images with the size of $900 \times 1600$ are resized and cropped to a resolution of $224 \times 480$ . We follow the training and evaluating settings used in previous methods [1,16,17,56] for fair comparisons, which use 1.0 second past states and current state to predict 2.0 seconds of the future states. It corresponds to predicting 4 future frames based on 3 observed frames. The size of the generated BEV grid map is $200 \times 200$ . Each grid has a range of $0.5m \times 0.5m$ , which means the perception and prediction range is $100m \times 100m$ . + +For training, we use AdamW [30] with a weight decay 0.01 to optimize the models. The learning rate is initialized as $10^{-4}$ and decays with a cosine annealing scheduler [29]. All models are trained on 4 NVIDIA A100 GPUs for 10 epochs. + +# 4.2. Metrics + +Following previous works [1, 16, 17, 56], we mainly use two metrics for evaluation. The first is Intersection over Union (IoU), which measures the quality of segmentation at each frame. The second is Video Panoptic Quality (VPQ), which is used to measure the consistency of the detected instances over time and the accuracy of the segmentation. + +
Exp.Warp.Sync.SLQSPEFuture semantic seg.Future instance seg.
Short (IoU)Long (IoU)Short (VPQ)Long (VPQ)
1✓58.738.450.631.8
2✓✓60.838.652.433.4
3✓✓✓62.040.753.234.3
4✓63.040.854.134.3
5✓✓63.841.155.835.8
6✓✓✓64.741.956.736.9
+ +Table 3. Ablation of our proposed architecture. Ablation results for our PoseSync BEV Encoder (Sync.), the learnable future queries, and the spatial embedding are presented. Exp. 1-3 use the traditional warping methods to align temporal BEV features. Separate learnable queries (SLQ) represent using separate learnable future queries instead of utilizing the same query with temporal positional encoding. Spatial positional embedding (SPE) represents using spatial scene representations in future prediction queries. + +
Temporal modelIoUVPQVRQVSQ
MotionNet†[53]35.430.643.171.1
FIERY†[16]38.332.145.470.7
BEVerse†[56]40.234.048.070.9
TBP-Former41.936.951.572.6
+ +Table 4. Ablation for the prediction model. †: We use MotionNet, FIERY and BEVerse to replace our prediction model for comparison, and the BEV encoder and task heads are the same. Besides IoU and VPQ, we also use Video Recognition Quality (VRQ) and Video Segmentation quality (VSQ) for evaluation. + +
AugmentationPerceptionPrediction
CamBEVVeh.Ped.IoUVPQ
45.017.740.534.4
✓44.818.540.935.3
✓45.318.641.435.6
✓✓46.218.641.936.9
+ +Table 5. Ablation for data augmentation strategies. Perception of Vehicles and Pedestrians with different data augmentation strategies are evaluated on segmentation IoU. Prediction results are evaluated on segmentation IoU and Video Panoptic Quality. + +The formula is shown below: + +$$ +\mathrm {V P Q} = \sum_ {t = 0} ^ {H} \frac {\sum_ {\left(p _ {t} , q _ {t}\right) \in T P _ {t}} \operatorname {I o U} \left(p _ {t} , q _ {t}\right)}{\left| T P _ {t} \right| + \frac {1}{2} \left| F P _ {t} \right| + \frac {1}{2} \left| F N _ {t} \right|} +$$ + +where $H$ is the sequence length, $TP_{t}$ represents the set of true positives, $FP_{t}$ represents the set of false positives and $FN_{t}$ represents the set of false negatives at timestamp $t$ . + +# 4.3. PnP results + +Perception and Prediction. Table 1 compares TBP-Former with other methods of perception and prediction task based on multi-view cameras. We see that i) we achieve state-of-the-art performance and exceed previous methods by a large margin. ii) Even though BEVerse has larger RGB resolutions, TBP-Former still surpasses their performance on IoU by $7.3\% / 8.3\%$ for short/long settings, respectively. + +TBP-Former also improves the VPQ by $12.1\% / 10.8\%$ . iii) Apart from the performance improvement, TBF-Former also has a larger FPS compared to other methods. Its inference speed is $25\%$ faster than BEVerse's. + +Fig. 5 shows the visualization results of our proposed method. We see that i) almost all the objects are detected correctly except for those occluded ones. ii) TBP-Former is capable of capturing the motion information in past frames and precisely predicting the vehicles' trajectories by occupancy and flow. Compared with FIERY [16], TBP-Former is closer to the ground truth. iii) TBP-Former does a better job than FIERY when predicting vehicles' turning. + +Perception Only. Table 2 compares the results of plenty of state-of-the-art methods on perception (segmentation) task. We see that our static model, which does not contain temporal information, can achieve 44.8 and 17.2 IoU of vehicles and pedestrians. With the input of temporal sequences, the performance improves further since auxiliary information is provided for better perception. The state-of-the-art results prove the effectiveness of the novel design of our Pose-synchronized BEV encoder. + +# 4.4. Ablation + +Effectiveness of PoseSync View Projection. The Exp. 1&4, 2&5, 3&6 in Table 3 compare the proposed PoseSync View Projection and the existing feature warping methods. We see that the proposed method always achieves better performance when other settings remain the same. The reasons are that: i) PoseSync View Projection based on Deformable Attention can guarantee the precise correspondence between BEV grids and image features. ii) Our projection method can alleviate distortion and our-of-range issues when synchronizing sequential BEV features. + +Effectiveness of the designed future queries. In Exp. 1&3 in Table 3, we utilize the identical query with temporal positional encoding for future queries. Exp. 2&4 in Table 3 demonstrate that using separate learnable embedding for future queries can achieve better performance. Exp. 3&6 in Table 3 validate the efficacy of the proposed spatial priors for future queries. The generated high + +![](images/05ace495861ad71b1b2312e77f5dacfff1efdf7fd70b78a3a7c7c32bc1eb1dca.jpg) +Figure 5. Demonstration of our results compared with FIERY and Ground Truth. Different vehicles are assigned with different colors in order to make a distinction. The darker parts represent the perception of the current frame, and the lighter parts represent the prediction of the vehicles in future frames. The visualization is based on the predicted occupancy and flow. + +dimensional map features provide the prediction model with useful geographic information. The additional spatial information can aid the prediction and lead to better scene forecasting. + +Effectiveness of STPT. Table 4 compares STPT with popular CNN-based [53] and RNN-based [16, 56] methods. We keep all the settings the same except for temporal modeling. To be specific, the size of input images $(224 \times 480)$ , image backbones and BEV feature extractor are the same. And then we plug their temporal models into our architecture. We see that i) STPT model performs better in all four metrics, including semantic segmentation IoU and three instance segmentation metrics from the video prediction area. ii) Our reproduced temporal models achieve higher performance than the original implements. This further validates the effectiveness and power of our BEV feature extractor. + +Data Augmentation. We perform both image-view and BEV augmentations. The image-view augmentations include random scaling, rotation and flip of the input images. The BEV augmentations include similar operations on both BEV representations and corresponding ground truth labels. Table 5 compares the results of different data augmentation + +strategies. We see that i) both augmentation methods improve the performance when used separately. ii) The combination of two methods works better than any single approach. Introducing data augmentation strategies is beneficial to the model's robustness and generalization ability. + +# 5. Conclusion + +This paper proposes a novel TBP-Former for vision-centric joint perception and prediction. We design a pose-synchronized BEV encoder module using a cross-view attention mechanism to solve the distortion issues in previous works. Furthermore, we propose a powerful spatial-temporal pyramid transformer for BEV feature extraction and BEV state prediction. Experiments show that i) TBP-Former improves the prediction performance over state-of-the-art methods significantly; and ii) both PoseSync BEV Encoder and STPT contribute to better performances. + +Acknowledgement This research is partially supported by National Natural Science Foundation of China under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005. + +# References + +[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022. 2, 3, 6 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6 +[3] Sergio Casas, Cole Gulino, Renjie Liao, and Raquel Urtasun. Spagnn: Spatially-aware graph neural networks for relational behavior forecasting from sensor data. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9491-9497. IEEE, 2020. 2 +[4] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 2, 3 +[5] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 3, 4 +[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 3 +[7] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 3 +[8] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597–1600. IEEE, 2017. 3 +[9] Artem Filatov, Andrey Rykov, and Viacheslav Murashkin. Any motion detector: Learning class-agnostic scene dynamics from a sequence of lidar point clouds. In 2020 IEEE international conference on robotics and automation (ICRA), pages 9498–9504. IEEE, 2020. 3 +[10] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17441-17451, 2022. 4 +[11] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2021. 4 +[12] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 4 + +[13] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. A simple baseline for bev perception without lidar. arXiv e-prints, pages arXiv-2206, 2022. 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[15] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3 +[16] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 2, 3, 6, 7, 8 +[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2, 3, 6 +[18] Yue Hu, Siheng Chen, Ya Zhang, and Xiao Gu. Collaborative motion predication via neural motion message passing. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3 +[19] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), November 2022. 3 +[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3 +[21] Kuan-Hui Lee, Matthew Kliemann, Adrien Gaidon, Jie Li, Chao Fang, Sudeep Pillai, and Wolfram Burgard. Pillarflow: End-to-end birds-eye-view flow estimation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2007-2013. IEEE, 2020. 3 +[22] Peixuan Li and Jieyu Jin. Time3d: End-to-end joint monocular 3d object detection and tracking for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3885-3894, 2022. 4 +[23] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework, 2021. 3 +[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2, 3, 4, 6 +[25] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1, 2, 3 +[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi + +Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3 +[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 4 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 5 +[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[31] Chenyang Lu, Marinus Jacobus Gerardus van de Molengraft, and Gijs Dubbelman. Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters, 4(2):445-452, 2019. 6 +[32] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 4 +[33] Chenxu Luo, Xiaodong Yang, and Alan Yuille. Self-supervised pillar motion learning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3183-3192, 2021. 3 +[34] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2, 4 +[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified multi-task model for behavior prediction and planning. arXiv e-prints, pages arXiv-2106, 2021. 4 +[36] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 3, 6 +[37] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 3 +[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 2, 3, 6 +[39] John Phillips, Julieta Martinez, Ioan Andrei Bársan, Sergio Casas, Abbas Sadat, and Raquel Urtasun. Deep multi-task learning for joint localization, perception, and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4679-4689, 2021. 2, 3 + +[40] Zequn Qin, Jingyu Chen, Chao Chen, Xiaozhi Chen, and Xi Li. Uniformer: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. arXiv preprint arXiv:2207.08536, 2022. 3 +[41] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent video transformer. arXiv preprint arXiv:2006.10704, 2020. 4 +[42] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 3, 6 +[43] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In European Conference on Computer Vision, pages 414-430. Springer, 2020. 1, 3 +[44] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In 2022 International Conference on Robotics and Automation (ICRA), pages 9200-9206. IEEE, 2022. 6 +[45] Marcel Schreiber, Vasileios Belagiannis, Claudius Gläser, and Klaus Dietmayer. Dynamic occupancy grid mapping with recurrent neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6717-6724. IEEE, 2021. 3 +[46] Meet Shah, Zhiling Huang, Ankit Laddha, Matthew Langford, Blake Barber, Sidney Zhang, Carlos Vallespi-Gonzalez, and Raquel Urtasun. Liranet: End-to-end trajectory prediction using spatio-temporal radar fusion. arXiv preprint arXiv:2010.00731, 2020. 1 +[47] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 4 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[49] Hengli Wang, Peide Cai, Yuxiang Sun, Lujia Wang, and Ming Liu. Learning interpretable end-to-end vision-based motion planning for autonomous driving with optical flow distillation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13731-13737. IEEE, 2021. 2, 6 +[50] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17093–17102, 2022. 4 +[51] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 4 +[52] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying unknown instances for au + +tonomous driving. In Conference on Robot Learning, pages 384-393. PMLR, 2020. 2 +[53] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 2, 3, 4, 7, 8 +[54] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3 +[55] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4 +[56] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2, 3, 6, 7, 8 +[57] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 3, 6 +[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5 \ No newline at end of file diff --git a/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/images.zip b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0056dd1fc4ab056970af051024d6ed184b851d94 --- /dev/null +++ b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1202247afad2cd03d65b0ed79ee86ec6e07a8fe990fb4784901538f3e84af75e +size 526505 diff --git a/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/layout.json b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe49345e63db4da652aa15ec9ed7bdccebe91fc1 --- /dev/null +++ b/2023/TBP-Former_ Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving/layout.json @@ -0,0 +1,8428 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 517, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 517, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 517, + 140 + ], + "type": "text", + "content": "TBP-Former: Learning Temporal Bird's-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "text", + "content": "Shaoheng Fang" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "text", + "content": " Zi Wang" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "text", + "content": " Yiqi Zhong" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "text", + "content": " Junhao Ge" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "text", + "content": " Siheng Chen" + }, + { + "bbox": [ + 107, + 160, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 176, + 474, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 176, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 176, + 474, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 176, + 474, + 190 + ], + "type": "text", + "content": "Cooperative Medianet Innovation Center, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 190, + 537, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 190, + 537, + 204 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 537, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 190, + 537, + 204 + ], + "type": "text", + "content": "Department of Computer Science, University of Southern California " + }, + { + "bbox": [ + 67, + 190, + 537, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 190, + 537, + 204 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 204, + 486, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 204, + 486, + 217 + ], + "spans": [ + { + "bbox": [ + 107, + 204, + 486, + 217 + ], + "type": "text", + "content": "1{shfang，w4ngz1，cancaries，sihengc}@sjtu.edu.cn 2{yiqizhon}@usc.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 290, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 522 + ], + "type": "text", + "content": "Vision-centric joint perception and prediction (PnP) has become an emerging trend in autonomous driving research. It predicts the future states of the traffic participants in the surrounding environment from raw RGB images. However, it is still a critical challenge to synchronize features obtained at multiple camera views and timestamps due to inevitable geometric distortions and further exploit those spatial-temporal features. To address this issue, we propose a temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP, which includes two novel designs. First, a pose-synchronized BEV encoder is proposed to map raw image inputs with any camera pose at any time to a shared and synchronized BEV space for better spatial-temporal synchronization. Second, a spatial-temporal pyramid transformer is introduced to comprehensively extract multi-scale BEV features and predict future BEV states with the support of spatial priors. Extensive experiments on nuScenes dataset show that our proposed framework overall outperforms all state-of-the-art vision-based prediction methods. Code is available at: https://github.com/MediaBrain-SJTU/TBP-Former" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 574, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 682 + ], + "type": "text", + "content": "As one of the most fascinating engineering projects, autonomous driving has been an aspiration for many researchers and engineers for decades. Although significant progress has been made, it is still an open question in designing a practical solution to achieve the goal of full self-driving. A traditional and common solution consists of a sequential stack of perception, prediction, planning, and control. Despite the idea of divide-and-conquer having achieved tremendous success in developing software sys" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 255, + 475, + 354 + ], + "blocks": [ + { + "bbox": [ + 365, + 245, + 421, + 254 + ], + "lines": [ + { + "bbox": [ + 365, + 245, + 421, + 254 + ], + "spans": [ + { + "bbox": [ + 365, + 245, + 421, + 254 + ], + "type": "text", + "content": "Previous works" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 255, + 475, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 475, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 475, + 354 + ], + "type": "image", + "image_path": "e14ef5292e54d78c4693f68ae053ebc277cef5b277c65fceee5ce62d613c6f82.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 481, + 255, + 542, + 354 + ], + "blocks": [ + { + "bbox": [ + 490, + 245, + 534, + 254 + ], + "lines": [ + { + "bbox": [ + 490, + 245, + 534, + 254 + ], + "spans": [ + { + "bbox": [ + 490, + 245, + 534, + 254 + ], + "type": "text", + "content": "TBP-Former" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 255, + 542, + 354 + ], + "lines": [ + { + "bbox": [ + 481, + 255, + 542, + 354 + ], + "spans": [ + { + "bbox": [ + 481, + 255, + 542, + 354 + ], + "type": "image", + "image_path": "143dbc6f80836fda5ead3f6fb745c36c14d7a4a92e17a75a79885c92e1f1e081.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 313, + 366, + 475, + 466 + ], + "blocks": [ + { + "bbox": [ + 380, + 356, + 492, + 365 + ], + "lines": [ + { + "bbox": [ + 380, + 356, + 492, + 365 + ], + "spans": [ + { + "bbox": [ + 380, + 356, + 492, + 365 + ], + "type": "text", + "content": "(a) Temporal BEV feature alignment" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 366, + 475, + 466 + ], + "lines": [ + { + "bbox": [ + 313, + 366, + 475, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 475, + 466 + ], + "type": "image", + "image_path": "8d43e3d772a538bacefdb6cfc5eacc582b42d4a95beb85b679b115888f99fc7e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 487, + 547, + 564 + ], + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 564 + ], + "type": "text", + "content": "Figure 1. Two major challenges in vision-based perception and prediction are (a) how to avoid distortion and deficiency when aggregating features across time and camera views; and (b) how to achieve spatial-temporal feature learning for prediction. Our Pose-Synchronized BEV Encoder can precisely map the visual features into synchronized BEV space, and Spatial-Temporal Pyramid Transformer extracts feature at multiple scales." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 481, + 366, + 544, + 466 + ], + "blocks": [ + { + "bbox": [ + 481, + 366, + 544, + 466 + ], + "lines": [ + { + "bbox": [ + 481, + 366, + 544, + 466 + ], + "spans": [ + { + "bbox": [ + 481, + 366, + 544, + 466 + ], + "type": "image", + "image_path": "002bf9e04e7761f1d780ff6b49709ad83265ad64d6e1d315907695cf48154154.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 469, + 517, + 479 + ], + "lines": [ + { + "bbox": [ + 353, + 469, + 517, + 479 + ], + "spans": [ + { + "bbox": [ + 353, + 469, + 517, + 479 + ], + "type": "text", + "content": "(b) Spatial-temporal feature modeling for prediction" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 590, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 674 + ], + "type": "text", + "content": "tems, a long stack could cause cascading failures in an autonomous system. Recently, there is a trend to combine multiple parts in an autonomous system to be a joint module, cutting down the stack. For example, [25, 46] consider joint perception and prediction and [5, 43] explore joint prediction and planning. This work focuses on joint perception and prediction." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "The task of joint perception and prediction (PnP) aims to predict the current and future states of the surrounding environment with the input of multi-frame raw sensor" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 215, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 215, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 215, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1368" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "data. The output current and future states would directly serve as the input for motion planning. Recently, many PnP methods are proposed based on diverse sensor input choices. For example, [4, 25, 34] take multi-frame LiDAR point clouds as input and achieve encouraging 3D detection and trajectory prediction performances simultaneously. Recently, the rapid development of vision-centric methods offers a new possibility to provide a cheaper and easy-to-deploy solution for PnP. For instance, [1, 16, 17] only uses RGB images collected by multiple cameras to build PnP systems. Meanwhile, without precise 3D measurements, vision-centric PnP is more technically challenging. Therefore, this work aims to advance this direction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 397 + ], + "type": "text", + "content": "The core of vision-centric PnP is to learn appropriate spatial-temporal feature representations from temporal image sequences. It is a crux and difficult from three aspects. First, since the input and the output of vision-centric PnP are supported in camera front-view (FV) and bird's-eye-view (BEV) respectively, one has to deal with distortion issues during geometric transformation between two views. Second, when the vehicle is moving, the view of the image input is time-varying and it is thus nontrivial to precisely map visual features across time into a shared and synchronized space. Third, since information in temporal image sequences is sufficiently rich for humans to accurately perceive the environment, we need a powerful learning model to comprehensively exploit spatial-temporal features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 604 + ], + "type": "text", + "content": "To tackle these issues, previous works on vision-centric PnP consider diverse strategies. For example, [16, 56] follows the method in [38] to map FV features to BEV features, then synchronizes BEV features across time via rigid transformation, and finally uses a recurrent network to exploit spatial-temporal features. However, due to the image discretization nature and depth estimation uncertainty, simply relying on rigid geometric transformations would cause inevitable distortion; see Fig. 1. Some other work [49] transforms the pseudo feature point cloud to current ego coordinates and then pools the pseudo-lidar to BEV features; however, this approach encounters deficiency due to the limited sensing range in perception. Meanwhile, many works [16, 17, 56] simply employ recurrent neural networks to learn the temporal features from multiple BEV representations, which is hard to comprehensively extract spatial-temporal features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "To promote more reliable and comprehensive feature learning across views and time, we propose the temporal bird's-eye-view pyramid transformer (TBP-Former) for vision-centric PnP. The proposed TBP-Former includes two key innovations: i) pose-synchronized BEV encoder, which leverages a pose-aware cross-attention mechanism to directly map a raw image input with any camera pose at any time to the corresponding feature map in a shared and synchronized BEV space; and ii) spatial-temporal pyra" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "mid transformer, which leverages a pyramid architecture with Swin-transformer [28] blocks to learn comprehensive spatial-temporal features from sequential BEV maps at multiple scales and predict future BEV states with a set of future queries equipped with spatial priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 546, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 287 + ], + "type": "text", + "content": "Compared to previous works, the proposed TBP-Former brings benefits from two aspects. First, previous works [16, 17, 24, 56] consider FV-to-BEV transformation and temporal synchronization as two separate steps, each of which could bring distortion due to discrete depth estimation and rigid transformation; while we merge them into one step and leverage both geometric transformation and attention-based learning ability to achieve spatial-temporal synchronization. Second, previous works [16, 53] use RNNs or 3D convolutions to learn spatial-temporal features; while we leverage a powerful pyramid transformer architecture to comprehensively capture spatial-temporal features, which makes prediction more effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 289, + 539, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 289, + 539, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 289, + 539, + 300 + ], + "type": "text", + "content": "To summarize, the main contributions of our work are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 308, + 546, + 564 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 308, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 308, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 308, + 545, + 369 + ], + "type": "text", + "content": "- To tackle the distortion issues in mapping temporal image sequences to a synchronized BEV space, we propose a pose-synchronized BEV encoder (PoseSync BEV Encoder) based on cross-view attention mechanism to extract quality temporal BEV features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 376, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 545, + 437 + ], + "type": "text", + "content": "- We propose a novel Spatial-Temporal Pyramid Transformer (STPT) to extract multi-scale spatial-temporal features from sequential BEV maps and predict future BEV states according to well-elaborated future queries integrated with spatial priors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 445, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 546, + 564 + ], + "type": "text", + "content": "- Overall, we propose TBP-Former, a vision-based joint perception and prediction framework for autonomous driving. TBP-Former achieves state-of-the-art performance on nuScenes [2] dataset for the vision-based prediction task. Extensive experiments show that both PoseSync BEV Encoder and STPT contribute greatly to the performance. Due to the decoupling property of the framework, both proposed modules can be easily utilized as alternative modules in any vision-based BEV prediction framework." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 579, + 392, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 579, + 392, + 591 + ], + "spans": [ + { + "bbox": [ + 306, + 579, + 392, + 591 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 599, + 477, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 477, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 477, + 612 + ], + "type": "text", + "content": "2.1. Joint Perception and Prediction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "content": "As the two core system modules of autonomous driving, how to conduct perception and prediction tasks jointly has received a lot of attention. Traditional approaches [3,4,25,34,39] formulate this joint task as a trajectory prediction problem that relies on the perception outputs of 3D object detection and tracking. The dependency on intermediate results tends to accumulate errors and lacks the capacity to perceive unknown objects [52,53]. Subsequently," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1369" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 539, + 254 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 539, + 254 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 539, + 254 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 539, + 254 + ], + "type": "image", + "image_path": "ecf270f42ebad10f4d13bcd27c06dd85e580b8c2d4aa88f99dd595d12d02bc97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 548, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 548, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 548, + 331 + ], + "type": "text", + "content": "Figure 2. An overview of TBP-Former architecture. Taking consecutive surrounding camera images as inputs, TBP-Former first generates image-space features and uses the PoseSync BEV Encoder to map front-view features to BEV features in a shared and synchronized BEV space. Then the BEV features from multiple frames are processed by the Spatial-Temporal Pyramid Transformer to extract BEV spatial-temporal features and predict future BEV states in order. In this process, high-level scene representations are generated from the last frame BEV feature as spatial priors to guide the prediction. Finally, the well-predicted future states are sent to decoder heads for joint perception and prediction tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 340, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 411 + ], + "type": "text", + "content": "instance-free methods [9,21,33,43,45,53] that predict dense future semantic occupancy and flow have become a growing trend to simplify the understanding of dynamic scenes. Also, several recent works [16, 17, 56] explore joint perception and prediction in the form of dense occupancy and flow using only surrounding camera input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 288, + 533 + ], + "type": "text", + "content": "In many previous works [4, 25, 39, 43], raster HD (high-definition) maps play an important role as input of the frameworks. HD maps can provide strong priors to guide the predicted results to follow the traffic lanes. However, in practice, HD maps are laborious and costly to produce and require frequent maintenance. Instead of using off-the-peg HD maps, we follow the philosophy of [5, 7, 23] in predicting online HD maps but propose to learn high-level scene geometry representations from real-time sensor inputs and take these representations as priors for the prediction task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 540, + 170, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 170, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 170, + 552 + ], + "type": "text", + "content": "2.2. BEV Representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "BEV representations provide a unified and physical-interpretable way to represent the rich information of road, moving objects and occlusion in a traffic scene, which can be easily utilized for downstream tasks such as motion prediction, planning and control, etc. For camera-based methods, how to solve the problem of projecting features from perspective view to BEV is a major challenge. Some learnable methods use MLP [23, 36, 42] or transformer network [37, 57] to implicitly reason the relationship between two different views. LSS [38] proposes the approach of predicting depth distribution per pixel on 2D features, then 'lifting' the 2D features according to the corresponding depth distribution to BEV space. Numerous works, aim" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 340, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 422 + ], + "type": "text", + "content": "ing at tasks of BEV perception [19, 20, 54], motion prediction [1, 16, 17, 56], lidar-camera fusion [26], etc., follow this form to generate BEV representations. Also, some methods [6, 13, 24] explicitly establish the correspondence from BEV location to image-view pixel using homography between image and BEV plane and achieve attractive performance in diverse tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 425, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 581 + ], + "type": "text", + "content": "However, when dealing with temporal information, most methods [1, 16, 17, 24, 56] warp history BEV representations according to the variation of ego poses. Due to the pre-defined fixed range and size of the BEV grid, rotation and translation operations may cause distortion and out-of-range problems when aligning history BEV maps to current ego coordinates. Though [40] introduces a similar operation to us to integrate historical information into the current frame, the design of their model is unable to predict future states. To alleviate these issues, we propose a PoseSync BEV Encoder module based on deformable attention to generate pose-synchronized BEV representations from temporally consecutive image-view input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 587, + 459, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 459, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 459, + 601 + ], + "type": "text", + "content": "2.3. Spatial-Temporal Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "In the BEV prediction field, how to design a temporal model to aggregate spatial-temporal information is a critical problem. Existing modeling methods can be classified into three categories: RNN-based, CNN-based and transformer-based. RNN-based methods [1, 16-18, 43, 56] utilize recurrent models such as LSTM [15], GRU [8] to predict the future latent states. Though the recurrent model is powerful to model temporal relationships, it is time-consuming for constraints in the parallelization of computation. Besides," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1370" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 81, + 303, + 218 + ], + "blocks": [ + { + "bbox": [ + 51, + 79, + 64, + 89 + ], + "lines": [ + { + "bbox": [ + 51, + 79, + 64, + 89 + ], + "spans": [ + { + "bbox": [ + 51, + 79, + 64, + 89 + ], + "type": "text", + "content": "(A)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 81, + 303, + 218 + ], + "lines": [ + { + "bbox": [ + 82, + 81, + 303, + 218 + ], + "spans": [ + { + "bbox": [ + 82, + 81, + 303, + 218 + ], + "type": "image", + "image_path": "2c7b4199d8890da50565946c771cae33c5f05eb0a2720c25da883f3f9515e474.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 79, + 331, + 90 + ], + "lines": [ + { + "bbox": [ + 317, + 79, + 331, + 90 + ], + "spans": [ + { + "bbox": [ + 317, + 79, + 331, + 90 + ], + "type": "text", + "content": "(B)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 345, + 81, + 545, + 211 + ], + "blocks": [ + { + "bbox": [ + 345, + 81, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 345, + 81, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 345, + 81, + 545, + 211 + ], + "type": "image", + "image_path": "999941865914e9fd53b96a15718d940b3b6514b1b6f1122f5345fd09a902cd3f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 546, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 546, + 264 + ], + "type": "text", + "content": "Figure 3. The PoseSync BEV Encoder (A) takes front-view features and camera poses as input and then maps to BEV space. The core to generate BEV features in a synchronized way is the Pose-Aware Cross Attention. Its cross-view attention mechanism is depicted in (B), where front-view features from different frames of a dynamic vehicle are projected into a uniform BEV space." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 273, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 309 + ], + "type": "text", + "content": "some CNN-based methods [5, 34, 50, 53] concatenate BEV features in the time dimension and take advantage of 3D convolution to extract spatial-temporal features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 311, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 288, + 479 + ], + "type": "text", + "content": "Due to the great power of transformer [48] in sequence modeling, it has shown promise in many temporal modeling tasks such as trajectory prediction [11, 35, 55], object tracking [22], video prediction [12, 41, 51], video interpolation [10, 32, 47], etc. For BEV perception, [24, 27] utilize self-attention to model temporal information from multiple frames to boost perception task. [22] leverage self-attention to aggregate spatial information and cross-attention to exploit affinities among sequence frames. To explore the capacity of transformer model in BEV spatial-temporal modeling, we propose a novel Spatial-Temporal Pyramid Transformer (STPT) architecture with future queries for BEV spatial-temporal features extraction and BEV future states prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 129, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 129, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 129, + 505 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 513, + 176, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 176, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 176, + 525 + ], + "type": "text", + "content": "3.1. Overview Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "The overall architecture of the proposed TBP-Former is illustrated in Fig. 2. It takes the input of multi-view images with the corresponding camera poses at consecutive " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " timestamps. The final output includes BEV map segmentation for current scene understanding and occupancy flow for motion prediction. The whole TBP-Former can be decoupled into three parts: (i) pose-synchronized BEV encoder, which maps raw image sequences into feature maps in a spatial-temporal-synchronized BEV space; (ii) spatial-temporal pyramid transformer, which achieves comprehensive feature learning at multiple spatial and temporal scales; and (iii) a multi-head decoder, which takes the spatial and temporal features to achieve scene understanding and motion prediction. We will elaborate on each part in the following subsections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 272, + 485, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 485, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 485, + 285 + ], + "type": "text", + "content": "3.2. Pose-Synchronized BEV Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 292, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 546, + 422 + ], + "type": "text", + "content": "Given images collected at multiple time stamps and from various camera poses, we aim to generate the corresponding feature maps in a shared and synchronized BEV space. Different from many previous works that synchronize spatial and temporal information in two separate steps, the proposed pose-synchronized BEV encoder leverage both geometric prior and learning ability to achieve one-step synchronization, alleviating distortion effects. Following the previous transformer-based method [24], this encoder adopts a transformer architecture whose core is a novel cross-view attention operation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": "Front-view feature map. Let " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{X}_i^{(-t)}\\}_{i=1,t=0}^{N,T}" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " be the input multi-frame multi-view images, where " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " is the number of cameras, " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " is the number of historical timestamps and " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^{(-t)} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " is the RBG image captured by the " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": "th camera at historical time stamp " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": ". Note that each front-view image is associated with a different camera pose. Let " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "S_i^{(-t)} = \\{(u_i^{(-t)}, v_i^{(-t)})\\}_{1,1}^{H,W}" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " be the pixel indices of the " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": "th camera's front-view space, whose image size is " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": ". We feed each RBG image " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^{(-t)}" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " into a shared backbone network (our implementation uses ResNet-101 [14]) and obtain the corresponding front-view feature map " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{(-t)} \\in \\mathbb{R}^{H' \\times W' \\times C}" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": " the channel number, which is also supported on the front-view space " + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "inline_equation", + "content": "S_i^{(-t)}" + }, + { + "bbox": [ + 304, + 423, + 546, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": "BEV queries. Let " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{BEV}} = \\{(x,y)\\}_{x = 1,y = 1}^{X,Y}" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " be the BEV grid indices, reflecting the " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "X\\times Y" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " BEV grid space based on the vehicle-ego pose at the current timestamp. Note that " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{BEV}}" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " is the only BEV space we work with in this paper. Let " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\in \\mathbb{R}^{X\\times Y\\times C}" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " be the trainable BEV queries whose element " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{x,y}\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " -dimensional query feature at the " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " th geo-location in the BEV space " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{BEV}}" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " .We use " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " as the input to query from the front-view feature map " + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{(-t)}" + }, + { + "bbox": [ + 304, + 590, + 545, + 700 + ], + "type": "text", + "content": " to produce the corresponding BEV feature map." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Cross-view attention. As the key operation in the pose" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1371" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "synchronized BEV encoder, the proposed cross-view attention constructs a feature map in the BEV space " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{BEV}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " by absorbing information from the corresponding pixels in the front-view feature map." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{(-t)}:S_{\\mathrm{BEV}}\\times \\mathcal{Z}\\to S_i^{(-t)}" + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "text", + "content": " be a project operation that maps a BEV index with a specific height index to a pixel index in the " + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "text", + "content": "th camera's front view at historical timestamp " + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 121, + 287, + 171 + ], + "type": "text", + "content": "; that is," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 95, + 182, + 238, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 182, + 238, + 202 + ], + "spans": [ + { + "bbox": [ + 95, + 182, + 238, + 202 + ], + "type": "interline_equation", + "content": "\\left(u _ {i} ^ {(- t)}, v _ {i} ^ {(- t)}\\right) = \\mathcal {P} _ {i} ^ {(- t)} \\Big ((x, y, z) \\Big),", + "image_path": "e4a0dfb9d15cd3ea8a5a9c3c56cc0f9d42fab1deba3d50605c571090de050645.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z} = \\{1, \\dots, Z\\}" + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "text", + "content": ". The project operation " + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{(-t)}" + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "text", + "content": " builds the geometric relationship between the BEV and a front view. The implementation of " + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{(-t)}" + }, + { + "bbox": [ + 46, + 214, + 287, + 255 + ], + "type": "text", + "content": " works as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 265, + 249, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 265, + 249, + 317 + ], + "spans": [ + { + "bbox": [ + 85, + 265, + 249, + 317 + ], + "type": "interline_equation", + "content": "z _ {i} ^ {(- t)} \\cdot \\left[ \\begin{array}{c} u _ {i} ^ {(- t)} \\\\ v _ {i} ^ {(- t)} \\\\ 1 \\end{array} \\right] = \\mathcal {T} _ {\\mathcal {S} _ {\\mathrm {B E V}} \\to \\mathcal {S} _ {i} ^ {(- t)}} \\cdot \\left[ \\begin{array}{c} x \\\\ y \\\\ z \\\\ 1 \\end{array} \\right],", + "image_path": "9315dc92e91eeef25d80f77a9e7f234ea9c3e79aa3bb6790470a18dc4837ed6f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 328, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 328, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{S}_{\\mathrm{BEV}}\\to \\mathcal{S}_i^{(-t)}}\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 46, + 328, + 287, + 366 + ], + "type": "text", + "content": " is a transformation matrix that can be calculated by camera's intrinsic/ extrinsic parameters and vehicle-ego pose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "text", + "content": "Based on the project operation, the cross-view attention can trace visual features in the front view through a BEV index. Let " + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i^{(-t)}\\in \\mathbb{R}^{X\\times Y\\times C}" + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "text", + "content": " be the BEV feature map associated with the RBG image " + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^{(-t)}" + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 368, + 287, + 432 + ], + "type": "text", + "content": "th element of the BEV feature map is obtained as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 443, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {B} _ {i} ^ {(- t)}\\right) _ {x, y} = \\sum_ {z} f _ {\\mathrm {D A}} \\left(\\mathbf {Q} _ {x, y}, \\mathcal {P} _ {i} ^ {(- t)} (x, y, z), \\mathbf {F} _ {i} ^ {(- t)}\\right), \\tag {1}", + "image_path": "29fcb3c53093a612bb342ad59b5bc45c801c4ccabf8b203541a63d8c3167fbce.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{DA}}(\\cdot)" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": " represents the deformable attention operation [58]. It allows BEV query " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "Q_{x,y}" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": " only to interact with the front-view feature " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{(-t)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": " within its regions of interest, which is sampled around the reference point calculated by " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{(-t)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": ". Since one BEV index might lead to multiple pixel indices in the front-view image because of various height possibilities in the 3D space. We thus sum over all possible heights in (1). To further aggregate BEV feature maps across all the " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": " camera views, we simply take the average; that is, the BEV feature map at historical timestamp " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{(-t)} = \\frac{1}{N}\\sum_{i}\\mathbf{B}_{i}^{(-t)}" + }, + { + "bbox": [ + 46, + 479, + 287, + 652 + ], + "type": "text", + "content": ". Note that all the front-view features across time and from multiple cameras are synchronized into the same BEV space in one step (1), leading to less information distortion or deficiency issues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "We can successively apply the cross-view attention followed by feed forward networks and normalization layers for multiple times. Finally, we order BEV feature maps at multiple timestamps and obtain a temporal BEV feature map " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = [\\mathbf{B}^{(0)},\\mathbf{B}^{(-1)},\\dots,\\mathbf{B}^{(-T)}]\\in \\mathbb{R}^{(T + 1)\\times X\\times Y\\times C}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 69, + 531, + 251 + ], + "blocks": [ + { + "bbox": [ + 322, + 69, + 531, + 251 + ], + "lines": [ + { + "bbox": [ + 322, + 69, + 531, + 251 + ], + "spans": [ + { + "bbox": [ + 322, + 69, + 531, + 251 + ], + "type": "image", + "image_path": "98c466355ad7037ecd88acf0f26e1b8e44eebe2b1f9dde4fd776cd2202957f94.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 258, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 335 + ], + "type": "text", + "content": "Figure 4. The network architecture of Spatial-Temporal Pyramid Transformer (STPT). Each encoder layer consists of an optional convolutional block for downsampling and Swin Transformer Blocks, while each decoder layer contains Swin Transformer Blocks and a deconvolutional block for upsampling. In the decoding process, we pre-define a set of future queries to represent future BEV states and query the features from encoders." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 354, + 515, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 515, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 515, + 367 + ], + "type": "text", + "content": "3.3. Spatial-Temporal Pyramid Transformer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 422 + ], + "type": "text", + "content": "We further propose a novel spatial-temporal pyramid transformer (STPT) to learn spatial-temporal features more comprehensively and produce the future BEV states. The detailed structure of STPT is depicted in Fig. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "text", + "content": "Temporal BEV pyramid feature learning. We encode the input temporal BEV feature map " + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "text", + "content": " with four hierarchical layers. Each encoder layer is composed of an optional convolution layer with stride 2 to downsample the features and a swim transformer encoder, which is a stack of Swin Transformer blocks [28]. In our implementation, the window size of Swin-T blocks is set as (4,4). We can then obtain multi-scale spatial-temporal features " + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_s\\in \\mathbb{R}^{(T + 1)\\times \\frac{X}{2^s}\\times \\frac{Y}{2^s}\\times C}" + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "inline_equation", + "content": "s = 0,1,2,3" + }, + { + "bbox": [ + 305, + 423, + 545, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": "Future BEV queries. Future BEV queries are defined to represent future BEV states and query the generated multiscale spatial-temporal features. There is a set of learnable future queries " + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}^{(t)}\\}" + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "inline_equation", + "content": "t = 0, \\dots, T'" + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{(t)}" + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": " has the same spatial dimension " + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\frac{X}{8} \\times \\frac{Y}{8}" + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_3" + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": ". Separate learning embeddings are employed for future queries to differentiate the predicted BEV states over time. Additionally, a map feature generator is applied to generate high-dimensional features from " + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{(T)}" + }, + { + "bbox": [ + 305, + 532, + 545, + 688 + ], + "type": "text", + "content": " in order to extract information about the scene's geometry. To be specific, the same structure and parameters of the hdmap decoder head (excluding the last linear layer) are reused. The resulting map feature is added to all future queries to provide spatial information priors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Future BEV state prediction. The decoding process contains corresponding four hierarchical layers as the en" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1372" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 504, + 168 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 504, + 168 + ], + "type": "table", + "html": "
MethodRGB ResolutionFuture semantic seg.Future instance seg.FPS
IoU (Short)IoU (Long)VPQ (Short)VPQ (Long)
FIERY [16]224×48059.436.750.229.91.56
StretchBEV [1]224×48055.537.146.029.01.56
ST-P3 [17]224×480-38.9-32.11.43
BEVerse [56]256×70460.338.752.233.31.96
TBP-Former224×48064.741.956.736.92.44
", + "image_path": "feb83b50d7aecfe810fcef3983927ca84336e2bd9a0a61808899555b459d3a1a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 232, + 279, + 402 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "type": "text", + "content": "Table 1. Prediction results on nuScenes [2] validation set. Intersection-over-Union (IoU) is used for future semantic segmentation and Video Panoptic Quality (VPQ) for future instance segmentation. Results are reported under two settings: short " + }, + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "type": "inline_equation", + "content": "(30m\\times 30m)" + }, + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "type": "text", + "content": " range and long " + }, + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "type": "inline_equation", + "content": "(100m\\times 100m)" + }, + { + "bbox": [ + 46, + 178, + 547, + 222 + ], + "type": "text", + "content": " range. Frame Per Second (FPS) means the inverse of inference time. All methods are tested under the same settings on a single NVIDIA A100. Our TBP-Former achieves SOTA performance and is still more computationally efficient than other methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 232, + 279, + 402 + ], + "lines": [ + { + "bbox": [ + 56, + 232, + 279, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 279, + 402 + ], + "type": "table", + "html": "
MethodTemp.Veh. IoUPed. IoU
VED [31]23.311.9
VPN [36]28.210.3
PON [42]27.913.9
LSS [38]34.615.0
CVT [57]36.0-
Image2Map [44]40.2-
BEVFormer [24]44.4-
IVMP [49]✓36.817.4
FIERY [16]✓38.217.2
ST-P3 [17]✓40.114.5
TBP-Former static44.817.2
TBP-Former✓46.218.6
", + "image_path": "2eec6cae78fea7371b0bacc3ac6dd1b2af0284cace4232bb0b83b62f43708a67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 445 + ], + "type": "text", + "content": "Table 2. Perception results on nuScenes [2] validation set. Results of vehicles and pedestrians are compared by segmentation IoU. Temp. indicates whether temporal information is involved." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "content": "coding process. Unlike the encoding process, " + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}^{(t)}\\}" + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "content": " is used as the query input of Swin-T block and performs cross attention with the encoded features " + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_3" + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "content": ". After the first decoding layer, the output of each layer is used as the query input of the next layer. Similar to encoding layers, the deconvolution layer is optionally applied to upsample the decoded features. The simplified process can be written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 544, + 271, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 271, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 271, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {s} = \\left\\{ \\begin{array}{l l} \\operatorname {S w i n T} (\\mathcal {B} _ {3}, \\{\\mathbf {Q} ^ {(t)} \\}), & s = 3 \\\\ \\operatorname {S w i n T} (\\mathcal {B} _ {s}, \\operatorname {D e C o n v} (\\mathbf {D} _ {s + 1})), & s = 0, 1, 2 \\end{array} \\right.", + "image_path": "0175df839c72ca35be11d329a4f1703a0d92bbda33e47a7fdb886913733eac10.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s\\in \\mathbb{R}^{(T' + 1)\\times \\frac{X}{2^s}\\times \\frac{Y}{2^s}\\times C}, s = 0,1,2,3" + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "text", + "content": " are decoded features. The future temporal BEV feature map at the 0th scale is a temporal sequence of final predicted BEV states; that is, " + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_0 = [\\mathbf{B}_*^{(0)},\\mathbf{B}_*^{(1)},\\dots,\\mathbf{B}_*^{(T')}]" + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_*^{(t)}\\in \\mathbb{R}^{X\\times Y\\times C}" + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "text", + "content": " is the BEV state at future time stamp " + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 586, + 286, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 162, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 162, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 162, + 670 + ], + "type": "text", + "content": "3.4. Multi-head decoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "The future temporal BEV feature map is fed into the multi-task decoder heads to generate various outputs for dynamic scene understanding, see Fig. 5. We follow the output" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 235, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 330 + ], + "type": "text", + "content": "setting in [16] that predicts BEV semantic segmentation, instance center, instance offset, and future flow for joint perception and prediction. Meanwhile, we set up an additional HD map decoder head to predict basic traffic scene elements including drivable areas and lanes. The map decoder head can not only provide scene information for subsequent planning and control modules but also give guidance to the prediction process, see sec. 3.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 335, + 388, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 388, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 388, + 347 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 354, + 422, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 422, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 422, + 367 + ], + "type": "text", + "content": "4.1. Dataset and settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": "We use nuScenes [2] datasets to evaluate our approach. NuScenes contains 1000 scenes, each of which has 20 seconds annotated at " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": ". In nuScenes, the images are captured by 6 cameras with a small overlap in the field of view, which guarantees the cameras cover the full " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": " field of view. For model input, raw camera images with the size of " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "900 \\times 1600" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": " are resized and cropped to a resolution of " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "224 \\times 480" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": ". We follow the training and evaluating settings used in previous methods [1,16,17,56] for fair comparisons, which use 1.0 second past states and current state to predict 2.0 seconds of the future states. It corresponds to predicting 4 future frames based on 3 observed frames. The size of the generated BEV grid map is " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": ". Each grid has a range of " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "0.5m \\times 0.5m" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": ", which means the perception and prediction range is " + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "inline_equation", + "content": "100m \\times 100m" + }, + { + "bbox": [ + 304, + 373, + 545, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "type": "text", + "content": "For training, we use AdamW [30] with a weight decay 0.01 to optimize the models. The learning rate is initialized as " + }, + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 554, + 545, + 613 + ], + "type": "text", + "content": " and decays with a cosine annealing scheduler [29]. All models are trained on 4 NVIDIA A100 GPUs for 10 epochs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 623, + 365, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 365, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 365, + 634 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Following previous works [1, 16, 17, 56], we mainly use two metrics for evaluation. The first is Intersection over Union (IoU), which measures the quality of segmentation at each frame. The second is Video Panoptic Quality (VPQ), which is used to measure the consistency of the detected instances over time and the accuracy of the segmentation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1373" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 500, + 178 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 500, + 178 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 500, + 178 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 500, + 178 + ], + "type": "table", + "html": "
Exp.Warp.Sync.SLQSPEFuture semantic seg.Future instance seg.
Short (IoU)Long (IoU)Short (VPQ)Long (VPQ)
1✓58.738.450.631.8
2✓✓60.838.652.433.4
3✓✓✓62.040.753.234.3
4✓63.040.854.134.3
5✓✓63.841.155.835.8
6✓✓✓64.741.956.736.9
", + "image_path": "92b53a1848df79cb682b19a90bad019779d318759c80ae325879ec7913eb8018.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 239, + 272, + 310 + ], + "blocks": [ + { + "bbox": [ + 46, + 187, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 547, + 232 + ], + "type": "text", + "content": "Table 3. Ablation of our proposed architecture. Ablation results for our PoseSync BEV Encoder (Sync.), the learnable future queries, and the spatial embedding are presented. Exp. 1-3 use the traditional warping methods to align temporal BEV features. Separate learnable queries (SLQ) represent using separate learnable future queries instead of utilizing the same query with temporal positional encoding. Spatial positional embedding (SPE) represents using spatial scene representations in future prediction queries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 239, + 272, + 310 + ], + "lines": [ + { + "bbox": [ + 63, + 239, + 272, + 310 + ], + "spans": [ + { + "bbox": [ + 63, + 239, + 272, + 310 + ], + "type": "table", + "html": "
Temporal modelIoUVPQVRQVSQ
MotionNet†[53]35.430.643.171.1
FIERY†[16]38.332.145.470.7
BEVerse†[56]40.234.048.070.9
TBP-Former41.936.951.572.6
", + "image_path": "086a6438eed1fa23a41a8454d9dfc1d037ca89c79985f33fabb0cb946824031f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 71, + 381, + 264, + 464 + ], + "blocks": [ + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "type": "text", + "content": "Table 4. Ablation for the prediction model. †: We use MotionNet, FIERY and BEVerse to replace our prediction model for comparison, and the BEV encoder and task heads are the same. Besides IoU and VPQ, we also use Video Recognition Quality (VRQ) and Video Segmentation quality (VSQ) for evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 381, + 264, + 464 + ], + "lines": [ + { + "bbox": [ + 71, + 381, + 264, + 464 + ], + "spans": [ + { + "bbox": [ + 71, + 381, + 264, + 464 + ], + "type": "table", + "html": "
AugmentationPerceptionPrediction
CamBEVVeh.Ped.IoUVPQ
45.017.740.534.4
✓44.818.540.935.3
✓45.318.641.435.6
✓✓46.218.641.936.9
", + "image_path": "dfd61a9af9512a84aac3f9e2849a25e9a8fd31db48a14e2ffdb69cc2413e2a4f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 517 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 517 + ], + "type": "text", + "content": "Table 5. Ablation for data augmentation strategies. Perception of Vehicles and Pedestrians with different data augmentation strategies are evaluated on segmentation IoU. Prediction results are evaluated on segmentation IoU and Video Panoptic Quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 525, + 167, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 167, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 167, + 536 + ], + "type": "text", + "content": "The formula is shown below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 543, + 250, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 543, + 250, + 574 + ], + "spans": [ + { + "bbox": [ + 83, + 543, + 250, + 574 + ], + "type": "interline_equation", + "content": "\\mathrm {V P Q} = \\sum_ {t = 0} ^ {H} \\frac {\\sum_ {\\left(p _ {t} , q _ {t}\\right) \\in T P _ {t}} \\operatorname {I o U} \\left(p _ {t} , q _ {t}\\right)}{\\left| T P _ {t} \\right| + \\frac {1}{2} \\left| F P _ {t} \\right| + \\frac {1}{2} \\left| F N _ {t} \\right|}", + "image_path": "7460e616947c0f9624ca01382e4121636aa7e2b5a122eb17dec3c74eac7fc755.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "content": " is the sequence length, " + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "inline_equation", + "content": "TP_{t}" + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "content": " represents the set of true positives, " + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "inline_equation", + "content": "FP_{t}" + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "content": " represents the set of false positives and " + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "inline_equation", + "content": "FN_{t}" + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "content": " represents the set of false negatives at timestamp " + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 574, + 287, + 611 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 614, + 124, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 124, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 124, + 625 + ], + "type": "text", + "content": "4.3. PnP results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Perception and Prediction. Table 1 compares TBP-Former with other methods of perception and prediction task based on multi-view cameras. We see that i) we achieve state-of-the-art performance and exceed previous methods by a large margin. ii) Even though BEVerse has larger RGB resolutions, TBP-Former still surpasses their performance on IoU by " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "7.3\\% / 8.3\\%" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " for short/long settings, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "content": "TBP-Former also improves the VPQ by " + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "inline_equation", + "content": "12.1\\% / 10.8\\%" + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "content": ". iii) Apart from the performance improvement, TBF-Former also has a larger FPS compared to other methods. Its inference speed is " + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "content": " faster than BEVerse's." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 384 + ], + "type": "text", + "content": "Fig. 5 shows the visualization results of our proposed method. We see that i) almost all the objects are detected correctly except for those occluded ones. ii) TBP-Former is capable of capturing the motion information in past frames and precisely predicting the vehicles' trajectories by occupancy and flow. Compared with FIERY [16], TBP-Former is closer to the ground truth. iii) TBP-Former does a better job than FIERY when predicting vehicles' turning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 492 + ], + "type": "text", + "content": "Perception Only. Table 2 compares the results of plenty of state-of-the-art methods on perception (segmentation) task. We see that our static model, which does not contain temporal information, can achieve 44.8 and 17.2 IoU of vehicles and pedestrians. With the input of temporal sequences, the performance improves further since auxiliary information is provided for better perception. The state-of-the-art results prove the effectiveness of the novel design of our Pose-synchronized BEV encoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 495, + 370, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 370, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 370, + 506 + ], + "type": "text", + "content": "4.4. Ablation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "content": "Effectiveness of PoseSync View Projection. The Exp. 1&4, 2&5, 3&6 in Table 3 compare the proposed PoseSync View Projection and the existing feature warping methods. We see that the proposed method always achieves better performance when other settings remain the same. The reasons are that: i) PoseSync View Projection based on Deformable Attention can guarantee the precise correspondence between BEV grids and image features. ii) Our projection method can alleviate distortion and our-of-range issues when synchronizing sequential BEV features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "Effectiveness of the designed future queries. In Exp. 1&3 in Table 3, we utilize the identical query with temporal positional encoding for future queries. Exp. 2&4 in Table 3 demonstrate that using separate learnable embedding for future queries can achieve better performance. Exp. 3&6 in Table 3 validate the efficacy of the proposed spatial priors for future queries. The generated high" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1374" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 539, + 393 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 539, + 393 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 539, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 539, + 393 + ], + "type": "image", + "image_path": "05ace495861ad71b1b2312e77f5dacfff1efdf7fd70b78a3a7c7c32bc1eb1dca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 547, + 437 + ], + "type": "text", + "content": "Figure 5. Demonstration of our results compared with FIERY and Ground Truth. Different vehicles are assigned with different colors in order to make a distinction. The darker parts represent the perception of the current frame, and the lighter parts represent the prediction of the vehicles in future frames. The visualization is based on the predicted occupancy and flow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 448, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 288, + 495 + ], + "type": "text", + "content": "dimensional map features provide the prediction model with useful geographic information. The additional spatial information can aid the prediction and lead to better scene forecasting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "content": "Effectiveness of STPT. Table 4 compares STPT with popular CNN-based [53] and RNN-based [16, 56] methods. We keep all the settings the same except for temporal modeling. To be specific, the size of input images " + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "inline_equation", + "content": "(224 \\times 480)" + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "content": ", image backbones and BEV feature extractor are the same. And then we plug their temporal models into our architecture. We see that i) STPT model performs better in all four metrics, including semantic segmentation IoU and three instance segmentation metrics from the video prediction area. ii) Our reproduced temporal models achieve higher performance than the original implements. This further validates the effectiveness and power of our BEV feature extractor." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Data Augmentation. We perform both image-view and BEV augmentations. The image-view augmentations include random scaling, rotation and flip of the input images. The BEV augmentations include similar operations on both BEV representations and corresponding ground truth labels. Table 5 compares the results of different data augmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 448, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 510 + ], + "type": "text", + "content": "strategies. We see that i) both augmentation methods improve the performance when used separately. ii) The combination of two methods works better than any single approach. Introducing data augmentation strategies is beneficial to the model's robustness and generalization ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 514, + 379, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 379, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 379, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 527, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 647 + ], + "type": "text", + "content": "This paper proposes a novel TBP-Former for vision-centric joint perception and prediction. We design a pose-synchronized BEV encoder module using a cross-view attention mechanism to solve the distortion issues in previous works. Furthermore, we propose a powerful spatial-temporal pyramid transformer for BEV feature extraction and BEV state prediction. Experiments show that i) TBP-Former improves the prediction performance over state-of-the-art methods significantly; and ii) both PoseSync BEV Encoder and STPT contribute to better performances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 647, + 547, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 547, + 706 + ], + "type": "text", + "content": "Acknowledgement This research is partially supported by National Natural Science Foundation of China under Grant 62171276 and the Science and Technology Commission of Shanghai Municipal under Grant 21511100900 and 22DZ2229005." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1375" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 190 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Sergio Casas, Cole Gulino, Renjie Liao, and Raquel Urtasun. Spagnn: Spatially-aware graph neural networks for relational behavior forecasting from sensor data. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9491-9497. IEEE, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "text", + "content": "[4] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[5] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 1, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 390 + ], + "type": "text", + "content": "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. arXiv preprint arXiv:2203.11089, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 446 + ], + "type": "text", + "content": "[7] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl Wellington. 3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception. IEEE Signal Processing Magazine, 38(1):68-86, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 491 + ], + "type": "text", + "content": "[8] Rahul Dey and Fathi M Salem. Gate-variants of gated recurrent unit (gru) neural networks. In 2017 IEEE 60th international midwest symposium on circuits and systems (MWS-CAS), pages 1597–1600. IEEE, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 546 + ], + "type": "text", + "content": "[9] Artem Filatov, Andrey Rykov, and Viacheslav Murashkin. Any motion detector: Learning class-agnostic scene dynamics from a sequence of lidar point clouds. In 2020 IEEE international conference on robotics and automation (ICRA), pages 9498–9504. IEEE, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 602 + ], + "type": "text", + "content": "[10] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17441-17451, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 667 + ], + "type": "text", + "content": "[11] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[12] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martín-Martín, and Li Fei-Fei. Maskvit: Masked visual pre-training for video prediction. arXiv preprint arXiv:2206.11894, 2022. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. A simple baseline for bev perception without lidar. arXiv e-prints, pages arXiv-2206, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 175 + ], + "type": "text", + "content": "[15] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 241 + ], + "type": "text", + "content": "[16] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "type": "text", + "content": "[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 342 + ], + "type": "text", + "content": "[18] Yue Hu, Siheng Chen, Ya Zhang, and Xiao Gu. Collaborative motion predication via neural motion message passing. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "text", + "content": "[19] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), November 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "text", + "content": "[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 498 + ], + "type": "text", + "content": "[21] Kuan-Hui Lee, Matthew Kliemann, Adrien Gaidon, Jie Li, Chao Fang, Sudeep Pillai, and Wolfram Burgard. Pillarflow: End-to-end birds-eye-view flow estimation for autonomous driving. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2007-2013. IEEE, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 499, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 545, + 544 + ], + "type": "text", + "content": "[22] Peixuan Li and Jieyu Jin. Time3d: End-to-end joint monocular 3d object detection and tracking for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3885-3894, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[23] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 633 + ], + "type": "text", + "content": "[24] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 634, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 690 + ], + "type": "text", + "content": "[25] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553-11562, 2020. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[26] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1376" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[27] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 250 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "text", + "content": "[31] Chenyang Lu, Marinus Jacobus Gerardus van de Molengraft, and Gijs Dubbelman. Monocular semantic occupancy grid mapping with convolutional variational encoder-decoder networks. IEEE Robotics and Automation Letters, 4(2):445-452, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "text", + "content": "[32] Liying Lu, Ruizheng Wu, Huaijia Lin, Jiangbo Lu, and Jiaya Jia. Video frame interpolation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3532-3542, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 393 + ], + "type": "text", + "content": "[33] Chenxu Luo, Xiaodong Yang, and Alan Yuille. Self-supervised pillar motion learning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3183-3192, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "text", + "content": "[34] Wenjie Luo, Bin Yang, and Raquel Urtasun. Fast and furious: Real time end-to-end 3d detection, tracking and motion forecasting with a single convolutional net. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 3569-3577, 2018. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 513 + ], + "type": "text", + "content": "[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified multi-task model for behavior prediction and planning. arXiv e-prints, pages arXiv-2106, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[36] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[37] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. arXiv preprint arXiv:2203.04050, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, pages 194-210. Springer, 2020. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[39] John Phillips, Julieta Martinez, Ioan Andrei Bársan, Sergio Casas, Abbas Sadat, and Raquel Urtasun. Deep multi-task learning for joint localization, perception, and prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4679-4689, 2021. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[40] Zequn Qin, Jingyu Chen, Chao Chen, Xiaozhi Chen, and Xi Li. Uniformer: Unified multi-view fusion transformer for spatial-temporal representation in bird's-eye-view. arXiv preprint arXiv:2207.08536, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "text", + "content": "[41] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent video transformer. arXiv preprint arXiv:2006.10704, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[42] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11138-11147, 2020. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[43] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In European Conference on Computer Vision, pages 414-430. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[44] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In 2022 International Conference on Robotics and Automation (ICRA), pages 9200-9206. IEEE, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 364 + ], + "type": "text", + "content": "[45] Marcel Schreiber, Vasileios Belagiannis, Claudius Gläser, and Klaus Dietmayer. Dynamic occupancy grid mapping with recurrent neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6717-6724. IEEE, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[46] Meet Shah, Zhiling Huang, Ankit Laddha, Matthew Langford, Blake Barber, Sidney Zhang, Carlos Vallespi-Gonzalez, and Raquel Urtasun. Liranet: End-to-end trajectory prediction using spatio-temporal radar fusion. arXiv preprint arXiv:2010.00731, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 475 + ], + "type": "text", + "content": "[47] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 522, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 588 + ], + "type": "text", + "content": "[49] Hengli Wang, Peide Cai, Yuxiang Sun, Lujia Wang, and Ming Liu. Learning interpretable end-to-end vision-based motion planning for autonomous driving with optical flow distillation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13731-13737. IEEE, 2021. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 655 + ], + "type": "text", + "content": "[50] Yunlong Wang, Hongyu Pan, Jun Zhu, Yu-Huan Wu, Xin Zhan, Kun Jiang, and Diange Yang. Be-sti: Spatial-temporal integrated network for class-agnostic motion prediction with bidirectional enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17093–17102, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 689 + ], + "type": "text", + "content": "[51] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "text", + "content": "[52] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying unknown instances for au" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1377" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 418 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "tonomous driving. In Conference on Robot Learning, pages 384-393. PMLR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 160 + ], + "type": "text", + "content": "[53] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "text", + "content": "[54] Enze Xie, Zhiding Yu, Daquan Zhou, Jonah Philion, Anima Anandkumar, Sanja Fidler, Ping Luo, and Jose M Alvarez. M^2bev: Multi-camera joint 3d detection and segmentation with unified birds-eye view representation. arXiv preprint arXiv:2204.05088, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 273 + ], + "type": "text", + "content": "[55] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 275, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 327 + ], + "type": "text", + "content": "[56] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 331, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 286, + 373 + ], + "type": "text", + "content": "[57] Brady Zhou and Philipp Kähenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13760-13769, 2022. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 375, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 375, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 375, + 286, + 418 + ], + "type": "text", + "content": "[58] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1378" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_content_list.json b/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c10901520026056a73a7a3de34bdc80cff1e5807 --- /dev/null +++ b/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_content_list.json @@ -0,0 +1,1409 @@ +[ + { + "type": "text", + "text": "TINC: Tree-structured Implicit Neural Compression", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runzhao Yang \nDepartment of Automation, Tsinghua University \nBeijing 100084, China \nyangrz20@mails.tsinghua.edu.cn", + "bbox": [ + 292, + 181, + 678, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 264, + 286, + 343, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit neural representation (INR) can describe the target scenes with high fidelity using a small number of parameters, and is emerging as a promising data compression technique. However, limited spectrum coverage is intrinsic to INR, and it is non-trivial to remove redundancy in diverse complex data effectively. Preliminary studies can only exploit either global or local correlation in the target data and thus of limited performance. In this paper, we propose a Tree-structured Implicit Neural Compression (TINC) to conduct compact representation for local regions and extract the shared features of these local representations in a hierarchical manner. Specifically, we use Multi-Layer Perceptrons (MLPs) to fit the partitioned local regions, and these MLPs are organized in tree structure to share parameters according to the spatial distance. The parameter sharing scheme not only ensures the continuity between adjacent regions, but also jointly removes the local and non-local redundancy. Extensive experiments show that TINC improves the compression fidelity of INR, and has shown impressive compression capabilities over commercial tools and other deep learning based methods. Besides, the approach is of high flexibility and can be tailored for different data and parameter settings. The source code can be found at https://github.com/RichealYoung/TINC.", + "bbox": [ + 76, + 316, + 472, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the big data era, there exist massive and continuously growing amount of visual data from different fields, from surveillance, entertainment, to biology and medical diagnosis. The urgent need for efficient data storage, sharing and transmission all requires effective compression techniques. Although image and video compression have been studied for decades and there are a number of widely used commercial tools, the compression of medical and biological data", + "bbox": [ + 76, + 733, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work was done in collaboration with Tingxiong Xiao and Yuxiao Cheng, under supervision of Prof. Jinli Suo and Prof. Qionghai Dai affiliated with Department of Automation, Tsinghua Univ.", + "bbox": [ + 75, + 862, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are not readily available.", + "bbox": [ + 500, + 287, + 665, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit neural representation (INR) is becoming widely-used for scene rendering [19, 21, 23, 31, 44], shape estimation [8, 11, 25, 29], and dynamics modeling [15, 26, 28]. Due to its powerful representation capability, INR can describe nature scenes at high fidelity with a much smaller number of parameters than raw discrete grid representation, and thus serves as a promising data compression technique.", + "bbox": [ + 496, + 305, + 893, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In spite of the big potential, INR based compression is quite limited confronted with large sized data [20], since INR is intrinsically of limited spectrum coverage and cannot envelop the spectrum of the target data, as analyzed in [41]. Two pioneering works using INR for data compression, including NeRV [7] and SCI [41], have attempted to handle this issue in their respective ways. Specifically, NeRV introduces the convolution operation into INR, which can reduce the required number of parameters using the weight sharing mechanism. However, convolution is spatially invariant and thus limits NeRV's representation accuracy on complex data with spatial varying feature distribution. Differently, to obtain high fidelity on complex data, SCI adopts divide-and-conquer strategy and partitions the data into blocks within INR's concentrated spectrum envelop. This improves the local fidelity, but cannot remove non-local redundancies for higher compression ratio and tend to cause blocking artifacts.", + "bbox": [ + 496, + 412, + 895, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To compress large and complex data with INR, in this paper we propose to build a tree-structured Multi-Layer Perceptrons (MLPs), which consists of a set of INRs to represent local regions in a compact manner and organizes them under a hierarchical architecture for parameter sharing and higher compression ratio. Specifically, we first draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress different regions with multiple MLPs separately. Then we incorporate these MLPs with a tree structure to extract their shared parameters hierarchically, following the rule that spatially closer regions are of higher similarity and share more parameters. Such a joint representation strategy can remove the redundancy both within each local region and among non-local", + "bbox": [ + 496, + 688, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18517", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "regions, and also ensures the continuity between adjacent regions. The scheme of the proposed network is illustrated in Fig. 1, and we name our approach TINC (Tree-structured Implicit Neural Compression).", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using the massive and diverse biomedical data, we conduct extensive experiments to validate that TINC greatly improves the capability of INR and even outperforms the commercial compression tools (H.264 and HEVC) under high compression ratios. The proposed TINC is also a general framework, which can be flexibly adapted to diverse data and varying settings.", + "bbox": [ + 75, + 151, + 468, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 271, + 218, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Compression. Data compression is a key technique for efficient data storage, transmission and sharing, especially in the big data era. In recent decades, many commercial compression tools for massive visual data have become increasingly mature, like JPEG [37], JPEG2000 [32], H.264 [39], and HEVC [33], while emerging deep learning-based approaches are also gaining momentum and demonstrating continuously improving compression capabilities [2, 3, 16, 22, 40, 42, 43]. Compression for nature images and videos have been widely and successfully used, however, there still lack effective compression techniques for the huge amount of three or higher dimensional data such as biomedical volume. Image/Video compressors cannot be applied directly, because these high dimensional data is a discrete sampling of 3D organism, instead of a video describing temporal evolution of 2D slices. Hence, the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured data [41], and tend to cause discontinuity and degenerate the downstream tasks. Our approach differs completely from existing image/video compression methodologies in that we encode the intrinsic data structure with INR. Besides, our approach searches for the best network structure and parameters representing the target data and is free of generalization issues.", + "bbox": [ + 75, + 296, + 470, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Neural Compression. Recent works [7, 17, 41] have introduced INR to the field of data compression with exciting results. They fit the original discrete grid-based data with continuous implicit functions in a transform-coding-like manner, which takes advantage of the powerful representation capabilities of neural networks while avoids the generalization problem. Obviously, the compression capability of INR is closely related to its representation capability, which can be improved from proper design of network structure [7, 10, 17], selection of activation function [14, 20, 30], embedding [4, 21, 23, 36, 45], and learning strategies such as manifold learning [9, 20, 30, 31], meta learning [5, 9, 35] and ensemble learning [1, 12, 24].", + "bbox": [ + 75, + 674, + 468, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, improving these capability does not necessarily head for a ready compressor for general scenes or ob", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "jects, since there exist large diversity in the target data and pursuing a general network with high fidelity for all the data under given parameter budget. In order to pursue a more compact and efficient representation, we draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress the data in each region separately. On top of this, we improve INR by introducing a hierarchical parameter sharing mechanism in terms of data compression.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter Sharing in INR. To improve the representation accuracy of INR for large and complex scenes, many algorithms use the parameter sharing strategy. One implementation is introducing convolution operations [6, 7, 27], which allows hierarchical extraction of local features, reduces the required parameters, and enables describing high resolution data. However, for those complex data with uneven distribution of features, the spatial invariance of convolution would lead to lower local representation accuracy. Another parameter sharing mechanism is positional encoding [21, 36], in which nearby spatial positions share partial mapping features hierarchically. However, this can only achieve higher representation accuracy when equipped with random or learnable mappings [4, 30, 36], which can be considered as an embedding that encodes only spatial distance. In addition, using embedding [18, 23, 34] intrinsically reduces the computational complexity at expenses of increased parameters, and thus not optimal for compression.", + "bbox": [ + 496, + 233, + 892, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To make use of advantageous and bypass the above issues, we borrow the idea of hierarchical parameter sharing into local INRs, and build our compressor on a widely used tree structured data partitioning [18, 34, 41]. On the one hand, the parameter sharing is applied among block-specific INRs to cope with the large data with high local fidelity. On the other hand, we share parameters on INR rather than the data itself or its embedding, to encode richer information other than spatial distance.", + "bbox": [ + 496, + 513, + 890, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Tree-structured Implicit Neural Compression", + "text_level": 1, + "bbox": [ + 500, + 684, + 888, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Ensemble of Implicit Neural Compressors", + "text_level": 1, + "bbox": [ + 500, + 734, + 857, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit neural compression uses a compact neural network to parameterize a function describing the target data defined over a spatial coordinates system. More specifically, for an $N$ dimensional data, let $\\mathbb{V} \\coloneqq [-1,1]^N$ denote its grid coordinates and $\\mathbf{d}(\\mathbf{v})$ with $\\mathbf{v} \\in \\mathbb{V}$ denote the intensity value at coordinate $\\mathbf{v}$ , one can learn a compact implicit function $f$ with parameter $\\Theta$ to estimate $\\mathbf{d}(\\mathbf{v})$ by $f(\\mathbf{v};\\Theta)$ and finally serialize them to achieve a compressed representation. Mathematically, we can formulate the compression", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18518", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f688fdac98302532d50afa6477736159751ed4539cb8c6909d20a5aeef44ccc.jpg", + "image_caption": [ + "Figure 1. The scheme of the proposed approach TINC. For a target data, we divide the volume into equal-size blocks via octree partitioning, with some neighboring and far-apart blocks of similar appearances. Then each block can be representation with an implicit neural function, implemented as an MLP. After sharing parameters among similar blocks, we can achieve a more compact neural network with a tree shaped structure. Here we highlight the similar blocks sharing network parameters with the same color." + ], + "image_footnote": [], + "bbox": [ + 98, + 109, + 883, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task as an optimization problem", + "bbox": [ + 76, + 419, + 290, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\Theta}} \\int_ {\\mathbf {v} \\in \\mathbb {V}} \\mathcal {L} (f (\\mathbf {v}; \\boldsymbol {\\Theta}), \\mathbf {d} (\\mathbf {v})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 440, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ measures the difference between the decompressed and target data. For decompression, one can retrieve $f$ and $\\Theta$ from the compressed file and then calculate function values $f(\\mathbf{v};\\Theta)$ over grid $\\mathbb{V}$ via simple forward propagation.", + "bbox": [ + 76, + 486, + 468, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to limited spectrum coverage, INR is incapable of compactly describe large and complex data [20,41], we borrow the idea of ensemble learning to partition the target volume into blocks and use multiple less expressive $f_{k}(\\cdot ,\\Theta_{k})$ to achieve a powerful representation. Specifically, we adopt the divide-and-conquer strategy as in [12, 20, 41] to ensemble all implicit functions $\\{f_k,k = 1,\\dots ,K\\}$ that represents data at its corresponding coordinate regions $\\{\\mathbb{V}_k,k = 1,\\dots ,K\\}$ and compose an $f$ , which is equivalent to approximating a complex function by a piece-wise function composed of $K$ simple functions", + "bbox": [ + 76, + 547, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {v}; \\boldsymbol {\\Theta}) := \\sum_ {k = 1} ^ {K} \\mathbf {1} _ {\\mathbb {V} _ {k}} \\left(\\mathbf {v}, f _ {k} \\left(\\mathbf {v}, \\boldsymbol {\\Theta} _ {k}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 719, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {1} _ {\\mathbb {V} _ {k}} (\\mathbf {v}, x) = \\left\\{ \\begin{array}{l} x, \\mathbf {v} \\in \\mathbb {V} _ {k} \\\\ 0, e l s e \\end{array} . \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 763, + 468, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we must constrain the total number of parameters to be consistent with the parameter budget, i.e., $|\\Theta| = \\sum_{k=1}^{K} |\\Theta_k|$ , where $|\\cdot|$ is the support operation. Here we omit the parameters for serializing $f_k$ which consists of only a few structural hyperparameters of neural network.", + "bbox": [ + 76, + 809, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a target data volume shown in the first column of Fig. 1, we use an $L$ level tree structure containing $K$ leaf nodes to organize $f_{k}$ and $\\mathbb{V}_k$ , as illustrated in the 2nd column of Fig. 1. The root node represents the whole coordinate region $\\mathbb{V}$ and each child node represents $\\frac{1}{2^N}$ region of its parent node. We use $f_{k}$ to compress the data in coordinate region $\\mathbb{V}_k$ represented by the $k$ th leaf node. This structure allows us to organize ensemble learning in high degree of flexibility.", + "bbox": [ + 496, + 419, + 890, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hierarchical Parameter Sharing Mechanism", + "text_level": 1, + "bbox": [ + 500, + 588, + 875, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We let these $\\{f_k\\}$ share their neural network parameters hierarchically with each other according to the spatial distance between corresponding regions $\\mathbb{V}_k$ . The sharing mechanism is defined on the octree structure. Specifically, for a leaf node at level $l$ , its corresponding MLP implemented $f_k$ 's hidden layers can be divided into $l$ segments, i.e. $f_k = f_k^{out} \\circ f_k^l \\circ f_k^{l-1} \\circ \\dots \\circ f_k^1 \\circ f_k^{in}$ . For a set of leaf nodes, we determine the number of shared segments based on the number of their common ancestor nodes. For example, if $f_i$ and $f_j$ share the same ancestor nodes at $1 \\sim 3$ levels, three pairs of hidden layer segments $(f_i^1, f_j^1)$ , $(f_i^2, f_j^2)$ , $(f_i^3, f_j^3)$ will share the same parameters.", + "bbox": [ + 496, + 619, + 890, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sharing mechanism is illustrated in the 3rd column in Fig. 1. Since the distance between two leaf nodes can also be inferred from the number of their shared ancestor nodes, above hierarchical parameter sharing mechanism implies that closer blocks share more parameters and are of higher similarity in the implicit representation space.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18519", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Tree-structured Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 90, + 406, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a tree-structured MLP based on the $L$ level octree partitioning, as illustrated in the 4th column in Fig. 1. Each node contains a hyper layer consisting of some fully connected layers and takes the output of its parent node's hyper layer as input. Root node and leaf nodes additionally contain the input and output layers respectively.", + "bbox": [ + 76, + 114, + 467, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the depth direction, from leaf nodes upward to the root of the tree, the output information of the leaf node is processed by the hyper layers in its ancestor nodes. If we consider the hyper layer as a part of hidden layers, then any leaf node can be considered as an MLP, whose hidden layers are explicitly divided into multiple segments contained in its ancestor nodes across different levels. In this way we can ensemble all $\\{f_k\\}$ into a tree-structured MLP.", + "bbox": [ + 75, + 205, + 467, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the same level, all sibling nodes share the same parent node and thus take the same information as input. Therefore, for any two leaf nodes sharing the same ancestor nodes, their output information have flown through the same hyper layers contained in the ancestor nodes. In other words, the MLPs of these two leaf nodes will share a part of hidden layers, which implements the hierarchical parameter sharing mechanism described in Sec. 3.2.", + "bbox": [ + 75, + 327, + 467, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Cross Level Parameter Allocation", + "text_level": 1, + "bbox": [ + 76, + 464, + 383, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the network design, we are left with some key implementation issues: (i) how to set a proper number of levels to finalize the tree structured network, (ii) how to allocate available parameters across MLPs at the tree nodes.", + "bbox": [ + 76, + 489, + 467, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Number of Tree Levels $L$ . The number of tree levels $L$ determines the number of divided segments of $f_{k}$ 's hidden layers, i.e., a larger $L$ implies the parameter sharing mechanism proceeds to higher spatial resolution. Setting a large $L$ has both advantageous and disadvantages. On the one hand, a larger $L$ means finer partitioning with a larger number of blocks sharing parameters (with the same hyper layers as in the ancestor nodes), which is beneficial for describing the regions with rich details and thus improve the compression fidelity. On the other hand, with the same total number of parameters, a larger $L$ will lead to more leaf nodes $K$ and less parameters $|\\Theta_k|$ each $f_{k}$ , which might reduce $f_{k}$ 's representation capability and might harm the compression fidelity in turn.", + "bbox": [ + 75, + 551, + 467, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we should set the tree level $L$ by trading off between increased fineness in partitioning and parameter allocation, and $f_{k}$ decreased representation capability. In Sec. 5.3, we experimentally prove that it is favourable using large tree levels for the data with rich details or under low compression ratios.", + "bbox": [ + 75, + 763, + 467, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inter-level Parameter Allocation. The parameters at the shallow level describe the shared features at large scale, i.e., the non-local redundancy distributed over distant regions.", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, from the perspective of data features, if some far apart sub-regions at a certain level are highly similar, one can allocate more parameters to their common ancestor nodes, and vice versa.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, for data with high global redundancy or repetitiveness, i.e., high non-local similarity among sub-regions at large scales, allocating more parameters to the shallow level will be more beneficial to improve its compression fidelity. Conversely, for data with weak global similarity, it is more efficient to allocate more parameters to the nodes at deep levels for better representation of the unique local features.", + "bbox": [ + 496, + 152, + 890, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intra-level Parameter Allocation. In addition to allocating parameters across levels, we can also allocate more parameters within a level to some nodes according to certain criteria instead of even allocation. For example, the most valuable information in neuronal data is often distributed in sparser regions. If we allocate more parameters to the more important regions, then we can sacrifice the unimportant regions in exchange for the improvement of compression fidelity of the important ones.", + "bbox": [ + 496, + 273, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A simple way to take advantage of the TINC's hierarchy and flexibility is to allocate the number of parameters per node at each level according to the importance of the corresponding regions. Following this allocation, at each spatial resolution, the nodes corresponding to more important regions will have more parameters compared to other nodes in the same level. We demonstrate in Sec. 5.3 that this simple approach can significantly improve the compression fidelity of biological data, especially in regions with very low compression fidelity in the even allocation case, if an appropriate measure of the importance of the data region is adopted. Obviously there are more complex and efficient allocation methods, but this preliminary exploration is sufficient to demonstrate the advantages of TINC.", + "bbox": [ + 496, + 411, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 638, + 633, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 664, + 715, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All the hyper layers in each node are implemented using MLP, where the activation functions are all sine function and the parameters are initialized following the scheme proposed in [30]. All MLPs are optimized by Adamax [13] with initiate learning rate of 0.001 and decreased by a factor of 0.2 at the 2000th and 5000th iterations, respectively. We adopted a complete octtree, which means $K = 8^{(L - 1)}$ . Before optimization, all the coordinates and the intensities are normalized to $[-1,1]$ and $[0,100]$ respectively. After optimization, all the parameters of each hyper layers are serialized and stored in binary. There are also some information that need to be saved for data decompression, such as MLP structure, original data bit depth, image size, inverse normalization basic information. Because all of the items are", + "bbox": [ + 496, + 688, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18520", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5670dc8f4b6df00fd72708d8e97e68803906e24c7ed7dcad4c4f55a89b673e35.jpg", + "table_caption": [ + "Table 1. The scores on each dataset as compression ratio ranges from 64 to 1024, with the #best and the ●second best marked by color and symbols. The metrics include PSNR(dB), SSIM, and Acc. $\\tau$ with $\\tau$ acts as a threshold for binarization. The “All” or “High” suffixes after the metrics name indicate that it is averaged for all compression ratios or only high compression ratios (around $1024 \\times$ )." + ], + "table_footnote": [], + "table_body": "
MethodMedical dataBiological data
PSNR All. (dB)SSIM All.PSNR High. (dB)SSIM High.Acc.200 All.Acc.500 All.Acc.200 High.Acc.500 High.
TINC (ours)•52.02•0.9897#50.59#0.9878•0.9945•0.9970#0.9934#0.9958
JPEG41.410.972230.490.93740.66120.98340.01970.9882
H.26451.180.989647.280.98600.99190.99590.98600.9926
HEVC#52.31#0.9903•50.51•0.9877#0.9955#0.99750.99170.9930
SCI51.900.989450.390.98760.99430.9965•0.9921•0.9951
NeRF50.930.987549.660.98630.99350.99620.99030.9940
NeRV47.110.985940.110.98000.98150.99010.97320.9867
DVC47.390.986545.740.98400.98270.99000.96920.9789
SGA+BB46.560.983643.020.98080.80380.98830.48170.9798
SSF46.250.980743.700.97730.72210.96030.77900.9542
", + "bbox": [ + 83, + 141, + 883, + 319 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/603eaac4398fd19d1861d67fb73f31a0f087cdd476fb18b35928fcf7f37abd8f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 335, + 285, + 445 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac4c74b1044fdfdc78f00d0733c68e363b122676e0f411f3f6d0978be2a8773c.jpg", + "image_caption": [ + "(b)", + "Figure 2. Comparison of different compression methods on medical (a,b) and biological data (c,d) at different compression ratios." + ], + "image_footnote": [], + "bbox": [ + 285, + 335, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/748e1e3b29dc7097da206165dc5c05ddfbe7a45c81a1e0386d4fa281119ce25e.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 483, + 335, + 681, + 445 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3c0f23b2c2a193454a213dc199bdd25a9533957ef635341e175347f0bec4a64.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 681, + 335, + 879, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "discrete, countable, and have a limited set of possible values, we can easily create a table listing all potential values and store their indices, to reduce the file size.", + "bbox": [ + 75, + 505, + 468, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated our approach on both massive medical and biological data. For the former we used HiP-CT dataset [38], which provides cellular level imaging of several organisms at multiple anatomical levels. We used four human organs (Lung, Heart, Kidney and Brain) from this dataset. For the latter we used the sub-micrometer resolution mouse whole-brain microscopic data (Neurons and Vessels) captured by our self-developed imaging system (see supplementary note 1 for details), which has varying brightness and structures across different regions. The access to these data are listed in supplementary note 2.", + "bbox": [ + 75, + 550, + 470, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For convenient and fair comparison, we cropped the data into the same size ( $256 \\times 256 \\times 256$ voxels for medical data and $64 \\times 512 \\times 512$ voxels for biological data).", + "bbox": [ + 75, + 717, + 470, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Performance Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 76, + 771, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Methods and Evaluation Metrics. The proposed method is comprehensively compared with state-of-the-art methods, which can be classified into three groups: (i) commercial compression tools, including JPEG [37], H.264 [39], and HEVC [33]; (ii) recently proposed data-driven deep learning based ones, including DVC [16],", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SGA+BB [43], and SSF [2]; (iii) recently proposed INR based ones, including SCI [41], NeRV [7], and NeRF [21]. We evaluated all the methods at 6 different compression ratios ranging from $64 \\times$ to $1024 \\times$ . It is worth noting that only INR based methods can precisely regulate the compression ratio. We adopted PSNR and SSIM for medical data including rich details, while Accuracy with threshold for binarization as evaluation metrics for biological data with sparse structures. For convenience we abbreviate Accuracy with threshold as \"Acc.threshold\" in the following descriptions.", + "bbox": [ + 496, + 505, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "JPEG was tested on OpenCV. H.264 and HEVC were tested with their FFmpeg implementations. For SGA+BB, we adopted the settings with the best performance mentioned in the original paper. For DVC, we used its Pytorch implementation, and compressed the \"I Frames\" using JPEG method with quality parameter set to 50. For SSF, we used its CompressAI's implementation. In addition, for the data-driven methods, i.e. DVC, SGA+BB, and SSF, we fine-tuned the provided pre-trained models on 300 volumes randomly selected from the dataset. For SCI, we set the hyper parameter $a_{max} = 32$ . For NeRF, the hyper parameter \"frequency\" was set to 10. For NeRV, we control the compression ratio by adjusting the number and width of the fully connected layers. For 2D images compression methods, i.e. JPEG and SGA+BB, we firstly segmented the 3D data int 2D images before compression.", + "bbox": [ + 496, + 657, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18521", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Performance Evaluation and Analysis. The average performance of each method under different compression ratios is plotted as a rate-distortion curve in Fig. 2. We can see that our algorithm outperforms almost all methods at high compression ratios and is comparable to HEVC at other compression ratios. For more detailed analysis, we summarize the scores on each dataset in Table 1. It should be noted that even if the PSNR of the medical data are higher than $40\\mathrm{dB}$ , it does not mean that there are no distortion in the decompressed data, which is different from the natural scene.", + "bbox": [ + 75, + 90, + 468, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Some visual comparisons of decompressed organ slices are presented in Fig. 3 (see supplementary figure 1 for more results). For better visual comparison, we applied the same post-processing to the decompressed data. Under compression ratio around $87 \\times$ (since HEVC and SGA+BB cannot precisely regulate compression ratio), we consistently outperform existing INR based (SCI and NeRV), data-driven (SGA+BB) and commercial (HEVC) compression methods. Although SCI achieved comparable results in the regions away from boundary, as shown in Fig. 3(a), it produced severe blocking artifacts at the boundary, as shown by the arrows in (b), which will affect the medical research as well as other downstream tasks. NeRV retained the general structures but suffered from the checkerboard artifacts caused by the aliasing in convolution. HEVC captured high fidelity details in \"I Frames\" but suffered from the artifacts in \"P Frames\", which distorted the edges in the volumes, such as the white matters in Brain data and muscles in Heart data, which would be unusable for medical research and other downstream tasks. SGA+BB struggled with retaining the rich details, since it cannot represent the similarity within the 3D volume as a 2D image compression algorithm. In addition, HEVC produced ghosting artifacts and position offset in neighbor slices, since the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured biomedical data. In contrast, our approach use INR to encode the intrinsic data structure.", + "bbox": [ + 75, + 258, + 472, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of Running Time. We summarized the time required for compression and decompression for each algorithm in our experiments, as shown in the supplementary tables 1 and 2. Comparatively, INR based compressor (TINC, SCI, NeRV, and NeRF) is slower in compression stage but can decompress faster. It is worth noting that biomedical data need to be compressed only once before saving or transmission, but decompressed frequently for processing and analysis. Therefore, TINC is advantageous in biomedical fields.", + "bbox": [ + 75, + 681, + 468, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Flexibility Settings for Different Data", + "text_level": 1, + "bbox": [ + 76, + 845, + 398, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setting of Tree Levels $L$ . We evaluated TINC at different tree levels $L$ on medical data under different com", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fc1755f1cf1667b217dd8c16719c101cfc7745e80f5a089430d5ab67f1c9841.jpg", + "image_caption": [ + "(a) Three ROIs from Brain data; compression ratio: $\\sim 87\\times$" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 887, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3c7e3676c9cb7992d190149b03dfe6d4882c7275262be0b10d320684534945c.jpg", + "image_caption": [ + "(b) Three ROIs from Heart data; compression ratio: $\\sim 87\\times$", + "Ground Truth TINC (ours) SCI NeRV HEVC SGA+BB", + "Figure 3. Visual comparisons of decompressed organs slices from Brain and Heart data between TINC and other benchmark methods under similar compression ratio around $87\\times$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 248, + 887, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pression ratios. As shown in Fig. 4(a), the performance of TINC was evaluated in terms of SSIM on 8 medical data, under $256 \\times$ compression ratio. To better demonstrate the effect of increasing tree levels, we used the growth rate of SSIM, calculated by normalizing all SSIM values by the one when $L = 1$ . The SSIM scores on all the test data have a steadily increasing growth rate (dashed lines) as the number of tree depth increases from 1 to 3. Experimental results on other compression ratios show consistent trend (supplementary figure 2).", + "bbox": [ + 496, + 507, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further investigated the relationship between the data complexity and the performance gain when increasing tree levels. For data complexity, we adopt the spectrum concentration defined in [41], which measures the percentage of high frequency information contained in the data. The results are plotted in Fig. 4(b), with the solid line being a quadratic fitting of the growth rate for the 8 medical data using least squares. From the plot one can see that the performance gain increases monotonously with data complexity. In other words, the benefit of using a large tree is more significant when the target data contains more high-frequency information. This trend is consistent with our reasoning in Sec. 4 that a large $L$ favours parameter sharing at coarser levels and leave more parameters for representing the details. Experimental results on other compression ratios and other data are consistent with this (supplementary figure 3).", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18522", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ded2a7b60ff6b048a82e6e6bf217bd846cb69b4211ffe5361f31475e0f1eec0e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 256, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a34db492e70e691d1a0885f417450fdcb15e8473910ba19f8d09b57960200e0.jpg", + "image_caption": [ + "(b)", + "Figure 4. Effect of tree levels $L$ on TINC's compression fidelity, in terms of SSIM on 8 medical data under $256 \\times$ compression ratio. (a) The growth rate of SSIM for each data when increasing the tree levels. The dashed lines represent the change in each data. The solid line represents the mean of changes, and the filled area represents the standard deviation. (b) The scatter plot of each data's complexity and the growth rate of SSIM when increasing tree levels from 1 to 3. The solid line represents the trend of change, estimated by a quadratic fit." + ], + "image_footnote": [], + "bbox": [ + 289, + 92, + 464, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we studied the performance gain of using a deep tree structure under limited bandwidth budget $|\\Theta|$ , e.g., at high compression ratio. Supplementary figure 4 plots the performance improvement at increasing compression ratios, calculated from 8 medical data volumes. Almost all data have a smaller growth rate of SSIM or PSNR when compression ratio increases, i.e. reducing the number of available parameters. At high compression ratio $(512 \\times)$ , some data are even of inferior fidelity using a deep tree, which is in line with our analysis in Sec. 4 that one should set a proper $L$ for choosing the right partitioning granularity under the constraint from available budget.", + "bbox": [ + 75, + 404, + 468, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inter-level Parameter Allocation. We tested the performance at different parameter allocation proportions across levels on medical data under different compression ratios. Assuming there are $K^l$ nodes at $l$ th level, we denote the number of allocated parameters at this level as $|\\Theta^l|$ and ratio with respect to that of its adjacent upper level as $r^l = \\frac{|\\Theta^l| / K^l}{|\\Theta^{l-1}| / K^{l-1}}$ . Simply speaking, $r^l = 1$ means even inter-level allocation, a $r^l$ larger than 1 means imposing higher priority to the shallow levels and vice versa. We let $r^l$ of all levels take 0.8, 1 and 1.2 to represent the three typical cases of parameter allocation.", + "bbox": [ + 75, + 588, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first investigated the effect of allocating more parameters to the shallow level, intending to share parameters among far apart similar blocks. As shown in Fig. 5(a), we divided a Brain data into 64 sub-regions, laid out in z-curve order, and selected 8 equally spaced distant regions. Then, we calculated the SSIM of these regions with respect to each other and used min-max normalization to obtain their non-local similarities. It can be seen that #1 and #33 regions are of high similarity. For convenient display, we also draw the", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91a36dac14b05d35ba23d46b1725b495453af2222d405a136f09e9c3a8a053d4.jpg", + "image_caption": [ + "(a)", + "Figure 5. Effect of allocating more parameters to the shallow level on TINC's compression fidelity for each sub-regions in a Brain data under $512 \\times$ compression ratio. All sub-figures share the same y-axis label. (a) The heatmap of similarities between 8 equally spaced distant sub-regions. The serial numbers of the regions represent their z-curve order. (b) The non-local similarity of each region. (c) The growth rate of SSIM for each region when allocating more parameters to the shallow level." + ], + "image_footnote": [], + "bbox": [ + 514, + 92, + 661, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6eda077fe1153665b0fa6ec45599548cc30578461a2d7e857475a514a2d6c240.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 697, + 95, + 785, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8202af729ba2662701429c6056c779272874c2d12a347b38bb0e3d13f3145945.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 803, + 95, + 883, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bar plot of the normalized marginal sums of the similarity matrix in Fig. 5(b), which tells that #1 and #33 are both of high similarity to other blocks, while #25 and #57 regions are highly distinctive from others.", + "bbox": [ + 496, + 386, + 890, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantitatively measure the performance change at different blocks when allocating more parameters to the shallow level, we plot the growth rate of SSIM under $512 \\times$ ratio, as shown in Fig. 5(c). Analyzing jointly with (b), we can find that the two regions with high non-local similarity, #1 and #33, gain a large boost, while those with the lowest similarity, #25 and #57, suffer a loss of fidelity. This is consistent with the intuition and analysis in Sec. 4 that allocating more parameters to the nodes at shallow levels would benefit representing distant but similar regions, but at expense of slight degeneration at very individual regions. Experimental results on other compression ratios and other data are shown in supplementary figures 5 and 6.", + "bbox": [ + 496, + 446, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then investigated how to allocate parameters across levels according to the global consistency of the medical data. We divided the target data into 64 sub-regions, calculated the marginal sums of the similarity matrix. We change the setting of $r^l$ and the results are shown in the supplementary figure 7. The results show allocating more parameters to the shallow level is more likely to improve fidelity when the global consistency of the data is greater than 0.7, and allocating more to the deep level is better when the global consistency lower than 0.6. Otherwise, an even allocation may be a better choice.", + "bbox": [ + 496, + 643, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intra-level Parameter Allocation. We tested the approach of allocating parameters to the nodes at the same level according to their importance proposed in Sec. 4, on biological data and under different compression ratios. Specifically, we let the number of allocated parameters be proportional to their importance. For example, with regard", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18523", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to the biological data such as Neurons the importance can be measured by the percentage of the number of valid signals, where the valid signal can be determined either by the result of traced neurons or by the intensity value. Here we chose the latter for simplicity.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first compare the compression fidelity of 6 Neurons data using the above two parameter allocation approaches under different compression ratios, as shown in supplementary figure 8. It can be seen that important allocation significantly improves the fidelity compared to even allocation.", + "bbox": [ + 75, + 167, + 467, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then look further into the performance at all the sub-regions in the same target data. As shown in Fig. 6(a), we divided a Neurons data into 64 sub-regions and compute their growth rate of Acc.2000 when switching from even allocation to important allocation. It should be noted that we omitted those regions that have reached 1 (the beast value) under both allocation approaches in the subsequent experiments. From the scatter diagram we can see three groups of data: (i) Some regions show non-significant change (less than $10\\%$ of the maximum change of the others, denoted as cross markers) using important allocation under $64\\times$ compression ratio, because they are already with sufficient parameters. (iii) Other regions show significant improvement after introducing important allocation, denoted with circles. We performed a linear fit using least squares, as shown by the solid line. Overall, allocating parameters according to importance is helpful for the compression fidelity of important regions while at expenses of little fidelity degeneration in other less important ones.", + "bbox": [ + 75, + 242, + 467, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For a statistical analysis, we plotted the Acc.2000 of all 64 sub-regions under even and important allocation, as shown in Fig. 6(b). We found that importance based allocation can significantly improve those regions that have very poor compression fidelity under even allocation. This is mainly because that those regions with poor fidelity tend to contain more complex structures (with more neural signals), and thus are of higher importance scores and thus allocated more parameters accordingly. Experimental results on other compression ratios are detailed in supplementary figure 9.", + "bbox": [ + 75, + 529, + 467, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 709, + 204, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary and Discussions. In this paper we propose a hierarchical parameter sharing mechanism for INR and implement it based on a tree-structured MLP for large and complex data compression tasks. The proposed approach TINC makes use of the local redundancy via block wise compression with separate MLPs and non-local redundancy via organizing these MLPs in tree structure to share parameters hierarchically. Such implementation forms a single network for the complex target data, not only incorporate the similarity distributed far apart and nonuniformly, and avoid discontinuity between adjacent regions as well.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46013685926776d1b8ad2ff8cea383ee9f9cd81ca3d9d1549efd4869b79cc59b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 681, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b765ba7b63ae12e00243303b71fd684e9b9d720be18c02d882b6ffb64260a0a.jpg", + "image_caption": [ + "(b)", + "Figure 6. Effect of important allocation on TINC's compression fidelity, in terms of Acc.2000 on a Neurons data under $64 \\times$ compression ratio. (a) The growth rate of Acc.2000 for each region. The regions with and without significant growth rate are represented as cross and square respectively. The solid line represents the trend of change, estimated by a linear fit. (b) The violin plots of Acc.2000 for all these regions. The shaded area represents the distribution of Acc.2000 across all regions. The three horizontal lines from top to bottom represent the maximum, average and minimum values respectively." + ], + "image_footnote": [], + "bbox": [ + 710, + 90, + 890, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extensive experiments show that this hierarchical parameter sharing mechanism introduced in TINC greatly improves the compression capability and wide applicability of INR. Our method outperforms state-of-the-art compressors including commercial tools (JPEG, H.264, HEVC), data-driven methods (DVC, SSF, SGA+BB), and existing INR based ones (SCI, NeRV, and NeRF) on diverse medical and biological data.", + "bbox": [ + 496, + 417, + 890, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also analyze TINC's flexibility to different cases via experimentally studying the effect of three key settings—number of tree levels, intra-level and inter-level parameter allocation methods, and provide highlights on proper parameter setting and strategy selection according to the features of target data.", + "bbox": [ + 496, + 539, + 890, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Extensions. Similar to all current INR based compression methods, TINC is of high decompression speed but slow in compression, since it takes time to pursue the MLPs matching the target data. We plan to combine meta-learning to find the best initialization parameters for each organ to speed up TINC.", + "bbox": [ + 496, + 632, + 890, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 743, + 666, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We acknowledge MoE Key Laboratory of Biomedical Photonics (Huazhong University of Science and Technology) for sharing their mouse brain-wide microscopic data.", + "bbox": [ + 496, + 770, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by Beijing Natural Science Foundation (Grant No. Z200021), the National Natural Science Foundation of China (Grant Nos. 61931012, 62088102) and Project of Medical Engineering Laboratory of Chinese PLA General Hospital (Grant No. 2022SYSZ-ZKY21).", + "bbox": [ + 496, + 816, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18524", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arya Aftab, Alireza Morsali, and Shahrokh Ghaemmaghami. Multi-Head Relu Implicit Neural Representation Networks. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2510-2514, 2022. 2", + "[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 5", + "[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational Image Compression with a Scale Hyperprior. In International Conference on Learning Representations, 2018. 2", + "[4] Nuri Benbarka, Timon Höfer, Hamd ul-moqeet Riaz, and Andreas Zell. Seeing Implicit Neural Representations as Fourier Series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2283-2292, 2022. 2", + "[5] Alexander William Bergman, Petr Kellnhofer, and Gordon Wetzstein. Fast Training of Neural Lumigraph Representations using Meta Learning. In Advances in Neural Information Processing Systems, 2021. 2", + "[6] Alexandre Boulch and Renaud Marlet. Poco: Point Convolution for Surface Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 2", + "[7] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. NeRV: Neural Representations for Videos. Advances in Neural Information Processing Systems, 2021. 1, 2, 5", + "[8] Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 1", + "[9] Yilun Du, Katherine Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning Signal-agnostic Manifolds of Neural Fields. Advances in Neural Information Processing Systems, 2021. 2", + "[10] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J. Zico Kolter. Multiplicative Filter Networks. In International Conference on Learning Representations, 2021. 2", + "[11] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning Shape Templates with Structured Implicit Functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7154-7164, 2019. 1", + "[12] Milad Soltany Kadarvish, Hesam Mojtahedi, Hossein Entezari Zarch, Amirhossein Kazerouni, Alireza Morsali, Azra Abtahi, and Farokh Marvasti. Ensemble Neural Representation Networks. arXiv preprint arXiv:2110.04124, 2021. 1, 2, 3", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980, 2014. 4", + "[14] Sylwester Klocek, Lukasz Maziarka, Maciej Wolczyk, Jacek Tabor, Jakub Nowak, and Marek Śmieja. Hypernetwork" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Functional Image Representation. In International Conference on Artificial Neural Networks, pages 496-510, 2019. 2", + "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1", + "[16] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: An End-to-end Deep Video Compression Framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2, 5", + "[17] Y. Lu, K. Jiang, J. A. Levine, and M. Berger. Compressive Neural Representations of Volumetric Scalar Fields. Computer Graphics Forum, 40(3):135-146, 2021. 2", + "[18] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive Coordinate Networks for Neural Scene Representation. ACM Trans. Graph., 40(4), jul 2021. 2", + "[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1", + "[20] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated Periodic Activations for Generalizable Local Functional Representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, October 2021. 1, 2, 3", + "[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2, 5", + "[22] David Minnen, Johannes Balle, and George D Toderici. Joint Autoregressive and Hierarchical Priors for Learned Image Compression. In Advances in Neural Information Processing Systems, volume 31, 2018. 2", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph., 41(4), jul 2022. 1, 2", + "[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing Scenes as Compositional Generative Neural Feature Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2", + "[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019.", + "[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18525", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1", + "[27] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional Occupancy Networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2", + "[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1", + "[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned Implicit Function for High-resolution Clothed Human Digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 1", + "[30] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit Neural Representations with Periodic Activation Functions. In Advances in Neural Information Processing Systems, 2020. 2, 4", + "[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3d-structure-aware Neural Scene Representations. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[32] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The JPEG 2000 Still Image Compression Standard. IEEE Signal Processing Magazine, 18(5):36-58, 2001. 2", + "[33] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the High Efficiency Video Coding (HEVC) Standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 2, 5", + "[34] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2", + "[35] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned Initializations for Optimizing Coordinate-based Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2021. 2", + "[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In Advances in Neural Information Processing Systems, 2020. 2", + "[37] Gregory K Wallace. The JPEG Still Picture Compression Standard. IEEE Transactions on Consumer Electronics, 38(1):xviii-xxxiv, 1992. 2, 5", + "[38] CL Walsh, P Tafforeau, WL Wagner, DJ Jafree, A Bellier, C Werlein, MP Kühnel, E Boller, S Walker-Samuel, JL Rober" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tus, et al. Imaging Intact Human Organs with Local Resolution of Cellular Structures using Hierarchical Phase-Contrast Tomography. Nature Methods, 18(12):1532-1541, 2021. 5", + "[39] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the H.264/AVC Video Coding Standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 2, 5", + "[40] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for Video Compression with Hierarchical Quality and Recurrent Enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6628-6637, 2020. 2", + "[41] Runzhao Yang, Tingxiong Xiao, Yuxiao Cheng, Qianni Cao, Jinyuan Qu, Jinli Suo, and Qionghai Dai. SCI: A Spectrum Concentrated Implicit Neural Compression for Biomedical Data. arXiv preprint arXiv:2209.15180, 2022. 1, 2, 3, 5, 6", + "[42] Ruihan Yang, Yibo Yang, Joseph Marino, and Stephan Mandt. Hierarchical Autoregressive Modeling for Neural Video Compression. In ICLR, 2021. 2", + "[43] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving Inference for Neural Image Compression. In Advances in Neural Information Processing Systems, volume 33, pages 573-584, 2020. 2, 5", + "[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields from One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1", + "[45] Ellen D Zhong, Tristan Bepler, Joseph H Davis, and Bonnie Berger. Reconstructing Continuous Distributions of 3D Protein Structure from Cryo-EM Images. arXiv preprint arXiv:1909.05215, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 527 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18526", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_model.json b/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..46bfc790c3483a6115203bebd5a51de2356f9a9d --- /dev/null +++ b/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "TINC: Tree-structured Implicit Neural Compression" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.182, + 0.679, + 0.253 + ], + "angle": 0, + "content": "Runzhao Yang \nDepartment of Automation, Tsinghua University \nBeijing 100084, China \nyangrz20@mails.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.287, + 0.344, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.473, + 0.683 + ], + "angle": 0, + "content": "Implicit neural representation (INR) can describe the target scenes with high fidelity using a small number of parameters, and is emerging as a promising data compression technique. However, limited spectrum coverage is intrinsic to INR, and it is non-trivial to remove redundancy in diverse complex data effectively. Preliminary studies can only exploit either global or local correlation in the target data and thus of limited performance. In this paper, we propose a Tree-structured Implicit Neural Compression (TINC) to conduct compact representation for local regions and extract the shared features of these local representations in a hierarchical manner. Specifically, we use Multi-Layer Perceptrons (MLPs) to fit the partitioned local regions, and these MLPs are organized in tree structure to share parameters according to the spatial distance. The parameter sharing scheme not only ensures the continuity between adjacent regions, but also jointly removes the local and non-local redundancy. Extensive experiments show that TINC improves the compression fidelity of INR, and has shown impressive compression capabilities over commercial tools and other deep learning based methods. Besides, the approach is of high flexibility and can be tailored for different data and parameter settings. The source code can be found at https://github.com/RichealYoung/TINC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.21, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.47, + 0.856 + ], + "angle": 0, + "content": "In the big data era, there exist massive and continuously growing amount of visual data from different fields, from surveillance, entertainment, to biology and medical diagnosis. The urgent need for efficient data storage, sharing and transmission all requires effective compression techniques. Although image and video compression have been studied for decades and there are a number of widely used commercial tools, the compression of medical and biological data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This work was done in collaboration with Tingxiong Xiao and Yuxiao Cheng, under supervision of Prof. Jinli Suo and Prof. Qionghai Dai affiliated with Department of Automation, Tsinghua Univ." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.666, + 0.304 + ], + "angle": 0, + "content": "are not readily available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.413 + ], + "angle": 0, + "content": "Implicit neural representation (INR) is becoming widely-used for scene rendering [19, 21, 23, 31, 44], shape estimation [8, 11, 25, 29], and dynamics modeling [15, 26, 28]. Due to its powerful representation capability, INR can describe nature scenes at high fidelity with a much smaller number of parameters than raw discrete grid representation, and thus serves as a promising data compression technique." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.896, + 0.688 + ], + "angle": 0, + "content": "In spite of the big potential, INR based compression is quite limited confronted with large sized data [20], since INR is intrinsically of limited spectrum coverage and cannot envelop the spectrum of the target data, as analyzed in [41]. Two pioneering works using INR for data compression, including NeRV [7] and SCI [41], have attempted to handle this issue in their respective ways. Specifically, NeRV introduces the convolution operation into INR, which can reduce the required number of parameters using the weight sharing mechanism. However, convolution is spatially invariant and thus limits NeRV's representation accuracy on complex data with spatial varying feature distribution. Differently, to obtain high fidelity on complex data, SCI adopts divide-and-conquer strategy and partitions the data into blocks within INR's concentrated spectrum envelop. This improves the local fidelity, but cannot remove non-local redundancies for higher compression ratio and tend to cause blocking artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.897, + 0.903 + ], + "angle": 0, + "content": "To compress large and complex data with INR, in this paper we propose to build a tree-structured Multi-Layer Perceptrons (MLPs), which consists of a set of INRs to represent local regions in a compact manner and organizes them under a hierarchical architecture for parameter sharing and higher compression ratio. Specifically, we first draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress different regions with multiple MLPs separately. Then we incorporate these MLPs with a tree structure to extract their shared parameters hierarchically, following the rule that spatially closer regions are of higher similarity and share more parameters. Such a joint representation strategy can remove the redundancy both within each local region and among non-local" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "18517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "regions, and also ensures the continuity between adjacent regions. The scheme of the proposed network is illustrated in Fig. 1, and we name our approach TINC (Tree-structured Implicit Neural Compression)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Using the massive and diverse biomedical data, we conduct extensive experiments to validate that TINC greatly improves the capability of INR and even outperforms the commercial compression tools (H.264 and HEVC) under high compression ratios. The proposed TINC is also a general framework, which can be flexibly adapted to diverse data and varying settings." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.272, + 0.22, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.674 + ], + "angle": 0, + "content": "Data Compression. Data compression is a key technique for efficient data storage, transmission and sharing, especially in the big data era. In recent decades, many commercial compression tools for massive visual data have become increasingly mature, like JPEG [37], JPEG2000 [32], H.264 [39], and HEVC [33], while emerging deep learning-based approaches are also gaining momentum and demonstrating continuously improving compression capabilities [2, 3, 16, 22, 40, 42, 43]. Compression for nature images and videos have been widely and successfully used, however, there still lack effective compression techniques for the huge amount of three or higher dimensional data such as biomedical volume. Image/Video compressors cannot be applied directly, because these high dimensional data is a discrete sampling of 3D organism, instead of a video describing temporal evolution of 2D slices. Hence, the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured data [41], and tend to cause discontinuity and degenerate the downstream tasks. Our approach differs completely from existing image/video compression methodologies in that we encode the intrinsic data structure with INR. Besides, our approach searches for the best network structure and parameters representing the target data and is free of generalization issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Implicit Neural Compression. Recent works [7, 17, 41] have introduced INR to the field of data compression with exciting results. They fit the original discrete grid-based data with continuous implicit functions in a transform-coding-like manner, which takes advantage of the powerful representation capabilities of neural networks while avoids the generalization problem. Obviously, the compression capability of INR is closely related to its representation capability, which can be improved from proper design of network structure [7, 10, 17], selection of activation function [14, 20, 30], embedding [4, 21, 23, 36, 45], and learning strategies such as manifold learning [9, 20, 30, 31], meta learning [5, 9, 35] and ensemble learning [1, 12, 24]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "However, improving these capability does not necessarily head for a ready compressor for general scenes or ob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "jects, since there exist large diversity in the target data and pursuing a general network with high fidelity for all the data under given parameter budget. In order to pursue a more compact and efficient representation, we draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress the data in each region separately. On top of this, we improve INR by introducing a hierarchical parameter sharing mechanism in terms of data compression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.893, + 0.506 + ], + "angle": 0, + "content": "Parameter Sharing in INR. To improve the representation accuracy of INR for large and complex scenes, many algorithms use the parameter sharing strategy. One implementation is introducing convolution operations [6, 7, 27], which allows hierarchical extraction of local features, reduces the required parameters, and enables describing high resolution data. However, for those complex data with uneven distribution of features, the spatial invariance of convolution would lead to lower local representation accuracy. Another parameter sharing mechanism is positional encoding [21, 36], in which nearby spatial positions share partial mapping features hierarchically. However, this can only achieve higher representation accuracy when equipped with random or learnable mappings [4, 30, 36], which can be considered as an embedding that encodes only spatial distance. In addition, using embedding [18, 23, 34] intrinsically reduces the computational complexity at expenses of increased parameters, and thus not optimal for compression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.65 + ], + "angle": 0, + "content": "To make use of advantageous and bypass the above issues, we borrow the idea of hierarchical parameter sharing into local INRs, and build our compressor on a widely used tree structured data partitioning [18, 34, 41]. On the one hand, the parameter sharing is applied among block-specific INRs to cope with the large data with high local fidelity. On the other hand, we share parameters on INR rather than the data itself or its embedding, to encode richer information other than spatial distance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.685, + 0.89, + 0.718 + ], + "angle": 0, + "content": "3. Tree-structured Implicit Neural Compression" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.859, + 0.751 + ], + "angle": 0, + "content": "3.1. Ensemble of Implicit Neural Compressors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implicit neural compression uses a compact neural network to parameterize a function describing the target data defined over a spatial coordinates system. More specifically, for an \\(N\\) dimensional data, let \\(\\mathbb{V} \\coloneqq [-1,1]^N\\) denote its grid coordinates and \\(\\mathbf{d}(\\mathbf{v})\\) with \\(\\mathbf{v} \\in \\mathbb{V}\\) denote the intensity value at coordinate \\(\\mathbf{v}\\), one can learn a compact implicit function \\(f\\) with parameter \\(\\Theta\\) to estimate \\(\\mathbf{d}(\\mathbf{v})\\) by \\(f(\\mathbf{v};\\Theta)\\) and finally serialize them to achieve a compressed representation. Mathematically, we can formulate the compression" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.11, + 0.885, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.893, + 0.394 + ], + "angle": 0, + "content": "Figure 1. The scheme of the proposed approach TINC. For a target data, we divide the volume into equal-size blocks via octree partitioning, with some neighboring and far-apart blocks of similar appearances. Then each block can be representation with an implicit neural function, implemented as an MLP. After sharing parameters among similar blocks, we can achieve a more compact neural network with a tree shaped structure. Here we highlight the similar blocks sharing network parameters with the same color." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.42, + 0.291, + 0.433 + ], + "angle": 0, + "content": "task as an optimization problem" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.441, + 0.469, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\Theta}} \\int_ {\\mathbf {v} \\in \\mathbb {V}} \\mathcal {L} (f (\\mathbf {v}; \\boldsymbol {\\Theta}), \\mathbf {d} (\\mathbf {v})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.487, + 0.469, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) measures the difference between the decompressed and target data. For decompression, one can retrieve \\(f\\) and \\(\\Theta\\) from the compressed file and then calculate function values \\(f(\\mathbf{v};\\Theta)\\) over grid \\(\\mathbb{V}\\) via simple forward propagation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.548, + 0.469, + 0.713 + ], + "angle": 0, + "content": "Due to limited spectrum coverage, INR is incapable of compactly describe large and complex data [20,41], we borrow the idea of ensemble learning to partition the target volume into blocks and use multiple less expressive \\( f_{k}(\\cdot ,\\Theta_{k}) \\) to achieve a powerful representation. Specifically, we adopt the divide-and-conquer strategy as in [12, 20, 41] to ensemble all implicit functions \\( \\{f_k,k = 1,\\dots ,K\\} \\) that represents data at its corresponding coordinate regions \\( \\{\\mathbb{V}_k,k = 1,\\dots ,K\\} \\) and compose an \\( f \\), which is equivalent to approximating a complex function by a piece-wise function composed of \\( K \\) simple functions" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {v}; \\boldsymbol {\\Theta}) := \\sum_ {k = 1} ^ {K} \\mathbf {1} _ {\\mathbb {V} _ {k}} \\left(\\mathbf {v}, f _ {k} \\left(\\mathbf {v}, \\boldsymbol {\\Theta} _ {k}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.765, + 0.469, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathbf {1} _ {\\mathbb {V} _ {k}} (\\mathbf {v}, x) = \\left\\{ \\begin{array}{l} x, \\mathbf {v} \\in \\mathbb {V} _ {k} \\\\ 0, e l s e \\end{array} . \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In addition, we must constrain the total number of parameters to be consistent with the parameter budget, i.e., \\(|\\Theta| = \\sum_{k=1}^{K} |\\Theta_k|\\), where \\(|\\cdot|\\) is the support operation. Here we omit the parameters for serializing \\(f_k\\) which consists of only a few structural hyperparameters of neural network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.555 + ], + "angle": 0, + "content": "For a target data volume shown in the first column of Fig. 1, we use an \\(L\\) level tree structure containing \\(K\\) leaf nodes to organize \\(f_{k}\\) and \\(\\mathbb{V}_k\\), as illustrated in the 2nd column of Fig. 1. The root node represents the whole coordinate region \\(\\mathbb{V}\\) and each child node represents \\(\\frac{1}{2^N}\\) region of its parent node. We use \\(f_{k}\\) to compress the data in coordinate region \\(\\mathbb{V}_k\\) represented by the \\(k\\)th leaf node. This structure allows us to organize ensemble learning in high degree of flexibility." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.877, + 0.606 + ], + "angle": 0, + "content": "3.2. Hierarchical Parameter Sharing Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.804 + ], + "angle": 0, + "content": "We let these \\(\\{f_k\\}\\) share their neural network parameters hierarchically with each other according to the spatial distance between corresponding regions \\(\\mathbb{V}_k\\). The sharing mechanism is defined on the octree structure. Specifically, for a leaf node at level \\(l\\), its corresponding MLP implemented \\(f_k\\)'s hidden layers can be divided into \\(l\\) segments, i.e. \\(f_k = f_k^{out} \\circ f_k^l \\circ f_k^{l-1} \\circ \\dots \\circ f_k^1 \\circ f_k^{in}\\). For a set of leaf nodes, we determine the number of shared segments based on the number of their common ancestor nodes. For example, if \\(f_i\\) and \\(f_j\\) share the same ancestor nodes at \\(1 \\sim 3\\) levels, three pairs of hidden layer segments \\((f_i^1, f_j^1)\\), \\((f_i^2, f_j^2)\\), \\((f_i^3, f_j^3)\\) will share the same parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The sharing mechanism is illustrated in the 3rd column in Fig. 1. Since the distance between two leaf nodes can also be inferred from the number of their shared ancestor nodes, above hierarchical parameter sharing mechanism implies that closer blocks share more parameters and are of higher similarity in the implicit representation space." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18519" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.408, + 0.107 + ], + "angle": 0, + "content": "3.3. Tree-structured Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.468, + 0.206 + ], + "angle": 0, + "content": "We propose a tree-structured MLP based on the \\(L\\) level octree partitioning, as illustrated in the 4th column in Fig. 1. Each node contains a hyper layer consisting of some fully connected layers and takes the output of its parent node's hyper layer as input. Root node and leaf nodes additionally contain the input and output layers respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.468, + 0.327 + ], + "angle": 0, + "content": "Following the depth direction, from leaf nodes upward to the root of the tree, the output information of the leaf node is processed by the hyper layers in its ancestor nodes. If we consider the hyper layer as a part of hidden layers, then any leaf node can be considered as an MLP, whose hidden layers are explicitly divided into multiple segments contained in its ancestor nodes across different levels. In this way we can ensemble all \\(\\{f_k\\}\\) into a tree-structured MLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.468, + 0.448 + ], + "angle": 0, + "content": "At the same level, all sibling nodes share the same parent node and thus take the same information as input. Therefore, for any two leaf nodes sharing the same ancestor nodes, their output information have flown through the same hyper layers contained in the ancestor nodes. In other words, the MLPs of these two leaf nodes will share a part of hidden layers, which implements the hierarchical parameter sharing mechanism described in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.465, + 0.384, + 0.481 + ], + "angle": 0, + "content": "4. Cross Level Parameter Allocation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.468, + 0.55 + ], + "angle": 0, + "content": "After the network design, we are left with some key implementation issues: (i) how to set a proper number of levels to finalize the tree structured network, (ii) how to allocate available parameters across MLPs at the tree nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.468, + 0.763 + ], + "angle": 0, + "content": "The Number of Tree Levels \\( L \\). The number of tree levels \\( L \\) determines the number of divided segments of \\( f_{k} \\)'s hidden layers, i.e., a larger \\( L \\) implies the parameter sharing mechanism proceeds to higher spatial resolution. Setting a large \\( L \\) has both advantageous and disadvantages. On the one hand, a larger \\( L \\) means finer partitioning with a larger number of blocks sharing parameters (with the same hyper layers as in the ancestor nodes), which is beneficial for describing the regions with rich details and thus improve the compression fidelity. On the other hand, with the same total number of parameters, a larger \\( L \\) will lead to more leaf nodes \\( K \\) and less parameters \\( |\\Theta_k| \\) each \\( f_{k} \\), which might reduce \\( f_{k} \\)'s representation capability and might harm the compression fidelity in turn." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Therefore, we should set the tree level \\( L \\) by trading off between increased fineness in partitioning and parameter allocation, and \\( f_{k} \\) decreased representation capability. In Sec. 5.3, we experimentally prove that it is favourable using large tree levels for the data with rich details or under low compression ratios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Inter-level Parameter Allocation. The parameters at the shallow level describe the shared features at large scale, i.e., the non-local redundancy distributed over distant regions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "Similarly, from the perspective of data features, if some far apart sub-regions at a certain level are highly similar, one can allocate more parameters to their common ancestor nodes, and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Therefore, for data with high global redundancy or repetitiveness, i.e., high non-local similarity among sub-regions at large scales, allocating more parameters to the shallow level will be more beneficial to improve its compression fidelity. Conversely, for data with weak global similarity, it is more efficient to allocate more parameters to the nodes at deep levels for better representation of the unique local features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Intra-level Parameter Allocation. In addition to allocating parameters across levels, we can also allocate more parameters within a level to some nodes according to certain criteria instead of even allocation. For example, the most valuable information in neuronal data is often distributed in sparser regions. If we allocate more parameters to the more important regions, then we can sacrifice the unimportant regions in exchange for the improvement of compression fidelity of the important ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.623 + ], + "angle": 0, + "content": "A simple way to take advantage of the TINC's hierarchy and flexibility is to allocate the number of parameters per node at each level according to the importance of the corresponding regions. Following this allocation, at each spatial resolution, the nodes corresponding to more important regions will have more parameters compared to other nodes in the same level. We demonstrate in Sec. 5.3 that this simple approach can significantly improve the compression fidelity of biological data, especially in regions with very low compression fidelity in the even allocation case, if an appropriate measure of the importance of the data region is adopted. Obviously there are more complex and efficient allocation methods, but this preliminary exploration is sufficient to demonstrate the advantages of TINC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.634, + 0.657 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.717, + 0.681 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.9 + ], + "angle": 0, + "content": "All the hyper layers in each node are implemented using MLP, where the activation functions are all sine function and the parameters are initialized following the scheme proposed in [30]. All MLPs are optimized by Adamax [13] with initiate learning rate of 0.001 and decreased by a factor of 0.2 at the 2000th and 5000th iterations, respectively. We adopted a complete octtree, which means \\( K = 8^{(L - 1)} \\). Before optimization, all the coordinates and the intensities are normalized to \\([-1,1]\\) and \\([0,100]\\) respectively. After optimization, all the parameters of each hyper layers are serialized and stored in binary. There are also some information that need to be saved for data decompression, such as MLP structure, original data bit depth, image size, inverse normalization basic information. Because all of the items are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18520" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. The scores on each dataset as compression ratio ranges from 64 to 1024, with the #best and the ●second best marked by color and symbols. The metrics include PSNR(dB), SSIM, and Acc. \\(\\tau\\) with \\(\\tau\\) acts as a threshold for binarization. The “All” or “High” suffixes after the metrics name indicate that it is averaged for all compression ratios or only high compression ratios (around \\(1024 \\times\\))." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.142, + 0.885, + 0.32 + ], + "angle": 0, + "content": "
MethodMedical dataBiological data
PSNR All. (dB)SSIM All.PSNR High. (dB)SSIM High.Acc.200 All.Acc.500 All.Acc.200 High.Acc.500 High.
TINC (ours)•52.02•0.9897#50.59#0.9878•0.9945•0.9970#0.9934#0.9958
JPEG41.410.972230.490.93740.66120.98340.01970.9882
H.26451.180.989647.280.98600.99190.99590.98600.9926
HEVC#52.31#0.9903•50.51•0.9877#0.9955#0.99750.99170.9930
SCI51.900.989450.390.98760.99430.9965•0.9921•0.9951
NeRF50.930.987549.660.98630.99350.99620.99030.9940
NeRV47.110.985940.110.98000.98150.99010.97320.9867
DVC47.390.986545.740.98400.98270.99000.96920.9789
SGA+BB46.560.983643.020.98080.80380.98830.48170.9798
SSF46.250.980743.700.97730.72210.96030.77900.9542
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.336, + 0.287, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.448, + 0.197, + 0.459 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.336, + 0.485, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.448, + 0.395, + 0.459 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.336, + 0.682, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.448, + 0.593, + 0.459 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.336, + 0.88, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.448, + 0.792, + 0.459 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.472, + 0.868, + 0.486 + ], + "angle": 0, + "content": "Figure 2. Comparison of different compression methods on medical (a,b) and biological data (c,d) at different compression ratios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.55 + ], + "angle": 0, + "content": "discrete, countable, and have a limited set of possible values, we can easily create a table listing all potential values and store their indices, to reduce the file size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.471, + 0.717 + ], + "angle": 0, + "content": "We evaluated our approach on both massive medical and biological data. For the former we used HiP-CT dataset [38], which provides cellular level imaging of several organisms at multiple anatomical levels. We used four human organs (Lung, Heart, Kidney and Brain) from this dataset. For the latter we used the sub-micrometer resolution mouse whole-brain microscopic data (Neurons and Vessels) captured by our self-developed imaging system (see supplementary note 1 for details), which has varying brightness and structures across different regions. The access to these data are listed in supplementary note 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.763 + ], + "angle": 0, + "content": "For convenient and fair comparison, we cropped the data into the same size (\\(256 \\times 256 \\times 256\\) voxels for medical data and \\(64 \\times 512 \\times 512\\) voxels for biological data)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.47, + 0.801 + ], + "angle": 0, + "content": "5.2. Performance Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Benchmark Methods and Evaluation Metrics. The proposed method is comprehensively compared with state-of-the-art methods, which can be classified into three groups: (i) commercial compression tools, including JPEG [37], H.264 [39], and HEVC [33]; (ii) recently proposed data-driven deep learning based ones, including DVC [16]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.657 + ], + "angle": 0, + "content": "SGA+BB [43], and SSF [2]; (iii) recently proposed INR based ones, including SCI [41], NeRV [7], and NeRF [21]. We evaluated all the methods at 6 different compression ratios ranging from \\(64 \\times\\) to \\(1024 \\times\\). It is worth noting that only INR based methods can precisely regulate the compression ratio. We adopted PSNR and SSIM for medical data including rich details, while Accuracy with threshold for binarization as evaluation metrics for biological data with sparse structures. For convenience we abbreviate Accuracy with threshold as \"Acc.threshold\" in the following descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.901 + ], + "angle": 0, + "content": "JPEG was tested on OpenCV. H.264 and HEVC were tested with their FFmpeg implementations. For SGA+BB, we adopted the settings with the best performance mentioned in the original paper. For DVC, we used its Pytorch implementation, and compressed the \"I Frames\" using JPEG method with quality parameter set to 50. For SSF, we used its CompressAI's implementation. In addition, for the data-driven methods, i.e. DVC, SGA+BB, and SSF, we fine-tuned the provided pre-trained models on 300 volumes randomly selected from the dataset. For SCI, we set the hyper parameter \\( a_{max} = 32 \\). For NeRF, the hyper parameter \"frequency\" was set to 10. For NeRV, we control the compression ratio by adjusting the number and width of the fully connected layers. For 2D images compression methods, i.e. JPEG and SGA+BB, we firstly segmented the 3D data int 2D images before compression." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18521" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.257 + ], + "angle": 0, + "content": "Quantitative Performance Evaluation and Analysis. The average performance of each method under different compression ratios is plotted as a rate-distortion curve in Fig. 2. We can see that our algorithm outperforms almost all methods at high compression ratios and is comparable to HEVC at other compression ratios. For more detailed analysis, we summarize the scores on each dataset in Table 1. It should be noted that even if the PSNR of the medical data are higher than \\(40\\mathrm{dB}\\), it does not mean that there are no distortion in the decompressed data, which is different from the natural scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Some visual comparisons of decompressed organ slices are presented in Fig. 3 (see supplementary figure 1 for more results). For better visual comparison, we applied the same post-processing to the decompressed data. Under compression ratio around \\(87 \\times\\) (since HEVC and SGA+BB cannot precisely regulate compression ratio), we consistently outperform existing INR based (SCI and NeRV), data-driven (SGA+BB) and commercial (HEVC) compression methods. Although SCI achieved comparable results in the regions away from boundary, as shown in Fig. 3(a), it produced severe blocking artifacts at the boundary, as shown by the arrows in (b), which will affect the medical research as well as other downstream tasks. NeRV retained the general structures but suffered from the checkerboard artifacts caused by the aliasing in convolution. HEVC captured high fidelity details in \"I Frames\" but suffered from the artifacts in \"P Frames\", which distorted the edges in the volumes, such as the white matters in Brain data and muscles in Heart data, which would be unusable for medical research and other downstream tasks. SGA+BB struggled with retaining the rich details, since it cannot represent the similarity within the 3D volume as a 2D image compression algorithm. In addition, HEVC produced ghosting artifacts and position offset in neighbor slices, since the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured biomedical data. In contrast, our approach use INR to encode the intrinsic data structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Comparison of Running Time. We summarized the time required for compression and decompression for each algorithm in our experiments, as shown in the supplementary tables 1 and 2. Comparatively, INR based compressor (TINC, SCI, NeRV, and NeRF) is slower in compression stage but can decompress faster. It is worth noting that biomedical data need to be compressed only once before saving or transmission, but decompressed frequently for processing and analysis. Therefore, TINC is advantageous in biomedical fields." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.4, + 0.863 + ], + "angle": 0, + "content": "5.3. Flexibility Settings for Different Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Setting of Tree Levels \\(L\\). We evaluated TINC at different tree levels \\(L\\) on medical data under different com" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.888, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.236, + 0.822, + 0.247 + ], + "angle": 0, + "content": "(a) Three ROIs from Brain data; compression ratio: \\(\\sim 87\\times\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.249, + 0.888, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.394, + 0.824, + 0.404 + ], + "angle": 0, + "content": "(b) Three ROIs from Heart data; compression ratio: \\(\\sim 87\\times\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.407, + 0.878, + 0.418 + ], + "angle": 0, + "content": "Ground Truth TINC (ours) SCI NeRV HEVC SGA+BB" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons of decompressed organs slices from Brain and Heart data between TINC and other benchmark methods under similar compression ratio around \\(87\\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.658 + ], + "angle": 0, + "content": "pression ratios. As shown in Fig. 4(a), the performance of TINC was evaluated in terms of SSIM on 8 medical data, under \\(256 \\times\\) compression ratio. To better demonstrate the effect of increasing tree levels, we used the growth rate of SSIM, calculated by normalizing all SSIM values by the one when \\(L = 1\\). The SSIM scores on all the test data have a steadily increasing growth rate (dashed lines) as the number of tree depth increases from 1 to 3. Experimental results on other compression ratios show consistent trend (supplementary figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We further investigated the relationship between the data complexity and the performance gain when increasing tree levels. For data complexity, we adopt the spectrum concentration defined in [41], which measures the percentage of high frequency information contained in the data. The results are plotted in Fig. 4(b), with the solid line being a quadratic fitting of the growth rate for the 8 medical data using least squares. From the plot one can see that the performance gain increases monotonously with data complexity. In other words, the benefit of using a large tree is more significant when the target data contains more high-frequency information. This trend is consistent with our reasoning in Sec. 4 that a large \\( L \\) favours parameter sharing at coarser levels and leave more parameters for representing the details. Experimental results on other compression ratios and other data are consistent with this (supplementary figure 3)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18522" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.257, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.232, + 0.18, + 0.242 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.093, + 0.465, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.231, + 0.388, + 0.243 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Effect of tree levels \\( L \\) on TINC's compression fidelity, in terms of SSIM on 8 medical data under \\( 256 \\times \\) compression ratio. (a) The growth rate of SSIM for each data when increasing the tree levels. The dashed lines represent the change in each data. The solid line represents the mean of changes, and the filled area represents the standard deviation. (b) The scatter plot of each data's complexity and the growth rate of SSIM when increasing tree levels from 1 to 3. The solid line represents the trend of change, estimated by a quadratic fit." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.469, + 0.587 + ], + "angle": 0, + "content": "In addition, we studied the performance gain of using a deep tree structure under limited bandwidth budget \\( |\\Theta| \\), e.g., at high compression ratio. Supplementary figure 4 plots the performance improvement at increasing compression ratios, calculated from 8 medical data volumes. Almost all data have a smaller growth rate of SSIM or PSNR when compression ratio increases, i.e. reducing the number of available parameters. At high compression ratio \\( (512 \\times) \\), some data are even of inferior fidelity using a deep tree, which is in line with our analysis in Sec. 4 that one should set a proper \\( L \\) for choosing the right partitioning granularity under the constraint from available budget." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Inter-level Parameter Allocation. We tested the performance at different parameter allocation proportions across levels on medical data under different compression ratios. Assuming there are \\( K^l \\) nodes at \\( l \\)th level, we denote the number of allocated parameters at this level as \\( |\\Theta^l| \\) and ratio with respect to that of its adjacent upper level as \\( r^l = \\frac{|\\Theta^l| / K^l}{|\\Theta^{l-1}| / K^{l-1}} \\). Simply speaking, \\( r^l = 1 \\) means even inter-level allocation, a \\( r^l \\) larger than 1 means imposing higher priority to the shallow levels and vice versa. We let \\( r^l \\) of all levels take 0.8, 1 and 1.2 to represent the three typical cases of parameter allocation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We first investigated the effect of allocating more parameters to the shallow level, intending to share parameters among far apart similar blocks. As shown in Fig. 5(a), we divided a Brain data into 64 sub-regions, laid out in z-curve order, and selected 8 equally spaced distant regions. Then, we calculated the SSIM of these regions with respect to each other and used min-max normalization to obtain their non-local similarities. It can be seen that #1 and #33 regions are of high similarity. For convenient display, we also draw the" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.093, + 0.663, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.232, + 0.602, + 0.242 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.096, + 0.787, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.232, + 0.75, + 0.242 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.096, + 0.885, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.231, + 0.854, + 0.242 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Figure 5. Effect of allocating more parameters to the shallow level on TINC's compression fidelity for each sub-regions in a Brain data under \\(512 \\times\\) compression ratio. All sub-figures share the same y-axis label. (a) The heatmap of similarities between 8 equally spaced distant sub-regions. The serial numbers of the regions represent their z-curve order. (b) The non-local similarity of each region. (c) The growth rate of SSIM for each region when allocating more parameters to the shallow level." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.447 + ], + "angle": 0, + "content": "bar plot of the normalized marginal sums of the similarity matrix in Fig. 5(b), which tells that #1 and #33 are both of high similarity to other blocks, while #25 and #57 regions are highly distinctive from others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.644 + ], + "angle": 0, + "content": "To quantitatively measure the performance change at different blocks when allocating more parameters to the shallow level, we plot the growth rate of SSIM under \\(512 \\times\\) ratio, as shown in Fig. 5(c). Analyzing jointly with (b), we can find that the two regions with high non-local similarity, #1 and #33, gain a large boost, while those with the lowest similarity, #25 and #57, suffer a loss of fidelity. This is consistent with the intuition and analysis in Sec. 4 that allocating more parameters to the nodes at shallow levels would benefit representing distant but similar regions, but at expense of slight degeneration at very individual regions. Experimental results on other compression ratios and other data are shown in supplementary figures 5 and 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We then investigated how to allocate parameters across levels according to the global consistency of the medical data. We divided the target data into 64 sub-regions, calculated the marginal sums of the similarity matrix. We change the setting of \\( r^l \\) and the results are shown in the supplementary figure 7. The results show allocating more parameters to the shallow level is more likely to improve fidelity when the global consistency of the data is greater than 0.7, and allocating more to the deep level is better when the global consistency lower than 0.6. Otherwise, an even allocation may be a better choice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Intra-level Parameter Allocation. We tested the approach of allocating parameters to the nodes at the same level according to their importance proposed in Sec. 4, on biological data and under different compression ratios. Specifically, we let the number of allocated parameters be proportional to their importance. For example, with regard" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18523" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "to the biological data such as Neurons the importance can be measured by the percentage of the number of valid signals, where the valid signal can be determined either by the result of traced neurons or by the intensity value. Here we chose the latter for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.243 + ], + "angle": 0, + "content": "We first compare the compression fidelity of 6 Neurons data using the above two parameter allocation approaches under different compression ratios, as shown in supplementary figure 8. It can be seen that important allocation significantly improves the fidelity compared to even allocation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.529 + ], + "angle": 0, + "content": "We then look further into the performance at all the sub-regions in the same target data. As shown in Fig. 6(a), we divided a Neurons data into 64 sub-regions and compute their growth rate of Acc.2000 when switching from even allocation to important allocation. It should be noted that we omitted those regions that have reached 1 (the beast value) under both allocation approaches in the subsequent experiments. From the scatter diagram we can see three groups of data: (i) Some regions show non-significant change (less than \\(10\\%\\) of the maximum change of the others, denoted as cross markers) using important allocation under \\(64\\times\\) compression ratio, because they are already with sufficient parameters. (iii) Other regions show significant improvement after introducing important allocation, denoted with circles. We performed a linear fit using least squares, as shown by the solid line. Overall, allocating parameters according to importance is helpful for the compression fidelity of important regions while at expenses of little fidelity degeneration in other less important ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.468, + 0.695 + ], + "angle": 0, + "content": "For a statistical analysis, we plotted the Acc.2000 of all 64 sub-regions under even and important allocation, as shown in Fig. 6(b). We found that importance based allocation can significantly improve those regions that have very poor compression fidelity under even allocation. This is mainly because that those regions with poor fidelity tend to contain more complex structures (with more neural signals), and thus are of higher importance scores and thus allocated more parameters accordingly. Experimental results on other compression ratios are detailed in supplementary figure 9." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.205, + 0.725 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Summary and Discussions. In this paper we propose a hierarchical parameter sharing mechanism for INR and implement it based on a tree-structured MLP for large and complex data compression tasks. The proposed approach TINC makes use of the local redundancy via block wise compression with separate MLPs and non-local redundancy via organizing these MLPs in tree structure to share parameters hierarchically. Such implementation forms a single network for the complex target data, not only incorporate the similarity distributed far apart and nonuniformly, and avoid discontinuity between adjacent regions as well." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.682, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.232, + 0.602, + 0.242 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.092, + 0.891, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.231, + 0.81, + 0.242 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Figure 6. Effect of important allocation on TINC's compression fidelity, in terms of Acc.2000 on a Neurons data under \\(64 \\times\\) compression ratio. (a) The growth rate of Acc.2000 for each region. The regions with and without significant growth rate are represented as cross and square respectively. The solid line represents the trend of change, estimated by a linear fit. (b) The violin plots of Acc.2000 for all these regions. The shaded area represents the distribution of Acc.2000 across all regions. The three horizontal lines from top to bottom represent the maximum, average and minimum values respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Extensive experiments show that this hierarchical parameter sharing mechanism introduced in TINC greatly improves the compression capability and wide applicability of INR. Our method outperforms state-of-the-art compressors including commercial tools (JPEG, H.264, HEVC), data-driven methods (DVC, SSF, SGA+BB), and existing INR based ones (SCI, NeRV, and NeRF) on diverse medical and biological data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.631 + ], + "angle": 0, + "content": "We also analyze TINC's flexibility to different cases via experimentally studying the effect of three key settings—number of tree levels, intra-level and inter-level parameter allocation methods, and provide highlights on proper parameter setting and strategy selection according to the features of target data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Limitations and Future Extensions. Similar to all current INR based compression methods, TINC is of high decompression speed but slow in compression, since it takes time to pursue the MLPs matching the target data. We plan to combine meta-learning to find the best initialization parameters for each organ to speed up TINC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.744, + 0.667, + 0.76 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.891, + 0.816 + ], + "angle": 0, + "content": "We acknowledge MoE Key Laboratory of Biomedical Photonics (Huazhong University of Science and Technology) for sharing their mouse brain-wide microscopic data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is supported by Beijing Natural Science Foundation (Grant No. Z200021), the National Natural Science Foundation of China (Grant Nos. 61931012, 62088102) and Project of Medical Engineering Laboratory of Chinese PLA General Hospital (Grant No. 2022SYSZ-ZKY21)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18524" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Arya Aftab, Alireza Morsali, and Shahrokh Ghaemmaghami. Multi-Head Relu Implicit Neural Representation Networks. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2510-2514, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational Image Compression with a Scale Hyperprior. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[4] Nuri Benbarka, Timon Höfer, Hamd ul-moqeet Riaz, and Andreas Zell. Seeing Implicit Neural Representations as Fourier Series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2283-2292, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Alexander William Bergman, Petr Kellnhofer, and Gordon Wetzstein. Fast Training of Neural Lumigraph Representations using Meta Learning. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.472, + 0.479 + ], + "angle": 0, + "content": "[6] Alexandre Boulch and Renaud Marlet. Poco: Point Convolution for Surface Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. NeRV: Neural Representations for Videos. Advances in Neural Information Processing Systems, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[9] Yilun Du, Katherine Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning Signal-agnostic Manifolds of Neural Fields. Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[10] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J. Zico Kolter. Multiplicative Filter Networks. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[11] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning Shape Templates with Structured Implicit Functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7154-7164, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[12] Milad Soltany Kadarvish, Hesam Mojtahedi, Hossein Entezari Zarch, Amirhossein Kazerouni, Alireza Morsali, Azra Abtahi, and Farokh Marvasti. Ensemble Neural Representation Networks. arXiv preprint arXiv:2110.04124, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.83, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Sylwester Klocek, Lukasz Maziarka, Maciej Wolczyk, Jacek Tabor, Jakub Nowak, and Marek Śmieja. Hypernetwork" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Functional Image Representation. In International Conference on Artificial Neural Networks, pages 496-510, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.204, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[16] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: An End-to-end Deep Video Compression Framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[17] Y. Lu, K. Jiang, J. A. Levine, and M. Berger. Compressive Neural Representations of Volumetric Scalar Fields. Computer Graphics Forum, 40(3):135-146, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[18] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive Coordinate Networks for Neural Scene Representation. ACM Trans. Graph., 40(4), jul 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[20] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated Periodic Activations for Generalizable Local Functional Representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, October 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] David Minnen, Johannes Balle, and George D Toderici. Joint Autoregressive and Hierarchical Priors for Learned Image Compression. In Advances in Neural Information Processing Systems, volume 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph., 41(4), jul 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing Scenes as Compositional Generative Neural Feature Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18525" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[27] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional Occupancy Networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned Implicit Function for High-resolution Clothed Human Digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[30] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit Neural Representations with Periodic Activation Functions. In Advances in Neural Information Processing Systems, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3d-structure-aware Neural Scene Representations. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[32] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The JPEG 2000 Still Image Compression Standard. IEEE Signal Processing Magazine, 18(5):36-58, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[33] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the High Efficiency Video Coding (HEVC) Standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[34] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[35] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned Initializations for Optimizing Coordinate-based Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[37] Gregory K Wallace. The JPEG Still Picture Compression Standard. IEEE Transactions on Consumer Electronics, 38(1):xviii-xxxiv, 1992. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[38] CL Walsh, P Tafforeau, WL Wagner, DJ Jafree, A Bellier, C Werlein, MP Kühnel, E Boller, S Walker-Samuel, JL Rober" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "tus, et al. Imaging Intact Human Organs with Local Resolution of Cellular Structures using Hierarchical Phase-Contrast Tomography. Nature Methods, 18(12):1532-1541, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the H.264/AVC Video Coding Standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[40] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for Video Compression with Hierarchical Quality and Recurrent Enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6628-6637, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[41] Runzhao Yang, Tingxiong Xiao, Yuxiao Cheng, Qianni Cao, Jinyuan Qu, Jinli Suo, and Qionghai Dai. SCI: A Spectrum Concentrated Implicit Neural Compression for Biomedical Data. arXiv preprint arXiv:2209.15180, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Ruihan Yang, Yibo Yang, Joseph Marino, and Stephan Mandt. Hierarchical Autoregressive Modeling for Neural Video Compression. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[43] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving Inference for Neural Image Compression. In Advances in Neural Information Processing Systems, volume 33, pages 573-584, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields from One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[45] Ellen D Zhong, Tristan Bepler, Joseph H Davis, and Bonnie Berger. Reconstructing Continuous Distributions of 3D Protein Structure from Cryo-EM Images. arXiv preprint arXiv:1909.05215, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18526" + } + ] +] \ No newline at end of file diff --git a/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_origin.pdf b/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..060ef60296985671e291c7455a2d816d8893d9dc --- /dev/null +++ b/2023/TINC_ Tree-Structured Implicit Neural Compression/8bbad193-1acf-4b1e-bca7-0998f55872d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b828c60440dc615153979ec67b68ed87b5da4d5522f674b60f8e0ac7814b353b +size 819963 diff --git a/2023/TINC_ Tree-Structured Implicit Neural Compression/full.md b/2023/TINC_ Tree-Structured Implicit Neural Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1e8472c4cb80c17d4a8fe51e2720b8d450bf6b4b --- /dev/null +++ b/2023/TINC_ Tree-Structured Implicit Neural Compression/full.md @@ -0,0 +1,280 @@ +# TINC: Tree-structured Implicit Neural Compression + +Runzhao Yang +Department of Automation, Tsinghua University +Beijing 100084, China +yangrz20@mails.tsinghua.edu.cn + +# Abstract + +Implicit neural representation (INR) can describe the target scenes with high fidelity using a small number of parameters, and is emerging as a promising data compression technique. However, limited spectrum coverage is intrinsic to INR, and it is non-trivial to remove redundancy in diverse complex data effectively. Preliminary studies can only exploit either global or local correlation in the target data and thus of limited performance. In this paper, we propose a Tree-structured Implicit Neural Compression (TINC) to conduct compact representation for local regions and extract the shared features of these local representations in a hierarchical manner. Specifically, we use Multi-Layer Perceptrons (MLPs) to fit the partitioned local regions, and these MLPs are organized in tree structure to share parameters according to the spatial distance. The parameter sharing scheme not only ensures the continuity between adjacent regions, but also jointly removes the local and non-local redundancy. Extensive experiments show that TINC improves the compression fidelity of INR, and has shown impressive compression capabilities over commercial tools and other deep learning based methods. Besides, the approach is of high flexibility and can be tailored for different data and parameter settings. The source code can be found at https://github.com/RichealYoung/TINC. + +# 1. Introduction + +In the big data era, there exist massive and continuously growing amount of visual data from different fields, from surveillance, entertainment, to biology and medical diagnosis. The urgent need for efficient data storage, sharing and transmission all requires effective compression techniques. Although image and video compression have been studied for decades and there are a number of widely used commercial tools, the compression of medical and biological data + +This work was done in collaboration with Tingxiong Xiao and Yuxiao Cheng, under supervision of Prof. Jinli Suo and Prof. Qionghai Dai affiliated with Department of Automation, Tsinghua Univ. + +are not readily available. + +Implicit neural representation (INR) is becoming widely-used for scene rendering [19, 21, 23, 31, 44], shape estimation [8, 11, 25, 29], and dynamics modeling [15, 26, 28]. Due to its powerful representation capability, INR can describe nature scenes at high fidelity with a much smaller number of parameters than raw discrete grid representation, and thus serves as a promising data compression technique. + +In spite of the big potential, INR based compression is quite limited confronted with large sized data [20], since INR is intrinsically of limited spectrum coverage and cannot envelop the spectrum of the target data, as analyzed in [41]. Two pioneering works using INR for data compression, including NeRV [7] and SCI [41], have attempted to handle this issue in their respective ways. Specifically, NeRV introduces the convolution operation into INR, which can reduce the required number of parameters using the weight sharing mechanism. However, convolution is spatially invariant and thus limits NeRV's representation accuracy on complex data with spatial varying feature distribution. Differently, to obtain high fidelity on complex data, SCI adopts divide-and-conquer strategy and partitions the data into blocks within INR's concentrated spectrum envelop. This improves the local fidelity, but cannot remove non-local redundancies for higher compression ratio and tend to cause blocking artifacts. + +To compress large and complex data with INR, in this paper we propose to build a tree-structured Multi-Layer Perceptrons (MLPs), which consists of a set of INRs to represent local regions in a compact manner and organizes them under a hierarchical architecture for parameter sharing and higher compression ratio. Specifically, we first draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress different regions with multiple MLPs separately. Then we incorporate these MLPs with a tree structure to extract their shared parameters hierarchically, following the rule that spatially closer regions are of higher similarity and share more parameters. Such a joint representation strategy can remove the redundancy both within each local region and among non-local + +regions, and also ensures the continuity between adjacent regions. The scheme of the proposed network is illustrated in Fig. 1, and we name our approach TINC (Tree-structured Implicit Neural Compression). + +Using the massive and diverse biomedical data, we conduct extensive experiments to validate that TINC greatly improves the capability of INR and even outperforms the commercial compression tools (H.264 and HEVC) under high compression ratios. The proposed TINC is also a general framework, which can be flexibly adapted to diverse data and varying settings. + +# 2. Related Work + +Data Compression. Data compression is a key technique for efficient data storage, transmission and sharing, especially in the big data era. In recent decades, many commercial compression tools for massive visual data have become increasingly mature, like JPEG [37], JPEG2000 [32], H.264 [39], and HEVC [33], while emerging deep learning-based approaches are also gaining momentum and demonstrating continuously improving compression capabilities [2, 3, 16, 22, 40, 42, 43]. Compression for nature images and videos have been widely and successfully used, however, there still lack effective compression techniques for the huge amount of three or higher dimensional data such as biomedical volume. Image/Video compressors cannot be applied directly, because these high dimensional data is a discrete sampling of 3D organism, instead of a video describing temporal evolution of 2D slices. Hence, the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured data [41], and tend to cause discontinuity and degenerate the downstream tasks. Our approach differs completely from existing image/video compression methodologies in that we encode the intrinsic data structure with INR. Besides, our approach searches for the best network structure and parameters representing the target data and is free of generalization issues. + +Implicit Neural Compression. Recent works [7, 17, 41] have introduced INR to the field of data compression with exciting results. They fit the original discrete grid-based data with continuous implicit functions in a transform-coding-like manner, which takes advantage of the powerful representation capabilities of neural networks while avoids the generalization problem. Obviously, the compression capability of INR is closely related to its representation capability, which can be improved from proper design of network structure [7, 10, 17], selection of activation function [14, 20, 30], embedding [4, 21, 23, 36, 45], and learning strategies such as manifold learning [9, 20, 30, 31], meta learning [5, 9, 35] and ensemble learning [1, 12, 24]. + +However, improving these capability does not necessarily head for a ready compressor for general scenes or ob + +jects, since there exist large diversity in the target data and pursuing a general network with high fidelity for all the data under given parameter budget. In order to pursue a more compact and efficient representation, we draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress the data in each region separately. On top of this, we improve INR by introducing a hierarchical parameter sharing mechanism in terms of data compression. + +Parameter Sharing in INR. To improve the representation accuracy of INR for large and complex scenes, many algorithms use the parameter sharing strategy. One implementation is introducing convolution operations [6, 7, 27], which allows hierarchical extraction of local features, reduces the required parameters, and enables describing high resolution data. However, for those complex data with uneven distribution of features, the spatial invariance of convolution would lead to lower local representation accuracy. Another parameter sharing mechanism is positional encoding [21, 36], in which nearby spatial positions share partial mapping features hierarchically. However, this can only achieve higher representation accuracy when equipped with random or learnable mappings [4, 30, 36], which can be considered as an embedding that encodes only spatial distance. In addition, using embedding [18, 23, 34] intrinsically reduces the computational complexity at expenses of increased parameters, and thus not optimal for compression. + +To make use of advantageous and bypass the above issues, we borrow the idea of hierarchical parameter sharing into local INRs, and build our compressor on a widely used tree structured data partitioning [18, 34, 41]. On the one hand, the parameter sharing is applied among block-specific INRs to cope with the large data with high local fidelity. On the other hand, we share parameters on INR rather than the data itself or its embedding, to encode richer information other than spatial distance. + +# 3. Tree-structured Implicit Neural Compression + +# 3.1. Ensemble of Implicit Neural Compressors + +Implicit neural compression uses a compact neural network to parameterize a function describing the target data defined over a spatial coordinates system. More specifically, for an $N$ dimensional data, let $\mathbb{V} \coloneqq [-1,1]^N$ denote its grid coordinates and $\mathbf{d}(\mathbf{v})$ with $\mathbf{v} \in \mathbb{V}$ denote the intensity value at coordinate $\mathbf{v}$ , one can learn a compact implicit function $f$ with parameter $\Theta$ to estimate $\mathbf{d}(\mathbf{v})$ by $f(\mathbf{v};\Theta)$ and finally serialize them to achieve a compressed representation. Mathematically, we can formulate the compression + +![](images/8f688fdac98302532d50afa6477736159751ed4539cb8c6909d20a5aeef44ccc.jpg) +Figure 1. The scheme of the proposed approach TINC. For a target data, we divide the volume into equal-size blocks via octree partitioning, with some neighboring and far-apart blocks of similar appearances. Then each block can be representation with an implicit neural function, implemented as an MLP. After sharing parameters among similar blocks, we can achieve a more compact neural network with a tree shaped structure. Here we highlight the similar blocks sharing network parameters with the same color. + +task as an optimization problem + +$$ +\min _ {\boldsymbol {\Theta}} \int_ {\mathbf {v} \in \mathbb {V}} \mathcal {L} (f (\mathbf {v}; \boldsymbol {\Theta}), \mathbf {d} (\mathbf {v})), \tag {1} +$$ + +where $\mathcal{L}$ measures the difference between the decompressed and target data. For decompression, one can retrieve $f$ and $\Theta$ from the compressed file and then calculate function values $f(\mathbf{v};\Theta)$ over grid $\mathbb{V}$ via simple forward propagation. + +Due to limited spectrum coverage, INR is incapable of compactly describe large and complex data [20,41], we borrow the idea of ensemble learning to partition the target volume into blocks and use multiple less expressive $f_{k}(\cdot ,\Theta_{k})$ to achieve a powerful representation. Specifically, we adopt the divide-and-conquer strategy as in [12, 20, 41] to ensemble all implicit functions $\{f_k,k = 1,\dots ,K\}$ that represents data at its corresponding coordinate regions $\{\mathbb{V}_k,k = 1,\dots ,K\}$ and compose an $f$ , which is equivalent to approximating a complex function by a piece-wise function composed of $K$ simple functions + +$$ +f (\mathbf {v}; \boldsymbol {\Theta}) := \sum_ {k = 1} ^ {K} \mathbf {1} _ {\mathbb {V} _ {k}} \left(\mathbf {v}, f _ {k} \left(\mathbf {v}, \boldsymbol {\Theta} _ {k}\right)\right), \tag {2} +$$ + +$$ +\mathbf {1} _ {\mathbb {V} _ {k}} (\mathbf {v}, x) = \left\{ \begin{array}{l} x, \mathbf {v} \in \mathbb {V} _ {k} \\ 0, e l s e \end{array} . \right. \tag {3} +$$ + +In addition, we must constrain the total number of parameters to be consistent with the parameter budget, i.e., $|\Theta| = \sum_{k=1}^{K} |\Theta_k|$ , where $|\cdot|$ is the support operation. Here we omit the parameters for serializing $f_k$ which consists of only a few structural hyperparameters of neural network. + +For a target data volume shown in the first column of Fig. 1, we use an $L$ level tree structure containing $K$ leaf nodes to organize $f_{k}$ and $\mathbb{V}_k$ , as illustrated in the 2nd column of Fig. 1. The root node represents the whole coordinate region $\mathbb{V}$ and each child node represents $\frac{1}{2^N}$ region of its parent node. We use $f_{k}$ to compress the data in coordinate region $\mathbb{V}_k$ represented by the $k$ th leaf node. This structure allows us to organize ensemble learning in high degree of flexibility. + +# 3.2. Hierarchical Parameter Sharing Mechanism + +We let these $\{f_k\}$ share their neural network parameters hierarchically with each other according to the spatial distance between corresponding regions $\mathbb{V}_k$ . The sharing mechanism is defined on the octree structure. Specifically, for a leaf node at level $l$ , its corresponding MLP implemented $f_k$ 's hidden layers can be divided into $l$ segments, i.e. $f_k = f_k^{out} \circ f_k^l \circ f_k^{l-1} \circ \dots \circ f_k^1 \circ f_k^{in}$ . For a set of leaf nodes, we determine the number of shared segments based on the number of their common ancestor nodes. For example, if $f_i$ and $f_j$ share the same ancestor nodes at $1 \sim 3$ levels, three pairs of hidden layer segments $(f_i^1, f_j^1)$ , $(f_i^2, f_j^2)$ , $(f_i^3, f_j^3)$ will share the same parameters. + +The sharing mechanism is illustrated in the 3rd column in Fig. 1. Since the distance between two leaf nodes can also be inferred from the number of their shared ancestor nodes, above hierarchical parameter sharing mechanism implies that closer blocks share more parameters and are of higher similarity in the implicit representation space. + +# 3.3. Tree-structured Network Architecture + +We propose a tree-structured MLP based on the $L$ level octree partitioning, as illustrated in the 4th column in Fig. 1. Each node contains a hyper layer consisting of some fully connected layers and takes the output of its parent node's hyper layer as input. Root node and leaf nodes additionally contain the input and output layers respectively. + +Following the depth direction, from leaf nodes upward to the root of the tree, the output information of the leaf node is processed by the hyper layers in its ancestor nodes. If we consider the hyper layer as a part of hidden layers, then any leaf node can be considered as an MLP, whose hidden layers are explicitly divided into multiple segments contained in its ancestor nodes across different levels. In this way we can ensemble all $\{f_k\}$ into a tree-structured MLP. + +At the same level, all sibling nodes share the same parent node and thus take the same information as input. Therefore, for any two leaf nodes sharing the same ancestor nodes, their output information have flown through the same hyper layers contained in the ancestor nodes. In other words, the MLPs of these two leaf nodes will share a part of hidden layers, which implements the hierarchical parameter sharing mechanism described in Sec. 3.2. + +# 4. Cross Level Parameter Allocation + +After the network design, we are left with some key implementation issues: (i) how to set a proper number of levels to finalize the tree structured network, (ii) how to allocate available parameters across MLPs at the tree nodes. + +The Number of Tree Levels $L$ . The number of tree levels $L$ determines the number of divided segments of $f_{k}$ 's hidden layers, i.e., a larger $L$ implies the parameter sharing mechanism proceeds to higher spatial resolution. Setting a large $L$ has both advantageous and disadvantages. On the one hand, a larger $L$ means finer partitioning with a larger number of blocks sharing parameters (with the same hyper layers as in the ancestor nodes), which is beneficial for describing the regions with rich details and thus improve the compression fidelity. On the other hand, with the same total number of parameters, a larger $L$ will lead to more leaf nodes $K$ and less parameters $|\Theta_k|$ each $f_{k}$ , which might reduce $f_{k}$ 's representation capability and might harm the compression fidelity in turn. + +Therefore, we should set the tree level $L$ by trading off between increased fineness in partitioning and parameter allocation, and $f_{k}$ decreased representation capability. In Sec. 5.3, we experimentally prove that it is favourable using large tree levels for the data with rich details or under low compression ratios. + +Inter-level Parameter Allocation. The parameters at the shallow level describe the shared features at large scale, i.e., the non-local redundancy distributed over distant regions. + +Similarly, from the perspective of data features, if some far apart sub-regions at a certain level are highly similar, one can allocate more parameters to their common ancestor nodes, and vice versa. + +Therefore, for data with high global redundancy or repetitiveness, i.e., high non-local similarity among sub-regions at large scales, allocating more parameters to the shallow level will be more beneficial to improve its compression fidelity. Conversely, for data with weak global similarity, it is more efficient to allocate more parameters to the nodes at deep levels for better representation of the unique local features. + +Intra-level Parameter Allocation. In addition to allocating parameters across levels, we can also allocate more parameters within a level to some nodes according to certain criteria instead of even allocation. For example, the most valuable information in neuronal data is often distributed in sparser regions. If we allocate more parameters to the more important regions, then we can sacrifice the unimportant regions in exchange for the improvement of compression fidelity of the important ones. + +A simple way to take advantage of the TINC's hierarchy and flexibility is to allocate the number of parameters per node at each level according to the importance of the corresponding regions. Following this allocation, at each spatial resolution, the nodes corresponding to more important regions will have more parameters compared to other nodes in the same level. We demonstrate in Sec. 5.3 that this simple approach can significantly improve the compression fidelity of biological data, especially in regions with very low compression fidelity in the even allocation case, if an appropriate measure of the importance of the data region is adopted. Obviously there are more complex and efficient allocation methods, but this preliminary exploration is sufficient to demonstrate the advantages of TINC. + +# 5. Experiments + +# 5.1. Implementation Details + +All the hyper layers in each node are implemented using MLP, where the activation functions are all sine function and the parameters are initialized following the scheme proposed in [30]. All MLPs are optimized by Adamax [13] with initiate learning rate of 0.001 and decreased by a factor of 0.2 at the 2000th and 5000th iterations, respectively. We adopted a complete octtree, which means $K = 8^{(L - 1)}$ . Before optimization, all the coordinates and the intensities are normalized to $[-1,1]$ and $[0,100]$ respectively. After optimization, all the parameters of each hyper layers are serialized and stored in binary. There are also some information that need to be saved for data decompression, such as MLP structure, original data bit depth, image size, inverse normalization basic information. Because all of the items are + +Table 1. The scores on each dataset as compression ratio ranges from 64 to 1024, with the #best and the ●second best marked by color and symbols. The metrics include PSNR(dB), SSIM, and Acc. $\tau$ with $\tau$ acts as a threshold for binarization. The “All” or “High” suffixes after the metrics name indicate that it is averaged for all compression ratios or only high compression ratios (around $1024 \times$ ). + +
MethodMedical dataBiological data
PSNR All. (dB)SSIM All.PSNR High. (dB)SSIM High.Acc.200 All.Acc.500 All.Acc.200 High.Acc.500 High.
TINC (ours)•52.02•0.9897#50.59#0.9878•0.9945•0.9970#0.9934#0.9958
JPEG41.410.972230.490.93740.66120.98340.01970.9882
H.26451.180.989647.280.98600.99190.99590.98600.9926
HEVC#52.31#0.9903•50.51•0.9877#0.9955#0.99750.99170.9930
SCI51.900.989450.390.98760.99430.9965•0.9921•0.9951
NeRF50.930.987549.660.98630.99350.99620.99030.9940
NeRV47.110.985940.110.98000.98150.99010.97320.9867
DVC47.390.986545.740.98400.98270.99000.96920.9789
SGA+BB46.560.983643.020.98080.80380.98830.48170.9798
SSF46.250.980743.700.97730.72210.96030.77900.9542
+ +![](images/603eaac4398fd19d1861d67fb73f31a0f087cdd476fb18b35928fcf7f37abd8f.jpg) +(a) + +![](images/ac4c74b1044fdfdc78f00d0733c68e363b122676e0f411f3f6d0978be2a8773c.jpg) +(b) +Figure 2. Comparison of different compression methods on medical (a,b) and biological data (c,d) at different compression ratios. + +![](images/748e1e3b29dc7097da206165dc5c05ddfbe7a45c81a1e0386d4fa281119ce25e.jpg) +(c) + +![](images/a3c0f23b2c2a193454a213dc199bdd25a9533957ef635341e175347f0bec4a64.jpg) +(d) + +discrete, countable, and have a limited set of possible values, we can easily create a table listing all potential values and store their indices, to reduce the file size. + +We evaluated our approach on both massive medical and biological data. For the former we used HiP-CT dataset [38], which provides cellular level imaging of several organisms at multiple anatomical levels. We used four human organs (Lung, Heart, Kidney and Brain) from this dataset. For the latter we used the sub-micrometer resolution mouse whole-brain microscopic data (Neurons and Vessels) captured by our self-developed imaging system (see supplementary note 1 for details), which has varying brightness and structures across different regions. The access to these data are listed in supplementary note 2. + +For convenient and fair comparison, we cropped the data into the same size ( $256 \times 256 \times 256$ voxels for medical data and $64 \times 512 \times 512$ voxels for biological data). + +# 5.2. Performance Comparison with State-of-the-arts + +Benchmark Methods and Evaluation Metrics. The proposed method is comprehensively compared with state-of-the-art methods, which can be classified into three groups: (i) commercial compression tools, including JPEG [37], H.264 [39], and HEVC [33]; (ii) recently proposed data-driven deep learning based ones, including DVC [16], + +SGA+BB [43], and SSF [2]; (iii) recently proposed INR based ones, including SCI [41], NeRV [7], and NeRF [21]. We evaluated all the methods at 6 different compression ratios ranging from $64 \times$ to $1024 \times$ . It is worth noting that only INR based methods can precisely regulate the compression ratio. We adopted PSNR and SSIM for medical data including rich details, while Accuracy with threshold for binarization as evaluation metrics for biological data with sparse structures. For convenience we abbreviate Accuracy with threshold as "Acc.threshold" in the following descriptions. + +JPEG was tested on OpenCV. H.264 and HEVC were tested with their FFmpeg implementations. For SGA+BB, we adopted the settings with the best performance mentioned in the original paper. For DVC, we used its Pytorch implementation, and compressed the "I Frames" using JPEG method with quality parameter set to 50. For SSF, we used its CompressAI's implementation. In addition, for the data-driven methods, i.e. DVC, SGA+BB, and SSF, we fine-tuned the provided pre-trained models on 300 volumes randomly selected from the dataset. For SCI, we set the hyper parameter $a_{max} = 32$ . For NeRF, the hyper parameter "frequency" was set to 10. For NeRV, we control the compression ratio by adjusting the number and width of the fully connected layers. For 2D images compression methods, i.e. JPEG and SGA+BB, we firstly segmented the 3D data int 2D images before compression. + +Quantitative Performance Evaluation and Analysis. The average performance of each method under different compression ratios is plotted as a rate-distortion curve in Fig. 2. We can see that our algorithm outperforms almost all methods at high compression ratios and is comparable to HEVC at other compression ratios. For more detailed analysis, we summarize the scores on each dataset in Table 1. It should be noted that even if the PSNR of the medical data are higher than $40\mathrm{dB}$ , it does not mean that there are no distortion in the decompressed data, which is different from the natural scene. + +Some visual comparisons of decompressed organ slices are presented in Fig. 3 (see supplementary figure 1 for more results). For better visual comparison, we applied the same post-processing to the decompressed data. Under compression ratio around $87 \times$ (since HEVC and SGA+BB cannot precisely regulate compression ratio), we consistently outperform existing INR based (SCI and NeRV), data-driven (SGA+BB) and commercial (HEVC) compression methods. Although SCI achieved comparable results in the regions away from boundary, as shown in Fig. 3(a), it produced severe blocking artifacts at the boundary, as shown by the arrows in (b), which will affect the medical research as well as other downstream tasks. NeRV retained the general structures but suffered from the checkerboard artifacts caused by the aliasing in convolution. HEVC captured high fidelity details in "I Frames" but suffered from the artifacts in "P Frames", which distorted the edges in the volumes, such as the white matters in Brain data and muscles in Heart data, which would be unusable for medical research and other downstream tasks. SGA+BB struggled with retaining the rich details, since it cannot represent the similarity within the 3D volume as a 2D image compression algorithm. In addition, HEVC produced ghosting artifacts and position offset in neighbor slices, since the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured biomedical data. In contrast, our approach use INR to encode the intrinsic data structure. + +Comparison of Running Time. We summarized the time required for compression and decompression for each algorithm in our experiments, as shown in the supplementary tables 1 and 2. Comparatively, INR based compressor (TINC, SCI, NeRV, and NeRF) is slower in compression stage but can decompress faster. It is worth noting that biomedical data need to be compressed only once before saving or transmission, but decompressed frequently for processing and analysis. Therefore, TINC is advantageous in biomedical fields. + +# 5.3. Flexibility Settings for Different Data + +Setting of Tree Levels $L$ . We evaluated TINC at different tree levels $L$ on medical data under different com + +![](images/9fc1755f1cf1667b217dd8c16719c101cfc7745e80f5a089430d5ab67f1c9841.jpg) +(a) Three ROIs from Brain data; compression ratio: $\sim 87\times$ + +![](images/f3c7e3676c9cb7992d190149b03dfe6d4882c7275262be0b10d320684534945c.jpg) +(b) Three ROIs from Heart data; compression ratio: $\sim 87\times$ +Ground Truth TINC (ours) SCI NeRV HEVC SGA+BB +Figure 3. Visual comparisons of decompressed organs slices from Brain and Heart data between TINC and other benchmark methods under similar compression ratio around $87\times$ . + +pression ratios. As shown in Fig. 4(a), the performance of TINC was evaluated in terms of SSIM on 8 medical data, under $256 \times$ compression ratio. To better demonstrate the effect of increasing tree levels, we used the growth rate of SSIM, calculated by normalizing all SSIM values by the one when $L = 1$ . The SSIM scores on all the test data have a steadily increasing growth rate (dashed lines) as the number of tree depth increases from 1 to 3. Experimental results on other compression ratios show consistent trend (supplementary figure 2). + +We further investigated the relationship between the data complexity and the performance gain when increasing tree levels. For data complexity, we adopt the spectrum concentration defined in [41], which measures the percentage of high frequency information contained in the data. The results are plotted in Fig. 4(b), with the solid line being a quadratic fitting of the growth rate for the 8 medical data using least squares. From the plot one can see that the performance gain increases monotonously with data complexity. In other words, the benefit of using a large tree is more significant when the target data contains more high-frequency information. This trend is consistent with our reasoning in Sec. 4 that a large $L$ favours parameter sharing at coarser levels and leave more parameters for representing the details. Experimental results on other compression ratios and other data are consistent with this (supplementary figure 3). + +![](images/ded2a7b60ff6b048a82e6e6bf217bd846cb69b4211ffe5361f31475e0f1eec0e.jpg) +(a) + +![](images/1a34db492e70e691d1a0885f417450fdcb15e8473910ba19f8d09b57960200e0.jpg) +(b) +Figure 4. Effect of tree levels $L$ on TINC's compression fidelity, in terms of SSIM on 8 medical data under $256 \times$ compression ratio. (a) The growth rate of SSIM for each data when increasing the tree levels. The dashed lines represent the change in each data. The solid line represents the mean of changes, and the filled area represents the standard deviation. (b) The scatter plot of each data's complexity and the growth rate of SSIM when increasing tree levels from 1 to 3. The solid line represents the trend of change, estimated by a quadratic fit. + +In addition, we studied the performance gain of using a deep tree structure under limited bandwidth budget $|\Theta|$ , e.g., at high compression ratio. Supplementary figure 4 plots the performance improvement at increasing compression ratios, calculated from 8 medical data volumes. Almost all data have a smaller growth rate of SSIM or PSNR when compression ratio increases, i.e. reducing the number of available parameters. At high compression ratio $(512 \times)$ , some data are even of inferior fidelity using a deep tree, which is in line with our analysis in Sec. 4 that one should set a proper $L$ for choosing the right partitioning granularity under the constraint from available budget. + +Inter-level Parameter Allocation. We tested the performance at different parameter allocation proportions across levels on medical data under different compression ratios. Assuming there are $K^l$ nodes at $l$ th level, we denote the number of allocated parameters at this level as $|\Theta^l|$ and ratio with respect to that of its adjacent upper level as $r^l = \frac{|\Theta^l| / K^l}{|\Theta^{l-1}| / K^{l-1}}$ . Simply speaking, $r^l = 1$ means even inter-level allocation, a $r^l$ larger than 1 means imposing higher priority to the shallow levels and vice versa. We let $r^l$ of all levels take 0.8, 1 and 1.2 to represent the three typical cases of parameter allocation. + +We first investigated the effect of allocating more parameters to the shallow level, intending to share parameters among far apart similar blocks. As shown in Fig. 5(a), we divided a Brain data into 64 sub-regions, laid out in z-curve order, and selected 8 equally spaced distant regions. Then, we calculated the SSIM of these regions with respect to each other and used min-max normalization to obtain their non-local similarities. It can be seen that #1 and #33 regions are of high similarity. For convenient display, we also draw the + +![](images/91a36dac14b05d35ba23d46b1725b495453af2222d405a136f09e9c3a8a053d4.jpg) +(a) +Figure 5. Effect of allocating more parameters to the shallow level on TINC's compression fidelity for each sub-regions in a Brain data under $512 \times$ compression ratio. All sub-figures share the same y-axis label. (a) The heatmap of similarities between 8 equally spaced distant sub-regions. The serial numbers of the regions represent their z-curve order. (b) The non-local similarity of each region. (c) The growth rate of SSIM for each region when allocating more parameters to the shallow level. + +![](images/6eda077fe1153665b0fa6ec45599548cc30578461a2d7e857475a514a2d6c240.jpg) +(b) + +![](images/8202af729ba2662701429c6056c779272874c2d12a347b38bb0e3d13f3145945.jpg) +(c) + +bar plot of the normalized marginal sums of the similarity matrix in Fig. 5(b), which tells that #1 and #33 are both of high similarity to other blocks, while #25 and #57 regions are highly distinctive from others. + +To quantitatively measure the performance change at different blocks when allocating more parameters to the shallow level, we plot the growth rate of SSIM under $512 \times$ ratio, as shown in Fig. 5(c). Analyzing jointly with (b), we can find that the two regions with high non-local similarity, #1 and #33, gain a large boost, while those with the lowest similarity, #25 and #57, suffer a loss of fidelity. This is consistent with the intuition and analysis in Sec. 4 that allocating more parameters to the nodes at shallow levels would benefit representing distant but similar regions, but at expense of slight degeneration at very individual regions. Experimental results on other compression ratios and other data are shown in supplementary figures 5 and 6. + +We then investigated how to allocate parameters across levels according to the global consistency of the medical data. We divided the target data into 64 sub-regions, calculated the marginal sums of the similarity matrix. We change the setting of $r^l$ and the results are shown in the supplementary figure 7. The results show allocating more parameters to the shallow level is more likely to improve fidelity when the global consistency of the data is greater than 0.7, and allocating more to the deep level is better when the global consistency lower than 0.6. Otherwise, an even allocation may be a better choice. + +Intra-level Parameter Allocation. We tested the approach of allocating parameters to the nodes at the same level according to their importance proposed in Sec. 4, on biological data and under different compression ratios. Specifically, we let the number of allocated parameters be proportional to their importance. For example, with regard + +to the biological data such as Neurons the importance can be measured by the percentage of the number of valid signals, where the valid signal can be determined either by the result of traced neurons or by the intensity value. Here we chose the latter for simplicity. + +We first compare the compression fidelity of 6 Neurons data using the above two parameter allocation approaches under different compression ratios, as shown in supplementary figure 8. It can be seen that important allocation significantly improves the fidelity compared to even allocation. + +We then look further into the performance at all the sub-regions in the same target data. As shown in Fig. 6(a), we divided a Neurons data into 64 sub-regions and compute their growth rate of Acc.2000 when switching from even allocation to important allocation. It should be noted that we omitted those regions that have reached 1 (the beast value) under both allocation approaches in the subsequent experiments. From the scatter diagram we can see three groups of data: (i) Some regions show non-significant change (less than $10\%$ of the maximum change of the others, denoted as cross markers) using important allocation under $64\times$ compression ratio, because they are already with sufficient parameters. (iii) Other regions show significant improvement after introducing important allocation, denoted with circles. We performed a linear fit using least squares, as shown by the solid line. Overall, allocating parameters according to importance is helpful for the compression fidelity of important regions while at expenses of little fidelity degeneration in other less important ones. + +For a statistical analysis, we plotted the Acc.2000 of all 64 sub-regions under even and important allocation, as shown in Fig. 6(b). We found that importance based allocation can significantly improve those regions that have very poor compression fidelity under even allocation. This is mainly because that those regions with poor fidelity tend to contain more complex structures (with more neural signals), and thus are of higher importance scores and thus allocated more parameters accordingly. Experimental results on other compression ratios are detailed in supplementary figure 9. + +# 6. Conclusions + +Summary and Discussions. In this paper we propose a hierarchical parameter sharing mechanism for INR and implement it based on a tree-structured MLP for large and complex data compression tasks. The proposed approach TINC makes use of the local redundancy via block wise compression with separate MLPs and non-local redundancy via organizing these MLPs in tree structure to share parameters hierarchically. Such implementation forms a single network for the complex target data, not only incorporate the similarity distributed far apart and nonuniformly, and avoid discontinuity between adjacent regions as well. + +![](images/46013685926776d1b8ad2ff8cea383ee9f9cd81ca3d9d1549efd4869b79cc59b.jpg) +(a) + +![](images/8b765ba7b63ae12e00243303b71fd684e9b9d720be18c02d882b6ffb64260a0a.jpg) +(b) +Figure 6. Effect of important allocation on TINC's compression fidelity, in terms of Acc.2000 on a Neurons data under $64 \times$ compression ratio. (a) The growth rate of Acc.2000 for each region. The regions with and without significant growth rate are represented as cross and square respectively. The solid line represents the trend of change, estimated by a linear fit. (b) The violin plots of Acc.2000 for all these regions. The shaded area represents the distribution of Acc.2000 across all regions. The three horizontal lines from top to bottom represent the maximum, average and minimum values respectively. + +Extensive experiments show that this hierarchical parameter sharing mechanism introduced in TINC greatly improves the compression capability and wide applicability of INR. Our method outperforms state-of-the-art compressors including commercial tools (JPEG, H.264, HEVC), data-driven methods (DVC, SSF, SGA+BB), and existing INR based ones (SCI, NeRV, and NeRF) on diverse medical and biological data. + +We also analyze TINC's flexibility to different cases via experimentally studying the effect of three key settings—number of tree levels, intra-level and inter-level parameter allocation methods, and provide highlights on proper parameter setting and strategy selection according to the features of target data. + +Limitations and Future Extensions. Similar to all current INR based compression methods, TINC is of high decompression speed but slow in compression, since it takes time to pursue the MLPs matching the target data. We plan to combine meta-learning to find the best initialization parameters for each organ to speed up TINC. + +# Acknowledgements + +We acknowledge MoE Key Laboratory of Biomedical Photonics (Huazhong University of Science and Technology) for sharing their mouse brain-wide microscopic data. + +This work is supported by Beijing Natural Science Foundation (Grant No. Z200021), the National Natural Science Foundation of China (Grant Nos. 61931012, 62088102) and Project of Medical Engineering Laboratory of Chinese PLA General Hospital (Grant No. 2022SYSZ-ZKY21). + +# References + +[1] Arya Aftab, Alireza Morsali, and Shahrokh Ghaemmaghami. Multi-Head Relu Implicit Neural Representation Networks. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2510-2514, 2022. 2 +[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 5 +[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational Image Compression with a Scale Hyperprior. In International Conference on Learning Representations, 2018. 2 +[4] Nuri Benbarka, Timon Höfer, Hamd ul-moqeet Riaz, and Andreas Zell. Seeing Implicit Neural Representations as Fourier Series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2283-2292, 2022. 2 +[5] Alexander William Bergman, Petr Kellnhofer, and Gordon Wetzstein. Fast Training of Neural Lumigraph Representations using Meta Learning. In Advances in Neural Information Processing Systems, 2021. 2 +[6] Alexandre Boulch and Renaud Marlet. Poco: Point Convolution for Surface Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 2 +[7] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. NeRV: Neural Representations for Videos. Advances in Neural Information Processing Systems, 2021. 1, 2, 5 +[8] Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 1 +[9] Yilun Du, Katherine Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning Signal-agnostic Manifolds of Neural Fields. Advances in Neural Information Processing Systems, 2021. 2 +[10] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J. Zico Kolter. Multiplicative Filter Networks. In International Conference on Learning Representations, 2021. 2 +[11] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning Shape Templates with Structured Implicit Functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7154-7164, 2019. 1 +[12] Milad Soltany Kadarvish, Hesam Mojtahedi, Hossein Entezari Zarch, Amirhossein Kazerouni, Alireza Morsali, Azra Abtahi, and Farokh Marvasti. Ensemble Neural Representation Networks. arXiv preprint arXiv:2110.04124, 2021. 1, 2, 3 +[13] Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980, 2014. 4 +[14] Sylwester Klocek, Lukasz Maziarka, Maciej Wolczyk, Jacek Tabor, Jakub Nowak, and Marek Śmieja. Hypernetwork + +Functional Image Representation. In International Conference on Artificial Neural Networks, pages 496-510, 2019. 2 +[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1 +[16] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: An End-to-end Deep Video Compression Framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2, 5 +[17] Y. Lu, K. Jiang, J. A. Levine, and M. Berger. Compressive Neural Representations of Volumetric Scalar Fields. Computer Graphics Forum, 40(3):135-146, 2021. 2 +[18] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive Coordinate Networks for Neural Scene Representation. ACM Trans. Graph., 40(4), jul 2021. 2 +[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1 +[20] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated Periodic Activations for Generalizable Local Functional Representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, October 2021. 1, 2, 3 +[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2, 5 +[22] David Minnen, Johannes Balle, and George D Toderici. Joint Autoregressive and Hierarchical Priors for Learned Image Compression. In Advances in Neural Information Processing Systems, volume 31, 2018. 2 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph., 41(4), jul 2022. 1, 2 +[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing Scenes as Compositional Generative Neural Feature Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2 +[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. +[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo + +Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1 +[27] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional Occupancy Networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2 +[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1 +[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned Implicit Function for High-resolution Clothed Human Digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 1 +[30] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit Neural Representations with Periodic Activation Functions. In Advances in Neural Information Processing Systems, 2020. 2, 4 +[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3d-structure-aware Neural Scene Representations. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[32] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The JPEG 2000 Still Image Compression Standard. IEEE Signal Processing Magazine, 18(5):36-58, 2001. 2 +[33] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the High Efficiency Video Coding (HEVC) Standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 2, 5 +[34] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2 +[35] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned Initializations for Optimizing Coordinate-based Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2021. 2 +[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In Advances in Neural Information Processing Systems, 2020. 2 +[37] Gregory K Wallace. The JPEG Still Picture Compression Standard. IEEE Transactions on Consumer Electronics, 38(1):xviii-xxxiv, 1992. 2, 5 +[38] CL Walsh, P Tafforeau, WL Wagner, DJ Jafree, A Bellier, C Werlein, MP Kühnel, E Boller, S Walker-Samuel, JL Rober + +tus, et al. Imaging Intact Human Organs with Local Resolution of Cellular Structures using Hierarchical Phase-Contrast Tomography. Nature Methods, 18(12):1532-1541, 2021. 5 +[39] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the H.264/AVC Video Coding Standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 2, 5 +[40] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for Video Compression with Hierarchical Quality and Recurrent Enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6628-6637, 2020. 2 +[41] Runzhao Yang, Tingxiong Xiao, Yuxiao Cheng, Qianni Cao, Jinyuan Qu, Jinli Suo, and Qionghai Dai. SCI: A Spectrum Concentrated Implicit Neural Compression for Biomedical Data. arXiv preprint arXiv:2209.15180, 2022. 1, 2, 3, 5, 6 +[42] Ruihan Yang, Yibo Yang, Joseph Marino, and Stephan Mandt. Hierarchical Autoregressive Modeling for Neural Video Compression. In ICLR, 2021. 2 +[43] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving Inference for Neural Image Compression. In Advances in Neural Information Processing Systems, volume 33, pages 573-584, 2020. 2, 5 +[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields from One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1 +[45] Ellen D Zhong, Tristan Bepler, Joseph H Davis, and Bonnie Berger. Reconstructing Continuous Distributions of 3D Protein Structure from Cryo-EM Images. arXiv preprint arXiv:1909.05215, 2019. 2 \ No newline at end of file diff --git a/2023/TINC_ Tree-Structured Implicit Neural Compression/images.zip b/2023/TINC_ Tree-Structured Implicit Neural Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0a65bca1a51ec8b2d20fbdaab999fbf3f908656 --- /dev/null +++ b/2023/TINC_ Tree-Structured Implicit Neural Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3125fa749cec496043f271e00a6101779770c65b75b95b5ddf651ed000dd54a0 +size 409477 diff --git a/2023/TINC_ Tree-Structured Implicit Neural Compression/layout.json b/2023/TINC_ Tree-Structured Implicit Neural Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4691bbfaef9139611627026835241c04db9bda8e --- /dev/null +++ b/2023/TINC_ Tree-Structured Implicit Neural Compression/layout.json @@ -0,0 +1,8216 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "type": "text", + "content": "TINC: Tree-structured Implicit Neural Compression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 144, + 415, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 144, + 415, + 200 + ], + "spans": [ + { + "bbox": [ + 179, + 144, + 415, + 200 + ], + "type": "text", + "content": "Runzhao Yang \nDepartment of Automation, Tsinghua University \nBeijing 100084, China \nyangrz20@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 227, + 210, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 227, + 210, + 239 + ], + "spans": [ + { + "bbox": [ + 162, + 227, + 210, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 289, + 540 + ], + "type": "text", + "content": "Implicit neural representation (INR) can describe the target scenes with high fidelity using a small number of parameters, and is emerging as a promising data compression technique. However, limited spectrum coverage is intrinsic to INR, and it is non-trivial to remove redundancy in diverse complex data effectively. Preliminary studies can only exploit either global or local correlation in the target data and thus of limited performance. In this paper, we propose a Tree-structured Implicit Neural Compression (TINC) to conduct compact representation for local regions and extract the shared features of these local representations in a hierarchical manner. Specifically, we use Multi-Layer Perceptrons (MLPs) to fit the partitioned local regions, and these MLPs are organized in tree structure to share parameters according to the spatial distance. The parameter sharing scheme not only ensures the continuity between adjacent regions, but also jointly removes the local and non-local redundancy. Extensive experiments show that TINC improves the compression fidelity of INR, and has shown impressive compression capabilities over commercial tools and other deep learning based methods. Besides, the approach is of high flexibility and can be tailored for different data and parameter settings. The source code can be found at https://github.com/RichealYoung/TINC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 581, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 677 + ], + "type": "text", + "content": "In the big data era, there exist massive and continuously growing amount of visual data from different fields, from surveillance, entertainment, to biology and medical diagnosis. The urgent need for efficient data storage, sharing and transmission all requires effective compression techniques. Although image and video compression have been studied for decades and there are a number of widely used commercial tools, the compression of medical and biological data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 715 + ], + "type": "text", + "content": "This work was done in collaboration with Tingxiong Xiao and Yuxiao Cheng, under supervision of Prof. Jinli Suo and Prof. Qionghai Dai affiliated with Department of Automation, Tsinghua Univ." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 407, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 407, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 407, + 240 + ], + "type": "text", + "content": "are not readily available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 242, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 327 + ], + "type": "text", + "content": "Implicit neural representation (INR) is becoming widely-used for scene rendering [19, 21, 23, 31, 44], shape estimation [8, 11, 25, 29], and dynamics modeling [15, 26, 28]. Due to its powerful representation capability, INR can describe nature scenes at high fidelity with a much smaller number of parameters than raw discrete grid representation, and thus serves as a promising data compression technique." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 327, + 548, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 548, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 548, + 544 + ], + "type": "text", + "content": "In spite of the big potential, INR based compression is quite limited confronted with large sized data [20], since INR is intrinsically of limited spectrum coverage and cannot envelop the spectrum of the target data, as analyzed in [41]. Two pioneering works using INR for data compression, including NeRV [7] and SCI [41], have attempted to handle this issue in their respective ways. Specifically, NeRV introduces the convolution operation into INR, which can reduce the required number of parameters using the weight sharing mechanism. However, convolution is spatially invariant and thus limits NeRV's representation accuracy on complex data with spatial varying feature distribution. Differently, to obtain high fidelity on complex data, SCI adopts divide-and-conquer strategy and partitions the data into blocks within INR's concentrated spectrum envelop. This improves the local fidelity, but cannot remove non-local redundancies for higher compression ratio and tend to cause blocking artifacts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 715 + ], + "type": "text", + "content": "To compress large and complex data with INR, in this paper we propose to build a tree-structured Multi-Layer Perceptrons (MLPs), which consists of a set of INRs to represent local regions in a compact manner and organizes them under a hierarchical architecture for parameter sharing and higher compression ratio. Specifically, we first draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress different regions with multiple MLPs separately. Then we incorporate these MLPs with a tree structure to extract their shared parameters hierarchically, following the rule that spatially closer regions are of higher similarity and share more parameters. Such a joint representation strategy can remove the redundancy both within each local region and among non-local" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18517" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "regions, and also ensures the continuity between adjacent regions. The scheme of the proposed network is illustrated in Fig. 1, and we name our approach TINC (Tree-structured Implicit Neural Compression)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 204 + ], + "type": "text", + "content": "Using the massive and diverse biomedical data, we conduct extensive experiments to validate that TINC greatly improves the capability of INR and even outperforms the commercial compression tools (H.264 and HEVC) under high compression ratios. The proposed TINC is also a general framework, which can be flexibly adapted to diverse data and varying settings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 215, + 134, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 134, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 134, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 235, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 533 + ], + "type": "text", + "content": "Data Compression. Data compression is a key technique for efficient data storage, transmission and sharing, especially in the big data era. In recent decades, many commercial compression tools for massive visual data have become increasingly mature, like JPEG [37], JPEG2000 [32], H.264 [39], and HEVC [33], while emerging deep learning-based approaches are also gaining momentum and demonstrating continuously improving compression capabilities [2, 3, 16, 22, 40, 42, 43]. Compression for nature images and videos have been widely and successfully used, however, there still lack effective compression techniques for the huge amount of three or higher dimensional data such as biomedical volume. Image/Video compressors cannot be applied directly, because these high dimensional data is a discrete sampling of 3D organism, instead of a video describing temporal evolution of 2D slices. Hence, the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured data [41], and tend to cause discontinuity and degenerate the downstream tasks. Our approach differs completely from existing image/video compression methodologies in that we encode the intrinsic data structure with INR. Besides, our approach searches for the best network structure and parameters representing the target data and is free of generalization issues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 689 + ], + "type": "text", + "content": "Implicit Neural Compression. Recent works [7, 17, 41] have introduced INR to the field of data compression with exciting results. They fit the original discrete grid-based data with continuous implicit functions in a transform-coding-like manner, which takes advantage of the powerful representation capabilities of neural networks while avoids the generalization problem. Obviously, the compression capability of INR is closely related to its representation capability, which can be improved from proper design of network structure [7, 10, 17], selection of activation function [14, 20, 30], embedding [4, 21, 23, 36, 45], and learning strategies such as manifold learning [9, 20, 30, 31], meta learning [5, 9, 35] and ensemble learning [1, 12, 24]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "However, improving these capability does not necessarily head for a ready compressor for general scenes or ob" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "jects, since there exist large diversity in the target data and pursuing a general network with high fidelity for all the data under given parameter budget. In order to pursue a more compact and efficient representation, we draw on the idea of ensemble learning [12] and use a divide-and-conquer strategy [20, 41] to compress the data in each region separately. On top of this, we improve INR by introducing a hierarchical parameter sharing mechanism in terms of data compression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 546, + 400 + ], + "type": "text", + "content": "Parameter Sharing in INR. To improve the representation accuracy of INR for large and complex scenes, many algorithms use the parameter sharing strategy. One implementation is introducing convolution operations [6, 7, 27], which allows hierarchical extraction of local features, reduces the required parameters, and enables describing high resolution data. However, for those complex data with uneven distribution of features, the spatial invariance of convolution would lead to lower local representation accuracy. Another parameter sharing mechanism is positional encoding [21, 36], in which nearby spatial positions share partial mapping features hierarchically. However, this can only achieve higher representation accuracy when equipped with random or learnable mappings [4, 30, 36], which can be considered as an embedding that encodes only spatial distance. In addition, using embedding [18, 23, 34] intrinsically reduces the computational complexity at expenses of increased parameters, and thus not optimal for compression." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": "To make use of advantageous and bypass the above issues, we borrow the idea of hierarchical parameter sharing into local INRs, and build our compressor on a widely used tree structured data partitioning [18, 34, 41]. On the one hand, the parameter sharing is applied among block-specific INRs to cope with the large data with high local fidelity. On the other hand, we share parameters on INR rather than the data itself or its embedding, to encode richer information other than spatial distance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 542, + 544, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 544, + 568 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 544, + 568 + ], + "type": "text", + "content": "3. Tree-structured Implicit Neural Compression" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 582, + 525, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 525, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 525, + 594 + ], + "type": "text", + "content": "3.1. Ensemble of Implicit Neural Compressors" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Implicit neural compression uses a compact neural network to parameterize a function describing the target data defined over a spatial coordinates system. More specifically, for an " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " dimensional data, let " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{V} \\coloneqq [-1,1]^N" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " denote its grid coordinates and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}(\\mathbf{v})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{V}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " denote the intensity value at coordinate " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", one can learn a compact implicit function " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to estimate " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}(\\mathbf{v})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v};\\Theta)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and finally serialize them to achieve a compressed representation. Mathematically, we can formulate the compression" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18518" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 87, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 60, + 87, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 87, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 87, + 541, + 236 + ], + "type": "image", + "image_path": "8f688fdac98302532d50afa6477736159751ed4539cb8c6909d20a5aeef44ccc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 546, + 312 + ], + "type": "text", + "content": "Figure 1. The scheme of the proposed approach TINC. For a target data, we divide the volume into equal-size blocks via octree partitioning, with some neighboring and far-apart blocks of similar appearances. Then each block can be representation with an implicit neural function, implemented as an MLP. After sharing parameters among similar blocks, we can achieve a more compact neural network with a tree shaped structure. Here we highlight the similar blocks sharing network parameters with the same color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 332, + 178, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 178, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 178, + 342 + ], + "type": "text", + "content": "task as an optimization problem" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 349, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 349, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 349, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\Theta}} \\int_ {\\mathbf {v} \\in \\mathbb {V}} \\mathcal {L} (f (\\mathbf {v}; \\boldsymbol {\\Theta}), \\mathbf {d} (\\mathbf {v})), \\tag {1}", + "image_path": "49787b9c72c110b92d0d3dbd066ab72ac3ec3a5f0298d814c7ef7597ac09ed56.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "content": " measures the difference between the decompressed and target data. For decompression, one can retrieve " + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "content": " from the compressed file and then calculate function values " + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v};\\Theta)" + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "content": " over grid " + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 47, + 385, + 287, + 433 + ], + "type": "text", + "content": " via simple forward propagation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "content": "Due to limited spectrum coverage, INR is incapable of compactly describe large and complex data [20,41], we borrow the idea of ensemble learning to partition the target volume into blocks and use multiple less expressive " + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "inline_equation", + "content": "f_{k}(\\cdot ,\\Theta_{k})" + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "content": " to achieve a powerful representation. Specifically, we adopt the divide-and-conquer strategy as in [12, 20, 41] to ensemble all implicit functions " + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\{f_k,k = 1,\\dots ,K\\}" + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "content": " that represents data at its corresponding coordinate regions " + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{V}_k,k = 1,\\dots ,K\\}" + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "content": " and compose an " + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "content": ", which is equivalent to approximating a complex function by a piece-wise function composed of " + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 434, + 287, + 564 + ], + "type": "text", + "content": " simple functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 570, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 95, + 570, + 287, + 602 + ], + "type": "interline_equation", + "content": "f (\\mathbf {v}; \\boldsymbol {\\Theta}) := \\sum_ {k = 1} ^ {K} \\mathbf {1} _ {\\mathbb {V} _ {k}} \\left(\\mathbf {v}, f _ {k} \\left(\\mathbf {v}, \\boldsymbol {\\Theta} _ {k}\\right)\\right), \\tag {2}", + "image_path": "a9d35f39dd07ffd941c10083d97fd4e6b9e4e106bb9f23f9175bb754bed81cb1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 605, + 287, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 605, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 92, + 605, + 287, + 637 + ], + "type": "interline_equation", + "content": "\\mathbf {1} _ {\\mathbb {V} _ {k}} (\\mathbf {v}, x) = \\left\\{ \\begin{array}{l} x, \\mathbf {v} \\in \\mathbb {V} _ {k} \\\\ 0, e l s e \\end{array} . \\right. \\tag {3}", + "image_path": "fd2325dc7f5fcfd4d2a7ec116ccf5b45733bc7cd72bc2025e9a6d08f31233d2a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": "In addition, we must constrain the total number of parameters to be consistent with the parameter budget, i.e., " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "|\\Theta| = \\sum_{k=1}^{K} |\\Theta_k|" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " is the support operation. Here we omit the parameters for serializing " + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_k" + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": " which consists of only a few structural hyperparameters of neural network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": "For a target data volume shown in the first column of Fig. 1, we use an " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " level tree structure containing " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " leaf nodes to organize " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_k" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": ", as illustrated in the 2nd column of Fig. 1. The root node represents the whole coordinate region " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " and each child node represents " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\frac{1}{2^N}" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " region of its parent node. We use " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " to compress the data in coordinate region " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_k" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": " represented by the " + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 332, + 545, + 439 + ], + "type": "text", + "content": "th leaf node. This structure allows us to organize ensemble learning in high degree of flexibility." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 466, + 536, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 536, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 536, + 479 + ], + "type": "text", + "content": "3.2. Hierarchical Parameter Sharing Mechanism" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": "We let these " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\{f_k\\}" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": " share their neural network parameters hierarchically with each other according to the spatial distance between corresponding regions " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_k" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": ". The sharing mechanism is defined on the octree structure. Specifically, for a leaf node at level " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": ", its corresponding MLP implemented " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "f_k" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": "'s hidden layers can be divided into " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": " segments, i.e. " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "f_k = f_k^{out} \\circ f_k^l \\circ f_k^{l-1} \\circ \\dots \\circ f_k^1 \\circ f_k^{in}" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": ". For a set of leaf nodes, we determine the number of shared segments based on the number of their common ancestor nodes. For example, if " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "f_j" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": " share the same ancestor nodes at " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "1 \\sim 3" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": " levels, three pairs of hidden layer segments " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "(f_i^1, f_j^1)" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "(f_i^2, f_j^2)" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "inline_equation", + "content": "(f_i^3, f_j^3)" + }, + { + "bbox": [ + 304, + 491, + 545, + 636 + ], + "type": "text", + "content": " will share the same parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "The sharing mechanism is illustrated in the 3rd column in Fig. 1. Since the distance between two leaf nodes can also be inferred from the number of their shared ancestor nodes, above hierarchical parameter sharing mechanism implies that closer blocks share more parameters and are of higher similarity in the implicit representation space." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18519" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 249, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 249, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 249, + 84 + ], + "type": "text", + "content": "3.3. Tree-structured Network Architecture" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 163 + ], + "type": "text", + "content": "We propose a tree-structured MLP based on the " + }, + { + "bbox": [ + 47, + 91, + 286, + 163 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 91, + 286, + 163 + ], + "type": "text", + "content": " level octree partitioning, as illustrated in the 4th column in Fig. 1. Each node contains a hyper layer consisting of some fully connected layers and takes the output of its parent node's hyper layer as input. Root node and leaf nodes additionally contain the input and output layers respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 286, + 258 + ], + "type": "text", + "content": "Following the depth direction, from leaf nodes upward to the root of the tree, the output information of the leaf node is processed by the hyper layers in its ancestor nodes. If we consider the hyper layer as a part of hidden layers, then any leaf node can be considered as an MLP, whose hidden layers are explicitly divided into multiple segments contained in its ancestor nodes across different levels. In this way we can ensemble all " + }, + { + "bbox": [ + 46, + 163, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\{f_k\\}" + }, + { + "bbox": [ + 46, + 163, + 286, + 258 + ], + "type": "text", + "content": " into a tree-structured MLP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 259, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 286, + 354 + ], + "type": "text", + "content": "At the same level, all sibling nodes share the same parent node and thus take the same information as input. Therefore, for any two leaf nodes sharing the same ancestor nodes, their output information have flown through the same hyper layers contained in the ancestor nodes. In other words, the MLPs of these two leaf nodes will share a part of hidden layers, which implements the hierarchical parameter sharing mechanism described in Sec. 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 368, + 235, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 235, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 235, + 380 + ], + "type": "text", + "content": "4. Cross Level Parameter Allocation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 388, + 286, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 286, + 435 + ], + "type": "text", + "content": "After the network design, we are left with some key implementation issues: (i) how to set a proper number of levels to finalize the tree structured network, (ii) how to allocate available parameters across MLPs at the tree nodes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": "The Number of Tree Levels " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": ". The number of tree levels " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " determines the number of divided segments of " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": "'s hidden layers, i.e., a larger " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " implies the parameter sharing mechanism proceeds to higher spatial resolution. Setting a large " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " has both advantageous and disadvantages. On the one hand, a larger " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " means finer partitioning with a larger number of blocks sharing parameters (with the same hyper layers as in the ancestor nodes), which is beneficial for describing the regions with rich details and thus improve the compression fidelity. On the other hand, with the same total number of parameters, a larger " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " will lead to more leaf nodes " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " and less parameters " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "|\\Theta_k|" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": " each " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": ", which might reduce " + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 437, + 286, + 604 + ], + "type": "text", + "content": "'s representation capability and might harm the compression fidelity in turn." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "content": "Therefore, we should set the tree level " + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "content": " by trading off between increased fineness in partitioning and parameter allocation, and " + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "content": " decreased representation capability. In Sec. 5.3, we experimentally prove that it is favourable using large tree levels for the data with rich details or under low compression ratios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Inter-level Parameter Allocation. The parameters at the shallow level describe the shared features at large scale, i.e., the non-local redundancy distributed over distant regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "Similarly, from the perspective of data features, if some far apart sub-regions at a certain level are highly similar, one can allocate more parameters to their common ancestor nodes, and vice versa." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 215 + ], + "type": "text", + "content": "Therefore, for data with high global redundancy or repetitiveness, i.e., high non-local similarity among sub-regions at large scales, allocating more parameters to the shallow level will be more beneficial to improve its compression fidelity. Conversely, for data with weak global similarity, it is more efficient to allocate more parameters to the nodes at deep levels for better representation of the unique local features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 217, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 324 + ], + "type": "text", + "content": "Intra-level Parameter Allocation. In addition to allocating parameters across levels, we can also allocate more parameters within a level to some nodes according to certain criteria instead of even allocation. For example, the most valuable information in neuronal data is often distributed in sparser regions. If we allocate more parameters to the more important regions, then we can sacrifice the unimportant regions in exchange for the improvement of compression fidelity of the important ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 326, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 493 + ], + "type": "text", + "content": "A simple way to take advantage of the TINC's hierarchy and flexibility is to allocate the number of parameters per node at each level according to the importance of the corresponding regions. Following this allocation, at each spatial resolution, the nodes corresponding to more important regions will have more parameters compared to other nodes in the same level. We demonstrate in Sec. 5.3 that this simple approach can significantly improve the compression fidelity of biological data, especially in regions with very low compression fidelity in the even allocation case, if an appropriate measure of the importance of the data region is adopted. Obviously there are more complex and efficient allocation methods, but this preliminary exploration is sufficient to demonstrate the advantages of TINC." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 506, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 388, + 520 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 526, + 438, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 438, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 438, + 539 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "content": "All the hyper layers in each node are implemented using MLP, where the activation functions are all sine function and the parameters are initialized following the scheme proposed in [30]. All MLPs are optimized by Adamax [13] with initiate learning rate of 0.001 and decreased by a factor of 0.2 at the 2000th and 5000th iterations, respectively. We adopted a complete octtree, which means " + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "inline_equation", + "content": "K = 8^{(L - 1)}" + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "content": ". Before optimization, all the coordinates and the intensities are normalized to " + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "inline_equation", + "content": "[0,100]" + }, + { + "bbox": [ + 304, + 545, + 545, + 712 + ], + "type": "text", + "content": " respectively. After optimization, all the parameters of each hyper layers are serialized and stored in binary. There are also some information that need to be saved for data decompression, such as MLP structure, original data bit depth, image size, inverse normalization basic information. Because all of the items are" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18520" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 112, + 541, + 253 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. The scores on each dataset as compression ratio ranges from 64 to 1024, with the #best and the ●second best marked by color and symbols. The metrics include PSNR(dB), SSIM, and Acc. " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " acts as a threshold for binarization. The “All” or “High” suffixes after the metrics name indicate that it is averaged for all compression ratios or only high compression ratios (around " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "1024 \\times" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 112, + 541, + 253 + ], + "lines": [ + { + "bbox": [ + 51, + 112, + 541, + 253 + ], + "spans": [ + { + "bbox": [ + 51, + 112, + 541, + 253 + ], + "type": "table", + "html": "
MethodMedical dataBiological data
PSNR All. (dB)SSIM All.PSNR High. (dB)SSIM High.Acc.200 All.Acc.500 All.Acc.200 High.Acc.500 High.
TINC (ours)•52.02•0.9897#50.59#0.9878•0.9945•0.9970#0.9934#0.9958
JPEG41.410.972230.490.93740.66120.98340.01970.9882
H.26451.180.989647.280.98600.99190.99590.98600.9926
HEVC#52.31#0.9903•50.51•0.9877#0.9955#0.99750.99170.9930
SCI51.900.989450.390.98760.99430.9965•0.9921•0.9951
NeRF50.930.987549.660.98630.99350.99620.99030.9940
NeRV47.110.985940.110.98000.98150.99010.97320.9867
DVC47.390.986545.740.98400.98270.99000.96920.9789
SGA+BB46.560.983643.020.98080.80380.98830.48170.9798
SSF46.250.980743.700.97730.72210.96030.77900.9542
", + "image_path": "5670dc8f4b6df00fd72708d8e97e68803906e24c7ed7dcad4c4f55a89b673e35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 266, + 175, + 353 + ], + "blocks": [ + { + "bbox": [ + 59, + 266, + 175, + 353 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 175, + 353 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 175, + 353 + ], + "type": "image", + "image_path": "603eaac4398fd19d1861d67fb73f31a0f087cdd476fb18b35928fcf7f37abd8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 354, + 120, + 363 + ], + "lines": [ + { + "bbox": [ + 109, + 354, + 120, + 363 + ], + "spans": [ + { + "bbox": [ + 109, + 354, + 120, + 363 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 266, + 296, + 353 + ], + "blocks": [ + { + "bbox": [ + 175, + 266, + 296, + 353 + ], + "lines": [ + { + "bbox": [ + 175, + 266, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 175, + 266, + 296, + 353 + ], + "type": "image", + "image_path": "ac4c74b1044fdfdc78f00d0733c68e363b122676e0f411f3f6d0978be2a8773c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 354, + 241, + 363 + ], + "lines": [ + { + "bbox": [ + 230, + 354, + 241, + 363 + ], + "spans": [ + { + "bbox": [ + 230, + 354, + 241, + 363 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 373, + 531, + 384 + ], + "lines": [ + { + "bbox": [ + 61, + 373, + 531, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 531, + 384 + ], + "type": "text", + "content": "Figure 2. Comparison of different compression methods on medical (a,b) and biological data (c,d) at different compression ratios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 266, + 417, + 353 + ], + "blocks": [ + { + "bbox": [ + 296, + 266, + 417, + 353 + ], + "lines": [ + { + "bbox": [ + 296, + 266, + 417, + 353 + ], + "spans": [ + { + "bbox": [ + 296, + 266, + 417, + 353 + ], + "type": "image", + "image_path": "748e1e3b29dc7097da206165dc5c05ddfbe7a45c81a1e0386d4fa281119ce25e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 354, + 362, + 363 + ], + "lines": [ + { + "bbox": [ + 352, + 354, + 362, + 363 + ], + "spans": [ + { + "bbox": [ + 352, + 354, + 362, + 363 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 417, + 266, + 538, + 353 + ], + "blocks": [ + { + "bbox": [ + 417, + 266, + 538, + 353 + ], + "lines": [ + { + "bbox": [ + 417, + 266, + 538, + 353 + ], + "spans": [ + { + "bbox": [ + 417, + 266, + 538, + 353 + ], + "type": "image", + "image_path": "a3c0f23b2c2a193454a213dc199bdd25a9533957ef635341e175347f0bec4a64.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 354, + 484, + 363 + ], + "lines": [ + { + "bbox": [ + 473, + 354, + 484, + 363 + ], + "spans": [ + { + "bbox": [ + 473, + 354, + 484, + 363 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 400, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 435 + ], + "type": "text", + "content": "discrete, countable, and have a limited set of possible values, we can easily create a table listing all potential values and store their indices, to reduce the file size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 436, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 288, + 567 + ], + "type": "text", + "content": "We evaluated our approach on both massive medical and biological data. For the former we used HiP-CT dataset [38], which provides cellular level imaging of several organisms at multiple anatomical levels. We used four human organs (Lung, Heart, Kidney and Brain) from this dataset. For the latter we used the sub-micrometer resolution mouse whole-brain microscopic data (Neurons and Vessels) captured by our self-developed imaging system (see supplementary note 1 for details), which has varying brightness and structures across different regions. The access to these data are listed in supplementary note 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "type": "text", + "content": "For convenient and fair comparison, we cropped the data into the same size (" + }, + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 256" + }, + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "type": "text", + "content": " voxels for medical data and " + }, + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "type": "inline_equation", + "content": "64 \\times 512 \\times 512" + }, + { + "bbox": [ + 46, + 568, + 288, + 604 + ], + "type": "text", + "content": " voxels for biological data)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 611, + 287, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 287, + 634 + ], + "type": "text", + "content": "5.2. Performance Comparison with State-of-the-arts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Benchmark Methods and Evaluation Metrics. The proposed method is comprehensively compared with state-of-the-art methods, which can be classified into three groups: (i) commercial compression tools, including JPEG [37], H.264 [39], and HEVC [33]; (ii) recently proposed data-driven deep learning based ones, including DVC [16]," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "type": "text", + "content": "SGA+BB [43], and SSF [2]; (iii) recently proposed INR based ones, including SCI [41], NeRV [7], and NeRF [21]. We evaluated all the methods at 6 different compression ratios ranging from " + }, + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "type": "inline_equation", + "content": "1024 \\times" + }, + { + "bbox": [ + 304, + 400, + 545, + 520 + ], + "type": "text", + "content": ". It is worth noting that only INR based methods can precisely regulate the compression ratio. We adopted PSNR and SSIM for medical data including rich details, while Accuracy with threshold for binarization as evaluation metrics for biological data with sparse structures. For convenience we abbreviate Accuracy with threshold as \"Acc.threshold\" in the following descriptions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": "JPEG was tested on OpenCV. H.264 and HEVC were tested with their FFmpeg implementations. For SGA+BB, we adopted the settings with the best performance mentioned in the original paper. For DVC, we used its Pytorch implementation, and compressed the \"I Frames\" using JPEG method with quality parameter set to 50. For SSF, we used its CompressAI's implementation. In addition, for the data-driven methods, i.e. DVC, SGA+BB, and SSF, we fine-tuned the provided pre-trained models on 300 volumes randomly selected from the dataset. For SCI, we set the hyper parameter " + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "inline_equation", + "content": "a_{max} = 32" + }, + { + "bbox": [ + 304, + 521, + 546, + 713 + ], + "type": "text", + "content": ". For NeRF, the hyper parameter \"frequency\" was set to 10. For NeRV, we control the compression ratio by adjusting the number and width of the fully connected layers. For 2D images compression methods, i.e. JPEG and SGA+BB, we firstly segmented the 3D data int 2D images before compression." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18521" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "text", + "content": "Quantitative Performance Evaluation and Analysis. The average performance of each method under different compression ratios is plotted as a rate-distortion curve in Fig. 2. We can see that our algorithm outperforms almost all methods at high compression ratios and is comparable to HEVC at other compression ratios. For more detailed analysis, we summarize the scores on each dataset in Table 1. It should be noted that even if the PSNR of the medical data are higher than " + }, + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "inline_equation", + "content": "40\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 72, + 287, + 203 + ], + "type": "text", + "content": ", it does not mean that there are no distortion in the decompressed data, which is different from the natural scene." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 539 + ], + "type": "text", + "content": "Some visual comparisons of decompressed organ slices are presented in Fig. 3 (see supplementary figure 1 for more results). For better visual comparison, we applied the same post-processing to the decompressed data. Under compression ratio around " + }, + { + "bbox": [ + 46, + 205, + 289, + 539 + ], + "type": "inline_equation", + "content": "87 \\times" + }, + { + "bbox": [ + 46, + 205, + 289, + 539 + ], + "type": "text", + "content": " (since HEVC and SGA+BB cannot precisely regulate compression ratio), we consistently outperform existing INR based (SCI and NeRV), data-driven (SGA+BB) and commercial (HEVC) compression methods. Although SCI achieved comparable results in the regions away from boundary, as shown in Fig. 3(a), it produced severe blocking artifacts at the boundary, as shown by the arrows in (b), which will affect the medical research as well as other downstream tasks. NeRV retained the general structures but suffered from the checkerboard artifacts caused by the aliasing in convolution. HEVC captured high fidelity details in \"I Frames\" but suffered from the artifacts in \"P Frames\", which distorted the edges in the volumes, such as the white matters in Brain data and muscles in Heart data, which would be unusable for medical research and other downstream tasks. SGA+BB struggled with retaining the rich details, since it cannot represent the similarity within the 3D volume as a 2D image compression algorithm. In addition, HEVC produced ghosting artifacts and position offset in neighbor slices, since the optical flow algorithms that are widely applicable in video compression do not work well on such inherently spatially structured biomedical data. In contrast, our approach use INR to encode the intrinsic data structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 660 + ], + "type": "text", + "content": "Comparison of Running Time. We summarized the time required for compression and decompression for each algorithm in our experiments, as shown in the supplementary tables 1 and 2. Comparatively, INR based compressor (TINC, SCI, NeRV, and NeRF) is slower in compression stage but can decompress faster. It is worth noting that biomedical data need to be compressed only once before saving or transmission, but decompressed frequently for processing and analysis. Therefore, TINC is advantageous in biomedical fields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 670, + 244, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 244, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 244, + 683 + ], + "type": "text", + "content": "5.3. Flexibility Settings for Different Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "Setting of Tree Levels " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": ". We evaluated TINC at different tree levels " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " on medical data under different com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 543, + 186 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 543, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 543, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 543, + 186 + ], + "type": "image", + "image_path": "9fc1755f1cf1667b217dd8c16719c101cfc7745e80f5a089430d5ab67f1c9841.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 186, + 503, + 195 + ], + "lines": [ + { + "bbox": [ + 350, + 186, + 503, + 195 + ], + "spans": [ + { + "bbox": [ + 350, + 186, + 503, + 195 + ], + "type": "text", + "content": "(a) Three ROIs from Brain data; compression ratio: " + }, + { + "bbox": [ + 350, + 186, + 503, + 195 + ], + "type": "inline_equation", + "content": "\\sim 87\\times" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 197, + 543, + 311 + ], + "blocks": [ + { + "bbox": [ + 309, + 197, + 543, + 311 + ], + "lines": [ + { + "bbox": [ + 309, + 197, + 543, + 311 + ], + "spans": [ + { + "bbox": [ + 309, + 197, + 543, + 311 + ], + "type": "image", + "image_path": "f3c7e3676c9cb7992d190149b03dfe6d4882c7275262be0b10d320684534945c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 312, + 504, + 319 + ], + "lines": [ + { + "bbox": [ + 350, + 312, + 504, + 319 + ], + "spans": [ + { + "bbox": [ + 350, + 312, + 504, + 319 + ], + "type": "text", + "content": "(b) Three ROIs from Heart data; compression ratio: " + }, + { + "bbox": [ + 350, + 312, + 504, + 319 + ], + "type": "inline_equation", + "content": "\\sim 87\\times" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 322, + 537, + 331 + ], + "lines": [ + { + "bbox": [ + 309, + 322, + 537, + 331 + ], + "spans": [ + { + "bbox": [ + 309, + 322, + 537, + 331 + ], + "type": "text", + "content": "Ground Truth TINC (ours) SCI NeRV HEVC SGA+BB" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "type": "text", + "content": "Figure 3. Visual comparisons of decompressed organs slices from Brain and Heart data between TINC and other benchmark methods under similar compression ratio around " + }, + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "type": "inline_equation", + "content": "87\\times" + }, + { + "bbox": [ + 305, + 346, + 545, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "text", + "content": "pression ratios. As shown in Fig. 4(a), the performance of TINC was evaluated in terms of SSIM on 8 medical data, under " + }, + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "inline_equation", + "content": "256 \\times" + }, + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "text", + "content": " compression ratio. To better demonstrate the effect of increasing tree levels, we used the growth rate of SSIM, calculated by normalizing all SSIM values by the one when " + }, + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 304, + 402, + 545, + 521 + ], + "type": "text", + "content": ". The SSIM scores on all the test data have a steadily increasing growth rate (dashed lines) as the number of tree depth increases from 1 to 3. Experimental results on other compression ratios show consistent trend (supplementary figure 2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "We further investigated the relationship between the data complexity and the performance gain when increasing tree levels. For data complexity, we adopt the spectrum concentration defined in [41], which measures the percentage of high frequency information contained in the data. The results are plotted in Fig. 4(b), with the solid line being a quadratic fitting of the growth rate for the 8 medical data using least squares. From the plot one can see that the performance gain increases monotonously with data complexity. In other words, the benefit of using a large tree is more significant when the target data contains more high-frequency information. This trend is consistent with our reasoning in Sec. 4 that a large " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " favours parameter sharing at coarser levels and leave more parameters for representing the details. Experimental results on other compression ratios and other data are consistent with this (supplementary figure 3)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18522" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 157, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 157, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 157, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 157, + 178 + ], + "type": "image", + "image_path": "ded2a7b60ff6b048a82e6e6bf217bd846cb69b4211ffe5361f31475e0f1eec0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 183, + 110, + 191 + ], + "lines": [ + { + "bbox": [ + 98, + 183, + 110, + 191 + ], + "spans": [ + { + "bbox": [ + 98, + 183, + 110, + 191 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 73, + 284, + 178 + ], + "blocks": [ + { + "bbox": [ + 177, + 73, + 284, + 178 + ], + "lines": [ + { + "bbox": [ + 177, + 73, + 284, + 178 + ], + "spans": [ + { + "bbox": [ + 177, + 73, + 284, + 178 + ], + "type": "image", + "image_path": "1a34db492e70e691d1a0885f417450fdcb15e8473910ba19f8d09b57960200e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 182, + 237, + 192 + ], + "lines": [ + { + "bbox": [ + 225, + 182, + 237, + 192 + ], + "spans": [ + { + "bbox": [ + 225, + 182, + 237, + 192 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "type": "text", + "content": "Figure 4. Effect of tree levels " + }, + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "type": "text", + "content": " on TINC's compression fidelity, in terms of SSIM on 8 medical data under " + }, + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "type": "inline_equation", + "content": "256 \\times" + }, + { + "bbox": [ + 46, + 201, + 287, + 301 + ], + "type": "text", + "content": " compression ratio. (a) The growth rate of SSIM for each data when increasing the tree levels. The dashed lines represent the change in each data. The solid line represents the mean of changes, and the filled area represents the standard deviation. (b) The scatter plot of each data's complexity and the growth rate of SSIM when increasing tree levels from 1 to 3. The solid line represents the trend of change, estimated by a quadratic fit." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "text", + "content": "In addition, we studied the performance gain of using a deep tree structure under limited bandwidth budget " + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "inline_equation", + "content": "|\\Theta|" + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "text", + "content": ", e.g., at high compression ratio. Supplementary figure 4 plots the performance improvement at increasing compression ratios, calculated from 8 medical data volumes. Almost all data have a smaller growth rate of SSIM or PSNR when compression ratio increases, i.e. reducing the number of available parameters. At high compression ratio " + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "inline_equation", + "content": "(512 \\times)" + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "text", + "content": ", some data are even of inferior fidelity using a deep tree, which is in line with our analysis in Sec. 4 that one should set a proper " + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 320, + 287, + 464 + ], + "type": "text", + "content": " for choosing the right partitioning granularity under the constraint from available budget." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": "Inter-level Parameter Allocation. We tested the performance at different parameter allocation proportions across levels on medical data under different compression ratios. Assuming there are " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "K^l" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": " nodes at " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": "th level, we denote the number of allocated parameters at this level as " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "|\\Theta^l|" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": " and ratio with respect to that of its adjacent upper level as " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "r^l = \\frac{|\\Theta^l| / K^l}{|\\Theta^{l-1}| / K^{l-1}}" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": ". Simply speaking, " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "r^l = 1" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": " means even inter-level allocation, a " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "r^l" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": " larger than 1 means imposing higher priority to the shallow levels and vice versa. We let " + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "inline_equation", + "content": "r^l" + }, + { + "bbox": [ + 46, + 466, + 287, + 604 + ], + "type": "text", + "content": " of all levels take 0.8, 1 and 1.2 to represent the three typical cases of parameter allocation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "We first investigated the effect of allocating more parameters to the shallow level, intending to share parameters among far apart similar blocks. As shown in Fig. 5(a), we divided a Brain data into 64 sub-regions, laid out in z-curve order, and selected 8 equally spaced distant regions. Then, we calculated the SSIM of these regions with respect to each other and used min-max normalization to obtain their non-local similarities. It can be seen that #1 and #33 regions are of high similarity. For convenient display, we also draw the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 73, + 405, + 178 + ], + "blocks": [ + { + "bbox": [ + 315, + 73, + 405, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 73, + 405, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 405, + 178 + ], + "type": "image", + "image_path": "91a36dac14b05d35ba23d46b1725b495453af2222d405a136f09e9c3a8a053d4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 183, + 368, + 191 + ], + "lines": [ + { + "bbox": [ + 358, + 183, + 368, + 191 + ], + "spans": [ + { + "bbox": [ + 358, + 183, + 368, + 191 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 201, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 289 + ], + "type": "text", + "content": "Figure 5. Effect of allocating more parameters to the shallow level on TINC's compression fidelity for each sub-regions in a Brain data under " + }, + { + "bbox": [ + 304, + 201, + 545, + 289 + ], + "type": "inline_equation", + "content": "512 \\times" + }, + { + "bbox": [ + 304, + 201, + 545, + 289 + ], + "type": "text", + "content": " compression ratio. All sub-figures share the same y-axis label. (a) The heatmap of similarities between 8 equally spaced distant sub-regions. The serial numbers of the regions represent their z-curve order. (b) The non-local similarity of each region. (c) The growth rate of SSIM for each region when allocating more parameters to the shallow level." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 76, + 481, + 178 + ], + "blocks": [ + { + "bbox": [ + 427, + 76, + 481, + 178 + ], + "lines": [ + { + "bbox": [ + 427, + 76, + 481, + 178 + ], + "spans": [ + { + "bbox": [ + 427, + 76, + 481, + 178 + ], + "type": "image", + "image_path": "6eda077fe1153665b0fa6ec45599548cc30578461a2d7e857475a514a2d6c240.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 183, + 459, + 191 + ], + "lines": [ + { + "bbox": [ + 448, + 183, + 459, + 191 + ], + "spans": [ + { + "bbox": [ + 448, + 183, + 459, + 191 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 492, + 76, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 492, + 76, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 492, + 76, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 492, + 76, + 541, + 178 + ], + "type": "image", + "image_path": "8202af729ba2662701429c6056c779272874c2d12a347b38bb0e3d13f3145945.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 182, + 522, + 191 + ], + "lines": [ + { + "bbox": [ + 512, + 182, + 522, + 191 + ], + "spans": [ + { + "bbox": [ + 512, + 182, + 522, + 191 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 306, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 354 + ], + "type": "text", + "content": "bar plot of the normalized marginal sums of the similarity matrix in Fig. 5(b), which tells that #1 and #33 are both of high similarity to other blocks, while #25 and #57 regions are highly distinctive from others." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 354, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 510 + ], + "type": "text", + "content": "To quantitatively measure the performance change at different blocks when allocating more parameters to the shallow level, we plot the growth rate of SSIM under " + }, + { + "bbox": [ + 304, + 354, + 546, + 510 + ], + "type": "inline_equation", + "content": "512 \\times" + }, + { + "bbox": [ + 304, + 354, + 546, + 510 + ], + "type": "text", + "content": " ratio, as shown in Fig. 5(c). Analyzing jointly with (b), we can find that the two regions with high non-local similarity, #1 and #33, gain a large boost, while those with the lowest similarity, #25 and #57, suffer a loss of fidelity. This is consistent with the intuition and analysis in Sec. 4 that allocating more parameters to the nodes at shallow levels would benefit representing distant but similar regions, but at expense of slight degeneration at very individual regions. Experimental results on other compression ratios and other data are shown in supplementary figures 5 and 6." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "content": "We then investigated how to allocate parameters across levels according to the global consistency of the medical data. We divided the target data into 64 sub-regions, calculated the marginal sums of the similarity matrix. We change the setting of " + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "inline_equation", + "content": "r^l" + }, + { + "bbox": [ + 304, + 510, + 545, + 640 + ], + "type": "text", + "content": " and the results are shown in the supplementary figure 7. The results show allocating more parameters to the shallow level is more likely to improve fidelity when the global consistency of the data is greater than 0.7, and allocating more to the deep level is better when the global consistency lower than 0.6. Otherwise, an even allocation may be a better choice." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "Intra-level Parameter Allocation. We tested the approach of allocating parameters to the nodes at the same level according to their importance proposed in Sec. 4, on biological data and under different compression ratios. Specifically, we let the number of allocated parameters be proportional to their importance. For example, with regard" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18523" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "to the biological data such as Neurons the importance can be measured by the percentage of the number of valid signals, where the valid signal can be determined either by the result of traced neurons or by the intensity value. Here we chose the latter for simplicity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 192 + ], + "type": "text", + "content": "We first compare the compression fidelity of 6 Neurons data using the above two parameter allocation approaches under different compression ratios, as shown in supplementary figure 8. It can be seen that important allocation significantly improves the fidelity compared to even allocation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "type": "text", + "content": "We then look further into the performance at all the sub-regions in the same target data. As shown in Fig. 6(a), we divided a Neurons data into 64 sub-regions and compute their growth rate of Acc.2000 when switching from even allocation to important allocation. It should be noted that we omitted those regions that have reached 1 (the beast value) under both allocation approaches in the subsequent experiments. From the scatter diagram we can see three groups of data: (i) Some regions show non-significant change (less than " + }, + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "type": "text", + "content": " of the maximum change of the others, denoted as cross markers) using important allocation under " + }, + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "type": "inline_equation", + "content": "64\\times" + }, + { + "bbox": [ + 46, + 192, + 286, + 418 + ], + "type": "text", + "content": " compression ratio, because they are already with sufficient parameters. (iii) Other regions show significant improvement after introducing important allocation, denoted with circles. We performed a linear fit using least squares, as shown by the solid line. Overall, allocating parameters according to importance is helpful for the compression fidelity of important regions while at expenses of little fidelity degeneration in other less important ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 286, + 550 + ], + "type": "text", + "content": "For a statistical analysis, we plotted the Acc.2000 of all 64 sub-regions under even and important allocation, as shown in Fig. 6(b). We found that importance based allocation can significantly improve those regions that have very poor compression fidelity under even allocation. This is mainly because that those regions with poor fidelity tend to contain more complex structures (with more neural signals), and thus are of higher importance scores and thus allocated more parameters accordingly. Experimental results on other compression ratios are detailed in supplementary figure 9." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 562, + 125, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 125, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 125, + 574 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Summary and Discussions. In this paper we propose a hierarchical parameter sharing mechanism for INR and implement it based on a tree-structured MLP for large and complex data compression tasks. The proposed approach TINC makes use of the local redundancy via block wise compression with separate MLPs and non-local redundancy via organizing these MLPs in tree structure to share parameters hierarchically. Such implementation forms a single network for the complex target data, not only incorporate the similarity distributed far apart and nonuniformly, and avoid discontinuity between adjacent regions as well." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 417, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 417, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 417, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 417, + 178 + ], + "type": "image", + "image_path": "46013685926776d1b8ad2ff8cea383ee9f9cd81ca3d9d1549efd4869b79cc59b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 183, + 368, + 191 + ], + "lines": [ + { + "bbox": [ + 358, + 183, + 368, + 191 + ], + "spans": [ + { + "bbox": [ + 358, + 183, + 368, + 191 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 545, + 178 + ], + "type": "image", + "image_path": "8b765ba7b63ae12e00243303b71fd684e9b9d720be18c02d882b6ffb64260a0a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 182, + 495, + 191 + ], + "lines": [ + { + "bbox": [ + 484, + 182, + 495, + 191 + ], + "spans": [ + { + "bbox": [ + 484, + 182, + 495, + 191 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 201, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 312 + ], + "type": "text", + "content": "Figure 6. Effect of important allocation on TINC's compression fidelity, in terms of Acc.2000 on a Neurons data under " + }, + { + "bbox": [ + 304, + 201, + 545, + 312 + ], + "type": "inline_equation", + "content": "64 \\times" + }, + { + "bbox": [ + 304, + 201, + 545, + 312 + ], + "type": "text", + "content": " compression ratio. (a) The growth rate of Acc.2000 for each region. The regions with and without significant growth rate are represented as cross and square respectively. The solid line represents the trend of change, estimated by a linear fit. (b) The violin plots of Acc.2000 for all these regions. The shaded area represents the distribution of Acc.2000 across all regions. The three horizontal lines from top to bottom represent the maximum, average and minimum values respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 426 + ], + "type": "text", + "content": "Extensive experiments show that this hierarchical parameter sharing mechanism introduced in TINC greatly improves the compression capability and wide applicability of INR. Our method outperforms state-of-the-art compressors including commercial tools (JPEG, H.264, HEVC), data-driven methods (DVC, SSF, SGA+BB), and existing INR based ones (SCI, NeRV, and NeRF) on diverse medical and biological data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": "We also analyze TINC's flexibility to different cases via experimentally studying the effect of three key settings—number of tree levels, intra-level and inter-level parameter allocation methods, and provide highlights on proper parameter setting and strategy selection according to the features of target data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": "Limitations and Future Extensions. Similar to all current INR based compression methods, TINC is of high decompression speed but slow in compression, since it takes time to pursue the MLPs matching the target data. We plan to combine meta-learning to find the best initialization parameters for each organ to speed up TINC." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 589, + 408, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 589, + 408, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 408, + 601 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 610, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 646 + ], + "type": "text", + "content": "We acknowledge MoE Key Laboratory of Biomedical Photonics (Huazhong University of Science and Technology) for sharing their mouse brain-wide microscopic data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 647, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 712 + ], + "type": "text", + "content": "This work is supported by Beijing Natural Science Foundation (Grant No. Z200021), the National Natural Science Foundation of China (Grant Nos. 61931012, 62088102) and Project of Medical Engineering Laboratory of Chinese PLA General Hospital (Grant No. 2022SYSZ-ZKY21)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18524" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Arya Aftab, Alireza Morsali, and Shahrokh Ghaemmaghami. Multi-Head Relu Implicit Neural Representation Networks. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2510-2514, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "text", + "content": "[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 246 + ], + "type": "text", + "content": "[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational Image Compression with a Scale Hyperprior. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[4] Nuri Benbarka, Timon Höfer, Hamd ul-moqeet Riaz, and Andreas Zell. Seeing Implicit Neural Representations as Fourier Series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2283-2292, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "text", + "content": "[5] Alexander William Bergman, Petr Kellnhofer, and Gordon Wetzstein. Fast Training of Neural Lumigraph Representations using Meta Learning. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "text", + "content": "[6] Alexandre Boulch and Renaud Marlet. Poco: Point Convolution for Surface Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[7] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. NeRV: Neural Representations for Videos. Advances in Neural Information Processing Systems, 2021. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[8] Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 512 + ], + "type": "text", + "content": "[9] Yilun Du, Katherine Collins, Josh Tenenbaum, and Vincent Sitzmann. Learning Signal-agnostic Manifolds of Neural Fields. Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 546 + ], + "type": "text", + "content": "[10] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J. Zico Kolter. Multiplicative Filter Networks. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "text", + "content": "[11] Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T Freeman, and Thomas Funkhouser. Learning Shape Templates with Structured Implicit Functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7154-7164, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "type": "text", + "content": "[12] Milad Soltany Kadarvish, Hesam Mojtahedi, Hossein Entezari Zarch, Amirhossein Kazerouni, Alireza Morsali, Azra Abtahi, and Farokh Marvasti. Ensemble Neural Representation Networks. arXiv preprint arXiv:2110.04124, 2021. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980, 2014. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[14] Sylwester Klocek, Lukasz Maziarka, Maciej Wolczyk, Jacek Tabor, Jakub Nowak, and Marek Śmieja. Hypernetwork" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Functional Image Representation. In International Conference on Artificial Neural Networks, pages 496-510, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 160 + ], + "type": "text", + "content": "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 161, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 161, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 547, + 216 + ], + "type": "text", + "content": "[16] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: An End-to-end Deep Video Compression Framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "type": "text", + "content": "[17] Y. Lu, K. Jiang, J. A. Levine, and M. Berger. Compressive Neural Representations of Volumetric Scalar Fields. Computer Graphics Forum, 40(3):135-146, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 251, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 294 + ], + "type": "text", + "content": "[18] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive Coordinate Networks for Neural Scene Representation. ACM Trans. Graph., 40(4), jul 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "text", + "content": "[19] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 425 + ], + "type": "text", + "content": "[20] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated Periodic Activations for Generalizable Local Functional Representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, October 2021. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 481 + ], + "type": "text", + "content": "[21] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[22] David Minnen, Johannes Balle, and George D Toderici. Joint Autoregressive and Hierarchical Priors for Learned Image Compression. In Advances in Neural Information Processing Systems, volume 31, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph., 41(4), jul 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing Scenes as Compositional Generative Neural Feature Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "type": "text", + "content": "[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 714 + ], + "type": "text", + "content": "[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18525" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[27] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional Occupancy Networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "type": "text", + "content": "[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 260 + ], + "type": "text", + "content": "[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned Implicit Function for High-resolution Clothed Human Digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "text", + "content": "[30] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit Neural Representations with Periodic Activation Functions. In Advances in Neural Information Processing Systems, 2020. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene Representation Networks: Continuous 3d-structure-aware Neural Scene Representations. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "text", + "content": "[32] Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi. The JPEG 2000 Still Image Compression Standard. IEEE Signal Processing Magazine, 18(5):36-58, 2001. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 458 + ], + "type": "text", + "content": "[33] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the High Efficiency Video Coding (HEVC) Standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 525 + ], + "type": "text", + "content": "[34] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural Geometric Level of Detail: Real-time Rendering with Implicit 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 591 + ], + "type": "text", + "content": "[35] Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned Initializations for Optimizing Coordinate-based Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "text", + "content": "[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. In Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[37] Gregory K Wallace. The JPEG Still Picture Compression Standard. IEEE Transactions on Consumer Electronics, 38(1):xviii-xxxiv, 1992. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[38] CL Walsh, P Tafforeau, WL Wagner, DJ Jafree, A Bellier, C Werlein, MP Kühnel, E Boller, S Walker-Samuel, JL Rober" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 418 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "tus, et al. Imaging Intact Human Organs with Local Resolution of Cellular Structures using Hierarchical Phase-Contrast Tomography. Nature Methods, 18(12):1532-1541, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[39] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the H.264/AVC Video Coding Standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "text", + "content": "[40] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for Video Compression with Hierarchical Quality and Recurrent Enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6628-6637, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[41] Runzhao Yang, Tingxiong Xiao, Yuxiao Cheng, Qianni Cao, Jinyuan Qu, Jinli Suo, and Qionghai Dai. SCI: A Spectrum Concentrated Implicit Neural Compression for Biomedical Data. arXiv preprint arXiv:2209.15180, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[42] Ruihan Yang, Yibo Yang, Joseph Marino, and Stephan Mandt. Hierarchical Autoregressive Modeling for Neural Video Compression. In ICLR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[43] Yibo Yang, Robert Bamler, and Stephan Mandt. Improving Inference for Neural Image Compression. In Advances in Neural Information Processing Systems, volume 33, pages 573-584, 2020. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 375 + ], + "type": "text", + "content": "[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields from One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 418 + ], + "type": "text", + "content": "[45] Ellen D Zhong, Tristan Bepler, Joseph H Davis, and Bonnie Berger. Reconstructing Continuous Distributions of 3D Protein Structure from Cryo-EM Images. arXiv preprint arXiv:1909.05215, 2019. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18526" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_content_list.json b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..52169ad8123bf409011b623bd5f10562f1ad731d --- /dev/null +++ b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "TIPI: Test Time Adaptation with Transformation Invariance", + "text_level": 1, + "bbox": [ + 179, + 130, + 790, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A. Tuan Nguyen1, Thanh Nguyen-Tang2, Ser-Nam Lim3*, Philip H.S. Torr1*", + "bbox": [ + 321, + 180, + 648, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Oxford, 2Johns Hopkins University, 3Meta AI", + "bbox": [ + 246, + 215, + 722, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tuan@robots.ox.ac.uk, nguyent@cs.jhu.edu", + "bbox": [ + 310, + 237, + 658, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sernamlim@meta.com, philip.torr@eng.ox.ac.uk", + "bbox": [ + 292, + 255, + 673, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When deploying a machine learning model to a new environment, we often encounter the distribution shift problem – meaning the target data distribution is different from the model's training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distribution can affect the model's performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a surrogate loss on the test-time unlabeled target data. In particular, minimizing the prediction's entropy on target samples [34] has received much interest as it is task-agnostic and does not require altering the model's training phase (e.g., does not require adding a self-supervised task during training on the source domain). However, as the target data's batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not optimal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an invariance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model's performance under input transformations. The resulting method (TIPI – Test tIme adaPtation with transformation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT [34] in all settings. Our code is released at https://github.com/atuannguyen/TIPI.", + "bbox": [ + 76, + 343, + 472, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 304, + 630, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distribution shift is a common problem and is often faced in real-world applications. Specifically, despite taking various precautions while training a machine learning model to ensure a better generalization (e.g., collecting and training on multiple source domains [17, 27], finding flat minima [5], training with meta-learning objectives [16] etc.), the model often still struggles when the test data distribution shifts slightly. Note that it is also common not to have labels for the new shifted domain during test time. To tackle this problem, test time adaptation (also known as online domain adaptation) is a framework that allows the model to adapt to the target distribution using unlabeled test data batches. This is necessary since we typically do not have time to annotate the data during test time and can only make use of the unlabeled data. In this framework, the model needs to give predictions for the target data while simultaneously updating itself to improve its performance on that particular target distribution. We assume a situation in which the target data only arrive in small batches, which makes the adaptation task extremely challenging and renders traditional domain adaptation techniques such as representation alignment (via a distance metric) ineffective.", + "bbox": [ + 496, + 335, + 893, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Within this test time adaptation framework, a common and effective approach is to optimize a surrogate objective function in lieu of the true loss function on the target data. The first group of surrogate objectives is the loss functions of self-supervised tasks. In particular, one would formulate a user-defined task (such as predicting the rotation angle of an image) and train it alongside the main task on the source domain; and keep training the self-supervised task on the test-time target data [19, 33]. However, these are not fully test-time adaptation methods, since they require altering the training procedure of the source domain. The second line of surrogate objectives is unsupervised loss functions. Among this group of unsupervised objectives, entropy minimization (TENT) [34] is the most successful method, and has been shown to be consistent across many benchmarks. Furthermore, different from the former group,", + "bbox": [ + 496, + 664, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The last two authors contributed equally", + "bbox": [ + 94, + 886, + 315, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24162", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TENT is a fully test-time adaptation method. For these reasons, TENT has received much interest and a lot of follow-up papers/discussions.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as pointed out by its authors, TENT is not robust when using small batch sizes, as it often collapses to a trivial solution (i.e., it always predicts the same class for all input). This is detrimental to real-world applications since test data often arrive in small batches. For example, autonomous driving systems only process a few frames in real time – they typically do not accumulate the frames (let's say within one minute) to form a bigger batch. Note that TENT is previously evaluated mainly for large batch sizes (such as 64,128 or 200).", + "bbox": [ + 75, + 137, + 467, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to tackle the aforementioned problem. Specifically, we aim to develop an unsupervised surrogate objective function for the test time adaptation problem such that it is task-agnostic, does not require altering the training procedure (e.g., does not require incorporating a self-supervised task into the training process), and is more resilient against small batch sizes.", + "bbox": [ + 75, + 286, + 467, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first provide theoretical results regarding a model's performance under input transformations. Specifically, we show that a model's loss on a data distribution is bounded by the KL distance on the predictive distribution of the data before and after the transformations (which we will use as a regularizer), and its loss on the transformed data distribution. Motivated by this result, we use small shifts in the input images that can simulate real source-target shifts, and enforce the network to be invariant under such data transformations. Our model outperforms TENT (and other relevant baselines) in all problem settings considered in the paper and is remarkably robust in the small-batch-size regime.", + "bbox": [ + 75, + 391, + 467, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions in this paper are threefold:", + "bbox": [ + 96, + 569, + 392, + 583 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide theoretical results regarding a model's performance under transformations of the input data. Specifically, a model's performance on the target domain is bounded by an invariance term (maximum KL divergence of the predictive distributions before and after the transformations) and its loss on the transformed domain.", + "- We propose to find input transformations that can simulate the domain shifts, and enforce the network to be invariant under such transformations, using a regularizer based on our derived bound. The resulting method is TIPI (Test tIme adaPtation with transformation Invariance).", + "- We perform extensive experiments on a wide range of datasets (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17) and settings (varying batch sizes) to validate our method. TIPI shows preferable performance compared to relevant baselines." + ], + "bbox": [ + 96, + 599, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 648, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distribution Shift is an important and commonly faced problem when deploying machine learning models in real-world applications. There are two main approaches that try to alleviate this problem. The first setting is domain generalization [1, 17, 23, 26, 27], where one aims to collect multiple source domains and learn a generalizable model based on those domains. Another setting is (unsupervised) domain adaptation [2, 6, 18, 21, 28], where a model is trained on a labeled source dataset and we attempt to adapt that model to a target distribution with the help of unlabeled target data. This paper focuses on the latter setting, in particular, the test-time domain adaptation variation.", + "bbox": [ + 496, + 113, + 890, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test Time Domain Adaptation: In this setting, the unlabeled target data is not available during the training phase of the source model. Instead, the model only has access to the target data when deployed in the target environment, as the stream of test data batches. As a consequence, the unlabeled target data is not available in large quantities, but only in small batches. This makes the adaptation task extremely challenging. So far, there are two main lines of work dealing with this online adaptation problem. The first approach is based on the batch normalization layer in the network, with the idea of using the test-time batch statistics for these layers instead of the source statistics [24, 32]. This can partially help with the representation misalignment between the source and target distributions but is only applicable when the network has BatchNorm layers and the batch size is large. The other line of work is to use an unsupervised surrogate loss to \"refine\" the model on the target data. For example, TENT [34] minimizes the prediction entropy of the model on the target data. However, as pointed out by its authors, TENT needs a large batch size to avoid collapsing to a trivial prediction (only output one class), due to the characteristics of this surrogate objective. Meanwhile, TTT [33] and TTT++ [19] train a self-supervised task alongside the main task on the source domain, and continue to train the self-supervised task during the online adaptation phase. Therefore, TTT and TTT++ are not fully off-the-shelf test-time adaptation methods, since they require modifications to the training phase of the source domain (to incorporate the self-supervised task). Recently, [8] generalizes TENT to other source loss functions (beyond the typical cross-entropy loss). There also have been several works [29] that study datapoints selection strategies for optimizing the unsupervised surrogate loss (e.g., TENT): only optimize the loss for reliable and informative datapoints. Note that this line of research is orthogonal/complementary to ours and we believe that our surrogate loss will also benefit from similar datapoints selection strategies. We leave exploring this direction to future work. Using a completely different approach, T3A [12] replaces the last linear layer (classification layer) with a nonparametric classifier and keeps updating the support set of each class with the new target", + "bbox": [ + 496, + 297, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24163", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datapoints. T3A is robust against small batch sizes (since it is not optimization-based), but is not as effective as other optimization-based methods (e.g., TENT, TTT, and ours).", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Connection between Adversarial Robustness and Robustness against Distribution shift: It has been observed that adversarily trained models generally perform better under certain types of domain shifts (for example, see the leaderboard of Robustbench [4] for Cifar10-C and Cifar100-C). Indeed, one might expect this to be the case if the space of adversarial attacks matches the space of distribution shifts. While it is not previously well understood, our paper sheds some light on this phenomenon. Note that different from previous works, our method enforces an unsupervised robustness loss during the online adaptation phase, not at the source training stage.", + "bbox": [ + 75, + 142, + 468, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 330, + 184, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem setting", + "text_level": 1, + "bbox": [ + 76, + 356, + 233, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume that we have a source domain with the distribution $p_S(x,y)$ , where $x \\in \\mathcal{X}$ is the input and $y \\in \\mathcal{Y}$ is the prediction target. Similar to prior works, we consider the classification problem, where $\\mathcal{Y} = \\{1,2,\\dots,C\\}$ (however, the approach can be straightforwardly extended to regression problems). Suppose that using data from the source domain, we already successfully trained a model $f_{\\theta}: \\mathcal{X} \\to \\mathbb{R}^{C}$ , where $\\theta$ is the parameter of the neural network and $f_{\\theta}(x)$ outputs the prediction probabilities for the $C$ classes $(\\|f_{\\theta}(x)\\|_1 = 1$ and $0 \\leq f_{\\theta}(x)_i \\leq 1, \\forall i \\in \\{1,2,\\dots,C\\})$ . This leads to a predictive distribution $\\hat{p}_{\\theta}(y|x)$ , which is the categorical distribution where $\\hat{p}_{\\theta}(y = i|x) = f_{\\theta}(x)_i$ . Note that, for our method, we will not assume or require any modifications to the training process of the model $f_{\\theta}$ on the source domain - a setting that is common and useful (e.g., a Resnet50 model downloaded directly from torchvision).", + "bbox": [ + 75, + 378, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After that, this model is deployed to the target environment with data distribution $p_T(x,y)$ . The model receives target data as a stream of unlabeled mini-batches (probably with small and varying sizes). The task is to adapt to this environment simultaneously and continually with only the unlabeled mini-batches.", + "bbox": [ + 75, + 628, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The expected loss of this model on a data distribution $q(x,y)$ (for example, $q(x,y) = p_T(x,y)$ ) is:", + "bbox": [ + 75, + 715, + 468, + 748 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\theta , q (x, y)) = \\mathbb {E} _ {q (x)} \\left[ l [ q (y | x), \\hat {p} _ {\\theta} (y | x) ] \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 756, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $l$ is a loss function (distance/divergence) between the true labeling distribution $q(y|x)$ and the predictive distribution $\\hat{p}_{\\theta}(y|x)$ . This loss is not actually computed in practice for both the source and target data (we simply do not have labeled data for both of those), and it is used for theoretical discussion only. To make the theoretical discussion simpler, in this paper, we will consider the simple $l_{1}$ distance between the two distributions. However,", + "bbox": [ + 75, + 782, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "note that our theoretical discussion also holds for other divergences such as the Jensen-Shannon divergence (or symmetric KL). Both of these distances satisfy the triangle inequality, which we use in our derivation.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Test Time Domain Adaptation with Transformation Invariance", + "text_level": 1, + "bbox": [ + 498, + 186, + 890, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall that the task is to gradually update the network parameter $\\theta$ so that the model can adapt to the target domain. In this setting, we do not have access to the source data nor a large amount of target data (only small batches), so traditional domain adaptation approaches such as representation alignment are often infeasible.", + "bbox": [ + 496, + 224, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given access only to the batches of unlabeled target data, it is natural to try to come up with a surrogate objective to optimize in lieu of the true loss function (which is not available without the labels). In this sense, this has a close resemblance to the semi-supervised learning problem. In both settings, we aim to find a surrogate function to optimize for unlabeled data. In fact, both the entropy-minimization and pseudo-labeling techniques have been borrowed from the semi-supervised literature, with varying degrees of success. It should be noted that, in semi-supervised learning, the surrogate loss on the unlabeled data is minimized jointly with the supervised objective for labeled data, thus model collapsing is often not a problem. It is much more challenging, however, in this online adaptation setting, since the unsupervised surrogate loss is optimized alone, which can lead to the catastrophic collapse of the network parameter (such that it can not perform prediction anymore, even for the source domain). For example, as pointed out by its authors, TENT [34] tends to collapse to a trivial model if the batch size is small, which is often the case in practice.", + "bbox": [ + 496, + 313, + 892, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we examine the invariance regularization technique in the setting of test time domain adaptation. Note that a similar regularizer ([3,22,31]) has been applied to semi-supervised learning. However, to the best of our knowledge, we are the first to investigate this type of regularization in the online domain adaptation problem. We aim to investigate both theoretically and empirically how this technique can help with adapting to the new target domain. We also find that this surrogate loss is much more robust (against small batch sizes) compared to TENT.", + "bbox": [ + 496, + 608, + 890, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, we propose to enforce the network to be invariant under a predefined space of transformations on the data on the target domain. The idea is that this space of transformations contains the actual domain shifts, or at least could simulate them. Specifically, we minimize $\\theta$ with the following objective:", + "bbox": [ + 496, + 753, + 890, + 843 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {p _ {T} (x)} \\left[ \\max _ {x ^ {\\prime} \\in A (x)} \\mathrm {K L} \\left[ \\hat {p} _ {\\theta} \\left(y \\mid x ^ {\\prime}\\right), \\hat {p} _ {\\bar {\\theta}} (y \\mid x) \\right] \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 842, + 890, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\overline{\\theta}$ is an identical copy of $\\theta$ , and $A(x) = \\{t(x)\\}_{t\\in S_t}$ .", + "bbox": [ + 500, + 883, + 888, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24164", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Adapt and predict. $\\theta$ is the network parameter, $\\{x_{i}\\}_{i = 1}^{n}$ is the batch data, $\\alpha$ is the learning rate for online adaptation", + "bbox": [ + 76, + 90, + 468, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input $\\theta ,\\{x_{1},x_{2},\\ldots ,x_{n}\\}$", + "bbox": [ + 84, + 138, + 279, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output $\\theta_{\\mathrm{new}}$ (newly adapted), $\\{\\hat{y}_1,\\hat{y}_2,\\dots ,\\hat{y}_n\\}$", + "bbox": [ + 86, + 154, + 403, + 170 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: function ADAPT-PREDICT $(\\theta ,\\{x_{1},x_{2},\\dots,x_{n}\\})$", + "2: $p_i \\gets f_\\theta(x_i)$ . detach() ▷ stop gradient", + "3: $x_{i}^{\\prime}\\gets \\arg \\max_{x^{\\prime}\\in A(x_{i})}\\mathrm{KL}[p_{\\theta}(y|x^{\\prime}),p_{i}]$", + "4: $l(\\theta) = \\frac{1}{n}\\sum_{i=1}^{n}\\mathrm{KL}[p_{\\theta}(y|x_i'), p_i]$", + "5: $\\theta_{\\mathrm{new}} = \\dot{\\theta} -\\alpha \\nabla_{\\theta}l(\\theta)$", + "6: $\\hat{y}_i = \\arg \\max f_{\\theta_{\\mathrm{new}}}(x_i)$", + "7: end function" + ], + "bbox": [ + 86, + 171, + 468, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(S_{t}$ is the predefined set of transformations).", + "bbox": [ + 76, + 303, + 369, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the motivations and derivations of this objective function in the next subsection. The adapting and predicting procedure is depicted in Algorithm 1.", + "bbox": [ + 73, + 316, + 468, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Invariance Regularization Technique", + "text_level": 1, + "bbox": [ + 76, + 378, + 387, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we introduce the invariance regularization technique for general transformations and provide theoretical results regarding the online domain adaptation problem, which will later motivate our method.", + "bbox": [ + 75, + 402, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let's consider a set of transformations $S_{t}$ , where each $t \\in S_{t}$ is a transformation of the data, i.e., $t: \\mathcal{X} \\to \\mathcal{X}$ . Ideally, $S_{t}$ should contain the transformations responsible for the domain shift. This set can contain transformations such as data augmentations, GAN-like cross-domain transformations, random noise perturbation, etc. For now, we shall not care much about the type of transformations in $S_{t}$ . The only important detail is that these transformations should preserve the class information (this is formalized in Assumption 1). Let $A(x) = \\{t(x)\\}_{t \\in S_t}$ and let $p_t(t)$ be a distribution over $S_{t}$ that assigns probability mass (or density) to each transformation $t \\in S_{t}$ . Each distribution $p_t(t)$ defines a conditional distribution $T(x^t | x)$ (with the support set $A(x)$ ), where $x^t = t(x), t \\sim p_t(t)$ . This in turn defines a joint distribution over $x, x^t$ and $y$ as:", + "bbox": [ + 75, + 462, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {T} (x, x ^ {t}, y) = p _ {T} (x, y) T \\left(x ^ {t} | x\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 686, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is also worth noting that we have transformed data distribution $p_T(x^t, y)$ of $x^t$ and $y$ (just marginalizing out $x$ from the above).", + "bbox": [ + 75, + 714, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 1.", + "text_level": 1, + "bbox": [ + 76, + 767, + 181, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {T} (x, y) = I _ {T} \\left(x ^ {t}, y\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 791, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{T}$ is the mutual information, calculated on the target domain (please refer to the Supplementary Material for their exact formula).", + "bbox": [ + 75, + 818, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, this assumption means that the transformed image $x^{t}$ contains the same amount of information about", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the label $y$ as the original image $x$ does. This assumption is equivalent to $\\mathbb{E}_{p_T(x,x^t)}[\\mathrm{KL}[p_T(y|x),p_T(y|x^t)]] = 0$ (a derivation can be found in the supplementary material), meaning that $p_T(y|x) = p_T(y|x^t)$ almost surely.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The transformation invariance regularizer on the target domain can be defined simply as:", + "bbox": [ + 498, + 154, + 890, + 184 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (\\theta , x) = \\max _ {x ^ {\\prime} \\in A (x)} D \\left[ \\hat {p} _ {\\bar {\\theta}} (y | x), \\hat {p} _ {\\theta} (y | x ^ {\\prime}) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 186, + 890, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D[\\cdot, \\cdot]$ is a divergence/distance of choice and $\\bar{\\theta}$ is an identical copy of $\\theta$ (so that the gradient is not backpropagated through it). We will see later (through an analysis of the gradient) that this is extremely important. Intuitively, it also helps to alleviate the network collapse problem, since, without this gradient blocking, the network would easily collapse to a constant output to minimize the regularizer term.", + "bbox": [ + 496, + 223, + 890, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, we can view this objective as using $\\hat{p}_{\\overline{\\theta}}(y|x)$ as the target to optimize $\\hat{p}_{\\theta}(y|t(x))$ for all transformation $t$ . This makes sense because, as discussed earlier, $t$ does not change the semantic meaning of $x$ significantly, so the predictive distribution of $x$ and $t(x)$ should be similar.", + "bbox": [ + 496, + 340, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this regularizer, we can bound the loss on the target domain as follows:", + "bbox": [ + 498, + 415, + 890, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. Under Assumption 1, if $D$ in Eq. 5 is either the forward or reverse KL divergence, we have:", + "bbox": [ + 498, + 452, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\theta , p _ {T} (x, y)) \\leq \\mathbb {E} _ {p _ {T} (x)} \\left[ \\sqrt {2 h (\\theta , x)} \\right] + \\ell (\\theta , p _ {T} (x ^ {t}, y)).\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 489, + 880, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Provided in the Supplementary Material.", + "bbox": [ + 500, + 523, + 815, + 537 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/902ea96dd6d39510cb0380d3819e426ff6444d283bdfd9027a42e19752302c8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 523, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This means that the loss on the target distribution is bounded by the invariance regularizer and the loss on the transformed distribution $p_T(x^t, y)$ .", + "bbox": [ + 496, + 547, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, assume for a moment (even if it is not realistic) that the set $S_{t}$ contains all the transformations that can successfully transform the target data distribution $p_{T}(x,y)$ back to the source distribution $p_{S}(x,y)$ (i.e., there exists a distribution $p_{t}(t)$ over $S_{t}$ such that the resulting transformed distribution $p_{T}(x^{t},y)$ is the same with $p_{S}(x,y)$ ). Then, we have that the loss on the target domain is bounded by the loss on the source domain and our invariance regularizer. Assume also that while adapting to the new target domain, our model does not suffer from catastrophic forgetting of the source domain (which is often the case if the network does not collapse since we are adapting to the same task in a different environment). Then we have that the loss on the source domain would be small, and the invariance regularizer would also be small (since we explicitly minimize it). As a consequence, the loss on the target domain would be small and we can guarantee a good adaptation performance.", + "bbox": [ + 496, + 590, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, knowing the exact transformations from the target domain back to the source domains is often infeasible (only possible if we have some prior domain knowledge). This motivates us to find other transformations that", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24165", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are task-agnostic and can simulate the distribution shift between the target and source domains.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Practical choice of the transformation set", + "text_level": 1, + "bbox": [ + 76, + 138, + 416, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We want to choose task-agnostic transformations such that they do not alter the semantic meaning of the image significantly, while they can simulate the distribution shift. The first idea that comes to mind is to use common image augmentations (e.g., random crop, random vertical flip, etc.) However, our early experiments suggest that this method is not consistent across corruption types/domain shifts. This makes sense since image augmentations are often heuristic, and they may be able to simulate certain types of distribution shifts, but not others. Therefore, we opt to use $l_{\\infty}$ -norm based perturbations. Empirically, this leads to better performance and also is consistent across domains/tasks.", + "bbox": [ + 75, + 161, + 468, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, we use the $l_{\\infty}$ -norm based perturbations as the transformations of images. Empirically, we also found out that using the reverse KL leads to better performance (although almost all distances/divergences have positive results). We provide explanations/intuitions as to why we should prefer the reverse KL, and why adversarial examples can help with the distribution shift problem in the next subsection. Our final optimization problem is:", + "bbox": [ + 75, + 347, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {p _ {T} (x)} \\left[ \\max _ {| | x ^ {\\prime} - x | | _ {\\infty} \\leq \\epsilon} \\mathrm {K L} \\left[ \\hat {p} _ {\\theta} \\left(y \\mid x ^ {\\prime}\\right), \\hat {p} _ {\\bar {\\theta}} (y \\mid x) \\right] \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 472, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We name our method TIPI (Test tIme adaPtation with transformation Invariance). To save computational cost, we use the single-step adversarial method FGSM [7] to find the adversarial example $x'$ . Note that one can also use more advanced methods such as PGD [20] to find the adversarial perturbation to get better results (but with the trade-off of speed). We set $\\epsilon$ to 2/255 in all of our experiments.", + "bbox": [ + 75, + 513, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Additional Discussions", + "text_level": 1, + "bbox": [ + 76, + 636, + 290, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Why preferring reverse KL over forward KL? Natural choices for $D$ are the KL-divergences (forward or reverse KL), since they can be computed straightforwardly for the predictive distributions, and also are commonly used in practice for similar problems (e.g., VAT [22] uses the forward KL). Our experiments empirically suggest that using the reverse KL as the divergence $D$ leads to better results compared to the forward KL. Here, we also give theoretical intuition as to why it is the case. First of all, let's consider the gradient of $\\theta$ with respect to both the forward and reverse KL:", + "bbox": [ + 75, + 657, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the forward KL:", + "bbox": [ + 76, + 820, + 212, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathrm {K L} [ \\hat {p} _ {\\bar {\\theta}} (y | x), \\hat {p} _ {\\theta} (y | x ^ {\\prime}) ] = \\sum_ {y \\in \\mathcal {Y}} - \\frac {\\hat {p} _ {\\bar {\\theta}} (y | x)}{\\hat {p} _ {\\theta} (y | x ^ {\\prime})} \\nabla_ {\\theta} \\hat {p} _ {\\theta} (y | x ^ {\\prime}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 842, + 468, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meanwhile, for the reverse KL:", + "bbox": [ + 500, + 90, + 709, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathrm {K L} [ \\hat {p} _ {\\theta} (y | x ^ {\\prime}), \\hat {p} _ {\\bar {\\theta}} (y | x) ] \\\\ = \\sum_ {y \\in \\mathcal {Y}} \\left(- \\log \\frac {\\hat {p} _ {\\bar {\\theta}} (y | x)}{\\hat {p} _ {\\theta} (y | x ^ {\\prime})} + 1\\right) \\nabla_ {\\theta} \\hat {p} _ {\\theta} (y | x ^ {\\prime}). \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 123, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compare Eq. 8 to Eq. 7, we can see that the coefficient of $\\nabla_{\\theta}\\hat{p}_{\\theta}(y|x^{\\prime})$ in the reverse KL gets squashed down by the log function, leading to much smaller variance. We conjecture that the small variance in gradient is the reason why reverse KL performs better than forward KL.", + "bbox": [ + 496, + 200, + 890, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "How can invariance enforcing with adversarial examples help with the distribution shift problem? In our previous subsection, we provide theoretical evidence that motivates us to find transformations that can simulate common distribution shifts in practice. We argue here that $l_{\\infty}$ adversarial perturbation is a good candidate for this. First of all, $l_{\\infty}$ adversarial perturbation is a task-agnostic transformation that does not require much tuning, and we do not need to design heuristic transformations per domain. Secondly, a lot of the common domain shifts in practice are image corruptions that add noises to the images (gaussian noise, gaussian blur, jpeg compression noise, etc.), which can be simulated very well by adversarial noises. Lastly, even for more global domain shifts as seen in the DIGITS and VisDA17 experiments (Section 4), we conjecture that training a network that is robust against local and small changes (adversarial noise) would also help to make it more robust against larger shifts. Our empirical experiments also agree with this, as TIPI performs consistently across multiple types of domain shifts.", + "bbox": [ + 496, + 300, + 892, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: Following TENT [34] and other test-time surrogate optimization methods, we only optimize the affine parameters of the model (i.e., the weight and bias parameters of the norm layers: BatchNorm, LayerNorm, InstanceNorm, or GroupNorm). This has been shown to lead to a more stable optimization, and also reduces the computational cost, as opposed to optimizing all parameters of the network. However, note that optimizing over the entire representation network (similar to SHOT [18]) is also possible. Furthermore, in case the network uses BatchNorm layers, we replace each BatchNorm with two BatchNorm layers (following [35]), one for the clean images and one for the adversarial ones. This is because the batch statistics of the adversarial examples are much different from that of the clean images [35], thus mixing them might cause performance degradation. Note that these two BatchNorm layers only keep separate batch statistics (feature mean and variance), while sharing the affine parameters (weight and bias) – the parameters being optimized.", + "bbox": [ + 496, + 619, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24166", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test the effectiveness of our method on a wide variety of tasks (robustness against image corruptions and domain adaptation). Especially, we performed extensive experiments on various batch sizes of the test-time data batches (as small as 2 and as large as 200). The highlight of our empirical findings is that: TIPI outperforms other test-time adaptation baselines in all scenarios and problem settings. We discuss these experiments in more detail in the following subsections. As a side note, we also found that using TIPI with TENT further improves the performance of both methods, and helps TENT avoid network collapse. Details of our experiments (such as learning rate) can be found in our publicly available source code.", + "bbox": [ + 75, + 114, + 472, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation Protocol and Baselines:", + "text_level": 1, + "bbox": [ + 76, + 320, + 380, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol: We consider a setting where the model is trained normally on the source domain. When the models are available online (e.g., ImageNet-trained Resnet50), we download the weights and keep them as is. Otherwise, we train the models ourselves. Following TENT [34], we perform online adaptation, and report the average performance across all batches (while the model is being adapted). This mimics a real-world scenario.", + "bbox": [ + 75, + 344, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: The main purpose of our experiments is to compare our invariance regularizer against predictive entropy as an unsupervised surrogate subjective. Therefore, TENT [34] is our main baseline in all of the experiments. We also consider recent and relevant fully test-time adaptation methods as our baselines, such as BatchNorm with test-time statistics (BN) [24, 32] and T3A [12]. We also report the performance of the source-only model (source-no adaptation). Note that we also have tested T3A with test-time batch norm statistics (T3A+BN), but this model's performance is almost identical to (or even a little worse than) that of BN for all batch sizes, so we do not include it here. For a fair comparison, we rerun all methods in our experiment (to eliminate artifacts such as source model training), using their official source code. However, our numbers for these methods are very close to the reported numbers.", + "bbox": [ + 75, + 487, + 468, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TTT [33] and $\\mathrm{TTT} + +$ [19] are not fully test-time adaptation methods since they require altering the training process of the source domain (to incorporate a self-supervised task). Thus, they are not optimal for off-the-shelf models that have not been trained with the self-supervised task (e.g., a ResNet50 trained on ImageNet downloaded directly from torchvision). For this reason, we do not include these two methods in the comparisons. Note that as reported in Table 2 of the $\\mathrm{TTT} + +$ paper [19], the variant of their model which has not been trained with the self-supervised task (which is the same as our evaluation protocol and most of the literature) largely underperforms our method.", + "bbox": [ + 75, + 723, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EATA [29] proposes a datapoints selection strategy for TENT, and achieves state-of-the-art performance on the test time adaptation task. As discussed before, we believe this line of research is complementary to ours. Indeed, we are able to incorporate TIPI into EATA, thereby improving its robustness (EATA only uses TENT and is not robust against small batchesizes). Due to the page limit, we defer this comparison to the Supplementary Material.", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Robustness to Image Corruptions", + "text_level": 1, + "bbox": [ + 498, + 218, + 794, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 239, + 709, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the literature, we conduct experiments on widely-used image-corruption datasets, namely:", + "bbox": [ + 496, + 263, + 890, + 295 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CIFAR10-C and CIFAR100-C [10]: Common image corruptions are applied to the CIFAR10 and CIFAR100 datasets [14]. We report the results for the highest level of severity, averaging over 15 types of corruption.", + "- ImageNet-C: Similar with the above two datasets, but for ImageNet [30]. We report the performance of all methods averaged over 15 types of corruption and 5 levels of severity." + ], + "bbox": [ + 517, + 303, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the CIFAR10-C and CIFAR100-C datasets, we use a WideResnet28-10 [36] model, which is the base model in Robustbench [4] and also used in TENT's official GitHub repository1. Note that in the original paper, TENT conducts experiments with a Resnet26 model [9]; however, the model/code was not publicly released, and there have been numerous reproducibility issues as reported in their GitHub page2. We use the Adam optimizer [13] for these two datasets.", + "bbox": [ + 496, + 455, + 890, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the ImageNet-C dataset, we follow previous works and use a Resnet50 network. We use SGD to optimize our invariance regularizer.", + "bbox": [ + 496, + 588, + 890, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Results", + "text_level": 1, + "bbox": [ + 500, + 651, + 606, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1, 2 and 3 show that our method is the best performer among the baselines. In the CIFAR10-C and CIFAR100-C experiments, our method performs similarly with TENT for large batch sizes (200). However, as the batch size decreases, TENT collapses and TIPI outperforms it significantly. For the challenging ImageNet-C dataset, our method performs better than TENT even for large batch sizes. TENT still collapses with small batch sizes while TIPI is very robust to this parameter. T3A is not an optimization-based adaptation approach so its performance is robust against the batch size, but it largely underperforms TIPI in all settings. Even more concerning, T3A performs extremely purely on ImageNet-C (worse than the source", + "bbox": [ + 496, + 674, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/DequanWang/tent", + "bbox": [ + 517, + 875, + 723, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/DequanWang/tent/issues", + "bbox": [ + 517, + 887, + 754, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24167", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9a80600e4c224729f72f1954e2b59aaa321b53f860ce9be6c80b00b9cc4711bc.jpg", + "table_caption": [ + "Table 1. CIFAR10-C: Results are averaged over 15 types of corruption with the highest level of severity" + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Source56.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A57.857.857.857.857.957.957.8
BN61.870.875.177.578.979.674.0
TENT17.460.774.078.681.081.465.5
TIPI (ours)78.679.880.381.081.281.580.4
", + "bbox": [ + 254, + 109, + 715, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91ae79be74bd2b8e5c170a17ab91de52dfee058902449e69289b2adbb02f508b.jpg", + "table_caption": [ + "Table 2. CIFAR100-C: Results are averaged over 15 types of corruption with the highest level of severity" + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Source37.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A39.039.039.039.039.039.039.0
BN26.048.054.157.459.560.750.9
TENT1.57.439.859.964.465.739.8
TIPI (ours)56.362.364.165.065.565.863.2
", + "bbox": [ + 254, + 297, + 715, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ae139580f1d99ee5ef0b06f5cf79e97322274162d3e3d743e75583001e16cfe.jpg", + "table_caption": [ + "Table 3. ImageNet-C: Results are averaged over 15 types of corruption and 5 levels of severity" + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Source39.5
Test Time Adaptation Batch SizeAverage
248163264
T3A30.230.230.230.330.330.330.3
BN13.933.442.846.948.949.947.2
TENT0.813.144.353.756.857.345.2
TIPI (ours)45.253.356.757.958.358.455.0
", + "bbox": [ + 254, + 483, + 715, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model) – we conjecture this is because the non-parametric classifier in T3A is not effective when there is a large number of classes (1000) and each class's support set has a few datapoints (50).", + "bbox": [ + 75, + 661, + 468, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Domain Adaptation", + "text_level": 1, + "bbox": [ + 76, + 729, + 266, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 751, + 285, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this problem, we consider two main adaptation tasks, with the following two datasets:", + "bbox": [ + 75, + 773, + 468, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- DIGITS: the task is digit classification (10 classes), and we have to adapt a model from SVHN [25] to MNIST [15], MNIST-M [6], and USPS [11]. For the MNIST and MNIST-M datasets, we only use the test set (10000 images) for the online adaptation. For the USPS dataset, we use both the training and test set to", + "bbox": [ + 94, + 811, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "have roughly the same number of images for adaptation (9298 in total).", + "bbox": [ + 529, + 662, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- VisDA17 (12-class classification)3: is a more challenging dataset, with the adaption task from synthetic to real images. The source dataset contains 152397 3D-rendered synthetic images, while the target dataset contains 55388 real images.", + "bbox": [ + 517, + 703, + 893, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the experiments performed in TENT, we use a ResNet26 for DIGITS and a ResNet50 for the VisDA17 dataset. For the test-time optimization of our surrogate loss, we use Adam for DIGITS and SGD for VisDA17. Other details such as the learning rate can be found in the provided source code.", + "bbox": [ + 496, + 789, + 893, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{3}$ Oxford solely obtained and processed the VisDA17 dataset", + "bbox": [ + 514, + 886, + 836, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24168", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab19eca82edab4197a5348e99fdc9e7bc7c5b614053fa495c32498bf0d857cd4.jpg", + "table_caption": [ + "Table 4. DIGITS: Online Domain Adaptation from SVHN to MNIST, MNIST-M and USPS" + ], + "table_footnote": [], + "table_body": "
MethodSVHN→MNISTSVHN→MNIST-MSVHN→USPS
B.Size 4B.Size 128B.Size 4B.Size 128B.Size 4B.Size 128
Source85.860.384.9
T3A86.386.360.760.784.984.9
BN75.683.556.060.972.480.8
TENT27.090.719.261.438.790.4
TIPI (ours)86.391.664.263.886.893.9
", + "bbox": [ + 187, + 109, + 782, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f477e1c95e33dba799d0bc95266f7b5175151ea8bd8eaa96ecbedc7872044c78.jpg", + "table_caption": [ + "Table 5. VisDA17: Online Domain Adaptation of the VisDA classification challenge" + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
Source46.4
Test Time Adaptation Batch SizeAverage
248163264128
T3A48.848.848.948.948.948.948.948.9
BN38.349.554.657.358.759.459.853.9
TENT12.326.553.059.962.162.062.048.3
TENTx210.022.145.959.262.563.163.246.6
TIPI (ours)51.260.362.764.064.364.664.561.7
", + "bbox": [ + 232, + 268, + 738, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Results", + "text_level": 1, + "bbox": [ + 76, + 462, + 181, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 and Table 5 show a similar trend as observed in the image corruption experiments. In particular, TIPI performs the best among all the baselines. Especially, our method does not collapse with small batch sizes as most other models (such as BN or TENT) do. It is clear that TIPI achieves preferable performance in all settings while being much more robust compared to TENT.", + "bbox": [ + 75, + 484, + 470, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Runtime and memory consumption", + "text_level": 1, + "bbox": [ + 76, + 598, + 385, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would also like to discuss the runtime and memory (VRAM) consumption of our method. As TIPI is also an optimization-based adaptation, it is natural to compare it against TENT. Comparing our implementation of TIPI with TENT's official implementation, we observe that TIPI and TENT require roughly the same amount of VRAM, while our method is about two times slower than TENT (because it needs to find the adversarial examples). We believe that this is worth the trade-off, since our method clearly outperforms TENT, and is much more robust against small batch sizes – which is often the case in practice. To be more specific, using an NVIDIA Quadro RTX 6000, TIPI can reach 160 FPS (including both adaptation and prediction) with the standard 224x224 images of ImageNet and VisDA17.", + "bbox": [ + 75, + 619, + 468, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, just to be completely fair, we also compare TIPI with a variant of TENT where we perform two adaptation steps for each batch (thus its runtime is roughly the same as TIPI's), namely TENTx2 (second-last row in Table 5). We perform this comparison on the challenging", + "bbox": [ + 75, + 827, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "adaptation dataset VisDA17 (for smaller datasets such as CIFAR and DIGITS, runtime and memory consumption are almost never a problem). TENTx2 does perform better than TENT (in the large-batch-size regime), but is still largely inferior to TIPI. Additionally, the collapsing problem for small batch sizes persists in this variant.", + "bbox": [ + 496, + 462, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 571, + 619, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose to use transformation invariance as an unsupervised surrogate loss function for online domain adaptation. Similar to TENT, our method (TIPI) is task-agnostic and does not require altering the source training process of the model, thus can be applied easily to a wide variety of off-the-shelf models. Remarkably, compared to TENT, our method is much more robust against small batch sizes, which proves to be extremely useful in challenging real-world scenarios. A promising future research direction is to investigate a data point selection strategy for optimizing the TIPI objective, to further improve the performance of the method.", + "bbox": [ + 496, + 598, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments Author A. Tuan Nguyen acknowledges Meta AI for funding his PhD study. Oxford lab headed by author Philip Torr is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1. Meta AI author Ser-Nam Lim is neither supported by the UKRI grant nor has any relationship to the grant.", + "bbox": [ + 496, + 804, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24169", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta regularization. Advances in neural information processing systems, 31, 2018. 2", + "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2", + "[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 3", + "[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020.3, 6", + "[5] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 1", + "[6] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 7", + "[7] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 5", + "[8] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and Zico Kolter. Test-time adaptation via conjugate pseudo-labels. arXiv preprint arXiv:2207.09640, 2022. 2", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[10] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6", + "[11] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(5):550-554, 1994. 7", + "[12] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 6", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[15] Yann LeCun, Corinna Cortes, and CJ Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 7", + "[16] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for do" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "main generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1", + "[17] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 1, 2", + "[18] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 5", + "[19] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808-21820, 2021. 1, 2, 6", + "[20] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 5", + "[21] Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430, 2009. 2", + "[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 3, 5", + "[23] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 2", + "[24] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020.2.6", + "[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 7", + "[26] A Tuan Nguyen, Ser-Nam Lim, and Philip HS Torr. FedSR: A simple and effective domain generalization method for federated learning. 2022. 2", + "[27] A Tuan Nguyen, Toan Tran, Yarin Gal, and Atilim Gunes Baydin. Domain invariant representation learning with domain density transformations. Advances in Neural Information Processing Systems, 34:5264-5275, 2021. 1, 2", + "[28] A Tuan Nguyen, Toan Tran, Yarin Gal, Philip HS Torr, and Atulim Güneş Baydin. KI guided domain adaptation. arXiv preprint arXiv:2106.07780, 2021. 2", + "[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. arXiv preprint arXiv:2204.02610, 2022. 2, 6", + "[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24170", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 3", + "[32] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2, 6", + "[33] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2, 6", + "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 5, 6", + "[35] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 5", + "[36] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + ], + "bbox": [ + 78, + 90, + 468, + 435 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_model.json b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f32a120d242be275e7ce4943f761fdcf36b2448f --- /dev/null +++ b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.792, + 0.152 + ], + "angle": 0, + "content": "TIPI: Test Time Adaptation with Transformation Invariance" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.181, + 0.649, + 0.217 + ], + "angle": 0, + "content": "A. Tuan Nguyen1, Thanh Nguyen-Tang2, Ser-Nam Lim3*, Philip H.S. Torr1*" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.217, + 0.723, + 0.235 + ], + "angle": 0, + "content": "1University of Oxford, 2Johns Hopkins University, 3Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.238, + 0.659, + 0.252 + ], + "angle": 0, + "content": "tuan@robots.ox.ac.uk, nguyent@cs.jhu.edu" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.256, + 0.674, + 0.269 + ], + "angle": 0, + "content": "sernamlim@meta.com, philip.torr@eng.ox.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.473, + 0.832 + ], + "angle": 0, + "content": "When deploying a machine learning model to a new environment, we often encounter the distribution shift problem – meaning the target data distribution is different from the model's training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distribution can affect the model's performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a surrogate loss on the test-time unlabeled target data. In particular, minimizing the prediction's entropy on target samples [34] has received much interest as it is task-agnostic and does not require altering the model's training phase (e.g., does not require adding a self-supervised task during training on the source domain). However, as the target data's batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not optimal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an invariance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model's performance under input transformations. The resulting method (TIPI – Test tIme adaPtation with transformation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT [34] in all settings. Our code is released at https://github.com/atuannguyen/TIPI." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.305, + 0.631, + 0.32 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.895, + 0.66 + ], + "angle": 0, + "content": "Distribution shift is a common problem and is often faced in real-world applications. Specifically, despite taking various precautions while training a machine learning model to ensure a better generalization (e.g., collecting and training on multiple source domains [17, 27], finding flat minima [5], training with meta-learning objectives [16] etc.), the model often still struggles when the test data distribution shifts slightly. Note that it is also common not to have labels for the new shifted domain during test time. To tackle this problem, test time adaptation (also known as online domain adaptation) is a framework that allows the model to adapt to the target distribution using unlabeled test data batches. This is necessary since we typically do not have time to annotate the data during test time and can only make use of the unlabeled data. In this framework, the model needs to give predictions for the target data while simultaneously updating itself to improve its performance on that particular target distribution. We assume a situation in which the target data only arrive in small batches, which makes the adaptation task extremely challenging and renders traditional domain adaptation techniques such as representation alignment (via a distance metric) ineffective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Within this test time adaptation framework, a common and effective approach is to optimize a surrogate objective function in lieu of the true loss function on the target data. The first group of surrogate objectives is the loss functions of self-supervised tasks. In particular, one would formulate a user-defined task (such as predicting the rotation angle of an image) and train it alongside the main task on the source domain; and keep training the self-supervised task on the test-time target data [19, 33]. However, these are not fully test-time adaptation methods, since they require altering the training procedure of the source domain. The second line of surrogate objectives is unsupervised loss functions. Among this group of unsupervised objectives, entropy minimization (TENT) [34] is the most successful method, and has been shown to be consistent across many benchmarks. Furthermore, different from the former group," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.316, + 0.901 + ], + "angle": 0, + "content": "*The last two authors contributed equally" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24162" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "TENT is a fully test-time adaptation method. For these reasons, TENT has received much interest and a lot of follow-up papers/discussions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.284 + ], + "angle": 0, + "content": "However, as pointed out by its authors, TENT is not robust when using small batch sizes, as it often collapses to a trivial solution (i.e., it always predicts the same class for all input). This is detrimental to real-world applications since test data often arrive in small batches. For example, autonomous driving systems only process a few frames in real time – they typically do not accumulate the frames (let's say within one minute) to form a bigger batch. Note that TENT is previously evaluated mainly for large batch sizes (such as 64,128 or 200)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.468, + 0.389 + ], + "angle": 0, + "content": "In this paper, we aim to tackle the aforementioned problem. Specifically, we aim to develop an unsupervised surrogate objective function for the test time adaptation problem such that it is task-agnostic, does not require altering the training procedure (e.g., does not require incorporating a self-supervised task into the training process), and is more resilient against small batch sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.468, + 0.568 + ], + "angle": 0, + "content": "We first provide theoretical results regarding a model's performance under input transformations. Specifically, we show that a model's loss on a data distribution is bounded by the KL distance on the predictive distribution of the data before and after the transformations (which we will use as a regularizer), and its loss on the transformed data distribution. Motivated by this result, we use small shifts in the input images that can simulate real source-target shifts, and enforce the network to be invariant under such data transformations. Our model outperforms TENT (and other relevant baselines) in all problem settings considered in the paper and is remarkably robust in the small-batch-size regime." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.57, + 0.393, + 0.584 + ], + "angle": 0, + "content": "Our contributions in this paper are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.601, + 0.468, + 0.703 + ], + "angle": 0, + "content": "- We provide theoretical results regarding a model's performance under transformations of the input data. Specifically, a model's performance on the target domain is bounded by an invariance term (maximum KL divergence of the predictive distributions before and after the transformations) and its loss on the transformed domain." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.721, + 0.468, + 0.809 + ], + "angle": 0, + "content": "- We propose to find input transformations that can simulate the domain shifts, and enforce the network to be invariant under such transformations, using a regularizer based on our derived bound. The resulting method is TIPI (Test tIme adaPtation with transformation Invariance)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.828, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- We perform extensive experiments on a wide range of datasets (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17) and settings (varying batch sizes) to validate our method. TIPI shows preferable performance compared to relevant baselines." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.601, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.65, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Distribution Shift is an important and commonly faced problem when deploying machine learning models in real-world applications. There are two main approaches that try to alleviate this problem. The first setting is domain generalization [1, 17, 23, 26, 27], where one aims to collect multiple source domains and learn a generalizable model based on those domains. Another setting is (unsupervised) domain adaptation [2, 6, 18, 21, 28], where a model is trained on a labeled source dataset and we attempt to adapt that model to a target distribution with the help of unlabeled target data. This paper focuses on the latter setting, in particular, the test-time domain adaptation variation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Test Time Domain Adaptation: In this setting, the unlabeled target data is not available during the training phase of the source model. Instead, the model only has access to the target data when deployed in the target environment, as the stream of test data batches. As a consequence, the unlabeled target data is not available in large quantities, but only in small batches. This makes the adaptation task extremely challenging. So far, there are two main lines of work dealing with this online adaptation problem. The first approach is based on the batch normalization layer in the network, with the idea of using the test-time batch statistics for these layers instead of the source statistics [24, 32]. This can partially help with the representation misalignment between the source and target distributions but is only applicable when the network has BatchNorm layers and the batch size is large. The other line of work is to use an unsupervised surrogate loss to \"refine\" the model on the target data. For example, TENT [34] minimizes the prediction entropy of the model on the target data. However, as pointed out by its authors, TENT needs a large batch size to avoid collapsing to a trivial prediction (only output one class), due to the characteristics of this surrogate objective. Meanwhile, TTT [33] and TTT++ [19] train a self-supervised task alongside the main task on the source domain, and continue to train the self-supervised task during the online adaptation phase. Therefore, TTT and TTT++ are not fully off-the-shelf test-time adaptation methods, since they require modifications to the training phase of the source domain (to incorporate the self-supervised task). Recently, [8] generalizes TENT to other source loss functions (beyond the typical cross-entropy loss). There also have been several works [29] that study datapoints selection strategies for optimizing the unsupervised surrogate loss (e.g., TENT): only optimize the loss for reliable and informative datapoints. Note that this line of research is orthogonal/complementary to ours and we believe that our surrogate loss will also benefit from similar datapoints selection strategies. We leave exploring this direction to future work. Using a completely different approach, T3A [12] replaces the last linear layer (classification layer) with a nonparametric classifier and keeps updating the support set of each class with the new target" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24163" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "datapoints. T3A is robust against small batch sizes (since it is not optimization-based), but is not as effective as other optimization-based methods (e.g., TENT, TTT, and ours)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.143, + 0.47, + 0.32 + ], + "angle": 0, + "content": "Connection between Adversarial Robustness and Robustness against Distribution shift: It has been observed that adversarily trained models generally perform better under certain types of domain shifts (for example, see the leaderboard of Robustbench [4] for Cifar10-C and Cifar100-C). Indeed, one might expect this to be the case if the space of adversarial attacks matches the space of distribution shifts. While it is not previously well understood, our paper sheds some light on this phenomenon. Note that different from previous works, our method enforces an unsupervised robustness loss during the online adaptation phase, not at the source training stage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.185, + 0.349 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.357, + 0.235, + 0.373 + ], + "angle": 0, + "content": "3.1. Problem setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.469, + 0.629 + ], + "angle": 0, + "content": "Assume that we have a source domain with the distribution \\( p_S(x,y) \\), where \\( x \\in \\mathcal{X} \\) is the input and \\( y \\in \\mathcal{Y} \\) is the prediction target. Similar to prior works, we consider the classification problem, where \\( \\mathcal{Y} = \\{1,2,\\dots,C\\} \\) (however, the approach can be straightforwardly extended to regression problems). Suppose that using data from the source domain, we already successfully trained a model \\( f_{\\theta}: \\mathcal{X} \\to \\mathbb{R}^{C} \\), where \\( \\theta \\) is the parameter of the neural network and \\( f_{\\theta}(x) \\) outputs the prediction probabilities for the \\( C \\) classes \\( (\\|f_{\\theta}(x)\\|_1 = 1 \\) and \\( 0 \\leq f_{\\theta}(x)_i \\leq 1, \\forall i \\in \\{1,2,\\dots,C\\}) \\). This leads to a predictive distribution \\( \\hat{p}_{\\theta}(y|x) \\), which is the categorical distribution where \\( \\hat{p}_{\\theta}(y = i|x) = f_{\\theta}(x)_i \\). Note that, for our method, we will not assume or require any modifications to the training process of the model \\( f_{\\theta} \\) on the source domain - a setting that is common and useful (e.g., a Resnet50 model downloaded directly from torchvision)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.716 + ], + "angle": 0, + "content": "After that, this model is deployed to the target environment with data distribution \\( p_T(x,y) \\). The model receives target data as a stream of unlabeled mini-batches (probably with small and varying sizes). The task is to adapt to this environment simultaneously and continually with only the unlabeled mini-batches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.749 + ], + "angle": 0, + "content": "The expected loss of this model on a data distribution \\( q(x,y) \\) (for example, \\( q(x,y) = p_T(x,y) \\)) is:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.757, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\ell (\\theta , q (x, y)) = \\mathbb {E} _ {q (x)} \\left[ l [ q (y | x), \\hat {p} _ {\\theta} (y | x) ] \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.783, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( l \\) is a loss function (distance/divergence) between the true labeling distribution \\( q(y|x) \\) and the predictive distribution \\( \\hat{p}_{\\theta}(y|x) \\). This loss is not actually computed in practice for both the source and target data (we simply do not have labeled data for both of those), and it is used for theoretical discussion only. To make the theoretical discussion simpler, in this paper, we will consider the simple \\( l_{1} \\) distance between the two distributions. However," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "note that our theoretical discussion also holds for other divergences such as the Jensen-Shannon divergence (or symmetric KL). Both of these distances satisfy the triangle inequality, which we use in our derivation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.188, + 0.892, + 0.218 + ], + "angle": 0, + "content": "3.2. Test Time Domain Adaptation with Transformation Invariance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.226, + 0.892, + 0.313 + ], + "angle": 0, + "content": "Recall that the task is to gradually update the network parameter \\(\\theta\\) so that the model can adapt to the target domain. In this setting, we do not have access to the source data nor a large amount of target data (only small batches), so traditional domain adaptation approaches such as representation alignment are often infeasible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.893, + 0.608 + ], + "angle": 0, + "content": "Given access only to the batches of unlabeled target data, it is natural to try to come up with a surrogate objective to optimize in lieu of the true loss function (which is not available without the labels). In this sense, this has a close resemblance to the semi-supervised learning problem. In both settings, we aim to find a surrogate function to optimize for unlabeled data. In fact, both the entropy-minimization and pseudo-labeling techniques have been borrowed from the semi-supervised literature, with varying degrees of success. It should be noted that, in semi-supervised learning, the surrogate loss on the unlabeled data is minimized jointly with the supervised objective for labeled data, thus model collapsing is often not a problem. It is much more challenging, however, in this online adaptation setting, since the unsupervised surrogate loss is optimized alone, which can lead to the catastrophic collapse of the network parameter (such that it can not perform prediction anymore, even for the source domain). For example, as pointed out by its authors, TENT [34] tends to collapse to a trivial model if the batch size is small, which is often the case in practice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.755 + ], + "angle": 0, + "content": "In this paper, we examine the invariance regularization technique in the setting of test time domain adaptation. Note that a similar regularizer ([3,22,31]) has been applied to semi-supervised learning. However, to the best of our knowledge, we are the first to investigate this type of regularization in the online domain adaptation problem. We aim to investigate both theoretically and empirically how this technique can help with adapting to the new target domain. We also find that this surrogate loss is much more robust (against small batch sizes) compared to TENT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.892, + 0.844 + ], + "angle": 0, + "content": "To summarize, we propose to enforce the network to be invariant under a predefined space of transformations on the data on the target domain. The idea is that this space of transformations contains the actual domain shifts, or at least could simulate them. Specifically, we minimize \\(\\theta\\) with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.843, + 0.892, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {p _ {T} (x)} \\left[ \\max _ {x ^ {\\prime} \\in A (x)} \\mathrm {K L} \\left[ \\hat {p} _ {\\theta} \\left(y \\mid x ^ {\\prime}\\right), \\hat {p} _ {\\bar {\\theta}} (y \\mid x) \\right] \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.89, + 0.903 + ], + "angle": 0, + "content": "where \\(\\overline{\\theta}\\) is an identical copy of \\(\\theta\\), and \\(A(x) = \\{t(x)\\}_{t\\in S_t}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.091, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Algorithm 1 Adapt and predict. \\(\\theta\\) is the network parameter, \\(\\{x_{i}\\}_{i = 1}^{n}\\) is the batch data, \\(\\alpha\\) is the learning rate for online adaptation" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.139, + 0.28, + 0.155 + ], + "angle": 0, + "content": "Input \\(\\theta ,\\{x_{1},x_{2},\\ldots ,x_{n}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.155, + 0.405, + 0.171 + ], + "angle": 0, + "content": "Output \\(\\theta_{\\mathrm{new}}\\) (newly adapted), \\(\\{\\hat{y}_1,\\hat{y}_2,\\dots ,\\hat{y}_n\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.172, + 0.419, + 0.187 + ], + "angle": 0, + "content": "1: function ADAPT-PREDICT \\((\\theta ,\\{x_{1},x_{2},\\dots,x_{n}\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.187, + 0.469, + 0.202 + ], + "angle": 0, + "content": "2: \\( p_i \\gets f_\\theta(x_i) \\). detach() ▷ stop gradient" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.202, + 0.394, + 0.217 + ], + "angle": 0, + "content": "3: \\(x_{i}^{\\prime}\\gets \\arg \\max_{x^{\\prime}\\in A(x_{i})}\\mathrm{KL}[p_{\\theta}(y|x^{\\prime}),p_{i}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.217, + 0.345, + 0.234 + ], + "angle": 0, + "content": "4: \\( l(\\theta) = \\frac{1}{n}\\sum_{i=1}^{n}\\mathrm{KL}[p_{\\theta}(y|x_i'), p_i] \\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.233, + 0.268, + 0.248 + ], + "angle": 0, + "content": "5: \\(\\theta_{\\mathrm{new}} = \\dot{\\theta} -\\alpha \\nabla_{\\theta}l(\\theta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.248, + 0.284, + 0.263 + ], + "angle": 0, + "content": "6: \\(\\hat{y}_i = \\arg \\max f_{\\theta_{\\mathrm{new}}}(x_i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.263, + 0.197, + 0.275 + ], + "angle": 0, + "content": "7: end function" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.172, + 0.469, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.304, + 0.37, + 0.318 + ], + "angle": 0, + "content": "\\((S_{t}\\) is the predefined set of transformations)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.318, + 0.469, + 0.363 + ], + "angle": 0, + "content": "We present the motivations and derivations of this objective function in the next subsection. The adapting and predicting procedure is depicted in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.388, + 0.396 + ], + "angle": 0, + "content": "3.2.1 Invariance Regularization Technique" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.469, + 0.462 + ], + "angle": 0, + "content": "In this subsection, we introduce the invariance regularization technique for general transformations and provide theoretical results regarding the online domain adaptation problem, which will later motivate our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.683 + ], + "angle": 0, + "content": "Let's consider a set of transformations \\( S_{t} \\), where each \\( t \\in S_{t} \\) is a transformation of the data, i.e., \\( t: \\mathcal{X} \\to \\mathcal{X} \\). Ideally, \\( S_{t} \\) should contain the transformations responsible for the domain shift. This set can contain transformations such as data augmentations, GAN-like cross-domain transformations, random noise perturbation, etc. For now, we shall not care much about the type of transformations in \\( S_{t} \\). The only important detail is that these transformations should preserve the class information (this is formalized in Assumption 1). Let \\( A(x) = \\{t(x)\\}_{t \\in S_t} \\) and let \\( p_t(t) \\) be a distribution over \\( S_{t} \\) that assigns probability mass (or density) to each transformation \\( t \\in S_{t} \\). Each distribution \\( p_t(t) \\) defines a conditional distribution \\( T(x^t | x) \\) (with the support set \\( A(x) \\)), where \\( x^t = t(x), t \\sim p_t(t) \\). This in turn defines a joint distribution over \\( x, x^t \\) and \\( y \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.688, + 0.469, + 0.706 + ], + "angle": 0, + "content": "\\[\np _ {T} (x, x ^ {t}, y) = p _ {T} (x, y) T \\left(x ^ {t} | x\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.469, + 0.759 + ], + "angle": 0, + "content": "It is also worth noting that we have transformed data distribution \\( p_T(x^t, y) \\) of \\( x^t \\) and \\( y \\) (just marginalizing out \\( x \\) from the above)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.768, + 0.182, + 0.782 + ], + "angle": 0, + "content": "Assumption 1." + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.792, + 0.469, + 0.809 + ], + "angle": 0, + "content": "\\[\nI _ {T} (x, y) = I _ {T} \\left(x ^ {t}, y\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.819, + 0.469, + 0.863 + ], + "angle": 0, + "content": "where \\( I_{T} \\) is the mutual information, calculated on the target domain (please refer to the Supplementary Material for their exact formula)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Intuitively, this assumption means that the transformed image \\( x^{t} \\) contains the same amount of information about" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "the label \\(y\\) as the original image \\(x\\) does. This assumption is equivalent to \\(\\mathbb{E}_{p_T(x,x^t)}[\\mathrm{KL}[p_T(y|x),p_T(y|x^t)]] = 0\\) (a derivation can be found in the supplementary material), meaning that \\(p_T(y|x) = p_T(y|x^t)\\) almost surely." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.155, + 0.892, + 0.185 + ], + "angle": 0, + "content": "The transformation invariance regularizer on the target domain can be defined simply as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.188, + 0.892, + 0.214 + ], + "angle": 0, + "content": "\\[\nh (\\theta , x) = \\max _ {x ^ {\\prime} \\in A (x)} D \\left[ \\hat {p} _ {\\bar {\\theta}} (y | x), \\hat {p} _ {\\theta} (y | x ^ {\\prime}) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.341 + ], + "angle": 0, + "content": "where \\( D[\\cdot, \\cdot] \\) is a divergence/distance of choice and \\( \\bar{\\theta} \\) is an identical copy of \\( \\theta \\) (so that the gradient is not backpropagated through it). We will see later (through an analysis of the gradient) that this is extremely important. Intuitively, it also helps to alleviate the network collapse problem, since, without this gradient blocking, the network would easily collapse to a constant output to minimize the regularizer term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Intuitively, we can view this objective as using \\(\\hat{p}_{\\overline{\\theta}}(y|x)\\) as the target to optimize \\(\\hat{p}_{\\theta}(y|t(x))\\) for all transformation \\(t\\). This makes sense because, as discussed earlier, \\(t\\) does not change the semantic meaning of \\(x\\) significantly, so the predictive distribution of \\(x\\) and \\(t(x)\\) should be similar." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.416, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Based on this regularizer, we can bound the loss on the target domain as follows:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Proposition 1. Under Assumption 1, if \\( D \\) in Eq. 5 is either the forward or reverse KL divergence, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.491, + 0.881, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\ell (\\theta , p _ {T} (x, y)) \\leq \\mathbb {E} _ {p _ {T} (x)} \\left[ \\sqrt {2 h (\\theta , x)} \\right] + \\ell (\\theta , p _ {T} (x ^ {t}, y)).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.524, + 0.816, + 0.539 + ], + "angle": 0, + "content": "Proof. Provided in the Supplementary Material." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.524, + 0.891, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.891, + 0.592 + ], + "angle": 0, + "content": "This means that the loss on the target distribution is bounded by the invariance regularizer and the loss on the transformed distribution \\( p_T(x^t, y) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.842 + ], + "angle": 0, + "content": "Now, assume for a moment (even if it is not realistic) that the set \\( S_{t} \\) contains all the transformations that can successfully transform the target data distribution \\( p_{T}(x,y) \\) back to the source distribution \\( p_{S}(x,y) \\) (i.e., there exists a distribution \\( p_{t}(t) \\) over \\( S_{t} \\) such that the resulting transformed distribution \\( p_{T}(x^{t},y) \\) is the same with \\( p_{S}(x,y) \\)). Then, we have that the loss on the target domain is bounded by the loss on the source domain and our invariance regularizer. Assume also that while adapting to the new target domain, our model does not suffer from catastrophic forgetting of the source domain (which is often the case if the network does not collapse since we are adapting to the same task in a different environment). Then we have that the loss on the source domain would be small, and the invariance regularizer would also be small (since we explicitly minimize it). As a consequence, the loss on the target domain would be small and we can guarantee a good adaptation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.842, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, knowing the exact transformations from the target domain back to the source domains is often infeasible (only possible if we have some prior domain knowledge). This motivates us to find other transformations that" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "are task-agnostic and can simulate the distribution shift between the target and source domains." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.139, + 0.418, + 0.154 + ], + "angle": 0, + "content": "3.2.2 Practical choice of the transformation set" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.162, + 0.47, + 0.339 + ], + "angle": 0, + "content": "We want to choose task-agnostic transformations such that they do not alter the semantic meaning of the image significantly, while they can simulate the distribution shift. The first idea that comes to mind is to use common image augmentations (e.g., random crop, random vertical flip, etc.) However, our early experiments suggest that this method is not consistent across corruption types/domain shifts. This makes sense since image augmentations are often heuristic, and they may be able to simulate certain types of distribution shifts, but not others. Therefore, we opt to use \\( l_{\\infty} \\)-norm based perturbations. Empirically, this leads to better performance and also is consistent across domains/tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.469, + 0.466 + ], + "angle": 0, + "content": "To summarize, we use the \\(l_{\\infty}\\)-norm based perturbations as the transformations of images. Empirically, we also found out that using the reverse KL leads to better performance (although almost all distances/divergences have positive results). We provide explanations/intuitions as to why we should prefer the reverse KL, and why adversarial examples can help with the distribution shift problem in the next subsection. Our final optimization problem is:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.473, + 0.47, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {p _ {T} (x)} \\left[ \\max _ {| | x ^ {\\prime} - x | | _ {\\infty} \\leq \\epsilon} \\mathrm {K L} \\left[ \\hat {p} _ {\\theta} \\left(y \\mid x ^ {\\prime}\\right), \\hat {p} _ {\\bar {\\theta}} (y \\mid x) \\right] \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.469, + 0.62 + ], + "angle": 0, + "content": "We name our method TIPI (Test tIme adaPtation with transformation Invariance). To save computational cost, we use the single-step adversarial method FGSM [7] to find the adversarial example \\( x' \\). Note that one can also use more advanced methods such as PGD [20] to find the adversarial perturbation to get better results (but with the trade-off of speed). We set \\( \\epsilon \\) to 2/255 in all of our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.637, + 0.291, + 0.65 + ], + "angle": 0, + "content": "3.2.3 Additional Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.82 + ], + "angle": 0, + "content": "Why preferring reverse KL over forward KL? Natural choices for \\( D \\) are the KL-divergences (forward or reverse KL), since they can be computed straightforwardly for the predictive distributions, and also are commonly used in practice for similar problems (e.g., VAT [22] uses the forward KL). Our experiments empirically suggest that using the reverse KL as the divergence \\( D \\) leads to better results compared to the forward KL. Here, we also give theoretical intuition as to why it is the case. First of all, let's consider the gradient of \\( \\theta \\) with respect to both the forward and reverse KL:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.821, + 0.214, + 0.835 + ], + "angle": 0, + "content": "For the forward KL:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.843, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathrm {K L} [ \\hat {p} _ {\\bar {\\theta}} (y | x), \\hat {p} _ {\\theta} (y | x ^ {\\prime}) ] = \\sum_ {y \\in \\mathcal {Y}} - \\frac {\\hat {p} _ {\\bar {\\theta}} (y | x)}{\\hat {p} _ {\\theta} (y | x ^ {\\prime})} \\nabla_ {\\theta} \\hat {p} _ {\\theta} (y | x ^ {\\prime}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.71, + 0.106 + ], + "angle": 0, + "content": "Meanwhile, for the reverse KL:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.124, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} \\mathrm {K L} [ \\hat {p} _ {\\theta} (y | x ^ {\\prime}), \\hat {p} _ {\\bar {\\theta}} (y | x) ] \\\\ = \\sum_ {y \\in \\mathcal {Y}} \\left(- \\log \\frac {\\hat {p} _ {\\bar {\\theta}} (y | x)}{\\hat {p} _ {\\theta} (y | x ^ {\\prime})} + 1\\right) \\nabla_ {\\theta} \\hat {p} _ {\\theta} (y | x ^ {\\prime}). \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.202, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Compare Eq. 8 to Eq. 7, we can see that the coefficient of \\(\\nabla_{\\theta}\\hat{p}_{\\theta}(y|x^{\\prime})\\) in the reverse KL gets squashed down by the log function, leading to much smaller variance. We conjecture that the small variance in gradient is the reason why reverse KL performs better than forward KL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.893, + 0.596 + ], + "angle": 0, + "content": "How can invariance enforcing with adversarial examples help with the distribution shift problem? In our previous subsection, we provide theoretical evidence that motivates us to find transformations that can simulate common distribution shifts in practice. We argue here that \\( l_{\\infty} \\) adversarial perturbation is a good candidate for this. First of all, \\( l_{\\infty} \\) adversarial perturbation is a task-agnostic transformation that does not require much tuning, and we do not need to design heuristic transformations per domain. Secondly, a lot of the common domain shifts in practice are image corruptions that add noises to the images (gaussian noise, gaussian blur, jpeg compression noise, etc.), which can be simulated very well by adversarial noises. Lastly, even for more global domain shifts as seen in the DIGITS and VisDA17 experiments (Section 4), we conjecture that training a network that is robust against local and small changes (adversarial noise) would also help to make it more robust against larger shifts. Our empirical experiments also agree with this, as TIPI performs consistently across multiple types of domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Implementation Details: Following TENT [34] and other test-time surrogate optimization methods, we only optimize the affine parameters of the model (i.e., the weight and bias parameters of the norm layers: BatchNorm, LayerNorm, InstanceNorm, or GroupNorm). This has been shown to lead to a more stable optimization, and also reduces the computational cost, as opposed to optimizing all parameters of the network. However, note that optimizing over the entire representation network (similar to SHOT [18]) is also possible. Furthermore, in case the network uses BatchNorm layers, we replace each BatchNorm with two BatchNorm layers (following [35]), one for the clean images and one for the adversarial ones. This is because the batch statistics of the adversarial examples are much different from that of the clean images [35], thus mixing them might cause performance degradation. Note that these two BatchNorm layers only keep separate batch statistics (feature mean and variance), while sharing the affine parameters (weight and bias) – the parameters being optimized." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24166" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.309 + ], + "angle": 0, + "content": "We test the effectiveness of our method on a wide variety of tasks (robustness against image corruptions and domain adaptation). Especially, we performed extensive experiments on various batch sizes of the test-time data batches (as small as 2 and as large as 200). The highlight of our empirical findings is that: TIPI outperforms other test-time adaptation baselines in all scenarios and problem settings. We discuss these experiments in more detail in the following subsections. As a side note, we also found that using TIPI with TENT further improves the performance of both methods, and helps TENT avoid network collapse. Details of our experiments (such as learning rate) can be found in our publicly available source code." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.321, + 0.382, + 0.337 + ], + "angle": 0, + "content": "4.1. Evaluation Protocol and Baselines:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.464 + ], + "angle": 0, + "content": "Evaluation Protocol: We consider a setting where the model is trained normally on the source domain. When the models are available online (e.g., ImageNet-trained Resnet50), we download the weights and keep them as is. Otherwise, we train the models ourselves. Following TENT [34], we perform online adaptation, and report the average performance across all batches (while the model is being adapted). This mimics a real-world scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.723 + ], + "angle": 0, + "content": "Baselines: The main purpose of our experiments is to compare our invariance regularizer against predictive entropy as an unsupervised surrogate subjective. Therefore, TENT [34] is our main baseline in all of the experiments. We also consider recent and relevant fully test-time adaptation methods as our baselines, such as BatchNorm with test-time statistics (BN) [24, 32] and T3A [12]. We also report the performance of the source-only model (source-no adaptation). Note that we also have tested T3A with test-time batch norm statistics (T3A+BN), but this model's performance is almost identical to (or even a little worse than) that of BN for all batch sizes, so we do not include it here. For a fair comparison, we rerun all methods in our experiment (to eliminate artifacts such as source model training), using their official source code. However, our numbers for these methods are very close to the reported numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.471, + 0.902 + ], + "angle": 0, + "content": "TTT [33] and \\(\\mathrm{TTT} + +\\) [19] are not fully test-time adaptation methods since they require altering the training process of the source domain (to incorporate a self-supervised task). Thus, they are not optimal for off-the-shelf models that have not been trained with the self-supervised task (e.g., a ResNet50 trained on ImageNet downloaded directly from torchvision). For this reason, we do not include these two methods in the comparisons. Note that as reported in Table 2 of the \\(\\mathrm{TTT} + +\\) paper [19], the variant of their model which has not been trained with the self-supervised task (which is the same as our evaluation protocol and most of the literature) largely underperforms our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.212 + ], + "angle": 0, + "content": "EATA [29] proposes a datapoints selection strategy for TENT, and achieves state-of-the-art performance on the test time adaptation task. As discussed before, we believe this line of research is complementary to ours. Indeed, we are able to incorporate TIPI into EATA, thereby improving its robustness (EATA only uses TENT and is not robust against small batchesizes). Due to the page limit, we defer this comparison to the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.219, + 0.795, + 0.236 + ], + "angle": 0, + "content": "4.2. Robustness to Image Corruptions" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.241, + 0.71, + 0.258 + ], + "angle": 0, + "content": "4.2.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Following the literature, we conduct experiments on widely-used image-corruption datasets, namely:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.304, + 0.892, + 0.379 + ], + "angle": 0, + "content": "- CIFAR10-C and CIFAR100-C [10]: Common image corruptions are applied to the CIFAR10 and CIFAR100 datasets [14]. We report the results for the highest level of severity, averaging over 15 types of corruption." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.388, + 0.892, + 0.448 + ], + "angle": 0, + "content": "- ImageNet-C: Similar with the above two datasets, but for ImageNet [30]. We report the performance of all methods averaged over 15 types of corruption and 5 levels of severity." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.304, + 0.892, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.588 + ], + "angle": 0, + "content": "For the CIFAR10-C and CIFAR100-C datasets, we use a WideResnet28-10 [36] model, which is the base model in Robustbench [4] and also used in TENT's official GitHub repository1. Note that in the original paper, TENT conducts experiments with a Resnet26 model [9]; however, the model/code was not publicly released, and there have been numerous reproducibility issues as reported in their GitHub page2. We use the Adam optimizer [13] for these two datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.634 + ], + "angle": 0, + "content": "For the ImageNet-C dataset, we follow previous works and use a Resnet50 network. We use SGD to optimize our invariance regularizer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.607, + 0.665 + ], + "angle": 0, + "content": "4.2.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Table 1, 2 and 3 show that our method is the best performer among the baselines. In the CIFAR10-C and CIFAR100-C experiments, our method performs similarly with TENT for large batch sizes (200). However, as the batch size decreases, TENT collapses and TIPI outperforms it significantly. For the challenging ImageNet-C dataset, our method performs better than TENT even for large batch sizes. TENT still collapses with small batch sizes while TIPI is very robust to this parameter. T3A is not an optimization-based adaptation approach so its performance is robust against the batch size, but it largely underperforms TIPI in all settings. Even more concerning, T3A performs extremely purely on ImageNet-C (worse than the source" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.875, + 0.724, + 0.888 + ], + "angle": 0, + "content": "1https://github.com/DequanWang/tent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.756, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/DequanWang/tent/issues" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.875, + 0.756, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24167" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.176, + 0.089, + 0.794, + 0.105 + ], + "angle": 0, + "content": "Table 1. CIFAR10-C: Results are averaged over 15 types of corruption with the highest level of severity" + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.111, + 0.716, + 0.264 + ], + "angle": 0, + "content": "
MethodAccuracy
Source56.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A57.857.857.857.857.957.957.8
BN61.870.875.177.578.979.674.0
TENT17.460.774.078.681.081.465.5
TIPI (ours)78.679.880.381.081.281.580.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.172, + 0.276, + 0.797, + 0.292 + ], + "angle": 0, + "content": "Table 2. CIFAR100-C: Results are averaged over 15 types of corruption with the highest level of severity" + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.298, + 0.716, + 0.451 + ], + "angle": 0, + "content": "
MethodAccuracy
Source37.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A39.039.039.039.039.039.039.0
BN26.048.054.157.459.560.750.9
TENT1.57.439.859.964.465.739.8
TIPI (ours)56.362.364.165.065.565.863.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.203, + 0.463, + 0.766, + 0.478 + ], + "angle": 0, + "content": "Table 3. ImageNet-C: Results are averaged over 15 types of corruption and 5 levels of severity" + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.484, + 0.716, + 0.638 + ], + "angle": 0, + "content": "
MethodAccuracy
Source39.5
Test Time Adaptation Batch SizeAverage
248163264
T3A30.230.230.230.330.330.330.3
BN13.933.442.846.948.949.947.2
TENT0.813.144.353.756.857.345.2
TIPI (ours)45.253.356.757.958.358.455.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.47, + 0.722 + ], + "angle": 0, + "content": "model) – we conjecture this is because the non-parametric classifier in T3A is not effective when there is a large number of classes (1000) and each class's support set has a few datapoints (50)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.73, + 0.267, + 0.745 + ], + "angle": 0, + "content": "4.3. Domain Adaptation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.752, + 0.287, + 0.768 + ], + "angle": 0, + "content": "4.3.1 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.469, + 0.806 + ], + "angle": 0, + "content": "In this problem, we consider two main adaptation tasks, with the following two datasets:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.812, + 0.473, + 0.902 + ], + "angle": 0, + "content": "- DIGITS: the task is digit classification (10 classes), and we have to adapt a model from SVHN [25] to MNIST [15], MNIST-M [6], and USPS [11]. For the MNIST and MNIST-M datasets, we only use the test set (10000 images) for the online adaptation. For the USPS dataset, we use both the training and test set to" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.663, + 0.892, + 0.692 + ], + "angle": 0, + "content": "have roughly the same number of images for adaptation (9298 in total)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.704, + 0.894, + 0.78 + ], + "angle": 0, + "content": "- VisDA17 (12-class classification)3: is a more challenging dataset, with the adaption task from synthetic to real images. The source dataset contains 152397 3D-rendered synthetic images, while the target dataset contains 55388 real images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.894, + 0.877 + ], + "angle": 0, + "content": "Following the experiments performed in TENT, we use a ResNet26 for DIGITS and a ResNet50 for the VisDA17 dataset. For the test-time optimization of our surrogate loss, we use Adam for DIGITS and SGD for VisDA17. Other details such as the learning rate can be found in the provided source code." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.837, + 0.902 + ], + "angle": 0, + "content": "\\(^{3}\\)Oxford solely obtained and processed the VisDA17 dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24168" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.21, + 0.089, + 0.759, + 0.104 + ], + "angle": 0, + "content": "Table 4. DIGITS: Online Domain Adaptation from SVHN to MNIST, MNIST-M and USPS" + }, + { + "type": "table", + "bbox": [ + 0.189, + 0.111, + 0.783, + 0.236 + ], + "angle": 0, + "content": "
MethodSVHN→MNISTSVHN→MNIST-MSVHN→USPS
B.Size 4B.Size 128B.Size 4B.Size 128B.Size 4B.Size 128
Source85.860.384.9
T3A86.386.360.760.784.984.9
BN75.683.556.060.972.480.8
TENT27.090.719.261.438.790.4
TIPI (ours)86.391.664.263.886.893.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.234, + 0.249, + 0.736, + 0.264 + ], + "angle": 0, + "content": "Table 5. VisDA17: Online Domain Adaptation of the VisDA classification challenge" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.269, + 0.739, + 0.438 + ], + "angle": 0, + "content": "
MethodAccuracy
Source46.4
Test Time Adaptation Batch SizeAverage
248163264128
T3A48.848.848.948.948.948.948.948.9
BN38.349.554.657.358.759.459.853.9
TENT12.326.553.059.962.162.062.048.3
TENTx210.022.145.959.262.563.163.246.6
TIPI (ours)51.260.362.764.064.364.664.561.7
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.463, + 0.183, + 0.477 + ], + "angle": 0, + "content": "4.3.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.471, + 0.59 + ], + "angle": 0, + "content": "Table 4 and Table 5 show a similar trend as observed in the image corruption experiments. In particular, TIPI performs the best among all the baselines. Especially, our method does not collapse with small batch sizes as most other models (such as BN or TENT) do. It is clear that TIPI achieves preferable performance in all settings while being much more robust compared to TENT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.599, + 0.387, + 0.615 + ], + "angle": 0, + "content": "4.4. Runtime and memory consumption" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.827 + ], + "angle": 0, + "content": "We would also like to discuss the runtime and memory (VRAM) consumption of our method. As TIPI is also an optimization-based adaptation, it is natural to compare it against TENT. Comparing our implementation of TIPI with TENT's official implementation, we observe that TIPI and TENT require roughly the same amount of VRAM, while our method is about two times slower than TENT (because it needs to find the adversarial examples). We believe that this is worth the trade-off, since our method clearly outperforms TENT, and is much more robust against small batch sizes – which is often the case in practice. To be more specific, using an NVIDIA Quadro RTX 6000, TIPI can reach 160 FPS (including both adaptation and prediction) with the standard 224x224 images of ImageNet and VisDA17." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.828, + 0.471, + 0.902 + ], + "angle": 0, + "content": "However, just to be completely fair, we also compare TIPI with a variant of TENT where we perform two adaptation steps for each batch (thus its runtime is roughly the same as TIPI's), namely TENTx2 (second-last row in Table 5). We perform this comparison on the challenging" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.552 + ], + "angle": 0, + "content": "adaptation dataset VisDA17 (for smaller datasets such as CIFAR and DIGITS, runtime and memory consumption are almost never a problem). TENTx2 does perform better than TENT (in the large-batch-size regime), but is still largely inferior to TIPI. Additionally, the collapsing problem for small batch sizes persists in this variant." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.62, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In this paper, we propose to use transformation invariance as an unsupervised surrogate loss function for online domain adaptation. Similar to TENT, our method (TIPI) is task-agnostic and does not require altering the source training process of the model, thus can be applied easily to a wide variety of off-the-shelf models. Remarkably, compared to TENT, our method is much more robust against small batch sizes, which proves to be extremely useful in challenging real-world scenarios. A promising future research direction is to investigate a data point selection strategy for optimizing the TIPI objective, to further improve the performance of the method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments Author A. Tuan Nguyen acknowledges Meta AI for funding his PhD study. Oxford lab headed by author Philip Torr is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1. Meta AI author Ser-Nam Lim is neither supported by the UKRI grant nor has any relationship to the grant." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24169" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.168 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta regularization. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.468, + 0.223 + ], + "angle": 0, + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.225, + 0.468, + 0.292 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.294, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020.3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.362, + 0.468, + 0.415 + ], + "angle": 0, + "content": "[5] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.468, + 0.484 + ], + "angle": 0, + "content": "[6] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.468, + 0.526 + ], + "angle": 0, + "content": "[7] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.528, + 0.468, + 0.567 + ], + "angle": 0, + "content": "[8] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and Zico Kolter. Test-time adaptation via conjugate pseudo-labels. arXiv preprint arXiv:2207.09640, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.569, + 0.468, + 0.622 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.624, + 0.468, + 0.664 + ], + "angle": 0, + "content": "[10] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.666, + 0.468, + 0.706 + ], + "angle": 0, + "content": "[11] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(5):550-554, 1994. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[12] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.831 + ], + "angle": 0, + "content": "[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[15] Yann LeCun, Corinna Cortes, and CJ Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[16] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for do" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "main generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[17] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.189, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[18] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.244, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[19] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808-21820, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.312, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[20] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.367, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[21] Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.409, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[23] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[24] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020.2.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.641, + 0.892, + 0.681 + ], + "angle": 0, + "content": "[26] A Tuan Nguyen, Ser-Nam Lim, and Philip HS Torr. FedSR: A simple and effective domain generalization method for federated learning. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.682, + 0.892, + 0.735 + ], + "angle": 0, + "content": "[27] A Tuan Nguyen, Toan Tran, Yarin Gal, and Atilim Gunes Baydin. Domain invariant representation learning with domain density transformations. Advances in Neural Information Processing Systems, 34:5264-5275, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.737, + 0.892, + 0.777 + ], + "angle": 0, + "content": "[28] A Tuan Nguyen, Toan Tran, Yarin Gal, Philip HS Torr, and Atulim Güneş Baydin. KI guided domain adaptation. arXiv preprint arXiv:2106.07780, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.779, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. arXiv preprint arXiv:2204.02610, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24170" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[32] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.217, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[33] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.285, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.341, + 0.47, + 0.408 + ], + "angle": 0, + "content": "[35] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[36] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "24171" + } + ] +] \ No newline at end of file diff --git a/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_origin.pdf b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0184cff73a9be780a17162bab5cca40f493c3800 --- /dev/null +++ b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/5131537e-5515-4827-8458-d96e395e71bb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52758dae209738cc1fa864d091b53799c3832e670c452cd75df07438012ecce7 +size 247852 diff --git a/2023/TIPI_ Test Time Adaptation With Transformation Invariance/full.md b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f20dd5b51781d505b27167a9639b13ab7c542a37 --- /dev/null +++ b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/full.md @@ -0,0 +1,310 @@ +# TIPI: Test Time Adaptation with Transformation Invariance + +A. Tuan Nguyen1, Thanh Nguyen-Tang2, Ser-Nam Lim3*, Philip H.S. Torr1* + +1University of Oxford, 2Johns Hopkins University, 3Meta AI + +tuan@robots.ox.ac.uk, nguyent@cs.jhu.edu + +sernamlim@meta.com, philip.torr@eng.ox.ac.uk + +# Abstract + +When deploying a machine learning model to a new environment, we often encounter the distribution shift problem – meaning the target data distribution is different from the model's training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distribution can affect the model's performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a surrogate loss on the test-time unlabeled target data. In particular, minimizing the prediction's entropy on target samples [34] has received much interest as it is task-agnostic and does not require altering the model's training phase (e.g., does not require adding a self-supervised task during training on the source domain). However, as the target data's batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not optimal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an invariance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model's performance under input transformations. The resulting method (TIPI – Test tIme adaPtation with transformation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT [34] in all settings. Our code is released at https://github.com/atuannguyen/TIPI. + +# 1. Introduction + +Distribution shift is a common problem and is often faced in real-world applications. Specifically, despite taking various precautions while training a machine learning model to ensure a better generalization (e.g., collecting and training on multiple source domains [17, 27], finding flat minima [5], training with meta-learning objectives [16] etc.), the model often still struggles when the test data distribution shifts slightly. Note that it is also common not to have labels for the new shifted domain during test time. To tackle this problem, test time adaptation (also known as online domain adaptation) is a framework that allows the model to adapt to the target distribution using unlabeled test data batches. This is necessary since we typically do not have time to annotate the data during test time and can only make use of the unlabeled data. In this framework, the model needs to give predictions for the target data while simultaneously updating itself to improve its performance on that particular target distribution. We assume a situation in which the target data only arrive in small batches, which makes the adaptation task extremely challenging and renders traditional domain adaptation techniques such as representation alignment (via a distance metric) ineffective. + +Within this test time adaptation framework, a common and effective approach is to optimize a surrogate objective function in lieu of the true loss function on the target data. The first group of surrogate objectives is the loss functions of self-supervised tasks. In particular, one would formulate a user-defined task (such as predicting the rotation angle of an image) and train it alongside the main task on the source domain; and keep training the self-supervised task on the test-time target data [19, 33]. However, these are not fully test-time adaptation methods, since they require altering the training procedure of the source domain. The second line of surrogate objectives is unsupervised loss functions. Among this group of unsupervised objectives, entropy minimization (TENT) [34] is the most successful method, and has been shown to be consistent across many benchmarks. Furthermore, different from the former group, + +TENT is a fully test-time adaptation method. For these reasons, TENT has received much interest and a lot of follow-up papers/discussions. + +However, as pointed out by its authors, TENT is not robust when using small batch sizes, as it often collapses to a trivial solution (i.e., it always predicts the same class for all input). This is detrimental to real-world applications since test data often arrive in small batches. For example, autonomous driving systems only process a few frames in real time – they typically do not accumulate the frames (let's say within one minute) to form a bigger batch. Note that TENT is previously evaluated mainly for large batch sizes (such as 64,128 or 200). + +In this paper, we aim to tackle the aforementioned problem. Specifically, we aim to develop an unsupervised surrogate objective function for the test time adaptation problem such that it is task-agnostic, does not require altering the training procedure (e.g., does not require incorporating a self-supervised task into the training process), and is more resilient against small batch sizes. + +We first provide theoretical results regarding a model's performance under input transformations. Specifically, we show that a model's loss on a data distribution is bounded by the KL distance on the predictive distribution of the data before and after the transformations (which we will use as a regularizer), and its loss on the transformed data distribution. Motivated by this result, we use small shifts in the input images that can simulate real source-target shifts, and enforce the network to be invariant under such data transformations. Our model outperforms TENT (and other relevant baselines) in all problem settings considered in the paper and is remarkably robust in the small-batch-size regime. + +Our contributions in this paper are threefold: + +- We provide theoretical results regarding a model's performance under transformations of the input data. Specifically, a model's performance on the target domain is bounded by an invariance term (maximum KL divergence of the predictive distributions before and after the transformations) and its loss on the transformed domain. +- We propose to find input transformations that can simulate the domain shifts, and enforce the network to be invariant under such transformations, using a regularizer based on our derived bound. The resulting method is TIPI (Test tIme adaPtation with transformation Invariance). +- We perform extensive experiments on a wide range of datasets (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17) and settings (varying batch sizes) to validate our method. TIPI shows preferable performance compared to relevant baselines. + +# 2. Related Works + +Distribution Shift is an important and commonly faced problem when deploying machine learning models in real-world applications. There are two main approaches that try to alleviate this problem. The first setting is domain generalization [1, 17, 23, 26, 27], where one aims to collect multiple source domains and learn a generalizable model based on those domains. Another setting is (unsupervised) domain adaptation [2, 6, 18, 21, 28], where a model is trained on a labeled source dataset and we attempt to adapt that model to a target distribution with the help of unlabeled target data. This paper focuses on the latter setting, in particular, the test-time domain adaptation variation. + +Test Time Domain Adaptation: In this setting, the unlabeled target data is not available during the training phase of the source model. Instead, the model only has access to the target data when deployed in the target environment, as the stream of test data batches. As a consequence, the unlabeled target data is not available in large quantities, but only in small batches. This makes the adaptation task extremely challenging. So far, there are two main lines of work dealing with this online adaptation problem. The first approach is based on the batch normalization layer in the network, with the idea of using the test-time batch statistics for these layers instead of the source statistics [24, 32]. This can partially help with the representation misalignment between the source and target distributions but is only applicable when the network has BatchNorm layers and the batch size is large. The other line of work is to use an unsupervised surrogate loss to "refine" the model on the target data. For example, TENT [34] minimizes the prediction entropy of the model on the target data. However, as pointed out by its authors, TENT needs a large batch size to avoid collapsing to a trivial prediction (only output one class), due to the characteristics of this surrogate objective. Meanwhile, TTT [33] and TTT++ [19] train a self-supervised task alongside the main task on the source domain, and continue to train the self-supervised task during the online adaptation phase. Therefore, TTT and TTT++ are not fully off-the-shelf test-time adaptation methods, since they require modifications to the training phase of the source domain (to incorporate the self-supervised task). Recently, [8] generalizes TENT to other source loss functions (beyond the typical cross-entropy loss). There also have been several works [29] that study datapoints selection strategies for optimizing the unsupervised surrogate loss (e.g., TENT): only optimize the loss for reliable and informative datapoints. Note that this line of research is orthogonal/complementary to ours and we believe that our surrogate loss will also benefit from similar datapoints selection strategies. We leave exploring this direction to future work. Using a completely different approach, T3A [12] replaces the last linear layer (classification layer) with a nonparametric classifier and keeps updating the support set of each class with the new target + +datapoints. T3A is robust against small batch sizes (since it is not optimization-based), but is not as effective as other optimization-based methods (e.g., TENT, TTT, and ours). + +Connection between Adversarial Robustness and Robustness against Distribution shift: It has been observed that adversarily trained models generally perform better under certain types of domain shifts (for example, see the leaderboard of Robustbench [4] for Cifar10-C and Cifar100-C). Indeed, one might expect this to be the case if the space of adversarial attacks matches the space of distribution shifts. While it is not previously well understood, our paper sheds some light on this phenomenon. Note that different from previous works, our method enforces an unsupervised robustness loss during the online adaptation phase, not at the source training stage. + +# 3. Approach + +# 3.1. Problem setting + +Assume that we have a source domain with the distribution $p_S(x,y)$ , where $x \in \mathcal{X}$ is the input and $y \in \mathcal{Y}$ is the prediction target. Similar to prior works, we consider the classification problem, where $\mathcal{Y} = \{1,2,\dots,C\}$ (however, the approach can be straightforwardly extended to regression problems). Suppose that using data from the source domain, we already successfully trained a model $f_{\theta}: \mathcal{X} \to \mathbb{R}^{C}$ , where $\theta$ is the parameter of the neural network and $f_{\theta}(x)$ outputs the prediction probabilities for the $C$ classes $(\|f_{\theta}(x)\|_1 = 1$ and $0 \leq f_{\theta}(x)_i \leq 1, \forall i \in \{1,2,\dots,C\})$ . This leads to a predictive distribution $\hat{p}_{\theta}(y|x)$ , which is the categorical distribution where $\hat{p}_{\theta}(y = i|x) = f_{\theta}(x)_i$ . Note that, for our method, we will not assume or require any modifications to the training process of the model $f_{\theta}$ on the source domain - a setting that is common and useful (e.g., a Resnet50 model downloaded directly from torchvision). + +After that, this model is deployed to the target environment with data distribution $p_T(x,y)$ . The model receives target data as a stream of unlabeled mini-batches (probably with small and varying sizes). The task is to adapt to this environment simultaneously and continually with only the unlabeled mini-batches. + +The expected loss of this model on a data distribution $q(x,y)$ (for example, $q(x,y) = p_T(x,y)$ ) is: + +$$ +\ell (\theta , q (x, y)) = \mathbb {E} _ {q (x)} \left[ l [ q (y | x), \hat {p} _ {\theta} (y | x) ] \right], \tag {1} +$$ + +where $l$ is a loss function (distance/divergence) between the true labeling distribution $q(y|x)$ and the predictive distribution $\hat{p}_{\theta}(y|x)$ . This loss is not actually computed in practice for both the source and target data (we simply do not have labeled data for both of those), and it is used for theoretical discussion only. To make the theoretical discussion simpler, in this paper, we will consider the simple $l_{1}$ distance between the two distributions. However, + +note that our theoretical discussion also holds for other divergences such as the Jensen-Shannon divergence (or symmetric KL). Both of these distances satisfy the triangle inequality, which we use in our derivation. + +# 3.2. Test Time Domain Adaptation with Transformation Invariance + +Recall that the task is to gradually update the network parameter $\theta$ so that the model can adapt to the target domain. In this setting, we do not have access to the source data nor a large amount of target data (only small batches), so traditional domain adaptation approaches such as representation alignment are often infeasible. + +Given access only to the batches of unlabeled target data, it is natural to try to come up with a surrogate objective to optimize in lieu of the true loss function (which is not available without the labels). In this sense, this has a close resemblance to the semi-supervised learning problem. In both settings, we aim to find a surrogate function to optimize for unlabeled data. In fact, both the entropy-minimization and pseudo-labeling techniques have been borrowed from the semi-supervised literature, with varying degrees of success. It should be noted that, in semi-supervised learning, the surrogate loss on the unlabeled data is minimized jointly with the supervised objective for labeled data, thus model collapsing is often not a problem. It is much more challenging, however, in this online adaptation setting, since the unsupervised surrogate loss is optimized alone, which can lead to the catastrophic collapse of the network parameter (such that it can not perform prediction anymore, even for the source domain). For example, as pointed out by its authors, TENT [34] tends to collapse to a trivial model if the batch size is small, which is often the case in practice. + +In this paper, we examine the invariance regularization technique in the setting of test time domain adaptation. Note that a similar regularizer ([3,22,31]) has been applied to semi-supervised learning. However, to the best of our knowledge, we are the first to investigate this type of regularization in the online domain adaptation problem. We aim to investigate both theoretically and empirically how this technique can help with adapting to the new target domain. We also find that this surrogate loss is much more robust (against small batch sizes) compared to TENT. + +To summarize, we propose to enforce the network to be invariant under a predefined space of transformations on the data on the target domain. The idea is that this space of transformations contains the actual domain shifts, or at least could simulate them. Specifically, we minimize $\theta$ with the following objective: + +$$ +\min _ {\theta} \mathbb {E} _ {p _ {T} (x)} \left[ \max _ {x ^ {\prime} \in A (x)} \mathrm {K L} \left[ \hat {p} _ {\theta} \left(y \mid x ^ {\prime}\right), \hat {p} _ {\bar {\theta}} (y \mid x) \right] \right], \tag {2} +$$ + +where $\overline{\theta}$ is an identical copy of $\theta$ , and $A(x) = \{t(x)\}_{t\in S_t}$ . + +Algorithm 1 Adapt and predict. $\theta$ is the network parameter, $\{x_{i}\}_{i = 1}^{n}$ is the batch data, $\alpha$ is the learning rate for online adaptation + +Input $\theta ,\{x_{1},x_{2},\ldots ,x_{n}\}$ + +Output $\theta_{\mathrm{new}}$ (newly adapted), $\{\hat{y}_1,\hat{y}_2,\dots ,\hat{y}_n\}$ + +1: function ADAPT-PREDICT $(\theta ,\{x_{1},x_{2},\dots,x_{n}\})$ +2: $p_i \gets f_\theta(x_i)$ . detach() ▷ stop gradient +3: $x_{i}^{\prime}\gets \arg \max_{x^{\prime}\in A(x_{i})}\mathrm{KL}[p_{\theta}(y|x^{\prime}),p_{i}]$ +4: $l(\theta) = \frac{1}{n}\sum_{i=1}^{n}\mathrm{KL}[p_{\theta}(y|x_i'), p_i]$ +5: $\theta_{\mathrm{new}} = \dot{\theta} -\alpha \nabla_{\theta}l(\theta)$ +6: $\hat{y}_i = \arg \max f_{\theta_{\mathrm{new}}}(x_i)$ +7: end function + +$(S_{t}$ is the predefined set of transformations). + +We present the motivations and derivations of this objective function in the next subsection. The adapting and predicting procedure is depicted in Algorithm 1. + +# 3.2.1 Invariance Regularization Technique + +In this subsection, we introduce the invariance regularization technique for general transformations and provide theoretical results regarding the online domain adaptation problem, which will later motivate our method. + +Let's consider a set of transformations $S_{t}$ , where each $t \in S_{t}$ is a transformation of the data, i.e., $t: \mathcal{X} \to \mathcal{X}$ . Ideally, $S_{t}$ should contain the transformations responsible for the domain shift. This set can contain transformations such as data augmentations, GAN-like cross-domain transformations, random noise perturbation, etc. For now, we shall not care much about the type of transformations in $S_{t}$ . The only important detail is that these transformations should preserve the class information (this is formalized in Assumption 1). Let $A(x) = \{t(x)\}_{t \in S_t}$ and let $p_t(t)$ be a distribution over $S_{t}$ that assigns probability mass (or density) to each transformation $t \in S_{t}$ . Each distribution $p_t(t)$ defines a conditional distribution $T(x^t | x)$ (with the support set $A(x)$ ), where $x^t = t(x), t \sim p_t(t)$ . This in turn defines a joint distribution over $x, x^t$ and $y$ as: + +$$ +p _ {T} (x, x ^ {t}, y) = p _ {T} (x, y) T \left(x ^ {t} | x\right). \tag {3} +$$ + +It is also worth noting that we have transformed data distribution $p_T(x^t, y)$ of $x^t$ and $y$ (just marginalizing out $x$ from the above). + +# Assumption 1. + +$$ +I _ {T} (x, y) = I _ {T} \left(x ^ {t}, y\right), \tag {4} +$$ + +where $I_{T}$ is the mutual information, calculated on the target domain (please refer to the Supplementary Material for their exact formula). + +Intuitively, this assumption means that the transformed image $x^{t}$ contains the same amount of information about + +the label $y$ as the original image $x$ does. This assumption is equivalent to $\mathbb{E}_{p_T(x,x^t)}[\mathrm{KL}[p_T(y|x),p_T(y|x^t)]] = 0$ (a derivation can be found in the supplementary material), meaning that $p_T(y|x) = p_T(y|x^t)$ almost surely. + +The transformation invariance regularizer on the target domain can be defined simply as: + +$$ +h (\theta , x) = \max _ {x ^ {\prime} \in A (x)} D \left[ \hat {p} _ {\bar {\theta}} (y | x), \hat {p} _ {\theta} (y | x ^ {\prime}) \right], \tag {5} +$$ + +where $D[\cdot, \cdot]$ is a divergence/distance of choice and $\bar{\theta}$ is an identical copy of $\theta$ (so that the gradient is not backpropagated through it). We will see later (through an analysis of the gradient) that this is extremely important. Intuitively, it also helps to alleviate the network collapse problem, since, without this gradient blocking, the network would easily collapse to a constant output to minimize the regularizer term. + +Intuitively, we can view this objective as using $\hat{p}_{\overline{\theta}}(y|x)$ as the target to optimize $\hat{p}_{\theta}(y|t(x))$ for all transformation $t$ . This makes sense because, as discussed earlier, $t$ does not change the semantic meaning of $x$ significantly, so the predictive distribution of $x$ and $t(x)$ should be similar. + +Based on this regularizer, we can bound the loss on the target domain as follows: + +Proposition 1. Under Assumption 1, if $D$ in Eq. 5 is either the forward or reverse KL divergence, we have: + +$$ +\ell (\theta , p _ {T} (x, y)) \leq \mathbb {E} _ {p _ {T} (x)} \left[ \sqrt {2 h (\theta , x)} \right] + \ell (\theta , p _ {T} (x ^ {t}, y)). +$$ + +Proof. Provided in the Supplementary Material. + +![](images/902ea96dd6d39510cb0380d3819e426ff6444d283bdfd9027a42e19752302c8b.jpg) + +This means that the loss on the target distribution is bounded by the invariance regularizer and the loss on the transformed distribution $p_T(x^t, y)$ . + +Now, assume for a moment (even if it is not realistic) that the set $S_{t}$ contains all the transformations that can successfully transform the target data distribution $p_{T}(x,y)$ back to the source distribution $p_{S}(x,y)$ (i.e., there exists a distribution $p_{t}(t)$ over $S_{t}$ such that the resulting transformed distribution $p_{T}(x^{t},y)$ is the same with $p_{S}(x,y)$ ). Then, we have that the loss on the target domain is bounded by the loss on the source domain and our invariance regularizer. Assume also that while adapting to the new target domain, our model does not suffer from catastrophic forgetting of the source domain (which is often the case if the network does not collapse since we are adapting to the same task in a different environment). Then we have that the loss on the source domain would be small, and the invariance regularizer would also be small (since we explicitly minimize it). As a consequence, the loss on the target domain would be small and we can guarantee a good adaptation performance. + +However, knowing the exact transformations from the target domain back to the source domains is often infeasible (only possible if we have some prior domain knowledge). This motivates us to find other transformations that + +are task-agnostic and can simulate the distribution shift between the target and source domains. + +# 3.2.2 Practical choice of the transformation set + +We want to choose task-agnostic transformations such that they do not alter the semantic meaning of the image significantly, while they can simulate the distribution shift. The first idea that comes to mind is to use common image augmentations (e.g., random crop, random vertical flip, etc.) However, our early experiments suggest that this method is not consistent across corruption types/domain shifts. This makes sense since image augmentations are often heuristic, and they may be able to simulate certain types of distribution shifts, but not others. Therefore, we opt to use $l_{\infty}$ -norm based perturbations. Empirically, this leads to better performance and also is consistent across domains/tasks. + +To summarize, we use the $l_{\infty}$ -norm based perturbations as the transformations of images. Empirically, we also found out that using the reverse KL leads to better performance (although almost all distances/divergences have positive results). We provide explanations/intuitions as to why we should prefer the reverse KL, and why adversarial examples can help with the distribution shift problem in the next subsection. Our final optimization problem is: + +$$ +\min _ {\theta} \mathbb {E} _ {p _ {T} (x)} \left[ \max _ {| | x ^ {\prime} - x | | _ {\infty} \leq \epsilon} \mathrm {K L} \left[ \hat {p} _ {\theta} \left(y \mid x ^ {\prime}\right), \hat {p} _ {\bar {\theta}} (y \mid x) \right] \right]. \tag {6} +$$ + +We name our method TIPI (Test tIme adaPtation with transformation Invariance). To save computational cost, we use the single-step adversarial method FGSM [7] to find the adversarial example $x'$ . Note that one can also use more advanced methods such as PGD [20] to find the adversarial perturbation to get better results (but with the trade-off of speed). We set $\epsilon$ to 2/255 in all of our experiments. + +# 3.2.3 Additional Discussions + +Why preferring reverse KL over forward KL? Natural choices for $D$ are the KL-divergences (forward or reverse KL), since they can be computed straightforwardly for the predictive distributions, and also are commonly used in practice for similar problems (e.g., VAT [22] uses the forward KL). Our experiments empirically suggest that using the reverse KL as the divergence $D$ leads to better results compared to the forward KL. Here, we also give theoretical intuition as to why it is the case. First of all, let's consider the gradient of $\theta$ with respect to both the forward and reverse KL: + +For the forward KL: + +$$ +\nabla_ {\theta} \mathrm {K L} [ \hat {p} _ {\bar {\theta}} (y | x), \hat {p} _ {\theta} (y | x ^ {\prime}) ] = \sum_ {y \in \mathcal {Y}} - \frac {\hat {p} _ {\bar {\theta}} (y | x)}{\hat {p} _ {\theta} (y | x ^ {\prime})} \nabla_ {\theta} \hat {p} _ {\theta} (y | x ^ {\prime}). \tag {7} +$$ + +Meanwhile, for the reverse KL: + +$$ +\begin{array}{l} \nabla_ {\theta} \mathrm {K L} [ \hat {p} _ {\theta} (y | x ^ {\prime}), \hat {p} _ {\bar {\theta}} (y | x) ] \\ = \sum_ {y \in \mathcal {Y}} \left(- \log \frac {\hat {p} _ {\bar {\theta}} (y | x)}{\hat {p} _ {\theta} (y | x ^ {\prime})} + 1\right) \nabla_ {\theta} \hat {p} _ {\theta} (y | x ^ {\prime}). \tag {8} \\ \end{array} +$$ + +Compare Eq. 8 to Eq. 7, we can see that the coefficient of $\nabla_{\theta}\hat{p}_{\theta}(y|x^{\prime})$ in the reverse KL gets squashed down by the log function, leading to much smaller variance. We conjecture that the small variance in gradient is the reason why reverse KL performs better than forward KL. + +How can invariance enforcing with adversarial examples help with the distribution shift problem? In our previous subsection, we provide theoretical evidence that motivates us to find transformations that can simulate common distribution shifts in practice. We argue here that $l_{\infty}$ adversarial perturbation is a good candidate for this. First of all, $l_{\infty}$ adversarial perturbation is a task-agnostic transformation that does not require much tuning, and we do not need to design heuristic transformations per domain. Secondly, a lot of the common domain shifts in practice are image corruptions that add noises to the images (gaussian noise, gaussian blur, jpeg compression noise, etc.), which can be simulated very well by adversarial noises. Lastly, even for more global domain shifts as seen in the DIGITS and VisDA17 experiments (Section 4), we conjecture that training a network that is robust against local and small changes (adversarial noise) would also help to make it more robust against larger shifts. Our empirical experiments also agree with this, as TIPI performs consistently across multiple types of domain shifts. + +Implementation Details: Following TENT [34] and other test-time surrogate optimization methods, we only optimize the affine parameters of the model (i.e., the weight and bias parameters of the norm layers: BatchNorm, LayerNorm, InstanceNorm, or GroupNorm). This has been shown to lead to a more stable optimization, and also reduces the computational cost, as opposed to optimizing all parameters of the network. However, note that optimizing over the entire representation network (similar to SHOT [18]) is also possible. Furthermore, in case the network uses BatchNorm layers, we replace each BatchNorm with two BatchNorm layers (following [35]), one for the clean images and one for the adversarial ones. This is because the batch statistics of the adversarial examples are much different from that of the clean images [35], thus mixing them might cause performance degradation. Note that these two BatchNorm layers only keep separate batch statistics (feature mean and variance), while sharing the affine parameters (weight and bias) – the parameters being optimized. + +# 4. Experiments + +We test the effectiveness of our method on a wide variety of tasks (robustness against image corruptions and domain adaptation). Especially, we performed extensive experiments on various batch sizes of the test-time data batches (as small as 2 and as large as 200). The highlight of our empirical findings is that: TIPI outperforms other test-time adaptation baselines in all scenarios and problem settings. We discuss these experiments in more detail in the following subsections. As a side note, we also found that using TIPI with TENT further improves the performance of both methods, and helps TENT avoid network collapse. Details of our experiments (such as learning rate) can be found in our publicly available source code. + +# 4.1. Evaluation Protocol and Baselines: + +Evaluation Protocol: We consider a setting where the model is trained normally on the source domain. When the models are available online (e.g., ImageNet-trained Resnet50), we download the weights and keep them as is. Otherwise, we train the models ourselves. Following TENT [34], we perform online adaptation, and report the average performance across all batches (while the model is being adapted). This mimics a real-world scenario. + +Baselines: The main purpose of our experiments is to compare our invariance regularizer against predictive entropy as an unsupervised surrogate subjective. Therefore, TENT [34] is our main baseline in all of the experiments. We also consider recent and relevant fully test-time adaptation methods as our baselines, such as BatchNorm with test-time statistics (BN) [24, 32] and T3A [12]. We also report the performance of the source-only model (source-no adaptation). Note that we also have tested T3A with test-time batch norm statistics (T3A+BN), but this model's performance is almost identical to (or even a little worse than) that of BN for all batch sizes, so we do not include it here. For a fair comparison, we rerun all methods in our experiment (to eliminate artifacts such as source model training), using their official source code. However, our numbers for these methods are very close to the reported numbers. + +TTT [33] and $\mathrm{TTT} + +$ [19] are not fully test-time adaptation methods since they require altering the training process of the source domain (to incorporate a self-supervised task). Thus, they are not optimal for off-the-shelf models that have not been trained with the self-supervised task (e.g., a ResNet50 trained on ImageNet downloaded directly from torchvision). For this reason, we do not include these two methods in the comparisons. Note that as reported in Table 2 of the $\mathrm{TTT} + +$ paper [19], the variant of their model which has not been trained with the self-supervised task (which is the same as our evaluation protocol and most of the literature) largely underperforms our method. + +EATA [29] proposes a datapoints selection strategy for TENT, and achieves state-of-the-art performance on the test time adaptation task. As discussed before, we believe this line of research is complementary to ours. Indeed, we are able to incorporate TIPI into EATA, thereby improving its robustness (EATA only uses TENT and is not robust against small batchesizes). Due to the page limit, we defer this comparison to the Supplementary Material. + +# 4.2. Robustness to Image Corruptions + +# 4.2.1 Experimental Settings + +Following the literature, we conduct experiments on widely-used image-corruption datasets, namely: + +- CIFAR10-C and CIFAR100-C [10]: Common image corruptions are applied to the CIFAR10 and CIFAR100 datasets [14]. We report the results for the highest level of severity, averaging over 15 types of corruption. +- ImageNet-C: Similar with the above two datasets, but for ImageNet [30]. We report the performance of all methods averaged over 15 types of corruption and 5 levels of severity. + +For the CIFAR10-C and CIFAR100-C datasets, we use a WideResnet28-10 [36] model, which is the base model in Robustbench [4] and also used in TENT's official GitHub repository1. Note that in the original paper, TENT conducts experiments with a Resnet26 model [9]; however, the model/code was not publicly released, and there have been numerous reproducibility issues as reported in their GitHub page2. We use the Adam optimizer [13] for these two datasets. + +For the ImageNet-C dataset, we follow previous works and use a Resnet50 network. We use SGD to optimize our invariance regularizer. + +# 4.2.2 Results + +Table 1, 2 and 3 show that our method is the best performer among the baselines. In the CIFAR10-C and CIFAR100-C experiments, our method performs similarly with TENT for large batch sizes (200). However, as the batch size decreases, TENT collapses and TIPI outperforms it significantly. For the challenging ImageNet-C dataset, our method performs better than TENT even for large batch sizes. TENT still collapses with small batch sizes while TIPI is very robust to this parameter. T3A is not an optimization-based adaptation approach so its performance is robust against the batch size, but it largely underperforms TIPI in all settings. Even more concerning, T3A performs extremely purely on ImageNet-C (worse than the source + +Table 1. CIFAR10-C: Results are averaged over 15 types of corruption with the highest level of severity + +
MethodAccuracy
Source56.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A57.857.857.857.857.957.957.8
BN61.870.875.177.578.979.674.0
TENT17.460.774.078.681.081.465.5
TIPI (ours)78.679.880.381.081.281.580.4
+ +Table 2. CIFAR100-C: Results are averaged over 15 types of corruption with the highest level of severity + +
MethodAccuracy
Source37.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A39.039.039.039.039.039.039.0
BN26.048.054.157.459.560.750.9
TENT1.57.439.859.964.465.739.8
TIPI (ours)56.362.364.165.065.565.863.2
+ +Table 3. ImageNet-C: Results are averaged over 15 types of corruption and 5 levels of severity + +
MethodAccuracy
Source39.5
Test Time Adaptation Batch SizeAverage
248163264
T3A30.230.230.230.330.330.330.3
BN13.933.442.846.948.949.947.2
TENT0.813.144.353.756.857.345.2
TIPI (ours)45.253.356.757.958.358.455.0
+ +model) – we conjecture this is because the non-parametric classifier in T3A is not effective when there is a large number of classes (1000) and each class's support set has a few datapoints (50). + +# 4.3. Domain Adaptation + +# 4.3.1 Experimental Settings + +In this problem, we consider two main adaptation tasks, with the following two datasets: + +- DIGITS: the task is digit classification (10 classes), and we have to adapt a model from SVHN [25] to MNIST [15], MNIST-M [6], and USPS [11]. For the MNIST and MNIST-M datasets, we only use the test set (10000 images) for the online adaptation. For the USPS dataset, we use both the training and test set to + +have roughly the same number of images for adaptation (9298 in total). + +- VisDA17 (12-class classification)3: is a more challenging dataset, with the adaption task from synthetic to real images. The source dataset contains 152397 3D-rendered synthetic images, while the target dataset contains 55388 real images. + +Following the experiments performed in TENT, we use a ResNet26 for DIGITS and a ResNet50 for the VisDA17 dataset. For the test-time optimization of our surrogate loss, we use Adam for DIGITS and SGD for VisDA17. Other details such as the learning rate can be found in the provided source code. + +Table 4. DIGITS: Online Domain Adaptation from SVHN to MNIST, MNIST-M and USPS + +
MethodSVHN→MNISTSVHN→MNIST-MSVHN→USPS
B.Size 4B.Size 128B.Size 4B.Size 128B.Size 4B.Size 128
Source85.860.384.9
T3A86.386.360.760.784.984.9
BN75.683.556.060.972.480.8
TENT27.090.719.261.438.790.4
TIPI (ours)86.391.664.263.886.893.9
+ +Table 5. VisDA17: Online Domain Adaptation of the VisDA classification challenge + +
MethodAccuracy
Source46.4
Test Time Adaptation Batch SizeAverage
248163264128
T3A48.848.848.948.948.948.948.948.9
BN38.349.554.657.358.759.459.853.9
TENT12.326.553.059.962.162.062.048.3
TENTx210.022.145.959.262.563.163.246.6
TIPI (ours)51.260.362.764.064.364.664.561.7
+ +# 4.3.2 Results + +Table 4 and Table 5 show a similar trend as observed in the image corruption experiments. In particular, TIPI performs the best among all the baselines. Especially, our method does not collapse with small batch sizes as most other models (such as BN or TENT) do. It is clear that TIPI achieves preferable performance in all settings while being much more robust compared to TENT. + +# 4.4. Runtime and memory consumption + +We would also like to discuss the runtime and memory (VRAM) consumption of our method. As TIPI is also an optimization-based adaptation, it is natural to compare it against TENT. Comparing our implementation of TIPI with TENT's official implementation, we observe that TIPI and TENT require roughly the same amount of VRAM, while our method is about two times slower than TENT (because it needs to find the adversarial examples). We believe that this is worth the trade-off, since our method clearly outperforms TENT, and is much more robust against small batch sizes – which is often the case in practice. To be more specific, using an NVIDIA Quadro RTX 6000, TIPI can reach 160 FPS (including both adaptation and prediction) with the standard 224x224 images of ImageNet and VisDA17. + +However, just to be completely fair, we also compare TIPI with a variant of TENT where we perform two adaptation steps for each batch (thus its runtime is roughly the same as TIPI's), namely TENTx2 (second-last row in Table 5). We perform this comparison on the challenging + +adaptation dataset VisDA17 (for smaller datasets such as CIFAR and DIGITS, runtime and memory consumption are almost never a problem). TENTx2 does perform better than TENT (in the large-batch-size regime), but is still largely inferior to TIPI. Additionally, the collapsing problem for small batch sizes persists in this variant. + +# 5. Conclusion + +In this paper, we propose to use transformation invariance as an unsupervised surrogate loss function for online domain adaptation. Similar to TENT, our method (TIPI) is task-agnostic and does not require altering the source training process of the model, thus can be applied easily to a wide variety of off-the-shelf models. Remarkably, compared to TENT, our method is much more robust against small batch sizes, which proves to be extremely useful in challenging real-world scenarios. A promising future research direction is to investigate a data point selection strategy for optimizing the TIPI objective, to further improve the performance of the method. + +Acknowledgments Author A. Tuan Nguyen acknowledges Meta AI for funding his PhD study. Oxford lab headed by author Philip Torr is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1. Meta AI author Ser-Nam Lim is neither supported by the UKRI grant nor has any relationship to the grant. + +# References + +[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta regularization. Advances in neural information processing systems, 31, 2018. 2 +[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2 +[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 3 +[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020.3, 6 +[5] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 1 +[6] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 7 +[7] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 5 +[8] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and Zico Kolter. Test-time adaptation via conjugate pseudo-labels. arXiv preprint arXiv:2207.09640, 2022. 2 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[10] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6 +[11] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(5):550-554, 1994. 7 +[12] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 6 +[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[15] Yann LeCun, Corinna Cortes, and CJ Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 7 +[16] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for do + +main generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1 +[17] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 1, 2 +[18] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 5 +[19] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808-21820, 2021. 1, 2, 6 +[20] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 5 +[21] Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430, 2009. 2 +[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 3, 5 +[23] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 2 +[24] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020.2.6 +[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 7 +[26] A Tuan Nguyen, Ser-Nam Lim, and Philip HS Torr. FedSR: A simple and effective domain generalization method for federated learning. 2022. 2 +[27] A Tuan Nguyen, Toan Tran, Yarin Gal, and Atilim Gunes Baydin. Domain invariant representation learning with domain density transformations. Advances in Neural Information Processing Systems, 34:5264-5275, 2021. 1, 2 +[28] A Tuan Nguyen, Toan Tran, Yarin Gal, Philip HS Torr, and Atulim Güneş Baydin. KI guided domain adaptation. arXiv preprint arXiv:2106.07780, 2021. 2 +[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. arXiv preprint arXiv:2204.02610, 2022. 2, 6 +[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6 + +[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 3 +[32] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2, 6 +[33] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2, 6 +[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 5, 6 +[35] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 5 +[36] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6 \ No newline at end of file diff --git a/2023/TIPI_ Test Time Adaptation With Transformation Invariance/images.zip b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..53c0c59a999a334c9cad8cd88b0ccceac076dc70 --- /dev/null +++ b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f29511bb303417a5c21acf6756f4c84a49a3e03ff4d15b09a33b281fbdd4063 +size 238022 diff --git a/2023/TIPI_ Test Time Adaptation With Transformation Invariance/layout.json b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d1f0d540197509965bf03fd3d41dbb0654703888 --- /dev/null +++ b/2023/TIPI_ Test Time Adaptation With Transformation Invariance/layout.json @@ -0,0 +1,8178 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 484, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 484, + 120 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 484, + 120 + ], + "type": "text", + "content": "TIPI: Test Time Adaptation with Transformation Invariance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 143, + 397, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 143, + 397, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 143, + 397, + 171 + ], + "type": "text", + "content": "A. Tuan Nguyen1, Thanh Nguyen-Tang2, Ser-Nam Lim3*, Philip H.S. Torr1*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 171, + 442, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 171, + 442, + 186 + ], + "spans": [ + { + "bbox": [ + 151, + 171, + 442, + 186 + ], + "type": "text", + "content": "1University of Oxford, 2Johns Hopkins University, 3Meta AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 188, + 403, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 188, + 403, + 199 + ], + "spans": [ + { + "bbox": [ + 190, + 188, + 403, + 199 + ], + "type": "text", + "content": "tuan@robots.ox.ac.uk, nguyent@cs.jhu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 179, + 202, + 412, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 202, + 412, + 213 + ], + "spans": [ + { + "bbox": [ + 179, + 202, + 412, + 213 + ], + "type": "text", + "content": "sernamlim@meta.com, philip.torr@eng.ox.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 272, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 289, + 658 + ], + "type": "text", + "content": "When deploying a machine learning model to a new environment, we often encounter the distribution shift problem – meaning the target data distribution is different from the model's training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distribution can affect the model's performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a surrogate loss on the test-time unlabeled target data. In particular, minimizing the prediction's entropy on target samples [34] has received much interest as it is task-agnostic and does not require altering the model's training phase (e.g., does not require adding a self-supervised task during training on the source domain). However, as the target data's batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not optimal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an invariance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model's performance under input transformations. The resulting method (TIPI – Test tIme adaPtation with transformation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT [34] in all settings. Our code is released at https://github.com/atuannguyen/TIPI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 241, + 386, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 386, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 386, + 253 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 266, + 547, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 522 + ], + "type": "text", + "content": "Distribution shift is a common problem and is often faced in real-world applications. Specifically, despite taking various precautions while training a machine learning model to ensure a better generalization (e.g., collecting and training on multiple source domains [17, 27], finding flat minima [5], training with meta-learning objectives [16] etc.), the model often still struggles when the test data distribution shifts slightly. Note that it is also common not to have labels for the new shifted domain during test time. To tackle this problem, test time adaptation (also known as online domain adaptation) is a framework that allows the model to adapt to the target distribution using unlabeled test data batches. This is necessary since we typically do not have time to annotate the data during test time and can only make use of the unlabeled data. In this framework, the model needs to give predictions for the target data while simultaneously updating itself to improve its performance on that particular target distribution. We assume a situation in which the target data only arrive in small batches, which makes the adaptation task extremely challenging and renders traditional domain adaptation techniques such as representation alignment (via a distance metric) ineffective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 526, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 547, + 715 + ], + "type": "text", + "content": "Within this test time adaptation framework, a common and effective approach is to optimize a surrogate objective function in lieu of the true loss function on the target data. The first group of surrogate objectives is the loss functions of self-supervised tasks. In particular, one would formulate a user-defined task (such as predicting the rotation angle of an image) and train it alongside the main task on the source domain; and keep training the self-supervised task on the test-time target data [19, 33]. However, these are not fully test-time adaptation methods, since they require altering the training procedure of the source domain. The second line of surrogate objectives is unsupervised loss functions. Among this group of unsupervised objectives, entropy minimization (TENT) [34] is the most successful method, and has been shown to be consistent across many benchmarks. Furthermore, different from the former group," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 193, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 193, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 193, + 713 + ], + "type": "text", + "content": "*The last two authors contributed equally" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24162" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "TENT is a fully test-time adaptation method. For these reasons, TENT has received much interest and a lot of follow-up papers/discussions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 224 + ], + "type": "text", + "content": "However, as pointed out by its authors, TENT is not robust when using small batch sizes, as it often collapses to a trivial solution (i.e., it always predicts the same class for all input). This is detrimental to real-world applications since test data often arrive in small batches. For example, autonomous driving systems only process a few frames in real time – they typically do not accumulate the frames (let's say within one minute) to form a bigger batch. Note that TENT is previously evaluated mainly for large batch sizes (such as 64,128 or 200)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 227, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 286, + 308 + ], + "type": "text", + "content": "In this paper, we aim to tackle the aforementioned problem. Specifically, we aim to develop an unsupervised surrogate objective function for the test time adaptation problem such that it is task-agnostic, does not require altering the training procedure (e.g., does not require incorporating a self-supervised task into the training process), and is more resilient against small batch sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 310, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 286, + 449 + ], + "type": "text", + "content": "We first provide theoretical results regarding a model's performance under input transformations. Specifically, we show that a model's loss on a data distribution is bounded by the KL distance on the predictive distribution of the data before and after the transformations (which we will use as a regularizer), and its loss on the transformed data distribution. Motivated by this result, we use small shifts in the input images that can simulate real source-target shifts, and enforce the network to be invariant under such data transformations. Our model outperforms TENT (and other relevant baselines) in all problem settings considered in the paper and is remarkably robust in the small-batch-size regime." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 451, + 240, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 451, + 240, + 462 + ], + "spans": [ + { + "bbox": [ + 59, + 451, + 240, + 462 + ], + "type": "text", + "content": "Our contributions in this paper are threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 475, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 59, + 475, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 475, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 59, + 475, + 286, + 556 + ], + "type": "text", + "content": "- We provide theoretical results regarding a model's performance under transformations of the input data. Specifically, a model's performance on the target domain is bounded by an invariance term (maximum KL divergence of the predictive distributions before and after the transformations) and its loss on the transformed domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 571, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 571, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 59, + 571, + 286, + 640 + ], + "type": "text", + "content": "- We propose to find input transformations that can simulate the domain shifts, and enforce the network to be invariant under such transformations, using a regularizer based on our derived bound. The resulting method is TIPI (Test tIme adaPtation with transformation Invariance)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 655, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 655, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 655, + 286, + 712 + ], + "type": "text", + "content": "- We perform extensive experiments on a wide range of datasets (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17) and settings (varying batch sizes) to validate our method. TIPI shows preferable performance compared to relevant baselines." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 90, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 230 + ], + "type": "text", + "content": "Distribution Shift is an important and commonly faced problem when deploying machine learning models in real-world applications. There are two main approaches that try to alleviate this problem. The first setting is domain generalization [1, 17, 23, 26, 27], where one aims to collect multiple source domains and learn a generalizable model based on those domains. Another setting is (unsupervised) domain adaptation [2, 6, 18, 21, 28], where a model is trained on a labeled source dataset and we attempt to adapt that model to a target distribution with the help of unlabeled target data. This paper focuses on the latter setting, in particular, the test-time domain adaptation variation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 236, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 713 + ], + "type": "text", + "content": "Test Time Domain Adaptation: In this setting, the unlabeled target data is not available during the training phase of the source model. Instead, the model only has access to the target data when deployed in the target environment, as the stream of test data batches. As a consequence, the unlabeled target data is not available in large quantities, but only in small batches. This makes the adaptation task extremely challenging. So far, there are two main lines of work dealing with this online adaptation problem. The first approach is based on the batch normalization layer in the network, with the idea of using the test-time batch statistics for these layers instead of the source statistics [24, 32]. This can partially help with the representation misalignment between the source and target distributions but is only applicable when the network has BatchNorm layers and the batch size is large. The other line of work is to use an unsupervised surrogate loss to \"refine\" the model on the target data. For example, TENT [34] minimizes the prediction entropy of the model on the target data. However, as pointed out by its authors, TENT needs a large batch size to avoid collapsing to a trivial prediction (only output one class), due to the characteristics of this surrogate objective. Meanwhile, TTT [33] and TTT++ [19] train a self-supervised task alongside the main task on the source domain, and continue to train the self-supervised task during the online adaptation phase. Therefore, TTT and TTT++ are not fully off-the-shelf test-time adaptation methods, since they require modifications to the training phase of the source domain (to incorporate the self-supervised task). Recently, [8] generalizes TENT to other source loss functions (beyond the typical cross-entropy loss). There also have been several works [29] that study datapoints selection strategies for optimizing the unsupervised surrogate loss (e.g., TENT): only optimize the loss for reliable and informative datapoints. Note that this line of research is orthogonal/complementary to ours and we believe that our surrogate loss will also benefit from similar datapoints selection strategies. We leave exploring this direction to future work. Using a completely different approach, T3A [12] replaces the last linear layer (classification layer) with a nonparametric classifier and keeps updating the support set of each class with the new target" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24163" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "datapoints. T3A is robust against small batch sizes (since it is not optimization-based), but is not as effective as other optimization-based methods (e.g., TENT, TTT, and ours)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 113, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 113, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 113, + 287, + 253 + ], + "type": "text", + "content": "Connection between Adversarial Robustness and Robustness against Distribution shift: It has been observed that adversarily trained models generally perform better under certain types of domain shifts (for example, see the leaderboard of Robustbench [4] for Cifar10-C and Cifar100-C). Indeed, one might expect this to be the case if the space of adversarial attacks matches the space of distribution shifts. While it is not previously well understood, our paper sheds some light on this phenomenon. Note that different from previous works, our method enforces an unsupervised robustness loss during the online adaptation phase, not at the source training stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 262, + 113, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 113, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 113, + 276 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 282, + 143, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 143, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 143, + 295 + ], + "type": "text", + "content": "3.1. Problem setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": "Assume that we have a source domain with the distribution " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "p_S(x,y)" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " is the input and " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " is the prediction target. Similar to prior works, we consider the classification problem, where " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " (however, the approach can be straightforwardly extended to regression problems). Suppose that using data from the source domain, we already successfully trained a model " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "f_{\\theta}: \\mathcal{X} \\to \\mathbb{R}^{C}" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " is the parameter of the neural network and " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x)" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " outputs the prediction probabilities for the " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " classes " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "(\\|f_{\\theta}(x)\\|_1 = 1" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "0 \\leq f_{\\theta}(x)_i \\leq 1, \\forall i \\in \\{1,2,\\dots,C\\})" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": ". This leads to a predictive distribution " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\theta}(y|x)" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": ", which is the categorical distribution where " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\theta}(y = i|x) = f_{\\theta}(x)_i" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": ". Note that, for our method, we will not assume or require any modifications to the training process of the model " + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 300, + 287, + 498 + ], + "type": "text", + "content": " on the source domain - a setting that is common and useful (e.g., a Resnet50 model downloaded directly from torchvision)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 567 + ], + "type": "text", + "content": "After that, this model is deployed to the target environment with data distribution " + }, + { + "bbox": [ + 46, + 498, + 287, + 567 + ], + "type": "inline_equation", + "content": "p_T(x,y)" + }, + { + "bbox": [ + 46, + 498, + 287, + 567 + ], + "type": "text", + "content": ". The model receives target data as a stream of unlabeled mini-batches (probably with small and varying sizes). The task is to adapt to this environment simultaneously and continually with only the unlabeled mini-batches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "type": "text", + "content": "The expected loss of this model on a data distribution " + }, + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "type": "inline_equation", + "content": "q(x,y)" + }, + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "type": "text", + "content": " (for example, " + }, + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "type": "inline_equation", + "content": "q(x,y) = p_T(x,y)" + }, + { + "bbox": [ + 46, + 567, + 287, + 593 + ], + "type": "text", + "content": ") is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 599, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 599, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 82, + 599, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\ell (\\theta , q (x, y)) = \\mathbb {E} _ {q (x)} \\left[ l [ q (y | x), \\hat {p} _ {\\theta} (y | x) ] \\right], \\tag {1}", + "image_path": "9a62550c0c208e72528f23fb4b36fa8a43e07e43e49de76378831351b56c15fe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "text", + "content": " is a loss function (distance/divergence) between the true labeling distribution " + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "inline_equation", + "content": "q(y|x)" + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "text", + "content": " and the predictive distribution " + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\theta}(y|x)" + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "text", + "content": ". This loss is not actually computed in practice for both the source and target data (we simply do not have labeled data for both of those), and it is used for theoretical discussion only. To make the theoretical discussion simpler, in this paper, we will consider the simple " + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 46, + 620, + 287, + 713 + ], + "type": "text", + "content": " distance between the two distributions. However," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "note that our theoretical discussion also holds for other divergences such as the Jensen-Shannon divergence (or symmetric KL). Both of these distances satisfy the triangle inequality, which we use in our derivation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 172 + ], + "type": "text", + "content": "3.2. Test Time Domain Adaptation with Transformation Invariance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 178, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 247 + ], + "type": "text", + "content": "Recall that the task is to gradually update the network parameter " + }, + { + "bbox": [ + 304, + 178, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 178, + 545, + 247 + ], + "type": "text", + "content": " so that the model can adapt to the target domain. In this setting, we do not have access to the source data nor a large amount of target data (only small batches), so traditional domain adaptation approaches such as representation alignment are often infeasible." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 248, + 546, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 546, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 546, + 481 + ], + "type": "text", + "content": "Given access only to the batches of unlabeled target data, it is natural to try to come up with a surrogate objective to optimize in lieu of the true loss function (which is not available without the labels). In this sense, this has a close resemblance to the semi-supervised learning problem. In both settings, we aim to find a surrogate function to optimize for unlabeled data. In fact, both the entropy-minimization and pseudo-labeling techniques have been borrowed from the semi-supervised literature, with varying degrees of success. It should be noted that, in semi-supervised learning, the surrogate loss on the unlabeled data is minimized jointly with the supervised objective for labeled data, thus model collapsing is often not a problem. It is much more challenging, however, in this online adaptation setting, since the unsupervised surrogate loss is optimized alone, which can lead to the catastrophic collapse of the network parameter (such that it can not perform prediction anymore, even for the source domain). For example, as pointed out by its authors, TENT [34] tends to collapse to a trivial model if the batch size is small, which is often the case in practice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 597 + ], + "type": "text", + "content": "In this paper, we examine the invariance regularization technique in the setting of test time domain adaptation. Note that a similar regularizer ([3,22,31]) has been applied to semi-supervised learning. However, to the best of our knowledge, we are the first to investigate this type of regularization in the online domain adaptation problem. We aim to investigate both theoretically and empirically how this technique can help with adapting to the new target domain. We also find that this surrogate loss is much more robust (against small batch sizes) compared to TENT." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 597, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 668 + ], + "type": "text", + "content": "To summarize, we propose to enforce the network to be invariant under a predefined space of transformations on the data on the target domain. The idea is that this space of transformations contains the actual domain shifts, or at least could simulate them. Specifically, we minimize " + }, + { + "bbox": [ + 304, + 597, + 545, + 668 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 597, + 545, + 668 + ], + "type": "text", + "content": " with the following objective:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 667, + 545, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 667, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 331, + 667, + 545, + 693 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {p _ {T} (x)} \\left[ \\max _ {x ^ {\\prime} \\in A (x)} \\mathrm {K L} \\left[ \\hat {p} _ {\\theta} \\left(y \\mid x ^ {\\prime}\\right), \\hat {p} _ {\\bar {\\theta}} (y \\mid x) \\right] \\right], \\tag {2}", + "image_path": "9b3a06fb53ccaa064f85705cbf5b5f78d5b9dffbf0d1bd98d0a6065f4cf9733c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "inline_equation", + "content": "\\overline{\\theta}" + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "text", + "content": " is an identical copy of " + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "inline_equation", + "content": "A(x) = \\{t(x)\\}_{t\\in S_t}" + }, + { + "bbox": [ + 306, + 700, + 544, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24164" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "content": "Algorithm 1 Adapt and predict. " + }, + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "content": " is the network parameter, " + }, + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "content": " is the batch data, " + }, + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "content": " is the learning rate for online adaptation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 110, + 171, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 110, + 171, + 122 + ], + "spans": [ + { + "bbox": [ + 52, + 110, + 171, + 122 + ], + "type": "text", + "content": "Input " + }, + { + "bbox": [ + 52, + 110, + 171, + 122 + ], + "type": "inline_equation", + "content": "\\theta ,\\{x_{1},x_{2},\\ldots ,x_{n}\\}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 122, + 247, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 122, + 247, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 122, + 247, + 135 + ], + "type": "text", + "content": "Output " + }, + { + "bbox": [ + 53, + 122, + 247, + 135 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{new}}" + }, + { + "bbox": [ + 53, + 122, + 247, + 135 + ], + "type": "text", + "content": " (newly adapted), " + }, + { + "bbox": [ + 53, + 122, + 247, + 135 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}_1,\\hat{y}_2,\\dots ,\\hat{y}_n\\}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 136, + 287, + 217 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 53, + 136, + 256, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 256, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 256, + 148 + ], + "type": "text", + "content": "1: function ADAPT-PREDICT " + }, + { + "bbox": [ + 53, + 136, + 256, + 148 + ], + "type": "inline_equation", + "content": "(\\theta ,\\{x_{1},x_{2},\\dots,x_{n}\\})" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 148, + 287, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 159 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 53, + 148, + 287, + 159 + ], + "type": "inline_equation", + "content": "p_i \\gets f_\\theta(x_i)" + }, + { + "bbox": [ + 53, + 148, + 287, + 159 + ], + "type": "text", + "content": ". detach() ▷ stop gradient" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 159, + 241, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 241, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 241, + 171 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 53, + 159, + 241, + 171 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}\\gets \\arg \\max_{x^{\\prime}\\in A(x_{i})}\\mathrm{KL}[p_{\\theta}(y|x^{\\prime}),p_{i}]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 171, + 211, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 211, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 211, + 185 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 53, + 171, + 211, + 185 + ], + "type": "inline_equation", + "content": "l(\\theta) = \\frac{1}{n}\\sum_{i=1}^{n}\\mathrm{KL}[p_{\\theta}(y|x_i'), p_i]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 184, + 164, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 164, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 164, + 196 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 53, + 184, + 164, + 196 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{new}} = \\dot{\\theta} -\\alpha \\nabla_{\\theta}l(\\theta)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 196, + 173, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 173, + 208 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 173, + 208 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 53, + 196, + 173, + 208 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = \\arg \\max f_{\\theta_{\\mathrm{new}}}(x_i)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 208, + 120, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 120, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 120, + 217 + ], + "type": "text", + "content": "7: end function" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 240, + 226, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 226, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 226, + 251 + ], + "type": "inline_equation", + "content": "(S_{t}" + }, + { + "bbox": [ + 47, + 240, + 226, + 251 + ], + "type": "text", + "content": " is the predefined set of transformations)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 251, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 251, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 45, + 251, + 287, + 287 + ], + "type": "text", + "content": "We present the motivations and derivations of this objective function in the next subsection. The adapting and predicting procedure is depicted in Algorithm 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 300, + 237, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 237, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 237, + 313 + ], + "type": "text", + "content": "3.2.1 Invariance Regularization Technique" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 319, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 365 + ], + "type": "text", + "content": "In this subsection, we introduce the invariance regularization technique for general transformations and provide theoretical results regarding the online domain adaptation problem, which will later motivate our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": "Let's consider a set of transformations " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "t \\in S_{t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " is a transformation of the data, i.e., " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "t: \\mathcal{X} \\to \\mathcal{X}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": ". Ideally, " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " should contain the transformations responsible for the domain shift. This set can contain transformations such as data augmentations, GAN-like cross-domain transformations, random noise perturbation, etc. For now, we shall not care much about the type of transformations in " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": ". The only important detail is that these transformations should preserve the class information (this is formalized in Assumption 1). Let " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "A(x) = \\{t(x)\\}_{t \\in S_t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "p_t(t)" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " be a distribution over " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " that assigns probability mass (or density) to each transformation " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "t \\in S_{t}" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": ". Each distribution " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "p_t(t)" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " defines a conditional distribution " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "T(x^t | x)" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " (with the support set " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "A(x)" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": "), where " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "x^t = t(x), t \\sim p_t(t)" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": ". This in turn defines a joint distribution over " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "x, x^t" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 366, + 287, + 540 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 99, + 544, + 287, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 544, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 99, + 544, + 287, + 559 + ], + "type": "interline_equation", + "content": "p _ {T} (x, x ^ {t}, y) = p _ {T} (x, y) T \\left(x ^ {t} | x\\right). \\tag {3}", + "image_path": "08a470cb453ed83098e6facd63048dabf39140ee4ce93ddbeb8128404446cd56.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "text", + "content": "It is also worth noting that we have transformed data distribution " + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "inline_equation", + "content": "p_T(x^t, y)" + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "text", + "content": " (just marginalizing out " + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 566, + 287, + 601 + ], + "type": "text", + "content": " from the above)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 608, + 111, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 111, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 111, + 619 + ], + "type": "text", + "content": "Assumption 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 121, + 627, + 287, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 627, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 121, + 627, + 287, + 640 + ], + "type": "interline_equation", + "content": "I _ {T} (x, y) = I _ {T} \\left(x ^ {t}, y\\right), \\tag {4}", + "image_path": "ebd16f0c6add4a888f0010add7bc981752853646d67f67202359409877406c95.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 648, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 648, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 287, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 648, + 287, + 683 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 46, + 648, + 287, + 683 + ], + "type": "text", + "content": " is the mutual information, calculated on the target domain (please refer to the Supplementary Material for their exact formula)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Intuitively, this assumption means that the transformed image " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " contains the same amount of information about" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "the label " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " as the original image " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " does. This assumption is equivalent to " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p_T(x,x^t)}[\\mathrm{KL}[p_T(y|x),p_T(y|x^t)]] = 0" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " (a derivation can be found in the supplementary material), meaning that " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "p_T(y|x) = p_T(y|x^t)" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " almost surely." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 122, + 545, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 122, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 122, + 545, + 146 + ], + "type": "text", + "content": "The transformation invariance regularizer on the target domain can be defined simply as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 341, + 148, + 545, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 148, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 341, + 148, + 545, + 169 + ], + "type": "interline_equation", + "content": "h (\\theta , x) = \\max _ {x ^ {\\prime} \\in A (x)} D \\left[ \\hat {p} _ {\\bar {\\theta}} (y | x), \\hat {p} _ {\\theta} (y | x ^ {\\prime}) \\right], \\tag {5}", + "image_path": "634980cdf062cbadb0bb1df2c9d792fbc7864a8ff62c612f566f89639bde33c6.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "inline_equation", + "content": "D[\\cdot, \\cdot]" + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "text", + "content": " is a divergence/distance of choice and " + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}" + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "text", + "content": " is an identical copy of " + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 177, + 545, + 270 + ], + "type": "text", + "content": " (so that the gradient is not backpropagated through it). We will see later (through an analysis of the gradient) that this is extremely important. Intuitively, it also helps to alleviate the network collapse problem, since, without this gradient blocking, the network would easily collapse to a constant output to minimize the regularizer term." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": "Intuitively, we can view this objective as using " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\overline{\\theta}}(y|x)" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": " as the target to optimize " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\hat{p}_{\\theta}(y|t(x))" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": " for all transformation " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": ". This makes sense because, as discussed earlier, " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": " does not change the semantic meaning of " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": " significantly, so the predictive distribution of " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "inline_equation", + "content": "t(x)" + }, + { + "bbox": [ + 304, + 270, + 545, + 329 + ], + "type": "text", + "content": " should be similar." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 329, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 545, + 352 + ], + "type": "text", + "content": "Based on this regularizer, we can bound the loss on the target domain as follows:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 358, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 545, + 381 + ], + "type": "text", + "content": "Proposition 1. Under Assumption 1, if " + }, + { + "bbox": [ + 305, + 358, + 545, + 381 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 358, + 545, + 381 + ], + "type": "text", + "content": " in Eq. 5 is either the forward or reverse KL divergence, we have:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 311, + 388, + 539, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 388, + 539, + 409 + ], + "spans": [ + { + "bbox": [ + 311, + 388, + 539, + 409 + ], + "type": "interline_equation", + "content": "\\ell (\\theta , p _ {T} (x, y)) \\leq \\mathbb {E} _ {p _ {T} (x)} \\left[ \\sqrt {2 h (\\theta , x)} \\right] + \\ell (\\theta , p _ {T} (x ^ {t}, y)).", + "image_path": "886c09045767695e6ee4e223f9081307e0912e49ac734b6538320b2864ea3e70.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 415, + 499, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 499, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 499, + 426 + ], + "type": "text", + "content": "Proof. Provided in the Supplementary Material." + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 535, + 415, + 545, + 424 + ], + "blocks": [ + { + "bbox": [ + 535, + 415, + 545, + 424 + ], + "lines": [ + { + "bbox": [ + 535, + 415, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 535, + 415, + 545, + 424 + ], + "type": "image", + "image_path": "902ea96dd6d39510cb0380d3819e426ff6444d283bdfd9027a42e19752302c8b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 434, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 468 + ], + "type": "text", + "content": "This means that the loss on the target distribution is bounded by the invariance regularizer and the loss on the transformed distribution " + }, + { + "bbox": [ + 304, + 434, + 545, + 468 + ], + "type": "inline_equation", + "content": "p_T(x^t, y)" + }, + { + "bbox": [ + 304, + 434, + 545, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": "Now, assume for a moment (even if it is not realistic) that the set " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": " contains all the transformations that can successfully transform the target data distribution " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "p_{T}(x,y)" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": " back to the source distribution " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "p_{S}(x,y)" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": " (i.e., there exists a distribution " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "p_{t}(t)" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": " such that the resulting transformed distribution " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "p_{T}(x^{t},y)" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": " is the same with " + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "inline_equation", + "content": "p_{S}(x,y)" + }, + { + "bbox": [ + 304, + 468, + 545, + 666 + ], + "type": "text", + "content": "). Then, we have that the loss on the target domain is bounded by the loss on the source domain and our invariance regularizer. Assume also that while adapting to the new target domain, our model does not suffer from catastrophic forgetting of the source domain (which is often the case if the network does not collapse since we are adapting to the same task in a different environment). Then we have that the loss on the source domain would be small, and the invariance regularizer would also be small (since we explicitly minimize it). As a consequence, the loss on the target domain would be small and we can guarantee a good adaptation performance." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "However, knowing the exact transformations from the target domain back to the source domains is often infeasible (only possible if we have some prior domain knowledge). This motivates us to find other transformations that" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24165" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "are task-agnostic and can simulate the distribution shift between the target and source domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 110, + 255, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 255, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 255, + 121 + ], + "type": "text", + "content": "3.2.2 Practical choice of the transformation set" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 128, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 128, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 128, + 287, + 268 + ], + "type": "text", + "content": "We want to choose task-agnostic transformations such that they do not alter the semantic meaning of the image significantly, while they can simulate the distribution shift. The first idea that comes to mind is to use common image augmentations (e.g., random crop, random vertical flip, etc.) However, our early experiments suggest that this method is not consistent across corruption types/domain shifts. This makes sense since image augmentations are often heuristic, and they may be able to simulate certain types of distribution shifts, but not others. Therefore, we opt to use " + }, + { + "bbox": [ + 46, + 128, + 287, + 268 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 46, + 128, + 287, + 268 + ], + "type": "text", + "content": "-norm based perturbations. Empirically, this leads to better performance and also is consistent across domains/tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 275, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 369 + ], + "type": "text", + "content": "To summarize, we use the " + }, + { + "bbox": [ + 46, + 275, + 287, + 369 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 46, + 275, + 287, + 369 + ], + "type": "text", + "content": "-norm based perturbations as the transformations of images. Empirically, we also found out that using the reverse KL leads to better performance (although almost all distances/divergences have positive results). We provide explanations/intuitions as to why we should prefer the reverse KL, and why adversarial examples can help with the distribution shift problem in the next subsection. Our final optimization problem is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 374, + 287, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 374, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 59, + 374, + 287, + 402 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {p _ {T} (x)} \\left[ \\max _ {| | x ^ {\\prime} - x | | _ {\\infty} \\leq \\epsilon} \\mathrm {K L} \\left[ \\hat {p} _ {\\theta} \\left(y \\mid x ^ {\\prime}\\right), \\hat {p} _ {\\bar {\\theta}} (y \\mid x) \\right] \\right]. \\tag {6}", + "image_path": "9807ab22436736825bd311167dfa1fa4ffb71f69d90120fe3702d2690e81741a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": "We name our method TIPI (Test tIme adaPtation with transformation Invariance). To save computational cost, we use the single-step adversarial method FGSM [7] to find the adversarial example " + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": ". Note that one can also use more advanced methods such as PGD [20] to find the adversarial perturbation to get better results (but with the trade-off of speed). We set " + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": " to 2/255 in all of our experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 504, + 178, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 178, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 178, + 514 + ], + "type": "text", + "content": "3.2.3 Additional Discussions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "text", + "content": "Why preferring reverse KL over forward KL? Natural choices for " + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "text", + "content": " are the KL-divergences (forward or reverse KL), since they can be computed straightforwardly for the predictive distributions, and also are commonly used in practice for similar problems (e.g., VAT [22] uses the forward KL). Our experiments empirically suggest that using the reverse KL as the divergence " + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "text", + "content": " leads to better results compared to the forward KL. Here, we also give theoretical intuition as to why it is the case. First of all, let's consider the gradient of " + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 521, + 287, + 649 + ], + "type": "text", + "content": " with respect to both the forward and reverse KL:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 650, + 130, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 130, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 130, + 661 + ], + "type": "text", + "content": "For the forward KL:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 667, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 667, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 52, + 667, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathrm {K L} [ \\hat {p} _ {\\bar {\\theta}} (y | x), \\hat {p} _ {\\theta} (y | x ^ {\\prime}) ] = \\sum_ {y \\in \\mathcal {Y}} - \\frac {\\hat {p} _ {\\bar {\\theta}} (y | x)}{\\hat {p} _ {\\theta} (y | x ^ {\\prime})} \\nabla_ {\\theta} \\hat {p} _ {\\theta} (y | x ^ {\\prime}). \\tag {7}", + "image_path": "0ad52104ada808fa04f15fa9dc909ea2808862325dc6158f63437c9772a33b13.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "text", + "content": "Meanwhile, for the reverse KL:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 98, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 98, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 321, + 98, + 545, + 144 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} \\mathrm {K L} [ \\hat {p} _ {\\theta} (y | x ^ {\\prime}), \\hat {p} _ {\\bar {\\theta}} (y | x) ] \\\\ = \\sum_ {y \\in \\mathcal {Y}} \\left(- \\log \\frac {\\hat {p} _ {\\bar {\\theta}} (y | x)}{\\hat {p} _ {\\theta} (y | x ^ {\\prime})} + 1\\right) \\nabla_ {\\theta} \\hat {p} _ {\\theta} (y | x ^ {\\prime}). \\tag {8} \\\\ \\end{array}", + "image_path": "2b4fbd0e5d3f0c6ed826b92549cdb26ede709232e494e908a22b4ff03142b6ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 159, + 545, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 218 + ], + "type": "text", + "content": "Compare Eq. 8 to Eq. 7, we can see that the coefficient of " + }, + { + "bbox": [ + 304, + 159, + 545, + 218 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\hat{p}_{\\theta}(y|x^{\\prime})" + }, + { + "bbox": [ + 304, + 159, + 545, + 218 + ], + "type": "text", + "content": " in the reverse KL gets squashed down by the log function, leading to much smaller variance. We conjecture that the small variance in gradient is the reason why reverse KL performs better than forward KL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "type": "text", + "content": "How can invariance enforcing with adversarial examples help with the distribution shift problem? In our previous subsection, we provide theoretical evidence that motivates us to find transformations that can simulate common distribution shifts in practice. We argue here that " + }, + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "type": "text", + "content": " adversarial perturbation is a good candidate for this. First of all, " + }, + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 238, + 546, + 472 + ], + "type": "text", + "content": " adversarial perturbation is a task-agnostic transformation that does not require much tuning, and we do not need to design heuristic transformations per domain. Secondly, a lot of the common domain shifts in practice are image corruptions that add noises to the images (gaussian noise, gaussian blur, jpeg compression noise, etc.), which can be simulated very well by adversarial noises. Lastly, even for more global domain shifts as seen in the DIGITS and VisDA17 experiments (Section 4), we conjecture that training a network that is robust against local and small changes (adversarial noise) would also help to make it more robust against larger shifts. Our empirical experiments also agree with this, as TIPI performs consistently across multiple types of domain shifts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 491, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 546, + 713 + ], + "type": "text", + "content": "Implementation Details: Following TENT [34] and other test-time surrogate optimization methods, we only optimize the affine parameters of the model (i.e., the weight and bias parameters of the norm layers: BatchNorm, LayerNorm, InstanceNorm, or GroupNorm). This has been shown to lead to a more stable optimization, and also reduces the computational cost, as opposed to optimizing all parameters of the network. However, note that optimizing over the entire representation network (similar to SHOT [18]) is also possible. Furthermore, in case the network uses BatchNorm layers, we replace each BatchNorm with two BatchNorm layers (following [35]), one for the clean images and one for the adversarial ones. This is because the batch statistics of the adversarial examples are much different from that of the clean images [35], thus mixing them might cause performance degradation. Note that these two BatchNorm layers only keep separate batch statistics (feature mean and variance), while sharing the affine parameters (weight and bias) – the parameters being optimized." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24166" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 244 + ], + "type": "text", + "content": "We test the effectiveness of our method on a wide variety of tasks (robustness against image corruptions and domain adaptation). Especially, we performed extensive experiments on various batch sizes of the test-time data batches (as small as 2 and as large as 200). The highlight of our empirical findings is that: TIPI outperforms other test-time adaptation baselines in all scenarios and problem settings. We discuss these experiments in more detail in the following subsections. As a side note, we also found that using TIPI with TENT further improves the performance of both methods, and helps TENT avoid network collapse. Details of our experiments (such as learning rate) can be found in our publicly available source code." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 254, + 233, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 233, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 233, + 266 + ], + "type": "text", + "content": "4.1. Evaluation Protocol and Baselines:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 273, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 367 + ], + "type": "text", + "content": "Evaluation Protocol: We consider a setting where the model is trained normally on the source domain. When the models are available online (e.g., ImageNet-trained Resnet50), we download the weights and keep them as is. Otherwise, we train the models ourselves. Following TENT [34], we perform online adaptation, and report the average performance across all batches (while the model is being adapted). This mimics a real-world scenario." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 386, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 572 + ], + "type": "text", + "content": "Baselines: The main purpose of our experiments is to compare our invariance regularizer against predictive entropy as an unsupervised surrogate subjective. Therefore, TENT [34] is our main baseline in all of the experiments. We also consider recent and relevant fully test-time adaptation methods as our baselines, such as BatchNorm with test-time statistics (BN) [24, 32] and T3A [12]. We also report the performance of the source-only model (source-no adaptation). Note that we also have tested T3A with test-time batch norm statistics (T3A+BN), but this model's performance is almost identical to (or even a little worse than) that of BN for all batch sizes, so we do not include it here. For a fair comparison, we rerun all methods in our experiment (to eliminate artifacts such as source model training), using their official source code. However, our numbers for these methods are very close to the reported numbers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "type": "text", + "content": "TTT [33] and " + }, + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TTT} + +" + }, + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "type": "text", + "content": " [19] are not fully test-time adaptation methods since they require altering the training process of the source domain (to incorporate a self-supervised task). Thus, they are not optimal for off-the-shelf models that have not been trained with the self-supervised task (e.g., a ResNet50 trained on ImageNet downloaded directly from torchvision). For this reason, we do not include these two methods in the comparisons. Note that as reported in Table 2 of the " + }, + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TTT} + +" + }, + { + "bbox": [ + 46, + 573, + 288, + 714 + ], + "type": "text", + "content": " paper [19], the variant of their model which has not been trained with the self-supervised task (which is the same as our evaluation protocol and most of the literature) largely underperforms our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "EATA [29] proposes a datapoints selection strategy for TENT, and achieves state-of-the-art performance on the test time adaptation task. As discussed before, we believe this line of research is complementary to ours. Indeed, we are able to incorporate TIPI into EATA, thereby improving its robustness (EATA only uses TENT and is not robust against small batchesizes). Due to the page limit, we defer this comparison to the Supplementary Material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 173, + 486, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 486, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 486, + 186 + ], + "type": "text", + "content": "4.2. Robustness to Image Corruptions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 190, + 434, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 434, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 434, + 204 + ], + "type": "text", + "content": "4.2.1 Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 209, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 234 + ], + "type": "text", + "content": "Following the literature, we conduct experiments on widely-used image-corruption datasets, namely:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 240, + 545, + 354 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 240, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 545, + 300 + ], + "type": "text", + "content": "- CIFAR10-C and CIFAR100-C [10]: Common image corruptions are applied to the CIFAR10 and CIFAR100 datasets [14]. We report the results for the highest level of severity, averaging over 15 types of corruption." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 307, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 307, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 307, + 545, + 354 + ], + "type": "text", + "content": "- ImageNet-C: Similar with the above two datasets, but for ImageNet [30]. We report the performance of all methods averaged over 15 types of corruption and 5 levels of severity." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 361, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 465 + ], + "type": "text", + "content": "For the CIFAR10-C and CIFAR100-C datasets, we use a WideResnet28-10 [36] model, which is the base model in Robustbench [4] and also used in TENT's official GitHub repository1. Note that in the original paper, TENT conducts experiments with a Resnet26 model [9]; however, the model/code was not publicly released, and there have been numerous reproducibility issues as reported in their GitHub page2. We use the Adam optimizer [13] for these two datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 502 + ], + "type": "text", + "content": "For the ImageNet-C dataset, we follow previous works and use a Resnet50 network. We use SGD to optimize our invariance regularizer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 516, + 371, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 371, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 371, + 526 + ], + "type": "text", + "content": "4.2.2 Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 686 + ], + "type": "text", + "content": "Table 1, 2 and 3 show that our method is the best performer among the baselines. In the CIFAR10-C and CIFAR100-C experiments, our method performs similarly with TENT for large batch sizes (200). However, as the batch size decreases, TENT collapses and TIPI outperforms it significantly. For the challenging ImageNet-C dataset, our method performs better than TENT even for large batch sizes. TENT still collapses with small batch sizes while TIPI is very robust to this parameter. T3A is not an optimization-based adaptation approach so its performance is robust against the batch size, but it largely underperforms TIPI in all settings. Even more concerning, T3A performs extremely purely on ImageNet-C (worse than the source" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 693, + 443, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 443, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 443, + 703 + ], + "type": "text", + "content": "1https://github.com/DequanWang/tent" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 703, + 462, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 462, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 462, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 317, + 703, + 462, + 713 + ], + "type": "text", + "content": "https://github.com/DequanWang/tent/issues" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24167" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 156, + 87, + 438, + 209 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 485, + 83 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 485, + 83 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 485, + 83 + ], + "type": "text", + "content": "Table 1. CIFAR10-C: Results are averaged over 15 types of corruption with the highest level of severity" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 87, + 438, + 209 + ], + "lines": [ + { + "bbox": [ + 156, + 87, + 438, + 209 + ], + "spans": [ + { + "bbox": [ + 156, + 87, + 438, + 209 + ], + "type": "table", + "html": "
MethodAccuracy
Source56.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A57.857.857.857.857.957.957.8
BN61.870.875.177.578.979.674.0
TENT17.460.774.078.681.081.465.5
TIPI (ours)78.679.880.381.081.281.580.4
", + "image_path": "9a80600e4c224729f72f1954e2b59aaa321b53f860ce9be6c80b00b9cc4711bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 156, + 236, + 438, + 357 + ], + "blocks": [ + { + "bbox": [ + 105, + 218, + 487, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 487, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 487, + 231 + ], + "type": "text", + "content": "Table 2. CIFAR100-C: Results are averaged over 15 types of corruption with the highest level of severity" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 236, + 438, + 357 + ], + "lines": [ + { + "bbox": [ + 156, + 236, + 438, + 357 + ], + "spans": [ + { + "bbox": [ + 156, + 236, + 438, + 357 + ], + "type": "table", + "html": "
MethodAccuracy
Source37.5
Test Time Adaptation Batch SizeAverage
25102050200
T3A39.039.039.039.039.039.039.0
BN26.048.054.157.459.560.750.9
TENT1.57.439.859.964.465.739.8
TIPI (ours)56.362.364.165.065.565.863.2
", + "image_path": "91ae79be74bd2b8e5c170a17ab91de52dfee058902449e69289b2adbb02f508b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 156, + 383, + 438, + 505 + ], + "blocks": [ + { + "bbox": [ + 124, + 366, + 468, + 378 + ], + "lines": [ + { + "bbox": [ + 124, + 366, + 468, + 378 + ], + "spans": [ + { + "bbox": [ + 124, + 366, + 468, + 378 + ], + "type": "text", + "content": "Table 3. ImageNet-C: Results are averaged over 15 types of corruption and 5 levels of severity" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 383, + 438, + 505 + ], + "lines": [ + { + "bbox": [ + 156, + 383, + 438, + 505 + ], + "spans": [ + { + "bbox": [ + 156, + 383, + 438, + 505 + ], + "type": "table", + "html": "
MethodAccuracy
Source39.5
Test Time Adaptation Batch SizeAverage
248163264
T3A30.230.230.230.330.330.330.3
BN13.933.442.846.948.949.947.2
TENT0.813.144.353.756.857.345.2
TIPI (ours)45.253.356.757.958.358.455.0
", + "image_path": "0ae139580f1d99ee5ef0b06f5cf79e97322274162d3e3d743e75583001e16cfe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 287, + 571 + ], + "type": "text", + "content": "model) – we conjecture this is because the non-parametric classifier in T3A is not effective when there is a large number of classes (1000) and each class's support set has a few datapoints (50)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 163, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 163, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 163, + 590 + ], + "type": "text", + "content": "4.3. Domain Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 595, + 175, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 595, + 175, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 595, + 175, + 608 + ], + "type": "text", + "content": "4.3.1 Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 613, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 638 + ], + "type": "text", + "content": "In this problem, we consider two main adaptation tasks, with the following two datasets:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 643, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 643, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 643, + 289, + 714 + ], + "type": "text", + "content": "- DIGITS: the task is digit classification (10 classes), and we have to adapt a model from SVHN [25] to MNIST [15], MNIST-M [6], and USPS [11]. For the MNIST and MNIST-M datasets, we only use the test set (10000 images) for the online adaptation. For the USPS dataset, we use both the training and test set to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 525, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 525, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 324, + 525, + 545, + 548 + ], + "type": "text", + "content": "have roughly the same number of images for adaptation (9298 in total)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 557, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 547, + 617 + ], + "type": "text", + "content": "- VisDA17 (12-class classification)3: is a more challenging dataset, with the adaption task from synthetic to real images. The source dataset contains 152397 3D-rendered synthetic images, while the target dataset contains 55388 real images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 625, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 547, + 694 + ], + "type": "text", + "content": "Following the experiments performed in TENT, we use a ResNet26 for DIGITS and a ResNet50 for the VisDA17 dataset. For the test-time optimization of our surrogate loss, we use Adam for DIGITS and SGD for VisDA17. Other details such as the learning rate can be found in the provided source code." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 512, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 512, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 512, + 714 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 315, + 702, + 512, + 714 + ], + "type": "text", + "content": "Oxford solely obtained and processed the VisDA17 dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24168" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 87, + 479, + 186 + ], + "blocks": [ + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "lines": [ + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "spans": [ + { + "bbox": [ + 128, + 70, + 464, + 82 + ], + "type": "text", + "content": "Table 4. DIGITS: Online Domain Adaptation from SVHN to MNIST, MNIST-M and USPS" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 87, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 115, + 87, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 87, + 479, + 186 + ], + "type": "table", + "html": "
MethodSVHN→MNISTSVHN→MNIST-MSVHN→USPS
B.Size 4B.Size 128B.Size 4B.Size 128B.Size 4B.Size 128
Source85.860.384.9
T3A86.386.360.760.784.984.9
BN75.683.556.060.972.480.8
TENT27.090.719.261.438.790.4
TIPI (ours)86.391.664.263.886.893.9
", + "image_path": "ab19eca82edab4197a5348e99fdc9e7bc7c5b614053fa495c32498bf0d857cd4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 142, + 213, + 452, + 346 + ], + "blocks": [ + { + "bbox": [ + 143, + 197, + 450, + 209 + ], + "lines": [ + { + "bbox": [ + 143, + 197, + 450, + 209 + ], + "spans": [ + { + "bbox": [ + 143, + 197, + 450, + 209 + ], + "type": "text", + "content": "Table 5. VisDA17: Online Domain Adaptation of the VisDA classification challenge" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 213, + 452, + 346 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 452, + 346 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 452, + 346 + ], + "type": "table", + "html": "
MethodAccuracy
Source46.4
Test Time Adaptation Batch SizeAverage
248163264128
T3A48.848.848.948.948.948.948.948.9
BN38.349.554.657.358.759.459.853.9
TENT12.326.553.059.962.162.062.048.3
TENTx210.022.145.959.262.563.163.246.6
TIPI (ours)51.260.362.764.064.364.664.561.7
", + "image_path": "f477e1c95e33dba799d0bc95266f7b5175151ea8bd8eaa96ecbedc7872044c78.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 366, + 111, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 111, + 377 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 111, + 377 + ], + "type": "text", + "content": "4.3.2 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 288, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 467 + ], + "type": "text", + "content": "Table 4 and Table 5 show a similar trend as observed in the image corruption experiments. In particular, TIPI performs the best among all the baselines. Especially, our method does not collapse with small batch sizes as most other models (such as BN or TENT) do. It is clear that TIPI achieves preferable performance in all settings while being much more robust compared to TENT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 474, + 236, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 236, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 236, + 487 + ], + "type": "text", + "content": "4.4. Runtime and memory consumption" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 491, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 654 + ], + "type": "text", + "content": "We would also like to discuss the runtime and memory (VRAM) consumption of our method. As TIPI is also an optimization-based adaptation, it is natural to compare it against TENT. Comparing our implementation of TIPI with TENT's official implementation, we observe that TIPI and TENT require roughly the same amount of VRAM, while our method is about two times slower than TENT (because it needs to find the adversarial examples). We believe that this is worth the trade-off, since our method clearly outperforms TENT, and is much more robust against small batch sizes – which is often the case in practice. To be more specific, using an NVIDIA Quadro RTX 6000, TIPI can reach 160 FPS (including both adaptation and prediction) with the standard 224x224 images of ImageNet and VisDA17." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": "However, just to be completely fair, we also compare TIPI with a variant of TENT where we perform two adaptation steps for each batch (thus its runtime is roughly the same as TIPI's), namely TENTx2 (second-last row in Table 5). We perform this comparison on the challenging" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "content": "adaptation dataset VisDA17 (for smaller datasets such as CIFAR and DIGITS, runtime and memory consumption are almost never a problem). TENTx2 does perform better than TENT (in the large-batch-size regime), but is still largely inferior to TIPI. Additionally, the collapsing problem for small batch sizes persists in this variant." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 453, + 379, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 379, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 379, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 615 + ], + "type": "text", + "content": "In this paper, we propose to use transformation invariance as an unsupervised surrogate loss function for online domain adaptation. Similar to TENT, our method (TIPI) is task-agnostic and does not require altering the source training process of the model, thus can be applied easily to a wide variety of off-the-shelf models. Remarkably, compared to TENT, our method is much more robust against small batch sizes, which proves to be extremely useful in challenging real-world scenarios. A promising future research direction is to investigate a data point selection strategy for optimizing the TIPI objective, to further improve the performance of the method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 637, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 637, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 637, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgments Author A. Tuan Nguyen acknowledges Meta AI for funding his PhD study. Oxford lab headed by author Philip Torr is supported by the UKRI grant: Turing AI Fellowship EP/W002981/1. Meta AI author Ser-Nam Lim is neither supported by the UKRI grant nor has any relationship to the grant." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24169" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using meta regularization. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 286, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 286, + 176 + ], + "type": "text", + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 178, + 286, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 286, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 286, + 231 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution alignment and augmentation anchoring. arXiv preprint arXiv:1911.09785, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 232, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 286, + 285 + ], + "type": "text", + "content": "[4] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020.3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 286, + 286, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 286, + 328 + ], + "type": "text", + "content": "[5] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 286, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 286, + 383 + ], + "type": "text", + "content": "[6] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 384, + 286, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 286, + 416 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 286, + 416 + ], + "type": "text", + "content": "[7] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 418, + 286, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 286, + 449 + ], + "type": "text", + "content": "[8] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and Zico Kolter. Test-time adaptation via conjugate pseudo-labels. arXiv preprint arXiv:2207.09640, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 450, + 286, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 450, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 53, + 450, + 286, + 492 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 494, + 286, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 494, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 494, + 286, + 525 + ], + "type": "text", + "content": "[10] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 527, + 286, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 527, + 286, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 527, + 286, + 559 + ], + "type": "text", + "content": "[11] J. J. Hull. A database for handwritten text recognition research. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(5):550-554, 1994. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "type": "text", + "content": "[12] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 604, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 286, + 634 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 286, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 658 + ], + "type": "text", + "content": "[14] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 659, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 286, + 689 + ], + "type": "text", + "content": "[15] Yann LeCun, Corinna Cortes, and CJ Burges. Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 712 + ], + "type": "text", + "content": "[16] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hesperides. Learning to generalize: Meta-learning for do" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 94 + ], + "type": "text", + "content": "main generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[17] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 149, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 149, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 149, + 545, + 191 + ], + "type": "text", + "content": "[18] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 193, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 193, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 308, + 193, + 545, + 245 + ], + "type": "text", + "content": "[19] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808-21820, 2021. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 247, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 247, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 545, + 289 + ], + "type": "text", + "content": "[20] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 290, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 290, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 290, + 545, + 322 + ], + "type": "text", + "content": "[21] Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms. arXiv preprint arXiv:0902.3430, 2009. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 323, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 323, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 545, + 375 + ], + "type": "text", + "content": "[22] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "text", + "content": "[23] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In International Conference on Machine Learning, pages 10-18. PMLR, 2013. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 420, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 472 + ], + "type": "text", + "content": "[24] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020.2.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 475, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 506 + ], + "type": "text", + "content": "[25] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 507, + 545, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 507, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 308, + 507, + 545, + 539 + ], + "type": "text", + "content": "[26] A Tuan Nguyen, Ser-Nam Lim, and Philip HS Torr. FedSR: A simple and effective domain generalization method for federated learning. 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 540, + 545, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 540, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 308, + 540, + 545, + 582 + ], + "type": "text", + "content": "[27] A Tuan Nguyen, Toan Tran, Yarin Gal, and Atilim Gunes Baydin. Domain invariant representation learning with domain density transformations. Advances in Neural Information Processing Systems, 34:5264-5275, 2021. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 583, + 545, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 583, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 308, + 583, + 545, + 615 + ], + "type": "text", + "content": "[28] A Tuan Nguyen, Toan Tran, Yarin Gal, Philip HS Torr, and Atulim Güneş Baydin. KI guided domain adaptation. arXiv preprint arXiv:2106.07780, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 616, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 616, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 616, + 545, + 658 + ], + "type": "text", + "content": "[29] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yafo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. arXiv preprint arXiv:2204.02610, 2022. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24170" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 345 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 170 + ], + "type": "text", + "content": "[32] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 33:11539-11551, 2020. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 171, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 287, + 224 + ], + "type": "text", + "content": "[33] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 225, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 225, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 225, + 287, + 267 + ], + "type": "text", + "content": "[34] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 287, + 323 + ], + "type": "text", + "content": "[35] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 324, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 345 + ], + "type": "text", + "content": "[36] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24171" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_content_list.json b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6015667ee71edb27d598e796a0ac5866b66a85e9 --- /dev/null +++ b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "TMO: Textured Mesh Acquisition of Objects with a Mobile Device by using Differentiable Rendering", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaehoon Choi $^{1,2}$ Dongki Jung $^{1}$ Taejae Lee $^{1}$ Sangwook Kim $^{1}$ Youngdong Jung $^{1}$ Dinesh Manocha $^{2}$ Donghwan Lee $^{1}$ $^{1}$ NAVER LABS $^{2}$ University of Maryland", + "bbox": [ + 137, + 202, + 830, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a new pipeline for acquiring a textured mesh in the wild with a single smartphone which offers access to images, depth maps, and valid poses. Our method first introduces an RGBD-aided structure from motion, which can yield filtered depth maps and refines camera poses guided by corresponding depth. Then, we adopt the neural implicit surface reconstruction method, which allows for high-quality mesh and develops a new training process for applying a regularization provided by classical multi-view stereo methods. Moreover, we apply a differentiable rendering to fine-tune incomplete texture maps and generate textures which are perceptually closer to the original scene. Our pipeline can be applied to any common objects in the real world without the need for either in-the-lab environments or accurate mask images. We demonstrate results of captured objects with complex shapes and validate our method numerically against existing 3D reconstruction and texture mapping methods.", + "bbox": [ + 75, + 324, + 473, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering the 3D geometry of objects and scenes is a longstanding challenge in computer vision and is essential to a broad range of applications. Depth sensing technologies range from highly specialized and expensive turn-table 3D scanners and structured-light scanners to commodity depth sensors. More recently, advances in mobile devices have developed a new method for 3D capture of real-world environments with high-resolution imaging and miniaturized LiDAR. Specifically, the modern smartphone such as iPhone 13 Pro are equipped with RGB camera, accelerometer, gyroscope, magnetometer, and LiDAR scanner. These various sensors can provide high-resolution images, low-resolution depth from the LiDAR scanner, and associated camera poses offered by off-the-shelf visual-inertial odometry (VIO) systems such as ARKit [1] and ARCore [2].", + "bbox": [ + 75, + 657, + 468, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Today's smartphones offer low-resolution depth maps", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/415abffe65e52adcfca8c622fe9a069aa2e7915b859afb4c130414484fed2770.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 506, + 291, + 602, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d87e713b938a2e0180aca13b15bb5108c6d26f186bb904b0e6e27811456f102a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 602, + 292, + 686, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eae39edf8883c9c41473a42704556ab8d563187a9117edaadc148e5bef3a0f5e.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 730, + 297, + 803, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64932692b2dde51a95e569123e6968423e566b687d6ac549099e26b4bf8fdc6a.jpg", + "image_caption": [ + "(d)", + "Figure 1. Example reconstruction results collected from a smartphone in the wild. (a) Data acquisition setup. (b) Images captured from a smartphone. (c) A reconstructed mesh. (d) A novel view of textured mesh. Our proposed method can reconstruct the high-quality geometric mesh with a visually realistic texture." + ], + "image_footnote": [], + "bbox": [ + 807, + 297, + 883, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[5] and valid poses. However, depth maps are very noisy and suffer from the limited range of depth sensors. Although this depth sensor can build a simple 3D structure such as a wall or floor, it cannot reconstruct objects with complex and varying shapes. Thus, the RGBD scanning methods [9, 23, 42, 57] are not suitable for these objects. Instead of the depth sensor, multi-view stereo (MVS) algorithms [13, 47, 62] reconstruct high-quality 3D geometry by matching feature correspondences across different RGB images and optimizing photometric consistency. While this pipeline is very robust in real-world environments, it misses the surface of low-textured areas [62]. Additionally, the resulting mesh generated by post-processing like Poisson reconstruction [27] heavily depends on the quality of matching, and the accumulated errors in correspondence matching often cause severe artifacts and missing surfaces. Because of the cumulative error from the above pipeline, the texture mapping process [53, 69] leads to undesirable results. Reconstructing high-fidelity texture and 3D geometry of real-world 3D objects remains an open challenge.", + "bbox": [ + 496, + 506, + 892, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Main Results: In this paper, we present a practical method to capture a high-quality textured mesh of a 3D object in the wild, shown in Fig. 1. We first develop a smartphone app based on ARKit [1] to collect images, LiDAR depths, and poses. Although the smartphone provides quite valid pose estimates, acquiring fine detail geometry and realis-", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16674", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8eb00b627d49c096dc461da7311abe5d33a0aff3fa1fbc9de7de57ab66ebde8.jpg", + "image_caption": [ + "Figure 2. Example objects collected by a smartphone in the wild" + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 462, + 157 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tic texture requires highly accurate camera poses. Thus, we present an RGBD-aided Structure from Motion (SfM) which combines the advantages of both VIO [12] and SfM [46]. Since ARKit based on VIO is robust to the degradation of visual information such as low-textured surface, we perform incremental triangulation with initial poses obtained from ARKit. We also propose a depth filtering method to handle noisy depth from the smartphone. These filtered depth points are exploited as an additional depth factor for bundle adjustment. Our RGBD-aided SfM can estimate highly precise camera poses due to the good initial poses and additional constraints from the filtered depth.", + "bbox": [ + 75, + 186, + 467, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our 3D geometry reconstruction process adopts a neural implicit representation [50, 55] for surface reconstruction with volumetric rendering. We observe that the neural implicit representation can perform more complete and smoother reconstruction than the classical methods which are known to be robust for 3D reconstruction in the wild. Furthermore, we introduce a new training method for neural implicit representations. In the early stage of training, we propose a regularization method that leverages the prior information from the classical MVS method. After obtaining the decent shape, we avoid using the prior information and generate sparse voxel octree to enable efficient sampling for training. Our training method can improve the performance of the neural implicit representations. Consequently, we show the generalization capabilities of our surface reconstruction method in real world objects collected by the smartphone.", + "bbox": [ + 75, + 368, + 467, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a mesh extracted from the trained neural implicit representation, we can run classical texture mapping algorithms [53, 69] to generate texture maps that often exhibit blurring artifacts and color misalignment. We propose applying differential rendering [31] to fine-tune these texture maps obtained from classical texture mapping via a photometric loss. Compared to classical 3D reconstruction [23, 47, 62] and texture mapping [53, 69] approaches, our method shows a better ability to reconstruct the 3D model and produce realistic textures. We evaluate our approach on the data collected by our smartphone application. The main contributions of this paper are summarized as follows:", + "bbox": [ + 75, + 625, + 467, + 805 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a unified framework to reconstruct the textured mesh using a smartphone.", + "- We propose a depth filtering scheme for noisy depths and refine initial poses from ARKit by using bundle adjustment with depth factor." + ], + "bbox": [ + 94, + 815, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our pipeline builds on classical 3D reconstruction and texture mapping. We leverage a neural geometry representation to enable surface reconstruction of complex shapes and a differentiable rendering to generate high-fidelity texture maps.", + "bbox": [ + 519, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 180, + 640, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classical 3D Reconstruction Traditional Structure from Motion (SfM) systems [46, 59] have been popular as a vision-based 3D reconstruction due to their robustness to various scenarios [48, 49, 58]. The SfM systems extract and match features [36] of adjacent images of the same scene and generate the camera poses and sparse 3D points via triangulation [20]. Then, Multi-view Stereo (MVS) algorithms [13, 52] compute the dense depth maps of each scene with multiple calibrated images. PatchMatch-based MVS [14, 47, 61, 62] methods are popular because they are proper for depth and normal optimization. After that, surface reconstruction methods [26, 27, 30] can be applied to reconstruct meshes from point clouds. However, all these methods often fail to reconstruct the 3D geometry of low-textured areas. RGB-D Reconstruction techniques fuse many overlapping depth maps into a single 3D model. Starting from the classical TSDF-fusion [8], several methods [9, 23, 42, 57] produce voxel-based TSDF with commodity-level depth cameras such as Kinect. However, these methods have difficulty handling outliers and thin geometry due to noisy depth measurements. Due to the noise and limited range of the LiDAR sensor in smartphone, RGB-D reconstruction methods fail to estimate the camera odometry.", + "bbox": [ + 496, + 205, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Reconstruction with differentiable rendering Recently, NeRF [38] and its variants [37, 39, 65, 68] have used volumetric implicit representations for rendering and learn $\\alpha$ -compositing of a radiance field along rays. These methods yield remarkable quality of novel view synthesis results and do not need ground truth masks. However, 3D geometry from volumetric representations is far from satisfactory due to their volumetric properties [68]. Some methods [41, 64, 67] reconstruct the surface by encoding the network as an implicit function such as a signed distance function (SDF) or occupancy function. Many hybrid approaches [4, 10, 43, 55, 63] unify surface and volumetric representations by regarding surface as defining volumes near the surface, enabling accurate surface reconstruction without masks. Their follow-up works [19, 34, 54, 67] import prior information from the learning-based network. We avoid exploiting prior hints obtained from learning-based models because they are expensive to acquire new training datasets and time-consuming to train other neural networks. Classical Texture Optimization The matching between texture and geometry is a significant step for building realistic 3D models. Lempitsky et al. [32] apply the pairwise Markov Random Field to choose an optimal view for each", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "face as texture. Waechter et al. [53] present a global color adjustment to mitigate the visual seams between face textures. Zhou et al. [69] optimize the camera poses using a color consistency measurement and geometric error via local image warping to obtain sharp texture. Bi et al. [6] employ a patch-based image synthesis to produce a texture per each face to optimize the camera pose and geometric error.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Texture Learning Recently, learning-based methods [17, 22, 25, 51] have been proposed to generate textures for a 3D model. Huang et al. [22] introduce adversarial learning for texture generation in RGB-D scans. Thies et al. [51] store neural textures for a known mesh with given UV mapping and decode them to color for novel views. Some works [7, 60] learn neural texture with neural volumetric rendering. Differentiable rendering methods [16, 21, 33, 40] usually focus on estimating surface radiometric properties from images. These methods can learn to predict not only texture but also geometry via image loss with ground truth masks. Unfortunately, differentiable rendering methods heavily rely on pixel-level accurate masks. Acquiring accurate ground-truth masks in the real-world environment is a difficult task [29]. Therefore, we apply the differentiable rendering to fine-tune texture maps with a fixed mesh.", + "bbox": [ + 75, + 196, + 470, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 76, + 450, + 207, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given images, low-resolution depths and corresponding camera poses from ARKit, our goal is to reconstruct both the texture and geometry of real-world 3D objects. In Fig. 4, our pipeline consists of three stages: RGBD-aided Structure from Motion (SfM), Geometry Reconstruction, and Texture Optimization. Since both depth maps and camera poses provided by ARKit are very noisy, we construct a SfM system to filter out depth maps and refine poses. We then formalize our task as training a neural model for zero-level set of a signed distance function (SDF) guided by classical 3D reconstruction. After training, we extract the mesh with Marching Cubes [35] and apply classical texture reconstruction [53] to generate initial texture maps. Next, we exploit differentiable rendering [31] to finetune the texture maps with the supervision of multi-view images.", + "bbox": [ + 75, + 476, + 468, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. RGBD-aided Structure from Motion", + "text_level": 1, + "bbox": [ + 76, + 710, + 393, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of this section is to refine initial poses from ARKit by performing a bundle adjustment with depth maps from the smartphone. Unlike simple structures like walls or floors, 3D reconstruction for capturing a high level of detail is required to refine inaccurate poses from ARKit.", + "bbox": [ + 75, + 734, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preprocessing We introduce two processes to obtain reliable depth estimates: single-view filtering and multi-view filtering. Single-view filtering utilizes a confidence map provided by ARKit, which has the same resolution as the depth map. Each pixel in the confidence map has 0, 1, or 2 with the value 0 as least confident and 2 serving as most", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ebf140c23abdec5ee01cfc3e5266ef65db80bbc5da8e39351c5ba4a0fc589bf1.jpg", + "image_caption": [ + "Figure 3. Overview of our pipeline. It is comprised of three steps: RGBD-aided Structure from Motion (Sec. 3.1), Geometry Reconstruction (Sec. 3.2), and Texture Optimization (Sec. 3.3). Our method builds on classical 3D reconstruction and texture mapping. To overcome the limitations of classical methods, we apply a neural geometry representation and neural texture fine-tuning to update 3D geometry and texture maps." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "confident with respect to depth-value accuracy. We choose depth pixels corresponding to 1 or 2 in the confidence map. A main concept of the multi-view filtering is to remove outliers in terms of the reprojection error. Given the reference image $I_{ref}$ with depth $D_{ref}$ , the source image $I_{src}$ with depth $D_{src}$ , the depth map $D_{src}'$ can be transformed from the source view to the reference view with accurate pixel-level correspondence. Then, the depth of pixel $p$ in the reference image will be considered as an inlier if the difference $\\delta_{depth}$ between $D_{src}$ and $D_{src}'$ is smaller than a small constant value $\\epsilon$ . Then, assuming a set of neighboring N source images, we compute the number of inliers which fulfills $\\delta_{depth} < \\epsilon$ for each source image, and keep depth pixels if the number of inliers is greater than a certain threshold. Finally, we apply both single-view and multi-view filtering to remove all depth outliers.", + "bbox": [ + 496, + 311, + 892, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose Refinement We consider camera poses obtained from ARKit as initial poses and exploit matching correspondences to perform incremental triangulation [46]. After combining filtered depth points from LiDAR and 3D points from incremental triangulation, we propose a depth factor and run bundle adjustment with the depth factor and reprojection error. Let $P_{i}$ denote the camera parameter of the $i$ -th camera and $X_{j}$ denote point parameters seen by the $i$ -th camera. We minimize the proposed joint objective function $E$ , which is formulated as follows:", + "bbox": [ + 496, + 554, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE = \\sum_ {i} \\sum_ {j} \\rho_ {i} \\left(\\left\\| \\phi \\left(P _ {i}, X _ {j}\\right) - x _ {i j} \\right\\| ^ {2}\\right) + \\frac {1}{\\eta} \\left| D _ {i} - z _ {i} \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 715, + 890, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi$ is the function that projects points to the image plane, and $\\rho$ denotes the Cauchy function as the robust loss function. $x_{ij}$ is the feature point corresponding to the 3D point $X_{j}$ seen by the $i$ -th camera. $z_{i}$ is the depth value of points projected from the 3D point $X_{j}$ and $D_{i}$ is the filtered depth of pixels corresponding to $z_{i}$ . We define $\\eta$ as sensor measurement noise and use it to down-weight noisy depth. We notice that the depth factor has two benefits. First, additional information from the depth sensor is a useful con", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16676", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strain for accurate pose estimation and reliable triangulation. Furthermore, this depth factor guides the scale of the SfM to follow the metric scale of the depth sensor.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Geometry Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 143, + 310, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Geometry Representations Instead of classical 3D reconstruction, we adopt NeuS [55], which combines the advantages of both volume and surface rendering. In other words, it can extract accurate surfaces even without pixel-level masks. Our scanned object is represented by two Multi-layer Perceptrons (MLPs): geometry network $f_{sdf}$ , which maps a 3D point $x \\in \\mathbb{R}^3$ to a signed distance, and color network $f_{color}$ which converts a 3D point $x$ and a viewing direction $v \\in \\mathbb{R}^3$ to a color $c$ . The 3D geometry of an object is represented implicitly as an SDF with zero-level set $\\{x \\in \\mathbb{R}^3 | f_{sdf}(x) = 0\\}$ . These two parameters of geometry and color are learned by the volume rendering technique. For a ray $\\{x(t) = o + tv | t \\geq 0\\}$ with camera center $o$ and viewing direction $d$ , the color is rendered with $n$ sampled points along the ray as follows:", + "bbox": [ + 75, + 167, + 470, + 395 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} = \\sum_ {i = 1} ^ {n} T _ {i} \\alpha_ {i} c (x (t), v), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 402, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})$ is the accumulated transmittance, and $\\alpha_{i}$ is the discrete opacity. Here, $\\alpha_{i} = 1 - \\exp(-\\int_{t_{i}}^{t_{i+1}} \\rho(t) dt)$ , and $\\rho(t)$ is opaque density function borrowed from the definition in NeuS [55].", + "bbox": [ + 76, + 450, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Process RGBD-aided SfM already estimates both poses and sparse depths by matching feature correspondences. Given this prior information, we leverage an MVS algorithm [62] to estimate the depth and normal for all images with less cost. It is worth noting that we avoid using learning-based MVS algorithms because they are not robust to diverse real-world scenes and are hard to evaluate reliability. In this section, we modify the training process of NeuS to avoid losing fine details of complex shape. Our training process is divided into two stages. In the first stage, we train a neural radiance field with a prior depth and normal obtained from the MVS [62] algorithm to supervise the training process. Specifically, we sample $K$ pixels $p_k$ and their corresponding color values $I(p_k)$ in each iteration. Here, we propose a regularization method that relies on prior depth $D$ and normal $N$ from MVS algorithm. We unproject depth under the given pose to obtain 3D point $X(p_k)$ . Then, we use this 3D point and its corresponding normal $N(p_k)$ from MVS to regularize the geometry network by minimizing", + "bbox": [ + 75, + 515, + 468, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e g} = \\sum_ {k} w _ {d} \\left| f _ {s d f} \\left(X \\left(p _ {k}\\right)\\right) \\right| + \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 825, + 468, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {n} (1 - \\langle \\nabla_ {p} f _ {s d f} (X (p _ {k})), N (p _ {k}) \\rangle),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 859, + 424, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{d}$ and $w_{n}$ are weights for each term. This regular-", + "bbox": [ + 81, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa4d083e001c0a421528df5dc4081def1d09d3bf1740c46cbde594ef330ced3c.jpg", + "image_caption": [ + "Figure 4. Overview of Geometry Reconstruction (Section 3.2) and Texture Optimization (Section 3.3). Our goal in geometry reconstruction is to train geometry network $f_{sdf}$ and color network $f_{color}$ via volumetric rendering with two training steps. In step 1, we leverage prior geometric information from the MVS algorithm [62]. Then, our method utilizes a sparse voxel octree to efficiently sample points near the surface. In texture optimization, we first utilize a classical texture reconstruction to generate texture maps for a given mesh. Then, our method applies a differentiable rendering to fine-tune texture maps." + ], + "image_footnote": [], + "bbox": [ + 513, + 94, + 883, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ization will guide the SDF to learn explicit supervision from the MVS algorithm. Although this regularization is helpful for learning prior geometry, it is susceptible to the noisy results derived from the MVS algorithm which tends to lose fine details and texture-less regions. Thus, we will only use this regularization in the first stage. The color of each pixel can be computed by E.q. 4. The color loss $L_{c}$ is defined as", + "bbox": [ + 496, + 383, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c} = \\sum_ {k} | \\hat {C} (p _ {k}) - I (p _ {k}) |. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 500, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Eikonal loss [18] to regularize the SDF is formulated as", + "bbox": [ + 498, + 541, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e i k} = \\sum_ {k, i} \\left(\\| \\nabla f _ {s d f} \\left(p _ {k, i}\\right) \\| _ {2} - 1\\right) ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 568, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall loss function $L$ in the first stage is defined as", + "bbox": [ + 500, + 606, + 874, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {c} + L _ {e i k} + L _ {r e g}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 632, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second stage, inspired by [50], we adopt a sparse voxel octree to guide the sampling process and capture fine details. After the first stage, we can extract a triangular mesh from the SDF via the Marching Cube [35] and define a sparse voxel volume based on this mesh. Then, based on the sparse voxel volume, we can avoid redundant point samples outside the sparse voxel and focus on point samples near the surface samples. Here, our total loss is $L = L_{c} + L_{eik}$ which is the same as E.q. 6 except for $L_{reg}$ . To capture fine details, we will not use $L_{reg}$ in E.q. 6.", + "bbox": [ + 496, + 659, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mesh Simplification A highly detailed 3D mesh from the SDF network with marching cube algorithms [35] leads to enormous memory requirements. To obtain a lightweight mesh, we explore popular quadric mesh simplification [15]. Although the mesh simplification generated small holes, the lightweight mesh enables us to reduce the computation", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "complexity of the texture mapping process. We set the decimation ratio to $0.4\\%$ .", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Texture Optimization", + "text_level": 1, + "bbox": [ + 76, + 128, + 279, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classical Texture Reconstruction For a triangular mesh $M$ provided by the previous section, we first utilize Waechter et al. [53] with poses and images across different views to construct a texture map $T_{orig}$ . Compared to the color network from Section 3.2, this classical method [53] can acquire visually realistic texture images. However, it cannot avoid texture bleeding on the boundary of different views and blurring artifacts.", + "bbox": [ + 75, + 152, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Texture Fine-tuning To tackle blurring and texture bleeding artifacts, we propose a fine-tuning step by optimizing the texture map. Here, we parameterize a texture map $T$ obtained from classical texture reconstruction. Given a 3D Mesh $M$ with the 2D texture map $T$ and multi-view images $I_{i}$ with the corresponding camera poses $P_{i}$ , we leverage a differentiable renderer $R$ to generate an image $\\hat{I}_i$ and depth $\\hat{D}_i$ , i.e. $\\hat{I}_i, \\hat{D}_i = R(M, P_i)$ . A differentiable renderer $R$ [31] performs mesh rasterization and then texture sampling to render an image $\\hat{I}_i$ . In other words, we project texture onto the image plane and compare it to the corresponding view of the image. Furthermore, we compare the rendered image with a different view image by minimizing multi-view photometric loss. We use the combination of the L1 and SSIM [56] as the photometric loss $L_{photo}$ in image space between the rendered image $\\hat{I}_i$ and the reference image $I_{i}$ as follows:", + "bbox": [ + 75, + 273, + 470, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p h o t o}} = \\sum_ {i} (1 - \\alpha) \\| \\hat {I} _ {i} - I _ {i} \\| + \\frac {\\alpha}{2} (1 - S S I M (\\hat {I} _ {i}, I _ {i})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 541, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is set to 0.85. Additionally, the differentiable render estimates depth $\\hat{D}_i$ by interpolating $z$ coordinates of each vertex. With the estimated depth $\\hat{D}_i$ and intrinsic parameter $K$ , we can generate a synthesized frame $I_i'$ by reprojecting adjacent frames $I_{i'}$ to the current frame $I_i$ with the camera intrinsic matrix $K$ , the estimated depth $\\hat{D}_i$ , and the relative pose $P_{i\\rightarrow i'}$ . The multi-view photometric loss $L_{multi}$ similar to Eq. 7 is formulated as:", + "bbox": [ + 76, + 585, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {\\prime} (p) = I _ {i ^ {\\prime}} \\langle \\pi (K P _ {i \\rightarrow i ^ {\\prime}} \\hat {D} _ {i} (p) K ^ {- 1} \\tilde {p}) \\rangle ,\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 718, + 369, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m u l t i} = (1 - \\alpha) \\| I _ {i} ^ {\\prime} - I _ {i} \\| + \\frac {\\alpha}{2} (1 - S S I M \\left(I _ {i} ^ {\\prime}, I _ {i}\\right)), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 738, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{p}$ is the homogeneous coordinate of $p$ , $\\pi$ means projection from homogeneous to image coordinates, and $\\langle \\cdot \\rangle$ indicates the bilinear sampling function. We define our total loss $L$ to be", + "bbox": [ + 76, + 787, + 468, + 847 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {p h o t o}} + L _ {\\text {m u l t i}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 849, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the differentiable rendering enables the loss gradients to back-propagate to update the textures parameters.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our method does not require ground truth masks, we avoid considering optimization for the geometry. We observe that joint optimization with geometry and texture fails to converge without the ground truth mask.", + "bbox": [ + 498, + 90, + 890, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 165, + 633, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 500, + 190, + 692, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Collection: Our smartphone application, running on an iPhone 13 Pro using ARKit 6, can record various sensor outputs: synchronized images, poses, confidence maps, dense depth maps from the LiDAR scanner, and IMU sensor information. We built two types of datasets: ARKit-video and AR-capture. ARKit-video collects long video sequences similar to datasets constructed from RGBD-SLAM [70]. The second type is collected by our smartphone application based on Object Capture1. AR-capture is a multiview dataset of objects consisting of high-quality 4K images with depth and gravity data. The ARKit-video collects thousands of frames sampled from the video with resolution of $1280 \\times 720$ . In contrast, AR-capture mostly contains less than fifty images at $4032 \\times 3096$ pixels. We collect 11 objects: robot arm 1, robot arm 2, plant, tree, sofa, bike, recliner chair, cafe stand, delivery robot, office chair, and camera stand. We randomly select $70 \\%$ as the training images and utilize the $30 \\%$ as test images for evaluating the novel view synthesis.", + "bbox": [ + 496, + 213, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DTU Dataset: The DTU dataset [24] is a MVS dataset that provides reference point clouds acquired with laser sensor in the controlled lab. Since AR-capture and ARKit-video do not include 3D point clouds acquired with specialized hardware, we use the DTU dataset to evaluate the performance of geometry reconstruction. We follow the evaluation of NeuS [55] and exploit the chamfer distance as metric.", + "bbox": [ + 496, + 500, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: We build the RGBD-aided Structure from Motion based on COLMAP [46], an existing SfM tool. Instead of SIFT [36], our system uses learning-based features [11, 44, 45] for feature extraction and matching to avoid failure in a featureless scene. ARKit offers an interface for accessing LiDAR data as a depth image. Following the work [55], the SDF network $f_{sdf}$ is implemented as an 8-layer MLP a hidden dimension of 256, and geometric initialization [3] for the network parameters. We construct the color network $f_{color}$ as MLPs with 4-layer MLP with hidden dimension of 256. We optimize networks in geometry reconstruction with Adam [28] and an initial learning rate of 5e-4. We train the first step for 60K iterations and the second step for 140K iterations. In terms of objects with complex shape, it takes 240K iterations for second step. In Texture Optimization, we utilize the public github code for classical method [53]. Then, for texture fine-tuning, the Adam optimizer is used with the learning rate for the tex", + "bbox": [ + 496, + 606, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://developer.apple.com/videos/play/wwwdc2021/10076/", + "bbox": [ + 517, + 886, + 836, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a43ebff7a14aacbfdb034db53c3b4d42633b3ba52185a5ec5bcdec724b02dd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 80, + 194, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ad56b6af3de92b010d7992935bb044e5daddeed01fe4da5732f849fac924f3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 87, + 263, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f774c18e07150992da3856495a45b6501a14dc2394c5380b4011c7269070aaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 82, + 387, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb34bd42d29220a5516e8de81e61709ebe4e124d0123d23fdff8c1ab0a473028.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 87, + 455, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8c4e27ddd1b0d3f7f8b3080cf0298ec5f7ca934f03afbc73b47aef5c5233af2.jpg", + "image_caption": [ + "a)" + ], + "image_footnote": [], + "bbox": [ + 84, + 167, + 186, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/009c1c16aeb5192a0b685495636aa5c5e9679853ca7d276c1c6d20cc0670e2ec.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 187, + 191, + 274, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5adda5009c718b94063d446a781c2dd8ca18b8cd202f3b0722f55f0462b920e9.jpg", + "image_caption": [ + "(c)", + "Figure 5. We show the benefits of RGBD-aided Structure from Motion. (a) illustrates the 3D point clouds obtained from initial depths with initial poses. In (c), we accumulate the 3D point clouds (by using filtered depths and refined poses) from RGBD-aided SfM. For a detailed description, we visualize the zoomed-in blue and red regions using filtered depth. Initial depth maps and poses from ARkit are input to the geometry reconstruction in Section 3.2 and its results are shown in (b). The results of our pipeline in (d) lead to higher quality reconstructions than (b) which does not apply RGBD-aided SfM." + ], + "image_footnote": [], + "bbox": [ + 277, + 174, + 379, + 261 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3df9ee72ff27b523475683c46fba4da9fa3b5113a6f9c341425abda80d87daa1.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 380, + 193, + 464, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ture parameter. This step takes 3000 epochs to fine-tune the texture parameter.", + "bbox": [ + 75, + 429, + 468, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 469, + 277, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 5, we visualize the effectiveness of RGBD-aided Structure from Motion in Section 3.1 on our dataset. Figure 5 (a) represents the 3D point clouds by initial depth maps and poses obtained from ARKit. Unfortunately, initial sensor data is too noisy to perform an accurate 3D reconstruction. After multi-view depth filtering as preprocessing, zoomed-in regions in (a) show the point clouds by reprojecting filtered depth maps with initial poses. Compared to initial data, multi-view depth filtering can handle noisy depth measurements. Figure 5 (c) provides the results of preprocessing and pose refinement. Pose refinement enables us to globally align 3D reconstruction without noticeable camera drift and with better local quality. For qualitative comparison, we take as input initial sensor data from ARKit and perform geometry reconstruction in Section 3.2 in Fig. 5 (b). Compared to our experimental results in Fig. 5 (d), it is difficult to achieve comparable quality and better reconstruction. Consequently, our proposed RGBD-aided SfM significantly improves geometry reconstruction performance.", + "bbox": [ + 75, + 492, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 qualitatively compares our method to VolSDF [63], TSDF-fusion [23], and ACMP [62]. We apply screened Poisson surface reconstruction [27] to generate a mesh from point clouds of ACMP. Among all four algorithms, we apply the same pose estimation algorithm from Section 3.1 for fair comparison. TSDF-fusion, which requires sufficient scene overlap, exploits more frames than our method and ACMP. Due to the low-textured regions,", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ffa55901bc8c1c33a5eb1ae805465561e29958d53d4ed05c2acd5a9da5b1c85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 80, + 591, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8709bba720f3bdecde117196720a3ebd42f76cf3ba0abee0054493f07273f846.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 80, + 656, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/563fe1f9863c6892e2a1056fae096294f805988b566e6b4e03b36d399616994e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 80, + 720, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/971a1e1073470c46a99f1b2bf5d7ec10d59a9ce7aed7fd95ce4f54b86049dab9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 80, + 784, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ba4a45691db645cb17ed7621ffa844886a6cc7c53fd6767e1d0cbd1162f7844.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 80, + 854, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f8ca75e22199423afbeeef829e7bafd8f000fdde66ce952bdf4c7dfdb2d859d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 152, + 594, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/45ca2ed330a1ffce7b5b955dc9c4de995b0f4cd1c306ecf117057873c655b03c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 154, + 660, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbac8ffdfc7283f16ca8e2d9707d5d5f2ac7702b0cf0b8f3cadd5aaef49ed51d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 154, + 725, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87cb4285c683ca52a3d797fb40dd99f3e8ccfea15311de885cf4a8a4d8f5c6fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 154, + 792, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15c44946e0a79b55493008634dab1f6f37c8578258ce2d27e6cbcbe5df8be44e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 154, + 859, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e32b689aa50550d61abeb5195d63483fb27a4ac587f90810b5d21c2c4763c88.jpg", + "image_caption": [ + "Scene" + ], + "image_footnote": [], + "bbox": [ + 535, + 220, + 594, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df29b3ed3de6da4219b148a6f97b48bd06a30eb8c53d7751107860ce4ae080a7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 601, + 220, + 661, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c559e72ff6957848b1e2bd828ab434fa464b6ffd9b8ae15662f5c38f74abfc42.jpg", + "image_caption": [ + "VolSDF" + ], + "image_footnote": [], + "bbox": [ + 665, + 220, + 725, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/edd37be9f214f70dd8f23839e1226e8ad31466be99db34b8d6d465f3af958481.jpg", + "image_caption": [ + "TSDF-Fusion", + "Figure 6. Reconstruction of our method, VolSDF [63], TSDF-Fusion [23], and a mesh from point clouds of ACMP [62] processed with screened Poisson surface reconstruction [27]. Note that the neural surface reconstruction is more perceptually convincing in the ARKit-video dataset." + ], + "image_footnote": [], + "bbox": [ + 730, + 220, + 785, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/616ec985860f07769c11792fb9fcc5eebfc904c7354f30847b91985a00070f5e.jpg", + "image_caption": [ + "ACMP" + ], + "image_footnote": [], + "bbox": [ + 797, + 220, + 859, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the surface of ACMP is very noisy. TSDF-fusion utilizes depth maps obtained from low-resolution LiDAR. Here, since this low-resolution depth sensor is very noisy, we only use filtered depth maps after preprocessing in Section 3.1. However, this model still cannot reconstruct either thin or challenging structures accurately (e.g., a chair armrest). Both implicit neural surface reconstruction approaches show comparable performance, and both are superior to classical 3D reconstruction. We observe that the difference between neural geometric representations (Ours and VolSDF) is negligible unlike pose quality. It is worth noting that using accurate poses as input significantly improves the 3D reconstruction quality as shown in Fig. 5.", + "bbox": [ + 496, + 460, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For qualitative comparison of textured meshes, we compare our proposed method with four different methods: 1) ACMP [62] is a state-of-the-art multi-view stereo algorithm: 2) Color Map Optimization (CMO) [69] is a texture mapping method for RGB-D scanning: 3) Waechter et al. [53] is a large-scale texture reconstruction based on global color adjustment, 4) NeuS [55] is the recent neural surface reconstruction based on volumetric rendering. We reconstruct a 3D mesh from the point cloud of ACMP with Screened Poisson Surface reconstruction [27]. In the case of CMO, since explicit depth maps from low-resolution LiDAR are too noisy, we utilize depth maps rendered by a triangular mesh from our geometry reconstruction step for a fair comparison. With these rendered depth maps and corresponding images, we run a color map optimization for mapping color images onto our reconstructed mesh. To train", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4a44161255f1a7f423d620e1645d81a203a3af34e19c6467ae74e872b1a1229.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison between (a) Ours, (b) Waechter et al. [53], (c) CMO [69], (d) ACMP [62], and (e) NeuS [55]. Compared to textured mesh (3-6 columns) obtained from different algorithms, the result of our method is perceptually closer to the original scene. Additional qualitative results are added in the supplementary material due to the limited space." + ], + "image_footnote": [], + "bbox": [ + 200, + 103, + 772, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NeuS, we follow a training process similar to that in Sec. 3.2. We then extract a mesh from the SDF network in NeuS and apply the Blender Smart UV Project tool to obtain a pervertex UV mapping. To fill a texture for this mesh, points densely sampled from triangle mesh with interpolated UV coordinates are fed to the color network, and eventually we can obtain texture maps for the given mesh.", + "bbox": [ + 75, + 729, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 7, we show the reconstructed texture for novel views. The visual results of our approach are visually sharper and perceptually closer to the ground truth scene. Results of NeuS and CMO exhibit blurring artifacts and", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "color misalignment. The reconstructed mesh from ACMP is very noisy due to incorrect geometry. Here, Waechter et al. (Fig. 7 (b)) is the result of classical texture reconstruction in Section 3.3. This method is more photorealistic and sharper than NeuS, CMO, and ACMP. Then, after texture fine-tuning, our method avoids the seams and texture misalignment often seen with Waechter et al. Table 1 reports the quantitative results on two datasets. Our method outperforms other texture optimization methods on all evaluation metrics. It indicates that our method satisfies both the high fidelity of the texture and its conformity to the ground truth.", + "bbox": [ + 496, + 729, + 890, + 896 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16680", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2456b5439aa3f2575f6df9969c4da4bca2a5f0d159dfaaf5b456a380d4a4041a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryDataNeuS [55]CMO [69]Waechter et al. [53]Ours
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
robot arm 1Ac20.2920.8890.08318.4730.8910.08319.9090.8720.06020.0490.8710.057
robot arm 2Ac20.7410.8820.08621.1070.8830.09220.1900.8620.06720.6520.8580.062
plantAc16.7480.6420.29518.4620.6270.31617.9210.6350.18619.9700.6910.146
treeAc18.4090.8690.11718.4480.8700.12717.5180.8580.09618.3310.8580.102
sofaAc19.9100.8720.13921.1700.8400.13619.7590.8290.11321.6080.8940.109
bikeAc20.9710.7780.18321.3110.7780.20619.8220.7400.08819.9430.7370.078
MeanAc19.5120.8220.15119.8250.8150.16019.1860.7990.10120.0950.8180.092
recliner chairAv24.9200.8930.09924.7610.8760.11223.6920.8640.05425.0450.8960.056
cafe standAv20.4980.7790.17322.9340.8280.15417.9150.7980.11222.6340.8180.095
delivery robotAv22.4120.8970.12623.5490.9080.12825.2460.9290.07323.4510.9060.086
office chairAv20.7220.8920.09621.1200.8990.10320.3150.8650.09420.9820.8920.083
camera standAv21.9010.9140.06921.8860.9070.07421.1210.8990.05922.2080.9190.060
MeanAv22.0920.8750.11322.8500.8840.11521.6580.8710.07922.8680.8860.076
", + "bbox": [ + 156, + 88, + 805, + 256 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/970cd86261685b2f1f557f0767f8a93777fc86822a21cb1c77412d9346814464.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on our dataset. Ac and Av denote the AR-catpure dataset and ARKit-video dataset, respectively. We improve the LPIPS metric by $8.9\\%$ in AR-capture and by $3.8\\%$ in ARKit-video." + ], + "table_footnote": [], + "table_body": "
Methodscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122Mean
COLMAP [47] with trim=70.450.910.370.370.901.000.541.221.080.640.480.590.320.430.450.65
UNISURF [43]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
NeuS [55]1.371.210.730.401.200.700.721.011.160.820.661.690.390.610.510.87
VolSDF [63]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Ours without 2nd step1.041.610.670.640.880.700.621.141.100.890.661.020.500.580.510.84
Ours0.961.370.670.600.880.700.451.141.100.840.661.110.500.580.500.80
", + "bbox": [ + 156, + 294, + 807, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison on DTU [24]. Chamfer distance is employed as the evaluation metric. COLMAP is superior to other neural geometry reconstructions. Our training process can improve the NeuS by $8\\%$ .", + "bbox": [ + 75, + 367, + 893, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/694af125746d1c40e1531a1b718526996fbd3a4d6e0c96dd7283463f0d615957.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons of texture optimization algorithms based on differentiable rendering: (a) our method, (b) Lin et al. [33], and (c) Munkberg et al. [40]." + ], + "image_footnote": [], + "bbox": [ + 106, + 407, + 163, + 520 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3ed6c897695308feeb73a9babae4f3932d6115695cbf1416298d82bee8d18da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 407, + 243, + 521 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6020174dca24b6bcfcc06acb4cedee350156da2f48a96132752e1d597ab4aee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 407, + 330, + 521 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d38370e4498bf8769ef5e4dbda7025adcaf01f519a17216c7499168fd12f270a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 407, + 437, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 8, we show the results of other texture optimization methods [33,40] based on differentiable rendering [31]. Both methods aim to learn geometry, textures, and lighting parameters. However, they heavily rely on ground truth masks which require additional annotation costs. Without ground truth masks, we note that joint optimization over geometry and texture is unstable and often has negative effects on geometry. For a fair comparison, we therefore provide the fixed mesh topology and train only texture and lighting parameters through differentiable rendering [31]. Since we capture objects in the wild under unknown environmental lighting conditions, it is difficult to disentangle materials and lighting via differentiable rendering. Neither method [33, 40] is able to generate realistic texture maps compared to either our method or classical texture reconstruction approaches.", + "bbox": [ + 75, + 582, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental Results on DTU dataset: Unlike our dataset, DTU [24] is captured by specialized hardware in the controlled lab environment. Thus, this dataset provides the ground truth poses and the RGBD-aided SfM is not necessary. As shown in Table 2, the surfaces obtained by", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "COLMAP [47] with trimming value 7 achieve high-quality reconstruction. Our method exploits MVS [47] to supervise the training NeuS and outperforms NeuS. For the ablation study, when we only train our network with the first step (ours without 2nd step), its result is not better than our method in some scans such as scan37. In our experiment, sparse voxel octree sampling does not have a significant effect on geometry performance. Since our training process can be applied to any other 3D neural surface reconstruction methods. In the future, we will apply this training process to other recent 3D neural surface reconstruction algorithms [10, 63, 66], which shows better results on DTU.", + "bbox": [ + 496, + 410, + 893, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Limitations and Future Work", + "text_level": 1, + "bbox": [ + 498, + 613, + 877, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have developed a practical framework to reconstruct high-quality textured meshes of 3D objects by using a smartphone. Given initial poses from ARKit, multi-view images, and low-resolution depth maps, our approach can refine initial poses and depth maps from ARKit. We adopt neural geometry representation [55] for surface reconstruction and improve its performance by modifying the training process. Finally, a differentiable rendering is designed for fine-tuning the texture map acquired from the classical texture mapping method. In the future, our goal is to reduce the training time which can be comparable with classical methods. Unfortunately, most neural geometry reconstruction [10, 55, 63, 66] requires a long training time. We would like to reduce the training time in our framework for practical purposes.", + "bbox": [ + 496, + 641, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was partly supported by Army Cooperative Agreement W911NF2120076.", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Apple arkit. https://developer.apple.com/documentation/arkit/. Accessed: 2022-11-09. 1", + "[2] Google arcore. https://developers.google.com/ar. Accessed: 2022-11-09. 1", + "[3] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5", + "[4] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 2", + "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1", + "[6] Sai Bi, Nima Khademi Kalantari, and Ravi Ramamoorthi. Patch-based optimization for image-based texture mapping. ACM Trans. Graph., 36(4):106-1, 2017. 3", + "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision (ECCV), 2022. 3", + "[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 2", + "[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 1, 2", + "[10] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 8", + "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR Deep Learning for Visual SLAM Workshop, 2018. 5", + "[12] Alex Flint, Oleg Naroditsky, Christopher P Broaddus, Andriy Grygorenko, Stergios Roumeliotis, and Oriel Bergig. Visual-based inertial navigation, Dec. 11 2018. US Patent 10,152,795. 2", + "[13] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2", + "[14] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. June 2015. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 4", + "[16] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 3", + "[17] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV, page 88-104, Berlin, Heidelberg, 2020. Springer-Verlag. 3", + "[18] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4", + "[19] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2", + "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2", + "[21] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7498-7507, 2020. 3", + "[22] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 3", + "[23] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 1, 2, 6", + "[24] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 5, 8", + "[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3", + "[26] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2", + "[27] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 1, 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16682", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[29] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3", + "[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2", + "[31] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5, 8", + "[32] Victor Lempitsky and Denis Ivanov. Seamless mosaicing of image-based texture maps. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-6. IEEE, 2007. 2", + "[33] Lixiang Lin, Jianke Zhu, and Yisu Zhang. Multiview textured mesh recovery by differentiable rendering. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 3, 8", + "[34] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 2", + "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 3, 4", + "[36] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. 2, 5", + "[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2", + "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2", + "[40] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. arXiv:2111.12503, 2021. 3, 8", + "[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 1, 2", + "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 8", + "[44] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019. 5", + "[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5", + "[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 2, 3, 5", + "[47] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 8", + "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2", + "[49] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80(2):189-210, 2008. 2", + "[50] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 4", + "[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Trans. Graph., 38(4), jul 2019. 3", + "[52] George Vogiatzis, Philip HS Torr, and Roberto Cipolla. Multi-view stereo via volumetric graph-cuts. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 391-398. IEEE, 2005. 2", + "[53] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 1, 2, 3, 5, 6, 7, 8", + "[54] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022. 2", + "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 4, 5, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16683", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[57] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2", + "[58] Aji Resindra Widya, Yusuke Monno, Masatoshi Okutomi, Sho Suzuki, Takuji Gotoda, and Kenji Miki. Stomach 3d reconstruction based on virtual chromoendoscopic image generation. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1848-1852. IEEE, 2020. 2", + "[59] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2", + "[60] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7119-7128, June 2021. 3", + "[61] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[62] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 1, 2, 4, 6, 7", + "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 6, 8", + "[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2", + "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2", + "[66] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 8", + "[67] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2", + "[68] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2", + "[69] Qian-Yi Zhou and Vladlen Koltun. Color map optimization for 3d reconstruction with consumer depth cameras. ACM" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transactions on Graphics (ToG), 33(4):1-10, 2014. 1, 2, 3, 6, 7, 8", + "[70] Michael Zollhöfer, Patrick Stotko, Andreas Görtlitz, Christian Theobalt, Matthias Nießner, Reinhard Klein, and Andreas Kolb. State of the art on 3d reconstruction with rgb-d cameras. In Computer graphics forum, volume 37, pages 625-652. Wiley Online Library, 2018. 5" + ], + "bbox": [ + 501, + 90, + 893, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16684", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_model.json b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc511f2320a52366b029e23bcdeac758fb5cc66b --- /dev/null +++ b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "TMO: Textured Mesh Acquisition of Objects with a Mobile Device by using Differentiable Rendering" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.203, + 0.831, + 0.258 + ], + "angle": 0, + "content": "Jaehoon Choi \\(^{1,2}\\) Dongki Jung \\(^{1}\\) Taejae Lee \\(^{1}\\) Sangwook Kim \\(^{1}\\) Youngdong Jung \\(^{1}\\) Dinesh Manocha \\(^{2}\\) Donghwan Lee \\(^{1}\\) \\(^{1}\\)NAVER LABS \\(^{2}\\)University of Maryland" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.598 + ], + "angle": 0, + "content": "We present a new pipeline for acquiring a textured mesh in the wild with a single smartphone which offers access to images, depth maps, and valid poses. Our method first introduces an RGBD-aided structure from motion, which can yield filtered depth maps and refines camera poses guided by corresponding depth. Then, we adopt the neural implicit surface reconstruction method, which allows for high-quality mesh and develops a new training process for applying a regularization provided by classical multi-view stereo methods. Moreover, we apply a differentiable rendering to fine-tune incomplete texture maps and generate textures which are perceptually closer to the original scene. Our pipeline can be applied to any common objects in the real world without the need for either in-the-lab environments or accurate mask images. We demonstrate results of captured objects with complex shapes and validate our method numerically against existing 3D reconstruction and texture mapping methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Recovering the 3D geometry of objects and scenes is a longstanding challenge in computer vision and is essential to a broad range of applications. Depth sensing technologies range from highly specialized and expensive turn-table 3D scanners and structured-light scanners to commodity depth sensors. More recently, advances in mobile devices have developed a new method for 3D capture of real-world environments with high-resolution imaging and miniaturized LiDAR. Specifically, the modern smartphone such as iPhone 13 Pro are equipped with RGB camera, accelerometer, gyroscope, magnetometer, and LiDAR scanner. These various sensors can provide high-resolution images, low-resolution depth from the LiDAR scanner, and associated camera poses offered by off-the-shelf visual-inertial odometry (VIO) systems such as ARKit [1] and ARCore [2]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Today's smartphones offer low-resolution depth maps" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.292, + 0.603, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.409, + 0.564, + 0.42 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.293, + 0.687, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.409, + 0.673, + 0.42 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.298, + 0.805, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.409, + 0.777, + 0.42 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.299, + 0.884, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.409, + 0.854, + 0.42 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Figure 1. Example reconstruction results collected from a smartphone in the wild. (a) Data acquisition setup. (b) Images captured from a smartphone. (c) A reconstructed mesh. (d) A novel view of textured mesh. Our proposed method can reconstruct the high-quality geometric mesh with a visually realistic texture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.809 + ], + "angle": 0, + "content": "[5] and valid poses. However, depth maps are very noisy and suffer from the limited range of depth sensors. Although this depth sensor can build a simple 3D structure such as a wall or floor, it cannot reconstruct objects with complex and varying shapes. Thus, the RGBD scanning methods [9, 23, 42, 57] are not suitable for these objects. Instead of the depth sensor, multi-view stereo (MVS) algorithms [13, 47, 62] reconstruct high-quality 3D geometry by matching feature correspondences across different RGB images and optimizing photometric consistency. While this pipeline is very robust in real-world environments, it misses the surface of low-textured areas [62]. Additionally, the resulting mesh generated by post-processing like Poisson reconstruction [27] heavily depends on the quality of matching, and the accumulated errors in correspondence matching often cause severe artifacts and missing surfaces. Because of the cumulative error from the above pipeline, the texture mapping process [53, 69] leads to undesirable results. Reconstructing high-fidelity texture and 3D geometry of real-world 3D objects remains an open challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Main Results: In this paper, we present a practical method to capture a high-quality textured mesh of a 3D object in the wild, shown in Fig. 1. We first develop a smartphone app based on ARKit [1] to collect images, LiDAR depths, and poses. Although the smartphone provides quite valid pose estimates, acquiring fine detail geometry and realis-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16674" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.464, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.162, + 0.466, + 0.175 + ], + "angle": 0, + "content": "Figure 2. Example objects collected by a smartphone in the wild" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.188, + 0.468, + 0.367 + ], + "angle": 0, + "content": "tic texture requires highly accurate camera poses. Thus, we present an RGBD-aided Structure from Motion (SfM) which combines the advantages of both VIO [12] and SfM [46]. Since ARKit based on VIO is robust to the degradation of visual information such as low-textured surface, we perform incremental triangulation with initial poses obtained from ARKit. We also propose a depth filtering method to handle noisy depth from the smartphone. These filtered depth points are exploited as an additional depth factor for bundle adjustment. Our RGBD-aided SfM can estimate highly precise camera poses due to the good initial poses and additional constraints from the filtered depth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.468, + 0.625 + ], + "angle": 0, + "content": "Our 3D geometry reconstruction process adopts a neural implicit representation [50, 55] for surface reconstruction with volumetric rendering. We observe that the neural implicit representation can perform more complete and smoother reconstruction than the classical methods which are known to be robust for 3D reconstruction in the wild. Furthermore, we introduce a new training method for neural implicit representations. In the early stage of training, we propose a regularization method that leverages the prior information from the classical MVS method. After obtaining the decent shape, we avoid using the prior information and generate sparse voxel octree to enable efficient sampling for training. Our training method can improve the performance of the neural implicit representations. Consequently, we show the generalization capabilities of our surface reconstruction method in real world objects collected by the smartphone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Given a mesh extracted from the trained neural implicit representation, we can run classical texture mapping algorithms [53, 69] to generate texture maps that often exhibit blurring artifacts and color misalignment. We propose applying differential rendering [31] to fine-tune these texture maps obtained from classical texture mapping via a photometric loss. Compared to classical 3D reconstruction [23, 47, 62] and texture mapping [53, 69] approaches, our method shows a better ability to reconstruct the 3D model and produce realistic textures. We evaluate our approach on the data collected by our smartphone application. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.816, + 0.468, + 0.845 + ], + "angle": 0, + "content": "- We present a unified framework to reconstruct the textured mesh using a smartphone." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We propose a depth filtering scheme for noisy depths and refine initial poses from ARKit by using bundle adjustment with depth factor." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.816, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "- Our pipeline builds on classical 3D reconstruction and texture mapping. We leverage a neural geometry representation to enable surface reconstruction of complex shapes and a differentiable rendering to generate high-fidelity texture maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.181, + 0.642, + 0.196 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Classical 3D Reconstruction Traditional Structure from Motion (SfM) systems [46, 59] have been popular as a vision-based 3D reconstruction due to their robustness to various scenarios [48, 49, 58]. The SfM systems extract and match features [36] of adjacent images of the same scene and generate the camera poses and sparse 3D points via triangulation [20]. Then, Multi-view Stereo (MVS) algorithms [13, 52] compute the dense depth maps of each scene with multiple calibrated images. PatchMatch-based MVS [14, 47, 61, 62] methods are popular because they are proper for depth and normal optimization. After that, surface reconstruction methods [26, 27, 30] can be applied to reconstruct meshes from point clouds. However, all these methods often fail to reconstruct the 3D geometry of low-textured areas. RGB-D Reconstruction techniques fuse many overlapping depth maps into a single 3D model. Starting from the classical TSDF-fusion [8], several methods [9, 23, 42, 57] produce voxel-based TSDF with commodity-level depth cameras such as Kinect. However, these methods have difficulty handling outliers and thin geometry due to noisy depth measurements. Due to the noise and limited range of the LiDAR sensor in smartphone, RGB-D reconstruction methods fail to estimate the camera odometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Reconstruction with differentiable rendering Recently, NeRF [38] and its variants [37, 39, 65, 68] have used volumetric implicit representations for rendering and learn \\(\\alpha\\)-compositing of a radiance field along rays. These methods yield remarkable quality of novel view synthesis results and do not need ground truth masks. However, 3D geometry from volumetric representations is far from satisfactory due to their volumetric properties [68]. Some methods [41, 64, 67] reconstruct the surface by encoding the network as an implicit function such as a signed distance function (SDF) or occupancy function. Many hybrid approaches [4, 10, 43, 55, 63] unify surface and volumetric representations by regarding surface as defining volumes near the surface, enabling accurate surface reconstruction without masks. Their follow-up works [19, 34, 54, 67] import prior information from the learning-based network. We avoid exploiting prior hints obtained from learning-based models because they are expensive to acquire new training datasets and time-consuming to train other neural networks. Classical Texture Optimization The matching between texture and geometry is a significant step for building realistic 3D models. Lempitsky et al. [32] apply the pairwise Markov Random Field to choose an optimal view for each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.197 + ], + "angle": 0, + "content": "face as texture. Waechter et al. [53] present a global color adjustment to mitigate the visual seams between face textures. Zhou et al. [69] optimize the camera poses using a color consistency measurement and geometric error via local image warping to obtain sharp texture. Bi et al. [6] employ a patch-based image synthesis to produce a texture per each face to optimize the camera pose and geometric error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.439 + ], + "angle": 0, + "content": "Texture Learning Recently, learning-based methods [17, 22, 25, 51] have been proposed to generate textures for a 3D model. Huang et al. [22] introduce adversarial learning for texture generation in RGB-D scans. Thies et al. [51] store neural textures for a known mesh with given UV mapping and decode them to color for novel views. Some works [7, 60] learn neural texture with neural volumetric rendering. Differentiable rendering methods [16, 21, 33, 40] usually focus on estimating surface radiometric properties from images. These methods can learn to predict not only texture but also geometry via image loss with ground truth masks. Unfortunately, differentiable rendering methods heavily rely on pixel-level accurate masks. Acquiring accurate ground-truth masks in the real-world environment is a difficult task [29]. Therefore, we apply the differentiable rendering to fine-tune texture maps with a fixed mesh." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.208, + 0.467 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Given images, low-resolution depths and corresponding camera poses from ARKit, our goal is to reconstruct both the texture and geometry of real-world 3D objects. In Fig. 4, our pipeline consists of three stages: RGBD-aided Structure from Motion (SfM), Geometry Reconstruction, and Texture Optimization. Since both depth maps and camera poses provided by ARKit are very noisy, we construct a SfM system to filter out depth maps and refine poses. We then formalize our task as training a neural model for zero-level set of a signed distance function (SDF) guided by classical 3D reconstruction. After training, we extract the mesh with Marching Cubes [35] and apply classical texture reconstruction [53] to generate initial texture maps. Next, we exploit differentiable rendering [31] to finetune the texture maps with the supervision of multi-view images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.395, + 0.726 + ], + "angle": 0, + "content": "3.1. RGBD-aided Structure from Motion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.81 + ], + "angle": 0, + "content": "The goal of this section is to refine initial poses from ARKit by performing a bundle adjustment with depth maps from the smartphone. Unlike simple structures like walls or floors, 3D reconstruction for capturing a high level of detail is required to refine inaccurate poses from ARKit." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Preprocessing We introduce two processes to obtain reliable depth estimates: single-view filtering and multi-view filtering. Single-view filtering utilizes a confidence map provided by ARKit, which has the same resolution as the depth map. Each pixel in the confidence map has 0, 1, or 2 with the value 0 as least confident and 2 serving as most" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.201, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Figure 3. Overview of our pipeline. It is comprised of three steps: RGBD-aided Structure from Motion (Sec. 3.1), Geometry Reconstruction (Sec. 3.2), and Texture Optimization (Sec. 3.3). Our method builds on classical 3D reconstruction and texture mapping. To overcome the limitations of classical methods, we apply a neural geometry representation and neural texture fine-tuning to update 3D geometry and texture maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.893, + 0.553 + ], + "angle": 0, + "content": "confident with respect to depth-value accuracy. We choose depth pixels corresponding to 1 or 2 in the confidence map. A main concept of the multi-view filtering is to remove outliers in terms of the reprojection error. Given the reference image \\( I_{ref} \\) with depth \\( D_{ref} \\), the source image \\( I_{src} \\) with depth \\( D_{src} \\), the depth map \\( D_{src}' \\) can be transformed from the source view to the reference view with accurate pixel-level correspondence. Then, the depth of pixel \\( p \\) in the reference image will be considered as an inlier if the difference \\( \\delta_{depth} \\) between \\( D_{src} \\) and \\( D_{src}' \\) is smaller than a small constant value \\( \\epsilon \\). Then, assuming a set of neighboring N source images, we compute the number of inliers which fulfills \\( \\delta_{depth} < \\epsilon \\) for each source image, and keep depth pixels if the number of inliers is greater than a certain threshold. Finally, we apply both single-view and multi-view filtering to remove all depth outliers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Pose Refinement We consider camera poses obtained from ARKit as initial poses and exploit matching correspondences to perform incremental triangulation [46]. After combining filtered depth points from LiDAR and 3D points from incremental triangulation, we propose a depth factor and run bundle adjustment with the depth factor and reprojection error. Let \\( P_{i} \\) denote the camera parameter of the \\( i \\)-th camera and \\( X_{j} \\) denote point parameters seen by the \\( i \\)-th camera. We minimize the proposed joint objective function \\( E \\), which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.716, + 0.892, + 0.753 + ], + "angle": 0, + "content": "\\[\nE = \\sum_ {i} \\sum_ {j} \\rho_ {i} \\left(\\left\\| \\phi \\left(P _ {i}, X _ {j}\\right) - x _ {i j} \\right\\| ^ {2}\\right) + \\frac {1}{\\eta} \\left| D _ {i} - z _ {i} \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Phi\\) is the function that projects points to the image plane, and \\(\\rho\\) denotes the Cauchy function as the robust loss function. \\(x_{ij}\\) is the feature point corresponding to the 3D point \\(X_{j}\\) seen by the \\(i\\)-th camera. \\(z_{i}\\) is the depth value of points projected from the 3D point \\(X_{j}\\) and \\(D_{i}\\) is the filtered depth of pixels corresponding to \\(z_{i}\\). We define \\(\\eta\\) as sensor measurement noise and use it to down-weight noisy depth. We notice that the depth factor has two benefits. First, additional information from the depth sensor is a useful con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16676" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "strain for accurate pose estimation and reliable triangulation. Furthermore, this depth factor guides the scale of the SfM to follow the metric scale of the depth sensor." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.145, + 0.312, + 0.161 + ], + "angle": 0, + "content": "3.2. Geometry Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.396 + ], + "angle": 0, + "content": "Neural Geometry Representations Instead of classical 3D reconstruction, we adopt NeuS [55], which combines the advantages of both volume and surface rendering. In other words, it can extract accurate surfaces even without pixel-level masks. Our scanned object is represented by two Multi-layer Perceptrons (MLPs): geometry network \\( f_{sdf} \\), which maps a 3D point \\( x \\in \\mathbb{R}^3 \\) to a signed distance, and color network \\( f_{color} \\) which converts a 3D point \\( x \\) and a viewing direction \\( v \\in \\mathbb{R}^3 \\) to a color \\( c \\). The 3D geometry of an object is represented implicitly as an SDF with zero-level set \\( \\{x \\in \\mathbb{R}^3 | f_{sdf}(x) = 0\\} \\). These two parameters of geometry and color are learned by the volume rendering technique. For a ray \\( \\{x(t) = o + tv | t \\geq 0\\} \\) with camera center \\( o \\) and viewing direction \\( d \\), the color is rendered with \\( n \\) sampled points along the ray as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.404, + 0.47, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\hat {C} = \\sum_ {i = 1} ^ {n} T _ {i} \\alpha_ {i} c (x (t), v), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.469, + 0.516 + ], + "angle": 0, + "content": "where \\( T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j}) \\) is the accumulated transmittance, and \\( \\alpha_{i} \\) is the discrete opacity. Here, \\( \\alpha_{i} = 1 - \\exp(-\\int_{t_{i}}^{t_{i+1}} \\rho(t) dt) \\), and \\( \\rho(t) \\) is opaque density function borrowed from the definition in NeuS [55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Training Process RGBD-aided SfM already estimates both poses and sparse depths by matching feature correspondences. Given this prior information, we leverage an MVS algorithm [62] to estimate the depth and normal for all images with less cost. It is worth noting that we avoid using learning-based MVS algorithms because they are not robust to diverse real-world scenes and are hard to evaluate reliability. In this section, we modify the training process of NeuS to avoid losing fine details of complex shape. Our training process is divided into two stages. In the first stage, we train a neural radiance field with a prior depth and normal obtained from the MVS [62] algorithm to supervise the training process. Specifically, we sample \\( K \\) pixels \\( p_k \\) and their corresponding color values \\( I(p_k) \\) in each iteration. Here, we propose a regularization method that relies on prior depth \\( D \\) and normal \\( N \\) from MVS algorithm. We unproject depth under the given pose to obtain 3D point \\( X(p_k) \\). Then, we use this 3D point and its corresponding normal \\( N(p_k) \\) from MVS to regularize the geometry network by minimizing" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.826, + 0.469, + 0.859 + ], + "angle": 0, + "content": "\\[\nL _ {r e g} = \\sum_ {k} w _ {d} \\left| f _ {s d f} \\left(X \\left(p _ {k}\\right)\\right) \\right| + \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.861, + 0.425, + 0.88 + ], + "angle": 0, + "content": "\\[\nw _ {n} (1 - \\langle \\nabla_ {p} f _ {s d f} (X (p _ {k})), N (p _ {k}) \\rangle),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( w_{d} \\) and \\( w_{n} \\) are weights for each term. This regular-" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.095, + 0.885, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Figure 4. Overview of Geometry Reconstruction (Section 3.2) and Texture Optimization (Section 3.3). Our goal in geometry reconstruction is to train geometry network \\( f_{sdf} \\) and color network \\( f_{color} \\) via volumetric rendering with two training steps. In step 1, we leverage prior geometric information from the MVS algorithm [62]. Then, our method utilizes a sparse voxel octree to efficiently sample points near the surface. In texture optimization, we first utilize a classical texture reconstruction to generate texture maps for a given mesh. Then, our method applies a differentiable rendering to fine-tune texture maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.492 + ], + "angle": 0, + "content": "ization will guide the SDF to learn explicit supervision from the MVS algorithm. Although this regularization is helpful for learning prior geometry, it is susceptible to the noisy results derived from the MVS algorithm which tends to lose fine details and texture-less regions. Thus, we will only use this regularization in the first stage. The color of each pixel can be computed by E.q. 4. The color loss \\( L_{c} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.501, + 0.891, + 0.533 + ], + "angle": 0, + "content": "\\[\nL _ {c} = \\sum_ {k} | \\hat {C} (p _ {k}) - I (p _ {k}) |. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.569 + ], + "angle": 0, + "content": "The Eikonal loss [18] to regularize the SDF is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.569, + 0.891, + 0.601 + ], + "angle": 0, + "content": "\\[\nL _ {e i k} = \\sum_ {k, i} \\left(\\| \\nabla f _ {s d f} \\left(p _ {k, i}\\right) \\| _ {2} - 1\\right) ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.607, + 0.875, + 0.622 + ], + "angle": 0, + "content": "The overall loss function \\(L\\) in the first stage is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.633, + 0.891, + 0.65 + ], + "angle": 0, + "content": "\\[\nL = L _ {c} + L _ {e i k} + L _ {r e g}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.811 + ], + "angle": 0, + "content": "In the second stage, inspired by [50], we adopt a sparse voxel octree to guide the sampling process and capture fine details. After the first stage, we can extract a triangular mesh from the SDF via the Marching Cube [35] and define a sparse voxel volume based on this mesh. Then, based on the sparse voxel volume, we can avoid redundant point samples outside the sparse voxel and focus on point samples near the surface samples. Here, our total loss is \\( L = L_{c} + L_{eik} \\) which is the same as E.q. 6 except for \\( L_{reg} \\). To capture fine details, we will not use \\( L_{reg} \\) in E.q. 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Mesh Simplification A highly detailed 3D mesh from the SDF network with marching cube algorithms [35] leads to enormous memory requirements. To obtain a lightweight mesh, we explore popular quadric mesh simplification [15]. Although the mesh simplification generated small holes, the lightweight mesh enables us to reduce the computation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16677" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "complexity of the texture mapping process. We set the decimation ratio to \\(0.4\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.13, + 0.281, + 0.148 + ], + "angle": 0, + "content": "3.3. Texture Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.47, + 0.274 + ], + "angle": 0, + "content": "Classical Texture Reconstruction For a triangular mesh \\(M\\) provided by the previous section, we first utilize Waechter et al. [53] with poses and images across different views to construct a texture map \\(T_{orig}\\). Compared to the color network from Section 3.2, this classical method [53] can acquire visually realistic texture images. However, it cannot avoid texture bleeding on the boundary of different views and blurring artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.471, + 0.534 + ], + "angle": 0, + "content": "Texture Fine-tuning To tackle blurring and texture bleeding artifacts, we propose a fine-tuning step by optimizing the texture map. Here, we parameterize a texture map \\( T \\) obtained from classical texture reconstruction. Given a 3D Mesh \\( M \\) with the 2D texture map \\( T \\) and multi-view images \\( I_{i} \\) with the corresponding camera poses \\( P_{i} \\), we leverage a differentiable renderer \\( R \\) to generate an image \\( \\hat{I}_i \\) and depth \\( \\hat{D}_i \\), i.e. \\( \\hat{I}_i, \\hat{D}_i = R(M, P_i) \\). A differentiable renderer \\( R \\) [31] performs mesh rasterization and then texture sampling to render an image \\( \\hat{I}_i \\). In other words, we project texture onto the image plane and compare it to the corresponding view of the image. Furthermore, we compare the rendered image with a different view image by minimizing multi-view photometric loss. We use the combination of the L1 and SSIM [56] as the photometric loss \\( L_{photo} \\) in image space between the rendered image \\( \\hat{I}_i \\) and the reference image \\( I_{i} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.542, + 0.469, + 0.587 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p h o t o}} = \\sum_ {i} (1 - \\alpha) \\| \\hat {I} _ {i} - I _ {i} \\| + \\frac {\\alpha}{2} (1 - S S I M (\\hat {I} _ {i}, I _ {i})), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.587, + 0.47, + 0.709 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is set to 0.85. Additionally, the differentiable render estimates depth \\(\\hat{D}_i\\) by interpolating \\(z\\) coordinates of each vertex. With the estimated depth \\(\\hat{D}_i\\) and intrinsic parameter \\(K\\), we can generate a synthesized frame \\(I_i'\\) by reprojecting adjacent frames \\(I_{i'}\\) to the current frame \\(I_i\\) with the camera intrinsic matrix \\(K\\), the estimated depth \\(\\hat{D}_i\\), and the relative pose \\(P_{i\\rightarrow i'}\\). The multi-view photometric loss \\(L_{multi}\\) similar to Eq. 7 is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.719, + 0.37, + 0.737 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {\\prime} (p) = I _ {i ^ {\\prime}} \\langle \\pi (K P _ {i \\rightarrow i ^ {\\prime}} \\hat {D} _ {i} (p) K ^ {- 1} \\tilde {p}) \\rangle ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.739, + 0.469, + 0.779 + ], + "angle": 0, + "content": "\\[\nL _ {m u l t i} = (1 - \\alpha) \\| I _ {i} ^ {\\prime} - I _ {i} \\| + \\frac {\\alpha}{2} (1 - S S I M \\left(I _ {i} ^ {\\prime}, I _ {i}\\right)), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.789, + 0.469, + 0.848 + ], + "angle": 0, + "content": "where \\(\\tilde{p}\\) is the homogeneous coordinate of \\(p\\), \\(\\pi\\) means projection from homogeneous to image coordinates, and \\(\\langle \\cdot \\rangle\\) indicates the bilinear sampling function. We define our total loss \\(L\\) to be" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.85, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {p h o t o}} + L _ {\\text {m u l t i}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Finally, the differentiable rendering enables the loss gradients to back-propagate to update the textures parameters." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "Since our method does not require ground truth masks, we avoid considering optimization for the geometry. We observe that joint optimization with geometry and texture fails to converge without the ground truth mask." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.166, + 0.634, + 0.183 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.191, + 0.694, + 0.208 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Data Collection: Our smartphone application, running on an iPhone 13 Pro using ARKit 6, can record various sensor outputs: synchronized images, poses, confidence maps, dense depth maps from the LiDAR scanner, and IMU sensor information. We built two types of datasets: ARKit-video and AR-capture. ARKit-video collects long video sequences similar to datasets constructed from RGBD-SLAM [70]. The second type is collected by our smartphone application based on Object Capture1. AR-capture is a multiview dataset of objects consisting of high-quality 4K images with depth and gravity data. The ARKit-video collects thousands of frames sampled from the video with resolution of \\(1280 \\times 720\\). In contrast, AR-capture mostly contains less than fifty images at \\(4032 \\times 3096\\) pixels. We collect 11 objects: robot arm 1, robot arm 2, plant, tree, sofa, bike, recliner chair, cafe stand, delivery robot, office chair, and camera stand. We randomly select \\(70 \\%\\) as the training images and utilize the \\(30 \\%\\) as test images for evaluating the novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.606 + ], + "angle": 0, + "content": "DTU Dataset: The DTU dataset [24] is a MVS dataset that provides reference point clouds acquired with laser sensor in the controlled lab. Since AR-capture and ARKit-video do not include 3D point clouds acquired with specialized hardware, we use the DTU dataset to evaluate the performance of geometry reconstruction. We follow the evaluation of NeuS [55] and exploit the chamfer distance as metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Implementation Details: We build the RGBD-aided Structure from Motion based on COLMAP [46], an existing SfM tool. Instead of SIFT [36], our system uses learning-based features [11, 44, 45] for feature extraction and matching to avoid failure in a featureless scene. ARKit offers an interface for accessing LiDAR data as a depth image. Following the work [55], the SDF network \\( f_{sdf} \\) is implemented as an 8-layer MLP a hidden dimension of 256, and geometric initialization [3] for the network parameters. We construct the color network \\( f_{color} \\) as MLPs with 4-layer MLP with hidden dimension of 256. We optimize networks in geometry reconstruction with Adam [28] and an initial learning rate of 5e-4. We train the first step for 60K iterations and the second step for 140K iterations. In terms of objects with complex shape, it takes 240K iterations for second step. In Texture Optimization, we utilize the public github code for classical method [53]. Then, for texture fine-tuning, the Adam optimizer is used with the learning rate for the tex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.838, + 0.901 + ], + "angle": 0, + "content": "1https://developer.apple.com/videos/play/wwwdc2021/10076/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.082, + 0.195, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.088, + 0.264, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.083, + 0.388, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.088, + 0.457, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.168, + 0.187, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.261, + 0.145, + 0.272 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.193, + 0.275, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.261, + 0.239, + 0.272 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.175, + 0.38, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.262, + 0.337, + 0.272 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.194, + 0.465, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.261, + 0.431, + 0.272 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.417 + ], + "angle": 0, + "content": "Figure 5. We show the benefits of RGBD-aided Structure from Motion. (a) illustrates the 3D point clouds obtained from initial depths with initial poses. In (c), we accumulate the 3D point clouds (by using filtered depths and refined poses) from RGBD-aided SfM. For a detailed description, we visualize the zoomed-in blue and red regions using filtered depth. Initial depth maps and poses from ARkit are input to the geometry reconstruction in Section 3.2 and its results are shown in (b). The results of our pipeline in (d) lead to higher quality reconstructions than (b) which does not apply RGBD-aided SfM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.469, + 0.46 + ], + "angle": 0, + "content": "ture parameter. This step takes 3000 epochs to fine-tune the texture parameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.279, + 0.485 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.78 + ], + "angle": 0, + "content": "In Fig. 5, we visualize the effectiveness of RGBD-aided Structure from Motion in Section 3.1 on our dataset. Figure 5 (a) represents the 3D point clouds by initial depth maps and poses obtained from ARKit. Unfortunately, initial sensor data is too noisy to perform an accurate 3D reconstruction. After multi-view depth filtering as preprocessing, zoomed-in regions in (a) show the point clouds by reprojecting filtered depth maps with initial poses. Compared to initial data, multi-view depth filtering can handle noisy depth measurements. Figure 5 (c) provides the results of preprocessing and pose refinement. Pose refinement enables us to globally align 3D reconstruction without noticeable camera drift and with better local quality. For qualitative comparison, we take as input initial sensor data from ARKit and perform geometry reconstruction in Section 3.2 in Fig. 5 (b). Compared to our experimental results in Fig. 5 (d), it is difficult to achieve comparable quality and better reconstruction. Consequently, our proposed RGBD-aided SfM significantly improves geometry reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Figure 6 qualitatively compares our method to VolSDF [63], TSDF-fusion [23], and ACMP [62]. We apply screened Poisson surface reconstruction [27] to generate a mesh from point clouds of ACMP. Among all four algorithms, we apply the same pose estimation algorithm from Section 3.1 for fair comparison. TSDF-fusion, which requires sufficient scene overlap, exploits more frames than our method and ACMP. Due to the low-textured regions," + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.082, + 0.593, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.082, + 0.658, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.082, + 0.721, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.082, + 0.785, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.082, + 0.855, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.154, + 0.596, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.155, + 0.661, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.155, + 0.727, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.155, + 0.793, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.155, + 0.86, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.222, + 0.596, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.364, + 0.58, + 0.372 + ], + "angle": 0, + "content": "Scene" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.222, + 0.663, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.363, + 0.642, + 0.372 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.222, + 0.726, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.363, + 0.714, + 0.372 + ], + "angle": 0, + "content": "VolSDF" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.222, + 0.787, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.363, + 0.788, + 0.372 + ], + "angle": 0, + "content": "TSDF-Fusion" + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.222, + 0.86, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.363, + 0.845, + 0.372 + ], + "angle": 0, + "content": "ACMP" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Figure 6. Reconstruction of our method, VolSDF [63], TSDF-Fusion [23], and a mesh from point clouds of ACMP [62] processed with screened Poisson surface reconstruction [27]. Note that the neural surface reconstruction is more perceptually convincing in the ARKit-video dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.658 + ], + "angle": 0, + "content": "the surface of ACMP is very noisy. TSDF-fusion utilizes depth maps obtained from low-resolution LiDAR. Here, since this low-resolution depth sensor is very noisy, we only use filtered depth maps after preprocessing in Section 3.1. However, this model still cannot reconstruct either thin or challenging structures accurately (e.g., a chair armrest). Both implicit neural surface reconstruction approaches show comparable performance, and both are superior to classical 3D reconstruction. We observe that the difference between neural geometric representations (Ours and VolSDF) is negligible unlike pose quality. It is worth noting that using accurate poses as input significantly improves the 3D reconstruction quality as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "For qualitative comparison of textured meshes, we compare our proposed method with four different methods: 1) ACMP [62] is a state-of-the-art multi-view stereo algorithm: 2) Color Map Optimization (CMO) [69] is a texture mapping method for RGB-D scanning: 3) Waechter et al. [53] is a large-scale texture reconstruction based on global color adjustment, 4) NeuS [55] is the recent neural surface reconstruction based on volumetric rendering. We reconstruct a 3D mesh from the point cloud of ACMP with Screened Poisson Surface reconstruction [27]. In the case of CMO, since explicit depth maps from low-resolution LiDAR are too noisy, we utilize depth maps rendered by a triangular mesh from our geometry reconstruction step for a fair comparison. With these rendered depth maps and corresponding images, we run a color map optimization for mapping color images onto our reconstructed mesh. To train" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.202, + 0.104, + 0.774, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.677, + 0.893, + 0.719 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison between (a) Ours, (b) Waechter et al. [53], (c) CMO [69], (d) ACMP [62], and (e) NeuS [55]. Compared to textured mesh (3-6 columns) obtained from different algorithms, the result of our method is perceptually closer to the original scene. Additional qualitative results are added in the supplementary material due to the limited space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.837 + ], + "angle": 0, + "content": "NeuS, we follow a training process similar to that in Sec. 3.2. We then extract a mesh from the SDF network in NeuS and apply the Blender Smart UV Project tool to obtain a pervertex UV mapping. To fill a texture for this mesh, points densely sampled from triangle mesh with interpolated UV coordinates are fed to the color network, and eventually we can obtain texture maps for the given mesh." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In Fig. 7, we show the reconstructed texture for novel views. The visual results of our approach are visually sharper and perceptually closer to the ground truth scene. Results of NeuS and CMO exhibit blurring artifacts and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.897 + ], + "angle": 0, + "content": "color misalignment. The reconstructed mesh from ACMP is very noisy due to incorrect geometry. Here, Waechter et al. (Fig. 7 (b)) is the result of classical texture reconstruction in Section 3.3. This method is more photorealistic and sharper than NeuS, CMO, and ACMP. Then, after texture fine-tuning, our method avoids the seams and texture misalignment often seen with Waechter et al. Table 1 reports the quantitative results on two datasets. Our method outperforms other texture optimization methods on all evaluation metrics. It indicates that our method satisfies both the high fidelity of the texture and its conformity to the ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16680" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.158, + 0.089, + 0.807, + 0.257 + ], + "angle": 0, + "content": "
CategoryDataNeuS [55]CMO [69]Waechter et al. [53]Ours
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
robot arm 1Ac20.2920.8890.08318.4730.8910.08319.9090.8720.06020.0490.8710.057
robot arm 2Ac20.7410.8820.08621.1070.8830.09220.1900.8620.06720.6520.8580.062
plantAc16.7480.6420.29518.4620.6270.31617.9210.6350.18619.9700.6910.146
treeAc18.4090.8690.11718.4480.8700.12717.5180.8580.09618.3310.8580.102
sofaAc19.9100.8720.13921.1700.8400.13619.7590.8290.11321.6080.8940.109
bikeAc20.9710.7780.18321.3110.7780.20619.8220.7400.08819.9430.7370.078
MeanAc19.5120.8220.15119.8250.8150.16019.1860.7990.10120.0950.8180.092
recliner chairAv24.9200.8930.09924.7610.8760.11223.6920.8640.05425.0450.8960.056
cafe standAv20.4980.7790.17322.9340.8280.15417.9150.7980.11222.6340.8180.095
delivery robotAv22.4120.8970.12623.5490.9080.12825.2460.9290.07323.4510.9060.086
office chairAv20.7220.8920.09621.1200.8990.10320.3150.8650.09420.9820.8920.083
camera standAv21.9010.9140.06921.8860.9070.07421.1210.8990.05922.2080.9190.060
MeanAv22.0920.8750.11322.8500.8840.11521.6580.8710.07922.8680.8860.076
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.26, + 0.895, + 0.289 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on our dataset. Ac and Av denote the AR-catpure dataset and ARKit-video dataset, respectively. We improve the LPIPS metric by \\(8.9\\%\\) in AR-capture and by \\(3.8\\%\\) in ARKit-video." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.295, + 0.808, + 0.364 + ], + "angle": 0, + "content": "
Methodscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122Mean
COLMAP [47] with trim=70.450.910.370.370.901.000.541.221.080.640.480.590.320.430.450.65
UNISURF [43]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
NeuS [55]1.371.210.730.401.200.700.721.011.160.820.661.690.390.610.510.87
VolSDF [63]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Ours without 2nd step1.041.610.670.640.880.700.621.141.100.890.661.020.500.580.510.84
Ours0.961.370.670.600.880.700.451.141.100.840.661.110.500.580.500.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.368, + 0.895, + 0.398 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on DTU [24]. Chamfer distance is employed as the evaluation metric. COLMAP is superior to other neural geometry reconstructions. Our training process can improve the NeuS by \\(8\\%\\)." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.409, + 0.165, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.409, + 0.245, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.409, + 0.331, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.409, + 0.438, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons of texture optimization algorithms based on differentiable rendering: (a) our method, (b) Lin et al. [33], and (c) Munkberg et al. [40]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.825 + ], + "angle": 0, + "content": "In Fig. 8, we show the results of other texture optimization methods [33,40] based on differentiable rendering [31]. Both methods aim to learn geometry, textures, and lighting parameters. However, they heavily rely on ground truth masks which require additional annotation costs. Without ground truth masks, we note that joint optimization over geometry and texture is unstable and often has negative effects on geometry. For a fair comparison, we therefore provide the fixed mesh topology and train only texture and lighting parameters through differentiable rendering [31]. Since we capture objects in the wild under unknown environmental lighting conditions, it is difficult to disentangle materials and lighting via differentiable rendering. Neither method [33, 40] is able to generate realistic texture maps compared to either our method or classical texture reconstruction approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Experimental Results on DTU dataset: Unlike our dataset, DTU [24] is captured by specialized hardware in the controlled lab environment. Thus, this dataset provides the ground truth poses and the RGBD-aided SfM is not necessary. As shown in Table 2, the surfaces obtained by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.895, + 0.594 + ], + "angle": 0, + "content": "COLMAP [47] with trimming value 7 achieve high-quality reconstruction. Our method exploits MVS [47] to supervise the training NeuS and outperforms NeuS. For the ablation study, when we only train our network with the first step (ours without 2nd step), its result is not better than our method in some scans such as scan37. In our experiment, sparse voxel octree sampling does not have a significant effect on geometry performance. Since our training process can be applied to any other 3D neural surface reconstruction methods. In the future, we will apply this training process to other recent 3D neural surface reconstruction algorithms [10, 63, 66], which shows better results on DTU." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.614, + 0.878, + 0.631 + ], + "angle": 0, + "content": "5. Conclusion, Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.869 + ], + "angle": 0, + "content": "In this paper, we have developed a practical framework to reconstruct high-quality textured meshes of 3D objects by using a smartphone. Given initial poses from ARKit, multi-view images, and low-resolution depth maps, our approach can refine initial poses and depth maps from ARKit. We adopt neural geometry representation [55] for surface reconstruction and improve its performance by modifying the training process. Finally, a differentiable rendering is designed for fine-tuning the texture map acquired from the classical texture mapping method. In the future, our goal is to reduce the training time which can be comparable with classical methods. Unfortunately, most neural geometry reconstruction [10, 55, 63, 66] requires a long training time. We would like to reduce the training time in our framework for practical purposes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements This research was partly supported by Army Cooperative Agreement W911NF2120076." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16681" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Apple arkit. https://developer.apple.com/documentation/arkit/. Accessed: 2022-11-09. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Google arcore. https://developers.google.com/ar. Accessed: 2022-11-09. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[3] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.468, + 0.296 + ], + "angle": 0, + "content": "[4] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.437 + ], + "angle": 0, + "content": "[6] Sai Bi, Nima Khademi Kalantari, and Ravi Ramamoorthi. Patch-based optimization for image-based texture mapping. ACM Trans. Graph., 36(4):106-1, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[10] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR Deep Learning for Visual SLAM Workshop, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[12] Alex Flint, Oleg Naroditsky, Christopher P Broaddus, Andriy Grygorenko, Stergios Roumeliotis, and Oriel Bergig. Visual-based inertial navigation, Dec. 11 2018. US Patent 10,152,795. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[13] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. June 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[16] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV, page 88-104, Berlin, Heidelberg, 2020. Springer-Verlag. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[19] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[21] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7498-7507, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[22] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[23] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[26] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16682" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[29] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[31] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[32] Victor Lempitsky and Denis Ivanov. Seamless mosaicing of image-based texture maps. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-6. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[33] Lixiang Lin, Jianke Zhu, and Yisu Zhang. Multiview textured mesh recovery by differentiable rendering. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.421, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[34] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.521, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[36] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.564, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.649, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[40] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. arXiv:2111.12503, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.833, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[44] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[47] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[49] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80(2):189-210, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[50] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.597, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Trans. Graph., 38(4), jul 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[52] George Vogiatzis, Philip HS Torr, and Roberto Cipolla. Multi-view stereo via volumetric graph-cuts. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 391-398. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[53] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.764, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[54] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.819, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[57] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[58] Aji Resindra Widya, Yusuke Monno, Masatoshi Okutomi, Sho Suzuki, Takuji Gotoda, and Kenji Miki. Stomach 3d reconstruction based on virtual chromoendoscopic image generation. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1848-1852. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[59] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[60] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7119-7128, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[61] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[62] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[66] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[67] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[68] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[69] Qian-Yi Zhou and Vladlen Koltun. Color map optimization for 3d reconstruction with consumer depth cameras. ACM" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Transactions on Graphics (ToG), 33(4):1-10, 2014. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[70] Michael Zollhöfer, Patrick Stotko, Andreas Görtlitz, Christian Theobalt, Matthias Nießner, Reinhard Klein, and Andreas Kolb. State of the art on 3d reconstruction with rgb-d cameras. In Computer graphics forum, volume 37, pages 625-652. Wiley Online Library, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16684" + } + ] +] \ No newline at end of file diff --git a/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_origin.pdf b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d85e122f974f5c536231c39352afbe9759f3875 --- /dev/null +++ b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/a9ab43cd-15ce-44b4-8413-0c6fe896d349_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:947637660b6b97e512d43e0f9d96bab8d789d64b75be7cbe4bb4644381bed88c +size 9131599 diff --git a/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/full.md b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..83bf2314de95585f07b40fcf714328370ddae305 --- /dev/null +++ b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/full.md @@ -0,0 +1,354 @@ +# TMO: Textured Mesh Acquisition of Objects with a Mobile Device by using Differentiable Rendering + +Jaehoon Choi $^{1,2}$ Dongki Jung $^{1}$ Taejae Lee $^{1}$ Sangwook Kim $^{1}$ Youngdong Jung $^{1}$ Dinesh Manocha $^{2}$ Donghwan Lee $^{1}$ $^{1}$ NAVER LABS $^{2}$ University of Maryland + +# Abstract + +We present a new pipeline for acquiring a textured mesh in the wild with a single smartphone which offers access to images, depth maps, and valid poses. Our method first introduces an RGBD-aided structure from motion, which can yield filtered depth maps and refines camera poses guided by corresponding depth. Then, we adopt the neural implicit surface reconstruction method, which allows for high-quality mesh and develops a new training process for applying a regularization provided by classical multi-view stereo methods. Moreover, we apply a differentiable rendering to fine-tune incomplete texture maps and generate textures which are perceptually closer to the original scene. Our pipeline can be applied to any common objects in the real world without the need for either in-the-lab environments or accurate mask images. We demonstrate results of captured objects with complex shapes and validate our method numerically against existing 3D reconstruction and texture mapping methods. + +# 1. Introduction + +Recovering the 3D geometry of objects and scenes is a longstanding challenge in computer vision and is essential to a broad range of applications. Depth sensing technologies range from highly specialized and expensive turn-table 3D scanners and structured-light scanners to commodity depth sensors. More recently, advances in mobile devices have developed a new method for 3D capture of real-world environments with high-resolution imaging and miniaturized LiDAR. Specifically, the modern smartphone such as iPhone 13 Pro are equipped with RGB camera, accelerometer, gyroscope, magnetometer, and LiDAR scanner. These various sensors can provide high-resolution images, low-resolution depth from the LiDAR scanner, and associated camera poses offered by off-the-shelf visual-inertial odometry (VIO) systems such as ARKit [1] and ARCore [2]. + +Today's smartphones offer low-resolution depth maps + +![](images/415abffe65e52adcfca8c622fe9a069aa2e7915b859afb4c130414484fed2770.jpg) +(a) + +![](images/d87e713b938a2e0180aca13b15bb5108c6d26f186bb904b0e6e27811456f102a.jpg) +(b) + +![](images/eae39edf8883c9c41473a42704556ab8d563187a9117edaadc148e5bef3a0f5e.jpg) +(c) + +![](images/64932692b2dde51a95e569123e6968423e566b687d6ac549099e26b4bf8fdc6a.jpg) +(d) +Figure 1. Example reconstruction results collected from a smartphone in the wild. (a) Data acquisition setup. (b) Images captured from a smartphone. (c) A reconstructed mesh. (d) A novel view of textured mesh. Our proposed method can reconstruct the high-quality geometric mesh with a visually realistic texture. + +[5] and valid poses. However, depth maps are very noisy and suffer from the limited range of depth sensors. Although this depth sensor can build a simple 3D structure such as a wall or floor, it cannot reconstruct objects with complex and varying shapes. Thus, the RGBD scanning methods [9, 23, 42, 57] are not suitable for these objects. Instead of the depth sensor, multi-view stereo (MVS) algorithms [13, 47, 62] reconstruct high-quality 3D geometry by matching feature correspondences across different RGB images and optimizing photometric consistency. While this pipeline is very robust in real-world environments, it misses the surface of low-textured areas [62]. Additionally, the resulting mesh generated by post-processing like Poisson reconstruction [27] heavily depends on the quality of matching, and the accumulated errors in correspondence matching often cause severe artifacts and missing surfaces. Because of the cumulative error from the above pipeline, the texture mapping process [53, 69] leads to undesirable results. Reconstructing high-fidelity texture and 3D geometry of real-world 3D objects remains an open challenge. + +Main Results: In this paper, we present a practical method to capture a high-quality textured mesh of a 3D object in the wild, shown in Fig. 1. We first develop a smartphone app based on ARKit [1] to collect images, LiDAR depths, and poses. Although the smartphone provides quite valid pose estimates, acquiring fine detail geometry and realis- + +![](images/b8eb00b627d49c096dc461da7311abe5d33a0aff3fa1fbc9de7de57ab66ebde8.jpg) +Figure 2. Example objects collected by a smartphone in the wild + +tic texture requires highly accurate camera poses. Thus, we present an RGBD-aided Structure from Motion (SfM) which combines the advantages of both VIO [12] and SfM [46]. Since ARKit based on VIO is robust to the degradation of visual information such as low-textured surface, we perform incremental triangulation with initial poses obtained from ARKit. We also propose a depth filtering method to handle noisy depth from the smartphone. These filtered depth points are exploited as an additional depth factor for bundle adjustment. Our RGBD-aided SfM can estimate highly precise camera poses due to the good initial poses and additional constraints from the filtered depth. + +Our 3D geometry reconstruction process adopts a neural implicit representation [50, 55] for surface reconstruction with volumetric rendering. We observe that the neural implicit representation can perform more complete and smoother reconstruction than the classical methods which are known to be robust for 3D reconstruction in the wild. Furthermore, we introduce a new training method for neural implicit representations. In the early stage of training, we propose a regularization method that leverages the prior information from the classical MVS method. After obtaining the decent shape, we avoid using the prior information and generate sparse voxel octree to enable efficient sampling for training. Our training method can improve the performance of the neural implicit representations. Consequently, we show the generalization capabilities of our surface reconstruction method in real world objects collected by the smartphone. + +Given a mesh extracted from the trained neural implicit representation, we can run classical texture mapping algorithms [53, 69] to generate texture maps that often exhibit blurring artifacts and color misalignment. We propose applying differential rendering [31] to fine-tune these texture maps obtained from classical texture mapping via a photometric loss. Compared to classical 3D reconstruction [23, 47, 62] and texture mapping [53, 69] approaches, our method shows a better ability to reconstruct the 3D model and produce realistic textures. We evaluate our approach on the data collected by our smartphone application. The main contributions of this paper are summarized as follows: + +- We present a unified framework to reconstruct the textured mesh using a smartphone. +- We propose a depth filtering scheme for noisy depths and refine initial poses from ARKit by using bundle adjustment with depth factor. + +- Our pipeline builds on classical 3D reconstruction and texture mapping. We leverage a neural geometry representation to enable surface reconstruction of complex shapes and a differentiable rendering to generate high-fidelity texture maps. + +# 2. Related Work + +Classical 3D Reconstruction Traditional Structure from Motion (SfM) systems [46, 59] have been popular as a vision-based 3D reconstruction due to their robustness to various scenarios [48, 49, 58]. The SfM systems extract and match features [36] of adjacent images of the same scene and generate the camera poses and sparse 3D points via triangulation [20]. Then, Multi-view Stereo (MVS) algorithms [13, 52] compute the dense depth maps of each scene with multiple calibrated images. PatchMatch-based MVS [14, 47, 61, 62] methods are popular because they are proper for depth and normal optimization. After that, surface reconstruction methods [26, 27, 30] can be applied to reconstruct meshes from point clouds. However, all these methods often fail to reconstruct the 3D geometry of low-textured areas. RGB-D Reconstruction techniques fuse many overlapping depth maps into a single 3D model. Starting from the classical TSDF-fusion [8], several methods [9, 23, 42, 57] produce voxel-based TSDF with commodity-level depth cameras such as Kinect. However, these methods have difficulty handling outliers and thin geometry due to noisy depth measurements. Due to the noise and limited range of the LiDAR sensor in smartphone, RGB-D reconstruction methods fail to estimate the camera odometry. + +3D Reconstruction with differentiable rendering Recently, NeRF [38] and its variants [37, 39, 65, 68] have used volumetric implicit representations for rendering and learn $\alpha$ -compositing of a radiance field along rays. These methods yield remarkable quality of novel view synthesis results and do not need ground truth masks. However, 3D geometry from volumetric representations is far from satisfactory due to their volumetric properties [68]. Some methods [41, 64, 67] reconstruct the surface by encoding the network as an implicit function such as a signed distance function (SDF) or occupancy function. Many hybrid approaches [4, 10, 43, 55, 63] unify surface and volumetric representations by regarding surface as defining volumes near the surface, enabling accurate surface reconstruction without masks. Their follow-up works [19, 34, 54, 67] import prior information from the learning-based network. We avoid exploiting prior hints obtained from learning-based models because they are expensive to acquire new training datasets and time-consuming to train other neural networks. Classical Texture Optimization The matching between texture and geometry is a significant step for building realistic 3D models. Lempitsky et al. [32] apply the pairwise Markov Random Field to choose an optimal view for each + +face as texture. Waechter et al. [53] present a global color adjustment to mitigate the visual seams between face textures. Zhou et al. [69] optimize the camera poses using a color consistency measurement and geometric error via local image warping to obtain sharp texture. Bi et al. [6] employ a patch-based image synthesis to produce a texture per each face to optimize the camera pose and geometric error. + +Texture Learning Recently, learning-based methods [17, 22, 25, 51] have been proposed to generate textures for a 3D model. Huang et al. [22] introduce adversarial learning for texture generation in RGB-D scans. Thies et al. [51] store neural textures for a known mesh with given UV mapping and decode them to color for novel views. Some works [7, 60] learn neural texture with neural volumetric rendering. Differentiable rendering methods [16, 21, 33, 40] usually focus on estimating surface radiometric properties from images. These methods can learn to predict not only texture but also geometry via image loss with ground truth masks. Unfortunately, differentiable rendering methods heavily rely on pixel-level accurate masks. Acquiring accurate ground-truth masks in the real-world environment is a difficult task [29]. Therefore, we apply the differentiable rendering to fine-tune texture maps with a fixed mesh. + +# 3. Our Method + +Given images, low-resolution depths and corresponding camera poses from ARKit, our goal is to reconstruct both the texture and geometry of real-world 3D objects. In Fig. 4, our pipeline consists of three stages: RGBD-aided Structure from Motion (SfM), Geometry Reconstruction, and Texture Optimization. Since both depth maps and camera poses provided by ARKit are very noisy, we construct a SfM system to filter out depth maps and refine poses. We then formalize our task as training a neural model for zero-level set of a signed distance function (SDF) guided by classical 3D reconstruction. After training, we extract the mesh with Marching Cubes [35] and apply classical texture reconstruction [53] to generate initial texture maps. Next, we exploit differentiable rendering [31] to finetune the texture maps with the supervision of multi-view images. + +# 3.1. RGBD-aided Structure from Motion + +The goal of this section is to refine initial poses from ARKit by performing a bundle adjustment with depth maps from the smartphone. Unlike simple structures like walls or floors, 3D reconstruction for capturing a high level of detail is required to refine inaccurate poses from ARKit. + +Preprocessing We introduce two processes to obtain reliable depth estimates: single-view filtering and multi-view filtering. Single-view filtering utilizes a confidence map provided by ARKit, which has the same resolution as the depth map. Each pixel in the confidence map has 0, 1, or 2 with the value 0 as least confident and 2 serving as most + +![](images/ebf140c23abdec5ee01cfc3e5266ef65db80bbc5da8e39351c5ba4a0fc589bf1.jpg) +Figure 3. Overview of our pipeline. It is comprised of three steps: RGBD-aided Structure from Motion (Sec. 3.1), Geometry Reconstruction (Sec. 3.2), and Texture Optimization (Sec. 3.3). Our method builds on classical 3D reconstruction and texture mapping. To overcome the limitations of classical methods, we apply a neural geometry representation and neural texture fine-tuning to update 3D geometry and texture maps. + +confident with respect to depth-value accuracy. We choose depth pixels corresponding to 1 or 2 in the confidence map. A main concept of the multi-view filtering is to remove outliers in terms of the reprojection error. Given the reference image $I_{ref}$ with depth $D_{ref}$ , the source image $I_{src}$ with depth $D_{src}$ , the depth map $D_{src}'$ can be transformed from the source view to the reference view with accurate pixel-level correspondence. Then, the depth of pixel $p$ in the reference image will be considered as an inlier if the difference $\delta_{depth}$ between $D_{src}$ and $D_{src}'$ is smaller than a small constant value $\epsilon$ . Then, assuming a set of neighboring N source images, we compute the number of inliers which fulfills $\delta_{depth} < \epsilon$ for each source image, and keep depth pixels if the number of inliers is greater than a certain threshold. Finally, we apply both single-view and multi-view filtering to remove all depth outliers. + +Pose Refinement We consider camera poses obtained from ARKit as initial poses and exploit matching correspondences to perform incremental triangulation [46]. After combining filtered depth points from LiDAR and 3D points from incremental triangulation, we propose a depth factor and run bundle adjustment with the depth factor and reprojection error. Let $P_{i}$ denote the camera parameter of the $i$ -th camera and $X_{j}$ denote point parameters seen by the $i$ -th camera. We minimize the proposed joint objective function $E$ , which is formulated as follows: + +$$ +E = \sum_ {i} \sum_ {j} \rho_ {i} \left(\left\| \phi \left(P _ {i}, X _ {j}\right) - x _ {i j} \right\| ^ {2}\right) + \frac {1}{\eta} \left| D _ {i} - z _ {i} \right|, \tag {1} +$$ + +where $\Phi$ is the function that projects points to the image plane, and $\rho$ denotes the Cauchy function as the robust loss function. $x_{ij}$ is the feature point corresponding to the 3D point $X_{j}$ seen by the $i$ -th camera. $z_{i}$ is the depth value of points projected from the 3D point $X_{j}$ and $D_{i}$ is the filtered depth of pixels corresponding to $z_{i}$ . We define $\eta$ as sensor measurement noise and use it to down-weight noisy depth. We notice that the depth factor has two benefits. First, additional information from the depth sensor is a useful con + +strain for accurate pose estimation and reliable triangulation. Furthermore, this depth factor guides the scale of the SfM to follow the metric scale of the depth sensor. + +# 3.2. Geometry Reconstruction + +Neural Geometry Representations Instead of classical 3D reconstruction, we adopt NeuS [55], which combines the advantages of both volume and surface rendering. In other words, it can extract accurate surfaces even without pixel-level masks. Our scanned object is represented by two Multi-layer Perceptrons (MLPs): geometry network $f_{sdf}$ , which maps a 3D point $x \in \mathbb{R}^3$ to a signed distance, and color network $f_{color}$ which converts a 3D point $x$ and a viewing direction $v \in \mathbb{R}^3$ to a color $c$ . The 3D geometry of an object is represented implicitly as an SDF with zero-level set $\{x \in \mathbb{R}^3 | f_{sdf}(x) = 0\}$ . These two parameters of geometry and color are learned by the volume rendering technique. For a ray $\{x(t) = o + tv | t \geq 0\}$ with camera center $o$ and viewing direction $d$ , the color is rendered with $n$ sampled points along the ray as follows: + +$$ +\hat {C} = \sum_ {i = 1} ^ {n} T _ {i} \alpha_ {i} c (x (t), v), \tag {2} +$$ + +where $T_{i} = \prod_{j=1}^{i-1}(1 - \alpha_{j})$ is the accumulated transmittance, and $\alpha_{i}$ is the discrete opacity. Here, $\alpha_{i} = 1 - \exp(-\int_{t_{i}}^{t_{i+1}} \rho(t) dt)$ , and $\rho(t)$ is opaque density function borrowed from the definition in NeuS [55]. + +Training Process RGBD-aided SfM already estimates both poses and sparse depths by matching feature correspondences. Given this prior information, we leverage an MVS algorithm [62] to estimate the depth and normal for all images with less cost. It is worth noting that we avoid using learning-based MVS algorithms because they are not robust to diverse real-world scenes and are hard to evaluate reliability. In this section, we modify the training process of NeuS to avoid losing fine details of complex shape. Our training process is divided into two stages. In the first stage, we train a neural radiance field with a prior depth and normal obtained from the MVS [62] algorithm to supervise the training process. Specifically, we sample $K$ pixels $p_k$ and their corresponding color values $I(p_k)$ in each iteration. Here, we propose a regularization method that relies on prior depth $D$ and normal $N$ from MVS algorithm. We unproject depth under the given pose to obtain 3D point $X(p_k)$ . Then, we use this 3D point and its corresponding normal $N(p_k)$ from MVS to regularize the geometry network by minimizing + +$$ +L _ {r e g} = \sum_ {k} w _ {d} \left| f _ {s d f} \left(X \left(p _ {k}\right)\right) \right| + \tag {3} +$$ + +$$ +w _ {n} (1 - \langle \nabla_ {p} f _ {s d f} (X (p _ {k})), N (p _ {k}) \rangle), +$$ + +where $w_{d}$ and $w_{n}$ are weights for each term. This regular- + +![](images/fa4d083e001c0a421528df5dc4081def1d09d3bf1740c46cbde594ef330ced3c.jpg) +Figure 4. Overview of Geometry Reconstruction (Section 3.2) and Texture Optimization (Section 3.3). Our goal in geometry reconstruction is to train geometry network $f_{sdf}$ and color network $f_{color}$ via volumetric rendering with two training steps. In step 1, we leverage prior geometric information from the MVS algorithm [62]. Then, our method utilizes a sparse voxel octree to efficiently sample points near the surface. In texture optimization, we first utilize a classical texture reconstruction to generate texture maps for a given mesh. Then, our method applies a differentiable rendering to fine-tune texture maps. + +ization will guide the SDF to learn explicit supervision from the MVS algorithm. Although this regularization is helpful for learning prior geometry, it is susceptible to the noisy results derived from the MVS algorithm which tends to lose fine details and texture-less regions. Thus, we will only use this regularization in the first stage. The color of each pixel can be computed by E.q. 4. The color loss $L_{c}$ is defined as + +$$ +L _ {c} = \sum_ {k} | \hat {C} (p _ {k}) - I (p _ {k}) |. \tag {4} +$$ + +The Eikonal loss [18] to regularize the SDF is formulated as + +$$ +L _ {e i k} = \sum_ {k, i} \left(\| \nabla f _ {s d f} \left(p _ {k, i}\right) \| _ {2} - 1\right) ^ {2}. \tag {5} +$$ + +The overall loss function $L$ in the first stage is defined as + +$$ +L = L _ {c} + L _ {e i k} + L _ {r e g}. \tag {6} +$$ + +In the second stage, inspired by [50], we adopt a sparse voxel octree to guide the sampling process and capture fine details. After the first stage, we can extract a triangular mesh from the SDF via the Marching Cube [35] and define a sparse voxel volume based on this mesh. Then, based on the sparse voxel volume, we can avoid redundant point samples outside the sparse voxel and focus on point samples near the surface samples. Here, our total loss is $L = L_{c} + L_{eik}$ which is the same as E.q. 6 except for $L_{reg}$ . To capture fine details, we will not use $L_{reg}$ in E.q. 6. + +Mesh Simplification A highly detailed 3D mesh from the SDF network with marching cube algorithms [35] leads to enormous memory requirements. To obtain a lightweight mesh, we explore popular quadric mesh simplification [15]. Although the mesh simplification generated small holes, the lightweight mesh enables us to reduce the computation + +complexity of the texture mapping process. We set the decimation ratio to $0.4\%$ . + +# 3.3. Texture Optimization + +Classical Texture Reconstruction For a triangular mesh $M$ provided by the previous section, we first utilize Waechter et al. [53] with poses and images across different views to construct a texture map $T_{orig}$ . Compared to the color network from Section 3.2, this classical method [53] can acquire visually realistic texture images. However, it cannot avoid texture bleeding on the boundary of different views and blurring artifacts. + +Texture Fine-tuning To tackle blurring and texture bleeding artifacts, we propose a fine-tuning step by optimizing the texture map. Here, we parameterize a texture map $T$ obtained from classical texture reconstruction. Given a 3D Mesh $M$ with the 2D texture map $T$ and multi-view images $I_{i}$ with the corresponding camera poses $P_{i}$ , we leverage a differentiable renderer $R$ to generate an image $\hat{I}_i$ and depth $\hat{D}_i$ , i.e. $\hat{I}_i, \hat{D}_i = R(M, P_i)$ . A differentiable renderer $R$ [31] performs mesh rasterization and then texture sampling to render an image $\hat{I}_i$ . In other words, we project texture onto the image plane and compare it to the corresponding view of the image. Furthermore, we compare the rendered image with a different view image by minimizing multi-view photometric loss. We use the combination of the L1 and SSIM [56] as the photometric loss $L_{photo}$ in image space between the rendered image $\hat{I}_i$ and the reference image $I_{i}$ as follows: + +$$ +L _ {\text {p h o t o}} = \sum_ {i} (1 - \alpha) \| \hat {I} _ {i} - I _ {i} \| + \frac {\alpha}{2} (1 - S S I M (\hat {I} _ {i}, I _ {i})), \tag {7} +$$ + +where $\alpha$ is set to 0.85. Additionally, the differentiable render estimates depth $\hat{D}_i$ by interpolating $z$ coordinates of each vertex. With the estimated depth $\hat{D}_i$ and intrinsic parameter $K$ , we can generate a synthesized frame $I_i'$ by reprojecting adjacent frames $I_{i'}$ to the current frame $I_i$ with the camera intrinsic matrix $K$ , the estimated depth $\hat{D}_i$ , and the relative pose $P_{i\rightarrow i'}$ . The multi-view photometric loss $L_{multi}$ similar to Eq. 7 is formulated as: + +$$ +I _ {i} ^ {\prime} (p) = I _ {i ^ {\prime}} \langle \pi (K P _ {i \rightarrow i ^ {\prime}} \hat {D} _ {i} (p) K ^ {- 1} \tilde {p}) \rangle , +$$ + +$$ +L _ {m u l t i} = (1 - \alpha) \| I _ {i} ^ {\prime} - I _ {i} \| + \frac {\alpha}{2} (1 - S S I M \left(I _ {i} ^ {\prime}, I _ {i}\right)), \tag {8} +$$ + +where $\tilde{p}$ is the homogeneous coordinate of $p$ , $\pi$ means projection from homogeneous to image coordinates, and $\langle \cdot \rangle$ indicates the bilinear sampling function. We define our total loss $L$ to be + +$$ +L = L _ {\text {p h o t o}} + L _ {\text {m u l t i}}. \tag {9} +$$ + +Finally, the differentiable rendering enables the loss gradients to back-propagate to update the textures parameters. + +Since our method does not require ground truth masks, we avoid considering optimization for the geometry. We observe that joint optimization with geometry and texture fails to converge without the ground truth mask. + +# 4. Experiments + +# 4.1. Experiment Settings + +Data Collection: Our smartphone application, running on an iPhone 13 Pro using ARKit 6, can record various sensor outputs: synchronized images, poses, confidence maps, dense depth maps from the LiDAR scanner, and IMU sensor information. We built two types of datasets: ARKit-video and AR-capture. ARKit-video collects long video sequences similar to datasets constructed from RGBD-SLAM [70]. The second type is collected by our smartphone application based on Object Capture1. AR-capture is a multiview dataset of objects consisting of high-quality 4K images with depth and gravity data. The ARKit-video collects thousands of frames sampled from the video with resolution of $1280 \times 720$ . In contrast, AR-capture mostly contains less than fifty images at $4032 \times 3096$ pixels. We collect 11 objects: robot arm 1, robot arm 2, plant, tree, sofa, bike, recliner chair, cafe stand, delivery robot, office chair, and camera stand. We randomly select $70 \%$ as the training images and utilize the $30 \%$ as test images for evaluating the novel view synthesis. + +DTU Dataset: The DTU dataset [24] is a MVS dataset that provides reference point clouds acquired with laser sensor in the controlled lab. Since AR-capture and ARKit-video do not include 3D point clouds acquired with specialized hardware, we use the DTU dataset to evaluate the performance of geometry reconstruction. We follow the evaluation of NeuS [55] and exploit the chamfer distance as metric. + +Implementation Details: We build the RGBD-aided Structure from Motion based on COLMAP [46], an existing SfM tool. Instead of SIFT [36], our system uses learning-based features [11, 44, 45] for feature extraction and matching to avoid failure in a featureless scene. ARKit offers an interface for accessing LiDAR data as a depth image. Following the work [55], the SDF network $f_{sdf}$ is implemented as an 8-layer MLP a hidden dimension of 256, and geometric initialization [3] for the network parameters. We construct the color network $f_{color}$ as MLPs with 4-layer MLP with hidden dimension of 256. We optimize networks in geometry reconstruction with Adam [28] and an initial learning rate of 5e-4. We train the first step for 60K iterations and the second step for 140K iterations. In terms of objects with complex shape, it takes 240K iterations for second step. In Texture Optimization, we utilize the public github code for classical method [53]. Then, for texture fine-tuning, the Adam optimizer is used with the learning rate for the tex + +![](images/4a43ebff7a14aacbfdb034db53c3b4d42633b3ba52185a5ec5bcdec724b02dd6.jpg) + +![](images/4ad56b6af3de92b010d7992935bb044e5daddeed01fe4da5732f849fac924f3f.jpg) + +![](images/9f774c18e07150992da3856495a45b6501a14dc2394c5380b4011c7269070aaf.jpg) + +![](images/fb34bd42d29220a5516e8de81e61709ebe4e124d0123d23fdff8c1ab0a473028.jpg) + +![](images/e8c4e27ddd1b0d3f7f8b3080cf0298ec5f7ca934f03afbc73b47aef5c5233af2.jpg) +a) + +![](images/009c1c16aeb5192a0b685495636aa5c5e9679853ca7d276c1c6d20cc0670e2ec.jpg) +(b) + +![](images/5adda5009c718b94063d446a781c2dd8ca18b8cd202f3b0722f55f0462b920e9.jpg) +(c) +Figure 5. We show the benefits of RGBD-aided Structure from Motion. (a) illustrates the 3D point clouds obtained from initial depths with initial poses. In (c), we accumulate the 3D point clouds (by using filtered depths and refined poses) from RGBD-aided SfM. For a detailed description, we visualize the zoomed-in blue and red regions using filtered depth. Initial depth maps and poses from ARkit are input to the geometry reconstruction in Section 3.2 and its results are shown in (b). The results of our pipeline in (d) lead to higher quality reconstructions than (b) which does not apply RGBD-aided SfM. + +![](images/3df9ee72ff27b523475683c46fba4da9fa3b5113a6f9c341425abda80d87daa1.jpg) +(d) + +ture parameter. This step takes 3000 epochs to fine-tune the texture parameter. + +# 4.2. Experimental Results + +In Fig. 5, we visualize the effectiveness of RGBD-aided Structure from Motion in Section 3.1 on our dataset. Figure 5 (a) represents the 3D point clouds by initial depth maps and poses obtained from ARKit. Unfortunately, initial sensor data is too noisy to perform an accurate 3D reconstruction. After multi-view depth filtering as preprocessing, zoomed-in regions in (a) show the point clouds by reprojecting filtered depth maps with initial poses. Compared to initial data, multi-view depth filtering can handle noisy depth measurements. Figure 5 (c) provides the results of preprocessing and pose refinement. Pose refinement enables us to globally align 3D reconstruction without noticeable camera drift and with better local quality. For qualitative comparison, we take as input initial sensor data from ARKit and perform geometry reconstruction in Section 3.2 in Fig. 5 (b). Compared to our experimental results in Fig. 5 (d), it is difficult to achieve comparable quality and better reconstruction. Consequently, our proposed RGBD-aided SfM significantly improves geometry reconstruction performance. + +Figure 6 qualitatively compares our method to VolSDF [63], TSDF-fusion [23], and ACMP [62]. We apply screened Poisson surface reconstruction [27] to generate a mesh from point clouds of ACMP. Among all four algorithms, we apply the same pose estimation algorithm from Section 3.1 for fair comparison. TSDF-fusion, which requires sufficient scene overlap, exploits more frames than our method and ACMP. Due to the low-textured regions, + +![](images/1ffa55901bc8c1c33a5eb1ae805465561e29958d53d4ed05c2acd5a9da5b1c85.jpg) + +![](images/8709bba720f3bdecde117196720a3ebd42f76cf3ba0abee0054493f07273f846.jpg) + +![](images/563fe1f9863c6892e2a1056fae096294f805988b566e6b4e03b36d399616994e.jpg) + +![](images/971a1e1073470c46a99f1b2bf5d7ec10d59a9ce7aed7fd95ce4f54b86049dab9.jpg) + +![](images/9ba4a45691db645cb17ed7621ffa844886a6cc7c53fd6767e1d0cbd1162f7844.jpg) + +![](images/f8ca75e22199423afbeeef829e7bafd8f000fdde66ce952bdf4c7dfdb2d859d6.jpg) + +![](images/45ca2ed330a1ffce7b5b955dc9c4de995b0f4cd1c306ecf117057873c655b03c.jpg) + +![](images/dbac8ffdfc7283f16ca8e2d9707d5d5f2ac7702b0cf0b8f3cadd5aaef49ed51d.jpg) + +![](images/87cb4285c683ca52a3d797fb40dd99f3e8ccfea15311de885cf4a8a4d8f5c6fd.jpg) + +![](images/15c44946e0a79b55493008634dab1f6f37c8578258ce2d27e6cbcbe5df8be44e.jpg) + +![](images/6e32b689aa50550d61abeb5195d63483fb27a4ac587f90810b5d21c2c4763c88.jpg) +Scene + +![](images/df29b3ed3de6da4219b148a6f97b48bd06a30eb8c53d7751107860ce4ae080a7.jpg) +Ours + +![](images/c559e72ff6957848b1e2bd828ab434fa464b6ffd9b8ae15662f5c38f74abfc42.jpg) +VolSDF + +![](images/edd37be9f214f70dd8f23839e1226e8ad31466be99db34b8d6d465f3af958481.jpg) +TSDF-Fusion +Figure 6. Reconstruction of our method, VolSDF [63], TSDF-Fusion [23], and a mesh from point clouds of ACMP [62] processed with screened Poisson surface reconstruction [27]. Note that the neural surface reconstruction is more perceptually convincing in the ARKit-video dataset. + +![](images/616ec985860f07769c11792fb9fcc5eebfc904c7354f30847b91985a00070f5e.jpg) +ACMP + +the surface of ACMP is very noisy. TSDF-fusion utilizes depth maps obtained from low-resolution LiDAR. Here, since this low-resolution depth sensor is very noisy, we only use filtered depth maps after preprocessing in Section 3.1. However, this model still cannot reconstruct either thin or challenging structures accurately (e.g., a chair armrest). Both implicit neural surface reconstruction approaches show comparable performance, and both are superior to classical 3D reconstruction. We observe that the difference between neural geometric representations (Ours and VolSDF) is negligible unlike pose quality. It is worth noting that using accurate poses as input significantly improves the 3D reconstruction quality as shown in Fig. 5. + +For qualitative comparison of textured meshes, we compare our proposed method with four different methods: 1) ACMP [62] is a state-of-the-art multi-view stereo algorithm: 2) Color Map Optimization (CMO) [69] is a texture mapping method for RGB-D scanning: 3) Waechter et al. [53] is a large-scale texture reconstruction based on global color adjustment, 4) NeuS [55] is the recent neural surface reconstruction based on volumetric rendering. We reconstruct a 3D mesh from the point cloud of ACMP with Screened Poisson Surface reconstruction [27]. In the case of CMO, since explicit depth maps from low-resolution LiDAR are too noisy, we utilize depth maps rendered by a triangular mesh from our geometry reconstruction step for a fair comparison. With these rendered depth maps and corresponding images, we run a color map optimization for mapping color images onto our reconstructed mesh. To train + +![](images/d4a44161255f1a7f423d620e1645d81a203a3af34e19c6467ae74e872b1a1229.jpg) +Figure 7. Qualitative comparison between (a) Ours, (b) Waechter et al. [53], (c) CMO [69], (d) ACMP [62], and (e) NeuS [55]. Compared to textured mesh (3-6 columns) obtained from different algorithms, the result of our method is perceptually closer to the original scene. Additional qualitative results are added in the supplementary material due to the limited space. + +NeuS, we follow a training process similar to that in Sec. 3.2. We then extract a mesh from the SDF network in NeuS and apply the Blender Smart UV Project tool to obtain a pervertex UV mapping. To fill a texture for this mesh, points densely sampled from triangle mesh with interpolated UV coordinates are fed to the color network, and eventually we can obtain texture maps for the given mesh. + +In Fig. 7, we show the reconstructed texture for novel views. The visual results of our approach are visually sharper and perceptually closer to the ground truth scene. Results of NeuS and CMO exhibit blurring artifacts and + +color misalignment. The reconstructed mesh from ACMP is very noisy due to incorrect geometry. Here, Waechter et al. (Fig. 7 (b)) is the result of classical texture reconstruction in Section 3.3. This method is more photorealistic and sharper than NeuS, CMO, and ACMP. Then, after texture fine-tuning, our method avoids the seams and texture misalignment often seen with Waechter et al. Table 1 reports the quantitative results on two datasets. Our method outperforms other texture optimization methods on all evaluation metrics. It indicates that our method satisfies both the high fidelity of the texture and its conformity to the ground truth. + +
CategoryDataNeuS [55]CMO [69]Waechter et al. [53]Ours
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
robot arm 1Ac20.2920.8890.08318.4730.8910.08319.9090.8720.06020.0490.8710.057
robot arm 2Ac20.7410.8820.08621.1070.8830.09220.1900.8620.06720.6520.8580.062
plantAc16.7480.6420.29518.4620.6270.31617.9210.6350.18619.9700.6910.146
treeAc18.4090.8690.11718.4480.8700.12717.5180.8580.09618.3310.8580.102
sofaAc19.9100.8720.13921.1700.8400.13619.7590.8290.11321.6080.8940.109
bikeAc20.9710.7780.18321.3110.7780.20619.8220.7400.08819.9430.7370.078
MeanAc19.5120.8220.15119.8250.8150.16019.1860.7990.10120.0950.8180.092
recliner chairAv24.9200.8930.09924.7610.8760.11223.6920.8640.05425.0450.8960.056
cafe standAv20.4980.7790.17322.9340.8280.15417.9150.7980.11222.6340.8180.095
delivery robotAv22.4120.8970.12623.5490.9080.12825.2460.9290.07323.4510.9060.086
office chairAv20.7220.8920.09621.1200.8990.10320.3150.8650.09420.9820.8920.083
camera standAv21.9010.9140.06921.8860.9070.07421.1210.8990.05922.2080.9190.060
MeanAv22.0920.8750.11322.8500.8840.11521.6580.8710.07922.8680.8860.076
+ +Table 1. Quantitative comparison on our dataset. Ac and Av denote the AR-catpure dataset and ARKit-video dataset, respectively. We improve the LPIPS metric by $8.9\%$ in AR-capture and by $3.8\%$ in ARKit-video. + +
Methodscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122Mean
COLMAP [47] with trim=70.450.910.370.370.901.000.541.221.080.640.480.590.320.430.450.65
UNISURF [43]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
NeuS [55]1.371.210.730.401.200.700.721.011.160.820.661.690.390.610.510.87
VolSDF [63]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Ours without 2nd step1.041.610.670.640.880.700.621.141.100.890.661.020.500.580.510.84
Ours0.961.370.670.600.880.700.451.141.100.840.661.110.500.580.500.80
+ +Table 2. Quantitative comparison on DTU [24]. Chamfer distance is employed as the evaluation metric. COLMAP is superior to other neural geometry reconstructions. Our training process can improve the NeuS by $8\%$ . + +![](images/694af125746d1c40e1531a1b718526996fbd3a4d6e0c96dd7283463f0d615957.jpg) +Figure 8. Qualitative comparisons of texture optimization algorithms based on differentiable rendering: (a) our method, (b) Lin et al. [33], and (c) Munkberg et al. [40]. + +![](images/f3ed6c897695308feeb73a9babae4f3932d6115695cbf1416298d82bee8d18da.jpg) + +![](images/a6020174dca24b6bcfcc06acb4cedee350156da2f48a96132752e1d597ab4aee.jpg) + +![](images/d38370e4498bf8769ef5e4dbda7025adcaf01f519a17216c7499168fd12f270a.jpg) + +In Fig. 8, we show the results of other texture optimization methods [33,40] based on differentiable rendering [31]. Both methods aim to learn geometry, textures, and lighting parameters. However, they heavily rely on ground truth masks which require additional annotation costs. Without ground truth masks, we note that joint optimization over geometry and texture is unstable and often has negative effects on geometry. For a fair comparison, we therefore provide the fixed mesh topology and train only texture and lighting parameters through differentiable rendering [31]. Since we capture objects in the wild under unknown environmental lighting conditions, it is difficult to disentangle materials and lighting via differentiable rendering. Neither method [33, 40] is able to generate realistic texture maps compared to either our method or classical texture reconstruction approaches. + +Experimental Results on DTU dataset: Unlike our dataset, DTU [24] is captured by specialized hardware in the controlled lab environment. Thus, this dataset provides the ground truth poses and the RGBD-aided SfM is not necessary. As shown in Table 2, the surfaces obtained by + +COLMAP [47] with trimming value 7 achieve high-quality reconstruction. Our method exploits MVS [47] to supervise the training NeuS and outperforms NeuS. For the ablation study, when we only train our network with the first step (ours without 2nd step), its result is not better than our method in some scans such as scan37. In our experiment, sparse voxel octree sampling does not have a significant effect on geometry performance. Since our training process can be applied to any other 3D neural surface reconstruction methods. In the future, we will apply this training process to other recent 3D neural surface reconstruction algorithms [10, 63, 66], which shows better results on DTU. + +# 5. Conclusion, Limitations and Future Work + +In this paper, we have developed a practical framework to reconstruct high-quality textured meshes of 3D objects by using a smartphone. Given initial poses from ARKit, multi-view images, and low-resolution depth maps, our approach can refine initial poses and depth maps from ARKit. We adopt neural geometry representation [55] for surface reconstruction and improve its performance by modifying the training process. Finally, a differentiable rendering is designed for fine-tuning the texture map acquired from the classical texture mapping method. In the future, our goal is to reduce the training time which can be comparable with classical methods. Unfortunately, most neural geometry reconstruction [10, 55, 63, 66] requires a long training time. We would like to reduce the training time in our framework for practical purposes. + +Acknowledgements This research was partly supported by Army Cooperative Agreement W911NF2120076. + +# References + +[1] Apple arkit. https://developer.apple.com/documentation/arkit/. Accessed: 2022-11-09. 1 +[2] Google arcore. https://developers.google.com/ar. Accessed: 2022-11-09. 1 +[3] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5 +[4] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 2 +[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1 +[6] Sai Bi, Nima Khademi Kalantari, and Ravi Ramamoorthi. Patch-based optimization for image-based texture mapping. ACM Trans. Graph., 36(4):106-1, 2017. 3 +[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision (ECCV), 2022. 3 +[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 2 +[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 1, 2 +[10] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 8 +[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR Deep Learning for Visual SLAM Workshop, 2018. 5 +[12] Alex Flint, Oleg Naroditsky, Christopher P Broaddus, Andriy Grygorenko, Stergios Roumeliotis, and Oriel Bergig. Visual-based inertial navigation, Dec. 11 2018. US Patent 10,152,795. 2 +[13] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2 +[14] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. June 2015. 2 + +[15] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 4 +[16] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 3 +[17] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV, page 88-104, Berlin, Heidelberg, 2020. Springer-Verlag. 3 +[18] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4 +[19] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2 +[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2 +[21] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7498-7507, 2020. 3 +[22] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 3 +[23] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 1, 2, 6 +[24] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 5, 8 +[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3 +[26] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2 +[27] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 1, 2, 6 + +[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[29] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3 +[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2 +[31] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5, 8 +[32] Victor Lempitsky and Denis Ivanov. Seamless mosaicing of image-based texture maps. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-6. IEEE, 2007. 2 +[33] Lixiang Lin, Jianke Zhu, and Yisu Zhang. Multiview textured mesh recovery by differentiable rendering. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 3, 8 +[34] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 2 +[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 3, 4 +[36] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. 2, 5 +[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2 +[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2 +[40] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. arXiv:2111.12503, 2021. 3, 8 +[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2 + +[42] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 1, 2 +[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 8 +[44] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019. 5 +[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5 +[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 2, 3, 5 +[47] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 8 +[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2 +[49] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80(2):189-210, 2008. 2 +[50] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 4 +[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Trans. Graph., 38(4), jul 2019. 3 +[52] George Vogiatzis, Philip HS Torr, and Roberto Cipolla. Multi-view stereo via volumetric graph-cuts. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 391-398. IEEE, 2005. 2 +[53] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 1, 2, 3, 5, 6, 7, 8 +[54] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022. 2 +[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 4, 5, 6, 7, 8 + +[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[57] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2 +[58] Aji Resindra Widya, Yusuke Monno, Masatoshi Okutomi, Sho Suzuki, Takuji Gotoda, and Kenji Miki. Stomach 3d reconstruction based on virtual chromoendoscopic image generation. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1848-1852. IEEE, 2020. 2 +[59] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2 +[60] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7119-7128, June 2021. 3 +[61] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[62] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 1, 2, 4, 6, 7 +[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 6, 8 +[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2 +[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2 +[66] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 8 +[67] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2 +[68] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2 +[69] Qian-Yi Zhou and Vladlen Koltun. Color map optimization for 3d reconstruction with consumer depth cameras. ACM + +Transactions on Graphics (ToG), 33(4):1-10, 2014. 1, 2, 3, 6, 7, 8 +[70] Michael Zollhöfer, Patrick Stotko, Andreas Görtlitz, Christian Theobalt, Matthias Nießner, Reinhard Klein, and Andreas Kolb. State of the art on 3d reconstruction with rgb-d cameras. In Computer graphics forum, volume 37, pages 625-652. Wiley Online Library, 2018. 5 \ No newline at end of file diff --git a/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/images.zip b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c9f33a4bcb4e2b18ac74eb81249d7d251eb92ae --- /dev/null +++ b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a43acc2990383027c4b74a8d5480993486e9f24fb18d165945eab8641b9e668 +size 636696 diff --git a/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/layout.json b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0767b5d6a03a9932972d1aba8b5f3580f36bf452 --- /dev/null +++ b/2023/TMO_ Textured Mesh Acquisition of Objects With a Mobile Device by Using Differentiable Rendering/layout.json @@ -0,0 +1,10264 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "text", + "content": "TMO: Textured Mesh Acquisition of Objects with a Mobile Device by using Differentiable Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "spans": [ + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": "Jaehoon Choi " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": " Dongki Jung " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": " Taejae Lee " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": " Sangwook Kim " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": " Youngdong Jung " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": " Dinesh Manocha " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": " Donghwan Lee " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": "NAVER LABS " + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 160, + 508, + 204 + ], + "type": "text", + "content": "University of Maryland" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 257, + 290, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 473 + ], + "type": "text", + "content": "We present a new pipeline for acquiring a textured mesh in the wild with a single smartphone which offers access to images, depth maps, and valid poses. Our method first introduces an RGBD-aided structure from motion, which can yield filtered depth maps and refines camera poses guided by corresponding depth. Then, we adopt the neural implicit surface reconstruction method, which allows for high-quality mesh and develops a new training process for applying a regularization provided by classical multi-view stereo methods. Moreover, we apply a differentiable rendering to fine-tune incomplete texture maps and generate textures which are perceptually closer to the original scene. Our pipeline can be applied to any common objects in the real world without the need for either in-the-lab environments or accurate mask images. We demonstrate results of captured objects with complex shapes and validate our method numerically against existing 3D reconstruction and texture mapping methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 700 + ], + "type": "text", + "content": "Recovering the 3D geometry of objects and scenes is a longstanding challenge in computer vision and is essential to a broad range of applications. Depth sensing technologies range from highly specialized and expensive turn-table 3D scanners and structured-light scanners to commodity depth sensors. More recently, advances in mobile devices have developed a new method for 3D capture of real-world environments with high-resolution imaging and miniaturized LiDAR. Specifically, the modern smartphone such as iPhone 13 Pro are equipped with RGB camera, accelerometer, gyroscope, magnetometer, and LiDAR scanner. These various sensors can provide high-resolution images, low-resolution depth from the LiDAR scanner, and associated camera poses offered by off-the-shelf visual-inertial odometry (VIO) systems such as ARKit [1] and ARCore [2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Today's smartphones offer low-resolution depth maps" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 231, + 369, + 322 + ], + "blocks": [ + { + "bbox": [ + 310, + 231, + 369, + 322 + ], + "lines": [ + { + "bbox": [ + 310, + 231, + 369, + 322 + ], + "spans": [ + { + "bbox": [ + 310, + 231, + 369, + 322 + ], + "type": "image", + "image_path": "415abffe65e52adcfca8c622fe9a069aa2e7915b859afb4c130414484fed2770.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 323, + 345, + 332 + ], + "lines": [ + { + "bbox": [ + 334, + 323, + 345, + 332 + ], + "spans": [ + { + "bbox": [ + 334, + 323, + 345, + 332 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 232, + 420, + 322 + ], + "blocks": [ + { + "bbox": [ + 369, + 232, + 420, + 322 + ], + "lines": [ + { + "bbox": [ + 369, + 232, + 420, + 322 + ], + "spans": [ + { + "bbox": [ + 369, + 232, + 420, + 322 + ], + "type": "image", + "image_path": "d87e713b938a2e0180aca13b15bb5108c6d26f186bb904b0e6e27811456f102a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 323, + 411, + 332 + ], + "lines": [ + { + "bbox": [ + 402, + 323, + 411, + 332 + ], + "spans": [ + { + "bbox": [ + 402, + 323, + 411, + 332 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 447, + 236, + 492, + 318 + ], + "blocks": [ + { + "bbox": [ + 447, + 236, + 492, + 318 + ], + "lines": [ + { + "bbox": [ + 447, + 236, + 492, + 318 + ], + "spans": [ + { + "bbox": [ + 447, + 236, + 492, + 318 + ], + "type": "image", + "image_path": "eae39edf8883c9c41473a42704556ab8d563187a9117edaadc148e5bef3a0f5e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 323, + 475, + 332 + ], + "lines": [ + { + "bbox": [ + 466, + 323, + 475, + 332 + ], + "spans": [ + { + "bbox": [ + 466, + 323, + 475, + 332 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 494, + 236, + 541, + 319 + ], + "blocks": [ + { + "bbox": [ + 494, + 236, + 541, + 319 + ], + "lines": [ + { + "bbox": [ + 494, + 236, + 541, + 319 + ], + "spans": [ + { + "bbox": [ + 494, + 236, + 541, + 319 + ], + "type": "image", + "image_path": "64932692b2dde51a95e569123e6968423e566b687d6ac549099e26b4bf8fdc6a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 323, + 522, + 332 + ], + "lines": [ + { + "bbox": [ + 512, + 323, + 522, + 332 + ], + "spans": [ + { + "bbox": [ + 512, + 323, + 522, + 332 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 334, + 545, + 390 + ], + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 390 + ], + "type": "text", + "content": "Figure 1. Example reconstruction results collected from a smartphone in the wild. (a) Data acquisition setup. (b) Images captured from a smartphone. (c) A reconstructed mesh. (d) A novel view of textured mesh. Our proposed method can reconstruct the high-quality geometric mesh with a visually realistic texture." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 401, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 640 + ], + "type": "text", + "content": "[5] and valid poses. However, depth maps are very noisy and suffer from the limited range of depth sensors. Although this depth sensor can build a simple 3D structure such as a wall or floor, it cannot reconstruct objects with complex and varying shapes. Thus, the RGBD scanning methods [9, 23, 42, 57] are not suitable for these objects. Instead of the depth sensor, multi-view stereo (MVS) algorithms [13, 47, 62] reconstruct high-quality 3D geometry by matching feature correspondences across different RGB images and optimizing photometric consistency. While this pipeline is very robust in real-world environments, it misses the surface of low-textured areas [62]. Additionally, the resulting mesh generated by post-processing like Poisson reconstruction [27] heavily depends on the quality of matching, and the accumulated errors in correspondence matching often cause severe artifacts and missing surfaces. Because of the cumulative error from the above pipeline, the texture mapping process [53, 69] leads to undesirable results. Reconstructing high-fidelity texture and 3D geometry of real-world 3D objects remains an open challenge." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Main Results: In this paper, we present a practical method to capture a high-quality textured mesh of a 3D object in the wild, shown in Fig. 1. We first develop a smartphone app based on ARKit [1] to collect images, LiDAR depths, and poses. Although the smartphone provides quite valid pose estimates, acquiring fine detail geometry and realis-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16674" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 283, + 125 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 283, + 125 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 283, + 125 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 283, + 125 + ], + "type": "image", + "image_path": "b8eb00b627d49c096dc461da7311abe5d33a0aff3fa1fbc9de7de57ab66ebde8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 128, + 285, + 138 + ], + "lines": [ + { + "bbox": [ + 49, + 128, + 285, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 128, + 285, + 138 + ], + "type": "text", + "content": "Figure 2. Example objects collected by a smartphone in the wild" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 148, + 286, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 286, + 290 + ], + "type": "text", + "content": "tic texture requires highly accurate camera poses. Thus, we present an RGBD-aided Structure from Motion (SfM) which combines the advantages of both VIO [12] and SfM [46]. Since ARKit based on VIO is robust to the degradation of visual information such as low-textured surface, we perform incremental triangulation with initial poses obtained from ARKit. We also propose a depth filtering method to handle noisy depth from the smartphone. These filtered depth points are exploited as an additional depth factor for bundle adjustment. Our RGBD-aided SfM can estimate highly precise camera poses due to the good initial poses and additional constraints from the filtered depth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 292, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 286, + 495 + ], + "type": "text", + "content": "Our 3D geometry reconstruction process adopts a neural implicit representation [50, 55] for surface reconstruction with volumetric rendering. We observe that the neural implicit representation can perform more complete and smoother reconstruction than the classical methods which are known to be robust for 3D reconstruction in the wild. Furthermore, we introduce a new training method for neural implicit representations. In the early stage of training, we propose a regularization method that leverages the prior information from the classical MVS method. After obtaining the decent shape, we avoid using the prior information and generate sparse voxel octree to enable efficient sampling for training. Our training method can improve the performance of the neural implicit representations. Consequently, we show the generalization capabilities of our surface reconstruction method in real world objects collected by the smartphone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 638 + ], + "type": "text", + "content": "Given a mesh extracted from the trained neural implicit representation, we can run classical texture mapping algorithms [53, 69] to generate texture maps that often exhibit blurring artifacts and color misalignment. We propose applying differential rendering [31] to fine-tune these texture maps obtained from classical texture mapping via a photometric loss. Compared to classical 3D reconstruction [23, 47, 62] and texture mapping [53, 69] approaches, our method shows a better ability to reconstruct the 3D model and produce realistic textures. We evaluate our approach on the data collected by our smartphone application. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 646, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 646, + 286, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 646, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 58, + 646, + 286, + 669 + ], + "type": "text", + "content": "- We present a unified framework to reconstruct the textured mesh using a smartphone." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "content": "- We propose a depth filtering scheme for noisy depths and refine initial poses from ARKit by using bundle adjustment with depth factor." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 545, + 131 + ], + "type": "text", + "content": "- Our pipeline builds on classical 3D reconstruction and texture mapping. We leverage a neural geometry representation to enable surface reconstruction of complex shapes and a differentiable rendering to generate high-fidelity texture maps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 143, + 392, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 392, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 392, + 155 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 163, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 437 + ], + "type": "text", + "content": "Classical 3D Reconstruction Traditional Structure from Motion (SfM) systems [46, 59] have been popular as a vision-based 3D reconstruction due to their robustness to various scenarios [48, 49, 58]. The SfM systems extract and match features [36] of adjacent images of the same scene and generate the camera poses and sparse 3D points via triangulation [20]. Then, Multi-view Stereo (MVS) algorithms [13, 52] compute the dense depth maps of each scene with multiple calibrated images. PatchMatch-based MVS [14, 47, 61, 62] methods are popular because they are proper for depth and normal optimization. After that, surface reconstruction methods [26, 27, 30] can be applied to reconstruct meshes from point clouds. However, all these methods often fail to reconstruct the 3D geometry of low-textured areas. RGB-D Reconstruction techniques fuse many overlapping depth maps into a single 3D model. Starting from the classical TSDF-fusion [8], several methods [9, 23, 42, 57] produce voxel-based TSDF with commodity-level depth cameras such as Kinect. However, these methods have difficulty handling outliers and thin geometry due to noisy depth measurements. Due to the noise and limited range of the LiDAR sensor in smartphone, RGB-D reconstruction methods fail to estimate the camera odometry." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "3D Reconstruction with differentiable rendering Recently, NeRF [38] and its variants [37, 39, 65, 68] have used volumetric implicit representations for rendering and learn " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "-compositing of a radiance field along rays. These methods yield remarkable quality of novel view synthesis results and do not need ground truth masks. However, 3D geometry from volumetric representations is far from satisfactory due to their volumetric properties [68]. Some methods [41, 64, 67] reconstruct the surface by encoding the network as an implicit function such as a signed distance function (SDF) or occupancy function. Many hybrid approaches [4, 10, 43, 55, 63] unify surface and volumetric representations by regarding surface as defining volumes near the surface, enabling accurate surface reconstruction without masks. Their follow-up works [19, 34, 54, 67] import prior information from the learning-based network. We avoid exploiting prior hints obtained from learning-based models because they are expensive to acquire new training datasets and time-consuming to train other neural networks. Classical Texture Optimization The matching between texture and geometry is a significant step for building realistic 3D models. Lempitsky et al. [32] apply the pairwise Markov Random Field to choose an optimal view for each" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16675" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "face as texture. Waechter et al. [53] present a global color adjustment to mitigate the visual seams between face textures. Zhou et al. [69] optimize the camera poses using a color consistency measurement and geometric error via local image warping to obtain sharp texture. Bi et al. [6] employ a patch-based image synthesis to produce a texture per each face to optimize the camera pose and geometric error." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 288, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 347 + ], + "type": "text", + "content": "Texture Learning Recently, learning-based methods [17, 22, 25, 51] have been proposed to generate textures for a 3D model. Huang et al. [22] introduce adversarial learning for texture generation in RGB-D scans. Thies et al. [51] store neural textures for a known mesh with given UV mapping and decode them to color for novel views. Some works [7, 60] learn neural texture with neural volumetric rendering. Differentiable rendering methods [16, 21, 33, 40] usually focus on estimating surface radiometric properties from images. These methods can learn to predict not only texture but also geometry via image loss with ground truth masks. Unfortunately, differentiable rendering methods heavily rely on pixel-level accurate masks. Acquiring accurate ground-truth masks in the real-world environment is a difficult task [29]. Therefore, we apply the differentiable rendering to fine-tune texture maps with a fixed mesh." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 357, + 127, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 127, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 127, + 369 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 557 + ], + "type": "text", + "content": "Given images, low-resolution depths and corresponding camera poses from ARKit, our goal is to reconstruct both the texture and geometry of real-world 3D objects. In Fig. 4, our pipeline consists of three stages: RGBD-aided Structure from Motion (SfM), Geometry Reconstruction, and Texture Optimization. Since both depth maps and camera poses provided by ARKit are very noisy, we construct a SfM system to filter out depth maps and refine poses. We then formalize our task as training a neural model for zero-level set of a signed distance function (SDF) guided by classical 3D reconstruction. After training, we extract the mesh with Marching Cubes [35] and apply classical texture reconstruction [53] to generate initial texture maps. Next, we exploit differentiable rendering [31] to finetune the texture maps with the supervision of multi-view images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 241, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 241, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 241, + 574 + ], + "type": "text", + "content": "3.1. RGBD-aided Structure from Motion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "The goal of this section is to refine initial poses from ARKit by performing a bundle adjustment with depth maps from the smartphone. Unlike simple structures like walls or floors, 3D reconstruction for capturing a high level of detail is required to refine inaccurate poses from ARKit." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Preprocessing We introduce two processes to obtain reliable depth estimates: single-view filtering and multi-view filtering. Single-view filtering utilizes a confidence map provided by ARKit, which has the same resolution as the depth map. Each pixel in the confidence map has 0, 1, or 2 with the value 0 as least confident and 2 serving as most" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 155 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 155 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 155 + ], + "type": "image", + "image_path": "ebf140c23abdec5ee01cfc3e5266ef65db80bbc5da8e39351c5ba4a0fc589bf1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 159, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 237 + ], + "type": "text", + "content": "Figure 3. Overview of our pipeline. It is comprised of three steps: RGBD-aided Structure from Motion (Sec. 3.1), Geometry Reconstruction (Sec. 3.2), and Texture Optimization (Sec. 3.3). Our method builds on classical 3D reconstruction and texture mapping. To overcome the limitations of classical methods, we apply a neural geometry representation and neural texture fine-tuning to update 3D geometry and texture maps." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": "confident with respect to depth-value accuracy. We choose depth pixels corresponding to 1 or 2 in the confidence map. A main concept of the multi-view filtering is to remove outliers in terms of the reprojection error. Given the reference image " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " with depth " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "D_{ref}" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": ", the source image " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "I_{src}" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " with depth " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "D_{src}" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": ", the depth map " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "D_{src}'" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " can be transformed from the source view to the reference view with accurate pixel-level correspondence. Then, the depth of pixel " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " in the reference image will be considered as an inlier if the difference " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\delta_{depth}" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "D_{src}" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "D_{src}'" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " is smaller than a small constant value " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": ". Then, assuming a set of neighboring N source images, we compute the number of inliers which fulfills " + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\delta_{depth} < \\epsilon" + }, + { + "bbox": [ + 304, + 247, + 546, + 437 + ], + "type": "text", + "content": " for each source image, and keep depth pixels if the number of inliers is greater than a certain threshold. Finally, we apply both single-view and multi-view filtering to remove all depth outliers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": "Pose Refinement We consider camera poses obtained from ARKit as initial poses and exploit matching correspondences to perform incremental triangulation [46]. After combining filtered depth points from LiDAR and 3D points from incremental triangulation, we propose a depth factor and run bundle adjustment with the depth factor and reprojection error. Let " + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": " denote the camera parameter of the " + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": "-th camera and " + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": " denote point parameters seen by the " + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": "-th camera. We minimize the proposed joint objective function " + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 439, + 545, + 559 + ], + "type": "text", + "content": ", which is formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 567, + 545, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 567, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 310, + 567, + 545, + 596 + ], + "type": "interline_equation", + "content": "E = \\sum_ {i} \\sum_ {j} \\rho_ {i} \\left(\\left\\| \\phi \\left(P _ {i}, X _ {j}\\right) - x _ {i j} \\right\\| ^ {2}\\right) + \\frac {1}{\\eta} \\left| D _ {i} - z _ {i} \\right|, \\tag {1}", + "image_path": "dd1e02804653ac682a0a1de8bb8b3d248d5053bc6704088900cfe4470fcd3565.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " is the function that projects points to the image plane, and " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " denotes the Cauchy function as the robust loss function. " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " is the feature point corresponding to the 3D point " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " seen by the " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "-th camera. " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " is the depth value of points projected from the 3D point " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " is the filtered depth of pixels corresponding to " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " as sensor measurement noise and use it to down-weight noisy depth. We notice that the depth factor has two benefits. First, additional information from the depth sensor is a useful con" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16676" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "strain for accurate pose estimation and reliable triangulation. Furthermore, this depth factor guides the scale of the SfM to follow the metric scale of the depth sensor." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 114, + 190, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 190, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 190, + 127 + ], + "type": "text", + "content": "3.2. Geometry Reconstruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": "Neural Geometry Representations Instead of classical 3D reconstruction, we adopt NeuS [55], which combines the advantages of both volume and surface rendering. In other words, it can extract accurate surfaces even without pixel-level masks. Our scanned object is represented by two Multi-layer Perceptrons (MLPs): geometry network " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": ", which maps a 3D point " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " to a signed distance, and color network " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " which converts a 3D point " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " to a color " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": ". The 3D geometry of an object is represented implicitly as an SDF with zero-level set " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "\\{x \\in \\mathbb{R}^3 | f_{sdf}(x) = 0\\}" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": ". These two parameters of geometry and color are learned by the volume rendering technique. For a ray " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "\\{x(t) = o + tv | t \\geq 0\\}" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " with camera center " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": ", the color is rendered with " + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 133, + 288, + 313 + ], + "type": "text", + "content": " sampled points along the ray as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 319, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 319, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 319, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\hat {C} = \\sum_ {i = 1} ^ {n} T _ {i} \\alpha_ {i} c (x (t), v), \\tag {2}", + "image_path": "ca59e528f604de05fc781ff43e724d906f8965390b9f4b41e718ae5926d9dbc4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "inline_equation", + "content": "T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "text", + "content": " is the accumulated transmittance, and " + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "text", + "content": " is the discrete opacity. Here, " + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\exp(-\\int_{t_{i}}^{t_{i+1}} \\rho(t) dt)" + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\rho(t)" + }, + { + "bbox": [ + 47, + 357, + 287, + 408 + ], + "type": "text", + "content": " is opaque density function borrowed from the definition in NeuS [55]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": "Training Process RGBD-aided SfM already estimates both poses and sparse depths by matching feature correspondences. Given this prior information, we leverage an MVS algorithm [62] to estimate the depth and normal for all images with less cost. It is worth noting that we avoid using learning-based MVS algorithms because they are not robust to diverse real-world scenes and are hard to evaluate reliability. In this section, we modify the training process of NeuS to avoid losing fine details of complex shape. Our training process is divided into two stages. In the first stage, we train a neural radiance field with a prior depth and normal obtained from the MVS [62] algorithm to supervise the training process. Specifically, we sample " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": " pixels " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": " and their corresponding color values " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "I(p_k)" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": " in each iteration. Here, we propose a regularization method that relies on prior depth " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": " from MVS algorithm. We unproject depth under the given pose to obtain 3D point " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "X(p_k)" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": ". Then, we use this 3D point and its corresponding normal " + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "inline_equation", + "content": "N(p_k)" + }, + { + "bbox": [ + 46, + 408, + 287, + 648 + ], + "type": "text", + "content": " from MVS to regularize the geometry network by minimizing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 654, + 287, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 654, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 62, + 654, + 287, + 680 + ], + "type": "interline_equation", + "content": "L _ {r e g} = \\sum_ {k} w _ {d} \\left| f _ {s d f} \\left(X \\left(p _ {k}\\right)\\right) \\right| + \\tag {3}", + "image_path": "854a424603e32f6ed01fc414b07a2b036279a2bbbb928c199c5fc22fc2db577f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 681, + 260, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 681, + 260, + 696 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 260, + 696 + ], + "type": "interline_equation", + "content": "w _ {n} (1 - \\langle \\nabla_ {p} f _ {s d f} (X (p _ {k})), N (p _ {k}) \\rangle),", + "image_path": "ca67f335e8fc575017cbd1c05ec4283898db41d119a2d026d5ec2e6efa69ac80.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_{d}" + }, + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 50, + 701, + 287, + 714 + ], + "type": "text", + "content": " are weights for each term. This regular-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 75, + 541, + 182 + ], + "blocks": [ + { + "bbox": [ + 314, + 75, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 314, + 75, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 75, + 541, + 182 + ], + "type": "image", + "image_path": "fa4d083e001c0a421528df5dc4081def1d09d3bf1740c46cbde594ef330ced3c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "type": "text", + "content": "Figure 4. Overview of Geometry Reconstruction (Section 3.2) and Texture Optimization (Section 3.3). Our goal in geometry reconstruction is to train geometry network " + }, + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "type": "text", + "content": " and color network " + }, + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 304, + 186, + 545, + 297 + ], + "type": "text", + "content": " via volumetric rendering with two training steps. In step 1, we leverage prior geometric information from the MVS algorithm [62]. Then, our method utilizes a sparse voxel octree to efficiently sample points near the surface. In texture optimization, we first utilize a classical texture reconstruction to generate texture maps for a given mesh. Then, our method applies a differentiable rendering to fine-tune texture maps." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 389 + ], + "type": "text", + "content": "ization will guide the SDF to learn explicit supervision from the MVS algorithm. Although this regularization is helpful for learning prior geometry, it is susceptible to the noisy results derived from the MVS algorithm which tends to lose fine details and texture-less regions. Thus, we will only use this regularization in the first stage. The color of each pixel can be computed by E.q. 4. The color loss " + }, + { + "bbox": [ + 304, + 304, + 545, + 389 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 304, + 304, + 545, + 389 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 396, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 396, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 369, + 396, + 545, + 422 + ], + "type": "interline_equation", + "content": "L _ {c} = \\sum_ {k} | \\hat {C} (p _ {k}) - I (p _ {k}) |. \\tag {4}", + "image_path": "82f54bcf8d895a67e981a9e5a9ca0c30fa77fd1d820f5269f77aa6b9e80953c7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 429, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 450 + ], + "type": "text", + "content": "The Eikonal loss [18] to regularize the SDF is formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 450, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 450, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 352, + 450, + 545, + 475 + ], + "type": "interline_equation", + "content": "L _ {e i k} = \\sum_ {k, i} \\left(\\| \\nabla f _ {s d f} \\left(p _ {k, i}\\right) \\| _ {2} - 1\\right) ^ {2}. \\tag {5}", + "image_path": "22a5178ad307f133ba0ac17f10c18a92ca7d64aee4a6b0b45953dfc692cbece8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 480, + 535, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 535, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 535, + 492 + ], + "type": "text", + "content": "The overall loss function " + }, + { + "bbox": [ + 306, + 480, + 535, + 492 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 306, + 480, + 535, + 492 + ], + "type": "text", + "content": " in the first stage is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 501, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 501, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 376, + 501, + 545, + 514 + ], + "type": "interline_equation", + "content": "L = L _ {c} + L _ {e i k} + L _ {r e g}. \\tag {6}", + "image_path": "8cc7ecf087f8022dd285c54473a178bba92e24d87ee5815bf53db08570d1dcba.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": "In the second stage, inspired by [50], we adopt a sparse voxel octree to guide the sampling process and capture fine details. After the first stage, we can extract a triangular mesh from the SDF via the Marching Cube [35] and define a sparse voxel volume based on this mesh. Then, based on the sparse voxel volume, we can avoid redundant point samples outside the sparse voxel and focus on point samples near the surface samples. Here, our total loss is " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "L = L_{c} + L_{eik}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " which is the same as E.q. 6 except for " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": ". To capture fine details, we will not use " + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 304, + 522, + 545, + 642 + ], + "type": "text", + "content": " in E.q. 6." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "Mesh Simplification A highly detailed 3D mesh from the SDF network with marching cube algorithms [35] leads to enormous memory requirements. To obtain a lightweight mesh, we explore popular quadric mesh simplification [15]. Although the mesh simplification generated small holes, the lightweight mesh enables us to reduce the computation" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16677" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "complexity of the texture mapping process. We set the decimation ratio to " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 171, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 171, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 171, + 117 + ], + "type": "text", + "content": "3.3. Texture Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "type": "text", + "content": "Classical Texture Reconstruction For a triangular mesh " + }, + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "type": "text", + "content": " provided by the previous section, we first utilize Waechter et al. [53] with poses and images across different views to construct a texture map " + }, + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "type": "inline_equation", + "content": "T_{orig}" + }, + { + "bbox": [ + 46, + 121, + 287, + 217 + ], + "type": "text", + "content": ". Compared to the color network from Section 3.2, this classical method [53] can acquire visually realistic texture images. However, it cannot avoid texture bleeding on the boundary of different views and blurring artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": "Texture Fine-tuning To tackle blurring and texture bleeding artifacts, we propose a fine-tuning step by optimizing the texture map. Here, we parameterize a texture map " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " obtained from classical texture reconstruction. Given a 3D Mesh " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " with the 2D texture map " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " and multi-view images " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " with the corresponding camera poses " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": ", we leverage a differentiable renderer " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " to generate an image " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{D}_i" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{I}_i, \\hat{D}_i = R(M, P_i)" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": ". A differentiable renderer " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " [31] performs mesh rasterization and then texture sampling to render an image " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": ". In other words, we project texture onto the image plane and compare it to the corresponding view of the image. Furthermore, we compare the rendered image with a different view image by minimizing multi-view photometric loss. We use the combination of the L1 and SSIM [56] as the photometric loss " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "L_{photo}" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " in image space between the rendered image " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " and the reference image " + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 217, + 288, + 422 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 429, + 287, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 287, + 464 + ], + "type": "interline_equation", + "content": "L _ {\\text {p h o t o}} = \\sum_ {i} (1 - \\alpha) \\| \\hat {I} _ {i} - I _ {i} \\| + \\frac {\\alpha}{2} (1 - S S I M (\\hat {I} _ {i}, I _ {i})), \\tag {7}", + "image_path": "5bd76db34ebfc1083c1ad83d1b722cc45b5a996e6eb3e203ca024b1f2941bd26.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " is set to 0.85. Additionally, the differentiable render estimates depth " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\hat{D}_i" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " by interpolating " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " coordinates of each vertex. With the estimated depth " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\hat{D}_i" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " and intrinsic parameter " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": ", we can generate a synthesized frame " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "I_i'" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " by reprojecting adjacent frames " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "I_{i'}" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " to the current frame " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " with the camera intrinsic matrix " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": ", the estimated depth " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "\\hat{D}_i" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": ", and the relative pose " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "P_{i\\rightarrow i'}" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": ". The multi-view photometric loss " + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "inline_equation", + "content": "L_{multi}" + }, + { + "bbox": [ + 47, + 464, + 287, + 561 + ], + "type": "text", + "content": " similar to Eq. 7 is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 569, + 226, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 569, + 226, + 583 + ], + "spans": [ + { + "bbox": [ + 75, + 569, + 226, + 583 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {\\prime} (p) = I _ {i ^ {\\prime}} \\langle \\pi (K P _ {i \\rightarrow i ^ {\\prime}} \\hat {D} _ {i} (p) K ^ {- 1} \\tilde {p}) \\rangle ,", + "image_path": "8b6f6219d94c623d2d0355ff68d4ceace9764bdef3d44087a3c0bc023805bd09.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 585, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 585, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 585, + 287, + 616 + ], + "type": "interline_equation", + "content": "L _ {m u l t i} = (1 - \\alpha) \\| I _ {i} ^ {\\prime} - I _ {i} \\| + \\frac {\\alpha}{2} (1 - S S I M \\left(I _ {i} ^ {\\prime}, I _ {i}\\right)), \\tag {8}", + "image_path": "d0440929667c674e5426e4675c2c0ebb47bfe180b6ef01e85cba30400db20bde.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "content": " is the homogeneous coordinate of " + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "content": " means projection from homogeneous to image coordinates, and " + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "content": " indicates the bilinear sampling function. We define our total loss " + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 624, + 287, + 671 + ], + "type": "text", + "content": " to be" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 121, + 673, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 673, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 121, + 673, + 287, + 685 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {p h o t o}} + L _ {\\text {m u l t i}}. \\tag {9}", + "image_path": "e19b83fa026c26cdc0cf36fb65b96310df09de5d4366c393434b48a5b4578a05.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Finally, the differentiable rendering enables the loss gradients to back-propagate to update the textures parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": "Since our method does not require ground truth masks, we avoid considering optimization for the geometry. We observe that joint optimization with geometry and texture fails to converge without the ground truth mask." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 131, + 388, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 131, + 388, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 388, + 144 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 151, + 424, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 151, + 424, + 164 + ], + "spans": [ + { + "bbox": [ + 306, + 151, + 424, + 164 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "text", + "content": "Data Collection: Our smartphone application, running on an iPhone 13 Pro using ARKit 6, can record various sensor outputs: synchronized images, poses, confidence maps, dense depth maps from the LiDAR scanner, and IMU sensor information. We built two types of datasets: ARKit-video and AR-capture. ARKit-video collects long video sequences similar to datasets constructed from RGBD-SLAM [70]. The second type is collected by our smartphone application based on Object Capture1. AR-capture is a multiview dataset of objects consisting of high-quality 4K images with depth and gravity data. The ARKit-video collects thousands of frames sampled from the video with resolution of " + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "text", + "content": ". In contrast, AR-capture mostly contains less than fifty images at " + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "inline_equation", + "content": "4032 \\times 3096" + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "text", + "content": " pixels. We collect 11 objects: robot arm 1, robot arm 2, plant, tree, sofa, bike, recliner chair, cafe stand, delivery robot, office chair, and camera stand. We randomly select " + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "inline_equation", + "content": "70 \\%" + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "text", + "content": " as the training images and utilize the " + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "inline_equation", + "content": "30 \\%" + }, + { + "bbox": [ + 304, + 169, + 545, + 396 + ], + "type": "text", + "content": " as test images for evaluating the novel view synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "content": "DTU Dataset: The DTU dataset [24] is a MVS dataset that provides reference point clouds acquired with laser sensor in the controlled lab. Since AR-capture and ARKit-video do not include 3D point clouds acquired with specialized hardware, we use the DTU dataset to evaluate the performance of geometry reconstruction. We follow the evaluation of NeuS [55] and exploit the chamfer distance as metric." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "type": "text", + "content": "Implementation Details: We build the RGBD-aided Structure from Motion based on COLMAP [46], an existing SfM tool. Instead of SIFT [36], our system uses learning-based features [11, 44, 45] for feature extraction and matching to avoid failure in a featureless scene. ARKit offers an interface for accessing LiDAR data as a depth image. Following the work [55], the SDF network " + }, + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "type": "text", + "content": " is implemented as an 8-layer MLP a hidden dimension of 256, and geometric initialization [3] for the network parameters. We construct the color network " + }, + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 304, + 480, + 545, + 696 + ], + "type": "text", + "content": " as MLPs with 4-layer MLP with hidden dimension of 256. We optimize networks in geometry reconstruction with Adam [28] and an initial learning rate of 5e-4. We train the first step for 60K iterations and the second step for 140K iterations. In terms of objects with complex shape, it takes 240K iterations for second step. In Texture Optimization, we utilize the public github code for classical method [53]. Then, for texture fine-tuning, the Adam optimizer is used with the learning rate for the tex" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 512, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 512, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 512, + 713 + ], + "type": "text", + "content": "1https://developer.apple.com/videos/play/wwwdc2021/10076/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16678" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 64, + 119, + 132 + ], + "blocks": [ + { + "bbox": [ + 49, + 64, + 119, + 132 + ], + "lines": [ + { + "bbox": [ + 49, + 64, + 119, + 132 + ], + "spans": [ + { + "bbox": [ + 49, + 64, + 119, + 132 + ], + "type": "image", + "image_path": "4a43ebff7a14aacbfdb034db53c3b4d42633b3ba52185a5ec5bcdec724b02dd6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 69, + 161, + 129 + ], + "blocks": [ + { + "bbox": [ + 126, + 69, + 161, + 129 + ], + "lines": [ + { + "bbox": [ + 126, + 69, + 161, + 129 + ], + "spans": [ + { + "bbox": [ + 126, + 69, + 161, + 129 + ], + "type": "image", + "image_path": "4ad56b6af3de92b010d7992935bb044e5daddeed01fe4da5732f849fac924f3f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 65, + 237, + 131 + ], + "blocks": [ + { + "bbox": [ + 167, + 65, + 237, + 131 + ], + "lines": [ + { + "bbox": [ + 167, + 65, + 237, + 131 + ], + "spans": [ + { + "bbox": [ + 167, + 65, + 237, + 131 + ], + "type": "image", + "image_path": "9f774c18e07150992da3856495a45b6501a14dc2394c5380b4011c7269070aaf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 244, + 69, + 279, + 129 + ], + "blocks": [ + { + "bbox": [ + 244, + 69, + 279, + 129 + ], + "lines": [ + { + "bbox": [ + 244, + 69, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 244, + 69, + 279, + 129 + ], + "type": "image", + "image_path": "fb34bd42d29220a5516e8de81e61709ebe4e124d0123d23fdff8c1ab0a473028.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 133, + 114, + 206 + ], + "blocks": [ + { + "bbox": [ + 52, + 133, + 114, + 206 + ], + "lines": [ + { + "bbox": [ + 52, + 133, + 114, + 206 + ], + "spans": [ + { + "bbox": [ + 52, + 133, + 114, + 206 + ], + "type": "image", + "image_path": "e8c4e27ddd1b0d3f7f8b3080cf0298ec5f7ca934f03afbc73b47aef5c5233af2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 206, + 88, + 215 + ], + "lines": [ + { + "bbox": [ + 83, + 206, + 88, + 215 + ], + "spans": [ + { + "bbox": [ + 83, + 206, + 88, + 215 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 115, + 152, + 168, + 206 + ], + "blocks": [ + { + "bbox": [ + 115, + 152, + 168, + 206 + ], + "lines": [ + { + "bbox": [ + 115, + 152, + 168, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 152, + 168, + 206 + ], + "type": "image", + "image_path": "009c1c16aeb5192a0b685495636aa5c5e9679853ca7d276c1c6d20cc0670e2ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 206, + 146, + 215 + ], + "lines": [ + { + "bbox": [ + 136, + 206, + 146, + 215 + ], + "spans": [ + { + "bbox": [ + 136, + 206, + 146, + 215 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 138, + 232, + 207 + ], + "blocks": [ + { + "bbox": [ + 170, + 138, + 232, + 207 + ], + "lines": [ + { + "bbox": [ + 170, + 138, + 232, + 207 + ], + "spans": [ + { + "bbox": [ + 170, + 138, + 232, + 207 + ], + "type": "image", + "image_path": "5adda5009c718b94063d446a781c2dd8ca18b8cd202f3b0722f55f0462b920e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 207, + 206, + 215 + ], + "lines": [ + { + "bbox": [ + 198, + 207, + 206, + 215 + ], + "spans": [ + { + "bbox": [ + 198, + 207, + 206, + 215 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 220, + 287, + 330 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 330 + ], + "type": "text", + "content": "Figure 5. We show the benefits of RGBD-aided Structure from Motion. (a) illustrates the 3D point clouds obtained from initial depths with initial poses. In (c), we accumulate the 3D point clouds (by using filtered depths and refined poses) from RGBD-aided SfM. For a detailed description, we visualize the zoomed-in blue and red regions using filtered depth. Initial depth maps and poses from ARkit are input to the geometry reconstruction in Section 3.2 and its results are shown in (b). The results of our pipeline in (d) lead to higher quality reconstructions than (b) which does not apply RGBD-aided SfM." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 233, + 153, + 284, + 206 + ], + "blocks": [ + { + "bbox": [ + 233, + 153, + 284, + 206 + ], + "lines": [ + { + "bbox": [ + 233, + 153, + 284, + 206 + ], + "spans": [ + { + "bbox": [ + 233, + 153, + 284, + 206 + ], + "type": "image", + "image_path": "3df9ee72ff27b523475683c46fba4da9fa3b5113a6f9c341425abda80d87daa1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 206, + 263, + 215 + ], + "lines": [ + { + "bbox": [ + 253, + 206, + 263, + 215 + ], + "spans": [ + { + "bbox": [ + 253, + 206, + 263, + 215 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 340, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 364 + ], + "type": "text", + "content": "ture parameter. This step takes 3000 epochs to fine-tune the texture parameter." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 372, + 170, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 170, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 170, + 384 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 390, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 617 + ], + "type": "text", + "content": "In Fig. 5, we visualize the effectiveness of RGBD-aided Structure from Motion in Section 3.1 on our dataset. Figure 5 (a) represents the 3D point clouds by initial depth maps and poses obtained from ARKit. Unfortunately, initial sensor data is too noisy to perform an accurate 3D reconstruction. After multi-view depth filtering as preprocessing, zoomed-in regions in (a) show the point clouds by reprojecting filtered depth maps with initial poses. Compared to initial data, multi-view depth filtering can handle noisy depth measurements. Figure 5 (c) provides the results of preprocessing and pose refinement. Pose refinement enables us to globally align 3D reconstruction without noticeable camera drift and with better local quality. For qualitative comparison, we take as input initial sensor data from ARKit and perform geometry reconstruction in Section 3.2 in Fig. 5 (b). Compared to our experimental results in Fig. 5 (d), it is difficult to achieve comparable quality and better reconstruction. Consequently, our proposed RGBD-aided SfM significantly improves geometry reconstruction performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Figure 6 qualitatively compares our method to VolSDF [63], TSDF-fusion [23], and ACMP [62]. We apply screened Poisson surface reconstruction [27] to generate a mesh from point clouds of ACMP. Among all four algorithms, we apply the same pose estimation algorithm from Section 3.1 for fair comparison. TSDF-fusion, which requires sufficient scene overlap, exploits more frames than our method and ACMP. Due to the low-textured regions," + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 326, + 64, + 362, + 121 + ], + "blocks": [ + { + "bbox": [ + 326, + 64, + 362, + 121 + ], + "lines": [ + { + "bbox": [ + 326, + 64, + 362, + 121 + ], + "spans": [ + { + "bbox": [ + 326, + 64, + 362, + 121 + ], + "type": "image", + "image_path": "1ffa55901bc8c1c33a5eb1ae805465561e29958d53d4ed05c2acd5a9da5b1c85.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 371, + 64, + 402, + 121 + ], + "blocks": [ + { + "bbox": [ + 371, + 64, + 402, + 121 + ], + "lines": [ + { + "bbox": [ + 371, + 64, + 402, + 121 + ], + "spans": [ + { + "bbox": [ + 371, + 64, + 402, + 121 + ], + "type": "image", + "image_path": "8709bba720f3bdecde117196720a3ebd42f76cf3ba0abee0054493f07273f846.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 64, + 441, + 121 + ], + "blocks": [ + { + "bbox": [ + 410, + 64, + 441, + 121 + ], + "lines": [ + { + "bbox": [ + 410, + 64, + 441, + 121 + ], + "spans": [ + { + "bbox": [ + 410, + 64, + 441, + 121 + ], + "type": "image", + "image_path": "563fe1f9863c6892e2a1056fae096294f805988b566e6b4e03b36d399616994e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 449, + 64, + 480, + 121 + ], + "blocks": [ + { + "bbox": [ + 449, + 64, + 480, + 121 + ], + "lines": [ + { + "bbox": [ + 449, + 64, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 449, + 64, + 480, + 121 + ], + "type": "image", + "image_path": "971a1e1073470c46a99f1b2bf5d7ec10d59a9ce7aed7fd95ce4f54b86049dab9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 489, + 64, + 523, + 121 + ], + "blocks": [ + { + "bbox": [ + 489, + 64, + 523, + 121 + ], + "lines": [ + { + "bbox": [ + 489, + 64, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 489, + 64, + 523, + 121 + ], + "type": "image", + "image_path": "9ba4a45691db645cb17ed7621ffa844886a6cc7c53fd6767e1d0cbd1162f7844.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 325, + 121, + 364, + 173 + ], + "blocks": [ + { + "bbox": [ + 325, + 121, + 364, + 173 + ], + "lines": [ + { + "bbox": [ + 325, + 121, + 364, + 173 + ], + "spans": [ + { + "bbox": [ + 325, + 121, + 364, + 173 + ], + "type": "image", + "image_path": "f8ca75e22199423afbeeef829e7bafd8f000fdde66ce952bdf4c7dfdb2d859d6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 365, + 122, + 404, + 173 + ], + "blocks": [ + { + "bbox": [ + 365, + 122, + 404, + 173 + ], + "lines": [ + { + "bbox": [ + 365, + 122, + 404, + 173 + ], + "spans": [ + { + "bbox": [ + 365, + 122, + 404, + 173 + ], + "type": "image", + "image_path": "45ca2ed330a1ffce7b5b955dc9c4de995b0f4cd1c306ecf117057873c655b03c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 406, + 122, + 444, + 173 + ], + "blocks": [ + { + "bbox": [ + 406, + 122, + 444, + 173 + ], + "lines": [ + { + "bbox": [ + 406, + 122, + 444, + 173 + ], + "spans": [ + { + "bbox": [ + 406, + 122, + 444, + 173 + ], + "type": "image", + "image_path": "dbac8ffdfc7283f16ca8e2d9707d5d5f2ac7702b0cf0b8f3cadd5aaef49ed51d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 446, + 122, + 485, + 173 + ], + "blocks": [ + { + "bbox": [ + 446, + 122, + 485, + 173 + ], + "lines": [ + { + "bbox": [ + 446, + 122, + 485, + 173 + ], + "spans": [ + { + "bbox": [ + 446, + 122, + 485, + 173 + ], + "type": "image", + "image_path": "87cb4285c683ca52a3d797fb40dd99f3e8ccfea15311de885cf4a8a4d8f5c6fd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 486, + 122, + 526, + 173 + ], + "blocks": [ + { + "bbox": [ + 486, + 122, + 526, + 173 + ], + "lines": [ + { + "bbox": [ + 486, + 122, + 526, + 173 + ], + "spans": [ + { + "bbox": [ + 486, + 122, + 526, + 173 + ], + "type": "image", + "image_path": "15c44946e0a79b55493008634dab1f6f37c8578258ce2d27e6cbcbe5df8be44e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 328, + 175, + 364, + 285 + ], + "blocks": [ + { + "bbox": [ + 328, + 175, + 364, + 285 + ], + "lines": [ + { + "bbox": [ + 328, + 175, + 364, + 285 + ], + "spans": [ + { + "bbox": [ + 328, + 175, + 364, + 285 + ], + "type": "image", + "image_path": "6e32b689aa50550d61abeb5195d63483fb27a4ac587f90810b5d21c2c4763c88.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 288, + 354, + 294 + ], + "lines": [ + { + "bbox": [ + 337, + 288, + 354, + 294 + ], + "spans": [ + { + "bbox": [ + 337, + 288, + 354, + 294 + ], + "type": "text", + "content": "Scene" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 368, + 175, + 405, + 285 + ], + "blocks": [ + { + "bbox": [ + 368, + 175, + 405, + 285 + ], + "lines": [ + { + "bbox": [ + 368, + 175, + 405, + 285 + ], + "spans": [ + { + "bbox": [ + 368, + 175, + 405, + 285 + ], + "type": "image", + "image_path": "df29b3ed3de6da4219b148a6f97b48bd06a30eb8c53d7751107860ce4ae080a7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 287, + 392, + 294 + ], + "lines": [ + { + "bbox": [ + 379, + 287, + 392, + 294 + ], + "spans": [ + { + "bbox": [ + 379, + 287, + 392, + 294 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 407, + 175, + 444, + 285 + ], + "blocks": [ + { + "bbox": [ + 407, + 175, + 444, + 285 + ], + "lines": [ + { + "bbox": [ + 407, + 175, + 444, + 285 + ], + "spans": [ + { + "bbox": [ + 407, + 175, + 444, + 285 + ], + "type": "image", + "image_path": "c559e72ff6957848b1e2bd828ab434fa464b6ffd9b8ae15662f5c38f74abfc42.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 287, + 436, + 294 + ], + "lines": [ + { + "bbox": [ + 416, + 287, + 436, + 294 + ], + "spans": [ + { + "bbox": [ + 416, + 287, + 436, + 294 + ], + "type": "text", + "content": "VolSDF" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 447, + 175, + 481, + 286 + ], + "blocks": [ + { + "bbox": [ + 447, + 175, + 481, + 286 + ], + "lines": [ + { + "bbox": [ + 447, + 175, + 481, + 286 + ], + "spans": [ + { + "bbox": [ + 447, + 175, + 481, + 286 + ], + "type": "image", + "image_path": "edd37be9f214f70dd8f23839e1226e8ad31466be99db34b8d6d465f3af958481.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 287, + 482, + 294 + ], + "lines": [ + { + "bbox": [ + 449, + 287, + 482, + 294 + ], + "spans": [ + { + "bbox": [ + 449, + 287, + 482, + 294 + ], + "type": "text", + "content": "TSDF-Fusion" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 298, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 353 + ], + "type": "text", + "content": "Figure 6. Reconstruction of our method, VolSDF [63], TSDF-Fusion [23], and a mesh from point clouds of ACMP [62] processed with screened Poisson surface reconstruction [27]. Note that the neural surface reconstruction is more perceptually convincing in the ARKit-video dataset." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 488, + 175, + 526, + 285 + ], + "blocks": [ + { + "bbox": [ + 488, + 175, + 526, + 285 + ], + "lines": [ + { + "bbox": [ + 488, + 175, + 526, + 285 + ], + "spans": [ + { + "bbox": [ + 488, + 175, + 526, + 285 + ], + "type": "image", + "image_path": "616ec985860f07769c11792fb9fcc5eebfc904c7354f30847b91985a00070f5e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 287, + 517, + 294 + ], + "lines": [ + { + "bbox": [ + 499, + 287, + 517, + 294 + ], + "spans": [ + { + "bbox": [ + 499, + 287, + 517, + 294 + ], + "type": "text", + "content": "ACMP" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 365, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 521 + ], + "type": "text", + "content": "the surface of ACMP is very noisy. TSDF-fusion utilizes depth maps obtained from low-resolution LiDAR. Here, since this low-resolution depth sensor is very noisy, we only use filtered depth maps after preprocessing in Section 3.1. However, this model still cannot reconstruct either thin or challenging structures accurately (e.g., a chair armrest). Both implicit neural surface reconstruction approaches show comparable performance, and both are superior to classical 3D reconstruction. We observe that the difference between neural geometric representations (Ours and VolSDF) is negligible unlike pose quality. It is worth noting that using accurate poses as input significantly improves the 3D reconstruction quality as shown in Fig. 5." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "For qualitative comparison of textured meshes, we compare our proposed method with four different methods: 1) ACMP [62] is a state-of-the-art multi-view stereo algorithm: 2) Color Map Optimization (CMO) [69] is a texture mapping method for RGB-D scanning: 3) Waechter et al. [53] is a large-scale texture reconstruction based on global color adjustment, 4) NeuS [55] is the recent neural surface reconstruction based on volumetric rendering. We reconstruct a 3D mesh from the point cloud of ACMP with Screened Poisson Surface reconstruction [27]. In the case of CMO, since explicit depth maps from low-resolution LiDAR are too noisy, we utilize depth maps rendered by a triangular mesh from our geometry reconstruction step for a fair comparison. With these rendered depth maps and corresponding images, we run a color map optimization for mapping color images onto our reconstructed mesh. To train" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16679" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 82, + 473, + 529 + ], + "blocks": [ + { + "bbox": [ + 123, + 82, + 473, + 529 + ], + "lines": [ + { + "bbox": [ + 123, + 82, + 473, + 529 + ], + "spans": [ + { + "bbox": [ + 123, + 82, + 473, + 529 + ], + "type": "image", + "image_path": "d4a44161255f1a7f423d620e1645d81a203a3af34e19c6467ae74e872b1a1229.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 536, + 546, + 569 + ], + "lines": [ + { + "bbox": [ + 46, + 536, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 546, + 569 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison between (a) Ours, (b) Waechter et al. [53], (c) CMO [69], (d) ACMP [62], and (e) NeuS [55]. Compared to textured mesh (3-6 columns) obtained from different algorithms, the result of our method is perceptually closer to the original scene. Additional qualitative results are added in the supplementary material due to the limited space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "type": "text", + "content": "NeuS, we follow a training process similar to that in Sec. 3.2. We then extract a mesh from the SDF network in NeuS and apply the Blender Smart UV Project tool to obtain a pervertex UV mapping. To fill a texture for this mesh, points densely sampled from triangle mesh with interpolated UV coordinates are fed to the color network, and eventually we can obtain texture maps for the given mesh." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "In Fig. 7, we show the reconstructed texture for novel views. The visual results of our approach are visually sharper and perceptually closer to the ground truth scene. Results of NeuS and CMO exhibit blurring artifacts and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 578, + 545, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 710 + ], + "type": "text", + "content": "color misalignment. The reconstructed mesh from ACMP is very noisy due to incorrect geometry. Here, Waechter et al. (Fig. 7 (b)) is the result of classical texture reconstruction in Section 3.3. This method is more photorealistic and sharper than NeuS, CMO, and ACMP. Then, after texture fine-tuning, our method avoids the seams and texture misalignment often seen with Waechter et al. Table 1 reports the quantitative results on two datasets. Our method outperforms other texture optimization methods on all evaluation metrics. It indicates that our method satisfies both the high fidelity of the texture and its conformity to the ground truth." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16680" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 70, + 493, + 203 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 493, + 203 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 493, + 203 + ], + "type": "table", + "html": "
CategoryDataNeuS [55]CMO [69]Waechter et al. [53]Ours
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
robot arm 1Ac20.2920.8890.08318.4730.8910.08319.9090.8720.06020.0490.8710.057
robot arm 2Ac20.7410.8820.08621.1070.8830.09220.1900.8620.06720.6520.8580.062
plantAc16.7480.6420.29518.4620.6270.31617.9210.6350.18619.9700.6910.146
treeAc18.4090.8690.11718.4480.8700.12717.5180.8580.09618.3310.8580.102
sofaAc19.9100.8720.13921.1700.8400.13619.7590.8290.11321.6080.8940.109
bikeAc20.9710.7780.18321.3110.7780.20619.8220.7400.08819.9430.7370.078
MeanAc19.5120.8220.15119.8250.8150.16019.1860.7990.10120.0950.8180.092
recliner chairAv24.9200.8930.09924.7610.8760.11223.6920.8640.05425.0450.8960.056
cafe standAv20.4980.7790.17322.9340.8280.15417.9150.7980.11222.6340.8180.095
delivery robotAv22.4120.8970.12623.5490.9080.12825.2460.9290.07323.4510.9060.086
office chairAv20.7220.8920.09621.1200.8990.10320.3150.8650.09420.9820.8920.083
camera standAv21.9010.9140.06921.8860.9070.07421.1210.8990.05922.2080.9190.060
MeanAv22.0920.8750.11322.8500.8840.11521.6580.8710.07922.8680.8860.076
", + "image_path": "2456b5439aa3f2575f6df9969c4da4bca2a5f0d159dfaaf5b456a380d4a4041a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 233, + 494, + 288 + ], + "blocks": [ + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on our dataset. Ac and Av denote the AR-catpure dataset and ARKit-video dataset, respectively. We improve the LPIPS metric by " + }, + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "type": "text", + "content": " in AR-capture and by " + }, + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 46, + 205, + 547, + 228 + ], + "type": "text", + "content": " in ARKit-video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 233, + 494, + 288 + ], + "lines": [ + { + "bbox": [ + 96, + 233, + 494, + 288 + ], + "spans": [ + { + "bbox": [ + 96, + 233, + 494, + 288 + ], + "type": "table", + "html": "
Methodscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122Mean
COLMAP [47] with trim=70.450.910.370.370.901.000.541.221.080.640.480.590.320.430.450.65
UNISURF [43]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
NeuS [55]1.371.210.730.401.200.700.721.011.160.820.661.690.390.610.510.87
VolSDF [63]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Ours without 2nd step1.041.610.670.640.880.700.621.141.100.890.661.020.500.580.510.84
Ours0.961.370.670.600.880.700.451.141.100.840.661.110.500.580.500.80
", + "image_path": "970cd86261685b2f1f557f0767f8a93777fc86822a21cb1c77412d9346814464.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 291, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 547, + 315 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on DTU [24]. Chamfer distance is employed as the evaluation metric. COLMAP is superior to other neural geometry reconstructions. Our training process can improve the NeuS by " + }, + { + "bbox": [ + 46, + 291, + 547, + 315 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 291, + 547, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 65, + 323, + 100, + 412 + ], + "blocks": [ + { + "bbox": [ + 65, + 323, + 100, + 412 + ], + "lines": [ + { + "bbox": [ + 65, + 323, + 100, + 412 + ], + "spans": [ + { + "bbox": [ + 65, + 323, + 100, + 412 + ], + "type": "image", + "image_path": "694af125746d1c40e1531a1b718526996fbd3a4d6e0c96dd7283463f0d615957.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 417, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 452 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons of texture optimization algorithms based on differentiable rendering: (a) our method, (b) Lin et al. [33], and (c) Munkberg et al. [40]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 115, + 323, + 149, + 413 + ], + "blocks": [ + { + "bbox": [ + 115, + 323, + 149, + 413 + ], + "lines": [ + { + "bbox": [ + 115, + 323, + 149, + 413 + ], + "spans": [ + { + "bbox": [ + 115, + 323, + 149, + 413 + ], + "type": "image", + "image_path": "f3ed6c897695308feeb73a9babae4f3932d6115695cbf1416298d82bee8d18da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 165, + 323, + 202, + 413 + ], + "blocks": [ + { + "bbox": [ + 165, + 323, + 202, + 413 + ], + "lines": [ + { + "bbox": [ + 165, + 323, + 202, + 413 + ], + "spans": [ + { + "bbox": [ + 165, + 323, + 202, + 413 + ], + "type": "image", + "image_path": "a6020174dca24b6bcfcc06acb4cedee350156da2f48a96132752e1d597ab4aee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 323, + 268, + 413 + ], + "blocks": [ + { + "bbox": [ + 217, + 323, + 268, + 413 + ], + "lines": [ + { + "bbox": [ + 217, + 323, + 268, + 413 + ], + "spans": [ + { + "bbox": [ + 217, + 323, + 268, + 413 + ], + "type": "image", + "image_path": "d38370e4498bf8769ef5e4dbda7025adcaf01f519a17216c7499168fd12f270a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 461, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 653 + ], + "type": "text", + "content": "In Fig. 8, we show the results of other texture optimization methods [33,40] based on differentiable rendering [31]. Both methods aim to learn geometry, textures, and lighting parameters. However, they heavily rely on ground truth masks which require additional annotation costs. Without ground truth masks, we note that joint optimization over geometry and texture is unstable and often has negative effects on geometry. For a fair comparison, we therefore provide the fixed mesh topology and train only texture and lighting parameters through differentiable rendering [31]. Since we capture objects in the wild under unknown environmental lighting conditions, it is difficult to disentangle materials and lighting via differentiable rendering. Neither method [33, 40] is able to generate realistic texture maps compared to either our method or classical texture reconstruction approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Experimental Results on DTU dataset: Unlike our dataset, DTU [24] is captured by specialized hardware in the controlled lab environment. Thus, this dataset provides the ground truth poses and the RGBD-aided SfM is not necessary. As shown in Table 2, the surfaces obtained by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 325, + 547, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 470 + ], + "type": "text", + "content": "COLMAP [47] with trimming value 7 achieve high-quality reconstruction. Our method exploits MVS [47] to supervise the training NeuS and outperforms NeuS. For the ablation study, when we only train our network with the first step (ours without 2nd step), its result is not better than our method in some scans such as scan37. In our experiment, sparse voxel octree sampling does not have a significant effect on geometry performance. Since our training process can be applied to any other 3D neural surface reconstruction methods. In the future, we will apply this training process to other recent 3D neural surface reconstruction algorithms [10, 63, 66], which shows better results on DTU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 486, + 537, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 537, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 537, + 499 + ], + "type": "text", + "content": "5. Conclusion, Limitations and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 508, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 688 + ], + "type": "text", + "content": "In this paper, we have developed a practical framework to reconstruct high-quality textured meshes of 3D objects by using a smartphone. Given initial poses from ARKit, multi-view images, and low-resolution depth maps, our approach can refine initial poses and depth maps from ARKit. We adopt neural geometry representation [55] for surface reconstruction and improve its performance by modifying the training process. Finally, a differentiable rendering is designed for fine-tuning the texture map acquired from the classical texture mapping method. In the future, our goal is to reduce the training time which can be comparable with classical methods. Unfortunately, most neural geometry reconstruction [10, 55, 63, 66] requires a long training time. We would like to reduce the training time in our framework for practical purposes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgements This research was partly supported by Army Cooperative Agreement W911NF2120076." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16681" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Apple arkit. https://developer.apple.com/documentation/arkit/. Accessed: 2022-11-09. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Google arcore. https://developers.google.com/ar. Accessed: 2022-11-09. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "type": "text", + "content": "[3] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 181, + 286, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 286, + 234 + ], + "type": "text", + "content": "[4] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 312 + ], + "type": "text", + "content": "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz, Tal Dimry, Brandon Joffe, Arik Schwartz, and Elad Shulman. ARK-scenes: A diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 346 + ], + "type": "text", + "content": "[6] Sai Bi, Nima Khademi Kalantari, and Ravi Ramamoorthi. Patch-based optimization for image-based texture mapping. ACM Trans. Graph., 36(4):106-1, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 401 + ], + "type": "text", + "content": "[7] Chong Bao and Bangbang Yang, Zeng Junyi, Bao Hujun, Zhang Yinda, Cui Zhaopeng, and Zhang Guofeng. Neumesh: Learning disentangled neural mesh-based implicit field for geometry and texture editing. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 446 + ], + "type": "text", + "content": "[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 448, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 286, + 501 + ], + "type": "text", + "content": "[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 557 + ], + "type": "text", + "content": "[10] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "text", + "content": "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPR Deep Learning for Visual SLAM Workshop, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "type": "text", + "content": "[12] Alex Flint, Oleg Naroditsky, Christopher P Broaddus, Andriy Grygorenko, Stergios Roumeliotis, and Oriel Bergig. Visual-based inertial navigation, Dec. 11 2018. US Patent 10,152,795. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "type": "text", + "content": "[13] Yasutaka Furukawa and Carlos Hernández. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[14] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. June 2015. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[16] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8635-8644, June 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "text", + "content": "[17] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV, page 88-104, Berlin, Heidelberg, 2020. Springer-Verlag. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[18] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 273, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 327 + ], + "type": "text", + "content": "[19] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 358 + ], + "type": "text", + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 404 + ], + "type": "text", + "content": "[21] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7498-7507, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 405, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 469 + ], + "type": "text", + "content": "[22] Jingwei Huang, Justus Thies, Angela Dai, Abhijit Kundu, Chiyu Jiang, Leonidas J Guibas, Matthias Nießner, Thomas Funkhouser, et al. Adversarial texture optimization from rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1559-1568, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 547 + ], + "type": "text", + "content": "[23] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[24] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "text", + "content": "[25] Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "text", + "content": "[26] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[27] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 1, 2, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16682" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[28] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[29] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "text", + "content": "[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 241 + ], + "type": "text", + "content": "[31] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 287, + 285 + ], + "type": "text", + "content": "[32] Victor Lempitsky and Denis Ivanov. Seamless mosaicing of image-based texture maps. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-6. IEEE, 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 331 + ], + "type": "text", + "content": "[33] Lixiang Lin, Jianke Zhu, and Yisu Zhang. Multiview textured mesh recovery by differentiable rendering. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 333, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 333, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 49, + 333, + 287, + 375 + ], + "type": "text", + "content": "[34] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 377, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 410 + ], + "type": "text", + "content": "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 412, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 412, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 412, + 287, + 443 + ], + "type": "text", + "content": "[36] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91-110, 2004. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 446, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 446, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 49, + 446, + 287, + 511 + ], + "type": "text", + "content": "[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 514, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 287, + 565 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 568, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 610 + ], + "type": "text", + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 657 + ], + "type": "text", + "content": "[40] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. arXiv:2111.12503, 2021. 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 659, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 659, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 659, + 287, + 712 + ], + "type": "text", + "content": "[41] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 172 + ], + "type": "text", + "content": "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 174, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 216 + ], + "type": "text", + "content": "[44] Jerome Revaud, Philippe Weinzaepfel, César De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2d2: repeatable and reliable detector and descriptor. arXiv preprint arXiv:1906.06195, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[45] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 273, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 315 + ], + "type": "text", + "content": "[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 317, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 360 + ], + "type": "text", + "content": "[47] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 1, 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "text", + "content": "[49] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80(2):189-210, 2008. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[50] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "text", + "content": "[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Trans. Graph., 38(4), jul 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 505, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 557 + ], + "type": "text", + "content": "[52] George Vogiatzis, Philip HS Torr, and Roberto Cipolla. Multi-view stereo via volumetric graph-cuts. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 391-398. IEEE, 2005. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[53] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In European conference on computer vision, pages 836-850. Springer, 2014. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 605, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 605, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 605, + 545, + 647 + ], + "type": "text", + "content": "[54] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 648, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 648, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 648, + 545, + 712 + ], + "type": "text", + "content": "[55] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16683" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "text", + "content": "[57] Thomas Whelan, Stefan Leutenegger, Renato Salas-Moreno, Ben Glocker, and Andrew Davison. Elasticfusion: Dense slam without a pose graph. Robotics: Science and Systems, 2015. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 228 + ], + "type": "text", + "content": "[58] Aji Resindra Widya, Yusuke Monno, Masatoshi Okutomi, Sho Suzuki, Takuji Gotoda, and Kenji Miki. Stomach 3d reconstruction based on virtual chromoendoscopic image generation. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1848-1852. IEEE, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 261 + ], + "type": "text", + "content": "[59] Changchang Wu. Towards linear-time incremental structure from motion. In 2013 International Conference on 3D Vision-3DV 2013, pages 127-134. IEEE, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "type": "text", + "content": "[60] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7119-7128, June 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "text", + "content": "[61] Qingshan Xu, Weihang Kong, Wenbing Tao, and Marc Pollefeys. Multi-scale geometric consistency guided and planar prior assisted multi-view stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "text", + "content": "[62] Qingshan Xu and Wenbing Tao. Planar prior assisted patch-match multi-view stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12516-12523, 2020. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "type": "text", + "content": "[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "text", + "content": "[66] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[67] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "text", + "content": "[68] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[69] Qian-Yi Zhou and Vladlen Koltun. Color map optimization for 3d reconstruction with consumer depth cameras. ACM" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 152 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 95 + ], + "type": "text", + "content": "Transactions on Graphics (ToG), 33(4):1-10, 2014. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "text", + "content": "[70] Michael Zollhöfer, Patrick Stotko, Andreas Görtlitz, Christian Theobalt, Matthias Nießner, Reinhard Klein, and Andreas Kolb. State of the art on 3d reconstruction with rgb-d cameras. In Computer graphics forum, volume 37, pages 625-652. Wiley Online Library, 2018. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16684" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_content_list.json b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71b296e1297e7b2d27ae3b085deaefb47ca886f5 --- /dev/null +++ b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_content_list.json @@ -0,0 +1,1613 @@ +[ + { + "type": "text", + "text": "TOPLight: Lightweight Neural Networks with Task-Oriented Pretraining for Visible-Infrared Recognition", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Yu $^{1}$ , Xu Cheng $^{1*}$ , Wei Peng $^{2}$", + "bbox": [ + 336, + 202, + 629, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Nanjing University of Information Science and Technology, China $^{2}$ Department of Psychiatry and Behavioral Sciences, Stanford University", + "bbox": [ + 101, + 220, + 867, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuhao,xcheng}@nuist.edu.cn，wepeng@stanford.edu", + "bbox": [ + 272, + 258, + 696, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visible-infrared recognition (VI recognition) is a challenging task due to the enormous visual difference across heterogeneous images. Most existing works achieve promising results by transfer learning, such as pretraining on the ImageNet, based on advanced neural architectures like ResNet and ViT. However, such methods ignore the negative influence of the pretrained colour prior knowledge, as well as their heavy computational burden makes them hard to deploy in actual scenarios with limited resources. In this paper, we propose a novel task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. Specifically, the TOPLight method simulates the domain conflict and sample variations with the proposed fake domain loss in the pretraining stage, which guides the network to learn how to handle those difficulties, such that a more general modality-shared feature representation is learned for the heterogeneous images. Moreover, an effective fine-grained dependency reconstruction module (FDR) is developed to discover substantial pattern dependencies shared in two modalities. Extensive experiments on VI person re-identification and VI face recognition datasets demonstrate the superiority of the proposed TOPLight, which significantly outperforms the current state of the arts while demanding fewer computational resources.", + "bbox": [ + 75, + 340, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Identity recognition technologies have provided numerous reliable solutions for monitoring systems, which strive to match the face (face recognition [6, 7]) or pedestrian (person re-identification [42]) images of the same identity. However, the majority of previous efforts only consider visible images. In real-life practice, many surveillance cameras can switch to infrared imaging mode at night. Thus, the essential cross-modality visible-infrared recognition (VI", + "bbox": [ + 75, + 756, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54bc120915c24eb7770ad9b0e317de0c86acda0d77e85122cb490881979aa087.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 311, + 867, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b0ae237ab44d55111eac04d4705e779f55e6177646d15303a0badc10d16678b.jpg", + "image_caption": [ + "Figure 1. (a) The task-oriented pretraining strategy; (b) Performance comparison of the standard ImageNet-1k pretraining scheme (SP) and the proposed task-oriented pretraining scheme (TOP) on the SYSU-MM01 dataset [35] (All-Search mode)." + ], + "image_footnote": [], + "bbox": [ + 517, + 416, + 869, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "recognition) technology has been developed to match the visible and infrared photographs of the same people.", + "bbox": [ + 498, + 598, + 890, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, visible-infrared person re-identification (VI-ReID) [26, 38] and visible-infrared face recognition [8, 13, 44] have been widely studied. The key issue is identifying the modality-shared patterns. To this end, several works [29, 30] use generative adversarial networks (GANs) to implement cross-modality alignment at the pixel and feature levels. Others [4, 26, 38] design the dual-path feature extraction network, coupled with inter-feature constraints, to close the embedding space of two modalities. However, these methods utilize at least one pretrained ResNet-50 [12] backbone to extract solid features, which makes them unsuitable for edge monitoring devices. Recent works [4, 38] employ auxiliary models (e.g., pose estimation, graph reasoning) to relieve the modality discrepancy, which enhances the performance on academic benchmarks but reduces the real-time inference speed. Compared with conventional deep networks (e.g., ResNet, ViT), lightweight networks [11, 15, 24] can extract basal features rapidly. In VI recogni", + "bbox": [ + 496, + 628, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author (Email: xcheng@nuist.edu.cn)", + "bbox": [ + 94, + 886, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3541", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion tasks, however, the vast modality discrepancy renders the performance of lightweight networks significantly inferior to that of conventional deep networks. The main reason is that lightweight networks lack the ability to identify modality-shared patterns from heterogeneous images.", + "bbox": [ + 75, + 90, + 467, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we present an effective task-oriented pretraining (TOP) strategy. As shown in Fig. 1(a), we first train a lightweight network on the ImageNet-1k dataset to learn vision prior knowledge. After that, the trained network is transformed into the dual-path network and further trained by using task-oriented data augmentations, identity consistency loss and fake domain loss on the ImageNet-mini dataset [18]. The task-oriented pretraining (TOP) strategy simulates the sample differences in VI scenes and teaches the network how to represent and embed discrepant features. Fig. 1(b) reports the performance of three lightweight networks in the VI-ReID task. Compared with the ImageNet-1k pretraining, our TOP strategy can remarkably improve the baseline performance.", + "bbox": [ + 75, + 167, + 467, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another weakness of lightweight networks is that few feature maps are learned from raw images for rapid inference. In the VI recognition scene, it is challenging to discover modality-shared patterns with so few learned feature maps. In practice, the network can focus on a group of aggregated modality-specific patterns that offer the most gradient for identity classification. In contrast, the fine-grained and modality-shared patterns, which are crucial for achieving robust cross-modality matching, are neglected.", + "bbox": [ + 75, + 378, + 467, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above observations, we present a novel fine-grained dependency reconstruction (FDR) module to help lightweight networks learn modality-shared and fine-grained patterns. Specifically, inspired by the horizontal slice scheme [1], we first slice feature maps horizontally and vertically to extract fine-grained patterns from diversified local regions. Then, the original spatial dependencies of these patterns are eliminated by using pooling operations. Further, the cross-modality dependencies are built by using up-sampling layers to amplify the modality-shared parts from these patterns. At last, to avoid overfitting, the shuffle attention is designed to re-weight the channel dependencies of all the feature maps, which spreads attention to local patterns as much as possible. In general, the major contributions of this paper can be summarized as follows.", + "bbox": [ + 75, + 515, + 467, + 739 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an effective task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. To the best of our knowledge, it is the first work to develop a paradigm for VI recognition on edge devices with an extremely low computation budget.", + "- An effective task-oriented pretraining strategy is proposed to enhance the heterogeneous feature learning capacity of lightweight networks with task-oriented augmentations and the proposed fake domain loss." + ], + "bbox": [ + 94, + 753, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A fine-grained dependency reconstruction module is designed to mine cross-modality dependencies.", + "- Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods on mainstream VI-ReID and VI face recognition datasets by a remarkable margin and extremely low complexity." + ], + "bbox": [ + 517, + 90, + 888, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 207, + 638, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visible-Infrared Face Recognition (VI-FR). It aims to match face images across visible and infrared modalities. Initially, handcraft features are designed for VI-FR. In [3,9], authors utilize the Local Radon Binary Pattern (LRBP) as the general representation to perform identity matching. Difference-of-Gaussian (DoG) filtering and Multiscale Block Local Binary Patterns (MB-LBP) are widely adopted [20, 50] to acquire robust representations for VI-FR. Recently, deep learning methods have achieved considerable success in VI-FR, focusing on learning modality-shared features with CNN networks [13, 14, 22] or disentangling the modality-invariant representations with GAN networks [46]. In [37], a dual-generation method is proposed to disentangle modality-invariant patterns from visible and infrared images, which helps relieve the modality discrepancy. After that, Duan et al. [8] proposed a pose-aligned cross-spectral hallucination (PACH) network to eliminate identity-independent patterns at multiple stages.", + "bbox": [ + 496, + 232, + 890, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visible-Infrared Person Re-identification (VI-ReID). It aims to match visible and infrared pedestrian images according to the feature similarity ranking [47]. The zeropadding scheme [35] was first proposed to handle VI-ReID. In the meantime, they established the first large-scale VI-ReID dataset, named SYSU-MM01. Afterwards, a bidirectional center-constrained network [39] was presented to simultaneously optimize intra- and inter-modality discrepancies. To help the network learn modality-invariant patterns, X-modality [19] is introduced to bridge the gap between visible and infrared modality. Recently, a dense alignment learning method [26] was presented to establish the cross-modality feature correspondence at the pixel level. Ye et al. [38] introduced an attentive graph to discover tri-level relations between visible and infrared images. In addition, several GAN-based methods implement cross-modality alignment [29, 30] or modality compensation [31, 45] to help the network learn robust and compact knowledge. However, despite the promising results, these works are too complex to be utilized in the real world.", + "bbox": [ + 496, + 505, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 821, + 633, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview of The Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 845, + 802, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The major pipeline of the proposed TOPLight is illustrated in Fig. 2, which can be summarized as three steps.", + "bbox": [ + 500, + 869, + 888, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3542", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4f675ff8de04960db9699beba0a92bcf39bdd9729b5f57efb858609afa3590a6.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of our method. We take three steps to adapt a lightweight network to the VI recognition task." + ], + "image_footnote": [], + "bbox": [ + 109, + 95, + 843, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step-1. To adapt the lightweight networks to VI recognition scenarios, we construct a dual-path network based on the lightweight model trained on ImageNet-1k (similar in [26]) and further train the dual-path network on ImageNet-mini [18] with task-oriented augmentations. The identity consistency loss $(L_{idc})$ and fake domain loss $(L_{fd})$ are designed to supervise the domain conflict learning process.", + "bbox": [ + 76, + 419, + 468, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step-2. We utilize the uniform soup scheme [34] to unify two pre-embedding block groups. This step aims to tune the structure of the pretrained network suited for VI training, which preserves two separate blocks (Block1-1 and Block1-2) to extract low-level features from two modalities.", + "bbox": [ + 76, + 527, + 468, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step-3. In this stage, the pretrained dual-path network is further trained on VI recognition datasets. A novel fine-grained dependency reconstruction (FDR) module is integrated into the network, which can construct fine-grained pattern dependencies between visible and infrared features.", + "bbox": [ + 76, + 604, + 468, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task-oriented Pretraining Strategy", + "text_level": 1, + "bbox": [ + 76, + 694, + 382, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core motivation of the task-oriented pretraining strategy is to enable lightweight networks to learn prior knowledge related to cross-modality matching during the pretraining stage. Specifically, we first transform the lightweight network trained on the ImageNet-1k dataset into the dual-path network, which consists of two stem blocks, two pre-embedding block groups and one post-embedding block group. Then, the dual-path network is further trained on the ImageNet-mini dataset. During this training, task-oriented data augmentation (DA) is utilized to create visual differences in training samples, which intends to simulate the discrepancy between visible and infrared", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fcbfacdd4e2c6c060f490d69c6ab24edbbb0b9275bddba7c85d5754619973a86.jpg", + "image_caption": [ + "Figure 3. The effects of each data augmentation." + ], + "image_footnote": [], + "bbox": [ + 519, + 417, + 867, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "images and enforce the network to pay more attention to the identity-aware but colour-irrelevant patterns. Thus, we design the generic, colour and texture data augmentation (DAs) to achieve the training process, as shown in Fig. 3.", + "bbox": [ + 498, + 613, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the generic DAs, we increase the overall sample diversity by using random crop, flip and deformation. In the colour DAs, we randomly perform colour jitter, channel shuffle and RGB shift for each sample in the Block1-1 path. The colour DAs are utilized to disturb the strong colour prior knowledge learned from the ImageNet-1k dataset, forcing the network to pay more attention to the structural patterns of visible images (which also exist in infrared images), not just colour patterns. In the texture DAs, we randomly perform compression, defocusing and sharpenGray for each sample in the Block1-2 path, which intends to remove the colour information and simulate the terrible imaging quality under night surveillance conditions. During the TOP, we combine these three DAs to create visual differences that are close to the VI recognition scenes.", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3543", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, in the above manner, the network can lazily learn from one path to avoid embedding visually discrepant samples. Meanwhile, the sample differences generated by augmentations are insufficient to simulate the serious domain conflict in VI recognition scenes. Thus, we develop a fake domain loss to aggravate the variation of mid-level features during the training and motivate the network to discover shared patterns among them.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assume $\\mathbf{x}$ denotes a raw sample from the ImageNet-mini dataset; after augmentations and feature learning, the $\\mathbf{x}_1$ and $\\mathbf{x}_2$ denote the outcome features of Block4-1 and Block4-2, respectively. $\\mathbf{y}$ is the shared ground-truth label. We concatenate $\\mathbf{x}_1$ and $\\mathbf{x}_2$ and feed them into Block5. The output of Block7 is denoted as $\\hat{\\mathbf{x}} = \\{\\hat{\\mathbf{x}}_1, \\hat{\\mathbf{x}}_2\\}$ . Based on the above definitions, the fake domain loss can be defined by:", + "bbox": [ + 75, + 212, + 470, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f d} ^ {1} = L _ {d} \\left(\\mathbf {x} _ {1}, \\mathbf {d} _ {1}\\right) + L _ {d} \\left(G R L \\left(\\hat {\\mathbf {x}} _ {1}\\right), \\mathbf {d} _ {1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 329, + 468, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f d} ^ {2} = L _ {d} \\left(\\mathbf {x} _ {2}, \\mathbf {d} _ {2}\\right) + L _ {d} \\left(G R L \\left(\\hat {\\mathbf {x}} _ {2}\\right), \\mathbf {d} _ {2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 372, + 468, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{d}_1$ and $\\mathbf{d}_2$ are the corresponding pseudo domain labels for $\\mathbf{x}_1$ and $\\mathbf{x}_2$ . We use the domain labels to pretend that $\\mathbf{x}_1$ is from one domain and $\\mathbf{x}_2$ is from the other. $L_{d}$ denotes the domain classification loss implemented with the logistic classifier. $GRL$ is the gradient reversal layer proposed in [10], which aims to reverse the optimization target of the final domain classifier. In this manner, $L_{fd}^{1}$ and $L_{fd}^{2}$ generate two contradictory learning procedures: self-domain constraints are individually set on $\\mathbf{x}_1$ and $\\mathbf{x}_2$ , which enforce them to be representative and specific for the two fake domains we pretended, respectively. The inverted-domain constraints are set on the final features after Block7 ( $\\hat{\\mathbf{x}}_1$ and $\\hat{\\mathbf{x}}_2$ ) by reversing gradients, which urges them to be similar so as to confound the domain classification loss $L_{d}$ .", + "bbox": [ + 75, + 396, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During this procedure, Block(2,3,4)-1 and Block(2,3,4)-2 are trained to extract two types of strongly distinguished \"heterogeneous features\". In comparison, Block(5,6,7) are trained to embed these discrepant features and learn their shared patterns. We manually create this conflict to force the lightweight network to learn how to represent heterogeneous features and discover shared patterns from them.", + "bbox": [ + 75, + 607, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that the embedded features are identity-related, the identity consistency loss $(L_{idc})$ is introduced as:", + "bbox": [ + 75, + 713, + 468, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i d c} = \\log [ P (y | p (\\hat {\\mathbf {x}} _ {1})) ] + \\log [ P (y | p (\\hat {\\mathbf {x}} _ {2})) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 758, + 468, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(\\hat{\\mathbf{x}}_1)$ and $p(\\hat{\\mathbf{x}}_2)$ denote the identity prediction results from the same linear classifier. $\\mathbf{y}$ is the ground-truth label for both $\\mathbf{x}_1$ and $\\mathbf{x}_2$ . Here, we utilize one classifier to predict two conflicted features with shared logits, aiming to further improve the embedding capacity of Block(5,6,7).", + "bbox": [ + 75, + 782, + 468, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall loss in the TOP is written as follows.", + "bbox": [ + 96, + 857, + 419, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {i d c} + L _ {f d} ^ {1} + L _ {f d} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the pretraining, we utilize the uniform soup scheme to unify two groups of pre-embedded blocks in turn (Step-2 in Fig. 2). The final lightweight network for VI-recognition training consists of two independent stem blocks (Block1-1 and 1-2), one pre-embedding block group (Block2,3,4) and one post-embedding block group (Block5,6,7), as shown in the Step-3 of Fig. 2.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Fine-grained Dependency Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 208, + 854, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the task-oriented pretraining, the lightweight network learns how to extract modality-invariant features from heterogeneous images. However, it lacks the ability to discover fine-grained patterns across the modalities. Therefore, we introduce a practical fine-grained dependency reconstruction module to help the lightweight network discover plentiful fine-grained patterns by constructing cross-modality feature dependencies in the VI training phase.", + "bbox": [ + 496, + 233, + 890, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, let us denote $S^v = \\{\\mathbf{x}_i^v | i = 1,2,\\dots,N^v\\}$ and $S^r = \\{\\mathbf{x}_i^r | i = 1,2,\\dots,N^r\\}$ as the visible and infrared training sets, respectively; $N^v$ and $N^r$ indicate the total numbers of visible and infrared samples, respectively. During the training, we feed $\\mathbf{x}^v$ and $\\mathbf{x}^r$ into Block1-1 and Block1-2 to extract low-level features of two modalities, respectively. Then, we concatenate these low-level features and feed them into the Block(2,3,4) for the initial feature embedding. Based on the outcome of Block4 $(\\mathbf{Z} = \\{\\mathbf{Z}^v,\\mathbf{Z}^r\\})$ , the fine-grained and cross-modality feature dependencies are constructed to help the network learn substantial modality-shared patterns by the FDR module.", + "bbox": [ + 496, + 354, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to extract fine-grained patterns from feature maps, we slice $\\mathbf{Z}$ into parts along the horizontal and vertical directions. For each spatial map $\\{\\mathbf{Z}\\}_{m=0}^{C-1} \\in \\mathbb{R}^{H \\times W}$ , the horizontal patterns are obtained by:", + "bbox": [ + 496, + 537, + 890, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} ^ {f} = \\operatorname {C o n c a t} [ \\operatorname {W i n d o w} (\\{\\mathbf {Z} ^ {f} \\} _ {m = 0} ^ {C - 1}) ], f \\in \\{v, r \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 608, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Z}^f\\in \\mathbb{R}^{C\\times H\\times W},f\\in \\{v,r\\}$ denote visible or infrared features. $C,H,W$ are the channel, height, and width dimensions, respectively. $\\mathbf{h}^f\\in \\mathbb{R}^{C\\times \\frac{H}{\\sigma}\\times \\sigma \\times W}$ denotes a group of fine-grained horizontal patterns extracted from $\\mathbf{Z}^f$ Similarly, the vertical patterns $\\mathbf{v}^f\\in \\mathbb{R}^{C\\times \\frac{W}{\\mu}\\times H\\times \\mu}$ can also be obtained by:", + "bbox": [ + 496, + 646, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {f} = \\operatorname {C o n c a t} \\left[ \\operatorname {W i n d o w} \\left(\\left\\{\\mathbf {Z} ^ {f} \\right\\} _ {m = 0} ^ {C - 1}\\right) \\right], f \\in \\{v, r \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 752, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ are used to control the slice granularity.", + "bbox": [ + 496, + 790, + 861, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, the spatial information of $\\mathbf{h}^f$ and $\\mathbf{v}^f$ is condensed by the average pooling operation, which produces $\\hat{\\mathbf{h}}^f\\in \\mathbb{R}^{C\\times \\frac{H}{\\sigma}}$ and $\\hat{\\mathbf{v}}^f\\in \\mathbb{R}^{C\\times \\frac{W}{\\mu}}$ . In the above manner, we refine the original spatial maps $\\{\\mathbf{Z}^f\\}_{m = 0}^{C - 1}\\in \\mathbb{R}^{H\\times W}$ into directional tensors $\\hat{\\mathbf{h}}^f$ and $\\hat{\\mathbf{v}}^f$ to cut off the original self-modality dependencies in spatial dimensions. Furthermore,", + "bbox": [ + 496, + 805, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3544", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the cross-modality dependencies are established as follows.", + "bbox": [ + 76, + 90, + 467, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {h} ^ {f} = \\underset {\\left(\\frac {H}{\\sigma} \\times W\\right)} {\\operatorname {T r a n s c o n v}} (\\hat {\\mathbf {h}} ^ {f}), f \\in \\{v, r \\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 119, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {v} ^ {f} = \\underset {\\left(H \\times \\frac {W}{\\mu}\\right)} {\\operatorname {T r a n s C o n v}} (\\hat {\\mathbf {v}} ^ {f}), f \\in \\{v, r \\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 164, + 468, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (7) and Eq. (8), we use two modality-shared transposed convolution layers to reconstruct the spatial maps of $\\mathbf{h}^f$ and $\\mathbf{v}^f$ , respectively. This process aims to establish the cross-modality spatial dependencies. The outputs are $\\mathbf{z}_h^f \\in \\mathbb{R}^{C \\times H \\times W}$ and $\\mathbf{z}_v^f \\in \\mathbb{R}^{C \\times H \\times W}$ .", + "bbox": [ + 76, + 202, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, there exist \"scheme stereotypes\" in $\\mathbf{Z}_h^f$ and $\\mathbf{Z}_v^f$ . The main reason is that these features are learned from two distinct directional schemes, which may lead to a skewed learning procedure. In contrast, we hope these two schemes can consistently provide diversified patterns that cover more valuable signals. Thus, the shuffle attention is designed to fuse $\\mathbf{Z}_h^f$ and $\\mathbf{Z}_v^f$ while decoupling them from two directions. Specifically, we concatenate them on the channel level and then shuffle the channel order to disarrange the spatial maps learned from two directions.", + "bbox": [ + 76, + 279, + 468, + 433 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} ^ {f} = \\operatorname {S h u f f l e} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {Z} _ {h} ^ {f}, \\mathbf {Z} _ {v} ^ {f}\\right)\\right), f \\in \\{v, r \\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 444, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{D}^f\\in \\mathbb{R}^{2C\\times H\\times W}$ denotes the concatenated feature with disarranged channel order. The new channel dependencies are established by weighting all the spatial maps.", + "bbox": [ + 76, + 474, + 468, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {D}} ^ {f} = \\mathbf {W} _ {2} R e L U \\left(\\mathbf {W} _ {1} G A P \\left(\\mathbf {D} ^ {f}\\right)\\right), f \\in \\{v, r \\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 547, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_1\\in \\mathbb{R}^{2C\\times \\frac{C}{4}}$ and $\\mathbf{W}_2\\in \\mathbb{R}^{\\frac{C}{4}\\times C}$ indicate projection matrices to compute the channel attention [16]. Finally, we introduce $\\hat{\\mathbf{D}}^f$ to refine the original features.", + "bbox": [ + 76, + 579, + 468, + 627 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} ^ {f} = \\hat {\\mathbf {D}} ^ {f} \\oplus \\mathbf {Z} ^ {f}, f \\in \\{v, r \\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 638, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{Z}}^f$ is used as the input of the next block; $\\oplus$ is the element-wise add operation. The proposed fine-grained dependency reconstruction module can effectively establish fine-grained and cross-modality dependencies in both visible and infrared features, which can help discover substantial small patterns that are vital in VI recognition.", + "bbox": [ + 76, + 669, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 772, + 220, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the hard-mining triplet loss $(L_{tri})$ [42] and cross-entropy loss $(L_{id})$ for metric learning and identity learning, respectively. The overall loss can be defined as:", + "bbox": [ + 76, + 797, + 468, + 842 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {i d} + \\gamma * L _ {t r i}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 859, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the parameter to balance each loss term.", + "bbox": [ + 76, + 885, + 413, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 90, + 705, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Settings for VI-ReID", + "text_level": 1, + "bbox": [ + 500, + 114, + 797, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate the proposed method for the visible-infrared person re-identification (VI-ReID) task on two public benchmarks: SYSU-MM01 [35] and RegDB [25].", + "bbox": [ + 498, + 137, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SYSU-MM01 [35] is the largest VI-ReID dataset, which contains 491 persons captured from six different cameras across scenes (indoor and outdoor). There are a total of 287628 visible images and 15792 infrared images. Following the evaluation protocol in [26,38], 395 persons are fixed for training and 96 persons for testing. The testing process includes all-search and indoor-search modes.", + "bbox": [ + 498, + 183, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RegDB [25] contains 412 persons, and each person has 10 visible and 10 infrared images. \"Visible to infrared\" and \"infrared to visible\" are two testing modes. The former uses visible as the query and infrared as the gallery. The latter does the opposite. Following [26, 30, 38], the final performance on the RegDB is made with the average results of 10 times the training and testing procedure. Each time, we randomly sample 206 identities for training, and the remaining 206 identities are used for testing.", + "bbox": [ + 498, + 287, + 890, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use the standard cumulative matching characteristics (CMC), mean average precision (mAP), and mean inverse negative penalty (mINP) [42] as the evaluation metrics.", + "bbox": [ + 498, + 424, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For the ablation studies, we select the MobileNetV3-L [15] as the lightweight baseline to validate each proposed component. As shown in Table 1, we encapsulate all the layers of MobileNetV3-L as Block1-7 and then initialize the first four blocks twice to construct the dual-path network. During the TOP, detailed settings are", + "bbox": [ + 498, + 484, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/65afeb3e414c73fb163027f9b4b5d077984c3438313ecfa9912a6b53c4dafe3a.jpg", + "table_caption": [ + "Table 1. Detailed structures for each Block. We package the entire MobileNetV3-Large into Block1-7 without overlap." + ], + "table_footnote": [], + "table_body": "
Block partitions on MobileNetV3-large
Layer NameStructuresOutput Size
Block1Conv(3×3, 2), Bneck(3×3, 1)16 × 1122
Block2Bneck(3×3, 2)24 × 562
Block3Bneck(3×3, 1)24 × 562
Block4Bneck(5×5, 2), Bneck(5×5, 1)×240 × 282
Block5Bneck(3×3, 2), Bneck(3×3, 1)×5112 × 142
Block6Bneck(5×5, 2), Bneck(5×5, 1)×2160 × 72
Block7Conv(1×1, 1)×31280 × 72
", + "bbox": [ + 503, + 611, + 890, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shown in Table 2(b). It is worth noticing that the \"Random pick\" means that for each image sample used in the Block1-1 path, we randomly pick one method from colour jitter, channel shuffle, or RGB shift to augment it.", + "bbox": [ + 498, + 763, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the VI-ReID training, detailed settings are presented in Table 2(a). Briefly, we randomly sample 6 identities with 6 visible and 6 infrared images per identity, which makes a total of 72 images in each batch. All images are resized to $288 \\times 144$ . Random erasing (RE) [49], flipping,", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3545", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e114a81b415a73e67aeb2910604622dc5cd06c71d2a36c9c98db898b8a85ec2b.jpg", + "table_caption": [ + "Table 2. Detailed training and augmentation settings." + ], + "table_footnote": [], + "table_body": "
(a): Training settings
SettingsOn ImageNet-1KOn ImageNet-miniOn VI-ReID
Epochs1505070
Batch size51236×236×2
Image size224×224224×224288×144
Augs.Stanard [15]Task-orientedRE, CE, Flip
OptimizerSGDSGDAdamW [23]
Lr.0.20.10.001
Decaycosinecosineconstant
Warm upFirst 5 epochNoneFirst 5 epoch
Weight decay5E-45E-51E-2
PlatformPytorch, FP16PytorchPytorch
GPUTesla V100RTX 3060RTX 3060
Cost57 h4 h-
", + "bbox": [ + 81, + 103, + 468, + 263 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9c2d4aafb28c31ea5533c06c99a63a7cc83653b305e742a5f30ca8a72f85698.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(b): Task-oriented augmentation settings
AugmentationsProbabilityTarget
Generic (G)Crop & Filp0.5All path
Deformation0.2
Colour (C)Colour jitterRandom pickBlock1-1 path
Channel shuffle
RGB shift
Texture (T)Compression0.2Block1-2 path
Defocusing0.2
Sharpen gray1
", + "bbox": [ + 81, + 265, + 468, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and channel erasing (CE) [40] are utilized against overfitting. The AdamW [23] optimizer with a learning rate (Lr) of 0.001, and a weight decay of 0.01 is adopted to train our network for 70 epochs. We use the Lr warm-up in the first 5 epochs and constant decay at the $20^{th}$ and $40^{th}$ epochs with a decay factor of 0.1. The parameters $\\mu$ and $\\sigma$ are set to 3 and 2 based on the fine-tuning results, respectively. The balance parameter $\\gamma$ in Eq. (12) is empirically set to $1/2$ .", + "bbox": [ + 75, + 392, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 526, + 230, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Task-oriented Pretraining. We first evaluate the effect of task-oriented pretraining on VI-ReID, as shown in Table 3. No.1 indicates the baseline performance obtained via direct fine-tuning of the ImageNet-1k pretrained model in VI-ReID datasets. When only using generic augmentations and identity loss to perform TOP on ImageNetmini, the \"catastrophic forgetting\" makes No.2 perform worse than the baseline. However, when applying the proposed task-oriented augmentations, the network learns to extract stable features from visually discrepant images in advance, thus improving the downstream VI-ReID performance (No.3). Coupled with $L_{idc}$ and $L_{fd}$ , we create a conflicting learning process that enforces the network to learn how to represent and embed heterogeneous features. No.5 shows significant improvements in both the SYSU-MM01 and RegDB datasets. By only using the MobileNetV3-large backbone, the performance reported in No.5 has already exceeded many ResNet-based works [19, 26, 38].", + "bbox": [ + 75, + 551, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of Fine-grained Dependency Reconstruction Module. In Table 3, No.6-No.9 show the ablation experiments about the fine-grained dependency reconstruction module. We intend to construct the cross-modality feature dependencies, and the first issue is how to model the", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "original feature relations. No.6 indicates directly applying the coarse-grained features to the shuffle attention module. The performance is slightly improved. When using the proposed directional scheme $(H_{s},V_{s})$ to extract fine-grained patterns and reconstruct their spatial relation using transposed convolution layers, No.7 and No.8 show meaningful improvements. The full power of the FDR module is shown in No.9, which adopts both fine-grained horizontal and vertical patterns to discover cross-modality feature similarities, achieving $66.14\\%$ and $63.80\\%$ in terms of Rank-1 and mAP performance on SYSU-MM01 under the all-search mode.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pointing at the difficulties in VI recognition tasks, we design the TOP strategy and FDR module to remedy the drawbacks of the lightweight network. Compared with the baseline (No.1), our method (No.9) considerably improves all metrics on two benchmarks. Meanwhile, the ablations reported in Table 3 demonstrate that all the proposed components bring advantages consistently to the accuracy gain.", + "bbox": [ + 496, + 258, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Applicability with Lightweight Networks. We evaluate our methods on three mainstream lightweight networks (ShuffleNetV2 [24], GhostNet [11] and MobileNetV3 [15]). As shown in Table 4, our method can remarkably improve VI Re-ID performance with ignorable complexity growth. Compared with conventional deep networks, we surpass them significantly with minimal complexity.", + "bbox": [ + 496, + 364, + 890, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detailed Ablations on the FDR Module. We validate the rationality of each component in the FDR module. Table 5(a) shows the impacts of different spatial modelling methods. Except for ours, the HAP [38] achieves the best performance. However, we extract patterns from both the horizontal and vertical regions and use up-sampling (u.) to discover modality-shared cues from re-enlarged spatial areas, which can be regarded as a more effective solution.", + "bbox": [ + 496, + 470, + 890, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5(b) shows the impacts of different channel relation reasoning methods. Our shuffle attention (SA) wins both SE [16] and CBAM [33]. The recipe is the channel shuffle (cs.) operation that builds scheme-decoupled dependencies. This crucial step can evenly correlate visible and infrared features from each fine-grained pattern.", + "bbox": [ + 496, + 592, + 890, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Feature Visualization", + "text_level": 1, + "bbox": [ + 500, + 694, + 700, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What Does TOP Do? We randomly track 10 identities during the TOP stage and visualize the feature distribution via T-SNE [28], as illustrated in Fig. 4.", + "bbox": [ + 496, + 718, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first use task-oriented augmentation to create two types of visual discrepant samples from each image, as shown in Fig. 4(a) and Fig. 4(b). Then, the proposed fake domain loss forces the features after Block4 to be domain-specific, as shown in Fig. 4(c). Concurrently, it forces the network to learn to embed those domain-specific features and provide domain-shared predictions. In Fig. 4(d), after the training, we can observe that the network knows how to extract the domain-shared representations and successfully", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3546", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8b0b05fc4857bf59dd2d0bcfebf3b496fe659b81c383baff117da99d9089aa26.jpg", + "table_caption": [ + "Table 3. Evaluation of each proposed component on two VI-ReID datasets. \"Augs.\" indicates the augmentations. G, C and T denote the generic, colour, and texture augmentations, respectively. In the FDR module, $H_{s}$ and $V_{s}$ denote the horizontal and vertical slices with up-sampling. SA is the shuffle attention module. Rank (r) (\\%), mAP (\\%) and mINP (\\%) are reported." + ], + "table_footnote": [], + "table_body": "
No.Task-oriented pretraining stageVI training stageSYSU-MM01 (all-search)RegDB (visible-to-infrared)
Augs.Loss functionsFDR module
GC+TLidLidcLfdHsVsSAr=1r=10mAPmINPr=1r=10mAPmINP
147.8189.7147.0633.4871.2689.9465.6648.50
2✓✓43.2885.9645.5631.1070.7388.5265.6548.41
3✓✓✓49.8589.7447.5635.5271.3289.9165.6748.49
4✓✓✓54.2892.1152.9441.2975.3192.6468.7852.16
5✓✓✓✓62.4194.1259.0645.1382.7594.1376.2161.84
6✓✓✓✓✓62.8994.2659.7945.8482.8894.1976.2461.92
7✓✓✓✓✓✓63.9595.2860.0946.8083.0794.4876.5562.20
8✓✓✓✓✓✓64.0495.4161.1246.9283.2294.6977.0163.16
9✓✓✓✓✓✓✓66.1496.0363.8049.7684.1594.9879.2663.86
", + "bbox": [ + 80, + 131, + 893, + 303 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d810e5f5ec8c44d9c3fc4d47057ef7a9b4c417c7cf9de012c68f1ba5a9477700.jpg", + "table_caption": [ + "Table 4. Experimental results on different lightweight networks and conventional deep networks." + ], + "table_footnote": [], + "table_body": "
MethodsFLOPs (M)SYSU-MM01RegDB
r=1mAPr=1mAP
ConventionResNet-50356256.9854.7276.8671.30
ConvNeXt-Tiny362058.7255.3178.2572.64
Vit-B568952.1751.8175.3170.37
Swin-Tiny328758.2455.1678.3972.68
LightweightShuffleNetV2-1.0×13941.8841.9467.8364.85
+TOP & FDR17755.7152.6379.8266.36
ShuffleNetV2-1.5×26547.3947.8170.1565.28
+TOP & FDR37163.3560.8184.1376.98
GhostNet-1.0×15042.5342.9471.2864.40
+TOP & FDR18958.5455.1983.2677.16
GhostNet-1.3×28150.8947.9272.5165.98
+TOP & FDR39566.7664.0185.5179.95
MobileNetV3-S10440.9242.5162.7758.31
+TOP & FDR13054.7550.2675.5370.17
MobileNetV3-L25047.8147.0671.2665.66
+TOP & FDR36266.1463.8084.1579.26
", + "bbox": [ + 81, + 340, + 468, + 571 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f6727feac59d6080d3cb37f2a68e155bb579df088b9c8a85f5bf87d670931ce.jpg", + "table_caption": [ + "Table 5. Evaluation of spatial modelling methods and channel relation reasoning methods in the FDR module. \"u.\" and \"cs.\" respectively denote the up-sampling and channel shuffle operations." + ], + "table_footnote": [], + "table_body": "
(a): Impact of different spatial modelling methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
GAP62.8959.7945.8482.8876.2461.92
Context [2]61.3757.5246.0881.5674.4962.11
HAP [38]62.9258.1347.8482.9876.2062.77
HS+VS(w/o u.)63.4559.7549.6183.9478.8263.21
HS+VS66.1463.8049.7684.1579.2663.86
(b): Impact of different channel relation reasoning methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
SE [16]62.9159.7045.7882.7977.4562.34
CBAM [33]61.7956.8845.2583.4177.2862.96
SA (w/o cs.)62.8158.1045.7682.7576.1262.56
SA66.1463.8049.7684.1579.2663.86
", + "bbox": [ + 81, + 619, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "embed all the features into their respective identity groups. Thus, the proposed task-oriented pretraining strategy makes lightweight networks fit cross-modality tasks.", + "bbox": [ + 76, + 821, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pattern Visualization. As demonstrated in Fig. 5, we visualize the learned patterns using Grad-CAM [27] to ex", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2baa24b60df4fee86f2fd1cb76f7bd15817972db5f26cfa5f8570c8b13645c16.jpg", + "image_caption": [ + "Figure 4. Feature distributions during task-oriented pretraining. Each colour denotes an identity. In (a), the star markers represent raw samples. In (b), (c) and (d), the circle and triangle markers represent colour and texture-augmented samples, respectively." + ], + "image_footnote": [], + "bbox": [ + 524, + 313, + 867, + 473 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee5eac419acd4b332838bda4aef0870d29194e08c82b3d3e8ee87875cce38cbc.jpg", + "image_caption": [ + "Figure 5. Visualization of learned patterns via Grad-CAM [27]. For a clear presentation, we average all the spatial maps into four. plain the effect of the FDR module. Clearly, the original patterns extracted by MobileNetV3-L are concentrated and overfit on the modality-specific parts. To remedy this issue, we use the FDR module to discover modality-shared patterns from fine-grained regions, which separates learned patterns to focus on considerable modality-shared cues. A good case is shown in Fig. 5. The visible patterns initially learned are concentrated in the high chest area, while the infrared patterns are all in the right arm. After re-weighting by the FDR module, modality-shared small regions like the face, arms, neck, and shoes are all taken into account, thus" + ], + "image_footnote": [], + "bbox": [ + 532, + 537, + 864, + 703 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3547", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df9d1fd59370fa41a70680fc7c1e467aa79b8ebcb942e71903a6ace465ff34c3.jpg", + "table_caption": [ + "Table 6. Comparison with the state-of-the-arts on SYSU-MM01 [35]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported." + ], + "table_footnote": [], + "table_body": "
DetailsAll-searchIndoor-search
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356214.8054.1271.3315.95-20.5868.3885.7926.92-
JSIA [30]ResNet50+GAN>413338.1080.7089.9036.90-43.8086.2094.2052.90-
AGW [42]ResNet50>356247.5084.3992.1447.6535.3054.1791.1495.9862.9759.23
X-Modal [19]ResNet50>356249.9089.8096.0050.70------
DMiR [38]ResNet50>356250.5488.1294.8649.29-53.9292.5097.0962.49-
FBP-AL [32]ResNet50>356254.1486.0493.0350.20------
DDAG [41]ResNet50>356254.7590.3995.8155.0239.6261.0294.0698.4167.9862.61
HAT [43]ResNet50>356255.2992.1497.3653.89-62.1095.7599.2070.84-
LBA [26]ResNet50>356255.41--54.14-58.46--66.33-
TSME [21]ResNet50>356264.2395.1998.7361.21-64.8096.9299.3171.53-
SPOT [4]ResNet50+ViT>481065.3492.7397.0462.2548.8669.4296.2299.1274.6370.48
TOLight (Ours)MobileNetV3-L= 36266.1496.0397.6863.8049.7672.4197.5499.2376.1171.43
TOLight (Ours)GhostNet-1.3×= 39566.7696.2398.7064.0150.1872.8997.9399.2876.7071.95
", + "bbox": [ + 81, + 103, + 890, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e8e66a564345f2bfc7608e999b789398a2eee7f403c60908ead0ee329c57bfc8.jpg", + "table_caption": [ + "Table 7. Comparison with the state-of-the-arts on RegDB [25]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported." + ], + "table_footnote": [], + "table_body": "
DetailsVisible-to-InfraredInfrared-to-Visible
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356217.7534.2144.3518.90-16.6334.6844.2517.82-
JSIA [30]ResNet50+GAN>413348.50--49.30-48.10--48.90-
AGW [42]ResNet50>356270.0586.2191.5566.3750.1970.4987.2191.8465.9051.24
X-Modal [19]ResNet50>356262.2183.1391.7260.18------
DMiR [38]ResNet50>356275.7989.8694.1869.97-73.9389.8793.9868.22-
FBP-AL [32]ResNet50>356273.9889.7193.6968.24-70.0589.2293.8866.61-
DDAG [41]ResNet50>356269.3486.1991.4963.4649.2468.0685.1590.3161.8048.62
HAT [43]ResNet50>356271.8387.1692.1667.56-70.0268.4591.6166.30-
LBA [26]ResNet50>356274.17--67.64-72.43--65.46-
SPOT [4]ResNet50+ViT>481080.3593.4896.4472.4656.1979.3792.7996.0172.2656.06
GECNet [48]ResNet50+GAN>435082.3392.7295.4978.45-78.9391.9995.4475.58-
TOLight (Ours)MobileNetV3-L=36284.1594.9896.5879.2663.8680.9492.8596.3776.1059.33
TOLight (Ours)GhostNet-1.3×=39585.5194.9996.7079.9563.8580.6592.8196.3275.9159.26
", + "bbox": [ + 81, + 294, + 890, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "improving the accuracy of the cross-modality matching.", + "bbox": [ + 76, + 476, + 444, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Comparison with State of The Arts", + "text_level": 1, + "bbox": [ + 76, + 500, + 382, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our method with recently proposed state-of-the-art VI-ReID methods, as shown in Table 6 and Table 7. They illustrate that we remarkably and efficiently exceed all the compared SOTAs under diverse evaluation settings. Specifically, based on the tiny GhostNet- $1.3\\times$ backbone, we attain the rank-1 of $66.76\\%$ and mAP of $64.01\\%$ on the all-search mode of the large-scale SYSU-MM01 dataset with only 395 FLOPs. The proposed method also achieves excellent results on the RegDB dataset, surpassing the previous SOTAs by a significant margin with lower FLOPs.", + "bbox": [ + 75, + 522, + 467, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Evaluation on two VI-FR datasets. CA is channel augmentation [40]. B is the LightCNN-29 baseline. Rank at 1 accuracy $(\\%)$ and false acceptance rate (F: $\\%$ ) are reported.", + "bbox": [ + 76, + 678, + 467, + 719 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46765ffac46e1a9082cb2252ca95b380e5aeef0782aaaa551e07c1795d729874.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsOulu [5]BUAA [17]
r=1F:1%F:0.1%r=1F:1%F:0.1%
IDR [13]94.373.446.294.393.484.7
VSA [44]99.996.882.398.098.292.5
PACH [8]10097.988.298.698.093.5
B [36]10097.987.098.097.793.7
B+CA [40]10098.991.798.398.294.5
B+TOP10098.891.598.398.194.5
B+TOP+FDR (Ours)10098.991.798.398.294.6
", + "bbox": [ + 81, + 720, + 467, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visible-Infrared Face Recognition", + "text_level": 1, + "bbox": [ + 76, + 847, + 372, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We follow [40] to examine our method on Oulu-CASIA NIR-VIS [5] and BUAA-VisNir face databases [17]. Fol", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lowing previous studies, we also adopt the LightCNN-29 [36] as our baseline network (B). Except for IDR [13], VSA [44], and PACH [8], we also compare recently proposed channel augmentation (CA) [40] with our method on the same baseline. Results are displayed in Table 8.", + "bbox": [ + 496, + 476, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evidently, the proposed TOP and FDR also enhance the performance of VI face recognition. These results further validate the effectiveness and generalization of our methods, which can solve VI recognition tasks effectively.", + "bbox": [ + 496, + 551, + 890, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 630, + 617, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents an effective task-oriented pretrained lightweight neural network (TOPLight) to solve visible-infrared recognition problems. First, the task-oriented pretraining strategy significantly improves the ability of lightweight networks to understand heterogeneous images by introducing domain conflict and sample variation during the pretraining phase. Second, the fine-grained dependency reconstruction module is utilized to destroy the modality-specific pattern dependencies and construct the fine-grained and modality-shared pattern dependencies between visible and infrared features. Extensive experiments demonstrate the superiority and effectiveness of the proposed method as well as the efficacy of each component of our framework.", + "bbox": [ + 496, + 657, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is funded by the National Natural Science Foundation of China (Grant No. 61802058, 61911530397).", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3548", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sun, et.al. Learning part-based convolutional features for person re-identification. TPAMI, 2019. 2", + "[2] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 7", + "[3] Srinivasa Rao Chalamala, Krishna Rao Kakkirala, and Jami Santosh Kumar. Face recognition using spatial pyramid matching and lrbp. In 2014 IEEE 10th International Colloquium on Signal Processing and its Applications, pages 67-70. IEEE, 2014. 2", + "[4] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Jianguo Jiang, and Chia-Wen Lin. Structure-aware positional transformer for visible-infrared person re-identification. IEEE Transactions on Image Processing, 31:2352-2364, 2022. 1, 8", + "[5] Jie Chen, Dong Yi, Jamei Yang, Guoying Zhao, Stan Z Li, and Matti Pietikainen. Learning mappings for face synthesis from near infrared to visual light images. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 156-163. IEEE, 2009. 8", + "[6] Changxing Ding and Dacheng Tao. A comprehensive survey on pose-invariant face recognition. ACM Transactions on intelligent systems and technology (TIST), 7(3):1-42, 2016. 1", + "[7] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 1", + "[8] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7930-7938, 2020. 1, 2, 8", + "[9] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In 2012 19th IEEE International Conference on Image Processing, pages 1837-1840. IEEE, 2012. 2", + "[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 4", + "[11] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 1, 6", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[13] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Learning invariant deep representation for nir-vis face recognition. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 1, 2, 8" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Wasserstein cnn: Learning invariant features for nir-vis face recognition. IEEE transactions on pattern analysis and machine intelligence, 41(7):1761-1773, 2018. 2", + "[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 5, 6", + "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5, 6, 7", + "[17] D Huang, J Sun, and Y Wang. The buaa-visnir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3:3, 2012. 8", + "[18] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 2, 3", + "[19] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4610–4617, 2020. 2, 6, 8", + "[20] Shengcai Liao, Dong Yi, Zhen Lei, Rui Qin, and Stan Z Li. Heterogeneous face recognition from local structures of normalized appearance. In International Conference on Biometrics, pages 209-218. Springer, 2009. 2", + "[21] Jianan Liu, Jialiang Wang, Nianchang Huang, Qiang Zhang, and Jungong Han. Revisiting modality-specific feature compensation for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 8", + "[22] Xiaoxiang Liu, Lingxiao Song, Xiang Wu, and Tieniu Tan. Transferring deep representation for nir-vis heterogeneous face recognition. In 2016 International Conference on Biometrics (ICB), pages 1-8. IEEE, 2016. 2", + "[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6", + "[24] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 1, 6", + "[25] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 17(3):605, 2017. 5, 8", + "[26] Hyunjong Park, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12046–12055, 2021. 1, 2, 3, 5, 6, 8", + "[27] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3549", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7", + "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6", + "[29] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3623-3632, 2019. 1, 2", + "[30] Guan-An Wang, Tianzhu Zhang, Yang Yang, Jian Cheng, Jianlong Chang, Xu Liang, and Zeng-Guang Hou. Cross-modality paired-images generation for rgb-infrared person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12144–12151, 2020. 1, 2, 5, 8", + "[31] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 618–626, 2019. 2", + "[32] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Flexible body partition-based adversarial learning for visible infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 2021. 8", + "[33] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 6, 7", + "[34] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 3", + "[35] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 5380-5389, 2017. 1, 2, 5, 8", + "[36] Xiang Wu, Ran He, Zhenan Sun, and Tieniu Tan. A light cnn for deep face representation with noisy labels. IEEE Transactions on Information Forensics and Security, 13(11):2884-2896, 2018. 8", + "[37] Xiang Wu, Huaibo Huang, Vishal M Patel, Ran He, and Zhenan Sun. Disentangled variational representation for heterogeneous face recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9005–9012, 2019. 2", + "[38] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE Transactions on Information Forensics and Security, 17:386–398, 2021. 1, 2, 5, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE Transactions on Information Forensics and Security, 15:407–419, 2019. 2", + "[40] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13567-13576, 2021. 6, 8", + "[41] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In European Conference on Computer Vision, pages 229–247. Springer, 2020. 8", + "[42] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 5, 8", + "[43] Mang Ye, Jianbing Shen, and Ling Shao. Visible-infrared person re-identification via homogeneous augmented trimodal learning. IEEE Transactions on Information Forensics and Security, 16:728-739, 2020. 8", + "[44] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-posed high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 1, 8", + "[45] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7349–7358, 2022. 2", + "[46] Teng Zhang, Arnold Wiliem, Siqi Yang, and Brian Lovell. Tv-gan: Generative adversarial network based thermal to visible face recognition. In 2018 international conference on biometrics (ICB), pages 174-181. IEEE, 2018. 2", + "[47] Huantao Zheng, Xian Zhong, Wenxin Huang, Kui Jiang, Wenxuan Liu, and Zheng Wang. Visible-infrared person re-identification: A comprehensive survey and a new setting. *Electronics*, 11(3):454, 2022. 2", + "[48] Xian Zhong, Tianyou Lu, Wenxin Huang, Mang Ye, Xuemei Jia, and Chia-Wen Lin. Grayscale enhancement colorization network for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1418–1430, 2021. 8", + "[49] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 5", + "[50] Jun-Yong Zhu, Wei-Shi Zheng, Jian-Huang Lai, and Stan Z Li. Matching nir face to vis face using transduction. IEEE Transactions on Information Forensics and Security, 9(3):501-514, 2014. 2" + ], + "bbox": [ + 501, + 92, + 890, + 852 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3550", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_model.json b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_model.json new file mode 100644 index 0000000000000000000000000000000000000000..53919ad5a3666a90a15c09a4804d70b5b1e3db37 --- /dev/null +++ b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.877, + 0.178 + ], + "angle": 0, + "content": "TOPLight: Lightweight Neural Networks with Task-Oriented Pretraining for Visible-Infrared Recognition" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.203, + 0.63, + 0.222 + ], + "angle": 0, + "content": "Hao Yu\\(^{1}\\), Xu Cheng\\(^{1*}\\), Wei Peng\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.222, + 0.868, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Nanjing University of Information Science and Technology, China \\(^{2}\\)Department of Psychiatry and Behavioral Sciences, Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.26, + 0.697, + 0.275 + ], + "angle": 0, + "content": "{yuhao,xcheng}@nuist.edu.cn，wepeng@stanford.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Visible-infrared recognition (VI recognition) is a challenging task due to the enormous visual difference across heterogeneous images. Most existing works achieve promising results by transfer learning, such as pretraining on the ImageNet, based on advanced neural architectures like ResNet and ViT. However, such methods ignore the negative influence of the pretrained colour prior knowledge, as well as their heavy computational burden makes them hard to deploy in actual scenarios with limited resources. In this paper, we propose a novel task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. Specifically, the TOPLight method simulates the domain conflict and sample variations with the proposed fake domain loss in the pretraining stage, which guides the network to learn how to handle those difficulties, such that a more general modality-shared feature representation is learned for the heterogeneous images. Moreover, an effective fine-grained dependency reconstruction module (FDR) is developed to discover substantial pattern dependencies shared in two modalities. Extensive experiments on VI person re-identification and VI face recognition datasets demonstrate the superiority of the proposed TOPLight, which significantly outperforms the current state of the arts while demanding fewer computational resources." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.732, + 0.21, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Identity recognition technologies have provided numerous reliable solutions for monitoring systems, which strive to match the face (face recognition [6, 7]) or pedestrian (person re-identification [42]) images of the same identity. However, the majority of previous efforts only consider visible images. In real-life practice, many surveillance cameras can switch to infrared imaging mode at night. Thus, the essential cross-modality visible-infrared recognition (VI" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.312, + 0.868, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.417, + 0.87, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Figure 1. (a) The task-oriented pretraining strategy; (b) Performance comparison of the standard ImageNet-1k pretraining scheme (SP) and the proposed task-oriented pretraining scheme (TOP) on the SYSU-MM01 dataset [35] (All-Search mode)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.891, + 0.629 + ], + "angle": 0, + "content": "recognition) technology has been developed to match the visible and infrared photographs of the same people." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Recently, visible-infrared person re-identification (VI-ReID) [26, 38] and visible-infrared face recognition [8, 13, 44] have been widely studied. The key issue is identifying the modality-shared patterns. To this end, several works [29, 30] use generative adversarial networks (GANs) to implement cross-modality alignment at the pixel and feature levels. Others [4, 26, 38] design the dual-path feature extraction network, coupled with inter-feature constraints, to close the embedding space of two modalities. However, these methods utilize at least one pretrained ResNet-50 [12] backbone to extract solid features, which makes them unsuitable for edge monitoring devices. Recent works [4, 38] employ auxiliary models (e.g., pose estimation, graph reasoning) to relieve the modality discrepancy, which enhances the performance on academic benchmarks but reduces the real-time inference speed. Compared with conventional deep networks (e.g., ResNet, ViT), lightweight networks [11, 15, 24] can extract basal features rapidly. In VI recogni" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.388, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author (Email: xcheng@nuist.edu.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "3541" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.168 + ], + "angle": 0, + "content": "tion tasks, however, the vast modality discrepancy renders the performance of lightweight networks significantly inferior to that of conventional deep networks. The main reason is that lightweight networks lack the ability to identify modality-shared patterns from heterogeneous images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.379 + ], + "angle": 0, + "content": "To address this issue, we present an effective task-oriented pretraining (TOP) strategy. As shown in Fig. 1(a), we first train a lightweight network on the ImageNet-1k dataset to learn vision prior knowledge. After that, the trained network is transformed into the dual-path network and further trained by using task-oriented data augmentations, identity consistency loss and fake domain loss on the ImageNet-mini dataset [18]. The task-oriented pretraining (TOP) strategy simulates the sample differences in VI scenes and teaches the network how to represent and embed discrepant features. Fig. 1(b) reports the performance of three lightweight networks in the VI-ReID task. Compared with the ImageNet-1k pretraining, our TOP strategy can remarkably improve the baseline performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.514 + ], + "angle": 0, + "content": "Another weakness of lightweight networks is that few feature maps are learned from raw images for rapid inference. In the VI recognition scene, it is challenging to discover modality-shared patterns with so few learned feature maps. In practice, the network can focus on a group of aggregated modality-specific patterns that offer the most gradient for identity classification. In contrast, the fine-grained and modality-shared patterns, which are crucial for achieving robust cross-modality matching, are neglected." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.468, + 0.741 + ], + "angle": 0, + "content": "Based on the above observations, we present a novel fine-grained dependency reconstruction (FDR) module to help lightweight networks learn modality-shared and fine-grained patterns. Specifically, inspired by the horizontal slice scheme [1], we first slice feature maps horizontally and vertically to extract fine-grained patterns from diversified local regions. Then, the original spatial dependencies of these patterns are eliminated by using pooling operations. Further, the cross-modality dependencies are built by using up-sampling layers to amplify the modality-shared parts from these patterns. At last, to avoid overfitting, the shuffle attention is designed to re-weight the channel dependencies of all the feature maps, which spreads attention to local patterns as much as possible. In general, the major contributions of this paper can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.755, + 0.468, + 0.83 + ], + "angle": 0, + "content": "- We propose an effective task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. To the best of our knowledge, it is the first work to develop a paradigm for VI recognition on edge devices with an extremely low computation budget." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- An effective task-oriented pretraining strategy is proposed to enhance the heterogeneous feature learning capacity of lightweight networks with task-oriented augmentations and the proposed fake domain loss." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.755, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "- A fine-grained dependency reconstruction module is designed to mine cross-modality dependencies." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.133, + 0.89, + 0.193 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods on mainstream VI-ReID and VI face recognition datasets by a remarkable margin and extremely low complexity." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.89, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.208, + 0.64, + 0.224 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Visible-Infrared Face Recognition (VI-FR). It aims to match face images across visible and infrared modalities. Initially, handcraft features are designed for VI-FR. In [3,9], authors utilize the Local Radon Binary Pattern (LRBP) as the general representation to perform identity matching. Difference-of-Gaussian (DoG) filtering and Multiscale Block Local Binary Patterns (MB-LBP) are widely adopted [20, 50] to acquire robust representations for VI-FR. Recently, deep learning methods have achieved considerable success in VI-FR, focusing on learning modality-shared features with CNN networks [13, 14, 22] or disentangling the modality-invariant representations with GAN networks [46]. In [37], a dual-generation method is proposed to disentangle modality-invariant patterns from visible and infrared images, which helps relieve the modality discrepancy. After that, Duan et al. [8] proposed a pose-aligned cross-spectral hallucination (PACH) network to eliminate identity-independent patterns at multiple stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Visible-Infrared Person Re-identification (VI-ReID). It aims to match visible and infrared pedestrian images according to the feature similarity ranking [47]. The zeropadding scheme [35] was first proposed to handle VI-ReID. In the meantime, they established the first large-scale VI-ReID dataset, named SYSU-MM01. Afterwards, a bidirectional center-constrained network [39] was presented to simultaneously optimize intra- and inter-modality discrepancies. To help the network learn modality-invariant patterns, X-modality [19] is introduced to bridge the gap between visible and infrared modality. Recently, a dense alignment learning method [26] was presented to establish the cross-modality feature correspondence at the pixel level. Ye et al. [38] introduced an attentive graph to discover tri-level relations between visible and infrared images. In addition, several GAN-based methods implement cross-modality alignment [29, 30] or modality compensation [31, 45] to help the network learn robust and compact knowledge. However, despite the promising results, these works are too complex to be utilized in the real world." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.634, + 0.839 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.803, + 0.862 + ], + "angle": 0, + "content": "3.1. Overview of The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.89, + 0.901 + ], + "angle": 0, + "content": "The major pipeline of the proposed TOPLight is illustrated in Fig. 2, which can be summarized as three steps." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3542" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.096, + 0.844, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.392, + 0.848, + 0.406 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of our method. We take three steps to adapt a lightweight network to the VI recognition task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.42, + 0.47, + 0.526 + ], + "angle": 0, + "content": "Step-1. To adapt the lightweight networks to VI recognition scenarios, we construct a dual-path network based on the lightweight model trained on ImageNet-1k (similar in [26]) and further train the dual-path network on ImageNet-mini [18] with task-oriented augmentations. The identity consistency loss \\((L_{idc})\\) and fake domain loss \\((L_{fd})\\) are designed to supervise the domain conflict learning process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.528, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Step-2. We utilize the uniform soup scheme [34] to unify two pre-embedding block groups. This step aims to tune the structure of the pretrained network suited for VI training, which preserves two separate blocks (Block1-1 and Block1-2) to extract low-level features from two modalities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.606, + 0.47, + 0.68 + ], + "angle": 0, + "content": "Step-3. In this stage, the pretrained dual-path network is further trained on VI recognition datasets. A novel fine-grained dependency reconstruction (FDR) module is integrated into the network, which can construct fine-grained pattern dependencies between visible and infrared features." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.383, + 0.712 + ], + "angle": 0, + "content": "3.2. Task-oriented Pretraining Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The core motivation of the task-oriented pretraining strategy is to enable lightweight networks to learn prior knowledge related to cross-modality matching during the pretraining stage. Specifically, we first transform the lightweight network trained on the ImageNet-1k dataset into the dual-path network, which consists of two stem blocks, two pre-embedding block groups and one post-embedding block group. Then, the dual-path network is further trained on the ImageNet-mini dataset. During this training, task-oriented data augmentation (DA) is utilized to create visual differences in training samples, which intends to simulate the discrepancy between visible and infrared" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.418, + 0.869, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.596, + 0.84, + 0.609 + ], + "angle": 0, + "content": "Figure 3. The effects of each data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.674 + ], + "angle": 0, + "content": "images and enforce the network to pay more attention to the identity-aware but colour-irrelevant patterns. Thus, we design the generic, colour and texture data augmentation (DAs) to achieve the training process, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "For the generic DAs, we increase the overall sample diversity by using random crop, flip and deformation. In the colour DAs, we randomly perform colour jitter, channel shuffle and RGB shift for each sample in the Block1-1 path. The colour DAs are utilized to disturb the strong colour prior knowledge learned from the ImageNet-1k dataset, forcing the network to pay more attention to the structural patterns of visible images (which also exist in infrared images), not just colour patterns. In the texture DAs, we randomly perform compression, defocusing and sharpenGray for each sample in the Block1-2 path, which intends to remove the colour information and simulate the terrible imaging quality under night surveillance conditions. During the TOP, we combine these three DAs to create visual differences that are close to the VI recognition scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3543" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "However, in the above manner, the network can lazily learn from one path to avoid embedding visually discrepant samples. Meanwhile, the sample differences generated by augmentations are insufficient to simulate the serious domain conflict in VI recognition scenes. Thus, we develop a fake domain loss to aggravate the variation of mid-level features during the training and motivate the network to discover shared patterns among them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Assume \\(\\mathbf{x}\\) denotes a raw sample from the ImageNet-mini dataset; after augmentations and feature learning, the \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\) denote the outcome features of Block4-1 and Block4-2, respectively. \\(\\mathbf{y}\\) is the shared ground-truth label. We concatenate \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\) and feed them into Block5. The output of Block7 is denoted as \\(\\hat{\\mathbf{x}} = \\{\\hat{\\mathbf{x}}_1, \\hat{\\mathbf{x}}_2\\}\\). Based on the above definitions, the fake domain loss can be defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.33, + 0.469, + 0.348 + ], + "angle": 0, + "content": "\\[\nL _ {f d} ^ {1} = L _ {d} \\left(\\mathbf {x} _ {1}, \\mathbf {d} _ {1}\\right) + L _ {d} \\left(G R L \\left(\\hat {\\mathbf {x}} _ {1}\\right), \\mathbf {d} _ {1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.373, + 0.469, + 0.391 + ], + "angle": 0, + "content": "\\[\nL _ {f d} ^ {2} = L _ {d} \\left(\\mathbf {x} _ {2}, \\mathbf {d} _ {2}\\right) + L _ {d} \\left(G R L \\left(\\hat {\\mathbf {x}} _ {2}\\right), \\mathbf {d} _ {2}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.607 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}_1\\) and \\(\\mathbf{d}_2\\) are the corresponding pseudo domain labels for \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\). We use the domain labels to pretend that \\(\\mathbf{x}_1\\) is from one domain and \\(\\mathbf{x}_2\\) is from the other. \\(L_{d}\\) denotes the domain classification loss implemented with the logistic classifier. \\(GRL\\) is the gradient reversal layer proposed in [10], which aims to reverse the optimization target of the final domain classifier. In this manner, \\(L_{fd}^{1}\\) and \\(L_{fd}^{2}\\) generate two contradictory learning procedures: self-domain constraints are individually set on \\(\\mathbf{x}_1\\) and \\(\\mathbf{x}_2\\), which enforce them to be representative and specific for the two fake domains we pretended, respectively. The inverted-domain constraints are set on the final features after Block7 (\\(\\hat{\\mathbf{x}}_1\\) and \\(\\hat{\\mathbf{x}}_2\\)) by reversing gradients, which urges them to be similar so as to confound the domain classification loss \\(L_{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.713 + ], + "angle": 0, + "content": "During this procedure, Block(2,3,4)-1 and Block(2,3,4)-2 are trained to extract two types of strongly distinguished \"heterogeneous features\". In comparison, Block(5,6,7) are trained to embed these discrepant features and learn their shared patterns. We manually create this conflict to force the lightweight network to learn how to represent heterogeneous features and discover shared patterns from them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.469, + 0.744 + ], + "angle": 0, + "content": "To ensure that the embedded features are identity-related, the identity consistency loss \\((L_{idc})\\) is introduced as:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.759, + 0.469, + 0.776 + ], + "angle": 0, + "content": "\\[\nL _ {i d c} = \\log [ P (y | p (\\hat {\\mathbf {x}} _ {1})) ] + \\log [ P (y | p (\\hat {\\mathbf {x}} _ {2})) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.783, + 0.469, + 0.857 + ], + "angle": 0, + "content": "where \\( p(\\hat{\\mathbf{x}}_1) \\) and \\( p(\\hat{\\mathbf{x}}_2) \\) denote the identity prediction results from the same linear classifier. \\( \\mathbf{y} \\) is the ground-truth label for both \\( \\mathbf{x}_1 \\) and \\( \\mathbf{x}_2 \\). Here, we utilize one classifier to predict two conflicted features with shared logits, aiming to further improve the embedding capacity of Block(5,6,7)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.858, + 0.42, + 0.872 + ], + "angle": 0, + "content": "The overall loss in the TOP is written as follows." + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nL = L _ {i d c} + L _ {f d} ^ {1} + L _ {f d} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "After the pretraining, we utilize the uniform soup scheme to unify two groups of pre-embedded blocks in turn (Step-2 in Fig. 2). The final lightweight network for VI-recognition training consists of two independent stem blocks (Block1-1 and 1-2), one pre-embedding block group (Block2,3,4) and one post-embedding block group (Block5,6,7), as shown in the Step-3 of Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.209, + 0.855, + 0.226 + ], + "angle": 0, + "content": "3.3. Fine-grained Dependency Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.354 + ], + "angle": 0, + "content": "After the task-oriented pretraining, the lightweight network learns how to extract modality-invariant features from heterogeneous images. However, it lacks the ability to discover fine-grained patterns across the modalities. Therefore, we introduce a practical fine-grained dependency reconstruction module to help the lightweight network discover plentiful fine-grained patterns by constructing cross-modality feature dependencies in the VI training phase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Specifically, let us denote \\( S^v = \\{\\mathbf{x}_i^v | i = 1,2,\\dots,N^v\\} \\) and \\( S^r = \\{\\mathbf{x}_i^r | i = 1,2,\\dots,N^r\\} \\) as the visible and infrared training sets, respectively; \\( N^v \\) and \\( N^r \\) indicate the total numbers of visible and infrared samples, respectively. During the training, we feed \\( \\mathbf{x}^v \\) and \\( \\mathbf{x}^r \\) into Block1-1 and Block1-2 to extract low-level features of two modalities, respectively. Then, we concatenate these low-level features and feed them into the Block(2,3,4) for the initial feature embedding. Based on the outcome of Block4 \\( (\\mathbf{Z} = \\{\\mathbf{Z}^v,\\mathbf{Z}^r\\}) \\), the fine-grained and cross-modality feature dependencies are constructed to help the network learn substantial modality-shared patterns by the FDR module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.598 + ], + "angle": 0, + "content": "In order to extract fine-grained patterns from feature maps, we slice \\(\\mathbf{Z}\\) into parts along the horizontal and vertical directions. For each spatial map \\(\\{\\mathbf{Z}\\}_{m=0}^{C-1} \\in \\mathbb{R}^{H \\times W}\\), the horizontal patterns are obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.609, + 0.891, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} ^ {f} = \\operatorname {C o n c a t} [ \\operatorname {W i n d o w} (\\{\\mathbf {Z} ^ {f} \\} _ {m = 0} ^ {C - 1}) ], f \\in \\{v, r \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.742 + ], + "angle": 0, + "content": "where \\(\\mathbf{Z}^f\\in \\mathbb{R}^{C\\times H\\times W},f\\in \\{v,r\\}\\) denote visible or infrared features. \\(C,H,W\\) are the channel, height, and width dimensions, respectively. \\(\\mathbf{h}^f\\in \\mathbb{R}^{C\\times \\frac{H}{\\sigma}\\times \\sigma \\times W}\\) denotes a group of fine-grained horizontal patterns extracted from \\(\\mathbf{Z}^f\\) Similarly, the vertical patterns \\(\\mathbf{v}^f\\in \\mathbb{R}^{C\\times \\frac{W}{\\mu}\\times H\\times \\mu}\\) can also be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.753, + 0.891, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {f} = \\operatorname {C o n c a t} \\left[ \\operatorname {W i n d o w} \\left(\\left\\{\\mathbf {Z} ^ {f} \\right\\} _ {m = 0} ^ {C - 1}\\right) \\right], f \\in \\{v, r \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.862, + 0.806 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) are used to control the slice granularity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Further, the spatial information of \\(\\mathbf{h}^f\\) and \\(\\mathbf{v}^f\\) is condensed by the average pooling operation, which produces \\(\\hat{\\mathbf{h}}^f\\in \\mathbb{R}^{C\\times \\frac{H}{\\sigma}}\\) and \\(\\hat{\\mathbf{v}}^f\\in \\mathbb{R}^{C\\times \\frac{W}{\\mu}}\\). In the above manner, we refine the original spatial maps \\(\\{\\mathbf{Z}^f\\}_{m = 0}^{C - 1}\\in \\mathbb{R}^{H\\times W}\\) into directional tensors \\(\\hat{\\mathbf{h}}^f\\) and \\(\\hat{\\mathbf{v}}^f\\) to cut off the original self-modality dependencies in spatial dimensions. Furthermore," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3544" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.468, + 0.107 + ], + "angle": 0, + "content": "the cross-modality dependencies are established as follows." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.121, + 0.47, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {h} ^ {f} = \\underset {\\left(\\frac {H}{\\sigma} \\times W\\right)} {\\operatorname {T r a n s c o n v}} (\\hat {\\mathbf {h}} ^ {f}), f \\in \\{v, r \\}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.165, + 0.469, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {v} ^ {f} = \\underset {\\left(H \\times \\frac {W}{\\mu}\\right)} {\\operatorname {T r a n s C o n v}} (\\hat {\\mathbf {v}} ^ {f}), f \\in \\{v, r \\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.203, + 0.469, + 0.279 + ], + "angle": 0, + "content": "In Eq. (7) and Eq. (8), we use two modality-shared transposed convolution layers to reconstruct the spatial maps of \\(\\mathbf{h}^f\\) and \\(\\mathbf{v}^f\\), respectively. This process aims to establish the cross-modality spatial dependencies. The outputs are \\(\\mathbf{z}_h^f \\in \\mathbb{R}^{C \\times H \\times W}\\) and \\(\\mathbf{z}_v^f \\in \\mathbb{R}^{C \\times H \\times W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.28, + 0.469, + 0.434 + ], + "angle": 0, + "content": "However, there exist \"scheme stereotypes\" in \\(\\mathbf{Z}_h^f\\) and \\(\\mathbf{Z}_v^f\\). The main reason is that these features are learned from two distinct directional schemes, which may lead to a skewed learning procedure. In contrast, we hope these two schemes can consistently provide diversified patterns that cover more valuable signals. Thus, the shuffle attention is designed to fuse \\(\\mathbf{Z}_h^f\\) and \\(\\mathbf{Z}_v^f\\) while decoupling them from two directions. Specifically, we concatenate them on the channel level and then shuffle the channel order to disarrange the spatial maps learned from two directions." + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.445, + 0.469, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} ^ {f} = \\operatorname {S h u f f l e} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {Z} _ {h} ^ {f}, \\mathbf {Z} _ {v} ^ {f}\\right)\\right), f \\in \\{v, r \\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.469, + 0.523 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}^f\\in \\mathbb{R}^{2C\\times H\\times W}\\) denotes the concatenated feature with disarranged channel order. The new channel dependencies are established by weighting all the spatial maps." + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.549, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {D}} ^ {f} = \\mathbf {W} _ {2} R e L U \\left(\\mathbf {W} _ {1} G A P \\left(\\mathbf {D} ^ {f}\\right)\\right), f \\in \\{v, r \\}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.58, + 0.469, + 0.628 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_1\\in \\mathbb{R}^{2C\\times \\frac{C}{4}}\\) and \\(\\mathbf{W}_2\\in \\mathbb{R}^{\\frac{C}{4}\\times C}\\) indicate projection matrices to compute the channel attention [16]. Finally, we introduce \\(\\hat{\\mathbf{D}}^f\\) to refine the original features." + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.639, + 0.469, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} ^ {f} = \\hat {\\mathbf {D}} ^ {f} \\oplus \\mathbf {Z} ^ {f}, f \\in \\{v, r \\}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.67, + 0.469, + 0.761 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{Z}}^f\\) is used as the input of the next block; \\(\\oplus\\) is the element-wise add operation. The proposed fine-grained dependency reconstruction module can effectively establish fine-grained and cross-modality dependencies in both visible and infrared features, which can help discover substantial small patterns that are vital in VI recognition." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.773, + 0.222, + 0.789 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.798, + 0.469, + 0.843 + ], + "angle": 0, + "content": "We utilize the hard-mining triplet loss \\((L_{tri})\\) [42] and cross-entropy loss \\((L_{id})\\) for metric learning and identity learning, respectively. The overall loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.861, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\nL = L _ {i d} + \\gamma * L _ {t r i}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.415, + 0.901 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the parameter to balance each loss term." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.706, + 0.107 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.799, + 0.131 + ], + "angle": 0, + "content": "4.1. Datasets and Settings for VI-ReID" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.892, + 0.183 + ], + "angle": 0, + "content": "We first evaluate the proposed method for the visible-infrared person re-identification (VI-ReID) task on two public benchmarks: SYSU-MM01 [35] and RegDB [25]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.288 + ], + "angle": 0, + "content": "SYSU-MM01 [35] is the largest VI-ReID dataset, which contains 491 persons captured from six different cameras across scenes (indoor and outdoor). There are a total of 287628 visible images and 15792 infrared images. Following the evaluation protocol in [26,38], 395 persons are fixed for training and 96 persons for testing. The testing process includes all-search and indoor-search modes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.425 + ], + "angle": 0, + "content": "RegDB [25] contains 412 persons, and each person has 10 visible and 10 infrared images. \"Visible to infrared\" and \"infrared to visible\" are two testing modes. The former uses visible as the query and infrared as the gallery. The latter does the opposite. Following [26, 30, 38], the final performance on the RegDB is made with the average results of 10 times the training and testing procedure. Each time, we randomly sample 206 identities for training, and the remaining 206 identities are used for testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Evaluation Metrics. We use the standard cumulative matching characteristics (CMC), mean average precision (mAP), and mean inverse negative penalty (mINP) [42] as the evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Implementation Details. For the ablation studies, we select the MobileNetV3-L [15] as the lightweight baseline to validate each proposed component. As shown in Table 1, we encapsulate all the layers of MobileNetV3-L as Block1-7 and then initialize the first four blocks twice to construct the dual-path network. During the TOP, detailed settings are" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.583, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Table 1. Detailed structures for each Block. We package the entire MobileNetV3-Large into Block1-7 without overlap." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.758 + ], + "angle": 0, + "content": "
Block partitions on MobileNetV3-large
Layer NameStructuresOutput Size
Block1Conv(3×3, 2), Bneck(3×3, 1)16 × 1122
Block2Bneck(3×3, 2)24 × 562
Block3Bneck(3×3, 1)24 × 562
Block4Bneck(5×5, 2), Bneck(5×5, 1)×240 × 282
Block5Bneck(3×3, 2), Bneck(3×3, 1)×5112 × 142
Block6Bneck(5×5, 2), Bneck(5×5, 1)×2160 × 72
Block7Conv(1×1, 1)×31280 × 72
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.891, + 0.825 + ], + "angle": 0, + "content": "shown in Table 2(b). It is worth noticing that the \"Random pick\" means that for each image sample used in the Block1-1 path, we randomly pick one method from colour jitter, channel shuffle, or RGB shift to augment it." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During the VI-ReID training, detailed settings are presented in Table 2(a). Briefly, we randomly sample 6 identities with 6 visible and 6 infrared images per identity, which makes a total of 72 images in each batch. All images are resized to \\(288 \\times 144\\). Random erasing (RE) [49], flipping," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3545" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.09, + 0.432, + 0.103 + ], + "angle": 0, + "content": "Table 2. Detailed training and augmentation settings." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.104, + 0.47, + 0.265 + ], + "angle": 0, + "content": "
(a): Training settings
SettingsOn ImageNet-1KOn ImageNet-miniOn VI-ReID
Epochs1505070
Batch size51236×236×2
Image size224×224224×224288×144
Augs.Stanard [15]Task-orientedRE, CE, Flip
OptimizerSGDSGDAdamW [23]
Lr.0.20.10.001
Decaycosinecosineconstant
Warm upFirst 5 epochNoneFirst 5 epoch
Weight decay5E-45E-51E-2
PlatformPytorch, FP16PytorchPytorch
GPUTesla V100RTX 3060RTX 3060
Cost57 h4 h-
" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.266, + 0.47, + 0.38 + ], + "angle": 0, + "content": "
(b): Task-oriented augmentation settings
AugmentationsProbabilityTarget
Generic (G)Crop & Filp0.5All path
Deformation0.2
Colour (C)Colour jitterRandom pickBlock1-1 path
Channel shuffle
RGB shift
Texture (T)Compression0.2Block1-2 path
Defocusing0.2
Sharpen gray1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.514 + ], + "angle": 0, + "content": "and channel erasing (CE) [40] are utilized against overfitting. The AdamW [23] optimizer with a learning rate (Lr) of 0.001, and a weight decay of 0.01 is adopted to train our network for 70 epochs. We use the Lr warm-up in the first 5 epochs and constant decay at the \\(20^{th}\\) and \\(40^{th}\\) epochs with a decay factor of 0.1. The parameters \\(\\mu\\) and \\(\\sigma\\) are set to 3 and 2 based on the fine-tuning results, respectively. The balance parameter \\(\\gamma\\) in Eq. (12) is empirically set to \\(1/2\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.527, + 0.232, + 0.543 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Effect of Task-oriented Pretraining. We first evaluate the effect of task-oriented pretraining on VI-ReID, as shown in Table 3. No.1 indicates the baseline performance obtained via direct fine-tuning of the ImageNet-1k pretrained model in VI-ReID datasets. When only using generic augmentations and identity loss to perform TOP on ImageNetmini, the \"catastrophic forgetting\" makes No.2 perform worse than the baseline. However, when applying the proposed task-oriented augmentations, the network learns to extract stable features from visually discrepant images in advance, thus improving the downstream VI-ReID performance (No.3). Coupled with \\( L_{idc} \\) and \\( L_{fd} \\), we create a conflicting learning process that enforces the network to learn how to represent and embed heterogeneous features. No.5 shows significant improvements in both the SYSU-MM01 and RegDB datasets. By only using the MobileNetV3-large backbone, the performance reported in No.5 has already exceeded many ResNet-based works [19, 26, 38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Effect of Fine-grained Dependency Reconstruction Module. In Table 3, No.6-No.9 show the ablation experiments about the fine-grained dependency reconstruction module. We intend to construct the cross-modality feature dependencies, and the first issue is how to model the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "original feature relations. No.6 indicates directly applying the coarse-grained features to the shuffle attention module. The performance is slightly improved. When using the proposed directional scheme \\((H_{s},V_{s})\\) to extract fine-grained patterns and reconstruct their spatial relation using transposed convolution layers, No.7 and No.8 show meaningful improvements. The full power of the FDR module is shown in No.9, which adopts both fine-grained horizontal and vertical patterns to discover cross-modality feature similarities, achieving \\(66.14\\%\\) and \\(63.80\\%\\) in terms of Rank-1 and mAP performance on SYSU-MM01 under the all-search mode." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Pointing at the difficulties in VI recognition tasks, we design the TOP strategy and FDR module to remedy the drawbacks of the lightweight network. Compared with the baseline (No.1), our method (No.9) considerably improves all metrics on two benchmarks. Meanwhile, the ablations reported in Table 3 demonstrate that all the proposed components bring advantages consistently to the accuracy gain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Applicability with Lightweight Networks. We evaluate our methods on three mainstream lightweight networks (ShuffleNetV2 [24], GhostNet [11] and MobileNetV3 [15]). As shown in Table 4, our method can remarkably improve VI Re-ID performance with ignorable complexity growth. Compared with conventional deep networks, we surpass them significantly with minimal complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.592 + ], + "angle": 0, + "content": "Detailed Ablations on the FDR Module. We validate the rationality of each component in the FDR module. Table 5(a) shows the impacts of different spatial modelling methods. Except for ours, the HAP [38] achieves the best performance. However, we extract patterns from both the horizontal and vertical regions and use up-sampling (u.) to discover modality-shared cues from re-enlarged spatial areas, which can be regarded as a more effective solution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Table 5(b) shows the impacts of different channel relation reasoning methods. Our shuffle attention (SA) wins both SE [16] and CBAM [33]. The recipe is the channel shuffle (cs.) operation that builds scheme-decoupled dependencies. This crucial step can evenly correlate visible and infrared features from each fine-grained pattern." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.702, + 0.71 + ], + "angle": 0, + "content": "4.3. Feature Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.764 + ], + "angle": 0, + "content": "What Does TOP Do? We randomly track 10 identities during the TOP stage and visualize the feature distribution via T-SNE [28], as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We first use task-oriented augmentation to create two types of visual discrepant samples from each image, as shown in Fig. 4(a) and Fig. 4(b). Then, the proposed fake domain loss forces the features after Block4 to be domain-specific, as shown in Fig. 4(c). Concurrently, it forces the network to learn to embed those domain-specific features and provide domain-shared predictions. In Fig. 4(d), after the training, we can observe that the network knows how to extract the domain-shared representations and successfully" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3546" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 3. Evaluation of each proposed component on two VI-ReID datasets. \"Augs.\" indicates the augmentations. G, C and T denote the generic, colour, and texture augmentations, respectively. In the FDR module, \\( H_{s} \\) and \\( V_{s} \\) denote the horizontal and vertical slices with up-sampling. SA is the shuffle attention module. Rank (r) (\\%), mAP (\\%) and mINP (\\%) are reported." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.132, + 0.894, + 0.304 + ], + "angle": 0, + "content": "
No.Task-oriented pretraining stageVI training stageSYSU-MM01 (all-search)RegDB (visible-to-infrared)
Augs.Loss functionsFDR module
GC+TLidLidcLfdHsVsSAr=1r=10mAPmINPr=1r=10mAPmINP
147.8189.7147.0633.4871.2689.9465.6648.50
2✓✓43.2885.9645.5631.1070.7388.5265.6548.41
3✓✓✓49.8589.7447.5635.5271.3289.9165.6748.49
4✓✓✓54.2892.1152.9441.2975.3192.6468.7852.16
5✓✓✓✓62.4194.1259.0645.1382.7594.1376.2161.84
6✓✓✓✓✓62.8994.2659.7945.8482.8894.1976.2461.92
7✓✓✓✓✓✓63.9595.2860.0946.8083.0794.4876.5562.20
8✓✓✓✓✓✓64.0495.4161.1246.9283.2294.6977.0163.16
9✓✓✓✓✓✓✓66.1496.0363.8049.7684.1594.9879.2663.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.314, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Table 4. Experimental results on different lightweight networks and conventional deep networks." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.342, + 0.47, + 0.573 + ], + "angle": 0, + "content": "
MethodsFLOPs (M)SYSU-MM01RegDB
r=1mAPr=1mAP
ConventionResNet-50356256.9854.7276.8671.30
ConvNeXt-Tiny362058.7255.3178.2572.64
Vit-B568952.1751.8175.3170.37
Swin-Tiny328758.2455.1678.3972.68
LightweightShuffleNetV2-1.0×13941.8841.9467.8364.85
+TOP & FDR17755.7152.6379.8266.36
ShuffleNetV2-1.5×26547.3947.8170.1565.28
+TOP & FDR37163.3560.8184.1376.98
GhostNet-1.0×15042.5342.9471.2864.40
+TOP & FDR18958.5455.1983.2677.16
GhostNet-1.3×28150.8947.9272.5165.98
+TOP & FDR39566.7664.0185.5179.95
MobileNetV3-S10440.9242.5162.7758.31
+TOP & FDR13054.7550.2675.5370.17
MobileNetV3-L25047.8147.0671.2665.66
+TOP & FDR36266.1463.8084.1579.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.577, + 0.47, + 0.619 + ], + "angle": 0, + "content": "Table 5. Evaluation of spatial modelling methods and channel relation reasoning methods in the FDR module. \"u.\" and \"cs.\" respectively denote the up-sampling and channel shuffle operations." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.621, + 0.47, + 0.81 + ], + "angle": 0, + "content": "
(a): Impact of different spatial modelling methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
GAP62.8959.7945.8482.8876.2461.92
Context [2]61.3757.5246.0881.5674.4962.11
HAP [38]62.9258.1347.8482.9876.2062.77
HS+VS(w/o u.)63.4559.7549.6183.9478.8263.21
HS+VS66.1463.8049.7684.1579.2663.86
(b): Impact of different channel relation reasoning methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
SE [16]62.9159.7045.7882.7977.4562.34
CBAM [33]61.7956.8845.2583.4177.2862.96
SA (w/o cs.)62.8158.1045.7682.7576.1262.56
SA66.1463.8049.7684.1579.2663.86
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.469, + 0.868 + ], + "angle": 0, + "content": "embed all the features into their respective identity groups. Thus, the proposed task-oriented pretraining strategy makes lightweight networks fit cross-modality tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Pattern Visualization. As demonstrated in Fig. 5, we visualize the learned patterns using Grad-CAM [27] to ex" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.314, + 0.868, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Figure 4. Feature distributions during task-oriented pretraining. Each colour denotes an identity. In (a), the star markers represent raw samples. In (b), (c) and (d), the circle and triangle markers represent colour and texture-augmented samples, respectively." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.538, + 0.865, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Figure 5. Visualization of learned patterns via Grad-CAM [27]. For a clear presentation, we average all the spatial maps into four. plain the effect of the FDR module. Clearly, the original patterns extracted by MobileNetV3-L are concentrated and overfit on the modality-specific parts. To remedy this issue, we use the FDR module to discover modality-shared patterns from fine-grained regions, which separates learned patterns to focus on considerable modality-shared cues. A good case is shown in Fig. 5. The visible patterns initially learned are concentrated in the high chest area, while the infrared patterns are all in the right arm. After re-weighting by the FDR module, modality-shared small regions like the face, arms, neck, and shoes are all taken into account, thus" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3547" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.88, + 0.103 + ], + "angle": 0, + "content": "Table 6. Comparison with the state-of-the-arts on SYSU-MM01 [35]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.104, + 0.892, + 0.276 + ], + "angle": 0, + "content": "
DetailsAll-searchIndoor-search
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356214.8054.1271.3315.95-20.5868.3885.7926.92-
JSIA [30]ResNet50+GAN>413338.1080.7089.9036.90-43.8086.2094.2052.90-
AGW [42]ResNet50>356247.5084.3992.1447.6535.3054.1791.1495.9862.9759.23
X-Modal [19]ResNet50>356249.9089.8096.0050.70------
DMiR [38]ResNet50>356250.5488.1294.8649.29-53.9292.5097.0962.49-
FBP-AL [32]ResNet50>356254.1486.0493.0350.20------
DDAG [41]ResNet50>356254.7590.3995.8155.0239.6261.0294.0698.4167.9862.61
HAT [43]ResNet50>356255.2992.1497.3653.89-62.1095.7599.2070.84-
LBA [26]ResNet50>356255.41--54.14-58.46--66.33-
TSME [21]ResNet50>356264.2395.1998.7361.21-64.8096.9299.3171.53-
SPOT [4]ResNet50+ViT>481065.3492.7397.0462.2548.8669.4296.2299.1274.6370.48
TOLight (Ours)MobileNetV3-L= 36266.1496.0397.6863.8049.7672.4197.5499.2376.1171.43
TOLight (Ours)GhostNet-1.3×= 39566.7696.2398.7064.0150.1872.8997.9399.2876.7071.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.281, + 0.859, + 0.294 + ], + "angle": 0, + "content": "Table 7. Comparison with the state-of-the-arts on RegDB [25]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.295, + 0.892, + 0.464 + ], + "angle": 0, + "content": "
DetailsVisible-to-InfraredInfrared-to-Visible
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356217.7534.2144.3518.90-16.6334.6844.2517.82-
JSIA [30]ResNet50+GAN>413348.50--49.30-48.10--48.90-
AGW [42]ResNet50>356270.0586.2191.5566.3750.1970.4987.2191.8465.9051.24
X-Modal [19]ResNet50>356262.2183.1391.7260.18------
DMiR [38]ResNet50>356275.7989.8694.1869.97-73.9389.8793.9868.22-
FBP-AL [32]ResNet50>356273.9889.7193.6968.24-70.0589.2293.8866.61-
DDAG [41]ResNet50>356269.3486.1991.4963.4649.2468.0685.1590.3161.8048.62
HAT [43]ResNet50>356271.8387.1692.1667.56-70.0268.4591.6166.30-
LBA [26]ResNet50>356274.17--67.64-72.43--65.46-
SPOT [4]ResNet50+ViT>481080.3593.4896.4472.4656.1979.3792.7996.0172.2656.06
GECNet [48]ResNet50+GAN>435082.3392.7295.4978.45-78.9391.9995.4475.58-
TOLight (Ours)MobileNetV3-L=36284.1594.9896.5879.2663.8680.9492.8596.3776.1059.33
TOLight (Ours)GhostNet-1.3×=39585.5194.9996.7079.9563.8580.6592.8196.3275.9159.26
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.477, + 0.445, + 0.492 + ], + "angle": 0, + "content": "improving the accuracy of the cross-modality matching." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.383, + 0.515 + ], + "angle": 0, + "content": "4.4. Comparison with State of The Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.468, + 0.674 + ], + "angle": 0, + "content": "We compare our method with recently proposed state-of-the-art VI-ReID methods, as shown in Table 6 and Table 7. They illustrate that we remarkably and efficiently exceed all the compared SOTAs under diverse evaluation settings. Specifically, based on the tiny GhostNet- \\(1.3\\times\\) backbone, we attain the rank-1 of \\(66.76\\%\\) and mAP of \\(64.01\\%\\) on the all-search mode of the large-scale SYSU-MM01 dataset with only 395 FLOPs. The proposed method also achieves excellent results on the RegDB dataset, surpassing the previous SOTAs by a significant margin with lower FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.679, + 0.468, + 0.72 + ], + "angle": 0, + "content": "Table 8. Evaluation on two VI-FR datasets. CA is channel augmentation [40]. B is the LightCNN-29 baseline. Rank at 1 accuracy \\((\\%)\\) and false acceptance rate (F: \\(\\%\\) ) are reported." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.721, + 0.468, + 0.829 + ], + "angle": 0, + "content": "
MethodsOulu [5]BUAA [17]
r=1F:1%F:0.1%r=1F:1%F:0.1%
IDR [13]94.373.446.294.393.484.7
VSA [44]99.996.882.398.098.292.5
PACH [8]10097.988.298.698.093.5
B [36]10097.987.098.097.793.7
B+CA [40]10098.991.798.398.294.5
B+TOP10098.891.598.398.194.5
B+TOP+FDR (Ours)10098.991.798.398.294.6
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.373, + 0.864 + ], + "angle": 0, + "content": "4.5. Visible-Infrared Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We follow [40] to examine our method on Oulu-CASIA NIR-VIS [5] and BUAA-VisNir face databases [17]. Fol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.552 + ], + "angle": 0, + "content": "lowing previous studies, we also adopt the LightCNN-29 [36] as our baseline network (B). Except for IDR [13], VSA [44], and PACH [8], we also compare recently proposed channel augmentation (CA) [40] with our method on the same baseline. Results are displayed in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Evidently, the proposed TOP and FDR also enhance the performance of VI face recognition. These results further validate the effectiveness and generalization of our methods, which can solve VI recognition tasks effectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.631, + 0.619, + 0.647 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.854 + ], + "angle": 0, + "content": "This work presents an effective task-oriented pretrained lightweight neural network (TOPLight) to solve visible-infrared recognition problems. First, the task-oriented pretraining strategy significantly improves the ability of lightweight networks to understand heterogeneous images by introducing domain conflict and sample variation during the pretraining phase. Second, the fine-grained dependency reconstruction module is utilized to destroy the modality-specific pattern dependencies and construct the fine-grained and modality-shared pattern dependencies between visible and infrared features. Extensive experiments demonstrate the superiority and effectiveness of the proposed method as well as the efficacy of each component of our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is funded by the National Natural Science Foundation of China (Grant No. 61802058, 61911530397)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3548" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] Sun, et.al. Learning part-based convolutional features for person re-identification. TPAMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Srinivasa Rao Chalamala, Krishna Rao Kakkirala, and Jami Santosh Kumar. Face recognition using spatial pyramid matching and lrbp. In 2014 IEEE 10th International Colloquium on Signal Processing and its Applications, pages 67-70. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.355 + ], + "angle": 0, + "content": "[4] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Jianguo Jiang, and Chia-Wen Lin. Structure-aware positional transformer for visible-infrared person re-identification. IEEE Transactions on Image Processing, 31:2352-2364, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[5] Jie Chen, Dong Yi, Jamei Yang, Guoying Zhao, Stan Z Li, and Matti Pietikainen. Learning mappings for face synthesis from near infrared to visual light images. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 156-163. IEEE, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[6] Changxing Ding and Dacheng Tao. A comprehensive survey on pose-invariant face recognition. ACM Transactions on intelligent systems and technology (TIST), 7(3):1-42, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[7] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[8] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7930-7938, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[9] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In 2012 19th IEEE International Conference on Image Processing, pages 1837-1840. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[11] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Learning invariant deep representation for nir-vis face recognition. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 1, 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Wasserstein cnn: Learning invariant features for nir-vis face recognition. IEEE transactions on pattern analysis and machine intelligence, 41(7):1761-1773, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[17] D Huang, J Sun, and Y Wang. The buaa-visnir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3:3, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[18] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[19] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4610–4617, 2020. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[20] Shengcai Liao, Dong Yi, Zhen Lei, Rui Qin, and Stan Z Li. Heterogeneous face recognition from local structures of normalized appearance. In International Conference on Biometrics, pages 209-218. Springer, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[21] Jianan Liu, Jialiang Wang, Nianchang Huang, Qiang Zhang, and Jungong Han. Revisiting modality-specific feature compensation for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[22] Xiaoxiang Liu, Lingxiao Song, Xiang Wu, and Tieniu Tan. Transferring deep representation for nir-vis heterogeneous face recognition. In 2016 International Conference on Biometrics (ICB), pages 1-8. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[24] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[25] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 17(3):605, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Hyunjong Park, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12046–12055, 2021. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3549" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[29] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3623-3632, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[30] Guan-An Wang, Tianzhu Zhang, Yang Yang, Jian Cheng, Jianlong Chang, Xu Liang, and Zeng-Guang Hou. Cross-modality paired-images generation for rgb-infrared person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12144–12151, 2020. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.418 + ], + "angle": 0, + "content": "[31] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 618–626, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.475 + ], + "angle": 0, + "content": "[32] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Flexible body partition-based adversarial learning for visible infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[33] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[34] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[35] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 5380-5389, 2017. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[36] Xiang Wu, Ran He, Zhenan Sun, and Tieniu Tan. A light cnn for deep face representation with noisy labels. IEEE Transactions on Information Forensics and Security, 13(11):2884-2896, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[37] Xiang Wu, Huaibo Huang, Vishal M Patel, Ran He, and Zhenan Sun. Disentangled variational representation for heterogeneous face recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9005–9012, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[38] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE Transactions on Information Forensics and Security, 17:386–398, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE Transactions on Information Forensics and Security, 15:407–419, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[40] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13567-13576, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[41] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In European Conference on Computer Vision, pages 229–247. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[42] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[43] Mang Ye, Jianbing Shen, and Ling Shao. Visible-infrared person re-identification via homogeneous augmented trimodal learning. IEEE Transactions on Information Forensics and Security, 16:728-739, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[44] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-posed high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[45] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7349–7358, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[46] Teng Zhang, Arnold Wiliem, Siqi Yang, and Brian Lovell. Tv-gan: Generative adversarial network based thermal to visible face recognition. In 2018 international conference on biometrics (ICB), pages 174-181. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[47] Huantao Zheng, Xian Zhong, Wenxin Huang, Kui Jiang, Wenxuan Liu, and Zheng Wang. Visible-infrared person re-identification: A comprehensive survey and a new setting. *Electronics*, 11(3):454, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[48] Xian Zhong, Tianyou Lu, Wenxin Huang, Mang Ye, Xuemei Jia, and Chia-Wen Lin. Grayscale enhancement colorization network for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1418–1430, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[49] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.799, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[50] Jun-Yong Zhu, Wei-Shi Zheng, Jian-Huang Lai, and Stan Z Li. Matching nir face to vis face using transduction. IEEE Transactions on Information Forensics and Security, 9(3):501-514, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3550" + } + ] +] \ No newline at end of file diff --git a/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_origin.pdf b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a2cc82ee986e18608793831fc3ece95d535c388f --- /dev/null +++ b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/fd571fdc-7e5b-4461-afda-4cb4debbc856_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b427848f3a2254d63111460f65a7ba5a28d44de5c3578ab96fe0f6c46d27e6d0 +size 2832751 diff --git a/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/full.md b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..97060165c0f57dc73d5a8a3794f8784f0a283911 --- /dev/null +++ b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/full.md @@ -0,0 +1,329 @@ +# TOPLight: Lightweight Neural Networks with Task-Oriented Pretraining for Visible-Infrared Recognition + +Hao Yu $^{1}$ , Xu Cheng $^{1*}$ , Wei Peng $^{2}$ + +$^{1}$ School of Computer Science, Nanjing University of Information Science and Technology, China $^{2}$ Department of Psychiatry and Behavioral Sciences, Stanford University + +{yuhao,xcheng}@nuist.edu.cn，wepeng@stanford.edu + +# Abstract + +Visible-infrared recognition (VI recognition) is a challenging task due to the enormous visual difference across heterogeneous images. Most existing works achieve promising results by transfer learning, such as pretraining on the ImageNet, based on advanced neural architectures like ResNet and ViT. However, such methods ignore the negative influence of the pretrained colour prior knowledge, as well as their heavy computational burden makes them hard to deploy in actual scenarios with limited resources. In this paper, we propose a novel task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. Specifically, the TOPLight method simulates the domain conflict and sample variations with the proposed fake domain loss in the pretraining stage, which guides the network to learn how to handle those difficulties, such that a more general modality-shared feature representation is learned for the heterogeneous images. Moreover, an effective fine-grained dependency reconstruction module (FDR) is developed to discover substantial pattern dependencies shared in two modalities. Extensive experiments on VI person re-identification and VI face recognition datasets demonstrate the superiority of the proposed TOPLight, which significantly outperforms the current state of the arts while demanding fewer computational resources. + +# 1. Introduction + +Identity recognition technologies have provided numerous reliable solutions for monitoring systems, which strive to match the face (face recognition [6, 7]) or pedestrian (person re-identification [42]) images of the same identity. However, the majority of previous efforts only consider visible images. In real-life practice, many surveillance cameras can switch to infrared imaging mode at night. Thus, the essential cross-modality visible-infrared recognition (VI + +![](images/54bc120915c24eb7770ad9b0e317de0c86acda0d77e85122cb490881979aa087.jpg) + +![](images/6b0ae237ab44d55111eac04d4705e779f55e6177646d15303a0badc10d16678b.jpg) +Figure 1. (a) The task-oriented pretraining strategy; (b) Performance comparison of the standard ImageNet-1k pretraining scheme (SP) and the proposed task-oriented pretraining scheme (TOP) on the SYSU-MM01 dataset [35] (All-Search mode). + +recognition) technology has been developed to match the visible and infrared photographs of the same people. + +Recently, visible-infrared person re-identification (VI-ReID) [26, 38] and visible-infrared face recognition [8, 13, 44] have been widely studied. The key issue is identifying the modality-shared patterns. To this end, several works [29, 30] use generative adversarial networks (GANs) to implement cross-modality alignment at the pixel and feature levels. Others [4, 26, 38] design the dual-path feature extraction network, coupled with inter-feature constraints, to close the embedding space of two modalities. However, these methods utilize at least one pretrained ResNet-50 [12] backbone to extract solid features, which makes them unsuitable for edge monitoring devices. Recent works [4, 38] employ auxiliary models (e.g., pose estimation, graph reasoning) to relieve the modality discrepancy, which enhances the performance on academic benchmarks but reduces the real-time inference speed. Compared with conventional deep networks (e.g., ResNet, ViT), lightweight networks [11, 15, 24] can extract basal features rapidly. In VI recogni + +tion tasks, however, the vast modality discrepancy renders the performance of lightweight networks significantly inferior to that of conventional deep networks. The main reason is that lightweight networks lack the ability to identify modality-shared patterns from heterogeneous images. + +To address this issue, we present an effective task-oriented pretraining (TOP) strategy. As shown in Fig. 1(a), we first train a lightweight network on the ImageNet-1k dataset to learn vision prior knowledge. After that, the trained network is transformed into the dual-path network and further trained by using task-oriented data augmentations, identity consistency loss and fake domain loss on the ImageNet-mini dataset [18]. The task-oriented pretraining (TOP) strategy simulates the sample differences in VI scenes and teaches the network how to represent and embed discrepant features. Fig. 1(b) reports the performance of three lightweight networks in the VI-ReID task. Compared with the ImageNet-1k pretraining, our TOP strategy can remarkably improve the baseline performance. + +Another weakness of lightweight networks is that few feature maps are learned from raw images for rapid inference. In the VI recognition scene, it is challenging to discover modality-shared patterns with so few learned feature maps. In practice, the network can focus on a group of aggregated modality-specific patterns that offer the most gradient for identity classification. In contrast, the fine-grained and modality-shared patterns, which are crucial for achieving robust cross-modality matching, are neglected. + +Based on the above observations, we present a novel fine-grained dependency reconstruction (FDR) module to help lightweight networks learn modality-shared and fine-grained patterns. Specifically, inspired by the horizontal slice scheme [1], we first slice feature maps horizontally and vertically to extract fine-grained patterns from diversified local regions. Then, the original spatial dependencies of these patterns are eliminated by using pooling operations. Further, the cross-modality dependencies are built by using up-sampling layers to amplify the modality-shared parts from these patterns. At last, to avoid overfitting, the shuffle attention is designed to re-weight the channel dependencies of all the feature maps, which spreads attention to local patterns as much as possible. In general, the major contributions of this paper can be summarized as follows. + +- We propose an effective task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. To the best of our knowledge, it is the first work to develop a paradigm for VI recognition on edge devices with an extremely low computation budget. +- An effective task-oriented pretraining strategy is proposed to enhance the heterogeneous feature learning capacity of lightweight networks with task-oriented augmentations and the proposed fake domain loss. + +- A fine-grained dependency reconstruction module is designed to mine cross-modality dependencies. +- Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods on mainstream VI-ReID and VI face recognition datasets by a remarkable margin and extremely low complexity. + +# 2. Related Work + +Visible-Infrared Face Recognition (VI-FR). It aims to match face images across visible and infrared modalities. Initially, handcraft features are designed for VI-FR. In [3,9], authors utilize the Local Radon Binary Pattern (LRBP) as the general representation to perform identity matching. Difference-of-Gaussian (DoG) filtering and Multiscale Block Local Binary Patterns (MB-LBP) are widely adopted [20, 50] to acquire robust representations for VI-FR. Recently, deep learning methods have achieved considerable success in VI-FR, focusing on learning modality-shared features with CNN networks [13, 14, 22] or disentangling the modality-invariant representations with GAN networks [46]. In [37], a dual-generation method is proposed to disentangle modality-invariant patterns from visible and infrared images, which helps relieve the modality discrepancy. After that, Duan et al. [8] proposed a pose-aligned cross-spectral hallucination (PACH) network to eliminate identity-independent patterns at multiple stages. + +Visible-Infrared Person Re-identification (VI-ReID). It aims to match visible and infrared pedestrian images according to the feature similarity ranking [47]. The zeropadding scheme [35] was first proposed to handle VI-ReID. In the meantime, they established the first large-scale VI-ReID dataset, named SYSU-MM01. Afterwards, a bidirectional center-constrained network [39] was presented to simultaneously optimize intra- and inter-modality discrepancies. To help the network learn modality-invariant patterns, X-modality [19] is introduced to bridge the gap between visible and infrared modality. Recently, a dense alignment learning method [26] was presented to establish the cross-modality feature correspondence at the pixel level. Ye et al. [38] introduced an attentive graph to discover tri-level relations between visible and infrared images. In addition, several GAN-based methods implement cross-modality alignment [29, 30] or modality compensation [31, 45] to help the network learn robust and compact knowledge. However, despite the promising results, these works are too complex to be utilized in the real world. + +# 3. Methodology + +# 3.1. Overview of The Proposed Method + +The major pipeline of the proposed TOPLight is illustrated in Fig. 2, which can be summarized as three steps. + +![](images/4f675ff8de04960db9699beba0a92bcf39bdd9729b5f57efb858609afa3590a6.jpg) +Figure 2. The overall pipeline of our method. We take three steps to adapt a lightweight network to the VI recognition task. + +Step-1. To adapt the lightweight networks to VI recognition scenarios, we construct a dual-path network based on the lightweight model trained on ImageNet-1k (similar in [26]) and further train the dual-path network on ImageNet-mini [18] with task-oriented augmentations. The identity consistency loss $(L_{idc})$ and fake domain loss $(L_{fd})$ are designed to supervise the domain conflict learning process. + +Step-2. We utilize the uniform soup scheme [34] to unify two pre-embedding block groups. This step aims to tune the structure of the pretrained network suited for VI training, which preserves two separate blocks (Block1-1 and Block1-2) to extract low-level features from two modalities. + +Step-3. In this stage, the pretrained dual-path network is further trained on VI recognition datasets. A novel fine-grained dependency reconstruction (FDR) module is integrated into the network, which can construct fine-grained pattern dependencies between visible and infrared features. + +# 3.2. Task-oriented Pretraining Strategy + +The core motivation of the task-oriented pretraining strategy is to enable lightweight networks to learn prior knowledge related to cross-modality matching during the pretraining stage. Specifically, we first transform the lightweight network trained on the ImageNet-1k dataset into the dual-path network, which consists of two stem blocks, two pre-embedding block groups and one post-embedding block group. Then, the dual-path network is further trained on the ImageNet-mini dataset. During this training, task-oriented data augmentation (DA) is utilized to create visual differences in training samples, which intends to simulate the discrepancy between visible and infrared + +![](images/fcbfacdd4e2c6c060f490d69c6ab24edbbb0b9275bddba7c85d5754619973a86.jpg) +Figure 3. The effects of each data augmentation. + +images and enforce the network to pay more attention to the identity-aware but colour-irrelevant patterns. Thus, we design the generic, colour and texture data augmentation (DAs) to achieve the training process, as shown in Fig. 3. + +For the generic DAs, we increase the overall sample diversity by using random crop, flip and deformation. In the colour DAs, we randomly perform colour jitter, channel shuffle and RGB shift for each sample in the Block1-1 path. The colour DAs are utilized to disturb the strong colour prior knowledge learned from the ImageNet-1k dataset, forcing the network to pay more attention to the structural patterns of visible images (which also exist in infrared images), not just colour patterns. In the texture DAs, we randomly perform compression, defocusing and sharpenGray for each sample in the Block1-2 path, which intends to remove the colour information and simulate the terrible imaging quality under night surveillance conditions. During the TOP, we combine these three DAs to create visual differences that are close to the VI recognition scenes. + +However, in the above manner, the network can lazily learn from one path to avoid embedding visually discrepant samples. Meanwhile, the sample differences generated by augmentations are insufficient to simulate the serious domain conflict in VI recognition scenes. Thus, we develop a fake domain loss to aggravate the variation of mid-level features during the training and motivate the network to discover shared patterns among them. + +Assume $\mathbf{x}$ denotes a raw sample from the ImageNet-mini dataset; after augmentations and feature learning, the $\mathbf{x}_1$ and $\mathbf{x}_2$ denote the outcome features of Block4-1 and Block4-2, respectively. $\mathbf{y}$ is the shared ground-truth label. We concatenate $\mathbf{x}_1$ and $\mathbf{x}_2$ and feed them into Block5. The output of Block7 is denoted as $\hat{\mathbf{x}} = \{\hat{\mathbf{x}}_1, \hat{\mathbf{x}}_2\}$ . Based on the above definitions, the fake domain loss can be defined by: + +$$ +L _ {f d} ^ {1} = L _ {d} \left(\mathbf {x} _ {1}, \mathbf {d} _ {1}\right) + L _ {d} \left(G R L \left(\hat {\mathbf {x}} _ {1}\right), \mathbf {d} _ {1}\right), \tag {1} +$$ + +$$ +L _ {f d} ^ {2} = L _ {d} \left(\mathbf {x} _ {2}, \mathbf {d} _ {2}\right) + L _ {d} \left(G R L \left(\hat {\mathbf {x}} _ {2}\right), \mathbf {d} _ {2}\right), \tag {2} +$$ + +where $\mathbf{d}_1$ and $\mathbf{d}_2$ are the corresponding pseudo domain labels for $\mathbf{x}_1$ and $\mathbf{x}_2$ . We use the domain labels to pretend that $\mathbf{x}_1$ is from one domain and $\mathbf{x}_2$ is from the other. $L_{d}$ denotes the domain classification loss implemented with the logistic classifier. $GRL$ is the gradient reversal layer proposed in [10], which aims to reverse the optimization target of the final domain classifier. In this manner, $L_{fd}^{1}$ and $L_{fd}^{2}$ generate two contradictory learning procedures: self-domain constraints are individually set on $\mathbf{x}_1$ and $\mathbf{x}_2$ , which enforce them to be representative and specific for the two fake domains we pretended, respectively. The inverted-domain constraints are set on the final features after Block7 ( $\hat{\mathbf{x}}_1$ and $\hat{\mathbf{x}}_2$ ) by reversing gradients, which urges them to be similar so as to confound the domain classification loss $L_{d}$ . + +During this procedure, Block(2,3,4)-1 and Block(2,3,4)-2 are trained to extract two types of strongly distinguished "heterogeneous features". In comparison, Block(5,6,7) are trained to embed these discrepant features and learn their shared patterns. We manually create this conflict to force the lightweight network to learn how to represent heterogeneous features and discover shared patterns from them. + +To ensure that the embedded features are identity-related, the identity consistency loss $(L_{idc})$ is introduced as: + +$$ +L _ {i d c} = \log [ P (y | p (\hat {\mathbf {x}} _ {1})) ] + \log [ P (y | p (\hat {\mathbf {x}} _ {2})) ], \tag {3} +$$ + +where $p(\hat{\mathbf{x}}_1)$ and $p(\hat{\mathbf{x}}_2)$ denote the identity prediction results from the same linear classifier. $\mathbf{y}$ is the ground-truth label for both $\mathbf{x}_1$ and $\mathbf{x}_2$ . Here, we utilize one classifier to predict two conflicted features with shared logits, aiming to further improve the embedding capacity of Block(5,6,7). + +The overall loss in the TOP is written as follows. + +$$ +L = L _ {i d c} + L _ {f d} ^ {1} + L _ {f d} ^ {2}. \tag {4} +$$ + +After the pretraining, we utilize the uniform soup scheme to unify two groups of pre-embedded blocks in turn (Step-2 in Fig. 2). The final lightweight network for VI-recognition training consists of two independent stem blocks (Block1-1 and 1-2), one pre-embedding block group (Block2,3,4) and one post-embedding block group (Block5,6,7), as shown in the Step-3 of Fig. 2. + +# 3.3. Fine-grained Dependency Reconstruction + +After the task-oriented pretraining, the lightweight network learns how to extract modality-invariant features from heterogeneous images. However, it lacks the ability to discover fine-grained patterns across the modalities. Therefore, we introduce a practical fine-grained dependency reconstruction module to help the lightweight network discover plentiful fine-grained patterns by constructing cross-modality feature dependencies in the VI training phase. + +Specifically, let us denote $S^v = \{\mathbf{x}_i^v | i = 1,2,\dots,N^v\}$ and $S^r = \{\mathbf{x}_i^r | i = 1,2,\dots,N^r\}$ as the visible and infrared training sets, respectively; $N^v$ and $N^r$ indicate the total numbers of visible and infrared samples, respectively. During the training, we feed $\mathbf{x}^v$ and $\mathbf{x}^r$ into Block1-1 and Block1-2 to extract low-level features of two modalities, respectively. Then, we concatenate these low-level features and feed them into the Block(2,3,4) for the initial feature embedding. Based on the outcome of Block4 $(\mathbf{Z} = \{\mathbf{Z}^v,\mathbf{Z}^r\})$ , the fine-grained and cross-modality feature dependencies are constructed to help the network learn substantial modality-shared patterns by the FDR module. + +In order to extract fine-grained patterns from feature maps, we slice $\mathbf{Z}$ into parts along the horizontal and vertical directions. For each spatial map $\{\mathbf{Z}\}_{m=0}^{C-1} \in \mathbb{R}^{H \times W}$ , the horizontal patterns are obtained by: + +$$ +\mathbf {h} ^ {f} = \operatorname {C o n c a t} [ \operatorname {W i n d o w} (\{\mathbf {Z} ^ {f} \} _ {m = 0} ^ {C - 1}) ], f \in \{v, r \}, \tag {5} +$$ + +where $\mathbf{Z}^f\in \mathbb{R}^{C\times H\times W},f\in \{v,r\}$ denote visible or infrared features. $C,H,W$ are the channel, height, and width dimensions, respectively. $\mathbf{h}^f\in \mathbb{R}^{C\times \frac{H}{\sigma}\times \sigma \times W}$ denotes a group of fine-grained horizontal patterns extracted from $\mathbf{Z}^f$ Similarly, the vertical patterns $\mathbf{v}^f\in \mathbb{R}^{C\times \frac{W}{\mu}\times H\times \mu}$ can also be obtained by: + +$$ +\mathbf {v} ^ {f} = \operatorname {C o n c a t} \left[ \operatorname {W i n d o w} \left(\left\{\mathbf {Z} ^ {f} \right\} _ {m = 0} ^ {C - 1}\right) \right], f \in \{v, r \}, \tag {6} +$$ + +where $\mu$ and $\sigma$ are used to control the slice granularity. + +Further, the spatial information of $\mathbf{h}^f$ and $\mathbf{v}^f$ is condensed by the average pooling operation, which produces $\hat{\mathbf{h}}^f\in \mathbb{R}^{C\times \frac{H}{\sigma}}$ and $\hat{\mathbf{v}}^f\in \mathbb{R}^{C\times \frac{W}{\mu}}$ . In the above manner, we refine the original spatial maps $\{\mathbf{Z}^f\}_{m = 0}^{C - 1}\in \mathbb{R}^{H\times W}$ into directional tensors $\hat{\mathbf{h}}^f$ and $\hat{\mathbf{v}}^f$ to cut off the original self-modality dependencies in spatial dimensions. Furthermore, + +the cross-modality dependencies are established as follows. + +$$ +\mathbf {Z} _ {h} ^ {f} = \underset {\left(\frac {H}{\sigma} \times W\right)} {\operatorname {T r a n s c o n v}} (\hat {\mathbf {h}} ^ {f}), f \in \{v, r \}, \tag {7} +$$ + +$$ +\mathbf {Z} _ {v} ^ {f} = \underset {\left(H \times \frac {W}{\mu}\right)} {\operatorname {T r a n s C o n v}} (\hat {\mathbf {v}} ^ {f}), f \in \{v, r \}. \tag {8} +$$ + +In Eq. (7) and Eq. (8), we use two modality-shared transposed convolution layers to reconstruct the spatial maps of $\mathbf{h}^f$ and $\mathbf{v}^f$ , respectively. This process aims to establish the cross-modality spatial dependencies. The outputs are $\mathbf{z}_h^f \in \mathbb{R}^{C \times H \times W}$ and $\mathbf{z}_v^f \in \mathbb{R}^{C \times H \times W}$ . + +However, there exist "scheme stereotypes" in $\mathbf{Z}_h^f$ and $\mathbf{Z}_v^f$ . The main reason is that these features are learned from two distinct directional schemes, which may lead to a skewed learning procedure. In contrast, we hope these two schemes can consistently provide diversified patterns that cover more valuable signals. Thus, the shuffle attention is designed to fuse $\mathbf{Z}_h^f$ and $\mathbf{Z}_v^f$ while decoupling them from two directions. Specifically, we concatenate them on the channel level and then shuffle the channel order to disarrange the spatial maps learned from two directions. + +$$ +\mathbf {D} ^ {f} = \operatorname {S h u f f l e} \left(\operatorname {C o n c a t} \left(\mathbf {Z} _ {h} ^ {f}, \mathbf {Z} _ {v} ^ {f}\right)\right), f \in \{v, r \}, \tag {9} +$$ + +where $\mathbf{D}^f\in \mathbb{R}^{2C\times H\times W}$ denotes the concatenated feature with disarranged channel order. The new channel dependencies are established by weighting all the spatial maps. + +$$ +\hat {\mathbf {D}} ^ {f} = \mathbf {W} _ {2} R e L U \left(\mathbf {W} _ {1} G A P \left(\mathbf {D} ^ {f}\right)\right), f \in \{v, r \}, \tag {10} +$$ + +where $\mathbf{W}_1\in \mathbb{R}^{2C\times \frac{C}{4}}$ and $\mathbf{W}_2\in \mathbb{R}^{\frac{C}{4}\times C}$ indicate projection matrices to compute the channel attention [16]. Finally, we introduce $\hat{\mathbf{D}}^f$ to refine the original features. + +$$ +\hat {\mathbf {Z}} ^ {f} = \hat {\mathbf {D}} ^ {f} \oplus \mathbf {Z} ^ {f}, f \in \{v, r \}, \tag {11} +$$ + +where $\hat{\mathbf{Z}}^f$ is used as the input of the next block; $\oplus$ is the element-wise add operation. The proposed fine-grained dependency reconstruction module can effectively establish fine-grained and cross-modality dependencies in both visible and infrared features, which can help discover substantial small patterns that are vital in VI recognition. + +# 3.4. Loss Function + +We utilize the hard-mining triplet loss $(L_{tri})$ [42] and cross-entropy loss $(L_{id})$ for metric learning and identity learning, respectively. The overall loss can be defined as: + +$$ +L = L _ {i d} + \gamma * L _ {t r i}, \tag {12} +$$ + +where $\gamma$ is the parameter to balance each loss term. + +# 4. Experimental Results + +# 4.1. Datasets and Settings for VI-ReID + +We first evaluate the proposed method for the visible-infrared person re-identification (VI-ReID) task on two public benchmarks: SYSU-MM01 [35] and RegDB [25]. + +SYSU-MM01 [35] is the largest VI-ReID dataset, which contains 491 persons captured from six different cameras across scenes (indoor and outdoor). There are a total of 287628 visible images and 15792 infrared images. Following the evaluation protocol in [26,38], 395 persons are fixed for training and 96 persons for testing. The testing process includes all-search and indoor-search modes. + +RegDB [25] contains 412 persons, and each person has 10 visible and 10 infrared images. "Visible to infrared" and "infrared to visible" are two testing modes. The former uses visible as the query and infrared as the gallery. The latter does the opposite. Following [26, 30, 38], the final performance on the RegDB is made with the average results of 10 times the training and testing procedure. Each time, we randomly sample 206 identities for training, and the remaining 206 identities are used for testing. + +Evaluation Metrics. We use the standard cumulative matching characteristics (CMC), mean average precision (mAP), and mean inverse negative penalty (mINP) [42] as the evaluation metrics. + +Implementation Details. For the ablation studies, we select the MobileNetV3-L [15] as the lightweight baseline to validate each proposed component. As shown in Table 1, we encapsulate all the layers of MobileNetV3-L as Block1-7 and then initialize the first four blocks twice to construct the dual-path network. During the TOP, detailed settings are + +Table 1. Detailed structures for each Block. We package the entire MobileNetV3-Large into Block1-7 without overlap. + +
Block partitions on MobileNetV3-large
Layer NameStructuresOutput Size
Block1Conv(3×3, 2), Bneck(3×3, 1)16 × 1122
Block2Bneck(3×3, 2)24 × 562
Block3Bneck(3×3, 1)24 × 562
Block4Bneck(5×5, 2), Bneck(5×5, 1)×240 × 282
Block5Bneck(3×3, 2), Bneck(3×3, 1)×5112 × 142
Block6Bneck(5×5, 2), Bneck(5×5, 1)×2160 × 72
Block7Conv(1×1, 1)×31280 × 72
+ +shown in Table 2(b). It is worth noticing that the "Random pick" means that for each image sample used in the Block1-1 path, we randomly pick one method from colour jitter, channel shuffle, or RGB shift to augment it. + +During the VI-ReID training, detailed settings are presented in Table 2(a). Briefly, we randomly sample 6 identities with 6 visible and 6 infrared images per identity, which makes a total of 72 images in each batch. All images are resized to $288 \times 144$ . Random erasing (RE) [49], flipping, + +Table 2. Detailed training and augmentation settings. + +
(a): Training settings
SettingsOn ImageNet-1KOn ImageNet-miniOn VI-ReID
Epochs1505070
Batch size51236×236×2
Image size224×224224×224288×144
Augs.Stanard [15]Task-orientedRE, CE, Flip
OptimizerSGDSGDAdamW [23]
Lr.0.20.10.001
Decaycosinecosineconstant
Warm upFirst 5 epochNoneFirst 5 epoch
Weight decay5E-45E-51E-2
PlatformPytorch, FP16PytorchPytorch
GPUTesla V100RTX 3060RTX 3060
Cost57 h4 h-
+ +
(b): Task-oriented augmentation settings
AugmentationsProbabilityTarget
Generic (G)Crop & Filp0.5All path
Deformation0.2
Colour (C)Colour jitterRandom pickBlock1-1 path
Channel shuffle
RGB shift
Texture (T)Compression0.2Block1-2 path
Defocusing0.2
Sharpen gray1
+ +and channel erasing (CE) [40] are utilized against overfitting. The AdamW [23] optimizer with a learning rate (Lr) of 0.001, and a weight decay of 0.01 is adopted to train our network for 70 epochs. We use the Lr warm-up in the first 5 epochs and constant decay at the $20^{th}$ and $40^{th}$ epochs with a decay factor of 0.1. The parameters $\mu$ and $\sigma$ are set to 3 and 2 based on the fine-tuning results, respectively. The balance parameter $\gamma$ in Eq. (12) is empirically set to $1/2$ . + +# 4.2. Ablation Study + +Effect of Task-oriented Pretraining. We first evaluate the effect of task-oriented pretraining on VI-ReID, as shown in Table 3. No.1 indicates the baseline performance obtained via direct fine-tuning of the ImageNet-1k pretrained model in VI-ReID datasets. When only using generic augmentations and identity loss to perform TOP on ImageNetmini, the "catastrophic forgetting" makes No.2 perform worse than the baseline. However, when applying the proposed task-oriented augmentations, the network learns to extract stable features from visually discrepant images in advance, thus improving the downstream VI-ReID performance (No.3). Coupled with $L_{idc}$ and $L_{fd}$ , we create a conflicting learning process that enforces the network to learn how to represent and embed heterogeneous features. No.5 shows significant improvements in both the SYSU-MM01 and RegDB datasets. By only using the MobileNetV3-large backbone, the performance reported in No.5 has already exceeded many ResNet-based works [19, 26, 38]. + +Effect of Fine-grained Dependency Reconstruction Module. In Table 3, No.6-No.9 show the ablation experiments about the fine-grained dependency reconstruction module. We intend to construct the cross-modality feature dependencies, and the first issue is how to model the + +original feature relations. No.6 indicates directly applying the coarse-grained features to the shuffle attention module. The performance is slightly improved. When using the proposed directional scheme $(H_{s},V_{s})$ to extract fine-grained patterns and reconstruct their spatial relation using transposed convolution layers, No.7 and No.8 show meaningful improvements. The full power of the FDR module is shown in No.9, which adopts both fine-grained horizontal and vertical patterns to discover cross-modality feature similarities, achieving $66.14\%$ and $63.80\%$ in terms of Rank-1 and mAP performance on SYSU-MM01 under the all-search mode. + +Pointing at the difficulties in VI recognition tasks, we design the TOP strategy and FDR module to remedy the drawbacks of the lightweight network. Compared with the baseline (No.1), our method (No.9) considerably improves all metrics on two benchmarks. Meanwhile, the ablations reported in Table 3 demonstrate that all the proposed components bring advantages consistently to the accuracy gain. + +Applicability with Lightweight Networks. We evaluate our methods on three mainstream lightweight networks (ShuffleNetV2 [24], GhostNet [11] and MobileNetV3 [15]). As shown in Table 4, our method can remarkably improve VI Re-ID performance with ignorable complexity growth. Compared with conventional deep networks, we surpass them significantly with minimal complexity. + +Detailed Ablations on the FDR Module. We validate the rationality of each component in the FDR module. Table 5(a) shows the impacts of different spatial modelling methods. Except for ours, the HAP [38] achieves the best performance. However, we extract patterns from both the horizontal and vertical regions and use up-sampling (u.) to discover modality-shared cues from re-enlarged spatial areas, which can be regarded as a more effective solution. + +Table 5(b) shows the impacts of different channel relation reasoning methods. Our shuffle attention (SA) wins both SE [16] and CBAM [33]. The recipe is the channel shuffle (cs.) operation that builds scheme-decoupled dependencies. This crucial step can evenly correlate visible and infrared features from each fine-grained pattern. + +# 4.3. Feature Visualization + +What Does TOP Do? We randomly track 10 identities during the TOP stage and visualize the feature distribution via T-SNE [28], as illustrated in Fig. 4. + +We first use task-oriented augmentation to create two types of visual discrepant samples from each image, as shown in Fig. 4(a) and Fig. 4(b). Then, the proposed fake domain loss forces the features after Block4 to be domain-specific, as shown in Fig. 4(c). Concurrently, it forces the network to learn to embed those domain-specific features and provide domain-shared predictions. In Fig. 4(d), after the training, we can observe that the network knows how to extract the domain-shared representations and successfully + +Table 3. Evaluation of each proposed component on two VI-ReID datasets. "Augs." indicates the augmentations. G, C and T denote the generic, colour, and texture augmentations, respectively. In the FDR module, $H_{s}$ and $V_{s}$ denote the horizontal and vertical slices with up-sampling. SA is the shuffle attention module. Rank (r) (\%), mAP (\%) and mINP (\%) are reported. + +
No.Task-oriented pretraining stageVI training stageSYSU-MM01 (all-search)RegDB (visible-to-infrared)
Augs.Loss functionsFDR module
GC+TLidLidcLfdHsVsSAr=1r=10mAPmINPr=1r=10mAPmINP
147.8189.7147.0633.4871.2689.9465.6648.50
2✓✓43.2885.9645.5631.1070.7388.5265.6548.41
3✓✓✓49.8589.7447.5635.5271.3289.9165.6748.49
4✓✓✓54.2892.1152.9441.2975.3192.6468.7852.16
5✓✓✓✓62.4194.1259.0645.1382.7594.1376.2161.84
6✓✓✓✓✓62.8994.2659.7945.8482.8894.1976.2461.92
7✓✓✓✓✓✓63.9595.2860.0946.8083.0794.4876.5562.20
8✓✓✓✓✓✓64.0495.4161.1246.9283.2294.6977.0163.16
9✓✓✓✓✓✓✓66.1496.0363.8049.7684.1594.9879.2663.86
+ +Table 4. Experimental results on different lightweight networks and conventional deep networks. + +
MethodsFLOPs (M)SYSU-MM01RegDB
r=1mAPr=1mAP
ConventionResNet-50356256.9854.7276.8671.30
ConvNeXt-Tiny362058.7255.3178.2572.64
Vit-B568952.1751.8175.3170.37
Swin-Tiny328758.2455.1678.3972.68
LightweightShuffleNetV2-1.0×13941.8841.9467.8364.85
+TOP & FDR17755.7152.6379.8266.36
ShuffleNetV2-1.5×26547.3947.8170.1565.28
+TOP & FDR37163.3560.8184.1376.98
GhostNet-1.0×15042.5342.9471.2864.40
+TOP & FDR18958.5455.1983.2677.16
GhostNet-1.3×28150.8947.9272.5165.98
+TOP & FDR39566.7664.0185.5179.95
MobileNetV3-S10440.9242.5162.7758.31
+TOP & FDR13054.7550.2675.5370.17
MobileNetV3-L25047.8147.0671.2665.66
+TOP & FDR36266.1463.8084.1579.26
+ +Table 5. Evaluation of spatial modelling methods and channel relation reasoning methods in the FDR module. "u." and "cs." respectively denote the up-sampling and channel shuffle operations. + +
(a): Impact of different spatial modelling methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
GAP62.8959.7945.8482.8876.2461.92
Context [2]61.3757.5246.0881.5674.4962.11
HAP [38]62.9258.1347.8482.9876.2062.77
HS+VS(w/o u.)63.4559.7549.6183.9478.8263.21
HS+VS66.1463.8049.7684.1579.2663.86
(b): Impact of different channel relation reasoning methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
SE [16]62.9159.7045.7882.7977.4562.34
CBAM [33]61.7956.8845.2583.4177.2862.96
SA (w/o cs.)62.8158.1045.7682.7576.1262.56
SA66.1463.8049.7684.1579.2663.86
+ +embed all the features into their respective identity groups. Thus, the proposed task-oriented pretraining strategy makes lightweight networks fit cross-modality tasks. + +Pattern Visualization. As demonstrated in Fig. 5, we visualize the learned patterns using Grad-CAM [27] to ex + +![](images/2baa24b60df4fee86f2fd1cb76f7bd15817972db5f26cfa5f8570c8b13645c16.jpg) +Figure 4. Feature distributions during task-oriented pretraining. Each colour denotes an identity. In (a), the star markers represent raw samples. In (b), (c) and (d), the circle and triangle markers represent colour and texture-augmented samples, respectively. + +![](images/ee5eac419acd4b332838bda4aef0870d29194e08c82b3d3e8ee87875cce38cbc.jpg) +Figure 5. Visualization of learned patterns via Grad-CAM [27]. For a clear presentation, we average all the spatial maps into four. plain the effect of the FDR module. Clearly, the original patterns extracted by MobileNetV3-L are concentrated and overfit on the modality-specific parts. To remedy this issue, we use the FDR module to discover modality-shared patterns from fine-grained regions, which separates learned patterns to focus on considerable modality-shared cues. A good case is shown in Fig. 5. The visible patterns initially learned are concentrated in the high chest area, while the infrared patterns are all in the right arm. After re-weighting by the FDR module, modality-shared small regions like the face, arms, neck, and shoes are all taken into account, thus + +Table 6. Comparison with the state-of-the-arts on SYSU-MM01 [35]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported. + +
DetailsAll-searchIndoor-search
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356214.8054.1271.3315.95-20.5868.3885.7926.92-
JSIA [30]ResNet50+GAN>413338.1080.7089.9036.90-43.8086.2094.2052.90-
AGW [42]ResNet50>356247.5084.3992.1447.6535.3054.1791.1495.9862.9759.23
X-Modal [19]ResNet50>356249.9089.8096.0050.70------
DMiR [38]ResNet50>356250.5488.1294.8649.29-53.9292.5097.0962.49-
FBP-AL [32]ResNet50>356254.1486.0493.0350.20------
DDAG [41]ResNet50>356254.7590.3995.8155.0239.6261.0294.0698.4167.9862.61
HAT [43]ResNet50>356255.2992.1497.3653.89-62.1095.7599.2070.84-
LBA [26]ResNet50>356255.41--54.14-58.46--66.33-
TSME [21]ResNet50>356264.2395.1998.7361.21-64.8096.9299.3171.53-
SPOT [4]ResNet50+ViT>481065.3492.7397.0462.2548.8669.4296.2299.1274.6370.48
TOLight (Ours)MobileNetV3-L= 36266.1496.0397.6863.8049.7672.4197.5499.2376.1171.43
TOLight (Ours)GhostNet-1.3×= 39566.7696.2398.7064.0150.1872.8997.9399.2876.7071.95
+ +Table 7. Comparison with the state-of-the-arts on RegDB [25]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported. + +
DetailsVisible-to-InfraredInfrared-to-Visible
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356217.7534.2144.3518.90-16.6334.6844.2517.82-
JSIA [30]ResNet50+GAN>413348.50--49.30-48.10--48.90-
AGW [42]ResNet50>356270.0586.2191.5566.3750.1970.4987.2191.8465.9051.24
X-Modal [19]ResNet50>356262.2183.1391.7260.18------
DMiR [38]ResNet50>356275.7989.8694.1869.97-73.9389.8793.9868.22-
FBP-AL [32]ResNet50>356273.9889.7193.6968.24-70.0589.2293.8866.61-
DDAG [41]ResNet50>356269.3486.1991.4963.4649.2468.0685.1590.3161.8048.62
HAT [43]ResNet50>356271.8387.1692.1667.56-70.0268.4591.6166.30-
LBA [26]ResNet50>356274.17--67.64-72.43--65.46-
SPOT [4]ResNet50+ViT>481080.3593.4896.4472.4656.1979.3792.7996.0172.2656.06
GECNet [48]ResNet50+GAN>435082.3392.7295.4978.45-78.9391.9995.4475.58-
TOLight (Ours)MobileNetV3-L=36284.1594.9896.5879.2663.8680.9492.8596.3776.1059.33
TOLight (Ours)GhostNet-1.3×=39585.5194.9996.7079.9563.8580.6592.8196.3275.9159.26
+ +improving the accuracy of the cross-modality matching. + +# 4.4. Comparison with State of The Arts + +We compare our method with recently proposed state-of-the-art VI-ReID methods, as shown in Table 6 and Table 7. They illustrate that we remarkably and efficiently exceed all the compared SOTAs under diverse evaluation settings. Specifically, based on the tiny GhostNet- $1.3\times$ backbone, we attain the rank-1 of $66.76\%$ and mAP of $64.01\%$ on the all-search mode of the large-scale SYSU-MM01 dataset with only 395 FLOPs. The proposed method also achieves excellent results on the RegDB dataset, surpassing the previous SOTAs by a significant margin with lower FLOPs. + +Table 8. Evaluation on two VI-FR datasets. CA is channel augmentation [40]. B is the LightCNN-29 baseline. Rank at 1 accuracy $(\%)$ and false acceptance rate (F: $\%$ ) are reported. + +
MethodsOulu [5]BUAA [17]
r=1F:1%F:0.1%r=1F:1%F:0.1%
IDR [13]94.373.446.294.393.484.7
VSA [44]99.996.882.398.098.292.5
PACH [8]10097.988.298.698.093.5
B [36]10097.987.098.097.793.7
B+CA [40]10098.991.798.398.294.5
B+TOP10098.891.598.398.194.5
B+TOP+FDR (Ours)10098.991.798.398.294.6
+ +# 4.5. Visible-Infrared Face Recognition + +We follow [40] to examine our method on Oulu-CASIA NIR-VIS [5] and BUAA-VisNir face databases [17]. Fol + +lowing previous studies, we also adopt the LightCNN-29 [36] as our baseline network (B). Except for IDR [13], VSA [44], and PACH [8], we also compare recently proposed channel augmentation (CA) [40] with our method on the same baseline. Results are displayed in Table 8. + +Evidently, the proposed TOP and FDR also enhance the performance of VI face recognition. These results further validate the effectiveness and generalization of our methods, which can solve VI recognition tasks effectively. + +# 5. Conclusion + +This work presents an effective task-oriented pretrained lightweight neural network (TOPLight) to solve visible-infrared recognition problems. First, the task-oriented pretraining strategy significantly improves the ability of lightweight networks to understand heterogeneous images by introducing domain conflict and sample variation during the pretraining phase. Second, the fine-grained dependency reconstruction module is utilized to destroy the modality-specific pattern dependencies and construct the fine-grained and modality-shared pattern dependencies between visible and infrared features. Extensive experiments demonstrate the superiority and effectiveness of the proposed method as well as the efficacy of each component of our framework. + +Acknowledgements. This work is funded by the National Natural Science Foundation of China (Grant No. 61802058, 61911530397). + +# References + +[1] Sun, et.al. Learning part-based convolutional features for person re-identification. TPAMI, 2019. 2 +[2] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 7 +[3] Srinivasa Rao Chalamala, Krishna Rao Kakkirala, and Jami Santosh Kumar. Face recognition using spatial pyramid matching and lrbp. In 2014 IEEE 10th International Colloquium on Signal Processing and its Applications, pages 67-70. IEEE, 2014. 2 +[4] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Jianguo Jiang, and Chia-Wen Lin. Structure-aware positional transformer for visible-infrared person re-identification. IEEE Transactions on Image Processing, 31:2352-2364, 2022. 1, 8 +[5] Jie Chen, Dong Yi, Jamei Yang, Guoying Zhao, Stan Z Li, and Matti Pietikainen. Learning mappings for face synthesis from near infrared to visual light images. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 156-163. IEEE, 2009. 8 +[6] Changxing Ding and Dacheng Tao. A comprehensive survey on pose-invariant face recognition. ACM Transactions on intelligent systems and technology (TIST), 7(3):1-42, 2016. 1 +[7] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 1 +[8] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7930-7938, 2020. 1, 2, 8 +[9] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In 2012 19th IEEE International Conference on Image Processing, pages 1837-1840. IEEE, 2012. 2 +[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 4 +[11] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 1, 6 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[13] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Learning invariant deep representation for nir-vis face recognition. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 1, 2, 8 + +[14] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Wasserstein cnn: Learning invariant features for nir-vis face recognition. IEEE transactions on pattern analysis and machine intelligence, 41(7):1761-1773, 2018. 2 +[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 5, 6 +[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5, 6, 7 +[17] D Huang, J Sun, and Y Wang. The buaa-visnir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3:3, 2012. 8 +[18] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 2, 3 +[19] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4610–4617, 2020. 2, 6, 8 +[20] Shengcai Liao, Dong Yi, Zhen Lei, Rui Qin, and Stan Z Li. Heterogeneous face recognition from local structures of normalized appearance. In International Conference on Biometrics, pages 209-218. Springer, 2009. 2 +[21] Jianan Liu, Jialiang Wang, Nianchang Huang, Qiang Zhang, and Jungong Han. Revisiting modality-specific feature compensation for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 8 +[22] Xiaoxiang Liu, Lingxiao Song, Xiang Wu, and Tieniu Tan. Transferring deep representation for nir-vis heterogeneous face recognition. In 2016 International Conference on Biometrics (ICB), pages 1-8. IEEE, 2016. 2 +[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6 +[24] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 1, 6 +[25] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 17(3):605, 2017. 5, 8 +[26] Hyunjong Park, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12046–12055, 2021. 1, 2, 3, 5, 6, 8 +[27] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. + +Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7 +[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6 +[29] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3623-3632, 2019. 1, 2 +[30] Guan-An Wang, Tianzhu Zhang, Yang Yang, Jian Cheng, Jianlong Chang, Xu Liang, and Zeng-Guang Hou. Cross-modality paired-images generation for rgb-infrared person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12144–12151, 2020. 1, 2, 5, 8 +[31] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 618–626, 2019. 2 +[32] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Flexible body partition-based adversarial learning for visible infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 2021. 8 +[33] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 6, 7 +[34] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 3 +[35] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 5380-5389, 2017. 1, 2, 5, 8 +[36] Xiang Wu, Ran He, Zhenan Sun, and Tieniu Tan. A light cnn for deep face representation with noisy labels. IEEE Transactions on Information Forensics and Security, 13(11):2884-2896, 2018. 8 +[37] Xiang Wu, Huaibo Huang, Vishal M Patel, Ran He, and Zhenan Sun. Disentangled variational representation for heterogeneous face recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9005–9012, 2019. 2 +[38] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE Transactions on Information Forensics and Security, 17:386–398, 2021. 1, 2, 5, 6, 7, 8 + +[39] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE Transactions on Information Forensics and Security, 15:407–419, 2019. 2 +[40] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13567-13576, 2021. 6, 8 +[41] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In European Conference on Computer Vision, pages 229–247. Springer, 2020. 8 +[42] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 5, 8 +[43] Mang Ye, Jianbing Shen, and Ling Shao. Visible-infrared person re-identification via homogeneous augmented trimodal learning. IEEE Transactions on Information Forensics and Security, 16:728-739, 2020. 8 +[44] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-posed high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 1, 8 +[45] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7349–7358, 2022. 2 +[46] Teng Zhang, Arnold Wiliem, Siqi Yang, and Brian Lovell. Tv-gan: Generative adversarial network based thermal to visible face recognition. In 2018 international conference on biometrics (ICB), pages 174-181. IEEE, 2018. 2 +[47] Huantao Zheng, Xian Zhong, Wenxin Huang, Kui Jiang, Wenxuan Liu, and Zheng Wang. Visible-infrared person re-identification: A comprehensive survey and a new setting. *Electronics*, 11(3):454, 2022. 2 +[48] Xian Zhong, Tianyou Lu, Wenxin Huang, Mang Ye, Xuemei Jia, and Chia-Wen Lin. Grayscale enhancement colorization network for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1418–1430, 2021. 8 +[49] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 5 +[50] Jun-Yong Zhu, Wei-Shi Zheng, Jian-Huang Lai, and Stan Z Li. Matching nir face to vis face using transduction. IEEE Transactions on Information Forensics and Security, 9(3):501-514, 2014. 2 \ No newline at end of file diff --git a/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/images.zip b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9825f52929f98213a6c82d71383effe74dc2c57 --- /dev/null +++ b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe3c7f7f67c7174d2cddb1b43e12cc08ac17e60992fa8f3681d1074804e517af +size 822969 diff --git a/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/layout.json b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b57db9c4b9f99cce510f868f52c495a3ec3237f4 --- /dev/null +++ b/2023/TOPLight_ Lightweight Neural Networks With Task-Oriented Pretraining for Visible-Infrared Recognition/layout.json @@ -0,0 +1,8630 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 140 + ], + "type": "text", + "content": "TOPLight: Lightweight Neural Networks with Task-Oriented Pretraining for Visible-Infrared Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "spans": [ + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "text", + "content": "Hao Yu" + }, + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "text", + "content": ", Xu Cheng" + }, + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "text", + "content": ", Wei Peng" + }, + { + "bbox": [ + 206, + 160, + 385, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 175, + 531, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 531, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 531, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 175, + 531, + 204 + ], + "type": "text", + "content": "School of Computer Science, Nanjing University of Information Science and Technology, China " + }, + { + "bbox": [ + 62, + 175, + 531, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 175, + 531, + 204 + ], + "type": "text", + "content": "Department of Psychiatry and Behavioral Sciences, Stanford University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 205, + 426, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 205, + 426, + 217 + ], + "spans": [ + { + "bbox": [ + 167, + 205, + 426, + 217 + ], + "type": "text", + "content": "{yuhao,xcheng}@nuist.edu.cn，wepeng@stanford.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "text", + "content": "Visible-infrared recognition (VI recognition) is a challenging task due to the enormous visual difference across heterogeneous images. Most existing works achieve promising results by transfer learning, such as pretraining on the ImageNet, based on advanced neural architectures like ResNet and ViT. However, such methods ignore the negative influence of the pretrained colour prior knowledge, as well as their heavy computational burden makes them hard to deploy in actual scenarios with limited resources. In this paper, we propose a novel task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. Specifically, the TOPLight method simulates the domain conflict and sample variations with the proposed fake domain loss in the pretraining stage, which guides the network to learn how to handle those difficulties, such that a more general modality-shared feature representation is learned for the heterogeneous images. Moreover, an effective fine-grained dependency reconstruction module (FDR) is developed to discover substantial pattern dependencies shared in two modalities. Extensive experiments on VI person re-identification and VI face recognition datasets demonstrate the superiority of the proposed TOPLight, which significantly outperforms the current state of the arts while demanding fewer computational resources." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "content": "Identity recognition technologies have provided numerous reliable solutions for monitoring systems, which strive to match the face (face recognition [6, 7]) or pedestrian (person re-identification [42]) images of the same identity. However, the majority of previous efforts only consider visible images. In real-life practice, many surveillance cameras can switch to infrared imaging mode at night. Thus, the essential cross-modality visible-infrared recognition (VI" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 247, + 531, + 330 + ], + "blocks": [ + { + "bbox": [ + 318, + 247, + 531, + 330 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 531, + 330 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 531, + 330 + ], + "type": "image", + "image_path": "54bc120915c24eb7770ad9b0e317de0c86acda0d77e85122cb490881979aa087.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 330, + 532, + 422 + ], + "blocks": [ + { + "bbox": [ + 317, + 330, + 532, + 422 + ], + "lines": [ + { + "bbox": [ + 317, + 330, + 532, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 330, + 532, + 422 + ], + "type": "image", + "image_path": "6b0ae237ab44d55111eac04d4705e779f55e6177646d15303a0badc10d16678b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 423, + 545, + 468 + ], + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 468 + ], + "type": "text", + "content": "Figure 1. (a) The task-oriented pretraining strategy; (b) Performance comparison of the standard ImageNet-1k pretraining scheme (SP) and the proposed task-oriented pretraining scheme (TOP) on the SYSU-MM01 dataset [35] (All-Search mode)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 474, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 498 + ], + "type": "text", + "content": "recognition) technology has been developed to match the visible and infrared photographs of the same people." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": "Recently, visible-infrared person re-identification (VI-ReID) [26, 38] and visible-infrared face recognition [8, 13, 44] have been widely studied. The key issue is identifying the modality-shared patterns. To this end, several works [29, 30] use generative adversarial networks (GANs) to implement cross-modality alignment at the pixel and feature levels. Others [4, 26, 38] design the dual-path feature extraction network, coupled with inter-feature constraints, to close the embedding space of two modalities. However, these methods utilize at least one pretrained ResNet-50 [12] backbone to extract solid features, which makes them unsuitable for edge monitoring devices. Recent works [4, 38] employ auxiliary models (e.g., pose estimation, graph reasoning) to relieve the modality discrepancy, which enhances the performance on academic benchmarks but reduces the real-time inference speed. Compared with conventional deep networks (e.g., ResNet, ViT), lightweight networks [11, 15, 24] can extract basal features rapidly. In VI recogni" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 237, + 713 + ], + "type": "text", + "content": "*Corresponding Author (Email: xcheng@nuist.edu.cn)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "3541" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 133 + ], + "type": "text", + "content": "tion tasks, however, the vast modality discrepancy renders the performance of lightweight networks significantly inferior to that of conventional deep networks. The main reason is that lightweight networks lack the ability to identify modality-shared patterns from heterogeneous images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 300 + ], + "type": "text", + "content": "To address this issue, we present an effective task-oriented pretraining (TOP) strategy. As shown in Fig. 1(a), we first train a lightweight network on the ImageNet-1k dataset to learn vision prior knowledge. After that, the trained network is transformed into the dual-path network and further trained by using task-oriented data augmentations, identity consistency loss and fake domain loss on the ImageNet-mini dataset [18]. The task-oriented pretraining (TOP) strategy simulates the sample differences in VI scenes and teaches the network how to represent and embed discrepant features. Fig. 1(b) reports the performance of three lightweight networks in the VI-ReID task. Compared with the ImageNet-1k pretraining, our TOP strategy can remarkably improve the baseline performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 407 + ], + "type": "text", + "content": "Another weakness of lightweight networks is that few feature maps are learned from raw images for rapid inference. In the VI recognition scene, it is challenging to discover modality-shared patterns with so few learned feature maps. In practice, the network can focus on a group of aggregated modality-specific patterns that offer the most gradient for identity classification. In contrast, the fine-grained and modality-shared patterns, which are crucial for achieving robust cross-modality matching, are neglected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 286, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 286, + 586 + ], + "type": "text", + "content": "Based on the above observations, we present a novel fine-grained dependency reconstruction (FDR) module to help lightweight networks learn modality-shared and fine-grained patterns. Specifically, inspired by the horizontal slice scheme [1], we first slice feature maps horizontally and vertically to extract fine-grained patterns from diversified local regions. Then, the original spatial dependencies of these patterns are eliminated by using pooling operations. Further, the cross-modality dependencies are built by using up-sampling layers to amplify the modality-shared parts from these patterns. At last, to avoid overfitting, the shuffle attention is designed to re-weight the channel dependencies of all the feature maps, which spreads attention to local patterns as much as possible. In general, the major contributions of this paper can be summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 597, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 597, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 597, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 58, + 597, + 286, + 657 + ], + "type": "text", + "content": "- We propose an effective task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. To the best of our knowledge, it is the first work to develop a paradigm for VI recognition on edge devices with an extremely low computation budget." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 666, + 286, + 713 + ], + "type": "text", + "content": "- An effective task-oriented pretraining strategy is proposed to enhance the heterogeneous feature learning capacity of lightweight networks with task-oriented augmentations and the proposed fake domain loss." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 544, + 152 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 544, + 96 + ], + "type": "text", + "content": "- A fine-grained dependency reconstruction module is designed to mine cross-modality dependencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 105, + 544, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 105, + 544, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 544, + 152 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods on mainstream VI-ReID and VI face recognition datasets by a remarkable margin and extremely low complexity." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 164, + 391, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 164, + 391, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 164, + 391, + 177 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 184, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 399 + ], + "type": "text", + "content": "Visible-Infrared Face Recognition (VI-FR). It aims to match face images across visible and infrared modalities. Initially, handcraft features are designed for VI-FR. In [3,9], authors utilize the Local Radon Binary Pattern (LRBP) as the general representation to perform identity matching. Difference-of-Gaussian (DoG) filtering and Multiscale Block Local Binary Patterns (MB-LBP) are widely adopted [20, 50] to acquire robust representations for VI-FR. Recently, deep learning methods have achieved considerable success in VI-FR, focusing on learning modality-shared features with CNN networks [13, 14, 22] or disentangling the modality-invariant representations with GAN networks [46]. In [37], a dual-generation method is proposed to disentangle modality-invariant patterns from visible and infrared images, which helps relieve the modality discrepancy. After that, Duan et al. [8] proposed a pose-aligned cross-spectral hallucination (PACH) network to eliminate identity-independent patterns at multiple stages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 400, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 639 + ], + "type": "text", + "content": "Visible-Infrared Person Re-identification (VI-ReID). It aims to match visible and infrared pedestrian images according to the feature similarity ranking [47]. The zeropadding scheme [35] was first proposed to handle VI-ReID. In the meantime, they established the first large-scale VI-ReID dataset, named SYSU-MM01. Afterwards, a bidirectional center-constrained network [39] was presented to simultaneously optimize intra- and inter-modality discrepancies. To help the network learn modality-invariant patterns, X-modality [19] is introduced to bridge the gap between visible and infrared modality. Recently, a dense alignment learning method [26] was presented to establish the cross-modality feature correspondence at the pixel level. Ye et al. [38] introduced an attentive graph to discover tri-level relations between visible and infrared images. In addition, several GAN-based methods implement cross-modality alignment [29, 30] or modality compensation [31, 45] to help the network learn robust and compact knowledge. However, despite the promising results, these works are too complex to be utilized in the real world." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 651, + 388, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 388, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 388, + 664 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 670, + 491, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 491, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 491, + 682 + ], + "type": "text", + "content": "3.1. Overview of The Proposed Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "content": "The major pipeline of the proposed TOPLight is illustrated in Fig. 2, which can be summarized as three steps." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3542" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 76, + 516, + 308 + ], + "blocks": [ + { + "bbox": [ + 67, + 76, + 516, + 308 + ], + "lines": [ + { + "bbox": [ + 67, + 76, + 516, + 308 + ], + "spans": [ + { + "bbox": [ + 67, + 76, + 516, + 308 + ], + "type": "image", + "image_path": "4f675ff8de04960db9699beba0a92bcf39bdd9729b5f57efb858609afa3590a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 310, + 518, + 321 + ], + "lines": [ + { + "bbox": [ + 72, + 310, + 518, + 321 + ], + "spans": [ + { + "bbox": [ + 72, + 310, + 518, + 321 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of our method. We take three steps to adapt a lightweight network to the VI recognition task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "type": "text", + "content": "Step-1. To adapt the lightweight networks to VI recognition scenarios, we construct a dual-path network based on the lightweight model trained on ImageNet-1k (similar in [26]) and further train the dual-path network on ImageNet-mini [18] with task-oriented augmentations. The identity consistency loss " + }, + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "type": "inline_equation", + "content": "(L_{idc})" + }, + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "type": "text", + "content": " and fake domain loss " + }, + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "type": "inline_equation", + "content": "(L_{fd})" + }, + { + "bbox": [ + 47, + 332, + 287, + 416 + ], + "type": "text", + "content": " are designed to supervise the domain conflict learning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 477 + ], + "type": "text", + "content": "Step-2. We utilize the uniform soup scheme [34] to unify two pre-embedding block groups. This step aims to tune the structure of the pretrained network suited for VI training, which preserves two separate blocks (Block1-1 and Block1-2) to extract low-level features from two modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 538 + ], + "type": "text", + "content": "Step-3. In this stage, the pretrained dual-path network is further trained on VI recognition datasets. A novel fine-grained dependency reconstruction (FDR) module is integrated into the network, which can construct fine-grained pattern dependencies between visible and infrared features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 550, + 234, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 234, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 234, + 563 + ], + "type": "text", + "content": "3.2. Task-oriented Pretraining Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "The core motivation of the task-oriented pretraining strategy is to enable lightweight networks to learn prior knowledge related to cross-modality matching during the pretraining stage. Specifically, we first transform the lightweight network trained on the ImageNet-1k dataset into the dual-path network, which consists of two stem blocks, two pre-embedding block groups and one post-embedding block group. Then, the dual-path network is further trained on the ImageNet-mini dataset. During this training, task-oriented data augmentation (DA) is utilized to create visual differences in training samples, which intends to simulate the discrepancy between visible and infrared" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 331, + 531, + 467 + ], + "blocks": [ + { + "bbox": [ + 318, + 331, + 531, + 467 + ], + "lines": [ + { + "bbox": [ + 318, + 331, + 531, + 467 + ], + "spans": [ + { + "bbox": [ + 318, + 331, + 531, + 467 + ], + "type": "image", + "image_path": "fcbfacdd4e2c6c060f490d69c6ab24edbbb0b9275bddba7c85d5754619973a86.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 472, + 514, + 482 + ], + "lines": [ + { + "bbox": [ + 336, + 472, + 514, + 482 + ], + "spans": [ + { + "bbox": [ + 336, + 472, + 514, + 482 + ], + "type": "text", + "content": "Figure 3. The effects of each data augmentation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 486, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 533 + ], + "type": "text", + "content": "images and enforce the network to pay more attention to the identity-aware but colour-irrelevant patterns. Thus, we design the generic, colour and texture data augmentation (DAs) to achieve the training process, as shown in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "For the generic DAs, we increase the overall sample diversity by using random crop, flip and deformation. In the colour DAs, we randomly perform colour jitter, channel shuffle and RGB shift for each sample in the Block1-1 path. The colour DAs are utilized to disturb the strong colour prior knowledge learned from the ImageNet-1k dataset, forcing the network to pay more attention to the structural patterns of visible images (which also exist in infrared images), not just colour patterns. In the texture DAs, we randomly perform compression, defocusing and sharpenGray for each sample in the Block1-2 path, which intends to remove the colour information and simulate the terrible imaging quality under night surveillance conditions. During the TOP, we combine these three DAs to create visual differences that are close to the VI recognition scenes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3543" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "However, in the above manner, the network can lazily learn from one path to avoid embedding visually discrepant samples. Meanwhile, the sample differences generated by augmentations are insufficient to simulate the serious domain conflict in VI recognition scenes. Thus, we develop a fake domain loss to aggravate the variation of mid-level features during the training and motivate the network to discover shared patterns among them." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": "Assume " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": " denotes a raw sample from the ImageNet-mini dataset; after augmentations and feature learning, the " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": " denote the outcome features of Block4-1 and Block4-2, respectively. " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": " is the shared ground-truth label. We concatenate " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": " and feed them into Block5. The output of Block7 is denoted as " + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\{\\hat{\\mathbf{x}}_1, \\hat{\\mathbf{x}}_2\\}" + }, + { + "bbox": [ + 46, + 168, + 288, + 251 + ], + "type": "text", + "content": ". Based on the above definitions, the fake domain loss can be defined by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 261, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 261, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 83, + 261, + 287, + 275 + ], + "type": "interline_equation", + "content": "L _ {f d} ^ {1} = L _ {d} \\left(\\mathbf {x} _ {1}, \\mathbf {d} _ {1}\\right) + L _ {d} \\left(G R L \\left(\\hat {\\mathbf {x}} _ {1}\\right), \\mathbf {d} _ {1}\\right), \\tag {1}", + "image_path": "99efc415a95774ca225a314fbdef6318f3466b49bd63186d4b1ba32c6503b1e8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 295, + 287, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 295, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 83, + 295, + 287, + 309 + ], + "type": "interline_equation", + "content": "L _ {f d} ^ {2} = L _ {d} \\left(\\mathbf {x} _ {2}, \\mathbf {d} _ {2}\\right) + L _ {d} \\left(G R L \\left(\\hat {\\mathbf {x}} _ {2}\\right), \\mathbf {d} _ {2}\\right), \\tag {2}", + "image_path": "03ec105b1fa194e9a3554c1375b74bc3b4642c9d1e608fa35c95d7449f47e0f4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_1" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_2" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " are the corresponding pseudo domain labels for " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": ". We use the domain labels to pretend that " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " is from one domain and " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " is from the other. " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " denotes the domain classification loss implemented with the logistic classifier. " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "GRL" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " is the gradient reversal layer proposed in [10], which aims to reverse the optimization target of the final domain classifier. In this manner, " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "L_{fd}^{1}" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "L_{fd}^{2}" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " generate two contradictory learning procedures: self-domain constraints are individually set on " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": ", which enforce them to be representative and specific for the two fake domains we pretended, respectively. The inverted-domain constraints are set on the final features after Block7 (" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_1" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_2" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": ") by reversing gradients, which urges them to be similar so as to confound the domain classification loss " + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 314, + 287, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 481, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 564 + ], + "type": "text", + "content": "During this procedure, Block(2,3,4)-1 and Block(2,3,4)-2 are trained to extract two types of strongly distinguished \"heterogeneous features\". In comparison, Block(5,6,7) are trained to embed these discrepant features and learn their shared patterns. We manually create this conflict to force the lightweight network to learn how to represent heterogeneous features and discover shared patterns from them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 565, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 589 + ], + "type": "text", + "content": "To ensure that the embedded features are identity-related, the identity consistency loss " + }, + { + "bbox": [ + 46, + 565, + 287, + 589 + ], + "type": "inline_equation", + "content": "(L_{idc})" + }, + { + "bbox": [ + 46, + 565, + 287, + 589 + ], + "type": "text", + "content": " is introduced as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 601, + 287, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 601, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 77, + 601, + 287, + 614 + ], + "type": "interline_equation", + "content": "L _ {i d c} = \\log [ P (y | p (\\hat {\\mathbf {x}} _ {1})) ] + \\log [ P (y | p (\\hat {\\mathbf {x}} _ {2})) ], \\tag {3}", + "image_path": "6caaddc12df90db2deae2a597c8f19ed6e80ab0c926fc1524a74015edafdeb4c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "inline_equation", + "content": "p(\\hat{\\mathbf{x}}_1)" + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "inline_equation", + "content": "p(\\hat{\\mathbf{x}}_2)" + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "content": " denote the identity prediction results from the same linear classifier. " + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "content": " is the ground-truth label for both " + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2" + }, + { + "bbox": [ + 46, + 620, + 287, + 678 + ], + "type": "text", + "content": ". Here, we utilize one classifier to predict two conflicted features with shared logits, aiming to further improve the embedding capacity of Block(5,6,7)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 679, + 257, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 679, + 257, + 690 + ], + "spans": [ + { + "bbox": [ + 59, + 679, + 257, + 690 + ], + "type": "text", + "content": "The overall loss in the TOP is written as follows." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 117, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "L = L _ {i d c} + L _ {f d} ^ {1} + L _ {f d} ^ {2}. \\tag {4}", + "image_path": "cf65ee22f169ce9a8eb495cb5edb9cba86d522c3870410efb3163806c72f1177.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "After the pretraining, we utilize the uniform soup scheme to unify two groups of pre-embedded blocks in turn (Step-2 in Fig. 2). The final lightweight network for VI-recognition training consists of two independent stem blocks (Block1-1 and 1-2), one pre-embedding block group (Block2,3,4) and one post-embedding block group (Block5,6,7), as shown in the Step-3 of Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 165, + 523, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 165, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 523, + 178 + ], + "type": "text", + "content": "3.3. Fine-grained Dependency Reconstruction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 185, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 280 + ], + "type": "text", + "content": "After the task-oriented pretraining, the lightweight network learns how to extract modality-invariant features from heterogeneous images. However, it lacks the ability to discover fine-grained patterns across the modalities. Therefore, we introduce a practical fine-grained dependency reconstruction module to help the lightweight network discover plentiful fine-grained patterns by constructing cross-modality feature dependencies in the VI training phase." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": "Specifically, let us denote " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "S^v = \\{\\mathbf{x}_i^v | i = 1,2,\\dots,N^v\\}" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "S^r = \\{\\mathbf{x}_i^r | i = 1,2,\\dots,N^r\\}" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " as the visible and infrared training sets, respectively; " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "N^v" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "N^r" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " indicate the total numbers of visible and infrared samples, respectively. During the training, we feed " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^v" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^r" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " into Block1-1 and Block1-2 to extract low-level features of two modalities, respectively. Then, we concatenate these low-level features and feed them into the Block(2,3,4) for the initial feature embedding. Based on the outcome of Block4 " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "(\\mathbf{Z} = \\{\\mathbf{Z}^v,\\mathbf{Z}^r\\})" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": ", the fine-grained and cross-modality feature dependencies are constructed to help the network learn substantial modality-shared patterns by the FDR module." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "text", + "content": "In order to extract fine-grained patterns from feature maps, we slice " + }, + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "text", + "content": " into parts along the horizontal and vertical directions. For each spatial map " + }, + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}\\}_{m=0}^{C-1} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 426, + 545, + 473 + ], + "type": "text", + "content": ", the horizontal patterns are obtained by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 482, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 482, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 319, + 482, + 545, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {h} ^ {f} = \\operatorname {C o n c a t} [ \\operatorname {W i n d o w} (\\{\\mathbf {Z} ^ {f} \\} _ {m = 0} ^ {C - 1}) ], f \\in \\{v, r \\}, \\tag {5}", + "image_path": "2d015fca48e10a1574dd5907b1993380287c6def56e9937a6ca37fcce94dc51d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^f\\in \\mathbb{R}^{C\\times H\\times W},f\\in \\{v,r\\}" + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "content": " denote visible or infrared features. " + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "inline_equation", + "content": "C,H,W" + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "content": " are the channel, height, and width dimensions, respectively. " + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^f\\in \\mathbb{R}^{C\\times \\frac{H}{\\sigma}\\times \\sigma \\times W}" + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "content": " denotes a group of fine-grained horizontal patterns extracted from " + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^f" + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "content": " Similarly, the vertical patterns " + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^f\\in \\mathbb{R}^{C\\times \\frac{W}{\\mu}\\times H\\times \\mu}" + }, + { + "bbox": [ + 304, + 512, + 545, + 587 + ], + "type": "text", + "content": " can also be obtained by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 596, + 545, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 596, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 319, + 596, + 545, + 616 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {f} = \\operatorname {C o n c a t} \\left[ \\operatorname {W i n d o w} \\left(\\left\\{\\mathbf {Z} ^ {f} \\right\\} _ {m = 0} ^ {C - 1}\\right) \\right], f \\in \\{v, r \\}, \\tag {6}", + "image_path": "f3d3b126f63524cc23404a13aa7c1a1fd8345dc5fc50216247929e980dd0ce6a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 626, + 527, + 638 + ], + "type": "text", + "content": " are used to control the slice granularity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": "Further, the spatial information of " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^f" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^f" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": " is condensed by the average pooling operation, which produces " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{h}}^f\\in \\mathbb{R}^{C\\times \\frac{H}{\\sigma}}" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}^f\\in \\mathbb{R}^{C\\times \\frac{W}{\\mu}}" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": ". In the above manner, we refine the original spatial maps " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}^f\\}_{m = 0}^{C - 1}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": " into directional tensors " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{h}}^f" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}^f" + }, + { + "bbox": [ + 304, + 638, + 545, + 713 + ], + "type": "text", + "content": " to cut off the original self-modality dependencies in spatial dimensions. Furthermore," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3544" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "type": "text", + "content": "the cross-modality dependencies are established as follows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 95, + 287, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 95, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 91, + 95, + 287, + 119 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {h} ^ {f} = \\underset {\\left(\\frac {H}{\\sigma} \\times W\\right)} {\\operatorname {T r a n s c o n v}} (\\hat {\\mathbf {h}} ^ {f}), f \\in \\{v, r \\}, \\tag {7}", + "image_path": "33dd4b48236a4d194b5dbe7cae032d28908feb59b2e3aa176d61d02044a00e50.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 92, + 130, + 287, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 130, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 92, + 130, + 287, + 153 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {v} ^ {f} = \\underset {\\left(H \\times \\frac {W}{\\mu}\\right)} {\\operatorname {T r a n s C o n v}} (\\hat {\\mathbf {v}} ^ {f}), f \\in \\{v, r \\}. \\tag {8}", + "image_path": "43316ec22a51fe33f82ab28c86386202e75a2b06cde30c0a8ddd2ec60bac6d19.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": "In Eq. (7) and Eq. (8), we use two modality-shared transposed convolution layers to reconstruct the spatial maps of " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^f" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^f" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": ", respectively. This process aims to establish the cross-modality spatial dependencies. The outputs are " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_h^f \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_v^f \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "text", + "content": "However, there exist \"scheme stereotypes\" in " + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_h^f" + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_v^f" + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "text", + "content": ". The main reason is that these features are learned from two distinct directional schemes, which may lead to a skewed learning procedure. In contrast, we hope these two schemes can consistently provide diversified patterns that cover more valuable signals. Thus, the shuffle attention is designed to fuse " + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_h^f" + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_v^f" + }, + { + "bbox": [ + 47, + 221, + 287, + 343 + ], + "type": "text", + "content": " while decoupling them from two directions. Specifically, we concatenate them on the channel level and then shuffle the channel order to disarrange the spatial maps learned from two directions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 352, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 352, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 64, + 352, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {D} ^ {f} = \\operatorname {S h u f f l e} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {Z} _ {h} ^ {f}, \\mathbf {Z} _ {v} ^ {f}\\right)\\right), f \\in \\{v, r \\}, \\tag {9}", + "image_path": "1898586c0697763e84e6065de4cb261dca87a73020b45305d9950599013af7f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 376, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 414 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 376, + 287, + 414 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^f\\in \\mathbb{R}^{2C\\times H\\times W}" + }, + { + "bbox": [ + 47, + 376, + 287, + 414 + ], + "type": "text", + "content": " denotes the concatenated feature with disarranged channel order. The new channel dependencies are established by weighting all the spatial maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {D}} ^ {f} = \\mathbf {W} _ {2} R e L U \\left(\\mathbf {W} _ {1} G A P \\left(\\mathbf {D} ^ {f}\\right)\\right), f \\in \\{v, r \\}, \\tag {10}", + "image_path": "f839978afc2b22360a67afe74ef6408afc8f9a6398a5ba06f81dff9e79d082b8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1\\in \\mathbb{R}^{2C\\times \\frac{C}{4}}" + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2\\in \\mathbb{R}^{\\frac{C}{4}\\times C}" + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "text", + "content": " indicate projection matrices to compute the channel attention [16]. Finally, we introduce " + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}^f" + }, + { + "bbox": [ + 47, + 459, + 287, + 497 + ], + "type": "text", + "content": " to refine the original features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 506, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 506, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 108, + 506, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} ^ {f} = \\hat {\\mathbf {D}} ^ {f} \\oplus \\mathbf {Z} ^ {f}, f \\in \\{v, r \\}, \\tag {11}", + "image_path": "101c19f955bd736edde72cfaa3fef61d454495b2962aef6595433eb39dcaa2ed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^f" + }, + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "type": "text", + "content": " is used as the input of the next block; " + }, + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 47, + 530, + 287, + 602 + ], + "type": "text", + "content": " is the element-wise add operation. The proposed fine-grained dependency reconstruction module can effectively establish fine-grained and cross-modality dependencies in both visible and infrared features, which can help discover substantial small patterns that are vital in VI recognition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 612, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 135, + 624 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "type": "text", + "content": "We utilize the hard-mining triplet loss " + }, + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "type": "inline_equation", + "content": "(L_{tri})" + }, + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "type": "text", + "content": " [42] and cross-entropy loss " + }, + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "type": "inline_equation", + "content": "(L_{id})" + }, + { + "bbox": [ + 47, + 632, + 287, + 667 + ], + "type": "text", + "content": " for metric learning and identity learning, respectively. The overall loss can be defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 124, + 681, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 681, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 124, + 681, + 287, + 694 + ], + "type": "interline_equation", + "content": "L = L _ {i d} + \\gamma * L _ {t r i}, \\tag {12}", + "image_path": "361f0487618f0043cc41e90697998706aee842f86f2a553e468d441324824f02.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 701, + 253, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 253, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 253, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 253, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 701, + 253, + 713 + ], + "type": "text", + "content": " is the parameter to balance each loss term." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 72, + 432, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 432, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 432, + 84 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 91, + 488, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 488, + 103 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 488, + 103 + ], + "type": "text", + "content": "4.1. Datasets and Settings for VI-ReID" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 109, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 545, + 144 + ], + "type": "text", + "content": "We first evaluate the proposed method for the visible-infrared person re-identification (VI-ReID) task on two public benchmarks: SYSU-MM01 [35] and RegDB [25]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 145, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 228 + ], + "type": "text", + "content": "SYSU-MM01 [35] is the largest VI-ReID dataset, which contains 491 persons captured from six different cameras across scenes (indoor and outdoor). There are a total of 287628 visible images and 15792 infrared images. Following the evaluation protocol in [26,38], 395 persons are fixed for training and 96 persons for testing. The testing process includes all-search and indoor-search modes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 228, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 336 + ], + "type": "text", + "content": "RegDB [25] contains 412 persons, and each person has 10 visible and 10 infrared images. \"Visible to infrared\" and \"infrared to visible\" are two testing modes. The former uses visible as the query and infrared as the gallery. The latter does the opposite. Following [26, 30, 38], the final performance on the RegDB is made with the average results of 10 times the training and testing procedure. Each time, we randomly sample 206 identities for training, and the remaining 206 identities are used for testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 336, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 384 + ], + "type": "text", + "content": "Evaluation Metrics. We use the standard cumulative matching characteristics (CMC), mean average precision (mAP), and mean inverse negative penalty (mINP) [42] as the evaluation metrics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 384, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 456 + ], + "type": "text", + "content": "Implementation Details. For the ablation studies, we select the MobileNetV3-L [15] as the lightweight baseline to validate each proposed component. As shown in Table 1, we encapsulate all the layers of MobileNetV3-L as Block1-7 and then initialize the first four blocks twice to construct the dual-path network. During the TOP, detailed settings are" + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 308, + 484, + 545, + 600 + ], + "blocks": [ + { + "bbox": [ + 306, + 461, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 306, + 461, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 545, + 483 + ], + "type": "text", + "content": "Table 1. Detailed structures for each Block. We package the entire MobileNetV3-Large into Block1-7 without overlap." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 484, + 545, + 600 + ], + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 600 + ], + "type": "table", + "html": "
Block partitions on MobileNetV3-large
Layer NameStructuresOutput Size
Block1Conv(3×3, 2), Bneck(3×3, 1)16 × 1122
Block2Bneck(3×3, 2)24 × 562
Block3Bneck(3×3, 1)24 × 562
Block4Bneck(5×5, 2), Bneck(5×5, 1)×240 × 282
Block5Bneck(3×3, 2), Bneck(3×3, 1)×5112 × 142
Block6Bneck(5×5, 2), Bneck(5×5, 1)×2160 × 72
Block7Conv(1×1, 1)×31280 × 72
", + "image_path": "65afeb3e414c73fb163027f9b4b5d077984c3438313ecfa9912a6b53c4dafe3a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 653 + ], + "type": "text", + "content": "shown in Table 2(b). It is worth noticing that the \"Random pick\" means that for each image sample used in the Block1-1 path, we randomly pick one method from colour jitter, channel shuffle, or RGB shift to augment it." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "During the VI-ReID training, detailed settings are presented in Table 2(a). Briefly, we randomly sample 6 identities with 6 visible and 6 infrared images per identity, which makes a total of 72 images in each batch. All images are resized to " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "288 \\times 144" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Random erasing (RE) [49], flipping," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3545" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 82, + 287, + 209 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 264, + 81 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 264, + 81 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 264, + 81 + ], + "type": "text", + "content": "Table 2. Detailed training and augmentation settings." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 82, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 287, + 209 + ], + "type": "table", + "html": "
(a): Training settings
SettingsOn ImageNet-1KOn ImageNet-miniOn VI-ReID
Epochs1505070
Batch size51236×236×2
Image size224×224224×224288×144
Augs.Stanard [15]Task-orientedRE, CE, Flip
OptimizerSGDSGDAdamW [23]
Lr.0.20.10.001
Decaycosinecosineconstant
Warm upFirst 5 epochNoneFirst 5 epoch
Weight decay5E-45E-51E-2
PlatformPytorch, FP16PytorchPytorch
GPUTesla V100RTX 3060RTX 3060
Cost57 h4 h-
", + "image_path": "e114a81b415a73e67aeb2910604622dc5cd06c71d2a36c9c98db898b8a85ec2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 210, + 287, + 300 + ], + "blocks": [ + { + "bbox": [ + 50, + 210, + 287, + 300 + ], + "lines": [ + { + "bbox": [ + 50, + 210, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 50, + 210, + 287, + 300 + ], + "type": "table", + "html": "
(b): Task-oriented augmentation settings
AugmentationsProbabilityTarget
Generic (G)Crop & Filp0.5All path
Deformation0.2
Colour (C)Colour jitterRandom pickBlock1-1 path
Channel shuffle
RGB shift
Texture (T)Compression0.2Block1-2 path
Defocusing0.2
Sharpen gray1
", + "image_path": "a9c2d4aafb28c31ea5533c06c99a63a7cc83653b305e742a5f30ca8a72f85698.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": "and channel erasing (CE) [40] are utilized against overfitting. The AdamW [23] optimizer with a learning rate (Lr) of 0.001, and a weight decay of 0.01 is adopted to train our network for 70 epochs. We use the Lr warm-up in the first 5 epochs and constant decay at the " + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "inline_equation", + "content": "20^{th}" + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "inline_equation", + "content": "40^{th}" + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": " epochs with a decay factor of 0.1. The parameters " + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": " are set to 3 and 2 based on the fine-tuning results, respectively. The balance parameter " + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": " in Eq. (12) is empirically set to " + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 46, + 311, + 287, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 417, + 141, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 141, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 141, + 430 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "type": "text", + "content": "Effect of Task-oriented Pretraining. We first evaluate the effect of task-oriented pretraining on VI-ReID, as shown in Table 3. No.1 indicates the baseline performance obtained via direct fine-tuning of the ImageNet-1k pretrained model in VI-ReID datasets. When only using generic augmentations and identity loss to perform TOP on ImageNetmini, the \"catastrophic forgetting\" makes No.2 perform worse than the baseline. However, when applying the proposed task-oriented augmentations, the network learns to extract stable features from visually discrepant images in advance, thus improving the downstream VI-ReID performance (No.3). Coupled with " + }, + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{idc}" + }, + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{fd}" + }, + { + "bbox": [ + 46, + 437, + 287, + 652 + ], + "type": "text", + "content": ", we create a conflicting learning process that enforces the network to learn how to represent and embed heterogeneous features. No.5 shows significant improvements in both the SYSU-MM01 and RegDB datasets. By only using the MobileNetV3-large backbone, the performance reported in No.5 has already exceeded many ResNet-based works [19, 26, 38]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Effect of Fine-grained Dependency Reconstruction Module. In Table 3, No.6-No.9 show the ablation experiments about the fine-grained dependency reconstruction module. We intend to construct the cross-modality feature dependencies, and the first issue is how to model the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "original feature relations. No.6 indicates directly applying the coarse-grained features to the shuffle attention module. The performance is slightly improved. When using the proposed directional scheme " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "(H_{s},V_{s})" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " to extract fine-grained patterns and reconstruct their spatial relation using transposed convolution layers, No.7 and No.8 show meaningful improvements. The full power of the FDR module is shown in No.9, which adopts both fine-grained horizontal and vertical patterns to discover cross-modality feature similarities, achieving " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "66.14\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "63.80\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " in terms of Rank-1 and mAP performance on SYSU-MM01 under the all-search mode." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "type": "text", + "content": "Pointing at the difficulties in VI recognition tasks, we design the TOP strategy and FDR module to remedy the drawbacks of the lightweight network. Compared with the baseline (No.1), our method (No.9) considerably improves all metrics on two benchmarks. Meanwhile, the ablations reported in Table 3 demonstrate that all the proposed components bring advantages consistently to the accuracy gain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": "Applicability with Lightweight Networks. We evaluate our methods on three mainstream lightweight networks (ShuffleNetV2 [24], GhostNet [11] and MobileNetV3 [15]). As shown in Table 4, our method can remarkably improve VI Re-ID performance with ignorable complexity growth. Compared with conventional deep networks, we surpass them significantly with minimal complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 468 + ], + "type": "text", + "content": "Detailed Ablations on the FDR Module. We validate the rationality of each component in the FDR module. Table 5(a) shows the impacts of different spatial modelling methods. Except for ours, the HAP [38] achieves the best performance. However, we extract patterns from both the horizontal and vertical regions and use up-sampling (u.) to discover modality-shared cues from re-enlarged spatial areas, which can be regarded as a more effective solution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": "Table 5(b) shows the impacts of different channel relation reasoning methods. Our shuffle attention (SA) wins both SE [16] and CBAM [33]. The recipe is the channel shuffle (cs.) operation that builds scheme-decoupled dependencies. This crucial step can evenly correlate visible and infrared features from each fine-grained pattern." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 550, + 429, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 429, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 429, + 562 + ], + "type": "text", + "content": "4.3. Feature Visualization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "type": "text", + "content": "What Does TOP Do? We randomly track 10 identities during the TOP stage and visualize the feature distribution via T-SNE [28], as illustrated in Fig. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We first use task-oriented augmentation to create two types of visual discrepant samples from each image, as shown in Fig. 4(a) and Fig. 4(b). Then, the proposed fake domain loss forces the features after Block4 to be domain-specific, as shown in Fig. 4(c). Concurrently, it forces the network to learn to embed those domain-specific features and provide domain-shared predictions. In Fig. 4(d), after the training, we can observe that the network knows how to extract the domain-shared representations and successfully" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3546" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 104, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 3. Evaluation of each proposed component on two VI-ReID datasets. \"Augs.\" indicates the augmentations. G, C and T denote the generic, colour, and texture augmentations, respectively. In the FDR module, " + }, + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 47, + 70, + 547, + 103 + ], + "type": "text", + "content": " denote the horizontal and vertical slices with up-sampling. SA is the shuffle attention module. Rank (r) (\\%), mAP (\\%) and mINP (\\%) are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 104, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 49, + 104, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 104, + 547, + 240 + ], + "type": "table", + "html": "
No.Task-oriented pretraining stageVI training stageSYSU-MM01 (all-search)RegDB (visible-to-infrared)
Augs.Loss functionsFDR module
GC+TLidLidcLfdHsVsSAr=1r=10mAPmINPr=1r=10mAPmINP
147.8189.7147.0633.4871.2689.9465.6648.50
2✓✓43.2885.9645.5631.1070.7388.5265.6548.41
3✓✓✓49.8589.7447.5635.5271.3289.9165.6748.49
4✓✓✓54.2892.1152.9441.2975.3192.6468.7852.16
5✓✓✓✓62.4194.1259.0645.1382.7594.1376.2161.84
6✓✓✓✓✓62.8994.2659.7945.8482.8894.1976.2461.92
7✓✓✓✓✓✓63.9595.2860.0946.8083.0794.4876.5562.20
8✓✓✓✓✓✓64.0495.4161.1246.9283.2294.6977.0163.16
9✓✓✓✓✓✓✓66.1496.0363.8049.7684.1594.9879.2663.86
", + "image_path": "8b0b05fc4857bf59dd2d0bcfebf3b496fe659b81c383baff117da99d9089aa26.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 270, + 287, + 453 + ], + "blocks": [ + { + "bbox": [ + 47, + 248, + 287, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 269 + ], + "type": "text", + "content": "Table 4. Experimental results on different lightweight networks and conventional deep networks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 270, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 50, + 270, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 50, + 270, + 287, + 453 + ], + "type": "table", + "html": "
MethodsFLOPs (M)SYSU-MM01RegDB
r=1mAPr=1mAP
ConventionResNet-50356256.9854.7276.8671.30
ConvNeXt-Tiny362058.7255.3178.2572.64
Vit-B568952.1751.8175.3170.37
Swin-Tiny328758.2455.1678.3972.68
LightweightShuffleNetV2-1.0×13941.8841.9467.8364.85
+TOP & FDR17755.7152.6379.8266.36
ShuffleNetV2-1.5×26547.3947.8170.1565.28
+TOP & FDR37163.3560.8184.1376.98
GhostNet-1.0×15042.5342.9471.2864.40
+TOP & FDR18958.5455.1983.2677.16
GhostNet-1.3×28150.8947.9272.5165.98
+TOP & FDR39566.7664.0185.5179.95
MobileNetV3-S10440.9242.5162.7758.31
+TOP & FDR13054.7550.2675.5370.17
MobileNetV3-L25047.8147.0671.2665.66
+TOP & FDR36266.1463.8084.1579.26
", + "image_path": "d810e5f5ec8c44d9c3fc4d47057ef7a9b4c417c7cf9de012c68f1ba5a9477700.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 491, + 287, + 641 + ], + "blocks": [ + { + "bbox": [ + 47, + 456, + 287, + 490 + ], + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 490 + ], + "type": "text", + "content": "Table 5. Evaluation of spatial modelling methods and channel relation reasoning methods in the FDR module. \"u.\" and \"cs.\" respectively denote the up-sampling and channel shuffle operations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 491, + 287, + 641 + ], + "lines": [ + { + "bbox": [ + 50, + 491, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 50, + 491, + 287, + 641 + ], + "type": "table", + "html": "
(a): Impact of different spatial modelling methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
GAP62.8959.7945.8482.8876.2461.92
Context [2]61.3757.5246.0881.5674.4962.11
HAP [38]62.9258.1347.8482.9876.2062.77
HS+VS(w/o u.)63.4559.7549.6183.9478.8263.21
HS+VS66.1463.8049.7684.1579.2663.86
(b): Impact of different channel relation reasoning methods
MethodsSYSU-MM01RegDB
r=1mAPmINPr=1mAPmINP
SE [16]62.9159.7045.7882.7977.4562.34
CBAM [33]61.7956.8845.2583.4177.2862.96
SA (w/o cs.)62.8158.1045.7682.7576.1262.56
SA66.1463.8049.7684.1579.2663.86
", + "image_path": "8f6727feac59d6080d3cb37f2a68e155bb579df088b9c8a85f5bf87d670931ce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 651, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 687 + ], + "type": "text", + "content": "embed all the features into their respective identity groups. Thus, the proposed task-oriented pretraining strategy makes lightweight networks fit cross-modality tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Pattern Visualization. As demonstrated in Fig. 5, we visualize the learned patterns using Grad-CAM [27] to ex" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 248, + 531, + 375 + ], + "blocks": [ + { + "bbox": [ + 321, + 248, + 531, + 375 + ], + "lines": [ + { + "bbox": [ + 321, + 248, + 531, + 375 + ], + "spans": [ + { + "bbox": [ + 321, + 248, + 531, + 375 + ], + "type": "image", + "image_path": "2baa24b60df4fee86f2fd1cb76f7bd15817972db5f26cfa5f8570c8b13645c16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 380, + 545, + 424 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 424 + ], + "type": "text", + "content": "Figure 4. Feature distributions during task-oriented pretraining. Each colour denotes an identity. In (a), the star markers represent raw samples. In (b), (c) and (d), the circle and triangle markers represent colour and texture-augmented samples, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 426, + 529, + 557 + ], + "blocks": [ + { + "bbox": [ + 326, + 426, + 529, + 557 + ], + "lines": [ + { + "bbox": [ + 326, + 426, + 529, + 557 + ], + "spans": [ + { + "bbox": [ + 326, + 426, + 529, + 557 + ], + "type": "image", + "image_path": "ee5eac419acd4b332838bda4aef0870d29194e08c82b3d3e8ee87875cce38cbc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "Figure 5. Visualization of learned patterns via Grad-CAM [27]. For a clear presentation, we average all the spatial maps into four. plain the effect of the FDR module. Clearly, the original patterns extracted by MobileNetV3-L are concentrated and overfit on the modality-specific parts. To remedy this issue, we use the FDR module to discover modality-shared patterns from fine-grained regions, which separates learned patterns to focus on considerable modality-shared cues. A good case is shown in Fig. 5. The visible patterns initially learned are concentrated in the high chest area, while the infrared patterns are all in the right arm. After re-weighting by the FDR module, modality-shared small regions like the face, arms, neck, and shoes are all taken into account, thus" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3547" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 82, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 538, + 81 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 538, + 81 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 538, + 81 + ], + "type": "text", + "content": "Table 6. Comparison with the state-of-the-arts on SYSU-MM01 [35]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 82, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 545, + 218 + ], + "type": "table", + "html": "
DetailsAll-searchIndoor-search
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356214.8054.1271.3315.95-20.5868.3885.7926.92-
JSIA [30]ResNet50+GAN>413338.1080.7089.9036.90-43.8086.2094.2052.90-
AGW [42]ResNet50>356247.5084.3992.1447.6535.3054.1791.1495.9862.9759.23
X-Modal [19]ResNet50>356249.9089.8096.0050.70------
DMiR [38]ResNet50>356250.5488.1294.8649.29-53.9292.5097.0962.49-
FBP-AL [32]ResNet50>356254.1486.0493.0350.20------
DDAG [41]ResNet50>356254.7590.3995.8155.0239.6261.0294.0698.4167.9862.61
HAT [43]ResNet50>356255.2992.1497.3653.89-62.1095.7599.2070.84-
LBA [26]ResNet50>356255.41--54.14-58.46--66.33-
TSME [21]ResNet50>356264.2395.1998.7361.21-64.8096.9299.3171.53-
SPOT [4]ResNet50+ViT>481065.3492.7397.0462.2548.8669.4296.2299.1274.6370.48
TOLight (Ours)MobileNetV3-L= 36266.1496.0397.6863.8049.7672.4197.5499.2376.1171.43
TOLight (Ours)GhostNet-1.3×= 39566.7696.2398.7064.0150.1872.8997.9399.2876.7071.95
", + "image_path": "df9d1fd59370fa41a70680fc7c1e467aa79b8ebcb942e71903a6ace465ff34c3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 233, + 545, + 367 + ], + "blocks": [ + { + "bbox": [ + 67, + 222, + 525, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 222, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 222, + 525, + 232 + ], + "type": "text", + "content": "Table 7. Comparison with the state-of-the-arts on RegDB [25]. Metrics of Rank at r (%), mAP (%), and mINP (% are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 233, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 50, + 233, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 50, + 233, + 545, + 367 + ], + "type": "table", + "html": "
DetailsVisible-to-InfraredInfrared-to-Visible
MethodsBackboneFLOPs (M)r=1r=10r=20mAPmINPr=1r=10r=20mAPmINP
Zero-pad [35]ResNet50>356217.7534.2144.3518.90-16.6334.6844.2517.82-
JSIA [30]ResNet50+GAN>413348.50--49.30-48.10--48.90-
AGW [42]ResNet50>356270.0586.2191.5566.3750.1970.4987.2191.8465.9051.24
X-Modal [19]ResNet50>356262.2183.1391.7260.18------
DMiR [38]ResNet50>356275.7989.8694.1869.97-73.9389.8793.9868.22-
FBP-AL [32]ResNet50>356273.9889.7193.6968.24-70.0589.2293.8866.61-
DDAG [41]ResNet50>356269.3486.1991.4963.4649.2468.0685.1590.3161.8048.62
HAT [43]ResNet50>356271.8387.1692.1667.56-70.0268.4591.6166.30-
LBA [26]ResNet50>356274.17--67.64-72.43--65.46-
SPOT [4]ResNet50+ViT>481080.3593.4896.4472.4656.1979.3792.7996.0172.2656.06
GECNet [48]ResNet50+GAN>435082.3392.7295.4978.45-78.9391.9995.4475.58-
TOLight (Ours)MobileNetV3-L=36284.1594.9896.5879.2663.8680.9492.8596.3776.1059.33
TOLight (Ours)GhostNet-1.3×=39585.5194.9996.7079.9563.8580.6592.8196.3275.9159.26
", + "image_path": "e8e66a564345f2bfc7608e999b789398a2eee7f403c60908ead0ee329c57bfc8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 377, + 272, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 272, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 272, + 389 + ], + "type": "text", + "content": "improving the accuracy of the cross-modality matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 234, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 234, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 234, + 407 + ], + "type": "text", + "content": "4.4. Comparison with State of The Arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": "We compare our method with recently proposed state-of-the-art VI-ReID methods, as shown in Table 6 and Table 7. They illustrate that we remarkably and efficiently exceed all the compared SOTAs under diverse evaluation settings. Specifically, based on the tiny GhostNet- " + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "inline_equation", + "content": "1.3\\times" + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": " backbone, we attain the rank-1 of " + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "inline_equation", + "content": "66.76\\%" + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": " and mAP of " + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "inline_equation", + "content": "64.01\\%" + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": " on the all-search mode of the large-scale SYSU-MM01 dataset with only 395 FLOPs. The proposed method also achieves excellent results on the RegDB dataset, surpassing the previous SOTAs by a significant margin with lower FLOPs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "type": "text", + "content": "Table 8. Evaluation on two VI-FR datasets. CA is channel augmentation [40]. B is the LightCNN-29 baseline. Rank at 1 accuracy " + }, + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "type": "text", + "content": " and false acceptance rate (F: " + }, + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 47, + 537, + 286, + 570 + ], + "type": "text", + "content": " ) are reported." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 571, + 286, + 656 + ], + "blocks": [ + { + "bbox": [ + 50, + 571, + 286, + 656 + ], + "lines": [ + { + "bbox": [ + 50, + 571, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 50, + 571, + 286, + 656 + ], + "type": "table", + "html": "
MethodsOulu [5]BUAA [17]
r=1F:1%F:0.1%r=1F:1%F:0.1%
IDR [13]94.373.446.294.393.484.7
VSA [44]99.996.882.398.098.292.5
PACH [8]10097.988.298.698.093.5
B [36]10097.987.098.097.793.7
B+CA [40]10098.991.798.398.294.5
B+TOP10098.891.598.398.194.5
B+TOP+FDR (Ours)10098.991.798.398.294.6
", + "image_path": "46765ffac46e1a9082cb2252ca95b380e5aeef0782aaaa551e07c1795d729874.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 228, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 228, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 228, + 684 + ], + "type": "text", + "content": "4.5. Visible-Infrared Face Recognition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "We follow [40] to examine our method on Oulu-CASIA NIR-VIS [5] and BUAA-VisNir face databases [17]. Fol" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 377, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 437 + ], + "type": "text", + "content": "lowing previous studies, we also adopt the LightCNN-29 [36] as our baseline network (B). Except for IDR [13], VSA [44], and PACH [8], we also compare recently proposed channel augmentation (CA) [40] with our method on the same baseline. Results are displayed in Table 8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 486 + ], + "type": "text", + "content": "Evidently, the proposed TOP and FDR also enhance the performance of VI face recognition. These results further validate the effectiveness and generalization of our methods, which can solve VI recognition tasks effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 499, + 378, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 499, + 378, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 499, + 378, + 512 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 676 + ], + "type": "text", + "content": "This work presents an effective task-oriented pretrained lightweight neural network (TOPLight) to solve visible-infrared recognition problems. First, the task-oriented pretraining strategy significantly improves the ability of lightweight networks to understand heterogeneous images by introducing domain conflict and sample variation during the pretraining phase. Second, the fine-grained dependency reconstruction module is utilized to destroy the modality-specific pattern dependencies and construct the fine-grained and modality-shared pattern dependencies between visible and infrared features. Extensive experiments demonstrate the superiority and effectiveness of the proposed method as well as the efficacy of each component of our framework." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is funded by the National Natural Science Foundation of China (Grant No. 61802058, 61911530397)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3548" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 113 + ], + "type": "text", + "content": "[1] Sun, et.al. Learning part-based convolutional features for person re-identification. TPAMI, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 288, + 169 + ], + "type": "text", + "content": "[2] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 225 + ], + "type": "text", + "content": "[3] Srinivasa Rao Chalamala, Krishna Rao Kakkirala, and Jami Santosh Kumar. Face recognition using spatial pyramid matching and lrbp. In 2014 IEEE 10th International Colloquium on Signal Processing and its Applications, pages 67-70. IEEE, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 281 + ], + "type": "text", + "content": "[4] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Jianguo Jiang, and Chia-Wen Lin. Structure-aware positional transformer for visible-infrared person re-identification. IEEE Transactions on Image Processing, 31:2352-2364, 2022. 1, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "type": "text", + "content": "[5] Jie Chen, Dong Yi, Jamei Yang, Guoying Zhao, Stan Z Li, and Matti Pietikainen. Learning mappings for face synthesis from near infrared to visual light images. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 156-163. IEEE, 2009. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 384 + ], + "type": "text", + "content": "[6] Changxing Ding and Dacheng Tao. A comprehensive survey on pose-invariant face recognition. ACM Transactions on intelligent systems and technology (TIST), 7(3):1-42, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 430 + ], + "type": "text", + "content": "[7] Hang Du, Hailin Shi, Dan Zeng, Xiao-Ping Zhang, and Tao Mei. The elements of end-to-end deep face recognition: A survey of recent advances. ACM Computing Surveys (CSUR), 54(10s):1-42, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 485 + ], + "type": "text", + "content": "[8] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7930-7938, 2020. 1, 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 488, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 532 + ], + "type": "text", + "content": "[9] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In 2012 19th IEEE International Conference on Image Processing, pages 1837-1840. IEEE, 2012. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "text", + "content": "[10] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "text", + "content": "[11] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[13] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Learning invariant deep representation for nir-vis face recognition. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 1, 2, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Ran He, Xiang Wu, Zhenan Sun, and Tieniu Tan. Wasserstein cnn: Learning invariant features for nir-vis face recognition. IEEE transactions on pattern analysis and machine intelligence, 41(7):1761-1773, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "text", + "content": "[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 1, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "text", + "content": "[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 263 + ], + "type": "text", + "content": "[17] D Huang, J Sun, and Y Wang. The buaa-visnir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3:3, 2012. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "type": "text", + "content": "[18] Kwonjoon Lee, Subhransu Maji, Avinash Ravichandran, and Stefano Soatto. Meta-learning with differentiable convex optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10657-10665, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "text", + "content": "[19] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an x modality. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4610–4617, 2020. 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 421 + ], + "type": "text", + "content": "[20] Shengcai Liao, Dong Yi, Zhen Lei, Rui Qin, and Stan Z Li. Heterogeneous face recognition from local structures of normalized appearance. In International Conference on Biometrics, pages 209-218. Springer, 2009. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 474 + ], + "type": "text", + "content": "[21] Jianan Liu, Jialiang Wang, Nianchang Huang, Qiang Zhang, and Jungong Han. Revisiting modality-specific feature compensation for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "type": "text", + "content": "[22] Xiaoxiang Liu, Lingxiao Song, Xiang Wu, and Tieniu Tan. Transferring deep representation for nir-vis heterogeneous face recognition. In 2016 International Conference on Biometrics (ICB), pages 1-8. IEEE, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 522, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 543 + ], + "type": "text", + "content": "[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[24] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "text", + "content": "[25] Dat Tien Nguyen, Hyung Gil Hong, Ki Wan Kim, and Kang Ryoung Park. Person recognition system based on a combination of body images from visible light and thermal cameras. Sensors, 17(3):605, 2017. 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[26] Hyunjong Park, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12046–12055, 2021. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "text", + "content": "[27] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3549" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 150 + ], + "type": "text", + "content": "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[29] Guan'an Wang, Tianzhu Zhang, Jian Cheng, Si Liu, Yang Yang, and Zengguang Hou. Rgb-infrared cross-modality person re-identification via joint pixel and feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3623-3632, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "type": "text", + "content": "[30] Guan-An Wang, Tianzhu Zhang, Yang Yang, Jian Cheng, Jianlong Chang, Xu Liang, and Zeng-Guang Hou. Cross-modality paired-images generation for rgb-infrared person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12144–12151, 2020. 1, 2, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 331 + ], + "type": "text", + "content": "[31] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 618–626, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "text", + "content": "[32] Ziyu Wei, Xi Yang, Nannan Wang, and Xinbo Gao. Flexible body partition-based adversarial learning for visible infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 421 + ], + "type": "text", + "content": "[33] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 499 + ], + "type": "text", + "content": "[34] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[35] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In Proceedings of the IEEE international conference on computer vision, pages 5380-5389, 2017. 1, 2, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 600 + ], + "type": "text", + "content": "[36] Xiang Wu, Ran He, Zhenan Sun, and Tieniu Tan. A light cnn for deep face representation with noisy labels. IEEE Transactions on Information Forensics and Security, 13(11):2884-2896, 2018. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 656 + ], + "type": "text", + "content": "[37] Xiang Wu, Huaibo Huang, Vishal M Patel, Ran He, and Zhenan Sun. Disentangled variational representation for heterogeneous face recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9005–9012, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[38] Mang Ye, Cuiqun Chen, Jianbing Shen, and Ling Shao. Dynamic tri-level relation mining with attentive graph for visible infrared re-identification. IEEE Transactions on Information Forensics and Security, 17:386–398, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 675 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Mang Ye, Xiangyuan Lan, Zheng Wang, and Pong C Yuen. Bi-directional center-constrained top-ranking for visible thermal person re-identification. IEEE Transactions on Information Forensics and Security, 15:407–419, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[40] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13567-13576, 2021. 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 227 + ], + "type": "text", + "content": "[41] Mang Ye, Jianbing Shen, David J Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. In European Conference on Computer Vision, pages 229–247. Springer, 2020. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 229, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 283 + ], + "type": "text", + "content": "[42] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 328 + ], + "type": "text", + "content": "[43] Mang Ye, Jianbing Shen, and Ling Shao. Visible-infrared person re-identification via homogeneous augmented trimodal learning. IEEE Transactions on Information Forensics and Security, 16:728-739, 2020. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "text", + "content": "[44] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-posed high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 441 + ], + "type": "text", + "content": "[45] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7349–7358, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 485 + ], + "type": "text", + "content": "[46] Teng Zhang, Arnold Wiliem, Siqi Yang, and Brian Lovell. Tv-gan: Generative adversarial network based thermal to visible face recognition. In 2018 international conference on biometrics (ICB), pages 174-181. IEEE, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 529 + ], + "type": "text", + "content": "[47] Huantao Zheng, Xian Zhong, Wenxin Huang, Kui Jiang, Wenxuan Liu, and Zheng Wang. Visible-infrared person re-identification: A comprehensive survey and a new setting. *Electronics*, 11(3):454, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 532, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 585 + ], + "type": "text", + "content": "[48] Xian Zhong, Tianyou Lu, Wenxin Huang, Mang Ye, Xuemei Jia, and Chia-Wen Lin. Grayscale enhancement colorization network for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1418–1430, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 587, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 631 + ], + "type": "text", + "content": "[49] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 13001-13008, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 632, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 545, + 675 + ], + "type": "text", + "content": "[50] Jun-Yong Zhu, Wei-Shi Zheng, Jian-Huang Lai, and Stan Z Li. Matching nir face to vis face using transduction. IEEE Transactions on Information Forensics and Security, 9(3):501-514, 2014. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3550" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_content_list.json b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9857909e5321ee6220cbf7b6539164d4e8dbde4d --- /dev/null +++ b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_content_list.json @@ -0,0 +1,1480 @@ +[ + { + "type": "text", + "text": "TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments", + "text_level": 1, + "bbox": [ + 258, + 130, + 712, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Sun $^{1*}$ Qian Bao $^{2\\dagger}$ Wu Liu $^{2\\dagger}$ Tao Mei $^{4}$ Michael J. Black $^{3}$", + "bbox": [ + 218, + 186, + 743, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Harbin Institute of Technology $^{2}$ Explore Academy of JD.com", + "bbox": [ + 227, + 205, + 741, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Max Planck Institute for Intelligent Systems HiDream.ai Inc.", + "bbox": [ + 227, + 224, + 741, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yusunhit@gmail.com, baoqian@jd.com, liuwul@jd.com", + "bbox": [ + 266, + 244, + 702, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tmei@hidream.ai, black@tuebingen.mpg.de", + "bbox": [ + 307, + 263, + 655, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5971fff5bed9d6d158c0495cedd0412794397823f881d9964ea67bfed7c7296.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 287, + 135, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic Monocular Camera", + "text_level": 1, + "bbox": [ + 138, + 296, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c61c8e122624cf0fc96aff32935004c837ea9533f8e03179629d700820206742.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 310, + 277, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf1fcb8fea368193fde60da9100333312b09de451a144b2314415a59fc11bada.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 309, + 419, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68ded7e47aa57dc68b9545efe0ce2abc58b36cdcd89b9f8b492ac993b61a9a8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 309, + 560, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2f7565c67777ab738977981f4af699e6e56a4165a08a5cc280c9a875bfca665.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 309, + 702, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9da159a468c0f4c2a27c14ebccb8b692b1b3cf9c4967a0af906e9fe40fa891b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 309, + 874, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ef673ece98846b17409ba65c428ceecd65d2001e3bb8e744dd8bf3417f38a34.jpg", + "image_caption": [ + "(a) Videos Captured by Dynamic Monocular Camera and Human 3D Recovery in Each Frame", + "(b) The 3D humans recovered by BEV [45] in camera coordinates", + "Figure 1. 5D temporal reconstruction of multiple 3D people in global coordinates with a dynamic camera. We introduce TRACE, a monocular one-stage method with a holistic 5D representation that enables the network to explicitly reason about human motion across time. Unlike previous methods [44,45] (b) that regress all 3D people from each frame in camera coordinates, TRACE (c) tracks the subjects (a) presented in the first frame through time and recovers their global trajectories in global coordinates in one shot." + ], + "image_footnote": [], + "bbox": [ + 91, + 445, + 482, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd6f2064535a36ec77850e054c10794e7d189078265c27a80c31296b43e33d24.jpg", + "image_caption": [ + "(c) The 5D humans recovered by TRACE in global coordinates" + ], + "image_footnote": [], + "bbox": [ + 486, + 445, + 879, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 689, + 313, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel ar", + "bbox": [ + 75, + 722, + 470, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chitectural components. Most importantly, it uses two new \"maps\" to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code1 and dataset2 are released for research purposes.", + "bbox": [ + 496, + 691, + 893, + 843 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Yu Sun was an intern at JD.com.", + "bbox": [ + 94, + 875, + 413, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://www.yusun.work/TRACE/TRACE.html", + "bbox": [ + 517, + 875, + 831, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2https://github.com/Arthur151/DynaCam", + "bbox": [ + 519, + 887, + 810, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8856", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 90, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The estimation of 3D human pose and shape (HPS) has many applications and there has been significant recent progress [6, 7, 21-24, 37, 38, 46, 49, 56, 57, 60]. Most methods, however, reason only about a single frame at a time and estimate humans in camera coordinates. Moreover, such methods do not track people and are unable to recover their global trajectories. The problem is even harder in typical hand-held videos, which are filmed with a dynamic, moving, camera. For many applications of HPS, single-frame estimates in camera coordinates are not sufficient. To capture human movement and then transfer it to a new 3D scene, we must have the movement in a coherent global coordinate system. This is a requirement for computer graphics, sports, video games, and extended reality (XR).", + "bbox": [ + 80, + 114, + 467, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is that most methods estimate humans in 3D, whereas the true problem is 5D. That is, a method needs to reason about 3D space, time, and subject identity. With a 5D representation, the problem becomes tractable, enabling a holistic solution that can exploit the full video to infer multiple people in a coherent global coordinate frame. As illustrated in Fig. 1, we develop a unified method to jointly regress the 3D pose, shape, identity, and global trajectory of the subjects in global coordinates from monocular videos captured by dynamic cameras (DC-videos).", + "bbox": [ + 80, + 327, + 467, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we deal with two main challenges. First, DC-videos contain both human motion and camera motion and these must be disentangled to recover the human trajectory in global coordinates. One idea would be to recover the camera motion relative to the rigid scene using structure-from-motion (SfM) methods (e.g. [31]). In scenes containing many people and human motion, however, such methods can be unreliable. An alternative approach is taken by GLAMR [58], which infers global human trajectories from local 3D human poses, without taking into account the full scene. By ignoring evidence from the full image, GLAMR fails to capture the correct global motion in common scenarios, such as biking, skating, boating, running on a treadmill, etc. Moreover, GLAMR is a multi-stage method, with each stage dependent on accurate estimates from the preceding one. Such approaches are more brittle than our holistic, end-to-end, method.", + "bbox": [ + 80, + 478, + 467, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The other challenge, as shown in the upper right corner of Fig. 1, is that severe occlusions are common in videos with multiple people. Currently, the most popular tracking strategy is to infer the association between 2D detections using a temporal prior (e.g. Kalman filter) [63]. However, in DC-videos, human motions are often irregular and can easily violate hand-crafted priors. PHALP [40] is one of the few methods to address this for 3D HPS. It uses a classical, multi-stage, detection-and-tracking formulation with heuristic temporal priors. It does not holistically reason about the sequence and is not trained end-to-end.", + "bbox": [ + 80, + 734, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we reason about people using a 5D representation and capture information from the full image and the motion of the scene. This holistic reasoning enables the reliable recovery of global human trajectories and subject tracking using a single-shot method. This is more reliable than multi-stage methods because the network can exploit more information to solve the task and is trained end-to-end. No hand-crafted priors are needed and the network is able to share information among modules.", + "bbox": [ + 503, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we develop TRACE, a unified one-stage method for Temporal Regression of Avatars with dynamic Cameras in 3D Environments. The architecture is inspired by BEV [45], which directly estimates multiple people in depth from a single image using multiple 2D maps. BEV uses a 2D map representing an imaginary, \"top down\", view of the scene. This is combined with an image-centric 2D map to reason about people in 3D. Our key insight is that the idea of maps can be extended to represent how people move in 3D. With this idea, TRACE introduces three new modules to holistically model 5D human states, performing multi-person temporal association, and inferring human trajectories in global coordinates; see Fig. 2.", + "bbox": [ + 503, + 229, + 890, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, to construct a holistic 5D representation of the video, we extract temporal image features by fusing single-frame feature maps from the image backbone with a temporal feature propagation module. We also compute the optical flow between adjacent frames with a motion backbone. The optical flow provides short-term motion features that carry information about the motion of the scene and the people. Second, to explicitly track human motions, we introduce a novel 3D motion offset map to establish the association of the same person across adjacent frames. This map contains a 3D offset vector at each position, which represents the difference between the 3D positions of the same subject from the previous frame to the current frame in camera coordinates. We also introduce a memory unit to keep track of subjects under long-term occlusion. Note that the 3D trajectories are built in camera space, and TRACE uses a novel world motion map that transfers the trajectories to global coordinates. At each position, this map contains a 6D vector to represent the difference between the 3D positions of the corresponding subject from the previous frame to the current frame and its 3D orientation in world coordinates. Taken together, this novel network architecture goes beyond prior work by taking information from the full video frames to address detection, pose estimation, tracking, and occlusion in a holistic network that is trained end-to-end.", + "bbox": [ + 503, + 429, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enable training and evaluation of global human trajectory estimation from in-the-wild DC-videos, we build a new dataset, DynaCam. Since collecting global human trajectories and camera poses with in-the-wild DC-videos is difficult, we simulate a moving camera using publicly available in-the-wild panoramic videos and regular videos cap", + "bbox": [ + 503, + 810, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8857", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tured by static cameras. In this way, we create more than 500 in-the-wild DC-videos with precise camera pose annotations. Then we generate pseudo-ground-truth 3D human annotations via fitting [20] SMPL [32] to detected 2D pose sequences [17,54,63]. With 2D/3D human pose and camera pose annotations, we can obtain the global human trajectories using the PnP algorithm [16]. This dataset is sufficient to train TRACE to deal with dynamic cameras.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate TRACE on two multi-person in-the-wild benchmarks (3DPW [50] and MuPoTS-3D [35]) and our DynaCam dataset. On 3DPW, TRACE achieves the state-of-the-art (SOTA) PA-MPJPE of $37.8\\mathrm{mm}$ , less the current best (42.7 [26]). On MuPoTS-3D, TRACE outperforms previous 3D-representation-based methods [39, 40] and tracking-by-detection methods [63] on tracking people under long-term occlusion. On DynaCam, TRACE outperforms GLAMR [58] in estimating the 3D human trajectory in global coordinates from DC-videos.", + "bbox": [ + 75, + 212, + 470, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our main contributions are: (1) We introduce a 5D representation and use it to learn holistic temporal cues related to both 3D human motions and the scene. (2) We introduce two novel motion offset representations to explicitly model temporal multi-subject association and global human trajectories from temporal clues in an end-to-end manner. (3) We estimate long-term 3D human motions over time in global coordinates, achieving SOTA results. (4) We collect the DynaCam dataset of DC-videos with pseudo ground truth, which facilitates the training and evaluation of global human trajectory estimation. The code and dataset are publicly available for research purposes.", + "bbox": [ + 75, + 363, + 470, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 558, + 218, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular 3D mesh regression with full images. Most existing methods [21-24, 26, 29, 37, 38, 46, 52, 53, 59, 60] take a multi-stage detection-based pipeline to estimate 3D HPS from cropped image patches, which exclude important cues, such as camera information and human-scene relationships. A few recent multi-stage [28] and one-stage [44, 45] methods have made steps towards using the full-image information. For instance, CIUFF [28] estimates 3D HPS by taking into account the bounding box locations, giving the method camera information and improving accuracy. To directly estimates multiple people at once from the full image, ROMP [44] introduces a 2D Center heatmap and a Mesh parameter map to represent 2D human locations and 3D human body meshes, respectively. BEV [45] goes beyond ROMP by introducing an imaginary bird's-eye-view map, which is combined with the front-view maps to construct a 3D view in camera coordinates. However, they only model 3D HPS in camera coordinates from a single image. Using \"maps\" like ROMP and BEV, TRACE also looks at the full image. We go further, however, by introducing novel maps that model human motions across a video se", + "bbox": [ + 75, + 583, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quence in global coordinates.", + "bbox": [ + 498, + 90, + 694, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tracking datasets. While there are many tracking datasets [8, 10, 11, 14, 18, 61] with 2D annotations, only a few [13, 15] capture the 3D trajectory of pedestrians. In both cases, the scene and human activities are limited. To address this, we use 3DPW [50] and MuPoTS-3D [35] for tracking evaluation. 3DPW is the most relevant dataset for our task and it provides a real-world test case. 3DPW contains videos that are captured by a moving camera that follows subjects to record their activities in many daily scenes. MuPoTS-3D contains rich multi-person interaction scenes with long-term occlusions for tracking evaluation.", + "bbox": [ + 496, + 114, + 890, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tracking 3D people through occlusions. Most existing methods [4, 42, 51, 63] perform tracking using 2D image cues. The classic tracking-by-detection paradigm focuses on associating the 2D detections using a temporal prior (e.g. Kalman filter). When applied to DC-videos containing rapid human and camera motions that violate the hand-crafted priors, such methods are brittle. Going beyond 2D, PHALP [40] separately extracts 3D human pose, appearance, and location with a multi-stage design from each video frame, and then assembles them for tracking. In contrast to these multi-stage methods, which are susceptible to errors in early stages, we explicitly learn the 3D human trajectory from temporal 5D cues in an end-to-end manner.", + "bbox": [ + 496, + 287, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular global 3D human trajectory reasoning. Most existing methods [64] that reason about global 3D human trajectories do so with static, calibrated, cameras in a multi-view setting. A few recent methods [31,58] have addressed the ill-posed problem of extracting the global motions of humans from monocular video. Liu et al. [31] employ a structure-from-motion (SfM) method [43] to estimate the camera poses from monocular videos captured by a dynamic camera. However, when the input video contains the movement of multiple subjects, it is hard for SfM methods to extract sufficiently many stable keypoints for reliable camera estimation. GLAMR [58] adopts a multi-stage pipeline to infer the global human trajectory from root-relative local human 3D poses estimated from each frame. The per-frame human pose estimates make it vulnerable to occlusion. Additionally, GLAMR relies on bounding boxes, ignoring scene-related information. Consequently, GLAMR fails in common scenarios like riding a bike or skating. In concurrent work (in this proceedings), SLAHMR [55] uses a multi-stage optimization-based approach that combines structure from motion with human motion priors to estimate 4D human trajectories in global coordinates; this is very computationally expensive. In contrast to previous multi-stage methods, TRACE simultaneously combines scene information and 3D human motions with a novel 5D representation to holistically exploit all temporal cues and to enable end-to-end training.", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8858", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aecf35d1f4ee96994fccd5c5d0b73fad77fa335a7ebeebdea5aa07a6a82b98a8.jpg", + "image_caption": [ + "Figure 2. TRACE Overview. TRACE takes a video sequence and regions in the first frame corresponding to the subjects to be tracked. TRACE encodes the video and its motion with temporal features and optical flow. A novel 3D Motion Offset map reasons about human trajectories in camera coordinates. The World Motion map represents the trajectory in global coordinates. A memory unit deals with occlusions by encoding the subject identities. TRACE is trained end-to-end to estimate the 3D shape and pose of multiple people throughout a video in global coordinates. See Sec. 3.1 for details." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 420, + 168, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 441, + 187, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall framework of TRACE is shown in Fig. 2. Given a video sequence captured with a dynamic camera $\\{f_i, i = 1, \\dots, N\\}$ with $N$ frames, the user specifies $K$ tracking subjects shown in the first frame. Our goal is to simultaneously recover the 3D pose, shape, identity, and trajectory of each subject in global coordinates. To achieve this, TRACE first extracts temporal features and then decodes each sub-task with a separate head network. First, via two parallel backbones, TRACE encodes the video and its motion into temporal image feature maps $F_i'$ and motion feature maps $O_i$ .", + "bbox": [ + 75, + 460, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Detection and Tracking branches take these features and perform multi-subject tracking in camera coordinates. Unlike BEV [45], our detection method takes temporal image features $\\pmb{F}_i'$ as input. It uses the features to detect the 3D human positions $\\pmb{t}_i$ and their confidence $c_i$ for all people in frame $f_i$ . The Mesh branch regresses all the human mesh parameters $(\\pmb{\\theta}_i, \\pmb{\\beta}_i)$ , in SMPL [32] format, from the input Feature maps. Unlike BEV, this branch takes both temporal image features and motion features.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The combined features $(F_{i}^{\\prime},O_{i})$ are fed to our novel Tracking branch to estimate the 3D Motion Offset map, indicating the 3D position change of each subject across frames. The new Memory Unit takes the 3D detection and its 3D motion offset as input. It then determines the subject identities and builds human trajectories $\\hat{t}_i$ of the $K$ subjects in camera coordinates. Note that, like BEV, our detection branch finds all the people in the video frames but our goal is to track only the $K$ input subjects. Consequently, the", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "memory unit filters out detected people who do not match the subject trajectories.", + "bbox": [ + 498, + 421, + 890, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, to estimate subject trajectories in global coordinates, the World branch estimates a world motion map, representing the 3D orientation $\\pmb{\\tau}_i$ and 3D translation offset $\\Delta T_{i}$ of the $K$ subjects in global coordinates. Accumulating $\\Delta T_{i}$ , starting with the 3D position $\\hat{t}_{1}$ of the tracked subjects in the first frame, gives their global 3D trajectory $\\pmb{T}$ . Note that the global (\"world\") coordinates are defined relative to the camera coordinates of the first frame.", + "bbox": [ + 498, + 452, + 892, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Holistic 5D Representation: Details", + "text_level": 1, + "bbox": [ + 500, + 587, + 808, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than directly estimating camera poses from environment keypoints [31] or inferring global human trajectories from local body poses [58], we develop a 5D representation to directly reason about human states, perform multiperson temporal association, and infer human trajectories in global coordinates. Learning a holistic 5D representation is the foundation of our one-stage framework. The representation has five main parts.", + "bbox": [ + 496, + 612, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "i. Temporal feature maps. To construct the temporal feature maps encoding 5D human states and scenes information, we need to extract both single-frame image features and the motion features between adjacent frames. Therefore, given frame $f_{i-1}$ and $f_i$ , we adopt a parallel two-branch structure to extract temporal image feature maps $\\pmb{F}_i'$ and motion feature maps $O_i$ for the current frame $f_i$ . First, in the image branch, we extract single-frame feature maps $F_{i-1}$ and $F_i$ with an image backbone (HRNet-32 [9]). To extract long-term and short-term motion features, we construct a temporal feature propagation module by combining", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a ConvGRU [47] module, a Deformable convolution [65] module, and a residual connection. With these, we fuse the image feature maps to generate a temporal image feature map $\\pmb{F}_i'$ . See Sup. Mat. for details and experimental analysis. Additionally, in the motion branch, we estimate the optical flow map $O_i$ between frames $f_{i-1}$ and $f_i$ with a motion backbone (RAFT [47]), to extract motion features of both people and scenes. From the combined temporal feature $(\\pmb{F}_i', O_i)$ , then we estimate five maps for the task.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ii. 3D detection maps. From the temporal image feature $\\pmb{F}_i'$ , we estimate a 3D Center map $M_C^{3D} \\in \\mathbb{R}^{1 \\times D \\times H \\times W}$ for coarse human detection and a 3D Localization map $M_L^{3D} \\in \\mathbb{R}^{1 \\times D \\times H \\times W}$ for fine localization. The two 3D detection maps are composited from the front-view maps and the 2D bird's-eye-view maps following BEV [45]. For $K$ subjects, we first parse out the detected 3D center positions $\\widetilde{\\pmb{t}}_i$ and their detection confidence $c_i$ from the 3D Center map. Then we sample the 3D Localization map at $\\widetilde{\\pmb{t}}_i$ to obtain the fine 3D localization offset vectors $\\Delta t_i$ . The predicted 3D translation of subjects to be tracked in camera coordinates is $\\pmb{t}_i = \\widetilde{\\pmb{t}}_i + \\Delta t_i$ .", + "bbox": [ + 76, + 229, + 472, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iii. 3D Motion Offset map. To track subjects between adjacent frames using $\\pmb{F}_i'$ and $O_i$ , we estimate a 3D Motion Offset map $M_M^{3D} \\in \\mathbb{R}^{3 \\times D \\times H \\times W}$ via a BEV-based 3D composition. We sample the $M_M^{3D}$ at detected 3D center positions $\\widetilde{\\pmb{t}}_i$ to obtain the 3D motion offset vectors $\\Delta m_i$ which represent the 3D position changes of the subjects from frame $f_{i-1}$ to frame $f_i$ in camera coordinates. With $\\Delta m_i, t_i, c_i$ as input, a memory unit (Sec. 3.3) predicts the tracked identities $D_i$ and the optimized 3D position $\\hat{t}_i$ in camera coordinates, which filters out low-confidence detections while keeping track of subjects through long-term occlusion.", + "bbox": [ + 75, + 412, + 472, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iv. Mesh Feature map. To regress the SMPL parameters of subjects, we first estimate a 2D Mesh Feature map $M_F^{2D} \\in \\mathbb{R}^{C \\times H \\times W}$ from temporal features $(F_i', O_i)$ . With the tracked 3D positions $\\widehat{t}_i = (u_i, v_i, d_i)$ , we sample a mesh feature vector from $M_F^{2D}$ at 2D positions $(u_i, v_i)$ . To differentiate the features of people at different depths, we map the $d_i$ to a 128-dim encoding vector via an embedding layer and add this with the mesh feature vector to regress SMPL [32] parameters $(\\theta_i, \\beta_i)$ . The SMPL model maps the pose and shape, $\\theta_i$ and $\\beta_i$ , to a 3D human body mesh $B \\in \\mathbb{R}^{6890 \\times 3}$ . With a sparse weight matrix $R \\in \\mathbb{R}^{Q \\times 6890}$ that describes the mapping from $B$ to $Q$ body keypoints, we can obtain the 3D positions of body keypoints $J \\in \\mathbb{R}^{Q \\times 3}$ via $RB$ . With the estimated 3D positions $\\widehat{t}_i$ and a predefined camera projection matrix $P$ , the projected 2D keypoints are $J_{2D} = P(J + \\widehat{t}_i)$ .", + "bbox": [ + 75, + 595, + 472, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "v. World Motion map. $M_W^{2D} \\in \\mathbb{R}^{6 \\times H \\times W}$ contains the 6D vectors that describe the 3D orientation and the relative translation of the subjects in global coordinates. We sample $M_W^{2D}$ at the 2D positions $(u_i, v_i) \\in \\widehat{t}_i$ of the tracked", + "bbox": [ + 75, + 838, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "subjects to obtain their 3D body orientation $\\pmb{\\tau}_{i}$ and 3D body motion offset $\\Delta T_{i}$ . For the tracked subject $k$ , $\\Delta T_{i}^{k}$ represents their 3D position change from frame $f_{i-1}$ to frame $f_{i}$ in global coordinates. We take the camera coordinate system of the first frame as the global coordinate system for the video. The global 3D trajectories of all tracked subjects are obtained by accumulating $\\Delta T_{i}$ to their 3D position $\\hat{t}_{1}$ in the first frame, $\\pmb{T} = \\{\\hat{t}_{1}, \\hat{t}_{1} + \\Delta T_{2}, \\hat{t}_{1} + \\Delta T_{2} + \\Delta T_{3}, \\dots\\}$ . Combining the 3D mesh, ID, and the global trajectories, the network reasons about 3D motions and trajectories of $K$ tracked subjects in global coordinates.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Tracking with a Memory Unit", + "text_level": 1, + "bbox": [ + 498, + 268, + 767, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct the 3D trajectory of each subject by associating the 3D detections $\\pmb{t}_i$ over time with a 3D motion offset $\\Delta m_i$ . To deal with long-term occlusions, we design the Memory Unit for persistent tracking, which will keep the memory for the full sequence. The memory unit stores the human states during inference and is not used for training. With predicted 3D positions $\\pmb{t}_i$ , detection confidence $c_i$ , and 3D motion offsets $\\Delta m_i$ as inputs, the memory unit can track online. In each process, we have three stages.", + "bbox": [ + 496, + 292, + 893, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "i. Initialization. First, we discard predicted 3D positions $\\pmb{t}_i = (\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i)$ whose detection confidence is below a threshold $\\lambda_c$ . We observe that our input video is usually shot by tracking the subjects, therefore, we discard the detection whose $1 / z_i$ is below the scale threshold $\\lambda_s$ . To suppress duplicate detections, we find the detection pairs whose Euclidean distance is below a pre-defined threshold $\\lambda_d$ and discard the detection with lower detection confidence. In the first frame, we use the 3D positions $\\pmb{t}_1$ and detection confidences $\\pmb{c}_i$ of $K$ subjects to initialize the $K$ memory nodes.", + "bbox": [ + 496, + 430, + 893, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ii. Memory node matching. The memory nodes store the 3D position $t_{i-1}^k$ of subject $k$ in previous frames. We match the memory node with the new filtered detections in the current frame. We select the predicted 3D translation $t_i$ that best matches the 3D trajectory as the optimized 3D position $\\hat{t}_i$ . Specifically, with filtered 3D positions $t_i$ and 3D motion offset $\\Delta m_i$ , we calculate the Euclidean distance matrix between $W t_{i-1}^k$ and $W (t_i - \\Delta m_i)$ where $W \\in \\mathbb{R}^3$ is a distance weight vector. To avoid the effects of the nonlinear relationship between depth $z_i$ and human scale in images, we convert the depth value $z$ to $1/(1 + z)$ when calculating the distance matrix. Using Hungarian matching, we keep the matched pairs whose matching distance is below a threshold $\\lambda_m$ and use them for memory update. We use ground truth tracks for training and only perform matching during inference.", + "bbox": [ + 496, + 583, + 895, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iii. Memory update. We update the successfully matched memory nodes with the new 3D position and detection confidence. For the memory nodes without a matched detection, we accumulate the time since failure. Then we remove the memory nodes whose failure time is", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8860", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "above a threshold $\\lambda_{f}$ . Tracking can be done in two modes: on-line or off-line. The former does not allow looking back in time, while the latter does. In the off-line mode, if a new detection re-activates a non-matched memory node, the non-matched part of the 3D trajectory is linearly interpolated. Finally, the memory unit outputs the latest 3D positions $\\widehat{t}_i$ and tracking IDs $D_{i}$ of all memory nodes.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. DynaCam Dataset", + "text_level": 1, + "bbox": [ + 76, + 207, + 253, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Even with a powerful 5D representation, we still lack in-the-wild data for training and evaluation of global human trajectory estimation. However, collecting global human trajectory and camera poses for natural DC-videos is difficult. Therefore, we create a new dataset, DynaCam, by simulating camera motions to convert in-the-wild videos captured by static cameras to DC-videos.", + "bbox": [ + 75, + 231, + 468, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use over 1000 video clips captured by static regular cameras from the MPII Human Pose Database [3] as well as videos from the InterNet [2]. We also use over 200 panoramic video clips that are either recorded by us with an Insta360 RS panoramic camera or are downloaded from the InterNet [1]. We manually design the 3D rotation and field of view (FOV) of dynamic cameras to track the subjects in panoramic videos. With the designed camera motions, we can project the panoramic frames into perspective views. Also, to simulate the 3D translation of dynamic cameras, we crop the videos captured by static cameras with sliding windows. In this way, we can obtain abundant in-the-wild DC-videos with accurate camera pose annotations. Then we perform 2D human detection, tracking, and 2D pose estimation viaYOLOX [17], ByteTrack [63], and ViT-Pose [54], respectively, to obtain 2D pose sequences of each subject. We estimate SMPL parameters by fitting the 2D poses using EFT [20] or ProHMR [25] and solve for their 3D positions in camera coordinates via the PnP algorithm (RANSAC [16]). Finally, we solve for the 3D human trajectories in the global coordinates with camera pose annotations. We manually filter out the failure cases. In this way, we generate more than 500 annotated DC-videos containing over 48K frames. More than half of video frames are generated from panoramic videos.", + "bbox": [ + 75, + 337, + 472, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limitations: The videos generated with our process only approximate real DC-videos shot in the wild since they lack perspective effects. Despite this they prove useful for training TRACE.", + "bbox": [ + 75, + 714, + 468, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 785, + 230, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TRACE is supervised by the weighted sum of 15 loss terms that fall into two groups: temporal motion losses and standard image losses. Here we focus on the novel temporal losses. Please refer to the Sup. Mat. for details of all losses.", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To learn the temporal motion, we introduce a 3D motion offset loss $\\mathcal{L}_m$ and a 6D world motion loss $\\mathcal{L}_W$ . $\\mathcal{L}_m$ is the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$L_{2}$ loss between the predicted 3D motion offset $\\Delta m_{i}$ and $(\\pmb{t}_i^{\\prime} - \\pmb{t}_{i - 1}^{\\prime})$ where $t_i^\\prime$ is the ground truth 3D human position at frame $f_{i}$ in our pre-defined camera coordinates (FOV=50°), which is solved for via the PnP algorithm [16]. $\\mathcal{L}_W$ consists of six parts, including an $L_{2}$ loss on the global 3D trajectory $\\pmb{T}$ , an $L_{2}$ loss on the velocity/acceleration of 3D trajectory nodes $\\dot{T} / \\ddot{T}$ , an $L_{2}$ loss of the velocity/acceleration of the 3D foot keypoints in global coordinates, and an $L_{2}$ loss on the global 3D body orientation $\\pmb{\\tau}_i$ .", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 242, + 633, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 267, + 718, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. During training, we directly use the ground truth trajectory of subjects to replace the estimated trajectory $\\hat{t}$ for sampling the parameters. We use the pretrained backbone of BEV [45] as the image backbone. We use RAFT [47] as the optical flow backbone. The training consists of two stages. In the first stage, we freeze the weights of the backbones and train the head network for 40 epochs with a learning rate of 5e-5. Then we train the image backbone and the head network together for 10 epochs with a learning rate of 1e-5. We use four V100-16GB GPUs for training. Limited by the GPU memory, we sample 4 video clips as a batch at each iteration; the clip length is 10 frames.", + "bbox": [ + 496, + 291, + 890, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and evaluation datasets. For training, we use three 3D human pose datasets (Human3.6M [19], MPIINF-3DHP [34, 35], and 3DPW [50]), two 2D human pose datasets (PennAction [62] and PoseTrack [12]), and our DynaCam dataset. We evaluate TRACE on two multi-person in-the-wild benchmarks, 3DPW [50] and MuPoTS-3D [35], and DynaCam. 3DPW videos are most consistent with our tracking scenario. Unfortunately, not all 3DPW videos have complete tracking annotations. We select the 16 videos that do and call this subset Dyna3DPW. We use this challenging subset to evaluate tracking and HPS accuracy in complex scenes with a moving camera.", + "bbox": [ + 496, + 473, + 892, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. For global 3D trajectory estimation, we compute the average 3D position (Traj) and velocity (Velocity) error of the predicted global 3D trajectory $T$ . For multi-object tracking, we report the ID switch (IDs), Multi-Object Tracking Accuracy (MOTA [5]), Identification F1-score (IDF1 [41]), and Higher Order Tracking Accuracy (HOTA [33]). To assess the accuracy of 3D human pose/shape estimation, we compute the Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PMPJPE), and Mean Vertex Error (MVE).", + "bbox": [ + 496, + 654, + 893, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Please refer to Sup. Mat. for more details.", + "bbox": [ + 517, + 805, + 795, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons to State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 500, + 830, + 852, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global 3D trajectory estimation. We aim to estimate the global human trajectory from dynamic cameras. We do not explicitly estimate the camera motion. Instead, we use a", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8861", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/064327ecdf6902a3f8ea4936e7a2ff7497f90a40a492bac4ca3ded7c8799c174.jpg", + "image_caption": [ + "(a) Tracking on 3DPW" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 356, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/023b77276145be755fef9e21f37451f08319d524896bf44609619ef6bb960fc5.jpg", + "image_caption": [ + "(b) Tracking on MuPoTS-3D" + ], + "image_footnote": [], + "bbox": [ + 359, + 88, + 594, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6043dc6c2841750742dabcce635b3220c2445b87d0468615723557c5418047d1.jpg", + "image_caption": [ + "(c) Global 3D trajectory estimation on DynaCam", + "Figure 3. Qualitative comparisons to SOTA methods (ByteTrack [63], PHALP [40], and GLAMR [58]) on 3DPW, MuPoTS-3D, and DynaCam. Red dash boxes highlight the ID switches. Yellow dot-dash boxes highlight the misses under occlusions." + ], + "image_footnote": [], + "bbox": [ + 611, + 87, + 883, + 431 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e82eb3e12ca5f1bcf3e0316cdfef1f92e1ea8ab384519fabe725c61e0b791a0f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTranslating CameraRotating CameraFPS
Traj↓Velocity↓Traj↓Velocity↓
BEV+DPVO2.0240.1543.6220.1195.7
GLAMR [58]1.7450.1051.7590.0551.5
TRACE1.4330.0821.0380.04521.2
", + "bbox": [ + 81, + 503, + 467, + 580 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ede17ea21fef5e7ef2f22646a9c005664e90427d7fdf0adcff1e00d4f5e1a8b.jpg", + "table_caption": [ + "Table 1. Errors in estimated 3D human trajectory in world coordinates (in $m$ ) and runtime efficiency on the DynaCam dataset." + ], + "table_footnote": [], + "table_body": "
MethodsIDs↓MOTA↑IDF1↑HOTA↑
Trackformer [36]4324.962.753.2
Tracktor [4]5351.570.950.3
FlowPose [51]4921.467.141.8
T3DP [39]3862.179.159.2
PHALP [40]2266.281.459.4
ByteTrack [63]1573.384.663.5
TRACE086.993.465.3
", + "bbox": [ + 102, + 627, + 439, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "world motion map to represent the global trajectory in world coordinates, which implies the camera motion. Therefore, we evaluate the global trajectory error, instead of the camera pose, in Tab. 3. First, we evaluate global 3D trajectory estimation on DynaCam. We compare TRACE with two baseline solutions. The first one uses BEV to estimate the subjects in camera coordinates and DPVO [48], a SLAM", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f3cac83016639bb3231781c103e6ac5dae99eb2a380eb8f8e206423fed80e00.jpg", + "table_caption": [ + "Table 2. Tracking results on MuPoTS-3D [35]. Results of the compared methods [4,36,39,40,51] are from [40]." + ], + "table_footnote": [], + "table_body": "
MethodsIDs↓MOTA↑IDF1↑HOTA↑
PHALP [40]597.993.774.2
YOLOX+ByteTrack397.398.273.1
BEV+ByteTrack3793.679.159.3
TRACE w/o MO495.897.372.7
TRACE199.399.774.7
", + "bbox": [ + 517, + 503, + 870, + 590 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/735f23b597df6aabf78ea09dfe56b21bc0bde3e6e5abf56a3dce28d9434d4f87.jpg", + "table_caption": [ + "Table 3. Tracking results on Dyna3DPW [50]. w/o MO means ablating the estimated 3D motion offsets for tracking." + ], + "table_footnote": [], + "table_body": "
MethodPAMPJPE↓MPJPE↓PVE↓
HybrIK [27]45.074.186.5
METRO [30]47.977.188.2
GLAMR [58]50.7--
CLIFF [28]43.069.081.2
D&D [26]42.773.788.6
ROMP [44]47.376.793.4
BEV [45]46.978.592.3
TRACE37.879.197.3
", + "bbox": [ + 547, + 630, + 841, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparisons with SOTA methods for 3D human pose and shape estimation on the 3DPW test set.", + "bbox": [ + 500, + 762, + 890, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method, to estimate the camera and its motion; we call this BEV+DPVO. As shown in Tab. 1, TRACE significantly outperforms BEV+DPVO in the accuracy of global 3D trajectory estimation. The moving people in the scene make it hard for a SLAM method to extract stable corresponding keypoints. Additionally, our synthetic camera motions", + "bbox": [ + 496, + 810, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8862", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "differ from real camera motions and this may hurt DPVO's performance. A more direct comparison is to GLAMR [58]. TRACE outperforms GLAMR in both accuracy and efficiency. In Fig. 3(c), we also perform visual comparisons with GLAMR on DynaCam. These results demonstrate the benefit of estimating global human trajectory using a holistic 5D representation. We provide more results in the supplemental video.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-subject tracking. To evaluate the performance of tracking subjects with dynamic cameras in real-world scenes, we compare TRACE with recent methods on Dyna3DPW. PHALP [40] is a recent (SOTA) method that uses 3D cues and appearance to track people using SMPL. YOLOX+ByteTrack [17, 63] is a recently proposed and popular tracking-by-detection solution. These methods are designed to track all the people in a scene. Therefore, we process their results to avoid them being penalized for tracking unlabeled passers-by; 3DPW has annotations for at most 2 people in a scene but some scenes contain many people. We first obtain their tracking results using their official code. We then select the tracking results that achieve maximum IoU with the labeled ground truth subjects; we use these tracks for evaluation. Note that, for a fair comparison with ByteTrack, TRACE runs in an on-line mode, without optimizing the past results. As shown in Tab. 3, TRACE outperforms PHALP, the previous 3D-representation-based method [40], and tracking-by-detection [17, 63] methods on Dyna3DPW. To evaluate the tracking robustness under long-term occlusion, we evaluate TRACE on MuPoTS-3D. Results of previous SOTA methods [4, 36, 39, 40, 51] are from [40]. Again, for a fair comparison, we filter out the unlabeled people in the ByteTrack results. As shown in Tab. 2, TRACE significantly outperforms previous SOTA methods. In particular, TRACE significantly reduces ID switches under long-term occlusions. These results illustrate the effectiveness and robustness of the proposed method for in-the-wild videos. The qualitative comparisons are presented in Fig. 3 and the supplemental video.", + "bbox": [ + 75, + 215, + 472, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D human pose and shape estimation. Finally, we evaluate 3D human regression performance in DC-video using the 3DPW test set. Because 3DPW does not provide ground-truth 3D translation in world coordinates, we evaluate root-relative 3D pose. We compare TRACE with the recent one-stage [44,45] and multi-stage [26-28,30,58] methods. Tab. 4 shows that TRACE significantly outperforms the previous method by $11.4\\%$ in terms of PMPJPE. These results demonstrate that learning a holistic 5D representation in an end-to-end manner is powerful. We adopt a fixed field of view $(\\mathrm{FOV} = 50^{\\circ})$ to define our camera space, which is very different from that of 3DPW video. This results in a misalignment between the predicted 3D orientation of the body mesh and the ground truth, resulting in higher MPJPE and MVE. This illustrates the importance of also estimating", + "bbox": [ + 75, + 674, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the camera so that our estimated global coordinate system is consistent with the real world. This is future work.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 498, + 133, + 663, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal 5D representation v.s. image-level 3D representation. We go beyond BEV's image-level 3D representation and build a temporal 5D representation. As shown in Tab. 1, 3, and 4, TRACE outperforms BEV or the multistage solutions using BEV on most metrics. This demonstrates the value of learning a holistic 5D representation.", + "bbox": [ + 496, + 155, + 890, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Motion Offset map. We also evaluate the effect of using predicted 3D motion offsets for tracking. As shown in Tab. 3, 3D motion offsets improve performance by $3.5\\%$ , $2.4\\%$ , and $2.7\\%$ in terms of MOTA, IDF1, and HOTA.", + "bbox": [ + 496, + 248, + 893, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 498, + 328, + 625, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human pose and shape estimation is not an end to itself. Rather, estimating the 3D human in motion is useful for many tasks from behavior analysis to computer graphics. However, to be useful, it is important to know the motion of humans with respect to the 3D scene and other people. This means that HPS methods must estimate humans in a global coordinate system and provide consistent tracks of people across time. For generality, they also need to be able to do this from arbitrary moving cameras.", + "bbox": [ + 496, + 352, + 890, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To tackle these challenging problems, we propose a novel 5D representation and a new neural architecture that reasons about people in 5D; that is, their 3D position, temporal trajectory, and identity. Moving to a 5D representation enables our method, TRACE, to take a holistic view of the video, processing full frames and incorporating temporal features. The core innovation of TRACE lies in its novel temporal representation in the form of new \"maps\" that represent the motion of people across time in the camera and global coordinates. These allow TRACE to be trained end-to-end, thus exploiting rich information from the video to solve the task. TRACE is the first such single-shot method for 3D HPS estimation from video and it achieves SOTA results on common benchmarks.", + "bbox": [ + 496, + 489, + 890, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work should look at explicitly estimating the camera, using training data like BEDLAM [6], which contains complex human motion, 3D scenes, and camera motions. We believe that camera motion and human motion provide complimentary information that can be used to recover human motion in world coordinates with metric accuracy.", + "bbox": [ + 496, + 704, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was partially supported by the National Key R&D Program of China under Grant (No. 2020AAA0103800) and Beijing Nova Program (No. 20220484063).", + "bbox": [ + 496, + 801, + 890, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MJB Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2023.txt", + "bbox": [ + 498, + 869, + 888, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bilibili. https://www.bilibili.com/.6", + "[2] Pexels. https://www.pexels.com.6", + "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6", + "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, pages 941-951, 2019. 3, 7, 8", + "[5] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, pages 1-10, 2008. 6", + "[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of 3d human bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 2, 8", + "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, pages 561-578, 2016. 2", + "[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 3", + "[9] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, pages 5386-5395, 2020. 4", + "[10] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, pages 436-454, 2020. 3", + "[11] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 3", + "[12] Andreas Doering, Di Chen, Shanshan Zhang, Bernt Schiele, and Juergen Gall. Posetrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In CVPR, pages 20963-20972, 2022. 6", + "[13] Mahsa Ehsanpour, Fatemeh Saleh, Silvio Savarese, Ian Reid, and Hamid Rezatofighi. Jrdb-act: A large-scale dataset for spatio-temporal action, social group and activity detection. In CVPR, pages 20983–20992, 2022. 3", + "[14] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. A mobile vision system for robust multi-person tracking. In CVPR, pages 1-8, 2008. 3", + "[15] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoşa Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, pages 10849-10859, 2021. 3", + "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 6", + "[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3, 6, 8", + "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 3", + "[19] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. TPAMI, 36(7):1325-1339, 2013. 6", + "[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In ECCV, 2020. 3, 6", + "[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 3", + "[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. VIBE: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3", + "[23] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In ICCV, pages 11035-11045, 2021. 2, 3", + "[24] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3", + "[25] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In CVPR, pages 11605-11614, 2021. 6", + "[26] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&d: Learning human dynamics from dynamic camera. In ECCV, 2022. 3, 7, 8", + "[27] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In CVPR, pages 3383-3393, 2021. 7, 8", + "[28] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, pages 590-606, 2022. 3, 7, 8", + "[29] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In CVPR, 2023. 3", + "[30] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 7, 8", + "[31] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, pages 930-939, 2021. 2, 3, 4", + "[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):1-16, 2015. 3, 4, 5" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, pages 548-578, 2021. 6", + "[34] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3D human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 6", + "[35] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 3, 6, 7", + "[36] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 7, 8", + "[37] Gyeongsik Moon and Kyoung Mu Lee. Pose2Pose: 3d positional pose-guided 3d rotational pose prediction for expressive 3d human pose and mesh estimation. arXiv, 2020. 2, 3", + "[38] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. TexturePose: Supervising human mesh estimation with texture consistency. In ICCV, pages 803-812, 2019. 2, 3", + "[39] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. In NeurIPS, 2021. 3, 7, 8", + "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3D appearance, location & pose. In CVPR, 2022. 2, 3, 7, 8", + "[41] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6", + "[42] Weijian Ruan, Wu Liu, Qian Bao, Jun Chen, Yuhao Cheng, and Tao Mei. Pointet: pose-guided ovonic insight network for multi-person pose tracking. In MM, pages 284-292, 2019. 3", + "[43] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 3", + "[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 1, 3, 7, 8", + "[45] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[46] Yu Sun, Yun Ye, Wu Liu, Wenpeng Gao, YiLi Fu, and Tao Mei. Human mesh recovery from monocular images via a skeleton-disentangled representation. In ICCV, pages 5348-5357, 2019. 2, 3", + "[47] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 5, 6", + "[48] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 7" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In CVPR, 2023. 2", + "[50] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 6, 7", + "[51] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018. 3, 7, 8", + "[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3", + "[53] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3", + "[54] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. arXiv preprint arXiv:2204.12484, 2022. 3, 6", + "[55] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 3", + "[56] Hongwei Yi, Chun-Hao P. Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J. Black. MIME: Human-aware 3D scene generation. In CVPR, 2023. 2", + "[57] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2", + "[58] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In CVPR, 2022. 2, 3, 4, 7, 8", + "[59] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3D pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 3", + "[60] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 2, 3", + "[61] Shanshan Zhang, Rodrigo Benenson, and Bernt Schiele. Citypersons: A diverse dataset for pedestrian detection. In CVPR, 2017. 3", + "[62] Weiyu Zhang, Menglong Zhu, and Konstantinos G Derpanis. From actemes to action: A strongly-supervised representation for detailed action understanding. In ICCV, pages 2248-2255, 2013. 6", + "[63] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 2, 3, 6, 7, 8", + "[64] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In ICCV, pages 6239-6249, 2021. 3" + ], + "bbox": [ + 501, + 90, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[65] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5", + "bbox": [ + 78, + 90, + 470, + 148 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8866", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_model.json b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_model.json new file mode 100644 index 0000000000000000000000000000000000000000..df16294803a312cdbc4c42960abdef9573643372 --- /dev/null +++ b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.131, + 0.713, + 0.175 + ], + "angle": 0, + "content": "TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.188, + 0.745, + 0.206 + ], + "angle": 0, + "content": "Yu Sun\\(^{1*}\\) Qian Bao\\(^{2\\dagger}\\) Wu Liu\\(^{2\\dagger}\\) Tao Mei\\(^{4}\\) Michael J. Black\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.207, + 0.742, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)Harbin Institute of Technology \\(^{2}\\)Explore Academy of JD.com" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.225, + 0.742, + 0.243 + ], + "angle": 0, + "content": "3Max Planck Institute for Intelligent Systems HiDream.ai Inc." + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.246, + 0.704, + 0.26 + ], + "angle": 0, + "content": "yusunhit@gmail.com, baoqian@jd.com, liuwul@jd.com" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.264, + 0.656, + 0.277 + ], + "angle": 0, + "content": "tmei@hidream.ai, black@tuebingen.mpg.de" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.288, + 0.136, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.297, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Dynamic Monocular Camera" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.311, + 0.278, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.31, + 0.421, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.31, + 0.562, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.31, + 0.704, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.31, + 0.875, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.435, + 0.748, + 0.446 + ], + "angle": 0, + "content": "(a) Videos Captured by Dynamic Monocular Camera and Human 3D Recovery in Each Frame" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.446, + 0.483, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.6, + 0.47, + 0.612 + ], + "angle": 0, + "content": "(b) The 3D humans recovered by BEV [45] in camera coordinates" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.446, + 0.88, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.6, + 0.858, + 0.612 + ], + "angle": 0, + "content": "(c) The 5D humans recovered by TRACE in global coordinates" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.616, + 0.893, + 0.673 + ], + "angle": 0, + "content": "Figure 1. 5D temporal reconstruction of multiple 3D people in global coordinates with a dynamic camera. We introduce TRACE, a monocular one-stage method with a holistic 5D representation that enables the network to explicitly reason about human motion across time. Unlike previous methods [44,45] (b) that regress all 3D people from each frame in camera coordinates, TRACE (c) tracks the subjects (a) presented in the first frame through time and recovers their global trajectories in global coordinates in one shot." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.69, + 0.314, + 0.706 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.471, + 0.86 + ], + "angle": 0, + "content": "Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel ar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.895, + 0.844 + ], + "angle": 0, + "content": "chitectural components. Most importantly, it uses two new \"maps\" to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code1 and dataset2 are released for research purposes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.415, + 0.888 + ], + "angle": 0, + "content": "*This work was done when Yu Sun was an intern at JD.com." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.415, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.875, + 0.833, + 0.888 + ], + "angle": 0, + "content": "1https://www.yusun.work/TRACE/TRACE.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.811, + 0.9 + ], + "angle": 0, + "content": "2https://github.com/Arthur151/DynaCam" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.875, + 0.833, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8856" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.116, + 0.468, + 0.326 + ], + "angle": 0, + "content": "The estimation of 3D human pose and shape (HPS) has many applications and there has been significant recent progress [6, 7, 21-24, 37, 38, 46, 49, 56, 57, 60]. Most methods, however, reason only about a single frame at a time and estimate humans in camera coordinates. Moreover, such methods do not track people and are unable to recover their global trajectories. The problem is even harder in typical hand-held videos, which are filmed with a dynamic, moving, camera. For many applications of HPS, single-frame estimates in camera coordinates are not sufficient. To capture human movement and then transfer it to a new 3D scene, we must have the movement in a coherent global coordinate system. This is a requirement for computer graphics, sports, video games, and extended reality (XR)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.328, + 0.468, + 0.477 + ], + "angle": 0, + "content": "Our key insight is that most methods estimate humans in 3D, whereas the true problem is 5D. That is, a method needs to reason about 3D space, time, and subject identity. With a 5D representation, the problem becomes tractable, enabling a holistic solution that can exploit the full video to infer multiple people in a coherent global coordinate frame. As illustrated in Fig. 1, we develop a unified method to jointly regress the 3D pose, shape, identity, and global trajectory of the subjects in global coordinates from monocular videos captured by dynamic cameras (DC-videos)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.479, + 0.468, + 0.733 + ], + "angle": 0, + "content": "To achieve this, we deal with two main challenges. First, DC-videos contain both human motion and camera motion and these must be disentangled to recover the human trajectory in global coordinates. One idea would be to recover the camera motion relative to the rigid scene using structure-from-motion (SfM) methods (e.g. [31]). In scenes containing many people and human motion, however, such methods can be unreliable. An alternative approach is taken by GLAMR [58], which infers global human trajectories from local 3D human poses, without taking into account the full scene. By ignoring evidence from the full image, GLAMR fails to capture the correct global motion in common scenarios, such as biking, skating, boating, running on a treadmill, etc. Moreover, GLAMR is a multi-stage method, with each stage dependent on accurate estimates from the preceding one. Such approaches are more brittle than our holistic, end-to-end, method." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.735, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The other challenge, as shown in the upper right corner of Fig. 1, is that severe occlusions are common in videos with multiple people. Currently, the most popular tracking strategy is to infer the association between 2D detections using a temporal prior (e.g. Kalman filter) [63]. However, in DC-videos, human motions are often irregular and can easily violate hand-crafted priors. PHALP [40] is one of the few methods to address this for 3D HPS. It uses a classical, multi-stage, detection-and-tracking formulation with heuristic temporal priors. It does not holistically reason about the sequence and is not trained end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.227 + ], + "angle": 0, + "content": "To address these issues, we reason about people using a 5D representation and capture information from the full image and the motion of the scene. This holistic reasoning enables the reliable recovery of global human trajectories and subject tracking using a single-shot method. This is more reliable than multi-stage methods because the network can exploit more information to solve the task and is trained end-to-end. No hand-crafted priors are needed and the network is able to share information among modules." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.231, + 0.891, + 0.426 + ], + "angle": 0, + "content": "Specifically, we develop TRACE, a unified one-stage method for Temporal Regression of Avatars with dynamic Cameras in 3D Environments. The architecture is inspired by BEV [45], which directly estimates multiple people in depth from a single image using multiple 2D maps. BEV uses a 2D map representing an imaginary, \"top down\", view of the scene. This is combined with an image-centric 2D map to reason about people in 3D. Our key insight is that the idea of maps can be extended to represent how people move in 3D. With this idea, TRACE introduces three new modules to holistically model 5D human states, performing multi-person temporal association, and inferring human trajectories in global coordinates; see Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.43, + 0.891, + 0.806 + ], + "angle": 0, + "content": "First, to construct a holistic 5D representation of the video, we extract temporal image features by fusing single-frame feature maps from the image backbone with a temporal feature propagation module. We also compute the optical flow between adjacent frames with a motion backbone. The optical flow provides short-term motion features that carry information about the motion of the scene and the people. Second, to explicitly track human motions, we introduce a novel 3D motion offset map to establish the association of the same person across adjacent frames. This map contains a 3D offset vector at each position, which represents the difference between the 3D positions of the same subject from the previous frame to the current frame in camera coordinates. We also introduce a memory unit to keep track of subjects under long-term occlusion. Note that the 3D trajectories are built in camera space, and TRACE uses a novel world motion map that transfers the trajectories to global coordinates. At each position, this map contains a 6D vector to represent the difference between the 3D positions of the corresponding subject from the previous frame to the current frame and its 3D orientation in world coordinates. Taken together, this novel network architecture goes beyond prior work by taking information from the full video frames to address detection, pose estimation, tracking, and occlusion in a holistic network that is trained end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.811, + 0.891, + 0.9 + ], + "angle": 0, + "content": "To enable training and evaluation of global human trajectory estimation from in-the-wild DC-videos, we build a new dataset, DynaCam. Since collecting global human trajectories and camera poses with in-the-wild DC-videos is difficult, we simulate a moving camera using publicly available in-the-wild panoramic videos and regular videos cap" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8857" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "tured by static cameras. In this way, we create more than 500 in-the-wild DC-videos with precise camera pose annotations. Then we generate pseudo-ground-truth 3D human annotations via fitting [20] SMPL [32] to detected 2D pose sequences [17,54,63]. With 2D/3D human pose and camera pose annotations, we can obtain the global human trajectories using the PnP algorithm [16]. This dataset is sufficient to train TRACE to deal with dynamic cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.363 + ], + "angle": 0, + "content": "We evaluate TRACE on two multi-person in-the-wild benchmarks (3DPW [50] and MuPoTS-3D [35]) and our DynaCam dataset. On 3DPW, TRACE achieves the state-of-the-art (SOTA) PA-MPJPE of \\(37.8\\mathrm{mm}\\), less the current best (42.7 [26]). On MuPoTS-3D, TRACE outperforms previous 3D-representation-based methods [39, 40] and tracking-by-detection methods [63] on tracking people under long-term occlusion. On DynaCam, TRACE outperforms GLAMR [58] in estimating the 3D human trajectory in global coordinates from DC-videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.471, + 0.545 + ], + "angle": 0, + "content": "In summary, our main contributions are: (1) We introduce a 5D representation and use it to learn holistic temporal cues related to both 3D human motions and the scene. (2) We introduce two novel motion offset representations to explicitly model temporal multi-subject association and global human trajectories from temporal clues in an end-to-end manner. (3) We estimate long-term 3D human motions over time in global coordinates, achieving SOTA results. (4) We collect the DynaCam dataset of DC-videos with pseudo ground truth, which facilitates the training and evaluation of global human trajectory estimation. The code and dataset are publicly available for research purposes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.22, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Monocular 3D mesh regression with full images. Most existing methods [21-24, 26, 29, 37, 38, 46, 52, 53, 59, 60] take a multi-stage detection-based pipeline to estimate 3D HPS from cropped image patches, which exclude important cues, such as camera information and human-scene relationships. A few recent multi-stage [28] and one-stage [44, 45] methods have made steps towards using the full-image information. For instance, CIUFF [28] estimates 3D HPS by taking into account the bounding box locations, giving the method camera information and improving accuracy. To directly estimates multiple people at once from the full image, ROMP [44] introduces a 2D Center heatmap and a Mesh parameter map to represent 2D human locations and 3D human body meshes, respectively. BEV [45] goes beyond ROMP by introducing an imaginary bird's-eye-view map, which is combined with the front-view maps to construct a 3D view in camera coordinates. However, they only model 3D HPS in camera coordinates from a single image. Using \"maps\" like ROMP and BEV, TRACE also looks at the full image. We go further, however, by introducing novel maps that model human motions across a video se" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.696, + 0.107 + ], + "angle": 0, + "content": "quence in global coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Tracking datasets. While there are many tracking datasets [8, 10, 11, 14, 18, 61] with 2D annotations, only a few [13, 15] capture the 3D trajectory of pedestrians. In both cases, the scene and human activities are limited. To address this, we use 3DPW [50] and MuPoTS-3D [35] for tracking evaluation. 3DPW is the most relevant dataset for our task and it provides a real-world test case. 3DPW contains videos that are captured by a moving camera that follows subjects to record their activities in many daily scenes. MuPoTS-3D contains rich multi-person interaction scenes with long-term occlusions for tracking evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Tracking 3D people through occlusions. Most existing methods [4, 42, 51, 63] perform tracking using 2D image cues. The classic tracking-by-detection paradigm focuses on associating the 2D detections using a temporal prior (e.g. Kalman filter). When applied to DC-videos containing rapid human and camera motions that violate the hand-crafted priors, such methods are brittle. Going beyond 2D, PHALP [40] separately extracts 3D human pose, appearance, and location with a multi-stage design from each video frame, and then assembles them for tracking. In contrast to these multi-stage methods, which are susceptible to errors in early stages, we explicitly learn the 3D human trajectory from temporal 5D cues in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Monocular global 3D human trajectory reasoning. Most existing methods [64] that reason about global 3D human trajectories do so with static, calibrated, cameras in a multi-view setting. A few recent methods [31,58] have addressed the ill-posed problem of extracting the global motions of humans from monocular video. Liu et al. [31] employ a structure-from-motion (SfM) method [43] to estimate the camera poses from monocular videos captured by a dynamic camera. However, when the input video contains the movement of multiple subjects, it is hard for SfM methods to extract sufficiently many stable keypoints for reliable camera estimation. GLAMR [58] adopts a multi-stage pipeline to infer the global human trajectory from root-relative local human 3D poses estimated from each frame. The per-frame human pose estimates make it vulnerable to occlusion. Additionally, GLAMR relies on bounding boxes, ignoring scene-related information. Consequently, GLAMR fails in common scenarios like riding a bike or skating. In concurrent work (in this proceedings), SLAHMR [55] uses a multi-stage optimization-based approach that combines structure from motion with human motion priors to estimate 4D human trajectories in global coordinates; this is very computationally expensive. In contrast to previous multi-stage methods, TRACE simultaneously combines scene information and 3D human motions with a novel 5D representation to holistically exploit all temporal cues and to enable end-to-end training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.894, + 0.404 + ], + "angle": 0, + "content": "Figure 2. TRACE Overview. TRACE takes a video sequence and regions in the first frame corresponding to the subjects to be tracked. TRACE encodes the video and its motion with temporal features and optical flow. A novel 3D Motion Offset map reasons about human trajectories in camera coordinates. The World Motion map represents the trajectory in global coordinates. A memory unit deals with occlusions by encoding the subject identities. TRACE is trained end-to-end to estimate the 3D shape and pose of multiple people throughout a video in global coordinates. See Sec. 3.1 for details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.421, + 0.169, + 0.436 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.443, + 0.188, + 0.457 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.629 + ], + "angle": 0, + "content": "The overall framework of TRACE is shown in Fig. 2. Given a video sequence captured with a dynamic camera \\(\\{f_i, i = 1, \\dots, N\\}\\) with \\(N\\) frames, the user specifies \\(K\\) tracking subjects shown in the first frame. Our goal is to simultaneously recover the 3D pose, shape, identity, and trajectory of each subject in global coordinates. To achieve this, TRACE first extracts temporal features and then decodes each sub-task with a separate head network. First, via two parallel backbones, TRACE encodes the video and its motion into temporal image feature maps \\(F_i'\\) and motion feature maps \\(O_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.764 + ], + "angle": 0, + "content": "The Detection and Tracking branches take these features and perform multi-subject tracking in camera coordinates. Unlike BEV [45], our detection method takes temporal image features \\( \\pmb{F}_i' \\) as input. It uses the features to detect the 3D human positions \\( \\pmb{t}_i \\) and their confidence \\( c_i \\) for all people in frame \\( f_i \\). The Mesh branch regresses all the human mesh parameters \\( (\\pmb{\\theta}_i, \\pmb{\\beta}_i) \\), in SMPL [32] format, from the input Feature maps. Unlike BEV, this branch takes both temporal image features and motion features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The combined features \\((F_{i}^{\\prime},O_{i})\\) are fed to our novel Tracking branch to estimate the 3D Motion Offset map, indicating the 3D position change of each subject across frames. The new Memory Unit takes the 3D detection and its 3D motion offset as input. It then determines the subject identities and builds human trajectories \\(\\hat{t}_i\\) of the \\(K\\) subjects in camera coordinates. Note that, like BEV, our detection branch finds all the people in the video frames but our goal is to track only the \\(K\\) input subjects. Consequently, the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.452 + ], + "angle": 0, + "content": "memory unit filters out detected people who do not match the subject trajectories." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.453, + 0.893, + 0.574 + ], + "angle": 0, + "content": "Finally, to estimate subject trajectories in global coordinates, the World branch estimates a world motion map, representing the 3D orientation \\(\\pmb{\\tau}_i\\) and 3D translation offset \\(\\Delta T_{i}\\) of the \\(K\\) subjects in global coordinates. Accumulating \\(\\Delta T_{i}\\), starting with the 3D position \\(\\hat{t}_{1}\\) of the tracked subjects in the first frame, gives their global 3D trajectory \\(\\pmb{T}\\). Note that the global (\"world\") coordinates are defined relative to the camera coordinates of the first frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.809, + 0.604 + ], + "angle": 0, + "content": "3.2. Holistic 5D Representation: Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Rather than directly estimating camera poses from environment keypoints [31] or inferring global human trajectories from local body poses [58], we develop a 5D representation to directly reason about human states, perform multiperson temporal association, and infer human trajectories in global coordinates. Learning a holistic 5D representation is the foundation of our one-stage framework. The representation has five main parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "i. Temporal feature maps. To construct the temporal feature maps encoding 5D human states and scenes information, we need to extract both single-frame image features and the motion features between adjacent frames. Therefore, given frame \\( f_{i-1} \\) and \\( f_i \\), we adopt a parallel two-branch structure to extract temporal image feature maps \\( \\pmb{F}_i' \\) and motion feature maps \\( O_i \\) for the current frame \\( f_i \\). First, in the image branch, we extract single-frame feature maps \\( F_{i-1} \\) and \\( F_i \\) with an image backbone (HRNet-32 [9]). To extract long-term and short-term motion features, we construct a temporal feature propagation module by combining" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.228 + ], + "angle": 0, + "content": "a ConvGRU [47] module, a Deformable convolution [65] module, and a residual connection. With these, we fuse the image feature maps to generate a temporal image feature map \\( \\pmb{F}_i' \\). See Sup. Mat. for details and experimental analysis. Additionally, in the motion branch, we estimate the optical flow map \\( O_i \\) between frames \\( f_{i-1} \\) and \\( f_i \\) with a motion backbone (RAFT [47]), to extract motion features of both people and scenes. From the combined temporal feature \\( (\\pmb{F}_i', O_i) \\), then we estimate five maps for the task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.23, + 0.473, + 0.412 + ], + "angle": 0, + "content": "ii. 3D detection maps. From the temporal image feature \\( \\pmb{F}_i' \\), we estimate a 3D Center map \\( M_C^{3D} \\in \\mathbb{R}^{1 \\times D \\times H \\times W} \\) for coarse human detection and a 3D Localization map \\( M_L^{3D} \\in \\mathbb{R}^{1 \\times D \\times H \\times W} \\) for fine localization. The two 3D detection maps are composited from the front-view maps and the 2D bird's-eye-view maps following BEV [45]. For \\( K \\) subjects, we first parse out the detected 3D center positions \\( \\widetilde{\\pmb{t}}_i \\) and their detection confidence \\( c_i \\) from the 3D Center map. Then we sample the 3D Localization map at \\( \\widetilde{\\pmb{t}}_i \\) to obtain the fine 3D localization offset vectors \\( \\Delta t_i \\). The predicted 3D translation of subjects to be tracked in camera coordinates is \\( \\pmb{t}_i = \\widetilde{\\pmb{t}}_i + \\Delta t_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.594 + ], + "angle": 0, + "content": "iii. 3D Motion Offset map. To track subjects between adjacent frames using \\( \\pmb{F}_i' \\) and \\( O_i \\), we estimate a 3D Motion Offset map \\( M_M^{3D} \\in \\mathbb{R}^{3 \\times D \\times H \\times W} \\) via a BEV-based 3D composition. We sample the \\( M_M^{3D} \\) at detected 3D center positions \\( \\widetilde{\\pmb{t}}_i \\) to obtain the 3D motion offset vectors \\( \\Delta m_i \\) which represent the 3D position changes of the subjects from frame \\( f_{i-1} \\) to frame \\( f_i \\) in camera coordinates. With \\( \\Delta m_i, t_i, c_i \\) as input, a memory unit (Sec. 3.3) predicts the tracked identities \\( D_i \\) and the optimized 3D position \\( \\hat{t}_i \\) in camera coordinates, which filters out low-confidence detections while keeping track of subjects through long-term occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.473, + 0.839 + ], + "angle": 0, + "content": "iv. Mesh Feature map. To regress the SMPL parameters of subjects, we first estimate a 2D Mesh Feature map \\( M_F^{2D} \\in \\mathbb{R}^{C \\times H \\times W} \\) from temporal features \\( (F_i', O_i) \\). With the tracked 3D positions \\( \\widehat{t}_i = (u_i, v_i, d_i) \\), we sample a mesh feature vector from \\( M_F^{2D} \\) at 2D positions \\( (u_i, v_i) \\). To differentiate the features of people at different depths, we map the \\( d_i \\) to a 128-dim encoding vector via an embedding layer and add this with the mesh feature vector to regress SMPL [32] parameters \\( (\\theta_i, \\beta_i) \\). The SMPL model maps the pose and shape, \\( \\theta_i \\) and \\( \\beta_i \\), to a 3D human body mesh \\( B \\in \\mathbb{R}^{6890 \\times 3} \\). With a sparse weight matrix \\( R \\in \\mathbb{R}^{Q \\times 6890} \\) that describes the mapping from \\( B \\) to \\( Q \\) body keypoints, we can obtain the 3D positions of body keypoints \\( J \\in \\mathbb{R}^{Q \\times 3} \\) via \\( RB \\). With the estimated 3D positions \\( \\widehat{t}_i \\) and a predefined camera projection matrix \\( P \\), the projected 2D keypoints are \\( J_{2D} = P(J + \\widehat{t}_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.473, + 0.903 + ], + "angle": 0, + "content": "v. World Motion map. \\( M_W^{2D} \\in \\mathbb{R}^{6 \\times H \\times W} \\) contains the 6D vectors that describe the 3D orientation and the relative translation of the subjects in global coordinates. We sample \\( M_W^{2D} \\) at the 2D positions \\( (u_i, v_i) \\in \\widehat{t}_i \\) of the tracked" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.26 + ], + "angle": 0, + "content": "subjects to obtain their 3D body orientation \\(\\pmb{\\tau}_{i}\\) and 3D body motion offset \\(\\Delta T_{i}\\). For the tracked subject \\(k\\), \\(\\Delta T_{i}^{k}\\) represents their 3D position change from frame \\(f_{i-1}\\) to frame \\(f_{i}\\) in global coordinates. We take the camera coordinate system of the first frame as the global coordinate system for the video. The global 3D trajectories of all tracked subjects are obtained by accumulating \\(\\Delta T_{i}\\) to their 3D position \\(\\hat{t}_{1}\\) in the first frame, \\(\\pmb{T} = \\{\\hat{t}_{1}, \\hat{t}_{1} + \\Delta T_{2}, \\hat{t}_{1} + \\Delta T_{2} + \\Delta T_{3}, \\dots\\}\\). Combining the 3D mesh, ID, and the global trajectories, the network reasons about 3D motions and trajectories of \\(K\\) tracked subjects in global coordinates." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.27, + 0.768, + 0.287 + ], + "angle": 0, + "content": "3.3. Tracking with a Memory Unit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.895, + 0.43 + ], + "angle": 0, + "content": "We construct the 3D trajectory of each subject by associating the 3D detections \\( \\pmb{t}_i \\) over time with a 3D motion offset \\( \\Delta m_i \\). To deal with long-term occlusions, we design the Memory Unit for persistent tracking, which will keep the memory for the full sequence. The memory unit stores the human states during inference and is not used for training. With predicted 3D positions \\( \\pmb{t}_i \\), detection confidence \\( c_i \\), and 3D motion offsets \\( \\Delta m_i \\) as inputs, the memory unit can track online. In each process, we have three stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.583 + ], + "angle": 0, + "content": "i. Initialization. First, we discard predicted 3D positions \\( \\pmb{t}_i = (\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i) \\) whose detection confidence is below a threshold \\( \\lambda_c \\). We observe that our input video is usually shot by tracking the subjects, therefore, we discard the detection whose \\( 1 / z_i \\) is below the scale threshold \\( \\lambda_s \\). To suppress duplicate detections, we find the detection pairs whose Euclidean distance is below a pre-defined threshold \\( \\lambda_d \\) and discard the detection with lower detection confidence. In the first frame, we use the 3D positions \\( \\pmb{t}_1 \\) and detection confidences \\( \\pmb{c}_i \\) of \\( K \\) subjects to initialize the \\( K \\) memory nodes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.896, + 0.825 + ], + "angle": 0, + "content": "ii. Memory node matching. The memory nodes store the 3D position \\( t_{i-1}^k \\) of subject \\( k \\) in previous frames. We match the memory node with the new filtered detections in the current frame. We select the predicted 3D translation \\( t_i \\) that best matches the 3D trajectory as the optimized 3D position \\( \\hat{t}_i \\). Specifically, with filtered 3D positions \\( t_i \\) and 3D motion offset \\( \\Delta m_i \\), we calculate the Euclidean distance matrix between \\( W t_{i-1}^k \\) and \\( W (t_i - \\Delta m_i) \\) where \\( W \\in \\mathbb{R}^3 \\) is a distance weight vector. To avoid the effects of the nonlinear relationship between depth \\( z_i \\) and human scale in images, we convert the depth value \\( z \\) to \\( 1/(1 + z) \\) when calculating the distance matrix. Using Hungarian matching, we keep the matched pairs whose matching distance is below a threshold \\( \\lambda_m \\) and use them for memory update. We use ground truth tracks for training and only perform matching during inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "iii. Memory update. We update the successfully matched memory nodes with the new 3D position and detection confidence. For the memory nodes without a matched detection, we accumulate the time since failure. Then we remove the memory nodes whose failure time is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8860" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "above a threshold \\(\\lambda_{f}\\). Tracking can be done in two modes: on-line or off-line. The former does not allow looking back in time, while the latter does. In the off-line mode, if a new detection re-activates a non-matched memory node, the non-matched part of the 3D trajectory is linearly interpolated. Finally, the memory unit outputs the latest 3D positions \\(\\widehat{t}_i\\) and tracking IDs \\(D_{i}\\) of all memory nodes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.208, + 0.254, + 0.224 + ], + "angle": 0, + "content": "3.4. DynaCam Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.337 + ], + "angle": 0, + "content": "Even with a powerful 5D representation, we still lack in-the-wild data for training and evaluation of global human trajectory estimation. However, collecting global human trajectory and camera poses for natural DC-videos is difficult. Therefore, we create a new dataset, DynaCam, by simulating camera motions to convert in-the-wild videos captured by static cameras to DC-videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.715 + ], + "angle": 0, + "content": "We use over 1000 video clips captured by static regular cameras from the MPII Human Pose Database [3] as well as videos from the InterNet [2]. We also use over 200 panoramic video clips that are either recorded by us with an Insta360 RS panoramic camera or are downloaded from the InterNet [1]. We manually design the 3D rotation and field of view (FOV) of dynamic cameras to track the subjects in panoramic videos. With the designed camera motions, we can project the panoramic frames into perspective views. Also, to simulate the 3D translation of dynamic cameras, we crop the videos captured by static cameras with sliding windows. In this way, we can obtain abundant in-the-wild DC-videos with accurate camera pose annotations. Then we perform 2D human detection, tracking, and 2D pose estimation viaYOLOX [17], ByteTrack [63], and ViT-Pose [54], respectively, to obtain 2D pose sequences of each subject. We estimate SMPL parameters by fitting the 2D poses using EFT [20] or ProHMR [25] and solve for their 3D positions in camera coordinates via the PnP algorithm (RANSAC [16]). Finally, we solve for the 3D human trajectories in the global coordinates with camera pose annotations. We manually filter out the failure cases. In this way, we generate more than 500 annotated DC-videos containing over 48K frames. More than half of video frames are generated from panoramic videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Limitations: The videos generated with our process only approximate real DC-videos shot in the wild since they lack perspective effects. Despite this they prove useful for training TRACE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.786, + 0.231, + 0.801 + ], + "angle": 0, + "content": "3.5. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.87 + ], + "angle": 0, + "content": "TRACE is supervised by the weighted sum of 15 loss terms that fall into two groups: temporal motion losses and standard image losses. Here we focus on the novel temporal losses. Please refer to the Sup. Mat. for details of all losses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To learn the temporal motion, we introduce a 3D motion offset loss \\(\\mathcal{L}_m\\) and a 6D world motion loss \\(\\mathcal{L}_W\\). \\(\\mathcal{L}_m\\) is the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "\\(L_{2}\\) loss between the predicted 3D motion offset \\(\\Delta m_{i}\\) and \\((\\pmb{t}_i^{\\prime} - \\pmb{t}_{i - 1}^{\\prime})\\) where \\(t_i^\\prime\\) is the ground truth 3D human position at frame \\(f_{i}\\) in our pre-defined camera coordinates (FOV=50°), which is solved for via the PnP algorithm [16]. \\(\\mathcal{L}_W\\) consists of six parts, including an \\(L_{2}\\) loss on the global 3D trajectory \\(\\pmb{T}\\), an \\(L_{2}\\) loss on the velocity/acceleration of 3D trajectory nodes \\(\\dot{T} / \\ddot{T}\\), an \\(L_{2}\\) loss of the velocity/acceleration of the 3D foot keypoints in global coordinates, and an \\(L_{2}\\) loss on the global 3D body orientation \\(\\pmb{\\tau}_i\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.243, + 0.634, + 0.26 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.719, + 0.284 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Training details. During training, we directly use the ground truth trajectory of subjects to replace the estimated trajectory \\(\\hat{t}\\) for sampling the parameters. We use the pretrained backbone of BEV [45] as the image backbone. We use RAFT [47] as the optical flow backbone. The training consists of two stages. In the first stage, we freeze the weights of the backbones and train the head network for 40 epochs with a learning rate of 5e-5. Then we train the image backbone and the head network together for 10 epochs with a learning rate of 1e-5. We use four V100-16GB GPUs for training. Limited by the GPU memory, we sample 4 video clips as a batch at each iteration; the clip length is 10 frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Training and evaluation datasets. For training, we use three 3D human pose datasets (Human3.6M [19], MPIINF-3DHP [34, 35], and 3DPW [50]), two 2D human pose datasets (PennAction [62] and PoseTrack [12]), and our DynaCam dataset. We evaluate TRACE on two multi-person in-the-wild benchmarks, 3DPW [50] and MuPoTS-3D [35], and DynaCam. 3DPW videos are most consistent with our tracking scenario. Unfortunately, not all 3DPW videos have complete tracking annotations. We select the 16 videos that do and call this subset Dyna3DPW. We use this challenging subset to evaluate tracking and HPS accuracy in complex scenes with a moving camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.894, + 0.806 + ], + "angle": 0, + "content": "Evaluation metrics. For global 3D trajectory estimation, we compute the average 3D position (Traj) and velocity (Velocity) error of the predicted global 3D trajectory \\( T \\). For multi-object tracking, we report the ID switch (IDs), Multi-Object Tracking Accuracy (MOTA [5]), Identification F1-score (IDF1 [41]), and Higher Order Tracking Accuracy (HOTA [33]). To assess the accuracy of 3D human pose/shape estimation, we compute the Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PMPJPE), and Mean Vertex Error (MVE)." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.806, + 0.797, + 0.821 + ], + "angle": 0, + "content": "Please refer to Sup. Mat. for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.854, + 0.848 + ], + "angle": 0, + "content": "4.2. Comparisons to State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Global 3D trajectory estimation. We aim to estimate the global human trajectory from dynamic cameras. We do not explicitly estimate the camera motion. Instead, we use a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8861" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.357, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.435, + 0.303, + 0.447 + ], + "angle": 0, + "content": "(a) Tracking on 3DPW" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.596, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.435, + 0.56, + 0.447 + ], + "angle": 0, + "content": "(b) Tracking on MuPoTS-3D" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.088, + 0.885, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.435, + 0.885, + 0.448 + ], + "angle": 0, + "content": "(c) Global 3D trajectory estimation on DynaCam" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.455, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons to SOTA methods (ByteTrack [63], PHALP [40], and GLAMR [58]) on 3DPW, MuPoTS-3D, and DynaCam. Red dash boxes highlight the ID switches. Yellow dot-dash boxes highlight the misses under occlusions." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.504, + 0.468, + 0.581 + ], + "angle": 0, + "content": "
MethodTranslating CameraRotating CameraFPS
Traj↓Velocity↓Traj↓Velocity↓
BEV+DPVO2.0240.1543.6220.1195.7
GLAMR [58]1.7450.1051.7590.0551.5
TRACE1.4330.0821.0380.04521.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.583, + 0.469, + 0.611 + ], + "angle": 0, + "content": "Table 1. Errors in estimated 3D human trajectory in world coordinates (in \\( m \\) ) and runtime efficiency on the DynaCam dataset." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.628, + 0.441, + 0.74 + ], + "angle": 0, + "content": "
MethodsIDs↓MOTA↑IDF1↑HOTA↑
Trackformer [36]4324.962.753.2
Tracktor [4]5351.570.950.3
FlowPose [51]4921.467.141.8
T3DP [39]3862.179.159.2
PHALP [40]2266.281.459.4
ByteTrack [63]1573.384.663.5
TRACE086.993.465.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.744, + 0.47, + 0.772 + ], + "angle": 0, + "content": "Table 2. Tracking results on MuPoTS-3D [35]. Results of the compared methods [4,36,39,40,51] are from [40]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "world motion map to represent the global trajectory in world coordinates, which implies the camera motion. Therefore, we evaluate the global trajectory error, instead of the camera pose, in Tab. 3. First, we evaluate global 3D trajectory estimation on DynaCam. We compare TRACE with two baseline solutions. The first one uses BEV to estimate the subjects in camera coordinates and DPVO [48], a SLAM" + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.504, + 0.872, + 0.591 + ], + "angle": 0, + "content": "
MethodsIDs↓MOTA↑IDF1↑HOTA↑
PHALP [40]597.993.774.2
YOLOX+ByteTrack397.398.273.1
BEV+ByteTrack3793.679.159.3
TRACE w/o MO495.897.372.7
TRACE199.399.774.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.596, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Table 3. Tracking results on Dyna3DPW [50]. w/o MO means ablating the estimated 3D motion offsets for tracking." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.631, + 0.843, + 0.761 + ], + "angle": 0, + "content": "
MethodPAMPJPE↓MPJPE↓PVE↓
HybrIK [27]45.074.186.5
METRO [30]47.977.188.2
GLAMR [58]50.7--
CLIFF [28]43.069.081.2
D&D [26]42.773.788.6
ROMP [44]47.376.793.4
BEV [45]46.978.592.3
TRACE37.879.197.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.763, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Table 4. Comparisons with SOTA methods for 3D human pose and shape estimation on the 3DPW test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.903 + ], + "angle": 0, + "content": "method, to estimate the camera and its motion; we call this BEV+DPVO. As shown in Tab. 1, TRACE significantly outperforms BEV+DPVO in the accuracy of global 3D trajectory estimation. The moving people in the scene make it hard for a SLAM method to extract stable corresponding keypoints. Additionally, our synthetic camera motions" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8862" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "differ from real camera motions and this may hurt DPVO's performance. A more direct comparison is to GLAMR [58]. TRACE outperforms GLAMR in both accuracy and efficiency. In Fig. 3(c), we also perform visual comparisons with GLAMR on DynaCam. These results demonstrate the benefit of estimating global human trajectory using a holistic 5D representation. We provide more results in the supplemental video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Multi-subject tracking. To evaluate the performance of tracking subjects with dynamic cameras in real-world scenes, we compare TRACE with recent methods on Dyna3DPW. PHALP [40] is a recent (SOTA) method that uses 3D cues and appearance to track people using SMPL. YOLOX+ByteTrack [17, 63] is a recently proposed and popular tracking-by-detection solution. These methods are designed to track all the people in a scene. Therefore, we process their results to avoid them being penalized for tracking unlabeled passers-by; 3DPW has annotations for at most 2 people in a scene but some scenes contain many people. We first obtain their tracking results using their official code. We then select the tracking results that achieve maximum IoU with the labeled ground truth subjects; we use these tracks for evaluation. Note that, for a fair comparison with ByteTrack, TRACE runs in an on-line mode, without optimizing the past results. As shown in Tab. 3, TRACE outperforms PHALP, the previous 3D-representation-based method [40], and tracking-by-detection [17, 63] methods on Dyna3DPW. To evaluate the tracking robustness under long-term occlusion, we evaluate TRACE on MuPoTS-3D. Results of previous SOTA methods [4, 36, 39, 40, 51] are from [40]. Again, for a fair comparison, we filter out the unlabeled people in the ByteTrack results. As shown in Tab. 2, TRACE significantly outperforms previous SOTA methods. In particular, TRACE significantly reduces ID switches under long-term occlusions. These results illustrate the effectiveness and robustness of the proposed method for in-the-wild videos. The qualitative comparisons are presented in Fig. 3 and the supplemental video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "3D human pose and shape estimation. Finally, we evaluate 3D human regression performance in DC-video using the 3DPW test set. Because 3DPW does not provide ground-truth 3D translation in world coordinates, we evaluate root-relative 3D pose. We compare TRACE with the recent one-stage [44,45] and multi-stage [26-28,30,58] methods. Tab. 4 shows that TRACE significantly outperforms the previous method by \\(11.4\\%\\) in terms of PMPJPE. These results demonstrate that learning a holistic 5D representation in an end-to-end manner is powerful. We adopt a fixed field of view \\((\\mathrm{FOV} = 50^{\\circ})\\) to define our camera space, which is very different from that of 3DPW video. This results in a misalignment between the predicted 3D orientation of the body mesh and the ground truth, resulting in higher MPJPE and MVE. This illustrates the importance of also estimating" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the camera so that our estimated global coordinate system is consistent with the real world. This is future work." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.134, + 0.665, + 0.149 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.247 + ], + "angle": 0, + "content": "Temporal 5D representation v.s. image-level 3D representation. We go beyond BEV's image-level 3D representation and build a temporal 5D representation. As shown in Tab. 1, 3, and 4, TRACE outperforms BEV or the multistage solutions using BEV on most metrics. This demonstrates the value of learning a holistic 5D representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.249, + 0.894, + 0.309 + ], + "angle": 0, + "content": "3D Motion Offset map. We also evaluate the effect of using predicted 3D motion offsets for tracking. As shown in Tab. 3, 3D motion offsets improve performance by \\(3.5\\%\\), \\(2.4\\%\\), and \\(2.7\\%\\) in terms of MOTA, IDF1, and HOTA." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.329, + 0.627, + 0.344 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Human pose and shape estimation is not an end to itself. Rather, estimating the 3D human in motion is useful for many tasks from behavior analysis to computer graphics. However, to be useful, it is important to know the motion of humans with respect to the 3D scene and other people. This means that HPS methods must estimate humans in a global coordinate system and provide consistent tracks of people across time. For generality, they also need to be able to do this from arbitrary moving cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.701 + ], + "angle": 0, + "content": "To tackle these challenging problems, we propose a novel 5D representation and a new neural architecture that reasons about people in 5D; that is, their 3D position, temporal trajectory, and identity. Moving to a 5D representation enables our method, TRACE, to take a holistic view of the video, processing full frames and incorporating temporal features. The core innovation of TRACE lies in its novel temporal representation in the form of new \"maps\" that represent the motion of people across time in the camera and global coordinates. These allow TRACE to be trained end-to-end, thus exploiting rich information from the video to solve the task. TRACE is the first such single-shot method for 3D HPS estimation from video and it achieves SOTA results on common benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Future work should look at explicitly estimating the camera, using training data like BEDLAM [6], which contains complex human motion, 3D scenes, and camera motions. We believe that camera motion and human motion provide complimentary information that can be used to recover human motion in world coordinates with metric accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Acknowledgements: This work was partially supported by the National Key R&D Program of China under Grant (No. 2020AAA0103800) and Beijing Nova Program (No. 20220484063)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.889, + 0.901 + ], + "angle": 0, + "content": "MJB Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2023.txt" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8863" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.403, + 0.129 + ], + "angle": 0, + "content": "[1] Bilibili. https://www.bilibili.com/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.373, + 0.145 + ], + "angle": 0, + "content": "[2] Pexels. https://www.pexels.com.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.469, + 0.201 + ], + "angle": 0, + "content": "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, pages 941-951, 2019. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[5] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, pages 1-10, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of 3d human bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, pages 561-578, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, pages 5386-5395, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[10] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, pages 436-454, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[11] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[12] Andreas Doering, Di Chen, Shanshan Zhang, Bernt Schiele, and Juergen Gall. Posetrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In CVPR, pages 20963-20972, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[13] Mahsa Ehsanpour, Fatemeh Saleh, Silvio Savarese, Ian Reid, and Hamid Rezatofighi. Jrdb-act: A large-scale dataset for spatio-temporal action, social group and activity detection. In CVPR, pages 20983–20992, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[14] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. A mobile vision system for robust multi-person tracking. In CVPR, pages 1-8, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoşa Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, pages 10849-10859, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[19] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. TPAMI, 36(7):1325-1339, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In ECCV, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. VIBE: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[23] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In ICCV, pages 11035-11045, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[24] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[25] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In CVPR, pages 11605-11614, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[26] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&d: Learning human dynamics from dynamic camera. In ECCV, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[27] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In CVPR, pages 3383-3393, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[28] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, pages 590-606, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[29] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[30] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, pages 930-939, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):1-16, 2015. 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8864" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[33] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, pages 548-578, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[34] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3D human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[35] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[36] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.397 + ], + "angle": 0, + "content": "[37] Gyeongsik Moon and Kyoung Mu Lee. Pose2Pose: 3d positional pose-guided 3d rotational pose prediction for expressive 3d human pose and mesh estimation. arXiv, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[38] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. TexturePose: Supervising human mesh estimation with texture consistency. In ICCV, pages 803-812, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[39] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. In NeurIPS, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.537 + ], + "angle": 0, + "content": "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3D appearance, location & pose. In CVPR, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[41] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.637 + ], + "angle": 0, + "content": "[42] Weijian Ruan, Wu Liu, Qian Bao, Jun Chen, Yuhao Cheng, and Tao Mei. Pointet: pose-guided ovonic insight network for multi-person pose tracking. In MM, pages 284-292, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[43] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[45] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[46] Yu Sun, Yun Ye, Wu Liu, Wenpeng Gao, YiLi Fu, and Tao Mei. Human mesh recovery from monocular images via a skeleton-disentangled representation. In ICCV, pages 5348-5357, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[47] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[48] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[49] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[50] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[51] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[53] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[54] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. arXiv preprint arXiv:2204.12484, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[55] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[56] Hongwei Yi, Chun-Hao P. Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J. Black. MIME: Human-aware 3D scene generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[57] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[58] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In CVPR, 2022. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[59] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3D pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[60] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[61] Shanshan Zhang, Rodrigo Benenson, and Bernt Schiele. Citypersons: A diverse dataset for pedestrian detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[62] Weiyu Zhang, Menglong Zhu, and Konstantinos G Derpanis. From actemes to action: A strongly-supervised representation for detailed action understanding. In ICCV, pages 2248-2255, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[63] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[64] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In ICCV, pages 6239-6249, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.15 + ], + "angle": 0, + "content": "[65] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8866" + } + ] +] \ No newline at end of file diff --git a/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_origin.pdf b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1af1a39b2d8857dc0d42c6d0d0c9463b00da08b0 --- /dev/null +++ b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/c8e8684e-51e8-46fa-8645-e310b1ba9073_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b31e7e68363877ae42d98ed55eba6350f786a7a74144ac66512da8b0a6a0527b +size 8665140 diff --git a/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/full.md b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb265787880059b38ef3c7235aa32831cee6a1a9 --- /dev/null +++ b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/full.md @@ -0,0 +1,283 @@ +# TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments + +Yu Sun $^{1*}$ Qian Bao $^{2\dagger}$ Wu Liu $^{2\dagger}$ Tao Mei $^{4}$ Michael J. Black $^{3}$ + +$^{1}$ Harbin Institute of Technology $^{2}$ Explore Academy of JD.com + +3Max Planck Institute for Intelligent Systems HiDream.ai Inc. + +yusunhit@gmail.com, baoqian@jd.com, liuwul@jd.com + +tmei@hidream.ai, black@tuebingen.mpg.de + +![](images/a5971fff5bed9d6d158c0495cedd0412794397823f881d9964ea67bfed7c7296.jpg) + +# Dynamic Monocular Camera + +![](images/c61c8e122624cf0fc96aff32935004c837ea9533f8e03179629d700820206742.jpg) + +![](images/cf1fcb8fea368193fde60da9100333312b09de451a144b2314415a59fc11bada.jpg) + +![](images/68ded7e47aa57dc68b9545efe0ce2abc58b36cdcd89b9f8b492ac993b61a9a8a.jpg) + +![](images/e2f7565c67777ab738977981f4af699e6e56a4165a08a5cc280c9a875bfca665.jpg) + +![](images/9da159a468c0f4c2a27c14ebccb8b692b1b3cf9c4967a0af906e9fe40fa891b9.jpg) + +![](images/2ef673ece98846b17409ba65c428ceecd65d2001e3bb8e744dd8bf3417f38a34.jpg) +(a) Videos Captured by Dynamic Monocular Camera and Human 3D Recovery in Each Frame +(b) The 3D humans recovered by BEV [45] in camera coordinates +Figure 1. 5D temporal reconstruction of multiple 3D people in global coordinates with a dynamic camera. We introduce TRACE, a monocular one-stage method with a holistic 5D representation that enables the network to explicitly reason about human motion across time. Unlike previous methods [44,45] (b) that regress all 3D people from each frame in camera coordinates, TRACE (c) tracks the subjects (a) presented in the first frame through time and recovers their global trajectories in global coordinates in one shot. + +![](images/bd6f2064535a36ec77850e054c10794e7d189078265c27a80c31296b43e33d24.jpg) +(c) The 5D humans recovered by TRACE in global coordinates + +# Abstract + +Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel ar + +chitectural components. Most importantly, it uses two new "maps" to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code1 and dataset2 are released for research purposes. + +# 1. Introduction + +The estimation of 3D human pose and shape (HPS) has many applications and there has been significant recent progress [6, 7, 21-24, 37, 38, 46, 49, 56, 57, 60]. Most methods, however, reason only about a single frame at a time and estimate humans in camera coordinates. Moreover, such methods do not track people and are unable to recover their global trajectories. The problem is even harder in typical hand-held videos, which are filmed with a dynamic, moving, camera. For many applications of HPS, single-frame estimates in camera coordinates are not sufficient. To capture human movement and then transfer it to a new 3D scene, we must have the movement in a coherent global coordinate system. This is a requirement for computer graphics, sports, video games, and extended reality (XR). + +Our key insight is that most methods estimate humans in 3D, whereas the true problem is 5D. That is, a method needs to reason about 3D space, time, and subject identity. With a 5D representation, the problem becomes tractable, enabling a holistic solution that can exploit the full video to infer multiple people in a coherent global coordinate frame. As illustrated in Fig. 1, we develop a unified method to jointly regress the 3D pose, shape, identity, and global trajectory of the subjects in global coordinates from monocular videos captured by dynamic cameras (DC-videos). + +To achieve this, we deal with two main challenges. First, DC-videos contain both human motion and camera motion and these must be disentangled to recover the human trajectory in global coordinates. One idea would be to recover the camera motion relative to the rigid scene using structure-from-motion (SfM) methods (e.g. [31]). In scenes containing many people and human motion, however, such methods can be unreliable. An alternative approach is taken by GLAMR [58], which infers global human trajectories from local 3D human poses, without taking into account the full scene. By ignoring evidence from the full image, GLAMR fails to capture the correct global motion in common scenarios, such as biking, skating, boating, running on a treadmill, etc. Moreover, GLAMR is a multi-stage method, with each stage dependent on accurate estimates from the preceding one. Such approaches are more brittle than our holistic, end-to-end, method. + +The other challenge, as shown in the upper right corner of Fig. 1, is that severe occlusions are common in videos with multiple people. Currently, the most popular tracking strategy is to infer the association between 2D detections using a temporal prior (e.g. Kalman filter) [63]. However, in DC-videos, human motions are often irregular and can easily violate hand-crafted priors. PHALP [40] is one of the few methods to address this for 3D HPS. It uses a classical, multi-stage, detection-and-tracking formulation with heuristic temporal priors. It does not holistically reason about the sequence and is not trained end-to-end. + +To address these issues, we reason about people using a 5D representation and capture information from the full image and the motion of the scene. This holistic reasoning enables the reliable recovery of global human trajectories and subject tracking using a single-shot method. This is more reliable than multi-stage methods because the network can exploit more information to solve the task and is trained end-to-end. No hand-crafted priors are needed and the network is able to share information among modules. + +Specifically, we develop TRACE, a unified one-stage method for Temporal Regression of Avatars with dynamic Cameras in 3D Environments. The architecture is inspired by BEV [45], which directly estimates multiple people in depth from a single image using multiple 2D maps. BEV uses a 2D map representing an imaginary, "top down", view of the scene. This is combined with an image-centric 2D map to reason about people in 3D. Our key insight is that the idea of maps can be extended to represent how people move in 3D. With this idea, TRACE introduces three new modules to holistically model 5D human states, performing multi-person temporal association, and inferring human trajectories in global coordinates; see Fig. 2. + +First, to construct a holistic 5D representation of the video, we extract temporal image features by fusing single-frame feature maps from the image backbone with a temporal feature propagation module. We also compute the optical flow between adjacent frames with a motion backbone. The optical flow provides short-term motion features that carry information about the motion of the scene and the people. Second, to explicitly track human motions, we introduce a novel 3D motion offset map to establish the association of the same person across adjacent frames. This map contains a 3D offset vector at each position, which represents the difference between the 3D positions of the same subject from the previous frame to the current frame in camera coordinates. We also introduce a memory unit to keep track of subjects under long-term occlusion. Note that the 3D trajectories are built in camera space, and TRACE uses a novel world motion map that transfers the trajectories to global coordinates. At each position, this map contains a 6D vector to represent the difference between the 3D positions of the corresponding subject from the previous frame to the current frame and its 3D orientation in world coordinates. Taken together, this novel network architecture goes beyond prior work by taking information from the full video frames to address detection, pose estimation, tracking, and occlusion in a holistic network that is trained end-to-end. + +To enable training and evaluation of global human trajectory estimation from in-the-wild DC-videos, we build a new dataset, DynaCam. Since collecting global human trajectories and camera poses with in-the-wild DC-videos is difficult, we simulate a moving camera using publicly available in-the-wild panoramic videos and regular videos cap + +tured by static cameras. In this way, we create more than 500 in-the-wild DC-videos with precise camera pose annotations. Then we generate pseudo-ground-truth 3D human annotations via fitting [20] SMPL [32] to detected 2D pose sequences [17,54,63]. With 2D/3D human pose and camera pose annotations, we can obtain the global human trajectories using the PnP algorithm [16]. This dataset is sufficient to train TRACE to deal with dynamic cameras. + +We evaluate TRACE on two multi-person in-the-wild benchmarks (3DPW [50] and MuPoTS-3D [35]) and our DynaCam dataset. On 3DPW, TRACE achieves the state-of-the-art (SOTA) PA-MPJPE of $37.8\mathrm{mm}$ , less the current best (42.7 [26]). On MuPoTS-3D, TRACE outperforms previous 3D-representation-based methods [39, 40] and tracking-by-detection methods [63] on tracking people under long-term occlusion. On DynaCam, TRACE outperforms GLAMR [58] in estimating the 3D human trajectory in global coordinates from DC-videos. + +In summary, our main contributions are: (1) We introduce a 5D representation and use it to learn holistic temporal cues related to both 3D human motions and the scene. (2) We introduce two novel motion offset representations to explicitly model temporal multi-subject association and global human trajectories from temporal clues in an end-to-end manner. (3) We estimate long-term 3D human motions over time in global coordinates, achieving SOTA results. (4) We collect the DynaCam dataset of DC-videos with pseudo ground truth, which facilitates the training and evaluation of global human trajectory estimation. The code and dataset are publicly available for research purposes. + +# 2. Related Work + +Monocular 3D mesh regression with full images. Most existing methods [21-24, 26, 29, 37, 38, 46, 52, 53, 59, 60] take a multi-stage detection-based pipeline to estimate 3D HPS from cropped image patches, which exclude important cues, such as camera information and human-scene relationships. A few recent multi-stage [28] and one-stage [44, 45] methods have made steps towards using the full-image information. For instance, CIUFF [28] estimates 3D HPS by taking into account the bounding box locations, giving the method camera information and improving accuracy. To directly estimates multiple people at once from the full image, ROMP [44] introduces a 2D Center heatmap and a Mesh parameter map to represent 2D human locations and 3D human body meshes, respectively. BEV [45] goes beyond ROMP by introducing an imaginary bird's-eye-view map, which is combined with the front-view maps to construct a 3D view in camera coordinates. However, they only model 3D HPS in camera coordinates from a single image. Using "maps" like ROMP and BEV, TRACE also looks at the full image. We go further, however, by introducing novel maps that model human motions across a video se + +quence in global coordinates. + +Tracking datasets. While there are many tracking datasets [8, 10, 11, 14, 18, 61] with 2D annotations, only a few [13, 15] capture the 3D trajectory of pedestrians. In both cases, the scene and human activities are limited. To address this, we use 3DPW [50] and MuPoTS-3D [35] for tracking evaluation. 3DPW is the most relevant dataset for our task and it provides a real-world test case. 3DPW contains videos that are captured by a moving camera that follows subjects to record their activities in many daily scenes. MuPoTS-3D contains rich multi-person interaction scenes with long-term occlusions for tracking evaluation. + +Tracking 3D people through occlusions. Most existing methods [4, 42, 51, 63] perform tracking using 2D image cues. The classic tracking-by-detection paradigm focuses on associating the 2D detections using a temporal prior (e.g. Kalman filter). When applied to DC-videos containing rapid human and camera motions that violate the hand-crafted priors, such methods are brittle. Going beyond 2D, PHALP [40] separately extracts 3D human pose, appearance, and location with a multi-stage design from each video frame, and then assembles them for tracking. In contrast to these multi-stage methods, which are susceptible to errors in early stages, we explicitly learn the 3D human trajectory from temporal 5D cues in an end-to-end manner. + +Monocular global 3D human trajectory reasoning. Most existing methods [64] that reason about global 3D human trajectories do so with static, calibrated, cameras in a multi-view setting. A few recent methods [31,58] have addressed the ill-posed problem of extracting the global motions of humans from monocular video. Liu et al. [31] employ a structure-from-motion (SfM) method [43] to estimate the camera poses from monocular videos captured by a dynamic camera. However, when the input video contains the movement of multiple subjects, it is hard for SfM methods to extract sufficiently many stable keypoints for reliable camera estimation. GLAMR [58] adopts a multi-stage pipeline to infer the global human trajectory from root-relative local human 3D poses estimated from each frame. The per-frame human pose estimates make it vulnerable to occlusion. Additionally, GLAMR relies on bounding boxes, ignoring scene-related information. Consequently, GLAMR fails in common scenarios like riding a bike or skating. In concurrent work (in this proceedings), SLAHMR [55] uses a multi-stage optimization-based approach that combines structure from motion with human motion priors to estimate 4D human trajectories in global coordinates; this is very computationally expensive. In contrast to previous multi-stage methods, TRACE simultaneously combines scene information and 3D human motions with a novel 5D representation to holistically exploit all temporal cues and to enable end-to-end training. + +![](images/aecf35d1f4ee96994fccd5c5d0b73fad77fa335a7ebeebdea5aa07a6a82b98a8.jpg) +Figure 2. TRACE Overview. TRACE takes a video sequence and regions in the first frame corresponding to the subjects to be tracked. TRACE encodes the video and its motion with temporal features and optical flow. A novel 3D Motion Offset map reasons about human trajectories in camera coordinates. The World Motion map represents the trajectory in global coordinates. A memory unit deals with occlusions by encoding the subject identities. TRACE is trained end-to-end to estimate the 3D shape and pose of multiple people throughout a video in global coordinates. See Sec. 3.1 for details. + +# 3. Method + +# 3.1. Overview + +The overall framework of TRACE is shown in Fig. 2. Given a video sequence captured with a dynamic camera $\{f_i, i = 1, \dots, N\}$ with $N$ frames, the user specifies $K$ tracking subjects shown in the first frame. Our goal is to simultaneously recover the 3D pose, shape, identity, and trajectory of each subject in global coordinates. To achieve this, TRACE first extracts temporal features and then decodes each sub-task with a separate head network. First, via two parallel backbones, TRACE encodes the video and its motion into temporal image feature maps $F_i'$ and motion feature maps $O_i$ . + +The Detection and Tracking branches take these features and perform multi-subject tracking in camera coordinates. Unlike BEV [45], our detection method takes temporal image features $\pmb{F}_i'$ as input. It uses the features to detect the 3D human positions $\pmb{t}_i$ and their confidence $c_i$ for all people in frame $f_i$ . The Mesh branch regresses all the human mesh parameters $(\pmb{\theta}_i, \pmb{\beta}_i)$ , in SMPL [32] format, from the input Feature maps. Unlike BEV, this branch takes both temporal image features and motion features. + +The combined features $(F_{i}^{\prime},O_{i})$ are fed to our novel Tracking branch to estimate the 3D Motion Offset map, indicating the 3D position change of each subject across frames. The new Memory Unit takes the 3D detection and its 3D motion offset as input. It then determines the subject identities and builds human trajectories $\hat{t}_i$ of the $K$ subjects in camera coordinates. Note that, like BEV, our detection branch finds all the people in the video frames but our goal is to track only the $K$ input subjects. Consequently, the + +memory unit filters out detected people who do not match the subject trajectories. + +Finally, to estimate subject trajectories in global coordinates, the World branch estimates a world motion map, representing the 3D orientation $\pmb{\tau}_i$ and 3D translation offset $\Delta T_{i}$ of the $K$ subjects in global coordinates. Accumulating $\Delta T_{i}$ , starting with the 3D position $\hat{t}_{1}$ of the tracked subjects in the first frame, gives their global 3D trajectory $\pmb{T}$ . Note that the global ("world") coordinates are defined relative to the camera coordinates of the first frame. + +# 3.2. Holistic 5D Representation: Details + +Rather than directly estimating camera poses from environment keypoints [31] or inferring global human trajectories from local body poses [58], we develop a 5D representation to directly reason about human states, perform multiperson temporal association, and infer human trajectories in global coordinates. Learning a holistic 5D representation is the foundation of our one-stage framework. The representation has five main parts. + +i. Temporal feature maps. To construct the temporal feature maps encoding 5D human states and scenes information, we need to extract both single-frame image features and the motion features between adjacent frames. Therefore, given frame $f_{i-1}$ and $f_i$ , we adopt a parallel two-branch structure to extract temporal image feature maps $\pmb{F}_i'$ and motion feature maps $O_i$ for the current frame $f_i$ . First, in the image branch, we extract single-frame feature maps $F_{i-1}$ and $F_i$ with an image backbone (HRNet-32 [9]). To extract long-term and short-term motion features, we construct a temporal feature propagation module by combining + +a ConvGRU [47] module, a Deformable convolution [65] module, and a residual connection. With these, we fuse the image feature maps to generate a temporal image feature map $\pmb{F}_i'$ . See Sup. Mat. for details and experimental analysis. Additionally, in the motion branch, we estimate the optical flow map $O_i$ between frames $f_{i-1}$ and $f_i$ with a motion backbone (RAFT [47]), to extract motion features of both people and scenes. From the combined temporal feature $(\pmb{F}_i', O_i)$ , then we estimate five maps for the task. + +ii. 3D detection maps. From the temporal image feature $\pmb{F}_i'$ , we estimate a 3D Center map $M_C^{3D} \in \mathbb{R}^{1 \times D \times H \times W}$ for coarse human detection and a 3D Localization map $M_L^{3D} \in \mathbb{R}^{1 \times D \times H \times W}$ for fine localization. The two 3D detection maps are composited from the front-view maps and the 2D bird's-eye-view maps following BEV [45]. For $K$ subjects, we first parse out the detected 3D center positions $\widetilde{\pmb{t}}_i$ and their detection confidence $c_i$ from the 3D Center map. Then we sample the 3D Localization map at $\widetilde{\pmb{t}}_i$ to obtain the fine 3D localization offset vectors $\Delta t_i$ . The predicted 3D translation of subjects to be tracked in camera coordinates is $\pmb{t}_i = \widetilde{\pmb{t}}_i + \Delta t_i$ . + +iii. 3D Motion Offset map. To track subjects between adjacent frames using $\pmb{F}_i'$ and $O_i$ , we estimate a 3D Motion Offset map $M_M^{3D} \in \mathbb{R}^{3 \times D \times H \times W}$ via a BEV-based 3D composition. We sample the $M_M^{3D}$ at detected 3D center positions $\widetilde{\pmb{t}}_i$ to obtain the 3D motion offset vectors $\Delta m_i$ which represent the 3D position changes of the subjects from frame $f_{i-1}$ to frame $f_i$ in camera coordinates. With $\Delta m_i, t_i, c_i$ as input, a memory unit (Sec. 3.3) predicts the tracked identities $D_i$ and the optimized 3D position $\hat{t}_i$ in camera coordinates, which filters out low-confidence detections while keeping track of subjects through long-term occlusion. + +iv. Mesh Feature map. To regress the SMPL parameters of subjects, we first estimate a 2D Mesh Feature map $M_F^{2D} \in \mathbb{R}^{C \times H \times W}$ from temporal features $(F_i', O_i)$ . With the tracked 3D positions $\widehat{t}_i = (u_i, v_i, d_i)$ , we sample a mesh feature vector from $M_F^{2D}$ at 2D positions $(u_i, v_i)$ . To differentiate the features of people at different depths, we map the $d_i$ to a 128-dim encoding vector via an embedding layer and add this with the mesh feature vector to regress SMPL [32] parameters $(\theta_i, \beta_i)$ . The SMPL model maps the pose and shape, $\theta_i$ and $\beta_i$ , to a 3D human body mesh $B \in \mathbb{R}^{6890 \times 3}$ . With a sparse weight matrix $R \in \mathbb{R}^{Q \times 6890}$ that describes the mapping from $B$ to $Q$ body keypoints, we can obtain the 3D positions of body keypoints $J \in \mathbb{R}^{Q \times 3}$ via $RB$ . With the estimated 3D positions $\widehat{t}_i$ and a predefined camera projection matrix $P$ , the projected 2D keypoints are $J_{2D} = P(J + \widehat{t}_i)$ . + +v. World Motion map. $M_W^{2D} \in \mathbb{R}^{6 \times H \times W}$ contains the 6D vectors that describe the 3D orientation and the relative translation of the subjects in global coordinates. We sample $M_W^{2D}$ at the 2D positions $(u_i, v_i) \in \widehat{t}_i$ of the tracked + +subjects to obtain their 3D body orientation $\pmb{\tau}_{i}$ and 3D body motion offset $\Delta T_{i}$ . For the tracked subject $k$ , $\Delta T_{i}^{k}$ represents their 3D position change from frame $f_{i-1}$ to frame $f_{i}$ in global coordinates. We take the camera coordinate system of the first frame as the global coordinate system for the video. The global 3D trajectories of all tracked subjects are obtained by accumulating $\Delta T_{i}$ to their 3D position $\hat{t}_{1}$ in the first frame, $\pmb{T} = \{\hat{t}_{1}, \hat{t}_{1} + \Delta T_{2}, \hat{t}_{1} + \Delta T_{2} + \Delta T_{3}, \dots\}$ . Combining the 3D mesh, ID, and the global trajectories, the network reasons about 3D motions and trajectories of $K$ tracked subjects in global coordinates. + +# 3.3. Tracking with a Memory Unit + +We construct the 3D trajectory of each subject by associating the 3D detections $\pmb{t}_i$ over time with a 3D motion offset $\Delta m_i$ . To deal with long-term occlusions, we design the Memory Unit for persistent tracking, which will keep the memory for the full sequence. The memory unit stores the human states during inference and is not used for training. With predicted 3D positions $\pmb{t}_i$ , detection confidence $c_i$ , and 3D motion offsets $\Delta m_i$ as inputs, the memory unit can track online. In each process, we have three stages. + +i. Initialization. First, we discard predicted 3D positions $\pmb{t}_i = (\pmb{x}_i, \pmb{y}_i, \pmb{z}_i)$ whose detection confidence is below a threshold $\lambda_c$ . We observe that our input video is usually shot by tracking the subjects, therefore, we discard the detection whose $1 / z_i$ is below the scale threshold $\lambda_s$ . To suppress duplicate detections, we find the detection pairs whose Euclidean distance is below a pre-defined threshold $\lambda_d$ and discard the detection with lower detection confidence. In the first frame, we use the 3D positions $\pmb{t}_1$ and detection confidences $\pmb{c}_i$ of $K$ subjects to initialize the $K$ memory nodes. + +ii. Memory node matching. The memory nodes store the 3D position $t_{i-1}^k$ of subject $k$ in previous frames. We match the memory node with the new filtered detections in the current frame. We select the predicted 3D translation $t_i$ that best matches the 3D trajectory as the optimized 3D position $\hat{t}_i$ . Specifically, with filtered 3D positions $t_i$ and 3D motion offset $\Delta m_i$ , we calculate the Euclidean distance matrix between $W t_{i-1}^k$ and $W (t_i - \Delta m_i)$ where $W \in \mathbb{R}^3$ is a distance weight vector. To avoid the effects of the nonlinear relationship between depth $z_i$ and human scale in images, we convert the depth value $z$ to $1/(1 + z)$ when calculating the distance matrix. Using Hungarian matching, we keep the matched pairs whose matching distance is below a threshold $\lambda_m$ and use them for memory update. We use ground truth tracks for training and only perform matching during inference. + +iii. Memory update. We update the successfully matched memory nodes with the new 3D position and detection confidence. For the memory nodes without a matched detection, we accumulate the time since failure. Then we remove the memory nodes whose failure time is + +above a threshold $\lambda_{f}$ . Tracking can be done in two modes: on-line or off-line. The former does not allow looking back in time, while the latter does. In the off-line mode, if a new detection re-activates a non-matched memory node, the non-matched part of the 3D trajectory is linearly interpolated. Finally, the memory unit outputs the latest 3D positions $\widehat{t}_i$ and tracking IDs $D_{i}$ of all memory nodes. + +# 3.4. DynaCam Dataset + +Even with a powerful 5D representation, we still lack in-the-wild data for training and evaluation of global human trajectory estimation. However, collecting global human trajectory and camera poses for natural DC-videos is difficult. Therefore, we create a new dataset, DynaCam, by simulating camera motions to convert in-the-wild videos captured by static cameras to DC-videos. + +We use over 1000 video clips captured by static regular cameras from the MPII Human Pose Database [3] as well as videos from the InterNet [2]. We also use over 200 panoramic video clips that are either recorded by us with an Insta360 RS panoramic camera or are downloaded from the InterNet [1]. We manually design the 3D rotation and field of view (FOV) of dynamic cameras to track the subjects in panoramic videos. With the designed camera motions, we can project the panoramic frames into perspective views. Also, to simulate the 3D translation of dynamic cameras, we crop the videos captured by static cameras with sliding windows. In this way, we can obtain abundant in-the-wild DC-videos with accurate camera pose annotations. Then we perform 2D human detection, tracking, and 2D pose estimation viaYOLOX [17], ByteTrack [63], and ViT-Pose [54], respectively, to obtain 2D pose sequences of each subject. We estimate SMPL parameters by fitting the 2D poses using EFT [20] or ProHMR [25] and solve for their 3D positions in camera coordinates via the PnP algorithm (RANSAC [16]). Finally, we solve for the 3D human trajectories in the global coordinates with camera pose annotations. We manually filter out the failure cases. In this way, we generate more than 500 annotated DC-videos containing over 48K frames. More than half of video frames are generated from panoramic videos. + +Limitations: The videos generated with our process only approximate real DC-videos shot in the wild since they lack perspective effects. Despite this they prove useful for training TRACE. + +# 3.5. Loss Functions + +TRACE is supervised by the weighted sum of 15 loss terms that fall into two groups: temporal motion losses and standard image losses. Here we focus on the novel temporal losses. Please refer to the Sup. Mat. for details of all losses. + +To learn the temporal motion, we introduce a 3D motion offset loss $\mathcal{L}_m$ and a 6D world motion loss $\mathcal{L}_W$ . $\mathcal{L}_m$ is the + +$L_{2}$ loss between the predicted 3D motion offset $\Delta m_{i}$ and $(\pmb{t}_i^{\prime} - \pmb{t}_{i - 1}^{\prime})$ where $t_i^\prime$ is the ground truth 3D human position at frame $f_{i}$ in our pre-defined camera coordinates (FOV=50°), which is solved for via the PnP algorithm [16]. $\mathcal{L}_W$ consists of six parts, including an $L_{2}$ loss on the global 3D trajectory $\pmb{T}$ , an $L_{2}$ loss on the velocity/acceleration of 3D trajectory nodes $\dot{T} / \ddot{T}$ , an $L_{2}$ loss of the velocity/acceleration of the 3D foot keypoints in global coordinates, and an $L_{2}$ loss on the global 3D body orientation $\pmb{\tau}_i$ . + +# 4. Experiments + +# 4.1. Implementation Details + +Training details. During training, we directly use the ground truth trajectory of subjects to replace the estimated trajectory $\hat{t}$ for sampling the parameters. We use the pretrained backbone of BEV [45] as the image backbone. We use RAFT [47] as the optical flow backbone. The training consists of two stages. In the first stage, we freeze the weights of the backbones and train the head network for 40 epochs with a learning rate of 5e-5. Then we train the image backbone and the head network together for 10 epochs with a learning rate of 1e-5. We use four V100-16GB GPUs for training. Limited by the GPU memory, we sample 4 video clips as a batch at each iteration; the clip length is 10 frames. + +Training and evaluation datasets. For training, we use three 3D human pose datasets (Human3.6M [19], MPIINF-3DHP [34, 35], and 3DPW [50]), two 2D human pose datasets (PennAction [62] and PoseTrack [12]), and our DynaCam dataset. We evaluate TRACE on two multi-person in-the-wild benchmarks, 3DPW [50] and MuPoTS-3D [35], and DynaCam. 3DPW videos are most consistent with our tracking scenario. Unfortunately, not all 3DPW videos have complete tracking annotations. We select the 16 videos that do and call this subset Dyna3DPW. We use this challenging subset to evaluate tracking and HPS accuracy in complex scenes with a moving camera. + +Evaluation metrics. For global 3D trajectory estimation, we compute the average 3D position (Traj) and velocity (Velocity) error of the predicted global 3D trajectory $T$ . For multi-object tracking, we report the ID switch (IDs), Multi-Object Tracking Accuracy (MOTA [5]), Identification F1-score (IDF1 [41]), and Higher Order Tracking Accuracy (HOTA [33]). To assess the accuracy of 3D human pose/shape estimation, we compute the Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PMPJPE), and Mean Vertex Error (MVE). + +Please refer to Sup. Mat. for more details. + +# 4.2. Comparisons to State-of-the-art Methods + +Global 3D trajectory estimation. We aim to estimate the global human trajectory from dynamic cameras. We do not explicitly estimate the camera motion. Instead, we use a + +![](images/064327ecdf6902a3f8ea4936e7a2ff7497f90a40a492bac4ca3ded7c8799c174.jpg) +(a) Tracking on 3DPW + +![](images/023b77276145be755fef9e21f37451f08319d524896bf44609619ef6bb960fc5.jpg) +(b) Tracking on MuPoTS-3D + +![](images/6043dc6c2841750742dabcce635b3220c2445b87d0468615723557c5418047d1.jpg) +(c) Global 3D trajectory estimation on DynaCam +Figure 3. Qualitative comparisons to SOTA methods (ByteTrack [63], PHALP [40], and GLAMR [58]) on 3DPW, MuPoTS-3D, and DynaCam. Red dash boxes highlight the ID switches. Yellow dot-dash boxes highlight the misses under occlusions. + +
MethodTranslating CameraRotating CameraFPS
Traj↓Velocity↓Traj↓Velocity↓
BEV+DPVO2.0240.1543.6220.1195.7
GLAMR [58]1.7450.1051.7590.0551.5
TRACE1.4330.0821.0380.04521.2
+ +Table 1. Errors in estimated 3D human trajectory in world coordinates (in $m$ ) and runtime efficiency on the DynaCam dataset. + +
MethodsIDs↓MOTA↑IDF1↑HOTA↑
Trackformer [36]4324.962.753.2
Tracktor [4]5351.570.950.3
FlowPose [51]4921.467.141.8
T3DP [39]3862.179.159.2
PHALP [40]2266.281.459.4
ByteTrack [63]1573.384.663.5
TRACE086.993.465.3
+ +world motion map to represent the global trajectory in world coordinates, which implies the camera motion. Therefore, we evaluate the global trajectory error, instead of the camera pose, in Tab. 3. First, we evaluate global 3D trajectory estimation on DynaCam. We compare TRACE with two baseline solutions. The first one uses BEV to estimate the subjects in camera coordinates and DPVO [48], a SLAM + +Table 2. Tracking results on MuPoTS-3D [35]. Results of the compared methods [4,36,39,40,51] are from [40]. + +
MethodsIDs↓MOTA↑IDF1↑HOTA↑
PHALP [40]597.993.774.2
YOLOX+ByteTrack397.398.273.1
BEV+ByteTrack3793.679.159.3
TRACE w/o MO495.897.372.7
TRACE199.399.774.7
+ +Table 3. Tracking results on Dyna3DPW [50]. w/o MO means ablating the estimated 3D motion offsets for tracking. + +
MethodPAMPJPE↓MPJPE↓PVE↓
HybrIK [27]45.074.186.5
METRO [30]47.977.188.2
GLAMR [58]50.7--
CLIFF [28]43.069.081.2
D&D [26]42.773.788.6
ROMP [44]47.376.793.4
BEV [45]46.978.592.3
TRACE37.879.197.3
+ +Table 4. Comparisons with SOTA methods for 3D human pose and shape estimation on the 3DPW test set. + +method, to estimate the camera and its motion; we call this BEV+DPVO. As shown in Tab. 1, TRACE significantly outperforms BEV+DPVO in the accuracy of global 3D trajectory estimation. The moving people in the scene make it hard for a SLAM method to extract stable corresponding keypoints. Additionally, our synthetic camera motions + +differ from real camera motions and this may hurt DPVO's performance. A more direct comparison is to GLAMR [58]. TRACE outperforms GLAMR in both accuracy and efficiency. In Fig. 3(c), we also perform visual comparisons with GLAMR on DynaCam. These results demonstrate the benefit of estimating global human trajectory using a holistic 5D representation. We provide more results in the supplemental video. + +Multi-subject tracking. To evaluate the performance of tracking subjects with dynamic cameras in real-world scenes, we compare TRACE with recent methods on Dyna3DPW. PHALP [40] is a recent (SOTA) method that uses 3D cues and appearance to track people using SMPL. YOLOX+ByteTrack [17, 63] is a recently proposed and popular tracking-by-detection solution. These methods are designed to track all the people in a scene. Therefore, we process their results to avoid them being penalized for tracking unlabeled passers-by; 3DPW has annotations for at most 2 people in a scene but some scenes contain many people. We first obtain their tracking results using their official code. We then select the tracking results that achieve maximum IoU with the labeled ground truth subjects; we use these tracks for evaluation. Note that, for a fair comparison with ByteTrack, TRACE runs in an on-line mode, without optimizing the past results. As shown in Tab. 3, TRACE outperforms PHALP, the previous 3D-representation-based method [40], and tracking-by-detection [17, 63] methods on Dyna3DPW. To evaluate the tracking robustness under long-term occlusion, we evaluate TRACE on MuPoTS-3D. Results of previous SOTA methods [4, 36, 39, 40, 51] are from [40]. Again, for a fair comparison, we filter out the unlabeled people in the ByteTrack results. As shown in Tab. 2, TRACE significantly outperforms previous SOTA methods. In particular, TRACE significantly reduces ID switches under long-term occlusions. These results illustrate the effectiveness and robustness of the proposed method for in-the-wild videos. The qualitative comparisons are presented in Fig. 3 and the supplemental video. + +3D human pose and shape estimation. Finally, we evaluate 3D human regression performance in DC-video using the 3DPW test set. Because 3DPW does not provide ground-truth 3D translation in world coordinates, we evaluate root-relative 3D pose. We compare TRACE with the recent one-stage [44,45] and multi-stage [26-28,30,58] methods. Tab. 4 shows that TRACE significantly outperforms the previous method by $11.4\%$ in terms of PMPJPE. These results demonstrate that learning a holistic 5D representation in an end-to-end manner is powerful. We adopt a fixed field of view $(\mathrm{FOV} = 50^{\circ})$ to define our camera space, which is very different from that of 3DPW video. This results in a misalignment between the predicted 3D orientation of the body mesh and the ground truth, resulting in higher MPJPE and MVE. This illustrates the importance of also estimating + +the camera so that our estimated global coordinate system is consistent with the real world. This is future work. + +# 4.3. Ablation Studies + +Temporal 5D representation v.s. image-level 3D representation. We go beyond BEV's image-level 3D representation and build a temporal 5D representation. As shown in Tab. 1, 3, and 4, TRACE outperforms BEV or the multistage solutions using BEV on most metrics. This demonstrates the value of learning a holistic 5D representation. + +3D Motion Offset map. We also evaluate the effect of using predicted 3D motion offsets for tracking. As shown in Tab. 3, 3D motion offsets improve performance by $3.5\%$ , $2.4\%$ , and $2.7\%$ in terms of MOTA, IDF1, and HOTA. + +# 5. Conclusions + +Human pose and shape estimation is not an end to itself. Rather, estimating the 3D human in motion is useful for many tasks from behavior analysis to computer graphics. However, to be useful, it is important to know the motion of humans with respect to the 3D scene and other people. This means that HPS methods must estimate humans in a global coordinate system and provide consistent tracks of people across time. For generality, they also need to be able to do this from arbitrary moving cameras. + +To tackle these challenging problems, we propose a novel 5D representation and a new neural architecture that reasons about people in 5D; that is, their 3D position, temporal trajectory, and identity. Moving to a 5D representation enables our method, TRACE, to take a holistic view of the video, processing full frames and incorporating temporal features. The core innovation of TRACE lies in its novel temporal representation in the form of new "maps" that represent the motion of people across time in the camera and global coordinates. These allow TRACE to be trained end-to-end, thus exploiting rich information from the video to solve the task. TRACE is the first such single-shot method for 3D HPS estimation from video and it achieves SOTA results on common benchmarks. + +Future work should look at explicitly estimating the camera, using training data like BEDLAM [6], which contains complex human motion, 3D scenes, and camera motions. We believe that camera motion and human motion provide complimentary information that can be used to recover human motion in world coordinates with metric accuracy. + +Acknowledgements: This work was partially supported by the National Key R&D Program of China under Grant (No. 2020AAA0103800) and Beijing Nova Program (No. 20220484063). + +MJB Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2023.txt + +# References + +[1] Bilibili. https://www.bilibili.com/.6 +[2] Pexels. https://www.pexels.com.6 +[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6 +[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, pages 941-951, 2019. 3, 7, 8 +[5] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, pages 1-10, 2008. 6 +[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of 3d human bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 2, 8 +[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, pages 561-578, 2016. 2 +[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 3 +[9] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, pages 5386-5395, 2020. 4 +[10] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, pages 436-454, 2020. 3 +[11] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 3 +[12] Andreas Doering, Di Chen, Shanshan Zhang, Bernt Schiele, and Juergen Gall. Posetrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In CVPR, pages 20963-20972, 2022. 6 +[13] Mahsa Ehsanpour, Fatemeh Saleh, Silvio Savarese, Ian Reid, and Hamid Rezatofighi. Jrdb-act: A large-scale dataset for spatio-temporal action, social group and activity detection. In CVPR, pages 20983–20992, 2022. 3 +[14] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. A mobile vision system for robust multi-person tracking. In CVPR, pages 1-8, 2008. 3 +[15] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoşa Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, pages 10849-10859, 2021. 3 +[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to + +image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 6 +[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3, 6, 8 +[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 3 +[19] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. TPAMI, 36(7):1325-1339, 2013. 6 +[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In ECCV, 2020. 3, 6 +[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 3 +[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. VIBE: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3 +[23] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In ICCV, pages 11035-11045, 2021. 2, 3 +[24] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3 +[25] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In CVPR, pages 11605-11614, 2021. 6 +[26] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&d: Learning human dynamics from dynamic camera. In ECCV, 2022. 3, 7, 8 +[27] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In CVPR, pages 3383-3393, 2021. 7, 8 +[28] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, pages 590-606, 2022. 3, 7, 8 +[29] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In CVPR, 2023. 3 +[30] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 7, 8 +[31] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, pages 930-939, 2021. 2, 3, 4 +[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):1-16, 2015. 3, 4, 5 + +[33] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, pages 548-578, 2021. 6 +[34] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3D human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 6 +[35] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 3, 6, 7 +[36] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 7, 8 +[37] Gyeongsik Moon and Kyoung Mu Lee. Pose2Pose: 3d positional pose-guided 3d rotational pose prediction for expressive 3d human pose and mesh estimation. arXiv, 2020. 2, 3 +[38] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. TexturePose: Supervising human mesh estimation with texture consistency. In ICCV, pages 803-812, 2019. 2, 3 +[39] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. In NeurIPS, 2021. 3, 7, 8 +[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3D appearance, location & pose. In CVPR, 2022. 2, 3, 7, 8 +[41] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6 +[42] Weijian Ruan, Wu Liu, Qian Bao, Jun Chen, Yuhao Cheng, and Tao Mei. Pointet: pose-guided ovonic insight network for multi-person pose tracking. In MM, pages 284-292, 2019. 3 +[43] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 3 +[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 1, 3, 7, 8 +[45] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[46] Yu Sun, Yun Ye, Wu Liu, Wenpeng Gao, YiLi Fu, and Tao Mei. Human mesh recovery from monocular images via a skeleton-disentangled representation. In ICCV, pages 5348-5357, 2019. 2, 3 +[47] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 5, 6 +[48] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 7 + +[49] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In CVPR, 2023. 2 +[50] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 6, 7 +[51] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018. 3, 7, 8 +[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3 +[53] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3 +[54] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. arXiv preprint arXiv:2204.12484, 2022. 3, 6 +[55] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 3 +[56] Hongwei Yi, Chun-Hao P. Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J. Black. MIME: Human-aware 3D scene generation. In CVPR, 2023. 2 +[57] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2 +[58] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In CVPR, 2022. 2, 3, 4, 7, 8 +[59] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3D pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 3 +[60] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 2, 3 +[61] Shanshan Zhang, Rodrigo Benenson, and Bernt Schiele. Citypersons: A diverse dataset for pedestrian detection. In CVPR, 2017. 3 +[62] Weiyu Zhang, Menglong Zhu, and Konstantinos G Derpanis. From actemes to action: A strongly-supervised representation for detailed action understanding. In ICCV, pages 2248-2255, 2013. 6 +[63] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 2, 3, 6, 7, 8 +[64] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In ICCV, pages 6239-6249, 2021. 3 + +[65] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5 \ No newline at end of file diff --git a/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/images.zip b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f2522c76e5a07c189605b1b05984dc2d70241171 --- /dev/null +++ b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b908233adf3c824caf959a140a5c0405ae7bdc7503d3194e523697bdc4d9e31 +size 526343 diff --git a/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/layout.json b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f731a09eba8b8500c4036fbb151997ac1e5fd31 --- /dev/null +++ b/2023/TRACE_ 5D Temporal Regression of Avatars With Dynamic Cameras in 3D Environments/layout.json @@ -0,0 +1,9651 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 158, + 103, + 436, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 103, + 436, + 138 + ], + "spans": [ + { + "bbox": [ + 158, + 103, + 436, + 138 + ], + "type": "text", + "content": "TRACE: 5D Temporal Regression of Avatars with Dynamic Cameras in 3D Environments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "spans": [ + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "text", + "content": "Yu Sun" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "text", + "content": " Qian Bao" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "text", + "content": " Wu Liu" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "text", + "content": " Tao Mei" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "text", + "content": " Michael J. Black" + }, + { + "bbox": [ + 134, + 148, + 455, + 163 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 163, + 454, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 163, + 454, + 178 + ], + "spans": [ + { + "bbox": [ + 139, + 163, + 454, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 163, + 454, + 178 + ], + "type": "text", + "content": "Harbin Institute of Technology " + }, + { + "bbox": [ + 139, + 163, + 454, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 163, + 454, + 178 + ], + "type": "text", + "content": "Explore Academy of JD.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 178, + 454, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 178, + 454, + 192 + ], + "spans": [ + { + "bbox": [ + 139, + 178, + 454, + 192 + ], + "type": "text", + "content": "3Max Planck Institute for Intelligent Systems HiDream.ai Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 163, + 194, + 430, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 194, + 430, + 205 + ], + "spans": [ + { + "bbox": [ + 163, + 194, + 430, + 205 + ], + "type": "text", + "content": "yusunhit@gmail.com, baoqian@jd.com, liuwul@jd.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 209, + 401, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 209, + 401, + 219 + ], + "spans": [ + { + "bbox": [ + 188, + 209, + 401, + 219 + ], + "type": "text", + "content": "tmei@hidream.ai, black@tuebingen.mpg.de" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 56, + 228, + 83, + 246 + ], + "blocks": [ + { + "bbox": [ + 56, + 228, + 83, + 246 + ], + "lines": [ + { + "bbox": [ + 56, + 228, + 83, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 83, + 246 + ], + "type": "image", + "image_path": "a5971fff5bed9d6d158c0495cedd0412794397823f881d9964ea67bfed7c7296.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 235, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 235, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 85, + 235, + 200, + 243 + ], + "type": "text", + "content": "Dynamic Monocular Camera" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 56, + 246, + 170, + 343 + ], + "blocks": [ + { + "bbox": [ + 56, + 246, + 170, + 343 + ], + "lines": [ + { + "bbox": [ + 56, + 246, + 170, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 170, + 343 + ], + "type": "image", + "image_path": "c61c8e122624cf0fc96aff32935004c837ea9533f8e03179629d700820206742.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 171, + 245, + 257, + 343 + ], + "blocks": [ + { + "bbox": [ + 171, + 245, + 257, + 343 + ], + "lines": [ + { + "bbox": [ + 171, + 245, + 257, + 343 + ], + "spans": [ + { + "bbox": [ + 171, + 245, + 257, + 343 + ], + "type": "image", + "image_path": "cf1fcb8fea368193fde60da9100333312b09de451a144b2314415a59fc11bada.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 259, + 245, + 343, + 343 + ], + "blocks": [ + { + "bbox": [ + 259, + 245, + 343, + 343 + ], + "lines": [ + { + "bbox": [ + 259, + 245, + 343, + 343 + ], + "spans": [ + { + "bbox": [ + 259, + 245, + 343, + 343 + ], + "type": "image", + "image_path": "68ded7e47aa57dc68b9545efe0ce2abc58b36cdcd89b9f8b492ac993b61a9a8a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 345, + 245, + 430, + 343 + ], + "blocks": [ + { + "bbox": [ + 345, + 245, + 430, + 343 + ], + "lines": [ + { + "bbox": [ + 345, + 245, + 430, + 343 + ], + "spans": [ + { + "bbox": [ + 345, + 245, + 430, + 343 + ], + "type": "image", + "image_path": "e2f7565c67777ab738977981f4af699e6e56a4165a08a5cc280c9a875bfca665.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 451, + 245, + 535, + 343 + ], + "blocks": [ + { + "bbox": [ + 451, + 245, + 535, + 343 + ], + "lines": [ + { + "bbox": [ + 451, + 245, + 535, + 343 + ], + "spans": [ + { + "bbox": [ + 451, + 245, + 535, + 343 + ], + "type": "image", + "image_path": "9da159a468c0f4c2a27c14ebccb8b692b1b3cf9c4967a0af906e9fe40fa891b9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 56, + 353, + 295, + 474 + ], + "blocks": [ + { + "bbox": [ + 134, + 344, + 457, + 353 + ], + "lines": [ + { + "bbox": [ + 134, + 344, + 457, + 353 + ], + "spans": [ + { + "bbox": [ + 134, + 344, + 457, + 353 + ], + "type": "text", + "content": "(a) Videos Captured by Dynamic Monocular Camera and Human 3D Recovery in Each Frame" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 353, + 295, + 474 + ], + "lines": [ + { + "bbox": [ + 56, + 353, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 295, + 474 + ], + "type": "image", + "image_path": "2ef673ece98846b17409ba65c428ceecd65d2001e3bb8e744dd8bf3417f38a34.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 475, + 287, + 484 + ], + "lines": [ + { + "bbox": [ + 63, + 475, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 63, + 475, + 287, + 484 + ], + "type": "text", + "content": "(b) The 3D humans recovered by BEV [45] in camera coordinates" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 487, + 546, + 533 + ], + "lines": [ + { + "bbox": [ + 45, + 487, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 45, + 487, + 546, + 533 + ], + "type": "text", + "content": "Figure 1. 5D temporal reconstruction of multiple 3D people in global coordinates with a dynamic camera. We introduce TRACE, a monocular one-stage method with a holistic 5D representation that enables the network to explicitly reason about human motion across time. Unlike previous methods [44,45] (b) that regress all 3D people from each frame in camera coordinates, TRACE (c) tracks the subjects (a) presented in the first frame through time and recovers their global trajectories in global coordinates in one shot." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 298, + 353, + 538, + 474 + ], + "blocks": [ + { + "bbox": [ + 298, + 353, + 538, + 474 + ], + "lines": [ + { + "bbox": [ + 298, + 353, + 538, + 474 + ], + "spans": [ + { + "bbox": [ + 298, + 353, + 538, + 474 + ], + "type": "image", + "image_path": "bd6f2064535a36ec77850e054c10794e7d189078265c27a80c31296b43e33d24.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 475, + 525, + 484 + ], + "lines": [ + { + "bbox": [ + 309, + 475, + 525, + 484 + ], + "spans": [ + { + "bbox": [ + 309, + 475, + 525, + 484 + ], + "type": "text", + "content": "(c) The 5D humans recovered by TRACE in global coordinates" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 546, + 192, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 546, + 192, + 559 + ], + "spans": [ + { + "bbox": [ + 143, + 546, + 192, + 559 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 572, + 288, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 288, + 681 + ], + "type": "text", + "content": "Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still cannot reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel ar" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 548, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 668 + ], + "type": "text", + "content": "chitectural components. Most importantly, it uses two new \"maps\" to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even during long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and using full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code1 and dataset2 are released for research purposes." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 253, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 253, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 253, + 703 + ], + "type": "text", + "content": "*This work was done when Yu Sun was an intern at JD.com." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 693, + 509, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 509, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 509, + 703 + ], + "type": "text", + "content": "1https://www.yusun.work/TRACE/TRACE.html" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 703, + 496, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 496, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 496, + 712 + ], + "type": "text", + "content": "2https://github.com/Arthur151/DynaCam" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8856" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 258 + ], + "type": "text", + "content": "The estimation of 3D human pose and shape (HPS) has many applications and there has been significant recent progress [6, 7, 21-24, 37, 38, 46, 49, 56, 57, 60]. Most methods, however, reason only about a single frame at a time and estimate humans in camera coordinates. Moreover, such methods do not track people and are unable to recover their global trajectories. The problem is even harder in typical hand-held videos, which are filmed with a dynamic, moving, camera. For many applications of HPS, single-frame estimates in camera coordinates are not sufficient. To capture human movement and then transfer it to a new 3D scene, we must have the movement in a coherent global coordinate system. This is a requirement for computer graphics, sports, video games, and extended reality (XR)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 259, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 259, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 49, + 259, + 286, + 377 + ], + "type": "text", + "content": "Our key insight is that most methods estimate humans in 3D, whereas the true problem is 5D. That is, a method needs to reason about 3D space, time, and subject identity. With a 5D representation, the problem becomes tractable, enabling a holistic solution that can exploit the full video to infer multiple people in a coherent global coordinate frame. As illustrated in Fig. 1, we develop a unified method to jointly regress the 3D pose, shape, identity, and global trajectory of the subjects in global coordinates from monocular videos captured by dynamic cameras (DC-videos)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 379, + 286, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 379, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 49, + 379, + 286, + 580 + ], + "type": "text", + "content": "To achieve this, we deal with two main challenges. First, DC-videos contain both human motion and camera motion and these must be disentangled to recover the human trajectory in global coordinates. One idea would be to recover the camera motion relative to the rigid scene using structure-from-motion (SfM) methods (e.g. [31]). In scenes containing many people and human motion, however, such methods can be unreliable. An alternative approach is taken by GLAMR [58], which infers global human trajectories from local 3D human poses, without taking into account the full scene. By ignoring evidence from the full image, GLAMR fails to capture the correct global motion in common scenarios, such as biking, skating, boating, running on a treadmill, etc. Moreover, GLAMR is a multi-stage method, with each stage dependent on accurate estimates from the preceding one. Such approaches are more brittle than our holistic, end-to-end, method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 582, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 582, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 582, + 286, + 712 + ], + "type": "text", + "content": "The other challenge, as shown in the upper right corner of Fig. 1, is that severe occlusions are common in videos with multiple people. Currently, the most popular tracking strategy is to infer the association between 2D detections using a temporal prior (e.g. Kalman filter) [63]. However, in DC-videos, human motions are often irregular and can easily violate hand-crafted priors. PHALP [40] is one of the few methods to address this for 3D HPS. It uses a classical, multi-stage, detection-and-tracking formulation with heuristic temporal priors. It does not holistically reason about the sequence and is not trained end-to-end." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 179 + ], + "type": "text", + "content": "To address these issues, we reason about people using a 5D representation and capture information from the full image and the motion of the scene. This holistic reasoning enables the reliable recovery of global human trajectories and subject tracking using a single-shot method. This is more reliable than multi-stage methods because the network can exploit more information to solve the task and is trained end-to-end. No hand-crafted priors are needed and the network is able to share information among modules." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 182, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 182, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 182, + 545, + 337 + ], + "type": "text", + "content": "Specifically, we develop TRACE, a unified one-stage method for Temporal Regression of Avatars with dynamic Cameras in 3D Environments. The architecture is inspired by BEV [45], which directly estimates multiple people in depth from a single image using multiple 2D maps. BEV uses a 2D map representing an imaginary, \"top down\", view of the scene. This is combined with an image-centric 2D map to reason about people in 3D. Our key insight is that the idea of maps can be extended to represent how people move in 3D. With this idea, TRACE introduces three new modules to holistically model 5D human states, performing multi-person temporal association, and inferring human trajectories in global coordinates; see Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 340, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 340, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 308, + 340, + 545, + 638 + ], + "type": "text", + "content": "First, to construct a holistic 5D representation of the video, we extract temporal image features by fusing single-frame feature maps from the image backbone with a temporal feature propagation module. We also compute the optical flow between adjacent frames with a motion backbone. The optical flow provides short-term motion features that carry information about the motion of the scene and the people. Second, to explicitly track human motions, we introduce a novel 3D motion offset map to establish the association of the same person across adjacent frames. This map contains a 3D offset vector at each position, which represents the difference between the 3D positions of the same subject from the previous frame to the current frame in camera coordinates. We also introduce a memory unit to keep track of subjects under long-term occlusion. Note that the 3D trajectories are built in camera space, and TRACE uses a novel world motion map that transfers the trajectories to global coordinates. At each position, this map contains a 6D vector to represent the difference between the 3D positions of the corresponding subject from the previous frame to the current frame and its 3D orientation in world coordinates. Taken together, this novel network architecture goes beyond prior work by taking information from the full video frames to address detection, pose estimation, tracking, and occlusion in a holistic network that is trained end-to-end." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 642, + 545, + 712 + ], + "type": "text", + "content": "To enable training and evaluation of global human trajectory estimation from in-the-wild DC-videos, we build a new dataset, DynaCam. Since collecting global human trajectories and camera poses with in-the-wild DC-videos is difficult, we simulate a moving camera using publicly available in-the-wild panoramic videos and regular videos cap" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "8857" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "tured by static cameras. In this way, we create more than 500 in-the-wild DC-videos with precise camera pose annotations. Then we generate pseudo-ground-truth 3D human annotations via fitting [20] SMPL [32] to detected 2D pose sequences [17,54,63]. With 2D/3D human pose and camera pose annotations, we can obtain the global human trajectories using the PnP algorithm [16]. This dataset is sufficient to train TRACE to deal with dynamic cameras." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "type": "text", + "content": "We evaluate TRACE on two multi-person in-the-wild benchmarks (3DPW [50] and MuPoTS-3D [35]) and our DynaCam dataset. On 3DPW, TRACE achieves the state-of-the-art (SOTA) PA-MPJPE of " + }, + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "type": "inline_equation", + "content": "37.8\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 168, + 288, + 287 + ], + "type": "text", + "content": ", less the current best (42.7 [26]). On MuPoTS-3D, TRACE outperforms previous 3D-representation-based methods [39, 40] and tracking-by-detection methods [63] on tracking people under long-term occlusion. On DynaCam, TRACE outperforms GLAMR [58] in estimating the 3D human trajectory in global coordinates from DC-videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 288, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 288, + 431 + ], + "type": "text", + "content": "In summary, our main contributions are: (1) We introduce a 5D representation and use it to learn holistic temporal cues related to both 3D human motions and the scene. (2) We introduce two novel motion offset representations to explicitly model temporal multi-subject association and global human trajectories from temporal clues in an end-to-end manner. (3) We estimate long-term 3D human motions over time in global coordinates, achieving SOTA results. (4) We collect the DynaCam dataset of DC-videos with pseudo ground truth, which facilitates the training and evaluation of global human trajectory estimation. The code and dataset are publicly available for research purposes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 134, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "content": "Monocular 3D mesh regression with full images. Most existing methods [21-24, 26, 29, 37, 38, 46, 52, 53, 59, 60] take a multi-stage detection-based pipeline to estimate 3D HPS from cropped image patches, which exclude important cues, such as camera information and human-scene relationships. A few recent multi-stage [28] and one-stage [44, 45] methods have made steps towards using the full-image information. For instance, CIUFF [28] estimates 3D HPS by taking into account the bounding box locations, giving the method camera information and improving accuracy. To directly estimates multiple people at once from the full image, ROMP [44] introduces a 2D Center heatmap and a Mesh parameter map to represent 2D human locations and 3D human body meshes, respectively. BEV [45] goes beyond ROMP by introducing an imaginary bird's-eye-view map, which is combined with the front-view maps to construct a 3D view in camera coordinates. However, they only model 3D HPS in camera coordinates from a single image. Using \"maps\" like ROMP and BEV, TRACE also looks at the full image. We go further, however, by introducing novel maps that model human motions across a video se" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 425, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 425, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 425, + 84 + ], + "type": "text", + "content": "quence in global coordinates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 224 + ], + "type": "text", + "content": "Tracking datasets. While there are many tracking datasets [8, 10, 11, 14, 18, 61] with 2D annotations, only a few [13, 15] capture the 3D trajectory of pedestrians. In both cases, the scene and human activities are limited. To address this, we use 3DPW [50] and MuPoTS-3D [35] for tracking evaluation. 3DPW is the most relevant dataset for our task and it provides a real-world test case. 3DPW contains videos that are captured by a moving camera that follows subjects to record their activities in many daily scenes. MuPoTS-3D contains rich multi-person interaction scenes with long-term occlusions for tracking evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 384 + ], + "type": "text", + "content": "Tracking 3D people through occlusions. Most existing methods [4, 42, 51, 63] perform tracking using 2D image cues. The classic tracking-by-detection paradigm focuses on associating the 2D detections using a temporal prior (e.g. Kalman filter). When applied to DC-videos containing rapid human and camera motions that violate the hand-crafted priors, such methods are brittle. Going beyond 2D, PHALP [40] separately extracts 3D human pose, appearance, and location with a multi-stage design from each video frame, and then assembles them for tracking. In contrast to these multi-stage methods, which are susceptible to errors in early stages, we explicitly learn the 3D human trajectory from temporal 5D cues in an end-to-end manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "Monocular global 3D human trajectory reasoning. Most existing methods [64] that reason about global 3D human trajectories do so with static, calibrated, cameras in a multi-view setting. A few recent methods [31,58] have addressed the ill-posed problem of extracting the global motions of humans from monocular video. Liu et al. [31] employ a structure-from-motion (SfM) method [43] to estimate the camera poses from monocular videos captured by a dynamic camera. However, when the input video contains the movement of multiple subjects, it is hard for SfM methods to extract sufficiently many stable keypoints for reliable camera estimation. GLAMR [58] adopts a multi-stage pipeline to infer the global human trajectory from root-relative local human 3D poses estimated from each frame. The per-frame human pose estimates make it vulnerable to occlusion. Additionally, GLAMR relies on bounding boxes, ignoring scene-related information. Consequently, GLAMR fails in common scenarios like riding a bike or skating. In concurrent work (in this proceedings), SLAHMR [55] uses a multi-stage optimization-based approach that combines structure from motion with human motion priors to estimate 4D human trajectories in global coordinates; this is very computationally expensive. In contrast to previous multi-stage methods, TRACE simultaneously combines scene information and 3D human motions with a novel 5D representation to holistically exploit all temporal cues and to enable end-to-end training." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8858" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 259 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 259 + ], + "type": "image", + "image_path": "aecf35d1f4ee96994fccd5c5d0b73fad77fa335a7ebeebdea5aa07a6a82b98a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 263, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. TRACE Overview. TRACE takes a video sequence and regions in the first frame corresponding to the subjects to be tracked. TRACE encodes the video and its motion with temporal features and optical flow. A novel 3D Motion Offset map reasons about human trajectories in camera coordinates. The World Motion map represents the trajectory in global coordinates. A memory unit deals with occlusions by encoding the subject identities. TRACE is trained end-to-end to estimate the 3D shape and pose of multiple people throughout a video in global coordinates. See Sec. 3.1 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 333, + 103, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 103, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 103, + 345 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 350, + 115, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 115, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 115, + 361 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "content": "The overall framework of TRACE is shown in Fig. 2. Given a video sequence captured with a dynamic camera " + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\{f_i, i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "content": " frames, the user specifies " + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "content": " tracking subjects shown in the first frame. Our goal is to simultaneously recover the 3D pose, shape, identity, and trajectory of each subject in global coordinates. To achieve this, TRACE first extracts temporal features and then decodes each sub-task with a separate head network. First, via two parallel backbones, TRACE encodes the video and its motion into temporal image feature maps " + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "inline_equation", + "content": "F_i'" + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "content": " and motion feature maps " + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 46, + 365, + 287, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "The Detection and Tracking branches take these features and perform multi-subject tracking in camera coordinates. Unlike BEV [45], our detection method takes temporal image features " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i'" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " as input. It uses the features to detect the 3D human positions " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " and their confidence " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " for all people in frame " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": ". The Mesh branch regresses all the human mesh parameters " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "(\\pmb{\\theta}_i, \\pmb{\\beta}_i)" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": ", in SMPL [32] format, from the input Feature maps. Unlike BEV, this branch takes both temporal image features and motion features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "The combined features " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "(F_{i}^{\\prime},O_{i})" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " are fed to our novel Tracking branch to estimate the 3D Motion Offset map, indicating the 3D position change of each subject across frames. The new Memory Unit takes the 3D detection and its 3D motion offset as input. It then determines the subject identities and builds human trajectories " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{t}_i" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " subjects in camera coordinates. Note that, like BEV, our detection branch finds all the people in the video frames but our goal is to track only the " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " input subjects. Consequently, the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 334, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 357 + ], + "type": "text", + "content": "memory unit filters out detected people who do not match the subject trajectories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": "Finally, to estimate subject trajectories in global coordinates, the World branch estimates a world motion map, representing the 3D orientation " + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_i" + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": " and 3D translation offset " + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\Delta T_{i}" + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": " subjects in global coordinates. Accumulating " + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\Delta T_{i}" + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": ", starting with the 3D position " + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\hat{t}_{1}" + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": " of the tracked subjects in the first frame, gives their global 3D trajectory " + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 305, + 358, + 546, + 454 + ], + "type": "text", + "content": ". Note that the global (\"world\") coordinates are defined relative to the camera coordinates of the first frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 465, + 495, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 495, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 495, + 478 + ], + "type": "text", + "content": "3.2. Holistic 5D Representation: Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "content": "Rather than directly estimating camera poses from environment keypoints [31] or inferring global human trajectories from local body poses [58], we develop a 5D representation to directly reason about human states, perform multiperson temporal association, and infer human trajectories in global coordinates. Learning a holistic 5D representation is the foundation of our one-stage framework. The representation has five main parts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "i. Temporal feature maps. To construct the temporal feature maps encoding 5D human states and scenes information, we need to extract both single-frame image features and the motion features between adjacent frames. Therefore, given frame " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "f_{i-1}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ", we adopt a parallel two-branch structure to extract temporal image feature maps " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i'" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " and motion feature maps " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " for the current frame " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ". First, in the image branch, we extract single-frame feature maps " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "F_{i-1}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " with an image backbone (HRNet-32 [9]). To extract long-term and short-term motion features, we construct a temporal feature propagation module by combining" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8859" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "a ConvGRU [47] module, a Deformable convolution [65] module, and a residual connection. With these, we fuse the image feature maps to generate a temporal image feature map " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i'" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". See Sup. Mat. for details and experimental analysis. Additionally, in the motion branch, we estimate the optical flow map " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " between frames " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "f_{i-1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " with a motion backbone (RAFT [47]), to extract motion features of both people and scenes. From the combined temporal feature " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "(\\pmb{F}_i', O_i)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", then we estimate five maps for the task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": "ii. 3D detection maps. From the temporal image feature " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i'" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": ", we estimate a 3D Center map " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "M_C^{3D} \\in \\mathbb{R}^{1 \\times D \\times H \\times W}" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": " for coarse human detection and a 3D Localization map " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "M_L^{3D} \\in \\mathbb{R}^{1 \\times D \\times H \\times W}" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": " for fine localization. The two 3D detection maps are composited from the front-view maps and the 2D bird's-eye-view maps following BEV [45]. For " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": " subjects, we first parse out the detected 3D center positions " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{t}}_i" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": " and their detection confidence " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": " from the 3D Center map. Then we sample the 3D Localization map at " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{t}}_i" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": " to obtain the fine 3D localization offset vectors " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\Delta t_i" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": ". The predicted 3D translation of subjects to be tracked in camera coordinates is " + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i = \\widetilde{\\pmb{t}}_i + \\Delta t_i" + }, + { + "bbox": [ + 47, + 182, + 289, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": "iii. 3D Motion Offset map. To track subjects between adjacent frames using " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i'" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": ", we estimate a 3D Motion Offset map " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "M_M^{3D} \\in \\mathbb{R}^{3 \\times D \\times H \\times W}" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " via a BEV-based 3D composition. We sample the " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "M_M^{3D}" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " at detected 3D center positions " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pmb{t}}_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " to obtain the 3D motion offset vectors " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "\\Delta m_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " which represent the 3D position changes of the subjects from frame " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "f_{i-1}" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " in camera coordinates. With " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "\\Delta m_i, t_i, c_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " as input, a memory unit (Sec. 3.3) predicts the tracked identities " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " and the optimized 3D position " + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "inline_equation", + "content": "\\hat{t}_i" + }, + { + "bbox": [ + 46, + 327, + 289, + 470 + ], + "type": "text", + "content": " in camera coordinates, which filters out low-confidence detections while keeping track of subjects through long-term occlusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": "iv. Mesh Feature map. To regress the SMPL parameters of subjects, we first estimate a 2D Mesh Feature map " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "M_F^{2D} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " from temporal features " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "(F_i', O_i)" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ". With the tracked 3D positions " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\widehat{t}_i = (u_i, v_i, d_i)" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ", we sample a mesh feature vector from " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "M_F^{2D}" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " at 2D positions " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "(u_i, v_i)" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ". To differentiate the features of people at different depths, we map the " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " to a 128-dim encoding vector via an embedding layer and add this with the mesh feature vector to regress SMPL [32] parameters " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "(\\theta_i, \\beta_i)" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ". The SMPL model maps the pose and shape, " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\beta_i" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ", to a 3D human body mesh " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{6890 \\times 3}" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ". With a sparse weight matrix " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{Q \\times 6890}" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " that describes the mapping from " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " body keypoints, we can obtain the 3D positions of body keypoints " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "J \\in \\mathbb{R}^{Q \\times 3}" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "RB" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ". With the estimated 3D positions " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\widehat{t}_i" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": " and a predefined camera projection matrix " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": ", the projected 2D keypoints are " + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "inline_equation", + "content": "J_{2D} = P(J + \\widehat{t}_i)" + }, + { + "bbox": [ + 46, + 472, + 289, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "text", + "content": "v. World Motion map. " + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "inline_equation", + "content": "M_W^{2D} \\in \\mathbb{R}^{6 \\times H \\times W}" + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "text", + "content": " contains the 6D vectors that describe the 3D orientation and the relative translation of the subjects in global coordinates. We sample " + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "inline_equation", + "content": "M_W^{2D}" + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "text", + "content": " at the 2D positions " + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "inline_equation", + "content": "(u_i, v_i) \\in \\widehat{t}_i" + }, + { + "bbox": [ + 46, + 664, + 289, + 715 + ], + "type": "text", + "content": " of the tracked" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "subjects to obtain their 3D body orientation " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " and 3D body motion offset " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\Delta T_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": ". For the tracked subject " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\Delta T_{i}^{k}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " represents their 3D position change from frame " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "f_{i-1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " in global coordinates. We take the camera coordinate system of the first frame as the global coordinate system for the video. The global 3D trajectories of all tracked subjects are obtained by accumulating " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\Delta T_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " to their 3D position " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\hat{t}_{1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " in the first frame, " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{T} = \\{\\hat{t}_{1}, \\hat{t}_{1} + \\Delta T_{2}, \\hat{t}_{1} + \\Delta T_{2} + \\Delta T_{3}, \\dots\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": ". Combining the 3D mesh, ID, and the global trajectories, the network reasons about 3D motions and trajectories of " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " tracked subjects in global coordinates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 213, + 470, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 470, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 470, + 227 + ], + "type": "text", + "content": "3.3. Tracking with a Memory Unit" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "content": "We construct the 3D trajectory of each subject by associating the 3D detections " + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "content": " over time with a 3D motion offset " + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\Delta m_i" + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "content": ". To deal with long-term occlusions, we design the Memory Unit for persistent tracking, which will keep the memory for the full sequence. The memory unit stores the human states during inference and is not used for training. With predicted 3D positions " + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i" + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "content": ", detection confidence " + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "content": ", and 3D motion offsets " + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "inline_equation", + "content": "\\Delta m_i" + }, + { + "bbox": [ + 304, + 232, + 547, + 340 + ], + "type": "text", + "content": " as inputs, the memory unit can track online. In each process, we have three stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": "i. Initialization. First, we discard predicted 3D positions " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{t}_i = (\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i)" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " whose detection confidence is below a threshold " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_c" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": ". We observe that our input video is usually shot by tracking the subjects, therefore, we discard the detection whose " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "1 / z_i" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " is below the scale threshold " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_s" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": ". To suppress duplicate detections, we find the detection pairs whose Euclidean distance is below a pre-defined threshold " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\lambda_d" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " and discard the detection with lower detection confidence. In the first frame, we use the 3D positions " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{t}_1" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " and detection confidences " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " subjects to initialize the " + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 341, + 547, + 461 + ], + "type": "text", + "content": " memory nodes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": "ii. Memory node matching. The memory nodes store the 3D position " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "t_{i-1}^k" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " of subject " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " in previous frames. We match the memory node with the new filtered detections in the current frame. We select the predicted 3D translation " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " that best matches the 3D trajectory as the optimized 3D position " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "\\hat{t}_i" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": ". Specifically, with filtered 3D positions " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " and 3D motion offset " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "\\Delta m_i" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": ", we calculate the Euclidean distance matrix between " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "W t_{i-1}^k" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "W (t_i - \\Delta m_i)" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " is a distance weight vector. To avoid the effects of the nonlinear relationship between depth " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " and human scale in images, we convert the depth value " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "1/(1 + z)" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " when calculating the distance matrix. Using Hungarian matching, we keep the matched pairs whose matching distance is below a threshold " + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_m" + }, + { + "bbox": [ + 304, + 462, + 548, + 653 + ], + "type": "text", + "content": " and use them for memory update. We use ground truth tracks for training and only perform matching during inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "iii. Memory update. We update the successfully matched memory nodes with the new 3D position and detection confidence. For the memory nodes without a matched detection, we accumulate the time since failure. Then we remove the memory nodes whose failure time is" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8860" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "above a threshold " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". Tracking can be done in two modes: on-line or off-line. The former does not allow looking back in time, while the latter does. In the off-line mode, if a new detection re-activates a non-matched memory node, the non-matched part of the 3D trajectory is linearly interpolated. Finally, the memory unit outputs the latest 3D positions " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\widehat{t}_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and tracking IDs " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " of all memory nodes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 164, + 155, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 155, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 155, + 177 + ], + "type": "text", + "content": "3.4. DynaCam Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 266 + ], + "type": "text", + "content": "Even with a powerful 5D representation, we still lack in-the-wild data for training and evaluation of global human trajectory estimation. However, collecting global human trajectory and camera poses for natural DC-videos is difficult. Therefore, we create a new dataset, DynaCam, by simulating camera motions to convert in-the-wild videos captured by static cameras to DC-videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 566 + ], + "type": "text", + "content": "We use over 1000 video clips captured by static regular cameras from the MPII Human Pose Database [3] as well as videos from the InterNet [2]. We also use over 200 panoramic video clips that are either recorded by us with an Insta360 RS panoramic camera or are downloaded from the InterNet [1]. We manually design the 3D rotation and field of view (FOV) of dynamic cameras to track the subjects in panoramic videos. With the designed camera motions, we can project the panoramic frames into perspective views. Also, to simulate the 3D translation of dynamic cameras, we crop the videos captured by static cameras with sliding windows. In this way, we can obtain abundant in-the-wild DC-videos with accurate camera pose annotations. Then we perform 2D human detection, tracking, and 2D pose estimation viaYOLOX [17], ByteTrack [63], and ViT-Pose [54], respectively, to obtain 2D pose sequences of each subject. We estimate SMPL parameters by fitting the 2D poses using EFT [20] or ProHMR [25] and solve for their 3D positions in camera coordinates via the PnP algorithm (RANSAC [16]). Finally, we solve for the 3D human trajectories in the global coordinates with camera pose annotations. We manually filter out the failure cases. In this way, we generate more than 500 annotated DC-videos containing over 48K frames. More than half of video frames are generated from panoramic videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 566, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 615 + ], + "type": "text", + "content": "Limitations: The videos generated with our process only approximate real DC-videos shot in the wild since they lack perspective effects. Despite this they prove useful for training TRACE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 141, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 141, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 141, + 634 + ], + "type": "text", + "content": "3.5. Loss Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "TRACE is supervised by the weighted sum of 15 loss terms that fall into two groups: temporal motion losses and standard image losses. Here we focus on the novel temporal losses. Please refer to the Sup. Mat. for details of all losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "To learn the temporal motion, we introduce a 3D motion offset loss " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and a 6D world motion loss " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_W" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " loss between the predicted 3D motion offset " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\Delta m_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "(\\pmb{t}_i^{\\prime} - \\pmb{t}_{i - 1}^{\\prime})" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "t_i^\\prime" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " is the ground truth 3D human position at frame " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " in our pre-defined camera coordinates (FOV=50°), which is solved for via the PnP algorithm [16]. " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_W" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " consists of six parts, including an " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " loss on the global 3D trajectory " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ", an " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " loss on the velocity/acceleration of 3D trajectory nodes " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\dot{T} / \\ddot{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ", an " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " loss of the velocity/acceleration of the 3D foot keypoints in global coordinates, and an " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " loss on the global 3D body orientation " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 192, + 388, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 388, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 388, + 205 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 212, + 440, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 440, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 440, + 224 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 231, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 374 + ], + "type": "text", + "content": "Training details. During training, we directly use the ground truth trajectory of subjects to replace the estimated trajectory " + }, + { + "bbox": [ + 304, + 231, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 304, + 231, + 545, + 374 + ], + "type": "text", + "content": " for sampling the parameters. We use the pretrained backbone of BEV [45] as the image backbone. We use RAFT [47] as the optical flow backbone. The training consists of two stages. In the first stage, we freeze the weights of the backbones and train the head network for 40 epochs with a learning rate of 5e-5. Then we train the image backbone and the head network together for 10 epochs with a learning rate of 1e-5. We use four V100-16GB GPUs for training. Limited by the GPU memory, we sample 4 video clips as a batch at each iteration; the clip length is 10 frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 375, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 518 + ], + "type": "text", + "content": "Training and evaluation datasets. For training, we use three 3D human pose datasets (Human3.6M [19], MPIINF-3DHP [34, 35], and 3DPW [50]), two 2D human pose datasets (PennAction [62] and PoseTrack [12]), and our DynaCam dataset. We evaluate TRACE on two multi-person in-the-wild benchmarks, 3DPW [50] and MuPoTS-3D [35], and DynaCam. 3DPW videos are most consistent with our tracking scenario. Unfortunately, not all 3DPW videos have complete tracking annotations. We select the 16 videos that do and call this subset Dyna3DPW. We use this challenging subset to evaluate tracking and HPS accuracy in complex scenes with a moving camera." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 518, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 638 + ], + "type": "text", + "content": "Evaluation metrics. For global 3D trajectory estimation, we compute the average 3D position (Traj) and velocity (Velocity) error of the predicted global 3D trajectory " + }, + { + "bbox": [ + 304, + 518, + 547, + 638 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 518, + 547, + 638 + ], + "type": "text", + "content": ". For multi-object tracking, we report the ID switch (IDs), Multi-Object Tracking Accuracy (MOTA [5]), Identification F1-score (IDF1 [41]), and Higher Order Tracking Accuracy (HOTA [33]). To assess the accuracy of 3D human pose/shape estimation, we compute the Mean Per Joint Position Error (MPJPE), Procrustes-aligned MPJPE (PMPJPE), and Mean Vertex Error (MVE)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 638, + 487, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 638, + 487, + 650 + ], + "spans": [ + { + "bbox": [ + 317, + 638, + 487, + 650 + ], + "type": "text", + "content": "Please refer to Sup. Mat. for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 658, + 522, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 522, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 522, + 671 + ], + "type": "text", + "content": "4.2. Comparisons to State-of-the-art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Global 3D trajectory estimation. We aim to estimate the global human trajectory from dynamic cameras. We do not explicitly estimate the camera motion. Instead, we use a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8861" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 218, + 342 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 218, + 342 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 218, + 342 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 218, + 342 + ], + "type": "image", + "image_path": "064327ecdf6902a3f8ea4936e7a2ff7497f90a40a492bac4ca3ded7c8799c174.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 344, + 185, + 354 + ], + "lines": [ + { + "bbox": [ + 104, + 344, + 185, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 344, + 185, + 354 + ], + "type": "text", + "content": "(a) Tracking on 3DPW" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 364, + 342 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 364, + 342 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 364, + 342 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 364, + 342 + ], + "type": "image", + "image_path": "023b77276145be755fef9e21f37451f08319d524896bf44609619ef6bb960fc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 344, + 342, + 354 + ], + "lines": [ + { + "bbox": [ + 240, + 344, + 342, + 354 + ], + "spans": [ + { + "bbox": [ + 240, + 344, + 342, + 354 + ], + "type": "text", + "content": "(b) Tracking on MuPoTS-3D" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 374, + 69, + 541, + 342 + ], + "blocks": [ + { + "bbox": [ + 374, + 69, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 374, + 69, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 374, + 69, + 541, + 342 + ], + "type": "image", + "image_path": "6043dc6c2841750742dabcce635b3220c2445b87d0468615723557c5418047d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 344, + 541, + 354 + ], + "lines": [ + { + "bbox": [ + 365, + 344, + 541, + 354 + ], + "spans": [ + { + "bbox": [ + 365, + 344, + 541, + 354 + ], + "type": "text", + "content": "(c) Global 3D trajectory estimation on DynaCam" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 360, + 546, + 383 + ], + "lines": [ + { + "bbox": [ + 47, + 360, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 546, + 383 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons to SOTA methods (ByteTrack [63], PHALP [40], and GLAMR [58]) on 3DPW, MuPoTS-3D, and DynaCam. Red dash boxes highlight the ID switches. Yellow dot-dash boxes highlight the misses under occlusions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 399, + 286, + 460 + ], + "blocks": [ + { + "bbox": [ + 50, + 399, + 286, + 460 + ], + "lines": [ + { + "bbox": [ + 50, + 399, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 50, + 399, + 286, + 460 + ], + "type": "table", + "html": "
MethodTranslating CameraRotating CameraFPS
Traj↓Velocity↓Traj↓Velocity↓
BEV+DPVO2.0240.1543.6220.1195.7
GLAMR [58]1.7450.1051.7590.0551.5
TRACE1.4330.0821.0380.04521.2
", + "image_path": "e82eb3e12ca5f1bcf3e0316cdfef1f92e1ea8ab384519fabe725c61e0b791a0f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 63, + 497, + 269, + 586 + ], + "blocks": [ + { + "bbox": [ + 47, + 461, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 47, + 461, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 287, + 483 + ], + "type": "text", + "content": "Table 1. Errors in estimated 3D human trajectory in world coordinates (in " + }, + { + "bbox": [ + 47, + 461, + 287, + 483 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 461, + 287, + 483 + ], + "type": "text", + "content": " ) and runtime efficiency on the DynaCam dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 497, + 269, + 586 + ], + "lines": [ + { + "bbox": [ + 63, + 497, + 269, + 586 + ], + "spans": [ + { + "bbox": [ + 63, + 497, + 269, + 586 + ], + "type": "table", + "html": "
MethodsIDs↓MOTA↑IDF1↑HOTA↑
Trackformer [36]4324.962.753.2
Tracktor [4]5351.570.950.3
FlowPose [51]4921.467.141.8
T3DP [39]3862.179.159.2
PHALP [40]2266.281.459.4
ByteTrack [63]1573.384.663.5
TRACE086.993.465.3
", + "image_path": "2ede17ea21fef5e7ef2f22646a9c005664e90427d7fdf0adcff1e00d4f5e1a8b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "world motion map to represent the global trajectory in world coordinates, which implies the camera motion. Therefore, we evaluate the global trajectory error, instead of the camera pose, in Tab. 3. First, we evaluate global 3D trajectory estimation on DynaCam. We compare TRACE with two baseline solutions. The first one uses BEV to estimate the subjects in camera coordinates and DPVO [48], a SLAM" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 399, + 533, + 468 + ], + "blocks": [ + { + "bbox": [ + 47, + 589, + 287, + 611 + ], + "lines": [ + { + "bbox": [ + 47, + 589, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 287, + 611 + ], + "type": "text", + "content": "Table 2. Tracking results on MuPoTS-3D [35]. Results of the compared methods [4,36,39,40,51] are from [40]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 399, + 533, + 468 + ], + "lines": [ + { + "bbox": [ + 317, + 399, + 533, + 468 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 533, + 468 + ], + "type": "table", + "html": "
MethodsIDs↓MOTA↑IDF1↑HOTA↑
PHALP [40]597.993.774.2
YOLOX+ByteTrack397.398.273.1
BEV+ByteTrack3793.679.159.3
TRACE w/o MO495.897.372.7
TRACE199.399.774.7
", + "image_path": "6f3cac83016639bb3231781c103e6ac5dae99eb2a380eb8f8e206423fed80e00.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 335, + 499, + 515, + 602 + ], + "blocks": [ + { + "bbox": [ + 306, + 472, + 545, + 494 + ], + "lines": [ + { + "bbox": [ + 306, + 472, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 545, + 494 + ], + "type": "text", + "content": "Table 3. Tracking results on Dyna3DPW [50]. w/o MO means ablating the estimated 3D motion offsets for tracking." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 499, + 515, + 602 + ], + "lines": [ + { + "bbox": [ + 335, + 499, + 515, + 602 + ], + "spans": [ + { + "bbox": [ + 335, + 499, + 515, + 602 + ], + "type": "table", + "html": "
MethodPAMPJPE↓MPJPE↓PVE↓
HybrIK [27]45.074.186.5
METRO [30]47.977.188.2
GLAMR [58]50.7--
CLIFF [28]43.069.081.2
D&D [26]42.773.788.6
ROMP [44]47.376.793.4
BEV [45]46.978.592.3
TRACE37.879.197.3
", + "image_path": "735f23b597df6aabf78ea09dfe56b21bc0bde3e6e5abf56a3dce28d9434d4f87.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 604, + 545, + 625 + ], + "lines": [ + { + "bbox": [ + 306, + 604, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 545, + 625 + ], + "type": "text", + "content": "Table 4. Comparisons with SOTA methods for 3D human pose and shape estimation on the 3DPW test set." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 715 + ], + "type": "text", + "content": "method, to estimate the camera and its motion; we call this BEV+DPVO. As shown in Tab. 1, TRACE significantly outperforms BEV+DPVO in the accuracy of global 3D trajectory estimation. The moving people in the scene make it hard for a SLAM method to extract stable corresponding keypoints. Additionally, our synthetic camera motions" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8862" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "differ from real camera motions and this may hurt DPVO's performance. A more direct comparison is to GLAMR [58]. TRACE outperforms GLAMR in both accuracy and efficiency. In Fig. 3(c), we also perform visual comparisons with GLAMR on DynaCam. These results demonstrate the benefit of estimating global human trajectory using a holistic 5D representation. We provide more results in the supplemental video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 529 + ], + "type": "text", + "content": "Multi-subject tracking. To evaluate the performance of tracking subjects with dynamic cameras in real-world scenes, we compare TRACE with recent methods on Dyna3DPW. PHALP [40] is a recent (SOTA) method that uses 3D cues and appearance to track people using SMPL. YOLOX+ByteTrack [17, 63] is a recently proposed and popular tracking-by-detection solution. These methods are designed to track all the people in a scene. Therefore, we process their results to avoid them being penalized for tracking unlabeled passers-by; 3DPW has annotations for at most 2 people in a scene but some scenes contain many people. We first obtain their tracking results using their official code. We then select the tracking results that achieve maximum IoU with the labeled ground truth subjects; we use these tracks for evaluation. Note that, for a fair comparison with ByteTrack, TRACE runs in an on-line mode, without optimizing the past results. As shown in Tab. 3, TRACE outperforms PHALP, the previous 3D-representation-based method [40], and tracking-by-detection [17, 63] methods on Dyna3DPW. To evaluate the tracking robustness under long-term occlusion, we evaluate TRACE on MuPoTS-3D. Results of previous SOTA methods [4, 36, 39, 40, 51] are from [40]. Again, for a fair comparison, we filter out the unlabeled people in the ByteTrack results. As shown in Tab. 2, TRACE significantly outperforms previous SOTA methods. In particular, TRACE significantly reduces ID switches under long-term occlusions. These results illustrate the effectiveness and robustness of the proposed method for in-the-wild videos. The qualitative comparisons are presented in Fig. 3 and the supplemental video." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": "3D human pose and shape estimation. Finally, we evaluate 3D human regression performance in DC-video using the 3DPW test set. Because 3DPW does not provide ground-truth 3D translation in world coordinates, we evaluate root-relative 3D pose. We compare TRACE with the recent one-stage [44,45] and multi-stage [26-28,30,58] methods. Tab. 4 shows that TRACE significantly outperforms the previous method by " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": " in terms of PMPJPE. These results demonstrate that learning a holistic 5D representation in an end-to-end manner is powerful. We adopt a fixed field of view " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{FOV} = 50^{\\circ})" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": " to define our camera space, which is very different from that of 3DPW video. This results in a misalignment between the predicted 3D orientation of the body mesh and the ground truth, resulting in higher MPJPE and MVE. This illustrates the importance of also estimating" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the camera so that our estimated global coordinate system is consistent with the real world. This is future work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 106, + 406, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 106, + 406, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 106, + 406, + 118 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 123, + 545, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 195 + ], + "type": "text", + "content": "Temporal 5D representation v.s. image-level 3D representation. We go beyond BEV's image-level 3D representation and build a temporal 5D representation. As shown in Tab. 1, 3, and 4, TRACE outperforms BEV or the multistage solutions using BEV on most metrics. This demonstrates the value of learning a holistic 5D representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "text", + "content": "3D Motion Offset map. We also evaluate the effect of using predicted 3D motion offsets for tracking. As shown in Tab. 3, 3D motion offsets improve performance by " + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 197, + 547, + 244 + ], + "type": "text", + "content": " in terms of MOTA, IDF1, and HOTA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 260, + 383, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 383, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 383, + 272 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 279, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 386 + ], + "type": "text", + "content": "Human pose and shape estimation is not an end to itself. Rather, estimating the 3D human in motion is useful for many tasks from behavior analysis to computer graphics. However, to be useful, it is important to know the motion of humans with respect to the 3D scene and other people. This means that HPS methods must estimate humans in a global coordinate system and provide consistent tracks of people across time. For generality, they also need to be able to do this from arbitrary moving cameras." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 555 + ], + "type": "text", + "content": "To tackle these challenging problems, we propose a novel 5D representation and a new neural architecture that reasons about people in 5D; that is, their 3D position, temporal trajectory, and identity. Moving to a 5D representation enables our method, TRACE, to take a holistic view of the video, processing full frames and incorporating temporal features. The core innovation of TRACE lies in its novel temporal representation in the form of new \"maps\" that represent the motion of people across time in the camera and global coordinates. These allow TRACE to be trained end-to-end, thus exploiting rich information from the video to solve the task. TRACE is the first such single-shot method for 3D HPS estimation from video and it achieves SOTA results on common benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": "Future work should look at explicitly estimating the camera, using training data like BEDLAM [6], which contains complex human motion, 3D scenes, and camera motions. We believe that camera motion and human motion provide complimentary information that can be used to recover human motion in world coordinates with metric accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 635, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 682 + ], + "type": "text", + "content": "Acknowledgements: This work was partially supported by the National Key R&D Program of China under Grant (No. 2020AAA0103800) and Beijing Nova Program (No. 20220484063)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 544, + 713 + ], + "type": "text", + "content": "MJB Disclosure: https://files.is.tue.mpg.de/black/CoI_CVPR_2023.txt" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8863" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 246, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 246, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 246, + 102 + ], + "type": "text", + "content": "[1] Bilibili. https://www.bilibili.com/.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 228, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 228, + 114 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 228, + 114 + ], + "type": "text", + "content": "[2] Pexels. https://www.pexels.com.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 116, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 287, + 159 + ], + "type": "text", + "content": "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 161, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 287, + 193 + ], + "type": "text", + "content": "[4] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, pages 941-951, 2019. 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 237 + ], + "type": "text", + "content": "[5] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, pages 1-10, 2008. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 240, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 287, + 283 + ], + "type": "text", + "content": "[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of 3d human bodies exhibiting detailed lifelike animated motion. In CVPR, 2023. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "type": "text", + "content": "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, pages 561-578, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "text", + "content": "[8] Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv:1905.00737, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "type": "text", + "content": "[9] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, pages 5386-5395, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "text", + "content": "[10] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, pages 436-454, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 455, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 508 + ], + "type": "text", + "content": "[11] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv:2003.09003, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[12] Andreas Doering, Di Chen, Shanshan Zhang, Bernt Schiele, and Juergen Gall. Posetrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In CVPR, pages 20963-20972, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "text", + "content": "[13] Mahsa Ehsanpour, Fatemeh Saleh, Silvio Savarese, Ian Reid, and Hamid Rezatofighi. Jrdb-act: A large-scale dataset for spatio-temporal action, social group and activity detection. In CVPR, pages 20983–20992, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "text", + "content": "[14] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. A mobile vision system for robust multi-person tracking. In CVPR, pages 1-8, 2008. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[15] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoşa Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, pages 10849-10859, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[17] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 162 + ], + "type": "text", + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 207 + ], + "type": "text", + "content": "[19] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments. TPAMI, 36(7):1325-1339, 2013. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "text", + "content": "[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In ECCV, 2020. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 243, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 275 + ], + "type": "text", + "content": "[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 277, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 308 + ], + "type": "text", + "content": "[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. VIBE: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 311, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 545, + 354 + ], + "type": "text", + "content": "[23] Muhammed Kocabas, Chun-Hao P. Huang, Joachim Tesch, Lea Müller, Otmar Hilliges, and Michael J. Black. SPEC: Seeing people in the wild with an estimated camera. In ICCV, pages 11035-11045, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 398 + ], + "type": "text", + "content": "[24] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 432 + ], + "type": "text", + "content": "[25] Nikos Kolotouros, Georgios Pavlakos, Dinesh Jayaraman, and Kostas Daniilidis. Probabilistic modeling for human mesh recovery. In CVPR, pages 11605-11614, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 434, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 465 + ], + "type": "text", + "content": "[26] Jiefeng Li, Siyuan Bian, Chao Xu, Gang Liu, Gang Yu, and Cewu Lu. D&d: Learning human dynamics from dynamic camera. In ECCV, 2022. 3, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[27] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In CVPR, pages 3383-3393, 2021. 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "text", + "content": "[28] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, pages 590-606, 2022. 3, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "text", + "content": "[29] Tingting Liao, Xiaomei Zhang, Yuliang Xiu, Hongwei Yi, Xudong Liu, Guo-Jun Qi, Yong Zhang, Xuan Wang, Xi-angyu Zhu, and Zhen Lei. High-Fidelity Clothed Avatar Reconstruction from a Single Image. In CVPR, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 602, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 634 + ], + "type": "text", + "content": "[30] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "text", + "content": "[31] Miao Liu, Dexin Yang, Yan Zhang, Zhaopeng Cui, James M Rehg, and Siyu Tang. 4d human body capture from egocentric video via 3d scene grounding. In 3DV, pages 930-939, 2021. 2, 3, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):1-16, 2015. 3, 4, 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8864" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[33] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, pages 548-578, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[34] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3D human pose estimation in the wild using improved cnn supervision. In 3DV, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[35] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[36] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 314 + ], + "type": "text", + "content": "[37] Gyeongsik Moon and Kyoung Mu Lee. Pose2Pose: 3d positional pose-guided 3d rotational pose prediction for expressive 3d human pose and mesh estimation. arXiv, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 315, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 287, + 349 + ], + "type": "text", + "content": "[38] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. TexturePose: Supervising human mesh estimation with texture consistency. In ICCV, pages 803-812, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[39] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. In NeurIPS, 2021. 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 425 + ], + "type": "text", + "content": "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3D appearance, location & pose. In CVPR, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "text", + "content": "[41] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV, pages 17-35, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 504 + ], + "type": "text", + "content": "[42] Weijian Ruan, Wu Liu, Qian Bao, Jun Chen, Yuhao Cheng, and Tao Mei. Pointet: pose-guided ovonic insight network for multi-person pose tracking. In MM, pages 284-292, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "text", + "content": "[43] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[44] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 1, 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[45] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[46] Yu Sun, Yun Ye, Wu Liu, Wenpeng Gao, YiLi Fu, and Tao Mei. Human mesh recovery from monocular images via a skeleton-disentangled representation. In ICCV, pages 5348-5357, 2019. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 690 + ], + "type": "text", + "content": "[47] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[48] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. arXiv preprint arXiv:2208.04726, 2022. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[49] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In CVPR, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[50] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[51] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose Flow: Efficient online pose tracking. In BMVC, 2018. 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[52] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 219, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 252 + ], + "type": "text", + "content": "[53] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 253, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 286 + ], + "type": "text", + "content": "[54] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. arXiv preprint arXiv:2204.12484, 2022. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "text", + "content": "[55] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "text", + "content": "[56] Hongwei Yi, Chun-Hao P. Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J. Black. MIME: Human-aware 3D scene generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[57] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[58] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. GLAMR: Global occlusion-aware human mesh recovery with dynamic cameras. In CVPR, 2022. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "type": "text", + "content": "[59] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3D pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "type": "text", + "content": "[60] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. PyMAF: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[61] Shanshan Zhang, Rodrigo Benenson, and Bernt Schiele. Citypersons: A diverse dataset for pedestrian detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[62] Weiyu Zhang, Menglong Zhu, and Konstantinos G Derpanis. From actemes to action: A strongly-supervised representation for detailed action understanding. In ICCV, pages 2248-2255, 2013. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[63] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In ECCV, pages 1-21, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 715 + ], + "type": "text", + "content": "[64] Yang Zheng, Ruizhi Shao, Yuxiang Zhang, Tao Yu, Zerong Zheng, Qionghai Dai, and Yebin Liu. Deepmulticap: Performance capture of multiple characters using sparse multiview cameras. In ICCV, pages 6239-6249, 2021. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8865" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 118 + ], + "type": "text", + "content": "[65] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "8866" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_content_list.json b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c916b6f3662069d9994602348b58ea981ad9967e --- /dev/null +++ b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_content_list.json @@ -0,0 +1,1554 @@ +[ + { + "type": "text", + "text": "TTA-COPE: Test-Time Adaptation for Category-Level Object Pose Estimation", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taeyeop Lee $^{1}$ Jonathan Tremblay $^{2}$ Valts Blukis $^{2}$ Bowen Wen $^{2}$ Byeong-Uk Lee $^{1}$ Inkyu Shin $^{1}$ Stan Birchfield $^{2}$ In So Kweon $^{1}$ Kuk-Jin Yoon $^{1}$ KAIST ${}^{2}$ NVIDIA", + "bbox": [ + 130, + 179, + 836, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test-time adaptation methods have been gaining attention recently as a practical solution for addressing source-to-target domain gaps by gradually updating the model without requiring labels on the target data. In this paper, we propose a method of test-time adaptation for category-level object pose estimation called TTA-COPE. We design a pose ensemble approach with a self-training loss using pose-aware confidence. Unlike previous unsupervised domain adaptation methods for category-level object pose estimation, our approach processes the test data in a sequential, online manner, and it does not require access to the source domain at runtime. Extensive experimental results demonstrate that the proposed pose ensemble and the self-training loss improve category-level object pose performance during test time under both semi-supervised and unsupervised settings.", + "bbox": [ + 75, + 301, + 473, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 571, + 209, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object pose estimation is a crucial problem in computer vision and robotics. Advanced methods that focus on diverse variations of object 6D pose estimation have been introduced, such as known 3D objects (instance-level) [28, 38], category-level [18, 36, 43], few-shot [52], and zero-shot pose estimation [13, 47]. These techniques are useful for downstream applications requiring an online operation, such as robotic manipulation [6, 25, 48] and augmented reality [23, 24, 32]. Our paper focuses on the category-level object pose estimation problem since it is more broadly applicable than the instance-level problem.", + "bbox": [ + 75, + 598, + 468, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many works on category-level object pose estimation [2, 3, 17, 18, 36, 43, 44] have been proposed recently. These approaches estimate multiple classes of object pose more efficiently in a single network compared to the instance-level object pose estimation methods [27, 38, 41, 49-51], which depend on known 3D shape knowledge and the size of the objects. Notably, Wang et al. [43] introduced a novel representation called Normalized Object Coordinate Space (NOCS) to align various object instances within each", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ef1e6bd5b3704e36771617886e8f39f7d9331b1e70e077523e0cc703f2436cf.jpg", + "image_caption": [ + "No Adaptation" + ], + "image_footnote": [], + "bbox": [ + 532, + 287, + 710, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3dbda982517c5473aedb4167fbacb6a9599b1521b7f18ec04a738358870136e.jpg", + "image_caption": [ + "Test-Time Adaptation" + ], + "image_footnote": [], + "bbox": [ + 715, + 287, + 890, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc3ea698be5ab3f71879da2fdc3a2beaa84e91fdf69ce9c75577a69518efe261.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 393, + 710, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/980c223f2f6155f6f0171b6eaa165cd5e91f23687d74a43d4a05297da5750bde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 393, + 890, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbe5bf23a24cc471ea48dd2491de9cb4b7e5fe2518cb029ef6983dea926ae2f2.jpg", + "image_caption": [ + "Figure 1. We propose a Test-Time Adaptation for Category-level Object Pose Estimation framework (TTA-COPE) that automatically improves the network in an online manner without labeled target data. As new image frames are processed, our method fine-tunes the network using the unlabeled data and simultaneously applies the network to perform pose estimation via inference. This approach successfully handles domain shifts compared with no adaptation, as seen here." + ], + "image_footnote": [], + "bbox": [ + 534, + 500, + 710, + 601 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01f64bb8e9793d621a001de91fa2b851caa9e0dc20e4981448cab24545eebcdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 500, + 890, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "category in a canonical 3D space. The strengths of the NOCS representation have led to its adoption by follow-up work [3, 17, 36].", + "bbox": [ + 496, + 734, + 890, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to obtain accurate category-level object pose methods in unseen real-world scenarios, it is desirable to fine-tune the models in the new environment with labeled target data. The model that is not fine-tuned on the target domain distribution will almost certainly exhibit lower performance than the fine-tuned model [37]. However, annotating 6D poses of objects in the target environment is an expensive process [1, 39, 43, 45] that we seek to avoid.", + "bbox": [ + 496, + 780, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21285", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/3ca988fda7aaed4e290bb6ae698a22957cefff6a35bf7726f7c9c213a2cf8456.jpg", + "table_caption": [ + "Table 1. Comparison with prior unsupervised works for category-level object pose estimation. Our unsupervised method trains models without 2D or 3D labels of target data, similar to Self-DPDN [16]. Unlike previous methods, our proposed approach updates the model online without offline batch processing. Moreover, we do not use the source data during test time (source-free) because it is impractical to train on a large amount of source data every iteration. There also may be privacy or legal constraints to access source data [21]." + ], + "table_footnote": [], + "table_body": "
MethodUnsupervisedTest-time Adaptation
Target 3DTarget 2DSource-FreeOnline Adaptation
SupervisedXXXX
SSC-6D [29]✓XXX
RePoNet [5]✓XXX
UDA-COPE [15]✓X✓X
Self-DPDN [16]✓✓XX
Ours✓✓✓✓
", + "bbox": [ + 81, + 223, + 465, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to annotating in 2D space, labeling in 3D space requires specific knowledge about geometry [7] from the annotator and is much more laborious, time-consuming, and error-prone due to the complex nature of $SE(3)$ space. Therefore, it is usually challenging to annotate real-world data with 3D annotations for fine-tuning.", + "bbox": [ + 75, + 353, + 468, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to solve the aforementioned problem of annotating object pose data in the real world, several recent methods [5, 15, 16] propose unsupervised domain adaptation (UDA) that aims to train the network without utilizing the ground truth of target pose labels. Although they show promising results using UDA techniques, these approaches still do not meet some of the requirements for online applications. For example, when a robot encounters a new environment, it is desirable to adapt the scene online manner while estimating object poses rather than waiting for enough data to be collected in the novel scene to train the model offline.", + "bbox": [ + 75, + 445, + 468, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This problem definition of online fine-tuning is more practical for real applications, where we desire to update the model instantly when new data becomes available for fast domain adaptation. This setting is known as test-time adaptation (TTA) [42]. For TTA, the requirements are as follows: 1) labeled source data should not be accessed at test time, 2) adaptation should be online (rather than offline batch processing), and 3) the method should be fully unsupervised, without using 2D or 3D target labels during online fine-tuning. Since we do not have access to labeled source data (source-free) at test time this problem is more challenging than existing unsupervised category-level object pose methods [5,15,16,29]. Table 1 summarizes the difference between our problem definition and existing methods, showing that test-time adaptation for category-level object pose estimation remains an open problem.", + "bbox": [ + 75, + 627, + 467, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Test-time Adaptation for Category-level Object Pose Estimation (TTA-COPE) to", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "handle domain shifts without any target domain annotations (see Fig. 1). Prior works on general test-time adaptation [42,46] propose self-training to minimize entropy loss. TENT [42] has shown improvement in 2D classification and segmentation tasks. We show, however, that simply extending TENT for the category-level object pose estimation is not effective. Another self-training strategy is the teacher-student framework [35] with pseudo labels. However, since pseudo labels are created without any noise filtering, naive pseudo labels may be unreliable and cause convergence to a suboptimal model.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this problem, we design a novel pose ensemble method to perform test-time adaptation for category-level object pose estimation by extending the pose-aware filtering of UDA-COPE [15]. The proposed method uses an ensemble of teacher-student predictions based on pose-aware confidence, which is used both for generating pseudo labels and inference. Also, the pose ensemble helps to train models with additional self-training loss to reduce the domain shift for category-level pose estimation by using pose-aware confidence. We demonstrate the advantages of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings. We show that our TTA-COPE framework achieves state-of-the-art performance compared to strong TTA baselines.", + "bbox": [ + 496, + 257, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of our work are as follows:", + "bbox": [ + 500, + 469, + 890, + 497 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE), which handles domain shifts without labeling target data and without accessing source data during test time.", + "- We introduce a pose ensemble with self-training loss that utilizes the teacher-student predictions to generate robust pseudo labels and estimates accurate poses for inference.", + "- We evaluate our framework with experimental comparisons against strong test-time baselines and state-of-the-art methods under both semi-supervised and unsupervised settings." + ], + "bbox": [ + 519, + 508, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 700, + 648, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Supervised Methods", + "text_level": 1, + "bbox": [ + 500, + 726, + 692, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully supervised learning methods for category-level object pose estimation [2,3,17,36,43,44] train their models using labeled source data (e.g., synthetic) and target data (e.g., real). Most category-level 6D object pose and size estimation approaches [3,14,36,43,44] use the dense correspondence via Normalized Object Coordinate Space (NOCS) representation as a common way to estimate pose and size. These correspondence-based methods initially estimate the NOCS map from RGB or depth images. Afterwards, the Umayama algorithm [40] with RANSAC is used to estimate", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21286", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "optimal poses and object sizes by minimizing distances between depth and estimated NOCS map.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some methods use category priors [3,36,44] as the representative 3D shape per class, jointly reconstructing the full 3D shape and estimating the NOCS map from the full shape. Results show that this prior category helps improve the accuracy of the NOCS map and enhance the pose estimation performance. Other methods directly regress the pose or jointly utilize the correspondence representations [2,11,17].", + "bbox": [ + 75, + 121, + 467, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Unsupervised Methods", + "text_level": 1, + "bbox": [ + 76, + 237, + 290, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that annotating the 6D object pose and 3D size labels in the real world is expensive, time-consuming, and laborious, RePoNet [5] and SSC-6D [29] propose semi-supervised approaches that reconstruct the entire shape and use differential rendering [19, 20] techniques for self-training signals. These are category-specific methods and utilize multiple models, as many models as the number of categories is required. UDA-COPE [15] proposes an unsupervised domain adaptation method to mitigate the domain shift from the source to the target domain with a single model to estimate all categories efficiently.", + "bbox": [ + 75, + 261, + 467, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although these works show reasonable performance without using pose labels, there is still a limitation in relying on 2D ground truth information (segmentation or bounding boxes) to train the segmentation network or pose network. Self-DPDN [16] shows a fully unsupervised method using inter/intra-consistency as a reconstructed shape in a self-supervised objective but requires a supervised loss for the source domain in the unsupervised learning process. Additionally, the self-training loss utilizes a full 3D shape and requires an additional reconstruction module for 3D shape. Most methods, except for UDA-COPE [15], jointly use the supervised loss using source label data during unsupervised learning to relax the unstable training. Furthermore, all the aforementioned unsupervised methods train their model in an offline manner and are unsuitable for online applications [24, 25].", + "bbox": [ + 75, + 428, + 467, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Test-time Adaptation Methods", + "text_level": 1, + "bbox": [ + 76, + 680, + 346, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA) aims to enable the online adaptation of a pretrained model to the target domain without access to the source domain (source-free) [26,34,42,46]. The source data is commonly inaccessible during inference time because it is inefficient to train on a huge amount of source data every iteration. There may also be privacy or legal constraints to accessing source data [21], thus TTA is a more difficult but practical task than UDA. From the point of view of real-world applications, it is necessary to adapt to the new scene in an online way. Accordingly, test-time adaptation is necessary for the success of practical, real-world computer vision applications. Wang et al. propose Test entropy minimization (TENT) [42], which trains a net-", + "bbox": [ + 75, + 703, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work using a labeled source dataset, and adapts it to the unlabeled target dataset by updating the network parameters in batch norm layers using entropy loss. CoTTA [46] proposes a continual test-time adaptation method on the 2D classification and semantic segmentation tasks, and it effectively reduces the error accumulations while continually changing target data. Not limited to the 2D tasks, test-time adaptation methods have been applied to other tasks such as 3D segmentation [31, 33] and robot manipulation [22].", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TTA-COPE", + "text_level": 1, + "bbox": [ + 500, + 239, + 624, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an RGB-D image, our approach aims to estimate the 6D pose $T \\in \\mathrm{SE}(3)$ and size $s \\in \\mathbb{R}_+^3$ of each object. The object pose $T$ is defined as the rigid transformation $[R|t]$ , with rotation $R \\in \\mathrm{SO}(3)$ and translation $t \\in \\mathbb{R}^3$ .", + "bbox": [ + 496, + 263, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method consists of a two-stage learning scheme using source and target data, respectively. In the first stage, we train the network using labeled (synthetic) source data in a supervised manner (Sec. 3.1). This step is the same as supervised methods and results in a pretrained model. We then utilize this pretrained model for test-time adaptation (TTA) using unlabeled target data without accessing the source data. To the best of our knowledge, we are the first to propose test-time adaptation for category-level object pose estimation. Therefore, we study and explore how TTA baselines (Sec. 3.2) might apply to category-level object pose estimation. Finally, our proposed method is presented and explained (Sec. 3.3).", + "bbox": [ + 496, + 325, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pretraining with Source Data", + "text_level": 1, + "bbox": [ + 500, + 530, + 761, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce an overview of our model and then describe how to train the model using labeled source data. We use UDA-COPE [15] network as a base model, which we supplement with batch normalization (BN) [10] in the 2D network to utilize BN updates for test-time adaptation. Fig. 2 shows an overview of our network, consisting of three steps to estimate the pose similar to correspondence methods [3,16,36]. First, we detect the bounding area of the object and segment the object surface using the off-the-shelf segmentation model from a single RGB image [8]. In the second stage, given a cropped 2D object image $I$ and segmented 3D object point cloud $D$ , our model estimates the correspondence as a NOCS map $N$ . We leverage the NOCS representation [43] to align diverse object instances within each class in a unified 3D space. In the final stage, our proposed pose ensemble method estimates the object pose $T$ . It simultaneously utilizes the predictions of the student and teacher model by calculating the pose confidence using the number of filtered points.", + "bbox": [ + 496, + 551, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The network is pretrained in a fully supervised manner using labeled source data. We minimize the predictions of the NOCS map $N$ using the cross-entropy loss $L_{CE}$ with ground truth labels $N^{GT}$ . We also use the consistency loss", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21287", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b540a372d764dd3d44e2b4d4fc56f261f2aa6e1b5fb0c28cbeb70bb6acd9374.jpg", + "image_caption": [ + "Figure 2. Overview of our Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE) framework. Our method consists of three steps to estimate category-level object pose, similar to correspondence methods. 1) We detect and segment the object region using an off-the-shelf segmentation model from a single RGB image. 2) Given a cropped 2D object image $I$ and segmented 3D object point cloud $D$ , our model estimates the correspondence points as a NOCS map $N$ . 3) Our proposed pose ensemble estimates the object pose $T$ . It simultaneously utilizes the predictions of the student $\\theta^S$ and teacher model $\\theta^T$ by calculating the pose-aware confidence using the number of inlier points. We operate point filtering for each student and teacher model and select a confident pose $T$ with higher confidence (i.e., more inlier points) for which the pose estimation is more accurate." + ], + "image_footnote": [], + "bbox": [ + 140, + 84, + 826, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{C}$ with 2D and 3D augmentation to make the network robust to noise in the input [15]. The supervised loss is formulated as:", + "bbox": [ + 76, + 446, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s u p} = \\lambda_ {C E} L _ {C E} (N, N ^ {G T}) + \\lambda_ {C} L _ {C} (N _ {a u g}, N), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 500, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{aug}$ is the estimated NOCS map from augmented 2D or 3D input, and $\\lambda_{CE},\\lambda_C\\in \\mathbb{R}_+$ are weights.", + "bbox": [ + 76, + 523, + 468, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Test-time Adaptation with Target Data", + "text_level": 1, + "bbox": [ + 76, + 561, + 410, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The model is updated every iteration while simultaneously estimating the object pose of the current scene. (Note that this is different from previous methods that update the model by running multiple epochs over the target data [16].)", + "bbox": [ + 76, + 584, + 468, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have a design choice for the objective loss when updating the model during test time. One of the widely used test-time objective losses is the entropy loss proposed by TENT [42],", + "bbox": [ + 75, + 645, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e n t} = - \\sum p \\left(x _ {t}\\right) \\log p \\left(x _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 710, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(x_{t})$ is the probability of predictions $\\theta^S (x_t)$ from target data $x_{t}$ . This simple objective loss encourages sharp distributions by assigning the most probability.", + "bbox": [ + 75, + 741, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another common approach is utilizing pseudo labels from a teacher-student framework [35] with momentum update. The teacher model $\\theta^T$ generates pseudo ground truth $\\hat{y} = \\theta^T(x_t)$ , where $x_t = (I, D)$ . The student model $\\theta^S$ then uses the predictions of the teacher as the ground truth signals (pseudo labels) with cross-entropy loss,", + "bbox": [ + 76, + 787, + 468, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p l} = L _ {C E} \\left(\\theta^ {S} \\left(x _ {t}\\right), \\hat {y}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After updating the student model $\\theta_{i}^{S}\\rightarrow \\theta_{i + 1}^{S}$ by minimizing Eq. (3), the teacher model $\\theta_{i + 1}^{T}$ is updated by momentum update,", + "bbox": [ + 498, + 446, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i + 1} ^ {T} \\leftarrow \\gamma \\theta_ {i} ^ {T} + (1 - \\gamma) \\theta_ {i + 1} ^ {S}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 491, + 890, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ stands for the time stamp of the iteration and $\\gamma$ is the momentum smoothing factor. The separation of the model structure mitigates some error accumulation by the pseudo label and momentum update.", + "bbox": [ + 498, + 517, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although these objective strategies from previous work potentially provide self-training signals for test-time adaptation, they become brittle when directly applied to a category-level object pose estimation task. The pseudo labels $\\hat{y}$ are still inherently exposed to noisy inputs and predictions, leading to errors in the student model $\\theta^S$ . Since $\\theta^T$ cannot guarantee clean pseudo labels for the student, we need the ability to filter noise in the input or predictions and to yield more reliable pseudo labels to the $\\theta^S$ .", + "bbox": [ + 498, + 579, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pose Ensemble", + "text_level": 1, + "bbox": [ + 500, + 724, + 650, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve this problem, we propose to extend the pseudolabel filtering proposed by UDA-COPE [15] with pose ensemble processing. We observe that the UDA-COPE framework and pose filtering, while showing notable results, is inefficient for two reasons. First, the teacher model is only used for generating pseudo labels and not for inference. Second, the student model is also only used for inference and has less influence in generating pseudo labels.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose instead to jointly utilize the teacher and student models for both generating pseudo labels and for infer", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21288", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ence, which we show in the experimental results produces a noticeable improvement. To this end, we design a pose ensemble module to simultaneously use teacher and student model predictions. Fig. 2 shows an overview of the pose ensemble with point filtering.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The point filtering has three steps. Unlike UDA-COPE [15], we repeat these steps for both the student and teacher: 1) We initially estimate $T^S$ using the Umeyama algorithm $\\theta_p$ within the region of object points $D$ and estimated NOCS map $N^S$ . 2) We transform observed depth to normalized object coordinate space $\\widehat{D}^S$ by multiplying $(T^S)^{-1}$ and $D$ . 3) We compute each matching point distance between $N^S$ and $\\widehat{D}$ , removing outliers that exceed a certain threshold $\\rho$ . This whole process of point-filtering is then repeated using $N^T$ to estimate $T^T$ . This process is represented as follows, where $N^{S,T}$ is either $N^S$ or $N^T$ , and similarly for $T^{S,T}$ :", + "bbox": [ + 75, + 167, + 470, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T ^ {S, T} = \\theta_ {p} (D, N ^ {S, T}), \\\\ \\widehat {D} ^ {S, T} = \\left(T ^ {S, T}\\right) ^ {- 1} D, \\tag {5} \\\\ e _ {j} ^ {S, T} = \\left\\{ \\begin{array}{l l} i n l i e r & \\text {i f} \\| \\widehat {D} _ {j} ^ {S, T} \\\\ o u t l i e r & \\text {o t h e r w i s e .} \\end{array} \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 361, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j = 1,\\ldots ,X$ , and $X = |D| = |\\widehat{D} |$", + "bbox": [ + 76, + 458, + 357, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed pose ensemble aims to fuse two predictions $(N^S, N^T)$ to estimate the optimal pose. Although this fusion could be conducted at the input level before obtaining the $T$ from $\\theta_p$ , we found that our proposal of fusing at the output level after getting $T$ produces slightly better results (see the ablation study, Sec. 4.5). We first estimate each pose $(T^S, T^T)$ from each prediction $(N^S, N^T)$ using $\\theta_p$ . We then calculate the number of inliers $(inliers^S, inliers^T)$ from each pose using Eq. (5) to evaluate the confidence of the pose. We assume that the higher this confidence (i.e., the more inliers) is, the more accurate the pose estimation will be. Therefore:", + "bbox": [ + 75, + 476, + 468, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {o u t} = \\left\\{ \\begin{array}{l l} T ^ {S} & \\text {i f i n l i e r s} ^ {S} > \\text {i n l i e r s} ^ {T}, \\\\ T ^ {T} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 667, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We found that this simple pose-aware ensemble method is more effective for generating high-quality pseudo labels and inference than other ensemble methods (Sec. 4.5).", + "bbox": [ + 75, + 719, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another crucial part of our method is a self-training loss. The total loss that we use to update the student network combines this self-training $L_{d}$ loss and the pseudo label loss $L_{pl}$ :", + "bbox": [ + 75, + 765, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t t a} = \\lambda_ {d} L _ {d} + \\lambda_ {p l} L _ {p l}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 839, + 468, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first term is self-training loss $L_{d}$ , which helps learn the distribution of the target domain and reduce the domain gap", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "through $T_{out}$ , which is obtained from the pose ensemble and observed $D$ . This loss is given by:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d} = L _ {C E} \\left(N _ {e} ^ {S}, \\operatorname {E n c} \\left(\\widehat {D} _ {e}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 132, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{e}^{S}$ is the inlier-only student NOCS map, $\\widehat{D}_{e}$ is the resulting inliner-only transformed depth point cloud in NOCS space, and $\\mathrm{Enc}(\\cdot)$ is the one-hot encoding to enable cross-entropy loss.", + "bbox": [ + 496, + 162, + 890, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second term", + "bbox": [ + 519, + 226, + 633, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p l} = L _ {C E} \\left(N _ {e} ^ {S}, N _ {e} ^ {T}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 255, + 890, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is the pseudo label loss from UDA-COPE [15], where $N_{e}^{T}$ is the inlier-only teacher NOCS map. Since $N^{S} = \\theta^{S}(x_{t})$ , and $\\hat{y} = \\theta^T (x_t)$ , $L_{pl}$ in Eq. (9) is the same as the loss in Eq. (3), except that here only inliers are considered.", + "bbox": [ + 496, + 286, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 363, + 630, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 500, + 388, + 594, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize two widely used category-level pose estimation datasets as the source and target domain, respectively. The source data is the Context-Aware MixEd Reality (CAMERA) dataset [43], generated by rendering and compositing synthetic objects into real scenes while considering the context. CAMERA consists of 275K RGB-D images as a training set with 1,085 object instances chosen from six categories: bottle, bowl, camera, can, laptop and mug. We use the REAL dataset [43] as the target domain. The target data consists of 4,300 real-world images of seven scenes for training and 2,750 real-world images of six scenes for evaluation. We refer to REAL evaluation set as REAL275. Our TTA methods do not use the target training set (4,300 images with seven scenes) and only use the evaluation set (2,750 images with six scenes, REAL275) for test-time adaptation, e.g., TENT [42] process evaluation set in a sequential, online manner. On the other hand, unsupervised methods such as [5, 16, 29] use the target training set for updating the models without any restrictions on how much and how long they use the information.", + "bbox": [ + 496, + 412, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 724, + 718, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Segmentation. We use Mask R-CNN [8] to obtain the object area in 2D image. We use the identical results of Mask R-CNN for a fair comparison with previous methods [3, 17, 36] for semi-supervised and unsupervised settings. For the semi-supervised setting, Mask R-CNN has trained on the source, and target domain supervised manner. For the unsupervised setting, Mask R-CNN is trained on only the source domain. The detected area of the image resizes to $192 \\times 192$ image resolution as the teacher or student model input.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21289", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bfca32a33947bdfdc633999d3afaaac04d47f4437d238a67464cb145cfdfc30b.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons with state-of-the-art methods on the REAL275 dataset." + ], + "table_footnote": [], + "table_body": "
MethodSupervisedUnsupervisedmAP (↑)
SourceTargetTargetOnlineIoU50IoU755° 2cm5° 5cm10° 2cm10° 5 cm
Source (Supervised)
Metric Scale [14]2D/3D54.68.42.25.410.125.0
SPD [36]2D/3D50.517.011.512.133.037.9
Source (Supervised) & Target (Supervised)
NOCS [43]2D/3D2D/3D47.29.47.210.013.825.2
SPD [36]2D/3D2D/3D68.527.019.521.643.554.0
DualPoseNet [17]2D/3D2D/3D—30.829.335.950.066.8
CR-Net [44]2D/3D2D/3D—33.227.834.347.260.8
SGPA [3]2D/3D2D/3D—37.135.939.661.370.7
Source (Supervised) & Target (Semi-Supervised)
SSC-6D [29]2D/3D2D3DX73.0—16.819.644.154.5
RePoNet [5]2D/3D2D3DX76.0—30.733.9—63.0
UDA-COPE [15]2D/3D2D3DX75.534.430.534.957.066.1
Self-DPDN [16]2D/3D2D3DX75.241.639.545.063.372.2
TTA-COPE (Ours)2D/3D2D3D✓78.743.533.338.164.375.1
Source (Supervised) & Target (Unsupervised)
Self-DPDN [16]2D/3D2D/3DX67.243.939.046.761.873.4
TTA-COPE (Ours)2D/3D2D/3D✓69.139.730.235.961.773.2
", + "bbox": [ + 91, + 109, + 877, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Student and Teacher. The student and teacher models have identical design in 2D and 3D branches from UDA-COPE [15]. We utilize the PSPNet [53] with ResNet34 [9] backbone for the 2D image feature extraction. For a 3D branch, we use the MinkowskiNet [4] and utilize sparse convolution operation with a $5\\mathrm{cm}$ voxel size. Our NOCS representation uses the classification with 32 bins [43] instead of direct regression. During the pretraining stage, we train our model on the source data for 50 epochs using the Adam optimizer by initializing the learning rate of 1e-4 with a batch size of 32. The learning rate was reduced by a ratio of 0.6 (at 15k iterations), 0.3 (at 30k iterations), 0.1 (at 45k iterations), and 0.01 (at 60k iterations). During test-time adaptation, our student model uses the same learning rate, and the teacher smoothly updates using momentum update with $\\gamma = 0.99$ . Given target data, we first update the model every iteration and then estimate the pose. We set $\\lambda_{\\mathrm{CE}} = 1.0$ , $\\lambda_{\\mathrm{C}} = 1\\mathrm{e - }6$ , $\\lambda_{\\mathrm{d}} = 1.0$ , $\\lambda_{\\mathrm{pl}} = 1.0$ for our experiments. The point filtering threshold $\\rho$ was set to 0.05 for all experiments.", + "bbox": [ + 75, + 444, + 472, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To evaluate the performance of 3D object detection and 6D pose estimation, we follow the previous pose and size evaluation metric from Wang et al. [43]. We report the mean average precision (mAP) at the $50\\%$ and $75\\%$ intersection over union (IoU) thresholds for 3D object detection. We also report mAP for 6D object pose evaluation w.r.t. rotation and translation errors, where e.g. the $5^{\\circ}5\\mathrm{cm}$ metric describes the percentage of pose predictions where the error is less than both $5^{\\circ}$ and $5\\mathrm{cm}$ and the same for other thresholds. We recalculated all the metrics with the improved code.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2e11649f2247962c8ad24264d39d9e6dd802a00ef7f1ef286566ef3fe266174.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons with TTA baselines for category-level object pose estimation on the REAL275 dataset." + ], + "table_footnote": [], + "table_body": "
MethodUnsupervisedIoU75mAP (↑)
Target 3DTarget 2D5° 2cm5° 5 cm
TENT [42] - Eq. (2)✓33.725.929.1
PL [35] - Eqs. (3)-(4)✓39.929.734.9
PL-F [15] - Eqs. (3)-(5)✓41.131.436.3
TTA-COPE (Ours)✓43.533.338.1
TENT [42] - Eq. (2)✓✓32.326.831.2
PL [35] - Eqs. (3)-(4)✓✓36.026.933.1
PL-F [15] - Eqs. (3)-(5)✓✓36.528.034.0
TTA-COPE (Ours)✓✓39.730.235.9
", + "bbox": [ + 503, + 481, + 890, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 500, + 616, + 790, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 summarizes quantitative results on the REAL275 dataset under different settings: 1) supervised only on source data, 2) supervised on source and target data, 3) supervised on source data and semi-supervised on target data, and 4) supervised on source data and unsupervised on target data. Not surprisingly, methods following 2) such as SPD [36] perform better than setting 1) since they have the benefit of accessing both source and target data. Recent unsupervised domain adaptation (UDA) methods [15, 16] show remarkable results compared to state-of-the-art supervised methods without using target 3D labels (semi-supervised setting). Our method shows state-of-the-art results in IoU metrics under the semi-supervised setting and even outperforms the recent supervised methods SGPA [3] by a large margin (6.4 mAP in $\\mathrm{IoU}_{75}$ ). Our TTA-COPE uses less time and data to train the target domain because of the advantage of test-time adaptation but achieves comparable", + "bbox": [ + 496, + 643, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4ca85968709e2232c726a0e2cdbf13f16ccbd641dc62e3d3f16a6ff4c3ae48f9.jpg", + "table_caption": [ + "Table 4. Ablation study on variants of pose ensemble methods and self-training loss under the semi-supervised setting." + ], + "table_footnote": [], + "table_body": "
MethodPose EnsemblemAP (↑)
InputOutputInferencePseudo LabelLd(8)IoU50IoU755° 2cm5° 5 cm10° 2cm10° 5 cm
LB76.237.529.134.662.173.2
(1)Argmax Match✓71.127.529.634.360.771.7
(2)Softmax Avg.✓78.740.432.137.364.074.7
(3)Softmax Max✓77.640.931.636.463.474.3
(4)Softmax Max✓77.541.132.137.164.374.6
(5)Inliers Max✓77.641.232.837.664.474.9
(6)Inliers Max✓✓78.242.932.437.864.174.7
(7)Argmax Match✓78.340.731.736.463.273.9
(8)Softmax Avg.✓78.641.131.536.263.674.5
(9)Softmax Max✓78.141.231.936.563.574.4
(10)Softmax Max✓78.641.231.836.363.574.4
(11)Inliers Max✓78.241.331.636.263.674.3
(12)Inliers Max✓✓78.743.032.637.063.573.9
OursInliers Max✓✓✓78.743.533.338.164.675.1
", + "bbox": [ + 83, + 107, + 879, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results to the SOTA method, Self-DPDN [16]. Ours takes about 31 minutes (Table 5) for TTA and is $58\\mathrm{x}$ faster than the Self-DPDN, which takes about 30 hours for training in the target domain, excluding inference time.", + "bbox": [ + 75, + 349, + 468, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with TTA baselines", + "text_level": 1, + "bbox": [ + 76, + 420, + 359, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 summarizes results for the different design test-time adaptation (TTA) baselines in semi-supervised and unsupervised settings. TENT [42] has been designed for general unsupervised settings, but it is not specially designed for the object pose estimation task and naturally observes overall poor performance compared to other TTA baselines. Mean teacher with pseudo labels (PL) [35] performs better than TENT [42] in semi-supervised and unsupervised settings, and we believe that training the model using pseudo labels with momentum update (Eqs. (3)-(4)) provides more stable training signals by reducing the error accumulation than entropy minimization (2). However, using pseudo labels without filtering (PL) provides unreliable labels as the ground truth, we found that pseudo label filtering (PL-F) [15] performs better than PL, which indicates that noise filtering (Eq. (5)) improves the results for the student as expected. Finally, our TTA-COPE achieves state-of-the-art (SOTA) performance among all TTA baselines under both semi-supervised and unsupervised settings.", + "bbox": [ + 73, + 443, + 470, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 739, + 230, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to evaluate the efficacy of the pose ensemble and self-training loss under a semi-supervised setting.", + "bbox": [ + 75, + 763, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input/Output-Level Pose Ensemble. As mentioned in the Sec. 3.3, we compare the different pose ensemble methods given predictions of teacher and student models $(N^T, N^S)$ and answer the following question: Which ensemble is the most effective for test-time adaptation? We categorize two ensemble techniques:", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Input-level ensemble that fuses two NOCS map predictions $(N^T, N^S)$ and makes an accurate single NOCS map $N$ to estimate poses using the Umayama algorithm $\\theta_p$ . For the input-level ensemble that fuses two predictions $(N^T, N^S)$ into one prediction $N$ , we compare three strategies: Argmax Match, Softmax Average, and Softmax Max operations [12, 33].", + "2) We ensemble output-level predictions $(N^T, N^S)$ to estimate each pose $(T^T, T^S)$ and choose the best pose predictions under specific criteria. Unlike the input ensemble, only the Softmax Max operation is valid for the output-level ensemble due to the nonlinearity of poses." + ], + "bbox": [ + 496, + 349, + 890, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 summarizes our ensemble ablation study results. Table 4-(1-3) shows that the Softmax Avg. and Softmax Max perform better than the Argmax Match operation among input-level pose ensembles. If we compare the same operation (Softmax Max) in the input-level, Table 4-(3), and output-level, Table 4-(4), the output-level ensemble shows slightly better performance since the output-level operation is directly related to the pose results. Our proposed pose ensemble, Table 4-(5), considers pose-aware confidence by choosing a higher number of inlier points and achieves the best performance in all pose ensembles in Table 4-(1-5).", + "bbox": [ + 496, + 534, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Use of Pose Ensemble. Our pose ensemble can be used for inference Table 4-(1-6, Ours) as well as generating pseudo labels Table 4-(7-12, Ours). The results show that our pose ensemble is helpful for both cases. In particular, applying our ensemble to inferences Table 4-(5) has been shown to be more effective than application to generating pseudo labels Table 4-(11) in pose metric. We also observe that our pose ensemble used for inference or pseudo labels improves overall metrics compared to the lower bound (LB) that is only trained on source data without test-time adaptation.", + "bbox": [ + 496, + 702, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of self-training loss. The results of Table 4-(6) and Table 4-(12) show the effect of self-training loss Eq. (8). It yields an improvement of more than $1.7\\mathrm{mAP}$ compared", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21291", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb14cff0bc335d4c526e0afbca505e83268ed605adcbeeccf42cf44e01fdc031.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison with the lower bound (LB) and our TTA-COPE under the semi-supervised setting." + ], + "image_footnote": [], + "bbox": [ + 91, + 79, + 883, + 417 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f692e43b64ef12a9a1b2f620b00ce7ad1deae0a5bf05ee97802cf937601503ce.jpg", + "table_caption": [ + "Table 5. Ablation study on different updating intervals." + ], + "table_footnote": [], + "table_body": "
MethodTTA TimeIoU755° 2cm5° 5 cm
Interval=1 (Ours)31 min43.533.338.1
Interval=1016 min41.832.737.7
Interval=2015 min41.131.937.1
", + "bbox": [ + 81, + 473, + 465, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to without self-training loss results of Table 4-(5, 11) in the $\\mathrm{IoU}_{75}$ metric. This indicates that the self-learning loss helps to learn the distribution of the target domain and reduces the domain shift by comparing the predicted NOCS map against the observed point cloud in the target domain. The result in the last column is our proposed model for test-time adaptation, Fig. 3 shows qualitative results against the lower bound method.", + "bbox": [ + 75, + 542, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Updating Interval. TTA is inevitably slower than simple inference since additional training is required. However, reducing the update interval enables faster TTA time than TTA baselines because of reduced training time. Table 5 shows the difference in TTA speed and performance according to the updating interval. Specifically, the interval of every 10 frames improves the speed of TTA roughly two-fold compared to Interval 1 (Ours), with a marginal performance drop. We also increase intervals to 20, but it does not show as much improvement as before since most of the bottleneck arises from inference time.", + "bbox": [ + 75, + 664, + 470, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed TTA-COPE, a test-time adaptation method for category-level object pose estimation that ad", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dresses the source-to-target domain shift without accessing source data at test time and without labeled target data. Specifically, we designed a pose ensemble method with self-training for test-time adaptation that simultaneously uses the teacher-student model to generate robust pseudo labels and estimate accurate poses for inference. We explore limitations of several test-time adaptation baselines and show that the proposed method achieves state-of-the-art performance. We demonstrate the benefits of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings.", + "bbox": [ + 496, + 460, + 890, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To the best of our knowledge, TTA-COPE is the first approach that tries to solve test-time adaptation for category-level object pose estimation. Since our method currently focuses on generating 6D pose labels, it does not affect 2D labels and the segmentation model. In future work, when jointly considering 2D label and Mask R-CNN, greater performance improvement and stable test-time adaptation would be possible. Also, our pose estimation relies on non-differentiable pose estimation, and as such we could benefit from a differentiable pose estimation method [30].", + "bbox": [ + 496, + 628, + 892, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 797, + 651, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was part of Taeyeop Lee's internship at NVIDIA and was also partially supported by an Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korean government(MSIT) (No.2020-0-00440).", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Time", + "bbox": [ + 836, + 65, + 872, + 77 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21292", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 1", + "[2] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11973-11982, 2020. 1, 2, 3", + "[3] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 1, 2, 3, 5, 6", + "[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3075-3084, 2019. 6", + "[5] Yang Fu and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. arXiv preprint arXiv:2206.15436, 2022. 2, 3, 5, 6", + "[6] Wei Gao and Russ Tedrake. kpm 2.0: Feedback control for category-level robotic manipulation. IEEE Robotics and Automation Letters (RA-L), 6(2):2962-2969, 2021. 1", + "[7] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2", + "[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 3, 5", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[10] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (ICML), pages 448-456, 2015. 3", + "[11] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. arXiv preprint arXiv:2207.13691, 2022. 3", + "[12] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12605-12614, 2020. 7", + "[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier," + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6d pose estimation of novel objects via render & compare. In 6th Annual Conference on Robot Learning (CoRL), 2022. 1", + "[14] Taeyeop Lee, Byeong-Uk Lee, Myungchul Kim, and In So Kweon. Category-level metric scale object shape and pose estimation. IEEE Robotics and Automation Letters (RA-L), 6(4):8575-8582, 2021. 2, 6", + "[15] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: Unsupervised domain adaptation for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14891-14900, 2022. 2, 3, 4, 5, 6, 7", + "[16] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2022. 2, 3, 4, 5, 6, 7", + "[17] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 1, 2, 3, 5, 6", + "[18] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an rgb image. In IEEE International Conference on Robotics and Automation (ICRA). ICRA, 2022. 1", + "[19] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 3", + "[20] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2020. 3", + "[21] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1215-1224, 2021. 2, 3", + "[22] Massimiliano Mancini, Hakan Karaoguz, Elisa Ricci, Patric Jensfelt, and Barbara Caputo. Kitting in the wild through online domain adaptation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1103-1109. IEEE, 2018. 3", + "[23] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics (TVCG), 22(12):2633-2651, 2015. 1", + "[24] Eitan Marder-Eppstein. Project Tango. In ACM SIGGRAPH Real-Time Live!, page 25, 2016. 1, 3", + "[25] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-DoF GraspNet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21293", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision (ICCV), pages 2901-2910, 2019. 1, 3", + "[26] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International Conference on Machine Learning (ICML), pages 16888-16905, 2022. 3", + "[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7668-7677, 2019. 1", + "[28] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-wise voting network for 6DoF pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4561-4570, 2019. 1", + "[29] Wanli Peng, Jianhang Yan, Hongtao Wen, and Yi Sun. Self-supervised category-level 6D object pose estimation with deep implicit shape representation. In Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI), volume 36, pages 2082-2090, 2022. 2, 3, 5, 6", + "[30] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky T. Q. Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 2022. 8", + "[31] Mihir Prabhudesai, Sujoy Paul, Sjoerd van Steenkiste, Mehdi SM Sajjadi, Anirudh Goyal, Deepak Pathak, Katerina Fragkiadaki, Gaurav Aggarwal, and Thomas Kipf. Test-time adaptation with slot-centric models. 2022. 3", + "[32] Martin Runz, Maud Buffer, and Lourdes Agapito. Mask-Fusion: Real-time recognition, tracking and reconstruction of multiple moving objects. In IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 10-20, 2018. 1", + "[33] Inkyu Shin, Yi-Hsuan Tsai, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Sparsh Garg, In So Kweon, and Kuk-Jin Yoon. MM-TTA: Multi-modal test-time adaptation for 3D semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16928-16937, 2022. 3, 7", + "[34] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. arXiv preprint arXiv:2303.01904, 2023. 3", + "[35] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 4, 6, 7", + "[36] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 5, 6", + "[37] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cam" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 969-977, 2018. 1", + "[38] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In Conference on Robot Learning (CoRL), 2018. 1", + "[39] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-DoF pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 1", + "[40] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence (TPAMI), 13(04):376-380, 1991. 2", + "[41] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3343–3352, 2019. 1", + "[42] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations (ICLR), 2021. 2, 3, 4, 5, 6, 7", + "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 1, 2, 3, 5, 6", + "[44] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 1, 2, 3, 6", + "[45] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21222-21231, 2022. 1", + "[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7201-7211, 2022. 2, 3", + "[47] Bowen Wen and Kostas Bekris. BundleTrack: 6D pose tracking for novel objects without instance or category-level 3D models. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074, 2021. 1", + "[48] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. Robotics: Science and Systems (RSS), 2022. 1" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21294", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1", + "[50] Bowen Wen, Chaitanya Mitash, Sruthi Soorian, Andrew Kimmel, Avishai Sintov, and Kostas E Bekris. Robust, occlusion-aware pose estimation for objects grasped by adaptive hands. In IEEE International Conference on Robotics and Automation (ICRA), pages 6210-6217. IEEE, 2020. 1", + "[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. Robotics: Science and Systems (RSS), 2018. 1", + "[52] He Yisheng, Wang Yao, Fan Haoqiang, Chen Qifeng, and Sun Jian. Fs6d: Few-shot 6d pose estimation of novel objects. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[53] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6" + ], + "bbox": [ + 78, + 90, + 468, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21295", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_model.json b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..63ae3e0cd8332d943e162a611db86ba502734130 --- /dev/null +++ b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.154 + ], + "angle": 0, + "content": "TTA-COPE: Test-Time Adaptation for Category-Level Object Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.18, + 0.838, + 0.233 + ], + "angle": 0, + "content": "Taeyeop Lee\\(^{1}\\) Jonathan Tremblay\\(^{2}\\) Valts Blukis\\(^{2}\\) Bowen Wen\\(^{2}\\) Byeong-Uk Lee\\(^{1}\\) Inkyu Shin\\(^{1}\\) Stan Birchfield\\(^{2}\\) In So Kweon\\(^{1}\\) Kuk-Jin Yoon\\(^{1}\\) KAIST \\({}^{2}\\)NVIDIA" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.543 + ], + "angle": 0, + "content": "Test-time adaptation methods have been gaining attention recently as a practical solution for addressing source-to-target domain gaps by gradually updating the model without requiring labels on the target data. In this paper, we propose a method of test-time adaptation for category-level object pose estimation called TTA-COPE. We design a pose ensemble approach with a self-training loss using pose-aware confidence. Unlike previous unsupervised domain adaptation methods for category-level object pose estimation, our approach processes the test data in a sequential, online manner, and it does not require access to the source domain at runtime. Extensive experimental results demonstrate that the proposed pose ensemble and the self-training loss improve category-level object pose performance during test time under both semi-supervised and unsupervised settings." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.573, + 0.21, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Object pose estimation is a crucial problem in computer vision and robotics. Advanced methods that focus on diverse variations of object 6D pose estimation have been introduced, such as known 3D objects (instance-level) [28, 38], category-level [18, 36, 43], few-shot [52], and zero-shot pose estimation [13, 47]. These techniques are useful for downstream applications requiring an online operation, such as robotic manipulation [6, 25, 48] and augmented reality [23, 24, 32]. Our paper focuses on the category-level object pose estimation problem since it is more broadly applicable than the instance-level problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Many works on category-level object pose estimation [2, 3, 17, 18, 36, 43, 44] have been proposed recently. These approaches estimate multiple classes of object pose more efficiently in a single network compared to the instance-level object pose estimation methods [27, 38, 41, 49-51], which depend on known 3D shape knowledge and the size of the objects. Notably, Wang et al. [43] introduced a novel representation called Normalized Object Coordinate Space (NOCS) to align various object instances within each" + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.274, + 0.674, + 0.288 + ], + "angle": 0, + "content": "No Adaptation" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.289, + 0.712, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.274, + 0.877, + 0.288 + ], + "angle": 0, + "content": "Test-Time Adaptation" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.289, + 0.892, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.395, + 0.712, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.395, + 0.892, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.5, + 0.712, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.5, + 0.892, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.606, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Figure 1. We propose a Test-Time Adaptation for Category-level Object Pose Estimation framework (TTA-COPE) that automatically improves the network in an online manner without labeled target data. As new image frames are processed, our method fine-tunes the network using the unlabeled data and simultaneously applies the network to perform pose estimation via inference. This approach successfully handles domain shifts compared with no adaptation, as seen here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.779 + ], + "angle": 0, + "content": "category in a canonical 3D space. The strengths of the NOCS representation have led to its adoption by follow-up work [3, 17, 36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In order to obtain accurate category-level object pose methods in unseen real-world scenarios, it is desirable to fine-tune the models in the new environment with labeled target data. The model that is not fine-tuned on the target domain distribution will almost certainly exhibit lower performance than the fine-tuned model [37]. However, annotating 6D poses of objects in the target environment is an expensive process [1, 39, 43, 45] that we seek to avoid." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21285" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.215 + ], + "angle": 0, + "content": "Table 1. Comparison with prior unsupervised works for category-level object pose estimation. Our unsupervised method trains models without 2D or 3D labels of target data, similar to Self-DPDN [16]. Unlike previous methods, our proposed approach updates the model online without offline batch processing. Moreover, we do not use the source data during test time (source-free) because it is impractical to train on a large amount of source data every iteration. There also may be privacy or legal constraints to access source data [21]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.224, + 0.467, + 0.332 + ], + "angle": 0, + "content": "
MethodUnsupervisedTest-time Adaptation
Target 3DTarget 2DSource-FreeOnline Adaptation
SupervisedXXXX
SSC-6D [29]✓XXX
RePoNet [5]✓XXX
UDA-COPE [15]✓X✓X
Self-DPDN [16]✓✓XX
Ours✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.446 + ], + "angle": 0, + "content": "Compared to annotating in 2D space, labeling in 3D space requires specific knowledge about geometry [7] from the annotator and is much more laborious, time-consuming, and error-prone due to the complex nature of \\(SE(3)\\) space. Therefore, it is usually challenging to annotate real-world data with 3D annotations for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.626 + ], + "angle": 0, + "content": "In order to solve the aforementioned problem of annotating object pose data in the real world, several recent methods [5, 15, 16] propose unsupervised domain adaptation (UDA) that aims to train the network without utilizing the ground truth of target pose labels. Although they show promising results using UDA techniques, these approaches still do not meet some of the requirements for online applications. For example, when a robot encounters a new environment, it is desirable to adapt the scene online manner while estimating object poses rather than waiting for enough data to be collected in the novel scene to train the model offline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.869 + ], + "angle": 0, + "content": "This problem definition of online fine-tuning is more practical for real applications, where we desire to update the model instantly when new data becomes available for fast domain adaptation. This setting is known as test-time adaptation (TTA) [42]. For TTA, the requirements are as follows: 1) labeled source data should not be accessed at test time, 2) adaptation should be online (rather than offline batch processing), and 3) the method should be fully unsupervised, without using 2D or 3D target labels during online fine-tuning. Since we do not have access to labeled source data (source-free) at test time this problem is more challenging than existing unsupervised category-level object pose methods [5,15,16,29]. Table 1 summarizes the difference between our problem definition and existing methods, showing that test-time adaptation for category-level object pose estimation remains an open problem." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose Test-time Adaptation for Category-level Object Pose Estimation (TTA-COPE) to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "handle domain shifts without any target domain annotations (see Fig. 1). Prior works on general test-time adaptation [42,46] propose self-training to minimize entropy loss. TENT [42] has shown improvement in 2D classification and segmentation tasks. We show, however, that simply extending TENT for the category-level object pose estimation is not effective. Another self-training strategy is the teacher-student framework [35] with pseudo labels. However, since pseudo labels are created without any noise filtering, naive pseudo labels may be unreliable and cause convergence to a suboptimal model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.469 + ], + "angle": 0, + "content": "To tackle this problem, we design a novel pose ensemble method to perform test-time adaptation for category-level object pose estimation by extending the pose-aware filtering of UDA-COPE [15]. The proposed method uses an ensemble of teacher-student predictions based on pose-aware confidence, which is used both for generating pseudo labels and inference. Also, the pose ensemble helps to train models with additional self-training loss to reduce the domain shift for category-level pose estimation by using pose-aware confidence. We demonstrate the advantages of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings. We show that our TTA-COPE framework achieves state-of-the-art performance compared to strong TTA baselines." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.47, + 0.892, + 0.498 + ], + "angle": 0, + "content": "In summary, the main contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.51, + 0.892, + 0.569 + ], + "angle": 0, + "content": "- We propose Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE), which handles domain shifts without labeling target data and without accessing source data during test time." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.569, + 0.892, + 0.627 + ], + "angle": 0, + "content": "- We introduce a pose ensemble with self-training loss that utilizes the teacher-student predictions to generate robust pseudo labels and estimates accurate poses for inference." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.628, + 0.892, + 0.688 + ], + "angle": 0, + "content": "- We evaluate our framework with experimental comparisons against strong test-time baselines and state-of-the-art methods under both semi-supervised and unsupervised settings." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.51, + 0.892, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.702, + 0.65, + 0.718 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.694, + 0.743 + ], + "angle": 0, + "content": "2.1. Supervised Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fully supervised learning methods for category-level object pose estimation [2,3,17,36,43,44] train their models using labeled source data (e.g., synthetic) and target data (e.g., real). Most category-level 6D object pose and size estimation approaches [3,14,36,43,44] use the dense correspondence via Normalized Object Coordinate Space (NOCS) representation as a common way to estimate pose and size. These correspondence-based methods initially estimate the NOCS map from RGB or depth images. Afterwards, the Umayama algorithm [40] with RANSAC is used to estimate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "optimal poses and object sizes by minimizing distances between depth and estimated NOCS map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Some methods use category priors [3,36,44] as the representative 3D shape per class, jointly reconstructing the full 3D shape and estimating the NOCS map from the full shape. Results show that this prior category helps improve the accuracy of the NOCS map and enhance the pose estimation performance. Other methods directly regress the pose or jointly utilize the correspondence representations [2,11,17]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.238, + 0.291, + 0.254 + ], + "angle": 0, + "content": "2.2. Unsupervised Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.468, + 0.428 + ], + "angle": 0, + "content": "Given that annotating the 6D object pose and 3D size labels in the real world is expensive, time-consuming, and laborious, RePoNet [5] and SSC-6D [29] propose semi-supervised approaches that reconstruct the entire shape and use differential rendering [19, 20] techniques for self-training signals. These are category-specific methods and utilize multiple models, as many models as the number of categories is required. UDA-COPE [15] proposes an unsupervised domain adaptation method to mitigate the domain shift from the source to the target domain with a single model to estimate all categories efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.468, + 0.671 + ], + "angle": 0, + "content": "Although these works show reasonable performance without using pose labels, there is still a limitation in relying on 2D ground truth information (segmentation or bounding boxes) to train the segmentation network or pose network. Self-DPDN [16] shows a fully unsupervised method using inter/intra-consistency as a reconstructed shape in a self-supervised objective but requires a supervised loss for the source domain in the unsupervised learning process. Additionally, the self-training loss utilizes a full 3D shape and requires an additional reconstruction module for 3D shape. Most methods, except for UDA-COPE [15], jointly use the supervised loss using source label data during unsupervised learning to relax the unstable training. Furthermore, all the aforementioned unsupervised methods train their model in an offline manner and are unsuitable for online applications [24, 25]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.347, + 0.697 + ], + "angle": 0, + "content": "2.3. Test-time Adaptation Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA) aims to enable the online adaptation of a pretrained model to the target domain without access to the source domain (source-free) [26,34,42,46]. The source data is commonly inaccessible during inference time because it is inefficient to train on a huge amount of source data every iteration. There may also be privacy or legal constraints to accessing source data [21], thus TTA is a more difficult but practical task than UDA. From the point of view of real-world applications, it is necessary to adapt to the new scene in an online way. Accordingly, test-time adaptation is necessary for the success of practical, real-world computer vision applications. Wang et al. propose Test entropy minimization (TENT) [42], which trains a net-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "work using a labeled source dataset, and adapts it to the unlabeled target dataset by updating the network parameters in batch norm layers using entropy loss. CoTTA [46] proposes a continual test-time adaptation method on the 2D classification and semantic segmentation tasks, and it effectively reduces the error accumulations while continually changing target data. Not limited to the 2D tasks, test-time adaptation methods have been applied to other tasks such as 3D segmentation [31, 33] and robot manipulation [22]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.24, + 0.625, + 0.256 + ], + "angle": 0, + "content": "3. TTA-COPE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Given an RGB-D image, our approach aims to estimate the 6D pose \\( T \\in \\mathrm{SE}(3) \\) and size \\( s \\in \\mathbb{R}_+^3 \\) of each object. The object pose \\( T \\) is defined as the rigid transformation \\( [R|t] \\), with rotation \\( R \\in \\mathrm{SO}(3) \\) and translation \\( t \\in \\mathbb{R}^3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Our method consists of a two-stage learning scheme using source and target data, respectively. In the first stage, we train the network using labeled (synthetic) source data in a supervised manner (Sec. 3.1). This step is the same as supervised methods and results in a pretrained model. We then utilize this pretrained model for test-time adaptation (TTA) using unlabeled target data without accessing the source data. To the best of our knowledge, we are the first to propose test-time adaptation for category-level object pose estimation. Therefore, we study and explore how TTA baselines (Sec. 3.2) might apply to category-level object pose estimation. Finally, our proposed method is presented and explained (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.762, + 0.547 + ], + "angle": 0, + "content": "3.1. Pretraining with Source Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this section, we introduce an overview of our model and then describe how to train the model using labeled source data. We use UDA-COPE [15] network as a base model, which we supplement with batch normalization (BN) [10] in the 2D network to utilize BN updates for test-time adaptation. Fig. 2 shows an overview of our network, consisting of three steps to estimate the pose similar to correspondence methods [3,16,36]. First, we detect the bounding area of the object and segment the object surface using the off-the-shelf segmentation model from a single RGB image [8]. In the second stage, given a cropped 2D object image \\(I\\) and segmented 3D object point cloud \\(D\\), our model estimates the correspondence as a NOCS map \\(N\\). We leverage the NOCS representation [43] to align diverse object instances within each class in a unified 3D space. In the final stage, our proposed pose ensemble method estimates the object pose \\(T\\). It simultaneously utilizes the predictions of the student and teacher model by calculating the pose confidence using the number of filtered points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The network is pretrained in a fully supervised manner using labeled source data. We minimize the predictions of the NOCS map \\( N \\) using the cross-entropy loss \\( L_{CE} \\) with ground truth labels \\( N^{GT} \\). We also use the consistency loss" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.085, + 0.827, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Figure 2. Overview of our Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE) framework. Our method consists of three steps to estimate category-level object pose, similar to correspondence methods. 1) We detect and segment the object region using an off-the-shelf segmentation model from a single RGB image. 2) Given a cropped 2D object image \\(I\\) and segmented 3D object point cloud \\(D\\), our model estimates the correspondence points as a NOCS map \\(N\\). 3) Our proposed pose ensemble estimates the object pose \\(T\\). It simultaneously utilizes the predictions of the student \\(\\theta^S\\) and teacher model \\(\\theta^T\\) by calculating the pose-aware confidence using the number of inlier points. We operate point filtering for each student and teacher model and select a confident pose \\(T\\) with higher confidence (i.e., more inlier points) for which the pose estimation is more accurate." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.448, + 0.47, + 0.492 + ], + "angle": 0, + "content": "\\(L_{C}\\) with 2D and 3D augmentation to make the network robust to noise in the input [15]. The supervised loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.5, + 0.469, + 0.517 + ], + "angle": 0, + "content": "\\[\nL _ {s u p} = \\lambda_ {C E} L _ {C E} (N, N ^ {G T}) + \\lambda_ {C} L _ {C} (N _ {a u g}, N), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.47, + 0.556 + ], + "angle": 0, + "content": "where \\(N_{aug}\\) is the estimated NOCS map from augmented 2D or 3D input, and \\(\\lambda_{CE},\\lambda_C\\in \\mathbb{R}_+\\) are weights." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.562, + 0.411, + 0.579 + ], + "angle": 0, + "content": "3.2. Test-time Adaptation with Target Data" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.585, + 0.469, + 0.646 + ], + "angle": 0, + "content": "The model is updated every iteration while simultaneously estimating the object pose of the current scene. (Note that this is different from previous methods that update the model by running multiple epochs over the target data [16].)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.469, + 0.706 + ], + "angle": 0, + "content": "We have a design choice for the objective loss when updating the model during test time. One of the widely used test-time objective losses is the entropy loss proposed by TENT [42]," + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.712, + 0.469, + 0.734 + ], + "angle": 0, + "content": "\\[\nL _ {e n t} = - \\sum p \\left(x _ {t}\\right) \\log p \\left(x _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.788 + ], + "angle": 0, + "content": "where \\( p(x_{t}) \\) is the probability of predictions \\( \\theta^S (x_t) \\) from target data \\( x_{t} \\). This simple objective loss encourages sharp distributions by assigning the most probability." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.788, + 0.469, + 0.878 + ], + "angle": 0, + "content": "Another common approach is utilizing pseudo labels from a teacher-student framework [35] with momentum update. The teacher model \\(\\theta^T\\) generates pseudo ground truth \\(\\hat{y} = \\theta^T(x_t)\\), where \\(x_t = (I, D)\\). The student model \\(\\theta^S\\) then uses the predictions of the teacher as the ground truth signals (pseudo labels) with cross-entropy loss," + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {p l} = L _ {C E} \\left(\\theta^ {S} \\left(x _ {t}\\right), \\hat {y}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.492 + ], + "angle": 0, + "content": "After updating the student model \\(\\theta_{i}^{S}\\rightarrow \\theta_{i + 1}^{S}\\) by minimizing Eq. (3), the teacher model \\(\\theta_{i + 1}^{T}\\) is updated by momentum update," + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.492, + 0.892, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i + 1} ^ {T} \\leftarrow \\gamma \\theta_ {i} ^ {T} + (1 - \\gamma) \\theta_ {i + 1} ^ {S}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.579 + ], + "angle": 0, + "content": "where \\( i \\) stands for the time stamp of the iteration and \\( \\gamma \\) is the momentum smoothing factor. The separation of the model structure mitigates some error accumulation by the pseudo label and momentum update." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.58, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Although these objective strategies from previous work potentially provide self-training signals for test-time adaptation, they become brittle when directly applied to a category-level object pose estimation task. The pseudo labels \\(\\hat{y}\\) are still inherently exposed to noisy inputs and predictions, leading to errors in the student model \\(\\theta^S\\). Since \\(\\theta^T\\) cannot guarantee clean pseudo labels for the student, we need the ability to filter noise in the input or predictions and to yield more reliable pseudo labels to the \\(\\theta^S\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.651, + 0.74 + ], + "angle": 0, + "content": "3.3. Pose Ensemble" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "To solve this problem, we propose to extend the pseudolabel filtering proposed by UDA-COPE [15] with pose ensemble processing. We observe that the UDA-COPE framework and pose filtering, while showing notable results, is inefficient for two reasons. First, the teacher model is only used for generating pseudo labels and not for inference. Second, the student model is also only used for inference and has less influence in generating pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We propose instead to jointly utilize the teacher and student models for both generating pseudo labels and for infer" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21288" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "ence, which we show in the experimental results produces a noticeable improvement. To this end, we design a pose ensemble module to simultaneously use teacher and student model predictions. Fig. 2 shows an overview of the pose ensemble with point filtering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.351 + ], + "angle": 0, + "content": "The point filtering has three steps. Unlike UDA-COPE [15], we repeat these steps for both the student and teacher: 1) We initially estimate \\( T^S \\) using the Umeyama algorithm \\( \\theta_p \\) within the region of object points \\( D \\) and estimated NOCS map \\( N^S \\). 2) We transform observed depth to normalized object coordinate space \\( \\widehat{D}^S \\) by multiplying \\( (T^S)^{-1} \\) and \\( D \\). 3) We compute each matching point distance between \\( N^S \\) and \\( \\widehat{D} \\), removing outliers that exceed a certain threshold \\( \\rho \\). This whole process of point-filtering is then repeated using \\( N^T \\) to estimate \\( T^T \\). This process is represented as follows, where \\( N^{S,T} \\) is either \\( N^S \\) or \\( N^T \\), and similarly for \\( T^{S,T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.362, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T ^ {S, T} = \\theta_ {p} (D, N ^ {S, T}), \\\\ \\widehat {D} ^ {S, T} = \\left(T ^ {S, T}\\right) ^ {- 1} D, \\tag {5} \\\\ e _ {j} ^ {S, T} = \\left\\{ \\begin{array}{l l} i n l i e r & \\text {i f} \\| \\widehat {D} _ {j} ^ {S, T} \\\\ o u t l i e r & \\text {o t h e r w i s e .} \\end{array} \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.359, + 0.476 + ], + "angle": 0, + "content": "where \\(j = 1,\\ldots ,X\\) , and \\(X = |D| = |\\widehat{D} |\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.657 + ], + "angle": 0, + "content": "Our proposed pose ensemble aims to fuse two predictions \\((N^S, N^T)\\) to estimate the optimal pose. Although this fusion could be conducted at the input level before obtaining the \\(T\\) from \\(\\theta_p\\), we found that our proposal of fusing at the output level after getting \\(T\\) produces slightly better results (see the ablation study, Sec. 4.5). We first estimate each pose \\((T^S, T^T)\\) from each prediction \\((N^S, N^T)\\) using \\(\\theta_p\\). We then calculate the number of inliers \\((inliers^S, inliers^T)\\) from each pose using Eq. (5) to evaluate the confidence of the pose. We assume that the higher this confidence (i.e., the more inliers) is, the more accurate the pose estimation will be. Therefore:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.669, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\nT _ {o u t} = \\left\\{ \\begin{array}{l l} T ^ {S} & \\text {i f i n l i e r s} ^ {S} > \\text {i n l i e r s} ^ {T}, \\\\ T ^ {T} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.765 + ], + "angle": 0, + "content": "We found that this simple pose-aware ensemble method is more effective for generating high-quality pseudo labels and inference than other ensemble methods (Sec. 4.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.469, + 0.827 + ], + "angle": 0, + "content": "Another crucial part of our method is a self-training loss. The total loss that we use to update the student network combines this self-training \\( L_{d} \\) loss and the pseudo label loss \\( L_{pl} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.84, + 0.469, + 0.857 + ], + "angle": 0, + "content": "\\[\nL _ {t t a} = \\lambda_ {d} L _ {d} + \\lambda_ {p l} L _ {p l}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The first term is self-training loss \\( L_{d} \\), which helps learn the distribution of the target domain and reduce the domain gap" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "through \\(T_{out}\\), which is obtained from the pose ensemble and observed \\(D\\). This loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.133, + 0.892, + 0.151 + ], + "angle": 0, + "content": "\\[\nL _ {d} = L _ {C E} \\left(N _ {e} ^ {S}, \\operatorname {E n c} \\left(\\widehat {D} _ {e}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.164, + 0.892, + 0.226 + ], + "angle": 0, + "content": "where \\( N_{e}^{S} \\) is the inlier-only student NOCS map, \\( \\widehat{D}_{e} \\) is the resulting inliner-only transformed depth point cloud in NOCS space, and \\( \\mathrm{Enc}(\\cdot) \\) is the one-hot encoding to enable cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.227, + 0.634, + 0.24 + ], + "angle": 0, + "content": "The second term" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.256, + 0.892, + 0.274 + ], + "angle": 0, + "content": "\\[\nL _ {p l} = L _ {C E} \\left(N _ {e} ^ {S}, N _ {e} ^ {T}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.35 + ], + "angle": 0, + "content": "is the pseudo label loss from UDA-COPE [15], where \\(N_{e}^{T}\\) is the inlier-only teacher NOCS map. Since \\(N^{S} = \\theta^{S}(x_{t})\\), and \\(\\hat{y} = \\theta^T (x_t)\\), \\(L_{pl}\\) in Eq. (9) is the same as the loss in Eq. (3), except that here only inliers are considered." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.364, + 0.632, + 0.381 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.389, + 0.596, + 0.404 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.716 + ], + "angle": 0, + "content": "We utilize two widely used category-level pose estimation datasets as the source and target domain, respectively. The source data is the Context-Aware MixEd Reality (CAMERA) dataset [43], generated by rendering and compositing synthetic objects into real scenes while considering the context. CAMERA consists of 275K RGB-D images as a training set with 1,085 object instances chosen from six categories: bottle, bowl, camera, can, laptop and mug. We use the REAL dataset [43] as the target domain. The target data consists of 4,300 real-world images of seven scenes for training and 2,750 real-world images of six scenes for evaluation. We refer to REAL evaluation set as REAL275. Our TTA methods do not use the target training set (4,300 images with seven scenes) and only use the evaluation set (2,750 images with six scenes, REAL275) for test-time adaptation, e.g., TENT [42] process evaluation set in a sequential, online manner. On the other hand, unsupervised methods such as [5, 16, 29] use the target training set for updating the models without any restrictions on how much and how long they use the information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.719, + 0.742 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object Segmentation. We use Mask R-CNN [8] to obtain the object area in 2D image. We use the identical results of Mask R-CNN for a fair comparison with previous methods [3, 17, 36] for semi-supervised and unsupervised settings. For the semi-supervised setting, Mask R-CNN has trained on the source, and target domain supervised manner. For the unsupervised setting, Mask R-CNN is trained on only the source domain. The detected area of the image resizes to \\(192 \\times 192\\) image resolution as the teacher or student model input." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21289" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.089, + 0.768, + 0.104 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons with state-of-the-art methods on the REAL275 dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.111, + 0.878, + 0.421 + ], + "angle": 0, + "content": "
MethodSupervisedUnsupervisedmAP (↑)
SourceTargetTargetOnlineIoU50IoU755° 2cm5° 5cm10° 2cm10° 5 cm
Source (Supervised)
Metric Scale [14]2D/3D54.68.42.25.410.125.0
SPD [36]2D/3D50.517.011.512.133.037.9
Source (Supervised) & Target (Supervised)
NOCS [43]2D/3D2D/3D47.29.47.210.013.825.2
SPD [36]2D/3D2D/3D68.527.019.521.643.554.0
DualPoseNet [17]2D/3D2D/3D—30.829.335.950.066.8
CR-Net [44]2D/3D2D/3D—33.227.834.347.260.8
SGPA [3]2D/3D2D/3D—37.135.939.661.370.7
Source (Supervised) & Target (Semi-Supervised)
SSC-6D [29]2D/3D2D3DX73.0—16.819.644.154.5
RePoNet [5]2D/3D2D3DX76.0—30.733.9—63.0
UDA-COPE [15]2D/3D2D3DX75.534.430.534.957.066.1
Self-DPDN [16]2D/3D2D3DX75.241.639.545.063.372.2
TTA-COPE (Ours)2D/3D2D3D✓78.743.533.338.164.375.1
Source (Supervised) & Target (Unsupervised)
Self-DPDN [16]2D/3D2D/3DX67.243.939.046.761.873.4
TTA-COPE (Ours)2D/3D2D/3D✓69.139.730.235.961.773.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.733 + ], + "angle": 0, + "content": "Student and Teacher. The student and teacher models have identical design in 2D and 3D branches from UDA-COPE [15]. We utilize the PSPNet [53] with ResNet34 [9] backbone for the 2D image feature extraction. For a 3D branch, we use the MinkowskiNet [4] and utilize sparse convolution operation with a \\(5\\mathrm{cm}\\) voxel size. Our NOCS representation uses the classification with 32 bins [43] instead of direct regression. During the pretraining stage, we train our model on the source data for 50 epochs using the Adam optimizer by initializing the learning rate of 1e-4 with a batch size of 32. The learning rate was reduced by a ratio of 0.6 (at 15k iterations), 0.3 (at 30k iterations), 0.1 (at 45k iterations), and 0.01 (at 60k iterations). During test-time adaptation, our student model uses the same learning rate, and the teacher smoothly updates using momentum update with \\(\\gamma = 0.99\\). Given target data, we first update the model every iteration and then estimate the pose. We set \\(\\lambda_{\\mathrm{CE}} = 1.0\\), \\(\\lambda_{\\mathrm{C}} = 1\\mathrm{e - }6\\), \\(\\lambda_{\\mathrm{d}} = 1.0\\), \\(\\lambda_{\\mathrm{pl}} = 1.0\\) for our experiments. The point filtering threshold \\(\\rho\\) was set to 0.05 for all experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Metrics. To evaluate the performance of 3D object detection and 6D pose estimation, we follow the previous pose and size evaluation metric from Wang et al. [43]. We report the mean average precision (mAP) at the \\(50\\%\\) and \\(75\\%\\) intersection over union (IoU) thresholds for 3D object detection. We also report mAP for 6D object pose evaluation w.r.t. rotation and translation errors, where e.g. the \\(5^{\\circ}5\\mathrm{cm}\\) metric describes the percentage of pose predictions where the error is less than both \\(5^{\\circ}\\) and \\(5\\mathrm{cm}\\) and the same for other thresholds. We recalculated all the metrics with the improved code." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.443, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons with TTA baselines for category-level object pose estimation on the REAL275 dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.482, + 0.891, + 0.597 + ], + "angle": 0, + "content": "
MethodUnsupervisedIoU75mAP (↑)
Target 3DTarget 2D5° 2cm5° 5 cm
TENT [42] - Eq. (2)✓33.725.929.1
PL [35] - Eqs. (3)-(4)✓39.929.734.9
PL-F [15] - Eqs. (3)-(5)✓41.131.436.3
TTA-COPE (Ours)✓43.533.338.1
TENT [42] - Eq. (2)✓✓32.326.831.2
PL [35] - Eqs. (3)-(4)✓✓36.026.933.1
PL-F [15] - Eqs. (3)-(5)✓✓36.528.034.0
TTA-COPE (Ours)✓✓39.730.235.9
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.617, + 0.791, + 0.633 + ], + "angle": 0, + "content": "4.3. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Table 2 summarizes quantitative results on the REAL275 dataset under different settings: 1) supervised only on source data, 2) supervised on source and target data, 3) supervised on source data and semi-supervised on target data, and 4) supervised on source data and unsupervised on target data. Not surprisingly, methods following 2) such as SPD [36] perform better than setting 1) since they have the benefit of accessing both source and target data. Recent unsupervised domain adaptation (UDA) methods [15, 16] show remarkable results compared to state-of-the-art supervised methods without using target 3D labels (semi-supervised setting). Our method shows state-of-the-art results in IoU metrics under the semi-supervised setting and even outperforms the recent supervised methods SGPA [3] by a large margin (6.4 mAP in \\(\\mathrm{IoU}_{75}\\)). Our TTA-COPE uses less time and data to train the target domain because of the advantage of test-time adaptation but achieves comparable" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21290" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.089, + 0.854, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation study on variants of pose ensemble methods and self-training loss under the semi-supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.108, + 0.88, + 0.338 + ], + "angle": 0, + "content": "
MethodPose EnsemblemAP (↑)
InputOutputInferencePseudo LabelLd(8)IoU50IoU755° 2cm5° 5 cm10° 2cm10° 5 cm
LB76.237.529.134.662.173.2
(1)Argmax Match✓71.127.529.634.360.771.7
(2)Softmax Avg.✓78.740.432.137.364.074.7
(3)Softmax Max✓77.640.931.636.463.474.3
(4)Softmax Max✓77.541.132.137.164.374.6
(5)Inliers Max✓77.641.232.837.664.474.9
(6)Inliers Max✓✓78.242.932.437.864.174.7
(7)Argmax Match✓78.340.731.736.463.273.9
(8)Softmax Avg.✓78.641.131.536.263.674.5
(9)Softmax Max✓78.141.231.936.563.574.4
(10)Softmax Max✓78.641.231.836.363.574.4
(11)Inliers Max✓78.241.331.636.263.674.3
(12)Inliers Max✓✓78.743.032.637.063.573.9
OursInliers Max✓✓✓78.743.533.338.164.675.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.412 + ], + "angle": 0, + "content": "results to the SOTA method, Self-DPDN [16]. Ours takes about 31 minutes (Table 5) for TTA and is \\(58\\mathrm{x}\\) faster than the Self-DPDN, which takes about 30 hours for training in the target domain, excluding inference time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.421, + 0.36, + 0.436 + ], + "angle": 0, + "content": "4.4. Comparison with TTA baselines" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.444, + 0.471, + 0.732 + ], + "angle": 0, + "content": "Table 3 summarizes results for the different design test-time adaptation (TTA) baselines in semi-supervised and unsupervised settings. TENT [42] has been designed for general unsupervised settings, but it is not specially designed for the object pose estimation task and naturally observes overall poor performance compared to other TTA baselines. Mean teacher with pseudo labels (PL) [35] performs better than TENT [42] in semi-supervised and unsupervised settings, and we believe that training the model using pseudo labels with momentum update (Eqs. (3)-(4)) provides more stable training signals by reducing the error accumulation than entropy minimization (2). However, using pseudo labels without filtering (PL) provides unreliable labels as the ground truth, we found that pseudo label filtering (PL-F) [15] performs better than PL, which indicates that noise filtering (Eq. (5)) improves the results for the student as expected. Finally, our TTA-COPE achieves state-of-the-art (SOTA) performance among all TTA baselines under both semi-supervised and unsupervised settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.231, + 0.758 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.81 + ], + "angle": 0, + "content": "In this section, we conduct experiments to evaluate the efficacy of the pose ensemble and self-training loss under a semi-supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Input/Output-Level Pose Ensemble. As mentioned in the Sec. 3.3, we compare the different pose ensemble methods given predictions of teacher and student models \\((N^T, N^S)\\) and answer the following question: Which ensemble is the most effective for test-time adaptation? We categorize two ensemble techniques:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.456 + ], + "angle": 0, + "content": "1) Input-level ensemble that fuses two NOCS map predictions \\((N^T, N^S)\\) and makes an accurate single NOCS map \\(N\\) to estimate poses using the Umayama algorithm \\(\\theta_p\\). For the input-level ensemble that fuses two predictions \\((N^T, N^S)\\) into one prediction \\(N\\), we compare three strategies: Argmax Match, Softmax Average, and Softmax Max operations [12, 33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.534 + ], + "angle": 0, + "content": "2) We ensemble output-level predictions \\((N^T, N^S)\\) to estimate each pose \\((T^T, T^S)\\) and choose the best pose predictions under specific criteria. Unlike the input ensemble, only the Softmax Max operation is valid for the output-level ensemble due to the nonlinearity of poses." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Table 4 summarizes our ensemble ablation study results. Table 4-(1-3) shows that the Softmax Avg. and Softmax Max perform better than the Argmax Match operation among input-level pose ensembles. If we compare the same operation (Softmax Max) in the input-level, Table 4-(3), and output-level, Table 4-(4), the output-level ensemble shows slightly better performance since the output-level operation is directly related to the pose results. Our proposed pose ensemble, Table 4-(5), considers pose-aware confidence by choosing a higher number of inlier points and achieves the best performance in all pose ensembles in Table 4-(1-5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Use of Pose Ensemble. Our pose ensemble can be used for inference Table 4-(1-6, Ours) as well as generating pseudo labels Table 4-(7-12, Ours). The results show that our pose ensemble is helpful for both cases. In particular, applying our ensemble to inferences Table 4-(5) has been shown to be more effective than application to generating pseudo labels Table 4-(11) in pose metric. We also observe that our pose ensemble used for inference or pseudo labels improves overall metrics compared to the lower bound (LB) that is only trained on source data without test-time adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of self-training loss. The results of Table 4-(6) and Table 4-(12) show the effect of self-training loss Eq. (8). It yields an improvement of more than \\(1.7\\mathrm{mAP}\\) compared" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21291" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.837, + 0.066, + 0.874, + 0.078 + ], + "angle": 0, + "content": "Time" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.08, + 0.885, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.428, + 0.848, + 0.443 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison with the lower bound (LB) and our TTA-COPE under the semi-supervised setting." + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.459, + 0.446, + 0.472 + ], + "angle": 0, + "content": "Table 5. Ablation study on different updating intervals." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.474, + 0.466, + 0.537 + ], + "angle": 0, + "content": "
MethodTTA TimeIoU755° 2cm5° 5 cm
Interval=1 (Ours)31 min43.533.338.1
Interval=1016 min41.832.737.7
Interval=2015 min41.131.937.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.664 + ], + "angle": 0, + "content": "to without self-training loss results of Table 4-(5, 11) in the \\(\\mathrm{IoU}_{75}\\) metric. This indicates that the self-learning loss helps to learn the distribution of the target domain and reduces the domain shift by comparing the predicted NOCS map against the observed point cloud in the target domain. The result in the last column is our proposed model for test-time adaptation, Fig. 3 shows qualitative results against the lower bound method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.471, + 0.831 + ], + "angle": 0, + "content": "Updating Interval. TTA is inevitably slower than simple inference since additional training is required. However, reducing the update interval enables faster TTA time than TTA baselines because of reduced training time. Table 5 shows the difference in TTA speed and performance according to the updating interval. Specifically, the interval of every 10 frames improves the speed of TTA roughly two-fold compared to Interval 1 (Ours), with a marginal performance drop. We also increase intervals to 20, but it does not show as much improvement as before since most of the bottleneck arises from inference time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We have proposed TTA-COPE, a test-time adaptation method for category-level object pose estimation that ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.627 + ], + "angle": 0, + "content": "dresses the source-to-target domain shift without accessing source data at test time and without labeled target data. Specifically, we designed a pose ensemble method with self-training for test-time adaptation that simultaneously uses the teacher-student model to generate robust pseudo labels and estimate accurate poses for inference. We explore limitations of several test-time adaptation baselines and show that the proposed method achieves state-of-the-art performance. We demonstrate the benefits of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.78 + ], + "angle": 0, + "content": "To the best of our knowledge, TTA-COPE is the first approach that tries to solve test-time adaptation for category-level object pose estimation. Since our method currently focuses on generating 6D pose labels, it does not affect 2D labels and the segmentation model. In future work, when jointly considering 2D label and Mask R-CNN, greater performance improvement and stable test-time adaptation would be possible. Also, our pose estimation relies on non-differentiable pose estimation, and as such we could benefit from a differentiable pose estimation method [30]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.799, + 0.652, + 0.815 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was part of Taeyeop Lee's internship at NVIDIA and was also partially supported by an Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korean government(MSIT) (No.2020-0-00440)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21292" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11973-11982, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[3] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3075-3084, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[5] Yang Fu and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. arXiv preprint arXiv:2206.15436, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Wei Gao and Russ Tedrake. kpm 2.0: Feedback control for category-level robotic manipulation. IEEE Robotics and Automation Letters (RA-L), 6(2):2962-2969, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[7] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.664, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (ICML), pages 448-456, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.72, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[11] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. arXiv preprint arXiv:2207.13691, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[12] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12605-12614, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier," + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6d pose estimation of novel objects via render & compare. In 6th Annual Conference on Robot Learning (CoRL), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Taeyeop Lee, Byeong-Uk Lee, Myungchul Kim, and In So Kweon. Category-level metric scale object shape and pose estimation. IEEE Robotics and Automation Letters (RA-L), 6(4):8575-8582, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[15] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: Unsupervised domain adaptation for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14891-14900, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[17] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an rgb image. In IEEE International Conference on Robotics and Automation (ICRA). ICRA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[19] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[20] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[21] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1215-1224, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[22] Massimiliano Mancini, Hakan Karaoguz, Elisa Ricci, Patric Jensfelt, and Barbara Caputo. Kitting in the wild through online domain adaptation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1103-1109. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics (TVCG), 22(12):2633-2651, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[24] Eitan Marder-Eppstein. Project Tango. In ACM SIGGRAPH Real-Time Live!, page 25, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-DoF GraspNet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision (ICCV), pages 2901-2910, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[26] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International Conference on Machine Learning (ICML), pages 16888-16905, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7668-7677, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[28] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-wise voting network for 6DoF pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4561-4570, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.331, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[29] Wanli Peng, Jianhang Yan, Hongtao Wen, and Yi Sun. Self-supervised category-level 6D object pose estimation with deep implicit shape representation. In Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI), volume 36, pages 2082-2090, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[30] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky T. Q. Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[31] Mihir Prabhudesai, Sujoy Paul, Sjoerd van Steenkiste, Mehdi SM Sajjadi, Anirudh Goyal, Deepak Pathak, Katerina Fragkiadaki, Gaurav Aggarwal, and Thomas Kipf. Test-time adaptation with slot-centric models. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[32] Martin Runz, Maud Buffer, and Lourdes Agapito. Mask-Fusion: Real-time recognition, tracking and reconstruction of multiple moving objects. In IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 10-20, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[33] Inkyu Shin, Yi-Hsuan Tsai, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Sparsh Garg, In So Kweon, and Kuk-Jin Yoon. MM-TTA: Multi-modal test-time adaptation for 3D semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16928-16937, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[34] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. arXiv preprint arXiv:2303.01904, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[35] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[36] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cam" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "eracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 969-977, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[38] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In Conference on Robot Learning (CoRL), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[39] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-DoF pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[40] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence (TPAMI), 13(04):376-380, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[41] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3343–3352, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[42] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations (ICLR), 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[44] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[45] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21222-21231, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7201-7211, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[47] Bowen Wen and Kostas Bekris. BundleTrack: 6D pose tracking for novel objects without instance or category-level 3D models. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[48] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. Robotics: Science and Systems (RSS), 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[49] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[50] Bowen Wen, Chaitanya Mitash, Sruthi Soorian, Andrew Kimmel, Avishai Sintov, and Kostas E Bekris. Robust, occlusion-aware pose estimation for objects grasped by adaptive hands. In IEEE International Conference on Robotics and Automation (ICRA), pages 6210-6217. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. Robotics: Science and Systems (RSS), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[52] He Yisheng, Wang Yao, Fan Haoqiang, Chen Qifeng, and Sun Jian. Fs6d: Few-shot 6d pose estimation of novel objects. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[53] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "21295" + } + ] +] \ No newline at end of file diff --git a/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_origin.pdf b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2232ab1d6852d92046af199be58c35b5d63f7d62 --- /dev/null +++ b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/e0dd8ebd-75c4-4890-8983-be1eb1c01bc4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a01f1e29442bf8875fad03456c6f81c20a1ff09a6173971ecf00bf7f7d8deb8 +size 1412317 diff --git a/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/full.md b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..434a13237f481973af306ba55c90277ab6034f9a --- /dev/null +++ b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/full.md @@ -0,0 +1,312 @@ +# TTA-COPE: Test-Time Adaptation for Category-Level Object Pose Estimation + +Taeyeop Lee $^{1}$ Jonathan Tremblay $^{2}$ Valts Blukis $^{2}$ Bowen Wen $^{2}$ Byeong-Uk Lee $^{1}$ Inkyu Shin $^{1}$ Stan Birchfield $^{2}$ In So Kweon $^{1}$ Kuk-Jin Yoon $^{1}$ KAIST ${}^{2}$ NVIDIA + +# Abstract + +Test-time adaptation methods have been gaining attention recently as a practical solution for addressing source-to-target domain gaps by gradually updating the model without requiring labels on the target data. In this paper, we propose a method of test-time adaptation for category-level object pose estimation called TTA-COPE. We design a pose ensemble approach with a self-training loss using pose-aware confidence. Unlike previous unsupervised domain adaptation methods for category-level object pose estimation, our approach processes the test data in a sequential, online manner, and it does not require access to the source domain at runtime. Extensive experimental results demonstrate that the proposed pose ensemble and the self-training loss improve category-level object pose performance during test time under both semi-supervised and unsupervised settings. + +# 1. Introduction + +Object pose estimation is a crucial problem in computer vision and robotics. Advanced methods that focus on diverse variations of object 6D pose estimation have been introduced, such as known 3D objects (instance-level) [28, 38], category-level [18, 36, 43], few-shot [52], and zero-shot pose estimation [13, 47]. These techniques are useful for downstream applications requiring an online operation, such as robotic manipulation [6, 25, 48] and augmented reality [23, 24, 32]. Our paper focuses on the category-level object pose estimation problem since it is more broadly applicable than the instance-level problem. + +Many works on category-level object pose estimation [2, 3, 17, 18, 36, 43, 44] have been proposed recently. These approaches estimate multiple classes of object pose more efficiently in a single network compared to the instance-level object pose estimation methods [27, 38, 41, 49-51], which depend on known 3D shape knowledge and the size of the objects. Notably, Wang et al. [43] introduced a novel representation called Normalized Object Coordinate Space (NOCS) to align various object instances within each + +![](images/8ef1e6bd5b3704e36771617886e8f39f7d9331b1e70e077523e0cc703f2436cf.jpg) +No Adaptation + +![](images/c3dbda982517c5473aedb4167fbacb6a9599b1521b7f18ec04a738358870136e.jpg) +Test-Time Adaptation + +![](images/fc3ea698be5ab3f71879da2fdc3a2beaa84e91fdf69ce9c75577a69518efe261.jpg) + +![](images/980c223f2f6155f6f0171b6eaa165cd5e91f23687d74a43d4a05297da5750bde.jpg) + +![](images/cbe5bf23a24cc471ea48dd2491de9cb4b7e5fe2518cb029ef6983dea926ae2f2.jpg) +Figure 1. We propose a Test-Time Adaptation for Category-level Object Pose Estimation framework (TTA-COPE) that automatically improves the network in an online manner without labeled target data. As new image frames are processed, our method fine-tunes the network using the unlabeled data and simultaneously applies the network to perform pose estimation via inference. This approach successfully handles domain shifts compared with no adaptation, as seen here. + +![](images/01f64bb8e9793d621a001de91fa2b851caa9e0dc20e4981448cab24545eebcdf.jpg) + +category in a canonical 3D space. The strengths of the NOCS representation have led to its adoption by follow-up work [3, 17, 36]. + +In order to obtain accurate category-level object pose methods in unseen real-world scenarios, it is desirable to fine-tune the models in the new environment with labeled target data. The model that is not fine-tuned on the target domain distribution will almost certainly exhibit lower performance than the fine-tuned model [37]. However, annotating 6D poses of objects in the target environment is an expensive process [1, 39, 43, 45] that we seek to avoid. + +Table 1. Comparison with prior unsupervised works for category-level object pose estimation. Our unsupervised method trains models without 2D or 3D labels of target data, similar to Self-DPDN [16]. Unlike previous methods, our proposed approach updates the model online without offline batch processing. Moreover, we do not use the source data during test time (source-free) because it is impractical to train on a large amount of source data every iteration. There also may be privacy or legal constraints to access source data [21]. + +
MethodUnsupervisedTest-time Adaptation
Target 3DTarget 2DSource-FreeOnline Adaptation
SupervisedXXXX
SSC-6D [29]✓XXX
RePoNet [5]✓XXX
UDA-COPE [15]✓X✓X
Self-DPDN [16]✓✓XX
Ours✓✓✓✓
+ +Compared to annotating in 2D space, labeling in 3D space requires specific knowledge about geometry [7] from the annotator and is much more laborious, time-consuming, and error-prone due to the complex nature of $SE(3)$ space. Therefore, it is usually challenging to annotate real-world data with 3D annotations for fine-tuning. + +In order to solve the aforementioned problem of annotating object pose data in the real world, several recent methods [5, 15, 16] propose unsupervised domain adaptation (UDA) that aims to train the network without utilizing the ground truth of target pose labels. Although they show promising results using UDA techniques, these approaches still do not meet some of the requirements for online applications. For example, when a robot encounters a new environment, it is desirable to adapt the scene online manner while estimating object poses rather than waiting for enough data to be collected in the novel scene to train the model offline. + +This problem definition of online fine-tuning is more practical for real applications, where we desire to update the model instantly when new data becomes available for fast domain adaptation. This setting is known as test-time adaptation (TTA) [42]. For TTA, the requirements are as follows: 1) labeled source data should not be accessed at test time, 2) adaptation should be online (rather than offline batch processing), and 3) the method should be fully unsupervised, without using 2D or 3D target labels during online fine-tuning. Since we do not have access to labeled source data (source-free) at test time this problem is more challenging than existing unsupervised category-level object pose methods [5,15,16,29]. Table 1 summarizes the difference between our problem definition and existing methods, showing that test-time adaptation for category-level object pose estimation remains an open problem. + +In this paper, we propose Test-time Adaptation for Category-level Object Pose Estimation (TTA-COPE) to + +handle domain shifts without any target domain annotations (see Fig. 1). Prior works on general test-time adaptation [42,46] propose self-training to minimize entropy loss. TENT [42] has shown improvement in 2D classification and segmentation tasks. We show, however, that simply extending TENT for the category-level object pose estimation is not effective. Another self-training strategy is the teacher-student framework [35] with pseudo labels. However, since pseudo labels are created without any noise filtering, naive pseudo labels may be unreliable and cause convergence to a suboptimal model. + +To tackle this problem, we design a novel pose ensemble method to perform test-time adaptation for category-level object pose estimation by extending the pose-aware filtering of UDA-COPE [15]. The proposed method uses an ensemble of teacher-student predictions based on pose-aware confidence, which is used both for generating pseudo labels and inference. Also, the pose ensemble helps to train models with additional self-training loss to reduce the domain shift for category-level pose estimation by using pose-aware confidence. We demonstrate the advantages of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings. We show that our TTA-COPE framework achieves state-of-the-art performance compared to strong TTA baselines. + +In summary, the main contributions of our work are as follows: + +- We propose Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE), which handles domain shifts without labeling target data and without accessing source data during test time. +- We introduce a pose ensemble with self-training loss that utilizes the teacher-student predictions to generate robust pseudo labels and estimates accurate poses for inference. +- We evaluate our framework with experimental comparisons against strong test-time baselines and state-of-the-art methods under both semi-supervised and unsupervised settings. + +# 2. Related Works + +# 2.1. Supervised Methods + +Fully supervised learning methods for category-level object pose estimation [2,3,17,36,43,44] train their models using labeled source data (e.g., synthetic) and target data (e.g., real). Most category-level 6D object pose and size estimation approaches [3,14,36,43,44] use the dense correspondence via Normalized Object Coordinate Space (NOCS) representation as a common way to estimate pose and size. These correspondence-based methods initially estimate the NOCS map from RGB or depth images. Afterwards, the Umayama algorithm [40] with RANSAC is used to estimate + +optimal poses and object sizes by minimizing distances between depth and estimated NOCS map. + +Some methods use category priors [3,36,44] as the representative 3D shape per class, jointly reconstructing the full 3D shape and estimating the NOCS map from the full shape. Results show that this prior category helps improve the accuracy of the NOCS map and enhance the pose estimation performance. Other methods directly regress the pose or jointly utilize the correspondence representations [2,11,17]. + +# 2.2. Unsupervised Methods + +Given that annotating the 6D object pose and 3D size labels in the real world is expensive, time-consuming, and laborious, RePoNet [5] and SSC-6D [29] propose semi-supervised approaches that reconstruct the entire shape and use differential rendering [19, 20] techniques for self-training signals. These are category-specific methods and utilize multiple models, as many models as the number of categories is required. UDA-COPE [15] proposes an unsupervised domain adaptation method to mitigate the domain shift from the source to the target domain with a single model to estimate all categories efficiently. + +Although these works show reasonable performance without using pose labels, there is still a limitation in relying on 2D ground truth information (segmentation or bounding boxes) to train the segmentation network or pose network. Self-DPDN [16] shows a fully unsupervised method using inter/intra-consistency as a reconstructed shape in a self-supervised objective but requires a supervised loss for the source domain in the unsupervised learning process. Additionally, the self-training loss utilizes a full 3D shape and requires an additional reconstruction module for 3D shape. Most methods, except for UDA-COPE [15], jointly use the supervised loss using source label data during unsupervised learning to relax the unstable training. Furthermore, all the aforementioned unsupervised methods train their model in an offline manner and are unsuitable for online applications [24, 25]. + +# 2.3. Test-time Adaptation Methods + +Test-Time Adaptation (TTA) aims to enable the online adaptation of a pretrained model to the target domain without access to the source domain (source-free) [26,34,42,46]. The source data is commonly inaccessible during inference time because it is inefficient to train on a huge amount of source data every iteration. There may also be privacy or legal constraints to accessing source data [21], thus TTA is a more difficult but practical task than UDA. From the point of view of real-world applications, it is necessary to adapt to the new scene in an online way. Accordingly, test-time adaptation is necessary for the success of practical, real-world computer vision applications. Wang et al. propose Test entropy minimization (TENT) [42], which trains a net- + +work using a labeled source dataset, and adapts it to the unlabeled target dataset by updating the network parameters in batch norm layers using entropy loss. CoTTA [46] proposes a continual test-time adaptation method on the 2D classification and semantic segmentation tasks, and it effectively reduces the error accumulations while continually changing target data. Not limited to the 2D tasks, test-time adaptation methods have been applied to other tasks such as 3D segmentation [31, 33] and robot manipulation [22]. + +# 3. TTA-COPE + +Given an RGB-D image, our approach aims to estimate the 6D pose $T \in \mathrm{SE}(3)$ and size $s \in \mathbb{R}_+^3$ of each object. The object pose $T$ is defined as the rigid transformation $[R|t]$ , with rotation $R \in \mathrm{SO}(3)$ and translation $t \in \mathbb{R}^3$ . + +Our method consists of a two-stage learning scheme using source and target data, respectively. In the first stage, we train the network using labeled (synthetic) source data in a supervised manner (Sec. 3.1). This step is the same as supervised methods and results in a pretrained model. We then utilize this pretrained model for test-time adaptation (TTA) using unlabeled target data without accessing the source data. To the best of our knowledge, we are the first to propose test-time adaptation for category-level object pose estimation. Therefore, we study and explore how TTA baselines (Sec. 3.2) might apply to category-level object pose estimation. Finally, our proposed method is presented and explained (Sec. 3.3). + +# 3.1. Pretraining with Source Data + +In this section, we introduce an overview of our model and then describe how to train the model using labeled source data. We use UDA-COPE [15] network as a base model, which we supplement with batch normalization (BN) [10] in the 2D network to utilize BN updates for test-time adaptation. Fig. 2 shows an overview of our network, consisting of three steps to estimate the pose similar to correspondence methods [3,16,36]. First, we detect the bounding area of the object and segment the object surface using the off-the-shelf segmentation model from a single RGB image [8]. In the second stage, given a cropped 2D object image $I$ and segmented 3D object point cloud $D$ , our model estimates the correspondence as a NOCS map $N$ . We leverage the NOCS representation [43] to align diverse object instances within each class in a unified 3D space. In the final stage, our proposed pose ensemble method estimates the object pose $T$ . It simultaneously utilizes the predictions of the student and teacher model by calculating the pose confidence using the number of filtered points. + +The network is pretrained in a fully supervised manner using labeled source data. We minimize the predictions of the NOCS map $N$ using the cross-entropy loss $L_{CE}$ with ground truth labels $N^{GT}$ . We also use the consistency loss + +![](images/5b540a372d764dd3d44e2b4d4fc56f261f2aa6e1b5fb0c28cbeb70bb6acd9374.jpg) +Figure 2. Overview of our Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE) framework. Our method consists of three steps to estimate category-level object pose, similar to correspondence methods. 1) We detect and segment the object region using an off-the-shelf segmentation model from a single RGB image. 2) Given a cropped 2D object image $I$ and segmented 3D object point cloud $D$ , our model estimates the correspondence points as a NOCS map $N$ . 3) Our proposed pose ensemble estimates the object pose $T$ . It simultaneously utilizes the predictions of the student $\theta^S$ and teacher model $\theta^T$ by calculating the pose-aware confidence using the number of inlier points. We operate point filtering for each student and teacher model and select a confident pose $T$ with higher confidence (i.e., more inlier points) for which the pose estimation is more accurate. + +$L_{C}$ with 2D and 3D augmentation to make the network robust to noise in the input [15]. The supervised loss is formulated as: + +$$ +L _ {s u p} = \lambda_ {C E} L _ {C E} (N, N ^ {G T}) + \lambda_ {C} L _ {C} (N _ {a u g}, N), \tag {1} +$$ + +where $N_{aug}$ is the estimated NOCS map from augmented 2D or 3D input, and $\lambda_{CE},\lambda_C\in \mathbb{R}_+$ are weights. + +# 3.2. Test-time Adaptation with Target Data + +The model is updated every iteration while simultaneously estimating the object pose of the current scene. (Note that this is different from previous methods that update the model by running multiple epochs over the target data [16].) + +We have a design choice for the objective loss when updating the model during test time. One of the widely used test-time objective losses is the entropy loss proposed by TENT [42], + +$$ +L _ {e n t} = - \sum p \left(x _ {t}\right) \log p \left(x _ {t}\right), \tag {2} +$$ + +where $p(x_{t})$ is the probability of predictions $\theta^S (x_t)$ from target data $x_{t}$ . This simple objective loss encourages sharp distributions by assigning the most probability. + +Another common approach is utilizing pseudo labels from a teacher-student framework [35] with momentum update. The teacher model $\theta^T$ generates pseudo ground truth $\hat{y} = \theta^T(x_t)$ , where $x_t = (I, D)$ . The student model $\theta^S$ then uses the predictions of the teacher as the ground truth signals (pseudo labels) with cross-entropy loss, + +$$ +L _ {p l} = L _ {C E} \left(\theta^ {S} \left(x _ {t}\right), \hat {y}\right). \tag {3} +$$ + +After updating the student model $\theta_{i}^{S}\rightarrow \theta_{i + 1}^{S}$ by minimizing Eq. (3), the teacher model $\theta_{i + 1}^{T}$ is updated by momentum update, + +$$ +\theta_ {i + 1} ^ {T} \leftarrow \gamma \theta_ {i} ^ {T} + (1 - \gamma) \theta_ {i + 1} ^ {S}, \tag {4} +$$ + +where $i$ stands for the time stamp of the iteration and $\gamma$ is the momentum smoothing factor. The separation of the model structure mitigates some error accumulation by the pseudo label and momentum update. + +Although these objective strategies from previous work potentially provide self-training signals for test-time adaptation, they become brittle when directly applied to a category-level object pose estimation task. The pseudo labels $\hat{y}$ are still inherently exposed to noisy inputs and predictions, leading to errors in the student model $\theta^S$ . Since $\theta^T$ cannot guarantee clean pseudo labels for the student, we need the ability to filter noise in the input or predictions and to yield more reliable pseudo labels to the $\theta^S$ . + +# 3.3. Pose Ensemble + +To solve this problem, we propose to extend the pseudolabel filtering proposed by UDA-COPE [15] with pose ensemble processing. We observe that the UDA-COPE framework and pose filtering, while showing notable results, is inefficient for two reasons. First, the teacher model is only used for generating pseudo labels and not for inference. Second, the student model is also only used for inference and has less influence in generating pseudo labels. + +We propose instead to jointly utilize the teacher and student models for both generating pseudo labels and for infer + +ence, which we show in the experimental results produces a noticeable improvement. To this end, we design a pose ensemble module to simultaneously use teacher and student model predictions. Fig. 2 shows an overview of the pose ensemble with point filtering. + +The point filtering has three steps. Unlike UDA-COPE [15], we repeat these steps for both the student and teacher: 1) We initially estimate $T^S$ using the Umeyama algorithm $\theta_p$ within the region of object points $D$ and estimated NOCS map $N^S$ . 2) We transform observed depth to normalized object coordinate space $\widehat{D}^S$ by multiplying $(T^S)^{-1}$ and $D$ . 3) We compute each matching point distance between $N^S$ and $\widehat{D}$ , removing outliers that exceed a certain threshold $\rho$ . This whole process of point-filtering is then repeated using $N^T$ to estimate $T^T$ . This process is represented as follows, where $N^{S,T}$ is either $N^S$ or $N^T$ , and similarly for $T^{S,T}$ : + +$$ +\begin{array}{l} T ^ {S, T} = \theta_ {p} (D, N ^ {S, T}), \\ \widehat {D} ^ {S, T} = \left(T ^ {S, T}\right) ^ {- 1} D, \tag {5} \\ e _ {j} ^ {S, T} = \left\{ \begin{array}{l l} i n l i e r & \text {i f} \| \widehat {D} _ {j} ^ {S, T} \\ o u t l i e r & \text {o t h e r w i s e .} \end{array} \right. \\ \end{array} +$$ + +where $j = 1,\ldots ,X$ , and $X = |D| = |\widehat{D} |$ + +Our proposed pose ensemble aims to fuse two predictions $(N^S, N^T)$ to estimate the optimal pose. Although this fusion could be conducted at the input level before obtaining the $T$ from $\theta_p$ , we found that our proposal of fusing at the output level after getting $T$ produces slightly better results (see the ablation study, Sec. 4.5). We first estimate each pose $(T^S, T^T)$ from each prediction $(N^S, N^T)$ using $\theta_p$ . We then calculate the number of inliers $(inliers^S, inliers^T)$ from each pose using Eq. (5) to evaluate the confidence of the pose. We assume that the higher this confidence (i.e., the more inliers) is, the more accurate the pose estimation will be. Therefore: + +$$ +T _ {o u t} = \left\{ \begin{array}{l l} T ^ {S} & \text {i f i n l i e r s} ^ {S} > \text {i n l i e r s} ^ {T}, \\ T ^ {T} & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +We found that this simple pose-aware ensemble method is more effective for generating high-quality pseudo labels and inference than other ensemble methods (Sec. 4.5). + +Another crucial part of our method is a self-training loss. The total loss that we use to update the student network combines this self-training $L_{d}$ loss and the pseudo label loss $L_{pl}$ : + +$$ +L _ {t t a} = \lambda_ {d} L _ {d} + \lambda_ {p l} L _ {p l}. \tag {7} +$$ + +The first term is self-training loss $L_{d}$ , which helps learn the distribution of the target domain and reduce the domain gap + +through $T_{out}$ , which is obtained from the pose ensemble and observed $D$ . This loss is given by: + +$$ +L _ {d} = L _ {C E} \left(N _ {e} ^ {S}, \operatorname {E n c} \left(\widehat {D} _ {e}\right)\right), \tag {8} +$$ + +where $N_{e}^{S}$ is the inlier-only student NOCS map, $\widehat{D}_{e}$ is the resulting inliner-only transformed depth point cloud in NOCS space, and $\mathrm{Enc}(\cdot)$ is the one-hot encoding to enable cross-entropy loss. + +The second term + +$$ +L _ {p l} = L _ {C E} \left(N _ {e} ^ {S}, N _ {e} ^ {T}\right) \tag {9} +$$ + +is the pseudo label loss from UDA-COPE [15], where $N_{e}^{T}$ is the inlier-only teacher NOCS map. Since $N^{S} = \theta^{S}(x_{t})$ , and $\hat{y} = \theta^T (x_t)$ , $L_{pl}$ in Eq. (9) is the same as the loss in Eq. (3), except that here only inliers are considered. + +# 4. Experiments + +# 4.1. Dataset + +We utilize two widely used category-level pose estimation datasets as the source and target domain, respectively. The source data is the Context-Aware MixEd Reality (CAMERA) dataset [43], generated by rendering and compositing synthetic objects into real scenes while considering the context. CAMERA consists of 275K RGB-D images as a training set with 1,085 object instances chosen from six categories: bottle, bowl, camera, can, laptop and mug. We use the REAL dataset [43] as the target domain. The target data consists of 4,300 real-world images of seven scenes for training and 2,750 real-world images of six scenes for evaluation. We refer to REAL evaluation set as REAL275. Our TTA methods do not use the target training set (4,300 images with seven scenes) and only use the evaluation set (2,750 images with six scenes, REAL275) for test-time adaptation, e.g., TENT [42] process evaluation set in a sequential, online manner. On the other hand, unsupervised methods such as [5, 16, 29] use the target training set for updating the models without any restrictions on how much and how long they use the information. + +# 4.2. Implementation Details + +Object Segmentation. We use Mask R-CNN [8] to obtain the object area in 2D image. We use the identical results of Mask R-CNN for a fair comparison with previous methods [3, 17, 36] for semi-supervised and unsupervised settings. For the semi-supervised setting, Mask R-CNN has trained on the source, and target domain supervised manner. For the unsupervised setting, Mask R-CNN is trained on only the source domain. The detected area of the image resizes to $192 \times 192$ image resolution as the teacher or student model input. + +Table 2. Quantitative comparisons with state-of-the-art methods on the REAL275 dataset. + +
MethodSupervisedUnsupervisedmAP (↑)
SourceTargetTargetOnlineIoU50IoU755° 2cm5° 5cm10° 2cm10° 5 cm
Source (Supervised)
Metric Scale [14]2D/3D54.68.42.25.410.125.0
SPD [36]2D/3D50.517.011.512.133.037.9
Source (Supervised) & Target (Supervised)
NOCS [43]2D/3D2D/3D47.29.47.210.013.825.2
SPD [36]2D/3D2D/3D68.527.019.521.643.554.0
DualPoseNet [17]2D/3D2D/3D—30.829.335.950.066.8
CR-Net [44]2D/3D2D/3D—33.227.834.347.260.8
SGPA [3]2D/3D2D/3D—37.135.939.661.370.7
Source (Supervised) & Target (Semi-Supervised)
SSC-6D [29]2D/3D2D3DX73.0—16.819.644.154.5
RePoNet [5]2D/3D2D3DX76.0—30.733.9—63.0
UDA-COPE [15]2D/3D2D3DX75.534.430.534.957.066.1
Self-DPDN [16]2D/3D2D3DX75.241.639.545.063.372.2
TTA-COPE (Ours)2D/3D2D3D✓78.743.533.338.164.375.1
Source (Supervised) & Target (Unsupervised)
Self-DPDN [16]2D/3D2D/3DX67.243.939.046.761.873.4
TTA-COPE (Ours)2D/3D2D/3D✓69.139.730.235.961.773.2
+ +Student and Teacher. The student and teacher models have identical design in 2D and 3D branches from UDA-COPE [15]. We utilize the PSPNet [53] with ResNet34 [9] backbone for the 2D image feature extraction. For a 3D branch, we use the MinkowskiNet [4] and utilize sparse convolution operation with a $5\mathrm{cm}$ voxel size. Our NOCS representation uses the classification with 32 bins [43] instead of direct regression. During the pretraining stage, we train our model on the source data for 50 epochs using the Adam optimizer by initializing the learning rate of 1e-4 with a batch size of 32. The learning rate was reduced by a ratio of 0.6 (at 15k iterations), 0.3 (at 30k iterations), 0.1 (at 45k iterations), and 0.01 (at 60k iterations). During test-time adaptation, our student model uses the same learning rate, and the teacher smoothly updates using momentum update with $\gamma = 0.99$ . Given target data, we first update the model every iteration and then estimate the pose. We set $\lambda_{\mathrm{CE}} = 1.0$ , $\lambda_{\mathrm{C}} = 1\mathrm{e - }6$ , $\lambda_{\mathrm{d}} = 1.0$ , $\lambda_{\mathrm{pl}} = 1.0$ for our experiments. The point filtering threshold $\rho$ was set to 0.05 for all experiments. + +Metrics. To evaluate the performance of 3D object detection and 6D pose estimation, we follow the previous pose and size evaluation metric from Wang et al. [43]. We report the mean average precision (mAP) at the $50\%$ and $75\%$ intersection over union (IoU) thresholds for 3D object detection. We also report mAP for 6D object pose evaluation w.r.t. rotation and translation errors, where e.g. the $5^{\circ}5\mathrm{cm}$ metric describes the percentage of pose predictions where the error is less than both $5^{\circ}$ and $5\mathrm{cm}$ and the same for other thresholds. We recalculated all the metrics with the improved code. + +Table 3. Quantitative comparisons with TTA baselines for category-level object pose estimation on the REAL275 dataset. + +
MethodUnsupervisedIoU75mAP (↑)
Target 3DTarget 2D5° 2cm5° 5 cm
TENT [42] - Eq. (2)✓33.725.929.1
PL [35] - Eqs. (3)-(4)✓39.929.734.9
PL-F [15] - Eqs. (3)-(5)✓41.131.436.3
TTA-COPE (Ours)✓43.533.338.1
TENT [42] - Eq. (2)✓✓32.326.831.2
PL [35] - Eqs. (3)-(4)✓✓36.026.933.1
PL-F [15] - Eqs. (3)-(5)✓✓36.528.034.0
TTA-COPE (Ours)✓✓39.730.235.9
+ +# 4.3. Comparison with state-of-the-art + +Table 2 summarizes quantitative results on the REAL275 dataset under different settings: 1) supervised only on source data, 2) supervised on source and target data, 3) supervised on source data and semi-supervised on target data, and 4) supervised on source data and unsupervised on target data. Not surprisingly, methods following 2) such as SPD [36] perform better than setting 1) since they have the benefit of accessing both source and target data. Recent unsupervised domain adaptation (UDA) methods [15, 16] show remarkable results compared to state-of-the-art supervised methods without using target 3D labels (semi-supervised setting). Our method shows state-of-the-art results in IoU metrics under the semi-supervised setting and even outperforms the recent supervised methods SGPA [3] by a large margin (6.4 mAP in $\mathrm{IoU}_{75}$ ). Our TTA-COPE uses less time and data to train the target domain because of the advantage of test-time adaptation but achieves comparable + +Table 4. Ablation study on variants of pose ensemble methods and self-training loss under the semi-supervised setting. + +
MethodPose EnsemblemAP (↑)
InputOutputInferencePseudo LabelLd(8)IoU50IoU755° 2cm5° 5 cm10° 2cm10° 5 cm
LB76.237.529.134.662.173.2
(1)Argmax Match✓71.127.529.634.360.771.7
(2)Softmax Avg.✓78.740.432.137.364.074.7
(3)Softmax Max✓77.640.931.636.463.474.3
(4)Softmax Max✓77.541.132.137.164.374.6
(5)Inliers Max✓77.641.232.837.664.474.9
(6)Inliers Max✓✓78.242.932.437.864.174.7
(7)Argmax Match✓78.340.731.736.463.273.9
(8)Softmax Avg.✓78.641.131.536.263.674.5
(9)Softmax Max✓78.141.231.936.563.574.4
(10)Softmax Max✓78.641.231.836.363.574.4
(11)Inliers Max✓78.241.331.636.263.674.3
(12)Inliers Max✓✓78.743.032.637.063.573.9
OursInliers Max✓✓✓78.743.533.338.164.675.1
+ +results to the SOTA method, Self-DPDN [16]. Ours takes about 31 minutes (Table 5) for TTA and is $58\mathrm{x}$ faster than the Self-DPDN, which takes about 30 hours for training in the target domain, excluding inference time. + +# 4.4. Comparison with TTA baselines + +Table 3 summarizes results for the different design test-time adaptation (TTA) baselines in semi-supervised and unsupervised settings. TENT [42] has been designed for general unsupervised settings, but it is not specially designed for the object pose estimation task and naturally observes overall poor performance compared to other TTA baselines. Mean teacher with pseudo labels (PL) [35] performs better than TENT [42] in semi-supervised and unsupervised settings, and we believe that training the model using pseudo labels with momentum update (Eqs. (3)-(4)) provides more stable training signals by reducing the error accumulation than entropy minimization (2). However, using pseudo labels without filtering (PL) provides unreliable labels as the ground truth, we found that pseudo label filtering (PL-F) [15] performs better than PL, which indicates that noise filtering (Eq. (5)) improves the results for the student as expected. Finally, our TTA-COPE achieves state-of-the-art (SOTA) performance among all TTA baselines under both semi-supervised and unsupervised settings. + +# 4.5. Ablation Study + +In this section, we conduct experiments to evaluate the efficacy of the pose ensemble and self-training loss under a semi-supervised setting. + +Input/Output-Level Pose Ensemble. As mentioned in the Sec. 3.3, we compare the different pose ensemble methods given predictions of teacher and student models $(N^T, N^S)$ and answer the following question: Which ensemble is the most effective for test-time adaptation? We categorize two ensemble techniques: + +1) Input-level ensemble that fuses two NOCS map predictions $(N^T, N^S)$ and makes an accurate single NOCS map $N$ to estimate poses using the Umayama algorithm $\theta_p$ . For the input-level ensemble that fuses two predictions $(N^T, N^S)$ into one prediction $N$ , we compare three strategies: Argmax Match, Softmax Average, and Softmax Max operations [12, 33]. +2) We ensemble output-level predictions $(N^T, N^S)$ to estimate each pose $(T^T, T^S)$ and choose the best pose predictions under specific criteria. Unlike the input ensemble, only the Softmax Max operation is valid for the output-level ensemble due to the nonlinearity of poses. + +Table 4 summarizes our ensemble ablation study results. Table 4-(1-3) shows that the Softmax Avg. and Softmax Max perform better than the Argmax Match operation among input-level pose ensembles. If we compare the same operation (Softmax Max) in the input-level, Table 4-(3), and output-level, Table 4-(4), the output-level ensemble shows slightly better performance since the output-level operation is directly related to the pose results. Our proposed pose ensemble, Table 4-(5), considers pose-aware confidence by choosing a higher number of inlier points and achieves the best performance in all pose ensembles in Table 4-(1-5). + +Use of Pose Ensemble. Our pose ensemble can be used for inference Table 4-(1-6, Ours) as well as generating pseudo labels Table 4-(7-12, Ours). The results show that our pose ensemble is helpful for both cases. In particular, applying our ensemble to inferences Table 4-(5) has been shown to be more effective than application to generating pseudo labels Table 4-(11) in pose metric. We also observe that our pose ensemble used for inference or pseudo labels improves overall metrics compared to the lower bound (LB) that is only trained on source data without test-time adaptation. + +Effect of self-training loss. The results of Table 4-(6) and Table 4-(12) show the effect of self-training loss Eq. (8). It yields an improvement of more than $1.7\mathrm{mAP}$ compared + +![](images/fb14cff0bc335d4c526e0afbca505e83268ed605adcbeeccf42cf44e01fdc031.jpg) +Figure 3. Qualitative comparison with the lower bound (LB) and our TTA-COPE under the semi-supervised setting. + +Table 5. Ablation study on different updating intervals. + +
MethodTTA TimeIoU755° 2cm5° 5 cm
Interval=1 (Ours)31 min43.533.338.1
Interval=1016 min41.832.737.7
Interval=2015 min41.131.937.1
+ +to without self-training loss results of Table 4-(5, 11) in the $\mathrm{IoU}_{75}$ metric. This indicates that the self-learning loss helps to learn the distribution of the target domain and reduces the domain shift by comparing the predicted NOCS map against the observed point cloud in the target domain. The result in the last column is our proposed model for test-time adaptation, Fig. 3 shows qualitative results against the lower bound method. + +Updating Interval. TTA is inevitably slower than simple inference since additional training is required. However, reducing the update interval enables faster TTA time than TTA baselines because of reduced training time. Table 5 shows the difference in TTA speed and performance according to the updating interval. Specifically, the interval of every 10 frames improves the speed of TTA roughly two-fold compared to Interval 1 (Ours), with a marginal performance drop. We also increase intervals to 20, but it does not show as much improvement as before since most of the bottleneck arises from inference time. + +# 5. Conclusion + +We have proposed TTA-COPE, a test-time adaptation method for category-level object pose estimation that ad + +dresses the source-to-target domain shift without accessing source data at test time and without labeled target data. Specifically, we designed a pose ensemble method with self-training for test-time adaptation that simultaneously uses the teacher-student model to generate robust pseudo labels and estimate accurate poses for inference. We explore limitations of several test-time adaptation baselines and show that the proposed method achieves state-of-the-art performance. We demonstrate the benefits of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings. + +To the best of our knowledge, TTA-COPE is the first approach that tries to solve test-time adaptation for category-level object pose estimation. Since our method currently focuses on generating 6D pose labels, it does not affect 2D labels and the segmentation model. In future work, when jointly considering 2D label and Mask R-CNN, greater performance improvement and stable test-time adaptation would be possible. Also, our pose estimation relies on non-differentiable pose estimation, and as such we could benefit from a differentiable pose estimation method [30]. + +# Acknowledgment + +This work was part of Taeyeop Lee's internship at NVIDIA and was also partially supported by an Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korean government(MSIT) (No.2020-0-00440). + +# References + +[1] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 1 +[2] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11973-11982, 2020. 1, 2, 3 +[3] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 1, 2, 3, 5, 6 +[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3075-3084, 2019. 6 +[5] Yang Fu and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. arXiv preprint arXiv:2206.15436, 2022. 2, 3, 5, 6 +[6] Wei Gao and Russ Tedrake. kpm 2.0: Feedback control for category-level robotic manipulation. IEEE Robotics and Automation Letters (RA-L), 6(2):2962-2969, 2021. 1 +[7] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2 +[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 3, 5 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[10] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (ICML), pages 448-456, 2015. 3 +[11] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. arXiv preprint arXiv:2207.13691, 2022. 3 +[12] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12605-12614, 2020. 7 +[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, + +Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6d pose estimation of novel objects via render & compare. In 6th Annual Conference on Robot Learning (CoRL), 2022. 1 +[14] Taeyeop Lee, Byeong-Uk Lee, Myungchul Kim, and In So Kweon. Category-level metric scale object shape and pose estimation. IEEE Robotics and Automation Letters (RA-L), 6(4):8575-8582, 2021. 2, 6 +[15] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: Unsupervised domain adaptation for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14891-14900, 2022. 2, 3, 4, 5, 6, 7 +[16] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2022. 2, 3, 4, 5, 6, 7 +[17] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 1, 2, 3, 5, 6 +[18] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an rgb image. In IEEE International Conference on Robotics and Automation (ICRA). ICRA, 2022. 1 +[19] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 3 +[20] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2020. 3 +[21] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1215-1224, 2021. 2, 3 +[22] Massimiliano Mancini, Hakan Karaoguz, Elisa Ricci, Patric Jensfelt, and Barbara Caputo. Kitting in the wild through online domain adaptation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1103-1109. IEEE, 2018. 3 +[23] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics (TVCG), 22(12):2633-2651, 2015. 1 +[24] Eitan Marder-Eppstein. Project Tango. In ACM SIGGRAPH Real-Time Live!, page 25, 2016. 1, 3 +[25] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-DoF GraspNet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision (ICCV), pages 2901-2910, 2019. 1, 3 +[26] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International Conference on Machine Learning (ICML), pages 16888-16905, 2022. 3 +[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7668-7677, 2019. 1 +[28] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-wise voting network for 6DoF pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4561-4570, 2019. 1 +[29] Wanli Peng, Jianhang Yan, Hongtao Wen, and Yi Sun. Self-supervised category-level 6D object pose estimation with deep implicit shape representation. In Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI), volume 36, pages 2082-2090, 2022. 2, 3, 5, 6 +[30] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky T. Q. Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 2022. 8 +[31] Mihir Prabhudesai, Sujoy Paul, Sjoerd van Steenkiste, Mehdi SM Sajjadi, Anirudh Goyal, Deepak Pathak, Katerina Fragkiadaki, Gaurav Aggarwal, and Thomas Kipf. Test-time adaptation with slot-centric models. 2022. 3 +[32] Martin Runz, Maud Buffer, and Lourdes Agapito. Mask-Fusion: Real-time recognition, tracking and reconstruction of multiple moving objects. In IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 10-20, 2018. 1 +[33] Inkyu Shin, Yi-Hsuan Tsai, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Sparsh Garg, In So Kweon, and Kuk-Jin Yoon. MM-TTA: Multi-modal test-time adaptation for 3D semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16928-16937, 2022. 3, 7 +[34] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. arXiv preprint arXiv:2303.01904, 2023. 3 +[35] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 4, 6, 7 +[36] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 5, 6 +[37] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cam + +eracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 969-977, 2018. 1 +[38] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In Conference on Robot Learning (CoRL), 2018. 1 +[39] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-DoF pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 1 +[40] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence (TPAMI), 13(04):376-380, 1991. 2 +[41] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3343–3352, 2019. 1 +[42] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations (ICLR), 2021. 2, 3, 4, 5, 6, 7 +[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 1, 2, 3, 5, 6 +[44] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 1, 2, 3, 6 +[45] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21222-21231, 2022. 1 +[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7201-7211, 2022. 2, 3 +[47] Bowen Wen and Kostas Bekris. BundleTrack: 6D pose tracking for novel objects without instance or category-level 3D models. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074, 2021. 1 +[48] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. Robotics: Science and Systems (RSS), 2022. 1 + +[49] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1 +[50] Bowen Wen, Chaitanya Mitash, Sruthi Soorian, Andrew Kimmel, Avishai Sintov, and Kostas E Bekris. Robust, occlusion-aware pose estimation for objects grasped by adaptive hands. In IEEE International Conference on Robotics and Automation (ICRA), pages 6210-6217. IEEE, 2020. 1 +[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. Robotics: Science and Systems (RSS), 2018. 1 +[52] He Yisheng, Wang Yao, Fan Haoqiang, Chen Qifeng, and Sun Jian. Fs6d: Few-shot 6d pose estimation of novel objects. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[53] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6 \ No newline at end of file diff --git a/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/images.zip b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a1b786dea2ccf61ca531ce4ad4d1b6f1a500396 --- /dev/null +++ b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03a52e50cbec10959ea2ffbdf4cbcca18bb7d622de59cce4f893705225baf004 +size 686525 diff --git a/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/layout.json b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a0f3319db9a673c4c11cb1ea094a6bca316ba9d --- /dev/null +++ b/2023/TTA-COPE_ Test-Time Adaptation for Category-Level Object Pose Estimation/layout.json @@ -0,0 +1,8918 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "text", + "content": "TTA-COPE: Test-Time Adaptation for Category-Level Object Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": "Taeyeop Lee" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Jonathan Tremblay" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Valts Blukis" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Bowen Wen" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Byeong-Uk Lee" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Inkyu Shin" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Stan Birchfield" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " In So Kweon" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " Kuk-Jin Yoon" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": " KAIST " + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 184 + ], + "type": "text", + "content": "NVIDIA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 430 + ], + "type": "text", + "content": "Test-time adaptation methods have been gaining attention recently as a practical solution for addressing source-to-target domain gaps by gradually updating the model without requiring labels on the target data. In this paper, we propose a method of test-time adaptation for category-level object pose estimation called TTA-COPE. We design a pose ensemble approach with a self-training loss using pose-aware confidence. Unlike previous unsupervised domain adaptation methods for category-level object pose estimation, our approach processes the test data in a sequential, online manner, and it does not require access to the source domain at runtime. Extensive experimental results demonstrate that the proposed pose ensemble and the self-training loss improve category-level object pose performance during test time under both semi-supervised and unsupervised settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 128, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 128, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 128, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 605 + ], + "type": "text", + "content": "Object pose estimation is a crucial problem in computer vision and robotics. Advanced methods that focus on diverse variations of object 6D pose estimation have been introduced, such as known 3D objects (instance-level) [28, 38], category-level [18, 36, 43], few-shot [52], and zero-shot pose estimation [13, 47]. These techniques are useful for downstream applications requiring an online operation, such as robotic manipulation [6, 25, 48] and augmented reality [23, 24, 32]. Our paper focuses on the category-level object pose estimation problem since it is more broadly applicable than the instance-level problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Many works on category-level object pose estimation [2, 3, 17, 18, 36, 43, 44] have been proposed recently. These approaches estimate multiple classes of object pose more efficiently in a single network compared to the instance-level object pose estimation methods [27, 38, 41, 49-51], which depend on known 3D shape knowledge and the size of the objects. Notably, Wang et al. [43] introduced a novel representation called Normalized Object Coordinate Space (NOCS) to align various object instances within each" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 228, + 435, + 310 + ], + "blocks": [ + { + "bbox": [ + 352, + 217, + 412, + 228 + ], + "lines": [ + { + "bbox": [ + 352, + 217, + 412, + 228 + ], + "spans": [ + { + "bbox": [ + 352, + 217, + 412, + 228 + ], + "type": "text", + "content": "No Adaptation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 228, + 435, + 310 + ], + "lines": [ + { + "bbox": [ + 326, + 228, + 435, + 310 + ], + "spans": [ + { + "bbox": [ + 326, + 228, + 435, + 310 + ], + "type": "image", + "image_path": "8ef1e6bd5b3704e36771617886e8f39f7d9331b1e70e077523e0cc703f2436cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 438, + 228, + 545, + 309 + ], + "blocks": [ + { + "bbox": [ + 447, + 217, + 536, + 228 + ], + "lines": [ + { + "bbox": [ + 447, + 217, + 536, + 228 + ], + "spans": [ + { + "bbox": [ + 447, + 217, + 536, + 228 + ], + "type": "text", + "content": "Test-Time Adaptation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 228, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 438, + 228, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 438, + 228, + 545, + 309 + ], + "type": "image", + "image_path": "c3dbda982517c5473aedb4167fbacb6a9599b1521b7f18ec04a738358870136e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 312, + 435, + 392 + ], + "blocks": [ + { + "bbox": [ + 328, + 312, + 435, + 392 + ], + "lines": [ + { + "bbox": [ + 328, + 312, + 435, + 392 + ], + "spans": [ + { + "bbox": [ + 328, + 312, + 435, + 392 + ], + "type": "image", + "image_path": "fc3ea698be5ab3f71879da2fdc3a2beaa84e91fdf69ce9c75577a69518efe261.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 438, + 312, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 438, + 312, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 438, + 312, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 438, + 312, + 545, + 392 + ], + "type": "image", + "image_path": "980c223f2f6155f6f0171b6eaa165cd5e91f23687d74a43d4a05297da5750bde.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 327, + 396, + 435, + 476 + ], + "blocks": [ + { + "bbox": [ + 327, + 396, + 435, + 476 + ], + "lines": [ + { + "bbox": [ + 327, + 396, + 435, + 476 + ], + "spans": [ + { + "bbox": [ + 327, + 396, + 435, + 476 + ], + "type": "image", + "image_path": "cbe5bf23a24cc471ea48dd2491de9cb4b7e5fe2518cb029ef6983dea926ae2f2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 479, + 547, + 567 + ], + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 567 + ], + "type": "text", + "content": "Figure 1. We propose a Test-Time Adaptation for Category-level Object Pose Estimation framework (TTA-COPE) that automatically improves the network in an online manner without labeled target data. As new image frames are processed, our method fine-tunes the network using the unlabeled data and simultaneously applies the network to perform pose estimation via inference. This approach successfully handles domain shifts compared with no adaptation, as seen here." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 438, + 396, + 545, + 476 + ], + "blocks": [ + { + "bbox": [ + 438, + 396, + 545, + 476 + ], + "lines": [ + { + "bbox": [ + 438, + 396, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 438, + 396, + 545, + 476 + ], + "type": "image", + "image_path": "01f64bb8e9793d621a001de91fa2b851caa9e0dc20e4981448cab24545eebcdf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 616 + ], + "type": "text", + "content": "category in a canonical 3D space. The strengths of the NOCS representation have led to its adoption by follow-up work [3, 17, 36]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": "In order to obtain accurate category-level object pose methods in unseen real-world scenarios, it is desirable to fine-tune the models in the new environment with labeled target data. The model that is not fine-tuned on the target domain distribution will almost certainly exhibit lower performance than the fine-tuned model [37]. However, annotating 6D poses of objects in the target environment is an expensive process [1, 39, 43, 45] that we seek to avoid." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21285" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 177, + 285, + 262 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 170 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 170 + ], + "type": "text", + "content": "Table 1. Comparison with prior unsupervised works for category-level object pose estimation. Our unsupervised method trains models without 2D or 3D labels of target data, similar to Self-DPDN [16]. Unlike previous methods, our proposed approach updates the model online without offline batch processing. Moreover, we do not use the source data during test time (source-free) because it is impractical to train on a large amount of source data every iteration. There also may be privacy or legal constraints to access source data [21]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 177, + 285, + 262 + ], + "lines": [ + { + "bbox": [ + 50, + 177, + 285, + 262 + ], + "spans": [ + { + "bbox": [ + 50, + 177, + 285, + 262 + ], + "type": "table", + "html": "
MethodUnsupervisedTest-time Adaptation
Target 3DTarget 2DSource-FreeOnline Adaptation
SupervisedXXXX
SSC-6D [29]✓XXX
RePoNet [5]✓XXX
UDA-COPE [15]✓X✓X
Self-DPDN [16]✓✓XX
Ours✓✓✓✓
", + "image_path": "3ca988fda7aaed4e290bb6ae698a22957cefff6a35bf7726f7c9c213a2cf8456.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "type": "text", + "content": "Compared to annotating in 2D space, labeling in 3D space requires specific knowledge about geometry [7] from the annotator and is much more laborious, time-consuming, and error-prone due to the complex nature of " + }, + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 46, + 280, + 287, + 353 + ], + "type": "text", + "content": " space. Therefore, it is usually challenging to annotate real-world data with 3D annotations for fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 495 + ], + "type": "text", + "content": "In order to solve the aforementioned problem of annotating object pose data in the real world, several recent methods [5, 15, 16] propose unsupervised domain adaptation (UDA) that aims to train the network without utilizing the ground truth of target pose labels. Although they show promising results using UDA techniques, these approaches still do not meet some of the requirements for online applications. For example, when a robot encounters a new environment, it is desirable to adapt the scene online manner while estimating object poses rather than waiting for enough data to be collected in the novel scene to train the model offline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 688 + ], + "type": "text", + "content": "This problem definition of online fine-tuning is more practical for real applications, where we desire to update the model instantly when new data becomes available for fast domain adaptation. This setting is known as test-time adaptation (TTA) [42]. For TTA, the requirements are as follows: 1) labeled source data should not be accessed at test time, 2) adaptation should be online (rather than offline batch processing), and 3) the method should be fully unsupervised, without using 2D or 3D target labels during online fine-tuning. Since we do not have access to labeled source data (source-free) at test time this problem is more challenging than existing unsupervised category-level object pose methods [5,15,16,29]. Table 1 summarizes the difference between our problem definition and existing methods, showing that test-time adaptation for category-level object pose estimation remains an open problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "In this paper, we propose Test-time Adaptation for Category-level Object Pose Estimation (TTA-COPE) to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "handle domain shifts without any target domain annotations (see Fig. 1). Prior works on general test-time adaptation [42,46] propose self-training to minimize entropy loss. TENT [42] has shown improvement in 2D classification and segmentation tasks. We show, however, that simply extending TENT for the category-level object pose estimation is not effective. Another self-training strategy is the teacher-student framework [35] with pseudo labels. However, since pseudo labels are created without any noise filtering, naive pseudo labels may be unreliable and cause convergence to a suboptimal model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 204, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 371 + ], + "type": "text", + "content": "To tackle this problem, we design a novel pose ensemble method to perform test-time adaptation for category-level object pose estimation by extending the pose-aware filtering of UDA-COPE [15]. The proposed method uses an ensemble of teacher-student predictions based on pose-aware confidence, which is used both for generating pseudo labels and inference. Also, the pose ensemble helps to train models with additional self-training loss to reduce the domain shift for category-level pose estimation by using pose-aware confidence. We demonstrate the advantages of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings. We show that our TTA-COPE framework achieves state-of-the-art performance compared to strong TTA baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 372, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 545, + 394 + ], + "type": "text", + "content": "In summary, the main contributions of our work are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 403, + 545, + 544 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 318, + 403, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 403, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 318, + 403, + 545, + 450 + ], + "type": "text", + "content": "- We propose Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE), which handles domain shifts without labeling target data and without accessing source data during test time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 450, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 450, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 318, + 450, + 545, + 496 + ], + "type": "text", + "content": "- We introduce a pose ensemble with self-training loss that utilizes the teacher-student predictions to generate robust pseudo labels and estimates accurate poses for inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 497, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 497, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 318, + 497, + 545, + 544 + ], + "type": "text", + "content": "- We evaluate our framework with experimental comparisons against strong test-time baselines and state-of-the-art methods under both semi-supervised and unsupervised settings." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 555, + 397, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 397, + 568 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 397, + 568 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 575, + 424, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 424, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 424, + 588 + ], + "type": "text", + "content": "2.1. Supervised Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Fully supervised learning methods for category-level object pose estimation [2,3,17,36,43,44] train their models using labeled source data (e.g., synthetic) and target data (e.g., real). Most category-level 6D object pose and size estimation approaches [3,14,36,43,44] use the dense correspondence via Normalized Object Coordinate Space (NOCS) representation as a common way to estimate pose and size. These correspondence-based methods initially estimate the NOCS map from RGB or depth images. Afterwards, the Umayama algorithm [40] with RANSAC is used to estimate" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21286" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "optimal poses and object sizes by minimizing distances between depth and estimated NOCS map." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 180 + ], + "type": "text", + "content": "Some methods use category priors [3,36,44] as the representative 3D shape per class, jointly reconstructing the full 3D shape and estimating the NOCS map from the full shape. Results show that this prior category helps improve the accuracy of the NOCS map and enhance the pose estimation performance. Other methods directly regress the pose or jointly utilize the correspondence representations [2,11,17]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 188, + 178, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 178, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 178, + 201 + ], + "type": "text", + "content": "2.2. Unsupervised Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 207, + 286, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 286, + 338 + ], + "type": "text", + "content": "Given that annotating the 6D object pose and 3D size labels in the real world is expensive, time-consuming, and laborious, RePoNet [5] and SSC-6D [29] propose semi-supervised approaches that reconstruct the entire shape and use differential rendering [19, 20] techniques for self-training signals. These are category-specific methods and utilize multiple models, as many models as the number of categories is required. UDA-COPE [15] proposes an unsupervised domain adaptation method to mitigate the domain shift from the source to the target domain with a single model to estimate all categories efficiently." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 286, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 286, + 531 + ], + "type": "text", + "content": "Although these works show reasonable performance without using pose labels, there is still a limitation in relying on 2D ground truth information (segmentation or bounding boxes) to train the segmentation network or pose network. Self-DPDN [16] shows a fully unsupervised method using inter/intra-consistency as a reconstructed shape in a self-supervised objective but requires a supervised loss for the source domain in the unsupervised learning process. Additionally, the self-training loss utilizes a full 3D shape and requires an additional reconstruction module for 3D shape. Most methods, except for UDA-COPE [15], jointly use the supervised loss using source label data during unsupervised learning to relax the unstable training. Furthermore, all the aforementioned unsupervised methods train their model in an offline manner and are unsuitable for online applications [24, 25]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 212, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 212, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 212, + 552 + ], + "type": "text", + "content": "2.3. Test-time Adaptation Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 713 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA) aims to enable the online adaptation of a pretrained model to the target domain without access to the source domain (source-free) [26,34,42,46]. The source data is commonly inaccessible during inference time because it is inefficient to train on a huge amount of source data every iteration. There may also be privacy or legal constraints to accessing source data [21], thus TTA is a more difficult but practical task than UDA. From the point of view of real-world applications, it is necessary to adapt to the new scene in an online way. Accordingly, test-time adaptation is necessary for the success of practical, real-world computer vision applications. Wang et al. propose Test entropy minimization (TENT) [42], which trains a net-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "work using a labeled source dataset, and adapts it to the unlabeled target dataset by updating the network parameters in batch norm layers using entropy loss. CoTTA [46] proposes a continual test-time adaptation method on the 2D classification and semantic segmentation tasks, and it effectively reduces the error accumulations while continually changing target data. Not limited to the 2D tasks, test-time adaptation methods have been applied to other tasks such as 3D segmentation [31, 33] and robot manipulation [22]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 190, + 382, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 382, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 382, + 202 + ], + "type": "text", + "content": "3. TTA-COPE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": "Given an RGB-D image, our approach aims to estimate the 6D pose " + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "inline_equation", + "content": "T \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": " and size " + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": " of each object. The object pose " + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": " is defined as the rigid transformation " + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "inline_equation", + "content": "[R|t]" + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": ", with rotation " + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "inline_equation", + "content": "R \\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 209, + 545, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 413 + ], + "type": "text", + "content": "Our method consists of a two-stage learning scheme using source and target data, respectively. In the first stage, we train the network using labeled (synthetic) source data in a supervised manner (Sec. 3.1). This step is the same as supervised methods and results in a pretrained model. We then utilize this pretrained model for test-time adaptation (TTA) using unlabeled target data without accessing the source data. To the best of our knowledge, we are the first to propose test-time adaptation for category-level object pose estimation. Therefore, we study and explore how TTA baselines (Sec. 3.2) might apply to category-level object pose estimation. Finally, our proposed method is presented and explained (Sec. 3.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 420, + 466, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 466, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 466, + 433 + ], + "type": "text", + "content": "3.1. Pretraining with Source Data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": "In this section, we introduce an overview of our model and then describe how to train the model using labeled source data. We use UDA-COPE [15] network as a base model, which we supplement with batch normalization (BN) [10] in the 2D network to utilize BN updates for test-time adaptation. Fig. 2 shows an overview of our network, consisting of three steps to estimate the pose similar to correspondence methods [3,16,36]. First, we detect the bounding area of the object and segment the object surface using the off-the-shelf segmentation model from a single RGB image [8]. In the second stage, given a cropped 2D object image " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": " and segmented 3D object point cloud " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", our model estimates the correspondence as a NOCS map " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ". We leverage the NOCS representation [43] to align diverse object instances within each class in a unified 3D space. In the final stage, our proposed pose ensemble method estimates the object pose " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ". It simultaneously utilizes the predictions of the student and teacher model by calculating the pose confidence using the number of filtered points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The network is pretrained in a fully supervised manner using labeled source data. We minimize the predictions of the NOCS map " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " using the cross-entropy loss " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{CE}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " with ground truth labels " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "N^{GT}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". We also use the consistency loss" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21287" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 67, + 506, + 257 + ], + "blocks": [ + { + "bbox": [ + 86, + 67, + 506, + 257 + ], + "lines": [ + { + "bbox": [ + 86, + 67, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 86, + 67, + 506, + 257 + ], + "type": "image", + "image_path": "5b540a372d764dd3d44e2b4d4fc56f261f2aa6e1b5fb0c28cbeb70bb6acd9374.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": "Figure 2. Overview of our Test-Time Adaptation for Category-level Object Pose Estimation (TTA-COPE) framework. Our method consists of three steps to estimate category-level object pose, similar to correspondence methods. 1) We detect and segment the object region using an off-the-shelf segmentation model from a single RGB image. 2) Given a cropped 2D object image " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": " and segmented 3D object point cloud " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": ", our model estimates the correspondence points as a NOCS map " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": ". 3) Our proposed pose ensemble estimates the object pose " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": ". It simultaneously utilizes the predictions of the student " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\theta^S" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": " and teacher model " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\theta^T" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": " by calculating the pose-aware confidence using the number of inlier points. We operate point filtering for each student and teacher model and select a confident pose " + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 264, + 545, + 342 + ], + "type": "text", + "content": " with higher confidence (i.e., more inlier points) for which the pose estimation is more accurate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "inline_equation", + "content": "L_{C}" + }, + { + "bbox": [ + 47, + 354, + 287, + 389 + ], + "type": "text", + "content": " with 2D and 3D augmentation to make the network robust to noise in the input [15]. The supervised loss is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 396, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 396, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 61, + 396, + 287, + 409 + ], + "type": "interline_equation", + "content": "L _ {s u p} = \\lambda_ {C E} L _ {C E} (N, N ^ {G T}) + \\lambda_ {C} L _ {C} (N _ {a u g}, N), \\tag {1}", + "image_path": "068fa0677c1dd3f10f76f47cf8fbd719a5b500921b940083b9fef1c18bf88a08.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "inline_equation", + "content": "N_{aug}" + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": " is the estimated NOCS map from augmented 2D or 3D input, and " + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\lambda_{CE},\\lambda_C\\in \\mathbb{R}_+" + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": " are weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 251, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 251, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 251, + 458 + ], + "type": "text", + "content": "3.2. Test-time Adaptation with Target Data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 287, + 511 + ], + "type": "text", + "content": "The model is updated every iteration while simultaneously estimating the object pose of the current scene. (Note that this is different from previous methods that update the model by running multiple epochs over the target data [16].)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 511, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 559 + ], + "type": "text", + "content": "We have a design choice for the objective loss when updating the model during test time. One of the widely used test-time objective losses is the entropy loss proposed by TENT [42]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 563, + 287, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 563, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 287, + 581 + ], + "type": "interline_equation", + "content": "L _ {e n t} = - \\sum p \\left(x _ {t}\\right) \\log p \\left(x _ {t}\\right), \\tag {2}", + "image_path": "bd51416b37e119c016e293db4e36e8db6b71a240f6ed2dae9bbeaa8b15017dc8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "p(x_{t})" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": " is the probability of predictions " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\theta^S (x_t)" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": " from target data " + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": ". This simple objective loss encourages sharp distributions by assigning the most probability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "text", + "content": "Another common approach is utilizing pseudo labels from a teacher-student framework [35] with momentum update. The teacher model " + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\theta^T" + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "text", + "content": " generates pseudo ground truth " + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\theta^T(x_t)" + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "inline_equation", + "content": "x_t = (I, D)" + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "text", + "content": ". The student model " + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "inline_equation", + "content": "\\theta^S" + }, + { + "bbox": [ + 47, + 624, + 287, + 695 + ], + "type": "text", + "content": " then uses the predictions of the teacher as the ground truth signals (pseudo labels) with cross-entropy loss," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 118, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "L _ {p l} = L _ {C E} \\left(\\theta^ {S} \\left(x _ {t}\\right), \\hat {y}\\right). \\tag {3}", + "image_path": "562ec654b019e66a392156e3c4eaf583b99cf0ef038ef8715877f8a8c2fc0fa0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "type": "text", + "content": "After updating the student model " + }, + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{S}\\rightarrow \\theta_{i + 1}^{S}" + }, + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "type": "text", + "content": " by minimizing Eq. (3), the teacher model " + }, + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\theta_{i + 1}^{T}" + }, + { + "bbox": [ + 305, + 354, + 545, + 389 + ], + "type": "text", + "content": " is updated by momentum update," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 389, + 545, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 389, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 367, + 389, + 545, + 404 + ], + "type": "interline_equation", + "content": "\\theta_ {i + 1} ^ {T} \\leftarrow \\gamma \\theta_ {i} ^ {T} + (1 - \\gamma) \\theta_ {i + 1} ^ {S}, \\tag {4}", + "image_path": "87d9b81d20afc10d527ead8cf49bcc3e97f0996914337ed565fe399974ae377a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "type": "text", + "content": " stands for the time stamp of the iteration and " + }, + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 410, + 545, + 458 + ], + "type": "text", + "content": " is the momentum smoothing factor. The separation of the model structure mitigates some error accumulation by the pseudo label and momentum update." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "text", + "content": "Although these objective strategies from previous work potentially provide self-training signals for test-time adaptation, they become brittle when directly applied to a category-level object pose estimation task. The pseudo labels " + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "text", + "content": " are still inherently exposed to noisy inputs and predictions, leading to errors in the student model " + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\theta^S" + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\theta^T" + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "text", + "content": " cannot guarantee clean pseudo labels for the student, we need the ability to filter noise in the input or predictions and to yield more reliable pseudo labels to the " + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\theta^S" + }, + { + "bbox": [ + 305, + 459, + 545, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 574, + 398, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 398, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 398, + 586 + ], + "type": "text", + "content": "3.3. Pose Ensemble" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "To solve this problem, we propose to extend the pseudolabel filtering proposed by UDA-COPE [15] with pose ensemble processing. We observe that the UDA-COPE framework and pose filtering, while showing notable results, is inefficient for two reasons. First, the teacher model is only used for generating pseudo labels and not for inference. Second, the student model is also only used for inference and has less influence in generating pseudo labels." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We propose instead to jointly utilize the teacher and student models for both generating pseudo labels and for infer" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21288" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "ence, which we show in the experimental results produces a noticeable improvement. To this end, we design a pose ensemble module to simultaneously use teacher and student model predictions. Fig. 2 shows an overview of the pose ensemble with point filtering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": "The point filtering has three steps. Unlike UDA-COPE [15], we repeat these steps for both the student and teacher: 1) We initially estimate " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "T^S" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " using the Umeyama algorithm " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " within the region of object points " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " and estimated NOCS map " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "N^S" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ". 2) We transform observed depth to normalized object coordinate space " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "\\widehat{D}^S" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " by multiplying " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "(T^S)^{-1}" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ". 3) We compute each matching point distance between " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "N^S" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "\\widehat{D}" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ", removing outliers that exceed a certain threshold " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ". This whole process of point-filtering is then repeated using " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "N^T" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " to estimate " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "T^T" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ". This process is represented as follows, where " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "N^{S,T}" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " is either " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "N^S" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "N^T" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ", and similarly for " + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "inline_equation", + "content": "T^{S,T}" + }, + { + "bbox": [ + 46, + 133, + 288, + 277 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 286, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 286, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 59, + 286, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T ^ {S, T} = \\theta_ {p} (D, N ^ {S, T}), \\\\ \\widehat {D} ^ {S, T} = \\left(T ^ {S, T}\\right) ^ {- 1} D, \\tag {5} \\\\ e _ {j} ^ {S, T} = \\left\\{ \\begin{array}{l l} i n l i e r & \\text {i f} \\| \\widehat {D} _ {j} ^ {S, T} \\\\ o u t l i e r & \\text {o t h e r w i s e .} \\end{array} \\right. \\\\ \\end{array}", + "image_path": "6b59ad3da650c6a74da45a0a71932f04229f14fdbc7213d11ecf6cc24937af90.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 363, + 219, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 219, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 219, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 363, + 219, + 376 + ], + "type": "inline_equation", + "content": "j = 1,\\ldots ,X" + }, + { + "bbox": [ + 47, + 363, + 219, + 376 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 47, + 363, + 219, + 376 + ], + "type": "inline_equation", + "content": "X = |D| = |\\widehat{D} |" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": "Our proposed pose ensemble aims to fuse two predictions " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "(N^S, N^T)" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " to estimate the optimal pose. Although this fusion could be conducted at the input level before obtaining the " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": ", we found that our proposal of fusing at the output level after getting " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " produces slightly better results (see the ablation study, Sec. 4.5). We first estimate each pose " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "(T^S, T^T)" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " from each prediction " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "(N^S, N^T)" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": ". We then calculate the number of inliers " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "(inliers^S, inliers^T)" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " from each pose using Eq. (5) to evaluate the confidence of the pose. We assume that the higher this confidence (i.e., the more inliers) is, the more accurate the pose estimation will be. Therefore:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 529, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 529, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 83, + 529, + 287, + 562 + ], + "type": "interline_equation", + "content": "T _ {o u t} = \\left\\{ \\begin{array}{l l} T ^ {S} & \\text {i f i n l i e r s} ^ {S} > \\text {i n l i e r s} ^ {T}, \\\\ T ^ {T} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "c2c480125ae4b78387490790f152fe5b01a598faa5070707c4b83c2a30b93786.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 605 + ], + "type": "text", + "content": "We found that this simple pose-aware ensemble method is more effective for generating high-quality pseudo labels and inference than other ensemble methods (Sec. 4.5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "text", + "content": "Another crucial part of our method is a self-training loss. The total loss that we use to update the student network combines this self-training " + }, + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "text", + "content": " loss and the pseudo label loss " + }, + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "inline_equation", + "content": "L_{pl}" + }, + { + "bbox": [ + 46, + 606, + 287, + 654 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 665, + 287, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 665, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 118, + 665, + 287, + 678 + ], + "type": "interline_equation", + "content": "L _ {t t a} = \\lambda_ {d} L _ {d} + \\lambda_ {p l} L _ {p l}. \\tag {7}", + "image_path": "bba5b6777ca0c1c2cffc61f9d0028cd59732ca78756b7411897a75a454542e9e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The first term is self-training loss " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{d}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", which helps learn the distribution of the target domain and reduce the domain gap" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "through " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "T_{out}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ", which is obtained from the pose ensemble and observed " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ". This loss is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 368, + 105, + 545, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 105, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 368, + 105, + 545, + 119 + ], + "type": "interline_equation", + "content": "L _ {d} = L _ {C E} \\left(N _ {e} ^ {S}, \\operatorname {E n c} \\left(\\widehat {D} _ {e}\\right)\\right), \\tag {8}", + "image_path": "be560b263aea9d243287a3391d715cea521a32aa95f2963b03cc5611754aba92.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "inline_equation", + "content": "N_{e}^{S}" + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "text", + "content": " is the inlier-only student NOCS map, " + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\widehat{D}_{e}" + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "text", + "content": " is the resulting inliner-only transformed depth point cloud in NOCS space, and " + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}(\\cdot)" + }, + { + "bbox": [ + 304, + 129, + 545, + 178 + ], + "type": "text", + "content": " is the one-hot encoding to enable cross-entropy loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 179, + 388, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 179, + 388, + 190 + ], + "spans": [ + { + "bbox": [ + 318, + 179, + 388, + 190 + ], + "type": "text", + "content": "The second term" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 202, + 545, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 202, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 380, + 202, + 545, + 217 + ], + "type": "interline_equation", + "content": "L _ {p l} = L _ {C E} \\left(N _ {e} ^ {S}, N _ {e} ^ {T}\\right) \\tag {9}", + "image_path": "6e44a4b79be077f582bd9ffd0fdd604bd1a90718d26394a9a82fc04304c08959.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "text", + "content": "is the pseudo label loss from UDA-COPE [15], where " + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "inline_equation", + "content": "N_{e}^{T}" + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "text", + "content": " is the inlier-only teacher NOCS map. Since " + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "inline_equation", + "content": "N^{S} = \\theta^{S}(x_{t})" + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\theta^T (x_t)" + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "inline_equation", + "content": "L_{pl}" + }, + { + "bbox": [ + 304, + 227, + 545, + 277 + ], + "type": "text", + "content": " in Eq. (9) is the same as the loss in Eq. (3), except that here only inliers are considered." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 288, + 386, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 386, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 386, + 301 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 308, + 364, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 364, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 364, + 319 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 327, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 567 + ], + "type": "text", + "content": "We utilize two widely used category-level pose estimation datasets as the source and target domain, respectively. The source data is the Context-Aware MixEd Reality (CAMERA) dataset [43], generated by rendering and compositing synthetic objects into real scenes while considering the context. CAMERA consists of 275K RGB-D images as a training set with 1,085 object instances chosen from six categories: bottle, bowl, camera, can, laptop and mug. We use the REAL dataset [43] as the target domain. The target data consists of 4,300 real-world images of seven scenes for training and 2,750 real-world images of six scenes for evaluation. We refer to REAL evaluation set as REAL275. Our TTA methods do not use the target training set (4,300 images with seven scenes) and only use the evaluation set (2,750 images with six scenes, REAL275) for test-time adaptation, e.g., TENT [42] process evaluation set in a sequential, online manner. On the other hand, unsupervised methods such as [5, 16, 29] use the target training set for updating the models without any restrictions on how much and how long they use the information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 574, + 440, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 440, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 440, + 587 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Object Segmentation. We use Mask R-CNN [8] to obtain the object area in 2D image. We use the identical results of Mask R-CNN for a fair comparison with previous methods [3, 17, 36] for semi-supervised and unsupervised settings. For the semi-supervised setting, Mask R-CNN has trained on the source, and target domain supervised manner. For the unsupervised setting, Mask R-CNN is trained on only the source domain. The detected area of the image resizes to " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "192 \\times 192" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " image resolution as the teacher or student model input." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21289" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 87, + 537, + 333 + ], + "blocks": [ + { + "bbox": [ + 122, + 70, + 470, + 82 + ], + "lines": [ + { + "bbox": [ + 122, + 70, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 122, + 70, + 470, + 82 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons with state-of-the-art methods on the REAL275 dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 87, + 537, + 333 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 537, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 537, + 333 + ], + "type": "table", + "html": "
MethodSupervisedUnsupervisedmAP (↑)
SourceTargetTargetOnlineIoU50IoU755° 2cm5° 5cm10° 2cm10° 5 cm
Source (Supervised)
Metric Scale [14]2D/3D54.68.42.25.410.125.0
SPD [36]2D/3D50.517.011.512.133.037.9
Source (Supervised) & Target (Supervised)
NOCS [43]2D/3D2D/3D47.29.47.210.013.825.2
SPD [36]2D/3D2D/3D68.527.019.521.643.554.0
DualPoseNet [17]2D/3D2D/3D—30.829.335.950.066.8
CR-Net [44]2D/3D2D/3D—33.227.834.347.260.8
SGPA [3]2D/3D2D/3D—37.135.939.661.370.7
Source (Supervised) & Target (Semi-Supervised)
SSC-6D [29]2D/3D2D3DX73.0—16.819.644.154.5
RePoNet [5]2D/3D2D3DX76.0—30.733.9—63.0
UDA-COPE [15]2D/3D2D3DX75.534.430.534.957.066.1
Self-DPDN [16]2D/3D2D3DX75.241.639.545.063.372.2
TTA-COPE (Ours)2D/3D2D3D✓78.743.533.338.164.375.1
Source (Supervised) & Target (Unsupervised)
Self-DPDN [16]2D/3D2D/3DX67.243.939.046.761.873.4
TTA-COPE (Ours)2D/3D2D/3D✓69.139.730.235.961.773.2
", + "image_path": "bfca32a33947bdfdc633999d3afaaac04d47f4437d238a67464cb145cfdfc30b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": "Student and Teacher. The student and teacher models have identical design in 2D and 3D branches from UDA-COPE [15]. We utilize the PSPNet [53] with ResNet34 [9] backbone for the 2D image feature extraction. For a 3D branch, we use the MinkowskiNet [4] and utilize sparse convolution operation with a " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": " voxel size. Our NOCS representation uses the classification with 32 bins [43] instead of direct regression. During the pretraining stage, we train our model on the source data for 50 epochs using the Adam optimizer by initializing the learning rate of 1e-4 with a batch size of 32. The learning rate was reduced by a ratio of 0.6 (at 15k iterations), 0.3 (at 30k iterations), 0.1 (at 45k iterations), and 0.01 (at 60k iterations). During test-time adaptation, our student model uses the same learning rate, and the teacher smoothly updates using momentum update with " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "\\gamma = 0.99" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": ". Given target data, we first update the model every iteration and then estimate the pose. We set " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CE}} = 1.0" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{C}} = 1\\mathrm{e - }6" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d}} = 1.0" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{pl}} = 1.0" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": " for our experiments. The point filtering threshold " + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 352, + 289, + 580 + ], + "type": "text", + "content": " was set to 0.05 for all experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Metrics. To evaluate the performance of 3D object detection and 6D pose estimation, we follow the previous pose and size evaluation metric from Wang et al. [43]. We report the mean average precision (mAP) at the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " intersection over union (IoU) thresholds for 3D object detection. We also report mAP for 6D object pose evaluation w.r.t. rotation and translation errors, where e.g. the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "5^{\\circ}5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " metric describes the percentage of pose predictions where the error is less than both " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and the same for other thresholds. We recalculated all the metrics with the improved code." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 308, + 381, + 545, + 472 + ], + "blocks": [ + { + "bbox": [ + 305, + 350, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 373 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons with TTA baselines for category-level object pose estimation on the REAL275 dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 381, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 308, + 381, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 381, + 545, + 472 + ], + "type": "table", + "html": "
MethodUnsupervisedIoU75mAP (↑)
Target 3DTarget 2D5° 2cm5° 5 cm
TENT [42] - Eq. (2)✓33.725.929.1
PL [35] - Eqs. (3)-(4)✓39.929.734.9
PL-F [15] - Eqs. (3)-(5)✓41.131.436.3
TTA-COPE (Ours)✓43.533.338.1
TENT [42] - Eq. (2)✓✓32.326.831.2
PL [35] - Eqs. (3)-(4)✓✓36.026.933.1
PL-F [15] - Eqs. (3)-(5)✓✓36.528.034.0
TTA-COPE (Ours)✓✓39.730.235.9
", + "image_path": "a2e11649f2247962c8ad24264d39d9e6dd802a00ef7f1ef286566ef3fe266174.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 488, + 484, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 484, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 484, + 501 + ], + "type": "text", + "content": "4.3. Comparison with state-of-the-art" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 510, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 715 + ], + "type": "text", + "content": "Table 2 summarizes quantitative results on the REAL275 dataset under different settings: 1) supervised only on source data, 2) supervised on source and target data, 3) supervised on source data and semi-supervised on target data, and 4) supervised on source data and unsupervised on target data. Not surprisingly, methods following 2) such as SPD [36] perform better than setting 1) since they have the benefit of accessing both source and target data. Recent unsupervised domain adaptation (UDA) methods [15, 16] show remarkable results compared to state-of-the-art supervised methods without using target 3D labels (semi-supervised setting). Our method shows state-of-the-art results in IoU metrics under the semi-supervised setting and even outperforms the recent supervised methods SGPA [3] by a large margin (6.4 mAP in " + }, + { + "bbox": [ + 304, + 510, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 304, + 510, + 545, + 715 + ], + "type": "text", + "content": "). Our TTA-COPE uses less time and data to train the target domain because of the advantage of test-time adaptation but achieves comparable" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21290" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 85, + 538, + 267 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 522, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 522, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 522, + 82 + ], + "type": "text", + "content": "Table 4. Ablation study on variants of pose ensemble methods and self-training loss under the semi-supervised setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 85, + 538, + 267 + ], + "lines": [ + { + "bbox": [ + 51, + 85, + 538, + 267 + ], + "spans": [ + { + "bbox": [ + 51, + 85, + 538, + 267 + ], + "type": "table", + "html": "
MethodPose EnsemblemAP (↑)
InputOutputInferencePseudo LabelLd(8)IoU50IoU755° 2cm5° 5 cm10° 2cm10° 5 cm
LB76.237.529.134.662.173.2
(1)Argmax Match✓71.127.529.634.360.771.7
(2)Softmax Avg.✓78.740.432.137.364.074.7
(3)Softmax Max✓77.640.931.636.463.474.3
(4)Softmax Max✓77.541.132.137.164.374.6
(5)Inliers Max✓77.641.232.837.664.474.9
(6)Inliers Max✓✓78.242.932.437.864.174.7
(7)Argmax Match✓78.340.731.736.463.273.9
(8)Softmax Avg.✓78.641.131.536.263.674.5
(9)Softmax Max✓78.141.231.936.563.574.4
(10)Softmax Max✓78.641.231.836.363.574.4
(11)Inliers Max✓78.241.331.636.263.674.3
(12)Inliers Max✓✓78.743.032.637.063.573.9
OursInliers Max✓✓✓78.743.533.338.164.675.1
", + "image_path": "4ca85968709e2232c726a0e2cdbf13f16ccbd641dc62e3d3f16a6ff4c3ae48f9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 326 + ], + "type": "text", + "content": "results to the SOTA method, Self-DPDN [16]. Ours takes about 31 minutes (Table 5) for TTA and is " + }, + { + "bbox": [ + 46, + 277, + 287, + 326 + ], + "type": "inline_equation", + "content": "58\\mathrm{x}" + }, + { + "bbox": [ + 46, + 277, + 287, + 326 + ], + "type": "text", + "content": " faster than the Self-DPDN, which takes about 30 hours for training in the target domain, excluding inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 333, + 220, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 220, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 220, + 345 + ], + "type": "text", + "content": "4.4. Comparison with TTA baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 351, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 351, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 45, + 351, + 288, + 579 + ], + "type": "text", + "content": "Table 3 summarizes results for the different design test-time adaptation (TTA) baselines in semi-supervised and unsupervised settings. TENT [42] has been designed for general unsupervised settings, but it is not specially designed for the object pose estimation task and naturally observes overall poor performance compared to other TTA baselines. Mean teacher with pseudo labels (PL) [35] performs better than TENT [42] in semi-supervised and unsupervised settings, and we believe that training the model using pseudo labels with momentum update (Eqs. (3)-(4)) provides more stable training signals by reducing the error accumulation than entropy minimization (2). However, using pseudo labels without filtering (PL) provides unreliable labels as the ground truth, we found that pseudo label filtering (PL-F) [15] performs better than PL, which indicates that noise filtering (Eq. (5)) improves the results for the student as expected. Finally, our TTA-COPE achieves state-of-the-art (SOTA) performance among all TTA baselines under both semi-supervised and unsupervised settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 141, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 141, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 141, + 600 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 641 + ], + "type": "text", + "content": "In this section, we conduct experiments to evaluate the efficacy of the pose ensemble and self-training loss under a semi-supervised setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Input/Output-Level Pose Ensemble. As mentioned in the Sec. 3.3, we compare the different pose ensemble methods given predictions of teacher and student models " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "(N^T, N^S)" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and answer the following question: Which ensemble is the most effective for test-time adaptation? We categorize two ensemble techniques:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 545, + 422 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": "1) Input-level ensemble that fuses two NOCS map predictions " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "(N^T, N^S)" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " and makes an accurate single NOCS map " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " to estimate poses using the Umayama algorithm " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": ". For the input-level ensemble that fuses two predictions " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "(N^T, N^S)" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " into one prediction " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": ", we compare three strategies: Argmax Match, Softmax Average, and Softmax Max operations [12, 33]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "type": "text", + "content": "2) We ensemble output-level predictions " + }, + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "type": "inline_equation", + "content": "(N^T, N^S)" + }, + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "type": "text", + "content": " to estimate each pose " + }, + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "type": "inline_equation", + "content": "(T^T, T^S)" + }, + { + "bbox": [ + 304, + 361, + 545, + 422 + ], + "type": "text", + "content": " and choose the best pose predictions under specific criteria. Unlike the input ensemble, only the Softmax Max operation is valid for the output-level ensemble due to the nonlinearity of poses." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 555 + ], + "type": "text", + "content": "Table 4 summarizes our ensemble ablation study results. Table 4-(1-3) shows that the Softmax Avg. and Softmax Max perform better than the Argmax Match operation among input-level pose ensembles. If we compare the same operation (Softmax Max) in the input-level, Table 4-(3), and output-level, Table 4-(4), the output-level ensemble shows slightly better performance since the output-level operation is directly related to the pose results. Our proposed pose ensemble, Table 4-(5), considers pose-aware confidence by choosing a higher number of inlier points and achieves the best performance in all pose ensembles in Table 4-(1-5)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 676 + ], + "type": "text", + "content": "Use of Pose Ensemble. Our pose ensemble can be used for inference Table 4-(1-6, Ours) as well as generating pseudo labels Table 4-(7-12, Ours). The results show that our pose ensemble is helpful for both cases. In particular, applying our ensemble to inferences Table 4-(5) has been shown to be more effective than application to generating pseudo labels Table 4-(11) in pose metric. We also observe that our pose ensemble used for inference or pseudo labels improves overall metrics compared to the lower bound (LB) that is only trained on source data without test-time adaptation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Effect of self-training loss. The results of Table 4-(6) and Table 4-(12) show the effect of self-training loss Eq. (8). It yields an improvement of more than " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " compared" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21291" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 63, + 541, + 331 + ], + "blocks": [ + { + "bbox": [ + 56, + 63, + 541, + 331 + ], + "lines": [ + { + "bbox": [ + 56, + 63, + 541, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 63, + 541, + 331 + ], + "type": "image", + "image_path": "fb14cff0bc335d4c526e0afbca505e83268ed605adcbeeccf42cf44e01fdc031.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 338, + 518, + 350 + ], + "lines": [ + { + "bbox": [ + 73, + 338, + 518, + 350 + ], + "spans": [ + { + "bbox": [ + 73, + 338, + 518, + 350 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison with the lower bound (LB) and our TTA-COPE under the semi-supervised setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 375, + 285, + 425 + ], + "blocks": [ + { + "bbox": [ + 61, + 363, + 272, + 373 + ], + "lines": [ + { + "bbox": [ + 61, + 363, + 272, + 373 + ], + "spans": [ + { + "bbox": [ + 61, + 363, + 272, + 373 + ], + "type": "text", + "content": "Table 5. Ablation study on different updating intervals." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 375, + 285, + 425 + ], + "lines": [ + { + "bbox": [ + 50, + 375, + 285, + 425 + ], + "spans": [ + { + "bbox": [ + 50, + 375, + 285, + 425 + ], + "type": "table", + "html": "
MethodTTA TimeIoU755° 2cm5° 5 cm
Interval=1 (Ours)31 min43.533.338.1
Interval=1016 min41.832.737.7
Interval=2015 min41.131.937.1
", + "image_path": "f692e43b64ef12a9a1b2f620b00ce7ad1deae0a5bf05ee97802cf937601503ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "content": "to without self-training loss results of Table 4-(5, 11) in the " + }, + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{75}" + }, + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "content": " metric. This indicates that the self-learning loss helps to learn the distribution of the target domain and reduces the domain shift by comparing the predicted NOCS map against the observed point cloud in the target domain. The result in the last column is our proposed model for test-time adaptation, Fig. 3 shows qualitative results against the lower bound method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 526, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 288, + 658 + ], + "type": "text", + "content": "Updating Interval. TTA is inevitably slower than simple inference since additional training is required. However, reducing the update interval enables faster TTA time than TTA baselines because of reduced training time. Table 5 shows the difference in TTA speed and performance according to the updating interval. Specifically, the interval of every 10 frames improves the speed of TTA roughly two-fold compared to Interval 1 (Ours), with a marginal performance drop. We also increase intervals to 20, but it does not show as much improvement as before since most of the bottleneck arises from inference time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "We have proposed TTA-COPE, a test-time adaptation method for category-level object pose estimation that ad" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 496 + ], + "type": "text", + "content": "dresses the source-to-target domain shift without accessing source data at test time and without labeled target data. Specifically, we designed a pose ensemble method with self-training for test-time adaptation that simultaneously uses the teacher-student model to generate robust pseudo labels and estimate accurate poses for inference. We explore limitations of several test-time adaptation baselines and show that the proposed method achieves state-of-the-art performance. We demonstrate the benefits of our proposed pose ensemble and self-training loss with extensive studies in both semi-supervised and unsupervised settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 617 + ], + "type": "text", + "content": "To the best of our knowledge, TTA-COPE is the first approach that tries to solve test-time adaptation for category-level object pose estimation. Since our method currently focuses on generating 6D pose labels, it does not affect 2D labels and the segmentation model. In future work, when jointly considering 2D label and Mask R-CNN, greater performance improvement and stable test-time adaptation would be possible. Also, our pose estimation relies on non-differentiable pose estimation, and as such we could benefit from a differentiable pose estimation method [30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 632, + 399, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 399, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 399, + 645 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "This work was part of Taeyeop Lee's internship at NVIDIA and was also partially supported by an Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korean government(MSIT) (No.2020-0-00440)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 512, + 52, + 534, + 61 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 512, + 52, + 534, + 61 + ], + "spans": [ + { + "bbox": [ + 512, + 52, + 534, + 61 + ], + "type": "text", + "content": "Time" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21292" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "text", + "content": "[2] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11973-11982, 2020. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "type": "text", + "content": "[3] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2773-2782, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 322 + ], + "type": "text", + "content": "[4] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3075-3084, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "text", + "content": "[5] Yang Fu and Xiaolong Wang. Category-level 6d object pose estimation in the wild: A semi-supervised learning approach and a new dataset. arXiv preprint arXiv:2206.15436, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 401 + ], + "type": "text", + "content": "[6] Wei Gao and Russ Tedrake. kpm 2.0: Feedback control for category-level robotic manipulation. IEEE Robotics and Automation Letters (RA-L), 6(2):2962-2969, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 434 + ], + "type": "text", + "content": "[7] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "text", + "content": "[8] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 524 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning (ICML), pages 448-456, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[11] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. ShAPO: Implicit representations for multi-object shape, appearance, and pose optimization. arXiv preprint arXiv:2207.13691, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[12] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12605-12614, 2020. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6d pose estimation of novel objects via render & compare. In 6th Annual Conference on Robot Learning (CoRL), 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 149 + ], + "type": "text", + "content": "[14] Taeyeop Lee, Byeong-Uk Lee, Myungchul Kim, and In So Kweon. Category-level metric scale object shape and pose estimation. IEEE Robotics and Automation Letters (RA-L), 6(4):8575-8582, 2021. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 216 + ], + "type": "text", + "content": "[15] Taeyeop Lee, Byeong-Uk Lee, Inkyu Shin, Jaesung Choe, Ukcheol Shin, In So Kweon, and Kuk-Jin Yoon. UDA-COPE: Unsupervised domain adaptation for category-level object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14891-14900, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[16] Jiehong Lin, Zewei Wei, Changxing Ding, and Kui Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 19-34, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 338 + ], + "type": "text", + "content": "[17] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3560-3569, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[18] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an rgb image. In IEEE International Conference on Robotics and Automation (ICRA). ICRA, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 446 + ], + "type": "text", + "content": "[19] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7708-7717, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 513 + ], + "type": "text", + "content": "[20] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. DIST: Rendering deep implicit signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[21] Yuang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1215-1224, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "text", + "content": "[22] Massimiliano Mancini, Hakan Karaoguz, Elisa Ricci, Patric Jensfelt, and Barbara Caputo. Kitting in the wild through online domain adaptation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1103-1109. IEEE, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 657 + ], + "type": "text", + "content": "[23] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Transactions on Visualization and Computer Graphics (TVCG), 22(12):2633-2651, 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 680 + ], + "type": "text", + "content": "[24] Eitan Marder-Eppstein. Project Tango. In ACM SIGGRAPH Real-Time Live!, page 25, 2016. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[25] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-DoF GraspNet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21293" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision (ICCV), pages 2901-2910, 2019. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[26] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International Conference on Machine Learning (ICML), pages 16888-16905, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 205 + ], + "type": "text", + "content": "[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7668-7677, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "text", + "content": "[28] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-wise voting network for 6DoF pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4561-4570, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 262, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 287, + 316 + ], + "type": "text", + "content": "[29] Wanli Peng, Jianhang Yan, Hongtao Wen, and Yi Sun. Self-supervised category-level 6D object pose estimation with deep implicit shape representation. In Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI), volume 36, pages 2082-2090, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 317, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 381 + ], + "type": "text", + "content": "[30] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky T. Q. Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A library for differentiable nonlinear optimization. Advances in Neural Information Processing Systems, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 426 + ], + "type": "text", + "content": "[31] Mihir Prabhudesai, Sujoy Paul, Sjoerd van Steenkiste, Mehdi SM Sajjadi, Anirudh Goyal, Deepak Pathak, Katerina Fragkiadaki, Gaurav Aggarwal, and Thomas Kipf. Test-time adaptation with slot-centric models. 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 427, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 480 + ], + "type": "text", + "content": "[32] Martin Runz, Maud Buffer, and Lourdes Agapito. Mask-Fusion: Real-time recognition, tracking and reconstruction of multiple moving objects. In IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 10-20, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 547 + ], + "type": "text", + "content": "[33] Inkyu Shin, Yi-Hsuan Tsai, Bingbing Zhuang, Samuel Schulter, Buyu Liu, Sparsh Garg, In So Kweon, and Kuk-Jin Yoon. MM-TTA: Multi-modal test-time adaptation for 3D semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16928-16937, 2022. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[34] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. arXiv preprint arXiv:2303.01904, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 593, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 645 + ], + "type": "text", + "content": "[35] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[36] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[37] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cam" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "text", + "content": "eracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 969-977, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[38] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In Conference on Robot Learning (CoRL), 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 174, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 545, + 237 + ], + "type": "text", + "content": "[39] Stephen Tyree, Jonathan Tremblay, Thang To, Jia Cheng, Terry Mosier, Jeffrey Smith, and Stan Birchfield. 6-DoF pose estimation of household objects for robotic manipulation: An accessible dataset and benchmark. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[40] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence (TPAMI), 13(04):376-380, 1991. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 338 + ], + "type": "text", + "content": "[41] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3343–3352, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 339, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 382 + ], + "type": "text", + "content": "[42] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations (ICLR), 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 384, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 448 + ], + "type": "text", + "content": "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2642-2651, 2019. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[44] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 494, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 569 + ], + "type": "text", + "content": "[45] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multi-modal dataset for category-level object pose estimation with photometrically challenging objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21222-21231, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[46] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7201-7211, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 667 + ], + "type": "text", + "content": "[47] Bowen Wen and Kostas Bekris. BundleTrack: 6D pose tracking for novel objects without instance or category-level 3D models. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[48] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. Robotics: Science and Systems (RSS), 2022. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21294" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 329 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[49] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 194 + ], + "type": "text", + "content": "[50] Bowen Wen, Chaitanya Mitash, Sruthi Soorian, Andrew Kimmel, Avishai Sintov, and Kostas E Bekris. Robust, occlusion-aware pose estimation for objects grasped by adaptive hands. In IEEE International Conference on Robotics and Automation (ICRA), pages 6210-6217. IEEE, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "text", + "content": "[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. Robotics: Science and Systems (RSS), 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 284 + ], + "type": "text", + "content": "[52] He Yisheng, Wang Yao, Fan Haoqiang, Chen Qifeng, and Sun Jian. Fs6d: Few-shot 6d pose estimation of novel objects. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "text", + "content": "[53] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2881-2890, 2017. 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21295" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_content_list.json b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5191712c7e4aa4754d3034f9310c8a3b33e47fa0 --- /dev/null +++ b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_content_list.json @@ -0,0 +1,1427 @@ +[ + { + "type": "text", + "text": "TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization", + "text_level": 1, + "bbox": [ + 78, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziquan Liu $^{1}$ , Yi Xu $^{2*}$ , Xiangyang Ji $^{3}$ and Antoni B. Chan $^{1*}$", + "bbox": [ + 251, + 202, + 712, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science, City University of Hong Kong", + "bbox": [ + 223, + 220, + 745, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, Dalian University of Technology", + "bbox": [ + 222, + 239, + 746, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Automation, Tsinghua University", + "bbox": [ + 287, + 257, + 679, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ziquanliu.cs@gmail.com, yxu@dlut.edu.cn, xyji@tsinghua.edu.cn, abchan@cityu.edu.hk", + "bbox": [ + 117, + 277, + 844, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have seen the ever-increasing importance of pre-trained models and their downstream training in deep learning research and applications. At the same time, the defense for adversarial examples has been mainly investigated in the context of training from random initialization on simple classification tasks. To better exploit the potential of pre-trained models in adversarial robustness, this paper focuses on the fine-tuning of an adversially pre-trained model in various classification tasks. Existing research has shown that since the robust pre-trained model has already learned a robust feature extractor, the crucial question is how to maintain the robustness in the pre-trained model when learning the downstream task. We study the model-based and data-based approaches for this goal and find that the two common approaches cannot achieve the objective of improving both generalization and adversarial robustness. Thus, we propose a novel statistics-based approach, TwoWing NormliSation (TWINS) fine-tuning framework, which consists of two neural networks where one of them keeps the population means and variances of pre-training data in the batch normalization layers. Besides the robust information transfer, TWINS increases the effective learning rate without hurting the training stability since the relationship between a weight norm and its gradient norm in standard batch normalization layer is broken, resulting in a faster escape from the sub-optimal initialization and alleviating the robust overfitting. Finally, TWINS is shown to be effective on a wide range of image classification datasets in terms of both generalization and robustness.", + "bbox": [ + 75, + 358, + 473, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The adversarial vulnerability of deep neural networks (DNNs) [59] is one of the major obstacles for their wide", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0da221f25917e7ef65d4a2563fe4e0898588b55aa9551d9aadb17eef084a780f.jpg", + "image_caption": [ + "Figure 1. The performance of fine-tuning robust and non-robust large-scale pre-trained (PT) ResNet50 [26, 55] on CIFAR10 [34] and Caltech-256 [22]. We compare standard adversarial training (AT), Learning without Forgetting (LwF) (model approach) [39], joint fine-tuning with UOT data selection (data approach) [46] and our TWINS fine-tuning. The robust accuracy is evaluated using $l_{\\infty}$ norm bounded AutoAttack [11] with $\\epsilon = 8 / 255$ . On CIFAR10, the data-based and model-based approach fail to improve clean and robust accuracy. On Caltech, both approaches improve the clean accuracy but hurt the robust accuracy. Our TWINS fine-tuning improves the clean and robust performance on both datasets. The pink triangle denotes the performance of standard AT with the non-robust pre-trained ResNet50, which drops considerably compared with fine-tuning starting from the robust pretrained model." + ], + "image_footnote": [], + "bbox": [ + 506, + 329, + 705, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d68393543ae72d019748e33d6fb39257d9b60d2a2d536d365a5f3644c161b2b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 329, + 890, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "applications in safety-critical scenarios such as self-driving cars [18] and medical diagnosis [19]. Thus, addressing this issue has been one focus of deep learning research in the past eight years. Existing works have proposed to improve adversarial robustness from different perspectives, including data augmentation [21, 48, 53, 56], regularization [37, 43, 44, 51] and neural architecture [23, 28]. However, most of existing works investigate the problem under the assumption that the training data is sufficient enough, and training from scratch gives a satisfactory performance, which is not realistic in the real world. There are a large number of computer vision tasks where training from scratch is inferior to training from pre-trained weights, such as fine-grained image classification (e.g., Caltech-", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16436", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98f62d62bc6ae35d3945d6926361d03650ee38345992915c930d1c0be0a55fb3.jpg", + "image_caption": [ + "(a) TWINS Structure", + "Figure 2. The TWINS structure and training pipeline. (a) The Frozen Net and Adaptive Net have the same structure and share the weight parameters, except for batch normalization (BN) layers. The Frozen Net uses pre-trained means and standard deviations (STD) in the normalization layer, while Adaptive Net uses the mean and STD computed from the current batch as in standard BN. (b) In each step of mini-batch stochastic gradient descent (SGD), we split the batch of adversarial examples, generated from attacking the Adaptive Net, into two sub-batches and feed them to the Adaptive Net and Frozen Net respectively. The loss of two networks are combined and backpropagated to their shared parameters to train the network. In the inference stage, only the Adaptive Net is used." + ], + "image_footnote": [], + "bbox": [ + 83, + 116, + 544, + 273 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/64cb32692987443b8900d6c4735c18d6bd1bb99180d034d6bad575cb6a6a6670.jpg", + "image_caption": [ + "(b) Training Pipeline" + ], + "image_footnote": [], + "bbox": [ + 558, + 119, + 885, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "UCSD Birds-200-2011 or CUB200 [60]), object detection [42] and semantic segmentation [49].", + "bbox": [ + 75, + 366, + 468, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, pre-trained models have been considered as the foundation models in deep learning [5] as a result of their strong performance and wide employment in computer vision [17, 24, 25, 45], as well as natural language processing [6, 13, 52]. Thus, how to better use the pre-trained model in downstream has emerged as a major research topic in many vision and language tasks, such as image classification under distribution shifts [47, 63], object detection [36] and semantic segmentation [29, 35]. There are a few papers that investigate the pre-trained model's robustness in target tasks [7, 8, 15, 31, 32, 57, 62]. [7, 57] mainly considers the transfer between small-scale datasets (e.g., CIFAR100 to CIFAR10), while [8, 32] use adversarial robust pre-training and fine-tuning on the same dataset, without considering a large-scale and general pre-trained model. Finally, [15, 62] investigate different kinds of robustness to corruption or out-of-distribution samples, and are not devoted to adversarial robustness.", + "bbox": [ + 75, + 398, + 470, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we consider how to transfer the adversarial robustness of a large-scale robust pre-trained model (e.g., a ResNet50 pre-trained on ImageNet [12] with adversarial training) on various downstream classification tasks when fine-tuning with adversarial training. This problem setting is becoming more important as the standard pretrained models do not learn robust representations from the pre-training data and are substantially weaker than the robust pre-trained counterparts in some challenging downstream tasks, e.g., fine-grained classification as shown in our experiment. Meanwhile, more large-scale robust pretrained models are released (e.g., ResNet [55] and ViT [4]), which makes the robust pre-trained models more accessible. However, naively applying adversarial training to fine-tune from the robustly pre-trained model will lead to subopti", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mal robustness, since the robust representations learned by the robust pre-trained model are not fully utilized. For example, [57] suggests that the robustness from a pre-trained model needs to be explicitly maintained for its better transfer to the downstream.", + "bbox": [ + 496, + 366, + 892, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the idea that the key to improving the transferability of robustness is to maintain the robustness of the pre-training stage during fine-tuning [57], we first evaluate the data-based and model-based approach on two representative datasets, CIFAR10 and Caltech-256. The database-based approach uses pre-trained data in the fine-tuning and keeps their performance under adversarial attack, while the model-based approach regularizes the distance of features of the fine-tuned and pre-trained model. Our experiment shows that both methods fail to improve the robustness and generalization (Fig. 1), since the two methods are too aggressive in retaining the robustness and hurt the learning in downstream. Thus, we propose a subtle approach that keeps the batch-norm (BN) statistics of pre-training for preserving the robustness, which we call Two-Wing NormaliSation (TWINS) fine-tuning. TWINS has two neural networks with fixed and adaptive BN layers respectively, where the fixed BN layers use the population means and STDs of pretraining for normalization, while the adaptive BN layers use the standard BN normalization. Our experiment first demonstrates the importance of pre-trained BN statistics in the robust fine-tuning and then finds the benefit of TWINS in adversarial training dynamics. As the relationship between weight norm and its gradient norm no longer holds in TWINS, it is able to increase the gradient magnitude without increasing the gradient variance. At the initial training stage, TWINS has a faster escaping speed from the suboptimal initialization than vanilla adversarial training [41]. At the final training stage, the gradient of TWINS is more stable than adversarial training, which alleviates the robust", + "bbox": [ + 496, + 446, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16437", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "overfitting effect [56]. In summary, the contributions of our paper are as follows:", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We focus on the fine-tuning of large-scale robust pretrained models as a result of their potential importance in various downstream tasks. We evaluate current approaches to retain the pre-training robustness in fine-tuning, and show that they cannot substantially improve the robustness.", + "2. We propose TWINS, a statistics-based approach for better transferability of robustness and generalization from the pre-training domain to the target domain. TWINS has two benefits: a) it keeps the robust statistics for downstream tasks, thus helps the transfer the robustness to downstream tasks and b) it enlarges the gradient magnitude without increasing gradient variance, thus helps the model escape from the initialization faster and mitigates robust overfitting. The mechanisms of these two benefits are validated by our empirical study.", + "3. The effectiveness of TWINS is corroborated on five downstream datasets by comparing with two popular adversarial training baselines, adversarial training (AT) [48] and TRADES [64]. On average, TWINS improves the clean and robust accuracy by $2.18\\%$ and $1.21\\%$ compared with AT, and by $1.46\\%$ and $0.69\\%$ compared with TRADES. The experiment shows the strong potential of robust pre-trained models in boosting downstream's robustness and generalization when using more effective fine-tuning methods." + ], + "bbox": [ + 88, + 121, + 468, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 542, + 218, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial defense. There are several major approaches to improving the adversarial robustness of DNNs. The training of DNNs can be regularized to induce biases that are beneficial to adversarial robustness, such as locally linear regularization [51], margin maximization [14, 43] and Jacobian regularization [30]. The most commonly used adversarial defense is adversarial training (AT) [48], which directly trains the DNN on adversarial examples generated from PGD attack. Later, TRADES [64] is proposed to add a KL regularization to AT and achieves stronger adversarial robustness. Our paper proposes TWINS to improve adversarial training in the fine-tuning stage when the initial model is adversarially pre-trained. We compare TWINS-AT and TWINS-TRADES with vanilla AT and TRADES in our experiment and show the strong effectiveness of TWINS in the robust fine-tuning setting.", + "bbox": [ + 75, + 568, + 468, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning for downstream robustness. Several aspects of robustness in pre-training and fine-tuning have been studied in existing works. Adversarial contrastive learning [8, 32] is proposed to pre-train on a dataset with contrastive learning and then fine-tune on the same dataset, without considering the transferability of robustness from a large-", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scale pre-trained model to a different downstream task. In contrast, our paper investigates a more general problem, where task-specific pre-training is not needed for a new task as we use one robust large-scale pre-trained model trained on ImageNet. [55] considers the robust pre-training on the large-scale ImageNet and its transfer to downstream tasks, but focuses on the performance on clean instead of adversarial images. The Learning-without-Forgotten (LwF) [39] approach for retaining robustness is shown to be effective in the small-scale transfer experiment [57], but is not effective in our experiment setting of transfer of large-scale models. [31] proposes a learning rate schedule to improve the adversarial robustness of fine-tuned models, and [16] proposes robust informative fine-tuning for pre-trained language models to robustly keep pre-training information in downstream. The difference between [16, 31] and our work is that they assume a standard pre-trained model instead of the adversarial pre-trained model. [15, 62] investigate the performance of pre-trained models in downstream tasks, but the focus is the robustness to out-of-distribution samples instead of adversarial perturbations.", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Batch normalization. There are existing papers proposing the two-branch BN structure for different purposes with different technical details. [58] proposes dual normalization for a better trade-off between accuracy and robustness, where the normalization is a weighted sum of normalized clean and adversarial input. [61] proposes a similar two-branch BN structure, where one branch is for adversarial examples and the other is for clean examples. The major difference between our work and [58, 61] is that both BN branches in TWINS are for adversarial examples and one branch (Frozen Net) has fixed BN statistics from pretraining so as to better maintain the pre-trained robustness, whereas [58, 61] uses clean examples in BN and aims to improve the accuracy for clean images.", + "bbox": [ + 496, + 410, + 890, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Model-based and Data-based Approach to Retaining Adversarial Robustness", + "text_level": 1, + "bbox": [ + 498, + 643, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces the two common approaches for keeping adversarial robustness of pre-training in downstream, model-based and data-based approaches. Denote the feature vector output of a neural network as $g_{\\theta}(\\boldsymbol{x})$ , the training sample in the downstream task as $\\{(x_i,y_i)\\}_{i=1}^N \\sim \\mathbb{P}$ , and the loss function as $\\mathcal{L}(\\boldsymbol{w}^T g_{\\theta}(\\boldsymbol{x}) + b, \\boldsymbol{y})$ , where $(\\boldsymbol{w}, b)$ are the parameters of last classification layer. We assume that the pre-trained model is trained on adversarial examples generated from the PGD attack [48], where the $l_{\\infty}$ norm of the adversarial attack is bounded by $\\epsilon$ , and during fine-tuning we use adversarial training with the same PGD attack to obtain adversarial robustness in downstream tasks. In short, we consider the robust pre-training and robust fine-tuning setting in this paper, if not specified otherwise.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16438", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model-based approaches. We first introduce the model-based approach, which keeps the pre-trained model $\\theta_{pt}$ during fine-tuning so as to maintain its robustness. The objective function is", + "bbox": [ + 76, + 90, + 468, + 145 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {\\left(\\boldsymbol {x} _ {n}, \\boldsymbol {y} _ {n}\\right) \\sim \\mathbb {P}} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) + b, \\boldsymbol {y} _ {n}\\right) + \\lambda_ {L w F} \\| g _ {\\boldsymbol {\\theta} _ {p t}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) - g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 151, + 480, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the adversarial example generated from $\\boldsymbol{x}_n$ is denoted as $\\tilde{\\boldsymbol{x}}_n$ . The regularization term of the loss aims to minimize the distance between the features from the pretrained and the fine-tuned models, which is expected to maintain the robustness of the pre-trained model. This approach is originally proposed in [39] to prevent the catastrophic forgetting in continual learning, and is used in [57] to preserve adversarial robustness in transfer learning. Note that [57] uses the LwF method in standard fine-tuning instead of robust fine-tuning as in our paper.", + "bbox": [ + 76, + 200, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data-based approaches. The objective function of the data-based approach is", + "bbox": [ + 76, + 368, + 468, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {\\left(\\boldsymbol {x} _ {n}, \\boldsymbol {y} _ {n}\\right) \\sim \\mathbb {P}} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) + b, \\boldsymbol {y} _ {n}\\right) \\\\ + \\lambda_ {U O T} \\sum_ {\\left(\\boldsymbol {x} _ {m}, \\boldsymbol {y} _ {m}\\right) \\sim \\mathbb {Q}} \\mathcal {L} \\left(\\boldsymbol {w} _ {q} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {m}\\right) + b _ {q}, \\boldsymbol {y} _ {m}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 401, + 468, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{P}$ and $\\mathbb{Q}$ are data distribution of the target and the pre-training tasks, $w_{q}, b_{q}$ are the classification layer for pre-trained data from $\\mathbb{Q}$ . This method regularizes the current fine-tuned model feature extractor so that its prediction is still robust on the pre-training data. The joint training method is proposed in [46] to improve the performance of fine-tuning in downstream tasks where the training data is not sufficient.", + "bbox": [ + 76, + 470, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we test the performance of these two approaches on two standard image classification datasets, CIFAR10 [34] and Caltech-256 [22], with results shown in Figure 1. We use a grid search for learning rate and $\\lambda_{LwF}(\\lambda_{UOT})$ and report the result of the model with the best robust accuracy. See Section 5 and the supplemental material for the experiment setting. On CIFAR10, both approaches fail to improve either clean or robust accuracy; on Caltech-256, the two approaches improve the clean accuracy by a small margin but deteriorate the robustness. One reason why the model- and data-based approaches fail is that the regularization term might be too strong, thus hurting the learning in the downstream.", + "bbox": [ + 76, + 592, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. TWINS Fine-Tuning", + "text_level": 1, + "bbox": [ + 76, + 800, + 274, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The previous section demonstrates that both data- and model-based approaches cannot substantially improve the adversarial robustness in the downstream task. Thus we propose the TWINS for the better fine-tuning of robust pretrained models for downstream adversarial robustness.", + "bbox": [ + 76, + 825, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cea56a67dcb0f72a2574e945b38f3b99df3eca33da144cb0c5c4831967c10967.jpg", + "image_caption": [ + "Figure 3. The mean and normalized STD of gradient norms in AT and TWINS-AT on CIFAR10, CIFAR100 and Caltech-256. The averaged $\\mu (\\| \\nabla \\pmb {w}\\|)$ and $\\sigma (\\| \\nabla \\pmb {w}\\|) / \\mu (\\| \\nabla \\pmb {w}\\|)$ over epochs are shown in each plot. The gradient magnitudes of TWINS-AT are substantially larger than those of AT, while the Normalized STDs of gradient norm in TWINS-AT are not obviously increased (CIFAR10) or even decreased (CIFAR100 and Caltech-256) compared with AT. This property leads to the faster escaping speed of TWINS-AT from the initial sub-optimum compared with AT (Fig. 4), and reduced robust overfitting (Tab. 1)." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 888, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 542, + 671, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Though BN layers contain only a few parameters compared to convolution and fully connected layers, they play an important role in the good performance of DNNs. [20] shows that even if we only train the BN layers, the performance of a DNN is already quite impressive. [50] finds that only training the parameters in BN layers in an image generator is effective for small datasets. [38] proposes adaptive BN for domain adaptation, which updates the BN statistics with data from a target domain. These works motivate us to propose a statistics-based approach for retaining pretraining information in the target task.", + "bbox": [ + 496, + 566, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typical BN layers track the mean and STD of the training set and save them for the inference stage. As this distribution information for each layer might be helpful for downstream robustness, we propose the TWINS robust finetuning, which maintains two networks, Frozen Net that uses the BN statistics from the robust pre-trained model, and Adaptive Net that learns its BN statistics from the downstream task. Instead of using two independent networks for the Frozen and Adaptive Net, we let the two networks share weight parameters, excluding the BN layers, to save the model size and inference time. At initialization, both", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16439", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26134a5960d82f6e7b6c7c25ab0464e79a784705b0138551bc6674497cd7727a.jpg", + "image_caption": [ + "Figure 4. The distance between the current step's weight $\\theta$ and the initialization $\\theta_{pt}$ . On the three datasets, TWINS-AT has a faster escaping speed from the sub-optimal initial model than AT, which is due to TWIN-AT's larger gradient norm as shown in Figure 3." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 470, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "networks and their BN statistics are initialized by the robust pre-trained models. During training, the Frozen Net uses the population means and STDs of pre-training data computed in the pre-training stage in the normalization operation, while the Adaptive Net uses the current batch's mean and STD in the normalization and updates its running mean and STD with the target training data. Fig. 2 shows the general pipeline of TWINS training and the network structure.", + "bbox": [ + 75, + 250, + 468, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training objective of TWINS with adversarial training (TWINS-AT) in one mini-batch is:", + "bbox": [ + 76, + 369, + 468, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {j = 1} ^ {B / 2} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta} _ {a}} \\left(\\tilde {\\boldsymbol {x}} _ {j} ^ {(a)}\\right) + b, \\boldsymbol {y} _ {j}\\right) + \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 404, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {t w i n s} \\sum_ {i = B / 2 + 1} ^ {B} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta} _ {f}} \\left(\\tilde {\\boldsymbol {x}} _ {i} ^ {(a)}\\right) + b, \\boldsymbol {y} _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 446, + 468, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{f}$ and $\\theta_{a}$ denote the Frozen and Adaptive Nets respectively, and $\\tilde{\\pmb{x}}^{(a)}$ is the adversarial image for $\\pmb{x}$ when attacking the Adaptive Net $\\theta_{a}$ . We split the batch into two different sub-batches to avoid doubled batch sizes in TWINS training. Since the Frozen and Adaptive Nets share weight parameters, the number of parameters in TWINS is only increased by a very small amount (i.e., BN parameters). Thus, TWINS-AT only has a negligible cost in terms of memory and training time compared with vanilla AT. The TWINS structure can also be used in the TRADES in a similar way. Similar to [38], we can use the target training set to update the BN statistics so that they are more relevant to the downstream task. We call this procedure warmup in TWINS fine-tuning. The pseudo codes of TWINS-AT and TWINS-TRADES are given in Alg. 1 of the supplemental.", + "bbox": [ + 75, + 491, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. The mechanism of TWINS", + "text_level": 1, + "bbox": [ + 76, + 726, + 318, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first benefit of TWINS is mentioned in the motivation of TWINS, i.e., the BN statistics have robustness information in the pre-training domain that can be leveraged by robust fine-tuning to improve the downstream robustness. This argument is validated by our ablation study in Section 5, where we initialize the means and STDs with 1.0 and 0.0 for Frozen Net instead of the pre-trained means and STDs and check the accuracy and robustness. Figure 5 shows that the TWINS with (1,0) initialization cannot match the performance with TWINS with pre-trained", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/010c98536299be578cc7a9d114a6bf2f2926e91104fb102e7386d21d88f1eef3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRob. Acc.C10C100CaltechCUBDogs
ATBest ↑51.8431.3849.0927.0821.19
Final ↑49.4128.5248.3726.6019.80
Gap ↓2.432.860.730.481.39
TWINS-ATBest ↑53.2331.6048.8029.2420.89
Final ↑52.4031.0848.4029.2420.58
Gap ↓0.830.520.400.000.29
", + "bbox": [ + 506, + 88, + 883, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The robust accuracy drop of AT and TWINS-AT, where the adversarial attack is PGD10. TWINS-AT has smaller accuracy drop compared with AT, indicating that the TWINS-AT is less prone to robust overfitting as a result of reduced variance of gradient norms and stable training as shown in Fig. 3.", + "bbox": [ + 498, + 164, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "statistics, indicating that the robustness information in pretraining is essential to the effectiveness of TWINS.", + "bbox": [ + 498, + 246, + 890, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is intriguing that even with the (1,0) initialization, TWINS still outperforms the AT baseline in terms of robustness or accuracy on some datasets, e.g., Stanford Dogs and CIFAR100. This suggests that besides retaining the pre-training information, TWINS provides some other benefits during robust fine-tuning. By analyzing the gradient of TWINS, we find that TWINS implicitly increases the effective learning rate without increasing the oscillation and empirically validate this finding. We detail this analysis next.", + "bbox": [ + 498, + 277, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effective learning rate We first write the gradient of a weight vector for TWINS training. Consider the $l$ -th layer's weight $\\boldsymbol{w}_j^{(l)}$ and its output after BN layer", + "bbox": [ + 498, + 412, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {h} _ {i j} ^ {(l)} = \\frac {\\hat {h} _ {i j} ^ {(l)} - \\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\hat {h} _ {k j} ^ {(l)}}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} (\\hat {h} _ {k j} ^ {(l)} - \\hat {\\mu} _ {j} ^ {(l)}) ^ {2}}} (5) \\\\ = \\frac {\\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\left(\\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)\\right) ^ {2}}}, (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 472, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we denote $h_{ij}^{(l)}, \\hat{h}_{ij}^{(l)}$ and $\\tilde{h}_{ij}^{(l)}$ as the output of ReLU, convolution or fully connected layer and BN layer respectively, for $i$ -th sample, $j$ -th output variable at $l$ -th layer. The gradient with respect to $\\boldsymbol{w}_j^{(l)}$ is", + "bbox": [ + 498, + 585, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} = \\frac {\\partial \\mathcal {L} _ {a} (\\tilde {\\boldsymbol {x}} _ {i})}{\\partial \\boldsymbol {w} _ {j} ^ {(l)}} = \\frac {\\nabla \\tilde {h} _ {i j} ^ {(l)} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\left(\\boldsymbol {w} _ {j} ^ {(l) T} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)\\right) ^ {2}}} \\\\ - \\frac {\\nabla \\tilde {h} _ {i j} ^ {(l)} \\boldsymbol {w} _ {j} ^ {(l) T} (\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)})}{(\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} (\\boldsymbol {w} _ {j} ^ {(l) T} (\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)})) ^ {2}) ^ {3 / 2}} \\boldsymbol {\\Sigma} ^ {(l - 1)} \\boldsymbol {w} _ {j} ^ {(l)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 659, + 895, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}^{(l-1)}$ and $\\pmb{\\Sigma}^{(l-1)}$ are the mean and covariance matrix from $(l-1)$ -th layer, $\\nabla \\tilde{h}_{ij}^{(l)}$ represents the gradient of loss with respect to $\\tilde{h}_{ij}^{(l)}$ . If we write the weight as its norm $\\| \\pmb{w}_j^{(l)}\\|$ multiplied by the unit vector $\\pmb{u}_j^{(l)}$ , there is a relationship between $\\nabla \\pmb{w}_j^{(l)}$ and $\\| \\pmb{w}_j^{(l)}\\|$ ,", + "bbox": [ + 498, + 755, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} \\right\\| = \\frac {1}{\\left\\| \\boldsymbol {w} _ {j} ^ {(l)} \\right\\|} \\| \\nabla_ {a} \\boldsymbol {u} _ {j} ^ {(l)} \\| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 859, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16440", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This relationship has been found in [2, 27] and extended to any scale-invariant layers such as layer normalization [3] by [40]. It means that there are two ways to increase the gradient magnitude or the effective learning rate: 1) find a steeper descent direction where $\\|\\nabla_{a}\\pmb{u}_{j}^{(l)}\\|$ is increased and 2) decrease the weight norm so that $1 / \\| \\pmb{w}_j^{(l)}\\|$ is increased. For a DNN with standard BN layers, the training will exploit this property to increase the gradient magnitude by reducing the weight norms with the help of weight decay regularization, which leads to a spurious increase in gradient magnitude and a larger variance of gradient estimation. In contrast, for a DNN with fixed BN layers such as the Frozen Net, the gradient norm is not correlated with weight norm,", + "bbox": [ + 75, + 90, + 472, + 309 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {f} \\boldsymbol {w} _ {j} ^ {(l)} = \\frac {\\partial \\mathcal {L} _ {f} (\\tilde {\\boldsymbol {x}} _ {i})}{\\partial \\boldsymbol {w} _ {j} ^ {(l)}} = \\nabla \\tilde {h} _ {i j} ^ {(l)} \\frac {\\boldsymbol {h} _ {i} ^ {(l - 1)}}{\\sigma_ {j , p t} ^ {(l)}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 318, + 468, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this gradient, the only way to increase the gradient magnitude is to find the actual steeper direction. The overall gradient for the weight is", + "bbox": [ + 75, + 369, + 468, + 415 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {w} _ {j} ^ {(l)} = \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} + \\lambda_ {t w i n s} \\nabla_ {f} \\boldsymbol {w} _ {j} ^ {(l)}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 425, + 468, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Empirical Study To see the difference between the gradients in AT and TWINS-AT, we record the gradient of all weight parameters in each step of the two training methods and compute the mean and STD of the gradient norm in each epoch. The model parameters and their gradients are treated as long vectors and we compute the $l_{2}$ norm of the weight and gradient vector. Figure 3 shows the mean and normalized STD of gradient norms in 60 training epochs for three datasets. Here we show the normalized STD, i.e., dividing the STD by the mean, to see the relative effect of variance. The major finding is that the gradient magnitude of TWINS-AT is substantially larger than that of AT, while the variance of TWINS-AT is lower than AT in most epochs. Note that in theory, the ratio between STD and mean should remain the same after down-scaling the weight norm in standard BN, but in practice we do observe the high normalized variance of DNNs with standard BNs since we use one epoch's gradients to approximate the variance and mean.", + "bbox": [ + 75, + 462, + 472, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One benefit of the larger gradient magnitude is that the model can escape from the initial sub-optimal point faster and find a better local optimum than the small gradient optimization [41]. We validate this hypothesis by recording the distance between the current model and the initial model during training. Figure 4 shows these weight distances on three datasets, where TWINS-AT moves away from the initial model much faster than the AT baseline. The robust overfitting effect of adversarial training [54] is partially a result of large gradient variance, especially at the final stage", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of training [9]. We compare the robust accuracy drop of AT and TWINS-AT in Table 1 and find that the small relative variance of TWINS-AT has the effect of reducing robust overfitting. See the experiment details in Sec. 5 and Supp.", + "bbox": [ + 496, + 90, + 890, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 500, + 167, + 624, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section presents our experiment with TWINS. We first introduce our experiment setting and then show our main result and ablation study.", + "bbox": [ + 496, + 193, + 890, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 498, + 250, + 692, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use five datasets in our experiment. CIFAR10 and CIFAR100 [34] are low-resolution image datasets, where the training and validation sets have 50,000 and 10,000 images, and CIFAR10 has 10 classes, while CIFAR100 has 100 classes. Caltech-256 [22] is a high-resolution dataset with 30,607 images and 257 classes, which is split into training and validation set with a ratio of 9:1. Caltech-UCSD Birds-200-2011 (CUB200) [60] is a high-resolution bird image dataset for fine-grained image classification, which contains 200 classes of birds, 5,994 training images and 5,794 validation images. Stanford Dogs [33] has high-resolution dog images from 120 dog categories, where the training and validation set has 12,000 and 8,580 images. For both low-resolution image datasets (CIFAR10 and CIFAR100) and high-resolution datasets (Caltech-256,CUB200 and Stanford Dogs), we resize the image to $224\\times 224$ so that the input sizes are the same for pre-training and fine-tuning. As with pre-training, the input image is normalized by the mean and STD of the pre-training set. Note that the resizing and normalization function is integrated into the model so we can attack the input image with the [0,1] bounds for pixel values as in standard adversarial attacks. We use the standard ImageNet data augmentation for high-resolution datasets [26]. For CIFAR datasets, we use random cropping with padding=4 and random horizontal flipping.", + "bbox": [ + 496, + 265, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adversarial Pre-Training. Large-scale adversarial pretraining on ImageNet is time-consuming, and thus [55] has released adversially pre-trained ResNet50 and WideResNet50-2, trained with $l_{2}$ and $l_{\\infty}$ norm bounded attacks. In this paper, we adopt the pre-trained ResNet50 models, trained with $l_{\\infty}$ attack with bound $\\epsilon_{pt} = 4 / 255$ . We test other robust pre-trained models in our ablation study.", + "bbox": [ + 496, + 657, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Setting. For baselines and our method, we train all parameters of the pre-trained model, i.e., full fine-tuning instead of linear probing [57], with PGD attacks of $l_{\\infty}$ norm. The PGD step is 10, $\\epsilon_{ft} = 8 / 255$ and stepsize $\\alpha = 2 / 255$ . We set the batch size as 128 and train the model for 60 epochs and divide the learning rate by 0.1 at 30th and 50th epoch. In TWINS with warmup, we initialize the means and STDs with their pre-trained values, and update the means and STDs using the target training set.", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16441", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/963ad8cf68b21cd1957afea37351500f2e60fe9cfc1a02b14ca809581c149471.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricMethodCIFAR10CIFAR100Caltech256CUB200Stanford Dogs
Clean Acc.AT89.7769.4875.9065.7460.09
TWINS-AT91.24(+1.47)70.72(+1.24)76.86(+0.96)68.09(+2.35)64.98(+4.89)
TWINS-AT+warmup91.95(+2.18)72.12(+2.64)77.35(+1.45)67.64(+1.90)66.12(+6.03)
TRADES87.0662.7669.7058.9259.99
TWINS-TRADES86.61(-0.45)66.72(+3.96)71.12(+1.42)60.72(+1.80)60.58(+0.59)
TWINS-TRADES+warmup86.60(-0.46)65.91(+3.15)73.39(+3.69)61.05(+2.13)63.96(+3.97)
PGD10AT52.2428.5248.3726.6019.80
TWINS-AT52.73(+0.49)31.08(+2.56)48.40(+0.03)29.24(+2.64)20.58(+0.78)
TWINS-AT+warmup52.46(+0.22)29.12(+0.60)49.13(+0.76)27.67(+1.07)19.48(-0.32)
TRADES54.0432.2047.2827.8721.36
TWINS-TRADES56.23(+2.19)33.51(+1.31)47.31(+0.03)27.05(-0.82)19.93(-1.43)
TWINS-TRADES+warmup55.81(+1.77)33.48(+1.28)48.53(+1.25)26.68(-1.19)19.46(-1.90)
AAAT48.4623.4743.8522.8212.30
TWINS-AT49.81(+1.35)26.73(+3.26)43.69(-0.16)22.33(-0.49)14.37(+2.07)
TWINS-AT+warmup49.02(+0.56)25.72(+2.25)43.92(+0.07)23.58(+0.75)13.80(+1.50)
TRADES50.3126.4043.3922.2112.05
TWINS-TRADES51.71(+1.40)28.29(+1.89)41.77(-1.62)22.68(+0.47)13.36(+1.31)
TWINS-TRADES+warmup51.10(+0.79)28.30(+1.90)43.55(+0.16)21.92(-0.29)10.94(-1.11)
", + "bbox": [ + 102, + 88, + 864, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/55c48de41814901a2abd95cbe25dbe937a1681e7a9b15d1ec020135bfd08e7df.jpg", + "table_caption": [ + "Table 2. The performance of our TWINS-AT and TWINS-TRADES on five image classification tasks compared with AT and TRADES. The clean accuracy means the accuracy when testing images are input without adversarial perturbations. PGD10 and AA denote the robust test accuracy under PGD10 and AutoAttack. The increase and decrease in performance are denoted with green and red numbers. The bold numbers denote the best robust accuracy under AA. The proposed TWINS achieves better robustness and clean accuracy compared with the baseline. Averaged over the datasets, the clean and robust accuracy of TWINS are increased by $2.18\\%$ and $1.21\\%$ compared with AT, and $1.46\\%$ and $0.69\\%$ compared with TRADES. The means and STDs of the performance are in the supplemental." + ], + "table_footnote": [], + "table_body": "
MetricPT ModelCaltech256CUB200Dogs
Clean AccRandom48.9912.387.27
Non-Robust64.6653.3041.99
Robust75.9065.7460.09
PGD10Random31.783.7283.59
Non-Robust39.8619.6813.32
Robust48.3726.6019.80
", + "bbox": [ + 78, + 446, + 464, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The momentum of updating statistics is 0.1, the batch size is 128, and the warmup only lasts one epoch. Note that in the warmup stage, the input samples are added with adversarial perturbations generated by the PGD attack, which has the same setting as the attack in training, and the classifier layer is the pre-trained classifier for the adversarial attack. Our pilot experiment shows that using adversarial examples as input is more effective than using clean exemplars in the warmup. The optimizer is SGD with momentum in all of our experiments. The learning rate, weight decay and regularization hyperparameter are determined by grid search, which is described in detail in the supplemental.", + "bbox": [ + 75, + 609, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adversarial robustness evaluation. Two standard adversarial attacks are used in our experiment, i.e., PGD and AutoAttack [11]. The adversarial perturbation is $l_{\\infty}$ norm bounded in our evaluation. The setting of PGD attack for validation set is the same as the PGD attack in training. The AutoAttack (AA) is a more reliable adversarial attack and more often used for evaluation than PGD in recent years.", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d91e62e11e982168dddb9e5ada1ab73b01f330fc22894d4d6425102f121dd986.jpg", + "table_caption": [ + "Table 3. Comparison of random initialization, non-robust and robust pre-trained model on three difficult classification tasks, when fine-tuned with AT. The robust pre-trained model is indispensable to downstream robustness." + ], + "table_footnote": [], + "table_body": "
MethodMetricλWD
1e-51e-41e-31e-2
ATClean Acc.89.9289.9489.7790.28
Robust Acc.46.3444.6748.4647.57
TWINS-ATClean Acc.91.9091.4291.2487.33
Robust Acc.47.1949.0749.8151.65
", + "bbox": [ + 501, + 446, + 893, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. The performance of TWINS-AT and AT on CIFAR10 when the hyperparameter for weight decay ${\\lambda }_{WD}$ is changed. The robust accuracy is evaluated using AutoAttack. Our TWINS-AT achieves better adversarial robustness than AT for different ${\\lambda }_{WD}$ .", + "bbox": [ + 498, + 536, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the standard attacks of AA, i.e., untargeted APGD-CE, targeted APGD-DLR, targeted FAB [10] and Square Attack [1], with $\\epsilon = 8 / 255$ . The robust accuracy in our experiment result denotes the accuracy under AA.", + "bbox": [ + 498, + 607, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental Result", + "text_level": 1, + "bbox": [ + 500, + 679, + 694, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the three high-resolution datasets, we compare the performance of fine-tuning different initialization models, i.e., random initialization, standard pre-trained ResNet50 and robust pre-trained ResNet50. Table 3 shows that the pre-trained models are essential to the accuracy and robustness in challenging downstream tasks, since the random initialization is much worse than the two pre-trained models. The robust pre-trained model has a clear benefit over the standard pre-trained one, indicating that robust pre-training is indispensable to downstream robustness.", + "bbox": [ + 496, + 703, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the result of TWINS-AT and TWINS-TRADES compared with the baselines. Since AA is a more reliable attack, we highlight the best robust accuracy un", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16442", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea33abe8f798ad25652a4208a8c1e235cbe4c6777310366e8782553eb682c466.jpg", + "image_caption": [ + "Figure 5. Ablation study where the BN statistics in TWINS are initialized with (0,1) for means and STDs, denoted as TWINS-Init. The population mean and STD of pre-training are crucial to TWINS." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "der AA on each dataset. The TWINS fine-tuning learns more robust DNNs, as well as achieves a better clean accuracy on all five datasets, demonstrating the strong effectiveness of TWINS in the robust pre-training and robust fine-tuning setting. On CIFAR10 and CIFAR100, both TWINS-AT and TWINS-TRADES achieve better robustness and clean accuracy than their baselines, while the warmup only improves the clean accuracy and hurts the robustness sometimes. On Caltech-256, TWINS-AT improves upon the vanilla AT in both robustness and accuracy, but TWINS-TRADES does not perform better than vanilla TRADES. However, the warmup helps boost the performance of TWINS-TRADES as well as TWINS-AT and makes the TWINS with warmup perform better than the baselines.", + "bbox": [ + 75, + 272, + 468, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the two fine-grained image classification datasets, TWINS-AT and TWINS-TRADES generally perform better than baselines in terms of accuracy and robustness, if we look at the robust accuracy under AA, where only TWINS-AT on CUB200 has a slightly worse robust accuracy than its baseline. We find that the warmup improves the clean accuracy but hurts the robustness in most cases, except for CIFAR100 and Caltech-256. This can be a result of noisy adversarial perturbation, since we use the pre-trained classifier layer in the adversarial attack, or the insufficient update steps. Nevertheless, we note that the warmup can be considered as an operation for achieving a trade-off between robustness and accuracy.", + "bbox": [ + 75, + 500, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 709, + 230, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TWINS initialization. We use the pre-trained BN statistics in the Frozen Net to keep the robust information learned during pre-training. To show the importance of the pretrained BN statistics, we use the standard initialization (mean=0 and STD=1) for BN statistics in the Frozen Net, denoted as TWINS-Init, and show the result on the five datasets in Figure 5. Both clean and robust accuracy drop when the (0,1) initialization is used in TWINS-Init, demonstrating the crucial role of pre-trained statistics in TWINS. The fact that TWINS-Init sometimes improves upon the AT baseline motivates us to investigate the effect of TWINS on", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4cff26a6315467f4263cd219a16d50c808eab067e8255efb1dbfd90cb7441e53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMetricεpt
1/2552/2554/2558/255
ATClean Acc.65.4767.0869.4869.93
Robust Acc.24.7625.7923.4727.71
TWINS-ATClean Acc.68.5570.4570.7272.59
Robust Acc.25.9726.6226.7328.62
", + "bbox": [ + 501, + 90, + 897, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The performance of TWINS-AT and AT on CIFAR100 when robust pre-trained ResNet50 with different $\\epsilon_{pt}$ are used.", + "bbox": [ + 500, + 178, + 890, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gradient norms in Section 4.", + "bbox": [ + 500, + 215, + 687, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of weight decay. Weight decay is the reason for the decreasing weight norm in DNNs with BN layers, so increasing the hyperparameter of weight decay $\\lambda_{WD}$ is also a way to increase the gradient magnitude. Table 4 shows the result of TWINS-AT and AT when different $\\lambda_{WD}$ are used when fine-tuning the robust pre-trained model on CIFAR10. The robust accuracy of TWINS-AT is consistently better than that of AT across different $\\lambda_{WD}$ 's. We draw the same conclusion on CIFAR100 (see supplemental). Note that the clean accuracy of TWINS-AT drops when a large $\\lambda_{WD}$ is used, suggesting that we should not use a too large $\\lambda_{WD}$ for TWINS-AT.", + "bbox": [ + 496, + 229, + 890, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different robust pre-trained models. The main experiment uses the robust pre-trained ResNet50 with $\\epsilon_{pt} = 4/255$ as the initial model. We try different robust pre-trained models with different $\\epsilon_{pt}$ in Table 5, which shows that a larger $\\epsilon_{pt}$ is beneficial to both clean and robust accuracy in the downstream, and the proposed TWINS-AT is better than AT in both metrics with different pre-trained models.", + "bbox": [ + 496, + 412, + 890, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 539, + 617, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper investigates the utility of robust pre-trained models in various downstream classification tasks. We first find that the commonly used data- and model-based approaches to maintain pre-training information do not work in the adversarial robust fine-tuning. We then propose a subtle statistics-based method, TWINS, for retaining the pretraining robustness in the downstream. In addition to the robustness preserving effect, we find that TWINS increases the gradient magnitude without sacrificing the training stability and improves the training dynamics of AT. Finally, the performance of TWINS is shown to be stronger than that of AT and TRADES on five datasets. One limitation of our work is that we only evaluate the robust supervised pre-trained ResNet50. Recently, robust pre-trained ViT's on ImageNet [4] have been released. Our statistics-based approach can be extended to the layer normalization, on which the increasing gradient magnitude argument also holds, and thus future work will extend TWINS to ViT.", + "bbox": [ + 496, + 553, + 890, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11215820) and the Fundamental Research Funds for the Central University of China (DUT No. 82232031).", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16443", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 7", + "[2] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. In International Conference on Learning Representations, 2019. 6", + "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6", + "[4] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 34:26831-26843, 2021. 2, 8", + "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[7] Dian Chen, Hongxin Hu, Qian Wang, Li Yinli, Cong Wang, Chao Shen, and Qi Li. Cartl: Cooperative adversarially-robust transfer learning. In International Conference on Machine Learning, pages 1640–1650. PMLR, 2021. 2", + "[8] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 699-708, 2020. 2, 3", + "[9] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In International Conference on Learning Representations, 2020. 6", + "[10] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In International Conference on Machine Learning, pages 2196-2205. PMLR, 2020. 7", + "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1, 7", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + ], + "bbox": [ + 76, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Gavin Weiguang Ding, Yash Sharma, Kry Yuk Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2019. 3", + "[15] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, et al. On robustness and transferability of convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16458-16468, 2021. 2, 3", + "[16] Xinshuai Dong, Anh Tuan Luu, Min Lin, Shuicheng Yan, and Hanwang Zhang. How should pre-trained language models be fine-tuned towards adversarial robustness? Advances in Neural Information Processing Systems, 34:4356-4369, 2021. 3", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2", + "[18] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature communications, 12(1):1-14, 2021. 1", + "[19] Samuel G Finlayson, John D Bowers, Joichi Ito, Jonathan L Zittrain, Andrew L Beam, and Isaac S Kohane. Adversarial attacks on medical machine learning. Science, 363(6433):1287-1289, 2019. 1", + "[20] Jonathan Frankle, David J. Schwab, and Ari S. Morcos. Training batchnorm and only batchnorm: On the expressive power of random features in {cnn}s. In International Conference on Learning Representations, 2021. 4", + "[21] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 1", + "[22] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 1, 4, 6", + "[23] Minghao Guo, Yuzhe Yang, Rui Xu, Ziwei Liu, and Dahua Lin. When nas meets robustness: In search of robust architectures against adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 631-640, 2020. 1", + "[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2", + "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16444", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6", + "[27] Elad Hoffer, Ron Banner, Itay Golan, and Daniel Soudry. Norm matters: efficient and accurate normalization schemes in deep networks. Advances in Neural Information Processing Systems, 31, 2018. 6", + "[28] Hanxun Huang, Yisen Wang, Sarah Erfani, Quanquan Gu, James Bailey, and Xingjun Ma. Exploring architectural ingredients of adversarially robust deep neural networks. Advances in Neural Information Processing Systems, 34:5545-5559, 2021. 1", + "[29] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 2", + "[30] Daniel Jakubovitz and Raja Giryes. Improving dnn robustness to adversarial attacks using jacobian regularization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 514-529, 2018. 3", + "[31] Ahmadreza Jeddi, Mohammad Javad Shafiee, and Alexander Wong. A simple fine-tuning is all you need: Towards robust deep learning via adversarial fine-tuning. arXiv preprint arXiv:2012.13628, 2020. 2, 3", + "[32] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 33:16199-16210, 2020. 2, 3", + "[33] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 6", + "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 4, 6", + "[35] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2", + "[36] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2", + "[37] Yao Li, Martin Renqiang Min, Thomas Lee, Wenchao Yu, Erik Kruus, Wei Wang, and Cho-Jui Hsieh. Towards robustness of deep neural networks via regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7496-7505, 2021. 1", + "[38] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779, 2016. 4, 5", + "[39] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 3, 4", + "[40] Zhiyuan Li, Kaifeng Lyu, and Sanjeev Arora. Reconciling modern deep learning with traditional optimization analyses: The intrinsic learning rate. Advances in Neural Information Processing Systems, 33:14544-14555, 2020. 6" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. Advances in Neural Information Processing Systems, 33:21476-21487, 2020. 2, 6", + "[42] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128(2):261-318, 2020. 2", + "[43] Ziquan Liu and Antoni B Chan. Boosting adversarial robustness from the perspective of effective margin regularization. British Machine Vision Conference (BMVC), 2022. 1, 3", + "[44] Ziquan Liu, Yufei Cui, and Antoni B Chan. Improve generalization and robustness of neural networks via weight scale shifting invariant regularizations. ICML Workshop on Adversarial Machine Learning, 2021. 1", + "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[46] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 4", + "[47] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, Xiangyang Ji, and Antoni B Chan. An empirical study on distribution shift robustness from the perspective of pre-training and data augmentation. arXiv preprint arXiv:2205.12753, 2022. 2", + "[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3", + "[49] Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 2", + "[50] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 4", + "[51] Chongli Qin, James Martens, Sven Gowal, Dilip Krishnan, Krishnamurthy Dvijotham, Alhussein Fawzi, Soham De, Robert Stanforth, and Pushmeet Kohli. Adversarial robustness through local linearization. Advances in Neural Information Processing Systems, 32, 2019. 1, 3", + "[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 2", + "[53] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Data augmentation can improve robustness. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16445", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093–8104. PMLR, 2020. 6", + "[55] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 1, 2, 3, 6", + "[56] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2021. 1, 3", + "[57] Ali Shafahi, Parsa Saadatpanah, Chen Zhu, Amin Ghiasi, Christoph Studer, David Jacobs, and Tom Goldstein. Adversarially robust transfer learning. In International Conference on Learning Representations, 2020. 2, 3, 4, 6", + "[58] Chawin Sitawarin, Arvind Sridhar, and David Wagner. Improving the accuracy-robustness trade-off for dual-domain adversarial training. ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning. 3", + "[59] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 1", + "[60] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6", + "[61] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 3", + "[62] Yutaro Yamada and Mayu Otani. Does robustness on image net transfer to downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9215-9224, 2022. 2, 3", + "[63] Yaodong Yu, Heinrich Jiang, Dara Bahri, Hossein Mobahi, Seungyeon Kim, Ankit Singh Rawat, Andreas Veit, and Yi Ma. An empirical study of pre-trained vision models on out-of-distribution generalization. In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021. 2", + "[64] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 781 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16446", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_model.json b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..febd9d29d2cb517032e7eec10f4ed06fc13d45f8 --- /dev/null +++ b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.203, + 0.714, + 0.221 + ], + "angle": 0, + "content": "Ziquan Liu\\(^{1}\\), Yi Xu\\(^{2*}\\), Xiangyang Ji\\(^{3}\\) and Antoni B. Chan\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.222, + 0.746, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computer Science, City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.24, + 0.748, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, Dalian University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.258, + 0.681, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Automation, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.278, + 0.846, + 0.293 + ], + "angle": 0, + "content": "ziquanliu.cs@gmail.com, yxu@dlut.edu.cn, xyji@tsinghua.edu.cn, abchan@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.797 + ], + "angle": 0, + "content": "Recent years have seen the ever-increasing importance of pre-trained models and their downstream training in deep learning research and applications. At the same time, the defense for adversarial examples has been mainly investigated in the context of training from random initialization on simple classification tasks. To better exploit the potential of pre-trained models in adversarial robustness, this paper focuses on the fine-tuning of an adversially pre-trained model in various classification tasks. Existing research has shown that since the robust pre-trained model has already learned a robust feature extractor, the crucial question is how to maintain the robustness in the pre-trained model when learning the downstream task. We study the model-based and data-based approaches for this goal and find that the two common approaches cannot achieve the objective of improving both generalization and adversarial robustness. Thus, we propose a novel statistics-based approach, TwoWing NormliSation (TWINS) fine-tuning framework, which consists of two neural networks where one of them keeps the population means and variances of pre-training data in the batch normalization layers. Besides the robust information transfer, TWINS increases the effective learning rate without hurting the training stability since the relationship between a weight norm and its gradient norm in standard batch normalization layer is broken, resulting in a faster escape from the sub-optimal initialization and alleviating the robust overfitting. Finally, TWINS is shown to be effective on a wide range of image classification datasets in terms of both generalization and robustness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.88 + ], + "angle": 0, + "content": "The adversarial vulnerability of deep neural networks (DNNs) [59] is one of the major obstacles for their wide" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.33, + 0.706, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.33, + 0.892, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.68 + ], + "angle": 0, + "content": "Figure 1. The performance of fine-tuning robust and non-robust large-scale pre-trained (PT) ResNet50 [26, 55] on CIFAR10 [34] and Caltech-256 [22]. We compare standard adversarial training (AT), Learning without Forgetting (LwF) (model approach) [39], joint fine-tuning with UOT data selection (data approach) [46] and our TWINS fine-tuning. The robust accuracy is evaluated using \\( l_{\\infty} \\) norm bounded AutoAttack [11] with \\( \\epsilon = 8 / 255 \\). On CIFAR10, the data-based and model-based approach fail to improve clean and robust accuracy. On Caltech, both approaches improve the clean accuracy but hurt the robust accuracy. Our TWINS fine-tuning improves the clean and robust performance on both datasets. The pink triangle denotes the performance of standard AT with the non-robust pre-trained ResNet50, which drops considerably compared with fine-tuning starting from the robust pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "applications in safety-critical scenarios such as self-driving cars [18] and medical diagnosis [19]. Thus, addressing this issue has been one focus of deep learning research in the past eight years. Existing works have proposed to improve adversarial robustness from different perspectives, including data augmentation [21, 48, 53, 56], regularization [37, 43, 44, 51] and neural architecture [23, 28]. However, most of existing works investigate the problem under the assumption that the training data is sufficient enough, and training from scratch gives a satisfactory performance, which is not realistic in the real world. There are a large number of computer vision tasks where training from scratch is inferior to training from pre-trained weights, such as fine-grained image classification (e.g., Caltech-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16436" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.096, + 0.389, + 0.113 + ], + "angle": 0, + "content": "(a) TWINS Structure" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.117, + 0.545, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.096, + 0.798, + 0.113 + ], + "angle": 0, + "content": "(b) Training Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.121, + 0.887, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.894, + 0.36 + ], + "angle": 0, + "content": "Figure 2. The TWINS structure and training pipeline. (a) The Frozen Net and Adaptive Net have the same structure and share the weight parameters, except for batch normalization (BN) layers. The Frozen Net uses pre-trained means and standard deviations (STD) in the normalization layer, while Adaptive Net uses the mean and STD computed from the current batch as in standard BN. (b) In each step of mini-batch stochastic gradient descent (SGD), we split the batch of adversarial examples, generated from attacking the Adaptive Net, into two sub-batches and feed them to the Adaptive Net and Frozen Net respectively. The loss of two networks are combined and backpropagated to their shared parameters to train the network. In the inference stage, only the Adaptive Net is used." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.398 + ], + "angle": 0, + "content": "UCSD Birds-200-2011 or CUB200 [60]), object detection [42] and semantic segmentation [49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.671 + ], + "angle": 0, + "content": "On the other hand, pre-trained models have been considered as the foundation models in deep learning [5] as a result of their strong performance and wide employment in computer vision [17, 24, 25, 45], as well as natural language processing [6, 13, 52]. Thus, how to better use the pre-trained model in downstream has emerged as a major research topic in many vision and language tasks, such as image classification under distribution shifts [47, 63], object detection [36] and semantic segmentation [29, 35]. There are a few papers that investigate the pre-trained model's robustness in target tasks [7, 8, 15, 31, 32, 57, 62]. [7, 57] mainly considers the transfer between small-scale datasets (e.g., CIFAR100 to CIFAR10), while [8, 32] use adversarial robust pre-training and fine-tuning on the same dataset, without considering a large-scale and general pre-trained model. Finally, [15, 62] investigate different kinds of robustness to corruption or out-of-distribution samples, and are not devoted to adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this paper, we consider how to transfer the adversarial robustness of a large-scale robust pre-trained model (e.g., a ResNet50 pre-trained on ImageNet [12] with adversarial training) on various downstream classification tasks when fine-tuning with adversarial training. This problem setting is becoming more important as the standard pretrained models do not learn robust representations from the pre-training data and are substantially weaker than the robust pre-trained counterparts in some challenging downstream tasks, e.g., fine-grained classification as shown in our experiment. Meanwhile, more large-scale robust pretrained models are released (e.g., ResNet [55] and ViT [4]), which makes the robust pre-trained models more accessible. However, naively applying adversarial training to fine-tune from the robustly pre-trained model will lead to subopti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.442 + ], + "angle": 0, + "content": "mal robustness, since the robust representations learned by the robust pre-trained model are not fully utilized. For example, [57] suggests that the robustness from a pre-trained model needs to be explicitly maintained for its better transfer to the downstream." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Following the idea that the key to improving the transferability of robustness is to maintain the robustness of the pre-training stage during fine-tuning [57], we first evaluate the data-based and model-based approach on two representative datasets, CIFAR10 and Caltech-256. The database-based approach uses pre-trained data in the fine-tuning and keeps their performance under adversarial attack, while the model-based approach regularizes the distance of features of the fine-tuned and pre-trained model. Our experiment shows that both methods fail to improve the robustness and generalization (Fig. 1), since the two methods are too aggressive in retaining the robustness and hurt the learning in downstream. Thus, we propose a subtle approach that keeps the batch-norm (BN) statistics of pre-training for preserving the robustness, which we call Two-Wing NormaliSation (TWINS) fine-tuning. TWINS has two neural networks with fixed and adaptive BN layers respectively, where the fixed BN layers use the population means and STDs of pretraining for normalization, while the adaptive BN layers use the standard BN normalization. Our experiment first demonstrates the importance of pre-trained BN statistics in the robust fine-tuning and then finds the benefit of TWINS in adversarial training dynamics. As the relationship between weight norm and its gradient norm no longer holds in TWINS, it is able to increase the gradient magnitude without increasing the gradient variance. At the initial training stage, TWINS has a faster escaping speed from the suboptimal initialization than vanilla adversarial training [41]. At the final training stage, the gradient of TWINS is more stable than adversarial training, which alleviates the robust" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16437" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "overfitting effect [56]. In summary, the contributions of our paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.47, + 0.212 + ], + "angle": 0, + "content": "1. We focus on the fine-tuning of large-scale robust pretrained models as a result of their potential importance in various downstream tasks. We evaluate current approaches to retain the pre-training robustness in fine-tuning, and show that they cannot substantially improve the robustness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.213, + 0.47, + 0.379 + ], + "angle": 0, + "content": "2. We propose TWINS, a statistics-based approach for better transferability of robustness and generalization from the pre-training domain to the target domain. TWINS has two benefits: a) it keeps the robust statistics for downstream tasks, thus helps the transfer the robustness to downstream tasks and b) it enlarges the gradient magnitude without increasing gradient variance, thus helps the model escape from the initialization faster and mitigates robust overfitting. The mechanisms of these two benefits are validated by our empirical study." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.47, + 0.53 + ], + "angle": 0, + "content": "3. The effectiveness of TWINS is corroborated on five downstream datasets by comparing with two popular adversarial training baselines, adversarial training (AT) [48] and TRADES [64]. On average, TWINS improves the clean and robust accuracy by \\(2.18\\%\\) and \\(1.21\\%\\) compared with AT, and by \\(1.46\\%\\) and \\(0.69\\%\\) compared with TRADES. The experiment shows the strong potential of robust pre-trained models in boosting downstream's robustness and generalization when using more effective fine-tuning methods." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.122, + 0.47, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.22, + 0.559 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.811 + ], + "angle": 0, + "content": "Adversarial defense. There are several major approaches to improving the adversarial robustness of DNNs. The training of DNNs can be regularized to induce biases that are beneficial to adversarial robustness, such as locally linear regularization [51], margin maximization [14, 43] and Jacobian regularization [30]. The most commonly used adversarial defense is adversarial training (AT) [48], which directly trains the DNN on adversarial examples generated from PGD attack. Later, TRADES [64] is proposed to add a KL regularization to AT and achieves stronger adversarial robustness. Our paper proposes TWINS to improve adversarial training in the fine-tuning stage when the initial model is adversarially pre-trained. We compare TWINS-AT and TWINS-TRADES with vanilla AT and TRADES in our experiment and show the strong effectiveness of TWINS in the robust fine-tuning setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fine-tuning for downstream robustness. Several aspects of robustness in pre-training and fine-tuning have been studied in existing works. Adversarial contrastive learning [8, 32] is proposed to pre-train on a dataset with contrastive learning and then fine-tune on the same dataset, without considering the transferability of robustness from a large-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.409 + ], + "angle": 0, + "content": "scale pre-trained model to a different downstream task. In contrast, our paper investigates a more general problem, where task-specific pre-training is not needed for a new task as we use one robust large-scale pre-trained model trained on ImageNet. [55] considers the robust pre-training on the large-scale ImageNet and its transfer to downstream tasks, but focuses on the performance on clean instead of adversarial images. The Learning-without-Forgotten (LwF) [39] approach for retaining robustness is shown to be effective in the small-scale transfer experiment [57], but is not effective in our experiment setting of transfer of large-scale models. [31] proposes a learning rate schedule to improve the adversarial robustness of fine-tuned models, and [16] proposes robust informative fine-tuning for pre-trained language models to robustly keep pre-training information in downstream. The difference between [16, 31] and our work is that they assume a standard pre-trained model instead of the adversarial pre-trained model. [15, 62] investigate the performance of pre-trained models in downstream tasks, but the focus is the robustness to out-of-distribution samples instead of adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Batch normalization. There are existing papers proposing the two-branch BN structure for different purposes with different technical details. [58] proposes dual normalization for a better trade-off between accuracy and robustness, where the normalization is a weighted sum of normalized clean and adversarial input. [61] proposes a similar two-branch BN structure, where one branch is for adversarial examples and the other is for clean examples. The major difference between our work and [58, 61] is that both BN branches in TWINS are for adversarial examples and one branch (Frozen Net) has fixed BN statistics from pretraining so as to better maintain the pre-trained robustness, whereas [58, 61] uses clean examples in BN and aims to improve the accuracy for clean images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.644, + 0.892, + 0.679 + ], + "angle": 0, + "content": "3. The Model-based and Data-based Approach to Retaining Adversarial Robustness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This section introduces the two common approaches for keeping adversarial robustness of pre-training in downstream, model-based and data-based approaches. Denote the feature vector output of a neural network as \\( g_{\\theta}(\\boldsymbol{x}) \\), the training sample in the downstream task as \\( \\{(x_i,y_i)\\}_{i=1}^N \\sim \\mathbb{P} \\), and the loss function as \\( \\mathcal{L}(\\boldsymbol{w}^T g_{\\theta}(\\boldsymbol{x}) + b, \\boldsymbol{y}) \\), where \\( (\\boldsymbol{w}, b) \\) are the parameters of last classification layer. We assume that the pre-trained model is trained on adversarial examples generated from the PGD attack [48], where the \\( l_{\\infty} \\) norm of the adversarial attack is bounded by \\( \\epsilon \\), and during fine-tuning we use adversarial training with the same PGD attack to obtain adversarial robustness in downstream tasks. In short, we consider the robust pre-training and robust fine-tuning setting in this paper, if not specified otherwise." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16438" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Model-based approaches. We first introduce the model-based approach, which keeps the pre-trained model \\(\\theta_{pt}\\) during fine-tuning so as to maintain its robustness. The objective function is" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.152, + 0.482, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\sum_ {\\left(\\boldsymbol {x} _ {n}, \\boldsymbol {y} _ {n}\\right) \\sim \\mathbb {P}} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) + b, \\boldsymbol {y} _ {n}\\right) + \\lambda_ {L w F} \\| g _ {\\boldsymbol {\\theta} _ {p t}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) - g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.202, + 0.47, + 0.352 + ], + "angle": 0, + "content": "where the adversarial example generated from \\( \\boldsymbol{x}_n \\) is denoted as \\( \\tilde{\\boldsymbol{x}}_n \\). The regularization term of the loss aims to minimize the distance between the features from the pretrained and the fine-tuned models, which is expected to maintain the robustness of the pre-trained model. This approach is originally proposed in [39] to prevent the catastrophic forgetting in continual learning, and is used in [57] to preserve adversarial robustness in transfer learning. Note that [57] uses the LwF method in standard fine-tuning instead of robust fine-tuning as in our paper." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.469, + 0.398 + ], + "angle": 0, + "content": "Data-based approaches. The objective function of the data-based approach is" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.402, + 0.47, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {\\left(\\boldsymbol {x} _ {n}, \\boldsymbol {y} _ {n}\\right) \\sim \\mathbb {P}} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) + b, \\boldsymbol {y} _ {n}\\right) \\\\ + \\lambda_ {U O T} \\sum_ {\\left(\\boldsymbol {x} _ {m}, \\boldsymbol {y} _ {m}\\right) \\sim \\mathbb {Q}} \\mathcal {L} \\left(\\boldsymbol {w} _ {q} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {m}\\right) + b _ {q}, \\boldsymbol {y} _ {m}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.469, + 0.59 + ], + "angle": 0, + "content": "where \\(\\mathbb{P}\\) and \\(\\mathbb{Q}\\) are data distribution of the target and the pre-training tasks, \\(w_{q}, b_{q}\\) are the classification layer for pre-trained data from \\(\\mathbb{Q}\\). This method regularizes the current fine-tuned model feature extractor so that its prediction is still robust on the pre-training data. The joint training method is proposed in [46] to improve the performance of fine-tuning in downstream tasks where the training data is not sufficient." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Next, we test the performance of these two approaches on two standard image classification datasets, CIFAR10 [34] and Caltech-256 [22], with results shown in Figure 1. We use a grid search for learning rate and \\(\\lambda_{LwF}(\\lambda_{UOT})\\) and report the result of the model with the best robust accuracy. See Section 5 and the supplemental material for the experiment setting. On CIFAR10, both approaches fail to improve either clean or robust accuracy; on Caltech-256, the two approaches improve the clean accuracy by a small margin but deteriorate the robustness. One reason why the model- and data-based approaches fail is that the regularization term might be too strong, thus hurting the learning in the downstream." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.275, + 0.817 + ], + "angle": 0, + "content": "4. TWINS Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.899 + ], + "angle": 0, + "content": "The previous section demonstrates that both data- and model-based approaches cannot substantially improve the adversarial robustness in the downstream task. Thus we propose the TWINS for the better fine-tuning of robust pretrained models for downstream adversarial robustness." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.889, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.541 + ], + "angle": 0, + "content": "Figure 3. The mean and normalized STD of gradient norms in AT and TWINS-AT on CIFAR10, CIFAR100 and Caltech-256. The averaged \\(\\mu (\\| \\nabla \\pmb {w}\\|)\\) and \\(\\sigma (\\| \\nabla \\pmb {w}\\|) / \\mu (\\| \\nabla \\pmb {w}\\|)\\) over epochs are shown in each plot. The gradient magnitudes of TWINS-AT are substantially larger than those of AT, while the Normalized STDs of gradient norm in TWINS-AT are not obviously increased (CIFAR10) or even decreased (CIFAR100 and Caltech-256) compared with AT. This property leads to the faster escaping speed of TWINS-AT from the initial sub-optimum compared with AT (Fig. 4), and reduced robust overfitting (Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.673, + 0.559 + ], + "angle": 0, + "content": "4.1. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Though BN layers contain only a few parameters compared to convolution and fully connected layers, they play an important role in the good performance of DNNs. [20] shows that even if we only train the BN layers, the performance of a DNN is already quite impressive. [50] finds that only training the parameters in BN layers in an image generator is effective for small datasets. [38] proposes adaptive BN for domain adaptation, which updates the BN statistics with data from a target domain. These works motivate us to propose a statistics-based approach for retaining pretraining information in the target task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Typical BN layers track the mean and STD of the training set and save them for the inference stage. As this distribution information for each layer might be helpful for downstream robustness, we propose the TWINS robust finetuning, which maintains two networks, Frozen Net that uses the BN statistics from the robust pre-trained model, and Adaptive Net that learns its BN statistics from the downstream task. Instead of using two independent networks for the Frozen and Adaptive Net, we let the two networks share weight parameters, excluding the BN layers, to save the model size and inference time. At initialization, both" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16439" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.472, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.181, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Figure 4. The distance between the current step's weight \\(\\theta\\) and the initialization \\(\\theta_{pt}\\). On the three datasets, TWINS-AT has a faster escaping speed from the sub-optimal initial model than AT, which is due to TWIN-AT's larger gradient norm as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.469, + 0.369 + ], + "angle": 0, + "content": "networks and their BN statistics are initialized by the robust pre-trained models. During training, the Frozen Net uses the population means and STDs of pre-training data computed in the pre-training stage in the normalization operation, while the Adaptive Net uses the current batch's mean and STD in the normalization and updates its running mean and STD with the target training data. Fig. 2 shows the general pipeline of TWINS training and the network structure." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.37, + 0.469, + 0.398 + ], + "angle": 0, + "content": "The training objective of TWINS with adversarial training (TWINS-AT) in one mini-batch is:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.405, + 0.469, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\sum_ {j = 1} ^ {B / 2} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta} _ {a}} \\left(\\tilde {\\boldsymbol {x}} _ {j} ^ {(a)}\\right) + b, \\boldsymbol {y} _ {j}\\right) + \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.447, + 0.469, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {t w i n s} \\sum_ {i = B / 2 + 1} ^ {B} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta} _ {f}} \\left(\\tilde {\\boldsymbol {x}} _ {i} ^ {(a)}\\right) + b, \\boldsymbol {y} _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.718 + ], + "angle": 0, + "content": "where \\(\\theta_{f}\\) and \\(\\theta_{a}\\) denote the Frozen and Adaptive Nets respectively, and \\(\\tilde{\\pmb{x}}^{(a)}\\) is the adversarial image for \\(\\pmb{x}\\) when attacking the Adaptive Net \\(\\theta_{a}\\). We split the batch into two different sub-batches to avoid doubled batch sizes in TWINS training. Since the Frozen and Adaptive Nets share weight parameters, the number of parameters in TWINS is only increased by a very small amount (i.e., BN parameters). Thus, TWINS-AT only has a negligible cost in terms of memory and training time compared with vanilla AT. The TWINS structure can also be used in the TRADES in a similar way. Similar to [38], we can use the target training set to update the BN statistics so that they are more relevant to the downstream task. We call this procedure warmup in TWINS fine-tuning. The pseudo codes of TWINS-AT and TWINS-TRADES are given in Alg. 1 of the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.319, + 0.74 + ], + "angle": 0, + "content": "4.2. The mechanism of TWINS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The first benefit of TWINS is mentioned in the motivation of TWINS, i.e., the BN statistics have robustness information in the pre-training domain that can be leveraged by robust fine-tuning to improve the downstream robustness. This argument is validated by our ablation study in Section 5, where we initialize the means and STDs with 1.0 and 0.0 for Frozen Net instead of the pre-trained means and STDs and check the accuracy and robustness. Figure 5 shows that the TWINS with (1,0) initialization cannot match the performance with TWINS with pre-trained" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.885, + 0.163 + ], + "angle": 0, + "content": "
MethodRob. Acc.C10C100CaltechCUBDogs
ATBest ↑51.8431.3849.0927.0821.19
Final ↑49.4128.5248.3726.6019.80
Gap ↓2.432.860.730.481.39
TWINS-ATBest ↑53.2331.6048.8029.2420.89
Final ↑52.4031.0848.4029.2420.58
Gap ↓0.830.520.400.000.29
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Table 1. The robust accuracy drop of AT and TWINS-AT, where the adversarial attack is PGD10. TWINS-AT has smaller accuracy drop compared with AT, indicating that the TWINS-AT is less prone to robust overfitting as a result of reduced variance of gradient norms and stable training as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.247, + 0.891, + 0.277 + ], + "angle": 0, + "content": "statistics, indicating that the robustness information in pretraining is essential to the effectiveness of TWINS." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.278, + 0.892, + 0.413 + ], + "angle": 0, + "content": "It is intriguing that even with the (1,0) initialization, TWINS still outperforms the AT baseline in terms of robustness or accuracy on some datasets, e.g., Stanford Dogs and CIFAR100. This suggests that besides retaining the pre-training information, TWINS provides some other benefits during robust fine-tuning. By analyzing the gradient of TWINS, we find that TWINS implicitly increases the effective learning rate without increasing the oscillation and empirically validate this finding. We detail this analysis next." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.414, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Effective learning rate We first write the gradient of a weight vector for TWINS training. Consider the \\(l\\)-th layer's weight \\(\\boldsymbol{w}_j^{(l)}\\) and its output after BN layer" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.473, + 0.891, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {h} _ {i j} ^ {(l)} = \\frac {\\hat {h} _ {i j} ^ {(l)} - \\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\hat {h} _ {k j} ^ {(l)}}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} (\\hat {h} _ {k j} ^ {(l)} - \\hat {\\mu} _ {j} ^ {(l)}) ^ {2}}} (5) \\\\ = \\frac {\\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\left(\\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)\\right) ^ {2}}}, (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.587, + 0.891, + 0.651 + ], + "angle": 0, + "content": "where we denote \\( h_{ij}^{(l)}, \\hat{h}_{ij}^{(l)} \\) and \\( \\tilde{h}_{ij}^{(l)} \\) as the output of ReLU, convolution or fully connected layer and BN layer respectively, for \\( i \\)-th sample, \\( j \\)-th output variable at \\( l \\)-th layer. The gradient with respect to \\( \\boldsymbol{w}_j^{(l)} \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.66, + 0.897, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} = \\frac {\\partial \\mathcal {L} _ {a} (\\tilde {\\boldsymbol {x}} _ {i})}{\\partial \\boldsymbol {w} _ {j} ^ {(l)}} = \\frac {\\nabla \\tilde {h} _ {i j} ^ {(l)} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\left(\\boldsymbol {w} _ {j} ^ {(l) T} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)\\right) ^ {2}}} \\\\ - \\frac {\\nabla \\tilde {h} _ {i j} ^ {(l)} \\boldsymbol {w} _ {j} ^ {(l) T} (\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)})}{(\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} (\\boldsymbol {w} _ {j} ^ {(l) T} (\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)})) ^ {2}) ^ {3 / 2}} \\boldsymbol {\\Sigma} ^ {(l - 1)} \\boldsymbol {w} _ {j} ^ {(l)}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.756, + 0.892, + 0.852 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}^{(l-1)}\\) and \\(\\pmb{\\Sigma}^{(l-1)}\\) are the mean and covariance matrix from \\((l-1)\\)-th layer, \\(\\nabla \\tilde{h}_{ij}^{(l)}\\) represents the gradient of loss with respect to \\(\\tilde{h}_{ij}^{(l)}\\). If we write the weight as its norm \\(\\| \\pmb{w}_j^{(l)}\\|\\) multiplied by the unit vector \\(\\pmb{u}_j^{(l)}\\), there is a relationship between \\(\\nabla \\pmb{w}_j^{(l)}\\) and \\(\\| \\pmb{w}_j^{(l)}\\|\\)," + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.861, + 0.891, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} \\right\\| = \\frac {1}{\\left\\| \\boldsymbol {w} _ {j} ^ {(l)} \\right\\|} \\| \\nabla_ {a} \\boldsymbol {u} _ {j} ^ {(l)} \\| \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16440" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.31 + ], + "angle": 0, + "content": "This relationship has been found in [2, 27] and extended to any scale-invariant layers such as layer normalization [3] by [40]. It means that there are two ways to increase the gradient magnitude or the effective learning rate: 1) find a steeper descent direction where \\( \\|\\nabla_{a}\\pmb{u}_{j}^{(l)}\\| \\) is increased and 2) decrease the weight norm so that \\( 1 / \\| \\pmb{w}_j^{(l)}\\| \\) is increased. For a DNN with standard BN layers, the training will exploit this property to increase the gradient magnitude by reducing the weight norms with the help of weight decay regularization, which leads to a spurious increase in gradient magnitude and a larger variance of gradient estimation. In contrast, for a DNN with fixed BN layers such as the Frozen Net, the gradient norm is not correlated with weight norm," + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.319, + 0.47, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {f} \\boldsymbol {w} _ {j} ^ {(l)} = \\frac {\\partial \\mathcal {L} _ {f} (\\tilde {\\boldsymbol {x}} _ {i})}{\\partial \\boldsymbol {w} _ {j} ^ {(l)}} = \\nabla \\tilde {h} _ {i j} ^ {(l)} \\frac {\\boldsymbol {h} _ {i} ^ {(l - 1)}}{\\sigma_ {j , p t} ^ {(l)}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.416 + ], + "angle": 0, + "content": "In this gradient, the only way to increase the gradient magnitude is to find the actual steeper direction. The overall gradient for the weight is" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.426, + 0.47, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {w} _ {j} ^ {(l)} = \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} + \\lambda_ {t w i n s} \\nabla_ {f} \\boldsymbol {w} _ {j} ^ {(l)}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.749 + ], + "angle": 0, + "content": "Empirical Study To see the difference between the gradients in AT and TWINS-AT, we record the gradient of all weight parameters in each step of the two training methods and compute the mean and STD of the gradient norm in each epoch. The model parameters and their gradients are treated as long vectors and we compute the \\( l_{2} \\) norm of the weight and gradient vector. Figure 3 shows the mean and normalized STD of gradient norms in 60 training epochs for three datasets. Here we show the normalized STD, i.e., dividing the STD by the mean, to see the relative effect of variance. The major finding is that the gradient magnitude of TWINS-AT is substantially larger than that of AT, while the variance of TWINS-AT is lower than AT in most epochs. Note that in theory, the ratio between STD and mean should remain the same after down-scaling the weight norm in standard BN, but in practice we do observe the high normalized variance of DNNs with standard BNs since we use one epoch's gradients to approximate the variance and mean." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.904 + ], + "angle": 0, + "content": "One benefit of the larger gradient magnitude is that the model can escape from the initial sub-optimal point faster and find a better local optimum than the small gradient optimization [41]. We validate this hypothesis by recording the distance between the current model and the initial model during training. Figure 4 shows these weight distances on three datasets, where TWINS-AT moves away from the initial model much faster than the AT baseline. The robust overfitting effect of adversarial training [54] is partially a result of large gradient variance, especially at the final stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "of training [9]. We compare the robust accuracy drop of AT and TWINS-AT in Table 1 and find that the small relative variance of TWINS-AT has the effect of reducing robust overfitting. See the experiment details in Sec. 5 and Supp." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.168, + 0.625, + 0.185 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.892, + 0.24 + ], + "angle": 0, + "content": "This section presents our experiment with TWINS. We first introduce our experiment setting and then show our main result and ablation study." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.251, + 0.694, + 0.266 + ], + "angle": 0, + "content": "5.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Dataset. We use five datasets in our experiment. CIFAR10 and CIFAR100 [34] are low-resolution image datasets, where the training and validation sets have 50,000 and 10,000 images, and CIFAR10 has 10 classes, while CIFAR100 has 100 classes. Caltech-256 [22] is a high-resolution dataset with 30,607 images and 257 classes, which is split into training and validation set with a ratio of 9:1. Caltech-UCSD Birds-200-2011 (CUB200) [60] is a high-resolution bird image dataset for fine-grained image classification, which contains 200 classes of birds, 5,994 training images and 5,794 validation images. Stanford Dogs [33] has high-resolution dog images from 120 dog categories, where the training and validation set has 12,000 and 8,580 images. For both low-resolution image datasets (CIFAR10 and CIFAR100) and high-resolution datasets (Caltech-256,CUB200 and Stanford Dogs), we resize the image to \\(224\\times 224\\) so that the input sizes are the same for pre-training and fine-tuning. As with pre-training, the input image is normalized by the mean and STD of the pre-training set. Note that the resizing and normalization function is integrated into the model so we can attack the input image with the [0,1] bounds for pixel values as in standard adversarial attacks. We use the standard ImageNet data augmentation for high-resolution datasets [26]. For CIFAR datasets, we use random cropping with padding=4 and random horizontal flipping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Adversarial Pre-Training. Large-scale adversarial pretraining on ImageNet is time-consuming, and thus [55] has released adversially pre-trained ResNet50 and WideResNet50-2, trained with \\( l_{2} \\) and \\( l_{\\infty} \\) norm bounded attacks. In this paper, we adopt the pre-trained ResNet50 models, trained with \\( l_{\\infty} \\) attack with bound \\( \\epsilon_{pt} = 4 / 255 \\). We test other robust pre-trained models in our ablation study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Training Setting. For baselines and our method, we train all parameters of the pre-trained model, i.e., full fine-tuning instead of linear probing [57], with PGD attacks of \\( l_{\\infty} \\) norm. The PGD step is 10, \\( \\epsilon_{ft} = 8 / 255 \\) and stepsize \\( \\alpha = 2 / 255 \\). We set the batch size as 128 and train the model for 60 epochs and divide the learning rate by 0.1 at 30th and 50th epoch. In TWINS with warmup, we initialize the means and STDs with their pre-trained values, and update the means and STDs using the target training set." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16441" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.103, + 0.089, + 0.865, + 0.356 + ], + "angle": 0, + "content": "
MetricMethodCIFAR10CIFAR100Caltech256CUB200Stanford Dogs
Clean Acc.AT89.7769.4875.9065.7460.09
TWINS-AT91.24(+1.47)70.72(+1.24)76.86(+0.96)68.09(+2.35)64.98(+4.89)
TWINS-AT+warmup91.95(+2.18)72.12(+2.64)77.35(+1.45)67.64(+1.90)66.12(+6.03)
TRADES87.0662.7669.7058.9259.99
TWINS-TRADES86.61(-0.45)66.72(+3.96)71.12(+1.42)60.72(+1.80)60.58(+0.59)
TWINS-TRADES+warmup86.60(-0.46)65.91(+3.15)73.39(+3.69)61.05(+2.13)63.96(+3.97)
PGD10AT52.2428.5248.3726.6019.80
TWINS-AT52.73(+0.49)31.08(+2.56)48.40(+0.03)29.24(+2.64)20.58(+0.78)
TWINS-AT+warmup52.46(+0.22)29.12(+0.60)49.13(+0.76)27.67(+1.07)19.48(-0.32)
TRADES54.0432.2047.2827.8721.36
TWINS-TRADES56.23(+2.19)33.51(+1.31)47.31(+0.03)27.05(-0.82)19.93(-1.43)
TWINS-TRADES+warmup55.81(+1.77)33.48(+1.28)48.53(+1.25)26.68(-1.19)19.46(-1.90)
AAAT48.4623.4743.8522.8212.30
TWINS-AT49.81(+1.35)26.73(+3.26)43.69(-0.16)22.33(-0.49)14.37(+2.07)
TWINS-AT+warmup49.02(+0.56)25.72(+2.25)43.92(+0.07)23.58(+0.75)13.80(+1.50)
TRADES50.3126.4043.3922.2112.05
TWINS-TRADES51.71(+1.40)28.29(+1.89)41.77(-1.62)22.68(+0.47)13.36(+1.31)
TWINS-TRADES+warmup51.10(+0.79)28.30(+1.90)43.55(+0.16)21.92(-0.29)10.94(-1.11)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.357, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Table 2. The performance of our TWINS-AT and TWINS-TRADES on five image classification tasks compared with AT and TRADES. The clean accuracy means the accuracy when testing images are input without adversarial perturbations. PGD10 and AA denote the robust test accuracy under PGD10 and AutoAttack. The increase and decrease in performance are denoted with green and red numbers. The bold numbers denote the best robust accuracy under AA. The proposed TWINS achieves better robustness and clean accuracy compared with the baseline. Averaged over the datasets, the clean and robust accuracy of TWINS are increased by \\(2.18\\%\\) and \\(1.21\\%\\) compared with AT, and \\(1.46\\%\\) and \\(0.69\\%\\) compared with TRADES. The means and STDs of the performance are in the supplemental." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.447, + 0.465, + 0.547 + ], + "angle": 0, + "content": "
MetricPT ModelCaltech256CUB200Dogs
Clean AccRandom48.9912.387.27
Non-Robust64.6653.3041.99
Robust75.9065.7460.09
PGD10Random31.783.7283.59
Non-Robust39.8619.6813.32
Robust48.3726.6019.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Table 3. Comparison of random initialization, non-robust and robust pre-trained model on three difficult classification tasks, when fine-tuned with AT. The robust pre-trained model is indispensable to downstream robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.469, + 0.793 + ], + "angle": 0, + "content": "The momentum of updating statistics is 0.1, the batch size is 128, and the warmup only lasts one epoch. Note that in the warmup stage, the input samples are added with adversarial perturbations generated by the PGD attack, which has the same setting as the attack in training, and the classifier layer is the pre-trained classifier for the adversarial attack. Our pilot experiment shows that using adversarial examples as input is more effective than using clean exemplars in the warmup. The optimizer is SGD with momentum in all of our experiments. The learning rate, weight decay and regularization hyperparameter are determined by grid search, which is described in detail in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Adversarial robustness evaluation. Two standard adversarial attacks are used in our experiment, i.e., PGD and AutoAttack [11]. The adversarial perturbation is \\( l_{\\infty} \\) norm bounded in our evaluation. The setting of PGD attack for validation set is the same as the PGD attack in training. The AutoAttack (AA) is a more reliable adversarial attack and more often used for evaluation than PGD in recent years." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.448, + 0.895, + 0.535 + ], + "angle": 0, + "content": "
MethodMetricλWD
1e-51e-41e-31e-2
ATClean Acc.89.9289.9489.7790.28
Robust Acc.46.3444.6748.4647.57
TWINS-ATClean Acc.91.9091.4291.2487.33
Robust Acc.47.1949.0749.8151.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Table 4. The performance of TWINS-AT and AT on CIFAR10 when the hyperparameter for weight decay \\( {\\lambda }_{WD} \\) is changed. The robust accuracy is evaluated using AutoAttack. Our TWINS-AT achieves better adversarial robustness than AT for different \\( {\\lambda }_{WD} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.608, + 0.892, + 0.669 + ], + "angle": 0, + "content": "We use the standard attacks of AA, i.e., untargeted APGD-CE, targeted APGD-DLR, targeted FAB [10] and Square Attack [1], with \\(\\epsilon = 8 / 255\\). The robust accuracy in our experiment result denotes the accuracy under AA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.695, + 0.696 + ], + "angle": 0, + "content": "5.2. Experimental Result" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.854 + ], + "angle": 0, + "content": "On the three high-resolution datasets, we compare the performance of fine-tuning different initialization models, i.e., random initialization, standard pre-trained ResNet50 and robust pre-trained ResNet50. Table 3 shows that the pre-trained models are essential to the accuracy and robustness in challenging downstream tasks, since the random initialization is much worse than the two pre-trained models. The robust pre-trained model has a clear benefit over the standard pre-trained one, indicating that robust pre-training is indispensable to downstream robustness." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows the result of TWINS-AT and TWINS-TRADES compared with the baselines. Since AA is a more reliable attack, we highlight the best robust accuracy un" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16442" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.211, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Figure 5. Ablation study where the BN statistics in TWINS are initialized with (0,1) for means and STDs, denoted as TWINS-Init. The population mean and STD of pre-training are crucial to TWINS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.499 + ], + "angle": 0, + "content": "der AA on each dataset. The TWINS fine-tuning learns more robust DNNs, as well as achieves a better clean accuracy on all five datasets, demonstrating the strong effectiveness of TWINS in the robust pre-training and robust fine-tuning setting. On CIFAR10 and CIFAR100, both TWINS-AT and TWINS-TRADES achieve better robustness and clean accuracy than their baselines, while the warmup only improves the clean accuracy and hurts the robustness sometimes. On Caltech-256, TWINS-AT improves upon the vanilla AT in both robustness and accuracy, but TWINS-TRADES does not perform better than vanilla TRADES. However, the warmup helps boost the performance of TWINS-TRADES as well as TWINS-AT and makes the TWINS with warmup perform better than the baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.47, + 0.698 + ], + "angle": 0, + "content": "On the two fine-grained image classification datasets, TWINS-AT and TWINS-TRADES generally perform better than baselines in terms of accuracy and robustness, if we look at the robust accuracy under AA, where only TWINS-AT on CUB200 has a slightly worse robust accuracy than its baseline. We find that the warmup improves the clean accuracy but hurts the robustness in most cases, except for CIFAR100 and Caltech-256. This can be a result of noisy adversarial perturbation, since we use the pre-trained classifier layer in the adversarial attack, or the insufficient update steps. Nevertheless, we note that the warmup can be considered as an operation for achieving a trade-off between robustness and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.232, + 0.727 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "TWINS initialization. We use the pre-trained BN statistics in the Frozen Net to keep the robust information learned during pre-training. To show the importance of the pretrained BN statistics, we use the standard initialization (mean=0 and STD=1) for BN statistics in the Frozen Net, denoted as TWINS-Init, and show the result on the five datasets in Figure 5. Both clean and robust accuracy drop when the (0,1) initialization is used in TWINS-Init, demonstrating the crucial role of pre-trained statistics in TWINS. The fact that TWINS-Init sometimes improves upon the AT baseline motivates us to investigate the effect of TWINS on" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.091, + 0.898, + 0.178 + ], + "angle": 0, + "content": "
MethodMetricεpt
1/2552/2554/2558/255
ATClean Acc.65.4767.0869.4869.93
Robust Acc.24.7625.7923.4727.71
TWINS-ATClean Acc.68.5570.4570.7272.59
Robust Acc.25.9726.6226.7328.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.179, + 0.892, + 0.207 + ], + "angle": 0, + "content": "Table 5. The performance of TWINS-AT and AT on CIFAR100 when robust pre-trained ResNet50 with different \\( \\epsilon_{pt} \\) are used." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.217, + 0.689, + 0.231 + ], + "angle": 0, + "content": "gradient norms in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Effect of weight decay. Weight decay is the reason for the decreasing weight norm in DNNs with BN layers, so increasing the hyperparameter of weight decay \\(\\lambda_{WD}\\) is also a way to increase the gradient magnitude. Table 4 shows the result of TWINS-AT and AT when different \\(\\lambda_{WD}\\) are used when fine-tuning the robust pre-trained model on CIFAR10. The robust accuracy of TWINS-AT is consistently better than that of AT across different \\(\\lambda_{WD}\\)'s. We draw the same conclusion on CIFAR100 (see supplemental). Note that the clean accuracy of TWINS-AT drops when a large \\(\\lambda_{WD}\\) is used, suggesting that we should not use a too large \\(\\lambda_{WD}\\) for TWINS-AT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Different robust pre-trained models. The main experiment uses the robust pre-trained ResNet50 with \\(\\epsilon_{pt} = 4/255\\) as the initial model. We try different robust pre-trained models with different \\(\\epsilon_{pt}\\) in Table 5, which shows that a larger \\(\\epsilon_{pt}\\) is beneficial to both clean and robust accuracy in the downstream, and the proposed TWINS-AT is better than AT in both metrics with different pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.54, + 0.619, + 0.554 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.825 + ], + "angle": 0, + "content": "This paper investigates the utility of robust pre-trained models in various downstream classification tasks. We first find that the commonly used data- and model-based approaches to maintain pre-training information do not work in the adversarial robust fine-tuning. We then propose a subtle statistics-based method, TWINS, for retaining the pretraining robustness in the downstream. In addition to the robustness preserving effect, we find that TWINS increases the gradient magnitude without sacrificing the training stability and improves the training dynamics of AT. Finally, the performance of TWINS is shown to be stronger than that of AT and TRADES on five datasets. One limitation of our work is that we only evaluate the robust supervised pre-trained ResNet50. Recently, robust pre-trained ViT's on ImageNet [4] have been released. Our statistics-based approach can be extended to the layer normalization, on which the increasing gradient magnitude argument also holds, and thus future work will extend TWINS to ViT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11215820) and the Fundamental Research Funds for the Central University of China (DUT No. 82232031)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16443" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 34:26831-26843, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Dian Chen, Hongxin Hu, Qian Wang, Li Yinli, Cong Wang, Chao Shen, and Qi Li. Cartl: Cooperative adversarially-robust transfer learning. In International Conference on Machine Learning, pages 1640–1650. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.472, + 0.614 + ], + "angle": 0, + "content": "[8] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 699-708, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[9] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In International Conference on Learning Representations, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.673, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[10] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In International Conference on Machine Learning, pages 2196-2205. PMLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.731, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.788, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[14] Gavin Weiguang Ding, Yash Sharma, Kry Yuk Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[15] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, et al. On robustness and transferability of convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16458-16468, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[16] Xinshuai Dong, Anh Tuan Luu, Min Lin, Shuicheng Yan, and Hanwang Zhang. How should pre-trained language models be fine-tuned towards adversarial robustness? Advances in Neural Information Processing Systems, 34:4356-4369, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.893, + 0.468 + ], + "angle": 0, + "content": "[18] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature communications, 12(1):1-14, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.893, + 0.523 + ], + "angle": 0, + "content": "[19] Samuel G Finlayson, John D Bowers, Joichi Ito, Jonathan L Zittrain, Andrew L Beam, and Isaac S Kohane. Adversarial attacks on medical machine learning. Science, 363(6433):1287-1289, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[20] Jonathan Frankle, David J. Schwab, and Ari S. Morcos. Training batchnorm and only batchnorm: On the expressive power of random features in {cnn}s. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.894, + 0.636 + ], + "angle": 0, + "content": "[21] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.893, + 0.664 + ], + "angle": 0, + "content": "[22] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.894, + 0.734 + ], + "angle": 0, + "content": "[23] Minghao Guo, Yuzhe Yang, Rui Xu, Ziwei Liu, and Dahua Lin. When nas meets robustness: In search of robust architectures against adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 631-640, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16444" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[27] Elad Hoffer, Ron Banner, Itay Golan, and Daniel Soudry. Norm matters: efficient and accurate normalization schemes in deep networks. Advances in Neural Information Processing Systems, 31, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[28] Hanxun Huang, Yisen Wang, Sarah Erfani, Quanquan Gu, James Bailey, and Xingjun Ma. Exploring architectural ingredients of adversarially robust deep neural networks. Advances in Neural Information Processing Systems, 34:5545-5559, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[29] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[30] Daniel Jakubovitz and Raja Giryes. Improving dnn robustness to adversarial attacks using jacobian regularization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 514-529, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[31] Ahmadreza Jeddi, Mohammad Javad Shafiee, and Alexander Wong. A simple fine-tuning is all you need: Towards robust deep learning via adversarial fine-tuning. arXiv preprint arXiv:2012.13628, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[32] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 33:16199-16210, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[33] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[35] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[36] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[37] Yao Li, Martin Renqiang Min, Thomas Lee, Wenchao Yu, Erik Kruus, Wei Wang, and Cho-Jui Hsieh. Towards robustness of deep neural networks via regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7496-7505, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[38] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[39] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Zhiyuan Li, Kaifeng Lyu, and Sanjeev Arora. Reconciling modern deep learning with traditional optimization analyses: The intrinsic learning rate. Advances in Neural Information Processing Systems, 33:14544-14555, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[41] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. Advances in Neural Information Processing Systems, 33:21476-21487, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[42] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128(2):261-318, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[43] Ziquan Liu and Antoni B Chan. Boosting adversarial robustness from the perspective of effective margin regularization. British Machine Vision Conference (BMVC), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[44] Ziquan Liu, Yufei Cui, and Antoni B Chan. Improve generalization and robustness of neural networks via weight scale shifting invariant regularizations. ICML Workshop on Adversarial Machine Learning, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[46] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[47] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, Xiangyang Ji, and Antoni B Chan. An empirical study on distribution shift robustness from the perspective of pre-training and data augmentation. arXiv preprint arXiv:2205.12753, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[49] Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[50] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[51] Chongli Qin, James Martens, Sven Gowal, Dilip Krishnan, Krishnamurthy Dvijotham, Alhussein Fawzi, Soham De, Robert Stanforth, and Pushmeet Kohli. Adversarial robustness through local linearization. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[53] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Data augmentation can improve robustness. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16445" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[54] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093–8104. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[55] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[56] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[57] Ali Shafahi, Parsa Saadatpanah, Chen Zhu, Amin Ghiasi, Christoph Studer, David Jacobs, and Tom Goldstein. Adversarially robust transfer learning. In International Conference on Learning Representations, 2020. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[58] Chawin Sitawarin, Arvind Sridhar, and David Wagner. Improving the accuracy-robustness trade-off for dual-domain adversarial training. ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[59] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[60] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[61] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[62] Yutaro Yamada and Mayu Otani. Does robustness on image net transfer to downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9215-9224, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[63] Yaodong Yu, Heinrich Jiang, Dara Bahri, Hossein Mobahi, Seungyeon Kim, Ankit Singh Rawat, Andreas Veit, and Yi Ma. An empirical study of pre-trained vision models on out-of-distribution generalization. In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.47, + 0.782 + ], + "angle": 0, + "content": "[64] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16446" + } + ] +] \ No newline at end of file diff --git a/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_origin.pdf b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec38e7134935a9bca22fb873249a3bb78e311443 --- /dev/null +++ b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/178a6a71-7aa4-4ba0-ae4e-9fff019b01f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ac6c2bff8f78e2897751f203e1dc61f90df66f7f63bc7f386c12d7ab133fdcb +size 842894 diff --git a/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/full.md b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4e57178aef4cc373d40070ef0218cd067989376 --- /dev/null +++ b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/full.md @@ -0,0 +1,294 @@ +# TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization + +Ziquan Liu $^{1}$ , Yi Xu $^{2*}$ , Xiangyang Ji $^{3}$ and Antoni B. Chan $^{1*}$ + +$^{1}$ Department of Computer Science, City University of Hong Kong + +$^{2}$ School of Artificial Intelligence, Dalian University of Technology + +$^{3}$ Department of Automation, Tsinghua University + +ziquanliu.cs@gmail.com, yxu@dlut.edu.cn, xyji@tsinghua.edu.cn, abchan@cityu.edu.hk + +# Abstract + +Recent years have seen the ever-increasing importance of pre-trained models and their downstream training in deep learning research and applications. At the same time, the defense for adversarial examples has been mainly investigated in the context of training from random initialization on simple classification tasks. To better exploit the potential of pre-trained models in adversarial robustness, this paper focuses on the fine-tuning of an adversially pre-trained model in various classification tasks. Existing research has shown that since the robust pre-trained model has already learned a robust feature extractor, the crucial question is how to maintain the robustness in the pre-trained model when learning the downstream task. We study the model-based and data-based approaches for this goal and find that the two common approaches cannot achieve the objective of improving both generalization and adversarial robustness. Thus, we propose a novel statistics-based approach, TwoWing NormliSation (TWINS) fine-tuning framework, which consists of two neural networks where one of them keeps the population means and variances of pre-training data in the batch normalization layers. Besides the robust information transfer, TWINS increases the effective learning rate without hurting the training stability since the relationship between a weight norm and its gradient norm in standard batch normalization layer is broken, resulting in a faster escape from the sub-optimal initialization and alleviating the robust overfitting. Finally, TWINS is shown to be effective on a wide range of image classification datasets in terms of both generalization and robustness. + +# 1. Introduction + +The adversarial vulnerability of deep neural networks (DNNs) [59] is one of the major obstacles for their wide + +![](images/0da221f25917e7ef65d4a2563fe4e0898588b55aa9551d9aadb17eef084a780f.jpg) +Figure 1. The performance of fine-tuning robust and non-robust large-scale pre-trained (PT) ResNet50 [26, 55] on CIFAR10 [34] and Caltech-256 [22]. We compare standard adversarial training (AT), Learning without Forgetting (LwF) (model approach) [39], joint fine-tuning with UOT data selection (data approach) [46] and our TWINS fine-tuning. The robust accuracy is evaluated using $l_{\infty}$ norm bounded AutoAttack [11] with $\epsilon = 8 / 255$ . On CIFAR10, the data-based and model-based approach fail to improve clean and robust accuracy. On Caltech, both approaches improve the clean accuracy but hurt the robust accuracy. Our TWINS fine-tuning improves the clean and robust performance on both datasets. The pink triangle denotes the performance of standard AT with the non-robust pre-trained ResNet50, which drops considerably compared with fine-tuning starting from the robust pretrained model. + +![](images/d68393543ae72d019748e33d6fb39257d9b60d2a2d536d365a5f3644c161b2b0.jpg) + +applications in safety-critical scenarios such as self-driving cars [18] and medical diagnosis [19]. Thus, addressing this issue has been one focus of deep learning research in the past eight years. Existing works have proposed to improve adversarial robustness from different perspectives, including data augmentation [21, 48, 53, 56], regularization [37, 43, 44, 51] and neural architecture [23, 28]. However, most of existing works investigate the problem under the assumption that the training data is sufficient enough, and training from scratch gives a satisfactory performance, which is not realistic in the real world. There are a large number of computer vision tasks where training from scratch is inferior to training from pre-trained weights, such as fine-grained image classification (e.g., Caltech- + +![](images/98f62d62bc6ae35d3945d6926361d03650ee38345992915c930d1c0be0a55fb3.jpg) +(a) TWINS Structure +Figure 2. The TWINS structure and training pipeline. (a) The Frozen Net and Adaptive Net have the same structure and share the weight parameters, except for batch normalization (BN) layers. The Frozen Net uses pre-trained means and standard deviations (STD) in the normalization layer, while Adaptive Net uses the mean and STD computed from the current batch as in standard BN. (b) In each step of mini-batch stochastic gradient descent (SGD), we split the batch of adversarial examples, generated from attacking the Adaptive Net, into two sub-batches and feed them to the Adaptive Net and Frozen Net respectively. The loss of two networks are combined and backpropagated to their shared parameters to train the network. In the inference stage, only the Adaptive Net is used. + +![](images/64cb32692987443b8900d6c4735c18d6bd1bb99180d034d6bad575cb6a6a6670.jpg) +(b) Training Pipeline + +UCSD Birds-200-2011 or CUB200 [60]), object detection [42] and semantic segmentation [49]. + +On the other hand, pre-trained models have been considered as the foundation models in deep learning [5] as a result of their strong performance and wide employment in computer vision [17, 24, 25, 45], as well as natural language processing [6, 13, 52]. Thus, how to better use the pre-trained model in downstream has emerged as a major research topic in many vision and language tasks, such as image classification under distribution shifts [47, 63], object detection [36] and semantic segmentation [29, 35]. There are a few papers that investigate the pre-trained model's robustness in target tasks [7, 8, 15, 31, 32, 57, 62]. [7, 57] mainly considers the transfer between small-scale datasets (e.g., CIFAR100 to CIFAR10), while [8, 32] use adversarial robust pre-training and fine-tuning on the same dataset, without considering a large-scale and general pre-trained model. Finally, [15, 62] investigate different kinds of robustness to corruption or out-of-distribution samples, and are not devoted to adversarial robustness. + +In this paper, we consider how to transfer the adversarial robustness of a large-scale robust pre-trained model (e.g., a ResNet50 pre-trained on ImageNet [12] with adversarial training) on various downstream classification tasks when fine-tuning with adversarial training. This problem setting is becoming more important as the standard pretrained models do not learn robust representations from the pre-training data and are substantially weaker than the robust pre-trained counterparts in some challenging downstream tasks, e.g., fine-grained classification as shown in our experiment. Meanwhile, more large-scale robust pretrained models are released (e.g., ResNet [55] and ViT [4]), which makes the robust pre-trained models more accessible. However, naively applying adversarial training to fine-tune from the robustly pre-trained model will lead to subopti + +mal robustness, since the robust representations learned by the robust pre-trained model are not fully utilized. For example, [57] suggests that the robustness from a pre-trained model needs to be explicitly maintained for its better transfer to the downstream. + +Following the idea that the key to improving the transferability of robustness is to maintain the robustness of the pre-training stage during fine-tuning [57], we first evaluate the data-based and model-based approach on two representative datasets, CIFAR10 and Caltech-256. The database-based approach uses pre-trained data in the fine-tuning and keeps their performance under adversarial attack, while the model-based approach regularizes the distance of features of the fine-tuned and pre-trained model. Our experiment shows that both methods fail to improve the robustness and generalization (Fig. 1), since the two methods are too aggressive in retaining the robustness and hurt the learning in downstream. Thus, we propose a subtle approach that keeps the batch-norm (BN) statistics of pre-training for preserving the robustness, which we call Two-Wing NormaliSation (TWINS) fine-tuning. TWINS has two neural networks with fixed and adaptive BN layers respectively, where the fixed BN layers use the population means and STDs of pretraining for normalization, while the adaptive BN layers use the standard BN normalization. Our experiment first demonstrates the importance of pre-trained BN statistics in the robust fine-tuning and then finds the benefit of TWINS in adversarial training dynamics. As the relationship between weight norm and its gradient norm no longer holds in TWINS, it is able to increase the gradient magnitude without increasing the gradient variance. At the initial training stage, TWINS has a faster escaping speed from the suboptimal initialization than vanilla adversarial training [41]. At the final training stage, the gradient of TWINS is more stable than adversarial training, which alleviates the robust + +overfitting effect [56]. In summary, the contributions of our paper are as follows: + +1. We focus on the fine-tuning of large-scale robust pretrained models as a result of their potential importance in various downstream tasks. We evaluate current approaches to retain the pre-training robustness in fine-tuning, and show that they cannot substantially improve the robustness. +2. We propose TWINS, a statistics-based approach for better transferability of robustness and generalization from the pre-training domain to the target domain. TWINS has two benefits: a) it keeps the robust statistics for downstream tasks, thus helps the transfer the robustness to downstream tasks and b) it enlarges the gradient magnitude without increasing gradient variance, thus helps the model escape from the initialization faster and mitigates robust overfitting. The mechanisms of these two benefits are validated by our empirical study. +3. The effectiveness of TWINS is corroborated on five downstream datasets by comparing with two popular adversarial training baselines, adversarial training (AT) [48] and TRADES [64]. On average, TWINS improves the clean and robust accuracy by $2.18\%$ and $1.21\%$ compared with AT, and by $1.46\%$ and $0.69\%$ compared with TRADES. The experiment shows the strong potential of robust pre-trained models in boosting downstream's robustness and generalization when using more effective fine-tuning methods. + +# 2. Related Work + +Adversarial defense. There are several major approaches to improving the adversarial robustness of DNNs. The training of DNNs can be regularized to induce biases that are beneficial to adversarial robustness, such as locally linear regularization [51], margin maximization [14, 43] and Jacobian regularization [30]. The most commonly used adversarial defense is adversarial training (AT) [48], which directly trains the DNN on adversarial examples generated from PGD attack. Later, TRADES [64] is proposed to add a KL regularization to AT and achieves stronger adversarial robustness. Our paper proposes TWINS to improve adversarial training in the fine-tuning stage when the initial model is adversarially pre-trained. We compare TWINS-AT and TWINS-TRADES with vanilla AT and TRADES in our experiment and show the strong effectiveness of TWINS in the robust fine-tuning setting. + +Fine-tuning for downstream robustness. Several aspects of robustness in pre-training and fine-tuning have been studied in existing works. Adversarial contrastive learning [8, 32] is proposed to pre-train on a dataset with contrastive learning and then fine-tune on the same dataset, without considering the transferability of robustness from a large- + +scale pre-trained model to a different downstream task. In contrast, our paper investigates a more general problem, where task-specific pre-training is not needed for a new task as we use one robust large-scale pre-trained model trained on ImageNet. [55] considers the robust pre-training on the large-scale ImageNet and its transfer to downstream tasks, but focuses on the performance on clean instead of adversarial images. The Learning-without-Forgotten (LwF) [39] approach for retaining robustness is shown to be effective in the small-scale transfer experiment [57], but is not effective in our experiment setting of transfer of large-scale models. [31] proposes a learning rate schedule to improve the adversarial robustness of fine-tuned models, and [16] proposes robust informative fine-tuning for pre-trained language models to robustly keep pre-training information in downstream. The difference between [16, 31] and our work is that they assume a standard pre-trained model instead of the adversarial pre-trained model. [15, 62] investigate the performance of pre-trained models in downstream tasks, but the focus is the robustness to out-of-distribution samples instead of adversarial perturbations. + +Batch normalization. There are existing papers proposing the two-branch BN structure for different purposes with different technical details. [58] proposes dual normalization for a better trade-off between accuracy and robustness, where the normalization is a weighted sum of normalized clean and adversarial input. [61] proposes a similar two-branch BN structure, where one branch is for adversarial examples and the other is for clean examples. The major difference between our work and [58, 61] is that both BN branches in TWINS are for adversarial examples and one branch (Frozen Net) has fixed BN statistics from pretraining so as to better maintain the pre-trained robustness, whereas [58, 61] uses clean examples in BN and aims to improve the accuracy for clean images. + +# 3. The Model-based and Data-based Approach to Retaining Adversarial Robustness + +This section introduces the two common approaches for keeping adversarial robustness of pre-training in downstream, model-based and data-based approaches. Denote the feature vector output of a neural network as $g_{\theta}(\boldsymbol{x})$ , the training sample in the downstream task as $\{(x_i,y_i)\}_{i=1}^N \sim \mathbb{P}$ , and the loss function as $\mathcal{L}(\boldsymbol{w}^T g_{\theta}(\boldsymbol{x}) + b, \boldsymbol{y})$ , where $(\boldsymbol{w}, b)$ are the parameters of last classification layer. We assume that the pre-trained model is trained on adversarial examples generated from the PGD attack [48], where the $l_{\infty}$ norm of the adversarial attack is bounded by $\epsilon$ , and during fine-tuning we use adversarial training with the same PGD attack to obtain adversarial robustness in downstream tasks. In short, we consider the robust pre-training and robust fine-tuning setting in this paper, if not specified otherwise. + +Model-based approaches. We first introduce the model-based approach, which keeps the pre-trained model $\theta_{pt}$ during fine-tuning so as to maintain its robustness. The objective function is + +$$ +\sum_ {\left(\boldsymbol {x} _ {n}, \boldsymbol {y} _ {n}\right) \sim \mathbb {P}} \mathcal {L} \left(\boldsymbol {w} ^ {T} g _ {\boldsymbol {\theta}} \left(\tilde {\boldsymbol {x}} _ {n}\right) + b, \boldsymbol {y} _ {n}\right) + \lambda_ {L w F} \| g _ {\boldsymbol {\theta} _ {p t}} \left(\tilde {\boldsymbol {x}} _ {n}\right) - g _ {\boldsymbol {\theta}} \left(\tilde {\boldsymbol {x}} _ {n}\right) \| _ {2}, \tag {1} +$$ + +where the adversarial example generated from $\boldsymbol{x}_n$ is denoted as $\tilde{\boldsymbol{x}}_n$ . The regularization term of the loss aims to minimize the distance between the features from the pretrained and the fine-tuned models, which is expected to maintain the robustness of the pre-trained model. This approach is originally proposed in [39] to prevent the catastrophic forgetting in continual learning, and is used in [57] to preserve adversarial robustness in transfer learning. Note that [57] uses the LwF method in standard fine-tuning instead of robust fine-tuning as in our paper. + +Data-based approaches. The objective function of the data-based approach is + +$$ +\begin{array}{l} \sum_ {\left(\boldsymbol {x} _ {n}, \boldsymbol {y} _ {n}\right) \sim \mathbb {P}} \mathcal {L} \left(\boldsymbol {w} ^ {T} g _ {\boldsymbol {\theta}} \left(\tilde {\boldsymbol {x}} _ {n}\right) + b, \boldsymbol {y} _ {n}\right) \\ + \lambda_ {U O T} \sum_ {\left(\boldsymbol {x} _ {m}, \boldsymbol {y} _ {m}\right) \sim \mathbb {Q}} \mathcal {L} \left(\boldsymbol {w} _ {q} ^ {T} g _ {\boldsymbol {\theta}} \left(\tilde {\boldsymbol {x}} _ {m}\right) + b _ {q}, \boldsymbol {y} _ {m}\right), \tag {2} \\ \end{array} +$$ + +where $\mathbb{P}$ and $\mathbb{Q}$ are data distribution of the target and the pre-training tasks, $w_{q}, b_{q}$ are the classification layer for pre-trained data from $\mathbb{Q}$ . This method regularizes the current fine-tuned model feature extractor so that its prediction is still robust on the pre-training data. The joint training method is proposed in [46] to improve the performance of fine-tuning in downstream tasks where the training data is not sufficient. + +Next, we test the performance of these two approaches on two standard image classification datasets, CIFAR10 [34] and Caltech-256 [22], with results shown in Figure 1. We use a grid search for learning rate and $\lambda_{LwF}(\lambda_{UOT})$ and report the result of the model with the best robust accuracy. See Section 5 and the supplemental material for the experiment setting. On CIFAR10, both approaches fail to improve either clean or robust accuracy; on Caltech-256, the two approaches improve the clean accuracy by a small margin but deteriorate the robustness. One reason why the model- and data-based approaches fail is that the regularization term might be too strong, thus hurting the learning in the downstream. + +# 4. TWINS Fine-Tuning + +The previous section demonstrates that both data- and model-based approaches cannot substantially improve the adversarial robustness in the downstream task. Thus we propose the TWINS for the better fine-tuning of robust pretrained models for downstream adversarial robustness. + +![](images/cea56a67dcb0f72a2574e945b38f3b99df3eca33da144cb0c5c4831967c10967.jpg) +Figure 3. The mean and normalized STD of gradient norms in AT and TWINS-AT on CIFAR10, CIFAR100 and Caltech-256. The averaged $\mu (\| \nabla \pmb {w}\|)$ and $\sigma (\| \nabla \pmb {w}\|) / \mu (\| \nabla \pmb {w}\|)$ over epochs are shown in each plot. The gradient magnitudes of TWINS-AT are substantially larger than those of AT, while the Normalized STDs of gradient norm in TWINS-AT are not obviously increased (CIFAR10) or even decreased (CIFAR100 and Caltech-256) compared with AT. This property leads to the faster escaping speed of TWINS-AT from the initial sub-optimum compared with AT (Fig. 4), and reduced robust overfitting (Tab. 1). + +# 4.1. Proposed Method + +Though BN layers contain only a few parameters compared to convolution and fully connected layers, they play an important role in the good performance of DNNs. [20] shows that even if we only train the BN layers, the performance of a DNN is already quite impressive. [50] finds that only training the parameters in BN layers in an image generator is effective for small datasets. [38] proposes adaptive BN for domain adaptation, which updates the BN statistics with data from a target domain. These works motivate us to propose a statistics-based approach for retaining pretraining information in the target task. + +Typical BN layers track the mean and STD of the training set and save them for the inference stage. As this distribution information for each layer might be helpful for downstream robustness, we propose the TWINS robust finetuning, which maintains two networks, Frozen Net that uses the BN statistics from the robust pre-trained model, and Adaptive Net that learns its BN statistics from the downstream task. Instead of using two independent networks for the Frozen and Adaptive Net, we let the two networks share weight parameters, excluding the BN layers, to save the model size and inference time. At initialization, both + +![](images/26134a5960d82f6e7b6c7c25ab0464e79a784705b0138551bc6674497cd7727a.jpg) +Figure 4. The distance between the current step's weight $\theta$ and the initialization $\theta_{pt}$ . On the three datasets, TWINS-AT has a faster escaping speed from the sub-optimal initial model than AT, which is due to TWIN-AT's larger gradient norm as shown in Figure 3. + +networks and their BN statistics are initialized by the robust pre-trained models. During training, the Frozen Net uses the population means and STDs of pre-training data computed in the pre-training stage in the normalization operation, while the Adaptive Net uses the current batch's mean and STD in the normalization and updates its running mean and STD with the target training data. Fig. 2 shows the general pipeline of TWINS training and the network structure. + +The training objective of TWINS with adversarial training (TWINS-AT) in one mini-batch is: + +$$ +\sum_ {j = 1} ^ {B / 2} \mathcal {L} \left(\boldsymbol {w} ^ {T} g _ {\boldsymbol {\theta} _ {a}} \left(\tilde {\boldsymbol {x}} _ {j} ^ {(a)}\right) + b, \boldsymbol {y} _ {j}\right) + \tag {3} +$$ + +$$ +\lambda_ {t w i n s} \sum_ {i = B / 2 + 1} ^ {B} \mathcal {L} \left(\boldsymbol {w} ^ {T} g _ {\boldsymbol {\theta} _ {f}} \left(\tilde {\boldsymbol {x}} _ {i} ^ {(a)}\right) + b, \boldsymbol {y} _ {i}\right), \tag {4} +$$ + +where $\theta_{f}$ and $\theta_{a}$ denote the Frozen and Adaptive Nets respectively, and $\tilde{\pmb{x}}^{(a)}$ is the adversarial image for $\pmb{x}$ when attacking the Adaptive Net $\theta_{a}$ . We split the batch into two different sub-batches to avoid doubled batch sizes in TWINS training. Since the Frozen and Adaptive Nets share weight parameters, the number of parameters in TWINS is only increased by a very small amount (i.e., BN parameters). Thus, TWINS-AT only has a negligible cost in terms of memory and training time compared with vanilla AT. The TWINS structure can also be used in the TRADES in a similar way. Similar to [38], we can use the target training set to update the BN statistics so that they are more relevant to the downstream task. We call this procedure warmup in TWINS fine-tuning. The pseudo codes of TWINS-AT and TWINS-TRADES are given in Alg. 1 of the supplemental. + +# 4.2. The mechanism of TWINS + +The first benefit of TWINS is mentioned in the motivation of TWINS, i.e., the BN statistics have robustness information in the pre-training domain that can be leveraged by robust fine-tuning to improve the downstream robustness. This argument is validated by our ablation study in Section 5, where we initialize the means and STDs with 1.0 and 0.0 for Frozen Net instead of the pre-trained means and STDs and check the accuracy and robustness. Figure 5 shows that the TWINS with (1,0) initialization cannot match the performance with TWINS with pre-trained + +
MethodRob. Acc.C10C100CaltechCUBDogs
ATBest ↑51.8431.3849.0927.0821.19
Final ↑49.4128.5248.3726.6019.80
Gap ↓2.432.860.730.481.39
TWINS-ATBest ↑53.2331.6048.8029.2420.89
Final ↑52.4031.0848.4029.2420.58
Gap ↓0.830.520.400.000.29
+ +Table 1. The robust accuracy drop of AT and TWINS-AT, where the adversarial attack is PGD10. TWINS-AT has smaller accuracy drop compared with AT, indicating that the TWINS-AT is less prone to robust overfitting as a result of reduced variance of gradient norms and stable training as shown in Fig. 3. + +statistics, indicating that the robustness information in pretraining is essential to the effectiveness of TWINS. + +It is intriguing that even with the (1,0) initialization, TWINS still outperforms the AT baseline in terms of robustness or accuracy on some datasets, e.g., Stanford Dogs and CIFAR100. This suggests that besides retaining the pre-training information, TWINS provides some other benefits during robust fine-tuning. By analyzing the gradient of TWINS, we find that TWINS implicitly increases the effective learning rate without increasing the oscillation and empirically validate this finding. We detail this analysis next. + +Effective learning rate We first write the gradient of a weight vector for TWINS training. Consider the $l$ -th layer's weight $\boldsymbol{w}_j^{(l)}$ and its output after BN layer + +$$ +\begin{array}{l} \tilde {h} _ {i j} ^ {(l)} = \frac {\hat {h} _ {i j} ^ {(l)} - \frac {2}{B} \sum_ {k = 1} ^ {B / 2} \hat {h} _ {k j} ^ {(l)}}{\sqrt {\frac {2}{B} \sum_ {k = 1} ^ {B / 2} (\hat {h} _ {k j} ^ {(l)} - \hat {\mu} _ {j} ^ {(l)}) ^ {2}}} (5) \\ = \frac {\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \left(\boldsymbol {h} _ {i} ^ {(l - 1)} - \boldsymbol {\mu} ^ {(l - 1)}\right)}{\sqrt {\frac {2}{B} \sum_ {k = 1} ^ {B / 2} \left(\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \left(\boldsymbol {h} _ {i} ^ {(l - 1)} - \boldsymbol {\mu} ^ {(l - 1)}\right)\right) ^ {2}}}, (6) \\ \end{array} +$$ + +where we denote $h_{ij}^{(l)}, \hat{h}_{ij}^{(l)}$ and $\tilde{h}_{ij}^{(l)}$ as the output of ReLU, convolution or fully connected layer and BN layer respectively, for $i$ -th sample, $j$ -th output variable at $l$ -th layer. The gradient with respect to $\boldsymbol{w}_j^{(l)}$ is + +$$ +\begin{array}{l} \nabla_ {a} \boldsymbol {w} _ {j} ^ {(l)} = \frac {\partial \mathcal {L} _ {a} (\tilde {\boldsymbol {x}} _ {i})}{\partial \boldsymbol {w} _ {j} ^ {(l)}} = \frac {\nabla \tilde {h} _ {i j} ^ {(l)} \left(\boldsymbol {h} _ {i} ^ {(l - 1)} - \boldsymbol {\mu} ^ {(l - 1)}\right)}{\sqrt {\frac {2}{B} \sum_ {k = 1} ^ {B / 2} \left(\boldsymbol {w} _ {j} ^ {(l) T} \left(\boldsymbol {h} _ {i} ^ {(l - 1)} - \boldsymbol {\mu} ^ {(l - 1)}\right)\right) ^ {2}}} \\ - \frac {\nabla \tilde {h} _ {i j} ^ {(l)} \boldsymbol {w} _ {j} ^ {(l) T} (\boldsymbol {h} _ {i} ^ {(l - 1)} - \boldsymbol {\mu} ^ {(l - 1)})}{(\frac {2}{B} \sum_ {k = 1} ^ {B / 2} (\boldsymbol {w} _ {j} ^ {(l) T} (\boldsymbol {h} _ {i} ^ {(l - 1)} - \boldsymbol {\mu} ^ {(l - 1)})) ^ {2}) ^ {3 / 2}} \boldsymbol {\Sigma} ^ {(l - 1)} \boldsymbol {w} _ {j} ^ {(l)}, \\ \end{array} +$$ + +where $\pmb{\mu}^{(l-1)}$ and $\pmb{\Sigma}^{(l-1)}$ are the mean and covariance matrix from $(l-1)$ -th layer, $\nabla \tilde{h}_{ij}^{(l)}$ represents the gradient of loss with respect to $\tilde{h}_{ij}^{(l)}$ . If we write the weight as its norm $\| \pmb{w}_j^{(l)}\|$ multiplied by the unit vector $\pmb{u}_j^{(l)}$ , there is a relationship between $\nabla \pmb{w}_j^{(l)}$ and $\| \pmb{w}_j^{(l)}\|$ , + +$$ +\left\| \nabla_ {a} \boldsymbol {w} _ {j} ^ {(l)} \right\| = \frac {1}{\left\| \boldsymbol {w} _ {j} ^ {(l)} \right\|} \| \nabla_ {a} \boldsymbol {u} _ {j} ^ {(l)} \| \tag {7} +$$ + +This relationship has been found in [2, 27] and extended to any scale-invariant layers such as layer normalization [3] by [40]. It means that there are two ways to increase the gradient magnitude or the effective learning rate: 1) find a steeper descent direction where $\|\nabla_{a}\pmb{u}_{j}^{(l)}\|$ is increased and 2) decrease the weight norm so that $1 / \| \pmb{w}_j^{(l)}\|$ is increased. For a DNN with standard BN layers, the training will exploit this property to increase the gradient magnitude by reducing the weight norms with the help of weight decay regularization, which leads to a spurious increase in gradient magnitude and a larger variance of gradient estimation. In contrast, for a DNN with fixed BN layers such as the Frozen Net, the gradient norm is not correlated with weight norm, + +$$ +\nabla_ {f} \boldsymbol {w} _ {j} ^ {(l)} = \frac {\partial \mathcal {L} _ {f} (\tilde {\boldsymbol {x}} _ {i})}{\partial \boldsymbol {w} _ {j} ^ {(l)}} = \nabla \tilde {h} _ {i j} ^ {(l)} \frac {\boldsymbol {h} _ {i} ^ {(l - 1)}}{\sigma_ {j , p t} ^ {(l)}}. \tag {8} +$$ + +In this gradient, the only way to increase the gradient magnitude is to find the actual steeper direction. The overall gradient for the weight is + +$$ +\Delta \boldsymbol {w} _ {j} ^ {(l)} = \nabla_ {a} \boldsymbol {w} _ {j} ^ {(l)} + \lambda_ {t w i n s} \nabla_ {f} \boldsymbol {w} _ {j} ^ {(l)}. \tag {9} +$$ + +Empirical Study To see the difference between the gradients in AT and TWINS-AT, we record the gradient of all weight parameters in each step of the two training methods and compute the mean and STD of the gradient norm in each epoch. The model parameters and their gradients are treated as long vectors and we compute the $l_{2}$ norm of the weight and gradient vector. Figure 3 shows the mean and normalized STD of gradient norms in 60 training epochs for three datasets. Here we show the normalized STD, i.e., dividing the STD by the mean, to see the relative effect of variance. The major finding is that the gradient magnitude of TWINS-AT is substantially larger than that of AT, while the variance of TWINS-AT is lower than AT in most epochs. Note that in theory, the ratio between STD and mean should remain the same after down-scaling the weight norm in standard BN, but in practice we do observe the high normalized variance of DNNs with standard BNs since we use one epoch's gradients to approximate the variance and mean. + +One benefit of the larger gradient magnitude is that the model can escape from the initial sub-optimal point faster and find a better local optimum than the small gradient optimization [41]. We validate this hypothesis by recording the distance between the current model and the initial model during training. Figure 4 shows these weight distances on three datasets, where TWINS-AT moves away from the initial model much faster than the AT baseline. The robust overfitting effect of adversarial training [54] is partially a result of large gradient variance, especially at the final stage + +of training [9]. We compare the robust accuracy drop of AT and TWINS-AT in Table 1 and find that the small relative variance of TWINS-AT has the effect of reducing robust overfitting. See the experiment details in Sec. 5 and Supp. + +# 5. Experiment + +This section presents our experiment with TWINS. We first introduce our experiment setting and then show our main result and ablation study. + +# 5.1. Experiment Settings + +Dataset. We use five datasets in our experiment. CIFAR10 and CIFAR100 [34] are low-resolution image datasets, where the training and validation sets have 50,000 and 10,000 images, and CIFAR10 has 10 classes, while CIFAR100 has 100 classes. Caltech-256 [22] is a high-resolution dataset with 30,607 images and 257 classes, which is split into training and validation set with a ratio of 9:1. Caltech-UCSD Birds-200-2011 (CUB200) [60] is a high-resolution bird image dataset for fine-grained image classification, which contains 200 classes of birds, 5,994 training images and 5,794 validation images. Stanford Dogs [33] has high-resolution dog images from 120 dog categories, where the training and validation set has 12,000 and 8,580 images. For both low-resolution image datasets (CIFAR10 and CIFAR100) and high-resolution datasets (Caltech-256,CUB200 and Stanford Dogs), we resize the image to $224\times 224$ so that the input sizes are the same for pre-training and fine-tuning. As with pre-training, the input image is normalized by the mean and STD of the pre-training set. Note that the resizing and normalization function is integrated into the model so we can attack the input image with the [0,1] bounds for pixel values as in standard adversarial attacks. We use the standard ImageNet data augmentation for high-resolution datasets [26]. For CIFAR datasets, we use random cropping with padding=4 and random horizontal flipping. + +Adversarial Pre-Training. Large-scale adversarial pretraining on ImageNet is time-consuming, and thus [55] has released adversially pre-trained ResNet50 and WideResNet50-2, trained with $l_{2}$ and $l_{\infty}$ norm bounded attacks. In this paper, we adopt the pre-trained ResNet50 models, trained with $l_{\infty}$ attack with bound $\epsilon_{pt} = 4 / 255$ . We test other robust pre-trained models in our ablation study. + +Training Setting. For baselines and our method, we train all parameters of the pre-trained model, i.e., full fine-tuning instead of linear probing [57], with PGD attacks of $l_{\infty}$ norm. The PGD step is 10, $\epsilon_{ft} = 8 / 255$ and stepsize $\alpha = 2 / 255$ . We set the batch size as 128 and train the model for 60 epochs and divide the learning rate by 0.1 at 30th and 50th epoch. In TWINS with warmup, we initialize the means and STDs with their pre-trained values, and update the means and STDs using the target training set. + +
MetricMethodCIFAR10CIFAR100Caltech256CUB200Stanford Dogs
Clean Acc.AT89.7769.4875.9065.7460.09
TWINS-AT91.24(+1.47)70.72(+1.24)76.86(+0.96)68.09(+2.35)64.98(+4.89)
TWINS-AT+warmup91.95(+2.18)72.12(+2.64)77.35(+1.45)67.64(+1.90)66.12(+6.03)
TRADES87.0662.7669.7058.9259.99
TWINS-TRADES86.61(-0.45)66.72(+3.96)71.12(+1.42)60.72(+1.80)60.58(+0.59)
TWINS-TRADES+warmup86.60(-0.46)65.91(+3.15)73.39(+3.69)61.05(+2.13)63.96(+3.97)
PGD10AT52.2428.5248.3726.6019.80
TWINS-AT52.73(+0.49)31.08(+2.56)48.40(+0.03)29.24(+2.64)20.58(+0.78)
TWINS-AT+warmup52.46(+0.22)29.12(+0.60)49.13(+0.76)27.67(+1.07)19.48(-0.32)
TRADES54.0432.2047.2827.8721.36
TWINS-TRADES56.23(+2.19)33.51(+1.31)47.31(+0.03)27.05(-0.82)19.93(-1.43)
TWINS-TRADES+warmup55.81(+1.77)33.48(+1.28)48.53(+1.25)26.68(-1.19)19.46(-1.90)
AAAT48.4623.4743.8522.8212.30
TWINS-AT49.81(+1.35)26.73(+3.26)43.69(-0.16)22.33(-0.49)14.37(+2.07)
TWINS-AT+warmup49.02(+0.56)25.72(+2.25)43.92(+0.07)23.58(+0.75)13.80(+1.50)
TRADES50.3126.4043.3922.2112.05
TWINS-TRADES51.71(+1.40)28.29(+1.89)41.77(-1.62)22.68(+0.47)13.36(+1.31)
TWINS-TRADES+warmup51.10(+0.79)28.30(+1.90)43.55(+0.16)21.92(-0.29)10.94(-1.11)
+ +Table 2. The performance of our TWINS-AT and TWINS-TRADES on five image classification tasks compared with AT and TRADES. The clean accuracy means the accuracy when testing images are input without adversarial perturbations. PGD10 and AA denote the robust test accuracy under PGD10 and AutoAttack. The increase and decrease in performance are denoted with green and red numbers. The bold numbers denote the best robust accuracy under AA. The proposed TWINS achieves better robustness and clean accuracy compared with the baseline. Averaged over the datasets, the clean and robust accuracy of TWINS are increased by $2.18\%$ and $1.21\%$ compared with AT, and $1.46\%$ and $0.69\%$ compared with TRADES. The means and STDs of the performance are in the supplemental. + +
MetricPT ModelCaltech256CUB200Dogs
Clean AccRandom48.9912.387.27
Non-Robust64.6653.3041.99
Robust75.9065.7460.09
PGD10Random31.783.7283.59
Non-Robust39.8619.6813.32
Robust48.3726.6019.80
+ +The momentum of updating statistics is 0.1, the batch size is 128, and the warmup only lasts one epoch. Note that in the warmup stage, the input samples are added with adversarial perturbations generated by the PGD attack, which has the same setting as the attack in training, and the classifier layer is the pre-trained classifier for the adversarial attack. Our pilot experiment shows that using adversarial examples as input is more effective than using clean exemplars in the warmup. The optimizer is SGD with momentum in all of our experiments. The learning rate, weight decay and regularization hyperparameter are determined by grid search, which is described in detail in the supplemental. + +Adversarial robustness evaluation. Two standard adversarial attacks are used in our experiment, i.e., PGD and AutoAttack [11]. The adversarial perturbation is $l_{\infty}$ norm bounded in our evaluation. The setting of PGD attack for validation set is the same as the PGD attack in training. The AutoAttack (AA) is a more reliable adversarial attack and more often used for evaluation than PGD in recent years. + +Table 3. Comparison of random initialization, non-robust and robust pre-trained model on three difficult classification tasks, when fine-tuned with AT. The robust pre-trained model is indispensable to downstream robustness. + +
MethodMetricλWD
1e-51e-41e-31e-2
ATClean Acc.89.9289.9489.7790.28
Robust Acc.46.3444.6748.4647.57
TWINS-ATClean Acc.91.9091.4291.2487.33
Robust Acc.47.1949.0749.8151.65
+ +Table 4. The performance of TWINS-AT and AT on CIFAR10 when the hyperparameter for weight decay ${\lambda }_{WD}$ is changed. The robust accuracy is evaluated using AutoAttack. Our TWINS-AT achieves better adversarial robustness than AT for different ${\lambda }_{WD}$ . + +We use the standard attacks of AA, i.e., untargeted APGD-CE, targeted APGD-DLR, targeted FAB [10] and Square Attack [1], with $\epsilon = 8 / 255$ . The robust accuracy in our experiment result denotes the accuracy under AA. + +# 5.2. Experimental Result + +On the three high-resolution datasets, we compare the performance of fine-tuning different initialization models, i.e., random initialization, standard pre-trained ResNet50 and robust pre-trained ResNet50. Table 3 shows that the pre-trained models are essential to the accuracy and robustness in challenging downstream tasks, since the random initialization is much worse than the two pre-trained models. The robust pre-trained model has a clear benefit over the standard pre-trained one, indicating that robust pre-training is indispensable to downstream robustness. + +Table 2 shows the result of TWINS-AT and TWINS-TRADES compared with the baselines. Since AA is a more reliable attack, we highlight the best robust accuracy un + +![](images/ea33abe8f798ad25652a4208a8c1e235cbe4c6777310366e8782553eb682c466.jpg) +Figure 5. Ablation study where the BN statistics in TWINS are initialized with (0,1) for means and STDs, denoted as TWINS-Init. The population mean and STD of pre-training are crucial to TWINS. + +der AA on each dataset. The TWINS fine-tuning learns more robust DNNs, as well as achieves a better clean accuracy on all five datasets, demonstrating the strong effectiveness of TWINS in the robust pre-training and robust fine-tuning setting. On CIFAR10 and CIFAR100, both TWINS-AT and TWINS-TRADES achieve better robustness and clean accuracy than their baselines, while the warmup only improves the clean accuracy and hurts the robustness sometimes. On Caltech-256, TWINS-AT improves upon the vanilla AT in both robustness and accuracy, but TWINS-TRADES does not perform better than vanilla TRADES. However, the warmup helps boost the performance of TWINS-TRADES as well as TWINS-AT and makes the TWINS with warmup perform better than the baselines. + +On the two fine-grained image classification datasets, TWINS-AT and TWINS-TRADES generally perform better than baselines in terms of accuracy and robustness, if we look at the robust accuracy under AA, where only TWINS-AT on CUB200 has a slightly worse robust accuracy than its baseline. We find that the warmup improves the clean accuracy but hurts the robustness in most cases, except for CIFAR100 and Caltech-256. This can be a result of noisy adversarial perturbation, since we use the pre-trained classifier layer in the adversarial attack, or the insufficient update steps. Nevertheless, we note that the warmup can be considered as an operation for achieving a trade-off between robustness and accuracy. + +# 5.3. Ablation Study + +TWINS initialization. We use the pre-trained BN statistics in the Frozen Net to keep the robust information learned during pre-training. To show the importance of the pretrained BN statistics, we use the standard initialization (mean=0 and STD=1) for BN statistics in the Frozen Net, denoted as TWINS-Init, and show the result on the five datasets in Figure 5. Both clean and robust accuracy drop when the (0,1) initialization is used in TWINS-Init, demonstrating the crucial role of pre-trained statistics in TWINS. The fact that TWINS-Init sometimes improves upon the AT baseline motivates us to investigate the effect of TWINS on + +
MethodMetricεpt
1/2552/2554/2558/255
ATClean Acc.65.4767.0869.4869.93
Robust Acc.24.7625.7923.4727.71
TWINS-ATClean Acc.68.5570.4570.7272.59
Robust Acc.25.9726.6226.7328.62
+ +Table 5. The performance of TWINS-AT and AT on CIFAR100 when robust pre-trained ResNet50 with different $\epsilon_{pt}$ are used. + +gradient norms in Section 4. + +Effect of weight decay. Weight decay is the reason for the decreasing weight norm in DNNs with BN layers, so increasing the hyperparameter of weight decay $\lambda_{WD}$ is also a way to increase the gradient magnitude. Table 4 shows the result of TWINS-AT and AT when different $\lambda_{WD}$ are used when fine-tuning the robust pre-trained model on CIFAR10. The robust accuracy of TWINS-AT is consistently better than that of AT across different $\lambda_{WD}$ 's. We draw the same conclusion on CIFAR100 (see supplemental). Note that the clean accuracy of TWINS-AT drops when a large $\lambda_{WD}$ is used, suggesting that we should not use a too large $\lambda_{WD}$ for TWINS-AT. + +Different robust pre-trained models. The main experiment uses the robust pre-trained ResNet50 with $\epsilon_{pt} = 4/255$ as the initial model. We try different robust pre-trained models with different $\epsilon_{pt}$ in Table 5, which shows that a larger $\epsilon_{pt}$ is beneficial to both clean and robust accuracy in the downstream, and the proposed TWINS-AT is better than AT in both metrics with different pre-trained models. + +# 6. Conclusion + +This paper investigates the utility of robust pre-trained models in various downstream classification tasks. We first find that the commonly used data- and model-based approaches to maintain pre-training information do not work in the adversarial robust fine-tuning. We then propose a subtle statistics-based method, TWINS, for retaining the pretraining robustness in the downstream. In addition to the robustness preserving effect, we find that TWINS increases the gradient magnitude without sacrificing the training stability and improves the training dynamics of AT. Finally, the performance of TWINS is shown to be stronger than that of AT and TRADES on five datasets. One limitation of our work is that we only evaluate the robust supervised pre-trained ResNet50. Recently, robust pre-trained ViT's on ImageNet [4] have been released. Our statistics-based approach can be extended to the layer normalization, on which the increasing gradient magnitude argument also holds, and thus future work will extend TWINS to ViT. + +Acknowledgement This work was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11215820) and the Fundamental Research Funds for the Central University of China (DUT No. 82232031). + +# References + +[1] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 7 +[2] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. In International Conference on Learning Representations, 2019. 6 +[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6 +[4] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 34:26831-26843, 2021. 2, 8 +[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[7] Dian Chen, Hongxin Hu, Qian Wang, Li Yinli, Cong Wang, Chao Shen, and Qi Li. Cartl: Cooperative adversarially-robust transfer learning. In International Conference on Machine Learning, pages 1640–1650. PMLR, 2021. 2 +[8] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 699-708, 2020. 2, 3 +[9] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In International Conference on Learning Representations, 2020. 6 +[10] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In International Conference on Machine Learning, pages 2196-2205. PMLR, 2020. 7 +[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1, 7 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 + +[14] Gavin Weiguang Ding, Yash Sharma, Kry Yuk Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2019. 3 +[15] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, et al. On robustness and transferability of convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16458-16468, 2021. 2, 3 +[16] Xinshuai Dong, Anh Tuan Luu, Min Lin, Shuicheng Yan, and Hanwang Zhang. How should pre-trained language models be fine-tuned towards adversarial robustness? Advances in Neural Information Processing Systems, 34:4356-4369, 2021. 3 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2 +[18] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature communications, 12(1):1-14, 2021. 1 +[19] Samuel G Finlayson, John D Bowers, Joichi Ito, Jonathan L Zittrain, Andrew L Beam, and Isaac S Kohane. Adversarial attacks on medical machine learning. Science, 363(6433):1287-1289, 2019. 1 +[20] Jonathan Frankle, David J. Schwab, and Ari S. Morcos. Training batchnorm and only batchnorm: On the expressive power of random features in {cnn}s. In International Conference on Learning Representations, 2021. 4 +[21] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 1 +[22] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 1, 4, 6 +[23] Minghao Guo, Yuzhe Yang, Rui Xu, Ziwei Liu, and Dahua Lin. When nas meets robustness: In search of robust architectures against adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 631-640, 2020. 1 +[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2 +[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6 +[27] Elad Hoffer, Ron Banner, Itay Golan, and Daniel Soudry. Norm matters: efficient and accurate normalization schemes in deep networks. Advances in Neural Information Processing Systems, 31, 2018. 6 +[28] Hanxun Huang, Yisen Wang, Sarah Erfani, Quanquan Gu, James Bailey, and Xingjun Ma. Exploring architectural ingredients of adversarially robust deep neural networks. Advances in Neural Information Processing Systems, 34:5545-5559, 2021. 1 +[29] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 2 +[30] Daniel Jakubovitz and Raja Giryes. Improving dnn robustness to adversarial attacks using jacobian regularization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 514-529, 2018. 3 +[31] Ahmadreza Jeddi, Mohammad Javad Shafiee, and Alexander Wong. A simple fine-tuning is all you need: Towards robust deep learning via adversarial fine-tuning. arXiv preprint arXiv:2012.13628, 2020. 2, 3 +[32] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 33:16199-16210, 2020. 2, 3 +[33] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 6 +[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 4, 6 +[35] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2 +[36] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2 +[37] Yao Li, Martin Renqiang Min, Thomas Lee, Wenchao Yu, Erik Kruus, Wei Wang, and Cho-Jui Hsieh. Towards robustness of deep neural networks via regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7496-7505, 2021. 1 +[38] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779, 2016. 4, 5 +[39] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 3, 4 +[40] Zhiyuan Li, Kaifeng Lyu, and Sanjeev Arora. Reconciling modern deep learning with traditional optimization analyses: The intrinsic learning rate. Advances in Neural Information Processing Systems, 33:14544-14555, 2020. 6 + +[41] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. Advances in Neural Information Processing Systems, 33:21476-21487, 2020. 2, 6 +[42] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128(2):261-318, 2020. 2 +[43] Ziquan Liu and Antoni B Chan. Boosting adversarial robustness from the perspective of effective margin regularization. British Machine Vision Conference (BMVC), 2022. 1, 3 +[44] Ziquan Liu, Yufei Cui, and Antoni B Chan. Improve generalization and robustness of neural networks via weight scale shifting invariant regularizations. ICML Workshop on Adversarial Machine Learning, 2021. 1 +[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[46] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 4 +[47] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, Xiangyang Ji, and Antoni B Chan. An empirical study on distribution shift robustness from the perspective of pre-training and data augmentation. arXiv preprint arXiv:2205.12753, 2022. 2 +[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3 +[49] Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 2 +[50] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 4 +[51] Chongli Qin, James Martens, Sven Gowal, Dilip Krishnan, Krishnamurthy Dvijotham, Alhussein Fawzi, Soham De, Robert Stanforth, and Pushmeet Kohli. Adversarial robustness through local linearization. Advances in Neural Information Processing Systems, 32, 2019. 1, 3 +[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 2 +[53] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Data augmentation can improve robustness. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1 + +[54] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093–8104. PMLR, 2020. 6 +[55] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 1, 2, 3, 6 +[56] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2021. 1, 3 +[57] Ali Shafahi, Parsa Saadatpanah, Chen Zhu, Amin Ghiasi, Christoph Studer, David Jacobs, and Tom Goldstein. Adversarially robust transfer learning. In International Conference on Learning Representations, 2020. 2, 3, 4, 6 +[58] Chawin Sitawarin, Arvind Sridhar, and David Wagner. Improving the accuracy-robustness trade-off for dual-domain adversarial training. ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning. 3 +[59] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 1 +[60] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6 +[61] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 3 +[62] Yutaro Yamada and Mayu Otani. Does robustness on image net transfer to downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9215-9224, 2022. 2, 3 +[63] Yaodong Yu, Heinrich Jiang, Dara Bahri, Hossein Mobahi, Seungyeon Kim, Ankit Singh Rawat, Andreas Veit, and Yi Ma. An empirical study of pre-trained vision models on out-of-distribution generalization. In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021. 2 +[64] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3 \ No newline at end of file diff --git a/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/images.zip b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0fdc4a373ac1536c7492219b8af92c2ee0001d78 --- /dev/null +++ b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bffc57c974c046f2a4d515cd3c22dec02428cb4c5558bf55fcdbb148164b4f4 +size 494521 diff --git a/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/layout.json b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dce2d26a236371c08796809e0068fd0fc4ff0c39 --- /dev/null +++ b/2023/TWINS_ A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization/layout.json @@ -0,0 +1,8120 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 545, + 138 + ], + "type": "text", + "content": "TWINS: A Fine-Tuning Framework for Improved Transferability of Adversarial Robustness and Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "spans": [ + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "text", + "content": "Ziquan Liu" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Yi Xu" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Xiangyang Ji" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "text", + "content": " and Antoni B. Chan" + }, + { + "bbox": [ + 154, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 175, + 456, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 175, + 456, + 190 + ], + "spans": [ + { + "bbox": [ + 137, + 175, + 456, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 175, + 456, + 190 + ], + "type": "text", + "content": "Department of Computer Science, City University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 190, + 457, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 190, + 457, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 457, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 190, + 457, + 204 + ], + "type": "text", + "content": "School of Artificial Intelligence, Dalian University of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 204, + 416, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 204, + 416, + 217 + ], + "spans": [ + { + "bbox": [ + 176, + 204, + 416, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 176, + 204, + 416, + 217 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 220, + 517, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 220, + 517, + 232 + ], + "spans": [ + { + "bbox": [ + 72, + 220, + 517, + 232 + ], + "type": "text", + "content": "ziquanliu.cs@gmail.com, yxu@dlut.edu.cn, xyji@tsinghua.edu.cn, abchan@cityu.edu.hk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 284, + 290, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 631 + ], + "type": "text", + "content": "Recent years have seen the ever-increasing importance of pre-trained models and their downstream training in deep learning research and applications. At the same time, the defense for adversarial examples has been mainly investigated in the context of training from random initialization on simple classification tasks. To better exploit the potential of pre-trained models in adversarial robustness, this paper focuses on the fine-tuning of an adversially pre-trained model in various classification tasks. Existing research has shown that since the robust pre-trained model has already learned a robust feature extractor, the crucial question is how to maintain the robustness in the pre-trained model when learning the downstream task. We study the model-based and data-based approaches for this goal and find that the two common approaches cannot achieve the objective of improving both generalization and adversarial robustness. Thus, we propose a novel statistics-based approach, TwoWing NormliSation (TWINS) fine-tuning framework, which consists of two neural networks where one of them keeps the population means and variances of pre-training data in the batch normalization layers. Besides the robust information transfer, TWINS increases the effective learning rate without hurting the training stability since the relationship between a weight norm and its gradient norm in standard batch normalization layer is broken, resulting in a faster escape from the sub-optimal initialization and alleviating the robust overfitting. Finally, TWINS is shown to be effective on a wide range of image classification datasets in terms of both generalization and robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "The adversarial vulnerability of deep neural networks (DNNs) [59] is one of the major obstacles for their wide" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 261, + 432, + 370 + ], + "blocks": [ + { + "bbox": [ + 310, + 261, + 432, + 370 + ], + "lines": [ + { + "bbox": [ + 310, + 261, + 432, + 370 + ], + "spans": [ + { + "bbox": [ + 310, + 261, + 432, + 370 + ], + "type": "image", + "image_path": "0da221f25917e7ef65d4a2563fe4e0898588b55aa9551d9aadb17eef084a780f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "type": "text", + "content": "Figure 1. The performance of fine-tuning robust and non-robust large-scale pre-trained (PT) ResNet50 [26, 55] on CIFAR10 [34] and Caltech-256 [22]. We compare standard adversarial training (AT), Learning without Forgetting (LwF) (model approach) [39], joint fine-tuning with UOT data selection (data approach) [46] and our TWINS fine-tuning. The robust accuracy is evaluated using " + }, + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "type": "text", + "content": " norm bounded AutoAttack [11] with " + }, + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 304, + 374, + 547, + 538 + ], + "type": "text", + "content": ". On CIFAR10, the data-based and model-based approach fail to improve clean and robust accuracy. On Caltech, both approaches improve the clean accuracy but hurt the robust accuracy. Our TWINS fine-tuning improves the clean and robust performance on both datasets. The pink triangle denotes the performance of standard AT with the non-robust pre-trained ResNet50, which drops considerably compared with fine-tuning starting from the robust pretrained model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 261, + 545, + 370 + ], + "blocks": [ + { + "bbox": [ + 432, + 261, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 432, + 261, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 432, + 261, + 545, + 370 + ], + "type": "image", + "image_path": "d68393543ae72d019748e33d6fb39257d9b60d2a2d536d365a5f3644c161b2b0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "applications in safety-critical scenarios such as self-driving cars [18] and medical diagnosis [19]. Thus, addressing this issue has been one focus of deep learning research in the past eight years. Existing works have proposed to improve adversarial robustness from different perspectives, including data augmentation [21, 48, 53, 56], regularization [37, 43, 44, 51] and neural architecture [23, 28]. However, most of existing works investigate the problem under the assumption that the training data is sufficient enough, and training from scratch gives a satisfactory performance, which is not realistic in the real world. There are a large number of computer vision tasks where training from scratch is inferior to training from pre-trained weights, such as fine-grained image classification (e.g., Caltech-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16436" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 92, + 333, + 217 + ], + "blocks": [ + { + "bbox": [ + 139, + 76, + 238, + 89 + ], + "lines": [ + { + "bbox": [ + 139, + 76, + 238, + 89 + ], + "spans": [ + { + "bbox": [ + 139, + 76, + 238, + 89 + ], + "type": "text", + "content": "(a) TWINS Structure" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 92, + 333, + 217 + ], + "lines": [ + { + "bbox": [ + 51, + 92, + 333, + 217 + ], + "spans": [ + { + "bbox": [ + 51, + 92, + 333, + 217 + ], + "type": "image", + "image_path": "98f62d62bc6ae35d3945d6926361d03650ee38345992915c930d1c0be0a55fb3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 547, + 285 + ], + "type": "text", + "content": "Figure 2. The TWINS structure and training pipeline. (a) The Frozen Net and Adaptive Net have the same structure and share the weight parameters, except for batch normalization (BN) layers. The Frozen Net uses pre-trained means and standard deviations (STD) in the normalization layer, while Adaptive Net uses the mean and STD computed from the current batch as in standard BN. (b) In each step of mini-batch stochastic gradient descent (SGD), we split the batch of adversarial examples, generated from attacking the Adaptive Net, into two sub-batches and feed them to the Adaptive Net and Frozen Net respectively. The loss of two networks are combined and backpropagated to their shared parameters to train the network. In the inference stage, only the Adaptive Net is used." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 342, + 95, + 542, + 211 + ], + "blocks": [ + { + "bbox": [ + 392, + 76, + 488, + 89 + ], + "lines": [ + { + "bbox": [ + 392, + 76, + 488, + 89 + ], + "spans": [ + { + "bbox": [ + 392, + 76, + 488, + 89 + ], + "type": "text", + "content": "(b) Training Pipeline" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 95, + 542, + 211 + ], + "lines": [ + { + "bbox": [ + 342, + 95, + 542, + 211 + ], + "spans": [ + { + "bbox": [ + 342, + 95, + 542, + 211 + ], + "type": "image", + "image_path": "64cb32692987443b8900d6c4735c18d6bd1bb99180d034d6bad575cb6a6a6670.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 290, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 315 + ], + "type": "text", + "content": "UCSD Birds-200-2011 or CUB200 [60]), object detection [42] and semantic segmentation [49]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 316, + 288, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 531 + ], + "type": "text", + "content": "On the other hand, pre-trained models have been considered as the foundation models in deep learning [5] as a result of their strong performance and wide employment in computer vision [17, 24, 25, 45], as well as natural language processing [6, 13, 52]. Thus, how to better use the pre-trained model in downstream has emerged as a major research topic in many vision and language tasks, such as image classification under distribution shifts [47, 63], object detection [36] and semantic segmentation [29, 35]. There are a few papers that investigate the pre-trained model's robustness in target tasks [7, 8, 15, 31, 32, 57, 62]. [7, 57] mainly considers the transfer between small-scale datasets (e.g., CIFAR100 to CIFAR10), while [8, 32] use adversarial robust pre-training and fine-tuning on the same dataset, without considering a large-scale and general pre-trained model. Finally, [15, 62] investigate different kinds of robustness to corruption or out-of-distribution samples, and are not devoted to adversarial robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we consider how to transfer the adversarial robustness of a large-scale robust pre-trained model (e.g., a ResNet50 pre-trained on ImageNet [12] with adversarial training) on various downstream classification tasks when fine-tuning with adversarial training. This problem setting is becoming more important as the standard pretrained models do not learn robust representations from the pre-training data and are substantially weaker than the robust pre-trained counterparts in some challenging downstream tasks, e.g., fine-grained classification as shown in our experiment. Meanwhile, more large-scale robust pretrained models are released (e.g., ResNet [55] and ViT [4]), which makes the robust pre-trained models more accessible. However, naively applying adversarial training to fine-tune from the robustly pre-trained model will lead to subopti" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 290, + 546, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 350 + ], + "type": "text", + "content": "mal robustness, since the robust representations learned by the robust pre-trained model are not fully utilized. For example, [57] suggests that the robustness from a pre-trained model needs to be explicitly maintained for its better transfer to the downstream." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 713 + ], + "type": "text", + "content": "Following the idea that the key to improving the transferability of robustness is to maintain the robustness of the pre-training stage during fine-tuning [57], we first evaluate the data-based and model-based approach on two representative datasets, CIFAR10 and Caltech-256. The database-based approach uses pre-trained data in the fine-tuning and keeps their performance under adversarial attack, while the model-based approach regularizes the distance of features of the fine-tuned and pre-trained model. Our experiment shows that both methods fail to improve the robustness and generalization (Fig. 1), since the two methods are too aggressive in retaining the robustness and hurt the learning in downstream. Thus, we propose a subtle approach that keeps the batch-norm (BN) statistics of pre-training for preserving the robustness, which we call Two-Wing NormaliSation (TWINS) fine-tuning. TWINS has two neural networks with fixed and adaptive BN layers respectively, where the fixed BN layers use the population means and STDs of pretraining for normalization, while the adaptive BN layers use the standard BN normalization. Our experiment first demonstrates the importance of pre-trained BN statistics in the robust fine-tuning and then finds the benefit of TWINS in adversarial training dynamics. As the relationship between weight norm and its gradient norm no longer holds in TWINS, it is able to increase the gradient magnitude without increasing the gradient variance. At the initial training stage, TWINS has a faster escaping speed from the suboptimal initialization than vanilla adversarial training [41]. At the final training stage, the gradient of TWINS is more stable than adversarial training, which alleviates the robust" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16437" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "overfitting effect [56]. In summary, the contributions of our paper are as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 287, + 419 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 96, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 287, + 167 + ], + "type": "text", + "content": "1. We focus on the fine-tuning of large-scale robust pretrained models as a result of their potential importance in various downstream tasks. We evaluate current approaches to retain the pre-training robustness in fine-tuning, and show that they cannot substantially improve the robustness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 168, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 168, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 168, + 287, + 300 + ], + "type": "text", + "content": "2. We propose TWINS, a statistics-based approach for better transferability of robustness and generalization from the pre-training domain to the target domain. TWINS has two benefits: a) it keeps the robust statistics for downstream tasks, thus helps the transfer the robustness to downstream tasks and b) it enlarges the gradient magnitude without increasing gradient variance, thus helps the model escape from the initialization faster and mitigates robust overfitting. The mechanisms of these two benefits are validated by our empirical study." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "text", + "content": "3. The effectiveness of TWINS is corroborated on five downstream datasets by comparing with two popular adversarial training baselines, adversarial training (AT) [48] and TRADES [64]. On average, TWINS improves the clean and robust accuracy by " + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "inline_equation", + "content": "2.18\\%" + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "text", + "content": " compared with AT, and by " + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "inline_equation", + "content": "1.46\\%" + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "inline_equation", + "content": "0.69\\%" + }, + { + "bbox": [ + 54, + 300, + 287, + 419 + ], + "type": "text", + "content": " compared with TRADES. The experiment shows the strong potential of robust pre-trained models in boosting downstream's robustness and generalization when using more effective fine-tuning methods." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 430, + 134, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 134, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 134, + 442 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 450, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 642 + ], + "type": "text", + "content": "Adversarial defense. There are several major approaches to improving the adversarial robustness of DNNs. The training of DNNs can be regularized to induce biases that are beneficial to adversarial robustness, such as locally linear regularization [51], margin maximization [14, 43] and Jacobian regularization [30]. The most commonly used adversarial defense is adversarial training (AT) [48], which directly trains the DNN on adversarial examples generated from PGD attack. Later, TRADES [64] is proposed to add a KL regularization to AT and achieves stronger adversarial robustness. Our paper proposes TWINS to improve adversarial training in the fine-tuning stage when the initial model is adversarially pre-trained. We compare TWINS-AT and TWINS-TRADES with vanilla AT and TRADES in our experiment and show the strong effectiveness of TWINS in the robust fine-tuning setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Fine-tuning for downstream robustness. Several aspects of robustness in pre-training and fine-tuning have been studied in existing works. Adversarial contrastive learning [8, 32] is proposed to pre-train on a dataset with contrastive learning and then fine-tune on the same dataset, without considering the transferability of robustness from a large-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "scale pre-trained model to a different downstream task. In contrast, our paper investigates a more general problem, where task-specific pre-training is not needed for a new task as we use one robust large-scale pre-trained model trained on ImageNet. [55] considers the robust pre-training on the large-scale ImageNet and its transfer to downstream tasks, but focuses on the performance on clean instead of adversarial images. The Learning-without-Forgotten (LwF) [39] approach for retaining robustness is shown to be effective in the small-scale transfer experiment [57], but is not effective in our experiment setting of transfer of large-scale models. [31] proposes a learning rate schedule to improve the adversarial robustness of fine-tuned models, and [16] proposes robust informative fine-tuning for pre-trained language models to robustly keep pre-training information in downstream. The difference between [16, 31] and our work is that they assume a standard pre-trained model instead of the adversarial pre-trained model. [15, 62] investigate the performance of pre-trained models in downstream tasks, but the focus is the robustness to out-of-distribution samples instead of adversarial perturbations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 325, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 494 + ], + "type": "text", + "content": "Batch normalization. There are existing papers proposing the two-branch BN structure for different purposes with different technical details. [58] proposes dual normalization for a better trade-off between accuracy and robustness, where the normalization is a weighted sum of normalized clean and adversarial input. [61] proposes a similar two-branch BN structure, where one branch is for adversarial examples and the other is for clean examples. The major difference between our work and [58, 61] is that both BN branches in TWINS are for adversarial examples and one branch (Frozen Net) has fixed BN statistics from pretraining so as to better maintain the pre-trained robustness, whereas [58, 61] uses clean examples in BN and aims to improve the accuracy for clean images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 510, + 545, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 537 + ], + "type": "text", + "content": "3. The Model-based and Data-based Approach to Retaining Adversarial Robustness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "This section introduces the two common approaches for keeping adversarial robustness of pre-training in downstream, model-based and data-based approaches. Denote the feature vector output of a neural network as " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\theta}(\\boldsymbol{x})" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", the training sample in the downstream task as " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{(x_i,y_i)\\}_{i=1}^N \\sim \\mathbb{P}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", and the loss function as " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\boldsymbol{w}^T g_{\\theta}(\\boldsymbol{x}) + b, \\boldsymbol{y})" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w}, b)" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " are the parameters of last classification layer. We assume that the pre-trained model is trained on adversarial examples generated from the PGD attack [48], where the " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " norm of the adversarial attack is bounded by " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", and during fine-tuning we use adversarial training with the same PGD attack to obtain adversarial robustness in downstream tasks. In short, we consider the robust pre-training and robust fine-tuning setting in this paper, if not specified otherwise." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16438" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 115 + ], + "type": "text", + "content": "Model-based approaches. We first introduce the model-based approach, which keeps the pre-trained model " + }, + { + "bbox": [ + 47, + 72, + 287, + 115 + ], + "type": "inline_equation", + "content": "\\theta_{pt}" + }, + { + "bbox": [ + 47, + 72, + 287, + 115 + ], + "type": "text", + "content": " during fine-tuning so as to maintain its robustness. The objective function is" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 294, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 294, + 155 + ], + "type": "interline_equation", + "content": "\\sum_ {\\left(\\boldsymbol {x} _ {n}, \\boldsymbol {y} _ {n}\\right) \\sim \\mathbb {P}} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) + b, \\boldsymbol {y} _ {n}\\right) + \\lambda_ {L w F} \\| g _ {\\boldsymbol {\\theta} _ {p t}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) - g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) \\| _ {2}, \\tag {1}", + "image_path": "f7fff46e83179a6ebd766d06a47fd6032c0098a67bd6eaf6828ac5f6e37f1cec.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "type": "text", + "content": "where the adversarial example generated from " + }, + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_n" + }, + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}}_n" + }, + { + "bbox": [ + 47, + 159, + 287, + 278 + ], + "type": "text", + "content": ". The regularization term of the loss aims to minimize the distance between the features from the pretrained and the fine-tuned models, which is expected to maintain the robustness of the pre-trained model. This approach is originally proposed in [39] to prevent the catastrophic forgetting in continual learning, and is used in [57] to preserve adversarial robustness in transfer learning. Note that [57] uses the LwF method in standard fine-tuning instead of robust fine-tuning as in our paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 292, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 315 + ], + "type": "text", + "content": "Data-based approaches. The objective function of the data-based approach is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 318, + 287, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 287, + 370 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {\\left(\\boldsymbol {x} _ {n}, \\boldsymbol {y} _ {n}\\right) \\sim \\mathbb {P}} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {n}\\right) + b, \\boldsymbol {y} _ {n}\\right) \\\\ + \\lambda_ {U O T} \\sum_ {\\left(\\boldsymbol {x} _ {m}, \\boldsymbol {y} _ {m}\\right) \\sim \\mathbb {Q}} \\mathcal {L} \\left(\\boldsymbol {w} _ {q} ^ {T} g _ {\\boldsymbol {\\theta}} \\left(\\tilde {\\boldsymbol {x}} _ {m}\\right) + b _ {q}, \\boldsymbol {y} _ {m}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "eeb0bba61f5e426614577d3da241ae5e7d6f2592e331457eec2b70468f555d1f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}" + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "text", + "content": " are data distribution of the target and the pre-training tasks, " + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "inline_equation", + "content": "w_{q}, b_{q}" + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "text", + "content": " are the classification layer for pre-trained data from " + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}" + }, + { + "bbox": [ + 47, + 373, + 287, + 467 + ], + "type": "text", + "content": ". This method regularizes the current fine-tuned model feature extractor so that its prediction is still robust on the pre-training data. The joint training method is proposed in [46] to improve the performance of fine-tuning in downstream tasks where the training data is not sufficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 469, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 624 + ], + "type": "text", + "content": "Next, we test the performance of these two approaches on two standard image classification datasets, CIFAR10 [34] and Caltech-256 [22], with results shown in Figure 1. We use a grid search for learning rate and " + }, + { + "bbox": [ + 47, + 469, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\lambda_{LwF}(\\lambda_{UOT})" + }, + { + "bbox": [ + 47, + 469, + 287, + 624 + ], + "type": "text", + "content": " and report the result of the model with the best robust accuracy. See Section 5 and the supplemental material for the experiment setting. On CIFAR10, both approaches fail to improve either clean or robust accuracy; on Caltech-256, the two approaches improve the clean accuracy by a small margin but deteriorate the robustness. One reason why the model- and data-based approaches fail is that the regularization term might be too strong, thus hurting the learning in the downstream." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 168, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 168, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 168, + 647 + ], + "type": "text", + "content": "4. TWINS Fine-Tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": "The previous section demonstrates that both data- and model-based approaches cannot substantially improve the adversarial robustness in the downstream task. Thus we propose the TWINS for the better fine-tuning of robust pretrained models for downstream adversarial robustness." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 544, + 319 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 544, + 319 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 544, + 319 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 544, + 319 + ], + "type": "image", + "image_path": "cea56a67dcb0f72a2574e945b38f3b99df3eca33da144cb0c5c4831967c10967.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "type": "text", + "content": "Figure 3. The mean and normalized STD of gradient norms in AT and TWINS-AT on CIFAR10, CIFAR100 and Caltech-256. The averaged " + }, + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "type": "inline_equation", + "content": "\\mu (\\| \\nabla \\pmb {w}\\|)" + }, + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "type": "inline_equation", + "content": "\\sigma (\\| \\nabla \\pmb {w}\\|) / \\mu (\\| \\nabla \\pmb {w}\\|)" + }, + { + "bbox": [ + 305, + 319, + 545, + 428 + ], + "type": "text", + "content": " over epochs are shown in each plot. The gradient magnitudes of TWINS-AT are substantially larger than those of AT, while the Normalized STDs of gradient norm in TWINS-AT are not obviously increased (CIFAR10) or even decreased (CIFAR100 and Caltech-256) compared with AT. This property leads to the faster escaping speed of TWINS-AT from the initial sub-optimum compared with AT (Fig. 4), and reduced robust overfitting (Tab. 1)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 430, + 411, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 411, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 411, + 442 + ], + "type": "text", + "content": "4.1. Proposed Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 581 + ], + "type": "text", + "content": "Though BN layers contain only a few parameters compared to convolution and fully connected layers, they play an important role in the good performance of DNNs. [20] shows that even if we only train the BN layers, the performance of a DNN is already quite impressive. [50] finds that only training the parameters in BN layers in an image generator is effective for small datasets. [38] proposes adaptive BN for domain adaptation, which updates the BN statistics with data from a target domain. These works motivate us to propose a statistics-based approach for retaining pretraining information in the target task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "Typical BN layers track the mean and STD of the training set and save them for the inference stage. As this distribution information for each layer might be helpful for downstream robustness, we propose the TWINS robust finetuning, which maintains two networks, Frozen Net that uses the BN statistics from the robust pre-trained model, and Adaptive Net that learns its BN statistics from the downstream task. Instead of using two independent networks for the Frozen and Adaptive Net, we let the two networks share weight parameters, excluding the BN layers, to save the model size and inference time. At initialization, both" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16439" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 288, + 141 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 288, + 141 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 288, + 141 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 288, + 141 + ], + "type": "image", + "image_path": "26134a5960d82f6e7b6c7c25ab0464e79a784705b0138551bc6674497cd7727a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "type": "text", + "content": "Figure 4. The distance between the current step's weight " + }, + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "type": "text", + "content": " and the initialization " + }, + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\theta_{pt}" + }, + { + "bbox": [ + 46, + 143, + 287, + 187 + ], + "type": "text", + "content": ". On the three datasets, TWINS-AT has a faster escaping speed from the sub-optimal initial model than AT, which is due to TWIN-AT's larger gradient norm as shown in Figure 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 198, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 292 + ], + "type": "text", + "content": "networks and their BN statistics are initialized by the robust pre-trained models. During training, the Frozen Net uses the population means and STDs of pre-training data computed in the pre-training stage in the normalization operation, while the Adaptive Net uses the current batch's mean and STD in the normalization and updates its running mean and STD with the target training data. Fig. 2 shows the general pipeline of TWINS training and the network structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 293, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 315 + ], + "type": "text", + "content": "The training objective of TWINS with adversarial training (TWINS-AT) in one mini-batch is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 320, + 287, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 320, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 85, + 320, + 287, + 350 + ], + "type": "interline_equation", + "content": "\\sum_ {j = 1} ^ {B / 2} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta} _ {a}} \\left(\\tilde {\\boldsymbol {x}} _ {j} ^ {(a)}\\right) + b, \\boldsymbol {y} _ {j}\\right) + \\tag {3}", + "image_path": "d69d79f58fd0fe36effc47c43d09fdeadaac4e6a8a1905f7171f645cfc923bf3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 354, + 287, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 354, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 93, + 354, + 287, + 384 + ], + "type": "interline_equation", + "content": "\\lambda_ {t w i n s} \\sum_ {i = B / 2 + 1} ^ {B} \\mathcal {L} \\left(\\boldsymbol {w} ^ {T} g _ {\\boldsymbol {\\theta} _ {f}} \\left(\\tilde {\\boldsymbol {x}} _ {i} ^ {(a)}\\right) + b, \\boldsymbol {y} _ {i}\\right), \\tag {4}", + "image_path": "888bba52b971fa6edeb6a6868f094e566afe1365ee71ca847b3119950f46df3d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\theta_{a}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " denote the Frozen and Adaptive Nets respectively, and " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}^{(a)}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " is the adversarial image for " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": " when attacking the Adaptive Net " + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\theta_{a}" + }, + { + "bbox": [ + 46, + 389, + 287, + 568 + ], + "type": "text", + "content": ". We split the batch into two different sub-batches to avoid doubled batch sizes in TWINS training. Since the Frozen and Adaptive Nets share weight parameters, the number of parameters in TWINS is only increased by a very small amount (i.e., BN parameters). Thus, TWINS-AT only has a negligible cost in terms of memory and training time compared with vanilla AT. The TWINS structure can also be used in the TRADES in a similar way. Similar to [38], we can use the target training set to update the BN statistics so that they are more relevant to the downstream task. We call this procedure warmup in TWINS fine-tuning. The pseudo codes of TWINS-AT and TWINS-TRADES are given in Alg. 1 of the supplemental." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 195, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 195, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 195, + 586 + ], + "type": "text", + "content": "4.2. The mechanism of TWINS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "The first benefit of TWINS is mentioned in the motivation of TWINS, i.e., the BN statistics have robustness information in the pre-training domain that can be leveraged by robust fine-tuning to improve the downstream robustness. This argument is validated by our ablation study in Section 5, where we initialize the means and STDs with 1.0 and 0.0 for Frozen Net instead of the pre-trained means and STDs and check the accuracy and robustness. Figure 5 shows that the TWINS with (1,0) initialization cannot match the performance with TWINS with pre-trained" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 541, + 129 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 541, + 129 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 541, + 129 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 541, + 129 + ], + "type": "table", + "html": "
MethodRob. Acc.C10C100CaltechCUBDogs
ATBest ↑51.8431.3849.0927.0821.19
Final ↑49.4128.5248.3726.6019.80
Gap ↓2.432.860.730.481.39
TWINS-ATBest ↑53.2331.6048.8029.2420.89
Final ↑52.4031.0848.4029.2420.58
Gap ↓0.830.520.400.000.29
", + "image_path": "010c98536299be578cc7a9d114a6bf2f2926e91104fb102e7386d21d88f1eef3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 130, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 545, + 185 + ], + "type": "text", + "content": "Table 1. The robust accuracy drop of AT and TWINS-AT, where the adversarial attack is PGD10. TWINS-AT has smaller accuracy drop compared with AT, indicating that the TWINS-AT is less prone to robust overfitting as a result of reduced variance of gradient norms and stable training as shown in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 195, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 195, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 545, + 219 + ], + "type": "text", + "content": "statistics, indicating that the robustness information in pretraining is essential to the effectiveness of TWINS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 220, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 327 + ], + "type": "text", + "content": "It is intriguing that even with the (1,0) initialization, TWINS still outperforms the AT baseline in terms of robustness or accuracy on some datasets, e.g., Stanford Dogs and CIFAR100. This suggests that besides retaining the pre-training information, TWINS provides some other benefits during robust fine-tuning. By analyzing the gradient of TWINS, we find that TWINS implicitly increases the effective learning rate without increasing the oscillation and empirically validate this finding. We detail this analysis next." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "type": "text", + "content": "Effective learning rate We first write the gradient of a weight vector for TWINS training. Consider the " + }, + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "type": "text", + "content": "-th layer's weight " + }, + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_j^{(l)}" + }, + { + "bbox": [ + 305, + 327, + 545, + 365 + ], + "type": "text", + "content": " and its output after BN layer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 374, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 374, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 329, + 374, + 545, + 455 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {h} _ {i j} ^ {(l)} = \\frac {\\hat {h} _ {i j} ^ {(l)} - \\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\hat {h} _ {k j} ^ {(l)}}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} (\\hat {h} _ {k j} ^ {(l)} - \\hat {\\mu} _ {j} ^ {(l)}) ^ {2}}} (5) \\\\ = \\frac {\\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\left(\\boldsymbol {w} _ {j} ^ {(l) ^ {T}} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)\\right) ^ {2}}}, (6) \\\\ \\end{array}", + "image_path": "3c7cf580fb6b2ef084f14ecdeb661805d8b9f8814c55b1032ca2faa33086d2e5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": "where we denote " + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "inline_equation", + "content": "h_{ij}^{(l)}, \\hat{h}_{ij}^{(l)}" + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\tilde{h}_{ij}^{(l)}" + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": " as the output of ReLU, convolution or fully connected layer and BN layer respectively, for " + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": "-th sample, " + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": "-th output variable at " + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": "-th layer. The gradient with respect to " + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}_j^{(l)}" + }, + { + "bbox": [ + 305, + 464, + 545, + 515 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 522, + 548, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 548, + 593 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 548, + 593 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} = \\frac {\\partial \\mathcal {L} _ {a} (\\tilde {\\boldsymbol {x}} _ {i})}{\\partial \\boldsymbol {w} _ {j} ^ {(l)}} = \\frac {\\nabla \\tilde {h} _ {i j} ^ {(l)} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)}{\\sqrt {\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} \\left(\\boldsymbol {w} _ {j} ^ {(l) T} \\left(\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)}\\right)\\right) ^ {2}}} \\\\ - \\frac {\\nabla \\tilde {h} _ {i j} ^ {(l)} \\boldsymbol {w} _ {j} ^ {(l) T} (\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)})}{(\\frac {2}{B} \\sum_ {k = 1} ^ {B / 2} (\\boldsymbol {w} _ {j} ^ {(l) T} (\\boldsymbol {h} _ {i} ^ {(l - 1)} - \\boldsymbol {\\mu} ^ {(l - 1)})) ^ {2}) ^ {3 / 2}} \\boldsymbol {\\Sigma} ^ {(l - 1)} \\boldsymbol {w} _ {j} ^ {(l)}, \\\\ \\end{array}", + "image_path": "d2ae08898e16c0733f578ad5df282402fcfae721afc370d485577417643d5eec.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(l-1)}" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{(l-1)}" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": " are the mean and covariance matrix from " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "(l-1)" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\nabla \\tilde{h}_{ij}^{(l)}" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": " represents the gradient of loss with respect to " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{h}_{ij}^{(l)}" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": ". If we write the weight as its norm " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\| \\pmb{w}_j^{(l)}\\|" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": " multiplied by the unit vector " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{u}_j^{(l)}" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": ", there is a relationship between " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\nabla \\pmb{w}_j^{(l)}" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\| \\pmb{w}_j^{(l)}\\|" + }, + { + "bbox": [ + 305, + 598, + 545, + 674 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 681, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 362, + 681, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\left\\| \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} \\right\\| = \\frac {1}{\\left\\| \\boldsymbol {w} _ {j} ^ {(l)} \\right\\|} \\| \\nabla_ {a} \\boldsymbol {u} _ {j} ^ {(l)} \\| \\tag {7}", + "image_path": "21d448abf3183fc79f949f6fbe5c77422f3d0c185b543fee154706e7a8af0581.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16440" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "type": "text", + "content": "This relationship has been found in [2, 27] and extended to any scale-invariant layers such as layer normalization [3] by [40]. It means that there are two ways to increase the gradient magnitude or the effective learning rate: 1) find a steeper descent direction where " + }, + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "type": "inline_equation", + "content": "\\|\\nabla_{a}\\pmb{u}_{j}^{(l)}\\|" + }, + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "type": "text", + "content": " is increased and 2) decrease the weight norm so that " + }, + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "type": "inline_equation", + "content": "1 / \\| \\pmb{w}_j^{(l)}\\|" + }, + { + "bbox": [ + 46, + 72, + 289, + 245 + ], + "type": "text", + "content": " is increased. For a DNN with standard BN layers, the training will exploit this property to increase the gradient magnitude by reducing the weight norms with the help of weight decay regularization, which leads to a spurious increase in gradient magnitude and a larger variance of gradient estimation. In contrast, for a DNN with fixed BN layers such as the Frozen Net, the gradient norm is not correlated with weight norm," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 252, + 287, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 252, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 89, + 252, + 287, + 287 + ], + "type": "interline_equation", + "content": "\\nabla_ {f} \\boldsymbol {w} _ {j} ^ {(l)} = \\frac {\\partial \\mathcal {L} _ {f} (\\tilde {\\boldsymbol {x}} _ {i})}{\\partial \\boldsymbol {w} _ {j} ^ {(l)}} = \\nabla \\tilde {h} _ {i j} ^ {(l)} \\frac {\\boldsymbol {h} _ {i} ^ {(l - 1)}}{\\sigma_ {j , p t} ^ {(l)}}. \\tag {8}", + "image_path": "2e420cc80a82dcc70da48848b62b03c4daad4987cb92172be6f07077fa376072.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "text", + "content": "In this gradient, the only way to increase the gradient magnitude is to find the actual steeper direction. The overall gradient for the weight is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 337, + 287, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 337, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 93, + 337, + 287, + 355 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {w} _ {j} ^ {(l)} = \\nabla_ {a} \\boldsymbol {w} _ {j} ^ {(l)} + \\lambda_ {t w i n s} \\nabla_ {f} \\boldsymbol {w} _ {j} ^ {(l)}. \\tag {9}", + "image_path": "f13606e08694ff635bc8c345add9d697b7dc670528d80a3612d9c546856046b4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 289, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 593 + ], + "type": "text", + "content": "Empirical Study To see the difference between the gradients in AT and TWINS-AT, we record the gradient of all weight parameters in each step of the two training methods and compute the mean and STD of the gradient norm in each epoch. The model parameters and their gradients are treated as long vectors and we compute the " + }, + { + "bbox": [ + 46, + 366, + 289, + 593 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 366, + 289, + 593 + ], + "type": "text", + "content": " norm of the weight and gradient vector. Figure 3 shows the mean and normalized STD of gradient norms in 60 training epochs for three datasets. Here we show the normalized STD, i.e., dividing the STD by the mean, to see the relative effect of variance. The major finding is that the gradient magnitude of TWINS-AT is substantially larger than that of AT, while the variance of TWINS-AT is lower than AT in most epochs. Note that in theory, the ratio between STD and mean should remain the same after down-scaling the weight norm in standard BN, but in practice we do observe the high normalized variance of DNNs with standard BNs since we use one epoch's gradients to approximate the variance and mean." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "One benefit of the larger gradient magnitude is that the model can escape from the initial sub-optimal point faster and find a better local optimum than the small gradient optimization [41]. We validate this hypothesis by recording the distance between the current model and the initial model during training. Figure 4 shows these weight distances on three datasets, where TWINS-AT moves away from the initial model much faster than the AT baseline. The robust overfitting effect of adversarial training [54] is partially a result of large gradient variance, especially at the final stage" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": "of training [9]. We compare the robust accuracy drop of AT and TWINS-AT in Table 1 and find that the small relative variance of TWINS-AT has the effect of reducing robust overfitting. See the experiment details in Sec. 5 and Supp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 133, + 382, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 382, + 146 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 382, + 146 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 153, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 545, + 190 + ], + "type": "text", + "content": "This section presents our experiment with TWINS. We first introduce our experiment setting and then show our main result and ablation study." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 198, + 424, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 424, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 424, + 210 + ], + "type": "text", + "content": "5.1. Experiment Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 210, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 521 + ], + "type": "text", + "content": "Dataset. We use five datasets in our experiment. CIFAR10 and CIFAR100 [34] are low-resolution image datasets, where the training and validation sets have 50,000 and 10,000 images, and CIFAR10 has 10 classes, while CIFAR100 has 100 classes. Caltech-256 [22] is a high-resolution dataset with 30,607 images and 257 classes, which is split into training and validation set with a ratio of 9:1. Caltech-UCSD Birds-200-2011 (CUB200) [60] is a high-resolution bird image dataset for fine-grained image classification, which contains 200 classes of birds, 5,994 training images and 5,794 validation images. Stanford Dogs [33] has high-resolution dog images from 120 dog categories, where the training and validation set has 12,000 and 8,580 images. For both low-resolution image datasets (CIFAR10 and CIFAR100) and high-resolution datasets (Caltech-256,CUB200 and Stanford Dogs), we resize the image to " + }, + { + "bbox": [ + 304, + 210, + 545, + 521 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 304, + 210, + 545, + 521 + ], + "type": "text", + "content": " so that the input sizes are the same for pre-training and fine-tuning. As with pre-training, the input image is normalized by the mean and STD of the pre-training set. Note that the resizing and normalization function is integrated into the model so we can attack the input image with the [0,1] bounds for pixel values as in standard adversarial attacks. We use the standard ImageNet data augmentation for high-resolution datasets [26]. For CIFAR datasets, we use random cropping with padding=4 and random horizontal flipping." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": "Adversarial Pre-Training. Large-scale adversarial pretraining on ImageNet is time-consuming, and thus [55] has released adversially pre-trained ResNet50 and WideResNet50-2, trained with " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " norm bounded attacks. In this paper, we adopt the pre-trained ResNet50 models, trained with " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " attack with bound " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon_{pt} = 4 / 255" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": ". We test other robust pre-trained models in our ablation study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Training Setting. For baselines and our method, we train all parameters of the pre-trained model, i.e., full fine-tuning instead of linear probing [57], with PGD attacks of " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " norm. The PGD step is 10, " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{ft} = 8 / 255" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " and stepsize " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 2 / 255" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ". We set the batch size as 128 and train the model for 60 epochs and divide the learning rate by 0.1 at 30th and 50th epoch. In TWINS with warmup, we initialize the means and STDs with their pre-trained values, and update the means and STDs using the target training set." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16441" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 529, + 281 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 529, + 281 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 529, + 281 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 529, + 281 + ], + "type": "table", + "html": "
MetricMethodCIFAR10CIFAR100Caltech256CUB200Stanford Dogs
Clean Acc.AT89.7769.4875.9065.7460.09
TWINS-AT91.24(+1.47)70.72(+1.24)76.86(+0.96)68.09(+2.35)64.98(+4.89)
TWINS-AT+warmup91.95(+2.18)72.12(+2.64)77.35(+1.45)67.64(+1.90)66.12(+6.03)
TRADES87.0662.7669.7058.9259.99
TWINS-TRADES86.61(-0.45)66.72(+3.96)71.12(+1.42)60.72(+1.80)60.58(+0.59)
TWINS-TRADES+warmup86.60(-0.46)65.91(+3.15)73.39(+3.69)61.05(+2.13)63.96(+3.97)
PGD10AT52.2428.5248.3726.6019.80
TWINS-AT52.73(+0.49)31.08(+2.56)48.40(+0.03)29.24(+2.64)20.58(+0.78)
TWINS-AT+warmup52.46(+0.22)29.12(+0.60)49.13(+0.76)27.67(+1.07)19.48(-0.32)
TRADES54.0432.2047.2827.8721.36
TWINS-TRADES56.23(+2.19)33.51(+1.31)47.31(+0.03)27.05(-0.82)19.93(-1.43)
TWINS-TRADES+warmup55.81(+1.77)33.48(+1.28)48.53(+1.25)26.68(-1.19)19.46(-1.90)
AAAT48.4623.4743.8522.8212.30
TWINS-AT49.81(+1.35)26.73(+3.26)43.69(-0.16)22.33(-0.49)14.37(+2.07)
TWINS-AT+warmup49.02(+0.56)25.72(+2.25)43.92(+0.07)23.58(+0.75)13.80(+1.50)
TRADES50.3126.4043.3922.2112.05
TWINS-TRADES51.71(+1.40)28.29(+1.89)41.77(-1.62)22.68(+0.47)13.36(+1.31)
TWINS-TRADES+warmup51.10(+0.79)28.30(+1.90)43.55(+0.16)21.92(-0.29)10.94(-1.11)
", + "image_path": "963ad8cf68b21cd1957afea37351500f2e60fe9cfc1a02b14ca809581c149471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 354, + 284, + 433 + ], + "blocks": [ + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "text", + "content": "Table 2. The performance of our TWINS-AT and TWINS-TRADES on five image classification tasks compared with AT and TRADES. The clean accuracy means the accuracy when testing images are input without adversarial perturbations. PGD10 and AA denote the robust test accuracy under PGD10 and AutoAttack. The increase and decrease in performance are denoted with green and red numbers. The bold numbers denote the best robust accuracy under AA. The proposed TWINS achieves better robustness and clean accuracy compared with the baseline. Averaged over the datasets, the clean and robust accuracy of TWINS are increased by " + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "inline_equation", + "content": "2.18\\%" + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "text", + "content": " compared with AT, and " + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "inline_equation", + "content": "1.46\\%" + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "inline_equation", + "content": "0.69\\%" + }, + { + "bbox": [ + 46, + 282, + 545, + 350 + ], + "type": "text", + "content": " compared with TRADES. The means and STDs of the performance are in the supplemental." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 354, + 284, + 433 + ], + "lines": [ + { + "bbox": [ + 48, + 354, + 284, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 284, + 433 + ], + "type": "table", + "html": "
MetricPT ModelCaltech256CUB200Dogs
Clean AccRandom48.9912.387.27
Non-Robust64.6653.3041.99
Robust75.9065.7460.09
PGD10Random31.783.7283.59
Non-Robust39.8619.6813.32
Robust48.3726.6019.80
", + "image_path": "55c48de41814901a2abd95cbe25dbe937a1681e7a9b15d1ec020135bfd08e7df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 483, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 628 + ], + "type": "text", + "content": "The momentum of updating statistics is 0.1, the batch size is 128, and the warmup only lasts one epoch. Note that in the warmup stage, the input samples are added with adversarial perturbations generated by the PGD attack, which has the same setting as the attack in training, and the classifier layer is the pre-trained classifier for the adversarial attack. Our pilot experiment shows that using adversarial examples as input is more effective than using clean exemplars in the warmup. The optimizer is SGD with momentum in all of our experiments. The learning rate, weight decay and regularization hyperparameter are determined by grid search, which is described in detail in the supplemental." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Adversarial robustness evaluation. Two standard adversarial attacks are used in our experiment, i.e., PGD and AutoAttack [11]. The adversarial perturbation is " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " norm bounded in our evaluation. The setting of PGD attack for validation set is the same as the PGD attack in training. The AutoAttack (AA) is a more reliable adversarial attack and more often used for evaluation than PGD in recent years." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 354, + 547, + 423 + ], + "blocks": [ + { + "bbox": [ + 46, + 434, + 287, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 477 + ], + "type": "text", + "content": "Table 3. Comparison of random initialization, non-robust and robust pre-trained model on three difficult classification tasks, when fine-tuned with AT. The robust pre-trained model is indispensable to downstream robustness." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 354, + 547, + 423 + ], + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 423 + ], + "type": "table", + "html": "
MethodMetricλWD
1e-51e-41e-31e-2
ATClean Acc.89.9289.9489.7790.28
Robust Acc.46.3444.6748.4647.57
TWINS-ATClean Acc.91.9091.4291.2487.33
Robust Acc.47.1949.0749.8151.65
", + "image_path": "d91e62e11e982168dddb9e5ada1ab73b01f330fc22894d4d6425102f121dd986.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "type": "text", + "content": "Table 4. The performance of TWINS-AT and AT on CIFAR10 when the hyperparameter for weight decay " + }, + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "type": "inline_equation", + "content": "{\\lambda }_{WD}" + }, + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "type": "text", + "content": " is changed. The robust accuracy is evaluated using AutoAttack. Our TWINS-AT achieves better adversarial robustness than AT for different " + }, + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "type": "inline_equation", + "content": "{\\lambda }_{WD}" + }, + { + "bbox": [ + 305, + 425, + 545, + 469 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 481, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 545, + 529 + ], + "type": "text", + "content": "We use the standard attacks of AA, i.e., untargeted APGD-CE, targeted APGD-DLR, targeted FAB [10] and Square Attack [1], with " + }, + { + "bbox": [ + 305, + 481, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 305, + 481, + 545, + 529 + ], + "type": "text", + "content": ". The robust accuracy in our experiment result denotes the accuracy under AA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 538, + 425, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 425, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 425, + 551 + ], + "type": "text", + "content": "5.2. Experimental Result" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "content": "On the three high-resolution datasets, we compare the performance of fine-tuning different initialization models, i.e., random initialization, standard pre-trained ResNet50 and robust pre-trained ResNet50. Table 3 shows that the pre-trained models are essential to the accuracy and robustness in challenging downstream tasks, since the random initialization is much worse than the two pre-trained models. The robust pre-trained model has a clear benefit over the standard pre-trained one, indicating that robust pre-training is indispensable to downstream robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Table 2 shows the result of TWINS-AT and TWINS-TRADES compared with the baselines. Since AA is a more reliable attack, we highlight the best robust accuracy un" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16442" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 166 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 166 + ], + "type": "image", + "image_path": "ea33abe8f798ad25652a4208a8c1e235cbe4c6777310366e8782553eb682c466.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 209 + ], + "type": "text", + "content": "Figure 5. Ablation study where the BN statistics in TWINS are initialized with (0,1) for means and STDs, denoted as TWINS-Init. The population mean and STD of pre-training are crucial to TWINS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 395 + ], + "type": "text", + "content": "der AA on each dataset. The TWINS fine-tuning learns more robust DNNs, as well as achieves a better clean accuracy on all five datasets, demonstrating the strong effectiveness of TWINS in the robust pre-training and robust fine-tuning setting. On CIFAR10 and CIFAR100, both TWINS-AT and TWINS-TRADES achieve better robustness and clean accuracy than their baselines, while the warmup only improves the clean accuracy and hurts the robustness sometimes. On Caltech-256, TWINS-AT improves upon the vanilla AT in both robustness and accuracy, but TWINS-TRADES does not perform better than vanilla TRADES. However, the warmup helps boost the performance of TWINS-TRADES as well as TWINS-AT and makes the TWINS with warmup perform better than the baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 552 + ], + "type": "text", + "content": "On the two fine-grained image classification datasets, TWINS-AT and TWINS-TRADES generally perform better than baselines in terms of accuracy and robustness, if we look at the robust accuracy under AA, where only TWINS-AT on CUB200 has a slightly worse robust accuracy than its baseline. We find that the warmup improves the clean accuracy but hurts the robustness in most cases, except for CIFAR100 and Caltech-256. This can be a result of noisy adversarial perturbation, since we use the pre-trained classifier layer in the adversarial attack, or the insufficient update steps. Nevertheless, we note that the warmup can be considered as an operation for achieving a trade-off between robustness and accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 562, + 141, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 141, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 141, + 575 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "TWINS initialization. We use the pre-trained BN statistics in the Frozen Net to keep the robust information learned during pre-training. To show the importance of the pretrained BN statistics, we use the standard initialization (mean=0 and STD=1) for BN statistics in the Frozen Net, denoted as TWINS-Init, and show the result on the five datasets in Figure 5. Both clean and robust accuracy drop when the (0,1) initialization is used in TWINS-Init, demonstrating the crucial role of pre-trained statistics in TWINS. The fact that TWINS-Init sometimes improves upon the AT baseline motivates us to investigate the effect of TWINS on" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 72, + 549, + 140 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 549, + 140 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 549, + 140 + ], + "type": "table", + "html": "
MethodMetricεpt
1/2552/2554/2558/255
ATClean Acc.65.4767.0869.4869.93
Robust Acc.24.7625.7923.4727.71
TWINS-ATClean Acc.68.5570.4570.7272.59
Robust Acc.25.9726.6226.7328.62
", + "image_path": "4cff26a6315467f4263cd219a16d50c808eab067e8255efb1dbfd90cb7441e53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "type": "text", + "content": "Table 5. The performance of TWINS-AT and AT on CIFAR100 when robust pre-trained ResNet50 with different " + }, + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\epsilon_{pt}" + }, + { + "bbox": [ + 306, + 141, + 545, + 163 + ], + "type": "text", + "content": " are used." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 171, + 421, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 421, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 421, + 182 + ], + "type": "text", + "content": "gradient norms in Section 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "content": "Effect of weight decay. Weight decay is the reason for the decreasing weight norm in DNNs with BN layers, so increasing the hyperparameter of weight decay " + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_{WD}" + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "content": " is also a way to increase the gradient magnitude. Table 4 shows the result of TWINS-AT and AT when different " + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_{WD}" + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "content": " are used when fine-tuning the robust pre-trained model on CIFAR10. The robust accuracy of TWINS-AT is consistently better than that of AT across different " + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_{WD}" + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "content": "'s. We draw the same conclusion on CIFAR100 (see supplemental). Note that the clean accuracy of TWINS-AT drops when a large " + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_{WD}" + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "content": " is used, suggesting that we should not use a too large " + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_{WD}" + }, + { + "bbox": [ + 304, + 182, + 545, + 327 + ], + "type": "text", + "content": " for TWINS-AT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": "Different robust pre-trained models. The main experiment uses the robust pre-trained ResNet50 with " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon_{pt} = 4/255" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " as the initial model. We try different robust pre-trained models with different " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon_{pt}" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " in Table 5, which shows that a larger " + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon_{pt}" + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": " is beneficial to both clean and robust accuracy in the downstream, and the proposed TWINS-AT is better than AT in both metrics with different pre-trained models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 427, + 378, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 378, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 378, + 438 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 653 + ], + "type": "text", + "content": "This paper investigates the utility of robust pre-trained models in various downstream classification tasks. We first find that the commonly used data- and model-based approaches to maintain pre-training information do not work in the adversarial robust fine-tuning. We then propose a subtle statistics-based method, TWINS, for retaining the pretraining robustness in the downstream. In addition to the robustness preserving effect, we find that TWINS increases the gradient magnitude without sacrificing the training stability and improves the training dynamics of AT. Finally, the performance of TWINS is shown to be stronger than that of AT and TRADES on five datasets. One limitation of our work is that we only evaluate the robust supervised pre-trained ResNet50. Recently, robust pre-trained ViT's on ImageNet [4] have been released. Our statistics-based approach can be extended to the layer normalization, on which the increasing gradient magnitude argument also holds, and thus future work will extend TWINS to ViT." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. CityU 11215820) and the Fundamental Research Funds for the Central University of China (DUT No. 82232031)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16443" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In European Conference on Computer Vision, pages 484-501. Springer, 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "text", + "content": "[2] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 225 + ], + "type": "text", + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 34:26831-26843, 2021. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "type": "text", + "content": "[5] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 288, + 385 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "text", + "content": "[7] Dian Chen, Hongxin Hu, Qian Wang, Li Yinli, Cong Wang, Chao Shen, and Qi Li. Cartl: Cooperative adversarially-robust transfer learning. In International Conference on Machine Learning, pages 1640–1650. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 486 + ], + "type": "text", + "content": "[8] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 699-708, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 532 + ], + "type": "text", + "content": "[9] Tianlong Chen, Zhenyu Zhang, Sijia Liu, Shiyu Chang, and Zhangyang Wang. Robust overfitting may be mitigated by properly learned smoothening. In International Conference on Learning Representations, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 288, + 577 + ], + "type": "text", + "content": "[10] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In International Conference on Machine Learning, pages 2196-2205. PMLR, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 288, + 623 + ], + "type": "text", + "content": "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 288, + 668 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[14] Gavin Weiguang Ding, Yash Sharma, Kry Yuk Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 194 + ], + "type": "text", + "content": "[15] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, et al. On robustness and transferability of convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16458-16468, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 248 + ], + "type": "text", + "content": "[16] Xinshuai Dong, Anh Tuan Luu, Min Lin, Shuicheng Yan, and Hanwang Zhang. How should pre-trained language models be fine-tuned towards adversarial robustness? Advances in Neural Information Processing Systems, 34:4356-4369, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 547, + 326 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 546, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 546, + 370 + ], + "type": "text", + "content": "[18] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature communications, 12(1):1-14, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 371, + 546, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 546, + 414 + ], + "type": "text", + "content": "[19] Samuel G Finlayson, John D Bowers, Joichi Ito, Jonathan L Zittrain, Andrew L Beam, and Isaac S Kohane. Adversarial attacks on medical machine learning. Science, 363(6433):1287-1289, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 415, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 546, + 459 + ], + "type": "text", + "content": "[20] Jonathan Frankle, David J. Schwab, and Ari S. Morcos. Training batchnorm and only batchnorm: On the expressive power of random features in {cnn}s. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 459, + 547, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 547, + 503 + ], + "type": "text", + "content": "[21] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 34:4218-4233, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 503, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 546, + 525 + ], + "type": "text", + "content": "[22] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 1, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 526, + 547, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 547, + 581 + ], + "type": "text", + "content": "[23] Minghao Guo, Yuzhe Yang, Rui Xu, Ziwei Liu, and Dahua Lin. When nas meets robustness: In search of robust architectures against adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 631-640, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 582, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 547, + 635 + ], + "type": "text", + "content": "[24] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 689 + ], + "type": "text", + "content": "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16444" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[27] Elad Hoffer, Ron Banner, Itay Golan, and Daniel Soudry. Norm matters: efficient and accurate normalization schemes in deep networks. Advances in Neural Information Processing Systems, 31, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[28] Hanxun Huang, Yisen Wang, Sarah Erfani, Quanquan Gu, James Bailey, and Xingjun Ma. Exploring architectural ingredients of adversarially robust deep neural networks. Advances in Neural Information Processing Systems, 34:5545-5559, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "text", + "content": "[29] Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, and Humphrey Shi. Semask: Semantically masked transformers for semantic segmentation. arXiv preprint arXiv:2112.12782, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "text", + "content": "[30] Daniel Jakubovitz and Raja Giryes. Improving dnn robustness to adversarial attacks using jacobian regularization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 514-529, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[31] Ahmadreza Jeddi, Mohammad Javad Shafiee, and Alexander Wong. A simple fine-tuning is all you need: Towards robust deep learning via adversarial fine-tuning. arXiv preprint arXiv:2012.13628, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[32] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 33:16199-16210, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "text", + "content": "[33] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2. Cite-seer, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 454 + ], + "type": "text", + "content": "[34] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[35] Feng Li, Hao Zhang, Shilong Liu, Lei Zhang, Lionel M Ni, Heung-Yeung Shum, et al. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. arXiv preprint arXiv:2206.02777, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[36] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. arXiv preprint arXiv:2203.16527, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "text", + "content": "[37] Yao Li, Martin Renqiang Min, Thomas Lee, Wenchao Yu, Erik Kruus, Wei Wang, and Cho-Jui Hsieh. Towards robustness of deep neural networks via regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7496-7505, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[38] Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779, 2016. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[39] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[40] Zhiyuan Li, Kaifeng Lyu, and Sanjeev Arora. Reconciling modern deep learning with traditional optimization analyses: The intrinsic learning rate. Advances in Neural Information Processing Systems, 33:14544-14555, 2020. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[41] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. Advances in Neural Information Processing Systems, 33:21476-21487, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[42] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128(2):261-318, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[43] Ziquan Liu and Antoni B Chan. Boosting adversarial robustness from the perspective of effective margin regularization. British Machine Vision Conference (BMVC), 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[44] Ziquan Liu, Yufei Cui, and Antoni B Chan. Improve generalization and robustness of neural networks via weight scale shifting invariant regularizations. ICML Workshop on Adversarial Machine Learning, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "text", + "content": "[46] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Antoni Chan, and Rong Jin. Improved fine-tuning by better leveraging pre-training data. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 408 + ], + "type": "text", + "content": "[47] Ziquan Liu, Yi Xu, Yuanhong Xu, Qi Qian, Hao Li, Rong Jin, Xiangyang Ji, and Antoni B Chan. An empirical study on distribution shift robustness from the perspective of pre-training and data augmentation. arXiv preprint arXiv:2205.12753, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[49] Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 501, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 545 + ], + "type": "text", + "content": "[50] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 546, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 601 + ], + "type": "text", + "content": "[51] Chongli Qin, James Martens, Sven Gowal, Dilip Krishnan, Krishnamurthy Dvijotham, Alhussein Fawzi, Soham De, Robert Stanforth, and Pushmeet Kohli. Adversarial robustness through local linearization. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 656 + ], + "type": "text", + "content": "[52] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[53] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Data augmentation can improve robustness. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16445" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 619 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[54] Leslie Rice, Eric Wong, and Zico Kolter. Overfitting in adversarially robust deep learning. In International Conference on Machine Learning, pages 8093–8104. PMLR, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[55] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 33:3533-3545, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[56] Vikash Sehwag, Saeed Mahloujifar, Tinashe Handina, Sihui Dai, Chong Xiang, Mung Chiang, and Prateek Mittal. Robust learning meets generative models: Can proxy distributions improve adversarial robustness? In International Conference on Learning Representations, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[57] Ali Shafahi, Parsa Saadatpanah, Chen Zhu, Amin Ghiasi, Christoph Studer, David Jacobs, and Tom Goldstein. Adversarially robust transfer learning. In International Conference on Learning Representations, 2020. 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[58] Chawin Sitawarin, Arvind Sridhar, and David Wagner. Improving the accuracy-robustness trade-off for dual-domain adversarial training. ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[59] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "type": "text", + "content": "[60] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[61] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 819-828, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "text", + "content": "[62] Yutaro Yamada and Mayu Otani. Does robustness on image net transfer to downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9215-9224, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 563 + ], + "type": "text", + "content": "[63] Yaodong Yu, Heinrich Jiang, Dara Bahri, Hossein Mobahi, Seungyeon Kim, Ankit Singh Rawat, Andreas Veit, and Yi Ma. An empirical study of pre-trained vision models on out-of-distribution generalization. In NeurIPS 2021 Workshop on Distribution Shifts: Connecting Methods and Applications, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 565, + 287, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 619 + ], + "type": "text", + "content": "[64] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16446" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_content_list.json b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5865fe08f6a8c5950567e26e89d13213a1b5b956 --- /dev/null +++ b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_content_list.json @@ -0,0 +1,1467 @@ +[ + { + "type": "text", + "text": "Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingting Zhu $^{1*}$ Xian Liu $^{2*}$ Xuanyu Liu $^{1}$ Rui Qian $^{2}$ Ziwei Liu $^{3}$ Lequan Yu $^{1\\dagger}$ $^{1}$ The University of Hong Kong \n $^{2}$ The Chinese University of Hong Kong \n $^{3}$ S-Lab, Nanyang Technological University", + "bbox": [ + 158, + 179, + 815, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ltzhu99，u3008631}@connect.hku.hk，lqyu@hku.hk, {alvinliu，qr021}@ie.cuhk.edu.hk，ziwei.liu@ntu.edu.sg", + "bbox": [ + 241, + 236, + 723, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animating virtual avatars to make co-speech gestures facilitates various applications in human-machine interaction. The existing methods mainly rely on generative adversarial networks (GANs), which typically suffer from notorious mode collapse and unstable training, thus making it difficult to learn accurate audio-gesture joint distributions. In this work, we propose a novel diffusion-based framework, named Diffusion Co-Speech Gesture (DiffGesture), to effectively capture the cross-modal audio-to-gesture associations and preserve temporal coherence for high-fidelity audio-driven co-speech gesture generation. Specifically, we first establish the diffusion-conditional generation process on clips of skeleton sequences and audio to enable the whole framework. Then, a novel Diffusion Audio-Gesture Transformer is devised to better attend to the information from multiple modalities and model the long-term temporal dependency. Moreover, to eliminate temporal inconsistency, we propose an effective Diffusion Gesture Stabilizer with an annealed noise sampling strategy. Benefiting from the architectural advantages of diffusion models, we further incorporate implicit classifier-free guidance to trade off between diversity and gesture quality. Extensive experiments demonstrate that DiffGesture achieves state-of-the-art performance, which renders coherent gestures with better mode coverage and stronger audio correlations. Code is available at https://github.com/Advocate99/DiffGesture.", + "bbox": [ + 75, + 335, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 760, + 209, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Making co-speech gestures is an innate human behavior in daily conversations, which helps the speakers to express their thoughts and the listeners to comprehend the meanings [10, 32, 38]. Previous linguistic studies verify that such non-verbal behaviors could liven up the atmosphere", + "bbox": [ + 75, + 786, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2400d7bf94a445e8333befea73469a2373b520a6be01c3f856550347bb91652e.jpg", + "image_caption": [ + "Figure 1. Illustration of Conditional Generation Process in Co-Speech Gesture Generation. The diffusion process $q$ gradually adds Gaussian noise to the gesture sequence (i.e., $\\pmb{x}_0$ sampled from the real data distribution). The generation process $p_{\\theta}$ learns to denoise the white noise (i.e., $\\pmb{x}_T$ sampled from the normal distribution) conditioned on context information $\\pmb{c}$ . Note that $\\pmb{x}_t$ denotes the corrupted gesture sequence at the $t$ -th diffusion step." + ], + "image_footnote": [], + "bbox": [ + 501, + 301, + 890, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and improve mutual intimacy [7, 8, 21]. Therefore, animating virtual avatars to gesticulate co-speech movements is crucial in embodied AI. To this end, recent researches focus on the problem of audio-driven co-speech gesture generation [16, 25, 30, 41], which synthesizes human upper body gesture sequences that are aligned to the speech audio.", + "bbox": [ + 496, + 517, + 892, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early attempts downgrade this task as a searching-and-connecting problem, where they predefine the corresponding gestures of each speech unit and stitch them together by optimizing the transitions between consecutive motions for coherent results [11,21,31]. In recent years, the compelling performance of deep neural networks has prompted data-driven approaches. Previous studies establish large-scale speech-gesture corpus to learn the mapping from speech audio to human skeletons in an end-to-end manner [4,5,25, 27,30,34,39]. To attain more expressive results, Ginosar et al. [16] and Yoon et al. [41] propose GAN-based methods to guarantee realism by adversarial mechanism, where the discriminator is trained to distinguish real gestures from the synthetic ones while the generator's objective is to fool the discriminator. However, such pipelines suffer from the inherent mode collapse and unstable training, making them difficult to capture the high-fidelity audio-conditioned gesture distribution, resulting in dull or unreasonable poses.", + "bbox": [ + 496, + 609, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent paradigm of diffusion probabilistic models", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10544", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "provides a new perspective for realistic generation [19, 37], facilitating high-fidelity synthesis with desirable properties such as good distribution coverage and stable training compared to GANs. However, it is non-trivial to adapt existing diffusion models for co-speech gesture generation. Most existing conditional diffusion models deal with static data and conditions [35, 36] (e.g., the image-text pairs without temporal dimension), while co-speech gesture generation requires generating temporally coherent gesture sequences conditioned on continual audio clips. Further, the commonly used denoising strategy in existing diffusion models samples independently and identically distributed (i.i.d.) noises in latent space to increase diversity. However, this strategy tends to introduce variation for each gesture frame and lead to temporal inconsistency in skeleton sequences. Therefore, how to generate high-fidelity co-speech gestures with strong audio correlations and temporal consistency is quite challenging within the diffusion paradigm.", + "bbox": [ + 75, + 90, + 472, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we propose a tailored Diffusion Co-Speech Gesture framework to capture the cross-modal audio-gesture associations while maintaining temporal coherence for high-fidelity audio-driven co-speech gesture generation, named DiffGesture. As shown in Figure 1, we formulate our task as a diffusion-conditional generation process on clips of skeleton and audio, where the diffusion phase is defined by gradually adding noise to gesture sequence, and the generation phase is referred as a parameterized Markov chain with conditional context features of audio clips to denoise the corrupted gestures. As we treat the multi-frame gesture clip as the diffusion latent space, the skeletons can be efficiently synthesized in a non-autoregressive manner to bypass error accumulation. To better attend to the sequential conditions from multiple modalities and enhance the temporal coherence, we then devise a novel Diffusion Audio-Gesture Transformer architecture to model audio-gesture long-term temporal dependency. Particularly, the per-frame skeleton and contextual features are concatenated in the aligned temporal dimension and embedded as individual input tokens to a Transformer block. Further, to eliminate the temporal inconsistency caused by the naive denoising strategy in the inference stage, we thus propose a new Diffusion Gesture Stabilizer module to gradually anneal down the noise discrepancy in the temporal dimension. Finally, we incorporate implicit classifier-free guidance by jointly training the conditional and unconditional models, which allows us to trade off between the diversity and sample quality during inference.", + "bbox": [ + 75, + 367, + 472, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on two benchmark datasets show that our synthesized results are coherent with stronger audio correlations and outperform the state-of-the-arts with superior performance on co-speech gesture generation. To summarize, our main contributions are three-fold: 1) As an early attempt at taming diffusion models for co-speech", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gesture generation, we formally define the diffusion and denoising process in gesture space, which synthesizes audio-aligned gestures of high-fidelity. 2) We devise the Diffusion Audio-Gesture Transformer with implicit classifier-free diffusion guidance to better deal with the input conditional information from multiple sequential modalities. 3) We propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy.", + "bbox": [ + 496, + 90, + 893, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 227, + 640, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Co-Speech Gesture Generation. Synthesizing co-speech gestures is crucial for a variety of applications. Conventional studies resort to rule-based pipelines [11, 21, 31], where linguistic experts pre-define the speech-gesture pairs and refine the transitions between different motions. Recent works exploit neural networks to learn the mapping from speech to gesture based on a large training corpus, where an off-the-shelf pose estimator is leveraged to label the online videos for pseudo annotations [1, 16, 30, 34, 41, 42]. Meanwhile, some works study the influence of input modality, verifying the connections between co-speech gesture and speech audio [25], text transcript [3], speaking style [2], and speaker identity [41]. To further improve the model's capacity, previous studies explore multiple architecture choices, including CNN [17], RNN [42], Transformer [6], and VQVAE [29, 40]. Notably, several recent works are based on GANs to guarantee realistic results [16, 30, 34, 41], which involve the adversarial training between the generator and the discriminator. However, the notorious mode collapse and unstable training of GANs prevent the high-fidelity gesture distribution learning conditioned on audio.", + "bbox": [ + 496, + 253, + 893, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Probabilistic Models. Diffusion probabilistic models have achieved promising results on unconditional image generation [19], which are further applied to conditional tasks like text-to-image [36]. Among the diffusion-based literature, previous works focus on static data and conditions. Besides, they mainly utilize explicit guidance like pretrained classifiers [13] and CLIP similarity [28, 33, 35, 36] to guide the generation process. In this work, we explore a more challenging co-speech gesture generation setting, where the gesture data and audio conditions are both sequential, and the audio-to-gesture mapping is implicit. To this end, we propose the Diffusion Audio-Gesture Transformer to guarantee temporally aligned generation. We further propose the Diffusion Gesture Stabilizer to simultaneously achieve diverse and temporally coherent gestures with an annealed noise sampling strategy.", + "bbox": [ + 496, + 571, + 895, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 500, + 828, + 648, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 2 depicts an overview of the proposed DiffGesture framework to generate co-speech gestures of high fidelity. In this section, we first introduce the problem", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10545", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/33f47d693166d371a24e87d4333ff5521d748911384902045219a7b08e41fa48.jpg", + "image_caption": [ + "Figure 2. Overview of the Diffusion Co-Speech Gesture (DiffGesture) Framework. Given the gesture sequence $\\pmb{x}_0$ , we first establish the forward diffusion (purple) and conditional denoising process (green) in gesture space. Then, we devise the Diffusion Audio-Gesture Transformer to attend to the input conditions of initial poses $\\pmb{p}_{(1:M)}$ , speech audio $\\pmb{a}$ , time embedding $t$ and corrupted gesture $\\pmb{x}_t$ from multiple modalities (blue). At the diffusion sampling stage (grey), we propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy. To further incorporate implicit classifier-free guidance, we jointly train the conditional $(1 - p_{\\text{uncond}})$ and unconditional $(p_{\\text{uncond}})$ models. This allows us to trade off between diversity and quality during inference." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 331, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/432316dc0fac48cbc18da6b2853d75cd9374a705e72851bedff75f5f61a4a010.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 90, + 723, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/10612c5ffe7a00e2688305fa64cd3490996b86ccc99125c08e3a3803274a9f9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 90, + 887, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formulation of audio-driven co-speech gesture generation (Section 3.1). We then establish the forward diffusion and the reverse conditional generation process in gesture space (Section 3.2). Furthermore, we elaborate the Diffusion Audio-Gesture Transformer to attend to the conditions from multiple modalities and enhance the speech-gesture correlations with temporal dependency (Section 3.3). To eliminate temporal inconsistency introduced by naive noises, we propose a novel Diffusion Gesture Stabilizer with annealed noise sampling strategies and describe this module in (Section 3.4). Finally, incorporating implicit classifier-free guidance in co-speech gestures is discussed in (Section 3.5).", + "bbox": [ + 75, + 386, + 468, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 585, + 279, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With a large-scale co-speech gesture training corpus, we leverage the speaking videos with clear co-speech upper body movements for model learning. In particular, for each video clip of $N$ frames, we extract the accompanying speech audio sequence $\\pmb{a} = \\{\\pmb{a}_1,\\dots ,\\pmb{a}_N\\}$ and use the offthe-shelf human pose estimator OpenPose [9] to annotate the per-frame human skeletons as $\\pmb{x} = \\{\\pmb{p}_1,\\dots ,\\pmb{p}_N\\}$ . We follow baseline methods [30, 41] to pre-process such skeletal representation as the concatenation of unit direction vectors as $\\pmb{p_i} = [\\pmb{d}_{i,1},\\pmb{d}_{i,2},\\dots ,\\pmb{d}_{i,J - 1}]$ , where $\\pmb{p_i}$ denotes the pose description coordinates of the $i$ -th frame, $J$ is the total joint number and $\\pmb{d}_{i,j}$ represents the $j$ -th unit direction vector among the $J$ joints of the $i$ -th image frame. The diffusion model's reverse denoising process $G$ parameterized by $\\theta$ is optimized to synthesize the human skeleton sequence $\\pmb{x}$ , which is further conditioned on the speech audio sequence $\\pmb{a}$ and the initial poses $\\{\\pmb{p}_1,\\dots ,\\pmb{p}_M\\}$ of the first $M$ frames. The learning objective of the overall framework can be formulated as $\\arg \\min_{\\theta}||\\pmb{x} - G_{\\theta}(\\pmb {a},\\pmb {p}_1,\\dots ,\\pmb {p}_M)||$", + "bbox": [ + 75, + 613, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Gesture Space Forward and Reverse Process", + "text_level": 1, + "bbox": [ + 500, + 386, + 877, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $\\pmb{x}_0 \\in \\mathbb{R}^{N \\times 3(J-1)}$ sampled from real data distribution $q(\\pmb{x}_0)$ , our goal is to learn a model distribution $p_{\\theta}(\\pmb{x}_0)$ parameterized by $\\theta$ that approximates $q(\\pmb{x}_0)$ . Specifically, denoising diffusion probabilistic models (DDPMs) [19] define the latent variable models of the form $p_{\\theta}(\\pmb{x}_0) = \\int p_{\\theta}(\\pmb{x}_{0:T}) d\\pmb{x}_{1:T}$ , where $\\pmb{x}_{1:T}$ are latent variables in the same sample space as $\\pmb{x}_0$ with the same dimensionality.", + "bbox": [ + 498, + 409, + 890, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Forward Diffusion Process. The forward process, which is also termed as the diffusion process, approximates the posterior distribution $q(\\pmb{x}_{1:T}|\\pmb{x}_0)$ . It is defined as a Markov chain that gradually adds Gaussian noise to the data sample according to a variance schedule $\\beta_1,\\dots ,\\beta_T$ :", + "bbox": [ + 498, + 518, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 609, + 890, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e \\quad q (\\boldsymbol {x} _ {t} | \\boldsymbol {x} _ {t - 1}) = \\mathcal {N} (\\boldsymbol {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\boldsymbol {x} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 654, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The variances $\\beta_{t}$ are constant hyperparameters to ease the modeling of the reverse process [19]. Through such a corruption scheme, the structural information of the original skeleton is gradually substituted by noises, which finally leads to a pure white noise when $T$ goes to infinity. Therefore, the prior latent distribution of $p(\\boldsymbol{x}_T)$ is $\\mathcal{N}(\\boldsymbol{x}_T; \\mathbf{0}, \\boldsymbol{I})$ with only information of Gaussian noise.", + "bbox": [ + 496, + 686, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reverse Conditional Gesture Generation. The reverse process, which is also termed as the generative process, estimates the joint distribution of $p_{\\theta}(\\pmb{x}_{0:T})$ . As proved in [15], the reverse process of the continuous diffusion process preserves the same transition distribution form, which motivates us to leverage a Gaussian transition to formulate $p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t)$ under an unconditional setting, which", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10546", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approximates the intractable process as:", + "bbox": [ + 76, + 90, + 341, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {x} _ {0: T}\\right) = p _ {\\theta} \\left(\\boldsymbol {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 118, + 468, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_{\\theta}(\\boldsymbol{x}_{t - 1}|\\boldsymbol{x}_t) = \\mathcal{N}(\\boldsymbol{x}_{t - 1};\\mu_{\\theta}(\\boldsymbol{x}_t,t),\\Sigma_{\\theta}(\\boldsymbol{x}_t,t))$ (4)", + "bbox": [ + 84, + 162, + 468, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The corrupted noisy data $\\pmb{x}_t$ is sampled by $q(\\pmb{x}_t|\\pmb{x}_0) = \\mathcal{N}(\\pmb{x}_t; \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0, (1 - \\bar{\\alpha}_t)\\pmb{I})$ , where $\\alpha_t = 1 - \\beta_t$ and $\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s$ . Note that we set the variances $\\Sigma_\\theta(\\pmb{x}_t, t) = \\beta_t\\pmb{I}$ to untrained time-dependent constants. The above diffusion model formulations show compelling performances on unconditional generation. To further adapt to the conditional co-gesture synthesis, we have to provide additional inputs to the model, including the audio and initial poses. Therefore, we treat the speech audio $\\pmb{a}$ and initial poses $\\pmb{p}_{1:M}$ as context information $\\pmb{c}$ and inject conditions into the generation process. The reverse process of each timestep (Eq. 4) can be thus updated as:", + "bbox": [ + 76, + 191, + 472, + 375 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}, \\boldsymbol {c}) = \\mathcal {N} (\\boldsymbol {x} _ {t - 1}; \\mu_ {\\theta} (\\boldsymbol {x} _ {t}, t, \\boldsymbol {c}), \\beta_ {t} \\boldsymbol {I}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 385, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we could start the generation process by firstly sample a Gaussian noise $\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\boldsymbol{I})$ and follow the Markov chain to iteratively denoise the latent variable $\\boldsymbol{x}_t$ via Eq. 5 to get the final results. The overview of conditional co-speech gesture process is illustrated in Figure 1.", + "bbox": [ + 75, + 412, + 468, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objective. To optimize the overall framework, we optimize the variational lower bound on negative log-likelihood: $\\mathbb{E}[-\\log p_{\\theta}(\\pmb{x}_0)] \\leq \\mathbb{E}_q[-\\log \\frac{p_{\\theta}(\\pmb{x}_0)}{q(\\pmb{x}_{1:T}|\\pmb{x}_0)}]$ . We rewrite the loss function conditioned on context $c$ and eliminate all the constant items that do not require training: $L(\\theta) = \\mathbb{E}_q[\\sum_{t=2}^{T} D_{KL}(q(\\pmb{x}_{t-1}|\\pmb{x}_t, \\pmb{x}_0)||p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t, c))]$ . With reparameterization, we can represent each term in $L_{\\theta}$ using MSE loss. We follow [19] to further simplify the training objective to the ensemble of MSE losses as:", + "bbox": [ + 76, + 489, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta) = \\mathbb {E} _ {q} [ \\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\boldsymbol {c}, t) \\right\\| ^ {2} ], (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 652, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is uniformly sampled between 1 and $T$ . As we jointly train the model under conditional and unconditional setting, a trainable masked embedding with probability $p_{\\text{uncond}}$ replaces context $c$ and the diffusion model predicts the noise in the unconditional setting. The detailed principles will be discussed in Section 3.5.", + "bbox": [ + 75, + 685, + 468, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Diffusion Audio-Gesture Transformer", + "text_level": 1, + "bbox": [ + 76, + 786, + 403, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the naive conditional generation scheme as specified in Section 3.2, we still confront a critical problem in the setting of co-speech gesture generation. Since $x_0$ denotes the skeleton sequence of $N$ frames, there exists temporal dependency among the target sequence and context information, making it more complex than time-invariant tasks", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "like image generation. Therefore, how to guarantee temporally coherent results in a non-autoregressive conditional generation process remains an unsolved problem.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to most previous studies that resort to recurrent networks [30, 41], we propose to make use of the Transformer's strong capacity in sequential data modeling. Specifically, since the noisy gesture sequence $\\boldsymbol{x}_t$ and the contextual information $c$ align in the temporal dimension, we concatenate them in the feature channel. In this way, the skeleton and context condition of each frame serve as an individual token, which captures the long-term dependency by the self-attention mechanism:", + "bbox": [ + 496, + 138, + 890, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\mathsf {T}}}{\\sqrt {\\ell}}\\right) \\mathbf {V}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 290, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q},\\mathbf{K},\\mathbf{V}$ are the query, key, and value matrix from input tokens, $\\ell$ is the channel dimension, and $\\mathsf{T}$ is the matrix transpose operation. Such a design also avoids severe error accumulation in autoregressive pipelines, enabling us to generate coherent gesture sequences.", + "bbox": [ + 496, + 339, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Diffusion Gesture Stabilizer", + "text_level": 1, + "bbox": [ + 500, + 431, + 750, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In DDPMs, the independent random variables $z$ introduced at the sampling stage promote diversity and thus systematically improve the task performance. However, the variation in temporal dimension introduced by $z$ , especially when timestep $t$ is small in the reverse process, can have a negative effect on temporal consistency. At the inference stage, to achieve the trade-off between diversity and temporal consistency, we propose a novel Diffusion Gesture Stabilizer without extra training expenses under two annealed scenarios, where the term \"annealed\" means that the process is transitioned from high variance and entropy (hot) to low variance and entropy (cold).", + "bbox": [ + 496, + 457, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thresholding. Since temporally independent Gaussian noises inevitably introduce inconsistency, restricting the temporal variation naturally helps to avoid inconsistency. And hard thresholding serves as an effective trick. In detail, we set a time threshold $t_0$ , and then use the same $z \\in \\mathbb{R}^{N \\times C}$ in the naive sampling strategy for $t > t_0$ and set $z = \\{z_0\\}_{i=1}^N$ for $t \\leq t_0$ , where $z_0 \\in \\mathbb{R}^C$ follows $\\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ which do not introduce variation in the temporal dimension.", + "bbox": [ + 496, + 641, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Smooth Sampling. We further modify $z(t) = \\{z_i(t)\\}_{i=1}^N$ to be a smooth annealing version via variance-aware sampling. In the original sampling rule of DDPMs, i.i.d.variables $z_i(t)$ are sampled from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . With smooth resampling, we first sample $z_0(t) \\sim \\mathcal{N}(\\mathbf{0},\\sigma_a^2(t)\\mathbf{I})$ only once for each timestep $t$ in the reverse process, then given $z_0(t)$ , we sample $z_i(t)|z_0(t) \\sim \\mathcal{N}(z_0(t),(1 - \\sigma_a^2(t))\\mathbf{I})$ for $i \\in \\{1,\\dots,N\\}$ , where $\\sigma_a(t) \\in [0,1]$ is a nondecreasing function to achieve variance annealing.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10547", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Training", + "text_level": 1, + "bbox": [ + 78, + 90, + 230, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: repeat", + "2: Sample $(\\pmb{x}_0,\\pmb{c})\\sim q(\\pmb{x}_0,\\pmb{c})$", + "3: Sample $\\tau \\sim \\mathrm{Uniform}(0,1)$ . Set $c = \\varnothing$ if $\\tau < p_{\\text{uncond}}$", + "4: Sample $t \\sim \\mathrm{Uniform}(\\{1, \\dots, T\\})$", + "5: Compute $\\nabla_{\\theta}\\left\\| \\pmb {\\epsilon} - \\pmb{\\epsilon}_{\\theta}(\\sqrt{\\bar{\\alpha}_{t}}\\pmb {x}_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\pmb {\\epsilon},\\pmb {c},t)\\right\\|^{2}$", + "6: Perform gradient descent", + "7: until converged" + ], + "bbox": [ + 86, + 109, + 436, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implicit Classifier-free Guidance", + "text_level": 1, + "bbox": [ + 76, + 234, + 362, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In co-speech gesture literature, the speech-to-gesture mapping is implicit, where the same audio corresponds to diverse gestures and different audiOs could incur the same motion [22, 25], making it difficult to utilize the commonly used explicit classifier guidance [13, 33]. Therefore, how can we further exploit practical guidance for better audio correlations and mode coverage? Our solution to this question is to train an extra \"unconditional\" diffusion model to implicitly guide the generation. Dhariwal et al. [13] first introduce the classifier guidance of cross-entropy gradient $\\nabla_{\\pmb{x}_t}\\log p_\\phi (y|\\pmb {x}_t)$ , where the pretrained classifier is parameterized by $\\phi$ and $y$ denotes the classification logits. This gradient term is further scaled by the covariance matrix to modify the mean value of transition distribution in Eq. 4. To adapt to the cases where no explicit guidance is available, we follow [20] to jointly train the conditional and unconditional models, termed as classifier-free guidance. In particular, according to the implicit classifier's property that $p(c|x_t)\\propto p(x_t|c) / p(x_t)$ , we could derive a gradient relationship in the implicit classifier as:", + "bbox": [ + 76, + 260, + 472, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {c} | \\boldsymbol {x} _ {t}) \\propto \\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {x} _ {t} | \\boldsymbol {c}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {x} _ {t}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 578, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is further proportional to $\\epsilon^{*}(\\boldsymbol{x}_{t}|\\boldsymbol{c}) - \\epsilon^{*}(\\boldsymbol{x}_{t})$ . Therefore, we use a single Transformer network to parameterize both settings by a mix-up training trick: for the probability of $p_{uncond}$ , we set the context information $\\boldsymbol{c}$ as masked embedding to train the unconditional setting, while for other cases, we train the original conditional counterpart. The training is shown in Algorithm 1.", + "bbox": [ + 76, + 609, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling with Classifier-free Guidance. Starting from Gaussian noise, we iteratively remove noises in $x_{t}$ . As we use implicit classifier-free guidance, similar to Equation 8, the predicted Gaussian noise is modified as:", + "bbox": [ + 76, + 717, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\epsilon}} _ {\\theta} = \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) + s \\cdot \\left(\\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 792, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ is the scale parameter to trade off the diversity and quality. With classifier-free guidance, Algorithm 2 reveals how to generate co-speech gestures given the trained diffusion model via the Diffusion Gesture Stabilizer with the Smooth Sampling annealed scenario.", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 Sampling", + "text_level": 1, + "bbox": [ + 501, + 90, + 661, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Trained diffusion model $\\theta, \\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$", + "2: for $t = T, \\dots, 1$ do", + "3: $\\hat{\\pmb{\\epsilon}}_{\\theta} = \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) + s\\cdot (\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,\\pmb {c},t) - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t))$", + "4: $\\mathbf{z}_0(t) \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_a^2(t)\\mathbf{I})$", + "5: for $i = 1,\\dots ,N$ do", + "6: $\\pmb{z}_i(t) \\sim \\mathcal{N}(\\pmb{z}_0(t), (1 - \\sigma_a^2(t))\\pmb{I})$", + "7: end for", + "8: $\\mathbf{z}(t) = \\{\\mathbf{z}_1(t),\\dots ,\\mathbf{z}_N(t)\\}$ , if $t > 1$ , else $\\mathbf{z}(t) = \\mathbf{0}$", + "9: $\\pmb{x}_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}} \\left( \\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}} \\hat{\\pmb{\\epsilon}}_\\theta \\right) + \\sigma_t \\pmb{z}(t)$", + "10: end for", + "11: return $x_0$" + ], + "bbox": [ + 503, + 109, + 849, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 291, + 632, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Co-Speech Gesture Datasets", + "text_level": 1, + "bbox": [ + 500, + 316, + 751, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TED Gesture. As a large-scale dataset for gesture generation research, TED Gesture dataset [41, 42] contains 1,766 TED videos of different narrators covering various topics. We follow the data process in former works [30, 41], where the poses are resampled with 15 FPS, and frame segments of length 34 are obtained with a stride of 10.", + "bbox": [ + 496, + 339, + 890, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TED Expressive. While the poses in TED Gesture only contain 10 upper body key points without vivid finger movements, the TED Expressive dataset [30] is further expressive of both finger and body movements. The state-of-art 3D pose estimator ExPose [12] is used to fully capture the pose information in data. As a result, TED Expressive annotates the 3D coordinates of 43 keypoints, including 13 upper body joints and 30 finger joints.", + "bbox": [ + 496, + 430, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 560, + 705, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. We compare our method on two benchmark datasets with the state-of-the-art methods in recent years. 1) Attention Seq2Seq [42] elaborates on the attention mechanism to generate pose sequences from speech text. 2) Speech2Gesture [16] uses spectrums of the speech audio segments as the input and generates speech gestures adversarially. 3) Joint Embedding [3] maps text and motion to the same embedding space, then generates outputs from motion description text. 4) Trimodal [41] serves as a strong baseline that learns from text, audio, and speaker identity to generate gestures, outperforming former methods by a large margin. 5) HA2G [30] introduces a hierarchical audio learner that captures information across different semantic granularities, achieving state-of-the-art performances. This method hierarchically extracts rich features at the cost of heavier GPU memory overhead, while our method requires much smaller expenses.", + "bbox": [ + 496, + 583, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For all the methods in both datasets, we set $N = 34$ and $M = 4$ to get $M$ -frame pose sequences where the first $N$ frames are used for reference, termed as initial poses. There are $J$ upper body joints in", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10548", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/30355dba6f476dc87ddb782b7dcc713fa2118a63436183d44d452990bdecd6ef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
Ground Truth00.698108.52500.703178.827
Attention Seq2Seq [42]18.1540.19682.77654.9200.152122.693
Speech2Gesture [16]19.2540.66893.80254.6500.679142.489
Joint Embedding [3]22.0830.20090.13864.5550.130120.627
Trimodal [41]3.7290.667101.24712.6130.563154.088
HA2G [30]3.0720.672104.3225.3060.641173.899
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
", + "bbox": [ + 183, + 88, + 787, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The Quantitative Results on TED Gesture [41] and TED Expressive [30]. We compare the proposed diffusion-based method against recent SOTA methods [3, 16, 30, 41, 42] and ground truth. For FGD, the lower, the better; for other metrics, the higher, the better.", + "bbox": [ + 75, + 270, + 890, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "all the frames of pose sequences, where $J = 10$ for TED Gesture and $J = 43$ for TED Expressive. Following [41], to eliminate the effect of the joint lengths and root motion, we represent the joints' positions using $J - 1$ directional vectors normalized to the unit vectors and train the model to learn the directional vectors. For the audio processing, we use the same audio encoder used in [41] to extract the feature of the raw audio clips directly. The audio clips are encoded as $N$ audio feature vectors of 32-D. The audio feature and initial poses are concatenated to form the conditional context information of the diffusion model. For the diffusion process, the number of timesteps is $T = 500$ , and the variances increase linearly from $\\beta_{1} = 1e - 4$ to $\\beta_{T} = 0.02$ . For the Stabilizer, $t_0$ can be adjusted from 20-30 for Thresholding, and a quadratic non-increasing function $\\sigma_{a}(t)$ is applied for Smooth Sampling. The hidden dimension of the transformer blocks, is set to 256 for TED Gesture and 512 for TED Expressive. We use 8 Transformer blocks, each of which comprises a multi-head self-attention block and a Feed-Forward Network. We use an Adam optimizer, and the learning rate is $5e - 4$ . It takes 10 hours to train the model on TED Gesture and 20 hours on TED Expressive on a single NVIDIA GeForce RTX 3090 GPU.", + "bbox": [ + 75, + 325, + 472, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 680, + 261, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In evaluation, we use three metrics that are used in cospeech gesture generation and relative fields [24, 30].", + "bbox": [ + 75, + 703, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fréchet Gesture Distance (FGD). Similar to the Fréchet Inception Distance (FID) metric [18], which is widely applied in image generation studies, FGD is used to measure the distance between the synthesized gesture distribution and the real data distribution. Yoon et al. [41] define FGD by training a skeleton sequence auto-encoder to extract the features of the real gesture sequences $X$ and the features of the generated gesture sequences $\\hat{X}$ :", + "bbox": [ + 75, + 734, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F G D} (X, \\hat {X}) = \\| \\mu_ {r} - \\mu_ {g} \\| ^ {2} + \\operatorname {T r} (\\Sigma_ {\\mathrm {r}} + \\Sigma_ {\\mathrm {g}} - 2 (\\Sigma_ {\\mathrm {r}} \\Sigma_ {\\mathrm {g}}) ^ {1 / 2}),\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 859, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu_r$ and $\\Sigma_r$ are the first and the second moments of the", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "latent feature distribution of the real gestures $X$ , and $\\mu_g$ and $\\Sigma_g$ are the first and the second moments of the latent feature distribution of the generated gestures $\\hat{X}$ . We intuitively find that among the three metrics, FGD tells the most whether the generated pose sequences are of high quality.", + "bbox": [ + 496, + 325, + 890, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Beat Consistency Score (BC). Proposed in [24, 26], BC measures motion-audio beat correlation. Considering that the kinematic velocities vary from different joints, we use the change of included angle between bones to track motion beats following [30]. Specifically, we can calculate the mean absolute angle change (MAAC) of angle $\\theta_{j}$ in adjacent frames by $\\mathrm{MAAC}(\\theta_j) = \\frac{1}{S}\\frac{1}{T - 1}\\sum_{s = 1}^{S}\\sum_{t = 1}^{T - 1}\\| \\theta_{j,s,t + 1} - \\theta_{j,s,t}\\| _1$ , where $S$ denotes the total number of clips in the dataset, $T$ denotes the number of frames in each clip, and $\\theta_{j,s,t}$ is the included angle between the $j$ -th and the $(j + 1)$ -th bone of the $s$ -th clip at time-step $t$ . Then, we can compute the angle change rate of frame $t$ for the $s$ -th clip as $\\frac{1}{J - 1}\\sum_{j = 1}^{J - 1}(\\| \\theta_{j,s,t + 1} - \\theta_{j,s,t}\\| _1 / \\mathrm{MAAC}(\\theta_j))$ . Then we extract the local optima whose first-order difference is higher than a threshold to get kinematic beats, which are used to compute BC later. Following [24] to detect audio beat by onset strength [14], we compute the average distance between each audio beat and its nearest motion beat as Beat Consistency Score:", + "bbox": [ + 496, + 402, + 892, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {B C} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\exp \\left(- \\frac {\\min _ {\\forall t _ {j} ^ {y} \\in B ^ {y}} \\| t _ {i} ^ {x} - t _ {j} ^ {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 699, + 890, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t_i^x$ is the $i$ -th audio beats, $B^y = \\{t_i^y\\}$ is the set of the kinematic beats, and $\\sigma$ is a parameter to normalize sequences, set to 0.1 empirically.", + "bbox": [ + 496, + 750, + 890, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity. This metric evaluates the variations among generated gestures corresponding to various inputs [23]. We use the same feature extractor when measuring FGD to map synthesized gestures into latent feature vectors and calculate the mean feature distance. In detail, we randomly pick 500 generated samples and compute the mean absolute error between the features and the shuffled features.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10549", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cdacf38770acbf609b3085c1b3fa88d95f5e6ed1d2b26ee14dea89b7e1612cba.jpg", + "image_caption": [ + "Figure 3. Visualization Results of Our DiffGesture on Two Datasets. Three cases are picked up, where (i) and (ii) are TED Expressive cases, and (iii) is a TED Gesture case. We highlight dull cases generated by comparison methods with rectangles, indicating the mode collapse phenomenon of baselines." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 467 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b81460fe9c6083b675e82904ef61c0818188974a7077ca73577a694693c53289.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsGTS2S. [42]S2G. [16]Joint. [3]Tri. [41]HA2G [30]DiffGesture(Ours)
Naturalness4.331.222.561.223.223.674.00
Smoothness3.943.501.613.443.443.393.89
Synchrony4.001.673.171.393.283.443.89
", + "bbox": [ + 129, + 531, + 841, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. User Study Results. The ratings of motion naturalness, smoothness, and synchrony, are on a scale of 1-5, with 5 being the best.", + "bbox": [ + 78, + 619, + 887, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Evaluation Results", + "text_level": 1, + "bbox": [ + 76, + 660, + 256, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. We compare our method with all the baselines with three metrics on TED Gesture and TED Expressive. The results are shown in Table 1. For the metrics of Ground Truth, we report the values in our implementation. For TED Gesture, we report FGD of all baselines in [30] and evaluate BC and Diversity on our own†. For TED Expressive, all the results of baselines are reported from [30]. Assuming the pseudo ground truth pose follows the real distribution, the FGD of Ground Truth in the table is 0. It is observed that our DiffGesture achieves state-of-the-art performance on both datasets, especially outperforming existing methods by a large margin on TED", + "bbox": [ + 75, + 683, + 470, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expressive. Besides, since BC and Diversity are proposed to measure motion-audio beat correlation and variation, these two metrics of Ground Truth cannot be treated as upper bounds, and it is worth noting that our results may be higher than the ones of Ground Truth, indicating that the generated gestures are of high quality. Qualitative Results. We show the keyframes of all the methods on two datasets in Figure 3. Since TED Expressive requires a higher ability of generative models, we pick up two cases for TED Expressive and one for TED Gesture. For each case, we select three keyframes (an early, a middle, and a late one) to show the pose motions. Comparison methods tend to generate slow and invariant poses and sometimes produce unreliable and stiff results. In contrast, DiffGesture produces diverse human-like poses without resulting in", + "bbox": [ + 496, + 660, + 890, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Since there exists an evaluation bug for the BC metric in HA2G [30], we report the re-implemented results from Liu et al.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10550", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5270cdced9380e9fc35b6ff2fb95e16427276d509b41d9957885b607c2c153e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
DiffGesture Base2.4500.632104.6883.8220.707174.377
DiffGesture w/o Stabilizer2.2190.674105.1922.7920.721180.125
DiffGesture w/o classifier-free1.8100.673105.6443.3260.717178.245
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
", + "bbox": [ + 161, + 88, + 807, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mean poses which are slow and rigid. Besides, as a primary drawback of GAN-based methods, mode collapse makes comparison methods often produce a single type of output, which is severe for pose motion generation. Such a phenomenon is shown in Fig. 3, where the generated frames with nearly the same pose are highlighted with rectangles.", + "bbox": [ + 75, + 267, + 468, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. To better validate the qualitative performance, we conduct a user study on the generated co-speech gestures. The study involves 18 participants, with 9 females and 9 males in the age range of 18-25 years old. The participants are required to grade the motion's quality and coherence, and all the clips are without labels. In total, we pick up 30 cases, 20 for TED-Expressive and 10 for TED Gesture. For each case, we show 7 videos with the order of the methods shuffling, including the ground truth. We adopt the Mean Opinion Scores rating protocol, and each participant is required to rate three aspects of generated motions: Naturalness; Smoothness; Synchrony between speech and generated gestures. The results are shown in Table 2 where the ratings are on a scale of 1 to 5, with 5 being the best. Our participants widely accept that our method produces high-fidelity results in all three aspects.", + "bbox": [ + 75, + 359, + 470, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 616, + 241, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on the Proposed Modules. To demonstrate the effectiveness of our proposed DiffGesture, we present ablation studies on the key modules in the framework. In detail, we conduct experiments as follows. 1) DiffGesture Base means we only use the proposed conditional diffusion generation process without further design. 2) DiffGesture w/o classifier-free, where classifier-free guidance is not implemented at both the training stage and inference stage. 3) DiffGesture w/o Stabilizer, where Diffusion Gesture Stabilizer is removed at the inference stage. The results are reported in Table 3. The results illustrate the effectiveness of the designed Diffusion Gesture Stabilizer and the implicit classifier-free guidance.", + "bbox": [ + 75, + 641, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on the Network Architectures. We investigate the performance of the GRU architecture in diffusion models, autoregressively generating poses in [30, 41]. We replace the Diffusion Audio-Gesture Transformer with the", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9289775af7ba2097cadb42b6bedc478d733a0bcbebbd5d5ee1363dbae51675d.jpg", + "table_caption": [ + "Table 3. Ablation Study on the Proposed Modules. We investigate effectiveness of the proposed modules, Diffusion Gesture Stabilizer and implicit classifier-free guidance. The results indicate that the proposed modules consistently improve performance on the benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodsFGD ↓BC ↑Diversity ↑
GRU on Da14.3430.65898.472
Transformer on Da1.5060.699106.722
GRU on Db17.4520.680172.168
Transformer on Db2.6000.718182.757
", + "bbox": [ + 514, + 263, + 879, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Study on the Network Architectures. We compare the performance of GRU and Transformer for diffusion-based backbone on TED Gesture $(D_{a})$ and TED Expressive $(D_{b})$ .", + "bbox": [ + 498, + 373, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GRU in the diffusion model. All the context inputs remain the same as our method and are concatenated before the diffusion network. Results are shown in Table 4. Though GRU serves as a strong baseline network in co-gesture learning, it fails to generate high-performance data like our designed Transformer-based network, which indicates the effectiveness of our Transformer-based network and that applying diffusion models in the audio-driven conditional generation is a non-trivial task.", + "bbox": [ + 496, + 438, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 588, + 617, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present a novel diffusion-based framework DiffGesture for co-speech gesture generation. To generate coherent gestures with strong audio correlations, we propose the Diffusion Audio-Gesture Transformer with the Diffusion Gesture Stabilizer to better attend to the condition information. Such a non-autoregressive pipeline helps to efficiently generate results and reduce error accumulation. We hope our method offers a new perspective for diffusion-based temporal generation and how to capture sequential cross-modal dependencies.", + "bbox": [ + 496, + 613, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. The work described in this paper was partially supported by grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (T45-401/22-N), the National Natural Science Fund (62201483), HKU Seed Fund for Basic Research (202009185079 and 202111159073), RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10551", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1884–1895, 2020. 2", + "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2", + "[3] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2, 5, 6, 7", + "[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, volume 39, pages 487-496. Wiley Online Library, 2020. 1", + "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1", + "[6] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 2", + "[7] Judee K Burgoon, Thomas Birk, and Michael Pfau. Nonverbal behaviors, persuasion, and credibility. Human communication research, 17(1):140-169, 1990. 1", + "[8] B. Butterworth and U. Hadar. Gesture, speech, and computational stages: A reply to mcneill. Psychological Review, 96(1):168-174, 1989. 1", + "[9] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 3", + "[10] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1", + "[11] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 1, 2", + "[12] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 5", + "[14] Daniel Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36:51-60, 03 2007. 6", + "[15] William Feller. On the theory of stochastic processes, with particular reference to applications. In *Berkeley Symposium on Mathematical Statistics and Probability*, pages 403-432, 1949. 3", + "[16] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 5, 6, 7", + "[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 6", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4", + "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5", + "[21] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In 2012 7th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 25-32. IEEE, 2012. 1, 2", + "[22] Dong Won Lee, Chaitanya Ahuja, and Louis-Philippe Morency. Crossmodal clustered contrastive learning: Grounding of spoken language to gesture. In GENEA: Generation and Evaluation of Non-verbal Behaviour for Embodied Agents Workshop 2021, 2021. 5", + "[23] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 6", + "[24] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 6", + "[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10552", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 5", + "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6", + "[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 612-630. Springer, 2022. 1", + "[28] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 2", + "[29] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. arXiv preprint arXiv:2212.02350, 2022. 2", + "[30] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10462-10472, June 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[31] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 25-35, 2013. 1, 2", + "[32] David McNeill. Hand and mind. De Gruyter Mouton, 2011. 1", + "[33] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2, 5", + "[34] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11077-11086, 2021. 1, 2", + "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[36] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2", + "[38] P. Wagner, Z. Malisz, and S. Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1", + "[39] Jing Xu, Wei Zhang, Yalong Bai, Qibin Sun, and Tao Mei. Freeform body motion generation from speech. arXiv preprint arXiv:2203.02291, 2022. 1", + "[40] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107. IEEE, 2022. 2", + "[41] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[42] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 5, 6, 7" + ], + "bbox": [ + 501, + 90, + 890, + 527 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10553", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_model.json b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a396c5d0f6aa8a46fc209e287cb84fdc56061e22 --- /dev/null +++ b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.18, + 0.816, + 0.235 + ], + "angle": 0, + "content": "Lingting Zhu\\(^{1*}\\) Xian Liu\\(^{2*}\\) Xuanyu Liu\\(^{1}\\) Rui Qian\\(^{2}\\) Ziwei Liu\\(^{3}\\) Lequan Yu\\(^{1\\dagger}\\) \n\\(^{1}\\)The University of Hong Kong \n\\(^{2}\\)The Chinese University of Hong Kong \n\\(^{3}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.237, + 0.724, + 0.27 + ], + "angle": 0, + "content": "{ltzhu99，u3008631}@connect.hku.hk，lqyu@hku.hk, {alvinliu，qr021}@ie.cuhk.edu.hk，ziwei.liu@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Animating virtual avatars to make co-speech gestures facilitates various applications in human-machine interaction. The existing methods mainly rely on generative adversarial networks (GANs), which typically suffer from notorious mode collapse and unstable training, thus making it difficult to learn accurate audio-gesture joint distributions. In this work, we propose a novel diffusion-based framework, named Diffusion Co-Speech Gesture (DiffGesture), to effectively capture the cross-modal audio-to-gesture associations and preserve temporal coherence for high-fidelity audio-driven co-speech gesture generation. Specifically, we first establish the diffusion-conditional generation process on clips of skeleton sequences and audio to enable the whole framework. Then, a novel Diffusion Audio-Gesture Transformer is devised to better attend to the information from multiple modalities and model the long-term temporal dependency. Moreover, to eliminate temporal inconsistency, we propose an effective Diffusion Gesture Stabilizer with an annealed noise sampling strategy. Benefiting from the architectural advantages of diffusion models, we further incorporate implicit classifier-free guidance to trade off between diversity and gesture quality. Extensive experiments demonstrate that DiffGesture achieves state-of-the-art performance, which renders coherent gestures with better mode coverage and stronger audio correlations. Code is available at https://github.com/Advocate99/DiffGesture." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.761, + 0.21, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Making co-speech gestures is an innate human behavior in daily conversations, which helps the speakers to express their thoughts and the listeners to comprehend the meanings [10, 32, 38]. Previous linguistic studies verify that such non-verbal behaviors could liven up the atmosphere" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.303, + 0.891, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.396, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Figure 1. Illustration of Conditional Generation Process in Co-Speech Gesture Generation. The diffusion process \\( q \\) gradually adds Gaussian noise to the gesture sequence (i.e., \\( \\pmb{x}_0 \\) sampled from the real data distribution). The generation process \\( p_{\\theta} \\) learns to denoise the white noise (i.e., \\( \\pmb{x}_T \\) sampled from the normal distribution) conditioned on context information \\( \\pmb{c} \\). Note that \\( \\pmb{x}_t \\) denotes the corrupted gesture sequence at the \\( t \\)-th diffusion step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.609 + ], + "angle": 0, + "content": "and improve mutual intimacy [7, 8, 21]. Therefore, animating virtual avatars to gesticulate co-speech movements is crucial in embodied AI. To this end, recent researches focus on the problem of audio-driven co-speech gesture generation [16, 25, 30, 41], which synthesizes human upper body gesture sequences that are aligned to the speech audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Early attempts downgrade this task as a searching-and-connecting problem, where they predefine the corresponding gestures of each speech unit and stitch them together by optimizing the transitions between consecutive motions for coherent results [11,21,31]. In recent years, the compelling performance of deep neural networks has prompted data-driven approaches. Previous studies establish large-scale speech-gesture corpus to learn the mapping from speech audio to human skeletons in an end-to-end manner [4,5,25, 27,30,34,39]. To attain more expressive results, Ginosar et al. [16] and Yoon et al. [41] propose GAN-based methods to guarantee realism by adversarial mechanism, where the discriminator is trained to distinguish real gestures from the synthetic ones while the generator's objective is to fool the discriminator. However, such pipelines suffer from the inherent mode collapse and unstable training, making them difficult to capture the high-fidelity audio-conditioned gesture distribution, resulting in dull or unreasonable poses." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The recent paradigm of diffusion probabilistic models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10544" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.365 + ], + "angle": 0, + "content": "provides a new perspective for realistic generation [19, 37], facilitating high-fidelity synthesis with desirable properties such as good distribution coverage and stable training compared to GANs. However, it is non-trivial to adapt existing diffusion models for co-speech gesture generation. Most existing conditional diffusion models deal with static data and conditions [35, 36] (e.g., the image-text pairs without temporal dimension), while co-speech gesture generation requires generating temporally coherent gesture sequences conditioned on continual audio clips. Further, the commonly used denoising strategy in existing diffusion models samples independently and identically distributed (i.i.d.) noises in latent space to increase diversity. However, this strategy tends to introduce variation for each gesture frame and lead to temporal inconsistency in skeleton sequences. Therefore, how to generate high-fidelity co-speech gestures with strong audio correlations and temporal consistency is quite challenging within the diffusion paradigm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.807 + ], + "angle": 0, + "content": "To address the above challenges, we propose a tailored Diffusion Co-Speech Gesture framework to capture the cross-modal audio-gesture associations while maintaining temporal coherence for high-fidelity audio-driven co-speech gesture generation, named DiffGesture. As shown in Figure 1, we formulate our task as a diffusion-conditional generation process on clips of skeleton and audio, where the diffusion phase is defined by gradually adding noise to gesture sequence, and the generation phase is referred as a parameterized Markov chain with conditional context features of audio clips to denoise the corrupted gestures. As we treat the multi-frame gesture clip as the diffusion latent space, the skeletons can be efficiently synthesized in a non-autoregressive manner to bypass error accumulation. To better attend to the sequential conditions from multiple modalities and enhance the temporal coherence, we then devise a novel Diffusion Audio-Gesture Transformer architecture to model audio-gesture long-term temporal dependency. Particularly, the per-frame skeleton and contextual features are concatenated in the aligned temporal dimension and embedded as individual input tokens to a Transformer block. Further, to eliminate the temporal inconsistency caused by the naive denoising strategy in the inference stage, we thus propose a new Diffusion Gesture Stabilizer module to gradually anneal down the noise discrepancy in the temporal dimension. Finally, we incorporate implicit classifier-free guidance by jointly training the conditional and unconditional models, which allows us to trade off between the diversity and sample quality during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Extensive experiments on two benchmark datasets show that our synthesized results are coherent with stronger audio correlations and outperform the state-of-the-arts with superior performance on co-speech gesture generation. To summarize, our main contributions are three-fold: 1) As an early attempt at taming diffusion models for co-speech" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.214 + ], + "angle": 0, + "content": "gesture generation, we formally define the diffusion and denoising process in gesture space, which synthesizes audio-aligned gestures of high-fidelity. 2) We devise the Diffusion Audio-Gesture Transformer with implicit classifier-free diffusion guidance to better deal with the input conditional information from multiple sequential modalities. 3) We propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.642, + 0.244 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Co-Speech Gesture Generation. Synthesizing co-speech gestures is crucial for a variety of applications. Conventional studies resort to rule-based pipelines [11, 21, 31], where linguistic experts pre-define the speech-gesture pairs and refine the transitions between different motions. Recent works exploit neural networks to learn the mapping from speech to gesture based on a large training corpus, where an off-the-shelf pose estimator is leveraged to label the online videos for pseudo annotations [1, 16, 30, 34, 41, 42]. Meanwhile, some works study the influence of input modality, verifying the connections between co-speech gesture and speech audio [25], text transcript [3], speaking style [2], and speaker identity [41]. To further improve the model's capacity, previous studies explore multiple architecture choices, including CNN [17], RNN [42], Transformer [6], and VQVAE [29, 40]. Notably, several recent works are based on GANs to guarantee realistic results [16, 30, 34, 41], which involve the adversarial training between the generator and the discriminator. However, the notorious mode collapse and unstable training of GANs prevent the high-fidelity gesture distribution learning conditioned on audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.896, + 0.816 + ], + "angle": 0, + "content": "Diffusion Probabilistic Models. Diffusion probabilistic models have achieved promising results on unconditional image generation [19], which are further applied to conditional tasks like text-to-image [36]. Among the diffusion-based literature, previous works focus on static data and conditions. Besides, they mainly utilize explicit guidance like pretrained classifiers [13] and CLIP similarity [28, 33, 35, 36] to guide the generation process. In this work, we explore a more challenging co-speech gesture generation setting, where the gesture data and audio conditions are both sequential, and the audio-to-gesture mapping is implicit. To this end, we propose the Diffusion Audio-Gesture Transformer to guarantee temporally aligned generation. We further propose the Diffusion Gesture Stabilizer to simultaneously achieve diverse and temporally coherent gestures with an annealed noise sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.649, + 0.848 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Figure 2 depicts an overview of the proposed DiffGesture framework to generate co-speech gestures of high fidelity. In this section, we first introduce the problem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "10545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.333, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.091, + 0.725, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.091, + 0.888, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Overview of the Diffusion Co-Speech Gesture (DiffGesture) Framework. Given the gesture sequence \\( \\pmb{x}_0 \\), we first establish the forward diffusion (purple) and conditional denoising process (green) in gesture space. Then, we devise the Diffusion Audio-Gesture Transformer to attend to the input conditions of initial poses \\( \\pmb{p}_{(1:M)} \\), speech audio \\( \\pmb{a} \\), time embedding \\( t \\) and corrupted gesture \\( \\pmb{x}_t \\) from multiple modalities (blue). At the diffusion sampling stage (grey), we propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy. To further incorporate implicit classifier-free guidance, we jointly train the conditional \\( (1 - p_{\\text{uncond}}) \\) and unconditional \\( (p_{\\text{uncond}}) \\) models. This allows us to trade off between diversity and quality during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.569 + ], + "angle": 0, + "content": "formulation of audio-driven co-speech gesture generation (Section 3.1). We then establish the forward diffusion and the reverse conditional generation process in gesture space (Section 3.2). Furthermore, we elaborate the Diffusion Audio-Gesture Transformer to attend to the conditions from multiple modalities and enhance the speech-gesture correlations with temporal dependency (Section 3.3). To eliminate temporal inconsistency introduced by naive noises, we propose a novel Diffusion Gesture Stabilizer with annealed noise sampling strategies and describe this module in (Section 3.4). Finally, incorporating implicit classifier-free guidance in co-speech gestures is discussed in (Section 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.28, + 0.602 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With a large-scale co-speech gesture training corpus, we leverage the speaking videos with clear co-speech upper body movements for model learning. In particular, for each video clip of \\(N\\) frames, we extract the accompanying speech audio sequence \\(\\pmb{a} = \\{\\pmb{a}_1,\\dots ,\\pmb{a}_N\\}\\) and use the offthe-shelf human pose estimator OpenPose [9] to annotate the per-frame human skeletons as \\(\\pmb{x} = \\{\\pmb{p}_1,\\dots ,\\pmb{p}_N\\}\\). We follow baseline methods [30, 41] to pre-process such skeletal representation as the concatenation of unit direction vectors as \\(\\pmb{p_i} = [\\pmb{d}_{i,1},\\pmb{d}_{i,2},\\dots ,\\pmb{d}_{i,J - 1}]\\), where \\(\\pmb{p_i}\\) denotes the pose description coordinates of the \\(i\\)-th frame, \\(J\\) is the total joint number and \\(\\pmb{d}_{i,j}\\) represents the \\(j\\)-th unit direction vector among the \\(J\\) joints of the \\(i\\)-th image frame. The diffusion model's reverse denoising process \\(G\\) parameterized by \\(\\theta\\) is optimized to synthesize the human skeleton sequence \\(\\pmb{x}\\), which is further conditioned on the speech audio sequence \\(\\pmb{a}\\) and the initial poses \\(\\{\\pmb{p}_1,\\dots ,\\pmb{p}_M\\}\\) of the first \\(M\\) frames. The learning objective of the overall framework can be formulated as \\(\\arg \\min_{\\theta}||\\pmb{x} - G_{\\theta}(\\pmb {a},\\pmb {p}_1,\\dots ,\\pmb {p}_M)||\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.879, + 0.403 + ], + "angle": 0, + "content": "3.2. Gesture Space Forward and Reverse Process" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.518 + ], + "angle": 0, + "content": "Given \\( \\pmb{x}_0 \\in \\mathbb{R}^{N \\times 3(J-1)} \\) sampled from real data distribution \\( q(\\pmb{x}_0) \\), our goal is to learn a model distribution \\( p_{\\theta}(\\pmb{x}_0) \\) parameterized by \\( \\theta \\) that approximates \\( q(\\pmb{x}_0) \\). Specifically, denoising diffusion probabilistic models (DDPMs) [19] define the latent variable models of the form \\( p_{\\theta}(\\pmb{x}_0) = \\int p_{\\theta}(\\pmb{x}_{0:T}) d\\pmb{x}_{1:T} \\), where \\( \\pmb{x}_{1:T} \\) are latent variables in the same sample space as \\( \\pmb{x}_0 \\) with the same dimensionality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.892, + 0.596 + ], + "angle": 0, + "content": "The Forward Diffusion Process. The forward process, which is also termed as the diffusion process, approximates the posterior distribution \\( q(\\pmb{x}_{1:T}|\\pmb{x}_0) \\). It is defined as a Markov chain that gradually adds Gaussian noise to the data sample according to a variance schedule \\( \\beta_1,\\dots ,\\beta_T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.61, + 0.891, + 0.651 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.655, + 0.891, + 0.673 + ], + "angle": 0, + "content": "\\[\nw h e r e \\quad q (\\boldsymbol {x} _ {t} | \\boldsymbol {x} _ {t - 1}) = \\mathcal {N} (\\boldsymbol {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\boldsymbol {x} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.793 + ], + "angle": 0, + "content": "The variances \\(\\beta_{t}\\) are constant hyperparameters to ease the modeling of the reverse process [19]. Through such a corruption scheme, the structural information of the original skeleton is gradually substituted by noises, which finally leads to a pure white noise when \\(T\\) goes to infinity. Therefore, the prior latent distribution of \\(p(\\boldsymbol{x}_T)\\) is \\(\\mathcal{N}(\\boldsymbol{x}_T; \\mathbf{0}, \\boldsymbol{I})\\) with only information of Gaussian noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Reverse Conditional Gesture Generation. The reverse process, which is also termed as the generative process, estimates the joint distribution of \\( p_{\\theta}(\\pmb{x}_{0:T}) \\). As proved in [15], the reverse process of the continuous diffusion process preserves the same transition distribution form, which motivates us to leverage a Gaussian transition to formulate \\( p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t) \\) under an unconditional setting, which" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10546" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.343, + 0.108 + ], + "angle": 0, + "content": "approximates the intractable process as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.119, + 0.47, + 0.16 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {x} _ {0: T}\\right) = p _ {\\theta} \\left(\\boldsymbol {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.163, + 0.469, + 0.181 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(\\boldsymbol{x}_{t - 1}|\\boldsymbol{x}_t) = \\mathcal{N}(\\boldsymbol{x}_{t - 1};\\mu_{\\theta}(\\boldsymbol{x}_t,t),\\Sigma_{\\theta}(\\boldsymbol{x}_t,t))\\) (4)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.192, + 0.473, + 0.375 + ], + "angle": 0, + "content": "The corrupted noisy data \\( \\pmb{x}_t \\) is sampled by \\( q(\\pmb{x}_t|\\pmb{x}_0) = \\mathcal{N}(\\pmb{x}_t; \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0, (1 - \\bar{\\alpha}_t)\\pmb{I}) \\), where \\( \\alpha_t = 1 - \\beta_t \\) and \\( \\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s \\). Note that we set the variances \\( \\Sigma_\\theta(\\pmb{x}_t, t) = \\beta_t\\pmb{I} \\) to untrained time-dependent constants. The above diffusion model formulations show compelling performances on unconditional generation. To further adapt to the conditional co-gesture synthesis, we have to provide additional inputs to the model, including the audio and initial poses. Therefore, we treat the speech audio \\( \\pmb{a} \\) and initial poses \\( \\pmb{p}_{1:M} \\) as context information \\( \\pmb{c} \\) and inject conditions into the generation process. The reverse process of each timestep (Eq. 4) can be thus updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.386, + 0.47, + 0.403 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}, \\boldsymbol {c}) = \\mathcal {N} (\\boldsymbol {x} _ {t - 1}; \\mu_ {\\theta} (\\boldsymbol {x} _ {t}, t, \\boldsymbol {c}), \\beta_ {t} \\boldsymbol {I}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.49 + ], + "angle": 0, + "content": "In this way, we could start the generation process by firstly sample a Gaussian noise \\( \\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\boldsymbol{I}) \\) and follow the Markov chain to iteratively denoise the latent variable \\( \\boldsymbol{x}_t \\) via Eq. 5 to get the final results. The overview of conditional co-speech gesture process is illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.49, + 0.47, + 0.63 + ], + "angle": 0, + "content": "Training Objective. To optimize the overall framework, we optimize the variational lower bound on negative log-likelihood: \\(\\mathbb{E}[-\\log p_{\\theta}(\\pmb{x}_0)] \\leq \\mathbb{E}_q[-\\log \\frac{p_{\\theta}(\\pmb{x}_0)}{q(\\pmb{x}_{1:T}|\\pmb{x}_0)}]\\). We rewrite the loss function conditioned on context \\(c\\) and eliminate all the constant items that do not require training: \\(L(\\theta) = \\mathbb{E}_q[\\sum_{t=2}^{T} D_{KL}(q(\\pmb{x}_{t-1}|\\pmb{x}_t, \\pmb{x}_0)||p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t, c))]\\). With reparameterization, we can represent each term in \\(L_{\\theta}\\) using MSE loss. We follow [19] to further simplify the training objective to the ensemble of MSE losses as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.654, + 0.47, + 0.677 + ], + "angle": 0, + "content": "\\[\nL (\\theta) = \\mathbb {E} _ {q} [ \\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\boldsymbol {c}, t) \\right\\| ^ {2} ], (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.777 + ], + "angle": 0, + "content": "where \\( t \\) is uniformly sampled between 1 and \\( T \\). As we jointly train the model under conditional and unconditional setting, a trainable masked embedding with probability \\( p_{\\text{uncond}} \\) replaces context \\( c \\) and the diffusion model predicts the noise in the unconditional setting. The detailed principles will be discussed in Section 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.405, + 0.803 + ], + "angle": 0, + "content": "3.3. Diffusion Audio-Gesture Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "With the naive conditional generation scheme as specified in Section 3.2, we still confront a critical problem in the setting of co-speech gesture generation. Since \\( x_0 \\) denotes the skeleton sequence of \\( N \\) frames, there exists temporal dependency among the target sequence and context information, making it more complex than time-invariant tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "like image generation. Therefore, how to guarantee temporally coherent results in a non-autoregressive conditional generation process remains an unsolved problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.892, + 0.275 + ], + "angle": 0, + "content": "In contrast to most previous studies that resort to recurrent networks [30, 41], we propose to make use of the Transformer's strong capacity in sequential data modeling. Specifically, since the noisy gesture sequence \\( \\boldsymbol{x}_t \\) and the contextual information \\( c \\) align in the temporal dimension, we concatenate them in the feature channel. In this way, the skeleton and context condition of each frame serve as an individual token, which captures the long-term dependency by the self-attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.291, + 0.892, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\mathsf {T}}}{\\sqrt {\\ell}}\\right) \\mathbf {V}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.416 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q},\\mathbf{K},\\mathbf{V}\\) are the query, key, and value matrix from input tokens, \\(\\ell\\) is the channel dimension, and \\(\\mathsf{T}\\) is the matrix transpose operation. Such a design also avoids severe error accumulation in autoregressive pipelines, enabling us to generate coherent gesture sequences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.75, + 0.448 + ], + "angle": 0, + "content": "3.4. Diffusion Gesture Stabilizer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.639 + ], + "angle": 0, + "content": "In DDPMs, the independent random variables \\( z \\) introduced at the sampling stage promote diversity and thus systematically improve the task performance. However, the variation in temporal dimension introduced by \\( z \\), especially when timestep \\( t \\) is small in the reverse process, can have a negative effect on temporal consistency. At the inference stage, to achieve the trade-off between diversity and temporal consistency, we propose a novel Diffusion Gesture Stabilizer without extra training expenses under two annealed scenarios, where the term \"annealed\" means that the process is transitioned from high variance and entropy (hot) to low variance and entropy (cold)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Thresholding. Since temporally independent Gaussian noises inevitably introduce inconsistency, restricting the temporal variation naturally helps to avoid inconsistency. And hard thresholding serves as an effective trick. In detail, we set a time threshold \\( t_0 \\), and then use the same \\( z \\in \\mathbb{R}^{N \\times C} \\) in the naive sampling strategy for \\( t > t_0 \\) and set \\( z = \\{z_0\\}_{i=1}^N \\) for \\( t \\leq t_0 \\), where \\( z_0 \\in \\mathbb{R}^C \\) follows \\( \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\) which do not introduce variation in the temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Smooth Sampling. We further modify \\( z(t) = \\{z_i(t)\\}_{i=1}^N \\) to be a smooth annealing version via variance-aware sampling. In the original sampling rule of DDPMs, i.i.d.variables \\( z_i(t) \\) are sampled from \\( \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\). With smooth resampling, we first sample \\( z_0(t) \\sim \\mathcal{N}(\\mathbf{0},\\sigma_a^2(t)\\mathbf{I}) \\) only once for each timestep \\( t \\) in the reverse process, then given \\( z_0(t) \\), we sample \\( z_i(t)|z_0(t) \\sim \\mathcal{N}(z_0(t),(1 - \\sigma_a^2(t))\\mathbf{I}) \\) for \\( i \\in \\{1,\\dots,N\\} \\), where \\( \\sigma_a(t) \\in [0,1] \\) is a nondecreasing function to achieve variance annealing." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10547" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.092, + 0.232, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Training" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.111, + 0.147, + 0.123 + ], + "angle": 0, + "content": "1: repeat" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.125, + 0.282, + 0.137 + ], + "angle": 0, + "content": "2: Sample \\((\\pmb{x}_0,\\pmb{c})\\sim q(\\pmb{x}_0,\\pmb{c})\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.138, + 0.437, + 0.151 + ], + "angle": 0, + "content": "3: Sample \\(\\tau \\sim \\mathrm{Uniform}(0,1)\\). Set \\(c = \\varnothing\\) if \\(\\tau < p_{\\text{uncond}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.152, + 0.327, + 0.165 + ], + "angle": 0, + "content": "4: Sample \\(t \\sim \\mathrm{Uniform}(\\{1, \\dots, T\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.166, + 0.426, + 0.182 + ], + "angle": 0, + "content": "5: Compute \\(\\nabla_{\\theta}\\left\\| \\pmb {\\epsilon} - \\pmb{\\epsilon}_{\\theta}(\\sqrt{\\bar{\\alpha}_{t}}\\pmb {x}_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\pmb {\\epsilon},\\pmb {c},t)\\right\\|^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.182, + 0.269, + 0.194 + ], + "angle": 0, + "content": "6: Perform gradient descent" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.196, + 0.2, + 0.208 + ], + "angle": 0, + "content": "7: until converged" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.111, + 0.437, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.236, + 0.364, + 0.252 + ], + "angle": 0, + "content": "3.5. Implicit Classifier-free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.261, + 0.473, + 0.564 + ], + "angle": 0, + "content": "In co-speech gesture literature, the speech-to-gesture mapping is implicit, where the same audio corresponds to diverse gestures and different audiOs could incur the same motion [22, 25], making it difficult to utilize the commonly used explicit classifier guidance [13, 33]. Therefore, how can we further exploit practical guidance for better audio correlations and mode coverage? Our solution to this question is to train an extra \"unconditional\" diffusion model to implicitly guide the generation. Dhariwal et al. [13] first introduce the classifier guidance of cross-entropy gradient \\(\\nabla_{\\pmb{x}_t}\\log p_\\phi (y|\\pmb {x}_t)\\) , where the pretrained classifier is parameterized by \\(\\phi\\) and \\(y\\) denotes the classification logits. This gradient term is further scaled by the covariance matrix to modify the mean value of transition distribution in Eq. 4. To adapt to the cases where no explicit guidance is available, we follow [20] to jointly train the conditional and unconditional models, termed as classifier-free guidance. In particular, according to the implicit classifier's property that \\(p(c|x_t)\\propto p(x_t|c) / p(x_t)\\) , we could derive a gradient relationship in the implicit classifier as:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {c} | \\boldsymbol {x} _ {t}) \\propto \\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {x} _ {t} | \\boldsymbol {c}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {x} _ {t}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.61, + 0.47, + 0.716 + ], + "angle": 0, + "content": "which is further proportional to \\(\\epsilon^{*}(\\boldsymbol{x}_{t}|\\boldsymbol{c}) - \\epsilon^{*}(\\boldsymbol{x}_{t})\\). Therefore, we use a single Transformer network to parameterize both settings by a mix-up training trick: for the probability of \\(p_{uncond}\\), we set the context information \\(\\boldsymbol{c}\\) as masked embedding to train the unconditional setting, while for other cases, we train the original conditional counterpart. The training is shown in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Sampling with Classifier-free Guidance. Starting from Gaussian noise, we iteratively remove noises in \\( x_{t} \\). As we use implicit classifier-free guidance, similar to Equation 8, the predicted Gaussian noise is modified as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.794, + 0.469, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\epsilon}} _ {\\theta} = \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) + s \\cdot \\left(\\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( s \\) is the scale parameter to trade off the diversity and quality. With classifier-free guidance, Algorithm 2 reveals how to generate co-speech gestures given the trained diffusion model via the Diffusion Gesture Stabilizer with the Smooth Sampling annealed scenario." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.092, + 0.662, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Sampling" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.111, + 0.78, + 0.125 + ], + "angle": 0, + "content": "1: Trained diffusion model \\(\\theta, \\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.126, + 0.649, + 0.138 + ], + "angle": 0, + "content": "2: for \\( t = T, \\dots, 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.139, + 0.819, + 0.153 + ], + "angle": 0, + "content": "3: \\(\\hat{\\pmb{\\epsilon}}_{\\theta} = \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) + s\\cdot (\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,\\pmb {c},t) - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t))\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.153, + 0.68, + 0.167 + ], + "angle": 0, + "content": "4: \\( \\mathbf{z}_0(t) \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_a^2(t)\\mathbf{I}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.167, + 0.666, + 0.18 + ], + "angle": 0, + "content": "5: for \\(i = 1,\\dots ,N\\) do" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.181, + 0.754, + 0.195 + ], + "angle": 0, + "content": "6: \\( \\pmb{z}_i(t) \\sim \\mathcal{N}(\\pmb{z}_0(t), (1 - \\sigma_a^2(t))\\pmb{I}) \\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.196, + 0.589, + 0.207 + ], + "angle": 0, + "content": "7: end for" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.208, + 0.85, + 0.222 + ], + "angle": 0, + "content": "8: \\( \\mathbf{z}(t) = \\{\\mathbf{z}_1(t),\\dots ,\\mathbf{z}_N(t)\\} \\), if \\( t > 1 \\), else \\( \\mathbf{z}(t) = \\mathbf{0} \\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.223, + 0.794, + 0.243 + ], + "angle": 0, + "content": "9: \\( \\pmb{x}_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}} \\left( \\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}} \\hat{\\pmb{\\epsilon}}_\\theta \\right) + \\sigma_t \\pmb{z}(t) \\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.243, + 0.575, + 0.254 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.256, + 0.59, + 0.268 + ], + "angle": 0, + "content": "11: return \\( x_0 \\)" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.111, + 0.85, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.292, + 0.633, + 0.309 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.753, + 0.333 + ], + "angle": 0, + "content": "4.1. Co-Speech Gesture Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.43 + ], + "angle": 0, + "content": "TED Gesture. As a large-scale dataset for gesture generation research, TED Gesture dataset [41, 42] contains 1,766 TED videos of different narrators covering various topics. We follow the data process in former works [30, 41], where the poses are resampled with 15 FPS, and frame segments of length 34 are obtained with a stride of 10." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.552 + ], + "angle": 0, + "content": "TED Expressive. While the poses in TED Gesture only contain 10 upper body key points without vivid finger movements, the TED Expressive dataset [30] is further expressive of both finger and body movements. The state-of-art 3D pose estimator ExPose [12] is used to fully capture the pose information in data. As a result, TED Expressive annotates the 3D coordinates of 43 keypoints, including 13 upper body joints and 30 finger joints." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.707, + 0.577 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Comparison Methods. We compare our method on two benchmark datasets with the state-of-the-art methods in recent years. 1) Attention Seq2Seq [42] elaborates on the attention mechanism to generate pose sequences from speech text. 2) Speech2Gesture [16] uses spectrums of the speech audio segments as the input and generates speech gestures adversarially. 3) Joint Embedding [3] maps text and motion to the same embedding space, then generates outputs from motion description text. 4) Trimodal [41] serves as a strong baseline that learns from text, audio, and speaker identity to generate gestures, outperforming former methods by a large margin. 5) HA2G [30] introduces a hierarchical audio learner that captures information across different semantic granularities, achieving state-of-the-art performances. This method hierarchically extracts rich features at the cost of heavier GPU memory overhead, while our method requires much smaller expenses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For all the methods in both datasets, we set \\( N = 34 \\) and \\( M = 4 \\) to get \\( M \\)-frame pose sequences where the first \\( N \\) frames are used for reference, termed as initial poses. There are \\( J \\) upper body joints in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10548" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.184, + 0.089, + 0.788, + 0.26 + ], + "angle": 0, + "content": "
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
Ground Truth00.698108.52500.703178.827
Attention Seq2Seq [42]18.1540.19682.77654.9200.152122.693
Speech2Gesture [16]19.2540.66893.80254.6500.679142.489
Joint Embedding [3]22.0830.20090.13864.5550.130120.627
Trimodal [41]3.7290.667101.24712.6130.563154.088
HA2G [30]3.0720.672104.3225.3060.641173.899
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.271, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Table 1. The Quantitative Results on TED Gesture [41] and TED Expressive [30]. We compare the proposed diffusion-based method against recent SOTA methods [3, 16, 30, 41, 42] and ground truth. For FGD, the lower, the better; for other metrics, the higher, the better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.473, + 0.673 + ], + "angle": 0, + "content": "all the frames of pose sequences, where \\( J = 10 \\) for TED Gesture and \\( J = 43 \\) for TED Expressive. Following [41], to eliminate the effect of the joint lengths and root motion, we represent the joints' positions using \\( J - 1 \\) directional vectors normalized to the unit vectors and train the model to learn the directional vectors. For the audio processing, we use the same audio encoder used in [41] to extract the feature of the raw audio clips directly. The audio clips are encoded as \\( N \\) audio feature vectors of 32-D. The audio feature and initial poses are concatenated to form the conditional context information of the diffusion model. For the diffusion process, the number of timesteps is \\( T = 500 \\), and the variances increase linearly from \\( \\beta_{1} = 1e - 4 \\) to \\( \\beta_{T} = 0.02 \\). For the Stabilizer, \\( t_0 \\) can be adjusted from 20-30 for Thresholding, and a quadratic non-increasing function \\( \\sigma_{a}(t) \\) is applied for Smooth Sampling. The hidden dimension of the transformer blocks, is set to 256 for TED Gesture and 512 for TED Expressive. We use 8 Transformer blocks, each of which comprises a multi-head self-attention block and a Feed-Forward Network. We use an Adam optimizer, and the learning rate is \\( 5e - 4 \\). It takes 10 hours to train the model on TED Gesture and 20 hours on TED Expressive on a single NVIDIA GeForce RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.262, + 0.695 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.734 + ], + "angle": 0, + "content": "In evaluation, we use three metrics that are used in cospeech gesture generation and relative fields [24, 30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Fréchet Gesture Distance (FGD). Similar to the Fréchet Inception Distance (FID) metric [18], which is widely applied in image generation studies, FGD is used to measure the distance between the synthesized gesture distribution and the real data distribution. Yoon et al. [41] define FGD by training a skeleton sequence auto-encoder to extract the features of the real gesture sequences \\( X \\) and the features of the generated gesture sequences \\( \\hat{X} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.861, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F G D} (X, \\hat {X}) = \\| \\mu_ {r} - \\mu_ {g} \\| ^ {2} + \\operatorname {T r} (\\Sigma_ {\\mathrm {r}} + \\Sigma_ {\\mathrm {g}} - 2 (\\Sigma_ {\\mathrm {r}} \\Sigma_ {\\mathrm {g}}) ^ {1 / 2}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mu_r\\) and \\(\\Sigma_r\\) are the first and the second moments of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.403 + ], + "angle": 0, + "content": "latent feature distribution of the real gestures \\( X \\), and \\( \\mu_g \\) and \\( \\Sigma_g \\) are the first and the second moments of the latent feature distribution of the generated gestures \\( \\hat{X} \\). We intuitively find that among the three metrics, FGD tells the most whether the generated pose sequences are of high quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.893, + 0.692 + ], + "angle": 0, + "content": "Beat Consistency Score (BC). Proposed in [24, 26], BC measures motion-audio beat correlation. Considering that the kinematic velocities vary from different joints, we use the change of included angle between bones to track motion beats following [30]. Specifically, we can calculate the mean absolute angle change (MAAC) of angle \\(\\theta_{j}\\) in adjacent frames by \\(\\mathrm{MAAC}(\\theta_j) = \\frac{1}{S}\\frac{1}{T - 1}\\sum_{s = 1}^{S}\\sum_{t = 1}^{T - 1}\\| \\theta_{j,s,t + 1} - \\theta_{j,s,t}\\| _1\\), where \\(S\\) denotes the total number of clips in the dataset, \\(T\\) denotes the number of frames in each clip, and \\(\\theta_{j,s,t}\\) is the included angle between the \\(j\\)-th and the \\((j + 1)\\)-th bone of the \\(s\\)-th clip at time-step \\(t\\). Then, we can compute the angle change rate of frame \\(t\\) for the \\(s\\)-th clip as \\(\\frac{1}{J - 1}\\sum_{j = 1}^{J - 1}(\\| \\theta_{j,s,t + 1} - \\theta_{j,s,t}\\| _1 / \\mathrm{MAAC}(\\theta_j))\\). Then we extract the local optima whose first-order difference is higher than a threshold to get kinematic beats, which are used to compute BC later. Following [24] to detect audio beat by onset strength [14], we compute the average distance between each audio beat and its nearest motion beat as Beat Consistency Score:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.7, + 0.892, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathrm {B C} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\exp \\left(- \\frac {\\min _ {\\forall t _ {j} ^ {y} \\in B ^ {y}} \\| t _ {i} ^ {x} - t _ {j} ^ {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.796 + ], + "angle": 0, + "content": "where \\( t_i^x \\) is the \\( i \\)-th audio beats, \\( B^y = \\{t_i^y\\} \\) is the set of the kinematic beats, and \\( \\sigma \\) is a parameter to normalize sequences, set to 0.1 empirically." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Diversity. This metric evaluates the variations among generated gestures corresponding to various inputs [23]. We use the same feature extractor when measuring FGD to map synthesized gestures into latent feature vectors and calculate the mean feature distance. In detail, we randomly pick 500 generated samples and compute the mean absolute error between the features and the shuffled features." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10549" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.477, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 3. Visualization Results of Our DiffGesture on Two Datasets. Three cases are picked up, where (i) and (ii) are TED Expressive cases, and (iii) is a TED Gesture case. We highlight dull cases generated by comparison methods with rectangles, indicating the mode collapse phenomenon of baselines." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.532, + 0.842, + 0.611 + ], + "angle": 0, + "content": "
MethodsGTS2S. [42]S2G. [16]Joint. [3]Tri. [41]HA2G [30]DiffGesture(Ours)
Naturalness4.331.222.561.223.223.674.00
Smoothness3.943.501.613.443.443.393.89
Synchrony4.001.673.171.393.283.443.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.621, + 0.888, + 0.636 + ], + "angle": 0, + "content": "Table 2. User Study Results. The ratings of motion naturalness, smoothness, and synchrony, are on a scale of 1-5, with 5 being the best." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.661, + 0.258, + 0.675 + ], + "angle": 0, + "content": "4.4. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Quantitative Results. We compare our method with all the baselines with three metrics on TED Gesture and TED Expressive. The results are shown in Table 1. For the metrics of Ground Truth, we report the values in our implementation. For TED Gesture, we report FGD of all baselines in [30] and evaluate BC and Diversity on our own†. For TED Expressive, all the results of baselines are reported from [30]. Assuming the pseudo ground truth pose follows the real distribution, the FGD of Ground Truth in the table is 0. It is observed that our DiffGesture achieves state-of-the-art performance on both datasets, especially outperforming existing methods by a large margin on TED" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Since there exists an evaluation bug for the BC metric in HA2G [30], we report the re-implemented results from Liu et al." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.89 + ], + "angle": 0, + "content": "Expressive. Besides, since BC and Diversity are proposed to measure motion-audio beat correlation and variation, these two metrics of Ground Truth cannot be treated as upper bounds, and it is worth noting that our results may be higher than the ones of Ground Truth, indicating that the generated gestures are of high quality. Qualitative Results. We show the keyframes of all the methods on two datasets in Figure 3. Since TED Expressive requires a higher ability of generative models, we pick up two cases for TED Expressive and one for TED Gesture. For each case, we select three keyframes (an early, a middle, and a late one) to show the pose motions. Comparison methods tend to generate slow and invariant poses and sometimes produce unreliable and stiff results. In contrast, DiffGesture produces diverse human-like poses without resulting in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10550" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.808, + 0.201 + ], + "angle": 0, + "content": "
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
DiffGesture Base2.4500.632104.6883.8220.707174.377
DiffGesture w/o Stabilizer2.2190.674105.1922.7920.721180.125
DiffGesture w/o classifier-free1.8100.673105.6443.3260.717178.245
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.212, + 0.895, + 0.242 + ], + "angle": 0, + "content": "Table 3. Ablation Study on the Proposed Modules. We investigate effectiveness of the proposed modules, Diffusion Gesture Stabilizer and implicit classifier-free guidance. The results indicate that the proposed modules consistently improve performance on the benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.359 + ], + "angle": 0, + "content": "mean poses which are slow and rigid. Besides, as a primary drawback of GAN-based methods, mode collapse makes comparison methods often produce a single type of output, which is severe for pose motion generation. Such a phenomenon is shown in Fig. 3, where the generated frames with nearly the same pose are highlighted with rectangles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.471, + 0.603 + ], + "angle": 0, + "content": "User Study. To better validate the qualitative performance, we conduct a user study on the generated co-speech gestures. The study involves 18 participants, with 9 females and 9 males in the age range of 18-25 years old. The participants are required to grade the motion's quality and coherence, and all the clips are without labels. In total, we pick up 30 cases, 20 for TED-Expressive and 10 for TED Gesture. For each case, we show 7 videos with the order of the methods shuffling, including the ground truth. We adopt the Mean Opinion Scores rating protocol, and each participant is required to rate three aspects of generated motions: Naturalness; Smoothness; Synchrony between speech and generated gestures. The results are shown in Table 2 where the ratings are on a scale of 1 to 5, with 5 being the best. Our participants widely accept that our method produces high-fidelity results in all three aspects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.242, + 0.632 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Ablation Study on the Proposed Modules. To demonstrate the effectiveness of our proposed DiffGesture, we present ablation studies on the key modules in the framework. In detail, we conduct experiments as follows. 1) DiffGesture Base means we only use the proposed conditional diffusion generation process without further design. 2) DiffGesture w/o classifier-free, where classifier-free guidance is not implemented at both the training stage and inference stage. 3) DiffGesture w/o Stabilizer, where Diffusion Gesture Stabilizer is removed at the inference stage. The results are reported in Table 3. The results illustrate the effectiveness of the designed Diffusion Gesture Stabilizer and the implicit classifier-free guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Ablation Study on the Network Architectures. We investigate the performance of the GRU architecture in diffusion models, autoregressively generating poses in [30, 41]. We replace the Diffusion Audio-Gesture Transformer with the" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.265, + 0.88, + 0.364 + ], + "angle": 0, + "content": "
MethodsFGD ↓BC ↑Diversity ↑
GRU on Da14.3430.65898.472
Transformer on Da1.5060.699106.722
GRU on Db17.4520.680172.168
Transformer on Db2.6000.718182.757
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Table 4. Ablation Study on the Network Architectures. We compare the performance of GRU and Transformer for diffusion-based backbone on TED Gesture \\((D_{a})\\) and TED Expressive \\((D_{b})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.577 + ], + "angle": 0, + "content": "GRU in the diffusion model. All the context inputs remain the same as our method and are concatenated before the diffusion network. Results are shown in Table 4. Though GRU serves as a strong baseline network in co-gesture learning, it fails to generate high-performance data like our designed Transformer-based network, which indicates the effectiveness of our Transformer-based network and that applying diffusion models in the audio-driven conditional generation is a non-trivial task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.619, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In this work, we present a novel diffusion-based framework DiffGesture for co-speech gesture generation. To generate coherent gestures with strong audio correlations, we propose the Diffusion Audio-Gesture Transformer with the Diffusion Gesture Stabilizer to better attend to the condition information. Such a non-autoregressive pipeline helps to efficiently generate results and reduce error accumulation. We hope our method offers a new perspective for diffusion-based temporal generation and how to capture sequential cross-modal dependencies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. The work described in this paper was partially supported by grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (T45-401/22-N), the National Natural Science Fund (62201483), HKU Seed Fund for Basic Research (202009185079 and 202111159073), RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10551" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1884–1895, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, volume 39, pages 487-496. Wiley Online Library, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Judee K Burgoon, Thomas Birk, and Michael Pfau. Nonverbal behaviors, persuasion, and credibility. Human communication research, 17(1):140-169, 1990. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] B. Butterworth and U. Hadar. Gesture, speech, and computational stages: A reply to mcneill. Psychological Review, 96(1):168-174, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.622, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[9] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.468, + 0.747 + ], + "angle": 0, + "content": "[10] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[11] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Daniel Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36:51-60, 03 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] William Feller. On the theory of stochastic processes, with particular reference to applications. In *Berkeley Symposium on Mathematical Statistics and Probability*, pages 403-432, 1949. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[16] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In 2012 7th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 25-32. IEEE, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[22] Dong Won Lee, Chaitanya Ahuja, and Louis-Philippe Morency. Crossmodal clustered contrastive learning: Grounding of spoken language to gesture. In GENEA: Generation and Evaluation of Non-verbal Behaviour for Embodied Agents Workshop 2021, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[24] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10552" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 612-630. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[28] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.471, + 0.432 + ], + "angle": 0, + "content": "[29] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. arXiv preprint arXiv:2212.02350, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[30] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10462-10472, June 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[31] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 25-35, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[32] David McNeill. Hand and mind. De Gruyter Mouton, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[33] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[34] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11077-11086, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[36] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[38] P. Wagner, Z. Malisz, and S. Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[39] Jing Xu, Wei Zhang, Yalong Bai, Qibin Sun, and Tao Mei. Freeform body motion generation from speech. arXiv preprint arXiv:2203.02291, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[40] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[41] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[42] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10553" + } + ] +] \ No newline at end of file diff --git a/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_origin.pdf b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..945d5a7f2d9caa0c331b62f73898d6e8297d8dcf --- /dev/null +++ b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/436e7035-a0f8-4cb2-9a25-a351d003ba3d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2bcf1f85d55069fbf1e74c3cabd399ccd4371a9f31700143ca6068a49c904fe +size 1187027 diff --git a/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/full.md b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..323a275f610049528f9a0d4d10b31e37e8719902 --- /dev/null +++ b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/full.md @@ -0,0 +1,294 @@ +# Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation + +Lingting Zhu $^{1*}$ Xian Liu $^{2*}$ Xuanyu Liu $^{1}$ Rui Qian $^{2}$ Ziwei Liu $^{3}$ Lequan Yu $^{1\dagger}$ $^{1}$ The University of Hong Kong + $^{2}$ The Chinese University of Hong Kong + $^{3}$ S-Lab, Nanyang Technological University + +{ltzhu99，u3008631}@connect.hku.hk，lqyu@hku.hk, {alvinliu，qr021}@ie.cuhk.edu.hk，ziwei.liu@ntu.edu.sg + +# Abstract + +Animating virtual avatars to make co-speech gestures facilitates various applications in human-machine interaction. The existing methods mainly rely on generative adversarial networks (GANs), which typically suffer from notorious mode collapse and unstable training, thus making it difficult to learn accurate audio-gesture joint distributions. In this work, we propose a novel diffusion-based framework, named Diffusion Co-Speech Gesture (DiffGesture), to effectively capture the cross-modal audio-to-gesture associations and preserve temporal coherence for high-fidelity audio-driven co-speech gesture generation. Specifically, we first establish the diffusion-conditional generation process on clips of skeleton sequences and audio to enable the whole framework. Then, a novel Diffusion Audio-Gesture Transformer is devised to better attend to the information from multiple modalities and model the long-term temporal dependency. Moreover, to eliminate temporal inconsistency, we propose an effective Diffusion Gesture Stabilizer with an annealed noise sampling strategy. Benefiting from the architectural advantages of diffusion models, we further incorporate implicit classifier-free guidance to trade off between diversity and gesture quality. Extensive experiments demonstrate that DiffGesture achieves state-of-the-art performance, which renders coherent gestures with better mode coverage and stronger audio correlations. Code is available at https://github.com/Advocate99/DiffGesture. + +# 1. Introduction + +Making co-speech gestures is an innate human behavior in daily conversations, which helps the speakers to express their thoughts and the listeners to comprehend the meanings [10, 32, 38]. Previous linguistic studies verify that such non-verbal behaviors could liven up the atmosphere + +![](images/2400d7bf94a445e8333befea73469a2373b520a6be01c3f856550347bb91652e.jpg) +Figure 1. Illustration of Conditional Generation Process in Co-Speech Gesture Generation. The diffusion process $q$ gradually adds Gaussian noise to the gesture sequence (i.e., $\pmb{x}_0$ sampled from the real data distribution). The generation process $p_{\theta}$ learns to denoise the white noise (i.e., $\pmb{x}_T$ sampled from the normal distribution) conditioned on context information $\pmb{c}$ . Note that $\pmb{x}_t$ denotes the corrupted gesture sequence at the $t$ -th diffusion step. + +and improve mutual intimacy [7, 8, 21]. Therefore, animating virtual avatars to gesticulate co-speech movements is crucial in embodied AI. To this end, recent researches focus on the problem of audio-driven co-speech gesture generation [16, 25, 30, 41], which synthesizes human upper body gesture sequences that are aligned to the speech audio. + +Early attempts downgrade this task as a searching-and-connecting problem, where they predefine the corresponding gestures of each speech unit and stitch them together by optimizing the transitions between consecutive motions for coherent results [11,21,31]. In recent years, the compelling performance of deep neural networks has prompted data-driven approaches. Previous studies establish large-scale speech-gesture corpus to learn the mapping from speech audio to human skeletons in an end-to-end manner [4,5,25, 27,30,34,39]. To attain more expressive results, Ginosar et al. [16] and Yoon et al. [41] propose GAN-based methods to guarantee realism by adversarial mechanism, where the discriminator is trained to distinguish real gestures from the synthetic ones while the generator's objective is to fool the discriminator. However, such pipelines suffer from the inherent mode collapse and unstable training, making them difficult to capture the high-fidelity audio-conditioned gesture distribution, resulting in dull or unreasonable poses. + +The recent paradigm of diffusion probabilistic models + +provides a new perspective for realistic generation [19, 37], facilitating high-fidelity synthesis with desirable properties such as good distribution coverage and stable training compared to GANs. However, it is non-trivial to adapt existing diffusion models for co-speech gesture generation. Most existing conditional diffusion models deal with static data and conditions [35, 36] (e.g., the image-text pairs without temporal dimension), while co-speech gesture generation requires generating temporally coherent gesture sequences conditioned on continual audio clips. Further, the commonly used denoising strategy in existing diffusion models samples independently and identically distributed (i.i.d.) noises in latent space to increase diversity. However, this strategy tends to introduce variation for each gesture frame and lead to temporal inconsistency in skeleton sequences. Therefore, how to generate high-fidelity co-speech gestures with strong audio correlations and temporal consistency is quite challenging within the diffusion paradigm. + +To address the above challenges, we propose a tailored Diffusion Co-Speech Gesture framework to capture the cross-modal audio-gesture associations while maintaining temporal coherence for high-fidelity audio-driven co-speech gesture generation, named DiffGesture. As shown in Figure 1, we formulate our task as a diffusion-conditional generation process on clips of skeleton and audio, where the diffusion phase is defined by gradually adding noise to gesture sequence, and the generation phase is referred as a parameterized Markov chain with conditional context features of audio clips to denoise the corrupted gestures. As we treat the multi-frame gesture clip as the diffusion latent space, the skeletons can be efficiently synthesized in a non-autoregressive manner to bypass error accumulation. To better attend to the sequential conditions from multiple modalities and enhance the temporal coherence, we then devise a novel Diffusion Audio-Gesture Transformer architecture to model audio-gesture long-term temporal dependency. Particularly, the per-frame skeleton and contextual features are concatenated in the aligned temporal dimension and embedded as individual input tokens to a Transformer block. Further, to eliminate the temporal inconsistency caused by the naive denoising strategy in the inference stage, we thus propose a new Diffusion Gesture Stabilizer module to gradually anneal down the noise discrepancy in the temporal dimension. Finally, we incorporate implicit classifier-free guidance by jointly training the conditional and unconditional models, which allows us to trade off between the diversity and sample quality during inference. + +Extensive experiments on two benchmark datasets show that our synthesized results are coherent with stronger audio correlations and outperform the state-of-the-arts with superior performance on co-speech gesture generation. To summarize, our main contributions are three-fold: 1) As an early attempt at taming diffusion models for co-speech + +gesture generation, we formally define the diffusion and denoising process in gesture space, which synthesizes audio-aligned gestures of high-fidelity. 2) We devise the Diffusion Audio-Gesture Transformer with implicit classifier-free diffusion guidance to better deal with the input conditional information from multiple sequential modalities. 3) We propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy. + +# 2. Related Work + +Co-Speech Gesture Generation. Synthesizing co-speech gestures is crucial for a variety of applications. Conventional studies resort to rule-based pipelines [11, 21, 31], where linguistic experts pre-define the speech-gesture pairs and refine the transitions between different motions. Recent works exploit neural networks to learn the mapping from speech to gesture based on a large training corpus, where an off-the-shelf pose estimator is leveraged to label the online videos for pseudo annotations [1, 16, 30, 34, 41, 42]. Meanwhile, some works study the influence of input modality, verifying the connections between co-speech gesture and speech audio [25], text transcript [3], speaking style [2], and speaker identity [41]. To further improve the model's capacity, previous studies explore multiple architecture choices, including CNN [17], RNN [42], Transformer [6], and VQVAE [29, 40]. Notably, several recent works are based on GANs to guarantee realistic results [16, 30, 34, 41], which involve the adversarial training between the generator and the discriminator. However, the notorious mode collapse and unstable training of GANs prevent the high-fidelity gesture distribution learning conditioned on audio. + +Diffusion Probabilistic Models. Diffusion probabilistic models have achieved promising results on unconditional image generation [19], which are further applied to conditional tasks like text-to-image [36]. Among the diffusion-based literature, previous works focus on static data and conditions. Besides, they mainly utilize explicit guidance like pretrained classifiers [13] and CLIP similarity [28, 33, 35, 36] to guide the generation process. In this work, we explore a more challenging co-speech gesture generation setting, where the gesture data and audio conditions are both sequential, and the audio-to-gesture mapping is implicit. To this end, we propose the Diffusion Audio-Gesture Transformer to guarantee temporally aligned generation. We further propose the Diffusion Gesture Stabilizer to simultaneously achieve diverse and temporally coherent gestures with an annealed noise sampling strategy. + +# 3. Our Approach + +Figure 2 depicts an overview of the proposed DiffGesture framework to generate co-speech gestures of high fidelity. In this section, we first introduce the problem + +![](images/33f47d693166d371a24e87d4333ff5521d748911384902045219a7b08e41fa48.jpg) +Figure 2. Overview of the Diffusion Co-Speech Gesture (DiffGesture) Framework. Given the gesture sequence $\pmb{x}_0$ , we first establish the forward diffusion (purple) and conditional denoising process (green) in gesture space. Then, we devise the Diffusion Audio-Gesture Transformer to attend to the input conditions of initial poses $\pmb{p}_{(1:M)}$ , speech audio $\pmb{a}$ , time embedding $t$ and corrupted gesture $\pmb{x}_t$ from multiple modalities (blue). At the diffusion sampling stage (grey), we propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy. To further incorporate implicit classifier-free guidance, we jointly train the conditional $(1 - p_{\text{uncond}})$ and unconditional $(p_{\text{uncond}})$ models. This allows us to trade off between diversity and quality during inference. + +![](images/432316dc0fac48cbc18da6b2853d75cd9374a705e72851bedff75f5f61a4a010.jpg) + +![](images/10612c5ffe7a00e2688305fa64cd3490996b86ccc99125c08e3a3803274a9f9f.jpg) + +formulation of audio-driven co-speech gesture generation (Section 3.1). We then establish the forward diffusion and the reverse conditional generation process in gesture space (Section 3.2). Furthermore, we elaborate the Diffusion Audio-Gesture Transformer to attend to the conditions from multiple modalities and enhance the speech-gesture correlations with temporal dependency (Section 3.3). To eliminate temporal inconsistency introduced by naive noises, we propose a novel Diffusion Gesture Stabilizer with annealed noise sampling strategies and describe this module in (Section 3.4). Finally, incorporating implicit classifier-free guidance in co-speech gestures is discussed in (Section 3.5). + +# 3.1. Problem Formulation + +With a large-scale co-speech gesture training corpus, we leverage the speaking videos with clear co-speech upper body movements for model learning. In particular, for each video clip of $N$ frames, we extract the accompanying speech audio sequence $\pmb{a} = \{\pmb{a}_1,\dots ,\pmb{a}_N\}$ and use the offthe-shelf human pose estimator OpenPose [9] to annotate the per-frame human skeletons as $\pmb{x} = \{\pmb{p}_1,\dots ,\pmb{p}_N\}$ . We follow baseline methods [30, 41] to pre-process such skeletal representation as the concatenation of unit direction vectors as $\pmb{p_i} = [\pmb{d}_{i,1},\pmb{d}_{i,2},\dots ,\pmb{d}_{i,J - 1}]$ , where $\pmb{p_i}$ denotes the pose description coordinates of the $i$ -th frame, $J$ is the total joint number and $\pmb{d}_{i,j}$ represents the $j$ -th unit direction vector among the $J$ joints of the $i$ -th image frame. The diffusion model's reverse denoising process $G$ parameterized by $\theta$ is optimized to synthesize the human skeleton sequence $\pmb{x}$ , which is further conditioned on the speech audio sequence $\pmb{a}$ and the initial poses $\{\pmb{p}_1,\dots ,\pmb{p}_M\}$ of the first $M$ frames. The learning objective of the overall framework can be formulated as $\arg \min_{\theta}||\pmb{x} - G_{\theta}(\pmb {a},\pmb {p}_1,\dots ,\pmb {p}_M)||$ + +# 3.2. Gesture Space Forward and Reverse Process + +Given $\pmb{x}_0 \in \mathbb{R}^{N \times 3(J-1)}$ sampled from real data distribution $q(\pmb{x}_0)$ , our goal is to learn a model distribution $p_{\theta}(\pmb{x}_0)$ parameterized by $\theta$ that approximates $q(\pmb{x}_0)$ . Specifically, denoising diffusion probabilistic models (DDPMs) [19] define the latent variable models of the form $p_{\theta}(\pmb{x}_0) = \int p_{\theta}(\pmb{x}_{0:T}) d\pmb{x}_{1:T}$ , where $\pmb{x}_{1:T}$ are latent variables in the same sample space as $\pmb{x}_0$ with the same dimensionality. + +The Forward Diffusion Process. The forward process, which is also termed as the diffusion process, approximates the posterior distribution $q(\pmb{x}_{1:T}|\pmb{x}_0)$ . It is defined as a Markov chain that gradually adds Gaussian noise to the data sample according to a variance schedule $\beta_1,\dots ,\beta_T$ : + +$$ +q \left(\boldsymbol {x} _ {1: T} \mid \boldsymbol {x} _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}\right), \tag {1} +$$ + +$$ +w h e r e \quad q (\boldsymbol {x} _ {t} | \boldsymbol {x} _ {t - 1}) = \mathcal {N} (\boldsymbol {x} _ {t}; \sqrt {1 - \beta_ {t}} \boldsymbol {x} _ {t - 1}, \beta_ {t} \boldsymbol {I}). \tag {2} +$$ + +The variances $\beta_{t}$ are constant hyperparameters to ease the modeling of the reverse process [19]. Through such a corruption scheme, the structural information of the original skeleton is gradually substituted by noises, which finally leads to a pure white noise when $T$ goes to infinity. Therefore, the prior latent distribution of $p(\boldsymbol{x}_T)$ is $\mathcal{N}(\boldsymbol{x}_T; \mathbf{0}, \boldsymbol{I})$ with only information of Gaussian noise. + +Reverse Conditional Gesture Generation. The reverse process, which is also termed as the generative process, estimates the joint distribution of $p_{\theta}(\pmb{x}_{0:T})$ . As proved in [15], the reverse process of the continuous diffusion process preserves the same transition distribution form, which motivates us to leverage a Gaussian transition to formulate $p_{\theta}(\pmb{x}_{t-1}|\pmb{x}_t)$ under an unconditional setting, which + +approximates the intractable process as: + +$$ +p _ {\theta} \left(\boldsymbol {x} _ {0: T}\right) = p _ {\theta} \left(\boldsymbol {x} _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}\right), \tag {3} +$$ + +where $p_{\theta}(\boldsymbol{x}_{t - 1}|\boldsymbol{x}_t) = \mathcal{N}(\boldsymbol{x}_{t - 1};\mu_{\theta}(\boldsymbol{x}_t,t),\Sigma_{\theta}(\boldsymbol{x}_t,t))$ (4) + +The corrupted noisy data $\pmb{x}_t$ is sampled by $q(\pmb{x}_t|\pmb{x}_0) = \mathcal{N}(\pmb{x}_t; \sqrt{\bar{\alpha}_t}\pmb{x}_0, (1 - \bar{\alpha}_t)\pmb{I})$ , where $\alpha_t = 1 - \beta_t$ and $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$ . Note that we set the variances $\Sigma_\theta(\pmb{x}_t, t) = \beta_t\pmb{I}$ to untrained time-dependent constants. The above diffusion model formulations show compelling performances on unconditional generation. To further adapt to the conditional co-gesture synthesis, we have to provide additional inputs to the model, including the audio and initial poses. Therefore, we treat the speech audio $\pmb{a}$ and initial poses $\pmb{p}_{1:M}$ as context information $\pmb{c}$ and inject conditions into the generation process. The reverse process of each timestep (Eq. 4) can be thus updated as: + +$$ +p _ {\theta} (\boldsymbol {x} _ {t - 1} | \boldsymbol {x} _ {t}, \boldsymbol {c}) = \mathcal {N} (\boldsymbol {x} _ {t - 1}; \mu_ {\theta} (\boldsymbol {x} _ {t}, t, \boldsymbol {c}), \beta_ {t} \boldsymbol {I}). \tag {5} +$$ + +In this way, we could start the generation process by firstly sample a Gaussian noise $\boldsymbol{x}_T \sim \mathcal{N}(\mathbf{0}, \boldsymbol{I})$ and follow the Markov chain to iteratively denoise the latent variable $\boldsymbol{x}_t$ via Eq. 5 to get the final results. The overview of conditional co-speech gesture process is illustrated in Figure 1. + +Training Objective. To optimize the overall framework, we optimize the variational lower bound on negative log-likelihood: $\mathbb{E}[-\log p_{\theta}(\pmb{x}_0)] \leq \mathbb{E}_q[-\log \frac{p_{\theta}(\pmb{x}_0)}{q(\pmb{x}_{1:T}|\pmb{x}_0)}]$ . We rewrite the loss function conditioned on context $c$ and eliminate all the constant items that do not require training: $L(\theta) = \mathbb{E}_q[\sum_{t=2}^{T} D_{KL}(q(\pmb{x}_{t-1}|\pmb{x}_t, \pmb{x}_0)||p_{\theta}(\pmb{x}_{t-1}|\pmb{x}_t, c))]$ . With reparameterization, we can represent each term in $L_{\theta}$ using MSE loss. We follow [19] to further simplify the training objective to the ensemble of MSE losses as: + +$$ +L (\theta) = \mathbb {E} _ {q} [ \left\| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}, \boldsymbol {c}, t) \right\| ^ {2} ], (6) +$$ + +where $t$ is uniformly sampled between 1 and $T$ . As we jointly train the model under conditional and unconditional setting, a trainable masked embedding with probability $p_{\text{uncond}}$ replaces context $c$ and the diffusion model predicts the noise in the unconditional setting. The detailed principles will be discussed in Section 3.5. + +# 3.3. Diffusion Audio-Gesture Transformer + +With the naive conditional generation scheme as specified in Section 3.2, we still confront a critical problem in the setting of co-speech gesture generation. Since $x_0$ denotes the skeleton sequence of $N$ frames, there exists temporal dependency among the target sequence and context information, making it more complex than time-invariant tasks + +like image generation. Therefore, how to guarantee temporally coherent results in a non-autoregressive conditional generation process remains an unsolved problem. + +In contrast to most previous studies that resort to recurrent networks [30, 41], we propose to make use of the Transformer's strong capacity in sequential data modeling. Specifically, since the noisy gesture sequence $\boldsymbol{x}_t$ and the contextual information $c$ align in the temporal dimension, we concatenate them in the feature channel. In this way, the skeleton and context condition of each frame serve as an individual token, which captures the long-term dependency by the self-attention mechanism: + +$$ +\operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q K} ^ {\mathsf {T}}}{\sqrt {\ell}}\right) \mathbf {V}, \tag {7} +$$ + +where $\mathbf{Q},\mathbf{K},\mathbf{V}$ are the query, key, and value matrix from input tokens, $\ell$ is the channel dimension, and $\mathsf{T}$ is the matrix transpose operation. Such a design also avoids severe error accumulation in autoregressive pipelines, enabling us to generate coherent gesture sequences. + +# 3.4. Diffusion Gesture Stabilizer + +In DDPMs, the independent random variables $z$ introduced at the sampling stage promote diversity and thus systematically improve the task performance. However, the variation in temporal dimension introduced by $z$ , especially when timestep $t$ is small in the reverse process, can have a negative effect on temporal consistency. At the inference stage, to achieve the trade-off between diversity and temporal consistency, we propose a novel Diffusion Gesture Stabilizer without extra training expenses under two annealed scenarios, where the term "annealed" means that the process is transitioned from high variance and entropy (hot) to low variance and entropy (cold). + +Thresholding. Since temporally independent Gaussian noises inevitably introduce inconsistency, restricting the temporal variation naturally helps to avoid inconsistency. And hard thresholding serves as an effective trick. In detail, we set a time threshold $t_0$ , and then use the same $z \in \mathbb{R}^{N \times C}$ in the naive sampling strategy for $t > t_0$ and set $z = \{z_0\}_{i=1}^N$ for $t \leq t_0$ , where $z_0 \in \mathbb{R}^C$ follows $\mathcal{N}(\mathbf{0}, \mathbf{I})$ which do not introduce variation in the temporal dimension. + +Smooth Sampling. We further modify $z(t) = \{z_i(t)\}_{i=1}^N$ to be a smooth annealing version via variance-aware sampling. In the original sampling rule of DDPMs, i.i.d.variables $z_i(t)$ are sampled from $\mathcal{N}(\mathbf{0},\mathbf{I})$ . With smooth resampling, we first sample $z_0(t) \sim \mathcal{N}(\mathbf{0},\sigma_a^2(t)\mathbf{I})$ only once for each timestep $t$ in the reverse process, then given $z_0(t)$ , we sample $z_i(t)|z_0(t) \sim \mathcal{N}(z_0(t),(1 - \sigma_a^2(t))\mathbf{I})$ for $i \in \{1,\dots,N\}$ , where $\sigma_a(t) \in [0,1]$ is a nondecreasing function to achieve variance annealing. + +# Algorithm 1 Training + +1: repeat +2: Sample $(\pmb{x}_0,\pmb{c})\sim q(\pmb{x}_0,\pmb{c})$ +3: Sample $\tau \sim \mathrm{Uniform}(0,1)$ . Set $c = \varnothing$ if $\tau < p_{\text{uncond}}$ +4: Sample $t \sim \mathrm{Uniform}(\{1, \dots, T\})$ +5: Compute $\nabla_{\theta}\left\| \pmb {\epsilon} - \pmb{\epsilon}_{\theta}(\sqrt{\bar{\alpha}_{t}}\pmb {x}_{0} + \sqrt{1 - \bar{\alpha}_{t}}\pmb {\epsilon},\pmb {c},t)\right\|^{2}$ +6: Perform gradient descent +7: until converged + +# 3.5. Implicit Classifier-free Guidance + +In co-speech gesture literature, the speech-to-gesture mapping is implicit, where the same audio corresponds to diverse gestures and different audiOs could incur the same motion [22, 25], making it difficult to utilize the commonly used explicit classifier guidance [13, 33]. Therefore, how can we further exploit practical guidance for better audio correlations and mode coverage? Our solution to this question is to train an extra "unconditional" diffusion model to implicitly guide the generation. Dhariwal et al. [13] first introduce the classifier guidance of cross-entropy gradient $\nabla_{\pmb{x}_t}\log p_\phi (y|\pmb {x}_t)$ , where the pretrained classifier is parameterized by $\phi$ and $y$ denotes the classification logits. This gradient term is further scaled by the covariance matrix to modify the mean value of transition distribution in Eq. 4. To adapt to the cases where no explicit guidance is available, we follow [20] to jointly train the conditional and unconditional models, termed as classifier-free guidance. In particular, according to the implicit classifier's property that $p(c|x_t)\propto p(x_t|c) / p(x_t)$ , we could derive a gradient relationship in the implicit classifier as: + +$$ +\nabla_ {\boldsymbol {x} _ {t}} \log p (\boldsymbol {c} | \boldsymbol {x} _ {t}) \propto \nabla_ {\boldsymbol {x} _ {t}} \log p (\boldsymbol {x} _ {t} | \boldsymbol {c}) - \nabla_ {\boldsymbol {x} _ {t}} \log p (\boldsymbol {x} _ {t}), \tag {8} +$$ + +which is further proportional to $\epsilon^{*}(\boldsymbol{x}_{t}|\boldsymbol{c}) - \epsilon^{*}(\boldsymbol{x}_{t})$ . Therefore, we use a single Transformer network to parameterize both settings by a mix-up training trick: for the probability of $p_{uncond}$ , we set the context information $\boldsymbol{c}$ as masked embedding to train the unconditional setting, while for other cases, we train the original conditional counterpart. The training is shown in Algorithm 1. + +Sampling with Classifier-free Guidance. Starting from Gaussian noise, we iteratively remove noises in $x_{t}$ . As we use implicit classifier-free guidance, similar to Equation 8, the predicted Gaussian noise is modified as: + +$$ +\hat {\boldsymbol {\epsilon}} _ {\theta} = \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t}, t) + s \cdot \left(\boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t}, \boldsymbol {c}, t\right) - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t}, t)\right), \tag {9} +$$ + +where $s$ is the scale parameter to trade off the diversity and quality. With classifier-free guidance, Algorithm 2 reveals how to generate co-speech gestures given the trained diffusion model via the Diffusion Gesture Stabilizer with the Smooth Sampling annealed scenario. + +# Algorithm 2 Sampling + +1: Trained diffusion model $\theta, \mathbf{x}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ +2: for $t = T, \dots, 1$ do +3: $\hat{\pmb{\epsilon}}_{\theta} = \pmb{\epsilon}_{\theta}(\pmb{x}_t,t) + s\cdot (\pmb{\epsilon}_{\theta}(\pmb{x}_t,\pmb {c},t) - \pmb{\epsilon}_{\theta}(\pmb{x}_t,t))$ +4: $\mathbf{z}_0(t) \sim \mathcal{N}(\mathbf{0}, \sigma_a^2(t)\mathbf{I})$ +5: for $i = 1,\dots ,N$ do +6: $\pmb{z}_i(t) \sim \mathcal{N}(\pmb{z}_0(t), (1 - \sigma_a^2(t))\pmb{I})$ +7: end for +8: $\mathbf{z}(t) = \{\mathbf{z}_1(t),\dots ,\mathbf{z}_N(t)\}$ , if $t > 1$ , else $\mathbf{z}(t) = \mathbf{0}$ +9: $\pmb{x}_{t-1} = \frac{1}{\sqrt{\alpha_t}} \left( \pmb{x}_t - \frac{1 - \alpha_t}{\sqrt{1 - \bar{\alpha}_t}} \hat{\pmb{\epsilon}}_\theta \right) + \sigma_t \pmb{z}(t)$ +10: end for +11: return $x_0$ + +# 4. Experiments + +# 4.1. Co-Speech Gesture Datasets + +TED Gesture. As a large-scale dataset for gesture generation research, TED Gesture dataset [41, 42] contains 1,766 TED videos of different narrators covering various topics. We follow the data process in former works [30, 41], where the poses are resampled with 15 FPS, and frame segments of length 34 are obtained with a stride of 10. + +TED Expressive. While the poses in TED Gesture only contain 10 upper body key points without vivid finger movements, the TED Expressive dataset [30] is further expressive of both finger and body movements. The state-of-art 3D pose estimator ExPose [12] is used to fully capture the pose information in data. As a result, TED Expressive annotates the 3D coordinates of 43 keypoints, including 13 upper body joints and 30 finger joints. + +# 4.2. Experimental Settings + +Comparison Methods. We compare our method on two benchmark datasets with the state-of-the-art methods in recent years. 1) Attention Seq2Seq [42] elaborates on the attention mechanism to generate pose sequences from speech text. 2) Speech2Gesture [16] uses spectrums of the speech audio segments as the input and generates speech gestures adversarially. 3) Joint Embedding [3] maps text and motion to the same embedding space, then generates outputs from motion description text. 4) Trimodal [41] serves as a strong baseline that learns from text, audio, and speaker identity to generate gestures, outperforming former methods by a large margin. 5) HA2G [30] introduces a hierarchical audio learner that captures information across different semantic granularities, achieving state-of-the-art performances. This method hierarchically extracts rich features at the cost of heavier GPU memory overhead, while our method requires much smaller expenses. + +Implementation Details. For all the methods in both datasets, we set $N = 34$ and $M = 4$ to get $M$ -frame pose sequences where the first $N$ frames are used for reference, termed as initial poses. There are $J$ upper body joints in + +
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
Ground Truth00.698108.52500.703178.827
Attention Seq2Seq [42]18.1540.19682.77654.9200.152122.693
Speech2Gesture [16]19.2540.66893.80254.6500.679142.489
Joint Embedding [3]22.0830.20090.13864.5550.130120.627
Trimodal [41]3.7290.667101.24712.6130.563154.088
HA2G [30]3.0720.672104.3225.3060.641173.899
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
+ +Table 1. The Quantitative Results on TED Gesture [41] and TED Expressive [30]. We compare the proposed diffusion-based method against recent SOTA methods [3, 16, 30, 41, 42] and ground truth. For FGD, the lower, the better; for other metrics, the higher, the better. + +all the frames of pose sequences, where $J = 10$ for TED Gesture and $J = 43$ for TED Expressive. Following [41], to eliminate the effect of the joint lengths and root motion, we represent the joints' positions using $J - 1$ directional vectors normalized to the unit vectors and train the model to learn the directional vectors. For the audio processing, we use the same audio encoder used in [41] to extract the feature of the raw audio clips directly. The audio clips are encoded as $N$ audio feature vectors of 32-D. The audio feature and initial poses are concatenated to form the conditional context information of the diffusion model. For the diffusion process, the number of timesteps is $T = 500$ , and the variances increase linearly from $\beta_{1} = 1e - 4$ to $\beta_{T} = 0.02$ . For the Stabilizer, $t_0$ can be adjusted from 20-30 for Thresholding, and a quadratic non-increasing function $\sigma_{a}(t)$ is applied for Smooth Sampling. The hidden dimension of the transformer blocks, is set to 256 for TED Gesture and 512 for TED Expressive. We use 8 Transformer blocks, each of which comprises a multi-head self-attention block and a Feed-Forward Network. We use an Adam optimizer, and the learning rate is $5e - 4$ . It takes 10 hours to train the model on TED Gesture and 20 hours on TED Expressive on a single NVIDIA GeForce RTX 3090 GPU. + +# 4.3. Evaluation Metrics + +In evaluation, we use three metrics that are used in cospeech gesture generation and relative fields [24, 30]. + +Fréchet Gesture Distance (FGD). Similar to the Fréchet Inception Distance (FID) metric [18], which is widely applied in image generation studies, FGD is used to measure the distance between the synthesized gesture distribution and the real data distribution. Yoon et al. [41] define FGD by training a skeleton sequence auto-encoder to extract the features of the real gesture sequences $X$ and the features of the generated gesture sequences $\hat{X}$ : + +$$ +\operatorname {F G D} (X, \hat {X}) = \| \mu_ {r} - \mu_ {g} \| ^ {2} + \operatorname {T r} (\Sigma_ {\mathrm {r}} + \Sigma_ {\mathrm {g}} - 2 (\Sigma_ {\mathrm {r}} \Sigma_ {\mathrm {g}}) ^ {1 / 2}), +$$ + +where $\mu_r$ and $\Sigma_r$ are the first and the second moments of the + +latent feature distribution of the real gestures $X$ , and $\mu_g$ and $\Sigma_g$ are the first and the second moments of the latent feature distribution of the generated gestures $\hat{X}$ . We intuitively find that among the three metrics, FGD tells the most whether the generated pose sequences are of high quality. + +Beat Consistency Score (BC). Proposed in [24, 26], BC measures motion-audio beat correlation. Considering that the kinematic velocities vary from different joints, we use the change of included angle between bones to track motion beats following [30]. Specifically, we can calculate the mean absolute angle change (MAAC) of angle $\theta_{j}$ in adjacent frames by $\mathrm{MAAC}(\theta_j) = \frac{1}{S}\frac{1}{T - 1}\sum_{s = 1}^{S}\sum_{t = 1}^{T - 1}\| \theta_{j,s,t + 1} - \theta_{j,s,t}\| _1$ , where $S$ denotes the total number of clips in the dataset, $T$ denotes the number of frames in each clip, and $\theta_{j,s,t}$ is the included angle between the $j$ -th and the $(j + 1)$ -th bone of the $s$ -th clip at time-step $t$ . Then, we can compute the angle change rate of frame $t$ for the $s$ -th clip as $\frac{1}{J - 1}\sum_{j = 1}^{J - 1}(\| \theta_{j,s,t + 1} - \theta_{j,s,t}\| _1 / \mathrm{MAAC}(\theta_j))$ . Then we extract the local optima whose first-order difference is higher than a threshold to get kinematic beats, which are used to compute BC later. Following [24] to detect audio beat by onset strength [14], we compute the average distance between each audio beat and its nearest motion beat as Beat Consistency Score: + +$$ +\mathrm {B C} = \frac {1}{n} \sum_ {i = 1} ^ {n} \exp \left(- \frac {\min _ {\forall t _ {j} ^ {y} \in B ^ {y}} \| t _ {i} ^ {x} - t _ {j} ^ {y} \| ^ {2}}{2 \sigma^ {2}}\right), \tag {10} +$$ + +where $t_i^x$ is the $i$ -th audio beats, $B^y = \{t_i^y\}$ is the set of the kinematic beats, and $\sigma$ is a parameter to normalize sequences, set to 0.1 empirically. + +Diversity. This metric evaluates the variations among generated gestures corresponding to various inputs [23]. We use the same feature extractor when measuring FGD to map synthesized gestures into latent feature vectors and calculate the mean feature distance. In detail, we randomly pick 500 generated samples and compute the mean absolute error between the features and the shuffled features. + +![](images/cdacf38770acbf609b3085c1b3fa88d95f5e6ed1d2b26ee14dea89b7e1612cba.jpg) +Figure 3. Visualization Results of Our DiffGesture on Two Datasets. Three cases are picked up, where (i) and (ii) are TED Expressive cases, and (iii) is a TED Gesture case. We highlight dull cases generated by comparison methods with rectangles, indicating the mode collapse phenomenon of baselines. + +
MethodsGTS2S. [42]S2G. [16]Joint. [3]Tri. [41]HA2G [30]DiffGesture(Ours)
Naturalness4.331.222.561.223.223.674.00
Smoothness3.943.501.613.443.443.393.89
Synchrony4.001.673.171.393.283.443.89
+ +Table 2. User Study Results. The ratings of motion naturalness, smoothness, and synchrony, are on a scale of 1-5, with 5 being the best. + +# 4.4. Evaluation Results + +Quantitative Results. We compare our method with all the baselines with three metrics on TED Gesture and TED Expressive. The results are shown in Table 1. For the metrics of Ground Truth, we report the values in our implementation. For TED Gesture, we report FGD of all baselines in [30] and evaluate BC and Diversity on our own†. For TED Expressive, all the results of baselines are reported from [30]. Assuming the pseudo ground truth pose follows the real distribution, the FGD of Ground Truth in the table is 0. It is observed that our DiffGesture achieves state-of-the-art performance on both datasets, especially outperforming existing methods by a large margin on TED + +Expressive. Besides, since BC and Diversity are proposed to measure motion-audio beat correlation and variation, these two metrics of Ground Truth cannot be treated as upper bounds, and it is worth noting that our results may be higher than the ones of Ground Truth, indicating that the generated gestures are of high quality. Qualitative Results. We show the keyframes of all the methods on two datasets in Figure 3. Since TED Expressive requires a higher ability of generative models, we pick up two cases for TED Expressive and one for TED Gesture. For each case, we select three keyframes (an early, a middle, and a late one) to show the pose motions. Comparison methods tend to generate slow and invariant poses and sometimes produce unreliable and stiff results. In contrast, DiffGesture produces diverse human-like poses without resulting in + +
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
DiffGesture Base2.4500.632104.6883.8220.707174.377
DiffGesture w/o Stabilizer2.2190.674105.1922.7920.721180.125
DiffGesture w/o classifier-free1.8100.673105.6443.3260.717178.245
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
+ +mean poses which are slow and rigid. Besides, as a primary drawback of GAN-based methods, mode collapse makes comparison methods often produce a single type of output, which is severe for pose motion generation. Such a phenomenon is shown in Fig. 3, where the generated frames with nearly the same pose are highlighted with rectangles. + +User Study. To better validate the qualitative performance, we conduct a user study on the generated co-speech gestures. The study involves 18 participants, with 9 females and 9 males in the age range of 18-25 years old. The participants are required to grade the motion's quality and coherence, and all the clips are without labels. In total, we pick up 30 cases, 20 for TED-Expressive and 10 for TED Gesture. For each case, we show 7 videos with the order of the methods shuffling, including the ground truth. We adopt the Mean Opinion Scores rating protocol, and each participant is required to rate three aspects of generated motions: Naturalness; Smoothness; Synchrony between speech and generated gestures. The results are shown in Table 2 where the ratings are on a scale of 1 to 5, with 5 being the best. Our participants widely accept that our method produces high-fidelity results in all three aspects. + +# 4.5. Ablation Studies + +Ablation Study on the Proposed Modules. To demonstrate the effectiveness of our proposed DiffGesture, we present ablation studies on the key modules in the framework. In detail, we conduct experiments as follows. 1) DiffGesture Base means we only use the proposed conditional diffusion generation process without further design. 2) DiffGesture w/o classifier-free, where classifier-free guidance is not implemented at both the training stage and inference stage. 3) DiffGesture w/o Stabilizer, where Diffusion Gesture Stabilizer is removed at the inference stage. The results are reported in Table 3. The results illustrate the effectiveness of the designed Diffusion Gesture Stabilizer and the implicit classifier-free guidance. + +Ablation Study on the Network Architectures. We investigate the performance of the GRU architecture in diffusion models, autoregressively generating poses in [30, 41]. We replace the Diffusion Audio-Gesture Transformer with the + +Table 3. Ablation Study on the Proposed Modules. We investigate effectiveness of the proposed modules, Diffusion Gesture Stabilizer and implicit classifier-free guidance. The results indicate that the proposed modules consistently improve performance on the benchmarks. + +
MethodsFGD ↓BC ↑Diversity ↑
GRU on Da14.3430.65898.472
Transformer on Da1.5060.699106.722
GRU on Db17.4520.680172.168
Transformer on Db2.6000.718182.757
+ +Table 4. Ablation Study on the Network Architectures. We compare the performance of GRU and Transformer for diffusion-based backbone on TED Gesture $(D_{a})$ and TED Expressive $(D_{b})$ . + +GRU in the diffusion model. All the context inputs remain the same as our method and are concatenated before the diffusion network. Results are shown in Table 4. Though GRU serves as a strong baseline network in co-gesture learning, it fails to generate high-performance data like our designed Transformer-based network, which indicates the effectiveness of our Transformer-based network and that applying diffusion models in the audio-driven conditional generation is a non-trivial task. + +# 5. Conclusion + +In this work, we present a novel diffusion-based framework DiffGesture for co-speech gesture generation. To generate coherent gestures with strong audio correlations, we propose the Diffusion Audio-Gesture Transformer with the Diffusion Gesture Stabilizer to better attend to the condition information. Such a non-autoregressive pipeline helps to efficiently generate results and reduce error accumulation. We hope our method offers a new perspective for diffusion-based temporal generation and how to capture sequential cross-modal dependencies. + +Acknowledgement. The work described in this paper was partially supported by grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (T45-401/22-N), the National Natural Science Fund (62201483), HKU Seed Fund for Basic Research (202009185079 and 202111159073), RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1884–1895, 2020. 2 +[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2 +[3] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2, 5, 6, 7 +[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, volume 39, pages 487-496. Wiley Online Library, 2020. 1 +[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1 +[6] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 2 +[7] Judee K Burgoon, Thomas Birk, and Michael Pfau. Nonverbal behaviors, persuasion, and credibility. Human communication research, 17(1):140-169, 1990. 1 +[8] B. Butterworth and U. Hadar. Gesture, speech, and computational stages: A reply to mcneill. Psychological Review, 96(1):168-174, 1989. 1 +[9] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 3 +[10] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1 +[11] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 1, 2 +[12] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 5 + +[13] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 5 +[14] Daniel Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36:51-60, 03 2007. 6 +[15] William Feller. On the theory of stochastic processes, with particular reference to applications. In *Berkeley Symposium on Mathematical Statistics and Probability*, pages 403-432, 1949. 3 +[16] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 5, 6, 7 +[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 6 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4 +[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5 +[21] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In 2012 7th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 25-32. IEEE, 2012. 1, 2 +[22] Dong Won Lee, Chaitanya Ahuja, and Louis-Philippe Morency. Crossmodal clustered contrastive learning: Grounding of spoken language to gesture. In GENEA: Generation and Evaluation of Non-verbal Behaviour for Embodied Agents Workshop 2021, 2021. 5 +[23] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 6 +[24] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 6 +[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional + +variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 5 +[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6 +[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 612-630. Springer, 2022. 1 +[28] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 2 +[29] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. arXiv preprint arXiv:2212.02350, 2022. 2 +[30] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10462-10472, June 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[31] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 25-35, 2013. 1, 2 +[32] David McNeill. Hand and mind. De Gruyter Mouton, 2011. 1 +[33] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2, 5 +[34] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11077-11086, 2021. 1, 2 +[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[36] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed + +Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2 +[38] P. Wagner, Z. Malisz, and S. Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1 +[39] Jing Xu, Wei Zhang, Yalong Bai, Qibin Sun, and Tao Mei. Freeform body motion generation from speech. arXiv preprint arXiv:2203.02291, 2022. 1 +[40] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107. IEEE, 2022. 2 +[41] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[42] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 5, 6, 7 \ No newline at end of file diff --git a/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/images.zip b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3779267fde943047bb69710435253d940aebd00 --- /dev/null +++ b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc2889be29140920d70f257b8862cbb8373eca95e04da20bdadc658cfea6744 +size 390097 diff --git a/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/layout.json b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e338077fd72179862c30acc736a06a9763926c17 --- /dev/null +++ b/2023/Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation/layout.json @@ -0,0 +1,9555 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "Taming Diffusion Models for Audio-Driven Co-Speech Gesture Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": "Lingting Zhu" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": " Xian Liu" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": " Xuanyu Liu" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": " Rui Qian" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": " Lequan Yu" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": "The University of Hong Kong \n" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": "The Chinese University of Hong Kong \n" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 499, + 186 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 187, + 443, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 187, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 148, + 187, + 443, + 213 + ], + "type": "text", + "content": "{ltzhu99，u3008631}@connect.hku.hk，lqyu@hku.hk, {alvinliu，qr021}@ie.cuhk.edu.hk，ziwei.liu@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 579 + ], + "type": "text", + "content": "Animating virtual avatars to make co-speech gestures facilitates various applications in human-machine interaction. The existing methods mainly rely on generative adversarial networks (GANs), which typically suffer from notorious mode collapse and unstable training, thus making it difficult to learn accurate audio-gesture joint distributions. In this work, we propose a novel diffusion-based framework, named Diffusion Co-Speech Gesture (DiffGesture), to effectively capture the cross-modal audio-to-gesture associations and preserve temporal coherence for high-fidelity audio-driven co-speech gesture generation. Specifically, we first establish the diffusion-conditional generation process on clips of skeleton sequences and audio to enable the whole framework. Then, a novel Diffusion Audio-Gesture Transformer is devised to better attend to the information from multiple modalities and model the long-term temporal dependency. Moreover, to eliminate temporal inconsistency, we propose an effective Diffusion Gesture Stabilizer with an annealed noise sampling strategy. Benefiting from the architectural advantages of diffusion models, we further incorporate implicit classifier-free guidance to trade off between diversity and gesture quality. Extensive experiments demonstrate that DiffGesture achieves state-of-the-art performance, which renders coherent gestures with better mode coverage and stronger audio correlations. Code is available at https://github.com/Advocate99/DiffGesture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 128, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 684 + ], + "type": "text", + "content": "Making co-speech gestures is an innate human behavior in daily conversations, which helps the speakers to express their thoughts and the listeners to comprehend the meanings [10, 32, 38]. Previous linguistic studies verify that such non-verbal behaviors could liven up the atmosphere" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 239, + 545, + 305 + ], + "blocks": [ + { + "bbox": [ + 307, + 239, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 305 + ], + "type": "image", + "image_path": "2400d7bf94a445e8333befea73469a2373b520a6be01c3f856550347bb91652e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": "Figure 1. Illustration of Conditional Generation Process in Co-Speech Gesture Generation. The diffusion process " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": " gradually adds Gaussian noise to the gesture sequence (i.e., " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": " sampled from the real data distribution). The generation process " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": " learns to denoise the white noise (i.e., " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": " sampled from the normal distribution) conditioned on context information " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": " denotes the corrupted gesture sequence at the " + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 313, + 546, + 392 + ], + "type": "text", + "content": "-th diffusion step." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 410, + 546, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 546, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 546, + 482 + ], + "type": "text", + "content": "and improve mutual intimacy [7, 8, 21]. Therefore, animating virtual avatars to gesticulate co-speech movements is crucial in embodied AI. To this end, recent researches focus on the problem of audio-driven co-speech gesture generation [16, 25, 30, 41], which synthesizes human upper body gesture sequences that are aligned to the speech audio." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 483, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 700 + ], + "type": "text", + "content": "Early attempts downgrade this task as a searching-and-connecting problem, where they predefine the corresponding gestures of each speech unit and stitch them together by optimizing the transitions between consecutive motions for coherent results [11,21,31]. In recent years, the compelling performance of deep neural networks has prompted data-driven approaches. Previous studies establish large-scale speech-gesture corpus to learn the mapping from speech audio to human skeletons in an end-to-end manner [4,5,25, 27,30,34,39]. To attain more expressive results, Ginosar et al. [16] and Yoon et al. [41] propose GAN-based methods to guarantee realism by adversarial mechanism, where the discriminator is trained to distinguish real gestures from the synthetic ones while the generator's objective is to fool the discriminator. However, such pipelines suffer from the inherent mode collapse and unstable training, making them difficult to capture the high-fidelity audio-conditioned gesture distribution, resulting in dull or unreasonable poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "The recent paradigm of diffusion probabilistic models" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10544" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 289 + ], + "type": "text", + "content": "provides a new perspective for realistic generation [19, 37], facilitating high-fidelity synthesis with desirable properties such as good distribution coverage and stable training compared to GANs. However, it is non-trivial to adapt existing diffusion models for co-speech gesture generation. Most existing conditional diffusion models deal with static data and conditions [35, 36] (e.g., the image-text pairs without temporal dimension), while co-speech gesture generation requires generating temporally coherent gesture sequences conditioned on continual audio clips. Further, the commonly used denoising strategy in existing diffusion models samples independently and identically distributed (i.i.d.) noises in latent space to increase diversity. However, this strategy tends to introduce variation for each gesture frame and lead to temporal inconsistency in skeleton sequences. Therefore, how to generate high-fidelity co-speech gestures with strong audio correlations and temporal consistency is quite challenging within the diffusion paradigm." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 639 + ], + "type": "text", + "content": "To address the above challenges, we propose a tailored Diffusion Co-Speech Gesture framework to capture the cross-modal audio-gesture associations while maintaining temporal coherence for high-fidelity audio-driven co-speech gesture generation, named DiffGesture. As shown in Figure 1, we formulate our task as a diffusion-conditional generation process on clips of skeleton and audio, where the diffusion phase is defined by gradually adding noise to gesture sequence, and the generation phase is referred as a parameterized Markov chain with conditional context features of audio clips to denoise the corrupted gestures. As we treat the multi-frame gesture clip as the diffusion latent space, the skeletons can be efficiently synthesized in a non-autoregressive manner to bypass error accumulation. To better attend to the sequential conditions from multiple modalities and enhance the temporal coherence, we then devise a novel Diffusion Audio-Gesture Transformer architecture to model audio-gesture long-term temporal dependency. Particularly, the per-frame skeleton and contextual features are concatenated in the aligned temporal dimension and embedded as individual input tokens to a Transformer block. Further, to eliminate the temporal inconsistency caused by the naive denoising strategy in the inference stage, we thus propose a new Diffusion Gesture Stabilizer module to gradually anneal down the noise discrepancy in the temporal dimension. Finally, we incorporate implicit classifier-free guidance by jointly training the conditional and unconditional models, which allows us to trade off between the diversity and sample quality during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "Extensive experiments on two benchmark datasets show that our synthesized results are coherent with stronger audio correlations and outperform the state-of-the-arts with superior performance on co-speech gesture generation. To summarize, our main contributions are three-fold: 1) As an early attempt at taming diffusion models for co-speech" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": "gesture generation, we formally define the diffusion and denoising process in gesture space, which synthesizes audio-aligned gestures of high-fidelity. 2) We devise the Diffusion Audio-Gesture Transformer with implicit classifier-free diffusion guidance to better deal with the input conditional information from multiple sequential modalities. 3) We propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 180, + 392, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 392, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 392, + 193 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 201, + 547, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 452 + ], + "type": "text", + "content": "Co-Speech Gesture Generation. Synthesizing co-speech gestures is crucial for a variety of applications. Conventional studies resort to rule-based pipelines [11, 21, 31], where linguistic experts pre-define the speech-gesture pairs and refine the transitions between different motions. Recent works exploit neural networks to learn the mapping from speech to gesture based on a large training corpus, where an off-the-shelf pose estimator is leveraged to label the online videos for pseudo annotations [1, 16, 30, 34, 41, 42]. Meanwhile, some works study the influence of input modality, verifying the connections between co-speech gesture and speech audio [25], text transcript [3], speaking style [2], and speaker identity [41]. To further improve the model's capacity, previous studies explore multiple architecture choices, including CNN [17], RNN [42], Transformer [6], and VQVAE [29, 40]. Notably, several recent works are based on GANs to guarantee realistic results [16, 30, 34, 41], which involve the adversarial training between the generator and the discriminator. However, the notorious mode collapse and unstable training of GANs prevent the high-fidelity gesture distribution learning conditioned on audio." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 453, + 548, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 548, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 548, + 646 + ], + "type": "text", + "content": "Diffusion Probabilistic Models. Diffusion probabilistic models have achieved promising results on unconditional image generation [19], which are further applied to conditional tasks like text-to-image [36]. Among the diffusion-based literature, previous works focus on static data and conditions. Besides, they mainly utilize explicit guidance like pretrained classifiers [13] and CLIP similarity [28, 33, 35, 36] to guide the generation process. In this work, we explore a more challenging co-speech gesture generation setting, where the gesture data and audio conditions are both sequential, and the audio-to-gesture mapping is implicit. To this end, we propose the Diffusion Audio-Gesture Transformer to guarantee temporally aligned generation. We further propose the Diffusion Gesture Stabilizer to simultaneously achieve diverse and temporally coherent gestures with an annealed noise sampling strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 656, + 397, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 397, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 397, + 671 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Figure 2 depicts an overview of the proposed DiffGesture framework to generate co-speech gestures of high fidelity. In this section, we first introduce the problem" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "10545" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 203, + 217 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 203, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 203, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 203, + 217 + ], + "type": "image", + "image_path": "33f47d693166d371a24e87d4333ff5521d748911384902045219a7b08e41fa48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of the Diffusion Co-Speech Gesture (DiffGesture) Framework. Given the gesture sequence " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": ", we first establish the forward diffusion (purple) and conditional denoising process (green) in gesture space. Then, we devise the Diffusion Audio-Gesture Transformer to attend to the input conditions of initial poses " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{(1:M)}" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": ", speech audio " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": ", time embedding " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": " and corrupted gesture " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": " from multiple modalities (blue). At the diffusion sampling stage (grey), we propose the Diffusion Gesture Stabilizer to eliminate temporal inconsistency with an annealed noise sampling strategy. To further incorporate implicit classifier-free guidance, we jointly train the conditional " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "(1 - p_{\\text{uncond}})" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": " and unconditional " + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "inline_equation", + "content": "(p_{\\text{uncond}})" + }, + { + "bbox": [ + 46, + 227, + 545, + 293 + ], + "type": "text", + "content": " models. This allows us to trade off between diversity and quality during inference." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 72, + 443, + 217 + ], + "blocks": [ + { + "bbox": [ + 205, + 72, + 443, + 217 + ], + "lines": [ + { + "bbox": [ + 205, + 72, + 443, + 217 + ], + "spans": [ + { + "bbox": [ + 205, + 72, + 443, + 217 + ], + "type": "image", + "image_path": "432316dc0fac48cbc18da6b2853d75cd9374a705e72851bedff75f5f61a4a010.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 447, + 72, + 543, + 217 + ], + "blocks": [ + { + "bbox": [ + 447, + 72, + 543, + 217 + ], + "lines": [ + { + "bbox": [ + 447, + 72, + 543, + 217 + ], + "spans": [ + { + "bbox": [ + 447, + 72, + 543, + 217 + ], + "type": "image", + "image_path": "10612c5ffe7a00e2688305fa64cd3490996b86ccc99125c08e3a3803274a9f9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 450 + ], + "type": "text", + "content": "formulation of audio-driven co-speech gesture generation (Section 3.1). We then establish the forward diffusion and the reverse conditional generation process in gesture space (Section 3.2). Furthermore, we elaborate the Diffusion Audio-Gesture Transformer to attend to the conditions from multiple modalities and enhance the speech-gesture correlations with temporal dependency (Section 3.3). To eliminate temporal inconsistency introduced by naive noises, we propose a novel Diffusion Gesture Stabilizer with annealed noise sampling strategies and describe this module in (Section 3.4). Finally, incorporating implicit classifier-free guidance in co-speech gestures is discussed in (Section 3.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 171, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 171, + 476 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 171, + 476 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": "With a large-scale co-speech gesture training corpus, we leverage the speaking videos with clear co-speech upper body movements for model learning. In particular, for each video clip of " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " frames, we extract the accompanying speech audio sequence " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{a} = \\{\\pmb{a}_1,\\dots ,\\pmb{a}_N\\}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " and use the offthe-shelf human pose estimator OpenPose [9] to annotate the per-frame human skeletons as " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x} = \\{\\pmb{p}_1,\\dots ,\\pmb{p}_N\\}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": ". We follow baseline methods [30, 41] to pre-process such skeletal representation as the concatenation of unit direction vectors as " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p_i} = [\\pmb{d}_{i,1},\\pmb{d}_{i,2},\\dots ,\\pmb{d}_{i,J - 1}]" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p_i}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " denotes the pose description coordinates of the " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " is the total joint number and " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{d}_{i,j}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": "-th unit direction vector among the " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " joints of the " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": "-th image frame. The diffusion model's reverse denoising process " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " is optimized to synthesize the human skeleton sequence " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": ", which is further conditioned on the speech audio sequence " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " and the initial poses " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\pmb{p}_1,\\dots ,\\pmb{p}_M\\}" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " of the first " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "text", + "content": " frames. The learning objective of the overall framework can be formulated as " + }, + { + "bbox": [ + 46, + 486, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\theta}||\\pmb{x} - G_{\\theta}(\\pmb {a},\\pmb {p}_1,\\dots ,\\pmb {p}_M)||" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 306, + 537, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 537, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 537, + 319 + ], + "type": "text", + "content": "3.2. Gesture Space Forward and Reverse Process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0 \\in \\mathbb{R}^{N \\times 3(J-1)}" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": " sampled from real data distribution " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_0)" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": ", our goal is to learn a model distribution " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": " that approximates " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_0)" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": ". Specifically, denoising diffusion probabilistic models (DDPMs) [19] define the latent variable models of the form " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_0) = \\int p_{\\theta}(\\pmb{x}_{0:T}) d\\pmb{x}_{1:T}" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{1:T}" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": " are latent variables in the same sample space as " + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 305, + 324, + 545, + 410 + ], + "type": "text", + "content": " with the same dimensionality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "content": "The Forward Diffusion Process. The forward process, which is also termed as the diffusion process, approximates the posterior distribution " + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{1:T}|\\pmb{x}_0)" + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "content": ". It is defined as a Markov chain that gradually adds Gaussian noise to the data sample according to a variance schedule " + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\beta_1,\\dots ,\\beta_T" + }, + { + "bbox": [ + 305, + 411, + 545, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 483, + 545, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 483, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 351, + 483, + 545, + 515 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}\\right), \\tag {1}", + "image_path": "050d3a90eb54587792ccd6cbe18d5a17e8bef55b88e64eda9cbf94583f4d4cae.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 518, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 518, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 518, + 545, + 533 + ], + "type": "interline_equation", + "content": "w h e r e \\quad q (\\boldsymbol {x} _ {t} | \\boldsymbol {x} _ {t - 1}) = \\mathcal {N} (\\boldsymbol {x} _ {t}; \\sqrt {1 - \\beta_ {t}} \\boldsymbol {x} _ {t - 1}, \\beta_ {t} \\boldsymbol {I}). \\tag {2}", + "image_path": "8044d9f2a8d53ac8b1a82cae55af992fb637d1b6cae04ba350a4a53505b419f9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": "The variances " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " are constant hyperparameters to ease the modeling of the reverse process [19]. Through such a corruption scheme, the structural information of the original skeleton is gradually substituted by noises, which finally leads to a pure white noise when " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " goes to infinity. Therefore, the prior latent distribution of " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{x}_T)" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\boldsymbol{x}_T; \\mathbf{0}, \\boldsymbol{I})" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " with only information of Gaussian noise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Reverse Conditional Gesture Generation. The reverse process, which is also termed as the generative process, estimates the joint distribution of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_{0:T})" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". As proved in [15], the reverse process of the continuous diffusion process preserves the same transition distribution form, which motivates us to leverage a Gaussian transition to formulate " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " under an unconditional setting, which" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10546" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 209, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 209, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 209, + 85 + ], + "type": "text", + "content": "approximates the intractable process as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 94, + 287, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 94, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 93, + 94, + 287, + 126 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {x} _ {0: T}\\right) = p _ {\\theta} \\left(\\boldsymbol {x} _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}\\right), \\tag {3}", + "image_path": "c0f0fd0d2108f53ac8ba6c19288b9a789ce0c69cfa738c57b2b80f1094c641e6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 129, + 287, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 129, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 52, + 129, + 287, + 143 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 52, + 129, + 287, + 143 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}_{t - 1}|\\boldsymbol{x}_t) = \\mathcal{N}(\\boldsymbol{x}_{t - 1};\\mu_{\\theta}(\\boldsymbol{x}_t,t),\\Sigma_{\\theta}(\\boldsymbol{x}_t,t))" + }, + { + "bbox": [ + 52, + 129, + 287, + 143 + ], + "type": "text", + "content": " (4)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": "The corrupted noisy data " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": " is sampled by " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t|\\pmb{x}_0) = \\mathcal{N}(\\pmb{x}_t; \\sqrt{\\bar{\\alpha}_t}\\pmb{x}_0, (1 - \\bar{\\alpha}_t)\\pmb{I})" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": ". Note that we set the variances " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\Sigma_\\theta(\\pmb{x}_t, t) = \\beta_t\\pmb{I}" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": " to untrained time-dependent constants. The above diffusion model formulations show compelling performances on unconditional generation. To further adapt to the conditional co-gesture synthesis, we have to provide additional inputs to the model, including the audio and initial poses. Therefore, we treat the speech audio " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": " and initial poses " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{1:M}" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": " as context information " + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 47, + 152, + 289, + 297 + ], + "type": "text", + "content": " and inject conditions into the generation process. The reverse process of each timestep (Eq. 4) can be thus updated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 305, + 287, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 305, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 73, + 305, + 287, + 319 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\boldsymbol {x} _ {t - 1} | \\boldsymbol {x} _ {t}, \\boldsymbol {c}) = \\mathcal {N} (\\boldsymbol {x} _ {t - 1}; \\mu_ {\\theta} (\\boldsymbol {x} _ {t}, t, \\boldsymbol {c}), \\beta_ {t} \\boldsymbol {I}). \\tag {5}", + "image_path": "f6b4f517f0c75e47507469ddafac16cb37fd22b1880ee0740d7b83bc7a9e0524.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": "In this way, we could start the generation process by firstly sample a Gaussian noise " + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\boldsymbol{I})" + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": " and follow the Markov chain to iteratively denoise the latent variable " + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 46, + 327, + 287, + 388 + ], + "type": "text", + "content": " via Eq. 5 to get the final results. The overview of conditional co-speech gesture process is illustrated in Figure 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "text", + "content": "Training Objective. To optimize the overall framework, we optimize the variational lower bound on negative log-likelihood: " + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[-\\log p_{\\theta}(\\pmb{x}_0)] \\leq \\mathbb{E}_q[-\\log \\frac{p_{\\theta}(\\pmb{x}_0)}{q(\\pmb{x}_{1:T}|\\pmb{x}_0)}]" + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "text", + "content": ". We rewrite the loss function conditioned on context " + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "text", + "content": " and eliminate all the constant items that do not require training: " + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "inline_equation", + "content": "L(\\theta) = \\mathbb{E}_q[\\sum_{t=2}^{T} D_{KL}(q(\\pmb{x}_{t-1}|\\pmb{x}_t, \\pmb{x}_0)||p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t, c))]" + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "text", + "content": ". With reparameterization, we can represent each term in " + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "inline_equation", + "content": "L_{\\theta}" + }, + { + "bbox": [ + 47, + 388, + 287, + 498 + ], + "type": "text", + "content": " using MSE loss. We follow [19] to further simplify the training objective to the ensemble of MSE losses as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 517, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 287, + 536 + ], + "type": "interline_equation", + "content": "L (\\theta) = \\mathbb {E} _ {q} [ \\left\\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\boldsymbol {c}, t) \\right\\| ^ {2} ], (6)", + "image_path": "0b3740110531d168c2157b32a8014ffc9a7c6262888357a9637e4611676843f0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": " is uniformly sampled between 1 and " + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": ". As we jointly train the model under conditional and unconditional setting, a trainable masked embedding with probability " + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "inline_equation", + "content": "p_{\\text{uncond}}" + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": " replaces context " + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": " and the diffusion model predicts the noise in the unconditional setting. The detailed principles will be discussed in Section 3.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 247, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 247, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 247, + 635 + ], + "type": "text", + "content": "3.3. Diffusion Audio-Gesture Transformer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "With the naive conditional generation scheme as specified in Section 3.2, we still confront a critical problem in the setting of co-speech gesture generation. Since " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " denotes the skeleton sequence of " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " frames, there exists temporal dependency among the target sequence and context information, making it more complex than time-invariant tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "like image generation. Therefore, how to guarantee temporally coherent results in a non-autoregressive conditional generation process remains an unsolved problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "type": "text", + "content": "In contrast to most previous studies that resort to recurrent networks [30, 41], we propose to make use of the Transformer's strong capacity in sequential data modeling. Specifically, since the noisy gesture sequence " + }, + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "type": "text", + "content": " and the contextual information " + }, + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 110, + 545, + 217 + ], + "type": "text", + "content": " align in the temporal dimension, we concatenate them in the feature channel. In this way, the skeleton and context condition of each frame serve as an individual token, which captures the long-term dependency by the self-attention mechanism:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 230, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 230, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 339, + 230, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\mathsf {T}}}{\\sqrt {\\ell}}\\right) \\mathbf {V}, \\tag {7}", + "image_path": "4692a0a5b2e88aad939f8a12281ec5dcd262eb8e762e46f8e65a7ce41806b562.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K},\\mathbf{V}" + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "text", + "content": " are the query, key, and value matrix from input tokens, " + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "text", + "content": " is the channel dimension, and " + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathsf{T}" + }, + { + "bbox": [ + 304, + 269, + 545, + 329 + ], + "type": "text", + "content": " is the matrix transpose operation. Such a design also avoids severe error accumulation in autoregressive pipelines, enabling us to generate coherent gesture sequences." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 342, + 459, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 459, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 459, + 354 + ], + "type": "text", + "content": "3.4. Diffusion Gesture Stabilizer" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "text", + "content": "In DDPMs, the independent random variables " + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "text", + "content": " introduced at the sampling stage promote diversity and thus systematically improve the task performance. However, the variation in temporal dimension introduced by " + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "text", + "content": ", especially when timestep " + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 362, + 545, + 506 + ], + "type": "text", + "content": " is small in the reverse process, can have a negative effect on temporal consistency. At the inference stage, to achieve the trade-off between diversity and temporal consistency, we propose a novel Diffusion Gesture Stabilizer without extra training expenses under two annealed scenarios, where the term \"annealed\" means that the process is transitioned from high variance and entropy (hot) to low variance and entropy (cold)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": "Thresholding. Since temporally independent Gaussian noises inevitably introduce inconsistency, restricting the temporal variation naturally helps to avoid inconsistency. And hard thresholding serves as an effective trick. In detail, we set a time threshold " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": ", and then use the same " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": " in the naive sampling strategy for " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "t > t_0" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": " and set " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "z = \\{z_0\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "t \\leq t_0" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "z_0 \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": " follows " + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 304, + 508, + 545, + 604 + ], + "type": "text", + "content": " which do not introduce variation in the temporal dimension." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Smooth Sampling. We further modify " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "z(t) = \\{z_i(t)\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to be a smooth annealing version via variance-aware sampling. In the original sampling rule of DDPMs, i.i.d.variables " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_i(t)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " are sampled from " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". With smooth resampling, we first sample " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_0(t) \\sim \\mathcal{N}(\\mathbf{0},\\sigma_a^2(t)\\mathbf{I})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " only once for each timestep " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " in the reverse process, then given " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_0(t)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_i(t)|z_0(t) \\sim \\mathcal{N}(z_0(t),(1 - \\sigma_a^2(t))\\mathbf{I})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "i \\in \\{1,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_a(t) \\in [0,1]" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " is a nondecreasing function to achieve variance annealing." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10547" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 141, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 141, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 141, + 84 + ], + "type": "text", + "content": "Algorithm 1 Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 87, + 267, + 164 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 53, + 87, + 89, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 87, + 89, + 97 + ], + "spans": [ + { + "bbox": [ + 53, + 87, + 89, + 97 + ], + "type": "text", + "content": "1: repeat" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 99, + 172, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 99, + 172, + 108 + ], + "spans": [ + { + "bbox": [ + 53, + 99, + 172, + 108 + ], + "type": "text", + "content": "2: Sample " + }, + { + "bbox": [ + 53, + 99, + 172, + 108 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_0,\\pmb{c})\\sim q(\\pmb{x}_0,\\pmb{c})" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "spans": [ + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "text", + "content": "3: Sample " + }, + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "inline_equation", + "content": "\\tau \\sim \\mathrm{Uniform}(0,1)" + }, + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "text", + "content": ". Set " + }, + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "inline_equation", + "content": "c = \\varnothing" + }, + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 53, + 109, + 267, + 119 + ], + "type": "inline_equation", + "content": "\\tau < p_{\\text{uncond}}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 120, + 200, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 120, + 200, + 130 + ], + "spans": [ + { + "bbox": [ + 53, + 120, + 200, + 130 + ], + "type": "text", + "content": "4: Sample " + }, + { + "bbox": [ + 53, + 120, + 200, + 130 + ], + "type": "inline_equation", + "content": "t \\sim \\mathrm{Uniform}(\\{1, \\dots, T\\})" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 131, + 260, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 131, + 260, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 131, + 260, + 144 + ], + "type": "text", + "content": "5: Compute " + }, + { + "bbox": [ + 53, + 131, + 260, + 144 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\left\\| \\pmb {\\epsilon} - \\pmb{\\epsilon}_{\\theta}(\\sqrt{\\bar{\\alpha}_{t}}\\pmb {x}_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\pmb {\\epsilon},\\pmb {c},t)\\right\\|^{2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 144, + 164, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 144, + 164, + 153 + ], + "spans": [ + { + "bbox": [ + 53, + 144, + 164, + 153 + ], + "type": "text", + "content": "6: Perform gradient descent" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 155, + 122, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 155, + 122, + 164 + ], + "spans": [ + { + "bbox": [ + 53, + 155, + 122, + 164 + ], + "type": "text", + "content": "7: until converged" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 186, + 222, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 222, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 222, + 199 + ], + "type": "text", + "content": "3.5. Implicit Classifier-free Guidance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "text", + "content": "In co-speech gesture literature, the speech-to-gesture mapping is implicit, where the same audio corresponds to diverse gestures and different audiOs could incur the same motion [22, 25], making it difficult to utilize the commonly used explicit classifier guidance [13, 33]. Therefore, how can we further exploit practical guidance for better audio correlations and mode coverage? Our solution to this question is to train an extra \"unconditional\" diffusion model to implicitly guide the generation. Dhariwal et al. [13] first introduce the classifier guidance of cross-entropy gradient " + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}_t}\\log p_\\phi (y|\\pmb {x}_t)" + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "text", + "content": " , where the pretrained classifier is parameterized by " + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "text", + "content": " denotes the classification logits. This gradient term is further scaled by the covariance matrix to modify the mean value of transition distribution in Eq. 4. To adapt to the cases where no explicit guidance is available, we follow [20] to jointly train the conditional and unconditional models, termed as classifier-free guidance. In particular, according to the implicit classifier's property that " + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "inline_equation", + "content": "p(c|x_t)\\propto p(x_t|c) / p(x_t)" + }, + { + "bbox": [ + 47, + 206, + 289, + 446 + ], + "type": "text", + "content": " , we could derive a gradient relationship in the implicit classifier as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 458, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {c} | \\boldsymbol {x} _ {t}) \\propto \\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {x} _ {t} | \\boldsymbol {c}) - \\nabla_ {\\boldsymbol {x} _ {t}} \\log p (\\boldsymbol {x} _ {t}), \\tag {8}", + "image_path": "3c0e642bbdfe18c0856cd773236fe39dc8a6a049210c9786814a4dd71766e843.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "text", + "content": "which is further proportional to " + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}(\\boldsymbol{x}_{t}|\\boldsymbol{c}) - \\epsilon^{*}(\\boldsymbol{x}_{t})" + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "text", + "content": ". Therefore, we use a single Transformer network to parameterize both settings by a mix-up training trick: for the probability of " + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "p_{uncond}" + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "text", + "content": ", we set the context information " + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}" + }, + { + "bbox": [ + 47, + 483, + 287, + 567 + ], + "type": "text", + "content": " as masked embedding to train the unconditional setting, while for other cases, we train the original conditional counterpart. The training is shown in Algorithm 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 568, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 616 + ], + "type": "text", + "content": "Sampling with Classifier-free Guidance. Starting from Gaussian noise, we iteratively remove noises in " + }, + { + "bbox": [ + 47, + 568, + 287, + 616 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 568, + 287, + 616 + ], + "type": "text", + "content": ". As we use implicit classifier-free guidance, similar to Equation 8, the predicted Gaussian noise is modified as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 73, + 628, + 287, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 628, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 73, + 628, + 287, + 642 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\epsilon}} _ {\\theta} = \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) + s \\cdot \\left(\\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {c}, t\\right) - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t)\\right), \\tag {9}", + "image_path": "174a93af0671e4cc4726d2649cf1c2629d1811a3ff20e7875a7630d71654e750.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " is the scale parameter to trade off the diversity and quality. With classifier-free guidance, Algorithm 2 reveals how to generate co-speech gestures given the trained diffusion model via the Diffusion Gesture Stabilizer with the Smooth Sampling annealed scenario." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 72, + 405, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 405, + 84 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 405, + 84 + ], + "type": "text", + "content": "Algorithm 2 Sampling" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 87, + 520, + 212 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 311, + 87, + 477, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 87, + 477, + 99 + ], + "spans": [ + { + "bbox": [ + 311, + 87, + 477, + 99 + ], + "type": "text", + "content": "1: Trained diffusion model " + }, + { + "bbox": [ + 311, + 87, + 477, + 99 + ], + "type": "inline_equation", + "content": "\\theta, \\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 99, + 397, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 99, + 397, + 109 + ], + "spans": [ + { + "bbox": [ + 311, + 99, + 397, + 109 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 311, + 99, + 397, + 109 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 311, + 99, + 397, + 109 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 110, + 501, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 110, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 311, + 110, + 501, + 121 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 311, + 110, + 501, + 121 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\epsilon}}_{\\theta} = \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t) + s\\cdot (\\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,\\pmb {c},t) - \\pmb{\\epsilon}_{\\theta}(\\pmb{x}_t,t))" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 121, + 416, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 121, + 416, + 132 + ], + "spans": [ + { + "bbox": [ + 311, + 121, + 416, + 132 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 311, + 121, + 416, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0(t) \\sim \\mathcal{N}(\\mathbf{0}, \\sigma_a^2(t)\\mathbf{I})" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 132, + 407, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 132, + 407, + 142 + ], + "spans": [ + { + "bbox": [ + 311, + 132, + 407, + 142 + ], + "type": "text", + "content": "5: for " + }, + { + "bbox": [ + 311, + 132, + 407, + 142 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,N" + }, + { + "bbox": [ + 311, + 132, + 407, + 142 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 143, + 461, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 143, + 461, + 154 + ], + "spans": [ + { + "bbox": [ + 311, + 143, + 461, + 154 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 311, + 143, + 461, + 154 + ], + "type": "inline_equation", + "content": "\\pmb{z}_i(t) \\sim \\mathcal{N}(\\pmb{z}_0(t), (1 - \\sigma_a^2(t))\\pmb{I})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 311, + 155, + 360, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 155, + 360, + 163 + ], + "spans": [ + { + "bbox": [ + 311, + 155, + 360, + 163 + ], + "type": "text", + "content": "7: end for" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "spans": [ + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{z}(t) = \\{\\mathbf{z}_1(t),\\dots ,\\mathbf{z}_N(t)\\}" + }, + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "text", + "content": ", else " + }, + { + "bbox": [ + 311, + 164, + 520, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{z}(t) = \\mathbf{0}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 176, + 485, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 176, + 485, + 192 + ], + "spans": [ + { + "bbox": [ + 311, + 176, + 485, + 192 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 311, + 176, + 485, + 192 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}} \\left( \\pmb{x}_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\bar{\\alpha}_t}} \\hat{\\pmb{\\epsilon}}_\\theta \\right) + \\sigma_t \\pmb{z}(t)" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 192, + 351, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 192, + 351, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 192, + 351, + 201 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 202, + 361, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 202, + 361, + 212 + ], + "spans": [ + { + "bbox": [ + 308, + 202, + 361, + 212 + ], + "type": "text", + "content": "11: return " + }, + { + "bbox": [ + 308, + 202, + 361, + 212 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 231, + 387, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 387, + 244 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 387, + 244 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 251, + 460, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 460, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 460, + 263 + ], + "type": "text", + "content": "4.1. Co-Speech Gesture Datasets" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": "TED Gesture. As a large-scale dataset for gesture generation research, TED Gesture dataset [41, 42] contains 1,766 TED videos of different narrators covering various topics. We follow the data process in former works [30, 41], where the poses are resampled with 15 FPS, and frame segments of length 34 are obtained with a stride of 10." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 341, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 437 + ], + "type": "text", + "content": "TED Expressive. While the poses in TED Gesture only contain 10 upper body key points without vivid finger movements, the TED Expressive dataset [30] is further expressive of both finger and body movements. The state-of-art 3D pose estimator ExPose [12] is used to fully capture the pose information in data. As a result, TED Expressive annotates the 3D coordinates of 43 keypoints, including 13 upper body joints and 30 finger joints." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 444, + 432, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 432, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 432, + 456 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 462, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 666 + ], + "type": "text", + "content": "Comparison Methods. We compare our method on two benchmark datasets with the state-of-the-art methods in recent years. 1) Attention Seq2Seq [42] elaborates on the attention mechanism to generate pose sequences from speech text. 2) Speech2Gesture [16] uses spectrums of the speech audio segments as the input and generates speech gestures adversarially. 3) Joint Embedding [3] maps text and motion to the same embedding space, then generates outputs from motion description text. 4) Trimodal [41] serves as a strong baseline that learns from text, audio, and speaker identity to generate gestures, outperforming former methods by a large margin. 5) HA2G [30] introduces a hierarchical audio learner that captures information across different semantic granularities, achieving state-of-the-art performances. This method hierarchically extracts rich features at the cost of heavier GPU memory overhead, while our method requires much smaller expenses." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. For all the methods in both datasets, we set " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 34" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "-frame pose sequences where the first " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " frames are used for reference, termed as initial poses. There are " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " upper body joints in" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10548" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 70, + 482, + 205 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 482, + 205 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 482, + 205 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 482, + 205 + ], + "type": "table", + "html": "
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
Ground Truth00.698108.52500.703178.827
Attention Seq2Seq [42]18.1540.19682.77654.9200.152122.693
Speech2Gesture [16]19.2540.66893.80254.6500.679142.489
Joint Embedding [3]22.0830.20090.13864.5550.130120.627
Trimodal [41]3.7290.667101.24712.6130.563154.088
HA2G [30]3.0720.672104.3225.3060.641173.899
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
", + "image_path": "30355dba6f476dc87ddb782b7dcc713fa2118a63436183d44d452990bdecd6ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 214, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 545, + 237 + ], + "type": "text", + "content": "Table 1. The Quantitative Results on TED Gesture [41] and TED Expressive [30]. We compare the proposed diffusion-based method against recent SOTA methods [3, 16, 30, 41, 42] and ground truth. For FGD, the lower, the better; for other metrics, the higher, the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": "all the frames of pose sequences, where " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "J = 10" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " for TED Gesture and " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "J = 43" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " for TED Expressive. Following [41], to eliminate the effect of the joint lengths and root motion, we represent the joints' positions using " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "J - 1" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " directional vectors normalized to the unit vectors and train the model to learn the directional vectors. For the audio processing, we use the same audio encoder used in [41] to extract the feature of the raw audio clips directly. The audio clips are encoded as " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " audio feature vectors of 32-D. The audio feature and initial poses are concatenated to form the conditional context information of the diffusion model. For the diffusion process, the number of timesteps is " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": ", and the variances increase linearly from " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 1e - 4" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\beta_{T} = 0.02" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": ". For the Stabilizer, " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " can be adjusted from 20-30 for Thresholding, and a quadratic non-increasing function " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\sigma_{a}(t)" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": " is applied for Smooth Sampling. The hidden dimension of the transformer blocks, is set to 256 for TED Gesture and 512 for TED Expressive. We use 8 Transformer blocks, each of which comprises a multi-head self-attention block and a Feed-Forward Network. We use an Adam optimizer, and the learning rate is " + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 46, + 258, + 289, + 533 + ], + "type": "text", + "content": ". It takes 10 hours to train the model on TED Gesture and 20 hours on TED Expressive on a single NVIDIA GeForce RTX 3090 GPU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 539, + 160, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 160, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 160, + 550 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 581 + ], + "type": "text", + "content": "In evaluation, we use three metrics that are used in cospeech gesture generation and relative fields [24, 30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "text", + "content": "Fréchet Gesture Distance (FGD). Similar to the Fréchet Inception Distance (FID) metric [18], which is widely applied in image generation studies, FGD is used to measure the distance between the synthesized gesture distribution and the real data distribution. Yoon et al. [41] define FGD by training a skeleton sequence auto-encoder to extract the features of the real gesture sequences " + }, + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "text", + "content": " and the features of the generated gesture sequences " + }, + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 681, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 681, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 681, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\operatorname {F G D} (X, \\hat {X}) = \\| \\mu_ {r} - \\mu_ {g} \\| ^ {2} + \\operatorname {T r} (\\Sigma_ {\\mathrm {r}} + \\Sigma_ {\\mathrm {g}} - 2 (\\Sigma_ {\\mathrm {r}} \\Sigma_ {\\mathrm {g}}) ^ {1 / 2}),", + "image_path": "dfc9e8eb5e96a029d6b74ed9f6cc4f1653923b5e790ed6cfbf428390eddf03ba.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mu_r" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma_r" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": " are the first and the second moments of the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "content": "latent feature distribution of the real gestures " + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mu_g" + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\Sigma_g" + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "content": " are the first and the second moments of the latent feature distribution of the generated gestures " + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 304, + 258, + 545, + 319 + ], + "type": "text", + "content": ". We intuitively find that among the three metrics, FGD tells the most whether the generated pose sequences are of high quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": "Beat Consistency Score (BC). Proposed in [24, 26], BC measures motion-audio beat correlation. Considering that the kinematic velocities vary from different joints, we use the change of included angle between bones to track motion beats following [30]. Specifically, we can calculate the mean absolute angle change (MAAC) of angle " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": " in adjacent frames by " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{MAAC}(\\theta_j) = \\frac{1}{S}\\frac{1}{T - 1}\\sum_{s = 1}^{S}\\sum_{t = 1}^{T - 1}\\| \\theta_{j,s,t + 1} - \\theta_{j,s,t}\\| _1" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": " denotes the total number of clips in the dataset, " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": " denotes the number of frames in each clip, and " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\theta_{j,s,t}" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": " is the included angle between the " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": "-th and the " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "(j + 1)" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": "-th bone of the " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": "-th clip at time-step " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": ". Then, we can compute the angle change rate of frame " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": "-th clip as " + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\frac{1}{J - 1}\\sum_{j = 1}^{J - 1}(\\| \\theta_{j,s,t + 1} - \\theta_{j,s,t}\\| _1 / \\mathrm{MAAC}(\\theta_j))" + }, + { + "bbox": [ + 304, + 319, + 546, + 548 + ], + "type": "text", + "content": ". Then we extract the local optima whose first-order difference is higher than a threshold to get kinematic beats, which are used to compute BC later. Following [24] to detect audio beat by onset strength [14], we compute the average distance between each audio beat and its nearest motion beat as Beat Consistency Score:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 554, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 325, + 554, + 545, + 586 + ], + "type": "interline_equation", + "content": "\\mathrm {B C} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\exp \\left(- \\frac {\\min _ {\\forall t _ {j} ^ {y} \\in B ^ {y}} \\| t _ {i} ^ {x} - t _ {j} ^ {y} \\| ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {10}", + "image_path": "a6f4fab3a6f6055142628b2d8e17bf1a6f4190f519d51e01b745ac8916eeec89.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "inline_equation", + "content": "t_i^x" + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "text", + "content": "-th audio beats, " + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "inline_equation", + "content": "B^y = \\{t_i^y\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "text", + "content": " is the set of the kinematic beats, and " + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 594, + 545, + 630 + ], + "type": "text", + "content": " is a parameter to normalize sequences, set to 0.1 empirically." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Diversity. This metric evaluates the variations among generated gestures corresponding to various inputs [23]. We use the same feature extractor when measuring FGD to map synthesized gestures into latent feature vectors and calculate the mean feature distance. In detail, we randomly pick 500 generated samples and compute the mean absolute error between the features and the shuffled features." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10549" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 370 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 370 + ], + "type": "image", + "image_path": "cdacf38770acbf609b3085c1b3fa88d95f5e6ed1d2b26ee14dea89b7e1612cba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 546, + 411 + ], + "type": "text", + "content": "Figure 3. Visualization Results of Our DiffGesture on Two Datasets. Three cases are picked up, where (i) and (ii) are TED Expressive cases, and (iii) is a TED Gesture case. We highlight dull cases generated by comparison methods with rectangles, indicating the mode collapse phenomenon of baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 79, + 421, + 515, + 483 + ], + "blocks": [ + { + "bbox": [ + 79, + 421, + 515, + 483 + ], + "lines": [ + { + "bbox": [ + 79, + 421, + 515, + 483 + ], + "spans": [ + { + "bbox": [ + 79, + 421, + 515, + 483 + ], + "type": "table", + "html": "
MethodsGTS2S. [42]S2G. [16]Joint. [3]Tri. [41]HA2G [30]DiffGesture(Ours)
Naturalness4.331.222.561.223.223.674.00
Smoothness3.943.501.613.443.443.393.89
Synchrony4.001.673.171.393.283.443.89
", + "image_path": "b81460fe9c6083b675e82904ef61c0818188974a7077ca73577a694693c53289.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 491, + 543, + 503 + ], + "lines": [ + { + "bbox": [ + 48, + 491, + 543, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 543, + 503 + ], + "type": "text", + "content": "Table 2. User Study Results. The ratings of motion naturalness, smoothness, and synchrony, are on a scale of 1-5, with 5 being the best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 523, + 157, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 157, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 157, + 534 + ], + "type": "text", + "content": "4.4. Evaluation Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 541, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 288, + 685 + ], + "type": "text", + "content": "Quantitative Results. We compare our method with all the baselines with three metrics on TED Gesture and TED Expressive. The results are shown in Table 1. For the metrics of Ground Truth, we report the values in our implementation. For TED Gesture, we report FGD of all baselines in [30] and evaluate BC and Diversity on our own†. For TED Expressive, all the results of baselines are reported from [30]. Assuming the pseudo ground truth pose follows the real distribution, the FGD of Ground Truth in the table is 0. It is observed that our DiffGesture achieves state-of-the-art performance on both datasets, especially outperforming existing methods by a large margin on TED" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 523, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 704 + ], + "type": "text", + "content": "Expressive. Besides, since BC and Diversity are proposed to measure motion-audio beat correlation and variation, these two metrics of Ground Truth cannot be treated as upper bounds, and it is worth noting that our results may be higher than the ones of Ground Truth, indicating that the generated gestures are of high quality. Qualitative Results. We show the keyframes of all the methods on two datasets in Figure 3. Since TED Expressive requires a higher ability of generative models, we pick up two cases for TED Expressive and one for TED Gesture. For each case, we select three keyframes (an early, a middle, and a late one) to show the pose motions. Comparison methods tend to generate slow and invariant poses and sometimes produce unreliable and stiff results. In contrast, DiffGesture produces diverse human-like poses without resulting in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": " Since there exists an evaluation bug for the BC metric in HA2G [30], we report the re-implemented results from Liu et al." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10550" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 494, + 159 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 494, + 159 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 494, + 159 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 494, + 159 + ], + "type": "table", + "html": "
MethodsTED Gesture [41]TED Expressive [30]
FGD ↓BC ↑Diversity ↑FGD ↓BC ↑Diversity ↑
DiffGesture Base2.4500.632104.6883.8220.707174.377
DiffGesture w/o Stabilizer2.2190.674105.1922.7920.721180.125
DiffGesture w/o classifier-free1.8100.673105.6443.3260.717178.245
DiffGesture (Ours)1.5060.699106.7222.6000.718182.757
", + "image_path": "5270cdced9380e9fc35b6ff2fb95e16427276d509b41d9957885b607c2c153e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 212, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 284 + ], + "type": "text", + "content": "mean poses which are slow and rigid. Besides, as a primary drawback of GAN-based methods, mode collapse makes comparison methods often produce a single type of output, which is severe for pose motion generation. Such a phenomenon is shown in Fig. 3, where the generated frames with nearly the same pose are highlighted with rectangles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 288, + 477 + ], + "type": "text", + "content": "User Study. To better validate the qualitative performance, we conduct a user study on the generated co-speech gestures. The study involves 18 participants, with 9 females and 9 males in the age range of 18-25 years old. The participants are required to grade the motion's quality and coherence, and all the clips are without labels. In total, we pick up 30 cases, 20 for TED-Expressive and 10 for TED Gesture. For each case, we show 7 videos with the order of the methods shuffling, including the ground truth. We adopt the Mean Opinion Scores rating protocol, and each participant is required to rate three aspects of generated motions: Naturalness; Smoothness; Synchrony between speech and generated gestures. The results are shown in Table 2 where the ratings are on a scale of 1 to 5, with 5 being the best. Our participants widely accept that our method produces high-fidelity results in all three aspects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 488, + 148, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 148, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 148, + 500 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": "Ablation Study on the Proposed Modules. To demonstrate the effectiveness of our proposed DiffGesture, we present ablation studies on the key modules in the framework. In detail, we conduct experiments as follows. 1) DiffGesture Base means we only use the proposed conditional diffusion generation process without further design. 2) DiffGesture w/o classifier-free, where classifier-free guidance is not implemented at both the training stage and inference stage. 3) DiffGesture w/o Stabilizer, where Diffusion Gesture Stabilizer is removed at the inference stage. The results are reported in Table 3. The results illustrate the effectiveness of the designed Diffusion Gesture Stabilizer and the implicit classifier-free guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Ablation Study on the Network Architectures. We investigate the performance of the GRU architecture in diffusion models, autoregressively generating poses in [30, 41]. We replace the Diffusion Audio-Gesture Transformer with the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 209, + 538, + 288 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 191 + ], + "type": "text", + "content": "Table 3. Ablation Study on the Proposed Modules. We investigate effectiveness of the proposed modules, Diffusion Gesture Stabilizer and implicit classifier-free guidance. The results indicate that the proposed modules consistently improve performance on the benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 209, + 538, + 288 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 538, + 288 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 538, + 288 + ], + "type": "table", + "html": "
MethodsFGD ↓BC ↑Diversity ↑
GRU on Da14.3430.65898.472
Transformer on Da1.5060.699106.722
GRU on Db17.4520.680172.168
Transformer on Db2.6000.718182.757
", + "image_path": "e9289775af7ba2097cadb42b6bedc478d733a0bcbebbd5d5ee1363dbae51675d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "text", + "content": "Table 4. Ablation Study on the Network Architectures. We compare the performance of GRU and Transformer for diffusion-based backbone on TED Gesture " + }, + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "inline_equation", + "content": "(D_{a})" + }, + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "text", + "content": " and TED Expressive " + }, + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "inline_equation", + "content": "(D_{b})" + }, + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 347, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 456 + ], + "type": "text", + "content": "GRU in the diffusion model. All the context inputs remain the same as our method and are concatenated before the diffusion network. Results are shown in Table 4. Though GRU serves as a strong baseline network in co-gesture learning, it fails to generate high-performance data like our designed Transformer-based network, which indicates the effectiveness of our Transformer-based network and that applying diffusion models in the audio-driven conditional generation is a non-trivial task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 378, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 605 + ], + "type": "text", + "content": "In this work, we present a novel diffusion-based framework DiffGesture for co-speech gesture generation. To generate coherent gestures with strong audio correlations, we propose the Diffusion Audio-Gesture Transformer with the Diffusion Gesture Stabilizer to better attend to the condition information. Such a non-autoregressive pipeline helps to efficiently generate results and reduce error accumulation. We hope our method offers a new perspective for diffusion-based temporal generation and how to capture sequential cross-modal dependencies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgement. The work described in this paper was partially supported by grants from the Research Grants Council of the Hong Kong Special Administrative Region, China (T45-401/22-N), the National Natural Science Fund (62201483), HKU Seed Fund for Basic Research (202009185079 and 202111159073), RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10551" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja, Dong Won Lee, Ryo Ishii, and Louis-Philippe Morency. No gestures left behind: Learning relationships between spoken language and freeform gestures. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1884–1895, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja, Dong Won Lee, Yukiko I Nakano, and Louis-Philippe Morency. Style transfer for co-speech gesture animation: A multi-speaker conditional-mixture approach. In European Conference on Computer Vision, pages 248–265. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "text", + "content": "[3] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[4] Simon Alexanderson, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, volume 39, pages 487-496. Wiley Online Library, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "text", + "content": "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6):1-19, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 423 + ], + "type": "text", + "content": "[6] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 456 + ], + "type": "text", + "content": "[7] Judee K Burgoon, Thomas Birk, and Michael Pfau. Nonverbal behaviors, persuasion, and credibility. Human communication research, 17(1):140-169, 1990. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 491 + ], + "type": "text", + "content": "[8] B. Butterworth and U. Hadar. Gesture, speech, and computational stages: A reply to mcneill. Psychological Review, 96(1):168-174, 1989. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 492, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 287, + 545 + ], + "type": "text", + "content": "[9] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: realtime multi-person 2d pose estimation using part affinity fields. IEEE transactions on pattern analysis and machine intelligence, 43(1):172-186, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 286, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 591 + ], + "type": "text", + "content": "[10] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 668 + ], + "type": "text", + "content": "[11] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[12] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision (ECCV), 2020. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 149 + ], + "type": "text", + "content": "[14] Daniel Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36:51-60, 03 2007. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[15] William Feller. On the theory of stochastic processes, with particular reference to applications. In *Berkeley Symposium on Mathematical Statistics and Probability*, pages 403-432, 1949. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 250 + ], + "type": "text", + "content": "[16] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 251, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 316 + ], + "type": "text", + "content": "[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 316, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 392 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 425 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 503 + ], + "type": "text", + "content": "[21] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In 2012 7th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 25-32. IEEE, 2012. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 559 + ], + "type": "text", + "content": "[22] Dong Won Lee, Chaitanya Ahuja, and Louis-Philippe Morency. Crossmodal clustered contrastive learning: Grounding of spoken language to gesture. In GENEA: Generation and Evaluation of Non-verbal Behaviour for Embodied Agents Workshop 2021, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 624 + ], + "type": "text", + "content": "[23] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 625, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 679 + ], + "type": "text", + "content": "[24] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[25] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10552" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[26] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 240 + ], + "type": "text", + "content": "[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 612-630. Springer, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 308 + ], + "type": "text", + "content": "[28] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 342 + ], + "type": "text", + "content": "[29] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. arXiv preprint arXiv:2212.02350, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 419 + ], + "type": "text", + "content": "[30] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10462-10472, June 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 421, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 476 + ], + "type": "text", + "content": "[31] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 25-35, 2013. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 477, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 498 + ], + "type": "text", + "content": "[32] David McNeill. Hand and mind. De Gruyter Mouton, 2011. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 588 + ], + "type": "text", + "content": "[33] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "text", + "content": "[34] Shenhan Qian, Zhi Tu, Yihao Zhi, Wen Liu, and Shenghua Gao. Speech drives templates: Co-speech gesture synthesis with learned templates. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11077-11086, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[36] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 418 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 118 + ], + "type": "text", + "content": "Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "type": "text", + "content": "[38] P. Wagner, Z. Malisz, and S. Kopp. Gesture and speech in interaction: An overview. Speech Communication, 57:209-232, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "text", + "content": "[39] Jing Xu, Wei Zhang, Yalong Bai, Qibin Sun, and Tao Mei. Freeform body motion generation from speech. arXiv preprint arXiv:2203.02291, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "text", + "content": "[40] Payam Jome Yazdian, Mo Chen, and Angelica Lim. Gesture2vec: Clustering gestures using representation learning methods for co-speech gesture generation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3100-3107. IEEE, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[41] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 353, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 418 + ], + "type": "text", + "content": "[42] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10553" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_content_list.json b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6e353d683eb73e500c7cfda8f73cd6d327a9228 --- /dev/null +++ b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_content_list.json @@ -0,0 +1,1768 @@ +[ + { + "type": "text", + "text": "Tangentially Elongated Gaussian Belief Propagation for Event-based Incremental Optical Flow Estimation", + "text_level": 1, + "bbox": [ + 212, + 130, + 754, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Nagata† and Yusuke Sekikawa† DENSO IT LAB., INC., Japan.", + "bbox": [ + 217, + 202, + 750, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 264, + 255, + 341, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optical flow estimation is a fundamental functionality in computer vision. An event-based camera, which asynchronously detects sparse intensity changes, is an ideal device for realizing low-latency estimation of the optical flow owing to its low-latency sensing mechanism. An existing method using local plane fitting of events could utilize the sparsity to realize incremental updates for low-latency estimation; however, its output is merely a normal component of the full optical flow. An alternative approach using a frame-based deep neural network could estimate the full flow; however, its intensive nonincremental dense operation prohibits the low-latency estimation. We propose tangentially elongated Gaussian (TEG) belief propagation (BP) that realizes incremental full-flow estimation. We model the probability of full flow as the joint distribution of TEGs from the normal flow measurements, such that the marginal of this distribution with correct prior equals the full flow. We formulate the marginalization using a message-passing based on the BP to realize efficient incremental updates using sparse measurements. In addition to the theoretical justification, we evaluate the effectiveness of the TEGBP in real-world datasets; it outperforms SOTA incremental quasi-full flow method by a large margin. (The code is available at https://github.com/DensoITLab/tegbp/)", + "bbox": [ + 75, + 287, + 473, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optical flow estimation, which computes the correspondence of pixels in different time measurements, is a fundamental building block of computer vision. One needs to estimate the flow at low latency in many practical applications, such as autonomous driving cars, unmanned aerial vehicles, and factory automation robots. Most of the existing optical flow algorithm utilizes dense video frames; it computes the flow by searching the similar intensity pattern [15, 29]. Recently, methods using deep neural network (DNN) [29] demonstrate impressive accuracy at the", + "bbox": [ + 75, + 722, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0553784f9a4b0d0b55180ab1194f8293bc38feac0ef3b0f32b87927b1e320d8b.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed TEGBP. We model a belief about the full flow from a normal flow measurement using TEG (RGB ellipse). The mean of the marginal distribution of each TEG with an appropriate prior equals the full flow (magenta arrow). The marginal (magenta ellipse) is computed incrementally using local message passing (RGB arrows) based on BP." + ], + "image_footnote": [], + "bbox": [ + 501, + 244, + 893, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cost of higher-computational cost. Either model-based or DNN-based, the frame-based algorithm needs to compute the entire pixel for every frame, even when there are subtle changes or no changes at all. This dense operation makes it difficult to realize low-latency estimation, especially on the resource-constrained edge device.", + "bbox": [ + 496, + 470, + 890, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The event-based camera is a bio-inspired vision sensor, which asynchronously detects intensity change on each pixel. Thanks to the novel sensing mechanism, the camera equips favorable characteristics for optical flow estimation, such as high dynamic range (HDR), blur-free measurement, and, most importantly, sparse low-latency data acquisition. Many researchers have explored the way to utilize sparsity to realize efficient low-latency estimation; one extends the well-known Lucas-Kanade algorithm [7], and the other exploits the local planer shape of the spatiotemporal event streams [6]. These methods could utilize the sparsity for efficient incremental processing; however, the optical flow computed in this way (e.g., by plane fitting) is the normal flow, which is a normal component of full flow and often different from them1 we want to obtain. The normal flow is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient). Some work tried to recover the full flow from the normal flow [2]; however, it does not precisely equal the full flow (refer to Sec. 2). There exist methods that could estimate full flow, such as a varia", + "bbox": [ + 496, + 561, + 892, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† These authors contributed equally to this work.", + "bbox": [ + 99, + 886, + 359, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 full flow is usually simply called optical flow or flow, yet, we use full flow when we want to highlight the difference with the normal flow.", + "bbox": [ + 498, + 875, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21940", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tional method [5], a multi-scale extension of contrast maximization [25], or a frame-based DNN [14]. However, they need to apply non-incremental dense operation for all the pixels of the event frame (dense representation constructed from sparse events) for every frame, which prohibits the low-latency estimation on the edge device.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our research goal is to realize an incremental full flow algorithm from sparse normal flow measurements. To this end, we propose Tangentially Elongated Gaussian (TEG) Belief Propagation (BP). We compute the full flow using the normal flow measurements, which can be observed directly from an event camera or computed cheaply using an existing algorithm2. Notice that given a single measurement of normal flow, there are infinite possibilities for the full flow along the tangential direction of the normal flow. We model this uncertainty using the TEG, Gaussian distribution, which has a large variance along the tangential direction of the normal flow. The probability density of full flow on each pixel is given as the margins of the joint distribution of TEG data factor and some prior factor on a sparse graph (Fig. 1, Sec. 3.3.2). We leverage the sparse graph to formulate the incremental full flow estimation algorithm using message-passing based on BP [9]. We evaluate the effectiveness of the TEGBP on real-world data captured from aerial drones and automobiles. TEGBP outperforms SOTA incremental method [2] by a large margin.", + "bbox": [ + 75, + 181, + 473, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 500, + 215, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section reviews the related literature on optical flow estimation using sparse event signals.", + "bbox": [ + 76, + 523, + 468, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Incremental Algorithm for Normal Flow", + "text_level": 1, + "bbox": [ + 76, + 564, + 423, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lucas-Kanade based method The traditional frame-based Lucas-Kanade method finds the flow by aligning the small patch of intensity between consecutive frames using its spatial gradient and temporal change. The event-based counterpart [7] uses the integration of events to approximate them for each incoming event. This approach realized highly efficient estimation by utilizing the sparsity of input events. However, it can only estimate the normal flow, which is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient) [11].", + "bbox": [ + 75, + 588, + 468, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Plane fitting based method When an edge undergoes liner motion, triggered events generate a time surface, which can be locally approximated by a plane in a spatiotemporal space. The local plane fitting method [3,6,17, 19] computes the optical flow by fitting the plane to the sparse event points. However, the flow obtained by plane fitting is also the normal component of the full flow.", + "bbox": [ + 75, + 758, + 470, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a method for computing the quasi-full flow has been proposed [2]; it estimates the full flow by considering the average normal flow in multiple resolutions. It incrementally updates the average and selects the flow with the largest average norm from the multiple resolutions. It expects the flow to have the largest norm to correspond to the full flow. The average operation induces bias on the estimated flow; therefore, it equals the full flow on the very limited scene where the true motion in the window is orthogonal to the edge direction (refer to Sec.5 in their paper). We'll also experimentally compare this point in Sec. 4.3.", + "bbox": [ + 500, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, current sparse incremental algorithms are limited to normal flow.", + "bbox": [ + 500, + 257, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Dense Algorithm for Full Flow", + "text_level": 1, + "bbox": [ + 500, + 297, + 769, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Variational method A variational optimization-based approach realizes the full flow estimation by incorporating intensity frame obtained either by the event/frame hybrid camera [22] or by simultaneously estimating image intensity [5]. It could estimate full flow; however, it requires additional intensity observation or intensive dense optimization to recover the intensity frame.", + "bbox": [ + 498, + 319, + 890, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrast maximization method Contrast maximization (CMax) [12, 27] computes the underlying motion (optical flow) of events stream by finding their alignment. It computes the alignment by maximizing the focus score (e.g., variance) of an image generated by warping the events (IWE) using the optical flow parameter. The original algorithm can not be used to estimate pixel-wise flow because it tends to converge to the degenerated solution (e.g., warps all events to a single point) [24, 35]. Recently, a hierarchical CMax approach avoiding the degenerated solution for pixel-wise estimation has been proposed [25]. Although this approach works well even in the practical scenario, it is not incremental; it must recompute the IWE for each sliding time window by using all events in the patch for every hierarchical scale (this involves intensive iteration by itself), making it difficult to operate in real-time.", + "bbox": [ + 498, + 444, + 890, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DNN based method Frame-based DNN is often utilized to compute the full flow [14, 31, 32, 34]. Recently, a frame-based high-accuracy model called RAFT [29] has been imported into the event-based vision literature showing the SOTA performance for the dense flow estimation [14]. These methods can not consume sparse events directly; the sparse input must be converted into dense frame representation. The dense convolution is computationally expensive because it needs to operate for all pixels, even when there are subtle changes or no changes at all, making it difficult to operate at a higher rate on the edge device.", + "bbox": [ + 498, + 703, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, some dense algorithms could estimate full flow; however, it compromises the event's sparse asyn", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2We consider the normal flow as measurement and discuss algorithms to process the sparse measurement of normal flow to estimate the full flow.", + "bbox": [ + 75, + 875, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21941", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "chronous nature, and their non-incremental dense operation prohibits them from realizing low-latency processing.", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Sparse Asynchronous Computing Architecture", + "text_level": 1, + "bbox": [ + 76, + 133, + 467, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Events are asynchronously triggered only from limited pixels which detect intensity changes. Several techniques have been developed for efficient processing of the sparse signals on CPUs by taking advantage of the highly optimized cache mechanism [16, 28]. Besides, sparse data-flow architectures [8, 9], which differ from prevalent Neuman-type computing architectures, are gaining attention for sparse signal processing. For example, bundle adjustment has been solved on the novel architecture using an algorithm based on Gaussian BP [21]. Our goal is to establish the incremental full flow estimation algorithm by taking advantage of the sparsity of events, which is expected to run efficiently on CPUs or emerging data-flow processors.", + "bbox": [ + 75, + 157, + 470, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed method", + "text_level": 1, + "bbox": [ + 76, + 369, + 246, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 395, + 205, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Normal flow", + "text_level": 1, + "bbox": [ + 76, + 419, + 217, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The normal flow $\\mathbf{v}_i^\\perp \\in \\mathbb{R}^2$ on pixel $i$ is a normal component of full flow $\\mathbf{v}_i \\in \\mathbb{R}^2$ , which we want to estimate. One can not decide the full flow uniquely from the normal flow, and there are infinitely many possibilities along the line perpendicular to the normal flow (Fig. 3 left). There are several options for obtaining the normal flow; some event-based cameras, such as CeleX-V [26] directly observe it, or it can be cheaply computed from the sparse events (Sec. 2.1).", + "bbox": [ + 75, + 443, + 468, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The moving edge in 3D space triggers events; we can compute the normal flow by finding the local plane on the triggered events. In this study, we adopt a naive least-squares fitting of the plane using events in a small spatiotemporal window. An erroneous normal flow may significantly deteriorate the full flow accuracy, especially on highly textured regions such as leaves on trees. One can improve the results by utilizing an advanced noise removal algorithm [4] or by adopting more sophisticated normal flow estimation methods, such as small neural networks.", + "bbox": [ + 75, + 566, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Belief propagation", + "text_level": 1, + "bbox": [ + 76, + 739, + 259, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Factor graphs geometrically represent the structure of probabilistic problems. A factor graph $G = (V, F, E)$ is composed of a set of variable nodes $V = \\{\\mathbf{v}_i\\}_{i=1:N_v}$ , a set of factor nodes $F = \\{f_s\\}_{s=1:N_f}$ and a set of edges $E$ . Each factor node $f_s$ represents a probabilistic constraint $f_s(V_s)$ between a subset of variables $V_s \\subset V$ . The factorization of the variables is explicitly represented in the graph by connecting factor nodes with the variable nodes they depend on. Probabilistically, these factors are the independent", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59f55debec1fc390bc0d1fbbc73c88e7024f8e93740ded52eff28fa7c4c044c8.jpg", + "image_caption": [ + "Figure 2. Processing pipeline. Observation set at $t$ is represented as $V_{t}^{\\perp}$ where the normal flow is sparsely assigned. The TEGBP incrementally update its estimation about full flow $V_{t + 1}$ using previous estimation $V_{t}$ and new observation $\\mathbf{v}_{i,t + 1}^{\\perp}$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 79, + 885, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "terms that make up the joint distribution:", + "bbox": [ + 500, + 296, + 769, + 310 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (V) = \\prod_ {s = 1} ^ {N _ {f}} f _ {s} \\left(V _ {s}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 321, + 890, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Belief propagation (BP) is a generic distributed algorithm for computing the marginal distribution of a set of variables from their joint distribution. The marginal for a single variable $\\mathbf{v}_i$ is computed by the integral over all the other variables. BP computed the marginals asynchronously using message-passing between connected nodes on the graph. It is substantially more efficient by leveraging the topology of the sparse graph than naive integration of the entire graph (which involves giant matrix inversion). Refer to the seminal work of [9] for more detail about the derivation, condition for convergence guarantee, and discussion about the stability.", + "bbox": [ + 498, + 380, + 890, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem formulation", + "text_level": 1, + "bbox": [ + 500, + 571, + 697, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given measurements of normal flow $V_{t}^{\\perp}$ on the the pixel grid where each normal flow is assigned to the corresponding pixels (within time interval $[t - \\tau, t]$ ), we want to estimate the posterior distribution of the corresponding full flow $p(V_{t} \\mid V_{t}^{\\perp})$ . Once the posterior distribution of full flow is obtained, the actual full flow output is given by the maximum a posteriori (MAP) estimate as:", + "bbox": [ + 498, + 594, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} _ {t} = \\underset {V _ {t}} {\\arg \\max } p \\left(V _ {t} \\mid V _ {t} ^ {\\perp}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 709, + 890, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to formulate an incremental full flow estimation algorithm using sparse observations of local normal flows (Fig. 2). Solving the arg max of the joint distribution of $V_{t}$ for every incoming normal flow measurement is infeasible both in terms of memory and computational perspective; therefore, we want to formulate $f$ as an incremental sparse update function, i.e., new observation interacts sparsely with the previous results $p(V_{t} \\mid V_{t}^{\\perp})$ as follows:", + "bbox": [ + 498, + 747, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf: \\left[ p \\left(V _ {t} \\mid V _ {t} ^ {\\perp}\\right), p \\left(\\mathbf {v} _ {i, t + 1} ^ {\\perp}\\right) \\right] \\mapsto p \\left(V _ {t + 1} \\mid V _ {t + 1} ^ {\\perp}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 878, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21942", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6ea688f256f599a2f3a2f06f94e7c744a8f2d74ec3dc782137026a9e9acf59f5.jpg", + "image_caption": [ + "Figure 3. The relation of normal flow, full flow, and TEG. There are infinitely possible candidates for full flows given a normal flow (left). TEG models this uncertainty using Gaussian having large variance along the tangential direction of the normal flow (right)." + ], + "image_footnote": [], + "bbox": [ + 107, + 78, + 277, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2c3de6766bfb9590737604dd11f71695831a4b4fd687b4172cba7607a6c833e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 78, + 439, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following section, we'll derive TEGBP, an algorithm based on BP for incremental full flow estimation.", + "bbox": [ + 76, + 297, + 468, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tangentially Elongated Gaussian BP", + "text_level": 1, + "bbox": [ + 76, + 343, + 393, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Tangentially Elongated Gaussian", + "text_level": 1, + "bbox": [ + 76, + 369, + 362, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The true full flow $\\mathbf{v}_i$ exists somewhere along the line perpendicular to the normal flow $\\mathbf{v}_i^\\perp$ (Fig. 3 left). In other words, we are uncertain about the full flow from the single normal flow measurement. We model this uncertainty using tangentially elongated Gaussian (TEG), A Gaussian distribution having large variance along the tangential direction of the observed normal flow (Fig. 3 right). We defined the precision matrix of TEG as follows:", + "bbox": [ + 75, + 397, + 470, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Lambda^ {\\mathrm {t e g}} = \\mathbb {R} (\\theta) \\Lambda^ {\\mathrm {m}} \\mathbb {R} (\\theta) ^ {\\top}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 534, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{R}(\\theta)$ is 2D rotation matrix parameterized by the rotation angle $\\theta$ of the normal flow where $\\theta = \\arctan (v_y^\\perp / v_x^\\perp)$ , and $\\Lambda^{\\mathrm{m}} = \\mathrm{diag}(1 / \\sigma_{\\mathrm{r}}^{2}, 1 / \\sigma_{\\mathrm{t}}^{2})$ . In principle, $\\sigma_{\\mathrm{t}} = \\infty$ and $\\sigma_{\\mathrm{r}} = 0$ correspond to the constraint of the full flow from a single noiseless normal flow measurement. In this case, one could parameterize the observation with a 1D projection of full flow (normal flow), and the uncertainty of the projected vector length is parameterized as normal variance, $\\sigma_{\\mathrm{r}}$ . In other words, this corresponds to the case when we are equally uncertain along the tangential direction. In a real-world scenario, the probability of full flow along the tangential direction is not uniform but highly concentrated around the normal flow (Supp-D). The knowledge about the probability density of full flow is a vital clue to recovering the accurate full flow from the inherently ill-posed observation of sparse normal flow; however, the 1D parameterization can not incorporate this vital information. Therefore, we propose to utilize 2D Gaussian (TEG) by incorporating additional parameters, i.e., finite $\\sigma_{\\mathrm{t}}$ . The 2D parameterization is crucial for estimating the accurate full flow from sparse and noisy measurements, and our TEG (2D) includes the 1D parameterization as a special case when $\\sigma_{\\mathrm{t}} = \\infty$ .", + "bbox": [ + 75, + 568, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Corner pixel** When the naive least-square planer fitting is adopted, the quality for the normal flow estimation might be low; we adjust $\\sigma_{\\mathrm{t}}, \\sigma_{\\mathrm{r}}$ depending on the quality. We use lower precision $\\Lambda^{\\mathrm{m}}$ depending on the mean square error between observed events and the estimated plane. Our BP algorithm automatically incorporates this confidence to estimate accurate flow even at the corner point.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark: If full flow observation is available at the corner (e.g., by corner tracking), TEGBP could utilize the reliable information by TEG with smaller $\\sigma_{\\mathrm{t}}$ to improve accuracy.", + "bbox": [ + 496, + 196, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Marginalization of TEG with Valid Prior", + "text_level": 1, + "bbox": [ + 498, + 260, + 839, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we saw earlier, the full flow estimation from the normal flow is inherently an ill-pose without some prior knowledge about the consistency with other pixels (when $\\sigma_{\\mathrm{t}} = \\infty$ and $\\sigma_{\\mathrm{r}} = 0$ ). Without a valid prior, there are infinitely many possibilities along the line perpendicular to the normal flow; any point on the line has the same likelihood when $\\sigma_{\\mathrm{t}} = \\infty$ . To obtain a correct solution, we need the correct prior. Our algorithms could incorporate any prior, e.g., simple uniform smoothness (e.g., total variation), an application-specific knowledge such as flow converges to the vanishing-point in an automotive scenario, or they could be learned using neural networks. In this study, we adopt uniform smoothness for simplicity. The important fact is that the mean of the marginal distribution of TEGs having $\\sigma_{\\mathrm{t}} = \\infty$ and the correct prior equals the correct full flow, and this still approximately holds when $\\sigma_{\\mathrm{t}}$ is moderately large (Sec. B.1).", + "bbox": [ + 496, + 282, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Example. To see this, let's consider the case of Fig. 1. We assume the corner having an edge angle of $\\theta_{1} + \\frac{\\pi}{2}$ and $\\theta_{3} + \\frac{\\pi}{2}$ moves linearly with uniform velocity (all pixels have the same velocity) toward the direction of $\\theta_{0}$ , i.e., $\\hat{\\mathbf{v}} = [\\cos \\theta_{0},\\sin \\theta_{0}]^{\\top}$ . The correct prior in this situation is uniform (infinitely strong smoothness prior; $\\sigma_{p} = 0.0$ ). In this scenario, we'll observed $\\mathbf{v}_{1}^{\\perp} = \\mathbf{v}_{2}^{\\perp} = [\\cos (\\theta_{1} - \\theta_{0})\\cos \\theta_{1},\\cos (\\theta_{1} - \\theta_{0})\\sin \\theta_{1}]^{\\top},\\mathbf{v}_{3}^{\\perp} = [\\cos (\\theta_{3} - \\theta_{0})\\cos \\theta_{3},\\cos (\\theta_{3} - \\theta_{0})\\sin \\theta_{3}]^{\\top}$ .", + "bbox": [ + 496, + 532, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The joint distribution of full flow $V \\coloneqq \\{\\mathbf{v}_1, \\mathbf{v}_2, \\mathbf{v}_3\\}$ from normal flow measurements and the prior is given as:", + "bbox": [ + 496, + 669, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (V) = \\mathcal {N} (\\mathbf {v} _ {1}; \\mathbf {v} _ {1} ^ {\\perp}, \\Sigma_ {\\theta_ {1}}) \\mathcal {N} (\\mathbf {v} _ {2}; \\mathbf {v} _ {2} ^ {\\perp}, \\Sigma_ {\\theta_ {2}}) \\mathcal {N} (\\mathbf {v} _ {3}; \\mathbf {v} _ {3} ^ {\\perp}, \\Sigma_ {\\theta_ {3}}) \\\\ \\mathcal {N} ([ \\mathbf {v} _ {1}; \\mathbf {v} _ {2} ]; \\sigma_ {p} ^ {2} \\mathrm {I}) \\mathcal {N} ([ \\mathbf {v} _ {2}; \\mathbf {v} _ {3} ]; \\sigma_ {p} ^ {2} \\mathrm {I}). \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 708, + 890, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Marginalizing for $\\mathbf{v}_2$ , $p(\\mathbf{v}_2) = \\int p(V)d\\mathbf{v}_1d\\mathbf{v}_3$ and putting $\\sigma_t = \\infty$ and $\\sigma_r = 0$ , we'll get $p(\\mathbf{v}_2) = \\mathcal{N}\\left(\\mathbf{v}_2;[\\cos \\theta_0,\\sin \\theta_0]^\\top ,\\mathrm{diag}(0,0)\\right)$ ; the MAP of this marginal equals to the true flow $\\hat{\\mathbf{v}}$ .", + "bbox": [ + 496, + 755, + 893, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One can also visually verify this result by considering the intersection of the lines of the major axis of the TEG. Notice that the marginal is equal to the full flow even when an average of the normal flow, such as employed in [2], differs from the full flow. In the", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21943", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eab2ad5eb23ce9bb3d9e701035eab2c030c3ffceec0a0f6f5dd675854191f250.jpg", + "image_caption": [ + "Figure 4. 2D factor graph of TEGBP. The topology changes dynamically depending on the active set $\\mathcal{A}_t$ . The measurement node and the active variable node on the same pixel are connected by measurement factor node, and each neighbor active variable node is connected via prior factor node." + ], + "image_footnote": [], + "bbox": [ + 81, + 78, + 467, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "above case, the average normal flow $(\\mathbf{v} = [(2\\cos(\\theta_1 - \\theta_0)\\cos\\theta_1 + \\cos(\\theta_3 - \\theta_0)\\cos\\theta_3)/3, (2\\cos(\\theta_1 - \\theta_0)\\sin\\theta_1 + \\cos(\\theta_3 - \\theta_0)\\sin\\theta_3)/3]^\\top)$ is different from the true flow $([\\cos\\theta_0,\\sin\\theta_0]^\\top)$ . This is a steak difference from the existing incremental sparse method using an average of normal flow and is a key to realizing a significant accuracy boost over the technique.", + "bbox": [ + 75, + 320, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Factor Graph of Full Flow", + "text_level": 1, + "bbox": [ + 76, + 444, + 316, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now we construct the sparse factor graph for full flow using the TEG. The joint of Eq. (2) can be factorized using Bayes theorem as follows:", + "bbox": [ + 75, + 467, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} _ {t} = \\underset {V _ {t}} {\\arg \\max } p \\left(V _ {t} ^ {\\perp} \\mid V _ {t}\\right) p \\left(V _ {t}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 521, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming independence of observations and dependence between neighboring-pixel variables, Eq. (6) is further factorized to the measurement factor $\\psi$ and the prior factor $\\phi$ defined on a sparse graph on the image grid (Fig. 4) as:", + "bbox": [ + 75, + 556, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (V) = \\prod_ {n \\in \\mathcal {A} _ {t}} \\psi (\\mathbf {v} _ {n}) \\prod_ {(i, j) \\in \\mathcal {E} _ {t}} \\phi (\\mathbf {v} _ {i}, \\mathbf {v} _ {j}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 627, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_t$ is the set of active nodes corresponding normal flow measurements $V_t^\\perp$ , and $\\mathcal{E}_t$ is the set of edges between active neighbor nodes. The measurements factor $\\psi$ is defined as TEG (refer to Sec. 3.3.1):", + "bbox": [ + 75, + 670, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (\\mathbf {v} _ {n}) = \\mathcal {N} ^ {- 1} \\left(; \\boldsymbol {\\eta} _ {n}, \\Lambda_ {n} ^ {\\mathrm {t e g}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 739, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{\\eta}_n = \\lambda_n^{\\mathrm{te}}\\mathbf{v}_n^\\perp$ . In this study, we assume the flow is smooth; therefore, use the following prior factor $\\phi$ :", + "bbox": [ + 75, + 768, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\left(\\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {j} \\end{array} \\right]\\right) = \\mathcal {N} ^ {- 1} (\\cdot ; \\mathbf {0}, \\Lambda^ {\\mathrm {p}}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 809, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Lambda^{\\mathrm{p}} = \\mathbf{J}_{\\mathrm{p}}^{\\top}\\mathrm{diag}(1 / \\sigma_{\\mathrm{p}}^{2},1 / \\sigma_{\\mathrm{p}}^{2})\\mathbf{J}_{\\mathrm{p}}$ and the jacobian $\\begin{array}{r}\\mathrm{J_p} = \\left[ \\begin{array}{rrrr} - 1 & 0 & 1 & 0\\\\ 0 & -1 & 0 & 1 \\end{array} \\right] \\end{array}$ for computing the difference between the neighbor nodes.", + "bbox": [ + 76, + 851, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fad96e619ebfcd82df818900c3c498686e218f7494909317d128e9bec2ce1665.jpg", + "table_caption": [ + "Table 1. Basic experimental setup." + ], + "table_footnote": [], + "table_body": "
ParameterESIMDVSMVSECDSEC
Window sizer [pix]5535
Δtpf [ms]40404040
Active durationτ [ms]5050100100
Prior stdσp [pix]0.110.51
Tangential stdσt [pix]10101010
Radial stdσr [pix]3333
Num. of layersL5555
Batch sizeNb10010010001000
", + "bbox": [ + 513, + 103, + 879, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To model the outlier that can not be expressed as a Gaussian (quadratic cost), we employed the Huber cost both for measurement and prior factors. We adopted the technique of [1, 9] to maintain the Gaussian form.", + "bbox": [ + 496, + 243, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4 TEGBP", + "text_level": 1, + "bbox": [ + 500, + 323, + 609, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equipped with the sparse probabilistic graph, we now formalize the incremental message-passing algorithm base on BP (Sec. 3.1.2) for the marginalization of the graph to estimate the full flow (Fig. 5). As the graph of active nodes is loopy, GBP needs to store a belief at each variable node for iterative belief update.", + "bbox": [ + 496, + 348, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Message-passing is asynchronously triggered upon a space observation of normal flow. The measurement factor sends a message to the connected variable node and updates its belief with messages that have been sent from the active node as follows:", + "bbox": [ + 496, + 459, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\eta} _ {i} ^ {\\mathrm {b}} = \\sum_ {j \\in \\mathcal {C} _ {i} \\cap \\mathcal {A} _ {t}} \\boldsymbol {\\eta} _ {j \\rightarrow i}, \\quad \\Lambda_ {i} ^ {\\mathrm {b}} = \\sum_ {j \\in \\mathcal {C} _ {i} \\cap \\mathcal {A} _ {t}} \\Lambda_ {j \\rightarrow i} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 544, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b(\\mathbf{v}_i) = \\mathcal{N}^{-1}(\\cdot ;\\pmb{\\eta}_i^{\\mathrm{b}},\\Lambda_i^{\\mathrm{b}})$ is the belief on node $i$ $\\mu_{j\\rightarrow i} = \\mathcal{N}^{-1}(\\cdot ;\\pmb{\\eta}_{j\\rightarrow i},\\Lambda_{j\\rightarrow i})$ is the message from node $j$ to node $i$ , and $\\mathcal{C}_i$ is the 4-neighbors connected node $j$ . After that, the variable node sends the following message about its updated beliefs to its neighbors:", + "bbox": [ + 496, + 590, + 890, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\eta} _ {j \\rightarrow i} = - \\Lambda_ {0 1} ^ {\\mathrm {p}} \\left(\\Lambda_ {1 1} ^ {\\mathrm {p}} + \\Lambda_ {j} ^ {\\mathrm {b}} - \\Lambda_ {j \\rightarrow i}\\right) ^ {- 1} \\left(\\boldsymbol {\\eta} _ {j} ^ {\\mathrm {b}} - \\boldsymbol {\\eta} _ {j \\rightarrow i}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 676, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Lambda_ {j \\rightarrow i} = \\Lambda_ {0 0} ^ {\\mathrm {p}} - \\Lambda_ {0 1} ^ {\\mathrm {p}} \\left(\\Lambda_ {1 1} ^ {\\mathrm {p}} + \\Lambda_ {j} ^ {\\mathrm {b}} - \\Lambda_ {j \\rightarrow i}\\right) ^ {- 1} \\Lambda_ {1 0} ^ {\\mathrm {p}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 699, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\Lambda_{\\alpha \\beta}^{\\mathrm{p}}$ is a $2\\times 2$ sub-matrix of $\\Lambda^{\\mathrm{p}}$ indexed by $\\alpha$ and $\\beta$ . For a single normal flow measurement, the belief is propagated for $K$ hops, which are repeated for $N_{\\mathrm{itr}}$ times (Fig. 5 correspond to $K = 2$ ).", + "bbox": [ + 496, + 728, + 890, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that our BP algorithm is guaranteed to converge to the correct MAP solution with sufficient iteration because all the belief is realized as a Gaussian (there is no guarantee for a variance because the graph is loopy) [30].", + "bbox": [ + 496, + 789, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse-to-fine message-passing scheme [10] is adopted to speed up the convergence. The measurement factor of", + "bbox": [ + 496, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21944", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/305cf6bb5a5d82c5711e03a37b4b38af0926dad776d25cbd02e8b85e81269759.jpg", + "image_caption": [ + "Normal flow observation" + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 205, + 146 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1de7223f645a716656a89b157485fa2d512269cf4ab9c75d66073f30f10116e.jpg", + "image_caption": [ + "Belief update", + "Figure 5. A schematic showing a single iteration of TEGBP. (a) When a single normal flow has been measured, the message from the measurement factor is sent to the variable node as TEG, (b) the node updates its belief using the message received from the neighbor active nodes. (c) The nodes broadcast the updated belief to the neighbor, and (d) the neighbor nodes update their beliefs. (a)-(d) is repeated until convergence. To prevent the blowup of the message communication, it is stopped after $K$ hops." + ], + "image_footnote": [], + "bbox": [ + 210, + 87, + 343, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/637f17ff8d66b9e36c917880457565e727db61e55f1969e0a8908bf05555e63d.jpg", + "image_caption": [ + "Message passing" + ], + "image_footnote": [], + "bbox": [ + 351, + 87, + 496, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/002a832e4a33fd70fd80e8f78b51cc9e9c1d898e756ed105f3fa9b759d16893b.jpg", + "image_caption": [ + "Belief update" + ], + "image_footnote": [], + "bbox": [ + 496, + 87, + 614, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/210c0528a1b25404df44cdb1ff5e92cd90a9f5a6b78b6cfaada8cbbf139722e5.jpg", + "image_caption": [ + "Message passing" + ], + "image_footnote": [], + "bbox": [ + 625, + 87, + 751, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3983fd458933343f1a93926c1d114fa71726d30c59251d3641984918a05f74bd.jpg", + "image_caption": [ + "Belief update" + ], + "image_footnote": [], + "bbox": [ + 763, + 87, + 887, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b545dc779e1b4b5317b65256ce07c2ce15b80d0771f37b20cc7b492cf91ab2f.jpg", + "image_caption": [ + "Figure 6. The result on ESIM-bricks. The estimated full flow from ARMS and TEGBP are compared with ground-truth flow indicated by the red arrow. They use the same normal flow measurement (left). TEGBP succeeds while ARMS fails." + ], + "image_footnote": [], + "bbox": [ + 78, + 244, + 467, + 345 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9b998636b6f33ffd39d910d02cbe01cdcb78557880f9cd8269036d8b4c159d9.jpg", + "image_caption": [ + "Time surface" + ], + "image_footnote": [], + "bbox": [ + 84, + 430, + 187, + 520 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7bb31ba4a087d3392dd028262d2b1cd903e5ab5c6ece193967f52655546472f3.jpg", + "image_caption": [ + "ARMS [2]", + "Figure 7. The results on DVS-stripes. The histograms of the flow $(v_{x})$ of ARMS and TEGBP are compared. The two red lines indicate the GT flow of the two stripes (the upper stripes have a smaller norm). The result from ARMS are shifted from the GT flow, especially on the area of upper stripes (smaller norm) On the contrary, our TEGBP successfully estimates individual flows." + ], + "image_footnote": [], + "bbox": [ + 207, + 430, + 336, + 529 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b93a9e7730c455e20220832ce8d093345b72ec472b7d1c72b9a2ad9049cb0b5f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 336, + 430, + 464, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the coarser layer $l$ is computed as the sum of active measurement factor messages in a $2 \\times 2$ block of finer layer $l - 1$ . The BP is executed coarse-to-fine; the coarse layer (message) estimation is copied to the finer layer as the initial guess at the start of the message passing on the finer layer (Supp-C.1). The short paths in the course graph improve efficiency for capturing long-range interactions and speed up convergence.", + "bbox": [ + 75, + 665, + 468, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 800, + 200, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted experiments to compare the accuracy of TEGBP with the state-of-the-art incremental flow estimation algorithm called aperture-robust multi-scale flow estimation (ARMS) [2]. Both algorithms are incremental that could asynchronously estimate flow from sparse normal", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "flow measurement. We used the same normal flow measurement for both algorithms for a fair comparison.", + "bbox": [ + 498, + 258, + 890, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setting", + "text_level": 1, + "bbox": [ + 500, + 300, + 697, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Normal flow measurement", + "text_level": 1, + "bbox": [ + 500, + 325, + 738, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize the simple plane fitting to compute the normal flow measurements that serve as inputs to TEGBP and ARMS. We first apply noise removal for the raw events using a refractory filter [17] of $40~\\mathrm{ms}$ duration. Then using the filtered events, we performed plane fitting for each event within a small spatiotemporal window of the size $r\\times r\\times \\Delta t_{\\mathrm{pf}}$ . During the plane fitting, we removed the outlier using the criteria proposed in [2]; we applied this process three times. We summarize the normal flow computation parameters in Tab. 1 (top).", + "bbox": [ + 496, + 349, + 890, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Full flow estimation", + "text_level": 1, + "bbox": [ + 500, + 522, + 691, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both TEGBP and ARMS use the same sparse normal flow measurement, and could be run asynchronously at the rate of normal flow; however, we process $N_{b}$ observation at once for computational efficiency on MATLAB [18] (interactive numerical computing environment) we used to compare the accuracy. The specific parameters for the (quasi) full flow computation are summarized in Tab. 1 (bottom).", + "bbox": [ + 496, + 547, + 890, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In TEGBP, the message should be propagated to all connected nodes, and the process needs to be repeated until convergence; however, we found a small number of hops $(K = 2)$ and single iterations $(N_{\\mathrm{itr}} = 1)$ works well in practice.", + "bbox": [ + 496, + 652, + 890, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. ESIM-bricks", + "text_level": 1, + "bbox": [ + 500, + 724, + 635, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this experiment, we compare the accuracy on a scene where the motion and edges are not perpendicular to the image edge. We synthesized the event stream by observing the image of bricks from a translating camera. We used ESIM [23] library for this simulation, which could synthesize the realistic noisy event stream from the event camera moving around the predefined 3D environment. We compute the normal flow and use the same measurements for both methods. Fig. 6 shows the result. We observe the drift in ARMS due to the average operation. TEGBP correctly", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21945", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/53dd1ab3e790e54c6cb8e7773962061d70dc4c95b32a1120981d6aa66a35b36c.jpg", + "table_caption": [ + "Table 2. The quantitative results on MVSEC. The flow is evaluated on the ground truth intervals." + ], + "table_footnote": [], + "table_body": "
indoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm2.3024.53.6142.93.1336.33.4443.0
ARMS [2]1.7112.72.6726.62.2821.62.6425.7
Ours1.146.251.8716.41.5411.81.4611.1
Indoor_flying1
Indoor_flying3
Outdoor_day1
Time surfaceNormal flowARMS [2]OursGround truth
", + "bbox": [ + 81, + 114, + 893, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8. The qualitative results on MVSEC. See the color wheel on the bottom left for the color encoding.", + "bbox": [ + 158, + 496, + 805, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "estimated the full flow from the normal flow measurement directed at a different angle from the full flow.", + "bbox": [ + 75, + 550, + 468, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.DVS-stirpes", + "text_level": 1, + "bbox": [ + 76, + 589, + 205, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we compare the accuracy on a scene where some region includes different flows. We used a scene from [19] where two stripes translating at different speeds are observed from a dynamic vision sensor (DVS). Fig. 7 shows the result. We observed a significant shift in the ARMS's estimation when motions of different speeds are adjacent. ARMS determines the flow by selecting the scale having the maximum average norm in the window (it falsely picks the large norm in the slow region). In contrast, the TEGBP correctly estimates individual motion.", + "bbox": [ + 75, + 612, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. MVSEC", + "text_level": 1, + "bbox": [ + 76, + 771, + 179, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We quantitatively compare with ARMS in practical robotic scenarios. We used outdoor scenes captured by automobiles and indoor scenes captured by aerial drones from the Multi-vehicle Stereo Event Camera (MVSEC) dataset [33]. This dataset includes ground truth (GT) optical flow computed as the motion filed from the camera motion and scene depth. We evaluate the accuracy using the average", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "endpoint error (AEE) and the percentage of pixels with endpoint error greater than 3 pixels (denoted by % Out). Both are measured on pixels where valid GT exists, and the plane fitting has been successful. We output flows asynchronously in batches with both methods and evaluate them at regular intervals (20Hz of GT rate). Tab. 2 and Fig. 8 show the quantitative and qualitative results. ARMS improves the normal flow, while TEGBP improves more. In contrast to the naive averaging in ARMS, our TEGBP properly models the distribution of flows using TEG and prior.", + "bbox": [ + 496, + 550, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. DSEC", + "text_level": 1, + "bbox": [ + 500, + 710, + 586, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the performance of TEGBP and ARMS on more diverse scenarios using more high-resolution input. To this end, we use the DSEC [13], a recently released dataset in automobile scenarios containing high-resolution event data of $480 \\times 640$ captured from the vehicle that undergoes more diverse motion. It is designed for dense optical flow estimation, and the GT of the test sequence is hidden for leaderboard evaluation. Therefore we use their train sequence to compare sparse AEE error with ARMS.", + "bbox": [ + 496, + 733, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 and Fig. 9 shows the quantitative and qualitative results. Similar to the MVSEC experiments, TEGBP sig", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21946", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef2c3702911d6e8ccd16cd9932c1e8143b60aa4c208737dd4c248d71bb73630f.jpg", + "table_caption": [ + "Table 3. The quantitative results on DSEC. The flow is evaluated on the ground truth intervals." + ], + "table_footnote": [], + "table_body": "
thun_00_azurich_city_01_azurich_city_02_azurich_city_08_azurich_city_11_a
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm7.0968.78.2637.912.886.88.4273.35.9063.4
ARMS [2]5.3160.06.3223.88.9878.76.2264.44.5751.4
Ours4.2645.55.6123.28.7072.65.0849.53.2732.2
", + "bbox": [ + 114, + 114, + 854, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7f6bdf484a396f0b3e25c9d2517d13fc279f7c677748e16859f8324cd6b5112.jpg", + "image_caption": [ + "Figure 9. The qualitative results on DSEC." + ], + "image_footnote": [], + "bbox": [ + 78, + 191, + 893, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nificantly boosts accuracy over AMRS.", + "bbox": [ + 76, + 454, + 336, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 486, + 194, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed TEGBP, a novel incremental full flow estimation algorithm for asynchronous event data. The proposed algorithm is backed up by the probabilistic model, which guarantees equality to the full flow when a valid prior is provided. It realizes efficient event-wise incremental updates based on the BP, where the MAP solution obtained by the local update is guaranteed to match the correct one. We demonstrate the effectiveness of the proposed method on a practical real-world dataset showing a significant boost in accuracy over the existing incremental algorithm.", + "bbox": [ + 75, + 513, + 468, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There is still much room for improvement in accuracy compared to the recent machine learning-based method (although it is computationally intensive). We suspect the gap mainly comes from the noise in the normal flow and poor prior. Toward the ultimate goal of realizing low-latency and high-accuracy full flow estimation in real-time on an edge device, we expect the gap to be closed by incorporating a more robust normal flow and application-specific prior.", + "bbox": [ + 75, + 666, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Future work", + "text_level": 1, + "bbox": [ + 76, + 800, + 210, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robust normal flow estimation As discussed in the experiment section, some scenes include strongly erroneous flow, e.g., normal flow from leaves on the tree. These measurement errors significantly degrade the accuracy. We expect the tiny CNN (e.g., input size of $7 \\times 7$ ) trained to regress", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the normal flow from noisy input will improve the accuracy without sacrificing the computational efficiency.", + "bbox": [ + 500, + 454, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application specific prior Even if the perfect normal flow is obtained, we cannot expect good accuracy without good prior (Sec. 3.3.2). The uniform smoothness prior we adopted may not optimal in many practical scenarios. The discontinuous depth region does not conform to the prior. Even if the planer object undergoes linear motion, we won't observe uniform flow when it is not perpendicular to the optical axis. We can utilize a better prior for a specific application to boost accuracy. For example, one may utilize a vanishing point for automotive applications (assuming the yaw rate is known from the vehicle sensor) similar to [20]. We left these explorations for future work.", + "bbox": [ + 496, + 502, + 892, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimized implementation Our algorithm is very efficient in principle, requiring lower FLOPS, and the entire process can be asynchronously parallelized. The normal flow could be cheaply obtained, e.g., directly from a camera or computed by the lightweight plane fitting (computation of normal flow is outside this study's scope). The core of our algorithm, message passing of (4), is also very cheap, requiring the communication of 6-dim vectors with neighbors. Therefore, an optimized CPU implementation or adaptation of processors supporting the sparse computation could realize low-energy, low-latency optical flow estimation in the edge device. We left the wall clock time and energy consumption evaluation as future work.", + "bbox": [ + 496, + 702, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21947", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pratik Agarwal, Gian Diego Tipaldi, Luciano Spinello, Cyril Stachniss, and Wolfram Burgard. Robust Map Optimization using Dynamic Covariance Scaling. IEEE International Conference on Robotics and Automation (ICRA), 2013. 5", + "[2] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):1-12, 2020. 1, 2, 4, 6, 7, 8, 11", + "[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. IEEE International Symposium on Circuits and Systems, 2018. 2", + "[4] R. Wes Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event Probability Mask (EPM) and Event Denoising Convolutional Neural Network (EDnCNN) for Neuromorphic Cameras. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1701-1710, 2020. 3", + "[5] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous Optical Flow and Intensity Estimation from an Event Camera. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2", + "[6] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, feb 2014. 1, 2", + "[7] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, mar 2012. 1, 2", + "[8] Andrew J. Davison. FutureMapping: The Computational Structure of Spatial AI Systems. arXiv, 2018. 3", + "[9] Andrew J. Davison and Joseph Ortiz. FutureMapping 2: Gaussian Belief Propagation for Spatial AI. arXiv, 2019. 2, 3, 5", + "[10] Pedro F Felzenszwalb and Daniel P Huttenlocher. Ef ci e n t Belief Propagation for Early Vision. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 5, 13", + "[11] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Joerg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1), 2020. 2", + "[12] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 2", + "[13] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. International Conference on 3D Vision (3DV), 2021. 2", + "[15] B. K. B. Horn and B. G. Schunck. Determining Optical Flow. Artificial Intelligence, 17(1-3):185-203, 1981. 1", + "[16] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Bill Nell, Nir Shavit, and Dan Alistarh. Inducing and Exploiting Activation Sparsity for Fast Neural Network Inference. Proceedings of Machine Learning Research, 119:5533-5543, 2020. 3", + "[17] Weng Fei Low, Zhi Gao, Cheng Xiang, and Bharath Ramesh. SOFEA: A non-iterative and robust optical flow estimation algorithm for dynamic vision sensors. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 2, 6", + "[18] MATLAB. version 9.13.0 (R2022b). The MathWorks Inc., Natick, Massachusetts, 2022. 6", + "[19] Elias Mueggler, Christian Forster, Nathan Baumli, Guillermo Gallego, and Davide Scaramuzza. Lifetime Estimation of Events from Dynamic Vision Sensors. IEEE International Conference on Robotics and Automation (ICRA), 2015. 2, 7", + "[20] Jun Nagata, Yusuke Sekikawa, Kosuke Hara, and Yoshimitsu Aoki. FOE-based Regularization for Optical Flow Estimation from an In-vehicle Event Camera. Proceedings of SPIE - The International Society for Optical Engineering, 11049, 2019. 8", + "[21] Joseph Ortiz, Mark Pupilli, Stefan Leutenegger, and Andrew J. Davison. Bundle Adjustment on a Graph Processor. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2413-2422, 2020. 3", + "[22] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single Image Optical Flow Estimation with an Event Camera. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[23] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an Open Event Camera Simulator. Conf. on Robotics Learning (CoRL), 87:969-982, 2018. 6", + "[24] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. MDPI Sensors, 22(14):1-20, 2022. 2", + "[25] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. European Conference on Computer Vision (ECCV), pages 1-23, 2022. 2", + "[26] Chen Shoushun and Guo Menghan. Live Demonstration : CeleX-V: a 1M Pixel Multi-Mode Event-based Sensor Chen Shoushun. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1–2, 2019. 3", + "[27] Timo Stoffregen and Lindsay Kleeman. Event Cameras, Contrast Maximization and Reward Functions: an Analysis. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12300-12308, 2019. 2", + "[28] Daniel C. Stumpp, Himanshu Akolkar, Alan D. George, and Ryad B. Benosman. hARMS: A Hardware Acceleration Architecture for Real-Time Event-Based Optical Flow. IEEE Access, 10:58181-58198, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21948", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. European Conference on Computer Vision (ECCV), 2020. 1, 2", + "[30] Yair Weiss and William T Freeman. Correctness of belief propagation in Gaussian graphical models of arbitrary topology. Neural Information Processing Systems (NIPS), pages 673-679, 2000. 5", + "[31] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion from Sparse Event Data. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 2", + "[32] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. Proceedings of Robotics: Science and Systems, jun 2018. 2", + "[33] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multi Vehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 7, 14", + "[34] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. European Conference on Computer Vision (ECCV) Workshop, pages 711-714, 2018. 2", + "[35] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-based Learning of Optical Flow, Depth, and Egomotion. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 989–997, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 513 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21949", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_model.json b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8cbd5ed43c8cecd83d2206e9a1d0973357e433c3 --- /dev/null +++ b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.756, + 0.175 + ], + "angle": 0, + "content": "Tangentially Elongated Gaussian Belief Propagation for Event-based Incremental Optical Flow Estimation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.75, + 0.222 + ], + "angle": 0, + "content": "Jun Nagata† and Yusuke Sekikawa† DENSO IT LAB., INC., Japan." + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.256, + 0.343, + 0.273 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.474, + 0.665 + ], + "angle": 0, + "content": "Optical flow estimation is a fundamental functionality in computer vision. An event-based camera, which asynchronously detects sparse intensity changes, is an ideal device for realizing low-latency estimation of the optical flow owing to its low-latency sensing mechanism. An existing method using local plane fitting of events could utilize the sparsity to realize incremental updates for low-latency estimation; however, its output is merely a normal component of the full optical flow. An alternative approach using a frame-based deep neural network could estimate the full flow; however, its intensive nonincremental dense operation prohibits the low-latency estimation. We propose tangentially elongated Gaussian (TEG) belief propagation (BP) that realizes incremental full-flow estimation. We model the probability of full flow as the joint distribution of TEGs from the normal flow measurements, such that the marginal of this distribution with correct prior equals the full flow. We formulate the marginalization using a message-passing based on the BP to realize efficient incremental updates using sparse measurements. In addition to the theoretical justification, we evaluate the effectiveness of the TEGBP in real-world datasets; it outperforms SOTA incremental quasi-full flow method by a large margin. (The code is available at https://github.com/DensoITLab/tegbp/)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.21, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Optical flow estimation, which computes the correspondence of pixels in different time measurements, is a fundamental building block of computer vision. One needs to estimate the flow at low latency in many practical applications, such as autonomous driving cars, unmanned aerial vehicles, and factory automation robots. Most of the existing optical flow algorithm utilizes dense video frames; it computes the flow by searching the similar intensity pattern [15, 29]. Recently, methods using deep neural network (DNN) [29] demonstrate impressive accuracy at the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.246, + 0.895, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.357, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed TEGBP. We model a belief about the full flow from a normal flow measurement using TEG (RGB ellipse). The mean of the marginal distribution of each TEG with an appropriate prior equals the full flow (magenta arrow). The marginal (magenta ellipse) is computed incrementally using local message passing (RGB arrows) based on BP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.561 + ], + "angle": 0, + "content": "cost of higher-computational cost. Either model-based or DNN-based, the frame-based algorithm needs to compute the entire pixel for every frame, even when there are subtle changes or no changes at all. This dense operation makes it difficult to realize low-latency estimation, especially on the resource-constrained edge device." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.893, + 0.865 + ], + "angle": 0, + "content": "The event-based camera is a bio-inspired vision sensor, which asynchronously detects intensity change on each pixel. Thanks to the novel sensing mechanism, the camera equips favorable characteristics for optical flow estimation, such as high dynamic range (HDR), blur-free measurement, and, most importantly, sparse low-latency data acquisition. Many researchers have explored the way to utilize sparsity to realize efficient low-latency estimation; one extends the well-known Lucas-Kanade algorithm [7], and the other exploits the local planer shape of the spatiotemporal event streams [6]. These methods could utilize the sparsity for efficient incremental processing; however, the optical flow computed in this way (e.g., by plane fitting) is the normal flow, which is a normal component of full flow and often different from them1 we want to obtain. The normal flow is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient). Some work tried to recover the full flow from the normal flow [2]; however, it does not precisely equal the full flow (refer to Sec. 2). There exist methods that could estimate full flow, such as a varia" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.361, + 0.9 + ], + "angle": 0, + "content": "† These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.875, + 0.893, + 0.901 + ], + "angle": 0, + "content": "1 full flow is usually simply called optical flow or flow, yet, we use full flow when we want to highlight the difference with the normal flow." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21940" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "tional method [5], a multi-scale extension of contrast maximization [25], or a frame-based DNN [14]. However, they need to apply non-incremental dense operation for all the pixels of the event frame (dense representation constructed from sparse events) for every frame, which prohibits the low-latency estimation on the edge device." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.474, + 0.486 + ], + "angle": 0, + "content": "Our research goal is to realize an incremental full flow algorithm from sparse normal flow measurements. To this end, we propose Tangentially Elongated Gaussian (TEG) Belief Propagation (BP). We compute the full flow using the normal flow measurements, which can be observed directly from an event camera or computed cheaply using an existing algorithm2. Notice that given a single measurement of normal flow, there are infinite possibilities for the full flow along the tangential direction of the normal flow. We model this uncertainty using the TEG, Gaussian distribution, which has a large variance along the tangential direction of the normal flow. The probability density of full flow on each pixel is given as the margins of the joint distribution of TEG data factor and some prior factor on a sparse graph (Fig. 1, Sec. 3.3.2). We leverage the sparse graph to formulate the incremental full flow estimation algorithm using message-passing based on BP [9]. We evaluate the effectiveness of the TEGBP on real-world data captured from aerial drones and automobiles. TEGBP outperforms SOTA incremental method [2] by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.5, + 0.216, + 0.515 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.47, + 0.556 + ], + "angle": 0, + "content": "This section reviews the related literature on optical flow estimation using sparse event signals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.424, + 0.581 + ], + "angle": 0, + "content": "2.1. Incremental Algorithm for Normal Flow" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Lucas-Kanade based method The traditional frame-based Lucas-Kanade method finds the flow by aligning the small patch of intensity between consecutive frames using its spatial gradient and temporal change. The event-based counterpart [7] uses the integration of events to approximate them for each incoming event. This approach realized highly efficient estimation by utilizing the sparsity of input events. However, it can only estimate the normal flow, which is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient) [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Plane fitting based method When an edge undergoes liner motion, triggered events generate a time surface, which can be locally approximated by a plane in a spatiotemporal space. The local plane fitting method [3,6,17, 19] computes the optical flow by fitting the plane to the sparse event points. However, the flow obtained by plane fitting is also the normal component of the full flow." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Recently, a method for computing the quasi-full flow has been proposed [2]; it estimates the full flow by considering the average normal flow in multiple resolutions. It incrementally updates the average and selects the flow with the largest average norm from the multiple resolutions. It expects the flow to have the largest norm to correspond to the full flow. The average operation induces bias on the estimated flow; therefore, it equals the full flow on the very limited scene where the true motion in the window is orthogonal to the edge direction (refer to Sec.5 in their paper). We'll also experimentally compare this point in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.258, + 0.892, + 0.287 + ], + "angle": 0, + "content": "In summary, current sparse incremental algorithms are limited to normal flow." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.298, + 0.771, + 0.314 + ], + "angle": 0, + "content": "2.2. Dense Algorithm for Full Flow" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.32, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Variational method A variational optimization-based approach realizes the full flow estimation by incorporating intensity frame obtained either by the event/frame hybrid camera [22] or by simultaneously estimating image intensity [5]. It could estimate full flow; however, it requires additional intensity observation or intensive dense optimization to recover the intensity frame." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.445, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Contrast maximization method Contrast maximization (CMax) [12, 27] computes the underlying motion (optical flow) of events stream by finding their alignment. It computes the alignment by maximizing the focus score (e.g., variance) of an image generated by warping the events (IWE) using the optical flow parameter. The original algorithm can not be used to estimate pixel-wise flow because it tends to converge to the degenerated solution (e.g., warps all events to a single point) [24, 35]. Recently, a hierarchical CMax approach avoiding the degenerated solution for pixel-wise estimation has been proposed [25]. Although this approach works well even in the practical scenario, it is not incremental; it must recompute the IWE for each sliding time window by using all events in the patch for every hierarchical scale (this involves intensive iteration by itself), making it difficult to operate in real-time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.87 + ], + "angle": 0, + "content": "DNN based method Frame-based DNN is often utilized to compute the full flow [14, 31, 32, 34]. Recently, a frame-based high-accuracy model called RAFT [29] has been imported into the event-based vision literature showing the SOTA performance for the dense flow estimation [14]. These methods can not consume sparse events directly; the sparse input must be converted into dense frame representation. The dense convolution is computationally expensive because it needs to operate for all pixels, even when there are subtle changes or no changes at all, making it difficult to operate at a higher rate on the edge device." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In summary, some dense algorithms could estimate full flow; however, it compromises the event's sparse asyn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "2We consider the normal flow as measurement and discuss algorithms to process the sparse measurement of normal flow to estimate the full flow." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21941" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "chronous nature, and their non-incremental dense operation prohibits them from realizing low-latency processing." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.468, + 0.15 + ], + "angle": 0, + "content": "2.3. Sparse Asynchronous Computing Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.158, + 0.471, + 0.355 + ], + "angle": 0, + "content": "Events are asynchronously triggered only from limited pixels which detect intensity changes. Several techniques have been developed for efficient processing of the sparse signals on CPUs by taking advantage of the highly optimized cache mechanism [16, 28]. Besides, sparse data-flow architectures [8, 9], which differ from prevalent Neuman-type computing architectures, are gaining attention for sparse signal processing. For example, bundle adjustment has been solved on the novel architecture using an algorithm based on Gaussian BP [21]. Our goal is to establish the incremental full flow estimation algorithm by taking advantage of the sparsity of events, which is expected to run efficiently on CPUs or emerging data-flow processors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.248, + 0.387 + ], + "angle": 0, + "content": "3. Proposed method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.396, + 0.206, + 0.412 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.42, + 0.218, + 0.433 + ], + "angle": 0, + "content": "3.1.1 Normal flow" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.566 + ], + "angle": 0, + "content": "The normal flow \\(\\mathbf{v}_i^\\perp \\in \\mathbb{R}^2\\) on pixel \\(i\\) is a normal component of full flow \\(\\mathbf{v}_i \\in \\mathbb{R}^2\\), which we want to estimate. One can not decide the full flow uniquely from the normal flow, and there are infinitely many possibilities along the line perpendicular to the normal flow (Fig. 3 left). There are several options for obtaining the normal flow; some event-based cameras, such as CeleX-V [26] directly observe it, or it can be cheaply computed from the sparse events (Sec. 2.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.469, + 0.716 + ], + "angle": 0, + "content": "The moving edge in 3D space triggers events; we can compute the normal flow by finding the local plane on the triggered events. In this study, we adopt a naive least-squares fitting of the plane using events in a small spatiotemporal window. An erroneous normal flow may significantly deteriorate the full flow accuracy, especially on highly textured regions such as leaves on trees. One can improve the results by utilizing an advanced noise removal algorithm [4] or by adopting more sophisticated normal flow estimation methods, such as small neural networks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.26, + 0.755 + ], + "angle": 0, + "content": "3.1.2 Belief propagation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Factor graphs geometrically represent the structure of probabilistic problems. A factor graph \\( G = (V, F, E) \\) is composed of a set of variable nodes \\( V = \\{\\mathbf{v}_i\\}_{i=1:N_v} \\), a set of factor nodes \\( F = \\{f_s\\}_{s=1:N_f} \\) and a set of edges \\( E \\). Each factor node \\( f_s \\) represents a probabilistic constraint \\( f_s(V_s) \\) between a subset of variables \\( V_s \\subset V \\). The factorization of the variables is explicitly represented in the graph by connecting factor nodes with the variable nodes they depend on. Probabilistically, these factors are the independent" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.08, + 0.887, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.201, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Figure 2. Processing pipeline. Observation set at \\( t \\) is represented as \\( V_{t}^{\\perp} \\) where the normal flow is sparsely assigned. The TEGBP incrementally update its estimation about full flow \\( V_{t + 1} \\) using previous estimation \\( V_{t} \\) and new observation \\( \\mathbf{v}_{i,t + 1}^{\\perp} \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.297, + 0.771, + 0.311 + ], + "angle": 0, + "content": "terms that make up the joint distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.322, + 0.892, + 0.365 + ], + "angle": 0, + "content": "\\[\np (V) = \\prod_ {s = 1} ^ {N _ {f}} f _ {s} \\left(V _ {s}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Belief propagation (BP) is a generic distributed algorithm for computing the marginal distribution of a set of variables from their joint distribution. The marginal for a single variable \\(\\mathbf{v}_i\\) is computed by the integral over all the other variables. BP computed the marginals asynchronously using message-passing between connected nodes on the graph. It is substantially more efficient by leveraging the topology of the sparse graph than naive integration of the entire graph (which involves giant matrix inversion). Refer to the seminal work of [9] for more detail about the derivation, condition for convergence guarantee, and discussion about the stability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.698, + 0.586 + ], + "angle": 0, + "content": "3.2. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.595, + 0.891, + 0.701 + ], + "angle": 0, + "content": "Given measurements of normal flow \\( V_{t}^{\\perp} \\) on the the pixel grid where each normal flow is assigned to the corresponding pixels (within time interval \\( [t - \\tau, t] \\)), we want to estimate the posterior distribution of the corresponding full flow \\( p(V_{t} \\mid V_{t}^{\\perp}) \\). Once the posterior distribution of full flow is obtained, the actual full flow output is given by the maximum a posteriori (MAP) estimate as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.71, + 0.891, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\hat {V} _ {t} = \\underset {V _ {t}} {\\arg \\max } p \\left(V _ {t} \\mid V _ {t} ^ {\\perp}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.748, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Our goal is to formulate an incremental full flow estimation algorithm using sparse observations of local normal flows (Fig. 2). Solving the arg max of the joint distribution of \\( V_{t} \\) for every incoming normal flow measurement is infeasible both in terms of memory and computational perspective; therefore, we want to formulate \\( f \\) as an incremental sparse update function, i.e., new observation interacts sparsely with the previous results \\( p(V_{t} \\mid V_{t}^{\\perp}) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.88, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\nf: \\left[ p \\left(V _ {t} \\mid V _ {t} ^ {\\perp}\\right), p \\left(\\mathbf {v} _ {i, t + 1} ^ {\\perp}\\right) \\right] \\mapsto p \\left(V _ {t + 1} \\mid V _ {t + 1} ^ {\\perp}\\right). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.079, + 0.278, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.079, + 0.44, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.198, + 0.471, + 0.255 + ], + "angle": 0, + "content": "Figure 3. The relation of normal flow, full flow, and TEG. There are infinitely possible candidates for full flows given a normal flow (left). TEG models this uncertainty using Gaussian having large variance along the tangential direction of the normal flow (right)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.298, + 0.47, + 0.328 + ], + "angle": 0, + "content": "In the following section, we'll derive TEGBP, an algorithm based on BP for incremental full flow estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.344, + 0.395, + 0.361 + ], + "angle": 0, + "content": "3.3. Tangentially Elongated Gaussian BP" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.371, + 0.363, + 0.387 + ], + "angle": 0, + "content": "3.3.1 Tangentially Elongated Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.519 + ], + "angle": 0, + "content": "The true full flow \\(\\mathbf{v}_i\\) exists somewhere along the line perpendicular to the normal flow \\(\\mathbf{v}_i^\\perp\\) (Fig. 3 left). In other words, we are uncertain about the full flow from the single normal flow measurement. We model this uncertainty using tangentially elongated Gaussian (TEG), A Gaussian distribution having large variance along the tangential direction of the observed normal flow (Fig. 3 right). We defined the precision matrix of TEG as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.535, + 0.47, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\Lambda^ {\\mathrm {t e g}} = \\mathbb {R} (\\theta) \\Lambda^ {\\mathrm {m}} \\mathbb {R} (\\theta) ^ {\\top}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{R}(\\theta)\\) is 2D rotation matrix parameterized by the rotation angle \\(\\theta\\) of the normal flow where \\(\\theta = \\arctan (v_y^\\perp / v_x^\\perp)\\), and \\(\\Lambda^{\\mathrm{m}} = \\mathrm{diag}(1 / \\sigma_{\\mathrm{r}}^{2}, 1 / \\sigma_{\\mathrm{t}}^{2})\\). In principle, \\(\\sigma_{\\mathrm{t}} = \\infty\\) and \\(\\sigma_{\\mathrm{r}} = 0\\) correspond to the constraint of the full flow from a single noiseless normal flow measurement. In this case, one could parameterize the observation with a 1D projection of full flow (normal flow), and the uncertainty of the projected vector length is parameterized as normal variance, \\(\\sigma_{\\mathrm{r}}\\). In other words, this corresponds to the case when we are equally uncertain along the tangential direction. In a real-world scenario, the probability of full flow along the tangential direction is not uniform but highly concentrated around the normal flow (Supp-D). The knowledge about the probability density of full flow is a vital clue to recovering the accurate full flow from the inherently ill-posed observation of sparse normal flow; however, the 1D parameterization can not incorporate this vital information. Therefore, we propose to utilize 2D Gaussian (TEG) by incorporating additional parameters, i.e., finite \\(\\sigma_{\\mathrm{t}}\\). The 2D parameterization is crucial for estimating the accurate full flow from sparse and noisy measurements, and our TEG (2D) includes the 1D parameterization as a special case when \\(\\sigma_{\\mathrm{t}} = \\infty\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "**Corner pixel** When the naive least-square planer fitting is adopted, the quality for the normal flow estimation might be low; we adjust \\(\\sigma_{\\mathrm{t}}, \\sigma_{\\mathrm{r}}\\) depending on the quality. We use lower precision \\(\\Lambda^{\\mathrm{m}}\\) depending on the mean square error between observed events and the estimated plane. Our BP algorithm automatically incorporates this confidence to estimate accurate flow even at the corner point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Remark: If full flow observation is available at the corner (e.g., by corner tracking), TEGBP could utilize the reliable information by TEG with smaller \\(\\sigma_{\\mathrm{t}}\\) to improve accuracy." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.261, + 0.84, + 0.275 + ], + "angle": 0, + "content": "3.3.2 Marginalization of TEG with Valid Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.527 + ], + "angle": 0, + "content": "As we saw earlier, the full flow estimation from the normal flow is inherently an ill-pose without some prior knowledge about the consistency with other pixels (when \\(\\sigma_{\\mathrm{t}} = \\infty\\) and \\(\\sigma_{\\mathrm{r}} = 0\\)). Without a valid prior, there are infinitely many possibilities along the line perpendicular to the normal flow; any point on the line has the same likelihood when \\(\\sigma_{\\mathrm{t}} = \\infty\\). To obtain a correct solution, we need the correct prior. Our algorithms could incorporate any prior, e.g., simple uniform smoothness (e.g., total variation), an application-specific knowledge such as flow converges to the vanishing-point in an automotive scenario, or they could be learned using neural networks. In this study, we adopt uniform smoothness for simplicity. The important fact is that the mean of the marginal distribution of TEGs having \\(\\sigma_{\\mathrm{t}} = \\infty\\) and the correct prior equals the correct full flow, and this still approximately holds when \\(\\sigma_{\\mathrm{t}}\\) is moderately large (Sec. B.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Example. To see this, let's consider the case of Fig. 1. We assume the corner having an edge angle of \\(\\theta_{1} + \\frac{\\pi}{2}\\) and \\(\\theta_{3} + \\frac{\\pi}{2}\\) moves linearly with uniform velocity (all pixels have the same velocity) toward the direction of \\(\\theta_{0}\\), i.e., \\(\\hat{\\mathbf{v}} = [\\cos \\theta_{0},\\sin \\theta_{0}]^{\\top}\\). The correct prior in this situation is uniform (infinitely strong smoothness prior; \\(\\sigma_{p} = 0.0\\)). In this scenario, we'll observed \\(\\mathbf{v}_{1}^{\\perp} = \\mathbf{v}_{2}^{\\perp} = [\\cos (\\theta_{1} - \\theta_{0})\\cos \\theta_{1},\\cos (\\theta_{1} - \\theta_{0})\\sin \\theta_{1}]^{\\top},\\mathbf{v}_{3}^{\\perp} = [\\cos (\\theta_{3} - \\theta_{0})\\cos \\theta_{3},\\cos (\\theta_{3} - \\theta_{0})\\sin \\theta_{3}]^{\\top}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.701 + ], + "angle": 0, + "content": "The joint distribution of full flow \\( V \\coloneqq \\{\\mathbf{v}_1, \\mathbf{v}_2, \\mathbf{v}_3\\} \\) from normal flow measurements and the prior is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.709, + 0.892, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (V) = \\mathcal {N} (\\mathbf {v} _ {1}; \\mathbf {v} _ {1} ^ {\\perp}, \\Sigma_ {\\theta_ {1}}) \\mathcal {N} (\\mathbf {v} _ {2}; \\mathbf {v} _ {2} ^ {\\perp}, \\Sigma_ {\\theta_ {2}}) \\mathcal {N} (\\mathbf {v} _ {3}; \\mathbf {v} _ {3} ^ {\\perp}, \\Sigma_ {\\theta_ {3}}) \\\\ \\mathcal {N} ([ \\mathbf {v} _ {1}; \\mathbf {v} _ {2} ]; \\sigma_ {p} ^ {2} \\mathrm {I}) \\mathcal {N} ([ \\mathbf {v} _ {2}; \\mathbf {v} _ {3} ]; \\sigma_ {p} ^ {2} \\mathrm {I}). \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.817 + ], + "angle": 0, + "content": "Marginalizing for \\(\\mathbf{v}_2\\), \\(p(\\mathbf{v}_2) = \\int p(V)d\\mathbf{v}_1d\\mathbf{v}_3\\) and putting \\(\\sigma_t = \\infty\\) and \\(\\sigma_r = 0\\), we'll get \\(p(\\mathbf{v}_2) = \\mathcal{N}\\left(\\mathbf{v}_2;[\\cos \\theta_0,\\sin \\theta_0]^\\top ,\\mathrm{diag}(0,0)\\right)\\); the MAP of this marginal equals to the true flow \\(\\hat{\\mathbf{v}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "One can also visually verify this result by considering the intersection of the lines of the major axis of the TEG. Notice that the marginal is equal to the full flow even when an average of the normal flow, such as employed in [2], differs from the full flow. In the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.079, + 0.468, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.214, + 0.471, + 0.283 + ], + "angle": 0, + "content": "Figure 4. 2D factor graph of TEGBP. The topology changes dynamically depending on the active set \\(\\mathcal{A}_t\\). The measurement node and the active variable node on the same pixel are connected by measurement factor node, and each neighbor active variable node is connected via prior factor node." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.47, + 0.428 + ], + "angle": 0, + "content": "above case, the average normal flow \\((\\mathbf{v} = [(2\\cos(\\theta_1 - \\theta_0)\\cos\\theta_1 + \\cos(\\theta_3 - \\theta_0)\\cos\\theta_3)/3, (2\\cos(\\theta_1 - \\theta_0)\\sin\\theta_1 + \\cos(\\theta_3 - \\theta_0)\\sin\\theta_3)/3]^\\top)\\) is different from the true flow \\(([\\cos\\theta_0,\\sin\\theta_0]^\\top)\\). This is a steak difference from the existing incremental sparse method using an average of normal flow and is a key to realizing a significant accuracy boost over the technique." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.445, + 0.317, + 0.459 + ], + "angle": 0, + "content": "3.3.3 Factor Graph of Full Flow" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.514 + ], + "angle": 0, + "content": "Now we construct the sparse factor graph for full flow using the TEG. The joint of Eq. (2) can be factorized using Bayes theorem as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.522, + 0.47, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\hat {V} _ {t} = \\underset {V _ {t}} {\\arg \\max } p \\left(V _ {t} ^ {\\perp} \\mid V _ {t}\\right) p \\left(V _ {t}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Assuming independence of observations and dependence between neighboring-pixel variables, Eq. (6) is further factorized to the measurement factor \\(\\psi\\) and the prior factor \\(\\phi\\) defined on a sparse graph on the image grid (Fig. 4) as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.628, + 0.47, + 0.662 + ], + "angle": 0, + "content": "\\[\np (V) = \\prod_ {n \\in \\mathcal {A} _ {t}} \\psi (\\mathbf {v} _ {n}) \\prod_ {(i, j) \\in \\mathcal {E} _ {t}} \\phi (\\mathbf {v} _ {i}, \\mathbf {v} _ {j}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.732 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_t\\) is the set of active nodes corresponding normal flow measurements \\(V_t^\\perp\\), and \\(\\mathcal{E}_t\\) is the set of edges between active neighbor nodes. The measurements factor \\(\\psi\\) is defined as TEG (refer to Sec. 3.3.1):" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.741, + 0.469, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\psi (\\mathbf {v} _ {n}) = \\mathcal {N} ^ {- 1} \\left(; \\boldsymbol {\\eta} _ {n}, \\Lambda_ {n} ^ {\\mathrm {t e g}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.802 + ], + "angle": 0, + "content": "where \\(\\pmb{\\eta}_n = \\lambda_n^{\\mathrm{te}}\\mathbf{v}_n^\\perp\\). In this study, we assume the flow is smooth; therefore, use the following prior factor \\(\\phi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.81, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\phi \\left(\\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {j} \\end{array} \\right]\\right) = \\mathcal {N} ^ {- 1} (\\cdot ; \\mathbf {0}, \\Lambda^ {\\mathrm {p}}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.852, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Lambda^{\\mathrm{p}} = \\mathbf{J}_{\\mathrm{p}}^{\\top}\\mathrm{diag}(1 / \\sigma_{\\mathrm{p}}^{2},1 / \\sigma_{\\mathrm{p}}^{2})\\mathbf{J}_{\\mathrm{p}}\\) and the jacobian \\(\\begin{array}{r}\\mathrm{J_p} = \\left[ \\begin{array}{rrrr} - 1 & 0 & 1 & 0\\\\ 0 & -1 & 0 & 1 \\end{array} \\right] \\end{array}\\) for computing the difference between the neighbor nodes." + }, + { + "type": "table_caption", + "bbox": [ + 0.588, + 0.079, + 0.804, + 0.094 + ], + "angle": 0, + "content": "Table 1. Basic experimental setup." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.104, + 0.88, + 0.219 + ], + "angle": 0, + "content": "
ParameterESIMDVSMVSECDSEC
Window sizer [pix]5535
Δtpf [ms]40404040
Active durationτ [ms]5050100100
Prior stdσp [pix]0.110.51
Tangential stdσt [pix]10101010
Radial stdσr [pix]3333
Num. of layersL5555
Batch sizeNb10010010001000
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.304 + ], + "angle": 0, + "content": "To model the outlier that can not be expressed as a Gaussian (quadratic cost), we employed the Huber cost both for measurement and prior factors. We adopted the technique of [1, 9] to maintain the Gaussian form." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.324, + 0.611, + 0.337 + ], + "angle": 0, + "content": "3.3.4 TEGBP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Equipped with the sparse probabilistic graph, we now formalize the incremental message-passing algorithm base on BP (Sec. 3.1.2) for the marginalization of the graph to estimate the full flow (Fig. 5). As the graph of active nodes is loopy, GBP needs to store a belief at each variable node for iterative belief update." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Message-passing is asynchronously triggered upon a space observation of normal flow. The measurement factor sends a message to the connected variable node and updates its belief with messages that have been sent from the active node as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.545, + 0.892, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\eta} _ {i} ^ {\\mathrm {b}} = \\sum_ {j \\in \\mathcal {C} _ {i} \\cap \\mathcal {A} _ {t}} \\boldsymbol {\\eta} _ {j \\rightarrow i}, \\quad \\Lambda_ {i} ^ {\\mathrm {b}} = \\sum_ {j \\in \\mathcal {C} _ {i} \\cap \\mathcal {A} _ {t}} \\Lambda_ {j \\rightarrow i} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.667 + ], + "angle": 0, + "content": "where \\(b(\\mathbf{v}_i) = \\mathcal{N}^{-1}(\\cdot ;\\pmb{\\eta}_i^{\\mathrm{b}},\\Lambda_i^{\\mathrm{b}})\\) is the belief on node \\(i\\) \\(\\mu_{j\\rightarrow i} = \\mathcal{N}^{-1}(\\cdot ;\\pmb{\\eta}_{j\\rightarrow i},\\Lambda_{j\\rightarrow i})\\) is the message from node \\(j\\) to node \\(i\\) , and \\(\\mathcal{C}_i\\) is the 4-neighbors connected node \\(j\\) . After that, the variable node sends the following message about its updated beliefs to its neighbors:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.678, + 0.891, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\eta} _ {j \\rightarrow i} = - \\Lambda_ {0 1} ^ {\\mathrm {p}} \\left(\\Lambda_ {1 1} ^ {\\mathrm {p}} + \\Lambda_ {j} ^ {\\mathrm {b}} - \\Lambda_ {j \\rightarrow i}\\right) ^ {- 1} \\left(\\boldsymbol {\\eta} _ {j} ^ {\\mathrm {b}} - \\boldsymbol {\\eta} _ {j \\rightarrow i}\\right) \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.7, + 0.891, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\Lambda_ {j \\rightarrow i} = \\Lambda_ {0 0} ^ {\\mathrm {p}} - \\Lambda_ {0 1} ^ {\\mathrm {p}} \\left(\\Lambda_ {1 1} ^ {\\mathrm {p}} + \\Lambda_ {j} ^ {\\mathrm {b}} - \\Lambda_ {j \\rightarrow i}\\right) ^ {- 1} \\Lambda_ {1 0} ^ {\\mathrm {p}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.789 + ], + "angle": 0, + "content": "where, \\(\\Lambda_{\\alpha \\beta}^{\\mathrm{p}}\\) is a \\(2\\times 2\\) sub-matrix of \\(\\Lambda^{\\mathrm{p}}\\) indexed by \\(\\alpha\\) and \\(\\beta\\). For a single normal flow measurement, the belief is propagated for \\(K\\) hops, which are repeated for \\(N_{\\mathrm{itr}}\\) times (Fig. 5 correspond to \\(K = 2\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Note that our BP algorithm is guaranteed to converge to the correct MAP solution with sufficient iteration because all the belief is realized as a Gaussian (there is no guarantee for a variance because the graph is loopy) [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Coarse-to-fine message-passing scheme [10] is adopted to speed up the convergence. The measurement factor of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.207, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.149, + 0.204, + 0.159 + ], + "angle": 0, + "content": "Normal flow observation" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.088, + 0.344, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.149, + 0.312, + 0.159 + ], + "angle": 0, + "content": "Belief update" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.088, + 0.498, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.149, + 0.456, + 0.16 + ], + "angle": 0, + "content": "Message passing" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.088, + 0.616, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.149, + 0.584, + 0.16 + ], + "angle": 0, + "content": "Belief update" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.088, + 0.752, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.149, + 0.73, + 0.16 + ], + "angle": 0, + "content": "Message passing" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.088, + 0.888, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.149, + 0.857, + 0.16 + ], + "angle": 0, + "content": "Belief update" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.164, + 0.894, + 0.222 + ], + "angle": 0, + "content": "Figure 5. A schematic showing a single iteration of TEGBP. (a) When a single normal flow has been measured, the message from the measurement factor is sent to the variable node as TEG, (b) the node updates its belief using the message received from the neighbor active nodes. (c) The nodes broadcast the updated belief to the neighbor, and (d) the neighbor nodes update their beliefs. (a)-(d) is repeated until convergence. To prevent the blowup of the message communication, it is stopped after \\( K \\) hops." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.246, + 0.468, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.404 + ], + "angle": 0, + "content": "Figure 6. The result on ESIM-bricks. The estimated full flow from ARMS and TEGBP are compared with ground-truth flow indicated by the red arrow. They use the same normal flow measurement (left). TEGBP succeeds while ARMS fails." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.431, + 0.188, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.531, + 0.168, + 0.542 + ], + "angle": 0, + "content": "Time surface" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.431, + 0.338, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.531, + 0.306, + 0.543 + ], + "angle": 0, + "content": "ARMS [2]" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.431, + 0.465, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.531, + 0.414, + 0.542 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.628 + ], + "angle": 0, + "content": "Figure 7. The results on DVS-stripes. The histograms of the flow \\((v_{x})\\) of ARMS and TEGBP are compared. The two red lines indicate the GT flow of the two stripes (the upper stripes have a smaller norm). The result from ARMS are shifted from the GT flow, especially on the area of upper stripes (smaller norm) On the contrary, our TEGBP successfully estimates individual flows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.788 + ], + "angle": 0, + "content": "the coarser layer \\( l \\) is computed as the sum of active measurement factor messages in a \\( 2 \\times 2 \\) block of finer layer \\( l - 1 \\). The BP is executed coarse-to-fine; the coarse layer (message) estimation is copied to the finer layer as the initial guess at the start of the message passing on the finer layer (Supp-C.1). The short paths in the course graph improve efficiency for capturing long-range interactions and speed up convergence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.201, + 0.817 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We conducted experiments to compare the accuracy of TEGBP with the state-of-the-art incremental flow estimation algorithm called aperture-robust multi-scale flow estimation (ARMS) [2]. Both algorithms are incremental that could asynchronously estimate flow from sparse normal" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.26, + 0.892, + 0.291 + ], + "angle": 0, + "content": "flow measurement. We used the same normal flow measurement for both algorithms for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.698, + 0.317 + ], + "angle": 0, + "content": "4.1. Experimental setting" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.326, + 0.739, + 0.339 + ], + "angle": 0, + "content": "4.1.1 Normal flow measurement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.502 + ], + "angle": 0, + "content": "We utilize the simple plane fitting to compute the normal flow measurements that serve as inputs to TEGBP and ARMS. We first apply noise removal for the raw events using a refractory filter [17] of \\(40~\\mathrm{ms}\\) duration. Then using the filtered events, we performed plane fitting for each event within a small spatiotemporal window of the size \\(r\\times r\\times \\Delta t_{\\mathrm{pf}}\\). During the plane fitting, we removed the outlier using the criteria proposed in [2]; we applied this process three times. We summarize the normal flow computation parameters in Tab. 1 (top)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.692, + 0.537 + ], + "angle": 0, + "content": "4.1.2 Full flow estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Both TEGBP and ARMS use the same sparse normal flow measurement, and could be run asynchronously at the rate of normal flow; however, we process \\( N_{b} \\) observation at once for computational efficiency on MATLAB [18] (interactive numerical computing environment) we used to compare the accuracy. The specific parameters for the (quasi) full flow computation are summarized in Tab. 1 (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.715 + ], + "angle": 0, + "content": "In TEGBP, the message should be propagated to all connected nodes, and the process needs to be repeated until convergence; however, we found a small number of hops \\((K = 2)\\) and single iterations \\((N_{\\mathrm{itr}} = 1)\\) works well in practice." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.637, + 0.74 + ], + "angle": 0, + "content": "4.2. ESIM-bricks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this experiment, we compare the accuracy on a scene where the motion and edges are not perpendicular to the image edge. We synthesized the event stream by observing the image of bricks from a translating camera. We used ESIM [23] library for this simulation, which could synthesize the realistic noisy event stream from the event camera moving around the predefined 3D environment. We compute the normal flow and use the same measurements for both methods. Fig. 6 shows the result. We observe the drift in ARMS due to the average operation. TEGBP correctly" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21945" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.191, + 0.09, + 0.777, + 0.104 + ], + "angle": 0, + "content": "Table 2. The quantitative results on MVSEC. The flow is evaluated on the ground truth intervals." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.115, + 0.895, + 0.489 + ], + "angle": 0, + "content": "
indoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm2.3024.53.6142.93.1336.33.4443.0
ARMS [2]1.7112.72.6726.62.2821.62.6425.7
Ours1.146.251.8716.41.5411.81.4611.1
Indoor_flying1
Indoor_flying3
Outdoor_day1
Time surfaceNormal flowARMS [2]OursGround truth
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.497, + 0.807, + 0.512 + ], + "angle": 0, + "content": "Figure 8. The qualitative results on MVSEC. See the color wheel on the bottom left for the color encoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.581 + ], + "angle": 0, + "content": "estimated the full flow from the normal flow measurement directed at a different angle from the full flow." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.206, + 0.606 + ], + "angle": 0, + "content": "4.3.DVS-stirpes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In this experiment, we compare the accuracy on a scene where some region includes different flows. We used a scene from [19] where two stripes translating at different speeds are observed from a dynamic vision sensor (DVS). Fig. 7 shows the result. We observed a significant shift in the ARMS's estimation when motions of different speeds are adjacent. ARMS determines the flow by selecting the scale having the maximum average norm in the window (it falsely picks the large norm in the slow region). In contrast, the TEGBP correctly estimates individual motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.18, + 0.787 + ], + "angle": 0, + "content": "4.4. MVSEC" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We quantitatively compare with ARMS in practical robotic scenarios. We used outdoor scenes captured by automobiles and indoor scenes captured by aerial drones from the Multi-vehicle Stereo Event Camera (MVSEC) dataset [33]. This dataset includes ground truth (GT) optical flow computed as the motion filed from the camera motion and scene depth. We evaluate the accuracy using the average" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.703 + ], + "angle": 0, + "content": "endpoint error (AEE) and the percentage of pixels with endpoint error greater than 3 pixels (denoted by % Out). Both are measured on pixels where valid GT exists, and the plane fitting has been successful. We output flows asynchronously in batches with both methods and evaluate them at regular intervals (20Hz of GT rate). Tab. 2 and Fig. 8 show the quantitative and qualitative results. ARMS improves the normal flow, while TEGBP improves more. In contrast to the naive averaging in ARMS, our TEGBP properly models the distribution of flows using TEG and prior." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.587, + 0.726 + ], + "angle": 0, + "content": "4.5. DSEC" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We compare the performance of TEGBP and ARMS on more diverse scenarios using more high-resolution input. To this end, we use the DSEC [13], a recently released dataset in automobile scenarios containing high-resolution event data of \\(480 \\times 640\\) captured from the vehicle that undergoes more diverse motion. It is designed for dense optical flow estimation, and the GT of the test sequence is hidden for leaderboard evaluation. Therefore we use their train sequence to compare sparse AEE error with ARMS." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tab. 3 and Fig. 9 shows the quantitative and qualitative results. Similar to the MVSEC experiments, TEGBP sig" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21946" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.198, + 0.09, + 0.772, + 0.104 + ], + "angle": 0, + "content": "Table 3. The quantitative results on DSEC. The flow is evaluated on the ground truth intervals." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.115, + 0.855, + 0.18 + ], + "angle": 0, + "content": "
thun_00_azurich_city_01_azurich_city_02_azurich_city_08_azurich_city_11_a
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm7.0968.78.2637.912.886.88.4273.35.9063.4
ARMS [2]5.3160.06.3223.88.9878.76.2264.44.5751.4
Ours4.2645.55.6123.28.7072.65.0849.53.2732.2
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.192, + 0.894, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.401, + 0.618, + 0.415 + ], + "angle": 0, + "content": "Figure 9. The qualitative results on DSEC." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.338, + 0.469 + ], + "angle": 0, + "content": "nificantly boosts accuracy over AMRS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.196, + 0.503 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.47, + 0.665 + ], + "angle": 0, + "content": "We proposed TEGBP, a novel incremental full flow estimation algorithm for asynchronous event data. The proposed algorithm is backed up by the probabilistic model, which guarantees equality to the full flow when a valid prior is provided. It realizes efficient event-wise incremental updates based on the BP, where the MAP solution obtained by the local update is guaranteed to match the correct one. We demonstrate the effectiveness of the proposed method on a practical real-world dataset showing a significant boost in accuracy over the existing incremental algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.788 + ], + "angle": 0, + "content": "There is still much room for improvement in accuracy compared to the recent machine learning-based method (although it is computationally intensive). We suspect the gap mainly comes from the noise in the normal flow and poor prior. Toward the ultimate goal of realizing low-latency and high-accuracy full flow estimation in real-time on an edge device, we expect the gap to be closed by incorporating a more robust normal flow and application-specific prior." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.212, + 0.815 + ], + "angle": 0, + "content": "5.1. Future work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Robust normal flow estimation As discussed in the experiment section, some scenes include strongly erroneous flow, e.g., normal flow from leaves on the tree. These measurement errors significantly degrade the accuracy. We expect the tiny CNN (e.g., input size of \\(7 \\times 7\\)) trained to regress" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.455, + 0.892, + 0.485 + ], + "angle": 0, + "content": "the normal flow from noisy input will improve the accuracy without sacrificing the computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.893, + 0.684 + ], + "angle": 0, + "content": "Application specific prior Even if the perfect normal flow is obtained, we cannot expect good accuracy without good prior (Sec. 3.3.2). The uniform smoothness prior we adopted may not optimal in many practical scenarios. The discontinuous depth region does not conform to the prior. Even if the planer object undergoes linear motion, we won't observe uniform flow when it is not perpendicular to the optical axis. We can utilize a better prior for a specific application to boost accuracy. For example, one may utilize a vanishing point for automotive applications (assuming the yaw rate is known from the vehicle sensor) similar to [20]. We left these explorations for future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.899 + ], + "angle": 0, + "content": "Optimized implementation Our algorithm is very efficient in principle, requiring lower FLOPS, and the entire process can be asynchronously parallelized. The normal flow could be cheaply obtained, e.g., directly from a camera or computed by the lightweight plane fitting (computation of normal flow is outside this study's scope). The core of our algorithm, message passing of (4), is also very cheap, requiring the communication of 6-dim vectors with neighbors. Therefore, an optimized CPU implementation or adaptation of processors supporting the sparse computation could realize low-energy, low-latency optical flow estimation in the edge device. We left the wall clock time and energy consumption evaluation as future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21947" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Pratik Agarwal, Gian Diego Tipaldi, Luciano Spinello, Cyril Stachniss, and Wolfram Burgard. Robust Map Optimization using Dynamic Covariance Scaling. IEEE International Conference on Robotics and Automation (ICRA), 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):1-12, 2020. 1, 2, 4, 6, 7, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. IEEE International Symposium on Circuits and Systems, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[4] R. Wes Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event Probability Mask (EPM) and Event Denoising Convolutional Neural Network (EDnCNN) for Neuromorphic Cameras. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1701-1710, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[5] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous Optical Flow and Intensity Estimation from an Event Camera. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.471, + 0.499 + ], + "angle": 0, + "content": "[6] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, feb 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.471, + 0.556 + ], + "angle": 0, + "content": "[7] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, mar 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.587 + ], + "angle": 0, + "content": "[8] Andrew J. Davison. FutureMapping: The Computational Structure of Spatial AI Systems. arXiv, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[9] Andrew J. Davison and Joseph Ortiz. FutureMapping 2: Gaussian Belief Propagation for Spatial AI. arXiv, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.686 + ], + "angle": 0, + "content": "[10] Pedro F Felzenszwalb and Daniel P Huttenlocher. Ef ci e n t Belief Propagation for Early Vision. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[11] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Joerg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. International Conference on 3D Vision (3DV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] B. K. B. Horn and B. G. Schunck. Determining Optical Flow. Artificial Intelligence, 17(1-3):185-203, 1981. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Bill Nell, Nir Shavit, and Dan Alistarh. Inducing and Exploiting Activation Sparsity for Fast Neural Network Inference. Proceedings of Machine Learning Research, 119:5533-5543, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Weng Fei Low, Zhi Gao, Cheng Xiang, and Bharath Ramesh. SOFEA: A non-iterative and robust optical flow estimation algorithm for dynamic vision sensors. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[18] MATLAB. version 9.13.0 (R2022b). The MathWorks Inc., Natick, Massachusetts, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Elias Mueggler, Christian Forster, Nathan Baumli, Guillermo Gallego, and Davide Scaramuzza. Lifetime Estimation of Events from Dynamic Vision Sensors. IEEE International Conference on Robotics and Automation (ICRA), 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[20] Jun Nagata, Yusuke Sekikawa, Kosuke Hara, and Yoshimitsu Aoki. FOE-based Regularization for Optical Flow Estimation from an In-vehicle Event Camera. Proceedings of SPIE - The International Society for Optical Engineering, 11049, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Joseph Ortiz, Mark Pupilli, Stefan Leutenegger, and Andrew J. Davison. Bundle Adjustment on a Graph Processor. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2413-2422, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single Image Optical Flow Estimation with an Event Camera. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[23] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an Open Event Camera Simulator. Conf. on Robotics Learning (CoRL), 87:969-982, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[24] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. MDPI Sensors, 22(14):1-20, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[25] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. European Conference on Computer Vision (ECCV), pages 1-23, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[26] Chen Shoushun and Guo Menghan. Live Demonstration : CeleX-V: a 1M Pixel Multi-Mode Event-based Sensor Chen Shoushun. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1–2, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[27] Timo Stoffregen and Lindsay Kleeman. Event Cameras, Contrast Maximization and Reward Functions: an Analysis. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12300-12308, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Daniel C. Stumpp, Himanshu Akolkar, Alan D. George, and Ryad B. Benosman. hARMS: A Hardware Acceleration Architecture for Real-Time Event-Based Optical Flow. IEEE Access, 10:58181-58198, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21948" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[29] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. European Conference on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[30] Yair Weiss and William T Freeman. Correctness of belief propagation in Gaussian graphical models of arbitrary topology. Neural Information Processing Systems (NIPS), pages 673-679, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[31] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion from Sparse Event Data. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[32] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. Proceedings of Robotics: Science and Systems, jun 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[33] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multi Vehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[34] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. European Conference on Computer Vision (ECCV) Workshop, pages 711-714, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[35] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-based Learning of Optical Flow, Depth, and Egomotion. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 989–997, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21949" + } + ] +] \ No newline at end of file diff --git a/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_origin.pdf b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8ff626d9c3159bfe1136dbd4c0ce797eb6e7341d --- /dev/null +++ b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/1e72f145-1dc6-4f05-89ce-8e276ebb5dbf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c6a81c543361f3da7abff1c475dc37d908b4116b982338f69e3eacb710c223c +size 5695555 diff --git a/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/full.md b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7347b5cb1db2b0b017bbdab9c855749b3aa3085e --- /dev/null +++ b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/full.md @@ -0,0 +1,340 @@ +# Tangentially Elongated Gaussian Belief Propagation for Event-based Incremental Optical Flow Estimation + +Jun Nagata† and Yusuke Sekikawa† DENSO IT LAB., INC., Japan. + +# Abstract + +Optical flow estimation is a fundamental functionality in computer vision. An event-based camera, which asynchronously detects sparse intensity changes, is an ideal device for realizing low-latency estimation of the optical flow owing to its low-latency sensing mechanism. An existing method using local plane fitting of events could utilize the sparsity to realize incremental updates for low-latency estimation; however, its output is merely a normal component of the full optical flow. An alternative approach using a frame-based deep neural network could estimate the full flow; however, its intensive nonincremental dense operation prohibits the low-latency estimation. We propose tangentially elongated Gaussian (TEG) belief propagation (BP) that realizes incremental full-flow estimation. We model the probability of full flow as the joint distribution of TEGs from the normal flow measurements, such that the marginal of this distribution with correct prior equals the full flow. We formulate the marginalization using a message-passing based on the BP to realize efficient incremental updates using sparse measurements. In addition to the theoretical justification, we evaluate the effectiveness of the TEGBP in real-world datasets; it outperforms SOTA incremental quasi-full flow method by a large margin. (The code is available at https://github.com/DensoITLab/tegbp/) + +# 1. Introduction + +Optical flow estimation, which computes the correspondence of pixels in different time measurements, is a fundamental building block of computer vision. One needs to estimate the flow at low latency in many practical applications, such as autonomous driving cars, unmanned aerial vehicles, and factory automation robots. Most of the existing optical flow algorithm utilizes dense video frames; it computes the flow by searching the similar intensity pattern [15, 29]. Recently, methods using deep neural network (DNN) [29] demonstrate impressive accuracy at the + +![](images/0553784f9a4b0d0b55180ab1194f8293bc38feac0ef3b0f32b87927b1e320d8b.jpg) +Figure 1. Overview of the proposed TEGBP. We model a belief about the full flow from a normal flow measurement using TEG (RGB ellipse). The mean of the marginal distribution of each TEG with an appropriate prior equals the full flow (magenta arrow). The marginal (magenta ellipse) is computed incrementally using local message passing (RGB arrows) based on BP. + +cost of higher-computational cost. Either model-based or DNN-based, the frame-based algorithm needs to compute the entire pixel for every frame, even when there are subtle changes or no changes at all. This dense operation makes it difficult to realize low-latency estimation, especially on the resource-constrained edge device. + +The event-based camera is a bio-inspired vision sensor, which asynchronously detects intensity change on each pixel. Thanks to the novel sensing mechanism, the camera equips favorable characteristics for optical flow estimation, such as high dynamic range (HDR), blur-free measurement, and, most importantly, sparse low-latency data acquisition. Many researchers have explored the way to utilize sparsity to realize efficient low-latency estimation; one extends the well-known Lucas-Kanade algorithm [7], and the other exploits the local planer shape of the spatiotemporal event streams [6]. These methods could utilize the sparsity for efficient incremental processing; however, the optical flow computed in this way (e.g., by plane fitting) is the normal flow, which is a normal component of full flow and often different from them1 we want to obtain. The normal flow is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient). Some work tried to recover the full flow from the normal flow [2]; however, it does not precisely equal the full flow (refer to Sec. 2). There exist methods that could estimate full flow, such as a varia + +tional method [5], a multi-scale extension of contrast maximization [25], or a frame-based DNN [14]. However, they need to apply non-incremental dense operation for all the pixels of the event frame (dense representation constructed from sparse events) for every frame, which prohibits the low-latency estimation on the edge device. + +Our research goal is to realize an incremental full flow algorithm from sparse normal flow measurements. To this end, we propose Tangentially Elongated Gaussian (TEG) Belief Propagation (BP). We compute the full flow using the normal flow measurements, which can be observed directly from an event camera or computed cheaply using an existing algorithm2. Notice that given a single measurement of normal flow, there are infinite possibilities for the full flow along the tangential direction of the normal flow. We model this uncertainty using the TEG, Gaussian distribution, which has a large variance along the tangential direction of the normal flow. The probability density of full flow on each pixel is given as the margins of the joint distribution of TEG data factor and some prior factor on a sparse graph (Fig. 1, Sec. 3.3.2). We leverage the sparse graph to formulate the incremental full flow estimation algorithm using message-passing based on BP [9]. We evaluate the effectiveness of the TEGBP on real-world data captured from aerial drones and automobiles. TEGBP outperforms SOTA incremental method [2] by a large margin. + +# 2. Related work + +This section reviews the related literature on optical flow estimation using sparse event signals. + +# 2.1. Incremental Algorithm for Normal Flow + +Lucas-Kanade based method The traditional frame-based Lucas-Kanade method finds the flow by aligning the small patch of intensity between consecutive frames using its spatial gradient and temporal change. The event-based counterpart [7] uses the integration of events to approximate them for each incoming event. This approach realized highly efficient estimation by utilizing the sparsity of input events. However, it can only estimate the normal flow, which is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient) [11]. + +Plane fitting based method When an edge undergoes liner motion, triggered events generate a time surface, which can be locally approximated by a plane in a spatiotemporal space. The local plane fitting method [3,6,17, 19] computes the optical flow by fitting the plane to the sparse event points. However, the flow obtained by plane fitting is also the normal component of the full flow. + +Recently, a method for computing the quasi-full flow has been proposed [2]; it estimates the full flow by considering the average normal flow in multiple resolutions. It incrementally updates the average and selects the flow with the largest average norm from the multiple resolutions. It expects the flow to have the largest norm to correspond to the full flow. The average operation induces bias on the estimated flow; therefore, it equals the full flow on the very limited scene where the true motion in the window is orthogonal to the edge direction (refer to Sec.5 in their paper). We'll also experimentally compare this point in Sec. 4.3. + +In summary, current sparse incremental algorithms are limited to normal flow. + +# 2.2. Dense Algorithm for Full Flow + +Variational method A variational optimization-based approach realizes the full flow estimation by incorporating intensity frame obtained either by the event/frame hybrid camera [22] or by simultaneously estimating image intensity [5]. It could estimate full flow; however, it requires additional intensity observation or intensive dense optimization to recover the intensity frame. + +Contrast maximization method Contrast maximization (CMax) [12, 27] computes the underlying motion (optical flow) of events stream by finding their alignment. It computes the alignment by maximizing the focus score (e.g., variance) of an image generated by warping the events (IWE) using the optical flow parameter. The original algorithm can not be used to estimate pixel-wise flow because it tends to converge to the degenerated solution (e.g., warps all events to a single point) [24, 35]. Recently, a hierarchical CMax approach avoiding the degenerated solution for pixel-wise estimation has been proposed [25]. Although this approach works well even in the practical scenario, it is not incremental; it must recompute the IWE for each sliding time window by using all events in the patch for every hierarchical scale (this involves intensive iteration by itself), making it difficult to operate in real-time. + +DNN based method Frame-based DNN is often utilized to compute the full flow [14, 31, 32, 34]. Recently, a frame-based high-accuracy model called RAFT [29] has been imported into the event-based vision literature showing the SOTA performance for the dense flow estimation [14]. These methods can not consume sparse events directly; the sparse input must be converted into dense frame representation. The dense convolution is computationally expensive because it needs to operate for all pixels, even when there are subtle changes or no changes at all, making it difficult to operate at a higher rate on the edge device. + +In summary, some dense algorithms could estimate full flow; however, it compromises the event's sparse asyn + +chronous nature, and their non-incremental dense operation prohibits them from realizing low-latency processing. + +# 2.3. Sparse Asynchronous Computing Architecture + +Events are asynchronously triggered only from limited pixels which detect intensity changes. Several techniques have been developed for efficient processing of the sparse signals on CPUs by taking advantage of the highly optimized cache mechanism [16, 28]. Besides, sparse data-flow architectures [8, 9], which differ from prevalent Neuman-type computing architectures, are gaining attention for sparse signal processing. For example, bundle adjustment has been solved on the novel architecture using an algorithm based on Gaussian BP [21]. Our goal is to establish the incremental full flow estimation algorithm by taking advantage of the sparsity of events, which is expected to run efficiently on CPUs or emerging data-flow processors. + +# 3. Proposed method + +# 3.1. Preliminary + +# 3.1.1 Normal flow + +The normal flow $\mathbf{v}_i^\perp \in \mathbb{R}^2$ on pixel $i$ is a normal component of full flow $\mathbf{v}_i \in \mathbb{R}^2$ , which we want to estimate. One can not decide the full flow uniquely from the normal flow, and there are infinitely many possibilities along the line perpendicular to the normal flow (Fig. 3 left). There are several options for obtaining the normal flow; some event-based cameras, such as CeleX-V [26] directly observe it, or it can be cheaply computed from the sparse events (Sec. 2.1). + +The moving edge in 3D space triggers events; we can compute the normal flow by finding the local plane on the triggered events. In this study, we adopt a naive least-squares fitting of the plane using events in a small spatiotemporal window. An erroneous normal flow may significantly deteriorate the full flow accuracy, especially on highly textured regions such as leaves on trees. One can improve the results by utilizing an advanced noise removal algorithm [4] or by adopting more sophisticated normal flow estimation methods, such as small neural networks. + +# 3.1.2 Belief propagation + +Factor graphs geometrically represent the structure of probabilistic problems. A factor graph $G = (V, F, E)$ is composed of a set of variable nodes $V = \{\mathbf{v}_i\}_{i=1:N_v}$ , a set of factor nodes $F = \{f_s\}_{s=1:N_f}$ and a set of edges $E$ . Each factor node $f_s$ represents a probabilistic constraint $f_s(V_s)$ between a subset of variables $V_s \subset V$ . The factorization of the variables is explicitly represented in the graph by connecting factor nodes with the variable nodes they depend on. Probabilistically, these factors are the independent + +![](images/59f55debec1fc390bc0d1fbbc73c88e7024f8e93740ded52eff28fa7c4c044c8.jpg) +Figure 2. Processing pipeline. Observation set at $t$ is represented as $V_{t}^{\perp}$ where the normal flow is sparsely assigned. The TEGBP incrementally update its estimation about full flow $V_{t + 1}$ using previous estimation $V_{t}$ and new observation $\mathbf{v}_{i,t + 1}^{\perp}$ . + +terms that make up the joint distribution: + +$$ +p (V) = \prod_ {s = 1} ^ {N _ {f}} f _ {s} \left(V _ {s}\right). \tag {1} +$$ + +Belief propagation (BP) is a generic distributed algorithm for computing the marginal distribution of a set of variables from their joint distribution. The marginal for a single variable $\mathbf{v}_i$ is computed by the integral over all the other variables. BP computed the marginals asynchronously using message-passing between connected nodes on the graph. It is substantially more efficient by leveraging the topology of the sparse graph than naive integration of the entire graph (which involves giant matrix inversion). Refer to the seminal work of [9] for more detail about the derivation, condition for convergence guarantee, and discussion about the stability. + +# 3.2. Problem formulation + +Given measurements of normal flow $V_{t}^{\perp}$ on the the pixel grid where each normal flow is assigned to the corresponding pixels (within time interval $[t - \tau, t]$ ), we want to estimate the posterior distribution of the corresponding full flow $p(V_{t} \mid V_{t}^{\perp})$ . Once the posterior distribution of full flow is obtained, the actual full flow output is given by the maximum a posteriori (MAP) estimate as: + +$$ +\hat {V} _ {t} = \underset {V _ {t}} {\arg \max } p \left(V _ {t} \mid V _ {t} ^ {\perp}\right). \tag {2} +$$ + +Our goal is to formulate an incremental full flow estimation algorithm using sparse observations of local normal flows (Fig. 2). Solving the arg max of the joint distribution of $V_{t}$ for every incoming normal flow measurement is infeasible both in terms of memory and computational perspective; therefore, we want to formulate $f$ as an incremental sparse update function, i.e., new observation interacts sparsely with the previous results $p(V_{t} \mid V_{t}^{\perp})$ as follows: + +$$ +f: \left[ p \left(V _ {t} \mid V _ {t} ^ {\perp}\right), p \left(\mathbf {v} _ {i, t + 1} ^ {\perp}\right) \right] \mapsto p \left(V _ {t + 1} \mid V _ {t + 1} ^ {\perp}\right). \tag {3} +$$ + +![](images/6ea688f256f599a2f3a2f06f94e7c744a8f2d74ec3dc782137026a9e9acf59f5.jpg) +Figure 3. The relation of normal flow, full flow, and TEG. There are infinitely possible candidates for full flows given a normal flow (left). TEG models this uncertainty using Gaussian having large variance along the tangential direction of the normal flow (right). + +![](images/2c3de6766bfb9590737604dd11f71695831a4b4fd687b4172cba7607a6c833e0.jpg) + +In the following section, we'll derive TEGBP, an algorithm based on BP for incremental full flow estimation. + +# 3.3. Tangentially Elongated Gaussian BP + +# 3.3.1 Tangentially Elongated Gaussian + +The true full flow $\mathbf{v}_i$ exists somewhere along the line perpendicular to the normal flow $\mathbf{v}_i^\perp$ (Fig. 3 left). In other words, we are uncertain about the full flow from the single normal flow measurement. We model this uncertainty using tangentially elongated Gaussian (TEG), A Gaussian distribution having large variance along the tangential direction of the observed normal flow (Fig. 3 right). We defined the precision matrix of TEG as follows: + +$$ +\Lambda^ {\mathrm {t e g}} = \mathbb {R} (\theta) \Lambda^ {\mathrm {m}} \mathbb {R} (\theta) ^ {\top}, \tag {4} +$$ + +where $\mathbb{R}(\theta)$ is 2D rotation matrix parameterized by the rotation angle $\theta$ of the normal flow where $\theta = \arctan (v_y^\perp / v_x^\perp)$ , and $\Lambda^{\mathrm{m}} = \mathrm{diag}(1 / \sigma_{\mathrm{r}}^{2}, 1 / \sigma_{\mathrm{t}}^{2})$ . In principle, $\sigma_{\mathrm{t}} = \infty$ and $\sigma_{\mathrm{r}} = 0$ correspond to the constraint of the full flow from a single noiseless normal flow measurement. In this case, one could parameterize the observation with a 1D projection of full flow (normal flow), and the uncertainty of the projected vector length is parameterized as normal variance, $\sigma_{\mathrm{r}}$ . In other words, this corresponds to the case when we are equally uncertain along the tangential direction. In a real-world scenario, the probability of full flow along the tangential direction is not uniform but highly concentrated around the normal flow (Supp-D). The knowledge about the probability density of full flow is a vital clue to recovering the accurate full flow from the inherently ill-posed observation of sparse normal flow; however, the 1D parameterization can not incorporate this vital information. Therefore, we propose to utilize 2D Gaussian (TEG) by incorporating additional parameters, i.e., finite $\sigma_{\mathrm{t}}$ . The 2D parameterization is crucial for estimating the accurate full flow from sparse and noisy measurements, and our TEG (2D) includes the 1D parameterization as a special case when $\sigma_{\mathrm{t}} = \infty$ . + +**Corner pixel** When the naive least-square planer fitting is adopted, the quality for the normal flow estimation might be low; we adjust $\sigma_{\mathrm{t}}, \sigma_{\mathrm{r}}$ depending on the quality. We use lower precision $\Lambda^{\mathrm{m}}$ depending on the mean square error between observed events and the estimated plane. Our BP algorithm automatically incorporates this confidence to estimate accurate flow even at the corner point. + +Remark: If full flow observation is available at the corner (e.g., by corner tracking), TEGBP could utilize the reliable information by TEG with smaller $\sigma_{\mathrm{t}}$ to improve accuracy. + +# 3.3.2 Marginalization of TEG with Valid Prior + +As we saw earlier, the full flow estimation from the normal flow is inherently an ill-pose without some prior knowledge about the consistency with other pixels (when $\sigma_{\mathrm{t}} = \infty$ and $\sigma_{\mathrm{r}} = 0$ ). Without a valid prior, there are infinitely many possibilities along the line perpendicular to the normal flow; any point on the line has the same likelihood when $\sigma_{\mathrm{t}} = \infty$ . To obtain a correct solution, we need the correct prior. Our algorithms could incorporate any prior, e.g., simple uniform smoothness (e.g., total variation), an application-specific knowledge such as flow converges to the vanishing-point in an automotive scenario, or they could be learned using neural networks. In this study, we adopt uniform smoothness for simplicity. The important fact is that the mean of the marginal distribution of TEGs having $\sigma_{\mathrm{t}} = \infty$ and the correct prior equals the correct full flow, and this still approximately holds when $\sigma_{\mathrm{t}}$ is moderately large (Sec. B.1). + +Example. To see this, let's consider the case of Fig. 1. We assume the corner having an edge angle of $\theta_{1} + \frac{\pi}{2}$ and $\theta_{3} + \frac{\pi}{2}$ moves linearly with uniform velocity (all pixels have the same velocity) toward the direction of $\theta_{0}$ , i.e., $\hat{\mathbf{v}} = [\cos \theta_{0},\sin \theta_{0}]^{\top}$ . The correct prior in this situation is uniform (infinitely strong smoothness prior; $\sigma_{p} = 0.0$ ). In this scenario, we'll observed $\mathbf{v}_{1}^{\perp} = \mathbf{v}_{2}^{\perp} = [\cos (\theta_{1} - \theta_{0})\cos \theta_{1},\cos (\theta_{1} - \theta_{0})\sin \theta_{1}]^{\top},\mathbf{v}_{3}^{\perp} = [\cos (\theta_{3} - \theta_{0})\cos \theta_{3},\cos (\theta_{3} - \theta_{0})\sin \theta_{3}]^{\top}$ . + +The joint distribution of full flow $V \coloneqq \{\mathbf{v}_1, \mathbf{v}_2, \mathbf{v}_3\}$ from normal flow measurements and the prior is given as: + +$$ +\begin{array}{l} p (V) = \mathcal {N} (\mathbf {v} _ {1}; \mathbf {v} _ {1} ^ {\perp}, \Sigma_ {\theta_ {1}}) \mathcal {N} (\mathbf {v} _ {2}; \mathbf {v} _ {2} ^ {\perp}, \Sigma_ {\theta_ {2}}) \mathcal {N} (\mathbf {v} _ {3}; \mathbf {v} _ {3} ^ {\perp}, \Sigma_ {\theta_ {3}}) \\ \mathcal {N} ([ \mathbf {v} _ {1}; \mathbf {v} _ {2} ]; \sigma_ {p} ^ {2} \mathrm {I}) \mathcal {N} ([ \mathbf {v} _ {2}; \mathbf {v} _ {3} ]; \sigma_ {p} ^ {2} \mathrm {I}). \tag {5} \\ \end{array} +$$ + +Marginalizing for $\mathbf{v}_2$ , $p(\mathbf{v}_2) = \int p(V)d\mathbf{v}_1d\mathbf{v}_3$ and putting $\sigma_t = \infty$ and $\sigma_r = 0$ , we'll get $p(\mathbf{v}_2) = \mathcal{N}\left(\mathbf{v}_2;[\cos \theta_0,\sin \theta_0]^\top ,\mathrm{diag}(0,0)\right)$ ; the MAP of this marginal equals to the true flow $\hat{\mathbf{v}}$ . + +One can also visually verify this result by considering the intersection of the lines of the major axis of the TEG. Notice that the marginal is equal to the full flow even when an average of the normal flow, such as employed in [2], differs from the full flow. In the + +![](images/eab2ad5eb23ce9bb3d9e701035eab2c030c3ffceec0a0f6f5dd675854191f250.jpg) +Figure 4. 2D factor graph of TEGBP. The topology changes dynamically depending on the active set $\mathcal{A}_t$ . The measurement node and the active variable node on the same pixel are connected by measurement factor node, and each neighbor active variable node is connected via prior factor node. + +above case, the average normal flow $(\mathbf{v} = [(2\cos(\theta_1 - \theta_0)\cos\theta_1 + \cos(\theta_3 - \theta_0)\cos\theta_3)/3, (2\cos(\theta_1 - \theta_0)\sin\theta_1 + \cos(\theta_3 - \theta_0)\sin\theta_3)/3]^\top)$ is different from the true flow $([\cos\theta_0,\sin\theta_0]^\top)$ . This is a steak difference from the existing incremental sparse method using an average of normal flow and is a key to realizing a significant accuracy boost over the technique. + +# 3.3.3 Factor Graph of Full Flow + +Now we construct the sparse factor graph for full flow using the TEG. The joint of Eq. (2) can be factorized using Bayes theorem as follows: + +$$ +\hat {V} _ {t} = \underset {V _ {t}} {\arg \max } p \left(V _ {t} ^ {\perp} \mid V _ {t}\right) p \left(V _ {t}\right). \tag {6} +$$ + +Assuming independence of observations and dependence between neighboring-pixel variables, Eq. (6) is further factorized to the measurement factor $\psi$ and the prior factor $\phi$ defined on a sparse graph on the image grid (Fig. 4) as: + +$$ +p (V) = \prod_ {n \in \mathcal {A} _ {t}} \psi (\mathbf {v} _ {n}) \prod_ {(i, j) \in \mathcal {E} _ {t}} \phi (\mathbf {v} _ {i}, \mathbf {v} _ {j}), \tag {7} +$$ + +where $\mathcal{A}_t$ is the set of active nodes corresponding normal flow measurements $V_t^\perp$ , and $\mathcal{E}_t$ is the set of edges between active neighbor nodes. The measurements factor $\psi$ is defined as TEG (refer to Sec. 3.3.1): + +$$ +\psi (\mathbf {v} _ {n}) = \mathcal {N} ^ {- 1} \left(; \boldsymbol {\eta} _ {n}, \Lambda_ {n} ^ {\mathrm {t e g}}\right), \tag {8} +$$ + +where $\pmb{\eta}_n = \lambda_n^{\mathrm{te}}\mathbf{v}_n^\perp$ . In this study, we assume the flow is smooth; therefore, use the following prior factor $\phi$ : + +$$ +\phi \left(\left[ \begin{array}{l} \mathbf {v} _ {i} \\ \mathbf {v} _ {j} \end{array} \right]\right) = \mathcal {N} ^ {- 1} (\cdot ; \mathbf {0}, \Lambda^ {\mathrm {p}}), \tag {9} +$$ + +where $\Lambda^{\mathrm{p}} = \mathbf{J}_{\mathrm{p}}^{\top}\mathrm{diag}(1 / \sigma_{\mathrm{p}}^{2},1 / \sigma_{\mathrm{p}}^{2})\mathbf{J}_{\mathrm{p}}$ and the jacobian $\begin{array}{r}\mathrm{J_p} = \left[ \begin{array}{rrrr} - 1 & 0 & 1 & 0\\ 0 & -1 & 0 & 1 \end{array} \right] \end{array}$ for computing the difference between the neighbor nodes. + +Table 1. Basic experimental setup. + +
ParameterESIMDVSMVSECDSEC
Window sizer [pix]5535
Δtpf [ms]40404040
Active durationτ [ms]5050100100
Prior stdσp [pix]0.110.51
Tangential stdσt [pix]10101010
Radial stdσr [pix]3333
Num. of layersL5555
Batch sizeNb10010010001000
+ +To model the outlier that can not be expressed as a Gaussian (quadratic cost), we employed the Huber cost both for measurement and prior factors. We adopted the technique of [1, 9] to maintain the Gaussian form. + +# 3.3.4 TEGBP + +Equipped with the sparse probabilistic graph, we now formalize the incremental message-passing algorithm base on BP (Sec. 3.1.2) for the marginalization of the graph to estimate the full flow (Fig. 5). As the graph of active nodes is loopy, GBP needs to store a belief at each variable node for iterative belief update. + +Message-passing is asynchronously triggered upon a space observation of normal flow. The measurement factor sends a message to the connected variable node and updates its belief with messages that have been sent from the active node as follows: + +$$ +\boldsymbol {\eta} _ {i} ^ {\mathrm {b}} = \sum_ {j \in \mathcal {C} _ {i} \cap \mathcal {A} _ {t}} \boldsymbol {\eta} _ {j \rightarrow i}, \quad \Lambda_ {i} ^ {\mathrm {b}} = \sum_ {j \in \mathcal {C} _ {i} \cap \mathcal {A} _ {t}} \Lambda_ {j \rightarrow i} \tag {10} +$$ + +where $b(\mathbf{v}_i) = \mathcal{N}^{-1}(\cdot ;\pmb{\eta}_i^{\mathrm{b}},\Lambda_i^{\mathrm{b}})$ is the belief on node $i$ $\mu_{j\rightarrow i} = \mathcal{N}^{-1}(\cdot ;\pmb{\eta}_{j\rightarrow i},\Lambda_{j\rightarrow i})$ is the message from node $j$ to node $i$ , and $\mathcal{C}_i$ is the 4-neighbors connected node $j$ . After that, the variable node sends the following message about its updated beliefs to its neighbors: + +$$ +\boldsymbol {\eta} _ {j \rightarrow i} = - \Lambda_ {0 1} ^ {\mathrm {p}} \left(\Lambda_ {1 1} ^ {\mathrm {p}} + \Lambda_ {j} ^ {\mathrm {b}} - \Lambda_ {j \rightarrow i}\right) ^ {- 1} \left(\boldsymbol {\eta} _ {j} ^ {\mathrm {b}} - \boldsymbol {\eta} _ {j \rightarrow i}\right) \tag {11} +$$ + +$$ +\Lambda_ {j \rightarrow i} = \Lambda_ {0 0} ^ {\mathrm {p}} - \Lambda_ {0 1} ^ {\mathrm {p}} \left(\Lambda_ {1 1} ^ {\mathrm {p}} + \Lambda_ {j} ^ {\mathrm {b}} - \Lambda_ {j \rightarrow i}\right) ^ {- 1} \Lambda_ {1 0} ^ {\mathrm {p}}, \tag {12} +$$ + +where, $\Lambda_{\alpha \beta}^{\mathrm{p}}$ is a $2\times 2$ sub-matrix of $\Lambda^{\mathrm{p}}$ indexed by $\alpha$ and $\beta$ . For a single normal flow measurement, the belief is propagated for $K$ hops, which are repeated for $N_{\mathrm{itr}}$ times (Fig. 5 correspond to $K = 2$ ). + +Note that our BP algorithm is guaranteed to converge to the correct MAP solution with sufficient iteration because all the belief is realized as a Gaussian (there is no guarantee for a variance because the graph is loopy) [30]. + +Coarse-to-fine message-passing scheme [10] is adopted to speed up the convergence. The measurement factor of + +![](images/305cf6bb5a5d82c5711e03a37b4b38af0926dad776d25cbd02e8b85e81269759.jpg) +Normal flow observation + +![](images/d1de7223f645a716656a89b157485fa2d512269cf4ab9c75d66073f30f10116e.jpg) +Belief update +Figure 5. A schematic showing a single iteration of TEGBP. (a) When a single normal flow has been measured, the message from the measurement factor is sent to the variable node as TEG, (b) the node updates its belief using the message received from the neighbor active nodes. (c) The nodes broadcast the updated belief to the neighbor, and (d) the neighbor nodes update their beliefs. (a)-(d) is repeated until convergence. To prevent the blowup of the message communication, it is stopped after $K$ hops. + +![](images/637f17ff8d66b9e36c917880457565e727db61e55f1969e0a8908bf05555e63d.jpg) +Message passing + +![](images/002a832e4a33fd70fd80e8f78b51cc9e9c1d898e756ed105f3fa9b759d16893b.jpg) +Belief update + +![](images/210c0528a1b25404df44cdb1ff5e92cd90a9f5a6b78b6cfaada8cbbf139722e5.jpg) +Message passing + +![](images/3983fd458933343f1a93926c1d114fa71726d30c59251d3641984918a05f74bd.jpg) +Belief update + +![](images/4b545dc779e1b4b5317b65256ce07c2ce15b80d0771f37b20cc7b492cf91ab2f.jpg) +Figure 6. The result on ESIM-bricks. The estimated full flow from ARMS and TEGBP are compared with ground-truth flow indicated by the red arrow. They use the same normal flow measurement (left). TEGBP succeeds while ARMS fails. + +![](images/c9b998636b6f33ffd39d910d02cbe01cdcb78557880f9cd8269036d8b4c159d9.jpg) +Time surface + +![](images/7bb31ba4a087d3392dd028262d2b1cd903e5ab5c6ece193967f52655546472f3.jpg) +ARMS [2] +Figure 7. The results on DVS-stripes. The histograms of the flow $(v_{x})$ of ARMS and TEGBP are compared. The two red lines indicate the GT flow of the two stripes (the upper stripes have a smaller norm). The result from ARMS are shifted from the GT flow, especially on the area of upper stripes (smaller norm) On the contrary, our TEGBP successfully estimates individual flows. + +![](images/b93a9e7730c455e20220832ce8d093345b72ec472b7d1c72b9a2ad9049cb0b5f.jpg) +Ours + +the coarser layer $l$ is computed as the sum of active measurement factor messages in a $2 \times 2$ block of finer layer $l - 1$ . The BP is executed coarse-to-fine; the coarse layer (message) estimation is copied to the finer layer as the initial guess at the start of the message passing on the finer layer (Supp-C.1). The short paths in the course graph improve efficiency for capturing long-range interactions and speed up convergence. + +# 4. Experiment + +We conducted experiments to compare the accuracy of TEGBP with the state-of-the-art incremental flow estimation algorithm called aperture-robust multi-scale flow estimation (ARMS) [2]. Both algorithms are incremental that could asynchronously estimate flow from sparse normal + +flow measurement. We used the same normal flow measurement for both algorithms for a fair comparison. + +# 4.1. Experimental setting + +# 4.1.1 Normal flow measurement + +We utilize the simple plane fitting to compute the normal flow measurements that serve as inputs to TEGBP and ARMS. We first apply noise removal for the raw events using a refractory filter [17] of $40~\mathrm{ms}$ duration. Then using the filtered events, we performed plane fitting for each event within a small spatiotemporal window of the size $r\times r\times \Delta t_{\mathrm{pf}}$ . During the plane fitting, we removed the outlier using the criteria proposed in [2]; we applied this process three times. We summarize the normal flow computation parameters in Tab. 1 (top). + +# 4.1.2 Full flow estimation + +Both TEGBP and ARMS use the same sparse normal flow measurement, and could be run asynchronously at the rate of normal flow; however, we process $N_{b}$ observation at once for computational efficiency on MATLAB [18] (interactive numerical computing environment) we used to compare the accuracy. The specific parameters for the (quasi) full flow computation are summarized in Tab. 1 (bottom). + +In TEGBP, the message should be propagated to all connected nodes, and the process needs to be repeated until convergence; however, we found a small number of hops $(K = 2)$ and single iterations $(N_{\mathrm{itr}} = 1)$ works well in practice. + +# 4.2. ESIM-bricks + +In this experiment, we compare the accuracy on a scene where the motion and edges are not perpendicular to the image edge. We synthesized the event stream by observing the image of bricks from a translating camera. We used ESIM [23] library for this simulation, which could synthesize the realistic noisy event stream from the event camera moving around the predefined 3D environment. We compute the normal flow and use the same measurements for both methods. Fig. 6 shows the result. We observe the drift in ARMS due to the average operation. TEGBP correctly + +Table 2. The quantitative results on MVSEC. The flow is evaluated on the ground truth intervals. + +
indoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm2.3024.53.6142.93.1336.33.4443.0
ARMS [2]1.7112.72.6726.62.2821.62.6425.7
Ours1.146.251.8716.41.5411.81.4611.1
Indoor_flying1
Indoor_flying3
Outdoor_day1
Time surfaceNormal flowARMS [2]OursGround truth
+ +Figure 8. The qualitative results on MVSEC. See the color wheel on the bottom left for the color encoding. + +estimated the full flow from the normal flow measurement directed at a different angle from the full flow. + +# 4.3.DVS-stirpes + +In this experiment, we compare the accuracy on a scene where some region includes different flows. We used a scene from [19] where two stripes translating at different speeds are observed from a dynamic vision sensor (DVS). Fig. 7 shows the result. We observed a significant shift in the ARMS's estimation when motions of different speeds are adjacent. ARMS determines the flow by selecting the scale having the maximum average norm in the window (it falsely picks the large norm in the slow region). In contrast, the TEGBP correctly estimates individual motion. + +# 4.4. MVSEC + +We quantitatively compare with ARMS in practical robotic scenarios. We used outdoor scenes captured by automobiles and indoor scenes captured by aerial drones from the Multi-vehicle Stereo Event Camera (MVSEC) dataset [33]. This dataset includes ground truth (GT) optical flow computed as the motion filed from the camera motion and scene depth. We evaluate the accuracy using the average + +endpoint error (AEE) and the percentage of pixels with endpoint error greater than 3 pixels (denoted by % Out). Both are measured on pixels where valid GT exists, and the plane fitting has been successful. We output flows asynchronously in batches with both methods and evaluate them at regular intervals (20Hz of GT rate). Tab. 2 and Fig. 8 show the quantitative and qualitative results. ARMS improves the normal flow, while TEGBP improves more. In contrast to the naive averaging in ARMS, our TEGBP properly models the distribution of flows using TEG and prior. + +# 4.5. DSEC + +We compare the performance of TEGBP and ARMS on more diverse scenarios using more high-resolution input. To this end, we use the DSEC [13], a recently released dataset in automobile scenarios containing high-resolution event data of $480 \times 640$ captured from the vehicle that undergoes more diverse motion. It is designed for dense optical flow estimation, and the GT of the test sequence is hidden for leaderboard evaluation. Therefore we use their train sequence to compare sparse AEE error with ARMS. + +Tab. 3 and Fig. 9 shows the quantitative and qualitative results. Similar to the MVSEC experiments, TEGBP sig + +Table 3. The quantitative results on DSEC. The flow is evaluated on the ground truth intervals. + +
thun_00_azurich_city_01_azurich_city_02_azurich_city_08_azurich_city_11_a
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm7.0968.78.2637.912.886.88.4273.35.9063.4
ARMS [2]5.3160.06.3223.88.9878.76.2264.44.5751.4
Ours4.2645.55.6123.28.7072.65.0849.53.2732.2
+ +![](images/b7f6bdf484a396f0b3e25c9d2517d13fc279f7c677748e16859f8324cd6b5112.jpg) +Figure 9. The qualitative results on DSEC. + +nificantly boosts accuracy over AMRS. + +# 5. Conclusion + +We proposed TEGBP, a novel incremental full flow estimation algorithm for asynchronous event data. The proposed algorithm is backed up by the probabilistic model, which guarantees equality to the full flow when a valid prior is provided. It realizes efficient event-wise incremental updates based on the BP, where the MAP solution obtained by the local update is guaranteed to match the correct one. We demonstrate the effectiveness of the proposed method on a practical real-world dataset showing a significant boost in accuracy over the existing incremental algorithm. + +There is still much room for improvement in accuracy compared to the recent machine learning-based method (although it is computationally intensive). We suspect the gap mainly comes from the noise in the normal flow and poor prior. Toward the ultimate goal of realizing low-latency and high-accuracy full flow estimation in real-time on an edge device, we expect the gap to be closed by incorporating a more robust normal flow and application-specific prior. + +# 5.1. Future work + +Robust normal flow estimation As discussed in the experiment section, some scenes include strongly erroneous flow, e.g., normal flow from leaves on the tree. These measurement errors significantly degrade the accuracy. We expect the tiny CNN (e.g., input size of $7 \times 7$ ) trained to regress + +the normal flow from noisy input will improve the accuracy without sacrificing the computational efficiency. + +Application specific prior Even if the perfect normal flow is obtained, we cannot expect good accuracy without good prior (Sec. 3.3.2). The uniform smoothness prior we adopted may not optimal in many practical scenarios. The discontinuous depth region does not conform to the prior. Even if the planer object undergoes linear motion, we won't observe uniform flow when it is not perpendicular to the optical axis. We can utilize a better prior for a specific application to boost accuracy. For example, one may utilize a vanishing point for automotive applications (assuming the yaw rate is known from the vehicle sensor) similar to [20]. We left these explorations for future work. + +Optimized implementation Our algorithm is very efficient in principle, requiring lower FLOPS, and the entire process can be asynchronously parallelized. The normal flow could be cheaply obtained, e.g., directly from a camera or computed by the lightweight plane fitting (computation of normal flow is outside this study's scope). The core of our algorithm, message passing of (4), is also very cheap, requiring the communication of 6-dim vectors with neighbors. Therefore, an optimized CPU implementation or adaptation of processors supporting the sparse computation could realize low-energy, low-latency optical flow estimation in the edge device. We left the wall clock time and energy consumption evaluation as future work. + +# References + +[1] Pratik Agarwal, Gian Diego Tipaldi, Luciano Spinello, Cyril Stachniss, and Wolfram Burgard. Robust Map Optimization using Dynamic Covariance Scaling. IEEE International Conference on Robotics and Automation (ICRA), 2013. 5 +[2] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):1-12, 2020. 1, 2, 4, 6, 7, 8, 11 +[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. IEEE International Symposium on Circuits and Systems, 2018. 2 +[4] R. Wes Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event Probability Mask (EPM) and Event Denoising Convolutional Neural Network (EDnCNN) for Neuromorphic Cameras. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1701-1710, 2020. 3 +[5] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous Optical Flow and Intensity Estimation from an Event Camera. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2 +[6] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, feb 2014. 1, 2 +[7] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, mar 2012. 1, 2 +[8] Andrew J. Davison. FutureMapping: The Computational Structure of Spatial AI Systems. arXiv, 2018. 3 +[9] Andrew J. Davison and Joseph Ortiz. FutureMapping 2: Gaussian Belief Propagation for Spatial AI. arXiv, 2019. 2, 3, 5 +[10] Pedro F Felzenszwalb and Daniel P Huttenlocher. Ef ci e n t Belief Propagation for Early Vision. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 5, 13 +[11] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Joerg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1), 2020. 2 +[12] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 2 +[13] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 7 + +[14] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. International Conference on 3D Vision (3DV), 2021. 2 +[15] B. K. B. Horn and B. G. Schunck. Determining Optical Flow. Artificial Intelligence, 17(1-3):185-203, 1981. 1 +[16] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Bill Nell, Nir Shavit, and Dan Alistarh. Inducing and Exploiting Activation Sparsity for Fast Neural Network Inference. Proceedings of Machine Learning Research, 119:5533-5543, 2020. 3 +[17] Weng Fei Low, Zhi Gao, Cheng Xiang, and Bharath Ramesh. SOFEA: A non-iterative and robust optical flow estimation algorithm for dynamic vision sensors. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 2, 6 +[18] MATLAB. version 9.13.0 (R2022b). The MathWorks Inc., Natick, Massachusetts, 2022. 6 +[19] Elias Mueggler, Christian Forster, Nathan Baumli, Guillermo Gallego, and Davide Scaramuzza. Lifetime Estimation of Events from Dynamic Vision Sensors. IEEE International Conference on Robotics and Automation (ICRA), 2015. 2, 7 +[20] Jun Nagata, Yusuke Sekikawa, Kosuke Hara, and Yoshimitsu Aoki. FOE-based Regularization for Optical Flow Estimation from an In-vehicle Event Camera. Proceedings of SPIE - The International Society for Optical Engineering, 11049, 2019. 8 +[21] Joseph Ortiz, Mark Pupilli, Stefan Leutenegger, and Andrew J. Davison. Bundle Adjustment on a Graph Processor. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2413-2422, 2020. 3 +[22] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single Image Optical Flow Estimation with an Event Camera. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[23] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an Open Event Camera Simulator. Conf. on Robotics Learning (CoRL), 87:969-982, 2018. 6 +[24] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. MDPI Sensors, 22(14):1-20, 2022. 2 +[25] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. European Conference on Computer Vision (ECCV), pages 1-23, 2022. 2 +[26] Chen Shoushun and Guo Menghan. Live Demonstration : CeleX-V: a 1M Pixel Multi-Mode Event-based Sensor Chen Shoushun. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1–2, 2019. 3 +[27] Timo Stoffregen and Lindsay Kleeman. Event Cameras, Contrast Maximization and Reward Functions: an Analysis. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12300-12308, 2019. 2 +[28] Daniel C. Stumpp, Himanshu Akolkar, Alan D. George, and Ryad B. Benosman. hARMS: A Hardware Acceleration Architecture for Real-Time Event-Based Optical Flow. IEEE Access, 10:58181-58198, 2022. 3 + +[29] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. European Conference on Computer Vision (ECCV), 2020. 1, 2 +[30] Yair Weiss and William T Freeman. Correctness of belief propagation in Gaussian graphical models of arbitrary topology. Neural Information Processing Systems (NIPS), pages 673-679, 2000. 5 +[31] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion from Sparse Event Data. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 2 +[32] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. Proceedings of Robotics: Science and Systems, jun 2018. 2 +[33] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multi Vehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 7, 14 +[34] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. European Conference on Computer Vision (ECCV) Workshop, pages 711-714, 2018. 2 +[35] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-based Learning of Optical Flow, Depth, and Egomotion. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 989–997, 2019. 2 \ No newline at end of file diff --git a/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/images.zip b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ac81bd3d2401d97709c1ddeb40213ec7f59ef452 --- /dev/null +++ b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74b97bbbdcb1b3997c729cc17b4c6305a54d8355573c55268405cffc73d9f081 +size 606111 diff --git a/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/layout.json b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..65cc85347eb662444b63effd8c575fb8ec4af81c --- /dev/null +++ b/2023/Tangentially Elongated Gaussian Belief Propagation for Event-Based Incremental Optical Flow Estimation/layout.json @@ -0,0 +1,8805 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 462, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 462, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 462, + 138 + ], + "type": "text", + "content": "Tangentially Elongated Gaussian Belief Propagation for Event-based Incremental Optical Flow Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 459, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 459, + 175 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 459, + 175 + ], + "type": "text", + "content": "Jun Nagata† and Yusuke Sekikawa† DENSO IT LAB., INC., Japan." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 202, + 209, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 202, + 209, + 216 + ], + "spans": [ + { + "bbox": [ + 162, + 202, + 209, + 216 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 290, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 526 + ], + "type": "text", + "content": "Optical flow estimation is a fundamental functionality in computer vision. An event-based camera, which asynchronously detects sparse intensity changes, is an ideal device for realizing low-latency estimation of the optical flow owing to its low-latency sensing mechanism. An existing method using local plane fitting of events could utilize the sparsity to realize incremental updates for low-latency estimation; however, its output is merely a normal component of the full optical flow. An alternative approach using a frame-based deep neural network could estimate the full flow; however, its intensive nonincremental dense operation prohibits the low-latency estimation. We propose tangentially elongated Gaussian (TEG) belief propagation (BP) that realizes incremental full-flow estimation. We model the probability of full flow as the joint distribution of TEGs from the normal flow measurements, such that the marginal of this distribution with correct prior equals the full flow. We formulate the marginalization using a message-passing based on the BP to realize efficient incremental updates using sparse measurements. In addition to the theoretical justification, we evaluate the effectiveness of the TEGBP in real-world datasets; it outperforms SOTA incremental quasi-full flow method by a large margin. (The code is available at https://github.com/DensoITLab/tegbp/)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "content": "Optical flow estimation, which computes the correspondence of pixels in different time measurements, is a fundamental building block of computer vision. One needs to estimate the flow at low latency in many practical applications, such as autonomous driving cars, unmanned aerial vehicles, and factory automation robots. Most of the existing optical flow algorithm utilizes dense video frames; it computes the flow by searching the similar intensity pattern [15, 29]. Recently, methods using deep neural network (DNN) [29] demonstrate impressive accuracy at the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 194, + 547, + 276 + ], + "blocks": [ + { + "bbox": [ + 307, + 194, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 307, + 194, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 547, + 276 + ], + "type": "image", + "image_path": "0553784f9a4b0d0b55180ab1194f8293bc38feac0ef3b0f32b87927b1e320d8b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 282, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 547, + 350 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed TEGBP. We model a belief about the full flow from a normal flow measurement using TEG (RGB ellipse). The mean of the marginal distribution of each TEG with an appropriate prior equals the full flow (magenta arrow). The marginal (magenta ellipse) is computed incrementally using local message passing (RGB arrows) based on BP." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 373, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 444 + ], + "type": "text", + "content": "cost of higher-computational cost. Either model-based or DNN-based, the frame-based algorithm needs to compute the entire pixel for every frame, even when there are subtle changes or no changes at all. This dense operation makes it difficult to realize low-latency estimation, especially on the resource-constrained edge device." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 445, + 546, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 685 + ], + "type": "text", + "content": "The event-based camera is a bio-inspired vision sensor, which asynchronously detects intensity change on each pixel. Thanks to the novel sensing mechanism, the camera equips favorable characteristics for optical flow estimation, such as high dynamic range (HDR), blur-free measurement, and, most importantly, sparse low-latency data acquisition. Many researchers have explored the way to utilize sparsity to realize efficient low-latency estimation; one extends the well-known Lucas-Kanade algorithm [7], and the other exploits the local planer shape of the spatiotemporal event streams [6]. These methods could utilize the sparsity for efficient incremental processing; however, the optical flow computed in this way (e.g., by plane fitting) is the normal flow, which is a normal component of full flow and often different from them1 we want to obtain. The normal flow is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient). Some work tried to recover the full flow from the normal flow [2]; however, it does not precisely equal the full flow (refer to Sec. 2). There exist methods that could estimate full flow, such as a varia" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 220, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 220, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 220, + 712 + ], + "type": "text", + "content": "† These authors contributed equally to this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 693, + 546, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 693, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 693, + 546, + 713 + ], + "type": "text", + "content": "1 full flow is usually simply called optical flow or flow, yet, we use full flow when we want to highlight the difference with the normal flow." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21940" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "tional method [5], a multi-scale extension of contrast maximization [25], or a frame-based DNN [14]. However, they need to apply non-incremental dense operation for all the pixels of the event frame (dense representation constructed from sparse events) for every frame, which prohibits the low-latency estimation on the edge device." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 290, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 290, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 290, + 384 + ], + "type": "text", + "content": "Our research goal is to realize an incremental full flow algorithm from sparse normal flow measurements. To this end, we propose Tangentially Elongated Gaussian (TEG) Belief Propagation (BP). We compute the full flow using the normal flow measurements, which can be observed directly from an event camera or computed cheaply using an existing algorithm2. Notice that given a single measurement of normal flow, there are infinite possibilities for the full flow along the tangential direction of the normal flow. We model this uncertainty using the TEG, Gaussian distribution, which has a large variance along the tangential direction of the normal flow. The probability density of full flow on each pixel is given as the margins of the joint distribution of TEG data factor and some prior factor on a sparse graph (Fig. 1, Sec. 3.3.2). We leverage the sparse graph to formulate the incremental full flow estimation algorithm using message-passing based on BP [9]. We evaluate the effectiveness of the TEGBP on real-world data captured from aerial drones and automobiles. TEGBP outperforms SOTA incremental method [2] by a large margin." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 396, + 132, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 132, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 132, + 407 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": "This section reviews the related literature on optical flow estimation using sparse event signals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 447, + 259, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 259, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 259, + 460 + ], + "type": "text", + "content": "2.1. Incremental Algorithm for Normal Flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 586 + ], + "type": "text", + "content": "Lucas-Kanade based method The traditional frame-based Lucas-Kanade method finds the flow by aligning the small patch of intensity between consecutive frames using its spatial gradient and temporal change. The event-based counterpart [7] uses the integration of events to approximate them for each incoming event. This approach realized highly efficient estimation by utilizing the sparsity of input events. However, it can only estimate the normal flow, which is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient) [11]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 685 + ], + "type": "text", + "content": "Plane fitting based method When an edge undergoes liner motion, triggered events generate a time surface, which can be locally approximated by a plane in a spatiotemporal space. The local plane fitting method [3,6,17, 19] computes the optical flow by fitting the plane to the sparse event points. However, the flow obtained by plane fitting is also the normal component of the full flow." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": "Recently, a method for computing the quasi-full flow has been proposed [2]; it estimates the full flow by considering the average normal flow in multiple resolutions. It incrementally updates the average and selects the flow with the largest average norm from the multiple resolutions. It expects the flow to have the largest norm to correspond to the full flow. The average operation induces bias on the estimated flow; therefore, it equals the full flow on the very limited scene where the true motion in the window is orthogonal to the edge direction (refer to Sec.5 in their paper). We'll also experimentally compare this point in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 545, + 227 + ], + "type": "text", + "content": "In summary, current sparse incremental algorithms are limited to normal flow." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 236, + 471, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 471, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 471, + 248 + ], + "type": "text", + "content": "2.2. Dense Algorithm for Full Flow" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 253, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 253, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 253, + 545, + 338 + ], + "type": "text", + "content": "Variational method A variational optimization-based approach realizes the full flow estimation by incorporating intensity frame obtained either by the event/frame hybrid camera [22] or by simultaneously estimating image intensity [5]. It could estimate full flow; however, it requires additional intensity observation or intensive dense optimization to recover the intensity frame." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 352, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 545, + 544 + ], + "type": "text", + "content": "Contrast maximization method Contrast maximization (CMax) [12, 27] computes the underlying motion (optical flow) of events stream by finding their alignment. It computes the alignment by maximizing the focus score (e.g., variance) of an image generated by warping the events (IWE) using the optical flow parameter. The original algorithm can not be used to estimate pixel-wise flow because it tends to converge to the degenerated solution (e.g., warps all events to a single point) [24, 35]. Recently, a hierarchical CMax approach avoiding the degenerated solution for pixel-wise estimation has been proposed [25]. Although this approach works well even in the practical scenario, it is not incremental; it must recompute the IWE for each sliding time window by using all events in the patch for every hierarchical scale (this involves intensive iteration by itself), making it difficult to operate in real-time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 557, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 689 + ], + "type": "text", + "content": "DNN based method Frame-based DNN is often utilized to compute the full flow [14, 31, 32, 34]. Recently, a frame-based high-accuracy model called RAFT [29] has been imported into the event-based vision literature showing the SOTA performance for the dense flow estimation [14]. These methods can not consume sparse events directly; the sparse input must be converted into dense frame representation. The dense convolution is computationally expensive because it needs to operate for all pixels, even when there are subtle changes or no changes at all, making it difficult to operate at a higher rate on the edge device." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In summary, some dense algorithms could estimate full flow; however, it compromises the event's sparse asyn" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "text", + "content": "2We consider the normal flow as measurement and discuss algorithms to process the sparse measurement of normal flow to estimate the full flow." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21941" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "chronous nature, and their non-incremental dense operation prohibits them from realizing low-latency processing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 286, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 286, + 118 + ], + "type": "text", + "content": "2.3. Sparse Asynchronous Computing Architecture" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 125, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 288, + 281 + ], + "type": "text", + "content": "Events are asynchronously triggered only from limited pixels which detect intensity changes. Several techniques have been developed for efficient processing of the sparse signals on CPUs by taking advantage of the highly optimized cache mechanism [16, 28]. Besides, sparse data-flow architectures [8, 9], which differ from prevalent Neuman-type computing architectures, are gaining attention for sparse signal processing. For example, bundle adjustment has been solved on the novel architecture using an algorithm based on Gaussian BP [21]. Our goal is to establish the incremental full flow estimation algorithm by taking advantage of the sparsity of events, which is expected to run efficiently on CPUs or emerging data-flow processors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 293, + 151, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 151, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 151, + 306 + ], + "type": "text", + "content": "3. Proposed method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 313, + 126, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 126, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 126, + 326 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 332, + 133, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 133, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 133, + 342 + ], + "type": "text", + "content": "3.1.1 Normal flow" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "text", + "content": "The normal flow " + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^\\perp \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "text", + "content": " on pixel " + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "text", + "content": " is a normal component of full flow " + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 351, + 287, + 448 + ], + "type": "text", + "content": ", which we want to estimate. One can not decide the full flow uniquely from the normal flow, and there are infinitely many possibilities along the line perpendicular to the normal flow (Fig. 3 left). There are several options for obtaining the normal flow; some event-based cameras, such as CeleX-V [26] directly observe it, or it can be cheaply computed from the sparse events (Sec. 2.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 567 + ], + "type": "text", + "content": "The moving edge in 3D space triggers events; we can compute the normal flow by finding the local plane on the triggered events. In this study, we adopt a naive least-squares fitting of the plane using events in a small spatiotemporal window. An erroneous normal flow may significantly deteriorate the full flow accuracy, especially on highly textured regions such as leaves on trees. One can improve the results by utilizing an advanced noise removal algorithm [4] or by adopting more sophisticated normal flow estimation methods, such as small neural networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 159, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 159, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 159, + 597 + ], + "type": "text", + "content": "3.1.2 Belief propagation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Factor graphs geometrically represent the structure of probabilistic problems. A factor graph " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "G = (V, F, E)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is composed of a set of variable nodes " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "V = \\{\\mathbf{v}_i\\}_{i=1:N_v}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", a set of factor nodes " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "F = \\{f_s\\}_{s=1:N_f}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " and a set of edges " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". Each factor node " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " represents a probabilistic constraint " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "f_s(V_s)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " between a subset of variables " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "V_s \\subset V" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". The factorization of the variables is explicitly represented in the graph by connecting factor nodes with the variable nodes they depend on. Probabilistically, these factors are the independent" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 63, + 542, + 155 + ], + "blocks": [ + { + "bbox": [ + 310, + 63, + 542, + 155 + ], + "lines": [ + { + "bbox": [ + 310, + 63, + 542, + 155 + ], + "spans": [ + { + "bbox": [ + 310, + 63, + 542, + 155 + ], + "type": "image", + "image_path": "59f55debec1fc390bc0d1fbbc73c88e7024f8e93740ded52eff28fa7c4c044c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "text", + "content": "Figure 2. Processing pipeline. Observation set at " + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "inline_equation", + "content": "V_{t}^{\\perp}" + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "text", + "content": " where the normal flow is sparsely assigned. The TEGBP incrementally update its estimation about full flow " + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "inline_equation", + "content": "V_{t + 1}" + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "text", + "content": " using previous estimation " + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "text", + "content": " and new observation " + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,t + 1}^{\\perp}" + }, + { + "bbox": [ + 305, + 159, + 545, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 235, + 471, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 471, + 246 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 471, + 246 + ], + "type": "text", + "content": "terms that make up the joint distribution:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 255, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 255, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 383, + 255, + 545, + 289 + ], + "type": "interline_equation", + "content": "p (V) = \\prod_ {s = 1} ^ {N _ {f}} f _ {s} \\left(V _ {s}\\right). \\tag {1}", + "image_path": "94ca24f7ae84f8716661a2b639d781113356b770c32e74c51a663510c85b928f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 301, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 445 + ], + "type": "text", + "content": "Belief propagation (BP) is a generic distributed algorithm for computing the marginal distribution of a set of variables from their joint distribution. The marginal for a single variable " + }, + { + "bbox": [ + 305, + 301, + 545, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 305, + 301, + 545, + 445 + ], + "type": "text", + "content": " is computed by the integral over all the other variables. BP computed the marginals asynchronously using message-passing between connected nodes on the graph. It is substantially more efficient by leveraging the topology of the sparse graph than naive integration of the entire graph (which involves giant matrix inversion). Refer to the seminal work of [9] for more detail about the derivation, condition for convergence guarantee, and discussion about the stability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 453, + 427, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 427, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 427, + 464 + ], + "type": "text", + "content": "3.2. Problem formulation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": "Given measurements of normal flow " + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "inline_equation", + "content": "V_{t}^{\\perp}" + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": " on the the pixel grid where each normal flow is assigned to the corresponding pixels (within time interval " + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "inline_equation", + "content": "[t - \\tau, t]" + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": "), we want to estimate the posterior distribution of the corresponding full flow " + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "inline_equation", + "content": "p(V_{t} \\mid V_{t}^{\\perp})" + }, + { + "bbox": [ + 305, + 471, + 545, + 555 + ], + "type": "text", + "content": ". Once the posterior distribution of full flow is obtained, the actual full flow output is given by the maximum a posteriori (MAP) estimate as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 562, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 562, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 370, + 562, + 545, + 583 + ], + "type": "interline_equation", + "content": "\\hat {V} _ {t} = \\underset {V _ {t}} {\\arg \\max } p \\left(V _ {t} \\mid V _ {t} ^ {\\perp}\\right). \\tag {2}", + "image_path": "6a8fd8128dcd30699a04d9cda3fc42fa6cc2250453945f803a6da206dc9e49bf.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "text", + "content": "Our goal is to formulate an incremental full flow estimation algorithm using sparse observations of local normal flows (Fig. 2). Solving the arg max of the joint distribution of " + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "text", + "content": " for every incoming normal flow measurement is infeasible both in terms of memory and computational perspective; therefore, we want to formulate " + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "text", + "content": " as an incremental sparse update function, i.e., new observation interacts sparsely with the previous results " + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "inline_equation", + "content": "p(V_{t} \\mid V_{t}^{\\perp})" + }, + { + "bbox": [ + 305, + 592, + 545, + 689 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 331, + 696, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 696, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 331, + 696, + 545, + 711 + ], + "type": "interline_equation", + "content": "f: \\left[ p \\left(V _ {t} \\mid V _ {t} ^ {\\perp}\\right), p \\left(\\mathbf {v} _ {i, t + 1} ^ {\\perp}\\right) \\right] \\mapsto p \\left(V _ {t + 1} \\mid V _ {t + 1} ^ {\\perp}\\right). \\tag {3}", + "image_path": "057db780f8b35a23299934a331757c6cad0354c984d778b4aa4643b971cef2c8.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21942" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 62, + 170, + 153 + ], + "blocks": [ + { + "bbox": [ + 66, + 62, + 170, + 153 + ], + "lines": [ + { + "bbox": [ + 66, + 62, + 170, + 153 + ], + "spans": [ + { + "bbox": [ + 66, + 62, + 170, + 153 + ], + "type": "image", + "image_path": "6ea688f256f599a2f3a2f06f94e7c744a8f2d74ec3dc782137026a9e9acf59f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 156, + 288, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 288, + 201 + ], + "type": "text", + "content": "Figure 3. The relation of normal flow, full flow, and TEG. There are infinitely possible candidates for full flows given a normal flow (left). TEG models this uncertainty using Gaussian having large variance along the tangential direction of the normal flow (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 62, + 269, + 153 + ], + "blocks": [ + { + "bbox": [ + 179, + 62, + 269, + 153 + ], + "lines": [ + { + "bbox": [ + 179, + 62, + 269, + 153 + ], + "spans": [ + { + "bbox": [ + 179, + 62, + 269, + 153 + ], + "type": "image", + "image_path": "2c3de6766bfb9590737604dd11f71695831a4b4fd687b4172cba7607a6c833e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 236, + 287, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 259 + ], + "type": "text", + "content": "In the following section, we'll derive TEGBP, an algorithm based on BP for incremental full flow estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 272, + 241, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 241, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 241, + 285 + ], + "type": "text", + "content": "3.3. Tangentially Elongated Gaussian BP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 222, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 222, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 222, + 306 + ], + "type": "text", + "content": "3.3.1 Tangentially Elongated Gaussian" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "type": "text", + "content": "The true full flow " + }, + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "type": "text", + "content": " exists somewhere along the line perpendicular to the normal flow " + }, + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^\\perp" + }, + { + "bbox": [ + 46, + 315, + 288, + 411 + ], + "type": "text", + "content": " (Fig. 3 left). In other words, we are uncertain about the full flow from the single normal flow measurement. We model this uncertainty using tangentially elongated Gaussian (TEG), A Gaussian distribution having large variance along the tangential direction of the observed normal flow (Fig. 3 right). We defined the precision matrix of TEG as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 423, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 423, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 123, + 423, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\Lambda^ {\\mathrm {t e g}} = \\mathbb {R} (\\theta) \\Lambda^ {\\mathrm {m}} \\mathbb {R} (\\theta) ^ {\\top}, \\tag {4}", + "image_path": "a5f5cedb2cc6b94ad3b0e8b1058a066a9a99ab69f06de64fdceca1bb038b182b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}(\\theta)" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " is 2D rotation matrix parameterized by the rotation angle " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " of the normal flow where " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta = \\arctan (v_y^\\perp / v_x^\\perp)" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Lambda^{\\mathrm{m}} = \\mathrm{diag}(1 / \\sigma_{\\mathrm{r}}^{2}, 1 / \\sigma_{\\mathrm{t}}^{2})" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": ". In principle, " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}} = \\infty" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{r}} = 0" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " correspond to the constraint of the full flow from a single noiseless normal flow measurement. In this case, one could parameterize the observation with a 1D projection of full flow (normal flow), and the uncertainty of the projected vector length is parameterized as normal variance, " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": ". In other words, this corresponds to the case when we are equally uncertain along the tangential direction. In a real-world scenario, the probability of full flow along the tangential direction is not uniform but highly concentrated around the normal flow (Supp-D). The knowledge about the probability density of full flow is a vital clue to recovering the accurate full flow from the inherently ill-posed observation of sparse normal flow; however, the 1D parameterization can not incorporate this vital information. Therefore, we propose to utilize 2D Gaussian (TEG) by incorporating additional parameters, i.e., finite " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}}" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": ". The 2D parameterization is crucial for estimating the accurate full flow from sparse and noisy measurements, and our TEG (2D) includes the 1D parameterization as a special case when " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}} = \\infty" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "**Corner pixel** When the naive least-square planer fitting is adopted, the quality for the normal flow estimation might be low; we adjust " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}}, \\sigma_{\\mathrm{r}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " depending on the quality. We use lower precision " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\Lambda^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " depending on the mean square error between observed events and the estimated plane. Our BP algorithm automatically incorporates this confidence to estimate accurate flow even at the corner point." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "content": "Remark: If full flow observation is available at the corner (e.g., by corner tracking), TEGBP could utilize the reliable information by TEG with smaller " + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 156, + 545, + 192 + ], + "type": "text", + "content": " to improve accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 206, + 514, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 206, + 514, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 514, + 217 + ], + "type": "text", + "content": "3.3.2 Marginalization of TEG with Valid Prior" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "content": "As we saw earlier, the full flow estimation from the normal flow is inherently an ill-pose without some prior knowledge about the consistency with other pixels (when " + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}} = \\infty" + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{r}} = 0" + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "content": "). Without a valid prior, there are infinitely many possibilities along the line perpendicular to the normal flow; any point on the line has the same likelihood when " + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}} = \\infty" + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "content": ". To obtain a correct solution, we need the correct prior. Our algorithms could incorporate any prior, e.g., simple uniform smoothness (e.g., total variation), an application-specific knowledge such as flow converges to the vanishing-point in an automotive scenario, or they could be learned using neural networks. In this study, we adopt uniform smoothness for simplicity. The important fact is that the mean of the marginal distribution of TEGs having " + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}} = \\infty" + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "content": " and the correct prior equals the correct full flow, and this still approximately holds when " + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 224, + 545, + 417 + ], + "type": "text", + "content": " is moderately large (Sec. B.1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": "Example. To see this, let's consider the case of Fig. 1. We assume the corner having an edge angle of " + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\theta_{1} + \\frac{\\pi}{2}" + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\theta_{3} + \\frac{\\pi}{2}" + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": " moves linearly with uniform velocity (all pixels have the same velocity) toward the direction of " + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\theta_{0}" + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}} = [\\cos \\theta_{0},\\sin \\theta_{0}]^{\\top}" + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": ". The correct prior in this situation is uniform (infinitely strong smoothness prior; " + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\sigma_{p} = 0.0" + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": "). In this scenario, we'll observed " + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{1}^{\\perp} = \\mathbf{v}_{2}^{\\perp} = [\\cos (\\theta_{1} - \\theta_{0})\\cos \\theta_{1},\\cos (\\theta_{1} - \\theta_{0})\\sin \\theta_{1}]^{\\top},\\mathbf{v}_{3}^{\\perp} = [\\cos (\\theta_{3} - \\theta_{0})\\cos \\theta_{3},\\cos (\\theta_{3} - \\theta_{0})\\sin \\theta_{3}]^{\\top}" + }, + { + "bbox": [ + 304, + 422, + 545, + 530 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 530, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 555 + ], + "type": "text", + "content": "The joint distribution of full flow " + }, + { + "bbox": [ + 304, + 530, + 545, + 555 + ], + "type": "inline_equation", + "content": "V \\coloneqq \\{\\mathbf{v}_1, \\mathbf{v}_2, \\mathbf{v}_3\\}" + }, + { + "bbox": [ + 304, + 530, + 545, + 555 + ], + "type": "text", + "content": " from normal flow measurements and the prior is given as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 309, + 561, + 545, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 561, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 309, + 561, + 545, + 593 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (V) = \\mathcal {N} (\\mathbf {v} _ {1}; \\mathbf {v} _ {1} ^ {\\perp}, \\Sigma_ {\\theta_ {1}}) \\mathcal {N} (\\mathbf {v} _ {2}; \\mathbf {v} _ {2} ^ {\\perp}, \\Sigma_ {\\theta_ {2}}) \\mathcal {N} (\\mathbf {v} _ {3}; \\mathbf {v} _ {3} ^ {\\perp}, \\Sigma_ {\\theta_ {3}}) \\\\ \\mathcal {N} ([ \\mathbf {v} _ {1}; \\mathbf {v} _ {2} ]; \\sigma_ {p} ^ {2} \\mathrm {I}) \\mathcal {N} ([ \\mathbf {v} _ {2}; \\mathbf {v} _ {3} ]; \\sigma_ {p} ^ {2} \\mathrm {I}). \\tag {5} \\\\ \\end{array}", + "image_path": "4afd5b54f68f989f0effcf24b6ce2a193ca2b3feb836f7a845c00fd59789d57f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": "Marginalizing for " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_2" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "p(\\mathbf{v}_2) = \\int p(V)d\\mathbf{v}_1d\\mathbf{v}_3" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " and putting " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\sigma_t = \\infty" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\sigma_r = 0" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": ", we'll get " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "p(\\mathbf{v}_2) = \\mathcal{N}\\left(\\mathbf{v}_2;[\\cos \\theta_0,\\sin \\theta_0]^\\top ,\\mathrm{diag}(0,0)\\right)" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": "; the MAP of this marginal equals to the true flow " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "One can also visually verify this result by considering the intersection of the lines of the major axis of the TEG. Notice that the marginal is equal to the full flow even when an average of the normal flow, such as employed in [2], differs from the full flow. In the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21943" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 62, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 62, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 62, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 62, + 286, + 167 + ], + "type": "image", + "image_path": "eab2ad5eb23ce9bb3d9e701035eab2c030c3ffceec0a0f6f5dd675854191f250.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "type": "text", + "content": "Figure 4. 2D factor graph of TEGBP. The topology changes dynamically depending on the active set " + }, + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 46, + 169, + 288, + 224 + ], + "type": "text", + "content": ". The measurement node and the active variable node on the same pixel are connected by measurement factor node, and each neighbor active variable node is connected via prior factor node." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "type": "text", + "content": "above case, the average normal flow " + }, + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "type": "inline_equation", + "content": "(\\mathbf{v} = [(2\\cos(\\theta_1 - \\theta_0)\\cos\\theta_1 + \\cos(\\theta_3 - \\theta_0)\\cos\\theta_3)/3, (2\\cos(\\theta_1 - \\theta_0)\\sin\\theta_1 + \\cos(\\theta_3 - \\theta_0)\\sin\\theta_3)/3]^\\top)" + }, + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "type": "text", + "content": " is different from the true flow " + }, + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "type": "inline_equation", + "content": "([\\cos\\theta_0,\\sin\\theta_0]^\\top)" + }, + { + "bbox": [ + 46, + 254, + 287, + 338 + ], + "type": "text", + "content": ". This is a steak difference from the existing incremental sparse method using an average of normal flow and is a key to realizing a significant accuracy boost over the technique." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 352, + 194, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 194, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 194, + 363 + ], + "type": "text", + "content": "3.3.3 Factor Graph of Full Flow" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 370, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 407 + ], + "type": "text", + "content": "Now we construct the sparse factor graph for full flow using the TEG. The joint of Eq. (2) can be factorized using Bayes theorem as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 413, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 413, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 101, + 413, + 287, + 434 + ], + "type": "interline_equation", + "content": "\\hat {V} _ {t} = \\underset {V _ {t}} {\\arg \\max } p \\left(V _ {t} ^ {\\perp} \\mid V _ {t}\\right) p \\left(V _ {t}\\right). \\tag {6}", + "image_path": "03de27ed2ebea54f136f9e208581448ac0d6f4f8dd836041d20bd6a5e88dd37b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "type": "text", + "content": "Assuming independence of observations and dependence between neighboring-pixel variables, Eq. (6) is further factorized to the measurement factor " + }, + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "type": "text", + "content": " and the prior factor " + }, + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 441, + 287, + 491 + ], + "type": "text", + "content": " defined on a sparse graph on the image grid (Fig. 4) as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 497, + 287, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 497, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 87, + 497, + 287, + 524 + ], + "type": "interline_equation", + "content": "p (V) = \\prod_ {n \\in \\mathcal {A} _ {t}} \\psi (\\mathbf {v} _ {n}) \\prod_ {(i, j) \\in \\mathcal {E} _ {t}} \\phi (\\mathbf {v} _ {i}, \\mathbf {v} _ {j}), \\tag {7}", + "image_path": "b27a0b5bf3f8f71f536dd7f85d1ef04a54a24fbc32973e70ab23b30fcdebadd2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_t" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": " is the set of active nodes corresponding normal flow measurements " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "V_t^\\perp" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": " is the set of edges between active neighbor nodes. The measurements factor " + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": " is defined as TEG (refer to Sec. 3.3.1):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 586, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 586, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 108, + 586, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\psi (\\mathbf {v} _ {n}) = \\mathcal {N} ^ {- 1} \\left(; \\boldsymbol {\\eta} _ {n}, \\Lambda_ {n} ^ {\\mathrm {t e g}}\\right), \\tag {8}", + "image_path": "6d7c19c74eef491e734ae89769d9c222880c5f22ee0517be5b03f8b2ee4c2e74.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}_n = \\lambda_n^{\\mathrm{te}}\\mathbf{v}_n^\\perp" + }, + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "type": "text", + "content": ". In this study, we assume the flow is smooth; therefore, use the following prior factor " + }, + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 609, + 287, + 635 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 641, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 641, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\phi \\left(\\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {j} \\end{array} \\right]\\right) = \\mathcal {N} ^ {- 1} (\\cdot ; \\mathbf {0}, \\Lambda^ {\\mathrm {p}}), \\tag {9}", + "image_path": "0e80610bc4fd7186d1a30e151b7123c397501783ef190028299b5af92dee1e7c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Lambda^{\\mathrm{p}} = \\mathbf{J}_{\\mathrm{p}}^{\\top}\\mathrm{diag}(1 / \\sigma_{\\mathrm{p}}^{2},1 / \\sigma_{\\mathrm{p}}^{2})\\mathbf{J}_{\\mathrm{p}}" + }, + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "type": "text", + "content": " and the jacobian " + }, + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\mathrm{J_p} = \\left[ \\begin{array}{rrrr} - 1 & 0 & 1 & 0\\\\ 0 & -1 & 0 & 1 \\end{array} \\right] \\end{array}" + }, + { + "bbox": [ + 47, + 674, + 287, + 714 + ], + "type": "text", + "content": " for computing the difference between the neighbor nodes." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 314, + 82, + 538, + 173 + ], + "blocks": [ + { + "bbox": [ + 359, + 62, + 492, + 74 + ], + "lines": [ + { + "bbox": [ + 359, + 62, + 492, + 74 + ], + "spans": [ + { + "bbox": [ + 359, + 62, + 492, + 74 + ], + "type": "text", + "content": "Table 1. Basic experimental setup." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 82, + 538, + 173 + ], + "lines": [ + { + "bbox": [ + 314, + 82, + 538, + 173 + ], + "spans": [ + { + "bbox": [ + 314, + 82, + 538, + 173 + ], + "type": "table", + "html": "
ParameterESIMDVSMVSECDSEC
Window sizer [pix]5535
Δtpf [ms]40404040
Active durationτ [ms]5050100100
Prior stdσp [pix]0.110.51
Tangential stdσt [pix]10101010
Radial stdσr [pix]3333
Num. of layersL5555
Batch sizeNb10010010001000
", + "image_path": "fad96e619ebfcd82df818900c3c498686e218f7494909317d128e9bec2ce1665.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 193, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 240 + ], + "type": "text", + "content": "To model the outlier that can not be expressed as a Gaussian (quadratic cost), we employed the Huber cost both for measurement and prior factors. We adopted the technique of [1, 9] to maintain the Gaussian form." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 256, + 373, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 373, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 373, + 266 + ], + "type": "text", + "content": "3.3.4 TEGBP" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 276, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 348 + ], + "type": "text", + "content": "Equipped with the sparse probabilistic graph, we now formalize the incremental message-passing algorithm base on BP (Sec. 3.1.2) for the marginalization of the graph to estimate the full flow (Fig. 5). As the graph of active nodes is loopy, GBP needs to store a belief at each variable node for iterative belief update." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 364, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 422 + ], + "type": "text", + "content": "Message-passing is asynchronously triggered upon a space observation of normal flow. The measurement factor sends a message to the connected variable node and updates its belief with messages that have been sent from the active node as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 339, + 431, + 545, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 431, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 339, + 431, + 545, + 458 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\eta} _ {i} ^ {\\mathrm {b}} = \\sum_ {j \\in \\mathcal {C} _ {i} \\cap \\mathcal {A} _ {t}} \\boldsymbol {\\eta} _ {j \\rightarrow i}, \\quad \\Lambda_ {i} ^ {\\mathrm {b}} = \\sum_ {j \\in \\mathcal {C} _ {i} \\cap \\mathcal {A} _ {t}} \\Lambda_ {j \\rightarrow i} \\tag {10}", + "image_path": "9c57ccc4e35bec5dfb73abec39ec1712636a8d63404f03e000cceb949e4995cd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "b(\\mathbf{v}_i) = \\mathcal{N}^{-1}(\\cdot ;\\pmb{\\eta}_i^{\\mathrm{b}},\\Lambda_i^{\\mathrm{b}})" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " is the belief on node " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\mu_{j\\rightarrow i} = \\mathcal{N}^{-1}(\\cdot ;\\pmb{\\eta}_{j\\rightarrow i},\\Lambda_{j\\rightarrow i})" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " is the message from node " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " to node " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " is the 4-neighbors connected node " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " . After that, the variable node sends the following message about its updated beliefs to its neighbors:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 536, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 536, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 318, + 536, + 545, + 552 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\eta} _ {j \\rightarrow i} = - \\Lambda_ {0 1} ^ {\\mathrm {p}} \\left(\\Lambda_ {1 1} ^ {\\mathrm {p}} + \\Lambda_ {j} ^ {\\mathrm {b}} - \\Lambda_ {j \\rightarrow i}\\right) ^ {- 1} \\left(\\boldsymbol {\\eta} _ {j} ^ {\\mathrm {b}} - \\boldsymbol {\\eta} _ {j \\rightarrow i}\\right) \\tag {11}", + "image_path": "282248a74ae78c961255e3e08f6bfcfeca84d036423fa72b72c05d0c0539cef0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 554, + 545, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 554, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 319, + 554, + 545, + 569 + ], + "type": "interline_equation", + "content": "\\Lambda_ {j \\rightarrow i} = \\Lambda_ {0 0} ^ {\\mathrm {p}} - \\Lambda_ {0 1} ^ {\\mathrm {p}} \\left(\\Lambda_ {1 1} ^ {\\mathrm {p}} + \\Lambda_ {j} ^ {\\mathrm {b}} - \\Lambda_ {j \\rightarrow i}\\right) ^ {- 1} \\Lambda_ {1 0} ^ {\\mathrm {p}}, \\tag {12}", + "image_path": "b6edb6b446588bfabd97fc2fdd6fb16acaf34be6609f0903b8c49e81f9e37024.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Lambda_{\\alpha \\beta}^{\\mathrm{p}}" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": " sub-matrix of " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Lambda^{\\mathrm{p}}" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": " indexed by " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": ". For a single normal flow measurement, the belief is propagated for " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": " hops, which are repeated for " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{itr}}" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": " times (Fig. 5 correspond to " + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 577, + 545, + 624 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 673 + ], + "type": "text", + "content": "Note that our BP algorithm is guaranteed to converge to the correct MAP solution with sufficient iteration because all the belief is realized as a Gaussian (there is no guarantee for a variance because the graph is loopy) [30]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 713 + ], + "type": "text", + "content": "Coarse-to-fine message-passing scheme [10] is adopted to speed up the convergence. The measurement factor of" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21944" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 126, + 116 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 126, + 116 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 126, + 116 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 126, + 116 + ], + "type": "image", + "image_path": "305cf6bb5a5d82c5711e03a37b4b38af0926dad776d25cbd02e8b85e81269759.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 118, + 124, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 118, + 124, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 124, + 125 + ], + "type": "text", + "content": "Normal flow observation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 69, + 210, + 117 + ], + "blocks": [ + { + "bbox": [ + 129, + 69, + 210, + 117 + ], + "lines": [ + { + "bbox": [ + 129, + 69, + 210, + 117 + ], + "spans": [ + { + "bbox": [ + 129, + 69, + 210, + 117 + ], + "type": "image", + "image_path": "d1de7223f645a716656a89b157485fa2d512269cf4ab9c75d66073f30f10116e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 118, + 190, + 125 + ], + "lines": [ + { + "bbox": [ + 153, + 118, + 190, + 125 + ], + "spans": [ + { + "bbox": [ + 153, + 118, + 190, + 125 + ], + "type": "text", + "content": "Belief update" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 129, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 129, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 547, + 175 + ], + "type": "text", + "content": "Figure 5. A schematic showing a single iteration of TEGBP. (a) When a single normal flow has been measured, the message from the measurement factor is sent to the variable node as TEG, (b) the node updates its belief using the message received from the neighbor active nodes. (c) The nodes broadcast the updated belief to the neighbor, and (d) the neighbor nodes update their beliefs. (a)-(d) is repeated until convergence. To prevent the blowup of the message communication, it is stopped after " + }, + { + "bbox": [ + 46, + 129, + 547, + 175 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 129, + 547, + 175 + ], + "type": "text", + "content": " hops." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 69, + 304, + 117 + ], + "blocks": [ + { + "bbox": [ + 215, + 69, + 304, + 117 + ], + "lines": [ + { + "bbox": [ + 215, + 69, + 304, + 117 + ], + "spans": [ + { + "bbox": [ + 215, + 69, + 304, + 117 + ], + "type": "image", + "image_path": "637f17ff8d66b9e36c917880457565e727db61e55f1969e0a8908bf05555e63d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 118, + 279, + 126 + ], + "lines": [ + { + "bbox": [ + 231, + 118, + 279, + 126 + ], + "spans": [ + { + "bbox": [ + 231, + 118, + 279, + 126 + ], + "type": "text", + "content": "Message passing" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 69, + 376, + 117 + ], + "blocks": [ + { + "bbox": [ + 304, + 69, + 376, + 117 + ], + "lines": [ + { + "bbox": [ + 304, + 69, + 376, + 117 + ], + "spans": [ + { + "bbox": [ + 304, + 69, + 376, + 117 + ], + "type": "image", + "image_path": "002a832e4a33fd70fd80e8f78b51cc9e9c1d898e756ed105f3fa9b759d16893b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 118, + 357, + 126 + ], + "lines": [ + { + "bbox": [ + 318, + 118, + 357, + 126 + ], + "spans": [ + { + "bbox": [ + 318, + 118, + 357, + 126 + ], + "type": "text", + "content": "Belief update" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 69, + 460, + 117 + ], + "blocks": [ + { + "bbox": [ + 383, + 69, + 460, + 117 + ], + "lines": [ + { + "bbox": [ + 383, + 69, + 460, + 117 + ], + "spans": [ + { + "bbox": [ + 383, + 69, + 460, + 117 + ], + "type": "image", + "image_path": "210c0528a1b25404df44cdb1ff5e92cd90a9f5a6b78b6cfaada8cbbf139722e5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 118, + 446, + 126 + ], + "lines": [ + { + "bbox": [ + 399, + 118, + 446, + 126 + ], + "spans": [ + { + "bbox": [ + 399, + 118, + 446, + 126 + ], + "type": "text", + "content": "Message passing" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 69, + 543, + 117 + ], + "blocks": [ + { + "bbox": [ + 467, + 69, + 543, + 117 + ], + "lines": [ + { + "bbox": [ + 467, + 69, + 543, + 117 + ], + "spans": [ + { + "bbox": [ + 467, + 69, + 543, + 117 + ], + "type": "image", + "image_path": "3983fd458933343f1a93926c1d114fa71726d30c59251d3641984918a05f74bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 118, + 524, + 126 + ], + "lines": [ + { + "bbox": [ + 486, + 118, + 524, + 126 + ], + "spans": [ + { + "bbox": [ + 486, + 118, + 524, + 126 + ], + "type": "text", + "content": "Belief update" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 48, + 194, + 286, + 274 + ], + "blocks": [ + { + "bbox": [ + 48, + 194, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 48, + 194, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 286, + 274 + ], + "type": "image", + "image_path": "4b545dc779e1b4b5317b65256ce07c2ce15b80d0771f37b20cc7b492cf91ab2f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 319 + ], + "type": "text", + "content": "Figure 6. The result on ESIM-bricks. The estimated full flow from ARMS and TEGBP are compared with ground-truth flow indicated by the red arrow. They use the same normal flow measurement (left). TEGBP succeeds while ARMS fails." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 52, + 341, + 115, + 412 + ], + "blocks": [ + { + "bbox": [ + 52, + 341, + 115, + 412 + ], + "lines": [ + { + "bbox": [ + 52, + 341, + 115, + 412 + ], + "spans": [ + { + "bbox": [ + 52, + 341, + 115, + 412 + ], + "type": "image", + "image_path": "c9b998636b6f33ffd39d910d02cbe01cdcb78557880f9cd8269036d8b4c159d9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 420, + 102, + 429 + ], + "lines": [ + { + "bbox": [ + 59, + 420, + 102, + 429 + ], + "spans": [ + { + "bbox": [ + 59, + 420, + 102, + 429 + ], + "type": "text", + "content": "Time surface" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 127, + 341, + 206, + 419 + ], + "blocks": [ + { + "bbox": [ + 127, + 341, + 206, + 419 + ], + "lines": [ + { + "bbox": [ + 127, + 341, + 206, + 419 + ], + "spans": [ + { + "bbox": [ + 127, + 341, + 206, + 419 + ], + "type": "image", + "image_path": "7bb31ba4a087d3392dd028262d2b1cd903e5ab5c6ece193967f52655546472f3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 420, + 187, + 430 + ], + "lines": [ + { + "bbox": [ + 153, + 420, + 187, + 430 + ], + "spans": [ + { + "bbox": [ + 153, + 420, + 187, + 430 + ], + "type": "text", + "content": "ARMS [2]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 431, + 287, + 497 + ], + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 497 + ], + "type": "text", + "content": "Figure 7. The results on DVS-stripes. The histograms of the flow " + }, + { + "bbox": [ + 46, + 431, + 287, + 497 + ], + "type": "inline_equation", + "content": "(v_{x})" + }, + { + "bbox": [ + 46, + 431, + 287, + 497 + ], + "type": "text", + "content": " of ARMS and TEGBP are compared. The two red lines indicate the GT flow of the two stripes (the upper stripes have a smaller norm). The result from ARMS are shifted from the GT flow, especially on the area of upper stripes (smaller norm) On the contrary, our TEGBP successfully estimates individual flows." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 206, + 341, + 284, + 419 + ], + "blocks": [ + { + "bbox": [ + 206, + 341, + 284, + 419 + ], + "lines": [ + { + "bbox": [ + 206, + 341, + 284, + 419 + ], + "spans": [ + { + "bbox": [ + 206, + 341, + 284, + 419 + ], + "type": "image", + "image_path": "b93a9e7730c455e20220832ce8d093345b72ec472b7d1c72b9a2ad9049cb0b5f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 420, + 253, + 429 + ], + "lines": [ + { + "bbox": [ + 236, + 420, + 253, + 429 + ], + "spans": [ + { + "bbox": [ + 236, + 420, + 253, + 429 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "text", + "content": "the coarser layer " + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "text", + "content": " is computed as the sum of active measurement factor messages in a " + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "text", + "content": " block of finer layer " + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 46, + 527, + 287, + 624 + ], + "type": "text", + "content": ". The BP is executed coarse-to-fine; the coarse layer (message) estimation is copied to the finer layer as the initial guess at the start of the message passing on the finer layer (Supp-C.1). The short paths in the course graph improve efficiency for capturing long-range interactions and speed up convergence." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 634, + 123, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 123, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 123, + 647 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "We conducted experiments to compare the accuracy of TEGBP with the state-of-the-art incremental flow estimation algorithm called aperture-robust multi-scale flow estimation (ARMS) [2]. Both algorithms are incremental that could asynchronously estimate flow from sparse normal" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 205, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 230 + ], + "type": "text", + "content": "flow measurement. We used the same normal flow measurement for both algorithms for a fair comparison." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 238, + 427, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 427, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 427, + 251 + ], + "type": "text", + "content": "4.1. Experimental setting" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 258, + 452, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 452, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 452, + 268 + ], + "type": "text", + "content": "4.1.1 Normal flow measurement" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "type": "text", + "content": "We utilize the simple plane fitting to compute the normal flow measurements that serve as inputs to TEGBP and ARMS. We first apply noise removal for the raw events using a refractory filter [17] of " + }, + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "type": "inline_equation", + "content": "40~\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "type": "text", + "content": " duration. Then using the filtered events, we performed plane fitting for each event within a small spatiotemporal window of the size " + }, + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "type": "inline_equation", + "content": "r\\times r\\times \\Delta t_{\\mathrm{pf}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 397 + ], + "type": "text", + "content": ". During the plane fitting, we removed the outlier using the criteria proposed in [2]; we applied this process three times. We summarize the normal flow computation parameters in Tab. 1 (top)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 414, + 423, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 423, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 423, + 425 + ], + "type": "text", + "content": "4.1.2 Full flow estimation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "type": "text", + "content": "Both TEGBP and ARMS use the same sparse normal flow measurement, and could be run asynchronously at the rate of normal flow; however, we process " + }, + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 304, + 434, + 545, + 517 + ], + "type": "text", + "content": " observation at once for computational efficiency on MATLAB [18] (interactive numerical computing environment) we used to compare the accuracy. The specific parameters for the (quasi) full flow computation are summarized in Tab. 1 (bottom)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "text", + "content": "In TEGBP, the message should be propagated to all connected nodes, and the process needs to be repeated until convergence; however, we found a small number of hops " + }, + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "inline_equation", + "content": "(K = 2)" + }, + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "text", + "content": " and single iterations " + }, + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "inline_equation", + "content": "(N_{\\mathrm{itr}} = 1)" + }, + { + "bbox": [ + 304, + 517, + 545, + 566 + ], + "type": "text", + "content": " works well in practice." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 574, + 389, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 389, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 389, + 586 + ], + "type": "text", + "content": "4.2. ESIM-bricks" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In this experiment, we compare the accuracy on a scene where the motion and edges are not perpendicular to the image edge. We synthesized the event stream by observing the image of bricks from a translating camera. We used ESIM [23] library for this simulation, which could synthesize the realistic noisy event stream from the event camera moving around the predefined 3D environment. We compute the normal flow and use the same measurements for both methods. Fig. 6 shows the result. We observe the drift in ARMS due to the average operation. TEGBP correctly" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21945" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 91, + 547, + 387 + ], + "blocks": [ + { + "bbox": [ + 116, + 71, + 475, + 82 + ], + "lines": [ + { + "bbox": [ + 116, + 71, + 475, + 82 + ], + "spans": [ + { + "bbox": [ + 116, + 71, + 475, + 82 + ], + "type": "text", + "content": "Table 2. The quantitative results on MVSEC. The flow is evaluated on the ground truth intervals." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 91, + 547, + 387 + ], + "lines": [ + { + "bbox": [ + 50, + 91, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 50, + 91, + 547, + 387 + ], + "type": "table", + "html": "
indoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm2.3024.53.6142.93.1336.33.4443.0
ARMS [2]1.7112.72.6726.62.2821.62.6425.7
Ours1.146.251.8716.41.5411.81.4611.1
Indoor_flying1
Indoor_flying3
Outdoor_day1
Time surfaceNormal flowARMS [2]OursGround truth
", + "image_path": "53dd1ab3e790e54c6cb8e7773962061d70dc4c95b32a1120981d6aa66a35b36c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 393, + 493, + 405 + ], + "lines": [ + { + "bbox": [ + 97, + 393, + 493, + 405 + ], + "spans": [ + { + "bbox": [ + 97, + 393, + 493, + 405 + ], + "type": "text", + "content": "Figure 8. The qualitative results on MVSEC. See the color wheel on the bottom left for the color encoding." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 436, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 460 + ], + "type": "text", + "content": "estimated the full flow from the normal flow measurement directed at a different angle from the full flow." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 467, + 126, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 126, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 126, + 479 + ], + "type": "text", + "content": "4.3.DVS-stirpes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "text", + "content": "In this experiment, we compare the accuracy on a scene where some region includes different flows. We used a scene from [19] where two stripes translating at different speeds are observed from a dynamic vision sensor (DVS). Fig. 7 shows the result. We observed a significant shift in the ARMS's estimation when motions of different speeds are adjacent. ARMS determines the flow by selecting the scale having the maximum average norm in the window (it falsely picks the large norm in the slow region). In contrast, the TEGBP correctly estimates individual motion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 611, + 110, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 110, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 110, + 623 + ], + "type": "text", + "content": "4.4. MVSEC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "We quantitatively compare with ARMS in practical robotic scenarios. We used outdoor scenes captured by automobiles and indoor scenes captured by aerial drones from the Multi-vehicle Stereo Event Camera (MVSEC) dataset [33]. This dataset includes ground truth (GT) optical flow computed as the motion filed from the camera motion and scene depth. We evaluate the accuracy using the average" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 436, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 556 + ], + "type": "text", + "content": "endpoint error (AEE) and the percentage of pixels with endpoint error greater than 3 pixels (denoted by % Out). Both are measured on pixels where valid GT exists, and the plane fitting has been successful. We output flows asynchronously in batches with both methods and evaluate them at regular intervals (20Hz of GT rate). Tab. 2 and Fig. 8 show the quantitative and qualitative results. ARMS improves the normal flow, while TEGBP improves more. In contrast to the naive averaging in ARMS, our TEGBP properly models the distribution of flows using TEG and prior." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 563, + 359, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 359, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 359, + 574 + ], + "type": "text", + "content": "4.5. DSEC" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": "We compare the performance of TEGBP and ARMS on more diverse scenarios using more high-resolution input. To this end, we use the DSEC [13], a recently released dataset in automobile scenarios containing high-resolution event data of " + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 304, + 581, + 545, + 689 + ], + "type": "text", + "content": " captured from the vehicle that undergoes more diverse motion. It is designed for dense optical flow estimation, and the GT of the test sequence is hidden for leaderboard evaluation. Therefore we use their train sequence to compare sparse AEE error with ARMS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Tab. 3 and Fig. 9 shows the quantitative and qualitative results. Similar to the MVSEC experiments, TEGBP sig" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21946" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 91, + 523, + 142 + ], + "blocks": [ + { + "bbox": [ + 121, + 71, + 472, + 82 + ], + "lines": [ + { + "bbox": [ + 121, + 71, + 472, + 82 + ], + "spans": [ + { + "bbox": [ + 121, + 71, + 472, + 82 + ], + "type": "text", + "content": "Table 3. The quantitative results on DSEC. The flow is evaluated on the ground truth intervals." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 91, + 523, + 142 + ], + "lines": [ + { + "bbox": [ + 70, + 91, + 523, + 142 + ], + "spans": [ + { + "bbox": [ + 70, + 91, + 523, + 142 + ], + "type": "table", + "html": "
thun_00_azurich_city_01_azurich_city_02_azurich_city_08_azurich_city_11_a
AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓AEE ↓% Out ↓
Norm7.0968.78.2637.912.886.88.4273.35.9063.4
ARMS [2]5.3160.06.3223.88.9878.76.2264.44.5751.4
Ours4.2645.55.6123.28.7072.65.0849.53.2732.2
", + "image_path": "ef2c3702911d6e8ccd16cd9932c1e8143b60aa4c208737dd4c248d71bb73630f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 152, + 547, + 310 + ], + "blocks": [ + { + "bbox": [ + 48, + 152, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 48, + 152, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 547, + 310 + ], + "type": "image", + "image_path": "b7f6bdf484a396f0b3e25c9d2517d13fc279f7c677748e16859f8324cd6b5112.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 317, + 378, + 328 + ], + "lines": [ + { + "bbox": [ + 214, + 317, + 378, + 328 + ], + "spans": [ + { + "bbox": [ + 214, + 317, + 378, + 328 + ], + "type": "text", + "content": "Figure 9. The qualitative results on DSEC." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 206, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 206, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 206, + 371 + ], + "type": "text", + "content": "nificantly boosts accuracy over AMRS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 385, + 119, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 119, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 119, + 398 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 407, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 526 + ], + "type": "text", + "content": "We proposed TEGBP, a novel incremental full flow estimation algorithm for asynchronous event data. The proposed algorithm is backed up by the probabilistic model, which guarantees equality to the full flow when a valid prior is provided. It realizes efficient event-wise incremental updates based on the BP, where the MAP solution obtained by the local update is guaranteed to match the correct one. We demonstrate the effectiveness of the proposed method on a practical real-world dataset showing a significant boost in accuracy over the existing incremental algorithm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 624 + ], + "type": "text", + "content": "There is still much room for improvement in accuracy compared to the recent machine learning-based method (although it is computationally intensive). We suspect the gap mainly comes from the noise in the normal flow and poor prior. Toward the ultimate goal of realizing low-latency and high-accuracy full flow estimation in real-time on an edge device, we expect the gap to be closed by incorporating a more robust normal flow and application-specific prior." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 129, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 129, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 129, + 645 + ], + "type": "text", + "content": "5.1. Future work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Robust normal flow estimation As discussed in the experiment section, some scenes include strongly erroneous flow, e.g., normal flow from leaves on the tree. These measurement errors significantly degrade the accuracy. We expect the tiny CNN (e.g., input size of " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": ") trained to regress" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 360, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 384 + ], + "type": "text", + "content": "the normal flow from noisy input will improve the accuracy without sacrificing the computational efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 398, + 546, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 546, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 546, + 541 + ], + "type": "text", + "content": "Application specific prior Even if the perfect normal flow is obtained, we cannot expect good accuracy without good prior (Sec. 3.3.2). The uniform smoothness prior we adopted may not optimal in many practical scenarios. The discontinuous depth region does not conform to the prior. Even if the planer object undergoes linear motion, we won't observe uniform flow when it is not perpendicular to the optical axis. We can utilize a better prior for a specific application to boost accuracy. For example, one may utilize a vanishing point for automotive applications (assuming the yaw rate is known from the vehicle sensor) similar to [20]. We left these explorations for future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 556, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 712 + ], + "type": "text", + "content": "Optimized implementation Our algorithm is very efficient in principle, requiring lower FLOPS, and the entire process can be asynchronously parallelized. The normal flow could be cheaply obtained, e.g., directly from a camera or computed by the lightweight plane fitting (computation of normal flow is outside this study's scope). The core of our algorithm, message passing of (4), is also very cheap, requiring the communication of 6-dim vectors with neighbors. Therefore, an optimized CPU implementation or adaptation of processors supporting the sparse computation could realize low-energy, low-latency optical flow estimation in the edge device. We left the wall clock time and energy consumption evaluation as future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21947" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Pratik Agarwal, Gian Diego Tipaldi, Luciano Spinello, Cyril Stachniss, and Wolfram Burgard. Robust Map Optimization using Dynamic Covariance Scaling. IEEE International Conference on Robotics and Automation (ICRA), 2013. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):1-12, 2020. 1, 2, 4, 6, 7, 8, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. IEEE International Symposium on Circuits and Systems, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 303 + ], + "type": "text", + "content": "[4] R. Wes Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event Probability Mask (EPM) and Event Denoising Convolutional Neural Network (EDnCNN) for Neuromorphic Cameras. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1701-1710, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 350 + ], + "type": "text", + "content": "[5] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous Optical Flow and Intensity Estimation from an Event Camera. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 395 + ], + "type": "text", + "content": "[6] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, feb 2014. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 440 + ], + "type": "text", + "content": "[7] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, mar 2012. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 464 + ], + "type": "text", + "content": "[8] Andrew J. Davison. FutureMapping: The Computational Structure of Spatial AI Systems. arXiv, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 498 + ], + "type": "text", + "content": "[9] Andrew J. Davison and Joseph Ortiz. FutureMapping 2: Gaussian Belief Propagation for Spatial AI. arXiv, 2019. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 543 + ], + "type": "text", + "content": "[10] Pedro F Felzenszwalb and Daniel P Huttenlocher. Ef ci e n t Belief Propagation for Early Vision. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 5, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 611 + ], + "type": "text", + "content": "[11] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Joerg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1), 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[12] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[13] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A Stereo Event Camera Dataset for Driving Scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. International Conference on 3D Vision (3DV), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[15] B. K. B. Horn and B. G. Schunck. Determining Optical Flow. Artificial Intelligence, 17(1-3):185-203, 1981. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 204 + ], + "type": "text", + "content": "[16] Mark Kurtz, Justin Kopinsky, Rati Gelashvili, Alexander Matveev, John Carr, Michael Goin, William Leiserson, Sage Moore, Bill Nell, Nir Shavit, and Dan Alistarh. Inducing and Exploiting Activation Sparsity for Fast Neural Network Inference. Proceedings of Machine Learning Research, 119:5533-5543, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[17] Weng Fei Low, Zhi Gao, Cheng Xiang, and Bharath Ramesh. SOFEA: A non-iterative and robust optical flow estimation algorithm for dynamic vision sensors. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 281 + ], + "type": "text", + "content": "[18] MATLAB. version 9.13.0 (R2022b). The MathWorks Inc., Natick, Massachusetts, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "text", + "content": "[19] Elias Mueggler, Christian Forster, Nathan Baumli, Guillermo Gallego, and Davide Scaramuzza. Lifetime Estimation of Events from Dynamic Vision Sensors. IEEE International Conference on Robotics and Automation (ICRA), 2015. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[20] Jun Nagata, Yusuke Sekikawa, Kosuke Hara, and Yoshimitsu Aoki. FOE-based Regularization for Optical Flow Estimation from an In-vehicle Event Camera. Proceedings of SPIE - The International Society for Optical Engineering, 11049, 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "text", + "content": "[21] Joseph Ortiz, Mark Pupilli, Stefan Leutenegger, and Andrew J. Davison. Bundle Adjustment on a Graph Processor. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2413-2422, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 480 + ], + "type": "text", + "content": "[22] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single Image Optical Flow Estimation with an Event Camera. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[23] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an Open Event Camera Simulator. Conf. on Robotics Learning (CoRL), 87:969-982, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "text", + "content": "[24] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. MDPI Sensors, 22(14):1-20, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[25] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. European Conference on Computer Vision (ECCV), pages 1-23, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[26] Chen Shoushun and Guo Menghan. Live Demonstration : CeleX-V: a 1M Pixel Multi-Mode Event-based Sensor Chen Shoushun. IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1–2, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[27] Timo Stoffregen and Lindsay Kleeman. Event Cameras, Contrast Maximization and Reward Functions: an Analysis. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12300-12308, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Daniel C. Stumpp, Himanshu Akolkar, Alan D. George, and Ryad B. Benosman. hARMS: A Hardware Acceleration Architecture for Real-Time Event-Based Optical Flow. IEEE Access, 10:58181-58198, 2022. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21948" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[29] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. European Conference on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[30] Yair Weiss and William T Freeman. Correctness of belief propagation in Gaussian graphical models of arbitrary topology. Neural Information Processing Systems (NIPS), pages 673-679, 2000. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[31] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion from Sparse Event Data. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[32] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. Proceedings of Robotics: Science and Systems, jun 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "text", + "content": "[33] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multi Vehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 7, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[34] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based optical flow using motion compensation. European Conference on Computer Vision (ECCV) Workshop, pages 711-714, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 407 + ], + "type": "text", + "content": "[35] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-based Learning of Optical Flow, Depth, and Egomotion. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 989–997, 2019. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21949" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_content_list.json b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5257ec933a4a1da9b60a42d702658748ff45fb5a --- /dev/null +++ b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_content_list.json @@ -0,0 +1,1341 @@ +[ + { + "type": "text", + "text": "TarViS: A Unified Approach for Target-based Video Segmentation", + "text_level": 1, + "bbox": [ + 150, + 118, + 818, + 140 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ali Athar $^{1}$ Alexander Hermans $^{1}$ Jonathon Luiten $^{1,2}$ Deva Ramanan $^{2}$ Bastian Leibe $^{1}$ $^{1}$ RWTH Aachen University, Germany $^{2}$ Carnegie Mellon University, USA \n{athar, hermans, luiten, leibe}@vision.rwth-aachen.de deva@cs.cmu.edu", + "bbox": [ + 120, + 167, + 846, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The general domain of video segmentation is currently fragmented into different tasks spanning multiple benchmarks. Despite rapid progress in the state-of-the-art, current methods are overwhelmingly task-specific and cannot conceptually generalize to other tasks. Inspired by recent approaches with multi-task capability, we propose TarViS: a novel, unified network architecture that can be applied to any task that requires segmenting a set of arbitrarily defined 'targets' in video. Our approach is flexible with respect to how tasks define these targets, since it models the latter as abstract 'queries' which are then used to predict pixel-precise target masks. A single TarViS model can be trained jointly on a collection of datasets spanning different tasks, and can hot-swap between tasks during inference without any task-specific retraining. To demonstrate its effectiveness, we apply TarViS to four different tasks, namely Video Instance Segmentation (VIS), Video Panoptic Segmentation (VPS), Video Object Segmentation (VOS) and Point Exemplar-guided Tracking (PET). Our unified, jointly trained model achieves state-of-the-art performance on 5/7 benchmarks spanning these four tasks, and competitive performance on the remaining two. Code and model weights are available at: https://github.com/Ali2500/TarViS", + "bbox": [ + 75, + 301, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 678, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to understand video scenes has been a longstanding goal of computer vision research because of wideranging applications in intelligent vehicles and robots. Early approaches tackled simpler tasks involving contour-based [33, 39] and box-level tracking [21, 25, 40, 52], background subtraction [20, 61], and motion segmentation [8, 49]. The deep learning boom then revolutionized the landscape by enabling methods to perform pixel-precise segmentation on challenging, real-world videos. In the past few years, a number of benchmarks have emerged, which evaluate how well methods can perform video segmentation according to various task formulations. Over time, these tasks/benchmarks have ballooned into separate re", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acfefb23785a7de608b92173bfe6d504186c65a7e93cee1b2da1e70280c49b67.jpg", + "image_caption": [ + "Figure 1. Predicted results from a jointly trained TarViS model for four different video segmentation tasks." + ], + "image_footnote": [], + "bbox": [ + 498, + 265, + 890, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "search sub-communities. Although existing methods are rapidly improving the state-of-the-art for these benchmarks, each of them typically tackles only one narrowly-defined task, and generalizing them is non-trivial since the task definition is baked into the core approach.", + "bbox": [ + 496, + 612, + 893, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that this fragmentation is unnecessary because video target segmentation tasks all require the same high-level capability, namely that of identifying, localizing and tracking rich semantic concepts. Meanwhile, recent progress on Transformer networks has enabled the wider AI research community to move towards unified, multi-task architectures [1, 30, 31, 38, 58], because the attention operation [62] is well-suited for processing feature sets with arbitrary structure and data modality. These developments give us the opportunity to unify the fractured landscape of target-based video segmentation. In this paper, we propose TarViS: a novel architecture which enables a single, unified model to be jointly trained for multiple video segmentation tasks. During inference, the same model can perform differ", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18738", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ent tasks at runtime by specifying the segmentation target.", + "bbox": [ + 76, + 90, + 464, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core idea is that TarViS tackles the generic task of segmenting a set of arbitrary targets in video (defined as semantic classes or as specific objects). These targets are encoded as queries which, together with the video features, are input to a Transformer-based model. The model iteratively refines these queries and produces a pixel-precise mask for each target entity. This formulation conceptually fuses all video segmentation tasks [3, 54, 66, 72] which fall under the umbrella of the above-mentioned generic task, because they differ only in how the targets are defined. During both training and inference, TarViS can hot-swap between tasks at run-time by providing the desired target query set.", + "bbox": [ + 76, + 107, + 468, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate our generalization capability, we tackle four different tasks: (1) Video Instance Segmentation (VIS) [54, 72], (2) Video Panoptic Segmentation (VPS) [35], (3) Video Object Segmentation [53], and (4) Point Exemplar-guided Tracking [3] (PET). For VIS, the segmentation targets are all objects in the video belonging to a predefined set of classes. The target set for VPS includes that for VIS, and additionally, a set of non-instantiable stuff semantic classes. For VOS, the targets are a specific set of objects for which the first-frame ground-truth mask is provided. PET is a more constrained version of VOS which only provides the location of a single point inside the object, rather than the full object mask.", + "bbox": [ + 76, + 289, + 470, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods for these tasks lack generalization capability because task-specific assumptions are typically baked into the approach (see Sec. 2 and 3 for details). In contrast, TarViS can tackle all four tasks with a unified model because we encode the task-specific targets as a set of queries, thus decoupling the network architecture from the task definition. Moreover, our approach can theoretically generalize further, e.g., one could potentially define the target set as all objects described by a given text prompt, though this is beyond the scope of this paper.", + "bbox": [ + 76, + 487, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows: we propose TarViS, a novel architecture that can perform any task requiring segmentation of a set of targets from video. For the first time, we are able to jointly train and infer a single model on a collection of datasets spanning the four aforementioned tasks (VIS, VPS, VOS, PET). Our experimental results show that TarViS performs competitively for VOS, and achieves state-of-the-art results for VIS, VPS and PET.", + "bbox": [ + 76, + 640, + 468, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 777, + 218, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-task Models. Multi-task learning has a long history [11] with several architectures and training strategies [24,36,37,51,59,76]. Earlier approaches mostly consist of a shared backbone with fixed task-specific heads, whereas we design a more general architecture for video segmentation with task-specific targets to specify what to", + "bbox": [ + 76, + 809, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segment. Our approach is inspired by recent attention-based models, e.g., PerceiverIO [30, 31], which can be trained on diverse data modalities and task-specific heads are replaced with output queries. UViM [38] follows a similar direction by creating a unified architecture for diverse dense prediction tasks. However, both of these models are trained separately for different tasks. Recent, powerful multi-task vision language models such as Flamingo [1] and GATO [58] tackle a multitude of tasks by requiring a sequence of task-specific input-output examples to prime the model. This is conceptually similar to our task-specific targets, however, our model does not require per-task priming. Moreover, our targets are not modeled as sequence prompts, and we aim for a video segmentation model which is several orders of magnitude smaller. In the realm of video tracking and segmentation, the recently proposed UNICORN [71] model tackles multiple object tracking-related tasks with a unified architecture. Unlike TarViS, however, UNICORN follows the task-specific output head approach and is generally oriented towards box-level tracking tasks [22, 45, 47, 75], thus requiring non-trivial modifications to tackle VPS or PET.", + "bbox": [ + 496, + 90, + 893, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query-based Transformer Architectures. Several works [2, 10, 13, 30, 31, 46, 65, 78] use query-based Transformer architectures for various tasks. The workhorse for task learning here is the iterative application of self- and cross-attention, where a set of query vectors (e.g., representing objects) are refined by interacting with each other, and with the input data (e.g., an image). Unlike existing methods which use queries in a task-specific context, TarViS adopts a query-based Transformer architecture in which the queries serve as a mechanism for decoupling the task definition from the architecture, i.e., our model can learn to tackle different tasks while being agnostic to their definition because the latter is abstracted behind a set of queries.", + "bbox": [ + 496, + 417, + 893, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task-specific Video Segmentation. Current Video Instance Segmentation (VIS) methods broadly work by predicting object tracks in the video, followed by classification into a pre-defined set of categories. Several approaches [6,9,23,28,34,42,54,64,69,72] are based on the tracking-by-detection paradigm, some model video as a joint spatiotemporal volume [4,5], whereas recent works [12,26,29,65, 68] adopt Transformer-based architectures.", + "bbox": [ + 496, + 623, + 893, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For Video Panoptic Segmentation (VPS), methods [35, 55, 66] generally extend image-level panoptic approaches [14] by employing multi-head architectures for semantic segmentation and instance mask regression, classification, and temporal association. In the Video Object Segmentation (VOS) community, state-of-the-art methods are broadly based on the seminal work of Oh et al. [50], which learns space-time correspondences between pixels in different video frames, and then uses these to propagate the first-frame masks across the video. Subsequent", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18739", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a444117723d665becfbbab1396b0a327d3e8a6f8f53b46db883986d98e33859.jpg", + "image_caption": [ + "Figure 2. TarViS Architecture. Segmentation targets for different tasks are represented by a set of abstract target queries $Q_{\\mathrm{in}}$ . The core network (in green) is agnostic to the task definitions. The inner product between the output queries $Q_{\\mathrm{out}}$ and video feature $F_{4}$ yields segmentation masks as required by the task." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 666, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c4f063716734fedb84b6c65cf46ab1dca5aea4ecb26d9a27bf07ea842d90657b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 88, + 890, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods [15-17, 60, 63, 70, 73, 73, 74] have significantly improved the performance and efficiency of this approach. Point Exemplar-guided Tracking (PET) [3, 27] is a fairly new task for which the current best approach [3] regresses a pseudo-ground-truth mask from the given point coordinates, and then applies a VOS method [17] to this mask.", + "bbox": [ + 75, + 375, + 467, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above methods thus incorporate task-specific assumptions into their core approach. This can benefit pertask performance, but makes it difficult for them to generalize across tasks. By contrast, TarViS can tackle all four aforementioned tasks, and generally any target-based video segmentation task, with a single, jointly trained model.", + "bbox": [ + 75, + 465, + 468, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 571, + 168, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TarViS can segment arbitrary targets in video since the architecture is flexible with respect to how these targets are defined, thus enabling us to conceptually unify and jointly tackle the four aforementioned tasks (VIS, VPS, VOS, PET). The architecture is illustrated in Fig. 2.", + "bbox": [ + 75, + 598, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For all tasks, the common input to the network is an RGB video clip of length $T$ denoted by $V \\in \\mathbb{R}^{H \\times W \\times T \\times 3}$ . This is input to a 2D backbone network which produces image-level feature maps, followed by a Temporal Neck, which enables feature interaction across time and outputs a set of temporally consistent, multi-scale, $D$ -dimensional feature maps $\\mathcal{F} = \\{F_{32}, F_{16}, F_{8}, F_{4}\\}$ where $F_{s} \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times T \\times D}$ . The feature maps $\\mathcal{F}$ are then fed to our Transformer decoder, together with a set of queries $Q_{\\mathrm{in}}$ which represent the segmentation targets. The decoder applies successive layers of self- and masked cross-attention wherein the queries are iteratively refined by attending to each other, and to the feature maps, respectively. The refined queries output by the decoder are denoted with $Q_{\\mathrm{out}}$ . The following subsections explain how TarViS tackles each task in detail.", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 500, + 375, + 759, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VIS defines the segmentation target set as all objects belonging to a set of predefined classes. Accordingly, the input query set $Q_{\\mathrm{in}}$ for VIS contains three types of queries: (1) semantic queries denoted by $Q_{\\mathrm{sem}} \\in \\mathbb{R}^{C \\times D}$ where $C$ is the number of classes defined by the dataset, i.e., each $D$ -dimensional vector in $Q_{\\mathrm{sem}}$ represents a particular semantic class. (2) instance queries denoted by $Q_{\\mathrm{inst}} \\in \\mathbb{R}^{I \\times D}$ where $I$ is assumed to be an upper bound on the number of instances in the video clip, and (3) a background query denoted by $Q_{\\mathrm{bg}} \\in \\mathbb{R}^{1 \\times D}$ to capture inactive instance queries.", + "bbox": [ + 496, + 401, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The three query sets are concatenated, i.e., $Q_{\\mathrm{in}} = \\operatorname{concat}(Q_{\\mathrm{sem}}, Q_{\\mathrm{inst}}, Q_{\\mathrm{bg}})$ , and input to the Transformer decoder, which refines their feature representation through successive attention layers and outputs a set of queries $Q_{\\mathrm{out}} = \\operatorname{concat}(Q_{\\mathrm{sem}}', Q_{\\mathrm{inst}}', Q_{\\mathrm{bg}}')$ . These are then used to produce temporally consistent instance mask logits by computing the inner product $\\langle F_4, Q_{\\mathrm{inst}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times I}$ . To obtain classification logits, we compute the inner product $\\langle Q_{\\mathrm{inst}}', \\operatorname{concat}(Q_{\\mathrm{sem}}', Q_{\\mathrm{bg}}') \\rangle \\in \\mathbb{R}^{I \\times (C + 1)}$ .", + "bbox": [ + 496, + 556, + 892, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The three types of queries are initialized randomly at the start of training and optimized thereafter. The instance queries $Q_{\\mathrm{inst}}$ enable us to segment a varying number of objects from the input clip. During training, we apply Hungarian matching between the predicted and ground-truth instance masks to assign instance queries to video instances, and then supervise their predicted masks and classification logits accordingly. When training on multiple datasets with heterogeneous classes, the semantic query sets are separately initialized per dataset, but $Q_{\\mathrm{inst}}$ and $Q_{\\mathrm{bg}}$ are shared.", + "bbox": [ + 496, + 696, + 890, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comparison to Instance Segmentation Methods. Several Transformer-based methods [10, 13, 65, 78] for image/video instance segmentation also use queries to seg", + "bbox": [ + 496, + 854, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18740", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment a variable number of input instances. The key difference to our approach is the handling of object classes: existing works employ only instance queries which are input to a fully-connected layer with a fan-out of $C + 1$ to obtain classification (and background) scores. The notion of class-guided instance segmentation is thus baked into the approach. By contrast, TarViS is agnostic to the task-specific notion of object classes because it models them as arbitrary queries which are dynamic inputs to the network. The semantic representation for these queries is thus decoupled from the core architecture and is only learned via loss supervision. An important enabler for this approach is the background query $Q_{\\mathrm{bg}}$ , which serves as a 'catch-all' class to represent everything that is not in $Q_{\\mathrm{sem}}$ . It is used to classify non-active instance queries, and its mask logits are supervised to segment all non-object input pixels.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 340, + 341, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VPS defines the segmentation targets as all objects belonging to a set of thing classes (e.g., 'person', 'car'), and additionally, a set of non-instantiable stuff classes (e.g., 'sky', 'grass') which cover all non-object pixels. TarViS can tackle VPS with virtually no modification to the workflow in Sec. 3.1. We can compute semantic segmentation masks for the input clip by simply taking the inner product between $Q_{\\mathrm{sem}}$ and the video features: $\\langle F_4, Q_{\\mathrm{sem}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times C}$ . Note that here, $Q_{\\mathrm{sem}}$ contains queries representing both thing and stuff classes.", + "bbox": [ + 75, + 364, + 472, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison to VPS Methods. Current VPS datasets [35, 66] involve driving scene videos captured from moving vehicles. Methods tackling this task [35, 55] are based on earlier image panoptic segmentation approaches [14] which involve multi-head networks for semantic and instance segmentation prediction. In terms of image-level panoptic segmentation, Mask2Former [13] uses a Transformer-based architecture, but it models stuff classes as instances which are Hungarian-matched to the ground-truth target during training, whereas TarViS models semantic classes and instances using separate, designated queries.", + "bbox": [ + 75, + 520, + 472, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Video Object Segmentation and Point Exemplar-guided Tracking", + "text_level": 1, + "bbox": [ + 76, + 696, + 470, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VOS and PET are instantiations of a general task where the segmentation targets are a set of $O$ objects for which some ground-truth cue $\\mathcal{G}$ is given. For VOS, $\\mathcal{G}$ is provided as the first-frame object masks $M_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times H \\times W}$ , whereas for PET, $\\mathcal{G}$ is provided as the $(x,y)$ coordinates $P_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times 2}$ of a point inside each of the objects. TarViS jointly tackles both tasks by adopting a generalized approach in which the $O$ target objects are encoded into a set of object queries $Q_{\\mathrm{obj}}$ . Thus, both VOS and PET boil down to designing a function Encode0objects(\\cdot) which regresses $Q_{\\mathrm{obj}}$ from the ground-truth cues $\\mathcal{G}$ and feature maps $\\mathcal{F}$ :", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\mathrm {o b j}} \\leftarrow \\operatorname {E n c o d e 0 b j e c t s} (\\mathcal {G}, \\mathcal {F}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 90, + 890, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $Q_{\\mathrm{obj}}$ is conceptually analogous to $Q_{\\mathrm{sem}}$ and $Q_{\\mathrm{inst}}$ used for VIS in that all three are abstract representations for their respective task-specific segmentation targets.", + "bbox": [ + 498, + 131, + 890, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Object Segmentation. We seek inspiration from HODOR [2] to implement EncodeObjects for VOS, a recent method for weakly-supervised VOS, which encodes objects into concise descriptors as follows: the descriptors are initialized by average pooling the image features inside the object masks, followed by iterative refinement where the descriptors attend to each other (self-attention) and to their respective soft-masked image features (cross-attention).", + "bbox": [ + 496, + 181, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For TarViS, we employ a lightweight Object Encoder with a similar workflow to encode the objects as a set of queries $Q_{\\mathrm{obj}}$ , but with two differences to HODOR [2]: instead of cross-attending to the entire image feature map $(H \\cdot W$ points) with soft-masked attention, we apply hard-masked cross-attention to at most $p_{\\mathrm{max}}$ feature points per object, where $p_{\\mathrm{max}} \\ll H \\cdot W$ . Object masks containing more than $p_{\\mathrm{max}}$ points are sub-sampled accordingly. This significantly improves the memory/run-time overhead of our Object Encoder. Secondly, we note that the process of distilling object features into a single descriptor involves a loss of object appearance information, which degrades performance. We therefore model each object with $q_{o}$ queries (instead of one) by spatially dividing each object mask into $q_{o}$ segments, i.e., $Q_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times q_{o} \\times D}$ (we use $q_{o} = 4$ ).", + "bbox": [ + 496, + 303, + 892, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to $Q_{\\mathrm{obj}}$ , we initialize a set of background queries $Q_{\\mathrm{bg}} \\in \\mathbb{R}^{B \\times D}$ to model the non-target pixels in the reference frame. Following HODOR [2], we employ multiple background queries, which are initialized dynamically by dividing the video frame containing the groundtruth masks $M_{\\mathrm{obj}}$ into a $4 \\times 4$ grid and average pooling the non-object pixels in each grid cell. The Object Encoder jointly refines the background and object queries to yield $Q_{\\mathrm{in}} = \\text{concat}(Q_{\\mathrm{obj}}, Q_{\\mathrm{bg}})$ . During training, the mask logits for the multiple background queries are aggregated per-pixel by applying $\\max(\\cdot)$ and supervised to segment all pixels not part of the target object set.", + "bbox": [ + 496, + 530, + 893, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The remaining workflow follows that for VIS and VPS: $Q_{\\mathrm{in}}$ is input to the Transformer decoder together with the video features $\\mathcal{F}$ . The refined output query set $Q_{\\mathrm{out}} = \\operatorname{concat}(Q_{\\mathrm{obj}}', Q_{\\mathrm{bg}}')$ is then used to compute the inner product $\\langle F_4, Q_{\\mathrm{obj}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times O \\times q_o}$ . Subsequently, $\\max(\\cdot)$ is applied on the $q_o$ -sized dimension to obtain the final mask logits for the $O$ target objects.", + "bbox": [ + 496, + 710, + 893, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Exemplar-guided Tracking. For PET we implement Encode0objects in the exactly same way as VOS: the given point coordinates $P_{\\mathrm{obj}}$ are converted into a mask with just one non-zero pixel, followed by iterative refinement by the Object Encoder (with shared weights for VOS and PET).", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18741", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The only difference is that here we represent each of the $O$ objects with just one query, i.e., $Q_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times D}$ ( $q_o = 1$ ). The subsequent workflow is also identical to that for VOS: the queries are refined by the Transformer decoder followed by an inner product with $F_4$ to obtain object mask logits.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison to VOS and PET Methods. Current state-of-the-art VOS methods are largely based on STM [50]. It involves learning pixel-to-pixel correspondences across video frames, which are then used to propagate the given object mask across the video. This approach is effective since every pixel in the given mask can be individually mapped to future frames, thus preserving fine-grained object details. The core approach is, however, task-specific since it assumes the availability of first-frame object masks, and does not generalize to the PET (see Sec. 4.2). PET can be viewed as a more constrained version of VOS, where only a single object point is provided instead of the full mask. Consequently, PET [3] is currently tackled by casting it as a VOS problem by using an image instance segmentation network [13] to regress pseudo-ground-truth object masks from the given point coordinates $P_{\\mathrm{obj}}$ .", + "bbox": [ + 75, + 172, + 468, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, our approach of encoding objects as concise queries causes loss of fine-grained object appearance information, but it has the advantage of being agnostic to how $\\mathcal{G}$ is defined. As evident from the unified workflow for VOS and PET, any variation of these tasks with arbitrary ground-truth cues $\\mathcal{G}$ can be seamlessly fused into our architecture as long as we can implement an effective Encode0bjects function to regress $Q_{\\mathrm{obj}}$ from the given $\\mathcal{G}$ .", + "bbox": [ + 75, + 417, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 553, + 282, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Neck. TarViS produces target masks by computing the inner product between $Q_{\\mathrm{out}}$ and the video feature map $F_{4}$ . For this to work, the per-pixel features $\\mathcal{F}$ must be aligned for the same, and dissimilar for different targets. Some image instance segmentation methods [13, 78] apply Deformable Attention [78] to the backbone feature maps to efficiently learn multi-scale image features. For TarViS, however, the features must also be temporally consistent across the entire input video clip. To achieve this, we propose a novel Temporal Neck architecture inspired from the work of Bertasius et al. [7] for video action classification. We enable efficient spatio-temporal feature interaction by applying two types of self-attention in an alternating fashion: the first is spatially global and temporally localized, whereas the second is spatially localized and temporally global. The first operation is implemented with Deformable Attention, following existing work [12, 78]. The second operation, Temporal Attention, involves dividing the input space-time volume into a grid along the spatial axes, and then applying self-attention to the space-time feature volume inside each grid cell. Both operations allow feature", + "bbox": [ + 75, + 583, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aacbac1eceba824eeabe1aa85079743ce7c28a8009455c1fca3a9c0ce70fbcdb.jpg", + "image_caption": [ + "Figure 3. Temporal Neck Layer. Colored regions denote the attention field w.r.t the selected pixel (darkened). Deformable Attention is spatially unrestricted but temporally limited to a single frame, whereas Temporal Attention is spatially localized, but temporally unrestricted. $F_{8}$ is inactive for the temporal attention." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 870, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "interaction across multiple scales. Both attention operations are illustrated in Fig. 3. We exclude $F_{8}$ from temporal attention since we found this to be more memory-efficient without negatively impacting prediction quality.", + "bbox": [ + 496, + 369, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer Decoder. The decoder architecture follows that of Mask2Former [13]: the input queries are iteratively refined over multiple layers. In each layer, the queries first cross-attend to their respective masked video features, then self-attend to each other, followed by feed-forward layers.", + "bbox": [ + 496, + 436, + 890, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference", + "text_level": 1, + "bbox": [ + 500, + 523, + 609, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To infer on videos with arbitrary length, we split videos into clips of length $T_{\\mathrm{clip}}$ with an overlap of $T_{\\mathrm{ov}}$ between successive clips. Object tracks are associated across clips based on their mask IoU in the overlapping frames. For the VOS tasks, the object queries for an intermediate clip are initialized by using the predicted masks in the overlapping frames from the previous clip as a pseudo-ground-truth. For VPS, we average the semantic segmentation logits in the overlapping frames. Our approach is thus near-online because the time delay in obtaining the output for a given frame is at most $T_{\\mathrm{clip}} - T_{\\mathrm{ov}} - 1$ (except for the first clip in the video).", + "bbox": [ + 496, + 547, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 729, + 632, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 755, + 717, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our Temporal Neck contains 6 layers of Deformable and Temporal Attention. We pretrain for $500\\mathrm{k}$ iterations on pseudo-video clips generated by applying on-the-fly augmentations to images from COCO [43], ADE20k [77], Mapillary [48] and Cityscapes [18]. The samples are either trained for VPS, VIS, VOS or PET. This is followed by fine-tuning for $90\\mathrm{k}$ iterations jointly on samples from YouTube-VIS [72], OVIS [54], KITTI-STEP [66],", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18742", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f7bd62b8b504abb59fbcc6cd1c40013fb5adf9389ffc2c095cc63fe19729171a.jpg", + "table_caption": [ + "Table 1. Results for Video Instance Segmentation (VIS) on the YouTube-VIS 2021 [72] and OVIS [54] validation sets." + ], + "table_footnote": [], + "table_body": "
MethodBackboneShared ModelYouTube-VIS 2021OVIS
APAP50AP75AR1AR10APAP50AP75AR1AR10
Mask2Former-VIS [12]R-50X40.660.941.8-------
IDOL [69]R-50X43.968.049.638.050.930.251.330.015.037.5
MinVIS [28]R-50X44.266.048.139.251.725.045.524.013.929.7
VITA [26]R-50X45.767.449.540.953.619.641.217.411.726.0
TarViSR-50✓48.369.653.240.555.931.152.530.415.939.9
Mask2Former-VIS [12]Swin-TX45.968.750.7-------
TarViSSwin-T✓50.971.656.642.257.234.055.034.416.140.9
IDOL [69]Swin-LX56.180.863.545.060.142.665.745.217.949.6
VITA [26]Swin-LX57.580.661.047.762.627.751.924.914.933.0
TarViSSwin-L✓60.281.467.647.664.843.267.844.618.050.4
", + "bbox": [ + 78, + 108, + 893, + 297 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b26befd008a442f377f814359271f0724bff7a9bad57089fd56de6fcafe24d0f.jpg", + "table_caption": [ + "Table 2. Video Panoptic Segmentation (VPS) results for validation sets of KITTI-STEP [66], CityscapesVPS [35] and VIPseg [44]." + ], + "table_footnote": [], + "table_body": "
MethodShared ModelKITTI-STEPCityscapesVPSVIPSeg
STQAQSQVPQ\\( VPQ^{Th} \\)\\( VPQ^{St} \\)VPQ\\( VPQ^{Th} \\)\\( VPQ^{St} \\)STQ
Mask Propagation [66]✘0.670.630.71------
Track [35]✘---55.943.764.8---
VPSNet [35]✘0.560.520.6157.044.766.014.014.014.220.8
VPSNet-SiamTrack [67]✘---57.344.766.417.217.317.321.1
VIP-Deeplab [55]✘---63.149.573.016.012.318.222.0
Clip-PanoFCN [44]✘------22.925.020.831.5
TarViS (R-50)✓0.700.700.6953.335.966.033.539.228.543.1
TarViS (Swin-T)✓0.710.710.7058.042.969.035.842.729.745.3
TarViS (Swin-L)✓0.720.720.7358.943.769.948.058.239.052.9
", + "bbox": [ + 78, + 329, + 892, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CityscapesVPS [35], VIPSeg [44], DAVIS [53] and BURST [3]. We train on 32 Nvidia A100 GPUs with batch size 1 per GPU. For each of the query types ( $Q_{\\mathrm{sem}}$ , $Q_{\\mathrm{inst}}$ , $Q_{\\mathrm{obj}}$ , $Q_{\\mathrm{bg}}$ ) discussed in Sec. 3, we employ a learned query embedding, which is used when computing the $\\mathbf{Key}^T$ Query affinity matrix for multi-head attention inside the decoder. We refer to the supplementary for more details.", + "bbox": [ + 75, + 521, + 470, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark Results", + "text_level": 1, + "bbox": [ + 76, + 642, + 264, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All results are computed with a single, jointly trained model which performs different tasks by simply providing the corresponding query set at run-time.", + "bbox": [ + 75, + 667, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Instance Segmentation (VIS). We evaluate on (1) YouTube-VIS 2021 [72] which covers 40 object classes and contains 2985/421 videos for training/validation, and (2) OVIS [54] which covers 25 object classes. It contains 607/140 videos for training/validation which are comparatively longer and more occluded. The AP scores for both are reported in Tab. 1. For all three backbones, TarViS achieves state-of-the-art results for both benchmarks even though other methods are trained separately per benchmark whereas we use a single model. On YouTube-VIS, TarViS achieves 48.3 AP with a ResNet-50 backbone compared to the 45.7 achieved by VITA [26]. With Swin-L, we achieve", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "60.2 AP which is also higher than the 57.5 by VITA. On OVIS with ResNet-50, our 31.1 AP is higher than the 30.2 for IDOL [69], and with Swin-L, TarViS (43.2 AP) outperforms the current state-of-the-art IDOL (42.6 AP).", + "bbox": [ + 496, + 521, + 893, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Panoptic Segmentation (VPS). We evaluate VPS on three datasets: (1) KITTI-STEP [66], which contains 12/9 lengthy driving scene videos for training/validation with 19 semantic classes (2 thing and 17 stuff classes), (2) CityscapesVPS [35], which contains 50 short driving scene clips, each with 6 annotated frames, and (3) VIPSeg [44], which is a larger dataset with 2806/343 in-the-wild videos for training/validation and 124 semantic classes. The results are reported in Tab. 2. For KITTI-STEP, TarViS achieves $70\\%$ STQ with a ResNet-50 backbone which is better than all existing approaches. The performance further improves to $72\\%$ with Swin-L. For CityscapesVPS, TarViS achieves 58.9 VPQ which is higher than all other methods except VIP-Deeplab [55] (63.1). However, VIP-Deeplab performs monocular depth estimation for additional guidance, and therefore requires ground-truth depth-maps for training.", + "bbox": [ + 496, + 592, + 892, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For VIPSeq, TarViS outperforms existing approaches by a significant margin. With a ResNet-50 backbone, our 33.5 VPQ is $10.6\\%$ higher than the 22.9 by Clip-PanoFCN [44]. With a Swin-Large backbone, TarViS achieves 48.0 VPQ", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18743", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/424a2a54db6d352fc6c7a2268ecc76238b4004e7dd5444449e3c47691f6bcc73.jpg", + "table_caption": [ + "Table 3. Results for VOS on DAVIS [53] and PET on BURST [3]. Detailed PET metrics are provided in supplementary." + ], + "table_footnote": [], + "table_body": "
MethodDAVIS (VOS)BURST (PET)
J&FJFHval allHtest all
UNICORN* [71]70.666.175.0--
HODOR [2]81.378.483.9--
STM [50]81.879.284.3--
CFBI [73]81.979.184.6--
HMMN [60]84.781.987.5--
AOT [74]84.982.387.5--
STCN [17]85.482.288.6--
XMem [15]86.282.989.5--
Box Tracker [32]---12.710.1
STCN+M2F [13, 17]---24.424.9
TarViS (R-50)82.679.385.930.932.1
TarViS (Swin-T)82.879.686.036.036.4
TarViS (Swin-L)85.381.788.537.536.1
", + "bbox": [ + 78, + 122, + 468, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is more than double that of Clip-PanoFCN (22.9). Note that VIP-Deeplab performs significantly worse for VIPSeg (16.0 VPQ), showing that TarViS generalizes better across benchmarks. Finally, we note that larger backbones results in significant performance gains for datasets with in-the-wild internet videos as in VIPSeg, but for specialized driving scene datasets (e.g. KITTI-STEP and Cityscapes-VPS), the improvements are much smaller.", + "bbox": [ + 75, + 359, + 468, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Object Segmentation (VOS). We evaluate VOS on the DAVIS 2017 [53] dataset, which contains 60/30 YouTube videos for training/validation. The results in Tab. 3 show that TarViS achieves 85.3 $\\mathcal{J}\\&\\mathcal{F}$ which is higher than all existing methods except STCN [17] (85.4) and XMem [15] (86.2). As mentioned in Sec. 3.3, encoding objects as queries incurs a loss of fine-grained information, which is detrimental to performance. On the other hand, space-time correspondence (STC) based approaches learn pixel-to-pixel affinities between frames, which enables them to propagate fine-grained object appearance information. We note, however, that TarViS is the first method not based on the STC paradigm which achieves this level is performance (85.3 $\\mathcal{J}\\&\\mathcal{F}$ ), outperforming several STC-based methods as well as all non-STC based methods e.g. HODOR [2] (81.5) and UNICORN [71] (70.6).", + "bbox": [ + 75, + 486, + 468, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point Exemplar-guided Tracking (PET). PET is evaluated on the recently introduced BURST benchmark [3] which contains 500/1000/1500 diverse videos for training/validation/testing. It is a constrained version of VOS which only provides the point coordinates of the object mask centroid instead of the full mask. Tab. 3 shows that existing methods can only tackle either VOS or PET. To verify this, we tried adapting STCN [17] for PET by training it with point masks, but the training did not converge. By contrast, TarViS encodes objects into queries, which enables it to tackle both tasks with a single model since the", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ce3d544c2c49354fbb19480fcf5eb55b9185888dcdb959db88667655a571c8e.jpg", + "image_caption": [ + "Figure 4. Qualitative results from a single TarViS model for all four tasks. Further results are shown in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 498, + 88, + 890, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "object guidance (point or mask) is abstracted behind the EncodeObjects(·) function.", + "bbox": [ + 498, + 342, + 890, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TarViS achieves a $\\mathrm{HOTA_{all}}$ score of 37.5 and 36.4 on the validation and test sets, respectively, which is significantly better than the 24.4 and 24.9 achieved by the best performing baseline method which casts PET as a VOS problem by regressing a pseudo-ground-truth mask from the given point, followed by applying a VOS approach (STCN [17]).", + "bbox": [ + 496, + 373, + 890, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 500, + 474, + 609, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows several architecture/training ablations.", + "bbox": [ + 517, + 497, + 867, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task-specific Training (row 1-3). The first three rows show results for task-specific models. We train a single model for VOS and PET since both tasks are closely related. We note that the VIS-only model performs worse than the multi-task model on YouTube-VIS (46.3 vs. 48.3) but slightly better on OVIS (31.5 vs. 31.1). For VPS, the performance on KITTI-STEP is unchanged, but Cityscapes-VPS and VIPSeg both show improvements with the multi-task model. Lastly, for VOS the task-specific model performs slightly worse on DAVIS (81.1 vs. 82.0) but significantly better on BURST for PET (34.7 vs. 30.9). To summarize, the final, multi-task model performs better on 4/7 benchmarks, worse on 2/7, and matches performance on 1/7 when compared to task-specific models. We thus conclude that the combination of multi-task supervision and more data is generally beneficial for performance.", + "bbox": [ + 496, + 517, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Queries for VIS (row 4). TarViS represents object classes as dynamic query inputs to the network ( $Q_{\\mathrm{sem}}$ , Sec. 3.1). We ablate this by modifying our network to work for only VIS by discarding the semantic/background queries and adopting a technique similar to existing methods [13, 65], i.e. using instance queries $Q_{\\mathrm{inst}}$ in conjunction with a linear layer for classification (separate for each dataset). Comparing the results with the VIS-only setting which is trained on similar data, we see that this architecture", + "bbox": [ + 496, + 763, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18744", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf511f89a9cce0a47c1e705f232d86c05d14513f509014ead367c3aad7a7c6b9.jpg", + "table_caption": [ + "Table 4. Ablation experiment results with ResNet-50 backbone. C-VPS: CityscapesVPS, YTVIS: YouTube-VIS, KITTI: KITTI-STEP." + ], + "table_footnote": [], + "table_body": "
SettingVideo Training DataVISVPSVOSPET
YTVISOVISKITTIC-VPSVIPSegYTVIS (mAP)OVIS (mAP)KITTI (STQ)C-VPS (VPQ)VIPSeg (VPQ)DAVIS (J&F)BURST (HOTAall)
1. VIS✓✓46.331.5-----
2. VPS✓✓✓--0.7049.732.4--
3. VOS + PET✓✓----81.134.7
4. No Semantic Queries✓✓44.729.8-----
5. No Temporal Neck✓✓✓✓✓42.822.30.6951.228.978.730.3
Final✓✓✓✓✓48.331.10.7053.333.582.030.9
", + "bbox": [ + 78, + 108, + 898, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performs worse than the VIS-only setting on both YouTube-VIS (44.7 vs. 46.3) and OVIS (29.8 vs. 31.5). Thus, our semantic query based classification makes the network architecture task-agnostic and also yields better performance.", + "bbox": [ + 75, + 290, + 468, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Neck (row 4). We validate our novel Temporal Neck (Sec. 3.4) by training a model with a simpler neck that contains only Deformable Attention layers [78], similar to Mask2Former [13], i.e. there is no feature interaction across frames. Doing this degrades performance across all datasets, with particularly large drops for YouTube-VIS (42.8 vs. 48.3) and OVIS (22.3 vs. 31.1). This shows that inter-frame feature interactions enabled by our Temporal Neck are highly beneficial for down-stream tasks.", + "bbox": [ + 75, + 356, + 468, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 507, + 191, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Training on multiple datasets/tasks does not necessarily improve performance on all benchmarks. For VOS, the model exhibits class bias and sometimes fails to track unusual objects which were not seen during training.", + "bbox": [ + 75, + 532, + 468, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Outlook. We jointly trained TarViS for four different tasks to validate its generalization capability. The architecture can, however, tackle any video segmentation task for which the targets can be encoded as queries. The recent emergence of joint language- vision models [41, 56, 57] thus makes it possible to perform multi-object segmentation based on a text prompt if the latter can be encoded as a target query using a language encoder [19]. Another interesting possibility is that TarViS could be applied to multiple tasks in the same forward pass by simply concatenating the task-specific queries. Fig. 5 offers a promising outlook for this; it shows our model's output for a video clip from a popular TV series where we perform VIS and VOS simultaneously by providing the semantic query for the 'person' class (from YouTube-VIS [72]), and the VOS-based object queries for the dragon by annotating its first frame mask, i.e. $Q_{\\mathrm{in}} = \\mathrm{concat}(Q_{\\mathrm{sem}}, Q_{\\mathrm{inst}}, Q_{\\mathrm{obj}}, Q_{\\mathrm{bg}})$ . TarViS successfully segments all four persons in the scene (VIS) and the dragon (VOS), even though our model was never trained to simultaneously tackle both tasks in a single forward pass.", + "bbox": [ + 75, + 598, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8ad2840a6ba0111007f04d5c2bddaaa31b66f876f9140351c49f9d69ef9d146.jpg", + "image_caption": [ + "Figure 5. TarViS performing VIS and VOS in a single forward pass. We provide the mask for the dragon on the left, and the semantic query for the 'person' class." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 696, + 459 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b8ef09b0ebbf85a3cd99440f9cbc3b49e31cd67c2b0d55f3460e48badb96cea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 287, + 890, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 513, + 617, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented TarViS: a novel, unified approach for tackling any task requiring pixel-precise segmentation of a set of targets in video. We adopt a generalized paradigm where the task-specific targets are encoded into a set of queries which are then input to our network together with the video features. The network is trained to produce segmentation masks for each target entity, but is inherently agnostic to the task-specific definition of these targets. To demonstrate the effectiveness of our approach, we applied it to four different video segmentation tasks (VIS, VPS, VOS, PET). We showed that a single TarViS model can be jointly trained for all tasks, and during inference can hot-swap between tasks without any task-specific fine-tuning. Our model achieved state-of-the-art performance on five benchmarks and has multiple, promising directions for future work.", + "bbox": [ + 496, + 539, + 890, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This project was partially funded by ERC CoG DeeVise (ERC-2017-COG-773161) and BMBF project NeuroSys-D (03ZU1106DA). Compute resources were granted by RWTH Aachen under project ID supp0003, and by the Gauss Centre for Supercomputing e.V. through the John von Neumann Institute for Computing on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre.", + "bbox": [ + 496, + 770, + 890, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18745", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Arxiv, 2022. 1, 2", + "[2] Ali Athar, Jonathon Luiten, Alexander Hermans, Deva Ramanan, and Bastian Leibe. Hodor: High-level object descriptors for object re-segmentation in video learned from static images. In CVPR, 2022. 2, 4, 7", + "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2, 3, 5, 6, 7", + "[4] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 2", + "[5] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. A single-stage, bottom-up approach for occluded vis using spatio-temporal embeddings. In ICCV-W, 2021. 2", + "[6] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2", + "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5", + "[8] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In ECCV, 2010. 1", + "[9] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2", + "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3", + "[11] Richard Caruana. Multitask Learning: A Knowledge-Based Source of Inductive Bias. In ICML, 1993. 2", + "[12] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In Arxiv, 2021. 2, 5, 6", + "[13] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In CVPR, 2022. 2, 3, 4, 5, 7, 8", + "[14] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2, 4", + "[15] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3, 7", + "[16] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 3" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 3, 7", + "[18] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5", + "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019. 8", + "[20] Ahmed Elgammal, David Harwood, and Larry Davis. Nonparametric model for background subtraction. In ECCV, 2000. 1", + "[21] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. Moving obstacle detection in highly dynamic scenes. In ICRA, 2009. 1", + "[22] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2", + "[23] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, 2021. 2", + "[24] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-Task Self-Training for Learning General Representations. In CVPR, 2021. 2", + "[25] Helmut Grabner, Michael Grabner, and Horst Bischof. Real-time tracking via on-line boosting. In BMVC, 2006. 1", + "[26] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6", + "[27] Namdar Homayounfar, Justin Liang, Wei-Chiu Ma, and Raquel Urtasun. Videoclick: Video object segmentation with a single click. Arxiv, 2021. 3", + "[28] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6", + "[29] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2", + "[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. *ICLR*, 2022. 1, 2", + "[31] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 1, 2", + "[32] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020.7", + "[33] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18746", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. In NeurIPS, 2021. 2", + "[35] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2, 4, 6", + "[36] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 2", + "[37] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2", + "[38] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. Uvim: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1, 2", + "[39] Dieter Koller, Joseph Weber, and Jitendra Malik. Robust multiple car tracking with occlusion reasoning. In ECCV, 1994. 1", + "[40] Bastian Leibe, Konrad Schindler, Nico Cornelis, and Luc Van Gool. Coupled object detection and tracking from static cameras and moving vehicles. PAMI, 2008. 1", + "[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 8", + "[42] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, 2021. 2", + "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5", + "[44] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 6", + "[45] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. In arXiv:1603.00831, 2016. 2", + "[46] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 2", + "[47] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2", + "[48] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5", + "[49] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. PAMI, 2013. 1", + "[50] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2, 5, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Kilian Pfeiffer, Alexander Hermans, István Sárándi, Mark Weber, and Bastian Leibe. Visual person understanding through multi-task and multi-dataset learning. In GCPR, 2019. 2", + "[52] Hamed Pirsiavash, Deva Ramanan, and Charless C Fowlkes. Globally-optimal greedy algorithms for tracking a variable number of objects. In CVPR, pages 1201-1208. IEEE, 2011. 1", + "[53] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. In *Arxiv*, 2017. 2, 6, 7", + "[54] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. In IJCV, 2022. 2, 5, 6", + "[55] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, 2021. 2, 4, 6", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8", + "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 8", + "[58] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. In *Arxiv*, 2022. 1, 2", + "[59] Sebastian Ruder. An Overview of Multi-Task Learning in Deep Neural Networks. arXiv:1706.05098, 2017. 2", + "[60] Hongje Seong, Seoung Wug Oh, Joon-Young Lee, Seongwon Lee, Suhyeon Lee, and Euntai Kim. Hierarchical memory matching network for video object segmentation. In ICCV, 2021. 3, 7", + "[61] C. Stauffer and W.E.L. Grimson. Adaptive background mixture models for real-time tracking. In CVPR, 1999. 1", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1", + "[63] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmentation. In CVPR, 2019. 3", + "[64] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 2", + "[65] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2, 3, 7", + "[66] Mark Weber, Jun Xie, Maxwell Collins, Yukun Zhu, Paul Voigtaender, Hartwig Adam, Bradley Green, Andreas" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18747", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Geiger, Bastian Leibe, Daniel Cremers, Aljosa Osep, Laura Leal-Taixe, and Liang-Chieh Chen. STEP: Segmenting and tracking every pixel. In NeurIPS, 2021. 2, 4, 5, 6", + "[67] Sanghyun Woo, Dahun Kim, Joon-Young Lee, and In So Kweon. Learning to associate every segment for video panoptic segmentation. In CVPR, 2021. 6", + "[68] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 2", + "[69] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6", + "[70] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, 2021. 3", + "[71] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2, 7", + "[72] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5, 6, 8", + "[73] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3, 7", + "[74] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 3, 7", + "[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2", + "[76] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 2", + "[77] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 5", + "[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICML, 2020. 2, 3, 5, 8" + ], + "bbox": [ + 78, + 90, + 468, + 646 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18748", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_model.json b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d99da39395da6ad571f58638d8ac4d2469eb89cd --- /dev/null +++ b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_model.json @@ -0,0 +1,2180 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.119, + 0.82, + 0.141 + ], + "angle": 0, + "content": "TarViS: A Unified Approach for Target-based Video Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.168, + 0.847, + 0.235 + ], + "angle": 0, + "content": "Ali Athar\\(^{1}\\) Alexander Hermans\\(^{1}\\) Jonathon Luiten\\(^{1,2}\\) Deva Ramanan\\(^{2}\\) Bastian Leibe\\(^{1}\\) \n\\(^{1}\\)RWTH Aachen University, Germany \\(^{2}\\)Carnegie Mellon University, USA \n{athar, hermans, luiten, leibe}@vision.rwth-aachen.de deva@cs.cmu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.65 + ], + "angle": 0, + "content": "The general domain of video segmentation is currently fragmented into different tasks spanning multiple benchmarks. Despite rapid progress in the state-of-the-art, current methods are overwhelmingly task-specific and cannot conceptually generalize to other tasks. Inspired by recent approaches with multi-task capability, we propose TarViS: a novel, unified network architecture that can be applied to any task that requires segmenting a set of arbitrarily defined 'targets' in video. Our approach is flexible with respect to how tasks define these targets, since it models the latter as abstract 'queries' which are then used to predict pixel-precise target masks. A single TarViS model can be trained jointly on a collection of datasets spanning different tasks, and can hot-swap between tasks during inference without any task-specific retraining. To demonstrate its effectiveness, we apply TarViS to four different tasks, namely Video Instance Segmentation (VIS), Video Panoptic Segmentation (VPS), Video Object Segmentation (VOS) and Point Exemplar-guided Tracking (PET). Our unified, jointly trained model achieves state-of-the-art performance on 5/7 benchmarks spanning these four tasks, and competitive performance on the remaining two. Code and model weights are available at: https://github.com/Ali2500/TarViS" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.679, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The ability to understand video scenes has been a longstanding goal of computer vision research because of wideranging applications in intelligent vehicles and robots. Early approaches tackled simpler tasks involving contour-based [33, 39] and box-level tracking [21, 25, 40, 52], background subtraction [20, 61], and motion segmentation [8, 49]. The deep learning boom then revolutionized the landscape by enabling methods to perform pixel-precise segmentation on challenging, real-world videos. In the past few years, a number of benchmarks have emerged, which evaluate how well methods can perform video segmentation according to various task formulations. Over time, these tasks/benchmarks have ballooned into separate re" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.266, + 0.892, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.572, + 0.894, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Predicted results from a jointly trained TarViS model for four different video segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.689 + ], + "angle": 0, + "content": "search sub-communities. Although existing methods are rapidly improving the state-of-the-art for these benchmarks, each of them typically tackles only one narrowly-defined task, and generalizing them is non-trivial since the task definition is baked into the core approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We argue that this fragmentation is unnecessary because video target segmentation tasks all require the same high-level capability, namely that of identifying, localizing and tracking rich semantic concepts. Meanwhile, recent progress on Transformer networks has enabled the wider AI research community to move towards unified, multi-task architectures [1, 30, 31, 38, 58], because the attention operation [62] is well-suited for processing feature sets with arbitrary structure and data modality. These developments give us the opportunity to unify the fractured landscape of target-based video segmentation. In this paper, we propose TarViS: a novel architecture which enables a single, unified model to be jointly trained for multiple video segmentation tasks. During inference, the same model can perform differ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18738" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.465, + 0.108 + ], + "angle": 0, + "content": "ent tasks at runtime by specifying the segmentation target." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.108, + 0.47, + 0.29 + ], + "angle": 0, + "content": "The core idea is that TarViS tackles the generic task of segmenting a set of arbitrary targets in video (defined as semantic classes or as specific objects). These targets are encoded as queries which, together with the video features, are input to a Transformer-based model. The model iteratively refines these queries and produces a pixel-precise mask for each target entity. This formulation conceptually fuses all video segmentation tasks [3, 54, 66, 72] which fall under the umbrella of the above-mentioned generic task, because they differ only in how the targets are defined. During both training and inference, TarViS can hot-swap between tasks at run-time by providing the desired target query set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.29, + 0.471, + 0.488 + ], + "angle": 0, + "content": "To demonstrate our generalization capability, we tackle four different tasks: (1) Video Instance Segmentation (VIS) [54, 72], (2) Video Panoptic Segmentation (VPS) [35], (3) Video Object Segmentation [53], and (4) Point Exemplar-guided Tracking [3] (PET). For VIS, the segmentation targets are all objects in the video belonging to a predefined set of classes. The target set for VPS includes that for VIS, and additionally, a set of non-instantiable stuff semantic classes. For VOS, the targets are a specific set of objects for which the first-frame ground-truth mask is provided. PET is a more constrained version of VOS which only provides the location of a single point inside the object, rather than the full object mask." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.47, + 0.64 + ], + "angle": 0, + "content": "Existing methods for these tasks lack generalization capability because task-specific assumptions are typically baked into the approach (see Sec. 2 and 3 for details). In contrast, TarViS can tackle all four tasks with a unified model because we encode the task-specific targets as a set of queries, thus decoupling the network architecture from the task definition. Moreover, our approach can theoretically generalize further, e.g., one could potentially define the target set as all objects described by a given text prompt, though this is beyond the scope of this paper." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.641, + 0.47, + 0.762 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows: we propose TarViS, a novel architecture that can perform any task requiring segmentation of a set of targets from video. For the first time, we are able to jointly train and infer a single model on a collection of datasets spanning the four aforementioned tasks (VIS, VPS, VOS, PET). Our experimental results show that TarViS performs competitively for VOS, and achieves state-of-the-art results for VIS, VPS and PET." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.779, + 0.22, + 0.794 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Multi-task Models. Multi-task learning has a long history [11] with several architectures and training strategies [24,36,37,51,59,76]. Earlier approaches mostly consist of a shared backbone with fixed task-specific heads, whereas we design a more general architecture for video segmentation with task-specific targets to specify what to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.41 + ], + "angle": 0, + "content": "segment. Our approach is inspired by recent attention-based models, e.g., PerceiverIO [30, 31], which can be trained on diverse data modalities and task-specific heads are replaced with output queries. UViM [38] follows a similar direction by creating a unified architecture for diverse dense prediction tasks. However, both of these models are trained separately for different tasks. Recent, powerful multi-task vision language models such as Flamingo [1] and GATO [58] tackle a multitude of tasks by requiring a sequence of task-specific input-output examples to prime the model. This is conceptually similar to our task-specific targets, however, our model does not require per-task priming. Moreover, our targets are not modeled as sequence prompts, and we aim for a video segmentation model which is several orders of magnitude smaller. In the realm of video tracking and segmentation, the recently proposed UNICORN [71] model tackles multiple object tracking-related tasks with a unified architecture. Unlike TarViS, however, UNICORN follows the task-specific output head approach and is generally oriented towards box-level tracking tasks [22, 45, 47, 75], thus requiring non-trivial modifications to tackle VPS or PET." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Query-based Transformer Architectures. Several works [2, 10, 13, 30, 31, 46, 65, 78] use query-based Transformer architectures for various tasks. The workhorse for task learning here is the iterative application of self- and cross-attention, where a set of query vectors (e.g., representing objects) are refined by interacting with each other, and with the input data (e.g., an image). Unlike existing methods which use queries in a task-specific context, TarViS adopts a query-based Transformer architecture in which the queries serve as a mechanism for decoupling the task definition from the architecture, i.e., our model can learn to tackle different tasks while being agnostic to their definition because the latter is abstracted behind a set of queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.894, + 0.746 + ], + "angle": 0, + "content": "Task-specific Video Segmentation. Current Video Instance Segmentation (VIS) methods broadly work by predicting object tracks in the video, followed by classification into a pre-defined set of categories. Several approaches [6,9,23,28,34,42,54,64,69,72] are based on the tracking-by-detection paradigm, some model video as a joint spatiotemporal volume [4,5], whereas recent works [12,26,29,65, 68] adopt Transformer-based architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For Video Panoptic Segmentation (VPS), methods [35, 55, 66] generally extend image-level panoptic approaches [14] by employing multi-head architectures for semantic segmentation and instance mask regression, classification, and temporal association. In the Video Object Segmentation (VOS) community, state-of-the-art methods are broadly based on the seminal work of Oh et al. [50], which learns space-time correspondences between pixels in different video frames, and then uses these to propagate the first-frame masks across the video. Subsequent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18739" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.668, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.089, + 0.891, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.893, + 0.351 + ], + "angle": 0, + "content": "Figure 2. TarViS Architecture. Segmentation targets for different tasks are represented by a set of abstract target queries \\( Q_{\\mathrm{in}} \\). The core network (in green) is agnostic to the task definitions. The inner product between the output queries \\( Q_{\\mathrm{out}} \\) and video feature \\( F_{4} \\) yields segmentation masks as required by the task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.468, + 0.466 + ], + "angle": 0, + "content": "methods [15-17, 60, 63, 70, 73, 73, 74] have significantly improved the performance and efficiency of this approach. Point Exemplar-guided Tracking (PET) [3, 27] is a fairly new task for which the current best approach [3] regresses a pseudo-ground-truth mask from the given point coordinates, and then applies a VOS method [17] to this mask." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.557 + ], + "angle": 0, + "content": "The above methods thus incorporate task-specific assumptions into their core approach. This can benefit pertask performance, but makes it difficult for them to generalize across tasks. By contrast, TarViS can tackle all four aforementioned tasks, and generally any target-based video segmentation task, with a single, jointly trained model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.169, + 0.588 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.674 + ], + "angle": 0, + "content": "TarViS can segment arbitrary targets in video since the architecture is flexible with respect to how these targets are defined, thus enabling us to conceptually unify and jointly tackle the four aforementioned tasks (VIS, VPS, VOS, PET). The architecture is illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "For all tasks, the common input to the network is an RGB video clip of length \\( T \\) denoted by \\( V \\in \\mathbb{R}^{H \\times W \\times T \\times 3} \\). This is input to a 2D backbone network which produces image-level feature maps, followed by a Temporal Neck, which enables feature interaction across time and outputs a set of temporally consistent, multi-scale, \\( D \\)-dimensional feature maps \\( \\mathcal{F} = \\{F_{32}, F_{16}, F_{8}, F_{4}\\} \\) where \\( F_{s} \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times T \\times D} \\). The feature maps \\( \\mathcal{F} \\) are then fed to our Transformer decoder, together with a set of queries \\( Q_{\\mathrm{in}} \\) which represent the segmentation targets. The decoder applies successive layers of self- and masked cross-attention wherein the queries are iteratively refined by attending to each other, and to the feature maps, respectively. The refined queries output by the decoder are denoted with \\( Q_{\\mathrm{out}} \\). The following subsections explain how TarViS tackles each task in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.76, + 0.391 + ], + "angle": 0, + "content": "3.1. Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.553 + ], + "angle": 0, + "content": "VIS defines the segmentation target set as all objects belonging to a set of predefined classes. Accordingly, the input query set \\( Q_{\\mathrm{in}} \\) for VIS contains three types of queries: (1) semantic queries denoted by \\( Q_{\\mathrm{sem}} \\in \\mathbb{R}^{C \\times D} \\) where \\( C \\) is the number of classes defined by the dataset, i.e., each \\( D \\)-dimensional vector in \\( Q_{\\mathrm{sem}} \\) represents a particular semantic class. (2) instance queries denoted by \\( Q_{\\mathrm{inst}} \\in \\mathbb{R}^{I \\times D} \\) where \\( I \\) is assumed to be an upper bound on the number of instances in the video clip, and (3) a background query denoted by \\( Q_{\\mathrm{bg}} \\in \\mathbb{R}^{1 \\times D} \\) to capture inactive instance queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.893, + 0.695 + ], + "angle": 0, + "content": "The three query sets are concatenated, i.e., \\( Q_{\\mathrm{in}} = \\operatorname{concat}(Q_{\\mathrm{sem}}, Q_{\\mathrm{inst}}, Q_{\\mathrm{bg}}) \\), and input to the Transformer decoder, which refines their feature representation through successive attention layers and outputs a set of queries \\( Q_{\\mathrm{out}} = \\operatorname{concat}(Q_{\\mathrm{sem}}', Q_{\\mathrm{inst}}', Q_{\\mathrm{bg}}') \\). These are then used to produce temporally consistent instance mask logits by computing the inner product \\( \\langle F_4, Q_{\\mathrm{inst}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times I} \\). To obtain classification logits, we compute the inner product \\( \\langle Q_{\\mathrm{inst}}', \\operatorname{concat}(Q_{\\mathrm{sem}}', Q_{\\mathrm{bg}}') \\rangle \\in \\mathbb{R}^{I \\times (C + 1)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.849 + ], + "angle": 0, + "content": "The three types of queries are initialized randomly at the start of training and optimized thereafter. The instance queries \\( Q_{\\mathrm{inst}} \\) enable us to segment a varying number of objects from the input clip. During training, we apply Hungarian matching between the predicted and ground-truth instance masks to assign instance queries to video instances, and then supervise their predicted masks and classification logits accordingly. When training on multiple datasets with heterogeneous classes, the semantic query sets are separately initialized per dataset, but \\( Q_{\\mathrm{inst}} \\) and \\( Q_{\\mathrm{bg}} \\) are shared." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Comparison to Instance Segmentation Methods. Several Transformer-based methods [10, 13, 65, 78] for image/video instance segmentation also use queries to seg" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18740" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "ment a variable number of input instances. The key difference to our approach is the handling of object classes: existing works employ only instance queries which are input to a fully-connected layer with a fan-out of \\( C + 1 \\) to obtain classification (and background) scores. The notion of class-guided instance segmentation is thus baked into the approach. By contrast, TarViS is agnostic to the task-specific notion of object classes because it models them as arbitrary queries which are dynamic inputs to the network. The semantic representation for these queries is thus decoupled from the core architecture and is only learned via loss supervision. An important enabler for this approach is the background query \\( Q_{\\mathrm{bg}} \\), which serves as a 'catch-all' class to represent everything that is not in \\( Q_{\\mathrm{sem}} \\). It is used to classify non-active instance queries, and its mask logits are supervised to segment all non-object input pixels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.342, + 0.342, + 0.359 + ], + "angle": 0, + "content": "3.2. Video Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.518 + ], + "angle": 0, + "content": "VPS defines the segmentation targets as all objects belonging to a set of thing classes (e.g., 'person', 'car'), and additionally, a set of non-instantiable stuff classes (e.g., 'sky', 'grass') which cover all non-object pixels. TarViS can tackle VPS with virtually no modification to the workflow in Sec. 3.1. We can compute semantic segmentation masks for the input clip by simply taking the inner product between \\( Q_{\\mathrm{sem}} \\) and the video features: \\( \\langle F_4, Q_{\\mathrm{sem}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times C} \\). Note that here, \\( Q_{\\mathrm{sem}} \\) contains queries representing both thing and stuff classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.473, + 0.689 + ], + "angle": 0, + "content": "Comparison to VPS Methods. Current VPS datasets [35, 66] involve driving scene videos captured from moving vehicles. Methods tackling this task [35, 55] are based on earlier image panoptic segmentation approaches [14] which involve multi-head networks for semantic and instance segmentation prediction. In terms of image-level panoptic segmentation, Mask2Former [13] uses a Transformer-based architecture, but it models stuff classes as instances which are Hungarian-matched to the ground-truth target during training, whereas TarViS models semantic classes and instances using separate, designated queries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.471, + 0.73 + ], + "angle": 0, + "content": "3.3. Video Object Segmentation and Point Exemplar-guided Tracking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.904 + ], + "angle": 0, + "content": "VOS and PET are instantiations of a general task where the segmentation targets are a set of \\(O\\) objects for which some ground-truth cue \\(\\mathcal{G}\\) is given. For VOS, \\(\\mathcal{G}\\) is provided as the first-frame object masks \\(M_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times H \\times W}\\), whereas for PET, \\(\\mathcal{G}\\) is provided as the \\((x,y)\\) coordinates \\(P_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times 2}\\) of a point inside each of the objects. TarViS jointly tackles both tasks by adopting a generalized approach in which the \\(O\\) target objects are encoded into a set of object queries \\(Q_{\\mathrm{obj}}\\). Thus, both VOS and PET boil down to designing a function Encode0objects(\\cdot) which regresses \\(Q_{\\mathrm{obj}}\\) from the ground-truth cues \\(\\mathcal{G}\\) and feature maps \\(\\mathcal{F}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.092, + 0.892, + 0.108 + ], + "angle": 0, + "content": "\\[\nQ _ {\\mathrm {o b j}} \\leftarrow \\operatorname {E n c o d e 0 b j e c t s} (\\mathcal {G}, \\mathcal {F}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.132, + 0.892, + 0.178 + ], + "angle": 0, + "content": "Note that \\( Q_{\\mathrm{obj}} \\) is conceptually analogous to \\( Q_{\\mathrm{sem}} \\) and \\( Q_{\\mathrm{inst}} \\) used for VIS in that all three are abstract representations for their respective task-specific segmentation targets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Video Object Segmentation. We seek inspiration from HODOR [2] to implement EncodeObjects for VOS, a recent method for weakly-supervised VOS, which encodes objects into concise descriptors as follows: the descriptors are initialized by average pooling the image features inside the object masks, followed by iterative refinement where the descriptors attend to each other (self-attention) and to their respective soft-masked image features (cross-attention)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.893, + 0.531 + ], + "angle": 0, + "content": "For TarViS, we employ a lightweight Object Encoder with a similar workflow to encode the objects as a set of queries \\( Q_{\\mathrm{obj}} \\), but with two differences to HODOR [2]: instead of cross-attending to the entire image feature map \\( (H \\cdot W \\) points) with soft-masked attention, we apply hard-masked cross-attention to at most \\( p_{\\mathrm{max}} \\) feature points per object, where \\( p_{\\mathrm{max}} \\ll H \\cdot W \\). Object masks containing more than \\( p_{\\mathrm{max}} \\) points are sub-sampled accordingly. This significantly improves the memory/run-time overhead of our Object Encoder. Secondly, we note that the process of distilling object features into a single descriptor involves a loss of object appearance information, which degrades performance. We therefore model each object with \\( q_{o} \\) queries (instead of one) by spatially dividing each object mask into \\( q_{o} \\) segments, i.e., \\( Q_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times q_{o} \\times D} \\) (we use \\( q_{o} = 4 \\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.894, + 0.712 + ], + "angle": 0, + "content": "In addition to \\( Q_{\\mathrm{obj}} \\), we initialize a set of background queries \\( Q_{\\mathrm{bg}} \\in \\mathbb{R}^{B \\times D} \\) to model the non-target pixels in the reference frame. Following HODOR [2], we employ multiple background queries, which are initialized dynamically by dividing the video frame containing the groundtruth masks \\( M_{\\mathrm{obj}} \\) into a \\( 4 \\times 4 \\) grid and average pooling the non-object pixels in each grid cell. The Object Encoder jointly refines the background and object queries to yield \\( Q_{\\mathrm{in}} = \\text{concat}(Q_{\\mathrm{obj}}, Q_{\\mathrm{bg}}) \\). During training, the mask logits for the multiple background queries are aggregated per-pixel by applying \\( \\max(\\cdot) \\) and supervised to segment all pixels not part of the target object set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.894, + 0.821 + ], + "angle": 0, + "content": "The remaining workflow follows that for VIS and VPS: \\( Q_{\\mathrm{in}} \\) is input to the Transformer decoder together with the video features \\( \\mathcal{F} \\). The refined output query set \\( Q_{\\mathrm{out}} = \\operatorname{concat}(Q_{\\mathrm{obj}}', Q_{\\mathrm{bg}}') \\) is then used to compute the inner product \\( \\langle F_4, Q_{\\mathrm{obj}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times O \\times q_o} \\). Subsequently, \\( \\max(\\cdot) \\) is applied on the \\( q_o \\)-sized dimension to obtain the final mask logits for the \\( O \\) target objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Point Exemplar-guided Tracking. For PET we implement Encode0objects in the exactly same way as VOS: the given point coordinates \\( P_{\\mathrm{obj}} \\) are converted into a mask with just one non-zero pixel, followed by iterative refinement by the Object Encoder (with shared weights for VOS and PET)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18741" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "The only difference is that here we represent each of the \\(O\\) objects with just one query, i.e., \\(Q_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times D}\\) (\\(q_o = 1\\)). The subsequent workflow is also identical to that for VOS: the queries are refined by the Transformer decoder followed by an inner product with \\(F_4\\) to obtain object mask logits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.174, + 0.47, + 0.416 + ], + "angle": 0, + "content": "Comparison to VOS and PET Methods. Current state-of-the-art VOS methods are largely based on STM [50]. It involves learning pixel-to-pixel correspondences across video frames, which are then used to propagate the given object mask across the video. This approach is effective since every pixel in the given mask can be individually mapped to future frames, thus preserving fine-grained object details. The core approach is, however, task-specific since it assumes the availability of first-frame object masks, and does not generalize to the PET (see Sec. 4.2). PET can be viewed as a more constrained version of VOS, where only a single object point is provided instead of the full mask. Consequently, PET [3] is currently tackled by casting it as a VOS problem by using an image instance segmentation network [13] to regress pseudo-ground-truth object masks from the given point coordinates \\( P_{\\mathrm{obj}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.54 + ], + "angle": 0, + "content": "On the other hand, our approach of encoding objects as concise queries causes loss of fine-grained object appearance information, but it has the advantage of being agnostic to how \\(\\mathcal{G}\\) is defined. As evident from the unified workflow for VOS and PET, any variation of these tasks with arbitrary ground-truth cues \\(\\mathcal{G}\\) can be seamlessly fused into our architecture as long as we can implement an effective Encode0bjects function to regress \\(Q_{\\mathrm{obj}}\\) from the given \\(\\mathcal{G}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.283, + 0.569 + ], + "angle": 0, + "content": "3.4. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Temporal Neck. TarViS produces target masks by computing the inner product between \\( Q_{\\mathrm{out}} \\) and the video feature map \\( F_{4} \\). For this to work, the per-pixel features \\( \\mathcal{F} \\) must be aligned for the same, and dissimilar for different targets. Some image instance segmentation methods [13, 78] apply Deformable Attention [78] to the backbone feature maps to efficiently learn multi-scale image features. For TarViS, however, the features must also be temporally consistent across the entire input video clip. To achieve this, we propose a novel Temporal Neck architecture inspired from the work of Bertasius et al. [7] for video action classification. We enable efficient spatio-temporal feature interaction by applying two types of self-attention in an alternating fashion: the first is spatially global and temporally localized, whereas the second is spatially localized and temporally global. The first operation is implemented with Deformable Attention, following existing work [12, 78]. The second operation, Temporal Attention, involves dividing the input space-time volume into a grid along the spatial axes, and then applying self-attention to the space-time feature volume inside each grid cell. Both operations allow feature" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.872, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.359 + ], + "angle": 0, + "content": "Figure 3. Temporal Neck Layer. Colored regions denote the attention field w.r.t the selected pixel (darkened). Deformable Attention is spatially unrestricted but temporally limited to a single frame, whereas Temporal Attention is spatially localized, but temporally unrestricted. \\( F_{8} \\) is inactive for the temporal attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.432 + ], + "angle": 0, + "content": "interaction across multiple scales. Both attention operations are illustrated in Fig. 3. We exclude \\( F_{8} \\) from temporal attention since we found this to be more memory-efficient without negatively impacting prediction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.513 + ], + "angle": 0, + "content": "Transformer Decoder. The decoder architecture follows that of Mask2Former [13]: the input queries are iteratively refined over multiple layers. In each layer, the queries first cross-attend to their respective masked video features, then self-attend to each other, followed by feed-forward layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.61, + 0.539 + ], + "angle": 0, + "content": "3.5. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.716 + ], + "angle": 0, + "content": "To infer on videos with arbitrary length, we split videos into clips of length \\( T_{\\mathrm{clip}} \\) with an overlap of \\( T_{\\mathrm{ov}} \\) between successive clips. Object tracks are associated across clips based on their mask IoU in the overlapping frames. For the VOS tasks, the object queries for an intermediate clip are initialized by using the predicted masks in the overlapping frames from the previous clip as a pseudo-ground-truth. For VPS, we average the semantic segmentation logits in the overlapping frames. Our approach is thus near-online because the time delay in obtaining the output for a given frame is at most \\( T_{\\mathrm{clip}} - T_{\\mathrm{ov}} - 1 \\) (except for the first clip in the video)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.731, + 0.633, + 0.748 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.718, + 0.772 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our Temporal Neck contains 6 layers of Deformable and Temporal Attention. We pretrain for \\(500\\mathrm{k}\\) iterations on pseudo-video clips generated by applying on-the-fly augmentations to images from COCO [43], ADE20k [77], Mapillary [48] and Cityscapes [18]. The samples are either trained for VPS, VIS, VOS or PET. This is followed by fine-tuning for \\(90\\mathrm{k}\\) iterations jointly on samples from YouTube-VIS [72], OVIS [54], KITTI-STEP [66]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18742" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.089, + 0.835, + 0.104 + ], + "angle": 0, + "content": "Table 1. Results for Video Instance Segmentation (VIS) on the YouTube-VIS 2021 [72] and OVIS [54] validation sets." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.109, + 0.894, + 0.299 + ], + "angle": 0, + "content": "
MethodBackboneShared ModelYouTube-VIS 2021OVIS
APAP50AP75AR1AR10APAP50AP75AR1AR10
Mask2Former-VIS [12]R-50X40.660.941.8-------
IDOL [69]R-50X43.968.049.638.050.930.251.330.015.037.5
MinVIS [28]R-50X44.266.048.139.251.725.045.524.013.929.7
VITA [26]R-50X45.767.449.540.953.619.641.217.411.726.0
TarViSR-50✓48.369.653.240.555.931.152.530.415.939.9
Mask2Former-VIS [12]Swin-TX45.968.750.7-------
TarViSSwin-T✓50.971.656.642.257.234.055.034.416.140.9
IDOL [69]Swin-LX56.180.863.545.060.142.665.745.217.949.6
VITA [26]Swin-LX57.580.661.047.762.627.751.924.914.933.0
TarViSSwin-L✓60.281.467.647.664.843.267.844.618.050.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.312, + 0.874, + 0.327 + ], + "angle": 0, + "content": "Table 2. Video Panoptic Segmentation (VPS) results for validation sets of KITTI-STEP [66], CityscapesVPS [35] and VIPseg [44]." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.33, + 0.893, + 0.498 + ], + "angle": 0, + "content": "
MethodShared ModelKITTI-STEPCityscapesVPSVIPSeg
STQAQSQVPQ\\( VPQ^{Th} \\)\\( VPQ^{St} \\)VPQ\\( VPQ^{Th} \\)\\( VPQ^{St} \\)STQ
Mask Propagation [66]✘0.670.630.71------
Track [35]✘---55.943.764.8---
VPSNet [35]✘0.560.520.6157.044.766.014.014.014.220.8
VPSNet-SiamTrack [67]✘---57.344.766.417.217.317.321.1
VIP-Deeplab [55]✘---63.149.573.016.012.318.222.0
Clip-PanoFCN [44]✘------22.925.020.831.5
TarViS (R-50)✓0.700.700.6953.335.966.033.539.228.543.1
TarViS (Swin-T)✓0.710.710.7058.042.969.035.842.729.745.3
TarViS (Swin-L)✓0.720.720.7358.943.769.948.058.239.052.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.471, + 0.629 + ], + "angle": 0, + "content": "CityscapesVPS [35], VIPSeg [44], DAVIS [53] and BURST [3]. We train on 32 Nvidia A100 GPUs with batch size 1 per GPU. For each of the query types (\\(Q_{\\mathrm{sem}}\\), \\(Q_{\\mathrm{inst}}\\), \\(Q_{\\mathrm{obj}}\\), \\(Q_{\\mathrm{bg}}\\)) discussed in Sec. 3, we employ a learned query embedding, which is used when computing the \\(\\mathbf{Key}^T\\) Query affinity matrix for multi-head attention inside the decoder. We refer to the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.265, + 0.658 + ], + "angle": 0, + "content": "4.2. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.713 + ], + "angle": 0, + "content": "All results are computed with a single, jointly trained model which performs different tasks by simply providing the corresponding query set at run-time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Video Instance Segmentation (VIS). We evaluate on (1) YouTube-VIS 2021 [72] which covers 40 object classes and contains 2985/421 videos for training/validation, and (2) OVIS [54] which covers 25 object classes. It contains 607/140 videos for training/validation which are comparatively longer and more occluded. The AP scores for both are reported in Tab. 1. For all three backbones, TarViS achieves state-of-the-art results for both benchmarks even though other methods are trained separately per benchmark whereas we use a single model. On YouTube-VIS, TarViS achieves 48.3 AP with a ResNet-50 backbone compared to the 45.7 achieved by VITA [26]. With Swin-L, we achieve" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.894, + 0.584 + ], + "angle": 0, + "content": "60.2 AP which is also higher than the 57.5 by VITA. On OVIS with ResNet-50, our 31.1 AP is higher than the 30.2 for IDOL [69], and with Swin-L, TarViS (43.2 AP) outperforms the current state-of-the-art IDOL (42.6 AP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.893, + 0.836 + ], + "angle": 0, + "content": "Video Panoptic Segmentation (VPS). We evaluate VPS on three datasets: (1) KITTI-STEP [66], which contains 12/9 lengthy driving scene videos for training/validation with 19 semantic classes (2 thing and 17 stuff classes), (2) CityscapesVPS [35], which contains 50 short driving scene clips, each with 6 annotated frames, and (3) VIPSeg [44], which is a larger dataset with 2806/343 in-the-wild videos for training/validation and 124 semantic classes. The results are reported in Tab. 2. For KITTI-STEP, TarViS achieves \\(70\\%\\) STQ with a ResNet-50 backbone which is better than all existing approaches. The performance further improves to \\(72\\%\\) with Swin-L. For CityscapesVPS, TarViS achieves 58.9 VPQ which is higher than all other methods except VIP-Deeplab [55] (63.1). However, VIP-Deeplab performs monocular depth estimation for additional guidance, and therefore requires ground-truth depth-maps for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "For VIPSeq, TarViS outperforms existing approaches by a significant margin. With a ResNet-50 backbone, our 33.5 VPQ is \\(10.6\\%\\) higher than the 22.9 by Clip-PanoFCN [44]. With a Swin-Large backbone, TarViS achieves 48.0 VPQ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18743" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.089, + 0.468, + 0.118 + ], + "angle": 0, + "content": "Table 3. Results for VOS on DAVIS [53] and PET on BURST [3]. Detailed PET metrics are provided in supplementary." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.123, + 0.47, + 0.351 + ], + "angle": 0, + "content": "
MethodDAVIS (VOS)BURST (PET)
J&FJFHval allHtest all
UNICORN* [71]70.666.175.0--
HODOR [2]81.378.483.9--
STM [50]81.879.284.3--
CFBI [73]81.979.184.6--
HMMN [60]84.781.987.5--
AOT [74]84.982.387.5--
STCN [17]85.482.288.6--
XMem [15]86.282.989.5--
Box Tracker [32]---12.710.1
STCN+M2F [13, 17]---24.424.9
TarViS (R-50)82.679.385.930.932.1
TarViS (Swin-T)82.879.686.036.036.4
TarViS (Swin-L)85.381.788.537.536.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.47, + 0.482 + ], + "angle": 0, + "content": "which is more than double that of Clip-PanoFCN (22.9). Note that VIP-Deeplab performs significantly worse for VIPSeg (16.0 VPQ), showing that TarViS generalizes better across benchmarks. Finally, we note that larger backbones results in significant performance gains for datasets with in-the-wild internet videos as in VIPSeg, but for specialized driving scene datasets (e.g. KITTI-STEP and Cityscapes-VPS), the improvements are much smaller." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Video Object Segmentation (VOS). We evaluate VOS on the DAVIS 2017 [53] dataset, which contains 60/30 YouTube videos for training/validation. The results in Tab. 3 show that TarViS achieves 85.3 \\(\\mathcal{J}\\&\\mathcal{F}\\) which is higher than all existing methods except STCN [17] (85.4) and XMem [15] (86.2). As mentioned in Sec. 3.3, encoding objects as queries incurs a loss of fine-grained information, which is detrimental to performance. On the other hand, space-time correspondence (STC) based approaches learn pixel-to-pixel affinities between frames, which enables them to propagate fine-grained object appearance information. We note, however, that TarViS is the first method not based on the STC paradigm which achieves this level is performance (85.3 \\(\\mathcal{J}\\&\\mathcal{F}\\)), outperforming several STC-based methods as well as all non-STC based methods e.g. HODOR [2] (81.5) and UNICORN [71] (70.6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Point Exemplar-guided Tracking (PET). PET is evaluated on the recently introduced BURST benchmark [3] which contains 500/1000/1500 diverse videos for training/validation/testing. It is a constrained version of VOS which only provides the point coordinates of the object mask centroid instead of the full mask. Tab. 3 shows that existing methods can only tackle either VOS or PET. To verify this, we tried adapting STCN [17] for PET by training it with point masks, but the training did not converge. By contrast, TarViS encodes objects into queries, which enables it to tackle both tasks with a single model since the" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 4. Qualitative results from a single TarViS model for all four tasks. Further results are shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.373 + ], + "angle": 0, + "content": "object guidance (point or mask) is abstracted behind the EncodeObjects(·) function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.466 + ], + "angle": 0, + "content": "TarViS achieves a \\(\\mathrm{HOTA_{all}}\\) score of 37.5 and 36.4 on the validation and test sets, respectively, which is significantly better than the 24.4 and 24.9 achieved by the best performing baseline method which casts PET as a VOS problem by regressing a pseudo-ground-truth mask from the given point, followed by applying a VOS approach (STCN [17])." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.475, + 0.611, + 0.489 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.498, + 0.869, + 0.513 + ], + "angle": 0, + "content": "Table 4 shows several architecture/training ablations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Task-specific Training (row 1-3). The first three rows show results for task-specific models. We train a single model for VOS and PET since both tasks are closely related. We note that the VIS-only model performs worse than the multi-task model on YouTube-VIS (46.3 vs. 48.3) but slightly better on OVIS (31.5 vs. 31.1). For VPS, the performance on KITTI-STEP is unchanged, but Cityscapes-VPS and VIPSeg both show improvements with the multi-task model. Lastly, for VOS the task-specific model performs slightly worse on DAVIS (81.1 vs. 82.0) but significantly better on BURST for PET (34.7 vs. 30.9). To summarize, the final, multi-task model performs better on 4/7 benchmarks, worse on 2/7, and matches performance on 1/7 when compared to task-specific models. We thus conclude that the combination of multi-task supervision and more data is generally beneficial for performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Semantic Queries for VIS (row 4). TarViS represents object classes as dynamic query inputs to the network (\\(Q_{\\mathrm{sem}}\\), Sec. 3.1). We ablate this by modifying our network to work for only VIS by discarding the semantic/background queries and adopting a technique similar to existing methods [13, 65], i.e. using instance queries \\(Q_{\\mathrm{inst}}\\) in conjunction with a linear layer for classification (separate for each dataset). Comparing the results with the VIS-only setting which is trained on similar data, we see that this architecture" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18744" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.089, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation experiment results with ResNet-50 backbone. C-VPS: CityscapesVPS, YTVIS: YouTube-VIS, KITTI: KITTI-STEP." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.109, + 0.9, + 0.267 + ], + "angle": 0, + "content": "
SettingVideo Training DataVISVPSVOSPET
YTVISOVISKITTIC-VPSVIPSegYTVIS (mAP)OVIS (mAP)KITTI (STQ)C-VPS (VPQ)VIPSeg (VPQ)DAVIS (J&F)BURST (HOTAall)
1. VIS✓✓46.331.5-----
2. VPS✓✓✓--0.7049.732.4--
3. VOS + PET✓✓----81.134.7
4. No Semantic Queries✓✓44.729.8-----
5. No Temporal Neck✓✓✓✓✓42.822.30.6951.228.978.730.3
Final✓✓✓✓✓48.331.10.7053.333.582.030.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.352 + ], + "angle": 0, + "content": "performs worse than the VIS-only setting on both YouTube-VIS (44.7 vs. 46.3) and OVIS (29.8 vs. 31.5). Thus, our semantic query based classification makes the network architecture task-agnostic and also yields better performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Temporal Neck (row 4). We validate our novel Temporal Neck (Sec. 3.4) by training a model with a simpler neck that contains only Deformable Attention layers [78], similar to Mask2Former [13], i.e. there is no feature interaction across frames. Doing this degrades performance across all datasets, with particularly large drops for YouTube-VIS (42.8 vs. 48.3) and OVIS (22.3 vs. 31.1). This shows that inter-frame feature interactions enabled by our Temporal Neck are highly beneficial for down-stream tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.192, + 0.525 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Limitations. Training on multiple datasets/tasks does not necessarily improve performance on all benchmarks. For VOS, the model exhibits class bias and sometimes fails to track unusual objects which were not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Future Outlook. We jointly trained TarViS for four different tasks to validate its generalization capability. The architecture can, however, tackle any video segmentation task for which the targets can be encoded as queries. The recent emergence of joint language- vision models [41, 56, 57] thus makes it possible to perform multi-object segmentation based on a text prompt if the latter can be encoded as a target query using a language encoder [19]. Another interesting possibility is that TarViS could be applied to multiple tasks in the same forward pass by simply concatenating the task-specific queries. Fig. 5 offers a promising outlook for this; it shows our model's output for a video clip from a popular TV series where we perform VIS and VOS simultaneously by providing the semantic query for the 'person' class (from YouTube-VIS [72]), and the VOS-based object queries for the dragon by annotating its first frame mask, i.e. \\( Q_{\\mathrm{in}} = \\mathrm{concat}(Q_{\\mathrm{sem}}, Q_{\\mathrm{inst}}, Q_{\\mathrm{obj}}, Q_{\\mathrm{bg}}) \\). TarViS successfully segments all four persons in the scene (VIS) and the dragon (VOS), even though our model was never trained to simultaneously tackle both tasks in a single forward pass." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.289, + 0.697, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.289, + 0.892, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.506 + ], + "angle": 0, + "content": "Figure 5. TarViS performing VIS and VOS in a single forward pass. We provide the mask for the dragon on the left, and the semantic query for the 'person' class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.619, + 0.53 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.767 + ], + "angle": 0, + "content": "We presented TarViS: a novel, unified approach for tackling any task requiring pixel-precise segmentation of a set of targets in video. We adopt a generalized paradigm where the task-specific targets are encoded into a set of queries which are then input to our network together with the video features. The network is trained to produce segmentation masks for each target entity, but is inherently agnostic to the task-specific definition of these targets. To demonstrate the effectiveness of our approach, we applied it to four different video segmentation tasks (VIS, VPS, VOS, PET). We showed that a single TarViS model can be jointly trained for all tasks, and during inference can hot-swap between tasks without any task-specific fine-tuning. Our model achieved state-of-the-art performance on five benchmarks and has multiple, promising directions for future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.892, + 0.893 + ], + "angle": 0, + "content": "Acknowledgments. This project was partially funded by ERC CoG DeeVise (ERC-2017-COG-773161) and BMBF project NeuroSys-D (03ZU1106DA). Compute resources were granted by RWTH Aachen under project ID supp0003, and by the Gauss Centre for Supercomputing e.V. through the John von Neumann Institute for Computing on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18745" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Arxiv, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Ali Athar, Jonathon Luiten, Alexander Hermans, Deva Ramanan, and Bastian Leibe. Hodor: High-level object descriptors for object re-segmentation in video learned from static images. In CVPR, 2022. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[4] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[5] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. A single-stage, bottom-up approach for occluded vis using spatio-temporal embeddings. In ICCV-W, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[6] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[8] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In ECCV, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[9] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[11] Richard Caruana. Multitask Learning: A Knowledge-Based Source of Inductive Bias. In ICML, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In Arxiv, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In CVPR, 2022. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[14] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[15] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[16] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[20] Ahmed Elgammal, David Harwood, and Larry Davis. Nonparametric model for background subtraction. In ECCV, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[21] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. Moving obstacle detection in highly dynamic scenes. In ICRA, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.351, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[22] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.409, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[23] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[24] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-Task Self-Training for Learning General Representations. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[25] Helmut Grabner, Michael Grabner, and Horst Bischof. Real-time tracking via on-line boosting. In BMVC, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[26] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.57, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[27] Namdar Homayounfar, Justin Liang, Wei-Chiu Ma, and Raquel Urtasun. Videoclick: Video object segmentation with a single click. Arxiv, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[28] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[29] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. *ICLR*, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[31] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[32] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18746" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[34] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[35] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[36] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[37] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[38] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. Uvim: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[39] Dieter Koller, Joseph Weber, and Jitendra Malik. Robust multiple car tracking with occlusion reasoning. In ECCV, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[40] Bastian Leibe, Konrad Schindler, Nico Cornelis, and Luc Van Gool. Coupled object detection and tracking from static cameras and moving vehicles. PAMI, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[42] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[44] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[45] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. In arXiv:1603.00831, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[46] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[47] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[48] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[49] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. PAMI, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[50] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[51] Kilian Pfeiffer, Alexander Hermans, István Sárándi, Mark Weber, and Bastian Leibe. Visual person understanding through multi-task and multi-dataset learning. In GCPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[52] Hamed Pirsiavash, Deva Ramanan, and Charless C Fowlkes. Globally-optimal greedy algorithms for tracking a variable number of objects. In CVPR, pages 1201-1208. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[53] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. In *Arxiv*, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[54] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. In IJCV, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[55] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[58] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. In *Arxiv*, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[59] Sebastian Ruder. An Overview of Multi-Task Learning in Deep Neural Networks. arXiv:1706.05098, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[60] Hongje Seong, Seoung Wug Oh, Joon-Young Lee, Seongwon Lee, Suhyeon Lee, and Euntai Kim. Hierarchical memory matching network for video object segmentation. In ICCV, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[61] C. Stauffer and W.E.L. Grimson. Adaptive background mixture models for real-time tracking. In CVPR, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[63] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmentation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[64] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[65] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[66] Mark Weber, Jun Xie, Maxwell Collins, Yukun Zhu, Paul Voigtaender, Hartwig Adam, Bradley Green, Andreas" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18747" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Geiger, Bastian Leibe, Daniel Cremers, Aljosa Osep, Laura Leal-Taixe, and Liang-Chieh Chen. STEP: Segmenting and tracking every pixel. In NeurIPS, 2021. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[67] Sanghyun Woo, Dahun Kim, Joon-Young Lee, and In So Kweon. Learning to associate every segment for video panoptic segmentation. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[68] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[69] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[70] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[71] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[72] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[73] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[74] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.561 + ], + "angle": 0, + "content": "[76] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.564, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[77] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICML, 2020. 2, 3, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18748" + } + ] +] \ No newline at end of file diff --git a/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_origin.pdf b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2b0af6b4527db167a14a603508bc31f3f5e8d97 --- /dev/null +++ b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/7480a5f9-95b4-4d94-9341-e301cfb19950_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b344cd12f1aac3c4e8154a7abce1c7564f2f37cbd1e35d41c5dd2f22e6f8be62 +size 8033737 diff --git a/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/full.md b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..43ede14d1c25b69591cb1ad8f036046ecca3846b --- /dev/null +++ b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/full.md @@ -0,0 +1,279 @@ +# TarViS: A Unified Approach for Target-based Video Segmentation + +Ali Athar $^{1}$ Alexander Hermans $^{1}$ Jonathon Luiten $^{1,2}$ Deva Ramanan $^{2}$ Bastian Leibe $^{1}$ $^{1}$ RWTH Aachen University, Germany $^{2}$ Carnegie Mellon University, USA +{athar, hermans, luiten, leibe}@vision.rwth-aachen.de deva@cs.cmu.edu + +# Abstract + +The general domain of video segmentation is currently fragmented into different tasks spanning multiple benchmarks. Despite rapid progress in the state-of-the-art, current methods are overwhelmingly task-specific and cannot conceptually generalize to other tasks. Inspired by recent approaches with multi-task capability, we propose TarViS: a novel, unified network architecture that can be applied to any task that requires segmenting a set of arbitrarily defined 'targets' in video. Our approach is flexible with respect to how tasks define these targets, since it models the latter as abstract 'queries' which are then used to predict pixel-precise target masks. A single TarViS model can be trained jointly on a collection of datasets spanning different tasks, and can hot-swap between tasks during inference without any task-specific retraining. To demonstrate its effectiveness, we apply TarViS to four different tasks, namely Video Instance Segmentation (VIS), Video Panoptic Segmentation (VPS), Video Object Segmentation (VOS) and Point Exemplar-guided Tracking (PET). Our unified, jointly trained model achieves state-of-the-art performance on 5/7 benchmarks spanning these four tasks, and competitive performance on the remaining two. Code and model weights are available at: https://github.com/Ali2500/TarViS + +# 1. Introduction + +The ability to understand video scenes has been a longstanding goal of computer vision research because of wideranging applications in intelligent vehicles and robots. Early approaches tackled simpler tasks involving contour-based [33, 39] and box-level tracking [21, 25, 40, 52], background subtraction [20, 61], and motion segmentation [8, 49]. The deep learning boom then revolutionized the landscape by enabling methods to perform pixel-precise segmentation on challenging, real-world videos. In the past few years, a number of benchmarks have emerged, which evaluate how well methods can perform video segmentation according to various task formulations. Over time, these tasks/benchmarks have ballooned into separate re + +![](images/acfefb23785a7de608b92173bfe6d504186c65a7e93cee1b2da1e70280c49b67.jpg) +Figure 1. Predicted results from a jointly trained TarViS model for four different video segmentation tasks. + +search sub-communities. Although existing methods are rapidly improving the state-of-the-art for these benchmarks, each of them typically tackles only one narrowly-defined task, and generalizing them is non-trivial since the task definition is baked into the core approach. + +We argue that this fragmentation is unnecessary because video target segmentation tasks all require the same high-level capability, namely that of identifying, localizing and tracking rich semantic concepts. Meanwhile, recent progress on Transformer networks has enabled the wider AI research community to move towards unified, multi-task architectures [1, 30, 31, 38, 58], because the attention operation [62] is well-suited for processing feature sets with arbitrary structure and data modality. These developments give us the opportunity to unify the fractured landscape of target-based video segmentation. In this paper, we propose TarViS: a novel architecture which enables a single, unified model to be jointly trained for multiple video segmentation tasks. During inference, the same model can perform differ + +ent tasks at runtime by specifying the segmentation target. + +The core idea is that TarViS tackles the generic task of segmenting a set of arbitrary targets in video (defined as semantic classes or as specific objects). These targets are encoded as queries which, together with the video features, are input to a Transformer-based model. The model iteratively refines these queries and produces a pixel-precise mask for each target entity. This formulation conceptually fuses all video segmentation tasks [3, 54, 66, 72] which fall under the umbrella of the above-mentioned generic task, because they differ only in how the targets are defined. During both training and inference, TarViS can hot-swap between tasks at run-time by providing the desired target query set. + +To demonstrate our generalization capability, we tackle four different tasks: (1) Video Instance Segmentation (VIS) [54, 72], (2) Video Panoptic Segmentation (VPS) [35], (3) Video Object Segmentation [53], and (4) Point Exemplar-guided Tracking [3] (PET). For VIS, the segmentation targets are all objects in the video belonging to a predefined set of classes. The target set for VPS includes that for VIS, and additionally, a set of non-instantiable stuff semantic classes. For VOS, the targets are a specific set of objects for which the first-frame ground-truth mask is provided. PET is a more constrained version of VOS which only provides the location of a single point inside the object, rather than the full object mask. + +Existing methods for these tasks lack generalization capability because task-specific assumptions are typically baked into the approach (see Sec. 2 and 3 for details). In contrast, TarViS can tackle all four tasks with a unified model because we encode the task-specific targets as a set of queries, thus decoupling the network architecture from the task definition. Moreover, our approach can theoretically generalize further, e.g., one could potentially define the target set as all objects described by a given text prompt, though this is beyond the scope of this paper. + +To summarize, our contributions are as follows: we propose TarViS, a novel architecture that can perform any task requiring segmentation of a set of targets from video. For the first time, we are able to jointly train and infer a single model on a collection of datasets spanning the four aforementioned tasks (VIS, VPS, VOS, PET). Our experimental results show that TarViS performs competitively for VOS, and achieves state-of-the-art results for VIS, VPS and PET. + +# 2. Related Work + +Multi-task Models. Multi-task learning has a long history [11] with several architectures and training strategies [24,36,37,51,59,76]. Earlier approaches mostly consist of a shared backbone with fixed task-specific heads, whereas we design a more general architecture for video segmentation with task-specific targets to specify what to + +segment. Our approach is inspired by recent attention-based models, e.g., PerceiverIO [30, 31], which can be trained on diverse data modalities and task-specific heads are replaced with output queries. UViM [38] follows a similar direction by creating a unified architecture for diverse dense prediction tasks. However, both of these models are trained separately for different tasks. Recent, powerful multi-task vision language models such as Flamingo [1] and GATO [58] tackle a multitude of tasks by requiring a sequence of task-specific input-output examples to prime the model. This is conceptually similar to our task-specific targets, however, our model does not require per-task priming. Moreover, our targets are not modeled as sequence prompts, and we aim for a video segmentation model which is several orders of magnitude smaller. In the realm of video tracking and segmentation, the recently proposed UNICORN [71] model tackles multiple object tracking-related tasks with a unified architecture. Unlike TarViS, however, UNICORN follows the task-specific output head approach and is generally oriented towards box-level tracking tasks [22, 45, 47, 75], thus requiring non-trivial modifications to tackle VPS or PET. + +Query-based Transformer Architectures. Several works [2, 10, 13, 30, 31, 46, 65, 78] use query-based Transformer architectures for various tasks. The workhorse for task learning here is the iterative application of self- and cross-attention, where a set of query vectors (e.g., representing objects) are refined by interacting with each other, and with the input data (e.g., an image). Unlike existing methods which use queries in a task-specific context, TarViS adopts a query-based Transformer architecture in which the queries serve as a mechanism for decoupling the task definition from the architecture, i.e., our model can learn to tackle different tasks while being agnostic to their definition because the latter is abstracted behind a set of queries. + +Task-specific Video Segmentation. Current Video Instance Segmentation (VIS) methods broadly work by predicting object tracks in the video, followed by classification into a pre-defined set of categories. Several approaches [6,9,23,28,34,42,54,64,69,72] are based on the tracking-by-detection paradigm, some model video as a joint spatiotemporal volume [4,5], whereas recent works [12,26,29,65, 68] adopt Transformer-based architectures. + +For Video Panoptic Segmentation (VPS), methods [35, 55, 66] generally extend image-level panoptic approaches [14] by employing multi-head architectures for semantic segmentation and instance mask regression, classification, and temporal association. In the Video Object Segmentation (VOS) community, state-of-the-art methods are broadly based on the seminal work of Oh et al. [50], which learns space-time correspondences between pixels in different video frames, and then uses these to propagate the first-frame masks across the video. Subsequent + +![](images/5a444117723d665becfbbab1396b0a327d3e8a6f8f53b46db883986d98e33859.jpg) +Figure 2. TarViS Architecture. Segmentation targets for different tasks are represented by a set of abstract target queries $Q_{\mathrm{in}}$ . The core network (in green) is agnostic to the task definitions. The inner product between the output queries $Q_{\mathrm{out}}$ and video feature $F_{4}$ yields segmentation masks as required by the task. + +![](images/c4f063716734fedb84b6c65cf46ab1dca5aea4ecb26d9a27bf07ea842d90657b.jpg) + +methods [15-17, 60, 63, 70, 73, 73, 74] have significantly improved the performance and efficiency of this approach. Point Exemplar-guided Tracking (PET) [3, 27] is a fairly new task for which the current best approach [3] regresses a pseudo-ground-truth mask from the given point coordinates, and then applies a VOS method [17] to this mask. + +The above methods thus incorporate task-specific assumptions into their core approach. This can benefit pertask performance, but makes it difficult for them to generalize across tasks. By contrast, TarViS can tackle all four aforementioned tasks, and generally any target-based video segmentation task, with a single, jointly trained model. + +# 3. Method + +TarViS can segment arbitrary targets in video since the architecture is flexible with respect to how these targets are defined, thus enabling us to conceptually unify and jointly tackle the four aforementioned tasks (VIS, VPS, VOS, PET). The architecture is illustrated in Fig. 2. + +For all tasks, the common input to the network is an RGB video clip of length $T$ denoted by $V \in \mathbb{R}^{H \times W \times T \times 3}$ . This is input to a 2D backbone network which produces image-level feature maps, followed by a Temporal Neck, which enables feature interaction across time and outputs a set of temporally consistent, multi-scale, $D$ -dimensional feature maps $\mathcal{F} = \{F_{32}, F_{16}, F_{8}, F_{4}\}$ where $F_{s} \in \mathbb{R}^{\frac{H}{s} \times \frac{W}{s} \times T \times D}$ . The feature maps $\mathcal{F}$ are then fed to our Transformer decoder, together with a set of queries $Q_{\mathrm{in}}$ which represent the segmentation targets. The decoder applies successive layers of self- and masked cross-attention wherein the queries are iteratively refined by attending to each other, and to the feature maps, respectively. The refined queries output by the decoder are denoted with $Q_{\mathrm{out}}$ . The following subsections explain how TarViS tackles each task in detail. + +# 3.1. Video Instance Segmentation + +VIS defines the segmentation target set as all objects belonging to a set of predefined classes. Accordingly, the input query set $Q_{\mathrm{in}}$ for VIS contains three types of queries: (1) semantic queries denoted by $Q_{\mathrm{sem}} \in \mathbb{R}^{C \times D}$ where $C$ is the number of classes defined by the dataset, i.e., each $D$ -dimensional vector in $Q_{\mathrm{sem}}$ represents a particular semantic class. (2) instance queries denoted by $Q_{\mathrm{inst}} \in \mathbb{R}^{I \times D}$ where $I$ is assumed to be an upper bound on the number of instances in the video clip, and (3) a background query denoted by $Q_{\mathrm{bg}} \in \mathbb{R}^{1 \times D}$ to capture inactive instance queries. + +The three query sets are concatenated, i.e., $Q_{\mathrm{in}} = \operatorname{concat}(Q_{\mathrm{sem}}, Q_{\mathrm{inst}}, Q_{\mathrm{bg}})$ , and input to the Transformer decoder, which refines their feature representation through successive attention layers and outputs a set of queries $Q_{\mathrm{out}} = \operatorname{concat}(Q_{\mathrm{sem}}', Q_{\mathrm{inst}}', Q_{\mathrm{bg}}')$ . These are then used to produce temporally consistent instance mask logits by computing the inner product $\langle F_4, Q_{\mathrm{inst}}' \rangle \in \mathbb{R}^{H \times W \times T \times I}$ . To obtain classification logits, we compute the inner product $\langle Q_{\mathrm{inst}}', \operatorname{concat}(Q_{\mathrm{sem}}', Q_{\mathrm{bg}}') \rangle \in \mathbb{R}^{I \times (C + 1)}$ . + +The three types of queries are initialized randomly at the start of training and optimized thereafter. The instance queries $Q_{\mathrm{inst}}$ enable us to segment a varying number of objects from the input clip. During training, we apply Hungarian matching between the predicted and ground-truth instance masks to assign instance queries to video instances, and then supervise their predicted masks and classification logits accordingly. When training on multiple datasets with heterogeneous classes, the semantic query sets are separately initialized per dataset, but $Q_{\mathrm{inst}}$ and $Q_{\mathrm{bg}}$ are shared. + +Comparison to Instance Segmentation Methods. Several Transformer-based methods [10, 13, 65, 78] for image/video instance segmentation also use queries to seg + +ment a variable number of input instances. The key difference to our approach is the handling of object classes: existing works employ only instance queries which are input to a fully-connected layer with a fan-out of $C + 1$ to obtain classification (and background) scores. The notion of class-guided instance segmentation is thus baked into the approach. By contrast, TarViS is agnostic to the task-specific notion of object classes because it models them as arbitrary queries which are dynamic inputs to the network. The semantic representation for these queries is thus decoupled from the core architecture and is only learned via loss supervision. An important enabler for this approach is the background query $Q_{\mathrm{bg}}$ , which serves as a 'catch-all' class to represent everything that is not in $Q_{\mathrm{sem}}$ . It is used to classify non-active instance queries, and its mask logits are supervised to segment all non-object input pixels. + +# 3.2. Video Panoptic Segmentation + +VPS defines the segmentation targets as all objects belonging to a set of thing classes (e.g., 'person', 'car'), and additionally, a set of non-instantiable stuff classes (e.g., 'sky', 'grass') which cover all non-object pixels. TarViS can tackle VPS with virtually no modification to the workflow in Sec. 3.1. We can compute semantic segmentation masks for the input clip by simply taking the inner product between $Q_{\mathrm{sem}}$ and the video features: $\langle F_4, Q_{\mathrm{sem}}' \rangle \in \mathbb{R}^{H \times W \times T \times C}$ . Note that here, $Q_{\mathrm{sem}}$ contains queries representing both thing and stuff classes. + +Comparison to VPS Methods. Current VPS datasets [35, 66] involve driving scene videos captured from moving vehicles. Methods tackling this task [35, 55] are based on earlier image panoptic segmentation approaches [14] which involve multi-head networks for semantic and instance segmentation prediction. In terms of image-level panoptic segmentation, Mask2Former [13] uses a Transformer-based architecture, but it models stuff classes as instances which are Hungarian-matched to the ground-truth target during training, whereas TarViS models semantic classes and instances using separate, designated queries. + +# 3.3. Video Object Segmentation and Point Exemplar-guided Tracking + +VOS and PET are instantiations of a general task where the segmentation targets are a set of $O$ objects for which some ground-truth cue $\mathcal{G}$ is given. For VOS, $\mathcal{G}$ is provided as the first-frame object masks $M_{\mathrm{obj}} \in \mathbb{R}^{O \times H \times W}$ , whereas for PET, $\mathcal{G}$ is provided as the $(x,y)$ coordinates $P_{\mathrm{obj}} \in \mathbb{R}^{O \times 2}$ of a point inside each of the objects. TarViS jointly tackles both tasks by adopting a generalized approach in which the $O$ target objects are encoded into a set of object queries $Q_{\mathrm{obj}}$ . Thus, both VOS and PET boil down to designing a function Encode0objects(\cdot) which regresses $Q_{\mathrm{obj}}$ from the ground-truth cues $\mathcal{G}$ and feature maps $\mathcal{F}$ : + +$$ +Q _ {\mathrm {o b j}} \leftarrow \operatorname {E n c o d e 0 b j e c t s} (\mathcal {G}, \mathcal {F}). \tag {1} +$$ + +Note that $Q_{\mathrm{obj}}$ is conceptually analogous to $Q_{\mathrm{sem}}$ and $Q_{\mathrm{inst}}$ used for VIS in that all three are abstract representations for their respective task-specific segmentation targets. + +Video Object Segmentation. We seek inspiration from HODOR [2] to implement EncodeObjects for VOS, a recent method for weakly-supervised VOS, which encodes objects into concise descriptors as follows: the descriptors are initialized by average pooling the image features inside the object masks, followed by iterative refinement where the descriptors attend to each other (self-attention) and to their respective soft-masked image features (cross-attention). + +For TarViS, we employ a lightweight Object Encoder with a similar workflow to encode the objects as a set of queries $Q_{\mathrm{obj}}$ , but with two differences to HODOR [2]: instead of cross-attending to the entire image feature map $(H \cdot W$ points) with soft-masked attention, we apply hard-masked cross-attention to at most $p_{\mathrm{max}}$ feature points per object, where $p_{\mathrm{max}} \ll H \cdot W$ . Object masks containing more than $p_{\mathrm{max}}$ points are sub-sampled accordingly. This significantly improves the memory/run-time overhead of our Object Encoder. Secondly, we note that the process of distilling object features into a single descriptor involves a loss of object appearance information, which degrades performance. We therefore model each object with $q_{o}$ queries (instead of one) by spatially dividing each object mask into $q_{o}$ segments, i.e., $Q_{\mathrm{obj}} \in \mathbb{R}^{O \times q_{o} \times D}$ (we use $q_{o} = 4$ ). + +In addition to $Q_{\mathrm{obj}}$ , we initialize a set of background queries $Q_{\mathrm{bg}} \in \mathbb{R}^{B \times D}$ to model the non-target pixels in the reference frame. Following HODOR [2], we employ multiple background queries, which are initialized dynamically by dividing the video frame containing the groundtruth masks $M_{\mathrm{obj}}$ into a $4 \times 4$ grid and average pooling the non-object pixels in each grid cell. The Object Encoder jointly refines the background and object queries to yield $Q_{\mathrm{in}} = \text{concat}(Q_{\mathrm{obj}}, Q_{\mathrm{bg}})$ . During training, the mask logits for the multiple background queries are aggregated per-pixel by applying $\max(\cdot)$ and supervised to segment all pixels not part of the target object set. + +The remaining workflow follows that for VIS and VPS: $Q_{\mathrm{in}}$ is input to the Transformer decoder together with the video features $\mathcal{F}$ . The refined output query set $Q_{\mathrm{out}} = \operatorname{concat}(Q_{\mathrm{obj}}', Q_{\mathrm{bg}}')$ is then used to compute the inner product $\langle F_4, Q_{\mathrm{obj}}' \rangle \in \mathbb{R}^{H \times W \times T \times O \times q_o}$ . Subsequently, $\max(\cdot)$ is applied on the $q_o$ -sized dimension to obtain the final mask logits for the $O$ target objects. + +Point Exemplar-guided Tracking. For PET we implement Encode0objects in the exactly same way as VOS: the given point coordinates $P_{\mathrm{obj}}$ are converted into a mask with just one non-zero pixel, followed by iterative refinement by the Object Encoder (with shared weights for VOS and PET). + +The only difference is that here we represent each of the $O$ objects with just one query, i.e., $Q_{\mathrm{obj}} \in \mathbb{R}^{O \times D}$ ( $q_o = 1$ ). The subsequent workflow is also identical to that for VOS: the queries are refined by the Transformer decoder followed by an inner product with $F_4$ to obtain object mask logits. + +Comparison to VOS and PET Methods. Current state-of-the-art VOS methods are largely based on STM [50]. It involves learning pixel-to-pixel correspondences across video frames, which are then used to propagate the given object mask across the video. This approach is effective since every pixel in the given mask can be individually mapped to future frames, thus preserving fine-grained object details. The core approach is, however, task-specific since it assumes the availability of first-frame object masks, and does not generalize to the PET (see Sec. 4.2). PET can be viewed as a more constrained version of VOS, where only a single object point is provided instead of the full mask. Consequently, PET [3] is currently tackled by casting it as a VOS problem by using an image instance segmentation network [13] to regress pseudo-ground-truth object masks from the given point coordinates $P_{\mathrm{obj}}$ . + +On the other hand, our approach of encoding objects as concise queries causes loss of fine-grained object appearance information, but it has the advantage of being agnostic to how $\mathcal{G}$ is defined. As evident from the unified workflow for VOS and PET, any variation of these tasks with arbitrary ground-truth cues $\mathcal{G}$ can be seamlessly fused into our architecture as long as we can implement an effective Encode0bjects function to regress $Q_{\mathrm{obj}}$ from the given $\mathcal{G}$ . + +# 3.4. Network Architecture + +Temporal Neck. TarViS produces target masks by computing the inner product between $Q_{\mathrm{out}}$ and the video feature map $F_{4}$ . For this to work, the per-pixel features $\mathcal{F}$ must be aligned for the same, and dissimilar for different targets. Some image instance segmentation methods [13, 78] apply Deformable Attention [78] to the backbone feature maps to efficiently learn multi-scale image features. For TarViS, however, the features must also be temporally consistent across the entire input video clip. To achieve this, we propose a novel Temporal Neck architecture inspired from the work of Bertasius et al. [7] for video action classification. We enable efficient spatio-temporal feature interaction by applying two types of self-attention in an alternating fashion: the first is spatially global and temporally localized, whereas the second is spatially localized and temporally global. The first operation is implemented with Deformable Attention, following existing work [12, 78]. The second operation, Temporal Attention, involves dividing the input space-time volume into a grid along the spatial axes, and then applying self-attention to the space-time feature volume inside each grid cell. Both operations allow feature + +![](images/aacbac1eceba824eeabe1aa85079743ce7c28a8009455c1fca3a9c0ce70fbcdb.jpg) +Figure 3. Temporal Neck Layer. Colored regions denote the attention field w.r.t the selected pixel (darkened). Deformable Attention is spatially unrestricted but temporally limited to a single frame, whereas Temporal Attention is spatially localized, but temporally unrestricted. $F_{8}$ is inactive for the temporal attention. + +interaction across multiple scales. Both attention operations are illustrated in Fig. 3. We exclude $F_{8}$ from temporal attention since we found this to be more memory-efficient without negatively impacting prediction quality. + +Transformer Decoder. The decoder architecture follows that of Mask2Former [13]: the input queries are iteratively refined over multiple layers. In each layer, the queries first cross-attend to their respective masked video features, then self-attend to each other, followed by feed-forward layers. + +# 3.5. Inference + +To infer on videos with arbitrary length, we split videos into clips of length $T_{\mathrm{clip}}$ with an overlap of $T_{\mathrm{ov}}$ between successive clips. Object tracks are associated across clips based on their mask IoU in the overlapping frames. For the VOS tasks, the object queries for an intermediate clip are initialized by using the predicted masks in the overlapping frames from the previous clip as a pseudo-ground-truth. For VPS, we average the semantic segmentation logits in the overlapping frames. Our approach is thus near-online because the time delay in obtaining the output for a given frame is at most $T_{\mathrm{clip}} - T_{\mathrm{ov}} - 1$ (except for the first clip in the video). + +# 4. Experiments + +# 4.1. Implementation Details + +Our Temporal Neck contains 6 layers of Deformable and Temporal Attention. We pretrain for $500\mathrm{k}$ iterations on pseudo-video clips generated by applying on-the-fly augmentations to images from COCO [43], ADE20k [77], Mapillary [48] and Cityscapes [18]. The samples are either trained for VPS, VIS, VOS or PET. This is followed by fine-tuning for $90\mathrm{k}$ iterations jointly on samples from YouTube-VIS [72], OVIS [54], KITTI-STEP [66], + +Table 1. Results for Video Instance Segmentation (VIS) on the YouTube-VIS 2021 [72] and OVIS [54] validation sets. + +
MethodBackboneShared ModelYouTube-VIS 2021OVIS
APAP50AP75AR1AR10APAP50AP75AR1AR10
Mask2Former-VIS [12]R-50X40.660.941.8-------
IDOL [69]R-50X43.968.049.638.050.930.251.330.015.037.5
MinVIS [28]R-50X44.266.048.139.251.725.045.524.013.929.7
VITA [26]R-50X45.767.449.540.953.619.641.217.411.726.0
TarViSR-50✓48.369.653.240.555.931.152.530.415.939.9
Mask2Former-VIS [12]Swin-TX45.968.750.7-------
TarViSSwin-T✓50.971.656.642.257.234.055.034.416.140.9
IDOL [69]Swin-LX56.180.863.545.060.142.665.745.217.949.6
VITA [26]Swin-LX57.580.661.047.762.627.751.924.914.933.0
TarViSSwin-L✓60.281.467.647.664.843.267.844.618.050.4
+ +Table 2. Video Panoptic Segmentation (VPS) results for validation sets of KITTI-STEP [66], CityscapesVPS [35] and VIPseg [44]. + +
MethodShared ModelKITTI-STEPCityscapesVPSVIPSeg
STQAQSQVPQ\( VPQ^{Th} \)\( VPQ^{St} \)VPQ\( VPQ^{Th} \)\( VPQ^{St} \)STQ
Mask Propagation [66]✘0.670.630.71------
Track [35]✘---55.943.764.8---
VPSNet [35]✘0.560.520.6157.044.766.014.014.014.220.8
VPSNet-SiamTrack [67]✘---57.344.766.417.217.317.321.1
VIP-Deeplab [55]✘---63.149.573.016.012.318.222.0
Clip-PanoFCN [44]✘------22.925.020.831.5
TarViS (R-50)✓0.700.700.6953.335.966.033.539.228.543.1
TarViS (Swin-T)✓0.710.710.7058.042.969.035.842.729.745.3
TarViS (Swin-L)✓0.720.720.7358.943.769.948.058.239.052.9
+ +CityscapesVPS [35], VIPSeg [44], DAVIS [53] and BURST [3]. We train on 32 Nvidia A100 GPUs with batch size 1 per GPU. For each of the query types ( $Q_{\mathrm{sem}}$ , $Q_{\mathrm{inst}}$ , $Q_{\mathrm{obj}}$ , $Q_{\mathrm{bg}}$ ) discussed in Sec. 3, we employ a learned query embedding, which is used when computing the $\mathbf{Key}^T$ Query affinity matrix for multi-head attention inside the decoder. We refer to the supplementary for more details. + +# 4.2. Benchmark Results + +All results are computed with a single, jointly trained model which performs different tasks by simply providing the corresponding query set at run-time. + +Video Instance Segmentation (VIS). We evaluate on (1) YouTube-VIS 2021 [72] which covers 40 object classes and contains 2985/421 videos for training/validation, and (2) OVIS [54] which covers 25 object classes. It contains 607/140 videos for training/validation which are comparatively longer and more occluded. The AP scores for both are reported in Tab. 1. For all three backbones, TarViS achieves state-of-the-art results for both benchmarks even though other methods are trained separately per benchmark whereas we use a single model. On YouTube-VIS, TarViS achieves 48.3 AP with a ResNet-50 backbone compared to the 45.7 achieved by VITA [26]. With Swin-L, we achieve + +60.2 AP which is also higher than the 57.5 by VITA. On OVIS with ResNet-50, our 31.1 AP is higher than the 30.2 for IDOL [69], and with Swin-L, TarViS (43.2 AP) outperforms the current state-of-the-art IDOL (42.6 AP). + +Video Panoptic Segmentation (VPS). We evaluate VPS on three datasets: (1) KITTI-STEP [66], which contains 12/9 lengthy driving scene videos for training/validation with 19 semantic classes (2 thing and 17 stuff classes), (2) CityscapesVPS [35], which contains 50 short driving scene clips, each with 6 annotated frames, and (3) VIPSeg [44], which is a larger dataset with 2806/343 in-the-wild videos for training/validation and 124 semantic classes. The results are reported in Tab. 2. For KITTI-STEP, TarViS achieves $70\%$ STQ with a ResNet-50 backbone which is better than all existing approaches. The performance further improves to $72\%$ with Swin-L. For CityscapesVPS, TarViS achieves 58.9 VPQ which is higher than all other methods except VIP-Deeplab [55] (63.1). However, VIP-Deeplab performs monocular depth estimation for additional guidance, and therefore requires ground-truth depth-maps for training. + +For VIPSeq, TarViS outperforms existing approaches by a significant margin. With a ResNet-50 backbone, our 33.5 VPQ is $10.6\%$ higher than the 22.9 by Clip-PanoFCN [44]. With a Swin-Large backbone, TarViS achieves 48.0 VPQ + +Table 3. Results for VOS on DAVIS [53] and PET on BURST [3]. Detailed PET metrics are provided in supplementary. + +
MethodDAVIS (VOS)BURST (PET)
J&FJFHval allHtest all
UNICORN* [71]70.666.175.0--
HODOR [2]81.378.483.9--
STM [50]81.879.284.3--
CFBI [73]81.979.184.6--
HMMN [60]84.781.987.5--
AOT [74]84.982.387.5--
STCN [17]85.482.288.6--
XMem [15]86.282.989.5--
Box Tracker [32]---12.710.1
STCN+M2F [13, 17]---24.424.9
TarViS (R-50)82.679.385.930.932.1
TarViS (Swin-T)82.879.686.036.036.4
TarViS (Swin-L)85.381.788.537.536.1
+ +which is more than double that of Clip-PanoFCN (22.9). Note that VIP-Deeplab performs significantly worse for VIPSeg (16.0 VPQ), showing that TarViS generalizes better across benchmarks. Finally, we note that larger backbones results in significant performance gains for datasets with in-the-wild internet videos as in VIPSeg, but for specialized driving scene datasets (e.g. KITTI-STEP and Cityscapes-VPS), the improvements are much smaller. + +Video Object Segmentation (VOS). We evaluate VOS on the DAVIS 2017 [53] dataset, which contains 60/30 YouTube videos for training/validation. The results in Tab. 3 show that TarViS achieves 85.3 $\mathcal{J}\&\mathcal{F}$ which is higher than all existing methods except STCN [17] (85.4) and XMem [15] (86.2). As mentioned in Sec. 3.3, encoding objects as queries incurs a loss of fine-grained information, which is detrimental to performance. On the other hand, space-time correspondence (STC) based approaches learn pixel-to-pixel affinities between frames, which enables them to propagate fine-grained object appearance information. We note, however, that TarViS is the first method not based on the STC paradigm which achieves this level is performance (85.3 $\mathcal{J}\&\mathcal{F}$ ), outperforming several STC-based methods as well as all non-STC based methods e.g. HODOR [2] (81.5) and UNICORN [71] (70.6). + +Point Exemplar-guided Tracking (PET). PET is evaluated on the recently introduced BURST benchmark [3] which contains 500/1000/1500 diverse videos for training/validation/testing. It is a constrained version of VOS which only provides the point coordinates of the object mask centroid instead of the full mask. Tab. 3 shows that existing methods can only tackle either VOS or PET. To verify this, we tried adapting STCN [17] for PET by training it with point masks, but the training did not converge. By contrast, TarViS encodes objects into queries, which enables it to tackle both tasks with a single model since the + +![](images/1ce3d544c2c49354fbb19480fcf5eb55b9185888dcdb959db88667655a571c8e.jpg) +Figure 4. Qualitative results from a single TarViS model for all four tasks. Further results are shown in the supplementary. + +object guidance (point or mask) is abstracted behind the EncodeObjects(·) function. + +TarViS achieves a $\mathrm{HOTA_{all}}$ score of 37.5 and 36.4 on the validation and test sets, respectively, which is significantly better than the 24.4 and 24.9 achieved by the best performing baseline method which casts PET as a VOS problem by regressing a pseudo-ground-truth mask from the given point, followed by applying a VOS approach (STCN [17]). + +# 4.3. Ablations + +Table 4 shows several architecture/training ablations. + +Task-specific Training (row 1-3). The first three rows show results for task-specific models. We train a single model for VOS and PET since both tasks are closely related. We note that the VIS-only model performs worse than the multi-task model on YouTube-VIS (46.3 vs. 48.3) but slightly better on OVIS (31.5 vs. 31.1). For VPS, the performance on KITTI-STEP is unchanged, but Cityscapes-VPS and VIPSeg both show improvements with the multi-task model. Lastly, for VOS the task-specific model performs slightly worse on DAVIS (81.1 vs. 82.0) but significantly better on BURST for PET (34.7 vs. 30.9). To summarize, the final, multi-task model performs better on 4/7 benchmarks, worse on 2/7, and matches performance on 1/7 when compared to task-specific models. We thus conclude that the combination of multi-task supervision and more data is generally beneficial for performance. + +Semantic Queries for VIS (row 4). TarViS represents object classes as dynamic query inputs to the network ( $Q_{\mathrm{sem}}$ , Sec. 3.1). We ablate this by modifying our network to work for only VIS by discarding the semantic/background queries and adopting a technique similar to existing methods [13, 65], i.e. using instance queries $Q_{\mathrm{inst}}$ in conjunction with a linear layer for classification (separate for each dataset). Comparing the results with the VIS-only setting which is trained on similar data, we see that this architecture + +Table 4. Ablation experiment results with ResNet-50 backbone. C-VPS: CityscapesVPS, YTVIS: YouTube-VIS, KITTI: KITTI-STEP. + +
SettingVideo Training DataVISVPSVOSPET
YTVISOVISKITTIC-VPSVIPSegYTVIS (mAP)OVIS (mAP)KITTI (STQ)C-VPS (VPQ)VIPSeg (VPQ)DAVIS (J&F)BURST (HOTAall)
1. VIS✓✓46.331.5-----
2. VPS✓✓✓--0.7049.732.4--
3. VOS + PET✓✓----81.134.7
4. No Semantic Queries✓✓44.729.8-----
5. No Temporal Neck✓✓✓✓✓42.822.30.6951.228.978.730.3
Final✓✓✓✓✓48.331.10.7053.333.582.030.9
+ +performs worse than the VIS-only setting on both YouTube-VIS (44.7 vs. 46.3) and OVIS (29.8 vs. 31.5). Thus, our semantic query based classification makes the network architecture task-agnostic and also yields better performance. + +Temporal Neck (row 4). We validate our novel Temporal Neck (Sec. 3.4) by training a model with a simpler neck that contains only Deformable Attention layers [78], similar to Mask2Former [13], i.e. there is no feature interaction across frames. Doing this degrades performance across all datasets, with particularly large drops for YouTube-VIS (42.8 vs. 48.3) and OVIS (22.3 vs. 31.1). This shows that inter-frame feature interactions enabled by our Temporal Neck are highly beneficial for down-stream tasks. + +# 5. Discussion + +Limitations. Training on multiple datasets/tasks does not necessarily improve performance on all benchmarks. For VOS, the model exhibits class bias and sometimes fails to track unusual objects which were not seen during training. + +Future Outlook. We jointly trained TarViS for four different tasks to validate its generalization capability. The architecture can, however, tackle any video segmentation task for which the targets can be encoded as queries. The recent emergence of joint language- vision models [41, 56, 57] thus makes it possible to perform multi-object segmentation based on a text prompt if the latter can be encoded as a target query using a language encoder [19]. Another interesting possibility is that TarViS could be applied to multiple tasks in the same forward pass by simply concatenating the task-specific queries. Fig. 5 offers a promising outlook for this; it shows our model's output for a video clip from a popular TV series where we perform VIS and VOS simultaneously by providing the semantic query for the 'person' class (from YouTube-VIS [72]), and the VOS-based object queries for the dragon by annotating its first frame mask, i.e. $Q_{\mathrm{in}} = \mathrm{concat}(Q_{\mathrm{sem}}, Q_{\mathrm{inst}}, Q_{\mathrm{obj}}, Q_{\mathrm{bg}})$ . TarViS successfully segments all four persons in the scene (VIS) and the dragon (VOS), even though our model was never trained to simultaneously tackle both tasks in a single forward pass. + +![](images/f8ad2840a6ba0111007f04d5c2bddaaa31b66f876f9140351c49f9d69ef9d146.jpg) +Figure 5. TarViS performing VIS and VOS in a single forward pass. We provide the mask for the dragon on the left, and the semantic query for the 'person' class. + +![](images/5b8ef09b0ebbf85a3cd99440f9cbc3b49e31cd67c2b0d55f3460e48badb96cea.jpg) + +# 6. Conclusion + +We presented TarViS: a novel, unified approach for tackling any task requiring pixel-precise segmentation of a set of targets in video. We adopt a generalized paradigm where the task-specific targets are encoded into a set of queries which are then input to our network together with the video features. The network is trained to produce segmentation masks for each target entity, but is inherently agnostic to the task-specific definition of these targets. To demonstrate the effectiveness of our approach, we applied it to four different video segmentation tasks (VIS, VPS, VOS, PET). We showed that a single TarViS model can be jointly trained for all tasks, and during inference can hot-swap between tasks without any task-specific fine-tuning. Our model achieved state-of-the-art performance on five benchmarks and has multiple, promising directions for future work. + +Acknowledgments. This project was partially funded by ERC CoG DeeVise (ERC-2017-COG-773161) and BMBF project NeuroSys-D (03ZU1106DA). Compute resources were granted by RWTH Aachen under project ID supp0003, and by the Gauss Centre for Supercomputing e.V. through the John von Neumann Institute for Computing on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Arxiv, 2022. 1, 2 +[2] Ali Athar, Jonathon Luiten, Alexander Hermans, Deva Ramanan, and Bastian Leibe. Hodor: High-level object descriptors for object re-segmentation in video learned from static images. In CVPR, 2022. 2, 4, 7 +[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2, 3, 5, 6, 7 +[4] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 2 +[5] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. A single-stage, bottom-up approach for occluded vis using spatio-temporal embeddings. In ICCV-W, 2021. 2 +[6] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2 +[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5 +[8] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In ECCV, 2010. 1 +[9] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2 +[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3 +[11] Richard Caruana. Multitask Learning: A Knowledge-Based Source of Inductive Bias. In ICML, 1993. 2 +[12] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In Arxiv, 2021. 2, 5, 6 +[13] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In CVPR, 2022. 2, 3, 4, 5, 7, 8 +[14] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2, 4 +[15] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3, 7 +[16] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 3 + +[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 3, 7 +[18] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5 +[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019. 8 +[20] Ahmed Elgammal, David Harwood, and Larry Davis. Nonparametric model for background subtraction. In ECCV, 2000. 1 +[21] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. Moving obstacle detection in highly dynamic scenes. In ICRA, 2009. 1 +[22] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2 +[23] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, 2021. 2 +[24] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-Task Self-Training for Learning General Representations. In CVPR, 2021. 2 +[25] Helmut Grabner, Michael Grabner, and Horst Bischof. Real-time tracking via on-line boosting. In BMVC, 2006. 1 +[26] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6 +[27] Namdar Homayounfar, Justin Liang, Wei-Chiu Ma, and Raquel Urtasun. Videoclick: Video object segmentation with a single click. Arxiv, 2021. 3 +[28] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6 +[29] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2 +[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. *ICLR*, 2022. 1, 2 +[31] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 1, 2 +[32] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020.7 +[33] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 1 + +[34] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. In NeurIPS, 2021. 2 +[35] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2, 4, 6 +[36] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 2 +[37] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2 +[38] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. Uvim: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1, 2 +[39] Dieter Koller, Joseph Weber, and Jitendra Malik. Robust multiple car tracking with occlusion reasoning. In ECCV, 1994. 1 +[40] Bastian Leibe, Konrad Schindler, Nico Cornelis, and Luc Van Gool. Coupled object detection and tracking from static cameras and moving vehicles. PAMI, 2008. 1 +[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 8 +[42] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, 2021. 2 +[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5 +[44] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 6 +[45] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. In arXiv:1603.00831, 2016. 2 +[46] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 2 +[47] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2 +[48] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5 +[49] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. PAMI, 2013. 1 +[50] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2, 5, 7 + +[51] Kilian Pfeiffer, Alexander Hermans, István Sárándi, Mark Weber, and Bastian Leibe. Visual person understanding through multi-task and multi-dataset learning. In GCPR, 2019. 2 +[52] Hamed Pirsiavash, Deva Ramanan, and Charless C Fowlkes. Globally-optimal greedy algorithms for tracking a variable number of objects. In CVPR, pages 1201-1208. IEEE, 2011. 1 +[53] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. In *Arxiv*, 2017. 2, 6, 7 +[54] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. In IJCV, 2022. 2, 5, 6 +[55] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, 2021. 2, 4, 6 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8 +[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 8 +[58] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. In *Arxiv*, 2022. 1, 2 +[59] Sebastian Ruder. An Overview of Multi-Task Learning in Deep Neural Networks. arXiv:1706.05098, 2017. 2 +[60] Hongje Seong, Seoung Wug Oh, Joon-Young Lee, Seongwon Lee, Suhyeon Lee, and Euntai Kim. Hierarchical memory matching network for video object segmentation. In ICCV, 2021. 3, 7 +[61] C. Stauffer and W.E.L. Grimson. Adaptive background mixture models for real-time tracking. In CVPR, 1999. 1 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1 +[63] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmentation. In CVPR, 2019. 3 +[64] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 2 +[65] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2, 3, 7 +[66] Mark Weber, Jun Xie, Maxwell Collins, Yukun Zhu, Paul Voigtaender, Hartwig Adam, Bradley Green, Andreas + +Geiger, Bastian Leibe, Daniel Cremers, Aljosa Osep, Laura Leal-Taixe, and Liang-Chieh Chen. STEP: Segmenting and tracking every pixel. In NeurIPS, 2021. 2, 4, 5, 6 +[67] Sanghyun Woo, Dahun Kim, Joon-Young Lee, and In So Kweon. Learning to associate every segment for video panoptic segmentation. In CVPR, 2021. 6 +[68] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 2 +[69] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6 +[70] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, 2021. 3 +[71] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2, 7 +[72] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5, 6, 8 +[73] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3, 7 +[74] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 3, 7 +[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2 +[76] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 2 +[77] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 5 +[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICML, 2020. 2, 3, 5, 8 \ No newline at end of file diff --git a/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/images.zip b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31edc5ad8ce87da5874bcd86d4f1adf01d03a355 --- /dev/null +++ b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4e7f3817feec85f33e48c6ea1ee1af34e9f966d1c5bdfe3c8e2f8d38f263a91 +size 539151 diff --git a/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/layout.json b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..44a1680ba1203312f44a53be919a3d24b280e728 --- /dev/null +++ b/2023/TarViS_ A Unified Approach for Target-Based Video Segmentation/layout.json @@ -0,0 +1,8827 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 92, + 94, + 501, + 111 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 94, + 501, + 111 + ], + "spans": [ + { + "bbox": [ + 92, + 94, + 501, + 111 + ], + "type": "text", + "content": "TarViS: A Unified Approach for Target-based Video Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": "Ali Athar" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": " Alexander Hermans" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": " Jonathon Luiten" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": " Deva Ramanan" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": " Bastian Leibe" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": "RWTH Aachen University, Germany " + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 133, + 518, + 186 + ], + "type": "text", + "content": "Carnegie Mellon University, USA \n{athar, hermans, luiten, leibe}@vision.rwth-aachen.de deva@cs.cmu.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "content": "The general domain of video segmentation is currently fragmented into different tasks spanning multiple benchmarks. Despite rapid progress in the state-of-the-art, current methods are overwhelmingly task-specific and cannot conceptually generalize to other tasks. Inspired by recent approaches with multi-task capability, we propose TarViS: a novel, unified network architecture that can be applied to any task that requires segmenting a set of arbitrarily defined 'targets' in video. Our approach is flexible with respect to how tasks define these targets, since it models the latter as abstract 'queries' which are then used to predict pixel-precise target masks. A single TarViS model can be trained jointly on a collection of datasets spanning different tasks, and can hot-swap between tasks during inference without any task-specific retraining. To demonstrate its effectiveness, we apply TarViS to four different tasks, namely Video Instance Segmentation (VIS), Video Panoptic Segmentation (VPS), Video Object Segmentation (VOS) and Point Exemplar-guided Tracking (PET). Our unified, jointly trained model achieves state-of-the-art performance on 5/7 benchmarks spanning these four tasks, and competitive performance on the remaining two. Code and model weights are available at: https://github.com/Ali2500/TarViS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "The ability to understand video scenes has been a longstanding goal of computer vision research because of wideranging applications in intelligent vehicles and robots. Early approaches tackled simpler tasks involving contour-based [33, 39] and box-level tracking [21, 25, 40, 52], background subtraction [20, 61], and motion segmentation [8, 49]. The deep learning boom then revolutionized the landscape by enabling methods to perform pixel-precise segmentation on challenging, real-world videos. In the past few years, a number of benchmarks have emerged, which evaluate how well methods can perform video segmentation according to various task formulations. Over time, these tasks/benchmarks have ballooned into separate re" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 210, + 545, + 449 + ], + "blocks": [ + { + "bbox": [ + 305, + 210, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 210, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 545, + 449 + ], + "type": "image", + "image_path": "acfefb23785a7de608b92173bfe6d504186c65a7e93cee1b2da1e70280c49b67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 453, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. Predicted results from a jointly trained TarViS model for four different video segmentation tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 545 + ], + "type": "text", + "content": "search sub-communities. Although existing methods are rapidly improving the state-of-the-art for these benchmarks, each of them typically tackles only one narrowly-defined task, and generalizing them is non-trivial since the task definition is baked into the core approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "We argue that this fragmentation is unnecessary because video target segmentation tasks all require the same high-level capability, namely that of identifying, localizing and tracking rich semantic concepts. Meanwhile, recent progress on Transformer networks has enabled the wider AI research community to move towards unified, multi-task architectures [1, 30, 31, 38, 58], because the attention operation [62] is well-suited for processing feature sets with arbitrary structure and data modality. These developments give us the opportunity to unify the fractured landscape of target-based video segmentation. In this paper, we propose TarViS: a novel architecture which enables a single, unified model to be jointly trained for multiple video segmentation tasks. During inference, the same model can perform differ" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18738" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 284, + 85 + ], + "type": "text", + "content": "ent tasks at runtime by specifying the segmentation target." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 85, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 85, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 85, + 287, + 229 + ], + "type": "text", + "content": "The core idea is that TarViS tackles the generic task of segmenting a set of arbitrary targets in video (defined as semantic classes or as specific objects). These targets are encoded as queries which, together with the video features, are input to a Transformer-based model. The model iteratively refines these queries and produces a pixel-precise mask for each target entity. This formulation conceptually fuses all video segmentation tasks [3, 54, 66, 72] which fall under the umbrella of the above-mentioned generic task, because they differ only in how the targets are defined. During both training and inference, TarViS can hot-swap between tasks at run-time by providing the desired target query set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 229, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 288, + 386 + ], + "type": "text", + "content": "To demonstrate our generalization capability, we tackle four different tasks: (1) Video Instance Segmentation (VIS) [54, 72], (2) Video Panoptic Segmentation (VPS) [35], (3) Video Object Segmentation [53], and (4) Point Exemplar-guided Tracking [3] (PET). For VIS, the segmentation targets are all objects in the video belonging to a predefined set of classes. The target set for VPS includes that for VIS, and additionally, a set of non-instantiable stuff semantic classes. For VOS, the targets are a specific set of objects for which the first-frame ground-truth mask is provided. PET is a more constrained version of VOS which only provides the location of a single point inside the object, rather than the full object mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 386, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 506 + ], + "type": "text", + "content": "Existing methods for these tasks lack generalization capability because task-specific assumptions are typically baked into the approach (see Sec. 2 and 3 for details). In contrast, TarViS can tackle all four tasks with a unified model because we encode the task-specific targets as a set of queries, thus decoupling the network architecture from the task definition. Moreover, our approach can theoretically generalize further, e.g., one could potentially define the target set as all objects described by a given text prompt, though this is beyond the scope of this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 507, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 287, + 603 + ], + "type": "text", + "content": "To summarize, our contributions are as follows: we propose TarViS, a novel architecture that can perform any task requiring segmentation of a set of targets from video. For the first time, we are able to jointly train and infer a single model on a collection of datasets spanning the four aforementioned tasks (VIS, VPS, VOS, PET). Our experimental results show that TarViS performs competitively for VOS, and achieves state-of-the-art results for VIS, VPS and PET." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 616, + 134, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 134, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 134, + 628 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 715 + ], + "type": "text", + "content": "Multi-task Models. Multi-task learning has a long history [11] with several architectures and training strategies [24,36,37,51,59,76]. Earlier approaches mostly consist of a shared backbone with fixed task-specific heads, whereas we design a more general architecture for video segmentation with task-specific targets to specify what to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": "segment. Our approach is inspired by recent attention-based models, e.g., PerceiverIO [30, 31], which can be trained on diverse data modalities and task-specific heads are replaced with output queries. UViM [38] follows a similar direction by creating a unified architecture for diverse dense prediction tasks. However, both of these models are trained separately for different tasks. Recent, powerful multi-task vision language models such as Flamingo [1] and GATO [58] tackle a multitude of tasks by requiring a sequence of task-specific input-output examples to prime the model. This is conceptually similar to our task-specific targets, however, our model does not require per-task priming. Moreover, our targets are not modeled as sequence prompts, and we aim for a video segmentation model which is several orders of magnitude smaller. In the realm of video tracking and segmentation, the recently proposed UNICORN [71] model tackles multiple object tracking-related tasks with a unified architecture. Unlike TarViS, however, UNICORN follows the task-specific output head approach and is generally oriented towards box-level tracking tasks [22, 45, 47, 75], thus requiring non-trivial modifications to tackle VPS or PET." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 331, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 487 + ], + "type": "text", + "content": "Query-based Transformer Architectures. Several works [2, 10, 13, 30, 31, 46, 65, 78] use query-based Transformer architectures for various tasks. The workhorse for task learning here is the iterative application of self- and cross-attention, where a set of query vectors (e.g., representing objects) are refined by interacting with each other, and with the input data (e.g., an image). Unlike existing methods which use queries in a task-specific context, TarViS adopts a query-based Transformer architecture in which the queries serve as a mechanism for decoupling the task definition from the architecture, i.e., our model can learn to tackle different tasks while being agnostic to their definition because the latter is abstracted behind a set of queries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 494, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 590 + ], + "type": "text", + "content": "Task-specific Video Segmentation. Current Video Instance Segmentation (VIS) methods broadly work by predicting object tracks in the video, followed by classification into a pre-defined set of categories. Several approaches [6,9,23,28,34,42,54,64,69,72] are based on the tracking-by-detection paradigm, some model video as a joint spatiotemporal volume [4,5], whereas recent works [12,26,29,65, 68] adopt Transformer-based architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "For Video Panoptic Segmentation (VPS), methods [35, 55, 66] generally extend image-level panoptic approaches [14] by employing multi-head architectures for semantic segmentation and instance mask regression, classification, and temporal association. In the Video Object Segmentation (VOS) community, state-of-the-art methods are broadly based on the seminal work of Oh et al. [50], which learns space-time correspondences between pixels in different video frames, and then uses these to propagate the first-frame masks across the video. Subsequent" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18739" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 408, + 240 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 408, + 240 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 408, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 408, + 240 + ], + "type": "image", + "image_path": "5a444117723d665becfbbab1396b0a327d3e8a6f8f53b46db883986d98e33859.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "text", + "content": "Figure 2. TarViS Architecture. Segmentation targets for different tasks are represented by a set of abstract target queries " + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "text", + "content": ". The core network (in green) is agnostic to the task definitions. The inner product between the output queries " + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "text", + "content": " and video feature " + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "inline_equation", + "content": "F_{4}" + }, + { + "bbox": [ + 46, + 243, + 546, + 277 + ], + "type": "text", + "content": " yields segmentation masks as required by the task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 545, + 239 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 545, + 239 + ], + "type": "image", + "image_path": "c4f063716734fedb84b6c65cf46ab1dca5aea4ecb26d9a27bf07ea842d90657b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 297, + 286, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 286, + 369 + ], + "type": "text", + "content": "methods [15-17, 60, 63, 70, 73, 73, 74] have significantly improved the performance and efficiency of this approach. Point Exemplar-guided Tracking (PET) [3, 27] is a fairly new task for which the current best approach [3] regresses a pseudo-ground-truth mask from the given point coordinates, and then applies a VOS method [17] to this mask." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 441 + ], + "type": "text", + "content": "The above methods thus incorporate task-specific assumptions into their core approach. This can benefit pertask performance, but makes it difficult for them to generalize across tasks. By contrast, TarViS can tackle all four aforementioned tasks, and generally any target-based video segmentation task, with a single, jointly trained model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 103, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 103, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 103, + 465 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 533 + ], + "type": "text", + "content": "TarViS can segment arbitrary targets in video since the architecture is flexible with respect to how these targets are defined, thus enabling us to conceptually unify and jointly tackle the four aforementioned tasks (VIS, VPS, VOS, PET). The architecture is illustrated in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "For all tasks, the common input to the network is an RGB video clip of length " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " denoted by " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{H \\times W \\times T \\times 3}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ". This is input to a 2D backbone network which produces image-level feature maps, followed by a Temporal Neck, which enables feature interaction across time and outputs a set of temporally consistent, multi-scale, " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "-dimensional feature maps " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{F_{32}, F_{16}, F_{8}, F_{4}\\}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "F_{s} \\in \\mathbb{R}^{\\frac{H}{s} \\times \\frac{W}{s} \\times T \\times D}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ". The feature maps " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " are then fed to our Transformer decoder, together with a set of queries " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " which represent the segmentation targets. The decoder applies successive layers of self- and masked cross-attention wherein the queries are iteratively refined by attending to each other, and to the feature maps, respectively. The refined queries output by the decoder are denoted with " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ". The following subsections explain how TarViS tackles each task in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 297, + 465, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 465, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 465, + 309 + ], + "type": "text", + "content": "3.1. Video Instance Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": "VIS defines the segmentation target set as all objects belonging to a set of predefined classes. Accordingly, the input query set " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " for VIS contains three types of queries: (1) semantic queries denoted by " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}} \\in \\mathbb{R}^{C \\times D}" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " is the number of classes defined by the dataset, i.e., each " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": "-dimensional vector in " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " represents a particular semantic class. (2) instance queries denoted by " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{inst}} \\in \\mathbb{R}^{I \\times D}" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " is assumed to be an upper bound on the number of instances in the video clip, and (3) a background query denoted by " + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{bg}} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 304, + 318, + 545, + 437 + ], + "type": "text", + "content": " to capture inactive instance queries." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "text", + "content": "The three query sets are concatenated, i.e., " + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}} = \\operatorname{concat}(Q_{\\mathrm{sem}}, Q_{\\mathrm{inst}}, Q_{\\mathrm{bg}})" + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "text", + "content": ", and input to the Transformer decoder, which refines their feature representation through successive attention layers and outputs a set of queries " + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{out}} = \\operatorname{concat}(Q_{\\mathrm{sem}}', Q_{\\mathrm{inst}}', Q_{\\mathrm{bg}}')" + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "text", + "content": ". These are then used to produce temporally consistent instance mask logits by computing the inner product " + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\langle F_4, Q_{\\mathrm{inst}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times I}" + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "text", + "content": ". To obtain classification logits, we compute the inner product " + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "inline_equation", + "content": "\\langle Q_{\\mathrm{inst}}', \\operatorname{concat}(Q_{\\mathrm{sem}}', Q_{\\mathrm{bg}}') \\rangle \\in \\mathbb{R}^{I \\times (C + 1)}" + }, + { + "bbox": [ + 304, + 441, + 546, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "text", + "content": "The three types of queries are initialized randomly at the start of training and optimized thereafter. The instance queries " + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{inst}}" + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "text", + "content": " enable us to segment a varying number of objects from the input clip. During training, we apply Hungarian matching between the predicted and ground-truth instance masks to assign instance queries to video instances, and then supervise their predicted masks and classification logits accordingly. When training on multiple datasets with heterogeneous classes, the semantic query sets are separately initialized per dataset, but " + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{inst}}" + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 552, + 545, + 672 + ], + "type": "text", + "content": " are shared." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 715 + ], + "type": "text", + "content": "Comparison to Instance Segmentation Methods. Several Transformer-based methods [10, 13, 65, 78] for image/video instance segmentation also use queries to seg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18740" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "ment a variable number of input instances. The key difference to our approach is the handling of object classes: existing works employ only instance queries which are input to a fully-connected layer with a fan-out of " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " to obtain classification (and background) scores. The notion of class-guided instance segmentation is thus baked into the approach. By contrast, TarViS is agnostic to the task-specific notion of object classes because it models them as arbitrary queries which are dynamic inputs to the network. The semantic representation for these queries is thus decoupled from the core architecture and is only learned via loss supervision. An important enabler for this approach is the background query " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ", which serves as a 'catch-all' class to represent everything that is not in " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ". It is used to classify non-active instance queries, and its mask logits are supervised to segment all non-object input pixels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 270, + 209, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 209, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 209, + 284 + ], + "type": "text", + "content": "3.2. Video Panoptic Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "text", + "content": "VPS defines the segmentation targets as all objects belonging to a set of thing classes (e.g., 'person', 'car'), and additionally, a set of non-instantiable stuff classes (e.g., 'sky', 'grass') which cover all non-object pixels. TarViS can tackle VPS with virtually no modification to the workflow in Sec. 3.1. We can compute semantic segmentation masks for the input clip by simply taking the inner product between " + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "text", + "content": " and the video features: " + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\langle F_4, Q_{\\mathrm{sem}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times C}" + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "text", + "content": ". Note that here, " + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 46, + 289, + 289, + 410 + ], + "type": "text", + "content": " contains queries representing both thing and stuff classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 289, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 289, + 545 + ], + "type": "text", + "content": "Comparison to VPS Methods. Current VPS datasets [35, 66] involve driving scene videos captured from moving vehicles. Methods tackling this task [35, 55] are based on earlier image panoptic segmentation approaches [14] which involve multi-head networks for semantic and instance segmentation prediction. In terms of image-level panoptic segmentation, Mask2Former [13] uses a Transformer-based architecture, but it models stuff classes as instances which are Hungarian-matched to the ground-truth target during training, whereas TarViS models semantic classes and instances using separate, designated queries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 552, + 288, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 288, + 578 + ], + "type": "text", + "content": "3.3. Video Object Segmentation and Point Exemplar-guided Tracking" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "VOS and PET are instantiations of a general task where the segmentation targets are a set of " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " objects for which some ground-truth cue " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is given. For VOS, " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is provided as the first-frame object masks " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times H \\times W}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", whereas for PET, " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is provided as the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " coordinates " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times 2}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " of a point inside each of the objects. TarViS jointly tackles both tasks by adopting a generalized approach in which the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " target objects are encoded into a set of object queries " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". Thus, both VOS and PET boil down to designing a function Encode0objects(\\cdot) which regresses " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " from the ground-truth cues " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and feature maps " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 359, + 72, + 545, + 85 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 545, + 85 + ], + "type": "interline_equation", + "content": "Q _ {\\mathrm {o b j}} \\leftarrow \\operatorname {E n c o d e 0 b j e c t s} (\\mathcal {G}, \\mathcal {F}). \\tag {1}", + "image_path": "5c35bafe9c4878dc6d32301ae54311609438445409d5a3e80c92f19b83210440.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "text", + "content": " is conceptually analogous to " + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{inst}}" + }, + { + "bbox": [ + 305, + 104, + 545, + 140 + ], + "type": "text", + "content": " used for VIS in that all three are abstract representations for their respective task-specific segmentation targets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 239 + ], + "type": "text", + "content": "Video Object Segmentation. We seek inspiration from HODOR [2] to implement EncodeObjects for VOS, a recent method for weakly-supervised VOS, which encodes objects into concise descriptors as follows: the descriptors are initialized by average pooling the image features inside the object masks, followed by iterative refinement where the descriptors attend to each other (self-attention) and to their respective soft-masked image features (cross-attention)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": "For TarViS, we employ a lightweight Object Encoder with a similar workflow to encode the objects as a set of queries " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": ", but with two differences to HODOR [2]: instead of cross-attending to the entire image feature map " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "(H \\cdot W" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": " points) with soft-masked attention, we apply hard-masked cross-attention to at most " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{max}}" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": " feature points per object, where " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{max}} \\ll H \\cdot W" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": ". Object masks containing more than " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{max}}" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": " points are sub-sampled accordingly. This significantly improves the memory/run-time overhead of our Object Encoder. Secondly, we note that the process of distilling object features into a single descriptor involves a loss of object appearance information, which degrades performance. We therefore model each object with " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "q_{o}" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": " queries (instead of one) by spatially dividing each object mask into " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "q_{o}" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": " segments, i.e., " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times q_{o} \\times D}" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": " (we use " + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "inline_equation", + "content": "q_{o} = 4" + }, + { + "bbox": [ + 304, + 240, + 546, + 420 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": "In addition to " + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": ", we initialize a set of background queries " + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{bg}} \\in \\mathbb{R}^{B \\times D}" + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": " to model the non-target pixels in the reference frame. Following HODOR [2], we employ multiple background queries, which are initialized dynamically by dividing the video frame containing the groundtruth masks " + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": " grid and average pooling the non-object pixels in each grid cell. The Object Encoder jointly refines the background and object queries to yield " + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}} = \\text{concat}(Q_{\\mathrm{obj}}, Q_{\\mathrm{bg}})" + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": ". During training, the mask logits for the multiple background queries are aggregated per-pixel by applying " + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "inline_equation", + "content": "\\max(\\cdot)" + }, + { + "bbox": [ + 304, + 420, + 547, + 563 + ], + "type": "text", + "content": " and supervised to segment all pixels not part of the target object set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": "The remaining workflow follows that for VIS and VPS: " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": " is input to the Transformer decoder together with the video features " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": ". The refined output query set " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{out}} = \\operatorname{concat}(Q_{\\mathrm{obj}}', Q_{\\mathrm{bg}}')" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": " is then used to compute the inner product " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\langle F_4, Q_{\\mathrm{obj}}' \\rangle \\in \\mathbb{R}^{H \\times W \\times T \\times O \\times q_o}" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "\\max(\\cdot)" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": " is applied on the " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "q_o" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": "-sized dimension to obtain the final mask logits for the " + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 563, + 547, + 650 + ], + "type": "text", + "content": " target objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Point Exemplar-guided Tracking. For PET we implement Encode0objects in the exactly same way as VOS: the given point coordinates " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{obj}}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " are converted into a mask with just one non-zero pixel, followed by iterative refinement by the Object Encoder (with shared weights for VOS and PET)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18741" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "The only difference is that here we represent each of the " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " objects with just one query, i.e., " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}} \\in \\mathbb{R}^{O \\times D}" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "q_o = 1" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "). The subsequent workflow is also identical to that for VOS: the queries are refined by the Transformer decoder followed by an inner product with " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "F_4" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " to obtain object mask logits." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 137, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 137, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 287, + 329 + ], + "type": "text", + "content": "Comparison to VOS and PET Methods. Current state-of-the-art VOS methods are largely based on STM [50]. It involves learning pixel-to-pixel correspondences across video frames, which are then used to propagate the given object mask across the video. This approach is effective since every pixel in the given mask can be individually mapped to future frames, thus preserving fine-grained object details. The core approach is, however, task-specific since it assumes the availability of first-frame object masks, and does not generalize to the PET (see Sec. 4.2). PET can be viewed as a more constrained version of VOS, where only a single object point is provided instead of the full mask. Consequently, PET [3] is currently tackled by casting it as a VOS problem by using an image instance segmentation network [13] to regress pseudo-ground-truth object masks from the given point coordinates " + }, + { + "bbox": [ + 46, + 137, + 287, + 329 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 137, + 287, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "text", + "content": "On the other hand, our approach of encoding objects as concise queries causes loss of fine-grained object appearance information, but it has the advantage of being agnostic to how " + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "text", + "content": " is defined. As evident from the unified workflow for VOS and PET, any variation of these tasks with arbitrary ground-truth cues " + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "text", + "content": " can be seamlessly fused into our architecture as long as we can implement an effective Encode0bjects function to regress " + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "text", + "content": " from the given " + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 331, + 287, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 438, + 173, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 173, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 173, + 450 + ], + "type": "text", + "content": "3.4. Network Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": "Temporal Neck. TarViS produces target masks by computing the inner product between " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": " and the video feature map " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "F_{4}" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": ". For this to work, the per-pixel features " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": " must be aligned for the same, and dissimilar for different targets. Some image instance segmentation methods [13, 78] apply Deformable Attention [78] to the backbone feature maps to efficiently learn multi-scale image features. For TarViS, however, the features must also be temporally consistent across the entire input video clip. To achieve this, we propose a novel Temporal Neck architecture inspired from the work of Bertasius et al. [7] for video action classification. We enable efficient spatio-temporal feature interaction by applying two types of self-attention in an alternating fashion: the first is spatially global and temporally localized, whereas the second is spatially localized and temporally global. The first operation is implemented with Deformable Attention, following existing work [12, 78]. The second operation, Temporal Attention, involves dividing the input space-time volume into a grid along the spatial axes, and then applying self-attention to the space-time feature volume inside each grid cell. Both operations allow feature" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 533, + 225 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 533, + 225 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 533, + 225 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 533, + 225 + ], + "type": "image", + "image_path": "aacbac1eceba824eeabe1aa85079743ce7c28a8009455c1fca3a9c0ce70fbcdb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 545, + 284 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 284 + ], + "type": "text", + "content": "Figure 3. Temporal Neck Layer. Colored regions denote the attention field w.r.t the selected pixel (darkened). Deformable Attention is spatially unrestricted but temporally limited to a single frame, whereas Temporal Attention is spatially localized, but temporally unrestricted. " + }, + { + "bbox": [ + 304, + 228, + 545, + 284 + ], + "type": "inline_equation", + "content": "F_{8}" + }, + { + "bbox": [ + 304, + 228, + 545, + 284 + ], + "type": "text", + "content": " is inactive for the temporal attention." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "type": "text", + "content": "interaction across multiple scales. Both attention operations are illustrated in Fig. 3. We exclude " + }, + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "type": "inline_equation", + "content": "F_{8}" + }, + { + "bbox": [ + 304, + 293, + 545, + 342 + ], + "type": "text", + "content": " from temporal attention since we found this to be more memory-efficient without negatively impacting prediction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 346, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 406 + ], + "type": "text", + "content": "Transformer Decoder. The decoder architecture follows that of Mask2Former [13]: the input queries are iteratively refined over multiple layers. In each layer, the queries first cross-attend to their respective masked video features, then self-attend to each other, followed by feed-forward layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 415, + 373, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 373, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 373, + 426 + ], + "type": "text", + "content": "3.5. Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": "To infer on videos with arbitrary length, we split videos into clips of length " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{clip}}" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " with an overlap of " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{ov}}" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " between successive clips. Object tracks are associated across clips based on their mask IoU in the overlapping frames. For the VOS tasks, the object queries for an intermediate clip are initialized by using the predicted masks in the overlapping frames from the previous clip as a pseudo-ground-truth. For VPS, we average the semantic segmentation logits in the overlapping frames. Our approach is thus near-online because the time delay in obtaining the output for a given frame is at most " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{clip}} - T_{\\mathrm{ov}} - 1" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " (except for the first clip in the video)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 578, + 387, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 387, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 387, + 592 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 598, + 439, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 439, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 439, + 611 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Our Temporal Neck contains 6 layers of Deformable and Temporal Attention. We pretrain for " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "500\\mathrm{k}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " iterations on pseudo-video clips generated by applying on-the-fly augmentations to images from COCO [43], ADE20k [77], Mapillary [48] and Cityscapes [18]. The samples are either trained for VPS, VIS, VOS or PET. This is followed by fine-tuning for " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\mathrm{k}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " iterations jointly on samples from YouTube-VIS [72], OVIS [54], KITTI-STEP [66]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18742" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 86, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 511, + 82 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 511, + 82 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 511, + 82 + ], + "type": "text", + "content": "Table 1. Results for Video Instance Segmentation (VIS) on the YouTube-VIS 2021 [72] and OVIS [54] validation sets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 86, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 48, + 86, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 86, + 547, + 236 + ], + "type": "table", + "html": "
MethodBackboneShared ModelYouTube-VIS 2021OVIS
APAP50AP75AR1AR10APAP50AP75AR1AR10
Mask2Former-VIS [12]R-50X40.660.941.8-------
IDOL [69]R-50X43.968.049.638.050.930.251.330.015.037.5
MinVIS [28]R-50X44.266.048.139.251.725.045.524.013.929.7
VITA [26]R-50X45.767.449.540.953.619.641.217.411.726.0
TarViSR-50✓48.369.653.240.555.931.152.530.415.939.9
Mask2Former-VIS [12]Swin-TX45.968.750.7-------
TarViSSwin-T✓50.971.656.642.257.234.055.034.416.140.9
IDOL [69]Swin-LX56.180.863.545.060.142.665.745.217.949.6
VITA [26]Swin-LX57.580.661.047.762.627.751.924.914.933.0
TarViSSwin-L✓60.281.467.647.664.843.267.844.618.050.4
", + "image_path": "f7bd62b8b504abb59fbcc6cd1c40013fb5adf9389ffc2c095cc63fe19729171a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 261, + 546, + 394 + ], + "blocks": [ + { + "bbox": [ + 58, + 247, + 534, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 534, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 534, + 258 + ], + "type": "text", + "content": "Table 2. Video Panoptic Segmentation (VPS) results for validation sets of KITTI-STEP [66], CityscapesVPS [35] and VIPseg [44]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 261, + 546, + 394 + ], + "lines": [ + { + "bbox": [ + 48, + 261, + 546, + 394 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 546, + 394 + ], + "type": "table", + "html": "
MethodShared ModelKITTI-STEPCityscapesVPSVIPSeg
STQAQSQVPQ\\( VPQ^{Th} \\)\\( VPQ^{St} \\)VPQ\\( VPQ^{Th} \\)\\( VPQ^{St} \\)STQ
Mask Propagation [66]✘0.670.630.71------
Track [35]✘---55.943.764.8---
VPSNet [35]✘0.560.520.6157.044.766.014.014.014.220.8
VPSNet-SiamTrack [67]✘---57.344.766.417.217.317.321.1
VIP-Deeplab [55]✘---63.149.573.016.012.318.222.0
Clip-PanoFCN [44]✘------22.925.020.831.5
TarViS (R-50)✓0.700.700.6953.335.966.033.539.228.543.1
TarViS (Swin-T)✓0.710.710.7058.042.969.035.842.729.745.3
TarViS (Swin-L)✓0.720.720.7358.943.769.948.058.239.052.9
", + "image_path": "b26befd008a442f377f814359271f0724bff7a9bad57089fd56de6fcafe24d0f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "content": "CityscapesVPS [35], VIPSeg [44], DAVIS [53] and BURST [3]. We train on 32 Nvidia A100 GPUs with batch size 1 per GPU. For each of the query types (" + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{inst}}" + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "content": ") discussed in Sec. 3, we employ a learned query embedding, which is used when computing the " + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{Key}^T" + }, + { + "bbox": [ + 46, + 413, + 288, + 498 + ], + "type": "text", + "content": " Query affinity matrix for multi-head attention inside the decoder. We refer to the supplementary for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 509, + 162, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 162, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 162, + 521 + ], + "type": "text", + "content": "4.2. Benchmark Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 564 + ], + "type": "text", + "content": "All results are computed with a single, jointly trained model which performs different tasks by simply providing the corresponding query set at run-time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "Video Instance Segmentation (VIS). We evaluate on (1) YouTube-VIS 2021 [72] which covers 40 object classes and contains 2985/421 videos for training/validation, and (2) OVIS [54] which covers 25 object classes. It contains 607/140 videos for training/validation which are comparatively longer and more occluded. The AP scores for both are reported in Tab. 1. For all three backbones, TarViS achieves state-of-the-art results for both benchmarks even though other methods are trained separately per benchmark whereas we use a single model. On YouTube-VIS, TarViS achieves 48.3 AP with a ResNet-50 backbone compared to the 45.7 achieved by VITA [26]. With Swin-L, we achieve" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 413, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 462 + ], + "type": "text", + "content": "60.2 AP which is also higher than the 57.5 by VITA. On OVIS with ResNet-50, our 31.1 AP is higher than the 30.2 for IDOL [69], and with Swin-L, TarViS (43.2 AP) outperforms the current state-of-the-art IDOL (42.6 AP)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "type": "text", + "content": "Video Panoptic Segmentation (VPS). We evaluate VPS on three datasets: (1) KITTI-STEP [66], which contains 12/9 lengthy driving scene videos for training/validation with 19 semantic classes (2 thing and 17 stuff classes), (2) CityscapesVPS [35], which contains 50 short driving scene clips, each with 6 annotated frames, and (3) VIPSeg [44], which is a larger dataset with 2806/343 in-the-wild videos for training/validation and 124 semantic classes. The results are reported in Tab. 2. For KITTI-STEP, TarViS achieves " + }, + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "type": "text", + "content": " STQ with a ResNet-50 backbone which is better than all existing approaches. The performance further improves to " + }, + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 304, + 469, + 546, + 662 + ], + "type": "text", + "content": " with Swin-L. For CityscapesVPS, TarViS achieves 58.9 VPQ which is higher than all other methods except VIP-Deeplab [55] (63.1). However, VIP-Deeplab performs monocular depth estimation for additional guidance, and therefore requires ground-truth depth-maps for training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "For VIPSeq, TarViS outperforms existing approaches by a significant margin. With a ResNet-50 backbone, our 33.5 VPQ is " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " higher than the 22.9 by Clip-PanoFCN [44]. With a Swin-Large backbone, TarViS achieves 48.0 VPQ" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18743" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 97, + 287, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 286, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 286, + 93 + ], + "type": "text", + "content": "Table 3. Results for VOS on DAVIS [53] and PET on BURST [3]. Detailed PET metrics are provided in supplementary." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 97, + 287, + 277 + ], + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 277 + ], + "type": "table", + "html": "
MethodDAVIS (VOS)BURST (PET)
J&FJFHval allHtest all
UNICORN* [71]70.666.175.0--
HODOR [2]81.378.483.9--
STM [50]81.879.284.3--
CFBI [73]81.979.184.6--
HMMN [60]84.781.987.5--
AOT [74]84.982.387.5--
STCN [17]85.482.288.6--
XMem [15]86.282.989.5--
Box Tracker [32]---12.710.1
STCN+M2F [13, 17]---24.424.9
TarViS (R-50)82.679.385.930.932.1
TarViS (Swin-T)82.879.686.036.036.4
TarViS (Swin-L)85.381.788.537.536.1
", + "image_path": "424a2a54db6d352fc6c7a2268ecc76238b4004e7dd5444449e3c47691f6bcc73.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 285, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 381 + ], + "type": "text", + "content": "which is more than double that of Clip-PanoFCN (22.9). Note that VIP-Deeplab performs significantly worse for VIPSeg (16.0 VPQ), showing that TarViS generalizes better across benchmarks. Finally, we note that larger backbones results in significant performance gains for datasets with in-the-wild internet videos as in VIPSeg, but for specialized driving scene datasets (e.g. KITTI-STEP and Cityscapes-VPS), the improvements are much smaller." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "type": "text", + "content": "Video Object Segmentation (VOS). We evaluate VOS on the DAVIS 2017 [53] dataset, which contains 60/30 YouTube videos for training/validation. The results in Tab. 3 show that TarViS achieves 85.3 " + }, + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "type": "text", + "content": " which is higher than all existing methods except STCN [17] (85.4) and XMem [15] (86.2). As mentioned in Sec. 3.3, encoding objects as queries incurs a loss of fine-grained information, which is detrimental to performance. On the other hand, space-time correspondence (STC) based approaches learn pixel-to-pixel affinities between frames, which enables them to propagate fine-grained object appearance information. We note, however, that TarViS is the first method not based on the STC paradigm which achieves this level is performance (85.3 " + }, + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 46, + 385, + 287, + 577 + ], + "type": "text", + "content": "), outperforming several STC-based methods as well as all non-STC based methods e.g. HODOR [2] (81.5) and UNICORN [71] (70.6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Point Exemplar-guided Tracking (PET). PET is evaluated on the recently introduced BURST benchmark [3] which contains 500/1000/1500 diverse videos for training/validation/testing. It is a constrained version of VOS which only provides the point coordinates of the object mask centroid instead of the full mask. Tab. 3 shows that existing methods can only tackle either VOS or PET. To verify this, we tried adapting STCN [17] for PET by training it with point masks, but the training did not converge. By contrast, TarViS encodes objects into queries, which enables it to tackle both tasks with a single model since the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 545, + 237 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 237 + ], + "type": "image", + "image_path": "1ce3d544c2c49354fbb19480fcf5eb55b9185888dcdb959db88667655a571c8e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 240, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 264 + ], + "type": "text", + "content": "Figure 4. Qualitative results from a single TarViS model for all four tasks. Further results are shown in the supplementary." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 271, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 295 + ], + "type": "text", + "content": "object guidance (point or mask) is abstracted behind the EncodeObjects(·) function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 296, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 369 + ], + "type": "text", + "content": "TarViS achieves a " + }, + { + "bbox": [ + 304, + 296, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{HOTA_{all}}" + }, + { + "bbox": [ + 304, + 296, + 545, + 369 + ], + "type": "text", + "content": " score of 37.5 and 36.4 on the validation and test sets, respectively, which is significantly better than the 24.4 and 24.9 achieved by the best performing baseline method which casts PET as a VOS problem by regressing a pseudo-ground-truth mask from the given point, followed by applying a VOS approach (STCN [17])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 376, + 373, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 373, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 373, + 387 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 394, + 531, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 531, + 406 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 531, + 406 + ], + "type": "text", + "content": "Table 4 shows several architecture/training ablations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 410, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 601 + ], + "type": "text", + "content": "Task-specific Training (row 1-3). The first three rows show results for task-specific models. We train a single model for VOS and PET since both tasks are closely related. We note that the VIS-only model performs worse than the multi-task model on YouTube-VIS (46.3 vs. 48.3) but slightly better on OVIS (31.5 vs. 31.1). For VPS, the performance on KITTI-STEP is unchanged, but Cityscapes-VPS and VIPSeg both show improvements with the multi-task model. Lastly, for VOS the task-specific model performs slightly worse on DAVIS (81.1 vs. 82.0) but significantly better on BURST for PET (34.7 vs. 30.9). To summarize, the final, multi-task model performs better on 4/7 benchmarks, worse on 2/7, and matches performance on 1/7 when compared to task-specific models. We thus conclude that the combination of multi-task supervision and more data is generally beneficial for performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": "Semantic Queries for VIS (row 4). TarViS represents object classes as dynamic query inputs to the network (" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{sem}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": ", Sec. 3.1). We ablate this by modifying our network to work for only VIS by discarding the semantic/background queries and adopting a technique similar to existing methods [13, 65], i.e. using instance queries " + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{inst}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": " in conjunction with a linear layer for classification (separate for each dataset). Comparing the results with the VIS-only setting which is trained on similar data, we see that this architecture" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18744" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 86, + 550, + 211 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 539, + 82 + ], + "type": "text", + "content": "Table 4. Ablation experiment results with ResNet-50 backbone. C-VPS: CityscapesVPS, YTVIS: YouTube-VIS, KITTI: KITTI-STEP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 86, + 550, + 211 + ], + "lines": [ + { + "bbox": [ + 48, + 86, + 550, + 211 + ], + "spans": [ + { + "bbox": [ + 48, + 86, + 550, + 211 + ], + "type": "table", + "html": "
SettingVideo Training DataVISVPSVOSPET
YTVISOVISKITTIC-VPSVIPSegYTVIS (mAP)OVIS (mAP)KITTI (STQ)C-VPS (VPQ)VIPSeg (VPQ)DAVIS (J&F)BURST (HOTAall)
1. VIS✓✓46.331.5-----
2. VPS✓✓✓--0.7049.732.4--
3. VOS + PET✓✓----81.134.7
4. No Semantic Queries✓✓44.729.8-----
5. No Temporal Neck✓✓✓✓✓42.822.30.6951.228.978.730.3
Final✓✓✓✓✓48.331.10.7053.333.582.030.9
", + "image_path": "bf511f89a9cce0a47c1e705f232d86c05d14513f509014ead367c3aad7a7c6b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 230, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 278 + ], + "type": "text", + "content": "performs worse than the VIS-only setting on both YouTube-VIS (44.7 vs. 46.3) and OVIS (29.8 vs. 31.5). Thus, our semantic query based classification makes the network architecture task-agnostic and also yields better performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": "Temporal Neck (row 4). We validate our novel Temporal Neck (Sec. 3.4) by training a model with a simpler neck that contains only Deformable Attention layers [78], similar to Mask2Former [13], i.e. there is no feature interaction across frames. Doing this degrades performance across all datasets, with particularly large drops for YouTube-VIS (42.8 vs. 48.3) and OVIS (22.3 vs. 31.1). This shows that inter-frame feature interactions enabled by our Temporal Neck are highly beneficial for down-stream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 402, + 117, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 117, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 117, + 415 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": "Limitations. Training on multiple datasets/tasks does not necessarily improve performance on all benchmarks. For VOS, the model exhibits class bias and sometimes fails to track unusual objects which were not seen during training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": "Future Outlook. We jointly trained TarViS for four different tasks to validate its generalization capability. The architecture can, however, tackle any video segmentation task for which the targets can be encoded as queries. The recent emergence of joint language- vision models [41, 56, 57] thus makes it possible to perform multi-object segmentation based on a text prompt if the latter can be encoded as a target query using a language encoder [19]. Another interesting possibility is that TarViS could be applied to multiple tasks in the same forward pass by simply concatenating the task-specific queries. Fig. 5 offers a promising outlook for this; it shows our model's output for a video clip from a popular TV series where we perform VIS and VOS simultaneously by providing the semantic query for the 'person' class (from YouTube-VIS [72]), and the VOS-based object queries for the dragon by annotating its first frame mask, i.e. " + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{in}} = \\mathrm{concat}(Q_{\\mathrm{sem}}, Q_{\\mathrm{inst}}, Q_{\\mathrm{obj}}, Q_{\\mathrm{bg}})" + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": ". TarViS successfully segments all four persons in the scene (VIS) and the dragon (VOS), even though our model was never trained to simultaneously tackle both tasks in a single forward pass." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 426, + 364 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 426, + 364 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 426, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 426, + 364 + ], + "type": "image", + "image_path": "f8ad2840a6ba0111007f04d5c2bddaaa31b66f876f9140351c49f9d69ef9d146.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 366, + 545, + 400 + ], + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 400 + ], + "type": "text", + "content": "Figure 5. TarViS performing VIS and VOS in a single forward pass. We provide the mask for the dragon on the left, and the semantic query for the 'person' class." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 228, + 545, + 364 + ], + "blocks": [ + { + "bbox": [ + 426, + 228, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 426, + 228, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 426, + 228, + 545, + 364 + ], + "type": "image", + "image_path": "5b8ef09b0ebbf85a3cd99440f9cbc3b49e31cd67c2b0d55f3460e48badb96cea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 427, + 545, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 607 + ], + "type": "text", + "content": "We presented TarViS: a novel, unified approach for tackling any task requiring pixel-precise segmentation of a set of targets in video. We adopt a generalized paradigm where the task-specific targets are encoded into a set of queries which are then input to our network together with the video features. The network is trained to produce segmentation masks for each target entity, but is inherently agnostic to the task-specific definition of these targets. To demonstrate the effectiveness of our approach, we applied it to four different video segmentation tasks (VIS, VPS, VOS, PET). We showed that a single TarViS model can be jointly trained for all tasks, and during inference can hot-swap between tasks without any task-specific fine-tuning. Our model achieved state-of-the-art performance on five benchmarks and has multiple, promising directions for future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 610, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 707 + ], + "type": "text", + "content": "Acknowledgments. This project was partially funded by ERC CoG DeeVise (ERC-2017-COG-773161) and BMBF project NeuroSys-D (03ZU1106DA). Compute resources were granted by RWTH Aachen under project ID supp0003, and by the Gauss Centre for Supercomputing e.V. through the John von Neumann Institute for Computing on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18745" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Arxiv, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "text", + "content": "[2] Ali Athar, Jonathon Luiten, Alexander Hermans, Deva Ramanan, and Bastian Leibe. Hodor: High-level object descriptors for object re-segmentation in video learned from static images. In CVPR, 2022. 2, 4, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "type": "text", + "content": "[3] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 267 + ], + "type": "text", + "content": "[4] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "type": "text", + "content": "[5] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. A single-stage, bottom-up approach for occluded vis using spatio-temporal embeddings. In ICCV-W, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 346 + ], + "type": "text", + "content": "[6] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 378 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 402 + ], + "type": "text", + "content": "[8] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In ECCV, 2010. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[9] Jiale Cao, Rao Muhammad Anwer, Hisham Cholakkal, Fahad Shahbaz Khan, Yanwei Pang, and Ling Shao. Sipmask: Spatial information preservation for fast image and video instance segmentation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 490 + ], + "type": "text", + "content": "[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "type": "text", + "content": "[11] Richard Caruana. Multitask Learning: A Knowledge-Based Source of Inductive Bias. In ICML, 1993. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[12] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In Arxiv, 2021. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[13] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. In CVPR, 2022. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[14] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[15] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[16] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[17] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[18] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "text", + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[20] Ahmed Elgammal, David Harwood, and Larry Davis. Nonparametric model for background subtraction. In ECCV, 2000. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 277 + ], + "type": "text", + "content": "[21] Andreas Ess, Bastian Leibe, Konrad Schindler, and Luc Van Gool. Moving obstacle detection in highly dynamic scenes. In ICRA, 2009. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 277, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 323 + ], + "type": "text", + "content": "[22] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 323, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 545, + 357 + ], + "type": "text", + "content": "[23] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 358, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 545, + 392 + ], + "type": "text", + "content": "[24] Golnaz Ghiasi, Barret Zoph, Ekin D Cubuk, Quoc V Le, and Tsung-Yi Lin. Multi-Task Self-Training for Learning General Representations. In CVPR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 393, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 415 + ], + "type": "text", + "content": "[25] Helmut Grabner, Michael Grabner, and Horst Bischof. Real-time tracking via on-line boosting. In BMVC, 2006. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 417, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 449 + ], + "type": "text", + "content": "[26] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 451, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 451, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 451, + 545, + 484 + ], + "type": "text", + "content": "[27] Namdar Homayounfar, Justin Liang, Wei-Chiu Ma, and Raquel Urtasun. Videoclick: Video object segmentation with a single click. Arxiv, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 545, + 518 + ], + "type": "text", + "content": "[28] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 520, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 553 + ], + "type": "text", + "content": "[29] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[30] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. *ICLR*, 2022. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "type": "text", + "content": "[31] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "text", + "content": "[32] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020.7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[33] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 1" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18746" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[34] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[35] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In CVPR, 2020. 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[36] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "text", + "content": "[37] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "text", + "content": "[38] Alexander Kolesnikov, André Susano Pinto, Lucas Beyer, Xiaohua Zhai, Jeremiah Harmsen, and Neil Houlsby. Uvim: A unified modeling approach for vision with learned guiding codes. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 308 + ], + "type": "text", + "content": "[39] Dieter Koller, Joseph Weber, and Jitendra Malik. Robust multiple car tracking with occlusion reasoning. In ECCV, 1994. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[40] Bastian Leibe, Konrad Schindler, Nico Cornelis, and Luc Van Gool. Coupled object detection and tracking from static cameras and moving vehicles. PAMI, 2008. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[41] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[42] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "text", + "content": "[44] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In CVPR, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[45] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. In arXiv:1603.00831, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 566 + ], + "type": "text", + "content": "[46] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[47] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[48] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 678 + ], + "type": "text", + "content": "[49] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. PAMI, 2013. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[50] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In ICCV, 2019. 2, 5, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[51] Kilian Pfeiffer, Alexander Hermans, István Sárándi, Mark Weber, and Bastian Leibe. Visual person understanding through multi-task and multi-dataset learning. In GCPR, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[52] Hamed Pirsiavash, Deva Ramanan, and Charless C Fowlkes. Globally-optimal greedy algorithms for tracking a variable number of objects. In CVPR, pages 1201-1208. IEEE, 2011. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[53] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. In *Arxiv*, 2017. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[54] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. In IJCV, 2022. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 252, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 295 + ], + "type": "text", + "content": "[55] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, 2021. 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 351 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 353, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 387 + ], + "type": "text", + "content": "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[58] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. In *Arxiv*, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 432, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 453 + ], + "type": "text", + "content": "[59] Sebastian Ruder. An Overview of Multi-Task Learning in Deep Neural Networks. arXiv:1706.05098, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "text", + "content": "[60] Hongje Seong, Seoung Wug Oh, Joon-Young Lee, Seongwon Lee, Suhyeon Lee, and Euntai Kim. Hierarchical memory matching network for video object segmentation. In ICCV, 2021. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 521 + ], + "type": "text", + "content": "[61] C. Stauffer and W.E.L. Grimson. Adaptive background mixture models for real-time tracking. In CVPR, 1999. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[63] Paul Voigtlaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast end-to-end embedding learning for video object segmentation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[64] Paul Voigtlaender, Michael Krause, Aljosa Osep, Jonathon Luiten, Berin Balachandar Gnana Sekar, Andreas Geiger, and Bastian Leibe. Mots: Multi-object tracking and segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[65] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 2, 3, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[66] Mark Weber, Jun Xie, Maxwell Collins, Yukun Zhu, Paul Voigtaender, Hartwig Adam, Bradley Green, Andreas" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18747" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 512 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Geiger, Bastian Leibe, Daniel Cremers, Aljosa Osep, Laura Leal-Taixe, and Liang-Chieh Chen. STEP: Segmenting and tracking every pixel. In NeurIPS, 2021. 2, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[67] Sanghyun Woo, Dahun Kim, Joon-Young Lee, and In So Kweon. Learning to associate every segment for video panoptic segmentation. In CVPR, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 173 + ], + "type": "text", + "content": "[68] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[69] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[70] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, and Wenxiu Sun. Efficient regional memory network for video object segmentation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[71] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV, 2022. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 277, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 297 + ], + "type": "text", + "content": "[72] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 300, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 331 + ], + "type": "text", + "content": "[73] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "text", + "content": "[74] Zongxin Yang, Yunchao Wei, and Yi Yang. Associating objects with transformers for video object segmentation. In NeurIPS, 2021. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 367, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 410 + ], + "type": "text", + "content": "[75] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 412, + 287, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 444 + ], + "type": "text", + "content": "[76] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 446, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 446, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 446, + 287, + 478 + ], + "type": "text", + "content": "[77] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 479, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 512 + ], + "type": "text", + "content": "[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICML, 2020. 2, 3, 5, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18748" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_content_list.json b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c9038931b49c0a4c737bf64bbf07b6830f75d01 --- /dev/null +++ b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_content_list.json @@ -0,0 +1,1506 @@ +[ + { + "type": "text", + "text": "Target-referenced Reactive Grasping for Dynamic Objects", + "text_level": 1, + "bbox": [ + 189, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jirong Liu $^{1,2}$ , Ruo Zhang $^{2}$ , Hao-Shu Fang $^{2}$ , Minghao Gou $^{2}$ , Hongjie Fang $^{2}$ , Chenxi Wang $^{2,3}$ , Sheng Xu $^{2}$ , Hengxu Yan $^{2}$ , Cewu Lu $^{1,2\\dagger}$ $^{1}$ Shanghai Qi Zhi institute, $^{2}$ Shanghai Jiao Tong University, $^{3}$ Flexiv Robotics, LTD jirong@sjtu.edu.cn, {ruozhang0608, fhaoshu}@gmail.com, {gmh2015, galaxies, wcx1997, xs1020, hengxuyan, lucewu}@sjtu.edu.cn", + "bbox": [ + 156, + 179, + 810, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reactive grasping, which enables the robot to successfully grasp dynamic moving objects, is of great interest in robotics. Current methods mainly focus on the temporal smoothness of the predicted grasp poses but few consider their semantic consistency. Consequently, the predicted grasps are not guaranteed to fall on the same part of the same object, especially in cluttered scenes. In this paper, we propose to solve reactive grasping in a target-referenced setting by tracking through generated grasp spaces. Given a targeted grasp pose on an object and detected grasp poses in a new observation, our method is composed of two stages: 1) discovering grasp pose correspondences through an attentional graph neural network and selecting the one with the highest similarity with respect to the target pose; 2) refining the selected grasp poses based on target and historical information. We evaluate our method on a large-scale benchmark GraspNet-1Billion. We also collect 30 scenes of dynamic objects for testing. The results suggest that our method outperforms other representative methods. Furthermore, our real robot experiments achieve an average success rate of over 80 percent. Code and demos are available at: https://graspnet.net/reactive.", + "bbox": [ + 75, + 335, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 695, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reactive grasping is in great demand in the industry. For instance, in places where human-robot collaboration is heavily required like factories, stress on laborers will be significantly relieved if robots can receive tools from humans and complete the harder work for laborers. Such a vision is based on reactive grasping.", + "bbox": [ + 75, + 719, + 468, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the contrary to static environments, in reactive grasping, dynamic task setting poses new challenges for algorithm design. Previous research in this area mainly focuses", + "bbox": [ + 75, + 811, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/605780c57469374cb35909e698f20e87ce52c6f039065b5a5da40e7ff9b62a59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 301, + 875, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b978c230c54871cba1ce8cf59944b78d8f13261f9f3a9fbd2d5654f2f458dff.jpg", + "image_caption": [ + "Figure 1. (a) Classic reactive grasping guarantees the smoothness of the grasp poses but cannot predicts grasps on the same part of the hammer. (b) Our target-referenced reactive grasping takes semantic consistency into consideration. The generated grasp poses across frames are illustrated with blue grippers." + ], + "image_footnote": [], + "bbox": [ + 522, + 440, + 870, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on planning temporally smooth grasps [22, 42] to avoid wavy and jerky robot motion. Few of them pay attention to its semantic consistency. In short, given a targeted grasp at the first frame, we want the robot to grasp the same part of the object in the following frames. Additionally, it is not guaranteed that grasp predictions made by classical methods fall on the same object in cluttered scenes. Hence, most of their experiments are conducted on single-object scenes. Unlike previous works, this work is aimed at achieving temporally smooth and semantically consistent reactive grasping in clutter given a targeted grasp. We refer to such a task setting as target-referenced reactive grasping as shown in Fig.1. Note that despite robot handover is a major application scenario of reactive grasping, this work focuses on a more general task setting - dynamic object grasping.", + "bbox": [ + 496, + 672, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Cewu Lu is the corresponding author, a member of Qing Yuan Research Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8824", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A naive idea to solve this task is to generate reference grasp poses for the initial scene and consecutively track the object's 6D pose. As the object moves, the initial grasp pose can be projected to a new coming frame based on the object's 6D pose. Although such an idea seems to be natural and valid, some bottleneck greatly degrades its viability. First of all, the solution to reactive grasping should be able to handle objects' motion in real-time, meaning that it requires fast inference speed and immediate response to continuous environmental changes. However, 6D pose tracking may be time-consuming due to commonly-used instance segmentation [11,40]. Second, 6D pose tracking usually requires objects' prior knowledge, such as CAD models [4,41], which is not always available in the real world as well or achieves only category-level generalization [37].", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from tracking objects, we propose to track grasps by a two-stage policy instead. We also comply with the restriction that no prior knowledge of the objects is allowed. Given a target grasp on a partial-view point cloud, we first discover its corresponding grasp among future frame's detected grasp poses as coarse estimation. These grasp poses can be given by an off-the-shelf grasp detector. Inspired by recent progress in local feature matching, which often uses image descriptors like SIFT [17] to describe interesting regions of images, we view grasp poses and their corresponding features as geometric descriptors on a partial-view point cloud. Based on such an assumption, we can simply estimate correspondences between two grasp sets from two different observation frames by matching the associated grasp features. Note that in opposition to classical local feature matching, features of the entire scene are also incorporated to help achieve global awareness. Furthermore, consecutive matching along an observation sequence may lead to the accumulation of error, on top of the coarse estimation through correspondence matching, we further use a memory-augmented coarse-to-fine module which uses both target grasp features and historical grasp features to refine the grasp tracking results for better temporal smoothness and semantic consistency.", + "bbox": [ + 75, + 318, + 472, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on two benchmarks to evaluate our method and demonstrate its effectiveness. The results show that our method outperforms two representative baseline methods. We also conduct real robot experiments on both single-object scenes and cluttered scenes. We report success rates of $81.25\\%$ for single-object scenes and $81.67\\%$ for multi-object scenes.", + "bbox": [ + 75, + 681, + 470, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 805, + 227, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Grasping in Static Scenes", + "text_level": 1, + "bbox": [ + 76, + 830, + 308, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the advances in the field of 3D perception [2, 25, 26], 6-DoF static grasping which takes point cloud as input is gaining increasing attention in both the", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "research community and industry. In general, there are two lines of work that have been explored. The first line adopts a sampling-evaluation manner [15, 18, 23, 34]. Grasp candidates are sampled or generated on the point clouds and then their qualities are evaluated by neural networks. The second line processes the point cloud of the entire scene and predicts grasp poses across the scene [9, 10, 24, 27, 36] in an end-to-end manner. Compared to the first line, the end-to-end strategy achieves better balance in terms of the speed-accuracy tradeoff. Some researchers also propose to predict actions in continuous space [14, 32, 38]. Nevertheless, these methods can be adversely impacted by distribution shifts. Our method uses a pre-trained grasp pose detector from [9] which belongs to the second research line.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Reactive Grasping", + "text_level": 1, + "bbox": [ + 500, + 323, + 679, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though reactive grasping remains much less explored, a small amount of literature has partially investigated this more difficult problem. Most of these papers rely on object motion tracking. [1] predicts object motion and adaptively plans the grasp for a reachable grasp candidate. However, the object's motion is limited to few prototypes and thus it cannot handle the case where the objects move unpredictably. [19] tracks the objects' poses, yet the grasp planning phase is limited by a set of fixed grasp trajectories. [12,20,28] requires prior knowledge of the objects which is not always available in the real-world.", + "bbox": [ + 496, + 349, + 893, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some attempts on reactive grasping without prior information have been conducted recently. [22] picks the nearest grasp poses across frames. [42] samples grasp candidates by adding disturbances to grasp poses from the previous frame and evaluates their qualities in the current frame's scene. These methods focused on the smoothness of the grasp pose sequence but did not impose semantic constraints. Their predicted grasp poses can switch between different objects. Thus they mainly demonstrate single object grasping, and [42] requires an extra human hand segmentation module to avoid grasping the human's hand during handover while our correspondence matching module mitigates such a problem since grasps on hands and grasps on objects have low correspondence scores. [8] proposed a temporal association module to alleviate this dilemma. However, it only considers correspondence across two frames which is prone to error accumulation. In this paper, we further adopt a refinement module to adjust the grasp pose according to the target and historical grasp poses.", + "bbox": [ + 496, + 518, + 895, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some other works adopt continuous action prediction [14, 32]. Even if it is not designed for dynamic environments, it can be executed on moving objects thanks to the closed-loop controller. These methods may suffer greatly from distribution shifts. Different from these methods, our method tracks grasp poses explicitly.", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8825", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6b513f18b4c1ad0ff11ab235f415d64b0b077abf86eaf78b5837e6418bc3198.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed method for target-referenced reactive grasping. Given a target grasp and colored point clouds of a new coming frame with $N$ points, our pipeline first detects $M$ grasp candidates of size $M \\times 12$ (3 for translation and 9 for rotation) across the scene. It then extracts grasp features of size $M \\times 4C$ for each detected grasp by cropping and embedding the points along the grasping direction. Furthermore, grasp features aggregation is used to estimate corresponding grasps between the previous frame and the current frame. After filtering, the correspondence score matrix produces $K$ valid pairs of grasp. The predicted corresponding grasps are then fed into a refinement network along with the target grasp and other frames selected from a memory pool to be updated toward the target. Contents in the memory pool are updated by the refined values as well." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 861, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Local Feature Matching in Computer Vision", + "text_level": 1, + "bbox": [ + 76, + 488, + 452, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike global features which summarize the entire scene, local features distribute densely across the scene and describe interesting regions. Under slightly different camera views, local feature matching is aimed at recognizing distinctive regions and establishing associations in images or point clouds. The mainstream of local feature matching includes detector-based matching and detector-free matching. For detector-based matching, classical hand-crafted features such as SIFT [17] and ORB [29] are widely adopted. With the development of deep learning, learned local features [5, 6] achieve satisfying results as well. Different from detector-based methods that extract sparse local features, detector-free methods establish pixel-wise or pointwise dense features [3, 13, 16, 31, 33]. With the extracted local features, nearest neighbor search or learning-based approach [30, 39] are often used in the matching phase. In our task, we consider each grasp brings rich geometric and visual patterns of the grasped local patch and our grasp correspondence are built upon these local features.", + "bbox": [ + 75, + 512, + 472, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 814, + 284, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first briefly introduce some notations and metrics for grasp pose and grasp distance in 3D space, followed by formulating the problem of target-referenced reactive grasping.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define a grasp pose $\\mathcal{G}$ as", + "bbox": [ + 519, + 489, + 705, + 505 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = (\\mathbf {R} \\mathbf {t}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 513, + 890, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ denotes the rotation of the grasp pose and $\\mathbf{t} \\in \\mathbb{R}^{3 \\times 1}$ denotes the translation. Consider two arbitrary grasps $\\mathcal{G}_1$ and $\\mathcal{G}_2$ , a grasp distance measure is defined as", + "bbox": [ + 498, + 537, + 892, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\left(\\mathcal {G} _ {1}, \\mathcal {G} _ {2}\\right) = \\Delta \\mathbf {R} + \\Delta \\mathbf {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 590, + 746, + 607 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {R} = \\operatorname {a r c c o s} \\frac {\\operatorname {t r a c e} \\left(\\mathbf {R} _ {1} ^ {\\top} \\mathbf {R} _ {2}\\right) - 1}{2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 611, + 890, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {t} = | | \\mathbf {t} _ {1} - \\mathbf {t} _ {2} | |.\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 643, + 697, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then we make a formal statement of target-referenced reactive grasping based on the background definition above. Given $\\mathcal{G}^1$ at first frame as reference grasp, for any timestep $j > 1$ , target-referenced reactive grasping is aimed at finding the grasp $\\mathcal{G}^j$ that minimizes the grasp distance:", + "bbox": [ + 496, + 667, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {j *} = \\underset {\\mathcal {G} ^ {j}} {\\arg \\min } \\mathcal {D} \\left(\\mathcal {G} ^ {j}, \\mathbf {T} ^ {j 1} \\mathcal {G} ^ {1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 750, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{T}^{j1}\\in \\mathbb{R}^{4\\times 4}$ means the ground-truth transformation matrix of the reference grasp from the initial frame to the frame at timestep $j$ in the camera's coordinate system.", + "bbox": [ + 496, + 786, + 890, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 844, + 589, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail our grasp tracking pipeline, as illustrated in Fig.2. For a target grasp, we first detect grasp", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8826", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2880805607c2eb692469878d5fbe94438f9cce41199c20fc4bccfc6bfd82d99e.jpg", + "image_caption": [ + "Figure 3. Visualization of grasp features representation." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 455, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "candidates across the scene. Then we discover strong correspondences between the target grasp and grasp candidates by attentional grasp features aggregation. We finally use a memory-augmented refinement network to correct the predictions towards the target grasp.", + "bbox": [ + 75, + 383, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Grasp Detector", + "text_level": 1, + "bbox": [ + 76, + 470, + 232, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each incoming new frame, we generate a grasp set along with it. This work uses an off-the-shelf GraspNet baseline [9] for proposals of grasp candidates. GraspNet is a learned grasp pose detector that takes the partial point cloud of the scene $\\mathcal{P}$ as input and outputs a grasp set $\\mathbf{G}$ across the entire scene. In GraspNet, a point cloud $\\mathcal{P}$ of size $N\\times 3$ is downsampled and transformed to feature vectors of size $M\\times C$ by the PointNet++ backbone network [26] where $C$ denotes the channel dimension of the features and $M$ is the number of points after farthest point sampling [7]. For each sampled point, GraspNet predicts a grasp $\\mathcal{G}$ and its confidence score. The detected grasp set is then sent to the following steps for grasp tracking. While GraspNet performs well in this work, the following steps are not limited to any specific grasp detector.", + "bbox": [ + 75, + 494, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Grasp Correspondence Network", + "text_level": 1, + "bbox": [ + 76, + 733, + 362, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the formulation of grasp correspondence in [8]. A many-to-many correspondence matrix is learned during training, given two grasp sets in consecutive frames. During testing, we can choose the grasp pose with the highest correspondence score with the target pose in the previous frame as the tracked pose.", + "bbox": [ + 75, + 757, + 468, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grasp Features Representation First we illustrate how we represent grasp features. For $M$ grasp poses in a scene,", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we follow [8] to extract seed features, geometric features and appearance features for them, each having a shape of $M \\times C$ . In addition, since it is quite common in the real world that multiple similar objects exist simultaneously in the scene, grasps on similar objects cannot be distinguished in such situations which results in correspondence switching between objects. Hence we further add global scene features of size $M \\times C$ to provide information about the scene state. This global feature is extracted by max-pooling pointwise features from the backbone output, thus is the same for each of $M$ grasp poses. Finally, we concatenate geometric features, visual features, seed point features and global scene features to construct grasp features of dimension $4C$ for each grasp, denoted by $\\mathbf{x}$ . We show the above process in Fig.3. We present more details of grasp features representation in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positional Information To provide positional information for grasps, we encode grasp poses by MLP layers of size $(12, 2C, 4C)$ . These feature vectors are then added to the raw grasp features $\\mathbf{x}$ , such that grasps can be treated in a similar way to words in a sentence.", + "bbox": [ + 496, + 354, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} \\leftarrow \\mathbf {x} + \\mathbf {M L P} ^ {(1)} (p), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 445, + 890, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ is the grasp pose of dimension 12, consisting of 9 parameters for the rotation matrix and 3 for translation. We use this feature as the final grasp features representation. Such a process makes our method position-dependent and benefits following the grasp features aggregation step.", + "bbox": [ + 496, + 472, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attentional Grasp Features Aggregation Thanks to the above-mentioned grasp features, it is now straightforward to reason about the visual, geometric and positional properties of grasps jointly. For a grasp $\\mathcal{G}_i$ , we first embed its associated grasp features using MLP layers and transform its feature size from $4C$ to $2C$ by MLP layers.", + "bbox": [ + 496, + 570, + 890, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {g _ {i}} = \\mathbf {M L P} ^ {(2)} \\left(\\mathbf {x} _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 676, + 890, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from the features of a grasp itself, given a pair of grasp sets, it is intuitive that interacting with contextual grasps within grasp sets or across grasp sets is critical to reduce ambiguities and increase the distinctiveness of grasps as well. [8] adopts vanilla MLP blocks which cannot model long-range relationships among grasp poses. In this paper, we take advantage of the global receptive field of the attentional graph neural network (GNN) to model such dependencies. Following [30], self-attention and cross-attention [35] are adopted for context aggregation within grasp sets and across grasp sets. In particular, features are aggregated unidirectionally in self-attention and bidirectionally in cross-attention. Given two grasp sets $\\mathbf{G}_1$ and $\\mathbf{G}_2$", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8827", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that each consists of $M$ grasps, grasp features aggregation generates matching descriptors $\\mathbf{f}^{m_1}$ , $\\mathbf{f}^{m_2} \\in \\mathbb{R}^{M \\times 2C}$ .", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we outline the workflow of grasp features aggregation. In self-attention, we only process grasps within $\\mathbf{G}_1$ or $\\mathbf{G}_2$ . Taking $\\mathbf{G}_1$ as an example, grasp features $\\mathbf{f}^{\\mathbf{G}_1}$ of size $(M \\times 2C)$ are first mapped to three elements which are named query, key and value in convention by learnable weights $\\mathbf{W}_{\\mathbf{Q}}$ , $\\mathbf{W}_{\\mathbf{K}}$ and $\\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{2C \\times 2C}$ . The attention computation can then be written as $\\mathbf{A} = (\\mathbf{Q} \\cdot \\mathbf{K}^T) / \\sqrt{C}$ where $\\mathbf{A}$ denotes the attention matrix and has a size of $(M \\times M)$ and $(\\cdot)$ is the dot product operator. $\\mathbf{A}$ actually measures the similarity between query and key. After that, the message propagation step can be defined as $\\mathbf{M} = \\mathrm{softmax}(\\mathbf{A}) \\cdot \\mathbf{V}$ . $\\mathbf{M}$ is later fed into MLP layers of size $(2C, 4C, 2C)$ to produce the output features which have the same size as $\\mathbf{f}^{\\mathbf{G}_1}$ . Since $\\mathbf{M}$ can be regarded as the sum of value vectors weighted by the similarity measure, therefore the output feature vectors represent correspondences among grasps in $\\mathbf{G}_1$ . In cross-attention, we consider grasp features from both $\\mathbf{G}_1$ and $\\mathbf{G}_2$ . To produce the output feature vector, the only thing different is that key and value are no longer generated from $\\mathbf{f}^{\\mathbf{G}_1}$ but from $\\mathbf{f}^{\\mathbf{G}_2}$ . Intuitively, in self-attention information flows only within $\\mathbf{G}_1$ or $\\mathbf{G}_2$ whereas in cross-attention information flows from $\\mathbf{G}_2$ to $\\mathbf{G}_1$ or vice versa. We consider the output of cross-attention layers as a representation of correspondence between $\\mathbf{G}_1$ and $\\mathbf{G}_2$ . We interleave self-attention and cross-attention layers by $N_c$ times and get the matching descriptors $\\mathbf{f}^{m_1}, \\mathbf{f}^{m_2}$ which are later used to compute correspondences within and across grasp sets.", + "bbox": [ + 75, + 122, + 472, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondence Estimation Given $\\mathbf{f}^{m_1},\\mathbf{f}^{m_2}$ , we can simply compute the correspondence matrix $S\\in \\mathbb{R}^{M\\times M}$ For each pair of grasps from these two sets, the correspondence score can be calculated by cosine similarity", + "bbox": [ + 75, + 568, + 468, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (i \\cdot j) = \\frac {\\mathbf {f} _ {i} ^ {m _ {1}} \\cdot \\mathbf {f} _ {j} ^ {m _ {2}}}{\\left| \\left| \\mathbf {f} _ {i} ^ {m _ {1}} \\right| \\right| _ {2} \\cdot \\left| \\left| \\mathbf {f} _ {j} ^ {m _ {2}} \\right| \\right| _ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 640, + 468, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since cosine similarity is naturally normalized, it makes training more stable. Taking advantage of contrastive learning, during training a target grasp can be assigned the same class to any number of grasps in the other grasp set. This is important because some grasps are close in terms of translation and some are close in terms of rotation. During training, our method allows parallel prediction of multiple target grasps. We simply pick grasps that have the highest correspondence scores with respect to target grasps. Before entering the next stage, a grasp set is filtered to $K$ grasps by removing upward grasps and grasps with low correspondence scores with respect to all other grasps. If the number of grasps is less than $K$ , we repeat the first grasp. In real robot experiments, only one target grasp is used.", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Memory-augmented Refinement Network", + "text_level": 1, + "bbox": [ + 498, + 90, + 856, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the correspondence is estimated between fixed grasp sets, the tracking performance greatly relies on the grasp detector which is not guaranteed to generate identical grasps across different frames. Also, consecutive correspondence matching may lead to an accumulation of errors. Therefore, to further improve the semantic consistency with the reference grasp pose, a refinement network is used to correct predictions made by correspondences towards the target grasp.", + "bbox": [ + 496, + 113, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be specific, we store the tracked grasp poses on past frames and their associated features in a memory pool. For each forward pass, the refinement network uses historical observations selected from the memory pool to refine the parameters of the selected grasp pose in the current frame. Here the historical frames from the memory pool are selected according to their time indexes. In practice, we select the reference pose in the first frame and apply uniform sampling for simplicity in the interval $(1,t - 1]$ to select past frames, where $t$ denotes the current time index.", + "bbox": [ + 496, + 250, + 892, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a $L$ frame grasping sequence that consists of 1 frame of tracking target, 1 current frame, and $L - 2$ frames selected from the memory pool, grasp features of size $L \\times (4C)$ are first transformed to $L \\times 2C$ by MLP layers. After that, we simply repeat the target frame and concatenate it with other frames, resulting in features of size $(L - 1) \\times (2C)$ . Such that all the selected frames are conditioned on the tracking target. We then apply temporal encoding which is similar to positional encoding introduced in [35]. These features are then fed into a refinement network which outputs $(L - 1)$ refined poses. Note that both historical grasp poses and the grasp pose in the current frame are updated. This refinement network has the same architecture as the grasp correspondence network, except that only self-attention layers are adopted. Moreover, the rotation is parameterized by the 6D representation introduced in [43]. Causal mask is used to avoid causality confusion as well.", + "bbox": [ + 496, + 402, + 892, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Supervision", + "text_level": 1, + "bbox": [ + 500, + 667, + 627, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective function used to train the above pipeline can be divided into two parts:", + "bbox": [ + 500, + 691, + 890, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c} + \\beta \\mathcal {L} _ {R, t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 734, + 890, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c$ denotes the loss for correspondence learning and $\\mathcal{L}_{R,t}$ denotes the loss for grasp pose refinement. $\\beta$ is the weighting term which we set as 1 during training.", + "bbox": [ + 498, + 763, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $\\mathcal{L}_c$ , we follow [8] and adopt supervised contrastive loss. For any two grasp poses with positive correspondence label, we denote them as $\\mathrm{corr}(\\mathcal{G}_i,\\mathcal{G}_k) = 1$ or vice versa. Given two arbitrary grasp sets $\\mathbf{G}_1$ and $\\mathbf{G}_2$ , we first locate the positive grasp set $\\mathbf{P}(i) = \\{\\mathcal{G}_k^2\\in \\mathbf{G}_2|\\mathrm{corr}(\\mathcal{G}_i,\\mathcal{G}_k) = 1\\}$ for any $\\mathcal{G}_i\\in \\mathbf{G}_1$ . This correspondence loss can then be", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8828", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "written as:", + "bbox": [ + 76, + 92, + 150, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\sum_ {\\mathcal {G} _ {i} ^ {1} \\in \\mathbf {G} _ {1}} \\frac {- 1}{| \\mathbf {P} (i) |} \\sum_ {\\mathcal {G} _ {k} ^ {2} \\in P (i)} \\log \\frac {\\exp (\\mathcal {S} (i \\cdot k) / \\tau)}{\\sum \\mathcal {G} _ {j} ^ {2} \\in \\mathbf {G} _ {2} \\exp (\\mathcal {S} (i \\cdot j) / \\tau)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 113, + 467, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $|\\mathbf{P}(i)|$ denotes the cardinality and $\\tau$ is a scalar temperature parameter. This loss pulls together the positive grasp pairs in embedding space, whereas pushes apart negative grasp pairs meanwhile.", + "bbox": [ + 76, + 167, + 468, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For grasp pose refinement, we adopt a transformation loss which consists two terms for optimizing rotation and translation respectively. Consider $\\mathcal{J} = \\{j_1,j_2,\\dots ,j_n\\}$ that stores arbitrary number of time index, for any time step $j\\in \\mathcal{J}$ , given target grasp $\\mathcal{G}^1$ , we first compute its transformed grasp in $j^{th}$ frame's coordinates, denoted by", + "bbox": [ + 76, + 228, + 468, + 319 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {j 1} = \\left(\\mathbf {R} ^ {j 1} \\mathbf {t} ^ {j 1}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 328, + 468, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then define the transformation loss as:", + "bbox": [ + 76, + 357, + 354, + 372 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {R, t} = \\frac {1}{| \\mathcal {I} |} \\sum_ {j \\in \\mathcal {J}} \\frac {1}{| \\mathbf {G} _ {j} |} \\sum_ {\\mathcal {G} _ {i} ^ {j} \\in \\mathbf {G} _ {j}} | | (\\mathbf {R} ^ {*}) ^ {- 1} \\mathbf {R} ^ {j 1} - \\mathbf {I} | | _ {2} (10) \\\\ + \\frac {1}{| \\mathcal {I} |} \\sum_ {j \\in \\mathcal {J}} \\frac {1}{| \\mathbf {G} _ {j} |} \\sum_ {\\mathcal {G} _ {i} ^ {j} \\in \\mathbf {G} _ {j}} \\| \\mathbf {t} ^ {j 1} - \\mathbf {t} ^ {*} \\| _ {2}, (10) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 378, + 468, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_j$ denotes grasp set at time step $j$ .", + "bbox": [ + 76, + 474, + 356, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 503, + 207, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 527, + 171, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed pipeline is trained on GraspNet-1Billion dataset [9] which is a large-scale real-world dataset for grasp detection. It includes 89 objects, 190 scenes and 256 camera views for each scene. We follow [8] to generate grasp correspondence labels across viewpoints on the GraspNet-1Billion dataset and refer readers to [8] for more details. In addition, we further collect 30 scenes of moving objects for testing. For each frame, we record the RGBD images and objects' 6D poses. We manually annotate 10 grasps for the first frame of each scene as targets and further project the annotated grasps according to the 6D poses for the future frames. This test set is named Moving GraspNet.", + "bbox": [ + 76, + 550, + 468, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Pre-processing and Augmentation", + "text_level": 1, + "bbox": [ + 76, + 756, + 372, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before being fed into the network, the point cloud is first downsampled to 20000 points and the RGB values are normalized. We also remove points outside a pre-defined workspace. Furthermore, we adopt massive data augmentation to avoid over-fitting and enrich objects' motion. For each object in the scene, its point cloud is augmented on-the-fly by a random translation of $[-0.2, 0.2]$ meter in all XYZ directions and a random rotation of $[-30, 30]$ degrees", + "bbox": [ + 76, + 779, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "around z-axis. Moreover the scene's point cloud is rotated by $[-30, 30]$ degrees around both x-axis and z-axis. Due to the space limitations, please refer to the supplementary materials for details about network implementation.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Performance of Grasp Tracking", + "text_level": 1, + "bbox": [ + 500, + 164, + 781, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric We propose a new Multiple Grasp Pose Tracking (MGTA) metric which follows the idea of Multiple Object Tracking Accuracy (MOTA) [21] in multi-object tracking task, except that the similarity measure between two grasps is computed using Eq.(2). Details are given in supplementary materials. Apart from MGTA, we also report the mean translation error and rotation error to the tracking target of each sequence.", + "bbox": [ + 498, + 188, + 890, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the same evaluation procedure on GraspNet1-Billion and Moving GraspNet. For each scene, we pick 10 reference grasps in the first frame which are later used for grasp tracking. To be specific, these reference grasps are detected by a grasp detector on GraspNet1-billion but manually labeled on Moving GraspNet. For each remaining frame, we compute the ground truth grasps using objects' 6D poses and segmentation mask of the scene. The metrics are calculated using the transformed reference grasps and predictions associated with each frame.", + "bbox": [ + 498, + 310, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing with Representative Methods We compare our methods with two other baseline methods. A heuristic method is implemented as the first baseline. Given a grasp pose from the previous frame, its nearest neighbor detected in the current camera coordinates is picked. For the second method, we adopt Bundle Track [40] which is one representative method in the field of unseen object 6D pose tracking. It fits our task setting well since it requires no objects' CAD model.", + "bbox": [ + 498, + 484, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As it is shown in Tab.1 and Tab.2, our method outperforms baseline methods on both GraspNet-1Billion and Moving GraspNet. It is found in experiments that, in many cases, grasps cannot fall on the same objects if we only pick the nearest neighbors. Also, it can be drawn from the results that tracking results are improved significantly after refinement, especially for rotation. Note that MGTA will be negative if there is a large number of false positives, false negatives, and ID switches.", + "bbox": [ + 498, + 622, + 890, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Real Robot Experiments", + "text_level": 1, + "bbox": [ + 500, + 770, + 723, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In real robot experiments, we use Flexiv Rizon robot arm with an Intel Realsense L515 depth camera mounted on the end-effector. A Robotiq-85 parallel-jaw gripper is used with 3D-printed extended fingertips attached to it. The real robot experiments run on an NVIDIA 3090 GPU at around 10 fps. Please refer to the supplementary materials for details of the objects.", + "bbox": [ + 498, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8829", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0d71b126b5b87bd840d8b3b4a98b1b416eeaf75b4a855098d59b2809b8bb4959.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSeenSimilarNovel
MGTA↑n (cm)↓n (°)↓MGTA↑n (cm)↓n (°)MGTA↑n (cm)↓n (°)↓
Nearest-0.8115.295.98-0.8015.497.61-0.8316.696.90
Bundle Track [40]0.741.5215.800.721.4716.370.721.9315.41
Ours, w/o refinement-0.322.3153.24-0.302.6857.59-0.392.9358.49
Ours, M=5120.801.5810.890.801.7410.620.752.1015.03
Ours, M=10240.841.3410.150.851.529.840.781.9012.96
Ours, M=20480.851.279.820.861.359.500.791.8411.48
", + "bbox": [ + 78, + 88, + 903, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b40619c72de1aebe5a2ca51abed56c6053ba4ed9f828fbd8e39f513154b090d2.jpg", + "image_caption": [ + "Figure 4. Illustration of tracking a toy dog and toy rabbit. We can see that the grasp poses are consistent, i.e., on the head of the dog and the ear of the rabbit. Yellow bounding box represents to the view limit of the camera. No box means the camera can see the entire scene." + ], + "image_footnote": [], + "bbox": [ + 104, + 255, + 861, + 526 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05d0391cb38cf3d070e1607d3c568da8df411e4a81e7212cd4b276aebd0de1ed.jpg", + "table_caption": [ + "Table 1. GraspNet-1Billion evaluation results on seen, similar and novel objects respectively. $M$ denotes the number of grasp detected,its default value is 1024." + ], + "table_footnote": [], + "table_body": "
MethodMGTA↑n (cm)↓n (°)↓
Nearest-0.478.0967
Bundle Track0.377.6023.52
Ours, w/o refinement-0.436.2462.33
Ours, M=5120.475.5322.63
Ours, M=10240.524.6020.37
Ours, M=20480.544.1418.28
", + "bbox": [ + 78, + 574, + 475, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Protocol For each frame, we receive RGBD images from the in-hand camera. We transform the point cloud to the camera coordinate system at ready state by the camera extrinsic for simplicity as well. Due to safety considerations, points outside a pre-defined workspace will be masked. After a grasp candidate generation phase, we select the grasp with the highest predicted score as a reference grasp and keep tracking it for each new frame. During the tracking phase, the robot moves to a pre-grasp pose which is 0.02", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6860b2b023cee55a6458b9fc092d7fb46367180379eb3c1d177f75e42e18f608.jpg", + "table_caption": [ + "Table 2. Moving GraspNet evaluation results. All objects in Moving GraspNet are novel. $M$ denotes the number of grasp detected, its default value is 1024." + ], + "table_footnote": [], + "table_body": "
Object ID#Attempt#SuccessSuccess Rate
11080.8
21090.9
31080.8
41070.7
61080.8
71080.8
91080.8
101090.9
Total80650.8125
", + "bbox": [ + 531, + 574, + 862, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Real robot experiments results on single object scenes.", + "bbox": [ + 504, + 744, + 883, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "meters higher than the actual grasp pose. The robot executes the grasp once three conditions are met simultaneously: 1) the translation distance between the gripper center and the target grasp in XY-plane is lower than 0.01 meter (the XYplane is parallel to the table); 2) the translation distance between the gripper center and the target grasp in 3D space is lower than 0.04 meters; 3) the rotation distance between the gripper and the target grasp in 3D space is lower than", + "bbox": [ + 496, + 780, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8830", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cbd3254b9f312b3045a5ed6916a8bc66e96fcde4a064eeb0b84449f41e2fb68c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Object ID#Attempt#SuccessSuccess Rate
7,7,8,81080.8
10,13,141090.9
1,11,15,16,17,181080.8
2,3,51090.9
9,121070.7
20,201080.8
Total60490.8167
", + "bbox": [ + 84, + 88, + 460, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b73592b9ec5cfb3a4256d3b6323b9d9bc6d943b52b8c4b8417283762fa02a948.jpg", + "image_caption": [ + "Figure 5. Correspondence switching between similar objects. Red arrow indicates correct correspondence and blue arrow indicates wrong correspondence." + ], + "image_footnote": [], + "bbox": [ + 86, + 258, + 460, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "15 degrees. The initial positions of the objects are chosen randomly.", + "bbox": [ + 76, + 577, + 468, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results As shown in Tab.3 and Tab.4, we execute 20 grasp attempts for each scene. In single-object scenes, we randomly sample 8 objects for testing and we achieve a success rate of 81 percent. In multi-object scenes, we only choose objects with similar visual and geometric patterns to test the temporal smoothness and semantic consistency of the predicted grasps. Results suggest that single-object scenes and multi-object scenes have nearly the same success rate which further proves the effectiveness of our method. We illustrate one case from real robot experiments in Fig.4.", + "bbox": [ + 75, + 626, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Analysis", + "text_level": 1, + "bbox": [ + 76, + 786, + 181, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Global Features In Sec.4.2, we incorporate global features to represent grasp features. While local grasp features are sufficient to identify corresponding local structures, grasp trials may fail due to correspondence switches among objects with similar visual and geometric patterns as we can see in Fig.5. Such phenomenon may be", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/843dfe2383abc7bbffb81c12e96f1d0bc240e6d6729927b3eb683172caea49ce.jpg", + "table_caption": [ + "Table 4. Real robot experiments results on multi-object scenes." + ], + "table_footnote": [], + "table_body": "
Object IDMethod#Attempt#SW
7,7,8,8w/o global features105
Full model101
1,17,17w/o global features104
Full model100
", + "bbox": [ + 506, + 88, + 887, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Statistics of Correspondence Switching. SW represents correspondence switching.", + "bbox": [ + 498, + 178, + 890, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "due to lacking awareness of the scene state and we propose to solve it by global features as what we state in Sec.4.2. In order to analyze the actual effectiveness of global features, we further record how frequent correspondence switching happens with or without global features. We conduct these experiments on identical objects which are placed close to each other in the scenes. Tab.5 suggests that the number of correspondence switches decreased significantly with global features.", + "bbox": [ + 496, + 220, + 890, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure Analysis There are three major types of failure cases found in real robot experiments. First, some predicted grasp candidates from the detector do not have good qualities. Naturally, tracking such targets may lead to failed grasps. Second, it is hard to heuristically plan the timing to close the gripper. Due to some grasps may fall on the edges of objects, it is common that the objects may slip away immediately when the robot receives the command to close the gripper. Furthermore, since we mount the camera on the wrist, the point cloud may become fragmented due to view limitation and occlusion, especially when the objects fall behind the gripper. In such cases, the predictions can be inaccurate.", + "bbox": [ + 496, + 375, + 892, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 617, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel solution to reactive grasping in a target-referenced setting. For a reference grasp, our method predicts temporally smooth and semantically consistent grasp poses in the future frames without any prior knowledge of instance. Experiments conducted on both single-object and multi-object scenes show that our method provides reliable grasp plans under various environments.", + "bbox": [ + 496, + 611, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the National Key R&D Program of China (No.2021ZD0110700), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Qi Zhi Institute, and Shanghai Science and Technology Commission (21511101200).", + "bbox": [ + 496, + 733, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contributions: J. Liu and H-S. Fang initiated the project. J. Liu designed and trained the network and conducted the evaluation together with R. Zhang. J. Liu wrote the paper. H-S. Fang devised and mentored this project and edited the paper. M. Gou, H. Fang, and S. Xu constructed the testing dataset. C. Wang and H. Yan helped with the evaluation. C. Lu supervised the project and provided hardware support.", + "bbox": [ + 496, + 803, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8831", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iretiayo Akinola, Jingxi Xu, Shuran Song, and Peter K. Allen. Dynamic grasping with reachability and motion awareness. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9422-9429, 2021. 2", + "[2] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2", + "[3] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3", + "[4] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 2", + "[5] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Toward geometric deep slam. ArXiv, abs/1707.07410, 2017.3", + "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 337-33712, 2018. 3", + "[7] Y. Eldar, M. Lindenbaum, M. Porat, and Y.Y. Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 4", + "[8] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. arXiv preprint arXiv:2212.08333, 2022. 2, 4, 5, 6", + "[9] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4, 6", + "[10] Minghao Gou, Hao-Shu Fang, Zhanda Zhu, Sheng Xu, Chenxi Wang, and Cewu Lu. Rgb matters: Learning 7-dof grasp poses on monocular rgbd images. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13459-13466. IEEE, 2021. 2", + "[11] Yinlin Hu, Joachim Hugonot, Pascal Fua, and Mathieu Salzmann. Segmentation-driven 6d object pose estimation. In CVPR, 2019. 2", + "[12] Seungsu Kim, Ashwini Shukla, and Aude Billard. Catching objects in flight. IEEE Transactions on Robotics, 30(5):1049-1065, 2014. 2", + "[13] Donghoon Lee, Onur C. Hamsici, Steven Feng, Prachee Sharma, and Thorsten Gernoth. Deeppro: Deep partial point cloud registration of objects. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5663-5672, 2021. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International journal of robotics research, 37(4-5):421-436, 2018. 2", + "[15] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Görner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In 2019 International Conference on Robotics and Automation (ICRA), pages 3629-3635. IEEE, 2019. 2", + "[16] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2011. 3", + "[17] David Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91–, 11 2004. 2, 3", + "[18] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26), 2019. 2", + "[19] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Aleš Leonardis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43(5):1241-1256, 2019. 2", + "[20] Arjun Menon, Benjamin Cohen, and Maxim Likhachev. Motion planning for smooth pickup of moving objects. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 453-460. IEEE, 2014. 2", + "[21] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. ArXiv, abs/1603.00831, 2016. 6", + "[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 1, 2", + "[23] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2901-2910, 2019. 2", + "[24] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3619-3625, 2020. 2", + "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[26] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4", + "[27] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4g: Amodal single-view single-shot se (3) grasp" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8832", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection in cluttered scenes. In Conference on robot learning, pages 53-65. PMLR, 2020. 2", + "[28] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2021. 2", + "[29] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3", + "[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 4", + "[31] Tanner Schmidt, Richard Newcombe, and Dieter Fox. Self-supervised visual descriptor learning for dense correspondence. IEEE Robotics and Automation Letters, 2(2):420-427, 2017. 3", + "[32] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 2", + "[33] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3", + "[34] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt. Grasp pose detection in point clouds. The International Journal of Robotics Research (IJRR), 36(13-14):1455-1473, 2017. 2", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4, 5", + "[36] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15964-15973, 2021. 2", + "[37] Chen Wang, Roberto Martín-Martín, Danfei Xu, Jun Lv, Cewu Lu, Li Fei-Fei, Silvio Savarese, and Yuke Zhu. 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10059–10066, 2020. 2", + "[38] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 2", + "[39] Yue Wang and Justin M. Solomon. Deep closest point: Learning representations for point cloud registration. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3", + "[40] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074. IEEE, 2021. 2, 6, 7", + "[41] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E. Bekris. se(3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373, 2020. 2", + "[42] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 1, 2", + "[43] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 5" + ], + "bbox": [ + 501, + 92, + 890, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8833", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_model.json b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_model.json new file mode 100644 index 0000000000000000000000000000000000000000..722ecfe0c4a912df75c7dcdde745854757b38100 --- /dev/null +++ b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.191, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "Target-referenced Reactive Grasping for Dynamic Objects" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.812, + 0.269 + ], + "angle": 0, + "content": "Jirong Liu\\(^{1,2}\\), Ruo Zhang\\(^{2}\\), Hao-Shu Fang\\(^{2}\\), Minghao Gou\\(^{2}\\), Hongjie Fang\\(^{2}\\), Chenxi Wang\\(^{2,3}\\), Sheng Xu\\(^{2}\\), Hengxu Yan\\(^{2}\\), Cewu Lu\\(^{1,2\\dagger}\\) \n\\(^{1}\\)Shanghai Qi Zhi institute, \\(^{2}\\)Shanghai Jiao Tong University, \\(^{3}\\)Flexiv Robotics, LTD jirong@sjtu.edu.cn, {ruozhang0608, fhaoshu}@gmail.com, {gmh2015, galaxies, wcx1997, xs1020, hengxuyan, lucewu}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.669 + ], + "angle": 0, + "content": "Reactive grasping, which enables the robot to successfully grasp dynamic moving objects, is of great interest in robotics. Current methods mainly focus on the temporal smoothness of the predicted grasp poses but few consider their semantic consistency. Consequently, the predicted grasps are not guaranteed to fall on the same part of the same object, especially in cluttered scenes. In this paper, we propose to solve reactive grasping in a target-referenced setting by tracking through generated grasp spaces. Given a targeted grasp pose on an object and detected grasp poses in a new observation, our method is composed of two stages: 1) discovering grasp pose correspondences through an attentional graph neural network and selecting the one with the highest similarity with respect to the target pose; 2) refining the selected grasp poses based on target and historical information. We evaluate our method on a large-scale benchmark GraspNet-1Billion. We also collect 30 scenes of dynamic objects for testing. The results suggest that our method outperforms other representative methods. Furthermore, our real robot experiments achieve an average success rate of over 80 percent. Code and demos are available at: https://graspnet.net/reactive." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.811 + ], + "angle": 0, + "content": "Reactive grasping is in great demand in the industry. For instance, in places where human-robot collaboration is heavily required like factories, stress on laborers will be significantly relieved if robots can receive tools from humans and complete the harder work for laborers. Such a vision is based on reactive grasping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.47, + 0.856 + ], + "angle": 0, + "content": "On the contrary to static environments, in reactive grasping, dynamic task setting poses new challenges for algorithm design. Previous research in this area mainly focuses" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.302, + 0.877, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.441, + 0.872, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Figure 1. (a) Classic reactive grasping guarantees the smoothness of the grasp poses but cannot predicts grasps on the same part of the hammer. (b) Our target-referenced reactive grasping takes semantic consistency into consideration. The generated grasp poses across frames are illustrated with blue grippers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.903 + ], + "angle": 0, + "content": "on planning temporally smooth grasps [22, 42] to avoid wavy and jerky robot motion. Few of them pay attention to its semantic consistency. In short, given a targeted grasp at the first frame, we want the robot to grasp the same part of the object in the following frames. Additionally, it is not guaranteed that grasp predictions made by classical methods fall on the same object in cluttered scenes. Hence, most of their experiments are conducted on single-object scenes. Unlike previous works, this work is aimed at achieving temporally smooth and semantically consistent reactive grasping in clutter given a targeted grasp. We refer to such a task setting as target-referenced reactive grasping as shown in Fig.1. Note that despite robot handover is a major application scenario of reactive grasping, this work focuses on a more general task setting - dynamic object grasping." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "† Cewu Lu is the corresponding author, a member of Qing Yuan Research Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8824" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "A naive idea to solve this task is to generate reference grasp poses for the initial scene and consecutively track the object's 6D pose. As the object moves, the initial grasp pose can be projected to a new coming frame based on the object's 6D pose. Although such an idea seems to be natural and valid, some bottleneck greatly degrades its viability. First of all, the solution to reactive grasping should be able to handle objects' motion in real-time, meaning that it requires fast inference speed and immediate response to continuous environmental changes. However, 6D pose tracking may be time-consuming due to commonly-used instance segmentation [11,40]. Second, 6D pose tracking usually requires objects' prior knowledge, such as CAD models [4,41], which is not always available in the real world as well or achieves only category-level generalization [37]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.682 + ], + "angle": 0, + "content": "Different from tracking objects, we propose to track grasps by a two-stage policy instead. We also comply with the restriction that no prior knowledge of the objects is allowed. Given a target grasp on a partial-view point cloud, we first discover its corresponding grasp among future frame's detected grasp poses as coarse estimation. These grasp poses can be given by an off-the-shelf grasp detector. Inspired by recent progress in local feature matching, which often uses image descriptors like SIFT [17] to describe interesting regions of images, we view grasp poses and their corresponding features as geometric descriptors on a partial-view point cloud. Based on such an assumption, we can simply estimate correspondences between two grasp sets from two different observation frames by matching the associated grasp features. Note that in opposition to classical local feature matching, features of the entire scene are also incorporated to help achieve global awareness. Furthermore, consecutive matching along an observation sequence may lead to the accumulation of error, on top of the coarse estimation through correspondence matching, we further use a memory-augmented coarse-to-fine module which uses both target grasp features and historical grasp features to refine the grasp tracking results for better temporal smoothness and semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.472, + 0.789 + ], + "angle": 0, + "content": "We conduct extensive experiments on two benchmarks to evaluate our method and demonstrate its effectiveness. The results show that our method outperforms two representative baseline methods. We also conduct real robot experiments on both single-object scenes and cluttered scenes. We report success rates of \\(81.25\\%\\) for single-object scenes and \\(81.67\\%\\) for multi-object scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.228, + 0.821 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.31, + 0.848 + ], + "angle": 0, + "content": "2.1. Grasping in Static Scenes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Thanks to the advances in the field of 3D perception [2, 25, 26], 6-DoF static grasping which takes point cloud as input is gaining increasing attention in both the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "research community and industry. In general, there are two lines of work that have been explored. The first line adopts a sampling-evaluation manner [15, 18, 23, 34]. Grasp candidates are sampled or generated on the point clouds and then their qualities are evaluated by neural networks. The second line processes the point cloud of the entire scene and predicts grasp poses across the scene [9, 10, 24, 27, 36] in an end-to-end manner. Compared to the first line, the end-to-end strategy achieves better balance in terms of the speed-accuracy tradeoff. Some researchers also propose to predict actions in continuous space [14, 32, 38]. Nevertheless, these methods can be adversely impacted by distribution shifts. Our method uses a pre-trained grasp pose detector from [9] which belongs to the second research line." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.324, + 0.68, + 0.34 + ], + "angle": 0, + "content": "2.2. Reactive Grasping" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.517 + ], + "angle": 0, + "content": "Though reactive grasping remains much less explored, a small amount of literature has partially investigated this more difficult problem. Most of these papers rely on object motion tracking. [1] predicts object motion and adaptively plans the grasp for a reachable grasp candidate. However, the object's motion is limited to few prototypes and thus it cannot handle the case where the objects move unpredictably. [19] tracks the objects' poses, yet the grasp planning phase is limited by a set of fixed grasp trajectories. [12,20,28] requires prior knowledge of the objects which is not always available in the real-world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.896, + 0.808 + ], + "angle": 0, + "content": "Some attempts on reactive grasping without prior information have been conducted recently. [22] picks the nearest grasp poses across frames. [42] samples grasp candidates by adding disturbances to grasp poses from the previous frame and evaluates their qualities in the current frame's scene. These methods focused on the smoothness of the grasp pose sequence but did not impose semantic constraints. Their predicted grasp poses can switch between different objects. Thus they mainly demonstrate single object grasping, and [42] requires an extra human hand segmentation module to avoid grasping the human's hand during handover while our correspondence matching module mitigates such a problem since grasps on hands and grasps on objects have low correspondence scores. [8] proposed a temporal association module to alleviate this dilemma. However, it only considers correspondence across two frames which is prone to error accumulation. In this paper, we further adopt a refinement module to adjust the grasp pose according to the target and historical grasp poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Some other works adopt continuous action prediction [14, 32]. Even if it is not designed for dynamic environments, it can be executed on moving objects thanks to the closed-loop controller. These methods may suffer greatly from distribution shifts. Different from these methods, our method tracks grasp poses explicitly." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8825" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.862, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.375, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed method for target-referenced reactive grasping. Given a target grasp and colored point clouds of a new coming frame with \\( N \\) points, our pipeline first detects \\( M \\) grasp candidates of size \\( M \\times 12 \\) (3 for translation and 9 for rotation) across the scene. It then extracts grasp features of size \\( M \\times 4C \\) for each detected grasp by cropping and embedding the points along the grasping direction. Furthermore, grasp features aggregation is used to estimate corresponding grasps between the previous frame and the current frame. After filtering, the correspondence score matrix produces \\( K \\) valid pairs of grasp. The predicted corresponding grasps are then fed into a refinement network along with the target grasp and other frames selected from a memory pool to be updated toward the target. Contents in the memory pool are updated by the refined values as well." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.489, + 0.453, + 0.506 + ], + "angle": 0, + "content": "2.3. Local Feature Matching in Computer Vision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.473, + 0.801 + ], + "angle": 0, + "content": "Unlike global features which summarize the entire scene, local features distribute densely across the scene and describe interesting regions. Under slightly different camera views, local feature matching is aimed at recognizing distinctive regions and establishing associations in images or point clouds. The mainstream of local feature matching includes detector-based matching and detector-free matching. For detector-based matching, classical hand-crafted features such as SIFT [17] and ORB [29] are widely adopted. With the development of deep learning, learned local features [5, 6] achieve satisfying results as well. Different from detector-based methods that extract sparse local features, detector-free methods establish pixel-wise or pointwise dense features [3, 13, 16, 31, 33]. With the extracted local features, nearest neighbor search or learning-based approach [30, 39] are often used in the matching phase. In our task, we consider each grasp brings rich geometric and visual patterns of the grasped local patch and our grasp correspondence are built upon these local features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.285, + 0.831 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we first briefly introduce some notations and metrics for grasp pose and grasp distance in 3D space, followed by formulating the problem of target-referenced reactive grasping." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.49, + 0.707, + 0.506 + ], + "angle": 0, + "content": "We define a grasp pose \\(\\mathcal{G}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.514, + 0.891, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = (\\mathbf {R} \\mathbf {t}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.893, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) denotes the rotation of the grasp pose and \\(\\mathbf{t} \\in \\mathbb{R}^{3 \\times 1}\\) denotes the translation. Consider two arbitrary grasps \\(\\mathcal{G}_1\\) and \\(\\mathcal{G}_2\\), a grasp distance measure is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.592, + 0.747, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\left(\\mathcal {G} _ {1}, \\mathcal {G} _ {2}\\right) = \\Delta \\mathbf {R} + \\Delta \\mathbf {t},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.612, + 0.892, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {R} = \\operatorname {a r c c o s} \\frac {\\operatorname {t r a c e} \\left(\\mathbf {R} _ {1} ^ {\\top} \\mathbf {R} _ {2}\\right) - 1}{2}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.645, + 0.698, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {t} = | | \\mathbf {t} _ {1} - \\mathbf {t} _ {2} | |.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Then we make a formal statement of target-referenced reactive grasping based on the background definition above. Given \\(\\mathcal{G}^1\\) at first frame as reference grasp, for any timestep \\(j > 1\\), target-referenced reactive grasping is aimed at finding the grasp \\(\\mathcal{G}^j\\) that minimizes the grasp distance:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.751, + 0.891, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {j *} = \\underset {\\mathcal {G} ^ {j}} {\\arg \\min } \\mathcal {D} \\left(\\mathcal {G} ^ {j}, \\mathbf {T} ^ {j 1} \\mathcal {G} ^ {1}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.891, + 0.834 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}^{j1}\\in \\mathbb{R}^{4\\times 4}\\) means the ground-truth transformation matrix of the reference grasp from the initial frame to the frame at timestep \\(j\\) in the camera's coordinate system." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.591, + 0.861 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this section, we detail our grasp tracking pipeline, as illustrated in Fig.2. For a target grasp, we first detect grasp" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8826" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.091, + 0.456, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.357, + 0.438, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Visualization of grasp features representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.462 + ], + "angle": 0, + "content": "candidates across the scene. Then we discover strong correspondences between the target grasp and grasp candidates by attentional grasp features aggregation. We finally use a memory-augmented refinement network to correct the predictions towards the target grasp." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.233, + 0.488 + ], + "angle": 0, + "content": "4.1. Grasp Detector" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.47, + 0.723 + ], + "angle": 0, + "content": "For each incoming new frame, we generate a grasp set along with it. This work uses an off-the-shelf GraspNet baseline [9] for proposals of grasp candidates. GraspNet is a learned grasp pose detector that takes the partial point cloud of the scene \\(\\mathcal{P}\\) as input and outputs a grasp set \\(\\mathbf{G}\\) across the entire scene. In GraspNet, a point cloud \\(\\mathcal{P}\\) of size \\(N\\times 3\\) is downsampled and transformed to feature vectors of size \\(M\\times C\\) by the PointNet++ backbone network [26] where \\(C\\) denotes the channel dimension of the features and \\(M\\) is the number of points after farthest point sampling [7]. For each sampled point, GraspNet predicts a grasp \\(\\mathcal{G}\\) and its confidence score. The detected grasp set is then sent to the following steps for grasp tracking. While GraspNet performs well in this work, the following steps are not limited to any specific grasp detector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.734, + 0.363, + 0.75 + ], + "angle": 0, + "content": "4.2. Grasp Correspondence Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.469, + 0.849 + ], + "angle": 0, + "content": "We follow the formulation of grasp correspondence in [8]. A many-to-many correspondence matrix is learned during training, given two grasp sets in consecutive frames. During testing, we can choose the grasp pose with the highest correspondence score with the target pose in the previous frame as the tracked pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Grasp Features Representation First we illustrate how we represent grasp features. For \\(M\\) grasp poses in a scene," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.334 + ], + "angle": 0, + "content": "we follow [8] to extract seed features, geometric features and appearance features for them, each having a shape of \\( M \\times C \\). In addition, since it is quite common in the real world that multiple similar objects exist simultaneously in the scene, grasps on similar objects cannot be distinguished in such situations which results in correspondence switching between objects. Hence we further add global scene features of size \\( M \\times C \\) to provide information about the scene state. This global feature is extracted by max-pooling pointwise features from the backbone output, thus is the same for each of \\( M \\) grasp poses. Finally, we concatenate geometric features, visual features, seed point features and global scene features to construct grasp features of dimension \\( 4C \\) for each grasp, denoted by \\( \\mathbf{x} \\). We show the above process in Fig.3. We present more details of grasp features representation in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Positional Information To provide positional information for grasps, we encode grasp poses by MLP layers of size \\((12, 2C, 4C)\\). These feature vectors are then added to the raw grasp features \\(\\mathbf{x}\\), such that grasps can be treated in a similar way to words in a sentence." + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.446, + 0.891, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} \\leftarrow \\mathbf {x} + \\mathbf {M L P} ^ {(1)} (p), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.55 + ], + "angle": 0, + "content": "where \\( p \\) is the grasp pose of dimension 12, consisting of 9 parameters for the rotation matrix and 3 for translation. We use this feature as the final grasp features representation. Such a process makes our method position-dependent and benefits following the grasp features aggregation step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.662 + ], + "angle": 0, + "content": "Attentional Grasp Features Aggregation Thanks to the above-mentioned grasp features, it is now straightforward to reason about the visual, geometric and positional properties of grasps jointly. For a grasp \\(\\mathcal{G}_i\\), we first embed its associated grasp features using MLP layers and transform its feature size from \\(4C\\) to \\(2C\\) by MLP layers." + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.677, + 0.891, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {g _ {i}} = \\mathbf {M L P} ^ {(2)} \\left(\\mathbf {x} _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Apart from the features of a grasp itself, given a pair of grasp sets, it is intuitive that interacting with contextual grasps within grasp sets or across grasp sets is critical to reduce ambiguities and increase the distinctiveness of grasps as well. [8] adopts vanilla MLP blocks which cannot model long-range relationships among grasp poses. In this paper, we take advantage of the global receptive field of the attentional graph neural network (GNN) to model such dependencies. Following [30], self-attention and cross-attention [35] are adopted for context aggregation within grasp sets and across grasp sets. In particular, features are aggregated unidirectionally in self-attention and bidirectionally in cross-attention. Given two grasp sets \\(\\mathbf{G}_1\\) and \\(\\mathbf{G}_2\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "that each consists of \\(M\\) grasps, grasp features aggregation generates matching descriptors \\(\\mathbf{f}^{m_1}\\), \\(\\mathbf{f}^{m_2} \\in \\mathbb{R}^{M \\times 2C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.473, + 0.548 + ], + "angle": 0, + "content": "Here we outline the workflow of grasp features aggregation. In self-attention, we only process grasps within \\(\\mathbf{G}_1\\) or \\(\\mathbf{G}_2\\). Taking \\(\\mathbf{G}_1\\) as an example, grasp features \\(\\mathbf{f}^{\\mathbf{G}_1}\\) of size \\((M \\times 2C)\\) are first mapped to three elements which are named query, key and value in convention by learnable weights \\(\\mathbf{W}_{\\mathbf{Q}}\\), \\(\\mathbf{W}_{\\mathbf{K}}\\) and \\(\\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{2C \\times 2C}\\). The attention computation can then be written as \\(\\mathbf{A} = (\\mathbf{Q} \\cdot \\mathbf{K}^T) / \\sqrt{C}\\) where \\(\\mathbf{A}\\) denotes the attention matrix and has a size of \\((M \\times M)\\) and \\((\\cdot)\\) is the dot product operator. \\(\\mathbf{A}\\) actually measures the similarity between query and key. After that, the message propagation step can be defined as \\(\\mathbf{M} = \\mathrm{softmax}(\\mathbf{A}) \\cdot \\mathbf{V}\\). \\(\\mathbf{M}\\) is later fed into MLP layers of size \\((2C, 4C, 2C)\\) to produce the output features which have the same size as \\(\\mathbf{f}^{\\mathbf{G}_1}\\). Since \\(\\mathbf{M}\\) can be regarded as the sum of value vectors weighted by the similarity measure, therefore the output feature vectors represent correspondences among grasps in \\(\\mathbf{G}_1\\). In cross-attention, we consider grasp features from both \\(\\mathbf{G}_1\\) and \\(\\mathbf{G}_2\\). To produce the output feature vector, the only thing different is that key and value are no longer generated from \\(\\mathbf{f}^{\\mathbf{G}_1}\\) but from \\(\\mathbf{f}^{\\mathbf{G}_2}\\). Intuitively, in self-attention information flows only within \\(\\mathbf{G}_1\\) or \\(\\mathbf{G}_2\\) whereas in cross-attention information flows from \\(\\mathbf{G}_2\\) to \\(\\mathbf{G}_1\\) or vice versa. We consider the output of cross-attention layers as a representation of correspondence between \\(\\mathbf{G}_1\\) and \\(\\mathbf{G}_2\\). We interleave self-attention and cross-attention layers by \\(N_c\\) times and get the matching descriptors \\(\\mathbf{f}^{m_1}, \\mathbf{f}^{m_2}\\) which are later used to compute correspondences within and across grasp sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Correspondence Estimation Given \\(\\mathbf{f}^{m_1},\\mathbf{f}^{m_2}\\) , we can simply compute the correspondence matrix \\(S\\in \\mathbb{R}^{M\\times M}\\) For each pair of grasps from these two sets, the correspondence score can be calculated by cosine similarity" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.641, + 0.47, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (i \\cdot j) = \\frac {\\mathbf {f} _ {i} ^ {m _ {1}} \\cdot \\mathbf {f} _ {j} ^ {m _ {2}}}{\\left| \\left| \\mathbf {f} _ {i} ^ {m _ {1}} \\right| \\right| _ {2} \\cdot \\left| \\left| \\mathbf {f} _ {j} ^ {m _ {2}} \\right| \\right| _ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Since cosine similarity is naturally normalized, it makes training more stable. Taking advantage of contrastive learning, during training a target grasp can be assigned the same class to any number of grasps in the other grasp set. This is important because some grasps are close in terms of translation and some are close in terms of rotation. During training, our method allows parallel prediction of multiple target grasps. We simply pick grasps that have the highest correspondence scores with respect to target grasps. Before entering the next stage, a grasp set is filtered to \\( K \\) grasps by removing upward grasps and grasps with low correspondence scores with respect to all other grasps. If the number of grasps is less than \\( K \\), we repeat the first grasp. In real robot experiments, only one target grasp is used." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.857, + 0.108 + ], + "angle": 0, + "content": "4.3. Memory-augmented Refinement Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Since the correspondence is estimated between fixed grasp sets, the tracking performance greatly relies on the grasp detector which is not guaranteed to generate identical grasps across different frames. Also, consecutive correspondence matching may lead to an accumulation of errors. Therefore, to further improve the semantic consistency with the reference grasp pose, a refinement network is used to correct predictions made by correspondences towards the target grasp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.893, + 0.401 + ], + "angle": 0, + "content": "To be specific, we store the tracked grasp poses on past frames and their associated features in a memory pool. For each forward pass, the refinement network uses historical observations selected from the memory pool to refine the parameters of the selected grasp pose in the current frame. Here the historical frames from the memory pool are selected according to their time indexes. In practice, we select the reference pose in the first frame and apply uniform sampling for simplicity in the interval \\((1,t - 1]\\) to select past frames, where \\(t\\) denotes the current time index." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.893, + 0.658 + ], + "angle": 0, + "content": "For a \\(L\\) frame grasping sequence that consists of 1 frame of tracking target, 1 current frame, and \\(L - 2\\) frames selected from the memory pool, grasp features of size \\(L \\times (4C)\\) are first transformed to \\(L \\times 2C\\) by MLP layers. After that, we simply repeat the target frame and concatenate it with other frames, resulting in features of size \\((L - 1) \\times (2C)\\). Such that all the selected frames are conditioned on the tracking target. We then apply temporal encoding which is similar to positional encoding introduced in [35]. These features are then fed into a refinement network which outputs \\((L - 1)\\) refined poses. Note that both historical grasp poses and the grasp pose in the current frame are updated. This refinement network has the same architecture as the grasp correspondence network, except that only self-attention layers are adopted. Moreover, the rotation is parameterized by the 6D representation introduced in [43]. Causal mask is used to avoid causality confusion as well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.628, + 0.685 + ], + "angle": 0, + "content": "4.4. Supervision" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.693, + 0.892, + 0.723 + ], + "angle": 0, + "content": "The objective function used to train the above pipeline can be divided into two parts:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.736, + 0.892, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c} + \\beta \\mathcal {L} _ {R, t}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c\\) denotes the loss for correspondence learning and \\(\\mathcal{L}_{R,t}\\) denotes the loss for grasp pose refinement. \\(\\beta\\) is the weighting term which we set as 1 during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For \\(\\mathcal{L}_c\\), we follow [8] and adopt supervised contrastive loss. For any two grasp poses with positive correspondence label, we denote them as \\(\\mathrm{corr}(\\mathcal{G}_i,\\mathcal{G}_k) = 1\\) or vice versa. Given two arbitrary grasp sets \\(\\mathbf{G}_1\\) and \\(\\mathbf{G}_2\\), we first locate the positive grasp set \\(\\mathbf{P}(i) = \\{\\mathcal{G}_k^2\\in \\mathbf{G}_2|\\mathrm{corr}(\\mathcal{G}_i,\\mathcal{G}_k) = 1\\}\\) for any \\(\\mathcal{G}_i\\in \\mathbf{G}_1\\). This correspondence loss can then be" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.151, + 0.105 + ], + "angle": 0, + "content": "written as:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.114, + 0.468, + 0.167 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\sum_ {\\mathcal {G} _ {i} ^ {1} \\in \\mathbf {G} _ {1}} \\frac {- 1}{| \\mathbf {P} (i) |} \\sum_ {\\mathcal {G} _ {k} ^ {2} \\in P (i)} \\log \\frac {\\exp (\\mathcal {S} (i \\cdot k) / \\tau)}{\\sum \\mathcal {G} _ {j} ^ {2} \\in \\mathbf {G} _ {2} \\exp (\\mathcal {S} (i \\cdot j) / \\tau)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.168, + 0.469, + 0.228 + ], + "angle": 0, + "content": "where \\( |\\mathbf{P}(i)| \\) denotes the cardinality and \\( \\tau \\) is a scalar temperature parameter. This loss pulls together the positive grasp pairs in embedding space, whereas pushes apart negative grasp pairs meanwhile." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.229, + 0.469, + 0.32 + ], + "angle": 0, + "content": "For grasp pose refinement, we adopt a transformation loss which consists two terms for optimizing rotation and translation respectively. Consider \\(\\mathcal{J} = \\{j_1,j_2,\\dots ,j_n\\}\\) that stores arbitrary number of time index, for any time step \\(j\\in \\mathcal{J}\\), given target grasp \\(\\mathcal{G}^1\\), we first compute its transformed grasp in \\(j^{th}\\) frame's coordinates, denoted by" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.329, + 0.469, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {j 1} = \\left(\\mathbf {R} ^ {j 1} \\mathbf {t} ^ {j 1}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.358, + 0.356, + 0.373 + ], + "angle": 0, + "content": "We then define the transformation loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.38, + 0.469, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {R, t} = \\frac {1}{| \\mathcal {I} |} \\sum_ {j \\in \\mathcal {J}} \\frac {1}{| \\mathbf {G} _ {j} |} \\sum_ {\\mathcal {G} _ {i} ^ {j} \\in \\mathbf {G} _ {j}} | | (\\mathbf {R} ^ {*}) ^ {- 1} \\mathbf {R} ^ {j 1} - \\mathbf {I} | | _ {2} (10) \\\\ + \\frac {1}{| \\mathcal {I} |} \\sum_ {j \\in \\mathcal {J}} \\frac {1}{| \\mathbf {G} _ {j} |} \\sum_ {\\mathcal {G} _ {i} ^ {j} \\in \\mathbf {G} _ {j}} \\| \\mathbf {t} ^ {j 1} - \\mathbf {t} ^ {*} \\| _ {2}, (10) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.475, + 0.357, + 0.491 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_j\\) denotes grasp set at time step \\(j\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.504, + 0.209, + 0.52 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.528, + 0.173, + 0.543 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.551, + 0.469, + 0.747 + ], + "angle": 0, + "content": "The proposed pipeline is trained on GraspNet-1Billion dataset [9] which is a large-scale real-world dataset for grasp detection. It includes 89 objects, 190 scenes and 256 camera views for each scene. We follow [8] to generate grasp correspondence labels across viewpoints on the GraspNet-1Billion dataset and refer readers to [8] for more details. In addition, we further collect 30 scenes of moving objects for testing. For each frame, we record the RGBD images and objects' 6D poses. We manually annotate 10 grasps for the first frame of each scene as targets and further project the annotated grasps according to the 6D poses for the future frames. This test set is named Moving GraspNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.373, + 0.773 + ], + "angle": 0, + "content": "5.2. Pre-processing and Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Before being fed into the network, the point cloud is first downsampled to 20000 points and the RGB values are normalized. We also remove points outside a pre-defined workspace. Furthermore, we adopt massive data augmentation to avoid over-fitting and enrich objects' motion. For each object in the scene, its point cloud is augmented on-the-fly by a random translation of \\([-0.2, 0.2]\\) meter in all XYZ directions and a random rotation of \\([-30, 30]\\) degrees" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "around z-axis. Moreover the scene's point cloud is rotated by \\([-30, 30]\\) degrees around both x-axis and z-axis. Due to the space limitations, please refer to the supplementary materials for details about network implementation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.165, + 0.782, + 0.181 + ], + "angle": 0, + "content": "5.3. Performance of Grasp Tracking" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Evaluation Metric We propose a new Multiple Grasp Pose Tracking (MGTA) metric which follows the idea of Multiple Object Tracking Accuracy (MOTA) [21] in multi-object tracking task, except that the similarity measure between two grasps is computed using Eq.(2). Details are given in supplementary materials. Apart from MGTA, we also report the mean translation error and rotation error to the tracking target of each sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.892, + 0.462 + ], + "angle": 0, + "content": "We follow the same evaluation procedure on GraspNet1-Billion and Moving GraspNet. For each scene, we pick 10 reference grasps in the first frame which are later used for grasp tracking. To be specific, these reference grasps are detected by a grasp detector on GraspNet1-billion but manually labeled on Moving GraspNet. For each remaining frame, we compute the ground truth grasps using objects' 6D poses and segmentation mask of the scene. The metrics are calculated using the transformed reference grasps and predictions associated with each frame." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Comparing with Representative Methods We compare our methods with two other baseline methods. A heuristic method is implemented as the first baseline. Given a grasp pose from the previous frame, its nearest neighbor detected in the current camera coordinates is picked. For the second method, we adopt Bundle Track [40] which is one representative method in the field of unseen object 6D pose tracking. It fits our task setting well since it requires no objects' CAD model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.623, + 0.892, + 0.758 + ], + "angle": 0, + "content": "As it is shown in Tab.1 and Tab.2, our method outperforms baseline methods on both GraspNet-1Billion and Moving GraspNet. It is found in experiments that, in many cases, grasps cannot fall on the same objects if we only pick the nearest neighbors. Also, it can be drawn from the results that tracking results are improved significantly after refinement, especially for rotation. Note that MGTA will be negative if there is a large number of false positives, false negatives, and ID switches." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.724, + 0.788 + ], + "angle": 0, + "content": "5.4. Real Robot Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In real robot experiments, we use Flexiv Rizon robot arm with an Intel Realsense L515 depth camera mounted on the end-effector. A Robotiq-85 parallel-jaw gripper is used with 3D-printed extended fingertips attached to it. The real robot experiments run on an NVIDIA 3090 GPU at around 10 fps. Please refer to the supplementary materials for details of the objects." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8829" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.905, + 0.215 + ], + "angle": 0, + "content": "
MethodSeenSimilarNovel
MGTA↑n (cm)↓n (°)↓MGTA↑n (cm)↓n (°)MGTA↑n (cm)↓n (°)↓
Nearest-0.8115.295.98-0.8015.497.61-0.8316.696.90
Bundle Track [40]0.741.5215.800.721.4716.370.721.9315.41
Ours, w/o refinement-0.322.3153.24-0.302.6857.59-0.392.9358.49
Ours, M=5120.801.5810.890.801.7410.620.752.1015.03
Ours, M=10240.841.3410.150.851.529.840.781.9012.96
Ours, M=20480.851.279.820.861.359.500.791.8411.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.225, + 0.893, + 0.252 + ], + "angle": 0, + "content": "Table 1. GraspNet-1Billion evaluation results on seen, similar and novel objects respectively. \\( M \\) denotes the number of grasp detected,its default value is 1024." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.256, + 0.862, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.537, + 0.893, + 0.566 + ], + "angle": 0, + "content": "Figure 4. Illustration of tracking a toy dog and toy rabbit. We can see that the grasp poses are consistent, i.e., on the head of the dog and the ear of the rabbit. Yellow bounding box represents to the view limit of the camera. No box means the camera can see the entire scene." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.575, + 0.477, + 0.687 + ], + "angle": 0, + "content": "
MethodMGTA↑n (cm)↓n (°)↓
Nearest-0.478.0967
Bundle Track0.377.6023.52
Ours, w/o refinement-0.436.2462.33
Ours, M=5120.475.5322.63
Ours, M=10240.524.6020.37
Ours, M=20480.544.1418.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.738 + ], + "angle": 0, + "content": "Table 2. Moving GraspNet evaluation results. All objects in Moving GraspNet are novel. \\( M \\) denotes the number of grasp detected, its default value is 1024." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Protocol For each frame, we receive RGBD images from the in-hand camera. We transform the point cloud to the camera coordinate system at ready state by the camera extrinsic for simplicity as well. Due to safety considerations, points outside a pre-defined workspace will be masked. After a grasp candidate generation phase, we select the grasp with the highest predicted score as a reference grasp and keep tracking it for each new frame. During the tracking phase, the robot moves to a pre-grasp pose which is 0.02" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.575, + 0.864, + 0.735 + ], + "angle": 0, + "content": "
Object ID#Attempt#SuccessSuccess Rate
11080.8
21090.9
31080.8
41070.7
61080.8
71080.8
91080.8
101090.9
Total80650.8125
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.745, + 0.885, + 0.76 + ], + "angle": 0, + "content": "Table 3. Real robot experiments results on single object scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.903 + ], + "angle": 0, + "content": "meters higher than the actual grasp pose. The robot executes the grasp once three conditions are met simultaneously: 1) the translation distance between the gripper center and the target grasp in XY-plane is lower than 0.01 meter (the XYplane is parallel to the table); 2) the translation distance between the gripper center and the target grasp in 3D space is lower than 0.04 meters; 3) the rotation distance between the gripper and the target grasp in 3D space is lower than" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8830" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.462, + 0.217 + ], + "angle": 0, + "content": "
Object ID#Attempt#SuccessSuccess Rate
7,7,8,81080.8
10,13,141090.9
1,11,15,16,17,181080.8
2,3,51090.9
9,121070.7
20,201080.8
Total60490.8167
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.227, + 0.462, + 0.243 + ], + "angle": 0, + "content": "Table 4. Real robot experiments results on multi-object scenes." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.259, + 0.462, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Figure 5. Correspondence switching between similar objects. Red arrow indicates correct correspondence and blue arrow indicates wrong correspondence." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.578, + 0.469, + 0.609 + ], + "angle": 0, + "content": "15 degrees. The initial positions of the objects are chosen randomly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Results As shown in Tab.3 and Tab.4, we execute 20 grasp attempts for each scene. In single-object scenes, we randomly sample 8 objects for testing and we achieve a success rate of 81 percent. In multi-object scenes, we only choose objects with similar visual and geometric patterns to test the temporal smoothness and semantic consistency of the predicted grasps. Results suggest that single-object scenes and multi-object scenes have nearly the same success rate which further proves the effectiveness of our method. We illustrate one case from real robot experiments in Fig.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.182, + 0.803 + ], + "angle": 0, + "content": "5.5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of Global Features In Sec.4.2, we incorporate global features to represent grasp features. While local grasp features are sufficient to identify corresponding local structures, grasp trials may fail due to correspondence switches among objects with similar visual and geometric patterns as we can see in Fig.5. Such phenomenon may be" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.888, + 0.169 + ], + "angle": 0, + "content": "
Object IDMethod#Attempt#SW
7,7,8,8w/o global features105
Full model101
1,17,17w/o global features104
Full model100
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.179, + 0.892, + 0.208 + ], + "angle": 0, + "content": "Table 5. Statistics of Correspondence Switching. SW represents correspondence switching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.892, + 0.357 + ], + "angle": 0, + "content": "due to lacking awareness of the scene state and we propose to solve it by global features as what we state in Sec.4.2. In order to analyze the actual effectiveness of global features, we further record how frequent correspondence switching happens with or without global features. We conduct these experiments on identical objects which are placed close to each other in the scenes. Tab.5 suggests that the number of correspondence switches decreased significantly with global features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.893, + 0.571 + ], + "angle": 0, + "content": "Failure Analysis There are three major types of failure cases found in real robot experiments. First, some predicted grasp candidates from the detector do not have good qualities. Naturally, tracking such targets may lead to failed grasps. Second, it is hard to heuristically plan the timing to close the gripper. Due to some grasps may fall on the edges of objects, it is common that the objects may slip away immediately when the robot receives the command to close the gripper. Furthermore, since we mount the camera on the wrist, the point cloud may become fragmented due to view limitation and occlusion, especially when the objects fall behind the gripper. In such cases, the predictions can be inaccurate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.619, + 0.603 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.718 + ], + "angle": 0, + "content": "We present a novel solution to reactive grasping in a target-referenced setting. For a reference grasp, our method predicts temporally smooth and semantically consistent grasp poses in the future frames without any prior knowledge of instance. Experiments conducted on both single-object and multi-object scenes show that our method provides reliable grasp plans under various environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the National Key R&D Program of China (No.2021ZD0110700), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Qi Zhi Institute, and Shanghai Science and Technology Commission (21511101200)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Contributions: J. Liu and H-S. Fang initiated the project. J. Liu designed and trained the network and conducted the evaluation together with R. Zhang. J. Liu wrote the paper. H-S. Fang devised and mentored this project and edited the paper. M. Gou, H. Fang, and S. Xu constructed the testing dataset. C. Wang and H. Yan helped with the evaluation. C. Lu supervised the project and provided hardware support." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8831" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Iretiayo Akinola, Jingxi Xu, Shuran Song, and Peter K. Allen. Dynamic grasping with reachability and motion awareness. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9422-9429, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[4] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[5] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Toward geometric deep slam. ArXiv, abs/1707.07410, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 337-33712, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[7] Y. Eldar, M. Lindenbaum, M. Porat, and Y.Y. Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[8] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. arXiv preprint arXiv:2212.08333, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[9] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[10] Minghao Gou, Hao-Shu Fang, Zhanda Zhu, Sheng Xu, Chenxi Wang, and Cewu Lu. Rgb matters: Learning 7-dof grasp poses on monocular rgbd images. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13459-13466. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[11] Yinlin Hu, Joachim Hugonot, Pascal Fua, and Mathieu Salzmann. Segmentation-driven 6d object pose estimation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[12] Seungsu Kim, Ashwini Shukla, and Aude Billard. Catching objects in flight. IEEE Transactions on Robotics, 30(5):1049-1065, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Donghoon Lee, Onur C. Hamsici, Steven Feng, Prachee Sharma, and Thorsten Gernoth. Deeppro: Deep partial point cloud registration of objects. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5663-5672, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International journal of robotics research, 37(4-5):421-436, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[15] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Görner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In 2019 International Conference on Robotics and Automation (ICRA), pages 3629-3635. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[16] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[17] David Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91–, 11 2004. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[19] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Aleš Leonardis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43(5):1241-1256, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[20] Arjun Menon, Benjamin Cohen, and Maxim Likhachev. Motion planning for smooth pickup of moving objects. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 453-460. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[21] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. ArXiv, abs/1603.00831, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[23] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2901-2910, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[24] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3619-3625, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4g: Amodal single-view single-shot se (3) grasp" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8832" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "detection in cluttered scenes. In Conference on robot learning, pages 53-65. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[28] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[29] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[31] Tanner Schmidt, Richard Newcombe, and Dieter Fox. Self-supervised visual descriptor learning for dense correspondence. IEEE Robotics and Automation Letters, 2(2):420-427, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[32] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[33] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[34] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt. Grasp pose detection in point clouds. The International Journal of Robotics Research (IJRR), 36(13-14):1455-1473, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[36] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15964-15973, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[37] Chen Wang, Roberto Martín-Martín, Danfei Xu, Jun Lv, Cewu Lu, Li Fei-Fei, Silvio Savarese, and Yuke Zhu. 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10059–10066, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[38] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[39] Yue Wang and Justin M. Solomon. Deep closest point: Learning representations for point cloud registration. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "models. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074. IEEE, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[41] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E. Bekris. se(3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[42] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[43] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8833" + } + ] +] \ No newline at end of file diff --git a/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_origin.pdf b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea55a5b3a52421fdb89e1299e16a33504f071884 --- /dev/null +++ b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/d5655194-9ce2-4b7e-be6f-e8df1e2c1030_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69e452dcf31bece490d60282544fd8bd40b900fec7a831c36ce0cdec73dc9875 +size 2285266 diff --git a/2023/Target-Referenced Reactive Grasping for Dynamic Objects/full.md b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c3351c4821647541dc0e3a32b1ffe92d0fbe5617 --- /dev/null +++ b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/full.md @@ -0,0 +1,299 @@ +# Target-referenced Reactive Grasping for Dynamic Objects + +Jirong Liu $^{1,2}$ , Ruo Zhang $^{2}$ , Hao-Shu Fang $^{2}$ , Minghao Gou $^{2}$ , Hongjie Fang $^{2}$ , Chenxi Wang $^{2,3}$ , Sheng Xu $^{2}$ , Hengxu Yan $^{2}$ , Cewu Lu $^{1,2\dagger}$ $^{1}$ Shanghai Qi Zhi institute, $^{2}$ Shanghai Jiao Tong University, $^{3}$ Flexiv Robotics, LTD jirong@sjtu.edu.cn, {ruozhang0608, fhaoshu}@gmail.com, {gmh2015, galaxies, wcx1997, xs1020, hengxuyan, lucewu}@sjtu.edu.cn + +# Abstract + +Reactive grasping, which enables the robot to successfully grasp dynamic moving objects, is of great interest in robotics. Current methods mainly focus on the temporal smoothness of the predicted grasp poses but few consider their semantic consistency. Consequently, the predicted grasps are not guaranteed to fall on the same part of the same object, especially in cluttered scenes. In this paper, we propose to solve reactive grasping in a target-referenced setting by tracking through generated grasp spaces. Given a targeted grasp pose on an object and detected grasp poses in a new observation, our method is composed of two stages: 1) discovering grasp pose correspondences through an attentional graph neural network and selecting the one with the highest similarity with respect to the target pose; 2) refining the selected grasp poses based on target and historical information. We evaluate our method on a large-scale benchmark GraspNet-1Billion. We also collect 30 scenes of dynamic objects for testing. The results suggest that our method outperforms other representative methods. Furthermore, our real robot experiments achieve an average success rate of over 80 percent. Code and demos are available at: https://graspnet.net/reactive. + +# 1. Introduction + +Reactive grasping is in great demand in the industry. For instance, in places where human-robot collaboration is heavily required like factories, stress on laborers will be significantly relieved if robots can receive tools from humans and complete the harder work for laborers. Such a vision is based on reactive grasping. + +On the contrary to static environments, in reactive grasping, dynamic task setting poses new challenges for algorithm design. Previous research in this area mainly focuses + +![](images/605780c57469374cb35909e698f20e87ce52c6f039065b5a5da40e7ff9b62a59.jpg) + +![](images/5b978c230c54871cba1ce8cf59944b78d8f13261f9f3a9fbd2d5654f2f458dff.jpg) +Figure 1. (a) Classic reactive grasping guarantees the smoothness of the grasp poses but cannot predicts grasps on the same part of the hammer. (b) Our target-referenced reactive grasping takes semantic consistency into consideration. The generated grasp poses across frames are illustrated with blue grippers. + +on planning temporally smooth grasps [22, 42] to avoid wavy and jerky robot motion. Few of them pay attention to its semantic consistency. In short, given a targeted grasp at the first frame, we want the robot to grasp the same part of the object in the following frames. Additionally, it is not guaranteed that grasp predictions made by classical methods fall on the same object in cluttered scenes. Hence, most of their experiments are conducted on single-object scenes. Unlike previous works, this work is aimed at achieving temporally smooth and semantically consistent reactive grasping in clutter given a targeted grasp. We refer to such a task setting as target-referenced reactive grasping as shown in Fig.1. Note that despite robot handover is a major application scenario of reactive grasping, this work focuses on a more general task setting - dynamic object grasping. + +A naive idea to solve this task is to generate reference grasp poses for the initial scene and consecutively track the object's 6D pose. As the object moves, the initial grasp pose can be projected to a new coming frame based on the object's 6D pose. Although such an idea seems to be natural and valid, some bottleneck greatly degrades its viability. First of all, the solution to reactive grasping should be able to handle objects' motion in real-time, meaning that it requires fast inference speed and immediate response to continuous environmental changes. However, 6D pose tracking may be time-consuming due to commonly-used instance segmentation [11,40]. Second, 6D pose tracking usually requires objects' prior knowledge, such as CAD models [4,41], which is not always available in the real world as well or achieves only category-level generalization [37]. + +Different from tracking objects, we propose to track grasps by a two-stage policy instead. We also comply with the restriction that no prior knowledge of the objects is allowed. Given a target grasp on a partial-view point cloud, we first discover its corresponding grasp among future frame's detected grasp poses as coarse estimation. These grasp poses can be given by an off-the-shelf grasp detector. Inspired by recent progress in local feature matching, which often uses image descriptors like SIFT [17] to describe interesting regions of images, we view grasp poses and their corresponding features as geometric descriptors on a partial-view point cloud. Based on such an assumption, we can simply estimate correspondences between two grasp sets from two different observation frames by matching the associated grasp features. Note that in opposition to classical local feature matching, features of the entire scene are also incorporated to help achieve global awareness. Furthermore, consecutive matching along an observation sequence may lead to the accumulation of error, on top of the coarse estimation through correspondence matching, we further use a memory-augmented coarse-to-fine module which uses both target grasp features and historical grasp features to refine the grasp tracking results for better temporal smoothness and semantic consistency. + +We conduct extensive experiments on two benchmarks to evaluate our method and demonstrate its effectiveness. The results show that our method outperforms two representative baseline methods. We also conduct real robot experiments on both single-object scenes and cluttered scenes. We report success rates of $81.25\%$ for single-object scenes and $81.67\%$ for multi-object scenes. + +# 2. Related Works + +# 2.1. Grasping in Static Scenes + +Thanks to the advances in the field of 3D perception [2, 25, 26], 6-DoF static grasping which takes point cloud as input is gaining increasing attention in both the + +research community and industry. In general, there are two lines of work that have been explored. The first line adopts a sampling-evaluation manner [15, 18, 23, 34]. Grasp candidates are sampled or generated on the point clouds and then their qualities are evaluated by neural networks. The second line processes the point cloud of the entire scene and predicts grasp poses across the scene [9, 10, 24, 27, 36] in an end-to-end manner. Compared to the first line, the end-to-end strategy achieves better balance in terms of the speed-accuracy tradeoff. Some researchers also propose to predict actions in continuous space [14, 32, 38]. Nevertheless, these methods can be adversely impacted by distribution shifts. Our method uses a pre-trained grasp pose detector from [9] which belongs to the second research line. + +# 2.2. Reactive Grasping + +Though reactive grasping remains much less explored, a small amount of literature has partially investigated this more difficult problem. Most of these papers rely on object motion tracking. [1] predicts object motion and adaptively plans the grasp for a reachable grasp candidate. However, the object's motion is limited to few prototypes and thus it cannot handle the case where the objects move unpredictably. [19] tracks the objects' poses, yet the grasp planning phase is limited by a set of fixed grasp trajectories. [12,20,28] requires prior knowledge of the objects which is not always available in the real-world. + +Some attempts on reactive grasping without prior information have been conducted recently. [22] picks the nearest grasp poses across frames. [42] samples grasp candidates by adding disturbances to grasp poses from the previous frame and evaluates their qualities in the current frame's scene. These methods focused on the smoothness of the grasp pose sequence but did not impose semantic constraints. Their predicted grasp poses can switch between different objects. Thus they mainly demonstrate single object grasping, and [42] requires an extra human hand segmentation module to avoid grasping the human's hand during handover while our correspondence matching module mitigates such a problem since grasps on hands and grasps on objects have low correspondence scores. [8] proposed a temporal association module to alleviate this dilemma. However, it only considers correspondence across two frames which is prone to error accumulation. In this paper, we further adopt a refinement module to adjust the grasp pose according to the target and historical grasp poses. + +Some other works adopt continuous action prediction [14, 32]. Even if it is not designed for dynamic environments, it can be executed on moving objects thanks to the closed-loop controller. These methods may suffer greatly from distribution shifts. Different from these methods, our method tracks grasp poses explicitly. + +![](images/e6b513f18b4c1ad0ff11ab235f415d64b0b077abf86eaf78b5837e6418bc3198.jpg) +Figure 2. An overview of the proposed method for target-referenced reactive grasping. Given a target grasp and colored point clouds of a new coming frame with $N$ points, our pipeline first detects $M$ grasp candidates of size $M \times 12$ (3 for translation and 9 for rotation) across the scene. It then extracts grasp features of size $M \times 4C$ for each detected grasp by cropping and embedding the points along the grasping direction. Furthermore, grasp features aggregation is used to estimate corresponding grasps between the previous frame and the current frame. After filtering, the correspondence score matrix produces $K$ valid pairs of grasp. The predicted corresponding grasps are then fed into a refinement network along with the target grasp and other frames selected from a memory pool to be updated toward the target. Contents in the memory pool are updated by the refined values as well. + +# 2.3. Local Feature Matching in Computer Vision + +Unlike global features which summarize the entire scene, local features distribute densely across the scene and describe interesting regions. Under slightly different camera views, local feature matching is aimed at recognizing distinctive regions and establishing associations in images or point clouds. The mainstream of local feature matching includes detector-based matching and detector-free matching. For detector-based matching, classical hand-crafted features such as SIFT [17] and ORB [29] are widely adopted. With the development of deep learning, learned local features [5, 6] achieve satisfying results as well. Different from detector-based methods that extract sparse local features, detector-free methods establish pixel-wise or pointwise dense features [3, 13, 16, 31, 33]. With the extracted local features, nearest neighbor search or learning-based approach [30, 39] are often used in the matching phase. In our task, we consider each grasp brings rich geometric and visual patterns of the grasped local patch and our grasp correspondence are built upon these local features. + +# 3. Problem Formulation + +In this section, we first briefly introduce some notations and metrics for grasp pose and grasp distance in 3D space, followed by formulating the problem of target-referenced reactive grasping. + +We define a grasp pose $\mathcal{G}$ as + +$$ +\mathcal {G} = (\mathbf {R} \mathbf {t}), \tag {1} +$$ + +where $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ denotes the rotation of the grasp pose and $\mathbf{t} \in \mathbb{R}^{3 \times 1}$ denotes the translation. Consider two arbitrary grasps $\mathcal{G}_1$ and $\mathcal{G}_2$ , a grasp distance measure is defined as + +$$ +\mathcal {D} \left(\mathcal {G} _ {1}, \mathcal {G} _ {2}\right) = \Delta \mathbf {R} + \Delta \mathbf {t}, +$$ + +$$ +\Delta \mathbf {R} = \operatorname {a r c c o s} \frac {\operatorname {t r a c e} \left(\mathbf {R} _ {1} ^ {\top} \mathbf {R} _ {2}\right) - 1}{2}, \tag {2} +$$ + +$$ +\Delta \mathbf {t} = | | \mathbf {t} _ {1} - \mathbf {t} _ {2} | |. +$$ + +Then we make a formal statement of target-referenced reactive grasping based on the background definition above. Given $\mathcal{G}^1$ at first frame as reference grasp, for any timestep $j > 1$ , target-referenced reactive grasping is aimed at finding the grasp $\mathcal{G}^j$ that minimizes the grasp distance: + +$$ +\mathcal {G} ^ {j *} = \underset {\mathcal {G} ^ {j}} {\arg \min } \mathcal {D} \left(\mathcal {G} ^ {j}, \mathbf {T} ^ {j 1} \mathcal {G} ^ {1}\right), \tag {3} +$$ + +where $\mathbf{T}^{j1}\in \mathbb{R}^{4\times 4}$ means the ground-truth transformation matrix of the reference grasp from the initial frame to the frame at timestep $j$ in the camera's coordinate system. + +# 4. Method + +In this section, we detail our grasp tracking pipeline, as illustrated in Fig.2. For a target grasp, we first detect grasp + +![](images/2880805607c2eb692469878d5fbe94438f9cce41199c20fc4bccfc6bfd82d99e.jpg) +Figure 3. Visualization of grasp features representation. + +candidates across the scene. Then we discover strong correspondences between the target grasp and grasp candidates by attentional grasp features aggregation. We finally use a memory-augmented refinement network to correct the predictions towards the target grasp. + +# 4.1. Grasp Detector + +For each incoming new frame, we generate a grasp set along with it. This work uses an off-the-shelf GraspNet baseline [9] for proposals of grasp candidates. GraspNet is a learned grasp pose detector that takes the partial point cloud of the scene $\mathcal{P}$ as input and outputs a grasp set $\mathbf{G}$ across the entire scene. In GraspNet, a point cloud $\mathcal{P}$ of size $N\times 3$ is downsampled and transformed to feature vectors of size $M\times C$ by the PointNet++ backbone network [26] where $C$ denotes the channel dimension of the features and $M$ is the number of points after farthest point sampling [7]. For each sampled point, GraspNet predicts a grasp $\mathcal{G}$ and its confidence score. The detected grasp set is then sent to the following steps for grasp tracking. While GraspNet performs well in this work, the following steps are not limited to any specific grasp detector. + +# 4.2. Grasp Correspondence Network + +We follow the formulation of grasp correspondence in [8]. A many-to-many correspondence matrix is learned during training, given two grasp sets in consecutive frames. During testing, we can choose the grasp pose with the highest correspondence score with the target pose in the previous frame as the tracked pose. + +Grasp Features Representation First we illustrate how we represent grasp features. For $M$ grasp poses in a scene, + +we follow [8] to extract seed features, geometric features and appearance features for them, each having a shape of $M \times C$ . In addition, since it is quite common in the real world that multiple similar objects exist simultaneously in the scene, grasps on similar objects cannot be distinguished in such situations which results in correspondence switching between objects. Hence we further add global scene features of size $M \times C$ to provide information about the scene state. This global feature is extracted by max-pooling pointwise features from the backbone output, thus is the same for each of $M$ grasp poses. Finally, we concatenate geometric features, visual features, seed point features and global scene features to construct grasp features of dimension $4C$ for each grasp, denoted by $\mathbf{x}$ . We show the above process in Fig.3. We present more details of grasp features representation in the supplementary material. + +Positional Information To provide positional information for grasps, we encode grasp poses by MLP layers of size $(12, 2C, 4C)$ . These feature vectors are then added to the raw grasp features $\mathbf{x}$ , such that grasps can be treated in a similar way to words in a sentence. + +$$ +\mathbf {x} \leftarrow \mathbf {x} + \mathbf {M L P} ^ {(1)} (p), \tag {4} +$$ + +where $p$ is the grasp pose of dimension 12, consisting of 9 parameters for the rotation matrix and 3 for translation. We use this feature as the final grasp features representation. Such a process makes our method position-dependent and benefits following the grasp features aggregation step. + +Attentional Grasp Features Aggregation Thanks to the above-mentioned grasp features, it is now straightforward to reason about the visual, geometric and positional properties of grasps jointly. For a grasp $\mathcal{G}_i$ , we first embed its associated grasp features using MLP layers and transform its feature size from $4C$ to $2C$ by MLP layers. + +$$ +\mathbf {f} ^ {g _ {i}} = \mathbf {M L P} ^ {(2)} \left(\mathbf {x} _ {i}\right), \tag {5} +$$ + +Apart from the features of a grasp itself, given a pair of grasp sets, it is intuitive that interacting with contextual grasps within grasp sets or across grasp sets is critical to reduce ambiguities and increase the distinctiveness of grasps as well. [8] adopts vanilla MLP blocks which cannot model long-range relationships among grasp poses. In this paper, we take advantage of the global receptive field of the attentional graph neural network (GNN) to model such dependencies. Following [30], self-attention and cross-attention [35] are adopted for context aggregation within grasp sets and across grasp sets. In particular, features are aggregated unidirectionally in self-attention and bidirectionally in cross-attention. Given two grasp sets $\mathbf{G}_1$ and $\mathbf{G}_2$ + +that each consists of $M$ grasps, grasp features aggregation generates matching descriptors $\mathbf{f}^{m_1}$ , $\mathbf{f}^{m_2} \in \mathbb{R}^{M \times 2C}$ . + +Here we outline the workflow of grasp features aggregation. In self-attention, we only process grasps within $\mathbf{G}_1$ or $\mathbf{G}_2$ . Taking $\mathbf{G}_1$ as an example, grasp features $\mathbf{f}^{\mathbf{G}_1}$ of size $(M \times 2C)$ are first mapped to three elements which are named query, key and value in convention by learnable weights $\mathbf{W}_{\mathbf{Q}}$ , $\mathbf{W}_{\mathbf{K}}$ and $\mathbf{W}_{\mathbf{V}} \in \mathbb{R}^{2C \times 2C}$ . The attention computation can then be written as $\mathbf{A} = (\mathbf{Q} \cdot \mathbf{K}^T) / \sqrt{C}$ where $\mathbf{A}$ denotes the attention matrix and has a size of $(M \times M)$ and $(\cdot)$ is the dot product operator. $\mathbf{A}$ actually measures the similarity between query and key. After that, the message propagation step can be defined as $\mathbf{M} = \mathrm{softmax}(\mathbf{A}) \cdot \mathbf{V}$ . $\mathbf{M}$ is later fed into MLP layers of size $(2C, 4C, 2C)$ to produce the output features which have the same size as $\mathbf{f}^{\mathbf{G}_1}$ . Since $\mathbf{M}$ can be regarded as the sum of value vectors weighted by the similarity measure, therefore the output feature vectors represent correspondences among grasps in $\mathbf{G}_1$ . In cross-attention, we consider grasp features from both $\mathbf{G}_1$ and $\mathbf{G}_2$ . To produce the output feature vector, the only thing different is that key and value are no longer generated from $\mathbf{f}^{\mathbf{G}_1}$ but from $\mathbf{f}^{\mathbf{G}_2}$ . Intuitively, in self-attention information flows only within $\mathbf{G}_1$ or $\mathbf{G}_2$ whereas in cross-attention information flows from $\mathbf{G}_2$ to $\mathbf{G}_1$ or vice versa. We consider the output of cross-attention layers as a representation of correspondence between $\mathbf{G}_1$ and $\mathbf{G}_2$ . We interleave self-attention and cross-attention layers by $N_c$ times and get the matching descriptors $\mathbf{f}^{m_1}, \mathbf{f}^{m_2}$ which are later used to compute correspondences within and across grasp sets. + +Correspondence Estimation Given $\mathbf{f}^{m_1},\mathbf{f}^{m_2}$ , we can simply compute the correspondence matrix $S\in \mathbb{R}^{M\times M}$ For each pair of grasps from these two sets, the correspondence score can be calculated by cosine similarity + +$$ +\mathcal {S} (i \cdot j) = \frac {\mathbf {f} _ {i} ^ {m _ {1}} \cdot \mathbf {f} _ {j} ^ {m _ {2}}}{\left| \left| \mathbf {f} _ {i} ^ {m _ {1}} \right| \right| _ {2} \cdot \left| \left| \mathbf {f} _ {j} ^ {m _ {2}} \right| \right| _ {2}}. \tag {6} +$$ + +Since cosine similarity is naturally normalized, it makes training more stable. Taking advantage of contrastive learning, during training a target grasp can be assigned the same class to any number of grasps in the other grasp set. This is important because some grasps are close in terms of translation and some are close in terms of rotation. During training, our method allows parallel prediction of multiple target grasps. We simply pick grasps that have the highest correspondence scores with respect to target grasps. Before entering the next stage, a grasp set is filtered to $K$ grasps by removing upward grasps and grasps with low correspondence scores with respect to all other grasps. If the number of grasps is less than $K$ , we repeat the first grasp. In real robot experiments, only one target grasp is used. + +# 4.3. Memory-augmented Refinement Network + +Since the correspondence is estimated between fixed grasp sets, the tracking performance greatly relies on the grasp detector which is not guaranteed to generate identical grasps across different frames. Also, consecutive correspondence matching may lead to an accumulation of errors. Therefore, to further improve the semantic consistency with the reference grasp pose, a refinement network is used to correct predictions made by correspondences towards the target grasp. + +To be specific, we store the tracked grasp poses on past frames and their associated features in a memory pool. For each forward pass, the refinement network uses historical observations selected from the memory pool to refine the parameters of the selected grasp pose in the current frame. Here the historical frames from the memory pool are selected according to their time indexes. In practice, we select the reference pose in the first frame and apply uniform sampling for simplicity in the interval $(1,t - 1]$ to select past frames, where $t$ denotes the current time index. + +For a $L$ frame grasping sequence that consists of 1 frame of tracking target, 1 current frame, and $L - 2$ frames selected from the memory pool, grasp features of size $L \times (4C)$ are first transformed to $L \times 2C$ by MLP layers. After that, we simply repeat the target frame and concatenate it with other frames, resulting in features of size $(L - 1) \times (2C)$ . Such that all the selected frames are conditioned on the tracking target. We then apply temporal encoding which is similar to positional encoding introduced in [35]. These features are then fed into a refinement network which outputs $(L - 1)$ refined poses. Note that both historical grasp poses and the grasp pose in the current frame are updated. This refinement network has the same architecture as the grasp correspondence network, except that only self-attention layers are adopted. Moreover, the rotation is parameterized by the 6D representation introduced in [43]. Causal mask is used to avoid causality confusion as well. + +# 4.4. Supervision + +The objective function used to train the above pipeline can be divided into two parts: + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \beta \mathcal {L} _ {R, t}, \tag {7} +$$ + +where $\mathcal{L}_c$ denotes the loss for correspondence learning and $\mathcal{L}_{R,t}$ denotes the loss for grasp pose refinement. $\beta$ is the weighting term which we set as 1 during training. + +For $\mathcal{L}_c$ , we follow [8] and adopt supervised contrastive loss. For any two grasp poses with positive correspondence label, we denote them as $\mathrm{corr}(\mathcal{G}_i,\mathcal{G}_k) = 1$ or vice versa. Given two arbitrary grasp sets $\mathbf{G}_1$ and $\mathbf{G}_2$ , we first locate the positive grasp set $\mathbf{P}(i) = \{\mathcal{G}_k^2\in \mathbf{G}_2|\mathrm{corr}(\mathcal{G}_i,\mathcal{G}_k) = 1\}$ for any $\mathcal{G}_i\in \mathbf{G}_1$ . This correspondence loss can then be + +written as: + +$$ +\mathcal {L} _ {c} = \sum_ {\mathcal {G} _ {i} ^ {1} \in \mathbf {G} _ {1}} \frac {- 1}{| \mathbf {P} (i) |} \sum_ {\mathcal {G} _ {k} ^ {2} \in P (i)} \log \frac {\exp (\mathcal {S} (i \cdot k) / \tau)}{\sum \mathcal {G} _ {j} ^ {2} \in \mathbf {G} _ {2} \exp (\mathcal {S} (i \cdot j) / \tau)}, \tag {8} +$$ + +where $|\mathbf{P}(i)|$ denotes the cardinality and $\tau$ is a scalar temperature parameter. This loss pulls together the positive grasp pairs in embedding space, whereas pushes apart negative grasp pairs meanwhile. + +For grasp pose refinement, we adopt a transformation loss which consists two terms for optimizing rotation and translation respectively. Consider $\mathcal{J} = \{j_1,j_2,\dots ,j_n\}$ that stores arbitrary number of time index, for any time step $j\in \mathcal{J}$ , given target grasp $\mathcal{G}^1$ , we first compute its transformed grasp in $j^{th}$ frame's coordinates, denoted by + +$$ +\mathcal {G} ^ {j 1} = \left(\mathbf {R} ^ {j 1} \mathbf {t} ^ {j 1}\right), \tag {9} +$$ + +We then define the transformation loss as: + +$$ +\begin{array}{l} \mathcal {L} _ {R, t} = \frac {1}{| \mathcal {I} |} \sum_ {j \in \mathcal {J}} \frac {1}{| \mathbf {G} _ {j} |} \sum_ {\mathcal {G} _ {i} ^ {j} \in \mathbf {G} _ {j}} | | (\mathbf {R} ^ {*}) ^ {- 1} \mathbf {R} ^ {j 1} - \mathbf {I} | | _ {2} (10) \\ + \frac {1}{| \mathcal {I} |} \sum_ {j \in \mathcal {J}} \frac {1}{| \mathbf {G} _ {j} |} \sum_ {\mathcal {G} _ {i} ^ {j} \in \mathbf {G} _ {j}} \| \mathbf {t} ^ {j 1} - \mathbf {t} ^ {*} \| _ {2}, (10) \\ \end{array} +$$ + +where $\mathbf{G}_j$ denotes grasp set at time step $j$ . + +# 5. Experiments + +# 5.1. Dataset + +The proposed pipeline is trained on GraspNet-1Billion dataset [9] which is a large-scale real-world dataset for grasp detection. It includes 89 objects, 190 scenes and 256 camera views for each scene. We follow [8] to generate grasp correspondence labels across viewpoints on the GraspNet-1Billion dataset and refer readers to [8] for more details. In addition, we further collect 30 scenes of moving objects for testing. For each frame, we record the RGBD images and objects' 6D poses. We manually annotate 10 grasps for the first frame of each scene as targets and further project the annotated grasps according to the 6D poses for the future frames. This test set is named Moving GraspNet. + +# 5.2. Pre-processing and Augmentation + +Before being fed into the network, the point cloud is first downsampled to 20000 points and the RGB values are normalized. We also remove points outside a pre-defined workspace. Furthermore, we adopt massive data augmentation to avoid over-fitting and enrich objects' motion. For each object in the scene, its point cloud is augmented on-the-fly by a random translation of $[-0.2, 0.2]$ meter in all XYZ directions and a random rotation of $[-30, 30]$ degrees + +around z-axis. Moreover the scene's point cloud is rotated by $[-30, 30]$ degrees around both x-axis and z-axis. Due to the space limitations, please refer to the supplementary materials for details about network implementation. + +# 5.3. Performance of Grasp Tracking + +Evaluation Metric We propose a new Multiple Grasp Pose Tracking (MGTA) metric which follows the idea of Multiple Object Tracking Accuracy (MOTA) [21] in multi-object tracking task, except that the similarity measure between two grasps is computed using Eq.(2). Details are given in supplementary materials. Apart from MGTA, we also report the mean translation error and rotation error to the tracking target of each sequence. + +We follow the same evaluation procedure on GraspNet1-Billion and Moving GraspNet. For each scene, we pick 10 reference grasps in the first frame which are later used for grasp tracking. To be specific, these reference grasps are detected by a grasp detector on GraspNet1-billion but manually labeled on Moving GraspNet. For each remaining frame, we compute the ground truth grasps using objects' 6D poses and segmentation mask of the scene. The metrics are calculated using the transformed reference grasps and predictions associated with each frame. + +Comparing with Representative Methods We compare our methods with two other baseline methods. A heuristic method is implemented as the first baseline. Given a grasp pose from the previous frame, its nearest neighbor detected in the current camera coordinates is picked. For the second method, we adopt Bundle Track [40] which is one representative method in the field of unseen object 6D pose tracking. It fits our task setting well since it requires no objects' CAD model. + +As it is shown in Tab.1 and Tab.2, our method outperforms baseline methods on both GraspNet-1Billion and Moving GraspNet. It is found in experiments that, in many cases, grasps cannot fall on the same objects if we only pick the nearest neighbors. Also, it can be drawn from the results that tracking results are improved significantly after refinement, especially for rotation. Note that MGTA will be negative if there is a large number of false positives, false negatives, and ID switches. + +# 5.4. Real Robot Experiments + +In real robot experiments, we use Flexiv Rizon robot arm with an Intel Realsense L515 depth camera mounted on the end-effector. A Robotiq-85 parallel-jaw gripper is used with 3D-printed extended fingertips attached to it. The real robot experiments run on an NVIDIA 3090 GPU at around 10 fps. Please refer to the supplementary materials for details of the objects. + +
MethodSeenSimilarNovel
MGTA↑n (cm)↓n (°)↓MGTA↑n (cm)↓n (°)MGTA↑n (cm)↓n (°)↓
Nearest-0.8115.295.98-0.8015.497.61-0.8316.696.90
Bundle Track [40]0.741.5215.800.721.4716.370.721.9315.41
Ours, w/o refinement-0.322.3153.24-0.302.6857.59-0.392.9358.49
Ours, M=5120.801.5810.890.801.7410.620.752.1015.03
Ours, M=10240.841.3410.150.851.529.840.781.9012.96
Ours, M=20480.851.279.820.861.359.500.791.8411.48
+ +![](images/b40619c72de1aebe5a2ca51abed56c6053ba4ed9f828fbd8e39f513154b090d2.jpg) +Figure 4. Illustration of tracking a toy dog and toy rabbit. We can see that the grasp poses are consistent, i.e., on the head of the dog and the ear of the rabbit. Yellow bounding box represents to the view limit of the camera. No box means the camera can see the entire scene. + +Table 1. GraspNet-1Billion evaluation results on seen, similar and novel objects respectively. $M$ denotes the number of grasp detected,its default value is 1024. + +
MethodMGTA↑n (cm)↓n (°)↓
Nearest-0.478.0967
Bundle Track0.377.6023.52
Ours, w/o refinement-0.436.2462.33
Ours, M=5120.475.5322.63
Ours, M=10240.524.6020.37
Ours, M=20480.544.1418.28
+ +Protocol For each frame, we receive RGBD images from the in-hand camera. We transform the point cloud to the camera coordinate system at ready state by the camera extrinsic for simplicity as well. Due to safety considerations, points outside a pre-defined workspace will be masked. After a grasp candidate generation phase, we select the grasp with the highest predicted score as a reference grasp and keep tracking it for each new frame. During the tracking phase, the robot moves to a pre-grasp pose which is 0.02 + +Table 2. Moving GraspNet evaluation results. All objects in Moving GraspNet are novel. $M$ denotes the number of grasp detected, its default value is 1024. + +
Object ID#Attempt#SuccessSuccess Rate
11080.8
21090.9
31080.8
41070.7
61080.8
71080.8
91080.8
101090.9
Total80650.8125
+ +Table 3. Real robot experiments results on single object scenes. + +meters higher than the actual grasp pose. The robot executes the grasp once three conditions are met simultaneously: 1) the translation distance between the gripper center and the target grasp in XY-plane is lower than 0.01 meter (the XYplane is parallel to the table); 2) the translation distance between the gripper center and the target grasp in 3D space is lower than 0.04 meters; 3) the rotation distance between the gripper and the target grasp in 3D space is lower than + +
Object ID#Attempt#SuccessSuccess Rate
7,7,8,81080.8
10,13,141090.9
1,11,15,16,17,181080.8
2,3,51090.9
9,121070.7
20,201080.8
Total60490.8167
+ +![](images/b73592b9ec5cfb3a4256d3b6323b9d9bc6d943b52b8c4b8417283762fa02a948.jpg) +Figure 5. Correspondence switching between similar objects. Red arrow indicates correct correspondence and blue arrow indicates wrong correspondence. + +15 degrees. The initial positions of the objects are chosen randomly. + +Results As shown in Tab.3 and Tab.4, we execute 20 grasp attempts for each scene. In single-object scenes, we randomly sample 8 objects for testing and we achieve a success rate of 81 percent. In multi-object scenes, we only choose objects with similar visual and geometric patterns to test the temporal smoothness and semantic consistency of the predicted grasps. Results suggest that single-object scenes and multi-object scenes have nearly the same success rate which further proves the effectiveness of our method. We illustrate one case from real robot experiments in Fig.4. + +# 5.5. Analysis + +Effect of Global Features In Sec.4.2, we incorporate global features to represent grasp features. While local grasp features are sufficient to identify corresponding local structures, grasp trials may fail due to correspondence switches among objects with similar visual and geometric patterns as we can see in Fig.5. Such phenomenon may be + +Table 4. Real robot experiments results on multi-object scenes. + +
Object IDMethod#Attempt#SW
7,7,8,8w/o global features105
Full model101
1,17,17w/o global features104
Full model100
+ +Table 5. Statistics of Correspondence Switching. SW represents correspondence switching. + +due to lacking awareness of the scene state and we propose to solve it by global features as what we state in Sec.4.2. In order to analyze the actual effectiveness of global features, we further record how frequent correspondence switching happens with or without global features. We conduct these experiments on identical objects which are placed close to each other in the scenes. Tab.5 suggests that the number of correspondence switches decreased significantly with global features. + +Failure Analysis There are three major types of failure cases found in real robot experiments. First, some predicted grasp candidates from the detector do not have good qualities. Naturally, tracking such targets may lead to failed grasps. Second, it is hard to heuristically plan the timing to close the gripper. Due to some grasps may fall on the edges of objects, it is common that the objects may slip away immediately when the robot receives the command to close the gripper. Furthermore, since we mount the camera on the wrist, the point cloud may become fragmented due to view limitation and occlusion, especially when the objects fall behind the gripper. In such cases, the predictions can be inaccurate. + +# 6. Conclusion + +We present a novel solution to reactive grasping in a target-referenced setting. For a reference grasp, our method predicts temporally smooth and semantically consistent grasp poses in the future frames without any prior knowledge of instance. Experiments conducted on both single-object and multi-object scenes show that our method provides reliable grasp plans under various environments. + +Acknowledgement This work was supported by the National Key R&D Program of China (No.2021ZD0110700), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Qi Zhi Institute, and Shanghai Science and Technology Commission (21511101200). + +Contributions: J. Liu and H-S. Fang initiated the project. J. Liu designed and trained the network and conducted the evaluation together with R. Zhang. J. Liu wrote the paper. H-S. Fang devised and mentored this project and edited the paper. M. Gou, H. Fang, and S. Xu constructed the testing dataset. C. Wang and H. Yan helped with the evaluation. C. Lu supervised the project and provided hardware support. + +# References + +[1] Iretiayo Akinola, Jingxi Xu, Shuran Song, and Peter K. Allen. Dynamic grasping with reachability and motion awareness. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9422-9429, 2021. 2 +[2] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2 +[3] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3 +[4] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 2 +[5] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Toward geometric deep slam. ArXiv, abs/1707.07410, 2017.3 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 337-33712, 2018. 3 +[7] Y. Eldar, M. Lindenbaum, M. Porat, and Y.Y. Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 4 +[8] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. arXiv preprint arXiv:2212.08333, 2022. 2, 4, 5, 6 +[9] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4, 6 +[10] Minghao Gou, Hao-Shu Fang, Zhanda Zhu, Sheng Xu, Chenxi Wang, and Cewu Lu. Rgb matters: Learning 7-dof grasp poses on monocular rgbd images. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13459-13466. IEEE, 2021. 2 +[11] Yinlin Hu, Joachim Hugonot, Pascal Fua, and Mathieu Salzmann. Segmentation-driven 6d object pose estimation. In CVPR, 2019. 2 +[12] Seungsu Kim, Ashwini Shukla, and Aude Billard. Catching objects in flight. IEEE Transactions on Robotics, 30(5):1049-1065, 2014. 2 +[13] Donghoon Lee, Onur C. Hamsici, Steven Feng, Prachee Sharma, and Thorsten Gernoth. Deeppro: Deep partial point cloud registration of objects. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5663-5672, 2021. 3 + +[14] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International journal of robotics research, 37(4-5):421-436, 2018. 2 +[15] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Görner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In 2019 International Conference on Robotics and Automation (ICRA), pages 3629-3635. IEEE, 2019. 2 +[16] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2011. 3 +[17] David Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91–, 11 2004. 2, 3 +[18] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26), 2019. 2 +[19] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Aleš Leonardis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43(5):1241-1256, 2019. 2 +[20] Arjun Menon, Benjamin Cohen, and Maxim Likhachev. Motion planning for smooth pickup of moving objects. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 453-460. IEEE, 2014. 2 +[21] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. ArXiv, abs/1603.00831, 2016. 6 +[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 1, 2 +[23] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2901-2910, 2019. 2 +[24] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3619-3625, 2020. 2 +[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[26] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4 +[27] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4g: Amodal single-view single-shot se (3) grasp + +detection in cluttered scenes. In Conference on robot learning, pages 53-65. PMLR, 2020. 2 +[28] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2021. 2 +[29] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3 +[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 4 +[31] Tanner Schmidt, Richard Newcombe, and Dieter Fox. Self-supervised visual descriptor learning for dense correspondence. IEEE Robotics and Automation Letters, 2(2):420-427, 2017. 3 +[32] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 2 +[33] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3 +[34] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt. Grasp pose detection in point clouds. The International Journal of Robotics Research (IJRR), 36(13-14):1455-1473, 2017. 2 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4, 5 +[36] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15964-15973, 2021. 2 +[37] Chen Wang, Roberto Martín-Martín, Danfei Xu, Jun Lv, Cewu Lu, Li Fei-Fei, Silvio Savarese, and Yuke Zhu. 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10059–10066, 2020. 2 +[38] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 2 +[39] Yue Wang and Justin M. Solomon. Deep closest point: Learning representations for point cloud registration. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3 +[40] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d + +models. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074. IEEE, 2021. 2, 6, 7 +[41] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E. Bekris. se(3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373, 2020. 2 +[42] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 1, 2 +[43] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 5 \ No newline at end of file diff --git a/2023/Target-Referenced Reactive Grasping for Dynamic Objects/images.zip b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c46d062c2a475d4a17aa1220e9d8c7e96e32d7d7 --- /dev/null +++ b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df383c7e4728266c775b256f40b4670ef3d186c66053bbfa89bb9f02f01b4266 +size 509841 diff --git a/2023/Target-Referenced Reactive Grasping for Dynamic Objects/layout.json b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00faab1543b25b0b39ed8191e6692e75d94302f5 --- /dev/null +++ b/2023/Target-Referenced Reactive Grasping for Dynamic Objects/layout.json @@ -0,0 +1,8423 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "type": "text", + "content": "Target-referenced Reactive Grasping for Dynamic Objects" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": "Jirong Liu" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Ruo Zhang" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Hao-Shu Fang" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Minghao Gou" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Hongjie Fang" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Chenxi Wang" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Sheng Xu" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Hengxu Yan" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": ", Cewu Lu" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": "Shanghai Qi Zhi institute, " + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, " + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 142, + 496, + 213 + ], + "type": "text", + "content": "Flexiv Robotics, LTD jirong@sjtu.edu.cn, {ruozhang0608, fhaoshu}@gmail.com, {gmh2015, galaxies, wcx1997, xs1020, hengxuyan, lucewu}@sjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 240, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 529 + ], + "type": "text", + "content": "Reactive grasping, which enables the robot to successfully grasp dynamic moving objects, is of great interest in robotics. Current methods mainly focus on the temporal smoothness of the predicted grasp poses but few consider their semantic consistency. Consequently, the predicted grasps are not guaranteed to fall on the same part of the same object, especially in cluttered scenes. In this paper, we propose to solve reactive grasping in a target-referenced setting by tracking through generated grasp spaces. Given a targeted grasp pose on an object and detected grasp poses in a new observation, our method is composed of two stages: 1) discovering grasp pose correspondences through an attentional graph neural network and selecting the one with the highest similarity with respect to the target pose; 2) refining the selected grasp poses based on target and historical information. We evaluate our method on a large-scale benchmark GraspNet-1Billion. We also collect 30 scenes of dynamic objects for testing. The results suggest that our method outperforms other representative methods. Furthermore, our real robot experiments achieve an average success rate of over 80 percent. Code and demos are available at: https://graspnet.net/reactive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "type": "text", + "content": "Reactive grasping is in great demand in the industry. For instance, in places where human-robot collaboration is heavily required like factories, stress on laborers will be significantly relieved if robots can receive tools from humans and complete the harder work for laborers. Such a vision is based on reactive grasping." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 643, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 677 + ], + "type": "text", + "content": "On the contrary to static environments, in reactive grasping, dynamic task setting poses new challenges for algorithm design. Previous research in this area mainly focuses" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 239, + 536, + 346 + ], + "blocks": [ + { + "bbox": [ + 320, + 239, + 536, + 346 + ], + "lines": [ + { + "bbox": [ + 320, + 239, + 536, + 346 + ], + "spans": [ + { + "bbox": [ + 320, + 239, + 536, + 346 + ], + "type": "image", + "image_path": "605780c57469374cb35909e698f20e87ce52c6f039065b5a5da40e7ff9b62a59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 349, + 533, + 453 + ], + "blocks": [ + { + "bbox": [ + 320, + 349, + 533, + 453 + ], + "lines": [ + { + "bbox": [ + 320, + 349, + 533, + 453 + ], + "spans": [ + { + "bbox": [ + 320, + 349, + 533, + 453 + ], + "type": "image", + "image_path": "5b978c230c54871cba1ce8cf59944b78d8f13261f9f3a9fbd2d5654f2f458dff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 462, + 547, + 519 + ], + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 519 + ], + "type": "text", + "content": "Figure 1. (a) Classic reactive grasping guarantees the smoothness of the grasp poses but cannot predicts grasps on the same part of the hammer. (b) Our target-referenced reactive grasping takes semantic consistency into consideration. The generated grasp poses across frames are illustrated with blue grippers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "type": "text", + "content": "on planning temporally smooth grasps [22, 42] to avoid wavy and jerky robot motion. Few of them pay attention to its semantic consistency. In short, given a targeted grasp at the first frame, we want the robot to grasp the same part of the object in the following frames. Additionally, it is not guaranteed that grasp predictions made by classical methods fall on the same object in cluttered scenes. Hence, most of their experiments are conducted on single-object scenes. Unlike previous works, this work is aimed at achieving temporally smooth and semantically consistent reactive grasping in clutter given a targeted grasp. We refer to such a task setting as target-referenced reactive grasping as shown in Fig.1. Note that despite robot handover is a major application scenario of reactive grasping, this work focuses on a more general task setting - dynamic object grasping." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "† Cewu Lu is the corresponding author, a member of Qing Yuan Research Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8824" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "A naive idea to solve this task is to generate reference grasp poses for the initial scene and consecutively track the object's 6D pose. As the object moves, the initial grasp pose can be projected to a new coming frame based on the object's 6D pose. Although such an idea seems to be natural and valid, some bottleneck greatly degrades its viability. First of all, the solution to reactive grasping should be able to handle objects' motion in real-time, meaning that it requires fast inference speed and immediate response to continuous environmental changes. However, 6D pose tracking may be time-consuming due to commonly-used instance segmentation [11,40]. Second, 6D pose tracking usually requires objects' prior knowledge, such as CAD models [4,41], which is not always available in the real world as well or achieves only category-level generalization [37]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 540 + ], + "type": "text", + "content": "Different from tracking objects, we propose to track grasps by a two-stage policy instead. We also comply with the restriction that no prior knowledge of the objects is allowed. Given a target grasp on a partial-view point cloud, we first discover its corresponding grasp among future frame's detected grasp poses as coarse estimation. These grasp poses can be given by an off-the-shelf grasp detector. Inspired by recent progress in local feature matching, which often uses image descriptors like SIFT [17] to describe interesting regions of images, we view grasp poses and their corresponding features as geometric descriptors on a partial-view point cloud. Based on such an assumption, we can simply estimate correspondences between two grasp sets from two different observation frames by matching the associated grasp features. Note that in opposition to classical local feature matching, features of the entire scene are also incorporated to help achieve global awareness. Furthermore, consecutive matching along an observation sequence may lead to the accumulation of error, on top of the coarse estimation through correspondence matching, we further use a memory-augmented coarse-to-fine module which uses both target grasp features and historical grasp features to refine the grasp tracking results for better temporal smoothness and semantic consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "type": "text", + "content": "We conduct extensive experiments on two benchmarks to evaluate our method and demonstrate its effectiveness. The results show that our method outperforms two representative baseline methods. We also conduct real robot experiments on both single-object scenes and cluttered scenes. We report success rates of " + }, + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "type": "inline_equation", + "content": "81.25\\%" + }, + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "type": "text", + "content": " for single-object scenes and " + }, + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "type": "inline_equation", + "content": "81.67\\%" + }, + { + "bbox": [ + 46, + 540, + 288, + 624 + ], + "type": "text", + "content": " for multi-object scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 638, + 139, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 139, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 139, + 650 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 658, + 189, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 189, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 189, + 671 + ], + "type": "text", + "content": "2.1. Grasping in Static Scenes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Thanks to the advances in the field of 3D perception [2, 25, 26], 6-DoF static grasping which takes point cloud as input is gaining increasing attention in both the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "research community and industry. In general, there are two lines of work that have been explored. The first line adopts a sampling-evaluation manner [15, 18, 23, 34]. Grasp candidates are sampled or generated on the point clouds and then their qualities are evaluated by neural networks. The second line processes the point cloud of the entire scene and predicts grasp poses across the scene [9, 10, 24, 27, 36] in an end-to-end manner. Compared to the first line, the end-to-end strategy achieves better balance in terms of the speed-accuracy tradeoff. Some researchers also propose to predict actions in continuous space [14, 32, 38]. Nevertheless, these methods can be adversely impacted by distribution shifts. Our method uses a pre-trained grasp pose detector from [9] which belongs to the second research line." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 256, + 416, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 416, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 416, + 269 + ], + "type": "text", + "content": "2.2. Reactive Grasping" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 409 + ], + "type": "text", + "content": "Though reactive grasping remains much less explored, a small amount of literature has partially investigated this more difficult problem. Most of these papers rely on object motion tracking. [1] predicts object motion and adaptively plans the grasp for a reachable grasp candidate. However, the object's motion is limited to few prototypes and thus it cannot handle the case where the objects move unpredictably. [19] tracks the objects' poses, yet the grasp planning phase is limited by a set of fixed grasp trajectories. [12,20,28] requires prior knowledge of the objects which is not always available in the real-world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 411, + 548, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 548, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 548, + 639 + ], + "type": "text", + "content": "Some attempts on reactive grasping without prior information have been conducted recently. [22] picks the nearest grasp poses across frames. [42] samples grasp candidates by adding disturbances to grasp poses from the previous frame and evaluates their qualities in the current frame's scene. These methods focused on the smoothness of the grasp pose sequence but did not impose semantic constraints. Their predicted grasp poses can switch between different objects. Thus they mainly demonstrate single object grasping, and [42] requires an extra human hand segmentation module to avoid grasping the human's hand during handover while our correspondence matching module mitigates such a problem since grasps on hands and grasps on objects have low correspondence scores. [8] proposed a temporal association module to alleviate this dilemma. However, it only considers correspondence across two frames which is prone to error accumulation. In this paper, we further adopt a refinement module to adjust the grasp pose according to the target and historical grasp poses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Some other works adopt continuous action prediction [14, 32]. Even if it is not designed for dynamic environments, it can be executed on moving objects thanks to the closed-loop controller. These methods may suffer greatly from distribution shifts. Different from these methods, our method tracks grasp poses explicitly." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8825" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 527, + 289 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 527, + 289 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 527, + 289 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 527, + 289 + ], + "type": "image", + "image_path": "e6b513f18b4c1ad0ff11ab235f415d64b0b077abf86eaf78b5837e6418bc3198.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed method for target-referenced reactive grasping. Given a target grasp and colored point clouds of a new coming frame with " + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "text", + "content": " points, our pipeline first detects " + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "text", + "content": " grasp candidates of size " + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "inline_equation", + "content": "M \\times 12" + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "text", + "content": " (3 for translation and 9 for rotation) across the scene. It then extracts grasp features of size " + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "inline_equation", + "content": "M \\times 4C" + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "text", + "content": " for each detected grasp by cropping and embedding the points along the grasping direction. Furthermore, grasp features aggregation is used to estimate corresponding grasps between the previous frame and the current frame. After filtering, the correspondence score matrix produces " + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 297, + 545, + 373 + ], + "type": "text", + "content": " valid pairs of grasp. The predicted corresponding grasps are then fed into a refinement network along with the target grasp and other frames selected from a memory pool to be updated toward the target. Contents in the memory pool are updated by the refined values as well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 387, + 277, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 277, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 277, + 400 + ], + "type": "text", + "content": "2.3. Local Feature Matching in Computer Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 406, + 289, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 289, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 289, + 634 + ], + "type": "text", + "content": "Unlike global features which summarize the entire scene, local features distribute densely across the scene and describe interesting regions. Under slightly different camera views, local feature matching is aimed at recognizing distinctive regions and establishing associations in images or point clouds. The mainstream of local feature matching includes detector-based matching and detector-free matching. For detector-based matching, classical hand-crafted features such as SIFT [17] and ORB [29] are widely adopted. With the development of deep learning, learned local features [5, 6] achieve satisfying results as well. Different from detector-based methods that extract sparse local features, detector-free methods establish pixel-wise or pointwise dense features [3, 13, 16, 31, 33]. With the extracted local features, nearest neighbor search or learning-based approach [30, 39] are often used in the matching phase. In our task, we consider each grasp brings rich geometric and visual patterns of the grasped local patch and our grasp correspondence are built upon these local features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 645, + 174, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 174, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 174, + 658 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "In this section, we first briefly introduce some notations and metrics for grasp pose and grasp distance in 3D space, followed by formulating the problem of target-referenced reactive grasping." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 388, + 432, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 388, + 432, + 400 + ], + "spans": [ + { + "bbox": [ + 318, + 388, + 432, + 400 + ], + "type": "text", + "content": "We define a grasp pose " + }, + { + "bbox": [ + 318, + 388, + 432, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 318, + 388, + 432, + 400 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 400, + 407, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 407, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 400, + 407, + 545, + 419 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = (\\mathbf {R} \\mathbf {t}), \\tag {1}", + "image_path": "7ca2cccb6b5951c6232fb42b56c0bbb83700ebde7fc74e751ce8fb474dd51ab7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "content": " denotes the rotation of the grasp pose and " + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "content": " denotes the translation. Consider two arbitrary grasps " + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2" + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "content": ", a grasp distance measure is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 468, + 457, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 468, + 457, + 481 + ], + "spans": [ + { + "bbox": [ + 354, + 468, + 457, + 481 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\left(\\mathcal {G} _ {1}, \\mathcal {G} _ {2}\\right) = \\Delta \\mathbf {R} + \\Delta \\mathbf {t},", + "image_path": "e2fdbf850d397e0414cf36511c24341ee77ba0146f3673df821948cf8269fca9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 484, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 484, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 354, + 484, + 545, + 509 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {R} = \\operatorname {a r c c o s} \\frac {\\operatorname {t r a c e} \\left(\\mathbf {R} _ {1} ^ {\\top} \\mathbf {R} _ {2}\\right) - 1}{2}, \\tag {2}", + "image_path": "964541d36ae393118f4d1eb6b796e151643312b3f9083694c7e64a15c16f3343.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 510, + 427, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 510, + 427, + 523 + ], + "spans": [ + { + "bbox": [ + 355, + 510, + 427, + 523 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {t} = | | \\mathbf {t} _ {1} - \\mathbf {t} _ {2} | |.", + "image_path": "139c1976d2ea8f834158491460ece88e66ab77505ef408291681620a4ea6e55a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "content": "Then we make a formal statement of target-referenced reactive grasping based on the background definition above. Given " + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^1" + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "content": " at first frame as reference grasp, for any timestep " + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "inline_equation", + "content": "j > 1" + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "content": ", target-referenced reactive grasping is aimed at finding the grasp " + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^j" + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "content": " that minimizes the grasp distance:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 594, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 594, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 362, + 594, + 545, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {j *} = \\underset {\\mathcal {G} ^ {j}} {\\arg \\min } \\mathcal {D} \\left(\\mathcal {G} ^ {j}, \\mathbf {T} ^ {j 1} \\mathcal {G} ^ {1}\\right), \\tag {3}", + "image_path": "dae43a481e672dd32db7208ee9af920899b3fe455037c525b0a6dd6f614ab3db.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{j1}\\in \\mathbb{R}^{4\\times 4}" + }, + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "type": "text", + "content": " means the ground-truth transformation matrix of the reference grasp from the initial frame to the frame at timestep " + }, + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 623, + 545, + 660 + ], + "type": "text", + "content": " in the camera's coordinate system." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "In this section, we detail our grasp tracking pipeline, as illustrated in Fig.2. For a target grasp, we first detect grasp" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8826" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 279, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 279, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 279, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 279, + 274 + ], + "type": "image", + "image_path": "2880805607c2eb692469878d5fbe94438f9cce41199c20fc4bccfc6bfd82d99e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 282, + 268, + 293 + ], + "lines": [ + { + "bbox": [ + 65, + 282, + 268, + 293 + ], + "spans": [ + { + "bbox": [ + 65, + 282, + 268, + 293 + ], + "type": "text", + "content": "Figure 3. Visualization of grasp features representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 365 + ], + "type": "text", + "content": "candidates across the scene. Then we discover strong correspondences between the target grasp and grasp candidates by attentional grasp features aggregation. We finally use a memory-augmented refinement network to correct the predictions towards the target grasp." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 373, + 142, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 142, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 142, + 386 + ], + "type": "text", + "content": "4.1. Grasp Detector" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": "For each incoming new frame, we generate a grasp set along with it. This work uses an off-the-shelf GraspNet baseline [9] for proposals of grasp candidates. GraspNet is a learned grasp pose detector that takes the partial point cloud of the scene " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " as input and outputs a grasp set " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " across the entire scene. In GraspNet, a point cloud " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "N\\times 3" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " is downsampled and transformed to feature vectors of size " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "M\\times C" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " by the PointNet++ backbone network [26] where " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " denotes the channel dimension of the features and " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " is the number of points after farthest point sampling [7]. For each sampled point, GraspNet predicts a grasp " + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 392, + 287, + 572 + ], + "type": "text", + "content": " and its confidence score. The detected grasp set is then sent to the following steps for grasp tracking. While GraspNet performs well in this work, the following steps are not limited to any specific grasp detector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 581, + 222, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 222, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 222, + 594 + ], + "type": "text", + "content": "4.2. Grasp Correspondence Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 600, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 672 + ], + "type": "text", + "content": "We follow the formulation of grasp correspondence in [8]. A many-to-many correspondence matrix is learned during training, given two grasp sets in consecutive frames. During testing, we can choose the grasp pose with the highest correspondence score with the target pose in the previous frame as the tracked pose." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Grasp Features Representation First we illustrate how we represent grasp features. For " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " grasp poses in a scene," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": "we follow [8] to extract seed features, geometric features and appearance features for them, each having a shape of " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "M \\times C" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": ". In addition, since it is quite common in the real world that multiple similar objects exist simultaneously in the scene, grasps on similar objects cannot be distinguished in such situations which results in correspondence switching between objects. Hence we further add global scene features of size " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "M \\times C" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " to provide information about the scene state. This global feature is extracted by max-pooling pointwise features from the backbone output, thus is the same for each of " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " grasp poses. Finally, we concatenate geometric features, visual features, seed point features and global scene features to construct grasp features of dimension " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "4C" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": " for each grasp, denoted by " + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": ". We show the above process in Fig.3. We present more details of grasp features representation in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "content": "Positional Information To provide positional information for grasps, we encode grasp poses by MLP layers of size " + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "inline_equation", + "content": "(12, 2C, 4C)" + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "content": ". These feature vectors are then added to the raw grasp features " + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "content": ", such that grasps can be treated in a similar way to words in a sentence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 377, + 353, + 545, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 353, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 377, + 353, + 545, + 368 + ], + "type": "interline_equation", + "content": "\\mathbf {x} \\leftarrow \\mathbf {x} + \\mathbf {M L P} ^ {(1)} (p), \\tag {4}", + "image_path": "95c5992e53cd96c16e75cabc58b8c11957b0cf81ff06b9dc3fa7d350d6e02a3c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 374, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 374, + 545, + 435 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 374, + 545, + 435 + ], + "type": "text", + "content": " is the grasp pose of dimension 12, consisting of 9 parameters for the rotation matrix and 3 for translation. We use this feature as the final grasp features representation. Such a process makes our method position-dependent and benefits following the grasp features aggregation step." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "text", + "content": "Attentional Grasp Features Aggregation Thanks to the above-mentioned grasp features, it is now straightforward to reason about the visual, geometric and positional properties of grasps jointly. For a grasp " + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "text", + "content": ", we first embed its associated grasp features using MLP layers and transform its feature size from " + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "inline_equation", + "content": "4C" + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "inline_equation", + "content": "2C" + }, + { + "bbox": [ + 304, + 452, + 545, + 524 + ], + "type": "text", + "content": " by MLP layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 536, + 545, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 536, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 382, + 536, + 545, + 551 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {g _ {i}} = \\mathbf {M L P} ^ {(2)} \\left(\\mathbf {x} _ {i}\\right), \\tag {5}", + "image_path": "aadbc76bd833067a296a71671ad621615f0237efd829f8a1830d6b7959397622.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Apart from the features of a grasp itself, given a pair of grasp sets, it is intuitive that interacting with contextual grasps within grasp sets or across grasp sets is critical to reduce ambiguities and increase the distinctiveness of grasps as well. [8] adopts vanilla MLP blocks which cannot model long-range relationships among grasp poses. In this paper, we take advantage of the global receptive field of the attentional graph neural network (GNN) to model such dependencies. Following [30], self-attention and cross-attention [35] are adopted for context aggregation within grasp sets and across grasp sets. In particular, features are aggregated unidirectionally in self-attention and bidirectionally in cross-attention. Given two grasp sets " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8827" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "that each consists of " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " grasps, grasp features aggregation generates matching descriptors " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{m_1}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{m_2} \\in \\mathbb{R}^{M \\times 2C}" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": "Here we outline the workflow of grasp features aggregation. In self-attention, we only process grasps within " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". Taking " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " as an example, grasp features " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathbf{G}_1}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "(M \\times 2C)" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " are first mapped to three elements which are named query, key and value in convention by learnable weights " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{K}}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{2C \\times 2C}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". The attention computation can then be written as " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = (\\mathbf{Q} \\cdot \\mathbf{K}^T) / \\sqrt{C}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " denotes the attention matrix and has a size of " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "(M \\times M)" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " is the dot product operator. " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " actually measures the similarity between query and key. After that, the message propagation step can be defined as " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\mathrm{softmax}(\\mathbf{A}) \\cdot \\mathbf{V}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " is later fed into MLP layers of size " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "(2C, 4C, 2C)" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " to produce the output features which have the same size as " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathbf{G}_1}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " can be regarded as the sum of value vectors weighted by the similarity measure, therefore the output feature vectors represent correspondences among grasps in " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". In cross-attention, we consider grasp features from both " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". To produce the output feature vector, the only thing different is that key and value are no longer generated from " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathbf{G}_1}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " but from " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathbf{G}_2}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". Intuitively, in self-attention information flows only within " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " whereas in cross-attention information flows from " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " or vice versa. We consider the output of cross-attention layers as a representation of correspondence between " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": ". We interleave self-attention and cross-attention layers by " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " times and get the matching descriptors " + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{m_1}, \\mathbf{f}^{m_2}" + }, + { + "bbox": [ + 46, + 97, + 289, + 434 + ], + "type": "text", + "content": " which are later used to compute correspondences within and across grasp sets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "type": "text", + "content": "Correspondence Estimation Given " + }, + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{m_1},\\mathbf{f}^{m_2}" + }, + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "type": "text", + "content": " , we can simply compute the correspondence matrix " + }, + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{M\\times M}" + }, + { + "bbox": [ + 46, + 450, + 287, + 499 + ], + "type": "text", + "content": " For each pair of grasps from these two sets, the correspondence score can be calculated by cosine similarity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 507, + 287, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 507, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 287, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (i \\cdot j) = \\frac {\\mathbf {f} _ {i} ^ {m _ {1}} \\cdot \\mathbf {f} _ {j} ^ {m _ {2}}}{\\left| \\left| \\mathbf {f} _ {i} ^ {m _ {1}} \\right| \\right| _ {2} \\cdot \\left| \\left| \\mathbf {f} _ {j} ^ {m _ {2}} \\right| \\right| _ {2}}. \\tag {6}", + "image_path": "24d421e66472ddb89b3dd64a9eb71c22862469946af010471b6c9a2be522c5cb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "Since cosine similarity is naturally normalized, it makes training more stable. Taking advantage of contrastive learning, during training a target grasp can be assigned the same class to any number of grasps in the other grasp set. This is important because some grasps are close in terms of translation and some are close in terms of rotation. During training, our method allows parallel prediction of multiple target grasps. We simply pick grasps that have the highest correspondence scores with respect to target grasps. Before entering the next stage, a grasp set is filtered to " + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": " grasps by removing upward grasps and grasps with low correspondence scores with respect to all other grasps. If the number of grasps is less than " + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": ", we repeat the first grasp. In real robot experiments, only one target grasp is used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 524, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 524, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 524, + 85 + ], + "type": "text", + "content": "4.3. Memory-augmented Refinement Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 198 + ], + "type": "text", + "content": "Since the correspondence is estimated between fixed grasp sets, the tracking performance greatly relies on the grasp detector which is not guaranteed to generate identical grasps across different frames. Also, consecutive correspondence matching may lead to an accumulation of errors. Therefore, to further improve the semantic consistency with the reference grasp pose, a refinement network is used to correct predictions made by correspondences towards the target grasp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "type": "text", + "content": "To be specific, we store the tracked grasp poses on past frames and their associated features in a memory pool. For each forward pass, the refinement network uses historical observations selected from the memory pool to refine the parameters of the selected grasp pose in the current frame. Here the historical frames from the memory pool are selected according to their time indexes. In practice, we select the reference pose in the first frame and apply uniform sampling for simplicity in the interval " + }, + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "type": "inline_equation", + "content": "(1,t - 1]" + }, + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "type": "text", + "content": " to select past frames, where " + }, + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 198, + 546, + 317 + ], + "type": "text", + "content": " denotes the current time index." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": "For a " + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": " frame grasping sequence that consists of 1 frame of tracking target, 1 current frame, and " + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "inline_equation", + "content": "L - 2" + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": " frames selected from the memory pool, grasp features of size " + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "inline_equation", + "content": "L \\times (4C)" + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": " are first transformed to " + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "inline_equation", + "content": "L \\times 2C" + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": " by MLP layers. After that, we simply repeat the target frame and concatenate it with other frames, resulting in features of size " + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "inline_equation", + "content": "(L - 1) \\times (2C)" + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": ". Such that all the selected frames are conditioned on the tracking target. We then apply temporal encoding which is similar to positional encoding introduced in [35]. These features are then fed into a refinement network which outputs " + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "inline_equation", + "content": "(L - 1)" + }, + { + "bbox": [ + 304, + 319, + 546, + 521 + ], + "type": "text", + "content": " refined poses. Note that both historical grasp poses and the grasp pose in the current frame are updated. This refinement network has the same architecture as the grasp correspondence network, except that only self-attention layers are adopted. Moreover, the rotation is parameterized by the 6D representation introduced in [43]. Causal mask is used to avoid causality confusion as well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 529, + 384, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 384, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 384, + 542 + ], + "type": "text", + "content": "4.4. Supervision" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 548, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 545, + 572 + ], + "type": "text", + "content": "The objective function used to train the above pipeline can be divided into two parts:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 582, + 545, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 582, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 388, + 582, + 545, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c} + \\beta \\mathcal {L} _ {R, t}, \\tag {7}", + "image_path": "20b02fcb891cc6cf0ff853601393a5ab6b117830159d2aba1bc242567d4a12d9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "text", + "content": " denotes the loss for correspondence learning and " + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{R,t}" + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "text", + "content": " denotes the loss for grasp pose refinement. " + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 605, + 545, + 641 + ], + "type": "text", + "content": " is the weighting term which we set as 1 during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ", we follow [8] and adopt supervised contrastive loss. For any two grasp poses with positive correspondence label, we denote them as " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{corr}(\\mathcal{G}_i,\\mathcal{G}_k) = 1" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " or vice versa. Given two arbitrary grasp sets " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ", we first locate the positive grasp set " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(i) = \\{\\mathcal{G}_k^2\\in \\mathbf{G}_2|\\mathrm{corr}(\\mathcal{G}_i,\\mathcal{G}_k) = 1\\}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i\\in \\mathbf{G}_1" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". This correspondence loss can then be" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8828" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 92, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 92, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 92, + 83 + ], + "type": "text", + "content": "written as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 90, + 286, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 90, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 48, + 90, + 286, + 132 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\sum_ {\\mathcal {G} _ {i} ^ {1} \\in \\mathbf {G} _ {1}} \\frac {- 1}{| \\mathbf {P} (i) |} \\sum_ {\\mathcal {G} _ {k} ^ {2} \\in P (i)} \\log \\frac {\\exp (\\mathcal {S} (i \\cdot k) / \\tau)}{\\sum \\mathcal {G} _ {j} ^ {2} \\in \\mathbf {G} _ {2} \\exp (\\mathcal {S} (i \\cdot j) / \\tau)}, \\tag {8}", + "image_path": "0682b3645eee8fd907e40e4a6cf639ac004eb0ce3638b2b6bd78b442abaa438c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "type": "inline_equation", + "content": "|\\mathbf{P}(i)|" + }, + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "type": "text", + "content": " denotes the cardinality and " + }, + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 133, + 287, + 180 + ], + "type": "text", + "content": " is a scalar temperature parameter. This loss pulls together the positive grasp pairs in embedding space, whereas pushes apart negative grasp pairs meanwhile." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "text", + "content": "For grasp pose refinement, we adopt a transformation loss which consists two terms for optimizing rotation and translation respectively. Consider " + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{J} = \\{j_1,j_2,\\dots ,j_n\\}" + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "text", + "content": " that stores arbitrary number of time index, for any time step " + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "inline_equation", + "content": "j\\in \\mathcal{J}" + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "text", + "content": ", given target grasp " + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^1" + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "text", + "content": ", we first compute its transformed grasp in " + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 47, + 181, + 287, + 253 + ], + "type": "text", + "content": " frame's coordinates, denoted by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 260, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 260, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 129, + 260, + 287, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {j 1} = \\left(\\mathbf {R} ^ {j 1} \\mathbf {t} ^ {j 1}\\right), \\tag {9}", + "image_path": "b103a894e775581517f747c4b1e299e43850fe5f96abd0c8b08358efbdaf812a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 283, + 217, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 217, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 217, + 295 + ], + "type": "text", + "content": "We then define the transformation loss as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 300, + 287, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 300, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 57, + 300, + 287, + 368 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {R, t} = \\frac {1}{| \\mathcal {I} |} \\sum_ {j \\in \\mathcal {J}} \\frac {1}{| \\mathbf {G} _ {j} |} \\sum_ {\\mathcal {G} _ {i} ^ {j} \\in \\mathbf {G} _ {j}} | | (\\mathbf {R} ^ {*}) ^ {- 1} \\mathbf {R} ^ {j 1} - \\mathbf {I} | | _ {2} (10) \\\\ + \\frac {1}{| \\mathcal {I} |} \\sum_ {j \\in \\mathcal {J}} \\frac {1}{| \\mathbf {G} _ {j} |} \\sum_ {\\mathcal {G} _ {i} ^ {j} \\in \\mathbf {G} _ {j}} \\| \\mathbf {t} ^ {j 1} - \\mathbf {t} ^ {*} \\| _ {2}, (10) \\\\ \\end{array}", + "image_path": "b90cbb39dd143a2c0ae324c16ec4ef52d3ca53a256dffe6854a1942cd77a5be0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j" + }, + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "type": "text", + "content": " denotes grasp set at time step " + }, + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 376, + 218, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 399, + 127, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 127, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 127, + 411 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 418, + 105, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 105, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 105, + 430 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 436, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 287, + 591 + ], + "type": "text", + "content": "The proposed pipeline is trained on GraspNet-1Billion dataset [9] which is a large-scale real-world dataset for grasp detection. It includes 89 objects, 190 scenes and 256 camera views for each scene. We follow [8] to generate grasp correspondence labels across viewpoints on the GraspNet-1Billion dataset and refer readers to [8] for more details. In addition, we further collect 30 scenes of moving objects for testing. For each frame, we record the RGBD images and objects' 6D poses. We manually annotate 10 grasps for the first frame of each scene as targets and further project the annotated grasps according to the 6D poses for the future frames. This test set is named Moving GraspNet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 599, + 228, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 228, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 228, + 612 + ], + "type": "text", + "content": "5.2. Pre-processing and Augmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": "Before being fed into the network, the point cloud is first downsampled to 20000 points and the RGB values are normalized. We also remove points outside a pre-defined workspace. Furthermore, we adopt massive data augmentation to avoid over-fitting and enrich objects' motion. For each object in the scene, its point cloud is augmented on-the-fly by a random translation of " + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "[-0.2, 0.2]" + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": " meter in all XYZ directions and a random rotation of " + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "[-30, 30]" + }, + { + "bbox": [ + 47, + 617, + 287, + 713 + ], + "type": "text", + "content": " degrees" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "around z-axis. Moreover the scene's point cloud is rotated by " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "[-30, 30]" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " degrees around both x-axis and z-axis. Due to the space limitations, please refer to the supplementary materials for details about network implementation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 130, + 478, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 130, + 478, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 130, + 478, + 143 + ], + "type": "text", + "content": "5.3. Performance of Grasp Tracking" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 149, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 245 + ], + "type": "text", + "content": "Evaluation Metric We propose a new Multiple Grasp Pose Tracking (MGTA) metric which follows the idea of Multiple Object Tracking Accuracy (MOTA) [21] in multi-object tracking task, except that the similarity measure between two grasps is computed using Eq.(2). Details are given in supplementary materials. Apart from MGTA, we also report the mean translation error and rotation error to the tracking target of each sequence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 246, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 545, + 365 + ], + "type": "text", + "content": "We follow the same evaluation procedure on GraspNet1-Billion and Moving GraspNet. For each scene, we pick 10 reference grasps in the first frame which are later used for grasp tracking. To be specific, these reference grasps are detected by a grasp detector on GraspNet1-billion but manually labeled on Moving GraspNet. For each remaining frame, we compute the ground truth grasps using objects' 6D poses and segmentation mask of the scene. The metrics are calculated using the transformed reference grasps and predictions associated with each frame." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 384, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 491 + ], + "type": "text", + "content": "Comparing with Representative Methods We compare our methods with two other baseline methods. A heuristic method is implemented as the first baseline. Given a grasp pose from the previous frame, its nearest neighbor detected in the current camera coordinates is picked. For the second method, we adopt Bundle Track [40] which is one representative method in the field of unseen object 6D pose tracking. It fits our task setting well since it requires no objects' CAD model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 493, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 493, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 493, + 545, + 600 + ], + "type": "text", + "content": "As it is shown in Tab.1 and Tab.2, our method outperforms baseline methods on both GraspNet-1Billion and Moving GraspNet. It is found in experiments that, in many cases, grasps cannot fall on the same objects if we only pick the nearest neighbors. Also, it can be drawn from the results that tracking results are improved significantly after refinement, especially for rotation. Note that MGTA will be negative if there is a large number of false positives, false negatives, and ID switches." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 610, + 443, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 443, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 443, + 624 + ], + "type": "text", + "content": "5.4. Real Robot Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "In real robot experiments, we use Flexiv Rizon robot arm with an Intel Realsense L515 depth camera mounted on the end-effector. A Robotiq-85 parallel-jaw gripper is used with 3D-printed extended fingertips attached to it. The real robot experiments run on an NVIDIA 3090 GPU at around 10 fps. Please refer to the supplementary materials for details of the objects." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8829" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 553, + 170 + ], + "type": "table", + "html": "
MethodSeenSimilarNovel
MGTA↑n (cm)↓n (°)↓MGTA↑n (cm)↓n (°)MGTA↑n (cm)↓n (°)↓
Nearest-0.8115.295.98-0.8015.497.61-0.8316.696.90
Bundle Track [40]0.741.5215.800.721.4716.370.721.9315.41
Ours, w/o refinement-0.322.3153.24-0.302.6857.59-0.392.9358.49
Ours, M=5120.801.5810.890.801.7410.620.752.1015.03
Ours, M=10240.841.3410.150.851.529.840.781.9012.96
Ours, M=20480.851.279.820.861.359.500.791.8411.48
", + "image_path": "0d71b126b5b87bd840d8b3b4a98b1b416eeaf75b4a855098d59b2809b8bb4959.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 64, + 202, + 527, + 417 + ], + "blocks": [ + { + "bbox": [ + 64, + 202, + 527, + 417 + ], + "lines": [ + { + "bbox": [ + 64, + 202, + 527, + 417 + ], + "spans": [ + { + "bbox": [ + 64, + 202, + 527, + 417 + ], + "type": "image", + "image_path": "b40619c72de1aebe5a2ca51abed56c6053ba4ed9f828fbd8e39f513154b090d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 425, + 546, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 425, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 546, + 448 + ], + "type": "text", + "content": "Figure 4. Illustration of tracking a toy dog and toy rabbit. We can see that the grasp poses are consistent, i.e., on the head of the dog and the ear of the rabbit. Yellow bounding box represents to the view limit of the camera. No box means the camera can see the entire scene." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 455, + 291, + 544 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 546, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 546, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 546, + 199 + ], + "type": "text", + "content": "Table 1. GraspNet-1Billion evaluation results on seen, similar and novel objects respectively. " + }, + { + "bbox": [ + 46, + 178, + 546, + 199 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 178, + 546, + 199 + ], + "type": "text", + "content": " denotes the number of grasp detected,its default value is 1024." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 455, + 291, + 544 + ], + "lines": [ + { + "bbox": [ + 48, + 455, + 291, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 291, + 544 + ], + "type": "table", + "html": "
MethodMGTA↑n (cm)↓n (°)↓
Nearest-0.478.0967
Bundle Track0.377.6023.52
Ours, w/o refinement-0.436.2462.33
Ours, M=5120.475.5322.63
Ours, M=10240.524.6020.37
Ours, M=20480.544.1418.28
", + "image_path": "05d0391cb38cf3d070e1607d3c568da8df411e4a81e7212cd4b276aebd0de1ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Protocol For each frame, we receive RGBD images from the in-hand camera. We transform the point cloud to the camera coordinate system at ready state by the camera extrinsic for simplicity as well. Due to safety considerations, points outside a pre-defined workspace will be masked. After a grasp candidate generation phase, we select the grasp with the highest predicted score as a reference grasp and keep tracking it for each new frame. During the tracking phase, the robot moves to a pre-grasp pose which is 0.02" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 325, + 455, + 528, + 582 + ], + "blocks": [ + { + "bbox": [ + 46, + 551, + 287, + 584 + ], + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 584 + ], + "type": "text", + "content": "Table 2. Moving GraspNet evaluation results. All objects in Moving GraspNet are novel. " + }, + { + "bbox": [ + 46, + 551, + 287, + 584 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 551, + 287, + 584 + ], + "type": "text", + "content": " denotes the number of grasp detected, its default value is 1024." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 455, + 528, + 582 + ], + "lines": [ + { + "bbox": [ + 325, + 455, + 528, + 582 + ], + "spans": [ + { + "bbox": [ + 325, + 455, + 528, + 582 + ], + "type": "table", + "html": "
Object ID#Attempt#SuccessSuccess Rate
11080.8
21090.9
31080.8
41070.7
61080.8
71080.8
91080.8
101090.9
Total80650.8125
", + "image_path": "6860b2b023cee55a6458b9fc092d7fb46367180379eb3c1d177f75e42e18f608.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 590, + 541, + 601 + ], + "lines": [ + { + "bbox": [ + 309, + 590, + 541, + 601 + ], + "spans": [ + { + "bbox": [ + 309, + 590, + 541, + 601 + ], + "type": "text", + "content": "Table 3. Real robot experiments results on single object scenes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 715 + ], + "type": "text", + "content": "meters higher than the actual grasp pose. The robot executes the grasp once three conditions are met simultaneously: 1) the translation distance between the gripper center and the target grasp in XY-plane is lower than 0.01 meter (the XYplane is parallel to the table); 2) the translation distance between the gripper center and the target grasp in 3D space is lower than 0.04 meters; 3) the rotation distance between the gripper and the target grasp in 3D space is lower than" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8830" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 282, + 171 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 171 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 171 + ], + "type": "table", + "html": "
Object ID#Attempt#SuccessSuccess Rate
7,7,8,81080.8
10,13,141090.9
1,11,15,16,17,181080.8
2,3,51090.9
9,121070.7
20,201080.8
Total60490.8167
", + "image_path": "cbd3254b9f312b3045a5ed6916a8bc66e96fcde4a064eeb0b84449f41e2fb68c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 205, + 282, + 402 + ], + "blocks": [ + { + "bbox": [ + 53, + 205, + 282, + 402 + ], + "lines": [ + { + "bbox": [ + 53, + 205, + 282, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 282, + 402 + ], + "type": "image", + "image_path": "b73592b9ec5cfb3a4256d3b6323b9d9bc6d943b52b8c4b8417283762fa02a948.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 414, + 287, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 448 + ], + "type": "text", + "content": "Figure 5. Correspondence switching between similar objects. Red arrow indicates correct correspondence and blue arrow indicates wrong correspondence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 457, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 287, + 482 + ], + "type": "text", + "content": "15 degrees. The initial positions of the objects are chosen randomly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "type": "text", + "content": "Results As shown in Tab.3 and Tab.4, we execute 20 grasp attempts for each scene. In single-object scenes, we randomly sample 8 objects for testing and we achieve a success rate of 81 percent. In multi-object scenes, we only choose objects with similar visual and geometric patterns to test the temporal smoothness and semantic consistency of the predicted grasps. Results suggest that single-object scenes and multi-object scenes have nearly the same success rate which further proves the effectiveness of our method. We illustrate one case from real robot experiments in Fig.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 111, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 111, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 111, + 635 + ], + "type": "text", + "content": "5.5. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Effect of Global Features In Sec.4.2, we incorporate global features to represent grasp features. While local grasp features are sufficient to identify corresponding local structures, grasp trials may fail due to correspondence switches among objects with similar visual and geometric patterns as we can see in Fig.5. Such phenomenon may be" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 543, + 133 + ], + "blocks": [ + { + "bbox": [ + 51, + 179, + 282, + 192 + ], + "lines": [ + { + "bbox": [ + 51, + 179, + 282, + 192 + ], + "spans": [ + { + "bbox": [ + 51, + 179, + 282, + 192 + ], + "type": "text", + "content": "Table 4. Real robot experiments results on multi-object scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 543, + 133 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 543, + 133 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 543, + 133 + ], + "type": "table", + "html": "
Object IDMethod#Attempt#SW
7,7,8,8w/o global features105
Full model101
1,17,17w/o global features104
Full model100
", + "image_path": "843dfe2383abc7bbffb81c12e96f1d0bc240e6d6729927b3eb683172caea49ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 141, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 305, + 141, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 141, + 545, + 164 + ], + "type": "text", + "content": "Table 5. Statistics of Correspondence Switching. SW represents correspondence switching." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 175, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 282 + ], + "type": "text", + "content": "due to lacking awareness of the scene state and we propose to solve it by global features as what we state in Sec.4.2. In order to analyze the actual effectiveness of global features, we further record how frequent correspondence switching happens with or without global features. We conduct these experiments on identical objects which are placed close to each other in the scenes. Tab.5 suggests that the number of correspondence switches decreased significantly with global features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "text", + "content": "Failure Analysis There are three major types of failure cases found in real robot experiments. First, some predicted grasp candidates from the detector do not have good qualities. Naturally, tracking such targets may lead to failed grasps. Second, it is hard to heuristically plan the timing to close the gripper. Due to some grasps may fall on the edges of objects, it is common that the objects may slip away immediately when the robot receives the command to close the gripper. Furthermore, since we mount the camera on the wrist, the point cloud may become fragmented due to view limitation and occlusion, especially when the objects fall behind the gripper. In such cases, the predictions can be inaccurate." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 568 + ], + "type": "text", + "content": "We present a novel solution to reactive grasping in a target-referenced setting. For a reference grasp, our method predicts temporally smooth and semantically consistent grasp poses in the future frames without any prior knowledge of instance. Experiments conducted on both single-object and multi-object scenes show that our method provides reliable grasp plans under various environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 581, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 635 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the National Key R&D Program of China (No.2021ZD0110700), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), Shanghai Qi Zhi Institute, and Shanghai Science and Technology Commission (21511101200)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 636, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 546, + 713 + ], + "type": "text", + "content": "Contributions: J. Liu and H-S. Fang initiated the project. J. Liu designed and trained the network and conducted the evaluation together with R. Zhang. J. Liu wrote the paper. H-S. Fang devised and mentored this project and edited the paper. M. Gou, H. Fang, and S. Xu constructed the testing dataset. C. Wang and H. Yan helped with the evaluation. C. Lu supervised the project and provided hardware support." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8831" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Iretiayo Akinola, Jingxi Xu, Shuran Song, and Peter K. Allen. Dynamic grasping with reachability and motion awareness. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9422-9429, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "text", + "content": "[2] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "text", + "content": "[3] Christopher B Choy, JunYoung Gwak, Silvio Savarese, and Manmohan Chandraker. Universal correspondence network. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[4] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "text", + "content": "[5] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Toward geometric deep slam. ArXiv, abs/1707.07410, 2017.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "type": "text", + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 337-33712, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 423 + ], + "type": "text", + "content": "[7] Y. Eldar, M. Lindenbaum, M. Porat, and Y.Y. Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "type": "text", + "content": "[8] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. arXiv preprint arXiv:2212.08333, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "text", + "content": "[9] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[10] Minghao Gou, Hao-Shu Fang, Zhanda Zhu, Sheng Xu, Chenxi Wang, and Cewu Lu. Rgb matters: Learning 7-dof grasp poses on monocular rgbd images. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13459-13466. IEEE, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 623 + ], + "type": "text", + "content": "[11] Yinlin Hu, Joachim Hugonot, Pascal Fua, and Mathieu Salzmann. Segmentation-driven 6d object pose estimation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "text", + "content": "[12] Seungsu Kim, Ashwini Shukla, and Aude Billard. Catching objects in flight. IEEE Transactions on Robotics, 30(5):1049-1065, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] Donghoon Lee, Onur C. Hamsici, Steven Feng, Prachee Sharma, and Thorsten Gernoth. Deeppro: Deep partial point cloud registration of objects. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5663-5672, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International journal of robotics research, 37(4-5):421-436, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[15] Hongzhuo Liang, Xiaojian Ma, Shuang Li, Michael Görner, Song Tang, Bin Fang, Fuchun Sun, and Jianwei Zhang. Pointnetgpd: Detecting grasp configurations from point sets. In 2019 International Conference on Robotics and Automation (ICRA), pages 3629-3635. IEEE, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 228 + ], + "type": "text", + "content": "[16] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2011. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "type": "text", + "content": "[17] David Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91–, 11 2004. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[18] Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley, and Ken Goldberg. Learning ambidextrous robot grasping policies. Science Robotics, 4(26), 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "text", + "content": "[19] Naresh Marturi, Marek Kopicki, Alireza Rastegarpanah, Vijaykumar Rajasekaran, Maxime Adjigble, Rustam Stolkin, Aleš Leonardis, and Yasemin Bekiroglu. Dynamic grasp and trajectory planning for moving objects. Autonomous Robots, 43(5):1241-1256, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "text", + "content": "[20] Arjun Menon, Benjamin Cohen, and Maxim Likhachev. Motion planning for smooth pickup of moving objects. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 453-460. IEEE, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 442 + ], + "type": "text", + "content": "[21] Anton Milan, Laura Leal-Taixe, Ian D. Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. ArXiv, abs/1603.00831, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "text", + "content": "[22] Douglas Morrison, Juxi Leitner, and Peter Corke. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. In Robotics: Science and Systems, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 533 + ], + "type": "text", + "content": "[23] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2901-2910, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 535, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 589 + ], + "type": "text", + "content": "[24] Peiyuan Ni, Wenguang Zhang, Xiaoxiao Zhu, and Qixin Cao. Pointnet++ grasping: Learning an end-to-end spatial grasp generation algorithm from sparse point clouds. 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3619-3625, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "text", + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 714 + ], + "type": "text", + "content": "[27] Yuzhe Qin, Rui Chen, Hao Zhu, Meng Song, Jing Xu, and Hao Su. S4g: Amodal single-view single-shot se (3) grasp" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8832" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "detection in cluttered scenes. In Conference on robot learning, pages 53-65. PMLR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[28] Patrick Rosenberger, Akansel Cosgun, Rhys Newbury, Jun Kwan, Valerio Ortenzi, Peter Corke, and Manfred Grafinger. Object-independent human-to-robot handovers using real time robotic vision. IEEE Robotics and Automation Letters, 6(1):17-23, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[29] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564-2571, 2011. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 249 + ], + "type": "text", + "content": "[30] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 293 + ], + "type": "text", + "content": "[31] Tanner Schmidt, Richard Newcombe, and Dieter Fox. Self-supervised visual descriptor learning for dense correspondence. IEEE Robotics and Automation Letters, 2(2):420-427, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 338 + ], + "type": "text", + "content": "[32] Shuran Song, Andy Zeng, Johnny Lee, and Thomas Funkhouser. Grasping in the wild: Learning 6dof closed-loop grasping from low-cost demonstrations. IEEE Robotics and Automation Letters, 5(3):4978-4985, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[33] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[34] Andreas ten Pas, Marcus Gualtieri, Kate Saenko, and Robert Platt. Grasp pose detection in point clouds. The International Journal of Robotics Research (IJRR), 36(13-14):1455-1473, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 491 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "text", + "content": "[36] Chenxi Wang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Jin Gao, and Cewu Lu. Graspness discovery in clutters for fast and accurate grasp detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15964-15973, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[37] Chen Wang, Roberto Martín-Martín, Danfei Xu, Jun Lv, Cewu Lu, Li Fei-Fei, Silvio Savarese, and Yuke Zhu. 6-pack: Category-level 6d pose tracker with anchor-based keypoints. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10059–10066, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 647 + ], + "type": "text", + "content": "[38] Lirui Wang, Yu Xiang, Wei Yang, Arsalan Mousavian, and Dieter Fox. Goal-auxiliary actor-critic for 6d robotic grasping with point clouds. In Conference on Robot Learning, pages 70-80. PMLR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[39] Yue Wang and Justin M. Solomon. Deep closest point: Learning representations for point cloud registration. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[40] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 262 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "models. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8067-8074. IEEE, 2021. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[41] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E. Bekris. se(3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[42] Wei Yang, Chris Paxton, Arsalan Mousavian, Yu-Wei Chao, Maya Cakmak, and Dieter Fox. Reactive human-to-robot handovers of arbitrary objects. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 3118–3124. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "type": "text", + "content": "[43] Yi Zhou, Connelly Barnes, Lu Jingwan, Yang Jimei, and Li Hao. On the continuity of rotation representations in neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8833" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_content_list.json b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..03c708a7c204a0c49c8a30b2fdc6ddf7740e88d3 --- /dev/null +++ b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning", + "text_level": 1, + "bbox": [ + 111, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjin Wang, Yunqing Hu, Qianglong Chen, Yin Zhang* Zhejiang University, Hangzhou, China", + "bbox": [ + 259, + 203, + 712, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangwenjin,yunqinghu,chenqianglong,zhangyin98}@zju.edu.cn", + "bbox": [ + 225, + 241, + 738, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Parameter regularization or allocation methods are effective in overcoming catastrophic forgetting in lifelong learning. However, they solve all tasks in a sequence uniformly and ignore the differences in the learning difficulty of different tasks. So parameter regularization methods face significant forgetting when learning a new task very different from learned tasks, and parameter allocation methods face unnecessary parameter overhead when learning simple tasks. In this paper, we propose the Parameter Allocation & Regularization (PAR), which adaptively select an appropriate strategy for each task from parameter allocation and regularization based on its learning difficulty. A task is easy for a model that has learned tasks related to it and vice versa. We propose a divergence estimation method based on the Nearest-Prototype distance to measure the task relatedness using only features of the new task. Moreover, we propose a time-efficient relatedness-aware sampling-based architecture search strategy to reduce the parameter overhead for allocation. Experimental results on multiple benchmarks demonstrate that, compared with SOTAs, our method is scalable and significantly reduces the model's redundancy while improving the model's performance. Further qualitative analysis indicates that PAR obtains reasonable task-relatedness.", + "bbox": [ + 73, + 323, + 473, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the lifelong learning [9] ability of neural networks, i.e., learning continuously from a continuous sequence of tasks, has been extensively studied. It is natural for human beings to constantly learn and accumulate knowledge from tasks and then use it to facilitate future learning. However, classical models [13, 18, 41] suffer catastrophic forgetting [12], i.e., the model's performance on learned tasks deteriorates rapidly after learning a new one.", + "bbox": [ + 73, + 739, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome the catastrophic forgetting, many param-", + "bbox": [ + 94, + 862, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f152f92326c8d478cc0b5bd471900e9e6c642ee2c3c219b8946dc471d4c411e.jpg", + "image_caption": [ + "Figure 1. PAR adaptively selects a strategy from regularization and allocation to handle each task according to its learning difficulty. The difficulty depends not only on the task itself, but also on whether the model has previously learned tasks related to it." + ], + "image_footnote": [], + "bbox": [ + 506, + 292, + 895, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eter regularization or allocation methods have been proposed. Parameter regularization methods [10, 16, 19, 21, 23, 25, 27] alleviate forgetting by adding a regularization term to the loss function and perform well when the new task does not differ much from learned tasks. Parameter allocation methods based on static models [7, 15, 31, 36] and dynamic models [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] allocate different parameters to different tasks and can adapt to new tasks quite different from learned tasks. However, the above methods solve all tasks in a sequence uniformly, and ignore the differences of learning difficulty of different tasks. This leads to the significant forgetting in parameter regularization methods when learning a new task which is quite different from learned tasks, and also leads to unnecessary parameter cost in parameter allocation methods when learning some simple tasks.", + "bbox": [ + 496, + 625, + 893, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a difficulty-aware method Parameter Allocation & Regularization (PAR). As shown", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Yin Zhang.", + "bbox": [ + 94, + 886, + 285, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7776", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in Fig. 1, we assume that the learning difficulty of a task in continual learning depends not only on the task itself, but also on the accumulated knowledge in the model. A new task is easy to adapt for a model if it has learned related tasks before and vice versa. Based on the assumption, the PAR adaptively adopts parameter allocation for difficult tasks and parameter regularization for easy tasks. Specifically, the PAR divides tasks into task groups and assigns each group a dedicated expert model. Given a new task, the PAR measures the relatedness between it and existing task groups at first. If the new task is related to one of the existing groups, it is easy for the corresponding expert. The PAR adds the task to the related group and learns it by the expert via the parameter regularization. Otherwise, the new task is difficult for all existing experts, and the PAR assigns it to a new task group and allocates a new expert to learn it.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two challenges in this work: the measurement of relatedness and the parameter explosion associated with parameter allocation. For the first one, we try to measure the relatedness by the KL divergence between feature distributions of tasks. However, the KL divergence is intractable and needs to be estimated since the feature distributions of tasks are usually unknown. In addition, the constraint of lifelong learning that only data of the current task are available exacerbates the difficulty of estimation. To solve above problems, inspired by the divergence estimation based on k-NN distance [44], we propose the divergence estimation method based on prototype distance, which only depends on the data of the current task. For the second one, we try to reduce parameter overhead per expert by searching compact architecture for it. However, the low time and memory efficiency is an obstacle to applying architecture search for a sequence of tasks in lifelong learning. To improve the efficiency of architecture search, we propose a relatedness-aware sampling-based hierarchical search. The main contributions of this work are as follows:", + "bbox": [ + 75, + 332, + 472, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a lifelong learning framework named Parameter Allocation & Regularization (PAR), which selects an appropriate strategy from parameter allocation and regularization for each task based on the learning difficulty. The difficulty depends on whether the model has learned related tasks before.", + "- We propose a divergence estimation method based on prototype distance to measure the distance between the new task and previous learned tasks with only data of the new task. Meanwhile, we propose a relatedness-aware sampling-based architecture search to reduce the parameter overhead of parameter allocation.", + "- Experimental results on CTrL, Mixed CIFAR100 and F-CelebA, CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20 demonstrate that PAR is scalable and significantly reduces the model redundancy" + ], + "bbox": [ + 96, + 641, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "while improving the model performance. Exhaustive ablation studies show the effectiveness of components in PAR and the visualizations show the reasonability of task distance in PAR.", + "bbox": [ + 531, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 166, + 640, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lifelong Learning", + "text_level": 1, + "bbox": [ + 500, + 191, + 674, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many methods have been proposed to overcome catastrophic forgetting. Replay methods try to replay samples of previous tasks when learning a new task from an episodic memory [32,40] or a generative memory [3,28,37]. Parameter regularization methods, including the prior-focused regularization [16, 19, 23, 27] and the data-focused regularization [10, 21, 25], try to alleviate forgetting by introducing a regularization term in the loss function of the new task. Parameter allocation methods based on the static model [7, 15, 31, 36] and the dynamic model [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] overcome catastrophic forgetting by allocating different parameters to different tasks.", + "bbox": [ + 496, + 215, + 890, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods with relatedness. Several methods also consider the utility of task relatedness [2, 15]. Expert Gate [2] assigns dedicated experts and auto-encoders for tasks and calculates the task relatedness by reconstruction error of auto-encoders. The task relatedness is used to transfer knowledge from the most related previous task. CAT [15] defines the task similarity by the positive knowledge transfer. It focuses on selectively transferring the knowledge from similar previous tasks and dealing with forgetting between dissimilar tasks by hard attention. In this paper, we propose a divergence estimation method based on prototype distance to calculate the task relatedness used to measure the task learning difficulty.", + "bbox": [ + 496, + 397, + 890, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods with NAS. Static model based methods [7, 31] try to search a sub-model for each task with neural architecture search (NAS). Dynamic model based methods [20, 42, 46] adopt NAS to select an appropriate model expansion strategy for each task and face the high GPU memory, parameter, and time overhead. Instead, in this paper, we propose a relatedness-aware hierarchical cell-based architecture search to search compact architecture for each expert with a lower GPU memory and time overhead. [8, 14] adopt NAS on multi-task learning.", + "bbox": [ + 496, + 594, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Cell-based NAS", + "text_level": 1, + "bbox": [ + 500, + 756, + 658, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural architecture search (NAS) [11, 33] aims to search for efficient neural network architectures from a pre-defined search space in a data-driven manner. To reduce the size of the search space, cell-based NAS methods [22, 48, 49] try to search for a cell architecture from a pre-defined cell search space, where a cell is a tiny convolutional network mapping an $H \\times W \\times F$ tensor to another $H' \\times W' \\times F'$ . The final model consists of a pre-defined number of stacked", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7777", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cells. The cell in NAS is similar to the residual block in the residual network (ResNet), but its architecture is more complex and is a directed acyclic graph (DAG). The operations in the search space are usually parameter-efficient. NAS methods usually produce more compact architectures than hand-crafted designs.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 196, + 169, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on the task-incremental scenario of lifelong learning. Specifically, the model learns a sequence of tasks $\\mathcal{T} = \\{T_1,\\dots ,T_t,\\dots ,T_N\\}$ one by one and the task id of sample is available during both training and inference. Each task $T_{t}$ has a training dataset, $D_{train}^{t} = \\{(x_{i}^{t},y_{i}^{t});i = 1,\\ldots ,n_{train}^{t}\\}$ , where $y_{i}^{t}$ is the true label and $n_{train}^{t}$ is the number of training examples. Similarly, we denote the validation and test dataset of task $T_{t}$ as $D_{valid}^{t}$ and $D_{test}^{t}$ .", + "bbox": [ + 75, + 220, + 468, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the PAR divides previously learned tasks into task groups $\\mathcal{G}$ and assigns each group $\\mathcal{G}_i$ a dedicated expert model $E_i$ . For a new task, PAR calculates the distances, which reflect the relatedness, between it and existing task groups at first. Then, it adopts an appropriate learning strategy from parameter allocation and regularization for the new task based on the distances.", + "bbox": [ + 75, + 344, + 470, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Distance via Nearest-Prototype Distance", + "text_level": 1, + "bbox": [ + 76, + 459, + 464, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The PAR calculates the distance between the new task and each existing task at first, then it averages the distances of tasks in the same group as the distance between the new task and the group.", + "bbox": [ + 75, + 482, + 468, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We calculate the distance between two tasks by the KL divergence between their distributions. However, the KL divergence is intractable because the true distributions of tasks are usually agnostic. Instead, we estimate the KL divergence by the features of data. To enhance estimation stability, we introduce an extra pre-trained extractor1 to generate a robust feature $X_{i}^{t}$ for each image $x_{i}^{t}$ in task $T_{t}$ . The pre-trained extractor is only used for divergence estimation and does not affect the learning of tasks. The number of extra parameters introduced by the model is a constant and does not affect the scalability of our method.", + "bbox": [ + 75, + 544, + 470, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$k$ -Nearest Neighbor Distance Given the features of distributions, a classical divergence estimation method for multidimensional densities is the $k$ -Nearest Neighbor ( $k$ -NN) distance [44]. Suppose the feature distributions of the new task $T_{i}$ and an existing task $T_{j}$ are $q_{i}$ and $q_{j}$ respectively, and the features $X^{i} = \\{X_{1}^{i},\\ldots ,X_{n}^{i}\\}$ and $X^{j} = \\{X_{1}^{j},\\ldots ,X_{m}^{j}\\}$ are drawn independently from distributions $q_{i}$ and $q_{j}$ , where $n$ and $m$ are the numbers of samples. The", + "bbox": [ + 76, + 729, + 468, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$k$ -NN distance of feature $X_{l}^{i}$ in $X^{i}$ and $X^{j}$ are denoted by $\\nu_{k}(l)$ and $\\rho_{k}(l)$ respectively. Specifically, the $\\rho_{k}(l)$ is the Euclidean distance from $X_{l}^{i}$ to its $k$ -NN in $\\{X_{z}^{i}\\}_{z\\neq l}$ . Similarly, the $\\nu_{k}(l)$ is the distance from $X_{l}^{i}$ to its $k$ -NN in $X^{j}$ . Then, the KL divergence between distributions $q_{i}$ and $q_{j}$ are estimated as follows:", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK L \\left(q _ {i} \\mid \\mid q _ {j}\\right) \\approx \\widehat {K L} \\left(q _ {i} \\mid \\mid q _ {j}\\right) = \\frac {d}{n} \\sum_ {l = 1} ^ {n} \\log \\frac {\\nu_ {k} (l)}{\\rho_ {k} (l)} + \\log \\frac {m}{n - 1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 186, + 890, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is the feature dimension.", + "bbox": [ + 500, + 239, + 720, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The $k$ -NN distance is asymmetric and this is consistent with the asymmetry of the KL divergence. The motivation behind Eq. (1) is that the $k$ -NN distance is the radius of a $d$ -dimensional open Euclidean ball used for $k$ -NN density estimation. Readers can refer to [44] for more details such as convergence analysis.", + "bbox": [ + 496, + 255, + 890, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nearest-Prototype Distance However, the calculation of the $k$ -NN distance involves features of two tasks $T_{i}$ and $T_{j}$ and violates the constraint of lifelong learning that only data of the current task $T_{i}$ are available. To overcome this problem, we propose the Nearest-Prototype distance to replace the $k$ -NN distance. Specifically, suppose the set of classes in task $T^{i}$ is $C^i$ , for each class $c \\in C^i$ , we maintain a prototype feature $U_{c}^{i}$ , which is the mean of features of samples belonging to the class $c$ . Then, the Nearest-Prototype distance of $X_{l}^{i}$ to $X^{i}$ is defined as follows:", + "bbox": [ + 496, + 363, + 890, + 515 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (l) = \\left\\| X _ {l} ^ {i} - U _ {c (l)} ^ {i} \\right\\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 522, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $||\\cdot ||$ is the Euclidean distance and $c(l)$ is the class of $X_{l}^{i}$ . Similarly, we denote the set of classes in task $T_{j}$ by $C^j$ and the prototype features of $X^{j}$ for each class $c\\in C^j$ by $U_{c}^{j}$ . Then, the Nearest-Prototype distance of $X_{l}^{i}$ to $X^{j}$ is denoted as", + "bbox": [ + 496, + 549, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nu (l) = \\min _ {c} \\| X _ {l} ^ {i} - U _ {c} ^ {j} \\| _ {2}, c \\in C ^ {j}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 633, + 890, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, by replacing the $k$ -NN distances $\\nu_{k}(l)$ and $\\rho_{k}(l)$ in Eq. (1) by the Nearest-Prototype distances $\\nu(l)$ and $\\rho(l)$ , the KL divergence estimated by the Nearest-Prototype distance is as follows:", + "bbox": [ + 496, + 664, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK L \\left(q _ {i} \\mid \\mid q _ {j}\\right) \\approx \\widetilde {K L} \\left(q _ {i} \\mid \\mid q _ {j}\\right) = \\sum_ {c \\in C ^ {i}} \\frac {1}{n _ {c}} \\sum_ {l = 1} ^ {n _ {c}} \\log \\frac {\\nu (l)}{\\rho (l)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 729, + 890, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where constant terms are omitted because we are only concerned with the relative magnitude of KL divergence. The Eqs. (2) to (4) only involve features of the new task $T_{i}$ and prototypes of the existing task $T_{j}$ , which satisfies the constraint of lifelong learning. The storage cost of prototypes is small and negligible.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The motivation behind Eq. (4) is intuitive. The Nearest-Prototype distance $\\rho(l)$ reflects the relationship between the", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We adopt a ResNet18 pre-trained on the ImageNet. It does not introduce unfairness because the generated features are not involved in model training and prediction.", + "bbox": [ + 75, + 862, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7778", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1affc81cb172d184e085eb37dc70dfa95b6a6392ebb0dccf79b88d75139e4776.jpg", + "image_caption": [ + "Figure 2. The architecture of PAR. PAR divides previously learned tasks into task groups and each group has a dedicated expert model. Given a new task $T_{5}$ , PAR calculates the distance between it and existing task groups to measure its learning difficulty. Then, PAR learns the new task by parameter regularization if it is easy to learn. Otherwise, PAR learns the new task by parameter allocation and search compact architecture for each expert by a relatedness-aware hierarchical architecture search." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sample $X_{l}^{i}$ in task $T_{i}$ and its class prototype in feature space. The Nearest-Prototype distance $\\nu (l)$ reflects the relationship between the sample $X_{l}^{i}$ and class prototypes of existing task $T_{j}$ in feature space. If the value of $\\rho (l)$ and $\\nu (l)$ are equal for each $X_{l}^{i}$ , the class prototypes of two tasks are close. In this case, the distribution of the two tasks are similar, and the estimated KL divergence by Eq. (4) is 0.", + "bbox": [ + 75, + 445, + 472, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Learning Strategy Given the distance between the new task $T_{i}$ and each existing tasks by KL divergence, we denote the distance $s_{i,g}^{\\prime}$ between the task $T_{i}$ and a task group $\\mathcal{G}_g$ by the average of distances of tasks in the group as follows:", + "bbox": [ + 76, + 569, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, g} ^ {\\prime} = \\frac {1}{| \\mathcal {G} _ {g} |} \\sum_ {j ^ {\\prime}} ^ {| \\mathcal {G} _ {g} |} K L (q _ {i} | | q _ {j ^ {\\prime}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 648, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{i}$ and $q_{j'}$ represent feature distributions of task $T_{i}$ and the $j'$ -th task in group $\\mathcal{G}_g$ . The $s_{i,g}' \\in [0, \\infty)$ since the range of the KL divergence is $[0, \\infty)$ . However, we try to use the distance to reflect the relatedness, which is usually measured by a value between 0 and 1. So we map the $s_{i,g}'$ into $[0, 1]$ by a monotone increasing function as follows:", + "bbox": [ + 75, + 700, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, g} = \\min \\left(s _ {i, g} ^ {\\prime}, 1 - e ^ {- 2 \\times s _ {i, g} ^ {\\prime}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 797, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e$ is the Euler's number.", + "bbox": [ + 76, + 825, + 279, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose the smallest distance between the new task $T_{t}$ and existing groups is $s_{t,g^{*}}$ and the corresponding task group is $\\mathcal{G}_{g^{*}}$ , the PAR selects the learning strategy by comparing the $s_{t,g^{*}}$ with a hyper-parameter $\\alpha$ . If $s_{t,g^{*}} \\leq \\alpha$ ,", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the new task is added to this related group $\\mathcal{G}_{g^*}$ and learned by parameter regularization. Otherwise, the new task is assigned to a new group and learned by parameter allocation because no existing task group is related to it.", + "bbox": [ + 496, + 445, + 892, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Parameter Regularization", + "text_level": 1, + "bbox": [ + 498, + 513, + 736, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A new task $T_{t}$ is easy for the expert model $E_{g^{*}}$ if it is related to the group $G_{g^{*}}$ . PAR reuses the expert $E_{g^{*}}$ to learn this task by parameter regularization. Inspired by LwF [21], we adopt a parameter regularization method based on knowledge distillation. Specifically, the loss function consists of a training loss $\\mathcal{L}_{\\mathrm{new}}$ and a distillation loss $\\mathcal{L}_{\\mathrm{old}}$ . The training loss encourages expert $E_{g^{*}}$ to adapt to the new task $T_{t}$ and is the cross-entropy for classification as follows:", + "bbox": [ + 496, + 537, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n e w}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\log \\left(p _ {E _ {g ^ {*}}} (y | x)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 667, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The distillation loss encourages the expert $E_{g^*}$ to maintain performance on previous tasks in the group $G_{g^*}$ . To calculate it, we record the logits $\\mathbf{y}^j$ of the output head of each previous task $T_j$ for each sample $x$ of task $T_t$ . The distillation loss is as follows:", + "bbox": [ + 496, + 710, + 890, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o l d}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\sum_ {T _ {j} \\in \\mathcal {G} _ {g ^ {*}}} \\mathbf {y} ^ {j} \\cdot \\log \\left(\\mathbf {p} _ {\\mathbf {E} _ {\\mathbf {g} ^ {*}}} (\\mathbf {x}) ^ {j}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 792, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{y}^j$ and $\\mathbf{p}_{\\mathbf{E}_{\\mathbf{g}^*}}(\\mathbf{x})^j$ are vectors with lengths equal to the number of categories of the previous task $T_{j}$ . The total loss is", + "bbox": [ + 496, + 839, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P R}} = \\mathcal {L} _ {\\text {n e w}} + \\lambda \\mathcal {L} _ {\\text {o l d}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7779", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/69815aea38a7996d96be27346c1bd4f37dd163962951631a80949f1635df6472.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodS-S+SinSoutSplSlong
APAFAPAFAPAFAPAFAPAFAPAF
Finetune0.180-0.3300.240-0.2300.180-0.3100.150-0.3700.210-0.3000.200-0.400
(1)EWC0.540-0.0100.370-0.0400.430-0.1200.510-0.0300.300-0.1700.270-0.300
(2)ER0.410-0.1300.430-0.0700.380-0.1700.540-0.0700.450-0.080--
(3)HAT*0.570-0.0100.5700.0000.580-0.0100.6000.0000.5800.0000.240-0.100
(4)Independent0.5600.0000.5700.0000.5700.0000.6100.0000.5900.0000.5700.000
PNN0.6200.0000.5200.0000.5700.0000.6200.0000.5400.000--
SG-F0.6360.0000.6150.0000.6550.0000.6410.0000.6200.000--
MNTDP-S0.6300.0000.5600.0000.5700.0000.6400.0000.5500.0000.6800.000
MNTDP-D0.6700.0000.6100.0000.6000.0000.6800.0000.6200.0000.7500.000
LMC0.6660.0000.601-0.0140.6950.0000.667-0.0100.616-0.0350.639-
PAR(our)0.7060.0000.670-0.0170.6990.0000.6940.0000.7000.0000.773-0.016
", + "bbox": [ + 98, + 88, + 874, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The average performance (AP) and average forgetting (AF) on the six streams $(S^{-}, S^{+}, S^{\\mathrm{in}}, S^{\\mathrm{out}}, S^{\\mathrm{pl}}, S^{\\mathrm{long}})$ in the CTRL benchmark. We compare the PAR with four types of baselines: (1) parameter regularization, (2) replay, (3) parameter allocation with static model, and (4) parameter allocation with dynamic model. The * corresponds to model using the AlexNet backbone.", + "bbox": [ + 75, + 325, + 893, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter to balance training and distillation loss. Note that our method is without memory and does not require storing samples for previous tasks.", + "bbox": [ + 75, + 396, + 470, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the expert $E_{g^{*}}$ may over-fit the new task whose sample size is far less than tasks in the group $\\mathcal{G}_{g^{*}}$ , even though the new task is related to the group. This leads to interference with previously accumulated knowledge in the expert $E_{g^{*}}$ . To avoid the above problem, PAR records the maximum sample size of tasks in each task group. Suppose the maximum sample size in the group $\\mathcal{G}_{g^{*}}$ is $Q$ , PAR freezes the parameters of expert $E_{g^{*}}$ except for the task-specific classification head during the learning if the sample size of the new task is less than 10 percent of $Q$ . By transferring the existing knowledge in expert $E_{g^{*}}$ , only optimizing the classification header is sufficient to adapt to the new task because the new task is related to the group $\\mathcal{G}_{g^{*}}$ .", + "bbox": [ + 75, + 441, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Parameter Allocation", + "text_level": 1, + "bbox": [ + 76, + 650, + 277, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If no existing groups are related to the new task $T_{t}$ , PAR assigns it to a new task group $\\mathcal{G}_{M + 1}$ with a new expert $E_{M + 1}$ , where $M$ is the number of existing groups. We adopt the cross-entropy loss for the task $T_{t}$ as follows:", + "bbox": [ + 76, + 672, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P A}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\log \\left(p _ {E _ {M + 1}} (y | x)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 746, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The number of experts and parameters in PAR is proportional to the number of task groups, mitigating the growth of parameters. To further reduce the parameter overhead of each expert, we adopt cell-based NAS (refer to Sec. 2.2 for details) to search for a compact architecture for it.", + "bbox": [ + 75, + 792, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each expert in PAR is stacked with multiple cells and the search for expert architecture is equivalent to the search", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for the cell architecture. Since the time overhead of NAS becomes unbearable as the number of tasks increases, we propose a relatedness-aware sampling-based architecture search strategy to improve efficiency.", + "bbox": [ + 496, + 396, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we construct a hierarchical search space. The coarse-grained search space contains cells used by existing experts and an unknown cell whose architecture will come from the fine-grained search space. Following common practice [22, 48], the fine-grained search space is a directed acyclic graph (DAG) with seven nodes (two input nodes $i_1$ , $i_2$ , an ordered sequence of intermediate nodes $n_1, n_2, n_3, n_4$ , and an output node). The input nodes represent the outputs of the previous two layers. The output is concatenated from all intermediate nodes. Each intermediate node has a directed candidate edge from each of its predecessors. Each candidate edge is associated with six parameter-efficient candidate operations.", + "bbox": [ + 496, + 458, + 892, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To search a cell for the new task, we introduce a hyperparameter $\\beta$ . If $s_{t,g^*} \\leq \\beta$ , PAR reuses the cell of expert $E_{g^*}$ for the task $T_t$ . A task distance greater than $\\alpha$ and less than $\\beta$ indicates that the new task is not related enough to share the same expert with tasks in the group $\\mathcal{G}_{g^*}$ but can use the same architecture. If $s_{t,g^*} > \\beta$ , PAR assigns the unknown cell to the new expert and adopts a sampling-based NAS method named MDL [48] to determine its architecture. Due to limited space, we leave the details of the MDL in the supplementary.", + "bbox": [ + 496, + 654, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 820, + 633, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 845, + 705, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks We adopt two benchmarks the CTrL [42] and the Mixed CIFAR100 and F-CelebA [15], which con", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7780", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ede08e4867f0d1ec8a14716198bd1007da5ea190a6ee60b388d7f9dab7f1cd02.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFinetuneHATCATPAR
AP0.61550.61780.68310.7122
", + "bbox": [ + 101, + 88, + 446, + 135 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/697856fa810b821e9133ab7bb062dfd3875e4723f60f7060f44c46ff475bf66b.jpg", + "table_caption": [ + "Table 2. We compare the average performance (AP) of PAR with two parameter allocation methods based on static model on the mixed CIFAR100 and F-CelebA benchmark." + ], + "table_footnote": [], + "table_body": "
#MethodCIFAR100-10MiniImageNet-20
AP(%)AF(%)AP(%)AF(%)
Finetune0.280-0.5600.378-0.314
1EWC0.608-0.0610.574-0.144
MAS0.579-0.0500.368-0.238
LwF0.655-0.0880.589-0.169
2GPM*0.725-0.604-
A-GEM0.677-0.1200.524-0.152
3Independent0.8200.0000.7870.000
PN0.8210.0000.7810.000
Learn to Grow0.7910.000--
MNTDP-S0.750-0.000--
MNTDP-D0.830-0.000--
PAR0.853-0.0200.816-0.022
", + "bbox": [ + 81, + 202, + 465, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tain mixed similar and dissimilar tasks. CTrL [42] includes 6 streams of image classification tasks: $S^{-}$ is used to evaluate the ability of direct transfer, $S^{+}$ is used to evaluate the ability of knowledge update, $S^{\\mathrm{in}}$ and $S^{\\mathrm{out}}$ are used to evaluate the transfer to similar input and output distributions respectively, $S^{\\mathrm{pl}}$ is used to evaluate the plasticity, $S^{\\mathrm{long}}$ consists of 100 tasks and is used to evaluate the scalability. Similarly, Mixed CIFAR100 and F-CelebA [15] including mixed 10 similar tasks from F-CelebA and 10 dissimilar tasks from CIFAR100 [17].", + "bbox": [ + 75, + 531, + 467, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further, we adopt classical task incremental learning benchmarks: CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20. CIFAR10-5 is constructed by dividing CIFAR10 [17] into 5 tasks and each task has 2 classes. Similarly, CIFAR100-10 and CIFAR100-20 are constructed by dividing CIFAR100 [17] into 10 10-classification tasks and 20 5-classification tasks respectively. MiniImageNet-20 is constructed by dividing MiniImageNet [43] into 20 tasks and each task has 5 classes. We leave the results on the CIFAR10-5 and CIFAR100-20 in the supplementary.", + "bbox": [ + 75, + 683, + 467, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines Firstly, we compare our method with two simple baselines Finetune and Independent. While Finetune learns tasks one-by-one without any constraints, Indepen", + "bbox": [ + 76, + 854, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f1dd6b7c7fa74567790f869eb15b5db707bb73f594ae3a776db77e0e3f615ff5.jpg", + "table_caption": [ + "Table 3. PAR outperforms (1) parameter regularization methods, (2) replay methods, and (3) parameter allocation methods with dynamic model on classical benchmarks. The * corresponds to model using the AlexNet backbone." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR100-10 (10 tasks)Slong (100 tasks)
APM(M)APM(M)
Finetune0.1800.6070.2000.607
Independent0.8206.0700.57060.705
PN0.82112.422--
Learn to Grow0.7911.926--
MNTDP-S0.7506.1000.68039.658
MNTDP-D0.8306.0750.75025.508
PAR0.8531.4000.77313.475
", + "bbox": [ + 501, + 88, + 890, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af8c8a4be59c91c76bf8375ef852d06001d591c783c72610811ddf5ca142a711.jpg", + "table_caption": [ + "Table 4. Compared with existing parameter allocation methods based on dynamic model, PAR obtains better average performance (AP) with fewer model parameters $(\\mathcal{M},\\mathrm{M} = 1\\mathrm{e}6)$ ." + ], + "table_footnote": [], + "table_body": "
#Allocation StrategyReg.APM(M)T(h)
FixedFSCS
1✓0.8552.2361.0
2✓0.8612.2771.9
3✓✓0.8701.9001.2
4✓0.7750.2441.0
5✓✓0.8341.1291.0
6✓✓0.8561.2101.4
PAR✓✓✓0.8531.4001.1
", + "bbox": [ + 501, + 305, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Ablation studies of components in PAR on the CIFAR100-10. The Fixed represents using a fixed cell from DARTs. The FS and CS represent fine-grained and coarse-grained search space respectively. The Reg. represents the parameter regularization. The PAR makes the trade off among average performance (AP), parameter overhead $(\\mathcal{M})$ , and time overhead (T).", + "bbox": [ + 498, + 463, + 890, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dent builds a model for each task independently. Then, we present parameter regularization methods, including EWC [16], LwF [21], and MAS [1], and replay methods, including ER [6], GPM [35], and A-GEM [5]. Meanwhile, we compare parameter allocation methods with the static model, including HAT [36] and CAT [15], and parameter allocation methods with the dynamic model, including PN [34], Learn to Grow [20], SG-F [24], MNTDP [42], LMC [29] with PAR.", + "bbox": [ + 496, + 575, + 890, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Denote the performance of the model on $T_{j}$ after learning task $T_{i}$ as $r_{i,j}$ where $j \\leq i$ . Suppose the current task is $T_{t}$ , the average performance (AP) and average forgetting (AF) are as follows:", + "bbox": [ + 496, + 739, + 890, + 800 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA P = \\frac {1}{t} \\sum_ {j = 1} ^ {t} r _ {t, j}, A F = \\frac {1}{t} \\sum_ {j = 1} ^ {t} r _ {t, j} - r _ {j, j}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 814, + 890, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the parameter overhead, we denote the total number of the model as $\\mathcal{M}$ .", + "bbox": [ + 498, + 869, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7781", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/040567f9d5e1e703148d1429295db47a15968074fcc410cc820dcee3759ecfd2.jpg", + "image_caption": [ + "(a) Average performance (AP)" + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 267, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d214bb5aaf067b6624392499f3fa2982f8bdef87d8426ec71467b7f05aa3e6e.jpg", + "image_caption": [ + "(b) Model size" + ], + "image_footnote": [], + "bbox": [ + 289, + 90, + 472, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31cef153349764ef8b39347a82832b851562e6a51052629f1366598bba3a4d68.jpg", + "image_caption": [ + "(c) Experts" + ], + "image_footnote": [], + "bbox": [ + 496, + 90, + 679, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee57c82820b029cb732756bbcdb6220014afc0890f516f0db7fffd338864bb5a.jpg", + "image_caption": [ + "(d) Time" + ], + "image_footnote": [], + "bbox": [ + 705, + 90, + 885, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9800dc49f02fcace30a9696a4513be5b30b09b0c48b0efdec577c1ba42f0fe92.jpg", + "image_caption": [ + "Figure 3. Ablation studies of $\\alpha$ (Alpha) when $\\beta = 0.75$ . The larger $\\alpha$ encourages the model to reuse experts, so the model size and number of experts decreases with the increase of $\\alpha$ . The $\\alpha$ is independent of the architecture search, so the learning time is stable relative to it.", + "(a) Average performance (AP)" + ], + "image_footnote": [], + "bbox": [ + 83, + 270, + 267, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7324a2243eb8ed2174afa8bad13bc536947a1bb43c325dd7d848d35d0f9dd386.jpg", + "image_caption": [ + "(b) Model size" + ], + "image_footnote": [], + "bbox": [ + 290, + 270, + 472, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2cb749a5a7ce995f6432259086e299f5c8577250f2d1d098aca5f27dd3f68b42.jpg", + "image_caption": [ + "(c) Cells" + ], + "image_footnote": [], + "bbox": [ + 498, + 270, + 679, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a26c713d8c8198f1bfd5bd95b3d2f94dbd27e708e517591f72c9ebc5abab1cc.jpg", + "image_caption": [ + "(d) Time" + ], + "image_footnote": [], + "bbox": [ + 705, + 270, + 887, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details We implement PAR by the PyTorch and open the source code2. We set the number of cells of each expert in PAR as 4, and set the $\\alpha$ to 0.5 and the $\\beta$ to 0.75 for all tasks. We adopt the SGD optimizer whose initial learning rate is 0.01 and anneals following a cosine schedule. We also set the momentum of SGD to 0.9 and search weight decay from [0.0003, 0.001, 0.003, 0.01] according to the validation performance. The results are averaged across three runs with different random seeds.", + "bbox": [ + 75, + 472, + 468, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without special instructions, following MNTDP [42], we adopt a light version of ResNet18 [13] with multiple output heads as the backbone for baselines. For the performance on CTrL, we report the results from MNTDP [42] and LMC [29]. For the performance on mixed CIFAR100 and F-CelebA, we report the results from CAT [15]. For the performance on classical benchmarks, we: report the results of RPSnet [31], InstAParam [7], and BNS [30] from origin papers; adopt the implementation of [4] for memory-based methods; and adopt our implementation for the others.", + "bbox": [ + 75, + 609, + 470, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with baselines", + "text_level": 1, + "bbox": [ + 76, + 773, + 320, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At first, we evaluate PAR on benchmarks with mixed similar and dissimilar tasks. As shown in Tab. 1, PAR outperforms baselines on all six streams in CTrL. One reason is that PAR allows knowledge transfer among related tasks while preventing interference among tasks not related. For", + "bbox": [ + 75, + 797, + 470, + 875 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/417121c303f687f8f5eb121e5652cbd5eeb90d98864a2c235259abb8e934696c.jpg", + "image_caption": [ + "Figure 4. Ablation studies of $\\beta$ (Beta) when $\\alpha = 0.5$ . The $\\beta$ has little effect on the average performance and size of model, and conversely has a great effect on the learning time because the determine the frequency of architecture search. The larger $\\beta$ encourages the model to reuse architectures of existing experts.", + "Figure 5. Visualization of task groups and experts on CIFAR100-Coarse. The PAR groups semantically related tasks into the same group. We distinguish the different groups by different colors." + ], + "image_footnote": [], + "bbox": [ + 503, + 474, + 890, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "example, performance on stream $S^{+}$ shows that PAR can update the knowledge of the expert for the task $t_{1}^{-}$ by the task $t_{1}^{+}$ with the same distribution and more samples. Performance on streams $S^{-}$ and $S^{\\mathrm{out}}$ shows that our method can transfer knowledge from experts of tasks related to the new task. Another reason is that a task-tailored architecture helps each expert perform better. Moreover, performance on stream $S^{\\mathrm{long}}$ shows that PAR is scalable for the long task sequence. Similarly, results in Tab. 2 show that PAR outperforms parameter allocation methods with static models.", + "bbox": [ + 496, + 731, + 890, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, we evaluate the PAR on classical benchmarks.", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2https://github.com/WenjinW/PAR", + "bbox": [ + 94, + 886, + 281, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7782", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ae64b8707c131b9cd33e82b2499d42459494137b5e9fd66f689f1b98a8cbba7.jpg", + "image_caption": [ + "(a) $S^{-}$" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 254, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d95c2b3a609831f66a61b5432002cc9402da89955360381b4c247e47de08d3a8.jpg", + "image_caption": [ + "(b) $S^{+}$" + ], + "image_footnote": [], + "bbox": [ + 295, + 90, + 455, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92e542b4c5033f189d6b7fad7eef66a87df44b27e66463e207ce458c5a3d9e75.jpg", + "image_caption": [ + "(c) $S^{\\mathrm{in}}$" + ], + "image_footnote": [], + "bbox": [ + 94, + 213, + 254, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/261de798ab628eff71a386dcf1090847872525dbb24d7ff9bd75b77f61a8a211.jpg", + "image_caption": [ + "(d) $S^{\\mathrm{out}}$", + "Figure 6. The heat-maps of task distances of four streams in CTrL benchmark. The PAR can recognize the related tasks in $S^{-}$ , $S^{+}$ , and $S^{\\mathrm{out}}$ exactly." + ], + "image_footnote": [], + "bbox": [ + 297, + 213, + 455, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results in Tab. 3 show that PAR outperforms parameter regularization methods, memory-based methods, and parameter allocation methods with dynamic model on the CIFAR100-10 and MiniImageNet-20.", + "bbox": [ + 75, + 412, + 468, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At last, we analyze the parameter overhead3 of PAR on CIFAR100-10 and $S^{\\mathrm{long}}$ . Compared with baselines, the PAR obtains better average performance with fewer model parameters. The reason is that the allocated experts in PAR have compact architecture and are parameter efficient. Performance on the stream $S^{\\mathrm{long}}$ shows that PAR is scalable.", + "bbox": [ + 75, + 474, + 468, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 573, + 241, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the effects of components in PAR and the results are listed in Tab. 5. First, we evaluate the impact of hierarchical architecture search on parameter allocation. Compared with using a fixed cell from DARTs [22] (#1), searching cells from fine-grained space can improve model performance (#2). Combined with coarse-grained space, searching from the hierarchical space can further improve performance while reducing the time and parameter overhead (#3). Then, we find that parameter regularization alone is time efficient, but its performance is low (#4). Finally, by selecting an appropriate strategy from parameter allocation and regularization based on the learning difficulty of each new task, PAR makes a trade off among average performance, parameter overhead, and time overhead.", + "bbox": [ + 75, + 597, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the impact of important hyper-parameters $\\alpha$ and $\\beta$ in PAR on the CIFAR100-10. The $\\alpha$ determines the reuse experts during the learning and the larger $\\alpha$ encour", + "bbox": [ + 75, + 809, + 468, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages the model to reuse experts. Results in Fig. 3 show that the model size and number of experts decreases with the increase of $\\alpha$ . The $\\beta$ determines the frequency of architecture search and the larger $\\beta$ encourages the model to reuse architectures of existing experts. Results in Fig. 4 show that the learning time decreases with the increase of $\\beta$ .", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization", + "text_level": 1, + "bbox": [ + 500, + 191, + 635, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To analyze the task distance in PAR, we construct a new benchmark CIFAR100-coarse by dividing CIFAR100 into 20 tasks based on the coarse-grained labels. The illustration in Fig. 5 shows that the PAR can obtain reasonable task groups. For example, the PAR adaptively puts tasks about animals into the same groups, such as the Aquatic mammals, Large omnivores and herbivores, Fish, and so on. The Food containers and Household electrical devices are divided into the same group since they both contain cylindrical and circular objects. It also finds the relation between Non-insect invertebrates and Insects.", + "bbox": [ + 496, + 214, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we visualize the heat maps of task distances obtained by PAR on four streams of CTrL in Fig. 6. The distance between the first and the last task on streams $S^{-}$ , $S^{+}$ and $S^{\\mathrm{out}}$ are small, which is consistent with the fact that they all come from the same data distribution. Instead, the distribution of the last task on stream $S^{\\mathrm{in}}$ is perturbed from the distribution of the first task, so their distance is large. The above results show that PAR can obtain reasonable task distance and relatedness.", + "bbox": [ + 496, + 380, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 530, + 617, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new method named Parameter Allocation & Regularization (PAR). It allows the model to adaptively select an appropriate strategy from parameter allocation and regularization for each task based on the task learning difficulty. Experimental results on multiple benchmarks demonstrate that PAR is scalable and significantly reduces the model redundancy while improving performance. Moreover, qualitative analysis shows that PAR can obtain reasonable task relatedness. Our method is flexible, and in the future, we will introduce more relatedness estimation, regularization, and allocation strategies into PAR.", + "bbox": [ + 496, + 555, + 890, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 734, + 658, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the NSFC projects (No. 62072399, No. U19B2042), Chinese Knowledge Center for Engineering Sciences and Technology, MoE Engineering Research Center of Digital Library, China Research Centre on Data and Knowledge for Engineering Sciences and Technology, and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 496, + 760, + 890, + 864 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3We ignore the parameter overhead of the extractor (11.18M) in divergence estimation, which is fixed and will not become a bottleneck with the increase of tasks.", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7783", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory Aware Synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 6", + "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert Gate: Lifelong Learning With a Network of Experts. In CVPR, pages 3366-3375, 2017. 1, 2", + "[3] Ali Ayub and Alan Wagner. EEC: Learning to Encode and Regenerate Images for Continual Learning. In ICLR 2021, 2020. 2", + "[4] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark Experience for General Continual Learning: A Strong, Simple Baseline. In NeurIPS 2020, page 11, 2020. 7", + "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient Lifelong Learning with A-GEM. In ICLR 2019, Sept. 2018. 6", + "[6] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. On Tiny Episodic Memories in Continual Learning. arXiv:1902.10486 [cs, stat], June 2019. 6", + "[7] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating Forgetting in Online Continual Learning via Instance-Aware Parameterization. In NeurIPS 2020, page 12, 2020. 1, 2, 7", + "[8] Zhi-Qi Cheng, Xiao Wu, Siyu Huang, Jun-Xiu Li, Alexander G Hauptmann, and Qiang Peng. Learning to transfer: Generalizable attribute learning with multitask neural model search. In Proceedings of the 26th ACM international conference on Multimedia, pages 90-98, 2018. 2", + "[9] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, pages 1-1, 2021. 1", + "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning Without Memorizing. In CVPR, pages 5138-5146, 2019. 1, 2", + "[11] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural Architecture Search: A Survey. JMLR, page 21, 2019. 2", + "[12] Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4):8, 1999. 1", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, June 2016. IEEE. 1, 7", + "[14] Siyu Huang, Xi Li, Zhi-Qi Cheng, Zhongfei Zhang, and Alexander Hauptmann. Gnas: A greedy neural architecture search method for multi-attribute learning. In Proceedings of the 26th ACM international conference on Multimedia, pages 2049–2057, 2018. 2", + "[15] Zixuan Ke, Bing Liu, and Xingchang Huang. Continual Learning of a Mixed Sequence of Similar and Dissimilar" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tasks. In NeurIPS 2020, volume 33, pages 18493-18504. Curran Associates, Inc., 2020. 1, 2, 5, 6, 7", + "[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 1, 2, 6", + "[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6", + "[18] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 1", + "[19] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming Catastrophic Forgetting by Incremental Moment Matching. In NeurlPS 2017, volume 30, pages 4652-4662, 2017. 1, 2", + "[20] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting. In ICML, page 10, 2019. 1, 2, 6", + "[21] Zhizhong Li and Derek Hoiem. Learning without Forgetting. IEEE transactions on pattern analysis and machine intelligence, page 14, 2017. 1, 2, 4, 6", + "[22] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: DIFFERENTIABLE ARCHITECTURE SEARCH. In ICLR 2019, page 13, 2018. 2, 5, 8", + "[23] Huihui Liu, Yiding Yang, and Xinchao Wang. Overcoming Catastrophic Forgetting in Graph Neural Networks. In AAAI 2021, 2020. 1, 2", + "[24] Jorge A. Mendez and Eric Eaton. Lifelong Learning of Compositional Structures. In ICLR 2021, Sept. 2020. 1, 2, 6", + "[25] Qier Meng and Satoh Shin'ichi. ADINet: Attribute Driven Incremental Network for Retinal Image Classification. In CVPR 2020, pages 4033-4042, 2020. 1, 2", + "[26] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual Learning with Filter Atom Swapping. In International Conference on Learning Representations, Sept. 2021. 1, 2", + "[27] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. VARIATIONAL CONTINUAL LEARNING. In ICLR, page 18, 2018. 1, 2", + "[28] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Johannichen, and Moin Nabi. Learning to Remember: A Synaptic Plasticity Driven Framework for Continual Learning. In CVPR, pages 11321-11329, 2019. 2", + "[29] Oleksiy Ostapenko, Pau Rodriguez, Massimo Caccia, and Laurent Charlin. Continual Learning via Local Module Composition. In NeurIPS 2021, May 2021. 1, 2, 6, 7", + "[30] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. BNS: Building Network Structures Dynamically for Continual Learning. In NeurIPS 2021, May 2021. 1, 2, 7", + "[31] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random Path Selection for Continual Learning. In H. Wallach, H. Larochelle, A." + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7784", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Beygelzimer, F. d\\text{text}quotingingle Alché-Buc, E. Fox, and R. Garnett, editors, NeurIPS 2019, pages 12669-12679. Curran Associates, Inc., 2019. 1, 2, 7", + "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental Classifier and Representation Learning. In CVPR, pages 2001–2010, 2017. 2", + "[33] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and Xin Wang. A Comprehensive Survey of Neural Architecture Search: Challenges and Solutions. ACM Computing Surveys, 2020. 2", + "[34] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive Neural Networks. arXiv:1606.04671 [cs], June 2016. 1, 2, 6", + "[35] Gobinda Saha and Kaushik Roy. Gradient Projection Memory for Continual Learning. In ICLR 2021, Sept. 2020. 6", + "[36] Joan Serrà, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming Catastrophic Forgetting with Hard Attention to the Task. In ICML, page 10, 2018. 1, 2, 6", + "[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual Learning with Deep Generative Replay. In NeurIPS 2017, pages 2990-2999. Curran Associates, Inc., 2017. 2", + "[38] Pravendra Singh, Pratik Mazumder, Piyush Rai, and Vinay P. Namboodiri. Rectification-Based Knowledge Retention for Continual Learning. In CVPR 2021, pages 15282-15291, 2021. 1, 2", + "[39] Pravendra Singh, Vinay Kumar Verma, Pratik Mazumder, Lawrence Carin, and Piyush Rai. Calibrating CNNs for Lifelong Learning. In NeurIPS 2020, volume 33, 2020. 1, 2", + "[40] Binh Tang and David S. Matteson. Graph-Based Continual Learning. In ICLR 2021, Sept. 2020. 2", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is All you Need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, NeurIPS 2017, pages 5998-6008. Curran Associates, Inc., 2017. 1", + "[42] Tom Veniat, Ludovic Denoyer, and MarcAurelio Ranzato. Efficient Continual Learning with Modular Networks and Task-Driven Priors. In ICLR 2021, Sept. 2020. 1, 2, 5, 6, 7", + "[43] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 6", + "[44] Qing Wang, Sanjeev R. Kulkarni, and Sergio Verdu. Divergence Estimation for Multidimensional Densities Via $\\$k$ -Nearest-Neighbor Distances. IEEE Transactions on Information Theory, 55(5):2392–2405, 2009. 2, 3", + "[45] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: An Alternative Approach to Efficient Ensemble and Lifelong Learning. In ICLR, Sept. 2019. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Ju Xu and Zhanxing Zhu. Reinforced Continual Learning. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, NeurIPS 2018, pages 899–908. Curran Associates, Inc., 2018. 2", + "[47] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically Expandable Representation for Class Incremental Learning. In CVPR 2021, pages 3014-3023, 2021. 1, 2", + "[48] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial Distribution Learning for Effective Neural Architecture Search. In ICCV, pages 1304-1313, 2019. 2, 5", + "[49] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning Transferable Architectures for Scalable Image Recognition. In CVPR 2018, pages 8697-8710, Salt Lake City, UT, 2018. IEEE. 2" + ], + "bbox": [ + 501, + 92, + 890, + 303 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7785", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_model.json b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..085eee06e925423be3c612159175981670683d78 --- /dev/null +++ b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.13, + 0.861, + 0.178 + ], + "angle": 0, + "content": "Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.204, + 0.713, + 0.24 + ], + "angle": 0, + "content": "Wenjin Wang, Yunqing Hu, Qianglong Chen, Yin Zhang* Zhejiang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.242, + 0.74, + 0.258 + ], + "angle": 0, + "content": "{wangwenjin,yunqinghu,chenqianglong,zhangyin98}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.324, + 0.474, + 0.686 + ], + "angle": 0, + "content": "Parameter regularization or allocation methods are effective in overcoming catastrophic forgetting in lifelong learning. However, they solve all tasks in a sequence uniformly and ignore the differences in the learning difficulty of different tasks. So parameter regularization methods face significant forgetting when learning a new task very different from learned tasks, and parameter allocation methods face unnecessary parameter overhead when learning simple tasks. In this paper, we propose the Parameter Allocation & Regularization (PAR), which adaptively select an appropriate strategy for each task from parameter allocation and regularization based on its learning difficulty. A task is easy for a model that has learned tasks related to it and vice versa. We propose a divergence estimation method based on the Nearest-Prototype distance to measure the task relatedness using only features of the new task. Moreover, we propose a time-efficient relatedness-aware sampling-based architecture search strategy to reduce the parameter overhead for allocation. Experimental results on multiple benchmarks demonstrate that, compared with SOTAs, our method is scalable and significantly reduces the model's redundancy while improving the model's performance. Further qualitative analysis indicates that PAR obtains reasonable task-relatedness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.716, + 0.21, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.741, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Recently, the lifelong learning [9] ability of neural networks, i.e., learning continuously from a continuous sequence of tasks, has been extensively studied. It is natural for human beings to constantly learn and accumulate knowledge from tasks and then use it to facilitate future learning. However, classical models [13, 18, 41] suffer catastrophic forgetting [12], i.e., the model's performance on learned tasks deteriorates rapidly after learning a new one." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.863, + 0.47, + 0.878 + ], + "angle": 0, + "content": "To overcome the catastrophic forgetting, many param-" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.293, + 0.897, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.533, + 0.895, + 0.588 + ], + "angle": 0, + "content": "Figure 1. PAR adaptively selects a strategy from regularization and allocation to handle each task according to its learning difficulty. The difficulty depends not only on the task itself, but also on whether the model has previously learned tasks related to it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.895, + 0.867 + ], + "angle": 0, + "content": "eter regularization or allocation methods have been proposed. Parameter regularization methods [10, 16, 19, 21, 23, 25, 27] alleviate forgetting by adding a regularization term to the loss function and perform well when the new task does not differ much from learned tasks. Parameter allocation methods based on static models [7, 15, 31, 36] and dynamic models [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] allocate different parameters to different tasks and can adapt to new tasks quite different from learned tasks. However, the above methods solve all tasks in a sequence uniformly, and ignore the differences of learning difficulty of different tasks. This leads to the significant forgetting in parameter regularization methods when learning a new task which is quite different from learned tasks, and also leads to unnecessary parameter cost in parameter allocation methods when learning some simple tasks." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a difficulty-aware method Parameter Allocation & Regularization (PAR). As shown" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.286, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Yin Zhang." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7776" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "in Fig. 1, we assume that the learning difficulty of a task in continual learning depends not only on the task itself, but also on the accumulated knowledge in the model. A new task is easy to adapt for a model if it has learned related tasks before and vice versa. Based on the assumption, the PAR adaptively adopts parameter allocation for difficult tasks and parameter regularization for easy tasks. Specifically, the PAR divides tasks into task groups and assigns each group a dedicated expert model. Given a new task, the PAR measures the relatedness between it and existing task groups at first. If the new task is related to one of the existing groups, it is easy for the corresponding expert. The PAR adds the task to the related group and learns it by the expert via the parameter regularization. Otherwise, the new task is difficult for all existing experts, and the PAR assigns it to a new task group and allocates a new expert to learn it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.635 + ], + "angle": 0, + "content": "There are two challenges in this work: the measurement of relatedness and the parameter explosion associated with parameter allocation. For the first one, we try to measure the relatedness by the KL divergence between feature distributions of tasks. However, the KL divergence is intractable and needs to be estimated since the feature distributions of tasks are usually unknown. In addition, the constraint of lifelong learning that only data of the current task are available exacerbates the difficulty of estimation. To solve above problems, inspired by the divergence estimation based on k-NN distance [44], we propose the divergence estimation method based on prototype distance, which only depends on the data of the current task. For the second one, we try to reduce parameter overhead per expert by searching compact architecture for it. However, the low time and memory efficiency is an obstacle to applying architecture search for a sequence of tasks in lifelong learning. To improve the efficiency of architecture search, we propose a relatedness-aware sampling-based hierarchical search. The main contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.642, + 0.47, + 0.732 + ], + "angle": 0, + "content": "- We propose a lifelong learning framework named Parameter Allocation & Regularization (PAR), which selects an appropriate strategy from parameter allocation and regularization for each task based on the learning difficulty. The difficulty depends on whether the model has learned related tasks before." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.741, + 0.47, + 0.832 + ], + "angle": 0, + "content": "- We propose a divergence estimation method based on prototype distance to measure the distance between the new task and previous learned tasks with only data of the new task. Meanwhile, we propose a relatedness-aware sampling-based architecture search to reduce the parameter overhead of parameter allocation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- Experimental results on CTrL, Mixed CIFAR100 and F-CelebA, CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20 demonstrate that PAR is scalable and significantly reduces the model redundancy" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.642, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "while improving the model performance. Exhaustive ablation studies show the effectiveness of components in PAR and the visualizations show the reasonability of task distance in PAR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.167, + 0.642, + 0.183 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.676, + 0.209 + ], + "angle": 0, + "content": "2.1. Lifelong Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Many methods have been proposed to overcome catastrophic forgetting. Replay methods try to replay samples of previous tasks when learning a new task from an episodic memory [32,40] or a generative memory [3,28,37]. Parameter regularization methods, including the prior-focused regularization [16, 19, 23, 27] and the data-focused regularization [10, 21, 25], try to alleviate forgetting by introducing a regularization term in the loss function of the new task. Parameter allocation methods based on the static model [7, 15, 31, 36] and the dynamic model [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] overcome catastrophic forgetting by allocating different parameters to different tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Methods with relatedness. Several methods also consider the utility of task relatedness [2, 15]. Expert Gate [2] assigns dedicated experts and auto-encoders for tasks and calculates the task relatedness by reconstruction error of auto-encoders. The task relatedness is used to transfer knowledge from the most related previous task. CAT [15] defines the task similarity by the positive knowledge transfer. It focuses on selectively transferring the knowledge from similar previous tasks and dealing with forgetting between dissimilar tasks by hard attention. In this paper, we propose a divergence estimation method based on prototype distance to calculate the task relatedness used to measure the task learning difficulty." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Methods with NAS. Static model based methods [7, 31] try to search a sub-model for each task with neural architecture search (NAS). Dynamic model based methods [20, 42, 46] adopt NAS to select an appropriate model expansion strategy for each task and face the high GPU memory, parameter, and time overhead. Instead, in this paper, we propose a relatedness-aware hierarchical cell-based architecture search to search compact architecture for each expert with a lower GPU memory and time overhead. [8, 14] adopt NAS on multi-task learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.659, + 0.771 + ], + "angle": 0, + "content": "2.2. Cell-based NAS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Neural architecture search (NAS) [11, 33] aims to search for efficient neural network architectures from a pre-defined search space in a data-driven manner. To reduce the size of the search space, cell-based NAS methods [22, 48, 49] try to search for a cell architecture from a pre-defined cell search space, where a cell is a tiny convolutional network mapping an \\( H \\times W \\times F \\) tensor to another \\( H' \\times W' \\times F' \\). The final model consists of a pre-defined number of stacked" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "cells. The cell in NAS is similar to the residual block in the residual network (ResNet), but its architecture is more complex and is a directed acyclic graph (DAG). The operations in the search space are usually parameter-efficient. NAS methods usually produce more compact architectures than hand-crafted designs." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.197, + 0.17, + 0.213 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.345 + ], + "angle": 0, + "content": "We focus on the task-incremental scenario of lifelong learning. Specifically, the model learns a sequence of tasks \\(\\mathcal{T} = \\{T_1,\\dots ,T_t,\\dots ,T_N\\}\\) one by one and the task id of sample is available during both training and inference. Each task \\(T_{t}\\) has a training dataset, \\(D_{train}^{t} = \\{(x_{i}^{t},y_{i}^{t});i = 1,\\ldots ,n_{train}^{t}\\}\\), where \\(y_{i}^{t}\\) is the true label and \\(n_{train}^{t}\\) is the number of training examples. Similarly, we denote the validation and test dataset of task \\(T_{t}\\) as \\(D_{valid}^{t}\\) and \\(D_{test}^{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.471, + 0.45 + ], + "angle": 0, + "content": "As shown in Fig. 2, the PAR divides previously learned tasks into task groups \\(\\mathcal{G}\\) and assigns each group \\(\\mathcal{G}_i\\) a dedicated expert model \\(E_i\\). For a new task, PAR calculates the distances, which reflect the relatedness, between it and existing task groups at first. Then, it adopts an appropriate learning strategy from parameter allocation and regularization for the new task based on the distances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.465, + 0.478 + ], + "angle": 0, + "content": "3.1. Task Distance via Nearest-Prototype Distance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.545 + ], + "angle": 0, + "content": "The PAR calculates the distance between the new task and each existing task at first, then it averages the distances of tasks in the same group as the distance between the new task and the group." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.471, + 0.711 + ], + "angle": 0, + "content": "We calculate the distance between two tasks by the KL divergence between their distributions. However, the KL divergence is intractable because the true distributions of tasks are usually agnostic. Instead, we estimate the KL divergence by the features of data. To enhance estimation stability, we introduce an extra pre-trained extractor1 to generate a robust feature \\( X_{i}^{t} \\) for each image \\( x_{i}^{t} \\) in task \\( T_{t} \\). The pre-trained extractor is only used for divergence estimation and does not affect the learning of tasks. The number of extra parameters introduced by the model is a constant and does not affect the scalability of our method." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.731, + 0.47, + 0.854 + ], + "angle": 0, + "content": "\\(k\\)-Nearest Neighbor Distance Given the features of distributions, a classical divergence estimation method for multidimensional densities is the \\(k\\)-Nearest Neighbor (\\(k\\)-NN) distance [44]. Suppose the feature distributions of the new task \\(T_{i}\\) and an existing task \\(T_{j}\\) are \\(q_{i}\\) and \\(q_{j}\\) respectively, and the features \\(X^{i} = \\{X_{1}^{i},\\ldots ,X_{n}^{i}\\}\\) and \\(X^{j} = \\{X_{1}^{j},\\ldots ,X_{m}^{j}\\}\\) are drawn independently from distributions \\(q_{i}\\) and \\(q_{j}\\), where \\(n\\) and \\(m\\) are the numbers of samples. The" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1We adopt a ResNet18 pre-trained on the ImageNet. It does not introduce unfairness because the generated features are not involved in model training and prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\(k\\)-NN distance of feature \\(X_{l}^{i}\\) in \\(X^{i}\\) and \\(X^{j}\\) are denoted by \\(\\nu_{k}(l)\\) and \\(\\rho_{k}(l)\\) respectively. Specifically, the \\(\\rho_{k}(l)\\) is the Euclidean distance from \\(X_{l}^{i}\\) to its \\(k\\)-NN in \\(\\{X_{z}^{i}\\}_{z\\neq l}\\). Similarly, the \\(\\nu_{k}(l)\\) is the distance from \\(X_{l}^{i}\\) to its \\(k\\)-NN in \\(X^{j}\\). Then, the KL divergence between distributions \\(q_{i}\\) and \\(q_{j}\\) are estimated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.188, + 0.892, + 0.24 + ], + "angle": 0, + "content": "\\[\nK L \\left(q _ {i} \\mid \\mid q _ {j}\\right) \\approx \\widehat {K L} \\left(q _ {i} \\mid \\mid q _ {j}\\right) = \\frac {d}{n} \\sum_ {l = 1} ^ {n} \\log \\frac {\\nu_ {k} (l)}{\\rho_ {k} (l)} + \\log \\frac {m}{n - 1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.241, + 0.722, + 0.255 + ], + "angle": 0, + "content": "where \\(d\\) is the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.347 + ], + "angle": 0, + "content": "The \\(k\\)-NN distance is asymmetric and this is consistent with the asymmetry of the KL divergence. The motivation behind Eq. (1) is that the \\(k\\)-NN distance is the radius of a \\(d\\)-dimensional open Euclidean ball used for \\(k\\)-NN density estimation. Readers can refer to [44] for more details such as convergence analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Nearest-Prototype Distance However, the calculation of the \\(k\\)-NN distance involves features of two tasks \\(T_{i}\\) and \\(T_{j}\\) and violates the constraint of lifelong learning that only data of the current task \\(T_{i}\\) are available. To overcome this problem, we propose the Nearest-Prototype distance to replace the \\(k\\)-NN distance. Specifically, suppose the set of classes in task \\(T^{i}\\) is \\(C^i\\), for each class \\(c \\in C^i\\), we maintain a prototype feature \\(U_{c}^{i}\\), which is the mean of features of samples belonging to the class \\(c\\). Then, the Nearest-Prototype distance of \\(X_{l}^{i}\\) to \\(X^{i}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.523, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\rho (l) = \\left\\| X _ {l} ^ {i} - U _ {c (l)} ^ {i} \\right\\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.625 + ], + "angle": 0, + "content": "where \\(||\\cdot ||\\) is the Euclidean distance and \\(c(l)\\) is the class of \\(X_{l}^{i}\\). Similarly, we denote the set of classes in task \\(T_{j}\\) by \\(C^j\\) and the prototype features of \\(X^{j}\\) for each class \\(c\\in C^j\\) by \\(U_{c}^{j}\\). Then, the Nearest-Prototype distance of \\(X_{l}^{i}\\) to \\(X^{j}\\) is denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.634, + 0.892, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\nu (l) = \\min _ {c} \\| X _ {l} ^ {i} - U _ {c} ^ {j} \\| _ {2}, c \\in C ^ {j}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.725 + ], + "angle": 0, + "content": "Then, by replacing the \\(k\\)-NN distances \\(\\nu_{k}(l)\\) and \\(\\rho_{k}(l)\\) in Eq. (1) by the Nearest-Prototype distances \\(\\nu(l)\\) and \\(\\rho(l)\\), the KL divergence estimated by the Nearest-Prototype distance is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "\\[\nK L \\left(q _ {i} \\mid \\mid q _ {j}\\right) \\approx \\widetilde {K L} \\left(q _ {i} \\mid \\mid q _ {j}\\right) = \\sum_ {c \\in C ^ {i}} \\frac {1}{n _ {c}} \\sum_ {l = 1} ^ {n _ {c}} \\log \\frac {\\nu (l)}{\\rho (l)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.871 + ], + "angle": 0, + "content": "where constant terms are omitted because we are only concerned with the relative magnitude of KL divergence. The Eqs. (2) to (4) only involve features of the new task \\( T_{i} \\) and prototypes of the existing task \\( T_{j} \\), which satisfies the constraint of lifelong learning. The storage cost of prototypes is small and negligible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The motivation behind Eq. (4) is intuitive. The Nearest-Prototype distance \\(\\rho(l)\\) reflects the relationship between the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.852, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.363, + 0.895, + 0.42 + ], + "angle": 0, + "content": "Figure 2. The architecture of PAR. PAR divides previously learned tasks into task groups and each group has a dedicated expert model. Given a new task \\( T_{5} \\), PAR calculates the distance between it and existing task groups to measure its learning difficulty. Then, PAR learns the new task by parameter regularization if it is easy to learn. Otherwise, PAR learns the new task by parameter allocation and search compact architecture for each expert by a relatedness-aware hierarchical architecture search." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.552 + ], + "angle": 0, + "content": "sample \\(X_{l}^{i}\\) in task \\(T_{i}\\) and its class prototype in feature space. The Nearest-Prototype distance \\(\\nu (l)\\) reflects the relationship between the sample \\(X_{l}^{i}\\) and class prototypes of existing task \\(T_{j}\\) in feature space. If the value of \\(\\rho (l)\\) and \\(\\nu (l)\\) are equal for each \\(X_{l}^{i}\\), the class prototypes of two tasks are close. In this case, the distribution of the two tasks are similar, and the estimated KL divergence by Eq. (4) is 0." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.57, + 0.47, + 0.645 + ], + "angle": 0, + "content": "Adaptive Learning Strategy Given the distance between the new task \\( T_{i} \\) and each existing tasks by KL divergence, we denote the distance \\( s_{i,g}^{\\prime} \\) between the task \\( T_{i} \\) and a task group \\( \\mathcal{G}_g \\) by the average of distances of tasks in the group as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.65, + 0.47, + 0.695 + ], + "angle": 0, + "content": "\\[\ns _ {i, g} ^ {\\prime} = \\frac {1}{| \\mathcal {G} _ {g} |} \\sum_ {j ^ {\\prime}} ^ {| \\mathcal {G} _ {g} |} K L (q _ {i} | | q _ {j ^ {\\prime}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.793 + ], + "angle": 0, + "content": "where \\(q_{i}\\) and \\(q_{j'}\\) represent feature distributions of task \\(T_{i}\\) and the \\(j'\\)-th task in group \\(\\mathcal{G}_g\\). The \\(s_{i,g}' \\in [0, \\infty)\\) since the range of the KL divergence is \\([0, \\infty)\\). However, we try to use the distance to reflect the relatedness, which is usually measured by a value between 0 and 1. So we map the \\(s_{i,g}'\\) into \\([0, 1]\\) by a monotone increasing function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.799, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\ns _ {i, g} = \\min \\left(s _ {i, g} ^ {\\prime}, 1 - e ^ {- 2 \\times s _ {i, g} ^ {\\prime}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.281, + 0.839 + ], + "angle": 0, + "content": "where \\(e\\) is the Euler's number." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Suppose the smallest distance between the new task \\( T_{t} \\) and existing groups is \\( s_{t,g^{*}} \\) and the corresponding task group is \\( \\mathcal{G}_{g^{*}} \\), the PAR selects the learning strategy by comparing the \\( s_{t,g^{*}} \\) with a hyper-parameter \\( \\alpha \\). If \\( s_{t,g^{*}} \\leq \\alpha \\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.507 + ], + "angle": 0, + "content": "the new task is added to this related group \\(\\mathcal{G}_{g^*}\\) and learned by parameter regularization. Otherwise, the new task is assigned to a new group and learned by parameter allocation because no existing task group is related to it." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.738, + 0.531 + ], + "angle": 0, + "content": "3.2. Parameter Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.659 + ], + "angle": 0, + "content": "A new task \\( T_{t} \\) is easy for the expert model \\( E_{g^{*}} \\) if it is related to the group \\( G_{g^{*}} \\). PAR reuses the expert \\( E_{g^{*}} \\) to learn this task by parameter regularization. Inspired by LwF [21], we adopt a parameter regularization method based on knowledge distillation. Specifically, the loss function consists of a training loss \\( \\mathcal{L}_{\\mathrm{new}} \\) and a distillation loss \\( \\mathcal{L}_{\\mathrm{old}} \\). The training loss encourages expert \\( E_{g^{*}} \\) to adapt to the new task \\( T_{t} \\) and is the cross-entropy for classification as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.668, + 0.892, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n e w}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\log \\left(p _ {E _ {g ^ {*}}} (y | x)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.787 + ], + "angle": 0, + "content": "The distillation loss encourages the expert \\( E_{g^*} \\) to maintain performance on previous tasks in the group \\( G_{g^*} \\). To calculate it, we record the logits \\( \\mathbf{y}^j \\) of the output head of each previous task \\( T_j \\) for each sample \\( x \\) of task \\( T_t \\). The distillation loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.793, + 0.892, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o l d}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\sum_ {T _ {j} \\in \\mathcal {G} _ {g ^ {*}}} \\mathbf {y} ^ {j} \\cdot \\log \\left(\\mathbf {p} _ {\\mathbf {E} _ {\\mathbf {g} ^ {*}}} (\\mathbf {x}) ^ {j}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.884 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}^j\\) and \\(\\mathbf{p}_{\\mathbf{E}_{\\mathbf{g}^*}}(\\mathbf{x})^j\\) are vectors with lengths equal to the number of categories of the previous task \\(T_{j}\\). The total loss is" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P R}} = \\mathcal {L} _ {\\text {n e w}} + \\lambda \\mathcal {L} _ {\\text {o l d}}, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7779" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.875, + 0.317 + ], + "angle": 0, + "content": "
MethodS-S+SinSoutSplSlong
APAFAPAFAPAFAPAFAPAFAPAF
Finetune0.180-0.3300.240-0.2300.180-0.3100.150-0.3700.210-0.3000.200-0.400
(1)EWC0.540-0.0100.370-0.0400.430-0.1200.510-0.0300.300-0.1700.270-0.300
(2)ER0.410-0.1300.430-0.0700.380-0.1700.540-0.0700.450-0.080--
(3)HAT*0.570-0.0100.5700.0000.580-0.0100.6000.0000.5800.0000.240-0.100
(4)Independent0.5600.0000.5700.0000.5700.0000.6100.0000.5900.0000.5700.000
PNN0.6200.0000.5200.0000.5700.0000.6200.0000.5400.000--
SG-F0.6360.0000.6150.0000.6550.0000.6410.0000.6200.000--
MNTDP-S0.6300.0000.5600.0000.5700.0000.6400.0000.5500.0000.6800.000
MNTDP-D0.6700.0000.6100.0000.6000.0000.6800.0000.6200.0000.7500.000
LMC0.6660.0000.601-0.0140.6950.0000.667-0.0100.616-0.0350.639-
PAR(our)0.7060.0000.670-0.0170.6990.0000.6940.0000.7000.0000.773-0.016
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.327, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Table 1. The average performance (AP) and average forgetting (AF) on the six streams \\((S^{-}, S^{+}, S^{\\mathrm{in}}, S^{\\mathrm{out}}, S^{\\mathrm{pl}}, S^{\\mathrm{long}})\\) in the CTRL benchmark. We compare the PAR with four types of baselines: (1) parameter regularization, (2) replay, (3) parameter allocation with static model, and (4) parameter allocation with dynamic model. The * corresponds to model using the AlexNet backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.442 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter to balance training and distillation loss. Note that our method is without memory and does not require storing samples for previous tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.641 + ], + "angle": 0, + "content": "However, the expert \\( E_{g^{*}} \\) may over-fit the new task whose sample size is far less than tasks in the group \\( \\mathcal{G}_{g^{*}} \\), even though the new task is related to the group. This leads to interference with previously accumulated knowledge in the expert \\( E_{g^{*}} \\). To avoid the above problem, PAR records the maximum sample size of tasks in each task group. Suppose the maximum sample size in the group \\( \\mathcal{G}_{g^{*}} \\) is \\( Q \\), PAR freezes the parameters of expert \\( E_{g^{*}} \\) except for the task-specific classification head during the learning if the sample size of the new task is less than 10 percent of \\( Q \\). By transferring the existing knowledge in expert \\( E_{g^{*}} \\), only optimizing the classification header is sufficient to adapt to the new task because the new task is related to the group \\( \\mathcal{G}_{g^{*}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.279, + 0.665 + ], + "angle": 0, + "content": "3.3. Parameter Allocation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.47, + 0.737 + ], + "angle": 0, + "content": "If no existing groups are related to the new task \\( T_{t} \\), PAR assigns it to a new task group \\( \\mathcal{G}_{M + 1} \\) with a new expert \\( E_{M + 1} \\), where \\( M \\) is the number of existing groups. We adopt the cross-entropy loss for the task \\( T_{t} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.747, + 0.47, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P A}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\log \\left(p _ {E _ {M + 1}} (y | x)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.47, + 0.87 + ], + "angle": 0, + "content": "The number of experts and parameters in PAR is proportional to the number of task groups, mitigating the growth of parameters. To further reduce the parameter overhead of each expert, we adopt cell-based NAS (refer to Sec. 2.2 for details) to search for a compact architecture for it." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Each expert in PAR is stacked with multiple cells and the search for expert architecture is equivalent to the search" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.457 + ], + "angle": 0, + "content": "for the cell architecture. Since the time overhead of NAS becomes unbearable as the number of tasks increases, we propose a relatedness-aware sampling-based architecture search strategy to improve efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.893, + 0.655 + ], + "angle": 0, + "content": "As shown in Fig. 2, we construct a hierarchical search space. The coarse-grained search space contains cells used by existing experts and an unknown cell whose architecture will come from the fine-grained search space. Following common practice [22, 48], the fine-grained search space is a directed acyclic graph (DAG) with seven nodes (two input nodes \\( i_1 \\), \\( i_2 \\), an ordered sequence of intermediate nodes \\( n_1, n_2, n_3, n_4 \\), and an output node). The input nodes represent the outputs of the previous two layers. The output is concatenated from all intermediate nodes. Each intermediate node has a directed candidate edge from each of its predecessors. Each candidate edge is associated with six parameter-efficient candidate operations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.807 + ], + "angle": 0, + "content": "To search a cell for the new task, we introduce a hyperparameter \\(\\beta\\). If \\(s_{t,g^*} \\leq \\beta\\), PAR reuses the cell of expert \\(E_{g^*}\\) for the task \\(T_t\\). A task distance greater than \\(\\alpha\\) and less than \\(\\beta\\) indicates that the new task is not related enough to share the same expert with tasks in the group \\(\\mathcal{G}_{g^*}\\) but can use the same architecture. If \\(s_{t,g^*} > \\beta\\), PAR assigns the unknown cell to the new expert and adopts a sampling-based NAS method named MDL [48] to determine its architecture. Due to limited space, we leave the details of the MDL in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.634, + 0.838 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.707, + 0.864 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Benchmarks We adopt two benchmarks the CTrL [42] and the Mixed CIFAR100 and F-CelebA [15], which con" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7780" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.447, + 0.136 + ], + "angle": 0, + "content": "
MethodFinetuneHATCATPAR
AP0.61550.61780.68310.7122
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.146, + 0.47, + 0.188 + ], + "angle": 0, + "content": "Table 2. We compare the average performance (AP) of PAR with two parameter allocation methods based on static model on the mixed CIFAR100 and F-CelebA benchmark." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.203, + 0.467, + 0.439 + ], + "angle": 0, + "content": "
#MethodCIFAR100-10MiniImageNet-20
AP(%)AF(%)AP(%)AF(%)
Finetune0.280-0.5600.378-0.314
1EWC0.608-0.0610.574-0.144
MAS0.579-0.0500.368-0.238
LwF0.655-0.0880.589-0.169
2GPM*0.725-0.604-
A-GEM0.677-0.1200.524-0.152
3Independent0.8200.0000.7870.000
PN0.8210.0000.7810.000
Learn to Grow0.7910.000--
MNTDP-S0.750-0.000--
MNTDP-D0.830-0.000--
PAR0.853-0.0200.816-0.022
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Table 3. PAR outperforms (1) parameter regularization methods, (2) replay methods, and (3) parameter allocation methods with dynamic model on classical benchmarks. The * corresponds to model using the AlexNet backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.468, + 0.683 + ], + "angle": 0, + "content": "tain mixed similar and dissimilar tasks. CTrL [42] includes 6 streams of image classification tasks: \\( S^{-} \\) is used to evaluate the ability of direct transfer, \\( S^{+} \\) is used to evaluate the ability of knowledge update, \\( S^{\\mathrm{in}} \\) and \\( S^{\\mathrm{out}} \\) are used to evaluate the transfer to similar input and output distributions respectively, \\( S^{\\mathrm{pl}} \\) is used to evaluate the plasticity, \\( S^{\\mathrm{long}} \\) consists of 100 tasks and is used to evaluate the scalability. Similarly, Mixed CIFAR100 and F-CelebA [15] including mixed 10 similar tasks from F-CelebA and 10 dissimilar tasks from CIFAR100 [17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.468, + 0.835 + ], + "angle": 0, + "content": "Further, we adopt classical task incremental learning benchmarks: CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20. CIFAR10-5 is constructed by dividing CIFAR10 [17] into 5 tasks and each task has 2 classes. Similarly, CIFAR100-10 and CIFAR100-20 are constructed by dividing CIFAR100 [17] into 10 10-classification tasks and 20 5-classification tasks respectively. MiniImageNet-20 is constructed by dividing MiniImageNet [43] into 20 tasks and each task has 5 classes. We leave the results on the CIFAR10-5 and CIFAR100-20 in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Baselines Firstly, we compare our method with two simple baselines Finetune and Independent. While Finetune learns tasks one-by-one without any constraints, Indepen" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.892, + 0.236 + ], + "angle": 0, + "content": "
MethodCIFAR100-10 (10 tasks)Slong (100 tasks)
APM(M)APM(M)
Finetune0.1800.6070.2000.607
Independent0.8206.0700.57060.705
PN0.82112.422--
Learn to Grow0.7911.926--
MNTDP-S0.7506.1000.68039.658
MNTDP-D0.8306.0750.75025.508
PAR0.8531.4000.77313.475
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.247, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Table 4. Compared with existing parameter allocation methods based on dynamic model, PAR obtains better average performance (AP) with fewer model parameters \\((\\mathcal{M},\\mathrm{M} = 1\\mathrm{e}6)\\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.306, + 0.892, + 0.453 + ], + "angle": 0, + "content": "
#Allocation StrategyReg.APM(M)T(h)
FixedFSCS
1✓0.8552.2361.0
2✓0.8612.2771.9
3✓✓0.8701.9001.2
4✓0.7750.2441.0
5✓✓0.8341.1291.0
6✓✓0.8561.2101.4
PAR✓✓✓0.8531.4001.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.464, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Table 5. Ablation studies of components in PAR on the CIFAR100-10. The Fixed represents using a fixed cell from DARTs. The FS and CS represent fine-grained and coarse-grained search space respectively. The Reg. represents the parameter regularization. The PAR makes the trade off among average performance (AP), parameter overhead \\((\\mathcal{M})\\), and time overhead (T)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.714 + ], + "angle": 0, + "content": "dent builds a model for each task independently. Then, we present parameter regularization methods, including EWC [16], LwF [21], and MAS [1], and replay methods, including ER [6], GPM [35], and A-GEM [5]. Meanwhile, we compare parameter allocation methods with the static model, including HAT [36] and CAT [15], and parameter allocation methods with the dynamic model, including PN [34], Learn to Grow [20], SG-F [24], MNTDP [42], LMC [29] with PAR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Metrics Denote the performance of the model on \\( T_{j} \\) after learning task \\( T_{i} \\) as \\( r_{i,j} \\) where \\( j \\leq i \\). Suppose the current task is \\( T_{t} \\), the average performance (AP) and average forgetting (AF) are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.815, + 0.892, + 0.858 + ], + "angle": 0, + "content": "\\[\nA P = \\frac {1}{t} \\sum_ {j = 1} ^ {t} r _ {t, j}, A F = \\frac {1}{t} \\sum_ {j = 1} ^ {t} r _ {t, j} - r _ {j, j}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "To evaluate the parameter overhead, we denote the total number of the model as \\(\\mathcal{M}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7781" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.268, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.201, + 0.255, + 0.214 + ], + "angle": 0, + "content": "(a) Average performance (AP)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.092, + 0.473, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.201, + 0.422, + 0.213 + ], + "angle": 0, + "content": "(b) Model size" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.092, + 0.681, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.201, + 0.62, + 0.214 + ], + "angle": 0, + "content": "(c) Experts" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.092, + 0.887, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.201, + 0.822, + 0.213 + ], + "angle": 0, + "content": "(d) Time" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Ablation studies of \\(\\alpha\\) (Alpha) when \\(\\beta = 0.75\\). The larger \\(\\alpha\\) encourages the model to reuse experts, so the model size and number of experts decreases with the increase of \\(\\alpha\\). The \\(\\alpha\\) is independent of the architecture search, so the learning time is stable relative to it." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.271, + 0.268, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.38, + 0.255, + 0.393 + ], + "angle": 0, + "content": "(a) Average performance (AP)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.271, + 0.473, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.38, + 0.422, + 0.392 + ], + "angle": 0, + "content": "(b) Model size" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.271, + 0.68, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.38, + 0.613, + 0.392 + ], + "angle": 0, + "content": "(c) Cells" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.271, + 0.888, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.38, + 0.822, + 0.392 + ], + "angle": 0, + "content": "(d) Time" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.894, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Ablation studies of \\(\\beta\\) (Beta) when \\(\\alpha = 0.5\\). The \\(\\beta\\) has little effect on the average performance and size of model, and conversely has a great effect on the learning time because the determine the frequency of architecture search. The larger \\(\\beta\\) encourages the model to reuse architectures of existing experts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Implementation details We implement PAR by the PyTorch and open the source code2. We set the number of cells of each expert in PAR as 4, and set the \\(\\alpha\\) to 0.5 and the \\(\\beta\\) to 0.75 for all tasks. We adopt the SGD optimizer whose initial learning rate is 0.01 and anneals following a cosine schedule. We also set the momentum of SGD to 0.9 and search weight decay from [0.0003, 0.001, 0.003, 0.01] according to the validation performance. The results are averaged across three runs with different random seeds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.471, + 0.762 + ], + "angle": 0, + "content": "Without special instructions, following MNTDP [42], we adopt a light version of ResNet18 [13] with multiple output heads as the backbone for baselines. For the performance on CTrL, we report the results from MNTDP [42] and LMC [29]. For the performance on mixed CIFAR100 and F-CelebA, we report the results from CAT [15]. For the performance on classical benchmarks, we: report the results of RPSnet [31], InstAParam [7], and BNS [30] from origin papers; adopt the implementation of [4] for memory-based methods; and adopt our implementation for the others." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.774, + 0.321, + 0.79 + ], + "angle": 0, + "content": "4.2. Comparison with baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.471, + 0.875 + ], + "angle": 0, + "content": "At first, we evaluate PAR on benchmarks with mixed similar and dissimilar tasks. As shown in Tab. 1, PAR outperforms baselines on all six streams in CTrL. One reason is that PAR allows knowledge transfer among related tasks while preventing interference among tasks not related. For" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.658, + 0.894, + 0.701 + ], + "angle": 0, + "content": "Figure 5. Visualization of task groups and experts on CIFAR100-Coarse. The PAR groups semantically related tasks into the same group. We distinguish the different groups by different colors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.883 + ], + "angle": 0, + "content": "example, performance on stream \\( S^{+} \\) shows that PAR can update the knowledge of the expert for the task \\( t_{1}^{-} \\) by the task \\( t_{1}^{+} \\) with the same distribution and more samples. Performance on streams \\( S^{-} \\) and \\( S^{\\mathrm{out}} \\) shows that our method can transfer knowledge from experts of tasks related to the new task. Another reason is that a task-tailored architecture helps each expert perform better. Moreover, performance on stream \\( S^{\\mathrm{long}} \\) shows that PAR is scalable for the long task sequence. Similarly, results in Tab. 2 show that PAR outperforms parameter allocation methods with static models." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Further, we evaluate the PAR on classical benchmarks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.282, + 0.901 + ], + "angle": 0, + "content": "2https://github.com/WenjinW/PAR" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7782" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.255, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.198, + 0.186, + 0.21 + ], + "angle": 0, + "content": "(a) \\(S^{-}\\)" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.091, + 0.457, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.198, + 0.395, + 0.21 + ], + "angle": 0, + "content": "(b) \\(S^{+}\\)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.214, + 0.255, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.321, + 0.19, + 0.333 + ], + "angle": 0, + "content": "(c) \\(S^{\\mathrm{in}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.214, + 0.457, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.321, + 0.398, + 0.333 + ], + "angle": 0, + "content": "(d) \\(S^{\\mathrm{out}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Figure 6. The heat-maps of task distances of four streams in CTrL benchmark. The PAR can recognize the related tasks in \\( S^{-} \\), \\( S^{+} \\), and \\( S^{\\mathrm{out}} \\) exactly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.469, + 0.474 + ], + "angle": 0, + "content": "Results in Tab. 3 show that PAR outperforms parameter regularization methods, memory-based methods, and parameter allocation methods with dynamic model on the CIFAR100-10 and MiniImageNet-20." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.469, + 0.565 + ], + "angle": 0, + "content": "At last, we analyze the parameter overhead3 of PAR on CIFAR100-10 and \\( S^{\\mathrm{long}} \\). Compared with baselines, the PAR obtains better average performance with fewer model parameters. The reason is that the allocated experts in PAR have compact architecture and are parameter efficient. Performance on the stream \\( S^{\\mathrm{long}} \\) shows that PAR is scalable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.242, + 0.589 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.809 + ], + "angle": 0, + "content": "We analyze the effects of components in PAR and the results are listed in Tab. 5. First, we evaluate the impact of hierarchical architecture search on parameter allocation. Compared with using a fixed cell from DARTs [22] (#1), searching cells from fine-grained space can improve model performance (#2). Combined with coarse-grained space, searching from the hierarchical space can further improve performance while reducing the time and parameter overhead (#3). Then, we find that parameter regularization alone is time efficient, but its performance is low (#4). Finally, by selecting an appropriate strategy from parameter allocation and regularization based on the learning difficulty of each new task, PAR makes a trade off among average performance, parameter overhead, and time overhead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.855 + ], + "angle": 0, + "content": "We analyze the impact of important hyper-parameters \\(\\alpha\\) and \\(\\beta\\) in PAR on the CIFAR100-10. The \\(\\alpha\\) determines the reuse experts during the learning and the larger \\(\\alpha\\) encour" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "ages the model to reuse experts. Results in Fig. 3 show that the model size and number of experts decreases with the increase of \\(\\alpha\\). The \\(\\beta\\) determines the frequency of architecture search and the larger \\(\\beta\\) encourages the model to reuse architectures of existing experts. Results in Fig. 4 show that the learning time decreases with the increase of \\(\\beta\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.192, + 0.637, + 0.206 + ], + "angle": 0, + "content": "4.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.38 + ], + "angle": 0, + "content": "To analyze the task distance in PAR, we construct a new benchmark CIFAR100-coarse by dividing CIFAR100 into 20 tasks based on the coarse-grained labels. The illustration in Fig. 5 shows that the PAR can obtain reasonable task groups. For example, the PAR adaptively puts tasks about animals into the same groups, such as the Aquatic mammals, Large omnivores and herbivores, Fish, and so on. The Food containers and Household electrical devices are divided into the same group since they both contain cylindrical and circular objects. It also finds the relation between Non-insect invertebrates and Insects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Moreover, we visualize the heat maps of task distances obtained by PAR on four streams of CTrL in Fig. 6. The distance between the first and the last task on streams \\( S^{-} \\), \\( S^{+} \\) and \\( S^{\\mathrm{out}} \\) are small, which is consistent with the fact that they all come from the same data distribution. Instead, the distribution of the last task on stream \\( S^{\\mathrm{in}} \\) is perturbed from the distribution of the first task, so their distance is large. The above results show that PAR can obtain reasonable task distance and relatedness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.619, + 0.545 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.723 + ], + "angle": 0, + "content": "In this paper, we propose a new method named Parameter Allocation & Regularization (PAR). It allows the model to adaptively select an appropriate strategy from parameter allocation and regularization for each task based on the task learning difficulty. Experimental results on multiple benchmarks demonstrate that PAR is scalable and significantly reduces the model redundancy while improving performance. Moreover, qualitative analysis shows that PAR can obtain reasonable task relatedness. Our method is flexible, and in the future, we will introduce more relatedness estimation, regularization, and allocation strategies into PAR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.736, + 0.659, + 0.752 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.866 + ], + "angle": 0, + "content": "This work was supported by the NSFC projects (No. 62072399, No. U19B2042), Chinese Knowledge Center for Engineering Sciences and Technology, MoE Engineering Research Center of Digital Library, China Research Centre on Data and Knowledge for Engineering Sciences and Technology, and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.9 + ], + "angle": 0, + "content": "3We ignore the parameter overhead of the extractor (11.18M) in divergence estimation, which is fixed and will not become a bottleneck with the increase of tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7783" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory Aware Synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert Gate: Lifelong Learning With a Network of Experts. In CVPR, pages 3366-3375, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[3] Ali Ayub and Alan Wagner. EEC: Learning to Encode and Regenerate Images for Continual Learning. In ICLR 2021, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.468, + 0.315 + ], + "angle": 0, + "content": "[4] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark Experience for General Continual Learning: A Strong, Simple Baseline. In NeurIPS 2020, page 11, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient Lifelong Learning with A-GEM. In ICLR 2019, Sept. 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.468, + 0.428 + ], + "angle": 0, + "content": "[6] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. On Tiny Episodic Memories in Continual Learning. arXiv:1902.10486 [cs, stat], June 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.468, + 0.485 + ], + "angle": 0, + "content": "[7] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating Forgetting in Online Continual Learning via Instance-Aware Parameterization. In NeurIPS 2020, page 12, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.468, + 0.557 + ], + "angle": 0, + "content": "[8] Zhi-Qi Cheng, Xiao Wu, Siyu Huang, Jun-Xiu Li, Alexander G Hauptmann, and Qiang Peng. Learning to transfer: Generalizable attribute learning with multitask neural model search. In Proceedings of the 26th ACM international conference on Multimedia, pages 90-98, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[9] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, pages 1-1, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning Without Memorizing. In CVPR, pages 5138-5146, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.699 + ], + "angle": 0, + "content": "[11] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural Architecture Search: A Survey. JMLR, page 21, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.468, + 0.729 + ], + "angle": 0, + "content": "[12] Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4):8, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, June 2016. IEEE. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[14] Siyu Huang, Xi Li, Zhi-Qi Cheng, Zhongfei Zhang, and Alexander Hauptmann. Gnas: A greedy neural architecture search method for multi-attribute learning. In Proceedings of the 26th ACM international conference on Multimedia, pages 2049–2057, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Zixuan Ke, Bing Liu, and Xingchang Huang. Continual Learning of a Mixed Sequence of Similar and Dissimilar" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Tasks. In NeurIPS 2020, volume 33, pages 18493-18504. Curran Associates, Inc., 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[18] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[19] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming Catastrophic Forgetting by Incremental Moment Matching. In NeurlPS 2017, volume 30, pages 4652-4662, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[20] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting. In ICML, page 10, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Zhizhong Li and Derek Hoiem. Learning without Forgetting. IEEE transactions on pattern analysis and machine intelligence, page 14, 2017. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[22] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: DIFFERENTIABLE ARCHITECTURE SEARCH. In ICLR 2019, page 13, 2018. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[23] Huihui Liu, Yiding Yang, and Xinchao Wang. Overcoming Catastrophic Forgetting in Graph Neural Networks. In AAAI 2021, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[24] Jorge A. Mendez and Eric Eaton. Lifelong Learning of Compositional Structures. In ICLR 2021, Sept. 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[25] Qier Meng and Satoh Shin'ichi. ADINet: Attribute Driven Incremental Network for Retinal Image Classification. In CVPR 2020, pages 4033-4042, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[26] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual Learning with Filter Atom Swapping. In International Conference on Learning Representations, Sept. 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[27] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. VARIATIONAL CONTINUAL LEARNING. In ICLR, page 18, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[28] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Johannichen, and Moin Nabi. Learning to Remember: A Synaptic Plasticity Driven Framework for Continual Learning. In CVPR, pages 11321-11329, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[29] Oleksiy Ostapenko, Pau Rodriguez, Massimo Caccia, and Laurent Charlin. Continual Learning via Local Module Composition. In NeurIPS 2021, May 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[30] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. BNS: Building Network Structures Dynamically for Continual Learning. In NeurIPS 2021, May 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random Path Selection for Continual Learning. In H. Wallach, H. Larochelle, A." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7784" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Beygelzimer, F. d\\text{text}quotingingle Alché-Buc, E. Fox, and R. Garnett, editors, NeurIPS 2019, pages 12669-12679. Curran Associates, Inc., 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental Classifier and Representation Learning. In CVPR, pages 2001–2010, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.25 + ], + "angle": 0, + "content": "[33] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and Xin Wang. A Comprehensive Survey of Neural Architecture Search: Challenges and Solutions. ACM Computing Surveys, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[34] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive Neural Networks. arXiv:1606.04671 [cs], June 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[35] Gobinda Saha and Kaushik Roy. Gradient Projection Memory for Continual Learning. In ICLR 2021, Sept. 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[36] Joan Serrà, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming Catastrophic Forgetting with Hard Attention to the Task. In ICML, page 10, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.469, + 0.452 + ], + "angle": 0, + "content": "[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual Learning with Deep Generative Replay. In NeurIPS 2017, pages 2990-2999. Curran Associates, Inc., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[38] Pravendra Singh, Pratik Mazumder, Piyush Rai, and Vinay P. Namboodiri. Rectification-Based Knowledge Retention for Continual Learning. In CVPR 2021, pages 15282-15291, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[39] Pravendra Singh, Vinay Kumar Verma, Pratik Mazumder, Lawrence Carin, and Piyush Rai. Calibrating CNNs for Lifelong Learning. In NeurIPS 2020, volume 33, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.557, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[40] Binh Tang and David S. Matteson. Graph-Based Continual Learning. In ICLR 2021, Sept. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is All you Need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, NeurIPS 2017, pages 5998-6008. Curran Associates, Inc., 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[42] Tom Veniat, Ludovic Denoyer, and MarcAurelio Ranzato. Efficient Continual Learning with Modular Networks and Task-Driven Priors. In ICLR 2021, Sept. 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.797 + ], + "angle": 0, + "content": "[43] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[44] Qing Wang, Sanjeev R. Kulkarni, and Sergio Verdu. Divergence Estimation for Multidimensional Densities Via \\(\\$k\\)-Nearest-Neighbor Distances. IEEE Transactions on Information Theory, 55(5):2392–2405, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: An Alternative Approach to Efficient Ensemble and Lifelong Learning. In ICLR, Sept. 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[46] Ju Xu and Zhanxing Zhu. Reinforced Continual Learning. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, NeurIPS 2018, pages 899–908. Curran Associates, Inc., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[47] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically Expandable Representation for Class Incremental Learning. In CVPR 2021, pages 3014-3023, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[48] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial Distribution Learning for Effective Neural Architecture Search. In ICCV, pages 1304-1313, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[49] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning Transferable Architectures for Scalable Image Recognition. In CVPR 2018, pages 8697-8710, Salt Lake City, UT, 2018. IEEE. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7785" + } + ] +] \ No newline at end of file diff --git a/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_origin.pdf b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a560785ed0a2db5c07ea386480c37c4280ddffc5 --- /dev/null +++ b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/a54864ea-e9cc-490e-a81a-988ea47d70e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f0fbee16338399167d4bee1c086337283f0d0459a88ac55a9ed32deb06aab79 +size 1427507 diff --git a/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/full.md b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..85752289a9759df4e89205872d90ef1a44a70462 --- /dev/null +++ b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/full.md @@ -0,0 +1,338 @@ +# Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning + +Wenjin Wang, Yunqing Hu, Qianglong Chen, Yin Zhang* Zhejiang University, Hangzhou, China + +{wangwenjin,yunqinghu,chenqianglong,zhangyin98}@zju.edu.cn + +# Abstract + +Parameter regularization or allocation methods are effective in overcoming catastrophic forgetting in lifelong learning. However, they solve all tasks in a sequence uniformly and ignore the differences in the learning difficulty of different tasks. So parameter regularization methods face significant forgetting when learning a new task very different from learned tasks, and parameter allocation methods face unnecessary parameter overhead when learning simple tasks. In this paper, we propose the Parameter Allocation & Regularization (PAR), which adaptively select an appropriate strategy for each task from parameter allocation and regularization based on its learning difficulty. A task is easy for a model that has learned tasks related to it and vice versa. We propose a divergence estimation method based on the Nearest-Prototype distance to measure the task relatedness using only features of the new task. Moreover, we propose a time-efficient relatedness-aware sampling-based architecture search strategy to reduce the parameter overhead for allocation. Experimental results on multiple benchmarks demonstrate that, compared with SOTAs, our method is scalable and significantly reduces the model's redundancy while improving the model's performance. Further qualitative analysis indicates that PAR obtains reasonable task-relatedness. + +# 1. Introduction + +Recently, the lifelong learning [9] ability of neural networks, i.e., learning continuously from a continuous sequence of tasks, has been extensively studied. It is natural for human beings to constantly learn and accumulate knowledge from tasks and then use it to facilitate future learning. However, classical models [13, 18, 41] suffer catastrophic forgetting [12], i.e., the model's performance on learned tasks deteriorates rapidly after learning a new one. + +To overcome the catastrophic forgetting, many param- + +![](images/6f152f92326c8d478cc0b5bd471900e9e6c642ee2c3c219b8946dc471d4c411e.jpg) +Figure 1. PAR adaptively selects a strategy from regularization and allocation to handle each task according to its learning difficulty. The difficulty depends not only on the task itself, but also on whether the model has previously learned tasks related to it. + +eter regularization or allocation methods have been proposed. Parameter regularization methods [10, 16, 19, 21, 23, 25, 27] alleviate forgetting by adding a regularization term to the loss function and perform well when the new task does not differ much from learned tasks. Parameter allocation methods based on static models [7, 15, 31, 36] and dynamic models [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] allocate different parameters to different tasks and can adapt to new tasks quite different from learned tasks. However, the above methods solve all tasks in a sequence uniformly, and ignore the differences of learning difficulty of different tasks. This leads to the significant forgetting in parameter regularization methods when learning a new task which is quite different from learned tasks, and also leads to unnecessary parameter cost in parameter allocation methods when learning some simple tasks. + +In this paper, we propose a difficulty-aware method Parameter Allocation & Regularization (PAR). As shown + +in Fig. 1, we assume that the learning difficulty of a task in continual learning depends not only on the task itself, but also on the accumulated knowledge in the model. A new task is easy to adapt for a model if it has learned related tasks before and vice versa. Based on the assumption, the PAR adaptively adopts parameter allocation for difficult tasks and parameter regularization for easy tasks. Specifically, the PAR divides tasks into task groups and assigns each group a dedicated expert model. Given a new task, the PAR measures the relatedness between it and existing task groups at first. If the new task is related to one of the existing groups, it is easy for the corresponding expert. The PAR adds the task to the related group and learns it by the expert via the parameter regularization. Otherwise, the new task is difficult for all existing experts, and the PAR assigns it to a new task group and allocates a new expert to learn it. + +There are two challenges in this work: the measurement of relatedness and the parameter explosion associated with parameter allocation. For the first one, we try to measure the relatedness by the KL divergence between feature distributions of tasks. However, the KL divergence is intractable and needs to be estimated since the feature distributions of tasks are usually unknown. In addition, the constraint of lifelong learning that only data of the current task are available exacerbates the difficulty of estimation. To solve above problems, inspired by the divergence estimation based on k-NN distance [44], we propose the divergence estimation method based on prototype distance, which only depends on the data of the current task. For the second one, we try to reduce parameter overhead per expert by searching compact architecture for it. However, the low time and memory efficiency is an obstacle to applying architecture search for a sequence of tasks in lifelong learning. To improve the efficiency of architecture search, we propose a relatedness-aware sampling-based hierarchical search. The main contributions of this work are as follows: + +- We propose a lifelong learning framework named Parameter Allocation & Regularization (PAR), which selects an appropriate strategy from parameter allocation and regularization for each task based on the learning difficulty. The difficulty depends on whether the model has learned related tasks before. +- We propose a divergence estimation method based on prototype distance to measure the distance between the new task and previous learned tasks with only data of the new task. Meanwhile, we propose a relatedness-aware sampling-based architecture search to reduce the parameter overhead of parameter allocation. +- Experimental results on CTrL, Mixed CIFAR100 and F-CelebA, CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20 demonstrate that PAR is scalable and significantly reduces the model redundancy + +while improving the model performance. Exhaustive ablation studies show the effectiveness of components in PAR and the visualizations show the reasonability of task distance in PAR. + +# 2. Related Work + +# 2.1. Lifelong Learning + +Many methods have been proposed to overcome catastrophic forgetting. Replay methods try to replay samples of previous tasks when learning a new task from an episodic memory [32,40] or a generative memory [3,28,37]. Parameter regularization methods, including the prior-focused regularization [16, 19, 23, 27] and the data-focused regularization [10, 21, 25], try to alleviate forgetting by introducing a regularization term in the loss function of the new task. Parameter allocation methods based on the static model [7, 15, 31, 36] and the dynamic model [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] overcome catastrophic forgetting by allocating different parameters to different tasks. + +Methods with relatedness. Several methods also consider the utility of task relatedness [2, 15]. Expert Gate [2] assigns dedicated experts and auto-encoders for tasks and calculates the task relatedness by reconstruction error of auto-encoders. The task relatedness is used to transfer knowledge from the most related previous task. CAT [15] defines the task similarity by the positive knowledge transfer. It focuses on selectively transferring the knowledge from similar previous tasks and dealing with forgetting between dissimilar tasks by hard attention. In this paper, we propose a divergence estimation method based on prototype distance to calculate the task relatedness used to measure the task learning difficulty. + +Methods with NAS. Static model based methods [7, 31] try to search a sub-model for each task with neural architecture search (NAS). Dynamic model based methods [20, 42, 46] adopt NAS to select an appropriate model expansion strategy for each task and face the high GPU memory, parameter, and time overhead. Instead, in this paper, we propose a relatedness-aware hierarchical cell-based architecture search to search compact architecture for each expert with a lower GPU memory and time overhead. [8, 14] adopt NAS on multi-task learning. + +# 2.2. Cell-based NAS + +Neural architecture search (NAS) [11, 33] aims to search for efficient neural network architectures from a pre-defined search space in a data-driven manner. To reduce the size of the search space, cell-based NAS methods [22, 48, 49] try to search for a cell architecture from a pre-defined cell search space, where a cell is a tiny convolutional network mapping an $H \times W \times F$ tensor to another $H' \times W' \times F'$ . The final model consists of a pre-defined number of stacked + +cells. The cell in NAS is similar to the residual block in the residual network (ResNet), but its architecture is more complex and is a directed acyclic graph (DAG). The operations in the search space are usually parameter-efficient. NAS methods usually produce more compact architectures than hand-crafted designs. + +# 3. Method + +We focus on the task-incremental scenario of lifelong learning. Specifically, the model learns a sequence of tasks $\mathcal{T} = \{T_1,\dots ,T_t,\dots ,T_N\}$ one by one and the task id of sample is available during both training and inference. Each task $T_{t}$ has a training dataset, $D_{train}^{t} = \{(x_{i}^{t},y_{i}^{t});i = 1,\ldots ,n_{train}^{t}\}$ , where $y_{i}^{t}$ is the true label and $n_{train}^{t}$ is the number of training examples. Similarly, we denote the validation and test dataset of task $T_{t}$ as $D_{valid}^{t}$ and $D_{test}^{t}$ . + +As shown in Fig. 2, the PAR divides previously learned tasks into task groups $\mathcal{G}$ and assigns each group $\mathcal{G}_i$ a dedicated expert model $E_i$ . For a new task, PAR calculates the distances, which reflect the relatedness, between it and existing task groups at first. Then, it adopts an appropriate learning strategy from parameter allocation and regularization for the new task based on the distances. + +# 3.1. Task Distance via Nearest-Prototype Distance + +The PAR calculates the distance between the new task and each existing task at first, then it averages the distances of tasks in the same group as the distance between the new task and the group. + +We calculate the distance between two tasks by the KL divergence between their distributions. However, the KL divergence is intractable because the true distributions of tasks are usually agnostic. Instead, we estimate the KL divergence by the features of data. To enhance estimation stability, we introduce an extra pre-trained extractor1 to generate a robust feature $X_{i}^{t}$ for each image $x_{i}^{t}$ in task $T_{t}$ . The pre-trained extractor is only used for divergence estimation and does not affect the learning of tasks. The number of extra parameters introduced by the model is a constant and does not affect the scalability of our method. + +$k$ -Nearest Neighbor Distance Given the features of distributions, a classical divergence estimation method for multidimensional densities is the $k$ -Nearest Neighbor ( $k$ -NN) distance [44]. Suppose the feature distributions of the new task $T_{i}$ and an existing task $T_{j}$ are $q_{i}$ and $q_{j}$ respectively, and the features $X^{i} = \{X_{1}^{i},\ldots ,X_{n}^{i}\}$ and $X^{j} = \{X_{1}^{j},\ldots ,X_{m}^{j}\}$ are drawn independently from distributions $q_{i}$ and $q_{j}$ , where $n$ and $m$ are the numbers of samples. The + +$k$ -NN distance of feature $X_{l}^{i}$ in $X^{i}$ and $X^{j}$ are denoted by $\nu_{k}(l)$ and $\rho_{k}(l)$ respectively. Specifically, the $\rho_{k}(l)$ is the Euclidean distance from $X_{l}^{i}$ to its $k$ -NN in $\{X_{z}^{i}\}_{z\neq l}$ . Similarly, the $\nu_{k}(l)$ is the distance from $X_{l}^{i}$ to its $k$ -NN in $X^{j}$ . Then, the KL divergence between distributions $q_{i}$ and $q_{j}$ are estimated as follows: + +$$ +K L \left(q _ {i} \mid \mid q _ {j}\right) \approx \widehat {K L} \left(q _ {i} \mid \mid q _ {j}\right) = \frac {d}{n} \sum_ {l = 1} ^ {n} \log \frac {\nu_ {k} (l)}{\rho_ {k} (l)} + \log \frac {m}{n - 1}, \tag {1} +$$ + +where $d$ is the feature dimension. + +The $k$ -NN distance is asymmetric and this is consistent with the asymmetry of the KL divergence. The motivation behind Eq. (1) is that the $k$ -NN distance is the radius of a $d$ -dimensional open Euclidean ball used for $k$ -NN density estimation. Readers can refer to [44] for more details such as convergence analysis. + +Nearest-Prototype Distance However, the calculation of the $k$ -NN distance involves features of two tasks $T_{i}$ and $T_{j}$ and violates the constraint of lifelong learning that only data of the current task $T_{i}$ are available. To overcome this problem, we propose the Nearest-Prototype distance to replace the $k$ -NN distance. Specifically, suppose the set of classes in task $T^{i}$ is $C^i$ , for each class $c \in C^i$ , we maintain a prototype feature $U_{c}^{i}$ , which is the mean of features of samples belonging to the class $c$ . Then, the Nearest-Prototype distance of $X_{l}^{i}$ to $X^{i}$ is defined as follows: + +$$ +\rho (l) = \left\| X _ {l} ^ {i} - U _ {c (l)} ^ {i} \right\|, \tag {2} +$$ + +where $||\cdot ||$ is the Euclidean distance and $c(l)$ is the class of $X_{l}^{i}$ . Similarly, we denote the set of classes in task $T_{j}$ by $C^j$ and the prototype features of $X^{j}$ for each class $c\in C^j$ by $U_{c}^{j}$ . Then, the Nearest-Prototype distance of $X_{l}^{i}$ to $X^{j}$ is denoted as + +$$ +\nu (l) = \min _ {c} \| X _ {l} ^ {i} - U _ {c} ^ {j} \| _ {2}, c \in C ^ {j}. \tag {3} +$$ + +Then, by replacing the $k$ -NN distances $\nu_{k}(l)$ and $\rho_{k}(l)$ in Eq. (1) by the Nearest-Prototype distances $\nu(l)$ and $\rho(l)$ , the KL divergence estimated by the Nearest-Prototype distance is as follows: + +$$ +K L \left(q _ {i} \mid \mid q _ {j}\right) \approx \widetilde {K L} \left(q _ {i} \mid \mid q _ {j}\right) = \sum_ {c \in C ^ {i}} \frac {1}{n _ {c}} \sum_ {l = 1} ^ {n _ {c}} \log \frac {\nu (l)}{\rho (l)}, \tag {4} +$$ + +where constant terms are omitted because we are only concerned with the relative magnitude of KL divergence. The Eqs. (2) to (4) only involve features of the new task $T_{i}$ and prototypes of the existing task $T_{j}$ , which satisfies the constraint of lifelong learning. The storage cost of prototypes is small and negligible. + +The motivation behind Eq. (4) is intuitive. The Nearest-Prototype distance $\rho(l)$ reflects the relationship between the + +![](images/1affc81cb172d184e085eb37dc70dfa95b6a6392ebb0dccf79b88d75139e4776.jpg) +Figure 2. The architecture of PAR. PAR divides previously learned tasks into task groups and each group has a dedicated expert model. Given a new task $T_{5}$ , PAR calculates the distance between it and existing task groups to measure its learning difficulty. Then, PAR learns the new task by parameter regularization if it is easy to learn. Otherwise, PAR learns the new task by parameter allocation and search compact architecture for each expert by a relatedness-aware hierarchical architecture search. + +sample $X_{l}^{i}$ in task $T_{i}$ and its class prototype in feature space. The Nearest-Prototype distance $\nu (l)$ reflects the relationship between the sample $X_{l}^{i}$ and class prototypes of existing task $T_{j}$ in feature space. If the value of $\rho (l)$ and $\nu (l)$ are equal for each $X_{l}^{i}$ , the class prototypes of two tasks are close. In this case, the distribution of the two tasks are similar, and the estimated KL divergence by Eq. (4) is 0. + +Adaptive Learning Strategy Given the distance between the new task $T_{i}$ and each existing tasks by KL divergence, we denote the distance $s_{i,g}^{\prime}$ between the task $T_{i}$ and a task group $\mathcal{G}_g$ by the average of distances of tasks in the group as follows: + +$$ +s _ {i, g} ^ {\prime} = \frac {1}{| \mathcal {G} _ {g} |} \sum_ {j ^ {\prime}} ^ {| \mathcal {G} _ {g} |} K L (q _ {i} | | q _ {j ^ {\prime}}), \tag {5} +$$ + +where $q_{i}$ and $q_{j'}$ represent feature distributions of task $T_{i}$ and the $j'$ -th task in group $\mathcal{G}_g$ . The $s_{i,g}' \in [0, \infty)$ since the range of the KL divergence is $[0, \infty)$ . However, we try to use the distance to reflect the relatedness, which is usually measured by a value between 0 and 1. So we map the $s_{i,g}'$ into $[0, 1]$ by a monotone increasing function as follows: + +$$ +s _ {i, g} = \min \left(s _ {i, g} ^ {\prime}, 1 - e ^ {- 2 \times s _ {i, g} ^ {\prime}}\right), \tag {6} +$$ + +where $e$ is the Euler's number. + +Suppose the smallest distance between the new task $T_{t}$ and existing groups is $s_{t,g^{*}}$ and the corresponding task group is $\mathcal{G}_{g^{*}}$ , the PAR selects the learning strategy by comparing the $s_{t,g^{*}}$ with a hyper-parameter $\alpha$ . If $s_{t,g^{*}} \leq \alpha$ , + +the new task is added to this related group $\mathcal{G}_{g^*}$ and learned by parameter regularization. Otherwise, the new task is assigned to a new group and learned by parameter allocation because no existing task group is related to it. + +# 3.2. Parameter Regularization + +A new task $T_{t}$ is easy for the expert model $E_{g^{*}}$ if it is related to the group $G_{g^{*}}$ . PAR reuses the expert $E_{g^{*}}$ to learn this task by parameter regularization. Inspired by LwF [21], we adopt a parameter regularization method based on knowledge distillation. Specifically, the loss function consists of a training loss $\mathcal{L}_{\mathrm{new}}$ and a distillation loss $\mathcal{L}_{\mathrm{old}}$ . The training loss encourages expert $E_{g^{*}}$ to adapt to the new task $T_{t}$ and is the cross-entropy for classification as follows: + +$$ +\mathcal {L} _ {\text {n e w}} = - \sum_ {(x, y) \in D _ {t r a i n} ^ {t}} \log \left(p _ {E _ {g ^ {*}}} (y | x)\right). \tag {7} +$$ + +The distillation loss encourages the expert $E_{g^*}$ to maintain performance on previous tasks in the group $G_{g^*}$ . To calculate it, we record the logits $\mathbf{y}^j$ of the output head of each previous task $T_j$ for each sample $x$ of task $T_t$ . The distillation loss is as follows: + +$$ +\mathcal {L} _ {\text {o l d}} = - \sum_ {(x, y) \in D _ {t r a i n} ^ {t}} \sum_ {T _ {j} \in \mathcal {G} _ {g ^ {*}}} \mathbf {y} ^ {j} \cdot \log \left(\mathbf {p} _ {\mathbf {E} _ {\mathbf {g} ^ {*}}} (\mathbf {x}) ^ {j}\right), \tag {8} +$$ + +where $\mathbf{y}^j$ and $\mathbf{p}_{\mathbf{E}_{\mathbf{g}^*}}(\mathbf{x})^j$ are vectors with lengths equal to the number of categories of the previous task $T_{j}$ . The total loss is + +$$ +\mathcal {L} _ {\mathrm {P R}} = \mathcal {L} _ {\text {n e w}} + \lambda \mathcal {L} _ {\text {o l d}}, \tag {9} +$$ + +
MethodS-S+SinSoutSplSlong
APAFAPAFAPAFAPAFAPAFAPAF
Finetune0.180-0.3300.240-0.2300.180-0.3100.150-0.3700.210-0.3000.200-0.400
(1)EWC0.540-0.0100.370-0.0400.430-0.1200.510-0.0300.300-0.1700.270-0.300
(2)ER0.410-0.1300.430-0.0700.380-0.1700.540-0.0700.450-0.080--
(3)HAT*0.570-0.0100.5700.0000.580-0.0100.6000.0000.5800.0000.240-0.100
(4)Independent0.5600.0000.5700.0000.5700.0000.6100.0000.5900.0000.5700.000
PNN0.6200.0000.5200.0000.5700.0000.6200.0000.5400.000--
SG-F0.6360.0000.6150.0000.6550.0000.6410.0000.6200.000--
MNTDP-S0.6300.0000.5600.0000.5700.0000.6400.0000.5500.0000.6800.000
MNTDP-D0.6700.0000.6100.0000.6000.0000.6800.0000.6200.0000.7500.000
LMC0.6660.0000.601-0.0140.6950.0000.667-0.0100.616-0.0350.639-
PAR(our)0.7060.0000.670-0.0170.6990.0000.6940.0000.7000.0000.773-0.016
+ +Table 1. The average performance (AP) and average forgetting (AF) on the six streams $(S^{-}, S^{+}, S^{\mathrm{in}}, S^{\mathrm{out}}, S^{\mathrm{pl}}, S^{\mathrm{long}})$ in the CTRL benchmark. We compare the PAR with four types of baselines: (1) parameter regularization, (2) replay, (3) parameter allocation with static model, and (4) parameter allocation with dynamic model. The * corresponds to model using the AlexNet backbone. + +where $\lambda$ is a hyper-parameter to balance training and distillation loss. Note that our method is without memory and does not require storing samples for previous tasks. + +However, the expert $E_{g^{*}}$ may over-fit the new task whose sample size is far less than tasks in the group $\mathcal{G}_{g^{*}}$ , even though the new task is related to the group. This leads to interference with previously accumulated knowledge in the expert $E_{g^{*}}$ . To avoid the above problem, PAR records the maximum sample size of tasks in each task group. Suppose the maximum sample size in the group $\mathcal{G}_{g^{*}}$ is $Q$ , PAR freezes the parameters of expert $E_{g^{*}}$ except for the task-specific classification head during the learning if the sample size of the new task is less than 10 percent of $Q$ . By transferring the existing knowledge in expert $E_{g^{*}}$ , only optimizing the classification header is sufficient to adapt to the new task because the new task is related to the group $\mathcal{G}_{g^{*}}$ . + +# 3.3. Parameter Allocation + +If no existing groups are related to the new task $T_{t}$ , PAR assigns it to a new task group $\mathcal{G}_{M + 1}$ with a new expert $E_{M + 1}$ , where $M$ is the number of existing groups. We adopt the cross-entropy loss for the task $T_{t}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {P A}} = - \sum_ {(x, y) \in D _ {t r a i n} ^ {t}} \log \left(p _ {E _ {M + 1}} (y | x)\right). \tag {10} +$$ + +The number of experts and parameters in PAR is proportional to the number of task groups, mitigating the growth of parameters. To further reduce the parameter overhead of each expert, we adopt cell-based NAS (refer to Sec. 2.2 for details) to search for a compact architecture for it. + +Each expert in PAR is stacked with multiple cells and the search for expert architecture is equivalent to the search + +for the cell architecture. Since the time overhead of NAS becomes unbearable as the number of tasks increases, we propose a relatedness-aware sampling-based architecture search strategy to improve efficiency. + +As shown in Fig. 2, we construct a hierarchical search space. The coarse-grained search space contains cells used by existing experts and an unknown cell whose architecture will come from the fine-grained search space. Following common practice [22, 48], the fine-grained search space is a directed acyclic graph (DAG) with seven nodes (two input nodes $i_1$ , $i_2$ , an ordered sequence of intermediate nodes $n_1, n_2, n_3, n_4$ , and an output node). The input nodes represent the outputs of the previous two layers. The output is concatenated from all intermediate nodes. Each intermediate node has a directed candidate edge from each of its predecessors. Each candidate edge is associated with six parameter-efficient candidate operations. + +To search a cell for the new task, we introduce a hyperparameter $\beta$ . If $s_{t,g^*} \leq \beta$ , PAR reuses the cell of expert $E_{g^*}$ for the task $T_t$ . A task distance greater than $\alpha$ and less than $\beta$ indicates that the new task is not related enough to share the same expert with tasks in the group $\mathcal{G}_{g^*}$ but can use the same architecture. If $s_{t,g^*} > \beta$ , PAR assigns the unknown cell to the new expert and adopts a sampling-based NAS method named MDL [48] to determine its architecture. Due to limited space, we leave the details of the MDL in the supplementary. + +# 4. Experiments + +# 4.1. Experimental Settings + +Benchmarks We adopt two benchmarks the CTrL [42] and the Mixed CIFAR100 and F-CelebA [15], which con + +
MethodFinetuneHATCATPAR
AP0.61550.61780.68310.7122
+ +Table 2. We compare the average performance (AP) of PAR with two parameter allocation methods based on static model on the mixed CIFAR100 and F-CelebA benchmark. + +
#MethodCIFAR100-10MiniImageNet-20
AP(%)AF(%)AP(%)AF(%)
Finetune0.280-0.5600.378-0.314
1EWC0.608-0.0610.574-0.144
MAS0.579-0.0500.368-0.238
LwF0.655-0.0880.589-0.169
2GPM*0.725-0.604-
A-GEM0.677-0.1200.524-0.152
3Independent0.8200.0000.7870.000
PN0.8210.0000.7810.000
Learn to Grow0.7910.000--
MNTDP-S0.750-0.000--
MNTDP-D0.830-0.000--
PAR0.853-0.0200.816-0.022
+ +tain mixed similar and dissimilar tasks. CTrL [42] includes 6 streams of image classification tasks: $S^{-}$ is used to evaluate the ability of direct transfer, $S^{+}$ is used to evaluate the ability of knowledge update, $S^{\mathrm{in}}$ and $S^{\mathrm{out}}$ are used to evaluate the transfer to similar input and output distributions respectively, $S^{\mathrm{pl}}$ is used to evaluate the plasticity, $S^{\mathrm{long}}$ consists of 100 tasks and is used to evaluate the scalability. Similarly, Mixed CIFAR100 and F-CelebA [15] including mixed 10 similar tasks from F-CelebA and 10 dissimilar tasks from CIFAR100 [17]. + +Further, we adopt classical task incremental learning benchmarks: CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20. CIFAR10-5 is constructed by dividing CIFAR10 [17] into 5 tasks and each task has 2 classes. Similarly, CIFAR100-10 and CIFAR100-20 are constructed by dividing CIFAR100 [17] into 10 10-classification tasks and 20 5-classification tasks respectively. MiniImageNet-20 is constructed by dividing MiniImageNet [43] into 20 tasks and each task has 5 classes. We leave the results on the CIFAR10-5 and CIFAR100-20 in the supplementary. + +Baselines Firstly, we compare our method with two simple baselines Finetune and Independent. While Finetune learns tasks one-by-one without any constraints, Indepen + +Table 3. PAR outperforms (1) parameter regularization methods, (2) replay methods, and (3) parameter allocation methods with dynamic model on classical benchmarks. The * corresponds to model using the AlexNet backbone. + +
MethodCIFAR100-10 (10 tasks)Slong (100 tasks)
APM(M)APM(M)
Finetune0.1800.6070.2000.607
Independent0.8206.0700.57060.705
PN0.82112.422--
Learn to Grow0.7911.926--
MNTDP-S0.7506.1000.68039.658
MNTDP-D0.8306.0750.75025.508
PAR0.8531.4000.77313.475
+ +Table 4. Compared with existing parameter allocation methods based on dynamic model, PAR obtains better average performance (AP) with fewer model parameters $(\mathcal{M},\mathrm{M} = 1\mathrm{e}6)$ . + +
#Allocation StrategyReg.APM(M)T(h)
FixedFSCS
1✓0.8552.2361.0
2✓0.8612.2771.9
3✓✓0.8701.9001.2
4✓0.7750.2441.0
5✓✓0.8341.1291.0
6✓✓0.8561.2101.4
PAR✓✓✓0.8531.4001.1
+ +Table 5. Ablation studies of components in PAR on the CIFAR100-10. The Fixed represents using a fixed cell from DARTs. The FS and CS represent fine-grained and coarse-grained search space respectively. The Reg. represents the parameter regularization. The PAR makes the trade off among average performance (AP), parameter overhead $(\mathcal{M})$ , and time overhead (T). + +dent builds a model for each task independently. Then, we present parameter regularization methods, including EWC [16], LwF [21], and MAS [1], and replay methods, including ER [6], GPM [35], and A-GEM [5]. Meanwhile, we compare parameter allocation methods with the static model, including HAT [36] and CAT [15], and parameter allocation methods with the dynamic model, including PN [34], Learn to Grow [20], SG-F [24], MNTDP [42], LMC [29] with PAR. + +Metrics Denote the performance of the model on $T_{j}$ after learning task $T_{i}$ as $r_{i,j}$ where $j \leq i$ . Suppose the current task is $T_{t}$ , the average performance (AP) and average forgetting (AF) are as follows: + +$$ +A P = \frac {1}{t} \sum_ {j = 1} ^ {t} r _ {t, j}, A F = \frac {1}{t} \sum_ {j = 1} ^ {t} r _ {t, j} - r _ {j, j}. \tag {11} +$$ + +To evaluate the parameter overhead, we denote the total number of the model as $\mathcal{M}$ . + +![](images/040567f9d5e1e703148d1429295db47a15968074fcc410cc820dcee3759ecfd2.jpg) +(a) Average performance (AP) + +![](images/6d214bb5aaf067b6624392499f3fa2982f8bdef87d8426ec71467b7f05aa3e6e.jpg) +(b) Model size + +![](images/31cef153349764ef8b39347a82832b851562e6a51052629f1366598bba3a4d68.jpg) +(c) Experts + +![](images/ee57c82820b029cb732756bbcdb6220014afc0890f516f0db7fffd338864bb5a.jpg) +(d) Time + +![](images/9800dc49f02fcace30a9696a4513be5b30b09b0c48b0efdec577c1ba42f0fe92.jpg) +Figure 3. Ablation studies of $\alpha$ (Alpha) when $\beta = 0.75$ . The larger $\alpha$ encourages the model to reuse experts, so the model size and number of experts decreases with the increase of $\alpha$ . The $\alpha$ is independent of the architecture search, so the learning time is stable relative to it. +(a) Average performance (AP) + +![](images/7324a2243eb8ed2174afa8bad13bc536947a1bb43c325dd7d848d35d0f9dd386.jpg) +(b) Model size + +![](images/2cb749a5a7ce995f6432259086e299f5c8577250f2d1d098aca5f27dd3f68b42.jpg) +(c) Cells + +![](images/4a26c713d8c8198f1bfd5bd95b3d2f94dbd27e708e517591f72c9ebc5abab1cc.jpg) +(d) Time + +Implementation details We implement PAR by the PyTorch and open the source code2. We set the number of cells of each expert in PAR as 4, and set the $\alpha$ to 0.5 and the $\beta$ to 0.75 for all tasks. We adopt the SGD optimizer whose initial learning rate is 0.01 and anneals following a cosine schedule. We also set the momentum of SGD to 0.9 and search weight decay from [0.0003, 0.001, 0.003, 0.01] according to the validation performance. The results are averaged across three runs with different random seeds. + +Without special instructions, following MNTDP [42], we adopt a light version of ResNet18 [13] with multiple output heads as the backbone for baselines. For the performance on CTrL, we report the results from MNTDP [42] and LMC [29]. For the performance on mixed CIFAR100 and F-CelebA, we report the results from CAT [15]. For the performance on classical benchmarks, we: report the results of RPSnet [31], InstAParam [7], and BNS [30] from origin papers; adopt the implementation of [4] for memory-based methods; and adopt our implementation for the others. + +# 4.2. Comparison with baselines + +At first, we evaluate PAR on benchmarks with mixed similar and dissimilar tasks. As shown in Tab. 1, PAR outperforms baselines on all six streams in CTrL. One reason is that PAR allows knowledge transfer among related tasks while preventing interference among tasks not related. For + +![](images/417121c303f687f8f5eb121e5652cbd5eeb90d98864a2c235259abb8e934696c.jpg) +Figure 4. Ablation studies of $\beta$ (Beta) when $\alpha = 0.5$ . The $\beta$ has little effect on the average performance and size of model, and conversely has a great effect on the learning time because the determine the frequency of architecture search. The larger $\beta$ encourages the model to reuse architectures of existing experts. +Figure 5. Visualization of task groups and experts on CIFAR100-Coarse. The PAR groups semantically related tasks into the same group. We distinguish the different groups by different colors. + +example, performance on stream $S^{+}$ shows that PAR can update the knowledge of the expert for the task $t_{1}^{-}$ by the task $t_{1}^{+}$ with the same distribution and more samples. Performance on streams $S^{-}$ and $S^{\mathrm{out}}$ shows that our method can transfer knowledge from experts of tasks related to the new task. Another reason is that a task-tailored architecture helps each expert perform better. Moreover, performance on stream $S^{\mathrm{long}}$ shows that PAR is scalable for the long task sequence. Similarly, results in Tab. 2 show that PAR outperforms parameter allocation methods with static models. + +Further, we evaluate the PAR on classical benchmarks. + +![](images/6ae64b8707c131b9cd33e82b2499d42459494137b5e9fd66f689f1b98a8cbba7.jpg) +(a) $S^{-}$ + +![](images/d95c2b3a609831f66a61b5432002cc9402da89955360381b4c247e47de08d3a8.jpg) +(b) $S^{+}$ + +![](images/92e542b4c5033f189d6b7fad7eef66a87df44b27e66463e207ce458c5a3d9e75.jpg) +(c) $S^{\mathrm{in}}$ + +![](images/261de798ab628eff71a386dcf1090847872525dbb24d7ff9bd75b77f61a8a211.jpg) +(d) $S^{\mathrm{out}}$ +Figure 6. The heat-maps of task distances of four streams in CTrL benchmark. The PAR can recognize the related tasks in $S^{-}$ , $S^{+}$ , and $S^{\mathrm{out}}$ exactly. + +Results in Tab. 3 show that PAR outperforms parameter regularization methods, memory-based methods, and parameter allocation methods with dynamic model on the CIFAR100-10 and MiniImageNet-20. + +At last, we analyze the parameter overhead3 of PAR on CIFAR100-10 and $S^{\mathrm{long}}$ . Compared with baselines, the PAR obtains better average performance with fewer model parameters. The reason is that the allocated experts in PAR have compact architecture and are parameter efficient. Performance on the stream $S^{\mathrm{long}}$ shows that PAR is scalable. + +# 4.3. Ablation Studies + +We analyze the effects of components in PAR and the results are listed in Tab. 5. First, we evaluate the impact of hierarchical architecture search on parameter allocation. Compared with using a fixed cell from DARTs [22] (#1), searching cells from fine-grained space can improve model performance (#2). Combined with coarse-grained space, searching from the hierarchical space can further improve performance while reducing the time and parameter overhead (#3). Then, we find that parameter regularization alone is time efficient, but its performance is low (#4). Finally, by selecting an appropriate strategy from parameter allocation and regularization based on the learning difficulty of each new task, PAR makes a trade off among average performance, parameter overhead, and time overhead. + +We analyze the impact of important hyper-parameters $\alpha$ and $\beta$ in PAR on the CIFAR100-10. The $\alpha$ determines the reuse experts during the learning and the larger $\alpha$ encour + +ages the model to reuse experts. Results in Fig. 3 show that the model size and number of experts decreases with the increase of $\alpha$ . The $\beta$ determines the frequency of architecture search and the larger $\beta$ encourages the model to reuse architectures of existing experts. Results in Fig. 4 show that the learning time decreases with the increase of $\beta$ . + +# 4.4. Visualization + +To analyze the task distance in PAR, we construct a new benchmark CIFAR100-coarse by dividing CIFAR100 into 20 tasks based on the coarse-grained labels. The illustration in Fig. 5 shows that the PAR can obtain reasonable task groups. For example, the PAR adaptively puts tasks about animals into the same groups, such as the Aquatic mammals, Large omnivores and herbivores, Fish, and so on. The Food containers and Household electrical devices are divided into the same group since they both contain cylindrical and circular objects. It also finds the relation between Non-insect invertebrates and Insects. + +Moreover, we visualize the heat maps of task distances obtained by PAR on four streams of CTrL in Fig. 6. The distance between the first and the last task on streams $S^{-}$ , $S^{+}$ and $S^{\mathrm{out}}$ are small, which is consistent with the fact that they all come from the same data distribution. Instead, the distribution of the last task on stream $S^{\mathrm{in}}$ is perturbed from the distribution of the first task, so their distance is large. The above results show that PAR can obtain reasonable task distance and relatedness. + +# 5. Conclusion + +In this paper, we propose a new method named Parameter Allocation & Regularization (PAR). It allows the model to adaptively select an appropriate strategy from parameter allocation and regularization for each task based on the task learning difficulty. Experimental results on multiple benchmarks demonstrate that PAR is scalable and significantly reduces the model redundancy while improving performance. Moreover, qualitative analysis shows that PAR can obtain reasonable task relatedness. Our method is flexible, and in the future, we will introduce more relatedness estimation, regularization, and allocation strategies into PAR. + +# Acknowledgments + +This work was supported by the NSFC projects (No. 62072399, No. U19B2042), Chinese Knowledge Center for Engineering Sciences and Technology, MoE Engineering Research Center of Digital Library, China Research Centre on Data and Knowledge for Engineering Sciences and Technology, and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory Aware Synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 6 +[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert Gate: Lifelong Learning With a Network of Experts. In CVPR, pages 3366-3375, 2017. 1, 2 +[3] Ali Ayub and Alan Wagner. EEC: Learning to Encode and Regenerate Images for Continual Learning. In ICLR 2021, 2020. 2 +[4] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark Experience for General Continual Learning: A Strong, Simple Baseline. In NeurIPS 2020, page 11, 2020. 7 +[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient Lifelong Learning with A-GEM. In ICLR 2019, Sept. 2018. 6 +[6] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. On Tiny Episodic Memories in Continual Learning. arXiv:1902.10486 [cs, stat], June 2019. 6 +[7] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating Forgetting in Online Continual Learning via Instance-Aware Parameterization. In NeurIPS 2020, page 12, 2020. 1, 2, 7 +[8] Zhi-Qi Cheng, Xiao Wu, Siyu Huang, Jun-Xiu Li, Alexander G Hauptmann, and Qiang Peng. Learning to transfer: Generalizable attribute learning with multitask neural model search. In Proceedings of the 26th ACM international conference on Multimedia, pages 90-98, 2018. 2 +[9] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, pages 1-1, 2021. 1 +[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning Without Memorizing. In CVPR, pages 5138-5146, 2019. 1, 2 +[11] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural Architecture Search: A Survey. JMLR, page 21, 2019. 2 +[12] Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4):8, 1999. 1 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, June 2016. IEEE. 1, 7 +[14] Siyu Huang, Xi Li, Zhi-Qi Cheng, Zhongfei Zhang, and Alexander Hauptmann. Gnas: A greedy neural architecture search method for multi-attribute learning. In Proceedings of the 26th ACM international conference on Multimedia, pages 2049–2057, 2018. 2 +[15] Zixuan Ke, Bing Liu, and Xingchang Huang. Continual Learning of a Mixed Sequence of Similar and Dissimilar + +Tasks. In NeurIPS 2020, volume 33, pages 18493-18504. Curran Associates, Inc., 2020. 1, 2, 5, 6, 7 +[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 1, 2, 6 +[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6 +[18] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 1 +[19] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming Catastrophic Forgetting by Incremental Moment Matching. In NeurlPS 2017, volume 30, pages 4652-4662, 2017. 1, 2 +[20] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting. In ICML, page 10, 2019. 1, 2, 6 +[21] Zhizhong Li and Derek Hoiem. Learning without Forgetting. IEEE transactions on pattern analysis and machine intelligence, page 14, 2017. 1, 2, 4, 6 +[22] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: DIFFERENTIABLE ARCHITECTURE SEARCH. In ICLR 2019, page 13, 2018. 2, 5, 8 +[23] Huihui Liu, Yiding Yang, and Xinchao Wang. Overcoming Catastrophic Forgetting in Graph Neural Networks. In AAAI 2021, 2020. 1, 2 +[24] Jorge A. Mendez and Eric Eaton. Lifelong Learning of Compositional Structures. In ICLR 2021, Sept. 2020. 1, 2, 6 +[25] Qier Meng and Satoh Shin'ichi. ADINet: Attribute Driven Incremental Network for Retinal Image Classification. In CVPR 2020, pages 4033-4042, 2020. 1, 2 +[26] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual Learning with Filter Atom Swapping. In International Conference on Learning Representations, Sept. 2021. 1, 2 +[27] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. VARIATIONAL CONTINUAL LEARNING. In ICLR, page 18, 2018. 1, 2 +[28] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Johannichen, and Moin Nabi. Learning to Remember: A Synaptic Plasticity Driven Framework for Continual Learning. In CVPR, pages 11321-11329, 2019. 2 +[29] Oleksiy Ostapenko, Pau Rodriguez, Massimo Caccia, and Laurent Charlin. Continual Learning via Local Module Composition. In NeurIPS 2021, May 2021. 1, 2, 6, 7 +[30] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. BNS: Building Network Structures Dynamically for Continual Learning. In NeurIPS 2021, May 2021. 1, 2, 7 +[31] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random Path Selection for Continual Learning. In H. Wallach, H. Larochelle, A. + +Beygelzimer, F. d\text{text}quotingingle Alché-Buc, E. Fox, and R. Garnett, editors, NeurIPS 2019, pages 12669-12679. Curran Associates, Inc., 2019. 1, 2, 7 +[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental Classifier and Representation Learning. In CVPR, pages 2001–2010, 2017. 2 +[33] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and Xin Wang. A Comprehensive Survey of Neural Architecture Search: Challenges and Solutions. ACM Computing Surveys, 2020. 2 +[34] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive Neural Networks. arXiv:1606.04671 [cs], June 2016. 1, 2, 6 +[35] Gobinda Saha and Kaushik Roy. Gradient Projection Memory for Continual Learning. In ICLR 2021, Sept. 2020. 6 +[36] Joan Serrà, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming Catastrophic Forgetting with Hard Attention to the Task. In ICML, page 10, 2018. 1, 2, 6 +[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual Learning with Deep Generative Replay. In NeurIPS 2017, pages 2990-2999. Curran Associates, Inc., 2017. 2 +[38] Pravendra Singh, Pratik Mazumder, Piyush Rai, and Vinay P. Namboodiri. Rectification-Based Knowledge Retention for Continual Learning. In CVPR 2021, pages 15282-15291, 2021. 1, 2 +[39] Pravendra Singh, Vinay Kumar Verma, Pratik Mazumder, Lawrence Carin, and Piyush Rai. Calibrating CNNs for Lifelong Learning. In NeurIPS 2020, volume 33, 2020. 1, 2 +[40] Binh Tang and David S. Matteson. Graph-Based Continual Learning. In ICLR 2021, Sept. 2020. 2 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is All you Need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, NeurIPS 2017, pages 5998-6008. Curran Associates, Inc., 2017. 1 +[42] Tom Veniat, Ludovic Denoyer, and MarcAurelio Ranzato. Efficient Continual Learning with Modular Networks and Task-Driven Priors. In ICLR 2021, Sept. 2020. 1, 2, 5, 6, 7 +[43] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 6 +[44] Qing Wang, Sanjeev R. Kulkarni, and Sergio Verdu. Divergence Estimation for Multidimensional Densities Via $\$k$ -Nearest-Neighbor Distances. IEEE Transactions on Information Theory, 55(5):2392–2405, 2009. 2, 3 +[45] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: An Alternative Approach to Efficient Ensemble and Lifelong Learning. In ICLR, Sept. 2019. 1, 2 + +[46] Ju Xu and Zhanxing Zhu. Reinforced Continual Learning. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, NeurIPS 2018, pages 899–908. Curran Associates, Inc., 2018. 2 +[47] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically Expandable Representation for Class Incremental Learning. In CVPR 2021, pages 3014-3023, 2021. 1, 2 +[48] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial Distribution Learning for Effective Neural Architecture Search. In ICCV, pages 1304-1313, 2019. 2, 5 +[49] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning Transferable Architectures for Scalable Image Recognition. In CVPR 2018, pages 8697-8710, Salt Lake City, UT, 2018. IEEE. 2 \ No newline at end of file diff --git a/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/images.zip b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e22c147b064eeb9d65ea4b1c92abc4302c6000e4 --- /dev/null +++ b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d258556bcbe1f6aff12a31692e87b2bbb692c372aabab2988ad9b4c22c3f3ce +size 601959 diff --git a/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/layout.json b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f56607d6577c1801d27510bd17616680013d9f6f --- /dev/null +++ b/2023/Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning/layout.json @@ -0,0 +1,11386 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 526, + 140 + ], + "type": "text", + "content": "Task Difficulty Aware Parameter Allocation & Regularization for Lifelong Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 161, + 436, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 161, + 436, + 190 + ], + "spans": [ + { + "bbox": [ + 159, + 161, + 436, + 190 + ], + "type": "text", + "content": "Wenjin Wang, Yunqing Hu, Qianglong Chen, Yin Zhang* Zhejiang University, Hangzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 191, + 452, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 191, + 452, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 191, + 452, + 204 + ], + "type": "text", + "content": "{wangwenjin,yunqinghu,chenqianglong,zhangyin98}@zju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 256, + 290, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 256, + 290, + 543 + ], + "spans": [ + { + "bbox": [ + 45, + 256, + 290, + 543 + ], + "type": "text", + "content": "Parameter regularization or allocation methods are effective in overcoming catastrophic forgetting in lifelong learning. However, they solve all tasks in a sequence uniformly and ignore the differences in the learning difficulty of different tasks. So parameter regularization methods face significant forgetting when learning a new task very different from learned tasks, and parameter allocation methods face unnecessary parameter overhead when learning simple tasks. In this paper, we propose the Parameter Allocation & Regularization (PAR), which adaptively select an appropriate strategy for each task from parameter allocation and regularization based on its learning difficulty. A task is easy for a model that has learned tasks related to it and vice versa. We propose a divergence estimation method based on the Nearest-Prototype distance to measure the task relatedness using only features of the new task. Moreover, we propose a time-efficient relatedness-aware sampling-based architecture search strategy to reduce the parameter overhead for allocation. Experimental results on multiple benchmarks demonstrate that, compared with SOTAs, our method is scalable and significantly reduces the model's redundancy while improving the model's performance. Further qualitative analysis indicates that PAR obtains reasonable task-relatedness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 586, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 586, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 45, + 586, + 287, + 682 + ], + "type": "text", + "content": "Recently, the lifelong learning [9] ability of neural networks, i.e., learning continuously from a continuous sequence of tasks, has been extensively studied. It is natural for human beings to constantly learn and accumulate knowledge from tasks and then use it to facilitate future learning. However, classical models [13, 18, 41] suffer catastrophic forgetting [12], i.e., the model's performance on learned tasks deteriorates rapidly after learning a new one." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 683, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 287, + 695 + ], + "type": "text", + "content": "To overcome the catastrophic forgetting, many param-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 232, + 548, + 414 + ], + "blocks": [ + { + "bbox": [ + 310, + 232, + 548, + 414 + ], + "lines": [ + { + "bbox": [ + 310, + 232, + 548, + 414 + ], + "spans": [ + { + "bbox": [ + 310, + 232, + 548, + 414 + ], + "type": "image", + "image_path": "6f152f92326c8d478cc0b5bd471900e9e6c642ee2c3c219b8946dc471d4c411e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 422, + 547, + 465 + ], + "lines": [ + { + "bbox": [ + 305, + 422, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 547, + 465 + ], + "type": "text", + "content": "Figure 1. PAR adaptively selects a strategy from regularization and allocation to handle each task according to its learning difficulty. The difficulty depends not only on the task itself, but also on whether the model has previously learned tasks related to it." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 686 + ], + "type": "text", + "content": "eter regularization or allocation methods have been proposed. Parameter regularization methods [10, 16, 19, 21, 23, 25, 27] alleviate forgetting by adding a regularization term to the loss function and perform well when the new task does not differ much from learned tasks. Parameter allocation methods based on static models [7, 15, 31, 36] and dynamic models [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] allocate different parameters to different tasks and can adapt to new tasks quite different from learned tasks. However, the above methods solve all tasks in a sequence uniformly, and ignore the differences of learning difficulty of different tasks. This leads to the significant forgetting in parameter regularization methods when learning a new task which is quite different from learned tasks, and also leads to unnecessary parameter cost in parameter allocation methods when learning some simple tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we propose a difficulty-aware method Parameter Allocation & Regularization (PAR). As shown" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 175, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 175, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 175, + 713 + ], + "type": "text", + "content": "*Corresponding author: Yin Zhang." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7776" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "in Fig. 1, we assume that the learning difficulty of a task in continual learning depends not only on the task itself, but also on the accumulated knowledge in the model. A new task is easy to adapt for a model if it has learned related tasks before and vice versa. Based on the assumption, the PAR adaptively adopts parameter allocation for difficult tasks and parameter regularization for easy tasks. Specifically, the PAR divides tasks into task groups and assigns each group a dedicated expert model. Given a new task, the PAR measures the relatedness between it and existing task groups at first. If the new task is related to one of the existing groups, it is easy for the corresponding expert. The PAR adds the task to the related group and learns it by the expert via the parameter regularization. Otherwise, the new task is difficult for all existing experts, and the PAR assigns it to a new task group and allocates a new expert to learn it." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 502 + ], + "type": "text", + "content": "There are two challenges in this work: the measurement of relatedness and the parameter explosion associated with parameter allocation. For the first one, we try to measure the relatedness by the KL divergence between feature distributions of tasks. However, the KL divergence is intractable and needs to be estimated since the feature distributions of tasks are usually unknown. In addition, the constraint of lifelong learning that only data of the current task are available exacerbates the difficulty of estimation. To solve above problems, inspired by the divergence estimation based on k-NN distance [44], we propose the divergence estimation method based on prototype distance, which only depends on the data of the current task. For the second one, we try to reduce parameter overhead per expert by searching compact architecture for it. However, the low time and memory efficiency is an obstacle to applying architecture search for a sequence of tasks in lifelong learning. To improve the efficiency of architecture search, we propose a relatedness-aware sampling-based hierarchical search. The main contributions of this work are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 508, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 508, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 508, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 59, + 508, + 287, + 579 + ], + "type": "text", + "content": "- We propose a lifelong learning framework named Parameter Allocation & Regularization (PAR), which selects an appropriate strategy from parameter allocation and regularization for each task based on the learning difficulty. The difficulty depends on whether the model has learned related tasks before." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 586, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 586, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 59, + 586, + 287, + 658 + ], + "type": "text", + "content": "- We propose a divergence estimation method based on prototype distance to measure the distance between the new task and previous learned tasks with only data of the new task. Meanwhile, we propose a relatedness-aware sampling-based architecture search to reduce the parameter overhead of parameter allocation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 287, + 714 + ], + "type": "text", + "content": "- Experimental results on CTrL, Mixed CIFAR100 and F-CelebA, CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20 demonstrate that PAR is scalable and significantly reduces the model redundancy" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 119 + ], + "type": "text", + "content": "while improving the model performance. Exhaustive ablation studies show the effectiveness of components in PAR and the visualizations show the reasonability of task distance in PAR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 132, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 132, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 132, + 392, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 152, + 413, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 413, + 165 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 413, + 165 + ], + "type": "text", + "content": "2.1. Lifelong Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 171, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 314 + ], + "type": "text", + "content": "Many methods have been proposed to overcome catastrophic forgetting. Replay methods try to replay samples of previous tasks when learning a new task from an episodic memory [32,40] or a generative memory [3,28,37]. Parameter regularization methods, including the prior-focused regularization [16, 19, 23, 27] and the data-focused regularization [10, 21, 25], try to alleviate forgetting by introducing a regularization term in the loss function of the new task. Parameter allocation methods based on the static model [7, 15, 31, 36] and the dynamic model [2, 20, 24, 26, 29, 30, 34, 38, 39, 42, 45, 47] overcome catastrophic forgetting by allocating different parameters to different tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 471 + ], + "type": "text", + "content": "Methods with relatedness. Several methods also consider the utility of task relatedness [2, 15]. Expert Gate [2] assigns dedicated experts and auto-encoders for tasks and calculates the task relatedness by reconstruction error of auto-encoders. The task relatedness is used to transfer knowledge from the most related previous task. CAT [15] defines the task similarity by the positive knowledge transfer. It focuses on selectively transferring the knowledge from similar previous tasks and dealing with forgetting between dissimilar tasks by hard attention. In this paper, we propose a divergence estimation method based on prototype distance to calculate the task relatedness used to measure the task learning difficulty." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 590 + ], + "type": "text", + "content": "Methods with NAS. Static model based methods [7, 31] try to search a sub-model for each task with neural architecture search (NAS). Dynamic model based methods [20, 42, 46] adopt NAS to select an appropriate model expansion strategy for each task and face the high GPU memory, parameter, and time overhead. Instead, in this paper, we propose a relatedness-aware hierarchical cell-based architecture search to search compact architecture for each expert with a lower GPU memory and time overhead. [8, 14] adopt NAS on multi-task learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 599, + 403, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 403, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 403, + 610 + ], + "type": "text", + "content": "2.2. Cell-based NAS" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Neural architecture search (NAS) [11, 33] aims to search for efficient neural network architectures from a pre-defined search space in a data-driven manner. To reduce the size of the search space, cell-based NAS methods [22, 48, 49] try to search for a cell architecture from a pre-defined cell search space, where a cell is a tiny convolutional network mapping an " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "H \\times W \\times F" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " tensor to another " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "H' \\times W' \\times F'" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". The final model consists of a pre-defined number of stacked" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7777" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "cells. The cell in NAS is similar to the residual block in the residual network (ResNet), but its architecture is more complex and is a directed acyclic graph (DAG). The operations in the search space are usually parameter-efficient. NAS methods usually produce more compact architectures than hand-crafted designs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 104, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 104, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 104, + 168 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": "We focus on the task-incremental scenario of lifelong learning. Specifically, the model learns a sequence of tasks " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{T_1,\\dots ,T_t,\\dots ,T_N\\}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": " one by one and the task id of sample is available during both training and inference. Each task " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": " has a training dataset, " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "D_{train}^{t} = \\{(x_{i}^{t},y_{i}^{t});i = 1,\\ldots ,n_{train}^{t}\\}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "y_{i}^{t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": " is the true label and " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "n_{train}^{t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": " is the number of training examples. Similarly, we denote the validation and test dataset of task " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "D_{valid}^{t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "inline_equation", + "content": "D_{test}^{t}" + }, + { + "bbox": [ + 46, + 175, + 287, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "text", + "content": "As shown in Fig. 2, the PAR divides previously learned tasks into task groups " + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "text", + "content": " and assigns each group " + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "text", + "content": " a dedicated expert model " + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 46, + 273, + 288, + 356 + ], + "type": "text", + "content": ". For a new task, PAR calculates the distances, which reflect the relatedness, between it and existing task groups at first. Then, it adopts an appropriate learning strategy from parameter allocation and regularization for the new task based on the distances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 364, + 284, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 284, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 284, + 378 + ], + "type": "text", + "content": "3.1. Task Distance via Nearest-Prototype Distance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 382, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 431 + ], + "type": "text", + "content": "The PAR calculates the distance between the new task and each existing task at first, then it averages the distances of tasks in the same group as the distance between the new task and the group." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "text", + "content": "We calculate the distance between two tasks by the KL divergence between their distributions. However, the KL divergence is intractable because the true distributions of tasks are usually agnostic. Instead, we estimate the KL divergence by the features of data. To enhance estimation stability, we introduce an extra pre-trained extractor1 to generate a robust feature " + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "inline_equation", + "content": "X_{i}^{t}" + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "text", + "content": " in task " + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 46, + 431, + 288, + 563 + ], + "type": "text", + "content": ". The pre-trained extractor is only used for divergence estimation and does not affect the learning of tasks. The number of extra parameters introduced by the model is a constant and does not affect the scalability of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": "-Nearest Neighbor Distance Given the features of distributions, a classical divergence estimation method for multidimensional densities is the " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": "-Nearest Neighbor (" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": "-NN) distance [44]. Suppose the feature distributions of the new task " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " and an existing task " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " respectively, and the features " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "X^{i} = \\{X_{1}^{i},\\ldots ,X_{n}^{i}\\}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "X^{j} = \\{X_{1}^{j},\\ldots ,X_{m}^{j}\\}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " are drawn independently from distributions " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 578, + 287, + 676 + ], + "type": "text", + "content": " are the numbers of samples. The" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "-NN distance of feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X^{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " are denoted by " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\nu_{k}(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\rho_{k}(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " respectively. Specifically, the " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\rho_{k}(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " is the Euclidean distance from " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to its " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "-NN in " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\{X_{z}^{i}\\}_{z\\neq l}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". Similarly, the " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\nu_{k}(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " is the distance from " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " to its " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "-NN in " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X^{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". Then, the KL divergence between distributions " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " are estimated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 148, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 148, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 545, + 190 + ], + "type": "interline_equation", + "content": "K L \\left(q _ {i} \\mid \\mid q _ {j}\\right) \\approx \\widehat {K L} \\left(q _ {i} \\mid \\mid q _ {j}\\right) = \\frac {d}{n} \\sum_ {l = 1} ^ {n} \\log \\frac {\\nu_ {k} (l)}{\\rho_ {k} (l)} + \\log \\frac {m}{n - 1}, \\tag {1}", + "image_path": "f8905477c1ef0e3a7f037250894cd9a538ab676b485e4d742c9b9bd83d56ff0f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 190, + 441, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 441, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 441, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 190, + 441, + 201 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 306, + 190, + 441, + 201 + ], + "type": "text", + "content": " is the feature dimension." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "-NN distance is asymmetric and this is consistent with the asymmetry of the KL divergence. The motivation behind Eq. (1) is that the " + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "-NN distance is the radius of a " + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "-dimensional open Euclidean ball used for " + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "-NN density estimation. Readers can refer to [44] for more details such as convergence analysis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": "Nearest-Prototype Distance However, the calculation of the " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": "-NN distance involves features of two tasks " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": " and violates the constraint of lifelong learning that only data of the current task " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": " are available. To overcome this problem, we propose the Nearest-Prototype distance to replace the " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": "-NN distance. Specifically, suppose the set of classes in task " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "T^{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": ", for each class " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "c \\in C^i" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": ", we maintain a prototype feature " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "U_{c}^{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": ", which is the mean of features of samples belonging to the class " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": ". Then, the Nearest-Prototype distance of " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 304, + 288, + 545, + 408 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 414, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 414, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 380, + 414, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\rho (l) = \\left\\| X _ {l} ^ {i} - U _ {c (l)} ^ {i} \\right\\|, \\tag {2}", + "image_path": "9a6ee0bbe949803b92cdf14973fcb61f50022fb6f83e6a93b55b454f7c8b67e5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "||\\cdot ||" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " is the Euclidean distance and " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "c(l)" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " is the class of " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ". Similarly, we denote the set of classes in task " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "C^j" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " and the prototype features of " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "X^{j}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "c\\in C^j" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "U_{c}^{j}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": ". Then, the Nearest-Prototype distance of " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "inline_equation", + "content": "X^{j}" + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": " is denoted as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 502, + 545, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 502, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 356, + 502, + 545, + 519 + ], + "type": "interline_equation", + "content": "\\nu (l) = \\min _ {c} \\| X _ {l} ^ {i} - U _ {c} ^ {j} \\| _ {2}, c \\in C ^ {j}. \\tag {3}", + "image_path": "cad000ab357833ef68ac69e9ce5392fa73a8e3dc5808a478c23ef5289a4d72a8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "content": "Then, by replacing the " + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "content": "-NN distances " + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\nu_{k}(l)" + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\rho_{k}(l)" + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "content": " in Eq. (1) by the Nearest-Prototype distances " + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\nu(l)" + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\rho(l)" + }, + { + "bbox": [ + 304, + 526, + 545, + 574 + ], + "type": "text", + "content": ", the KL divergence estimated by the Nearest-Prototype distance is as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 578, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 315, + 578, + 545, + 611 + ], + "type": "interline_equation", + "content": "K L \\left(q _ {i} \\mid \\mid q _ {j}\\right) \\approx \\widetilde {K L} \\left(q _ {i} \\mid \\mid q _ {j}\\right) = \\sum_ {c \\in C ^ {i}} \\frac {1}{n _ {c}} \\sum_ {l = 1} ^ {n _ {c}} \\log \\frac {\\nu (l)}{\\rho (l)}, \\tag {4}", + "image_path": "e67acfb606807af76faa4dd301451c6cac5fd6a1b6af95db42a18da0c36ecbb8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "where constant terms are omitted because we are only concerned with the relative magnitude of KL divergence. The Eqs. (2) to (4) only involve features of the new task " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " and prototypes of the existing task " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", which satisfies the constraint of lifelong learning. The storage cost of prototypes is small and negligible." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "The motivation behind Eq. (4) is intuitive. The Nearest-Prototype distance " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rho(l)" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " reflects the relationship between the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": "1We adopt a ResNet18 pre-trained on the ImageNet. It does not introduce unfairness because the generated features are not involved in model training and prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7778" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 280 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 280 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 280 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 280 + ], + "type": "image", + "image_path": "1affc81cb172d184e085eb37dc70dfa95b6a6392ebb0dccf79b88d75139e4776.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 287, + 547, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 547, + 332 + ], + "type": "text", + "content": "Figure 2. The architecture of PAR. PAR divides previously learned tasks into task groups and each group has a dedicated expert model. Given a new task " + }, + { + "bbox": [ + 46, + 287, + 547, + 332 + ], + "type": "inline_equation", + "content": "T_{5}" + }, + { + "bbox": [ + 46, + 287, + 547, + 332 + ], + "type": "text", + "content": ", PAR calculates the distance between it and existing task groups to measure its learning difficulty. Then, PAR learns the new task by parameter regularization if it is easy to learn. Otherwise, PAR learns the new task by parameter allocation and search compact architecture for each expert by a relatedness-aware hierarchical architecture search." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": "sample " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " in task " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " and its class prototype in feature space. The Nearest-Prototype distance " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\nu (l)" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " reflects the relationship between the sample " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " and class prototypes of existing task " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " in feature space. If the value of " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\rho (l)" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\nu (l)" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": " are equal for each " + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "inline_equation", + "content": "X_{l}^{i}" + }, + { + "bbox": [ + 46, + 353, + 289, + 437 + ], + "type": "text", + "content": ", the class prototypes of two tasks are close. In this case, the distribution of the two tasks are similar, and the estimated KL divergence by Eq. (4) is 0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "text", + "content": "Adaptive Learning Strategy Given the distance between the new task " + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "text", + "content": " and each existing tasks by KL divergence, we denote the distance " + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "inline_equation", + "content": "s_{i,g}^{\\prime}" + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "text", + "content": " between the task " + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "text", + "content": " and a task group " + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_g" + }, + { + "bbox": [ + 47, + 451, + 287, + 510 + ], + "type": "text", + "content": " by the average of distances of tasks in the group as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 514, + 287, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 514, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 107, + 514, + 287, + 550 + ], + "type": "interline_equation", + "content": "s _ {i, g} ^ {\\prime} = \\frac {1}{| \\mathcal {G} _ {g} |} \\sum_ {j ^ {\\prime}} ^ {| \\mathcal {G} _ {g} |} K L (q _ {i} | | q _ {j ^ {\\prime}}), \\tag {5}", + "image_path": "e2c502ca570dfa1f414a1964c3a13ac638526d8137158a5dcb8abab39738e347.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "q_{j'}" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": " represent feature distributions of task " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "j'" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": "-th task in group " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_g" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "s_{i,g}' \\in [0, \\infty)" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": " since the range of the KL divergence is " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "[0, \\infty)" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": ". However, we try to use the distance to reflect the relatedness, which is usually measured by a value between 0 and 1. So we map the " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "s_{i,g}'" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 46, + 555, + 287, + 628 + ], + "type": "text", + "content": " by a monotone increasing function as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 632, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 632, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 102, + 632, + 287, + 649 + ], + "type": "interline_equation", + "content": "s _ {i, g} = \\min \\left(s _ {i, g} ^ {\\prime}, 1 - e ^ {- 2 \\times s _ {i, g} ^ {\\prime}}\\right), \\tag {6}", + "image_path": "8b29996eca451e038359550074d9d7e2b15291a66a78b6d7b557f1a226ddd00a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 654, + 171, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 171, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 171, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 171, + 664 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 47, + 654, + 171, + 664 + ], + "type": "text", + "content": " is the Euler's number." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Suppose the smallest distance between the new task " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and existing groups is " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "s_{t,g^{*}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and the corresponding task group is " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{g^{*}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ", the PAR selects the learning strategy by comparing the " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "s_{t,g^{*}}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " with a hyper-parameter " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "s_{t,g^{*}} \\leq \\alpha" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 353, + 546, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 401 + ], + "type": "text", + "content": "the new task is added to this related group " + }, + { + "bbox": [ + 304, + 353, + 546, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{g^*}" + }, + { + "bbox": [ + 304, + 353, + 546, + 401 + ], + "type": "text", + "content": " and learned by parameter regularization. Otherwise, the new task is assigned to a new group and learned by parameter allocation because no existing task group is related to it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 407, + 451, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 451, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 451, + 420 + ], + "type": "text", + "content": "3.2. Parameter Regularization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": "A new task " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " is easy for the expert model " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " if it is related to the group " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "G_{g^{*}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": ". PAR reuses the expert " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " to learn this task by parameter regularization. Inspired by LwF [21], we adopt a parameter regularization method based on knowledge distillation. Specifically, the loss function consists of a training loss " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{new}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " and a distillation loss " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{old}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": ". The training loss encourages expert " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " to adapt to the new task " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " and is the cross-entropy for classification as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 346, + 529, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 529, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 346, + 529, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n e w}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\log \\left(p _ {E _ {g ^ {*}}} (y | x)\\right). \\tag {7}", + "image_path": "588a040096ae83a3cecc721d0f82f52f5831cd32db90b8388051cba46e56580e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": "The distillation loss encourages the expert " + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "inline_equation", + "content": "E_{g^*}" + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": " to maintain performance on previous tasks in the group " + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "inline_equation", + "content": "G_{g^*}" + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": ". To calculate it, we record the logits " + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^j" + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": " of the output head of each previous task " + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "inline_equation", + "content": "T_j" + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": " for each sample " + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": " of task " + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "inline_equation", + "content": "T_t" + }, + { + "bbox": [ + 304, + 563, + 545, + 623 + ], + "type": "text", + "content": ". The distillation loss is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 628, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 628, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 319, + 628, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o l d}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\sum_ {T _ {j} \\in \\mathcal {G} _ {g ^ {*}}} \\mathbf {y} ^ {j} \\cdot \\log \\left(\\mathbf {p} _ {\\mathbf {E} _ {\\mathbf {g} ^ {*}}} (\\mathbf {x}) ^ {j}\\right), \\tag {8}", + "image_path": "2e6250ddb0d3f7535605ca1dd1c5bc76f558beb3122b796666ba3648e85da529.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^j" + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{E}_{\\mathbf{g}^*}}(\\mathbf{x})^j" + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "content": " are vectors with lengths equal to the number of categories of the previous task " + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 665, + 545, + 700 + ], + "type": "text", + "content": ". The total loss is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 381, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P R}} = \\mathcal {L} _ {\\text {n e w}} + \\lambda \\mathcal {L} _ {\\text {o l d}}, \\tag {9}", + "image_path": "09c358cade831971d276b8b91c0f9af263adc47feb7cca53b20cac39ad467b5b.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7779" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 535, + 251 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 535, + 251 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 535, + 251 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 535, + 251 + ], + "type": "table", + "html": "
MethodS-S+SinSoutSplSlong
APAFAPAFAPAFAPAFAPAFAPAF
Finetune0.180-0.3300.240-0.2300.180-0.3100.150-0.3700.210-0.3000.200-0.400
(1)EWC0.540-0.0100.370-0.0400.430-0.1200.510-0.0300.300-0.1700.270-0.300
(2)ER0.410-0.1300.430-0.0700.380-0.1700.540-0.0700.450-0.080--
(3)HAT*0.570-0.0100.5700.0000.580-0.0100.6000.0000.5800.0000.240-0.100
(4)Independent0.5600.0000.5700.0000.5700.0000.6100.0000.5900.0000.5700.000
PNN0.6200.0000.5200.0000.5700.0000.6200.0000.5400.000--
SG-F0.6360.0000.6150.0000.6550.0000.6410.0000.6200.000--
MNTDP-S0.6300.0000.5600.0000.5700.0000.6400.0000.5500.0000.6800.000
MNTDP-D0.6700.0000.6100.0000.6000.0000.6800.0000.6200.0000.7500.000
LMC0.6660.0000.601-0.0140.6950.0000.667-0.0100.616-0.0350.639-
PAR(our)0.7060.0000.670-0.0170.6990.0000.6940.0000.7000.0000.773-0.016
", + "image_path": "69815aea38a7996d96be27346c1bd4f37dd163962951631a80949f1635df6472.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "text", + "content": "Table 1. The average performance (AP) and average forgetting (AF) on the six streams " + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "inline_equation", + "content": "(S^{-}, S^{+}, S^{\\mathrm{in}}, S^{\\mathrm{out}}, S^{\\mathrm{pl}}, S^{\\mathrm{long}})" + }, + { + "bbox": [ + 46, + 258, + 547, + 293 + ], + "type": "text", + "content": " in the CTRL benchmark. We compare the PAR with four types of baselines: (1) parameter regularization, (2) replay, (3) parameter allocation with static model, and (4) parameter allocation with dynamic model. The * corresponds to model using the AlexNet backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 314, + 288, + 350 + ], + "type": "text", + "content": " is a hyper-parameter to balance training and distillation loss. Note that our method is without memory and does not require storing samples for previous tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": "However, the expert " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": " may over-fit the new task whose sample size is far less than tasks in the group " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": ", even though the new task is related to the group. This leads to interference with previously accumulated knowledge in the expert " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": ". To avoid the above problem, PAR records the maximum sample size of tasks in each task group. Suppose the maximum sample size in the group " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": ", PAR freezes the parameters of expert " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": " except for the task-specific classification head during the learning if the sample size of the new task is less than 10 percent of " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": ". By transferring the existing knowledge in expert " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "E_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": ", only optimizing the classification header is sufficient to adapt to the new task because the new task is related to the group " + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{g^{*}}" + }, + { + "bbox": [ + 46, + 350, + 287, + 507 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 170, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 170, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 170, + 526 + ], + "type": "text", + "content": "3.3. Parameter Allocation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "content": "If no existing groups are related to the new task " + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "content": ", PAR assigns it to a new task group " + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{M + 1}" + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "content": " with a new expert " + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "inline_equation", + "content": "E_{M + 1}" + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "content": " is the number of existing groups. We adopt the cross-entropy loss for the task " + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 47, + 533, + 287, + 583 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 591, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 591, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 85, + 591, + 287, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P A}} = - \\sum_ {(x, y) \\in D _ {t r a i n} ^ {t}} \\log \\left(p _ {E _ {M + 1}} (y | x)\\right). \\tag {10}", + "image_path": "75bdf8e97b05f041e8caf091a4dcf8500359ddf21d8449f95a1b56f6da9eed7e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "type": "text", + "content": "The number of experts and parameters in PAR is proportional to the number of task groups, mitigating the growth of parameters. To further reduce the parameter overhead of each expert, we adopt cell-based NAS (refer to Sec. 2.2 for details) to search for a compact architecture for it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Each expert in PAR is stacked with multiple cells and the search for expert architecture is equivalent to the search" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 361 + ], + "type": "text", + "content": "for the cell architecture. Since the time overhead of NAS becomes unbearable as the number of tasks increases, we propose a relatedness-aware sampling-based architecture search strategy to improve efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "content": "As shown in Fig. 2, we construct a hierarchical search space. The coarse-grained search space contains cells used by existing experts and an unknown cell whose architecture will come from the fine-grained search space. Following common practice [22, 48], the fine-grained search space is a directed acyclic graph (DAG) with seven nodes (two input nodes " + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "inline_equation", + "content": "i_1" + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "inline_equation", + "content": "i_2" + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "content": ", an ordered sequence of intermediate nodes " + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "inline_equation", + "content": "n_1, n_2, n_3, n_4" + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "content": ", and an output node). The input nodes represent the outputs of the previous two layers. The output is concatenated from all intermediate nodes. Each intermediate node has a directed candidate edge from each of its predecessors. Each candidate edge is associated with six parameter-efficient candidate operations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": "To search a cell for the new task, we introduce a hyperparameter " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "s_{t,g^*} \\leq \\beta" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": ", PAR reuses the cell of expert " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "E_{g^*}" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": " for the task " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "T_t" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": ". A task distance greater than " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": " and less than " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": " indicates that the new task is not related enough to share the same expert with tasks in the group " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{g^*}" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": " but can use the same architecture. If " + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "inline_equation", + "content": "s_{t,g^*} > \\beta" + }, + { + "bbox": [ + 304, + 518, + 545, + 639 + ], + "type": "text", + "content": ", PAR assigns the unknown cell to the new expert and adopts a sampling-based NAS method named MDL [48] to determine its architecture. Due to limited space, we leave the details of the MDL in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 650, + 388, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 388, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 388, + 663 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 670, + 432, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 432, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 432, + 684 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Benchmarks We adopt two benchmarks the CTrL [42] and the Mixed CIFAR100 and F-CelebA [15], which con" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7780" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 273, + 107 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 273, + 107 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 273, + 107 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 273, + 107 + ], + "type": "table", + "html": "
MethodFinetuneHATCATPAR
AP0.61550.61780.68310.7122
", + "image_path": "ede08e4867f0d1ec8a14716198bd1007da5ea190a6ee60b388d7f9dab7f1cd02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 160, + 285, + 347 + ], + "blocks": [ + { + "bbox": [ + 47, + 115, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 47, + 115, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 287, + 148 + ], + "type": "text", + "content": "Table 2. We compare the average performance (AP) of PAR with two parameter allocation methods based on static model on the mixed CIFAR100 and F-CelebA benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 160, + 285, + 347 + ], + "lines": [ + { + "bbox": [ + 50, + 160, + 285, + 347 + ], + "spans": [ + { + "bbox": [ + 50, + 160, + 285, + 347 + ], + "type": "table", + "html": "
#MethodCIFAR100-10MiniImageNet-20
AP(%)AF(%)AP(%)AF(%)
Finetune0.280-0.5600.378-0.314
1EWC0.608-0.0610.574-0.144
MAS0.579-0.0500.368-0.238
LwF0.655-0.0880.589-0.169
2GPM*0.725-0.604-
A-GEM0.677-0.1200.524-0.152
3Independent0.8200.0000.7870.000
PN0.8210.0000.7810.000
Learn to Grow0.7910.000--
MNTDP-S0.750-0.000--
MNTDP-D0.830-0.000--
PAR0.853-0.0200.816-0.022
", + "image_path": "697856fa810b821e9133ab7bb062dfd3875e4723f60f7060f44c46ff475bf66b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": "tain mixed similar and dissimilar tasks. CTrL [42] includes 6 streams of image classification tasks: " + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": " is used to evaluate the ability of direct transfer, " + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": " is used to evaluate the ability of knowledge update, " + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{in}}" + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": " are used to evaluate the transfer to similar input and output distributions respectively, " + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": " is used to evaluate the plasticity, " + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{long}}" + }, + { + "bbox": [ + 46, + 421, + 286, + 540 + ], + "type": "text", + "content": " consists of 100 tasks and is used to evaluate the scalability. Similarly, Mixed CIFAR100 and F-CelebA [15] including mixed 10 similar tasks from F-CelebA and 10 dissimilar tasks from CIFAR100 [17]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 286, + 661 + ], + "type": "text", + "content": "Further, we adopt classical task incremental learning benchmarks: CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20. CIFAR10-5 is constructed by dividing CIFAR10 [17] into 5 tasks and each task has 2 classes. Similarly, CIFAR100-10 and CIFAR100-20 are constructed by dividing CIFAR100 [17] into 10 10-classification tasks and 20 5-classification tasks respectively. MiniImageNet-20 is constructed by dividing MiniImageNet [43] into 20 tasks and each task has 5 classes. We leave the results on the CIFAR10-5 and CIFAR100-20 in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 714 + ], + "type": "text", + "content": "Baselines Firstly, we compare our method with two simple baselines Finetune and Independent. While Finetune learns tasks one-by-one without any constraints, Indepen" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 46, + 355, + 287, + 400 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 400 + ], + "type": "text", + "content": "Table 3. PAR outperforms (1) parameter regularization methods, (2) replay methods, and (3) parameter allocation methods with dynamic model on classical benchmarks. The * corresponds to model using the AlexNet backbone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 186 + ], + "type": "table", + "html": "
MethodCIFAR100-10 (10 tasks)Slong (100 tasks)
APM(M)APM(M)
Finetune0.1800.6070.2000.607
Independent0.8206.0700.57060.705
PN0.82112.422--
Learn to Grow0.7911.926--
MNTDP-S0.7506.1000.68039.658
MNTDP-D0.8306.0750.75025.508
PAR0.8531.4000.77313.475
", + "image_path": "f1dd6b7c7fa74567790f869eb15b5db707bb73f594ae3a776db77e0e3f615ff5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 242, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 305, + 195, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 305, + 195, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 545, + 229 + ], + "type": "text", + "content": "Table 4. Compared with existing parameter allocation methods based on dynamic model, PAR obtains better average performance (AP) with fewer model parameters " + }, + { + "bbox": [ + 305, + 195, + 545, + 229 + ], + "type": "inline_equation", + "content": "(\\mathcal{M},\\mathrm{M} = 1\\mathrm{e}6)" + }, + { + "bbox": [ + 305, + 195, + 545, + 229 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 242, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 358 + ], + "type": "table", + "html": "
#Allocation StrategyReg.APM(M)T(h)
FixedFSCS
1✓0.8552.2361.0
2✓0.8612.2771.9
3✓✓0.8701.9001.2
4✓0.7750.2441.0
5✓✓0.8341.1291.0
6✓✓0.8561.2101.4
PAR✓✓✓0.8531.4001.1
", + "image_path": "af8c8a4be59c91c76bf8375ef852d06001d591c783c72610811ddf5ca142a711.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 367, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 545, + 434 + ], + "type": "text", + "content": "Table 5. Ablation studies of components in PAR on the CIFAR100-10. The Fixed represents using a fixed cell from DARTs. The FS and CS represent fine-grained and coarse-grained search space respectively. The Reg. represents the parameter regularization. The PAR makes the trade off among average performance (AP), parameter overhead " + }, + { + "bbox": [ + 305, + 367, + 545, + 434 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 305, + 367, + 545, + 434 + ], + "type": "text", + "content": ", and time overhead (T)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 456, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 565 + ], + "type": "text", + "content": "dent builds a model for each task independently. Then, we present parameter regularization methods, including EWC [16], LwF [21], and MAS [1], and replay methods, including ER [6], GPM [35], and A-GEM [5]. Meanwhile, we compare parameter allocation methods with the static model, including HAT [36] and CAT [15], and parameter allocation methods with the dynamic model, including PN [34], Learn to Grow [20], SG-F [24], MNTDP [42], LMC [29] with PAR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": "Metrics Denote the performance of the model on " + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": " after learning task " + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "inline_equation", + "content": "r_{i,j}" + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "inline_equation", + "content": "j \\leq i" + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": ". Suppose the current task is " + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 304, + 586, + 545, + 634 + ], + "type": "text", + "content": ", the average performance (AP) and average forgetting (AF) are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 645, + 545, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 645, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 339, + 645, + 545, + 679 + ], + "type": "interline_equation", + "content": "A P = \\frac {1}{t} \\sum_ {j = 1} ^ {t} r _ {t, j}, A F = \\frac {1}{t} \\sum_ {j = 1} ^ {t} r _ {t, j} - r _ {j, j}. \\tag {11}", + "image_path": "fbd0ac2662543f908ca544c454724ed381b1ef9b13c234356653edf2f77a353e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "To evaluate the parameter overhead, we denote the total number of the model as " + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 689, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7781" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 164, + 155 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 164, + 155 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 164, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 164, + 155 + ], + "type": "image", + "image_path": "040567f9d5e1e703148d1429295db47a15968074fcc410cc820dcee3759ecfd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 159, + 156, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 159, + 156, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 156, + 169 + ], + "type": "text", + "content": "(a) Average performance (AP)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 72, + 289, + 154 + ], + "blocks": [ + { + "bbox": [ + 177, + 72, + 289, + 154 + ], + "lines": [ + { + "bbox": [ + 177, + 72, + 289, + 154 + ], + "spans": [ + { + "bbox": [ + 177, + 72, + 289, + 154 + ], + "type": "image", + "image_path": "6d214bb5aaf067b6624392499f3fa2982f8bdef87d8426ec71467b7f05aa3e6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 159, + 258, + 168 + ], + "lines": [ + { + "bbox": [ + 209, + 159, + 258, + 168 + ], + "spans": [ + { + "bbox": [ + 209, + 159, + 258, + 168 + ], + "type": "text", + "content": "(b) Model size" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 416, + 155 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 416, + 155 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 416, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 416, + 155 + ], + "type": "image", + "image_path": "31cef153349764ef8b39347a82832b851562e6a51052629f1366598bba3a4d68.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 159, + 379, + 169 + ], + "lines": [ + { + "bbox": [ + 342, + 159, + 379, + 169 + ], + "spans": [ + { + "bbox": [ + 342, + 159, + 379, + 169 + ], + "type": "text", + "content": "(c) Experts" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 432, + 72, + 542, + 155 + ], + "blocks": [ + { + "bbox": [ + 432, + 72, + 542, + 155 + ], + "lines": [ + { + "bbox": [ + 432, + 72, + 542, + 155 + ], + "spans": [ + { + "bbox": [ + 432, + 72, + 542, + 155 + ], + "type": "image", + "image_path": "ee57c82820b029cb732756bbcdb6220014afc0890f516f0db7fffd338864bb5a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 159, + 503, + 168 + ], + "lines": [ + { + "bbox": [ + 473, + 159, + 503, + 168 + ], + "spans": [ + { + "bbox": [ + 473, + 159, + 503, + 168 + ], + "type": "text", + "content": "(d) Time" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 51, + 214, + 164, + 296 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": "Figure 3. Ablation studies of " + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": " (Alpha) when " + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\beta = 0.75" + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": ". The larger " + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": " encourages the model to reuse experts, so the model size and number of experts decreases with the increase of " + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 178, + 547, + 201 + ], + "type": "text", + "content": " is independent of the architecture search, so the learning time is stable relative to it." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 214, + 164, + 296 + ], + "lines": [ + { + "bbox": [ + 51, + 214, + 164, + 296 + ], + "spans": [ + { + "bbox": [ + 51, + 214, + 164, + 296 + ], + "type": "image", + "image_path": "9800dc49f02fcace30a9696a4513be5b30b09b0c48b0efdec577c1ba42f0fe92.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 300, + 156, + 311 + ], + "lines": [ + { + "bbox": [ + 56, + 300, + 156, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 156, + 311 + ], + "type": "text", + "content": "(a) Average performance (AP)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 178, + 214, + 289, + 296 + ], + "blocks": [ + { + "bbox": [ + 178, + 214, + 289, + 296 + ], + "lines": [ + { + "bbox": [ + 178, + 214, + 289, + 296 + ], + "spans": [ + { + "bbox": [ + 178, + 214, + 289, + 296 + ], + "type": "image", + "image_path": "7324a2243eb8ed2174afa8bad13bc536947a1bb43c325dd7d848d35d0f9dd386.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 300, + 258, + 310 + ], + "lines": [ + { + "bbox": [ + 209, + 300, + 258, + 310 + ], + "spans": [ + { + "bbox": [ + 209, + 300, + 258, + 310 + ], + "type": "text", + "content": "(b) Model size" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 214, + 416, + 296 + ], + "blocks": [ + { + "bbox": [ + 305, + 214, + 416, + 296 + ], + "lines": [ + { + "bbox": [ + 305, + 214, + 416, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 416, + 296 + ], + "type": "image", + "image_path": "2cb749a5a7ce995f6432259086e299f5c8577250f2d1d098aca5f27dd3f68b42.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 300, + 375, + 310 + ], + "lines": [ + { + "bbox": [ + 346, + 300, + 375, + 310 + ], + "spans": [ + { + "bbox": [ + 346, + 300, + 375, + 310 + ], + "type": "text", + "content": "(c) Cells" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 214, + 543, + 296 + ], + "blocks": [ + { + "bbox": [ + 432, + 214, + 543, + 296 + ], + "lines": [ + { + "bbox": [ + 432, + 214, + 543, + 296 + ], + "spans": [ + { + "bbox": [ + 432, + 214, + 543, + 296 + ], + "type": "image", + "image_path": "4a26c713d8c8198f1bfd5bd95b3d2f94dbd27e708e517591f72c9ebc5abab1cc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 300, + 503, + 310 + ], + "lines": [ + { + "bbox": [ + 473, + 300, + 503, + 310 + ], + "spans": [ + { + "bbox": [ + 473, + 300, + 503, + 310 + ], + "type": "text", + "content": "(d) Time" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": "Implementation details We implement PAR by the PyTorch and open the source code2. We set the number of cells of each expert in PAR as 4, and set the " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": " to 0.5 and the " + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 374, + 287, + 481 + ], + "type": "text", + "content": " to 0.75 for all tasks. We adopt the SGD optimizer whose initial learning rate is 0.01 and anneals following a cosine schedule. We also set the momentum of SGD to 0.9 and search weight decay from [0.0003, 0.001, 0.003, 0.01] according to the validation performance. The results are averaged across three runs with different random seeds." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 483, + 288, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 603 + ], + "type": "text", + "content": "Without special instructions, following MNTDP [42], we adopt a light version of ResNet18 [13] with multiple output heads as the backbone for baselines. For the performance on CTrL, we report the results from MNTDP [42] and LMC [29]. For the performance on mixed CIFAR100 and F-CelebA, we report the results from CAT [15]. For the performance on classical benchmarks, we: report the results of RPSnet [31], InstAParam [7], and BNS [30] from origin papers; adopt the implementation of [4] for memory-based methods; and adopt our implementation for the others." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 613, + 196, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 196, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 196, + 625 + ], + "type": "text", + "content": "4.2. Comparison with baselines" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 632, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 288, + 693 + ], + "type": "text", + "content": "At first, we evaluate PAR on benchmarks with mixed similar and dissimilar tasks. As shown in Tab. 1, PAR outperforms baselines on all six streams in CTrL. One reason is that PAR allows knowledge transfer among related tasks while preventing interference among tasks not related. For" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 308, + 376, + 545, + 511 + ], + "blocks": [ + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "text", + "content": "Figure 4. Ablation studies of " + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "text", + "content": " (Beta) when " + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "text", + "content": " has little effect on the average performance and size of model, and conversely has a great effect on the learning time because the determine the frequency of architecture search. The larger " + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 320, + 547, + 354 + ], + "type": "text", + "content": " encourages the model to reuse architectures of existing experts." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 376, + 545, + 511 + ], + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 511 + ], + "type": "image", + "image_path": "417121c303f687f8f5eb121e5652cbd5eeb90d98864a2c235259abb8e934696c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 521, + 547, + 555 + ], + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 555 + ], + "type": "text", + "content": "Figure 5. Visualization of task groups and experts on CIFAR100-Coarse. The PAR groups semantically related tasks into the same group. We distinguish the different groups by different colors." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": "example, performance on stream " + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": " shows that PAR can update the knowledge of the expert for the task " + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "inline_equation", + "content": "t_{1}^{-}" + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": " by the task " + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "inline_equation", + "content": "t_{1}^{+}" + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": " with the same distribution and more samples. Performance on streams " + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{out}}" + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": " shows that our method can transfer knowledge from experts of tasks related to the new task. Another reason is that a task-tailored architecture helps each expert perform better. Moreover, performance on stream " + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{long}}" + }, + { + "bbox": [ + 304, + 579, + 545, + 699 + ], + "type": "text", + "content": " shows that PAR is scalable for the long task sequence. Similarly, results in Tab. 2 show that PAR outperforms parameter allocation methods with static models." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Further, we evaluate the PAR on classical benchmarks." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 172, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 172, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 172, + 713 + ], + "type": "text", + "content": "2https://github.com/WenjinW/PAR" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7782" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 156, + 153 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 156, + 153 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 156, + 153 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 156, + 153 + ], + "type": "image", + "image_path": "6ae64b8707c131b9cd33e82b2499d42459494137b5e9fd66f689f1b98a8cbba7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 156, + 113, + 166 + ], + "lines": [ + { + "bbox": [ + 93, + 156, + 113, + 166 + ], + "spans": [ + { + "bbox": [ + 93, + 156, + 113, + 166 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 93, + 156, + 113, + 166 + ], + "type": "inline_equation", + "content": "S^{-}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 72, + 279, + 153 + ], + "blocks": [ + { + "bbox": [ + 181, + 72, + 279, + 153 + ], + "lines": [ + { + "bbox": [ + 181, + 72, + 279, + 153 + ], + "spans": [ + { + "bbox": [ + 181, + 72, + 279, + 153 + ], + "type": "image", + "image_path": "d95c2b3a609831f66a61b5432002cc9402da89955360381b4c247e47de08d3a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 156, + 241, + 166 + ], + "lines": [ + { + "bbox": [ + 218, + 156, + 241, + 166 + ], + "spans": [ + { + "bbox": [ + 218, + 156, + 241, + 166 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 218, + 156, + 241, + 166 + ], + "type": "inline_equation", + "content": "S^{+}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 169, + 156, + 251 + ], + "blocks": [ + { + "bbox": [ + 58, + 169, + 156, + 251 + ], + "lines": [ + { + "bbox": [ + 58, + 169, + 156, + 251 + ], + "spans": [ + { + "bbox": [ + 58, + 169, + 156, + 251 + ], + "type": "image", + "image_path": "92e542b4c5033f189d6b7fad7eef66a87df44b27e66463e207ce458c5a3d9e75.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 254, + 116, + 263 + ], + "lines": [ + { + "bbox": [ + 93, + 254, + 116, + 263 + ], + "spans": [ + { + "bbox": [ + 93, + 254, + 116, + 263 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 93, + 254, + 116, + 263 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{in}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 182, + 169, + 279, + 251 + ], + "blocks": [ + { + "bbox": [ + 182, + 169, + 279, + 251 + ], + "lines": [ + { + "bbox": [ + 182, + 169, + 279, + 251 + ], + "spans": [ + { + "bbox": [ + 182, + 169, + 279, + 251 + ], + "type": "image", + "image_path": "261de798ab628eff71a386dcf1090847872525dbb24d7ff9bd75b77f61a8a211.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 254, + 243, + 263 + ], + "lines": [ + { + "bbox": [ + 217, + 254, + 243, + 263 + ], + "spans": [ + { + "bbox": [ + 217, + 254, + 243, + 263 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 217, + 254, + 243, + 263 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{out}}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "text", + "content": "Figure 6. The heat-maps of task distances of four streams in CTrL benchmark. The PAR can recognize the related tasks in " + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{out}}" + }, + { + "bbox": [ + 46, + 273, + 287, + 307 + ], + "type": "text", + "content": " exactly." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 375 + ], + "type": "text", + "content": "Results in Tab. 3 show that PAR outperforms parameter regularization methods, memory-based methods, and parameter allocation methods with dynamic model on the CIFAR100-10 and MiniImageNet-20." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "type": "text", + "content": "At last, we analyze the parameter overhead3 of PAR on CIFAR100-10 and " + }, + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{long}}" + }, + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "type": "text", + "content": ". Compared with baselines, the PAR obtains better average performance with fewer model parameters. The reason is that the allocated experts in PAR have compact architecture and are parameter efficient. Performance on the stream " + }, + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{long}}" + }, + { + "bbox": [ + 46, + 376, + 287, + 447 + ], + "type": "text", + "content": " shows that PAR is scalable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 454, + 148, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 148, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 148, + 466 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 640 + ], + "type": "text", + "content": "We analyze the effects of components in PAR and the results are listed in Tab. 5. First, we evaluate the impact of hierarchical architecture search on parameter allocation. Compared with using a fixed cell from DARTs [22] (#1), searching cells from fine-grained space can improve model performance (#2). Combined with coarse-grained space, searching from the hierarchical space can further improve performance while reducing the time and parameter overhead (#3). Then, we find that parameter regularization alone is time efficient, but its performance is low (#4). Finally, by selecting an appropriate strategy from parameter allocation and regularization based on the learning difficulty of each new task, PAR makes a trade off among average performance, parameter overhead, and time overhead." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": "We analyze the impact of important hyper-parameters " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " in PAR on the CIFAR100-10. The " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " determines the reuse experts during the learning and the larger " + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 641, + 287, + 677 + ], + "type": "text", + "content": " encour" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "ages the model to reuse experts. Results in Fig. 3 show that the model size and number of experts decreases with the increase of " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " determines the frequency of architecture search and the larger " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " encourages the model to reuse architectures of existing experts. Results in Fig. 4 show that the learning time decreases with the increase of " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 152, + 389, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 389, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 389, + 163 + ], + "type": "text", + "content": "4.4. Visualization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 170, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 300 + ], + "type": "text", + "content": "To analyze the task distance in PAR, we construct a new benchmark CIFAR100-coarse by dividing CIFAR100 into 20 tasks based on the coarse-grained labels. The illustration in Fig. 5 shows that the PAR can obtain reasonable task groups. For example, the PAR adaptively puts tasks about animals into the same groups, such as the Aquatic mammals, Large omnivores and herbivores, Fish, and so on. The Food containers and Household electrical devices are divided into the same group since they both contain cylindrical and circular objects. It also finds the relation between Non-insect invertebrates and Insects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "text", + "content": "Moreover, we visualize the heat maps of task distances obtained by PAR on four streams of CTrL in Fig. 6. The distance between the first and the last task on streams " + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "inline_equation", + "content": "S^{-}" + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{out}}" + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "text", + "content": " are small, which is consistent with the fact that they all come from the same data distribution. Instead, the distribution of the last task on stream " + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 301, + 545, + 408 + ], + "type": "text", + "content": " is perturbed from the distribution of the first task, so their distance is large. The above results show that PAR can obtain reasonable task distance and relatedness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 420, + 378, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 378, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 378, + 431 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 440, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 572 + ], + "type": "text", + "content": "In this paper, we propose a new method named Parameter Allocation & Regularization (PAR). It allows the model to adaptively select an appropriate strategy from parameter allocation and regularization for each task based on the task learning difficulty. Experimental results on multiple benchmarks demonstrate that PAR is scalable and significantly reduces the model redundancy while improving performance. Moreover, qualitative analysis shows that PAR can obtain reasonable task relatedness. Our method is flexible, and in the future, we will introduce more relatedness estimation, regularization, and allocation strategies into PAR." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 582, + 403, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 403, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 403, + 595 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": "This work was supported by the NSFC projects (No. 62072399, No. U19B2042), Chinese Knowledge Center for Engineering Sciences and Technology, MoE Engineering Research Center of Digital Library, China Research Centre on Data and Knowledge for Engineering Sciences and Technology, and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "3We ignore the parameter overhead of the extractor (11.18M) in divergence estimation, which is fixed and will not become a bottleneck with the increase of tasks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7783" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory Aware Synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert Gate: Lifelong Learning With a Network of Experts. In CVPR, pages 3366-3375, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "text", + "content": "[3] Ali Ayub and Alan Wagner. EEC: Learning to Encode and Regenerate Images for Continual Learning. In ICLR 2021, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 249 + ], + "type": "text", + "content": "[4] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark Experience for General Continual Learning: A Strong, Simple Baseline. In NeurIPS 2020, page 11, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 286, + 282 + ], + "type": "text", + "content": "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient Lifelong Learning with A-GEM. In ICLR 2019, Sept. 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 338 + ], + "type": "text", + "content": "[6] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc'Aurelio Ranzato. On Tiny Episodic Memories in Continual Learning. arXiv:1902.10486 [cs, stat], June 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 286, + 384 + ], + "type": "text", + "content": "[7] Hung-Jen Chen, An-Chieh Cheng, Da-Cheng Juan, Wei Wei, and Min Sun. Mitigating Forgetting in Online Continual Learning via Instance-Aware Parameterization. In NeurIPS 2020, page 12, 2020. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 286, + 441 + ], + "type": "text", + "content": "[8] Zhi-Qi Cheng, Xiao Wu, Siyu Huang, Jun-Xiu Li, Alexander G Hauptmann, and Qiang Peng. Learning to transfer: Generalizable attribute learning with multitask neural model search. In Proceedings of the 26th ACM international conference on Multimedia, pages 90-98, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 286, + 485 + ], + "type": "text", + "content": "[9] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, pages 1-1, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "text", + "content": "[10] Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning Without Memorizing. In CVPR, pages 5138-5146, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 522, + 286, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 553 + ], + "type": "text", + "content": "[11] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural Architecture Search: A Survey. JMLR, page 21, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 556, + 286, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 286, + 577 + ], + "type": "text", + "content": "[12] Robert M French. Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences, 3(4):8, 1999. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 286, + 632 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, Las Vegas, NV, USA, June 2016. IEEE. 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 689 + ], + "type": "text", + "content": "[14] Siyu Huang, Xi Li, Zhi-Qi Cheng, Zhongfei Zhang, and Alexander Hauptmann. Gnas: A greedy neural architecture search method for multi-attribute learning. In Proceedings of the 26th ACM international conference on Multimedia, pages 2049–2057, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[15] Zixuan Ke, Bing Liu, and Xingchang Huang. Continual Learning of a Mixed Sequence of Similar and Dissimilar" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Tasks. In NeurIPS 2020, volume 33, pages 18493-18504. Curran Associates, Inc., 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "type": "text", + "content": "[16] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 173, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 194 + ], + "type": "text", + "content": "[17] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "text", + "content": "[18] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[19] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming Catastrophic Forgetting by Incremental Moment Matching. In NeurlPS 2017, volume 30, pages 4652-4662, 2017. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 305, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 348 + ], + "type": "text", + "content": "[20] Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher, and Caiming Xiong. Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting. In ICML, page 10, 2019. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "text", + "content": "[21] Zhizhong Li and Derek Hoiem. Learning without Forgetting. IEEE transactions on pattern analysis and machine intelligence, page 14, 2017. 1, 2, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[22] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: DIFFERENTIABLE ARCHITECTURE SEARCH. In ICLR 2019, page 13, 2018. 2, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 416, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 446 + ], + "type": "text", + "content": "[23] Huihui Liu, Yiding Yang, and Xinchao Wang. Overcoming Catastrophic Forgetting in Graph Neural Networks. In AAAI 2021, 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "text", + "content": "[24] Jorge A. Mendez and Eric Eaton. Lifelong Learning of Compositional Structures. In ICLR 2021, Sept. 2020. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "text", + "content": "[25] Qier Meng and Satoh Shin'ichi. ADINet: Attribute Driven Incremental Network for Retinal Image Classification. In CVPR 2020, pages 4033-4042, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "text", + "content": "[26] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual Learning with Filter Atom Swapping. In International Conference on Learning Representations, Sept. 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 568 + ], + "type": "text", + "content": "[27] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. VARIATIONAL CONTINUAL LEARNING. In ICLR, page 18, 2018. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[28] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Johannichen, and Moin Nabi. Learning to Remember: A Synaptic Plasticity Driven Framework for Continual Learning. In CVPR, pages 11321-11329, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "text", + "content": "[29] Oleksiy Ostapenko, Pau Rodriguez, Massimo Caccia, and Laurent Charlin. Continual Learning via Local Module Composition. In NeurIPS 2021, May 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[30] Qi Qin, Wenpeng Hu, Han Peng, Dongyan Zhao, and Bing Liu. BNS: Building Network Structures Dynamically for Continual Learning. In NeurIPS 2021, May 2021. 1, 2, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[31] Jathushan Rajasegaran, Munawar Hayat, Salman H Khan, Fahad Shahbaz Khan, and Ling Shao. Random Path Selection for Continual Learning. In H. Wallach, H. Larochelle, A." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7784" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Beygelzimer, F. d\\text{text}quotingingle Alché-Buc, E. Fox, and R. Garnett, editors, NeurIPS 2019, pages 12669-12679. Curran Associates, Inc., 2019. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 151 + ], + "type": "text", + "content": "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. iCaRL: Incremental Classifier and Representation Learning. In CVPR, pages 2001–2010, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 198 + ], + "type": "text", + "content": "[33] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and Xin Wang. A Comprehensive Survey of Neural Architecture Search: Challenges and Solutions. ACM Computing Surveys, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "type": "text", + "content": "[34] Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive Neural Networks. arXiv:1606.04671 [cs], June 2016. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 286, + 267 + ], + "type": "text", + "content": "[35] Gobinda Saha and Kaushik Roy. Gradient Projection Memory for Continual Learning. In ICLR 2021, Sept. 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 312 + ], + "type": "text", + "content": "[36] Joan Serrà, Dídac Surís, Marius Miron, and Alexandros Karatzoglou. Overcoming Catastrophic Forgetting with Hard Attention to the Task. In ICML, page 10, 2018. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 315, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 287, + 357 + ], + "type": "text", + "content": "[37] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual Learning with Deep Generative Replay. In NeurIPS 2017, pages 2990-2999. Curran Associates, Inc., 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 403 + ], + "type": "text", + "content": "[38] Pravendra Singh, Pratik Mazumder, Piyush Rai, and Vinay P. Namboodiri. Rectification-Based Knowledge Retention for Continual Learning. In CVPR 2021, pages 15282-15291, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 406, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 438 + ], + "type": "text", + "content": "[39] Pravendra Singh, Vinay Kumar Verma, Pratik Mazumder, Lawrence Carin, and Piyush Rai. Calibrating CNNs for Lifelong Learning. In NeurIPS 2020, volume 33, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 441, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 287, + 462 + ], + "type": "text", + "content": "[40] Binh Tang and David S. Matteson. Graph-Based Continual Learning. In ICLR 2021, Sept. 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 529 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is All you Need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, NeurIPS 2017, pages 5998-6008. Curran Associates, Inc., 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 574 + ], + "type": "text", + "content": "[42] Tom Veniat, Ludovic Denoyer, and MarcAurelio Ranzato. Efficient Continual Learning with Modular Networks and Task-Driven Priors. In ICLR 2021, Sept. 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 631 + ], + "type": "text", + "content": "[43] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[44] Qing Wang, Sanjeev R. Kulkarni, and Sergio Verdu. Divergence Estimation for Multidimensional Densities Via " + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\$k" + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "-Nearest-Neighbor Distances. IEEE Transactions on Information Theory, 55(5):2392–2405, 2009. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[45] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: An Alternative Approach to Efficient Ensemble and Lifelong Learning. In ICLR, Sept. 2019. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 240 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[46] Ju Xu and Zhanxing Zhu. Reinforced Continual Learning. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, NeurIPS 2018, pages 899–908. Curran Associates, Inc., 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "text", + "content": "[47] Shipeng Yan, Jiangwei Xie, and Xuming He. DER: Dynamically Expandable Representation for Class Incremental Learning. In CVPR 2021, pages 3014-3023, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[48] Xiawu Zheng, Rongrong Ji, Lang Tang, Baochang Zhang, Jianzhuang Liu, and Qi Tian. Multinomial Distribution Learning for Effective Neural Architecture Search. In ICCV, pages 1304-1313, 2019. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[49] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning Transferable Architectures for Scalable Image Recognition. In CVPR 2018, pages 8697-8710, Salt Lake City, UT, 2018. IEEE. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7785" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_content_list.json b/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..623141eb25b565ac2a7f4e5f93960f6b7d67a61d --- /dev/null +++ b/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_content_list.json @@ -0,0 +1,1681 @@ +[ + { + "type": "text", + "text": "Task Residual for Tuning Vision-Language Models", + "text_level": 1, + "bbox": [ + 236, + 130, + 750, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Yu $^{1,2*}$ Zhihe Lu $^{1*}$ Xin Jin $^{1,2}$ Zhibo Chen $^{2}$ Xinchao Wang $^{1\\dagger}$ $^{1}$ National University of Singapore $^{2}$ University of Science and Technology of China", + "bbox": [ + 161, + 179, + 821, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yutao666@mail.ustc.edu.cn, zhihelu@nus.edu.sg, jinxustc@mail.ustc.edu.cn, chenzhibo@ustc.edu.cn, xinchao@nus.edu.sg", + "bbox": [ + 166, + 219, + 812, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 236, + 286, + 316, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes.", + "bbox": [ + 75, + 318, + 478, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decade, deep learning-based visual recognition models [10, 18, 28, 55, 58] have achieved great success. These state-of-the-art models are often trained on a large amount of image and discrete label pairs. The discrete label is generated by converting a detailed textual description, e.g., \"American curl cat\", into a simple scalar, which strikingly eases the computation of loss function. However, this also results in two evident limitations: (i) the rich semantics", + "bbox": [ + 75, + 742, + 475, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec5630472625830893480b27ff0857a22d4d426ea19f25fcd61d1b27c78071fe.jpg", + "image_caption": [ + "Figure 1. Performance comparison between Zero-shot CLIP [49], CoOp [73], CLIP-Adapter [15], Tip-Adapter-F [71] and our TaskRes on ImageNet with few shot settings." + ], + "image_footnote": [], + "bbox": [ + 540, + 287, + 874, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in the textual description are underused, and (ii) the trained models are limited to recognizing the close-set classes only.", + "bbox": [ + 504, + 550, + 906, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent large-scale vision-language model (VLM) pretraining [1, 23, 33, 49, 69] eliminates those limitations by learning visual representations via textual supervision. For instance, texts and images are encoded and mapped into a unified space via a contrastive loss during pre-training [49]. The pre-trained text encoder can then be used to synthesize a text-based classifier for image recognition given the corresponding natural language descriptions as shown in Figure 2 (a). Those pre-trained VLMs have demonstrated a powerful transferability on a variety of downstream tasks in a zero-shot manner. However, the effectiveness of the aforementioned models heavily relies on their large-scale architectures and training datasets. For instance, CLIP [49] has up to 428 million parameters and is trained on 0.4 billion text-image pairs, while Flamingo [1] boasts up to 80 billion parameters and is trained on a staggering 2.1 billion pairs. This makes it impractical to fully fine-tune the model on downstream tasks in a low-data regime.", + "bbox": [ + 504, + 582, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For that reason, efficient transfer learning (ETL) [15, 71-73] on pre-trained VLMs has gained popularity. ETL represents transfer learning to downstream tasks in both", + "bbox": [ + 504, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10899", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b34b2bbb64c09659bf9d7071e34c573312b6e19ec4dcef6b02745c33246da132.jpg", + "image_caption": [ + "(a) Zero-shot CLIP" + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 457, + 169 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2086ad88a7c944b314b658e80da43ffcbf9d48a5eaeb1f050ee86b65e1327d52.jpg", + "image_caption": [ + "(b) Prompt Tuning" + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 851, + 169 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c0ae072c1d6bee0fcc19c326e4d83304df18069ccc354f174291732cbb2d830f.jpg", + "image_caption": [ + "(c) Adapter-style Tuning" + ], + "image_footnote": [], + "bbox": [ + 133, + 191, + 514, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fc3168e0f3d22bbaf220920fb5eccc3310dd2c71387d13b24ca4ed97827b6ea.jpg", + "image_caption": [ + "(d) Task Residual Tuning", + "Figure 2. Illustration of (a) Zero-shot CLIP, (b) prompt tuning, (c) adapter-style tuning and (d) our proposed Task Residual Tuning (TaskRes). Our method introduces a prior-independent task residual to the fixed pre-trained classifier (i.e., text embeddings of CLIP), being free of running the text encoder every time or extra architecture design." + ], + "image_footnote": [], + "bbox": [ + 526, + 191, + 849, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "parameter- and data-efficient manner. The core of ETL is twofold: (i) properly inheriting the well-learned knowledge structure of VLMs, which is already transferable; (ii) effectively exploring the task-specific knowledge given limited data. However, most existing ETL approaches, i.e., prompt tuning (PT) [72, 73] and adapter-style tuning (AT) [15, 71], either damage the prior knowledge of VLMs or learn the new knowledge of a task in an inappropriate/insufficient way. For example, instead of using the pre-trained text-based classifier, CoOp [73] (in Figure 2 (b)) is proposed to learn a continuous prompt for synthesizing a completely new one, which inevitably causes the loss of previous knowledge. Consequently, CoOp underperforms Zero-shot CLIP by $1.03\\% / 0.37\\%$ in 1-/2-shot learning on ImageNet (see Figure 1). In contrast, CLIP-Adapter [15] preserves the pretrained classifier, but is excessively biased towards the prior knowledge when learning a new task, i.e., it transforms the pre-trained classifier weights to be task-specific as illustrated in Figure 2 (c). This results in an inferior new knowledge exploration, thereby a lower accuracy as shown in Figure 1.", + "bbox": [ + 75, + 375, + 475, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For better ETL on pre-trained VLMs, we propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the old knowledge of the pretrained models and the new knowledge for a target task. The rationale is that the decoupling enables a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration, i.e., the learned knowledge w.r.t. the task is independent on the old knowledge. Specifically, TaskRes keeps the original classifier weights frozen and introduces a set of prior-independent parameters that are added to the weights. These additive parameters, tuned for adaptation to the target task, are thus named \"task residual\".", + "bbox": [ + 75, + 681, + 477, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To gain insight into how TaskRes works, we perform ex", + "bbox": [ + 96, + 885, + 475, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tensive experiments across 11 benchmark datasets [73] and conduct a systematic investigation of learned task residuals. The experimental results demonstrate that introducing task residual can significantly enhance the transfer performance. We visualize the correlation between the magnitude of learned task residual and the difficulty of transferring a pre-trained model to a downstream task, and observe that the magnitude increases with the transfer difficulty. This suggests the residual is automatically adapted to the task to fully explore the new knowledge, thereby achieving a new state-of-the-art performance on 11 diverse datasets. Furthermore, it is worth noting that our method requires minimal effort for the implementation, i.e., technically adding one line of code only. Our contributions are summarized below:", + "bbox": [ + 504, + 375, + 906, + 585 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We for the first time emphasize the necessity of a proper knowledge inheritance from pre-trained VLMs to downstream tasks via ETL, reveal the pitfalls of existing tuning paradigms, and conduct an in-depth analysis to manifest that decoupling the old pre-trained knowledge and the new task-specific knowledge is the key.", + "- We propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which achieves a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration.", + "- TaskRes is convenient for use, which needs a few tuning parameters and effortless implementation." + ], + "bbox": [ + 524, + 594, + 906, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 506, + 806, + 648, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-Language Models", + "text_level": 1, + "bbox": [ + 506, + 832, + 733, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We mainly review the literature of vision-language models (VLMs) on language-driven visual representation learning (LDVRL) [2,9,11,14,17,25,29,31,49,52,56]. The core of", + "bbox": [ + 504, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10900", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LDVRL is to map texts (language) and images into a common space so that the text representations can be used for visual classification. To achieve this, two encoders for texts and images respectively, and specific loss functions as regularization are needed.", + "bbox": [ + 75, + 90, + 473, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works explore unsupervised pre-trained models [56] or the skip-gram text modeling [14, 44, 45] for text embedding while sparse coding and vector quantization [7, 56] or Classeme features [11, 60] for visual encoding. For training objectives, MSE [56], self-supervised topic probabilities matching [17] or multi-class logistic loss are adopted. However, those methods are limited to small datasets [9, 25, 31] and weakly representative encoding backbones [7, 56], which tremendously hinders their transferability.", + "bbox": [ + 75, + 167, + 475, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, recent works [23, 35, 49] have advanced previous ones by leveraging billion-level data of image-text pairs from internet and super powerful neural networks, e.g., Transformers [10, 61, 62, 68, 70], for representation learning. By end-to-end pre-training with a contrastive loss, large-scale VLMs demonstrate remarkable transferability on various downstream tasks in a zero-shot evaluation manner. Standing on the strong transfer capability of large-scale vision-language pre-training, we further explore its potential in efficient transfer learning.", + "bbox": [ + 75, + 304, + 475, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Efficient Transfer Learning", + "text_level": 1, + "bbox": [ + 76, + 467, + 323, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-training a neural network on large-scale datasets, e.g., ImageNet [8] and WebImageText [49], then fine-tuning it on downstream tasks has been a common step of transfer learning [36-38, 65-67]. Here, we mainly focus on efficient transfer learning (ETL) on pre-trained VLMs. ETL represents both parameter- and data-efficient transfer learning, in which a small number of parameters are tuned and a small set of data is utilized. Existing works on ETL can be grouped into two categories: prompt tuning [40, 72, 73] and adapter-style tuning (AT) [15, 71]. Specifically, prompt engineering [24, 54] is first explored to generate appropriate discrete prompts for downstream tasks. Later, learning continuous prompts [39, 73] that can better adapt to a new task shows a more advanced performance. However, these methods face two issues: (i) they need the pre-trained text encoder to engage in the whole training, limiting its scalability and increasing the computational overhead; (ii) they abandon the well-pre-trained text-based classifier and generate a new one, which causes the loss of prior knowledge of VLMs.", + "bbox": [ + 75, + 491, + 475, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AT-based methods [15, 71] solve the above issues by using the text encoder once for text-based classifier generation and then focusing on adapting text/image features only. This straightforward design can achieve even better performance, but it heavily relies on the prior knowledge/pre-trained features, which causes an inferior new knowledge exploration. To address this, we propose a new efficient tuning approach, Task Residual Tuning (TaskRes), for a better and more flexi", + "bbox": [ + 75, + 779, + 475, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ble task-specific knowledge learning by a prior-independent \"task residual\".", + "bbox": [ + 504, + 90, + 906, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Few-Shot Learning", + "text_level": 1, + "bbox": [ + 506, + 130, + 691, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot learning (FSL) aims to adapt a model to novel tasks/classes with a handful of labeled examples. Conventional FSL methods [5, 13, 16, 41, 42, 53] often meta-learn on abundant data from base classes for adaptive capability. However, the need of learning on base datasets restricts their scalability. Recent VLM pre-training works [23, 49] offer an effective alternative, which does not require a base dataset. They show that the pre-trained models can already achieve astonishing performance on many downstream tasks in a zero-shot manner. ETL [15, 73] can further boost the performance. In this work, we propose a novel ETL method for adapting VLMs to downstream tasks and evaluate its effectiveness on few-shot tasks.", + "bbox": [ + 504, + 152, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 506, + 363, + 643, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We briefly introduce the adopted VLM, i.e., contrastive language-image pre-training (CLIP) [49], and recap two mainstream approaches for ETL on VLMs, i.e., prompt tuning and adapter-style tuning.", + "bbox": [ + 504, + 388, + 906, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contrastive Language-Image Pre-training", + "text_level": 1, + "bbox": [ + 506, + 458, + 864, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The CLIP model [49] is designed to obtain visual representations through natural language supervision. It is trained on 0.4 billion image-text pairs, where image features from an image encoder and text features from a text encoder are aligned within a unified embedding space using a contrastive learning loss, allowing CLIP to effectively capture broad visual concepts and learn general visual representations. When testing, CLIP can classify a query image into $K$ possible categories. This is achieved by calculating the cosine similarity between a query image embedding $\\mathbf{z}$ , obtained from the image encoder with projection, and the text embeddings $\\{\\mathbf{t}_i\\}_{i=1}^K$ , which are derived by inputting texts (e.g., \"a photo of a {class}\") into the text branch. The predicted probability for class $i$ is formulated as", + "bbox": [ + 504, + 481, + 908, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (y = i | \\mathbf {z}) = \\frac {\\exp (\\sin (\\mathbf {z} , \\mathbf {t} _ {i}) / \\tau)}{\\sum_ {j = 1} ^ {K} \\exp (\\sin (\\mathbf {z} , \\mathbf {t} _ {j}) / \\tau)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 699, + 906, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sin (\\cdot ,\\cdot)$ indicates cosine similarity and $\\tau$ is the learned temperature of CLIP.", + "bbox": [ + 504, + 747, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Revisiting Previous Tuning Paradigms", + "text_level": 1, + "bbox": [ + 506, + 786, + 836, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the success of ETL approaches in natural language processing, e.g., prompt tuning [30, 34] and Adapter [22], recent advances (e.g., CoOp [73] and CLIP-Adapter [15]) borrow their ideas to ETL on VLMs.", + "bbox": [ + 504, + 809, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CoOp, for the first time, introduces prompt tuning to VLMs. Instead of using fixed text prompt contexts such as", + "bbox": [ + 506, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "\"a photo of a\", CoOp proposes to use $M$ learnable context vectors $\\{\\mathbf{v}_m\\}_{m=1}^M$ as the task-specific templates. The prompt given to the text encoder then becomes $\\{\\mathbf{v}_1, \\mathbf{v}_2, \\dots, \\mathbf{v}_M, \\mathbf{c}_i\\}$ where $\\mathbf{c}_i$ is the embedding of class $i$ . During the whole training process, CoOp keeps the parameters of pre-trained VLMs frozen and only tunes the learnable vectors $\\{\\mathbf{v}_m\\}_{m=1}^M$ .", + "bbox": [ + 75, + 90, + 473, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adapter-style tuning introduces additional modules $\\phi_{\\omega}(\\cdot)$ with tunable parameters $\\omega$ to the pre-trained models to transform the pre-trained features $\\mathbf{f}$ to new ones $\\mathbf{f}'$ . In general, adapter-style tuning can be formulated as", + "bbox": [ + 75, + 181, + 475, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {\\prime} = \\mathbf {f} + \\alpha \\phi_ {\\omega} (\\mathbf {f}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 253, + 475, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling factor. In CLIP-Adapter, the adapter module $\\phi_{\\omega}$ is composed of two linear transformation layers and a ReLU activation between them. CLIP-Adapter investigates both visual and text adapters, i.e., applying the adapter module to the image and text branches of CLIP, respectively, and shows that they have comparable performance. During training on downstream tasks, adapter-style methods only tune their adapter modules.", + "bbox": [ + 75, + 280, + 475, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 76, + 414, + 184, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we start by identifying the pitfalls of the existing ETL paradigms on VLMs. Those pitfalls motivate us to propose a new ETL approach for VLMs, named Task Residual Tuning (TaskRes). The proposed TaskRes is simple yet can effectively avoid the pitfalls.", + "bbox": [ + 75, + 439, + 475, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Pitfalls of Existing ETL Paradigms on VLMs", + "text_level": 1, + "bbox": [ + 76, + 523, + 457, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We rethink the usage of the prior knowledge from pretrained VLMs and the acquisition of the new knowledge regarding downstream tasks. On the one hand, large-scale VLMs trained with a vast amount of data have learned broad visual concepts that are general for a wide range of downstream vision tasks, enabling homogenization [3]. The prior knowledge should be well preserved when performing a transfer. On the other hand, despite the huge data used in the pre-training, inevitably, there are domain shifts or uncertain concepts in downstream tasks. New knowledge specific to the downstream tasks should be appropriately supplemented to the prior knowledge. However, existing ETL paradigms do not well consider the above principles and have the following two issues.", + "bbox": [ + 75, + 547, + 475, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pitfall 1: Lack of guarantees of prior knowledge preservation in prompt tuning. While the weights of pre-trained text branch modules (e.g., text encoder and projection) are frozen in the prompt tuning paradigm, the original well-learned classification boundary is more or less damaged. This is because the tuning of input prompts ends up with a new boundary that may forget the old knowledge without explicit regularization. As a result, the performance of prompt tuning is limited. For instance, the performance of", + "bbox": [ + 75, + 763, + 478, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CoOp [73] is not as good as Zero-shot CLIP in 1-/2-shot learning on ImageNet as shown in Figure 1.", + "bbox": [ + 504, + 90, + 906, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pitfall 2: Limited flexibility of new knowledge exploration in adapter-style tuning. The data distributions in downstream tasks often shift from the pre-training distribution, and some task-specific or fine-grained visual concepts/representations may not be well learned by pre-trained VLMs, e.g., from CLIP [49] to satellite image dataset EuroSAT [19]. Thus, the new knowledge regarding downstream tasks needs to be appropriately explored. We observe that adapter-style tuning may not sufficiently explore the task-specific knowledge as the input of the adapter is strictly limited to old/pre-trained features, as shown in Figure 2 (c). Regardless of whether the pre-trained features are suitable for the task, the results of the adapter depend only on them, making adapter-style tuning have limited flexibility for learning new knowledge.", + "bbox": [ + 504, + 128, + 908, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Task Residual Tuning", + "text_level": 1, + "bbox": [ + 506, + 364, + 710, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that the existing ETL paradigms face the aforementioned issues, we propose Task Residual Tuning (TaskRes) to address the issues in a simple way. TaskRes explicitly decouples the maintenance of the old knowledge from the pre-trained VLMs and the learning of task-specific knowledge that is not overly biased on the pre-trained features. We elaborate our TaskRes below.", + "bbox": [ + 504, + 388, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fixed base classifier. As illustrated in Figure 2 (d), our TaskRes performs the tuning directly on the text-based classifier (i.e., text embeddings)1. Base classifier is the text embeddings of a pre-trained vision-language model, e.g., CLIP. We denote the base classifier as $\\mathbf{t} \\in \\mathbb{R}^{K \\times D}$ where $K$ is the number of categories and $D$ is the feature dimension. We keep the base classifier weights frozen to explicitly prevent the base classifier from being damaged.", + "bbox": [ + 504, + 501, + 908, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior-independent task residual. To learn task-specific knowledge without being restricted by prior knowledge, we propose task residual, which is a set of tunable parameters $\\mathbf{x} \\in \\mathbb{R}^{K \\times D}$ that are not dependent on the base classifier. Our task residual is scaled by a factor $\\alpha$ and element-wisely added to the base classifier to formulate a new classifier $\\mathbf{t}'$ for the target task, written as", + "bbox": [ + 504, + 630, + 908, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} ^ {\\prime} = \\mathbf {t} + \\alpha \\mathbf {x}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 742, + 906, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tuning for downstream tasks. During tuning, we only tune the prior-independent task residual while keeping the base classifier (together with the image branch) fixed, enabling reliable old knowledge preservation and flexible new", + "bbox": [ + 504, + 768, + 908, + 829 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1While we can tune on image embeddings, it may encounter overfitting since the testing image embeddings are different from the training ones. Tuning on text embeddings is free of the issue. Besides, large diversity between image embeddings is not conducive to learning stable task-level information.", + "bbox": [ + 504, + 838, + 908, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10902", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "knowledge exploration. Given an image, the fixed image branch of CLIP extracts its embeddings $\\mathbf{z} \\in \\mathbb{R}^D$ . The predicted probability for class $i$ is then calculated as", + "bbox": [ + 75, + 90, + 477, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (y = i | \\mathbf {z}) = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} , \\mathbf {t} _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} , \\mathbf {t} _ {j} ^ {\\prime}\\right) / \\tau\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 143, + 475, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the predicted probabilities, the downstream task loss (e.g., cross-entropy loss) only updates the task residual through standard backpropagation.", + "bbox": [ + 75, + 188, + 477, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 247, + 200, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 271, + 158, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow prior efficient transfer learning (ETL) work [15, 71, 73] to conduct a few-shot evaluation for ETL models on 11 benchmark datasets i.e., ImageNet [8], Caltech101 [12], OxfordPets [48], StanfordCars [27], Flowers102 [47], Food101 [4], FGVCAircraft [43], SUN397 [64], DTD [6], EuroSAT [19] and UCF101 [57]. Those datasets include a wide range of visual recognition tasks, e.g., classifying general objects, fine-grained objects, actions, scenes, etc. Specifically, ETL models are trained using $1/2/4/8/16$ shots per class from the training sets, respectively, and are tested on the full test sets. Additionally, following CoOp [73], we test the generalization performance of our models from ImageNet to its variants (ImageNetV2 [51], ImageNet-Sketch [63], ImageNet-A [21] and ImageNet-R [20]).", + "bbox": [ + 75, + 294, + 477, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Implementation", + "text_level": 1, + "bbox": [ + 76, + 513, + 236, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach has two main components, i.e., base classifier and task residual. For the base classifier, we develop two versions: regular base classifier that directly adopts the weights of the text embeddings of the pre-trained CLIP (using pre-defined text templates following prior work [49, 71]), and enhanced base classifier obtained by tuning the text projection layer of CLIP on the target task before starting our task residual tuning $^2$ . Task residual is a matrix filled with learnable parameters that are initialized by zeros. Task residual has the same shape as the base classifier. We scale the task residual and element-wisely add it to the base classifier, as in Eq. 3. By default, the scaling factor $\\alpha$ is set to 0.5 for all datasets except for Flowers102 using 1. Additionally, we explore the use of a learnable $\\alpha$ later in our ablation study. All models in this work are built upon the pre-trained CLIP models. Unless otherwise specified, we use the ResNet-50 [18] version of CLIP which has a ResNet-50 backbone in the image branch. We train our models for 100 epochs for 1-/2-/4-shot experiments and 200 epochs for 8-/16-shot experiments, with batch size 256. Besides, the aforementioned enhanced base classifier is tuned for 50 epochs. Our models", + "bbox": [ + 75, + 537, + 477, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are optimized by Adam [26] with an initial learning rate of $2e - 3$ except for ImageNet using $2e - 4$ . Following CoOp [73], our optimization process adopts a cosine learning rate decay schedule and a warmup scheme (i.e., fixing the learning rate in the first epoch to $1e - 5$ ). We test the performance of models after training, and all experimental results are averaged over three random seeds.", + "bbox": [ + 504, + 90, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Performance Comparison", + "text_level": 1, + "bbox": [ + 506, + 208, + 741, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3.1 Few-Shot Learning", + "text_level": 1, + "bbox": [ + 506, + 232, + 696, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We develop two versions of our TaskRes, i.e., Ours/Ours* using the regular/enhanced base classifier, respectively. We compare our models with Zero-shot CLIP [49], together with the state-of-the-art (SOTA) ETL methods including CoOp [73], CLIP-Adapter [15] and Tip-Adapter-F $[71]^3$ . The comparison results of the 11 benchmark datasets are shown in Figure 3. Overall, both Ours and Ours* achieve the SOTA/SOTA-comparable performance on the 11 datasets across all few-shot settings and significantly surpass Zero-shot CLIP [49]. Besides, a few observations can be made. First, with more shots, Tip-Adapter-F reaches a closer performance to ours, but as shown in Figure 1 the number of its tunable parameters increases linearly with that of shots, substantially limiting its scalability. Second, our method shows inferior performance on OxfordPets and Food101. Training with a few shots on these two fine-grained datasets is prone to be overfitting, which is also found in CoOp [73]. Finally, we found the proposed method is superior when needed to recognize a large amount of classes, e.g., 1000 classes of ImageNet and 397 classes of SUN397 in 1-shot case. This is because our TaskRes, not excessively biased to the pre-trained features, can better explore the task-specific knowledge even in such extreme scenarios.", + "bbox": [ + 504, + 257, + 908, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3.2 Domain Generalization", + "text_level": 1, + "bbox": [ + 506, + 614, + 723, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As pointed out by CoOp [73], ETL models that are trained on a specific domain are at risk of learning spurious correlations when generalizing to unseen domains. We evaluate our models concerning domain generalization. We train our models on 16-shot ImageNet and test the generalization performance of the trained models on four unseen ImageNet variant datasets (ImageNet-V2, -Sketch, -A and -R). As shown in Table 1, our models consistently outperform the compared models (Zero-shot CLIP [49], Linear Probe CLIP [49] and CoOp [73]) across various CLIP visual backbones (ResNet-50 [18], ResNet-101 [18], ViT-B/32 [10] and ViT-B/16 [10]). We also observe that our TaskRes using the enhanced base classifier (Ours*) achieves higher accuracy than that using the regular base classifier (Ours) on the source dataset while", + "bbox": [ + 504, + 638, + 908, + 851 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{3}$ Original Tip-Adapter [71] has two versions: training-free Tip-Adapter and a fine-tuning version called Tip-Adapter-F. We report the one with higher performance, i.e., Tip-Adapter-F.", + "bbox": [ + 504, + 862, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2The enhanced base classifier is introduced to help us investigate whether there are consistent performance improvements when the base classifier is enhanced.", + "bbox": [ + 75, + 862, + 477, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6afc4c5d957958795bd4b21505c85970a89696ffaff6333c66cd5658c77763cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 89, + 369, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4490e6b3ce7dfe329a760fea469a63a4fd9a33ad15b0453cc61619991bbccd91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 90, + 609, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2cc10b7a01c9b612c6160b35dd7948090f2192a61e243fd79bebf3d03f0637d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 90, + 854, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b7624f0d791b0a8cb0f20b00af6dd5db577eb6906b00667ac1ffc476975784c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 243, + 367, + 391 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bee1ed9a2cf628ea56787d01fd0a17b103f4f7de2e0f139642820ee49c5b6975.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 243, + 609, + 390 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20689d8536aabae615cbcf8939ac94588553abe587ab383309f31a6484beee6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 243, + 854, + 390 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56cd09bd822811baa963386ed0a7772fe94eea0e9ad11a1477972841d63f421a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 396, + 367, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e74d59325f2293f8d11a23d3e6ee17d57b8b69056021cb9e101dd0f8cbf6cfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 396, + 609, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f14bf08663271e41629dd0a5ec2fe7113f7704d09d8944ec3a1d9347d2671d08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 396, + 854, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d36ea553e962bd64ad5891c35bd98e7cd94f00c57ee77974a284e4509492b066.jpg", + "image_caption": [ + "Figure 3. Performance comparison on few-shot learning, i.e., 1-/2-/4-/8-/16-shot, on 11 benchmark datasets. The top-left is the averaged accuracy over the 11 datasets. The full numerical results can be found in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 130, + 547, + 367, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b9614b1d271812ede7ebf39fb32cf34d8b0a7bcf5e303754e8aaf4fb1a31e7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 549, + 609, + 696 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/633f7e1b0e220eff428c83eef895a978db2766201c16f3184b888df3588369da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 551, + 854, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reducing a certain degree of generalization, which is due to a slight overfitting on the source dataset.", + "bbox": [ + 75, + 752, + 478, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 799, + 230, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importance of prior knowledge preservation and prior-independent tuning. We build four target classifiers $\\mathbf{t}'$ to investigate the importance of prior knowledge preservation and prior-independent tuning (where the tunable parameters are not dependent on the pre-trained features), i.e., original", + "bbox": [ + 75, + 825, + 477, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pre-trained classifier $\\mathbf{t}' = \\mathbf{t}$ (i.e., our regular base classifier), directly-adapted classifier $\\mathbf{t}' = \\phi_{\\omega}(\\mathbf{t})$ , adapter-style classifier $\\mathbf{t}' = \\mathbf{t} + \\alpha \\phi_{\\omega}(\\mathbf{t})$ and our TaskRes classifier $\\mathbf{t}' = \\mathbf{t} + \\alpha \\mathbf{x}$ . We conduct experiments on ImageNet across all few-shot settings, using the same scaling factor $\\alpha = 0.5$ . The results in Table 2 support our motivations that (i) explicit preservation of prior knowledge matters as the directly-adapted classifier performs much worse than the original pre-trained classifier and adapter-style classifier, and (ii) tuning prior-", + "bbox": [ + 504, + 752, + 908, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10904", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cc91907262720a5b1b9dcc5179596856eb3283f81c9948db91ff75b7be1dc944.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVisual BackboneSourceTarget
ImageNet-V2-Sketch-A-RAverage
Zero-Shot CLIP [49]58.1851.3433.3221.6556.0040.58
Linear Probe CLIP [49]55.8745.9719.0712.7434.8628.16
CoOp [73]ResNet-5062.9555.1132.7422.1254.9641.23
Ours64.7556.4735.8322.8060.7043.95
Ours*65.7357.0034.4321.5058.1342.77
Zero-Shot CLIP [49]61.6254.8138.7128.0564.3846.49
Linear Probe CLIP [49]59.7550.0526.8019.4447.1935.87
CoOp [73]ResNet-10166.6058.6639.0828.8963.0047.41
Ours67.7059.5041.7029.8768.0749.79
Ours*68.7360.0040.328.0064.8048.28
Zero-Shot CLIP [49]62.0554.7940.8229.5765.9947.79
Linear Probe CLIP [49]59.5849.7328.0619.6747.2036.17
CoOp [73]ViT-B/3266.8558.0840.4430.6264.4548.40
Ours68.2059.2042.5031.4369.3350.62
Ours*69.1759.4740.8729.7066.2749.08
Zero-Shot CLIP [49]66.7360.8346.1547.7773.9657.18
Linear Probe CLIP [49]65.8556.2634.7735.6858.4346.29
CoOp [73]ViT-B/1671.9264.1846.7148.4174.3258.41
Ours73.0765.3049.1350.3777.7060.63
Ours*73.9065.8547.7049.1775.2359.49
", + "bbox": [ + 233, + 87, + 750, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "independent parameters is more effective than the dependent one as the TaskRes classifier obviously surpasses the adapter-style classifier.", + "bbox": [ + 80, + 402, + 473, + 446 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f53f397305b90a7a12e5f926d407fec5f90003ed1ba97220908effda465ade7.jpg", + "table_caption": [ + "Table 1. Performance comparison on generalization (from ImageNet to ImageNet-V2/-Sketch/-A/-R) with multiple CLIP visual backbones." + ], + "table_footnote": [], + "table_body": "
Target Classifier t'1-shot2-shot4-shot8-shot16-shot
t60.3360.3360.3360.3360.33
φω(t)28.0334.9042.5748.1754.13
t + αφω(t)61.0361.2361.2762.0063.17
t + αx61.4362.1762.9364.0364.75
", + "bbox": [ + 109, + 460, + 442, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of task residual learning. We conduct an ablation study of the effect of introducing our task residual learning (TaskRes). Table 3 shows the few-shot results averaged over 11 datasets of the regular base classifier and that equipped with our TaskRes (Ours), together with the enhanced counterparts. Our TaskRes notably improves the regular base, e.g., with $16.14\\%$ accuracy gain in 16-shot case. Similarly, the performance gain is also impressive even on the enhanced base. This indicates that TaskRes can always benefit a pre-trained base when adapted to downstream tasks.", + "bbox": [ + 80, + 611, + 473, + 762 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/822a3152d11c2df1a213105ad253b9851a257694375d1752b27876d8cbf829fe.jpg", + "table_caption": [ + "Table 2. Ablation study of various combinations of three main components: base classifier $\\mathbf{t}$ , adapter module ${\\phi }_{\\omega }\\left( \\cdot \\right)$ and our \"task residual\" $\\mathbf{x}$ ,for building a target classifier ${\\mathbf{t}}^{\\prime }$ . The experiment is conducted on ImageNet." + ], + "table_footnote": [], + "table_body": "
Classifier1-shot2-shot4-shot8-shot16-shot
Regular Base58.9658.9658.9658.9658.96
Regular Base + TaskRes64.0467.0269.6172.7675.10
Enhanced Base61.1162.8264.4469.9173.09
Enhanced Base + TaskRes64.2867.5570.2873.3575.78
", + "bbox": [ + 89, + 775, + 464, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Variants of TaskRes. The proposed \"task residual\" can also be applied to the image branch of CLIP, denoted as TaskRes-I. We also denote the default TaskRes applied on the text branch as TaskRes-T (i.e., Ours) and on the both branches as TaskRes-I&T. We compare their performance on ImageNet in Table 4. TaskRes-T consistently outperforms TaskRes-I, as shown in Table 4. The performance of TaskRes-I increases much slower than TaskRes-T as the number of shots increases. This is because TaskRes-I tends to encounter overfitting where training and testing image embeddings are different. Tuning on text embeddings can avoid the issue. Additionally, the diversity of image embeddings is larger than that of text embeddings, making it difficult to learn task-level information. TaskRes-I&T performs slightly worse than TaskRes-T. This might be caused by a misalignment between the updated image embeddings and our base classifier.", + "bbox": [ + 509, + 402, + 903, + 657 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dea6a54af21c43a181c9ad2b219865f5c35b250c722ca76c1ffc6ed78553c9a0.jpg", + "table_caption": [ + "Table 3. Exploring the effectiveness of the proposed task residual tuning (TaskRes). The accuracy results averaged on 11 datasets are reported under all shots. Refer to Appendix 2.3 for the full comparison results of the 11 datasets." + ], + "table_footnote": [], + "table_body": "
Model1-shot2-shot4-shot8-shot16-shot
Regular Base60.3360.3360.3360.3360.33
TaskRes-I61.2061.2761.4761.5361.60
TaskRes-T61.4362.1762.9364.0364.75
TaskRes-I&T61.5362.1062.7363.8364.67
", + "bbox": [ + 547, + 669, + 864, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance comparison of three TaskRes variants on ImageNet: TaskRes applied on CLIP's image branch (TaskRes-I), text branch (TaskRes-T) and both branches (TaskRes-I&T).", + "bbox": [ + 509, + 750, + 903, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling factor. We investigate the effect of scaling factor $\\alpha$ and show the results of both manually assigned values and a learned value in Table 5. For the manually set $\\alpha$ , we observe a notable improvement with TaskRes scaled even by 0.1, while the highest accuracy is obtained with $\\alpha = 0.5$ . In contrast, the learnable $\\alpha$ can further boost the performance. We show more analyses and results in Appendix 2.3.", + "bbox": [ + 509, + 794, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10905", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02561acb5c6b8c5780896fa2f6ef3afa95b075d85fa76bad458f611904ddad89.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
α00.10.30.50.71.0Learned
Acc.58.9661.8663.7663.9763.6763.0264.01
", + "bbox": [ + 102, + 88, + 452, + 125 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual backbone. We further evaluate models on various CLIP visual backbones, namely, ResNet-50 (RN50), ResNet-101 (RN101), ViT-B/32 and ViT-B/16, as shown in Table 6. Our method shows consistent superiority against other alternatives regardless of which backbone is used.", + "bbox": [ + 76, + 178, + 475, + 253 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2956db81da2fc83b74cd9273973b551979670d8861a1c31c74ac7f2c2f9658d7.jpg", + "table_caption": [ + "Table 5. Ablation study of scaling factor $\\alpha$ . The results are averaged over 11 datasets in 1-shot case." + ], + "table_footnote": [], + "table_body": "
MethodRN50RN101ViT-B/32ViT-B/16
Zero-shot CLIP [49]58.1861.6262.0566.73
CoOp [73]62.9566.6066.8571.92
CLIP-Adapter [15]63.5965.3966.1971.13
Tip-Adapter-F [71]65.4468.5668.6573.69
Ours64.7567.7068.2073.07
Ours*65.7368.7369.1773.90
", + "bbox": [ + 109, + 266, + 444, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Results of CLIP visual backbones on 16-shot ImageNet.", + "bbox": [ + 81, + 369, + 468, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Investigation of Learned Task Residual", + "text_level": 1, + "bbox": [ + 76, + 398, + 413, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate whether the learned task residuals are in fact related to the difficulty of transferring CLIP to downstream tasks. We first define relative transfer difficulty to assess the difficulty of transferring a pre-trained model to a target task. Concretely, the relative transfer difficulty is defined as the ratio of baseline classifier precision to the zero-shot precision of the pre-trained model. We use a random classifier to serve as the baseline classifier where the precision of it equals to $1 / K$ ( $K$ is the number of classes of the downstream task). We then can calculate the relative transfer difficulty of CLIP regarding the 11 benchmark datasets based on the random classifier and Zero-shot CLIP. We count the average magnitudes of the learned task residual of each dataset in 16-shot setting and find that they are highly related to relative transfer difficulty, as shown in Figure 4. This verifies the reliability of the proposed task residual. There are some interesting findings that may be counter-intuitive, e.g., transferring CLIP to ImageNet (with 1000 classes) is the simplest and to EuroSAT (with 10 classes) is the most difficult. This is because the inherent difficulty of ImageNet is large but the data distribution and containing objects are \"familiar\" to CLIP, whereas EuroSAT is the opposite. Besides, there is more space for TaskRes to play its role in the tasks with large relative transfer difficulty, e.g., EuroSAT (1-shot accuracy gain: $23.71\\%$ ) and DTD (1-shot accuracy gain: $7.85\\%$ ). More results and analyses can be found in Appendix 2.4.", + "bbox": [ + 75, + 421, + 475, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion, Limitation and Future Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 444, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a new approach for tuning VLMs, i.e., TaskRes. The proposed TaskRes performs efficient transfer learning (ETL) on VLMs by explicitly de", + "bbox": [ + 76, + 854, + 475, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25727f268e1c570ff846a98cfe4debd7b16cc218d992ae03e5b14a0cd5fc1f43.jpg", + "image_caption": [ + "Figure 4. The magnitude of learned task residual increases with relative transfer difficulty (in logarithmic scale). The shadow indicates the standard deviation regarding random seeds." + ], + "image_footnote": [], + "bbox": [ + 557, + 90, + 857, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coupling the classifier into two key parts: undamaged base classifier with rich prior knowledge and task residual independent to the base classifier for better exploring task-specific knowledge. Interestingly, the magnitude of learned task residual is highly related to the difficulty of transferring pre-trained VLMs to a target downstream task. This may inspire the community to consider ETL from a new perspective, e.g., modeling \"task-to-task transfer difficulty\" for ETL. Extensive experiments have demonstrated the effectiveness of our TaskRes, despite its simplicity.", + "bbox": [ + 504, + 319, + 906, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Nevertheless, our method bears some limitations. For instance, our method encounters negative transfer on two datasets, i.e., OxfordPets (1-shot) and Food101 (1-/2-/4-/8-shot). We conjecture that this happens with two conditions: (i) the downstream tasks are of high relative transfer difficulty as shown in Figure 4, and (ii) Zero-shot CLIP has already achieved a fair precision on them. Besides, the assessment of transfer difficulty in this work is heuristic.", + "bbox": [ + 504, + 469, + 908, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the rapid advancement of foundation models [3], it is becoming increasingly crucial to establish precise and reliable metrics for assessing the transfer difficulty of pretrained foundation models to downstream tasks. A comprehensive study of transfer difficulty, which includes distribution analyses [32], is in high demand. Moreover, we can extend the transfer difficulty to a concept-wise level, and investigate the correlation between performance and the occurrence frequency of visual concepts by exploring the CLIP models trained on specific datasets, e.g., SLIP [46] trained on YFCC15M [50, 59]. We will explore these in future work.", + "bbox": [ + 504, + 590, + 908, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 506, + 768, + 660, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006), the National Research Foundation, Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, NSFC under Grant U1908209 and 62021001, and ZJNSFC under Grant LQ23F010008.", + "bbox": [ + 504, + 794, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10906", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1", + "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 2", + "[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 4, 8", + "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461. Springer, 2014. 5", + "[5] Qi Cai, Yingwei Pan, Ting Yao, Chenggang Yan, and Tao Mei. Memory matching networks for one-shot image recognition. In CVPR, pages 4080-4088, 2018. 3", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5", + "[7] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, 2011. 3", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3, 5", + "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2, 3", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 3, 5", + "[11] Mohamed Elhoseiny, Babak Saleh, and Ahmed Elgammal. Write a classifier: Zero-shot learning using purely textual descriptions. In ICCV, pages 2584-2591, 2013. 2, 3", + "[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, pages 1126–1135. PMLR, 2017. 3", + "[14] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. NeurIPS, 26, 2013. 2, 3", + "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 1, 2, 3, 5, 8" + ], + "bbox": [ + 78, + 114, + 475, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Micah Goldblum, Steven Reich, Liam Fowl, Renkun Ni, Valeriiia Cherepanova, and Tom Goldstein. Unraveling meta-learning: Understanding feature representations for few-shot tasks. In ICML, pages 3607-3616. PMLR, 2020. 3", + "[17] Lluis Gomez, Yash Patel, Marçal Rusinol, Dimosthenis Karatzas, and CV Jawahar. Self-supervised learning of visual features through embedding images into text topic spaces. In CVPR, pages 4230-4239, 2017. 2, 3", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5", + "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 4, 5", + "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5", + "[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5", + "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799. PMLR, 2019. 3", + "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 3", + "[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3", + "[25] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In ECCV, pages 67-84. Springer, 2016. 2, 3", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2015.5", + "[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IC-CVW, pages 554–561, 2013. 5", + "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1", + "[29] Jimmy Lei Ba, Kevin Swersky, Sanja Fidler, et al. Predicting deep zero-shot convolutional neural networks using textual descriptions. In ICCV, pages 4247-4255, 2015. 2", + "[30] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, pages 3045-3059, Online and Punta Cana, Dominican Republic, Nov. 2021. ACL. 3" + ], + "bbox": [ + 509, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Ang Li, Allan Jabri, Armand Joulin, and Laurens Van Der Maaten. Learning visual n-grams from web data. In ICCV, pages 4183–4192, 2017. 2, 3", + "[32] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. arXiv preprint arXiv:2204.08790, 2022. 8", + "[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 1", + "[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, pages 4582-4597, Online, Aug. 2021. ACL. 3", + "[35] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 3", + "[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3", + "[37] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In NeurIPS, 2022. 3", + "[38] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In CVPR, 2023. 3", + "[39] Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In CVPR, pages 375-383, 2017. 3", + "[40] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3", + "[41] Zhihe Lu, Sen He, Da Li, Yi-Zhe Song, and Tao Xiang. Prediction calibration for generalized few-shot semantic segmentation. arXiv preprint arXiv:2210.08290, 2022. 3", + "[42] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In ICCV, pages 8741-8750, 2021. 3", + "[43] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 3", + "[45] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. NeurIPS, 26, 2013. 3", + "[46] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8", + "[47] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In IC-CVGIP, pages 722-729. IEEE, 2008. 5" + ], + "bbox": [ + 78, + 90, + 475, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, pages 3498-3505. IEEE, 2012. 5", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7, 8", + "[50] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 8", + "[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, pages 5389-5400. PMLR, 2019. 5", + "[52] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In ECCV, pages 153-170. Springer, 2020. 2", + "[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3", + "[54] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In EMNLP, 2020. 3", + "[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[56] Richard Socher, Milind Ganjoo, Christopher D Manning, and Andrew Ng. Zero-shot learning through cross-modal transfer. NeurIPS, 26, 2013. 2, 3", + "[57] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 1", + "[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 8", + "[60] Lorenzo Torresani, Martin Szummer, and Andrew Fitzgibbon. Efficient object category recognition using classemes. In ECCV, pages 776-789. Springer, 2010. 3", + "[61] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 3", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3", + "[63] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5", + "[64] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society" + ], + "bbox": [ + 509, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10908", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5", + "[65] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In ECCV, 2022. 3", + "[66] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In NeurIPS, 2022. 3", + "[67] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In CVPR, 2023. 3", + "[68] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, pages 10819-10829, 2022. 3", + "[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1", + "[70] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 3", + "[71] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 1, 2, 3, 5, 8", + "[72] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3", + "[73] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 5, 7, 8" + ], + "bbox": [ + 78, + 92, + 475, + 532 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10909", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_model.json b/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..12522b7f0fcf4df1f23f1c1c35e43d78b2dff4fe --- /dev/null +++ b/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.131, + 0.75, + 0.154 + ], + "angle": 0, + "content": "Task Residual for Tuning Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.18, + 0.823, + 0.218 + ], + "angle": 0, + "content": "Tao Yu\\(^{1,2*}\\) Zhihe Lu\\(^{1*}\\) Xin Jin\\(^{1,2}\\) Zhibo Chen\\(^{2}\\) Xinchao Wang\\(^{1\\dagger}\\) \n\\(^{1}\\)National University of Singapore \\(^{2}\\)University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.22, + 0.813, + 0.252 + ], + "angle": 0, + "content": "yutao666@mail.ustc.edu.cn, zhihelu@nus.edu.sg, jinxustc@mail.ustc.edu.cn, chenzhibo@ustc.edu.cn, xinchao@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.238, + 0.287, + 0.317, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.479, + 0.697 + ], + "angle": 0, + "content": "Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.476, + 0.864 + ], + "angle": 0, + "content": "Over the past decade, deep learning-based visual recognition models [10, 18, 28, 55, 58] have achieved great success. These state-of-the-art models are often trained on a large amount of image and discrete label pairs. The discrete label is generated by converting a detailed textual description, e.g., \"American curl cat\", into a simple scalar, which strikingly eases the computation of loss function. However, this also results in two evident limitations: (i) the rich semantics" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.288, + 0.875, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.495, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Performance comparison between Zero-shot CLIP [49], CoOp [73], CLIP-Adapter [15], Tip-Adapter-F [71] and our TaskRes on ImageNet with few shot settings." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.551, + 0.907, + 0.581 + ], + "angle": 0, + "content": "in the textual description are underused, and (ii) the trained models are limited to recognizing the close-set classes only." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.583, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Recent large-scale vision-language model (VLM) pretraining [1, 23, 33, 49, 69] eliminates those limitations by learning visual representations via textual supervision. For instance, texts and images are encoded and mapped into a unified space via a contrastive loss during pre-training [49]. The pre-trained text encoder can then be used to synthesize a text-based classifier for image recognition given the corresponding natural language descriptions as shown in Figure 2 (a). Those pre-trained VLMs have demonstrated a powerful transferability on a variety of downstream tasks in a zero-shot manner. However, the effectiveness of the aforementioned models heavily relies on their large-scale architectures and training datasets. For instance, CLIP [49] has up to 428 million parameters and is trained on 0.4 billion text-image pairs, while Flamingo [1] boasts up to 80 billion parameters and is trained on a staggering 2.1 billion pairs. This makes it impractical to fully fine-tune the model on downstream tasks in a low-data regime." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For that reason, efficient transfer learning (ETL) [15, 71-73] on pre-trained VLMs has gained popularity. ETL represents transfer learning to downstream tasks in both" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10899" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.458, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.176, + 0.361, + 0.189 + ], + "angle": 0, + "content": "(a) Zero-shot CLIP" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.089, + 0.852, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.177, + 0.741, + 0.189 + ], + "angle": 0, + "content": "(b) Prompt Tuning" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.192, + 0.516, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.296, + 0.38, + 0.309 + ], + "angle": 0, + "content": "(c) Adapter-style Tuning" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.192, + 0.851, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.296, + 0.768, + 0.308 + ], + "angle": 0, + "content": "(d) Task Residual Tuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.908, + 0.362 + ], + "angle": 0, + "content": "Figure 2. Illustration of (a) Zero-shot CLIP, (b) prompt tuning, (c) adapter-style tuning and (d) our proposed Task Residual Tuning (TaskRes). Our method introduces a prior-independent task residual to the fixed pre-trained classifier (i.e., text embeddings of CLIP), being free of running the text encoder every time or extra architecture design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.477, + 0.678 + ], + "angle": 0, + "content": "parameter- and data-efficient manner. The core of ETL is twofold: (i) properly inheriting the well-learned knowledge structure of VLMs, which is already transferable; (ii) effectively exploring the task-specific knowledge given limited data. However, most existing ETL approaches, i.e., prompt tuning (PT) [72, 73] and adapter-style tuning (AT) [15, 71], either damage the prior knowledge of VLMs or learn the new knowledge of a task in an inappropriate/insufficient way. For example, instead of using the pre-trained text-based classifier, CoOp [73] (in Figure 2 (b)) is proposed to learn a continuous prompt for synthesizing a completely new one, which inevitably causes the loss of previous knowledge. Consequently, CoOp underperforms Zero-shot CLIP by \\(1.03\\% / 0.37\\%\\) in 1-/2-shot learning on ImageNet (see Figure 1). In contrast, CLIP-Adapter [15] preserves the pretrained classifier, but is excessively biased towards the prior knowledge when learning a new task, i.e., it transforms the pre-trained classifier weights to be task-specific as illustrated in Figure 2 (c). This results in an inferior new knowledge exploration, thereby a lower accuracy as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.478, + 0.88 + ], + "angle": 0, + "content": "For better ETL on pre-trained VLMs, we propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the old knowledge of the pretrained models and the new knowledge for a target task. The rationale is that the decoupling enables a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration, i.e., the learned knowledge w.r.t. the task is independent on the old knowledge. Specifically, TaskRes keeps the original classifier weights frozen and introduces a set of prior-independent parameters that are added to the weights. These additive parameters, tuned for adaptation to the target task, are thus named \"task residual\"." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.476, + 0.901 + ], + "angle": 0, + "content": "To gain insight into how TaskRes works, we perform ex" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.375, + 0.908, + 0.587 + ], + "angle": 0, + "content": "tensive experiments across 11 benchmark datasets [73] and conduct a systematic investigation of learned task residuals. The experimental results demonstrate that introducing task residual can significantly enhance the transfer performance. We visualize the correlation between the magnitude of learned task residual and the difficulty of transferring a pre-trained model to a downstream task, and observe that the magnitude increases with the transfer difficulty. This suggests the residual is automatically adapted to the task to fully explore the new knowledge, thereby achieving a new state-of-the-art performance on 11 diverse datasets. Furthermore, it is worth noting that our method requires minimal effort for the implementation, i.e., technically adding one line of code only. Our contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.595, + 0.907, + 0.686 + ], + "angle": 0, + "content": "- We for the first time emphasize the necessity of a proper knowledge inheritance from pre-trained VLMs to downstream tasks via ETL, reveal the pitfalls of existing tuning paradigms, and conduct an in-depth analysis to manifest that decoupling the old pre-trained knowledge and the new task-specific knowledge is the key." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.695, + 0.907, + 0.755 + ], + "angle": 0, + "content": "- We propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which achieves a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.765, + 0.905, + 0.795 + ], + "angle": 0, + "content": "- TaskRes is convenient for use, which needs a few tuning parameters and effortless implementation." + }, + { + "type": "list", + "bbox": [ + 0.526, + 0.595, + 0.907, + 0.795 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.808, + 0.649, + 0.823 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.833, + 0.734, + 0.849 + ], + "angle": 0, + "content": "2.1. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We mainly review the literature of vision-language models (VLMs) on language-driven visual representation learning (LDVRL) [2,9,11,14,17,25,29,31,49,52,56]. The core of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10900" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.475, + 0.166 + ], + "angle": 0, + "content": "LDVRL is to map texts (language) and images into a common space so that the text representations can be used for visual classification. To achieve this, two encoders for texts and images respectively, and specific loss functions as regularization are needed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.476, + 0.304 + ], + "angle": 0, + "content": "Early works explore unsupervised pre-trained models [56] or the skip-gram text modeling [14, 44, 45] for text embedding while sparse coding and vector quantization [7, 56] or Classeme features [11, 60] for visual encoding. For training objectives, MSE [56], self-supervised topic probabilities matching [17] or multi-class logistic loss are adopted. However, those methods are limited to small datasets [9, 25, 31] and weakly representative encoding backbones [7, 56], which tremendously hinders their transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.476, + 0.456 + ], + "angle": 0, + "content": "In contrast, recent works [23, 35, 49] have advanced previous ones by leveraging billion-level data of image-text pairs from internet and super powerful neural networks, e.g., Transformers [10, 61, 62, 68, 70], for representation learning. By end-to-end pre-training with a contrastive loss, large-scale VLMs demonstrate remarkable transferability on various downstream tasks in a zero-shot evaluation manner. Standing on the strong transfer capability of large-scale vision-language pre-training, we further explore its potential in efficient transfer learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.325, + 0.484 + ], + "angle": 0, + "content": "2.2. Efficient Transfer Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.476, + 0.779 + ], + "angle": 0, + "content": "Pre-training a neural network on large-scale datasets, e.g., ImageNet [8] and WebImageText [49], then fine-tuning it on downstream tasks has been a common step of transfer learning [36-38, 65-67]. Here, we mainly focus on efficient transfer learning (ETL) on pre-trained VLMs. ETL represents both parameter- and data-efficient transfer learning, in which a small number of parameters are tuned and a small set of data is utilized. Existing works on ETL can be grouped into two categories: prompt tuning [40, 72, 73] and adapter-style tuning (AT) [15, 71]. Specifically, prompt engineering [24, 54] is first explored to generate appropriate discrete prompts for downstream tasks. Later, learning continuous prompts [39, 73] that can better adapt to a new task shows a more advanced performance. However, these methods face two issues: (i) they need the pre-trained text encoder to engage in the whole training, limiting its scalability and increasing the computational overhead; (ii) they abandon the well-pre-trained text-based classifier and generate a new one, which causes the loss of prior knowledge of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.476, + 0.901 + ], + "angle": 0, + "content": "AT-based methods [15, 71] solve the above issues by using the text encoder once for text-based classifier generation and then focusing on adapting text/image features only. This straightforward design can achieve even better performance, but it heavily relies on the prior knowledge/pre-trained features, which causes an inferior new knowledge exploration. To address this, we propose a new efficient tuning approach, Task Residual Tuning (TaskRes), for a better and more flexi" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "ble task-specific knowledge learning by a prior-independent \"task residual\"." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.131, + 0.692, + 0.146 + ], + "angle": 0, + "content": "2.3. Few-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.154, + 0.907, + 0.35 + ], + "angle": 0, + "content": "Few-shot learning (FSL) aims to adapt a model to novel tasks/classes with a handful of labeled examples. Conventional FSL methods [5, 13, 16, 41, 42, 53] often meta-learn on abundant data from base classes for adaptive capability. However, the need of learning on base datasets restricts their scalability. Recent VLM pre-training works [23, 49] offer an effective alternative, which does not require a base dataset. They show that the pre-trained models can already achieve astonishing performance on many downstream tasks in a zero-shot manner. ETL [15, 73] can further boost the performance. In this work, we propose a novel ETL method for adapting VLMs to downstream tasks and evaluate its effectiveness on few-shot tasks." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.364, + 0.644, + 0.379 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.389, + 0.907, + 0.45 + ], + "angle": 0, + "content": "We briefly introduce the adopted VLM, i.e., contrastive language-image pre-training (CLIP) [49], and recap two mainstream approaches for ETL on VLMs, i.e., prompt tuning and adapter-style tuning." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.459, + 0.866, + 0.475 + ], + "angle": 0, + "content": "3.1. Contrastive Language-Image Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.482, + 0.909, + 0.693 + ], + "angle": 0, + "content": "The CLIP model [49] is designed to obtain visual representations through natural language supervision. It is trained on 0.4 billion image-text pairs, where image features from an image encoder and text features from a text encoder are aligned within a unified embedding space using a contrastive learning loss, allowing CLIP to effectively capture broad visual concepts and learn general visual representations. When testing, CLIP can classify a query image into \\( K \\) possible categories. This is achieved by calculating the cosine similarity between a query image embedding \\( \\mathbf{z} \\), obtained from the image encoder with projection, and the text embeddings \\( \\{\\mathbf{t}_i\\}_{i=1}^K \\), which are derived by inputting texts (e.g., \"a photo of a {class}\") into the text branch. The predicted probability for class \\( i \\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.7, + 0.907, + 0.739 + ], + "angle": 0, + "content": "\\[\np (y = i | \\mathbf {z}) = \\frac {\\exp (\\sin (\\mathbf {z} , \\mathbf {t} _ {i}) / \\tau)}{\\sum_ {j = 1} ^ {K} \\exp (\\sin (\\mathbf {z} , \\mathbf {t} _ {j}) / \\tau)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.748, + 0.908, + 0.779 + ], + "angle": 0, + "content": "where \\(\\sin (\\cdot ,\\cdot)\\) indicates cosine similarity and \\(\\tau\\) is the learned temperature of CLIP." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.787, + 0.837, + 0.804 + ], + "angle": 0, + "content": "3.2. Revisiting Previous Tuning Paradigms" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.81, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Inspired by the success of ETL approaches in natural language processing, e.g., prompt tuning [30, 34] and Adapter [22], recent advances (e.g., CoOp [73] and CLIP-Adapter [15]) borrow their ideas to ETL on VLMs." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CoOp, for the first time, introduces prompt tuning to VLMs. Instead of using fixed text prompt contexts such as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10901" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.475, + 0.183 + ], + "angle": 0, + "content": "\"a photo of a\", CoOp proposes to use \\(M\\) learnable context vectors \\(\\{\\mathbf{v}_m\\}_{m=1}^M\\) as the task-specific templates. The prompt given to the text encoder then becomes \\(\\{\\mathbf{v}_1, \\mathbf{v}_2, \\dots, \\mathbf{v}_M, \\mathbf{c}_i\\}\\) where \\(\\mathbf{c}_i\\) is the embedding of class \\(i\\). During the whole training process, CoOp keeps the parameters of pre-trained VLMs frozen and only tunes the learnable vectors \\(\\{\\mathbf{v}_m\\}_{m=1}^M\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.476, + 0.243 + ], + "angle": 0, + "content": "Adapter-style tuning introduces additional modules \\(\\phi_{\\omega}(\\cdot)\\) with tunable parameters \\(\\omega\\) to the pre-trained models to transform the pre-trained features \\(\\mathbf{f}\\) to new ones \\(\\mathbf{f}'\\). In general, adapter-style tuning can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.254, + 0.476, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {\\prime} = \\mathbf {f} + \\alpha \\phi_ {\\omega} (\\mathbf {f}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.476, + 0.401 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling factor. In CLIP-Adapter, the adapter module \\(\\phi_{\\omega}\\) is composed of two linear transformation layers and a ReLU activation between them. CLIP-Adapter investigates both visual and text adapters, i.e., applying the adapter module to the image and text branches of CLIP, respectively, and shows that they have comparable performance. During training on downstream tasks, adapter-style methods only tune their adapter modules." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.185, + 0.433 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.476, + 0.516 + ], + "angle": 0, + "content": "In this section, we start by identifying the pitfalls of the existing ETL paradigms on VLMs. Those pitfalls motivate us to propose a new ETL approach for VLMs, named Task Residual Tuning (TaskRes). The proposed TaskRes is simple yet can effectively avoid the pitfalls." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.459, + 0.541 + ], + "angle": 0, + "content": "4.1. Pitfalls of Existing ETL Paradigms on VLMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.476, + 0.76 + ], + "angle": 0, + "content": "We rethink the usage of the prior knowledge from pretrained VLMs and the acquisition of the new knowledge regarding downstream tasks. On the one hand, large-scale VLMs trained with a vast amount of data have learned broad visual concepts that are general for a wide range of downstream vision tasks, enabling homogenization [3]. The prior knowledge should be well preserved when performing a transfer. On the other hand, despite the huge data used in the pre-training, inevitably, there are domain shifts or uncertain concepts in downstream tasks. New knowledge specific to the downstream tasks should be appropriately supplemented to the prior knowledge. However, existing ETL paradigms do not well consider the above principles and have the following two issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.479, + 0.901 + ], + "angle": 0, + "content": "Pitfall 1: Lack of guarantees of prior knowledge preservation in prompt tuning. While the weights of pre-trained text branch modules (e.g., text encoder and projection) are frozen in the prompt tuning paradigm, the original well-learned classification boundary is more or less damaged. This is because the tuning of input prompts ends up with a new boundary that may forget the old knowledge without explicit regularization. As a result, the performance of prompt tuning is limited. For instance, the performance of" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "CoOp [73] is not as good as Zero-shot CLIP in 1-/2-shot learning on ImageNet as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.129, + 0.909, + 0.356 + ], + "angle": 0, + "content": "Pitfall 2: Limited flexibility of new knowledge exploration in adapter-style tuning. The data distributions in downstream tasks often shift from the pre-training distribution, and some task-specific or fine-grained visual concepts/representations may not be well learned by pre-trained VLMs, e.g., from CLIP [49] to satellite image dataset EuroSAT [19]. Thus, the new knowledge regarding downstream tasks needs to be appropriately explored. We observe that adapter-style tuning may not sufficiently explore the task-specific knowledge as the input of the adapter is strictly limited to old/pre-trained features, as shown in Figure 2 (c). Regardless of whether the pre-trained features are suitable for the task, the results of the adapter depend only on them, making adapter-style tuning have limited flexibility for learning new knowledge." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.366, + 0.712, + 0.382 + ], + "angle": 0, + "content": "4.2. Task Residual Tuning" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.389, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Given that the existing ETL paradigms face the aforementioned issues, we propose Task Residual Tuning (TaskRes) to address the issues in a simple way. TaskRes explicitly decouples the maintenance of the old knowledge from the pre-trained VLMs and the learning of task-specific knowledge that is not overly biased on the pre-trained features. We elaborate our TaskRes below." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.502, + 0.909, + 0.624 + ], + "angle": 0, + "content": "Fixed base classifier. As illustrated in Figure 2 (d), our TaskRes performs the tuning directly on the text-based classifier (i.e., text embeddings)1. Base classifier is the text embeddings of a pre-trained vision-language model, e.g., CLIP. We denote the base classifier as \\(\\mathbf{t} \\in \\mathbb{R}^{K \\times D}\\) where \\(K\\) is the number of categories and \\(D\\) is the feature dimension. We keep the base classifier weights frozen to explicitly prevent the base classifier from being damaged." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.631, + 0.909, + 0.737 + ], + "angle": 0, + "content": "Prior-independent task residual. To learn task-specific knowledge without being restricted by prior knowledge, we propose task residual, which is a set of tunable parameters \\(\\mathbf{x} \\in \\mathbb{R}^{K \\times D}\\) that are not dependent on the base classifier. Our task residual is scaled by a factor \\(\\alpha\\) and element-wisely added to the base classifier to formulate a new classifier \\(\\mathbf{t}'\\) for the target task, written as" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.743, + 0.907, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} ^ {\\prime} = \\mathbf {t} + \\alpha \\mathbf {x}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.769, + 0.909, + 0.83 + ], + "angle": 0, + "content": "Tuning for downstream tasks. During tuning, we only tune the prior-independent task residual while keeping the base classifier (together with the image branch) fixed, enabling reliable old knowledge preservation and flexible new" + }, + { + "type": "page_footnote", + "bbox": [ + 0.506, + 0.839, + 0.909, + 0.9 + ], + "angle": 0, + "content": "1While we can tune on image embeddings, it may encounter overfitting since the testing image embeddings are different from the training ones. Tuning on text embeddings is free of the issue. Besides, large diversity between image embeddings is not conducive to learning stable task-level information." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.478, + 0.138 + ], + "angle": 0, + "content": "knowledge exploration. Given an image, the fixed image branch of CLIP extracts its embeddings \\(\\mathbf{z} \\in \\mathbb{R}^D\\). The predicted probability for class \\(i\\) is then calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.145, + 0.476, + 0.183 + ], + "angle": 0, + "content": "\\[\np (y = i | \\mathbf {z}) = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} , \\mathbf {t} _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} , \\mathbf {t} _ {j} ^ {\\prime}\\right) / \\tau\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.478, + 0.235 + ], + "angle": 0, + "content": "Based on the predicted probabilities, the downstream task loss (e.g., cross-entropy loss) only updates the task residual through standard backpropagation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.248, + 0.201, + 0.265 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.272, + 0.159, + 0.287 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.478, + 0.508 + ], + "angle": 0, + "content": "We follow prior efficient transfer learning (ETL) work [15, 71, 73] to conduct a few-shot evaluation for ETL models on 11 benchmark datasets i.e., ImageNet [8], Caltech101 [12], OxfordPets [48], StanfordCars [27], Flowers102 [47], Food101 [4], FGVCAircraft [43], SUN397 [64], DTD [6], EuroSAT [19] and UCF101 [57]. Those datasets include a wide range of visual recognition tasks, e.g., classifying general objects, fine-grained objects, actions, scenes, etc. Specifically, ETL models are trained using \\(1/2/4/8/16\\) shots per class from the training sets, respectively, and are tested on the full test sets. Additionally, following CoOp [73], we test the generalization performance of our models from ImageNet to its variants (ImageNetV2 [51], ImageNet-Sketch [63], ImageNet-A [21] and ImageNet-R [20])." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.238, + 0.531 + ], + "angle": 0, + "content": "5.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.478, + 0.856 + ], + "angle": 0, + "content": "Our approach has two main components, i.e., base classifier and task residual. For the base classifier, we develop two versions: regular base classifier that directly adopts the weights of the text embeddings of the pre-trained CLIP (using pre-defined text templates following prior work [49, 71]), and enhanced base classifier obtained by tuning the text projection layer of CLIP on the target task before starting our task residual tuning\\(^2\\). Task residual is a matrix filled with learnable parameters that are initialized by zeros. Task residual has the same shape as the base classifier. We scale the task residual and element-wisely add it to the base classifier, as in Eq. 3. By default, the scaling factor \\(\\alpha\\) is set to 0.5 for all datasets except for Flowers102 using 1. Additionally, we explore the use of a learnable \\(\\alpha\\) later in our ablation study. All models in this work are built upon the pre-trained CLIP models. Unless otherwise specified, we use the ResNet-50 [18] version of CLIP which has a ResNet-50 backbone in the image branch. We train our models for 100 epochs for 1-/2-/4-shot experiments and 200 epochs for 8-/16-shot experiments, with batch size 256. Besides, the aforementioned enhanced base classifier is tuned for 50 epochs. Our models" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "are optimized by Adam [26] with an initial learning rate of \\( 2e - 3 \\) except for ImageNet using \\( 2e - 4 \\). Following CoOp [73], our optimization process adopts a cosine learning rate decay schedule and a warmup scheme (i.e., fixing the learning rate in the first epoch to \\( 1e - 5 \\)). We test the performance of models after training, and all experimental results are averaged over three random seeds." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.209, + 0.742, + 0.225 + ], + "angle": 0, + "content": "5.3. Performance Comparison" + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.233, + 0.697, + 0.248 + ], + "angle": 0, + "content": "5.3.1 Few-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.258, + 0.909, + 0.605 + ], + "angle": 0, + "content": "We develop two versions of our TaskRes, i.e., Ours/Ours* using the regular/enhanced base classifier, respectively. We compare our models with Zero-shot CLIP [49], together with the state-of-the-art (SOTA) ETL methods including CoOp [73], CLIP-Adapter [15] and Tip-Adapter-F \\([71]^3\\). The comparison results of the 11 benchmark datasets are shown in Figure 3. Overall, both Ours and Ours* achieve the SOTA/SOTA-comparable performance on the 11 datasets across all few-shot settings and significantly surpass Zero-shot CLIP [49]. Besides, a few observations can be made. First, with more shots, Tip-Adapter-F reaches a closer performance to ours, but as shown in Figure 1 the number of its tunable parameters increases linearly with that of shots, substantially limiting its scalability. Second, our method shows inferior performance on OxfordPets and Food101. Training with a few shots on these two fine-grained datasets is prone to be overfitting, which is also found in CoOp [73]. Finally, we found the proposed method is superior when needed to recognize a large amount of classes, e.g., 1000 classes of ImageNet and 397 classes of SUN397 in 1-shot case. This is because our TaskRes, not excessively biased to the pre-trained features, can better explore the task-specific knowledge even in such extreme scenarios." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.615, + 0.725, + 0.629 + ], + "angle": 0, + "content": "5.3.2 Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.64, + 0.909, + 0.852 + ], + "angle": 0, + "content": "As pointed out by CoOp [73], ETL models that are trained on a specific domain are at risk of learning spurious correlations when generalizing to unseen domains. We evaluate our models concerning domain generalization. We train our models on 16-shot ImageNet and test the generalization performance of the trained models on four unseen ImageNet variant datasets (ImageNet-V2, -Sketch, -A and -R). As shown in Table 1, our models consistently outperform the compared models (Zero-shot CLIP [49], Linear Probe CLIP [49] and CoOp [73]) across various CLIP visual backbones (ResNet-50 [18], ResNet-101 [18], ViT-B/32 [10] and ViT-B/16 [10]). We also observe that our TaskRes using the enhanced base classifier (Ours*) achieves higher accuracy than that using the regular base classifier (Ours) on the source dataset while" + }, + { + "type": "page_footnote", + "bbox": [ + 0.506, + 0.863, + 0.908, + 0.901 + ], + "angle": 0, + "content": "\\(^{3}\\)Original Tip-Adapter [71] has two versions: training-free Tip-Adapter and a fine-tuning version called Tip-Adapter-F. We report the one with higher performance, i.e., Tip-Adapter-F." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.478, + 0.9 + ], + "angle": 0, + "content": "2The enhanced base classifier is introduced to help us investigate whether there are consistent performance improvements when the base classifier is enhanced." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10903" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.09, + 0.37, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.091, + 0.611, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.091, + 0.855, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.244, + 0.368, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.244, + 0.611, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.244, + 0.855, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.397, + 0.369, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.397, + 0.611, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.397, + 0.855, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.549, + 0.368, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.55, + 0.61, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.552, + 0.855, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.71, + 0.908, + 0.74 + ], + "angle": 0, + "content": "Figure 3. Performance comparison on few-shot learning, i.e., 1-/2-/4-/8-/16-shot, on 11 benchmark datasets. The top-left is the averaged accuracy over the 11 datasets. The full numerical results can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.479, + 0.784 + ], + "angle": 0, + "content": "reducing a certain degree of generalization, which is due to a slight overfitting on the source dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.231, + 0.816 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.478, + 0.902 + ], + "angle": 0, + "content": "Importance of prior knowledge preservation and prior-independent tuning. We build four target classifiers \\(\\mathbf{t}'\\) to investigate the importance of prior knowledge preservation and prior-independent tuning (where the tunable parameters are not dependent on the pre-trained features), i.e., original" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.753, + 0.909, + 0.89 + ], + "angle": 0, + "content": "pre-trained classifier \\(\\mathbf{t}' = \\mathbf{t}\\) (i.e., our regular base classifier), directly-adapted classifier \\(\\mathbf{t}' = \\phi_{\\omega}(\\mathbf{t})\\), adapter-style classifier \\(\\mathbf{t}' = \\mathbf{t} + \\alpha \\phi_{\\omega}(\\mathbf{t})\\) and our TaskRes classifier \\(\\mathbf{t}' = \\mathbf{t} + \\alpha \\mathbf{x}\\). We conduct experiments on ImageNet across all few-shot settings, using the same scaling factor \\(\\alpha = 0.5\\). The results in Table 2 support our motivations that (i) explicit preservation of prior knowledge matters as the directly-adapted classifier performs much worse than the original pre-trained classifier and adapter-style classifier, and (ii) tuning prior-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10904" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.235, + 0.088, + 0.75, + 0.362 + ], + "angle": 0, + "content": "
MethodVisual BackboneSourceTarget
ImageNet-V2-Sketch-A-RAverage
Zero-Shot CLIP [49]58.1851.3433.3221.6556.0040.58
Linear Probe CLIP [49]55.8745.9719.0712.7434.8628.16
CoOp [73]ResNet-5062.9555.1132.7422.1254.9641.23
Ours64.7556.4735.8322.8060.7043.95
Ours*65.7357.0034.4321.5058.1342.77
Zero-Shot CLIP [49]61.6254.8138.7128.0564.3846.49
Linear Probe CLIP [49]59.7550.0526.8019.4447.1935.87
CoOp [73]ResNet-10166.6058.6639.0828.8963.0047.41
Ours67.7059.5041.7029.8768.0749.79
Ours*68.7360.0040.328.0064.8048.28
Zero-Shot CLIP [49]62.0554.7940.8229.5765.9947.79
Linear Probe CLIP [49]59.5849.7328.0619.6747.2036.17
CoOp [73]ViT-B/3266.8558.0840.4430.6264.4548.40
Ours68.2059.2042.5031.4369.3350.62
Ours*69.1759.4740.8729.7066.2749.08
Zero-Shot CLIP [49]66.7360.8346.1547.7773.9657.18
Linear Probe CLIP [49]65.8556.2634.7735.6858.4346.29
CoOp [73]ViT-B/1671.9264.1846.7148.4174.3258.41
Ours73.0765.3049.1350.3777.7060.63
Ours*73.9065.8547.7049.1775.2359.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.373, + 0.903, + 0.387 + ], + "angle": 0, + "content": "Table 1. Performance comparison on generalization (from ImageNet to ImageNet-V2/-Sketch/-A/-R) with multiple CLIP visual backbones." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.404, + 0.475, + 0.448 + ], + "angle": 0, + "content": "independent parameters is more effective than the dependent one as the TaskRes classifier obviously surpasses the adapter-style classifier." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.462, + 0.443, + 0.529 + ], + "angle": 0, + "content": "
Target Classifier t'1-shot2-shot4-shot8-shot16-shot
t60.3360.3360.3360.3360.33
φω(t)28.0334.9042.5748.1754.13
t + αφω(t)61.0361.2361.2762.0063.17
t + αx61.4362.1762.9364.0364.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.54, + 0.475, + 0.596 + ], + "angle": 0, + "content": "Table 2. Ablation study of various combinations of three main components: base classifier \\( \\mathbf{t} \\), adapter module \\( {\\phi }_{\\omega }\\left( \\cdot \\right) \\) and our \"task residual\" \\( \\mathbf{x} \\) ,for building a target classifier \\( {\\mathbf{t}}^{\\prime } \\) . The experiment is conducted on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.612, + 0.475, + 0.763 + ], + "angle": 0, + "content": "Effectiveness of task residual learning. We conduct an ablation study of the effect of introducing our task residual learning (TaskRes). Table 3 shows the few-shot results averaged over 11 datasets of the regular base classifier and that equipped with our TaskRes (Ours), together with the enhanced counterparts. Our TaskRes notably improves the regular base, e.g., with \\(16.14\\%\\) accuracy gain in 16-shot case. Similarly, the performance gain is also impressive even on the enhanced base. This indicates that TaskRes can always benefit a pre-trained base when adapted to downstream tasks." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.776, + 0.465, + 0.848 + ], + "angle": 0, + "content": "
Classifier1-shot2-shot4-shot8-shot16-shot
Regular Base58.9658.9658.9658.9658.96
Regular Base + TaskRes64.0467.0269.6172.7675.10
Enhanced Base61.1162.8264.4469.9173.09
Enhanced Base + TaskRes64.2867.5570.2873.3575.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.86, + 0.475, + 0.915 + ], + "angle": 0, + "content": "Table 3. Exploring the effectiveness of the proposed task residual tuning (TaskRes). The accuracy results averaged on 11 datasets are reported under all shots. Refer to Appendix 2.3 for the full comparison results of the 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.404, + 0.905, + 0.658 + ], + "angle": 0, + "content": "Variants of TaskRes. The proposed \"task residual\" can also be applied to the image branch of CLIP, denoted as TaskRes-I. We also denote the default TaskRes applied on the text branch as TaskRes-T (i.e., Ours) and on the both branches as TaskRes-I&T. We compare their performance on ImageNet in Table 4. TaskRes-T consistently outperforms TaskRes-I, as shown in Table 4. The performance of TaskRes-I increases much slower than TaskRes-T as the number of shots increases. This is because TaskRes-I tends to encounter overfitting where training and testing image embeddings are different. Tuning on text embeddings can avoid the issue. Additionally, the diversity of image embeddings is larger than that of text embeddings, making it difficult to learn task-level information. TaskRes-I&T performs slightly worse than TaskRes-T. This might be caused by a misalignment between the updated image embeddings and our base classifier." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.67, + 0.865, + 0.738 + ], + "angle": 0, + "content": "
Model1-shot2-shot4-shot8-shot16-shot
Regular Base60.3360.3360.3360.3360.33
TaskRes-I61.2061.2761.4761.5361.60
TaskRes-T61.4362.1762.9364.0364.75
TaskRes-I&T61.5362.1062.7363.8364.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.75, + 0.905, + 0.791 + ], + "angle": 0, + "content": "Table 4. Performance comparison of three TaskRes variants on ImageNet: TaskRes applied on CLIP's image branch (TaskRes-I), text branch (TaskRes-T) and both branches (TaskRes-I&T)." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.795, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Scaling factor. We investigate the effect of scaling factor \\(\\alpha\\) and show the results of both manually assigned values and a learned value in Table 5. For the manually set \\(\\alpha\\), we observe a notable improvement with TaskRes scaled even by 0.1, while the highest accuracy is obtained with \\(\\alpha = 0.5\\). In contrast, the learnable \\(\\alpha\\) can further boost the performance. We show more analyses and results in Appendix 2.3." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10905" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.103, + 0.089, + 0.454, + 0.126 + ], + "angle": 0, + "content": "
α00.10.30.50.71.0Learned
Acc.58.9661.8663.7663.9763.6763.0264.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.135, + 0.477, + 0.163 + ], + "angle": 0, + "content": "Table 5. Ablation study of scaling factor \\( \\alpha \\). The results are averaged over 11 datasets in 1-shot case." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.179, + 0.476, + 0.254 + ], + "angle": 0, + "content": "Visual backbone. We further evaluate models on various CLIP visual backbones, namely, ResNet-50 (RN50), ResNet-101 (RN101), ViT-B/32 and ViT-B/16, as shown in Table 6. Our method shows consistent superiority against other alternatives regardless of which backbone is used." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.267, + 0.445, + 0.36 + ], + "angle": 0, + "content": "
MethodRN50RN101ViT-B/32ViT-B/16
Zero-shot CLIP [49]58.1861.6262.0566.73
CoOp [73]62.9566.6066.8571.92
CLIP-Adapter [15]63.5965.3966.1971.13
Tip-Adapter-F [71]65.4468.5668.6573.69
Ours64.7567.7068.2073.07
Ours*65.7368.7369.1773.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.083, + 0.37, + 0.47, + 0.385 + ], + "angle": 0, + "content": "Table 6. Results of CLIP visual backbones on 16-shot ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.414, + 0.416 + ], + "angle": 0, + "content": "5.5. Investigation of Learned Task Residual" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.477, + 0.816 + ], + "angle": 0, + "content": "We investigate whether the learned task residuals are in fact related to the difficulty of transferring CLIP to downstream tasks. We first define relative transfer difficulty to assess the difficulty of transferring a pre-trained model to a target task. Concretely, the relative transfer difficulty is defined as the ratio of baseline classifier precision to the zero-shot precision of the pre-trained model. We use a random classifier to serve as the baseline classifier where the precision of it equals to \\( 1 / K \\) (\\( K \\) is the number of classes of the downstream task). We then can calculate the relative transfer difficulty of CLIP regarding the 11 benchmark datasets based on the random classifier and Zero-shot CLIP. We count the average magnitudes of the learned task residual of each dataset in 16-shot setting and find that they are highly related to relative transfer difficulty, as shown in Figure 4. This verifies the reliability of the proposed task residual. There are some interesting findings that may be counter-intuitive, e.g., transferring CLIP to ImageNet (with 1000 classes) is the simplest and to EuroSAT (with 10 classes) is the most difficult. This is because the inherent difficulty of ImageNet is large but the data distribution and containing objects are \"familiar\" to CLIP, whereas EuroSAT is the opposite. Besides, there is more space for TaskRes to play its role in the tasks with large relative transfer difficulty, e.g., EuroSAT (1-shot accuracy gain: \\( 23.71\\% \\)) and DTD (1-shot accuracy gain: \\( 7.85\\% \\)). More results and analyses can be found in Appendix 2.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.446, + 0.846 + ], + "angle": 0, + "content": "6. Conclusion, Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.476, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose a new approach for tuning VLMs, i.e., TaskRes. The proposed TaskRes performs efficient transfer learning (ETL) on VLMs by explicitly de" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.092, + 0.858, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.267, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Figure 4. The magnitude of learned task residual increases with relative transfer difficulty (in logarithmic scale). The shadow indicates the standard deviation regarding random seeds." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.32, + 0.907, + 0.47 + ], + "angle": 0, + "content": "coupling the classifier into two key parts: undamaged base classifier with rich prior knowledge and task residual independent to the base classifier for better exploring task-specific knowledge. Interestingly, the magnitude of learned task residual is highly related to the difficulty of transferring pre-trained VLMs to a target downstream task. This may inspire the community to consider ETL from a new perspective, e.g., modeling \"task-to-task transfer difficulty\" for ETL. Extensive experiments have demonstrated the effectiveness of our TaskRes, despite its simplicity." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.47, + 0.909, + 0.59 + ], + "angle": 0, + "content": "Nevertheless, our method bears some limitations. For instance, our method encounters negative transfer on two datasets, i.e., OxfordPets (1-shot) and Food101 (1-/2-/4-/8-shot). We conjecture that this happens with two conditions: (i) the downstream tasks are of high relative transfer difficulty as shown in Figure 4, and (ii) Zero-shot CLIP has already achieved a fair precision on them. Besides, the assessment of transfer difficulty in this work is heuristic." + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.591, + 0.909, + 0.756 + ], + "angle": 0, + "content": "With the rapid advancement of foundation models [3], it is becoming increasingly crucial to establish precise and reliable metrics for assessing the transfer difficulty of pretrained foundation models to downstream tasks. A comprehensive study of transfer difficulty, which includes distribution analyses [32], is in high demand. Moreover, we can extend the transfer difficulty to a concept-wise level, and investigate the correlation between performance and the occurrence frequency of visual concepts by exploring the CLIP models trained on specific datasets, e.g., SLIP [46] trained on YFCC15M [50, 59]. We will explore these in future work." + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.77, + 0.661, + 0.788 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006), the National Research Foundation, Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, NSFC under Grant U1908209 and 62021001, and ZJNSFC under Grant LQ23F010008." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10906" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.476, + 0.183 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.476, + 0.24 + ], + "angle": 0, + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.476, + 0.311 + ], + "angle": 0, + "content": "[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.476, + 0.354 + ], + "angle": 0, + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.476, + 0.395 + ], + "angle": 0, + "content": "[5] Qi Cai, Yingwei Pan, Ting Yao, Chenggang Yan, and Tao Mei. Memory matching networks for one-shot image recognition. In CVPR, pages 4080-4088, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.476, + 0.437 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.476, + 0.478 + ], + "angle": 0, + "content": "[7] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.476, + 0.522 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.476, + 0.563 + ], + "angle": 0, + "content": "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.476, + 0.647 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.476, + 0.69 + ], + "angle": 0, + "content": "[11] Mohamed Elhoseiny, Babak Saleh, and Ahmed Elgammal. Write a classifier: Zero-shot learning using purely textual descriptions. In ICCV, pages 2584-2591, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.476, + 0.746 + ], + "angle": 0, + "content": "[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.476, + 0.788 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, pages 1126–1135. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.476, + 0.843 + ], + "angle": 0, + "content": "[14] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. NeurIPS, 26, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.476, + 0.901 + ], + "angle": 0, + "content": "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 1, 2, 3, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.476, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[16] Micah Goldblum, Steven Reich, Liam Fowl, Renkun Ni, Valeriiia Cherepanova, and Tom Goldstein. Unraveling meta-learning: Understanding feature representations for few-shot tasks. In ICML, pages 3607-3616. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.15, + 0.906, + 0.206 + ], + "angle": 0, + "content": "[17] Lluis Gomez, Yash Patel, Marçal Rusinol, Dimosthenis Karatzas, and CV Jawahar. Self-supervised learning of visual features through embedding images into text topic spaces. In CVPR, pages 4230-4239, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.208, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.251, + 0.906, + 0.32 + ], + "angle": 0, + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.324, + 0.906, + 0.392 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.395, + 0.906, + 0.437 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.439, + 0.906, + 0.506 + ], + "angle": 0, + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.51, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.582, + 0.906, + 0.623 + ], + "angle": 0, + "content": "[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.626, + 0.906, + 0.668 + ], + "angle": 0, + "content": "[25] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In ECCV, pages 67-84. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.67, + 0.906, + 0.711 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2015.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.714, + 0.906, + 0.755 + ], + "angle": 0, + "content": "[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IC-CVW, pages 554–561, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.758, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.802, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[29] Jimmy Lei Ba, Kevin Swersky, Sanja Fidler, et al. Predicting deep zero-shot convolutional neural networks using textual descriptions. In ICCV, pages 4247-4255, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.845, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[30] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, pages 3045-3059, Online and Punta Cana, Dominican Republic, Nov. 2021. ACL. 3" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10907" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.476, + 0.134 + ], + "angle": 0, + "content": "[31] Ang Li, Allan Jabri, Armand Joulin, and Laurens Van Der Maaten. Learning visual n-grams from web data. In ICCV, pages 4183–4192, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.476, + 0.205 + ], + "angle": 0, + "content": "[32] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. arXiv preprint arXiv:2204.08790, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.476, + 0.261 + ], + "angle": 0, + "content": "[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.476, + 0.303 + ], + "angle": 0, + "content": "[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, pages 4582-4597, Online, Aug. 2021. ACL. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.476, + 0.374 + ], + "angle": 0, + "content": "[35] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.476, + 0.417 + ], + "angle": 0, + "content": "[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.476, + 0.459 + ], + "angle": 0, + "content": "[37] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.476, + 0.488 + ], + "angle": 0, + "content": "[38] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.476, + 0.544 + ], + "angle": 0, + "content": "[39] Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In CVPR, pages 375-383, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.476, + 0.587 + ], + "angle": 0, + "content": "[40] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.476, + 0.63 + ], + "angle": 0, + "content": "[41] Zhihe Lu, Sen He, Da Li, Yi-Zhe Song, and Tao Xiang. Prediction calibration for generalized few-shot semantic segmentation. arXiv preprint arXiv:2210.08290, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.476, + 0.686 + ], + "angle": 0, + "content": "[42] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In ICCV, pages 8741-8750, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.476, + 0.73 + ], + "angle": 0, + "content": "[43] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.476, + 0.772 + ], + "angle": 0, + "content": "[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.476, + 0.815 + ], + "angle": 0, + "content": "[45] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. NeurIPS, 26, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.476, + 0.858 + ], + "angle": 0, + "content": "[46] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.476, + 0.901 + ], + "angle": 0, + "content": "[47] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In IC-CVGIP, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.476, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.907, + 0.133 + ], + "angle": 0, + "content": "[48] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, pages 3498-3505. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.135, + 0.907, + 0.204 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[50] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.247, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, pages 5389-5400. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.288, + 0.907, + 0.329 + ], + "angle": 0, + "content": "[52] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In ECCV, pages 153-170. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.33, + 0.907, + 0.383 + ], + "angle": 0, + "content": "[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.385, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[54] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In EMNLP, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.442, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.484, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[56] Richard Socher, Milind Ganjoo, Christopher D Manning, and Andrew Ng. Zero-shot learning through cross-modal transfer. NeurIPS, 26, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.525, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[57] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.567, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.623, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.679, + 0.907, + 0.719 + ], + "angle": 0, + "content": "[60] Lorenzo Torresani, Martin Szummer, and Andrew Fitzgibbon. Efficient object category recognition using classemes. In ECCV, pages 776-789. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.72, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[61] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.776, + 0.907, + 0.817 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.818, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[63] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[64] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10908" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.476, + 0.119 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.476, + 0.148 + ], + "angle": 0, + "content": "[65] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.151, + 0.476, + 0.178 + ], + "angle": 0, + "content": "[66] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.18, + 0.476, + 0.207 + ], + "angle": 0, + "content": "[67] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.209, + 0.476, + 0.263 + ], + "angle": 0, + "content": "[68] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, pages 10819-10829, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.266, + 0.476, + 0.333 + ], + "angle": 0, + "content": "[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.336, + 0.476, + 0.39 + ], + "angle": 0, + "content": "[70] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.476, + 0.447 + ], + "angle": 0, + "content": "[71] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.476, + 0.49 + ], + "angle": 0, + "content": "[72] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.492, + 0.476, + 0.533 + ], + "angle": 0, + "content": "[73] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.476, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10909" + } + ] +] \ No newline at end of file diff --git a/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_origin.pdf b/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e4363fdc46b192f98b3ec97336720572a4df5217 --- /dev/null +++ b/2023/Task Residual for Tuning Vision-Language Models/7e947285-2912-4e13-91da-b9b910f110fa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:835ee347476f57ed4d5997b646d7662bed3682dab0a200a6d73efa9997b2262c +size 565868 diff --git a/2023/Task Residual for Tuning Vision-Language Models/full.md b/2023/Task Residual for Tuning Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44004986939312751fb7d934bf89e6942df9f012 --- /dev/null +++ b/2023/Task Residual for Tuning Vision-Language Models/full.md @@ -0,0 +1,328 @@ +# Task Residual for Tuning Vision-Language Models + +Tao Yu $^{1,2*}$ Zhihe Lu $^{1*}$ Xin Jin $^{1,2}$ Zhibo Chen $^{2}$ Xinchao Wang $^{1\dagger}$ $^{1}$ National University of Singapore $^{2}$ University of Science and Technology of China + +yutao666@mail.ustc.edu.cn, zhihelu@nus.edu.sg, jinxustc@mail.ustc.edu.cn, chenzhibo@ustc.edu.cn, xinchao@nus.edu.sg + +# Abstract + +Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes. + +# 1. Introduction + +Over the past decade, deep learning-based visual recognition models [10, 18, 28, 55, 58] have achieved great success. These state-of-the-art models are often trained on a large amount of image and discrete label pairs. The discrete label is generated by converting a detailed textual description, e.g., "American curl cat", into a simple scalar, which strikingly eases the computation of loss function. However, this also results in two evident limitations: (i) the rich semantics + +![](images/ec5630472625830893480b27ff0857a22d4d426ea19f25fcd61d1b27c78071fe.jpg) +Figure 1. Performance comparison between Zero-shot CLIP [49], CoOp [73], CLIP-Adapter [15], Tip-Adapter-F [71] and our TaskRes on ImageNet with few shot settings. + +in the textual description are underused, and (ii) the trained models are limited to recognizing the close-set classes only. + +Recent large-scale vision-language model (VLM) pretraining [1, 23, 33, 49, 69] eliminates those limitations by learning visual representations via textual supervision. For instance, texts and images are encoded and mapped into a unified space via a contrastive loss during pre-training [49]. The pre-trained text encoder can then be used to synthesize a text-based classifier for image recognition given the corresponding natural language descriptions as shown in Figure 2 (a). Those pre-trained VLMs have demonstrated a powerful transferability on a variety of downstream tasks in a zero-shot manner. However, the effectiveness of the aforementioned models heavily relies on their large-scale architectures and training datasets. For instance, CLIP [49] has up to 428 million parameters and is trained on 0.4 billion text-image pairs, while Flamingo [1] boasts up to 80 billion parameters and is trained on a staggering 2.1 billion pairs. This makes it impractical to fully fine-tune the model on downstream tasks in a low-data regime. + +For that reason, efficient transfer learning (ETL) [15, 71-73] on pre-trained VLMs has gained popularity. ETL represents transfer learning to downstream tasks in both + +![](images/b34b2bbb64c09659bf9d7071e34c573312b6e19ec4dcef6b02745c33246da132.jpg) +(a) Zero-shot CLIP + +![](images/2086ad88a7c944b314b658e80da43ffcbf9d48a5eaeb1f050ee86b65e1327d52.jpg) +(b) Prompt Tuning + +![](images/c0ae072c1d6bee0fcc19c326e4d83304df18069ccc354f174291732cbb2d830f.jpg) +(c) Adapter-style Tuning + +![](images/6fc3168e0f3d22bbaf220920fb5eccc3310dd2c71387d13b24ca4ed97827b6ea.jpg) +(d) Task Residual Tuning +Figure 2. Illustration of (a) Zero-shot CLIP, (b) prompt tuning, (c) adapter-style tuning and (d) our proposed Task Residual Tuning (TaskRes). Our method introduces a prior-independent task residual to the fixed pre-trained classifier (i.e., text embeddings of CLIP), being free of running the text encoder every time or extra architecture design. + +parameter- and data-efficient manner. The core of ETL is twofold: (i) properly inheriting the well-learned knowledge structure of VLMs, which is already transferable; (ii) effectively exploring the task-specific knowledge given limited data. However, most existing ETL approaches, i.e., prompt tuning (PT) [72, 73] and adapter-style tuning (AT) [15, 71], either damage the prior knowledge of VLMs or learn the new knowledge of a task in an inappropriate/insufficient way. For example, instead of using the pre-trained text-based classifier, CoOp [73] (in Figure 2 (b)) is proposed to learn a continuous prompt for synthesizing a completely new one, which inevitably causes the loss of previous knowledge. Consequently, CoOp underperforms Zero-shot CLIP by $1.03\% / 0.37\%$ in 1-/2-shot learning on ImageNet (see Figure 1). In contrast, CLIP-Adapter [15] preserves the pretrained classifier, but is excessively biased towards the prior knowledge when learning a new task, i.e., it transforms the pre-trained classifier weights to be task-specific as illustrated in Figure 2 (c). This results in an inferior new knowledge exploration, thereby a lower accuracy as shown in Figure 1. + +For better ETL on pre-trained VLMs, we propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the old knowledge of the pretrained models and the new knowledge for a target task. The rationale is that the decoupling enables a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration, i.e., the learned knowledge w.r.t. the task is independent on the old knowledge. Specifically, TaskRes keeps the original classifier weights frozen and introduces a set of prior-independent parameters that are added to the weights. These additive parameters, tuned for adaptation to the target task, are thus named "task residual". + +To gain insight into how TaskRes works, we perform ex + +tensive experiments across 11 benchmark datasets [73] and conduct a systematic investigation of learned task residuals. The experimental results demonstrate that introducing task residual can significantly enhance the transfer performance. We visualize the correlation between the magnitude of learned task residual and the difficulty of transferring a pre-trained model to a downstream task, and observe that the magnitude increases with the transfer difficulty. This suggests the residual is automatically adapted to the task to fully explore the new knowledge, thereby achieving a new state-of-the-art performance on 11 diverse datasets. Furthermore, it is worth noting that our method requires minimal effort for the implementation, i.e., technically adding one line of code only. Our contributions are summarized below: + +- We for the first time emphasize the necessity of a proper knowledge inheritance from pre-trained VLMs to downstream tasks via ETL, reveal the pitfalls of existing tuning paradigms, and conduct an in-depth analysis to manifest that decoupling the old pre-trained knowledge and the new task-specific knowledge is the key. +- We propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which achieves a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration. +- TaskRes is convenient for use, which needs a few tuning parameters and effortless implementation. + +# 2. Related Work + +# 2.1. Vision-Language Models + +We mainly review the literature of vision-language models (VLMs) on language-driven visual representation learning (LDVRL) [2,9,11,14,17,25,29,31,49,52,56]. The core of + +LDVRL is to map texts (language) and images into a common space so that the text representations can be used for visual classification. To achieve this, two encoders for texts and images respectively, and specific loss functions as regularization are needed. + +Early works explore unsupervised pre-trained models [56] or the skip-gram text modeling [14, 44, 45] for text embedding while sparse coding and vector quantization [7, 56] or Classeme features [11, 60] for visual encoding. For training objectives, MSE [56], self-supervised topic probabilities matching [17] or multi-class logistic loss are adopted. However, those methods are limited to small datasets [9, 25, 31] and weakly representative encoding backbones [7, 56], which tremendously hinders their transferability. + +In contrast, recent works [23, 35, 49] have advanced previous ones by leveraging billion-level data of image-text pairs from internet and super powerful neural networks, e.g., Transformers [10, 61, 62, 68, 70], for representation learning. By end-to-end pre-training with a contrastive loss, large-scale VLMs demonstrate remarkable transferability on various downstream tasks in a zero-shot evaluation manner. Standing on the strong transfer capability of large-scale vision-language pre-training, we further explore its potential in efficient transfer learning. + +# 2.2. Efficient Transfer Learning + +Pre-training a neural network on large-scale datasets, e.g., ImageNet [8] and WebImageText [49], then fine-tuning it on downstream tasks has been a common step of transfer learning [36-38, 65-67]. Here, we mainly focus on efficient transfer learning (ETL) on pre-trained VLMs. ETL represents both parameter- and data-efficient transfer learning, in which a small number of parameters are tuned and a small set of data is utilized. Existing works on ETL can be grouped into two categories: prompt tuning [40, 72, 73] and adapter-style tuning (AT) [15, 71]. Specifically, prompt engineering [24, 54] is first explored to generate appropriate discrete prompts for downstream tasks. Later, learning continuous prompts [39, 73] that can better adapt to a new task shows a more advanced performance. However, these methods face two issues: (i) they need the pre-trained text encoder to engage in the whole training, limiting its scalability and increasing the computational overhead; (ii) they abandon the well-pre-trained text-based classifier and generate a new one, which causes the loss of prior knowledge of VLMs. + +AT-based methods [15, 71] solve the above issues by using the text encoder once for text-based classifier generation and then focusing on adapting text/image features only. This straightforward design can achieve even better performance, but it heavily relies on the prior knowledge/pre-trained features, which causes an inferior new knowledge exploration. To address this, we propose a new efficient tuning approach, Task Residual Tuning (TaskRes), for a better and more flexi + +ble task-specific knowledge learning by a prior-independent "task residual". + +# 2.3. Few-Shot Learning + +Few-shot learning (FSL) aims to adapt a model to novel tasks/classes with a handful of labeled examples. Conventional FSL methods [5, 13, 16, 41, 42, 53] often meta-learn on abundant data from base classes for adaptive capability. However, the need of learning on base datasets restricts their scalability. Recent VLM pre-training works [23, 49] offer an effective alternative, which does not require a base dataset. They show that the pre-trained models can already achieve astonishing performance on many downstream tasks in a zero-shot manner. ETL [15, 73] can further boost the performance. In this work, we propose a novel ETL method for adapting VLMs to downstream tasks and evaluate its effectiveness on few-shot tasks. + +# 3. Preliminaries + +We briefly introduce the adopted VLM, i.e., contrastive language-image pre-training (CLIP) [49], and recap two mainstream approaches for ETL on VLMs, i.e., prompt tuning and adapter-style tuning. + +# 3.1. Contrastive Language-Image Pre-training + +The CLIP model [49] is designed to obtain visual representations through natural language supervision. It is trained on 0.4 billion image-text pairs, where image features from an image encoder and text features from a text encoder are aligned within a unified embedding space using a contrastive learning loss, allowing CLIP to effectively capture broad visual concepts and learn general visual representations. When testing, CLIP can classify a query image into $K$ possible categories. This is achieved by calculating the cosine similarity between a query image embedding $\mathbf{z}$ , obtained from the image encoder with projection, and the text embeddings $\{\mathbf{t}_i\}_{i=1}^K$ , which are derived by inputting texts (e.g., "a photo of a {class}") into the text branch. The predicted probability for class $i$ is formulated as + +$$ +p (y = i | \mathbf {z}) = \frac {\exp (\sin (\mathbf {z} , \mathbf {t} _ {i}) / \tau)}{\sum_ {j = 1} ^ {K} \exp (\sin (\mathbf {z} , \mathbf {t} _ {j}) / \tau)}, \tag {1} +$$ + +where $\sin (\cdot ,\cdot)$ indicates cosine similarity and $\tau$ is the learned temperature of CLIP. + +# 3.2. Revisiting Previous Tuning Paradigms + +Inspired by the success of ETL approaches in natural language processing, e.g., prompt tuning [30, 34] and Adapter [22], recent advances (e.g., CoOp [73] and CLIP-Adapter [15]) borrow their ideas to ETL on VLMs. + +CoOp, for the first time, introduces prompt tuning to VLMs. Instead of using fixed text prompt contexts such as + +"a photo of a", CoOp proposes to use $M$ learnable context vectors $\{\mathbf{v}_m\}_{m=1}^M$ as the task-specific templates. The prompt given to the text encoder then becomes $\{\mathbf{v}_1, \mathbf{v}_2, \dots, \mathbf{v}_M, \mathbf{c}_i\}$ where $\mathbf{c}_i$ is the embedding of class $i$ . During the whole training process, CoOp keeps the parameters of pre-trained VLMs frozen and only tunes the learnable vectors $\{\mathbf{v}_m\}_{m=1}^M$ . + +Adapter-style tuning introduces additional modules $\phi_{\omega}(\cdot)$ with tunable parameters $\omega$ to the pre-trained models to transform the pre-trained features $\mathbf{f}$ to new ones $\mathbf{f}'$ . In general, adapter-style tuning can be formulated as + +$$ +\mathbf {f} ^ {\prime} = \mathbf {f} + \alpha \phi_ {\omega} (\mathbf {f}), \tag {2} +$$ + +where $\alpha$ is a scaling factor. In CLIP-Adapter, the adapter module $\phi_{\omega}$ is composed of two linear transformation layers and a ReLU activation between them. CLIP-Adapter investigates both visual and text adapters, i.e., applying the adapter module to the image and text branches of CLIP, respectively, and shows that they have comparable performance. During training on downstream tasks, adapter-style methods only tune their adapter modules. + +# 4. Approach + +In this section, we start by identifying the pitfalls of the existing ETL paradigms on VLMs. Those pitfalls motivate us to propose a new ETL approach for VLMs, named Task Residual Tuning (TaskRes). The proposed TaskRes is simple yet can effectively avoid the pitfalls. + +# 4.1. Pitfalls of Existing ETL Paradigms on VLMs + +We rethink the usage of the prior knowledge from pretrained VLMs and the acquisition of the new knowledge regarding downstream tasks. On the one hand, large-scale VLMs trained with a vast amount of data have learned broad visual concepts that are general for a wide range of downstream vision tasks, enabling homogenization [3]. The prior knowledge should be well preserved when performing a transfer. On the other hand, despite the huge data used in the pre-training, inevitably, there are domain shifts or uncertain concepts in downstream tasks. New knowledge specific to the downstream tasks should be appropriately supplemented to the prior knowledge. However, existing ETL paradigms do not well consider the above principles and have the following two issues. + +Pitfall 1: Lack of guarantees of prior knowledge preservation in prompt tuning. While the weights of pre-trained text branch modules (e.g., text encoder and projection) are frozen in the prompt tuning paradigm, the original well-learned classification boundary is more or less damaged. This is because the tuning of input prompts ends up with a new boundary that may forget the old knowledge without explicit regularization. As a result, the performance of prompt tuning is limited. For instance, the performance of + +CoOp [73] is not as good as Zero-shot CLIP in 1-/2-shot learning on ImageNet as shown in Figure 1. + +Pitfall 2: Limited flexibility of new knowledge exploration in adapter-style tuning. The data distributions in downstream tasks often shift from the pre-training distribution, and some task-specific or fine-grained visual concepts/representations may not be well learned by pre-trained VLMs, e.g., from CLIP [49] to satellite image dataset EuroSAT [19]. Thus, the new knowledge regarding downstream tasks needs to be appropriately explored. We observe that adapter-style tuning may not sufficiently explore the task-specific knowledge as the input of the adapter is strictly limited to old/pre-trained features, as shown in Figure 2 (c). Regardless of whether the pre-trained features are suitable for the task, the results of the adapter depend only on them, making adapter-style tuning have limited flexibility for learning new knowledge. + +# 4.2. Task Residual Tuning + +Given that the existing ETL paradigms face the aforementioned issues, we propose Task Residual Tuning (TaskRes) to address the issues in a simple way. TaskRes explicitly decouples the maintenance of the old knowledge from the pre-trained VLMs and the learning of task-specific knowledge that is not overly biased on the pre-trained features. We elaborate our TaskRes below. + +Fixed base classifier. As illustrated in Figure 2 (d), our TaskRes performs the tuning directly on the text-based classifier (i.e., text embeddings)1. Base classifier is the text embeddings of a pre-trained vision-language model, e.g., CLIP. We denote the base classifier as $\mathbf{t} \in \mathbb{R}^{K \times D}$ where $K$ is the number of categories and $D$ is the feature dimension. We keep the base classifier weights frozen to explicitly prevent the base classifier from being damaged. + +Prior-independent task residual. To learn task-specific knowledge without being restricted by prior knowledge, we propose task residual, which is a set of tunable parameters $\mathbf{x} \in \mathbb{R}^{K \times D}$ that are not dependent on the base classifier. Our task residual is scaled by a factor $\alpha$ and element-wisely added to the base classifier to formulate a new classifier $\mathbf{t}'$ for the target task, written as + +$$ +\mathbf {t} ^ {\prime} = \mathbf {t} + \alpha \mathbf {x}. \tag {3} +$$ + +Tuning for downstream tasks. During tuning, we only tune the prior-independent task residual while keeping the base classifier (together with the image branch) fixed, enabling reliable old knowledge preservation and flexible new + +knowledge exploration. Given an image, the fixed image branch of CLIP extracts its embeddings $\mathbf{z} \in \mathbb{R}^D$ . The predicted probability for class $i$ is then calculated as + +$$ +p (y = i | \mathbf {z}) = \frac {\exp \left(\operatorname {s i m} \left(\mathbf {z} , \mathbf {t} _ {i} ^ {\prime}\right) / \tau\right)}{\sum_ {j = 1} ^ {K} \exp \left(\operatorname {s i m} \left(\mathbf {z} , \mathbf {t} _ {j} ^ {\prime}\right) / \tau\right)}. \tag {4} +$$ + +Based on the predicted probabilities, the downstream task loss (e.g., cross-entropy loss) only updates the task residual through standard backpropagation. + +# 5. Experiment + +# 5.1. Setup + +We follow prior efficient transfer learning (ETL) work [15, 71, 73] to conduct a few-shot evaluation for ETL models on 11 benchmark datasets i.e., ImageNet [8], Caltech101 [12], OxfordPets [48], StanfordCars [27], Flowers102 [47], Food101 [4], FGVCAircraft [43], SUN397 [64], DTD [6], EuroSAT [19] and UCF101 [57]. Those datasets include a wide range of visual recognition tasks, e.g., classifying general objects, fine-grained objects, actions, scenes, etc. Specifically, ETL models are trained using $1/2/4/8/16$ shots per class from the training sets, respectively, and are tested on the full test sets. Additionally, following CoOp [73], we test the generalization performance of our models from ImageNet to its variants (ImageNetV2 [51], ImageNet-Sketch [63], ImageNet-A [21] and ImageNet-R [20]). + +# 5.2. Implementation + +Our approach has two main components, i.e., base classifier and task residual. For the base classifier, we develop two versions: regular base classifier that directly adopts the weights of the text embeddings of the pre-trained CLIP (using pre-defined text templates following prior work [49, 71]), and enhanced base classifier obtained by tuning the text projection layer of CLIP on the target task before starting our task residual tuning $^2$ . Task residual is a matrix filled with learnable parameters that are initialized by zeros. Task residual has the same shape as the base classifier. We scale the task residual and element-wisely add it to the base classifier, as in Eq. 3. By default, the scaling factor $\alpha$ is set to 0.5 for all datasets except for Flowers102 using 1. Additionally, we explore the use of a learnable $\alpha$ later in our ablation study. All models in this work are built upon the pre-trained CLIP models. Unless otherwise specified, we use the ResNet-50 [18] version of CLIP which has a ResNet-50 backbone in the image branch. We train our models for 100 epochs for 1-/2-/4-shot experiments and 200 epochs for 8-/16-shot experiments, with batch size 256. Besides, the aforementioned enhanced base classifier is tuned for 50 epochs. Our models + +are optimized by Adam [26] with an initial learning rate of $2e - 3$ except for ImageNet using $2e - 4$ . Following CoOp [73], our optimization process adopts a cosine learning rate decay schedule and a warmup scheme (i.e., fixing the learning rate in the first epoch to $1e - 5$ ). We test the performance of models after training, and all experimental results are averaged over three random seeds. + +# 5.3. Performance Comparison + +# 5.3.1 Few-Shot Learning + +We develop two versions of our TaskRes, i.e., Ours/Ours* using the regular/enhanced base classifier, respectively. We compare our models with Zero-shot CLIP [49], together with the state-of-the-art (SOTA) ETL methods including CoOp [73], CLIP-Adapter [15] and Tip-Adapter-F $[71]^3$ . The comparison results of the 11 benchmark datasets are shown in Figure 3. Overall, both Ours and Ours* achieve the SOTA/SOTA-comparable performance on the 11 datasets across all few-shot settings and significantly surpass Zero-shot CLIP [49]. Besides, a few observations can be made. First, with more shots, Tip-Adapter-F reaches a closer performance to ours, but as shown in Figure 1 the number of its tunable parameters increases linearly with that of shots, substantially limiting its scalability. Second, our method shows inferior performance on OxfordPets and Food101. Training with a few shots on these two fine-grained datasets is prone to be overfitting, which is also found in CoOp [73]. Finally, we found the proposed method is superior when needed to recognize a large amount of classes, e.g., 1000 classes of ImageNet and 397 classes of SUN397 in 1-shot case. This is because our TaskRes, not excessively biased to the pre-trained features, can better explore the task-specific knowledge even in such extreme scenarios. + +# 5.3.2 Domain Generalization + +As pointed out by CoOp [73], ETL models that are trained on a specific domain are at risk of learning spurious correlations when generalizing to unseen domains. We evaluate our models concerning domain generalization. We train our models on 16-shot ImageNet and test the generalization performance of the trained models on four unseen ImageNet variant datasets (ImageNet-V2, -Sketch, -A and -R). As shown in Table 1, our models consistently outperform the compared models (Zero-shot CLIP [49], Linear Probe CLIP [49] and CoOp [73]) across various CLIP visual backbones (ResNet-50 [18], ResNet-101 [18], ViT-B/32 [10] and ViT-B/16 [10]). We also observe that our TaskRes using the enhanced base classifier (Ours*) achieves higher accuracy than that using the regular base classifier (Ours) on the source dataset while + +![](images/6afc4c5d957958795bd4b21505c85970a89696ffaff6333c66cd5658c77763cc.jpg) + +![](images/4490e6b3ce7dfe329a760fea469a63a4fd9a33ad15b0453cc61619991bbccd91.jpg) + +![](images/2cc10b7a01c9b612c6160b35dd7948090f2192a61e243fd79bebf3d03f0637d5.jpg) + +![](images/9b7624f0d791b0a8cb0f20b00af6dd5db577eb6906b00667ac1ffc476975784c.jpg) + +![](images/bee1ed9a2cf628ea56787d01fd0a17b103f4f7de2e0f139642820ee49c5b6975.jpg) + +![](images/20689d8536aabae615cbcf8939ac94588553abe587ab383309f31a6484beee6d.jpg) + +![](images/56cd09bd822811baa963386ed0a7772fe94eea0e9ad11a1477972841d63f421a.jpg) + +![](images/5e74d59325f2293f8d11a23d3e6ee17d57b8b69056021cb9e101dd0f8cbf6cfe.jpg) + +![](images/f14bf08663271e41629dd0a5ec2fe7113f7704d09d8944ec3a1d9347d2671d08.jpg) + +![](images/d36ea553e962bd64ad5891c35bd98e7cd94f00c57ee77974a284e4509492b066.jpg) +Figure 3. Performance comparison on few-shot learning, i.e., 1-/2-/4-/8-/16-shot, on 11 benchmark datasets. The top-left is the averaged accuracy over the 11 datasets. The full numerical results can be found in the supplementary. + +![](images/0b9614b1d271812ede7ebf39fb32cf34d8b0a7bcf5e303754e8aaf4fb1a31e7d.jpg) + +![](images/633f7e1b0e220eff428c83eef895a978db2766201c16f3184b888df3588369da.jpg) + +reducing a certain degree of generalization, which is due to a slight overfitting on the source dataset. + +# 5.4. Ablation Study + +Importance of prior knowledge preservation and prior-independent tuning. We build four target classifiers $\mathbf{t}'$ to investigate the importance of prior knowledge preservation and prior-independent tuning (where the tunable parameters are not dependent on the pre-trained features), i.e., original + +pre-trained classifier $\mathbf{t}' = \mathbf{t}$ (i.e., our regular base classifier), directly-adapted classifier $\mathbf{t}' = \phi_{\omega}(\mathbf{t})$ , adapter-style classifier $\mathbf{t}' = \mathbf{t} + \alpha \phi_{\omega}(\mathbf{t})$ and our TaskRes classifier $\mathbf{t}' = \mathbf{t} + \alpha \mathbf{x}$ . We conduct experiments on ImageNet across all few-shot settings, using the same scaling factor $\alpha = 0.5$ . The results in Table 2 support our motivations that (i) explicit preservation of prior knowledge matters as the directly-adapted classifier performs much worse than the original pre-trained classifier and adapter-style classifier, and (ii) tuning prior- + +
MethodVisual BackboneSourceTarget
ImageNet-V2-Sketch-A-RAverage
Zero-Shot CLIP [49]58.1851.3433.3221.6556.0040.58
Linear Probe CLIP [49]55.8745.9719.0712.7434.8628.16
CoOp [73]ResNet-5062.9555.1132.7422.1254.9641.23
Ours64.7556.4735.8322.8060.7043.95
Ours*65.7357.0034.4321.5058.1342.77
Zero-Shot CLIP [49]61.6254.8138.7128.0564.3846.49
Linear Probe CLIP [49]59.7550.0526.8019.4447.1935.87
CoOp [73]ResNet-10166.6058.6639.0828.8963.0047.41
Ours67.7059.5041.7029.8768.0749.79
Ours*68.7360.0040.328.0064.8048.28
Zero-Shot CLIP [49]62.0554.7940.8229.5765.9947.79
Linear Probe CLIP [49]59.5849.7328.0619.6747.2036.17
CoOp [73]ViT-B/3266.8558.0840.4430.6264.4548.40
Ours68.2059.2042.5031.4369.3350.62
Ours*69.1759.4740.8729.7066.2749.08
Zero-Shot CLIP [49]66.7360.8346.1547.7773.9657.18
Linear Probe CLIP [49]65.8556.2634.7735.6858.4346.29
CoOp [73]ViT-B/1671.9264.1846.7148.4174.3258.41
Ours73.0765.3049.1350.3777.7060.63
Ours*73.9065.8547.7049.1775.2359.49
+ +independent parameters is more effective than the dependent one as the TaskRes classifier obviously surpasses the adapter-style classifier. + +Table 1. Performance comparison on generalization (from ImageNet to ImageNet-V2/-Sketch/-A/-R) with multiple CLIP visual backbones. + +
Target Classifier t'1-shot2-shot4-shot8-shot16-shot
t60.3360.3360.3360.3360.33
φω(t)28.0334.9042.5748.1754.13
t + αφω(t)61.0361.2361.2762.0063.17
t + αx61.4362.1762.9364.0364.75
+ +Effectiveness of task residual learning. We conduct an ablation study of the effect of introducing our task residual learning (TaskRes). Table 3 shows the few-shot results averaged over 11 datasets of the regular base classifier and that equipped with our TaskRes (Ours), together with the enhanced counterparts. Our TaskRes notably improves the regular base, e.g., with $16.14\%$ accuracy gain in 16-shot case. Similarly, the performance gain is also impressive even on the enhanced base. This indicates that TaskRes can always benefit a pre-trained base when adapted to downstream tasks. + +Table 2. Ablation study of various combinations of three main components: base classifier $\mathbf{t}$ , adapter module ${\phi }_{\omega }\left( \cdot \right)$ and our "task residual" $\mathbf{x}$ ,for building a target classifier ${\mathbf{t}}^{\prime }$ . The experiment is conducted on ImageNet. + +
Classifier1-shot2-shot4-shot8-shot16-shot
Regular Base58.9658.9658.9658.9658.96
Regular Base + TaskRes64.0467.0269.6172.7675.10
Enhanced Base61.1162.8264.4469.9173.09
Enhanced Base + TaskRes64.2867.5570.2873.3575.78
+ +Variants of TaskRes. The proposed "task residual" can also be applied to the image branch of CLIP, denoted as TaskRes-I. We also denote the default TaskRes applied on the text branch as TaskRes-T (i.e., Ours) and on the both branches as TaskRes-I&T. We compare their performance on ImageNet in Table 4. TaskRes-T consistently outperforms TaskRes-I, as shown in Table 4. The performance of TaskRes-I increases much slower than TaskRes-T as the number of shots increases. This is because TaskRes-I tends to encounter overfitting where training and testing image embeddings are different. Tuning on text embeddings can avoid the issue. Additionally, the diversity of image embeddings is larger than that of text embeddings, making it difficult to learn task-level information. TaskRes-I&T performs slightly worse than TaskRes-T. This might be caused by a misalignment between the updated image embeddings and our base classifier. + +Table 3. Exploring the effectiveness of the proposed task residual tuning (TaskRes). The accuracy results averaged on 11 datasets are reported under all shots. Refer to Appendix 2.3 for the full comparison results of the 11 datasets. + +
Model1-shot2-shot4-shot8-shot16-shot
Regular Base60.3360.3360.3360.3360.33
TaskRes-I61.2061.2761.4761.5361.60
TaskRes-T61.4362.1762.9364.0364.75
TaskRes-I&T61.5362.1062.7363.8364.67
+ +Table 4. Performance comparison of three TaskRes variants on ImageNet: TaskRes applied on CLIP's image branch (TaskRes-I), text branch (TaskRes-T) and both branches (TaskRes-I&T). + +Scaling factor. We investigate the effect of scaling factor $\alpha$ and show the results of both manually assigned values and a learned value in Table 5. For the manually set $\alpha$ , we observe a notable improvement with TaskRes scaled even by 0.1, while the highest accuracy is obtained with $\alpha = 0.5$ . In contrast, the learnable $\alpha$ can further boost the performance. We show more analyses and results in Appendix 2.3. + +
α00.10.30.50.71.0Learned
Acc.58.9661.8663.7663.9763.6763.0264.01
+ +Visual backbone. We further evaluate models on various CLIP visual backbones, namely, ResNet-50 (RN50), ResNet-101 (RN101), ViT-B/32 and ViT-B/16, as shown in Table 6. Our method shows consistent superiority against other alternatives regardless of which backbone is used. + +Table 5. Ablation study of scaling factor $\alpha$ . The results are averaged over 11 datasets in 1-shot case. + +
MethodRN50RN101ViT-B/32ViT-B/16
Zero-shot CLIP [49]58.1861.6262.0566.73
CoOp [73]62.9566.6066.8571.92
CLIP-Adapter [15]63.5965.3966.1971.13
Tip-Adapter-F [71]65.4468.5668.6573.69
Ours64.7567.7068.2073.07
Ours*65.7368.7369.1773.90
+ +Table 6. Results of CLIP visual backbones on 16-shot ImageNet. + +# 5.5. Investigation of Learned Task Residual + +We investigate whether the learned task residuals are in fact related to the difficulty of transferring CLIP to downstream tasks. We first define relative transfer difficulty to assess the difficulty of transferring a pre-trained model to a target task. Concretely, the relative transfer difficulty is defined as the ratio of baseline classifier precision to the zero-shot precision of the pre-trained model. We use a random classifier to serve as the baseline classifier where the precision of it equals to $1 / K$ ( $K$ is the number of classes of the downstream task). We then can calculate the relative transfer difficulty of CLIP regarding the 11 benchmark datasets based on the random classifier and Zero-shot CLIP. We count the average magnitudes of the learned task residual of each dataset in 16-shot setting and find that they are highly related to relative transfer difficulty, as shown in Figure 4. This verifies the reliability of the proposed task residual. There are some interesting findings that may be counter-intuitive, e.g., transferring CLIP to ImageNet (with 1000 classes) is the simplest and to EuroSAT (with 10 classes) is the most difficult. This is because the inherent difficulty of ImageNet is large but the data distribution and containing objects are "familiar" to CLIP, whereas EuroSAT is the opposite. Besides, there is more space for TaskRes to play its role in the tasks with large relative transfer difficulty, e.g., EuroSAT (1-shot accuracy gain: $23.71\%$ ) and DTD (1-shot accuracy gain: $7.85\%$ ). More results and analyses can be found in Appendix 2.4. + +# 6. Conclusion, Limitation and Future Work + +In this work, we propose a new approach for tuning VLMs, i.e., TaskRes. The proposed TaskRes performs efficient transfer learning (ETL) on VLMs by explicitly de + +![](images/25727f268e1c570ff846a98cfe4debd7b16cc218d992ae03e5b14a0cd5fc1f43.jpg) +Figure 4. The magnitude of learned task residual increases with relative transfer difficulty (in logarithmic scale). The shadow indicates the standard deviation regarding random seeds. + +coupling the classifier into two key parts: undamaged base classifier with rich prior knowledge and task residual independent to the base classifier for better exploring task-specific knowledge. Interestingly, the magnitude of learned task residual is highly related to the difficulty of transferring pre-trained VLMs to a target downstream task. This may inspire the community to consider ETL from a new perspective, e.g., modeling "task-to-task transfer difficulty" for ETL. Extensive experiments have demonstrated the effectiveness of our TaskRes, despite its simplicity. + +Nevertheless, our method bears some limitations. For instance, our method encounters negative transfer on two datasets, i.e., OxfordPets (1-shot) and Food101 (1-/2-/4-/8-shot). We conjecture that this happens with two conditions: (i) the downstream tasks are of high relative transfer difficulty as shown in Figure 4, and (ii) Zero-shot CLIP has already achieved a fair precision on them. Besides, the assessment of transfer difficulty in this work is heuristic. + +With the rapid advancement of foundation models [3], it is becoming increasingly crucial to establish precise and reliable metrics for assessing the transfer difficulty of pretrained foundation models to downstream tasks. A comprehensive study of transfer difficulty, which includes distribution analyses [32], is in high demand. Moreover, we can extend the transfer difficulty to a concept-wise level, and investigate the correlation between performance and the occurrence frequency of visual concepts by exploring the CLIP models trained on specific datasets, e.g., SLIP [46] trained on YFCC15M [50, 59]. We will explore these in future work. + +# Acknowledgment + +This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006), the National Research Foundation, Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, NSFC under Grant U1908209 and 62021001, and ZJNSFC under Grant LQ23F010008. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1 +[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 2 +[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 4, 8 +[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461. Springer, 2014. 5 +[5] Qi Cai, Yingwei Pan, Ting Yao, Chenggang Yan, and Tao Mei. Memory matching networks for one-shot image recognition. In CVPR, pages 4080-4088, 2018. 3 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5 +[7] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, 2011. 3 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3, 5 +[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2, 3 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 3, 5 +[11] Mohamed Elhoseiny, Babak Saleh, and Ahmed Elgammal. Write a classifier: Zero-shot learning using purely textual descriptions. In ICCV, pages 2584-2591, 2013. 2, 3 +[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, pages 1126–1135. PMLR, 2017. 3 +[14] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. NeurIPS, 26, 2013. 2, 3 +[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 1, 2, 3, 5, 8 + +[16] Micah Goldblum, Steven Reich, Liam Fowl, Renkun Ni, Valeriiia Cherepanova, and Tom Goldstein. Unraveling meta-learning: Understanding feature representations for few-shot tasks. In ICML, pages 3607-3616. PMLR, 2020. 3 +[17] Lluis Gomez, Yash Patel, Marçal Rusinol, Dimosthenis Karatzas, and CV Jawahar. Self-supervised learning of visual features through embedding images into text topic spaces. In CVPR, pages 4230-4239, 2017. 2, 3 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5 +[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 4, 5 +[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5 +[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5 +[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799. PMLR, 2019. 3 +[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 3 +[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3 +[25] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In ECCV, pages 67-84. Springer, 2016. 2, 3 +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2015.5 +[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IC-CVW, pages 554–561, 2013. 5 +[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1 +[29] Jimmy Lei Ba, Kevin Swersky, Sanja Fidler, et al. Predicting deep zero-shot convolutional neural networks using textual descriptions. In ICCV, pages 4247-4255, 2015. 2 +[30] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, pages 3045-3059, Online and Punta Cana, Dominican Republic, Nov. 2021. ACL. 3 + +[31] Ang Li, Allan Jabri, Armand Joulin, and Laurens Van Der Maaten. Learning visual n-grams from web data. In ICCV, pages 4183–4192, 2017. 2, 3 +[32] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. arXiv preprint arXiv:2204.08790, 2022. 8 +[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 1 +[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, pages 4582-4597, Online, Aug. 2021. ACL. 3 +[35] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 3 +[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3 +[37] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In NeurIPS, 2022. 3 +[38] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In CVPR, 2023. 3 +[39] Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In CVPR, pages 375-383, 2017. 3 +[40] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3 +[41] Zhihe Lu, Sen He, Da Li, Yi-Zhe Song, and Tao Xiang. Prediction calibration for generalized few-shot semantic segmentation. arXiv preprint arXiv:2210.08290, 2022. 3 +[42] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In ICCV, pages 8741-8750, 2021. 3 +[43] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 3 +[45] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. NeurIPS, 26, 2013. 3 +[46] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8 +[47] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In IC-CVGIP, pages 722-729. IEEE, 2008. 5 + +[48] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, pages 3498-3505. IEEE, 2012. 5 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7, 8 +[50] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 8 +[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, pages 5389-5400. PMLR, 2019. 5 +[52] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In ECCV, pages 153-170. Springer, 2020. 2 +[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3 +[54] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In EMNLP, 2020. 3 +[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[56] Richard Socher, Milind Ganjoo, Christopher D Manning, and Andrew Ng. Zero-shot learning through cross-modal transfer. NeurIPS, 26, 2013. 2, 3 +[57] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 1 +[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 8 +[60] Lorenzo Torresani, Martin Szummer, and Andrew Fitzgibbon. Efficient object category recognition using classemes. In ECCV, pages 776-789. Springer, 2010. 3 +[61] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 3 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3 +[63] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5 +[64] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society + +conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5 +[65] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In ECCV, 2022. 3 +[66] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In NeurIPS, 2022. 3 +[67] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In CVPR, 2023. 3 +[68] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, pages 10819-10829, 2022. 3 +[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1 +[70] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 3 +[71] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 1, 2, 3, 5, 8 +[72] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3 +[73] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 5, 7, 8 \ No newline at end of file diff --git a/2023/Task Residual for Tuning Vision-Language Models/images.zip b/2023/Task Residual for Tuning Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0d78065c2776f518db26bc37eaee12058afd75c9 --- /dev/null +++ b/2023/Task Residual for Tuning Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd8cc92751713768e41adf1b97d2259cb476ce13040109de86623560a2370136 +size 513520 diff --git a/2023/Task Residual for Tuning Vision-Language Models/layout.json b/2023/Task Residual for Tuning Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3bcf4b18d21337399999c1db11f4c3f53fd7f91a --- /dev/null +++ b/2023/Task Residual for Tuning Vision-Language Models/layout.json @@ -0,0 +1,8943 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 459, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 459, + 121 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 459, + 121 + ], + "type": "text", + "content": "Task Residual for Tuning Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": "Tao Yu" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": " Zhihe Lu" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": " Xin Jin" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": " Zhibo Chen" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": " Xinchao Wang" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 503, + 172 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 174, + 497, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 174, + 497, + 199 + ], + "spans": [ + { + "bbox": [ + 102, + 174, + 497, + 199 + ], + "type": "text", + "content": "yutao666@mail.ustc.edu.cn, zhihelu@nus.edu.sg, jinxustc@mail.ustc.edu.cn, chenzhibo@ustc.edu.cn, xinchao@nus.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 227, + 194, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 227, + 194, + 239 + ], + "spans": [ + { + "bbox": [ + 145, + 227, + 194, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 293, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 293, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 293, + 552 + ], + "type": "text", + "content": "Large-scale vision-language models (VLMs) pre-trained on billion-level data have learned general visual representations and broad visual concepts. In principle, the well-learned knowledge structure of the VLMs should be inherited appropriately when being transferred to downstream tasks with limited data. However, most existing efficient transfer learning (ETL) approaches for VLMs either damage or are excessively biased towards the prior knowledge, e.g., prompt tuning (PT) discards the pre-trained text-based classifier and builds a new one while adapter-style tuning (AT) fully relies on the pre-trained features. To address this, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task. Specifically, TaskRes keeps the original classifier weights from the VLMs frozen and obtains a new classifier for the target task by tuning a set of prior-independent parameters as a residual to the original one, which enables reliable prior knowledge preservation and flexible task-specific knowledge exploration. The proposed TaskRes is simple yet effective, which significantly outperforms previous ETL methods (e.g., PT and AT) on 11 benchmark datasets while requiring minimal effort for the implementation. Our code is available at https://github.com/geekyutao/TaskRes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 588, + 291, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 291, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 291, + 684 + ], + "type": "text", + "content": "Over the past decade, deep learning-based visual recognition models [10, 18, 28, 55, 58] have achieved great success. These state-of-the-art models are often trained on a large amount of image and discrete label pairs. The discrete label is generated by converting a detailed textual description, e.g., \"American curl cat\", into a simple scalar, which strikingly eases the computation of loss function. However, this also results in two evident limitations: (i) the rich semantics" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 228, + 535, + 382 + ], + "blocks": [ + { + "bbox": [ + 331, + 228, + 535, + 382 + ], + "lines": [ + { + "bbox": [ + 331, + 228, + 535, + 382 + ], + "spans": [ + { + "bbox": [ + 331, + 228, + 535, + 382 + ], + "type": "image", + "image_path": "ec5630472625830893480b27ff0857a22d4d426ea19f25fcd61d1b27c78071fe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 392, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 309, + 392, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 392, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Performance comparison between Zero-shot CLIP [49], CoOp [73], CLIP-Adapter [15], Tip-Adapter-F [71] and our TaskRes on ImageNet with few shot settings." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 436, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 436, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 309, + 436, + 555, + 460 + ], + "type": "text", + "content": "in the textual description are underused, and (ii) the trained models are limited to recognizing the close-set classes only." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 461, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 461, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 309, + 461, + 556, + 675 + ], + "type": "text", + "content": "Recent large-scale vision-language model (VLM) pretraining [1, 23, 33, 49, 69] eliminates those limitations by learning visual representations via textual supervision. For instance, texts and images are encoded and mapped into a unified space via a contrastive loss during pre-training [49]. The pre-trained text encoder can then be used to synthesize a text-based classifier for image recognition given the corresponding natural language descriptions as shown in Figure 2 (a). Those pre-trained VLMs have demonstrated a powerful transferability on a variety of downstream tasks in a zero-shot manner. However, the effectiveness of the aforementioned models heavily relies on their large-scale architectures and training datasets. For instance, CLIP [49] has up to 428 million parameters and is trained on 0.4 billion text-image pairs, while Flamingo [1] boasts up to 80 billion parameters and is trained on a staggering 2.1 billion pairs. This makes it impractical to fully fine-tune the model on downstream tasks in a low-data regime." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 309, + 677, + 556, + 715 + ], + "type": "text", + "content": "For that reason, efficient transfer learning (ETL) [15, 71-73] on pre-trained VLMs has gained popularity. ETL represents transfer learning to downstream tasks in both" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10899" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 280, + 134 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 280, + 134 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 280, + 134 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 280, + 134 + ], + "type": "image", + "image_path": "b34b2bbb64c09659bf9d7071e34c573312b6e19ec4dcef6b02745c33246da132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 139, + 220, + 149 + ], + "lines": [ + { + "bbox": [ + 164, + 139, + 220, + 149 + ], + "spans": [ + { + "bbox": [ + 164, + 139, + 220, + 149 + ], + "type": "text", + "content": "(a) Zero-shot CLIP" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 521, + 134 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 521, + 134 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 521, + 134 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 521, + 134 + ], + "type": "image", + "image_path": "2086ad88a7c944b314b658e80da43ffcbf9d48a5eaeb1f050ee86b65e1327d52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 140, + 453, + 149 + ], + "lines": [ + { + "bbox": [ + 396, + 140, + 453, + 149 + ], + "spans": [ + { + "bbox": [ + 396, + 140, + 453, + 149 + ], + "type": "text", + "content": "(b) Prompt Tuning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 152, + 315, + 231 + ], + "blocks": [ + { + "bbox": [ + 82, + 152, + 315, + 231 + ], + "lines": [ + { + "bbox": [ + 82, + 152, + 315, + 231 + ], + "spans": [ + { + "bbox": [ + 82, + 152, + 315, + 231 + ], + "type": "image", + "image_path": "c0ae072c1d6bee0fcc19c326e4d83304df18069ccc354f174291732cbb2d830f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 234, + 232, + 244 + ], + "lines": [ + { + "bbox": [ + 156, + 234, + 232, + 244 + ], + "spans": [ + { + "bbox": [ + 156, + 234, + 232, + 244 + ], + "type": "text", + "content": "(c) Adapter-style Tuning" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 152, + 520, + 228 + ], + "blocks": [ + { + "bbox": [ + 322, + 152, + 520, + 228 + ], + "lines": [ + { + "bbox": [ + 322, + 152, + 520, + 228 + ], + "spans": [ + { + "bbox": [ + 322, + 152, + 520, + 228 + ], + "type": "image", + "image_path": "6fc3168e0f3d22bbaf220920fb5eccc3310dd2c71387d13b24ca4ed97827b6ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 234, + 470, + 243 + ], + "lines": [ + { + "bbox": [ + 394, + 234, + 470, + 243 + ], + "spans": [ + { + "bbox": [ + 394, + 234, + 470, + 243 + ], + "type": "text", + "content": "(d) Task Residual Tuning" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 253, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 555, + 286 + ], + "type": "text", + "content": "Figure 2. Illustration of (a) Zero-shot CLIP, (b) prompt tuning, (c) adapter-style tuning and (d) our proposed Task Residual Tuning (TaskRes). Our method introduces a prior-independent task residual to the fixed pre-trained classifier (i.e., text embeddings of CLIP), being free of running the text encoder every time or extra architecture design." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 297, + 291, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 291, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 291, + 536 + ], + "type": "text", + "content": "parameter- and data-efficient manner. The core of ETL is twofold: (i) properly inheriting the well-learned knowledge structure of VLMs, which is already transferable; (ii) effectively exploring the task-specific knowledge given limited data. However, most existing ETL approaches, i.e., prompt tuning (PT) [72, 73] and adapter-style tuning (AT) [15, 71], either damage the prior knowledge of VLMs or learn the new knowledge of a task in an inappropriate/insufficient way. For example, instead of using the pre-trained text-based classifier, CoOp [73] (in Figure 2 (b)) is proposed to learn a continuous prompt for synthesizing a completely new one, which inevitably causes the loss of previous knowledge. Consequently, CoOp underperforms Zero-shot CLIP by " + }, + { + "bbox": [ + 46, + 297, + 291, + 536 + ], + "type": "inline_equation", + "content": "1.03\\% / 0.37\\%" + }, + { + "bbox": [ + 46, + 297, + 291, + 536 + ], + "type": "text", + "content": " in 1-/2-shot learning on ImageNet (see Figure 1). In contrast, CLIP-Adapter [15] preserves the pretrained classifier, but is excessively biased towards the prior knowledge when learning a new task, i.e., it transforms the pre-trained classifier weights to be task-specific as illustrated in Figure 2 (c). This results in an inferior new knowledge exploration, thereby a lower accuracy as shown in Figure 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 540, + 292, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 292, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 292, + 696 + ], + "type": "text", + "content": "For better ETL on pre-trained VLMs, we propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the old knowledge of the pretrained models and the new knowledge for a target task. The rationale is that the decoupling enables a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration, i.e., the learned knowledge w.r.t. the task is independent on the old knowledge. Specifically, TaskRes keeps the original classifier weights frozen and introduces a set of prior-independent parameters that are added to the weights. These additive parameters, tuned for adaptation to the target task, are thus named \"task residual\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 701, + 291, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 291, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 291, + 713 + ], + "type": "text", + "content": "To gain insight into how TaskRes works, we perform ex" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 297, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 297, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 309, + 297, + 555, + 464 + ], + "type": "text", + "content": "tensive experiments across 11 benchmark datasets [73] and conduct a systematic investigation of learned task residuals. The experimental results demonstrate that introducing task residual can significantly enhance the transfer performance. We visualize the correlation between the magnitude of learned task residual and the difficulty of transferring a pre-trained model to a downstream task, and observe that the magnitude increases with the transfer difficulty. This suggests the residual is automatically adapted to the task to fully explore the new knowledge, thereby achieving a new state-of-the-art performance on 11 diverse datasets. Furthermore, it is worth noting that our method requires minimal effort for the implementation, i.e., technically adding one line of code only. Our contributions are summarized below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 471, + 555, + 629 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 321, + 471, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 471, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 321, + 471, + 555, + 543 + ], + "type": "text", + "content": "- We for the first time emphasize the necessity of a proper knowledge inheritance from pre-trained VLMs to downstream tasks via ETL, reveal the pitfalls of existing tuning paradigms, and conduct an in-depth analysis to manifest that decoupling the old pre-trained knowledge and the new task-specific knowledge is the key." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 550, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 550, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 321, + 550, + 555, + 597 + ], + "type": "text", + "content": "- We propose a new efficient tuning approach named Task Residual Tuning (TaskRes), which achieves a better old knowledge inheritance from VLMs and a more flexible task-specific knowledge exploration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 605, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 605, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 321, + 605, + 553, + 629 + ], + "type": "text", + "content": "- TaskRes is convenient for use, which needs a few tuning parameters and effortless implementation." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 310, + 639, + 397, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 639, + 397, + 651 + ], + "spans": [ + { + "bbox": [ + 310, + 639, + 397, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 659, + 449, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 659, + 449, + 672 + ], + "spans": [ + { + "bbox": [ + 310, + 659, + 449, + 672 + ], + "type": "text", + "content": "2.1. Vision-Language Models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 309, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 677, + 555, + 713 + ], + "type": "text", + "content": "We mainly review the literature of vision-language models (VLMs) on language-driven visual representation learning (LDVRL) [2,9,11,14,17,25,29,31,49,52,56]. The core of" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10900" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 290, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 290, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 290, + 131 + ], + "type": "text", + "content": "LDVRL is to map texts (language) and images into a common space so that the text representations can be used for visual classification. To achieve this, two encoders for texts and images respectively, and specific loss functions as regularization are needed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 291, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 291, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 291, + 240 + ], + "type": "text", + "content": "Early works explore unsupervised pre-trained models [56] or the skip-gram text modeling [14, 44, 45] for text embedding while sparse coding and vector quantization [7, 56] or Classeme features [11, 60] for visual encoding. For training objectives, MSE [56], self-supervised topic probabilities matching [17] or multi-class logistic loss are adopted. However, those methods are limited to small datasets [9, 25, 31] and weakly representative encoding backbones [7, 56], which tremendously hinders their transferability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 291, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 291, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 291, + 361 + ], + "type": "text", + "content": "In contrast, recent works [23, 35, 49] have advanced previous ones by leveraging billion-level data of image-text pairs from internet and super powerful neural networks, e.g., Transformers [10, 61, 62, 68, 70], for representation learning. By end-to-end pre-training with a contrastive loss, large-scale VLMs demonstrate remarkable transferability on various downstream tasks in a zero-shot evaluation manner. Standing on the strong transfer capability of large-scale vision-language pre-training, we further explore its potential in efficient transfer learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 370, + 198, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 198, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 198, + 383 + ], + "type": "text", + "content": "2.2. Efficient Transfer Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 291, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 291, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 291, + 616 + ], + "type": "text", + "content": "Pre-training a neural network on large-scale datasets, e.g., ImageNet [8] and WebImageText [49], then fine-tuning it on downstream tasks has been a common step of transfer learning [36-38, 65-67]. Here, we mainly focus on efficient transfer learning (ETL) on pre-trained VLMs. ETL represents both parameter- and data-efficient transfer learning, in which a small number of parameters are tuned and a small set of data is utilized. Existing works on ETL can be grouped into two categories: prompt tuning [40, 72, 73] and adapter-style tuning (AT) [15, 71]. Specifically, prompt engineering [24, 54] is first explored to generate appropriate discrete prompts for downstream tasks. Later, learning continuous prompts [39, 73] that can better adapt to a new task shows a more advanced performance. However, these methods face two issues: (i) they need the pre-trained text encoder to engage in the whole training, limiting its scalability and increasing the computational overhead; (ii) they abandon the well-pre-trained text-based classifier and generate a new one, which causes the loss of prior knowledge of VLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 291, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 291, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 291, + 713 + ], + "type": "text", + "content": "AT-based methods [15, 71] solve the above issues by using the text encoder once for text-based classifier generation and then focusing on adapting text/image features only. This straightforward design can achieve even better performance, but it heavily relies on the prior knowledge/pre-trained features, which causes an inferior new knowledge exploration. To address this, we propose a new efficient tuning approach, Task Residual Tuning (TaskRes), for a better and more flexi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 555, + 95 + ], + "type": "text", + "content": "ble task-specific knowledge learning by a prior-independent \"task residual\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 103, + 423, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 103, + 423, + 115 + ], + "spans": [ + { + "bbox": [ + 310, + 103, + 423, + 115 + ], + "type": "text", + "content": "2.3. Few-Shot Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 121, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 121, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 309, + 121, + 555, + 277 + ], + "type": "text", + "content": "Few-shot learning (FSL) aims to adapt a model to novel tasks/classes with a handful of labeled examples. Conventional FSL methods [5, 13, 16, 41, 42, 53] often meta-learn on abundant data from base classes for adaptive capability. However, the need of learning on base datasets restricts their scalability. Recent VLM pre-training works [23, 49] offer an effective alternative, which does not require a base dataset. They show that the pre-trained models can already achieve astonishing performance on many downstream tasks in a zero-shot manner. ETL [15, 73] can further boost the performance. In this work, we propose a novel ETL method for adapting VLMs to downstream tasks and evaluate its effectiveness on few-shot tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 288, + 394, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 288, + 394, + 300 + ], + "spans": [ + { + "bbox": [ + 310, + 288, + 394, + 300 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 308, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 308, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 555, + 356 + ], + "type": "text", + "content": "We briefly introduce the adopted VLM, i.e., contrastive language-image pre-training (CLIP) [49], and recap two mainstream approaches for ETL on VLMs, i.e., prompt tuning and adapter-style tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 363, + 529, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 363, + 529, + 376 + ], + "spans": [ + { + "bbox": [ + 310, + 363, + 529, + 376 + ], + "type": "text", + "content": "3.1. Contrastive Language-Image Pre-training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "text", + "content": "The CLIP model [49] is designed to obtain visual representations through natural language supervision. It is trained on 0.4 billion image-text pairs, where image features from an image encoder and text features from a text encoder are aligned within a unified embedding space using a contrastive learning loss, allowing CLIP to effectively capture broad visual concepts and learn general visual representations. When testing, CLIP can classify a query image into " + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "text", + "content": " possible categories. This is achieved by calculating the cosine similarity between a query image embedding " + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "text", + "content": ", obtained from the image encoder with projection, and the text embeddings " + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "text", + "content": ", which are derived by inputting texts (e.g., \"a photo of a {class}\") into the text branch. The predicted probability for class " + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 309, + 381, + 556, + 548 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 554, + 555, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 554, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 348, + 554, + 555, + 585 + ], + "type": "interline_equation", + "content": "p (y = i | \\mathbf {z}) = \\frac {\\exp (\\sin (\\mathbf {z} , \\mathbf {t} _ {i}) / \\tau)}{\\sum_ {j = 1} ^ {K} \\exp (\\sin (\\mathbf {z} , \\mathbf {t} _ {j}) / \\tau)}, \\tag {1}", + "image_path": "624c22dd9a3e45fdf54cd93f136d389a794a9c85de4d2a1d3f24336554cb7f0c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\sin (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "type": "text", + "content": " indicates cosine similarity and " + }, + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 309, + 592, + 555, + 616 + ], + "type": "text", + "content": " is the learned temperature of CLIP." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 623, + 512, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 623, + 512, + 636 + ], + "spans": [ + { + "bbox": [ + 310, + 623, + 512, + 636 + ], + "type": "text", + "content": "3.2. Revisiting Previous Tuning Paradigms" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 641, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 641, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 309, + 641, + 555, + 689 + ], + "type": "text", + "content": "Inspired by the success of ETL approaches in natural language processing, e.g., prompt tuning [30, 34] and Adapter [22], recent advances (e.g., CoOp [73] and CLIP-Adapter [15]) borrow their ideas to ETL on VLMs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 689, + 555, + 713 + ], + "type": "text", + "content": "CoOp, for the first time, introduces prompt tuning to VLMs. Instead of using fixed text prompt contexts such as" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10901" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": "\"a photo of a\", CoOp proposes to use " + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": " learnable context vectors " + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_m\\}_{m=1}^M" + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": " as the task-specific templates. The prompt given to the text encoder then becomes " + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_1, \\mathbf{v}_2, \\dots, \\mathbf{v}_M, \\mathbf{c}_i\\}" + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": " is the embedding of class " + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": ". During the whole training process, CoOp keeps the parameters of pre-trained VLMs frozen and only tunes the learnable vectors " + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_m\\}_{m=1}^M" + }, + { + "bbox": [ + 46, + 72, + 290, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "text", + "content": "Adapter-style tuning introduces additional modules " + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "inline_equation", + "content": "\\phi_{\\omega}(\\cdot)" + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "text", + "content": " with tunable parameters " + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "text", + "content": " to the pre-trained models to transform the pre-trained features " + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "text", + "content": " to new ones " + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 46, + 144, + 291, + 192 + ], + "type": "text", + "content": ". In general, adapter-style tuning can be formulated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 201, + 291, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 291, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 291, + 213 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {\\prime} = \\mathbf {f} + \\alpha \\phi_ {\\omega} (\\mathbf {f}), \\tag {2}", + "image_path": "7a3d25350e25991c6dc2216c606d19d7273d87f571993df01254d6831fda3362.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "type": "text", + "content": " is a scaling factor. In CLIP-Adapter, the adapter module " + }, + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "type": "inline_equation", + "content": "\\phi_{\\omega}" + }, + { + "bbox": [ + 46, + 222, + 291, + 317 + ], + "type": "text", + "content": " is composed of two linear transformation layers and a ReLU activation between them. CLIP-Adapter investigates both visual and text adapters, i.e., applying the adapter module to the image and text branches of CLIP, respectively, and shows that they have comparable performance. During training on downstream tasks, adapter-style methods only tune their adapter modules." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 328, + 113, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 113, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 113, + 342 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 291, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 291, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 291, + 408 + ], + "type": "text", + "content": "In this section, we start by identifying the pitfalls of the existing ETL paradigms on VLMs. Those pitfalls motivate us to propose a new ETL approach for VLMs, named Task Residual Tuning (TaskRes). The proposed TaskRes is simple yet can effectively avoid the pitfalls." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 415, + 280, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 280, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 280, + 428 + ], + "type": "text", + "content": "4.1. Pitfalls of Existing ETL Paradigms on VLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 434, + 291, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 291, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 291, + 601 + ], + "type": "text", + "content": "We rethink the usage of the prior knowledge from pretrained VLMs and the acquisition of the new knowledge regarding downstream tasks. On the one hand, large-scale VLMs trained with a vast amount of data have learned broad visual concepts that are general for a wide range of downstream vision tasks, enabling homogenization [3]. The prior knowledge should be well preserved when performing a transfer. On the other hand, despite the huge data used in the pre-training, inevitably, there are domain shifts or uncertain concepts in downstream tasks. New knowledge specific to the downstream tasks should be appropriately supplemented to the prior knowledge. However, existing ETL paradigms do not well consider the above principles and have the following two issues." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 293, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 293, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 293, + 713 + ], + "type": "text", + "content": "Pitfall 1: Lack of guarantees of prior knowledge preservation in prompt tuning. While the weights of pre-trained text branch modules (e.g., text encoder and projection) are frozen in the prompt tuning paradigm, the original well-learned classification boundary is more or less damaged. This is because the tuning of input prompts ends up with a new boundary that may forget the old knowledge without explicit regularization. As a result, the performance of prompt tuning is limited. For instance, the performance of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 555, + 96 + ], + "type": "text", + "content": "CoOp [73] is not as good as Zero-shot CLIP in 1-/2-shot learning on ImageNet as shown in Figure 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 102, + 556, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 102, + 556, + 281 + ], + "spans": [ + { + "bbox": [ + 309, + 102, + 556, + 281 + ], + "type": "text", + "content": "Pitfall 2: Limited flexibility of new knowledge exploration in adapter-style tuning. The data distributions in downstream tasks often shift from the pre-training distribution, and some task-specific or fine-grained visual concepts/representations may not be well learned by pre-trained VLMs, e.g., from CLIP [49] to satellite image dataset EuroSAT [19]. Thus, the new knowledge regarding downstream tasks needs to be appropriately explored. We observe that adapter-style tuning may not sufficiently explore the task-specific knowledge as the input of the adapter is strictly limited to old/pre-trained features, as shown in Figure 2 (c). Regardless of whether the pre-trained features are suitable for the task, the results of the adapter depend only on them, making adapter-style tuning have limited flexibility for learning new knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 289, + 435, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 289, + 435, + 302 + ], + "spans": [ + { + "bbox": [ + 310, + 289, + 435, + 302 + ], + "type": "text", + "content": "4.2. Task Residual Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 308, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 308, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 555, + 392 + ], + "type": "text", + "content": "Given that the existing ETL paradigms face the aforementioned issues, we propose Task Residual Tuning (TaskRes) to address the issues in a simple way. TaskRes explicitly decouples the maintenance of the old knowledge from the pre-trained VLMs and the learning of task-specific knowledge that is not overly biased on the pre-trained features. We elaborate our TaskRes below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "text", + "content": "Fixed base classifier. As illustrated in Figure 2 (d), our TaskRes performs the tuning directly on the text-based classifier (i.e., text embeddings)1. Base classifier is the text embeddings of a pre-trained vision-language model, e.g., CLIP. We denote the base classifier as " + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "text", + "content": " is the number of categories and " + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 309, + 397, + 556, + 494 + ], + "type": "text", + "content": " is the feature dimension. We keep the base classifier weights frozen to explicitly prevent the base classifier from being damaged." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "spans": [ + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "text", + "content": "Prior-independent task residual. To learn task-specific knowledge without being restricted by prior knowledge, we propose task residual, which is a set of tunable parameters " + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{K \\times D}" + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "text", + "content": " that are not dependent on the base classifier. Our task residual is scaled by a factor " + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "text", + "content": " and element-wisely added to the base classifier to formulate a new classifier " + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{t}'" + }, + { + "bbox": [ + 309, + 499, + 556, + 583 + ], + "type": "text", + "content": " for the target task, written as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 403, + 588, + 555, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 588, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 403, + 588, + 555, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {t} ^ {\\prime} = \\mathbf {t} + \\alpha \\mathbf {x}. \\tag {3}", + "image_path": "ae50187c4014bbc3f295f50c5d6fbc480a15e067784377a26484bc67d89a5cdb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 609, + 556, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 609, + 556, + 657 + ], + "spans": [ + { + "bbox": [ + 309, + 609, + 556, + 657 + ], + "type": "text", + "content": "Tuning for downstream tasks. During tuning, we only tune the prior-independent task residual while keeping the base classifier (together with the image branch) fixed, enabling reliable old knowledge preservation and flexible new" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 309, + 664, + 556, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 664, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 309, + 664, + 556, + 712 + ], + "type": "text", + "content": "1While we can tune on image embeddings, it may encounter overfitting since the testing image embeddings are different from the training ones. Tuning on text embeddings is free of the issue. Besides, large diversity between image embeddings is not conducive to learning stable task-level information." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10902" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "type": "text", + "content": "knowledge exploration. Given an image, the fixed image branch of CLIP extracts its embeddings " + }, + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "type": "text", + "content": ". The predicted probability for class " + }, + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 292, + 109 + ], + "type": "text", + "content": " is then calculated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 114, + 291, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 114, + 291, + 144 + ], + "spans": [ + { + "bbox": [ + 85, + 114, + 291, + 144 + ], + "type": "interline_equation", + "content": "p (y = i | \\mathbf {z}) = \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} , \\mathbf {t} _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\operatorname {s i m} \\left(\\mathbf {z} , \\mathbf {t} _ {j} ^ {\\prime}\\right) / \\tau\\right)}. \\tag {4}", + "image_path": "622b3f1d07f6ab6959eb6498aa973a53dd91a9025fbc835cadf68309c317a8ba.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 292, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 292, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 292, + 186 + ], + "type": "text", + "content": "Based on the predicted probabilities, the downstream task loss (e.g., cross-entropy loss) only updates the task residual through standard backpropagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 196, + 123, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 196, + 123, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 196, + 123, + 209 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 215, + 97, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 97, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 97, + 227 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 233, + 292, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 292, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 292, + 402 + ], + "type": "text", + "content": "We follow prior efficient transfer learning (ETL) work [15, 71, 73] to conduct a few-shot evaluation for ETL models on 11 benchmark datasets i.e., ImageNet [8], Caltech101 [12], OxfordPets [48], StanfordCars [27], Flowers102 [47], Food101 [4], FGVCAircraft [43], SUN397 [64], DTD [6], EuroSAT [19] and UCF101 [57]. Those datasets include a wide range of visual recognition tasks, e.g., classifying general objects, fine-grained objects, actions, scenes, etc. Specifically, ETL models are trained using " + }, + { + "bbox": [ + 46, + 233, + 292, + 402 + ], + "type": "inline_equation", + "content": "1/2/4/8/16" + }, + { + "bbox": [ + 46, + 233, + 292, + 402 + ], + "type": "text", + "content": " shots per class from the training sets, respectively, and are tested on the full test sets. Additionally, following CoOp [73], we test the generalization performance of our models from ImageNet to its variants (ImageNetV2 [51], ImageNet-Sketch [63], ImageNet-A [21] and ImageNet-R [20])." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 407, + 145, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 145, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 145, + 420 + ], + "type": "text", + "content": "5.2. Implementation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "text", + "content": "Our approach has two main components, i.e., base classifier and task residual. For the base classifier, we develop two versions: regular base classifier that directly adopts the weights of the text embeddings of the pre-trained CLIP (using pre-defined text templates following prior work [49, 71]), and enhanced base classifier obtained by tuning the text projection layer of CLIP on the target task before starting our task residual tuning" + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "text", + "content": ". Task residual is a matrix filled with learnable parameters that are initialized by zeros. Task residual has the same shape as the base classifier. We scale the task residual and element-wisely add it to the base classifier, as in Eq. 3. By default, the scaling factor " + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "text", + "content": " is set to 0.5 for all datasets except for Flowers102 using 1. Additionally, we explore the use of a learnable " + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 426, + 292, + 677 + ], + "type": "text", + "content": " later in our ablation study. All models in this work are built upon the pre-trained CLIP models. Unless otherwise specified, we use the ResNet-50 [18] version of CLIP which has a ResNet-50 backbone in the image branch. We train our models for 100 epochs for 1-/2-/4-shot experiments and 200 epochs for 8-/16-shot experiments, with batch size 256. Besides, the aforementioned enhanced base classifier is tuned for 50 epochs. Our models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "text", + "content": "are optimized by Adam [26] with an initial learning rate of " + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "2e - 3" + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "text", + "content": " except for ImageNet using " + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "2e - 4" + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "text", + "content": ". Following CoOp [73], our optimization process adopts a cosine learning rate decay schedule and a warmup scheme (i.e., fixing the learning rate in the first epoch to " + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 309, + 72, + 555, + 156 + ], + "type": "text", + "content": "). We test the performance of models after training, and all experimental results are averaged over three random seeds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 165, + 454, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 165, + 454, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 165, + 454, + 178 + ], + "type": "text", + "content": "5.3. Performance Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 184, + 426, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 184, + 426, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 184, + 426, + 196 + ], + "type": "text", + "content": "5.3.1 Few-Shot Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 309, + 204, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 204, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 309, + 204, + 556, + 479 + ], + "type": "text", + "content": "We develop two versions of our TaskRes, i.e., Ours/Ours* using the regular/enhanced base classifier, respectively. We compare our models with Zero-shot CLIP [49], together with the state-of-the-art (SOTA) ETL methods including CoOp [73], CLIP-Adapter [15] and Tip-Adapter-F " + }, + { + "bbox": [ + 309, + 204, + 556, + 479 + ], + "type": "inline_equation", + "content": "[71]^3" + }, + { + "bbox": [ + 309, + 204, + 556, + 479 + ], + "type": "text", + "content": ". The comparison results of the 11 benchmark datasets are shown in Figure 3. Overall, both Ours and Ours* achieve the SOTA/SOTA-comparable performance on the 11 datasets across all few-shot settings and significantly surpass Zero-shot CLIP [49]. Besides, a few observations can be made. First, with more shots, Tip-Adapter-F reaches a closer performance to ours, but as shown in Figure 1 the number of its tunable parameters increases linearly with that of shots, substantially limiting its scalability. Second, our method shows inferior performance on OxfordPets and Food101. Training with a few shots on these two fine-grained datasets is prone to be overfitting, which is also found in CoOp [73]. Finally, we found the proposed method is superior when needed to recognize a large amount of classes, e.g., 1000 classes of ImageNet and 397 classes of SUN397 in 1-shot case. This is because our TaskRes, not excessively biased to the pre-trained features, can better explore the task-specific knowledge even in such extreme scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 487, + 443, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 487, + 443, + 498 + ], + "spans": [ + { + "bbox": [ + 310, + 487, + 443, + 498 + ], + "type": "text", + "content": "5.3.2 Domain Generalization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 506, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 506, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 309, + 506, + 556, + 674 + ], + "type": "text", + "content": "As pointed out by CoOp [73], ETL models that are trained on a specific domain are at risk of learning spurious correlations when generalizing to unseen domains. We evaluate our models concerning domain generalization. We train our models on 16-shot ImageNet and test the generalization performance of the trained models on four unseen ImageNet variant datasets (ImageNet-V2, -Sketch, -A and -R). As shown in Table 1, our models consistently outperform the compared models (Zero-shot CLIP [49], Linear Probe CLIP [49] and CoOp [73]) across various CLIP visual backbones (ResNet-50 [18], ResNet-101 [18], ViT-B/32 [10] and ViT-B/16 [10]). We also observe that our TaskRes using the enhanced base classifier (Ours*) achieves higher accuracy than that using the regular base classifier (Ours) on the source dataset while" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 309, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 309, + 683, + 555, + 713 + ], + "type": "text", + "content": "Original Tip-Adapter [71] has two versions: training-free Tip-Adapter and a fine-tuning version called Tip-Adapter-F. We report the one with higher performance, i.e., Tip-Adapter-F." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 683, + 292, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 292, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 292, + 712 + ], + "type": "text", + "content": "2The enhanced base classifier is introduced to help us investigate whether there are consistent performance improvements when the base classifier is enhanced." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10903" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 226, + 189 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 226, + 189 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 226, + 189 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 226, + 189 + ], + "type": "image", + "image_path": "6afc4c5d957958795bd4b21505c85970a89696ffaff6333c66cd5658c77763cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 72, + 373, + 188 + ], + "blocks": [ + { + "bbox": [ + 228, + 72, + 373, + 188 + ], + "lines": [ + { + "bbox": [ + 228, + 72, + 373, + 188 + ], + "spans": [ + { + "bbox": [ + 228, + 72, + 373, + 188 + ], + "type": "image", + "image_path": "4490e6b3ce7dfe329a760fea469a63a4fd9a33ad15b0453cc61619991bbccd91.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 523, + 188 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 523, + 188 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 523, + 188 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 523, + 188 + ], + "type": "image", + "image_path": "2cc10b7a01c9b612c6160b35dd7948090f2192a61e243fd79bebf3d03f0637d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 78, + 193, + 225, + 310 + ], + "blocks": [ + { + "bbox": [ + 78, + 193, + 225, + 310 + ], + "lines": [ + { + "bbox": [ + 78, + 193, + 225, + 310 + ], + "spans": [ + { + "bbox": [ + 78, + 193, + 225, + 310 + ], + "type": "image", + "image_path": "9b7624f0d791b0a8cb0f20b00af6dd5db577eb6906b00667ac1ffc476975784c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 228, + 193, + 373, + 309 + ], + "blocks": [ + { + "bbox": [ + 228, + 193, + 373, + 309 + ], + "lines": [ + { + "bbox": [ + 228, + 193, + 373, + 309 + ], + "spans": [ + { + "bbox": [ + 228, + 193, + 373, + 309 + ], + "type": "image", + "image_path": "bee1ed9a2cf628ea56787d01fd0a17b103f4f7de2e0f139642820ee49c5b6975.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 377, + 193, + 523, + 309 + ], + "blocks": [ + { + "bbox": [ + 377, + 193, + 523, + 309 + ], + "lines": [ + { + "bbox": [ + 377, + 193, + 523, + 309 + ], + "spans": [ + { + "bbox": [ + 377, + 193, + 523, + 309 + ], + "type": "image", + "image_path": "20689d8536aabae615cbcf8939ac94588553abe587ab383309f31a6484beee6d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 80, + 314, + 225, + 430 + ], + "blocks": [ + { + "bbox": [ + 80, + 314, + 225, + 430 + ], + "lines": [ + { + "bbox": [ + 80, + 314, + 225, + 430 + ], + "spans": [ + { + "bbox": [ + 80, + 314, + 225, + 430 + ], + "type": "image", + "image_path": "56cd09bd822811baa963386ed0a7772fe94eea0e9ad11a1477972841d63f421a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 314, + 373, + 430 + ], + "blocks": [ + { + "bbox": [ + 228, + 314, + 373, + 430 + ], + "lines": [ + { + "bbox": [ + 228, + 314, + 373, + 430 + ], + "spans": [ + { + "bbox": [ + 228, + 314, + 373, + 430 + ], + "type": "image", + "image_path": "5e74d59325f2293f8d11a23d3e6ee17d57b8b69056021cb9e101dd0f8cbf6cfe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 377, + 314, + 523, + 430 + ], + "blocks": [ + { + "bbox": [ + 377, + 314, + 523, + 430 + ], + "lines": [ + { + "bbox": [ + 377, + 314, + 523, + 430 + ], + "spans": [ + { + "bbox": [ + 377, + 314, + 523, + 430 + ], + "type": "image", + "image_path": "f14bf08663271e41629dd0a5ec2fe7113f7704d09d8944ec3a1d9347d2671d08.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 80, + 434, + 225, + 552 + ], + "blocks": [ + { + "bbox": [ + 80, + 434, + 225, + 552 + ], + "lines": [ + { + "bbox": [ + 80, + 434, + 225, + 552 + ], + "spans": [ + { + "bbox": [ + 80, + 434, + 225, + 552 + ], + "type": "image", + "image_path": "d36ea553e962bd64ad5891c35bd98e7cd94f00c57ee77974a284e4509492b066.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 562, + 555, + 586 + ], + "lines": [ + { + "bbox": [ + 45, + 562, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 45, + 562, + 555, + 586 + ], + "type": "text", + "content": "Figure 3. Performance comparison on few-shot learning, i.e., 1-/2-/4-/8-/16-shot, on 11 benchmark datasets. The top-left is the averaged accuracy over the 11 datasets. The full numerical results can be found in the supplementary." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 435, + 373, + 552 + ], + "blocks": [ + { + "bbox": [ + 228, + 435, + 373, + 552 + ], + "lines": [ + { + "bbox": [ + 228, + 435, + 373, + 552 + ], + "spans": [ + { + "bbox": [ + 228, + 435, + 373, + 552 + ], + "type": "image", + "image_path": "0b9614b1d271812ede7ebf39fb32cf34d8b0a7bcf5e303754e8aaf4fb1a31e7d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 377, + 437, + 523, + 552 + ], + "blocks": [ + { + "bbox": [ + 377, + 437, + 523, + 552 + ], + "lines": [ + { + "bbox": [ + 377, + 437, + 523, + 552 + ], + "spans": [ + { + "bbox": [ + 377, + 437, + 523, + 552 + ], + "type": "image", + "image_path": "633f7e1b0e220eff428c83eef895a978db2766201c16f3184b888df3588369da.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 596, + 293, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 293, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 293, + 620 + ], + "type": "text", + "content": "reducing a certain degree of generalization, which is due to a slight overfitting on the source dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 633, + 141, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 141, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 141, + 646 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 654, + 292, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 292, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 292, + 714 + ], + "type": "text", + "content": "Importance of prior knowledge preservation and prior-independent tuning. We build four target classifiers " + }, + { + "bbox": [ + 46, + 654, + 292, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{t}'" + }, + { + "bbox": [ + 46, + 654, + 292, + 714 + ], + "type": "text", + "content": " to investigate the importance of prior knowledge preservation and prior-independent tuning (where the tunable parameters are not dependent on the pre-trained features), i.e., original" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "content": "pre-trained classifier " + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "inline_equation", + "content": "\\mathbf{t}' = \\mathbf{t}" + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "content": " (i.e., our regular base classifier), directly-adapted classifier " + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "inline_equation", + "content": "\\mathbf{t}' = \\phi_{\\omega}(\\mathbf{t})" + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "content": ", adapter-style classifier " + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "inline_equation", + "content": "\\mathbf{t}' = \\mathbf{t} + \\alpha \\phi_{\\omega}(\\mathbf{t})" + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "content": " and our TaskRes classifier " + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "inline_equation", + "content": "\\mathbf{t}' = \\mathbf{t} + \\alpha \\mathbf{x}" + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "content": ". We conduct experiments on ImageNet across all few-shot settings, using the same scaling factor " + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 309, + 596, + 556, + 704 + ], + "type": "text", + "content": ". The results in Table 2 support our motivations that (i) explicit preservation of prior knowledge matters as the directly-adapted classifier performs much worse than the original pre-trained classifier and adapter-style classifier, and (ii) tuning prior-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10904" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 143, + 69, + 459, + 286 + ], + "blocks": [ + { + "bbox": [ + 143, + 69, + 459, + 286 + ], + "lines": [ + { + "bbox": [ + 143, + 69, + 459, + 286 + ], + "spans": [ + { + "bbox": [ + 143, + 69, + 459, + 286 + ], + "type": "table", + "html": "
MethodVisual BackboneSourceTarget
ImageNet-V2-Sketch-A-RAverage
Zero-Shot CLIP [49]58.1851.3433.3221.6556.0040.58
Linear Probe CLIP [49]55.8745.9719.0712.7434.8628.16
CoOp [73]ResNet-5062.9555.1132.7422.1254.9641.23
Ours64.7556.4735.8322.8060.7043.95
Ours*65.7357.0034.4321.5058.1342.77
Zero-Shot CLIP [49]61.6254.8138.7128.0564.3846.49
Linear Probe CLIP [49]59.7550.0526.8019.4447.1935.87
CoOp [73]ResNet-10166.6058.6639.0828.8963.0047.41
Ours67.7059.5041.7029.8768.0749.79
Ours*68.7360.0040.328.0064.8048.28
Zero-Shot CLIP [49]62.0554.7940.8229.5765.9947.79
Linear Probe CLIP [49]59.5849.7328.0619.6747.2036.17
CoOp [73]ViT-B/3266.8558.0840.4430.6264.4548.40
Ours68.2059.2042.5031.4369.3350.62
Ours*69.1759.4740.8729.7066.2749.08
Zero-Shot CLIP [49]66.7360.8346.1547.7773.9657.18
Linear Probe CLIP [49]65.8556.2634.7735.6858.4346.29
CoOp [73]ViT-B/1671.9264.1846.7148.4174.3258.41
Ours73.0765.3049.1350.3777.7060.63
Ours*73.9065.8547.7049.1775.2359.49
", + "image_path": "cc91907262720a5b1b9dcc5179596856eb3283f81c9948db91ff75b7be1dc944.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 319, + 290, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 290, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 290, + 354 + ], + "type": "text", + "content": "independent parameters is more effective than the dependent one as the TaskRes classifier obviously surpasses the adapter-style classifier." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 365, + 271, + 418 + ], + "blocks": [ + { + "bbox": [ + 50, + 295, + 552, + 306 + ], + "lines": [ + { + "bbox": [ + 50, + 295, + 552, + 306 + ], + "spans": [ + { + "bbox": [ + 50, + 295, + 552, + 306 + ], + "type": "text", + "content": "Table 1. Performance comparison on generalization (from ImageNet to ImageNet-V2/-Sketch/-A/-R) with multiple CLIP visual backbones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 365, + 271, + 418 + ], + "lines": [ + { + "bbox": [ + 67, + 365, + 271, + 418 + ], + "spans": [ + { + "bbox": [ + 67, + 365, + 271, + 418 + ], + "type": "table", + "html": "
Target Classifier t'1-shot2-shot4-shot8-shot16-shot
t60.3360.3360.3360.3360.33
φω(t)28.0334.9042.5748.1754.13
t + αφω(t)61.0361.2361.2762.0063.17
t + αx61.4362.1762.9364.0364.75
", + "image_path": "7f53f397305b90a7a12e5f926d407fec5f90003ed1ba97220908effda465ade7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 484, + 290, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 484, + 290, + 604 + ], + "spans": [ + { + "bbox": [ + 49, + 484, + 290, + 604 + ], + "type": "text", + "content": "Effectiveness of task residual learning. We conduct an ablation study of the effect of introducing our task residual learning (TaskRes). Table 3 shows the few-shot results averaged over 11 datasets of the regular base classifier and that equipped with our TaskRes (Ours), together with the enhanced counterparts. Our TaskRes notably improves the regular base, e.g., with " + }, + { + "bbox": [ + 49, + 484, + 290, + 604 + ], + "type": "inline_equation", + "content": "16.14\\%" + }, + { + "bbox": [ + 49, + 484, + 290, + 604 + ], + "type": "text", + "content": " accuracy gain in 16-shot case. Similarly, the performance gain is also impressive even on the enhanced base. This indicates that TaskRes can always benefit a pre-trained base when adapted to downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 55, + 614, + 284, + 671 + ], + "blocks": [ + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "lines": [ + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "text", + "content": "Table 2. Ablation study of various combinations of three main components: base classifier " + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "text", + "content": ", adapter module " + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "inline_equation", + "content": "{\\phi }_{\\omega }\\left( \\cdot \\right)" + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "text", + "content": " and our \"task residual\" " + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "text", + "content": " ,for building a target classifier " + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "inline_equation", + "content": "{\\mathbf{t}}^{\\prime }" + }, + { + "bbox": [ + 49, + 427, + 290, + 472 + ], + "type": "text", + "content": " . The experiment is conducted on ImageNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 614, + 284, + 671 + ], + "lines": [ + { + "bbox": [ + 55, + 614, + 284, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 284, + 671 + ], + "type": "table", + "html": "
Classifier1-shot2-shot4-shot8-shot16-shot
Regular Base58.9658.9658.9658.9658.96
Regular Base + TaskRes64.0467.0269.6172.7675.10
Enhanced Base61.1162.8264.4469.9173.09
Enhanced Base + TaskRes64.2867.5570.2873.3575.78
", + "image_path": "822a3152d11c2df1a213105ad253b9851a257694375d1752b27876d8cbf829fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 319, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 319, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 319, + 553, + 521 + ], + "type": "text", + "content": "Variants of TaskRes. The proposed \"task residual\" can also be applied to the image branch of CLIP, denoted as TaskRes-I. We also denote the default TaskRes applied on the text branch as TaskRes-T (i.e., Ours) and on the both branches as TaskRes-I&T. We compare their performance on ImageNet in Table 4. TaskRes-T consistently outperforms TaskRes-I, as shown in Table 4. The performance of TaskRes-I increases much slower than TaskRes-T as the number of shots increases. This is because TaskRes-I tends to encounter overfitting where training and testing image embeddings are different. Tuning on text embeddings can avoid the issue. Additionally, the diversity of image embeddings is larger than that of text embeddings, making it difficult to learn task-level information. TaskRes-I&T performs slightly worse than TaskRes-T. This might be caused by a misalignment between the updated image embeddings and our base classifier." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 335, + 530, + 529, + 584 + ], + "blocks": [ + { + "bbox": [ + 49, + 681, + 290, + 724 + ], + "lines": [ + { + "bbox": [ + 49, + 681, + 290, + 724 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 290, + 724 + ], + "type": "text", + "content": "Table 3. Exploring the effectiveness of the proposed task residual tuning (TaskRes). The accuracy results averaged on 11 datasets are reported under all shots. Refer to Appendix 2.3 for the full comparison results of the 11 datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 530, + 529, + 584 + ], + "lines": [ + { + "bbox": [ + 335, + 530, + 529, + 584 + ], + "spans": [ + { + "bbox": [ + 335, + 530, + 529, + 584 + ], + "type": "table", + "html": "
Model1-shot2-shot4-shot8-shot16-shot
Regular Base60.3360.3360.3360.3360.33
TaskRes-I61.2061.2761.4761.5361.60
TaskRes-T61.4362.1762.9364.0364.75
TaskRes-I&T61.5362.1062.7363.8364.67
", + "image_path": "dea6a54af21c43a181c9ad2b219865f5c35b250c722ca76c1ffc6ed78553c9a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 594, + 553, + 626 + ], + "lines": [ + { + "bbox": [ + 312, + 594, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 553, + 626 + ], + "type": "text", + "content": "Table 4. Performance comparison of three TaskRes variants on ImageNet: TaskRes applied on CLIP's image branch (TaskRes-I), text branch (TaskRes-T) and both branches (TaskRes-I&T)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "text", + "content": "Scaling factor. We investigate the effect of scaling factor " + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "text", + "content": " and show the results of both manually assigned values and a learned value in Table 5. For the manually set " + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "text", + "content": ", we observe a notable improvement with TaskRes scaled even by 0.1, while the highest accuracy is obtained with " + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "text", + "content": ". In contrast, the learnable " + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 312, + 629, + 553, + 712 + ], + "type": "text", + "content": " can further boost the performance. We show more analyses and results in Appendix 2.3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "10905" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 277, + 99 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 277, + 99 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 277, + 99 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 277, + 99 + ], + "type": "table", + "html": "
α00.10.30.50.71.0Learned
Acc.58.9661.8663.7663.9763.6763.0264.01
", + "image_path": "02561acb5c6b8c5780896fa2f6ef3afa95b075d85fa76bad458f611904ddad89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 141, + 291, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 291, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 291, + 201 + ], + "type": "text", + "content": "Visual backbone. We further evaluate models on various CLIP visual backbones, namely, ResNet-50 (RN50), ResNet-101 (RN101), ViT-B/32 and ViT-B/16, as shown in Table 6. Our method shows consistent superiority against other alternatives regardless of which backbone is used." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 211, + 272, + 285 + ], + "blocks": [ + { + "bbox": [ + 47, + 106, + 291, + 129 + ], + "lines": [ + { + "bbox": [ + 47, + 106, + 291, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 291, + 129 + ], + "type": "text", + "content": "Table 5. Ablation study of scaling factor " + }, + { + "bbox": [ + 47, + 106, + 291, + 129 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 106, + 291, + 129 + ], + "type": "text", + "content": ". The results are averaged over 11 datasets in 1-shot case." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 211, + 272, + 285 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 272, + 285 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 272, + 285 + ], + "type": "table", + "html": "
MethodRN50RN101ViT-B/32ViT-B/16
Zero-shot CLIP [49]58.1861.6262.0566.73
CoOp [73]62.9566.6066.8571.92
CLIP-Adapter [15]63.5965.3966.1971.13
Tip-Adapter-F [71]65.4468.5668.6573.69
Ours64.7567.7068.2073.07
Ours*65.7368.7369.1773.90
", + "image_path": "2956db81da2fc83b74cd9273973b551979670d8861a1c31c74ac7f2c2f9658d7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 293, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 50, + 293, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 50, + 293, + 287, + 304 + ], + "type": "text", + "content": "Table 6. Results of CLIP visual backbones on 16-shot ImageNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 316, + 253, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 253, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 253, + 329 + ], + "type": "text", + "content": "5.5. Investigation of Learned Task Residual" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "text", + "content": "We investigate whether the learned task residuals are in fact related to the difficulty of transferring CLIP to downstream tasks. We first define relative transfer difficulty to assess the difficulty of transferring a pre-trained model to a target task. Concretely, the relative transfer difficulty is defined as the ratio of baseline classifier precision to the zero-shot precision of the pre-trained model. We use a random classifier to serve as the baseline classifier where the precision of it equals to " + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "inline_equation", + "content": "1 / K" + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "text", + "content": " is the number of classes of the downstream task). We then can calculate the relative transfer difficulty of CLIP regarding the 11 benchmark datasets based on the random classifier and Zero-shot CLIP. We count the average magnitudes of the learned task residual of each dataset in 16-shot setting and find that they are highly related to relative transfer difficulty, as shown in Figure 4. This verifies the reliability of the proposed task residual. There are some interesting findings that may be counter-intuitive, e.g., transferring CLIP to ImageNet (with 1000 classes) is the simplest and to EuroSAT (with 10 classes) is the most difficult. This is because the inherent difficulty of ImageNet is large but the data distribution and containing objects are \"familiar\" to CLIP, whereas EuroSAT is the opposite. Besides, there is more space for TaskRes to play its role in the tasks with large relative transfer difficulty, e.g., EuroSAT (1-shot accuracy gain: " + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "inline_equation", + "content": "23.71\\%" + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "text", + "content": ") and DTD (1-shot accuracy gain: " + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "inline_equation", + "content": "7.85\\%" + }, + { + "bbox": [ + 46, + 334, + 291, + 646 + ], + "type": "text", + "content": "). More results and analyses can be found in Appendix 2.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 657, + 272, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 272, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 272, + 670 + ], + "type": "text", + "content": "6. Conclusion, Limitation and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 291, + 714 + ], + "type": "text", + "content": "In this work, we propose a new approach for tuning VLMs, i.e., TaskRes. The proposed TaskRes performs efficient transfer learning (ETL) on VLMs by explicitly de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 341, + 72, + 525, + 208 + ], + "blocks": [ + { + "bbox": [ + 341, + 72, + 525, + 208 + ], + "lines": [ + { + "bbox": [ + 341, + 72, + 525, + 208 + ], + "spans": [ + { + "bbox": [ + 341, + 72, + 525, + 208 + ], + "type": "image", + "image_path": "25727f268e1c570ff846a98cfe4debd7b16cc218d992ae03e5b14a0cd5fc1f43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 211, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 309, + 211, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 309, + 211, + 555, + 245 + ], + "type": "text", + "content": "Figure 4. The magnitude of learned task residual increases with relative transfer difficulty (in logarithmic scale). The shadow indicates the standard deviation regarding random seeds." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 253, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 253, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 253, + 555, + 372 + ], + "type": "text", + "content": "coupling the classifier into two key parts: undamaged base classifier with rich prior knowledge and task residual independent to the base classifier for better exploring task-specific knowledge. Interestingly, the magnitude of learned task residual is highly related to the difficulty of transferring pre-trained VLMs to a target downstream task. This may inspire the community to consider ETL from a new perspective, e.g., modeling \"task-to-task transfer difficulty\" for ETL. Extensive experiments have demonstrated the effectiveness of our TaskRes, despite its simplicity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 372, + 556, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 372, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 309, + 372, + 556, + 467 + ], + "type": "text", + "content": "Nevertheless, our method bears some limitations. For instance, our method encounters negative transfer on two datasets, i.e., OxfordPets (1-shot) and Food101 (1-/2-/4-/8-shot). We conjecture that this happens with two conditions: (i) the downstream tasks are of high relative transfer difficulty as shown in Figure 4, and (ii) Zero-shot CLIP has already achieved a fair precision on them. Besides, the assessment of transfer difficulty in this work is heuristic." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 468, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 468, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 309, + 468, + 556, + 598 + ], + "type": "text", + "content": "With the rapid advancement of foundation models [3], it is becoming increasingly crucial to establish precise and reliable metrics for assessing the transfer difficulty of pretrained foundation models to downstream tasks. A comprehensive study of transfer difficulty, which includes distribution analyses [32], is in high demand. Moreover, we can extend the transfer difficulty to a concept-wise level, and investigate the correlation between performance and the occurrence frequency of visual concepts by exploring the CLIP models trained on specific datasets, e.g., SLIP [46] trained on YFCC15M [50, 59]. We will explore these in future work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 609, + 404, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 609, + 404, + 624 + ], + "spans": [ + { + "bbox": [ + 310, + 609, + 404, + 624 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 309, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 309, + 629, + 556, + 715 + ], + "type": "text", + "content": "This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006), the National Research Foundation, Singapore under its Medium Sized Center for Advanced Robotics Technology Innovation, NSFC under Grant U1908209 and 62021001, and ZJNSFC under Grant LQ23F010008." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10906" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 291, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 291, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 291, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 291, + 144 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 291, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 291, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 291, + 190 + ], + "type": "text", + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 291, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 291, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 291, + 246 + ], + "type": "text", + "content": "[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 291, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 291, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 291, + 280 + ], + "type": "text", + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 291, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 291, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 291, + 312 + ], + "type": "text", + "content": "[5] Qi Cai, Yingwei Pan, Ting Yao, Chenggang Yan, and Tao Mei. Memory matching networks for one-shot image recognition. In CVPR, pages 4080-4088, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 291, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 291, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 291, + 346 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 291, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 291, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 291, + 378 + ], + "type": "text", + "content": "[7] Adam Coates and Andrew Y Ng. The importance of encoding versus training with sparse coding and vector quantization. In ICML, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 291, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 291, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 291, + 413 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 291, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 291, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 291, + 445 + ], + "type": "text", + "content": "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 291, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 291, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 291, + 512 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 291, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 291, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 291, + 546 + ], + "type": "text", + "content": "[11] Mohamed Elhoseiny, Babak Saleh, and Ahmed Elgammal. Write a classifier: Zero-shot learning using purely textual descriptions. In ICCV, pages 2584-2591, 2013. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 291, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 291, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 291, + 590 + ], + "type": "text", + "content": "[12] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 291, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 291, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 291, + 624 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, pages 1126–1135. PMLR, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 291, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 291, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 291, + 667 + ], + "type": "text", + "content": "[14] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. NeurIPS, 26, 2013. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 291, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 291, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 291, + 713 + ], + "type": "text", + "content": "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 312, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 554, + 116 + ], + "type": "text", + "content": "[16] Micah Goldblum, Steven Reich, Liam Fowl, Renkun Ni, Valeriiia Cherepanova, and Tom Goldstein. Unraveling meta-learning: Understanding feature representations for few-shot tasks. In ICML, pages 3607-3616. PMLR, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 118, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 554, + 163 + ], + "type": "text", + "content": "[17] Lluis Gomez, Yash Patel, Marçal Rusinol, Dimosthenis Karatzas, and CV Jawahar. Self-supervised learning of visual features through embedding images into text topic spaces. In CVPR, pages 4230-4239, 2017. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 164, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 554, + 196 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 198, + 554, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 554, + 253 + ], + "type": "text", + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 256, + 554, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 256, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 554, + 310 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 312, + 554, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 312, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 312, + 312, + 554, + 346 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 347, + 554, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 347, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 312, + 347, + 554, + 400 + ], + "type": "text", + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, pages 2790-2799. PMLR, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 403, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 403, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 403, + 554, + 458 + ], + "type": "text", + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 460, + 554, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 554, + 493 + ], + "type": "text", + "content": "[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 495, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 495, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 312, + 495, + 554, + 529 + ], + "type": "text", + "content": "[25] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual features from large weakly supervised data. In ECCV, pages 67-84. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 530, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 530, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 312, + 530, + 554, + 563 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2015.5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 565, + 554, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 565, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 312, + 565, + 554, + 597 + ], + "type": "text", + "content": "[27] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IC-CVW, pages 554–561, 2013. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 600, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 554, + 632 + ], + "type": "text", + "content": "[28] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 635, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 554, + 667 + ], + "type": "text", + "content": "[29] Jimmy Lei Ba, Kevin Swersky, Sanja Fidler, et al. Predicting deep zero-shot convolutional neural networks using textual descriptions. In ICCV, pages 4247-4255, 2015. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 669, + 554, + 712 + ], + "type": "text", + "content": "[30] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, pages 3045-3059, Online and Punta Cana, Dominican Republic, Nov. 2021. ACL. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10907" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 291, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 291, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 291, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 291, + 106 + ], + "type": "text", + "content": "[31] Ang Li, Allan Jabri, Armand Joulin, and Laurens Van Der Maaten. Learning visual n-grams from web data. In ICCV, pages 4183–4192, 2017. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 291, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 291, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 291, + 162 + ], + "type": "text", + "content": "[32] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. arXiv preprint arXiv:2204.08790, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 291, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 291, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 291, + 206 + ], + "type": "text", + "content": "[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34:9694-9705, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 291, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 291, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 291, + 239 + ], + "type": "text", + "content": "[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, pages 4582-4597, Online, Aug. 2021. ACL. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 291, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 291, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 291, + 296 + ], + "type": "text", + "content": "[35] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 291, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 291, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 291, + 330 + ], + "type": "text", + "content": "[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 291, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 291, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 291, + 363 + ], + "type": "text", + "content": "[37] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 291, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 291, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 291, + 386 + ], + "type": "text", + "content": "[38] Songhua Liu, Jingwen Ye, Runpeng Yu, and Xinchao Wang. Slimmable dataset condensation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 291, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 291, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 291, + 430 + ], + "type": "text", + "content": "[39] Jiasen Lu, Caiming Xiong, Devi Parikh, and Richard Socher. Knowing when to look: Adaptive attention via a visual sentinel for image captioning. In CVPR, pages 375-383, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 291, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 291, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 291, + 464 + ], + "type": "text", + "content": "[40] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In CVPR, pages 5206-5215, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 291, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 291, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 291, + 498 + ], + "type": "text", + "content": "[41] Zhihe Lu, Sen He, Da Li, Yi-Zhe Song, and Tao Xiang. Prediction calibration for generalized few-shot semantic segmentation. arXiv preprint arXiv:2210.08290, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 291, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 291, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 291, + 543 + ], + "type": "text", + "content": "[42] Zhihe Lu, Sen He, Xiatian Zhu, Li Zhang, Yi-Zhe Song, and Tao Xiang. Simpler is better: Few-shot semantic segmentation with classifier weight transformer. In ICCV, pages 8741-8750, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 291, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 291, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 291, + 578 + ], + "type": "text", + "content": "[43] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 579, + 291, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 291, + 611 + ], + "type": "text", + "content": "[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 291, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 291, + 645 + ], + "type": "text", + "content": "[45] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. NeurIPS, 26, 2013. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 291, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 291, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 291, + 679 + ], + "type": "text", + "content": "[46] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, pages 529-544. Springer, 2022. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 291, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 291, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 291, + 713 + ], + "type": "text", + "content": "[47] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In IC-CVGIP, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 72, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 555, + 105 + ], + "type": "text", + "content": "[48] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, pages 3498-3505. IEEE, 2012. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 106, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 555, + 161 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 162, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 555, + 194 + ], + "type": "text", + "content": "[50] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 195, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 555, + 227 + ], + "type": "text", + "content": "[51] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, pages 5389-5400. PMLR, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 228, + 555, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 555, + 260 + ], + "type": "text", + "content": "[52] Mert Bulent Sariyildiz, Julien Perez, and Diane Larlus. Learning visual representations with caption annotations. In ECCV, pages 153-170. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 261, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 555, + 303 + ], + "type": "text", + "content": "[53] Jürgen Schmidhuber. Evolutionary principles in self-referential learning, or on learning how to learn: the metameta... hook. PhD thesis, Technische Universität München, 1987. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 304, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 304, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 555, + 348 + ], + "type": "text", + "content": "[54] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In EMNLP, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 350, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 555, + 381 + ], + "type": "text", + "content": "[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 383, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 555, + 414 + ], + "type": "text", + "content": "[56] Richard Socher, Milind Ganjoo, Christopher D Manning, and Andrew Ng. Zero-shot learning through cross-modal transfer. NeurIPS, 26, 2013. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 415, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 555, + 448 + ], + "type": "text", + "content": "[57] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 449, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 492 + ], + "type": "text", + "content": "[58] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 493, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 555, + 536 + ], + "type": "text", + "content": "[59] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 537, + 555, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 555, + 569 + ], + "type": "text", + "content": "[60] Lorenzo Torresani, Martin Szummer, and Andrew Fitzgibbon. Efficient object category recognition using classemes. In ECCV, pages 776-789. Springer, 2010. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 613 + ], + "type": "text", + "content": "[61] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 614, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 555, + 647 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 647, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 555, + 679 + ], + "type": "text", + "content": "[63] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 555, + 713 + ], + "type": "text", + "content": "[64] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10908" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 291, + 422 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 73, + 291, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 291, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 291, + 94 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 291, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 291, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 291, + 117 + ], + "type": "text", + "content": "[65] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In ECCV, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 119, + 291, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 119, + 291, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 119, + 291, + 140 + ], + "type": "text", + "content": "[66] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 142, + 291, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 142, + 291, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 142, + 291, + 163 + ], + "type": "text", + "content": "[67] Jingwen Ye, Songhua Liu, and Xinchao Wang. Partial network cloning. In CVPR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 165, + 291, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 165, + 291, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 165, + 291, + 208 + ], + "type": "text", + "content": "[68] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, pages 10819-10829, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 210, + 291, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 210, + 291, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 210, + 291, + 263 + ], + "type": "text", + "content": "[69] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 266, + 291, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 266, + 291, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 266, + 291, + 308 + ], + "type": "text", + "content": "[70] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 310, + 291, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 291, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 291, + 354 + ], + "type": "text", + "content": "[71] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 355, + 291, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 291, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 291, + 388 + ], + "type": "text", + "content": "[72] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 389, + 291, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 389, + 291, + 422 + ], + "spans": [ + { + "bbox": [ + 49, + 389, + 291, + 422 + ], + "type": "text", + "content": "[73] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10909" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_content_list.json b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..109d2507e63f43cc3c5166059737aeb3cfabef15 --- /dev/null +++ b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_content_list.json @@ -0,0 +1,1691 @@ +[ + { + "type": "text", + "text": "Task-specific Fine-tuning via Variational Information Bottleneck for Weakly-supervised Pathology Whole Slide Image Classification", + "text_level": 1, + "bbox": [ + 142, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Honglin Li $^{1,3}$ , Chenglu Zhu $^{2,3}$ , Yunlong Zhang $^{1,3}$ , Yuxuan Sun $^{1,3}$", + "bbox": [ + 225, + 202, + 743, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongyi Shui $^{1,3}$ , Wenwei Kuang $^{3,4}$ , Sunyi Zheng $^{2,3}$ , Lin Yang $^{2,3*}$", + "bbox": [ + 228, + 220, + 745, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ College of Computer Science and Technology, Zhejiang University", + "bbox": [ + 217, + 238, + 751, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Research Center for Industries of the Future and $^{3}$ School of Engineering, Westlake University", + "bbox": [ + 112, + 256, + 856, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The University of Hong Kong", + "bbox": [ + 362, + 273, + 607, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lihonglin, yanglin}@westlake.edu.cn", + "bbox": [ + 328, + 294, + 635, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While Multiple Instance Learning (MIL) has shown promising results in digital Pathology Whole Slide Image (WSI) analysis, such a paradigm still faces performance and generalization problems due to high computational costs and limited supervision of Gigapixel WSIs. To deal with the computation problem, previous methods utilize a frozen model pretrained from ImageNet to obtain representations, however, it may lose key information owing to the large domain gap and hinder the generalization ability without image-level training-time augmentation. Though Self-supervised Learning (SSL) proposes viable representation learning schemes, the downstream task-specific features via partial label tuning are not explored. To alleviate this problem, we propose an efficient WSI fine-tuning framework motivated by the Information Bottleneck theory. The theory enables the framework to find the minimal sufficient statistics of WSI, thus supporting us to fine-tune the backbone into a task-specific representation only depending on WSI-level weak labels. The WSI-MIL problem is further analyzed to theoretically deduce our fine-tuning method. We evaluate the method on five pathological WSI datasets on various WSI heads. The experimental results show significant improvements in both accuracy and generalization compared with previous works. Source code will be available at https://github.com/invoker-LL/WSI-finetuning.", + "bbox": [ + 75, + 377, + 473, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Digital Pathology or microscopic images have been widely used for the diagnosis of cancers such as Breast Cancer [13] and Prostate Cancer [6]. However, the reading of Whole Slide Images (WSIs) with gigapixel resolution is", + "bbox": [ + 75, + 813, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6924865837e9beb58e361df843e2650f9342cb52ad2accdf50e53de468bb4a9.jpg", + "image_caption": [ + "a. ImageNet-1k v-score: 0.285" + ], + "image_footnote": [], + "bbox": [ + 526, + 369, + 691, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b16c017a48df0351af1861e5a76d45e2840b8169638d446149845fb224d0ab5d.jpg", + "image_caption": [ + "b. Full Supervision v-score: 0.745" + ], + "image_footnote": [], + "bbox": [ + 700, + 367, + 864, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8dd7bdc4b45348a884325912f55e7268cc1c8e2c01c4ddad7575050ca58800dc.jpg", + "image_caption": [ + "c. Self-supervised v-score: 0.394" + ], + "image_footnote": [], + "bbox": [ + 526, + 479, + 691, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34c05d368cfa1dafa6cfdbce0d49836e0029f6ed549d71c4eaf2271c900ecb11.jpg", + "image_caption": [ + "d. IB Fine-tuning (ours) \nv-score: 0.538", + "Figure 1. T-SNE visualization of different representations on patches. Our method converts chaotic ImageNet-1K and SSL features into a more task-specific and separable distribution. The cluster evaluation measurement, v-scores, show weakly supervised fine-tuned features are more close to full supervision compared to others. a. ImageNet-1k pretraining. b. Full patch supervision. c. Self-supervised Learning. d. Fine-tuning with WSI labels." + ], + "image_footnote": [], + "bbox": [ + 700, + 479, + 864, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "time-consuming which poses an urgent need for automatic computer-assisted diagnosis. Though computers can boost the speed of the diagnosis process, the enormous size of resolution, over 100M [45], makes it infeasible to acquire precise and exhaustive annotations for model training, let alone the current hardware can hardly support the parallel training on all patches of a WSI. Hence, an annotation-efficient learning scheme with light computation is increasingly desirable to cope with those problems. In pathology WSI analysis, the heavy annotation cost is usually alleviated by Multiple Instance Learning (MIL) with only WSI-level weak supervision, which makes a comprehensive decision on a series of instances as a bag sample [19, 31]. Intuitively, all", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7454", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "small patches in the WSI are regarded as instances to constitute a bag sample [7, 30, 38], where the WSI's category corresponds to the max lesion level of all patch instances.", + "bbox": [ + 78, + 90, + 465, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most methods pay much effort to design WSI architectures while overlooking the instance-level representation ability. Because of the computational limitation, the gradient at the WSI-level is impossible to parallelly backpropagate to instance encoders with more than 10k instances of a bag. Thus parameters of the pretrained backbone from ImageNet-1k (IN-1K) are frozen to obtain invariant embeddings. Due to the large domain gap between IN-1K and pathological images, some essential information may be discarded by layers of frozen convolutional filters, which constrains the accuracy of previous WSI methods. To address the constraint, recent works [9, 25] make efforts to learn a good feature representation at the patchlevel by leveraging Self-supervised Learning (SSL). However, such task-agnostic features are dominated by the proxy objective of SSL, e.g. Contrastive Learning in [8, 11, 17] may push away the distance between two instances within the same category, thus only performs slightly better than IN-1K pretraining in WSI classification. Nearly all SSL methods [8, 11, 16, 17] proposed on natural image recognition utilize a small portion of annotations to get promising fine-tuning accuracy compared to full supervision, which is higher than Linear Probing [17] by a large margin.", + "bbox": [ + 80, + 138, + 467, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These findings illuminate us to design a fine-tuning scheme for WSI analysis to convert IN-K or SSL task-agnostic representations into task-specifics. Motivated by the Information Bottleneck (IB) theory [1, 2], we argue that pretraining is limited to downstream tasks, therefore fine-tuning is necessary for WSI analysis. In addition, we develop a solution based on Variational IB to tackle the dilemma of fine-tuning and computational limitation by its minimal sufficient statistics and attribution properties [1, 23]. The differences among the above three feature representations are depicted in Figure 1, where the feature representation under full patch-level supervision is considered as the upper bound.", + "bbox": [ + 80, + 489, + 467, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are in 3 folds: 1) We propose a simple agent task of WSI-MIL by introducing an IB module that distills over $10\\mathrm{k}$ redundant instances within a bag into less than $1\\mathrm{k}$ of the most supported instances. Thus the parallel computation cost of gradient-based training on Gigapixel Images is over ten times relieved. By learning and making classification on the simplified bag, we find that there are trivial information losses due to the low-rank property of pathological WSI, and the distilled bag makes it possible to train a WSI-MIL model with the feature extractor on patches end-to-end, thus boosting the final performance. 2) We argue that the performance can be further improved by combining with the SSL pretraining since we could convert the task-agnostic representation from SSL into task-specific", + "bbox": [ + 80, + 688, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "one by well-designed fine-tuning. The proposed framework only relies on annotations at WSI levels, which is similar to recent SSL approaches [8, 11, 16, 17]. Note that our method only utilizes less than a $1\\%$ fraction of full patch annotation to achieve competitive accuracy compared to counterparts. 3) Versatile training-time augmentations can be incorporated with our proposed fine-tuning scheme, thus resulting in better generalization in various real-world or simulated datasets with domain shift, which previous works ignore to validate. These empirical results show that our method advances accuracy and generalization simultaneously, and thus would be more practical for real-world applications.", + "bbox": [ + 503, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 305, + 638, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multiple Instance Learning for WSI Analysis", + "text_level": 1, + "bbox": [ + 503, + 335, + 880, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple Instance Learning (MIL) is a well-defined task and has been explored extensively. Currently, there are mainly two lines of methods for WSI analysis: 1) explicit modeling the MIL definition $Y = \\text{pooling}\\{y_1, y_2, \\dots, y_n\\}$ that WSI level prediction is aggregated by the probability of all patches with Mean or Max-pooling [7,45]. 2) implicit learning WSI level representation by aggregating all patches embeddings via a WSI classifier with Recurrent Neural Network (RNN) [7] or in an attention [19] mechanism. The latter shows superior performance since such modeling includes less inductive bias compared to the former processing with fixed weights.", + "bbox": [ + 503, + 366, + 890, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RNN [7] treats WSI as a sequence of patches, but its permutation variation and defects on long-term dependency do not match the nature of WSI well. By contrast, attention-based MIL (AB-MIL) [19] learns the weights of instance representations adaptively, and these methods can be flexibly inserted into the classification framework as a plug-and-play module to generalize the full-field map according to the task. To make patches' representation more discriminative, CLAM [30] introduces an auxiliary task in the MIL framework to distinguish its corresponding availability according to the size of instance attention while training the WSI classification. Trans-MIL [38] is designed to model the relation among patches via Self-Attention [41] and solves the softmax computational complexity in the long sequence by linear Self-Attention [42, 43]. To embed the multi-scale information of WSI, DS-MIL [25] simply concatenates patch features in three scales, and HIPT [9] builds hierarchical multi-scale features. To mitigate overfitting in complex attention model trained on the limited number of WSIs, DTFD-MIL [45] resamples instances from the original bag to generate various sub-bags for augmentation. However, all these works pay too much attention on WSI head to explore the backbone.", + "bbox": [ + 503, + 553, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7455", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fac218ab8ba49211d7bcd248afda7952532317d0a5308b6a7335c3d2bfc25d5a.jpg", + "image_caption": [ + "Figure 2. An illustration of the tuning scheme. To maximize $R_{IB}$ , enlarging the first term helps latent representation vector $Z$ become more predictive for label $Y$ (More overlap between $Y$ and $Z2$ compared to $Z1$ , which denotes higher mutual information), while by reducing the second term to filter irrelevant features (smaller overlap between image $X$ and $Z2$ compared to $Z1$ )." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 851, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Self-supervised Learning and Fine-tuning", + "text_level": 1, + "bbox": [ + 76, + 349, + 431, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised Learning (SSL) has shown to be a promising paradigm both in computer vision [4,8,11,16,17] and natural language processing [12,33,44]. However, fine-tuning (FT) is quite necessary for the pretrained model of SSL to the downstream tasks, e.g. in [16,17], the vanilla FT method can improve about 15 percent of accuracy in IN-1K compared to vanilla Linear Probing [17]. More interesting and parameter-efficient FT methods [14,18,24,26,27,47,48] are proposed according to the similarity between pretraining and downstream in natural language processing and few-shot image generation. Unfortunately, the computational limitation on redundant instances prevents FT from being directly introduced into WSI analysis. Although SSL achieves optimistic performance in some works [9, 25] for pathology WSI analysis but only shows a small margin of improvement. There may be much potential to be discovered in WSI analysis if an appropriate FT method can be used to address the training cost from surplus information of WSI.", + "bbox": [ + 75, + 375, + 472, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Information Bottleneck and Attributions", + "text_level": 1, + "bbox": [ + 76, + 676, + 426, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Information Bottleneck (IB) conception is introduced in [40] as an information-theoretic framework for learning, which is currently used in deep learning both theoretically and practically. [39] proposes to unveil deep neural network in an information flow perspective by estimating Mutual Information (MI) between the outputs of two layers, then in followed-up works [15, 35], they revisit the MI compression process of IB to propose complements and better measurements of MI in Deep Neural Network (DNN).", + "bbox": [ + 75, + 702, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IB not only reveals how the traditional DNN training finishes information compression implicitly in the above works, but can also be employed explicitly as an objective to intervene in DNN's training: [2] derives a variational", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bound of the IB objective, claiming that the proxy objective provides an additional regularization term which improves the robustness to adversarial samples compared to the traditional training of DNN. [32] adopts the variational bound in [2] and replaces the distribution of latent features from Gaussian into Bernoulli. In this way, it generates masks to control the conciseness of the rationale extraction of sentences. Achille et al. [1] claim that IB solutions approximate minimal sufficient statistics (MSS) [23], and by penalizing the redundancy of representations, they heuristically argue that the model's sensitivity to nuisances is mitigated. Similar to some visual attribution methods like CAM [37, 49], LRP [3] and patch masking [4, 16, 28] in ViT, an IB-based attribution method is proposed in [36] by adding noise to intermediate feature maps, restricting the flow of information, then how much information image regions provide can be quantified. Intuitively, the IB objective reveals the limitation of pretraining for downstream tasks and the imperative of FT for task-specific representation in WSI analysis, as shown in Figure 2. Moreover, it motivates us to solve the dilemma of FT and computational limitation by its' MSS and visual attribution properties.", + "bbox": [ + 496, + 349, + 893, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 695, + 591, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview on MIL-based WSI Analysis", + "text_level": 1, + "bbox": [ + 500, + 720, + 831, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a WSI $X$ , the goal is to make slide-level prediction $Y$ by learning a classifier $f(X; \\theta)$ . Due to its extremely high resolution, $X$ is patched into a huge bag of small instances $X = \\{x_1, \\dots, x_N\\}$ , where $N$ is the number of instance. The slide-level supervision $\\hat{Y}$ is given by a Max-pooling operation of the latent label $\\hat{y}_i$ for each instance $x_i$ , which can be defined as:", + "bbox": [ + 496, + 742, + 890, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} = \\max \\left\\{\\hat {y} _ {1}, \\dots , \\hat {y} _ {N} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 847, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since all latent labels of instances $\\hat{y}_i$ are unknown under the WSI-level supervision, conventional approaches con", + "bbox": [ + 496, + 869, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7456", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vert this problem into a MIL formulation in the following two steps: 1) Processing images into feature representations $Z = \\{z_{1},\\dots,z_{N}\\}$ with a backbone $h$ as $z_{i} = h(x_{i};\\theta_{1})$ where $h$ is a model of any architecture such as CNN or ViT with parameters $\\theta_{1}$ . 2) Aggregating all patches' features within a slide and producing the slide-level prediction $Y = g(Z;\\theta_2)$ , where $g$ is an attention-based pooling function followed by a linear classifier head as:", + "bbox": [ + 76, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (Z; \\theta_ {2}) = \\sigma \\left(\\sum_ {i = 1} ^ {N} a _ {i} z _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 220, + 468, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a_{i}$ is attention weights and $\\sigma (\\cdot)$ is a linear head. Limited by the computational cost, the parameters $\\theta_{1}$ and $\\theta_{2}$ in $f(X;\\theta) = g\\{h(X;\\theta_1);\\theta_2\\}$ are learned separately by following steps: 1) Initializing $\\theta_{1}$ from the pretrained model, which refers to general features from public IN-1K, or learned by SSL on the related dataset to extract the domain-specific representations. 2) Freezing $\\theta_{1}$ and learning $\\theta_{2}$ under slide-level supervision.", + "bbox": [ + 76, + 271, + 472, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Information Bottleneck for MIL Sparsity", + "text_level": 1, + "bbox": [ + 76, + 401, + 429, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background of the Information Bottleneck", + "text_level": 1, + "bbox": [ + 76, + 424, + 380, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Information Bottleneck (IB) can work as an information compression role to intervene in DNN's training [2]. The objective function of IB to be maximized is given in [40] as,", + "bbox": [ + 76, + 439, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {I B} = I (Z, Y) - \\beta I (Z, X), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 500, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I(\\cdot, \\cdot)$ indicates the Mutual Information (MI) and $\\beta$ is a Lagrange multiplier controlling the trade-off between the information that the representation variable $Z$ shares with the label $Y$ and its shares with input $X$ . Since the computation of MI is intractable during the training of the neural networks, to maximize IB objective can be transferred to minimize a variational bound of Eq.(3) derived in [2] follows:", + "bbox": [ + 76, + 521, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {z \\sim p _ {\\theta} (z | x _ {n})} [ - \\log q _ {\\phi} (y _ {n} | z) ] + \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 648, + 468, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta K L [ p _ {\\theta} (z | x _ {n}), r (z) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 272, + 691, + 423, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ denotes the number of samples, $q_{\\phi}(y|z)$ is a parametric approximation to the likelihood $p(y|z), r(z)$ is the prior probability of $z$ to variational approximate the marginal $p(z)$ , and $p_{\\theta}(z|x)$ is the parametric posterior distribution over $z$ .", + "bbox": [ + 76, + 719, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learn MIL Sparsity via Variational Bound of IB", + "text_level": 1, + "bbox": [ + 76, + 795, + 419, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To trade off the dilemma of computational limitation and task-specific representation learning via end-to-end backpropagation, we propose to utilize the IB module to filter most task-irrelevant instances for task-specific fine-tuning.", + "bbox": [ + 76, + 810, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above filtering process can be implemented by optimizing the second term of in Eq.(3) which controls the", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compression. There are two ways that compress $X$ to $Z$ by decreasing the KL divergence between $p(z|x)$ and $r(z)$ in Eq.(4) variational method: reducing the dimension of representation $Z$ compared to $X$ in [2], or converting input $X$ into a sparse one in [32].", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the setting of our long instance sequenced MIL, we reduce $I(X,Z)$ into a degree so that the gradients can be back-propagated to the backbone encoder, which needs us to convert a WSI of bag size over 10k into 1k for the sake of sparsity. Considering MIL for tumor v.s. normal binary classification without loss of generality and the latent label $y_{i}$ of each instance $x_{i}$ in Eq.(1), we argue that it is sufficient enough to make the WSI level prediction if one tumor area is detected. With the above understanding, we propose to learn compressed components similar to [32] by defining a IB module as:", + "bbox": [ + 496, + 167, + 892, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = m \\odot x, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 335, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ is a Bernoulli $(\\pi)$ distributed binary mask and in this way $KL[p_{\\theta}(z|x),r(z)]$ in Eq.(4) can be decomposed as,", + "bbox": [ + 496, + 357, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK L \\left[ p _ {\\theta} \\left(m _ {i} | x\\right), r \\left(m _ {i}\\right) \\right] + \\pi H (X), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 402, + 890, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H(X)$ is the entropy of $X$ , which can be omitted during the minimization due to its constant value. Please check Supplementary for above proof.", + "bbox": [ + 496, + 426, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Bernoulli $(\\pi)$ distribution for $m$ fits the definition of MIL empirically: we can treat $m$ as a latent weak prediction $\\hat{y}$ describing whether the patch contains tumor or not, denoting $P_{set} = \\{p(m_1|x_1),\\dots,p(m_N|x_N)\\}$ , then during inference Eq.(1) can be derived as:", + "bbox": [ + 496, + 473, + 892, + 549 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} = \\max \\left\\{P _ {s e t} \\right\\} = \\max \\left\\{P _ {s u b s e t} \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 560, + 890, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{\\text{subset}} \\in P_{\\text{set}}$ , generated by select top-K elements in $P_{\\text{set}}$ . The patch classifier trained with only slide-level supervision shows low accuracy [25], so we only use it to generate mask for sparse sub-bag and still utilize attention-based MIL on the sub-bag for decision making.", + "bbox": [ + 496, + 590, + 890, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Loss Function and its Implementation", + "text_level": 1, + "bbox": [ + 500, + 676, + 828, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Derived from Eq.(4) and Eq.(6), our Variational IB module should be optimized with the following loss function,", + "bbox": [ + 496, + 700, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {l o s s} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {z \\sim p _ {\\theta} (z | x _ {n})} [ - \\log q _ {\\phi} (y _ {n} | z) ] + \\tag {8} \\\\ \\beta K L [ p _ {\\theta} (m | x _ {n}), r (m) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 741, + 890, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first term is the task loss to learn task-specific features, and it can be treated as the cross entropy same to prior works [2, 32] in a sampling perspective. The second term is the information loss to filter out task-irrelevant instances by minimizing the Kullback-Leibler divergence between the distribution of mask and the prior Bernoulli.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7457", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c54e0a4389595a0c7e8e568f600bc565c1e36bc3fd910646357eda573e0de30a.jpg", + "image_caption": [ + "Figure 3. Workflow of WSI-MIL task-specific fine-tuning. 1) Initialize the backbone with pretrained parameters and set frozen, then learn the IB module to generate instance masks. 2) fix the mask to distill a sparse bag, then fine-tune the WSI head and patch the backbone end-2-end. 3) utilize all fine-tuned instance features within a bag and train the WSI-MIL classifier head." + ], + "image_footnote": [], + "bbox": [ + 102, + 92, + 867, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $p(m_{n}|x_{n})$ is first generated by a linear layer from the representation of $x_{n}$ and followed by a sigmoid activation. If the mask $m$ is directly converted by thresholding $p(m_{n}|x_{n})$ , the gradients from $z = m\\odot x$ can not be backpropagated since such an operation is not differentiable. So, during training, $m$ is generated via Monte Carlo sampling, and $p(m_n|x_n)$ can be learned with the reparameterization trick [21, 22] for gradient's estimation.", + "bbox": [ + 75, + 381, + 472, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Task-specific Fine-tuning", + "text_level": 1, + "bbox": [ + 76, + 515, + 307, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To realize task-specific fine-tuning and better performance with only slide-level labels, the workflow of our method is depicted in Figure 3, which includes 3 stages: 1) By minimizing the loss function in section 3.3 with a frozen backbone, a sparse instance set within a bag is obtained. 2) Assisted by the small portion of instances in a bag generated from stage 1, it is possible to fine-tune the backbone for better task-specific features and here we directly backpropagate the loss gradients from slide-level supervision into top-K patch instances. 3) Since the appropriateness of the sparse instance set heavily relies on the performance of the pretrained backbone, the first training IB module is insufficient and may result in small or even zero top-K recall. Moreover, the sparse instance set may lose the contextual dependency of WSI modeling. Thus in this stage, We utilize all fine-tuned instance features within a bag to train a traditional attention-based WSI-MIL classifier.", + "bbox": [ + 75, + 539, + 472, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 828, + 209, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the performance of the proposed method incorporated with the latest WSI-MIL frameworks, where the IN-1K and SSL features are used for com", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "parisons. Ablation experiments are performed to further study the proposed method and for paper length, more experimental results are presented in the Supplementary.", + "bbox": [ + 496, + 381, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Tasks.", + "text_level": 1, + "bbox": [ + 500, + 428, + 640, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use five datasets to evaluate our method. The slide-level classification performance of IN-1K, SSL, and our method is evaluated on three datasets including both histopathology and cytopathology images: two public histopathological WSI datasets, Camelyon-16 [5] for tumor / normal binary classification, The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) [34] for tumors subtyping. One internal cytopathology WSI dataset is introduced to validate the universality of our method on both histo- and cyto-pathology, which is Liquid-based Preparation cytology for Cervical Cancer's early screening (LBP-CECA).", + "bbox": [ + 496, + 444, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite the original evaluation on the closed dataset, we also evaluate the generalization of our method on Camelyon-16-C (generated with random synthetic domain shift from Camelyon-16), and Camelyon-17 [29] from five different centers, which occurs frequently in practical pathological diagnosis and has been hindering the application of automatical WSI analysis to the real world. Details of five datasets can be found in the supplementary.", + "bbox": [ + 496, + 611, + 890, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For pre-processing, we follow the operations in CLAMSB [30] which mainly includes HSV, Blur, Thresholding, and Contours methods to localize the tissue regions in each WSI. Then non-overlapping patches with size $256\\times 256$ on the $20\\mathrm{X}$ magnification are extracted from the tissue regions.", + "bbox": [ + 496, + 732, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining and Fine-tuning.", + "text_level": 1, + "bbox": [ + 500, + 824, + 705, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our work mainly focuses on the method of fine-tuning. Because good pretrained initialization results in better finetuning performance, we employ mainstream pretraining methods: 1) ImageNet-1k (IN-1K) data pretraining. 2) SSL", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7458", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9d8e728b497cc15aeca67b46288a4fcff18695e12e6660992e5adc8d9540a423.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCamelyon-16TCGA-BRCALBP-CECA
F1AUCF1AUCF1AUC
Full Supervision0.967±0.0050.992±0.003--0.741±0.0060.942±0.002
RNN-MIL [7]0.834±0.0170.861±0.0210.776±0.0350.871±0.033--
AB-MIL [19]0.828±0.0130.851±0.0250.771±0.0400.869±0.0370.525±0.0170.845±0.002
DS-MIL [25]0.857±0.0230.892±0.0120.775±0.0440.875±0.041--
CLAM-SB [30]0.839±0.0180.875±0.0280.797±0.0460.879±0.0190.587±0.0140.860±0.005
TransMIL [38]0.846±0.0130.883±0.0090.806±0.0460.889±0.0360.533±0.0060.850±0.007
DTFD-MIL [45]0.882±0.0080.932±0.0160.816±0.0450.895±0.0420.569±0.0260.847±0.003
FT+ CLAM-SB0.911±0.0170.956±0.0130.845±0.0320.935±0.0270.718±0.0100.907±0.005
FT+ TransMIL0.923±0.0120.967±0.0030.848±0.0440.945±0.0200.720±0.0240.918±0.004
FT+ DTFD-MIL0.921±0.0070.962±0.0060.849±0.0270.951±0.0160.723±0.0080.922±0.005
Mean-pooling0.629±0.0290.591±0.0120.818±0.0220.910±0.0320.350±0.0170.735±0.006
Max-pooling0.805±0.0120.824±0.0160.644±0.1790.826±0.0960.636±0.0640.893±0.019
KNN (Mean)0.468±0.0000.506±0.0000.633±0.0660.749±0.0550.393±0.0000.650±0.000
KNN (Max)0.559±0.0000.535±0.0000.524±0.0320.639±0.0630.477±0.0000.743±0.000
FT+ Mean-pooling0.842±0.0060.831±0.0070.866±0.0350.952±0.0180.685±0.0140.900±0.002
FT+ Max-pooling0.927±0.0110.969±0.0040.852±0.0430.948±0.0190.695±0.0130.912±0.004
FT+ KNN (Mean)0.505±0.0000.526±0.0000.784±0.0440.907±0.0340.529±0.0000.737±0.000
FT+ KNN (Max)0.905±0.0000.916±0.0000.802±0.0630.882±0.0360.676±0.0000.875±0.000
", + "bbox": [ + 83, + 90, + 885, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Slide-Level Classification by using the IN-1K pre-trained backbone or the proposed fine-tuned (FT) in three datasets. Top Rows. Different MIL architectures are compared to select the top 3 SOTA methods to validate the transfer learning performance using the IN-1K pre-trained backbone or the FT. Bottom Rows. The competition of various traditional aggregation and feature evaluation methods by using pre-trained IN-1K or the FT.", + "bbox": [ + 75, + 444, + 890, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pretraining with SimCLR [11], MoCo [17] and DINO [8]. 3) for Camelyon-16 and LBP-CECA, since there are comprehensive tumor area annotations, we use the annotation to pretrain the patch backbone, which acts as the upper-bound of our method.", + "bbox": [ + 75, + 522, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different from prior works backbone frozen after pretraining, we perform backbone fine-tuning with the method proposed in 3.4. In this stage, we fine-tune the backbone and WSI model end-to-end with 25 epochs using the AdamW optimizer, batch size of 1 for WSI with bag size of 512, a learning rate of 1e-5 for backbone and 1e-3 for WSI head. For the BN layers in ResNet, we turn it to eval mode to fix statistics during fine-tuning since we find that the distribution of the top-K instances in a bag is limited to estimate statistics because of the similarity among instances. Shallow layers of backbone are frozen since they only focus on morphological features while deep layer focus on semantics.", + "bbox": [ + 75, + 597, + 470, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Slide-level Classification", + "text_level": 1, + "bbox": [ + 76, + 801, + 299, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For all the experiments, macro-AUC and macro-F1 scores are reported since all the 3 datasets are class imbalanced. For Camelyon-16, the official training set is randomly split into training and validation sets with a ratio of 9:1. The experiment is conducted 5 times and", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the results of the official test set are reported. For TCGA-BRCA, we perform the 10 folder cross-validation with the same running setting adopted in HIPT [9]. Besides the dataset LBP-CECA is randomly split with a ratio of 6:1:3 for training, validation, and testing. The experiment is conducted 5 times. The mean and standard variance values of performance metrics are reported for multi-runs or cross-validation runs.", + "bbox": [ + 496, + 522, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with baselines. Classification results are summarized in Table 1. We first show full patch supervision results as an upper bound, then we directly evaluate several classic WSI-MIL methods, including RNN-MIL [7], AB-MIL [19], DS-MIL [25], CLAM-SB [30], TransMIL [38], DTFD-MIL [45]. All the WSI-MIL baselines across the three classification tasks suffer from relatively low performance due to the inappropriate backbone features from the pretrained ResNet-50 in IN-1K. Then we apply FT to the backbone to obtain features for WSI classification by three WSI architectures including CLAM-SB (pure global attention, no intersection among instances), TransMIL (self-attention for long sequence), and DTFD-MIL (multi-tier attention paths followed with aggregation, good for patch imbalance).", + "bbox": [ + 496, + 643, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results show clearly consistent improvements equipped with FT features under AUC metric. CLAM-SB", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7459", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "achieves a performance relative increase of $9.26\\%$ , $6.37\\%$ , $5.47\\%$ on Camelyon-16, TCGA-BRCA, and LBP-CECA respectively with vanilla FT based on IN-1K pretraining. In the same conditions, TransMIL and DTFD-MIL achieve new SOTAs on all three datasets and can obtain better relative improvement, especially in LBP-CECA with an $8.00\\%$ and $8.85\\%$ growth compared to CLAM-SB. The results show similar improvements under the F1 metric. For the bottom two rows of Table 1, we compare simple Mean/Max-pooling training and KNN evaluations.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The competitive results demonstrate that our proposed FT method in the backbone of the MIL framework can boost the performance of WSI classification, even for the most simple feature-level Mean/Max-pooling. In addition, task-specific features help us better unearth the properties of different WSI-MIL architectures. Apparently, the more complex architecture, TransMIL, and DTFD-MIL, can not perform desirable performance with the frozen pretrained parameters in LBP-CECE. In contrast, its capabilities can be further enhanced by fine-tuning the backbone for the defined target. Moreover, the effects of different feature aggregation methods vary depending on the task. The better performance is shown by using Max-pooling aggregation in Camelyon-16 containing smaller tumor areas, while Mean-pooling is better in TCGA-BRCA.", + "bbox": [ + 75, + 242, + 472, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Combination of SSL and Fine-tuning", + "text_level": 1, + "bbox": [ + 76, + 477, + 398, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we further improve the slide-level classification performance by combining SSL with FT.", + "bbox": [ + 75, + 500, + 467, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. TCGA-BRCA is used for the evaluation of its task complexity in tumor subtyping with the same setting to 4.1.", + "bbox": [ + 75, + 530, + 467, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with baselines. Since there is no apparent single semantic object in small histopathology patches, we", + "bbox": [ + 75, + 575, + 467, + 606 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21093e556b52e04337c059319f3408bbf0acbf474e1236e21d0a442963596969.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodF1AUC
IN-1K$-0.884±0.059
IN-1K0.797±0.0460.879±0.019
/w FT0.845±0.0320.935±0.027
SimCLR [11]$-0.879±0.069
MoCo [17]0.804±0.0420.904±0.030
/w FT0.851±0.0290.948±0.026
DINO [8]$-0.886±0.059
DINO0.801±0.0450.891±0.043
/w FT0.848±0.0270.944±0.036
", + "bbox": [ + 78, + 628, + 467, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mainly compare SSLs by Contrastive Learning or augmentation like MoCo [17] and DINO [8], and SimCLR [11]. We show the results of SSLs performed in previous works [9, 10] for fair comparison. All results are performed with the same WSI architecture of CLAM-SB [30].", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental results are summarized in Table 2. Compared to vanilla IN-1K, MoCo and DINO (combined with FT) achieve consistent growth of $7.85\\%$ and $7.39\\%$ , respectively. Besides, they have a slight increase of $1.39\\%$ and $0.96\\%$ in comparison to IN-1K with FT. By SSL, the intrinsic task-agnostic features are learned from all patches in WSIs, and after applying our proposed FT for all modules, task-specific features can be distilled from the label and partial data. In such an SSL with an FT paradigm, the data and the label are explored comprehensively to produce a state-of-the-art WSI analysis.", + "bbox": [ + 496, + 167, + 890, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Generalization on Domain Shift", + "text_level": 1, + "bbox": [ + 498, + 340, + 779, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the generalization of slide-level classification models on domain shift [46] [20], which is crucial for real-world clinical applications due to the diversity in staining, preparation, and imaging devices for pathological image processing among hospitals.", + "bbox": [ + 496, + 364, + 890, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. The Camelyon-16-C dataset is generated from Camelyon-16 test set by synthetic domain shift, which adopts a random combination of Brightness, JPEG, and Hue proposed in [46]. For Camelyon-17, since it shares similarities to Camelyon-16 but is collected from five different medical centers, we randomly collect 30 samples from each center to evaluate the robustness of natural domain shift. We directly evaluate models on Camelyon-16-C and 150 extra data from Camelyon-17 with five times of running same as the previous setting in section 4.1. The standard variance values of results is omitted due to the constraint on", + "bbox": [ + 496, + 440, + 890, + 604 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f071be5fbe5a534fbb2a1ed38d5d6f348eeb3d60c9a454498b0cf80b4339f512.jpg", + "table_caption": [ + "Table 2. Combination of SSL and Fine-tuning. We compare SSLs with IN-1K and their further improvement via fine-tuning (FT) on TCGA-BRCA. The symbol $\\S$ indicates the result released in previous publication [9, 10]." + ], + "table_footnote": [], + "table_body": "
MethodCamelyon-16-CCamelyon-17
F1AUCF1AUC
Max-pooling0.6890.7420.5780.670
/w FT0.8160.8920.6870.720
CLAM-SB [30]0.7420.8360.6240.702
/w FT0.8230.8620.6760.725
TransMIL [38]0.7480.8420.6570.706
/w FT0.7950.8570.6840.717
DTFD-MIL [45]0.7750.7990.5760.676
/w FT0.8040.8380.6890.717
", + "bbox": [ + 500, + 628, + 893, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Generalization on Domain Shift. The generalization ability of all methods is compared between fine-tuning(FT) and IN-1K features on two datasets with domain shift. Camelyon-16-C and Camelyon-17 are synthetic and real corruptions respectively.", + "bbox": [ + 496, + 828, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7460", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the paper length and a more clear comprehensive comparison can be found in Supplementary.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with baselines. The evaluation of generalization results is summarized in Table 3. For WSI model architecture, despite CLAM-SB [30], TransMIL [38], DTFDMIL [45], we also compare Max-pooling on feature level since it shows promising results on Camelyon-16 in 4.1.", + "bbox": [ + 75, + 121, + 467, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results demonstrate consistent improvements after equipping the WSI head with fine-tuned features under the F1 metric: In general, all models can resist domain shifts to a certain extent by utilizing the proposed fine-tuning. Compared with freezing IN-1K parameters, Max-pooling obtains a noticeable performance increase of $18.43\\%$ and $18.86\\%$ on Camelyon-16-C and Camelyon-17, respectively. CLAM-SBachieves an increase of $10.91\\%$ , $8.33\\%$ respectively. TransMIL achieves an increase of $6.28\\%$ , $4.11\\%$ respectively. DTFD-MIL achieves an increase of $2.90\\%$ , $11.30\\%$ respectively. In contrast, the generalization of DTFD-MIL can still be further improved by our proposed FT optimization, even though it is already robust to different domains.", + "bbox": [ + 75, + 196, + 467, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interestingly, all improvements in performance under the AUC metric are much less than F1. The three Attention-based pooling methods respectively achieve the average improvement of $6.70\\%$ and $7.91\\%$ on Camelyon-16-C and Camelyon-17 under the F1 metric, as compared to $2.93\\%$ and $2.98\\%$ under AUC, which indicates that the classification ability of WSI-MIL model (measured by F1) may be much weaker than ranking (measured by AUC). However, doctors may much more care about F1 instead of AUC in clinical diagnosis from sensitivity and specificity perspectives. Most importantly, the classification threshold usually keeps fixed after deploying the model in the real world, which reflects the contribution of our method is more meaningful in practical applications.", + "bbox": [ + 75, + 407, + 467, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further Ablation Experiments", + "text_level": 1, + "bbox": [ + 76, + 627, + 346, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional ablations are included in the Supplementary Materials, with the main focus on the effects of learning rate on the backbone, Number selection of Top-K, and Value selection of Lagrange multiplier, results of the 3 stages.", + "bbox": [ + 75, + 650, + 467, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Interpretability and Visualization", + "text_level": 1, + "bbox": [ + 76, + 719, + 369, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we further show the interpretability improvements with FT features. As shown in Figure 4, attention scores from CLAM-SB [30] were visualized as a heatmap to determine the ROI and interpret the important morphology used for diagnosis. Obviously, the model focuses more on tumors with FT features.", + "bbox": [ + 75, + 742, + 467, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present a fine-tuning method for WSI classification under the weak supervision of slide-level la", + "bbox": [ + 76, + 869, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b211ae4ea0c3503d27702e6b6d7b236f40c72927c643e08ee51d54ff9c54a673.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 92, + 630, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99b9450acde0b21e7790cf12353b56f38823a2511e9a8004c2a1ea5021447cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 92, + 758, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20de2c06e6cfb7ba2f8d76ba52a51339dc310b9caaac6fc28a7a48a77b2372ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 92, + 885, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc20e19c7581bb63c363320446a19eb1812a80c084ac873642276f76a7279f77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 191, + 630, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ab4e29709964280adfead2dd5a0b2c9bfe7b591a9d60f845e66df2aad708758.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 191, + 758, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25dcc4119d2965536db446e2d47396faba673305d449e36531fc62b8ea664020.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 191, + 885, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/221afd1a32da494642922b79bce36520da32aaa1d646130939b95b75cbac5f20.jpg", + "image_caption": [ + "Figure 4. Heatmap comparison between pretraining and FT. The first row shows the full annotations, where red contours denote the tumor area and green contours filter the nonsense background. The second and third row shows the WSI attention map on pretraining and FT features respectively." + ], + "image_footnote": [], + "bbox": [ + 504, + 290, + 630, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abb2288ea1e58537763e53aea0130eba9751189ec26d37fc4a9e6af666063e28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 290, + 758, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17a386dc455de9ced873f15ee0330f12b3411f430ecb56ec3dac28b1cfb74e8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 290, + 885, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "belts. Initially, an effective IB module is introduced to mitigate the training cost of Gigapixel WSI, which distills the oversized bag into a sparse one. Then the backbone of the instance is able to be trained end-to-end in the MIL framework by learning and making classifications on the distilled bag. Thus the WSI classification performance is improved through the retention of mainly task-specific information. In addition, SSL can be combined with the proposed framework for further improvement. Compared with fully supervised learning, our methods can achieve competitive accuracy by utilizing extremely weak WSI labels. Furthermore, our training scheme can introduce versatile training-time augmentations for better generalization on datasets with domain shift, which is an inevitable challenge for previous work. The experimental results reflect the advances of our method in both accuracy and generalization. Overall, Our proposed approach shows strong potential for MIL application in real-world pathology diagnosis, with better performance, faster convergence and annotation efficiency.", + "bbox": [ + 496, + 488, + 890, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 789, + 658, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study was partially supported by the National Natural Science Foundation of China (Grant no. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant no. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University.", + "bbox": [ + 498, + 814, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7461", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy computation. IEEE transactions on pattern analysis and machine intelligence, 40(12):2897-2905, 2018. 2, 3", + "[2] Alexander A. Alemi, Ian Fischer, Joshua V. Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 2, 3, 4", + "[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10(7):e0130140, 2015. 3", + "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021.3", + "[5] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 5", + "[6] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature medicine, 28(1):154-163, 2022. 1", + "[7] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2, 6", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. CoRR, abs/2104.14294, 2021. 2, 3, 6, 7", + "[9] Richard J. Chen and et al. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In CVPR, pages 16144-16155, June 2022. 2, 3, 6, 7", + "[10] Richard J Chen and Rahul G Krishnan. Self-supervised vision transformers learn visual concepts in histopathology. arXiv preprint arXiv:2203.00585, 2022. 7", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6, 7", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2018. cite arxiv:1810.04805Comment: 13 pages. 3", + "[13] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Littjens, Jeroen A. W. M. van der Laak, , and the CAME-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "LYON16 Consortium. Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA, 318(22):2199-2210, 12 2017. 1", + "[14] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30(4):681–694, 2020. 3", + "[15] Ziv Goldfeld, Ewout van den Berg, Kristjan Greenewald, Igor Melnyk, Nam Nguyen, Brian Kingsbury, and Yury Polyanskiy. Estimating information flow in deep neural networks. arXiv preprint arXiv:1810.05728, 2018. 3", + "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 2, 3", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3, 6, 7", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[19] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 2127-2136. PMLR, 10-15 Jul 2018. 1, 2, 6", + "[20] Syed Ashar Javed, Dinkar Juyal, Zahl Shanis, Shreya Chakraborty, Harsha Pokkalla, and Aaditya Prakash. Rethinking machine learning model evaluation in pathology. arXiv preprint arXiv:2204.05205, 2022. 7", + "[21] Durk P Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. Advances in neural information processing systems, 28, 2015. 5", + "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5", + "[23] Erich Leo Lehmann and Henry Scheffé. Completeness, similar regions, and unbiased estimation-part i. In Selected works of EL Lehmann, pages 233-268. Springer, 2012. 2, 3", + "[24] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3", + "[25] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2021. 2, 3, 4, 6", + "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[27] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 3", + "[29] Geert Litjens, Peter Bandi, Babak Ehteshami Bejnordi, Oscar Geessink, Maschenka Balkenhol, Peter Bult, Altuna Halilovic, Meyke Hermsen, Rob van de Loo, Rob Vogels, et al. 1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset. *GigaScience*, 7(6):giy065, 2018. 5", + "[30] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 2, 5, 6, 7, 8", + "[31] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1", + "[32] Bhargavi Paranjape, Mandar Joshi, John Thickstun, Hannaneh Hajishirzi, and Luke Zettlemoyer. An information bottleneck approach for controlling conciseness in rationale extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1938-1952, Online, Nov. 2020. Association for Computational Linguistics. 3, 4", + "[33] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. 3", + "[34] Nicholas A Petrick, Shazia Akbar, Kenny HH Cha, Sharon Nofech-Mozes, Berkman Sahiner, Marios A Gavrielides, Jayashree Kalpathy-Cramer, Karen Drukker, Anne LL Martel, et al. Spie-aapm-nci breastpathq challenge: an image analysis challenge for quantitative tumor cellularity assessment in breast cancer histology images following neoadjuvant treatment. Journal of Medical Imaging, 8(3):034501, 2021. 5", + "[35] Andrew M Saxe, Yamini Bansal, Joel Dapello, Madhu Advani, Artemy Kolchinsky, Brendan D Tracey, and David D Cox. On the information bottleneck theory of deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124020, 2019. 3", + "[36] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. arXiv preprint arXiv:2001.00396, 2020. 3", + "[37] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 3", + "[38] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, and yongbing zhang. Transmil: Transformer based correlated multiple instance learning for whole" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "slide image classification. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2136-2147. Curran Associates, Inc., 2021. 2, 6, 7, 8", + "[39] Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017. 3", + "[40] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3, 4", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2", + "[42] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 2", + "[43] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. 2021. 2", + "[44] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. Xlnet: Generalized autoregressive pretraining for language understanding. In NeurIPS, 2019. 3", + "[45] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E. Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. ArXiv, abs/2203.12081, 2022. 1, 2, 6, 7, 8", + "[46] Yunlong Zhang, Yuxuan Sun, Honglin Li, Sunyi Zheng, Chenglu Zhu, and Lin Yang. Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2022. 7", + "[47] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdolahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3", + "[48] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9140-9150, June 2022. 3", + "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016. 3" + ], + "bbox": [ + 501, + 92, + 890, + 852 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7463", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_model.json b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4cafd6365be52987d28275b52f9b8c6faff8489 --- /dev/null +++ b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "Task-specific Fine-tuning via Variational Information Bottleneck for Weakly-supervised Pathology Whole Slide Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.203, + 0.744, + 0.221 + ], + "angle": 0, + "content": "Honglin Li\\(^{1,3}\\), Chenglu Zhu\\(^{2,3}\\), Yunlong Zhang\\(^{1,3}\\), Yuxuan Sun\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.221, + 0.746, + 0.24 + ], + "angle": 0, + "content": "Zhongyi Shui\\(^{1,3}\\), Wenwei Kuang\\(^{3,4}\\), Sunyi Zheng\\(^{2,3}\\), Lin Yang\\(^{2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.239, + 0.753, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)College of Computer Science and Technology, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.257, + 0.857, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Research Center for Industries of the Future and \\(^{3}\\)School of Engineering, Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.275, + 0.609, + 0.293 + ], + "angle": 0, + "content": "4The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.295, + 0.637, + 0.31 + ], + "angle": 0, + "content": "{lihonglin, yanglin}@westlake.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.474, + 0.756 + ], + "angle": 0, + "content": "While Multiple Instance Learning (MIL) has shown promising results in digital Pathology Whole Slide Image (WSI) analysis, such a paradigm still faces performance and generalization problems due to high computational costs and limited supervision of Gigapixel WSIs. To deal with the computation problem, previous methods utilize a frozen model pretrained from ImageNet to obtain representations, however, it may lose key information owing to the large domain gap and hinder the generalization ability without image-level training-time augmentation. Though Self-supervised Learning (SSL) proposes viable representation learning schemes, the downstream task-specific features via partial label tuning are not explored. To alleviate this problem, we propose an efficient WSI fine-tuning framework motivated by the Information Bottleneck theory. The theory enables the framework to find the minimal sufficient statistics of WSI, thus supporting us to fine-tune the backbone into a task-specific representation only depending on WSI-level weak labels. The WSI-MIL problem is further analyzed to theoretically deduce our fine-tuning method. We evaluate the method on five pathological WSI datasets on various WSI heads. The experimental results show significant improvements in both accuracy and generalization compared with previous works. Source code will be available at https://github.com/invoker-LL/WSI-finetuning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Digital Pathology or microscopic images have been widely used for the diagnosis of cancers such as Breast Cancer [13] and Prostate Cancer [6]. However, the reading of Whole Slide Images (WSIs) with gigapixel resolution is" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.343, + 0.658, + 0.365 + ], + "angle": 0, + "content": "a. ImageNet-1k v-score: 0.285" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.37, + 0.692, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.456, + 0.664, + 0.477 + ], + "angle": 0, + "content": "c. Self-supervised v-score: 0.394" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.343, + 0.839, + 0.365 + ], + "angle": 0, + "content": "b. Full Supervision v-score: 0.745" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.368, + 0.866, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.456, + 0.852, + 0.477 + ], + "angle": 0, + "content": "d. IB Fine-tuning (ours) \nv-score: 0.538" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.481, + 0.692, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.481, + 0.865, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.677 + ], + "angle": 0, + "content": "Figure 1. T-SNE visualization of different representations on patches. Our method converts chaotic ImageNet-1K and SSL features into a more task-specific and separable distribution. The cluster evaluation measurement, v-scores, show weakly supervised fine-tuned features are more close to full supervision compared to others. a. ImageNet-1k pretraining. b. Full patch supervision. c. Self-supervised Learning. d. Fine-tuning with WSI labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "time-consuming which poses an urgent need for automatic computer-assisted diagnosis. Though computers can boost the speed of the diagnosis process, the enormous size of resolution, over 100M [45], makes it infeasible to acquire precise and exhaustive annotations for model training, let alone the current hardware can hardly support the parallel training on all patches of a WSI. Hence, an annotation-efficient learning scheme with light computation is increasingly desirable to cope with those problems. In pathology WSI analysis, the heavy annotation cost is usually alleviated by Multiple Instance Learning (MIL) with only WSI-level weak supervision, which makes a comprehensive decision on a series of instances as a bag sample [19, 31]. Intuitively, all" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7454" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.467, + 0.137 + ], + "angle": 0, + "content": "small patches in the WSI are regarded as instances to constitute a bag sample [7, 30, 38], where the WSI's category corresponds to the max lesion level of all patch instances." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.14, + 0.468, + 0.487 + ], + "angle": 0, + "content": "However, most methods pay much effort to design WSI architectures while overlooking the instance-level representation ability. Because of the computational limitation, the gradient at the WSI-level is impossible to parallelly backpropagate to instance encoders with more than 10k instances of a bag. Thus parameters of the pretrained backbone from ImageNet-1k (IN-1K) are frozen to obtain invariant embeddings. Due to the large domain gap between IN-1K and pathological images, some essential information may be discarded by layers of frozen convolutional filters, which constrains the accuracy of previous WSI methods. To address the constraint, recent works [9, 25] make efforts to learn a good feature representation at the patchlevel by leveraging Self-supervised Learning (SSL). However, such task-agnostic features are dominated by the proxy objective of SSL, e.g. Contrastive Learning in [8, 11, 17] may push away the distance between two instances within the same category, thus only performs slightly better than IN-1K pretraining in WSI classification. Nearly all SSL methods [8, 11, 16, 17] proposed on natural image recognition utilize a small portion of annotations to get promising fine-tuning accuracy compared to full supervision, which is higher than Linear Probing [17] by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.491, + 0.468, + 0.686 + ], + "angle": 0, + "content": "These findings illuminate us to design a fine-tuning scheme for WSI analysis to convert IN-K or SSL task-agnostic representations into task-specifics. Motivated by the Information Bottleneck (IB) theory [1, 2], we argue that pretraining is limited to downstream tasks, therefore fine-tuning is necessary for WSI analysis. In addition, we develop a solution based on Variational IB to tackle the dilemma of fine-tuning and computational limitation by its minimal sufficient statistics and attribution properties [1, 23]. The differences among the above three feature representations are depicted in Figure 1, where the feature representation under full patch-level supervision is considered as the upper bound." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.689, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Our main contributions are in 3 folds: 1) We propose a simple agent task of WSI-MIL by introducing an IB module that distills over \\(10\\mathrm{k}\\) redundant instances within a bag into less than \\(1\\mathrm{k}\\) of the most supported instances. Thus the parallel computation cost of gradient-based training on Gigapixel Images is over ten times relieved. By learning and making classification on the simplified bag, we find that there are trivial information losses due to the low-rank property of pathological WSI, and the distilled bag makes it possible to train a WSI-MIL model with the feature extractor on patches end-to-end, thus boosting the final performance. 2) We argue that the performance can be further improved by combining with the SSL pretraining since we could convert the task-agnostic representation from SSL into task-specific" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.272 + ], + "angle": 0, + "content": "one by well-designed fine-tuning. The proposed framework only relies on annotations at WSI levels, which is similar to recent SSL approaches [8, 11, 16, 17]. Note that our method only utilizes less than a \\(1\\%\\) fraction of full patch annotation to achieve competitive accuracy compared to counterparts. 3) Versatile training-time augmentations can be incorporated with our proposed fine-tuning scheme, thus resulting in better generalization in various real-world or simulated datasets with domain shift, which previous works ignore to validate. These empirical results show that our method advances accuracy and generalization simultaneously, and thus would be more practical for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.306, + 0.64, + 0.321 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.337, + 0.881, + 0.352 + ], + "angle": 0, + "content": "2.1. Multiple Instance Learning for WSI Analysis" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.367, + 0.891, + 0.547 + ], + "angle": 0, + "content": "Multiple Instance Learning (MIL) is a well-defined task and has been explored extensively. Currently, there are mainly two lines of methods for WSI analysis: 1) explicit modeling the MIL definition \\( Y = \\text{pooling}\\{y_1, y_2, \\dots, y_n\\} \\) that WSI level prediction is aggregated by the probability of all patches with Mean or Max-pooling [7,45]. 2) implicit learning WSI level representation by aggregating all patches embeddings via a WSI classifier with Recurrent Neural Network (RNN) [7] or in an attention [19] mechanism. The latter shows superior performance since such modeling includes less inductive bias compared to the former processing with fixed weights." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.554, + 0.891, + 0.9 + ], + "angle": 0, + "content": "RNN [7] treats WSI as a sequence of patches, but its permutation variation and defects on long-term dependency do not match the nature of WSI well. By contrast, attention-based MIL (AB-MIL) [19] learns the weights of instance representations adaptively, and these methods can be flexibly inserted into the classification framework as a plug-and-play module to generalize the full-field map according to the task. To make patches' representation more discriminative, CLAM [30] introduces an auxiliary task in the MIL framework to distinguish its corresponding availability according to the size of instance attention while training the WSI classification. Trans-MIL [38] is designed to model the relation among patches via Self-Attention [41] and solves the softmax computational complexity in the long sequence by linear Self-Attention [42, 43]. To embed the multi-scale information of WSI, DS-MIL [25] simply concatenates patch features in three scales, and HIPT [9] builds hierarchical multi-scale features. To mitigate overfitting in complex attention model trained on the limited number of WSIs, DTFD-MIL [45] resamples instances from the original bag to generate various sub-bags for augmentation. However, all these works pay too much attention on WSI head to explore the backbone." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.092, + 0.852, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Figure 2. An illustration of the tuning scheme. To maximize \\( R_{IB} \\), enlarging the first term helps latent representation vector \\( Z \\) become more predictive for label \\( Y \\) (More overlap between \\( Y \\) and \\( Z2 \\) compared to \\( Z1 \\), which denotes higher mutual information), while by reducing the second term to filter irrelevant features (smaller overlap between image \\( X \\) and \\( Z2 \\) compared to \\( Z1 \\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.35, + 0.433, + 0.367 + ], + "angle": 0, + "content": "2.2. Self-supervised Learning and Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.661 + ], + "angle": 0, + "content": "Self-supervised Learning (SSL) has shown to be a promising paradigm both in computer vision [4,8,11,16,17] and natural language processing [12,33,44]. However, fine-tuning (FT) is quite necessary for the pretrained model of SSL to the downstream tasks, e.g. in [16,17], the vanilla FT method can improve about 15 percent of accuracy in IN-1K compared to vanilla Linear Probing [17]. More interesting and parameter-efficient FT methods [14,18,24,26,27,47,48] are proposed according to the similarity between pretraining and downstream in natural language processing and few-shot image generation. Unfortunately, the computational limitation on redundant instances prevents FT from being directly introduced into WSI analysis. Although SSL achieves optimistic performance in some works [9, 25] for pathology WSI analysis but only shows a small margin of improvement. There may be much potential to be discovered in WSI analysis if an appropriate FT method can be used to address the training cost from surplus information of WSI." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.427, + 0.693 + ], + "angle": 0, + "content": "2.3. Information Bottleneck and Attributions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.839 + ], + "angle": 0, + "content": "The Information Bottleneck (IB) conception is introduced in [40] as an information-theoretic framework for learning, which is currently used in deep learning both theoretically and practically. [39] proposes to unveil deep neural network in an information flow perspective by estimating Mutual Information (MI) between the outputs of two layers, then in followed-up works [15, 35], they revisit the MI compression process of IB to propose complements and better measurements of MI in Deep Neural Network (DNN)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "IB not only reveals how the traditional DNN training finishes information compression implicitly in the above works, but can also be employed explicitly as an objective to intervene in DNN's training: [2] derives a variational" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.685 + ], + "angle": 0, + "content": "bound of the IB objective, claiming that the proxy objective provides an additional regularization term which improves the robustness to adversarial samples compared to the traditional training of DNN. [32] adopts the variational bound in [2] and replaces the distribution of latent features from Gaussian into Bernoulli. In this way, it generates masks to control the conciseness of the rationale extraction of sentences. Achille et al. [1] claim that IB solutions approximate minimal sufficient statistics (MSS) [23], and by penalizing the redundancy of representations, they heuristically argue that the model's sensitivity to nuisances is mitigated. Similar to some visual attribution methods like CAM [37, 49], LRP [3] and patch masking [4, 16, 28] in ViT, an IB-based attribution method is proposed in [36] by adding noise to intermediate feature maps, restricting the flow of information, then how much information image regions provide can be quantified. Intuitively, the IB objective reveals the limitation of pretraining for downstream tasks and the imperative of FT for task-specific representation in WSI analysis, as shown in Figure 2. Moreover, it motivates us to solve the dilemma of FT and computational limitation by its' MSS and visual attribution properties." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.593, + 0.711 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.833, + 0.737 + ], + "angle": 0, + "content": "3.1. Overview on MIL-based WSI Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.848 + ], + "angle": 0, + "content": "Given a WSI \\(X\\), the goal is to make slide-level prediction \\(Y\\) by learning a classifier \\(f(X; \\theta)\\). Due to its extremely high resolution, \\(X\\) is patched into a huge bag of small instances \\(X = \\{x_1, \\dots, x_N\\}\\), where \\(N\\) is the number of instance. The slide-level supervision \\(\\hat{Y}\\) is given by a Max-pooling operation of the latent label \\(\\hat{y}_i\\) for each instance \\(x_i\\), which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.848, + 0.892, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} = \\max \\left\\{\\hat {y} _ {1}, \\dots , \\hat {y} _ {N} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Since all latent labels of instances \\(\\hat{y}_i\\) are unknown under the WSI-level supervision, conventional approaches con" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "vert this problem into a MIL formulation in the following two steps: 1) Processing images into feature representations \\( Z = \\{z_{1},\\dots,z_{N}\\} \\) with a backbone \\( h \\) as \\( z_{i} = h(x_{i};\\theta_{1}) \\) where \\( h \\) is a model of any architecture such as CNN or ViT with parameters \\( \\theta_{1} \\). 2) Aggregating all patches' features within a slide and producing the slide-level prediction \\( Y = g(Z;\\theta_2) \\), where \\( g \\) is an attention-based pooling function followed by a linear classifier head as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.222, + 0.47, + 0.263 + ], + "angle": 0, + "content": "\\[\ng (Z; \\theta_ {2}) = \\sigma \\left(\\sum_ {i = 1} ^ {N} a _ {i} z _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.272, + 0.473, + 0.395 + ], + "angle": 0, + "content": "where \\(a_{i}\\) is attention weights and \\(\\sigma (\\cdot)\\) is a linear head. Limited by the computational cost, the parameters \\(\\theta_{1}\\) and \\(\\theta_{2}\\) in \\(f(X;\\theta) = g\\{h(X;\\theta_1);\\theta_2\\}\\) are learned separately by following steps: 1) Initializing \\(\\theta_{1}\\) from the pretrained model, which refers to general features from public IN-1K, or learned by SSL on the related dataset to extract the domain-specific representations. 2) Freezing \\(\\theta_{1}\\) and learning \\(\\theta_{2}\\) under slide-level supervision." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.402, + 0.43, + 0.419 + ], + "angle": 0, + "content": "3.2. Information Bottleneck for MIL Sparsity" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.425, + 0.381, + 0.439 + ], + "angle": 0, + "content": "Background of the Information Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.47, + 0.5 + ], + "angle": 0, + "content": "The Information Bottleneck (IB) can work as an information compression role to intervene in DNN's training [2]. The objective function of IB to be maximized is given in [40] as," + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.5, + 0.469, + 0.517 + ], + "angle": 0, + "content": "\\[\nR _ {I B} = I (Z, Y) - \\beta I (Z, X), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.47, + 0.643 + ], + "angle": 0, + "content": "where \\( I(\\cdot, \\cdot) \\) indicates the Mutual Information (MI) and \\( \\beta \\) is a Lagrange multiplier controlling the trade-off between the information that the representation variable \\( Z \\) shares with the label \\( Y \\) and its shares with input \\( X \\). Since the computation of MI is intractable during the training of the neural networks, to maximize IB objective can be transferred to minimize a variational bound of Eq.(3) derived in [2] follows:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.649, + 0.47, + 0.69 + ], + "angle": 0, + "content": "\\[\nJ _ {I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {z \\sim p _ {\\theta} (z | x _ {n})} [ - \\log q _ {\\phi} (y _ {n} | z) ] + \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.274, + 0.693, + 0.424, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\beta K L [ p _ {\\theta} (z | x _ {n}), r (z) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.795 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of samples, \\(q_{\\phi}(y|z)\\) is a parametric approximation to the likelihood \\(p(y|z), r(z)\\) is the prior probability of \\(z\\) to variational approximate the marginal \\(p(z)\\), and \\(p_{\\theta}(z|x)\\) is the parametric posterior distribution over \\(z\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.796, + 0.421, + 0.81 + ], + "angle": 0, + "content": "Learn MIL Sparsity via Variational Bound of IB" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.871 + ], + "angle": 0, + "content": "To trade off the dilemma of computational limitation and task-specific representation learning via end-to-end backpropagation, we propose to utilize the IB module to filter most task-irrelevant instances for task-specific fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The above filtering process can be implemented by optimizing the second term of in Eq.(3) which controls the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "compression. There are two ways that compress \\( X \\) to \\( Z \\) by decreasing the KL divergence between \\( p(z|x) \\) and \\( r(z) \\) in Eq.(4) variational method: reducing the dimension of representation \\( Z \\) compared to \\( X \\) in [2], or converting input \\( X \\) into a sparse one in [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.332 + ], + "angle": 0, + "content": "For the setting of our long instance sequenced MIL, we reduce \\( I(X,Z) \\) into a degree so that the gradients can be back-propagated to the backbone encoder, which needs us to convert a WSI of bag size over 10k into 1k for the sake of sparsity. Considering MIL for tumor v.s. normal binary classification without loss of generality and the latent label \\( y_{i} \\) of each instance \\( x_{i} \\) in Eq.(1), we argue that it is sufficient enough to make the WSI level prediction if one tumor area is detected. With the above understanding, we propose to learn compressed components similar to [32] by defining a IB module as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.336, + 0.892, + 0.351 + ], + "angle": 0, + "content": "\\[\nz = m \\odot x, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.403 + ], + "angle": 0, + "content": "where \\(m\\) is a Bernoulli \\((\\pi)\\) distributed binary mask and in this way \\(KL[p_{\\theta}(z|x),r(z)]\\) in Eq.(4) can be decomposed as," + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.404, + 0.892, + 0.421 + ], + "angle": 0, + "content": "\\[\nK L \\left[ p _ {\\theta} \\left(m _ {i} | x\\right), r \\left(m _ {i}\\right) \\right] + \\pi H (X), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.473 + ], + "angle": 0, + "content": "where \\( H(X) \\) is the entropy of \\( X \\), which can be omitted during the minimization due to its constant value. Please check Supplementary for above proof." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.55 + ], + "angle": 0, + "content": "The Bernoulli \\((\\pi)\\) distribution for \\(m\\) fits the definition of MIL empirically: we can treat \\(m\\) as a latent weak prediction \\(\\hat{y}\\) describing whether the patch contains tumor or not, denoting \\(P_{set} = \\{p(m_1|x_1),\\dots,p(m_N|x_N)\\}\\), then during inference Eq.(1) can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.561, + 0.892, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} = \\max \\left\\{P _ {s e t} \\right\\} = \\max \\left\\{P _ {s u b s e t} \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.667 + ], + "angle": 0, + "content": "where \\( P_{\\text{subset}} \\in P_{\\text{set}} \\), generated by select top-K elements in \\( P_{\\text{set}} \\). The patch classifier trained with only slide-level supervision shows low accuracy [25], so we only use it to generate mask for sparse sub-bag and still utilize attention-based MIL on the sub-bag for decision making." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.829, + 0.693 + ], + "angle": 0, + "content": "3.3. Loss Function and its Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Derived from Eq.(4) and Eq.(6), our Variational IB module should be optimized with the following loss function," + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.742, + 0.892, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {l o s s} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {z \\sim p _ {\\theta} (z | x _ {n})} [ - \\log q _ {\\phi} (y _ {n} | z) ] + \\tag {8} \\\\ \\beta K L [ p _ {\\theta} (m | x _ {n}), r (m) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where the first term is the task loss to learn task-specific features, and it can be treated as the cross entropy same to prior works [2, 32] in a sampling perspective. The second term is the information loss to filter out task-irrelevant instances by minimizing the Kullback-Leibler divergence between the distribution of mask and the prior Bernoulli." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7457" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.093, + 0.868, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Workflow of WSI-MIL task-specific fine-tuning. 1) Initialize the backbone with pretrained parameters and set frozen, then learn the IB module to generate instance masks. 2) fix the mask to distill a sparse bag, then fine-tune the WSI head and patch the backbone end-2-end. 3) utilize all fine-tuned instance features within a bag and train the WSI-MIL classifier head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.473, + 0.504 + ], + "angle": 0, + "content": "The \\( p(m_{n}|x_{n}) \\) is first generated by a linear layer from the representation of \\( x_{n} \\) and followed by a sigmoid activation. If the mask \\( m \\) is directly converted by thresholding \\( p(m_{n}|x_{n}) \\), the gradients from \\( z = m\\odot x \\) can not be backpropagated since such an operation is not differentiable. So, during training, \\( m \\) is generated via Monte Carlo sampling, and \\( p(m_n|x_n) \\) can be learned with the reparameterization trick [21, 22] for gradient's estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.516, + 0.308, + 0.533 + ], + "angle": 0, + "content": "3.4. Task-specific Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.473, + 0.812 + ], + "angle": 0, + "content": "To realize task-specific fine-tuning and better performance with only slide-level labels, the workflow of our method is depicted in Figure 3, which includes 3 stages: 1) By minimizing the loss function in section 3.3 with a frozen backbone, a sparse instance set within a bag is obtained. 2) Assisted by the small portion of instances in a bag generated from stage 1, it is possible to fine-tune the backbone for better task-specific features and here we directly backpropagate the loss gradients from slide-level supervision into top-K patch instances. 3) Since the appropriateness of the sparse instance set heavily relies on the performance of the pretrained backbone, the first training IB module is insufficient and may result in small or even zero top-K recall. Moreover, the sparse instance set may lose the contextual dependency of WSI modeling. Thus in this stage, We utilize all fine-tuned instance features within a bag to train a traditional attention-based WSI-MIL classifier." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.21, + 0.846 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we present the performance of the proposed method incorporated with the latest WSI-MIL frameworks, where the IN-1K and SSL features are used for com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.428 + ], + "angle": 0, + "content": "parisons. Ablation experiments are performed to further study the proposed method and for paper length, more experimental results are presented in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.642, + 0.442 + ], + "angle": 0, + "content": "Datasets and Tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.61 + ], + "angle": 0, + "content": "We use five datasets to evaluate our method. The slide-level classification performance of IN-1K, SSL, and our method is evaluated on three datasets including both histopathology and cytopathology images: two public histopathological WSI datasets, Camelyon-16 [5] for tumor / normal binary classification, The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) [34] for tumors subtyping. One internal cytopathology WSI dataset is introduced to validate the universality of our method on both histo- and cyto-pathology, which is Liquid-based Preparation cytology for Cervical Cancer's early screening (LBP-CECA)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Despite the original evaluation on the closed dataset, we also evaluate the generalization of our method on Camelyon-16-C (generated with random synthetic domain shift from Camelyon-16), and Camelyon-17 [29] from five different centers, which occurs frequently in practical pathological diagnosis and has been hindering the application of automatical WSI analysis to the real world. Details of five datasets can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.809 + ], + "angle": 0, + "content": "For pre-processing, we follow the operations in CLAMSB [30] which mainly includes HSV, Blur, Thresholding, and Contours methods to localize the tissue regions in each WSI. Then non-overlapping patches with size \\(256\\times 256\\) on the \\(20\\mathrm{X}\\) magnification are extracted from the tissue regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.825, + 0.706, + 0.84 + ], + "angle": 0, + "content": "Pretraining and Fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our work mainly focuses on the method of fine-tuning. Because good pretrained initialization results in better finetuning performance, we employ mainstream pretraining methods: 1) ImageNet-1k (IN-1K) data pretraining. 2) SSL" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7458" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.091, + 0.887, + 0.432 + ], + "angle": 0, + "content": "
MethodCamelyon-16TCGA-BRCALBP-CECA
F1AUCF1AUCF1AUC
Full Supervision0.967±0.0050.992±0.003--0.741±0.0060.942±0.002
RNN-MIL [7]0.834±0.0170.861±0.0210.776±0.0350.871±0.033--
AB-MIL [19]0.828±0.0130.851±0.0250.771±0.0400.869±0.0370.525±0.0170.845±0.002
DS-MIL [25]0.857±0.0230.892±0.0120.775±0.0440.875±0.041--
CLAM-SB [30]0.839±0.0180.875±0.0280.797±0.0460.879±0.0190.587±0.0140.860±0.005
TransMIL [38]0.846±0.0130.883±0.0090.806±0.0460.889±0.0360.533±0.0060.850±0.007
DTFD-MIL [45]0.882±0.0080.932±0.0160.816±0.0450.895±0.0420.569±0.0260.847±0.003
FT+ CLAM-SB0.911±0.0170.956±0.0130.845±0.0320.935±0.0270.718±0.0100.907±0.005
FT+ TransMIL0.923±0.0120.967±0.0030.848±0.0440.945±0.0200.720±0.0240.918±0.004
FT+ DTFD-MIL0.921±0.0070.962±0.0060.849±0.0270.951±0.0160.723±0.0080.922±0.005
Mean-pooling0.629±0.0290.591±0.0120.818±0.0220.910±0.0320.350±0.0170.735±0.006
Max-pooling0.805±0.0120.824±0.0160.644±0.1790.826±0.0960.636±0.0640.893±0.019
KNN (Mean)0.468±0.0000.506±0.0000.633±0.0660.749±0.0550.393±0.0000.650±0.000
KNN (Max)0.559±0.0000.535±0.0000.524±0.0320.639±0.0630.477±0.0000.743±0.000
FT+ Mean-pooling0.842±0.0060.831±0.0070.866±0.0350.952±0.0180.685±0.0140.900±0.002
FT+ Max-pooling0.927±0.0110.969±0.0040.852±0.0430.948±0.0190.695±0.0130.912±0.004
FT+ KNN (Mean)0.505±0.0000.526±0.0000.784±0.0440.907±0.0340.529±0.0000.737±0.000
FT+ KNN (Max)0.905±0.0000.916±0.0000.802±0.0630.882±0.0360.676±0.0000.875±0.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.445, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Table 1. Slide-Level Classification by using the IN-1K pre-trained backbone or the proposed fine-tuned (FT) in three datasets. Top Rows. Different MIL architectures are compared to select the top 3 SOTA methods to validate the transfer learning performance using the IN-1K pre-trained backbone or the FT. Bottom Rows. The competition of various traditional aggregation and feature evaluation methods by using pre-trained IN-1K or the FT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.597 + ], + "angle": 0, + "content": "pretraining with SimCLR [11], MoCo [17] and DINO [8]. 3) for Camelyon-16 and LBP-CECA, since there are comprehensive tumor area annotations, we use the annotation to pretrain the patch backbone, which acts as the upper-bound of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.793 + ], + "angle": 0, + "content": "Different from prior works backbone frozen after pretraining, we perform backbone fine-tuning with the method proposed in 3.4. In this stage, we fine-tune the backbone and WSI model end-to-end with 25 epochs using the AdamW optimizer, batch size of 1 for WSI with bag size of 512, a learning rate of 1e-5 for backbone and 1e-3 for WSI head. For the BN layers in ResNet, we turn it to eval mode to fix statistics during fine-tuning since we find that the distribution of the top-K instances in a bag is limited to estimate statistics because of the similarity among instances. Shallow layers of backbone are frozen since they only focus on morphological features while deep layer focus on semantics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.3, + 0.817 + ], + "angle": 0, + "content": "4.1. Slide-level Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. For all the experiments, macro-AUC and macro-F1 scores are reported since all the 3 datasets are class imbalanced. For Camelyon-16, the official training set is randomly split into training and validation sets with a ratio of 9:1. The experiment is conducted 5 times and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.643 + ], + "angle": 0, + "content": "the results of the official test set are reported. For TCGA-BRCA, we perform the 10 folder cross-validation with the same running setting adopted in HIPT [9]. Besides the dataset LBP-CECA is randomly split with a ratio of 6:1:3 for training, validation, and testing. The experiment is conducted 5 times. The mean and standard variance values of performance metrics are reported for multi-runs or cross-validation runs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Comparison with baselines. Classification results are summarized in Table 1. We first show full patch supervision results as an upper bound, then we directly evaluate several classic WSI-MIL methods, including RNN-MIL [7], AB-MIL [19], DS-MIL [25], CLAM-SB [30], TransMIL [38], DTFD-MIL [45]. All the WSI-MIL baselines across the three classification tasks suffer from relatively low performance due to the inappropriate backbone features from the pretrained ResNet-50 in IN-1K. Then we apply FT to the backbone to obtain features for WSI classification by three WSI architectures including CLAM-SB (pure global attention, no intersection among instances), TransMIL (self-attention for long sequence), and DTFD-MIL (multi-tier attention paths followed with aggregation, good for patch imbalance)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The results show clearly consistent improvements equipped with FT features under AUC metric. CLAM-SB" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7459" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "achieves a performance relative increase of \\(9.26\\%\\), \\(6.37\\%\\), \\(5.47\\%\\) on Camelyon-16, TCGA-BRCA, and LBP-CECA respectively with vanilla FT based on IN-1K pretraining. In the same conditions, TransMIL and DTFD-MIL achieve new SOTAs on all three datasets and can obtain better relative improvement, especially in LBP-CECA with an \\(8.00\\%\\) and \\(8.85\\%\\) growth compared to CLAM-SB. The results show similar improvements under the F1 metric. For the bottom two rows of Table 1, we compare simple Mean/Max-pooling training and KNN evaluations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.471 + ], + "angle": 0, + "content": "The competitive results demonstrate that our proposed FT method in the backbone of the MIL framework can boost the performance of WSI classification, even for the most simple feature-level Mean/Max-pooling. In addition, task-specific features help us better unearth the properties of different WSI-MIL architectures. Apparently, the more complex architecture, TransMIL, and DTFD-MIL, can not perform desirable performance with the frozen pretrained parameters in LBP-CECE. In contrast, its capabilities can be further enhanced by fine-tuning the backbone for the defined target. Moreover, the effects of different feature aggregation methods vary depending on the task. The better performance is shown by using Max-pooling aggregation in Camelyon-16 containing smaller tumor areas, while Mean-pooling is better in TCGA-BRCA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.4, + 0.494 + ], + "angle": 0, + "content": "4.2. Combination of SSL and Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.468, + 0.531 + ], + "angle": 0, + "content": "In this section, we further improve the slide-level classification performance by combining SSL with FT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.468, + 0.576 + ], + "angle": 0, + "content": "Evaluation Metrics. TCGA-BRCA is used for the evaluation of its task complexity in tumor subtyping with the same setting to 4.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.468, + 0.607 + ], + "angle": 0, + "content": "Comparison with baselines. Since there is no apparent single semantic object in small histopathology patches, we" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.63, + 0.468, + 0.817 + ], + "angle": 0, + "content": "
MethodF1AUC
IN-1K$-0.884±0.059
IN-1K0.797±0.0460.879±0.019
/w FT0.845±0.0320.935±0.027
SimCLR [11]$-0.879±0.069
MoCo [17]0.804±0.0420.904±0.030
/w FT0.851±0.0290.948±0.026
DINO [8]$-0.886±0.059
DINO0.801±0.0450.891±0.043
/w FT0.848±0.0270.944±0.036
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.829, + 0.468, + 0.884 + ], + "angle": 0, + "content": "Table 2. Combination of SSL and Fine-tuning. We compare SSLs with IN-1K and their further improvement via fine-tuning (FT) on TCGA-BRCA. The symbol \\(\\S\\) indicates the result released in previous publication [9, 10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "mainly compare SSLs by Contrastive Learning or augmentation like MoCo [17] and DINO [8], and SimCLR [11]. We show the results of SSLs performed in previous works [9, 10] for fair comparison. All results are performed with the same WSI architecture of CLAM-SB [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Experimental results are summarized in Table 2. Compared to vanilla IN-1K, MoCo and DINO (combined with FT) achieve consistent growth of \\(7.85\\%\\) and \\(7.39\\%\\), respectively. Besides, they have a slight increase of \\(1.39\\%\\) and \\(0.96\\%\\) in comparison to IN-1K with FT. By SSL, the intrinsic task-agnostic features are learned from all patches in WSIs, and after applying our proposed FT for all modules, task-specific features can be distilled from the label and partial data. In such an SSL with an FT paradigm, the data and the label are explored comprehensively to produce a state-of-the-art WSI analysis." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.342, + 0.78, + 0.357 + ], + "angle": 0, + "content": "4.3. Generalization on Domain Shift" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.44 + ], + "angle": 0, + "content": "In this section, we evaluate the generalization of slide-level classification models on domain shift [46] [20], which is crucial for real-world clinical applications due to the diversity in staining, preparation, and imaging devices for pathological image processing among hospitals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Evaluation Metrics. The Camelyon-16-C dataset is generated from Camelyon-16 test set by synthetic domain shift, which adopts a random combination of Brightness, JPEG, and Hue proposed in [46]. For Camelyon-17, since it shares similarities to Camelyon-16 but is collected from five different medical centers, we randomly collect 30 samples from each center to evaluate the robustness of natural domain shift. We directly evaluate models on Camelyon-16-C and 150 extra data from Camelyon-17 with five times of running same as the previous setting in section 4.1. The standard variance values of results is omitted due to the constraint on" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.63, + 0.894, + 0.817 + ], + "angle": 0, + "content": "
MethodCamelyon-16-CCamelyon-17
F1AUCF1AUC
Max-pooling0.6890.7420.5780.670
/w FT0.8160.8920.6870.720
CLAM-SB [30]0.7420.8360.6240.702
/w FT0.8230.8620.6760.725
TransMIL [38]0.7480.8420.6570.706
/w FT0.7950.8570.6840.717
DTFD-MIL [45]0.7750.7990.5760.676
/w FT0.8040.8380.6890.717
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.829, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Table 3. Generalization on Domain Shift. The generalization ability of all methods is compared between fine-tuning(FT) and IN-1K features on two datasets with domain shift. Camelyon-16-C and Camelyon-17 are synthetic and real corruptions respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7460" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "the paper length and a more clear comprehensive comparison can be found in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.197 + ], + "angle": 0, + "content": "Comparison with baselines. The evaluation of generalization results is summarized in Table 3. For WSI model architecture, despite CLAM-SB [30], TransMIL [38], DTFDMIL [45], we also compare Max-pooling on feature level since it shows promising results on Camelyon-16 in 4.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.407 + ], + "angle": 0, + "content": "The results demonstrate consistent improvements after equipping the WSI head with fine-tuned features under the F1 metric: In general, all models can resist domain shifts to a certain extent by utilizing the proposed fine-tuning. Compared with freezing IN-1K parameters, Max-pooling obtains a noticeable performance increase of \\(18.43\\%\\) and \\(18.86\\%\\) on Camelyon-16-C and Camelyon-17, respectively. CLAM-SBachieves an increase of \\(10.91\\%\\), \\(8.33\\%\\) respectively. TransMIL achieves an increase of \\(6.28\\%\\), \\(4.11\\%\\) respectively. DTFD-MIL achieves an increase of \\(2.90\\%\\), \\(11.30\\%\\) respectively. In contrast, the generalization of DTFD-MIL can still be further improved by our proposed FT optimization, even though it is already robust to different domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.62 + ], + "angle": 0, + "content": "Interestingly, all improvements in performance under the AUC metric are much less than F1. The three Attention-based pooling methods respectively achieve the average improvement of \\(6.70\\%\\) and \\(7.91\\%\\) on Camelyon-16-C and Camelyon-17 under the F1 metric, as compared to \\(2.93\\%\\) and \\(2.98\\%\\) under AUC, which indicates that the classification ability of WSI-MIL model (measured by F1) may be much weaker than ranking (measured by AUC). However, doctors may much more care about F1 instead of AUC in clinical diagnosis from sensitivity and specificity perspectives. Most importantly, the classification threshold usually keeps fixed after deploying the model in the real world, which reflects the contribution of our method is more meaningful in practical applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.628, + 0.348, + 0.644 + ], + "angle": 0, + "content": "4.4. Further Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.468, + 0.711 + ], + "angle": 0, + "content": "Additional ablations are included in the Supplementary Materials, with the main focus on the effects of learning rate on the backbone, Number selection of Top-K, and Value selection of Lagrange multiplier, results of the 3 stages." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.37, + 0.735 + ], + "angle": 0, + "content": "4.5. Interpretability and Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.468, + 0.831 + ], + "angle": 0, + "content": "Here, we further show the interpretability improvements with FT features. As shown in Figure 4, attention scores from CLAM-SB [30] were visualized as a heatmap to determine the ROI and interpret the important morphology used for diagnosis. Obviously, the model focuses more on tumors with FT features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.195, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In this work, we present a fine-tuning method for WSI classification under the weak supervision of slide-level la" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.632, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.093, + 0.759, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.093, + 0.887, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.192, + 0.631, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.192, + 0.759, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.192, + 0.887, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.291, + 0.631, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.291, + 0.759, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.291, + 0.887, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.471 + ], + "angle": 0, + "content": "Figure 4. Heatmap comparison between pretraining and FT. The first row shows the full annotations, where red contours denote the tumor area and green contours filter the nonsense background. The second and third row shows the WSI attention map on pretraining and FT features respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.776 + ], + "angle": 0, + "content": "belts. Initially, an effective IB module is introduced to mitigate the training cost of Gigapixel WSI, which distills the oversized bag into a sparse one. Then the backbone of the instance is able to be trained end-to-end in the MIL framework by learning and making classifications on the distilled bag. Thus the WSI classification performance is improved through the retention of mainly task-specific information. In addition, SSL can be combined with the proposed framework for further improvement. Compared with fully supervised learning, our methods can achieve competitive accuracy by utilizing extremely weak WSI labels. Furthermore, our training scheme can introduce versatile training-time augmentations for better generalization on datasets with domain shift, which is an inevitable challenge for previous work. The experimental results reflect the advances of our method in both accuracy and generalization. Overall, Our proposed approach shows strong potential for MIL application in real-world pathology diagnosis, with better performance, faster convergence and annotation efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.79, + 0.66, + 0.807 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.815, + 0.892, + 0.89 + ], + "angle": 0, + "content": "This study was partially supported by the National Natural Science Foundation of China (Grant no. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant no. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7461" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy computation. IEEE transactions on pattern analysis and machine intelligence, 40(12):2897-2905, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Alexander A. Alemi, Ian Fischer, Joshua V. Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.296 + ], + "angle": 0, + "content": "[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10(7):e0130140, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[5] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[6] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature medicine, 28(1):154-163, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[7] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. CoRR, abs/2104.14294, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[9] Richard J. Chen and et al. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In CVPR, pages 16144-16155, June 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[10] Richard J Chen and Rahul G Krishnan. Self-supervised vision transformers learn visual concepts in histopathology. arXiv preprint arXiv:2203.00585, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2018. cite arxiv:1810.04805Comment: 13 pages. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Littjens, Jeroen A. W. M. van der Laak, , and the CAME-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "LYON16 Consortium. Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA, 318(22):2199-2210, 12 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[14] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30(4):681–694, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[15] Ziv Goldfeld, Ewout van den Berg, Kristjan Greenewald, Igor Melnyk, Nam Nguyen, Brian Kingsbury, and Yury Polyanskiy. Estimating information flow in deep neural networks. arXiv preprint arXiv:1810.05728, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.252, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.367, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[19] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 2127-2136. PMLR, 10-15 Jul 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.51, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[20] Syed Ashar Javed, Dinkar Juyal, Zahl Shanis, Shreya Chakraborty, Harsha Pokkalla, and Aaditya Prakash. Rethinking machine learning model evaluation in pathology. arXiv preprint arXiv:2204.05205, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[21] Durk P Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. Advances in neural information processing systems, 28, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.625, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.655, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[23] Erich Leo Lehmann and Henry Scheffé. Completeness, similar regions, and unbiased estimation-part i. In Selected works of EL Lehmann, pages 233-268. Springer, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[24] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[25] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7462" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[29] Geert Litjens, Peter Bandi, Babak Ehteshami Bejnordi, Oscar Geessink, Maschenka Balkenhol, Peter Bult, Altuna Halilovic, Meyke Hermsen, Rob van de Loo, Rob Vogels, et al. 1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset. *GigaScience*, 7(6):giy065, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[30] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[31] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[32] Bhargavi Paranjape, Mandar Joshi, John Thickstun, Hannaneh Hajishirzi, and Luke Zettlemoyer. An information bottleneck approach for controlling conciseness in rationale extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1938-1952, Online, Nov. 2020. Association for Computational Linguistics. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.551 + ], + "angle": 0, + "content": "[33] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[34] Nicholas A Petrick, Shazia Akbar, Kenny HH Cha, Sharon Nofech-Mozes, Berkman Sahiner, Marios A Gavrielides, Jayashree Kalpathy-Cramer, Karen Drukker, Anne LL Martel, et al. Spie-aapm-nci breastpathq challenge: an image analysis challenge for quantitative tumor cellularity assessment in breast cancer histology images following neoadjuvant treatment. Journal of Medical Imaging, 8(3):034501, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[35] Andrew M Saxe, Yamini Bansal, Joel Dapello, Madhu Advani, Artemy Kolchinsky, Brendan D Tracey, and David D Cox. On the information bottleneck theory of deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124020, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[36] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. arXiv preprint arXiv:2001.00396, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[37] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[38] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, and yongbing zhang. Transmil: Transformer based correlated multiple instance learning for whole" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "slide image classification. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2136-2147. Curran Associates, Inc., 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[39] Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[40] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[42] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[43] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[44] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. Xlnet: Generalized autoregressive pretraining for language understanding. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[45] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E. Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. ArXiv, abs/2203.12081, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.656 + ], + "angle": 0, + "content": "[46] Yunlong Zhang, Yuxuan Sun, Honglin Li, Sunyi Zheng, Chenglu Zhu, and Lin Yang. Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[47] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdolahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.782 + ], + "angle": 0, + "content": "[48] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9140-9150, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.785, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7463" + } + ] +] \ No newline at end of file diff --git a/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_origin.pdf b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e95dde67cd2cfdf24f972f8a6f6bc94e32df671d --- /dev/null +++ b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/1b72dc26-76ea-4a15-b4ba-65955df2cecc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec2c7171af1f0be09cd6e56de2618aa66f0ae1e26c2ac8f3f1f6d1caa41a85bf +size 2244097 diff --git a/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/full.md b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d452590f0d3a658eb4c155ae49ac8942f1852bd --- /dev/null +++ b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/full.md @@ -0,0 +1,329 @@ +# Task-specific Fine-tuning via Variational Information Bottleneck for Weakly-supervised Pathology Whole Slide Image Classification + +Honglin Li $^{1,3}$ , Chenglu Zhu $^{2,3}$ , Yunlong Zhang $^{1,3}$ , Yuxuan Sun $^{1,3}$ + +Zhongyi Shui $^{1,3}$ , Wenwei Kuang $^{3,4}$ , Sunyi Zheng $^{2,3}$ , Lin Yang $^{2,3*}$ + +$^{1}$ College of Computer Science and Technology, Zhejiang University + +$^{2}$ Research Center for Industries of the Future and $^{3}$ School of Engineering, Westlake University + +4The University of Hong Kong + +{lihonglin, yanglin}@westlake.edu.cn + +# Abstract + +While Multiple Instance Learning (MIL) has shown promising results in digital Pathology Whole Slide Image (WSI) analysis, such a paradigm still faces performance and generalization problems due to high computational costs and limited supervision of Gigapixel WSIs. To deal with the computation problem, previous methods utilize a frozen model pretrained from ImageNet to obtain representations, however, it may lose key information owing to the large domain gap and hinder the generalization ability without image-level training-time augmentation. Though Self-supervised Learning (SSL) proposes viable representation learning schemes, the downstream task-specific features via partial label tuning are not explored. To alleviate this problem, we propose an efficient WSI fine-tuning framework motivated by the Information Bottleneck theory. The theory enables the framework to find the minimal sufficient statistics of WSI, thus supporting us to fine-tune the backbone into a task-specific representation only depending on WSI-level weak labels. The WSI-MIL problem is further analyzed to theoretically deduce our fine-tuning method. We evaluate the method on five pathological WSI datasets on various WSI heads. The experimental results show significant improvements in both accuracy and generalization compared with previous works. Source code will be available at https://github.com/invoker-LL/WSI-finetuning. + +# 1. Introduction + +Digital Pathology or microscopic images have been widely used for the diagnosis of cancers such as Breast Cancer [13] and Prostate Cancer [6]. However, the reading of Whole Slide Images (WSIs) with gigapixel resolution is + +![](images/e6924865837e9beb58e361df843e2650f9342cb52ad2accdf50e53de468bb4a9.jpg) +a. ImageNet-1k v-score: 0.285 + +![](images/b16c017a48df0351af1861e5a76d45e2840b8169638d446149845fb224d0ab5d.jpg) +b. Full Supervision v-score: 0.745 + +![](images/8dd7bdc4b45348a884325912f55e7268cc1c8e2c01c4ddad7575050ca58800dc.jpg) +c. Self-supervised v-score: 0.394 + +![](images/34c05d368cfa1dafa6cfdbce0d49836e0029f6ed549d71c4eaf2271c900ecb11.jpg) +d. IB Fine-tuning (ours) +v-score: 0.538 +Figure 1. T-SNE visualization of different representations on patches. Our method converts chaotic ImageNet-1K and SSL features into a more task-specific and separable distribution. The cluster evaluation measurement, v-scores, show weakly supervised fine-tuned features are more close to full supervision compared to others. a. ImageNet-1k pretraining. b. Full patch supervision. c. Self-supervised Learning. d. Fine-tuning with WSI labels. + +time-consuming which poses an urgent need for automatic computer-assisted diagnosis. Though computers can boost the speed of the diagnosis process, the enormous size of resolution, over 100M [45], makes it infeasible to acquire precise and exhaustive annotations for model training, let alone the current hardware can hardly support the parallel training on all patches of a WSI. Hence, an annotation-efficient learning scheme with light computation is increasingly desirable to cope with those problems. In pathology WSI analysis, the heavy annotation cost is usually alleviated by Multiple Instance Learning (MIL) with only WSI-level weak supervision, which makes a comprehensive decision on a series of instances as a bag sample [19, 31]. Intuitively, all + +small patches in the WSI are regarded as instances to constitute a bag sample [7, 30, 38], where the WSI's category corresponds to the max lesion level of all patch instances. + +However, most methods pay much effort to design WSI architectures while overlooking the instance-level representation ability. Because of the computational limitation, the gradient at the WSI-level is impossible to parallelly backpropagate to instance encoders with more than 10k instances of a bag. Thus parameters of the pretrained backbone from ImageNet-1k (IN-1K) are frozen to obtain invariant embeddings. Due to the large domain gap between IN-1K and pathological images, some essential information may be discarded by layers of frozen convolutional filters, which constrains the accuracy of previous WSI methods. To address the constraint, recent works [9, 25] make efforts to learn a good feature representation at the patchlevel by leveraging Self-supervised Learning (SSL). However, such task-agnostic features are dominated by the proxy objective of SSL, e.g. Contrastive Learning in [8, 11, 17] may push away the distance between two instances within the same category, thus only performs slightly better than IN-1K pretraining in WSI classification. Nearly all SSL methods [8, 11, 16, 17] proposed on natural image recognition utilize a small portion of annotations to get promising fine-tuning accuracy compared to full supervision, which is higher than Linear Probing [17] by a large margin. + +These findings illuminate us to design a fine-tuning scheme for WSI analysis to convert IN-K or SSL task-agnostic representations into task-specifics. Motivated by the Information Bottleneck (IB) theory [1, 2], we argue that pretraining is limited to downstream tasks, therefore fine-tuning is necessary for WSI analysis. In addition, we develop a solution based on Variational IB to tackle the dilemma of fine-tuning and computational limitation by its minimal sufficient statistics and attribution properties [1, 23]. The differences among the above three feature representations are depicted in Figure 1, where the feature representation under full patch-level supervision is considered as the upper bound. + +Our main contributions are in 3 folds: 1) We propose a simple agent task of WSI-MIL by introducing an IB module that distills over $10\mathrm{k}$ redundant instances within a bag into less than $1\mathrm{k}$ of the most supported instances. Thus the parallel computation cost of gradient-based training on Gigapixel Images is over ten times relieved. By learning and making classification on the simplified bag, we find that there are trivial information losses due to the low-rank property of pathological WSI, and the distilled bag makes it possible to train a WSI-MIL model with the feature extractor on patches end-to-end, thus boosting the final performance. 2) We argue that the performance can be further improved by combining with the SSL pretraining since we could convert the task-agnostic representation from SSL into task-specific + +one by well-designed fine-tuning. The proposed framework only relies on annotations at WSI levels, which is similar to recent SSL approaches [8, 11, 16, 17]. Note that our method only utilizes less than a $1\%$ fraction of full patch annotation to achieve competitive accuracy compared to counterparts. 3) Versatile training-time augmentations can be incorporated with our proposed fine-tuning scheme, thus resulting in better generalization in various real-world or simulated datasets with domain shift, which previous works ignore to validate. These empirical results show that our method advances accuracy and generalization simultaneously, and thus would be more practical for real-world applications. + +# 2. Related Work + +# 2.1. Multiple Instance Learning for WSI Analysis + +Multiple Instance Learning (MIL) is a well-defined task and has been explored extensively. Currently, there are mainly two lines of methods for WSI analysis: 1) explicit modeling the MIL definition $Y = \text{pooling}\{y_1, y_2, \dots, y_n\}$ that WSI level prediction is aggregated by the probability of all patches with Mean or Max-pooling [7,45]. 2) implicit learning WSI level representation by aggregating all patches embeddings via a WSI classifier with Recurrent Neural Network (RNN) [7] or in an attention [19] mechanism. The latter shows superior performance since such modeling includes less inductive bias compared to the former processing with fixed weights. + +RNN [7] treats WSI as a sequence of patches, but its permutation variation and defects on long-term dependency do not match the nature of WSI well. By contrast, attention-based MIL (AB-MIL) [19] learns the weights of instance representations adaptively, and these methods can be flexibly inserted into the classification framework as a plug-and-play module to generalize the full-field map according to the task. To make patches' representation more discriminative, CLAM [30] introduces an auxiliary task in the MIL framework to distinguish its corresponding availability according to the size of instance attention while training the WSI classification. Trans-MIL [38] is designed to model the relation among patches via Self-Attention [41] and solves the softmax computational complexity in the long sequence by linear Self-Attention [42, 43]. To embed the multi-scale information of WSI, DS-MIL [25] simply concatenates patch features in three scales, and HIPT [9] builds hierarchical multi-scale features. To mitigate overfitting in complex attention model trained on the limited number of WSIs, DTFD-MIL [45] resamples instances from the original bag to generate various sub-bags for augmentation. However, all these works pay too much attention on WSI head to explore the backbone. + +![](images/fac218ab8ba49211d7bcd248afda7952532317d0a5308b6a7335c3d2bfc25d5a.jpg) +Figure 2. An illustration of the tuning scheme. To maximize $R_{IB}$ , enlarging the first term helps latent representation vector $Z$ become more predictive for label $Y$ (More overlap between $Y$ and $Z2$ compared to $Z1$ , which denotes higher mutual information), while by reducing the second term to filter irrelevant features (smaller overlap between image $X$ and $Z2$ compared to $Z1$ ). + +# 2.2. Self-supervised Learning and Fine-tuning + +Self-supervised Learning (SSL) has shown to be a promising paradigm both in computer vision [4,8,11,16,17] and natural language processing [12,33,44]. However, fine-tuning (FT) is quite necessary for the pretrained model of SSL to the downstream tasks, e.g. in [16,17], the vanilla FT method can improve about 15 percent of accuracy in IN-1K compared to vanilla Linear Probing [17]. More interesting and parameter-efficient FT methods [14,18,24,26,27,47,48] are proposed according to the similarity between pretraining and downstream in natural language processing and few-shot image generation. Unfortunately, the computational limitation on redundant instances prevents FT from being directly introduced into WSI analysis. Although SSL achieves optimistic performance in some works [9, 25] for pathology WSI analysis but only shows a small margin of improvement. There may be much potential to be discovered in WSI analysis if an appropriate FT method can be used to address the training cost from surplus information of WSI. + +# 2.3. Information Bottleneck and Attributions + +The Information Bottleneck (IB) conception is introduced in [40] as an information-theoretic framework for learning, which is currently used in deep learning both theoretically and practically. [39] proposes to unveil deep neural network in an information flow perspective by estimating Mutual Information (MI) between the outputs of two layers, then in followed-up works [15, 35], they revisit the MI compression process of IB to propose complements and better measurements of MI in Deep Neural Network (DNN). + +IB not only reveals how the traditional DNN training finishes information compression implicitly in the above works, but can also be employed explicitly as an objective to intervene in DNN's training: [2] derives a variational + +bound of the IB objective, claiming that the proxy objective provides an additional regularization term which improves the robustness to adversarial samples compared to the traditional training of DNN. [32] adopts the variational bound in [2] and replaces the distribution of latent features from Gaussian into Bernoulli. In this way, it generates masks to control the conciseness of the rationale extraction of sentences. Achille et al. [1] claim that IB solutions approximate minimal sufficient statistics (MSS) [23], and by penalizing the redundancy of representations, they heuristically argue that the model's sensitivity to nuisances is mitigated. Similar to some visual attribution methods like CAM [37, 49], LRP [3] and patch masking [4, 16, 28] in ViT, an IB-based attribution method is proposed in [36] by adding noise to intermediate feature maps, restricting the flow of information, then how much information image regions provide can be quantified. Intuitively, the IB objective reveals the limitation of pretraining for downstream tasks and the imperative of FT for task-specific representation in WSI analysis, as shown in Figure 2. Moreover, it motivates us to solve the dilemma of FT and computational limitation by its' MSS and visual attribution properties. + +# 3. Method + +# 3.1. Overview on MIL-based WSI Analysis + +Given a WSI $X$ , the goal is to make slide-level prediction $Y$ by learning a classifier $f(X; \theta)$ . Due to its extremely high resolution, $X$ is patched into a huge bag of small instances $X = \{x_1, \dots, x_N\}$ , where $N$ is the number of instance. The slide-level supervision $\hat{Y}$ is given by a Max-pooling operation of the latent label $\hat{y}_i$ for each instance $x_i$ , which can be defined as: + +$$ +\hat {Y} = \max \left\{\hat {y} _ {1}, \dots , \hat {y} _ {N} \right\}. \tag {1} +$$ + +Since all latent labels of instances $\hat{y}_i$ are unknown under the WSI-level supervision, conventional approaches con + +vert this problem into a MIL formulation in the following two steps: 1) Processing images into feature representations $Z = \{z_{1},\dots,z_{N}\}$ with a backbone $h$ as $z_{i} = h(x_{i};\theta_{1})$ where $h$ is a model of any architecture such as CNN or ViT with parameters $\theta_{1}$ . 2) Aggregating all patches' features within a slide and producing the slide-level prediction $Y = g(Z;\theta_2)$ , where $g$ is an attention-based pooling function followed by a linear classifier head as: + +$$ +g (Z; \theta_ {2}) = \sigma \left(\sum_ {i = 1} ^ {N} a _ {i} z _ {i}\right), \tag {2} +$$ + +where $a_{i}$ is attention weights and $\sigma (\cdot)$ is a linear head. Limited by the computational cost, the parameters $\theta_{1}$ and $\theta_{2}$ in $f(X;\theta) = g\{h(X;\theta_1);\theta_2\}$ are learned separately by following steps: 1) Initializing $\theta_{1}$ from the pretrained model, which refers to general features from public IN-1K, or learned by SSL on the related dataset to extract the domain-specific representations. 2) Freezing $\theta_{1}$ and learning $\theta_{2}$ under slide-level supervision. + +# 3.2. Information Bottleneck for MIL Sparsity + +# Background of the Information Bottleneck + +The Information Bottleneck (IB) can work as an information compression role to intervene in DNN's training [2]. The objective function of IB to be maximized is given in [40] as, + +$$ +R _ {I B} = I (Z, Y) - \beta I (Z, X), \tag {3} +$$ + +where $I(\cdot, \cdot)$ indicates the Mutual Information (MI) and $\beta$ is a Lagrange multiplier controlling the trade-off between the information that the representation variable $Z$ shares with the label $Y$ and its shares with input $X$ . Since the computation of MI is intractable during the training of the neural networks, to maximize IB objective can be transferred to minimize a variational bound of Eq.(3) derived in [2] follows: + +$$ +J _ {I B} = \frac {1}{N} \sum_ {n = 1} ^ {N} \mathbb {E} _ {z \sim p _ {\theta} (z | x _ {n})} [ - \log q _ {\phi} (y _ {n} | z) ] + \tag {4} +$$ + +$$ +\beta K L [ p _ {\theta} (z | x _ {n}), r (z) ], +$$ + +where $N$ denotes the number of samples, $q_{\phi}(y|z)$ is a parametric approximation to the likelihood $p(y|z), r(z)$ is the prior probability of $z$ to variational approximate the marginal $p(z)$ , and $p_{\theta}(z|x)$ is the parametric posterior distribution over $z$ . + +# Learn MIL Sparsity via Variational Bound of IB + +To trade off the dilemma of computational limitation and task-specific representation learning via end-to-end backpropagation, we propose to utilize the IB module to filter most task-irrelevant instances for task-specific fine-tuning. + +The above filtering process can be implemented by optimizing the second term of in Eq.(3) which controls the + +compression. There are two ways that compress $X$ to $Z$ by decreasing the KL divergence between $p(z|x)$ and $r(z)$ in Eq.(4) variational method: reducing the dimension of representation $Z$ compared to $X$ in [2], or converting input $X$ into a sparse one in [32]. + +For the setting of our long instance sequenced MIL, we reduce $I(X,Z)$ into a degree so that the gradients can be back-propagated to the backbone encoder, which needs us to convert a WSI of bag size over 10k into 1k for the sake of sparsity. Considering MIL for tumor v.s. normal binary classification without loss of generality and the latent label $y_{i}$ of each instance $x_{i}$ in Eq.(1), we argue that it is sufficient enough to make the WSI level prediction if one tumor area is detected. With the above understanding, we propose to learn compressed components similar to [32] by defining a IB module as: + +$$ +z = m \odot x, \tag {5} +$$ + +where $m$ is a Bernoulli $(\pi)$ distributed binary mask and in this way $KL[p_{\theta}(z|x),r(z)]$ in Eq.(4) can be decomposed as, + +$$ +K L \left[ p _ {\theta} \left(m _ {i} | x\right), r \left(m _ {i}\right) \right] + \pi H (X), \tag {6} +$$ + +where $H(X)$ is the entropy of $X$ , which can be omitted during the minimization due to its constant value. Please check Supplementary for above proof. + +The Bernoulli $(\pi)$ distribution for $m$ fits the definition of MIL empirically: we can treat $m$ as a latent weak prediction $\hat{y}$ describing whether the patch contains tumor or not, denoting $P_{set} = \{p(m_1|x_1),\dots,p(m_N|x_N)\}$ , then during inference Eq.(1) can be derived as: + +$$ +\hat {Y} = \max \left\{P _ {s e t} \right\} = \max \left\{P _ {s u b s e t} \right\}, \tag {7} +$$ + +where $P_{\text{subset}} \in P_{\text{set}}$ , generated by select top-K elements in $P_{\text{set}}$ . The patch classifier trained with only slide-level supervision shows low accuracy [25], so we only use it to generate mask for sparse sub-bag and still utilize attention-based MIL on the sub-bag for decision making. + +# 3.3. Loss Function and its Implementation + +Derived from Eq.(4) and Eq.(6), our Variational IB module should be optimized with the following loss function, + +$$ +\begin{array}{l} \operatorname {l o s s} = \frac {1}{N} \sum_ {n = 1} ^ {N} \mathbb {E} _ {z \sim p _ {\theta} (z | x _ {n})} [ - \log q _ {\phi} (y _ {n} | z) ] + \tag {8} \\ \beta K L [ p _ {\theta} (m | x _ {n}), r (m) ], \\ \end{array} +$$ + +where the first term is the task loss to learn task-specific features, and it can be treated as the cross entropy same to prior works [2, 32] in a sampling perspective. The second term is the information loss to filter out task-irrelevant instances by minimizing the Kullback-Leibler divergence between the distribution of mask and the prior Bernoulli. + +![](images/c54e0a4389595a0c7e8e568f600bc565c1e36bc3fd910646357eda573e0de30a.jpg) +Figure 3. Workflow of WSI-MIL task-specific fine-tuning. 1) Initialize the backbone with pretrained parameters and set frozen, then learn the IB module to generate instance masks. 2) fix the mask to distill a sparse bag, then fine-tune the WSI head and patch the backbone end-2-end. 3) utilize all fine-tuned instance features within a bag and train the WSI-MIL classifier head. + +The $p(m_{n}|x_{n})$ is first generated by a linear layer from the representation of $x_{n}$ and followed by a sigmoid activation. If the mask $m$ is directly converted by thresholding $p(m_{n}|x_{n})$ , the gradients from $z = m\odot x$ can not be backpropagated since such an operation is not differentiable. So, during training, $m$ is generated via Monte Carlo sampling, and $p(m_n|x_n)$ can be learned with the reparameterization trick [21, 22] for gradient's estimation. + +# 3.4. Task-specific Fine-tuning + +To realize task-specific fine-tuning and better performance with only slide-level labels, the workflow of our method is depicted in Figure 3, which includes 3 stages: 1) By minimizing the loss function in section 3.3 with a frozen backbone, a sparse instance set within a bag is obtained. 2) Assisted by the small portion of instances in a bag generated from stage 1, it is possible to fine-tune the backbone for better task-specific features and here we directly backpropagate the loss gradients from slide-level supervision into top-K patch instances. 3) Since the appropriateness of the sparse instance set heavily relies on the performance of the pretrained backbone, the first training IB module is insufficient and may result in small or even zero top-K recall. Moreover, the sparse instance set may lose the contextual dependency of WSI modeling. Thus in this stage, We utilize all fine-tuned instance features within a bag to train a traditional attention-based WSI-MIL classifier. + +# 4. Experiments + +In this section, we present the performance of the proposed method incorporated with the latest WSI-MIL frameworks, where the IN-1K and SSL features are used for com + +parisons. Ablation experiments are performed to further study the proposed method and for paper length, more experimental results are presented in the Supplementary. + +# Datasets and Tasks. + +We use five datasets to evaluate our method. The slide-level classification performance of IN-1K, SSL, and our method is evaluated on three datasets including both histopathology and cytopathology images: two public histopathological WSI datasets, Camelyon-16 [5] for tumor / normal binary classification, The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) [34] for tumors subtyping. One internal cytopathology WSI dataset is introduced to validate the universality of our method on both histo- and cyto-pathology, which is Liquid-based Preparation cytology for Cervical Cancer's early screening (LBP-CECA). + +Despite the original evaluation on the closed dataset, we also evaluate the generalization of our method on Camelyon-16-C (generated with random synthetic domain shift from Camelyon-16), and Camelyon-17 [29] from five different centers, which occurs frequently in practical pathological diagnosis and has been hindering the application of automatical WSI analysis to the real world. Details of five datasets can be found in the supplementary. + +For pre-processing, we follow the operations in CLAMSB [30] which mainly includes HSV, Blur, Thresholding, and Contours methods to localize the tissue regions in each WSI. Then non-overlapping patches with size $256\times 256$ on the $20\mathrm{X}$ magnification are extracted from the tissue regions. + +# Pretraining and Fine-tuning. + +Our work mainly focuses on the method of fine-tuning. Because good pretrained initialization results in better finetuning performance, we employ mainstream pretraining methods: 1) ImageNet-1k (IN-1K) data pretraining. 2) SSL + +
MethodCamelyon-16TCGA-BRCALBP-CECA
F1AUCF1AUCF1AUC
Full Supervision0.967±0.0050.992±0.003--0.741±0.0060.942±0.002
RNN-MIL [7]0.834±0.0170.861±0.0210.776±0.0350.871±0.033--
AB-MIL [19]0.828±0.0130.851±0.0250.771±0.0400.869±0.0370.525±0.0170.845±0.002
DS-MIL [25]0.857±0.0230.892±0.0120.775±0.0440.875±0.041--
CLAM-SB [30]0.839±0.0180.875±0.0280.797±0.0460.879±0.0190.587±0.0140.860±0.005
TransMIL [38]0.846±0.0130.883±0.0090.806±0.0460.889±0.0360.533±0.0060.850±0.007
DTFD-MIL [45]0.882±0.0080.932±0.0160.816±0.0450.895±0.0420.569±0.0260.847±0.003
FT+ CLAM-SB0.911±0.0170.956±0.0130.845±0.0320.935±0.0270.718±0.0100.907±0.005
FT+ TransMIL0.923±0.0120.967±0.0030.848±0.0440.945±0.0200.720±0.0240.918±0.004
FT+ DTFD-MIL0.921±0.0070.962±0.0060.849±0.0270.951±0.0160.723±0.0080.922±0.005
Mean-pooling0.629±0.0290.591±0.0120.818±0.0220.910±0.0320.350±0.0170.735±0.006
Max-pooling0.805±0.0120.824±0.0160.644±0.1790.826±0.0960.636±0.0640.893±0.019
KNN (Mean)0.468±0.0000.506±0.0000.633±0.0660.749±0.0550.393±0.0000.650±0.000
KNN (Max)0.559±0.0000.535±0.0000.524±0.0320.639±0.0630.477±0.0000.743±0.000
FT+ Mean-pooling0.842±0.0060.831±0.0070.866±0.0350.952±0.0180.685±0.0140.900±0.002
FT+ Max-pooling0.927±0.0110.969±0.0040.852±0.0430.948±0.0190.695±0.0130.912±0.004
FT+ KNN (Mean)0.505±0.0000.526±0.0000.784±0.0440.907±0.0340.529±0.0000.737±0.000
FT+ KNN (Max)0.905±0.0000.916±0.0000.802±0.0630.882±0.0360.676±0.0000.875±0.000
+ +Table 1. Slide-Level Classification by using the IN-1K pre-trained backbone or the proposed fine-tuned (FT) in three datasets. Top Rows. Different MIL architectures are compared to select the top 3 SOTA methods to validate the transfer learning performance using the IN-1K pre-trained backbone or the FT. Bottom Rows. The competition of various traditional aggregation and feature evaluation methods by using pre-trained IN-1K or the FT. + +pretraining with SimCLR [11], MoCo [17] and DINO [8]. 3) for Camelyon-16 and LBP-CECA, since there are comprehensive tumor area annotations, we use the annotation to pretrain the patch backbone, which acts as the upper-bound of our method. + +Different from prior works backbone frozen after pretraining, we perform backbone fine-tuning with the method proposed in 3.4. In this stage, we fine-tune the backbone and WSI model end-to-end with 25 epochs using the AdamW optimizer, batch size of 1 for WSI with bag size of 512, a learning rate of 1e-5 for backbone and 1e-3 for WSI head. For the BN layers in ResNet, we turn it to eval mode to fix statistics during fine-tuning since we find that the distribution of the top-K instances in a bag is limited to estimate statistics because of the similarity among instances. Shallow layers of backbone are frozen since they only focus on morphological features while deep layer focus on semantics. + +# 4.1. Slide-level Classification + +Evaluation Metrics. For all the experiments, macro-AUC and macro-F1 scores are reported since all the 3 datasets are class imbalanced. For Camelyon-16, the official training set is randomly split into training and validation sets with a ratio of 9:1. The experiment is conducted 5 times and + +the results of the official test set are reported. For TCGA-BRCA, we perform the 10 folder cross-validation with the same running setting adopted in HIPT [9]. Besides the dataset LBP-CECA is randomly split with a ratio of 6:1:3 for training, validation, and testing. The experiment is conducted 5 times. The mean and standard variance values of performance metrics are reported for multi-runs or cross-validation runs. + +Comparison with baselines. Classification results are summarized in Table 1. We first show full patch supervision results as an upper bound, then we directly evaluate several classic WSI-MIL methods, including RNN-MIL [7], AB-MIL [19], DS-MIL [25], CLAM-SB [30], TransMIL [38], DTFD-MIL [45]. All the WSI-MIL baselines across the three classification tasks suffer from relatively low performance due to the inappropriate backbone features from the pretrained ResNet-50 in IN-1K. Then we apply FT to the backbone to obtain features for WSI classification by three WSI architectures including CLAM-SB (pure global attention, no intersection among instances), TransMIL (self-attention for long sequence), and DTFD-MIL (multi-tier attention paths followed with aggregation, good for patch imbalance). + +The results show clearly consistent improvements equipped with FT features under AUC metric. CLAM-SB + +achieves a performance relative increase of $9.26\%$ , $6.37\%$ , $5.47\%$ on Camelyon-16, TCGA-BRCA, and LBP-CECA respectively with vanilla FT based on IN-1K pretraining. In the same conditions, TransMIL and DTFD-MIL achieve new SOTAs on all three datasets and can obtain better relative improvement, especially in LBP-CECA with an $8.00\%$ and $8.85\%$ growth compared to CLAM-SB. The results show similar improvements under the F1 metric. For the bottom two rows of Table 1, we compare simple Mean/Max-pooling training and KNN evaluations. + +The competitive results demonstrate that our proposed FT method in the backbone of the MIL framework can boost the performance of WSI classification, even for the most simple feature-level Mean/Max-pooling. In addition, task-specific features help us better unearth the properties of different WSI-MIL architectures. Apparently, the more complex architecture, TransMIL, and DTFD-MIL, can not perform desirable performance with the frozen pretrained parameters in LBP-CECE. In contrast, its capabilities can be further enhanced by fine-tuning the backbone for the defined target. Moreover, the effects of different feature aggregation methods vary depending on the task. The better performance is shown by using Max-pooling aggregation in Camelyon-16 containing smaller tumor areas, while Mean-pooling is better in TCGA-BRCA. + +# 4.2. Combination of SSL and Fine-tuning + +In this section, we further improve the slide-level classification performance by combining SSL with FT. + +Evaluation Metrics. TCGA-BRCA is used for the evaluation of its task complexity in tumor subtyping with the same setting to 4.1. + +Comparison with baselines. Since there is no apparent single semantic object in small histopathology patches, we + +
MethodF1AUC
IN-1K$-0.884±0.059
IN-1K0.797±0.0460.879±0.019
/w FT0.845±0.0320.935±0.027
SimCLR [11]$-0.879±0.069
MoCo [17]0.804±0.0420.904±0.030
/w FT0.851±0.0290.948±0.026
DINO [8]$-0.886±0.059
DINO0.801±0.0450.891±0.043
/w FT0.848±0.0270.944±0.036
+ +mainly compare SSLs by Contrastive Learning or augmentation like MoCo [17] and DINO [8], and SimCLR [11]. We show the results of SSLs performed in previous works [9, 10] for fair comparison. All results are performed with the same WSI architecture of CLAM-SB [30]. + +Experimental results are summarized in Table 2. Compared to vanilla IN-1K, MoCo and DINO (combined with FT) achieve consistent growth of $7.85\%$ and $7.39\%$ , respectively. Besides, they have a slight increase of $1.39\%$ and $0.96\%$ in comparison to IN-1K with FT. By SSL, the intrinsic task-agnostic features are learned from all patches in WSIs, and after applying our proposed FT for all modules, task-specific features can be distilled from the label and partial data. In such an SSL with an FT paradigm, the data and the label are explored comprehensively to produce a state-of-the-art WSI analysis. + +# 4.3. Generalization on Domain Shift + +In this section, we evaluate the generalization of slide-level classification models on domain shift [46] [20], which is crucial for real-world clinical applications due to the diversity in staining, preparation, and imaging devices for pathological image processing among hospitals. + +Evaluation Metrics. The Camelyon-16-C dataset is generated from Camelyon-16 test set by synthetic domain shift, which adopts a random combination of Brightness, JPEG, and Hue proposed in [46]. For Camelyon-17, since it shares similarities to Camelyon-16 but is collected from five different medical centers, we randomly collect 30 samples from each center to evaluate the robustness of natural domain shift. We directly evaluate models on Camelyon-16-C and 150 extra data from Camelyon-17 with five times of running same as the previous setting in section 4.1. The standard variance values of results is omitted due to the constraint on + +Table 2. Combination of SSL and Fine-tuning. We compare SSLs with IN-1K and their further improvement via fine-tuning (FT) on TCGA-BRCA. The symbol $\S$ indicates the result released in previous publication [9, 10]. + +
MethodCamelyon-16-CCamelyon-17
F1AUCF1AUC
Max-pooling0.6890.7420.5780.670
/w FT0.8160.8920.6870.720
CLAM-SB [30]0.7420.8360.6240.702
/w FT0.8230.8620.6760.725
TransMIL [38]0.7480.8420.6570.706
/w FT0.7950.8570.6840.717
DTFD-MIL [45]0.7750.7990.5760.676
/w FT0.8040.8380.6890.717
+ +Table 3. Generalization on Domain Shift. The generalization ability of all methods is compared between fine-tuning(FT) and IN-1K features on two datasets with domain shift. Camelyon-16-C and Camelyon-17 are synthetic and real corruptions respectively. + +the paper length and a more clear comprehensive comparison can be found in Supplementary. + +Comparison with baselines. The evaluation of generalization results is summarized in Table 3. For WSI model architecture, despite CLAM-SB [30], TransMIL [38], DTFDMIL [45], we also compare Max-pooling on feature level since it shows promising results on Camelyon-16 in 4.1. + +The results demonstrate consistent improvements after equipping the WSI head with fine-tuned features under the F1 metric: In general, all models can resist domain shifts to a certain extent by utilizing the proposed fine-tuning. Compared with freezing IN-1K parameters, Max-pooling obtains a noticeable performance increase of $18.43\%$ and $18.86\%$ on Camelyon-16-C and Camelyon-17, respectively. CLAM-SBachieves an increase of $10.91\%$ , $8.33\%$ respectively. TransMIL achieves an increase of $6.28\%$ , $4.11\%$ respectively. DTFD-MIL achieves an increase of $2.90\%$ , $11.30\%$ respectively. In contrast, the generalization of DTFD-MIL can still be further improved by our proposed FT optimization, even though it is already robust to different domains. + +Interestingly, all improvements in performance under the AUC metric are much less than F1. The three Attention-based pooling methods respectively achieve the average improvement of $6.70\%$ and $7.91\%$ on Camelyon-16-C and Camelyon-17 under the F1 metric, as compared to $2.93\%$ and $2.98\%$ under AUC, which indicates that the classification ability of WSI-MIL model (measured by F1) may be much weaker than ranking (measured by AUC). However, doctors may much more care about F1 instead of AUC in clinical diagnosis from sensitivity and specificity perspectives. Most importantly, the classification threshold usually keeps fixed after deploying the model in the real world, which reflects the contribution of our method is more meaningful in practical applications. + +# 4.4. Further Ablation Experiments + +Additional ablations are included in the Supplementary Materials, with the main focus on the effects of learning rate on the backbone, Number selection of Top-K, and Value selection of Lagrange multiplier, results of the 3 stages. + +# 4.5. Interpretability and Visualization + +Here, we further show the interpretability improvements with FT features. As shown in Figure 4, attention scores from CLAM-SB [30] were visualized as a heatmap to determine the ROI and interpret the important morphology used for diagnosis. Obviously, the model focuses more on tumors with FT features. + +# 5. Conclusion + +In this work, we present a fine-tuning method for WSI classification under the weak supervision of slide-level la + +![](images/b211ae4ea0c3503d27702e6b6d7b236f40c72927c643e08ee51d54ff9c54a673.jpg) + +![](images/99b9450acde0b21e7790cf12353b56f38823a2511e9a8004c2a1ea5021447cd1.jpg) + +![](images/20de2c06e6cfb7ba2f8d76ba52a51339dc310b9caaac6fc28a7a48a77b2372ba.jpg) + +![](images/cc20e19c7581bb63c363320446a19eb1812a80c084ac873642276f76a7279f77.jpg) + +![](images/3ab4e29709964280adfead2dd5a0b2c9bfe7b591a9d60f845e66df2aad708758.jpg) + +![](images/25dcc4119d2965536db446e2d47396faba673305d449e36531fc62b8ea664020.jpg) + +![](images/221afd1a32da494642922b79bce36520da32aaa1d646130939b95b75cbac5f20.jpg) +Figure 4. Heatmap comparison between pretraining and FT. The first row shows the full annotations, where red contours denote the tumor area and green contours filter the nonsense background. The second and third row shows the WSI attention map on pretraining and FT features respectively. + +![](images/abb2288ea1e58537763e53aea0130eba9751189ec26d37fc4a9e6af666063e28.jpg) + +![](images/17a386dc455de9ced873f15ee0330f12b3411f430ecb56ec3dac28b1cfb74e8c.jpg) + +belts. Initially, an effective IB module is introduced to mitigate the training cost of Gigapixel WSI, which distills the oversized bag into a sparse one. Then the backbone of the instance is able to be trained end-to-end in the MIL framework by learning and making classifications on the distilled bag. Thus the WSI classification performance is improved through the retention of mainly task-specific information. In addition, SSL can be combined with the proposed framework for further improvement. Compared with fully supervised learning, our methods can achieve competitive accuracy by utilizing extremely weak WSI labels. Furthermore, our training scheme can introduce versatile training-time augmentations for better generalization on datasets with domain shift, which is an inevitable challenge for previous work. The experimental results reflect the advances of our method in both accuracy and generalization. Overall, Our proposed approach shows strong potential for MIL application in real-world pathology diagnosis, with better performance, faster convergence and annotation efficiency. + +# Acknowledgement + +This study was partially supported by the National Natural Science Foundation of China (Grant no. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant no. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University. + +# References + +[1] Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy computation. IEEE transactions on pattern analysis and machine intelligence, 40(12):2897-2905, 2018. 2, 3 +[2] Alexander A. Alemi, Ian Fischer, Joshua V. Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 2, 3, 4 +[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10(7):e0130140, 2015. 3 +[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021.3 +[5] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 5 +[6] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature medicine, 28(1):154-163, 2022. 1 +[7] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2, 6 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. CoRR, abs/2104.14294, 2021. 2, 3, 6, 7 +[9] Richard J. Chen and et al. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In CVPR, pages 16144-16155, June 2022. 2, 3, 6, 7 +[10] Richard J Chen and Rahul G Krishnan. Self-supervised vision transformers learn visual concepts in histopathology. arXiv preprint arXiv:2203.00585, 2022. 7 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6, 7 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2018. cite arxiv:1810.04805Comment: 13 pages. 3 +[13] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Littjens, Jeroen A. W. M. van der Laak, , and the CAME- + +LYON16 Consortium. Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA, 318(22):2199-2210, 12 2017. 1 +[14] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30(4):681–694, 2020. 3 +[15] Ziv Goldfeld, Ewout van den Berg, Kristjan Greenewald, Igor Melnyk, Nam Nguyen, Brian Kingsbury, and Yury Polyanskiy. Estimating information flow in deep neural networks. arXiv preprint arXiv:1810.05728, 2018. 3 +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 2, 3 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3, 6, 7 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[19] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 2127-2136. PMLR, 10-15 Jul 2018. 1, 2, 6 +[20] Syed Ashar Javed, Dinkar Juyal, Zahl Shanis, Shreya Chakraborty, Harsha Pokkalla, and Aaditya Prakash. Rethinking machine learning model evaluation in pathology. arXiv preprint arXiv:2204.05205, 2022. 7 +[21] Durk P Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. Advances in neural information processing systems, 28, 2015. 5 +[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5 +[23] Erich Leo Lehmann and Henry Scheffé. Completeness, similar regions, and unbiased estimation-part i. In Selected works of EL Lehmann, pages 233-268. Springer, 2012. 2, 3 +[24] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3 +[25] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2021. 2, 3, 4, 6 +[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[27] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3 + +[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 3 +[29] Geert Litjens, Peter Bandi, Babak Ehteshami Bejnordi, Oscar Geessink, Maschenka Balkenhol, Peter Bult, Altuna Halilovic, Meyke Hermsen, Rob van de Loo, Rob Vogels, et al. 1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset. *GigaScience*, 7(6):giy065, 2018. 5 +[30] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 2, 5, 6, 7, 8 +[31] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1 +[32] Bhargavi Paranjape, Mandar Joshi, John Thickstun, Hannaneh Hajishirzi, and Luke Zettlemoyer. An information bottleneck approach for controlling conciseness in rationale extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1938-1952, Online, Nov. 2020. Association for Computational Linguistics. 3, 4 +[33] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. 3 +[34] Nicholas A Petrick, Shazia Akbar, Kenny HH Cha, Sharon Nofech-Mozes, Berkman Sahiner, Marios A Gavrielides, Jayashree Kalpathy-Cramer, Karen Drukker, Anne LL Martel, et al. Spie-aapm-nci breastpathq challenge: an image analysis challenge for quantitative tumor cellularity assessment in breast cancer histology images following neoadjuvant treatment. Journal of Medical Imaging, 8(3):034501, 2021. 5 +[35] Andrew M Saxe, Yamini Bansal, Joel Dapello, Madhu Advani, Artemy Kolchinsky, Brendan D Tracey, and David D Cox. On the information bottleneck theory of deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124020, 2019. 3 +[36] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. arXiv preprint arXiv:2001.00396, 2020. 3 +[37] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 3 +[38] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, and yongbing zhang. Transmil: Transformer based correlated multiple instance learning for whole + +slide image classification. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2136-2147. Curran Associates, Inc., 2021. 2, 6, 7, 8 +[39] Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017. 3 +[40] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3, 4 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2 +[42] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 2 +[43] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. 2021. 2 +[44] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. Xlnet: Generalized autoregressive pretraining for language understanding. In NeurIPS, 2019. 3 +[45] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E. Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. ArXiv, abs/2203.12081, 2022. 1, 2, 6, 7, 8 +[46] Yunlong Zhang, Yuxuan Sun, Honglin Li, Sunyi Zheng, Chenglu Zhu, and Lin Yang. Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2022. 7 +[47] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdolahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3 +[48] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9140-9150, June 2022. 3 +[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016. 3 \ No newline at end of file diff --git a/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/images.zip b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c073217768e11517c0c472885910899fb411a6aa --- /dev/null +++ b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12d1f978270a998b4f769bf7a3894332d0ca41e38a72e2cdf85f531e1a670e6a +size 565332 diff --git a/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/layout.json b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6204a7d355237c600d0f0c53bbec3f18f6505649 --- /dev/null +++ b/2023/Task-Specific Fine-Tuning via Variational Information Bottleneck for Weakly-Supervised Pathology Whole Slide Image Classification/layout.json @@ -0,0 +1,9108 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 507, + 140 + ], + "type": "text", + "content": "Task-specific Fine-tuning via Variational Information Bottleneck for Weakly-supervised Pathology Whole Slide Image Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "text", + "content": "Honglin Li" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "text", + "content": ", Chenglu Zhu" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "text", + "content": ", Yunlong Zhang" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "text", + "content": ", Yuxuan Sun" + }, + { + "bbox": [ + 138, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "spans": [ + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "text", + "content": "Zhongyi Shui" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "text", + "content": ", Wenwei Kuang" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "text", + "content": ", Sunyi Zheng" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "text", + "content": ", Lin Yang" + }, + { + "bbox": [ + 140, + 175, + 456, + 190 + ], + "type": "inline_equation", + "content": "^{2,3*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 189, + 460, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 189, + 460, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 189, + 460, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 189, + 460, + 204 + ], + "type": "text", + "content": "College of Computer Science and Technology, Zhejiang University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 203, + 524, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 203, + 524, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 524, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 203, + 524, + 217 + ], + "type": "text", + "content": "Research Center for Industries of the Future and " + }, + { + "bbox": [ + 69, + 203, + 524, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 203, + 524, + 217 + ], + "type": "text", + "content": "School of Engineering, Westlake University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 222, + 217, + 372, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 217, + 372, + 232 + ], + "spans": [ + { + "bbox": [ + 222, + 217, + 372, + 232 + ], + "type": "text", + "content": "4The University of Hong Kong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 201, + 233, + 389, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 233, + 389, + 245 + ], + "spans": [ + { + "bbox": [ + 201, + 233, + 389, + 245 + ], + "type": "text", + "content": "{lihonglin, yanglin}@westlake.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 299, + 290, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 290, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 290, + 598 + ], + "type": "text", + "content": "While Multiple Instance Learning (MIL) has shown promising results in digital Pathology Whole Slide Image (WSI) analysis, such a paradigm still faces performance and generalization problems due to high computational costs and limited supervision of Gigapixel WSIs. To deal with the computation problem, previous methods utilize a frozen model pretrained from ImageNet to obtain representations, however, it may lose key information owing to the large domain gap and hinder the generalization ability without image-level training-time augmentation. Though Self-supervised Learning (SSL) proposes viable representation learning schemes, the downstream task-specific features via partial label tuning are not explored. To alleviate this problem, we propose an efficient WSI fine-tuning framework motivated by the Information Bottleneck theory. The theory enables the framework to find the minimal sufficient statistics of WSI, thus supporting us to fine-tune the backbone into a task-specific representation only depending on WSI-level weak labels. The WSI-MIL problem is further analyzed to theoretically deduce our fine-tuning method. We evaluate the method on five pathological WSI datasets on various WSI heads. The experimental results show significant improvements in both accuracy and generalization compared with previous works. Source code will be available at https://github.com/invoker-LL/WSI-finetuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "content": "Digital Pathology or microscopic images have been widely used for the diagnosis of cancers such as Breast Cancer [13] and Prostate Cancer [6]. However, the reading of Whole Slide Images (WSIs) with gigapixel resolution is" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 293, + 423, + 357 + ], + "blocks": [ + { + "bbox": [ + 345, + 271, + 402, + 289 + ], + "lines": [ + { + "bbox": [ + 345, + 271, + 402, + 289 + ], + "spans": [ + { + "bbox": [ + 345, + 271, + 402, + 289 + ], + "type": "text", + "content": "a. ImageNet-1k v-score: 0.285" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 293, + 423, + 357 + ], + "lines": [ + { + "bbox": [ + 322, + 293, + 423, + 357 + ], + "spans": [ + { + "bbox": [ + 322, + 293, + 423, + 357 + ], + "type": "image", + "image_path": "e6924865837e9beb58e361df843e2650f9342cb52ad2accdf50e53de468bb4a9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 429, + 291, + 529, + 357 + ], + "blocks": [ + { + "bbox": [ + 446, + 271, + 513, + 289 + ], + "lines": [ + { + "bbox": [ + 446, + 271, + 513, + 289 + ], + "spans": [ + { + "bbox": [ + 446, + 271, + 513, + 289 + ], + "type": "text", + "content": "b. Full Supervision v-score: 0.745" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 291, + 529, + 357 + ], + "lines": [ + { + "bbox": [ + 429, + 291, + 529, + 357 + ], + "spans": [ + { + "bbox": [ + 429, + 291, + 529, + 357 + ], + "type": "image", + "image_path": "b16c017a48df0351af1861e5a76d45e2840b8169638d446149845fb224d0ab5d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 322, + 380, + 423, + 447 + ], + "blocks": [ + { + "bbox": [ + 340, + 361, + 406, + 377 + ], + "lines": [ + { + "bbox": [ + 340, + 361, + 406, + 377 + ], + "spans": [ + { + "bbox": [ + 340, + 361, + 406, + 377 + ], + "type": "text", + "content": "c. Self-supervised v-score: 0.394" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 380, + 423, + 447 + ], + "lines": [ + { + "bbox": [ + 322, + 380, + 423, + 447 + ], + "spans": [ + { + "bbox": [ + 322, + 380, + 423, + 447 + ], + "type": "image", + "image_path": "8dd7bdc4b45348a884325912f55e7268cc1c8e2c01c4ddad7575050ca58800dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 380, + 529, + 447 + ], + "blocks": [ + { + "bbox": [ + 437, + 361, + 521, + 377 + ], + "lines": [ + { + "bbox": [ + 437, + 361, + 521, + 377 + ], + "spans": [ + { + "bbox": [ + 437, + 361, + 521, + 377 + ], + "type": "text", + "content": "d. IB Fine-tuning (ours) \nv-score: 0.538" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 380, + 529, + 447 + ], + "lines": [ + { + "bbox": [ + 429, + 380, + 529, + 447 + ], + "spans": [ + { + "bbox": [ + 429, + 380, + 529, + 447 + ], + "type": "image", + "image_path": "34c05d368cfa1dafa6cfdbce0d49836e0029f6ed549d71c4eaf2271c900ecb11.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 457, + 547, + 536 + ], + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 536 + ], + "type": "text", + "content": "Figure 1. T-SNE visualization of different representations on patches. Our method converts chaotic ImageNet-1K and SSL features into a more task-specific and separable distribution. The cluster evaluation measurement, v-scores, show weakly supervised fine-tuned features are more close to full supervision compared to others. a. ImageNet-1k pretraining. b. Full patch supervision. c. Self-supervised Learning. d. Fine-tuning with WSI labels." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "time-consuming which poses an urgent need for automatic computer-assisted diagnosis. Though computers can boost the speed of the diagnosis process, the enormous size of resolution, over 100M [45], makes it infeasible to acquire precise and exhaustive annotations for model training, let alone the current hardware can hardly support the parallel training on all patches of a WSI. Hence, an annotation-efficient learning scheme with light computation is increasingly desirable to cope with those problems. In pathology WSI analysis, the heavy annotation cost is usually alleviated by Multiple Instance Learning (MIL) with only WSI-level weak supervision, which makes a comprehensive decision on a series of instances as a bag sample [19, 31]. Intuitively, all" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7454" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 108 + ], + "type": "text", + "content": "small patches in the WSI are regarded as instances to constitute a bag sample [7, 30, 38], where the WSI's category corresponds to the max lesion level of all patch instances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 110, + 286, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 110, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 110, + 286, + 385 + ], + "type": "text", + "content": "However, most methods pay much effort to design WSI architectures while overlooking the instance-level representation ability. Because of the computational limitation, the gradient at the WSI-level is impossible to parallelly backpropagate to instance encoders with more than 10k instances of a bag. Thus parameters of the pretrained backbone from ImageNet-1k (IN-1K) are frozen to obtain invariant embeddings. Due to the large domain gap between IN-1K and pathological images, some essential information may be discarded by layers of frozen convolutional filters, which constrains the accuracy of previous WSI methods. To address the constraint, recent works [9, 25] make efforts to learn a good feature representation at the patchlevel by leveraging Self-supervised Learning (SSL). However, such task-agnostic features are dominated by the proxy objective of SSL, e.g. Contrastive Learning in [8, 11, 17] may push away the distance between two instances within the same category, thus only performs slightly better than IN-1K pretraining in WSI classification. Nearly all SSL methods [8, 11, 16, 17] proposed on natural image recognition utilize a small portion of annotations to get promising fine-tuning accuracy compared to full supervision, which is higher than Linear Probing [17] by a large margin." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 388, + 286, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 543 + ], + "type": "text", + "content": "These findings illuminate us to design a fine-tuning scheme for WSI analysis to convert IN-K or SSL task-agnostic representations into task-specifics. Motivated by the Information Bottleneck (IB) theory [1, 2], we argue that pretraining is limited to downstream tasks, therefore fine-tuning is necessary for WSI analysis. In addition, we develop a solution based on Variational IB to tackle the dilemma of fine-tuning and computational limitation by its minimal sufficient statistics and attribution properties [1, 23]. The differences among the above three feature representations are depicted in Figure 1, where the feature representation under full patch-level supervision is considered as the upper bound." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "content": "Our main contributions are in 3 folds: 1) We propose a simple agent task of WSI-MIL by introducing an IB module that distills over " + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "content": " redundant instances within a bag into less than " + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "inline_equation", + "content": "1\\mathrm{k}" + }, + { + "bbox": [ + 49, + 545, + 286, + 712 + ], + "type": "text", + "content": " of the most supported instances. Thus the parallel computation cost of gradient-based training on Gigapixel Images is over ten times relieved. By learning and making classification on the simplified bag, we find that there are trivial information losses due to the low-rank property of pathological WSI, and the distilled bag makes it possible to train a WSI-MIL model with the feature extractor on patches end-to-end, thus boosting the final performance. 2) We argue that the performance can be further improved by combining with the SSL pretraining since we could convert the task-agnostic representation from SSL into task-specific" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "type": "text", + "content": "one by well-designed fine-tuning. The proposed framework only relies on annotations at WSI levels, which is similar to recent SSL approaches [8, 11, 16, 17]. Note that our method only utilizes less than a " + }, + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "type": "text", + "content": " fraction of full patch annotation to achieve competitive accuracy compared to counterparts. 3) Versatile training-time augmentations can be incorporated with our proposed fine-tuning scheme, thus resulting in better generalization in various real-world or simulated datasets with domain shift, which previous works ignore to validate. These empirical results show that our method advances accuracy and generalization simultaneously, and thus would be more practical for real-world applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 242, + 391, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 391, + 254 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 391, + 254 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 266, + 539, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 539, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 539, + 278 + ], + "type": "text", + "content": "2.1. Multiple Instance Learning for WSI Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 290, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 290, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 308, + 290, + 545, + 433 + ], + "type": "text", + "content": "Multiple Instance Learning (MIL) is a well-defined task and has been explored extensively. Currently, there are mainly two lines of methods for WSI analysis: 1) explicit modeling the MIL definition " + }, + { + "bbox": [ + 308, + 290, + 545, + 433 + ], + "type": "inline_equation", + "content": "Y = \\text{pooling}\\{y_1, y_2, \\dots, y_n\\}" + }, + { + "bbox": [ + 308, + 290, + 545, + 433 + ], + "type": "text", + "content": " that WSI level prediction is aggregated by the probability of all patches with Mean or Max-pooling [7,45]. 2) implicit learning WSI level representation by aggregating all patches embeddings via a WSI classifier with Recurrent Neural Network (RNN) [7] or in an attention [19] mechanism. The latter shows superior performance since such modeling includes less inductive bias compared to the former processing with fixed weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 438, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 712 + ], + "type": "text", + "content": "RNN [7] treats WSI as a sequence of patches, but its permutation variation and defects on long-term dependency do not match the nature of WSI well. By contrast, attention-based MIL (AB-MIL) [19] learns the weights of instance representations adaptively, and these methods can be flexibly inserted into the classification framework as a plug-and-play module to generalize the full-field map according to the task. To make patches' representation more discriminative, CLAM [30] introduces an auxiliary task in the MIL framework to distinguish its corresponding availability according to the size of instance attention while training the WSI classification. Trans-MIL [38] is designed to model the relation among patches via Self-Attention [41] and solves the softmax computational complexity in the long sequence by linear Self-Attention [42, 43]. To embed the multi-scale information of WSI, DS-MIL [25] simply concatenates patch features in three scales, and HIPT [9] builds hierarchical multi-scale features. To mitigate overfitting in complex attention model trained on the limited number of WSIs, DTFD-MIL [45] resamples instances from the original bag to generate various sub-bags for augmentation. However, all these works pay too much attention on WSI head to explore the backbone." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "7455" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 521, + 216 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 521, + 216 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 521, + 216 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 521, + 216 + ], + "type": "image", + "image_path": "fac218ab8ba49211d7bcd248afda7952532317d0a5308b6a7335c3d2bfc25d5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": "Figure 2. An illustration of the tuning scheme. To maximize " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "R_{IB}" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": ", enlarging the first term helps latent representation vector " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": " become more predictive for label " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": " (More overlap between " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Z2" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Z1" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": ", which denotes higher mutual information), while by reducing the second term to filter irrelevant features (smaller overlap between image " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Z2" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "inline_equation", + "content": "Z1" + }, + { + "bbox": [ + 46, + 228, + 547, + 264 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 264, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 264, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 264, + 290 + ], + "type": "text", + "content": "2.2. Self-supervised Learning and Fine-tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 523 + ], + "type": "text", + "content": "Self-supervised Learning (SSL) has shown to be a promising paradigm both in computer vision [4,8,11,16,17] and natural language processing [12,33,44]. However, fine-tuning (FT) is quite necessary for the pretrained model of SSL to the downstream tasks, e.g. in [16,17], the vanilla FT method can improve about 15 percent of accuracy in IN-1K compared to vanilla Linear Probing [17]. More interesting and parameter-efficient FT methods [14,18,24,26,27,47,48] are proposed according to the similarity between pretraining and downstream in natural language processing and few-shot image generation. Unfortunately, the computational limitation on redundant instances prevents FT from being directly introduced into WSI analysis. Although SSL achieves optimistic performance in some works [9, 25] for pathology WSI analysis but only shows a small margin of improvement. There may be much potential to be discovered in WSI analysis if an appropriate FT method can be used to address the training cost from surplus information of WSI." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 536, + 261, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 261, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 261, + 548 + ], + "type": "text", + "content": "2.3. Information Bottleneck and Attributions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": "The Information Bottleneck (IB) conception is introduced in [40] as an information-theoretic framework for learning, which is currently used in deep learning both theoretically and practically. [39] proposes to unveil deep neural network in an information flow perspective by estimating Mutual Information (MI) between the outputs of two layers, then in followed-up works [15, 35], they revisit the MI compression process of IB to propose complements and better measurements of MI in Deep Neural Network (DNN)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "IB not only reveals how the traditional DNN training finishes information compression implicitly in the above works, but can also be employed explicitly as an objective to intervene in DNN's training: [2] derives a variational" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 277, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 542 + ], + "type": "text", + "content": "bound of the IB objective, claiming that the proxy objective provides an additional regularization term which improves the robustness to adversarial samples compared to the traditional training of DNN. [32] adopts the variational bound in [2] and replaces the distribution of latent features from Gaussian into Bernoulli. In this way, it generates masks to control the conciseness of the rationale extraction of sentences. Achille et al. [1] claim that IB solutions approximate minimal sufficient statistics (MSS) [23], and by penalizing the redundancy of representations, they heuristically argue that the model's sensitivity to nuisances is mitigated. Similar to some visual attribution methods like CAM [37, 49], LRP [3] and patch masking [4, 16, 28] in ViT, an IB-based attribution method is proposed in [36] by adding noise to intermediate feature maps, restricting the flow of information, then how much information image regions provide can be quantified. Intuitively, the IB objective reveals the limitation of pretraining for downstream tasks and the imperative of FT for task-specific representation in WSI analysis, as shown in Figure 2. Moreover, it motivates us to solve the dilemma of FT and computational limitation by its' MSS and visual attribution properties." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 551, + 362, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 362, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 362, + 563 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 571, + 509, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 509, + 583 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 509, + 583 + ], + "type": "text", + "content": "3.1. Overview on MIL-based WSI Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": "Given a WSI " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": ", the goal is to make slide-level prediction " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": " by learning a classifier " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "f(X; \\theta)" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": ". Due to its extremely high resolution, " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": " is patched into a huge bag of small instances " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "X = \\{x_1, \\dots, x_N\\}" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": " is the number of instance. The slide-level supervision " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": " is given by a Max-pooling operation of the latent label " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": " for each instance " + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 588, + 545, + 671 + ], + "type": "text", + "content": ", which can be defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 671, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 671, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 378, + 671, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\hat {Y} = \\max \\left\\{\\hat {y} _ {1}, \\dots , \\hat {y} _ {N} \\right\\}. \\tag {1}", + "image_path": "91f346cebeabb7bfae86efe0656bc86db25411f5830df757f08c7946071307ce.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": "Since all latent labels of instances " + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": " are unknown under the WSI-level supervision, conventional approaches con" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7456" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": "vert this problem into a MIL formulation in the following two steps: 1) Processing images into feature representations " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "Z = \\{z_{1},\\dots,z_{N}\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": " with a backbone " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "z_{i} = h(x_{i};\\theta_{1})" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": " is a model of any architecture such as CNN or ViT with parameters " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": ". 2) Aggregating all patches' features within a slide and producing the slide-level prediction " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "Y = g(Z;\\theta_2)" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 72, + 289, + 168 + ], + "type": "text", + "content": " is an attention-based pooling function followed by a linear classifier head as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 175, + 287, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 175, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 287, + 208 + ], + "type": "interline_equation", + "content": "g (Z; \\theta_ {2}) = \\sigma \\left(\\sum_ {i = 1} ^ {N} a _ {i} z _ {i}\\right), \\tag {2}", + "image_path": "ba4f6baf868686f1ee23426534d49e518a9d2bc05eede0292a9030cc16192110.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " is attention weights and " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " is a linear head. Limited by the computational cost, the parameters " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "f(X;\\theta) = g\\{h(X;\\theta_1);\\theta_2\\}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " are learned separately by following steps: 1) Initializing " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " from the pretrained model, which refers to general features from public IN-1K, or learned by SSL on the related dataset to extract the domain-specific representations. 2) Freezing " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " and learning " + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 47, + 215, + 289, + 312 + ], + "type": "text", + "content": " under slide-level supervision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 263, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 263, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 263, + 331 + ], + "type": "text", + "content": "3.2. Information Bottleneck for MIL Sparsity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 336, + 233, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 233, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 233, + 347 + ], + "type": "text", + "content": "Background of the Information Bottleneck" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 348, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 396 + ], + "type": "text", + "content": "The Information Bottleneck (IB) can work as an information compression role to intervene in DNN's training [2]. The objective function of IB to be maximized is given in [40] as," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 396, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 396, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 287, + 409 + ], + "type": "interline_equation", + "content": "R _ {I B} = I (Z, Y) - \\beta I (Z, X), \\tag {3}", + "image_path": "38548a0b1ffd7641341313c702af68d10444950ad972d084caf35cd1176d2fe6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "I(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "content": " indicates the Mutual Information (MI) and " + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "content": " is a Lagrange multiplier controlling the trade-off between the information that the representation variable " + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "content": " shares with the label " + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "content": " and its shares with input " + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 413, + 287, + 509 + ], + "type": "text", + "content": ". Since the computation of MI is intractable during the training of the neural networks, to maximize IB objective can be transferred to minimize a variational bound of Eq.(3) derived in [2] follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 514, + 287, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 75, + 514, + 287, + 546 + ], + "type": "interline_equation", + "content": "J _ {I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {z \\sim p _ {\\theta} (z | x _ {n})} [ - \\log q _ {\\phi} (y _ {n} | z) ] + \\tag {4}", + "image_path": "a57b33bf0455e94a8eef8cf0e2f3ddb3f5f2082f8b5172c57e180d577f1347b5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 167, + 548, + 259, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 548, + 259, + 563 + ], + "spans": [ + { + "bbox": [ + 167, + 548, + 259, + 563 + ], + "type": "interline_equation", + "content": "\\beta K L [ p _ {\\theta} (z | x _ {n}), r (z) ],", + "image_path": "51a74b4c244bd398cc3725ec2ff6c0c758922797a61ab6b04eb2d527df19c04e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": " denotes the number of samples, " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "q_{\\phi}(y|z)" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": " is a parametric approximation to the likelihood " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "p(y|z), r(z)" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": " is the prior probability of " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": " to variational approximate the marginal " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "p_{\\theta}(z|x)" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": " is the parametric posterior distribution over " + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 570, + 287, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 630, + 257, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 257, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 257, + 641 + ], + "type": "text", + "content": "Learn MIL Sparsity via Variational Bound of IB" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 689 + ], + "type": "text", + "content": "To trade off the dilemma of computational limitation and task-specific representation learning via end-to-end backpropagation, we propose to utilize the IB module to filter most task-irrelevant instances for task-specific fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The above filtering process can be implemented by optimizing the second term of in Eq.(3) which controls the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "compression. There are two ways that compress " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " by decreasing the KL divergence between " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "p(z|x)" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "r(z)" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " in Eq.(4) variational method: reducing the dimension of representation " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " in [2], or converting input " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " into a sparse one in [32]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "text", + "content": "For the setting of our long instance sequenced MIL, we reduce " + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "inline_equation", + "content": "I(X,Z)" + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "text", + "content": " into a degree so that the gradients can be back-propagated to the backbone encoder, which needs us to convert a WSI of bag size over 10k into 1k for the sake of sparsity. Considering MIL for tumor v.s. normal binary classification without loss of generality and the latent label " + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "text", + "content": " of each instance " + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 133, + 546, + 262 + ], + "type": "text", + "content": " in Eq.(1), we argue that it is sufficient enough to make the WSI level prediction if one tumor area is detected. With the above understanding, we propose to learn compressed components similar to [32] by defining a IB module as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 400, + 266, + 545, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 266, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 400, + 266, + 545, + 277 + ], + "type": "interline_equation", + "content": "z = m \\odot x, \\tag {5}", + "image_path": "daf2f90ee292412321868589759bbbcc81438f13eb8d1ccec570ec037f8ea534.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": " is a Bernoulli " + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "inline_equation", + "content": "(\\pi)" + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": " distributed binary mask and in this way " + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "inline_equation", + "content": "KL[p_{\\theta}(z|x),r(z)]" + }, + { + "bbox": [ + 304, + 283, + 545, + 319 + ], + "type": "text", + "content": " in Eq.(4) can be decomposed as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 356, + 319, + 545, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 319, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 356, + 319, + 545, + 333 + ], + "type": "interline_equation", + "content": "K L \\left[ p _ {\\theta} \\left(m _ {i} | x\\right), r \\left(m _ {i}\\right) \\right] + \\pi H (X), \\tag {6}", + "image_path": "3e0ce812707f7c9367073c5aa17cc9b53e8fef3fde79bf7456c3e7a34a0ba77b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "type": "inline_equation", + "content": "H(X)" + }, + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "type": "text", + "content": " is the entropy of " + }, + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 338, + 545, + 374 + ], + "type": "text", + "content": ", which can be omitted during the minimization due to its constant value. Please check Supplementary for above proof." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "content": "The Bernoulli " + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "inline_equation", + "content": "(\\pi)" + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "content": " distribution for " + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "content": " fits the definition of MIL empirically: we can treat " + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "content": " as a latent weak prediction " + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "content": " describing whether the patch contains tumor or not, denoting " + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "inline_equation", + "content": "P_{set} = \\{p(m_1|x_1),\\dots,p(m_N|x_N)\\}" + }, + { + "bbox": [ + 304, + 375, + 546, + 435 + ], + "type": "text", + "content": ", then during inference Eq.(1) can be derived as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 353, + 444, + 545, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 444, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 353, + 444, + 545, + 459 + ], + "type": "interline_equation", + "content": "\\hat {Y} = \\max \\left\\{P _ {s e t} \\right\\} = \\max \\left\\{P _ {s u b s e t} \\right\\}, \\tag {7}", + "image_path": "f922cb9831be4df9b1df858f9e2f213710fc70d8fb0b9733baa7c50a96872ee1.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "P_{\\text{subset}} \\in P_{\\text{set}}" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": ", generated by select top-K elements in " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "P_{\\text{set}}" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": ". The patch classifier trained with only slide-level supervision shows low accuracy [25], so we only use it to generate mask for sparse sub-bag and still utilize attention-based MIL on the sub-bag for decision making." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 536, + 507, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 507, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 507, + 548 + ], + "type": "text", + "content": "3.3. Loss Function and its Implementation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 555, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 579 + ], + "type": "text", + "content": "Derived from Eq.(4) and Eq.(6), our Variational IB module should be optimized with the following loss function," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 332, + 587, + 545, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 587, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 332, + 587, + 545, + 636 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {l o s s} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {z \\sim p _ {\\theta} (z | x _ {n})} [ - \\log q _ {\\phi} (y _ {n} | z) ] + \\tag {8} \\\\ \\beta K L [ p _ {\\theta} (m | x _ {n}), r (m) ], \\\\ \\end{array}", + "image_path": "71a32b147b476d70aeadc25d02cd1c60acc9920f6d6ba8a0ef4f6f5c5843b3cf.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "where the first term is the task loss to learn task-specific features, and it can be treated as the cross entropy same to prior works [2, 32] in a sampling perspective. The second term is the information loss to filter out task-irrelevant instances by minimizing the Kullback-Leibler divergence between the distribution of mask and the prior Bernoulli." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7457" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 73, + 531, + 246 + ], + "blocks": [ + { + "bbox": [ + 63, + 73, + 531, + 246 + ], + "lines": [ + { + "bbox": [ + 63, + 73, + 531, + 246 + ], + "spans": [ + { + "bbox": [ + 63, + 73, + 531, + 246 + ], + "type": "image", + "image_path": "c54e0a4389595a0c7e8e568f600bc565c1e36bc3fd910646357eda573e0de30a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 256, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 547, + 291 + ], + "type": "text", + "content": "Figure 3. Workflow of WSI-MIL task-specific fine-tuning. 1) Initialize the backbone with pretrained parameters and set frozen, then learn the IB module to generate instance masks. 2) fix the mask to distill a sparse bag, then fine-tune the WSI head and patch the backbone end-2-end. 3) utilize all fine-tuned instance features within a bag and train the WSI-MIL classifier head." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "p(m_{n}|x_{n})" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": " is first generated by a linear layer from the representation of " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": " and followed by a sigmoid activation. If the mask " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": " is directly converted by thresholding " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "p(m_{n}|x_{n})" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": ", the gradients from " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "z = m\\odot x" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": " can not be backpropagated since such an operation is not differentiable. So, during training, " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": " is generated via Monte Carlo sampling, and " + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "inline_equation", + "content": "p(m_n|x_n)" + }, + { + "bbox": [ + 46, + 302, + 289, + 399 + ], + "type": "text", + "content": " can be learned with the reparameterization trick [21, 22] for gradient's estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 408, + 188, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 188, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 188, + 422 + ], + "type": "text", + "content": "3.4. Task-specific Fine-tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 427, + 289, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 289, + 643 + ], + "type": "text", + "content": "To realize task-specific fine-tuning and better performance with only slide-level labels, the workflow of our method is depicted in Figure 3, which includes 3 stages: 1) By minimizing the loss function in section 3.3 with a frozen backbone, a sparse instance set within a bag is obtained. 2) Assisted by the small portion of instances in a bag generated from stage 1, it is possible to fine-tune the backbone for better task-specific features and here we directly backpropagate the loss gradients from slide-level supervision into top-K patch instances. 3) Since the appropriateness of the sparse instance set heavily relies on the performance of the pretrained backbone, the first training IB module is insufficient and may result in small or even zero top-K recall. Moreover, the sparse instance set may lose the contextual dependency of WSI modeling. Thus in this stage, We utilize all fine-tuned instance features within a bag to train a traditional attention-based WSI-MIL classifier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 656, + 128, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 128, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 128, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "In this section, we present the performance of the proposed method incorporated with the latest WSI-MIL frameworks, where the IN-1K and SSL features are used for com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 302, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 338 + ], + "type": "text", + "content": "parisons. Ablation experiments are performed to further study the proposed method and for paper length, more experimental results are presented in the Supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 339, + 392, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 392, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 392, + 350 + ], + "type": "text", + "content": "Datasets and Tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "type": "text", + "content": "We use five datasets to evaluate our method. The slide-level classification performance of IN-1K, SSL, and our method is evaluated on three datasets including both histopathology and cytopathology images: two public histopathological WSI datasets, Camelyon-16 [5] for tumor / normal binary classification, The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) [34] for tumors subtyping. One internal cytopathology WSI dataset is introduced to validate the universality of our method on both histo- and cyto-pathology, which is Liquid-based Preparation cytology for Cervical Cancer's early screening (LBP-CECA)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 579 + ], + "type": "text", + "content": "Despite the original evaluation on the closed dataset, we also evaluate the generalization of our method on Camelyon-16-C (generated with random synthetic domain shift from Camelyon-16), and Camelyon-17 [29] from five different centers, which occurs frequently in practical pathological diagnosis and has been hindering the application of automatical WSI analysis to the real world. Details of five datasets can be found in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": "For pre-processing, we follow the operations in CLAMSB [30] which mainly includes HSV, Blur, Thresholding, and Contours methods to localize the tissue regions in each WSI. Then non-overlapping patches with size " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "inline_equation", + "content": "20\\mathrm{X}" + }, + { + "bbox": [ + 304, + 580, + 545, + 640 + ], + "type": "text", + "content": " magnification are extracted from the tissue regions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 653, + 432, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 432, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 432, + 665 + ], + "type": "text", + "content": "Pretraining and Fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Our work mainly focuses on the method of fine-tuning. Because good pretrained initialization results in better finetuning performance, we employ mainstream pretraining methods: 1) ImageNet-1k (IN-1K) data pretraining. 2) SSL" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7458" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 72, + 542, + 342 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 542, + 342 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 542, + 342 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 542, + 342 + ], + "type": "table", + "html": "
MethodCamelyon-16TCGA-BRCALBP-CECA
F1AUCF1AUCF1AUC
Full Supervision0.967±0.0050.992±0.003--0.741±0.0060.942±0.002
RNN-MIL [7]0.834±0.0170.861±0.0210.776±0.0350.871±0.033--
AB-MIL [19]0.828±0.0130.851±0.0250.771±0.0400.869±0.0370.525±0.0170.845±0.002
DS-MIL [25]0.857±0.0230.892±0.0120.775±0.0440.875±0.041--
CLAM-SB [30]0.839±0.0180.875±0.0280.797±0.0460.879±0.0190.587±0.0140.860±0.005
TransMIL [38]0.846±0.0130.883±0.0090.806±0.0460.889±0.0360.533±0.0060.850±0.007
DTFD-MIL [45]0.882±0.0080.932±0.0160.816±0.0450.895±0.0420.569±0.0260.847±0.003
FT+ CLAM-SB0.911±0.0170.956±0.0130.845±0.0320.935±0.0270.718±0.0100.907±0.005
FT+ TransMIL0.923±0.0120.967±0.0030.848±0.0440.945±0.0200.720±0.0240.918±0.004
FT+ DTFD-MIL0.921±0.0070.962±0.0060.849±0.0270.951±0.0160.723±0.0080.922±0.005
Mean-pooling0.629±0.0290.591±0.0120.818±0.0220.910±0.0320.350±0.0170.735±0.006
Max-pooling0.805±0.0120.824±0.0160.644±0.1790.826±0.0960.636±0.0640.893±0.019
KNN (Mean)0.468±0.0000.506±0.0000.633±0.0660.749±0.0550.393±0.0000.650±0.000
KNN (Max)0.559±0.0000.535±0.0000.524±0.0320.639±0.0630.477±0.0000.743±0.000
FT+ Mean-pooling0.842±0.0060.831±0.0070.866±0.0350.952±0.0180.685±0.0140.900±0.002
FT+ Max-pooling0.927±0.0110.969±0.0040.852±0.0430.948±0.0190.695±0.0130.912±0.004
FT+ KNN (Mean)0.505±0.0000.526±0.0000.784±0.0440.907±0.0340.529±0.0000.737±0.000
FT+ KNN (Max)0.905±0.0000.916±0.0000.802±0.0630.882±0.0360.676±0.0000.875±0.000
", + "image_path": "9d8e728b497cc15aeca67b46288a4fcff18695e12e6660992e5adc8d9540a423.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 352, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 545, + 398 + ], + "type": "text", + "content": "Table 1. Slide-Level Classification by using the IN-1K pre-trained backbone or the proposed fine-tuned (FT) in three datasets. Top Rows. Different MIL architectures are compared to select the top 3 SOTA methods to validate the transfer learning performance using the IN-1K pre-trained backbone or the FT. Bottom Rows. The competition of various traditional aggregation and feature evaluation methods by using pre-trained IN-1K or the FT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "text", + "content": "pretraining with SimCLR [11], MoCo [17] and DINO [8]. 3) for Camelyon-16 and LBP-CECA, since there are comprehensive tumor area annotations, we use the annotation to pretrain the patch backbone, which acts as the upper-bound of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 473, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 628 + ], + "type": "text", + "content": "Different from prior works backbone frozen after pretraining, we perform backbone fine-tuning with the method proposed in 3.4. In this stage, we fine-tune the backbone and WSI model end-to-end with 25 epochs using the AdamW optimizer, batch size of 1 for WSI with bag size of 512, a learning rate of 1e-5 for backbone and 1e-3 for WSI head. For the BN layers in ResNet, we turn it to eval mode to fix statistics during fine-tuning since we find that the distribution of the top-K instances in a bag is limited to estimate statistics because of the similarity among instances. Shallow layers of backbone are frozen since they only focus on morphological features while deep layer focus on semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 183, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 183, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 183, + 647 + ], + "type": "text", + "content": "4.1. Slide-level Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. For all the experiments, macro-AUC and macro-F1 scores are reported since all the 3 datasets are class imbalanced. For Camelyon-16, the official training set is randomly split into training and validation sets with a ratio of 9:1. The experiment is conducted 5 times and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 414, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 509 + ], + "type": "text", + "content": "the results of the official test set are reported. For TCGA-BRCA, we perform the 10 folder cross-validation with the same running setting adopted in HIPT [9]. Besides the dataset LBP-CECA is randomly split with a ratio of 6:1:3 for training, validation, and testing. The experiment is conducted 5 times. The mean and standard variance values of performance metrics are reported for multi-runs or cross-validation runs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 510, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 688 + ], + "type": "text", + "content": "Comparison with baselines. Classification results are summarized in Table 1. We first show full patch supervision results as an upper bound, then we directly evaluate several classic WSI-MIL methods, including RNN-MIL [7], AB-MIL [19], DS-MIL [25], CLAM-SB [30], TransMIL [38], DTFD-MIL [45]. All the WSI-MIL baselines across the three classification tasks suffer from relatively low performance due to the inappropriate backbone features from the pretrained ResNet-50 in IN-1K. Then we apply FT to the backbone to obtain features for WSI classification by three WSI architectures including CLAM-SB (pure global attention, no intersection among instances), TransMIL (self-attention for long sequence), and DTFD-MIL (multi-tier attention paths followed with aggregation, good for patch imbalance)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "The results show clearly consistent improvements equipped with FT features under AUC metric. CLAM-SB" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7459" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "achieves a performance relative increase of " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "9.26\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "6.37\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "5.47\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " on Camelyon-16, TCGA-BRCA, and LBP-CECA respectively with vanilla FT based on IN-1K pretraining. In the same conditions, TransMIL and DTFD-MIL achieve new SOTAs on all three datasets and can obtain better relative improvement, especially in LBP-CECA with an " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "8.00\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "8.85\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " growth compared to CLAM-SB. The results show similar improvements under the F1 metric. For the bottom two rows of Table 1, we compare simple Mean/Max-pooling training and KNN evaluations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 373 + ], + "type": "text", + "content": "The competitive results demonstrate that our proposed FT method in the backbone of the MIL framework can boost the performance of WSI classification, even for the most simple feature-level Mean/Max-pooling. In addition, task-specific features help us better unearth the properties of different WSI-MIL architectures. Apparently, the more complex architecture, TransMIL, and DTFD-MIL, can not perform desirable performance with the frozen pretrained parameters in LBP-CECE. In contrast, its capabilities can be further enhanced by fine-tuning the backbone for the defined target. Moreover, the effects of different feature aggregation methods vary depending on the task. The better performance is shown by using Max-pooling aggregation in Camelyon-16 containing smaller tumor areas, while Mean-pooling is better in TCGA-BRCA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 378, + 244, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 244, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 244, + 391 + ], + "type": "text", + "content": "4.2. Combination of SSL and Fine-tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 286, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 286, + 420 + ], + "type": "text", + "content": "In this section, we further improve the slide-level classification performance by combining SSL with FT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 420, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 286, + 456 + ], + "type": "text", + "content": "Evaluation Metrics. TCGA-BRCA is used for the evaluation of its task complexity in tumor subtyping with the same setting to 4.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 286, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 480 + ], + "type": "text", + "content": "Comparison with baselines. Since there is no apparent single semantic object in small histopathology patches, we" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 498, + 286, + 647 + ], + "blocks": [ + { + "bbox": [ + 48, + 498, + 286, + 647 + ], + "lines": [ + { + "bbox": [ + 48, + 498, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 286, + 647 + ], + "type": "table", + "html": "
MethodF1AUC
IN-1K$-0.884±0.059
IN-1K0.797±0.0460.879±0.019
/w FT0.845±0.0320.935±0.027
SimCLR [11]$-0.879±0.069
MoCo [17]0.804±0.0420.904±0.030
/w FT0.851±0.0290.948±0.026
DINO [8]$-0.886±0.059
DINO0.801±0.0450.891±0.043
/w FT0.848±0.0270.944±0.036
", + "image_path": "21093e556b52e04337c059319f3408bbf0acbf474e1236e21d0a442963596969.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "mainly compare SSLs by Contrastive Learning or augmentation like MoCo [17] and DINO [8], and SimCLR [11]. We show the results of SSLs performed in previous works [9, 10] for fair comparison. All results are performed with the same WSI architecture of CLAM-SB [30]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": "Experimental results are summarized in Table 2. Compared to vanilla IN-1K, MoCo and DINO (combined with FT) achieve consistent growth of " + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "inline_equation", + "content": "7.85\\%" + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "inline_equation", + "content": "7.39\\%" + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": ", respectively. Besides, they have a slight increase of " + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "inline_equation", + "content": "1.39\\%" + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "inline_equation", + "content": "0.96\\%" + }, + { + "bbox": [ + 304, + 133, + 545, + 263 + ], + "type": "text", + "content": " in comparison to IN-1K with FT. By SSL, the intrinsic task-agnostic features are learned from all patches in WSIs, and after applying our proposed FT for all modules, task-specific features can be distilled from the label and partial data. In such an SSL with an FT paradigm, the data and the label are explored comprehensively to produce a state-of-the-art WSI analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 270, + 477, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 477, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 477, + 282 + ], + "type": "text", + "content": "4.3. Generalization on Domain Shift" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 289, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 348 + ], + "type": "text", + "content": "In this section, we evaluate the generalization of slide-level classification models on domain shift [46] [20], which is crucial for real-world clinical applications due to the diversity in staining, preparation, and imaging devices for pathological image processing among hospitals." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 349, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 479 + ], + "type": "text", + "content": "Evaluation Metrics. The Camelyon-16-C dataset is generated from Camelyon-16 test set by synthetic domain shift, which adopts a random combination of Brightness, JPEG, and Hue proposed in [46]. For Camelyon-17, since it shares similarities to Camelyon-16 but is collected from five different medical centers, we randomly collect 30 samples from each center to evaluate the robustness of natural domain shift. We directly evaluate models on Camelyon-16-C and 150 extra data from Camelyon-17 with five times of running same as the previous setting in section 4.1. The standard variance values of results is omitted due to the constraint on" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 306, + 498, + 547, + 647 + ], + "blocks": [ + { + "bbox": [ + 46, + 656, + 286, + 700 + ], + "lines": [ + { + "bbox": [ + 46, + 656, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 656, + 286, + 700 + ], + "type": "text", + "content": "Table 2. Combination of SSL and Fine-tuning. We compare SSLs with IN-1K and their further improvement via fine-tuning (FT) on TCGA-BRCA. The symbol " + }, + { + "bbox": [ + 46, + 656, + 286, + 700 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 46, + 656, + 286, + 700 + ], + "type": "text", + "content": " indicates the result released in previous publication [9, 10]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 498, + 547, + 647 + ], + "lines": [ + { + "bbox": [ + 306, + 498, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 547, + 647 + ], + "type": "table", + "html": "
MethodCamelyon-16-CCamelyon-17
F1AUCF1AUC
Max-pooling0.6890.7420.5780.670
/w FT0.8160.8920.6870.720
CLAM-SB [30]0.7420.8360.6240.702
/w FT0.8230.8620.6760.725
TransMIL [38]0.7480.8420.6570.706
/w FT0.7950.8570.6840.717
DTFD-MIL [45]0.7750.7990.5760.676
/w FT0.8040.8380.6890.717
", + "image_path": "f071be5fbe5a534fbb2a1ed38d5d6f348eeb3d60c9a454498b0cf80b4339f512.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 656, + 545, + 700 + ], + "lines": [ + { + "bbox": [ + 304, + 656, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 656, + 545, + 700 + ], + "type": "text", + "content": "Table 3. Generalization on Domain Shift. The generalization ability of all methods is compared between fine-tuning(FT) and IN-1K features on two datasets with domain shift. Camelyon-16-C and Camelyon-17 are synthetic and real corruptions respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7460" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "the paper length and a more clear comprehensive comparison can be found in Supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 156 + ], + "type": "text", + "content": "Comparison with baselines. The evaluation of generalization results is summarized in Table 3. For WSI model architecture, despite CLAM-SB [30], TransMIL [38], DTFDMIL [45], we also compare Max-pooling on feature level since it shows promising results on Camelyon-16 in 4.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": "The results demonstrate consistent improvements after equipping the WSI head with fine-tuned features under the F1 metric: In general, all models can resist domain shifts to a certain extent by utilizing the proposed fine-tuning. Compared with freezing IN-1K parameters, Max-pooling obtains a noticeable performance increase of " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "18.43\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "18.86\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": " on Camelyon-16-C and Camelyon-17, respectively. CLAM-SBachieves an increase of " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "10.91\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "8.33\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": " respectively. TransMIL achieves an increase of " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "6.28\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "4.11\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": " respectively. DTFD-MIL achieves an increase of " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "2.90\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "inline_equation", + "content": "11.30\\%" + }, + { + "bbox": [ + 46, + 156, + 286, + 322 + ], + "type": "text", + "content": " respectively. In contrast, the generalization of DTFD-MIL can still be further improved by our proposed FT optimization, even though it is already robust to different domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "text", + "content": "Interestingly, all improvements in performance under the AUC metric are much less than F1. The three Attention-based pooling methods respectively achieve the average improvement of " + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "inline_equation", + "content": "6.70\\%" + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "inline_equation", + "content": "7.91\\%" + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "text", + "content": " on Camelyon-16-C and Camelyon-17 under the F1 metric, as compared to " + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "inline_equation", + "content": "2.93\\%" + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "inline_equation", + "content": "2.98\\%" + }, + { + "bbox": [ + 46, + 323, + 286, + 491 + ], + "type": "text", + "content": " under AUC, which indicates that the classification ability of WSI-MIL model (measured by F1) may be much weaker than ranking (measured by AUC). However, doctors may much more care about F1 instead of AUC in clinical diagnosis from sensitivity and specificity perspectives. Most importantly, the classification threshold usually keeps fixed after deploying the model in the real world, which reflects the contribution of our method is more meaningful in practical applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 497, + 212, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 212, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 212, + 510 + ], + "type": "text", + "content": "4.4. Further Ablation Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 515, + 286, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 286, + 563 + ], + "type": "text", + "content": "Additional ablations are included in the Supplementary Materials, with the main focus on the effects of learning rate on the backbone, Number selection of Top-K, and Value selection of Lagrange multiplier, results of the 3 stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 570, + 226, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 226, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 226, + 582 + ], + "type": "text", + "content": "4.5. Interpretability and Visualization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 588, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 286, + 658 + ], + "type": "text", + "content": "Here, we further show the interpretability improvements with FT features. As shown in Figure 4, attention scores from CLAM-SB [30] were visualized as a heatmap to determine the ROI and interpret the important morphology used for diagnosis. Obviously, the model focuses more on tumors with FT features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "content": "In this work, we present a fine-tuning method for WSI classification under the weak supervision of slide-level la" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 386, + 150 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 386, + 150 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 386, + 150 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 386, + 150 + ], + "type": "image", + "image_path": "b211ae4ea0c3503d27702e6b6d7b236f40c72927c643e08ee51d54ff9c54a673.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 388, + 73, + 464, + 150 + ], + "blocks": [ + { + "bbox": [ + 388, + 73, + 464, + 150 + ], + "lines": [ + { + "bbox": [ + 388, + 73, + 464, + 150 + ], + "spans": [ + { + "bbox": [ + 388, + 73, + 464, + 150 + ], + "type": "image", + "image_path": "99b9450acde0b21e7790cf12353b56f38823a2511e9a8004c2a1ea5021447cd1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 466, + 73, + 542, + 150 + ], + "blocks": [ + { + "bbox": [ + 466, + 73, + 542, + 150 + ], + "lines": [ + { + "bbox": [ + 466, + 73, + 542, + 150 + ], + "spans": [ + { + "bbox": [ + 466, + 73, + 542, + 150 + ], + "type": "image", + "image_path": "20de2c06e6cfb7ba2f8d76ba52a51339dc310b9caaac6fc28a7a48a77b2372ba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 152, + 386, + 228 + ], + "blocks": [ + { + "bbox": [ + 309, + 152, + 386, + 228 + ], + "lines": [ + { + "bbox": [ + 309, + 152, + 386, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 152, + 386, + 228 + ], + "type": "image", + "image_path": "cc20e19c7581bb63c363320446a19eb1812a80c084ac873642276f76a7279f77.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 388, + 152, + 464, + 228 + ], + "blocks": [ + { + "bbox": [ + 388, + 152, + 464, + 228 + ], + "lines": [ + { + "bbox": [ + 388, + 152, + 464, + 228 + ], + "spans": [ + { + "bbox": [ + 388, + 152, + 464, + 228 + ], + "type": "image", + "image_path": "3ab4e29709964280adfead2dd5a0b2c9bfe7b591a9d60f845e66df2aad708758.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 466, + 152, + 542, + 228 + ], + "blocks": [ + { + "bbox": [ + 466, + 152, + 542, + 228 + ], + "lines": [ + { + "bbox": [ + 466, + 152, + 542, + 228 + ], + "spans": [ + { + "bbox": [ + 466, + 152, + 542, + 228 + ], + "type": "image", + "image_path": "25dcc4119d2965536db446e2d47396faba673305d449e36531fc62b8ea664020.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 309, + 230, + 386, + 306 + ], + "blocks": [ + { + "bbox": [ + 309, + 230, + 386, + 306 + ], + "lines": [ + { + "bbox": [ + 309, + 230, + 386, + 306 + ], + "spans": [ + { + "bbox": [ + 309, + 230, + 386, + 306 + ], + "type": "image", + "image_path": "221afd1a32da494642922b79bce36520da32aaa1d646130939b95b75cbac5f20.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 318, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 373 + ], + "type": "text", + "content": "Figure 4. Heatmap comparison between pretraining and FT. The first row shows the full annotations, where red contours denote the tumor area and green contours filter the nonsense background. The second and third row shows the WSI attention map on pretraining and FT features respectively." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 388, + 230, + 464, + 306 + ], + "blocks": [ + { + "bbox": [ + 388, + 230, + 464, + 306 + ], + "lines": [ + { + "bbox": [ + 388, + 230, + 464, + 306 + ], + "spans": [ + { + "bbox": [ + 388, + 230, + 464, + 306 + ], + "type": "image", + "image_path": "abb2288ea1e58537763e53aea0130eba9751189ec26d37fc4a9e6af666063e28.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 466, + 230, + 542, + 306 + ], + "blocks": [ + { + "bbox": [ + 466, + 230, + 542, + 306 + ], + "lines": [ + { + "bbox": [ + 466, + 230, + 542, + 306 + ], + "spans": [ + { + "bbox": [ + 466, + 230, + 542, + 306 + ], + "type": "image", + "image_path": "17a386dc455de9ced873f15ee0330f12b3411f430ecb56ec3dac28b1cfb74e8c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 387, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 614 + ], + "type": "text", + "content": "belts. Initially, an effective IB module is introduced to mitigate the training cost of Gigapixel WSI, which distills the oversized bag into a sparse one. Then the backbone of the instance is able to be trained end-to-end in the MIL framework by learning and making classifications on the distilled bag. Thus the WSI classification performance is improved through the retention of mainly task-specific information. In addition, SSL can be combined with the proposed framework for further improvement. Compared with fully supervised learning, our methods can achieve competitive accuracy by utilizing extremely weak WSI labels. Furthermore, our training scheme can introduce versatile training-time augmentations for better generalization on datasets with domain shift, which is an inevitable challenge for previous work. The experimental results reflect the advances of our method in both accuracy and generalization. Overall, Our proposed approach shows strong potential for MIL application in real-world pathology diagnosis, with better performance, faster convergence and annotation efficiency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 625, + 403, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 625, + 403, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 625, + 403, + 639 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 645, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 645, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 545, + 704 + ], + "type": "text", + "content": "This study was partially supported by the National Natural Science Foundation of China (Grant no. 92270108), Zhejiang Provincial Natural Science Foundation of China (Grant no. XHD23F0201), and the Research Center for Industries of the Future (RCIF) at Westlake University." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7461" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Alessandro Achille and Stefano Soatto. Information dropout: Learning optimal representations through noisy computation. IEEE transactions on pattern analysis and machine intelligence, 40(12):2897-2905, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Alexander A. Alemi, Ian Fischer, Joshua V. Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 234 + ], + "type": "text", + "content": "[3] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10(7):e0130140, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 267 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 345 + ], + "type": "text", + "content": "[5] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 412 + ], + "type": "text", + "content": "[6] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature medicine, 28(1):154-163, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 479 + ], + "type": "text", + "content": "[7] Gabriele Campanella, Matthew G Hanna, Luke Geneslaw, Allen Miraflor, Vitor Werneck Krauss Silva, Klaus J Busam, Edi Brogi, Victor E Reuter, David S Klimstra, and Thomas J Fuchs. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine, 25(8):1301-1309, 2019. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 479, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 287, + 523 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. CoRR, abs/2104.14294, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 524, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 287, + 557 + ], + "type": "text", + "content": "[9] Richard J. Chen and et al. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In CVPR, pages 16144-16155, June 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[10] Richard J Chen and Rahul G Krishnan. Self-supervised vision transformers learn visual concepts in histopathology. arXiv preprint arXiv:2203.00585, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2018. cite arxiv:1810.04805Comment: 13 pages. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "text", + "content": "[13] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Littjens, Jeroen A. W. M. van der Laak, , and the CAME-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "LYON16 Consortium. Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA, 318(22):2199-2210, 12 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "text", + "content": "[14] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30(4):681–694, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 198 + ], + "type": "text", + "content": "[15] Ziv Goldfeld, Ewout van den Berg, Kristjan Greenewald, Igor Melnyk, Nam Nguyen, Brian Kingsbury, and Yury Polyanskiy. Estimating information flow in deep neural networks. arXiv preprint arXiv:1810.05728, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 199, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 232 + ], + "type": "text", + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 234, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 288 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 290, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 290, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 545, + 334 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 335, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 402 + ], + "type": "text", + "content": "[19] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 2127-2136. PMLR, 10-15 Jul 2018. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 403, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 545, + 448 + ], + "type": "text", + "content": "[20] Syed Ashar Javed, Dinkar Juyal, Zahl Shanis, Shreya Chakraborty, Harsha Pokkalla, and Aaditya Prakash. Rethinking machine learning model evaluation in pathology. arXiv preprint arXiv:2204.05205, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "text", + "content": "[21] Durk P Kingma, Tim Salimans, and Max Welling. Variational dropout and the local reparameterization trick. Advances in neural information processing systems, 28, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 495, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 495, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 495, + 545, + 517 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 518, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 518, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 518, + 545, + 552 + ], + "type": "text", + "content": "[23] Erich Leo Lehmann and Henry Scheffé. Completeness, similar regions, and unbiased estimation-part i. In Selected works of EL Lehmann, pages 233-268. Springer, 2012. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "text", + "content": "[24] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 588, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 643 + ], + "type": "text", + "content": "[25] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318-14328, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "text", + "content": "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[27] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7462" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 183 + ], + "type": "text", + "content": "[29] Geert Litjens, Peter Bandi, Babak Ehteshami Bejnordi, Oscar Geessink, Maschenka Balkenhol, Peter Bult, Altuna Halilovic, Meyke Hermsen, Rob van de Loo, Rob Vogels, et al. 1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset. *GigaScience*, 7(6):giy065, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 237 + ], + "type": "text", + "content": "[30] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 271 + ], + "type": "text", + "content": "[31] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 348 + ], + "type": "text", + "content": "[32] Bhargavi Paranjape, Mandar Joshi, John Thickstun, Hannaneh Hajishirzi, and Luke Zettlemoyer. An information bottleneck approach for controlling conciseness in rationale extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1938-1952, Online, Nov. 2020. Association for Computational Linguistics. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 350, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 436 + ], + "type": "text", + "content": "[33] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 437, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 524 + ], + "type": "text", + "content": "[34] Nicholas A Petrick, Shazia Akbar, Kenny HH Cha, Sharon Nofech-Mozes, Berkman Sahiner, Marios A Gavrielides, Jayashree Kalpathy-Cramer, Karen Drukker, Anne LL Martel, et al. Spie-aapm-nci breastpathq challenge: an image analysis challenge for quantitative tumor cellularity assessment in breast cancer histology images following neoadjuvant treatment. Journal of Medical Imaging, 8(3):034501, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "text", + "content": "[35] Andrew M Saxe, Yamini Bansal, Joel Dapello, Madhu Advani, Artemy Kolchinsky, Brendan D Tracey, and David D Cox. On the information bottleneck theory of deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124020, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 581, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 613 + ], + "type": "text", + "content": "[36] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. arXiv preprint arXiv:2001.00396, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 614, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 678 + ], + "type": "text", + "content": "[37] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[38] Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, and yongbing zhang. Transmil: Transformer based correlated multiple instance learning for whole" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 675 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 127 + ], + "type": "text", + "content": "slide image classification. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2136-2147. Curran Associates, Inc., 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[39] Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information. arXiv preprint arXiv:1703.00810, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[40] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 197, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 272 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 307 + ], + "type": "text", + "content": "[42] Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 308, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 351 + ], + "type": "text", + "content": "[43] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[44] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. Xlnet: Generalized autoregressive pretraining for language understanding. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 398, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 452 + ], + "type": "text", + "content": "[45] Hongrun Zhang, Yanda Meng, Yitian Zhao, Yihong Qiao, Xiaoyun Yang, Sarah E. Coupland, and Yalin Zheng. Dtfmil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification. ArXiv, abs/2203.12081, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 454, + 545, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 519 + ], + "type": "text", + "content": "[46] Yunlong Zhang, Yuxuan Sun, Honglin Li, Sunyi Zheng, Chenglu Zhu, and Lin Yang. Benchmarking the robustness of deep neural networks to common corruptions in digital pathology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "text", + "content": "[47] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdolahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 565, + 545, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 619 + ], + "type": "text", + "content": "[48] Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9140-9150, June 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 621, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 621, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 307, + 621, + 545, + 675 + ], + "type": "text", + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2016. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7463" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_content_list.json b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6d227add13d9c32b6e18f23f37828ff4426888ed --- /dev/null +++ b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_content_list.json @@ -0,0 +1,1565 @@ +[ + { + "type": "text", + "text": "TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation", + "text_level": 1, + "bbox": [ + 99, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Devavrat Tomar1 Guillaume Vray1 Behzad Bozorgtabar1,2 Jean-Philippe Thiran1,2 EPFL CHUV", + "bbox": [ + 130, + 179, + 836, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{$ firstname\\}. $\\{$ lastname\\}@epfl.ch", + "bbox": [ + 346, + 217, + 614, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 310, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with automatic Adversarial augmentation dubbed TeSLA for adapting a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning methods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowledge distillation by simulating high entropy augmented images. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several desirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available at https://github.com/devavratTomar/TeSLA.", + "bbox": [ + 75, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 207, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) perform exceptionally well when the training (source) and test (target) data follow the same distribution. However, distribution shifts are inevitable in real-world settings and propose a major challenge to the performance of deep networks after deployment. Also, access to the labeled training data may be infeasible at test time due to privacy concerns or transmission bandwidth. In such scenarios, source-free domain adaptation (SFDA) [1, 23, 25] and test-time adaptation (TTA) methods [19, 28, 39] aim to adapt the pre-trained source model to the unlabeled distributionally shifted target domain while easing access to source data. While SFDA methods have access to all full target data through multiple training epochs", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/606c4613313034c2fb162c2724bbb58c1b37abc42126bae8d18860dca17451d4.jpg", + "image_caption": [ + "Figure 1. Knowledge distillation with adversarial augmentations. (a) Easy images with confident soft-pseudo labels and (b) Hard images with unconfident soft-pseudo labels are adversarily augmented and pushed to the uncertainty region (high entropy) near the decision boundary. The model is updated for (a) to match its output on the augmented views with non-augmented views of Easy test images using KL-Divergence $\\mathcal{L}_{\\mathrm{kd}} \\neq 0$ , while not updated for (b) as $\\mathcal{L}_{\\mathrm{kd}} \\sim 0$ between Hard images and their augmented views." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 645, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/792aacf96515bab2dd2ea598a8a614d2dd2a04833e3da270cdc529d583ed32b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 267, + 808, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76aca1b51c9e128429a15268ad368e38b13e9cc94b8f01981ddd8c2d8351a6c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 267, + 892, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(offline setup), TTA methods usually process test images in an online streaming fashion and represent a more realistic domain adaptation. However, most of these methods are applied: (i) only to classification tasks, (ii) evaluated on the non-real-world domain shifts, e.g., the non-measurement shift; (iii) destroy model calibration—entropy minimizing with overconfident predictions [45] on incorrectly classified samples, and (iv) use specialized network architectures or rely on the source dataset feature statistics [28].", + "bbox": [ + 496, + 506, + 892, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address these issues by proposing a new test-time adaptation method with automatic adversarial augmentation called TeSLA, under which we further define realistic TTA protocols. Self-learning methods often supervise the model adaptation on the unlabeled test images using their predicted pseudo-labels. As the model can easily overfit on its own pseudo-labels, a weight-averaged teacher model (slowly updated by the student model) is employed for obtaining the pseudo-labels [41, 47]. The student model is then trained with cross-entropy $(\\mathbb{C}\\mathbb{E})$ loss between the one-hot pseudolabels and its predictions on the test images. In this paper, we instead propose to minimize flipped cross-entropy between the student model's predictions and the soft pseudo-labels (notice the reverse order) with the negative entropy of its marginalized predictions over the test images. In Sec. 3, we show that the proposed formulation is an equivalence to mutual information maximization implicitly corrected", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by the teacher-student knowledge distillation via pseudolabels, yielding performance improvement on various test-time adaptation protocols and compared to the basic $\\mathbb{C}\\mathbb{E}$ optimization (cf. ablation in Fig. 5-(1)).", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by teacher-student knowledge distillation, another central tenet of our method is to assist the student model during adaptation in improving its performance on the hard-to-classify (high entropy) test images. For this purpose, we propose learning automatic adversarial augmentations (see Fig. 1) as a proxy for simulating images in the uncertainty region of the feature space. The model is then updated to ensure consistency between predictions of high entropy augmented images and soft-pseudo labels from the respective non-augmented versions. Consequently, the model is self-distilled on Easy test images with confident soft-pseudo labels (Fig. 1a). In contrast, the model update on Hard test images is discarded (Fig. 1b), resulting in better class-wise feature separation.", + "bbox": [ + 75, + 152, + 470, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (i) we propose a novel test-time self-learning method based on flipped cross-entropy $(f\\text{-}\\mathbb{C}\\mathbb{E})$ through the tight connection with the mutual information between the model's predictions and the test images; (ii) we propose an efficient plug-in test-time automatic adversarial augmentation module used for online knowledge distillation from the teacher to the student network that consistently improves the performance of test-time adaptation methods, including ours; and (iii) TeSLA achieves new state-of-the-art results on several benchmarks, from common image corruption to realistic measurement shifts for classification and segmentation tasks. Furthermore, TeSLA outperforms existing TTA methods in terms of calibration and uncertainty metrics while making no assumptions about the network architecture and source domain's information, e.g., feature statistics or training strategy.", + "bbox": [ + 75, + 363, + 470, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 618, + 218, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, domain adaptation methods aim to distill knowledge from source data that are well-generalizable to target data and relax the assumption of i.i.d. between source and target datasets. To circumvent expensive and cumbersome annotation of new target data, unsupervised domain adaptation (UDA) emerges, and there has been a large corpus of UDAs [13, 17, 32, 51] to match the distribution of domains on both labeled source data and unlabeled target data. Nonetheless, the above-mentioned methods demand source data to achieve the domain adaptation process, which is often impractical in real-world scenarios, e.g., due to privacy restrictions. The above issue motivates research into source-free domain adaptation (SFDA) [1, 23, 25, 26] and test-time training or adaptation (TTA) [19, 28, 39, 45], which are more closely relevant to our problem setup.", + "bbox": [ + 75, + 643, + 468, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SFDA approaches formulate domain adaptation through pseudo labeling [23, 26], target feature clustering [49], syn", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thesizing extra training samples [24], or feature restoration [12]. SHOT [26] proposed feature clustering via information maximization while incorporating pseudo-labeling as additional supervision. BAIT [48] leverages the fixed source classifier as source anchors and uses them to achieve feature alignment between the source and target domain. CPGA [35] proposed the SFDA method based on matching feature prototypes between source and target domains. AaD-SFDA [50] proposed to optimize an objective by encouraging prediction consistency of local neighbors in feature space. Nevertheless, SFDA methods require apriori access to all target data in advance, and the current pseudo-labeling approach, e.g., SHOT [26], used offline pseudo-label refinement on a perepoch basis. In a more realistic domain adaptation scenario, SFDA is still incompetent to perform inference and adaptation simultaneously.", + "bbox": [ + 496, + 90, + 893, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TTA methods [4, 45] propose alleviating the domina shift by online (or streaming) adapting a model at test time. Still, we argue that there are ambiguities over the problem setup of TTA in the literature, particularly on whether sequential inference on target test data is feasible upon arrival [19, 45] or whether training objectives must be adapted [28, 40]. TTT [40] proposed fine-tuning the model parameters via rotation classification task as a proxy. On-target adaptation [44] used pseudo-labeling and contrastive learning to initialize the target-domain feature; each performed independently in their method. T3A [19] utilized pseudo-labeling to adjust the classifier prototype. More recently, TTAC [39] leveraged the clustering scheme to match target domain clusters to source domain categories. Nevertheless, existing TTA methods rely on specialized neural network architectures or only update a fraction of model weights yielding limited performance gain on the target data. For instance, TENT [45] proposed updating affine parameters in the batchnorm layers of convolutional neural networks (CNN), while AdaContrast [4] and SHOT [26] used an additional weight normalization classification layer with a projection head. In contrast, we show our method's superiority for various neural network architectures, including CNNs and vision transformers (ViTs) [11], without additional architectural requirements or different source model training strategies. Moreover, we update all model parameters, and our method is stable over a wide range of hyper-parameters, e.g., learning rate.", + "bbox": [ + 496, + 337, + 893, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test time augmentation methods [2, 29] are another popular line of domain adaptation research. GPS [29] learns optimal augmentation sub-policies by combining image transformations of RandAugment [10] that minimize calibrated log-likelihood loss [2] on the validation set. OptTTA [42] optimized the magnitudes of augmentation sub-policies using gradient descent to maximize mutual information and match feature statistics over augmented images. Though effective, these methods [29, 42, 46] are computationally expensive as all augmentation sub-policies need to be evaluated, making", + "bbox": [ + 496, + 750, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20342", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "their real-time application difficult. Instead, our proposed adversarial augmentation policies can be learned online and are several orders faster than the existing learnable test-time augmentation strategies, e.g., [42].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 157, + 210, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. We first introduce our flipped cross-entropy $(f\\text{-}\\mathbb{C}\\mathbb{E})$ loss through the tight connection with the mutual information between the model's predictions and the test images. (Sec. 3.1). Using this equivalence, we derive our test-time loss function that implicitly incorporates teacher-student knowledge distillation. In Sec. 3.2, we propose to enhance the test-time teacher-student knowledge distillation by utilizing the consistency of the student model's predictions on the proposed adversarial augmentations with their corresponding refined soft-pseudo labels from the teacher model. We refine the soft-pseudo labels by averaging the teacher model's predictions on (1) weakly augmented test images; (2) nearest neighbors in the feature space. Furthermore, we propose an efficient online algorithm for learning adversarial augmentations (Sec. 3.3). Finally, in Sec. 3.4, we summarize our online test-time adaptation method, TeSLA, based on self-learning with the proposed automatic adversarial augmentation. The overall framework is shown in Fig. 2.", + "bbox": [ + 75, + 183, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setup. Given an existing model $f_{\\theta_0}$ , parametrized by $\\theta_0$ and pre-trained on the inaccessible source data, we aim to improve its performance by updating all its parameters using the unlabeled streaming test data $\\mathbf{X}$ during adaptation. Motivated by the concept of mean teacher [41] as the weight-averaged model over training steps that can yield a more accurate model, we build our online knowledge distillation framework upon teacher-student models. We utilize identical network architecture for teacher and student models. Each model $f = h \\circ g$ comprises a backbone encoder $g: \\mathbf{X} \\to \\mathbb{R}^D$ mapping unlabeled test image $\\mathbf{x} \\in \\mathbf{X}$ to the feature representation $\\mathbf{z} = g(\\mathbf{x}) \\in \\mathbb{R}^D$ and a classifier (hypothesis) head $h: \\mathbb{R}^D \\to \\mathbb{R}^K$ mapping $\\mathbf{z}$ to the class prediction $\\mathbf{y} \\in \\mathbb{R}^K$ , where $D$ and $K$ denote the feature dimension and number of classes. The parameters of teacher $\\theta_t$ and student $\\theta_s$ are first initialized from the source model parameters $\\theta_0$ . Then the teacher model's parameters $\\theta_t$ are updated from the student model's parameters $\\theta_s$ using an exponential moving average (EMA) with momentum coefficient $\\alpha$ as: $\\theta_t \\gets \\alpha \\cdot \\theta_t + (1 - \\alpha) \\cdot \\theta_s$ . Following [26], we freeze the classifier $h$ 's parameters and only update the encoder $g$ 's parameters during test-time adaptation.", + "bbox": [ + 75, + 460, + 472, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rationale Behind the $f$ -CE Objective", + "text_level": 1, + "bbox": [ + 76, + 801, + 397, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by analyzing the rationale behind the proposed $f$ -CE loss and its benefit for self-learning through the tight connection with the mutual information between the model's predictions and unlabeled test images. Before that, we first define the notations.", + "bbox": [ + 73, + 825, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notations. We define the random variables of unlabeled test images and the predictions from the student model $f_{s}$ as $\\mathbf{X}$ and $\\mathbf{Y}$ and those of the teacher model $f_{t}$ 's soft pseudo-label predictions as $\\hat{\\mathbf{Y}}$ . Furthermore, let $p_{\\mathbf{Y}}$ denotes the marginal distribution over $\\mathbf{Y}$ , $p_{(\\mathbf{Y},\\mathbf{X})}$ be the joint distribution of $\\mathbf{Y}$ and $\\mathbf{X}$ , and $p_{\\mathbf{Y}|\\mathbf{X}}$ be the conditional distribution of $\\mathbf{Y}$ given $\\mathbf{X}$ . Then, the entropy of $\\mathbf{Y}$ and the conditional entropy of $\\mathbf{Y}$ given $\\mathbf{X}$ can be defined as $\\mathcal{H}(\\mathbf{Y}) := \\mathbb{E}_{p_{\\mathbf{Y}}}[-\\log p_{\\mathbf{Y}}(\\mathbf{Y})]$ and $\\mathcal{H}(\\mathbf{Y}|\\mathbf{X}) := \\mathbb{E}_{p_{(\\mathbf{Y},\\mathbf{X})}}[-\\log p_{\\mathbf{Y}|\\mathbf{X}}(\\mathbf{Y}|\\mathbf{X})]$ , respectively. Besides, let the flipped cross-entropy ( $f-\\mathbb{C}\\mathbb{E}$ ) given $\\mathbf{X}$ be $\\mathcal{H}(\\mathbf{Y};\\hat{\\mathbf{Y}}|\\mathbf{X}) := \\mathbb{E}_{p_{(\\mathbf{Y},\\mathbf{X})}}[-\\log p_{\\hat{\\mathbf{Y}}|\\mathbf{X}}(\\mathbf{Y}|\\mathbf{X})]$ .", + "bbox": [ + 496, + 90, + 893, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-learning and mutual information. Based on the above definitions, the $f$ - $\\mathbb{C}\\mathbb{E}$ between $\\mathbf{Y}$ and $\\hat{\\mathbf{Y}}$ conditioned over $\\mathbf{X}$ has the tight connection with the mutual information $\\mathcal{I}(\\mathbf{Y};\\mathbf{X})$ between unlabeled test images $\\mathbf{X}$ and the predictions from the student model $\\mathbf{Y}$ as follows:", + "bbox": [ + 496, + 270, + 893, + 347 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {H} (\\mathbf {Y}; \\hat {\\mathbf {Y}} \\mid \\mathbf {X}) = \\mathcal {H} (\\mathbf {Y} \\mid \\mathbf {X}) + \\mathcal {D} _ {\\mathrm {K L}} (\\mathbf {Y} \\| \\hat {\\mathbf {Y}} \\mid \\mathbf {X}) \\tag {1} \\\\ = - \\mathcal {I} (\\mathbf {Y}; \\mathbf {X}) + \\mathcal {H} (\\mathbf {Y}) + \\mathcal {D} _ {\\mathrm {K L}} \\left(\\mathbf {Y} \\| \\hat {\\mathbf {Y}} | \\mathbf {X}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 369, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This implies that minimizing the $f$ -CE $(\\mathcal{H}(\\mathbf{Y};\\hat{\\mathbf{Y}}\\mid \\mathbf{X})$ in Eq. 1) and maximizing the entropy of class-marginal prediction $\\mathcal{H}(\\mathbf{Y})$ is equivalent to maximizing the mutual information between the test images and the student model's predictions $\\mathcal{I}(\\mathbf{Y};\\mathbf{X})$ with a correction KL-divergence term $\\mathcal{D}_{\\mathrm{KL}}(\\mathbf{Y}\\parallel \\hat{\\mathbf{Y}}\\mid \\mathbf{X})$ involving soft-pseudo labels as:", + "bbox": [ + 496, + 436, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\mathcal {H} (\\mathbf {Y} ; \\hat {\\mathbf {Y}} \\mid \\mathbf {X})} _ {f \\cdot \\mathbb {C E}} - \\mathcal {H} (\\mathbf {Y}) = - \\underbrace {\\mathcal {I} (\\mathbf {Y} ; \\mathbf {X})} _ {\\text {M u t u a l I n f o .}} + \\mathcal {D} _ {\\mathrm {K L}} (\\mathbf {Y} \\| \\hat {\\mathbf {Y}} | \\mathbf {X}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 561, + 900, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, minimizing the left side of Eq. 2 with respect to the student model's parameters $\\theta_{s}$ allows the student model to cluster test images using mutual information and criterion-corrected knowledge distillation from soft-pseudo labels. Using the above formulation, we define the following test-time objective to train the student model $f_{s}$ on a batch of $B$ test images $X = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_B\\}$ with their corresponding soft-pseudo labels from teacher model $\\hat{Y} = \\{\\hat{y}_1,\\dots ,\\hat{y}_B\\}$ .", + "bbox": [ + 496, + 619, + 893, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p l}} (X, \\hat {Y}) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {k = 1} ^ {K} f _ {s} \\left(\\boldsymbol {x} _ {i}\\right) _ {k} \\log \\left(\\left(\\hat {\\boldsymbol {y}} _ {i}\\right) _ {k}\\right) \\\\ + \\sum_ {k = 1} ^ {K} \\hat {f} _ {s} (X) _ {k} \\log (\\hat {f} _ {s} (X)) _ {k} \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 768, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{f}_s(X) = \\frac{1}{B}\\sum_{i=1}^B f_s(\\pmb{x}_i)$ denotes the marginal class distribution over the batch of test images $X$ .", + "bbox": [ + 496, + 867, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d368dff16a867de483049f5c93f9de3352026a3a91076cbc03a38f08811b65f8.jpg", + "image_caption": [ + "Figure 2. Overview of TeSLA Framework. (a) The student model $f_{s}$ is adapted on the test images by minimizing the proposed test-time objective $\\mathcal{L}_{\\mathrm{pl}}$ . The high-quality soft-pseudo labels required by $\\mathcal{L}_{\\mathrm{pl}}$ are obtained from the exponentially weighted averaged teacher model $f_{t}$ and refined using the proposed Soft-Pseudo Label Refinement (PLR) on the corresponding test images. The soft-pseudo labels are further utilized for teacher-student knowledge distillation via $\\mathcal{L}_{\\mathrm{kd}}$ on the adversarially augmented views of the test images. (b) The adversarial augmentations are obtained by applying learned sub-policies sampled i.i.d from $\\mathbb{P}$ using the probability distribution $P$ with their corresponding magnitudes selected from $M$ . The parameters $M$ and $P$ of the augmentation module are updated by the unbiased gradient estimator (Eq. 13) of the loss $\\mathcal{L}_{\\mathrm{Aug}}$ computed on the augmented test images." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 872, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Knowledge Distillation via Augmentation", + "text_level": 1, + "bbox": [ + 76, + 353, + 429, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-learning from adversarial augmentations. As illustrated in Fig. 1, adversarial augmentations are learned by pushing their feature representations toward the decision boundary (Expansion phase), followed by updating the student model $f_{s}$ to match its prediction on these augmented images with their corresponding soft-pseudo label (Separation phase), yielding better separation of features into their respective classes. In our method setup, we continually learn automatic augmentations of the test images (Sec. 3.3) that are adversarial to the current teacher model $f_{t}$ and enforce consistency between the student model's $f_{s}$ predictions on the augmented views and the corresponding soft-pseudo labels. Since we freeze the classifier module, $f_{s}$ and $f_{t}$ share the common decision boundary. Let $\\tilde{\\boldsymbol{x}}$ denote the learned adversarial augmented view of $\\boldsymbol{x}$ and $\\hat{\\boldsymbol{y}}$ denote the corresponding soft-pseudo label. We minimize the following knowledge distillation loss $\\mathcal{L}_{\\mathrm{kd}}$ using KL-divergence to distill knowledge from the teacher model to the student model:", + "bbox": [ + 75, + 376, + 473, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k d}} (\\tilde {\\boldsymbol {x}}, \\hat {\\boldsymbol {y}}) = \\mathcal {D} _ {\\mathrm {K L}} (\\hat {\\boldsymbol {y}} \\| f _ {s} (\\tilde {\\boldsymbol {x}})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 664, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Soft pseudo label refinement (PLR). We refine the quality of soft pseudo-labels and the feature representations by averaging the teacher model's outputs on multiple weakly augmented image views $\\rho_w(\\pmb{x})$ (via FLIPPING, CROPPING augmentations) of the same test image $\\pmb{x}$ as follows:", + "bbox": [ + 75, + 694, + 470, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t}, \\mathbf {y} _ {t} \\leftarrow \\mathbb {E} _ {\\mathbf {u} \\in \\rho_ {w} (\\boldsymbol {x})} \\left[ g _ {t} (\\mathbf {u}), h _ {t} \\left(g _ {t} (\\mathbf {u})\\right) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 785, + 468, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The refined feature representations $\\mathbf{z}_t$ and the softmaxed pseudo labels $\\mathbf{y}_t$ from the encoder $g_t$ and the classifier $h_t$ are further stored in an online memory queue $\\mathbf{Q}$ of fixed size. The final soft pseudo-label is computed by averaging the refined soft pseudo-labels of $n$ -nearest neighbors of the current test image in the feature space as follows:", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Q} [ \\arg \\max (\\boldsymbol {y} _ {t}) ]. \\text {a p p e n d} \\left(\\left\\{\\boldsymbol {z} _ {t}, \\boldsymbol {y} _ {t} \\right\\}\\right) \\\\ \\hat {\\boldsymbol {y}} = \\frac {1}{n} \\sum \\mathcal {N} _ {\\mathbf {Q}, n} (\\mathbf {z} _ {t}) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 366, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}$ denotes an online memory of class-balanced queues (with size $|\\mathbf{Q}[\\arg \\max (\\pmb {y}_t)]| \\leq N_{\\mathbf{Q}}$ ) of the refined feature representations and the softmaxed label predictions on the previously seen test images, and $\\mathcal{N}_{\\mathbf{Q},n}(\\mathbf{z}_t)$ denotes soft labels of $n$ -nearest neighbors of $\\mathbf{z}_t$ from $\\mathbf{Q}$ .", + "bbox": [ + 496, + 421, + 893, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learning Adversarial Data Augmentation", + "text_level": 1, + "bbox": [ + 500, + 507, + 856, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section presents our efficient online method for learning adversarial data augmentation for the current teacher $f_{t}$ . We first introduce our adversarial augmentation search space. Then, we describe our differentiable strategy to optimize and sample adversarial augmentations that push the feature representations of the test images toward the uncertain region in the close vicinity of the decision boundary.", + "bbox": [ + 496, + 532, + 893, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Policy search space. Let $\\mathbb{O}$ be a set of all image transformation operations $\\mathcal{O}:\\mathbb{R}^{H\\times W\\times 3}\\to \\mathbb{R}^{H\\times W\\times 3}$ defined in our search space. In particular, we use image transformations AUTO-CONTRAST, EQUALIZE, INVERT, SOLARIZE, POSTERIZE, CONTRAST, BRIGHTNESS, COLOR, SHEARX, SHEARY, TRANSLATEX, TRANSLATEY, ROTATE, AND SHARPNESS. Each transformation $\\mathcal{O}$ is specified with its magnitude parameter $m\\in [0,1]$ . Since the output of some image operations may not be conditional on its magnitude (e.g., EQUALIZE) or may not be differentiable (e.g., POSTERIZE), for such image operations, we use the straight-through gradient estimate w.r.t. each pixel as: $\\partial \\mathcal{O}(\\boldsymbol{x}_{ij}) / \\partial m = 1$ for magnitude optimization.", + "bbox": [ + 496, + 643, + 893, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define a sub-policy $\\rho$ as a combination of $N$ image operations (sub-policy dimension) from $\\mathbb{O}$ that is applied sequentially to a given image $\\pmb{x}$ as:", + "bbox": [ + 496, + 825, + 893, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\left(\\boldsymbol {x}, m ^ {\\rho}\\right) = \\mathcal {O} _ {1} ^ {\\rho} \\dots \\mathcal {O} _ {N} ^ {\\rho} \\left(\\boldsymbol {x}, m _ {N} ^ {\\rho}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 883, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20344", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m^{\\rho} = [m_{1}^{\\rho},\\dots,m_{N}^{\\rho}]$ denotes the magnitude set of the sub-policy $\\rho$ . We also define $\\mathbb{P} = \\{\\rho_1,\\rho_2,\\ldots \\}$ as the set of all possible sub-policies.", + "bbox": [ + 75, + 90, + 472, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy evaluation. We propose learnable adversarial augmentation, aiming to optimize and sample data augmentations as a proxy for simulating data in the uncertain region of the feature space. Given the teacher model $f_{t}$ , a sub-policy $\\rho$ with magnitude $m$ is evaluated for a test image $x$ using the following adversarial objective:", + "bbox": [ + 75, + 142, + 472, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho) = \\sum_ {k = 1} ^ {K} f _ {t} (\\tilde {\\boldsymbol {x}}) \\log \\left(f _ {t} (\\tilde {\\boldsymbol {x}})\\right) + \\lambda_ {1} r (\\tilde {\\boldsymbol {x}}, \\boldsymbol {x}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 256, + 470, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\pmb{x}} = \\rho (\\pmb {x},m)$ denotes the adversarial augmented image and $r(\\cdot ,\\cdot)$ regularizes augmentation severity. The hyperparameter $\\lambda_{1}$ controls the augmentation severity. We optimize the adversarial augmentation loss $\\mathcal{L}_{\\mathrm{aug}}$ with respect to the sub-policy $\\rho$ 's parameters. Minimizing $\\mathcal{L}_{\\mathrm{aug}}$ is equivalent to maximizing the entropy of the teacher model's prediction on the augmented image $\\tilde{\\pmb{x}}$ thus pushing its feature representation towards the decision boundary (first term). For the augmentation regularization (second term), we use the mean squared distance function between the teacher encoder's $L$ internal layers' activations of the adversarial augmentation and non-augmented versions of image $\\pmb{x}$ defined as:", + "bbox": [ + 75, + 309, + 472, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (\\tilde {\\boldsymbol {x}}, \\boldsymbol {x}) = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\| \\mu_ {l} (\\tilde {\\boldsymbol {x}}) - \\mu_ {l} (\\boldsymbol {x}) \\| ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 502, + 470, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_l(\\cdot)$ denotes the mean activation of the $l^{th}$ layer of the teacher encoder $g_t$ .", + "bbox": [ + 75, + 550, + 470, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy optimization. Let $M = [m^{\\rho_1}, \\ldots, m^{\\rho_{\\| \\mathbb{P} \\|}}]$ denote the set of magnitudes $m^{\\rho}$ of all sub-policies $\\rho \\in \\mathbb{P}$ and $P = [p_1, \\ldots, p_{\\| \\mathbb{P} \\|}]$ denote the probability of selecting a sub-policy from $\\mathbb{P}$ . The expected policy evaluation loss (Eq. 8) for an image $x$ over the policy search space is given by:", + "bbox": [ + 76, + 585, + 472, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}) \\right] = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} p _ {i} \\cdot \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 672, + 470, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluating the gradient of $\\mathbb{E}[\\mathcal{L}_{\\mathrm{aug}}(\\pmb{x})]$ w.r.t. $M$ and $P$ can become computationally expensive as $\\| \\mathbb{P}\\| \\sim \\binom{\\|\\mathbb{O}\\|}{N}$ . Thus, we use the following re-parameterization trick to estimate its unbiased gradient:", + "bbox": [ + 75, + 722, + 470, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla \\mathbb {E} [ \\mathcal {L} _ {\\mathrm {a u g}} (x) ] = \\delta (\\boldsymbol {x}, \\mathbb {P}) = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} \\nabla \\left(p _ {i} \\cdot \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i})\\right) \\tag {11} \\\\ = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} p _ {i} \\left(\\nabla \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) + \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\cdot \\nabla \\log p _ {i}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 809, + 470, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, an unbiased estimator of $\\delta (\\pmb {x},\\mathbb{P})$ can be written as:", + "bbox": [ + 500, + 90, + 877, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\delta} (\\boldsymbol {x}, \\rho_ {i}) = \\nabla \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) + \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\cdot \\nabla \\log p _ {i} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 118, + 893, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where index $i$ is sampled from the probability distribution $P$ . For an online batch of $B$ test images $\\{\\pmb{x}_1, \\pmb{x}_2, \\dots, \\pmb{x}_B\\}$ , we apply augmentation sub-policies $\\{\\rho_{i_1}, \\rho_{i_2}, \\dots, \\rho_{i_B}\\}$ from $\\mathbb{P}$ where $\\{i_1, i_2, \\dots, i_B\\}$ are sampled i.i.d from distribution $P$ , and update the parameters $P$ and $M$ with the following stochastic gradient update rule:", + "bbox": [ + 498, + 148, + 893, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ P, M ] \\leftarrow [ P, M ] - \\frac {\\gamma}{B} \\sum_ {j = 1} ^ {B} \\hat {\\delta} \\left(\\boldsymbol {x} _ {j}, \\rho_ {i _ {j}}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 251, + 892, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the learning rate. We set $\\gamma = 0.1$ for all experiments without the need for hyperparameter tuning.", + "bbox": [ + 498, + 305, + 893, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Self-Learning With Adversarial Augmentation", + "text_level": 1, + "bbox": [ + 500, + 345, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We minimize the following overall objective $\\mathcal{L}_{\\mathrm{TeSLA}}$ for training the student model $f_{s}$ on a batch of $B$ test images $X = \\{\\pmb {x}_1,\\dots ,\\pmb {x}_B\\}$ and their adversarial augmented views $\\hat{X} = \\{\\tilde{x}_1,\\dots ,\\tilde{x}_B\\}$ with the corresponding refined softpseudo labels from the teacher model $\\hat{Y} = \\{\\hat{y}_1,\\dots ,\\hat{y}_B\\}$ :", + "bbox": [ + 500, + 368, + 893, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T e S L A}} (X, \\tilde {X}, \\hat {Y}) = \\mathcal {L} _ {\\mathrm {p l}} (X, \\hat {Y}) + \\frac {\\lambda_ {2}}{B} \\sum_ {i = 1} ^ {B} \\mathcal {L} _ {\\mathrm {k d}} (\\tilde {\\boldsymbol {x}} _ {i}, \\hat {\\boldsymbol {y}} _ {i}) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 468, + 892, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{2}$ is a hyper-parameter for the knowledge distillation. The adversarial augmented views $\\tilde{X}$ are obtained by sampling i.i.d. augmentation sub-policies from $\\mathbb{P}$ for every test image in $X$ with probability $P$ and applying the corresponding magnitude from $M$ . Concurrently, we also optimize $M$ and $P$ using the policy optimization mentioned in Sec. 3.3.", + "bbox": [ + 498, + 520, + 893, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 625, + 633, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 650, + 808, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate and compare TeSLA against state-of-the-art (SOTA) test-time adaptation algorithms for both classification and segmentation tasks under three types of test-time distribution shifts resulting from (1) common image corruption, (2) synthetic to real data transfer, and (3) measurement shifts on medical images. The latter is characterized by a change in medical imaging systems, e.g., different scanners across hospitals or various staining techniques.", + "bbox": [ + 498, + 674, + 893, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TTA protocols. We adopt the TTA protocols in [39] and categorize competing methods based on two factors: (i) source training objective and (ii) sequential or non-sequential inference. First, unlike [39], we use $\\mathbf{Y}$ to indicate if access to the source domain's information, e.g., feature statistics is possible or if the source training objective is allowed to be modified; otherwise, we use $\\mathbf{N}$ . Next, we use $\\mathbf{O}$ to indicate a", + "bbox": [ + 498, + 795, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20345", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "one-pass adaptation and evaluation protocol (one epoch) for sequential test data and $\\mathbf{M}$ to show a multi-pass adaptation on all test data and inference after several epochs. Thus, we end up with four possible TTA protocols, namely N-M, Y-M, N-O, and Y-O. Unlike previous works [28,39], TeSLA does not rely on any source domain feature distribution, yielding two realistic TTA protocols: N-M and N-O.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameters for all experiments. We set $\\lambda_1 = \\lambda_2 = 1.0$ for all experiments (cf. ablation in Appendix E). For the adversarial augmentation module, we use sub-policy dimension $N = 2$ (except for VisDA-C [33], $N = 4/3$ for N-O/N-M protocols) (cf. Appendix E) and Adam optimizer [22]. For pseudo-label refinement (PLR), we use five weak augmentations composed of random flips and resize crop. More details of hyperparameters used for each experiment can be found in Appendix A.", + "bbox": [ + 75, + 199, + 470, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Common image corruptions. To evaluate TeSLA's efficacy for the classification task, we use CIFAR10-C/CIFAR100-C [16] and large-scale ImageNet-C [16] datasets each containing 19 types of corruptions applied to the clean test set with five levels of severity. We perform validation on 4 out of 19 types of corruption (SPATTER, GAUSSIAN BLUR, SPECKLE NOISE, SATURATE) to select hyperparameters and test on the remaining 15 corruptions at the maximum severity level of 5. Following [28], we train the ResNet50 [15] on the clean CIFAR10/CIFAR100/ImageNet training set and adapt it to classify the unlabeled corrupted test set.", + "bbox": [ + 75, + 335, + 470, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic to real data adaption. We use challenging and large-scale VisDA-C, and VisDA-S [33] datasets for evaluating synthetic-to-real data adaptation at test-time for classification and segmentation tasks, respectively. Following [4, 26], we adapt the ResNet101 network pre-trained on synthetic images to classify 12 vehicle classes on the photo-realistic images of VisDA-C. While for VisDA-S, we adapt the DeepLab-v3 [5] backbone pre-trained on synthetic GTA5 [36] to the Cityscapes [7] to segment 19 classes.", + "bbox": [ + 75, + 503, + 470, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Measurement shifts on medical images. We access TeSLA on staining variations for the tissue type classification of hematoxylin & eosin (H&E) stained patches from colorectal cancer tissue slides. We use MobileNetV2 [38] trained on the source Kather-19 dataset [20] and adapt it to the target Kather-16 dataset [21] on four tissue categories: tumor, stroma, lymphocyte, and mucosa. We also evaluate TeSLA for variations in scanners and imaging protocols in multi-site medical images on two magnetic resonance imaging (MRI) datasets for the segmentation task, namely the multi-site prostate MRI [27] and spinal cord grey matter segmentation (SCGM) [34]. Following [42], we adapt the U-Net [37] from site 1 to sites $\\{2,3,4\\}$ of the spinal cord dataset, and from sites $\\{\\mathrm{A,B}\\}$ to sites $\\{\\mathrm{D,E,F}\\}$ of the prostate dataset.", + "bbox": [ + 75, + 642, + 470, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competing baselines. We compare TeSLA with the following SFDA and TTA baselines, including direct inference of the trained source model on the target test data without", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c94b321f2ebfd4eea560c891857dca2c4b0f4eb11dba94009e0c10369b54a064.jpg", + "table_caption": [ + "Table 1. Comparison of SOTA TTA methods under different protocols evaluated on CIFAR-10/100-C, ImageNet-C, VisDA-C and Kather-16 datasets. We report the average error computed over 15 test corruptions for the common image corruption shifts. We also report Class Avg. in % error rates for synthetic-to-real and measurement shifts over 3 and 10 seeds, respectively." + ], + "table_footnote": [], + "table_body": "
MethodProtocolCommon Image CorruptionsSyn-to-RealMeasurement Shift
CIFAR10-CCIFAR100-CImageNet-CVisDA-CKather-16
OMOMOMOMOM
SourceN29.160.481.851.532.0
BN [18,30]N15.615.443.743.367.767.635.435.018.318.2
TENT [45]N14.112.939.036.557.454.233.529.316.212.0
SHOT [26]N13.914.239.238.768.768.229.424.514.712.0
AdaContrast [4]N-23.120.2-
TTT++ [28]Y15.89.844.434.159.3-35.234.116.77.9
TTAC [39]Y13.49.441.733.658.7-32.231.19.65.5
TeSLAN12.59.738.232.955.051.517.813.59.23.3
TeSLA-sY12.19.737.332.653.1-24.017.99.93.1
", + "bbox": [ + 503, + 183, + 893, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adaptation (Source). We also implement the SFDA-based pseudo-labeling baselines from a basic pseudo-labeling approach (PL) to the SHOT approach [26] based on training the feature extraction module by adopting the information maximization loss to make globally diverse but individually certain predictions on the target domain. For recent TTA methods, we compare our method against $\\mathrm{TTT}++$ [28], Ada-Contrast [4], CoTTA [47], basic TENT [45] and BN [18] as representative methods based on feature distribution alignment via stored statistics, contrastive learning, augmentation-averaged predictions, entropy minimization, and batch normalization statistic. Furthermore, we compare TeSLA to more recent methods based on test-time augmentation policy learning, OptTTA [42], and anchor clustering TTAC [39]. By default, [4,18,26] follow the N-M protocol, and we adapt them to the N-O protocol, while [42,45,47], by default, follow the N-O protocol and adapt them to the N-M protocol setting. Using source features distribution, TTAC and $\\mathrm{TTT}++$ follow Y-O and Y-M. We report error rates on classification tasks and Mean Intersection over Union (mIoU)/ Dice scores on segmentation tasks. For consistency across baselines and TTA protocols, we do not use any specialized model architectures (e.g., projection head, weight-normalized classifier layer) for SHOT [26] and AdaContrast [4].", + "bbox": [ + 496, + 334, + 895, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 709, + 591, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification task. In Table 1, we summarize the average (Avg) classification error rates $(\\%)$ of TeSLA against several SOTA SFDA and TTA baselines on the common image corruptions of CIFAR10-C, CIFAR100-C, ImageNet-C; synthetic to real data domain shifts of VisDA-C; and measurement shifts on the Kather-16 dataset. We also report per-class top-1 accuracies for VisDA-C and Kather-16 datasets and corruption-wise error rates on CIFAR-10-C/CIFAR-100-C and ImageNet-C datasets in Appendix B. TeSLA easily surpasses all competing methods under the protocols N-O and N-M on all the datasets. In particular, TeSLA outper", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20346", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5481e6f89e003844907e5351db3ec8d7f9fbe209cb6f605e759a00edcd5361d1.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 143, + 89, + 472, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51a933acca7eabc2edb1b3dca9907571ec6ab9ca6b60f8439724348436d09b4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 85, + 630, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89ef494eb32bc41feff3ecfad170cab501eecd14ca45eaeedc89557b927532f1.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 638, + 85, + 776, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dc96ef47a6b961c8efc22b5ca63c6d510532cfa72c6860e7c1a408e7fc548fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 85, + 821, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa75a6905b65285834f8f181ef46ea44c25bd404a58608b3b943a9dee2ed7683.jpg", + "image_caption": [ + "Figure 3. Ablation experiments on the VisDA-C. (a) Source training strategy and model architecture. TeSLA outperforms the competing TTA methods across varied source training strategies (Supervised, SimCLRV2 [6], and DINO [3]) and model architectures (ResNet50, ResNet101 [15], and ViT-B/16 [11]) using the same set of hyperparameters for both protocols (N-O/N-M). Each vertex represents the mean class avg accuracy over 3 seeds. (b) t-SNE visualization [43] comparison of feature embeddings for AdaContrast [4] and TeSLA." + ], + "image_footnote": [], + "bbox": [ + 76, + 311, + 267, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/733c05ae857a8f2ac9c803d3f672b8f392aa59ff42b2f12cb264332f3671cab8.jpg", + "image_caption": [ + "Figure 4. Model calibration and uncertainty estimation. Model calibration using reliability diagrams [31] and model uncertainty estimation using Brier Score and Negative Log Likelihood (NLL) of the Source Model, TENT [45], OptTTA [42] and, TeSLA on the prostate dataset for test-time adaptation." + ], + "image_footnote": [], + "bbox": [ + 279, + 311, + 470, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/850278dec17a7f33c413d5d9711da22eedcb00a679ca0bc3a4dc640abae35b43.jpg", + "table_caption": [ + "Table 2. Semantic segmentation results (Class Avg. mIoU in %) on the VisDA-S (GTA5 → Cityscapes) test-time adaptation task." + ], + "table_footnote": [], + "table_body": "
ProtocolSourceBN [30]TENT [45]PLCoTTA [47]TeSLA
NO35.336.738.338.837.044.5
NM38.439.238.639.946.0
", + "bbox": [ + 98, + 561, + 447, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "forms the second-best baseline under N-O/N-M protocol by a $(\\%)$ margin 1.4/3.2 on CIFAR10-C, 0.8/3.6 on CIFAR100-C, 2.4/2.7 on ImageNet-C, 5.3/6.7 on VisDA-C, and 5.5/8.7 on the Kather-16 datasets. Despite not utilizing any source feature statistic, TeSLA (N-O/N-M) is either competitive or surpasses TTAC [39] and TTT++ [28] (Y-O/Y-M) that benefit from the source dataset feature statistics during adaptation. Including the global source feature alignment module of TTAC with TeSLA (TeSLA-s) could further improve its performance under (Y-O/Y-M) protocols for slight domain shifts (e.g., common image corruptions). We report the results of AdaContrast for VisDA-C only as the implementation for other datasets are not provided. We do not report TeSLA-s results on ImageNet-C under the Y-M protocol, as previous baselines were evaluated only Y-O protocol.", + "bbox": [ + 75, + 613, + 472, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation task. Unlike TeSLA, several TTA methods e.g., AdaContrast [4], SHOT [26], TTT++ [28], TTAC [39], have only been evaluated on the classification task. Therefore, we compare TeSLA against the current SOTA TTA", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6418cfad5fc98528e22d4a4ad06fd3f77e25b68c79776683c75cea719190e5e0.jpg", + "table_caption": [ + "Table 3. Semantic segmentation results (Class Avg. volume-wise mean Dice in %) on the cross-site spinal cord and prostate MRI test-time adaptation tasks." + ], + "table_footnote": [], + "table_body": "
ProtocolSourceBN [30]TENT [45]PLOptTTA [42]TeSLA
Spinal CordSite{1} → Sites{2,3,4}
NO76.0±11.881.6±8.381.1±9.181.7±8.684.1±4.885.3±5.8
NM84.3±4.884.4±4.784.3±4.784.3±4.485.4±4.4
ProstateSites{A,B} → Sites{D,E,F}
NO60.5±27.072.1±15.274.7±17.972.4±15.283.1±7.783.5±6.5
NM73.1±18.081.2±9.381.1±9.283.4±7.784.3±5.8
", + "bbox": [ + 503, + 364, + 893, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods applied to the segmentation task on synthetic-to-real data transfer of the VisDA-S dataset in Table 2. TeSLA significantly outperforms the competing methods and achieves the best mIOU scores for both N-O and N-M protocols, beating the second-best baseline by a margin of $+5.7\\%$ and $+6.1\\%$ , respectively. In Table 3, we compare TeSLA against the recent SOTA test-time augmentation policy method, OptTTA [42] and other TTA baselines for the inter-site adaptation on two challenging MRI datasets mentioned in Sec. 4.1. TeSLA convincingly outperforms all other methods on severe measurement shifts across sites under N-O and N-M protocols.", + "bbox": [ + 496, + 470, + 893, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational cost for adversarial augmentations. As shown in Table 4, TeSLA learns the adversarial augmentation in an online manner, and thus its runtime is several orders faster than learnable test-time augmentation OptTTA [42], while comparable to that of static augmentation policies with an additional overhead of only 0.10 GPU hours/epoch on the VisDA-C dataset.", + "bbox": [ + 496, + 654, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Test-time feature visualization. Fig. 3b compares t-SNE projection [43] of the encoder's features of AdaContrast [4] with TeSLA on the VisDA-C. TeSLA shows better inter-class separation for features than AdaContrast, as supported by an improved silhouette score from 0.149 to 0.271.", + "bbox": [ + 496, + 762, + 893, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model calibration and uncertainty estimation. Entropy minimization-based TTA methods [26, 45] explicitly make the model confident in their predictions, resulting in poor model calibration. For trusting the adapted model, it should", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20347", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d88846d76a7c506b0779c94fc0476e693ab0251a53c6b0fa4fd18e277251acc.jpg", + "table_caption": [ + "Table 4. Performance comparison of adversarial augmentation learned by TeSLA during adaptation against the prior art augmentation methods on runtime (GPU hours/epoch on NVIDIA GeForce RTX 3090) and Class Avg. accuracy in % on the VisDA-C dataset." + ], + "table_footnote": [], + "table_body": "
Augmentation TypeOptTTATeSLARA [10]TeSLAAA [8]TeSLAN=2TeSLAN=3
Runtime4.500.050.060.160.18
Class Avg. Acc. (N-O)-80.281.281.582.2
Class Avg. Acc. (N-M)-84.785.786.386.5
", + "bbox": [ + 81, + 151, + 465, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "output reliable confidence estimates matching its true underlying performance on the test images. Such a well-behaved model is characterized by expected calibration error (ECE) through reliability diagram [31], and uncertainty metrics of brier score and negative log-likelihood (NLL) [14]. In Fig. 4, we plot the reliability diagram (dividing the probability range [0, 0.5] into 10 bins), and uncertainty metrics of the Source model, TENT [45] and OptTTA [42] against TeSLA for the inter-site model adaptation on the prostate dataset [27]. We observe that TeSLA achieves the best model calibration (lowest ECE of $1.80\\%$ ) and lowest Brier and NLL scores of 0.12 and 0.24 ( $p < 0.006$ against TENT).", + "bbox": [ + 75, + 228, + 470, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 421, + 240, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we scrutinize the roles played by different components of TeSLA. All ablations are performed on the synthetic-to-real test data adaptation of the VisDA-C dataset. Loss functions for self-learning. We compare TTA performance of our self-learning objective $\\mathcal{L}_{\\mathrm{pl}}$ with the proposed flipped cross-entropy loss $f$ -CE and the basic cross-entropy loss $\\mathbb{C}\\mathbb{E}$ in Fig. 5-(1). We do not use the proposed augmentation module and pseudo-label refinement in this experiment. We observe that $f$ -CE alone improves the accuracy of the Source model in the N-M protocol by $+23.1\\%$ compared to $+16.6\\%$ improvement by $\\mathbb{C}\\mathbb{E}$ . Using $\\mathcal{L}_{\\mathrm{pl}}$ (Eq. 3) further improves the accuracy to a margin of $+23.8\\%$ .", + "bbox": [ + 75, + 445, + 468, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contribution of individual components: PLR, $\\mathcal{L}_{\\mathrm{pl}}$ , and $\\mathcal{L}_{\\mathrm{kd}}$ . Our test-time objective $\\mathcal{L}_{\\mathrm{TeSLA}}$ has three components (i) self-learning loss $\\mathcal{L}_{\\mathrm{pl}}$ , (ii) pseudo-label refinement (PLR), and (iii) knowledge distillation $\\mathcal{L}_{\\mathrm{kd}}$ with adversarial augmentation. In Fig. 5-(3), we study the effect of accruing individual components to $\\mathcal{L}_{\\mathrm{TeSLA}}$ (starting from source trained model) for the test-time adaptation accuracy. We observe that $\\mathcal{L}_{\\mathrm{pl}}$ alone improves the model's accuracy from $48.5\\%$ to $72.3\\%$ while adding PLR and $\\mathcal{L}_{\\mathrm{kd}}$ further boosts it to $81.6\\%$ $(+9.3)$ and $86.3\\%$ $(+4.7)$ , respectively.", + "bbox": [ + 75, + 628, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PLR: ensembling on weak augmentations (Ens) and nearest neighbor averaging (NN). We also conduct an ablation on the two pseudo-label refinement approaches - (i) ensembling the teacher model's output on five weak augmentations. (ii) averaging soft pseudo labels of $n = 10$ nearest neighbors in the feature space (cf. sensitivity test in Appendix E). Fig. 5-(2) shows that combining both approaches gives better results than using them individually.", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d3f41fc7804936003a516f0bb3545a2b728a6817358273bbe0a209dcc23190b.jpg", + "image_caption": [ + "Figure 5. Ablation experiments on the VisDA-C. (left to right) (1) Ablation results for self-learning loss functions, including $f\\text{-}\\mathbb{C}\\mathbb{E}$ , $\\mathbb{C}\\mathbb{E}$ , and $\\mathcal{L}_{\\mathrm{pl}}$ ; (2) ablation on different PLR methods: assembling predictions on 5 weak augmentations, averaging predictions of the 10 nearest neighbors in the feature space; (3) ablation on different components of TeSLA's objective function: $\\mathcal{L}_{\\mathrm{pl}}$ , PLR, and $\\mathcal{L}_{\\mathrm{kd}}$ ." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 629, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e54867bba1455f679010f097741a460ee3810386624bea3ffd5d1af8a9ba4043.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 90, + 751, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b204bfa589e44d85f6571bcdea810af261121e04f110a54d4ecd08a1c7d81a96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 90, + 880, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adversarial augmentation. Table 4 shows the benefit of using our adversarial augmentation $(N = 2,3)$ with TeSLA against (i) RandAugment (RA) [10], (ii) AutoAugment (AA) [9] (optimized for ImageNet), and (iii) OptTTA [42]. We achieve far better Class Avg. accuracy on the VisDA-C at the cost of slight runtime overhead compared to static augmentation policies for protocols (N-O, N-M). In addition, TeSLA augmentation (TeAA) consistently improves the performance of other TTA methods (Table 5).", + "bbox": [ + 496, + 297, + 893, + 434 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a9dbea93fcc28b84628336ebc80e3a0791ba8b69931650cea705212ddc4ed460.jpg", + "table_caption": [ + "Table 5. Class Avg. accuracy (%) of TeSLA augmentation (TeAA) with other TTA objectives under N-O protocol on VisDA-C dataset." + ], + "table_footnote": [], + "table_body": "
TENT+TeAA| SHOT+TeAA| AdaContrast+TeAA
66.572.0| 70.673.1| 76.979.1
", + "bbox": [ + 517, + 481, + 870, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Source training strategies and model architectures. The ablation results (Fig. 3a) on the VisDA-C show that TeSLA surpasses competing TTA baselines for different source training strategies, including Supervised and self-supervised (SimCLRV2 [6], DINO [3]). Moreover, we ablate TeSLA over different architectures, ranging from ResNet-50/101 [15] to the recent vision transformer ViT-B [11] using the same set of hyperparameters, showing the merits of no reliance on the network architecture or source training strategy.", + "bbox": [ + 496, + 527, + 893, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, in Appendix E, we provide detailed sensitivity results with respect to all hyperparameters.", + "bbox": [ + 496, + 664, + 890, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced TeSLA, a novel self-learning algorithm for test-time adaptation that utilizes automatic adversarial augmentation. TeSLA is agnostic to the model architecture and source training strategies and gives better model calibration than other TTA methods. Through extensive experiments on various domain shifts (measurement, common corruption, synthetic to real), we show TeSLA's superiority over previous TTA methods on classification and segmentation tasks. Note that TeSLA assumes class uniformly when implicitly maximizing mutual information and can be improved by incorporating prior, e.g., class label distribution statistics.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20348", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009-2018, 2022. 1, 2", + "[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. In International Conference on Learning Representations, 2019. 2", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 7, 8", + "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2, 6, 7", + "[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7, 8", + "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[8] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 8", + "[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 8", + "[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 8", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 7, 8", + "[12] Cian Eastwood, Ian Mason, Christopher K. I. Williams, and Bernhard Scholkopf. Source-free adaptation to measurement shift via bottom-up feature restoration. In International Conference on Learning Representations, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 2", + "[14] Alvaro Gomariz, Tiziano Portenier, César Nombela-Arrieta, and Orcun Goksel. Probabilistic spatial analysis in quantitative microscopy with uncertainty-aware cell detection using deep bayesian regression of density maps. arXiv preprint arXiv:2102.11865, 2021. 8", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7, 8", + "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 6", + "[17] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 2", + "[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 6", + "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1, 2", + "[20] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 6", + "[21] Jakob Nikolas Kather, Frank Gerrit Zollner, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Cleo-Aron Weis. Collection of textures in colorectal cancer histology, May 2016. 6", + "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[23] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 1, 2", + "[24] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12376-12385, 2020. 2", + "[25] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 1, 2", + "[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20349", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 3, 6, 7", + "[27] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. Msnet: Multi-site network for improving prostate segmentation with heterogeneous mri data. IEEE Transactions on Medical Imaging, 2020. 6, 8", + "[28] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 7", + "[29] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy policy search: A simple baseline for learnable test-time augmentation. In Conference on Uncertainty in Artificial Intelligence, pages 1308-1317. PMLR, 2020. 2", + "[30] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 6, 7", + "[31] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In Proceedings of the 22nd international conference on Machine learning, pages 625-632, 2005. 7, 8", + "[32] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 2", + "[33] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6", + "[34] Ferran Prados, John Ashburner, Claudia Blaiotta, Tom Brosch, Julio Carballido-Gamio, Manuel Jorge Cardoso, Benjamin N Conrad, Esha Datta, Gergely David, Benjamin De Leener, et al. Spinal cord grey matter segmentation challenge. Neuroimage, 152:312-329, 2017. 6", + "[35] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In International Joint Conference on Artificial Intelligence, 2021. 2", + "[36] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, European Conference on Computer Vision (ECCV), volume 9906 of LNCS, pages 102-118. Springer International Publishing, 2016. 6", + "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6", + "[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6", + "[39] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7", + "[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2", + "[41] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3", + "[42] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozortbar. OptTTA: Learnable test-time augmentation for source-free medical image segmentation under domain shift. In Medical Imaging with Deep Learning, 2022. 2, 3, 6, 7, 8", + "[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7", + "[44] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 2", + "[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8", + "[46] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 338:34-45, 2019. 2", + "[47] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 6, 7", + "[48] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Casting a bait for offline and online source-free domain adaptation. arXiv preprint arXiv:2010.12427, 2020. 2", + "[49] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8978-8987, 2021. 2", + "[50] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. arXiv preprint arXiv:2205.04183, 2022. 2", + "[51] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29:7834-7844, 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20350", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_model.json b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c477d82d5ce3d578bb74cdeefd68409d34b83fac --- /dev/null +++ b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.18, + 0.837, + 0.215 + ], + "angle": 0, + "content": "Devavrat Tomar1 Guillaume Vray1 Behzad Bozorgtabar1,2 Jean-Philippe Thiran1,2 EPFL CHUV" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.218, + 0.616, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\{\\) firstname\\}. \\(\\{\\)lastname\\}@epfl.ch" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.312, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with automatic Adversarial augmentation dubbed TeSLA for adapting a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning methods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowledge distillation by simulating high entropy augmented images. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several desirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available at https://github.com/devavratTomar/TeSLA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.208, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) perform exceptionally well when the training (source) and test (target) data follow the same distribution. However, distribution shifts are inevitable in real-world settings and propose a major challenge to the performance of deep networks after deployment. Also, access to the labeled training data may be infeasible at test time due to privacy concerns or transmission bandwidth. In such scenarios, source-free domain adaptation (SFDA) [1, 23, 25] and test-time adaptation (TTA) methods [19, 28, 39] aim to adapt the pre-trained source model to the unlabeled distributionally shifted target domain while easing access to source data. While SFDA methods have access to all full target data through multiple training epochs" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.646, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.268, + 0.81, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.268, + 0.893, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.385, + 0.895, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Knowledge distillation with adversarial augmentations. (a) Easy images with confident soft-pseudo labels and (b) Hard images with unconfident soft-pseudo labels are adversarily augmented and pushed to the uncertainty region (high entropy) near the decision boundary. The model is updated for (a) to match its output on the augmented views with non-augmented views of Easy test images using KL-Divergence \\(\\mathcal{L}_{\\mathrm{kd}} \\neq 0\\), while not updated for (b) as \\(\\mathcal{L}_{\\mathrm{kd}} \\sim 0\\) between Hard images and their augmented views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.643 + ], + "angle": 0, + "content": "(offline setup), TTA methods usually process test images in an online streaming fashion and represent a more realistic domain adaptation. However, most of these methods are applied: (i) only to classification tasks, (ii) evaluated on the non-real-world domain shifts, e.g., the non-measurement shift; (iii) destroy model calibration—entropy minimizing with overconfident predictions [45] on incorrectly classified samples, and (iv) use specialized network architectures or rely on the source dataset feature statistics [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We address these issues by proposing a new test-time adaptation method with automatic adversarial augmentation called TeSLA, under which we further define realistic TTA protocols. Self-learning methods often supervise the model adaptation on the unlabeled test images using their predicted pseudo-labels. As the model can easily overfit on its own pseudo-labels, a weight-averaged teacher model (slowly updated by the student model) is employed for obtaining the pseudo-labels [41, 47]. The student model is then trained with cross-entropy \\((\\mathbb{C}\\mathbb{E})\\) loss between the one-hot pseudolabels and its predictions on the test images. In this paper, we instead propose to minimize flipped cross-entropy between the student model's predictions and the soft pseudo-labels (notice the reverse order) with the negative entropy of its marginalized predictions over the test images. In Sec. 3, we show that the proposed formulation is an equivalence to mutual information maximization implicitly corrected" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20341" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "by the teacher-student knowledge distillation via pseudolabels, yielding performance improvement on various test-time adaptation protocols and compared to the basic \\(\\mathbb{C}\\mathbb{E}\\) optimization (cf. ablation in Fig. 5-(1))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.363 + ], + "angle": 0, + "content": "Motivated by teacher-student knowledge distillation, another central tenet of our method is to assist the student model during adaptation in improving its performance on the hard-to-classify (high entropy) test images. For this purpose, we propose learning automatic adversarial augmentations (see Fig. 1) as a proxy for simulating images in the uncertainty region of the feature space. The model is then updated to ensure consistency between predictions of high entropy augmented images and soft-pseudo labels from the respective non-augmented versions. Consequently, the model is self-distilled on Easy test images with confident soft-pseudo labels (Fig. 1a). In contrast, the model update on Hard test images is discarded (Fig. 1b), resulting in better class-wise feature separation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.471, + 0.606 + ], + "angle": 0, + "content": "In summary, our contributions are: (i) we propose a novel test-time self-learning method based on flipped cross-entropy \\((f\\text{-}\\mathbb{C}\\mathbb{E})\\) through the tight connection with the mutual information between the model's predictions and the test images; (ii) we propose an efficient plug-in test-time automatic adversarial augmentation module used for online knowledge distillation from the teacher to the student network that consistently improves the performance of test-time adaptation methods, including ours; and (iii) TeSLA achieves new state-of-the-art results on several benchmarks, from common image corruption to realistic measurement shifts for classification and segmentation tasks. Furthermore, TeSLA outperforms existing TTA methods in terms of calibration and uncertainty metrics while making no assumptions about the network architecture and source domain's information, e.g., feature statistics or training strategy." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.619, + 0.22, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.87 + ], + "angle": 0, + "content": "In general, domain adaptation methods aim to distill knowledge from source data that are well-generalizable to target data and relax the assumption of i.i.d. between source and target datasets. To circumvent expensive and cumbersome annotation of new target data, unsupervised domain adaptation (UDA) emerges, and there has been a large corpus of UDAs [13, 17, 32, 51] to match the distribution of domains on both labeled source data and unlabeled target data. Nonetheless, the above-mentioned methods demand source data to achieve the domain adaptation process, which is often impractical in real-world scenarios, e.g., due to privacy restrictions. The above issue motivates research into source-free domain adaptation (SFDA) [1, 23, 25, 26] and test-time training or adaptation (TTA) [19, 28, 39, 45], which are more closely relevant to our problem setup." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "SFDA approaches formulate domain adaptation through pseudo labeling [23, 26], target feature clustering [49], syn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.333 + ], + "angle": 0, + "content": "thesizing extra training samples [24], or feature restoration [12]. SHOT [26] proposed feature clustering via information maximization while incorporating pseudo-labeling as additional supervision. BAIT [48] leverages the fixed source classifier as source anchors and uses them to achieve feature alignment between the source and target domain. CPGA [35] proposed the SFDA method based on matching feature prototypes between source and target domains. AaD-SFDA [50] proposed to optimize an objective by encouraging prediction consistency of local neighbors in feature space. Nevertheless, SFDA methods require apriori access to all target data in advance, and the current pseudo-labeling approach, e.g., SHOT [26], used offline pseudo-label refinement on a perepoch basis. In a more realistic domain adaptation scenario, SFDA is still incompetent to perform inference and adaptation simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.746 + ], + "angle": 0, + "content": "TTA methods [4, 45] propose alleviating the domina shift by online (or streaming) adapting a model at test time. Still, we argue that there are ambiguities over the problem setup of TTA in the literature, particularly on whether sequential inference on target test data is feasible upon arrival [19, 45] or whether training objectives must be adapted [28, 40]. TTT [40] proposed fine-tuning the model parameters via rotation classification task as a proxy. On-target adaptation [44] used pseudo-labeling and contrastive learning to initialize the target-domain feature; each performed independently in their method. T3A [19] utilized pseudo-labeling to adjust the classifier prototype. More recently, TTAC [39] leveraged the clustering scheme to match target domain clusters to source domain categories. Nevertheless, existing TTA methods rely on specialized neural network architectures or only update a fraction of model weights yielding limited performance gain on the target data. For instance, TENT [45] proposed updating affine parameters in the batchnorm layers of convolutional neural networks (CNN), while AdaContrast [4] and SHOT [26] used an additional weight normalization classification layer with a projection head. In contrast, we show our method's superiority for various neural network architectures, including CNNs and vision transformers (ViTs) [11], without additional architectural requirements or different source model training strategies. Moreover, we update all model parameters, and our method is stable over a wide range of hyper-parameters, e.g., learning rate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Test time augmentation methods [2, 29] are another popular line of domain adaptation research. GPS [29] learns optimal augmentation sub-policies by combining image transformations of RandAugment [10] that minimize calibrated log-likelihood loss [2] on the validation set. OptTTA [42] optimized the magnitudes of augmentation sub-policies using gradient descent to maximize mutual information and match feature statistics over augmented images. Though effective, these methods [29, 42, 46] are computationally expensive as all augmentation sub-policies need to be evaluated, making" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20342" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "their real-time application difficult. Instead, our proposed adversarial augmentation policies can be learned online and are several orders faster than the existing learnable test-time augmentation strategies, e.g., [42]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.159, + 0.212, + 0.177 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Overview. We first introduce our flipped cross-entropy \\((f\\text{-}\\mathbb{C}\\mathbb{E})\\) loss through the tight connection with the mutual information between the model's predictions and the test images. (Sec. 3.1). Using this equivalence, we derive our test-time loss function that implicitly incorporates teacher-student knowledge distillation. In Sec. 3.2, we propose to enhance the test-time teacher-student knowledge distillation by utilizing the consistency of the student model's predictions on the proposed adversarial augmentations with their corresponding refined soft-pseudo labels from the teacher model. We refine the soft-pseudo labels by averaging the teacher model's predictions on (1) weakly augmented test images; (2) nearest neighbors in the feature space. Furthermore, we propose an efficient online algorithm for learning adversarial augmentations (Sec. 3.3). Finally, in Sec. 3.4, we summarize our online test-time adaptation method, TeSLA, based on self-learning with the proposed automatic adversarial augmentation. The overall framework is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.473, + 0.795 + ], + "angle": 0, + "content": "Problem setup. Given an existing model \\( f_{\\theta_0} \\), parametrized by \\( \\theta_0 \\) and pre-trained on the inaccessible source data, we aim to improve its performance by updating all its parameters using the unlabeled streaming test data \\( \\mathbf{X} \\) during adaptation. Motivated by the concept of mean teacher [41] as the weight-averaged model over training steps that can yield a more accurate model, we build our online knowledge distillation framework upon teacher-student models. We utilize identical network architecture for teacher and student models. Each model \\( f = h \\circ g \\) comprises a backbone encoder \\( g: \\mathbf{X} \\to \\mathbb{R}^D \\) mapping unlabeled test image \\( \\mathbf{x} \\in \\mathbf{X} \\) to the feature representation \\( \\mathbf{z} = g(\\mathbf{x}) \\in \\mathbb{R}^D \\) and a classifier (hypothesis) head \\( h: \\mathbb{R}^D \\to \\mathbb{R}^K \\) mapping \\( \\mathbf{z} \\) to the class prediction \\( \\mathbf{y} \\in \\mathbb{R}^K \\), where \\( D \\) and \\( K \\) denote the feature dimension and number of classes. The parameters of teacher \\( \\theta_t \\) and student \\( \\theta_s \\) are first initialized from the source model parameters \\( \\theta_0 \\). Then the teacher model's parameters \\( \\theta_t \\) are updated from the student model's parameters \\( \\theta_s \\) using an exponential moving average (EMA) with momentum coefficient \\( \\alpha \\) as: \\( \\theta_t \\gets \\alpha \\cdot \\theta_t + (1 - \\alpha) \\cdot \\theta_s \\). Following [26], we freeze the classifier \\( h \\)'s parameters and only update the encoder \\( g \\)'s parameters during test-time adaptation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.398, + 0.818 + ], + "angle": 0, + "content": "3.1. Rationale Behind the \\( f \\)-CE Objective" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "We start by analyzing the rationale behind the proposed \\(f\\)-CE loss and its benefit for self-learning through the tight connection with the mutual information between the model's predictions and unlabeled test images. Before that, we first define the notations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.267 + ], + "angle": 0, + "content": "Notations. We define the random variables of unlabeled test images and the predictions from the student model \\( f_{s} \\) as \\( \\mathbf{X} \\) and \\( \\mathbf{Y} \\) and those of the teacher model \\( f_{t} \\)'s soft pseudo-label predictions as \\( \\hat{\\mathbf{Y}} \\). Furthermore, let \\( p_{\\mathbf{Y}} \\) denotes the marginal distribution over \\( \\mathbf{Y} \\), \\( p_{(\\mathbf{Y},\\mathbf{X})} \\) be the joint distribution of \\( \\mathbf{Y} \\) and \\( \\mathbf{X} \\), and \\( p_{\\mathbf{Y}|\\mathbf{X}} \\) be the conditional distribution of \\( \\mathbf{Y} \\) given \\( \\mathbf{X} \\). Then, the entropy of \\( \\mathbf{Y} \\) and the conditional entropy of \\( \\mathbf{Y} \\) given \\( \\mathbf{X} \\) can be defined as \\( \\mathcal{H}(\\mathbf{Y}) := \\mathbb{E}_{p_{\\mathbf{Y}}}[-\\log p_{\\mathbf{Y}}(\\mathbf{Y})] \\) and \\( \\mathcal{H}(\\mathbf{Y}|\\mathbf{X}) := \\mathbb{E}_{p_{(\\mathbf{Y},\\mathbf{X})}}[-\\log p_{\\mathbf{Y}|\\mathbf{X}}(\\mathbf{Y}|\\mathbf{X})] \\), respectively. Besides, let the flipped cross-entropy (\\( f-\\mathbb{C}\\mathbb{E} \\)) given \\( \\mathbf{X} \\) be \\( \\mathcal{H}(\\mathbf{Y};\\hat{\\mathbf{Y}}|\\mathbf{X}) := \\mathbb{E}_{p_{(\\mathbf{Y},\\mathbf{X})}}[-\\log p_{\\hat{\\mathbf{Y}}|\\mathbf{X}}(\\mathbf{Y}|\\mathbf{X})] \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Self-learning and mutual information. Based on the above definitions, the \\(f\\)-\\(\\mathbb{C}\\mathbb{E}\\) between \\(\\mathbf{Y}\\) and \\(\\hat{\\mathbf{Y}}\\) conditioned over \\(\\mathbf{X}\\) has the tight connection with the mutual information \\(\\mathcal{I}(\\mathbf{Y};\\mathbf{X})\\) between unlabeled test images \\(\\mathbf{X}\\) and the predictions from the student model \\(\\mathbf{Y}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.37, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {H} (\\mathbf {Y}; \\hat {\\mathbf {Y}} \\mid \\mathbf {X}) = \\mathcal {H} (\\mathbf {Y} \\mid \\mathbf {X}) + \\mathcal {D} _ {\\mathrm {K L}} (\\mathbf {Y} \\| \\hat {\\mathbf {Y}} \\mid \\mathbf {X}) \\tag {1} \\\\ = - \\mathcal {I} (\\mathbf {Y}; \\mathbf {X}) + \\mathcal {H} (\\mathbf {Y}) + \\mathcal {D} _ {\\mathrm {K L}} \\left(\\mathbf {Y} \\| \\hat {\\mathbf {Y}} | \\mathbf {X}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.543 + ], + "angle": 0, + "content": "This implies that minimizing the \\(f\\)-CE \\((\\mathcal{H}(\\mathbf{Y};\\hat{\\mathbf{Y}}\\mid \\mathbf{X})\\) in Eq. 1) and maximizing the entropy of class-marginal prediction \\(\\mathcal{H}(\\mathbf{Y})\\) is equivalent to maximizing the mutual information between the test images and the student model's predictions \\(\\mathcal{I}(\\mathbf{Y};\\mathbf{X})\\) with a correction KL-divergence term \\(\\mathcal{D}_{\\mathrm{KL}}(\\mathbf{Y}\\parallel \\hat{\\mathbf{Y}}\\mid \\mathbf{X})\\) involving soft-pseudo labels as:" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.562, + 0.901, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\mathcal {H} (\\mathbf {Y} ; \\hat {\\mathbf {Y}} \\mid \\mathbf {X})} _ {f \\cdot \\mathbb {C E}} - \\mathcal {H} (\\mathbf {Y}) = - \\underbrace {\\mathcal {I} (\\mathbf {Y} ; \\mathbf {X})} _ {\\text {M u t u a l I n f o .}} + \\mathcal {D} _ {\\mathrm {K L}} (\\mathbf {Y} \\| \\hat {\\mathbf {Y}} | \\mathbf {X}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.895, + 0.743 + ], + "angle": 0, + "content": "Thus, minimizing the left side of Eq. 2 with respect to the student model's parameters \\(\\theta_{s}\\) allows the student model to cluster test images using mutual information and criterion-corrected knowledge distillation from soft-pseudo labels. Using the above formulation, we define the following test-time objective to train the student model \\(f_{s}\\) on a batch of \\(B\\) test images \\(X = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_B\\}\\) with their corresponding soft-pseudo labels from teacher model \\(\\hat{Y} = \\{\\hat{y}_1,\\dots ,\\hat{y}_B\\}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.77, + 0.892, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p l}} (X, \\hat {Y}) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {k = 1} ^ {K} f _ {s} \\left(\\boldsymbol {x} _ {i}\\right) _ {k} \\log \\left(\\left(\\hat {\\boldsymbol {y}} _ {i}\\right) _ {k}\\right) \\\\ + \\sum_ {k = 1} ^ {K} \\hat {f} _ {s} (X) _ {k} \\log (\\hat {f} _ {s} (X)) _ {k} \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.868, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\hat{f}_s(X) = \\frac{1}{B}\\sum_{i=1}^B f_s(\\pmb{x}_i)\\) denotes the marginal class distribution over the batch of test images \\(X\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20343" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.088, + 0.873, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.236, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Figure 2. Overview of TeSLA Framework. (a) The student model \\( f_{s} \\) is adapted on the test images by minimizing the proposed test-time objective \\( \\mathcal{L}_{\\mathrm{pl}} \\). The high-quality soft-pseudo labels required by \\( \\mathcal{L}_{\\mathrm{pl}} \\) are obtained from the exponentially weighted averaged teacher model \\( f_{t} \\) and refined using the proposed Soft-Pseudo Label Refinement (PLR) on the corresponding test images. The soft-pseudo labels are further utilized for teacher-student knowledge distillation via \\( \\mathcal{L}_{\\mathrm{kd}} \\) on the adversarially augmented views of the test images. (b) The adversarial augmentations are obtained by applying learned sub-policies sampled i.i.d from \\( \\mathbb{P} \\) using the probability distribution \\( P \\) with their corresponding magnitudes selected from \\( M \\). The parameters \\( M \\) and \\( P \\) of the augmentation module are updated by the unbiased gradient estimator (Eq. 13) of the loss \\( \\mathcal{L}_{\\mathrm{Aug}} \\) computed on the augmented test images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.354, + 0.43, + 0.37 + ], + "angle": 0, + "content": "3.2. Knowledge Distillation via Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Self-learning from adversarial augmentations. As illustrated in Fig. 1, adversarial augmentations are learned by pushing their feature representations toward the decision boundary (Expansion phase), followed by updating the student model \\( f_{s} \\) to match its prediction on these augmented images with their corresponding soft-pseudo label (Separation phase), yielding better separation of features into their respective classes. In our method setup, we continually learn automatic augmentations of the test images (Sec. 3.3) that are adversarial to the current teacher model \\( f_{t} \\) and enforce consistency between the student model's \\( f_{s} \\) predictions on the augmented views and the corresponding soft-pseudo labels. Since we freeze the classifier module, \\( f_{s} \\) and \\( f_{t} \\) share the common decision boundary. Let \\( \\tilde{\\boldsymbol{x}} \\) denote the learned adversarial augmented view of \\( \\boldsymbol{x} \\) and \\( \\hat{\\boldsymbol{y}} \\) denote the corresponding soft-pseudo label. We minimize the following knowledge distillation loss \\( \\mathcal{L}_{\\mathrm{kd}} \\) using KL-divergence to distill knowledge from the teacher model to the student model:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.665, + 0.47, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k d}} (\\tilde {\\boldsymbol {x}}, \\hat {\\boldsymbol {y}}) = \\mathcal {D} _ {\\mathrm {K L}} (\\hat {\\boldsymbol {y}} \\| f _ {s} (\\tilde {\\boldsymbol {x}})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.472, + 0.772 + ], + "angle": 0, + "content": "Soft pseudo label refinement (PLR). We refine the quality of soft pseudo-labels and the feature representations by averaging the teacher model's outputs on multiple weakly augmented image views \\(\\rho_w(\\pmb{x})\\) (via FLIPPING, CROPPING augmentations) of the same test image \\(\\pmb{x}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.786, + 0.47, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t}, \\mathbf {y} _ {t} \\leftarrow \\mathbb {E} _ {\\mathbf {u} \\in \\rho_ {w} (\\boldsymbol {x})} \\left[ g _ {t} (\\mathbf {u}), h _ {t} \\left(g _ {t} (\\mathbf {u})\\right) \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The refined feature representations \\(\\mathbf{z}_t\\) and the softmaxed pseudo labels \\(\\mathbf{y}_t\\) from the encoder \\(g_t\\) and the classifier \\(h_t\\) are further stored in an online memory queue \\(\\mathbf{Q}\\) of fixed size. The final soft pseudo-label is computed by averaging the refined soft pseudo-labels of \\(n\\)-nearest neighbors of the current test image in the feature space as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.367, + 0.892, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Q} [ \\arg \\max (\\boldsymbol {y} _ {t}) ]. \\text {a p p e n d} \\left(\\left\\{\\boldsymbol {z} _ {t}, \\boldsymbol {y} _ {t} \\right\\}\\right) \\\\ \\hat {\\boldsymbol {y}} = \\frac {1}{n} \\sum \\mathcal {N} _ {\\mathbf {Q}, n} (\\mathbf {z} _ {t}) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.498 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}\\) denotes an online memory of class-balanced queues (with size \\(|\\mathbf{Q}[\\arg \\max (\\pmb {y}_t)]| \\leq N_{\\mathbf{Q}}\\)) of the refined feature representations and the softmaxed label predictions on the previously seen test images, and \\(\\mathcal{N}_{\\mathbf{Q},n}(\\mathbf{z}_t)\\) denotes soft labels of \\(n\\)-nearest neighbors of \\(\\mathbf{z}_t\\) from \\(\\mathbf{Q}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.508, + 0.857, + 0.525 + ], + "angle": 0, + "content": "3.3. Learning Adversarial Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.895, + 0.639 + ], + "angle": 0, + "content": "This section presents our efficient online method for learning adversarial data augmentation for the current teacher \\( f_{t} \\). We first introduce our adversarial augmentation search space. Then, we describe our differentiable strategy to optimize and sample adversarial augmentations that push the feature representations of the test images toward the uncertain region in the close vicinity of the decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.827 + ], + "angle": 0, + "content": "Policy search space. Let \\(\\mathbb{O}\\) be a set of all image transformation operations \\(\\mathcal{O}:\\mathbb{R}^{H\\times W\\times 3}\\to \\mathbb{R}^{H\\times W\\times 3}\\) defined in our search space. In particular, we use image transformations AUTO-CONTRAST, EQUALIZE, INVERT, SOLARIZE, POSTERIZE, CONTRAST, BRIGHTNESS, COLOR, SHEARX, SHEARY, TRANSLATEX, TRANSLATEY, ROTATE, AND SHARPNESS. Each transformation \\(\\mathcal{O}\\) is specified with its magnitude parameter \\(m\\in [0,1]\\). Since the output of some image operations may not be conditional on its magnitude (e.g., EQUALIZE) or may not be differentiable (e.g., POSTERIZE), for such image operations, we use the straight-through gradient estimate w.r.t. each pixel as: \\(\\partial \\mathcal{O}(\\boldsymbol{x}_{ij}) / \\partial m = 1\\) for magnitude optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.894, + 0.872 + ], + "angle": 0, + "content": "We define a sub-policy \\(\\rho\\) as a combination of \\(N\\) image operations (sub-policy dimension) from \\(\\mathbb{O}\\) that is applied sequentially to a given image \\(\\pmb{x}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.885, + 0.893, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\rho \\left(\\boldsymbol {x}, m ^ {\\rho}\\right) = \\mathcal {O} _ {1} ^ {\\rho} \\dots \\mathcal {O} _ {N} ^ {\\rho} \\left(\\boldsymbol {x}, m _ {N} ^ {\\rho}\\right) \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20344" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.138 + ], + "angle": 0, + "content": "where \\(m^{\\rho} = [m_{1}^{\\rho},\\dots,m_{N}^{\\rho}]\\) denotes the magnitude set of the sub-policy \\(\\rho\\). We also define \\(\\mathbb{P} = \\{\\rho_1,\\rho_2,\\ldots \\}\\) as the set of all possible sub-policies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.143, + 0.473, + 0.234 + ], + "angle": 0, + "content": "Policy evaluation. We propose learnable adversarial augmentation, aiming to optimize and sample data augmentations as a proxy for simulating data in the uncertain region of the feature space. Given the teacher model \\( f_{t} \\), a sub-policy \\( \\rho \\) with magnitude \\( m \\) is evaluated for a test image \\( x \\) using the following adversarial objective:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.257, + 0.472, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho) = \\sum_ {k = 1} ^ {K} f _ {t} (\\tilde {\\boldsymbol {x}}) \\log \\left(f _ {t} (\\tilde {\\boldsymbol {x}})\\right) + \\lambda_ {1} r (\\tilde {\\boldsymbol {x}}, \\boldsymbol {x}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.473, + 0.492 + ], + "angle": 0, + "content": "where \\(\\tilde{\\pmb{x}} = \\rho (\\pmb {x},m)\\) denotes the adversarial augmented image and \\(r(\\cdot ,\\cdot)\\) regularizes augmentation severity. The hyperparameter \\(\\lambda_{1}\\) controls the augmentation severity. We optimize the adversarial augmentation loss \\(\\mathcal{L}_{\\mathrm{aug}}\\) with respect to the sub-policy \\(\\rho\\) 's parameters. Minimizing \\(\\mathcal{L}_{\\mathrm{aug}}\\) is equivalent to maximizing the entropy of the teacher model's prediction on the augmented image \\(\\tilde{\\pmb{x}}\\) thus pushing its feature representation towards the decision boundary (first term). For the augmentation regularization (second term), we use the mean squared distance function between the teacher encoder's \\(L\\) internal layers' activations of the adversarial augmentation and non-augmented versions of image \\(\\pmb{x}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.503, + 0.472, + 0.544 + ], + "angle": 0, + "content": "\\[\nr (\\tilde {\\boldsymbol {x}}, \\boldsymbol {x}) = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\| \\mu_ {l} (\\tilde {\\boldsymbol {x}}) - \\mu_ {l} (\\boldsymbol {x}) \\| ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.472, + 0.582 + ], + "angle": 0, + "content": "where \\(\\mu_l(\\cdot)\\) denotes the mean activation of the \\(l^{th}\\) layer of the teacher encoder \\(g_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.587, + 0.473, + 0.664 + ], + "angle": 0, + "content": "Policy optimization. Let \\( M = [m^{\\rho_1}, \\ldots, m^{\\rho_{\\| \\mathbb{P} \\|}}] \\) denote the set of magnitudes \\( m^{\\rho} \\) of all sub-policies \\( \\rho \\in \\mathbb{P} \\) and \\( P = [p_1, \\ldots, p_{\\| \\mathbb{P} \\|}] \\) denote the probability of selecting a sub-policy from \\( \\mathbb{P} \\). The expected policy evaluation loss (Eq. 8) for an image \\( x \\) over the policy search space is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.674, + 0.472, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}) \\right] = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} p _ {i} \\cdot \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.472, + 0.786 + ], + "angle": 0, + "content": "Evaluating the gradient of \\(\\mathbb{E}[\\mathcal{L}_{\\mathrm{aug}}(\\pmb{x})]\\) w.r.t. \\(M\\) and \\(P\\) can become computationally expensive as \\(\\| \\mathbb{P}\\| \\sim \\binom{\\|\\mathbb{O}\\|}{N}\\). Thus, we use the following re-parameterization trick to estimate its unbiased gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.81, + 0.472, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla \\mathbb {E} [ \\mathcal {L} _ {\\mathrm {a u g}} (x) ] = \\delta (\\boldsymbol {x}, \\mathbb {P}) = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} \\nabla \\left(p _ {i} \\cdot \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i})\\right) \\tag {11} \\\\ = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} p _ {i} \\left(\\nabla \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) + \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\cdot \\nabla \\log p _ {i}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.878, + 0.107 + ], + "angle": 0, + "content": "Thus, an unbiased estimator of \\(\\delta (\\pmb {x},\\mathbb{P})\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.119, + 0.894, + 0.138 + ], + "angle": 0, + "content": "\\[\n\\hat {\\delta} (\\boldsymbol {x}, \\rho_ {i}) = \\nabla \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) + \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\cdot \\nabla \\log p _ {i} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.149, + 0.894, + 0.24 + ], + "angle": 0, + "content": "where index \\(i\\) is sampled from the probability distribution \\(P\\). For an online batch of \\(B\\) test images \\(\\{\\pmb{x}_1, \\pmb{x}_2, \\dots, \\pmb{x}_B\\}\\), we apply augmentation sub-policies \\(\\{\\rho_{i_1}, \\rho_{i_2}, \\dots, \\rho_{i_B}\\}\\) from \\(\\mathbb{P}\\) where \\(\\{i_1, i_2, \\dots, i_B\\}\\) are sampled i.i.d from distribution \\(P\\), and update the parameters \\(P\\) and \\(M\\) with the following stochastic gradient update rule:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.252, + 0.893, + 0.295 + ], + "angle": 0, + "content": "\\[\n[ P, M ] \\leftarrow [ P, M ] - \\frac {\\gamma}{B} \\sum_ {j = 1} ^ {B} \\hat {\\delta} \\left(\\boldsymbol {x} _ {j}, \\rho_ {i _ {j}}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.894, + 0.336 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the learning rate. We set \\(\\gamma = 0.1\\) for all experiments without the need for hyperparameter tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.346, + 0.892, + 0.362 + ], + "angle": 0, + "content": "3.4. Self-Learning With Adversarial Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.369, + 0.894, + 0.447 + ], + "angle": 0, + "content": "We minimize the following overall objective \\(\\mathcal{L}_{\\mathrm{TeSLA}}\\) for training the student model \\(f_{s}\\) on a batch of \\(B\\) test images \\(X = \\{\\pmb {x}_1,\\dots ,\\pmb {x}_B\\}\\) and their adversarial augmented views \\(\\hat{X} = \\{\\tilde{x}_1,\\dots ,\\tilde{x}_B\\}\\) with the corresponding refined softpseudo labels from the teacher model \\(\\hat{Y} = \\{\\hat{y}_1,\\dots ,\\hat{y}_B\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.469, + 0.893, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T e S L A}} (X, \\tilde {X}, \\hat {Y}) = \\mathcal {L} _ {\\mathrm {p l}} (X, \\hat {Y}) + \\frac {\\lambda_ {2}}{B} \\sum_ {i = 1} ^ {B} \\mathcal {L} _ {\\mathrm {k d}} (\\tilde {\\boldsymbol {x}} _ {i}, \\hat {\\boldsymbol {y}} _ {i}) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.894, + 0.612 + ], + "angle": 0, + "content": "where \\(\\lambda_{2}\\) is a hyper-parameter for the knowledge distillation. The adversarial augmented views \\(\\tilde{X}\\) are obtained by sampling i.i.d. augmentation sub-policies from \\(\\mathbb{P}\\) for every test image in \\(X\\) with probability \\(P\\) and applying the corresponding magnitude from \\(M\\). Concurrently, we also optimize \\(M\\) and \\(P\\) using the policy optimization mentioned in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.626, + 0.634, + 0.643 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.809, + 0.667 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.675, + 0.894, + 0.795 + ], + "angle": 0, + "content": "We evaluate and compare TeSLA against state-of-the-art (SOTA) test-time adaptation algorithms for both classification and segmentation tasks under three types of test-time distribution shifts resulting from (1) common image corruption, (2) synthetic to real data transfer, and (3) measurement shifts on medical images. The latter is characterized by a change in medical imaging systems, e.g., different scanners across hospitals or various staining techniques." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "TTA protocols. We adopt the TTA protocols in [39] and categorize competing methods based on two factors: (i) source training objective and (ii) sequential or non-sequential inference. First, unlike [39], we use \\(\\mathbf{Y}\\) to indicate if access to the source domain's information, e.g., feature statistics is possible or if the source training objective is allowed to be modified; otherwise, we use \\(\\mathbf{N}\\). Next, we use \\(\\mathbf{O}\\) to indicate a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20345" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "one-pass adaptation and evaluation protocol (one epoch) for sequential test data and \\(\\mathbf{M}\\) to show a multi-pass adaptation on all test data and inference after several epochs. Thus, we end up with four possible TTA protocols, namely N-M, Y-M, N-O, and Y-O. Unlike previous works [28,39], TeSLA does not rely on any source domain feature distribution, yielding two realistic TTA protocols: N-M and N-O." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.472, + 0.336 + ], + "angle": 0, + "content": "Hyperparameters for all experiments. We set \\(\\lambda_1 = \\lambda_2 = 1.0\\) for all experiments (cf. ablation in Appendix E). For the adversarial augmentation module, we use sub-policy dimension \\(N = 2\\) (except for VisDA-C [33], \\(N = 4/3\\) for N-O/N-M protocols) (cf. Appendix E) and Adam optimizer [22]. For pseudo-label refinement (PLR), we use five weak augmentations composed of random flips and resize crop. More details of hyperparameters used for each experiment can be found in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.472, + 0.502 + ], + "angle": 0, + "content": "Common image corruptions. To evaluate TeSLA's efficacy for the classification task, we use CIFAR10-C/CIFAR100-C [16] and large-scale ImageNet-C [16] datasets each containing 19 types of corruptions applied to the clean test set with five levels of severity. We perform validation on 4 out of 19 types of corruption (SPATTER, GAUSSIAN BLUR, SPECKLE NOISE, SATURATE) to select hyperparameters and test on the remaining 15 corruptions at the maximum severity level of 5. Following [28], we train the ResNet50 [15] on the clean CIFAR10/CIFAR100/ImageNet training set and adapt it to classify the unlabeled corrupted test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.472, + 0.642 + ], + "angle": 0, + "content": "Synthetic to real data adaption. We use challenging and large-scale VisDA-C, and VisDA-S [33] datasets for evaluating synthetic-to-real data adaptation at test-time for classification and segmentation tasks, respectively. Following [4, 26], we adapt the ResNet101 network pre-trained on synthetic images to classify 12 vehicle classes on the photo-realistic images of VisDA-C. While for VisDA-S, we adapt the DeepLab-v3 [5] backbone pre-trained on synthetic GTA5 [36] to the Cityscapes [7] to segment 19 classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.472, + 0.854 + ], + "angle": 0, + "content": "Measurement shifts on medical images. We access TeSLA on staining variations for the tissue type classification of hematoxylin & eosin (H&E) stained patches from colorectal cancer tissue slides. We use MobileNetV2 [38] trained on the source Kather-19 dataset [20] and adapt it to the target Kather-16 dataset [21] on four tissue categories: tumor, stroma, lymphocyte, and mucosa. We also evaluate TeSLA for variations in scanners and imaging protocols in multi-site medical images on two magnetic resonance imaging (MRI) datasets for the segmentation task, namely the multi-site prostate MRI [27] and spinal cord grey matter segmentation (SCGM) [34]. Following [42], we adapt the U-Net [37] from site 1 to sites \\(\\{2,3,4\\}\\) of the spinal cord dataset, and from sites \\(\\{\\mathrm{A,B}\\}\\) to sites \\(\\{\\mathrm{D,E,F}\\}\\) of the prostate dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Competing baselines. We compare TeSLA with the following SFDA and TTA baselines, including direct inference of the trained source model on the target test data without" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.895, + 0.175 + ], + "angle": 0, + "content": "Table 1. Comparison of SOTA TTA methods under different protocols evaluated on CIFAR-10/100-C, ImageNet-C, VisDA-C and Kather-16 datasets. We report the average error computed over 15 test corruptions for the common image corruption shifts. We also report Class Avg. in % error rates for synthetic-to-real and measurement shifts over 3 and 10 seeds, respectively." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.184, + 0.895, + 0.326 + ], + "angle": 0, + "content": "
MethodProtocolCommon Image CorruptionsSyn-to-RealMeasurement Shift
CIFAR10-CCIFAR100-CImageNet-CVisDA-CKather-16
OMOMOMOMOM
SourceN29.160.481.851.532.0
BN [18,30]N15.615.443.743.367.767.635.435.018.318.2
TENT [45]N14.112.939.036.557.454.233.529.316.212.0
SHOT [26]N13.914.239.238.768.768.229.424.514.712.0
AdaContrast [4]N-23.120.2-
TTT++ [28]Y15.89.844.434.159.3-35.234.116.77.9
TTAC [39]Y13.49.441.733.658.7-32.231.19.65.5
TeSLAN12.59.738.232.955.051.517.813.59.23.3
TeSLA-sY12.19.737.332.653.1-24.017.99.93.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.897, + 0.699 + ], + "angle": 0, + "content": "adaptation (Source). We also implement the SFDA-based pseudo-labeling baselines from a basic pseudo-labeling approach (PL) to the SHOT approach [26] based on training the feature extraction module by adopting the information maximization loss to make globally diverse but individually certain predictions on the target domain. For recent TTA methods, we compare our method against \\(\\mathrm{TTT}++\\) [28], Ada-Contrast [4], CoTTA [47], basic TENT [45] and BN [18] as representative methods based on feature distribution alignment via stored statistics, contrastive learning, augmentation-averaged predictions, entropy minimization, and batch normalization statistic. Furthermore, we compare TeSLA to more recent methods based on test-time augmentation policy learning, OptTTA [42], and anchor clustering TTAC [39]. By default, [4,18,26] follow the N-M protocol, and we adapt them to the N-O protocol, while [42,45,47], by default, follow the N-O protocol and adapt them to the N-M protocol setting. Using source features distribution, TTAC and \\(\\mathrm{TTT}++\\) follow Y-O and Y-M. We report error rates on classification tasks and Mean Intersection over Union (mIoU)/ Dice scores on segmentation tasks. For consistency across baselines and TTA protocols, we do not use any specialized model architectures (e.g., projection head, weight-normalized classifier layer) for SHOT [26] and AdaContrast [4]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.71, + 0.593, + 0.725 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Classification task. In Table 1, we summarize the average (Avg) classification error rates \\((\\%)\\) of TeSLA against several SOTA SFDA and TTA baselines on the common image corruptions of CIFAR10-C, CIFAR100-C, ImageNet-C; synthetic to real data domain shifts of VisDA-C; and measurement shifts on the Kather-16 dataset. We also report per-class top-1 accuracies for VisDA-C and Kather-16 datasets and corruption-wise error rates on CIFAR-10-C/CIFAR-100-C and ImageNet-C datasets in Appendix B. TeSLA easily surpasses all competing methods under the protocols N-O and N-M on all the datasets. In particular, TeSLA outper" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20346" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.09, + 0.473, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.215, + 0.319, + 0.226 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.087, + 0.631, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.087, + 0.777, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.087, + 0.822, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.215, + 0.667, + 0.226 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Figure 3. Ablation experiments on the VisDA-C. (a) Source training strategy and model architecture. TeSLA outperforms the competing TTA methods across varied source training strategies (Supervised, SimCLRV2 [6], and DINO [3]) and model architectures (ResNet50, ResNet101 [15], and ViT-B/16 [11]) using the same set of hyperparameters for both protocols (N-O/N-M). Each vertex represents the mean class avg accuracy over 3 seeds. (b) t-SNE visualization [43] comparison of feature embeddings for AdaContrast [4] and TeSLA." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.313, + 0.268, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.313, + 0.471, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.452, + 0.471, + 0.522 + ], + "angle": 0, + "content": "Figure 4. Model calibration and uncertainty estimation. Model calibration using reliability diagrams [31] and model uncertainty estimation using Brier Score and Negative Log Likelihood (NLL) of the Source Model, TENT [45], OptTTA [42] and, TeSLA on the prostate dataset for test-time adaptation." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.53, + 0.471, + 0.559 + ], + "angle": 0, + "content": "Table 2. Semantic segmentation results (Class Avg. mIoU in %) on the VisDA-S (GTA5 → Cityscapes) test-time adaptation task." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.563, + 0.449, + 0.605 + ], + "angle": 0, + "content": "
ProtocolSourceBN [30]TENT [45]PLCoTTA [47]TeSLA
NO35.336.738.338.837.044.5
NM38.439.238.639.946.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.84 + ], + "angle": 0, + "content": "forms the second-best baseline under N-O/N-M protocol by a \\((\\%)\\) margin 1.4/3.2 on CIFAR10-C, 0.8/3.6 on CIFAR100-C, 2.4/2.7 on ImageNet-C, 5.3/6.7 on VisDA-C, and 5.5/8.7 on the Kather-16 datasets. Despite not utilizing any source feature statistic, TeSLA (N-O/N-M) is either competitive or surpasses TTAC [39] and TTT++ [28] (Y-O/Y-M) that benefit from the source dataset feature statistics during adaptation. Including the global source feature alignment module of TTAC with TeSLA (TeSLA-s) could further improve its performance under (Y-O/Y-M) protocols for slight domain shifts (e.g., common image corruptions). We report the results of AdaContrast for VisDA-C only as the implementation for other datasets are not provided. We do not report TeSLA-s results on ImageNet-C under the Y-M protocol, as previous baselines were evaluated only Y-O protocol." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Segmentation task. Unlike TeSLA, several TTA methods e.g., AdaContrast [4], SHOT [26], TTT++ [28], TTAC [39], have only been evaluated on the classification task. Therefore, we compare TeSLA against the current SOTA TTA" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.313, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Table 3. Semantic segmentation results (Class Avg. volume-wise mean Dice in %) on the cross-site spinal cord and prostate MRI test-time adaptation tasks." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.365, + 0.894, + 0.462 + ], + "angle": 0, + "content": "
ProtocolSourceBN [30]TENT [45]PLOptTTA [42]TeSLA
Spinal CordSite{1} → Sites{2,3,4}
NO76.0±11.881.6±8.381.1±9.181.7±8.684.1±4.885.3±5.8
NM84.3±4.884.4±4.784.3±4.784.3±4.485.4±4.4
ProstateSites{A,B} → Sites{D,E,F}
NO60.5±27.072.1±15.274.7±17.972.4±15.283.1±7.783.5±6.5
NM73.1±18.081.2±9.381.1±9.283.4±7.784.3±5.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.895, + 0.652 + ], + "angle": 0, + "content": "methods applied to the segmentation task on synthetic-to-real data transfer of the VisDA-S dataset in Table 2. TeSLA significantly outperforms the competing methods and achieves the best mIOU scores for both N-O and N-M protocols, beating the second-best baseline by a margin of \\(+5.7\\%\\) and \\(+6.1\\%\\), respectively. In Table 3, we compare TeSLA against the recent SOTA test-time augmentation policy method, OptTTA [42] and other TTA baselines for the inter-site adaptation on two challenging MRI datasets mentioned in Sec. 4.1. TeSLA convincingly outperforms all other methods on severe measurement shifts across sites under N-O and N-M protocols." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.894, + 0.759 + ], + "angle": 0, + "content": "Computational cost for adversarial augmentations. As shown in Table 4, TeSLA learns the adversarial augmentation in an online manner, and thus its runtime is several orders faster than learnable test-time augmentation OptTTA [42], while comparable to that of static augmentation policies with an additional overhead of only 0.10 GPU hours/epoch on the VisDA-C dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.894, + 0.838 + ], + "angle": 0, + "content": "Test-time feature visualization. Fig. 3b compares t-SNE projection [43] of the encoder's features of AdaContrast [4] with TeSLA on the VisDA-C. TeSLA shows better inter-class separation for features than AdaContrast, as supported by an improved silhouette score from 0.149 to 0.271." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Model calibration and uncertainty estimation. Entropy minimization-based TTA methods [26, 45] explicitly make the model confident in their predictions, resulting in poor model calibration. For trusting the adapted model, it should" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20347" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.146 + ], + "angle": 0, + "content": "Table 4. Performance comparison of adversarial augmentation learned by TeSLA during adaptation against the prior art augmentation methods on runtime (GPU hours/epoch on NVIDIA GeForce RTX 3090) and Class Avg. accuracy in % on the VisDA-C dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.152, + 0.466, + 0.219 + ], + "angle": 0, + "content": "
Augmentation TypeOptTTATeSLARA [10]TeSLAAA [8]TeSLAN=2TeSLAN=3
Runtime4.500.050.060.160.18
Class Avg. Acc. (N-O)-80.281.281.582.2
Class Avg. Acc. (N-M)-84.785.786.386.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.471, + 0.412 + ], + "angle": 0, + "content": "output reliable confidence estimates matching its true underlying performance on the test images. Such a well-behaved model is characterized by expected calibration error (ECE) through reliability diagram [31], and uncertainty metrics of brier score and negative log-likelihood (NLL) [14]. In Fig. 4, we plot the reliability diagram (dividing the probability range [0, 0.5] into 10 bins), and uncertainty metrics of the Source model, TENT [45] and OptTTA [42] against TeSLA for the inter-site model adaptation on the prostate dataset [27]. We observe that TeSLA achieves the best model calibration (lowest ECE of \\(1.80\\%\\)) and lowest Brier and NLL scores of 0.12 and 0.24 (\\(p < 0.006\\) against TENT)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.241, + 0.436 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.627 + ], + "angle": 0, + "content": "In this section, we scrutinize the roles played by different components of TeSLA. All ablations are performed on the synthetic-to-real test data adaptation of the VisDA-C dataset. Loss functions for self-learning. We compare TTA performance of our self-learning objective \\(\\mathcal{L}_{\\mathrm{pl}}\\) with the proposed flipped cross-entropy loss \\(f\\)-CE and the basic cross-entropy loss \\(\\mathbb{C}\\mathbb{E}\\) in Fig. 5-(1). We do not use the proposed augmentation module and pseudo-label refinement in this experiment. We observe that \\(f\\)-CE alone improves the accuracy of the Source model in the N-M protocol by \\(+23.1\\%\\) compared to \\(+16.6\\%\\) improvement by \\(\\mathbb{C}\\mathbb{E}\\). Using \\(\\mathcal{L}_{\\mathrm{pl}}\\) (Eq. 3) further improves the accuracy to a margin of \\(+23.8\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Contribution of individual components: PLR, \\(\\mathcal{L}_{\\mathrm{pl}}\\), and \\(\\mathcal{L}_{\\mathrm{kd}}\\). Our test-time objective \\(\\mathcal{L}_{\\mathrm{TeSLA}}\\) has three components (i) self-learning loss \\(\\mathcal{L}_{\\mathrm{pl}}\\), (ii) pseudo-label refinement (PLR), and (iii) knowledge distillation \\(\\mathcal{L}_{\\mathrm{kd}}\\) with adversarial augmentation. In Fig. 5-(3), we study the effect of accruing individual components to \\(\\mathcal{L}_{\\mathrm{TeSLA}}\\) (starting from source trained model) for the test-time adaptation accuracy. We observe that \\(\\mathcal{L}_{\\mathrm{pl}}\\) alone improves the model's accuracy from \\(48.5\\%\\) to \\(72.3\\%\\) while adding PLR and \\(\\mathcal{L}_{\\mathrm{kd}}\\) further boosts it to \\(81.6\\%\\) \\((+9.3)\\) and \\(86.3\\%\\) \\((+4.7)\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "PLR: ensembling on weak augmentations (Ens) and nearest neighbor averaging (NN). We also conduct an ablation on the two pseudo-label refinement approaches - (i) ensembling the teacher model's output on five weak augmentations. (ii) averaging soft pseudo labels of \\( n = 10 \\) nearest neighbors in the feature space (cf. sensitivity test in Appendix E). Fig. 5-(2) shows that combining both approaches gives better results than using them individually." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.63, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.092, + 0.753, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.092, + 0.882, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.205, + 0.894, + 0.289 + ], + "angle": 0, + "content": "Figure 5. Ablation experiments on the VisDA-C. (left to right) (1) Ablation results for self-learning loss functions, including \\( f\\text{-}\\mathbb{C}\\mathbb{E} \\), \\( \\mathbb{C}\\mathbb{E} \\), and \\( \\mathcal{L}_{\\mathrm{pl}} \\); (2) ablation on different PLR methods: assembling predictions on 5 weak augmentations, averaging predictions of the 10 nearest neighbors in the feature space; (3) ablation on different components of TeSLA's objective function: \\( \\mathcal{L}_{\\mathrm{pl}} \\), PLR, and \\( \\mathcal{L}_{\\mathrm{kd}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.894, + 0.435 + ], + "angle": 0, + "content": "Adversarial augmentation. Table 4 shows the benefit of using our adversarial augmentation \\((N = 2,3)\\) with TeSLA against (i) RandAugment (RA) [10], (ii) AutoAugment (AA) [9] (optimized for ImageNet), and (iii) OptTTA [42]. We achieve far better Class Avg. accuracy on the VisDA-C at the cost of slight runtime overhead compared to static augmentation policies for protocols (N-O, N-M). In addition, TeSLA augmentation (TeAA) consistently improves the performance of other TTA methods (Table 5)." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.476 + ], + "angle": 0, + "content": "Table 5. Class Avg. accuracy (%) of TeSLA augmentation (TeAA) with other TTA objectives under N-O protocol on VisDA-C dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.482, + 0.871, + 0.519 + ], + "angle": 0, + "content": "
TENT+TeAA| SHOT+TeAA| AdaContrast+TeAA
66.572.0| 70.673.1| 76.979.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.894, + 0.664 + ], + "angle": 0, + "content": "Source training strategies and model architectures. The ablation results (Fig. 3a) on the VisDA-C show that TeSLA surpasses competing TTA baselines for different source training strategies, including Supervised and self-supervised (SimCLRV2 [6], DINO [3]). Moreover, we ablate TeSLA over different architectures, ranging from ResNet-50/101 [15] to the recent vision transformer ViT-B [11] using the same set of hyperparameters, showing the merits of no reliance on the network architecture or source training strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Finally, in Appendix E, we provide detailed sensitivity results with respect to all hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We introduced TeSLA, a novel self-learning algorithm for test-time adaptation that utilizes automatic adversarial augmentation. TeSLA is agnostic to the model architecture and source training strategies and gives better model calibration than other TTA methods. Through extensive experiments on various domain shifts (measurement, common corruption, synthetic to real), we show TeSLA's superiority over previous TTA methods on classification and segmentation tasks. Note that TeSLA assumes class uniformly when implicitly maximizing mutual information and can be improved by incorporating prior, e.g., class label distribution statistics." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20348" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009-2018, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.472, + 0.57 + ], + "angle": 0, + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[8] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.758 + ], + "angle": 0, + "content": "[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[12] Cian Eastwood, Ian Mason, Christopher K. I. Williams, and Bernhard Scholkopf. Source-free adaptation to measurement shift via bottom-up feature restoration. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[13] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[14] Alvaro Gomariz, Tiziano Portenier, César Nombela-Arrieta, and Orcun Goksel. Probabilistic spatial analysis in quantitative microscopy with uncertainty-aware cell detection using deep bayesian regression of density maps. arXiv preprint arXiv:2102.11865, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.894, + 0.383 + ], + "angle": 0, + "content": "[17] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.894, + 0.579 + ], + "angle": 0, + "content": "[20] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.894, + 0.635 + ], + "angle": 0, + "content": "[21] Jakob Nikolas Kather, Frank Gerrit Zollner, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Cleo-Aron Weis. Collection of textures in colorectal cancer histology, May 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.894, + 0.676 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.894, + 0.734 + ], + "angle": 0, + "content": "[23] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[24] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12376-12385, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[25] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20349" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.471, + 0.133 + ], + "angle": 0, + "content": "unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[27] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. Msnet: Multi-site network for improving prostate segmentation with heterogeneous mri data. IEEE Transactions on Medical Imaging, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[28] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[29] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy policy search: A simple baseline for learnable test-time augmentation. In Conference on Uncertainty in Artificial Intelligence, pages 1308-1317. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.472, + 0.411 + ], + "angle": 0, + "content": "[30] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[31] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In Proceedings of the 22nd international conference on Machine learning, pages 625-632, 2005. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.537 + ], + "angle": 0, + "content": "[32] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[33] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[34] Ferran Prados, John Ashburner, Claudia Blaiotta, Tom Brosch, Julio Carballido-Gamio, Manuel Jorge Cardoso, Benjamin N Conrad, Esha Datta, Gergely David, Benjamin De Leener, et al. Spinal cord grey matter segmentation challenge. Neuroimage, 152:312-329, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[35] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In International Joint Conference on Artificial Intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[36] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, European Conference on Computer Vision (ECCV), volume 9906 of LNCS, pages 102-118. Springer International Publishing, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.895, + 0.26 + ], + "angle": 0, + "content": "[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.895, + 0.317 + ], + "angle": 0, + "content": "[41] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.895, + 0.386 + ], + "angle": 0, + "content": "[42] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozortbar. OptTTA: Learnable test-time augmentation for source-free medical image segmentation under domain shift. In Medical Imaging with Deep Learning, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.43, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[44] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.529, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[46] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 338:34-45, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[47] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.71 + ], + "angle": 0, + "content": "[48] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Casting a bait for offline and online source-free domain adaptation. arXiv preprint arXiv:2010.12427, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[49] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8978-8987, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.769, + 0.892, + 0.823 + ], + "angle": 0, + "content": "[50] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. arXiv preprint arXiv:2205.04183, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.825, + 0.895, + 0.893 + ], + "angle": 0, + "content": "[51] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29:7834-7844, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20350" + } + ] +] \ No newline at end of file diff --git a/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_origin.pdf b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bdd0dfef9ec077624e2df383c6ee622129e2391b --- /dev/null +++ b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/ad821131-dbc0-46a1-909c-72d909b80828_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6894ea0d1bf837c0bc2d584e223640ed09f6f22a83409defc832410565eba71 +size 1458069 diff --git a/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/full.md b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b5e0523274075b4fef888dc78406433748d69622 --- /dev/null +++ b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/full.md @@ -0,0 +1,325 @@ +# TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation + +Devavrat Tomar1 Guillaume Vray1 Behzad Bozorgtabar1,2 Jean-Philippe Thiran1,2 EPFL CHUV + +$^{1}\{$ firstname\}. $\{$ lastname\}@epfl.ch + +# Abstract + +Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with automatic Adversarial augmentation dubbed TeSLA for adapting a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning methods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowledge distillation by simulating high entropy augmented images. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several desirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available at https://github.com/devavratTomar/TeSLA. + +# 1. Introduction + +Deep neural networks (DNNs) perform exceptionally well when the training (source) and test (target) data follow the same distribution. However, distribution shifts are inevitable in real-world settings and propose a major challenge to the performance of deep networks after deployment. Also, access to the labeled training data may be infeasible at test time due to privacy concerns or transmission bandwidth. In such scenarios, source-free domain adaptation (SFDA) [1, 23, 25] and test-time adaptation (TTA) methods [19, 28, 39] aim to adapt the pre-trained source model to the unlabeled distributionally shifted target domain while easing access to source data. While SFDA methods have access to all full target data through multiple training epochs + +![](images/606c4613313034c2fb162c2724bbb58c1b37abc42126bae8d18860dca17451d4.jpg) +Figure 1. Knowledge distillation with adversarial augmentations. (a) Easy images with confident soft-pseudo labels and (b) Hard images with unconfident soft-pseudo labels are adversarily augmented and pushed to the uncertainty region (high entropy) near the decision boundary. The model is updated for (a) to match its output on the augmented views with non-augmented views of Easy test images using KL-Divergence $\mathcal{L}_{\mathrm{kd}} \neq 0$ , while not updated for (b) as $\mathcal{L}_{\mathrm{kd}} \sim 0$ between Hard images and their augmented views. + +![](images/792aacf96515bab2dd2ea598a8a614d2dd2a04833e3da270cdc529d583ed32b4.jpg) + +![](images/76aca1b51c9e128429a15268ad368e38b13e9cc94b8f01981ddd8c2d8351a6c9.jpg) + +(offline setup), TTA methods usually process test images in an online streaming fashion and represent a more realistic domain adaptation. However, most of these methods are applied: (i) only to classification tasks, (ii) evaluated on the non-real-world domain shifts, e.g., the non-measurement shift; (iii) destroy model calibration—entropy minimizing with overconfident predictions [45] on incorrectly classified samples, and (iv) use specialized network architectures or rely on the source dataset feature statistics [28]. + +We address these issues by proposing a new test-time adaptation method with automatic adversarial augmentation called TeSLA, under which we further define realistic TTA protocols. Self-learning methods often supervise the model adaptation on the unlabeled test images using their predicted pseudo-labels. As the model can easily overfit on its own pseudo-labels, a weight-averaged teacher model (slowly updated by the student model) is employed for obtaining the pseudo-labels [41, 47]. The student model is then trained with cross-entropy $(\mathbb{C}\mathbb{E})$ loss between the one-hot pseudolabels and its predictions on the test images. In this paper, we instead propose to minimize flipped cross-entropy between the student model's predictions and the soft pseudo-labels (notice the reverse order) with the negative entropy of its marginalized predictions over the test images. In Sec. 3, we show that the proposed formulation is an equivalence to mutual information maximization implicitly corrected + +by the teacher-student knowledge distillation via pseudolabels, yielding performance improvement on various test-time adaptation protocols and compared to the basic $\mathbb{C}\mathbb{E}$ optimization (cf. ablation in Fig. 5-(1)). + +Motivated by teacher-student knowledge distillation, another central tenet of our method is to assist the student model during adaptation in improving its performance on the hard-to-classify (high entropy) test images. For this purpose, we propose learning automatic adversarial augmentations (see Fig. 1) as a proxy for simulating images in the uncertainty region of the feature space. The model is then updated to ensure consistency between predictions of high entropy augmented images and soft-pseudo labels from the respective non-augmented versions. Consequently, the model is self-distilled on Easy test images with confident soft-pseudo labels (Fig. 1a). In contrast, the model update on Hard test images is discarded (Fig. 1b), resulting in better class-wise feature separation. + +In summary, our contributions are: (i) we propose a novel test-time self-learning method based on flipped cross-entropy $(f\text{-}\mathbb{C}\mathbb{E})$ through the tight connection with the mutual information between the model's predictions and the test images; (ii) we propose an efficient plug-in test-time automatic adversarial augmentation module used for online knowledge distillation from the teacher to the student network that consistently improves the performance of test-time adaptation methods, including ours; and (iii) TeSLA achieves new state-of-the-art results on several benchmarks, from common image corruption to realistic measurement shifts for classification and segmentation tasks. Furthermore, TeSLA outperforms existing TTA methods in terms of calibration and uncertainty metrics while making no assumptions about the network architecture and source domain's information, e.g., feature statistics or training strategy. + +# 2. Related Work + +In general, domain adaptation methods aim to distill knowledge from source data that are well-generalizable to target data and relax the assumption of i.i.d. between source and target datasets. To circumvent expensive and cumbersome annotation of new target data, unsupervised domain adaptation (UDA) emerges, and there has been a large corpus of UDAs [13, 17, 32, 51] to match the distribution of domains on both labeled source data and unlabeled target data. Nonetheless, the above-mentioned methods demand source data to achieve the domain adaptation process, which is often impractical in real-world scenarios, e.g., due to privacy restrictions. The above issue motivates research into source-free domain adaptation (SFDA) [1, 23, 25, 26] and test-time training or adaptation (TTA) [19, 28, 39, 45], which are more closely relevant to our problem setup. + +SFDA approaches formulate domain adaptation through pseudo labeling [23, 26], target feature clustering [49], syn + +thesizing extra training samples [24], or feature restoration [12]. SHOT [26] proposed feature clustering via information maximization while incorporating pseudo-labeling as additional supervision. BAIT [48] leverages the fixed source classifier as source anchors and uses them to achieve feature alignment between the source and target domain. CPGA [35] proposed the SFDA method based on matching feature prototypes between source and target domains. AaD-SFDA [50] proposed to optimize an objective by encouraging prediction consistency of local neighbors in feature space. Nevertheless, SFDA methods require apriori access to all target data in advance, and the current pseudo-labeling approach, e.g., SHOT [26], used offline pseudo-label refinement on a perepoch basis. In a more realistic domain adaptation scenario, SFDA is still incompetent to perform inference and adaptation simultaneously. + +TTA methods [4, 45] propose alleviating the domina shift by online (or streaming) adapting a model at test time. Still, we argue that there are ambiguities over the problem setup of TTA in the literature, particularly on whether sequential inference on target test data is feasible upon arrival [19, 45] or whether training objectives must be adapted [28, 40]. TTT [40] proposed fine-tuning the model parameters via rotation classification task as a proxy. On-target adaptation [44] used pseudo-labeling and contrastive learning to initialize the target-domain feature; each performed independently in their method. T3A [19] utilized pseudo-labeling to adjust the classifier prototype. More recently, TTAC [39] leveraged the clustering scheme to match target domain clusters to source domain categories. Nevertheless, existing TTA methods rely on specialized neural network architectures or only update a fraction of model weights yielding limited performance gain on the target data. For instance, TENT [45] proposed updating affine parameters in the batchnorm layers of convolutional neural networks (CNN), while AdaContrast [4] and SHOT [26] used an additional weight normalization classification layer with a projection head. In contrast, we show our method's superiority for various neural network architectures, including CNNs and vision transformers (ViTs) [11], without additional architectural requirements or different source model training strategies. Moreover, we update all model parameters, and our method is stable over a wide range of hyper-parameters, e.g., learning rate. + +Test time augmentation methods [2, 29] are another popular line of domain adaptation research. GPS [29] learns optimal augmentation sub-policies by combining image transformations of RandAugment [10] that minimize calibrated log-likelihood loss [2] on the validation set. OptTTA [42] optimized the magnitudes of augmentation sub-policies using gradient descent to maximize mutual information and match feature statistics over augmented images. Though effective, these methods [29, 42, 46] are computationally expensive as all augmentation sub-policies need to be evaluated, making + +their real-time application difficult. Instead, our proposed adversarial augmentation policies can be learned online and are several orders faster than the existing learnable test-time augmentation strategies, e.g., [42]. + +# 3. Methodology + +Overview. We first introduce our flipped cross-entropy $(f\text{-}\mathbb{C}\mathbb{E})$ loss through the tight connection with the mutual information between the model's predictions and the test images. (Sec. 3.1). Using this equivalence, we derive our test-time loss function that implicitly incorporates teacher-student knowledge distillation. In Sec. 3.2, we propose to enhance the test-time teacher-student knowledge distillation by utilizing the consistency of the student model's predictions on the proposed adversarial augmentations with their corresponding refined soft-pseudo labels from the teacher model. We refine the soft-pseudo labels by averaging the teacher model's predictions on (1) weakly augmented test images; (2) nearest neighbors in the feature space. Furthermore, we propose an efficient online algorithm for learning adversarial augmentations (Sec. 3.3). Finally, in Sec. 3.4, we summarize our online test-time adaptation method, TeSLA, based on self-learning with the proposed automatic adversarial augmentation. The overall framework is shown in Fig. 2. + +Problem setup. Given an existing model $f_{\theta_0}$ , parametrized by $\theta_0$ and pre-trained on the inaccessible source data, we aim to improve its performance by updating all its parameters using the unlabeled streaming test data $\mathbf{X}$ during adaptation. Motivated by the concept of mean teacher [41] as the weight-averaged model over training steps that can yield a more accurate model, we build our online knowledge distillation framework upon teacher-student models. We utilize identical network architecture for teacher and student models. Each model $f = h \circ g$ comprises a backbone encoder $g: \mathbf{X} \to \mathbb{R}^D$ mapping unlabeled test image $\mathbf{x} \in \mathbf{X}$ to the feature representation $\mathbf{z} = g(\mathbf{x}) \in \mathbb{R}^D$ and a classifier (hypothesis) head $h: \mathbb{R}^D \to \mathbb{R}^K$ mapping $\mathbf{z}$ to the class prediction $\mathbf{y} \in \mathbb{R}^K$ , where $D$ and $K$ denote the feature dimension and number of classes. The parameters of teacher $\theta_t$ and student $\theta_s$ are first initialized from the source model parameters $\theta_0$ . Then the teacher model's parameters $\theta_t$ are updated from the student model's parameters $\theta_s$ using an exponential moving average (EMA) with momentum coefficient $\alpha$ as: $\theta_t \gets \alpha \cdot \theta_t + (1 - \alpha) \cdot \theta_s$ . Following [26], we freeze the classifier $h$ 's parameters and only update the encoder $g$ 's parameters during test-time adaptation. + +# 3.1. Rationale Behind the $f$ -CE Objective + +We start by analyzing the rationale behind the proposed $f$ -CE loss and its benefit for self-learning through the tight connection with the mutual information between the model's predictions and unlabeled test images. Before that, we first define the notations. + +Notations. We define the random variables of unlabeled test images and the predictions from the student model $f_{s}$ as $\mathbf{X}$ and $\mathbf{Y}$ and those of the teacher model $f_{t}$ 's soft pseudo-label predictions as $\hat{\mathbf{Y}}$ . Furthermore, let $p_{\mathbf{Y}}$ denotes the marginal distribution over $\mathbf{Y}$ , $p_{(\mathbf{Y},\mathbf{X})}$ be the joint distribution of $\mathbf{Y}$ and $\mathbf{X}$ , and $p_{\mathbf{Y}|\mathbf{X}}$ be the conditional distribution of $\mathbf{Y}$ given $\mathbf{X}$ . Then, the entropy of $\mathbf{Y}$ and the conditional entropy of $\mathbf{Y}$ given $\mathbf{X}$ can be defined as $\mathcal{H}(\mathbf{Y}) := \mathbb{E}_{p_{\mathbf{Y}}}[-\log p_{\mathbf{Y}}(\mathbf{Y})]$ and $\mathcal{H}(\mathbf{Y}|\mathbf{X}) := \mathbb{E}_{p_{(\mathbf{Y},\mathbf{X})}}[-\log p_{\mathbf{Y}|\mathbf{X}}(\mathbf{Y}|\mathbf{X})]$ , respectively. Besides, let the flipped cross-entropy ( $f-\mathbb{C}\mathbb{E}$ ) given $\mathbf{X}$ be $\mathcal{H}(\mathbf{Y};\hat{\mathbf{Y}}|\mathbf{X}) := \mathbb{E}_{p_{(\mathbf{Y},\mathbf{X})}}[-\log p_{\hat{\mathbf{Y}}|\mathbf{X}}(\mathbf{Y}|\mathbf{X})]$ . + +Self-learning and mutual information. Based on the above definitions, the $f$ - $\mathbb{C}\mathbb{E}$ between $\mathbf{Y}$ and $\hat{\mathbf{Y}}$ conditioned over $\mathbf{X}$ has the tight connection with the mutual information $\mathcal{I}(\mathbf{Y};\mathbf{X})$ between unlabeled test images $\mathbf{X}$ and the predictions from the student model $\mathbf{Y}$ as follows: + +$$ +\begin{array}{l} \mathcal {H} (\mathbf {Y}; \hat {\mathbf {Y}} \mid \mathbf {X}) = \mathcal {H} (\mathbf {Y} \mid \mathbf {X}) + \mathcal {D} _ {\mathrm {K L}} (\mathbf {Y} \| \hat {\mathbf {Y}} \mid \mathbf {X}) \tag {1} \\ = - \mathcal {I} (\mathbf {Y}; \mathbf {X}) + \mathcal {H} (\mathbf {Y}) + \mathcal {D} _ {\mathrm {K L}} \left(\mathbf {Y} \| \hat {\mathbf {Y}} | \mathbf {X}\right) \\ \end{array} +$$ + +This implies that minimizing the $f$ -CE $(\mathcal{H}(\mathbf{Y};\hat{\mathbf{Y}}\mid \mathbf{X})$ in Eq. 1) and maximizing the entropy of class-marginal prediction $\mathcal{H}(\mathbf{Y})$ is equivalent to maximizing the mutual information between the test images and the student model's predictions $\mathcal{I}(\mathbf{Y};\mathbf{X})$ with a correction KL-divergence term $\mathcal{D}_{\mathrm{KL}}(\mathbf{Y}\parallel \hat{\mathbf{Y}}\mid \mathbf{X})$ involving soft-pseudo labels as: + +$$ +\underbrace {\mathcal {H} (\mathbf {Y} ; \hat {\mathbf {Y}} \mid \mathbf {X})} _ {f \cdot \mathbb {C E}} - \mathcal {H} (\mathbf {Y}) = - \underbrace {\mathcal {I} (\mathbf {Y} ; \mathbf {X})} _ {\text {M u t u a l I n f o .}} + \mathcal {D} _ {\mathrm {K L}} (\mathbf {Y} \| \hat {\mathbf {Y}} | \mathbf {X}) \tag {2} +$$ + +Thus, minimizing the left side of Eq. 2 with respect to the student model's parameters $\theta_{s}$ allows the student model to cluster test images using mutual information and criterion-corrected knowledge distillation from soft-pseudo labels. Using the above formulation, we define the following test-time objective to train the student model $f_{s}$ on a batch of $B$ test images $X = \{\pmb{x}_1,\dots ,\pmb{x}_B\}$ with their corresponding soft-pseudo labels from teacher model $\hat{Y} = \{\hat{y}_1,\dots ,\hat{y}_B\}$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {p l}} (X, \hat {Y}) = - \frac {1}{B} \sum_ {i = 1} ^ {B} \sum_ {k = 1} ^ {K} f _ {s} \left(\boldsymbol {x} _ {i}\right) _ {k} \log \left(\left(\hat {\boldsymbol {y}} _ {i}\right) _ {k}\right) \\ + \sum_ {k = 1} ^ {K} \hat {f} _ {s} (X) _ {k} \log (\hat {f} _ {s} (X)) _ {k} \tag {3} \\ \end{array} +$$ + +where $\hat{f}_s(X) = \frac{1}{B}\sum_{i=1}^B f_s(\pmb{x}_i)$ denotes the marginal class distribution over the batch of test images $X$ . + +![](images/d368dff16a867de483049f5c93f9de3352026a3a91076cbc03a38f08811b65f8.jpg) +Figure 2. Overview of TeSLA Framework. (a) The student model $f_{s}$ is adapted on the test images by minimizing the proposed test-time objective $\mathcal{L}_{\mathrm{pl}}$ . The high-quality soft-pseudo labels required by $\mathcal{L}_{\mathrm{pl}}$ are obtained from the exponentially weighted averaged teacher model $f_{t}$ and refined using the proposed Soft-Pseudo Label Refinement (PLR) on the corresponding test images. The soft-pseudo labels are further utilized for teacher-student knowledge distillation via $\mathcal{L}_{\mathrm{kd}}$ on the adversarially augmented views of the test images. (b) The adversarial augmentations are obtained by applying learned sub-policies sampled i.i.d from $\mathbb{P}$ using the probability distribution $P$ with their corresponding magnitudes selected from $M$ . The parameters $M$ and $P$ of the augmentation module are updated by the unbiased gradient estimator (Eq. 13) of the loss $\mathcal{L}_{\mathrm{Aug}}$ computed on the augmented test images. + +# 3.2. Knowledge Distillation via Augmentation + +Self-learning from adversarial augmentations. As illustrated in Fig. 1, adversarial augmentations are learned by pushing their feature representations toward the decision boundary (Expansion phase), followed by updating the student model $f_{s}$ to match its prediction on these augmented images with their corresponding soft-pseudo label (Separation phase), yielding better separation of features into their respective classes. In our method setup, we continually learn automatic augmentations of the test images (Sec. 3.3) that are adversarial to the current teacher model $f_{t}$ and enforce consistency between the student model's $f_{s}$ predictions on the augmented views and the corresponding soft-pseudo labels. Since we freeze the classifier module, $f_{s}$ and $f_{t}$ share the common decision boundary. Let $\tilde{\boldsymbol{x}}$ denote the learned adversarial augmented view of $\boldsymbol{x}$ and $\hat{\boldsymbol{y}}$ denote the corresponding soft-pseudo label. We minimize the following knowledge distillation loss $\mathcal{L}_{\mathrm{kd}}$ using KL-divergence to distill knowledge from the teacher model to the student model: + +$$ +\mathcal {L} _ {\mathrm {k d}} (\tilde {\boldsymbol {x}}, \hat {\boldsymbol {y}}) = \mathcal {D} _ {\mathrm {K L}} (\hat {\boldsymbol {y}} \| f _ {s} (\tilde {\boldsymbol {x}})) \tag {4} +$$ + +Soft pseudo label refinement (PLR). We refine the quality of soft pseudo-labels and the feature representations by averaging the teacher model's outputs on multiple weakly augmented image views $\rho_w(\pmb{x})$ (via FLIPPING, CROPPING augmentations) of the same test image $\pmb{x}$ as follows: + +$$ +\mathbf {z} _ {t}, \mathbf {y} _ {t} \leftarrow \mathbb {E} _ {\mathbf {u} \in \rho_ {w} (\boldsymbol {x})} \left[ g _ {t} (\mathbf {u}), h _ {t} \left(g _ {t} (\mathbf {u})\right) \right] \tag {5} +$$ + +The refined feature representations $\mathbf{z}_t$ and the softmaxed pseudo labels $\mathbf{y}_t$ from the encoder $g_t$ and the classifier $h_t$ are further stored in an online memory queue $\mathbf{Q}$ of fixed size. The final soft pseudo-label is computed by averaging the refined soft pseudo-labels of $n$ -nearest neighbors of the current test image in the feature space as follows: + +$$ +\begin{array}{l} \mathbf {Q} [ \arg \max (\boldsymbol {y} _ {t}) ]. \text {a p p e n d} \left(\left\{\boldsymbol {z} _ {t}, \boldsymbol {y} _ {t} \right\}\right) \\ \hat {\boldsymbol {y}} = \frac {1}{n} \sum \mathcal {N} _ {\mathbf {Q}, n} (\mathbf {z} _ {t}) \tag {6} \\ \end{array} +$$ + +where $\mathbf{Q}$ denotes an online memory of class-balanced queues (with size $|\mathbf{Q}[\arg \max (\pmb {y}_t)]| \leq N_{\mathbf{Q}}$ ) of the refined feature representations and the softmaxed label predictions on the previously seen test images, and $\mathcal{N}_{\mathbf{Q},n}(\mathbf{z}_t)$ denotes soft labels of $n$ -nearest neighbors of $\mathbf{z}_t$ from $\mathbf{Q}$ . + +# 3.3. Learning Adversarial Data Augmentation + +This section presents our efficient online method for learning adversarial data augmentation for the current teacher $f_{t}$ . We first introduce our adversarial augmentation search space. Then, we describe our differentiable strategy to optimize and sample adversarial augmentations that push the feature representations of the test images toward the uncertain region in the close vicinity of the decision boundary. + +Policy search space. Let $\mathbb{O}$ be a set of all image transformation operations $\mathcal{O}:\mathbb{R}^{H\times W\times 3}\to \mathbb{R}^{H\times W\times 3}$ defined in our search space. In particular, we use image transformations AUTO-CONTRAST, EQUALIZE, INVERT, SOLARIZE, POSTERIZE, CONTRAST, BRIGHTNESS, COLOR, SHEARX, SHEARY, TRANSLATEX, TRANSLATEY, ROTATE, AND SHARPNESS. Each transformation $\mathcal{O}$ is specified with its magnitude parameter $m\in [0,1]$ . Since the output of some image operations may not be conditional on its magnitude (e.g., EQUALIZE) or may not be differentiable (e.g., POSTERIZE), for such image operations, we use the straight-through gradient estimate w.r.t. each pixel as: $\partial \mathcal{O}(\boldsymbol{x}_{ij}) / \partial m = 1$ for magnitude optimization. + +We define a sub-policy $\rho$ as a combination of $N$ image operations (sub-policy dimension) from $\mathbb{O}$ that is applied sequentially to a given image $\pmb{x}$ as: + +$$ +\rho \left(\boldsymbol {x}, m ^ {\rho}\right) = \mathcal {O} _ {1} ^ {\rho} \dots \mathcal {O} _ {N} ^ {\rho} \left(\boldsymbol {x}, m _ {N} ^ {\rho}\right) \tag {7} +$$ + +where $m^{\rho} = [m_{1}^{\rho},\dots,m_{N}^{\rho}]$ denotes the magnitude set of the sub-policy $\rho$ . We also define $\mathbb{P} = \{\rho_1,\rho_2,\ldots \}$ as the set of all possible sub-policies. + +Policy evaluation. We propose learnable adversarial augmentation, aiming to optimize and sample data augmentations as a proxy for simulating data in the uncertain region of the feature space. Given the teacher model $f_{t}$ , a sub-policy $\rho$ with magnitude $m$ is evaluated for a test image $x$ using the following adversarial objective: + +$$ +\mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho) = \sum_ {k = 1} ^ {K} f _ {t} (\tilde {\boldsymbol {x}}) \log \left(f _ {t} (\tilde {\boldsymbol {x}})\right) + \lambda_ {1} r (\tilde {\boldsymbol {x}}, \boldsymbol {x}) \tag {8} +$$ + +where $\tilde{\pmb{x}} = \rho (\pmb {x},m)$ denotes the adversarial augmented image and $r(\cdot ,\cdot)$ regularizes augmentation severity. The hyperparameter $\lambda_{1}$ controls the augmentation severity. We optimize the adversarial augmentation loss $\mathcal{L}_{\mathrm{aug}}$ with respect to the sub-policy $\rho$ 's parameters. Minimizing $\mathcal{L}_{\mathrm{aug}}$ is equivalent to maximizing the entropy of the teacher model's prediction on the augmented image $\tilde{\pmb{x}}$ thus pushing its feature representation towards the decision boundary (first term). For the augmentation regularization (second term), we use the mean squared distance function between the teacher encoder's $L$ internal layers' activations of the adversarial augmentation and non-augmented versions of image $\pmb{x}$ defined as: + +$$ +r (\tilde {\boldsymbol {x}}, \boldsymbol {x}) = \frac {1}{L} \sum_ {l = 1} ^ {L} \| \mu_ {l} (\tilde {\boldsymbol {x}}) - \mu_ {l} (\boldsymbol {x}) \| ^ {2} \tag {9} +$$ + +where $\mu_l(\cdot)$ denotes the mean activation of the $l^{th}$ layer of the teacher encoder $g_t$ . + +Policy optimization. Let $M = [m^{\rho_1}, \ldots, m^{\rho_{\| \mathbb{P} \|}}]$ denote the set of magnitudes $m^{\rho}$ of all sub-policies $\rho \in \mathbb{P}$ and $P = [p_1, \ldots, p_{\| \mathbb{P} \|}]$ denote the probability of selecting a sub-policy from $\mathbb{P}$ . The expected policy evaluation loss (Eq. 8) for an image $x$ over the policy search space is given by: + +$$ +\mathbb {E} \left[ \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}) \right] = \sum_ {i = 1} ^ {\| \mathbb {P} \|} p _ {i} \cdot \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho_ {i}) \tag {10} +$$ + +Evaluating the gradient of $\mathbb{E}[\mathcal{L}_{\mathrm{aug}}(\pmb{x})]$ w.r.t. $M$ and $P$ can become computationally expensive as $\| \mathbb{P}\| \sim \binom{\|\mathbb{O}\|}{N}$ . Thus, we use the following re-parameterization trick to estimate its unbiased gradient: + +$$ +\begin{array}{l} \nabla \mathbb {E} [ \mathcal {L} _ {\mathrm {a u g}} (x) ] = \delta (\boldsymbol {x}, \mathbb {P}) = \sum_ {i = 1} ^ {\| \mathbb {P} \|} \nabla \left(p _ {i} \cdot \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho_ {i})\right) \tag {11} \\ = \sum_ {i = 1} ^ {\| \mathbb {P} \|} p _ {i} \left(\nabla \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho_ {i}) + \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho_ {i}) \cdot \nabla \log p _ {i}\right) \\ \end{array} +$$ + +Thus, an unbiased estimator of $\delta (\pmb {x},\mathbb{P})$ can be written as: + +$$ +\hat {\delta} (\boldsymbol {x}, \rho_ {i}) = \nabla \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho_ {i}) + \mathcal {L} _ {\mathrm {a u g}} (\boldsymbol {x}, \rho_ {i}) \cdot \nabla \log p _ {i} \tag {12} +$$ + +where index $i$ is sampled from the probability distribution $P$ . For an online batch of $B$ test images $\{\pmb{x}_1, \pmb{x}_2, \dots, \pmb{x}_B\}$ , we apply augmentation sub-policies $\{\rho_{i_1}, \rho_{i_2}, \dots, \rho_{i_B}\}$ from $\mathbb{P}$ where $\{i_1, i_2, \dots, i_B\}$ are sampled i.i.d from distribution $P$ , and update the parameters $P$ and $M$ with the following stochastic gradient update rule: + +$$ +[ P, M ] \leftarrow [ P, M ] - \frac {\gamma}{B} \sum_ {j = 1} ^ {B} \hat {\delta} \left(\boldsymbol {x} _ {j}, \rho_ {i _ {j}}\right) \tag {13} +$$ + +where $\gamma$ is the learning rate. We set $\gamma = 0.1$ for all experiments without the need for hyperparameter tuning. + +# 3.4. Self-Learning With Adversarial Augmentation + +We minimize the following overall objective $\mathcal{L}_{\mathrm{TeSLA}}$ for training the student model $f_{s}$ on a batch of $B$ test images $X = \{\pmb {x}_1,\dots ,\pmb {x}_B\}$ and their adversarial augmented views $\hat{X} = \{\tilde{x}_1,\dots ,\tilde{x}_B\}$ with the corresponding refined softpseudo labels from the teacher model $\hat{Y} = \{\hat{y}_1,\dots ,\hat{y}_B\}$ : + +$$ +\mathcal {L} _ {\mathrm {T e S L A}} (X, \tilde {X}, \hat {Y}) = \mathcal {L} _ {\mathrm {p l}} (X, \hat {Y}) + \frac {\lambda_ {2}}{B} \sum_ {i = 1} ^ {B} \mathcal {L} _ {\mathrm {k d}} (\tilde {\boldsymbol {x}} _ {i}, \hat {\boldsymbol {y}} _ {i}) \tag {14} +$$ + +where $\lambda_{2}$ is a hyper-parameter for the knowledge distillation. The adversarial augmented views $\tilde{X}$ are obtained by sampling i.i.d. augmentation sub-policies from $\mathbb{P}$ for every test image in $X$ with probability $P$ and applying the corresponding magnitude from $M$ . Concurrently, we also optimize $M$ and $P$ using the policy optimization mentioned in Sec. 3.3. + +# 4. Experiments + +# 4.1. Datasets and Experimental Settings + +We evaluate and compare TeSLA against state-of-the-art (SOTA) test-time adaptation algorithms for both classification and segmentation tasks under three types of test-time distribution shifts resulting from (1) common image corruption, (2) synthetic to real data transfer, and (3) measurement shifts on medical images. The latter is characterized by a change in medical imaging systems, e.g., different scanners across hospitals or various staining techniques. + +TTA protocols. We adopt the TTA protocols in [39] and categorize competing methods based on two factors: (i) source training objective and (ii) sequential or non-sequential inference. First, unlike [39], we use $\mathbf{Y}$ to indicate if access to the source domain's information, e.g., feature statistics is possible or if the source training objective is allowed to be modified; otherwise, we use $\mathbf{N}$ . Next, we use $\mathbf{O}$ to indicate a + +one-pass adaptation and evaluation protocol (one epoch) for sequential test data and $\mathbf{M}$ to show a multi-pass adaptation on all test data and inference after several epochs. Thus, we end up with four possible TTA protocols, namely N-M, Y-M, N-O, and Y-O. Unlike previous works [28,39], TeSLA does not rely on any source domain feature distribution, yielding two realistic TTA protocols: N-M and N-O. + +Hyperparameters for all experiments. We set $\lambda_1 = \lambda_2 = 1.0$ for all experiments (cf. ablation in Appendix E). For the adversarial augmentation module, we use sub-policy dimension $N = 2$ (except for VisDA-C [33], $N = 4/3$ for N-O/N-M protocols) (cf. Appendix E) and Adam optimizer [22]. For pseudo-label refinement (PLR), we use five weak augmentations composed of random flips and resize crop. More details of hyperparameters used for each experiment can be found in Appendix A. + +Common image corruptions. To evaluate TeSLA's efficacy for the classification task, we use CIFAR10-C/CIFAR100-C [16] and large-scale ImageNet-C [16] datasets each containing 19 types of corruptions applied to the clean test set with five levels of severity. We perform validation on 4 out of 19 types of corruption (SPATTER, GAUSSIAN BLUR, SPECKLE NOISE, SATURATE) to select hyperparameters and test on the remaining 15 corruptions at the maximum severity level of 5. Following [28], we train the ResNet50 [15] on the clean CIFAR10/CIFAR100/ImageNet training set and adapt it to classify the unlabeled corrupted test set. + +Synthetic to real data adaption. We use challenging and large-scale VisDA-C, and VisDA-S [33] datasets for evaluating synthetic-to-real data adaptation at test-time for classification and segmentation tasks, respectively. Following [4, 26], we adapt the ResNet101 network pre-trained on synthetic images to classify 12 vehicle classes on the photo-realistic images of VisDA-C. While for VisDA-S, we adapt the DeepLab-v3 [5] backbone pre-trained on synthetic GTA5 [36] to the Cityscapes [7] to segment 19 classes. + +Measurement shifts on medical images. We access TeSLA on staining variations for the tissue type classification of hematoxylin & eosin (H&E) stained patches from colorectal cancer tissue slides. We use MobileNetV2 [38] trained on the source Kather-19 dataset [20] and adapt it to the target Kather-16 dataset [21] on four tissue categories: tumor, stroma, lymphocyte, and mucosa. We also evaluate TeSLA for variations in scanners and imaging protocols in multi-site medical images on two magnetic resonance imaging (MRI) datasets for the segmentation task, namely the multi-site prostate MRI [27] and spinal cord grey matter segmentation (SCGM) [34]. Following [42], we adapt the U-Net [37] from site 1 to sites $\{2,3,4\}$ of the spinal cord dataset, and from sites $\{\mathrm{A,B}\}$ to sites $\{\mathrm{D,E,F}\}$ of the prostate dataset. + +Competing baselines. We compare TeSLA with the following SFDA and TTA baselines, including direct inference of the trained source model on the target test data without + +Table 1. Comparison of SOTA TTA methods under different protocols evaluated on CIFAR-10/100-C, ImageNet-C, VisDA-C and Kather-16 datasets. We report the average error computed over 15 test corruptions for the common image corruption shifts. We also report Class Avg. in % error rates for synthetic-to-real and measurement shifts over 3 and 10 seeds, respectively. + +
MethodProtocolCommon Image CorruptionsSyn-to-RealMeasurement Shift
CIFAR10-CCIFAR100-CImageNet-CVisDA-CKather-16
OMOMOMOMOM
SourceN29.160.481.851.532.0
BN [18,30]N15.615.443.743.367.767.635.435.018.318.2
TENT [45]N14.112.939.036.557.454.233.529.316.212.0
SHOT [26]N13.914.239.238.768.768.229.424.514.712.0
AdaContrast [4]N-23.120.2-
TTT++ [28]Y15.89.844.434.159.3-35.234.116.77.9
TTAC [39]Y13.49.441.733.658.7-32.231.19.65.5
TeSLAN12.59.738.232.955.051.517.813.59.23.3
TeSLA-sY12.19.737.332.653.1-24.017.99.93.1
+ +adaptation (Source). We also implement the SFDA-based pseudo-labeling baselines from a basic pseudo-labeling approach (PL) to the SHOT approach [26] based on training the feature extraction module by adopting the information maximization loss to make globally diverse but individually certain predictions on the target domain. For recent TTA methods, we compare our method against $\mathrm{TTT}++$ [28], Ada-Contrast [4], CoTTA [47], basic TENT [45] and BN [18] as representative methods based on feature distribution alignment via stored statistics, contrastive learning, augmentation-averaged predictions, entropy minimization, and batch normalization statistic. Furthermore, we compare TeSLA to more recent methods based on test-time augmentation policy learning, OptTTA [42], and anchor clustering TTAC [39]. By default, [4,18,26] follow the N-M protocol, and we adapt them to the N-O protocol, while [42,45,47], by default, follow the N-O protocol and adapt them to the N-M protocol setting. Using source features distribution, TTAC and $\mathrm{TTT}++$ follow Y-O and Y-M. We report error rates on classification tasks and Mean Intersection over Union (mIoU)/ Dice scores on segmentation tasks. For consistency across baselines and TTA protocols, we do not use any specialized model architectures (e.g., projection head, weight-normalized classifier layer) for SHOT [26] and AdaContrast [4]. + +# 4.2. Results + +Classification task. In Table 1, we summarize the average (Avg) classification error rates $(\%)$ of TeSLA against several SOTA SFDA and TTA baselines on the common image corruptions of CIFAR10-C, CIFAR100-C, ImageNet-C; synthetic to real data domain shifts of VisDA-C; and measurement shifts on the Kather-16 dataset. We also report per-class top-1 accuracies for VisDA-C and Kather-16 datasets and corruption-wise error rates on CIFAR-10-C/CIFAR-100-C and ImageNet-C datasets in Appendix B. TeSLA easily surpasses all competing methods under the protocols N-O and N-M on all the datasets. In particular, TeSLA outper + +![](images/5481e6f89e003844907e5351db3ec8d7f9fbe209cb6f605e759a00edcd5361d1.jpg) +(a) + +![](images/51a933acca7eabc2edb1b3dca9907571ec6ab9ca6b60f8439724348436d09b4b.jpg) + +![](images/89ef494eb32bc41feff3ecfad170cab501eecd14ca45eaeedc89557b927532f1.jpg) +(b) + +![](images/5dc96ef47a6b961c8efc22b5ca63c6d510532cfa72c6860e7c1a408e7fc548fa.jpg) + +![](images/aa75a6905b65285834f8f181ef46ea44c25bd404a58608b3b943a9dee2ed7683.jpg) +Figure 3. Ablation experiments on the VisDA-C. (a) Source training strategy and model architecture. TeSLA outperforms the competing TTA methods across varied source training strategies (Supervised, SimCLRV2 [6], and DINO [3]) and model architectures (ResNet50, ResNet101 [15], and ViT-B/16 [11]) using the same set of hyperparameters for both protocols (N-O/N-M). Each vertex represents the mean class avg accuracy over 3 seeds. (b) t-SNE visualization [43] comparison of feature embeddings for AdaContrast [4] and TeSLA. + +![](images/733c05ae857a8f2ac9c803d3f672b8f392aa59ff42b2f12cb264332f3671cab8.jpg) +Figure 4. Model calibration and uncertainty estimation. Model calibration using reliability diagrams [31] and model uncertainty estimation using Brier Score and Negative Log Likelihood (NLL) of the Source Model, TENT [45], OptTTA [42] and, TeSLA on the prostate dataset for test-time adaptation. + +Table 2. Semantic segmentation results (Class Avg. mIoU in %) on the VisDA-S (GTA5 → Cityscapes) test-time adaptation task. + +
ProtocolSourceBN [30]TENT [45]PLCoTTA [47]TeSLA
NO35.336.738.338.837.044.5
NM38.439.238.639.946.0
+ +forms the second-best baseline under N-O/N-M protocol by a $(\%)$ margin 1.4/3.2 on CIFAR10-C, 0.8/3.6 on CIFAR100-C, 2.4/2.7 on ImageNet-C, 5.3/6.7 on VisDA-C, and 5.5/8.7 on the Kather-16 datasets. Despite not utilizing any source feature statistic, TeSLA (N-O/N-M) is either competitive or surpasses TTAC [39] and TTT++ [28] (Y-O/Y-M) that benefit from the source dataset feature statistics during adaptation. Including the global source feature alignment module of TTAC with TeSLA (TeSLA-s) could further improve its performance under (Y-O/Y-M) protocols for slight domain shifts (e.g., common image corruptions). We report the results of AdaContrast for VisDA-C only as the implementation for other datasets are not provided. We do not report TeSLA-s results on ImageNet-C under the Y-M protocol, as previous baselines were evaluated only Y-O protocol. + +Segmentation task. Unlike TeSLA, several TTA methods e.g., AdaContrast [4], SHOT [26], TTT++ [28], TTAC [39], have only been evaluated on the classification task. Therefore, we compare TeSLA against the current SOTA TTA + +Table 3. Semantic segmentation results (Class Avg. volume-wise mean Dice in %) on the cross-site spinal cord and prostate MRI test-time adaptation tasks. + +
ProtocolSourceBN [30]TENT [45]PLOptTTA [42]TeSLA
Spinal CordSite{1} → Sites{2,3,4}
NO76.0±11.881.6±8.381.1±9.181.7±8.684.1±4.885.3±5.8
NM84.3±4.884.4±4.784.3±4.784.3±4.485.4±4.4
ProstateSites{A,B} → Sites{D,E,F}
NO60.5±27.072.1±15.274.7±17.972.4±15.283.1±7.783.5±6.5
NM73.1±18.081.2±9.381.1±9.283.4±7.784.3±5.8
+ +methods applied to the segmentation task on synthetic-to-real data transfer of the VisDA-S dataset in Table 2. TeSLA significantly outperforms the competing methods and achieves the best mIOU scores for both N-O and N-M protocols, beating the second-best baseline by a margin of $+5.7\%$ and $+6.1\%$ , respectively. In Table 3, we compare TeSLA against the recent SOTA test-time augmentation policy method, OptTTA [42] and other TTA baselines for the inter-site adaptation on two challenging MRI datasets mentioned in Sec. 4.1. TeSLA convincingly outperforms all other methods on severe measurement shifts across sites under N-O and N-M protocols. + +Computational cost for adversarial augmentations. As shown in Table 4, TeSLA learns the adversarial augmentation in an online manner, and thus its runtime is several orders faster than learnable test-time augmentation OptTTA [42], while comparable to that of static augmentation policies with an additional overhead of only 0.10 GPU hours/epoch on the VisDA-C dataset. + +Test-time feature visualization. Fig. 3b compares t-SNE projection [43] of the encoder's features of AdaContrast [4] with TeSLA on the VisDA-C. TeSLA shows better inter-class separation for features than AdaContrast, as supported by an improved silhouette score from 0.149 to 0.271. + +Model calibration and uncertainty estimation. Entropy minimization-based TTA methods [26, 45] explicitly make the model confident in their predictions, resulting in poor model calibration. For trusting the adapted model, it should + +Table 4. Performance comparison of adversarial augmentation learned by TeSLA during adaptation against the prior art augmentation methods on runtime (GPU hours/epoch on NVIDIA GeForce RTX 3090) and Class Avg. accuracy in % on the VisDA-C dataset. + +
Augmentation TypeOptTTATeSLARA [10]TeSLAAA [8]TeSLAN=2TeSLAN=3
Runtime4.500.050.060.160.18
Class Avg. Acc. (N-O)-80.281.281.582.2
Class Avg. Acc. (N-M)-84.785.786.386.5
+ +output reliable confidence estimates matching its true underlying performance on the test images. Such a well-behaved model is characterized by expected calibration error (ECE) through reliability diagram [31], and uncertainty metrics of brier score and negative log-likelihood (NLL) [14]. In Fig. 4, we plot the reliability diagram (dividing the probability range [0, 0.5] into 10 bins), and uncertainty metrics of the Source model, TENT [45] and OptTTA [42] against TeSLA for the inter-site model adaptation on the prostate dataset [27]. We observe that TeSLA achieves the best model calibration (lowest ECE of $1.80\%$ ) and lowest Brier and NLL scores of 0.12 and 0.24 ( $p < 0.006$ against TENT). + +# 4.3. Ablation Studies + +In this section, we scrutinize the roles played by different components of TeSLA. All ablations are performed on the synthetic-to-real test data adaptation of the VisDA-C dataset. Loss functions for self-learning. We compare TTA performance of our self-learning objective $\mathcal{L}_{\mathrm{pl}}$ with the proposed flipped cross-entropy loss $f$ -CE and the basic cross-entropy loss $\mathbb{C}\mathbb{E}$ in Fig. 5-(1). We do not use the proposed augmentation module and pseudo-label refinement in this experiment. We observe that $f$ -CE alone improves the accuracy of the Source model in the N-M protocol by $+23.1\%$ compared to $+16.6\%$ improvement by $\mathbb{C}\mathbb{E}$ . Using $\mathcal{L}_{\mathrm{pl}}$ (Eq. 3) further improves the accuracy to a margin of $+23.8\%$ . + +Contribution of individual components: PLR, $\mathcal{L}_{\mathrm{pl}}$ , and $\mathcal{L}_{\mathrm{kd}}$ . Our test-time objective $\mathcal{L}_{\mathrm{TeSLA}}$ has three components (i) self-learning loss $\mathcal{L}_{\mathrm{pl}}$ , (ii) pseudo-label refinement (PLR), and (iii) knowledge distillation $\mathcal{L}_{\mathrm{kd}}$ with adversarial augmentation. In Fig. 5-(3), we study the effect of accruing individual components to $\mathcal{L}_{\mathrm{TeSLA}}$ (starting from source trained model) for the test-time adaptation accuracy. We observe that $\mathcal{L}_{\mathrm{pl}}$ alone improves the model's accuracy from $48.5\%$ to $72.3\%$ while adding PLR and $\mathcal{L}_{\mathrm{kd}}$ further boosts it to $81.6\%$ $(+9.3)$ and $86.3\%$ $(+4.7)$ , respectively. + +PLR: ensembling on weak augmentations (Ens) and nearest neighbor averaging (NN). We also conduct an ablation on the two pseudo-label refinement approaches - (i) ensembling the teacher model's output on five weak augmentations. (ii) averaging soft pseudo labels of $n = 10$ nearest neighbors in the feature space (cf. sensitivity test in Appendix E). Fig. 5-(2) shows that combining both approaches gives better results than using them individually. + +![](images/8d3f41fc7804936003a516f0bb3545a2b728a6817358273bbe0a209dcc23190b.jpg) +Figure 5. Ablation experiments on the VisDA-C. (left to right) (1) Ablation results for self-learning loss functions, including $f\text{-}\mathbb{C}\mathbb{E}$ , $\mathbb{C}\mathbb{E}$ , and $\mathcal{L}_{\mathrm{pl}}$ ; (2) ablation on different PLR methods: assembling predictions on 5 weak augmentations, averaging predictions of the 10 nearest neighbors in the feature space; (3) ablation on different components of TeSLA's objective function: $\mathcal{L}_{\mathrm{pl}}$ , PLR, and $\mathcal{L}_{\mathrm{kd}}$ . + +![](images/e54867bba1455f679010f097741a460ee3810386624bea3ffd5d1af8a9ba4043.jpg) + +![](images/b204bfa589e44d85f6571bcdea810af261121e04f110a54d4ecd08a1c7d81a96.jpg) + +Adversarial augmentation. Table 4 shows the benefit of using our adversarial augmentation $(N = 2,3)$ with TeSLA against (i) RandAugment (RA) [10], (ii) AutoAugment (AA) [9] (optimized for ImageNet), and (iii) OptTTA [42]. We achieve far better Class Avg. accuracy on the VisDA-C at the cost of slight runtime overhead compared to static augmentation policies for protocols (N-O, N-M). In addition, TeSLA augmentation (TeAA) consistently improves the performance of other TTA methods (Table 5). + +Table 5. Class Avg. accuracy (%) of TeSLA augmentation (TeAA) with other TTA objectives under N-O protocol on VisDA-C dataset. + +
TENT+TeAA| SHOT+TeAA| AdaContrast+TeAA
66.572.0| 70.673.1| 76.979.1
+ +Source training strategies and model architectures. The ablation results (Fig. 3a) on the VisDA-C show that TeSLA surpasses competing TTA baselines for different source training strategies, including Supervised and self-supervised (SimCLRV2 [6], DINO [3]). Moreover, we ablate TeSLA over different architectures, ranging from ResNet-50/101 [15] to the recent vision transformer ViT-B [11] using the same set of hyperparameters, showing the merits of no reliance on the network architecture or source training strategy. + +Finally, in Appendix E, we provide detailed sensitivity results with respect to all hyperparameters. + +# 5. Conclusion + +We introduced TeSLA, a novel self-learning algorithm for test-time adaptation that utilizes automatic adversarial augmentation. TeSLA is agnostic to the model architecture and source training strategies and gives better model calibration than other TTA methods. Through extensive experiments on various domain shifts (measurement, common corruption, synthetic to real), we show TeSLA's superiority over previous TTA methods on classification and segmentation tasks. Note that TeSLA assumes class uniformly when implicitly maximizing mutual information and can be improved by incorporating prior, e.g., class label distribution statistics. + +# References + +[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009-2018, 2022. 1, 2 +[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. In International Conference on Learning Representations, 2019. 2 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 7, 8 +[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2, 6, 7 +[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7, 8 +[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[8] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 8 +[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 8 +[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 8 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 7, 8 +[12] Cian Eastwood, Ian Mason, Christopher K. I. Williams, and Bernhard Scholkopf. Source-free adaptation to measurement shift via bottom-up feature restoration. In International Conference on Learning Representations, 2022. 2 + +[13] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 2 +[14] Alvaro Gomariz, Tiziano Portenier, César Nombela-Arrieta, and Orcun Goksel. Probabilistic spatial analysis in quantitative microscopy with uncertainty-aware cell detection using deep bayesian regression of density maps. arXiv preprint arXiv:2102.11865, 2021. 8 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7, 8 +[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 6 +[17] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 2 +[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 6 +[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1, 2 +[20] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 6 +[21] Jakob Nikolas Kather, Frank Gerrit Zollner, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Cleo-Aron Weis. Collection of textures in colorectal cancer histology, May 2016. 6 +[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[23] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 1, 2 +[24] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12376-12385, 2020. 2 +[25] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 1, 2 +[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for + +unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 3, 6, 7 +[27] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. Msnet: Multi-site network for improving prostate segmentation with heterogeneous mri data. IEEE Transactions on Medical Imaging, 2020. 6, 8 +[28] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 7 +[29] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy policy search: A simple baseline for learnable test-time augmentation. In Conference on Uncertainty in Artificial Intelligence, pages 1308-1317. PMLR, 2020. 2 +[30] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 6, 7 +[31] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In Proceedings of the 22nd international conference on Machine learning, pages 625-632, 2005. 7, 8 +[32] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 2 +[33] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6 +[34] Ferran Prados, John Ashburner, Claudia Blaiotta, Tom Brosch, Julio Carballido-Gamio, Manuel Jorge Cardoso, Benjamin N Conrad, Esha Datta, Gergely David, Benjamin De Leener, et al. Spinal cord grey matter segmentation challenge. Neuroimage, 152:312-329, 2017. 6 +[35] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In International Joint Conference on Artificial Intelligence, 2021. 2 +[36] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, European Conference on Computer Vision (ECCV), volume 9906 of LNCS, pages 102-118. Springer International Publishing, 2016. 6 +[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6 +[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted + +residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6 +[39] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7 +[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2 +[41] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3 +[42] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozortbar. OptTTA: Learnable test-time augmentation for source-free medical image segmentation under domain shift. In Medical Imaging with Deep Learning, 2022. 2, 3, 6, 7, 8 +[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7 +[44] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 2 +[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8 +[46] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 338:34-45, 2019. 2 +[47] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 6, 7 +[48] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Casting a bait for offline and online source-free domain adaptation. arXiv preprint arXiv:2010.12427, 2020. 2 +[49] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8978-8987, 2021. 2 +[50] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. arXiv preprint arXiv:2205.04183, 2022. 2 +[51] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29:7834-7844, 2020. 2 \ No newline at end of file diff --git a/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/images.zip b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a686fc0b45b1dc93369a432086f2a31e9932552d --- /dev/null +++ b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e466481b882003566f228705597a44e2348ef555d306c1f3c58fcb76255d5884 +size 409470 diff --git a/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/layout.json b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..83d1d0e26f95e5f834818021a3b28ef908eba845 --- /dev/null +++ b/2023/TeSLA_ Test-Time Self-Learning With Automatic Adversarial Augmentation/layout.json @@ -0,0 +1,11014 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "text", + "content": "TeSLA: Test-Time Self-Learning With Automatic Adversarial Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 512, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 512, + 170 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 512, + 170 + ], + "type": "text", + "content": "Devavrat Tomar1 Guillaume Vray1 Behzad Bozorgtabar1,2 Jean-Philippe Thiran1,2 EPFL CHUV" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 172, + 376, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 172, + 376, + 185 + ], + "spans": [ + { + "bbox": [ + 212, + 172, + 376, + 185 + ], + "type": "inline_equation", + "content": "^{1}\\{" + }, + { + "bbox": [ + 212, + 172, + 376, + 185 + ], + "type": "text", + "content": " firstname\\}. " + }, + { + "bbox": [ + 212, + 172, + 376, + 185 + ], + "type": "inline_equation", + "content": "\\{" + }, + { + "bbox": [ + 212, + 172, + 376, + 185 + ], + "type": "text", + "content": "lastname\\}@epfl.ch" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 515 + ], + "type": "text", + "content": "Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with automatic Adversarial augmentation dubbed TeSLA for adapting a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning methods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowledge distillation by simulating high entropy augmented images. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several desirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available at https://github.com/devavratTomar/TeSLA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 127, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 127, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 127, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Deep neural networks (DNNs) perform exceptionally well when the training (source) and test (target) data follow the same distribution. However, distribution shifts are inevitable in real-world settings and propose a major challenge to the performance of deep networks after deployment. Also, access to the labeled training data may be infeasible at test time due to privacy concerns or transmission bandwidth. In such scenarios, source-free domain adaptation (SFDA) [1, 23, 25] and test-time adaptation (TTA) methods [19, 28, 39] aim to adapt the pre-trained source model to the unlabeled distributionally shifted target domain while easing access to source data. While SFDA methods have access to all full target data through multiple training epochs" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 395, + 297 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 395, + 297 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 395, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 395, + 297 + ], + "type": "image", + "image_path": "606c4613313034c2fb162c2724bbb58c1b37abc42126bae8d18860dca17451d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "type": "text", + "content": "Figure 1. Knowledge distillation with adversarial augmentations. (a) Easy images with confident soft-pseudo labels and (b) Hard images with unconfident soft-pseudo labels are adversarily augmented and pushed to the uncertainty region (high entropy) near the decision boundary. The model is updated for (a) to match its output on the augmented views with non-augmented views of Easy test images using KL-Divergence " + }, + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}} \\neq 0" + }, + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "type": "text", + "content": ", while not updated for (b) as " + }, + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}} \\sim 0" + }, + { + "bbox": [ + 305, + 304, + 547, + 393 + ], + "type": "text", + "content": " between Hard images and their augmented views." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 406, + 212, + 495, + 297 + ], + "blocks": [ + { + "bbox": [ + 406, + 212, + 495, + 297 + ], + "lines": [ + { + "bbox": [ + 406, + 212, + 495, + 297 + ], + "spans": [ + { + "bbox": [ + 406, + 212, + 495, + 297 + ], + "type": "image", + "image_path": "792aacf96515bab2dd2ea598a8a614d2dd2a04833e3da270cdc529d583ed32b4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 507, + 212, + 546, + 297 + ], + "blocks": [ + { + "bbox": [ + 507, + 212, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 507, + 212, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 507, + 212, + 546, + 297 + ], + "type": "image", + "image_path": "76aca1b51c9e128429a15268ad368e38b13e9cc94b8f01981ddd8c2d8351a6c9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 401, + 546, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 509 + ], + "type": "text", + "content": "(offline setup), TTA methods usually process test images in an online streaming fashion and represent a more realistic domain adaptation. However, most of these methods are applied: (i) only to classification tasks, (ii) evaluated on the non-real-world domain shifts, e.g., the non-measurement shift; (iii) destroy model calibration—entropy minimizing with overconfident predictions [45] on incorrectly classified samples, and (iv) use specialized network architectures or rely on the source dataset feature statistics [28]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "We address these issues by proposing a new test-time adaptation method with automatic adversarial augmentation called TeSLA, under which we further define realistic TTA protocols. Self-learning methods often supervise the model adaptation on the unlabeled test images using their predicted pseudo-labels. As the model can easily overfit on its own pseudo-labels, a weight-averaged teacher model (slowly updated by the student model) is employed for obtaining the pseudo-labels [41, 47]. The student model is then trained with cross-entropy " + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbb{C}\\mathbb{E})" + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": " loss between the one-hot pseudolabels and its predictions on the test images. In this paper, we instead propose to minimize flipped cross-entropy between the student model's predictions and the soft pseudo-labels (notice the reverse order) with the negative entropy of its marginalized predictions over the test images. In Sec. 3, we show that the proposed formulation is an equivalence to mutual information maximization implicitly corrected" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20341" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "by the teacher-student knowledge distillation via pseudolabels, yielding performance improvement on various test-time adaptation protocols and compared to the basic " + }, + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "inline_equation", + "content": "\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": " optimization (cf. ablation in Fig. 5-(1))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 287 + ], + "type": "text", + "content": "Motivated by teacher-student knowledge distillation, another central tenet of our method is to assist the student model during adaptation in improving its performance on the hard-to-classify (high entropy) test images. For this purpose, we propose learning automatic adversarial augmentations (see Fig. 1) as a proxy for simulating images in the uncertainty region of the feature space. The model is then updated to ensure consistency between predictions of high entropy augmented images and soft-pseudo labels from the respective non-augmented versions. Consequently, the model is self-distilled on Easy test images with confident soft-pseudo labels (Fig. 1a). In contrast, the model update on Hard test images is discarded (Fig. 1b), resulting in better class-wise feature separation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 288, + 479 + ], + "type": "text", + "content": "In summary, our contributions are: (i) we propose a novel test-time self-learning method based on flipped cross-entropy " + }, + { + "bbox": [ + 46, + 288, + 288, + 479 + ], + "type": "inline_equation", + "content": "(f\\text{-}\\mathbb{C}\\mathbb{E})" + }, + { + "bbox": [ + 46, + 288, + 288, + 479 + ], + "type": "text", + "content": " through the tight connection with the mutual information between the model's predictions and the test images; (ii) we propose an efficient plug-in test-time automatic adversarial augmentation module used for online knowledge distillation from the teacher to the student network that consistently improves the performance of test-time adaptation methods, including ours; and (iii) TeSLA achieves new state-of-the-art results on several benchmarks, from common image corruption to realistic measurement shifts for classification and segmentation tasks. Furthermore, TeSLA outperforms existing TTA methods in terms of calibration and uncertainty metrics while making no assumptions about the network architecture and source domain's information, e.g., feature statistics or training strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": "In general, domain adaptation methods aim to distill knowledge from source data that are well-generalizable to target data and relax the assumption of i.i.d. between source and target datasets. To circumvent expensive and cumbersome annotation of new target data, unsupervised domain adaptation (UDA) emerges, and there has been a large corpus of UDAs [13, 17, 32, 51] to match the distribution of domains on both labeled source data and unlabeled target data. Nonetheless, the above-mentioned methods demand source data to achieve the domain adaptation process, which is often impractical in real-world scenarios, e.g., due to privacy restrictions. The above issue motivates research into source-free domain adaptation (SFDA) [1, 23, 25, 26] and test-time training or adaptation (TTA) [19, 28, 39, 45], which are more closely relevant to our problem setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "SFDA approaches formulate domain adaptation through pseudo labeling [23, 26], target feature clustering [49], syn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": "thesizing extra training samples [24], or feature restoration [12]. SHOT [26] proposed feature clustering via information maximization while incorporating pseudo-labeling as additional supervision. BAIT [48] leverages the fixed source classifier as source anchors and uses them to achieve feature alignment between the source and target domain. CPGA [35] proposed the SFDA method based on matching feature prototypes between source and target domains. AaD-SFDA [50] proposed to optimize an objective by encouraging prediction consistency of local neighbors in feature space. Nevertheless, SFDA methods require apriori access to all target data in advance, and the current pseudo-labeling approach, e.g., SHOT [26], used offline pseudo-label refinement on a perepoch basis. In a more realistic domain adaptation scenario, SFDA is still incompetent to perform inference and adaptation simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 267, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 590 + ], + "type": "text", + "content": "TTA methods [4, 45] propose alleviating the domina shift by online (or streaming) adapting a model at test time. Still, we argue that there are ambiguities over the problem setup of TTA in the literature, particularly on whether sequential inference on target test data is feasible upon arrival [19, 45] or whether training objectives must be adapted [28, 40]. TTT [40] proposed fine-tuning the model parameters via rotation classification task as a proxy. On-target adaptation [44] used pseudo-labeling and contrastive learning to initialize the target-domain feature; each performed independently in their method. T3A [19] utilized pseudo-labeling to adjust the classifier prototype. More recently, TTAC [39] leveraged the clustering scheme to match target domain clusters to source domain categories. Nevertheless, existing TTA methods rely on specialized neural network architectures or only update a fraction of model weights yielding limited performance gain on the target data. For instance, TENT [45] proposed updating affine parameters in the batchnorm layers of convolutional neural networks (CNN), while AdaContrast [4] and SHOT [26] used an additional weight normalization classification layer with a projection head. In contrast, we show our method's superiority for various neural network architectures, including CNNs and vision transformers (ViTs) [11], without additional architectural requirements or different source model training strategies. Moreover, we update all model parameters, and our method is stable over a wide range of hyper-parameters, e.g., learning rate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": "Test time augmentation methods [2, 29] are another popular line of domain adaptation research. GPS [29] learns optimal augmentation sub-policies by combining image transformations of RandAugment [10] that minimize calibrated log-likelihood loss [2] on the validation set. OptTTA [42] optimized the magnitudes of augmentation sub-policies using gradient descent to maximize mutual information and match feature statistics over augmented images. Though effective, these methods [29, 42, 46] are computationally expensive as all augmentation sub-policies need to be evaluated, making" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20342" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "their real-time application difficult. Instead, our proposed adversarial augmentation policies can be learned online and are several orders faster than the existing learnable test-time augmentation strategies, e.g., [42]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 125, + 129, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 125, + 129, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 129, + 140 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 145, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 289, + 361 + ], + "type": "text", + "content": "Overview. We first introduce our flipped cross-entropy " + }, + { + "bbox": [ + 46, + 145, + 289, + 361 + ], + "type": "inline_equation", + "content": "(f\\text{-}\\mathbb{C}\\mathbb{E})" + }, + { + "bbox": [ + 46, + 145, + 289, + 361 + ], + "type": "text", + "content": " loss through the tight connection with the mutual information between the model's predictions and the test images. (Sec. 3.1). Using this equivalence, we derive our test-time loss function that implicitly incorporates teacher-student knowledge distillation. In Sec. 3.2, we propose to enhance the test-time teacher-student knowledge distillation by utilizing the consistency of the student model's predictions on the proposed adversarial augmentations with their corresponding refined soft-pseudo labels from the teacher model. We refine the soft-pseudo labels by averaging the teacher model's predictions on (1) weakly augmented test images; (2) nearest neighbors in the feature space. Furthermore, we propose an efficient online algorithm for learning adversarial augmentations (Sec. 3.3). Finally, in Sec. 3.4, we summarize our online test-time adaptation method, TeSLA, based on self-learning with the proposed automatic adversarial augmentation. The overall framework is shown in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": "Problem setup. Given an existing model " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "f_{\\theta_0}" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": ", parametrized by " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " and pre-trained on the inaccessible source data, we aim to improve its performance by updating all its parameters using the unlabeled streaming test data " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " during adaptation. Motivated by the concept of mean teacher [41] as the weight-averaged model over training steps that can yield a more accurate model, we build our online knowledge distillation framework upon teacher-student models. We utilize identical network architecture for teacher and student models. Each model " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "f = h \\circ g" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " comprises a backbone encoder " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "g: \\mathbf{X} \\to \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " mapping unlabeled test image " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbf{X}" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " to the feature representation " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g(\\mathbf{x}) \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " and a classifier (hypothesis) head " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "h: \\mathbb{R}^D \\to \\mathbb{R}^K" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " mapping " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " to the class prediction " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " denote the feature dimension and number of classes. The parameters of teacher " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " and student " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " are first initialized from the source model parameters " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": ". Then the teacher model's parameters " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " are updated from the student model's parameters " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " using an exponential moving average (EMA) with momentum coefficient " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "\\theta_t \\gets \\alpha \\cdot \\theta_t + (1 - \\alpha) \\cdot \\theta_s" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": ". Following [26], we freeze the classifier " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": "'s parameters and only update the encoder " + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 365, + 289, + 629 + ], + "type": "text", + "content": "'s parameters during test-time adaptation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 243, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 243, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 243, + 647 + ], + "type": "text", + "content": "3.1. Rationale Behind the " + }, + { + "bbox": [ + 47, + 635, + 243, + 647 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 635, + 243, + 647 + ], + "type": "text", + "content": "-CE Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 654, + 287, + 712 + ], + "type": "text", + "content": "We start by analyzing the rationale behind the proposed " + }, + { + "bbox": [ + 45, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 45, + 654, + 287, + 712 + ], + "type": "text", + "content": "-CE loss and its benefit for self-learning through the tight connection with the mutual information between the model's predictions and unlabeled test images. Before that, we first define the notations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": "Notations. We define the random variables of unlabeled test images and the predictions from the student model " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " and those of the teacher model " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": "'s soft pseudo-label predictions as " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": ". Furthermore, let " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " denotes the marginal distribution over " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "p_{(\\mathbf{Y},\\mathbf{X})}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " be the joint distribution of " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{Y}|\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " be the conditional distribution of " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": ". Then, the entropy of " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " and the conditional entropy of " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\mathbf{Y}) := \\mathbb{E}_{p_{\\mathbf{Y}}}[-\\log p_{\\mathbf{Y}}(\\mathbf{Y})]" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\mathbf{Y}|\\mathbf{X}) := \\mathbb{E}_{p_{(\\mathbf{Y},\\mathbf{X})}}[-\\log p_{\\mathbf{Y}|\\mathbf{X}}(\\mathbf{Y}|\\mathbf{X})]" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": ", respectively. Besides, let the flipped cross-entropy (" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "f-\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": ") given " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\mathbf{Y};\\hat{\\mathbf{Y}}|\\mathbf{X}) := \\mathbb{E}_{p_{(\\mathbf{Y},\\mathbf{X})}}[-\\log p_{\\hat{\\mathbf{Y}}|\\mathbf{X}}(\\mathbf{Y}|\\mathbf{X})]" + }, + { + "bbox": [ + 304, + 72, + 547, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": "Self-learning and mutual information. Based on the above definitions, the " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " conditioned over " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " has the tight connection with the mutual information " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\mathbf{Y};\\mathbf{X})" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " between unlabeled test images " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " and the predictions from the student model " + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 214, + 547, + 275 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 293, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 293, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {H} (\\mathbf {Y}; \\hat {\\mathbf {Y}} \\mid \\mathbf {X}) = \\mathcal {H} (\\mathbf {Y} \\mid \\mathbf {X}) + \\mathcal {D} _ {\\mathrm {K L}} (\\mathbf {Y} \\| \\hat {\\mathbf {Y}} \\mid \\mathbf {X}) \\tag {1} \\\\ = - \\mathcal {I} (\\mathbf {Y}; \\mathbf {X}) + \\mathcal {H} (\\mathbf {Y}) + \\mathcal {D} _ {\\mathrm {K L}} \\left(\\mathbf {Y} \\| \\hat {\\mathbf {Y}} | \\mathbf {X}\\right) \\\\ \\end{array}", + "image_path": "42790894b5d7270e5c7394153935447a866ea715b66fec36c99088eac3d192af.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": "This implies that minimizing the " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": "-CE " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "(\\mathcal{H}(\\mathbf{Y};\\hat{\\mathbf{Y}}\\mid \\mathbf{X})" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " in Eq. 1) and maximizing the entropy of class-marginal prediction " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\mathbf{Y})" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " is equivalent to maximizing the mutual information between the test images and the student model's predictions " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\mathbf{Y};\\mathbf{X})" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " with a correction KL-divergence term " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{KL}}(\\mathbf{Y}\\parallel \\hat{\\mathbf{Y}}\\mid \\mathbf{X})" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " involving soft-pseudo labels as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 551, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 551, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 551, + 491 + ], + "type": "interline_equation", + "content": "\\underbrace {\\mathcal {H} (\\mathbf {Y} ; \\hat {\\mathbf {Y}} \\mid \\mathbf {X})} _ {f \\cdot \\mathbb {C E}} - \\mathcal {H} (\\mathbf {Y}) = - \\underbrace {\\mathcal {I} (\\mathbf {Y} ; \\mathbf {X})} _ {\\text {M u t u a l I n f o .}} + \\mathcal {D} _ {\\mathrm {K L}} (\\mathbf {Y} \\| \\hat {\\mathbf {Y}} | \\mathbf {X}) \\tag {2}", + "image_path": "2715af9ba51edc5678b54a5fc263268796c2c6de1b45ac08f616b6d1d5bf86b8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "content": "Thus, minimizing the left side of Eq. 2 with respect to the student model's parameters " + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "content": " allows the student model to cluster test images using mutual information and criterion-corrected knowledge distillation from soft-pseudo labels. Using the above formulation, we define the following test-time objective to train the student model " + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "content": " on a batch of " + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "content": " test images " + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "inline_equation", + "content": "X = \\{\\pmb{x}_1,\\dots ,\\pmb{x}_B\\}" + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "content": " with their corresponding soft-pseudo labels from teacher model " + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "inline_equation", + "content": "\\hat{Y} = \\{\\hat{y}_1,\\dots ,\\hat{y}_B\\}" + }, + { + "bbox": [ + 304, + 491, + 547, + 588 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 609, + 545, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 609, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 609, + 545, + 677 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {p l}} (X, \\hat {Y}) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {k = 1} ^ {K} f _ {s} \\left(\\boldsymbol {x} _ {i}\\right) _ {k} \\log \\left(\\left(\\hat {\\boldsymbol {y}} _ {i}\\right) _ {k}\\right) \\\\ + \\sum_ {k = 1} ^ {K} \\hat {f} _ {s} (X) _ {k} \\log (\\hat {f} _ {s} (X)) _ {k} \\tag {3} \\\\ \\end{array}", + "image_path": "5853e6b3d7b162b1038e88c8677139e08386d49bf4c044ed83a12b5f24e5e890.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{f}_s(X) = \\frac{1}{B}\\sum_{i=1}^B f_s(\\pmb{x}_i)" + }, + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "type": "text", + "content": " denotes the marginal class distribution over the batch of test images " + }, + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 687, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20343" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 534, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 534, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 534, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 534, + 178 + ], + "type": "image", + "image_path": "d368dff16a867de483049f5c93f9de3352026a3a91076cbc03a38f08811b65f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": "Figure 2. Overview of TeSLA Framework. (a) The student model " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " is adapted on the test images by minimizing the proposed test-time objective " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": ". The high-quality soft-pseudo labels required by " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " are obtained from the exponentially weighted averaged teacher model " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " and refined using the proposed Soft-Pseudo Label Refinement (PLR) on the corresponding test images. The soft-pseudo labels are further utilized for teacher-student knowledge distillation via " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " on the adversarially augmented views of the test images. (b) The adversarial augmentations are obtained by applying learned sub-policies sampled i.i.d from " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " using the probability distribution " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " with their corresponding magnitudes selected from " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": ". The parameters " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " of the augmentation module are updated by the unbiased gradient estimator (Eq. 13) of the loss " + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Aug}}" + }, + { + "bbox": [ + 46, + 186, + 547, + 264 + ], + "type": "text", + "content": " computed on the augmented test images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 280, + 263, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 263, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 263, + 293 + ], + "type": "text", + "content": "3.2. Knowledge Distillation via Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": "Self-learning from adversarial augmentations. As illustrated in Fig. 1, adversarial augmentations are learned by pushing their feature representations toward the decision boundary (Expansion phase), followed by updating the student model " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " to match its prediction on these augmented images with their corresponding soft-pseudo label (Separation phase), yielding better separation of features into their respective classes. In our method setup, we continually learn automatic augmentations of the test images (Sec. 3.3) that are adversarial to the current teacher model " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " and enforce consistency between the student model's " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " predictions on the augmented views and the corresponding soft-pseudo labels. Since we freeze the classifier module, " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " share the common decision boundary. Let " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " denote the learned adversarial augmented view of " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{y}}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " denote the corresponding soft-pseudo label. We minimize the following knowledge distillation loss " + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 298, + 290, + 514 + ], + "type": "text", + "content": " using KL-divergence to distill knowledge from the teacher model to the student model:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 526, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 526, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 109, + 526, + 287, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k d}} (\\tilde {\\boldsymbol {x}}, \\hat {\\boldsymbol {y}}) = \\mathcal {D} _ {\\mathrm {K L}} (\\hat {\\boldsymbol {y}} \\| f _ {s} (\\tilde {\\boldsymbol {x}})) \\tag {4}", + "image_path": "c0ff18221a3e63b5cd1078669d7f0b9fc5f8f0e1ab2f3162816a50b04329b0da.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "type": "text", + "content": "Soft pseudo label refinement (PLR). We refine the quality of soft pseudo-labels and the feature representations by averaging the teacher model's outputs on multiple weakly augmented image views " + }, + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "type": "inline_equation", + "content": "\\rho_w(\\pmb{x})" + }, + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "type": "text", + "content": " (via FLIPPING, CROPPING augmentations) of the same test image " + }, + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 550, + 288, + 611 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 622, + 287, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 622, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 90, + 622, + 287, + 636 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t}, \\mathbf {y} _ {t} \\leftarrow \\mathbb {E} _ {\\mathbf {u} \\in \\rho_ {w} (\\boldsymbol {x})} \\left[ g _ {t} (\\mathbf {u}), h _ {t} \\left(g _ {t} (\\mathbf {u})\\right) \\right] \\tag {5}", + "image_path": "7c29b22e0d2350bf407a57218262237eee07a38af54a4e13ec3bd18b5d8673ac.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "The refined feature representations " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and the softmaxed pseudo labels " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " from the encoder " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and the classifier " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " are further stored in an online memory queue " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " of fixed size. The final soft pseudo-label is computed by averaging the refined soft pseudo-labels of " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "-nearest neighbors of the current test image in the feature space as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 290, + 545, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 290, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 354, + 290, + 545, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Q} [ \\arg \\max (\\boldsymbol {y} _ {t}) ]. \\text {a p p e n d} \\left(\\left\\{\\boldsymbol {z} _ {t}, \\boldsymbol {y} _ {t} \\right\\}\\right) \\\\ \\hat {\\boldsymbol {y}} = \\frac {1}{n} \\sum \\mathcal {N} _ {\\mathbf {Q}, n} (\\mathbf {z} _ {t}) \\tag {6} \\\\ \\end{array}", + "image_path": "0396d4768458debe4371d2bc54580b9c822ef054429938c4724c07c3a57f0242.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": " denotes an online memory of class-balanced queues (with size " + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "inline_equation", + "content": "|\\mathbf{Q}[\\arg \\max (\\pmb {y}_t)]| \\leq N_{\\mathbf{Q}}" + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": ") of the refined feature representations and the softmaxed label predictions on the previously seen test images, and " + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\mathbf{Q},n}(\\mathbf{z}_t)" + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": " denotes soft labels of " + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 334, + 547, + 394 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 402, + 524, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 524, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 524, + 415 + ], + "type": "text", + "content": "3.3. Learning Adversarial Data Augmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": "This section presents our efficient online method for learning adversarial data augmentation for the current teacher " + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": ". We first introduce our adversarial augmentation search space. Then, we describe our differentiable strategy to optimize and sample adversarial augmentations that push the feature representations of the test images toward the uncertain region in the close vicinity of the decision boundary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "content": "Policy search space. Let " + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "content": " be a set of all image transformation operations " + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{O}:\\mathbb{R}^{H\\times W\\times 3}\\to \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "content": " defined in our search space. In particular, we use image transformations AUTO-CONTRAST, EQUALIZE, INVERT, SOLARIZE, POSTERIZE, CONTRAST, BRIGHTNESS, COLOR, SHEARX, SHEARY, TRANSLATEX, TRANSLATEY, ROTATE, AND SHARPNESS. Each transformation " + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "content": " is specified with its magnitude parameter " + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "inline_equation", + "content": "m\\in [0,1]" + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "content": ". Since the output of some image operations may not be conditional on its magnitude (e.g., EQUALIZE) or may not be differentiable (e.g., POSTERIZE), for such image operations, we use the straight-through gradient estimate w.r.t. each pixel as: " + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{O}(\\boldsymbol{x}_{ij}) / \\partial m = 1" + }, + { + "bbox": [ + 304, + 510, + 547, + 654 + ], + "type": "text", + "content": " for magnitude optimization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "text", + "content": "We define a sub-policy " + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "text", + "content": " as a combination of " + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "text", + "content": " image operations (sub-policy dimension) from " + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "text", + "content": " that is applied sequentially to a given image " + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 654, + 547, + 690 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 700, + 546, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 700, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 358, + 700, + 546, + 715 + ], + "type": "interline_equation", + "content": "\\rho \\left(\\boldsymbol {x}, m ^ {\\rho}\\right) = \\mathcal {O} _ {1} ^ {\\rho} \\dots \\mathcal {O} _ {N} ^ {\\rho} \\left(\\boldsymbol {x}, m _ {N} ^ {\\rho}\\right) \\tag {7}", + "image_path": "345b28f4036704b6e9f147417d5371a3865eaafc65189c52d4c1aeb98104970a.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20344" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "m^{\\rho} = [m_{1}^{\\rho},\\dots,m_{N}^{\\rho}]" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": " denotes the magnitude set of the sub-policy " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": ". We also define " + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "inline_equation", + "content": "\\mathbb{P} = \\{\\rho_1,\\rho_2,\\ldots \\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 109 + ], + "type": "text", + "content": " as the set of all possible sub-policies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "spans": [ + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "text", + "content": "Policy evaluation. We propose learnable adversarial augmentation, aiming to optimize and sample data augmentations as a proxy for simulating data in the uncertain region of the feature space. Given the teacher model " + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "text", + "content": ", a sub-policy " + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "text", + "content": " with magnitude " + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "text", + "content": " is evaluated for a test image " + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 113, + 289, + 185 + ], + "type": "text", + "content": " using the following adversarial objective:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 203, + 288, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 288, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho) = \\sum_ {k = 1} ^ {K} f _ {t} (\\tilde {\\boldsymbol {x}}) \\log \\left(f _ {t} (\\tilde {\\boldsymbol {x}})\\right) + \\lambda_ {1} r (\\tilde {\\boldsymbol {x}}, \\boldsymbol {x}) \\tag {8}", + "image_path": "90904255e1bc2cc9d3892244601506d7002799611209b8f602fddaafb335282f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}} = \\rho (\\pmb {x},m)" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " denotes the adversarial augmented image and " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "r(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " regularizes augmentation severity. The hyperparameter " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " controls the augmentation severity. We optimize the adversarial augmentation loss " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " with respect to the sub-policy " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " 's parameters. Minimizing " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aug}}" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " is equivalent to maximizing the entropy of the teacher model's prediction on the augmented image " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " thus pushing its feature representation towards the decision boundary (first term). For the augmentation regularization (second term), we use the mean squared distance function between the teacher encoder's " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " internal layers' activations of the adversarial augmentation and non-augmented versions of image " + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 245, + 289, + 389 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 398, + 288, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 398, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 93, + 398, + 288, + 430 + ], + "type": "interline_equation", + "content": "r (\\tilde {\\boldsymbol {x}}, \\boldsymbol {x}) = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\| \\mu_ {l} (\\tilde {\\boldsymbol {x}}) - \\mu_ {l} (\\boldsymbol {x}) \\| ^ {2} \\tag {9}", + "image_path": "9f83f34dff24e8ef9a01eb75400da780d57697c9556bf1ba7e6a50dced19f20a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "inline_equation", + "content": "\\mu_l(\\cdot)" + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "text", + "content": " denotes the mean activation of the " + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "text", + "content": " layer of the teacher encoder " + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 46, + 436, + 288, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": "Policy optimization. Let " + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "inline_equation", + "content": "M = [m^{\\rho_1}, \\ldots, m^{\\rho_{\\| \\mathbb{P} \\|}}]" + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": " denote the set of magnitudes " + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "inline_equation", + "content": "m^{\\rho}" + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": " of all sub-policies " + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "inline_equation", + "content": "\\rho \\in \\mathbb{P}" + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "inline_equation", + "content": "P = [p_1, \\ldots, p_{\\| \\mathbb{P} \\|}]" + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": " denote the probability of selecting a sub-policy from " + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": ". The expected policy evaluation loss (Eq. 8) for an image " + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 464, + 289, + 525 + ], + "type": "text", + "content": " over the policy search space is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 533, + 288, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 533, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 97, + 533, + 288, + 567 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}) \\right] = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} p _ {i} \\cdot \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\tag {10}", + "image_path": "a172e6fd1aefdb8113c9c487e1c8fa54eae11e2eecd1fb7c2d62363d0a69105a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "content": "Evaluating the gradient of " + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathcal{L}_{\\mathrm{aug}}(\\pmb{x})]" + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "content": " can become computationally expensive as " + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "inline_equation", + "content": "\\| \\mathbb{P}\\| \\sim \\binom{\\|\\mathbb{O}\\|}{N}" + }, + { + "bbox": [ + 46, + 572, + 288, + 622 + ], + "type": "text", + "content": ". Thus, we use the following re-parameterization trick to estimate its unbiased gradient:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 641, + 288, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 641, + 288, + 711 + ], + "spans": [ + { + "bbox": [ + 59, + 641, + 288, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla \\mathbb {E} [ \\mathcal {L} _ {\\mathrm {a u g}} (x) ] = \\delta (\\boldsymbol {x}, \\mathbb {P}) = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} \\nabla \\left(p _ {i} \\cdot \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i})\\right) \\tag {11} \\\\ = \\sum_ {i = 1} ^ {\\| \\mathbb {P} \\|} p _ {i} \\left(\\nabla \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) + \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\cdot \\nabla \\log p _ {i}\\right) \\\\ \\end{array}", + "image_path": "91244cf72885d3ffc062a935e17647f84ea3a8a708a0a691c534e143bec6ffd7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 537, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 537, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 537, + 84 + ], + "type": "text", + "content": "Thus, an unbiased estimator of " + }, + { + "bbox": [ + 306, + 72, + 537, + 84 + ], + "type": "inline_equation", + "content": "\\delta (\\pmb {x},\\mathbb{P})" + }, + { + "bbox": [ + 306, + 72, + 537, + 84 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 94, + 547, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 94, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 318, + 94, + 547, + 109 + ], + "type": "interline_equation", + "content": "\\hat {\\delta} (\\boldsymbol {x}, \\rho_ {i}) = \\nabla \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) + \\mathcal {L} _ {\\mathrm {a u g}} (\\boldsymbol {x}, \\rho_ {i}) \\cdot \\nabla \\log p _ {i} \\tag {12}", + "image_path": "52892114baa6d7910a5da1673e8bca2f85f360336104ae3c005badfc3cbca299.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": "where index " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " is sampled from the probability distribution " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": ". For an online batch of " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " test images " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}_1, \\pmb{x}_2, \\dots, \\pmb{x}_B\\}" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": ", we apply augmentation sub-policies " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "\\{\\rho_{i_1}, \\rho_{i_2}, \\dots, \\rho_{i_B}\\}" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "\\{i_1, i_2, \\dots, i_B\\}" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " are sampled i.i.d from distribution " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": ", and update the parameters " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 118, + 547, + 190 + ], + "type": "text", + "content": " with the following stochastic gradient update rule:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 199, + 546, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 199, + 546, + 233 + ], + "spans": [ + { + "bbox": [ + 349, + 199, + 546, + 233 + ], + "type": "interline_equation", + "content": "[ P, M ] \\leftarrow [ P, M ] - \\frac {\\gamma}{B} \\sum_ {j = 1} ^ {B} \\hat {\\delta} \\left(\\boldsymbol {x} _ {j}, \\rho_ {i _ {j}}\\right) \\tag {13}", + "image_path": "e836bfc2ebb59fe9ef33f4ab10ae022a44f523e6c5501ebb04a0d5ce9eed6e90.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "type": "text", + "content": " is the learning rate. We set " + }, + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\gamma = 0.1" + }, + { + "bbox": [ + 305, + 242, + 547, + 266 + ], + "type": "text", + "content": " for all experiments without the need for hyperparameter tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 274, + 545, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 545, + 286 + ], + "type": "text", + "content": "3.4. Self-Learning With Adversarial Augmentation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": "We minimize the following overall objective " + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TeSLA}}" + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": " for training the student model " + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": " on a batch of " + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": " test images " + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "inline_equation", + "content": "X = \\{\\pmb {x}_1,\\dots ,\\pmb {x}_B\\}" + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": " and their adversarial augmented views " + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\hat{X} = \\{\\tilde{x}_1,\\dots ,\\tilde{x}_B\\}" + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": " with the corresponding refined softpseudo labels from the teacher model " + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\hat{Y} = \\{\\hat{y}_1,\\dots ,\\hat{y}_B\\}" + }, + { + "bbox": [ + 306, + 292, + 547, + 354 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 371, + 546, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 371, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 310, + 371, + 546, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T e S L A}} (X, \\tilde {X}, \\hat {Y}) = \\mathcal {L} _ {\\mathrm {p l}} (X, \\hat {Y}) + \\frac {\\lambda_ {2}}{B} \\sum_ {i = 1} ^ {B} \\mathcal {L} _ {\\mathrm {k d}} (\\tilde {\\boldsymbol {x}} _ {i}, \\hat {\\boldsymbol {y}} _ {i}) \\tag {14}", + "image_path": "253e0a2b832f11a296115c79be178768dfbb087bd4d7b7761650f36714017622.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " is a hyper-parameter for the knowledge distillation. The adversarial augmented views " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\tilde{X}" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " are obtained by sampling i.i.d. augmentation sub-policies from " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " for every test image in " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " with probability " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " and applying the corresponding magnitude from " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": ". Concurrently, we also optimize " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 412, + 547, + 484 + ], + "type": "text", + "content": " using the policy optimization mentioned in Sec. 3.3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 495, + 388, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 388, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 388, + 509 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 515, + 495, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 495, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 495, + 528 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 534, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 534, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 305, + 534, + 547, + 629 + ], + "type": "text", + "content": "We evaluate and compare TeSLA against state-of-the-art (SOTA) test-time adaptation algorithms for both classification and segmentation tasks under three types of test-time distribution shifts resulting from (1) common image corruption, (2) synthetic to real data transfer, and (3) measurement shifts on medical images. The latter is characterized by a change in medical imaging systems, e.g., different scanners across hospitals or various staining techniques." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "content": "TTA protocols. We adopt the TTA protocols in [39] and categorize competing methods based on two factors: (i) source training objective and (ii) sequential or non-sequential inference. First, unlike [39], we use " + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "content": " to indicate if access to the source domain's information, e.g., feature statistics is possible or if the source training objective is allowed to be modified; otherwise, we use " + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "content": ". Next, we use " + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 305, + 630, + 547, + 713 + ], + "type": "text", + "content": " to indicate a" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20345" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "one-pass adaptation and evaluation protocol (one epoch) for sequential test data and " + }, + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": " to show a multi-pass adaptation on all test data and inference after several epochs. Thus, we end up with four possible TTA protocols, namely N-M, Y-M, N-O, and Y-O. Unlike previous works [28,39], TeSLA does not rely on any source domain feature distribution, yielding two realistic TTA protocols: N-M and N-O." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "text", + "content": "Hyperparameters for all experiments. We set " + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_2 = 1.0" + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "text", + "content": " for all experiments (cf. ablation in Appendix E). For the adversarial augmentation module, we use sub-policy dimension " + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "text", + "content": " (except for VisDA-C [33], " + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "inline_equation", + "content": "N = 4/3" + }, + { + "bbox": [ + 46, + 158, + 288, + 266 + ], + "type": "text", + "content": " for N-O/N-M protocols) (cf. Appendix E) and Adam optimizer [22]. For pseudo-label refinement (PLR), we use five weak augmentations composed of random flips and resize crop. More details of hyperparameters used for each experiment can be found in Appendix A." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 397 + ], + "type": "text", + "content": "Common image corruptions. To evaluate TeSLA's efficacy for the classification task, we use CIFAR10-C/CIFAR100-C [16] and large-scale ImageNet-C [16] datasets each containing 19 types of corruptions applied to the clean test set with five levels of severity. We perform validation on 4 out of 19 types of corruption (SPATTER, GAUSSIAN BLUR, SPECKLE NOISE, SATURATE) to select hyperparameters and test on the remaining 15 corruptions at the maximum severity level of 5. Following [28], we train the ResNet50 [15] on the clean CIFAR10/CIFAR100/ImageNet training set and adapt it to classify the unlabeled corrupted test set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 508 + ], + "type": "text", + "content": "Synthetic to real data adaption. We use challenging and large-scale VisDA-C, and VisDA-S [33] datasets for evaluating synthetic-to-real data adaptation at test-time for classification and segmentation tasks, respectively. Following [4, 26], we adapt the ResNet101 network pre-trained on synthetic images to classify 12 vehicle classes on the photo-realistic images of VisDA-C. While for VisDA-S, we adapt the DeepLab-v3 [5] backbone pre-trained on synthetic GTA5 [36] to the Cityscapes [7] to segment 19 classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "text", + "content": "Measurement shifts on medical images. We access TeSLA on staining variations for the tissue type classification of hematoxylin & eosin (H&E) stained patches from colorectal cancer tissue slides. We use MobileNetV2 [38] trained on the source Kather-19 dataset [20] and adapt it to the target Kather-16 dataset [21] on four tissue categories: tumor, stroma, lymphocyte, and mucosa. We also evaluate TeSLA for variations in scanners and imaging protocols in multi-site medical images on two magnetic resonance imaging (MRI) datasets for the segmentation task, namely the multi-site prostate MRI [27] and spinal cord grey matter segmentation (SCGM) [34]. Following [42], we adapt the U-Net [37] from site 1 to sites " + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\{2,3,4\\}" + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "text", + "content": " of the spinal cord dataset, and from sites " + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{A,B}\\}" + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "text", + "content": " to sites " + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{D,E,F}\\}" + }, + { + "bbox": [ + 46, + 509, + 288, + 676 + ], + "type": "text", + "content": " of the prostate dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Competing baselines. We compare TeSLA with the following SFDA and TTA baselines, including direct inference of the trained source model on the target test data without" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 145, + 547, + 258 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 138 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 138 + ], + "type": "text", + "content": "Table 1. Comparison of SOTA TTA methods under different protocols evaluated on CIFAR-10/100-C, ImageNet-C, VisDA-C and Kather-16 datasets. We report the average error computed over 15 test corruptions for the common image corruption shifts. We also report Class Avg. in % error rates for synthetic-to-real and measurement shifts over 3 and 10 seeds, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 145, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 308, + 145, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 145, + 547, + 258 + ], + "type": "table", + "html": "
MethodProtocolCommon Image CorruptionsSyn-to-RealMeasurement Shift
CIFAR10-CCIFAR100-CImageNet-CVisDA-CKather-16
OMOMOMOMOM
SourceN29.160.481.851.532.0
BN [18,30]N15.615.443.743.367.767.635.435.018.318.2
TENT [45]N14.112.939.036.557.454.233.529.316.212.0
SHOT [26]N13.914.239.238.768.768.229.424.514.712.0
AdaContrast [4]N-23.120.2-
TTT++ [28]Y15.89.844.434.159.3-35.234.116.77.9
TTAC [39]Y13.49.441.733.658.7-32.231.19.65.5
TeSLAN12.59.738.232.955.051.517.813.59.23.3
TeSLA-sY12.19.737.332.653.1-24.017.99.93.1
", + "image_path": "c94b321f2ebfd4eea560c891857dca2c4b0f4eb11dba94009e0c10369b54a064.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "type": "text", + "content": "adaptation (Source). We also implement the SFDA-based pseudo-labeling baselines from a basic pseudo-labeling approach (PL) to the SHOT approach [26] based on training the feature extraction module by adopting the information maximization loss to make globally diverse but individually certain predictions on the target domain. For recent TTA methods, we compare our method against " + }, + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{TTT}++" + }, + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "type": "text", + "content": " [28], Ada-Contrast [4], CoTTA [47], basic TENT [45] and BN [18] as representative methods based on feature distribution alignment via stored statistics, contrastive learning, augmentation-averaged predictions, entropy minimization, and batch normalization statistic. Furthermore, we compare TeSLA to more recent methods based on test-time augmentation policy learning, OptTTA [42], and anchor clustering TTAC [39]. By default, [4,18,26] follow the N-M protocol, and we adapt them to the N-O protocol, while [42,45,47], by default, follow the N-O protocol and adapt them to the N-M protocol setting. Using source features distribution, TTAC and " + }, + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{TTT}++" + }, + { + "bbox": [ + 304, + 265, + 548, + 553 + ], + "type": "text", + "content": " follow Y-O and Y-M. We report error rates on classification tasks and Mean Intersection over Union (mIoU)/ Dice scores on segmentation tasks. For consistency across baselines and TTA protocols, we do not use any specialized model architectures (e.g., projection head, weight-normalized classifier layer) for SHOT [26] and AdaContrast [4]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 562, + 362, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 362, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 362, + 574 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Classification task. In Table 1, we summarize the average (Avg) classification error rates " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " of TeSLA against several SOTA SFDA and TTA baselines on the common image corruptions of CIFAR10-C, CIFAR100-C, ImageNet-C; synthetic to real data domain shifts of VisDA-C; and measurement shifts on the Kather-16 dataset. We also report per-class top-1 accuracies for VisDA-C and Kather-16 datasets and corruption-wise error rates on CIFAR-10-C/CIFAR-100-C and ImageNet-C datasets in Appendix B. TeSLA easily surpasses all competing methods under the protocols N-O and N-M on all the datasets. In particular, TeSLA outper" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20346" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 71, + 289, + 168 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 289, + 168 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 289, + 168 + ], + "type": "image", + "image_path": "5481e6f89e003844907e5351db3ec8d7f9fbe209cb6f605e759a00edcd5361d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 170, + 195, + 178 + ], + "lines": [ + { + "bbox": [ + 184, + 170, + 195, + 178 + ], + "spans": [ + { + "bbox": [ + 184, + 170, + 195, + 178 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 68, + 386, + 168 + ], + "blocks": [ + { + "bbox": [ + 301, + 68, + 386, + 168 + ], + "lines": [ + { + "bbox": [ + 301, + 68, + 386, + 168 + ], + "spans": [ + { + "bbox": [ + 301, + 68, + 386, + 168 + ], + "type": "image", + "image_path": "51a933acca7eabc2edb1b3dca9907571ec6ab9ca6b60f8439724348436d09b4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 68, + 475, + 168 + ], + "blocks": [ + { + "bbox": [ + 391, + 68, + 475, + 168 + ], + "lines": [ + { + "bbox": [ + 391, + 68, + 475, + 168 + ], + "spans": [ + { + "bbox": [ + 391, + 68, + 475, + 168 + ], + "type": "image", + "image_path": "89ef494eb32bc41feff3ecfad170cab501eecd14ca45eaeedc89557b927532f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 170, + 408, + 178 + ], + "lines": [ + { + "bbox": [ + 396, + 170, + 408, + 178 + ], + "spans": [ + { + "bbox": [ + 396, + 170, + 408, + 178 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 479, + 68, + 503, + 160 + ], + "blocks": [ + { + "bbox": [ + 479, + 68, + 503, + 160 + ], + "lines": [ + { + "bbox": [ + 479, + 68, + 503, + 160 + ], + "spans": [ + { + "bbox": [ + 479, + 68, + 503, + 160 + ], + "type": "image", + "image_path": "5dc96ef47a6b961c8efc22b5ca63c6d510532cfa72c6860e7c1a408e7fc548fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 247, + 164, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 189, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 234 + ], + "type": "text", + "content": "Figure 3. Ablation experiments on the VisDA-C. (a) Source training strategy and model architecture. TeSLA outperforms the competing TTA methods across varied source training strategies (Supervised, SimCLRV2 [6], and DINO [3]) and model architectures (ResNet50, ResNet101 [15], and ViT-B/16 [11]) using the same set of hyperparameters for both protocols (N-O/N-M). Each vertex represents the mean class avg accuracy over 3 seeds. (b) t-SNE visualization [43] comparison of feature embeddings for AdaContrast [4] and TeSLA." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 247, + 164, + 350 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 164, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 164, + 350 + ], + "type": "image", + "image_path": "aa75a6905b65285834f8f181ef46ea44c25bd404a58608b3b943a9dee2ed7683.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 171, + 247, + 288, + 350 + ], + "blocks": [ + { + "bbox": [ + 171, + 247, + 288, + 350 + ], + "lines": [ + { + "bbox": [ + 171, + 247, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 171, + 247, + 288, + 350 + ], + "type": "image", + "image_path": "733c05ae857a8f2ac9c803d3f672b8f392aa59ff42b2f12cb264332f3671cab8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 288, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 413 + ], + "type": "text", + "content": "Figure 4. Model calibration and uncertainty estimation. Model calibration using reliability diagrams [31] and model uncertainty estimation using Brier Score and Negative Log Likelihood (NLL) of the Source Model, TENT [45], OptTTA [42] and, TeSLA on the prostate dataset for test-time adaptation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 60, + 445, + 274, + 479 + ], + "blocks": [ + { + "bbox": [ + 46, + 419, + 288, + 442 + ], + "lines": [ + { + "bbox": [ + 46, + 419, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 288, + 442 + ], + "type": "text", + "content": "Table 2. Semantic segmentation results (Class Avg. mIoU in %) on the VisDA-S (GTA5 → Cityscapes) test-time adaptation task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 445, + 274, + 479 + ], + "lines": [ + { + "bbox": [ + 60, + 445, + 274, + 479 + ], + "spans": [ + { + "bbox": [ + 60, + 445, + 274, + 479 + ], + "type": "table", + "html": "
ProtocolSourceBN [30]TENT [45]PLCoTTA [47]TeSLA
NO35.336.738.338.837.044.5
NM38.439.238.639.946.0
", + "image_path": "850278dec17a7f33c413d5d9711da22eedcb00a679ca0bc3a4dc640abae35b43.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 486, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 665 + ], + "type": "text", + "content": "forms the second-best baseline under N-O/N-M protocol by a " + }, + { + "bbox": [ + 46, + 486, + 289, + 665 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 486, + 289, + 665 + ], + "type": "text", + "content": " margin 1.4/3.2 on CIFAR10-C, 0.8/3.6 on CIFAR100-C, 2.4/2.7 on ImageNet-C, 5.3/6.7 on VisDA-C, and 5.5/8.7 on the Kather-16 datasets. Despite not utilizing any source feature statistic, TeSLA (N-O/N-M) is either competitive or surpasses TTAC [39] and TTT++ [28] (Y-O/Y-M) that benefit from the source dataset feature statistics during adaptation. Including the global source feature alignment module of TTAC with TeSLA (TeSLA-s) could further improve its performance under (Y-O/Y-M) protocols for slight domain shifts (e.g., common image corruptions). We report the results of AdaContrast for VisDA-C only as the implementation for other datasets are not provided. We do not report TeSLA-s results on ImageNet-C under the Y-M protocol, as previous baselines were evaluated only Y-O protocol." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Segmentation task. Unlike TeSLA, several TTA methods e.g., AdaContrast [4], SHOT [26], TTT++ [28], TTAC [39], have only been evaluated on the classification task. Therefore, we compare TeSLA against the current SOTA TTA" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 308, + 289, + 547, + 365 + ], + "blocks": [ + { + "bbox": [ + 304, + 247, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 281 + ], + "type": "text", + "content": "Table 3. Semantic segmentation results (Class Avg. volume-wise mean Dice in %) on the cross-site spinal cord and prostate MRI test-time adaptation tasks." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 289, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 308, + 289, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 547, + 365 + ], + "type": "table", + "html": "
ProtocolSourceBN [30]TENT [45]PLOptTTA [42]TeSLA
Spinal CordSite{1} → Sites{2,3,4}
NO76.0±11.881.6±8.381.1±9.181.7±8.684.1±4.885.3±5.8
NM84.3±4.884.4±4.784.3±4.784.3±4.485.4±4.4
ProstateSites{A,B} → Sites{D,E,F}
NO60.5±27.072.1±15.274.7±17.972.4±15.283.1±7.783.5±6.5
NM73.1±18.081.2±9.381.1±9.283.4±7.784.3±5.8
", + "image_path": "6418cfad5fc98528e22d4a4ad06fd3f77e25b68c79776683c75cea719190e5e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "type": "text", + "content": "methods applied to the segmentation task on synthetic-to-real data transfer of the VisDA-S dataset in Table 2. TeSLA significantly outperforms the competing methods and achieves the best mIOU scores for both N-O and N-M protocols, beating the second-best baseline by a margin of " + }, + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "type": "inline_equation", + "content": "+5.7\\%" + }, + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "type": "inline_equation", + "content": "+6.1\\%" + }, + { + "bbox": [ + 304, + 373, + 547, + 516 + ], + "type": "text", + "content": ", respectively. In Table 3, we compare TeSLA against the recent SOTA test-time augmentation policy method, OptTTA [42] and other TTA baselines for the inter-site adaptation on two challenging MRI datasets mentioned in Sec. 4.1. TeSLA convincingly outperforms all other methods on severe measurement shifts across sites under N-O and N-M protocols." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 518, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 601 + ], + "type": "text", + "content": "Computational cost for adversarial augmentations. As shown in Table 4, TeSLA learns the adversarial augmentation in an online manner, and thus its runtime is several orders faster than learnable test-time augmentation OptTTA [42], while comparable to that of static augmentation policies with an additional overhead of only 0.10 GPU hours/epoch on the VisDA-C dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 604, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 663 + ], + "type": "text", + "content": "Test-time feature visualization. Fig. 3b compares t-SNE projection [43] of the encoder's features of AdaContrast [4] with TeSLA on the VisDA-C. TeSLA shows better inter-class separation for features than AdaContrast, as supported by an improved silhouette score from 0.149 to 0.271." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Model calibration and uncertainty estimation. Entropy minimization-based TTA methods [26, 45] explicitly make the model confident in their predictions, resulting in poor model calibration. For trusting the adapted model, it should" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20347" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 120, + 285, + 173 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 115 + ], + "type": "text", + "content": "Table 4. Performance comparison of adversarial augmentation learned by TeSLA during adaptation against the prior art augmentation methods on runtime (GPU hours/epoch on NVIDIA GeForce RTX 3090) and Class Avg. accuracy in % on the VisDA-C dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 120, + 285, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 120, + 285, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 120, + 285, + 173 + ], + "type": "table", + "html": "
Augmentation TypeOptTTATeSLARA [10]TeSLAAA [8]TeSLAN=2TeSLAN=3
Runtime4.500.050.060.160.18
Class Avg. Acc. (N-O)-80.281.281.582.2
Class Avg. Acc. (N-M)-84.785.786.386.5
", + "image_path": "8d88846d76a7c506b0779c94fc0476e693ab0251a53c6b0fa4fd18e277251acc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "type": "text", + "content": "output reliable confidence estimates matching its true underlying performance on the test images. Such a well-behaved model is characterized by expected calibration error (ECE) through reliability diagram [31], and uncertainty metrics of brier score and negative log-likelihood (NLL) [14]. In Fig. 4, we plot the reliability diagram (dividing the probability range [0, 0.5] into 10 bins), and uncertainty metrics of the Source model, TENT [45] and OptTTA [42] against TeSLA for the inter-site model adaptation on the prostate dataset [27]. We observe that TeSLA achieves the best model calibration (lowest ECE of " + }, + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "type": "inline_equation", + "content": "1.80\\%" + }, + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "type": "text", + "content": ") and lowest Brier and NLL scores of 0.12 and 0.24 (" + }, + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "type": "inline_equation", + "content": "p < 0.006" + }, + { + "bbox": [ + 46, + 181, + 288, + 326 + ], + "type": "text", + "content": " against TENT)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 334, + 147, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 147, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 147, + 345 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": "In this section, we scrutinize the roles played by different components of TeSLA. All ablations are performed on the synthetic-to-real test data adaptation of the VisDA-C dataset. Loss functions for self-learning. We compare TTA performance of our self-learning objective " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": " with the proposed flipped cross-entropy loss " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": "-CE and the basic cross-entropy loss " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": " in Fig. 5-(1). We do not use the proposed augmentation module and pseudo-label refinement in this experiment. We observe that " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": "-CE alone improves the accuracy of the Source model in the N-M protocol by " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "+23.1\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "+16.6\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": " improvement by " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": " (Eq. 3) further improves the accuracy to a margin of " + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "inline_equation", + "content": "+23.8\\%" + }, + { + "bbox": [ + 46, + 353, + 287, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": "Contribution of individual components: PLR, " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": ". Our test-time objective " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TeSLA}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " has three components (i) self-learning loss " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": ", (ii) pseudo-label refinement (PLR), and (iii) knowledge distillation " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " with adversarial augmentation. In Fig. 5-(3), we study the effect of accruing individual components to " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TeSLA}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " (starting from source trained model) for the test-time adaptation accuracy. We observe that " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " alone improves the model's accuracy from " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "48.5\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "72.3\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " while adding PLR and " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " further boosts it to " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "81.6\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "(+9.3)" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "86.3\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "(+4.7)" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "PLR: ensembling on weak augmentations (Ens) and nearest neighbor averaging (NN). We also conduct an ablation on the two pseudo-label refinement approaches - (i) ensembling the teacher model's output on five weak augmentations. (ii) averaging soft pseudo labels of " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "n = 10" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " nearest neighbors in the feature space (cf. sensitivity test in Appendix E). Fig. 5-(2) shows that combining both approaches gives better results than using them individually." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 385, + 156 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 385, + 156 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 385, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 385, + 156 + ], + "type": "image", + "image_path": "8d3f41fc7804936003a516f0bb3545a2b728a6817358273bbe0a209dcc23190b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "text", + "content": "Figure 5. Ablation experiments on the VisDA-C. (left to right) (1) Ablation results for self-learning loss functions, including " + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "inline_equation", + "content": "f\\text{-}\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathbb{C}\\mathbb{E}" + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "text", + "content": "; (2) ablation on different PLR methods: assembling predictions on 5 weak augmentations, averaging predictions of the 10 nearest neighbors in the feature space; (3) ablation on different components of TeSLA's objective function: " + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pl}}" + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "text", + "content": ", PLR, and " + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 304, + 162, + 547, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 460, + 156 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 460, + 156 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 460, + 156 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 460, + 156 + ], + "type": "image", + "image_path": "e54867bba1455f679010f097741a460ee3810386624bea3ffd5d1af8a9ba4043.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 72, + 539, + 156 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 539, + 156 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 539, + 156 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 539, + 156 + ], + "type": "image", + "image_path": "b204bfa589e44d85f6571bcdea810af261121e04f110a54d4ecd08a1c7d81a96.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "text", + "content": "Adversarial augmentation. Table 4 shows the benefit of using our adversarial augmentation " + }, + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "inline_equation", + "content": "(N = 2,3)" + }, + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "text", + "content": " with TeSLA against (i) RandAugment (RA) [10], (ii) AutoAugment (AA) [9] (optimized for ImageNet), and (iii) OptTTA [42]. We achieve far better Class Avg. accuracy on the VisDA-C at the cost of slight runtime overhead compared to static augmentation policies for protocols (N-O, N-M). In addition, TeSLA augmentation (TeAA) consistently improves the performance of other TTA methods (Table 5)." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 381, + 533, + 411 + ], + "blocks": [ + { + "bbox": [ + 304, + 354, + 546, + 376 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 376 + ], + "type": "text", + "content": "Table 5. Class Avg. accuracy (%) of TeSLA augmentation (TeAA) with other TTA objectives under N-O protocol on VisDA-C dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 381, + 533, + 411 + ], + "lines": [ + { + "bbox": [ + 317, + 381, + 533, + 411 + ], + "spans": [ + { + "bbox": [ + 317, + 381, + 533, + 411 + ], + "type": "table", + "html": "
TENT+TeAA| SHOT+TeAA| AdaContrast+TeAA
66.572.0| 70.673.1| 76.979.1
", + "image_path": "a9dbea93fcc28b84628336ebc80e3a0791ba8b69931650cea705212ddc4ed460.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 418, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 525 + ], + "type": "text", + "content": "Source training strategies and model architectures. The ablation results (Fig. 3a) on the VisDA-C show that TeSLA surpasses competing TTA baselines for different source training strategies, including Supervised and self-supervised (SimCLRV2 [6], DINO [3]). Moreover, we ablate TeSLA over different architectures, ranging from ResNet-50/101 [15] to the recent vision transformer ViT-B [11] using the same set of hyperparameters, showing the merits of no reliance on the network architecture or source training strategy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 526, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 550 + ], + "type": "text", + "content": "Finally, in Appendix E, we provide detailed sensitivity results with respect to all hyperparameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "We introduced TeSLA, a novel self-learning algorithm for test-time adaptation that utilizes automatic adversarial augmentation. TeSLA is agnostic to the model architecture and source training strategies and gives better model calibration than other TTA methods. Through extensive experiments on various domain shifts (measurement, common corruption, synthetic to real), we show TeSLA's superiority over previous TTA methods on classification and segmentation tasks. Note that TeSLA assumes class uniformly when implicitly maximizing mutual information and can be improved by incorporating prior, e.g., class label distribution statistics." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20348" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised robust domain adaptation without source data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009-2018, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Armenii Ashukha, Alexander Lyzhov, Dmitry Molchanov, and Dmitry Vetrov. Pitfalls of in-domain uncertainty estimation and ensembling in deep learning. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 294 + ], + "type": "text", + "content": "[4] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 295-305, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 339 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 387, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 288, + 451 + ], + "type": "text", + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 487 + ], + "type": "text", + "content": "[8] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 542 + ], + "type": "text", + "content": "[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 600 + ], + "type": "text", + "content": "[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 667 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[12] Cian Eastwood, Ian Mason, Christopher K. I. Williams, and Bernhard Scholkopf. Source-free adaptation to measurement shift via bottom-up feature restoration. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 160 + ], + "type": "text", + "content": "[14] Alvaro Gomariz, Tiziano Portenier, César Nombela-Arrieta, and Orcun Goksel. Probabilistic spatial analysis in quantitative microscopy with uncertainty-aware cell detection using deep bayesian regression of density maps. arXiv preprint arXiv:2102.11865, 2021. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 547, + 205 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 547, + 248 + ], + "type": "text", + "content": "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 547, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 547, + 303 + ], + "type": "text", + "content": "[17] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In International conference on machine learning, pages 1989-1998. Pmlr, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 305, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 547, + 348 + ], + "type": "text", + "content": "[18] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 547, + 392 + ], + "type": "text", + "content": "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 547, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 547, + 458 + ], + "type": "text", + "content": "[20] Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Herpel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospective multicenter study. PLoS medicine, 16(1):e1002730, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 547, + 502 + ], + "type": "text", + "content": "[21] Jakob Nikolas Kather, Frank Gerrit Zollner, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Cleo-Aron Weis. Collection of textures in colorectal cancer histology, May 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 504, + 547, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 547, + 535 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 547, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 547, + 581 + ], + "type": "text", + "content": "[23] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4544-4553, 2020. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 547, + 635 + ], + "type": "text", + "content": "[24] Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, R Venkatesh Babu, et al. Towards inheritable models for open-set domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12376-12385, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 547, + 689 + ], + "type": "text", + "content": "[25] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9641-9650, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[26] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20349" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 288, + 105 + ], + "type": "text", + "content": "unsupervised domain adaptation. In International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 149 + ], + "type": "text", + "content": "[27] Quande Liu, Qi Dou, Lequan Yu, and Pheng Ann Heng. Msnet: Multi-site network for improving prostate segmentation with heterogeneous mri data. IEEE Transactions on Medical Imaging, 2020. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 216 + ], + "type": "text", + "content": "[28] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "text", + "content": "[29] Alexander Lyzhov, Yuliya Molchanova, Arsenii Ashukha, Dmitry Molchanov, and Dmitry Vetrov. Greedy policy search: A simple baseline for learnable test-time augmentation. In Conference on Uncertainty in Artificial Intelligence, pages 1308-1317. PMLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 288, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 325 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 325 + ], + "type": "text", + "content": "[30] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[31] Alexandru Niculescu-Mizil and Rich Caruana. Predicting good probabilities with supervised learning. In Proceedings of the 22nd international conference on Machine learning, pages 625-632, 2005. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "type": "text", + "content": "[32] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 459 + ], + "type": "text", + "content": "[33] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 514 + ], + "type": "text", + "content": "[34] Ferran Prados, John Ashburner, Claudia Blaiotta, Tom Brosch, Julio Carballido-Gamio, Manuel Jorge Cardoso, Benjamin N Conrad, Esha Datta, Gergely David, Benjamin De Leener, et al. Spinal cord grey matter segmentation challenge. Neuroimage, 152:312-329, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[35] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In International Joint Conference on Artificial Intelligence, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "type": "text", + "content": "[36] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, European Conference on Computer Vision (ECCV), volume 9906 of LNCS, pages 102-118. Springer International Publishing, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "text", + "content": "[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 150 + ], + "type": "text", + "content": "[39] Yongyi Su, Xun Xu, and Kui Jia. Revisiting realistic test-time training: Sequential inference and adaptation by anchored clustering. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 205 + ], + "type": "text", + "content": "[40] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 251 + ], + "type": "text", + "content": "[41] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 251, + 547, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 547, + 305 + ], + "type": "text", + "content": "[42] Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran, and Behzad Bozortbar. OptTTA: Learnable test-time augmentation for source-free medical image segmentation under domain shift. In Medical Imaging with Deep Learning, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 307, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 307, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 545, + 338 + ], + "type": "text", + "content": "[43] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 340, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 340, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 340, + 545, + 373 + ], + "type": "text", + "content": "[44] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 374, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 545, + 418 + ], + "type": "text", + "content": "[45] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 418, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 473 + ], + "type": "text", + "content": "[46] Guotai Wang, Wenqi Li, Michael Aertsen, Jan Deprest, Sébastien Ourselin, and Tom Vercauteren. Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks. Neurocomputing, 338:34-45, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "text", + "content": "[47] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 519, + 545, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 562 + ], + "type": "text", + "content": "[48] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Casting a bait for offline and online source-free domain adaptation. arXiv preprint arXiv:2010.12427, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 563, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 545, + 608 + ], + "type": "text", + "content": "[49] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8978-8987, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 609, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 651 + ], + "type": "text", + "content": "[50] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation. arXiv preprint arXiv:2205.04183, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 653, + 547, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 653, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 653, + 547, + 707 + ], + "type": "text", + "content": "[51] Yifan Zhang, Ying Wei, Qingyao Wu, Peilin Zhao, Shuaicheng Niu, Junzhou Huang, and Mingkui Tan. Collaborative unsupervised domain adaptation for medical image diagnosis. IEEE Transactions on Image Processing, 29:7834-7844, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20350" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_content_list.json b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2b09ce32ff6a5ff47cf3c483659e363d1df465e --- /dev/null +++ b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "text", + "text": "Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models", + "text_level": 1, + "bbox": [ + 106, + 130, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yushi Yao*", + "bbox": [ + 117, + 203, + 209, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Waymo", + "bbox": [ + 132, + 222, + 196, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yushiy@waymo.com", + "bbox": [ + 91, + 242, + 235, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chang Ye*", + "bbox": [ + 307, + 203, + 395, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google", + "bbox": [ + 321, + 222, + 383, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yechang@google.com", + "bbox": [ + 272, + 242, + 433, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junfeng He†", + "bbox": [ + 509, + 203, + 607, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google", + "bbox": [ + 529, + 222, + 589, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "junfenghe@google.com", + "bbox": [ + 470, + 242, + 648, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gamaleldin F. Elsayed†", + "bbox": [ + 687, + 203, + 874, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google", + "bbox": [ + 751, + 222, + 812, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gamaleldin@google.com", + "bbox": [ + 684, + 242, + 872, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 312, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human spatial attention conveys information about the regions of visual scenes that are important for performing visual tasks. Prior work has shown that the information about human attention can be leveraged to benefit various supervised vision tasks. Might providing this weak form of supervision be useful for self-supervised representation learning? Addressing this question requires collecting large datasets with human attention labels. Yet, collecting such large scale data is very expensive. To address this challenge, we construct an auxiliary teacher model to predict human attention, trained on a relatively small labeled dataset. This teacher model allows us to generate image (pseudo) attention labels for ImageNet. We then train a model with a primary contrastive objective; to this standard configuration, we add a simple output head trained to predict the attention map for each image, guided by the pseudo labels from teacher model. We measure the quality of learned representations by evaluating classification performance from the frozen learned embeddings as well as performance on image retrieval tasks (see supplementary material). We find that the spatial-attention maps predicted from the contrastive model trained with teacher guidance aligns better with human attention compared to vanilla contrastive models. Moreover, we find that our approach improves classification accuracy and robustness of the contrastive models on ImageNet and ImageNet-C. Further, we find that model representations become more useful for image retrieval task as measured by precision-recall performance on ImageNet, ImageNet-C, CIFAR10, and CIFAR10-C datasets.", + "bbox": [ + 73, + 372, + 473, + 828 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/322434ca0869a5340264ba1200a146318688434775e26bb4924014b017ec42fb.jpg", + "image_caption": [ + "Figure 1. Illustration. A teacher model is trained to predict human spatial-attention from a small dataset. Then the model is used to provide attention labels for larger dataset, which are used as additional targets for contrastive models." + ], + "image_footnote": [], + "bbox": [ + 517, + 290, + 874, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 590, + 632, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models have made significant progress and obtained notable success on various vision tasks. Despite these promising results, humans continue to perform better than deep learning models in many applications. A notable reason is that deep learning models have a tendency to learn \"short-cuts\", i.e., giving significance to physically meaningless patterns or exploiting features which are predictive in some settings, but not causal [20]. Examples include focusing on less significant features such as background and textures [13]. These models yield representations that are less generalizable and lead to models that are highly sensitive to small pixel modulations [42].", + "bbox": [ + 496, + 617, + 892, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human vision on the other hand is known to be much more robust and generalizable. One major difference between human and machine vision is that humans tend to", + "bbox": [ + 498, + 797, + 893, + 843 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal technical contribution.", + "bbox": [ + 500, + 851, + 663, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal leadership and advising contribution", + "bbox": [ + 500, + 863, + 733, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Correspondence to:", + "bbox": [ + 500, + 876, + 606, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "junfenghe@google.com & gamaleldin@google.com", + "bbox": [ + 500, + 888, + 841, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23282", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "focus on specific regions in visual scene [45]. These locations often reflect regions salient or useful to perform a specific vision task. Machines, instead, initially place equal significance to all regions. A natural question is: will it be beneficial if machine vision models is guided by human spatial attention?", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human spatial attention has been shown to benefit computer vision models in supervised tasks, such as classification [32]. Yet, it is still a question whether adding a form of weak supervision in the form of human spatial attention could similarly benefit self-supervised models that are trained end-to-end. Self-supervised models typically need a large amount of data to yield good representations. To test if training weakly supervised models with human spatial attention cues, we will need to collect a large volume of human spatial attention labels, which is a very expensive process that requires either using trackers to record eye movements [5,43,52] or asking humans to highlight regions that they attend to [25, 27]. This process is prohibitively tedious and costly for datasets with millions of examples.", + "bbox": [ + 75, + 181, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we test the hypothesis that a weak supervision in the form of human spatial attention is beneficial for representation learning for models trained with a contrastive objective. Inspired by knowledge distillation and self-training using teacher models [47, 49], we address the challenge of obtaining spatial attention labels on large scale image datasets by using machine pseudo-labeling. We train a teacher model on a set of limited ground truth human spatial attention labels, and use this teacher model to generate spatial attention pseudo-labels for the large ImageNet benchmark. We are then able to utilize the generated spatial attention maps in the contrastive models, and discover that this approach yields representations that are highly predictive of human spatial attention. Further, we find that the learned representations are better as measured by higher accuracy and robustness on classification downstream tasks, and higher precision and recall on image retrieval tasks. Interestingly, we find that the gains from using teacher models to provide pseudo labels are larger than using the limited ground truth human labels directly when training contrastive models, and the gains are larger for contrastive models than when applying same method to supervised models. In summary, our contributions are as follows:", + "bbox": [ + 75, + 393, + 470, + 741 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We create a dataset with spatial attention maps for the ImageNet [37] benchmark by first training a teacher model to predict human spatial attention labels from Silicon dataset [25] and then use the model to label ImageNet examples", + "- We use spatial-attention labels from the teacher model as an additional prediction target to models trained" + ], + "bbox": [ + 93, + 750, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with contrastive objective.", + "bbox": [ + 532, + 90, + 707, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We find that the proposed method can learn better representation, leading to better accuracy and robustness for downstream classification tasks (on ImageNet and ImageNet-C), and better performance on retrieval tasks (on ImageNet, ImageNet-C, CIFAR-10, and CIFAR10-C).", + "bbox": [ + 516, + 118, + 890, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 224, + 637, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning: Contrastive learning has gained popularity in the past few years for self-supervised and semi-supervised representation learning. In general, contrastive learning aims to learn similar representations for similar data pairs and different representations for different pairs. SimCLR [6] utilized MLP projection heads and strong data augmentation for constructing similar pairs and have demonstrated great gains in image classification downstream tasks. To form the contrastive loss for each minibatch with N examples, the similar data pairs are constructed from two augmentations of the same image and different pairs from the other images within a batch, and then computing the NT-Xent loss. A different formulation is used in [50] by encouraging the empirical cross correlation of the representations of two versions of augmented mini-batch to be close to identity. In [17], it is further proposed to build large dictionaries for self-supervised learning (MOCO), and moreover in [7], better results are achieved on image classification and object detection tasks when combining advances from SimCLR and MOCO. Follow up work by [14] managed to obtain good performance without the need for dissimilar pairs by encouraging the representation of similar pairs across two versions of the network (trained network and exponential moving average version) to be similar. Further, in [8] it is shown that simple Siamese networks can still learn good representations without the need for dissimilar pairs, large batches or momentum encoders, etc.", + "bbox": [ + 496, + 250, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human spatial attention: Human visual system has developed an attention mechanism that focuses on regions in the visual space that are of interest or highly informative to the vision task [12, 48]. Eye trackers are often used to collect human spatial attention [5, 43, 52]. Many gaze data sets [2] have been collected with these eye trackers. Besides eye trackers, human spatial attention data can also be collected via mouse tracking [25, 27], e.g., users see a blurry version of an image, then click on regions they want to see more clearly, mimicking human's peripheral vision based on neurophysiological studies [19, 27]. Both eye tracking and mouse tracking methods are very expensive, which limit the number of examples in those datasets. Due to the relatively lower cost of mouse tracking, one can often generate relatively larger attention data from this method than", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Trained teacher model is available at: https://github.com/google-research/google-research/tree/master/humanattention/", + "bbox": [ + 76, + 875, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23283", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eye tracking. For example, Silicon [25] dataset is one of the largest spatial attention datasets, contains around 20K images, each with attention labels from 50-60 participants, via a mouse tracking system, under free viewing setting. Yet, this data is still orders of magnitude smaller than those needed to train contrastive representation learning models.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial attention/saliency prediction models aim at predicting which areas in an image will be salient to human attention and attract eye fixation, usually with collected gaze/attention data as ground-truth. Early works in saliency prediction usually define saliency through a set of hand crafted features such as color difference, contrast, intensities, etc. [21, 28]. Recent works [1, 19, 24, 30, 31, 36] leverage the power of deep neural networks and are often trained/fine-tuned on large scale gaze data sets like Silicon [2, 25].", + "bbox": [ + 75, + 185, + 470, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial attention of computer vision models: Spatial attention in neural networks can be mainly categorized into post-hoc attention like class activation map (CAM) ([51]), and trainable attention (e.g., [15, 23, 46]). Post-hoc spatial attention methods have been proposed to estimate regions in the image that are important or give rise to model decisions, often for model interpretation. In supervised settings where classification labels are known, the simplest and most direct method is class activation map (CAM) [51]. CAM uses class labels to extract the feature map that is most informative about the true class of an image. Grad-CAM [39] generalizes the CAM to apply to any model with any downstream task. [38] proposed to use Grad-CAM to design augmentation policies in self-supervised learning to tackle weak performance in complex scene images with many objects. ContraCAM [35] applies Grad-CAM assuming downstream task of contrastive learning, thus allowing computing spatial attention maps with no class label supervision. In [35], it is proposed to utilize the spatial attention information learned from ContraCAM to design data augmentation strategies to discourage contextual and background biases in a scene. Unlike [38] and [35] that uses spatial attention to design augmentation policies, here we focus on an end-to-end framework to predict spatial attention targets instead.", + "bbox": [ + 75, + 339, + 470, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, there are also several papers [10, 11, 16, 33, 41, 44]) exploring the similarity/difference of model spatial attention vs human spatial attention, for the task of VQA (visual question answering) [10], object detection [11], reinforcement learning [16], etc. Among existing works, [32] is the most relevant to ours, since it conducted experiments to use human spatial attention to supervise model spatial attention, for three tasks (salient object segmentation, video action recognition and fine-grained image classification) and demonstrated that human spatial attention is beneficial. However, it still remains a question whether such benefits could be extended to contrastive representation learning. In", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "terms of strategy to utilize spatial attention labels, in [32], attention labels are used as spatial weighting, while we designed auxiliary task to predict spatial attention labels.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Teacher model pseudo-labeling: Previous work on knowledge distillation and machine self training has demonstrated that machine teaching machines approaches may address the challenge of labeling large datasets. In image classification, [49] demonstrated that training a model to classify images then use that model to provide pseudolabels improved classification performance. Related idea is applied in [47] for language models, which showed that the knowledge learned by language models pretrained on text corpus could be distilled to generate new datasets for common reasoning, and that training common reasoning models on this new data largely improve performance. Inspired by these successes, we train a teacher model on smaller human attention data and use this model to generate new spatial attention pseudo labels for ImageNet benchmark (see Figure 1 and Figure 2).", + "bbox": [ + 496, + 136, + 892, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 387, + 598, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The Contrastive Learning Framework", + "text_level": 1, + "bbox": [ + 500, + 411, + 830, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning is one of the most popular self-supervised representation learning methods. It learns an embedding space so that similar data (positive) pairs are mapped to be close in the embedding space and different (negative) pairs are mapped to be far away. In practice, positive pairs are often generated by applying data augmentation to one image like adding noises, cropping, etc and negative pairs are different examples in the mini-batch. Among contrastive learning methods, SimCLR framework [6] has shown solid performance, which we choose here as our main method. For each batch, the images are augmented in two different ways. Then we feed them into ResNet feature extractor backbone and compute NT-Xent loss to minimize the difference between augmentations of the same images and maximize the differences between different images. NT-Xent loss can be computed as follows [6]:", + "bbox": [ + 496, + 435, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i, j} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {j}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {2 N} 1 _ {k \\neq i} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {k}\\right) / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 686, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{i}$ is the embedding of ith example, and $\\operatorname{sim}(\\cdot ,\\cdot)$ is the cosine similarity between two embeddings.", + "bbox": [ + 496, + 731, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. ImageNet-Attn: spatial attention maps for ImageNet generated by a teacher model", + "text_level": 1, + "bbox": [ + 500, + 771, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train contrastive learning models, we need large datasets with millions of examples. Thus, to address the question whether human attention is beneficial for contrastive models, we will need large human spatial attention labeled dataset. However, there is no spatial attention dataset available with that size on typical large image", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23284", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/93fa50206b0d3cbe786cbf9ad40bcec24b5a6cbb038725bcb734ba052fc9a579.jpg", + "image_caption": [ + "a", + "Teacher model", + "b", + "C" + ], + "image_footnote": [], + "bbox": [ + 178, + 114, + 825, + 246 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ea18cfcc421aaec73cd1885287a1fb4d55a91c07e8270484dcf100a13a57fa7.jpg", + "image_caption": [ + "Sallcon dataset image" + ], + "image_footnote": [], + "bbox": [ + 161, + 291, + 274, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a3768fe338159b9ab323fb7fe4c83cbd97773ba9f8b9a024ed1bf2e6d7cffa08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 380, + 274, + 467 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b0c07aaea506276e13a400e8b827692afdae1d147790c8df966c239e5410b92.jpg", + "image_caption": [ + "Figure 2. (a) The architecture of teacher model to predict attention. The teacher model was trained from random initialization to predict ground truth (G.T.) human spatial attention labels. (b) Examples from Salicon dataset, with G.T. human attention heatmaps, and teacher model predictions. (c) Examples of spatial-attention pseudo-labels generated by the teacher model for ImageNet dataset (ImageNet-Attn)." + ], + "image_footnote": [], + "bbox": [ + 161, + 467, + 274, + 556 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1755ab91f9f4189d4abeb98133b473ee9329c5f854e24ff4137cd0426bd3cc5f.jpg", + "image_caption": [ + "ground truth human spatial attention" + ], + "image_footnote": [], + "bbox": [ + 294, + 291, + 406, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f3181742b530666673515591cb612661a6d81ae096a0b81856c8e1889370377.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 380, + 406, + 465 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8c95a5c3463ab36352c93dbf1d1a7eeb299c849967865c5a2914ea986749c88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 468, + 406, + 556 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/89d78b025c249a28fae17397f08ca1b33aa52c7b1632f052736d60eac8fdab9c.jpg", + "image_caption": [ + "teacher model predictions" + ], + "image_footnote": [], + "bbox": [ + 424, + 291, + 537, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/574387cf417c34711f8395c2b0906971b0b95162958167806d02b86f105810dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 380, + 537, + 465 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/71cafe5f403f4e8fd4d126bcef999bb009c1d4417a991953e2004132f5c78329.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 468, + 537, + 556 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e0574f871272e15f4f1fd1a2d12a754c2afe77495dbe90ec41c3172de78c207d.jpg", + "image_caption": [ + "ImageNet dataset image" + ], + "image_footnote": [], + "bbox": [ + 576, + 292, + 689, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/376c356029e8c7641a85a9d170af3ed4b7726fa1f828194ba10f67a263af4f35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 380, + 689, + 467 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66963753b7c8fb7087a4029952f3b08637b4dc57463f9eddc3fc8ddaf91f8380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 469, + 689, + 556 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8907546e08968dbf24fe5c0b0f03b1acb32454e82b1f8c7d686f3e42a3ee3bd.jpg", + "image_caption": [ + "spatial-attention labels from the teacher model" + ], + "image_footnote": [], + "bbox": [ + 717, + 292, + 831, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4d20847130eeb8ac6624910389ad3c7e00051a253297e68fdf56d3f33f2478a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 380, + 830, + 467 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9d2174013c7ecfccbb3d60ba276f079aff4be0b4285359ef7f8fe2541a317fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 469, + 830, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "benchmark such as ImageNet. The largest attention data set available is Silicon [25], consisting of twenty thousand examples only, orders of magnitude smaller than ImageNet, which has around 14 million examples.", + "bbox": [ + 75, + 619, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate attention labels for ImageNet images, we train a teacher model on Silicon attention data set to predict human attention ground truth labels (see Figure 2b). Then we use this teacher model to create a new dataset (ImageNet-Attn) with spatial attention maps for ImageNet (see Figure 2c for some examples). The teacher model architecture is illustrated in Figure 2a. Specifically, we used MobileNet-V3-small as the backbone, and embeddings from 4 layers (conv 2,4,6,8) are extracted. On each embedding, we applied two conv layers (the first conv layer with $3 \\times 3$ kernel, number of channels matching the input embedding, max pooling of 3 and relu; the second conv layer with $1 \\times 1$ kernel, 1 channel and relu). We bilinearly resized the output to the same resolution of input. The", + "bbox": [ + 75, + 689, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "result of the 4 branches are then summed to yield 1 feature, followed up by a sigmoid function to obtain spatial-attention map. Note that using multiple intermediate (both early and late) layers is one of the key ingredient to make the spatial attention teacher model successful. This strategy mimics human visual attention, which is known to be affected by both low-level characteristics such as color, intensity, texture, as well as high level characteristics like shape and object, etc. Essentially, our teacher model has a similar architecture as in [36], but is further simplified with more efficient backbone, less channels/layers, etc, so that it can be trained from scratch on Silicon data with randomly initialized backbone to avoid any leak of class label information, while most existing attention prediction models [1,36] needs to finetune with a pretrained classification backbone network.", + "bbox": [ + 496, + 619, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23285", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training contrastive models with spatial attention maps", + "text_level": 1, + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Attention Branch The overall training framework is shown in Figure 1. This spatial attention prediction framework consists of two branches: the contrastive branch and the spatial attention branch. The contrastive branch is the same as the original SimCLR method [6], which applies augmentations to image $x$ to get different variants $x_{i}$ and $x_{j}$ , and learns the representation $h_{i}$ and $h_{j}$ via a feature extractor backbone network (e.g., ResNet), then use a projection head to map $h_{i} / h_{j}$ to $z_{i} / z_{j}$ , where the contrastive loss is applied.", + "bbox": [ + 75, + 128, + 467, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the spatial-attention branch, it takes as an input not only the final embedding $h$ , but also early intermediate layer embeddings (following the same reason of including both high level and low level visual cues as in teacher model), and predicts an spatial attention heatmap $m_p$ . More specifically, we apply a global average pooling on the output of the last three blocks of the ResNet backbone. Then, we select the max channel for each of three block output (after average pooling), and resize with bilinear interpolation to the image resolution, which has some resemblance to CAM approaches [51]. Finally we stack the representations together, pass them into a linear readout layer (with a bias term), and use the output as our final spatial attention prediction $m_p$ . We keep the spatial attention head as simple as possible so that the guide on attention head output can be back propagated to representation more directly.", + "bbox": [ + 75, + 280, + 467, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the ImageNet-Attn data generated by the teacher model as target for each ImageNet example, denoted as $m_t$ , we can then train the network spatial attention output $m_p$ to be close to $m_t$ . We hypothesize that this method regularizes the training of the feature extractor backbone rather than explicitly enforce the network to generate masked representations that match the spatial attention maps. Note that for attention branch, there is no augmentation applied to each image $x$ , since human attention is not invariant to transformation (e.g., a human looking at a cropped image may attend to different regions compared to a consistent crop of human attention map of the original image).", + "bbox": [ + 75, + 521, + 467, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function The loss function $L$ consists of two terms:", + "bbox": [ + 76, + 702, + 467, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {c o n t r a}} + L _ {\\text {a t t n}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 733, + 467, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$L_{\\text{contra}}$ is the contrastive loss, or more specifically $L_{\\text{contra}} = \\sum_{i,j} l_{i,j}$ and $l_{i,j}$ is defined in Equation (1), the same as in [6]. $L_{\\text{attr}}$ is the attention regularization loss, and more specifically", + "bbox": [ + 76, + 752, + 467, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a t t n} = \\sum_ {i} \\left(\\lambda K L D \\left(m _ {i} ^ {p}, m _ {i} ^ {t}\\right) - \\beta N S S \\left(m _ {i} ^ {p}, p _ {i} ^ {t}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 818, + 467, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\beta$ are two weighting parameters ( $\\lambda = 1.0$ and $\\beta = 0.1$ in our work). $m_{i}^{p}$ is the predicted 2D spatial-attention map from the attention head of our proposed", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a9b10a12a9df8bb7e0fabaccd40d072c535616b64e15a72095a869a368154c79.jpg", + "table_caption": [ + "Table 1. ImageNet Top-1 classification accuracy for different models (mean $\\pm$ SE for 3 seeds, except for * which means best result from all runs)." + ], + "table_footnote": [], + "table_body": "
ModelAccuracy (%)
Contrastive67.61 ± 0.04
Contrastive attn. teacher68.23 ± 0.08
Contrastive attn. co-train66.35 ± 0.12
Contrastive attn. with explicit attention block [32]61.15*
Contrastive attn. with self attention mask [32]49.65*
Contrastive attn. with ContraCAM [35] as attention67.71*
Supervised75.91 ± 0.10
Supervised attn. teacher76.02 ± 0.04
Supervised (ResNet-18)69.17 ± 0.07
Supervised (ResNet-18) attn. teacher69.30 ± 0.04
", + "bbox": [ + 504, + 143, + 888, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model for $i$ -th example (original image, not augmented ones), $m_i^t$ is the pseudo spatial-attention target map predicted by the teacher model for $i$ -th example, and $KLD()$ is the KL divergence1. Besides KLD loss, we also use $NSS()$ , the Normalized Scanpath Saliency loss [4]. Those two losses are typically used for human attention predictions [3,9]. KLD is often used to match the spatial-attention target distribution, and NSS is typically added on top as it is generally observed to help generate attention maps that are in perceptual agreement with human judgements (see [3,34]). For $NSS$ loss, the larger the better, so there is a negative sign before it. $NSS$ loss needs gaze/attention points instead of heatmap as ground-truth, so we extract pseudo gaze/attention points $p_i^t$ from the attention heatmap $m_i^t$ . To obtain $p_i^t$ , we first extract the point with highest value in current spatial-attention map $m_i^t$ , then generate a new spatial-attention map by subtracting a Gaussian blur around the extracted point from the current attention map. The process is repeated with the new attention map until the maximal value of the attention map is smaller than a threshold (more specifically, $0.2 * \\max\\_ sal$ , where $\\max\\_ sal$ is the maximal value in $m_i^t$ ).2", + "bbox": [ + 496, + 325, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 672, + 632, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 696, + 712, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, in Eq.1, we choose $\\tau = 0.1$ and $N = 2048$ . In Eq.2, we used equal weighting of 1.0 for the SimCLR and attention losses.", + "bbox": [ + 500, + 720, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The weight for KLD and NSS losses are 1 and 0.1, re", + "bbox": [ + 517, + 766, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "${}^{1}{m}_{i}^{p}$ is normalized when computing KLD, i.e.,divided by its pixel value sum,so that it becomes a distribution with all pixel value sum equal to 1. Similarly for ${m}_{i}^{t}$ .", + "bbox": [ + 500, + 790, + 890, + 827 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Our process of extracting gaze points from heatmaps is to inverse the process of generating heatmaps from gaze points. Due to the uncertainty about the gaze locations, researchers typically apply Gaussian blur for each gaze point, then sum all Gaussians to generate the heatmap (see [3, 26]). We follow the inverse of this process to obtain gaze points from heatmap, similarly as done in [22].", + "bbox": [ + 500, + 827, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23286", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/32d7e5374b151f79b5cb8a57482cedc8232264f0992237cd7e8a14ad84159635.jpg", + "image_caption": [ + "a", + "image" + ], + "image_footnote": [], + "bbox": [ + 124, + 143, + 253, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2eecca999a2c947038b270bef71548eb6649dea9201ac5ae400bb29351fe1285.jpg", + "image_caption": [ + "ground truth human spatial attention" + ], + "image_footnote": [], + "bbox": [ + 274, + 142, + 403, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/846b7e8c3341bff98bef74217b447a2b18dbaa29226df152eace79dcd4cf4f5d.jpg", + "image_caption": [ + "teacher model spatial attention" + ], + "image_footnote": [], + "bbox": [ + 426, + 143, + 555, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/539fbb18ff73cf7649d2a6653796aee1e0b4187cf05d2c7128a02397947ec962.jpg", + "image_caption": [ + "spatial attention guided contrastive model" + ], + "image_footnote": [], + "bbox": [ + 578, + 142, + 707, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eaf653822539f8748e0132a7dbe5e18f56f00e4b46a529919778f1287b405ded.jpg", + "image_caption": [ + "baseline trastive model" + ], + "image_footnote": [], + "bbox": [ + 733, + 142, + 862, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a945762d74bebc49c3a6083f0a783e9040a178ad6e3ef8170cb3d7f826e23f03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 244, + 251, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/45d66b37c81c7ea81df24645f90460ad72fb3484b062276c1688f843f84515ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 345, + 251, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/642d6b9bdc2ab1f0f8f178624888629995f74415b2b5995cf545f63ed0acf79c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 244, + 401, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/736137c06046f970352bf1a27a8aeec06985b2ca244f1d79af159d04e98105e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 345, + 401, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e69720b1fcb4eb30c868b50bfacfe9a8090aed7f1b35fe46931c49b39c4087ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 244, + 553, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d082c4709c76c8911510e6ae8101c69d06a1b24a93c94a76c433f9fdf77c82d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 345, + 552, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f9463b54bfb9044d1eab9c31c10c71d034fb0cd5baab8747b43241ba864a9bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 244, + 705, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1be15bb1b92d287fe25227d4a63aeeb56c5377dbccf79ffe32f0e94756e900bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 345, + 705, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7b688f628f4f15ccfaf1b70a84c4b737b12d46893e61d4698ed198b008255a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 244, + 862, + 345 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c85bff58326ab2ea0aa7f50d5c220a869fbc3fa580de36be75793ade749e5826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 345, + 862, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c3ddacaa39d16ca0e6be2230e0bc9b0daf621bf30c33814a8bc83f594917518.jpg", + "image_caption": [ + "b" + ], + "image_footnote": [], + "bbox": [ + 107, + 455, + 478, + 623 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b3c64989230165649d08fbd2ea5c4944425e7a4354d50107fa8d2d7ecf74867.jpg", + "image_caption": [ + "correlation coeff. (model vs human)", + "Figure 3. (a) Examples comparing spatial attention maps predicted by different models vs ground truth human attention data on OSIE dataset [43]. (b) Distribution of correlation coefficients between model predicted attention maps vs ground truth human attention maps on Silicon [25] validation set (left) and OSIE dataset (right) [43]. Three models are shown: (1) Teacher model trained on Silicon [25] training set. (2) Baseline SimCLR model (3) The proposed spatial attention guided SimCLR model trained with pseudo-attention labels provided by the teacher model. Please note that both baseline and our model are trained to predict target attention labels, with main difference that the baseline model has a stop gradient placed between the prediction layer and network features to prevent spatial-attention targets to inform network features. Also note that our teacher model is trained on Silicon where the attention is collected by mouse tracking, while the results in (a) and (c) here are on OSIE data, where the attention are collected by eye trackers. The results demonstrated both the teacher model and the attention head in proposed model has a good generalization capacity for new kinds of attention labels." + ], + "image_footnote": [], + "bbox": [ + 496, + 455, + 870, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spectively. We obtained those weights by hyperparameter search from 0.1-5 for KLD and 0.01-1 for NSS, based on ImageNet accuracy of a separate dev set (80%-20% split on training set for parameter search with 20% as dev set).", + "bbox": [ + 75, + 786, + 468, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Spatial attention guided models are highly predictive of human attention", + "text_level": 1, + "bbox": [ + 500, + 786, + 890, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we explore whether the use of auxiliary teacher model to provide spatial attention pseudo labels on ImageNet better aligns contrastive model's attention with human attention. We define aligning model spatial attention here as the ability to predict spatial attention mask from", + "bbox": [ + 498, + 825, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23287", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the model backbone features by a simple readout layer discussed in Section 3. We trained two ResNet-50 backbones using the SimCLR objective from [6]. We added additional spatial-attention losses as discussed in Section 3. For the baseline model, we placed a stop gradient operation between the backbone features and the attention projection head to prevent attention information from leaking to the backbone features, whereas for the attention guided model, we allowed the learned attention gradients to flow back to the backbone.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated the degree the predicted attention maps is aligned with human attention by performing correlation analysis, which is typically used to compare human and model spatial attention. We measured Pearson's correlation coefficient between the model predicted attention and ground truth human spatial-attention maps from Silicon [25] validation set (Note that the teacher model is trained with Silicon training set). We find that both teacher model and spatial attention guided contrastive model are highly correlated with human attention (Figure 3b), while the baseline contrastive model is much less correlated.", + "bbox": [ + 75, + 244, + 472, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test whether this alignment with human attention is general, beyond just Salicon data, we obtained another attention dataset on OSIE images [43] (Figure 3a). This data more faithfully reflects human attention as it collected from (mobile) eye tracker [43]. We find that the baseline model is weakly positively correlated with human attention (ttest: $\\rho = 0.07$ $p < 0.001$ ) suggesting that the contrastive loss produces features that are predictive of human attention to some extent. Yet, the correlation was generally close to 0 and explains only $0.5\\%$ of data variance. The spatial attention guided model has a much stronger correlation to human attention (ttest: $\\rho = 0.48$ $p < 0.001$ ) than the baseline model (Two samples ttest: $p < 0.001$ ), and thus more faithfully reflecting human visual attention (See Fig 3a for qualitative examples and Fig 3b bottom for quantitative analysis). Further, the correlation with human attention for the teacher model and the attention guided contrastive model were quite similar (Two samples ttest: $p = 0.7$ ).", + "bbox": [ + 75, + 411, + 472, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Spatial attention guided models are more accurate than baselines", + "text_level": 1, + "bbox": [ + 76, + 699, + 468, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the quality of the representations learned by spatial attention guidance framework using the typical contrastive learning evaluation criteria: fitting an ImageNet [37] linear classifier on top of the frozen representation (in practice we place stop gradient at the end of the backbone and train the classifier concurrently while training the backbone). We compute Top 1 accuracy on ImageNet validation set and compare the results with baselines. As shown in Table 1, we observe around $0.6\\%$ accuracy gain on ImageNet compared to vanilla SimCLR. We further explore an alternative way of incorporating human attention data.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rather than using pseudo attention labels on ImageNet from the teacher model, we add Silicon data to the training data, and directly predict attention ground truth labels for Silicon data with spatial attention head (Contrastive attn. co-train in Table 1). More specifically, we pass 2048 ImageNet images and N=2048 Silicon images into our backbone (N is obtained by hyperparameter search 512-2048 based on evaluation on the separate dev set) to compute backbone features. Then, we compute the SimCLR loss from ImageNet features and use Silicon images features to predict spatial attention maps, and we use Silicon attention ground truth labels to compute the KLD and NSS losses. Interestingly, we find this method to lead to worse performance compared to using the teacher model generated spatial-attention labels, which gives evidence that the teacher model is generalizing its knowledge about human attention data beyond the limited Silicon training data.", + "bbox": [ + 496, + 90, + 893, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One important question is whether the classic trainable attention method applies to contrastive learning. For example, in [32], an attention block/map is used to adjust the weights of the embedding, where the attention block can be self learned implicitly, or guided by human attention ground-truth. It shows both implicit self-learned attention block and attention block guided by human attention can improve the accuracy of fine-grained classification (Table VII in [32]). We adopted the approach described in their paper: add an attention block after the representation embedding, and use the attention block's output to mask the representation embedding. We apply it to the contrastive models, and the results are shown in Table 1 as \"explicit attention mask\" (the attention block is guided by human attention) and \"self attention mask\" (self attention). As observed, both yield poor performance, since the contrastive attention with self-attention mask method learns an attention mask from the contrastive loss rather than to predict the teacher targets, the model may be getting weaker supervision.", + "bbox": [ + 496, + 349, + 893, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another important question is whether our current spatial attention targets can be replaced by other spatial attention approach like ContraCAM [35]. In our method, we predicted the spatial attention with a linear layer. Whereas in ContraCAM a contrastive loss is used to construct the attention prediction. In both cases, we trained the predictions to match the target attention maps obtained from the teacher model. The result is shown in Table 1 as \"Contrastive with ContraCAM as attention\". However, we don't see much accuracy improvements with ContraCAM as the model attention, showing that the linear simple prediction leads to better performance.", + "bbox": [ + 496, + 656, + 893, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate whether the spatial attention guidance framework benefits supervised models in the same way as contrastive models, we applied the same approach for supervised models. Supervised models similarly benefit from", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23288", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this framework, yet the gain is limited compared to the contrastive models perhaps due to the higher accuracy the supervised models achieves compared to the contrastive models. To control for accuracy, we used a supervised model trained with smaller backbone (ResNet-18), which gives comparable accuracy to the bigger ResNet-50 backbone that is trained with contrastive objective. Even when controlling the mismatch in accuracy, the supervised model gain from providing spatial-attention supervision is limited compared to contrastive models (Table 1).", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Spatial attention guided models are more robust than baselines", + "text_level": 1, + "bbox": [ + 76, + 253, + 470, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human vision is very robust to noise or small pixel modulations, compared to computer vision models. Here, we hypothesize that models that aligns better with human attention may learn more robust representations. We test this hypothesis using ImageNet-C dataset [18]. We take the representations learned from the proposed model and train a linear classifier on ImageNet training data. We then evaluate classification performance on ImageNet-C at various corruption/noise types. Table Supp.1 in the supplementary compares the classification performance on ImageNet-C for baseline contrastive model and the proposed models trained with spatial-attention teacher guidance, average over the 5 different corruption/noise magnitudes in ImageNet-C. We find that contrastive models trained with teacher guidance outperforms baseline consistently, suggesting that the representation learned by spatial-attention guided model is indeed more robust.", + "bbox": [ + 75, + 292, + 472, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Spatial attention guided models generate better representation for retrieval", + "text_level": 1, + "bbox": [ + 76, + 559, + 468, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides classification, we tested the quality of the representations for another downstream task: image retrieval. We use the model to extract representation for ImageNet validation set, and use the representations to run image retrieval. 5000 randomly chosen validation images are used as query, while the rest validation images are database images. For each query image, top k retrieval results are returned by sorting the cosine distance between representations of database images and the query. The retrieval accuracy is measured with the standard precision recall curve [40], where precision and recall are computed by checking whether results have the same class label as the query. Mean precision and recall is obtained as the average across queries. We choose different k values, and draw the mean precision-recall curve. Besides ImageNet (\"clean\" in Fig Supp.1a in the supplementary), we also use ImageNet-C (\"noise level 1-5\"). Results from \"fog\" corruption is shown in In Figure Supp.1a in the supplementary., while results for other corruptions can be found in the supplementary too. The representation extracted with", + "bbox": [ + 75, + 598, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the Contrastive attn teacher model outperforms the baseline Constractive model in this retrieval task, on both ImageNet images, and ImageNet-C images for most corruption types and levels.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we evaluate transfer to other dataset for the retrieval task. We take the model trained on ImageNet to extract representations for CIFAR-10 [29] test set and CIFAR-10-C [18] (\"clean\" and \"noise level 1-5\" in Fig Supp.1b in the supplementary respectively), without retraining/finetuning. 1000 images are used as query and the rest are used as database. Results for \"fog\" noise are shown in Fig Supp.1b in the supplementary, while results for other noises are shown in supplementary too. As shown, the proposed attention guided model outperforms the baseline model on both clean images and almost all noise/corruption types and levels.", + "bbox": [ + 496, + 152, + 893, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 340, + 617, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we tested the hypothesis that utilizing human spatial attention can be beneficial for obtaining better representation for contrastive models. We overcome the challenge of obtaining human spatial attention labels for large dataset by utilizing a teacher model trained on limited human attention labels to provide pseudo-attention labels for ImageNet. We augmented training of a commonly-used network (ResNet50) trained with SimCLR contrastive objective with pseudo-spatial attention labels from the teacher model. Our results demonstrate that contrastive models trained with those pseudo-attention labels are more predictive of human attention and we obtain better representations.", + "bbox": [ + 496, + 364, + 890, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the gains observed on downstream tasks are not large, the gains were consistent across 3 tasks (classification, robustness, and image retrieval). Thus, taken together, our findings support the above hypothesis. The limited gains in classification may suggest that spatial attention alone may not be enough to achieve SOTA, but perhaps in conjunction with other architectural and methodological improvements. Another limitation is that our exploration was conducted on natural images only, but other interesting applications may include other domains like medical images or autonomous driving. When domains become very different, it may be required to retrain teacher models to capture those new domains (e.g., train on attention data that comes from medical experts for medical images). These are interesting and important questions that may be addressed in future work.", + "bbox": [ + 496, + 561, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 815, + 679, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We are grateful to Yifan Tian, Josh McAdams, Zhuode Liu, and Ethan Steinberg for assistance with analyses. We also thank Mike Mozer, Kai Kohlhoff, Simon Kornblith, and Ting Chen for useful discussions.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23289", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Borji. Saliency prediction in the deep learning era: Successes and limitations. IEEE transactions on pattern analysis and machine intelligence, 2019. 3, 4", + "[2] Zoya Bylinskii, Tilke Judd, Ali Borji, Laurent Itti, Frédo Durand, Aude Oliva, and Antonio Torralba. Mit saliency benchmark. http://saliency.mit.edu, 2012. 2, 3", + "[3] Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, and F. Durand. What do different evaluation metrics tell us about saliency models? IEEE TPAMI, 41(3):740-757, 2018. 5", + "[4] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5", + "[5] Jorge Paolo Casas and Chandramouli Chandrasekaran. openeyetrack-a high speed multi-threaded eye tracker for head-fixed applications. Journal of open source software, 4(42), 2019. 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 5, 7", + "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2", + "[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an lstmbased saliency attentive model. IEEE TIP, 27(10):5142- 5154, 2018. 5", + "[10] Abhishek Das, Harsh Agrawal, Larry Zitnick, Devi Parikh, and Dhruv Batra. Human attention in visual question answering: Do humans and deep networks look at the same regions? Computer Vision and Image Understanding, 163:90-100, 2017. 3", + "[11] Mohammad K Ebrahimpour, J Ben Falandays, Samuel Spevack, and David C Noelle. Do humans look where deep convolutional neural networks \"attend\"? In International Symposium on Visual Computing, pages 53-65. Springer, 2019. 3", + "[12] Simone Frintrop, Erich Rome, and Henrik I Christensen. Computational visual attention systems and their cognitive foundations: A survey. ACM Transactions on Applied Perception (TAP), 7(1):1-39, 2010. 2", + "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1", + "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 2", + "[15] Meng-Hao Guo, Tian-Xing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph R Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational Visual Media, pages 1-38, 2022. 3", + "[16] Suna Sihang Guo, Ruohan Zhang, Bo Liu, Yifeng Zhu, Dana Ballard, Mary Hayhoe, and Peter Stone. Machine versus human attention in deep reinforcement learning tasks. Advances in Neural Information Processing Systems, 34:25370-25385, 2021. 3", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 8", + "[19] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 262-270, 2015. 2, 3", + "[20] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 1", + "[21] Laurent Itti and Christof Koch. Computational modelling of visual attention. Nature reviews neuroscience, 2(3):194-203, 2001. 3", + "[22] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE TPAMI, 20(11):1254-1259, 1998. 5", + "[23] Saumya Jetley, Nicholas A Lord, Namhoon Lee, and Philip HS Torr. Learn to pay attention. arXiv preprint arXiv:1804.02391, 2018. 3", + "[24] Sen Jia and Neil DB Bruce. Eml-net: An expandable multilayer network for saliency prediction. Image and Vision Computing, 95:103887, 2020. 3", + "[25] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Salicon: Saliency in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1072-1080, 2015. 2, 3, 4, 6, 7", + "[26] T. Judd, F. Durand, and A. Torralba. A benchmark of computational models of saliency to predict human fixations. 2012. 5", + "[27] Nam Wook Kim, Zoya Bylinskii, Michelle A Borkin, Krzysztof Z Gajos, Aude Oliva, Fredo Durand, and Hanspeter Pfister. Bubbleview: an interface for crowdsourcing image importance maps and tracking visual attention. ACM Transactions on Computer-Human Interaction (TOCHI), 24(5):1-40, 2017. 2", + "[28] Christof Koch and Shimon Ullman. Shifts in selective visual attention: towards the underlying neural circuitry. In Matters of intelligence, pages 115-141. Springer, 1987. 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23290", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8", + "[30] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. arXiv preprint arXiv:1411.1045, 2014. 3", + "[31] Matthias Kummerer, Thomas SA Wallis, Leon A Gatys, and Matthias Bethge. Understanding low-and high-level contributions to fixation prediction. In Proceedings of the IEEE International Conference on Computer Vision, pages 4789-4798, 2017. 3", + "[32] Qixia Lai, Salman Khan, Yongwei Nie, Hanqiu Sun, Jianbing Shen, and Ling Shao. Understanding more about human and machine attention in deep neural networks. IEEE Transactions on Multimedia, 23:2086-2099, 2020. 2, 3, 5, 7", + "[33] Thomas Langlois, Haicheng Zhao, Erin Grant, Ishita Dasgupta, Tom Griffiths, and Nori Jacoby. Passive attention in artificial neural networks predicts human visual selectivity. Advances in Neural Information Processing Systems, 34, 2021. 3", + "[34] J. Lou, H. Lin, D. Marshall, D. Saupe, and H. Liu. Transal-net: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 5", + "[35] Sangwoo Mo, Hyunwoo Kang, Kihyuk Sohn, Chun-Liang Li, and Jinwoo Shin. Object-aware contrastive learning for debiased scene representation. Advances in Neural Information Processing Systems, 34, 2021. 3, 5, 7", + "[36] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10241-10247. IEEE, 2020. 3, 4", + "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 7", + "[38] Ramprasaath R Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. Casting your model: Learning to localize improves self-supervised representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11058-11067, 2021. 3", + "[39] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a hint: Leveraging explanations to make vision and language models more grounded. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2591-2600, 2019. 3", + "[40] Nikhil V Shirahatti and Kobus Barnard. Evaluating image retrieval. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 1, pages 955-961. IEEE, 2005. 8", + "[41] Ekta Sood, Fabian Kögel, Florian Strohm, Prajit Dhar, and Andreas Bulling. Vqa-mhug: A gaze dataset to study multimodal neural attention in visual question answering. arXiv preprint arXiv:2109.13116, 2021. 3", + "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[43] Nachiappan Valliappan, Na Dai, Ethan Steinberg, Junfeng He, Kantwon Rogers, Venky Ramachandran, Pingmei Xu, Mina Shojaeizadeh, Li Guo, Kai Kohlhoff, et al. Accelerating eye movement research via accurate and affordable smartphone eye tracking. Nature communications, 11(1):1-12, 2020. 2, 6, 7", + "[44] Leonard Elia Van Dyck, Roland Kwitt, Sebastian Jochen Denzler, and Walter Roland Gruber. Comparing object recognition in humans and deep convolutional neural networks—an eye tracking study. Frontiers in Neuroscience, page 1326, 2021. 3", + "[45] Brian A Wandell. Foundations of vision. Sinauer Associates, 1995. 2", + "[46] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2017. 3", + "[47] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2, 3", + "[48] Jeremy M Wolfe and Todd S Horowitz. What attributes guide the deployment of visual attention and how do they do it? Nature reviews neuroscience, 5(6):495-501, 2004. 2", + "[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2, 3", + "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2", + "[51] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 5", + "[52] Jan Zimmermann, Yuriria Vazquez, Paul W Glimcher, Bijan Pesaran, and Kenway Louie. Oculomatic: high speed, reliable, and accurate open-source eye tracking for humans and non-human primates. Journal of neuroscience methods, 270:138-146, 2016. 2" + ], + "bbox": [ + 501, + 92, + 890, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23291", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_model.json b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41045a8f474472f1f3e042aefbcaddee1c90173c --- /dev/null +++ b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.868, + 0.175 + ], + "angle": 0, + "content": "Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.204, + 0.21, + 0.22 + ], + "angle": 0, + "content": "Yushi Yao*" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.223, + 0.197, + 0.239 + ], + "angle": 0, + "content": "Waymo" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.243, + 0.236, + 0.256 + ], + "angle": 0, + "content": "yushiy@waymo.com" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.204, + 0.397, + 0.222 + ], + "angle": 0, + "content": "Chang Ye*" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.223, + 0.385, + 0.24 + ], + "angle": 0, + "content": "Google" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.243, + 0.434, + 0.257 + ], + "angle": 0, + "content": "yechang@google.com" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.204, + 0.609, + 0.222 + ], + "angle": 0, + "content": "Junfeng He†" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.223, + 0.591, + 0.24 + ], + "angle": 0, + "content": "Google" + }, + { + "type": "text", + "bbox": [ + 0.471, + 0.243, + 0.65, + 0.257 + ], + "angle": 0, + "content": "junfenghe@google.com" + }, + { + "type": "text", + "bbox": [ + 0.688, + 0.204, + 0.875, + 0.222 + ], + "angle": 0, + "content": "Gamaleldin F. Elsayed†" + }, + { + "type": "text", + "bbox": [ + 0.752, + 0.223, + 0.813, + 0.24 + ], + "angle": 0, + "content": "Google" + }, + { + "type": "text", + "bbox": [ + 0.686, + 0.243, + 0.874, + 0.257 + ], + "angle": 0, + "content": "gamaleldin@google.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.313, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.373, + 0.474, + 0.829 + ], + "angle": 0, + "content": "Human spatial attention conveys information about the regions of visual scenes that are important for performing visual tasks. Prior work has shown that the information about human attention can be leveraged to benefit various supervised vision tasks. Might providing this weak form of supervision be useful for self-supervised representation learning? Addressing this question requires collecting large datasets with human attention labels. Yet, collecting such large scale data is very expensive. To address this challenge, we construct an auxiliary teacher model to predict human attention, trained on a relatively small labeled dataset. This teacher model allows us to generate image (pseudo) attention labels for ImageNet. We then train a model with a primary contrastive objective; to this standard configuration, we add a simple output head trained to predict the attention map for each image, guided by the pseudo labels from teacher model. We measure the quality of learned representations by evaluating classification performance from the frozen learned embeddings as well as performance on image retrieval tasks (see supplementary material). We find that the spatial-attention maps predicted from the contrastive model trained with teacher guidance aligns better with human attention compared to vanilla contrastive models. Moreover, we find that our approach improves classification accuracy and robustness of the contrastive models on ImageNet and ImageNet-C. Further, we find that model representations become more useful for image retrieval task as measured by precision-recall performance on ImageNet, ImageNet-C, CIFAR10, and CIFAR10-C datasets." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.875, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.528, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Illustration. A teacher model is trained to predict human spatial-attention from a small dataset. Then the model is used to provide attention labels for larger dataset, which are used as additional targets for contrastive models." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.592, + 0.633, + 0.608 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.893, + 0.799 + ], + "angle": 0, + "content": "Deep learning models have made significant progress and obtained notable success on various vision tasks. Despite these promising results, humans continue to perform better than deep learning models in many applications. A notable reason is that deep learning models have a tendency to learn \"short-cuts\", i.e., giving significance to physically meaningless patterns or exploiting features which are predictive in some settings, but not causal [20]. Examples include focusing on less significant features such as background and textures [13]. These models yield representations that are less generalizable and lead to models that are highly sensitive to small pixel modulations [42]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.799, + 0.894, + 0.844 + ], + "angle": 0, + "content": "Human vision on the other hand is known to be much more robust and generalizable. One major difference between human and machine vision is that humans tend to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.852, + 0.664, + 0.864 + ], + "angle": 0, + "content": "*Equal technical contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.864, + 0.735, + 0.877 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal leadership and advising contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.877, + 0.607, + 0.888 + ], + "angle": 0, + "content": "Correspondence to:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.889, + 0.843, + 0.901 + ], + "angle": 0, + "content": "junfenghe@google.com & gamaleldin@google.com" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.852, + 0.843, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23282" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "focus on specific regions in visual scene [45]. These locations often reflect regions salient or useful to perform a specific vision task. Machines, instead, initially place equal significance to all regions. A natural question is: will it be beneficial if machine vision models is guided by human spatial attention?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.394 + ], + "angle": 0, + "content": "Human spatial attention has been shown to benefit computer vision models in supervised tasks, such as classification [32]. Yet, it is still a question whether adding a form of weak supervision in the form of human spatial attention could similarly benefit self-supervised models that are trained end-to-end. Self-supervised models typically need a large amount of data to yield good representations. To test if training weakly supervised models with human spatial attention cues, we will need to collect a large volume of human spatial attention labels, which is a very expensive process that requires either using trackers to record eye movements [5,43,52] or asking humans to highlight regions that they attend to [25, 27]. This process is prohibitively tedious and costly for datasets with millions of examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.742 + ], + "angle": 0, + "content": "In this work, we test the hypothesis that a weak supervision in the form of human spatial attention is beneficial for representation learning for models trained with a contrastive objective. Inspired by knowledge distillation and self-training using teacher models [47, 49], we address the challenge of obtaining spatial attention labels on large scale image datasets by using machine pseudo-labeling. We train a teacher model on a set of limited ground truth human spatial attention labels, and use this teacher model to generate spatial attention pseudo-labels for the large ImageNet benchmark. We are then able to utilize the generated spatial attention maps in the contrastive models, and discover that this approach yields representations that are highly predictive of human spatial attention. Further, we find that the learned representations are better as measured by higher accuracy and robustness on classification downstream tasks, and higher precision and recall on image retrieval tasks. Interestingly, we find that the gains from using teacher models to provide pseudo labels are larger than using the limited ground truth human labels directly when training contrastive models, and the gains are larger for contrastive models than when applying same method to supervised models. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.751, + 0.47, + 0.828 + ], + "angle": 0, + "content": "- We create a dataset with spatial attention maps for the ImageNet [37] benchmark by first training a teacher model to predict human spatial attention labels from Silicon dataset [25] and then use the model to label ImageNet examples" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.837, + 0.47, + 0.867 + ], + "angle": 0, + "content": "- We use spatial-attention labels from the teacher model as an additional prediction target to models trained" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.751, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.708, + 0.107 + ], + "angle": 0, + "content": "with contrastive objective." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.119, + 0.892, + 0.21 + ], + "angle": 0, + "content": "- We find that the proposed method can learn better representation, leading to better accuracy and robustness for downstream classification tasks (on ImageNet and ImageNet-C), and better performance on retrieval tasks (on ImageNet, ImageNet-C, CIFAR-10, and CIFAR10-C)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.638, + 0.241 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Contrastive learning: Contrastive learning has gained popularity in the past few years for self-supervised and semi-supervised representation learning. In general, contrastive learning aims to learn similar representations for similar data pairs and different representations for different pairs. SimCLR [6] utilized MLP projection heads and strong data augmentation for constructing similar pairs and have demonstrated great gains in image classification downstream tasks. To form the contrastive loss for each minibatch with N examples, the similar data pairs are constructed from two augmentations of the same image and different pairs from the other images within a batch, and then computing the NT-Xent loss. A different formulation is used in [50] by encouraging the empirical cross correlation of the representations of two versions of augmented mini-batch to be close to identity. In [17], it is further proposed to build large dictionaries for self-supervised learning (MOCO), and moreover in [7], better results are achieved on image classification and object detection tasks when combining advances from SimCLR and MOCO. Follow up work by [14] managed to obtain good performance without the need for dissimilar pairs by encouraging the representation of similar pairs across two versions of the network (trained network and exponential moving average version) to be similar. Further, in [8] it is shown that simple Siamese networks can still learn good representations without the need for dissimilar pairs, large batches or momentum encoders, etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Human spatial attention: Human visual system has developed an attention mechanism that focuses on regions in the visual space that are of interest or highly informative to the vision task [12, 48]. Eye trackers are often used to collect human spatial attention [5, 43, 52]. Many gaze data sets [2] have been collected with these eye trackers. Besides eye trackers, human spatial attention data can also be collected via mouse tracking [25, 27], e.g., users see a blurry version of an image, then click on regions they want to see more clearly, mimicking human's peripheral vision based on neurophysiological studies [19, 27]. Both eye tracking and mouse tracking methods are very expensive, which limit the number of examples in those datasets. Due to the relatively lower cost of mouse tracking, one can often generate relatively larger attention data from this method than" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Trained teacher model is available at: https://github.com/google-research/google-research/tree/master/humanattention/" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23283" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "eye tracking. For example, Silicon [25] dataset is one of the largest spatial attention datasets, contains around 20K images, each with attention labels from 50-60 participants, via a mouse tracking system, under free viewing setting. Yet, this data is still orders of magnitude smaller than those needed to train contrastive representation learning models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.337 + ], + "angle": 0, + "content": "Spatial attention/saliency prediction models aim at predicting which areas in an image will be salient to human attention and attract eye fixation, usually with collected gaze/attention data as ground-truth. Early works in saliency prediction usually define saliency through a set of hand crafted features such as color difference, contrast, intensities, etc. [21, 28]. Recent works [1, 19, 24, 30, 31, 36] leverage the power of deep neural networks and are often trained/fine-tuned on large scale gaze data sets like Silicon [2, 25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.471, + 0.717 + ], + "angle": 0, + "content": "Spatial attention of computer vision models: Spatial attention in neural networks can be mainly categorized into post-hoc attention like class activation map (CAM) ([51]), and trainable attention (e.g., [15, 23, 46]). Post-hoc spatial attention methods have been proposed to estimate regions in the image that are important or give rise to model decisions, often for model interpretation. In supervised settings where classification labels are known, the simplest and most direct method is class activation map (CAM) [51]. CAM uses class labels to extract the feature map that is most informative about the true class of an image. Grad-CAM [39] generalizes the CAM to apply to any model with any downstream task. [38] proposed to use Grad-CAM to design augmentation policies in self-supervised learning to tackle weak performance in complex scene images with many objects. ContraCAM [35] applies Grad-CAM assuming downstream task of contrastive learning, thus allowing computing spatial attention maps with no class label supervision. In [35], it is proposed to utilize the spatial attention information learned from ContraCAM to design data augmentation strategies to discourage contextual and background biases in a scene. Unlike [38] and [35] that uses spatial attention to design augmentation policies, here we focus on an end-to-end framework to predict spatial attention targets instead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Recently, there are also several papers [10, 11, 16, 33, 41, 44]) exploring the similarity/difference of model spatial attention vs human spatial attention, for the task of VQA (visual question answering) [10], object detection [11], reinforcement learning [16], etc. Among existing works, [32] is the most relevant to ours, since it conducted experiments to use human spatial attention to supervise model spatial attention, for three tasks (salient object segmentation, video action recognition and fine-grained image classification) and demonstrated that human spatial attention is beneficial. However, it still remains a question whether such benefits could be extended to contrastive representation learning. In" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "terms of strategy to utilize spatial attention labels, in [32], attention labels are used as spatial weighting, while we designed auxiliary task to predict spatial attention labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.379 + ], + "angle": 0, + "content": "Teacher model pseudo-labeling: Previous work on knowledge distillation and machine self training has demonstrated that machine teaching machines approaches may address the challenge of labeling large datasets. In image classification, [49] demonstrated that training a model to classify images then use that model to provide pseudolabels improved classification performance. Related idea is applied in [47] for language models, which showed that the knowledge learned by language models pretrained on text corpus could be distilled to generate new datasets for common reasoning, and that training common reasoning models on this new data largely improve performance. Inspired by these successes, we train a teacher model on smaller human attention data and use this model to generate new spatial attention pseudo labels for ImageNet benchmark (see Figure 1 and Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.599, + 0.403 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.412, + 0.831, + 0.429 + ], + "angle": 0, + "content": "3.1. The Contrastive Learning Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.679 + ], + "angle": 0, + "content": "Contrastive learning is one of the most popular self-supervised representation learning methods. It learns an embedding space so that similar data (positive) pairs are mapped to be close in the embedding space and different (negative) pairs are mapped to be far away. In practice, positive pairs are often generated by applying data augmentation to one image like adding noises, cropping, etc and negative pairs are different examples in the mini-batch. Among contrastive learning methods, SimCLR framework [6] has shown solid performance, which we choose here as our main method. For each batch, the images are augmented in two different ways. Then we feed them into ResNet feature extractor backbone and compute NT-Xent loss to minimize the difference between augmentations of the same images and maximize the differences between different images. NT-Xent loss can be computed as follows [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.687, + 0.892, + 0.724 + ], + "angle": 0, + "content": "\\[\nl _ {i, j} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {j}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {2 N} 1 _ {k \\neq i} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {k}\\right) / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where \\(z_{i}\\) is the embedding of ith example, and \\(\\operatorname{sim}(\\cdot ,\\cdot)\\) is the cosine similarity between two embeddings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.892, + 0.803 + ], + "angle": 0, + "content": "3.2. ImageNet-Attn: spatial attention maps for ImageNet generated by a teacher model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To train contrastive learning models, we need large datasets with millions of examples. Thus, to address the question whether human attention is beneficial for contrastive models, we will need large human spatial attention labeled dataset. However, there is no spatial attention dataset available with that size on typical large image" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23284" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.093, + 0.133, + 0.102 + ], + "angle": 0, + "content": "a" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.105, + 0.236, + 0.115 + ], + "angle": 0, + "content": "Teacher model" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.116, + 0.826, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.251, + 0.133, + 0.261 + ], + "angle": 0, + "content": "b" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.258, + 0.244, + 0.282 + ], + "angle": 0, + "content": "Sallcon dataset image" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.292, + 0.276, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.381, + 0.276, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.468, + 0.276, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.269, + 0.404, + 0.291 + ], + "angle": 0, + "content": "ground truth human spatial attention" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.292, + 0.408, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.381, + 0.408, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.469, + 0.408, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.269, + 0.523, + 0.291 + ], + "angle": 0, + "content": "teacher model predictions" + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.292, + 0.539, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.381, + 0.539, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.469, + 0.539, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.255, + 0.569, + 0.264 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.258, + 0.685, + 0.284 + ], + "angle": 0, + "content": "ImageNet dataset image" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.293, + 0.69, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.381, + 0.69, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.47, + 0.69, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.271, + 0.841, + 0.293 + ], + "angle": 0, + "content": "spatial-attention labels from the teacher model" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.294, + 0.832, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.381, + 0.831, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.47, + 0.831, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.566, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Figure 2. (a) The architecture of teacher model to predict attention. The teacher model was trained from random initialization to predict ground truth (G.T.) human spatial attention labels. (b) Examples from Salicon dataset, with G.T. human attention heatmaps, and teacher model predictions. (c) Examples of spatial-attention pseudo-labels generated by the teacher model for ImageNet dataset (ImageNet-Attn)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.68 + ], + "angle": 0, + "content": "benchmark such as ImageNet. The largest attention data set available is Silicon [25], consisting of twenty thousand examples only, orders of magnitude smaller than ImageNet, which has around 14 million examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To generate attention labels for ImageNet images, we train a teacher model on Silicon attention data set to predict human attention ground truth labels (see Figure 2b). Then we use this teacher model to create a new dataset (ImageNet-Attn) with spatial attention maps for ImageNet (see Figure 2c for some examples). The teacher model architecture is illustrated in Figure 2a. Specifically, we used MobileNet-V3-small as the backbone, and embeddings from 4 layers (conv 2,4,6,8) are extracted. On each embedding, we applied two conv layers (the first conv layer with \\(3 \\times 3\\) kernel, number of channels matching the input embedding, max pooling of 3 and relu; the second conv layer with \\(1 \\times 1\\) kernel, 1 channel and relu). We bilinearly resized the output to the same resolution of input. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.861 + ], + "angle": 0, + "content": "result of the 4 branches are then summed to yield 1 feature, followed up by a sigmoid function to obtain spatial-attention map. Note that using multiple intermediate (both early and late) layers is one of the key ingredient to make the spatial attention teacher model successful. This strategy mimics human visual attention, which is known to be affected by both low-level characteristics such as color, intensity, texture, as well as high level characteristics like shape and object, etc. Essentially, our teacher model has a similar architecture as in [36], but is further simplified with more efficient backbone, less channels/layers, etc, so that it can be trained from scratch on Silicon data with randomly initialized backbone to avoid any leak of class label information, while most existing attention prediction models [1,36] needs to finetune with a pretrained classification backbone network." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23285" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "3.3. Training contrastive models with spatial attention maps" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.129, + 0.468, + 0.279 + ], + "angle": 0, + "content": "Spatial Attention Branch The overall training framework is shown in Figure 1. This spatial attention prediction framework consists of two branches: the contrastive branch and the spatial attention branch. The contrastive branch is the same as the original SimCLR method [6], which applies augmentations to image \\( x \\) to get different variants \\( x_{i} \\) and \\( x_{j} \\), and learns the representation \\( h_{i} \\) and \\( h_{j} \\) via a feature extractor backbone network (e.g., ResNet), then use a projection head to map \\( h_{i} / h_{j} \\) to \\( z_{i} / z_{j} \\), where the contrastive loss is applied." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.468, + 0.521 + ], + "angle": 0, + "content": "For the spatial-attention branch, it takes as an input not only the final embedding \\( h \\), but also early intermediate layer embeddings (following the same reason of including both high level and low level visual cues as in teacher model), and predicts an spatial attention heatmap \\( m_p \\). More specifically, we apply a global average pooling on the output of the last three blocks of the ResNet backbone. Then, we select the max channel for each of three block output (after average pooling), and resize with bilinear interpolation to the image resolution, which has some resemblance to CAM approaches [51]. Finally we stack the representations together, pass them into a linear readout layer (with a bias term), and use the output as our final spatial attention prediction \\( m_p \\). We keep the spatial attention head as simple as possible so that the guide on attention head output can be back propagated to representation more directly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.702 + ], + "angle": 0, + "content": "We use the ImageNet-Attn data generated by the teacher model as target for each ImageNet example, denoted as \\( m_t \\), we can then train the network spatial attention output \\( m_p \\) to be close to \\( m_t \\). We hypothesize that this method regularizes the training of the feature extractor backbone rather than explicitly enforce the network to generate masked representations that match the spatial attention maps. Note that for attention branch, there is no augmentation applied to each image \\( x \\), since human attention is not invariant to transformation (e.g., a human looking at a cropped image may attend to different regions compared to a consistent crop of human attention map of the original image)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.703, + 0.468, + 0.731 + ], + "angle": 0, + "content": "Loss function The loss function \\(L\\) consists of two terms:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.734, + 0.468, + 0.748 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {c o n t r a}} + L _ {\\text {a t t n}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.468, + 0.813 + ], + "angle": 0, + "content": "\\(L_{\\text{contra}}\\) is the contrastive loss, or more specifically \\(L_{\\text{contra}} = \\sum_{i,j} l_{i,j}\\) and \\(l_{i,j}\\) is defined in Equation (1), the same as in [6]. \\(L_{\\text{attr}}\\) is the attention regularization loss, and more specifically" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.819, + 0.468, + 0.849 + ], + "angle": 0, + "content": "\\[\nL _ {a t t n} = \\sum_ {i} \\left(\\lambda K L D \\left(m _ {i} ^ {p}, m _ {i} ^ {t}\\right) - \\beta N S S \\left(m _ {i} ^ {p}, p _ {i} ^ {t}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\beta\\) are two weighting parameters (\\(\\lambda = 1.0\\) and \\(\\beta = 0.1\\) in our work). \\(m_{i}^{p}\\) is the predicted 2D spatial-attention map from the attention head of our proposed" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 1. ImageNet Top-1 classification accuracy for different models (mean \\(\\pm\\) SE for 3 seeds, except for * which means best result from all runs)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.144, + 0.89, + 0.301 + ], + "angle": 0, + "content": "
ModelAccuracy (%)
Contrastive67.61 ± 0.04
Contrastive attn. teacher68.23 ± 0.08
Contrastive attn. co-train66.35 ± 0.12
Contrastive attn. with explicit attention block [32]61.15*
Contrastive attn. with self attention mask [32]49.65*
Contrastive attn. with ContraCAM [35] as attention67.71*
Supervised75.91 ± 0.10
Supervised attn. teacher76.02 ± 0.04
Supervised (ResNet-18)69.17 ± 0.07
Supervised (ResNet-18) attn. teacher69.30 ± 0.04
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.659 + ], + "angle": 0, + "content": "model for \\(i\\)-th example (original image, not augmented ones), \\(m_i^t\\) is the pseudo spatial-attention target map predicted by the teacher model for \\(i\\)-th example, and \\(KLD()\\) is the KL divergence1. Besides KLD loss, we also use \\(NSS()\\), the Normalized Scanpath Saliency loss [4]. Those two losses are typically used for human attention predictions [3,9]. KLD is often used to match the spatial-attention target distribution, and NSS is typically added on top as it is generally observed to help generate attention maps that are in perceptual agreement with human judgements (see [3,34]). For \\(NSS\\) loss, the larger the better, so there is a negative sign before it. \\(NSS\\) loss needs gaze/attention points instead of heatmap as ground-truth, so we extract pseudo gaze/attention points \\(p_i^t\\) from the attention heatmap \\(m_i^t\\). To obtain \\(p_i^t\\), we first extract the point with highest value in current spatial-attention map \\(m_i^t\\), then generate a new spatial-attention map by subtracting a Gaussian blur around the extracted point from the current attention map. The process is repeated with the new attention map until the maximal value of the attention map is smaller than a threshold (more specifically, \\(0.2 * \\max\\_ sal\\), where \\(\\max\\_ sal\\) is the maximal value in \\(m_i^t\\)).2" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.633, + 0.69 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.714, + 0.714 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.722, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In our experiments, in Eq.1, we choose \\(\\tau = 0.1\\) and \\(N = 2048\\). In Eq.2, we used equal weighting of 1.0 for the SimCLR and attention losses." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.767, + 0.892, + 0.782 + ], + "angle": 0, + "content": "The weight for KLD and NSS losses are 1 and 0.1, re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.791, + 0.892, + 0.828 + ], + "angle": 0, + "content": "\\( {}^{1}{m}_{i}^{p} \\) is normalized when computing KLD, i.e.,divided by its pixel value sum,so that it becomes a distribution with all pixel value sum equal to 1. Similarly for \\( {m}_{i}^{t} \\) ." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.828, + 0.892, + 0.9 + ], + "angle": 0, + "content": "2Our process of extracting gaze points from heatmaps is to inverse the process of generating heatmaps from gaze points. Due to the uncertainty about the gaze locations, researchers typically apply Gaussian blur for each gaze point, then sum all Gaussians to generate the heatmap (see [3, 26]). We follow the inverse of this process to obtain gaze points from heatmap, similarly as done in [22]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.791, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23286" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.108, + 0.114, + 0.117 + ], + "angle": 0, + "content": "a" + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.117, + 0.208, + 0.129 + ], + "angle": 0, + "content": "image" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.144, + 0.254, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.117, + 0.396, + 0.141 + ], + "angle": 0, + "content": "ground truth human spatial attention" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.143, + 0.404, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.117, + 0.543, + 0.141 + ], + "angle": 0, + "content": "teacher model spatial attention" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.144, + 0.556, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.117, + 0.716, + 0.14 + ], + "angle": 0, + "content": "spatial attention guided contrastive model" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.143, + 0.708, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.117, + 0.85, + 0.14 + ], + "angle": 0, + "content": "baseline trastive model" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.143, + 0.864, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.245, + 0.253, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.347, + 0.253, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.245, + 0.403, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.347, + 0.403, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.245, + 0.555, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.347, + 0.553, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.245, + 0.707, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.347, + 0.706, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.246, + 0.864, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.347, + 0.864, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.446, + 0.114, + 0.457 + ], + "angle": 0, + "content": "b" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.457, + 0.48, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.626, + 0.602, + 0.638 + ], + "angle": 0, + "content": "correlation coeff. (model vs human)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.456, + 0.871, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.646, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Figure 3. (a) Examples comparing spatial attention maps predicted by different models vs ground truth human attention data on OSIE dataset [43]. (b) Distribution of correlation coefficients between model predicted attention maps vs ground truth human attention maps on Silicon [25] validation set (left) and OSIE dataset (right) [43]. Three models are shown: (1) Teacher model trained on Silicon [25] training set. (2) Baseline SimCLR model (3) The proposed spatial attention guided SimCLR model trained with pseudo-attention labels provided by the teacher model. Please note that both baseline and our model are trained to predict target attention labels, with main difference that the baseline model has a stop gradient placed between the prediction layer and network features to prevent spatial-attention targets to inform network features. Also note that our teacher model is trained on Silicon where the attention is collected by mouse tracking, while the results in (a) and (c) here are on OSIE data, where the attention are collected by eye trackers. The results demonstrated both the teacher model and the attention head in proposed model has a good generalization capacity for new kinds of attention labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.848 + ], + "angle": 0, + "content": "spectively. We obtained those weights by hyperparameter search from 0.1-5 for KLD and 0.01-1 for NSS, based on ImageNet accuracy of a separate dev set (80%-20% split on training set for parameter search with 20% as dev set)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.892, + 0.816 + ], + "angle": 0, + "content": "4.2. Spatial attention guided models are highly predictive of human attention" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this section, we explore whether the use of auxiliary teacher model to provide spatial attention pseudo labels on ImageNet better aligns contrastive model's attention with human attention. We define aligning model spatial attention here as the ability to predict spatial attention mask from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23287" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "the model backbone features by a simple readout layer discussed in Section 3. We trained two ResNet-50 backbones using the SimCLR objective from [6]. We added additional spatial-attention losses as discussed in Section 3. For the baseline model, we placed a stop gradient operation between the backbone features and the attention projection head to prevent attention information from leaking to the backbone features, whereas for the attention guided model, we allowed the learned attention gradients to flow back to the backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.473, + 0.41 + ], + "angle": 0, + "content": "We evaluated the degree the predicted attention maps is aligned with human attention by performing correlation analysis, which is typically used to compare human and model spatial attention. We measured Pearson's correlation coefficient between the model predicted attention and ground truth human spatial-attention maps from Silicon [25] validation set (Note that the teacher model is trained with Silicon training set). We find that both teacher model and spatial attention guided contrastive model are highly correlated with human attention (Figure 3b), while the baseline contrastive model is much less correlated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.473, + 0.685 + ], + "angle": 0, + "content": "To test whether this alignment with human attention is general, beyond just Salicon data, we obtained another attention dataset on OSIE images [43] (Figure 3a). This data more faithfully reflects human attention as it collected from (mobile) eye tracker [43]. We find that the baseline model is weakly positively correlated with human attention (ttest: \\(\\rho = 0.07\\) \\(p < 0.001\\)) suggesting that the contrastive loss produces features that are predictive of human attention to some extent. Yet, the correlation was generally close to 0 and explains only \\(0.5\\%\\) of data variance. The spatial attention guided model has a much stronger correlation to human attention (ttest: \\(\\rho = 0.48\\) \\(p < 0.001\\)) than the baseline model (Two samples ttest: \\(p < 0.001\\)), and thus more faithfully reflecting human visual attention (See Fig 3a for qualitative examples and Fig 3b bottom for quantitative analysis). Further, the correlation with human attention for the teacher model and the attention guided contrastive model were quite similar (Two samples ttest: \\(p = 0.7\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.47, + 0.731 + ], + "angle": 0, + "content": "4.3. Spatial attention guided models are more accurate than baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We evaluate the quality of the representations learned by spatial attention guidance framework using the typical contrastive learning evaluation criteria: fitting an ImageNet [37] linear classifier on top of the frozen representation (in practice we place stop gradient at the end of the backbone and train the classifier concurrently while training the backbone). We compute Top 1 accuracy on ImageNet validation set and compare the results with baselines. As shown in Table 1, we observe around \\(0.6\\%\\) accuracy gain on ImageNet compared to vanilla SimCLR. We further explore an alternative way of incorporating human attention data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Rather than using pseudo attention labels on ImageNet from the teacher model, we add Silicon data to the training data, and directly predict attention ground truth labels for Silicon data with spatial attention head (Contrastive attn. co-train in Table 1). More specifically, we pass 2048 ImageNet images and N=2048 Silicon images into our backbone (N is obtained by hyperparameter search 512-2048 based on evaluation on the separate dev set) to compute backbone features. Then, we compute the SimCLR loss from ImageNet features and use Silicon images features to predict spatial attention maps, and we use Silicon attention ground truth labels to compute the KLD and NSS losses. Interestingly, we find this method to lead to worse performance compared to using the teacher model generated spatial-attention labels, which gives evidence that the teacher model is generalizing its knowledge about human attention data beyond the limited Silicon training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.653 + ], + "angle": 0, + "content": "One important question is whether the classic trainable attention method applies to contrastive learning. For example, in [32], an attention block/map is used to adjust the weights of the embedding, where the attention block can be self learned implicitly, or guided by human attention ground-truth. It shows both implicit self-learned attention block and attention block guided by human attention can improve the accuracy of fine-grained classification (Table VII in [32]). We adopted the approach described in their paper: add an attention block after the representation embedding, and use the attention block's output to mask the representation embedding. We apply it to the contrastive models, and the results are shown in Table 1 as \"explicit attention mask\" (the attention block is guided by human attention) and \"self attention mask\" (self attention). As observed, both yield poor performance, since the contrastive attention with self-attention mask method learns an attention mask from the contrastive loss rather than to predict the teacher targets, the model may be getting weaker supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Another important question is whether our current spatial attention targets can be replaced by other spatial attention approach like ContraCAM [35]. In our method, we predicted the spatial attention with a linear layer. Whereas in ContraCAM a contrastive loss is used to construct the attention prediction. In both cases, we trained the predictions to match the target attention maps obtained from the teacher model. The result is shown in Table 1 as \"Contrastive with ContraCAM as attention\". However, we don't see much accuracy improvements with ContraCAM as the model attention, showing that the linear simple prediction leads to better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To investigate whether the spatial attention guidance framework benefits supervised models in the same way as contrastive models, we applied the same approach for supervised models. Supervised models similarly benefit from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23288" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "this framework, yet the gain is limited compared to the contrastive models perhaps due to the higher accuracy the supervised models achieves compared to the contrastive models. To control for accuracy, we used a supervised model trained with smaller backbone (ResNet-18), which gives comparable accuracy to the bigger ResNet-50 backbone that is trained with contrastive objective. Even when controlling the mismatch in accuracy, the supervised model gain from providing spatial-attention supervision is limited compared to contrastive models (Table 1)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.254, + 0.471, + 0.284 + ], + "angle": 0, + "content": "4.4. Spatial attention guided models are more robust than baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.473, + 0.549 + ], + "angle": 0, + "content": "Human vision is very robust to noise or small pixel modulations, compared to computer vision models. Here, we hypothesize that models that aligns better with human attention may learn more robust representations. We test this hypothesis using ImageNet-C dataset [18]. We take the representations learned from the proposed model and train a linear classifier on ImageNet training data. We then evaluate classification performance on ImageNet-C at various corruption/noise types. Table Supp.1 in the supplementary compares the classification performance on ImageNet-C for baseline contrastive model and the proposed models trained with spatial-attention teacher guidance, average over the 5 different corruption/noise magnitudes in ImageNet-C. We find that contrastive models trained with teacher guidance outperforms baseline consistently, suggesting that the representation learned by spatial-attention guided model is indeed more robust." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.47, + 0.592 + ], + "angle": 0, + "content": "4.5. Spatial attention guided models generate better representation for retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Besides classification, we tested the quality of the representations for another downstream task: image retrieval. We use the model to extract representation for ImageNet validation set, and use the representations to run image retrieval. 5000 randomly chosen validation images are used as query, while the rest validation images are database images. For each query image, top k retrieval results are returned by sorting the cosine distance between representations of database images and the query. The retrieval accuracy is measured with the standard precision recall curve [40], where precision and recall are computed by checking whether results have the same class label as the query. Mean precision and recall is obtained as the average across queries. We choose different k values, and draw the mean precision-recall curve. Besides ImageNet (\"clean\" in Fig Supp.1a in the supplementary), we also use ImageNet-C (\"noise level 1-5\"). Results from \"fog\" corruption is shown in In Figure Supp.1a in the supplementary., while results for other corruptions can be found in the supplementary too. The representation extracted with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "the Contrastive attn teacher model outperforms the baseline Constractive model in this retrieval task, on both ImageNet images, and ImageNet-C images for most corruption types and levels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.895, + 0.333 + ], + "angle": 0, + "content": "Moreover, we evaluate transfer to other dataset for the retrieval task. We take the model trained on ImageNet to extract representations for CIFAR-10 [29] test set and CIFAR-10-C [18] (\"clean\" and \"noise level 1-5\" in Fig Supp.1b in the supplementary respectively), without retraining/finetuning. 1000 images are used as query and the rest are used as database. Results for \"fog\" noise are shown in Fig Supp.1b in the supplementary, while results for other noises are shown in supplementary too. As shown, the proposed attention guided model outperforms the baseline model on both clean images and almost all noise/corruption types and levels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.341, + 0.619, + 0.356 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.56 + ], + "angle": 0, + "content": "In this work, we tested the hypothesis that utilizing human spatial attention can be beneficial for obtaining better representation for contrastive models. We overcome the challenge of obtaining human spatial attention labels for large dataset by utilizing a teacher model trained on limited human attention labels to provide pseudo-attention labels for ImageNet. We augmented training of a commonly-used network (ResNet50) trained with SimCLR contrastive objective with pseudo-spatial attention labels from the teacher model. Our results demonstrate that contrastive models trained with those pseudo-attention labels are more predictive of human attention and we obtain better representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Despite the gains observed on downstream tasks are not large, the gains were consistent across 3 tasks (classification, robustness, and image retrieval). Thus, taken together, our findings support the above hypothesis. The limited gains in classification may suggest that spatial attention alone may not be enough to achieve SOTA, but perhaps in conjunction with other architectural and methodological improvements. Another limitation is that our exploration was conducted on natural images only, but other interesting applications may include other domains like medical images or autonomous driving. When domains become very different, it may be required to retrain teacher models to capture those new domains (e.g., train on attention data that comes from medical experts for medical images). These are interesting and important questions that may be addressed in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.681, + 0.833 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We are grateful to Yifan Tian, Josh McAdams, Zhuode Liu, and Ethan Steinberg for assistance with analyses. We also thank Mike Mozer, Kai Kohlhoff, Simon Kornblith, and Ting Chen for useful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23289" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Ali Borji. Saliency prediction in the deep learning era: Successes and limitations. IEEE transactions on pattern analysis and machine intelligence, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.47, + 0.201 + ], + "angle": 0, + "content": "[2] Zoya Bylinskii, Tilke Judd, Ali Borji, Laurent Itti, Frédo Durand, Aude Oliva, and Antonio Torralba. Mit saliency benchmark. http://saliency.mit.edu, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.471, + 0.243 + ], + "angle": 0, + "content": "[3] Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, and F. Durand. What do different evaluation metrics tell us about saliency models? IEEE TPAMI, 41(3):740-757, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[4] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[5] Jorge Paolo Casas and Chandramouli Chandrasekaran. openeyetrack-a high speed multi-threaded eye tracker for head-fixed applications. Journal of open source software, 4(42), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.47, + 0.416 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.461, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an lstmbased saliency attentive model. IEEE TIP, 27(10):5142- 5154, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[10] Abhishek Das, Harsh Agrawal, Larry Zitnick, Devi Parikh, and Dhruv Batra. Human attention in visual question answering: Do humans and deep networks look at the same regions? Computer Vision and Image Understanding, 163:90-100, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[11] Mohammad K Ebrahimpour, J Ben Falandays, Samuel Spevack, and David C Noelle. Do humans look where deep convolutional neural networks \"attend\"? In International Symposium on Visual Computing, pages 53-65. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[12] Simone Frintrop, Erich Rome, and Henrik I Christensen. Computational visual attention systems and their cognitive foundations: A survey. ACM Transactions on Applied Perception (TAP), 7(1):1-39, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Meng-Hao Guo, Tian-Xing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph R Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational Visual Media, pages 1-38, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Suna Sihang Guo, Ruohan Zhang, Bo Liu, Yifeng Zhu, Dana Ballard, Mary Hayhoe, and Peter Stone. Machine versus human attention in deep reinforcement learning tasks. Advances in Neural Information Processing Systems, 34:25370-25385, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 262-270, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[20] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[21] Laurent Itti and Christof Koch. Computational modelling of visual attention. Nature reviews neuroscience, 2(3):194-203, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[22] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE TPAMI, 20(11):1254-1259, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[23] Saumya Jetley, Nicholas A Lord, Namhoon Lee, and Philip HS Torr. Learn to pay attention. arXiv preprint arXiv:1804.02391, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[24] Sen Jia and Neil DB Bruce. Eml-net: An expandable multilayer network for saliency prediction. Image and Vision Computing, 95:103887, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[25] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Salicon: Saliency in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1072-1080, 2015. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[26] T. Judd, F. Durand, and A. Torralba. A benchmark of computational models of saliency to predict human fixations. 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[27] Nam Wook Kim, Zoya Bylinskii, Michelle A Borkin, Krzysztof Z Gajos, Aude Oliva, Fredo Durand, and Hanspeter Pfister. Bubbleview: an interface for crowdsourcing image importance maps and tracking visual attention. ACM Transactions on Computer-Human Interaction (TOCHI), 24(5):1-40, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Christof Koch and Shimon Ullman. Shifts in selective visual attention: towards the underlying neural circuitry. In Matters of intelligence, pages 115-141. Springer, 1987. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23290" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[30] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. arXiv preprint arXiv:1411.1045, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[31] Matthias Kummerer, Thomas SA Wallis, Leon A Gatys, and Matthias Bethge. Understanding low-and high-level contributions to fixation prediction. In Proceedings of the IEEE International Conference on Computer Vision, pages 4789-4798, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[32] Qixia Lai, Salman Khan, Yongwei Nie, Hanqiu Sun, Jianbing Shen, and Ling Shao. Understanding more about human and machine attention in deep neural networks. IEEE Transactions on Multimedia, 23:2086-2099, 2020. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[33] Thomas Langlois, Haicheng Zhao, Erin Grant, Ishita Dasgupta, Tom Griffiths, and Nori Jacoby. Passive attention in artificial neural networks predicts human visual selectivity. Advances in Neural Information Processing Systems, 34, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[34] J. Lou, H. Lin, D. Marshall, D. Saupe, and H. Liu. Transal-net: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[35] Sangwoo Mo, Hyunwoo Kang, Kihyuk Sohn, Chun-Liang Li, and Jinwoo Shin. Object-aware contrastive learning for debiased scene representation. Advances in Neural Information Processing Systems, 34, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[36] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10241-10247. IEEE, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[38] Ramprasaath R Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. Casting your model: Learning to localize improves self-supervised representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11058-11067, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[39] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a hint: Leveraging explanations to make vision and language models more grounded. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2591-2600, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[40] Nikhil V Shirahatti and Kobus Barnard. Evaluating image retrieval. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 1, pages 955-961. IEEE, 2005. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[41] Ekta Sood, Fabian Kögel, Florian Strohm, Prajit Dhar, and Andreas Bulling. Vqa-mhug: A gaze dataset to study multimodal neural attention in visual question answering. arXiv preprint arXiv:2109.13116, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[43] Nachiappan Valliappan, Na Dai, Ethan Steinberg, Junfeng He, Kantwon Rogers, Venky Ramachandran, Pingmei Xu, Mina Shojaeizadeh, Li Guo, Kai Kohlhoff, et al. Accelerating eye movement research via accurate and affordable smartphone eye tracking. Nature communications, 11(1):1-12, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[44] Leonard Elia Van Dyck, Roland Kwitt, Sebastian Jochen Denzler, and Walter Roland Gruber. Comparing object recognition in humans and deep convolutional neural networks—an eye tracking study. Frontiers in Neuroscience, page 1326, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.891, + 0.303 + ], + "angle": 0, + "content": "[45] Brian A Wandell. Foundations of vision. Sinauer Associates, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[46] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[47] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[48] Jeremy M Wolfe and Todd S Horowitz. What attributes guide the deployment of visual attention and how do they do it? Nature reviews neuroscience, 5(6):495-501, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[51] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[52] Jan Zimmermann, Yuriria Vazquez, Paul W Glimcher, Bijan Pesaran, and Kenway Louie. Oculomatic: high speed, reliable, and accurate open-source eye tracking for humans and non-human primates. Journal of neuroscience methods, 270:138-146, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23291" + } + ] +] \ No newline at end of file diff --git a/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_origin.pdf b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a826ee7ff67ab3765f5e3aae3e8d8b3f120cd18 --- /dev/null +++ b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/7ff3c0aa-d380-4920-a7cd-d62a1a9e78dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d23954b35479ecfe35f6786bf3232e79596124c7ade77fe242e49aff62c2c4a +size 9353021 diff --git a/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/full.md b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76166f51a64b57307b41c1bdd278d6059d569602 --- /dev/null +++ b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/full.md @@ -0,0 +1,320 @@ +# Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models + +Yushi Yao* + +Waymo + +yushiy@waymo.com + +Chang Ye* + +Google + +yechang@google.com + +Junfeng He† + +Google + +junfenghe@google.com + +Gamaleldin F. Elsayed† + +Google + +gamaleldin@google.com + +# Abstract + +Human spatial attention conveys information about the regions of visual scenes that are important for performing visual tasks. Prior work has shown that the information about human attention can be leveraged to benefit various supervised vision tasks. Might providing this weak form of supervision be useful for self-supervised representation learning? Addressing this question requires collecting large datasets with human attention labels. Yet, collecting such large scale data is very expensive. To address this challenge, we construct an auxiliary teacher model to predict human attention, trained on a relatively small labeled dataset. This teacher model allows us to generate image (pseudo) attention labels for ImageNet. We then train a model with a primary contrastive objective; to this standard configuration, we add a simple output head trained to predict the attention map for each image, guided by the pseudo labels from teacher model. We measure the quality of learned representations by evaluating classification performance from the frozen learned embeddings as well as performance on image retrieval tasks (see supplementary material). We find that the spatial-attention maps predicted from the contrastive model trained with teacher guidance aligns better with human attention compared to vanilla contrastive models. Moreover, we find that our approach improves classification accuracy and robustness of the contrastive models on ImageNet and ImageNet-C. Further, we find that model representations become more useful for image retrieval task as measured by precision-recall performance on ImageNet, ImageNet-C, CIFAR10, and CIFAR10-C datasets. + +![](images/322434ca0869a5340264ba1200a146318688434775e26bb4924014b017ec42fb.jpg) +Figure 1. Illustration. A teacher model is trained to predict human spatial-attention from a small dataset. Then the model is used to provide attention labels for larger dataset, which are used as additional targets for contrastive models. + +# 1. Introduction + +Deep learning models have made significant progress and obtained notable success on various vision tasks. Despite these promising results, humans continue to perform better than deep learning models in many applications. A notable reason is that deep learning models have a tendency to learn "short-cuts", i.e., giving significance to physically meaningless patterns or exploiting features which are predictive in some settings, but not causal [20]. Examples include focusing on less significant features such as background and textures [13]. These models yield representations that are less generalizable and lead to models that are highly sensitive to small pixel modulations [42]. + +Human vision on the other hand is known to be much more robust and generalizable. One major difference between human and machine vision is that humans tend to + +focus on specific regions in visual scene [45]. These locations often reflect regions salient or useful to perform a specific vision task. Machines, instead, initially place equal significance to all regions. A natural question is: will it be beneficial if machine vision models is guided by human spatial attention? + +Human spatial attention has been shown to benefit computer vision models in supervised tasks, such as classification [32]. Yet, it is still a question whether adding a form of weak supervision in the form of human spatial attention could similarly benefit self-supervised models that are trained end-to-end. Self-supervised models typically need a large amount of data to yield good representations. To test if training weakly supervised models with human spatial attention cues, we will need to collect a large volume of human spatial attention labels, which is a very expensive process that requires either using trackers to record eye movements [5,43,52] or asking humans to highlight regions that they attend to [25, 27]. This process is prohibitively tedious and costly for datasets with millions of examples. + +In this work, we test the hypothesis that a weak supervision in the form of human spatial attention is beneficial for representation learning for models trained with a contrastive objective. Inspired by knowledge distillation and self-training using teacher models [47, 49], we address the challenge of obtaining spatial attention labels on large scale image datasets by using machine pseudo-labeling. We train a teacher model on a set of limited ground truth human spatial attention labels, and use this teacher model to generate spatial attention pseudo-labels for the large ImageNet benchmark. We are then able to utilize the generated spatial attention maps in the contrastive models, and discover that this approach yields representations that are highly predictive of human spatial attention. Further, we find that the learned representations are better as measured by higher accuracy and robustness on classification downstream tasks, and higher precision and recall on image retrieval tasks. Interestingly, we find that the gains from using teacher models to provide pseudo labels are larger than using the limited ground truth human labels directly when training contrastive models, and the gains are larger for contrastive models than when applying same method to supervised models. In summary, our contributions are as follows: + +- We create a dataset with spatial attention maps for the ImageNet [37] benchmark by first training a teacher model to predict human spatial attention labels from Silicon dataset [25] and then use the model to label ImageNet examples +- We use spatial-attention labels from the teacher model as an additional prediction target to models trained + +with contrastive objective. + +- We find that the proposed method can learn better representation, leading to better accuracy and robustness for downstream classification tasks (on ImageNet and ImageNet-C), and better performance on retrieval tasks (on ImageNet, ImageNet-C, CIFAR-10, and CIFAR10-C). + +# 2. Related work + +Contrastive learning: Contrastive learning has gained popularity in the past few years for self-supervised and semi-supervised representation learning. In general, contrastive learning aims to learn similar representations for similar data pairs and different representations for different pairs. SimCLR [6] utilized MLP projection heads and strong data augmentation for constructing similar pairs and have demonstrated great gains in image classification downstream tasks. To form the contrastive loss for each minibatch with N examples, the similar data pairs are constructed from two augmentations of the same image and different pairs from the other images within a batch, and then computing the NT-Xent loss. A different formulation is used in [50] by encouraging the empirical cross correlation of the representations of two versions of augmented mini-batch to be close to identity. In [17], it is further proposed to build large dictionaries for self-supervised learning (MOCO), and moreover in [7], better results are achieved on image classification and object detection tasks when combining advances from SimCLR and MOCO. Follow up work by [14] managed to obtain good performance without the need for dissimilar pairs by encouraging the representation of similar pairs across two versions of the network (trained network and exponential moving average version) to be similar. Further, in [8] it is shown that simple Siamese networks can still learn good representations without the need for dissimilar pairs, large batches or momentum encoders, etc. + +Human spatial attention: Human visual system has developed an attention mechanism that focuses on regions in the visual space that are of interest or highly informative to the vision task [12, 48]. Eye trackers are often used to collect human spatial attention [5, 43, 52]. Many gaze data sets [2] have been collected with these eye trackers. Besides eye trackers, human spatial attention data can also be collected via mouse tracking [25, 27], e.g., users see a blurry version of an image, then click on regions they want to see more clearly, mimicking human's peripheral vision based on neurophysiological studies [19, 27]. Both eye tracking and mouse tracking methods are very expensive, which limit the number of examples in those datasets. Due to the relatively lower cost of mouse tracking, one can often generate relatively larger attention data from this method than + +eye tracking. For example, Silicon [25] dataset is one of the largest spatial attention datasets, contains around 20K images, each with attention labels from 50-60 participants, via a mouse tracking system, under free viewing setting. Yet, this data is still orders of magnitude smaller than those needed to train contrastive representation learning models. + +Spatial attention/saliency prediction models aim at predicting which areas in an image will be salient to human attention and attract eye fixation, usually with collected gaze/attention data as ground-truth. Early works in saliency prediction usually define saliency through a set of hand crafted features such as color difference, contrast, intensities, etc. [21, 28]. Recent works [1, 19, 24, 30, 31, 36] leverage the power of deep neural networks and are often trained/fine-tuned on large scale gaze data sets like Silicon [2, 25]. + +Spatial attention of computer vision models: Spatial attention in neural networks can be mainly categorized into post-hoc attention like class activation map (CAM) ([51]), and trainable attention (e.g., [15, 23, 46]). Post-hoc spatial attention methods have been proposed to estimate regions in the image that are important or give rise to model decisions, often for model interpretation. In supervised settings where classification labels are known, the simplest and most direct method is class activation map (CAM) [51]. CAM uses class labels to extract the feature map that is most informative about the true class of an image. Grad-CAM [39] generalizes the CAM to apply to any model with any downstream task. [38] proposed to use Grad-CAM to design augmentation policies in self-supervised learning to tackle weak performance in complex scene images with many objects. ContraCAM [35] applies Grad-CAM assuming downstream task of contrastive learning, thus allowing computing spatial attention maps with no class label supervision. In [35], it is proposed to utilize the spatial attention information learned from ContraCAM to design data augmentation strategies to discourage contextual and background biases in a scene. Unlike [38] and [35] that uses spatial attention to design augmentation policies, here we focus on an end-to-end framework to predict spatial attention targets instead. + +Recently, there are also several papers [10, 11, 16, 33, 41, 44]) exploring the similarity/difference of model spatial attention vs human spatial attention, for the task of VQA (visual question answering) [10], object detection [11], reinforcement learning [16], etc. Among existing works, [32] is the most relevant to ours, since it conducted experiments to use human spatial attention to supervise model spatial attention, for three tasks (salient object segmentation, video action recognition and fine-grained image classification) and demonstrated that human spatial attention is beneficial. However, it still remains a question whether such benefits could be extended to contrastive representation learning. In + +terms of strategy to utilize spatial attention labels, in [32], attention labels are used as spatial weighting, while we designed auxiliary task to predict spatial attention labels. + +Teacher model pseudo-labeling: Previous work on knowledge distillation and machine self training has demonstrated that machine teaching machines approaches may address the challenge of labeling large datasets. In image classification, [49] demonstrated that training a model to classify images then use that model to provide pseudolabels improved classification performance. Related idea is applied in [47] for language models, which showed that the knowledge learned by language models pretrained on text corpus could be distilled to generate new datasets for common reasoning, and that training common reasoning models on this new data largely improve performance. Inspired by these successes, we train a teacher model on smaller human attention data and use this model to generate new spatial attention pseudo labels for ImageNet benchmark (see Figure 1 and Figure 2). + +# 3. Methods + +# 3.1. The Contrastive Learning Framework + +Contrastive learning is one of the most popular self-supervised representation learning methods. It learns an embedding space so that similar data (positive) pairs are mapped to be close in the embedding space and different (negative) pairs are mapped to be far away. In practice, positive pairs are often generated by applying data augmentation to one image like adding noises, cropping, etc and negative pairs are different examples in the mini-batch. Among contrastive learning methods, SimCLR framework [6] has shown solid performance, which we choose here as our main method. For each batch, the images are augmented in two different ways. Then we feed them into ResNet feature extractor backbone and compute NT-Xent loss to minimize the difference between augmentations of the same images and maximize the differences between different images. NT-Xent loss can be computed as follows [6]: + +$$ +l _ {i, j} = - \log \frac {\exp \left(\operatorname {s i m} \left(z _ {i} , z _ {j}\right) / \tau\right)}{\sum_ {k = 1} ^ {2 N} 1 _ {k \neq i} \exp \left(\operatorname {s i m} \left(z _ {i} , z _ {k}\right) / \tau\right)} \tag {1} +$$ + +where $z_{i}$ is the embedding of ith example, and $\operatorname{sim}(\cdot ,\cdot)$ is the cosine similarity between two embeddings. + +# 3.2. ImageNet-Attn: spatial attention maps for ImageNet generated by a teacher model + +To train contrastive learning models, we need large datasets with millions of examples. Thus, to address the question whether human attention is beneficial for contrastive models, we will need large human spatial attention labeled dataset. However, there is no spatial attention dataset available with that size on typical large image + +![](images/93fa50206b0d3cbe786cbf9ad40bcec24b5a6cbb038725bcb734ba052fc9a579.jpg) +a +Teacher model +b +C + +![](images/5ea18cfcc421aaec73cd1885287a1fb4d55a91c07e8270484dcf100a13a57fa7.jpg) +Sallcon dataset image + +![](images/a3768fe338159b9ab323fb7fe4c83cbd97773ba9f8b9a024ed1bf2e6d7cffa08.jpg) + +![](images/9b0c07aaea506276e13a400e8b827692afdae1d147790c8df966c239e5410b92.jpg) +Figure 2. (a) The architecture of teacher model to predict attention. The teacher model was trained from random initialization to predict ground truth (G.T.) human spatial attention labels. (b) Examples from Salicon dataset, with G.T. human attention heatmaps, and teacher model predictions. (c) Examples of spatial-attention pseudo-labels generated by the teacher model for ImageNet dataset (ImageNet-Attn). + +![](images/1755ab91f9f4189d4abeb98133b473ee9329c5f854e24ff4137cd0426bd3cc5f.jpg) +ground truth human spatial attention + +![](images/9f3181742b530666673515591cb612661a6d81ae096a0b81856c8e1889370377.jpg) + +![](images/c8c95a5c3463ab36352c93dbf1d1a7eeb299c849967865c5a2914ea986749c88.jpg) + +![](images/89d78b025c249a28fae17397f08ca1b33aa52c7b1632f052736d60eac8fdab9c.jpg) +teacher model predictions + +![](images/574387cf417c34711f8395c2b0906971b0b95162958167806d02b86f105810dc.jpg) + +![](images/71cafe5f403f4e8fd4d126bcef999bb009c1d4417a991953e2004132f5c78329.jpg) + +![](images/e0574f871272e15f4f1fd1a2d12a754c2afe77495dbe90ec41c3172de78c207d.jpg) +ImageNet dataset image + +![](images/376c356029e8c7641a85a9d170af3ed4b7726fa1f828194ba10f67a263af4f35.jpg) + +![](images/66963753b7c8fb7087a4029952f3b08637b4dc57463f9eddc3fc8ddaf91f8380.jpg) + +![](images/f8907546e08968dbf24fe5c0b0f03b1acb32454e82b1f8c7d686f3e42a3ee3bd.jpg) +spatial-attention labels from the teacher model + +![](images/d4d20847130eeb8ac6624910389ad3c7e00051a253297e68fdf56d3f33f2478a.jpg) + +![](images/c9d2174013c7ecfccbb3d60ba276f079aff4be0b4285359ef7f8fe2541a317fc.jpg) + +benchmark such as ImageNet. The largest attention data set available is Silicon [25], consisting of twenty thousand examples only, orders of magnitude smaller than ImageNet, which has around 14 million examples. + +To generate attention labels for ImageNet images, we train a teacher model on Silicon attention data set to predict human attention ground truth labels (see Figure 2b). Then we use this teacher model to create a new dataset (ImageNet-Attn) with spatial attention maps for ImageNet (see Figure 2c for some examples). The teacher model architecture is illustrated in Figure 2a. Specifically, we used MobileNet-V3-small as the backbone, and embeddings from 4 layers (conv 2,4,6,8) are extracted. On each embedding, we applied two conv layers (the first conv layer with $3 \times 3$ kernel, number of channels matching the input embedding, max pooling of 3 and relu; the second conv layer with $1 \times 1$ kernel, 1 channel and relu). We bilinearly resized the output to the same resolution of input. The + +result of the 4 branches are then summed to yield 1 feature, followed up by a sigmoid function to obtain spatial-attention map. Note that using multiple intermediate (both early and late) layers is one of the key ingredient to make the spatial attention teacher model successful. This strategy mimics human visual attention, which is known to be affected by both low-level characteristics such as color, intensity, texture, as well as high level characteristics like shape and object, etc. Essentially, our teacher model has a similar architecture as in [36], but is further simplified with more efficient backbone, less channels/layers, etc, so that it can be trained from scratch on Silicon data with randomly initialized backbone to avoid any leak of class label information, while most existing attention prediction models [1,36] needs to finetune with a pretrained classification backbone network. + +# 3.3. Training contrastive models with spatial attention maps + +Spatial Attention Branch The overall training framework is shown in Figure 1. This spatial attention prediction framework consists of two branches: the contrastive branch and the spatial attention branch. The contrastive branch is the same as the original SimCLR method [6], which applies augmentations to image $x$ to get different variants $x_{i}$ and $x_{j}$ , and learns the representation $h_{i}$ and $h_{j}$ via a feature extractor backbone network (e.g., ResNet), then use a projection head to map $h_{i} / h_{j}$ to $z_{i} / z_{j}$ , where the contrastive loss is applied. + +For the spatial-attention branch, it takes as an input not only the final embedding $h$ , but also early intermediate layer embeddings (following the same reason of including both high level and low level visual cues as in teacher model), and predicts an spatial attention heatmap $m_p$ . More specifically, we apply a global average pooling on the output of the last three blocks of the ResNet backbone. Then, we select the max channel for each of three block output (after average pooling), and resize with bilinear interpolation to the image resolution, which has some resemblance to CAM approaches [51]. Finally we stack the representations together, pass them into a linear readout layer (with a bias term), and use the output as our final spatial attention prediction $m_p$ . We keep the spatial attention head as simple as possible so that the guide on attention head output can be back propagated to representation more directly. + +We use the ImageNet-Attn data generated by the teacher model as target for each ImageNet example, denoted as $m_t$ , we can then train the network spatial attention output $m_p$ to be close to $m_t$ . We hypothesize that this method regularizes the training of the feature extractor backbone rather than explicitly enforce the network to generate masked representations that match the spatial attention maps. Note that for attention branch, there is no augmentation applied to each image $x$ , since human attention is not invariant to transformation (e.g., a human looking at a cropped image may attend to different regions compared to a consistent crop of human attention map of the original image). + +Loss function The loss function $L$ consists of two terms: + +$$ +L = L _ {\text {c o n t r a}} + L _ {\text {a t t n}} \tag {2} +$$ + +$L_{\text{contra}}$ is the contrastive loss, or more specifically $L_{\text{contra}} = \sum_{i,j} l_{i,j}$ and $l_{i,j}$ is defined in Equation (1), the same as in [6]. $L_{\text{attr}}$ is the attention regularization loss, and more specifically + +$$ +L _ {a t t n} = \sum_ {i} \left(\lambda K L D \left(m _ {i} ^ {p}, m _ {i} ^ {t}\right) - \beta N S S \left(m _ {i} ^ {p}, p _ {i} ^ {t}\right)\right) \tag {3} +$$ + +where $\lambda$ and $\beta$ are two weighting parameters ( $\lambda = 1.0$ and $\beta = 0.1$ in our work). $m_{i}^{p}$ is the predicted 2D spatial-attention map from the attention head of our proposed + +Table 1. ImageNet Top-1 classification accuracy for different models (mean $\pm$ SE for 3 seeds, except for * which means best result from all runs). + +
ModelAccuracy (%)
Contrastive67.61 ± 0.04
Contrastive attn. teacher68.23 ± 0.08
Contrastive attn. co-train66.35 ± 0.12
Contrastive attn. with explicit attention block [32]61.15*
Contrastive attn. with self attention mask [32]49.65*
Contrastive attn. with ContraCAM [35] as attention67.71*
Supervised75.91 ± 0.10
Supervised attn. teacher76.02 ± 0.04
Supervised (ResNet-18)69.17 ± 0.07
Supervised (ResNet-18) attn. teacher69.30 ± 0.04
+ +model for $i$ -th example (original image, not augmented ones), $m_i^t$ is the pseudo spatial-attention target map predicted by the teacher model for $i$ -th example, and $KLD()$ is the KL divergence1. Besides KLD loss, we also use $NSS()$ , the Normalized Scanpath Saliency loss [4]. Those two losses are typically used for human attention predictions [3,9]. KLD is often used to match the spatial-attention target distribution, and NSS is typically added on top as it is generally observed to help generate attention maps that are in perceptual agreement with human judgements (see [3,34]). For $NSS$ loss, the larger the better, so there is a negative sign before it. $NSS$ loss needs gaze/attention points instead of heatmap as ground-truth, so we extract pseudo gaze/attention points $p_i^t$ from the attention heatmap $m_i^t$ . To obtain $p_i^t$ , we first extract the point with highest value in current spatial-attention map $m_i^t$ , then generate a new spatial-attention map by subtracting a Gaussian blur around the extracted point from the current attention map. The process is repeated with the new attention map until the maximal value of the attention map is smaller than a threshold (more specifically, $0.2 * \max\_ sal$ , where $\max\_ sal$ is the maximal value in $m_i^t$ ).2 + +# 4. Experiments + +# 4.1. Implementation details + +In our experiments, in Eq.1, we choose $\tau = 0.1$ and $N = 2048$ . In Eq.2, we used equal weighting of 1.0 for the SimCLR and attention losses. + +The weight for KLD and NSS losses are 1 and 0.1, re + +![](images/32d7e5374b151f79b5cb8a57482cedc8232264f0992237cd7e8a14ad84159635.jpg) +a +image + +![](images/2eecca999a2c947038b270bef71548eb6649dea9201ac5ae400bb29351fe1285.jpg) +ground truth human spatial attention + +![](images/846b7e8c3341bff98bef74217b447a2b18dbaa29226df152eace79dcd4cf4f5d.jpg) +teacher model spatial attention + +![](images/539fbb18ff73cf7649d2a6653796aee1e0b4187cf05d2c7128a02397947ec962.jpg) +spatial attention guided contrastive model + +![](images/eaf653822539f8748e0132a7dbe5e18f56f00e4b46a529919778f1287b405ded.jpg) +baseline trastive model + +![](images/a945762d74bebc49c3a6083f0a783e9040a178ad6e3ef8170cb3d7f826e23f03.jpg) + +![](images/45d66b37c81c7ea81df24645f90460ad72fb3484b062276c1688f843f84515ba.jpg) + +![](images/642d6b9bdc2ab1f0f8f178624888629995f74415b2b5995cf545f63ed0acf79c.jpg) + +![](images/736137c06046f970352bf1a27a8aeec06985b2ca244f1d79af159d04e98105e3.jpg) + +![](images/e69720b1fcb4eb30c868b50bfacfe9a8090aed7f1b35fe46931c49b39c4087ad.jpg) + +![](images/0d082c4709c76c8911510e6ae8101c69d06a1b24a93c94a76c433f9fdf77c82d.jpg) + +![](images/2f9463b54bfb9044d1eab9c31c10c71d034fb0cd5baab8747b43241ba864a9bf.jpg) + +![](images/1be15bb1b92d287fe25227d4a63aeeb56c5377dbccf79ffe32f0e94756e900bc.jpg) + +![](images/d7b688f628f4f15ccfaf1b70a84c4b737b12d46893e61d4698ed198b008255a1.jpg) + +![](images/c85bff58326ab2ea0aa7f50d5c220a869fbc3fa580de36be75793ade749e5826.jpg) + +![](images/1c3ddacaa39d16ca0e6be2230e0bc9b0daf621bf30c33814a8bc83f594917518.jpg) +b + +![](images/8b3c64989230165649d08fbd2ea5c4944425e7a4354d50107fa8d2d7ecf74867.jpg) +correlation coeff. (model vs human) +Figure 3. (a) Examples comparing spatial attention maps predicted by different models vs ground truth human attention data on OSIE dataset [43]. (b) Distribution of correlation coefficients between model predicted attention maps vs ground truth human attention maps on Silicon [25] validation set (left) and OSIE dataset (right) [43]. Three models are shown: (1) Teacher model trained on Silicon [25] training set. (2) Baseline SimCLR model (3) The proposed spatial attention guided SimCLR model trained with pseudo-attention labels provided by the teacher model. Please note that both baseline and our model are trained to predict target attention labels, with main difference that the baseline model has a stop gradient placed between the prediction layer and network features to prevent spatial-attention targets to inform network features. Also note that our teacher model is trained on Silicon where the attention is collected by mouse tracking, while the results in (a) and (c) here are on OSIE data, where the attention are collected by eye trackers. The results demonstrated both the teacher model and the attention head in proposed model has a good generalization capacity for new kinds of attention labels. + +spectively. We obtained those weights by hyperparameter search from 0.1-5 for KLD and 0.01-1 for NSS, based on ImageNet accuracy of a separate dev set (80%-20% split on training set for parameter search with 20% as dev set). + +# 4.2. Spatial attention guided models are highly predictive of human attention + +In this section, we explore whether the use of auxiliary teacher model to provide spatial attention pseudo labels on ImageNet better aligns contrastive model's attention with human attention. We define aligning model spatial attention here as the ability to predict spatial attention mask from + +the model backbone features by a simple readout layer discussed in Section 3. We trained two ResNet-50 backbones using the SimCLR objective from [6]. We added additional spatial-attention losses as discussed in Section 3. For the baseline model, we placed a stop gradient operation between the backbone features and the attention projection head to prevent attention information from leaking to the backbone features, whereas for the attention guided model, we allowed the learned attention gradients to flow back to the backbone. + +We evaluated the degree the predicted attention maps is aligned with human attention by performing correlation analysis, which is typically used to compare human and model spatial attention. We measured Pearson's correlation coefficient between the model predicted attention and ground truth human spatial-attention maps from Silicon [25] validation set (Note that the teacher model is trained with Silicon training set). We find that both teacher model and spatial attention guided contrastive model are highly correlated with human attention (Figure 3b), while the baseline contrastive model is much less correlated. + +To test whether this alignment with human attention is general, beyond just Salicon data, we obtained another attention dataset on OSIE images [43] (Figure 3a). This data more faithfully reflects human attention as it collected from (mobile) eye tracker [43]. We find that the baseline model is weakly positively correlated with human attention (ttest: $\rho = 0.07$ $p < 0.001$ ) suggesting that the contrastive loss produces features that are predictive of human attention to some extent. Yet, the correlation was generally close to 0 and explains only $0.5\%$ of data variance. The spatial attention guided model has a much stronger correlation to human attention (ttest: $\rho = 0.48$ $p < 0.001$ ) than the baseline model (Two samples ttest: $p < 0.001$ ), and thus more faithfully reflecting human visual attention (See Fig 3a for qualitative examples and Fig 3b bottom for quantitative analysis). Further, the correlation with human attention for the teacher model and the attention guided contrastive model were quite similar (Two samples ttest: $p = 0.7$ ). + +# 4.3. Spatial attention guided models are more accurate than baselines + +We evaluate the quality of the representations learned by spatial attention guidance framework using the typical contrastive learning evaluation criteria: fitting an ImageNet [37] linear classifier on top of the frozen representation (in practice we place stop gradient at the end of the backbone and train the classifier concurrently while training the backbone). We compute Top 1 accuracy on ImageNet validation set and compare the results with baselines. As shown in Table 1, we observe around $0.6\%$ accuracy gain on ImageNet compared to vanilla SimCLR. We further explore an alternative way of incorporating human attention data. + +Rather than using pseudo attention labels on ImageNet from the teacher model, we add Silicon data to the training data, and directly predict attention ground truth labels for Silicon data with spatial attention head (Contrastive attn. co-train in Table 1). More specifically, we pass 2048 ImageNet images and N=2048 Silicon images into our backbone (N is obtained by hyperparameter search 512-2048 based on evaluation on the separate dev set) to compute backbone features. Then, we compute the SimCLR loss from ImageNet features and use Silicon images features to predict spatial attention maps, and we use Silicon attention ground truth labels to compute the KLD and NSS losses. Interestingly, we find this method to lead to worse performance compared to using the teacher model generated spatial-attention labels, which gives evidence that the teacher model is generalizing its knowledge about human attention data beyond the limited Silicon training data. + +One important question is whether the classic trainable attention method applies to contrastive learning. For example, in [32], an attention block/map is used to adjust the weights of the embedding, where the attention block can be self learned implicitly, or guided by human attention ground-truth. It shows both implicit self-learned attention block and attention block guided by human attention can improve the accuracy of fine-grained classification (Table VII in [32]). We adopted the approach described in their paper: add an attention block after the representation embedding, and use the attention block's output to mask the representation embedding. We apply it to the contrastive models, and the results are shown in Table 1 as "explicit attention mask" (the attention block is guided by human attention) and "self attention mask" (self attention). As observed, both yield poor performance, since the contrastive attention with self-attention mask method learns an attention mask from the contrastive loss rather than to predict the teacher targets, the model may be getting weaker supervision. + +Another important question is whether our current spatial attention targets can be replaced by other spatial attention approach like ContraCAM [35]. In our method, we predicted the spatial attention with a linear layer. Whereas in ContraCAM a contrastive loss is used to construct the attention prediction. In both cases, we trained the predictions to match the target attention maps obtained from the teacher model. The result is shown in Table 1 as "Contrastive with ContraCAM as attention". However, we don't see much accuracy improvements with ContraCAM as the model attention, showing that the linear simple prediction leads to better performance. + +To investigate whether the spatial attention guidance framework benefits supervised models in the same way as contrastive models, we applied the same approach for supervised models. Supervised models similarly benefit from + +this framework, yet the gain is limited compared to the contrastive models perhaps due to the higher accuracy the supervised models achieves compared to the contrastive models. To control for accuracy, we used a supervised model trained with smaller backbone (ResNet-18), which gives comparable accuracy to the bigger ResNet-50 backbone that is trained with contrastive objective. Even when controlling the mismatch in accuracy, the supervised model gain from providing spatial-attention supervision is limited compared to contrastive models (Table 1). + +# 4.4. Spatial attention guided models are more robust than baselines + +Human vision is very robust to noise or small pixel modulations, compared to computer vision models. Here, we hypothesize that models that aligns better with human attention may learn more robust representations. We test this hypothesis using ImageNet-C dataset [18]. We take the representations learned from the proposed model and train a linear classifier on ImageNet training data. We then evaluate classification performance on ImageNet-C at various corruption/noise types. Table Supp.1 in the supplementary compares the classification performance on ImageNet-C for baseline contrastive model and the proposed models trained with spatial-attention teacher guidance, average over the 5 different corruption/noise magnitudes in ImageNet-C. We find that contrastive models trained with teacher guidance outperforms baseline consistently, suggesting that the representation learned by spatial-attention guided model is indeed more robust. + +# 4.5. Spatial attention guided models generate better representation for retrieval + +Besides classification, we tested the quality of the representations for another downstream task: image retrieval. We use the model to extract representation for ImageNet validation set, and use the representations to run image retrieval. 5000 randomly chosen validation images are used as query, while the rest validation images are database images. For each query image, top k retrieval results are returned by sorting the cosine distance between representations of database images and the query. The retrieval accuracy is measured with the standard precision recall curve [40], where precision and recall are computed by checking whether results have the same class label as the query. Mean precision and recall is obtained as the average across queries. We choose different k values, and draw the mean precision-recall curve. Besides ImageNet ("clean" in Fig Supp.1a in the supplementary), we also use ImageNet-C ("noise level 1-5"). Results from "fog" corruption is shown in In Figure Supp.1a in the supplementary., while results for other corruptions can be found in the supplementary too. The representation extracted with + +the Contrastive attn teacher model outperforms the baseline Constractive model in this retrieval task, on both ImageNet images, and ImageNet-C images for most corruption types and levels. + +Moreover, we evaluate transfer to other dataset for the retrieval task. We take the model trained on ImageNet to extract representations for CIFAR-10 [29] test set and CIFAR-10-C [18] ("clean" and "noise level 1-5" in Fig Supp.1b in the supplementary respectively), without retraining/finetuning. 1000 images are used as query and the rest are used as database. Results for "fog" noise are shown in Fig Supp.1b in the supplementary, while results for other noises are shown in supplementary too. As shown, the proposed attention guided model outperforms the baseline model on both clean images and almost all noise/corruption types and levels. + +# 5. Conclusion + +In this work, we tested the hypothesis that utilizing human spatial attention can be beneficial for obtaining better representation for contrastive models. We overcome the challenge of obtaining human spatial attention labels for large dataset by utilizing a teacher model trained on limited human attention labels to provide pseudo-attention labels for ImageNet. We augmented training of a commonly-used network (ResNet50) trained with SimCLR contrastive objective with pseudo-spatial attention labels from the teacher model. Our results demonstrate that contrastive models trained with those pseudo-attention labels are more predictive of human attention and we obtain better representations. + +Despite the gains observed on downstream tasks are not large, the gains were consistent across 3 tasks (classification, robustness, and image retrieval). Thus, taken together, our findings support the above hypothesis. The limited gains in classification may suggest that spatial attention alone may not be enough to achieve SOTA, but perhaps in conjunction with other architectural and methodological improvements. Another limitation is that our exploration was conducted on natural images only, but other interesting applications may include other domains like medical images or autonomous driving. When domains become very different, it may be required to retrain teacher models to capture those new domains (e.g., train on attention data that comes from medical experts for medical images). These are interesting and important questions that may be addressed in future work. + +# 6. Acknowledgement + +We are grateful to Yifan Tian, Josh McAdams, Zhuode Liu, and Ethan Steinberg for assistance with analyses. We also thank Mike Mozer, Kai Kohlhoff, Simon Kornblith, and Ting Chen for useful discussions. + +# References + +[1] Ali Borji. Saliency prediction in the deep learning era: Successes and limitations. IEEE transactions on pattern analysis and machine intelligence, 2019. 3, 4 +[2] Zoya Bylinskii, Tilke Judd, Ali Borji, Laurent Itti, Frédo Durand, Aude Oliva, and Antonio Torralba. Mit saliency benchmark. http://saliency.mit.edu, 2012. 2, 3 +[3] Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, and F. Durand. What do different evaluation metrics tell us about saliency models? IEEE TPAMI, 41(3):740-757, 2018. 5 +[4] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5 +[5] Jorge Paolo Casas and Chandramouli Chandrasekaran. openeyetrack-a high speed multi-threaded eye tracker for head-fixed applications. Journal of open source software, 4(42), 2019. 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 5, 7 +[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2 +[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an lstmbased saliency attentive model. IEEE TIP, 27(10):5142- 5154, 2018. 5 +[10] Abhishek Das, Harsh Agrawal, Larry Zitnick, Devi Parikh, and Dhruv Batra. Human attention in visual question answering: Do humans and deep networks look at the same regions? Computer Vision and Image Understanding, 163:90-100, 2017. 3 +[11] Mohammad K Ebrahimpour, J Ben Falandays, Samuel Spevack, and David C Noelle. Do humans look where deep convolutional neural networks "attend"? In International Symposium on Visual Computing, pages 53-65. Springer, 2019. 3 +[12] Simone Frintrop, Erich Rome, and Henrik I Christensen. Computational visual attention systems and their cognitive foundations: A survey. ACM Transactions on Applied Perception (TAP), 7(1):1-39, 2010. 2 +[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1 +[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach + +to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 2 +[15] Meng-Hao Guo, Tian-Xing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph R Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational Visual Media, pages 1-38, 2022. 3 +[16] Suna Sihang Guo, Ruohan Zhang, Bo Liu, Yifeng Zhu, Dana Ballard, Mary Hayhoe, and Peter Stone. Machine versus human attention in deep reinforcement learning tasks. Advances in Neural Information Processing Systems, 34:25370-25385, 2021. 3 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 8 +[19] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 262-270, 2015. 2, 3 +[20] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 1 +[21] Laurent Itti and Christof Koch. Computational modelling of visual attention. Nature reviews neuroscience, 2(3):194-203, 2001. 3 +[22] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE TPAMI, 20(11):1254-1259, 1998. 5 +[23] Saumya Jetley, Nicholas A Lord, Namhoon Lee, and Philip HS Torr. Learn to pay attention. arXiv preprint arXiv:1804.02391, 2018. 3 +[24] Sen Jia and Neil DB Bruce. Eml-net: An expandable multilayer network for saliency prediction. Image and Vision Computing, 95:103887, 2020. 3 +[25] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Salicon: Saliency in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1072-1080, 2015. 2, 3, 4, 6, 7 +[26] T. Judd, F. Durand, and A. Torralba. A benchmark of computational models of saliency to predict human fixations. 2012. 5 +[27] Nam Wook Kim, Zoya Bylinskii, Michelle A Borkin, Krzysztof Z Gajos, Aude Oliva, Fredo Durand, and Hanspeter Pfister. Bubbleview: an interface for crowdsourcing image importance maps and tracking visual attention. ACM Transactions on Computer-Human Interaction (TOCHI), 24(5):1-40, 2017. 2 +[28] Christof Koch and Shimon Ullman. Shifts in selective visual attention: towards the underlying neural circuitry. In Matters of intelligence, pages 115-141. Springer, 1987. 3 + +[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8 +[30] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. arXiv preprint arXiv:1411.1045, 2014. 3 +[31] Matthias Kummerer, Thomas SA Wallis, Leon A Gatys, and Matthias Bethge. Understanding low-and high-level contributions to fixation prediction. In Proceedings of the IEEE International Conference on Computer Vision, pages 4789-4798, 2017. 3 +[32] Qixia Lai, Salman Khan, Yongwei Nie, Hanqiu Sun, Jianbing Shen, and Ling Shao. Understanding more about human and machine attention in deep neural networks. IEEE Transactions on Multimedia, 23:2086-2099, 2020. 2, 3, 5, 7 +[33] Thomas Langlois, Haicheng Zhao, Erin Grant, Ishita Dasgupta, Tom Griffiths, and Nori Jacoby. Passive attention in artificial neural networks predicts human visual selectivity. Advances in Neural Information Processing Systems, 34, 2021. 3 +[34] J. Lou, H. Lin, D. Marshall, D. Saupe, and H. Liu. Transal-net: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 5 +[35] Sangwoo Mo, Hyunwoo Kang, Kihyuk Sohn, Chun-Liang Li, and Jinwoo Shin. Object-aware contrastive learning for debiased scene representation. Advances in Neural Information Processing Systems, 34, 2021. 3, 5, 7 +[36] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10241-10247. IEEE, 2020. 3, 4 +[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 7 +[38] Ramprasaath R Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. Casting your model: Learning to localize improves self-supervised representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11058-11067, 2021. 3 +[39] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a hint: Leveraging explanations to make vision and language models more grounded. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2591-2600, 2019. 3 +[40] Nikhil V Shirahatti and Kobus Barnard. Evaluating image retrieval. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 1, pages 955-961. IEEE, 2005. 8 +[41] Ekta Sood, Fabian Kögel, Florian Strohm, Prajit Dhar, and Andreas Bulling. Vqa-mhug: A gaze dataset to study multimodal neural attention in visual question answering. arXiv preprint arXiv:2109.13116, 2021. 3 +[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. + +Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[43] Nachiappan Valliappan, Na Dai, Ethan Steinberg, Junfeng He, Kantwon Rogers, Venky Ramachandran, Pingmei Xu, Mina Shojaeizadeh, Li Guo, Kai Kohlhoff, et al. Accelerating eye movement research via accurate and affordable smartphone eye tracking. Nature communications, 11(1):1-12, 2020. 2, 6, 7 +[44] Leonard Elia Van Dyck, Roland Kwitt, Sebastian Jochen Denzler, and Walter Roland Gruber. Comparing object recognition in humans and deep convolutional neural networks—an eye tracking study. Frontiers in Neuroscience, page 1326, 2021. 3 +[45] Brian A Wandell. Foundations of vision. Sinauer Associates, 1995. 2 +[46] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2017. 3 +[47] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2, 3 +[48] Jeremy M Wolfe and Todd S Horowitz. What attributes guide the deployment of visual attention and how do they do it? Nature reviews neuroscience, 5(6):495-501, 2004. 2 +[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2, 3 +[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2 +[51] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 5 +[52] Jan Zimmermann, Yuriria Vazquez, Paul W Glimcher, Bijan Pesaran, and Kenway Louie. Oculomatic: high speed, reliable, and accurate open-source eye tracking for humans and non-human primates. Journal of neuroscience methods, 270:138-146, 2016. 2 \ No newline at end of file diff --git a/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/images.zip b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03be3848de6832b4dbfe7ed3953d74929d884c7a --- /dev/null +++ b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9bf303334d08b34d1ac5be1bffe7bf1e605391d99bb16d836a275bbd137fa489 +size 337062 diff --git a/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/layout.json b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d5650ca7a0aecfb1e6a089ade62a4e70004e0c23 --- /dev/null +++ b/2023/Teacher-Generated Spatial-Attention Labels Boost Robustness and Accuracy of Contrastive Models/layout.json @@ -0,0 +1,8676 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 531, + 138 + ], + "type": "text", + "content": "Teacher-generated spatial-attention labels boost robustness and accuracy of contrastive models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 161, + 128, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 161, + 128, + 174 + ], + "spans": [ + { + "bbox": [ + 72, + 161, + 128, + 174 + ], + "type": "text", + "content": "Yushi Yao*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 176, + 120, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 176, + 120, + 189 + ], + "spans": [ + { + "bbox": [ + 81, + 176, + 120, + 189 + ], + "type": "text", + "content": "Waymo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 192, + 144, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 192, + 144, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 192, + 144, + 202 + ], + "type": "text", + "content": "yushiy@waymo.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 188, + 161, + 242, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 161, + 242, + 175 + ], + "spans": [ + { + "bbox": [ + 188, + 161, + 242, + 175 + ], + "type": "text", + "content": "Chang Ye*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 176, + 235, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 176, + 235, + 190 + ], + "spans": [ + { + "bbox": [ + 197, + 176, + 235, + 190 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 192, + 265, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 192, + 265, + 203 + ], + "spans": [ + { + "bbox": [ + 167, + 192, + 265, + 203 + ], + "type": "text", + "content": "yechang@google.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 161, + 372, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 161, + 372, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 161, + 372, + 175 + ], + "type": "text", + "content": "Junfeng He†" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 176, + 361, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 176, + 361, + 190 + ], + "spans": [ + { + "bbox": [ + 324, + 176, + 361, + 190 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 288, + 192, + 397, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 192, + 397, + 203 + ], + "spans": [ + { + "bbox": [ + 288, + 192, + 397, + 203 + ], + "type": "text", + "content": "junfenghe@google.com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 421, + 161, + 535, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 161, + 535, + 175 + ], + "spans": [ + { + "bbox": [ + 421, + 161, + 535, + 175 + ], + "type": "text", + "content": "Gamaleldin F. Elsayed†" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 460, + 176, + 497, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 460, + 176, + 497, + 190 + ], + "spans": [ + { + "bbox": [ + 460, + 176, + 497, + 190 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 419, + 192, + 534, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 192, + 534, + 203 + ], + "spans": [ + { + "bbox": [ + 419, + 192, + 534, + 203 + ], + "type": "text", + "content": "gamaleldin@google.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 191, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 295, + 290, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 295, + 290, + 656 + ], + "spans": [ + { + "bbox": [ + 45, + 295, + 290, + 656 + ], + "type": "text", + "content": "Human spatial attention conveys information about the regions of visual scenes that are important for performing visual tasks. Prior work has shown that the information about human attention can be leveraged to benefit various supervised vision tasks. Might providing this weak form of supervision be useful for self-supervised representation learning? Addressing this question requires collecting large datasets with human attention labels. Yet, collecting such large scale data is very expensive. To address this challenge, we construct an auxiliary teacher model to predict human attention, trained on a relatively small labeled dataset. This teacher model allows us to generate image (pseudo) attention labels for ImageNet. We then train a model with a primary contrastive objective; to this standard configuration, we add a simple output head trained to predict the attention map for each image, guided by the pseudo labels from teacher model. We measure the quality of learned representations by evaluating classification performance from the frozen learned embeddings as well as performance on image retrieval tasks (see supplementary material). We find that the spatial-attention maps predicted from the contrastive model trained with teacher guidance aligns better with human attention compared to vanilla contrastive models. Moreover, we find that our approach improves classification accuracy and robustness of the contrastive models on ImageNet and ImageNet-C. Further, we find that model representations become more useful for image retrieval task as measured by precision-recall performance on ImageNet, ImageNet-C, CIFAR10, and CIFAR10-C datasets." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 535, + 413 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 535, + 413 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 535, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 535, + 413 + ], + "type": "image", + "image_path": "322434ca0869a5340264ba1200a146318688434775e26bb4924014b017ec42fb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 418, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 547, + 462 + ], + "type": "text", + "content": "Figure 1. Illustration. A teacher model is trained to predict human spatial-attention from a small dataset. Then the model is used to provide attention labels for larger dataset, which are used as additional targets for contrastive models." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 468, + 387, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 387, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 387, + 481 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 489, + 546, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 546, + 632 + ], + "type": "text", + "content": "Deep learning models have made significant progress and obtained notable success on various vision tasks. Despite these promising results, humans continue to perform better than deep learning models in many applications. A notable reason is that deep learning models have a tendency to learn \"short-cuts\", i.e., giving significance to physically meaningless patterns or exploiting features which are predictive in some settings, but not causal [20]. Examples include focusing on less significant features such as background and textures [13]. These models yield representations that are less generalizable and lead to models that are highly sensitive to small pixel modulations [42]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 632, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 547, + 668 + ], + "type": "text", + "content": "Human vision on the other hand is known to be much more robust and generalizable. One major difference between human and machine vision is that humans tend to" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 674, + 406, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 674, + 406, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 674, + 406, + 684 + ], + "type": "text", + "content": "*Equal technical contribution." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 684, + 449, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 684, + 449, + 694 + ], + "spans": [ + { + "bbox": [ + 306, + 684, + 449, + 694 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 306, + 684, + 449, + 694 + ], + "type": "text", + "content": " Equal leadership and advising contribution" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 694, + 371, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 694, + 371, + 703 + ], + "spans": [ + { + "bbox": [ + 306, + 694, + 371, + 703 + ], + "type": "text", + "content": "Correspondence to:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 704, + 515, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 704, + 515, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 704, + 515, + 713 + ], + "type": "text", + "content": "junfenghe@google.com & gamaleldin@google.com" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23282" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "focus on specific regions in visual scene [45]. These locations often reflect regions salient or useful to perform a specific vision task. Machines, instead, initially place equal significance to all regions. A natural question is: will it be beneficial if machine vision models is guided by human spatial attention?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 312 + ], + "type": "text", + "content": "Human spatial attention has been shown to benefit computer vision models in supervised tasks, such as classification [32]. Yet, it is still a question whether adding a form of weak supervision in the form of human spatial attention could similarly benefit self-supervised models that are trained end-to-end. Self-supervised models typically need a large amount of data to yield good representations. To test if training weakly supervised models with human spatial attention cues, we will need to collect a large volume of human spatial attention labels, which is a very expensive process that requires either using trackers to record eye movements [5,43,52] or asking humans to highlight regions that they attend to [25, 27]. This process is prohibitively tedious and costly for datasets with millions of examples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 587 + ], + "type": "text", + "content": "In this work, we test the hypothesis that a weak supervision in the form of human spatial attention is beneficial for representation learning for models trained with a contrastive objective. Inspired by knowledge distillation and self-training using teacher models [47, 49], we address the challenge of obtaining spatial attention labels on large scale image datasets by using machine pseudo-labeling. We train a teacher model on a set of limited ground truth human spatial attention labels, and use this teacher model to generate spatial attention pseudo-labels for the large ImageNet benchmark. We are then able to utilize the generated spatial attention maps in the contrastive models, and discover that this approach yields representations that are highly predictive of human spatial attention. Further, we find that the learned representations are better as measured by higher accuracy and robustness on classification downstream tasks, and higher precision and recall on image retrieval tasks. Interestingly, we find that the gains from using teacher models to provide pseudo labels are larger than using the limited ground truth human labels directly when training contrastive models, and the gains are larger for contrastive models than when applying same method to supervised models. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 594, + 287, + 686 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 57, + 594, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 594, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 594, + 287, + 655 + ], + "type": "text", + "content": "- We create a dataset with spatial attention maps for the ImageNet [37] benchmark by first training a teacher model to predict human spatial attention labels from Silicon dataset [25] and then use the model to label ImageNet examples" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 662, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 662, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 57, + 662, + 287, + 686 + ], + "type": "text", + "content": "- We use spatial-attention labels from the teacher model as an additional prediction target to models trained" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 433, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 433, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 433, + 84 + ], + "type": "text", + "content": "with contrastive objective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 94, + 545, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 94, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 316, + 94, + 545, + 166 + ], + "type": "text", + "content": "- We find that the proposed method can learn better representation, leading to better accuracy and robustness for downstream classification tasks (on ImageNet and ImageNet-C), and better performance on retrieval tasks (on ImageNet, ImageNet-C, CIFAR-10, and CIFAR10-C)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 178, + 390, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 390, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 390, + 190 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 533 + ], + "type": "text", + "content": "Contrastive learning: Contrastive learning has gained popularity in the past few years for self-supervised and semi-supervised representation learning. In general, contrastive learning aims to learn similar representations for similar data pairs and different representations for different pairs. SimCLR [6] utilized MLP projection heads and strong data augmentation for constructing similar pairs and have demonstrated great gains in image classification downstream tasks. To form the contrastive loss for each minibatch with N examples, the similar data pairs are constructed from two augmentations of the same image and different pairs from the other images within a batch, and then computing the NT-Xent loss. A different formulation is used in [50] by encouraging the empirical cross correlation of the representations of two versions of augmented mini-batch to be close to identity. In [17], it is further proposed to build large dictionaries for self-supervised learning (MOCO), and moreover in [7], better results are achieved on image classification and object detection tasks when combining advances from SimCLR and MOCO. Follow up work by [14] managed to obtain good performance without the need for dissimilar pairs by encouraging the representation of similar pairs across two versions of the network (trained network and exponential moving average version) to be similar. Further, in [8] it is shown that simple Siamese networks can still learn good representations without the need for dissimilar pairs, large batches or momentum encoders, etc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Human spatial attention: Human visual system has developed an attention mechanism that focuses on regions in the visual space that are of interest or highly informative to the vision task [12, 48]. Eye trackers are often used to collect human spatial attention [5, 43, 52]. Many gaze data sets [2] have been collected with these eye trackers. Besides eye trackers, human spatial attention data can also be collected via mouse tracking [25, 27], e.g., users see a blurry version of an image, then click on regions they want to see more clearly, mimicking human's peripheral vision based on neurophysiological studies [19, 27]. Both eye tracking and mouse tracking methods are very expensive, which limit the number of examples in those datasets. Due to the relatively lower cost of mouse tracking, one can often generate relatively larger attention data from this method than" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "text", + "content": "Trained teacher model is available at: https://github.com/google-research/google-research/tree/master/humanattention/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23283" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "eye tracking. For example, Silicon [25] dataset is one of the largest spatial attention datasets, contains around 20K images, each with attention labels from 50-60 participants, via a mouse tracking system, under free viewing setting. Yet, this data is still orders of magnitude smaller than those needed to train contrastive representation learning models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 266 + ], + "type": "text", + "content": "Spatial attention/saliency prediction models aim at predicting which areas in an image will be salient to human attention and attract eye fixation, usually with collected gaze/attention data as ground-truth. Early works in saliency prediction usually define saliency through a set of hand crafted features such as color difference, contrast, intensities, etc. [21, 28]. Recent works [1, 19, 24, 30, 31, 36] leverage the power of deep neural networks and are often trained/fine-tuned on large scale gaze data sets like Silicon [2, 25]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 288, + 567 + ], + "type": "text", + "content": "Spatial attention of computer vision models: Spatial attention in neural networks can be mainly categorized into post-hoc attention like class activation map (CAM) ([51]), and trainable attention (e.g., [15, 23, 46]). Post-hoc spatial attention methods have been proposed to estimate regions in the image that are important or give rise to model decisions, often for model interpretation. In supervised settings where classification labels are known, the simplest and most direct method is class activation map (CAM) [51]. CAM uses class labels to extract the feature map that is most informative about the true class of an image. Grad-CAM [39] generalizes the CAM to apply to any model with any downstream task. [38] proposed to use Grad-CAM to design augmentation policies in self-supervised learning to tackle weak performance in complex scene images with many objects. ContraCAM [35] applies Grad-CAM assuming downstream task of contrastive learning, thus allowing computing spatial attention maps with no class label supervision. In [35], it is proposed to utilize the spatial attention information learned from ContraCAM to design data augmentation strategies to discourage contextual and background biases in a scene. Unlike [38] and [35] that uses spatial attention to design augmentation policies, here we focus on an end-to-end framework to predict spatial attention targets instead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Recently, there are also several papers [10, 11, 16, 33, 41, 44]) exploring the similarity/difference of model spatial attention vs human spatial attention, for the task of VQA (visual question answering) [10], object detection [11], reinforcement learning [16], etc. Among existing works, [32] is the most relevant to ours, since it conducted experiments to use human spatial attention to supervise model spatial attention, for three tasks (salient object segmentation, video action recognition and fine-grained image classification) and demonstrated that human spatial attention is beneficial. However, it still remains a question whether such benefits could be extended to contrastive representation learning. In" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "terms of strategy to utilize spatial attention labels, in [32], attention labels are used as spatial weighting, while we designed auxiliary task to predict spatial attention labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 108, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 300 + ], + "type": "text", + "content": "Teacher model pseudo-labeling: Previous work on knowledge distillation and machine self training has demonstrated that machine teaching machines approaches may address the challenge of labeling large datasets. In image classification, [49] demonstrated that training a model to classify images then use that model to provide pseudolabels improved classification performance. Related idea is applied in [47] for language models, which showed that the knowledge learned by language models pretrained on text corpus could be distilled to generate new datasets for common reasoning, and that training common reasoning models on this new data largely improve performance. Inspired by these successes, we train a teacher model on smaller human attention data and use this model to generate new spatial attention pseudo labels for ImageNet benchmark (see Figure 1 and Figure 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 307, + 366, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 366, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 366, + 319 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 326, + 508, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 326, + 508, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 326, + 508, + 339 + ], + "type": "text", + "content": "3.1. The Contrastive Learning Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 345, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 537 + ], + "type": "text", + "content": "Contrastive learning is one of the most popular self-supervised representation learning methods. It learns an embedding space so that similar data (positive) pairs are mapped to be close in the embedding space and different (negative) pairs are mapped to be far away. In practice, positive pairs are often generated by applying data augmentation to one image like adding noises, cropping, etc and negative pairs are different examples in the mini-batch. Among contrastive learning methods, SimCLR framework [6] has shown solid performance, which we choose here as our main method. For each batch, the images are augmented in two different ways. Then we feed them into ResNet feature extractor backbone and compute NT-Xent loss to minimize the difference between augmentations of the same images and maximize the differences between different images. NT-Xent loss can be computed as follows [6]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 333, + 544, + 545, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 544, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 333, + 544, + 545, + 573 + ], + "type": "interline_equation", + "content": "l _ {i, j} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {j}\\right) / \\tau\\right)}{\\sum_ {k = 1} ^ {2 N} 1 _ {k \\neq i} \\exp \\left(\\operatorname {s i m} \\left(z _ {i} , z _ {k}\\right) / \\tau\\right)} \\tag {1}", + "image_path": "3f586d840734188c1d4c5b02c917718bb2f1c66a5eda8b22a3928aef20906809.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "type": "text", + "content": " is the embedding of ith example, and " + }, + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\operatorname{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 579, + 545, + 604 + ], + "type": "text", + "content": " is the cosine similarity between two embeddings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 611, + 545, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 545, + 635 + ], + "type": "text", + "content": "3.2. ImageNet-Attn: spatial attention maps for ImageNet generated by a teacher model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "To train contrastive learning models, we need large datasets with millions of examples. Thus, to address the question whether human attention is beneficial for contrastive models, we will need large human spatial attention labeled dataset. However, there is no spatial attention dataset available with that size on typical large image" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23284" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 91, + 505, + 195 + ], + "blocks": [ + { + "bbox": [ + 74, + 73, + 81, + 80 + ], + "lines": [ + { + "bbox": [ + 74, + 73, + 81, + 80 + ], + "spans": [ + { + "bbox": [ + 74, + 73, + 81, + 80 + ], + "type": "text", + "content": "a" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 83, + 144, + 91 + ], + "lines": [ + { + "bbox": [ + 83, + 83, + 144, + 91 + ], + "spans": [ + { + "bbox": [ + 83, + 83, + 144, + 91 + ], + "type": "text", + "content": "Teacher model" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 91, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 109, + 91, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 109, + 91, + 505, + 195 + ], + "type": "image", + "image_path": "93fa50206b0d3cbe786cbf9ad40bcec24b5a6cbb038725bcb734ba052fc9a579.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 198, + 81, + 206 + ], + "lines": [ + { + "bbox": [ + 74, + 198, + 81, + 206 + ], + "spans": [ + { + "bbox": [ + 74, + 198, + 81, + 206 + ], + "type": "text", + "content": "b" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 201, + 348, + 209 + ], + "lines": [ + { + "bbox": [ + 342, + 201, + 348, + 209 + ], + "spans": [ + { + "bbox": [ + 342, + 201, + 348, + 209 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 99, + 231, + 168, + 300 + ], + "blocks": [ + { + "bbox": [ + 85, + 204, + 149, + 223 + ], + "lines": [ + { + "bbox": [ + 85, + 204, + 149, + 223 + ], + "spans": [ + { + "bbox": [ + 85, + 204, + 149, + 223 + ], + "type": "text", + "content": "Sallcon dataset image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 231, + 168, + 300 + ], + "lines": [ + { + "bbox": [ + 99, + 231, + 168, + 300 + ], + "spans": [ + { + "bbox": [ + 99, + 231, + 168, + 300 + ], + "type": "image", + "image_path": "5ea18cfcc421aaec73cd1885287a1fb4d55a91c07e8270484dcf100a13a57fa7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 99, + 301, + 168, + 370 + ], + "blocks": [ + { + "bbox": [ + 99, + 301, + 168, + 370 + ], + "lines": [ + { + "bbox": [ + 99, + 301, + 168, + 370 + ], + "spans": [ + { + "bbox": [ + 99, + 301, + 168, + 370 + ], + "type": "image", + "image_path": "a3768fe338159b9ab323fb7fe4c83cbd97773ba9f8b9a024ed1bf2e6d7cffa08.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 99, + 370, + 168, + 441 + ], + "blocks": [ + { + "bbox": [ + 99, + 370, + 168, + 441 + ], + "lines": [ + { + "bbox": [ + 99, + 370, + 168, + 441 + ], + "spans": [ + { + "bbox": [ + 99, + 370, + 168, + 441 + ], + "type": "image", + "image_path": "9b0c07aaea506276e13a400e8b827692afdae1d147790c8df966c239e5410b92.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 448, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 46, + 448, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 545, + 481 + ], + "type": "text", + "content": "Figure 2. (a) The architecture of teacher model to predict attention. The teacher model was trained from random initialization to predict ground truth (G.T.) human spatial attention labels. (b) Examples from Salicon dataset, with G.T. human attention heatmaps, and teacher model predictions. (c) Examples of spatial-attention pseudo-labels generated by the teacher model for ImageNet dataset (ImageNet-Attn)." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 180, + 231, + 249, + 300 + ], + "blocks": [ + { + "bbox": [ + 177, + 213, + 247, + 230 + ], + "lines": [ + { + "bbox": [ + 177, + 213, + 247, + 230 + ], + "spans": [ + { + "bbox": [ + 177, + 213, + 247, + 230 + ], + "type": "text", + "content": "ground truth human spatial attention" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 231, + 249, + 300 + ], + "lines": [ + { + "bbox": [ + 180, + 231, + 249, + 300 + ], + "spans": [ + { + "bbox": [ + 180, + 231, + 249, + 300 + ], + "type": "image", + "image_path": "1755ab91f9f4189d4abeb98133b473ee9329c5f854e24ff4137cd0426bd3cc5f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 180, + 301, + 249, + 369 + ], + "blocks": [ + { + "bbox": [ + 180, + 301, + 249, + 369 + ], + "lines": [ + { + "bbox": [ + 180, + 301, + 249, + 369 + ], + "spans": [ + { + "bbox": [ + 180, + 301, + 249, + 369 + ], + "type": "image", + "image_path": "9f3181742b530666673515591cb612661a6d81ae096a0b81856c8e1889370377.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 180, + 371, + 249, + 441 + ], + "blocks": [ + { + "bbox": [ + 180, + 371, + 249, + 441 + ], + "lines": [ + { + "bbox": [ + 180, + 371, + 249, + 441 + ], + "spans": [ + { + "bbox": [ + 180, + 371, + 249, + 441 + ], + "type": "image", + "image_path": "c8c95a5c3463ab36352c93dbf1d1a7eeb299c849967865c5a2914ea986749c88.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 260, + 231, + 329, + 300 + ], + "blocks": [ + { + "bbox": [ + 269, + 213, + 320, + 230 + ], + "lines": [ + { + "bbox": [ + 269, + 213, + 320, + 230 + ], + "spans": [ + { + "bbox": [ + 269, + 213, + 320, + 230 + ], + "type": "text", + "content": "teacher model predictions" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 231, + 329, + 300 + ], + "lines": [ + { + "bbox": [ + 260, + 231, + 329, + 300 + ], + "spans": [ + { + "bbox": [ + 260, + 231, + 329, + 300 + ], + "type": "image", + "image_path": "89d78b025c249a28fae17397f08ca1b33aa52c7b1632f052736d60eac8fdab9c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 260, + 301, + 329, + 369 + ], + "blocks": [ + { + "bbox": [ + 260, + 301, + 329, + 369 + ], + "lines": [ + { + "bbox": [ + 260, + 301, + 329, + 369 + ], + "spans": [ + { + "bbox": [ + 260, + 301, + 329, + 369 + ], + "type": "image", + "image_path": "574387cf417c34711f8395c2b0906971b0b95162958167806d02b86f105810dc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 260, + 371, + 329, + 441 + ], + "blocks": [ + { + "bbox": [ + 260, + 371, + 329, + 441 + ], + "lines": [ + { + "bbox": [ + 260, + 371, + 329, + 441 + ], + "spans": [ + { + "bbox": [ + 260, + 371, + 329, + 441 + ], + "type": "image", + "image_path": "71cafe5f403f4e8fd4d126bcef999bb009c1d4417a991953e2004132f5c78329.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 353, + 232, + 422, + 300 + ], + "blocks": [ + { + "bbox": [ + 352, + 204, + 419, + 224 + ], + "lines": [ + { + "bbox": [ + 352, + 204, + 419, + 224 + ], + "spans": [ + { + "bbox": [ + 352, + 204, + 419, + 224 + ], + "type": "text", + "content": "ImageNet dataset image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 232, + 422, + 300 + ], + "lines": [ + { + "bbox": [ + 353, + 232, + 422, + 300 + ], + "spans": [ + { + "bbox": [ + 353, + 232, + 422, + 300 + ], + "type": "image", + "image_path": "e0574f871272e15f4f1fd1a2d12a754c2afe77495dbe90ec41c3172de78c207d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 353, + 301, + 422, + 370 + ], + "blocks": [ + { + "bbox": [ + 353, + 301, + 422, + 370 + ], + "lines": [ + { + "bbox": [ + 353, + 301, + 422, + 370 + ], + "spans": [ + { + "bbox": [ + 353, + 301, + 422, + 370 + ], + "type": "image", + "image_path": "376c356029e8c7641a85a9d170af3ed4b7726fa1f828194ba10f67a263af4f35.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 353, + 372, + 422, + 441 + ], + "blocks": [ + { + "bbox": [ + 353, + 372, + 422, + 441 + ], + "lines": [ + { + "bbox": [ + 353, + 372, + 422, + 441 + ], + "spans": [ + { + "bbox": [ + 353, + 372, + 422, + 441 + ], + "type": "image", + "image_path": "66963753b7c8fb7087a4029952f3b08637b4dc57463f9eddc3fc8ddaf91f8380.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 439, + 232, + 509, + 300 + ], + "blocks": [ + { + "bbox": [ + 433, + 214, + 514, + 232 + ], + "lines": [ + { + "bbox": [ + 433, + 214, + 514, + 232 + ], + "spans": [ + { + "bbox": [ + 433, + 214, + 514, + 232 + ], + "type": "text", + "content": "spatial-attention labels from the teacher model" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 439, + 232, + 509, + 300 + ], + "lines": [ + { + "bbox": [ + 439, + 232, + 509, + 300 + ], + "spans": [ + { + "bbox": [ + 439, + 232, + 509, + 300 + ], + "type": "image", + "image_path": "f8907546e08968dbf24fe5c0b0f03b1acb32454e82b1f8c7d686f3e42a3ee3bd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 439, + 301, + 508, + 370 + ], + "blocks": [ + { + "bbox": [ + 439, + 301, + 508, + 370 + ], + "lines": [ + { + "bbox": [ + 439, + 301, + 508, + 370 + ], + "spans": [ + { + "bbox": [ + 439, + 301, + 508, + 370 + ], + "type": "image", + "image_path": "d4d20847130eeb8ac6624910389ad3c7e00051a253297e68fdf56d3f33f2478a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 439, + 372, + 508, + 441 + ], + "blocks": [ + { + "bbox": [ + 439, + 372, + 508, + 441 + ], + "lines": [ + { + "bbox": [ + 439, + 372, + 508, + 441 + ], + "spans": [ + { + "bbox": [ + 439, + 372, + 508, + 441 + ], + "type": "image", + "image_path": "c9d2174013c7ecfccbb3d60ba276f079aff4be0b4285359ef7f8fe2541a317fc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 491, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 538 + ], + "type": "text", + "content": "benchmark such as ImageNet. The largest attention data set available is Silicon [25], consisting of twenty thousand examples only, orders of magnitude smaller than ImageNet, which has around 14 million examples." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": "To generate attention labels for ImageNet images, we train a teacher model on Silicon attention data set to predict human attention ground truth labels (see Figure 2b). Then we use this teacher model to create a new dataset (ImageNet-Attn) with spatial attention maps for ImageNet (see Figure 2c for some examples). The teacher model architecture is illustrated in Figure 2a. Specifically, we used MobileNet-V3-small as the backbone, and embeddings from 4 layers (conv 2,4,6,8) are extracted. On each embedding, we applied two conv layers (the first conv layer with " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " kernel, number of channels matching the input embedding, max pooling of 3 and relu; the second conv layer with " + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 546, + 288, + 713 + ], + "type": "text", + "content": " kernel, 1 channel and relu). We bilinearly resized the output to the same resolution of input. The" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 491, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 681 + ], + "type": "text", + "content": "result of the 4 branches are then summed to yield 1 feature, followed up by a sigmoid function to obtain spatial-attention map. Note that using multiple intermediate (both early and late) layers is one of the key ingredient to make the spatial attention teacher model successful. This strategy mimics human visual attention, which is known to be affected by both low-level characteristics such as color, intensity, texture, as well as high level characteristics like shape and object, etc. Essentially, our teacher model has a similar architecture as in [36], but is further simplified with more efficient backbone, less channels/layers, etc, so that it can be trained from scratch on Silicon data with randomly initialized backbone to avoid any leak of class label information, while most existing attention prediction models [1,36] needs to finetune with a pretrained classification backbone network." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23285" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "3.3. Training contrastive models with spatial attention maps" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": "Spatial Attention Branch The overall training framework is shown in Figure 1. This spatial attention prediction framework consists of two branches: the contrastive branch and the spatial attention branch. The contrastive branch is the same as the original SimCLR method [6], which applies augmentations to image " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": " to get different variants " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": ", and learns the representation " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "h_{i}" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "h_{j}" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": " via a feature extractor backbone network (e.g., ResNet), then use a projection head to map " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "h_{i} / h_{j}" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "inline_equation", + "content": "z_{i} / z_{j}" + }, + { + "bbox": [ + 46, + 102, + 286, + 220 + ], + "type": "text", + "content": ", where the contrastive loss is applied." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "text", + "content": "For the spatial-attention branch, it takes as an input not only the final embedding " + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "text", + "content": ", but also early intermediate layer embeddings (following the same reason of including both high level and low level visual cues as in teacher model), and predicts an spatial attention heatmap " + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "inline_equation", + "content": "m_p" + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "text", + "content": ". More specifically, we apply a global average pooling on the output of the last three blocks of the ResNet backbone. Then, we select the max channel for each of three block output (after average pooling), and resize with bilinear interpolation to the image resolution, which has some resemblance to CAM approaches [51]. Finally we stack the representations together, pass them into a linear readout layer (with a bias term), and use the output as our final spatial attention prediction " + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "inline_equation", + "content": "m_p" + }, + { + "bbox": [ + 46, + 222, + 286, + 412 + ], + "type": "text", + "content": ". We keep the spatial attention head as simple as possible so that the guide on attention head output can be back propagated to representation more directly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "text", + "content": "We use the ImageNet-Attn data generated by the teacher model as target for each ImageNet example, denoted as " + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "text", + "content": ", we can then train the network spatial attention output " + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "inline_equation", + "content": "m_p" + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "text", + "content": " to be close to " + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "text", + "content": ". We hypothesize that this method regularizes the training of the feature extractor backbone rather than explicitly enforce the network to generate masked representations that match the spatial attention maps. Note that for attention branch, there is no augmentation applied to each image " + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 413, + 286, + 555 + ], + "type": "text", + "content": ", since human attention is not invariant to transformation (e.g., a human looking at a cropped image may attend to different regions compared to a consistent crop of human attention map of the original image)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 556, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 286, + 578 + ], + "type": "text", + "content": "Loss function The loss function " + }, + { + "bbox": [ + 47, + 556, + 286, + 578 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 556, + 286, + 578 + ], + "type": "text", + "content": " consists of two terms:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 581, + 286, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 581, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 123, + 581, + 286, + 592 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {c o n t r a}} + L _ {\\text {a t t n}} \\tag {2}", + "image_path": "2b40bb3d081063ec966ed2b97e943f81b27c8dc1f37b8fbe438f62e2c0f1497a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "inline_equation", + "content": "L_{\\text{contra}}" + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "text", + "content": " is the contrastive loss, or more specifically " + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "inline_equation", + "content": "L_{\\text{contra}} = \\sum_{i,j} l_{i,j}" + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "inline_equation", + "content": "l_{i,j}" + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "text", + "content": " is defined in Equation (1), the same as in [6]. " + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "inline_equation", + "content": "L_{\\text{attr}}" + }, + { + "bbox": [ + 47, + 596, + 286, + 643 + ], + "type": "text", + "content": " is the attention regularization loss, and more specifically" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 648, + 286, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 648, + 286, + 672 + ], + "spans": [ + { + "bbox": [ + 59, + 648, + 286, + 672 + ], + "type": "interline_equation", + "content": "L _ {a t t n} = \\sum_ {i} \\left(\\lambda K L D \\left(m _ {i} ^ {p}, m _ {i} ^ {t}\\right) - \\beta N S S \\left(m _ {i} ^ {p}, p _ {i} ^ {t}\\right)\\right) \\tag {3}", + "image_path": "cd98e6ed29e12105c885517afbac86ed8d58978014dcd5bac5657d51f64c6d9f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " are two weighting parameters (" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " in our work). " + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "m_{i}^{p}" + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": " is the predicted 2D spatial-attention map from the attention head of our proposed" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 114, + 544, + 238 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "type": "text", + "content": "Table 1. ImageNet Top-1 classification accuracy for different models (mean " + }, + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 305, + 70, + 545, + 103 + ], + "type": "text", + "content": " SE for 3 seeds, except for * which means best result from all runs)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 114, + 544, + 238 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 544, + 238 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 544, + 238 + ], + "type": "table", + "html": "
ModelAccuracy (%)
Contrastive67.61 ± 0.04
Contrastive attn. teacher68.23 ± 0.08
Contrastive attn. co-train66.35 ± 0.12
Contrastive attn. with explicit attention block [32]61.15*
Contrastive attn. with self attention mask [32]49.65*
Contrastive attn. with ContraCAM [35] as attention67.71*
Supervised75.91 ± 0.10
Supervised attn. teacher76.02 ± 0.04
Supervised (ResNet-18)69.17 ± 0.07
Supervised (ResNet-18) attn. teacher69.30 ± 0.04
", + "image_path": "a9b10a12a9df8bb7e0fabaccd40d072c535616b64e15a72095a869a368154c79.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": "model for " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": "-th example (original image, not augmented ones), " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "m_i^t" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": " is the pseudo spatial-attention target map predicted by the teacher model for " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": "-th example, and " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "KLD()" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": " is the KL divergence1. Besides KLD loss, we also use " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "NSS()" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": ", the Normalized Scanpath Saliency loss [4]. Those two losses are typically used for human attention predictions [3,9]. KLD is often used to match the spatial-attention target distribution, and NSS is typically added on top as it is generally observed to help generate attention maps that are in perceptual agreement with human judgements (see [3,34]). For " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "NSS" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": " loss, the larger the better, so there is a negative sign before it. " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "NSS" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": " loss needs gaze/attention points instead of heatmap as ground-truth, so we extract pseudo gaze/attention points " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "p_i^t" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": " from the attention heatmap " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "m_i^t" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": ". To obtain " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "p_i^t" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": ", we first extract the point with highest value in current spatial-attention map " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "m_i^t" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": ", then generate a new spatial-attention map by subtracting a Gaussian blur around the extracted point from the current attention map. The process is repeated with the new attention map until the maximal value of the attention map is smaller than a threshold (more specifically, " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "0.2 * \\max\\_ sal" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\max\\_ sal" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": " is the maximal value in " + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "inline_equation", + "content": "m_i^t" + }, + { + "bbox": [ + 304, + 258, + 545, + 521 + ], + "type": "text", + "content": ").2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 533, + 387, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 387, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 387, + 546 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 552, + 436, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 436, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 436, + 565 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "type": "text", + "content": "In our experiments, in Eq.1, we choose " + }, + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "type": "inline_equation", + "content": "N = 2048" + }, + { + "bbox": [ + 306, + 571, + 545, + 605 + ], + "type": "text", + "content": ". In Eq.2, we used equal weighting of 1.0 for the SimCLR and attention losses." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 607, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 607, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 317, + 607, + 545, + 619 + ], + "type": "text", + "content": "The weight for KLD and NSS losses are 1 and 0.1, re" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 626, + 545, + 655 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 545, + 655 + ], + "type": "inline_equation", + "content": "{}^{1}{m}_{i}^{p}" + }, + { + "bbox": [ + 306, + 626, + 545, + 655 + ], + "type": "text", + "content": " is normalized when computing KLD, i.e.,divided by its pixel value sum,so that it becomes a distribution with all pixel value sum equal to 1. Similarly for " + }, + { + "bbox": [ + 306, + 626, + 545, + 655 + ], + "type": "inline_equation", + "content": "{m}_{i}^{t}" + }, + { + "bbox": [ + 306, + 626, + 545, + 655 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 655, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 655, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 655, + 545, + 712 + ], + "type": "text", + "content": "2Our process of extracting gaze points from heatmaps is to inverse the process of generating heatmaps from gaze points. Due to the uncertainty about the gaze locations, researchers typically apply Gaussian blur for each gaze point, then sum all Gaussians to generate the heatmap (see [3, 26]). We follow the inverse of this process to obtain gaze points from heatmap, similarly as done in [22]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23286" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 114, + 155, + 192 + ], + "blocks": [ + { + "bbox": [ + 62, + 85, + 69, + 92 + ], + "lines": [ + { + "bbox": [ + 62, + 85, + 69, + 92 + ], + "spans": [ + { + "bbox": [ + 62, + 85, + 69, + 92 + ], + "type": "text", + "content": "a" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 103, + 92, + 127, + 102 + ], + "lines": [ + { + "bbox": [ + 103, + 92, + 127, + 102 + ], + "spans": [ + { + "bbox": [ + 103, + 92, + 127, + 102 + ], + "type": "text", + "content": "image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 114, + 155, + 192 + ], + "lines": [ + { + "bbox": [ + 76, + 114, + 155, + 192 + ], + "spans": [ + { + "bbox": [ + 76, + 114, + 155, + 192 + ], + "type": "image", + "image_path": "32d7e5374b151f79b5cb8a57482cedc8232264f0992237cd7e8a14ad84159635.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 113, + 247, + 192 + ], + "blocks": [ + { + "bbox": [ + 169, + 92, + 242, + 111 + ], + "lines": [ + { + "bbox": [ + 169, + 92, + 242, + 111 + ], + "spans": [ + { + "bbox": [ + 169, + 92, + 242, + 111 + ], + "type": "text", + "content": "ground truth human spatial attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 113, + 247, + 192 + ], + "lines": [ + { + "bbox": [ + 168, + 113, + 247, + 192 + ], + "spans": [ + { + "bbox": [ + 168, + 113, + 247, + 192 + ], + "type": "image", + "image_path": "2eecca999a2c947038b270bef71548eb6649dea9201ac5ae400bb29351fe1285.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 261, + 114, + 340, + 192 + ], + "blocks": [ + { + "bbox": [ + 274, + 92, + 332, + 111 + ], + "lines": [ + { + "bbox": [ + 274, + 92, + 332, + 111 + ], + "spans": [ + { + "bbox": [ + 274, + 92, + 332, + 111 + ], + "type": "text", + "content": "teacher model spatial attention" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 114, + 340, + 192 + ], + "lines": [ + { + "bbox": [ + 261, + 114, + 340, + 192 + ], + "spans": [ + { + "bbox": [ + 261, + 114, + 340, + 192 + ], + "type": "image", + "image_path": "846b7e8c3341bff98bef74217b447a2b18dbaa29226df152eace79dcd4cf4f5d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 354, + 113, + 433, + 192 + ], + "blocks": [ + { + "bbox": [ + 353, + 92, + 438, + 110 + ], + "lines": [ + { + "bbox": [ + 353, + 92, + 438, + 110 + ], + "spans": [ + { + "bbox": [ + 353, + 92, + 438, + 110 + ], + "type": "text", + "content": "spatial attention guided contrastive model" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 113, + 433, + 192 + ], + "lines": [ + { + "bbox": [ + 354, + 113, + 433, + 192 + ], + "spans": [ + { + "bbox": [ + 354, + 113, + 433, + 192 + ], + "type": "image", + "image_path": "539fbb18ff73cf7649d2a6653796aee1e0b4187cf05d2c7128a02397947ec962.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 449, + 113, + 528, + 192 + ], + "blocks": [ + { + "bbox": [ + 470, + 92, + 520, + 110 + ], + "lines": [ + { + "bbox": [ + 470, + 92, + 520, + 110 + ], + "spans": [ + { + "bbox": [ + 470, + 92, + 520, + 110 + ], + "type": "text", + "content": "baseline trastive model" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 113, + 528, + 192 + ], + "lines": [ + { + "bbox": [ + 449, + 113, + 528, + 192 + ], + "spans": [ + { + "bbox": [ + 449, + 113, + 528, + 192 + ], + "type": "image", + "image_path": "eaf653822539f8748e0132a7dbe5e18f56f00e4b46a529919778f1287b405ded.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 76, + 194, + 154, + 273 + ], + "blocks": [ + { + "bbox": [ + 76, + 194, + 154, + 273 + ], + "lines": [ + { + "bbox": [ + 76, + 194, + 154, + 273 + ], + "spans": [ + { + "bbox": [ + 76, + 194, + 154, + 273 + ], + "type": "image", + "image_path": "a945762d74bebc49c3a6083f0a783e9040a178ad6e3ef8170cb3d7f826e23f03.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 76, + 274, + 154, + 353 + ], + "blocks": [ + { + "bbox": [ + 76, + 274, + 154, + 353 + ], + "lines": [ + { + "bbox": [ + 76, + 274, + 154, + 353 + ], + "spans": [ + { + "bbox": [ + 76, + 274, + 154, + 353 + ], + "type": "image", + "image_path": "45d66b37c81c7ea81df24645f90460ad72fb3484b062276c1688f843f84515ba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 168, + 194, + 246, + 272 + ], + "blocks": [ + { + "bbox": [ + 168, + 194, + 246, + 272 + ], + "lines": [ + { + "bbox": [ + 168, + 194, + 246, + 272 + ], + "spans": [ + { + "bbox": [ + 168, + 194, + 246, + 272 + ], + "type": "image", + "image_path": "642d6b9bdc2ab1f0f8f178624888629995f74415b2b5995cf545f63ed0acf79c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 168, + 274, + 246, + 353 + ], + "blocks": [ + { + "bbox": [ + 168, + 274, + 246, + 353 + ], + "lines": [ + { + "bbox": [ + 168, + 274, + 246, + 353 + ], + "spans": [ + { + "bbox": [ + 168, + 274, + 246, + 353 + ], + "type": "image", + "image_path": "736137c06046f970352bf1a27a8aeec06985b2ca244f1d79af159d04e98105e3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 261, + 194, + 339, + 272 + ], + "blocks": [ + { + "bbox": [ + 261, + 194, + 339, + 272 + ], + "lines": [ + { + "bbox": [ + 261, + 194, + 339, + 272 + ], + "spans": [ + { + "bbox": [ + 261, + 194, + 339, + 272 + ], + "type": "image", + "image_path": "e69720b1fcb4eb30c868b50bfacfe9a8090aed7f1b35fe46931c49b39c4087ad.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 260, + 274, + 338, + 353 + ], + "blocks": [ + { + "bbox": [ + 260, + 274, + 338, + 353 + ], + "lines": [ + { + "bbox": [ + 260, + 274, + 338, + 353 + ], + "spans": [ + { + "bbox": [ + 260, + 274, + 338, + 353 + ], + "type": "image", + "image_path": "0d082c4709c76c8911510e6ae8101c69d06a1b24a93c94a76c433f9fdf77c82d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 354, + 194, + 432, + 272 + ], + "blocks": [ + { + "bbox": [ + 354, + 194, + 432, + 272 + ], + "lines": [ + { + "bbox": [ + 354, + 194, + 432, + 272 + ], + "spans": [ + { + "bbox": [ + 354, + 194, + 432, + 272 + ], + "type": "image", + "image_path": "2f9463b54bfb9044d1eab9c31c10c71d034fb0cd5baab8747b43241ba864a9bf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 353, + 274, + 432, + 353 + ], + "blocks": [ + { + "bbox": [ + 353, + 274, + 432, + 353 + ], + "lines": [ + { + "bbox": [ + 353, + 274, + 432, + 353 + ], + "spans": [ + { + "bbox": [ + 353, + 274, + 432, + 353 + ], + "type": "image", + "image_path": "1be15bb1b92d287fe25227d4a63aeeb56c5377dbccf79ffe32f0e94756e900bc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 450, + 194, + 528, + 274 + ], + "blocks": [ + { + "bbox": [ + 450, + 194, + 528, + 274 + ], + "lines": [ + { + "bbox": [ + 450, + 194, + 528, + 274 + ], + "spans": [ + { + "bbox": [ + 450, + 194, + 528, + 274 + ], + "type": "image", + "image_path": "d7b688f628f4f15ccfaf1b70a84c4b737b12d46893e61d4698ed198b008255a1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 450, + 274, + 528, + 353 + ], + "blocks": [ + { + "bbox": [ + 450, + 274, + 528, + 353 + ], + "lines": [ + { + "bbox": [ + 450, + 274, + 528, + 353 + ], + "spans": [ + { + "bbox": [ + 450, + 274, + 528, + 353 + ], + "type": "image", + "image_path": "c85bff58326ab2ea0aa7f50d5c220a869fbc3fa580de36be75793ade749e5826.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 66, + 361, + 293, + 494 + ], + "blocks": [ + { + "bbox": [ + 62, + 353, + 69, + 361 + ], + "lines": [ + { + "bbox": [ + 62, + 353, + 69, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 69, + 361 + ], + "type": "text", + "content": "b" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 361, + 293, + 494 + ], + "lines": [ + { + "bbox": [ + 66, + 361, + 293, + 494 + ], + "spans": [ + { + "bbox": [ + 66, + 361, + 293, + 494 + ], + "type": "image", + "image_path": "1c3ddacaa39d16ca0e6be2230e0bc9b0daf621bf30c33814a8bc83f594917518.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 304, + 361, + 533, + 495 + ], + "blocks": [ + { + "bbox": [ + 239, + 495, + 368, + 505 + ], + "lines": [ + { + "bbox": [ + 239, + 495, + 368, + 505 + ], + "spans": [ + { + "bbox": [ + 239, + 495, + 368, + 505 + ], + "type": "text", + "content": "correlation coeff. (model vs human)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 361, + 533, + 495 + ], + "lines": [ + { + "bbox": [ + 304, + 361, + 533, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 533, + 495 + ], + "type": "image", + "image_path": "8b3c64989230165649d08fbd2ea5c4944425e7a4354d50107fa8d2d7ecf74867.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 511, + 545, + 611 + ], + "lines": [ + { + "bbox": [ + 46, + 511, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 545, + 611 + ], + "type": "text", + "content": "Figure 3. (a) Examples comparing spatial attention maps predicted by different models vs ground truth human attention data on OSIE dataset [43]. (b) Distribution of correlation coefficients between model predicted attention maps vs ground truth human attention maps on Silicon [25] validation set (left) and OSIE dataset (right) [43]. Three models are shown: (1) Teacher model trained on Silicon [25] training set. (2) Baseline SimCLR model (3) The proposed spatial attention guided SimCLR model trained with pseudo-attention labels provided by the teacher model. Please note that both baseline and our model are trained to predict target attention labels, with main difference that the baseline model has a stop gradient placed between the prediction layer and network features to prevent spatial-attention targets to inform network features. Also note that our teacher model is trained on Silicon where the attention is collected by mouse tracking, while the results in (a) and (c) here are on OSIE data, where the attention are collected by eye trackers. The results demonstrated both the teacher model and the attention head in proposed model has a good generalization capacity for new kinds of attention labels." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 623, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 671 + ], + "type": "text", + "content": "spectively. We obtained those weights by hyperparameter search from 0.1-5 for KLD and 0.01-1 for NSS, based on ImageNet accuracy of a separate dev set (80%-20% split on training set for parameter search with 20% as dev set)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 623, + 545, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 545, + 646 + ], + "type": "text", + "content": "4.2. Spatial attention guided models are highly predictive of human attention" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 546, + 713 + ], + "type": "text", + "content": "In this section, we explore whether the use of auxiliary teacher model to provide spatial attention pseudo labels on ImageNet better aligns contrastive model's attention with human attention. We define aligning model spatial attention here as the ability to predict spatial attention mask from" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23287" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "the model backbone features by a simple readout layer discussed in Section 3. We trained two ResNet-50 backbones using the SimCLR objective from [6]. We added additional spatial-attention losses as discussed in Section 3. For the baseline model, we placed a stop gradient operation between the backbone features and the attention projection head to prevent attention information from leaking to the backbone features, whereas for the attention guided model, we allowed the learned attention gradients to flow back to the backbone." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 324 + ], + "type": "text", + "content": "We evaluated the degree the predicted attention maps is aligned with human attention by performing correlation analysis, which is typically used to compare human and model spatial attention. We measured Pearson's correlation coefficient between the model predicted attention and ground truth human spatial-attention maps from Silicon [25] validation set (Note that the teacher model is trained with Silicon training set). We find that both teacher model and spatial attention guided contrastive model are highly correlated with human attention (Figure 3b), while the baseline contrastive model is much less correlated." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "content": "To test whether this alignment with human attention is general, beyond just Salicon data, we obtained another attention dataset on OSIE images [43] (Figure 3a). This data more faithfully reflects human attention as it collected from (mobile) eye tracker [43]. We find that the baseline model is weakly positively correlated with human attention (ttest: " + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "\\rho = 0.07" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "content": ") suggesting that the contrastive loss produces features that are predictive of human attention to some extent. Yet, the correlation was generally close to 0 and explains only " + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "content": " of data variance. The spatial attention guided model has a much stronger correlation to human attention (ttest: " + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "\\rho = 0.48" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "content": ") than the baseline model (Two samples ttest: " + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "content": "), and thus more faithfully reflecting human visual attention (See Fig 3a for qualitative examples and Fig 3b bottom for quantitative analysis). Further, the correlation with human attention for the teacher model and the attention guided contrastive model were quite similar (Two samples ttest: " + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "inline_equation", + "content": "p = 0.7" + }, + { + "bbox": [ + 46, + 326, + 289, + 542 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 554, + 287, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 578 + ], + "type": "text", + "content": "4.3. Spatial attention guided models are more accurate than baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "We evaluate the quality of the representations learned by spatial attention guidance framework using the typical contrastive learning evaluation criteria: fitting an ImageNet [37] linear classifier on top of the frozen representation (in practice we place stop gradient at the end of the backbone and train the classifier concurrently while training the backbone). We compute Top 1 accuracy on ImageNet validation set and compare the results with baselines. As shown in Table 1, we observe around " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " accuracy gain on ImageNet compared to vanilla SimCLR. We further explore an alternative way of incorporating human attention data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 276 + ], + "type": "text", + "content": "Rather than using pseudo attention labels on ImageNet from the teacher model, we add Silicon data to the training data, and directly predict attention ground truth labels for Silicon data with spatial attention head (Contrastive attn. co-train in Table 1). More specifically, we pass 2048 ImageNet images and N=2048 Silicon images into our backbone (N is obtained by hyperparameter search 512-2048 based on evaluation on the separate dev set) to compute backbone features. Then, we compute the SimCLR loss from ImageNet features and use Silicon images features to predict spatial attention maps, and we use Silicon attention ground truth labels to compute the KLD and NSS losses. Interestingly, we find this method to lead to worse performance compared to using the teacher model generated spatial-attention labels, which gives evidence that the teacher model is generalizing its knowledge about human attention data beyond the limited Silicon training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 277, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 517 + ], + "type": "text", + "content": "One important question is whether the classic trainable attention method applies to contrastive learning. For example, in [32], an attention block/map is used to adjust the weights of the embedding, where the attention block can be self learned implicitly, or guided by human attention ground-truth. It shows both implicit self-learned attention block and attention block guided by human attention can improve the accuracy of fine-grained classification (Table VII in [32]). We adopted the approach described in their paper: add an attention block after the representation embedding, and use the attention block's output to mask the representation embedding. We apply it to the contrastive models, and the results are shown in Table 1 as \"explicit attention mask\" (the attention block is guided by human attention) and \"self attention mask\" (self attention). As observed, both yield poor performance, since the contrastive attention with self-attention mask method learns an attention mask from the contrastive loss rather than to predict the teacher targets, the model may be getting weaker supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 663 + ], + "type": "text", + "content": "Another important question is whether our current spatial attention targets can be replaced by other spatial attention approach like ContraCAM [35]. In our method, we predicted the spatial attention with a linear layer. Whereas in ContraCAM a contrastive loss is used to construct the attention prediction. In both cases, we trained the predictions to match the target attention maps obtained from the teacher model. The result is shown in Table 1 as \"Contrastive with ContraCAM as attention\". However, we don't see much accuracy improvements with ContraCAM as the model attention, showing that the linear simple prediction leads to better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "To investigate whether the spatial attention guidance framework benefits supervised models in the same way as contrastive models, we applied the same approach for supervised models. Supervised models similarly benefit from" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23288" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "this framework, yet the gain is limited compared to the contrastive models perhaps due to the higher accuracy the supervised models achieves compared to the contrastive models. To control for accuracy, we used a supervised model trained with smaller backbone (ResNet-18), which gives comparable accuracy to the bigger ResNet-50 backbone that is trained with contrastive objective. Even when controlling the mismatch in accuracy, the supervised model gain from providing spatial-attention supervision is limited compared to contrastive models (Table 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 201, + 288, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 288, + 224 + ], + "type": "text", + "content": "4.4. Spatial attention guided models are more robust than baselines" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 434 + ], + "type": "text", + "content": "Human vision is very robust to noise or small pixel modulations, compared to computer vision models. Here, we hypothesize that models that aligns better with human attention may learn more robust representations. We test this hypothesis using ImageNet-C dataset [18]. We take the representations learned from the proposed model and train a linear classifier on ImageNet training data. We then evaluate classification performance on ImageNet-C at various corruption/noise types. Table Supp.1 in the supplementary compares the classification performance on ImageNet-C for baseline contrastive model and the proposed models trained with spatial-attention teacher guidance, average over the 5 different corruption/noise magnitudes in ImageNet-C. We find that contrastive models trained with teacher guidance outperforms baseline consistently, suggesting that the representation learned by spatial-attention guided model is indeed more robust." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 287, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 468 + ], + "type": "text", + "content": "4.5. Spatial attention guided models generate better representation for retrieval" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 714 + ], + "type": "text", + "content": "Besides classification, we tested the quality of the representations for another downstream task: image retrieval. We use the model to extract representation for ImageNet validation set, and use the representations to run image retrieval. 5000 randomly chosen validation images are used as query, while the rest validation images are database images. For each query image, top k retrieval results are returned by sorting the cosine distance between representations of database images and the query. The retrieval accuracy is measured with the standard precision recall curve [40], where precision and recall are computed by checking whether results have the same class label as the query. Mean precision and recall is obtained as the average across queries. We choose different k values, and draw the mean precision-recall curve. Besides ImageNet (\"clean\" in Fig Supp.1a in the supplementary), we also use ImageNet-C (\"noise level 1-5\"). Results from \"fog\" corruption is shown in In Figure Supp.1a in the supplementary., while results for other corruptions can be found in the supplementary too. The representation extracted with" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "the Contrastive attn teacher model outperforms the baseline Constractive model in this retrieval task, on both ImageNet images, and ImageNet-C images for most corruption types and levels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 263 + ], + "type": "text", + "content": "Moreover, we evaluate transfer to other dataset for the retrieval task. We take the model trained on ImageNet to extract representations for CIFAR-10 [29] test set and CIFAR-10-C [18] (\"clean\" and \"noise level 1-5\" in Fig Supp.1b in the supplementary respectively), without retraining/finetuning. 1000 images are used as query and the rest are used as database. Results for \"fog\" noise are shown in Fig Supp.1b in the supplementary, while results for other noises are shown in supplementary too. As shown, the proposed attention guided model outperforms the baseline model on both clean images and almost all noise/corruption types and levels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 270, + 378, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 378, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 378, + 281 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 289, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 443 + ], + "type": "text", + "content": "In this work, we tested the hypothesis that utilizing human spatial attention can be beneficial for obtaining better representation for contrastive models. We overcome the challenge of obtaining human spatial attention labels for large dataset by utilizing a teacher model trained on limited human attention labels to provide pseudo-attention labels for ImageNet. We augmented training of a commonly-used network (ResNet50) trained with SimCLR contrastive objective with pseudo-spatial attention labels from the teacher model. Our results demonstrate that contrastive models trained with those pseudo-attention labels are more predictive of human attention and we obtain better representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 445, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 635 + ], + "type": "text", + "content": "Despite the gains observed on downstream tasks are not large, the gains were consistent across 3 tasks (classification, robustness, and image retrieval). Thus, taken together, our findings support the above hypothesis. The limited gains in classification may suggest that spatial attention alone may not be enough to achieve SOTA, but perhaps in conjunction with other architectural and methodological improvements. Another limitation is that our exploration was conducted on natural images only, but other interesting applications may include other domains like medical images or autonomous driving. When domains become very different, it may be required to retrain teacher models to capture those new domains (e.g., train on attention data that comes from medical experts for medical images). These are interesting and important questions that may be addressed in future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 646, + 416, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 416, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 416, + 659 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We are grateful to Yifan Tian, Josh McAdams, Zhuode Liu, and Ethan Steinberg for assistance with analyses. We also thank Mike Mozer, Kai Kohlhoff, Simon Kornblith, and Ting Chen for useful discussions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23289" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Ali Borji. Saliency prediction in the deep learning era: Successes and limitations. IEEE transactions on pattern analysis and machine intelligence, 2019. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 159 + ], + "type": "text", + "content": "[2] Zoya Bylinskii, Tilke Judd, Ali Borji, Laurent Itti, Frédo Durand, Aude Oliva, and Antonio Torralba. Mit saliency benchmark. http://saliency.mit.edu, 2012. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 192 + ], + "type": "text", + "content": "[3] Z. Bylinskii, T. Judd, A. Oliva, A. Torralba, and F. Durand. What do different evaluation metrics tell us about saliency models? IEEE TPAMI, 41(3):740-757, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "type": "text", + "content": "[4] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Frédo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(3):740-757, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 240, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 288, + 283 + ], + "type": "text", + "content": "[5] Jorge Paolo Casas and Chandramouli Chandrasekaran. openeyetrack-a high speed multi-threaded eye tracker for head-fixed applications. Journal of open source software, 4(42), 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 329 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 363 + ], + "type": "text", + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 287, + 408 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750-15758, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 410, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 453 + ], + "type": "text", + "content": "[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an lstmbased saliency attentive model. IEEE TIP, 27(10):5142- 5154, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[10] Abhishek Das, Harsh Agrawal, Larry Zitnick, Devi Parikh, and Dhruv Batra. Human attention in visual question answering: Do humans and deep networks look at the same regions? Computer Vision and Image Understanding, 163:90-100, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 565 + ], + "type": "text", + "content": "[11] Mohammad K Ebrahimpour, J Ben Falandays, Samuel Spevack, and David C Noelle. Do humans look where deep convolutional neural networks \"attend\"? In International Symposium on Visual Computing, pages 53-65. Springer, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[12] Simone Frintrop, Erich Rome, and Henrik I Christensen. Computational visual attention systems and their cognitive foundations: A survey. ACM Transactions on Applied Perception (TAP), 7(1):1-39, 2010. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "to self-supervised learning. Advances in Neural Information Processing Systems, 33:21271-21284, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[15] Meng-Hao Guo, Tian-Xing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph R Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational Visual Media, pages 1-38, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 204 + ], + "type": "text", + "content": "[16] Suna Sihang Guo, Ruohan Zhang, Bo Liu, Yifeng Zhu, Dana Ballard, Mary Hayhoe, and Peter Stone. Machine versus human attention in deep reinforcement learning tasks. Advances in Neural Information Processing Systems, 34:25370-25385, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 303 + ], + "type": "text", + "content": "[18] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2019. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 304, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 358 + ], + "type": "text", + "content": "[19] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. Silicon: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 262-270, 2015. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "type": "text", + "content": "[20] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "text", + "content": "[21] Laurent Itti and Christof Koch. Computational modelling of visual attention. Nature reviews neuroscience, 2(3):194-203, 2001. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[22] L. Itti, C. Koch, and E. Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE TPAMI, 20(11):1254-1259, 1998. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "type": "text", + "content": "[23] Saumya Jetley, Nicholas A Lord, Namhoon Lee, and Philip HS Torr. Learn to pay attention. arXiv preprint arXiv:1804.02391, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[24] Sen Jia and Neil DB Bruce. Eml-net: An expandable multilayer network for saliency prediction. Image and Vision Computing, 95:103887, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[25] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Salicon: Saliency in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1072-1080, 2015. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "type": "text", + "content": "[26] T. Judd, F. Durand, and A. Torralba. A benchmark of computational models of saliency to predict human fixations. 2012. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 679 + ], + "type": "text", + "content": "[27] Nam Wook Kim, Zoya Bylinskii, Michelle A Borkin, Krzysztof Z Gajos, Aude Oliva, Fredo Durand, and Hanspeter Pfister. Bubbleview: an interface for crowdsourcing image importance maps and tracking visual attention. ACM Transactions on Computer-Human Interaction (TOCHI), 24(5):1-40, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Christof Koch and Shimon Ullman. Shifts in selective visual attention: towards the underlying neural circuitry. In Matters of intelligence, pages 115-141. Springer, 1987. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23290" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "text", + "content": "[29] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[30] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze i: Boosting saliency prediction with feature maps trained onImagenet. arXiv preprint arXiv:1411.1045, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[31] Matthias Kummerer, Thomas SA Wallis, Leon A Gatys, and Matthias Bethge. Understanding low-and high-level contributions to fixation prediction. In Proceedings of the IEEE International Conference on Computer Vision, pages 4789-4798, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[32] Qixia Lai, Salman Khan, Yongwei Nie, Hanqiu Sun, Jianbing Shen, and Ling Shao. Understanding more about human and machine attention in deep neural networks. IEEE Transactions on Multimedia, 23:2086-2099, 2020. 2, 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[33] Thomas Langlois, Haicheng Zhao, Erin Grant, Ishita Dasgupta, Tom Griffiths, and Nori Jacoby. Passive attention in artificial neural networks predicts human visual selectivity. Advances in Neural Information Processing Systems, 34, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 326 + ], + "type": "text", + "content": "[34] J. Lou, H. Lin, D. Marshall, D. Saupe, and H. Liu. Transal-net: Towards perceptually relevant visual saliency prediction. Neurocomputing, 494:455-467, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 371 + ], + "type": "text", + "content": "[35] Sangwoo Mo, Hyunwoo Kang, Kihyuk Sohn, Chun-Liang Li, and Jinwoo Shin. Object-aware contrastive learning for debiased scene representation. Advances in Neural Information Processing Systems, 34, 2021. 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 372, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 425 + ], + "type": "text", + "content": "[36] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10241-10247. IEEE, 2020. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[37] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 536 + ], + "type": "text", + "content": "[38] Ramprasaath R Selvaraju, Karan Desai, Justin Johnson, and Nikhil Naik. Casting your model: Learning to localize improves self-supervised representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11058-11067, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 602 + ], + "type": "text", + "content": "[39] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a hint: Leveraging explanations to make vision and language models more grounded. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2591-2600, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 647 + ], + "type": "text", + "content": "[40] Nikhil V Shirahatti and Kobus Barnard. Evaluating image retrieval. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 1, pages 955-961. IEEE, 2005. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[41] Ekta Sood, Fabian Kögel, Florian Strohm, Prajit Dhar, and Andreas Bulling. Vqa-mhug: A gaze dataset to study multimodal neural attention in visual question answering. arXiv preprint arXiv:2109.13116, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 597 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "type": "text", + "content": "[43] Nachiappan Valliappan, Na Dai, Ethan Steinberg, Junfeng He, Kantwon Rogers, Venky Ramachandran, Pingmei Xu, Mina Shojaeizadeh, Li Guo, Kai Kohlhoff, et al. Accelerating eye movement research via accurate and affordable smartphone eye tracking. Nature communications, 11(1):1-12, 2020. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[44] Leonard Elia Van Dyck, Roland Kwitt, Sebastian Jochen Denzler, and Walter Roland Gruber. Comparing object recognition in humans and deep convolutional neural networks—an eye tracking study. Frontiers in Neuroscience, page 1326, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 239 + ], + "type": "text", + "content": "[45] Brian A Wandell. Foundations of vision. Sinauer Associates, 1995. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 297 + ], + "type": "text", + "content": "[46] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 352 + ], + "type": "text", + "content": "[47] Peter West, Chandra Bhagavatula, Jack Hessel, Jena D Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. arXiv preprint arXiv:2110.07178, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "text", + "content": "[48] Jeremy M Wolfe and Todd S Horowitz. What attributes guide the deployment of visual attention and how do they do it? Nature reviews neuroscience, 5(6):495-501, 2004. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "text", + "content": "[49] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 442, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 487 + ], + "type": "text", + "content": "[50] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 487, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 541 + ], + "type": "text", + "content": "[51] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 597 + ], + "type": "text", + "content": "[52] Jan Zimmermann, Yuriria Vazquez, Paul W Glimcher, Bijan Pesaran, and Kenway Louie. Oculomatic: high speed, reliable, and accurate open-source eye tracking for humans and non-human primates. Journal of neuroscience methods, 270:138-146, 2016. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23291" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_content_list.json b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4355268da364234114933a168c88b9ea73e5a408 --- /dev/null +++ b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_content_list.json @@ -0,0 +1,1719 @@ +[ + { + "type": "text", + "text": "Teaching Matters: Investigating the Role of Supervision in Vision Transformers", + "text_level": 1, + "bbox": [ + 81, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthew Walmer* Saksham Suri* Kamal Gupta Abhinav Shrivastava University of Maryland, College Park", + "bbox": [ + 189, + 171, + 779, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 267, + 242, + 346, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have gained significant popularity in recent years and have proliferated into many applications. However, their behavior under different learning paradigms is not well explored. We compare ViTs trained through different methods of supervision, and show that they learn a diverse range of behaviors in terms of their attention, representations, and downstream performance. We also discover ViT behaviors that are consistent across supervision, including the emergence of Offset Local Attention Heads. These are self-attention heads that attend to a token adjacent to the current token with a fixed directional offset, a phenomenon that to the best of our knowledge has not been highlighted in any prior work. Our analysis shows that ViTs are highly flexible and learn to process local and global information in different orders depending on their training method. We find that contrastive self-supervised methods learn features that are competitive with explicitly supervised features, and they can even be superior for part-level tasks. We also find that the representations of reconstruction-based models show non-trivial similarity to contrastive self-supervised models.", + "bbox": [ + 75, + 273, + 473, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 611, + 209, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of Computer Vision has advanced massively in the past decade, largely built on the backbone of Convolutional Neural Networks (CNNs). More recently, Vision Transformers (ViTs) [18] have shown the potential to overtake CNNs as the go-to visual processing model. Prior works have asked the question do ViTs see like CNNs do? [52], but in this work, we ask: how do ViTs learn under different supervision? Past examinations of ViTs have largely focused on models trained through full supervision. Instead, we aim to characterize the differences and similarities of ViTs trained through varying training methods, including self-supervised methods. Unlike CNNs, the ViT architecture imposes few structural biases to guide the learning of representations. This gives them the flexibility to", + "bbox": [ + 75, + 637, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab9b2c8aacf56c3858e9ca9f701ed5374df875564c974002b277ec2cadf39e2e.jpg", + "image_caption": [ + "Figure 1. ViTs exhibit highly varied behaviors depending on their method of training. In this work, we compare ViTs through three domains of analysis representing the How, What, and Why of ViTs. How do ViTs process information through attention? (Top) Attention maps averaged over 5000 images show clear differences in the mid-to-late layers. What do ViTs learn to represent? (Left) Contrastive self-supervised ViTs have a greater feature similarity to explicitly supervised ViTs, but also have some similarity with ViTs trained through masked reconstruction. Why do we care about using ViTs? (Right) We evaluate ViTs on a variety of global and local tasks and show that the best model and layer vary greatly." + ], + "image_footnote": [], + "bbox": [ + 501, + 244, + 893, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learn diverse information processing strategies, and through our analyses, we uncover a wide array of ViT behaviors.", + "bbox": [ + 496, + 777, + 890, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are countless ways to analyze ViTs, so to guide this analysis we choose three major domains which correspond to the How, What, and Why of ViTs. For the How, we focus on how ViTs process information through Attention. Multi-Headed Attention (MHA) layers are arguably the key element of ViTs, and they most distinguish them", + "bbox": [ + 496, + 810, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributors. Web: www.cs.umd.edu/~sakshams/vit_analysis Code: www.github.com/mwalmer-umd/vit_analysis", + "bbox": [ + 94, + 862, + 447, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7486", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from CNNs. For the What, we examine the Features of ViTs, as these are typically what practitioners take away from them. Finally for the Why, we focus on Downstream Tasks, which are why we care about using ViTs.", + "bbox": [ + 76, + 90, + 470, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work unveils that a powerful aspect of the ViT architecture is its local-global dual nature, which plays a role in all three aspects of our analyses. While standard CNNs are restricted to building representations hierarchically from local to global, in a ViT each token can attend to information from any other image region at any time. And unlike popular CNN modifications like Spatial Pyramids [20, 28, 33, 35] and top-down strategies [6, 47, 56], ViTs have the freedom to decide when and where global information should be integrated. In this study, we show that the order and the relative ratio of local and global attention in ViTs varies dramatically based on the method of supervision. We also find clearly different trends in the allocation of attention in the mid-to-late layers of these networks, as highlighted in Figure 2. This local-global dual nature is also embedded into the structure and features of the ViT, which encodes both local spatial tokens and a non-local classifier (CLS) token throughout its entire depth. We analyze the features of ViTs for both the CLS and spatial tokens, and assess how they align with semantics at the image, object, part, and pixel-level. We perform this analysis at every layer of the ViT to show the emergence of different levels of semantic information. Finally, we assess ViTs on a number of local and global downstream tasks.", + "bbox": [ + 76, + 154, + 472, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions are: [1] A detailed comparison of ViTs trained with six different methods, including both fully supervised and self-supervised training. [2] A cross-cutting analysis spanning three major domains: Attention, Features, and Downstream Tasks. [3] Multiple insights into the inner workings of ViTs to guide future development of ViT variants, training strategies, and applications.", + "bbox": [ + 76, + 518, + 468, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we summarize some of our key observations about ViT behavior: [1] The attention maps of explicitly supervised ViTs devolve into Sparse Repeating Patterns in the mid-to-late layers, but the quality of features continues to improve in these layers (Section 4.1). [2] All ViTs studied learn to use Offset Local Attention Heads, suggesting they are fundamentally necessary in ViTs (Section 4.2). To the best of our knowledge, no prior work has brought attention to this phenomenon. [3] ViTs learn to process local and global information in different orders depending on their method of supervision (Section 4.3). [4] All ViTs studied differentiate salient foreground objects by the early-to-mid layers (Section 4.4). [5] Reconstruction-based self-supervised methods can learn semantically meaningful CLS representations, even when the CLS token is only a placeholder (Section 5.1, 5.2). [6] Supervised method's features are the most semantically rich, but contrastive self-supervised methods are comparable or even superior in", + "bbox": [ + 76, + 628, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0962b2816c9e68af5a40832021161591bc7b9c2c8d41a895f94f74bd226f5e7.jpg", + "image_caption": [ + "Average CLS Token Attention over 5000 Images" + ], + "image_footnote": [], + "bbox": [ + 500, + 107, + 893, + 296 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/413a1b747a127219530b61b87f06f3a862f5bb7f2f70ca9a26d4f3cec6666878.jpg", + "image_caption": [ + "FS & CLIP Layers 7-12: Sparse Repeating Patterns" + ], + "image_footnote": [], + "bbox": [ + 527, + 313, + 866, + 377 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a03d89c8f641acb2a133b50c97e0b77506e5d15444b4fa5d766667252d6ca9a7.jpg", + "image_caption": [ + "DINO & MoCo Layers 7-12: Object Centered Blobs" + ], + "image_footnote": [], + "bbox": [ + 527, + 393, + 866, + 455 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac5ec47af50a17553e70f1b8dd2adb824389acb623802580934d47675354d6b5.jpg", + "image_caption": [ + "MAE & BEiT Layers 7-12: Diverse Attention Maps", + "Figure 2. Clear differences in attention emerge in the mid-to-late layers under different supervision methods. These plots show the attention maps of CLS tokens averaged over 5000 ImageNet images. Rows indicate layers and columns indicate heads. For brevity, we show only three heads per layer. The bracketed numbers in the lower half denote the layer and head." + ], + "image_footnote": [], + "bbox": [ + 527, + 474, + 866, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "some cases (Section 5.2, 5.3). [7] For localized tasks, the best performance often comes from a mid-to-late layer (Section 6.2). [8] There is no single \"best\" training method or layer for all downstream tasks (Section 6.3).", + "bbox": [ + 498, + 635, + 890, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 709, + 640, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works have attempted to understand the representation quality for both supervised and self-supervised training for Convolutional Neural Networks (CNNs). [5] focuses on understanding the concepts learned by individual neurons while [41] looks at explaining their compositionality in the case of supervised networks. Simultaneously, due to the popularity of self-supervised learning methods [3, 4, 9-11, 26, 27, 29, 40, 43, 60, 73] multiple works have analyzed these representations learned from no labels. Under this umbrella, [13, 62] tried to understand the effect of training data in terms of both the number and type", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of samples. Some works [66, 67] analyze the alignment, separability, and uniformity of features while [49] looks at invariance to augmentations like occlusion, illumination, and viewpoint change in the learned representation. [64] looks at the downstream performance of self-supervised networks on fine-grained tasks. Finally, [21, 23, 32] analyze multiple self-supervised methods and compare their performance based on representation similarity and downstream task performance over multiple datasets along with comparisons to supervised methods.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the proliferation of ViTs, a number of works have tried to understand and explore the different properties of the representations learned by these networks. A few works [45, 55, 75] have analyzed the robustness of ViT features against corruptions, perturbations, distribution shifts, and adversarial examples while also analyzing the role of self-attention for robustness. [34] benchmarks different pretrained ViTs as backbones for object detection. [7] provides a theoretical understanding of how MAEs work while [53] analyzes attention using convex duality. [61] gives insights to train and use ViTs more efficiently. [44] gives a deeper understanding of how Multi-Headed Attention layers work while comparing and contrasting to how convolution layers behave in terms of the loss landscapes and low-pass/high-pass filtering. [52] compares fully supervised ViTs and ResNets in terms of the local and global information encoded at different depths, the role of skip connections, and the uniformity of representations.", + "bbox": [ + 75, + 242, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All these prior works either examine the impact of supervision on CNNs or compare CNNs and ViTs trained with full supervision. Some recent and concurrent works have compared the properties of differently supervised ViTs, though typically focused on a particular task and only two methods of supervision at a time. [1] compares the properties of fully supervised and DINO ViT features in the context of dense feature descriptors, and [2] further compares these two across several semantic correspondence tasks. [19] compares fully supervised and CLIP ViTs through feature visualizations. To the best of our knowledge, we present what is to date the broadest and the most in-depth comparison of ViTs with varying supervision, including six different methods covering three supervision subcategories. Additionally, we propose new attention-based analysis methods along with evaluations on multiple downstream tasks focused on both local and global information.", + "bbox": [ + 75, + 516, + 472, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Experimental Design", + "text_level": 1, + "bbox": [ + 76, + 789, + 277, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. A Primer on Vision Transformers", + "text_level": 1, + "bbox": [ + 76, + 815, + 372, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) [18] are adapted from Transformers [65] for the Natural Language Processing domain. A ViT consists of an array of tokens, each representing an image patch. In addition, most ViTs include an ex", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tra \"classifier\" or \"CLS\" token, which is connected to the task-specific output layers during training. ViTs use MultiHeaded Attention (MHA) layers [65], which use a Query-Key-Value system that allows each token to attend to all other tokens with a variable intensity attention map. This is in stark contrast to the limited receptive fields of convolutions. These layers are \"multi-headed\" because they repeat this process multiple times in parallel, allowing tokens to apply multiple attention strategies concurrently. A ViT architecture includes multiple blocks, each with one MHA layer followed by a position-wise fully connected layer. Unlike CNNs, which usually get narrower in deeper layers, ViTs maintain the same \"width\" (number of tokens) throughout. There are some transformer variants, like SWiN transformers [37], that introduce a narrowing width, but for our analyses, we focus on only traditional ViTs. Specifically, our primary analysis focuses on ViT-Base models with patch size $16 \\times 16$ (ViT-B/16) and input size $224 \\times 224$ , which results in a $14 \\times 14$ spatial token array. ViT-Base has 12 blocks and 12 attention heads per MHA layer. In the Appendix, we provide additional results on a wider range of ViTs, including variations in architecture size and patch size.", + "bbox": [ + 496, + 90, + 893, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Methods of Supervision", + "text_level": 1, + "bbox": [ + 498, + 450, + 718, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although a large number of ViT training methods have been proposed in a short span of three years, many of the most popular methods can be loosely categorized into the following three groups. From each group, we select two representative models for in-depth analysis. We further discuss these models' details in Appendix B.2.", + "bbox": [ + 496, + 474, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Explicit Supervision. These models are trained with an explicit objective that is defined either by human annotations or by labels derived from another source, like paired image captions. For this category, we use a Fully Supervised (FS) ViT pretrained on ImageNet21k and fine-tuned on ImageNet1k [58, 69], as well as a CLIP ViT [51].", + "bbox": [ + 496, + 566, + 890, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervision (Contrastive). Self-supervised learning methods broadly attempt to train a model through a pretext task that can be directly derived from the input data. Among the more popular pretext tasks are contrastive learning methods [27, 70] which generally present a model with multiple augmented views of the same image alongside distractor views of other images. The model must learn to identify which of the views came from the same image. For this category, we select DINO [9] and MoCo-v3 [12] which we denote simply as MoCo for the rest of this paper.", + "bbox": [ + 496, + 657, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervision (Reconstruction). Another popular category of self-supervision is reconstruction methods, which train models to predict the missing content from masked or otherwise corrupted images. We select MAE [26] and BEiT [4] for this category. Note that MAE has a separate decoder which is discarded after pretraining,", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while BEiT's decoder is learned in the same ViT. This has a strong impact on the behavior of the later layers of BEiT.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Datasets", + "text_level": 1, + "bbox": [ + 76, + 128, + 179, + 142 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We study the ViTs on multiple datasets and downstream tasks. Unless otherwise specified, we use ImageNet-50 [63], a subset of ImageNet [16] which narrows the dataset down to 50 representative categories. We sample 100 images per class to create a diverse collection of 5000 images. We additionally use PartImageNet [25] to measure Attention Saliency and part-level feature purity, as well as COCO [36] to measure object-level feature purity. We use revisited [50] Oxford [46] (ROxford5k) for evaluating image retrieval, DAVIS [48] for video segmentation, and SPair-71k [39] for keypoint correspondence.", + "bbox": [ + 75, + 152, + 470, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Proposed Analyses", + "text_level": 1, + "bbox": [ + 76, + 325, + 256, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our analysis is broadly divided into three domains covering the How, What, and Why of ViTs:", + "bbox": [ + 75, + 349, + 468, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How ViTs process local/global information (Attention). Do self-attention heads learn to operate in different ways depending on their method of training? Are there distinctive modes of attention behavior? How does supervision impact the processing order of local and global information?", + "bbox": [ + 75, + 378, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What we take away from ViTs (Features). How do the final and intermediate representations of a ViT change depending on the method of supervision? Are these trends similar or different for CLS vs. spatial tokens?", + "bbox": [ + 76, + 455, + 468, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Why we use ViTs (Downstream Tasks). Which forms of supervision are best suited for different downstream tasks? Which layers of a ViT produce features that are best for different local and global tasks?", + "bbox": [ + 76, + 515, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Attention Analysis", + "text_level": 1, + "bbox": [ + 76, + 587, + 256, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Headed Attention layers are one of the defining components of the Transformer architecture, and the attention maps they generate can give key insights into what is similar or different about ViTs trained through different methods. We perform an in-depth examination of the self-attention maps of ViT-B/16 models at every layer. Through this study, we uncover a diverse range of attention head behavioral modalities. Additional visualizations are provided in Appendix C for a wide range of ViT variants.", + "bbox": [ + 75, + 613, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Attention Visualizations", + "text_level": 1, + "bbox": [ + 76, + 756, + 297, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start by examining the attention maps of the CLS tokens of each head and layer. To gain a comprehensive understanding of each head's behavior, we compute the average attention maps over 5000 ImageNet images, as shown in Figure 2. For brevity, we display only three heads per layer, but complete plots can be found in Appendix C.1, along with additional visualizations for spatial token attention and individual input images.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aligned Aggregated Attention Maps for Spatial Tokens", + "bbox": [ + 522, + 90, + 869, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Strict Local Attention Heads", + "bbox": [ + 606, + 107, + 785, + 119 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2327a29ce828d09acaa5337f1667f8186e3f62e1ace66c919e1716603c7f88a1.jpg", + "image_caption": [ + "FS", + "CLIP", + "[1,1]", + "1", + "[1,6]", + "1", + "[2,4]", + "1", + "[2,9]", + "BE" + ], + "image_footnote": [], + "bbox": [ + 527, + 119, + 580, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/638697a81c7f1ec96804679e5d95892a014916dc08ae679fcf62a33ca8ba2f6c.jpg", + "image_caption": [ + "[1,7]", + "Soft Local Attention Heads", + "FS", + "[3,6]", + "CLIP", + "[2,3]", + "1", + "[2,9]", + "1" + ], + "image_footnote": [], + "bbox": [ + 527, + 199, + 580, + 238 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c8d5b16488fa968162a8203e7ac1522c6c5e164aa13a3baec91f99eccd3538c.jpg", + "image_caption": [ + "Axial Local Attention Heads", + "FS", + "[5,7]", + "CLIP" + ], + "image_footnote": [], + "bbox": [ + 527, + 277, + 580, + 318 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/943e69af392a718b2d3d9f4784ac56a586ce68dee3c1eef5e0322116f2881348.jpg", + "image_caption": [ + "FS", + "[3,7]", + "Figure 3. Multiple distinct forms of local attention exist. We visualize spatial token attention using Aligned Aggregated Attention Maps, and highlight different types of local attention heads, including Strict, Soft, Axial, and Offset Local Attention Heads. In row 4 we draw the mid-lines in red as a visual aid." + ], + "image_footnote": [], + "bbox": [ + 527, + 358, + 580, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/acaeed370ca87bcd31f1f12c3acd942fd369e2ba1a50a0023828a5bb72e01d14.jpg", + "image_caption": [ + "[5,6]", + "Offset Local Attention Heads", + "CLIP", + "[4,1]" + ], + "image_footnote": [], + "bbox": [ + 586, + 358, + 637, + 397 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6f76749a2f6c951fa1e3ae4683153f17b85ca119cfb3f9f2a216c27de396b6bd.jpg", + "image_caption": [ + "1", + "[6,3]", + "DINO", + "[4,1]" + ], + "image_footnote": [], + "bbox": [ + 640, + 358, + 692, + 397 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/83149018e84bcd3ef1fd80743dad9eedc28d1b7f111ec4542669e37aebc9e374.jpg", + "image_caption": [ + "[2,6]", + "1", + "[5,5]", + "MoCo", + "[3,2]" + ], + "image_footnote": [], + "bbox": [ + 697, + 358, + 723, + 397 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64803b7c3d725314a70336e42453d4b16bd2dd785ad483cd5d697e74382b63ed.jpg", + "image_caption": [ + "1", + "[1,12]", + "1", + "[6,9]", + "MAE", + "[5,5]" + ], + "image_footnote": [], + "bbox": [ + 754, + 358, + 779, + 397 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/241349858d62828debf30be886ac74d33cf3bcaed8cf4d5810ebf39e8122dda9.jpg", + "image_caption": [ + "[2,3]", + "BE", + "[3,9]", + "B", + "[3,12]", + "BEIT", + "[2,7]" + ], + "image_footnote": [], + "bbox": [ + 813, + 358, + 864, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One of the clearest differences can be seen by comparing the mid-to-final layers. For the contrastive self-supervised methods, DINO and MoCo, the attention maps tend to be centered blobs. These heads tend to focus on salient foreground objects, so these blobs simply reflect object-centered photography bias. For the reconstruction-based methods, MAE and BEiT, we see a more diverse group of attention maps. This is likely because these methods must reconstruct all image regions, and thus their attention in the final layers must be more diverse and cover more of the image. Finally, for the explicitly supervised methods, FS and CLIP, the mid-to-final layers do not focus on salient object regions and instead focus on Sparse Repeating Patterns with seemingly no spatial meaning. This occurs for both the CLS tokens and spatial tokens, and the patterns are repeated across both heads and layers. We hypothesize that these patterns occur because the mid-to-late layers are no longer focused on parsing the scene structure, and instead are using their processing power to generate their final decisions for their respective tasks. This phenomenon helps to explain why the attention maps of the later layers of fully-supervised ViTs are poorly suited for segmentation tasks, as was observed by [9].", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Emergence of Offset Local Attention Heads", + "text_level": 1, + "bbox": [ + 76, + 90, + 446, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It has been shown that ViTs use a mixture of short-range \"local\" and long-range \"global\" attention heads in any given layer [9, 18]. To gain a better understanding of local attention, we propose a visualization strategy of Aligned Aggregated Attention Maps (AAAMs). We extract all spatial token attention maps for 5000 ImageNet images, but before averaging them, we first realigned them so the current spatial token is always in the center of the array. Additional samples of AAAMs are provided in Appendix C.1. Studying these aligned views reveals multiple forms of local attention, shown in Figure 3. We find Strict Local Attention Heads, which attend almost completely to their own position, as well as Soft Local Attention Heads, which attend to a wider neighborhood around them. We also find Axial Local Attention Heads, which are elongated to attend to the local neighborhood along one or both spatial axes.", + "bbox": [ + 75, + 113, + 468, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "But perhaps the most noteworthy type of head we observe is the Offset Local Attention Head. These are heads that attend locally, but to a point or region offset from the current token in a vertical or horizontal direction. We find instances of Offset Local Attention Heads in all the models examined, suggesting they are fundamentally necessary for ViTs. To the best of our knowledge, ours is the first work to draw attention to this phenomenon. We believe that such heads are absolutely necessary because ViTs, unlike CNNs, do not have an easy built-in way to test if two features occur next to each other with a particular spatial arrangement. In a CNN, this type of check is naturally embedded into the convolution operator. But in a ViT, there is no inductive bias to induce such a check. For comparison, Soft Local Attention Heads are able to identify if a certain feature is near another feature, but they cannot identify their specific directional arrangement due to their symmetrical attention pattern. The existence of Offset Local Attention Heads implies one possible path for improvement for the ViT architecture, possibly by adding a self-attention variant that introduces some implicit directional structure.", + "bbox": [ + 75, + 356, + 470, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Average Attention Distance", + "text_level": 1, + "bbox": [ + 76, + 681, + 323, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We measure the Average Attention Distance [18, 52] of each head to assess if particular heads have a short-range \"local\" focus or a long-range \"global\" focus. This metric is computed by measuring the distance from each spatial token to all other tokens and taking a weighted average using the attention map. We normalize the distances so the token grid is embedded on a $1 \\times 1$ square. [52] observed that for a well-trained fully-supervised ViT, the early layers have a mixture of local and global attention heads, while the later layers have only global attention heads.", + "bbox": [ + 75, + 703, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 (left) shows the Average Attention Distances of all heads organized by layer and model. Like [52], we see that most layers use a mixture of local and global attention", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/55ccde2c27836ef39ead72527fc1066c5c83943dda02ad91e2253f22fb12a7cb.jpg", + "image_caption": [ + "Figure 4. Different methods of supervision lead to different orderings and ratios of local and global processing. We show the Average Attention Distance of all ViT attention heads organized by layer (left), and the per-layer averages (right)." + ], + "image_footnote": [ + "FS CLIP DINO MoCo MAE BEiT" + ], + "bbox": [ + 504, + 90, + 694, + 226 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fdb8e3d6d034a8a370d4a2dcfc18c2566db141c46eb741abad6389fe7c6a5b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 90, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MoCo MAE BEiT", + "bbox": [ + 699, + 233, + 846, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2570959cb9d1ca8b3f3f4a4f398bad321eba47e1e9c92b99a9f564b330ca5fff.jpg", + "image_caption": [ + "Figure 5. Attention IoU with salient content plateaus early for all ViTs evaluated. We calculate the alignment of ground-truth segmentation masks with CLS token attention maps (left) and the average of spatial token attention maps (right)." + ], + "image_footnote": [ + "FS CLIP DINO MoCo MAE BEiT" + ], + "bbox": [ + 503, + 328, + 696, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eb6bc27b4ec22aad2b0df63940aa8d613413300756edb761395eac987ad13526.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 328, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MoCo MAE BEiT", + "bbox": [ + 689, + 470, + 846, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "heads, however, we also find that the ordering of local and global processing varies greatly with the supervision type. FS, CLIP, DINO, and MoCo all use exclusively global attention heads in the last layers, but the reconstruction-based methods MAE and BEiT use a diverse range of heads in their later layers. Figure 4 (right) compares the combined Average Attention Distances at the per-layer level. In all models, we observe a greater number of global attention heads in the initial layers, followed by decreased distances around layers 3-6. This result is again in contrast to [52]. The behaviors diverge in the mid-to-late layers. For the models trained with explicit or contrastive supervision, the Attention Distance trends upward in the later layers. For the reconstruction-based methods, the Average Attention Distances stay lower. These results show that, unlike CNNs, ViTs can learn a variable local/global processing order depending on the training method used.", + "bbox": [ + 496, + 559, + 890, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Attention Alignment with Salient Content", + "text_level": 1, + "bbox": [ + 500, + 830, + 856, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the most desirable (and exploitable) features of DINO is that the CLS token attention maps of the last layer tend to be well-aligned with salient foreground objects [9].", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7490", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5b144029558877b69cd52d2f0d8e67bc7fb5396a933a5873dbcd7f15d3b50c6.jpg", + "image_caption": [ + "Figure 6. CKA similarity between final layer features of different ViTs for their CLS tokens (left) and spatial tokens (right)." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 272, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91542c7060b528dee1eb1f232faa662447128b040739e34a8758831a728936c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 89, + 459, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several methods propose to use DINO attention maps, feature maps, or a combination of the two to generate segmentations in a self-supervised manner [24, 57, 68]. We conduct a quantitative analysis of this property at all layers of the ViTs, both to measure the usefulness of masks and to assess how early the ViTs differentiate salient object regions. Like [9], we threshold the CLS token attention masks keeping $60\\%$ of the total attention mass. We then compute the Intersection over Union (IoU) of said masks with ground-truth segmentations from PartImageNet [25]. As an alternative to CLS token attention, we also extract masks using the average of spatial token attention maps. We present results for the single \"best\" head per layer in Figure 5.", + "bbox": [ + 75, + 268, + 468, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We see a clear drop in FS and CLIP mask IoU around the middle of the network, which directly corresponds to the emergence of the Sparse Repeating Patterns observed in Section 4.1. We also find that the IoUs plateau around layers 3-6 for all networks. This demonstrates that ViT models already have a solid understanding of foreground/background separation by the middle layers. While the later attention maps of FS and CLIP are much worse than their self-supervised counterparts, their early-to-mid layers are more comparable. We find that MoCo, MAE, and BEiT can all produce attention maps with IoUs that are comparable with DINO. In addition, we see that the average of spatial tokens produces maps that are comparable with the CLS token, and for CLIP the IoU increases greatly in the final layer.", + "bbox": [ + 75, + 465, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Feature Analysis", + "text_level": 1, + "bbox": [ + 76, + 691, + 241, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we directly compare ViT features across models and layers using Centered Kernel Alignment (CKA) [14, 31]. We also study unsupervised clustering performance to compare global and local semantic information in the learned representations. We present addition analysis in Appendix D.4 focused on ViT residual connections.", + "bbox": [ + 75, + 715, + 467, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Last Block Feature Comparisons", + "text_level": 1, + "bbox": [ + 76, + 816, + 366, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing representations is non-trivial due to varying feature sizes, large feature representations, and lack of alignment between them. To overcome this, we use batched Centered Kernel Alignment (CKA) [14, 31, 42] which can", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "align features and compute a similarity score. We compare the last layer outputs for each model.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 (left) shows that the CLS token representations are usually similar for similar supervision strategies (explicit, contrastive, reconstruction). The contrastive methods, MoCo and DINO, show very high similarity to each other, indicating that the CLS token encodes the same type of information for both these methods. There is also an increased level of similarity between the explicitly supervised methods, FS and CLIP, and the contrastive methods. Interestingly, we see that MAE has as high a similarity with DINO and MoCo as it does with BEiT. This result is surprising because MAE's CLS token has no explicit training objective or loss, and the way these approaches are trained is very different. This presents evidence that training autoencoders with a high masking percentage indeed forces the model to learn image-level semantics.", + "bbox": [ + 496, + 122, + 890, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 6 (right), we look at the similarity of the last-layer spatial token representations. Unlike the CLS token representations, CLIP and FS have low similarity in their spatial representations. The self-supervised methods DINO, MAE, and MoCo show a high level of similarity to each other, and a lower level of similarity to BEiT. MoCo and DINO show the highest similarity due to their similar kind of self-supervision. Once again, MAE has a high similarity to MoCo and DINO despite their very different supervision.", + "bbox": [ + 496, + 349, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Feature Clustering for Global Semantics", + "text_level": 1, + "bbox": [ + 500, + 497, + 844, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through this analysis, we aim to test how well the learned CLS and spatial token representations encode global (image-level) semantic information at every layer. We extract the CLS token features from the end of each block for 5000 ImageNet images, and we generate k-Means cluster assignments with $k = 50$ . We present results for cluster purity measured with respect to ground truth image labels, but additional clustering metrics are also presented in Appendix D.5. For the spatial tokens, we follow the same process except we average-pool over all positions before clustering. We also compute a random chance score by replacing the ViT features with Gaussian random noise.", + "bbox": [ + 496, + 522, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CLS token features, shown in Figure 7 (left), cluster purity improves with depth with the exception of the last layers of BEiT. This is likely because the last layers of BEiT serve as a task-specific decoder, unlike MAE, where the decoder is separate and discarded after pretraining. Unsurprisingly, FS achieves the best cluster purity, followed by CLIP. The contrastive methods, DINO and MoCo, achieve scores close to the explicitly supervised methods. The reconstruction-based methods, MAE and BEiT, have the lowest cluster purity, but they are still above random chance, which again indicates that they do learn to encode some image-level semantic information in their CLS tokens. Also, we find that semantic information emerges earlier for", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7491", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e73f13a940cb786fa28fa1b6ea032f5593708b1644091a01e6d1861956ca1ef4.jpg", + "image_caption": [ + "Figure 7. Clustering purity analysis with image-level labels in ImageNet-50 for CLS features (left) and average-pooled spatial token features (right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 472, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be76f898de49b7ef4068642da8488cc6c5316a7c131bd1cf94030f0a15a394de.jpg", + "image_caption": [ + "Figure 8. Clustering purity of spatial token features gathered at the object-level in COCO (left) and part-level in PartImageNet (right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 313, + 468, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DINO and MoCo. For the spatial token features, shown in Figure 7 (right), the cluster purity of FS rises earlier compared with the FS CLS token. This suggests that the FS spatial tokens do more work gathering semantic information in the early layers. For all other ViTs, the spatial feature purity is lower in the final layers, but is comparable in layers 1-7.", + "bbox": [ + 75, + 518, + 467, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Feature Clustering for Local Semantics", + "text_level": 1, + "bbox": [ + 76, + 619, + 415, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We next measure how well the spatial token features differentiate salient image content at the object or part-level using COCO [36] and PartImageNet [25] respectively. We use a tiling strategy to extract a denser array of features, which we detail in Appendix B.4. Using ground truth segmentation masks, we extract and average the features of the tokens overlapping with the masks. This generates a collection of object-level or part-level features which we cluster just like Section 5.2. The results for object-level features are shown in Figure 8 (left). We see that the supervised methods CLIP and FS have the highest feature purity by far, followed by the contrastive methods DINO and MoCo. The purity is much lower for the reconstruction methods MAE and BEiT. For part-level features, shown in Figure 8 (right), FS achieves the best purity, but the contrastive methods are very competitive in this case, surpassing CLIP completely. In addition, while still being the lowest scoring, MAE and", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BEiT are much more competitive at the part-level. Like the image-level purity, the object and part-level feature purity tend to improve with depth, but the purity peaks early around layers 9 to 11. The peak for BEiT is even earlier, likely due to its integrated decoder.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Downstream Task Analysis", + "text_level": 1, + "bbox": [ + 498, + 179, + 751, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we analyze the performance of these models on downstream tasks that can be performed directly without any fine-tuning or training. We follow the evaluation protocols of [9, 30] for $k$ -NN classification, image retrieval, and video object segmentation. We also perform keypoint correspondence as a more local-focused task. Again we compute random chance scores by replacing all ViT features with Gaussian noise.", + "bbox": [ + 496, + 203, + 890, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Global Tasks", + "text_level": 1, + "bbox": [ + 500, + 333, + 635, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImageNet Classification. We perform $k$ -Nearest Neighbor ( $k$ -NN) image classification on ImageNet [16] with $k = 20$ . We use the CLS token features from each network and assign the label for a test sample based on the training set features and labels. As can be seen from Figure 9 (left), FS performs the best as it has been trained to classify the same dataset. DINO and MoCo follow a similar trend as Section 5.2 and better encode semantic information in the earlier layers. FS and CLIP also follow a similar trend where their performance shoots up in the last few layers. It is also interesting to see how MAE and BEiT, for which the CLS tokens have no explicit objective, do better than chance, although MAE is considerably better than BEiT.", + "bbox": [ + 496, + 354, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Retrieval. Similar to $k$ -NN classification, we utilize the CLS token representation for retrieval. We evaluate on ROxford5k [50] for the Medium (M) split and report the Mean Average Precision (mAP). In Appendix E.2 we also report results for the Hard (H) split and the RParis6k [50] dataset, which follow similar trends. The results, shown in Figure 9 (right), align closely with those for $k$ -NN Classification on ImageNet. FS performs the best followed by CLIP and then DINO and MoCo, and finally by MAE and BEiT with the lowest performance. We hypothesize that the local/global crops used in DINO training help it perform competitively in these global tasks.", + "bbox": [ + 496, + 551, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Local Tasks", + "text_level": 1, + "bbox": [ + 500, + 741, + 627, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DAVIS Segmentation Propagation. DAVIS Segmentation Propagation is a dense prediction-based video localized task where frame-by-frame features are used to propagate the first frame segmentation mask to subsequent frames. Like Section 5.3, we use a tiling-based dense feature extraction strategy. Results are shown in Figure 10 (left). The contrastive techniques of DINO and MoCo perform the best while FS and CLIP, which are more image-level approaches, face a drop in performance towards the", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7492", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/936de7d6abd4083539243b9877b2cb9bb09286027b82a12a444d544ece520940.jpg", + "image_caption": [], + "image_footnote": [ + "FS CLIP DINO MoCo MAE BEiT -Random" + ], + "bbox": [ + 81, + 89, + 276, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/041bc83afe63922ed3a6d5ef14d74430bbf7b41f1c551cdcaca60c6a2676ff2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 90, + 467, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58ee801ec5b6cd8d9c3532baf64dbd31e0342dd1f2364439adb33a36b62194a4.jpg", + "image_caption": [ + "Figure 9. Global (image-level) downstream task analysis using the CLS token. We present $k$ -NN classifier Top-1 Accuracy on ImageNet (left) and image retrieval mAP on ROxford5k (right)." + ], + "image_footnote": [ + "FS CLIP DINO MoCo MAE BEiT -Random" + ], + "bbox": [ + 81, + 303, + 274, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffe40f7ed68cbe91e11cf0c357b9c12406812438ead6e24deaa1549715b261cc.jpg", + "image_caption": [ + "Figure 10. Local (pixel-level) downstream task analysis using the dense spatial token features. We perform DAVIS video segmentation (left) and SPair-71k keypoint correspondence (right)." + ], + "image_footnote": [], + "bbox": [ + 276, + 303, + 467, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "later layers. The local reconstruction-based methods, MAE and BEiT, are also much more competitive in this task. These results show that for purposes of constructing highly descriptive local features, contrastive methods like DINO and MoCo and reconstructive methods like BEiT can surpass features trained with explicit image-level supervision.", + "bbox": [ + 75, + 520, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Keypoint Correspondence. We choose keypoint correspondence as an additional local-focused downstream task. Given an image with annotated keypoints, the model must predict the position of corresponding keypoints in a paired image with similar content. We use the SPair-71k [39] dataset and follow the evaluation protocol of [1] and report the Percentage of Correct Keypoints (PCK) [72]. The results are summarized in Figure 10 (right). CLIP excels at this task, closely followed by both FS and DINO. Meanwhile, MoCo, BEiT, and MAE are all very competitive also. The position of the best layer varies significantly, from 8 for CLIP and BEiT to 11 for MAE.", + "bbox": [ + 75, + 611, + 468, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Summary of Downstream Tasks", + "text_level": 1, + "bbox": [ + 76, + 801, + 357, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We summarize the best results for all downstream tasks in Table 1. We denote the best-performing layers in parenthesis. These results show that ViTs with different supervision methods [1] peak at different layers, and [2] perform best at different tasks. In image-level tasks like $k$ -NN clas", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f865bc0be458f9f2804cd846f54241265695f52d5be51785d91988b105a83e85.jpg", + "table_caption": [ + "Table 1. Best performance for each ViT on each downstream task with the corresponding best layer in parenthesis." + ], + "table_footnote": [], + "table_body": "
ModelTask Performance (Best Performing Layer)
Dataset MetricImageNet Top-1↑ROxford5k (M) mAP↑Davis J and F Mean↑SPair-71k PCK@0.1↑
FS83.79 (12)0.45 (12)0.59 (8)28.56 (9)
CLIP75.75 (12)0.40 (12)0.60 (9)30.70 (8)
DINO76.06 (12)0.37 (12)0.60 (12)28.28 (9)
MoCo71.59 (12)0.31 (12)0.61 (11)25.85 (9)
MAE45.19 (12)0.15 (10)0.54 (12)22.74 (11)
BEiT26.84 (8)0.14 (8)0.58 (9)24.11 (8)
Random0.100.020.061.32
", + "bbox": [ + 501, + 128, + 890, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sification and retrieval, usually, the last layer works the best. For localized tasks like keypoint correspondence and video object segmentation, most models' peak performance happens a few layers before the last one. This shows that always picking the last layer output is not optimal.", + "bbox": [ + 496, + 292, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 396, + 617, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have performed an in-depth comparison of ViTs trained through different methodologies by examining their attention patterns, learned representations, and downstream task performance. We review some of the key findings of our analyses. First, different methods of supervision lead to ViTs that process local and global information in different orders. All ViTs have heads that align well with salient image content, but for the explicitly supervised models, the late-layer attention maps change into Sparse Repeating Patterns. In addition, all ViTs examined have learned to use Offset Local Attention Heads in multiple layers. While explicitly supervised ViTs have the most semantically rich representations at the image level, contrastive methods are competitive, and reconstruction-based methods can also learn meaningful CLS token representations even though said token is a placeholder and has no explicit supervision. Finally, there is no single best model for all the downstream tasks, and the best layer to extract representations from also varies greatly by task and model, so one should not simply take the last layer representation. ViTs have shown a great deal of potential, and we expect they will become more widely used in the coming years. We hope these insights can help with the future development of losses and architectures for Vision Transformers.", + "bbox": [ + 496, + 426, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and DARPA GARD (HR00112020007) programs. We would like to thank our colleagues Vatsal Agarwal, Matthew Gwilliam, Alex Hanson, Pulkit Kumar, Saketh Rambhatla, and Gaurav Shrivastava for their feedback on this work.", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7493", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 3, 8, 24", + "[2] Mehmet Aygün and Oisin Mac Aodha. Demystifying unsupervised semantic correspondence estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 125-142. Springer, 2022. 3", + "[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2", + "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3, 12, 14", + "[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2", + "[6] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 2956-2964, 2015. 2", + "[7] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022. 3", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 14", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 12", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 3, 12", + "[13] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 2", + "[14] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment." + ], + "bbox": [ + 78, + 114, + 470, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The Journal of Machine Learning Research, 13:795-828, 2012. 6, 18", + "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 12", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 7", + "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 14", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 5", + "[19] Amin Ghiasi, Hamid Kazemi, Eitan Borgnia, Steven Reich, Manli Shu, Micah Goldblum, Andrew Gordon Wilson, and Tom Goldstein. What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727, 2022. 3", + "[20] Kristen Grauman and Trevor Darrell. The pyramid match kernel: Discriminative classification with sets of image features. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1458-1465. IEEE, 2005. 2", + "[21] Tom George Grigg, Dan Busbridge, Jason Ramapuram, and Russ Webb. Do self-supervised and supervised methods learn similar visual representations? arXiv preprint arXiv:2110.00528, 2021. 3", + "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 14", + "[23] Matthew Gwilliam and Abhinav Shrivastava. Beyond supervised vs. unsupervised: Representative benchmarking and analysis of image representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9642-9652, 2022. 3", + "[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 6", + "[25] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021. 4, 6, 7, 18", + "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3, 12, 14", + "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross" + ], + "bbox": [ + 503, + 92, + 893, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7494", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 12", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2", + "[29] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2", + "[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7", + "[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 6, 18", + "[32] Klemen Kotar, Gabriel Ilharco, Ludwig Schmidt, Kiana Ehsani, and Roozbeh Mottaghi. Contrasting contrastive self-supervised representation learning pipelines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9949-9959, 2021. 3", + "[33] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 2169-2178. IEEE, 2006. 2", + "[34] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3", + "[35] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 7, 18", + "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3", + "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 12", + "[39] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 4, 8, 24", + "[40] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 892 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2", + "[42] Thao Nguyen, Maithra Raghu, and Simon Kornblith. Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth. arXiv preprint arXiv:2010.15327, 2020. 6, 18", + "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[44] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 3", + "[45] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2071-2081, 2022. 3", + "[46] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 4", + "[47] Pedro O Pinheiro, Tsung-Yi Lin, Ronan Collobert, and Piotr Dóllár. Learning to refine object segments. In European conference on computer vision, pages 75-91. Springer, 2016. 2", + "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4", + "[49] Senthil Purushwalkam and Abhinav Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. Advances in Neural Information Processing Systems, 33:3407-3418, 2020. 3", + "[50] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5706-5715, 2018. 4, 7", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12", + "[52] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 1, 3, 5, 23", + "[53] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. arXiv preprint arXiv:2205.08078, 2022. 3", + "[54] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 24", + "[55] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 891 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7495", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Abhinav Shrivastava, Rahul Sukthankar, Jitendra Malik, and Abhinav Gupta. Beyond skip connections: Top-down modulation for object detection. arXiv preprint arXiv:1612.06851, 2016. 2", + "[57] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 6", + "[58] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 3, 12", + "[59] Anand Subramanian. Torch cka. https://github.com/AntixK/PyTorch-Model-Compare, 2021.18", + "[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2", + "[61] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 3", + "[62] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34:16238-16250, 2021. 2", + "[63] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 4", + "[64] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking representation learning for natural world image collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12884-12893, 2021. 3", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[66] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3", + "[67] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3", + "[68] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 6", + "[69] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.3, 12", + "[70] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin." + ], + "bbox": [ + 78, + 90, + 468, + 890 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3", + "[71] Kun Xu, Yong Li, Tao Ju, Shi-Min Hu, and Tian-Qiang Liu. Efficient affinity-based edit propagation using kd tree. ACM Transactions on Graphics (TOG), 28(5):1-6, 2009. 24", + "[72] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 8, 24", + "[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2", + "[74] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 12", + "[75] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In International Conference on Machine Learning, pages 27378-27394. PMLR, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 402 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7496", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_model.json b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_model.json new file mode 100644 index 0000000000000000000000000000000000000000..580a5752739f335b3971b5d3805f3a1da536deec --- /dev/null +++ b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_model.json @@ -0,0 +1,3104 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.131, + 0.884, + 0.153 + ], + "angle": 0, + "content": "Teaching Matters: Investigating the Role of Supervision in Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.172, + 0.781, + 0.208 + ], + "angle": 0, + "content": "Matthew Walmer* Saksham Suri* Kamal Gupta Abhinav Shrivastava University of Maryland, College Park" + }, + { + "type": "title", + "bbox": [ + 0.269, + 0.243, + 0.347, + 0.258 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.474, + 0.593 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have gained significant popularity in recent years and have proliferated into many applications. However, their behavior under different learning paradigms is not well explored. We compare ViTs trained through different methods of supervision, and show that they learn a diverse range of behaviors in terms of their attention, representations, and downstream performance. We also discover ViT behaviors that are consistent across supervision, including the emergence of Offset Local Attention Heads. These are self-attention heads that attend to a token adjacent to the current token with a fixed directional offset, a phenomenon that to the best of our knowledge has not been highlighted in any prior work. Our analysis shows that ViTs are highly flexible and learn to process local and global information in different orders depending on their training method. We find that contrastive self-supervised methods learn features that are competitive with explicitly supervised features, and they can even be superior for part-level tasks. We also find that the representations of reconstruction-based models show non-trivial similarity to contrastive self-supervised models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.612, + 0.21, + 0.628 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.851 + ], + "angle": 0, + "content": "The field of Computer Vision has advanced massively in the past decade, largely built on the backbone of Convolutional Neural Networks (CNNs). More recently, Vision Transformers (ViTs) [18] have shown the potential to overtake CNNs as the go-to visual processing model. Prior works have asked the question do ViTs see like CNNs do? [52], but in this work, we ask: how do ViTs learn under different supervision? Past examinations of ViTs have largely focused on models trained through full supervision. Instead, we aim to characterize the differences and similarities of ViTs trained through varying training methods, including self-supervised methods. Unlike CNNs, the ViT architecture imposes few structural biases to guide the learning of representations. This gives them the flexibility to" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.246, + 0.895, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.615, + 0.895, + 0.77 + ], + "angle": 0, + "content": "Figure 1. ViTs exhibit highly varied behaviors depending on their method of training. In this work, we compare ViTs through three domains of analysis representing the How, What, and Why of ViTs. How do ViTs process information through attention? (Top) Attention maps averaged over 5000 images show clear differences in the mid-to-late layers. What do ViTs learn to represent? (Left) Contrastive self-supervised ViTs have a greater feature similarity to explicitly supervised ViTs, but also have some similarity with ViTs trained through masked reconstruction. Why do we care about using ViTs? (Right) We evaluate ViTs on a variety of global and local tasks and show that the best model and layer vary greatly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.809 + ], + "angle": 0, + "content": "learn diverse information processing strategies, and through our analyses, we uncover a wide array of ViT behaviors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.901 + ], + "angle": 0, + "content": "There are countless ways to analyze ViTs, so to guide this analysis we choose three major domains which correspond to the How, What, and Why of ViTs. For the How, we focus on how ViTs process information through Attention. Multi-Headed Attention (MHA) layers are arguably the key element of ViTs, and they most distinguish them" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.448, + 0.901 + ], + "angle": 0, + "content": "*Equal contributors. Web: www.cs.umd.edu/~sakshams/vit_analysis Code: www.github.com/mwalmer-umd/vit_analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7486" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "from CNNs. For the What, we examine the Features of ViTs, as these are typically what practitioners take away from them. Finally for the Why, we focus on Downstream Tasks, which are why we care about using ViTs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.473, + 0.517 + ], + "angle": 0, + "content": "Our work unveils that a powerful aspect of the ViT architecture is its local-global dual nature, which plays a role in all three aspects of our analyses. While standard CNNs are restricted to building representations hierarchically from local to global, in a ViT each token can attend to information from any other image region at any time. And unlike popular CNN modifications like Spatial Pyramids [20, 28, 33, 35] and top-down strategies [6, 47, 56], ViTs have the freedom to decide when and where global information should be integrated. In this study, we show that the order and the relative ratio of local and global attention in ViTs varies dramatically based on the method of supervision. We also find clearly different trends in the allocation of attention in the mid-to-late layers of these networks, as highlighted in Figure 2. This local-global dual nature is also embedded into the structure and features of the ViT, which encodes both local spatial tokens and a non-local classifier (CLS) token throughout its entire depth. We analyze the features of ViTs for both the CLS and spatial tokens, and assess how they align with semantics at the image, object, part, and pixel-level. We perform this analysis at every layer of the ViT to show the emergence of different levels of semantic information. Finally, we assess ViTs on a number of local and global downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Overall, our contributions are: [1] A detailed comparison of ViTs trained with six different methods, including both fully supervised and self-supervised training. [2] A cross-cutting analysis spanning three major domains: Attention, Features, and Downstream Tasks. [3] Multiple insights into the inner workings of ViTs to guide future development of ViT variants, training strategies, and applications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In addition, we summarize some of our key observations about ViT behavior: [1] The attention maps of explicitly supervised ViTs devolve into Sparse Repeating Patterns in the mid-to-late layers, but the quality of features continues to improve in these layers (Section 4.1). [2] All ViTs studied learn to use Offset Local Attention Heads, suggesting they are fundamentally necessary in ViTs (Section 4.2). To the best of our knowledge, no prior work has brought attention to this phenomenon. [3] ViTs learn to process local and global information in different orders depending on their method of supervision (Section 4.3). [4] All ViTs studied differentiate salient foreground objects by the early-to-mid layers (Section 4.4). [5] Reconstruction-based self-supervised methods can learn semantically meaningful CLS representations, even when the CLS token is only a placeholder (Section 5.1, 5.2). [6] Supervised method's features are the most semantically rich, but contrastive self-supervised methods are comparable or even superior in" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.092, + 0.848, + 0.106 + ], + "angle": 0, + "content": "Average CLS Token Attention over 5000 Images" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.108, + 0.895, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.302, + 0.854, + 0.314 + ], + "angle": 0, + "content": "FS & CLIP Layers 7-12: Sparse Repeating Patterns" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.314, + 0.867, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.382, + 0.857, + 0.395 + ], + "angle": 0, + "content": "DINO & MoCo Layers 7-12: Object Centered Blobs" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.395, + 0.867, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.462, + 0.853, + 0.475 + ], + "angle": 0, + "content": "MAE & BEiT Layers 7-12: Diverse Attention Maps" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.475, + 0.867, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.545, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Figure 2. Clear differences in attention emerge in the mid-to-late layers under different supervision methods. These plots show the attention maps of CLS tokens averaged over 5000 ImageNet images. Rows indicate layers and columns indicate heads. For brevity, we show only three heads per layer. The bracketed numbers in the lower half denote the layer and head." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.697 + ], + "angle": 0, + "content": "some cases (Section 5.2, 5.3). [7] For localized tasks, the best performance often comes from a mid-to-late layer (Section 6.2). [8] There is no single \"best\" training method or layer for all downstream tasks (Section 6.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.641, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Previous works have attempted to understand the representation quality for both supervised and self-supervised training for Convolutional Neural Networks (CNNs). [5] focuses on understanding the concepts learned by individual neurons while [41] looks at explaining their compositionality in the case of supervised networks. Simultaneously, due to the popularity of self-supervised learning methods [3, 4, 9-11, 26, 27, 29, 40, 43, 60, 73] multiple works have analyzed these representations learned from no labels. Under this umbrella, [13, 62] tried to understand the effect of training data in terms of both the number and type" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7487" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "of samples. Some works [66, 67] analyze the alignment, separability, and uniformity of features while [49] looks at invariance to augmentations like occlusion, illumination, and viewpoint change in the learned representation. [64] looks at the downstream performance of self-supervised networks on fine-grained tasks. Finally, [21, 23, 32] analyze multiple self-supervised methods and compare their performance based on representation similarity and downstream task performance over multiple datasets along with comparisons to supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.515 + ], + "angle": 0, + "content": "Since the proliferation of ViTs, a number of works have tried to understand and explore the different properties of the representations learned by these networks. A few works [45, 55, 75] have analyzed the robustness of ViT features against corruptions, perturbations, distribution shifts, and adversarial examples while also analyzing the role of self-attention for robustness. [34] benchmarks different pretrained ViTs as backbones for object detection. [7] provides a theoretical understanding of how MAEs work while [53] analyzes attention using convex duality. [61] gives insights to train and use ViTs more efficiently. [44] gives a deeper understanding of how Multi-Headed Attention layers work while comparing and contrasting to how convolution layers behave in terms of the loss landscapes and low-pass/high-pass filtering. [52] compares fully supervised ViTs and ResNets in terms of the local and global information encoded at different depths, the role of skip connections, and the uniformity of representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.473, + 0.775 + ], + "angle": 0, + "content": "All these prior works either examine the impact of supervision on CNNs or compare CNNs and ViTs trained with full supervision. Some recent and concurrent works have compared the properties of differently supervised ViTs, though typically focused on a particular task and only two methods of supervision at a time. [1] compares the properties of fully supervised and DINO ViT features in the context of dense feature descriptors, and [2] further compares these two across several semantic correspondence tasks. [19] compares fully supervised and CLIP ViTs through feature visualizations. To the best of our knowledge, we present what is to date the broadest and the most in-depth comparison of ViTs with varying supervision, including six different methods covering three supervision subcategories. Additionally, we propose new attention-based analysis methods along with evaluations on multiple downstream tasks focused on both local and global information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.279, + 0.808 + ], + "angle": 0, + "content": "3. Experimental Design" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.373, + 0.832 + ], + "angle": 0, + "content": "3.1. A Primer on Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) [18] are adapted from Transformers [65] for the Natural Language Processing domain. A ViT consists of an array of tokens, each representing an image patch. In addition, most ViTs include an ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.44 + ], + "angle": 0, + "content": "tra \"classifier\" or \"CLS\" token, which is connected to the task-specific output layers during training. ViTs use MultiHeaded Attention (MHA) layers [65], which use a Query-Key-Value system that allows each token to attend to all other tokens with a variable intensity attention map. This is in stark contrast to the limited receptive fields of convolutions. These layers are \"multi-headed\" because they repeat this process multiple times in parallel, allowing tokens to apply multiple attention strategies concurrently. A ViT architecture includes multiple blocks, each with one MHA layer followed by a position-wise fully connected layer. Unlike CNNs, which usually get narrower in deeper layers, ViTs maintain the same \"width\" (number of tokens) throughout. There are some transformer variants, like SWiN transformers [37], that introduce a narrowing width, but for our analyses, we focus on only traditional ViTs. Specifically, our primary analysis focuses on ViT-Base models with patch size \\(16 \\times 16\\) (ViT-B/16) and input size \\(224 \\times 224\\), which results in a \\(14 \\times 14\\) spatial token array. ViT-Base has 12 blocks and 12 attention heads per MHA layer. In the Appendix, we provide additional results on a wider range of ViTs, including variations in architecture size and patch size." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.451, + 0.72, + 0.468 + ], + "angle": 0, + "content": "3.2. Methods of Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Although a large number of ViT training methods have been proposed in a short span of three years, many of the most popular methods can be loosely categorized into the following three groups. From each group, we select two representative models for in-depth analysis. We further discuss these models' details in Appendix B.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Explicit Supervision. These models are trained with an explicit objective that is defined either by human annotations or by labels derived from another source, like paired image captions. For this category, we use a Fully Supervised (FS) ViT pretrained on ImageNet21k and fine-tuned on ImageNet1k [58, 69], as well as a CLIP ViT [51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Self-Supervision (Contrastive). Self-supervised learning methods broadly attempt to train a model through a pretext task that can be directly derived from the input data. Among the more popular pretext tasks are contrastive learning methods [27, 70] which generally present a model with multiple augmented views of the same image alongside distractor views of other images. The model must learn to identify which of the views came from the same image. For this category, we select DINO [9] and MoCo-v3 [12] which we denote simply as MoCo for the rest of this paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Self-Supervision (Reconstruction). Another popular category of self-supervision is reconstruction methods, which train models to predict the missing content from masked or otherwise corrupted images. We select MAE [26] and BEiT [4] for this category. Note that MAE has a separate decoder which is discarded after pretraining," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7488" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "while BEiT's decoder is learned in the same ViT. This has a strong impact on the behavior of the later layers of BEiT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.129, + 0.18, + 0.143 + ], + "angle": 0, + "content": "3.3. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.319 + ], + "angle": 0, + "content": "We study the ViTs on multiple datasets and downstream tasks. Unless otherwise specified, we use ImageNet-50 [63], a subset of ImageNet [16] which narrows the dataset down to 50 representative categories. We sample 100 images per class to create a diverse collection of 5000 images. We additionally use PartImageNet [25] to measure Attention Saliency and part-level feature purity, as well as COCO [36] to measure object-level feature purity. We use revisited [50] Oxford [46] (ROxford5k) for evaluating image retrieval, DAVIS [48] for video segmentation, and SPair-71k [39] for keypoint correspondence." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.327, + 0.258, + 0.343 + ], + "angle": 0, + "content": "3.4. Proposed Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.379 + ], + "angle": 0, + "content": "Our analysis is broadly divided into three domains covering the How, What, and Why of ViTs:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.469, + 0.456 + ], + "angle": 0, + "content": "How ViTs process local/global information (Attention). Do self-attention heads learn to operate in different ways depending on their method of training? Are there distinctive modes of attention behavior? How does supervision impact the processing order of local and global information?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.456, + 0.469, + 0.516 + ], + "angle": 0, + "content": "What we take away from ViTs (Features). How do the final and intermediate representations of a ViT change depending on the method of supervision? Are these trends similar or different for CLS vs. spatial tokens?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.516, + 0.469, + 0.576 + ], + "angle": 0, + "content": "Why we use ViTs (Downstream Tasks). Which forms of supervision are best suited for different downstream tasks? Which layers of a ViT produce features that are best for different local and global tasks?" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.258, + 0.605 + ], + "angle": 0, + "content": "4. Attention Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Multi-Headed Attention layers are one of the defining components of the Transformer architecture, and the attention maps they generate can give key insights into what is similar or different about ViTs trained through different methods. We perform an in-depth examination of the self-attention maps of ViT-B/16 models at every layer. Through this study, we uncover a diverse range of attention head behavioral modalities. Additional visualizations are provided in Appendix C for a wide range of ViT variants." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.299, + 0.772 + ], + "angle": 0, + "content": "4.1. Attention Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We start by examining the attention maps of the CLS tokens of each head and layer. To gain a comprehensive understanding of each head's behavior, we compute the average attention maps over 5000 ImageNet images, as shown in Figure 2. For brevity, we display only three heads per layer, but complete plots can be found in Appendix C.1, along with additional visualizations for spatial token attention and individual input images." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.092, + 0.87, + 0.106 + ], + "angle": 0, + "content": "Aligned Aggregated Attention Maps for Spatial Tokens" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.108, + 0.786, + 0.12 + ], + "angle": 0, + "content": "Strict Local Attention Heads" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.12, + 0.581, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.16, + 0.562, + 0.169 + ], + "angle": 0, + "content": "FS" + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.171, + 0.568, + 0.182 + ], + "angle": 0, + "content": "[1,7]" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.16, + 0.623, + 0.17 + ], + "angle": 0, + "content": "CLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.171, + 0.624, + 0.182 + ], + "angle": 0, + "content": "[1,1]" + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.16, + 0.657, + 0.169 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.171, + 0.682, + 0.182 + ], + "angle": 0, + "content": "[1,6]" + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.16, + 0.712, + 0.169 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.171, + 0.739, + 0.182 + ], + "angle": 0, + "content": "[2,4]" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.16, + 0.772, + 0.169 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.171, + 0.796, + 0.182 + ], + "angle": 0, + "content": "[2,9]" + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.16, + 0.839, + 0.169 + ], + "angle": 0, + "content": "BE" + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.171, + 0.852, + 0.182 + ], + "angle": 0, + "content": "[2,3]" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.188, + 0.783, + 0.199 + ], + "angle": 0, + "content": "Soft Local Attention Heads" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.2, + 0.581, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.24, + 0.562, + 0.25 + ], + "angle": 0, + "content": "FS" + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.251, + 0.567, + 0.262 + ], + "angle": 0, + "content": "[3,6]" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.24, + 0.623, + 0.25 + ], + "angle": 0, + "content": "CLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.251, + 0.624, + 0.262 + ], + "angle": 0, + "content": "[2,3]" + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.24, + 0.657, + 0.25 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.251, + 0.681, + 0.262 + ], + "angle": 0, + "content": "[2,9]" + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.24, + 0.712, + 0.25 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.251, + 0.739, + 0.262 + ], + "angle": 0, + "content": "[2,6]" + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.24, + 0.773, + 0.25 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.251, + 0.797, + 0.262 + ], + "angle": 0, + "content": "[1,12]" + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.24, + 0.839, + 0.25 + ], + "angle": 0, + "content": "BE" + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.251, + 0.852, + 0.262 + ], + "angle": 0, + "content": "[3,9]" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.268, + 0.785, + 0.279 + ], + "angle": 0, + "content": "Axial Local Attention Heads" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.279, + 0.581, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.32, + 0.562, + 0.33 + ], + "angle": 0, + "content": "FS" + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.331, + 0.567, + 0.342 + ], + "angle": 0, + "content": "[5,7]" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.32, + 0.623, + 0.33 + ], + "angle": 0, + "content": "CLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.331, + 0.624, + 0.342 + ], + "angle": 0, + "content": "[5,6]" + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.32, + 0.657, + 0.33 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.331, + 0.681, + 0.342 + ], + "angle": 0, + "content": "[6,3]" + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.32, + 0.712, + 0.33 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.331, + 0.739, + 0.342 + ], + "angle": 0, + "content": "[5,5]" + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.32, + 0.773, + 0.33 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.331, + 0.796, + 0.342 + ], + "angle": 0, + "content": "[6,9]" + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.32, + 0.839, + 0.33 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.331, + 0.855, + 0.342 + ], + "angle": 0, + "content": "[3,12]" + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.347, + 0.788, + 0.359 + ], + "angle": 0, + "content": "Offset Local Attention Heads" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.359, + 0.581, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.4, + 0.562, + 0.41 + ], + "angle": 0, + "content": "FS" + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.41, + 0.567, + 0.421 + ], + "angle": 0, + "content": "[3,7]" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.359, + 0.638, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.4, + 0.624, + 0.41 + ], + "angle": 0, + "content": "CLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.41, + 0.624, + 0.421 + ], + "angle": 0, + "content": "[4,1]" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.359, + 0.694, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.4, + 0.683, + 0.41 + ], + "angle": 0, + "content": "DINO" + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.41, + 0.682, + 0.421 + ], + "angle": 0, + "content": "[4,1]" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.359, + 0.724, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.4, + 0.743, + 0.41 + ], + "angle": 0, + "content": "MoCo" + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.41, + 0.739, + 0.421 + ], + "angle": 0, + "content": "[3,2]" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.359, + 0.78, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.4, + 0.796, + 0.41 + ], + "angle": 0, + "content": "MAE" + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.41, + 0.796, + 0.421 + ], + "angle": 0, + "content": "[5,5]" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.359, + 0.866, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.4, + 0.852, + 0.409 + ], + "angle": 0, + "content": "BEIT" + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.41, + 0.852, + 0.421 + ], + "angle": 0, + "content": "[2,7]" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.43, + 0.892, + 0.498 + ], + "angle": 0, + "content": "Figure 3. Multiple distinct forms of local attention exist. We visualize spatial token attention using Aligned Aggregated Attention Maps, and highlight different types of local attention heads, including Strict, Soft, Axial, and Offset Local Attention Heads. In row 4 we draw the mid-lines in red as a visual aid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "One of the clearest differences can be seen by comparing the mid-to-final layers. For the contrastive self-supervised methods, DINO and MoCo, the attention maps tend to be centered blobs. These heads tend to focus on salient foreground objects, so these blobs simply reflect object-centered photography bias. For the reconstruction-based methods, MAE and BEiT, we see a more diverse group of attention maps. This is likely because these methods must reconstruct all image regions, and thus their attention in the final layers must be more diverse and cover more of the image. Finally, for the explicitly supervised methods, FS and CLIP, the mid-to-final layers do not focus on salient object regions and instead focus on Sparse Repeating Patterns with seemingly no spatial meaning. This occurs for both the CLS tokens and spatial tokens, and the patterns are repeated across both heads and layers. We hypothesize that these patterns occur because the mid-to-late layers are no longer focused on parsing the scene structure, and instead are using their processing power to generate their final decisions for their respective tasks. This phenomenon helps to explain why the attention maps of the later layers of fully-supervised ViTs are poorly suited for segmentation tasks, as was observed by [9]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7489" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.447, + 0.107 + ], + "angle": 0, + "content": "4.2. Emergence of Offset Local Attention Heads" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.355 + ], + "angle": 0, + "content": "It has been shown that ViTs use a mixture of short-range \"local\" and long-range \"global\" attention heads in any given layer [9, 18]. To gain a better understanding of local attention, we propose a visualization strategy of Aligned Aggregated Attention Maps (AAAMs). We extract all spatial token attention maps for 5000 ImageNet images, but before averaging them, we first realigned them so the current spatial token is always in the center of the array. Additional samples of AAAMs are provided in Appendix C.1. Studying these aligned views reveals multiple forms of local attention, shown in Figure 3. We find Strict Local Attention Heads, which attend almost completely to their own position, as well as Soft Local Attention Heads, which attend to a wider neighborhood around them. We also find Axial Local Attention Heads, which are elongated to attend to the local neighborhood along one or both spatial axes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.471, + 0.673 + ], + "angle": 0, + "content": "But perhaps the most noteworthy type of head we observe is the Offset Local Attention Head. These are heads that attend locally, but to a point or region offset from the current token in a vertical or horizontal direction. We find instances of Offset Local Attention Heads in all the models examined, suggesting they are fundamentally necessary for ViTs. To the best of our knowledge, ours is the first work to draw attention to this phenomenon. We believe that such heads are absolutely necessary because ViTs, unlike CNNs, do not have an easy built-in way to test if two features occur next to each other with a particular spatial arrangement. In a CNN, this type of check is naturally embedded into the convolution operator. But in a ViT, there is no inductive bias to induce such a check. For comparison, Soft Local Attention Heads are able to identify if a certain feature is near another feature, but they cannot identify their specific directional arrangement due to their symmetrical attention pattern. The existence of Offset Local Attention Heads implies one possible path for improvement for the ViT architecture, possibly by adding a self-attention variant that introduces some implicit directional structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.325, + 0.697 + ], + "angle": 0, + "content": "4.3. Average Attention Distance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.856 + ], + "angle": 0, + "content": "We measure the Average Attention Distance [18, 52] of each head to assess if particular heads have a short-range \"local\" focus or a long-range \"global\" focus. This metric is computed by measuring the distance from each spatial token to all other tokens and taking a weighted average using the attention map. We normalize the distances so the token grid is embedded on a \\(1 \\times 1\\) square. [52] observed that for a well-trained fully-supervised ViT, the early layers have a mixture of local and global attention heads, while the later layers have only global attention heads." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Figure 4 (left) shows the Average Attention Distances of all heads organized by layer and model. Like [52], we see that most layers use a mixture of local and global attention" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.696, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.545, + 0.234, + 0.847, + 0.246 + ], + "angle": 0, + "content": "FS CLIP DINO MoCo MAE BEiT" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.091, + 0.892, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.7, + 0.234, + 0.848, + 0.246 + ], + "angle": 0, + "content": "MoCo MAE BEiT" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.256, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 4. Different methods of supervision lead to different orderings and ratios of local and global processing. We show the Average Attention Distance of all ViT attention heads organized by layer (left), and the per-layer averages (right)." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.329, + 0.697, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.545, + 0.471, + 0.847, + 0.484 + ], + "angle": 0, + "content": "FS CLIP DINO MoCo MAE BEiT" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.329, + 0.892, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.691, + 0.471, + 0.847, + 0.483 + ], + "angle": 0, + "content": "MoCo MAE BEiT" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.549 + ], + "angle": 0, + "content": "Figure 5. Attention IoU with salient content plateaus early for all ViTs evaluated. We calculate the alignment of ground-truth segmentation masks with CLS token attention maps (left) and the average of spatial token attention maps (right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.817 + ], + "angle": 0, + "content": "heads, however, we also find that the ordering of local and global processing varies greatly with the supervision type. FS, CLIP, DINO, and MoCo all use exclusively global attention heads in the last layers, but the reconstruction-based methods MAE and BEiT use a diverse range of heads in their later layers. Figure 4 (right) compares the combined Average Attention Distances at the per-layer level. In all models, we observe a greater number of global attention heads in the initial layers, followed by decreased distances around layers 3-6. This result is again in contrast to [52]. The behaviors diverge in the mid-to-late layers. For the models trained with explicit or contrastive supervision, the Attention Distance trends upward in the later layers. For the reconstruction-based methods, the Average Attention Distances stay lower. These results show that, unlike CNNs, ViTs can learn a variable local/global processing order depending on the training method used." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.857, + 0.847 + ], + "angle": 0, + "content": "4.4. Attention Alignment with Salient Content" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "One of the most desirable (and exploitable) features of DINO is that the CLS token attention maps of the last layer tend to be well-aligned with salient foreground objects [9]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7490" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.273, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.09, + 0.46, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.228, + 0.47, + 0.257 + ], + "angle": 0, + "content": "Figure 6. CKA similarity between final layer features of different ViTs for their CLS tokens (left) and spatial tokens (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.27, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Several methods propose to use DINO attention maps, feature maps, or a combination of the two to generate segmentations in a self-supervised manner [24, 57, 68]. We conduct a quantitative analysis of this property at all layers of the ViTs, both to measure the usefulness of masks and to assess how early the ViTs differentiate salient object regions. Like [9], we threshold the CLS token attention masks keeping \\(60\\%\\) of the total attention mass. We then compute the Intersection over Union (IoU) of said masks with ground-truth segmentations from PartImageNet [25]. As an alternative to CLS token attention, we also extract masks using the average of spatial token attention maps. We present results for the single \"best\" head per layer in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.679 + ], + "angle": 0, + "content": "We see a clear drop in FS and CLIP mask IoU around the middle of the network, which directly corresponds to the emergence of the Sparse Repeating Patterns observed in Section 4.1. We also find that the IoUs plateau around layers 3-6 for all networks. This demonstrates that ViT models already have a solid understanding of foreground/background separation by the middle layers. While the later attention maps of FS and CLIP are much worse than their self-supervised counterparts, their early-to-mid layers are more comparable. We find that MoCo, MAE, and BEiT can all produce attention maps with IoUs that are comparable with DINO. In addition, we see that the average of spatial tokens produces maps that are comparable with the CLS token, and for CLIP the IoU increases greatly in the final layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.242, + 0.709 + ], + "angle": 0, + "content": "5. Feature Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.468, + 0.807 + ], + "angle": 0, + "content": "In this section, we directly compare ViT features across models and layers using Centered Kernel Alignment (CKA) [14, 31]. We also study unsupervised clustering performance to compare global and local semantic information in the learned representations. We present addition analysis in Appendix D.4 focused on ViT residual connections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.367, + 0.833 + ], + "angle": 0, + "content": "5.1. Last Block Feature Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Comparing representations is non-trivial due to varying feature sizes, large feature representations, and lack of alignment between them. To overcome this, we use batched Centered Kernel Alignment (CKA) [14, 31, 42] which can" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "align features and compute a similarity score. We compare the last layer outputs for each model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Figure 6 (left) shows that the CLS token representations are usually similar for similar supervision strategies (explicit, contrastive, reconstruction). The contrastive methods, MoCo and DINO, show very high similarity to each other, indicating that the CLS token encodes the same type of information for both these methods. There is also an increased level of similarity between the explicitly supervised methods, FS and CLIP, and the contrastive methods. Interestingly, we see that MAE has as high a similarity with DINO and MoCo as it does with BEiT. This result is surprising because MAE's CLS token has no explicit training objective or loss, and the way these approaches are trained is very different. This presents evidence that training autoencoders with a high masking percentage indeed forces the model to learn image-level semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.486 + ], + "angle": 0, + "content": "In Figure 6 (right), we look at the similarity of the last-layer spatial token representations. Unlike the CLS token representations, CLIP and FS have low similarity in their spatial representations. The self-supervised methods DINO, MAE, and MoCo show a high level of similarity to each other, and a lower level of similarity to BEiT. MoCo and DINO show the highest similarity due to their similar kind of self-supervision. Once again, MAE has a high similarity to MoCo and DINO despite their very different supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.846, + 0.515 + ], + "angle": 0, + "content": "5.2. Feature Clustering for Global Semantics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Through this analysis, we aim to test how well the learned CLS and spatial token representations encode global (image-level) semantic information at every layer. We extract the CLS token features from the end of each block for 5000 ImageNet images, and we generate k-Means cluster assignments with \\( k = 50 \\). We present results for cluster purity measured with respect to ground truth image labels, but additional clustering metrics are also presented in Appendix D.5. For the spatial tokens, we follow the same process except we average-pool over all positions before clustering. We also compute a random chance score by replacing the ViT features with Gaussian random noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For CLS token features, shown in Figure 7 (left), cluster purity improves with depth with the exception of the last layers of BEiT. This is likely because the last layers of BEiT serve as a task-specific decoder, unlike MAE, where the decoder is separate and discarded after pretraining. Unsurprisingly, FS achieves the best cluster purity, followed by CLIP. The contrastive methods, DINO and MoCo, achieve scores close to the explicitly supervised methods. The reconstruction-based methods, MAE and BEiT, have the lowest cluster purity, but they are still above random chance, which again indicates that they do learn to encode some image-level semantic information in their CLS tokens. Also, we find that semantic information emerges earlier for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7491" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.299 + ], + "angle": 0, + "content": "Figure 7. Clustering purity analysis with image-level labels in ImageNet-50 for CLS features (left) and average-pooled spatial token features (right)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.314, + 0.47, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Figure 8. Clustering purity of spatial token features gathered at the object-level in COCO (left) and part-level in PartImageNet (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.61 + ], + "angle": 0, + "content": "DINO and MoCo. For the spatial token features, shown in Figure 7 (right), the cluster purity of FS rises earlier compared with the FS CLS token. This suggests that the FS spatial tokens do more work gathering semantic information in the early layers. For all other ViTs, the spatial feature purity is lower in the final layers, but is comparable in layers 1-7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.416, + 0.637 + ], + "angle": 0, + "content": "5.3. Feature Clustering for Local Semantics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We next measure how well the spatial token features differentiate salient image content at the object or part-level using COCO [36] and PartImageNet [25] respectively. We use a tiling strategy to extract a denser array of features, which we detail in Appendix B.4. Using ground truth segmentation masks, we extract and average the features of the tokens overlapping with the masks. This generates a collection of object-level or part-level features which we cluster just like Section 5.2. The results for object-level features are shown in Figure 8 (left). We see that the supervised methods CLIP and FS have the highest feature purity by far, followed by the contrastive methods DINO and MoCo. The purity is much lower for the reconstruction methods MAE and BEiT. For part-level features, shown in Figure 8 (right), FS achieves the best purity, but the contrastive methods are very competitive in this case, surpassing CLIP completely. In addition, while still being the lowest scoring, MAE and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "BEiT are much more competitive at the part-level. Like the image-level purity, the object and part-level feature purity tend to improve with depth, but the purity peaks early around layers 9 to 11. The peak for BEiT is even earlier, likely due to its integrated decoder." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.18, + 0.753, + 0.197 + ], + "angle": 0, + "content": "6. Downstream Task Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Finally, we analyze the performance of these models on downstream tasks that can be performed directly without any fine-tuning or training. We follow the evaluation protocols of [9, 30] for \\( k \\)-NN classification, image retrieval, and video object segmentation. We also perform keypoint correspondence as a more local-focused task. Again we compute random chance scores by replacing all ViT features with Gaussian noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.334, + 0.637, + 0.348 + ], + "angle": 0, + "content": "6.1. Global Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.552 + ], + "angle": 0, + "content": "ImageNet Classification. We perform \\( k \\)-Nearest Neighbor (\\( k \\)-NN) image classification on ImageNet [16] with \\( k = 20 \\). We use the CLS token features from each network and assign the label for a test sample based on the training set features and labels. As can be seen from Figure 9 (left), FS performs the best as it has been trained to classify the same dataset. DINO and MoCo follow a similar trend as Section 5.2 and better encode semantic information in the earlier layers. FS and CLIP also follow a similar trend where their performance shoots up in the last few layers. It is also interesting to see how MAE and BEiT, for which the CLS tokens have no explicit objective, do better than chance, although MAE is considerably better than BEiT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Image Retrieval. Similar to \\( k \\)-NN classification, we utilize the CLS token representation for retrieval. We evaluate on ROxford5k [50] for the Medium (M) split and report the Mean Average Precision (mAP). In Appendix E.2 we also report results for the Hard (H) split and the RParis6k [50] dataset, which follow similar trends. The results, shown in Figure 9 (right), align closely with those for \\( k \\)-NN Classification on ImageNet. FS performs the best followed by CLIP and then DINO and MoCo, and finally by MAE and BEiT with the lowest performance. We hypothesize that the local/global crops used in DINO training help it perform competitively in these global tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.628, + 0.757 + ], + "angle": 0, + "content": "6.2. Local Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "DAVIS Segmentation Propagation. DAVIS Segmentation Propagation is a dense prediction-based video localized task where frame-by-frame features are used to propagate the first frame segmentation mask to subsequent frames. Like Section 5.3, we use a tiling-based dense feature extraction strategy. Results are shown in Figure 10 (left). The contrastive techniques of DINO and MoCo perform the best while FS and CLIP, which are more image-level approaches, face a drop in performance towards the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7492" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.277, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.091, + 0.468, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.083, + 0.231, + 0.463, + 0.244 + ], + "angle": 0, + "content": "FS CLIP DINO MoCo MAE BEiT -Random" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.299 + ], + "angle": 0, + "content": "Figure 9. Global (image-level) downstream task analysis using the CLS token. We present \\( k \\)-NN classifier Top-1 Accuracy on ImageNet (left) and image retrieval mAP on ROxford5k (right)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.304, + 0.275, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.304, + 0.468, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.083, + 0.446, + 0.463, + 0.458 + ], + "angle": 0, + "content": "FS CLIP DINO MoCo MAE BEiT -Random" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.514 + ], + "angle": 0, + "content": "Figure 10. Local (pixel-level) downstream task analysis using the dense spatial token features. We perform DAVIS video segmentation (left) and SPair-71k keypoint correspondence (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.469, + 0.612 + ], + "angle": 0, + "content": "later layers. The local reconstruction-based methods, MAE and BEiT, are also much more competitive in this task. These results show that for purposes of constructing highly descriptive local features, contrastive methods like DINO and MoCo and reconstructive methods like BEiT can surpass features trained with explicit image-level supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.792 + ], + "angle": 0, + "content": "Keypoint Correspondence. We choose keypoint correspondence as an additional local-focused downstream task. Given an image with annotated keypoints, the model must predict the position of corresponding keypoints in a paired image with similar content. We use the SPair-71k [39] dataset and follow the evaluation protocol of [1] and report the Percentage of Correct Keypoints (PCK) [72]. The results are summarized in Figure 10 (right). CLIP excels at this task, closely followed by both FS and DINO. Meanwhile, MoCo, BEiT, and MAE are all very competitive also. The position of the best layer varies significantly, from 8 for CLIP and BEiT to 11 for MAE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.358, + 0.818 + ], + "angle": 0, + "content": "6.3. Summary of Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We summarize the best results for all downstream tasks in Table 1. We denote the best-performing layers in parenthesis. These results show that ViTs with different supervision methods [1] peak at different layers, and [2] perform best at different tasks. In image-level tasks like \\( k \\)-NN clas" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Best performance for each ViT on each downstream task with the corresponding best layer in parenthesis." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.129, + 0.892, + 0.269 + ], + "angle": 0, + "content": "
ModelTask Performance (Best Performing Layer)
Dataset MetricImageNet Top-1↑ROxford5k (M) mAP↑Davis J and F Mean↑SPair-71k PCK@0.1↑
FS83.79 (12)0.45 (12)0.59 (8)28.56 (9)
CLIP75.75 (12)0.40 (12)0.60 (9)30.70 (8)
DINO76.06 (12)0.37 (12)0.60 (12)28.28 (9)
MoCo71.59 (12)0.31 (12)0.61 (11)25.85 (9)
MAE45.19 (12)0.15 (10)0.54 (12)22.74 (11)
BEiT26.84 (8)0.14 (8)0.58 (9)24.11 (8)
Random0.100.020.061.32
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.37 + ], + "angle": 0, + "content": "sification and retrieval, usually, the last layer works the best. For localized tasks like keypoint correspondence and video object segmentation, most models' peak performance happens a few layers before the last one. This shows that always picking the last layer output is not optimal." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.397, + 0.619, + 0.412 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.789 + ], + "angle": 0, + "content": "In this work, we have performed an in-depth comparison of ViTs trained through different methodologies by examining their attention patterns, learned representations, and downstream task performance. We review some of the key findings of our analyses. First, different methods of supervision lead to ViTs that process local and global information in different orders. All ViTs have heads that align well with salient image content, but for the explicitly supervised models, the late-layer attention maps change into Sparse Repeating Patterns. In addition, all ViTs examined have learned to use Offset Local Attention Heads in multiple layers. While explicitly supervised ViTs have the most semantically rich representations at the image level, contrastive methods are competitive, and reconstruction-based methods can also learn meaningful CLS token representations even though said token is a placeholder and has no explicit supervision. Finally, there is no single best model for all the downstream tasks, and the best layer to extract representations from also varies greatly by task and model, so one should not simply take the last layer representation. ViTs have shown a great deal of potential, and we expect they will become more widely used in the coming years. We hope these insights can help with the future development of losses and architectures for Vision Transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and DARPA GARD (HR00112020007) programs. We would like to thank our colleagues Vatsal Agarwal, Matthew Gwilliam, Alex Hanson, Pulkit Kumar, Saketh Rambhatla, and Gaurav Shrivastava for their feedback on this work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7493" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 3, 8, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.224 + ], + "angle": 0, + "content": "[2] Mehmet Aygün and Oisin Mac Aodha. Demystifying unsupervised semantic correspondence estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 125-142. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.318 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.472, + 0.387 + ], + "angle": 0, + "content": "[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[6] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 2956-2964, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[7] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.577, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.47, + 0.739 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.739, + 0.47, + 0.793 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.793, + 0.471, + 0.861 + ], + "angle": 0, + "content": "[13] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.861, + 0.47, + 0.888 + ], + "angle": 0, + "content": "[14] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "The Journal of Machine Learning Research, 13:795-828, 2012. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.188, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.242, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[19] Amin Ghiasi, Hamid Kazemi, Eitan Borgnia, Steven Reich, Manli Shu, Micah Goldblum, Andrew Gordon Wilson, and Tom Goldstein. What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.432, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[20] Kristen Grauman and Trevor Darrell. The pyramid match kernel: Discriminative classification with sets of image features. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1458-1465. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.5, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[21] Tom George Grigg, Dan Busbridge, Jason Ramapuram, and Russ Webb. Do self-supervised and supervised methods learn similar visual representations? arXiv preprint arXiv:2110.00528, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.894, + 0.703 + ], + "angle": 0, + "content": "[23] Matthew Gwilliam and Abhinav Shrivastava. Beyond supervised vs. unsupervised: Representative benchmarking and analysis of image representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9642-9652, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[25] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021. 4, 6, 7, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.812, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.88, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7494" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.201 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[29] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.471, + 0.378 + ], + "angle": 0, + "content": "[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[32] Klemen Kotar, Gabriel Ilharco, Ludwig Schmidt, Kiana Ehsani, and Roozbeh Mottaghi. Contrasting contrastive self-supervised representation learning pipelines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9949-9959, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[33] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 2169-2178. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.514, + 0.47, + 0.568 + ], + "angle": 0, + "content": "[34] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 7, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.799, + 0.469, + 0.838 + ], + "angle": 0, + "content": "[39] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 4, 8, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.838, + 0.469, + 0.893 + ], + "angle": 0, + "content": "[40] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[41] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[42] Thao Nguyen, Maithra Raghu, and Simon Kornblith. Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth. arXiv preprint arXiv:2010.15327, 2020. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.228, + 0.892, + 0.254 + ], + "angle": 0, + "content": "[44] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.254, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[45] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2071-2081, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[46] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[47] Pedro O Pinheiro, Tsung-Yi Lin, Ronan Collobert, and Piotr Dóllár. Learning to refine object segments. In European conference on computer vision, pages 75-91. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[49] Senthil Purushwalkam and Abhinav Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. Advances in Neural Information Processing Systems, 33:3407-3418, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[50] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5706-5715, 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.593, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[52] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 1, 3, 5, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[53] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. arXiv preprint arXiv:2205.08078, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.797, + 0.892, + 0.85 + ], + "angle": 0, + "content": "[54] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.851, + 0.892, + 0.892 + ], + "angle": 0, + "content": "[55] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7495" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[56] Abhinav Shrivastava, Rahul Sukthankar, Jitendra Malik, and Abhinav Gupta. Beyond skip connections: Top-down modulation for object detection. arXiv preprint arXiv:1612.06851, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.147, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[57] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.215, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[58] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.27, + 0.467, + 0.296 + ], + "angle": 0, + "content": "[59] Anand Subramanian. Torch cka. https://github.com/AntixK/PyTorch-Model-Compare, 2021.18" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.297, + 0.469, + 0.336 + ], + "angle": 0, + "content": "[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.337, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[61] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[62] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34:16238-16250, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.458, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[63] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.526, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[64] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking representation learning for natural world image collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12884-12893, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.649, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[66] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.702, + 0.469, + 0.756 + ], + "angle": 0, + "content": "[67] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.757, + 0.469, + 0.837 + ], + "angle": 0, + "content": "[68] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.838, + 0.468, + 0.878 + ], + "angle": 0, + "content": "[69] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.879, + 0.468, + 0.891 + ], + "angle": 0, + "content": "[70] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[71] Kun Xu, Yong Li, Tao Ju, Shi-Min Hu, and Tian-Qiang Liu. Efficient affinity-based edit propagation using kd tree. ACM Transactions on Graphics (TOG), 28(5):1-6, 2009. 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.892, + 0.24 + ], + "angle": 0, + "content": "[72] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 8, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[74] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.337, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[75] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In International Conference on Machine Learning, pages 27378-27394. PMLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7496" + } + ] +] \ No newline at end of file diff --git a/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_origin.pdf b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bbca7d718903de55daec29476c014ea9b76f6a3 --- /dev/null +++ b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/25874b86-4464-466c-bbd4-91f62653cf36_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a605c004eae7f98f7c58713ebee9c05b95919ec970a375070a10292354c7bb5 +size 899895 diff --git a/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/full.md b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6433490705688072c97597b7825882a60aad6b52 --- /dev/null +++ b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/full.md @@ -0,0 +1,378 @@ +# Teaching Matters: Investigating the Role of Supervision in Vision Transformers + +Matthew Walmer* Saksham Suri* Kamal Gupta Abhinav Shrivastava University of Maryland, College Park + +# Abstract + +Vision Transformers (ViTs) have gained significant popularity in recent years and have proliferated into many applications. However, their behavior under different learning paradigms is not well explored. We compare ViTs trained through different methods of supervision, and show that they learn a diverse range of behaviors in terms of their attention, representations, and downstream performance. We also discover ViT behaviors that are consistent across supervision, including the emergence of Offset Local Attention Heads. These are self-attention heads that attend to a token adjacent to the current token with a fixed directional offset, a phenomenon that to the best of our knowledge has not been highlighted in any prior work. Our analysis shows that ViTs are highly flexible and learn to process local and global information in different orders depending on their training method. We find that contrastive self-supervised methods learn features that are competitive with explicitly supervised features, and they can even be superior for part-level tasks. We also find that the representations of reconstruction-based models show non-trivial similarity to contrastive self-supervised models. + +# 1. Introduction + +The field of Computer Vision has advanced massively in the past decade, largely built on the backbone of Convolutional Neural Networks (CNNs). More recently, Vision Transformers (ViTs) [18] have shown the potential to overtake CNNs as the go-to visual processing model. Prior works have asked the question do ViTs see like CNNs do? [52], but in this work, we ask: how do ViTs learn under different supervision? Past examinations of ViTs have largely focused on models trained through full supervision. Instead, we aim to characterize the differences and similarities of ViTs trained through varying training methods, including self-supervised methods. Unlike CNNs, the ViT architecture imposes few structural biases to guide the learning of representations. This gives them the flexibility to + +![](images/ab9b2c8aacf56c3858e9ca9f701ed5374df875564c974002b277ec2cadf39e2e.jpg) +Figure 1. ViTs exhibit highly varied behaviors depending on their method of training. In this work, we compare ViTs through three domains of analysis representing the How, What, and Why of ViTs. How do ViTs process information through attention? (Top) Attention maps averaged over 5000 images show clear differences in the mid-to-late layers. What do ViTs learn to represent? (Left) Contrastive self-supervised ViTs have a greater feature similarity to explicitly supervised ViTs, but also have some similarity with ViTs trained through masked reconstruction. Why do we care about using ViTs? (Right) We evaluate ViTs on a variety of global and local tasks and show that the best model and layer vary greatly. + +learn diverse information processing strategies, and through our analyses, we uncover a wide array of ViT behaviors. + +There are countless ways to analyze ViTs, so to guide this analysis we choose three major domains which correspond to the How, What, and Why of ViTs. For the How, we focus on how ViTs process information through Attention. Multi-Headed Attention (MHA) layers are arguably the key element of ViTs, and they most distinguish them + +from CNNs. For the What, we examine the Features of ViTs, as these are typically what practitioners take away from them. Finally for the Why, we focus on Downstream Tasks, which are why we care about using ViTs. + +Our work unveils that a powerful aspect of the ViT architecture is its local-global dual nature, which plays a role in all three aspects of our analyses. While standard CNNs are restricted to building representations hierarchically from local to global, in a ViT each token can attend to information from any other image region at any time. And unlike popular CNN modifications like Spatial Pyramids [20, 28, 33, 35] and top-down strategies [6, 47, 56], ViTs have the freedom to decide when and where global information should be integrated. In this study, we show that the order and the relative ratio of local and global attention in ViTs varies dramatically based on the method of supervision. We also find clearly different trends in the allocation of attention in the mid-to-late layers of these networks, as highlighted in Figure 2. This local-global dual nature is also embedded into the structure and features of the ViT, which encodes both local spatial tokens and a non-local classifier (CLS) token throughout its entire depth. We analyze the features of ViTs for both the CLS and spatial tokens, and assess how they align with semantics at the image, object, part, and pixel-level. We perform this analysis at every layer of the ViT to show the emergence of different levels of semantic information. Finally, we assess ViTs on a number of local and global downstream tasks. + +Overall, our contributions are: [1] A detailed comparison of ViTs trained with six different methods, including both fully supervised and self-supervised training. [2] A cross-cutting analysis spanning three major domains: Attention, Features, and Downstream Tasks. [3] Multiple insights into the inner workings of ViTs to guide future development of ViT variants, training strategies, and applications. + +In addition, we summarize some of our key observations about ViT behavior: [1] The attention maps of explicitly supervised ViTs devolve into Sparse Repeating Patterns in the mid-to-late layers, but the quality of features continues to improve in these layers (Section 4.1). [2] All ViTs studied learn to use Offset Local Attention Heads, suggesting they are fundamentally necessary in ViTs (Section 4.2). To the best of our knowledge, no prior work has brought attention to this phenomenon. [3] ViTs learn to process local and global information in different orders depending on their method of supervision (Section 4.3). [4] All ViTs studied differentiate salient foreground objects by the early-to-mid layers (Section 4.4). [5] Reconstruction-based self-supervised methods can learn semantically meaningful CLS representations, even when the CLS token is only a placeholder (Section 5.1, 5.2). [6] Supervised method's features are the most semantically rich, but contrastive self-supervised methods are comparable or even superior in + +![](images/d0962b2816c9e68af5a40832021161591bc7b9c2c8d41a895f94f74bd226f5e7.jpg) +Average CLS Token Attention over 5000 Images + +![](images/413a1b747a127219530b61b87f06f3a862f5bb7f2f70ca9a26d4f3cec6666878.jpg) +FS & CLIP Layers 7-12: Sparse Repeating Patterns + +![](images/a03d89c8f641acb2a133b50c97e0b77506e5d15444b4fa5d766667252d6ca9a7.jpg) +DINO & MoCo Layers 7-12: Object Centered Blobs + +![](images/ac5ec47af50a17553e70f1b8dd2adb824389acb623802580934d47675354d6b5.jpg) +MAE & BEiT Layers 7-12: Diverse Attention Maps +Figure 2. Clear differences in attention emerge in the mid-to-late layers under different supervision methods. These plots show the attention maps of CLS tokens averaged over 5000 ImageNet images. Rows indicate layers and columns indicate heads. For brevity, we show only three heads per layer. The bracketed numbers in the lower half denote the layer and head. + +some cases (Section 5.2, 5.3). [7] For localized tasks, the best performance often comes from a mid-to-late layer (Section 6.2). [8] There is no single "best" training method or layer for all downstream tasks (Section 6.3). + +# 2. Related Work + +Previous works have attempted to understand the representation quality for both supervised and self-supervised training for Convolutional Neural Networks (CNNs). [5] focuses on understanding the concepts learned by individual neurons while [41] looks at explaining their compositionality in the case of supervised networks. Simultaneously, due to the popularity of self-supervised learning methods [3, 4, 9-11, 26, 27, 29, 40, 43, 60, 73] multiple works have analyzed these representations learned from no labels. Under this umbrella, [13, 62] tried to understand the effect of training data in terms of both the number and type + +of samples. Some works [66, 67] analyze the alignment, separability, and uniformity of features while [49] looks at invariance to augmentations like occlusion, illumination, and viewpoint change in the learned representation. [64] looks at the downstream performance of self-supervised networks on fine-grained tasks. Finally, [21, 23, 32] analyze multiple self-supervised methods and compare their performance based on representation similarity and downstream task performance over multiple datasets along with comparisons to supervised methods. + +Since the proliferation of ViTs, a number of works have tried to understand and explore the different properties of the representations learned by these networks. A few works [45, 55, 75] have analyzed the robustness of ViT features against corruptions, perturbations, distribution shifts, and adversarial examples while also analyzing the role of self-attention for robustness. [34] benchmarks different pretrained ViTs as backbones for object detection. [7] provides a theoretical understanding of how MAEs work while [53] analyzes attention using convex duality. [61] gives insights to train and use ViTs more efficiently. [44] gives a deeper understanding of how Multi-Headed Attention layers work while comparing and contrasting to how convolution layers behave in terms of the loss landscapes and low-pass/high-pass filtering. [52] compares fully supervised ViTs and ResNets in terms of the local and global information encoded at different depths, the role of skip connections, and the uniformity of representations. + +All these prior works either examine the impact of supervision on CNNs or compare CNNs and ViTs trained with full supervision. Some recent and concurrent works have compared the properties of differently supervised ViTs, though typically focused on a particular task and only two methods of supervision at a time. [1] compares the properties of fully supervised and DINO ViT features in the context of dense feature descriptors, and [2] further compares these two across several semantic correspondence tasks. [19] compares fully supervised and CLIP ViTs through feature visualizations. To the best of our knowledge, we present what is to date the broadest and the most in-depth comparison of ViTs with varying supervision, including six different methods covering three supervision subcategories. Additionally, we propose new attention-based analysis methods along with evaluations on multiple downstream tasks focused on both local and global information. + +# 3. Experimental Design + +# 3.1. A Primer on Vision Transformers + +Vision Transformers (ViTs) [18] are adapted from Transformers [65] for the Natural Language Processing domain. A ViT consists of an array of tokens, each representing an image patch. In addition, most ViTs include an ex + +tra "classifier" or "CLS" token, which is connected to the task-specific output layers during training. ViTs use MultiHeaded Attention (MHA) layers [65], which use a Query-Key-Value system that allows each token to attend to all other tokens with a variable intensity attention map. This is in stark contrast to the limited receptive fields of convolutions. These layers are "multi-headed" because they repeat this process multiple times in parallel, allowing tokens to apply multiple attention strategies concurrently. A ViT architecture includes multiple blocks, each with one MHA layer followed by a position-wise fully connected layer. Unlike CNNs, which usually get narrower in deeper layers, ViTs maintain the same "width" (number of tokens) throughout. There are some transformer variants, like SWiN transformers [37], that introduce a narrowing width, but for our analyses, we focus on only traditional ViTs. Specifically, our primary analysis focuses on ViT-Base models with patch size $16 \times 16$ (ViT-B/16) and input size $224 \times 224$ , which results in a $14 \times 14$ spatial token array. ViT-Base has 12 blocks and 12 attention heads per MHA layer. In the Appendix, we provide additional results on a wider range of ViTs, including variations in architecture size and patch size. + +# 3.2. Methods of Supervision + +Although a large number of ViT training methods have been proposed in a short span of three years, many of the most popular methods can be loosely categorized into the following three groups. From each group, we select two representative models for in-depth analysis. We further discuss these models' details in Appendix B.2. + +Explicit Supervision. These models are trained with an explicit objective that is defined either by human annotations or by labels derived from another source, like paired image captions. For this category, we use a Fully Supervised (FS) ViT pretrained on ImageNet21k and fine-tuned on ImageNet1k [58, 69], as well as a CLIP ViT [51]. + +Self-Supervision (Contrastive). Self-supervised learning methods broadly attempt to train a model through a pretext task that can be directly derived from the input data. Among the more popular pretext tasks are contrastive learning methods [27, 70] which generally present a model with multiple augmented views of the same image alongside distractor views of other images. The model must learn to identify which of the views came from the same image. For this category, we select DINO [9] and MoCo-v3 [12] which we denote simply as MoCo for the rest of this paper. + +Self-Supervision (Reconstruction). Another popular category of self-supervision is reconstruction methods, which train models to predict the missing content from masked or otherwise corrupted images. We select MAE [26] and BEiT [4] for this category. Note that MAE has a separate decoder which is discarded after pretraining, + +while BEiT's decoder is learned in the same ViT. This has a strong impact on the behavior of the later layers of BEiT. + +# 3.3. Datasets + +We study the ViTs on multiple datasets and downstream tasks. Unless otherwise specified, we use ImageNet-50 [63], a subset of ImageNet [16] which narrows the dataset down to 50 representative categories. We sample 100 images per class to create a diverse collection of 5000 images. We additionally use PartImageNet [25] to measure Attention Saliency and part-level feature purity, as well as COCO [36] to measure object-level feature purity. We use revisited [50] Oxford [46] (ROxford5k) for evaluating image retrieval, DAVIS [48] for video segmentation, and SPair-71k [39] for keypoint correspondence. + +# 3.4. Proposed Analyses + +Our analysis is broadly divided into three domains covering the How, What, and Why of ViTs: + +How ViTs process local/global information (Attention). Do self-attention heads learn to operate in different ways depending on their method of training? Are there distinctive modes of attention behavior? How does supervision impact the processing order of local and global information? + +What we take away from ViTs (Features). How do the final and intermediate representations of a ViT change depending on the method of supervision? Are these trends similar or different for CLS vs. spatial tokens? + +Why we use ViTs (Downstream Tasks). Which forms of supervision are best suited for different downstream tasks? Which layers of a ViT produce features that are best for different local and global tasks? + +# 4. Attention Analysis + +Multi-Headed Attention layers are one of the defining components of the Transformer architecture, and the attention maps they generate can give key insights into what is similar or different about ViTs trained through different methods. We perform an in-depth examination of the self-attention maps of ViT-B/16 models at every layer. Through this study, we uncover a diverse range of attention head behavioral modalities. Additional visualizations are provided in Appendix C for a wide range of ViT variants. + +# 4.1. Attention Visualizations + +We start by examining the attention maps of the CLS tokens of each head and layer. To gain a comprehensive understanding of each head's behavior, we compute the average attention maps over 5000 ImageNet images, as shown in Figure 2. For brevity, we display only three heads per layer, but complete plots can be found in Appendix C.1, along with additional visualizations for spatial token attention and individual input images. + +Aligned Aggregated Attention Maps for Spatial Tokens + +Strict Local Attention Heads + +![](images/2327a29ce828d09acaa5337f1667f8186e3f62e1ace66c919e1716603c7f88a1.jpg) +FS +CLIP +[1,1] +1 +[1,6] +1 +[2,4] +1 +[2,9] +BE + +![](images/638697a81c7f1ec96804679e5d95892a014916dc08ae679fcf62a33ca8ba2f6c.jpg) +[1,7] +Soft Local Attention Heads +FS +[3,6] +CLIP +[2,3] +1 +[2,9] +1 + +![](images/9c8d5b16488fa968162a8203e7ac1522c6c5e164aa13a3baec91f99eccd3538c.jpg) +Axial Local Attention Heads +FS +[5,7] +CLIP + +![](images/943e69af392a718b2d3d9f4784ac56a586ce68dee3c1eef5e0322116f2881348.jpg) +FS +[3,7] +Figure 3. Multiple distinct forms of local attention exist. We visualize spatial token attention using Aligned Aggregated Attention Maps, and highlight different types of local attention heads, including Strict, Soft, Axial, and Offset Local Attention Heads. In row 4 we draw the mid-lines in red as a visual aid. + +![](images/acaeed370ca87bcd31f1f12c3acd942fd369e2ba1a50a0023828a5bb72e01d14.jpg) +[5,6] +Offset Local Attention Heads +CLIP +[4,1] + +![](images/6f76749a2f6c951fa1e3ae4683153f17b85ca119cfb3f9f2a216c27de396b6bd.jpg) +1 +[6,3] +DINO +[4,1] + +![](images/83149018e84bcd3ef1fd80743dad9eedc28d1b7f111ec4542669e37aebc9e374.jpg) +[2,6] +1 +[5,5] +MoCo +[3,2] + +![](images/64803b7c3d725314a70336e42453d4b16bd2dd785ad483cd5d697e74382b63ed.jpg) +1 +[1,12] +1 +[6,9] +MAE +[5,5] + +![](images/241349858d62828debf30be886ac74d33cf3bcaed8cf4d5810ebf39e8122dda9.jpg) +[2,3] +BE +[3,9] +B +[3,12] +BEIT +[2,7] + +One of the clearest differences can be seen by comparing the mid-to-final layers. For the contrastive self-supervised methods, DINO and MoCo, the attention maps tend to be centered blobs. These heads tend to focus on salient foreground objects, so these blobs simply reflect object-centered photography bias. For the reconstruction-based methods, MAE and BEiT, we see a more diverse group of attention maps. This is likely because these methods must reconstruct all image regions, and thus their attention in the final layers must be more diverse and cover more of the image. Finally, for the explicitly supervised methods, FS and CLIP, the mid-to-final layers do not focus on salient object regions and instead focus on Sparse Repeating Patterns with seemingly no spatial meaning. This occurs for both the CLS tokens and spatial tokens, and the patterns are repeated across both heads and layers. We hypothesize that these patterns occur because the mid-to-late layers are no longer focused on parsing the scene structure, and instead are using their processing power to generate their final decisions for their respective tasks. This phenomenon helps to explain why the attention maps of the later layers of fully-supervised ViTs are poorly suited for segmentation tasks, as was observed by [9]. + +# 4.2. Emergence of Offset Local Attention Heads + +It has been shown that ViTs use a mixture of short-range "local" and long-range "global" attention heads in any given layer [9, 18]. To gain a better understanding of local attention, we propose a visualization strategy of Aligned Aggregated Attention Maps (AAAMs). We extract all spatial token attention maps for 5000 ImageNet images, but before averaging them, we first realigned them so the current spatial token is always in the center of the array. Additional samples of AAAMs are provided in Appendix C.1. Studying these aligned views reveals multiple forms of local attention, shown in Figure 3. We find Strict Local Attention Heads, which attend almost completely to their own position, as well as Soft Local Attention Heads, which attend to a wider neighborhood around them. We also find Axial Local Attention Heads, which are elongated to attend to the local neighborhood along one or both spatial axes. + +But perhaps the most noteworthy type of head we observe is the Offset Local Attention Head. These are heads that attend locally, but to a point or region offset from the current token in a vertical or horizontal direction. We find instances of Offset Local Attention Heads in all the models examined, suggesting they are fundamentally necessary for ViTs. To the best of our knowledge, ours is the first work to draw attention to this phenomenon. We believe that such heads are absolutely necessary because ViTs, unlike CNNs, do not have an easy built-in way to test if two features occur next to each other with a particular spatial arrangement. In a CNN, this type of check is naturally embedded into the convolution operator. But in a ViT, there is no inductive bias to induce such a check. For comparison, Soft Local Attention Heads are able to identify if a certain feature is near another feature, but they cannot identify their specific directional arrangement due to their symmetrical attention pattern. The existence of Offset Local Attention Heads implies one possible path for improvement for the ViT architecture, possibly by adding a self-attention variant that introduces some implicit directional structure. + +# 4.3. Average Attention Distance + +We measure the Average Attention Distance [18, 52] of each head to assess if particular heads have a short-range "local" focus or a long-range "global" focus. This metric is computed by measuring the distance from each spatial token to all other tokens and taking a weighted average using the attention map. We normalize the distances so the token grid is embedded on a $1 \times 1$ square. [52] observed that for a well-trained fully-supervised ViT, the early layers have a mixture of local and global attention heads, while the later layers have only global attention heads. + +Figure 4 (left) shows the Average Attention Distances of all heads organized by layer and model. Like [52], we see that most layers use a mixture of local and global attention + +Figure 4. Different methods of supervision lead to different orderings and ratios of local and global processing. We show the Average Attention Distance of all ViT attention heads organized by layer (left), and the per-layer averages (right). +![](images/55ccde2c27836ef39ead72527fc1066c5c83943dda02ad91e2253f22fb12a7cb.jpg) +FS CLIP DINO MoCo MAE BEiT + +![](images/3fdb8e3d6d034a8a370d4a2dcfc18c2566db141c46eb741abad6389fe7c6a5b0.jpg) + +MoCo MAE BEiT + +Figure 5. Attention IoU with salient content plateaus early for all ViTs evaluated. We calculate the alignment of ground-truth segmentation masks with CLS token attention maps (left) and the average of spatial token attention maps (right). +![](images/2570959cb9d1ca8b3f3f4a4f398bad321eba47e1e9c92b99a9f564b330ca5fff.jpg) +FS CLIP DINO MoCo MAE BEiT + +![](images/eb6bc27b4ec22aad2b0df63940aa8d613413300756edb761395eac987ad13526.jpg) + +MoCo MAE BEiT + +heads, however, we also find that the ordering of local and global processing varies greatly with the supervision type. FS, CLIP, DINO, and MoCo all use exclusively global attention heads in the last layers, but the reconstruction-based methods MAE and BEiT use a diverse range of heads in their later layers. Figure 4 (right) compares the combined Average Attention Distances at the per-layer level. In all models, we observe a greater number of global attention heads in the initial layers, followed by decreased distances around layers 3-6. This result is again in contrast to [52]. The behaviors diverge in the mid-to-late layers. For the models trained with explicit or contrastive supervision, the Attention Distance trends upward in the later layers. For the reconstruction-based methods, the Average Attention Distances stay lower. These results show that, unlike CNNs, ViTs can learn a variable local/global processing order depending on the training method used. + +# 4.4. Attention Alignment with Salient Content + +One of the most desirable (and exploitable) features of DINO is that the CLS token attention maps of the last layer tend to be well-aligned with salient foreground objects [9]. + +![](images/a5b144029558877b69cd52d2f0d8e67bc7fb5396a933a5873dbcd7f15d3b50c6.jpg) +Figure 6. CKA similarity between final layer features of different ViTs for their CLS tokens (left) and spatial tokens (right). + +![](images/91542c7060b528dee1eb1f232faa662447128b040739e34a8758831a728936c9.jpg) + +Several methods propose to use DINO attention maps, feature maps, or a combination of the two to generate segmentations in a self-supervised manner [24, 57, 68]. We conduct a quantitative analysis of this property at all layers of the ViTs, both to measure the usefulness of masks and to assess how early the ViTs differentiate salient object regions. Like [9], we threshold the CLS token attention masks keeping $60\%$ of the total attention mass. We then compute the Intersection over Union (IoU) of said masks with ground-truth segmentations from PartImageNet [25]. As an alternative to CLS token attention, we also extract masks using the average of spatial token attention maps. We present results for the single "best" head per layer in Figure 5. + +We see a clear drop in FS and CLIP mask IoU around the middle of the network, which directly corresponds to the emergence of the Sparse Repeating Patterns observed in Section 4.1. We also find that the IoUs plateau around layers 3-6 for all networks. This demonstrates that ViT models already have a solid understanding of foreground/background separation by the middle layers. While the later attention maps of FS and CLIP are much worse than their self-supervised counterparts, their early-to-mid layers are more comparable. We find that MoCo, MAE, and BEiT can all produce attention maps with IoUs that are comparable with DINO. In addition, we see that the average of spatial tokens produces maps that are comparable with the CLS token, and for CLIP the IoU increases greatly in the final layer. + +# 5. Feature Analysis + +In this section, we directly compare ViT features across models and layers using Centered Kernel Alignment (CKA) [14, 31]. We also study unsupervised clustering performance to compare global and local semantic information in the learned representations. We present addition analysis in Appendix D.4 focused on ViT residual connections. + +# 5.1. Last Block Feature Comparisons + +Comparing representations is non-trivial due to varying feature sizes, large feature representations, and lack of alignment between them. To overcome this, we use batched Centered Kernel Alignment (CKA) [14, 31, 42] which can + +align features and compute a similarity score. We compare the last layer outputs for each model. + +Figure 6 (left) shows that the CLS token representations are usually similar for similar supervision strategies (explicit, contrastive, reconstruction). The contrastive methods, MoCo and DINO, show very high similarity to each other, indicating that the CLS token encodes the same type of information for both these methods. There is also an increased level of similarity between the explicitly supervised methods, FS and CLIP, and the contrastive methods. Interestingly, we see that MAE has as high a similarity with DINO and MoCo as it does with BEiT. This result is surprising because MAE's CLS token has no explicit training objective or loss, and the way these approaches are trained is very different. This presents evidence that training autoencoders with a high masking percentage indeed forces the model to learn image-level semantics. + +In Figure 6 (right), we look at the similarity of the last-layer spatial token representations. Unlike the CLS token representations, CLIP and FS have low similarity in their spatial representations. The self-supervised methods DINO, MAE, and MoCo show a high level of similarity to each other, and a lower level of similarity to BEiT. MoCo and DINO show the highest similarity due to their similar kind of self-supervision. Once again, MAE has a high similarity to MoCo and DINO despite their very different supervision. + +# 5.2. Feature Clustering for Global Semantics + +Through this analysis, we aim to test how well the learned CLS and spatial token representations encode global (image-level) semantic information at every layer. We extract the CLS token features from the end of each block for 5000 ImageNet images, and we generate k-Means cluster assignments with $k = 50$ . We present results for cluster purity measured with respect to ground truth image labels, but additional clustering metrics are also presented in Appendix D.5. For the spatial tokens, we follow the same process except we average-pool over all positions before clustering. We also compute a random chance score by replacing the ViT features with Gaussian random noise. + +For CLS token features, shown in Figure 7 (left), cluster purity improves with depth with the exception of the last layers of BEiT. This is likely because the last layers of BEiT serve as a task-specific decoder, unlike MAE, where the decoder is separate and discarded after pretraining. Unsurprisingly, FS achieves the best cluster purity, followed by CLIP. The contrastive methods, DINO and MoCo, achieve scores close to the explicitly supervised methods. The reconstruction-based methods, MAE and BEiT, have the lowest cluster purity, but they are still above random chance, which again indicates that they do learn to encode some image-level semantic information in their CLS tokens. Also, we find that semantic information emerges earlier for + +![](images/e73f13a940cb786fa28fa1b6ea032f5593708b1644091a01e6d1861956ca1ef4.jpg) +Figure 7. Clustering purity analysis with image-level labels in ImageNet-50 for CLS features (left) and average-pooled spatial token features (right). + +![](images/be76f898de49b7ef4068642da8488cc6c5316a7c131bd1cf94030f0a15a394de.jpg) +Figure 8. Clustering purity of spatial token features gathered at the object-level in COCO (left) and part-level in PartImageNet (right). + +DINO and MoCo. For the spatial token features, shown in Figure 7 (right), the cluster purity of FS rises earlier compared with the FS CLS token. This suggests that the FS spatial tokens do more work gathering semantic information in the early layers. For all other ViTs, the spatial feature purity is lower in the final layers, but is comparable in layers 1-7. + +# 5.3. Feature Clustering for Local Semantics + +We next measure how well the spatial token features differentiate salient image content at the object or part-level using COCO [36] and PartImageNet [25] respectively. We use a tiling strategy to extract a denser array of features, which we detail in Appendix B.4. Using ground truth segmentation masks, we extract and average the features of the tokens overlapping with the masks. This generates a collection of object-level or part-level features which we cluster just like Section 5.2. The results for object-level features are shown in Figure 8 (left). We see that the supervised methods CLIP and FS have the highest feature purity by far, followed by the contrastive methods DINO and MoCo. The purity is much lower for the reconstruction methods MAE and BEiT. For part-level features, shown in Figure 8 (right), FS achieves the best purity, but the contrastive methods are very competitive in this case, surpassing CLIP completely. In addition, while still being the lowest scoring, MAE and + +BEiT are much more competitive at the part-level. Like the image-level purity, the object and part-level feature purity tend to improve with depth, but the purity peaks early around layers 9 to 11. The peak for BEiT is even earlier, likely due to its integrated decoder. + +# 6. Downstream Task Analysis + +Finally, we analyze the performance of these models on downstream tasks that can be performed directly without any fine-tuning or training. We follow the evaluation protocols of [9, 30] for $k$ -NN classification, image retrieval, and video object segmentation. We also perform keypoint correspondence as a more local-focused task. Again we compute random chance scores by replacing all ViT features with Gaussian noise. + +# 6.1. Global Tasks + +ImageNet Classification. We perform $k$ -Nearest Neighbor ( $k$ -NN) image classification on ImageNet [16] with $k = 20$ . We use the CLS token features from each network and assign the label for a test sample based on the training set features and labels. As can be seen from Figure 9 (left), FS performs the best as it has been trained to classify the same dataset. DINO and MoCo follow a similar trend as Section 5.2 and better encode semantic information in the earlier layers. FS and CLIP also follow a similar trend where their performance shoots up in the last few layers. It is also interesting to see how MAE and BEiT, for which the CLS tokens have no explicit objective, do better than chance, although MAE is considerably better than BEiT. + +Image Retrieval. Similar to $k$ -NN classification, we utilize the CLS token representation for retrieval. We evaluate on ROxford5k [50] for the Medium (M) split and report the Mean Average Precision (mAP). In Appendix E.2 we also report results for the Hard (H) split and the RParis6k [50] dataset, which follow similar trends. The results, shown in Figure 9 (right), align closely with those for $k$ -NN Classification on ImageNet. FS performs the best followed by CLIP and then DINO and MoCo, and finally by MAE and BEiT with the lowest performance. We hypothesize that the local/global crops used in DINO training help it perform competitively in these global tasks. + +# 6.2. Local Tasks + +DAVIS Segmentation Propagation. DAVIS Segmentation Propagation is a dense prediction-based video localized task where frame-by-frame features are used to propagate the first frame segmentation mask to subsequent frames. Like Section 5.3, we use a tiling-based dense feature extraction strategy. Results are shown in Figure 10 (left). The contrastive techniques of DINO and MoCo perform the best while FS and CLIP, which are more image-level approaches, face a drop in performance towards the + +![](images/936de7d6abd4083539243b9877b2cb9bb09286027b82a12a444d544ece520940.jpg) +FS CLIP DINO MoCo MAE BEiT -Random + +![](images/041bc83afe63922ed3a6d5ef14d74430bbf7b41f1c551cdcaca60c6a2676ff2b.jpg) + +Figure 9. Global (image-level) downstream task analysis using the CLS token. We present $k$ -NN classifier Top-1 Accuracy on ImageNet (left) and image retrieval mAP on ROxford5k (right). +![](images/58ee801ec5b6cd8d9c3532baf64dbd31e0342dd1f2364439adb33a36b62194a4.jpg) +FS CLIP DINO MoCo MAE BEiT -Random + +![](images/ffe40f7ed68cbe91e11cf0c357b9c12406812438ead6e24deaa1549715b261cc.jpg) +Figure 10. Local (pixel-level) downstream task analysis using the dense spatial token features. We perform DAVIS video segmentation (left) and SPair-71k keypoint correspondence (right). + +later layers. The local reconstruction-based methods, MAE and BEiT, are also much more competitive in this task. These results show that for purposes of constructing highly descriptive local features, contrastive methods like DINO and MoCo and reconstructive methods like BEiT can surpass features trained with explicit image-level supervision. + +Keypoint Correspondence. We choose keypoint correspondence as an additional local-focused downstream task. Given an image with annotated keypoints, the model must predict the position of corresponding keypoints in a paired image with similar content. We use the SPair-71k [39] dataset and follow the evaluation protocol of [1] and report the Percentage of Correct Keypoints (PCK) [72]. The results are summarized in Figure 10 (right). CLIP excels at this task, closely followed by both FS and DINO. Meanwhile, MoCo, BEiT, and MAE are all very competitive also. The position of the best layer varies significantly, from 8 for CLIP and BEiT to 11 for MAE. + +# 6.3. Summary of Downstream Tasks + +We summarize the best results for all downstream tasks in Table 1. We denote the best-performing layers in parenthesis. These results show that ViTs with different supervision methods [1] peak at different layers, and [2] perform best at different tasks. In image-level tasks like $k$ -NN clas + +Table 1. Best performance for each ViT on each downstream task with the corresponding best layer in parenthesis. + +
ModelTask Performance (Best Performing Layer)
Dataset MetricImageNet Top-1↑ROxford5k (M) mAP↑Davis J and F Mean↑SPair-71k PCK@0.1↑
FS83.79 (12)0.45 (12)0.59 (8)28.56 (9)
CLIP75.75 (12)0.40 (12)0.60 (9)30.70 (8)
DINO76.06 (12)0.37 (12)0.60 (12)28.28 (9)
MoCo71.59 (12)0.31 (12)0.61 (11)25.85 (9)
MAE45.19 (12)0.15 (10)0.54 (12)22.74 (11)
BEiT26.84 (8)0.14 (8)0.58 (9)24.11 (8)
Random0.100.020.061.32
+ +sification and retrieval, usually, the last layer works the best. For localized tasks like keypoint correspondence and video object segmentation, most models' peak performance happens a few layers before the last one. This shows that always picking the last layer output is not optimal. + +# 7. Conclusion + +In this work, we have performed an in-depth comparison of ViTs trained through different methodologies by examining their attention patterns, learned representations, and downstream task performance. We review some of the key findings of our analyses. First, different methods of supervision lead to ViTs that process local and global information in different orders. All ViTs have heads that align well with salient image content, but for the explicitly supervised models, the late-layer attention maps change into Sparse Repeating Patterns. In addition, all ViTs examined have learned to use Offset Local Attention Heads in multiple layers. While explicitly supervised ViTs have the most semantically rich representations at the image level, contrastive methods are competitive, and reconstruction-based methods can also learn meaningful CLS token representations even though said token is a placeholder and has no explicit supervision. Finally, there is no single best model for all the downstream tasks, and the best layer to extract representations from also varies greatly by task and model, so one should not simply take the last layer representation. ViTs have shown a great deal of potential, and we expect they will become more widely used in the coming years. We hope these insights can help with the future development of losses and architectures for Vision Transformers. + +Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and DARPA GARD (HR00112020007) programs. We would like to thank our colleagues Vatsal Agarwal, Matthew Gwilliam, Alex Hanson, Pulkit Kumar, Saketh Rambhatla, and Gaurav Shrivastava for their feedback on this work. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 3, 8, 24 +[2] Mehmet Aygün and Oisin Mac Aodha. Demystifying unsupervised semantic correspondence estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 125-142. Springer, 2022. 3 +[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2 +[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3, 12, 14 +[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2 +[6] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 2956-2964, 2015. 2 +[7] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022. 3 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 14 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 12 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 3, 12 +[13] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 2 +[14] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment. + +The Journal of Machine Learning Research, 13:795-828, 2012. 6, 18 +[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 12 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 7 +[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 14 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 5 +[19] Amin Ghiasi, Hamid Kazemi, Eitan Borgnia, Steven Reich, Manli Shu, Micah Goldblum, Andrew Gordon Wilson, and Tom Goldstein. What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727, 2022. 3 +[20] Kristen Grauman and Trevor Darrell. The pyramid match kernel: Discriminative classification with sets of image features. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1458-1465. IEEE, 2005. 2 +[21] Tom George Grigg, Dan Busbridge, Jason Ramapuram, and Russ Webb. Do self-supervised and supervised methods learn similar visual representations? arXiv preprint arXiv:2110.00528, 2021. 3 +[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 14 +[23] Matthew Gwilliam and Abhinav Shrivastava. Beyond supervised vs. unsupervised: Representative benchmarking and analysis of image representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9642-9652, 2022. 3 +[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 6 +[25] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021. 4, 6, 7, 18 +[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3, 12, 14 +[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross + +Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 12 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2 +[29] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2 +[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7 +[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 6, 18 +[32] Klemen Kotar, Gabriel Ilharco, Ludwig Schmidt, Kiana Ehsani, and Roozbeh Mottaghi. Contrasting contrastive self-supervised representation learning pipelines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9949-9959, 2021. 3 +[33] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 2169-2178. IEEE, 2006. 2 +[34] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3 +[35] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 7, 18 +[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3 +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 12 +[39] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 4, 8, 24 +[40] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020. 2 + +[41] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2 +[42] Thao Nguyen, Maithra Raghu, and Simon Kornblith. Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth. arXiv preprint arXiv:2010.15327, 2020. 6, 18 +[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[44] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 3 +[45] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2071-2081, 2022. 3 +[46] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 4 +[47] Pedro O Pinheiro, Tsung-Yi Lin, Ronan Collobert, and Piotr Dóllár. Learning to refine object segments. In European conference on computer vision, pages 75-91. Springer, 2016. 2 +[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4 +[49] Senthil Purushwalkam and Abhinav Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. Advances in Neural Information Processing Systems, 33:3407-3418, 2020. 3 +[50] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5706-5715, 2018. 4, 7 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12 +[52] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 1, 3, 5, 23 +[53] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. arXiv preprint arXiv:2205.08078, 2022. 3 +[54] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 24 +[55] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 3 + +[56] Abhinav Shrivastava, Rahul Sukthankar, Jitendra Malik, and Abhinav Gupta. Beyond skip connections: Top-down modulation for object detection. arXiv preprint arXiv:1612.06851, 2016. 2 +[57] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 6 +[58] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 3, 12 +[59] Anand Subramanian. Torch cka. https://github.com/AntixK/PyTorch-Model-Compare, 2021.18 +[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2 +[61] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 3 +[62] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34:16238-16250, 2021. 2 +[63] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 4 +[64] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking representation learning for natural world image collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12884-12893, 2021. 3 +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[66] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3 +[67] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3 +[68] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 6 +[69] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.3, 12 +[70] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. + +Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3 +[71] Kun Xu, Yong Li, Tao Ju, Shi-Min Hu, and Tian-Qiang Liu. Efficient affinity-based edit propagation using kd tree. ACM Transactions on Graphics (TOG), 28(5):1-6, 2009. 24 +[72] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 8, 24 +[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2 +[74] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 12 +[75] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In International Conference on Machine Learning, pages 27378-27394. PMLR, 2022. 3 \ No newline at end of file diff --git a/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/images.zip b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b172016b76770941c4c5ae11c58ac43507dcdc75 --- /dev/null +++ b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2ba7769ac05ceb73daa14f28997a7589065131eab49fd38ef2bfa49dbc57a31 +size 526579 diff --git a/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/layout.json b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..34e1daa4ff0743c6bd8e0b17fc78be4d1f8e8432 --- /dev/null +++ b/2023/Teaching Matters_ Investigating the Role of Supervision in Vision Transformers/layout.json @@ -0,0 +1,9858 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 541, + 121 + ], + "type": "text", + "content": "Teaching Matters: Investigating the Role of Supervision in Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 136, + 477, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 136, + 477, + 164 + ], + "spans": [ + { + "bbox": [ + 116, + 136, + 477, + 164 + ], + "type": "text", + "content": "Matthew Walmer* Saksham Suri* Kamal Gupta Abhinav Shrivastava University of Maryland, College Park" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 192, + 212, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 192, + 212, + 204 + ], + "spans": [ + { + "bbox": [ + 164, + 192, + 212, + 204 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 217, + 290, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 290, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 290, + 469 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have gained significant popularity in recent years and have proliferated into many applications. However, their behavior under different learning paradigms is not well explored. We compare ViTs trained through different methods of supervision, and show that they learn a diverse range of behaviors in terms of their attention, representations, and downstream performance. We also discover ViT behaviors that are consistent across supervision, including the emergence of Offset Local Attention Heads. These are self-attention heads that attend to a token adjacent to the current token with a fixed directional offset, a phenomenon that to the best of our knowledge has not been highlighted in any prior work. Our analysis shows that ViTs are highly flexible and learn to process local and global information in different orders depending on their training method. We find that contrastive self-supervised methods learn features that are competitive with explicitly supervised features, and they can even be superior for part-level tasks. We also find that the representations of reconstruction-based models show non-trivial similarity to contrastive self-supervised models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 484, + 128, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 128, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 128, + 497 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 505, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 673 + ], + "type": "text", + "content": "The field of Computer Vision has advanced massively in the past decade, largely built on the backbone of Convolutional Neural Networks (CNNs). More recently, Vision Transformers (ViTs) [18] have shown the potential to overtake CNNs as the go-to visual processing model. Prior works have asked the question do ViTs see like CNNs do? [52], but in this work, we ask: how do ViTs learn under different supervision? Past examinations of ViTs have largely focused on models trained through full supervision. Instead, we aim to characterize the differences and similarities of ViTs trained through varying training methods, including self-supervised methods. Unlike CNNs, the ViT architecture imposes few structural biases to guide the learning of representations. This gives them the flexibility to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 194, + 547, + 479 + ], + "blocks": [ + { + "bbox": [ + 307, + 194, + 547, + 479 + ], + "lines": [ + { + "bbox": [ + 307, + 194, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 547, + 479 + ], + "type": "image", + "image_path": "ab9b2c8aacf56c3858e9ca9f701ed5374df875564c974002b277ec2cadf39e2e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 487, + 547, + 609 + ], + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 609 + ], + "type": "text", + "content": "Figure 1. ViTs exhibit highly varied behaviors depending on their method of training. In this work, we compare ViTs through three domains of analysis representing the How, What, and Why of ViTs. How do ViTs process information through attention? (Top) Attention maps averaged over 5000 images show clear differences in the mid-to-late layers. What do ViTs learn to represent? (Left) Contrastive self-supervised ViTs have a greater feature similarity to explicitly supervised ViTs, but also have some similarity with ViTs trained through masked reconstruction. Why do we care about using ViTs? (Right) We evaluate ViTs on a variety of global and local tasks and show that the best model and layer vary greatly." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 616, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 640 + ], + "type": "text", + "content": "learn diverse information processing strategies, and through our analyses, we uncover a wide array of ViT behaviors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 713 + ], + "type": "text", + "content": "There are countless ways to analyze ViTs, so to guide this analysis we choose three major domains which correspond to the How, What, and Why of ViTs. For the How, we focus on how ViTs process information through Attention. Multi-Headed Attention (MHA) layers are arguably the key element of ViTs, and they most distinguish them" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 274, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 274, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 274, + 713 + ], + "type": "text", + "content": "*Equal contributors. Web: www.cs.umd.edu/~sakshams/vit_analysis Code: www.github.com/mwalmer-umd/vit_analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7486" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 120 + ], + "type": "text", + "content": "from CNNs. For the What, we examine the Features of ViTs, as these are typically what practitioners take away from them. Finally for the Why, we focus on Downstream Tasks, which are why we care about using ViTs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 122, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 289, + 409 + ], + "type": "text", + "content": "Our work unveils that a powerful aspect of the ViT architecture is its local-global dual nature, which plays a role in all three aspects of our analyses. While standard CNNs are restricted to building representations hierarchically from local to global, in a ViT each token can attend to information from any other image region at any time. And unlike popular CNN modifications like Spatial Pyramids [20, 28, 33, 35] and top-down strategies [6, 47, 56], ViTs have the freedom to decide when and where global information should be integrated. In this study, we show that the order and the relative ratio of local and global attention in ViTs varies dramatically based on the method of supervision. We also find clearly different trends in the allocation of attention in the mid-to-late layers of these networks, as highlighted in Figure 2. This local-global dual nature is also embedded into the structure and features of the ViT, which encodes both local spatial tokens and a non-local classifier (CLS) token throughout its entire depth. We analyze the features of ViTs for both the CLS and spatial tokens, and assess how they align with semantics at the image, object, part, and pixel-level. We perform this analysis at every layer of the ViT to show the emergence of different levels of semantic information. Finally, we assess ViTs on a number of local and global downstream tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 411, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 496 + ], + "type": "text", + "content": "Overall, our contributions are: [1] A detailed comparison of ViTs trained with six different methods, including both fully supervised and self-supervised training. [2] A cross-cutting analysis spanning three major domains: Attention, Features, and Downstream Tasks. [3] Multiple insights into the inner workings of ViTs to guide future development of ViT variants, training strategies, and applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 288, + 714 + ], + "type": "text", + "content": "In addition, we summarize some of our key observations about ViT behavior: [1] The attention maps of explicitly supervised ViTs devolve into Sparse Repeating Patterns in the mid-to-late layers, but the quality of features continues to improve in these layers (Section 4.1). [2] All ViTs studied learn to use Offset Local Attention Heads, suggesting they are fundamentally necessary in ViTs (Section 4.2). To the best of our knowledge, no prior work has brought attention to this phenomenon. [3] ViTs learn to process local and global information in different orders depending on their method of supervision (Section 4.3). [4] All ViTs studied differentiate salient foreground objects by the early-to-mid layers (Section 4.4). [5] Reconstruction-based self-supervised methods can learn semantically meaningful CLS representations, even when the CLS token is only a placeholder (Section 5.1, 5.2). [6] Supervised method's features are the most semantically rich, but contrastive self-supervised methods are comparable or even superior in" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 85, + 547, + 235 + ], + "blocks": [ + { + "bbox": [ + 333, + 72, + 518, + 83 + ], + "lines": [ + { + "bbox": [ + 333, + 72, + 518, + 83 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 518, + 83 + ], + "type": "text", + "content": "Average CLS Token Attention over 5000 Images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 85, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 547, + 235 + ], + "type": "image", + "image_path": "d0962b2816c9e68af5a40832021161591bc7b9c2c8d41a895f94f74bd226f5e7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 248, + 530, + 299 + ], + "blocks": [ + { + "bbox": [ + 331, + 239, + 522, + 248 + ], + "lines": [ + { + "bbox": [ + 331, + 239, + 522, + 248 + ], + "spans": [ + { + "bbox": [ + 331, + 239, + 522, + 248 + ], + "type": "text", + "content": "FS & CLIP Layers 7-12: Sparse Repeating Patterns" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 248, + 530, + 299 + ], + "lines": [ + { + "bbox": [ + 323, + 248, + 530, + 299 + ], + "spans": [ + { + "bbox": [ + 323, + 248, + 530, + 299 + ], + "type": "image", + "image_path": "413a1b747a127219530b61b87f06f3a862f5bb7f2f70ca9a26d4f3cec6666878.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 312, + 530, + 361 + ], + "blocks": [ + { + "bbox": [ + 329, + 302, + 524, + 312 + ], + "lines": [ + { + "bbox": [ + 329, + 302, + 524, + 312 + ], + "spans": [ + { + "bbox": [ + 329, + 302, + 524, + 312 + ], + "type": "text", + "content": "DINO & MoCo Layers 7-12: Object Centered Blobs" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 312, + 530, + 361 + ], + "lines": [ + { + "bbox": [ + 323, + 312, + 530, + 361 + ], + "spans": [ + { + "bbox": [ + 323, + 312, + 530, + 361 + ], + "type": "image", + "image_path": "a03d89c8f641acb2a133b50c97e0b77506e5d15444b4fa5d766667252d6ca9a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 376, + 530, + 425 + ], + "blocks": [ + { + "bbox": [ + 331, + 365, + 522, + 376 + ], + "lines": [ + { + "bbox": [ + 331, + 365, + 522, + 376 + ], + "spans": [ + { + "bbox": [ + 331, + 365, + 522, + 376 + ], + "type": "text", + "content": "MAE & BEiT Layers 7-12: Diverse Attention Maps" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 376, + 530, + 425 + ], + "lines": [ + { + "bbox": [ + 323, + 376, + 530, + 425 + ], + "spans": [ + { + "bbox": [ + 323, + 376, + 530, + 425 + ], + "type": "image", + "image_path": "ac5ec47af50a17553e70f1b8dd2adb824389acb623802580934d47675354d6b5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 431, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 305, + 431, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 545, + 498 + ], + "type": "text", + "content": "Figure 2. Clear differences in attention emerge in the mid-to-late layers under different supervision methods. These plots show the attention maps of CLS tokens averaged over 5000 ImageNet images. Rows indicate layers and columns indicate heads. For brevity, we show only three heads per layer. The bracketed numbers in the lower half denote the layer and head." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 503, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 552 + ], + "type": "text", + "content": "some cases (Section 5.2, 5.3). [7] For localized tasks, the best performance often comes from a mid-to-late layer (Section 6.2). [8] There is no single \"best\" training method or layer for all downstream tasks (Section 6.3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 562, + 392, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 392, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 392, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "Previous works have attempted to understand the representation quality for both supervised and self-supervised training for Convolutional Neural Networks (CNNs). [5] focuses on understanding the concepts learned by individual neurons while [41] looks at explaining their compositionality in the case of supervised networks. Simultaneously, due to the popularity of self-supervised learning methods [3, 4, 9-11, 26, 27, 29, 40, 43, 60, 73] multiple works have analyzed these representations learned from no labels. Under this umbrella, [13, 62] tried to understand the effect of training data in terms of both the number and type" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7487" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "of samples. Some works [66, 67] analyze the alignment, separability, and uniformity of features while [49] looks at invariance to augmentations like occlusion, illumination, and viewpoint change in the learned representation. [64] looks at the downstream performance of self-supervised networks on fine-grained tasks. Finally, [21, 23, 32] analyze multiple self-supervised methods and compare their performance based on representation similarity and downstream task performance over multiple datasets along with comparisons to supervised methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 407 + ], + "type": "text", + "content": "Since the proliferation of ViTs, a number of works have tried to understand and explore the different properties of the representations learned by these networks. A few works [45, 55, 75] have analyzed the robustness of ViT features against corruptions, perturbations, distribution shifts, and adversarial examples while also analyzing the role of self-attention for robustness. [34] benchmarks different pretrained ViTs as backbones for object detection. [7] provides a theoretical understanding of how MAEs work while [53] analyzes attention using convex duality. [61] gives insights to train and use ViTs more efficiently. [44] gives a deeper understanding of how Multi-Headed Attention layers work while comparing and contrasting to how convolution layers behave in terms of the loss landscapes and low-pass/high-pass filtering. [52] compares fully supervised ViTs and ResNets in terms of the local and global information encoded at different depths, the role of skip connections, and the uniformity of representations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 409, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 289, + 613 + ], + "type": "text", + "content": "All these prior works either examine the impact of supervision on CNNs or compare CNNs and ViTs trained with full supervision. Some recent and concurrent works have compared the properties of differently supervised ViTs, though typically focused on a particular task and only two methods of supervision at a time. [1] compares the properties of fully supervised and DINO ViT features in the context of dense feature descriptors, and [2] further compares these two across several semantic correspondence tasks. [19] compares fully supervised and CLIP ViTs through feature visualizations. To the best of our knowledge, we present what is to date the broadest and the most in-depth comparison of ViTs with varying supervision, including six different methods covering three supervision subcategories. Additionally, we propose new attention-based analysis methods along with evaluations on multiple downstream tasks focused on both local and global information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 625, + 170, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 170, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 170, + 639 + ], + "type": "text", + "content": "3. Experimental Design" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 646, + 228, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 228, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 228, + 658 + ], + "type": "text", + "content": "3.1. A Primer on Vision Transformers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Vision Transformers (ViTs) [18] are adapted from Transformers [65] for the Natural Language Processing domain. A ViT consists of an array of tokens, each representing an image patch. In addition, most ViTs include an ex" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "content": "tra \"classifier\" or \"CLS\" token, which is connected to the task-specific output layers during training. ViTs use MultiHeaded Attention (MHA) layers [65], which use a Query-Key-Value system that allows each token to attend to all other tokens with a variable intensity attention map. This is in stark contrast to the limited receptive fields of convolutions. These layers are \"multi-headed\" because they repeat this process multiple times in parallel, allowing tokens to apply multiple attention strategies concurrently. A ViT architecture includes multiple blocks, each with one MHA layer followed by a position-wise fully connected layer. Unlike CNNs, which usually get narrower in deeper layers, ViTs maintain the same \"width\" (number of tokens) throughout. There are some transformer variants, like SWiN transformers [37], that introduce a narrowing width, but for our analyses, we focus on only traditional ViTs. Specifically, our primary analysis focuses on ViT-Base models with patch size " + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "content": " (ViT-B/16) and input size " + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "content": ", which results in a " + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 304, + 72, + 547, + 348 + ], + "type": "text", + "content": " spatial token array. ViT-Base has 12 blocks and 12 attention heads per MHA layer. In the Appendix, we provide additional results on a wider range of ViTs, including variations in architecture size and patch size." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 357, + 440, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 440, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 440, + 370 + ], + "type": "text", + "content": "3.2. Methods of Supervision" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": "Although a large number of ViT training methods have been proposed in a short span of three years, many of the most popular methods can be loosely categorized into the following three groups. From each group, we select two representative models for in-depth analysis. We further discuss these models' details in Appendix B.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": "Explicit Supervision. These models are trained with an explicit objective that is defined either by human annotations or by labels derived from another source, like paired image captions. For this category, we use a Fully Supervised (FS) ViT pretrained on ImageNet21k and fine-tuned on ImageNet1k [58, 69], as well as a CLIP ViT [51]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": "Self-Supervision (Contrastive). Self-supervised learning methods broadly attempt to train a model through a pretext task that can be directly derived from the input data. Among the more popular pretext tasks are contrastive learning methods [27, 70] which generally present a model with multiple augmented views of the same image alongside distractor views of other images. The model must learn to identify which of the views came from the same image. For this category, we select DINO [9] and MoCo-v3 [12] which we denote simply as MoCo for the rest of this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Self-Supervision (Reconstruction). Another popular category of self-supervision is reconstruction methods, which train models to predict the missing content from masked or otherwise corrupted images. We select MAE [26] and BEiT [4] for this category. Note that MAE has a separate decoder which is discarded after pretraining," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7488" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "while BEiT's decoder is learned in the same ViT. This has a strong impact on the behavior of the later layers of BEiT." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 110, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 110, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 110, + 113 + ], + "type": "text", + "content": "3.3. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 252 + ], + "type": "text", + "content": "We study the ViTs on multiple datasets and downstream tasks. Unless otherwise specified, we use ImageNet-50 [63], a subset of ImageNet [16] which narrows the dataset down to 50 representative categories. We sample 100 images per class to create a diverse collection of 5000 images. We additionally use PartImageNet [25] to measure Attention Saliency and part-level feature purity, as well as COCO [36] to measure object-level feature purity. We use revisited [50] Oxford [46] (ROxford5k) for evaluating image retrieval, DAVIS [48] for video segmentation, and SPair-71k [39] for keypoint correspondence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 258, + 157, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 157, + 271 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 157, + 271 + ], + "type": "text", + "content": "3.4. Proposed Analyses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 300 + ], + "type": "text", + "content": "Our analysis is broadly divided into three domains covering the How, What, and Why of ViTs:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 300, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 361 + ], + "type": "text", + "content": "How ViTs process local/global information (Attention). Do self-attention heads learn to operate in different ways depending on their method of training? Are there distinctive modes of attention behavior? How does supervision impact the processing order of local and global information?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 361, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 287, + 408 + ], + "type": "text", + "content": "What we take away from ViTs (Features). How do the final and intermediate representations of a ViT change depending on the method of supervision? Are these trends similar or different for CLS vs. spatial tokens?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 408, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 287, + 456 + ], + "type": "text", + "content": "Why we use ViTs (Downstream Tasks). Which forms of supervision are best suited for different downstream tasks? Which layers of a ViT produce features that are best for different local and global tasks?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 465, + 157, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 157, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 157, + 479 + ], + "type": "text", + "content": "4. Attention Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": "Multi-Headed Attention layers are one of the defining components of the Transformer architecture, and the attention maps they generate can give key insights into what is similar or different about ViTs trained through different methods. We perform an in-depth examination of the self-attention maps of ViT-B/16 models at every layer. Through this study, we uncover a diverse range of attention head behavioral modalities. Additional visualizations are provided in Appendix C for a wide range of ViT variants." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 599, + 182, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 182, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 182, + 611 + ], + "type": "text", + "content": "4.1. Attention Visualizations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "We start by examining the attention maps of the CLS tokens of each head and layer. To gain a comprehensive understanding of each head's behavior, we compute the average attention maps over 5000 ImageNet images, as shown in Figure 2. For brevity, we display only three heads per layer, but complete plots can be found in Appendix C.1, along with additional visualizations for spatial token attention and individual input images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 72, + 532, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 72, + 532, + 83 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 532, + 83 + ], + "type": "text", + "content": "Aligned Aggregated Attention Maps for Spatial Tokens" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 371, + 85, + 481, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 85, + 481, + 95 + ], + "spans": [ + { + "bbox": [ + 371, + 85, + 481, + 95 + ], + "type": "text", + "content": "Strict Local Attention Heads" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 323, + 95, + 355, + 125 + ], + "blocks": [ + { + "bbox": [ + 323, + 95, + 355, + 125 + ], + "lines": [ + { + "bbox": [ + 323, + 95, + 355, + 125 + ], + "spans": [ + { + "bbox": [ + 323, + 95, + 355, + 125 + ], + "type": "image", + "image_path": "2327a29ce828d09acaa5337f1667f8186e3f62e1ace66c919e1716603c7f88a1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 126, + 343, + 133 + ], + "lines": [ + { + "bbox": [ + 335, + 126, + 343, + 133 + ], + "spans": [ + { + "bbox": [ + 335, + 126, + 343, + 133 + ], + "type": "text", + "content": "FS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 126, + 381, + 134 + ], + "lines": [ + { + "bbox": [ + 366, + 126, + 381, + 134 + ], + "spans": [ + { + "bbox": [ + 366, + 126, + 381, + 134 + ], + "type": "text", + "content": "CLIP" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 135, + 381, + 144 + ], + "lines": [ + { + "bbox": [ + 366, + 135, + 381, + 144 + ], + "spans": [ + { + "bbox": [ + 366, + 135, + 381, + 144 + ], + "type": "text", + "content": "[1,1]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 126, + 402, + 133 + ], + "lines": [ + { + "bbox": [ + 393, + 126, + 402, + 133 + ], + "spans": [ + { + "bbox": [ + 393, + 126, + 402, + 133 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 135, + 417, + 144 + ], + "lines": [ + { + "bbox": [ + 400, + 135, + 417, + 144 + ], + "spans": [ + { + "bbox": [ + 400, + 135, + 417, + 144 + ], + "type": "text", + "content": "[1,6]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 126, + 435, + 133 + ], + "lines": [ + { + "bbox": [ + 427, + 126, + 435, + 133 + ], + "spans": [ + { + "bbox": [ + 427, + 126, + 435, + 133 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 135, + 452, + 144 + ], + "lines": [ + { + "bbox": [ + 435, + 135, + 452, + 144 + ], + "spans": [ + { + "bbox": [ + 435, + 135, + 452, + 144 + ], + "type": "text", + "content": "[2,4]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 126, + 472, + 133 + ], + "lines": [ + { + "bbox": [ + 463, + 126, + 472, + 133 + ], + "spans": [ + { + "bbox": [ + 463, + 126, + 472, + 133 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 135, + 487, + 144 + ], + "lines": [ + { + "bbox": [ + 470, + 135, + 487, + 144 + ], + "spans": [ + { + "bbox": [ + 470, + 135, + 487, + 144 + ], + "type": "text", + "content": "[2,9]" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 126, + 513, + 133 + ], + "lines": [ + { + "bbox": [ + 498, + 126, + 513, + 133 + ], + "spans": [ + { + "bbox": [ + 498, + 126, + 513, + 133 + ], + "type": "text", + "content": "BE" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 323, + 158, + 355, + 189 + ], + "blocks": [ + { + "bbox": [ + 331, + 135, + 347, + 144 + ], + "lines": [ + { + "bbox": [ + 331, + 135, + 347, + 144 + ], + "spans": [ + { + "bbox": [ + 331, + 135, + 347, + 144 + ], + "type": "text", + "content": "[1,7]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 374, + 148, + 479, + 157 + ], + "lines": [ + { + "bbox": [ + 374, + 148, + 479, + 157 + ], + "spans": [ + { + "bbox": [ + 374, + 148, + 479, + 157 + ], + "type": "text", + "content": "Soft Local Attention Heads" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 158, + 355, + 189 + ], + "lines": [ + { + "bbox": [ + 323, + 158, + 355, + 189 + ], + "spans": [ + { + "bbox": [ + 323, + 158, + 355, + 189 + ], + "type": "image", + "image_path": "638697a81c7f1ec96804679e5d95892a014916dc08ae679fcf62a33ca8ba2f6c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 190, + 343, + 198 + ], + "lines": [ + { + "bbox": [ + 335, + 190, + 343, + 198 + ], + "spans": [ + { + "bbox": [ + 335, + 190, + 343, + 198 + ], + "type": "text", + "content": "FS" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 198, + 347, + 207 + ], + "lines": [ + { + "bbox": [ + 331, + 198, + 347, + 207 + ], + "spans": [ + { + "bbox": [ + 331, + 198, + 347, + 207 + ], + "type": "text", + "content": "[3,6]" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 190, + 381, + 198 + ], + "lines": [ + { + "bbox": [ + 366, + 190, + 381, + 198 + ], + "spans": [ + { + "bbox": [ + 366, + 190, + 381, + 198 + ], + "type": "text", + "content": "CLIP" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 198, + 381, + 207 + ], + "lines": [ + { + "bbox": [ + 366, + 198, + 381, + 207 + ], + "spans": [ + { + "bbox": [ + 366, + 198, + 381, + 207 + ], + "type": "text", + "content": "[2,3]" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 190, + 402, + 198 + ], + "lines": [ + { + "bbox": [ + 393, + 190, + 402, + 198 + ], + "spans": [ + { + "bbox": [ + 393, + 190, + 402, + 198 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 198, + 416, + 207 + ], + "lines": [ + { + "bbox": [ + 400, + 198, + 416, + 207 + ], + "spans": [ + { + "bbox": [ + 400, + 198, + 416, + 207 + ], + "type": "text", + "content": "[2,9]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 190, + 435, + 198 + ], + "lines": [ + { + "bbox": [ + 427, + 190, + 435, + 198 + ], + "spans": [ + { + "bbox": [ + 427, + 190, + 435, + 198 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 323, + 220, + 355, + 252 + ], + "blocks": [ + { + "bbox": [ + 372, + 212, + 480, + 220 + ], + "lines": [ + { + "bbox": [ + 372, + 212, + 480, + 220 + ], + "spans": [ + { + "bbox": [ + 372, + 212, + 480, + 220 + ], + "type": "text", + "content": "Axial Local Attention Heads" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 220, + 355, + 252 + ], + "lines": [ + { + "bbox": [ + 323, + 220, + 355, + 252 + ], + "spans": [ + { + "bbox": [ + 323, + 220, + 355, + 252 + ], + "type": "image", + "image_path": "9c8d5b16488fa968162a8203e7ac1522c6c5e164aa13a3baec91f99eccd3538c.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 253, + 343, + 261 + ], + "lines": [ + { + "bbox": [ + 335, + 253, + 343, + 261 + ], + "spans": [ + { + "bbox": [ + 335, + 253, + 343, + 261 + ], + "type": "text", + "content": "FS" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 262, + 347, + 270 + ], + "lines": [ + { + "bbox": [ + 331, + 262, + 347, + 270 + ], + "spans": [ + { + "bbox": [ + 331, + 262, + 347, + 270 + ], + "type": "text", + "content": "[5,7]" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 253, + 381, + 261 + ], + "lines": [ + { + "bbox": [ + 366, + 253, + 381, + 261 + ], + "spans": [ + { + "bbox": [ + 366, + 253, + 381, + 261 + ], + "type": "text", + "content": "CLIP" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 323, + 284, + 355, + 316 + ], + "blocks": [ + { + "bbox": [ + 323, + 284, + 355, + 316 + ], + "lines": [ + { + "bbox": [ + 323, + 284, + 355, + 316 + ], + "spans": [ + { + "bbox": [ + 323, + 284, + 355, + 316 + ], + "type": "image", + "image_path": "943e69af392a718b2d3d9f4784ac56a586ce68dee3c1eef5e0322116f2881348.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 316, + 343, + 324 + ], + "lines": [ + { + "bbox": [ + 331, + 316, + 343, + 324 + ], + "spans": [ + { + "bbox": [ + 331, + 316, + 343, + 324 + ], + "type": "text", + "content": "FS" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 324, + 347, + 333 + ], + "lines": [ + { + "bbox": [ + 331, + 324, + 347, + 333 + ], + "spans": [ + { + "bbox": [ + 331, + 324, + 347, + 333 + ], + "type": "text", + "content": "[3,7]" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 340, + 545, + 394 + ], + "lines": [ + { + "bbox": [ + 305, + 340, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 545, + 394 + ], + "type": "text", + "content": "Figure 3. Multiple distinct forms of local attention exist. We visualize spatial token attention using Aligned Aggregated Attention Maps, and highlight different types of local attention heads, including Strict, Soft, Axial, and Offset Local Attention Heads. In row 4 we draw the mid-lines in red as a visual aid." + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 359, + 284, + 390, + 315 + ], + "blocks": [ + { + "bbox": [ + 366, + 262, + 381, + 270 + ], + "lines": [ + { + "bbox": [ + 366, + 262, + 381, + 270 + ], + "spans": [ + { + "bbox": [ + 366, + 262, + 381, + 270 + ], + "type": "text", + "content": "[5,6]" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 274, + 482, + 284 + ], + "lines": [ + { + "bbox": [ + 370, + 274, + 482, + 284 + ], + "spans": [ + { + "bbox": [ + 370, + 274, + 482, + 284 + ], + "type": "text", + "content": "Offset Local Attention Heads" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 284, + 390, + 315 + ], + "lines": [ + { + "bbox": [ + 359, + 284, + 390, + 315 + ], + "spans": [ + { + "bbox": [ + 359, + 284, + 390, + 315 + ], + "type": "image", + "image_path": "acaeed370ca87bcd31f1f12c3acd942fd369e2ba1a50a0023828a5bb72e01d14.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 316, + 381, + 324 + ], + "lines": [ + { + "bbox": [ + 366, + 316, + 381, + 324 + ], + "spans": [ + { + "bbox": [ + 366, + 316, + 381, + 324 + ], + "type": "text", + "content": "CLIP" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 324, + 381, + 333 + ], + "lines": [ + { + "bbox": [ + 366, + 324, + 381, + 333 + ], + "spans": [ + { + "bbox": [ + 366, + 324, + 381, + 333 + ], + "type": "text", + "content": "[4,1]" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 392, + 284, + 424, + 315 + ], + "blocks": [ + { + "bbox": [ + 393, + 253, + 402, + 261 + ], + "lines": [ + { + "bbox": [ + 393, + 253, + 402, + 261 + ], + "spans": [ + { + "bbox": [ + 393, + 253, + 402, + 261 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 262, + 416, + 270 + ], + "lines": [ + { + "bbox": [ + 400, + 262, + 416, + 270 + ], + "spans": [ + { + "bbox": [ + 400, + 262, + 416, + 270 + ], + "type": "text", + "content": "[6,3]" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 284, + 424, + 315 + ], + "lines": [ + { + "bbox": [ + 392, + 284, + 424, + 315 + ], + "spans": [ + { + "bbox": [ + 392, + 284, + 424, + 315 + ], + "type": "image", + "image_path": "6f76749a2f6c951fa1e3ae4683153f17b85ca119cfb3f9f2a216c27de396b6bd.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 316, + 417, + 324 + ], + "lines": [ + { + "bbox": [ + 399, + 316, + 417, + 324 + ], + "spans": [ + { + "bbox": [ + 399, + 316, + 417, + 324 + ], + "type": "text", + "content": "DINO" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 324, + 417, + 333 + ], + "lines": [ + { + "bbox": [ + 400, + 324, + 417, + 333 + ], + "spans": [ + { + "bbox": [ + 400, + 324, + 417, + 333 + ], + "type": "text", + "content": "[4,1]" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 427, + 284, + 443, + 315 + ], + "blocks": [ + { + "bbox": [ + 435, + 198, + 452, + 207 + ], + "lines": [ + { + "bbox": [ + 435, + 198, + 452, + 207 + ], + "spans": [ + { + "bbox": [ + 435, + 198, + 452, + 207 + ], + "type": "text", + "content": "[2,6]" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 253, + 435, + 261 + ], + "lines": [ + { + "bbox": [ + 427, + 253, + 435, + 261 + ], + "spans": [ + { + "bbox": [ + 427, + 253, + 435, + 261 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 262, + 452, + 270 + ], + "lines": [ + { + "bbox": [ + 435, + 262, + 452, + 270 + ], + "spans": [ + { + "bbox": [ + 435, + 262, + 452, + 270 + ], + "type": "text", + "content": "[5,5]" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 284, + 443, + 315 + ], + "lines": [ + { + "bbox": [ + 427, + 284, + 443, + 315 + ], + "spans": [ + { + "bbox": [ + 427, + 284, + 443, + 315 + ], + "type": "image", + "image_path": "83149018e84bcd3ef1fd80743dad9eedc28d1b7f111ec4542669e37aebc9e374.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 316, + 454, + 324 + ], + "lines": [ + { + "bbox": [ + 433, + 316, + 454, + 324 + ], + "spans": [ + { + "bbox": [ + 433, + 316, + 454, + 324 + ], + "type": "text", + "content": "MoCo" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 324, + 452, + 333 + ], + "lines": [ + { + "bbox": [ + 436, + 324, + 452, + 333 + ], + "spans": [ + { + "bbox": [ + 436, + 324, + 452, + 333 + ], + "type": "text", + "content": "[3,2]" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 462, + 284, + 477, + 315 + ], + "blocks": [ + { + "bbox": [ + 463, + 190, + 473, + 198 + ], + "lines": [ + { + "bbox": [ + 463, + 190, + 473, + 198 + ], + "spans": [ + { + "bbox": [ + 463, + 190, + 473, + 198 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 198, + 487, + 207 + ], + "lines": [ + { + "bbox": [ + 469, + 198, + 487, + 207 + ], + "spans": [ + { + "bbox": [ + 469, + 198, + 487, + 207 + ], + "type": "text", + "content": "[1,12]" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 253, + 473, + 261 + ], + "lines": [ + { + "bbox": [ + 463, + 253, + 473, + 261 + ], + "spans": [ + { + "bbox": [ + 463, + 253, + 473, + 261 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 262, + 487, + 270 + ], + "lines": [ + { + "bbox": [ + 469, + 262, + 487, + 270 + ], + "spans": [ + { + "bbox": [ + 469, + 262, + 487, + 270 + ], + "type": "text", + "content": "[6,9]" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 284, + 477, + 315 + ], + "lines": [ + { + "bbox": [ + 462, + 284, + 477, + 315 + ], + "spans": [ + { + "bbox": [ + 462, + 284, + 477, + 315 + ], + "type": "image", + "image_path": "64803b7c3d725314a70336e42453d4b16bd2dd785ad483cd5d697e74382b63ed.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 316, + 487, + 324 + ], + "lines": [ + { + "bbox": [ + 470, + 316, + 487, + 324 + ], + "spans": [ + { + "bbox": [ + 470, + 316, + 487, + 324 + ], + "type": "text", + "content": "MAE" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 324, + 487, + 333 + ], + "lines": [ + { + "bbox": [ + 470, + 324, + 487, + 333 + ], + "spans": [ + { + "bbox": [ + 470, + 324, + 487, + 333 + ], + "type": "text", + "content": "[5,5]" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 498, + 284, + 529, + 315 + ], + "blocks": [ + { + "bbox": [ + 505, + 135, + 521, + 144 + ], + "lines": [ + { + "bbox": [ + 505, + 135, + 521, + 144 + ], + "spans": [ + { + "bbox": [ + 505, + 135, + 521, + 144 + ], + "type": "text", + "content": "[2,3]" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 190, + 513, + 198 + ], + "lines": [ + { + "bbox": [ + 498, + 190, + 513, + 198 + ], + "spans": [ + { + "bbox": [ + 498, + 190, + 513, + 198 + ], + "type": "text", + "content": "BE" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 198, + 521, + 207 + ], + "lines": [ + { + "bbox": [ + 504, + 198, + 521, + 207 + ], + "spans": [ + { + "bbox": [ + 504, + 198, + 521, + 207 + ], + "type": "text", + "content": "[3,9]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 253, + 513, + 261 + ], + "lines": [ + { + "bbox": [ + 498, + 253, + 513, + 261 + ], + "spans": [ + { + "bbox": [ + 498, + 253, + 513, + 261 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 262, + 523, + 270 + ], + "lines": [ + { + "bbox": [ + 504, + 262, + 523, + 270 + ], + "spans": [ + { + "bbox": [ + 504, + 262, + 523, + 270 + ], + "type": "text", + "content": "[3,12]" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 284, + 529, + 315 + ], + "lines": [ + { + "bbox": [ + 498, + 284, + 529, + 315 + ], + "spans": [ + { + "bbox": [ + 498, + 284, + 529, + 315 + ], + "type": "image", + "image_path": "241349858d62828debf30be886ac74d33cf3bcaed8cf4d5810ebf39e8122dda9.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 316, + 521, + 323 + ], + "lines": [ + { + "bbox": [ + 505, + 316, + 521, + 323 + ], + "spans": [ + { + "bbox": [ + 505, + 316, + 521, + 323 + ], + "type": "text", + "content": "BEIT" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 505, + 324, + 521, + 333 + ], + "lines": [ + { + "bbox": [ + 505, + 324, + 521, + 333 + ], + "spans": [ + { + "bbox": [ + 505, + 324, + 521, + 333 + ], + "type": "text", + "content": "[2,7]" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "One of the clearest differences can be seen by comparing the mid-to-final layers. For the contrastive self-supervised methods, DINO and MoCo, the attention maps tend to be centered blobs. These heads tend to focus on salient foreground objects, so these blobs simply reflect object-centered photography bias. For the reconstruction-based methods, MAE and BEiT, we see a more diverse group of attention maps. This is likely because these methods must reconstruct all image regions, and thus their attention in the final layers must be more diverse and cover more of the image. Finally, for the explicitly supervised methods, FS and CLIP, the mid-to-final layers do not focus on salient object regions and instead focus on Sparse Repeating Patterns with seemingly no spatial meaning. This occurs for both the CLS tokens and spatial tokens, and the patterns are repeated across both heads and layers. We hypothesize that these patterns occur because the mid-to-late layers are no longer focused on parsing the scene structure, and instead are using their processing power to generate their final decisions for their respective tasks. This phenomenon helps to explain why the attention maps of the later layers of fully-supervised ViTs are poorly suited for segmentation tasks, as was observed by [9]." + } + ] + } + ], + "index": 75 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7489" + } + ] + } + ], + "index": 76 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 273, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 273, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 273, + 84 + ], + "type": "text", + "content": "4.2. Emergence of Offset Local Attention Heads" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 281 + ], + "type": "text", + "content": "It has been shown that ViTs use a mixture of short-range \"local\" and long-range \"global\" attention heads in any given layer [9, 18]. To gain a better understanding of local attention, we propose a visualization strategy of Aligned Aggregated Attention Maps (AAAMs). We extract all spatial token attention maps for 5000 ImageNet images, but before averaging them, we first realigned them so the current spatial token is always in the center of the array. Additional samples of AAAMs are provided in Appendix C.1. Studying these aligned views reveals multiple forms of local attention, shown in Figure 3. We find Strict Local Attention Heads, which attend almost completely to their own position, as well as Soft Local Attention Heads, which attend to a wider neighborhood around them. We also find Axial Local Attention Heads, which are elongated to attend to the local neighborhood along one or both spatial axes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 288, + 533 + ], + "type": "text", + "content": "But perhaps the most noteworthy type of head we observe is the Offset Local Attention Head. These are heads that attend locally, but to a point or region offset from the current token in a vertical or horizontal direction. We find instances of Offset Local Attention Heads in all the models examined, suggesting they are fundamentally necessary for ViTs. To the best of our knowledge, ours is the first work to draw attention to this phenomenon. We believe that such heads are absolutely necessary because ViTs, unlike CNNs, do not have an easy built-in way to test if two features occur next to each other with a particular spatial arrangement. In a CNN, this type of check is naturally embedded into the convolution operator. But in a ViT, there is no inductive bias to induce such a check. For comparison, Soft Local Attention Heads are able to identify if a certain feature is near another feature, but they cannot identify their specific directional arrangement due to their symmetrical attention pattern. The existence of Offset Local Attention Heads implies one possible path for improvement for the ViT architecture, possibly by adding a self-attention variant that introduces some implicit directional structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 540, + 198, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 198, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 198, + 552 + ], + "type": "text", + "content": "4.3. Average Attention Distance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": "We measure the Average Attention Distance [18, 52] of each head to assess if particular heads have a short-range \"local\" focus or a long-range \"global\" focus. This metric is computed by measuring the distance from each spatial token to all other tokens and taking a weighted average using the attention map. We normalize the distances so the token grid is embedded on a " + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 557, + 287, + 677 + ], + "type": "text", + "content": " square. [52] observed that for a well-trained fully-supervised ViT, the early layers have a mixture of local and global attention heads, while the later layers have only global attention heads." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Figure 4 (left) shows the Average Attention Distances of all heads organized by layer and model. Like [52], we see that most layers use a mixture of local and global attention" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 425, + 179 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 425, + 179 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 425, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 425, + 179 + ], + "type": "image", + "image_path": "55ccde2c27836ef39ead72527fc1066c5c83943dda02ad91e2253f22fb12a7cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 185, + 518, + 194 + ], + "lines": [ + { + "bbox": [ + 333, + 185, + 518, + 194 + ], + "spans": [ + { + "bbox": [ + 333, + 185, + 518, + 194 + ], + "type": "text", + "content": "FS CLIP DINO MoCo MAE BEiT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 304, + 202, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 247 + ], + "type": "text", + "content": "Figure 4. Different methods of supervision lead to different orderings and ratios of local and global processing. We show the Average Attention Distance of all ViT attention heads organized by layer (left), and the per-layer averages (right)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 545, + 179 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 545, + 179 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 545, + 179 + ], + "type": "image", + "image_path": "3fdb8e3d6d034a8a370d4a2dcfc18c2566db141c46eb741abad6389fe7c6a5b0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 428, + 185, + 518, + 194 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 428, + 185, + 518, + 194 + ], + "spans": [ + { + "bbox": [ + 428, + 185, + 518, + 194 + ], + "type": "text", + "content": "MoCo MAE BEiT" + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 260, + 426, + 369 + ], + "blocks": [ + { + "bbox": [ + 308, + 260, + 426, + 369 + ], + "lines": [ + { + "bbox": [ + 308, + 260, + 426, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 426, + 369 + ], + "type": "image", + "image_path": "2570959cb9d1ca8b3f3f4a4f398bad321eba47e1e9c92b99a9f564b330ca5fff.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 373, + 518, + 383 + ], + "lines": [ + { + "bbox": [ + 333, + 373, + 518, + 383 + ], + "spans": [ + { + "bbox": [ + 333, + 373, + 518, + 383 + ], + "type": "text", + "content": "FS CLIP DINO MoCo MAE BEiT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 304, + 390, + 546, + 434 + ], + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 434 + ], + "type": "text", + "content": "Figure 5. Attention IoU with salient content plateaus early for all ViTs evaluated. We calculate the alignment of ground-truth segmentation masks with CLS token attention maps (left) and the average of spatial token attention maps (right)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 428, + 260, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 428, + 260, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 428, + 260, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 428, + 260, + 545, + 369 + ], + "type": "image", + "image_path": "eb6bc27b4ec22aad2b0df63940aa8d613413300756edb761395eac987ad13526.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 422, + 373, + 518, + 382 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 373, + 518, + 382 + ], + "spans": [ + { + "bbox": [ + 422, + 373, + 518, + 382 + ], + "type": "text", + "content": "MoCo MAE BEiT" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 304, + 443, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 647 + ], + "type": "text", + "content": "heads, however, we also find that the ordering of local and global processing varies greatly with the supervision type. FS, CLIP, DINO, and MoCo all use exclusively global attention heads in the last layers, but the reconstruction-based methods MAE and BEiT use a diverse range of heads in their later layers. Figure 4 (right) compares the combined Average Attention Distances at the per-layer level. In all models, we observe a greater number of global attention heads in the initial layers, followed by decreased distances around layers 3-6. This result is again in contrast to [52]. The behaviors diverge in the mid-to-late layers. For the models trained with explicit or contrastive supervision, the Attention Distance trends upward in the later layers. For the reconstruction-based methods, the Average Attention Distances stay lower. These results show that, unlike CNNs, ViTs can learn a variable local/global processing order depending on the training method used." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 658, + 524, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 524, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 524, + 670 + ], + "type": "text", + "content": "4.4. Attention Alignment with Salient Content" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "One of the most desirable (and exploitable) features of DINO is that the CLS token attention maps of the last layer tend to be well-aligned with salient foreground objects [9]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7490" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 167, + 174 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 167, + 174 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 167, + 174 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 167, + 174 + ], + "type": "image", + "image_path": "a5b144029558877b69cd52d2f0d8e67bc7fb5396a933a5873dbcd7f15d3b50c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 180, + 287, + 203 + ], + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 203 + ], + "type": "text", + "content": "Figure 6. CKA similarity between final layer features of different ViTs for their CLS tokens (left) and spatial tokens (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 71, + 281, + 174 + ], + "blocks": [ + { + "bbox": [ + 168, + 71, + 281, + 174 + ], + "lines": [ + { + "bbox": [ + 168, + 71, + 281, + 174 + ], + "spans": [ + { + "bbox": [ + 168, + 71, + 281, + 174 + ], + "type": "image", + "image_path": "91542c7060b528dee1eb1f232faa662447128b040739e34a8758831a728936c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 369 + ], + "type": "text", + "content": "Several methods propose to use DINO attention maps, feature maps, or a combination of the two to generate segmentations in a self-supervised manner [24, 57, 68]. We conduct a quantitative analysis of this property at all layers of the ViTs, both to measure the usefulness of masks and to assess how early the ViTs differentiate salient object regions. Like [9], we threshold the CLS token attention masks keeping " + }, + { + "bbox": [ + 46, + 213, + 287, + 369 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 213, + 287, + 369 + ], + "type": "text", + "content": " of the total attention mass. We then compute the Intersection over Union (IoU) of said masks with ground-truth segmentations from PartImageNet [25]. As an alternative to CLS token attention, we also extract masks using the average of spatial token attention maps. We present results for the single \"best\" head per layer in Figure 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 537 + ], + "type": "text", + "content": "We see a clear drop in FS and CLIP mask IoU around the middle of the network, which directly corresponds to the emergence of the Sparse Repeating Patterns observed in Section 4.1. We also find that the IoUs plateau around layers 3-6 for all networks. This demonstrates that ViT models already have a solid understanding of foreground/background separation by the middle layers. While the later attention maps of FS and CLIP are much worse than their self-supervised counterparts, their early-to-mid layers are more comparable. We find that MoCo, MAE, and BEiT can all produce attention maps with IoUs that are comparable with DINO. In addition, we see that the average of spatial tokens produces maps that are comparable with the CLS token, and for CLIP the IoU increases greatly in the final layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 148, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 148, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 148, + 561 + ], + "type": "text", + "content": "5. Feature Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 286, + 639 + ], + "type": "text", + "content": "In this section, we directly compare ViT features across models and layers using Centered Kernel Alignment (CKA) [14, 31]. We also study unsupervised clustering performance to compare global and local semantic information in the learned representations. We present addition analysis in Appendix D.4 focused on ViT residual connections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 224, + 659 + ], + "type": "text", + "content": "5.1. Last Block Feature Comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Comparing representations is non-trivial due to varying feature sizes, large feature representations, and lack of alignment between them. To overcome this, we use batched Centered Kernel Alignment (CKA) [14, 31, 42] which can" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "align features and compute a similarity score. We compare the last layer outputs for each model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 97, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 545, + 276 + ], + "type": "text", + "content": "Figure 6 (left) shows that the CLS token representations are usually similar for similar supervision strategies (explicit, contrastive, reconstruction). The contrastive methods, MoCo and DINO, show very high similarity to each other, indicating that the CLS token encodes the same type of information for both these methods. There is also an increased level of similarity between the explicitly supervised methods, FS and CLIP, and the contrastive methods. Interestingly, we see that MAE has as high a similarity with DINO and MoCo as it does with BEiT. This result is surprising because MAE's CLS token has no explicit training objective or loss, and the way these approaches are trained is very different. This presents evidence that training autoencoders with a high masking percentage indeed forces the model to learn image-level semantics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 384 + ], + "type": "text", + "content": "In Figure 6 (right), we look at the similarity of the last-layer spatial token representations. Unlike the CLS token representations, CLIP and FS have low similarity in their spatial representations. The self-supervised methods DINO, MAE, and MoCo show a high level of similarity to each other, and a lower level of similarity to BEiT. MoCo and DINO show the highest similarity due to their similar kind of self-supervision. Once again, MAE has a high similarity to MoCo and DINO despite their very different supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 394, + 517, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 517, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 517, + 407 + ], + "type": "text", + "content": "5.2. Feature Clustering for Global Semantics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 414, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 556 + ], + "type": "text", + "content": "Through this analysis, we aim to test how well the learned CLS and spatial token representations encode global (image-level) semantic information at every layer. We extract the CLS token features from the end of each block for 5000 ImageNet images, and we generate k-Means cluster assignments with " + }, + { + "bbox": [ + 304, + 414, + 545, + 556 + ], + "type": "inline_equation", + "content": "k = 50" + }, + { + "bbox": [ + 304, + 414, + 545, + 556 + ], + "type": "text", + "content": ". We present results for cluster purity measured with respect to ground truth image labels, but additional clustering metrics are also presented in Appendix D.5. For the spatial tokens, we follow the same process except we average-pool over all positions before clustering. We also compute a random chance score by replacing the ViT features with Gaussian random noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "For CLS token features, shown in Figure 7 (left), cluster purity improves with depth with the exception of the last layers of BEiT. This is likely because the last layers of BEiT serve as a task-specific decoder, unlike MAE, where the decoder is separate and discarded after pretraining. Unsurprisingly, FS achieves the best cluster purity, followed by CLIP. The contrastive methods, DINO and MoCo, achieve scores close to the explicitly supervised methods. The reconstruction-based methods, MAE and BEiT, have the lowest cluster purity, but they are still above random chance, which again indicates that they do learn to encode some image-level semantic information in their CLS tokens. Also, we find that semantic information emerges earlier for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7491" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 289, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 194 + ], + "type": "image", + "image_path": "e73f13a940cb786fa28fa1b6ea032f5593708b1644091a01e6d1861956ca1ef4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "type": "text", + "content": "Figure 7. Clustering purity analysis with image-level labels in ImageNet-50 for CLS features (left) and average-pooled spatial token features (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 248, + 287, + 372 + ], + "blocks": [ + { + "bbox": [ + 50, + 248, + 287, + 372 + ], + "lines": [ + { + "bbox": [ + 50, + 248, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 50, + 248, + 287, + 372 + ], + "type": "image", + "image_path": "be76f898de49b7ef4068642da8488cc6c5316a7c131bd1cf94030f0a15a394de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 403 + ], + "type": "text", + "content": "Figure 8. Clustering purity of spatial token features gathered at the object-level in COCO (left) and part-level in PartImageNet (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 286, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 483 + ], + "type": "text", + "content": "DINO and MoCo. For the spatial token features, shown in Figure 7 (right), the cluster purity of FS rises earlier compared with the FS CLS token. This suggests that the FS spatial tokens do more work gathering semantic information in the early layers. For all other ViTs, the spatial feature purity is lower in the final layers, but is comparable in layers 1-7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 491, + 254, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 254, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 254, + 504 + ], + "type": "text", + "content": "5.3. Feature Clustering for Local Semantics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "We next measure how well the spatial token features differentiate salient image content at the object or part-level using COCO [36] and PartImageNet [25] respectively. We use a tiling strategy to extract a denser array of features, which we detail in Appendix B.4. Using ground truth segmentation masks, we extract and average the features of the tokens overlapping with the masks. This generates a collection of object-level or part-level features which we cluster just like Section 5.2. The results for object-level features are shown in Figure 8 (left). We see that the supervised methods CLIP and FS have the highest feature purity by far, followed by the contrastive methods DINO and MoCo. The purity is much lower for the reconstruction methods MAE and BEiT. For part-level features, shown in Figure 8 (right), FS achieves the best purity, but the contrastive methods are very competitive in this case, surpassing CLIP completely. In addition, while still being the lowest scoring, MAE and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "BEiT are much more competitive at the part-level. Like the image-level purity, the object and part-level feature purity tend to improve with depth, but the purity peaks early around layers 9 to 11. The peak for BEiT is even earlier, likely due to its integrated decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 142, + 460, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 142, + 460, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 460, + 156 + ], + "type": "text", + "content": "6. Downstream Task Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 161, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 257 + ], + "type": "text", + "content": "Finally, we analyze the performance of these models on downstream tasks that can be performed directly without any fine-tuning or training. We follow the evaluation protocols of [9, 30] for " + }, + { + "bbox": [ + 304, + 161, + 545, + 257 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 161, + 545, + 257 + ], + "type": "text", + "content": "-NN classification, image retrieval, and video object segmentation. We also perform keypoint correspondence as a more local-focused task. Again we compute random chance scores by replacing all ViT features with Gaussian noise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 264, + 389, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 389, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 389, + 275 + ], + "type": "text", + "content": "6.1. Global Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": "ImageNet Classification. We perform " + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": "-Nearest Neighbor (" + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": "-NN) image classification on ImageNet [16] with " + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 304, + 281, + 545, + 437 + ], + "type": "text", + "content": ". We use the CLS token features from each network and assign the label for a test sample based on the training set features and labels. As can be seen from Figure 9 (left), FS performs the best as it has been trained to classify the same dataset. DINO and MoCo follow a similar trend as Section 5.2 and better encode semantic information in the earlier layers. FS and CLIP also follow a similar trend where their performance shoots up in the last few layers. It is also interesting to see how MAE and BEiT, for which the CLS tokens have no explicit objective, do better than chance, although MAE is considerably better than BEiT." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": "Image Retrieval. Similar to " + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": "-NN classification, we utilize the CLS token representation for retrieval. We evaluate on ROxford5k [50] for the Medium (M) split and report the Mean Average Precision (mAP). In Appendix E.2 we also report results for the Hard (H) split and the RParis6k [50] dataset, which follow similar trends. The results, shown in Figure 9 (right), align closely with those for " + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": "-NN Classification on ImageNet. FS performs the best followed by CLIP and then DINO and MoCo, and finally by MAE and BEiT with the lowest performance. We hypothesize that the local/global crops used in DINO training help it perform competitively in these global tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 587, + 384, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 384, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 384, + 599 + ], + "type": "text", + "content": "6.2. Local Tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "DAVIS Segmentation Propagation. DAVIS Segmentation Propagation is a dense prediction-based video localized task where frame-by-frame features are used to propagate the first frame segmentation mask to subsequent frames. Like Section 5.3, we use a tiling-based dense feature extraction strategy. Results are shown in Figure 10 (left). The contrastive techniques of DINO and MoCo perform the best while FS and CLIP, which are more image-level approaches, face a drop in performance towards the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7492" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 169, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 169, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 169, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 169, + 178 + ], + "type": "image", + "image_path": "936de7d6abd4083539243b9877b2cb9bb09286027b82a12a444d544ece520940.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 182, + 283, + 193 + ], + "lines": [ + { + "bbox": [ + 50, + 182, + 283, + 193 + ], + "spans": [ + { + "bbox": [ + 50, + 182, + 283, + 193 + ], + "type": "text", + "content": "FS CLIP DINO MoCo MAE BEiT -Random" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 286, + 178 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 286, + 178 + ], + "type": "image", + "image_path": "041bc83afe63922ed3a6d5ef14d74430bbf7b41f1c551cdcaca60c6a2676ff2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 240, + 168, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "type": "text", + "content": "Figure 9. Global (image-level) downstream task analysis using the CLS token. We present " + }, + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 202, + 287, + 236 + ], + "type": "text", + "content": "-NN classifier Top-1 Accuracy on ImageNet (left) and image retrieval mAP on ROxford5k (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 240, + 168, + 350 + ], + "lines": [ + { + "bbox": [ + 50, + 240, + 168, + 350 + ], + "spans": [ + { + "bbox": [ + 50, + 240, + 168, + 350 + ], + "type": "image", + "image_path": "58ee801ec5b6cd8d9c3532baf64dbd31e0342dd1f2364439adb33a36b62194a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 353, + 283, + 362 + ], + "lines": [ + { + "bbox": [ + 50, + 353, + 283, + 362 + ], + "spans": [ + { + "bbox": [ + 50, + 353, + 283, + 362 + ], + "type": "text", + "content": "FS CLIP DINO MoCo MAE BEiT -Random" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 169, + 240, + 286, + 350 + ], + "blocks": [ + { + "bbox": [ + 169, + 240, + 286, + 350 + ], + "lines": [ + { + "bbox": [ + 169, + 240, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 169, + 240, + 286, + 350 + ], + "type": "image", + "image_path": "ffe40f7ed68cbe91e11cf0c357b9c12406812438ead6e24deaa1549715b261cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "type": "text", + "content": "Figure 10. Local (pixel-level) downstream task analysis using the dense spatial token features. We perform DAVIS video segmentation (left) and SPair-71k keypoint correspondence (right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 484 + ], + "type": "text", + "content": "later layers. The local reconstruction-based methods, MAE and BEiT, are also much more competitive in this task. These results show that for purposes of constructing highly descriptive local features, contrastive methods like DINO and MoCo and reconstructive methods like BEiT can surpass features trained with explicit image-level supervision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 484, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 627 + ], + "type": "text", + "content": "Keypoint Correspondence. We choose keypoint correspondence as an additional local-focused downstream task. Given an image with annotated keypoints, the model must predict the position of corresponding keypoints in a paired image with similar content. We use the SPair-71k [39] dataset and follow the evaluation protocol of [1] and report the Percentage of Correct Keypoints (PCK) [72]. The results are summarized in Figure 10 (right). CLIP excels at this task, closely followed by both FS and DINO. Meanwhile, MoCo, BEiT, and MAE are all very competitive also. The position of the best layer varies significantly, from 8 for CLIP and BEiT to 11 for MAE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 219, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 219, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 219, + 647 + ], + "type": "text", + "content": "6.3. Summary of Downstream Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "We summarize the best results for all downstream tasks in Table 1. We denote the best-performing layers in parenthesis. These results show that ViTs with different supervision methods [1] peak at different layers, and [2] perform best at different tasks. In image-level tasks like " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "-NN clas" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 102, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Best performance for each ViT on each downstream task with the corresponding best layer in parenthesis." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 102, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 307, + 102, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 102, + 545, + 213 + ], + "type": "table", + "html": "
ModelTask Performance (Best Performing Layer)
Dataset MetricImageNet Top-1↑ROxford5k (M) mAP↑Davis J and F Mean↑SPair-71k PCK@0.1↑
FS83.79 (12)0.45 (12)0.59 (8)28.56 (9)
CLIP75.75 (12)0.40 (12)0.60 (9)30.70 (8)
DINO76.06 (12)0.37 (12)0.60 (12)28.28 (9)
MoCo71.59 (12)0.31 (12)0.61 (11)25.85 (9)
MAE45.19 (12)0.15 (10)0.54 (12)22.74 (11)
BEiT26.84 (8)0.14 (8)0.58 (9)24.11 (8)
Random0.100.020.061.32
", + "image_path": "f865bc0be458f9f2804cd846f54241265695f52d5be51785d91988b105a83e85.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 232, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 293 + ], + "type": "text", + "content": "sification and retrieval, usually, the last layer works the best. For localized tasks like keypoint correspondence and video object segmentation, most models' peak performance happens a few layers before the last one. This shows that always picking the last layer output is not optimal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 314, + 378, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 378, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 378, + 326 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 338, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 624 + ], + "type": "text", + "content": "In this work, we have performed an in-depth comparison of ViTs trained through different methodologies by examining their attention patterns, learned representations, and downstream task performance. We review some of the key findings of our analyses. First, different methods of supervision lead to ViTs that process local and global information in different orders. All ViTs have heads that align well with salient image content, but for the explicitly supervised models, the late-layer attention maps change into Sparse Repeating Patterns. In addition, all ViTs examined have learned to use Offset Local Attention Heads in multiple layers. While explicitly supervised ViTs have the most semantically rich representations at the image level, contrastive methods are competitive, and reconstruction-based methods can also learn meaningful CLS token representations even though said token is a placeholder and has no explicit supervision. Finally, there is no single best model for all the downstream tasks, and the best layer to extract representations from also varies greatly by task and model, so one should not simply take the last layer representation. ViTs have shown a great deal of potential, and we expect they will become more widely used in the coming years. We hope these insights can help with the future development of losses and architectures for Vision Transformers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and DARPA GARD (HR00112020007) programs. We would like to thank our colleagues Vatsal Agarwal, Matthew Gwilliam, Alex Hanson, Pulkit Kumar, Saketh Rambhatla, and Gaurav Shrivastava for their feedback on this work." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7493" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. ECCVW What is Motion For?, 2022. 3, 8, 24" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 177 + ], + "type": "text", + "content": "[2] Mehmet Aygün and Oisin Mac Aodha. Demystifying unsupervised semantic correspondence estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXX, pages 125-142. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 220 + ], + "type": "text", + "content": "[3] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 251 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 3, 12, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 306 + ], + "type": "text", + "content": "[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 371 + ], + "type": "text", + "content": "[6] Chunshui Cao, Xianming Liu, Yi Yang, Yinan Yu, Jiang Wang, Zilei Wang, Yongzhen Huang, Liang Wang, Chang Huang, Wei Xu, et al. Look and think twice: Capturing top-down visual attention with feedback convolutional neural networks. In Proceedings of the IEEE international conference on computer vision, pages 2956-2964, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 371, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 402 + ], + "type": "text", + "content": "[7] Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 456 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 510 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 2, 3, 4, 5, 6, 7, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 553, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 585 + ], + "type": "text", + "content": "[11] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 585, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 585, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 585, + 287, + 628 + ], + "type": "text", + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 3, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 628, + 288, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 288, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 288, + 681 + ], + "type": "text", + "content": "[13] Elijah Cole, Xuan Yang, Kimberly Wilber, Oisin Mac Aodha, and Serge Belongie. When does contrastive visual representation learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14755-14764, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 703 + ], + "type": "text", + "content": "[14] Corinna Cortes, Mehryar Mohri, and Afshin Rostamizadeh. Algorithms for learning kernels based on centered alignment." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "The Journal of Machine Learning Research, 13:795-828, 2012. 6, 18" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "type": "text", + "content": "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 148, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 191 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 191, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 191, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 545, + 234 + ], + "type": "text", + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 234, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 298 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "text", + "content": "[19] Amin Ghiasi, Hamid Kazemi, Eitan Borgnia, Steven Reich, Manli Shu, Micah Goldblum, Andrew Gordon Wilson, and Tom Goldstein. What do vision transformers learn? a visual exploration. arXiv preprint arXiv:2212.06727, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 342, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 395 + ], + "type": "text", + "content": "[20] Kristen Grauman and Trevor Darrell. The pyramid match kernel: Discriminative classification with sets of image features. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1458-1465. IEEE, 2005. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 396, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 437 + ], + "type": "text", + "content": "[21] Tom George Grigg, Dan Busbridge, Jason Ramapuram, and Russ Webb. Do self-supervised and supervised methods learn similar visual representations? arXiv preprint arXiv:2110.00528, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 504 + ], + "type": "text", + "content": "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 503, + 547, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 547, + 556 + ], + "type": "text", + "content": "[23] Matthew Gwilliam and Abhinav Shrivastava. Beyond supervised vs. unsupervised: Representative benchmarking and analysis of image representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9642-9652, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 600 + ], + "type": "text", + "content": "[24] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "text", + "content": "[25] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021. 4, 6, 7, 18" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 643, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 643, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 643, + 545, + 696 + ], + "type": "text", + "content": "[26] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 3, 12, 14" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 696, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 696, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 696, + 545, + 707 + ], + "type": "text", + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7494" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 707 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 159 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904-1916, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "type": "text", + "content": "[29] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 213, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 287, + 255 + ], + "type": "text", + "content": "[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 288, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 288, + 299 + ], + "type": "text", + "content": "[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 6, 18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "text", + "content": "[32] Klemen Kotar, Gabriel Ilharco, Ludwig Schmidt, Kiana Ehsani, and Roozbeh Mottaghi. Contrasting contrastive self-supervised representation learning pipelines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9949-9959, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "text", + "content": "[33] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 2169-2178. IEEE, 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 407, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 449 + ], + "type": "text", + "content": "[34] Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaiming He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 503 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 7, 18" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "text", + "content": "[37] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 610, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 287, + 632 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 632, + 287, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 663 + ], + "type": "text", + "content": "[39] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 4, 8, 24" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 663, + 287, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 663, + 287, + 707 + ], + "spans": [ + { + "bbox": [ + 48, + 663, + 287, + 707 + ], + "type": "text", + "content": "[40] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6707-6717, 2020. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 706 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[41] Jesse Mu and Jacob Andreas. Compositional explanations of neurons. Advances in Neural Information Processing Systems, 33:17153-17163, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 148 + ], + "type": "text", + "content": "[42] Thao Nguyen, Maithra Raghu, and Simon Kornblith. Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth. arXiv preprint arXiv:2010.15327, 2020. 6, 18" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 148, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 180 + ], + "type": "text", + "content": "[43] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 180, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 180, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 545, + 201 + ], + "type": "text", + "content": "[44] Namuk Park and Songkuk Kim. How do vision transformers work? arXiv preprint arXiv:2202.06709, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 201, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 233 + ], + "type": "text", + "content": "[45] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2071-2081, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 233, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 286 + ], + "type": "text", + "content": "[46] James Philbin, Ondrej Chum, Michael Isard, Josef Sivic, and Andrew Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 287, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 328 + ], + "type": "text", + "content": "[47] Pedro O Pinheiro, Tsung-Yi Lin, Ronan Collobert, and Piotr Dóllár. Learning to refine object segments. In European conference on computer vision, pages 75-91. Springer, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 329, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 372 + ], + "type": "text", + "content": "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "type": "text", + "content": "[49] Senthil Purushwalkam and Abhinav Gupta. Demystifying contrastive self-supervised learning: Invariances, augmentations and dataset biases. Advances in Neural Information Processing Systems, 33:3407-3418, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 469 + ], + "type": "text", + "content": "[50] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5706-5715, 2018. 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 469, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 469, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 469, + 545, + 534 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 534, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 587 + ], + "type": "text", + "content": "[52] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 1, 3, 5, 23" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 587, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 631 + ], + "type": "text", + "content": "[53] Arda Sahiner, Tolga Ergen, Batu Ozturkler, John Pauly, Morteza Mardani, and Mert Pilanci. Unraveling attention via convex duality: Analysis and interpretations of vision transformers. arXiv preprint arXiv:2205.08078, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 631, + 545, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 631, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 307, + 631, + 545, + 673 + ], + "type": "text", + "content": "[54] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 24" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 673, + 545, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 673, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 307, + 673, + 545, + 706 + ], + "type": "text", + "content": "[55] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7495" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 705 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[56] Abhinav Shrivastava, Rahul Sukthankar, Jitendra Malik, and Abhinav Gupta. Beyond skip connections: Top-down modulation for object detection. arXiv preprint arXiv:1612.06851, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 116, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 116, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 49, + 116, + 287, + 169 + ], + "type": "text", + "content": "[57] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 170, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 170, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 170, + 287, + 213 + ], + "type": "text", + "content": "[58] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 3, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 213, + 285, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 213, + 285, + 234 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 285, + 234 + ], + "type": "text", + "content": "[59] Anand Subramanian. Torch cka. https://github.com/AntixK/PyTorch-Model-Compare, 2021.18" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 235, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 235, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 49, + 235, + 287, + 266 + ], + "type": "text", + "content": "[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In European conference on computer vision, pages 776-794. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 266, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 266, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 266, + 287, + 308 + ], + "type": "text", + "content": "[61] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 361 + ], + "type": "text", + "content": "[62] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34:16238-16250, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 362, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 362, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 49, + 362, + 287, + 415 + ], + "type": "text", + "content": "[63] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 416, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 416, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 416, + 287, + 470 + ], + "type": "text", + "content": "[64] Grant Van Horn, Elijah Cole, Sara Beery, Kimberly Wilber, Serge Belongie, and Oisin Mac Aodha. Benchmarking representation learning for natural world image collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12884-12893, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 471, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 513 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 514, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 287, + 555 + ], + "type": "text", + "content": "[66] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 555, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 555, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 49, + 555, + 287, + 598 + ], + "type": "text", + "content": "[67] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 599, + 287, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 599, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 49, + 599, + 287, + 662 + ], + "type": "text", + "content": "[68] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14543-14553, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 663, + 286, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 663, + 286, + 695 + ], + "spans": [ + { + "bbox": [ + 49, + 663, + 286, + 695 + ], + "type": "text", + "content": "[69] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.3, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 696, + 286, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 696, + 286, + 705 + ], + "spans": [ + { + "bbox": [ + 49, + 696, + 286, + 705 + ], + "type": "text", + "content": "[70] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 319 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 115 + ], + "type": "text", + "content": "Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 117, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 148 + ], + "type": "text", + "content": "[71] Kun Xu, Yong Li, Tao Ju, Shi-Min Hu, and Tian-Qiang Liu. Efficient affinity-based edit propagation using kd tree. ACM Transactions on Graphics (TOG), 28(5):1-6, 2009. 24" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 148, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 190 + ], + "type": "text", + "content": "[72] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 8, 24" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 191, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 545, + 234 + ], + "type": "text", + "content": "[73] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 235, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 545, + 266 + ], + "type": "text", + "content": "[74] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 266, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 319 + ], + "type": "text", + "content": "[75] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In International Conference on Machine Learning, pages 27378-27394. PMLR, 2022. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7496" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_content_list.json b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f0a3df2d35f7f249b720d79efef34d01ceb3058 --- /dev/null +++ b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_content_list.json @@ -0,0 +1,1536 @@ +[ + { + "type": "text", + "text": "Teaching Structured Vision & Language Concepts to Vision & Language Models", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sivan Doveh $^{1,2}$ , Assaf Arbelle $^{1}$ , Sivan Harary $^{1}$ , Eli Schwartz $^{1,3}$ , Roei Herzig $^{1,3}$ , Raja Giryes $^{3}$ , Rogerio Feris $^{4}$ , Rameswar Panda $^{4}$ , Shimon Ullman $^{*2}$ , Leonid Karlinsky $^{*4}$", + "bbox": [ + 140, + 179, + 841, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1IBM Research, 2Weizmann Institute of Science, 3Tel-Aviv University, 4MIT-IBM Watson AI Lab", + "bbox": [ + 84, + 234, + 883, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to $15\\%$ in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC", + "bbox": [ + 75, + 316, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 750, + 209, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent Vision & Language (VL) models [19, 31, 43, 44, 47, 57] achieve excellent zero-shot performance with respect to various computer-vision tasks such as detection, classification, segmentation, etc. However, recent studies [68, 82] have demonstrated that even the strongest VL models struggle with the compositional understanding of some basic Structured VL Concepts (SVLC) such as ob-", + "bbox": [ + 75, + 773, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/130aaaae7f82b9506eb17458c9f4fa69eccce7e95d52727c6908bd568f38a1ad.jpg", + "image_caption": [ + "Figure 1. Teaching language structure to VL models. (a) Standard contrastive text-to-image loss (e.g. CLIP [57]) tends to under-emphasize SVLC content of the text, likely due to the random nature of the training batches; (b) We generate modified versions of corresponding texts and use them to add losses to explicitly teach language structure (SVLC) to VL models." + ], + "image_footnote": [], + "bbox": [ + 557, + 284, + 830, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ject attributes, inter-object relations, transitive actions, object states and more. Collecting specialized large scale data to teach VL models these missing 'skills' is impractical, as finding specialized text-image pairs for each kind and possible value of the different attributes, relations, or states, is both difficult and expensive.", + "bbox": [ + 496, + 672, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another important challenge in training VL models with new concepts is catastrophic forgetting, which is a common property to all neural models [7, 34, 37, 49, 58] and has been explored for VL models in a recent concurrent work [14]. Large VL models such as CLIP [57] and CyCLIP [19] have exhibited excellent zero-shot learning abilities in many tasks. Therefore, even given a large dataset with new concepts, it is important not to lose these abilities when performing the adaptation to the new data.", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 887, + 202, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2657", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4eb3b49556f4bacdf3f366b210e8a3a0c60d28259ceb52609f8cf3fbecf243a4.jpg", + "image_caption": [ + "Figure 2. Teaching structured image understanding to VL models via structured textual data manipulation harnessing the power of language modeling. (1) Generating Rule-Based negative texts (Sec. 3.1.1); (2) Generating negatives using Large Language Model (LLM) unmasking (Sec. 3.1.2); (3) Generating analogies (positives) via LLM prompting (Sec. 3.1.3)." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 890, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a way to leverage existing (off-the-shelf) VL pre-training data sources in order to improve the SVLC understanding skills of a given model, while at the same time maintaining its zero-shot object recognition accuracy. Naturally, succeeding in this goal would lead to potential improvement w.r.t. SVLC understanding in a wide variety of downstream tasks building upon pre-trained VL models, such as zeros-shot detection, segmentation, image generation, and many more.", + "bbox": [ + 75, + 410, + 470, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research [68, 82] has shown that VL models exhibit an 'object bias' partially due to the contrastive text-to-image loss used in their pre-training. For example, the popular CLIP-loss [57] is computed over a random batch of text-image pairs sampled from a large-scale and diverse VL dataset with the chance of two images in the same batch containing the same set of objects being very low. For such a loss, representing just a 'bag of objects' in each image or text is sufficient for matching the corresponding pairs. Intuitively, this leads to the 'object bias' where SVLCs like attributes, states, and relations are being underrepresented (e.g. having a much smaller amplitude in the resulting feature superposition), consequently causing the aforementioned issues with SVLC understanding.", + "bbox": [ + 75, + 549, + 468, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this intuition, we propose a simple data-driven technique that harnesses existing language parsing and modeling capabilities to enhance the importance of SVLCs in the VL model training losses. For each text in the training batch, we automatically generate alternative negative or positive text by manipulating its content to be opposite or equivalent to the original text. Using the newly generated texts, we explicitly teach SVLC to the model via additional losses (see Fig. 1) that enforce differentiating be", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tween different (original and generated) SVLC texts and are no longer satisfiable by the 'bag of objects' representation.", + "bbox": [ + 498, + 410, + 890, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards this end, we propose several techniques for implementing this approach, including (i) rule-based priors based on classical NLP parsing and word substitution vocabulary according to attribute/relation type; (ii) prompting a Large Language Model (LLM) (e.g. [53]) for analogous text; (iii) generating different meaning (negative) alternatives by LLM-based unmasking of parsed text entities of different kinds; (iv) combinations of these methods.", + "bbox": [ + 496, + 441, + 892, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate that all these techniques can lead to significant improvements of up to $15\\%$ percent when measuring the VL models' SVLC understanding. We verify this on 5 datasets: VG [39], HAKE [48], VAW [54], SWIG [55], and all combined, using the protocol recently proposed in VL-Checklist [82]. In addition, we show that the resulting VL models largely preserve their zero-shot object recognition performance. For the latter, we also propose a variant of efficient LLM fine-tuning using low-rank residual adapters (LoRA) [27] adjusted to VL models. Finally, we show that our framework allows better harnessing of the standard available VL data, e.g. CC3M [62] and LAION [61]. This is exhibited by the aforementioned gains, both in new VL models trained from scratch, as well as in models fine-tuned using strong available VL models such as CLIP [57] and CyCLIP [19].", + "bbox": [ + 496, + 566, + 892, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we offer the following contributions: (i) We propose a data-driven approach for better harnessing the standard available VL data to improve VL models' SVLC understanding skills, such as understanding object attributes, inter-object relations, transitive actions, object states, and more, without sacrificing zero-shot object recog", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2658", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/0bbf2336bb1f88532cc8e4a37a2f30a75e36171cde93bb058538854057a3550b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelPreArchObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CC3MCLIP [57]✓Vit-B/3281.58%67.60%63.05%56.37%(a)
CLIP + LoRA✓Vit-B/3280.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41% (+0.04%)
CLIP + Ours RB Neg✓Vit-B/3283.89% (+2.30%)73.35% (+5.75%)75.33% (+12.28%)54.32% (-2.05%)
CLIP + Ours LLM Neg✓Vit-B/3284.44% (+2.85%)71.63% (+4.03%)74.82% (+11.77%)55.60% (-0.77%)
CLIP + Ours RB+LLM Negs✓Vit-B/3285.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
CLIP + Ours Combined✓Vit-B/3285.00% (+3.42%)71.97% (+4.37%)68.95% (+5.90%)54.77% (-1.60%)
CLIP [57]✓Vit-B/1682.91%67.32%61.80%60.00%(b)
CLIP + Ours RB+LLM Negs✓Vit-B/1685.82% (+2.91%)73.92% (+6.6%)77.40% (+15.6%)59.37% (-0.63%)
CLIP + Ours Combined✓Vit-B/1684.75% (+1.84%)71.18% (+3.86%)69.68% (+7.88%)59.87% (-0.13%)
CLIPXVit-B/3271.17%57.86%45.20%21.96%(c)
CLIP + Ours CombinedXVit-B/3271.79% (+0.62%)63.29% (+5.43%)58.13% (+12.93%)20.96% (-1.00%)
CLIP [57]XVit-B/1664.01%54.27%41.57%15.49%(d)
CLIP + Ours RB+LLM NegsXVit-B/1673.11% (+9.1%)65.32% (+11.05)71.93% (+30.36)20.78% (+5.29%)
CLIP + Ours CombinedXVit-B/1672.99% (+8.98%)63.01% (+8.74%)62.95% (+21.38)20.61% (+5.12%)
CyCLIP [19]✓R5073.49%59.33%53.83%26.00%(e)
CyCLIP + LoRA✓R5073.30% (-0.19%)58.89% (-0.44%)53.03% (-0.80%)26.30% (+0.30%)
CyCLIP + Ours Combined✓R5074.20% (+0.71%)63.52% (+4.20%)59.47% (+5.63%)26.31% (+0.31%)
CyCLIPXR5069.41%57.59%53.70%21.02%(f)
CyCLIP + Ours CombinedXR5071.50% (+2.09%)65.69% (+8.10%)70.20% (+16.50%)20.44% (-0.42%)
LAIONCLIP [57]✓Vit-B/3281.58%67.6%63.05%56.37%(g)
CLIP + LoRA✓Vit-B/3282.18% (+0.60%)68.48% (+0.88%)62.72% (-0.33%)57.15% (+0.78%)
CLIP + Ours Combined✓Vit-B/3282.54% (+0.96%)69.64% (+2.04%)66.05% (+3.00%)56.71% (+0.34%)
", + "bbox": [ + 76, + 88, + 906, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. VL-Checklist and Zero-Shot classification evaluation (ImageNet + 20 datasets). Finetuned models (for 5 epochs as detailed in Sec. 3.3, starting from officially released CLIP [57] and CyCLIP [19] weights) are marked with $\\checkmark$ in the Pre-trained (Pre) column, while models trained from scratch (for 10 epochs) are marked with $\\times$ . The gains and losses of our approach (+Ours) are in color and are computed w.r.t. to corresponding baselines in each section. CLIP/CyCLIP + LoRA indicate finetuning in the same way and on the same data, but without our approach. Finetuning without LoRA on the same data yields significantly worse performance in all metrics. Sections are separated by bold horizontal lines. (a) CC3M fine-tuning - CLIP - Vit-B/32: we significantly improve the SVLC understanding, observing only small ZS performance drops, $0.77\\%$ ; (b) CC3M fine-tuning - CLIP - Vit-B/16: we can observe similar improvements as in (a), with an even smaller impact on ZS performance; (c) CC3M from scratch - CLIP - Vit-B/32: we significantly improve SVLC understanding with only a small ( $1\\%$ ) decrease in ZS performance; (d) CC3M from scratch - CLIP - Vit-B/16: compared to (c), even greater SVLC understanding improvement is observed (up to $30.36\\%$ ), at no cost in ZS performance, and even improvement of over $5\\%$ ; (e) CC3M fine-tuning - CyCLIP: we use CyCLIP original code (with LoRA integration) and losses, as can be seen - adding our techniques improves CyCLIP SVLC performance considerably without sacrificing ZS performance; (f) CC3M from scratch - CyCLIP: observing even largergains in SVLC understanding compared to (e) (up to $16.5\\%$ ), with small reduction in ZS performance of $0.42\\%$ ; (g) LAION fine-tuning - CLIP - Vit-B/32: we improve SVLC understanding without any decrease in ZS performance.", + "bbox": [ + 75, + 521, + 893, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nition performance; (ii) More specifically, we propose to leverage the well-understood and well-modeled structure of language, through classical NLP parsing and/or use of the modern pre-trained LLMs, for manipulating the text part of the standard VL paired datasets to regularize VL training and teach SVLC understanding to VL models. (iii) We further propose an adaptation of efficient LLM fine-tuning technique of [27] for fine-tuning VL models, allowing for only minimal reduction in zero-shot object recognition performance after fine-tuning, while still obtaining the aforementioned SVLC understanding gains. (iv) Empirically, for", + "bbox": [ + 75, + 729, + 472, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the popular CLIP [57] and its most recent extension CyCLIP [19], we demonstrate SVLC understanding average improvements of up to $13\\%$ when training from scratch, and $15\\%$ when fine-tuning from a pretrained model.", + "bbox": [ + 496, + 729, + 893, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 800, + 640, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language (VL) Models. (e.g., CLIP [57] and ALIGN [31]) show significant advances in diverse zero-shot downstream tasks. They are pre-trained using contrastive image-text alignment on a large-scale noisy dataset of text-image pairs collected from the web. Several methods [9,46,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2659", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[65] additionally employ off-the-shelf object detectors to extract region features. In order to relax this limitation, some methods [31,36,43,76] propose to use cross-attention layers with self-supervised learning objectives including image-text matching and masked/autoregressive language modeling. BLIP [43] generates synthetic captions from the language modeling head and filters noisy captions based on the image-text matching score. Recently, there have been attempts to learn finer-level alignment and relations between image and text [15, 17, 19, 47, 77]. FILIP proposes fine-grained contrastive learning to maximize the token-wise similarity between visual and textual tokens. CyClip [19] imposes additional geometrical consistency on the image and text embeddings. DeCLIP [47] introduces additional positives from the nearest neighbors. While these methods improve image-text retrieval tasks on the existing benchmarks, such as ImageNet [60] and MS-COCO [50], recent studies such as VL-CheckList [83] and the Winoground Challenge [68], show that these models cannot distinguish fine-grained language details or understand structured concepts (SVLCs) such as object attributes and relations. In this paper, we focus on the latter and propose orthogonal data-driven techniques that have the potential to improve the SVLC understanding for all VL models.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning Structured Representations. A full understanding of the semantics of rich visual scenes requires the ability to understand visual concepts, such as detecting individual entities and reasoning about their interactions and attributes. Structured representations have played an important role in achieving this goal, having been successfully applied to a wide range of computer vision applications: vision and language [10, 45, 46, 66], scene graphs [25, 29, 38, 56, 75], relational reasoning [4, 5], human-object interactions [16, 33, 74], action recognition [1, 2, 23, 24, 30, 52, 70], and even image & video generation from graphs [3, 22, 32]. However, most of these works rely on detailed, manually curated, supervision, often involving annotation of location information and structural details, which are very expensive to collect and scale, resulting in limited-size or synthetic data sources for training. In contrast, in our work, we focus on methods for teaching SVLC understanding to large VL models while only leveraging the available large-scale noisy VL data sources collected from the web without any use of expensive manual curation.", + "bbox": [ + 76, + 457, + 472, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Augmentation. Augmentation plays a key role in many computer vision applications [8, 63]. Several advanced image augmentation methods (CutMix [78], mixup [79], AutoAugment [11], RandAugment [12], etc) have been proposed and greatly improved computer vision task performance. Text augmentation has been tackled through back-translation [73], word and frame-semantic embedding augmentations [69], word replacement [81], random word insertion/deletion/swap [71], or using a text gen", + "bbox": [ + 76, + 763, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "erative model [41] in diverse NLP applications. In VL tasks, previous work explores machine translation between different languages [6, 35], generating synthetic captions [43], adversarial/synthetic data augmentation for VQA [59, 67] or mixup for VL [20]. We focus on leveraging the well-understood and modeled language structure for manipulating text in a way that explicitly targets teaching SVLC semantics to VL models. To the best of our knowledge, this has not been attempted before.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 237, + 591, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we discuss the proposed framework for improving SVLC performance of VL models using already available VL data. Sec. 3.1 and Sec. 3.2 present our main techniques for teaching SVLC to VL models. These approaches can be effectively applied both for fine-tuning existing strong VL Pre-trained models, as well as for training VL models from scratch. In both cases, they exhibit significantly improved SVLC performance as demonstrated in our experiments in Sec. 4. Sec. 3.3 presents our strategy for fine-tuning VL models on SVLC-enhanced VL data, while at the same time being parameter efficient and significantly reducing forgetting, thus, maintaining the VL model Zero-Shot (ZS) performance. Qualitative examples showing improvements attained by our proposed approach, as well as some examples of failure cases, are provided in the supplementary material.", + "bbox": [ + 496, + 261, + 893, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Teaching SVLC Using NLP and LLMs", + "text_level": 1, + "bbox": [ + 500, + 508, + 831, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present several ways in which the data of existing VL pre-training paired datasets can be enhanced to emphasize SVLC in the texts, and teaching them to the VL model. We propose two kinds of data enhancements - generating negative and positive text alternatives. When generating negative examples, only one word of the sentence is changed such that the semantic meaning of the sentence changes. We propose two methods for the generation of the negatives: (i) rule-based (Sec. 3.1.1); and (ii) LLM-based (Sec. 3.1.2). Positive alternatives are generated as sentences with semantically similar meanings, but different wording (Sec. 3.1.3). We then present the losses which properly take these two types of generated textual data into account during training in Sec. 3.2.", + "bbox": [ + 496, + 530, + 893, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1 Generating Rule-Based (RB) Negatives", + "text_level": 1, + "bbox": [ + 500, + 758, + 823, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One simple yet effective method for negative text generation is using a collection of pre-defined language rules which match and replace words of a specific entity type or a pattern, such as color, material, size, etc. This method is especially useful when one has prior knowledge of a specific aspect of the language that needs to be taught. For example, if we know that our model lacks the ability to understand the colors of objects, we can easily create a rule for detecting", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2660", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and replacing color words in the text, for generating negative text that does not correspond to its paired image. To employ the generation of the rule-based negative, for each taught SVLC we define a list of words belonging to its characteristic. We then scan the VL data texts searching for the words within these lists. If a word is located, we simply replace it with a randomly selected word from the same list to generate a negative pair. For example, applying the color-rule to a sentence: \"A big brown dog\" can lead to \"A big yellow dog\". If a text has multiple candidates of words to be replaced, one of them is chosen randomly. We perform this process multiple times for the full list of SVLC characteristics of interest such as color, size, material, spatial relations, etc. These generated negative texts are SVLC specific and differ from the original text in only one word. For a detailed description and more examples of the RB negative generation, please refer to the supplementary material.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.2 LLM-based Negative Generation via Unmasking", + "text_level": 1, + "bbox": [ + 76, + 367, + 470, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A natural extension of rule-based negatives technique is the generation of negatives using Large Language Models (LLMs) unmasking. Recent LLMs are explicitly trained in a self-supervised manner with the objective of \"unmasking\" parts of the text. Given a sentence with one missing word, models such as BERT [13], can suggest multiple words that fit the context of the sentence. Using this useful property of LLMs, we can therefore automatically create plausible negative examples without the need for prior knowledge of the SVLC characteristics of interest. In order to focus the randomly selected masked words to be likely to belong to SVLCs of interest, we use common NLP parsing techniques (such as spacy [26]) to parse the sentence into its components such as nouns, verbs, adjectives, adverbs, etc. We then randomly choose a type of sentence part and a word belonging to this part type, mask out the selected word and replace it with one of the options suggested by the LLM's unmasking. These negative examples, when used properly in the loss function (Eq. (3)) focus the network on the important details that affect the SVLC understanding. As we show in Sec. 4, this method is extremely useful and can significantly improve the VL model's understanding of different SVLCs. Further details and examples of the LLM negative generation are provided in the supplementary material.", + "bbox": [ + 75, + 388, + 472, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3 Generating Text Analogies via LLM Prompting", + "text_level": 1, + "bbox": [ + 76, + 771, + 464, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the goal of the negative text generation (Sec. 3.1.1 and Sec. 3.1.2) was to make minor perturbations to a given text such that the meaning changes, the goal here is exactly the opposite. We would like to make major changes to the text, while still keeping the same semantic meaning. For example “A woman standing left to a sitting cat” and “A cat sitting to the right of a standing woman” are two very", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different texts describing the exact same scene. One effective way to generate such semantically similar texts is by prompting the foundational LLMs. Specifically, we use the open access BLOOM [53] model. In the spirit of recently popular in-context learning [18], we present the model with a textual prompt with examples of semantically similar texts (see Fig. 2). We then append the current image caption and retrieve the BLOOMs prediction of a semantically similar text. For a detailed description, we refer the reader to the supplementary material.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Losses", + "text_level": 1, + "bbox": [ + 500, + 251, + 586, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All of our evaluated models (CLIP [57] and CyCLIP [19]) admit a text & image pair $(T, I)$ and are comprised of two parts: (i) image encoder $e_I = \\mathcal{E}_I(I)$ ; (ii) text encoder $e_T = \\mathcal{E}_T(T)$ . In this notation, the text-to-image similarity score is therefore computed as:", + "bbox": [ + 498, + 272, + 893, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (T, I) = \\exp \\left(\\frac {\\tau e _ {T} ^ {T} e _ {I}}{| | e _ {T} | | ^ {2} | | e _ {I} | | ^ {2}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 359, + 890, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a learned temperature parameter.", + "bbox": [ + 498, + 405, + 792, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive Loss. As most contemporary VL models, we employ the contrastive CLIP-loss [57] as one of our losses for each batch $\\mathcal{B}$ .", + "bbox": [ + 496, + 420, + 890, + 464 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t}} = \\sum_ {i} \\log \\left(\\frac {S \\left(T _ {i} , I _ {i}\\right)}{\\sum_ {j} S \\left(T _ {i} , I _ {j}\\right)}\\right) + \\log \\left(\\frac {S \\left(T _ {i} , I _ {i}\\right)}{\\sum_ {k} S \\left(T _ {k} , I _ {i}\\right)}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 474, + 890, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Negatives Loss. In our ablation study in Sec. 5, we show that for a given text $T_{i}$ simply adding the corresponding generated negative text $T_{i}^{neg}$ to the contrastive loss $\\mathcal{L}_{cont}$ is much less effective than having a separate loss individually attending to the similarity difference of $T_{i}$ and $T_{i}^{neg}$ w.r.t. the image $I_{i}$ corresponding to $T_{i}$ . We, therefore, employ the following negatives loss:", + "bbox": [ + 496, + 527, + 893, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n e g} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} , I _ {i}\\right)}{\\mathcal {S} \\left(T _ {i} , I _ {i}\\right) + \\mathcal {S} \\left(T _ {i} ^ {n e g} , I _ {i}\\right)}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 645, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analogy Loss. For the generated analogy texts $T_{i}^{sim}$ produced from the text $T_{i}$ which corresponds to image $I_{i}$ , we employ the combination of the two following losses:", + "bbox": [ + 496, + 696, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i m} ^ {t e x t} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} ^ {s i m} , T _ {i}\\right)}{\\sum_ {j} \\mathcal {S} \\left(T _ {i} ^ {s i m} , T _ {j}\\right)}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 753, + 890, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where with some abuse of notation, $S(T_{1}, T_{2})$ denotes the exponent cosine similarity between text $T_{1}$ and text $T_{2}$ text-embeddings, and:", + "bbox": [ + 496, + 805, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i m} ^ {i m g} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} ^ {s i m} , I _ {i}\\right)}{\\sum_ {j} \\mathcal {S} \\left(T _ {i} ^ {s i m} , I _ {j}\\right)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 863, + 890, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2661", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which simply corresponds to the second summand of $\\mathcal{L}_{cont}$ in Eq. (2), with replacing $T_{i}$ by $T_{i}^{sim}$ .", + "bbox": [ + 76, + 90, + 467, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our final loss is, therefore:", + "bbox": [ + 96, + 121, + 276, + 135 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t}} + \\alpha \\cdot \\mathcal {L} _ {\\text {n e g}} + \\beta \\cdot \\left(\\mathcal {L} _ {\\text {s i m}} ^ {\\text {t e x t}} + \\mathcal {L} _ {\\text {s i m}} ^ {\\text {i m g}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 146, + 468, + 164 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Fine-tuning Pre-trained VL Models", + "text_level": 1, + "bbox": [ + 76, + 180, + 387, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Each of the $\\mathcal{E}_T$ and $\\mathcal{E}_I$ networks is comprised of a mix of non-parametric functions and two types of parametric functions: linear layers and embedding layers. Roughly, each of those functions, $\\mathcal{F}_k^{lin}(x)$ and $\\mathcal{F}_k^{emb}(x)$ where $k$ is layer index, is parameterized by a weight matrix $\\mathcal{W}_k$ so that:", + "bbox": [ + 76, + 202, + 468, + 277 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {k} ^ {\\text {l i n}} (x) = \\mathcal {W} _ {k} \\cdot x \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 287, + 468, + 305 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {k} ^ {e m b} (x) = E M B (x; \\mathcal {W} _ {k}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 308, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $EMB$ is the embedding operator assuming $x$ is a stream of integers and picking the respective columns of $\\mathcal{W}_k$ . Following the idea proposed in LoRA [27] for efficient LLM fine-tuning using low-rank residual adapters, when adapting a pre-trained VL model $\\mathcal{M} = (\\mathcal{E}_T, \\mathcal{E}_I)$ we parameterize the adapted weights $\\mathcal{W}_k^*$ of the model $\\mathcal{M}^*$ fine-tuned from $\\mathcal{M}$ as:", + "bbox": [ + 76, + 335, + 468, + 439 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {k} ^ {*} = \\mathcal {W} _ {k} + \\mathcal {A} _ {k} \\cdot \\mathcal {B} _ {k} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 441, + 468, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where for $\\mathcal{W}_k$ of size $m\\times l$ , $\\mathcal{A}_k$ and $\\mathcal{B}_k$ are rank- $r$ matrices of sizes $m\\times r$ and $r\\times l$ respectively. These low-rank residual adapters can be applied efficiently as:", + "bbox": [ + 76, + 463, + 468, + 508 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {k} ^ {*, l i n} (x) = \\mathcal {F} _ {k} ^ {l i n} (x) + \\mathcal {A} _ {k} \\cdot (\\mathcal {B} _ {k} \\cdot x) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 517, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {k} ^ {*}, e m b (x) = \\mathcal {F} _ {k} ^ {e m b} (x) + \\mathcal {A} _ {k} \\cdot E M B (x; \\mathcal {B} _ {k}) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 539, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the fine-tuning of $\\mathcal{M}^*$ , we freeze all the base model $\\mathcal{M}$ parameters $\\forall k, \\{\\mathcal{W}_k\\}$ and only the LoRA adapters $\\forall k, \\{(\\mathcal{A}_k, \\mathcal{B}_k)\\}$ are being learned. In the above notation we disregard possible bias terms of the linear functions, if they are present, since we keep them frozen too.", + "bbox": [ + 76, + 568, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are several interesting things to note about the proposed architecture: (i) as opposed to [64], who evaluated the use of efficient LLM fine-tuning techniques for VL models adaptation, we add our LoRA adapters everywhere, i.e. to all layers of both the text and image encoders, and not only to the text encoder/decoder as done in [64]; (ii) as opposed to [80] who attach a small side network only to the output of the adapted model, our LoRA adapters are added to all the parametric functions inside the model and affect all the intermediate computations; (iii) same as noted in the original [27] paper, at inference all the LoRA adapters can be folded back into the original weight matrices by simple summation, thus returning the number of total parameters to be the same as in the original model and hence have the same inference speed; (iv) with rank $r$ kept low, the number of extra parameters added by all the LoRA adapters can be very low making adaptation fast and efficient; finally,", + "bbox": [ + 75, + 643, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(v) such form of fine low-rank adaptation allows for significantly mitigating the zero-shot performance forgetting effects as demonstrated in our results and explored in the corresponding ablation (Sec. 5).", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 188, + 630, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 220, + 601, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the model using common Image-Text pair datasets, namely Conceptual Captions 3M [62] and LAION 400M [61] and test using the VL-Checklist [82] datasets, which will be described below.", + "bbox": [ + 498, + 250, + 890, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conceptual Captions 3M (CC3M) [62] is a dataset of three million image-text pairs automatically crawled from the internet where image descriptions are harvested from Alt-text attributes and then processed and filtered to create relatively clean descriptions.", + "bbox": [ + 498, + 310, + 890, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LAION 400M [61] is a very large scale image-text pair dataset which, similarly to CC3M has been automatically harvested from the internet. One major difference between the two, apart from the size, is that LAION examples have been filtered using the pretrained CLIP model, such that the CLIP image-text similarity is high by design. Recent re-implementations of the original CLIP paper have successfully used LAION 400M to reproduce similar capabilities as the original CLIP model.", + "bbox": [ + 498, + 386, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VL-Checklist [82] combines images and annotations from the Visual Genome [39], SWiG [55], VAW [54], and HAKE [48] datasets. It is processed such that each image is annotated with two captions, one positive and one negative. The positive caption originates in its source dataset and corresponds to the image. The negative caption is constructed from the positive caption so that only one word, corresponding to the tested SVLC of interest, is changed to negate the SVLC (e.g., color, size, material, etc.). VL-Checklist evaluates three main types of VL concepts further subdivided into 7 types of SVLCs total: (1) Object: its spatial location and size, (2) Attribute: color, material, size, state, and action, and (3) Relation: spatial or action relation between two objects. In the following sections we report results on a combined VL-Checklist dataset. The results on individual comprising datasets (Visual Genome [39], SWiG [55], VAW [54], and HAKE [48]) are provided in the supplementary material.", + "bbox": [ + 498, + 522, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-Shot Classification We evaluated our method on 21 classification dataset using the Zero-Shot classification protocol described in the ELEVATER Image Classification Toolkit [42]. The evaluation includes 21 different datasets, including common classification datasets such as ImageNet [60], CIFAR100 [40], EuroSat [21], and others. In Tables 1-3 we report the average results over the 21 tasks.", + "bbox": [ + 498, + 794, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2662", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c72db18680a7699ba4fef32b47fe0087d4293b283810155ba3875ea38fe92ca0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CLIP [57]81.58%67.60%63.05%56.37%
CLIP +LoRA80.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41%(+0.04%)
w/o Neg Loss82.27% (+0.69%)67.58% (-0.02%)55.17% (-7.88%)55.37% (-1.00%)
Ours RB+LLM Negs85.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
", + "bbox": [ + 151, + 88, + 820, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study - separate Negative Losses (Sec. 3.2) vs adding negatives to contrastive loss.", + "bbox": [ + 173, + 204, + 795, + 219 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 232, + 294, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For CLIP we use the ML-Foundation Open-CLIP repository [28] and for CyCLIP [19] we use its original code repository, which is also based on Open-CLIP. In most experiments, unless stated otherwise, we train our model for five epochs on 4 V100 NVIDIA GPUs, with a total batch size of 512. When starting from a pre-trained model, we use rank 4 LoRA adapters (Sec. 3.3) and the learning rate is set to $5E - 6$ . When training from scratch, for CLIP we use the default parameters set in the open-CLIP library, and for CyCLIP the defaults of its original implementation. For all CLIP experiments, we use VIT/32-B as the model architecture and ResNet-50 when training CyCLIP (following [19]). When fine-tuning we initialize with the original model weights released by the respective authors. In all experiments involving a combination of CyCLIP [19] and our method (in Sec. 4.4-4.5), in addition to Eq. (6) loss we also employ all the extra losses proposed in CyCLIP [19].", + "bbox": [ + 75, + 253, + 470, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Baselines", + "text_level": 1, + "bbox": [ + 76, + 516, + 184, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method to two strong baselines under several configurations. The first is the CLIP [57] OpenAI pretrained model trained on 400M image-text pairs which achieves high ZS performance. The second is the very recent CyCLIP [19] method which, similarly to us, improves over the original CLIP loss. For a fair comparison all methods use the same network architecture and the same initialization from the OpenAI pretrained model. For the CyCLIP baseline we continue training from the pretrained initialization using LoRA and the CyCLIP losses. As CyCLIP losses are orthogonal to ours, we also show a unified version of the two methods (CyCLIP + Ours Combined).", + "bbox": [ + 75, + 537, + 468, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Fine-tuning VL Models", + "text_level": 1, + "bbox": [ + 76, + 726, + 294, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following experiments we show the effects of fin-tuning a pre-trained VL model using our additional data enhancement methods and losses. All experiments are initialized from the official OpenAI CLIP. We compare our method to two baselines on the VL-Checklist tasks, and on the Zero-Shot image classification task. The first baseline is the original pretrained model without any further training. The second, is the same model with the additional LoRA parameters, trained on CC3M (Table 1a) and LAION (Table 1g) using the original CLIP loss function. Table 1a", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6d2f5a157ac665955aa3d39ef8f23d02b2101b355f705b7fd77ac3772737f17.jpg", + "image_caption": [ + "Figure 3. CC3M fine-tuning - detailed. Detailed results of the baselines (CLIP, CLIP+LoRA) and our models (RB+LLM Neg, RB+LLM Negs+Pos) initialized from OpenAI pretrained CLIP ViT-32B and fine-tuned (except CLIP) on CC3M." + ], + "image_footnote": [], + "bbox": [ + 555, + 229, + 836, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows several configurations of our method compared to the baselines. We see that our method shows significant improvements on all VL-Checklist tasks reaching up to $15\\%$ improvement. Figure 3 displays the relative gains on all tasks of the \"Attribute\" and \"Relation\" tests. It is clear that our gains are across all tests. These improvements come at a price of some minor degradation to the Zero-Shot performance compared to CLIP. Moreover, when fine-tuning CLIP on the LAION dataset (Table 1g) we do not see these degradations. In Tab. 1b we show consistent gains to ones in Tab. 1a when finetuning with a stronger (higher ZS) VitB/16 CLIP [57] pre-trained image encoder, observing an even lower drop in ZS performance. In Tab. 1e we show significant gains in fine-tuning CyCLIP on CC3M, comparing fine-tuning before and after integrating our proposed approach, this also comes at no ZS performance cost.", + "bbox": [ + 496, + 513, + 892, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Training from scratch", + "text_level": 1, + "bbox": [ + 500, + 768, + 705, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In these experiments, we show the advantage of integrating our approach throughout the whole training procedure. To this end, we train from scratch on CC3M - both CLIP and CyCLIP, either on their own, or together with our proposed method integrated. Tables - Tab. 1c, Tab. 1d, and Tab. 1f show that, similar to finetining, our method significantly improves both CLIP and CyCLIP SVLC capabilities.", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2663", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/036d6a0d76816cc7a750759277fceb719efcd1bf111741bc168dcf2656d4aef5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodColorMaterialAttribute SizeStateActionRelation21 Zero-Shot Tasks Average
ActionSpatial
Ours Pos72.35%69.25%69.80%59.35%66.08%70.97%39.20%55.37%
Ours RB Neg78.45%83.20%69.50%65.95%69.66%75.97%74.70%54.66%
Ours LLM Neg76.00%79.70%72.75%61.35%68.36%77.23%72.40%55.05%
Ours RB+LLM Negs79.25%84.25%72.15%64.05%69.82%79.03%78.40%54.66%%
Ours Combined77.45%77.35%73.35%62.30%69.39%74.70%63.20%54.77%%
", + "bbox": [ + 107, + 80, + 864, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study - component analysis. Detailed evaluation on Attribute & Relation SVLCs.", + "bbox": [ + 184, + 205, + 782, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ties, while still keeping similar ZS performance. Figure 4 shows the detailed parsing of the \"Attribute\" and \"Relation\" tests with consistent gains across the specific tasks. These findings suggest that had we trained our method for many epochs on a large dataset, such as the full LAION-9B, we would reach similar ZS performance while greatly enhancing the SVLC performance of the model.", + "bbox": [ + 75, + 233, + 470, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablations", + "text_level": 1, + "bbox": [ + 76, + 352, + 184, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section we will examine several aspects and components of our proposed methods.", + "bbox": [ + 76, + 375, + 468, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Negative Losses", + "text_level": 1, + "bbox": [ + 76, + 414, + 235, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Section 3.2 details our additional loss functions which utilize our generated texts. Specifically, Eq. (3) describes the loss which contrasts a given example only with its generated negatives. Table 2 clearly shows the importance of this loss as opposed to simply adding the negative examples to the original contrastive loss (Eq. (2)). Without explicitly forcing the network to attend to the small changes in the text (Eq. (3)), the generated negative examples, when inserted only to the contrastive loss (Eq. (2)) do not provide the desired gains over the baseline.", + "bbox": [ + 75, + 435, + 470, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Component analysis", + "text_level": 1, + "bbox": [ + 76, + 595, + 269, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Section 3 we describe several components of our proposed method. Specifically we present our RB negative generation (Sec. 3.1.1), our LLM-based negative genera", + "bbox": [ + 75, + 616, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de41b8340c047d239b087b68f831ecbe3f93a12b553ad11bbe5d23118219e5ea.jpg", + "image_caption": [ + "Figure 4. CC3M train from scratch - detailed. Detailed results of the baseline CLIP and our models (RB+LLM Neg, RB+LLM Negs+Pos), all CLIP ViT-32B, trained from scratch on CC3M." + ], + "image_footnote": [], + "bbox": [ + 151, + 681, + 395, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion (Sec. 3.1.2), and our LLM-based analogy generation, referred to here as \"Pos\" (Sec. 3.1.3). Tab. 3 provides a detailed analysis and comparison of the contribution of each component to the final result. Through this analysis we see two contradicting forces between the SVLC capabilities and the original Zero-Shot performance. We can see that each negative generation method plays a crucial role in some of the tasks while the use of both is usually the best performing option. On the other hand, the LLM-based analogy generation stabilizes the Zero-Shot performance and mitigates the drop with respect to the baseline. The joint version of all components (\"Ours Combined\") exhibits a good trade-off between the two contradicting forces.", + "bbox": [ + 496, + 233, + 893, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 441, + 627, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a data-driven technique for enhancing the performance of VL models in the important task of SVLC understanding without sacrificing their impressive ZS object recognition capabilities. Our proposed method attains significant gains in multiple experiments on a variety of base VL models and datasets. It builds upon the modeling strength and knowledge of language structure to teach this structure to VL models in an orthogonal way, suggesting wide applicability to existing or future VL models.", + "bbox": [ + 496, + 465, + 890, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While attaining impressive gains in SVLC understanding, we believe the small drop observed in ZS performance could be further reduced. An additional possible extension of our work is using more sophisticated sequence generation techniques for improving our batch data. One may combine annotation efforts with a language model to get improved data for training and evaluation [51]. Another possibility is adding a corrector [72] in the training that validates whether the VL model learns the correct concepts or not. We leave all of these exciting directions to future work.", + "bbox": [ + 496, + 601, + 890, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 766, + 668, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would like to thank Donghyun Kim for his invaluable help in this work. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. FA8750-19-C-1001. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA). This material is based upon work supported by the 'Data Science grant from the Israeli Council of Higher Education'. This research or RG was supported by ERC-StG SPADE grant no. 75749.", + "bbox": [ + 496, + 789, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2664", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 4", + "[2] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 4", + "[3] Amir Bar, Roei Herzig, Xiaolong Wang, Anna Rohrbach, Gal Chechik, Trevor Darrell, and A. Globerson. Compositional video synthesis with action graphs. In ICML, 2021. 4", + "[4] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, pages 105-121, 2018. 4", + "[5] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.4", + "[6] Andrea Burns, Donghyun Kim, Derry Wijaya, Kate Saenko, and Bryan A Plummer. Learning to scale multilingual representations for vision-language tasks. In European Conference on Computer Vision, pages 197-213. Springer, 2020. 4", + "[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 1", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4", + "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 3", + "[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 4", + "[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 4", + "[12] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 4", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[14] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't stop learning: Towards continual learning for the clip model. arXiv:2207.09248, 2022. 1", + "[15] Andreas Fürst, Elisabeth Rumetshofer, Viet Tran, Hubert Ramsauer, Fei Tang, Johannes Lehner, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto-Nemling, et al. Cloob: Modern hopfield networks with infoloob outperform clip. arXiv preprint arXiv:2110.11316, 2021. 4", + "[16] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. ArXiv, abs/2008.11714, 2020. 4", + "[17] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022. 4", + "[18] Shivam Garg, Dimitris Tsipras, Percy Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes, 2022. 5", + "[19] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022. 1, 2, 3, 4, 5, 7", + "[20] Xiaoshuai Hao, Yi Zhu, Srikar Appalaraju, Aston Zhang, Wanqian Zhang, Bo Li, and Mu Li. Mixgen: A new multimodal data augmentation. arXiv preprint arXiv:2206.08358, 2022.4", + "[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6", + "[22] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, 2020. 4", + "[23] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4", + "[24] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0–0, 2019. 4", + "[25] Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, and Amir Globerson. Mapping images to scene graphs with permutation-invariant structured prediction. In Advances in Neural Information Processing Systems (NIPS), 2018. 4", + "[26] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 5", + "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2665", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.2,3,6", + "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 7", + "[29] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 4", + "[30] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as composition of spatiotemporal scene graphs. arXiv preprint arXiv:1912.06992, 2019.4", + "[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 3, 4", + "[32] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 4", + "[33] Keizo Kato, Yin Li, and Abhinav Gupta. Compositional learning for human object interaction. In ECCV, 2018. 4", + "[34] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 1", + "[35] Donghyun Kim, Kuniaki Saito, Kate Saenko, Stan Sclaroff, and Bryan Plummer. Mule: Multimodal universal language embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11254-11261, 2020. 4", + "[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 4", + "[37] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 1", + "[38] Ranjay Krishna, Ines Chami, Michael S. Bernstein, and Li Fei-Fei. Referring relationships. *ECCV*, 2018. 4", + "[39] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2, 6", + "[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 4", + "[42] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. Neural Information Processing Systems, 2022. 6", + "[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 1, 4", + "[44] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1", + "[45] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. ArXiv, abs/1908.03557, 2019. 4", + "[46] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 3, 4", + "[47] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1, 4", + "[48] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Mingyang Chen, Ze Ma, Shiyi Wang, Hao-Shu Fang, and Cewu Lu. Hake: Human activity knowledge engine. arXiv preprint arXiv:1904.06539, 2019. 2, 6", + "[49] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1", + "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4", + "[51] Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. Wanli: Worker and ai collaboration for natural language inference dataset creation, Jan. 2022. 8", + "[52] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4", + "[53] Margaret Mitchell, Giada Pistilli, Yacine Jernite, Ezinwanne Ozoani, Marissa Gerchick, Nazneen Rajani, Sasha Luccioni, Irene Solaiman, Maraim Masoud, Somaieh Nikpoor, Munoz Ferrandis Carlos, Stas Bekman, Christopher Akiki, Danish Contractor, David Lansky, Angelina McMillan-Major, Tristan Thrush, Suzana Ilic, Gerard Dupont, Shayne Longpre," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2666", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Manan Dey, Stella Biderman, Douwe Kiela, Emi Baylor, Teven Le Scao, Aaron Gokaslan, Julien Launay, and Niklas Muennighoff. Bigscience, bigscience language open-science open-access multilingual (bloom) language model. International, May 2021-May 2022. 2, 5", + "[54] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13018-13028, 2021. 2, 6", + "[55] Sarah Pratt, Mark Yatskar, Luca Weihs, Ali Farhadi, and Aniruddha Kembhavi. Grounded situation recognition. In European Conference on Computer Vision, pages 314-332. Springer, 2020. 2, 6", + "[56] Moshiko Raboh, Roei Herzig, Gal Chechik, Jonathan Berant, and Amir Globerson. Differentiable scene graphs. In WACV, 2020. 4", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 7", + "[58] Vinay Venkatesh Ramasesh, Ethan Dyer, and Maithra Raghu. Anatomy of catastrophic forgetting: Hidden representations and task semantics. In International Conference on Learning Representations, 2021. 1", + "[59] Arijit Ray, Karan Sikka, Ajay Divakaran, Stefan Lee, and Giedrius Burachas. Sunny and dark outside?! improving answer consistency in vqa through entailed question generation. arXiv preprint arXiv:1909.04696, 2019. 4", + "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 6", + "[61] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 6", + "[62] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2, 6", + "[63] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020.4", + "[64] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3", + "[66] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 4", + "[67] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In European Conference on Computer Vision, pages 437-453. Springer, 2020. 4", + "[68] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2, 4", + "[69] William Yang Wang and Diyi Yang. That's so annoying!!! A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using# petpeeve tweets. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 2557–2563, 2015. 4", + "[70] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 4", + "[71] Jason Wei and Kai Zou. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196, 2019. 4", + "[72] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generating sequences by learning to self-correct, 2022. 8", + "[73] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 4", + "[74] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and M. Kankanhalli. Learning to detect human-object interactions with knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2019. 4", + "[75] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene Graph Generation by Iterative Message Passing. In CVPR, pages 3097–3106, 2017. 4", + "[76] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 4", + "[77] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 4", + "[78] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2667", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4", + "[80] Jeffrey O Zhang, Alexander Sax, Amir Zamir, Leonidas Guibas, and Jitendra Malik. Side-tuning: a baseline for network adaptation via additive side networks. In European Conference on Computer Vision, pages 698-714. Springer, 2020. 6", + "[81] Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015. 4", + "[82] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. V1-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6", + "[83] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022.4" + ], + "bbox": [ + 78, + 90, + 467, + 386 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2668", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_model.json b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3093610708a2ff9a1325016922dd2d01133dd480 --- /dev/null +++ b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_model.json @@ -0,0 +1,2391 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.131, + 0.892, + 0.155 + ], + "angle": 0, + "content": "Teaching Structured Vision & Language Concepts to Vision & Language Models" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.18, + 0.842, + 0.22 + ], + "angle": 0, + "content": "Sivan Doveh\\(^{1,2}\\), Assaf Arbelle\\(^{1}\\), Sivan Harary\\(^{1}\\), Eli Schwartz\\(^{1,3}\\), Roei Herzig\\(^{1,3}\\), Raja Giryes\\(^{3}\\), Rogerio Feris\\(^{4}\\), Rameswar Panda\\(^{4}\\), Shimon Ullman\\(^{*2}\\), Leonid Karlinsky\\(^{*4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.236, + 0.885, + 0.252 + ], + "angle": 0, + "content": "1IBM Research, 2Weizmann Institute of Science, 3Tel-Aviv University, 4MIT-IBM Watson AI Lab" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to \\(15\\%\\) in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.751, + 0.21, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Recent Vision & Language (VL) models [19, 31, 43, 44, 47, 57] achieve excellent zero-shot performance with respect to various computer-vision tasks such as detection, classification, segmentation, etc. However, recent studies [68, 82] have demonstrated that even the strongest VL models struggle with the compositional understanding of some basic Structured VL Concepts (SVLC) such as ob-" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.285, + 0.831, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.571, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Figure 1. Teaching language structure to VL models. (a) Standard contrastive text-to-image loss (e.g. CLIP [57]) tends to under-emphasize SVLC content of the text, likely due to the random nature of the training batches; (b) We generate modified versions of corresponding texts and use them to add losses to explicitly teach language structure (SVLC) to VL models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.764 + ], + "angle": 0, + "content": "ject attributes, inter-object relations, transitive actions, object states and more. Collecting specialized large scale data to teach VL models these missing 'skills' is impractical, as finding specialized text-image pairs for each kind and possible value of the different attributes, relations, or states, is both difficult and expensive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Another important challenge in training VL models with new concepts is catastrophic forgetting, which is a common property to all neural models [7, 34, 37, 49, 58] and has been explored for VL models in a recent concurrent work [14]. Large VL models such as CLIP [57] and CyCLIP [19] have exhibited excellent zero-shot learning abilities in many tasks. Therefore, even given a large dataset with new concepts, it is important not to lose these abilities when performing the adaptation to the new data." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.204, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2657" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.891, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.895, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Teaching structured image understanding to VL models via structured textual data manipulation harnessing the power of language modeling. (1) Generating Rule-Based negative texts (Sec. 3.1.1); (2) Generating negatives using Large Language Model (LLM) unmasking (Sec. 3.1.2); (3) Generating analogies (positives) via LLM prompting (Sec. 3.1.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.471, + 0.546 + ], + "angle": 0, + "content": "In this paper, we propose a way to leverage existing (off-the-shelf) VL pre-training data sources in order to improve the SVLC understanding skills of a given model, while at the same time maintaining its zero-shot object recognition accuracy. Naturally, succeeding in this goal would lead to potential improvement w.r.t. SVLC understanding in a wide variety of downstream tasks building upon pre-trained VL models, such as zeros-shot detection, segmentation, image generation, and many more." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Recent research [68, 82] has shown that VL models exhibit an 'object bias' partially due to the contrastive text-to-image loss used in their pre-training. For example, the popular CLIP-loss [57] is computed over a random batch of text-image pairs sampled from a large-scale and diverse VL dataset with the chance of two images in the same batch containing the same set of objects being very low. For such a loss, representing just a 'bag of objects' in each image or text is sufficient for matching the corresponding pairs. Intuitively, this leads to the 'object bias' where SVLCs like attributes, states, and relations are being underrepresented (e.g. having a much smaller amplitude in the resulting feature superposition), consequently causing the aforementioned issues with SVLC understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Based on this intuition, we propose a simple data-driven technique that harnesses existing language parsing and modeling capabilities to enhance the importance of SVLCs in the VL model training losses. For each text in the training batch, we automatically generate alternative negative or positive text by manipulating its content to be opposite or equivalent to the original text. Using the newly generated texts, we explicitly teach SVLC to the model via additional losses (see Fig. 1) that enforce differentiating be" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.441 + ], + "angle": 0, + "content": "tween different (original and generated) SVLC texts and are no longer satisfiable by the 'bag of objects' representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.893, + 0.564 + ], + "angle": 0, + "content": "Towards this end, we propose several techniques for implementing this approach, including (i) rule-based priors based on classical NLP parsing and word substitution vocabulary according to attribute/relation type; (ii) prompting a Large Language Model (LLM) (e.g. [53]) for analogous text; (iii) generating different meaning (negative) alternatives by LLM-based unmasking of parsed text entities of different kinds; (iv) combinations of these methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.809 + ], + "angle": 0, + "content": "We demonstrate that all these techniques can lead to significant improvements of up to \\(15\\%\\) percent when measuring the VL models' SVLC understanding. We verify this on 5 datasets: VG [39], HAKE [48], VAW [54], SWIG [55], and all combined, using the protocol recently proposed in VL-Checklist [82]. In addition, we show that the resulting VL models largely preserve their zero-shot object recognition performance. For the latter, we also propose a variant of efficient LLM fine-tuning using low-rank residual adapters (LoRA) [27] adjusted to VL models. Finally, we show that our framework allows better harnessing of the standard available VL data, e.g. CC3M [62] and LAION [61]. This is exhibited by the aforementioned gains, both in new VL models trained from scratch, as well as in models fine-tuned using strong available VL models such as CLIP [57] and CyCLIP [19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "To summarize, we offer the following contributions: (i) We propose a data-driven approach for better harnessing the standard available VL data to improve VL models' SVLC understanding skills, such as understanding object attributes, inter-object relations, transitive actions, object states, and more, without sacrificing zero-shot object recog" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2658" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.907, + 0.511 + ], + "angle": 0, + "content": "
DatasetModelPreArchObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CC3MCLIP [57]✓Vit-B/3281.58%67.60%63.05%56.37%(a)
CLIP + LoRA✓Vit-B/3280.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41% (+0.04%)
CLIP + Ours RB Neg✓Vit-B/3283.89% (+2.30%)73.35% (+5.75%)75.33% (+12.28%)54.32% (-2.05%)
CLIP + Ours LLM Neg✓Vit-B/3284.44% (+2.85%)71.63% (+4.03%)74.82% (+11.77%)55.60% (-0.77%)
CLIP + Ours RB+LLM Negs✓Vit-B/3285.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
CLIP + Ours Combined✓Vit-B/3285.00% (+3.42%)71.97% (+4.37%)68.95% (+5.90%)54.77% (-1.60%)
CLIP [57]✓Vit-B/1682.91%67.32%61.80%60.00%(b)
CLIP + Ours RB+LLM Negs✓Vit-B/1685.82% (+2.91%)73.92% (+6.6%)77.40% (+15.6%)59.37% (-0.63%)
CLIP + Ours Combined✓Vit-B/1684.75% (+1.84%)71.18% (+3.86%)69.68% (+7.88%)59.87% (-0.13%)
CLIPXVit-B/3271.17%57.86%45.20%21.96%(c)
CLIP + Ours CombinedXVit-B/3271.79% (+0.62%)63.29% (+5.43%)58.13% (+12.93%)20.96% (-1.00%)
CLIP [57]XVit-B/1664.01%54.27%41.57%15.49%(d)
CLIP + Ours RB+LLM NegsXVit-B/1673.11% (+9.1%)65.32% (+11.05)71.93% (+30.36)20.78% (+5.29%)
CLIP + Ours CombinedXVit-B/1672.99% (+8.98%)63.01% (+8.74%)62.95% (+21.38)20.61% (+5.12%)
CyCLIP [19]✓R5073.49%59.33%53.83%26.00%(e)
CyCLIP + LoRA✓R5073.30% (-0.19%)58.89% (-0.44%)53.03% (-0.80%)26.30% (+0.30%)
CyCLIP + Ours Combined✓R5074.20% (+0.71%)63.52% (+4.20%)59.47% (+5.63%)26.31% (+0.31%)
CyCLIPXR5069.41%57.59%53.70%21.02%(f)
CyCLIP + Ours CombinedXR5071.50% (+2.09%)65.69% (+8.10%)70.20% (+16.50%)20.44% (-0.42%)
LAIONCLIP [57]✓Vit-B/3281.58%67.6%63.05%56.37%(g)
CLIP + LoRA✓Vit-B/3282.18% (+0.60%)68.48% (+0.88%)62.72% (-0.33%)57.15% (+0.78%)
CLIP + Ours Combined✓Vit-B/3282.54% (+0.96%)69.64% (+2.04%)66.05% (+3.00%)56.71% (+0.34%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.522, + 0.895, + 0.718 + ], + "angle": 0, + "content": "Table 1. VL-Checklist and Zero-Shot classification evaluation (ImageNet + 20 datasets). Finetuned models (for 5 epochs as detailed in Sec. 3.3, starting from officially released CLIP [57] and CyCLIP [19] weights) are marked with \\(\\checkmark\\) in the Pre-trained (Pre) column, while models trained from scratch (for 10 epochs) are marked with \\(\\times\\). The gains and losses of our approach (+Ours) are in color and are computed w.r.t. to corresponding baselines in each section. CLIP/CyCLIP + LoRA indicate finetuning in the same way and on the same data, but without our approach. Finetuning without LoRA on the same data yields significantly worse performance in all metrics. Sections are separated by bold horizontal lines. (a) CC3M fine-tuning - CLIP - Vit-B/32: we significantly improve the SVLC understanding, observing only small ZS performance drops, \\(0.77\\%\\); (b) CC3M fine-tuning - CLIP - Vit-B/16: we can observe similar improvements as in (a), with an even smaller impact on ZS performance; (c) CC3M from scratch - CLIP - Vit-B/32: we significantly improve SVLC understanding with only a small (\\(1\\%\\)) decrease in ZS performance; (d) CC3M from scratch - CLIP - Vit-B/16: compared to (c), even greater SVLC understanding improvement is observed (up to \\(30.36\\%\\)), at no cost in ZS performance, and even improvement of over \\(5\\%\\); (e) CC3M fine-tuning - CyCLIP: we use CyCLIP original code (with LoRA integration) and losses, as can be seen - adding our techniques improves CyCLIP SVLC performance considerably without sacrificing ZS performance; (f) CC3M from scratch - CyCLIP: observing even largergains in SVLC understanding compared to (e) (up to \\(16.5\\%\\)), with small reduction in ZS performance of \\(0.42\\%\\); (g) LAION fine-tuning - CLIP - Vit-B/32: we improve SVLC understanding without any decrease in ZS performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.473, + 0.898 + ], + "angle": 0, + "content": "nition performance; (ii) More specifically, we propose to leverage the well-understood and well-modeled structure of language, through classical NLP parsing and/or use of the modern pre-trained LLMs, for manipulating the text part of the standard VL paired datasets to regularize VL training and teach SVLC understanding to VL models. (iii) We further propose an adaptation of efficient LLM fine-tuning technique of [27] for fine-tuning VL models, allowing for only minimal reduction in zero-shot object recognition performance after fine-tuning, while still obtaining the aforementioned SVLC understanding gains. (iv) Empirically, for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.895, + 0.793 + ], + "angle": 0, + "content": "the popular CLIP [57] and its most recent extension CyCLIP [19], we demonstrate SVLC understanding average improvements of up to \\(13\\%\\) when training from scratch, and \\(15\\%\\) when fine-tuning from a pretrained model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.642, + 0.817 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Vision-language (VL) Models. (e.g., CLIP [57] and ALIGN [31]) show significant advances in diverse zero-shot downstream tasks. They are pre-trained using contrastive image-text alignment on a large-scale noisy dataset of text-image pairs collected from the web. Several methods [9,46," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2659" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "[65] additionally employ off-the-shelf object detectors to extract region features. In order to relax this limitation, some methods [31,36,43,76] propose to use cross-attention layers with self-supervised learning objectives including image-text matching and masked/autoregressive language modeling. BLIP [43] generates synthetic captions from the language modeling head and filters noisy captions based on the image-text matching score. Recently, there have been attempts to learn finer-level alignment and relations between image and text [15, 17, 19, 47, 77]. FILIP proposes fine-grained contrastive learning to maximize the token-wise similarity between visual and textual tokens. CyClip [19] imposes additional geometrical consistency on the image and text embeddings. DeCLIP [47] introduces additional positives from the nearest neighbors. While these methods improve image-text retrieval tasks on the existing benchmarks, such as ImageNet [60] and MS-COCO [50], recent studies such as VL-CheckList [83] and the Winoground Challenge [68], show that these models cannot distinguish fine-grained language details or understand structured concepts (SVLCs) such as object attributes and relations. In this paper, we focus on the latter and propose orthogonal data-driven techniques that have the potential to improve the SVLC understanding for all VL models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.473, + 0.761 + ], + "angle": 0, + "content": "Learning Structured Representations. A full understanding of the semantics of rich visual scenes requires the ability to understand visual concepts, such as detecting individual entities and reasoning about their interactions and attributes. Structured representations have played an important role in achieving this goal, having been successfully applied to a wide range of computer vision applications: vision and language [10, 45, 46, 66], scene graphs [25, 29, 38, 56, 75], relational reasoning [4, 5], human-object interactions [16, 33, 74], action recognition [1, 2, 23, 24, 30, 52, 70], and even image & video generation from graphs [3, 22, 32]. However, most of these works rely on detailed, manually curated, supervision, often involving annotation of location information and structural details, which are very expensive to collect and scale, resulting in limited-size or synthetic data sources for training. In contrast, in our work, we focus on methods for teaching SVLC understanding to large VL models while only leveraging the available large-scale noisy VL data sources collected from the web without any use of expensive manual curation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Data Augmentation. Augmentation plays a key role in many computer vision applications [8, 63]. Several advanced image augmentation methods (CutMix [78], mixup [79], AutoAugment [11], RandAugment [12], etc) have been proposed and greatly improved computer vision task performance. Text augmentation has been tackled through back-translation [73], word and frame-semantic embedding augmentations [69], word replacement [81], random word insertion/deletion/swap [71], or using a text gen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "erative model [41] in diverse NLP applications. In VL tasks, previous work explores machine translation between different languages [6, 35], generating synthetic captions [43], adversarial/synthetic data augmentation for VQA [59, 67] or mixup for VL [20]. We focus on leveraging the well-understood and modeled language structure for manipulating text in a way that explicitly targets teaching SVLC semantics to VL models. To the best of our knowledge, this has not been attempted before." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.238, + 0.593, + 0.254 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.895, + 0.504 + ], + "angle": 0, + "content": "In this section, we discuss the proposed framework for improving SVLC performance of VL models using already available VL data. Sec. 3.1 and Sec. 3.2 present our main techniques for teaching SVLC to VL models. These approaches can be effectively applied both for fine-tuning existing strong VL Pre-trained models, as well as for training VL models from scratch. In both cases, they exhibit significantly improved SVLC performance as demonstrated in our experiments in Sec. 4. Sec. 3.3 presents our strategy for fine-tuning VL models on SVLC-enhanced VL data, while at the same time being parameter efficient and significantly reducing forgetting, thus, maintaining the VL model Zero-Shot (ZS) performance. Qualitative examples showing improvements attained by our proposed approach, as well as some examples of failure cases, are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.509, + 0.833, + 0.525 + ], + "angle": 0, + "content": "3.1. Teaching SVLC Using NLP and LLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.743 + ], + "angle": 0, + "content": "In this section, we present several ways in which the data of existing VL pre-training paired datasets can be enhanced to emphasize SVLC in the texts, and teaching them to the VL model. We propose two kinds of data enhancements - generating negative and positive text alternatives. When generating negative examples, only one word of the sentence is changed such that the semantic meaning of the sentence changes. We propose two methods for the generation of the negatives: (i) rule-based (Sec. 3.1.1); and (ii) LLM-based (Sec. 3.1.2). Positive alternatives are generated as sentences with semantically similar meanings, but different wording (Sec. 3.1.3). We then present the losses which properly take these two types of generated textual data into account during training in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.759, + 0.825, + 0.774 + ], + "angle": 0, + "content": "3.1.1 Generating Rule-Based (RB) Negatives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "One simple yet effective method for negative text generation is using a collection of pre-defined language rules which match and replace words of a specific entity type or a pattern, such as color, material, size, etc. This method is especially useful when one has prior knowledge of a specific aspect of the language that needs to be taught. For example, if we know that our model lacks the ability to understand the colors of objects, we can easily create a rule for detecting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2660" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.351 + ], + "angle": 0, + "content": "and replacing color words in the text, for generating negative text that does not correspond to its paired image. To employ the generation of the rule-based negative, for each taught SVLC we define a list of words belonging to its characteristic. We then scan the VL data texts searching for the words within these lists. If a word is located, we simply replace it with a randomly selected word from the same list to generate a negative pair. For example, applying the color-rule to a sentence: \"A big brown dog\" can lead to \"A big yellow dog\". If a text has multiple candidates of words to be replaced, one of them is chosen randomly. We perform this process multiple times for the full list of SVLC characteristics of interest such as color, size, material, spatial relations, etc. These generated negative texts are SVLC specific and differ from the original text in only one word. For a detailed description and more examples of the RB negative generation, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.471, + 0.384 + ], + "angle": 0, + "content": "3.1.2 LLM-based Negative Generation via Unmasking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.755 + ], + "angle": 0, + "content": "A natural extension of rule-based negatives technique is the generation of negatives using Large Language Models (LLMs) unmasking. Recent LLMs are explicitly trained in a self-supervised manner with the objective of \"unmasking\" parts of the text. Given a sentence with one missing word, models such as BERT [13], can suggest multiple words that fit the context of the sentence. Using this useful property of LLMs, we can therefore automatically create plausible negative examples without the need for prior knowledge of the SVLC characteristics of interest. In order to focus the randomly selected masked words to be likely to belong to SVLCs of interest, we use common NLP parsing techniques (such as spacy [26]) to parse the sentence into its components such as nouns, verbs, adjectives, adverbs, etc. We then randomly choose a type of sentence part and a word belonging to this part type, mask out the selected word and replace it with one of the options suggested by the LLM's unmasking. These negative examples, when used properly in the loss function (Eq. (3)) focus the network on the important details that affect the SVLC understanding. As we show in Sec. 4, this method is extremely useful and can significantly improve the VL model's understanding of different SVLCs. Further details and examples of the LLM negative generation are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.465, + 0.789 + ], + "angle": 0, + "content": "3.1.3 Generating Text Analogies via LLM Prompting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "While the goal of the negative text generation (Sec. 3.1.1 and Sec. 3.1.2) was to make minor perturbations to a given text such that the meaning changes, the goal here is exactly the opposite. We would like to make major changes to the text, while still keeping the same semantic meaning. For example “A woman standing left to a sitting cat” and “A cat sitting to the right of a standing woman” are two very" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "different texts describing the exact same scene. One effective way to generate such semantically similar texts is by prompting the foundational LLMs. Specifically, we use the open access BLOOM [53] model. In the spirit of recently popular in-context learning [18], we present the model with a textual prompt with examples of semantically similar texts (see Fig. 2). We then append the current image caption and retrieve the BLOOMs prediction of a semantically similar text. For a detailed description, we refer the reader to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.587, + 0.266 + ], + "angle": 0, + "content": "3.2. Losses" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.273, + 0.894, + 0.35 + ], + "angle": 0, + "content": "All of our evaluated models (CLIP [57] and CyCLIP [19]) admit a text & image pair \\((T, I)\\) and are comprised of two parts: (i) image encoder \\(e_I = \\mathcal{E}_I(I)\\); (ii) text encoder \\(e_T = \\mathcal{E}_T(T)\\). In this notation, the text-to-image similarity score is therefore computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.36, + 0.892, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (T, I) = \\exp \\left(\\frac {\\tau e _ {T} ^ {T} e _ {I}}{| | e _ {T} | | ^ {2} | | e _ {I} | | ^ {2}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.793, + 0.42 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a learned temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Contrastive Loss. As most contemporary VL models, we employ the contrastive CLIP-loss [57] as one of our losses for each batch \\(\\mathcal{B}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.476, + 0.892, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t}} = \\sum_ {i} \\log \\left(\\frac {S \\left(T _ {i} , I _ {i}\\right)}{\\sum_ {j} S \\left(T _ {i} , I _ {j}\\right)}\\right) + \\log \\left(\\frac {S \\left(T _ {i} , I _ {i}\\right)}{\\sum_ {k} S \\left(T _ {k} , I _ {i}\\right)}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.894, + 0.636 + ], + "angle": 0, + "content": "Negatives Loss. In our ablation study in Sec. 5, we show that for a given text \\( T_{i} \\) simply adding the corresponding generated negative text \\( T_{i}^{neg} \\) to the contrastive loss \\( \\mathcal{L}_{cont} \\) is much less effective than having a separate loss individually attending to the similarity difference of \\( T_{i} \\) and \\( T_{i}^{neg} \\) w.r.t. the image \\( I_{i} \\) corresponding to \\( T_{i} \\). We, therefore, employ the following negatives loss:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.646, + 0.892, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n e g} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} , I _ {i}\\right)}{\\mathcal {S} \\left(T _ {i} , I _ {i}\\right) + \\mathcal {S} \\left(T _ {i} ^ {n e g} , I _ {i}\\right)}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Analogy Loss. For the generated analogy texts \\( T_{i}^{sim} \\) produced from the text \\( T_{i} \\) which corresponds to image \\( I_{i} \\), we employ the combination of the two following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.754, + 0.892, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i m} ^ {t e x t} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} ^ {s i m} , T _ {i}\\right)}{\\sum_ {j} \\mathcal {S} \\left(T _ {i} ^ {s i m} , T _ {j}\\right)}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.892, + 0.852 + ], + "angle": 0, + "content": "where with some abuse of notation, \\( S(T_{1}, T_{2}) \\) denotes the exponent cosine similarity between text \\( T_{1} \\) and text \\( T_{2} \\) text-embeddings, and:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.864, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i m} ^ {i m g} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} ^ {s i m} , I _ {i}\\right)}{\\sum_ {j} \\mathcal {S} \\left(T _ {i} ^ {s i m} , I _ {j}\\right)}\\right), \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2661" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.122 + ], + "angle": 0, + "content": "which simply corresponds to the second summand of \\(\\mathcal{L}_{cont}\\) in Eq. (2), with replacing \\(T_{i}\\) by \\(T_{i}^{sim}\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.122, + 0.277, + 0.136 + ], + "angle": 0, + "content": "Our final loss is, therefore:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.147, + 0.469, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t}} + \\alpha \\cdot \\mathcal {L} _ {\\text {n e g}} + \\beta \\cdot \\left(\\mathcal {L} _ {\\text {s i m}} ^ {\\text {t e x t}} + \\mathcal {L} _ {\\text {s i m}} ^ {\\text {i m g}}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.181, + 0.388, + 0.198 + ], + "angle": 0, + "content": "3.3. Fine-tuning Pre-trained VL Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.203, + 0.47, + 0.279 + ], + "angle": 0, + "content": "Each of the \\(\\mathcal{E}_T\\) and \\(\\mathcal{E}_I\\) networks is comprised of a mix of non-parametric functions and two types of parametric functions: linear layers and embedding layers. Roughly, each of those functions, \\(\\mathcal{F}_k^{lin}(x)\\) and \\(\\mathcal{F}_k^{emb}(x)\\) where \\(k\\) is layer index, is parameterized by a weight matrix \\(\\mathcal{W}_k\\) so that:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.288, + 0.469, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {k} ^ {\\text {l i n}} (x) = \\mathcal {W} _ {k} \\cdot x \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.309, + 0.469, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {k} ^ {e m b} (x) = E M B (x; \\mathcal {W} _ {k}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.469, + 0.44 + ], + "angle": 0, + "content": "where \\( EMB \\) is the embedding operator assuming \\( x \\) is a stream of integers and picking the respective columns of \\( \\mathcal{W}_k \\). Following the idea proposed in LoRA [27] for efficient LLM fine-tuning using low-rank residual adapters, when adapting a pre-trained VL model \\( \\mathcal{M} = (\\mathcal{E}_T, \\mathcal{E}_I) \\) we parameterize the adapted weights \\( \\mathcal{W}_k^* \\) of the model \\( \\mathcal{M}^* \\) fine-tuned from \\( \\mathcal{M} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.443, + 0.469, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {k} ^ {*} = \\mathcal {W} _ {k} + \\mathcal {A} _ {k} \\cdot \\mathcal {B} _ {k} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.464, + 0.469, + 0.51 + ], + "angle": 0, + "content": "where for \\(\\mathcal{W}_k\\) of size \\(m\\times l\\), \\(\\mathcal{A}_k\\) and \\(\\mathcal{B}_k\\) are rank- \\(r\\) matrices of sizes \\(m\\times r\\) and \\(r\\times l\\) respectively. These low-rank residual adapters can be applied efficiently as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.518, + 0.469, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {k} ^ {*, l i n} (x) = \\mathcal {F} _ {k} ^ {l i n} (x) + \\mathcal {A} _ {k} \\cdot (\\mathcal {B} _ {k} \\cdot x) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.54, + 0.469, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {k} ^ {*}, e m b (x) = \\mathcal {F} _ {k} ^ {e m b} (x) + \\mathcal {A} _ {k} \\cdot E M B (x; \\mathcal {B} _ {k}) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.469, + 0.644 + ], + "angle": 0, + "content": "During the fine-tuning of \\(\\mathcal{M}^*\\), we freeze all the base model \\(\\mathcal{M}\\) parameters \\(\\forall k, \\{\\mathcal{W}_k\\}\\) and only the LoRA adapters \\(\\forall k, \\{(\\mathcal{A}_k, \\mathcal{B}_k)\\}\\) are being learned. In the above notation we disregard possible bias terms of the linear functions, if they are present, since we keep them frozen too." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.902 + ], + "angle": 0, + "content": "There are several interesting things to note about the proposed architecture: (i) as opposed to [64], who evaluated the use of efficient LLM fine-tuning techniques for VL models adaptation, we add our LoRA adapters everywhere, i.e. to all layers of both the text and image encoders, and not only to the text encoder/decoder as done in [64]; (ii) as opposed to [80] who attach a small side network only to the output of the adapted model, our LoRA adapters are added to all the parametric functions inside the model and affect all the intermediate computations; (iii) same as noted in the original [27] paper, at inference all the LoRA adapters can be folded back into the original weight matrices by simple summation, thus returning the number of total parameters to be the same as in the original model and hence have the same inference speed; (iv) with rank \\( r \\) kept low, the number of extra parameters added by all the LoRA adapters can be very low making adaptation fast and efficient; finally," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "(v) such form of fine low-rank adaptation allows for significantly mitigating the zero-shot performance forgetting effects as demonstrated in our results and explored in the corresponding ablation (Sec. 5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.189, + 0.632, + 0.207 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.221, + 0.602, + 0.235 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.31 + ], + "angle": 0, + "content": "We train the model using common Image-Text pair datasets, namely Conceptual Captions 3M [62] and LAION 400M [61] and test using the VL-Checklist [82] datasets, which will be described below." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Conceptual Captions 3M (CC3M) [62] is a dataset of three million image-text pairs automatically crawled from the internet where image descriptions are harvested from Alt-text attributes and then processed and filtered to create relatively clean descriptions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.522 + ], + "angle": 0, + "content": "LAION 400M [61] is a very large scale image-text pair dataset which, similarly to CC3M has been automatically harvested from the internet. One major difference between the two, apart from the size, is that LAION examples have been filtered using the pretrained CLIP model, such that the CLIP image-text similarity is high by design. Recent re-implementations of the original CLIP paper have successfully used LAION 400M to reproduce similar capabilities as the original CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.892, + 0.793 + ], + "angle": 0, + "content": "VL-Checklist [82] combines images and annotations from the Visual Genome [39], SWiG [55], VAW [54], and HAKE [48] datasets. It is processed such that each image is annotated with two captions, one positive and one negative. The positive caption originates in its source dataset and corresponds to the image. The negative caption is constructed from the positive caption so that only one word, corresponding to the tested SVLC of interest, is changed to negate the SVLC (e.g., color, size, material, etc.). VL-Checklist evaluates three main types of VL concepts further subdivided into 7 types of SVLCs total: (1) Object: its spatial location and size, (2) Attribute: color, material, size, state, and action, and (3) Relation: spatial or action relation between two objects. In the following sections we report results on a combined VL-Checklist dataset. The results on individual comprising datasets (Visual Genome [39], SWiG [55], VAW [54], and HAKE [48]) are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Zero-Shot Classification We evaluated our method on 21 classification dataset using the Zero-Shot classification protocol described in the ELEVATER Image Classification Toolkit [42]. The evaluation includes 21 different datasets, including common classification datasets such as ImageNet [60], CIFAR100 [40], EuroSat [21], and others. In Tables 1-3 we report the average results over the 21 tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2662" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.821, + 0.196 + ], + "angle": 0, + "content": "
ObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CLIP [57]81.58%67.60%63.05%56.37%
CLIP +LoRA80.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41%(+0.04%)
w/o Neg Loss82.27% (+0.69%)67.58% (-0.02%)55.17% (-7.88%)55.37% (-1.00%)
Ours RB+LLM Negs85.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.174, + 0.205, + 0.796, + 0.22 + ], + "angle": 0, + "content": "Table 2. Ablation study - separate Negative Losses (Sec. 3.2) vs adding negatives to contrastive loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.233, + 0.295, + 0.249 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.471, + 0.513 + ], + "angle": 0, + "content": "For CLIP we use the ML-Foundation Open-CLIP repository [28] and for CyCLIP [19] we use its original code repository, which is also based on Open-CLIP. In most experiments, unless stated otherwise, we train our model for five epochs on 4 V100 NVIDIA GPUs, with a total batch size of 512. When starting from a pre-trained model, we use rank 4 LoRA adapters (Sec. 3.3) and the learning rate is set to \\(5E - 6\\). When training from scratch, for CLIP we use the default parameters set in the open-CLIP library, and for CyCLIP the defaults of its original implementation. For all CLIP experiments, we use VIT/32-B as the model architecture and ResNet-50 when training CyCLIP (following [19]). When fine-tuning we initialize with the original model weights released by the respective authors. In all experiments involving a combination of CyCLIP [19] and our method (in Sec. 4.4-4.5), in addition to Eq. (6) loss we also employ all the extra losses proposed in CyCLIP [19]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.517, + 0.185, + 0.532 + ], + "angle": 0, + "content": "4.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.72 + ], + "angle": 0, + "content": "We compare our method to two strong baselines under several configurations. The first is the CLIP [57] OpenAI pretrained model trained on 400M image-text pairs which achieves high ZS performance. The second is the very recent CyCLIP [19] method which, similarly to us, improves over the original CLIP loss. For a fair comparison all methods use the same network architecture and the same initialization from the OpenAI pretrained model. For the CyCLIP baseline we continue training from the pretrained initialization using LoRA and the CyCLIP losses. As CyCLIP losses are orthogonal to ours, we also show a unified version of the two methods (CyCLIP + Ours Combined)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.295, + 0.743 + ], + "angle": 0, + "content": "4.4. Fine-tuning VL Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the following experiments we show the effects of fin-tuning a pre-trained VL model using our additional data enhancement methods and losses. All experiments are initialized from the official OpenAI CLIP. We compare our method to two baselines on the VL-Checklist tasks, and on the Zero-Shot image classification task. The first baseline is the original pretrained model without any further training. The second, is the same model with the additional LoRA parameters, trained on CC3M (Table 1a) and LAION (Table 1g) using the original CLIP loss function. Table 1a" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.23, + 0.838, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.439, + 0.894, + 0.496 + ], + "angle": 0, + "content": "Figure 3. CC3M fine-tuning - detailed. Detailed results of the baselines (CLIP, CLIP+LoRA) and our models (RB+LLM Neg, RB+LLM Negs+Pos) initialized from OpenAI pretrained CLIP ViT-32B and fine-tuned (except CLIP) on CC3M." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.893, + 0.756 + ], + "angle": 0, + "content": "shows several configurations of our method compared to the baselines. We see that our method shows significant improvements on all VL-Checklist tasks reaching up to \\(15\\%\\) improvement. Figure 3 displays the relative gains on all tasks of the \"Attribute\" and \"Relation\" tests. It is clear that our gains are across all tests. These improvements come at a price of some minor degradation to the Zero-Shot performance compared to CLIP. Moreover, when fine-tuning CLIP on the LAION dataset (Table 1g) we do not see these degradations. In Tab. 1b we show consistent gains to ones in Tab. 1a when finetuning with a stronger (higher ZS) VitB/16 CLIP [57] pre-trained image encoder, observing an even lower drop in ZS performance. In Tab. 1e we show significant gains in fine-tuning CyCLIP on CC3M, comparing fine-tuning before and after integrating our proposed approach, this also comes at no ZS performance cost." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.707, + 0.787 + ], + "angle": 0, + "content": "4.5. Training from scratch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In these experiments, we show the advantage of integrating our approach throughout the whole training procedure. To this end, we train from scratch on CC3M - both CLIP and CyCLIP, either on their own, or together with our proposed method integrated. Tables - Tab. 1c, Tab. 1d, and Tab. 1f show that, similar to finetining, our method significantly improves both CLIP and CyCLIP SVLC capabilities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2663" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.082, + 0.865, + 0.196 + ], + "angle": 0, + "content": "
MethodColorMaterialAttribute SizeStateActionRelation21 Zero-Shot Tasks Average
ActionSpatial
Ours Pos72.35%69.25%69.80%59.35%66.08%70.97%39.20%55.37%
Ours RB Neg78.45%83.20%69.50%65.95%69.66%75.97%74.70%54.66%
Ours LLM Neg76.00%79.70%72.75%61.35%68.36%77.23%72.40%55.05%
Ours RB+LLM Negs79.25%84.25%72.15%64.05%69.82%79.03%78.40%54.66%%
Ours Combined77.45%77.35%73.35%62.30%69.39%74.70%63.20%54.77%%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.185, + 0.206, + 0.784, + 0.22 + ], + "angle": 0, + "content": "Table 3. Ablation study - component analysis. Detailed evaluation on Attribute & Relation SVLCs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.471, + 0.34 + ], + "angle": 0, + "content": "ties, while still keeping similar ZS performance. Figure 4 shows the detailed parsing of the \"Attribute\" and \"Relation\" tests with consistent gains across the specific tasks. These findings suggest that had we trained our method for many epochs on a large dataset, such as the full LAION-9B, we would reach similar ZS performance while greatly enhancing the SVLC performance of the model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.353, + 0.185, + 0.368 + ], + "angle": 0, + "content": "5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.376, + 0.47, + 0.407 + ], + "angle": 0, + "content": "In this section we will examine several aspects and components of our proposed methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.236, + 0.431 + ], + "angle": 0, + "content": "5.1. Negative Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.471, + 0.587 + ], + "angle": 0, + "content": "Section 3.2 details our additional loss functions which utilize our generated texts. Specifically, Eq. (3) describes the loss which contrasts a given example only with its generated negatives. Table 2 clearly shows the importance of this loss as opposed to simply adding the negative examples to the original contrastive loss (Eq. (2)). Without explicitly forcing the network to attend to the small changes in the text (Eq. (3)), the generated negative examples, when inserted only to the contrastive loss (Eq. (2)) do not provide the desired gains over the baseline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.596, + 0.27, + 0.612 + ], + "angle": 0, + "content": "5.2. Component analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.664 + ], + "angle": 0, + "content": "In Section 3 we describe several components of our proposed method. Specifically we present our RB negative generation (Sec. 3.1.1), our LLM-based negative genera" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.683, + 0.396, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.866, + 0.47, + 0.907 + ], + "angle": 0, + "content": "Figure 4. CC3M train from scratch - detailed. Detailed results of the baseline CLIP and our models (RB+LLM Neg, RB+LLM Negs+Pos), all CLIP ViT-32B, trained from scratch on CC3M." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.895, + 0.432 + ], + "angle": 0, + "content": "tion (Sec. 3.1.2), and our LLM-based analogy generation, referred to here as \"Pos\" (Sec. 3.1.3). Tab. 3 provides a detailed analysis and comparison of the contribution of each component to the final result. Through this analysis we see two contradicting forces between the SVLC capabilities and the original Zero-Shot performance. We can see that each negative generation method plays a crucial role in some of the tasks while the use of both is usually the best performing option. On the other hand, the LLM-based analogy generation stabilizes the Zero-Shot performance and mitigates the drop with respect to the baseline. The joint version of all components (\"Ours Combined\") exhibits a good trade-off between the two contradicting forces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.443, + 0.628, + 0.458 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.602 + ], + "angle": 0, + "content": "We have presented a data-driven technique for enhancing the performance of VL models in the important task of SVLC understanding without sacrificing their impressive ZS object recognition capabilities. Our proposed method attains significant gains in multiple experiments on a variety of base VL models and datasets. It builds upon the modeling strength and knowledge of language structure to teach this structure to VL models in an orthogonal way, suggesting wide applicability to existing or future VL models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.753 + ], + "angle": 0, + "content": "While attaining impressive gains in SVLC understanding, we believe the small drop observed in ZS performance could be further reduced. An additional possible extension of our work is using more sophisticated sequence generation techniques for improving our batch data. One may combine annotation efforts with a language model to get improved data for training and evaluation [51]. Another possibility is adding a corrector [72] in the training that validates whether the VL model learns the correct concepts or not. We leave all of these exciting directions to future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.767, + 0.669, + 0.784 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We would like to thank Donghyun Kim for his invaluable help in this work. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. FA8750-19-C-1001. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA). This material is based upon work supported by the 'Data Science grant from the Israeli Council of Higher Education'. This research or RG was supported by ERC-StG SPADE grant no. 75749." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2664" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.241 + ], + "angle": 0, + "content": "[2] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[3] Amir Bar, Roei Herzig, Xiaolong Wang, Anna Rohrbach, Gal Chechik, Trevor Darrell, and A. Globerson. Compositional video synthesis with action graphs. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, pages 105-121, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[5] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[6] Andrea Burns, Donghyun Kim, Derry Wijaya, Kate Saenko, and Bryan A Plummer. Learning to scale multilingual representations for vision-language tasks. In European Conference on Computer Vision, pages 197-213. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.471, + 0.684 + ], + "angle": 0, + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.688, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.471, + 0.799 + ], + "angle": 0, + "content": "[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[14] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't stop learning: Towards continual learning for the clip model. arXiv:2207.09248, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.166, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[15] Andreas Fürst, Elisabeth Rumetshofer, Viet Tran, Hubert Ramsauer, Fei Tang, Johannes Lehner, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto-Nemling, et al. Cloob: Modern hopfield networks with infoloob outperform clip. arXiv preprint arXiv:2110.11316, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[16] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. ArXiv, abs/2008.11714, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[18] Shivam Garg, Dimitris Tsipras, Percy Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[19] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.436, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[20] Xiaoshuai Hao, Yi Zhu, Srikar Appalaraju, Aston Zhang, Wanqian Zhang, Bo Li, and Mu Li. Mixgen: A new multimodal data augmentation. arXiv preprint arXiv:2206.08358, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[22] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.62, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[23] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0–0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[25] Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, and Amir Globerson. Mapping images to scene graphs with permutation-invariant structured prediction. In Advances in Neural Information Processing Systems (NIPS), 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2665" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.2,3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[29] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[30] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as composition of spatiotemporal scene graphs. arXiv preprint arXiv:1912.06992, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[32] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[33] Keizo Kato, Yin Li, and Abhinav Gupta. Compositional learning for human object interaction. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[34] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[35] Donghyun Kim, Kuniaki Saito, Kate Saenko, Stan Sclaroff, and Bryan Plummer. Mule: Multimodal universal language embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11254-11261, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[37] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[38] Ranjay Krishna, Ines Chami, Michael S. Bernstein, and Li Fei-Fei. Referring relationships. *ECCV*, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[39] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[41] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[42] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. Neural Information Processing Systems, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[44] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[45] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. ArXiv, abs/1908.03557, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[46] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[47] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[48] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Mingyang Chen, Ze Ma, Shiyi Wang, Hao-Shu Fang, and Cewu Lu. Hake: Human activity knowledge engine. arXiv preprint arXiv:1904.06539, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[49] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[51] Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. Wanli: Worker and ai collaboration for natural language inference dataset creation, Jan. 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[52] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Margaret Mitchell, Giada Pistilli, Yacine Jernite, Ezinwanne Ozoani, Marissa Gerchick, Nazneen Rajani, Sasha Luccioni, Irene Solaiman, Maraim Masoud, Somaieh Nikpoor, Munoz Ferrandis Carlos, Stas Bekman, Christopher Akiki, Danish Contractor, David Lansky, Angelina McMillan-Major, Tristan Thrush, Suzana Ilic, Gerard Dupont, Shayne Longpre," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2666" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "Manan Dey, Stella Biderman, Douwe Kiela, Emi Baylor, Teven Le Scao, Aaron Gokaslan, Julien Launay, and Niklas Muennighoff. Bigscience, bigscience language open-science open-access multilingual (bloom) language model. International, May 2021-May 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[54] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13018-13028, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[55] Sarah Pratt, Mark Yatskar, Luca Weihs, Ali Farhadi, and Aniruddha Kembhavi. Grounded situation recognition. In European Conference on Computer Vision, pages 314-332. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[56] Moshiko Raboh, Roei Herzig, Gal Chechik, Jonathan Berant, and Amir Globerson. Differentiable scene graphs. In WACV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.47, + 0.419 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.476 + ], + "angle": 0, + "content": "[58] Vinay Venkatesh Ramasesh, Ethan Dyer, and Maithra Raghu. Anatomy of catastrophic forgetting: Hidden representations and task semantics. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[59] Arijit Ray, Karan Sikka, Ajay Divakaran, Stefan Lee, and Giedrius Burachas. Sunny and dark outside?! improving answer consistency in vqa through entailed question generation. arXiv preprint arXiv:1909.04696, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[61] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[62] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[63] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[64] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[65] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[66] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[67] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In European Conference on Computer Vision, pages 437-453. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[68] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[69] William Yang Wang and Diyi Yang. That's so annoying!!! A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using# petpeeve tweets. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 2557–2563, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[70] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[71] Jason Wei and Kai Zou. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[72] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generating sequences by learning to self-correct, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[73] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[74] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and M. Kankanhalli. Learning to detect human-object interactions with knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[75] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene Graph Generation by Iterative Message Passing. In CVPR, pages 3097–3106, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[76] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.832 + ], + "angle": 0, + "content": "[77] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[78] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2667" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.203 + ], + "angle": 0, + "content": "[80] Jeffrey O Zhang, Alexander Sax, Amir Zamir, Leonidas Guibas, and Jitendra Malik. Side-tuning: a baseline for network adaptation via additive side networks. In European Conference on Computer Vision, pages 698-714. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.206, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[81] Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[82] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. V1-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[83] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022.4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2668" + } + ] +] \ No newline at end of file diff --git a/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_origin.pdf b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0120d7a6bf7601243b6c04955bf645578fe38513 --- /dev/null +++ b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/9a8c6788-2d0c-421a-8a6a-0e2eb6239bd0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87b641965a3de2cba35ca28d37db96482cef5de999f59cd695304954f4b7b2a0 +size 911552 diff --git a/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/full.md b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f07b247cf2c3a405164978d61af44c42bcaba37 --- /dev/null +++ b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/full.md @@ -0,0 +1,324 @@ +# Teaching Structured Vision & Language Concepts to Vision & Language Models + +Sivan Doveh $^{1,2}$ , Assaf Arbelle $^{1}$ , Sivan Harary $^{1}$ , Eli Schwartz $^{1,3}$ , Roei Herzig $^{1,3}$ , Raja Giryes $^{3}$ , Rogerio Feris $^{4}$ , Rameswar Panda $^{4}$ , Shimon Ullman $^{*2}$ , Leonid Karlinsky $^{*4}$ + +1IBM Research, 2Weizmann Institute of Science, 3Tel-Aviv University, 4MIT-IBM Watson AI Lab + +# Abstract + +Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to $15\%$ in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC + +# 1. Introduction + +Recent Vision & Language (VL) models [19, 31, 43, 44, 47, 57] achieve excellent zero-shot performance with respect to various computer-vision tasks such as detection, classification, segmentation, etc. However, recent studies [68, 82] have demonstrated that even the strongest VL models struggle with the compositional understanding of some basic Structured VL Concepts (SVLC) such as ob- + +![](images/130aaaae7f82b9506eb17458c9f4fa69eccce7e95d52727c6908bd568f38a1ad.jpg) +Figure 1. Teaching language structure to VL models. (a) Standard contrastive text-to-image loss (e.g. CLIP [57]) tends to under-emphasize SVLC content of the text, likely due to the random nature of the training batches; (b) We generate modified versions of corresponding texts and use them to add losses to explicitly teach language structure (SVLC) to VL models. + +ject attributes, inter-object relations, transitive actions, object states and more. Collecting specialized large scale data to teach VL models these missing 'skills' is impractical, as finding specialized text-image pairs for each kind and possible value of the different attributes, relations, or states, is both difficult and expensive. + +Another important challenge in training VL models with new concepts is catastrophic forgetting, which is a common property to all neural models [7, 34, 37, 49, 58] and has been explored for VL models in a recent concurrent work [14]. Large VL models such as CLIP [57] and CyCLIP [19] have exhibited excellent zero-shot learning abilities in many tasks. Therefore, even given a large dataset with new concepts, it is important not to lose these abilities when performing the adaptation to the new data. + +![](images/4eb3b49556f4bacdf3f366b210e8a3a0c60d28259ceb52609f8cf3fbecf243a4.jpg) +Figure 2. Teaching structured image understanding to VL models via structured textual data manipulation harnessing the power of language modeling. (1) Generating Rule-Based negative texts (Sec. 3.1.1); (2) Generating negatives using Large Language Model (LLM) unmasking (Sec. 3.1.2); (3) Generating analogies (positives) via LLM prompting (Sec. 3.1.3). + +In this paper, we propose a way to leverage existing (off-the-shelf) VL pre-training data sources in order to improve the SVLC understanding skills of a given model, while at the same time maintaining its zero-shot object recognition accuracy. Naturally, succeeding in this goal would lead to potential improvement w.r.t. SVLC understanding in a wide variety of downstream tasks building upon pre-trained VL models, such as zeros-shot detection, segmentation, image generation, and many more. + +Recent research [68, 82] has shown that VL models exhibit an 'object bias' partially due to the contrastive text-to-image loss used in their pre-training. For example, the popular CLIP-loss [57] is computed over a random batch of text-image pairs sampled from a large-scale and diverse VL dataset with the chance of two images in the same batch containing the same set of objects being very low. For such a loss, representing just a 'bag of objects' in each image or text is sufficient for matching the corresponding pairs. Intuitively, this leads to the 'object bias' where SVLCs like attributes, states, and relations are being underrepresented (e.g. having a much smaller amplitude in the resulting feature superposition), consequently causing the aforementioned issues with SVLC understanding. + +Based on this intuition, we propose a simple data-driven technique that harnesses existing language parsing and modeling capabilities to enhance the importance of SVLCs in the VL model training losses. For each text in the training batch, we automatically generate alternative negative or positive text by manipulating its content to be opposite or equivalent to the original text. Using the newly generated texts, we explicitly teach SVLC to the model via additional losses (see Fig. 1) that enforce differentiating be + +tween different (original and generated) SVLC texts and are no longer satisfiable by the 'bag of objects' representation. + +Towards this end, we propose several techniques for implementing this approach, including (i) rule-based priors based on classical NLP parsing and word substitution vocabulary according to attribute/relation type; (ii) prompting a Large Language Model (LLM) (e.g. [53]) for analogous text; (iii) generating different meaning (negative) alternatives by LLM-based unmasking of parsed text entities of different kinds; (iv) combinations of these methods. + +We demonstrate that all these techniques can lead to significant improvements of up to $15\%$ percent when measuring the VL models' SVLC understanding. We verify this on 5 datasets: VG [39], HAKE [48], VAW [54], SWIG [55], and all combined, using the protocol recently proposed in VL-Checklist [82]. In addition, we show that the resulting VL models largely preserve their zero-shot object recognition performance. For the latter, we also propose a variant of efficient LLM fine-tuning using low-rank residual adapters (LoRA) [27] adjusted to VL models. Finally, we show that our framework allows better harnessing of the standard available VL data, e.g. CC3M [62] and LAION [61]. This is exhibited by the aforementioned gains, both in new VL models trained from scratch, as well as in models fine-tuned using strong available VL models such as CLIP [57] and CyCLIP [19]. + +To summarize, we offer the following contributions: (i) We propose a data-driven approach for better harnessing the standard available VL data to improve VL models' SVLC understanding skills, such as understanding object attributes, inter-object relations, transitive actions, object states, and more, without sacrificing zero-shot object recog + +
DatasetModelPreArchObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CC3MCLIP [57]✓Vit-B/3281.58%67.60%63.05%56.37%(a)
CLIP + LoRA✓Vit-B/3280.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41% (+0.04%)
CLIP + Ours RB Neg✓Vit-B/3283.89% (+2.30%)73.35% (+5.75%)75.33% (+12.28%)54.32% (-2.05%)
CLIP + Ours LLM Neg✓Vit-B/3284.44% (+2.85%)71.63% (+4.03%)74.82% (+11.77%)55.60% (-0.77%)
CLIP + Ours RB+LLM Negs✓Vit-B/3285.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
CLIP + Ours Combined✓Vit-B/3285.00% (+3.42%)71.97% (+4.37%)68.95% (+5.90%)54.77% (-1.60%)
CLIP [57]✓Vit-B/1682.91%67.32%61.80%60.00%(b)
CLIP + Ours RB+LLM Negs✓Vit-B/1685.82% (+2.91%)73.92% (+6.6%)77.40% (+15.6%)59.37% (-0.63%)
CLIP + Ours Combined✓Vit-B/1684.75% (+1.84%)71.18% (+3.86%)69.68% (+7.88%)59.87% (-0.13%)
CLIPXVit-B/3271.17%57.86%45.20%21.96%(c)
CLIP + Ours CombinedXVit-B/3271.79% (+0.62%)63.29% (+5.43%)58.13% (+12.93%)20.96% (-1.00%)
CLIP [57]XVit-B/1664.01%54.27%41.57%15.49%(d)
CLIP + Ours RB+LLM NegsXVit-B/1673.11% (+9.1%)65.32% (+11.05)71.93% (+30.36)20.78% (+5.29%)
CLIP + Ours CombinedXVit-B/1672.99% (+8.98%)63.01% (+8.74%)62.95% (+21.38)20.61% (+5.12%)
CyCLIP [19]✓R5073.49%59.33%53.83%26.00%(e)
CyCLIP + LoRA✓R5073.30% (-0.19%)58.89% (-0.44%)53.03% (-0.80%)26.30% (+0.30%)
CyCLIP + Ours Combined✓R5074.20% (+0.71%)63.52% (+4.20%)59.47% (+5.63%)26.31% (+0.31%)
CyCLIPXR5069.41%57.59%53.70%21.02%(f)
CyCLIP + Ours CombinedXR5071.50% (+2.09%)65.69% (+8.10%)70.20% (+16.50%)20.44% (-0.42%)
LAIONCLIP [57]✓Vit-B/3281.58%67.6%63.05%56.37%(g)
CLIP + LoRA✓Vit-B/3282.18% (+0.60%)68.48% (+0.88%)62.72% (-0.33%)57.15% (+0.78%)
CLIP + Ours Combined✓Vit-B/3282.54% (+0.96%)69.64% (+2.04%)66.05% (+3.00%)56.71% (+0.34%)
+ +Table 1. VL-Checklist and Zero-Shot classification evaluation (ImageNet + 20 datasets). Finetuned models (for 5 epochs as detailed in Sec. 3.3, starting from officially released CLIP [57] and CyCLIP [19] weights) are marked with $\checkmark$ in the Pre-trained (Pre) column, while models trained from scratch (for 10 epochs) are marked with $\times$ . The gains and losses of our approach (+Ours) are in color and are computed w.r.t. to corresponding baselines in each section. CLIP/CyCLIP + LoRA indicate finetuning in the same way and on the same data, but without our approach. Finetuning without LoRA on the same data yields significantly worse performance in all metrics. Sections are separated by bold horizontal lines. (a) CC3M fine-tuning - CLIP - Vit-B/32: we significantly improve the SVLC understanding, observing only small ZS performance drops, $0.77\%$ ; (b) CC3M fine-tuning - CLIP - Vit-B/16: we can observe similar improvements as in (a), with an even smaller impact on ZS performance; (c) CC3M from scratch - CLIP - Vit-B/32: we significantly improve SVLC understanding with only a small ( $1\%$ ) decrease in ZS performance; (d) CC3M from scratch - CLIP - Vit-B/16: compared to (c), even greater SVLC understanding improvement is observed (up to $30.36\%$ ), at no cost in ZS performance, and even improvement of over $5\%$ ; (e) CC3M fine-tuning - CyCLIP: we use CyCLIP original code (with LoRA integration) and losses, as can be seen - adding our techniques improves CyCLIP SVLC performance considerably without sacrificing ZS performance; (f) CC3M from scratch - CyCLIP: observing even largergains in SVLC understanding compared to (e) (up to $16.5\%$ ), with small reduction in ZS performance of $0.42\%$ ; (g) LAION fine-tuning - CLIP - Vit-B/32: we improve SVLC understanding without any decrease in ZS performance. + +nition performance; (ii) More specifically, we propose to leverage the well-understood and well-modeled structure of language, through classical NLP parsing and/or use of the modern pre-trained LLMs, for manipulating the text part of the standard VL paired datasets to regularize VL training and teach SVLC understanding to VL models. (iii) We further propose an adaptation of efficient LLM fine-tuning technique of [27] for fine-tuning VL models, allowing for only minimal reduction in zero-shot object recognition performance after fine-tuning, while still obtaining the aforementioned SVLC understanding gains. (iv) Empirically, for + +the popular CLIP [57] and its most recent extension CyCLIP [19], we demonstrate SVLC understanding average improvements of up to $13\%$ when training from scratch, and $15\%$ when fine-tuning from a pretrained model. + +# 2. Related Work + +Vision-language (VL) Models. (e.g., CLIP [57] and ALIGN [31]) show significant advances in diverse zero-shot downstream tasks. They are pre-trained using contrastive image-text alignment on a large-scale noisy dataset of text-image pairs collected from the web. Several methods [9,46, + +[65] additionally employ off-the-shelf object detectors to extract region features. In order to relax this limitation, some methods [31,36,43,76] propose to use cross-attention layers with self-supervised learning objectives including image-text matching and masked/autoregressive language modeling. BLIP [43] generates synthetic captions from the language modeling head and filters noisy captions based on the image-text matching score. Recently, there have been attempts to learn finer-level alignment and relations between image and text [15, 17, 19, 47, 77]. FILIP proposes fine-grained contrastive learning to maximize the token-wise similarity between visual and textual tokens. CyClip [19] imposes additional geometrical consistency on the image and text embeddings. DeCLIP [47] introduces additional positives from the nearest neighbors. While these methods improve image-text retrieval tasks on the existing benchmarks, such as ImageNet [60] and MS-COCO [50], recent studies such as VL-CheckList [83] and the Winoground Challenge [68], show that these models cannot distinguish fine-grained language details or understand structured concepts (SVLCs) such as object attributes and relations. In this paper, we focus on the latter and propose orthogonal data-driven techniques that have the potential to improve the SVLC understanding for all VL models. + +Learning Structured Representations. A full understanding of the semantics of rich visual scenes requires the ability to understand visual concepts, such as detecting individual entities and reasoning about their interactions and attributes. Structured representations have played an important role in achieving this goal, having been successfully applied to a wide range of computer vision applications: vision and language [10, 45, 46, 66], scene graphs [25, 29, 38, 56, 75], relational reasoning [4, 5], human-object interactions [16, 33, 74], action recognition [1, 2, 23, 24, 30, 52, 70], and even image & video generation from graphs [3, 22, 32]. However, most of these works rely on detailed, manually curated, supervision, often involving annotation of location information and structural details, which are very expensive to collect and scale, resulting in limited-size or synthetic data sources for training. In contrast, in our work, we focus on methods for teaching SVLC understanding to large VL models while only leveraging the available large-scale noisy VL data sources collected from the web without any use of expensive manual curation. + +Data Augmentation. Augmentation plays a key role in many computer vision applications [8, 63]. Several advanced image augmentation methods (CutMix [78], mixup [79], AutoAugment [11], RandAugment [12], etc) have been proposed and greatly improved computer vision task performance. Text augmentation has been tackled through back-translation [73], word and frame-semantic embedding augmentations [69], word replacement [81], random word insertion/deletion/swap [71], or using a text gen + +erative model [41] in diverse NLP applications. In VL tasks, previous work explores machine translation between different languages [6, 35], generating synthetic captions [43], adversarial/synthetic data augmentation for VQA [59, 67] or mixup for VL [20]. We focus on leveraging the well-understood and modeled language structure for manipulating text in a way that explicitly targets teaching SVLC semantics to VL models. To the best of our knowledge, this has not been attempted before. + +# 3. Method + +In this section, we discuss the proposed framework for improving SVLC performance of VL models using already available VL data. Sec. 3.1 and Sec. 3.2 present our main techniques for teaching SVLC to VL models. These approaches can be effectively applied both for fine-tuning existing strong VL Pre-trained models, as well as for training VL models from scratch. In both cases, they exhibit significantly improved SVLC performance as demonstrated in our experiments in Sec. 4. Sec. 3.3 presents our strategy for fine-tuning VL models on SVLC-enhanced VL data, while at the same time being parameter efficient and significantly reducing forgetting, thus, maintaining the VL model Zero-Shot (ZS) performance. Qualitative examples showing improvements attained by our proposed approach, as well as some examples of failure cases, are provided in the supplementary material. + +# 3.1. Teaching SVLC Using NLP and LLMs + +In this section, we present several ways in which the data of existing VL pre-training paired datasets can be enhanced to emphasize SVLC in the texts, and teaching them to the VL model. We propose two kinds of data enhancements - generating negative and positive text alternatives. When generating negative examples, only one word of the sentence is changed such that the semantic meaning of the sentence changes. We propose two methods for the generation of the negatives: (i) rule-based (Sec. 3.1.1); and (ii) LLM-based (Sec. 3.1.2). Positive alternatives are generated as sentences with semantically similar meanings, but different wording (Sec. 3.1.3). We then present the losses which properly take these two types of generated textual data into account during training in Sec. 3.2. + +# 3.1.1 Generating Rule-Based (RB) Negatives + +One simple yet effective method for negative text generation is using a collection of pre-defined language rules which match and replace words of a specific entity type or a pattern, such as color, material, size, etc. This method is especially useful when one has prior knowledge of a specific aspect of the language that needs to be taught. For example, if we know that our model lacks the ability to understand the colors of objects, we can easily create a rule for detecting + +and replacing color words in the text, for generating negative text that does not correspond to its paired image. To employ the generation of the rule-based negative, for each taught SVLC we define a list of words belonging to its characteristic. We then scan the VL data texts searching for the words within these lists. If a word is located, we simply replace it with a randomly selected word from the same list to generate a negative pair. For example, applying the color-rule to a sentence: "A big brown dog" can lead to "A big yellow dog". If a text has multiple candidates of words to be replaced, one of them is chosen randomly. We perform this process multiple times for the full list of SVLC characteristics of interest such as color, size, material, spatial relations, etc. These generated negative texts are SVLC specific and differ from the original text in only one word. For a detailed description and more examples of the RB negative generation, please refer to the supplementary material. + +# 3.1.2 LLM-based Negative Generation via Unmasking + +A natural extension of rule-based negatives technique is the generation of negatives using Large Language Models (LLMs) unmasking. Recent LLMs are explicitly trained in a self-supervised manner with the objective of "unmasking" parts of the text. Given a sentence with one missing word, models such as BERT [13], can suggest multiple words that fit the context of the sentence. Using this useful property of LLMs, we can therefore automatically create plausible negative examples without the need for prior knowledge of the SVLC characteristics of interest. In order to focus the randomly selected masked words to be likely to belong to SVLCs of interest, we use common NLP parsing techniques (such as spacy [26]) to parse the sentence into its components such as nouns, verbs, adjectives, adverbs, etc. We then randomly choose a type of sentence part and a word belonging to this part type, mask out the selected word and replace it with one of the options suggested by the LLM's unmasking. These negative examples, when used properly in the loss function (Eq. (3)) focus the network on the important details that affect the SVLC understanding. As we show in Sec. 4, this method is extremely useful and can significantly improve the VL model's understanding of different SVLCs. Further details and examples of the LLM negative generation are provided in the supplementary material. + +# 3.1.3 Generating Text Analogies via LLM Prompting + +While the goal of the negative text generation (Sec. 3.1.1 and Sec. 3.1.2) was to make minor perturbations to a given text such that the meaning changes, the goal here is exactly the opposite. We would like to make major changes to the text, while still keeping the same semantic meaning. For example “A woman standing left to a sitting cat” and “A cat sitting to the right of a standing woman” are two very + +different texts describing the exact same scene. One effective way to generate such semantically similar texts is by prompting the foundational LLMs. Specifically, we use the open access BLOOM [53] model. In the spirit of recently popular in-context learning [18], we present the model with a textual prompt with examples of semantically similar texts (see Fig. 2). We then append the current image caption and retrieve the BLOOMs prediction of a semantically similar text. For a detailed description, we refer the reader to the supplementary material. + +# 3.2. Losses + +All of our evaluated models (CLIP [57] and CyCLIP [19]) admit a text & image pair $(T, I)$ and are comprised of two parts: (i) image encoder $e_I = \mathcal{E}_I(I)$ ; (ii) text encoder $e_T = \mathcal{E}_T(T)$ . In this notation, the text-to-image similarity score is therefore computed as: + +$$ +\mathcal {S} (T, I) = \exp \left(\frac {\tau e _ {T} ^ {T} e _ {I}}{| | e _ {T} | | ^ {2} | | e _ {I} | | ^ {2}}\right), \tag {1} +$$ + +where $\tau$ is a learned temperature parameter. + +Contrastive Loss. As most contemporary VL models, we employ the contrastive CLIP-loss [57] as one of our losses for each batch $\mathcal{B}$ . + +$$ +\mathcal {L} _ {\text {c o n t}} = \sum_ {i} \log \left(\frac {S \left(T _ {i} , I _ {i}\right)}{\sum_ {j} S \left(T _ {i} , I _ {j}\right)}\right) + \log \left(\frac {S \left(T _ {i} , I _ {i}\right)}{\sum_ {k} S \left(T _ {k} , I _ {i}\right)}\right). \tag {2} +$$ + +Negatives Loss. In our ablation study in Sec. 5, we show that for a given text $T_{i}$ simply adding the corresponding generated negative text $T_{i}^{neg}$ to the contrastive loss $\mathcal{L}_{cont}$ is much less effective than having a separate loss individually attending to the similarity difference of $T_{i}$ and $T_{i}^{neg}$ w.r.t. the image $I_{i}$ corresponding to $T_{i}$ . We, therefore, employ the following negatives loss: + +$$ +\mathcal {L} _ {n e g} = \sum_ {i} - l o g \left(\frac {\mathcal {S} \left(T _ {i} , I _ {i}\right)}{\mathcal {S} \left(T _ {i} , I _ {i}\right) + \mathcal {S} \left(T _ {i} ^ {n e g} , I _ {i}\right)}\right). \tag {3} +$$ + +Analogy Loss. For the generated analogy texts $T_{i}^{sim}$ produced from the text $T_{i}$ which corresponds to image $I_{i}$ , we employ the combination of the two following losses: + +$$ +\mathcal {L} _ {s i m} ^ {t e x t} = \sum_ {i} - l o g \left(\frac {\mathcal {S} \left(T _ {i} ^ {s i m} , T _ {i}\right)}{\sum_ {j} \mathcal {S} \left(T _ {i} ^ {s i m} , T _ {j}\right)}\right), \tag {4} +$$ + +where with some abuse of notation, $S(T_{1}, T_{2})$ denotes the exponent cosine similarity between text $T_{1}$ and text $T_{2}$ text-embeddings, and: + +$$ +\mathcal {L} _ {s i m} ^ {i m g} = \sum_ {i} - l o g \left(\frac {\mathcal {S} \left(T _ {i} ^ {s i m} , I _ {i}\right)}{\sum_ {j} \mathcal {S} \left(T _ {i} ^ {s i m} , I _ {j}\right)}\right), \tag {5} +$$ + +which simply corresponds to the second summand of $\mathcal{L}_{cont}$ in Eq. (2), with replacing $T_{i}$ by $T_{i}^{sim}$ . + +Our final loss is, therefore: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o n t}} + \alpha \cdot \mathcal {L} _ {\text {n e g}} + \beta \cdot \left(\mathcal {L} _ {\text {s i m}} ^ {\text {t e x t}} + \mathcal {L} _ {\text {s i m}} ^ {\text {i m g}}\right). \tag {6} +$$ + +# 3.3. Fine-tuning Pre-trained VL Models + +Each of the $\mathcal{E}_T$ and $\mathcal{E}_I$ networks is comprised of a mix of non-parametric functions and two types of parametric functions: linear layers and embedding layers. Roughly, each of those functions, $\mathcal{F}_k^{lin}(x)$ and $\mathcal{F}_k^{emb}(x)$ where $k$ is layer index, is parameterized by a weight matrix $\mathcal{W}_k$ so that: + +$$ +\mathcal {F} _ {k} ^ {\text {l i n}} (x) = \mathcal {W} _ {k} \cdot x \tag {7} +$$ + +$$ +\mathcal {F} _ {k} ^ {e m b} (x) = E M B (x; \mathcal {W} _ {k}) \tag {8} +$$ + +where $EMB$ is the embedding operator assuming $x$ is a stream of integers and picking the respective columns of $\mathcal{W}_k$ . Following the idea proposed in LoRA [27] for efficient LLM fine-tuning using low-rank residual adapters, when adapting a pre-trained VL model $\mathcal{M} = (\mathcal{E}_T, \mathcal{E}_I)$ we parameterize the adapted weights $\mathcal{W}_k^*$ of the model $\mathcal{M}^*$ fine-tuned from $\mathcal{M}$ as: + +$$ +\mathcal {W} _ {k} ^ {*} = \mathcal {W} _ {k} + \mathcal {A} _ {k} \cdot \mathcal {B} _ {k} \tag {9} +$$ + +where for $\mathcal{W}_k$ of size $m\times l$ , $\mathcal{A}_k$ and $\mathcal{B}_k$ are rank- $r$ matrices of sizes $m\times r$ and $r\times l$ respectively. These low-rank residual adapters can be applied efficiently as: + +$$ +\mathcal {F} _ {k} ^ {*, l i n} (x) = \mathcal {F} _ {k} ^ {l i n} (x) + \mathcal {A} _ {k} \cdot (\mathcal {B} _ {k} \cdot x) \tag {10} +$$ + +$$ +\mathcal {F} _ {k} ^ {*}, e m b (x) = \mathcal {F} _ {k} ^ {e m b} (x) + \mathcal {A} _ {k} \cdot E M B (x; \mathcal {B} _ {k}) \tag {11} +$$ + +During the fine-tuning of $\mathcal{M}^*$ , we freeze all the base model $\mathcal{M}$ parameters $\forall k, \{\mathcal{W}_k\}$ and only the LoRA adapters $\forall k, \{(\mathcal{A}_k, \mathcal{B}_k)\}$ are being learned. In the above notation we disregard possible bias terms of the linear functions, if they are present, since we keep them frozen too. + +There are several interesting things to note about the proposed architecture: (i) as opposed to [64], who evaluated the use of efficient LLM fine-tuning techniques for VL models adaptation, we add our LoRA adapters everywhere, i.e. to all layers of both the text and image encoders, and not only to the text encoder/decoder as done in [64]; (ii) as opposed to [80] who attach a small side network only to the output of the adapted model, our LoRA adapters are added to all the parametric functions inside the model and affect all the intermediate computations; (iii) same as noted in the original [27] paper, at inference all the LoRA adapters can be folded back into the original weight matrices by simple summation, thus returning the number of total parameters to be the same as in the original model and hence have the same inference speed; (iv) with rank $r$ kept low, the number of extra parameters added by all the LoRA adapters can be very low making adaptation fast and efficient; finally, + +(v) such form of fine low-rank adaptation allows for significantly mitigating the zero-shot performance forgetting effects as demonstrated in our results and explored in the corresponding ablation (Sec. 5). + +# 4. Experiments + +# 4.1. Datasets + +We train the model using common Image-Text pair datasets, namely Conceptual Captions 3M [62] and LAION 400M [61] and test using the VL-Checklist [82] datasets, which will be described below. + +Conceptual Captions 3M (CC3M) [62] is a dataset of three million image-text pairs automatically crawled from the internet where image descriptions are harvested from Alt-text attributes and then processed and filtered to create relatively clean descriptions. + +LAION 400M [61] is a very large scale image-text pair dataset which, similarly to CC3M has been automatically harvested from the internet. One major difference between the two, apart from the size, is that LAION examples have been filtered using the pretrained CLIP model, such that the CLIP image-text similarity is high by design. Recent re-implementations of the original CLIP paper have successfully used LAION 400M to reproduce similar capabilities as the original CLIP model. + +VL-Checklist [82] combines images and annotations from the Visual Genome [39], SWiG [55], VAW [54], and HAKE [48] datasets. It is processed such that each image is annotated with two captions, one positive and one negative. The positive caption originates in its source dataset and corresponds to the image. The negative caption is constructed from the positive caption so that only one word, corresponding to the tested SVLC of interest, is changed to negate the SVLC (e.g., color, size, material, etc.). VL-Checklist evaluates three main types of VL concepts further subdivided into 7 types of SVLCs total: (1) Object: its spatial location and size, (2) Attribute: color, material, size, state, and action, and (3) Relation: spatial or action relation between two objects. In the following sections we report results on a combined VL-Checklist dataset. The results on individual comprising datasets (Visual Genome [39], SWiG [55], VAW [54], and HAKE [48]) are provided in the supplementary material. + +Zero-Shot Classification We evaluated our method on 21 classification dataset using the Zero-Shot classification protocol described in the ELEVATER Image Classification Toolkit [42]. The evaluation includes 21 different datasets, including common classification datasets such as ImageNet [60], CIFAR100 [40], EuroSat [21], and others. In Tables 1-3 we report the average results over the 21 tasks. + +
ObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CLIP [57]81.58%67.60%63.05%56.37%
CLIP +LoRA80.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41%(+0.04%)
w/o Neg Loss82.27% (+0.69%)67.58% (-0.02%)55.17% (-7.88%)55.37% (-1.00%)
Ours RB+LLM Negs85.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
+ +Table 2. Ablation study - separate Negative Losses (Sec. 3.2) vs adding negatives to contrastive loss. + +# 4.2. Implementation Details + +For CLIP we use the ML-Foundation Open-CLIP repository [28] and for CyCLIP [19] we use its original code repository, which is also based on Open-CLIP. In most experiments, unless stated otherwise, we train our model for five epochs on 4 V100 NVIDIA GPUs, with a total batch size of 512. When starting from a pre-trained model, we use rank 4 LoRA adapters (Sec. 3.3) and the learning rate is set to $5E - 6$ . When training from scratch, for CLIP we use the default parameters set in the open-CLIP library, and for CyCLIP the defaults of its original implementation. For all CLIP experiments, we use VIT/32-B as the model architecture and ResNet-50 when training CyCLIP (following [19]). When fine-tuning we initialize with the original model weights released by the respective authors. In all experiments involving a combination of CyCLIP [19] and our method (in Sec. 4.4-4.5), in addition to Eq. (6) loss we also employ all the extra losses proposed in CyCLIP [19]. + +# 4.3. Baselines + +We compare our method to two strong baselines under several configurations. The first is the CLIP [57] OpenAI pretrained model trained on 400M image-text pairs which achieves high ZS performance. The second is the very recent CyCLIP [19] method which, similarly to us, improves over the original CLIP loss. For a fair comparison all methods use the same network architecture and the same initialization from the OpenAI pretrained model. For the CyCLIP baseline we continue training from the pretrained initialization using LoRA and the CyCLIP losses. As CyCLIP losses are orthogonal to ours, we also show a unified version of the two methods (CyCLIP + Ours Combined). + +# 4.4. Fine-tuning VL Models + +In the following experiments we show the effects of fin-tuning a pre-trained VL model using our additional data enhancement methods and losses. All experiments are initialized from the official OpenAI CLIP. We compare our method to two baselines on the VL-Checklist tasks, and on the Zero-Shot image classification task. The first baseline is the original pretrained model without any further training. The second, is the same model with the additional LoRA parameters, trained on CC3M (Table 1a) and LAION (Table 1g) using the original CLIP loss function. Table 1a + +![](images/f6d2f5a157ac665955aa3d39ef8f23d02b2101b355f705b7fd77ac3772737f17.jpg) +Figure 3. CC3M fine-tuning - detailed. Detailed results of the baselines (CLIP, CLIP+LoRA) and our models (RB+LLM Neg, RB+LLM Negs+Pos) initialized from OpenAI pretrained CLIP ViT-32B and fine-tuned (except CLIP) on CC3M. + +shows several configurations of our method compared to the baselines. We see that our method shows significant improvements on all VL-Checklist tasks reaching up to $15\%$ improvement. Figure 3 displays the relative gains on all tasks of the "Attribute" and "Relation" tests. It is clear that our gains are across all tests. These improvements come at a price of some minor degradation to the Zero-Shot performance compared to CLIP. Moreover, when fine-tuning CLIP on the LAION dataset (Table 1g) we do not see these degradations. In Tab. 1b we show consistent gains to ones in Tab. 1a when finetuning with a stronger (higher ZS) VitB/16 CLIP [57] pre-trained image encoder, observing an even lower drop in ZS performance. In Tab. 1e we show significant gains in fine-tuning CyCLIP on CC3M, comparing fine-tuning before and after integrating our proposed approach, this also comes at no ZS performance cost. + +# 4.5. Training from scratch + +In these experiments, we show the advantage of integrating our approach throughout the whole training procedure. To this end, we train from scratch on CC3M - both CLIP and CyCLIP, either on their own, or together with our proposed method integrated. Tables - Tab. 1c, Tab. 1d, and Tab. 1f show that, similar to finetining, our method significantly improves both CLIP and CyCLIP SVLC capabilities. + +
MethodColorMaterialAttribute SizeStateActionRelation21 Zero-Shot Tasks Average
ActionSpatial
Ours Pos72.35%69.25%69.80%59.35%66.08%70.97%39.20%55.37%
Ours RB Neg78.45%83.20%69.50%65.95%69.66%75.97%74.70%54.66%
Ours LLM Neg76.00%79.70%72.75%61.35%68.36%77.23%72.40%55.05%
Ours RB+LLM Negs79.25%84.25%72.15%64.05%69.82%79.03%78.40%54.66%%
Ours Combined77.45%77.35%73.35%62.30%69.39%74.70%63.20%54.77%%
+ +Table 3. Ablation study - component analysis. Detailed evaluation on Attribute & Relation SVLCs. + +ties, while still keeping similar ZS performance. Figure 4 shows the detailed parsing of the "Attribute" and "Relation" tests with consistent gains across the specific tasks. These findings suggest that had we trained our method for many epochs on a large dataset, such as the full LAION-9B, we would reach similar ZS performance while greatly enhancing the SVLC performance of the model. + +# 5. Ablations + +In this section we will examine several aspects and components of our proposed methods. + +# 5.1. Negative Losses + +Section 3.2 details our additional loss functions which utilize our generated texts. Specifically, Eq. (3) describes the loss which contrasts a given example only with its generated negatives. Table 2 clearly shows the importance of this loss as opposed to simply adding the negative examples to the original contrastive loss (Eq. (2)). Without explicitly forcing the network to attend to the small changes in the text (Eq. (3)), the generated negative examples, when inserted only to the contrastive loss (Eq. (2)) do not provide the desired gains over the baseline. + +# 5.2. Component analysis + +In Section 3 we describe several components of our proposed method. Specifically we present our RB negative generation (Sec. 3.1.1), our LLM-based negative genera + +![](images/de41b8340c047d239b087b68f831ecbe3f93a12b553ad11bbe5d23118219e5ea.jpg) +Figure 4. CC3M train from scratch - detailed. Detailed results of the baseline CLIP and our models (RB+LLM Neg, RB+LLM Negs+Pos), all CLIP ViT-32B, trained from scratch on CC3M. + +tion (Sec. 3.1.2), and our LLM-based analogy generation, referred to here as "Pos" (Sec. 3.1.3). Tab. 3 provides a detailed analysis and comparison of the contribution of each component to the final result. Through this analysis we see two contradicting forces between the SVLC capabilities and the original Zero-Shot performance. We can see that each negative generation method plays a crucial role in some of the tasks while the use of both is usually the best performing option. On the other hand, the LLM-based analogy generation stabilizes the Zero-Shot performance and mitigates the drop with respect to the baseline. The joint version of all components ("Ours Combined") exhibits a good trade-off between the two contradicting forces. + +# 6. Conclusions + +We have presented a data-driven technique for enhancing the performance of VL models in the important task of SVLC understanding without sacrificing their impressive ZS object recognition capabilities. Our proposed method attains significant gains in multiple experiments on a variety of base VL models and datasets. It builds upon the modeling strength and knowledge of language structure to teach this structure to VL models in an orthogonal way, suggesting wide applicability to existing or future VL models. + +While attaining impressive gains in SVLC understanding, we believe the small drop observed in ZS performance could be further reduced. An additional possible extension of our work is using more sophisticated sequence generation techniques for improving our batch data. One may combine annotation efforts with a language model to get improved data for training and evaluation [51]. Another possibility is adding a corrector [72] in the training that validates whether the VL model learns the correct concepts or not. We leave all of these exciting directions to future work. + +# Acknowledgements + +We would like to thank Donghyun Kim for his invaluable help in this work. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. FA8750-19-C-1001. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA). This material is based upon work supported by the 'Data Science grant from the Israeli Council of Higher Education'. This research or RG was supported by ERC-StG SPADE grant no. 75749. + +# References + +[1] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 4 +[2] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 4 +[3] Amir Bar, Roei Herzig, Xiaolong Wang, Anna Rohrbach, Gal Chechik, Trevor Darrell, and A. Globerson. Compositional video synthesis with action graphs. In ICML, 2021. 4 +[4] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, pages 105-121, 2018. 4 +[5] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.4 +[6] Andrea Burns, Donghyun Kim, Derry Wijaya, Kate Saenko, and Bryan A Plummer. Learning to scale multilingual representations for vision-language tasks. In European Conference on Computer Vision, pages 197-213. Springer, 2020. 4 +[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 1 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4 +[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 3 +[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 4 +[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 4 +[12] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 4 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional + +transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[14] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't stop learning: Towards continual learning for the clip model. arXiv:2207.09248, 2022. 1 +[15] Andreas Fürst, Elisabeth Rumetshofer, Viet Tran, Hubert Ramsauer, Fei Tang, Johannes Lehner, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto-Nemling, et al. Cloob: Modern hopfield networks with infoloob outperform clip. arXiv preprint arXiv:2110.11316, 2021. 4 +[16] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. ArXiv, abs/2008.11714, 2020. 4 +[17] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022. 4 +[18] Shivam Garg, Dimitris Tsipras, Percy Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes, 2022. 5 +[19] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022. 1, 2, 3, 4, 5, 7 +[20] Xiaoshuai Hao, Yi Zhu, Srikar Appalaraju, Aston Zhang, Wanqian Zhang, Bo Li, and Mu Li. Mixgen: A new multimodal data augmentation. arXiv preprint arXiv:2206.08358, 2022.4 +[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6 +[22] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, 2020. 4 +[23] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4 +[24] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0–0, 2019. 4 +[25] Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, and Amir Globerson. Mapping images to scene graphs with permutation-invariant structured prediction. In Advances in Neural Information Processing Systems (NIPS), 2018. 4 +[26] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 5 +[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. + +Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.2,3,6 +[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 7 +[29] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 4 +[30] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as composition of spatiotemporal scene graphs. arXiv preprint arXiv:1912.06992, 2019.4 +[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 3, 4 +[32] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 4 +[33] Keizo Kato, Yin Li, and Abhinav Gupta. Compositional learning for human object interaction. In ECCV, 2018. 4 +[34] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 1 +[35] Donghyun Kim, Kuniaki Saito, Kate Saenko, Stan Sclaroff, and Bryan Plummer. Mule: Multimodal universal language embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11254-11261, 2020. 4 +[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 4 +[37] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 1 +[38] Ranjay Krishna, Ines Chami, Michael S. Bernstein, and Li Fei-Fei. Referring relationships. *ECCV*, 2018. 4 +[39] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2, 6 +[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 + +[41] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 4 +[42] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. Neural Information Processing Systems, 2022. 6 +[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 1, 4 +[44] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1 +[45] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. ArXiv, abs/1908.03557, 2019. 4 +[46] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 3, 4 +[47] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1, 4 +[48] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Mingyang Chen, Ze Ma, Shiyi Wang, Hao-Shu Fang, and Cewu Lu. Hake: Human activity knowledge engine. arXiv preprint arXiv:1904.06539, 2019. 2, 6 +[49] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1 +[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4 +[51] Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. Wanli: Worker and ai collaboration for natural language inference dataset creation, Jan. 2022. 8 +[52] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4 +[53] Margaret Mitchell, Giada Pistilli, Yacine Jernite, Ezinwanne Ozoani, Marissa Gerchick, Nazneen Rajani, Sasha Luccioni, Irene Solaiman, Maraim Masoud, Somaieh Nikpoor, Munoz Ferrandis Carlos, Stas Bekman, Christopher Akiki, Danish Contractor, David Lansky, Angelina McMillan-Major, Tristan Thrush, Suzana Ilic, Gerard Dupont, Shayne Longpre, + +Manan Dey, Stella Biderman, Douwe Kiela, Emi Baylor, Teven Le Scao, Aaron Gokaslan, Julien Launay, and Niklas Muennighoff. Bigscience, bigscience language open-science open-access multilingual (bloom) language model. International, May 2021-May 2022. 2, 5 +[54] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13018-13028, 2021. 2, 6 +[55] Sarah Pratt, Mark Yatskar, Luca Weihs, Ali Farhadi, and Aniruddha Kembhavi. Grounded situation recognition. In European Conference on Computer Vision, pages 314-332. Springer, 2020. 2, 6 +[56] Moshiko Raboh, Roei Herzig, Gal Chechik, Jonathan Berant, and Amir Globerson. Differentiable scene graphs. In WACV, 2020. 4 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 7 +[58] Vinay Venkatesh Ramasesh, Ethan Dyer, and Maithra Raghu. Anatomy of catastrophic forgetting: Hidden representations and task semantics. In International Conference on Learning Representations, 2021. 1 +[59] Arijit Ray, Karan Sikka, Ajay Divakaran, Stefan Lee, and Giedrius Burachas. Sunny and dark outside?! improving answer consistency in vqa through entailed question generation. arXiv preprint arXiv:1909.04696, 2019. 4 +[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 6 +[61] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 6 +[62] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2, 6 +[63] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020.4 +[64] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 6 + +[65] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3 +[66] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 4 +[67] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In European Conference on Computer Vision, pages 437-453. Springer, 2020. 4 +[68] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2, 4 +[69] William Yang Wang and Diyi Yang. That's so annoying!!! A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using# petpeeve tweets. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 2557–2563, 2015. 4 +[70] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 4 +[71] Jason Wei and Kai Zou. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196, 2019. 4 +[72] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generating sequences by learning to self-correct, 2022. 8 +[73] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 4 +[74] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and M. Kankanhalli. Learning to detect human-object interactions with knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2019. 4 +[75] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene Graph Generation by Iterative Message Passing. In CVPR, pages 3097–3106, 2017. 4 +[76] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 4 +[77] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 4 +[78] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4 + +[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4 +[80] Jeffrey O Zhang, Alexander Sax, Amir Zamir, Leonidas Guibas, and Jitendra Malik. Side-tuning: a baseline for network adaptation via additive side networks. In European Conference on Computer Vision, pages 698-714. Springer, 2020. 6 +[81] Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015. 4 +[82] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. V1-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6 +[83] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022.4 \ No newline at end of file diff --git a/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/images.zip b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c7b29aadd3229e7fa1a1cfc11a746e8bd493a4d --- /dev/null +++ b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a32e3b807b7d275f65076cbcedd53764c9c08178651df9bd07582549b1db7c14 +size 555767 diff --git a/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/layout.json b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2f156aa7cc45e6a5a48773ed3b2fe5d315618ff3 --- /dev/null +++ b/2023/Teaching Structured Vision & Language Concepts to Vision & Language Models/layout.json @@ -0,0 +1,8588 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "text", + "content": "Teaching Structured Vision & Language Concepts to Vision & Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": "Sivan Doveh" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Assaf Arbelle" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Sivan Harary" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Eli Schwartz" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Roei Herzig" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Raja Giryes" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Rogerio Feris" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Rameswar Panda" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Shimon Ullman" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{*2}" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "text", + "content": ", Leonid Karlinsky" + }, + { + "bbox": [ + 86, + 142, + 515, + 174 + ], + "type": "inline_equation", + "content": "^{*4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 186, + 541, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 186, + 541, + 199 + ], + "spans": [ + { + "bbox": [ + 52, + 186, + 541, + 199 + ], + "type": "text", + "content": "1IBM Research, 2Weizmann Institute of Science, 3Tel-Aviv University, 4MIT-IBM Watson AI Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 251, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 574 + ], + "type": "text", + "content": "Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to " + }, + { + "bbox": [ + 46, + 251, + 290, + 574 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 251, + 290, + 574 + ], + "type": "text", + "content": " in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 287, + 696 + ], + "type": "text", + "content": "Recent Vision & Language (VL) models [19, 31, 43, 44, 47, 57] achieve excellent zero-shot performance with respect to various computer-vision tasks such as detection, classification, segmentation, etc. However, recent studies [68, 82] have demonstrated that even the strongest VL models struggle with the compositional understanding of some basic Structured VL Concepts (SVLC) such as ob-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 341, + 225, + 508, + 441 + ], + "blocks": [ + { + "bbox": [ + 341, + 225, + 508, + 441 + ], + "lines": [ + { + "bbox": [ + 341, + 225, + 508, + 441 + ], + "spans": [ + { + "bbox": [ + 341, + 225, + 508, + 441 + ], + "type": "image", + "image_path": "130aaaae7f82b9506eb17458c9f4fa69eccce7e95d52727c6908bd568f38a1ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 452, + 547, + 519 + ], + "lines": [ + { + "bbox": [ + 304, + 452, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 547, + 519 + ], + "type": "text", + "content": "Figure 1. Teaching language structure to VL models. (a) Standard contrastive text-to-image loss (e.g. CLIP [57]) tends to under-emphasize SVLC content of the text, likely due to the random nature of the training batches; (b) We generate modified versions of corresponding texts and use them to add losses to explicitly teach language structure (SVLC) to VL models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 605 + ], + "type": "text", + "content": "ject attributes, inter-object relations, transitive actions, object states and more. Collecting specialized large scale data to teach VL models these missing 'skills' is impractical, as finding specialized text-image pairs for each kind and possible value of the different attributes, relations, or states, is both difficult and expensive." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Another important challenge in training VL models with new concepts is catastrophic forgetting, which is a common property to all neural models [7, 34, 37, 49, 58] and has been explored for VL models in a recent concurrent work [14]. Large VL models such as CLIP [57] and CyCLIP [19] have exhibited excellent zero-shot learning abilities in many tasks. Therefore, even given a large dataset with new concepts, it is important not to lose these abilities when performing the adaptation to the new data." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 124, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 124, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 124, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2657" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 545, + 270 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 545, + 270 + ], + "type": "image", + "image_path": "4eb3b49556f4bacdf3f366b210e8a3a0c60d28259ceb52609f8cf3fbecf243a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 313 + ], + "type": "text", + "content": "Figure 2. Teaching structured image understanding to VL models via structured textual data manipulation harnessing the power of language modeling. (1) Generating Rule-Based negative texts (Sec. 3.1.1); (2) Generating negatives using Large Language Model (LLM) unmasking (Sec. 3.1.2); (3) Generating analogies (positives) via LLM prompting (Sec. 3.1.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 325, + 288, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 432 + ], + "type": "text", + "content": "In this paper, we propose a way to leverage existing (off-the-shelf) VL pre-training data sources in order to improve the SVLC understanding skills of a given model, while at the same time maintaining its zero-shot object recognition accuracy. Naturally, succeeding in this goal would lead to potential improvement w.r.t. SVLC understanding in a wide variety of downstream tasks building upon pre-trained VL models, such as zeros-shot detection, segmentation, image generation, and many more." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 604 + ], + "type": "text", + "content": "Recent research [68, 82] has shown that VL models exhibit an 'object bias' partially due to the contrastive text-to-image loss used in their pre-training. For example, the popular CLIP-loss [57] is computed over a random batch of text-image pairs sampled from a large-scale and diverse VL dataset with the chance of two images in the same batch containing the same set of objects being very low. For such a loss, representing just a 'bag of objects' in each image or text is sufficient for matching the corresponding pairs. Intuitively, this leads to the 'object bias' where SVLCs like attributes, states, and relations are being underrepresented (e.g. having a much smaller amplitude in the resulting feature superposition), consequently causing the aforementioned issues with SVLC understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Based on this intuition, we propose a simple data-driven technique that harnesses existing language parsing and modeling capabilities to enhance the importance of SVLCs in the VL model training losses. For each text in the training batch, we automatically generate alternative negative or positive text by manipulating its content to be opposite or equivalent to the original text. Using the newly generated texts, we explicitly teach SVLC to the model via additional losses (see Fig. 1) that enforce differentiating be" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 325, + 545, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 349 + ], + "type": "text", + "content": "tween different (original and generated) SVLC texts and are no longer satisfiable by the 'bag of objects' representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 350, + 546, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 546, + 446 + ], + "type": "text", + "content": "Towards this end, we propose several techniques for implementing this approach, including (i) rule-based priors based on classical NLP parsing and word substitution vocabulary according to attribute/relation type; (ii) prompting a Large Language Model (LLM) (e.g. [53]) for analogous text; (iii) generating different meaning (negative) alternatives by LLM-based unmasking of parsed text entities of different kinds; (iv) combinations of these methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 449, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 640 + ], + "type": "text", + "content": "We demonstrate that all these techniques can lead to significant improvements of up to " + }, + { + "bbox": [ + 304, + 449, + 546, + 640 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 449, + 546, + 640 + ], + "type": "text", + "content": " percent when measuring the VL models' SVLC understanding. We verify this on 5 datasets: VG [39], HAKE [48], VAW [54], SWIG [55], and all combined, using the protocol recently proposed in VL-Checklist [82]. In addition, we show that the resulting VL models largely preserve their zero-shot object recognition performance. For the latter, we also propose a variant of efficient LLM fine-tuning using low-rank residual adapters (LoRA) [27] adjusted to VL models. Finally, we show that our framework allows better harnessing of the standard available VL data, e.g. CC3M [62] and LAION [61]. This is exhibited by the aforementioned gains, both in new VL models trained from scratch, as well as in models fine-tuned using strong available VL models such as CLIP [57] and CyCLIP [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "To summarize, we offer the following contributions: (i) We propose a data-driven approach for better harnessing the standard available VL data to improve VL models' SVLC understanding skills, such as understanding object attributes, inter-object relations, transitive actions, object states, and more, without sacrificing zero-shot object recog" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2658" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 555, + 404 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 555, + 404 + ], + "type": "table", + "html": "
DatasetModelPreArchObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CC3MCLIP [57]✓Vit-B/3281.58%67.60%63.05%56.37%(a)
CLIP + LoRA✓Vit-B/3280.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41% (+0.04%)
CLIP + Ours RB Neg✓Vit-B/3283.89% (+2.30%)73.35% (+5.75%)75.33% (+12.28%)54.32% (-2.05%)
CLIP + Ours LLM Neg✓Vit-B/3284.44% (+2.85%)71.63% (+4.03%)74.82% (+11.77%)55.60% (-0.77%)
CLIP + Ours RB+LLM Negs✓Vit-B/3285.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
CLIP + Ours Combined✓Vit-B/3285.00% (+3.42%)71.97% (+4.37%)68.95% (+5.90%)54.77% (-1.60%)
CLIP [57]✓Vit-B/1682.91%67.32%61.80%60.00%(b)
CLIP + Ours RB+LLM Negs✓Vit-B/1685.82% (+2.91%)73.92% (+6.6%)77.40% (+15.6%)59.37% (-0.63%)
CLIP + Ours Combined✓Vit-B/1684.75% (+1.84%)71.18% (+3.86%)69.68% (+7.88%)59.87% (-0.13%)
CLIPXVit-B/3271.17%57.86%45.20%21.96%(c)
CLIP + Ours CombinedXVit-B/3271.79% (+0.62%)63.29% (+5.43%)58.13% (+12.93%)20.96% (-1.00%)
CLIP [57]XVit-B/1664.01%54.27%41.57%15.49%(d)
CLIP + Ours RB+LLM NegsXVit-B/1673.11% (+9.1%)65.32% (+11.05)71.93% (+30.36)20.78% (+5.29%)
CLIP + Ours CombinedXVit-B/1672.99% (+8.98%)63.01% (+8.74%)62.95% (+21.38)20.61% (+5.12%)
CyCLIP [19]✓R5073.49%59.33%53.83%26.00%(e)
CyCLIP + LoRA✓R5073.30% (-0.19%)58.89% (-0.44%)53.03% (-0.80%)26.30% (+0.30%)
CyCLIP + Ours Combined✓R5074.20% (+0.71%)63.52% (+4.20%)59.47% (+5.63%)26.31% (+0.31%)
CyCLIPXR5069.41%57.59%53.70%21.02%(f)
CyCLIP + Ours CombinedXR5071.50% (+2.09%)65.69% (+8.10%)70.20% (+16.50%)20.44% (-0.42%)
LAIONCLIP [57]✓Vit-B/3281.58%67.6%63.05%56.37%(g)
CLIP + LoRA✓Vit-B/3282.18% (+0.60%)68.48% (+0.88%)62.72% (-0.33%)57.15% (+0.78%)
CLIP + Ours Combined✓Vit-B/3282.54% (+0.96%)69.64% (+2.04%)66.05% (+3.00%)56.71% (+0.34%)
", + "image_path": "0bbf2336bb1f88532cc8e4a37a2f30a75e36171cde93bb058538854057a3550b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": "Table 1. VL-Checklist and Zero-Shot classification evaluation (ImageNet + 20 datasets). Finetuned models (for 5 epochs as detailed in Sec. 3.3, starting from officially released CLIP [57] and CyCLIP [19] weights) are marked with " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": " in the Pre-trained (Pre) column, while models trained from scratch (for 10 epochs) are marked with " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": ". The gains and losses of our approach (+Ours) are in color and are computed w.r.t. to corresponding baselines in each section. CLIP/CyCLIP + LoRA indicate finetuning in the same way and on the same data, but without our approach. Finetuning without LoRA on the same data yields significantly worse performance in all metrics. Sections are separated by bold horizontal lines. (a) CC3M fine-tuning - CLIP - Vit-B/32: we significantly improve the SVLC understanding, observing only small ZS performance drops, " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "0.77\\%" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": "; (b) CC3M fine-tuning - CLIP - Vit-B/16: we can observe similar improvements as in (a), with an even smaller impact on ZS performance; (c) CC3M from scratch - CLIP - Vit-B/32: we significantly improve SVLC understanding with only a small (" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": ") decrease in ZS performance; (d) CC3M from scratch - CLIP - Vit-B/16: compared to (c), even greater SVLC understanding improvement is observed (up to " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "30.36\\%" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": "), at no cost in ZS performance, and even improvement of over " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": "; (e) CC3M fine-tuning - CyCLIP: we use CyCLIP original code (with LoRA integration) and losses, as can be seen - adding our techniques improves CyCLIP SVLC performance considerably without sacrificing ZS performance; (f) CC3M from scratch - CyCLIP: observing even largergains in SVLC understanding compared to (e) (up to " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": "), with small reduction in ZS performance of " + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "inline_equation", + "content": "0.42\\%" + }, + { + "bbox": [ + 46, + 413, + 547, + 568 + ], + "type": "text", + "content": "; (g) LAION fine-tuning - CLIP - Vit-B/32: we improve SVLC understanding without any decrease in ZS performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 578, + 289, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 289, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 289, + 711 + ], + "type": "text", + "content": "nition performance; (ii) More specifically, we propose to leverage the well-understood and well-modeled structure of language, through classical NLP parsing and/or use of the modern pre-trained LLMs, for manipulating the text part of the standard VL paired datasets to regularize VL training and teach SVLC understanding to VL models. (iii) We further propose an adaptation of efficient LLM fine-tuning technique of [27] for fine-tuning VL models, allowing for only minimal reduction in zero-shot object recognition performance after fine-tuning, while still obtaining the aforementioned SVLC understanding gains. (iv) Empirically, for" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "type": "text", + "content": "the popular CLIP [57] and its most recent extension CyCLIP [19], we demonstrate SVLC understanding average improvements of up to " + }, + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "type": "text", + "content": " when training from scratch, and " + }, + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 628 + ], + "type": "text", + "content": " when fine-tuning from a pretrained model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 634, + 392, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 392, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 392, + 647 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Vision-language (VL) Models. (e.g., CLIP [57] and ALIGN [31]) show significant advances in diverse zero-shot downstream tasks. They are pre-trained using contrastive image-text alignment on a large-scale noisy dataset of text-image pairs collected from the web. Several methods [9,46," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2659" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "[65] additionally employ off-the-shelf object detectors to extract region features. In order to relax this limitation, some methods [31,36,43,76] propose to use cross-attention layers with self-supervised learning objectives including image-text matching and masked/autoregressive language modeling. BLIP [43] generates synthetic captions from the language modeling head and filters noisy captions based on the image-text matching score. Recently, there have been attempts to learn finer-level alignment and relations between image and text [15, 17, 19, 47, 77]. FILIP proposes fine-grained contrastive learning to maximize the token-wise similarity between visual and textual tokens. CyClip [19] imposes additional geometrical consistency on the image and text embeddings. DeCLIP [47] introduces additional positives from the nearest neighbors. While these methods improve image-text retrieval tasks on the existing benchmarks, such as ImageNet [60] and MS-COCO [50], recent studies such as VL-CheckList [83] and the Winoground Challenge [68], show that these models cannot distinguish fine-grained language details or understand structured concepts (SVLCs) such as object attributes and relations. In this paper, we focus on the latter and propose orthogonal data-driven techniques that have the potential to improve the SVLC understanding for all VL models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 362, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 289, + 602 + ], + "type": "text", + "content": "Learning Structured Representations. A full understanding of the semantics of rich visual scenes requires the ability to understand visual concepts, such as detecting individual entities and reasoning about their interactions and attributes. Structured representations have played an important role in achieving this goal, having been successfully applied to a wide range of computer vision applications: vision and language [10, 45, 46, 66], scene graphs [25, 29, 38, 56, 75], relational reasoning [4, 5], human-object interactions [16, 33, 74], action recognition [1, 2, 23, 24, 30, 52, 70], and even image & video generation from graphs [3, 22, 32]. However, most of these works rely on detailed, manually curated, supervision, often involving annotation of location information and structural details, which are very expensive to collect and scale, resulting in limited-size or synthetic data sources for training. In contrast, in our work, we focus on methods for teaching SVLC understanding to large VL models while only leveraging the available large-scale noisy VL data sources collected from the web without any use of expensive manual curation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "type": "text", + "content": "Data Augmentation. Augmentation plays a key role in many computer vision applications [8, 63]. Several advanced image augmentation methods (CutMix [78], mixup [79], AutoAugment [11], RandAugment [12], etc) have been proposed and greatly improved computer vision task performance. Text augmentation has been tackled through back-translation [73], word and frame-semantic embedding augmentations [69], word replacement [81], random word insertion/deletion/swap [71], or using a text gen" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "erative model [41] in diverse NLP applications. In VL tasks, previous work explores machine translation between different languages [6, 35], generating synthetic captions [43], adversarial/synthetic data augmentation for VQA [59, 67] or mixup for VL [20]. We focus on leveraging the well-understood and modeled language structure for manipulating text in a way that explicitly targets teaching SVLC semantics to VL models. To the best of our knowledge, this has not been attempted before." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 188, + 362, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 188, + 362, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 362, + 201 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 207, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 547, + 399 + ], + "type": "text", + "content": "In this section, we discuss the proposed framework for improving SVLC performance of VL models using already available VL data. Sec. 3.1 and Sec. 3.2 present our main techniques for teaching SVLC to VL models. These approaches can be effectively applied both for fine-tuning existing strong VL Pre-trained models, as well as for training VL models from scratch. In both cases, they exhibit significantly improved SVLC performance as demonstrated in our experiments in Sec. 4. Sec. 3.3 presents our strategy for fine-tuning VL models on SVLC-enhanced VL data, while at the same time being parameter efficient and significantly reducing forgetting, thus, maintaining the VL model Zero-Shot (ZS) performance. Qualitative examples showing improvements attained by our proposed approach, as well as some examples of failure cases, are provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 403, + 509, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 509, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 509, + 415 + ], + "type": "text", + "content": "3.1. Teaching SVLC Using NLP and LLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 420, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 588 + ], + "type": "text", + "content": "In this section, we present several ways in which the data of existing VL pre-training paired datasets can be enhanced to emphasize SVLC in the texts, and teaching them to the VL model. We propose two kinds of data enhancements - generating negative and positive text alternatives. When generating negative examples, only one word of the sentence is changed such that the semantic meaning of the sentence changes. We propose two methods for the generation of the negatives: (i) rule-based (Sec. 3.1.1); and (ii) LLM-based (Sec. 3.1.2). Positive alternatives are generated as sentences with semantically similar meanings, but different wording (Sec. 3.1.3). We then present the losses which properly take these two types of generated textual data into account during training in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 601, + 504, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 504, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 504, + 613 + ], + "type": "text", + "content": "3.1.1 Generating Rule-Based (RB) Negatives" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "One simple yet effective method for negative text generation is using a collection of pre-defined language rules which match and replace words of a specific entity type or a pattern, such as color, material, size, etc. This method is especially useful when one has prior knowledge of a specific aspect of the language that needs to be taught. For example, if we know that our model lacks the ability to understand the colors of objects, we can easily create a rule for detecting" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2660" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "and replacing color words in the text, for generating negative text that does not correspond to its paired image. To employ the generation of the rule-based negative, for each taught SVLC we define a list of words belonging to its characteristic. We then scan the VL data texts searching for the words within these lists. If a word is located, we simply replace it with a randomly selected word from the same list to generate a negative pair. For example, applying the color-rule to a sentence: \"A big brown dog\" can lead to \"A big yellow dog\". If a text has multiple candidates of words to be replaced, one of them is chosen randomly. We perform this process multiple times for the full list of SVLC characteristics of interest such as color, size, material, spatial relations, etc. These generated negative texts are SVLC specific and differ from the original text in only one word. For a detailed description and more examples of the RB negative generation, please refer to the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 291, + 288, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 288, + 304 + ], + "type": "text", + "content": "3.1.2 LLM-based Negative Generation via Unmasking" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 308, + 289, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 597 + ], + "type": "text", + "content": "A natural extension of rule-based negatives technique is the generation of negatives using Large Language Models (LLMs) unmasking. Recent LLMs are explicitly trained in a self-supervised manner with the objective of \"unmasking\" parts of the text. Given a sentence with one missing word, models such as BERT [13], can suggest multiple words that fit the context of the sentence. Using this useful property of LLMs, we can therefore automatically create plausible negative examples without the need for prior knowledge of the SVLC characteristics of interest. In order to focus the randomly selected masked words to be likely to belong to SVLCs of interest, we use common NLP parsing techniques (such as spacy [26]) to parse the sentence into its components such as nouns, verbs, adjectives, adverbs, etc. We then randomly choose a type of sentence part and a word belonging to this part type, mask out the selected word and replace it with one of the options suggested by the LLM's unmasking. These negative examples, when used properly in the loss function (Eq. (3)) focus the network on the important details that affect the SVLC understanding. As we show in Sec. 4, this method is extremely useful and can significantly improve the VL model's understanding of different SVLCs. Further details and examples of the LLM negative generation are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 611, + 284, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 284, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 284, + 624 + ], + "type": "text", + "content": "3.1.3 Generating Text Analogies via LLM Prompting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "While the goal of the negative text generation (Sec. 3.1.1 and Sec. 3.1.2) was to make minor perturbations to a given text such that the meaning changes, the goal here is exactly the opposite. We would like to make major changes to the text, while still keeping the same semantic meaning. For example “A woman standing left to a sitting cat” and “A cat sitting to the right of a standing woman” are two very" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "different texts describing the exact same scene. One effective way to generate such semantically similar texts is by prompting the foundational LLMs. Specifically, we use the open access BLOOM [53] model. In the spirit of recently popular in-context learning [18], we present the model with a textual prompt with examples of semantically similar texts (see Fig. 2). We then append the current image caption and retrieve the BLOOMs prediction of a semantically similar text. For a detailed description, we refer the reader to the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 199, + 359, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 359, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 359, + 210 + ], + "type": "text", + "content": "3.2. Losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "text", + "content": "All of our evaluated models (CLIP [57] and CyCLIP [19]) admit a text & image pair " + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "inline_equation", + "content": "(T, I)" + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "text", + "content": " and are comprised of two parts: (i) image encoder " + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "inline_equation", + "content": "e_I = \\mathcal{E}_I(I)" + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "text", + "content": "; (ii) text encoder " + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "inline_equation", + "content": "e_T = \\mathcal{E}_T(T)" + }, + { + "bbox": [ + 305, + 216, + 547, + 277 + ], + "type": "text", + "content": ". In this notation, the text-to-image similarity score is therefore computed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 357, + 285, + 545, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 285, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 357, + 285, + 545, + 313 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (T, I) = \\exp \\left(\\frac {\\tau e _ {T} ^ {T} e _ {I}}{| | e _ {T} | | ^ {2} | | e _ {I} | | ^ {2}}\\right), \\tag {1}", + "image_path": "a94c85ac6774d12e80c39e06528c3c6e8ea04398a295a7d3a22fa327cf0ba391.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 321, + 485, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 485, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 485, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 321, + 485, + 332 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 321, + 485, + 332 + ], + "type": "text", + "content": " is a learned temperature parameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 333, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 368 + ], + "type": "text", + "content": "Contrastive Loss. As most contemporary VL models, we employ the contrastive CLIP-loss [57] as one of our losses for each batch " + }, + { + "bbox": [ + 304, + 333, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 333, + 545, + 368 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 376, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t}} = \\sum_ {i} \\log \\left(\\frac {S \\left(T _ {i} , I _ {i}\\right)}{\\sum_ {j} S \\left(T _ {i} , I _ {j}\\right)}\\right) + \\log \\left(\\frac {S \\left(T _ {i} , I _ {i}\\right)}{\\sum_ {k} S \\left(T _ {k} , I _ {i}\\right)}\\right). \\tag {2}", + "image_path": "49afdb8bc18684caaea2f42cb2686b39fc2bfab34b21672ebc7c778748598183.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": "Negatives Loss. In our ablation study in Sec. 5, we show that for a given text " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": " simply adding the corresponding generated negative text " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "T_{i}^{neg}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": " to the contrastive loss " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cont}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": " is much less effective than having a separate loss individually attending to the similarity difference of " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "T_{i}^{neg}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": " w.r.t. the image " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 418, + 547, + 503 + ], + "type": "text", + "content": ". We, therefore, employ the following negatives loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 511, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 511, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 329, + 511, + 545, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n e g} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} , I _ {i}\\right)}{\\mathcal {S} \\left(T _ {i} , I _ {i}\\right) + \\mathcal {S} \\left(T _ {i} ^ {n e g} , I _ {i}\\right)}\\right). \\tag {3}", + "image_path": "729c83a0f7de2845dbca8adcbd80edef16370743fdcdfeadc23e9359483ce702.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "text", + "content": "Analogy Loss. For the generated analogy texts " + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "inline_equation", + "content": "T_{i}^{sim}" + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "text", + "content": " produced from the text " + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "text", + "content": " which corresponds to image " + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 552, + 545, + 589 + ], + "type": "text", + "content": ", we employ the combination of the two following losses:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 597, + 545, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 597, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 343, + 597, + 545, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i m} ^ {t e x t} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} ^ {s i m} , T _ {i}\\right)}{\\sum_ {j} \\mathcal {S} \\left(T _ {i} ^ {s i m} , T _ {j}\\right)}\\right), \\tag {4}", + "image_path": "a98e7292d6a8c57eda03eb8a1e60ed5b8432a75183832bc83c2f1d64dfff7625.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "text", + "content": "where with some abuse of notation, " + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "inline_equation", + "content": "S(T_{1}, T_{2})" + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "text", + "content": " denotes the exponent cosine similarity between text " + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 304, + 638, + 545, + 674 + ], + "type": "text", + "content": " text-embeddings, and:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 684, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 684, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 344, + 684, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i m} ^ {i m g} = \\sum_ {i} - l o g \\left(\\frac {\\mathcal {S} \\left(T _ {i} ^ {s i m} , I _ {i}\\right)}{\\sum_ {j} \\mathcal {S} \\left(T _ {i} ^ {s i m} , I _ {j}\\right)}\\right), \\tag {5}", + "image_path": "f6ebf88683cc858fd6a9ba9facbc0149f4c0e951c4deca7bf3179c9f014c46a3.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2661" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "which simply corresponds to the second summand of " + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cont}" + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": " in Eq. (2), with replacing " + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "T_{i}^{sim}" + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 96, + 169, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 96, + 169, + 107 + ], + "spans": [ + { + "bbox": [ + 59, + 96, + 169, + 107 + ], + "type": "text", + "content": "Our final loss is, therefore:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 116, + 287, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 116, + 287, + 130 + ], + "spans": [ + { + "bbox": [ + 75, + 116, + 287, + 130 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o n t}} + \\alpha \\cdot \\mathcal {L} _ {\\text {n e g}} + \\beta \\cdot \\left(\\mathcal {L} _ {\\text {s i m}} ^ {\\text {t e x t}} + \\mathcal {L} _ {\\text {s i m}} ^ {\\text {i m g}}\\right). \\tag {6}", + "image_path": "40b5bcb6e7922de9440d00c688a21cd1d4900ee399ce26cab71e17a3261fab3e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 143, + 237, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 237, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 237, + 156 + ], + "type": "text", + "content": "3.3. Fine-tuning Pre-trained VL Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": "Each of the " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_T" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_I" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " networks is comprised of a mix of non-parametric functions and two types of parametric functions: linear layers and embedding layers. Roughly, each of those functions, " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_k^{lin}(x)" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_k^{emb}(x)" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " is layer index, is parameterized by a weight matrix " + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 47, + 160, + 287, + 220 + ], + "type": "text", + "content": " so that:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 228, + 287, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 228, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 110, + 228, + 287, + 242 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {k} ^ {\\text {l i n}} (x) = \\mathcal {W} _ {k} \\cdot x \\tag {7}", + "image_path": "cf370e2003f73df3a47beeaeac1927b970cf4cadf7bfa07edfc2d1e4c68fb017.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 244, + 287, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 244, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 111, + 244, + 287, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {k} ^ {e m b} (x) = E M B (x; \\mathcal {W} _ {k}) \\tag {8}", + "image_path": "0c291f046ade023e28d077edec0085a1bf1ac803ffa5bf4a17ffd241e391158f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "EMB" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": " is the embedding operator assuming " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": " is a stream of integers and picking the respective columns of " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": ". Following the idea proposed in LoRA [27] for efficient LLM fine-tuning using low-rank residual adapters, when adapting a pre-trained VL model " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{E}_T, \\mathcal{E}_I)" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": " we parameterize the adapted weights " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k^*" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": " of the model " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^*" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": " fine-tuned from " + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 266, + 287, + 348 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 350, + 287, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 350, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 121, + 350, + 287, + 363 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {k} ^ {*} = \\mathcal {W} _ {k} + \\mathcal {A} _ {k} \\cdot \\mathcal {B} _ {k} \\tag {9}", + "image_path": "53575c320ab3d217fdd91166afdc1c59e5e90297c11f4fa9333f04d7784b0347.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": "where for " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_k" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "m\\times l" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_k" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": " are rank- " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": " matrices of sizes " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "m\\times r" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "inline_equation", + "content": "r\\times l" + }, + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": " respectively. These low-rank residual adapters can be applied efficiently as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 410, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 410, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 66, + 410, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {k} ^ {*, l i n} (x) = \\mathcal {F} _ {k} ^ {l i n} (x) + \\mathcal {A} _ {k} \\cdot (\\mathcal {B} _ {k} \\cdot x) \\tag {10}", + "image_path": "8b0d699e18bcde36a841ecac51b6e8de2bf458f825754b489a24aca7844d1134.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 427, + 287, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 427, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 67, + 427, + 287, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {k} ^ {*}, e m b (x) = \\mathcal {F} _ {k} ^ {e m b} (x) + \\mathcal {A} _ {k} \\cdot E M B (x; \\mathcal {B} _ {k}) \\tag {11}", + "image_path": "9c066dd9f7a42bfa18893bdc4ef71c3270e64327c6d7c14895b88e6fbc00712b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "text", + "content": "During the fine-tuning of " + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^*" + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "text", + "content": ", we freeze all the base model " + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "text", + "content": " parameters " + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\forall k, \\{\\mathcal{W}_k\\}" + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "text", + "content": " and only the LoRA adapters " + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\forall k, \\{(\\mathcal{A}_k, \\mathcal{B}_k)\\}" + }, + { + "bbox": [ + 47, + 450, + 287, + 510 + ], + "type": "text", + "content": " are being learned. In the above notation we disregard possible bias terms of the linear functions, if they are present, since we keep them frozen too." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": "There are several interesting things to note about the proposed architecture: (i) as opposed to [64], who evaluated the use of efficient LLM fine-tuning techniques for VL models adaptation, we add our LoRA adapters everywhere, i.e. to all layers of both the text and image encoders, and not only to the text encoder/decoder as done in [64]; (ii) as opposed to [80] who attach a small side network only to the output of the adapted model, our LoRA adapters are added to all the parametric functions inside the model and affect all the intermediate computations; (iii) same as noted in the original [27] paper, at inference all the LoRA adapters can be folded back into the original weight matrices by simple summation, thus returning the number of total parameters to be the same as in the original model and hence have the same inference speed; (iv) with rank " + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": " kept low, the number of extra parameters added by all the LoRA adapters can be very low making adaptation fast and efficient; finally," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "(v) such form of fine low-rank adaptation allows for significantly mitigating the zero-shot performance forgetting effects as demonstrated in our results and explored in the corresponding ablation (Sec. 5)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 149, + 386, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 149, + 386, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 149, + 386, + 163 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 175, + 368, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 368, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 368, + 186 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 198, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 245 + ], + "type": "text", + "content": "We train the model using common Image-Text pair datasets, namely Conceptual Captions 3M [62] and LAION 400M [61] and test using the VL-Checklist [82] datasets, which will be described below." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 246, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 545, + 305 + ], + "type": "text", + "content": "Conceptual Captions 3M (CC3M) [62] is a dataset of three million image-text pairs automatically crawled from the internet where image descriptions are harvested from Alt-text attributes and then processed and filtered to create relatively clean descriptions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 306, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 413 + ], + "type": "text", + "content": "LAION 400M [61] is a very large scale image-text pair dataset which, similarly to CC3M has been automatically harvested from the internet. One major difference between the two, apart from the size, is that LAION examples have been filtered using the pretrained CLIP model, such that the CLIP image-text similarity is high by design. Recent re-implementations of the original CLIP paper have successfully used LAION 400M to reproduce similar capabilities as the original CLIP model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 414, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 628 + ], + "type": "text", + "content": "VL-Checklist [82] combines images and annotations from the Visual Genome [39], SWiG [55], VAW [54], and HAKE [48] datasets. It is processed such that each image is annotated with two captions, one positive and one negative. The positive caption originates in its source dataset and corresponds to the image. The negative caption is constructed from the positive caption so that only one word, corresponding to the tested SVLC of interest, is changed to negate the SVLC (e.g., color, size, material, etc.). VL-Checklist evaluates three main types of VL concepts further subdivided into 7 types of SVLCs total: (1) Object: its spatial location and size, (2) Attribute: color, material, size, state, and action, and (3) Relation: spatial or action relation between two objects. In the following sections we report results on a combined VL-Checklist dataset. The results on individual comprising datasets (Visual Genome [39], SWiG [55], VAW [54], and HAKE [48]) are provided in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 629, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 712 + ], + "type": "text", + "content": "Zero-Shot Classification We evaluated our method on 21 classification dataset using the Zero-Shot classification protocol described in the ELEVATER Image Classification Toolkit [42]. The evaluation includes 21 different datasets, including common classification datasets such as ImageNet [60], CIFAR100 [40], EuroSat [21], and others. In Tables 1-3 we report the average results over the 21 tasks." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2662" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 502, + 155 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 502, + 155 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 502, + 155 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 502, + 155 + ], + "type": "table", + "html": "
ObjectVL-Checklist AttributeRelation21 Zero-Shot Tasks Average
CLIP [57]81.58%67.60%63.05%56.37%
CLIP +LoRA80.93% (-0.66%)66.28% (-1.32%)55.52% (-7.53%)56.41%(+0.04%)
w/o Neg Loss82.27% (+0.69%)67.58% (-0.02%)55.17% (-7.88%)55.37% (-1.00%)
Ours RB+LLM Negs85.09% (+3.50%)73.90% (+6.30%)78.72% (+15.67%)54.66% (-1.71%)
", + "image_path": "c72db18680a7699ba4fef32b47fe0087d4293b283810155ba3875ea38fe92ca0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 162, + 487, + 174 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 487, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 487, + 174 + ], + "type": "text", + "content": "Table 2. Ablation study - separate Negative Losses (Sec. 3.2) vs adding negatives to contrastive loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 184, + 180, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 180, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 180, + 197 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 201, + 288, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 406 + ], + "type": "text", + "content": "For CLIP we use the ML-Foundation Open-CLIP repository [28] and for CyCLIP [19] we use its original code repository, which is also based on Open-CLIP. In most experiments, unless stated otherwise, we train our model for five epochs on 4 V100 NVIDIA GPUs, with a total batch size of 512. When starting from a pre-trained model, we use rank 4 LoRA adapters (Sec. 3.3) and the learning rate is set to " + }, + { + "bbox": [ + 46, + 201, + 288, + 406 + ], + "type": "inline_equation", + "content": "5E - 6" + }, + { + "bbox": [ + 46, + 201, + 288, + 406 + ], + "type": "text", + "content": ". When training from scratch, for CLIP we use the default parameters set in the open-CLIP library, and for CyCLIP the defaults of its original implementation. For all CLIP experiments, we use VIT/32-B as the model architecture and ResNet-50 when training CyCLIP (following [19]). When fine-tuning we initialize with the original model weights released by the respective authors. In all experiments involving a combination of CyCLIP [19] and our method (in Sec. 4.4-4.5), in addition to Eq. (6) loss we also employ all the extra losses proposed in CyCLIP [19]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 409, + 113, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 113, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 113, + 421 + ], + "type": "text", + "content": "4.3. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 570 + ], + "type": "text", + "content": "We compare our method to two strong baselines under several configurations. The first is the CLIP [57] OpenAI pretrained model trained on 400M image-text pairs which achieves high ZS performance. The second is the very recent CyCLIP [19] method which, similarly to us, improves over the original CLIP loss. For a fair comparison all methods use the same network architecture and the same initialization from the OpenAI pretrained model. For the CyCLIP baseline we continue training from the pretrained initialization using LoRA and the CyCLIP losses. As CyCLIP losses are orthogonal to ours, we also show a unified version of the two methods (CyCLIP + Ours Combined)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 180, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 180, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 180, + 588 + ], + "type": "text", + "content": "4.4. Fine-tuning VL Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "In the following experiments we show the effects of fin-tuning a pre-trained VL model using our additional data enhancement methods and losses. All experiments are initialized from the official OpenAI CLIP. We compare our method to two baselines on the VL-Checklist tasks, and on the Zero-Shot image classification task. The first baseline is the original pretrained model without any further training. The second, is the same model with the additional LoRA parameters, trained on CC3M (Table 1a) and LAION (Table 1g) using the original CLIP loss function. Table 1a" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 182, + 512, + 338 + ], + "blocks": [ + { + "bbox": [ + 340, + 182, + 512, + 338 + ], + "lines": [ + { + "bbox": [ + 340, + 182, + 512, + 338 + ], + "spans": [ + { + "bbox": [ + 340, + 182, + 512, + 338 + ], + "type": "image", + "image_path": "f6d2f5a157ac665955aa3d39ef8f23d02b2101b355f705b7fd77ac3772737f17.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 347, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 347, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 547, + 392 + ], + "type": "text", + "content": "Figure 3. CC3M fine-tuning - detailed. Detailed results of the baselines (CLIP, CLIP+LoRA) and our models (RB+LLM Neg, RB+LLM Negs+Pos) initialized from OpenAI pretrained CLIP ViT-32B and fine-tuned (except CLIP) on CC3M." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 407, + 546, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 546, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 546, + 598 + ], + "type": "text", + "content": "shows several configurations of our method compared to the baselines. We see that our method shows significant improvements on all VL-Checklist tasks reaching up to " + }, + { + "bbox": [ + 304, + 407, + 546, + 598 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 407, + 546, + 598 + ], + "type": "text", + "content": " improvement. Figure 3 displays the relative gains on all tasks of the \"Attribute\" and \"Relation\" tests. It is clear that our gains are across all tests. These improvements come at a price of some minor degradation to the Zero-Shot performance compared to CLIP. Moreover, when fine-tuning CLIP on the LAION dataset (Table 1g) we do not see these degradations. In Tab. 1b we show consistent gains to ones in Tab. 1a when finetuning with a stronger (higher ZS) VitB/16 CLIP [57] pre-trained image encoder, observing an even lower drop in ZS performance. In Tab. 1e we show significant gains in fine-tuning CyCLIP on CC3M, comparing fine-tuning before and after integrating our proposed approach, this also comes at no ZS performance cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 609, + 432, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 432, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 432, + 623 + ], + "type": "text", + "content": "4.5. Training from scratch" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "In these experiments, we show the advantage of integrating our approach throughout the whole training procedure. To this end, we train from scratch on CC3M - both CLIP and CyCLIP, either on their own, or together with our proposed method integrated. Tables - Tab. 1c, Tab. 1d, and Tab. 1f show that, similar to finetining, our method significantly improves both CLIP and CyCLIP SVLC capabilities." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2663" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 64, + 529, + 155 + ], + "blocks": [ + { + "bbox": [ + 66, + 64, + 529, + 155 + ], + "lines": [ + { + "bbox": [ + 66, + 64, + 529, + 155 + ], + "spans": [ + { + "bbox": [ + 66, + 64, + 529, + 155 + ], + "type": "table", + "html": "
MethodColorMaterialAttribute SizeStateActionRelation21 Zero-Shot Tasks Average
ActionSpatial
Ours Pos72.35%69.25%69.80%59.35%66.08%70.97%39.20%55.37%
Ours RB Neg78.45%83.20%69.50%65.95%69.66%75.97%74.70%54.66%
Ours LLM Neg76.00%79.70%72.75%61.35%68.36%77.23%72.40%55.05%
Ours RB+LLM Negs79.25%84.25%72.15%64.05%69.82%79.03%78.40%54.66%%
Ours Combined77.45%77.35%73.35%62.30%69.39%74.70%63.20%54.77%%
", + "image_path": "036d6a0d76816cc7a750759277fceb719efcd1bf111741bc168dcf2656d4aef5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 163, + 479, + 174 + ], + "lines": [ + { + "bbox": [ + 113, + 163, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 113, + 163, + 479, + 174 + ], + "type": "text", + "content": "Table 3. Ablation study - component analysis. Detailed evaluation on Attribute & Relation SVLCs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 185, + 288, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 288, + 269 + ], + "type": "text", + "content": "ties, while still keeping similar ZS performance. Figure 4 shows the detailed parsing of the \"Attribute\" and \"Relation\" tests with consistent gains across the specific tasks. These findings suggest that had we trained our method for many epochs on a large dataset, such as the full LAION-9B, we would reach similar ZS performance while greatly enhancing the SVLC performance of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 279, + 113, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 113, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 113, + 291 + ], + "type": "text", + "content": "5. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 297, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 322 + ], + "type": "text", + "content": "In this section we will examine several aspects and components of our proposed methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 328, + 144, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 144, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 144, + 341 + ], + "type": "text", + "content": "5.1. Negative Losses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 345, + 288, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 288, + 464 + ], + "type": "text", + "content": "Section 3.2 details our additional loss functions which utilize our generated texts. Specifically, Eq. (3) describes the loss which contrasts a given example only with its generated negatives. Table 2 clearly shows the importance of this loss as opposed to simply adding the negative examples to the original contrastive loss (Eq. (2)). Without explicitly forcing the network to attend to the small changes in the text (Eq. (3)), the generated negative examples, when inserted only to the contrastive loss (Eq. (2)) do not provide the desired gains over the baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 472, + 165, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 165, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 165, + 484 + ], + "type": "text", + "content": "5.2. Component analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 488, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 525 + ], + "type": "text", + "content": "In Section 3 we describe several components of our proposed method. Specifically we present our RB negative generation (Sec. 3.1.1), our LLM-based negative genera" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 93, + 540, + 242, + 676 + ], + "blocks": [ + { + "bbox": [ + 93, + 540, + 242, + 676 + ], + "lines": [ + { + "bbox": [ + 93, + 540, + 242, + 676 + ], + "spans": [ + { + "bbox": [ + 93, + 540, + 242, + 676 + ], + "type": "image", + "image_path": "de41b8340c047d239b087b68f831ecbe3f93a12b553ad11bbe5d23118219e5ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 685, + 287, + 718 + ], + "lines": [ + { + "bbox": [ + 46, + 685, + 287, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 685, + 287, + 718 + ], + "type": "text", + "content": "Figure 4. CC3M train from scratch - detailed. Detailed results of the baseline CLIP and our models (RB+LLM Neg, RB+LLM Negs+Pos), all CLIP ViT-32B, trained from scratch on CC3M." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 185, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 342 + ], + "type": "text", + "content": "tion (Sec. 3.1.2), and our LLM-based analogy generation, referred to here as \"Pos\" (Sec. 3.1.3). Tab. 3 provides a detailed analysis and comparison of the contribution of each component to the final result. Through this analysis we see two contradicting forces between the SVLC capabilities and the original Zero-Shot performance. We can see that each negative generation method plays a crucial role in some of the tasks while the use of both is usually the best performing option. On the other hand, the LLM-based analogy generation stabilizes the Zero-Shot performance and mitigates the drop with respect to the baseline. The joint version of all components (\"Ours Combined\") exhibits a good trade-off between the two contradicting forces." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 350, + 384, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 384, + 362 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 384, + 362 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 476 + ], + "type": "text", + "content": "We have presented a data-driven technique for enhancing the performance of VL models in the important task of SVLC understanding without sacrificing their impressive ZS object recognition capabilities. Our proposed method attains significant gains in multiple experiments on a variety of base VL models and datasets. It builds upon the modeling strength and knowledge of language structure to teach this structure to VL models in an orthogonal way, suggesting wide applicability to existing or future VL models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 476, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 596 + ], + "type": "text", + "content": "While attaining impressive gains in SVLC understanding, we believe the small drop observed in ZS performance could be further reduced. An additional possible extension of our work is using more sophisticated sequence generation techniques for improving our batch data. One may combine annotation efforts with a language model to get improved data for training and evaluation [51]. Another possibility is adding a corrector [72] in the training that validates whether the VL model learns the correct concepts or not. We leave all of these exciting directions to future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 607, + 409, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 607, + 409, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 607, + 409, + 620 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 625, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 545, + 712 + ], + "type": "text", + "content": "We would like to thank Donghyun Kim for his invaluable help in this work. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. FA8750-19-C-1001. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA). This material is based upon work supported by the 'Data Science grant from the Israeli Council of Higher Education'. This research or RG was supported by ERC-StG SPADE grant no. 75749." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2664" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Anurag Arnab, Chen Sun, and Cordelia Schmid. Unified graph structured models for video understanding. In ICCV, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 190 + ], + "type": "text", + "content": "[2] Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, and Amir Globerson. Bringing image scene structure to video via frame-clip consistency of object tokens. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "type": "text", + "content": "[3] Amir Bar, Roei Herzig, Xiaolong Wang, Anna Rohrbach, Gal Chechik, Trevor Darrell, and A. Globerson. Compositional video synthesis with action graphs. In ICML, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "text", + "content": "[4] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV, pages 105-121, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 338 + ], + "type": "text", + "content": "[5] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 393 + ], + "type": "text", + "content": "[6] Andrea Burns, Donghyun Kim, Derry Wijaya, Kate Saenko, and Bryan A Plummer. Learning to scale multilingual representations for vision-language tasks. In European Conference on Computer Vision, pages 197-213. Springer, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[7] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 233-248, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 443, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 288, + 485 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 488, + 288, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 288, + 541 + ], + "type": "text", + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "type": "text", + "content": "[10] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 632 + ], + "type": "text", + "content": "[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[12] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[14] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't stop learning: Towards continual learning for the clip model. arXiv:2207.09248, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 131, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 131, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 131, + 545, + 185 + ], + "type": "text", + "content": "[15] Andreas Fürst, Elisabeth Rumetshofer, Viet Tran, Hubert Ramsauer, Fei Tang, Johannes Lehner, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto-Nemling, et al. Cloob: Modern hopfield networks with infoloob outperform clip. arXiv preprint arXiv:2110.11316, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 217 + ], + "type": "text", + "content": "[16] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. ArXiv, abs/2008.11714, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[17] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "type": "text", + "content": "[18] Shivam Garg, Dimitris Tsipras, Percy Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "text", + "content": "[19] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan A Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 345, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 345, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 345, + 545, + 387 + ], + "type": "text", + "content": "[20] Xiaoshuai Hao, Yi Zhu, Srikar Appalaraju, Aston Zhang, Wanqian Zhang, Bo Li, and Mu Li. Mixgen: A new multimodal data augmentation. arXiv preprint arXiv:2206.08358, 2022.4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 389, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 443 + ], + "type": "text", + "content": "[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[22] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darrell, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 544 + ], + "type": "text", + "content": "[23] Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam, Amir Bar, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Object-region video transformers. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "type": "text", + "content": "[24] Roei Herzig, Elad Levi, Huijuan Xu, Hang Gao, Eli Brosh, Xiaolong Wang, Amir Globerson, and Trevor Darrell. Spatio-temporal action graph networks. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 0–0, 2019. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "text", + "content": "[25] Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, and Amir Globerson. Mapping images to scene graphs with permutation-invariant structured prediction. In Advances in Neural Information Processing Systems (NIPS), 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[26] Matthew Honnibal and Ines Montani. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear, 2017. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2665" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.2,3,6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 161 + ], + "type": "text", + "content": "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "text", + "content": "[29] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "text", + "content": "[30] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as composition of spatiotemporal scene graphs. arXiv preprint arXiv:1912.06992, 2019.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 318 + ], + "type": "text", + "content": "[31] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "type": "text", + "content": "[32] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1219–1228, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "type": "text", + "content": "[33] Keizo Kato, Yin Li, and Abhinav Gupta. Compositional learning for human object interaction. In ECCV, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "text", + "content": "[34] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. AAAI Conference on Artificial Intelligence, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "text", + "content": "[35] Donghyun Kim, Kuniaki Saito, Kate Saenko, Stan Sclaroff, and Bryan Plummer. Mule: Multimodal universal language embedding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11254-11261, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[36] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 599 + ], + "type": "text", + "content": "[37] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 623 + ], + "type": "text", + "content": "[38] Ranjay Krishna, Ines Chami, Michael S. Bernstein, and Li Fei-Fei. Referring relationships. *ECCV*, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[39] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[41] Varun Kumar, Ashutosh Choudhary, and Eunah Cho. Data augmentation using pre-trained transformer models. arXiv preprint arXiv:2003.02245, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 171 + ], + "type": "text", + "content": "[42] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. Neural Information Processing Systems, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[43] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 1, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 271 + ], + "type": "text", + "content": "[44] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 273, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 314 + ], + "type": "text", + "content": "[45] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. ArXiv, abs/1908.03557, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "text", + "content": "[46] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "text", + "content": "[47] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "text", + "content": "[48] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Mingyang Chen, Ze Ma, Shiyi Wang, Hao-Shu Fang, and Cewu Lu. Hake: Human activity knowledge engine. arXiv preprint arXiv:1904.06539, 2019. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 503 + ], + "type": "text", + "content": "[49] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 505, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 545, + 558 + ], + "type": "text", + "content": "[50] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 592 + ], + "type": "text", + "content": "[51] Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. Wanli: Worker and ai collaboration for natural language inference dataset creation, Jan. 2022. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 593, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 647 + ], + "type": "text", + "content": "[52] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "text", + "content": "[53] Margaret Mitchell, Giada Pistilli, Yacine Jernite, Ezinwanne Ozoani, Marissa Gerchick, Nazneen Rajani, Sasha Luccioni, Irene Solaiman, Maraim Masoud, Somaieh Nikpoor, Munoz Ferrandis Carlos, Stas Bekman, Christopher Akiki, Danish Contractor, David Lansky, Angelina McMillan-Major, Tristan Thrush, Suzana Ilic, Gerard Dupont, Shayne Longpre," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2666" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 128 + ], + "type": "text", + "content": "Manan Dey, Stella Biderman, Douwe Kiela, Emi Baylor, Teven Le Scao, Aaron Gokaslan, Julien Launay, and Niklas Muennighoff. Bigscience, bigscience language open-science open-access multilingual (bloom) language model. International, May 2021-May 2022. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[54] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13018-13028, 2021. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "text", + "content": "[55] Sarah Pratt, Mark Yatskar, Luca Weihs, Ali Farhadi, and Aniruddha Kembhavi. Grounded situation recognition. In European Conference on Computer Vision, pages 314-332. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 263 + ], + "type": "text", + "content": "[56] Moshiko Raboh, Roei Herzig, Gal Chechik, Jonathan Berant, and Amir Globerson. Differentiable scene graphs. In WACV, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 265, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 331 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[58] Vinay Venkatesh Ramasesh, Ethan Dyer, and Maithra Raghu. Anatomy of catastrophic forgetting: Hidden representations and task semantics. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "text", + "content": "[59] Arijit Ray, Karan Sikka, Ajay Divakaran, Stefan Lee, and Giedrius Burachas. Sunny and dark outside?! improving answer consistency in vqa through entailed question generation. arXiv preprint arXiv:1909.04696, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 477 + ], + "type": "text", + "content": "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "text", + "content": "[61] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 535, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 601 + ], + "type": "text", + "content": "[62] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[63] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. arXiv preprint arXiv:2001.07685, 2020.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[64] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5237, 2022. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[65] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[66] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "text", + "content": "[67] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In European Conference on Computer Vision, pages 437-453. Springer, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 194, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 260 + ], + "type": "text", + "content": "[68] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 1, 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 327 + ], + "type": "text", + "content": "[69] William Yang Wang and Diyi Yang. That's so annoying!!! A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using# petpeeve tweets. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 2557–2563, 2015. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 349 + ], + "type": "text", + "content": "[70] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV, 2018. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "text", + "content": "[71] Jason Wei and Kai Zou. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[72] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. Generating sequences by learning to self-correct, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "text", + "content": "[73] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[74] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and M. Kankanhalli. Learning to detect human-object interactions with knowledge. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2019-2028, 2019. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "text", + "content": "[75] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene Graph Generation by Iterative Message Passing. In CVPR, pages 3097–3106, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[76] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[77] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 659, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 714 + ], + "type": "text", + "content": "[78] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2667" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 306 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[79] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 160 + ], + "type": "text", + "content": "[80] Jeffrey O Zhang, Alexander Sax, Amir Zamir, Leonidas Guibas, and Jitendra Malik. Side-tuning: a baseline for network adaptation via additive side networks. In European Conference on Computer Vision, pages 698-714. Springer, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "text", + "content": "[81] Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 251 + ], + "type": "text", + "content": "[82] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. V1-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 286, + 306 + ], + "type": "text", + "content": "[83] Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022.4" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2668" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_content_list.json b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a54ce5657931666c1ed598ad83b9b33afec1c7ca --- /dev/null +++ b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_content_list.json @@ -0,0 +1,2270 @@ +[ + { + "type": "text", + "text": "Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22ce7a0da53799a283a040951791c760f89a62c4d805631443c2a3890330a18a.jpg", + "image_caption": [ + "Figure 1. Teleidoscopic Imaging System. (a) We propose a novel system for microscale 3D imaging using an imaging setup that consists of a camera, a kaleidoscopic multi-facet mirror with a monocentric front lens, and a projector. (b) Our system realizes a dense close-up surrounding multi-view capture of the target. (c) Our method enables fuller and dense reconstruction of microscale objects." + ], + "image_footnote": [], + "bbox": [ + 89, + 244, + 356, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d93f9aa0dd94cd442af7962111403cd3b0ebbedd03dd492586eca05959dcd061.jpg", + "image_caption": [ + "Ryo Kawahara \nKyushu Institute of Technology \nrkawahara@ai.kyutech.ac.jp" + ], + "image_footnote": [], + "bbox": [ + 367, + 243, + 589, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cfedaef57b0e38b012017ff3114aa0cfb48426e45e7fcea13244e7c9d0bcf83.jpg", + "image_caption": [ + "Meng-Yu Jennifer Kuo University of Minnesota kuo00052@umn.edu" + ], + "image_footnote": [], + "bbox": [ + 593, + 243, + 781, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/668d72d9a3758679c630e887d42512a46a2f017998456eb55bf41a8152504fd7.jpg", + "image_caption": [ + "Shohei Nobuhara Kyoto University nob@i.kyoto-u.ac.jp" + ], + "image_footnote": [], + "bbox": [ + 794, + 243, + 883, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 446, + 313, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a practical method of microscale 3D shape capturing by a teleidoscopic imaging system. The main challenge in microscale 3D shape reconstruction is to capture the target from multiple viewpoints with a large enough depth-of-field. Our idea is to employ a teleidoscopic measurement system consisting of three planar mirrors and monocentric lens. The planar mirrors virtually define multiple viewpoints by multiple reflections, and the monocentric lens realizes a high magnification with less blurry and surround view even in closeup imaging. Our contributions include, a structured ray-pixel camera model which handles refractive and reflective projection rays efficiently, analytical evaluations of depth of field of our teleidoscopic imaging system, and a practical calibration algorithm of the teleidoscopic imaging system. Evaluations with real images prove the concept of our measurement system.", + "bbox": [ + 75, + 479, + 470, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 738, + 207, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Microscale 3D reconstruction has found profound applications in a wide range of domains including medical imaging, life science, and aquaculture, due to its non-constrained and non-invasive measurements. The main challenges in image-based microscopic 3D shape measurement is its shallow depth of field and camera arrangement in the closeup scenario. Applying conventional multiple camera system designed for human-size capture [14, 30] cannot be a feasible solution due to limitations on camera placement. Con-", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ventional multiple mirror system [33] also have difficulties inevitably in depth-of-focus due to differences in their optical path lengths with varying numbers of bounces.", + "bbox": [ + 496, + 448, + 890, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we show that the fuller 3D shape of a microscale object can be recovered. Our key idea is to employ a catadioptric imaging system which realizes a practical closeup multi-view imaging. The point of our design is that the system has a monocentric front lens like a teleidoscope, instead of using microscopic system in the camera side. That is, as shown in Fig. 1(a), we introduce a kaleidoscopic multi-facet mirror between the front lens and the camera, where the design realizes a deeper depth-of-field and results in less blurring imaging. Unlike conventional microscopic imaging system such as differential phase contrast (DPC) microscopy [4, 34] and multi-focus approaches [15, 24, 26], our method realizes a multi-view capture of the target from a single physical viewpoint which can contribute to free-viewpoint rendering, 3D shape reconstruction, and reflection analysis. Our system also has an advantage over existing imaging systems that build multiple views behind the main lens [7, 21] in a closeup environment. The wide FoV with a monocentric main lens and virtual multiple views with mirrors allow closeup and surrounding view capturing in focus.", + "bbox": [ + 496, + 500, + 892, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We call our system teleidoscopic imaging system and show that the system can be compactly modeled by a structured ray-pixel camera model [11], which handles refractive and reflective projection rays efficiently. Based on our ray-pixel camera model, we derive a practical calibration algo", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rithm that estimates the positions of the monocentric lens and multi-facet mirror w.r.t. the camera by using a single reference planar pattern (i.e., a checkerboard). Given the calibration parameters, a scene point can then be linearly triangulated from its teleidoscopic projection in a direct linear transform (DLT) manner [13].", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We implement our method with an imaging system consisting of a camera, three planar mirrors, a monocentric lens, and a projector placed outside the mirrors for structured-light casting. We quantitatively evaluate the computation cost of numerical projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also validate the effectiveness of our method qualitatively on a number of real-world microscale objects. These results demonstrate that our method enables holistic and dense reconstruction of microscale objects. We believe our method expands the avenues of three-dimensional analysis of microscale objects and scenes in real world scenarios.", + "bbox": [ + 75, + 181, + 467, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 390, + 218, + 405 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our teleidoscopic imaging system is a catadioptric system consisting of a monocentric lens and kaleidoscopic mirrors (Fig. 1(a)). While catadioptric system has a wide variety of applications such as omnidirectional observation and panoramic stereo [1, 10, 25, 27], this section reviews studies with multiple flat mirrors and monocentric lenses.", + "bbox": [ + 75, + 415, + 467, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Imaging with Multiple Flat Mirrors", + "text_level": 1, + "bbox": [ + 76, + 513, + 385, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A fundamental motivation of introducing mirrors in observation system is to increase the number of viewpoints without installing additional cameras for multi-view capture of a target [3, 8, 20, 28, 32, 33]. Takahashi et al. [33] have proposed a kaleidoscopic imaging system and demonstrated a 3D shape reconstruction using multiple reflections. Tagawa et al. [32] have proposed a multi-facet imaging system that observes a target from an equally distributed virtual cameras for reflectance analysis.", + "bbox": [ + 75, + 537, + 467, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Imaging with Monocentric Lens", + "text_level": 1, + "bbox": [ + 76, + 681, + 359, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the context of imaging system, the monocentric lens is often used to obtain a wide field-of-view [17, 35], in particular for endoscopes, as a short focal length lens. The monocentric lens, however, has additional useful characteristics: its symmetric structure and magnifying power. Cossair et al. [5] have proposed a camera array which captures a same scene through a single monocentric lens so that the images can be stitched into a single high-resolution image. Similarly, Dansereau et al. [6] have proposed a lightfield camera which captures omnidirectional lightfield images through a single monocentric lens with a camera orbiting around it.", + "bbox": [ + 75, + 704, + 467, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A typical use of the monocentric lens as a magnifier can be found in the Leeuwenhoek's microscope in the 17th cen", + "bbox": [ + 76, + 869, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4917e17c664ac641403725b7043a4d08968a5a6af8531e5094fb63e1dfa86bb4.jpg", + "image_caption": [ + "Figure 2. Measurement through Monocentric Lens. The segments of projection path $\\ell_{o} - \\ell_{g} - \\ell_{i}$ have an axially symmetric structure around the axis which directs to the monocentric lens center $o_{D}$ ." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 883, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tury. It utilizes a single monocentric lens and realized over $100 \\times$ magnifications [9].", + "bbox": [ + 500, + 316, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Ray-pixel Camera Model", + "text_level": 1, + "bbox": [ + 500, + 358, + 728, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to handle light paths captured via complex imaging system such as reflection and refraction, Grossberg and Nayar [11] have proposed a ray-pixel (or raxel) camera model which describes the projection by end-to-end mappings between the pixels and their corresponding emergent rays. While this pixel-wise representation has a great flexibility to describe complicated projections [22, 31], it is also possible to exploit a structure of the rays to simplify the model [16, 22, 36-38]. Kawahara et al. [16] have proposed a ray-pixel camera model for cameras observing the scene through flat surfaces. Compared with analytical forward projections through flat surfaces which requires solving a high order equation [2], it realized a practical numerical forward projection computation by exploiting the axially symmetric structure of the rays.", + "bbox": [ + 498, + 382, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Ray-Pixel Camera Model for Teleidoscopic Imaging System", + "text_level": 1, + "bbox": [ + 500, + 625, + 890, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our teleidoscopic system, the cameras mirrored by the kaleidoscopic multi-facet mirrors observe the scene through the single monocentric lens (Fig. 1(a)) as done in [5,6], and each of the monocentric lens and the camera pairs form a microscope capturing the object from a different direction.", + "bbox": [ + 498, + 667, + 890, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we show that this design realizes a microscale multi-view capturing with a deep depth-of-field, and that the projections by the teleidoscopic imaging system can be modeled compactly by ray-pixel cameras.", + "bbox": [ + 498, + 744, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Measurement through Monocentric Lens", + "text_level": 1, + "bbox": [ + 500, + 816, + 849, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, suppose a perspective camera whose camera center is located at $\\pmb{o}_0$ observes the scene through a monocentric lens located at $\\pmb{o}_D$ . Here the perspective camera can be assumed to be directed to the center", + "bbox": [ + 498, + 840, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20814", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of the monocentric lens $\\mathbf{o}_D$ without loss of generality, since we can calibrate the rotation to align the optical axis of the camera to the line $\\mathbf{o}_D - \\mathbf{o}_0$ as described later. We denote this normalized camera by $C_n$ .", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Obviously, the rays backprojected through pixels of $C_n$ have an axially symmetric structure around the optical axis, and an incident ray $\\ell_i$ through a pixel $p_p$ and its refraction $\\ell_g$ and emergent ray $\\ell_o$ are on a single plane-of-refraction [2]. Therefore, we can describe the ray through a single pixel by a 2D $(r,z)^{\\top}$ coordinate system centered at $o_0$ .", + "bbox": [ + 75, + 152, + 467, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hereafter, let $r_{\\alpha}$ and $z_{\\alpha}$ denote the $r$ and $z$ elements of the vector $\\alpha$ in general. For example, point $\\pmb{p}_p$ is described as $\\pmb{p}_p = (r_{\\pmb{p}_p}, z_{\\pmb{p}_p})^\\top$ . Also let $\\pmb{v}_X = (r_{\\pmb{v}_X}, z_{\\pmb{v}_X})^\\top$ denote the direction vector of line $\\ell_X$ towards the scene from the camera, where $X = \\{\\mathrm{i}, \\mathrm{g}, \\mathrm{o}\\}$ represents the label of the light path segment.", + "bbox": [ + 76, + 247, + 467, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider the projection path $\\ell_o - \\ell_g - \\ell_i$ from $o_0$ through the point $\\pmb{p}_p$ . The incident point $\\pmb{p}_i$ on the sphere is described as a function of $\\theta_p = \\tan^{-1}(r_{\\pmb{p}_p} / f_c)$ as", + "bbox": [ + 76, + 339, + 467, + 387 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {p _ {i}} = \\frac {d _ {D} - \\sqrt {d _ {D} ^ {2} - (1 + \\tan^ {2} \\theta_ {p}) \\left(d _ {D} ^ {2} - r _ {D} ^ {2}\\right)}}{1 + \\tan^ {2} \\theta_ {p}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 402, + 467, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr _ {p _ {i}} = z _ {p _ {i}} \\tan \\theta_ {p},\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 433, + 267, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d_{D}$ and $r_D$ are the distances to the monocentric lens and the lens radius.", + "bbox": [ + 76, + 464, + 467, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The refraction angle $\\theta_{g}$ is then given by Snell's law with assuming the refractive indices of the air $\\mu_{a}$ and the lens $\\mu_{g}$ are known:", + "bbox": [ + 76, + 496, + 467, + 540 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {g} \\sin \\theta_ {g} = \\mu_ {a} \\sin \\left(\\theta_ {p} + \\theta_ {D i}\\right), \\\\ \\Leftrightarrow \\sin \\theta_ {g} = \\frac {\\mu_ {a}}{\\mu_ {g}} \\frac {d _ {D} r _ {p _ {i}}}{r _ {D} \\sqrt {r _ {p _ {i}} ^ {2} + z _ {p _ {i}} ^ {2}}}, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 555, + 468, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{D_i}$ is the angle of $o_0 - o_D - p_i$ .", + "bbox": [ + 76, + 619, + 316, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The point of emergence $\\pmb{p}_o$ is the intersection of the path $\\ell_g$ and the sphere. It is also obtained as the mirror of the point $\\pmb{p}_i$ because of the symmetrical relationship between incident path $\\ell_i$ and outgoing $\\ell_o$ w.r.t. the monocentric lens. As shown in Fig. 2, $\\ell_i$ and $\\ell_o$ are in line symmetry to the line that is perpendicular to the vector $\\pmb{n}_p$ through the point $\\pmb{o}_D$ . That is, given the point $\\pmb{p}_i$ , its reflection $\\pmb{p}_o$ is described as", + "bbox": [ + 76, + 636, + 467, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {p} _ {o} = H _ {p} \\boldsymbol {p} _ {i} + \\boldsymbol {t} _ {p}, \\tag {3} \\\\ \\Leftrightarrow \\boldsymbol {p} _ {o} = \\left(\\boldsymbol {I} - 2 \\boldsymbol {n} _ {p} \\boldsymbol {n} _ {p} ^ {\\top}\\right) \\boldsymbol {p} _ {i} + 2 (\\boldsymbol {o} _ {D} ^ {\\top} \\boldsymbol {n} _ {p}) \\boldsymbol {n} _ {p}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 771, + 467, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H_{p}$ is the Householder matrix and $\\pmb{t}_{p}$ denotes the center of reflection. Besides, the direction $\\pmb{n}_p$ is given as", + "bbox": [ + 76, + 823, + 467, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {n} _ {p} = \\left[ \\begin{array}{l} \\cos \\left(\\pi / 2 + \\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\sin \\left(\\pi / 2 + \\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right] = \\left[ \\begin{array}{c} - \\sin \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\cos \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 869, + 467, + 905 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5db582d9d8e83d9dba075e94906f27ea39cb5edc2c385ae58ed0a8493cbdf543.jpg", + "image_caption": [ + "Figure 3. 3D-2D Projection. The black line $\\ell_{o}$ is the correct projection path that intersects with the optical axis at the camera center $o_0$ . Our 3D-2D projection hypothesizes a line $\\ell_{\\lambda}$ as an initial guess of projection, and then optimizes it by verifying if the point and the angle of intersection with the optical axis is consistent with the ray-pixel mapping $\\langle \\theta_q, f_q \\rangle$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 99, + 867, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "therefore $H_{p}$ and $\\pmb{t}_{p}$ are rewritten to", + "bbox": [ + 500, + 338, + 738, + 354 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} H _ {p} = \\left[ \\begin{array}{l l} \\cos 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) & \\sin 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\sin 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) & - \\cos 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right], \\tag {5} \\\\ \\boldsymbol {t} _ {p} = \\left[ \\begin{array}{c} - d _ {D} \\sin 2 (\\theta_ {D i} - \\theta_ {g}) \\\\ d _ {D} \\cos 2 (\\theta_ {D i} - \\theta_ {g}) + d _ {D} \\end{array} \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 364, + 890, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The direction $\\pmb{v}_o$ is also given by Householder matrix as", + "bbox": [ + 500, + 443, + 869, + 458 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {o} = H _ {p} \\boldsymbol {v} _ {i}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 470, + 888, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a result, the intersection $\\pmb{o}_{f_q} = (0,f_q)^\\top$ of the ray $\\ell_o$ and the optical axis is given as follows:", + "bbox": [ + 500, + 498, + 888, + 530 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {o} _ {f _ {q}} = t _ {o} \\boldsymbol {v} _ {o} + \\boldsymbol {p} _ {o}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 542, + 888, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Leftrightarrow t _ {o} = - \\frac {r _ {\\boldsymbol {p} _ {o}}}{r _ {\\boldsymbol {v} _ {o}}}, \\quad f _ {q} = z _ {\\boldsymbol {p} _ {o}} + \\left(\\frac {- r _ {\\boldsymbol {v} _ {o}}}{z _ {\\boldsymbol {v} _ {o}}}\\right) ^ {- 1} r _ {\\boldsymbol {p} _ {o}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 561, + 888, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This equation describes the ray $\\ell_{o}$ backprojecting from the pixel $p_p$ as a function of $\\theta_{p}$ . Inversely, the forward projection can be obtained for a given 3D point in front of the monocentric lens. The analytical 3D-2D projection, however, requires solving a 10th-degree equation [2]. The next section introduces a ray-pixel camera which exploits the axial symmetric structure of the rays to provide an efficient numerical 3D-2D projection.", + "bbox": [ + 496, + 599, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Spherical Refraction Ray-Pixel Camera Model", + "text_level": 1, + "bbox": [ + 500, + 729, + 877, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Up to this point, we observed that the rays emitted from a perspective camera through a monocentric lens show an axially symmetric structure around the line from the camera to the lens center. As illustrated in Fig. 2, given a pixel $p_p$ by specifying $\\theta_p$ , the corresponding ray $\\ell_o$ can be determined uniquely by Eq. (8). This section introduces our spherical refraction ray-pixel camera model by representing the ray-pixel mapping as follows.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the light path from a pixel $\\pmb{p}_p$ on the plane-of-refraction. Suppose the emergent ray $\\ell_o$ intersects", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20815", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/75a5ba2f4d5c9c7e6ea8a05cdc068a1b9b9be251dc72abc4ce1a1568bbe0c4f3.jpg", + "table_caption": [ + "Table 1. The Pixel-Ray Mapping in Spherical Refraction Ray-Pixel Camera. A virtual pixel parameterized by $\\theta_q$ is associated with a pixel-wise focal length $f_q(\\theta_q) \\in \\mathbb{R}$ . The derivative $\\frac{\\mathrm{d}f_q(\\theta_q)}{\\mathrm{d}\\theta_q}$ is also stored for our numerical 3D-2D projection." + ], + "table_footnote": [], + "table_body": "
pixelθq0θq1...θqn
rayfq(θq0)fq(θq1)...fq(θqn)
dfq(θq)/dθqf′q(θq0)f′q(θq1)...f′q(θqn)
", + "bbox": [ + 111, + 152, + 436, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the optical axis at $\\mathbf{o}_{f_q} = (0, f_q)^\\top$ with angle $\\theta_q$ . Obviously, changing the position of $\\mathbf{p}_p$ in $r$ direction results in changing the corresponding $\\langle f_q, \\theta_q \\rangle$ pair. That is, since the mapping between the pixel $\\mathbf{p}_p$ i.e., $\\ell_i$ , and the ray $\\ell_o$ , i.e., $\\langle f_q, \\theta_q \\rangle$ is bijective because of the reversibility of light, representing the light paths from the pixels in $r$ space is identical to knowing all possible $\\langle f_q, \\theta_q \\rangle$ pairs. In other words, the system of Fig. 3 as a whole can be seen as a pixel-wise varifocal camera which changes the focal length $f_q$ for each virtual pixel parameterized by $\\theta_q$ . In fact, the mapping $\\theta_q \\mapsto f_q$ is a monotonic function due to the spherical structure of the lens.", + "bbox": [ + 75, + 238, + 468, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows our ray-pixel mapping. Due to the spherical structure of the lens, this is a discretization of the monotonic function $\\theta_q \\mapsto f_q$ by $\\theta_q$ . In practice, we sample $\\theta_q$ so that their interval results in a sub-pixel sampling in the original image domain. Notice that the derivative $\\frac{\\mathrm{d}f_q(\\theta_q)}{\\mathrm{d}\\theta_q}$ is also stored for our numerical 3D-2D projection as described below.", + "bbox": [ + 75, + 422, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 3D-2D Projection", + "text_level": 1, + "bbox": [ + 76, + 553, + 253, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of the analytical forward projection which requires solving a 10th-degree equation [2], this section introduces a numerical forward projection using our ray-pixel mapping table (Tab. 1). The key idea of our numerical forward projection is to hypothesize a projection line $\\ell_{\\lambda}$ in Fig. 3 first, and check if it intersects with the optical axis at the identical virtual focal length stored in the ray-pixel mapping table (Tab. 1). If the focal lengths are not identical, then $\\ell_{\\lambda}$ , i.e., the virtual pixel $\\theta_{q}$ equivalently, is refined to minimize the difference. We can formulate this process as a Gauss-Newton optimization as follows.", + "bbox": [ + 75, + 575, + 468, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gauss-Newton Method As shown in Fig. 3, let us consider the ray $\\ell_{\\lambda}$ from a 3D point to be projected. By hypothesizing that $\\ell_{\\lambda}$ intersects the optical axis at an angle $\\theta_{q}$ , i.e., by hypothesizing that the 3D point is projected to a virtual pixel $\\theta_{q}$ , we can compute the intersection $\\Lambda_{q}(\\theta_{q}) = (0, \\lambda_{q}(\\theta_{q}))^{\\top}$ of $\\ell_{\\lambda}$ and the optical axis. If $\\lambda_{q}$ is equal to the virtual focal length $f_{q}(\\theta_{q})$ stored in the ray-pixel mapping, the ray $\\ell_{\\lambda}$ is identical to $\\ell_{o}$ , and hence that can intersect with the optical axis at the camera center $o_{0}$ .", + "bbox": [ + 75, + 746, + 468, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "That is, the numerical 3D-2D projection can be achieved", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a36862c957beaaf26ce9a13cc4815d931ac3952aa8a80fae46246aedf5bd2b61.jpg", + "image_caption": [ + "Figure 4. Kaleidoscopic Imaging. (a) Mirrored Cameras. (b) Chambers in the real camera image. The label $i$ of $C_i$ and $\\pmb{u}^{(i)}$ denote the reflections of $C$ and $u$ by the $i$ th mirror respectively." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 681, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/47819e3d5136a0a930b81dc57518d1c877fd451ce490bdfe93f36a0ff7d640c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 95, + 870, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by solving the following optimization:", + "bbox": [ + 498, + 292, + 754, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {q} = \\underset {\\theta_ {q}} {\\operatorname {a r g m i n}} G \\left(\\theta_ {q}\\right) = \\underset {\\theta_ {q}} {\\operatorname {a r g m i n}} \\left(f _ {q} \\left(\\theta_ {q}\\right) - \\lambda_ {q} \\left(\\theta_ {q}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 319, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $G(\\theta_q)$ is a monotonic function and hence we can refine $\\theta_q$ iteratively by Gauss-Newton method as", + "bbox": [ + 498, + 359, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta_ {q} ^ {(k + 1)} = \\theta_ {q} ^ {(k)} - G \\frac {\\Delta \\theta_ {q}}{\\Delta G} \\\\ = \\theta_ {q} ^ {(k)} - \\frac {f _ {q} \\left(\\theta_ {q} ^ {(k)}\\right) - \\lambda_ {q} \\left(\\theta_ {q} ^ {(k)}\\right)}{f _ {q} ^ {\\prime} \\left(\\theta_ {q} ^ {(k)}\\right) - \\lambda_ {q} ^ {\\prime} \\left(\\theta_ {q} ^ {(k)}\\right)}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 401, + 888, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_q^{(k)}$ denotes $\\theta_{q}$ of the $k$ -th iteration.", + "bbox": [ + 498, + 467, + 782, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, if we compute the derivative $f_{q}^{\\prime}$ beforehand as shown in the third row of Tab. 1, then this 3D-2D projection can be computed efficiently.", + "bbox": [ + 498, + 484, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multifacet Mirror", + "text_level": 1, + "bbox": [ + 500, + 542, + 674, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As is well known, observing the scene via a multi-facet mirror or a kaleidoscope is identical to observing the scene by virtual multi-view cameras, and in particular, kaleidoscopes with three mirrors are known to be reasonable in terms of less overlaps of mirrored images called discontinuities [29,33]. In our teleidoscopic imaging system, we use a three-facet mirror which defines reflections of the spherical refraction ray-pixel camera introduced in Sec. 3.1.1.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reflection $\\pmb{p}^{\\prime}$ of a 3D point $\\pmb{p}$ by a mirror of normal $\\pmb{n}_i$ and distance $d$ is given by", + "bbox": [ + 500, + 686, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {p} ^ {\\prime} = H _ {i} \\boldsymbol {p} + \\boldsymbol {t} _ {i} \\\\ \\Leftrightarrow \\boldsymbol {p} ^ {\\prime} = \\left(\\boldsymbol {I} - 2 \\boldsymbol {n} _ {i} \\boldsymbol {n} _ {i} ^ {\\top}\\right) \\boldsymbol {p} + 2 d _ {i} \\boldsymbol {n} _ {i}, \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 729, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H_{i}$ the Householder transformation matrix.", + "bbox": [ + 500, + 779, + 828, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the case of kaleidoscopic imaging, the mirrors generate bouncing reflections as shown in Fig. 4. The reflection of $\\pmb{p}^{\\prime}$ by another mirror of normal $\\pmb{n}_j$ and distance $d$ is given simply by", + "bbox": [ + 496, + 795, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {p} ^ {\\prime \\prime} = H _ {i j} \\boldsymbol {p} + \\boldsymbol {t} _ {i j} \\\\ = H _ {j} H _ {i} \\boldsymbol {p} + 2 d _ {j} \\boldsymbol {n} _ {j} + 2 H _ {j} d _ {i} \\boldsymbol {n} _ {i}. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 867, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20816", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/37b04a7540fbc4886328706ad8ad31dcbd2b10f2182fda818b2beb54037eac76.jpg", + "image_caption": [ + "Figure 5. Depth-of-Field with Monocentric Lens. $D_{F}$ and $D_{N}$ denote the depth-of-field of the camera itself without the monocentric lens. The effective depth-of-field of the entire system $L_{F}$ and $L_{N}$ can be obtained by projecting $D_{N}$ and $D_{F}$ through the monocentric lens (the red and blue dashed lines)." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 454, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, our teleidoscopic imaging system can be modeled as multi-view spherical refraction ray-pixel cameras. They share the single monocentric lens, and the pose of each virtual cameras can be computed by Eq. (12) if the mirror parameters are given. The later section describes our calibration algorithm to estimate such parameters.", + "bbox": [ + 75, + 281, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Depth-of-Field of Teleidoscopic Imaging System", + "text_level": 1, + "bbox": [ + 76, + 385, + 470, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section describes analytical evaluations on the depth-of-field of our teleidoscopic imaging system. We first review the depth-of-field of monocentric lens.", + "bbox": [ + 75, + 428, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Depth-of-Field with Monocentric Lens", + "text_level": 1, + "bbox": [ + 76, + 482, + 408, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5, depth-of-field with a monocentric lens can be simply defined as an extension of the path of the thin lens camera (cf. supplemental material). This is because the points in the near and the far depth-of-field $D_N$ and $D_F$ of thin lens are projected within the permissible circle-of-confusion even though the monocentric lens itself introduces the spherical aberration. Hence, we define the depth-of-field with a monocentric lens $L_N$ and $L_F$ as the intersection of the backprojection path through the edge of aperture (the red and blue dash lines in Fig. 5) and the backprojection path through the lens center $o_0$ .", + "bbox": [ + 75, + 506, + 468, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because of the symmetric structure of the monocentric lens, the depth-of-field $L_{N} + L_{F}$ can be deeper than the original depth-of-field $D_{N} + D_{F}$ if the subject distance of the monocentric lens $s_{D} - r_{D}$ is larger than its focal distance $t_{D} - r_{D}$ . That is, there is a trade-off between magnification and depth-of-field even in a monocentric lens. The reason for the deep depth-of-field is due to the significant contribution of the monocentric lens to the far depth of field $L_{F}$ , and thus imaging with less blurring even when designed for high magnification (cf. Sec. 6.3).", + "bbox": [ + 75, + 672, + 470, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Depth-of-Field of Teleidoscopic Imaging System", + "text_level": 1, + "bbox": [ + 76, + 830, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 6 illustrates the near and the far depth-of-fields of a virtual camera $C_i$ defined as a mirror of the original camera.", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a18d5b47c6ffef7520bc7ce301c581cdd35cb2819de0d3f9bdb07ad734e28a10.jpg", + "image_caption": [ + "Figure 6. Depth of Field of Teleidoscopic Imaging System. $D_N$ and $D_F$ denote the near and far depth-of-field of the thin lens camera $C$ respectively. All the points once imaged within $D_N - D_F$ should be originally emitted from the points within $L_N - L_F$ ." + ], + "image_footnote": [], + "bbox": [ + 527, + 92, + 846, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the aperture, the near and the far depth-of-fields form a curve respectively (the red and the blue dashed lines).", + "bbox": [ + 498, + 354, + 890, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the teleidoscopic imaging system has multiple virtual cameras as shown in Fig. 4, each of them has a different depth-of-field according to their relative pose to the monocentric lens. That is, the intersections of such depth-of-fields can be used for multi-view stereo reconstruction.", + "bbox": [ + 498, + 386, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Calibration of Teleidoscopic Imaging System", + "text_level": 1, + "bbox": [ + 500, + 477, + 890, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section introduces our calibration algorithm of teleidoscopic imaging system which requires capturing a single reference planar patterns. Fig. 7 shows the measurement model where the real camera $C$ observes reference points $\\pmb{p}_w$ such as chessboard corners on a reference board II via a monocentric lens and three mirrors. Our calibration estimates the mirror normals $\\pmb{n}_i$ , their distances $d_i$ from the camera, and the position and the radius of the monocentric lens, with assuming that the 2D positions of the reference points $\\pmb{p}_w$ on the reference plane II, the intrinsic parameters of the camera, and the refraction index of the lens are given beforehand.", + "bbox": [ + 496, + 521, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A challenge in this calibration is the fact that the mirrors require observing 3D points and their reflections to estimate their poses [33], while the observation in the teleidoscopic system does not include such mirrored points that follow Eq. (12) due to the refraction by the monocentric lens. Similarly, the rays reflected and then refracted through the projection in teleidoscopic imaging system do not satisfy the coplanarity constraint. In addition, as mentioned in [2], the estimation of the monocentric lens parameters requires multiple viewpoints or multiple monocentric lenses. These points result in a chicken-and-egg problem.", + "bbox": [ + 496, + 703, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve this problem, we utilize the fact that the positions of the centers of the monocentric lens and its mirrors", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20817", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7c5db9c3e0095bcf17f4d780f894beecc7c9bf8ade56c6513dde23420475e205.jpg", + "image_caption": [ + "Figure 7. Calibration of Teleidoscopic Imaging System. Cyan lines denote the light paths for a mirrored camera. The calibration can be conducted by capturing reference points on a plane of a known geometry through the monocentric lens." + ], + "image_footnote": [], + "bbox": [ + 112, + 94, + 433, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are always captured without refraction by definition, and satisfy Eq. (12). Therefore, we start by estimating the center of the monocentric lens from the captured reference points.", + "bbox": [ + 75, + 335, + 468, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In what follows, $\\pmb{p}_p^{(0)}$ denotes the real image of a reference point $\\pmb{p}_w$ in the camera $C$ . Similarly $\\pmb{p}_p^{(i)}$ ( $i = \\{1,2,3\\}$ ) denotes the image of its first reflection by the mirror $i$ , and $\\pmb{p}_p^{(ij)}$ ( $ij = \\{12,13,21,23,31,32\\}$ ) denotes its second reflection by the mirrors $i$ and $j$ .", + "bbox": [ + 76, + 381, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Axis of Monocentric Lens The axis to the center of monocentric lens $\\pmb{a}_v^{(0)}$ from the camera $C$ can be estimated by $\\pmb{p}_w - \\pmb{p}_P^{(0)}$ correspondences. Similarly to [2], we estimate $\\pmb{a}_v^{(0)}$ by the coplanarity constraint of the three vectors $\\pmb{a}_v^{(0)}$ , $\\pmb{p}_p^{(0)}$ , and $\\pmb{p}_w$ on a single plane-of-refraction:", + "bbox": [ + 76, + 470, + 470, + 551 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {p} _ {p} ^ {(0) \\top} \\left(\\boldsymbol {a} _ {v} ^ {(0)} \\times \\left(R _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {t} _ {\\Pi}\\right)\\right) = 0, \\tag {13} \\\\ \\Leftrightarrow \\boldsymbol {p} _ {p} ^ {(0) \\top} \\left(E _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {s} _ {\\Pi}\\right)) = 0, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 559, + 468, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $R_{\\Pi}, t_{\\Pi}$ are the pose of the reference plane in the camera coordinate system and $E_{\\Pi} = \\boldsymbol{a}_v^{(0)} \\times R_{\\Pi}$ and $s_{\\Pi} = \\boldsymbol{a}_v^{(0)} \\times t_{\\Pi}$ . Since this is a linear equation with 9 unknown parameters of $E_{\\Pi}$ and $s_{\\Pi}$ , we can obtain $E_{\\Pi}$ and $s_{\\Pi}$ up to scale by observing at least 8 points on the reference plane. Using the estimated $E_{\\Pi}$ , the axis $\\boldsymbol{a}_v^{(0)}$ is given by", + "bbox": [ + 75, + 614, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {v} ^ {(0)} = \\frac {E _ {\\Pi} (: , 1) \\times E _ {\\Pi} (: , 2)}{| | E _ {\\Pi} (: , 1) \\times E _ {\\Pi} (: , 2) | |}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 718, + 468, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, the axis to the center of the mirrored monocentric lens $\\pmb{a}_v^{(i)}$ can be obtained by the mirrored points $\\pmb{p}_p^{(i)}$ .", + "bbox": [ + 75, + 748, + 468, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mirror Normals The axis to the center of the monocentric lens $\\pmb{a}_{v}^{(0)} = (x_{a_{v}}^{(0)},y_{a_{v}}^{(0)},z_{a_{v}}^{(0)})^{\\top}$ and its mirror $\\pmb{a}_{v}^{(i)} = (x_{a_{v}}^{(i)},y_{a_{v}}^{(i)},z_{a_{v}}^{(i)})^{\\top}$ satisfies", + "bbox": [ + 76, + 787, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {v} ^ {(0) \\top} [ \\boldsymbol {n} _ {i} ] _ {\\times} \\boldsymbol {a} _ {v} ^ {(i)} = 0, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 843, + 468, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $[\\pmb{n}_i]_{\\times}$ denotes the skew-symmetric matrix defined by the normal $\\pmb{n}_i = (x_{n_i},y_{n_i},z_{n_i})^\\top$ of the mirror $i$ .", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The same constraint holds for each of the first-second reflection pairs $\\pmb{a}_v^{(i)} - \\pmb{a}_v^{(ij)}$ about the same mirror normal $\\pmb{n}_i$ [33]. Therefore $\\pmb{n}_i$ can be obtained linearly only from the axes to the centers of the monocentric lenses.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mirror Distances Once the mirror normals are estimated, we can utilize the kaleidoscopic triangulation [33] to obtain linear constraints on the mirror distances $d_{i}$ (See supp. Sec. 2 for details).", + "bbox": [ + 498, + 157, + 890, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose of Reference Plane Similarly to Eq. (13), the plane-of-refraction constraint holds for the mirrored cameras $C_i$ :", + "bbox": [ + 498, + 227, + 890, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {p} _ {p} ^ {(i)}\\right) ^ {\\top} \\left(\\boldsymbol {a} _ {v} ^ {(i)} \\times \\left(H _ {i} \\left(R _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {t} _ {\\Pi}\\right) + \\boldsymbol {t} _ {i}\\right)\\right) = 0. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 279, + 890, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This constraint allows us estimating the pose of the reference plane $R_{\\Pi}$ and $\\pmb{t}_{\\Pi}$ linearly.", + "bbox": [ + 498, + 314, + 890, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Monocentric Lens Parameters The calibration algorithm up to this point does not require the monocentric lens parameters $d_{D}, r_{D}$ , and $\\mu_{g}$ . We estimate these parameters by the coplanarity constraint of the ray through $\\pmb{p}_{p}^{(i)}$ of $C_{i}$ :", + "bbox": [ + 498, + 351, + 890, + 415 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {o} ^ {(i)} \\times \\left(\\boldsymbol {p} _ {w} ^ {(i)} - \\boldsymbol {p} _ {o} ^ {(i)}\\right) = 0. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 426, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This is a nonlinear constraint for the monocentric lens parameters as described in Sec. 3.1 and we solve this as a nonlinear optimization problem with assuming their rough estimates are available in practice.", + "bbox": [ + 498, + 457, + 890, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Bundle Adjustment** The last step of our calibration is to refine the parameters $\\boldsymbol{a}_v^{(0)}, d_D, r_D, \\mu_g, \\boldsymbol{n}_i, d_i, R_{\\Pi}, t_{\\Pi}$ ( $i = 1, 2, 3$ ) by minimizing the reprojection errors of the reference points $\\boldsymbol{p}_w$ as a nonlinear optimization problem. Here the forward projection is computed using the method described in Sec. 3.1.2.", + "bbox": [ + 498, + 525, + 890, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the teleidoscopic imaging system parameters are calibrated, a scene point can be linearly triangulated in a DLT manner [13] from its teleidoscopic projections (See supp. Sec. 3 for details).", + "bbox": [ + 498, + 618, + 890, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 691, + 614, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section provides quantitative evaluations of the proposed numerical 3D-2D projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also demonstrate the effectiveness of our method by reconstructing real-world microscale objects.", + "bbox": [ + 498, + 717, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Computational Efficiency of 3D-2D Projection", + "text_level": 1, + "bbox": [ + 500, + 801, + 888, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The average computation cost of a single forward projection by our method and by the analytical solution solving a 10th-degree equation for each projection [2] are $0.62\\mathrm{ms}$ and $3.93\\mathrm{ms}$ respectively. These runtime costs are the average values of 100 trials of 10K points forward projections", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20818", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03031ee35174fe2150d6e749dfa273a836c69c3dcdc6d5503a054d2751f0ec68.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 272, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de6b2965b32579d49160ab7c99ead40090053d152a4407775fd497c3b08e5d06.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 281, + 90, + 455, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b7d6ec1ded5a65e174362d74ed5a867250a9f86437daa10a7e95ee4602f41a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 286, + 200, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93b89c7865b2d4e25fe093997d934f490fdc280288507f4a9faf76efc13829f7.jpg", + "image_caption": [ + "Figure 8. Teleidoscopic System for Quantitative Evaluation. (a) Simulated measurement environment. (b) Reprojection errors at different noise levels. The bars denote the standard deviation of the errors." + ], + "image_footnote": [], + "bbox": [ + 205, + 285, + 338, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/804146f7ae0a0ab5d7205f8da0c8f32e9aa91c41ea8ebb537a1c9d6ed263fdca.jpg", + "image_caption": [ + "Figure 11. Changes in Depth-of-Field by the Monocentric Lens. (a), (b) and (c) show the total, near, and far depth-of-field with a monocentric lens. The red and blue plots indicate the results with and without the monocentric lens respectively. The green lines indicate the subject distance where the subject and the focal distances of the monocentric lens itself are same." + ], + "image_footnote": [], + "bbox": [ + 341, + 285, + 462, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbe50aa59c1daf80a7c3d094de19608c47dab964072df6bb9eda6568d254e337.jpg", + "image_caption": [ + "(a)", + "Figure 9. Calibration Errors. From left to right, the upper figures show the (a) angular error of the axis to the monocentric lens center (degree), (b) angular error of the mirror normal (degree), and (c) rotation error (Riemannian distance) of reference plane. The lower ones are the distance errors normalized by the ground truth. Here the distance of the monocentric lens is only estimated in a nonlinear manner (Sec. 5)." + ], + "image_footnote": [], + "bbox": [ + 81, + 396, + 200, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37b0a7a601c404758d8d79bdf8e2b25cadb137d39fe8be824a9e41cae583508d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 202, + 400, + 338, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23d0b9f4d92190129dcd0363d47aa707486abb095172c74d10ad44e25a2d06d8.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 341, + 396, + 460, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "run in Matlab on an Intel Core-i7 2.6GHz PC. This result clearly verifies that our numerical forward projection based on our ray-pixel model can realize an efficient projection.", + "bbox": [ + 75, + 632, + 468, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Calibration", + "text_level": 1, + "bbox": [ + 76, + 689, + 202, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.1 Quantitative Evaluation using Synthesized Data", + "text_level": 1, + "bbox": [ + 75, + 713, + 464, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8(a) shows the measurement environment which simulates the real capture system used in Sec. 6.2.2. The system has a kaleidoscope with three $10 \\times 30\\mathrm{mm}$ mirrors in front of the camera $C$ . The mirrors are at slightly off-perpendicular angle of $1.4^{\\circ}$ to the camera image plane so that the mirrors define virtual cameras around the target with less overlaps of the mirrored images. The system also has a monocentric lens of $10\\mathrm{mm}$ diameter in front of the mirrors, at $40\\mathrm{mm}$ distance from the camera. The refraction index $\\mu_{g}$ is set to 2.0. The system captures 48 reference points (blue dots in Fig. 8(a)) to calibrate its parameters. By injecting Gaussian", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9859d492b60fb7b88a47003e31d41c09930e0848df079b3586fc222e6d53d8ed.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 700, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cc7d065929938b3447e28105a16ae5bc9b9424c6118b50f63dad7e3555ea0e4.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 705, + 90, + 877, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce54496830517d8b910b8e59a4d940185f48f7a08350852b88d66a965e05be5b.jpg", + "image_caption": [ + "Figure 10. Experimental Setup and Calibration. (a) Our teleidoscopic imaging system consists of a camera, a projector, three mirrors $(10 \\times 30\\mathrm{mm})$ , and a monocentric lens (5mm radius, refraction index $\\mu_g = 2.0$ ). (b) Input image for calibration.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 500, + 277, + 632, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3546a9cce26579489d9b7920f9b824b83ab86a45d15f57a15cf7575ef81b8588.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 635, + 277, + 763, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2016e58bc7025fcec3e565772d928baaa8617c757bdcaca3c43efb5256c6c8b0.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 761, + 277, + 892, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "noise of different standard deviations $\\sigma$ to the 2D positions of their projections, we evaluate the robustness of our calibration procedure. Fig. 8(b) shows average reprojection errors in pixel of 100 trials at each pixel noise level $\\sigma$ . We can observe that the reprojection errors increase linearly against the pixel noise level.", + "bbox": [ + 496, + 474, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 9 shows the estimation errors of the monocentric lens parameters, the mirror parameters, and the reference plane parameters respectively. These results indicate that our calibration algorithm performs reasonably under realistic observation noise.", + "bbox": [ + 496, + 566, + 890, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.2 Quantitative Evaluation using Real Data", + "text_level": 1, + "bbox": [ + 500, + 651, + 834, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 10(a) shows an overview of our teleidoscopic imaging system. The system consists of a FLIR Blackfly BFS-U3-89S6C-C camera (4096×2160 resolution, pixel size 3.45μm) with a f=4mm / F8 lens, three 10 × 30mm mirrors, a monocentric lens of 10mm diameter whose refraction index is $\\mu_g = 2.0$ , and a LED projector ASUS E1-R for casting Gray code patterns onto the target surface to identify corresponding points [12, 23]. Note that the effective resolution of each \"virtual camera\" image is approximately 300×200 (16 \"virtual cameras\" in total), which is calculated from the area inside the base chamber with one horizontal edge of the triangle and inside the spherical lens.", + "bbox": [ + 496, + 672, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We used a single image (Fig. 10(b)) capturing $5 \\times 8$ corners of a checkerboard for calibration, and used up to second reflections in the image. The average reprojection er", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20819", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e84516d714b8c73c2142fd0f160d7400f800b813363becc3af92ef5974be9c2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 106, + 184, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29442c3ae27b96808aed4d56cdb952e1993fe90ac6d6daeecb0af5381f12ef1f.jpg", + "image_caption": [ + "(b)", + "Figure 12. Evaluation with a sphere (radius: $3\\mathrm{mm}$ ). (a) Ground truth. (b) Recovered 3D shape. The colors visualize the $z$ -coordinate of the reconstructed shape. (c) Error map. $E_{d}$ is the mean absolute error in millimeter." + ], + "image_footnote": [], + "bbox": [ + 196, + 107, + 284, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86fae6deb7684254fb2f0ca1c2117761dbe701f5eacedc374726bc50e95af33f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 90, + 336, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a41e30a812ea681b3cd2731f5197c9da69cb242e354fa8752fbeb895161e04c6.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 331, + 90, + 418, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea5c35744b6ecea20b8f6dad9641c08b4065274b818ed69f9de81f5115219d36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 90, + 460, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ror is 0.62px, and the average distance error of the triangulated 3D points to their ground truth positions was $0.17\\mathrm{mm}$ . The size of reference checkerboard is around $15\\mathrm{mm}$ width, hence the relative error of the triangulation is around $1\\%$ . These results show our method provide a practical calibration and the valid scale of the target object also depends on size of the checkerboard.", + "bbox": [ + 75, + 236, + 467, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Depth-of-Field", + "text_level": 1, + "bbox": [ + 76, + 354, + 225, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 11 shows the changes in the depth-of-field by the monocentric lens capturing an object of $0.5\\mathrm{mm}$ height at the same magnification given the configuration same as the real capture in Sec. 6.2.2. Fig. 11(a) shows the total depth-of-field with and without the monocentric lens $L_{N} + L_{F}$ (red and blue respectively). Fig. 11(b) and (c) show $L_{N}$ and $L_{F}$ in the same manner. In these figures, the horizontal axis indicates the apparent size of the target in pixel, with assuming the measurement model used in Sec. 6.2. These plots clearly indicate that the monocentric lens contributes to increase the depth-of-field successfully.", + "bbox": [ + 75, + 378, + 468, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Teleidoscopic 3D Shape Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 556, + 413, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the proposed teleidoscopic system as a multiview camera system for 3D shape reconstruction, this section demonstrates 3D reconstruction of small objects of approximately $6\\mathrm{mm}$ size.", + "bbox": [ + 75, + 580, + 468, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We used the system calibrated in Sec. 6.2.2. Fig. 12 visualizes our 3D reconstruction of the front side of a real sphere of $3\\mathrm{mm}$ radius with the ground-truth and the error map. The mean absolute error $E_{d}$ of the estimated depth is $0.097\\mathrm{mm}$ . Compared with its size, the relative error is about $1.6\\%$ .", + "bbox": [ + 75, + 642, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 13 shows additional 3D reconstructions. The 1st and the 2nd rows demonstrate that our imaging system has the ability to recover complex shapes. The proposed imaging system well recovers the complex bumpy parts of the shell and the sharp edges and corners of the object. The 2nd and the 3rd rows compare \"with\" and \"w/o\" the monocentric lens cases. As the input image (b) shows, the monocentric lens creates a surrounding view while preserving the area of each chamber, resulting in a fuller and more accurate geometry, as shown in (c). We can verify this quantitatively by comparing the numbers of pixels covering the object area in the captured image (145,217 vs. 12,730) and the numbers", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e840fa03f603c82fcaf928e8bdf7c04a9544cc1b12cf990cf12606354ade632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 126, + 573, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c70bd838d0cd6cdab51399b0df5a82a973c9cf33474578042c407494162d3c4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 74, + 684, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1190d819cbdbc82425c5883254016964c4fff978aa3088bec2a748233951388c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 147, + 684, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08b5901c897b76425a4eb169698d777219de370683da0de430003ab576c992b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 93, + 766, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9713e13f344a192e93d1f4bf477753e01a2f46be450fda7b6def1b6ba5523c44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 92, + 851, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d76b5caa8d46a9081d1b61545d94e610dcb7e85015a6ec5ffed84de36963e6ef.jpg", + "image_caption": [ + "40,639 pts recovered" + ], + "image_footnote": [], + "bbox": [ + 705, + 160, + 774, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36f1a76d08c9302bf58ab2e161b52e9e87a3cce888471e1f2b586db192f1cbeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 160, + 859, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/011618e35a7c6f8b9d23e0ca58837101d5f693a6e5b266310edfbf05dc51cfe2.jpg", + "image_caption": [ + "w/o monocentric lens", + "(b)", + "Figure 13. Results of 3D Shape Estimation. The second and the third row compares \"with\" and \"w/o\" the monocentric lens. (a) Room-light appearance. (b) Teleidoscopic image. (c) Reconstructed 3D shapes." + ], + "image_footnote": [], + "bbox": [ + 580, + 233, + 684, + 304 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c88bbca32efa14ac1fbdf8f46b9490602bbb756aed8546a90df57fd3cdcf9592.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 253, + 777, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/444d6283fdb898dd71ccf9d0bff27a2e3e7237613ae2a3cb5e603ab77c458349.jpg", + "image_caption": [ + "8,264 pts recovered", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 777, + 253, + 859, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of triangulated 3D points using the ground-truth correspondences as shown in the figure (40,639 vs. 8,264). See the supplementary material for more results.", + "bbox": [ + 496, + 388, + 890, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations The major limitations of our method are twofold. The monocentric lens has a relatively large spherical aberration, and the optical resolution of the entire system is limited by such aberrations. One possible solution is to introduce nested concentric glass spheres with different refractive indices for the monocentric lens [6, 35]. The other limitation is the trade-off between the number of virtual viewpoints and the effective resolution of each view because the use of a kaleidoscope requires zooming-out to capture the reflected images. While larger number of the cameras contributes to robustify the triangulation, it inevitably reduces the resolution of the reconstruction as the number of pixels per view decreases.", + "bbox": [ + 496, + 436, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 640, + 617, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposed the concept of image-based 3D measurement by a teleidoscopic imaging system for microscale objects. By combining a monocentric lens with a kaleidoscope, our system realized a closeup and surrounding multi-view capture of the target. Our contributions are twofold. We proposed a ray-pixel camera model that realized an efficient numerical 3D-2D projection computation, and also proposed a calibration method for the teleidoscopic imaging system. Our future work includes aberration and diffraction blur handling such as the inverse filter proposed in [5], and also refraction/attenuation modeling in the case of capturing objects in water [18, 19].", + "bbox": [ + 496, + 661, + 890, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 856, + 666, + 871 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is in part supported by JSPS KAKENHI JP22K17914.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20820", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rajat Aggarwal, Amrisha Vohra, and Anoop M. Namboodiri. Panoramic stereo videos with a single camera. In CVPR, pages 3755-3763, 2016. 2", + "[2] Amit Agrawal and Srikumar Ramalingam. Single image calibration of multi-axial imaging systems. In CVPR, pages 1399–1406, 2013. 2, 3, 4, 5, 6", + "[3] Byeongjoo Ahn, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Kaleidoscopic structured light. ACM TOG, 40(6):1-15, 2021. 2", + "[4] Michael Chen, Zachary F. Phillips, and Laura Waller. Quantitative differential phase contrast (dpc) microscopy with computational aberration correction. Optics Express, (25):32888-32899. 1", + "[5] O. S. Cossairt, D. Miau, and S. K. Nayar. Gigapixel computational imaging. In ICCP, pages 1-8, 2011. 2, 8", + "[6] Donald G. Dansereau, Glenn Schuster, Joseph Ford, and Gordon Wetzstein. A wide-field-of-view monocentric light field camera. In CVPR, 2017. 2, 8", + "[7] Michael De Zeeuw and Aswin C Sankaranarayanan. Widebaseline light fields using ellipsoidal mirrors. 2022. 1", + "[8] Keith Forbes, Fred Nicolls, Gerhard De Jager, and Anthon Voigt. Shape-from-silhouette with two mirrors and an uncalibrated camera. In ECCV, 2006. 2", + "[9] Brian J. Ford. Single lens: The story of the simple microscope. Journal of the History of Biology, 19(2):320-321, 1986. 2", + "[10] Joshua Gluckman, Shree K. Nayar, and Keith J. Thoresz. Real-time omnidirectional and panoramic stereo. In Proceedings of the 1998 DARPA Image Understanding Workshop, pages 299-303. Morgan Kaufmann, 1998. 2", + "[11] Michael D. Grossberg and Shree K. Nayar. The raxel imaging model and ray-based calibration. *IJCV*, 61(2):119–137, 2005. 1, 2", + "[12] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G. Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR, pages 713-720, 2011. 7", + "[13] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 2, 6", + "[14] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE TPAMI, 41(1):190-204, 2019. 1", + "[15] Srinivasa Karthik and Ambasamudram N. Rajagopalan. Underwater microscopic shape from focus. In ICPR, pages 2107-2112, 2014. 1", + "[16] Ryo Kawahara, Shohei Nobuhara, and Takashi Matsuyama. A pixel-wise varifocal camera model for efficient forward projection and linear extrinsic calibration of underwater cameras with flat housings. In Proc. of ICCV 2013 Underwater Vision Workshop, pages 819-824, 2013. 2", + "[17] Gurunandan Krishnan and Shree K. Nayar. Towards a true spherical camera. In SPIE, volume 7240, 724002, pages 1-13, 2009. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Meng-Yu Jennifer Kuo, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Non-rigid shape from water. IEEE TPAMI, 43(7):2220-2232, 2021. 8", + "[19] Meng-Yu Jennifer Kuo, Satoshi Murai, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Surface normals and shape from water. IEEE TPAMI, 44(12):9150-9162, 2021. 8", + "[20] Douglas Lanman, Daniel Crispell, and Gabriel Taubin. Surround structured lightning: 3-d scanning with orthographic illumination. In CVIU, pages 1107-1117, 2009. 2", + "[21] A. Manakov, J. Restrepo, O. Klehm, R. Hegedus, E. Eisemann, H. P. Seidel, and I. Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM TOG, 32(4):47-1, 2013. 1", + "[22] P. Miraldo and H. Araujo. Calibration of smooth camera models. IEEE TPAMI, 35(9):2091-2103, 2013. 2", + "[23] Daniel Moreno and Gabriel Taubin. Simple, accurate, and robust projector-camera calibration. In 3DPVT, pages 464-471, 2012. 7", + "[24] Andrew D. Mullen, Tali Treibitz, Paul LD Roberts, Emily L. A. Kelly, Rael Horwitz, Jennifer E. Smith, and Jules S. Jaffe. Underwater microscopy for in situ studies of benthic ecosystems. Nature Communications, 7, 2016/07 2016. 1", + "[25] Shree K. Nayar. Omnidirectional video camera. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 235-241, 1997. 2", + "[26] Markus Niederost, Jana Niederost, and Jiri Scucka. Automatic 3d reconstruction and visualization of microscopic objects from a monoscopic multifocus image sequence. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XXXIV-5/W10:5-10, 01 2002. 1", + "[27] Ryoichiro Nishi, Takahito Aoto, Norihiko Kawai, Tomokazu Sato, Yasuhiro Mukaigawa, and Naokazu Yokoya. *Ultrashallow dof imaging using faced paraboloidal mirrors*. In ACCV, pages 115–128, 2016. 2", + "[28] Ilya Reshetouski and Ivo Ihrke. Mirrors in Computer Graphics, Computer Vision and Time-of-Flight Imaging, pages 77–104. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013.", + "[29] Ilya Reshetouski, Alkhazur Manakov, Hans-Peter Seidel, and Ivo Ihrke. Three-dimensional kaleidoscopic imaging. In CVPR, pages 353-360, 2011. 4", + "[30] J. Starck, A. Maki, S. Nobuhara, A. Hilton, and T. Matsuyama. The multiple-camera 3-d production studio. IEEE TCSVT, 19(6):856-869, 2009. 1", + "[31] Peter Sturm. Multi-view geometry for general camera models. In CVPR, volume 1, pages 206-212, 2005. 2", + "[32] Seiichi Tagawa, Yasuhiro Mukaigawa, Jaewon Kim, Ramesh Raskar, Yasuyuki Matsushita, and Yasushi Yagi. Hemispherical confocal imaging. IPSJ Transactions on Computer Vision and Applications, 3:222-235, 2011. 2", + "[33] Kosuke Takahashi, Akihiro Miyata, Shohei Nobuhara, and Takashi Matsuyama. A linear extrinsic calibration of kaleidoscopic imaging system from single 3d point. In CVPR, 2017. 1, 2, 4, 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20821", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Lei Tian, Jingyan Wang, and Laura Waller. 3d differential phase-contrast microscopy with computational illumination using an led array. Optics letters, 39:1326-9, 2014. 1", + "[35] Eric J Tremblay, Daniel L Marks, David J Brady, and Joseph E Ford. Design and scaling of monocentric multiscale imagers. Applied Optics, 51(20):4691-4702, 2012. 2, 8", + "[36] Borislav Trifonov, Derek Bradley, and Wolfgang Heidrich. Tomographic reconstruction of transparent objects. In Proc. of Eurographics Conference on Rendering Techniques, pages 51-60, 2006. 2", + "[37] David Uhlig and Michael Heizmann. A calibration method for the generalized imaging model with uncertain calibration target coordinates. In ACCV. 2", + "[38] Frank Verbiest, Marc Proesmans, and Luc Van Gool. Modeling the effects of windshield refraction for camera calibration. In ECCV, pages 397-412, 2020. 2" + ], + "bbox": [ + 78, + 90, + 467, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20822", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_model.json b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9a7619fe74c701d12ec7bfcc72264f145f9a6a2 --- /dev/null +++ b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_model.json @@ -0,0 +1,2860 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.843, + 0.154 + ], + "angle": 0, + "content": "Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.182, + 0.384, + 0.234 + ], + "angle": 0, + "content": "Ryo Kawahara \nKyushu Institute of Technology \nrkawahara@ai.kyutech.ac.jp" + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.182, + 0.621, + 0.233 + ], + "angle": 0, + "content": "Meng-Yu Jennifer Kuo University of Minnesota kuo00052@umn.edu" + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.182, + 0.834, + 0.234 + ], + "angle": 0, + "content": "Shohei Nobuhara Kyoto University nob@i.kyoto-u.ac.jp" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.245, + 0.357, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.244, + 0.591, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.244, + 0.782, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.244, + 0.885, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 1. Teleidoscopic Imaging System. (a) We propose a novel system for microscale 3D imaging using an imaging setup that consists of a camera, a kaleidoscopic multi-facet mirror with a monocentric front lens, and a projector. (b) Our system realizes a dense close-up surrounding multi-view capture of the target. (c) Our method enables fuller and dense reconstruction of microscale objects." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.448, + 0.314, + 0.463 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.472, + 0.722 + ], + "angle": 0, + "content": "This paper proposes a practical method of microscale 3D shape capturing by a teleidoscopic imaging system. The main challenge in microscale 3D shape reconstruction is to capture the target from multiple viewpoints with a large enough depth-of-field. Our idea is to employ a teleidoscopic measurement system consisting of three planar mirrors and monocentric lens. The planar mirrors virtually define multiple viewpoints by multiple reflections, and the monocentric lens realizes a high magnification with less blurry and surround view even in closeup imaging. Our contributions include, a structured ray-pixel camera model which handles refractive and reflective projection rays efficiently, analytical evaluations of depth of field of our teleidoscopic imaging system, and a practical calibration algorithm of the teleidoscopic imaging system. Evaluations with real images prove the concept of our measurement system." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.208, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Microscale 3D reconstruction has found profound applications in a wide range of domains including medical imaging, life science, and aquaculture, due to its non-constrained and non-invasive measurements. The main challenges in image-based microscopic 3D shape measurement is its shallow depth of field and camera arrangement in the closeup scenario. Applying conventional multiple camera system designed for human-size capture [14, 30] cannot be a feasible solution due to limitations on camera placement. Con-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.495 + ], + "angle": 0, + "content": "ventional multiple mirror system [33] also have difficulties inevitably in depth-of-focus due to differences in their optical path lengths with varying numbers of bounces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.818 + ], + "angle": 0, + "content": "In this paper, we show that the fuller 3D shape of a microscale object can be recovered. Our key idea is to employ a catadioptric imaging system which realizes a practical closeup multi-view imaging. The point of our design is that the system has a monocentric front lens like a teleidoscope, instead of using microscopic system in the camera side. That is, as shown in Fig. 1(a), we introduce a kaleidoscopic multi-facet mirror between the front lens and the camera, where the design realizes a deeper depth-of-field and results in less blurring imaging. Unlike conventional microscopic imaging system such as differential phase contrast (DPC) microscopy [4, 34] and multi-focus approaches [15, 24, 26], our method realizes a multi-view capture of the target from a single physical viewpoint which can contribute to free-viewpoint rendering, 3D shape reconstruction, and reflection analysis. Our system also has an advantage over existing imaging systems that build multiple views behind the main lens [7, 21] in a closeup environment. The wide FoV with a monocentric main lens and virtual multiple views with mirrors allow closeup and surrounding view capturing in focus." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We call our system teleidoscopic imaging system and show that the system can be compactly modeled by a structured ray-pixel camera model [11], which handles refractive and reflective projection rays efficiently. Based on our ray-pixel camera model, we derive a practical calibration algo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20813" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "rithm that estimates the positions of the monocentric lens and multi-facet mirror w.r.t. the camera by using a single reference planar pattern (i.e., a checkerboard). Given the calibration parameters, a scene point can then be linearly triangulated from its teleidoscopic projection in a direct linear transform (DLT) manner [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.378 + ], + "angle": 0, + "content": "We implement our method with an imaging system consisting of a camera, three planar mirrors, a monocentric lens, and a projector placed outside the mirrors for structured-light casting. We quantitatively evaluate the computation cost of numerical projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also validate the effectiveness of our method qualitatively on a number of real-world microscale objects. These results demonstrate that our method enables holistic and dense reconstruction of microscale objects. We believe our method expands the avenues of three-dimensional analysis of microscale objects and scenes in real world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.391, + 0.219, + 0.406 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.468, + 0.506 + ], + "angle": 0, + "content": "Our teleidoscopic imaging system is a catadioptric system consisting of a monocentric lens and kaleidoscopic mirrors (Fig. 1(a)). While catadioptric system has a wide variety of applications such as omnidirectional observation and panoramic stereo [1, 10, 25, 27], this section reviews studies with multiple flat mirrors and monocentric lenses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.387, + 0.53 + ], + "angle": 0, + "content": "2.1. Imaging with Multiple Flat Mirrors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.673 + ], + "angle": 0, + "content": "A fundamental motivation of introducing mirrors in observation system is to increase the number of viewpoints without installing additional cameras for multi-view capture of a target [3, 8, 20, 28, 32, 33]. Takahashi et al. [33] have proposed a kaleidoscopic imaging system and demonstrated a 3D shape reconstruction using multiple reflections. Tagawa et al. [32] have proposed a multi-facet imaging system that observes a target from an equally distributed virtual cameras for reflectance analysis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.36, + 0.697 + ], + "angle": 0, + "content": "2.2. Imaging with Monocentric Lens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.87 + ], + "angle": 0, + "content": "In the context of imaging system, the monocentric lens is often used to obtain a wide field-of-view [17, 35], in particular for endoscopes, as a short focal length lens. The monocentric lens, however, has additional useful characteristics: its symmetric structure and magnifying power. Cossair et al. [5] have proposed a camera array which captures a same scene through a single monocentric lens so that the images can be stitched into a single high-resolution image. Similarly, Dansereau et al. [6] have proposed a lightfield camera which captures omnidirectional lightfield images through a single monocentric lens with a camera orbiting around it." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "A typical use of the monocentric lens as a magnifier can be found in the Leeuwenhoek's microscope in the 17th cen" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.092, + 0.885, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.258, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Measurement through Monocentric Lens. The segments of projection path \\(\\ell_{o} - \\ell_{g} - \\ell_{i}\\) have an axially symmetric structure around the axis which directs to the monocentric lens center \\(o_{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.318, + 0.892, + 0.348 + ], + "angle": 0, + "content": "tury. It utilizes a single monocentric lens and realized over \\(100 \\times\\) magnifications [9]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.73, + 0.375 + ], + "angle": 0, + "content": "2.3. Ray-pixel Camera Model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.61 + ], + "angle": 0, + "content": "In order to handle light paths captured via complex imaging system such as reflection and refraction, Grossberg and Nayar [11] have proposed a ray-pixel (or raxel) camera model which describes the projection by end-to-end mappings between the pixels and their corresponding emergent rays. While this pixel-wise representation has a great flexibility to describe complicated projections [22, 31], it is also possible to exploit a structure of the rays to simplify the model [16, 22, 36-38]. Kawahara et al. [16] have proposed a ray-pixel camera model for cameras observing the scene through flat surfaces. Compared with analytical forward projections through flat surfaces which requires solving a high order equation [2], it realized a practical numerical forward projection computation by exploiting the axially symmetric structure of the rays." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.625, + 0.891, + 0.66 + ], + "angle": 0, + "content": "3. Ray-Pixel Camera Model for Teleidoscopic Imaging System" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.669, + 0.892, + 0.744 + ], + "angle": 0, + "content": "In our teleidoscopic system, the cameras mirrored by the kaleidoscopic multi-facet mirrors observe the scene through the single monocentric lens (Fig. 1(a)) as done in [5,6], and each of the monocentric lens and the camera pairs form a microscope capturing the object from a different direction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.745, + 0.891, + 0.806 + ], + "angle": 0, + "content": "In this section, we show that this design realizes a microscale multi-view capturing with a deep depth-of-field, and that the projections by the teleidoscopic imaging system can be modeled compactly by ray-pixel cameras." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.851, + 0.832 + ], + "angle": 0, + "content": "3.1. Measurement through Monocentric Lens" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, suppose a perspective camera whose camera center is located at \\( \\pmb{o}_0 \\) observes the scene through a monocentric lens located at \\( \\pmb{o}_D \\). Here the perspective camera can be assumed to be directed to the center" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20814" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "of the monocentric lens \\( \\mathbf{o}_D \\) without loss of generality, since we can calibrate the rotation to align the optical axis of the camera to the line \\( \\mathbf{o}_D - \\mathbf{o}_0 \\) as described later. We denote this normalized camera by \\( C_n \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.468, + 0.246 + ], + "angle": 0, + "content": "Obviously, the rays backprojected through pixels of \\( C_n \\) have an axially symmetric structure around the optical axis, and an incident ray \\( \\ell_i \\) through a pixel \\( p_p \\) and its refraction \\( \\ell_g \\) and emergent ray \\( \\ell_o \\) are on a single plane-of-refraction [2]. Therefore, we can describe the ray through a single pixel by a 2D \\( (r,z)^{\\top} \\) coordinate system centered at \\( o_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.248, + 0.468, + 0.338 + ], + "angle": 0, + "content": "Hereafter, let \\( r_{\\alpha} \\) and \\( z_{\\alpha} \\) denote the \\( r \\) and \\( z \\) elements of the vector \\( \\alpha \\) in general. For example, point \\( \\pmb{p}_p \\) is described as \\( \\pmb{p}_p = (r_{\\pmb{p}_p}, z_{\\pmb{p}_p})^\\top \\). Also let \\( \\pmb{v}_X = (r_{\\pmb{v}_X}, z_{\\pmb{v}_X})^\\top \\) denote the direction vector of line \\( \\ell_X \\) towards the scene from the camera, where \\( X = \\{\\mathrm{i}, \\mathrm{g}, \\mathrm{o}\\} \\) represents the label of the light path segment." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.468, + 0.388 + ], + "angle": 0, + "content": "Consider the projection path \\(\\ell_o - \\ell_g - \\ell_i\\) from \\(o_0\\) through the point \\(\\pmb{p}_p\\). The incident point \\(\\pmb{p}_i\\) on the sphere is described as a function of \\(\\theta_p = \\tan^{-1}(r_{\\pmb{p}_p} / f_c)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.403, + 0.468, + 0.433 + ], + "angle": 0, + "content": "\\[\nz _ {p _ {i}} = \\frac {d _ {D} - \\sqrt {d _ {D} ^ {2} - (1 + \\tan^ {2} \\theta_ {p}) \\left(d _ {D} ^ {2} - r _ {D} ^ {2}\\right)}}{1 + \\tan^ {2} \\theta_ {p}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.434, + 0.269, + 0.45 + ], + "angle": 0, + "content": "\\[\nr _ {p _ {i}} = z _ {p _ {i}} \\tan \\theta_ {p},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.465, + 0.468, + 0.493 + ], + "angle": 0, + "content": "where \\(d_{D}\\) and \\(r_D\\) are the distances to the monocentric lens and the lens radius." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.468, + 0.541 + ], + "angle": 0, + "content": "The refraction angle \\(\\theta_{g}\\) is then given by Snell's law with assuming the refractive indices of the air \\(\\mu_{a}\\) and the lens \\(\\mu_{g}\\) are known:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.556, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {g} \\sin \\theta_ {g} = \\mu_ {a} \\sin \\left(\\theta_ {p} + \\theta_ {D i}\\right), \\\\ \\Leftrightarrow \\sin \\theta_ {g} = \\frac {\\mu_ {a}}{\\mu_ {g}} \\frac {d _ {D} r _ {p _ {i}}}{r _ {D} \\sqrt {r _ {p _ {i}} ^ {2} + z _ {p _ {i}} ^ {2}}}, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.62, + 0.318, + 0.635 + ], + "angle": 0, + "content": "where \\(\\theta_{D_i}\\) is the angle of \\(o_0 - o_D - p_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.637, + 0.468, + 0.756 + ], + "angle": 0, + "content": "The point of emergence \\( \\pmb{p}_o \\) is the intersection of the path \\( \\ell_g \\) and the sphere. It is also obtained as the mirror of the point \\( \\pmb{p}_i \\) because of the symmetrical relationship between incident path \\( \\ell_i \\) and outgoing \\( \\ell_o \\) w.r.t. the monocentric lens. As shown in Fig. 2, \\( \\ell_i \\) and \\( \\ell_o \\) are in line symmetry to the line that is perpendicular to the vector \\( \\pmb{n}_p \\) through the point \\( \\pmb{o}_D \\). That is, given the point \\( \\pmb{p}_i \\), its reflection \\( \\pmb{p}_o \\) is described as" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.772, + 0.468, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {p} _ {o} = H _ {p} \\boldsymbol {p} _ {i} + \\boldsymbol {t} _ {p}, \\tag {3} \\\\ \\Leftrightarrow \\boldsymbol {p} _ {o} = \\left(\\boldsymbol {I} - 2 \\boldsymbol {n} _ {p} \\boldsymbol {n} _ {p} ^ {\\top}\\right) \\boldsymbol {p} _ {i} + 2 (\\boldsymbol {o} _ {D} ^ {\\top} \\boldsymbol {n} _ {p}) \\boldsymbol {n} _ {p}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.824, + 0.468, + 0.856 + ], + "angle": 0, + "content": "where \\(H_{p}\\) is the Householder matrix and \\(\\pmb{t}_{p}\\) denotes the center of reflection. Besides, the direction \\(\\pmb{n}_p\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.871, + 0.468, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {n} _ {p} = \\left[ \\begin{array}{l} \\cos \\left(\\pi / 2 + \\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\sin \\left(\\pi / 2 + \\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right] = \\left[ \\begin{array}{c} - \\sin \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\cos \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right], \\tag {4}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.101, + 0.868, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 3. 3D-2D Projection. The black line \\(\\ell_{o}\\) is the correct projection path that intersects with the optical axis at the camera center \\(o_0\\). Our 3D-2D projection hypothesizes a line \\(\\ell_{\\lambda}\\) as an initial guess of projection, and then optimizes it by verifying if the point and the angle of intersection with the optical axis is consistent with the ray-pixel mapping \\(\\langle \\theta_q, f_q \\rangle\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.739, + 0.355 + ], + "angle": 0, + "content": "therefore \\(H_{p}\\) and \\(\\pmb{t}_{p}\\) are rewritten to" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.365, + 0.892, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} H _ {p} = \\left[ \\begin{array}{l l} \\cos 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) & \\sin 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\sin 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) & - \\cos 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right], \\tag {5} \\\\ \\boldsymbol {t} _ {p} = \\left[ \\begin{array}{c} - d _ {D} \\sin 2 (\\theta_ {D i} - \\theta_ {g}) \\\\ d _ {D} \\cos 2 (\\theta_ {D i} - \\theta_ {g}) + d _ {D} \\end{array} \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.444, + 0.87, + 0.459 + ], + "angle": 0, + "content": "The direction \\(\\pmb{v}_o\\) is also given by Householder matrix as" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.472, + 0.89, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {o} = H _ {p} \\boldsymbol {v} _ {i}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.499, + 0.89, + 0.531 + ], + "angle": 0, + "content": "As a result, the intersection \\( \\pmb{o}_{f_q} = (0,f_q)^\\top \\) of the ray \\( \\ell_o \\) and the optical axis is given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.544, + 0.89, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {o} _ {f _ {q}} = t _ {o} \\boldsymbol {v} _ {o} + \\boldsymbol {p} _ {o}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.562, + 0.89, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\Leftrightarrow t _ {o} = - \\frac {r _ {\\boldsymbol {p} _ {o}}}{r _ {\\boldsymbol {v} _ {o}}}, \\quad f _ {q} = z _ {\\boldsymbol {p} _ {o}} + \\left(\\frac {- r _ {\\boldsymbol {v} _ {o}}}{z _ {\\boldsymbol {v} _ {o}}}\\right) ^ {- 1} r _ {\\boldsymbol {p} _ {o}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.722 + ], + "angle": 0, + "content": "This equation describes the ray \\(\\ell_{o}\\) backprojecting from the pixel \\(p_p\\) as a function of \\(\\theta_{p}\\). Inversely, the forward projection can be obtained for a given 3D point in front of the monocentric lens. The analytical 3D-2D projection, however, requires solving a 10th-degree equation [2]. The next section introduces a ray-pixel camera which exploits the axial symmetric structure of the rays to provide an efficient numerical 3D-2D projection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.73, + 0.879, + 0.745 + ], + "angle": 0, + "content": "3.1.1 Spherical Refraction Ray-Pixel Camera Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Up to this point, we observed that the rays emitted from a perspective camera through a monocentric lens show an axially symmetric structure around the line from the camera to the lens center. As illustrated in Fig. 2, given a pixel \\( p_p \\) by specifying \\( \\theta_p \\), the corresponding ray \\( \\ell_o \\) can be determined uniquely by Eq. (8). This section introduces our spherical refraction ray-pixel camera model by representing the ray-pixel mapping as follows." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 illustrates the light path from a pixel \\( \\pmb{p}_p \\) on the plane-of-refraction. Suppose the emergent ray \\( \\ell_o \\) intersects" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20815" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.47, + 0.149 + ], + "angle": 0, + "content": "Table 1. The Pixel-Ray Mapping in Spherical Refraction Ray-Pixel Camera. A virtual pixel parameterized by \\(\\theta_q\\) is associated with a pixel-wise focal length \\(f_q(\\theta_q) \\in \\mathbb{R}\\). The derivative \\(\\frac{\\mathrm{d}f_q(\\theta_q)}{\\mathrm{d}\\theta_q}\\) is also stored for our numerical 3D-2D projection." + }, + { + "type": "table", + "bbox": [ + 0.112, + 0.153, + 0.437, + 0.208 + ], + "angle": 0, + "content": "
pixelθq0θq1...θqn
rayfq(θq0)fq(θq1)...fq(θqn)
dfq(θq)/dθqf′q(θq0)f′q(θq1)...f′q(θqn)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.47, + 0.42 + ], + "angle": 0, + "content": "with the optical axis at \\( \\mathbf{o}_{f_q} = (0, f_q)^\\top \\) with angle \\( \\theta_q \\). Obviously, changing the position of \\( \\mathbf{p}_p \\) in \\( r \\) direction results in changing the corresponding \\( \\langle f_q, \\theta_q \\rangle \\) pair. That is, since the mapping between the pixel \\( \\mathbf{p}_p \\) i.e., \\( \\ell_i \\), and the ray \\( \\ell_o \\), i.e., \\( \\langle f_q, \\theta_q \\rangle \\) is bijective because of the reversibility of light, representing the light paths from the pixels in \\( r \\) space is identical to knowing all possible \\( \\langle f_q, \\theta_q \\rangle \\) pairs. In other words, the system of Fig. 3 as a whole can be seen as a pixel-wise varifocal camera which changes the focal length \\( f_q \\) for each virtual pixel parameterized by \\( \\theta_q \\). In fact, the mapping \\( \\theta_q \\mapsto f_q \\) is a monotonic function due to the spherical structure of the lens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.533 + ], + "angle": 0, + "content": "Table 1 shows our ray-pixel mapping. Due to the spherical structure of the lens, this is a discretization of the monotonic function \\(\\theta_q \\mapsto f_q\\) by \\(\\theta_q\\). In practice, we sample \\(\\theta_q\\) so that their interval results in a sub-pixel sampling in the original image domain. Notice that the derivative \\(\\frac{\\mathrm{d}f_q(\\theta_q)}{\\mathrm{d}\\theta_q}\\) is also stored for our numerical 3D-2D projection as described below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.554, + 0.254, + 0.57 + ], + "angle": 0, + "content": "3.1.2 3D-2D Projection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.469, + 0.744 + ], + "angle": 0, + "content": "Instead of the analytical forward projection which requires solving a 10th-degree equation [2], this section introduces a numerical forward projection using our ray-pixel mapping table (Tab. 1). The key idea of our numerical forward projection is to hypothesize a projection line \\(\\ell_{\\lambda}\\) in Fig. 3 first, and check if it intersects with the optical axis at the identical virtual focal length stored in the ray-pixel mapping table (Tab. 1). If the focal lengths are not identical, then \\(\\ell_{\\lambda}\\), i.e., the virtual pixel \\(\\theta_{q}\\) equivalently, is refined to minimize the difference. We can formulate this process as a Gauss-Newton optimization as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.883 + ], + "angle": 0, + "content": "Gauss-Newton Method As shown in Fig. 3, let us consider the ray \\(\\ell_{\\lambda}\\) from a 3D point to be projected. By hypothesizing that \\(\\ell_{\\lambda}\\) intersects the optical axis at an angle \\(\\theta_{q}\\), i.e., by hypothesizing that the 3D point is projected to a virtual pixel \\(\\theta_{q}\\), we can compute the intersection \\(\\Lambda_{q}(\\theta_{q}) = (0, \\lambda_{q}(\\theta_{q}))^{\\top}\\) of \\(\\ell_{\\lambda}\\) and the optical axis. If \\(\\lambda_{q}\\) is equal to the virtual focal length \\(f_{q}(\\theta_{q})\\) stored in the ray-pixel mapping, the ray \\(\\ell_{\\lambda}\\) is identical to \\(\\ell_{o}\\), and hence that can intersect with the optical axis at the camera center \\(o_{0}\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "That is, the numerical 3D-2D projection can be achieved" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.682, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.096, + 0.871, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Figure 4. Kaleidoscopic Imaging. (a) Mirrored Cameras. (b) Chambers in the real camera image. The label \\(i\\) of \\(C_i\\) and \\(\\pmb{u}^{(i)}\\) denote the reflections of \\(C\\) and \\(u\\) by the \\(i\\)th mirror respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.756, + 0.308 + ], + "angle": 0, + "content": "by solving the following optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.32, + 0.892, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\theta_ {q} = \\underset {\\theta_ {q}} {\\operatorname {a r g m i n}} G \\left(\\theta_ {q}\\right) = \\underset {\\theta_ {q}} {\\operatorname {a r g m i n}} \\left(f _ {q} \\left(\\theta_ {q}\\right) - \\lambda_ {q} \\left(\\theta_ {q}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.36, + 0.892, + 0.391 + ], + "angle": 0, + "content": "Here \\( G(\\theta_q) \\) is a monotonic function and hence we can refine \\( \\theta_q \\) iteratively by Gauss-Newton method as" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.402, + 0.89, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta_ {q} ^ {(k + 1)} = \\theta_ {q} ^ {(k)} - G \\frac {\\Delta \\theta_ {q}}{\\Delta G} \\\\ = \\theta_ {q} ^ {(k)} - \\frac {f _ {q} \\left(\\theta_ {q} ^ {(k)}\\right) - \\lambda_ {q} \\left(\\theta_ {q} ^ {(k)}\\right)}{f _ {q} ^ {\\prime} \\left(\\theta_ {q} ^ {(k)}\\right) - \\lambda_ {q} ^ {\\prime} \\left(\\theta_ {q} ^ {(k)}\\right)}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.783, + 0.487 + ], + "angle": 0, + "content": "where \\(\\theta_q^{(k)}\\) denotes \\(\\theta_{q}\\) of the \\(k\\)-th iteration." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.891, + 0.532 + ], + "angle": 0, + "content": "Therefore, if we compute the derivative \\( f_{q}^{\\prime} \\) beforehand as shown in the third row of Tab. 1, then this 3D-2D projection can be computed efficiently." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.675, + 0.557 + ], + "angle": 0, + "content": "3.2. Multifacet Mirror" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.687 + ], + "angle": 0, + "content": "As is well known, observing the scene via a multi-facet mirror or a kaleidoscope is identical to observing the scene by virtual multi-view cameras, and in particular, kaleidoscopes with three mirrors are known to be reasonable in terms of less overlaps of mirrored images called discontinuities [29,33]. In our teleidoscopic imaging system, we use a three-facet mirror which defines reflections of the spherical refraction ray-pixel camera introduced in Sec. 3.1.1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.688, + 0.891, + 0.718 + ], + "angle": 0, + "content": "The reflection \\( \\pmb{p}^{\\prime} \\) of a 3D point \\( \\pmb{p} \\) by a mirror of normal \\( \\pmb{n}_i \\) and distance \\( d \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.73, + 0.891, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {p} ^ {\\prime} = H _ {i} \\boldsymbol {p} + \\boldsymbol {t} _ {i} \\\\ \\Leftrightarrow \\boldsymbol {p} ^ {\\prime} = \\left(\\boldsymbol {I} - 2 \\boldsymbol {n} _ {i} \\boldsymbol {n} _ {i} ^ {\\top}\\right) \\boldsymbol {p} + 2 d _ {i} \\boldsymbol {n} _ {i}, \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.78, + 0.829, + 0.794 + ], + "angle": 0, + "content": "where \\(H_{i}\\) the Householder transformation matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.891, + 0.856 + ], + "angle": 0, + "content": "In the case of kaleidoscopic imaging, the mirrors generate bouncing reflections as shown in Fig. 4. The reflection of \\( \\pmb{p}^{\\prime} \\) by another mirror of normal \\( \\pmb{n}_j \\) and distance \\( d \\) is given simply by" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.868, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {p} ^ {\\prime \\prime} = H _ {i j} \\boldsymbol {p} + \\boldsymbol {t} _ {i j} \\\\ = H _ {j} H _ {i} \\boldsymbol {p} + 2 d _ {j} \\boldsymbol {n} _ {j} + 2 H _ {j} d _ {i} \\boldsymbol {n} _ {i}. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.455, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Depth-of-Field with Monocentric Lens. \\( D_{F} \\) and \\( D_{N} \\) denote the depth-of-field of the camera itself without the monocentric lens. The effective depth-of-field of the entire system \\( L_{F} \\) and \\( L_{N} \\) can be obtained by projecting \\( D_{N} \\) and \\( D_{F} \\) through the monocentric lens (the red and blue dashed lines)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.375 + ], + "angle": 0, + "content": "As a result, our teleidoscopic imaging system can be modeled as multi-view spherical refraction ray-pixel cameras. They share the single monocentric lens, and the pose of each virtual cameras can be computed by Eq. (12) if the mirror parameters are given. The later section describes our calibration algorithm to estimate such parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.386, + 0.471, + 0.421 + ], + "angle": 0, + "content": "4. Depth-of-Field of Teleidoscopic Imaging System" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.475 + ], + "angle": 0, + "content": "This section describes analytical evaluations on the depth-of-field of our teleidoscopic imaging system. We first review the depth-of-field of monocentric lens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.409, + 0.499 + ], + "angle": 0, + "content": "4.1. Depth-of-Field with Monocentric Lens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.672 + ], + "angle": 0, + "content": "As shown in Fig. 5, depth-of-field with a monocentric lens can be simply defined as an extension of the path of the thin lens camera (cf. supplemental material). This is because the points in the near and the far depth-of-field \\( D_N \\) and \\( D_F \\) of thin lens are projected within the permissible circle-of-confusion even though the monocentric lens itself introduces the spherical aberration. Hence, we define the depth-of-field with a monocentric lens \\( L_N \\) and \\( L_F \\) as the intersection of the backprojection path through the edge of aperture (the red and blue dash lines in Fig. 5) and the backprojection path through the lens center \\( o_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.825 + ], + "angle": 0, + "content": "Because of the symmetric structure of the monocentric lens, the depth-of-field \\( L_{N} + L_{F} \\) can be deeper than the original depth-of-field \\( D_{N} + D_{F} \\) if the subject distance of the monocentric lens \\( s_{D} - r_{D} \\) is larger than its focal distance \\( t_{D} - r_{D} \\). That is, there is a trade-off between magnification and depth-of-field even in a monocentric lens. The reason for the deep depth-of-field is due to the significant contribution of the monocentric lens to the far depth of field \\( L_{F} \\), and thus imaging with less blurring even when designed for high magnification (cf. Sec. 6.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.863 + ], + "angle": 0, + "content": "4.2. Depth-of-Field of Teleidoscopic Imaging System" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Fig. 6 illustrates the near and the far depth-of-fields of a virtual camera \\( C_i \\) defined as a mirror of the original camera." + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.093, + 0.847, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 6. Depth of Field of Teleidoscopic Imaging System. \\(D_N\\) and \\(D_F\\) denote the near and far depth-of-field of the thin lens camera \\(C\\) respectively. All the points once imaged within \\(D_N - D_F\\) should be originally emitted from the points within \\(L_N - L_F\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Due to the aperture, the near and the far depth-of-fields form a curve respectively (the red and the blue dashed lines)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Since the teleidoscopic imaging system has multiple virtual cameras as shown in Fig. 4, each of them has a different depth-of-field according to their relative pose to the monocentric lens. That is, the intersections of such depth-of-fields can be used for multi-view stereo reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.891, + 0.512 + ], + "angle": 0, + "content": "5. Calibration of Teleidoscopic Imaging System" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.702 + ], + "angle": 0, + "content": "This section introduces our calibration algorithm of teleidoscopic imaging system which requires capturing a single reference planar patterns. Fig. 7 shows the measurement model where the real camera \\(C\\) observes reference points \\(\\pmb{p}_w\\) such as chessboard corners on a reference board II via a monocentric lens and three mirrors. Our calibration estimates the mirror normals \\(\\pmb{n}_i\\), their distances \\(d_i\\) from the camera, and the position and the radius of the monocentric lens, with assuming that the 2D positions of the reference points \\(\\pmb{p}_w\\) on the reference plane II, the intrinsic parameters of the camera, and the refraction index of the lens are given beforehand." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.87 + ], + "angle": 0, + "content": "A challenge in this calibration is the fact that the mirrors require observing 3D points and their reflections to estimate their poses [33], while the observation in the teleidoscopic system does not include such mirrored points that follow Eq. (12) due to the refraction by the monocentric lens. Similarly, the rays reflected and then refracted through the projection in teleidoscopic imaging system do not satisfy the coplanarity constraint. In addition, as mentioned in [2], the estimation of the monocentric lens parameters requires multiple viewpoints or multiple monocentric lenses. These points result in a chicken-and-egg problem." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "To solve this problem, we utilize the fact that the positions of the centers of the monocentric lens and its mirrors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20817" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.095, + 0.434, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.471, + 0.323 + ], + "angle": 0, + "content": "Figure 7. Calibration of Teleidoscopic Imaging System. Cyan lines denote the light paths for a mirrored camera. The calibration can be conducted by capturing reference points on a plane of a known geometry through the monocentric lens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.381 + ], + "angle": 0, + "content": "are always captured without refraction by definition, and satisfy Eq. (12). Therefore, we start by estimating the center of the monocentric lens from the captured reference points." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.382, + 0.47, + 0.464 + ], + "angle": 0, + "content": "In what follows, \\( \\pmb{p}_p^{(0)} \\) denotes the real image of a reference point \\( \\pmb{p}_w \\) in the camera \\( C \\). Similarly \\( \\pmb{p}_p^{(i)} \\) (\\( i = \\{1,2,3\\} \\)) denotes the image of its first reflection by the mirror \\( i \\), and \\( \\pmb{p}_p^{(ij)} \\) (\\( ij = \\{12,13,21,23,31,32\\} \\)) denotes its second reflection by the mirrors \\( i \\) and \\( j \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.471, + 0.471, + 0.553 + ], + "angle": 0, + "content": "Axis of Monocentric Lens The axis to the center of monocentric lens \\( \\pmb{a}_v^{(0)} \\) from the camera \\( C \\) can be estimated by \\( \\pmb{p}_w - \\pmb{p}_P^{(0)} \\) correspondences. Similarly to [2], we estimate \\( \\pmb{a}_v^{(0)} \\) by the coplanarity constraint of the three vectors \\( \\pmb{a}_v^{(0)} \\), \\( \\pmb{p}_p^{(0)} \\), and \\( \\pmb{p}_w \\) on a single plane-of-refraction:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.56, + 0.47, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {p} _ {p} ^ {(0) \\top} \\left(\\boldsymbol {a} _ {v} ^ {(0)} \\times \\left(R _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {t} _ {\\Pi}\\right)\\right) = 0, \\tag {13} \\\\ \\Leftrightarrow \\boldsymbol {p} _ {p} ^ {(0) \\top} \\left(E _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {s} _ {\\Pi}\\right)) = 0, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.469, + 0.711 + ], + "angle": 0, + "content": "where \\(R_{\\Pi}, t_{\\Pi}\\) are the pose of the reference plane in the camera coordinate system and \\(E_{\\Pi} = \\boldsymbol{a}_v^{(0)} \\times R_{\\Pi}\\) and \\(s_{\\Pi} = \\boldsymbol{a}_v^{(0)} \\times t_{\\Pi}\\). Since this is a linear equation with 9 unknown parameters of \\(E_{\\Pi}\\) and \\(s_{\\Pi}\\), we can obtain \\(E_{\\Pi}\\) and \\(s_{\\Pi}\\) up to scale by observing at least 8 points on the reference plane. Using the estimated \\(E_{\\Pi}\\), the axis \\(\\boldsymbol{a}_v^{(0)}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.719, + 0.47, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {v} ^ {(0)} = \\frac {E _ {\\Pi} (: , 1) \\times E _ {\\Pi} (: , 2)}{| | E _ {\\Pi} (: , 1) \\times E _ {\\Pi} (: , 2) | |}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.782 + ], + "angle": 0, + "content": "Similarly, the axis to the center of the mirrored monocentric lens \\( \\pmb{a}_v^{(i)} \\) can be obtained by the mirrored points \\( \\pmb{p}_p^{(i)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.788, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Mirror Normals The axis to the center of the monocentric lens \\( \\pmb{a}_{v}^{(0)} = (x_{a_{v}}^{(0)},y_{a_{v}}^{(0)},z_{a_{v}}^{(0)})^{\\top} \\) and its mirror \\( \\pmb{a}_{v}^{(i)} = (x_{a_{v}}^{(i)},y_{a_{v}}^{(i)},z_{a_{v}}^{(i)})^{\\top} \\) satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.844, + 0.469, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {v} ^ {(0) \\top} [ \\boldsymbol {n} _ {i} ] _ {\\times} \\boldsymbol {a} _ {v} ^ {(i)} = 0, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "where \\([\\pmb{n}_i]_{\\times}\\) denotes the skew-symmetric matrix defined by the normal \\(\\pmb{n}_i = (x_{n_i},y_{n_i},z_{n_i})^\\top\\) of the mirror \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "The same constraint holds for each of the first-second reflection pairs \\( \\pmb{a}_v^{(i)} - \\pmb{a}_v^{(ij)} \\) about the same mirror normal \\( \\pmb{n}_i \\) [33]. Therefore \\( \\pmb{n}_i \\) can be obtained linearly only from the axes to the centers of the monocentric lenses." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.159, + 0.892, + 0.22 + ], + "angle": 0, + "content": "Mirror Distances Once the mirror normals are estimated, we can utilize the kaleidoscopic triangulation [33] to obtain linear constraints on the mirror distances \\( d_{i} \\) (See supp. Sec. 2 for details)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Pose of Reference Plane Similarly to Eq. (13), the plane-of-refraction constraint holds for the mirrored cameras \\( C_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.28, + 0.892, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {p} _ {p} ^ {(i)}\\right) ^ {\\top} \\left(\\boldsymbol {a} _ {v} ^ {(i)} \\times \\left(H _ {i} \\left(R _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {t} _ {\\Pi}\\right) + \\boldsymbol {t} _ {i}\\right)\\right) = 0. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.315, + 0.892, + 0.345 + ], + "angle": 0, + "content": "This constraint allows us estimating the pose of the reference plane \\(R_{\\Pi}\\) and \\(\\pmb{t}_{\\Pi}\\) linearly." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.352, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Monocentric Lens Parameters The calibration algorithm up to this point does not require the monocentric lens parameters \\(d_{D}, r_{D}\\), and \\(\\mu_{g}\\). We estimate these parameters by the coplanarity constraint of the ray through \\(\\pmb{p}_{p}^{(i)}\\) of \\(C_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.427, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {o} ^ {(i)} \\times \\left(\\boldsymbol {p} _ {w} ^ {(i)} - \\boldsymbol {p} _ {o} ^ {(i)}\\right) = 0. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.458, + 0.892, + 0.518 + ], + "angle": 0, + "content": "This is a nonlinear constraint for the monocentric lens parameters as described in Sec. 3.1 and we solve this as a nonlinear optimization problem with assuming their rough estimates are available in practice." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.618 + ], + "angle": 0, + "content": "**Bundle Adjustment** The last step of our calibration is to refine the parameters \\( \\boldsymbol{a}_v^{(0)}, d_D, r_D, \\mu_g, \\boldsymbol{n}_i, d_i, R_{\\Pi}, t_{\\Pi} \\) (\\( i = 1, 2, 3 \\)) by minimizing the reprojection errors of the reference points \\( \\boldsymbol{p}_w \\) as a nonlinear optimization problem. Here the forward projection is computed using the method described in Sec. 3.1.2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.619, + 0.892, + 0.679 + ], + "angle": 0, + "content": "Once the teleidoscopic imaging system parameters are calibrated, a scene point can be linearly triangulated in a DLT manner [13] from its teleidoscopic projections (See supp. Sec. 3 for details)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.692, + 0.616, + 0.708 + ], + "angle": 0, + "content": "6. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.718, + 0.892, + 0.794 + ], + "angle": 0, + "content": "This section provides quantitative evaluations of the proposed numerical 3D-2D projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also demonstrate the effectiveness of our method by reconstructing real-world microscale objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.89, + 0.818 + ], + "angle": 0, + "content": "6.1. Computational Efficiency of 3D-2D Projection" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The average computation cost of a single forward projection by our method and by the analytical solution solving a 10th-degree equation for each projection [2] are \\(0.62\\mathrm{ms}\\) and \\(3.93\\mathrm{ms}\\) respectively. These runtime costs are the average values of 100 trials of 10K points forward projections" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.094, + 0.273, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.199, + 0.17, + 0.211 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.091, + 0.457, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.199, + 0.392, + 0.21 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.47, + 0.27 + ], + "angle": 0, + "content": "Figure 8. Teleidoscopic System for Quantitative Evaluation. (a) Simulated measurement environment. (b) Reprojection errors at different noise levels. The bars denote the standard deviation of the errors." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.287, + 0.201, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.286, + 0.339, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.286, + 0.464, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.397, + 0.201, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.507, + 0.158, + 0.518 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.401, + 0.339, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.507, + 0.295, + 0.519 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.397, + 0.462, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.507, + 0.424, + 0.519 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Figure 9. Calibration Errors. From left to right, the upper figures show the (a) angular error of the axis to the monocentric lens center (degree), (b) angular error of the mirror normal (degree), and (c) rotation error (Riemannian distance) of reference plane. The lower ones are the distance errors normalized by the ground truth. Here the distance of the monocentric lens is only estimated in a nonlinear manner (Sec. 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.68 + ], + "angle": 0, + "content": "run in Matlab on an Intel Core-i7 2.6GHz PC. This result clearly verifies that our numerical forward projection based on our ray-pixel model can realize an efficient projection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.69, + 0.203, + 0.705 + ], + "angle": 0, + "content": "6.2. Calibration" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.714, + 0.465, + 0.729 + ], + "angle": 0, + "content": "6.2.1 Quantitative Evaluation using Synthesized Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Fig. 8(a) shows the measurement environment which simulates the real capture system used in Sec. 6.2.2. The system has a kaleidoscope with three \\(10 \\times 30\\mathrm{mm}\\) mirrors in front of the camera \\(C\\). The mirrors are at slightly off-perpendicular angle of \\(1.4^{\\circ}\\) to the camera image plane so that the mirrors define virtual cameras around the target with less overlaps of the mirrored images. The system also has a monocentric lens of \\(10\\mathrm{mm}\\) diameter in front of the mirrors, at \\(40\\mathrm{mm}\\) distance from the camera. The refraction index \\(\\mu_{g}\\) is set to 2.0. The system captures 48 reference points (blue dots in Fig. 8(a)) to calibrate its parameters. By injecting Gaussian" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.701, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.191, + 0.612, + 0.202 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.091, + 0.878, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.191, + 0.803, + 0.201 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.207, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 10. Experimental Setup and Calibration. (a) Our teleidoscopic imaging system consists of a camera, a projector, three mirrors \\((10 \\times 30\\mathrm{mm})\\), and a monocentric lens (5mm radius, refraction index \\(\\mu_g = 2.0\\)). (b) Input image for calibration." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.279, + 0.633, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.363, + 0.587, + 0.374 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.279, + 0.764, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.362, + 0.719, + 0.373 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.279, + 0.893, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.362, + 0.845, + 0.373 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Figure 11. Changes in Depth-of-Field by the Monocentric Lens. (a), (b) and (c) show the total, near, and far depth-of-field with a monocentric lens. The red and blue plots indicate the results with and without the monocentric lens respectively. The green lines indicate the subject distance where the subject and the focal distances of the monocentric lens itself are same." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.566 + ], + "angle": 0, + "content": "noise of different standard deviations \\(\\sigma\\) to the 2D positions of their projections, we evaluate the robustness of our calibration procedure. Fig. 8(b) shows average reprojection errors in pixel of 100 trials at each pixel noise level \\(\\sigma\\). We can observe that the reprojection errors increase linearly against the pixel noise level." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Fig. 9 shows the estimation errors of the monocentric lens parameters, the mirror parameters, and the reference plane parameters respectively. These results indicate that our calibration algorithm performs reasonably under realistic observation noise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.836, + 0.667 + ], + "angle": 0, + "content": "6.2.2 Quantitative Evaluation using Real Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Fig. 10(a) shows an overview of our teleidoscopic imaging system. The system consists of a FLIR Blackfly BFS-U3-89S6C-C camera (4096×2160 resolution, pixel size 3.45μm) with a f=4mm / F8 lens, three 10 × 30mm mirrors, a monocentric lens of 10mm diameter whose refraction index is \\(\\mu_g = 2.0\\), and a LED projector ASUS E1-R for casting Gray code patterns onto the target surface to identify corresponding points [12, 23]. Note that the effective resolution of each \"virtual camera\" image is approximately 300×200 (16 \"virtual cameras\" in total), which is calculated from the area inside the base chamber with one horizontal edge of the triangle and inside the spherical lens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We used a single image (Fig. 10(b)) capturing \\(5 \\times 8\\) corners of a checkerboard for calibration, and used up to second reflections in the image. The average reprojection er" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.107, + 0.186, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.15, + 0.147, + 0.162 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.108, + 0.285, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.151, + 0.25, + 0.162 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.091, + 0.338, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.092, + 0.419, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.092, + 0.462, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.151, + 0.385, + 0.162 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.167, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Figure 12. Evaluation with a sphere (radius: \\(3\\mathrm{mm}\\)). (a) Ground truth. (b) Recovered 3D shape. The colors visualize the \\(z\\)-coordinate of the reconstructed shape. (c) Error map. \\(E_{d}\\) is the mean absolute error in millimeter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.468, + 0.342 + ], + "angle": 0, + "content": "ror is 0.62px, and the average distance error of the triangulated 3D points to their ground truth positions was \\(0.17\\mathrm{mm}\\). The size of reference checkerboard is around \\(15\\mathrm{mm}\\) width, hence the relative error of the triangulation is around \\(1\\%\\). These results show our method provide a practical calibration and the valid scale of the target object also depends on size of the checkerboard." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.227, + 0.371 + ], + "angle": 0, + "content": "6.3. Depth-of-Field" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Fig. 11 shows the changes in the depth-of-field by the monocentric lens capturing an object of \\(0.5\\mathrm{mm}\\) height at the same magnification given the configuration same as the real capture in Sec. 6.2.2. Fig. 11(a) shows the total depth-of-field with and without the monocentric lens \\(L_{N} + L_{F}\\) (red and blue respectively). Fig. 11(b) and (c) show \\(L_{N}\\) and \\(L_{F}\\) in the same manner. In these figures, the horizontal axis indicates the apparent size of the target in pixel, with assuming the measurement model used in Sec. 6.2. These plots clearly indicate that the monocentric lens contributes to increase the depth-of-field successfully." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.415, + 0.574 + ], + "angle": 0, + "content": "6.4. Teleidoscopic 3D Shape Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.642 + ], + "angle": 0, + "content": "To validate the proposed teleidoscopic system as a multiview camera system for 3D shape reconstruction, this section demonstrates 3D reconstruction of small objects of approximately \\(6\\mathrm{mm}\\) size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.469, + 0.718 + ], + "angle": 0, + "content": "We used the system calibrated in Sec. 6.2.2. Fig. 12 visualizes our 3D reconstruction of the front side of a real sphere of \\(3\\mathrm{mm}\\) radius with the ground-truth and the error map. The mean absolute error \\(E_{d}\\) of the estimated depth is \\(0.097\\mathrm{mm}\\). Compared with its size, the relative error is about \\(1.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fig. 13 shows additional 3D reconstructions. The 1st and the 2nd rows demonstrate that our imaging system has the ability to recover complex shapes. The proposed imaging system well recovers the complex bumpy parts of the shell and the sharp edges and corners of the object. The 2nd and the 3rd rows compare \"with\" and \"w/o\" the monocentric lens cases. As the input image (b) shows, the monocentric lens creates a surrounding view while preserving the area of each chamber, resulting in a fuller and more accurate geometry, as shown in (c). We can verify this quantitatively by comparing the numbers of pixels covering the object area in the captured image (145,217 vs. 12,730) and the numbers" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.127, + 0.574, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.075, + 0.685, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.148, + 0.685, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.094, + 0.767, + 0.129 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.093, + 0.852, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.161, + 0.775, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.161, + 0.861, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.206, + 0.808, + 0.217 + ], + "angle": 0, + "content": "40,639 pts recovered" + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.223, + 0.785, + 0.233 + ], + "angle": 0, + "content": "w/o monocentric lens" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.234, + 0.685, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.306, + 0.643, + 0.318 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.255, + 0.778, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.293, + 0.803, + 0.305 + ], + "angle": 0, + "content": "8,264 pts recovered" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.254, + 0.86, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.306, + 0.795, + 0.318 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.326, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Figure 13. Results of 3D Shape Estimation. The second and the third row compares \"with\" and \"w/o\" the monocentric lens. (a) Room-light appearance. (b) Teleidoscopic image. (c) Reconstructed 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.891, + 0.434 + ], + "angle": 0, + "content": "of triangulated 3D points using the ground-truth correspondences as shown in the figure (40,639 vs. 8,264). See the supplementary material for more results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Limitations The major limitations of our method are twofold. The monocentric lens has a relatively large spherical aberration, and the optical resolution of the entire system is limited by such aberrations. One possible solution is to introduce nested concentric glass spheres with different refractive indices for the monocentric lens [6, 35]. The other limitation is the trade-off between the number of virtual viewpoints and the effective resolution of each view because the use of a kaleidoscope requires zooming-out to capture the reflected images. While larger number of the cameras contributes to robustify the triangulation, it inevitably reduces the resolution of the reconstruction as the number of pixels per view decreases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.641, + 0.619, + 0.656 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.844 + ], + "angle": 0, + "content": "This paper proposed the concept of image-based 3D measurement by a teleidoscopic imaging system for microscale objects. By combining a monocentric lens with a kaleidoscope, our system realized a closeup and surrounding multi-view capture of the target. Our contributions are twofold. We proposed a ray-pixel camera model that realized an efficient numerical 3D-2D projection computation, and also proposed a calibration method for the teleidoscopic imaging system. Our future work includes aberration and diffraction blur handling such as the inverse filter proposed in [5], and also refraction/attenuation modeling in the case of capturing objects in water [18, 19]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.857, + 0.667, + 0.872 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "This work is in part supported by JSPS KAKENHI JP22K17914." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20820" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Rajat Aggarwal, Amrisha Vohra, and Anoop M. Namboodiri. Panoramic stereo videos with a single camera. In CVPR, pages 3755-3763, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.198 + ], + "angle": 0, + "content": "[2] Amit Agrawal and Srikumar Ramalingam. Single image calibration of multi-axial imaging systems. In CVPR, pages 1399–1406, 2013. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[3] Byeongjoo Ahn, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Kaleidoscopic structured light. ACM TOG, 40(6):1-15, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.296 + ], + "angle": 0, + "content": "[4] Michael Chen, Zachary F. Phillips, and Laura Waller. Quantitative differential phase contrast (dpc) microscopy with computational aberration correction. Optics Express, (25):32888-32899. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.325 + ], + "angle": 0, + "content": "[5] O. S. Cossairt, D. Miau, and S. K. Nayar. Gigapixel computational imaging. In ICCP, pages 1-8, 2011. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[6] Donald G. Dansereau, Glenn Schuster, Joseph Ford, and Gordon Wetzstein. A wide-field-of-view monocentric light field camera. In CVPR, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.396 + ], + "angle": 0, + "content": "[7] Michael De Zeeuw and Aswin C Sankaranarayanan. Widebaseline light fields using ellipsoidal mirrors. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.468, + 0.437 + ], + "angle": 0, + "content": "[8] Keith Forbes, Fred Nicolls, Gerhard De Jager, and Anthon Voigt. Shape-from-silhouette with two mirrors and an uncalibrated camera. In ECCV, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[9] Brian J. Ford. Single lens: The story of the simple microscope. Journal of the History of Biology, 19(2):320-321, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.481, + 0.468, + 0.536 + ], + "angle": 0, + "content": "[10] Joshua Gluckman, Shree K. Nayar, and Keith J. Thoresz. Real-time omnidirectional and panoramic stereo. In Proceedings of the 1998 DARPA Image Understanding Workshop, pages 299-303. Morgan Kaufmann, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[11] Michael D. Grossberg and Shree K. Nayar. The raxel imaging model and ray-based calibration. *IJCV*, 61(2):119–137, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[12] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G. Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR, pages 713-720, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.468, + 0.663 + ], + "angle": 0, + "content": "[13] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[14] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE TPAMI, 41(1):190-204, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[15] Srinivasa Karthik and Ambasamudram N. Rajagopalan. Underwater microscopic shape from focus. In ICPR, pages 2107-2112, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[16] Ryo Kawahara, Shohei Nobuhara, and Takashi Matsuyama. A pixel-wise varifocal camera model for efficient forward projection and linear extrinsic calibration of underwater cameras with flat housings. In Proc. of ICCV 2013 Underwater Vision Workshop, pages 819-824, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[17] Gurunandan Krishnan and Shree K. Nayar. Towards a true spherical camera. In SPIE, volume 7240, 724002, pages 1-13, 2009. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Meng-Yu Jennifer Kuo, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Non-rigid shape from water. IEEE TPAMI, 43(7):2220-2232, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Meng-Yu Jennifer Kuo, Satoshi Murai, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Surface normals and shape from water. IEEE TPAMI, 44(12):9150-9162, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[20] Douglas Lanman, Daniel Crispell, and Gabriel Taubin. Surround structured lightning: 3-d scanning with orthographic illumination. In CVIU, pages 1107-1117, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[21] A. Manakov, J. Restrepo, O. Klehm, R. Hegedus, E. Eisemann, H. P. Seidel, and I. Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM TOG, 32(4):47-1, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[22] P. Miraldo and H. Araujo. Calibration of smooth camera models. IEEE TPAMI, 35(9):2091-2103, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.326, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[23] Daniel Moreno and Gabriel Taubin. Simple, accurate, and robust projector-camera calibration. In 3DPVT, pages 464-471, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[24] Andrew D. Mullen, Tali Treibitz, Paul LD Roberts, Emily L. A. Kelly, Rael Horwitz, Jennifer E. Smith, and Jules S. Jaffe. Underwater microscopy for in situ studies of benthic ecosystems. Nature Communications, 7, 2016/07 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[25] Shree K. Nayar. Omnidirectional video camera. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 235-241, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[26] Markus Niederost, Jana Niederost, and Jiri Scucka. Automatic 3d reconstruction and visualization of microscopic objects from a monoscopic multifocus image sequence. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XXXIV-5/W10:5-10, 01 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[27] Ryoichiro Nishi, Takahito Aoto, Norihiko Kawai, Tomokazu Sato, Yasuhiro Mukaigawa, and Naokazu Yokoya. *Ultrashallow dof imaging using faced paraboloidal mirrors*. In ACCV, pages 115–128, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[28] Ilya Reshetouski and Ivo Ihrke. Mirrors in Computer Graphics, Computer Vision and Time-of-Flight Imaging, pages 77–104. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[29] Ilya Reshetouski, Alkhazur Manakov, Hans-Peter Seidel, and Ivo Ihrke. Three-dimensional kaleidoscopic imaging. In CVPR, pages 353-360, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[30] J. Starck, A. Maki, S. Nobuhara, A. Hilton, and T. Matsuyama. The multiple-camera 3-d production studio. IEEE TCSVT, 19(6):856-869, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[31] Peter Sturm. Multi-view geometry for general camera models. In CVPR, volume 1, pages 206-212, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[32] Seiichi Tagawa, Yasuhiro Mukaigawa, Jaewon Kim, Ramesh Raskar, Yasuyuki Matsushita, and Yasushi Yagi. Hemispherical confocal imaging. IPSJ Transactions on Computer Vision and Applications, 3:222-235, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[33] Kosuke Takahashi, Akihiro Miyata, Shohei Nobuhara, and Takashi Matsuyama. A linear extrinsic calibration of kaleidoscopic imaging system from single 3d point. In CVPR, 2017. 1, 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20821" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[34] Lei Tian, Jingyan Wang, and Laura Waller. 3d differential phase-contrast microscopy with computational illumination using an led array. Optics letters, 39:1326-9, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.188 + ], + "angle": 0, + "content": "[35] Eric J Tremblay, Daniel L Marks, David J Brady, and Joseph E Ford. Design and scaling of monocentric multiscale imagers. Applied Optics, 51(20):4691-4702, 2012. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.246 + ], + "angle": 0, + "content": "[36] Borislav Trifonov, Derek Bradley, and Wolfgang Heidrich. Tomographic reconstruction of transparent objects. In Proc. of Eurographics Conference on Rendering Techniques, pages 51-60, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[37] David Uhlig and Michael Heizmann. A calibration method for the generalized imaging model with uncertain calibration target coordinates. In ACCV. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[38] Frank Verbiest, Marc Proesmans, and Luc Van Gool. Modeling the effects of windshield refraction for camera calibration. In ECCV, pages 397-412, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20822" + } + ] +] \ No newline at end of file diff --git a/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_origin.pdf b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..788fe6b6d8cbeb8d2f317c0d86cd2b289d61c0e8 --- /dev/null +++ b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/d72f0f54-2daa-4009-a052-255877447821_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb13b109e92e4bad028dc167e6d3882d31209843c992ecfde3483a878035e12e +size 1593431 diff --git a/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/full.md b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c06137327ce01a7b6dc2411bf00b66fc3add5fed --- /dev/null +++ b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/full.md @@ -0,0 +1,454 @@ +# Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction + +![](images/22ce7a0da53799a283a040951791c760f89a62c4d805631443c2a3890330a18a.jpg) +Figure 1. Teleidoscopic Imaging System. (a) We propose a novel system for microscale 3D imaging using an imaging setup that consists of a camera, a kaleidoscopic multi-facet mirror with a monocentric front lens, and a projector. (b) Our system realizes a dense close-up surrounding multi-view capture of the target. (c) Our method enables fuller and dense reconstruction of microscale objects. + +![](images/d93f9aa0dd94cd442af7962111403cd3b0ebbedd03dd492586eca05959dcd061.jpg) +Ryo Kawahara +Kyushu Institute of Technology +rkawahara@ai.kyutech.ac.jp + +![](images/0cfedaef57b0e38b012017ff3114aa0cfb48426e45e7fcea13244e7c9d0bcf83.jpg) +Meng-Yu Jennifer Kuo University of Minnesota kuo00052@umn.edu + +![](images/668d72d9a3758679c630e887d42512a46a2f017998456eb55bf41a8152504fd7.jpg) +Shohei Nobuhara Kyoto University nob@i.kyoto-u.ac.jp + +# Abstract + +This paper proposes a practical method of microscale 3D shape capturing by a teleidoscopic imaging system. The main challenge in microscale 3D shape reconstruction is to capture the target from multiple viewpoints with a large enough depth-of-field. Our idea is to employ a teleidoscopic measurement system consisting of three planar mirrors and monocentric lens. The planar mirrors virtually define multiple viewpoints by multiple reflections, and the monocentric lens realizes a high magnification with less blurry and surround view even in closeup imaging. Our contributions include, a structured ray-pixel camera model which handles refractive and reflective projection rays efficiently, analytical evaluations of depth of field of our teleidoscopic imaging system, and a practical calibration algorithm of the teleidoscopic imaging system. Evaluations with real images prove the concept of our measurement system. + +# 1. Introduction + +Microscale 3D reconstruction has found profound applications in a wide range of domains including medical imaging, life science, and aquaculture, due to its non-constrained and non-invasive measurements. The main challenges in image-based microscopic 3D shape measurement is its shallow depth of field and camera arrangement in the closeup scenario. Applying conventional multiple camera system designed for human-size capture [14, 30] cannot be a feasible solution due to limitations on camera placement. Con- + +ventional multiple mirror system [33] also have difficulties inevitably in depth-of-focus due to differences in their optical path lengths with varying numbers of bounces. + +In this paper, we show that the fuller 3D shape of a microscale object can be recovered. Our key idea is to employ a catadioptric imaging system which realizes a practical closeup multi-view imaging. The point of our design is that the system has a monocentric front lens like a teleidoscope, instead of using microscopic system in the camera side. That is, as shown in Fig. 1(a), we introduce a kaleidoscopic multi-facet mirror between the front lens and the camera, where the design realizes a deeper depth-of-field and results in less blurring imaging. Unlike conventional microscopic imaging system such as differential phase contrast (DPC) microscopy [4, 34] and multi-focus approaches [15, 24, 26], our method realizes a multi-view capture of the target from a single physical viewpoint which can contribute to free-viewpoint rendering, 3D shape reconstruction, and reflection analysis. Our system also has an advantage over existing imaging systems that build multiple views behind the main lens [7, 21] in a closeup environment. The wide FoV with a monocentric main lens and virtual multiple views with mirrors allow closeup and surrounding view capturing in focus. + +We call our system teleidoscopic imaging system and show that the system can be compactly modeled by a structured ray-pixel camera model [11], which handles refractive and reflective projection rays efficiently. Based on our ray-pixel camera model, we derive a practical calibration algo + +rithm that estimates the positions of the monocentric lens and multi-facet mirror w.r.t. the camera by using a single reference planar pattern (i.e., a checkerboard). Given the calibration parameters, a scene point can then be linearly triangulated from its teleidoscopic projection in a direct linear transform (DLT) manner [13]. + +We implement our method with an imaging system consisting of a camera, three planar mirrors, a monocentric lens, and a projector placed outside the mirrors for structured-light casting. We quantitatively evaluate the computation cost of numerical projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also validate the effectiveness of our method qualitatively on a number of real-world microscale objects. These results demonstrate that our method enables holistic and dense reconstruction of microscale objects. We believe our method expands the avenues of three-dimensional analysis of microscale objects and scenes in real world scenarios. + +# 2. Related Work + +Our teleidoscopic imaging system is a catadioptric system consisting of a monocentric lens and kaleidoscopic mirrors (Fig. 1(a)). While catadioptric system has a wide variety of applications such as omnidirectional observation and panoramic stereo [1, 10, 25, 27], this section reviews studies with multiple flat mirrors and monocentric lenses. + +# 2.1. Imaging with Multiple Flat Mirrors + +A fundamental motivation of introducing mirrors in observation system is to increase the number of viewpoints without installing additional cameras for multi-view capture of a target [3, 8, 20, 28, 32, 33]. Takahashi et al. [33] have proposed a kaleidoscopic imaging system and demonstrated a 3D shape reconstruction using multiple reflections. Tagawa et al. [32] have proposed a multi-facet imaging system that observes a target from an equally distributed virtual cameras for reflectance analysis. + +# 2.2. Imaging with Monocentric Lens + +In the context of imaging system, the monocentric lens is often used to obtain a wide field-of-view [17, 35], in particular for endoscopes, as a short focal length lens. The monocentric lens, however, has additional useful characteristics: its symmetric structure and magnifying power. Cossair et al. [5] have proposed a camera array which captures a same scene through a single monocentric lens so that the images can be stitched into a single high-resolution image. Similarly, Dansereau et al. [6] have proposed a lightfield camera which captures omnidirectional lightfield images through a single monocentric lens with a camera orbiting around it. + +A typical use of the monocentric lens as a magnifier can be found in the Leeuwenhoek's microscope in the 17th cen + +![](images/4917e17c664ac641403725b7043a4d08968a5a6af8531e5094fb63e1dfa86bb4.jpg) +Figure 2. Measurement through Monocentric Lens. The segments of projection path $\ell_{o} - \ell_{g} - \ell_{i}$ have an axially symmetric structure around the axis which directs to the monocentric lens center $o_{D}$ . + +tury. It utilizes a single monocentric lens and realized over $100 \times$ magnifications [9]. + +# 2.3. Ray-pixel Camera Model + +In order to handle light paths captured via complex imaging system such as reflection and refraction, Grossberg and Nayar [11] have proposed a ray-pixel (or raxel) camera model which describes the projection by end-to-end mappings between the pixels and their corresponding emergent rays. While this pixel-wise representation has a great flexibility to describe complicated projections [22, 31], it is also possible to exploit a structure of the rays to simplify the model [16, 22, 36-38]. Kawahara et al. [16] have proposed a ray-pixel camera model for cameras observing the scene through flat surfaces. Compared with analytical forward projections through flat surfaces which requires solving a high order equation [2], it realized a practical numerical forward projection computation by exploiting the axially symmetric structure of the rays. + +# 3. Ray-Pixel Camera Model for Teleidoscopic Imaging System + +In our teleidoscopic system, the cameras mirrored by the kaleidoscopic multi-facet mirrors observe the scene through the single monocentric lens (Fig. 1(a)) as done in [5,6], and each of the monocentric lens and the camera pairs form a microscope capturing the object from a different direction. + +In this section, we show that this design realizes a microscale multi-view capturing with a deep depth-of-field, and that the projections by the teleidoscopic imaging system can be modeled compactly by ray-pixel cameras. + +# 3.1. Measurement through Monocentric Lens + +As illustrated in Fig. 2, suppose a perspective camera whose camera center is located at $\pmb{o}_0$ observes the scene through a monocentric lens located at $\pmb{o}_D$ . Here the perspective camera can be assumed to be directed to the center + +of the monocentric lens $\mathbf{o}_D$ without loss of generality, since we can calibrate the rotation to align the optical axis of the camera to the line $\mathbf{o}_D - \mathbf{o}_0$ as described later. We denote this normalized camera by $C_n$ . + +Obviously, the rays backprojected through pixels of $C_n$ have an axially symmetric structure around the optical axis, and an incident ray $\ell_i$ through a pixel $p_p$ and its refraction $\ell_g$ and emergent ray $\ell_o$ are on a single plane-of-refraction [2]. Therefore, we can describe the ray through a single pixel by a 2D $(r,z)^{\top}$ coordinate system centered at $o_0$ . + +Hereafter, let $r_{\alpha}$ and $z_{\alpha}$ denote the $r$ and $z$ elements of the vector $\alpha$ in general. For example, point $\pmb{p}_p$ is described as $\pmb{p}_p = (r_{\pmb{p}_p}, z_{\pmb{p}_p})^\top$ . Also let $\pmb{v}_X = (r_{\pmb{v}_X}, z_{\pmb{v}_X})^\top$ denote the direction vector of line $\ell_X$ towards the scene from the camera, where $X = \{\mathrm{i}, \mathrm{g}, \mathrm{o}\}$ represents the label of the light path segment. + +Consider the projection path $\ell_o - \ell_g - \ell_i$ from $o_0$ through the point $\pmb{p}_p$ . The incident point $\pmb{p}_i$ on the sphere is described as a function of $\theta_p = \tan^{-1}(r_{\pmb{p}_p} / f_c)$ as + +$$ +z _ {p _ {i}} = \frac {d _ {D} - \sqrt {d _ {D} ^ {2} - (1 + \tan^ {2} \theta_ {p}) \left(d _ {D} ^ {2} - r _ {D} ^ {2}\right)}}{1 + \tan^ {2} \theta_ {p}}, \tag {1} +$$ + +$$ +r _ {p _ {i}} = z _ {p _ {i}} \tan \theta_ {p}, +$$ + +where $d_{D}$ and $r_D$ are the distances to the monocentric lens and the lens radius. + +The refraction angle $\theta_{g}$ is then given by Snell's law with assuming the refractive indices of the air $\mu_{a}$ and the lens $\mu_{g}$ are known: + +$$ +\begin{array}{l} \mu_ {g} \sin \theta_ {g} = \mu_ {a} \sin \left(\theta_ {p} + \theta_ {D i}\right), \\ \Leftrightarrow \sin \theta_ {g} = \frac {\mu_ {a}}{\mu_ {g}} \frac {d _ {D} r _ {p _ {i}}}{r _ {D} \sqrt {r _ {p _ {i}} ^ {2} + z _ {p _ {i}} ^ {2}}}, \tag {2} \\ \end{array} +$$ + +where $\theta_{D_i}$ is the angle of $o_0 - o_D - p_i$ . + +The point of emergence $\pmb{p}_o$ is the intersection of the path $\ell_g$ and the sphere. It is also obtained as the mirror of the point $\pmb{p}_i$ because of the symmetrical relationship between incident path $\ell_i$ and outgoing $\ell_o$ w.r.t. the monocentric lens. As shown in Fig. 2, $\ell_i$ and $\ell_o$ are in line symmetry to the line that is perpendicular to the vector $\pmb{n}_p$ through the point $\pmb{o}_D$ . That is, given the point $\pmb{p}_i$ , its reflection $\pmb{p}_o$ is described as + +$$ +\begin{array}{l} \boldsymbol {p} _ {o} = H _ {p} \boldsymbol {p} _ {i} + \boldsymbol {t} _ {p}, \tag {3} \\ \Leftrightarrow \boldsymbol {p} _ {o} = \left(\boldsymbol {I} - 2 \boldsymbol {n} _ {p} \boldsymbol {n} _ {p} ^ {\top}\right) \boldsymbol {p} _ {i} + 2 (\boldsymbol {o} _ {D} ^ {\top} \boldsymbol {n} _ {p}) \boldsymbol {n} _ {p}, \\ \end{array} +$$ + +where $H_{p}$ is the Householder matrix and $\pmb{t}_{p}$ denotes the center of reflection. Besides, the direction $\pmb{n}_p$ is given as + +$$ +\boldsymbol {n} _ {p} = \left[ \begin{array}{l} \cos \left(\pi / 2 + \theta_ {D i} - \theta_ {g}\right) \\ \sin \left(\pi / 2 + \theta_ {D i} - \theta_ {g}\right) \end{array} \right] = \left[ \begin{array}{c} - \sin \left(\theta_ {D i} - \theta_ {g}\right) \\ \cos \left(\theta_ {D i} - \theta_ {g}\right) \end{array} \right], \tag {4} +$$ + +![](images/5db582d9d8e83d9dba075e94906f27ea39cb5edc2c385ae58ed0a8493cbdf543.jpg) +Figure 3. 3D-2D Projection. The black line $\ell_{o}$ is the correct projection path that intersects with the optical axis at the camera center $o_0$ . Our 3D-2D projection hypothesizes a line $\ell_{\lambda}$ as an initial guess of projection, and then optimizes it by verifying if the point and the angle of intersection with the optical axis is consistent with the ray-pixel mapping $\langle \theta_q, f_q \rangle$ . + +therefore $H_{p}$ and $\pmb{t}_{p}$ are rewritten to + +$$ +\begin{array}{l} H _ {p} = \left[ \begin{array}{l l} \cos 2 \left(\theta_ {D i} - \theta_ {g}\right) & \sin 2 \left(\theta_ {D i} - \theta_ {g}\right) \\ \sin 2 \left(\theta_ {D i} - \theta_ {g}\right) & - \cos 2 \left(\theta_ {D i} - \theta_ {g}\right) \end{array} \right], \tag {5} \\ \boldsymbol {t} _ {p} = \left[ \begin{array}{c} - d _ {D} \sin 2 (\theta_ {D i} - \theta_ {g}) \\ d _ {D} \cos 2 (\theta_ {D i} - \theta_ {g}) + d _ {D} \end{array} \right]. \\ \end{array} +$$ + +The direction $\pmb{v}_o$ is also given by Householder matrix as + +$$ +\boldsymbol {v} _ {o} = H _ {p} \boldsymbol {v} _ {i}. \tag {6} +$$ + +As a result, the intersection $\pmb{o}_{f_q} = (0,f_q)^\top$ of the ray $\ell_o$ and the optical axis is given as follows: + +$$ +\boldsymbol {o} _ {f _ {q}} = t _ {o} \boldsymbol {v} _ {o} + \boldsymbol {p} _ {o}, \tag {7} +$$ + +$$ +\Leftrightarrow t _ {o} = - \frac {r _ {\boldsymbol {p} _ {o}}}{r _ {\boldsymbol {v} _ {o}}}, \quad f _ {q} = z _ {\boldsymbol {p} _ {o}} + \left(\frac {- r _ {\boldsymbol {v} _ {o}}}{z _ {\boldsymbol {v} _ {o}}}\right) ^ {- 1} r _ {\boldsymbol {p} _ {o}}. \tag {8} +$$ + +This equation describes the ray $\ell_{o}$ backprojecting from the pixel $p_p$ as a function of $\theta_{p}$ . Inversely, the forward projection can be obtained for a given 3D point in front of the monocentric lens. The analytical 3D-2D projection, however, requires solving a 10th-degree equation [2]. The next section introduces a ray-pixel camera which exploits the axial symmetric structure of the rays to provide an efficient numerical 3D-2D projection. + +# 3.1.1 Spherical Refraction Ray-Pixel Camera Model + +Up to this point, we observed that the rays emitted from a perspective camera through a monocentric lens show an axially symmetric structure around the line from the camera to the lens center. As illustrated in Fig. 2, given a pixel $p_p$ by specifying $\theta_p$ , the corresponding ray $\ell_o$ can be determined uniquely by Eq. (8). This section introduces our spherical refraction ray-pixel camera model by representing the ray-pixel mapping as follows. + +Fig. 3 illustrates the light path from a pixel $\pmb{p}_p$ on the plane-of-refraction. Suppose the emergent ray $\ell_o$ intersects + +Table 1. The Pixel-Ray Mapping in Spherical Refraction Ray-Pixel Camera. A virtual pixel parameterized by $\theta_q$ is associated with a pixel-wise focal length $f_q(\theta_q) \in \mathbb{R}$ . The derivative $\frac{\mathrm{d}f_q(\theta_q)}{\mathrm{d}\theta_q}$ is also stored for our numerical 3D-2D projection. + +
pixelθq0θq1...θqn
rayfq(θq0)fq(θq1)...fq(θqn)
dfq(θq)/dθqf′q(θq0)f′q(θq1)...f′q(θqn)
+ +with the optical axis at $\mathbf{o}_{f_q} = (0, f_q)^\top$ with angle $\theta_q$ . Obviously, changing the position of $\mathbf{p}_p$ in $r$ direction results in changing the corresponding $\langle f_q, \theta_q \rangle$ pair. That is, since the mapping between the pixel $\mathbf{p}_p$ i.e., $\ell_i$ , and the ray $\ell_o$ , i.e., $\langle f_q, \theta_q \rangle$ is bijective because of the reversibility of light, representing the light paths from the pixels in $r$ space is identical to knowing all possible $\langle f_q, \theta_q \rangle$ pairs. In other words, the system of Fig. 3 as a whole can be seen as a pixel-wise varifocal camera which changes the focal length $f_q$ for each virtual pixel parameterized by $\theta_q$ . In fact, the mapping $\theta_q \mapsto f_q$ is a monotonic function due to the spherical structure of the lens. + +Table 1 shows our ray-pixel mapping. Due to the spherical structure of the lens, this is a discretization of the monotonic function $\theta_q \mapsto f_q$ by $\theta_q$ . In practice, we sample $\theta_q$ so that their interval results in a sub-pixel sampling in the original image domain. Notice that the derivative $\frac{\mathrm{d}f_q(\theta_q)}{\mathrm{d}\theta_q}$ is also stored for our numerical 3D-2D projection as described below. + +# 3.1.2 3D-2D Projection + +Instead of the analytical forward projection which requires solving a 10th-degree equation [2], this section introduces a numerical forward projection using our ray-pixel mapping table (Tab. 1). The key idea of our numerical forward projection is to hypothesize a projection line $\ell_{\lambda}$ in Fig. 3 first, and check if it intersects with the optical axis at the identical virtual focal length stored in the ray-pixel mapping table (Tab. 1). If the focal lengths are not identical, then $\ell_{\lambda}$ , i.e., the virtual pixel $\theta_{q}$ equivalently, is refined to minimize the difference. We can formulate this process as a Gauss-Newton optimization as follows. + +Gauss-Newton Method As shown in Fig. 3, let us consider the ray $\ell_{\lambda}$ from a 3D point to be projected. By hypothesizing that $\ell_{\lambda}$ intersects the optical axis at an angle $\theta_{q}$ , i.e., by hypothesizing that the 3D point is projected to a virtual pixel $\theta_{q}$ , we can compute the intersection $\Lambda_{q}(\theta_{q}) = (0, \lambda_{q}(\theta_{q}))^{\top}$ of $\ell_{\lambda}$ and the optical axis. If $\lambda_{q}$ is equal to the virtual focal length $f_{q}(\theta_{q})$ stored in the ray-pixel mapping, the ray $\ell_{\lambda}$ is identical to $\ell_{o}$ , and hence that can intersect with the optical axis at the camera center $o_{0}$ . + +That is, the numerical 3D-2D projection can be achieved + +![](images/a36862c957beaaf26ce9a13cc4815d931ac3952aa8a80fae46246aedf5bd2b61.jpg) +Figure 4. Kaleidoscopic Imaging. (a) Mirrored Cameras. (b) Chambers in the real camera image. The label $i$ of $C_i$ and $\pmb{u}^{(i)}$ denote the reflections of $C$ and $u$ by the $i$ th mirror respectively. + +![](images/47819e3d5136a0a930b81dc57518d1c877fd451ce490bdfe93f36a0ff7d640c5.jpg) + +by solving the following optimization: + +$$ +\theta_ {q} = \underset {\theta_ {q}} {\operatorname {a r g m i n}} G \left(\theta_ {q}\right) = \underset {\theta_ {q}} {\operatorname {a r g m i n}} \left(f _ {q} \left(\theta_ {q}\right) - \lambda_ {q} \left(\theta_ {q}\right)\right). \tag {9} +$$ + +Here $G(\theta_q)$ is a monotonic function and hence we can refine $\theta_q$ iteratively by Gauss-Newton method as + +$$ +\begin{array}{l} \theta_ {q} ^ {(k + 1)} = \theta_ {q} ^ {(k)} - G \frac {\Delta \theta_ {q}}{\Delta G} \\ = \theta_ {q} ^ {(k)} - \frac {f _ {q} \left(\theta_ {q} ^ {(k)}\right) - \lambda_ {q} \left(\theta_ {q} ^ {(k)}\right)}{f _ {q} ^ {\prime} \left(\theta_ {q} ^ {(k)}\right) - \lambda_ {q} ^ {\prime} \left(\theta_ {q} ^ {(k)}\right)}, \tag {10} \\ \end{array} +$$ + +where $\theta_q^{(k)}$ denotes $\theta_{q}$ of the $k$ -th iteration. + +Therefore, if we compute the derivative $f_{q}^{\prime}$ beforehand as shown in the third row of Tab. 1, then this 3D-2D projection can be computed efficiently. + +# 3.2. Multifacet Mirror + +As is well known, observing the scene via a multi-facet mirror or a kaleidoscope is identical to observing the scene by virtual multi-view cameras, and in particular, kaleidoscopes with three mirrors are known to be reasonable in terms of less overlaps of mirrored images called discontinuities [29,33]. In our teleidoscopic imaging system, we use a three-facet mirror which defines reflections of the spherical refraction ray-pixel camera introduced in Sec. 3.1.1. + +The reflection $\pmb{p}^{\prime}$ of a 3D point $\pmb{p}$ by a mirror of normal $\pmb{n}_i$ and distance $d$ is given by + +$$ +\begin{array}{l} \boldsymbol {p} ^ {\prime} = H _ {i} \boldsymbol {p} + \boldsymbol {t} _ {i} \\ \Leftrightarrow \boldsymbol {p} ^ {\prime} = \left(\boldsymbol {I} - 2 \boldsymbol {n} _ {i} \boldsymbol {n} _ {i} ^ {\top}\right) \boldsymbol {p} + 2 d _ {i} \boldsymbol {n} _ {i}, \tag {11} \\ \end{array} +$$ + +where $H_{i}$ the Householder transformation matrix. + +In the case of kaleidoscopic imaging, the mirrors generate bouncing reflections as shown in Fig. 4. The reflection of $\pmb{p}^{\prime}$ by another mirror of normal $\pmb{n}_j$ and distance $d$ is given simply by + +$$ +\begin{array}{l} \boldsymbol {p} ^ {\prime \prime} = H _ {i j} \boldsymbol {p} + \boldsymbol {t} _ {i j} \\ = H _ {j} H _ {i} \boldsymbol {p} + 2 d _ {j} \boldsymbol {n} _ {j} + 2 H _ {j} d _ {i} \boldsymbol {n} _ {i}. \tag {12} \\ \end{array} +$$ + +![](images/37b04a7540fbc4886328706ad8ad31dcbd2b10f2182fda818b2beb54037eac76.jpg) +Figure 5. Depth-of-Field with Monocentric Lens. $D_{F}$ and $D_{N}$ denote the depth-of-field of the camera itself without the monocentric lens. The effective depth-of-field of the entire system $L_{F}$ and $L_{N}$ can be obtained by projecting $D_{N}$ and $D_{F}$ through the monocentric lens (the red and blue dashed lines). + +As a result, our teleidoscopic imaging system can be modeled as multi-view spherical refraction ray-pixel cameras. They share the single monocentric lens, and the pose of each virtual cameras can be computed by Eq. (12) if the mirror parameters are given. The later section describes our calibration algorithm to estimate such parameters. + +# 4. Depth-of-Field of Teleidoscopic Imaging System + +This section describes analytical evaluations on the depth-of-field of our teleidoscopic imaging system. We first review the depth-of-field of monocentric lens. + +# 4.1. Depth-of-Field with Monocentric Lens + +As shown in Fig. 5, depth-of-field with a monocentric lens can be simply defined as an extension of the path of the thin lens camera (cf. supplemental material). This is because the points in the near and the far depth-of-field $D_N$ and $D_F$ of thin lens are projected within the permissible circle-of-confusion even though the monocentric lens itself introduces the spherical aberration. Hence, we define the depth-of-field with a monocentric lens $L_N$ and $L_F$ as the intersection of the backprojection path through the edge of aperture (the red and blue dash lines in Fig. 5) and the backprojection path through the lens center $o_0$ . + +Because of the symmetric structure of the monocentric lens, the depth-of-field $L_{N} + L_{F}$ can be deeper than the original depth-of-field $D_{N} + D_{F}$ if the subject distance of the monocentric lens $s_{D} - r_{D}$ is larger than its focal distance $t_{D} - r_{D}$ . That is, there is a trade-off between magnification and depth-of-field even in a monocentric lens. The reason for the deep depth-of-field is due to the significant contribution of the monocentric lens to the far depth of field $L_{F}$ , and thus imaging with less blurring even when designed for high magnification (cf. Sec. 6.3). + +# 4.2. Depth-of-Field of Teleidoscopic Imaging System + +Fig. 6 illustrates the near and the far depth-of-fields of a virtual camera $C_i$ defined as a mirror of the original camera. + +![](images/a18d5b47c6ffef7520bc7ce301c581cdd35cb2819de0d3f9bdb07ad734e28a10.jpg) +Figure 6. Depth of Field of Teleidoscopic Imaging System. $D_N$ and $D_F$ denote the near and far depth-of-field of the thin lens camera $C$ respectively. All the points once imaged within $D_N - D_F$ should be originally emitted from the points within $L_N - L_F$ . + +Due to the aperture, the near and the far depth-of-fields form a curve respectively (the red and the blue dashed lines). + +Since the teleidoscopic imaging system has multiple virtual cameras as shown in Fig. 4, each of them has a different depth-of-field according to their relative pose to the monocentric lens. That is, the intersections of such depth-of-fields can be used for multi-view stereo reconstruction. + +# 5. Calibration of Teleidoscopic Imaging System + +This section introduces our calibration algorithm of teleidoscopic imaging system which requires capturing a single reference planar patterns. Fig. 7 shows the measurement model where the real camera $C$ observes reference points $\pmb{p}_w$ such as chessboard corners on a reference board II via a monocentric lens and three mirrors. Our calibration estimates the mirror normals $\pmb{n}_i$ , their distances $d_i$ from the camera, and the position and the radius of the monocentric lens, with assuming that the 2D positions of the reference points $\pmb{p}_w$ on the reference plane II, the intrinsic parameters of the camera, and the refraction index of the lens are given beforehand. + +A challenge in this calibration is the fact that the mirrors require observing 3D points and their reflections to estimate their poses [33], while the observation in the teleidoscopic system does not include such mirrored points that follow Eq. (12) due to the refraction by the monocentric lens. Similarly, the rays reflected and then refracted through the projection in teleidoscopic imaging system do not satisfy the coplanarity constraint. In addition, as mentioned in [2], the estimation of the monocentric lens parameters requires multiple viewpoints or multiple monocentric lenses. These points result in a chicken-and-egg problem. + +To solve this problem, we utilize the fact that the positions of the centers of the monocentric lens and its mirrors + +![](images/7c5db9c3e0095bcf17f4d780f894beecc7c9bf8ade56c6513dde23420475e205.jpg) +Figure 7. Calibration of Teleidoscopic Imaging System. Cyan lines denote the light paths for a mirrored camera. The calibration can be conducted by capturing reference points on a plane of a known geometry through the monocentric lens. + +are always captured without refraction by definition, and satisfy Eq. (12). Therefore, we start by estimating the center of the monocentric lens from the captured reference points. + +In what follows, $\pmb{p}_p^{(0)}$ denotes the real image of a reference point $\pmb{p}_w$ in the camera $C$ . Similarly $\pmb{p}_p^{(i)}$ ( $i = \{1,2,3\}$ ) denotes the image of its first reflection by the mirror $i$ , and $\pmb{p}_p^{(ij)}$ ( $ij = \{12,13,21,23,31,32\}$ ) denotes its second reflection by the mirrors $i$ and $j$ . + +Axis of Monocentric Lens The axis to the center of monocentric lens $\pmb{a}_v^{(0)}$ from the camera $C$ can be estimated by $\pmb{p}_w - \pmb{p}_P^{(0)}$ correspondences. Similarly to [2], we estimate $\pmb{a}_v^{(0)}$ by the coplanarity constraint of the three vectors $\pmb{a}_v^{(0)}$ , $\pmb{p}_p^{(0)}$ , and $\pmb{p}_w$ on a single plane-of-refraction: + +$$ +\begin{array}{l} \boldsymbol {p} _ {p} ^ {(0) \top} \left(\boldsymbol {a} _ {v} ^ {(0)} \times \left(R _ {\Pi} \boldsymbol {p} _ {w} + \boldsymbol {t} _ {\Pi}\right)\right) = 0, \tag {13} \\ \Leftrightarrow \boldsymbol {p} _ {p} ^ {(0) \top} \left(E _ {\Pi} \boldsymbol {p} _ {w} + \boldsymbol {s} _ {\Pi}\right)) = 0, \\ \end{array} +$$ + +where $R_{\Pi}, t_{\Pi}$ are the pose of the reference plane in the camera coordinate system and $E_{\Pi} = \boldsymbol{a}_v^{(0)} \times R_{\Pi}$ and $s_{\Pi} = \boldsymbol{a}_v^{(0)} \times t_{\Pi}$ . Since this is a linear equation with 9 unknown parameters of $E_{\Pi}$ and $s_{\Pi}$ , we can obtain $E_{\Pi}$ and $s_{\Pi}$ up to scale by observing at least 8 points on the reference plane. Using the estimated $E_{\Pi}$ , the axis $\boldsymbol{a}_v^{(0)}$ is given by + +$$ +\boldsymbol {a} _ {v} ^ {(0)} = \frac {E _ {\Pi} (: , 1) \times E _ {\Pi} (: , 2)}{| | E _ {\Pi} (: , 1) \times E _ {\Pi} (: , 2) | |}. \tag {14} +$$ + +Similarly, the axis to the center of the mirrored monocentric lens $\pmb{a}_v^{(i)}$ can be obtained by the mirrored points $\pmb{p}_p^{(i)}$ . + +Mirror Normals The axis to the center of the monocentric lens $\pmb{a}_{v}^{(0)} = (x_{a_{v}}^{(0)},y_{a_{v}}^{(0)},z_{a_{v}}^{(0)})^{\top}$ and its mirror $\pmb{a}_{v}^{(i)} = (x_{a_{v}}^{(i)},y_{a_{v}}^{(i)},z_{a_{v}}^{(i)})^{\top}$ satisfies + +$$ +\boldsymbol {a} _ {v} ^ {(0) \top} [ \boldsymbol {n} _ {i} ] _ {\times} \boldsymbol {a} _ {v} ^ {(i)} = 0, \tag {15} +$$ + +where $[\pmb{n}_i]_{\times}$ denotes the skew-symmetric matrix defined by the normal $\pmb{n}_i = (x_{n_i},y_{n_i},z_{n_i})^\top$ of the mirror $i$ . + +The same constraint holds for each of the first-second reflection pairs $\pmb{a}_v^{(i)} - \pmb{a}_v^{(ij)}$ about the same mirror normal $\pmb{n}_i$ [33]. Therefore $\pmb{n}_i$ can be obtained linearly only from the axes to the centers of the monocentric lenses. + +Mirror Distances Once the mirror normals are estimated, we can utilize the kaleidoscopic triangulation [33] to obtain linear constraints on the mirror distances $d_{i}$ (See supp. Sec. 2 for details). + +Pose of Reference Plane Similarly to Eq. (13), the plane-of-refraction constraint holds for the mirrored cameras $C_i$ : + +$$ +\left(\boldsymbol {p} _ {p} ^ {(i)}\right) ^ {\top} \left(\boldsymbol {a} _ {v} ^ {(i)} \times \left(H _ {i} \left(R _ {\Pi} \boldsymbol {p} _ {w} + \boldsymbol {t} _ {\Pi}\right) + \boldsymbol {t} _ {i}\right)\right) = 0. \tag {16} +$$ + +This constraint allows us estimating the pose of the reference plane $R_{\Pi}$ and $\pmb{t}_{\Pi}$ linearly. + +Monocentric Lens Parameters The calibration algorithm up to this point does not require the monocentric lens parameters $d_{D}, r_{D}$ , and $\mu_{g}$ . We estimate these parameters by the coplanarity constraint of the ray through $\pmb{p}_{p}^{(i)}$ of $C_{i}$ : + +$$ +\boldsymbol {v} _ {o} ^ {(i)} \times \left(\boldsymbol {p} _ {w} ^ {(i)} - \boldsymbol {p} _ {o} ^ {(i)}\right) = 0. \tag {17} +$$ + +This is a nonlinear constraint for the monocentric lens parameters as described in Sec. 3.1 and we solve this as a nonlinear optimization problem with assuming their rough estimates are available in practice. + +**Bundle Adjustment** The last step of our calibration is to refine the parameters $\boldsymbol{a}_v^{(0)}, d_D, r_D, \mu_g, \boldsymbol{n}_i, d_i, R_{\Pi}, t_{\Pi}$ ( $i = 1, 2, 3$ ) by minimizing the reprojection errors of the reference points $\boldsymbol{p}_w$ as a nonlinear optimization problem. Here the forward projection is computed using the method described in Sec. 3.1.2. + +Once the teleidoscopic imaging system parameters are calibrated, a scene point can be linearly triangulated in a DLT manner [13] from its teleidoscopic projections (See supp. Sec. 3 for details). + +# 6. Evaluation + +This section provides quantitative evaluations of the proposed numerical 3D-2D projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also demonstrate the effectiveness of our method by reconstructing real-world microscale objects. + +# 6.1. Computational Efficiency of 3D-2D Projection + +The average computation cost of a single forward projection by our method and by the analytical solution solving a 10th-degree equation for each projection [2] are $0.62\mathrm{ms}$ and $3.93\mathrm{ms}$ respectively. These runtime costs are the average values of 100 trials of 10K points forward projections + +![](images/03031ee35174fe2150d6e749dfa273a836c69c3dcdc6d5503a054d2751f0ec68.jpg) +(a) + +![](images/de6b2965b32579d49160ab7c99ead40090053d152a4407775fd497c3b08e5d06.jpg) +(b) + +![](images/5b7d6ec1ded5a65e174362d74ed5a867250a9f86437daa10a7e95ee4602f41a7.jpg) + +![](images/93b89c7865b2d4e25fe093997d934f490fdc280288507f4a9faf76efc13829f7.jpg) +Figure 8. Teleidoscopic System for Quantitative Evaluation. (a) Simulated measurement environment. (b) Reprojection errors at different noise levels. The bars denote the standard deviation of the errors. + +![](images/804146f7ae0a0ab5d7205f8da0c8f32e9aa91c41ea8ebb537a1c9d6ed263fdca.jpg) +Figure 11. Changes in Depth-of-Field by the Monocentric Lens. (a), (b) and (c) show the total, near, and far depth-of-field with a monocentric lens. The red and blue plots indicate the results with and without the monocentric lens respectively. The green lines indicate the subject distance where the subject and the focal distances of the monocentric lens itself are same. + +![](images/fbe50aa59c1daf80a7c3d094de19608c47dab964072df6bb9eda6568d254e337.jpg) +(a) +Figure 9. Calibration Errors. From left to right, the upper figures show the (a) angular error of the axis to the monocentric lens center (degree), (b) angular error of the mirror normal (degree), and (c) rotation error (Riemannian distance) of reference plane. The lower ones are the distance errors normalized by the ground truth. Here the distance of the monocentric lens is only estimated in a nonlinear manner (Sec. 5). + +![](images/37b0a7a601c404758d8d79bdf8e2b25cadb137d39fe8be824a9e41cae583508d.jpg) +(b) + +![](images/23d0b9f4d92190129dcd0363d47aa707486abb095172c74d10ad44e25a2d06d8.jpg) +(c) + +run in Matlab on an Intel Core-i7 2.6GHz PC. This result clearly verifies that our numerical forward projection based on our ray-pixel model can realize an efficient projection. + +# 6.2. Calibration + +# 6.2.1 Quantitative Evaluation using Synthesized Data + +Fig. 8(a) shows the measurement environment which simulates the real capture system used in Sec. 6.2.2. The system has a kaleidoscope with three $10 \times 30\mathrm{mm}$ mirrors in front of the camera $C$ . The mirrors are at slightly off-perpendicular angle of $1.4^{\circ}$ to the camera image plane so that the mirrors define virtual cameras around the target with less overlaps of the mirrored images. The system also has a monocentric lens of $10\mathrm{mm}$ diameter in front of the mirrors, at $40\mathrm{mm}$ distance from the camera. The refraction index $\mu_{g}$ is set to 2.0. The system captures 48 reference points (blue dots in Fig. 8(a)) to calibrate its parameters. By injecting Gaussian + +![](images/9859d492b60fb7b88a47003e31d41c09930e0848df079b3586fc222e6d53d8ed.jpg) +(a) + +![](images/7cc7d065929938b3447e28105a16ae5bc9b9424c6118b50f63dad7e3555ea0e4.jpg) +(b) + +![](images/ce54496830517d8b910b8e59a4d940185f48f7a08350852b88d66a965e05be5b.jpg) +Figure 10. Experimental Setup and Calibration. (a) Our teleidoscopic imaging system consists of a camera, a projector, three mirrors $(10 \times 30\mathrm{mm})$ , and a monocentric lens (5mm radius, refraction index $\mu_g = 2.0$ ). (b) Input image for calibration. +(a) + +![](images/3546a9cce26579489d9b7920f9b824b83ab86a45d15f57a15cf7575ef81b8588.jpg) +(b) + +![](images/2016e58bc7025fcec3e565772d928baaa8617c757bdcaca3c43efb5256c6c8b0.jpg) +(c) + +noise of different standard deviations $\sigma$ to the 2D positions of their projections, we evaluate the robustness of our calibration procedure. Fig. 8(b) shows average reprojection errors in pixel of 100 trials at each pixel noise level $\sigma$ . We can observe that the reprojection errors increase linearly against the pixel noise level. + +Fig. 9 shows the estimation errors of the monocentric lens parameters, the mirror parameters, and the reference plane parameters respectively. These results indicate that our calibration algorithm performs reasonably under realistic observation noise. + +# 6.2.2 Quantitative Evaluation using Real Data + +Fig. 10(a) shows an overview of our teleidoscopic imaging system. The system consists of a FLIR Blackfly BFS-U3-89S6C-C camera (4096×2160 resolution, pixel size 3.45μm) with a f=4mm / F8 lens, three 10 × 30mm mirrors, a monocentric lens of 10mm diameter whose refraction index is $\mu_g = 2.0$ , and a LED projector ASUS E1-R for casting Gray code patterns onto the target surface to identify corresponding points [12, 23]. Note that the effective resolution of each "virtual camera" image is approximately 300×200 (16 "virtual cameras" in total), which is calculated from the area inside the base chamber with one horizontal edge of the triangle and inside the spherical lens. + +We used a single image (Fig. 10(b)) capturing $5 \times 8$ corners of a checkerboard for calibration, and used up to second reflections in the image. The average reprojection er + +![](images/2e84516d714b8c73c2142fd0f160d7400f800b813363becc3af92ef5974be9c2.jpg) +(a) + +![](images/29442c3ae27b96808aed4d56cdb952e1993fe90ac6d6daeecb0af5381f12ef1f.jpg) +(b) +Figure 12. Evaluation with a sphere (radius: $3\mathrm{mm}$ ). (a) Ground truth. (b) Recovered 3D shape. The colors visualize the $z$ -coordinate of the reconstructed shape. (c) Error map. $E_{d}$ is the mean absolute error in millimeter. + +![](images/86fae6deb7684254fb2f0ca1c2117761dbe701f5eacedc374726bc50e95af33f.jpg) + +![](images/a41e30a812ea681b3cd2731f5197c9da69cb242e354fa8752fbeb895161e04c6.jpg) +(c) + +![](images/ea5c35744b6ecea20b8f6dad9641c08b4065274b818ed69f9de81f5115219d36.jpg) + +ror is 0.62px, and the average distance error of the triangulated 3D points to their ground truth positions was $0.17\mathrm{mm}$ . The size of reference checkerboard is around $15\mathrm{mm}$ width, hence the relative error of the triangulation is around $1\%$ . These results show our method provide a practical calibration and the valid scale of the target object also depends on size of the checkerboard. + +# 6.3. Depth-of-Field + +Fig. 11 shows the changes in the depth-of-field by the monocentric lens capturing an object of $0.5\mathrm{mm}$ height at the same magnification given the configuration same as the real capture in Sec. 6.2.2. Fig. 11(a) shows the total depth-of-field with and without the monocentric lens $L_{N} + L_{F}$ (red and blue respectively). Fig. 11(b) and (c) show $L_{N}$ and $L_{F}$ in the same manner. In these figures, the horizontal axis indicates the apparent size of the target in pixel, with assuming the measurement model used in Sec. 6.2. These plots clearly indicate that the monocentric lens contributes to increase the depth-of-field successfully. + +# 6.4. Teleidoscopic 3D Shape Reconstruction + +To validate the proposed teleidoscopic system as a multiview camera system for 3D shape reconstruction, this section demonstrates 3D reconstruction of small objects of approximately $6\mathrm{mm}$ size. + +We used the system calibrated in Sec. 6.2.2. Fig. 12 visualizes our 3D reconstruction of the front side of a real sphere of $3\mathrm{mm}$ radius with the ground-truth and the error map. The mean absolute error $E_{d}$ of the estimated depth is $0.097\mathrm{mm}$ . Compared with its size, the relative error is about $1.6\%$ . + +Fig. 13 shows additional 3D reconstructions. The 1st and the 2nd rows demonstrate that our imaging system has the ability to recover complex shapes. The proposed imaging system well recovers the complex bumpy parts of the shell and the sharp edges and corners of the object. The 2nd and the 3rd rows compare "with" and "w/o" the monocentric lens cases. As the input image (b) shows, the monocentric lens creates a surrounding view while preserving the area of each chamber, resulting in a fuller and more accurate geometry, as shown in (c). We can verify this quantitatively by comparing the numbers of pixels covering the object area in the captured image (145,217 vs. 12,730) and the numbers + +![](images/1e840fa03f603c82fcaf928e8bdf7c04a9544cc1b12cf990cf12606354ade632.jpg) + +![](images/c70bd838d0cd6cdab51399b0df5a82a973c9cf33474578042c407494162d3c4d.jpg) + +![](images/1190d819cbdbc82425c5883254016964c4fff978aa3088bec2a748233951388c.jpg) + +![](images/08b5901c897b76425a4eb169698d777219de370683da0de430003ab576c992b3.jpg) + +![](images/9713e13f344a192e93d1f4bf477753e01a2f46be450fda7b6def1b6ba5523c44.jpg) + +![](images/d76b5caa8d46a9081d1b61545d94e610dcb7e85015a6ec5ffed84de36963e6ef.jpg) +40,639 pts recovered + +![](images/36f1a76d08c9302bf58ab2e161b52e9e87a3cce888471e1f2b586db192f1cbeb.jpg) + +![](images/011618e35a7c6f8b9d23e0ca58837101d5f693a6e5b266310edfbf05dc51cfe2.jpg) +w/o monocentric lens +(b) +Figure 13. Results of 3D Shape Estimation. The second and the third row compares "with" and "w/o" the monocentric lens. (a) Room-light appearance. (b) Teleidoscopic image. (c) Reconstructed 3D shapes. + +![](images/c88bbca32efa14ac1fbdf8f46b9490602bbb756aed8546a90df57fd3cdcf9592.jpg) + +![](images/444d6283fdb898dd71ccf9d0bff27a2e3e7237613ae2a3cb5e603ab77c458349.jpg) +8,264 pts recovered +(c) + +of triangulated 3D points using the ground-truth correspondences as shown in the figure (40,639 vs. 8,264). See the supplementary material for more results. + +Limitations The major limitations of our method are twofold. The monocentric lens has a relatively large spherical aberration, and the optical resolution of the entire system is limited by such aberrations. One possible solution is to introduce nested concentric glass spheres with different refractive indices for the monocentric lens [6, 35]. The other limitation is the trade-off between the number of virtual viewpoints and the effective resolution of each view because the use of a kaleidoscope requires zooming-out to capture the reflected images. While larger number of the cameras contributes to robustify the triangulation, it inevitably reduces the resolution of the reconstruction as the number of pixels per view decreases. + +# 7. Conclusion + +This paper proposed the concept of image-based 3D measurement by a teleidoscopic imaging system for microscale objects. By combining a monocentric lens with a kaleidoscope, our system realized a closeup and surrounding multi-view capture of the target. Our contributions are twofold. We proposed a ray-pixel camera model that realized an efficient numerical 3D-2D projection computation, and also proposed a calibration method for the teleidoscopic imaging system. Our future work includes aberration and diffraction blur handling such as the inverse filter proposed in [5], and also refraction/attenuation modeling in the case of capturing objects in water [18, 19]. + +# Acknowledgements + +This work is in part supported by JSPS KAKENHI JP22K17914. + +# References + +[1] Rajat Aggarwal, Amrisha Vohra, and Anoop M. Namboodiri. Panoramic stereo videos with a single camera. In CVPR, pages 3755-3763, 2016. 2 +[2] Amit Agrawal and Srikumar Ramalingam. Single image calibration of multi-axial imaging systems. In CVPR, pages 1399–1406, 2013. 2, 3, 4, 5, 6 +[3] Byeongjoo Ahn, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Kaleidoscopic structured light. ACM TOG, 40(6):1-15, 2021. 2 +[4] Michael Chen, Zachary F. Phillips, and Laura Waller. Quantitative differential phase contrast (dpc) microscopy with computational aberration correction. Optics Express, (25):32888-32899. 1 +[5] O. S. Cossairt, D. Miau, and S. K. Nayar. Gigapixel computational imaging. In ICCP, pages 1-8, 2011. 2, 8 +[6] Donald G. Dansereau, Glenn Schuster, Joseph Ford, and Gordon Wetzstein. A wide-field-of-view monocentric light field camera. In CVPR, 2017. 2, 8 +[7] Michael De Zeeuw and Aswin C Sankaranarayanan. Widebaseline light fields using ellipsoidal mirrors. 2022. 1 +[8] Keith Forbes, Fred Nicolls, Gerhard De Jager, and Anthon Voigt. Shape-from-silhouette with two mirrors and an uncalibrated camera. In ECCV, 2006. 2 +[9] Brian J. Ford. Single lens: The story of the simple microscope. Journal of the History of Biology, 19(2):320-321, 1986. 2 +[10] Joshua Gluckman, Shree K. Nayar, and Keith J. Thoresz. Real-time omnidirectional and panoramic stereo. In Proceedings of the 1998 DARPA Image Understanding Workshop, pages 299-303. Morgan Kaufmann, 1998. 2 +[11] Michael D. Grossberg and Shree K. Nayar. The raxel imaging model and ray-based calibration. *IJCV*, 61(2):119–137, 2005. 1, 2 +[12] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G. Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR, pages 713-720, 2011. 7 +[13] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 2, 6 +[14] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE TPAMI, 41(1):190-204, 2019. 1 +[15] Srinivasa Karthik and Ambasamudram N. Rajagopalan. Underwater microscopic shape from focus. In ICPR, pages 2107-2112, 2014. 1 +[16] Ryo Kawahara, Shohei Nobuhara, and Takashi Matsuyama. A pixel-wise varifocal camera model for efficient forward projection and linear extrinsic calibration of underwater cameras with flat housings. In Proc. of ICCV 2013 Underwater Vision Workshop, pages 819-824, 2013. 2 +[17] Gurunandan Krishnan and Shree K. Nayar. Towards a true spherical camera. In SPIE, volume 7240, 724002, pages 1-13, 2009. 2 + +[18] Meng-Yu Jennifer Kuo, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Non-rigid shape from water. IEEE TPAMI, 43(7):2220-2232, 2021. 8 +[19] Meng-Yu Jennifer Kuo, Satoshi Murai, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Surface normals and shape from water. IEEE TPAMI, 44(12):9150-9162, 2021. 8 +[20] Douglas Lanman, Daniel Crispell, and Gabriel Taubin. Surround structured lightning: 3-d scanning with orthographic illumination. In CVIU, pages 1107-1117, 2009. 2 +[21] A. Manakov, J. Restrepo, O. Klehm, R. Hegedus, E. Eisemann, H. P. Seidel, and I. Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM TOG, 32(4):47-1, 2013. 1 +[22] P. Miraldo and H. Araujo. Calibration of smooth camera models. IEEE TPAMI, 35(9):2091-2103, 2013. 2 +[23] Daniel Moreno and Gabriel Taubin. Simple, accurate, and robust projector-camera calibration. In 3DPVT, pages 464-471, 2012. 7 +[24] Andrew D. Mullen, Tali Treibitz, Paul LD Roberts, Emily L. A. Kelly, Rael Horwitz, Jennifer E. Smith, and Jules S. Jaffe. Underwater microscopy for in situ studies of benthic ecosystems. Nature Communications, 7, 2016/07 2016. 1 +[25] Shree K. Nayar. Omnidirectional video camera. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 235-241, 1997. 2 +[26] Markus Niederost, Jana Niederost, and Jiri Scucka. Automatic 3d reconstruction and visualization of microscopic objects from a monoscopic multifocus image sequence. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XXXIV-5/W10:5-10, 01 2002. 1 +[27] Ryoichiro Nishi, Takahito Aoto, Norihiko Kawai, Tomokazu Sato, Yasuhiro Mukaigawa, and Naokazu Yokoya. *Ultrashallow dof imaging using faced paraboloidal mirrors*. In ACCV, pages 115–128, 2016. 2 +[28] Ilya Reshetouski and Ivo Ihrke. Mirrors in Computer Graphics, Computer Vision and Time-of-Flight Imaging, pages 77–104. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013. +[29] Ilya Reshetouski, Alkhazur Manakov, Hans-Peter Seidel, and Ivo Ihrke. Three-dimensional kaleidoscopic imaging. In CVPR, pages 353-360, 2011. 4 +[30] J. Starck, A. Maki, S. Nobuhara, A. Hilton, and T. Matsuyama. The multiple-camera 3-d production studio. IEEE TCSVT, 19(6):856-869, 2009. 1 +[31] Peter Sturm. Multi-view geometry for general camera models. In CVPR, volume 1, pages 206-212, 2005. 2 +[32] Seiichi Tagawa, Yasuhiro Mukaigawa, Jaewon Kim, Ramesh Raskar, Yasuyuki Matsushita, and Yasushi Yagi. Hemispherical confocal imaging. IPSJ Transactions on Computer Vision and Applications, 3:222-235, 2011. 2 +[33] Kosuke Takahashi, Akihiro Miyata, Shohei Nobuhara, and Takashi Matsuyama. A linear extrinsic calibration of kaleidoscopic imaging system from single 3d point. In CVPR, 2017. 1, 2, 4, 5, 6 + +[34] Lei Tian, Jingyan Wang, and Laura Waller. 3d differential phase-contrast microscopy with computational illumination using an led array. Optics letters, 39:1326-9, 2014. 1 +[35] Eric J Tremblay, Daniel L Marks, David J Brady, and Joseph E Ford. Design and scaling of monocentric multiscale imagers. Applied Optics, 51(20):4691-4702, 2012. 2, 8 +[36] Borislav Trifonov, Derek Bradley, and Wolfgang Heidrich. Tomographic reconstruction of transparent objects. In Proc. of Eurographics Conference on Rendering Techniques, pages 51-60, 2006. 2 +[37] David Uhlig and Michael Heizmann. A calibration method for the generalized imaging model with uncertain calibration target coordinates. In ACCV. 2 +[38] Frank Verbiest, Marc Proesmans, and Luc Van Gool. Modeling the effects of windshield refraction for camera calibration. In ECCV, pages 397-412, 2020. 2 \ No newline at end of file diff --git a/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/images.zip b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d2ab8a28f33af721f306202eefa6f1afec8db3cf --- /dev/null +++ b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c570e5bfcdd2d9fe243613d24bc1dae367339ba09664fc34e03ec20b9d211ef +size 467628 diff --git a/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/layout.json b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..58ae2b900aa42975cd0e4de658c817acf6dd2b11 --- /dev/null +++ b/2023/Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction/layout.json @@ -0,0 +1,13592 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "Teleidoscopic Imaging System for Microscale 3D Shape Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 194, + 218, + 307 + ], + "blocks": [ + { + "bbox": [ + 55, + 194, + 218, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 218, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 218, + 307 + ], + "type": "image", + "image_path": "22ce7a0da53799a283a040951791c760f89a62c4d805631443c2a3890330a18a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 546, + 344 + ], + "type": "text", + "content": "Figure 1. Teleidoscopic Imaging System. (a) We propose a novel system for microscale 3D imaging using an imaging setup that consists of a camera, a kaleidoscopic multi-facet mirror with a monocentric front lens, and a projector. (b) Our system realizes a dense close-up surrounding multi-view capture of the target. (c) Our method enables fuller and dense reconstruction of microscale objects." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 193, + 361, + 307 + ], + "blocks": [ + { + "bbox": [ + 80, + 144, + 235, + 185 + ], + "lines": [ + { + "bbox": [ + 80, + 144, + 235, + 185 + ], + "spans": [ + { + "bbox": [ + 80, + 144, + 235, + 185 + ], + "type": "text", + "content": "Ryo Kawahara \nKyushu Institute of Technology \nrkawahara@ai.kyutech.ac.jp" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 193, + 361, + 307 + ], + "lines": [ + { + "bbox": [ + 225, + 193, + 361, + 307 + ], + "spans": [ + { + "bbox": [ + 225, + 193, + 361, + 307 + ], + "type": "image", + "image_path": "d93f9aa0dd94cd442af7962111403cd3b0ebbedd03dd492586eca05959dcd061.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 363, + 193, + 478, + 307 + ], + "blocks": [ + { + "bbox": [ + 261, + 144, + 380, + 184 + ], + "lines": [ + { + "bbox": [ + 261, + 144, + 380, + 184 + ], + "spans": [ + { + "bbox": [ + 261, + 144, + 380, + 184 + ], + "type": "text", + "content": "Meng-Yu Jennifer Kuo University of Minnesota kuo00052@umn.edu" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 363, + 193, + 478, + 307 + ], + "lines": [ + { + "bbox": [ + 363, + 193, + 478, + 307 + ], + "spans": [ + { + "bbox": [ + 363, + 193, + 478, + 307 + ], + "type": "image", + "image_path": "0cfedaef57b0e38b012017ff3114aa0cfb48426e45e7fcea13244e7c9d0bcf83.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 486, + 193, + 541, + 294 + ], + "blocks": [ + { + "bbox": [ + 405, + 144, + 510, + 185 + ], + "lines": [ + { + "bbox": [ + 405, + 144, + 510, + 185 + ], + "spans": [ + { + "bbox": [ + 405, + 144, + 510, + 185 + ], + "type": "text", + "content": "Shohei Nobuhara Kyoto University nob@i.kyoto-u.ac.jp" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 486, + 193, + 541, + 294 + ], + "lines": [ + { + "bbox": [ + 486, + 193, + 541, + 294 + ], + "spans": [ + { + "bbox": [ + 486, + 193, + 541, + 294 + ], + "type": "image", + "image_path": "668d72d9a3758679c630e887d42512a46a2f017998456eb55bf41a8152504fd7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 354, + 192, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 354, + 192, + 366 + ], + "spans": [ + { + "bbox": [ + 143, + 354, + 192, + 366 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 380, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 571 + ], + "type": "text", + "content": "This paper proposes a practical method of microscale 3D shape capturing by a teleidoscopic imaging system. The main challenge in microscale 3D shape reconstruction is to capture the target from multiple viewpoints with a large enough depth-of-field. Our idea is to employ a teleidoscopic measurement system consisting of three planar mirrors and monocentric lens. The planar mirrors virtually define multiple viewpoints by multiple reflections, and the monocentric lens realizes a high magnification with less blurry and surround view even in closeup imaging. Our contributions include, a structured ray-pixel camera model which handles refractive and reflective projection rays efficiently, analytical evaluations of depth of field of our teleidoscopic imaging system, and a practical calibration algorithm of the teleidoscopic imaging system. Evaluations with real images prove the concept of our measurement system." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 585, + 127, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 127, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 127, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Microscale 3D reconstruction has found profound applications in a wide range of domains including medical imaging, life science, and aquaculture, due to its non-constrained and non-invasive measurements. The main challenges in image-based microscopic 3D shape measurement is its shallow depth of field and camera arrangement in the closeup scenario. Applying conventional multiple camera system designed for human-size capture [14, 30] cannot be a feasible solution due to limitations on camera placement. Con-" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 355, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 392 + ], + "type": "text", + "content": "ventional multiple mirror system [33] also have difficulties inevitably in depth-of-focus due to differences in their optical path lengths with varying numbers of bounces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 396, + 546, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 647 + ], + "type": "text", + "content": "In this paper, we show that the fuller 3D shape of a microscale object can be recovered. Our key idea is to employ a catadioptric imaging system which realizes a practical closeup multi-view imaging. The point of our design is that the system has a monocentric front lens like a teleidoscope, instead of using microscopic system in the camera side. That is, as shown in Fig. 1(a), we introduce a kaleidoscopic multi-facet mirror between the front lens and the camera, where the design realizes a deeper depth-of-field and results in less blurring imaging. Unlike conventional microscopic imaging system such as differential phase contrast (DPC) microscopy [4, 34] and multi-focus approaches [15, 24, 26], our method realizes a multi-view capture of the target from a single physical viewpoint which can contribute to free-viewpoint rendering, 3D shape reconstruction, and reflection analysis. Our system also has an advantage over existing imaging systems that build multiple views behind the main lens [7, 21] in a closeup environment. The wide FoV with a monocentric main lens and virtual multiple views with mirrors allow closeup and surrounding view capturing in focus." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "We call our system teleidoscopic imaging system and show that the system can be compactly modeled by a structured ray-pixel camera model [11], which handles refractive and reflective projection rays efficiently. Based on our ray-pixel camera model, we derive a practical calibration algo" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20813" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "rithm that estimates the positions of the monocentric lens and multi-facet mirror w.r.t. the camera by using a single reference planar pattern (i.e., a checkerboard). Given the calibration parameters, a scene point can then be linearly triangulated from its teleidoscopic projection in a direct linear transform (DLT) manner [13]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 299 + ], + "type": "text", + "content": "We implement our method with an imaging system consisting of a camera, three planar mirrors, a monocentric lens, and a projector placed outside the mirrors for structured-light casting. We quantitatively evaluate the computation cost of numerical projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also validate the effectiveness of our method qualitatively on a number of real-world microscale objects. These results demonstrate that our method enables holistic and dense reconstruction of microscale objects. We believe our method expands the avenues of three-dimensional analysis of microscale objects and scenes in real world scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 309, + 134, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 134, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 134, + 321 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "type": "text", + "content": "Our teleidoscopic imaging system is a catadioptric system consisting of a monocentric lens and kaleidoscopic mirrors (Fig. 1(a)). While catadioptric system has a wide variety of applications such as omnidirectional observation and panoramic stereo [1, 10, 25, 27], this section reviews studies with multiple flat mirrors and monocentric lenses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 236, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 236, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 236, + 419 + ], + "type": "text", + "content": "2.1. Imaging with Multiple Flat Mirrors" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 533 + ], + "type": "text", + "content": "A fundamental motivation of introducing mirrors in observation system is to increase the number of viewpoints without installing additional cameras for multi-view capture of a target [3, 8, 20, 28, 32, 33]. Takahashi et al. [33] have proposed a kaleidoscopic imaging system and demonstrated a 3D shape reconstruction using multiple reflections. Tagawa et al. [32] have proposed a multi-facet imaging system that observes a target from an equally distributed virtual cameras for reflectance analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 220, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 220, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 220, + 552 + ], + "type": "text", + "content": "2.2. Imaging with Monocentric Lens" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 689 + ], + "type": "text", + "content": "In the context of imaging system, the monocentric lens is often used to obtain a wide field-of-view [17, 35], in particular for endoscopes, as a short focal length lens. The monocentric lens, however, has additional useful characteristics: its symmetric structure and magnifying power. Cossair et al. [5] have proposed a camera array which captures a same scene through a single monocentric lens so that the images can be stitched into a single high-resolution image. Similarly, Dansereau et al. [6] have proposed a lightfield camera which captures omnidirectional lightfield images through a single monocentric lens with a camera orbiting around it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 712 + ], + "type": "text", + "content": "A typical use of the monocentric lens as a magnifier can be found in the Leeuwenhoek's microscope in the 17th cen" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 541, + 191 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 541, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 191 + ], + "type": "image", + "image_path": "4917e17c664ac641403725b7043a4d08968a5a6af8531e5094fb63e1dfa86bb4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "type": "text", + "content": "Figure 2. Measurement through Monocentric Lens. The segments of projection path " + }, + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "type": "inline_equation", + "content": "\\ell_{o} - \\ell_{g} - \\ell_{i}" + }, + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "type": "text", + "content": " have an axially symmetric structure around the axis which directs to the monocentric lens center " + }, + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "type": "inline_equation", + "content": "o_{D}" + }, + { + "bbox": [ + 305, + 204, + 545, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 251, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 275 + ], + "type": "text", + "content": "tury. It utilizes a single monocentric lens and realized over " + }, + { + "bbox": [ + 306, + 251, + 545, + 275 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 306, + 251, + 545, + 275 + ], + "type": "text", + "content": " magnifications [9]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 284, + 446, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 446, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 446, + 297 + ], + "type": "text", + "content": "2.3. Ray-pixel Camera Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 303, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 483 + ], + "type": "text", + "content": "In order to handle light paths captured via complex imaging system such as reflection and refraction, Grossberg and Nayar [11] have proposed a ray-pixel (or raxel) camera model which describes the projection by end-to-end mappings between the pixels and their corresponding emergent rays. While this pixel-wise representation has a great flexibility to describe complicated projections [22, 31], it is also possible to exploit a structure of the rays to simplify the model [16, 22, 36-38]. Kawahara et al. [16] have proposed a ray-pixel camera model for cameras observing the scene through flat surfaces. Compared with analytical forward projections through flat surfaces which requires solving a high order equation [2], it realized a practical numerical forward projection computation by exploiting the axially symmetric structure of the rays." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 495, + 545, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 545, + 522 + ], + "type": "text", + "content": "3. Ray-Pixel Camera Model for Teleidoscopic Imaging System" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 529, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 589 + ], + "type": "text", + "content": "In our teleidoscopic system, the cameras mirrored by the kaleidoscopic multi-facet mirrors observe the scene through the single monocentric lens (Fig. 1(a)) as done in [5,6], and each of the monocentric lens and the camera pairs form a microscope capturing the object from a different direction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 545, + 638 + ], + "type": "text", + "content": "In this section, we show that this design realizes a microscale multi-view capturing with a deep depth-of-field, and that the projections by the teleidoscopic imaging system can be modeled compactly by ray-pixel cameras." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 520, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 520, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 520, + 658 + ], + "type": "text", + "content": "3.1. Measurement through Monocentric Lens" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 2, suppose a perspective camera whose camera center is located at " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{o}_0" + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": " observes the scene through a monocentric lens located at " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{o}_D" + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": ". Here the perspective camera can be assumed to be directed to the center" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20814" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "of the monocentric lens " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_D" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " without loss of generality, since we can calibrate the rotation to align the optical axis of the camera to the line " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_D - \\mathbf{o}_0" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " as described later. We denote this normalized camera by " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": "Obviously, the rays backprojected through pixels of " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": " have an axially symmetric structure around the optical axis, and an incident ray " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "\\ell_i" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": " through a pixel " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "p_p" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": " and its refraction " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "\\ell_g" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": " and emergent ray " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": " are on a single plane-of-refraction [2]. Therefore, we can describe the ray through a single pixel by a 2D " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "(r,z)^{\\top}" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": " coordinate system centered at " + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "inline_equation", + "content": "o_0" + }, + { + "bbox": [ + 46, + 121, + 286, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": "Hereafter, let " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "r_{\\alpha}" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "z_{\\alpha}" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " elements of the vector " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " in general. For example, point " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " is described as " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p = (r_{\\pmb{p}_p}, z_{\\pmb{p}_p})^\\top" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": ". Also let " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\pmb{v}_X = (r_{\\pmb{v}_X}, z_{\\pmb{v}_X})^\\top" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " denote the direction vector of line " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "\\ell_X" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " towards the scene from the camera, where " + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "inline_equation", + "content": "X = \\{\\mathrm{i}, \\mathrm{g}, \\mathrm{o}\\}" + }, + { + "bbox": [ + 47, + 196, + 286, + 267 + ], + "type": "text", + "content": " represents the label of the light path segment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "content": "Consider the projection path " + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\ell_o - \\ell_g - \\ell_i" + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "inline_equation", + "content": "o_0" + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "content": " through the point " + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p" + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "content": ". The incident point " + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "content": " on the sphere is described as a function of " + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "inline_equation", + "content": "\\theta_p = \\tan^{-1}(r_{\\pmb{p}_p} / f_c)" + }, + { + "bbox": [ + 47, + 269, + 286, + 307 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 319, + 286, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 319, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 91, + 319, + 286, + 342 + ], + "type": "interline_equation", + "content": "z _ {p _ {i}} = \\frac {d _ {D} - \\sqrt {d _ {D} ^ {2} - (1 + \\tan^ {2} \\theta_ {p}) \\left(d _ {D} ^ {2} - r _ {D} ^ {2}\\right)}}{1 + \\tan^ {2} \\theta_ {p}}, \\tag {1}", + "image_path": "d6aeb9360cc8a54ef3c4011cb0cf28c6cd65649b49ab1a1916e85a579a6ed923.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 343, + 164, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 343, + 164, + 356 + ], + "spans": [ + { + "bbox": [ + 93, + 343, + 164, + 356 + ], + "type": "interline_equation", + "content": "r _ {p _ {i}} = z _ {p _ {i}} \\tan \\theta_ {p},", + "image_path": "33ecd1d106ea26e1687454ecc7eeeb89368c1b7f9450ba507adc6b1144e538c1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "inline_equation", + "content": "d_{D}" + }, + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "inline_equation", + "content": "r_D" + }, + { + "bbox": [ + 47, + 368, + 286, + 390 + ], + "type": "text", + "content": " are the distances to the monocentric lens and the lens radius." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "text", + "content": "The refraction angle " + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "text", + "content": " is then given by Snell's law with assuming the refractive indices of the air " + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\mu_{a}" + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "text", + "content": " and the lens " + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "inline_equation", + "content": "\\mu_{g}" + }, + { + "bbox": [ + 47, + 393, + 286, + 428 + ], + "type": "text", + "content": " are known:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 440, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 440, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 440, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {g} \\sin \\theta_ {g} = \\mu_ {a} \\sin \\left(\\theta_ {p} + \\theta_ {D i}\\right), \\\\ \\Leftrightarrow \\sin \\theta_ {g} = \\frac {\\mu_ {a}}{\\mu_ {g}} \\frac {d _ {D} r _ {p _ {i}}}{r _ {D} \\sqrt {r _ {p _ {i}} ^ {2} + z _ {p _ {i}} ^ {2}}}, \\tag {2} \\\\ \\end{array}", + "image_path": "c32195d9edb7ce258ac70d5dbeb2b49efa4822d73369d80f26690543a6eda029.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "type": "inline_equation", + "content": "\\theta_{D_i}" + }, + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "type": "text", + "content": " is the angle of " + }, + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "type": "inline_equation", + "content": "o_0 - o_D - p_i" + }, + { + "bbox": [ + 47, + 491, + 194, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": "The point of emergence " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{p}_o" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " is the intersection of the path " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\ell_g" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " and the sphere. It is also obtained as the mirror of the point " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " because of the symmetrical relationship between incident path " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\ell_i" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " and outgoing " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " w.r.t. the monocentric lens. As shown in Fig. 2, " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\ell_i" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " are in line symmetry to the line that is perpendicular to the vector " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{n}_p" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " through the point " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{o}_D" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": ". That is, given the point " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": ", its reflection " + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{p}_o" + }, + { + "bbox": [ + 47, + 504, + 286, + 598 + ], + "type": "text", + "content": " is described as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 82, + 611, + 286, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 611, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 82, + 611, + 286, + 641 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {p} _ {o} = H _ {p} \\boldsymbol {p} _ {i} + \\boldsymbol {t} _ {p}, \\tag {3} \\\\ \\Leftrightarrow \\boldsymbol {p} _ {o} = \\left(\\boldsymbol {I} - 2 \\boldsymbol {n} _ {p} \\boldsymbol {n} _ {p} ^ {\\top}\\right) \\boldsymbol {p} _ {i} + 2 (\\boldsymbol {o} _ {D} ^ {\\top} \\boldsymbol {n} _ {p}) \\boldsymbol {n} _ {p}, \\\\ \\end{array}", + "image_path": "e09e835da36f74d8429c7d3891838703e514de44201bf887704f06f3c3c8796d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "inline_equation", + "content": "H_{p}" + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "text", + "content": " is the Householder matrix and " + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{p}" + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "text", + "content": " denotes the center of reflection. Besides, the direction " + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{n}_p" + }, + { + "bbox": [ + 47, + 652, + 286, + 677 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 52, + 689, + 286, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 689, + 286, + 717 + ], + "spans": [ + { + "bbox": [ + 52, + 689, + 286, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {n} _ {p} = \\left[ \\begin{array}{l} \\cos \\left(\\pi / 2 + \\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\sin \\left(\\pi / 2 + \\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right] = \\left[ \\begin{array}{c} - \\sin \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\cos \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right], \\tag {4}", + "image_path": "a5c75b46c9c7b46dc5ac361864bb3c300a1eb4b565ecb65e1a01e4d47bd4f716.jpg" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 79, + 531, + 174 + ], + "blocks": [ + { + "bbox": [ + 317, + 79, + 531, + 174 + ], + "lines": [ + { + "bbox": [ + 317, + 79, + 531, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 79, + 531, + 174 + ], + "type": "image", + "image_path": "5db582d9d8e83d9dba075e94906f27ea39cb5edc2c385ae58ed0a8493cbdf543.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "text", + "content": "Figure 3. 3D-2D Projection. The black line " + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\ell_{o}" + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "text", + "content": " is the correct projection path that intersects with the optical axis at the camera center " + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "inline_equation", + "content": "o_0" + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "text", + "content": ". Our 3D-2D projection hypothesizes a line " + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "text", + "content": " as an initial guess of projection, and then optimizes it by verifying if the point and the angle of intersection with the optical axis is consistent with the ray-pixel mapping " + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\langle \\theta_q, f_q \\rangle" + }, + { + "bbox": [ + 305, + 189, + 545, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "type": "text", + "content": "therefore " + }, + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "type": "inline_equation", + "content": "H_{p}" + }, + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{p}" + }, + { + "bbox": [ + 306, + 268, + 452, + 281 + ], + "type": "text", + "content": " are rewritten to" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 289, + 545, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 289, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 323, + 289, + 545, + 343 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} H _ {p} = \\left[ \\begin{array}{l l} \\cos 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) & \\sin 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\\\ \\sin 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) & - \\cos 2 \\left(\\theta_ {D i} - \\theta_ {g}\\right) \\end{array} \\right], \\tag {5} \\\\ \\boldsymbol {t} _ {p} = \\left[ \\begin{array}{c} - d _ {D} \\sin 2 (\\theta_ {D i} - \\theta_ {g}) \\\\ d _ {D} \\cos 2 (\\theta_ {D i} - \\theta_ {g}) + d _ {D} \\end{array} \\right]. \\\\ \\end{array}", + "image_path": "2d88b322dfa5538b4c3c7827c8dbeb53a0d54317f841ed7ebe12183369fb1dc8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 351, + 532, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 532, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 532, + 363 + ], + "type": "text", + "content": "The direction " + }, + { + "bbox": [ + 306, + 351, + 532, + 363 + ], + "type": "inline_equation", + "content": "\\pmb{v}_o" + }, + { + "bbox": [ + 306, + 351, + 532, + 363 + ], + "type": "text", + "content": " is also given by Householder matrix as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 399, + 373, + 544, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 373, + 544, + 387 + ], + "spans": [ + { + "bbox": [ + 399, + 373, + 544, + 387 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {o} = H _ {p} \\boldsymbol {v} _ {i}. \\tag {6}", + "image_path": "4a8afee0c2a7a1185f9620e6e7d913e4d10f599af6fdf081c86ed02d0902b9ec.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "type": "text", + "content": "As a result, the intersection " + }, + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "type": "inline_equation", + "content": "\\pmb{o}_{f_q} = (0,f_q)^\\top" + }, + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "type": "text", + "content": " of the ray " + }, + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 306, + 395, + 544, + 420 + ], + "type": "text", + "content": " and the optical axis is given as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 430, + 544, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 430, + 544, + 443 + ], + "spans": [ + { + "bbox": [ + 339, + 430, + 544, + 443 + ], + "type": "interline_equation", + "content": "\\boldsymbol {o} _ {f _ {q}} = t _ {o} \\boldsymbol {v} _ {o} + \\boldsymbol {p} _ {o}, \\tag {7}", + "image_path": "269b17acb85dc32ea5436a35a33ba8a8d4e10938f0617fb3e94654dc5dfcb88c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 333, + 445, + 544, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 445, + 544, + 467 + ], + "spans": [ + { + "bbox": [ + 333, + 445, + 544, + 467 + ], + "type": "interline_equation", + "content": "\\Leftrightarrow t _ {o} = - \\frac {r _ {\\boldsymbol {p} _ {o}}}{r _ {\\boldsymbol {v} _ {o}}}, \\quad f _ {q} = z _ {\\boldsymbol {p} _ {o}} + \\left(\\frac {- r _ {\\boldsymbol {v} _ {o}}}{z _ {\\boldsymbol {v} _ {o}}}\\right) ^ {- 1} r _ {\\boldsymbol {p} _ {o}}. \\tag {8}", + "image_path": "1e4e42a26fd1e4b8860d3a0fc7ee08ae7ca09e187aa04016a290f8a8bb288105.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "text", + "content": "This equation describes the ray " + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\ell_{o}" + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "text", + "content": " backprojecting from the pixel " + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "inline_equation", + "content": "p_p" + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "text", + "content": " as a function of " + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\theta_{p}" + }, + { + "bbox": [ + 304, + 475, + 545, + 571 + ], + "type": "text", + "content": ". Inversely, the forward projection can be obtained for a given 3D point in front of the monocentric lens. The analytical 3D-2D projection, however, requires solving a 10th-degree equation [2]. The next section introduces a ray-pixel camera which exploits the axial symmetric structure of the rays to provide an efficient numerical 3D-2D projection." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 578, + 537, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 537, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 537, + 590 + ], + "type": "text", + "content": "3.1.1 Spherical Refraction Ray-Pixel Camera Model" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Up to this point, we observed that the rays emitted from a perspective camera through a monocentric lens show an axially symmetric structure around the line from the camera to the lens center. As illustrated in Fig. 2, given a pixel " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "p_p" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " by specifying " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\theta_p" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ", the corresponding ray " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " can be determined uniquely by Eq. (8). This section introduces our spherical refraction ray-pixel camera model by representing the ray-pixel mapping as follows." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Fig. 3 illustrates the light path from a pixel " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " on the plane-of-refraction. Suppose the emergent ray " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " intersects" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20815" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 121, + 267, + 164 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "text", + "content": "Table 1. The Pixel-Ray Mapping in Spherical Refraction Ray-Pixel Camera. A virtual pixel parameterized by " + }, + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "inline_equation", + "content": "\\theta_q" + }, + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "text", + "content": " is associated with a pixel-wise focal length " + }, + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "inline_equation", + "content": "f_q(\\theta_q) \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "text", + "content": ". The derivative " + }, + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{d}f_q(\\theta_q)}{\\mathrm{d}\\theta_q}" + }, + { + "bbox": [ + 46, + 70, + 287, + 118 + ], + "type": "text", + "content": " is also stored for our numerical 3D-2D projection." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 68, + 121, + 267, + 164 + ], + "lines": [ + { + "bbox": [ + 68, + 121, + 267, + 164 + ], + "spans": [ + { + "bbox": [ + 68, + 121, + 267, + 164 + ], + "type": "table", + "html": "
pixelθq0θq1...θqn
rayfq(θq0)fq(θq1)...fq(θqn)
dfq(θq)/dθqf′q(θq0)f′q(θq1)...f′q(θqn)
", + "image_path": "75a5ba2f4d5c9c7e6ea8a05cdc068a1b9b9be251dc72abc4ce1a1568bbe0c4f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": "with the optical axis at " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{f_q} = (0, f_q)^\\top" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " with angle " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\theta_q" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": ". Obviously, changing the position of " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_p" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " direction results in changing the corresponding " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\langle f_q, \\theta_q \\rangle" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " pair. That is, since the mapping between the pixel " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_p" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\ell_i" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": ", and the ray " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\ell_o" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\langle f_q, \\theta_q \\rangle" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " is bijective because of the reversibility of light, representing the light paths from the pixels in " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " space is identical to knowing all possible " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\langle f_q, \\theta_q \\rangle" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " pairs. In other words, the system of Fig. 3 as a whole can be seen as a pixel-wise varifocal camera which changes the focal length " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "f_q" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " for each virtual pixel parameterized by " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\theta_q" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": ". In fact, the mapping " + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\theta_q \\mapsto f_q" + }, + { + "bbox": [ + 46, + 189, + 287, + 332 + ], + "type": "text", + "content": " is a monotonic function due to the spherical structure of the lens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "text", + "content": "Table 1 shows our ray-pixel mapping. Due to the spherical structure of the lens, this is a discretization of the monotonic function " + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\theta_q \\mapsto f_q" + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\theta_q" + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "text", + "content": ". In practice, we sample " + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\theta_q" + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "text", + "content": " so that their interval results in a sub-pixel sampling in the original image domain. Notice that the derivative " + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{d}f_q(\\theta_q)}{\\mathrm{d}\\theta_q}" + }, + { + "bbox": [ + 46, + 335, + 287, + 422 + ], + "type": "text", + "content": " is also stored for our numerical 3D-2D projection as described below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 438, + 155, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 155, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 155, + 451 + ], + "type": "text", + "content": "3.1.2 3D-2D Projection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "text", + "content": "Instead of the analytical forward projection which requires solving a 10th-degree equation [2], this section introduces a numerical forward projection using our ray-pixel mapping table (Tab. 1). The key idea of our numerical forward projection is to hypothesize a projection line " + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "text", + "content": " in Fig. 3 first, and check if it intersects with the optical axis at the identical virtual focal length stored in the ray-pixel mapping table (Tab. 1). If the focal lengths are not identical, then " + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "text", + "content": ", i.e., the virtual pixel " + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 46, + 456, + 287, + 589 + ], + "type": "text", + "content": " equivalently, is refined to minimize the difference. We can formulate this process as a Gauss-Newton optimization as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": "Gauss-Newton Method As shown in Fig. 3, let us consider the ray " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " from a 3D point to be projected. By hypothesizing that " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " intersects the optical axis at an angle " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": ", i.e., by hypothesizing that the 3D point is projected to a virtual pixel " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": ", we can compute the intersection " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\Lambda_{q}(\\theta_{q}) = (0, \\lambda_{q}(\\theta_{q}))^{\\top}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " and the optical axis. If " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\lambda_{q}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " is equal to the virtual focal length " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "f_{q}(\\theta_{q})" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " stored in the ray-pixel mapping, the ray " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\ell_{\\lambda}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": " is identical to " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "\\ell_{o}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": ", and hence that can intersect with the optical axis at the camera center " + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "inline_equation", + "content": "o_{0}" + }, + { + "bbox": [ + 46, + 591, + 287, + 699 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "That is, the numerical 3D-2D projection can be achieved" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 417, + 175 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 417, + 175 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 417, + 175 + ], + "type": "image", + "image_path": "a36862c957beaaf26ce9a13cc4815d931ac3952aa8a80fae46246aedf5bd2b61.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": "Figure 4. Kaleidoscopic Imaging. (a) Mirrored Cameras. (b) Chambers in the real camera image. The label " + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "inline_equation", + "content": "\\pmb{u}^{(i)}" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": " denote the reflections of " + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": " by the " + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 184, + 545, + 218 + ], + "type": "text", + "content": "th mirror respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 76, + 533, + 175 + ], + "blocks": [ + { + "bbox": [ + 425, + 76, + 533, + 175 + ], + "lines": [ + { + "bbox": [ + 425, + 76, + 533, + 175 + ], + "spans": [ + { + "bbox": [ + 425, + 76, + 533, + 175 + ], + "type": "image", + "image_path": "47819e3d5136a0a930b81dc57518d1c877fd451ce490bdfe93f36a0ff7d640c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 232, + 462, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 462, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 462, + 243 + ], + "type": "text", + "content": "by solving the following optimization:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 253, + 545, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 253, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 545, + 275 + ], + "type": "interline_equation", + "content": "\\theta_ {q} = \\underset {\\theta_ {q}} {\\operatorname {a r g m i n}} G \\left(\\theta_ {q}\\right) = \\underset {\\theta_ {q}} {\\operatorname {a r g m i n}} \\left(f _ {q} \\left(\\theta_ {q}\\right) - \\lambda_ {q} \\left(\\theta_ {q}\\right)\\right). \\tag {9}", + "image_path": "13255d34f3fc966f40a9dfe9ea4ff16f94cd9fde9cb74360b6fcca2c9f978053.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "type": "inline_equation", + "content": "G(\\theta_q)" + }, + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "type": "text", + "content": " is a monotonic function and hence we can refine " + }, + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\theta_q" + }, + { + "bbox": [ + 305, + 285, + 545, + 309 + ], + "type": "text", + "content": " iteratively by Gauss-Newton method as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 318, + 544, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 318, + 544, + 360 + ], + "spans": [ + { + "bbox": [ + 354, + 318, + 544, + 360 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta_ {q} ^ {(k + 1)} = \\theta_ {q} ^ {(k)} - G \\frac {\\Delta \\theta_ {q}}{\\Delta G} \\\\ = \\theta_ {q} ^ {(k)} - \\frac {f _ {q} \\left(\\theta_ {q} ^ {(k)}\\right) - \\lambda_ {q} \\left(\\theta_ {q} ^ {(k)}\\right)}{f _ {q} ^ {\\prime} \\left(\\theta_ {q} ^ {(k)}\\right) - \\lambda_ {q} ^ {\\prime} \\left(\\theta_ {q} ^ {(k)}\\right)}, \\tag {10} \\\\ \\end{array}", + "image_path": "264b551948bd1715d7d791fa5effd240f6202e919349f2e75ce0865b6a3716b1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "inline_equation", + "content": "\\theta_q^{(k)}" + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "inline_equation", + "content": "\\theta_{q}" + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 370, + 479, + 385 + ], + "type": "text", + "content": "-th iteration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 384, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 421 + ], + "type": "text", + "content": "Therefore, if we compute the derivative " + }, + { + "bbox": [ + 305, + 384, + 545, + 421 + ], + "type": "inline_equation", + "content": "f_{q}^{\\prime}" + }, + { + "bbox": [ + 305, + 384, + 545, + 421 + ], + "type": "text", + "content": " beforehand as shown in the third row of Tab. 1, then this 3D-2D projection can be computed efficiently." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 430, + 413, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 413, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 413, + 441 + ], + "type": "text", + "content": "3.2. Multifacet Mirror" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "As is well known, observing the scene via a multi-facet mirror or a kaleidoscope is identical to observing the scene by virtual multi-view cameras, and in particular, kaleidoscopes with three mirrors are known to be reasonable in terms of less overlaps of mirrored images called discontinuities [29,33]. In our teleidoscopic imaging system, we use a three-facet mirror which defines reflections of the spherical refraction ray-pixel camera introduced in Sec. 3.1.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "text", + "content": "The reflection " + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{\\prime}" + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "text", + "content": " of a 3D point " + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "text", + "content": " by a mirror of normal " + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "text", + "content": " and distance " + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 306, + 544, + 545, + 568 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 354, + 578, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 578, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 354, + 578, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {p} ^ {\\prime} = H _ {i} \\boldsymbol {p} + \\boldsymbol {t} _ {i} \\\\ \\Leftrightarrow \\boldsymbol {p} ^ {\\prime} = \\left(\\boldsymbol {I} - 2 \\boldsymbol {n} _ {i} \\boldsymbol {n} _ {i} ^ {\\top}\\right) \\boldsymbol {p} + 2 d _ {i} \\boldsymbol {n} _ {i}, \\tag {11} \\\\ \\end{array}", + "image_path": "ad9e51106e14b2b85d558dea6a800f73fb910e45e3a0eb168f4eaa221d82ca93.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 617, + 507, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 507, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 507, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 617, + 507, + 628 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 306, + 617, + 507, + 628 + ], + "type": "text", + "content": " the Householder transformation matrix." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": "In the case of kaleidoscopic imaging, the mirrors generate bouncing reflections as shown in Fig. 4. The reflection of " + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{\\prime}" + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": " by another mirror of normal " + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{n}_j" + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": " and distance " + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": " is given simply by" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 350, + 687, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 687, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 687, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {p} ^ {\\prime \\prime} = H _ {i j} \\boldsymbol {p} + \\boldsymbol {t} _ {i j} \\\\ = H _ {j} H _ {i} \\boldsymbol {p} + 2 d _ {j} \\boldsymbol {n} _ {j} + 2 H _ {j} d _ {i} \\boldsymbol {n} _ {i}. \\tag {12} \\\\ \\end{array}", + "image_path": "8cae350eed4f74aedea09a7b67fd80a01b49b76b210efd59b305012905d95402.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20816" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 278, + 144 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 278, + 144 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 278, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 278, + 144 + ], + "type": "image", + "image_path": "37b04a7540fbc4886328706ad8ad31dcbd2b10f2182fda818b2beb54037eac76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": "Figure 5. Depth-of-Field with Monocentric Lens. " + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "inline_equation", + "content": "D_{F}" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "inline_equation", + "content": "D_{N}" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": " denote the depth-of-field of the camera itself without the monocentric lens. The effective depth-of-field of the entire system " + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "inline_equation", + "content": "L_{F}" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "inline_equation", + "content": "L_{N}" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": " can be obtained by projecting " + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "inline_equation", + "content": "D_{N}" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "inline_equation", + "content": "D_{F}" + }, + { + "bbox": [ + 46, + 156, + 288, + 212 + ], + "type": "text", + "content": " through the monocentric lens (the red and blue dashed lines)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 297 + ], + "type": "text", + "content": "As a result, our teleidoscopic imaging system can be modeled as multi-view spherical refraction ray-pixel cameras. They share the single monocentric lens, and the pose of each virtual cameras can be computed by Eq. (12) if the mirror parameters are given. The later section describes our calibration algorithm to estimate such parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 305, + 288, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 288, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 288, + 333 + ], + "type": "text", + "content": "4. Depth-of-Field of Teleidoscopic Imaging System" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 376 + ], + "type": "text", + "content": "This section describes analytical evaluations on the depth-of-field of our teleidoscopic imaging system. We first review the depth-of-field of monocentric lens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 382, + 250, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 250, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 250, + 395 + ], + "type": "text", + "content": "4.1. Depth-of-Field with Monocentric Lens" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "content": "As shown in Fig. 5, depth-of-field with a monocentric lens can be simply defined as an extension of the path of the thin lens camera (cf. supplemental material). This is because the points in the near and the far depth-of-field " + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "inline_equation", + "content": "D_N" + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "inline_equation", + "content": "D_F" + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "content": " of thin lens are projected within the permissible circle-of-confusion even though the monocentric lens itself introduces the spherical aberration. Hence, we define the depth-of-field with a monocentric lens " + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "inline_equation", + "content": "L_N" + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "inline_equation", + "content": "L_F" + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "content": " as the intersection of the backprojection path through the edge of aperture (the red and blue dash lines in Fig. 5) and the backprojection path through the lens center " + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "inline_equation", + "content": "o_0" + }, + { + "bbox": [ + 46, + 401, + 287, + 532 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "content": "Because of the symmetric structure of the monocentric lens, the depth-of-field " + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "inline_equation", + "content": "L_{N} + L_{F}" + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "content": " can be deeper than the original depth-of-field " + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "inline_equation", + "content": "D_{N} + D_{F}" + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "content": " if the subject distance of the monocentric lens " + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "inline_equation", + "content": "s_{D} - r_{D}" + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "content": " is larger than its focal distance " + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "inline_equation", + "content": "t_{D} - r_{D}" + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "content": ". That is, there is a trade-off between magnification and depth-of-field even in a monocentric lens. The reason for the deep depth-of-field is due to the significant contribution of the monocentric lens to the far depth of field " + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "inline_equation", + "content": "L_{F}" + }, + { + "bbox": [ + 46, + 533, + 288, + 653 + ], + "type": "text", + "content": ", and thus imaging with less blurring even when designed for high magnification (cf. Sec. 6.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "content": "4.2. Depth-of-Field of Teleidoscopic Imaging System" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Fig. 6 illustrates the near and the far depth-of-fields of a virtual camera " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " defined as a mirror of the original camera." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 73, + 518, + 211 + ], + "blocks": [ + { + "bbox": [ + 323, + 73, + 518, + 211 + ], + "lines": [ + { + "bbox": [ + 323, + 73, + 518, + 211 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 518, + 211 + ], + "type": "image", + "image_path": "a18d5b47c6ffef7520bc7ce301c581cdd35cb2819de0d3f9bdb07ad734e28a10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "text", + "content": "Figure 6. Depth of Field of Teleidoscopic Imaging System. " + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "inline_equation", + "content": "D_N" + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "inline_equation", + "content": "D_F" + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "text", + "content": " denote the near and far depth-of-field of the thin lens camera " + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "text", + "content": " respectively. All the points once imaged within " + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "inline_equation", + "content": "D_N - D_F" + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "text", + "content": " should be originally emitted from the points within " + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "inline_equation", + "content": "L_N - L_F" + }, + { + "bbox": [ + 304, + 223, + 545, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 281, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 306 + ], + "type": "text", + "content": "Due to the aperture, the near and the far depth-of-fields form a curve respectively (the red and the blue dashed lines)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 306, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 365 + ], + "type": "text", + "content": "Since the teleidoscopic imaging system has multiple virtual cameras as shown in Fig. 4, each of them has a different depth-of-field according to their relative pose to the monocentric lens. That is, the intersections of such depth-of-fields can be used for multi-view stereo reconstruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 378, + 545, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 545, + 405 + ], + "type": "text", + "content": "5. Calibration of Teleidoscopic Imaging System" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "content": "This section introduces our calibration algorithm of teleidoscopic imaging system which requires capturing a single reference planar patterns. Fig. 7 shows the measurement model where the real camera " + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "content": " observes reference points " + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{p}_w" + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "content": " such as chessboard corners on a reference board II via a monocentric lens and three mirrors. Our calibration estimates the mirror normals " + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "content": ", their distances " + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "content": " from the camera, and the position and the radius of the monocentric lens, with assuming that the 2D positions of the reference points " + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{p}_w" + }, + { + "bbox": [ + 304, + 413, + 545, + 555 + ], + "type": "text", + "content": " on the reference plane II, the intrinsic parameters of the camera, and the refraction index of the lens are given beforehand." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 689 + ], + "type": "text", + "content": "A challenge in this calibration is the fact that the mirrors require observing 3D points and their reflections to estimate their poses [33], while the observation in the teleidoscopic system does not include such mirrored points that follow Eq. (12) due to the refraction by the monocentric lens. Similarly, the rays reflected and then refracted through the projection in teleidoscopic imaging system do not satisfy the coplanarity constraint. In addition, as mentioned in [2], the estimation of the monocentric lens parameters requires multiple viewpoints or multiple monocentric lenses. These points result in a chicken-and-egg problem." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "To solve this problem, we utilize the fact that the positions of the centers of the monocentric lens and its mirrors" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20817" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 75, + 265, + 202 + ], + "blocks": [ + { + "bbox": [ + 69, + 75, + 265, + 202 + ], + "lines": [ + { + "bbox": [ + 69, + 75, + 265, + 202 + ], + "spans": [ + { + "bbox": [ + 69, + 75, + 265, + 202 + ], + "type": "image", + "image_path": "7c5db9c3e0095bcf17f4d780f894beecc7c9bf8ade56c6513dde23420475e205.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 288, + 255 + ], + "type": "text", + "content": "Figure 7. Calibration of Teleidoscopic Imaging System. Cyan lines denote the light paths for a mirrored camera. The calibration can be conducted by capturing reference points on a plane of a known geometry through the monocentric lens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 301 + ], + "type": "text", + "content": "are always captured without refraction by definition, and satisfy Eq. (12). Therefore, we start by estimating the center of the monocentric lens from the captured reference points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": "In what follows, " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p^{(0)}" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": " denotes the real image of a reference point " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{p}_w" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": " in the camera " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": ". Similarly " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p^{(i)}" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "i = \\{1,2,3\\}" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": ") denotes the image of its first reflection by the mirror " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p^{(ij)}" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "ij = \\{12,13,21,23,31,32\\}" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": ") denotes its second reflection by the mirrors " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 302, + 287, + 367 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": "Axis of Monocentric Lens The axis to the center of monocentric lens " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{a}_v^{(0)}" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": " from the camera " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": " can be estimated by " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{p}_w - \\pmb{p}_P^{(0)}" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": " correspondences. Similarly to [2], we estimate " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{a}_v^{(0)}" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": " by the coplanarity constraint of the three vectors " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{a}_v^{(0)}" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p^{(0)}" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{p}_w" + }, + { + "bbox": [ + 47, + 373, + 288, + 437 + ], + "type": "text", + "content": " on a single plane-of-refraction:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 443, + 287, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 443, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 91, + 443, + 287, + 481 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {p} _ {p} ^ {(0) \\top} \\left(\\boldsymbol {a} _ {v} ^ {(0)} \\times \\left(R _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {t} _ {\\Pi}\\right)\\right) = 0, \\tag {13} \\\\ \\Leftrightarrow \\boldsymbol {p} _ {p} ^ {(0) \\top} \\left(E _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {s} _ {\\Pi}\\right)) = 0, \\\\ \\end{array}", + "image_path": "db76aa1f277151522015512eaf7213d7f6f89dc5a8e82ca9a847c98517558b9b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "R_{\\Pi}, t_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": " are the pose of the reference plane in the camera coordinate system and " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "E_{\\Pi} = \\boldsymbol{a}_v^{(0)} \\times R_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "s_{\\Pi} = \\boldsymbol{a}_v^{(0)} \\times t_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": ". Since this is a linear equation with 9 unknown parameters of " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "E_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "s_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": ", we can obtain " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "E_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "s_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": " up to scale by observing at least 8 points on the reference plane. Using the estimated " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "E_{\\Pi}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": ", the axis " + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\boldsymbol{a}_v^{(0)}" + }, + { + "bbox": [ + 46, + 487, + 287, + 563 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 569, + 287, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 569, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 112, + 569, + 287, + 588 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {v} ^ {(0)} = \\frac {E _ {\\Pi} (: , 1) \\times E _ {\\Pi} (: , 2)}{| | E _ {\\Pi} (: , 1) \\times E _ {\\Pi} (: , 2) | |}. \\tag {14}", + "image_path": "12fa19ee7b0bdef6a091891f676397a2b0a517de66154e60a060466632f6cf1a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "type": "text", + "content": "Similarly, the axis to the center of the mirrored monocentric lens " + }, + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{a}_v^{(i)}" + }, + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "type": "text", + "content": " can be obtained by the mirrored points " + }, + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\pmb{p}_p^{(i)}" + }, + { + "bbox": [ + 46, + 593, + 287, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "type": "text", + "content": "Mirror Normals The axis to the center of the monocentric lens " + }, + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{v}^{(0)} = (x_{a_{v}}^{(0)},y_{a_{v}}^{(0)},z_{a_{v}}^{(0)})^{\\top}" + }, + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "type": "text", + "content": " and its mirror " + }, + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{v}^{(i)} = (x_{a_{v}}^{(i)},y_{a_{v}}^{(i)},z_{a_{v}}^{(i)})^{\\top}" + }, + { + "bbox": [ + 47, + 624, + 287, + 662 + ], + "type": "text", + "content": " satisfies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 668, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 668, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 123, + 668, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {v} ^ {(0) \\top} [ \\boldsymbol {n} _ {i} ] _ {\\times} \\boldsymbol {a} _ {v} ^ {(i)} = 0, \\tag {15}", + "image_path": "e0e40b8529840a9c5c82095247f594861ec250c70bff585fe40b2adbea4445bb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "[\\pmb{n}_i]_{\\times}" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " denotes the skew-symmetric matrix defined by the normal " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i = (x_{n_i},y_{n_i},z_{n_i})^\\top" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " of the mirror " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "The same constraint holds for each of the first-second reflection pairs " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\pmb{a}_v^{(i)} - \\pmb{a}_v^{(ij)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " about the same mirror normal " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " [33]. Therefore " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\pmb{n}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " can be obtained linearly only from the axes to the centers of the monocentric lenses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 125, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 125, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 545, + 174 + ], + "type": "text", + "content": "Mirror Distances Once the mirror normals are estimated, we can utilize the kaleidoscopic triangulation [33] to obtain linear constraints on the mirror distances " + }, + { + "bbox": [ + 305, + 125, + 545, + 174 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 305, + 125, + 545, + 174 + ], + "type": "text", + "content": " (See supp. Sec. 2 for details)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 180, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 204 + ], + "type": "text", + "content": "Pose of Reference Plane Similarly to Eq. (13), the plane-of-refraction constraint holds for the mirrored cameras " + }, + { + "bbox": [ + 305, + 180, + 545, + 204 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 305, + 180, + 545, + 204 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 221, + 545, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 221, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 321, + 221, + 545, + 242 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {p} _ {p} ^ {(i)}\\right) ^ {\\top} \\left(\\boldsymbol {a} _ {v} ^ {(i)} \\times \\left(H _ {i} \\left(R _ {\\Pi} \\boldsymbol {p} _ {w} + \\boldsymbol {t} _ {\\Pi}\\right) + \\boldsymbol {t} _ {i}\\right)\\right) = 0. \\tag {16}", + "image_path": "ad4595552bf8069be51ab8b3931d7af795025bf9014b1301625d3cc52d4fed5c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "content": "This constraint allows us estimating the pose of the reference plane " + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "inline_equation", + "content": "R_{\\Pi}" + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{\\Pi}" + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "content": " linearly." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "text", + "content": "Monocentric Lens Parameters The calibration algorithm up to this point does not require the monocentric lens parameters " + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "inline_equation", + "content": "d_{D}, r_{D}" + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mu_{g}" + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "text", + "content": ". We estimate these parameters by the coplanarity constraint of the ray through " + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{p}^{(i)}" + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 305, + 278, + 545, + 329 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 373, + 338, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 338, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 373, + 338, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {o} ^ {(i)} \\times \\left(\\boldsymbol {p} _ {w} ^ {(i)} - \\boldsymbol {p} _ {o} ^ {(i)}\\right) = 0. \\tag {17}", + "image_path": "1c32c09eb58cba870bf72665fd25f838f409c6a2e7ab3542a5dade85cf2e7894.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 362, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 545, + 410 + ], + "type": "text", + "content": "This is a nonlinear constraint for the monocentric lens parameters as described in Sec. 3.1 and we solve this as a nonlinear optimization problem with assuming their rough estimates are available in practice." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "content": "**Bundle Adjustment** The last step of our calibration is to refine the parameters " + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\boldsymbol{a}_v^{(0)}, d_D, r_D, \\mu_g, \\boldsymbol{n}_i, d_i, R_{\\Pi}, t_{\\Pi}" + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "inline_equation", + "content": "i = 1, 2, 3" + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "content": ") by minimizing the reprojection errors of the reference points " + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}_w" + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "content": " as a nonlinear optimization problem. Here the forward projection is computed using the method described in Sec. 3.1.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 490, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 490, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 490, + 545, + 537 + ], + "type": "text", + "content": "Once the teleidoscopic imaging system parameters are calibrated, a scene point can be linearly triangulated in a DLT manner [13] from its teleidoscopic projections (See supp. Sec. 3 for details)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 548, + 376, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 376, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 376, + 560 + ], + "type": "text", + "content": "6. Evaluation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 568, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 568, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 568, + 545, + 628 + ], + "type": "text", + "content": "This section provides quantitative evaluations of the proposed numerical 3D-2D projections, the robustness of the calibration, and the depth of field of our teleidoscopic imaging system. We also demonstrate the effectiveness of our method by reconstructing real-world microscale objects." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 635, + 544, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 544, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 544, + 647 + ], + "type": "text", + "content": "6.1. Computational Efficiency of 3D-2D Projection" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "The average computation cost of a single forward projection by our method and by the analytical solution solving a 10th-degree equation for each projection [2] are " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.62\\mathrm{ms}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.93\\mathrm{ms}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " respectively. These runtime costs are the average values of 100 trials of 10K points forward projections" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20818" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 167, + 152 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 167, + 152 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 167, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 167, + 152 + ], + "type": "image", + "image_path": "03031ee35174fe2150d6e749dfa273a836c69c3dcdc6d5503a054d2751f0ec68.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 157, + 104, + 167 + ], + "lines": [ + { + "bbox": [ + 93, + 157, + 104, + 167 + ], + "spans": [ + { + "bbox": [ + 93, + 157, + 104, + 167 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 72, + 279, + 155 + ], + "blocks": [ + { + "bbox": [ + 172, + 72, + 279, + 155 + ], + "lines": [ + { + "bbox": [ + 172, + 72, + 279, + 155 + ], + "spans": [ + { + "bbox": [ + 172, + 72, + 279, + 155 + ], + "type": "image", + "image_path": "de6b2965b32579d49160ab7c99ead40090053d152a4407775fd497c3b08e5d06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 157, + 239, + 166 + ], + "lines": [ + { + "bbox": [ + 229, + 157, + 239, + 166 + ], + "spans": [ + { + "bbox": [ + 229, + 157, + 239, + 166 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 227, + 123, + 312 + ], + "blocks": [ + { + "bbox": [ + 50, + 227, + 123, + 312 + ], + "lines": [ + { + "bbox": [ + 50, + 227, + 123, + 312 + ], + "spans": [ + { + "bbox": [ + 50, + 227, + 123, + 312 + ], + "type": "image", + "image_path": "5b7d6ec1ded5a65e174362d74ed5a867250a9f86437daa10a7e95ee4602f41a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 126, + 226, + 207, + 313 + ], + "blocks": [ + { + "bbox": [ + 46, + 170, + 287, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 213 + ], + "type": "text", + "content": "Figure 8. Teleidoscopic System for Quantitative Evaluation. (a) Simulated measurement environment. (b) Reprojection errors at different noise levels. The bars denote the standard deviation of the errors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 226, + 207, + 313 + ], + "lines": [ + { + "bbox": [ + 126, + 226, + 207, + 313 + ], + "spans": [ + { + "bbox": [ + 126, + 226, + 207, + 313 + ], + "type": "image", + "image_path": "93b89c7865b2d4e25fe093997d934f490fdc280288507f4a9faf76efc13829f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 209, + 226, + 283, + 312 + ], + "blocks": [ + { + "bbox": [ + 209, + 226, + 283, + 312 + ], + "lines": [ + { + "bbox": [ + 209, + 226, + 283, + 312 + ], + "spans": [ + { + "bbox": [ + 209, + 226, + 283, + 312 + ], + "type": "image", + "image_path": "804146f7ae0a0ab5d7205f8da0c8f32e9aa91c41ea8ebb537a1c9d6ed263fdca.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 299, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 365 + ], + "type": "text", + "content": "Figure 11. Changes in Depth-of-Field by the Monocentric Lens. (a), (b) and (c) show the total, near, and far depth-of-field with a monocentric lens. The red and blue plots indicate the results with and without the monocentric lens respectively. The green lines indicate the subject distance where the subject and the focal distances of the monocentric lens itself are same." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 314, + 123, + 399 + ], + "blocks": [ + { + "bbox": [ + 50, + 314, + 123, + 399 + ], + "lines": [ + { + "bbox": [ + 50, + 314, + 123, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 314, + 123, + 399 + ], + "type": "image", + "image_path": "fbe50aa59c1daf80a7c3d094de19608c47dab964072df6bb9eda6568d254e337.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 401, + 96, + 410 + ], + "lines": [ + { + "bbox": [ + 86, + 401, + 96, + 410 + ], + "spans": [ + { + "bbox": [ + 86, + 401, + 96, + 410 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 414, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 491 + ], + "type": "text", + "content": "Figure 9. Calibration Errors. From left to right, the upper figures show the (a) angular error of the axis to the monocentric lens center (degree), (b) angular error of the mirror normal (degree), and (c) rotation error (Riemannian distance) of reference plane. The lower ones are the distance errors normalized by the ground truth. Here the distance of the monocentric lens is only estimated in a nonlinear manner (Sec. 5)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 124, + 317, + 207, + 399 + ], + "blocks": [ + { + "bbox": [ + 124, + 317, + 207, + 399 + ], + "lines": [ + { + "bbox": [ + 124, + 317, + 207, + 399 + ], + "spans": [ + { + "bbox": [ + 124, + 317, + 207, + 399 + ], + "type": "image", + "image_path": "37b0a7a601c404758d8d79bdf8e2b25cadb137d39fe8be824a9e41cae583508d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 401, + 180, + 411 + ], + "lines": [ + { + "bbox": [ + 169, + 401, + 180, + 411 + ], + "spans": [ + { + "bbox": [ + 169, + 401, + 180, + 411 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 209, + 314, + 282, + 399 + ], + "blocks": [ + { + "bbox": [ + 209, + 314, + 282, + 399 + ], + "lines": [ + { + "bbox": [ + 209, + 314, + 282, + 399 + ], + "spans": [ + { + "bbox": [ + 209, + 314, + 282, + 399 + ], + "type": "image", + "image_path": "23d0b9f4d92190129dcd0363d47aa707486abb095172c74d10ad44e25a2d06d8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 401, + 259, + 411 + ], + "lines": [ + { + "bbox": [ + 249, + 401, + 259, + 411 + ], + "spans": [ + { + "bbox": [ + 249, + 401, + 259, + 411 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 501, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 538 + ], + "type": "text", + "content": "run in Matlab on an Intel Core-i7 2.6GHz PC. This result clearly verifies that our numerical forward projection based on our ray-pixel model can realize an efficient projection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 546, + 124, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 124, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 124, + 558 + ], + "type": "text", + "content": "6.2. Calibration" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 565, + 284, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 284, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 284, + 577 + ], + "type": "text", + "content": "6.2.1 Quantitative Evaluation using Synthesized Data" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "Fig. 8(a) shows the measurement environment which simulates the real capture system used in Sec. 6.2.2. The system has a kaleidoscope with three " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "10 \\times 30\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " mirrors in front of the camera " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ". The mirrors are at slightly off-perpendicular angle of " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "1.4^{\\circ}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " to the camera image plane so that the mirrors define virtual cameras around the target with less overlaps of the mirrored images. The system also has a monocentric lens of " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "10\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " diameter in front of the mirrors, at " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "40\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " distance from the camera. The refraction index " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{g}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " is set to 2.0. The system captures 48 reference points (blue dots in Fig. 8(a)) to calibrate its parameters. By injecting Gaussian" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 429, + 150 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 429, + 150 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 429, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 429, + 150 + ], + "type": "image", + "image_path": "9859d492b60fb7b88a47003e31d41c09930e0848df079b3586fc222e6d53d8ed.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 151, + 374, + 159 + ], + "lines": [ + { + "bbox": [ + 364, + 151, + 374, + 159 + ], + "spans": [ + { + "bbox": [ + 364, + 151, + 374, + 159 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 432, + 72, + 537, + 150 + ], + "blocks": [ + { + "bbox": [ + 432, + 72, + 537, + 150 + ], + "lines": [ + { + "bbox": [ + 432, + 72, + 537, + 150 + ], + "spans": [ + { + "bbox": [ + 432, + 72, + 537, + 150 + ], + "type": "image", + "image_path": "7cc7d065929938b3447e28105a16ae5bc9b9424c6118b50f63dad7e3555ea0e4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 151, + 491, + 159 + ], + "lines": [ + { + "bbox": [ + 481, + 151, + 491, + 159 + ], + "spans": [ + { + "bbox": [ + 481, + 151, + 491, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 306, + 220, + 387, + 285 + ], + "blocks": [ + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "type": "text", + "content": "Figure 10. Experimental Setup and Calibration. (a) Our teleidoscopic imaging system consists of a camera, a projector, three mirrors " + }, + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "(10 \\times 30\\mathrm{mm})" + }, + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "type": "text", + "content": ", and a monocentric lens (5mm radius, refraction index " + }, + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "type": "inline_equation", + "content": "\\mu_g = 2.0" + }, + { + "bbox": [ + 304, + 163, + 545, + 208 + ], + "type": "text", + "content": "). (b) Input image for calibration." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 220, + 387, + 285 + ], + "lines": [ + { + "bbox": [ + 306, + 220, + 387, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 387, + 285 + ], + "type": "image", + "image_path": "ce54496830517d8b910b8e59a4d940185f48f7a08350852b88d66a965e05be5b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 287, + 359, + 296 + ], + "lines": [ + { + "bbox": [ + 349, + 287, + 359, + 296 + ], + "spans": [ + { + "bbox": [ + 349, + 287, + 359, + 296 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 389, + 220, + 467, + 285 + ], + "blocks": [ + { + "bbox": [ + 389, + 220, + 467, + 285 + ], + "lines": [ + { + "bbox": [ + 389, + 220, + 467, + 285 + ], + "spans": [ + { + "bbox": [ + 389, + 220, + 467, + 285 + ], + "type": "image", + "image_path": "3546a9cce26579489d9b7920f9b824b83ab86a45d15f57a15cf7575ef81b8588.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 286, + 440, + 295 + ], + "lines": [ + { + "bbox": [ + 429, + 286, + 440, + 295 + ], + "spans": [ + { + "bbox": [ + 429, + 286, + 440, + 295 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 466, + 220, + 546, + 285 + ], + "blocks": [ + { + "bbox": [ + 466, + 220, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 466, + 220, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 466, + 220, + 546, + 285 + ], + "type": "image", + "image_path": "2016e58bc7025fcec3e565772d928baaa8617c757bdcaca3c43efb5256c6c8b0.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 286, + 517, + 295 + ], + "lines": [ + { + "bbox": [ + 506, + 286, + 517, + 295 + ], + "spans": [ + { + "bbox": [ + 506, + 286, + 517, + 295 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": "noise of different standard deviations " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": " to the 2D positions of their projections, we evaluate the robustness of our calibration procedure. Fig. 8(b) shows average reprojection errors in pixel of 100 trials at each pixel noise level " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": ". We can observe that the reprojection errors increase linearly against the pixel noise level." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 449, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 507 + ], + "type": "text", + "content": "Fig. 9 shows the estimation errors of the monocentric lens parameters, the mirror parameters, and the reference plane parameters respectively. These results indicate that our calibration algorithm performs reasonably under realistic observation noise." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 516, + 511, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 511, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 511, + 528 + ], + "type": "text", + "content": "6.2.2 Quantitative Evaluation using Real Data" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "Fig. 10(a) shows an overview of our teleidoscopic imaging system. The system consists of a FLIR Blackfly BFS-U3-89S6C-C camera (4096×2160 resolution, pixel size 3.45μm) with a f=4mm / F8 lens, three 10 × 30mm mirrors, a monocentric lens of 10mm diameter whose refraction index is " + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mu_g = 2.0" + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": ", and a LED projector ASUS E1-R for casting Gray code patterns onto the target surface to identify corresponding points [12, 23]. Note that the effective resolution of each \"virtual camera\" image is approximately 300×200 (16 \"virtual cameras\" in total), which is calculated from the area inside the base chamber with one horizontal edge of the triangle and inside the spherical lens." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We used a single image (Fig. 10(b)) capturing " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "5 \\times 8" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " corners of a checkerboard for calibration, and used up to second reflections in the image. The average reprojection er" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20819" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 84, + 113, + 115 + ], + "blocks": [ + { + "bbox": [ + 56, + 84, + 113, + 115 + ], + "lines": [ + { + "bbox": [ + 56, + 84, + 113, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 84, + 113, + 115 + ], + "type": "image", + "image_path": "2e84516d714b8c73c2142fd0f160d7400f800b813363becc3af92ef5974be9c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 118, + 89, + 128 + ], + "lines": [ + { + "bbox": [ + 78, + 118, + 89, + 128 + ], + "spans": [ + { + "bbox": [ + 78, + 118, + 89, + 128 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 85, + 174, + 117 + ], + "blocks": [ + { + "bbox": [ + 120, + 85, + 174, + 117 + ], + "lines": [ + { + "bbox": [ + 120, + 85, + 174, + 117 + ], + "spans": [ + { + "bbox": [ + 120, + 85, + 174, + 117 + ], + "type": "image", + "image_path": "29442c3ae27b96808aed4d56cdb952e1993fe90ac6d6daeecb0af5381f12ef1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 119, + 153, + 128 + ], + "lines": [ + { + "bbox": [ + 141, + 119, + 153, + 128 + ], + "spans": [ + { + "bbox": [ + 141, + 119, + 153, + 128 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "text", + "content": "Figure 12. Evaluation with a sphere (radius: " + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "inline_equation", + "content": "3\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "text", + "content": "). (a) Ground truth. (b) Recovered 3D shape. The colors visualize the " + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "text", + "content": "-coordinate of the reconstructed shape. (c) Error map. " + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "inline_equation", + "content": "E_{d}" + }, + { + "bbox": [ + 46, + 132, + 287, + 175 + ], + "type": "text", + "content": " is the mean absolute error in millimeter." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 72, + 206, + 117 + ], + "blocks": [ + { + "bbox": [ + 175, + 72, + 206, + 117 + ], + "lines": [ + { + "bbox": [ + 175, + 72, + 206, + 117 + ], + "spans": [ + { + "bbox": [ + 175, + 72, + 206, + 117 + ], + "type": "image", + "image_path": "86fae6deb7684254fb2f0ca1c2117761dbe701f5eacedc374726bc50e95af33f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 256, + 117 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 256, + 117 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 256, + 117 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 256, + 117 + ], + "type": "image", + "image_path": "a41e30a812ea681b3cd2731f5197c9da69cb242e354fa8752fbeb895161e04c6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 119, + 235, + 128 + ], + "lines": [ + { + "bbox": [ + 224, + 119, + 235, + 128 + ], + "spans": [ + { + "bbox": [ + 224, + 119, + 235, + 128 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 258, + 72, + 282, + 117 + ], + "blocks": [ + { + "bbox": [ + 258, + 72, + 282, + 117 + ], + "lines": [ + { + "bbox": [ + 258, + 72, + 282, + 117 + ], + "spans": [ + { + "bbox": [ + 258, + 72, + 282, + 117 + ], + "type": "image", + "image_path": "ea5c35744b6ecea20b8f6dad9641c08b4065274b818ed69f9de81f5115219d36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "text", + "content": "ror is 0.62px, and the average distance error of the triangulated 3D points to their ground truth positions was " + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "inline_equation", + "content": "0.17\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "text", + "content": ". The size of reference checkerboard is around " + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "inline_equation", + "content": "15\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "text", + "content": " width, hence the relative error of the triangulation is around " + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 187, + 286, + 270 + ], + "type": "text", + "content": ". These results show our method provide a practical calibration and the valid scale of the target object also depends on size of the checkerboard." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 281, + 138, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 138, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 138, + 293 + ], + "type": "text", + "content": "6.3. Depth-of-Field" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "content": "Fig. 11 shows the changes in the depth-of-field by the monocentric lens capturing an object of " + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "content": " height at the same magnification given the configuration same as the real capture in Sec. 6.2.2. Fig. 11(a) shows the total depth-of-field with and without the monocentric lens " + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "inline_equation", + "content": "L_{N} + L_{F}" + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "content": " (red and blue respectively). Fig. 11(b) and (c) show " + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "inline_equation", + "content": "L_{N}" + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "inline_equation", + "content": "L_{F}" + }, + { + "bbox": [ + 46, + 300, + 287, + 431 + ], + "type": "text", + "content": " in the same manner. In these figures, the horizontal axis indicates the apparent size of the target in pixel, with assuming the measurement model used in Sec. 6.2. These plots clearly indicate that the monocentric lens contributes to increase the depth-of-field successfully." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 441, + 253, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 253, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 253, + 454 + ], + "type": "text", + "content": "6.4. Teleidoscopic 3D Shape Reconstruction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "text", + "content": "To validate the proposed teleidoscopic system as a multiview camera system for 3D shape reconstruction, this section demonstrates 3D reconstruction of small objects of approximately " + }, + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "inline_equation", + "content": "6\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 460, + 287, + 508 + ], + "type": "text", + "content": " size." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "content": "We used the system calibrated in Sec. 6.2.2. Fig. 12 visualizes our 3D reconstruction of the front side of a real sphere of " + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "inline_equation", + "content": "3\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "content": " radius with the ground-truth and the error map. The mean absolute error " + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "inline_equation", + "content": "E_{d}" + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "content": " of the estimated depth is " + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "inline_equation", + "content": "0.097\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "content": ". Compared with its size, the relative error is about " + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 509, + 287, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Fig. 13 shows additional 3D reconstructions. The 1st and the 2nd rows demonstrate that our imaging system has the ability to recover complex shapes. The proposed imaging system well recovers the complex bumpy parts of the shell and the sharp edges and corners of the object. The 2nd and the 3rd rows compare \"with\" and \"w/o\" the monocentric lens cases. As the input image (b) shows, the monocentric lens creates a surrounding view while preserving the area of each chamber, resulting in a fuller and more accurate geometry, as shown in (c). We can verify this quantitatively by comparing the numbers of pixels covering the object area in the captured image (145,217 vs. 12,730) and the numbers" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 100, + 351, + 198 + ], + "blocks": [ + { + "bbox": [ + 307, + 100, + 351, + 198 + ], + "lines": [ + { + "bbox": [ + 307, + 100, + 351, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 100, + 351, + 198 + ], + "type": "image", + "image_path": "1e840fa03f603c82fcaf928e8bdf7c04a9544cc1b12cf990cf12606354ade632.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 355, + 59, + 419, + 115 + ], + "blocks": [ + { + "bbox": [ + 355, + 59, + 419, + 115 + ], + "lines": [ + { + "bbox": [ + 355, + 59, + 419, + 115 + ], + "spans": [ + { + "bbox": [ + 355, + 59, + 419, + 115 + ], + "type": "image", + "image_path": "c70bd838d0cd6cdab51399b0df5a82a973c9cf33474578042c407494162d3c4d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 355, + 117, + 419, + 173 + ], + "blocks": [ + { + "bbox": [ + 355, + 117, + 419, + 173 + ], + "lines": [ + { + "bbox": [ + 355, + 117, + 419, + 173 + ], + "spans": [ + { + "bbox": [ + 355, + 117, + 419, + 173 + ], + "type": "image", + "image_path": "1190d819cbdbc82425c5883254016964c4fff978aa3088bec2a748233951388c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 436, + 74, + 469, + 102 + ], + "blocks": [ + { + "bbox": [ + 436, + 74, + 469, + 102 + ], + "lines": [ + { + "bbox": [ + 436, + 74, + 469, + 102 + ], + "spans": [ + { + "bbox": [ + 436, + 74, + 469, + 102 + ], + "type": "image", + "image_path": "08b5901c897b76425a4eb169698d777219de370683da0de430003ab576c992b3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 493, + 73, + 521, + 102 + ], + "blocks": [ + { + "bbox": [ + 493, + 73, + 521, + 102 + ], + "lines": [ + { + "bbox": [ + 493, + 73, + 521, + 102 + ], + "spans": [ + { + "bbox": [ + 493, + 73, + 521, + 102 + ], + "type": "image", + "image_path": "9713e13f344a192e93d1f4bf477753e01a2f46be450fda7b6def1b6ba5523c44.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 432, + 127, + 474, + 161 + ], + "blocks": [ + { + "bbox": [ + 432, + 127, + 474, + 161 + ], + "lines": [ + { + "bbox": [ + 432, + 127, + 474, + 161 + ], + "spans": [ + { + "bbox": [ + 432, + 127, + 474, + 161 + ], + "type": "image", + "image_path": "d76b5caa8d46a9081d1b61545d94e610dcb7e85015a6ec5ffed84de36963e6ef.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 163, + 494, + 171 + ], + "lines": [ + { + "bbox": [ + 427, + 163, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 427, + 163, + 494, + 171 + ], + "type": "text", + "content": "40,639 pts recovered" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 491, + 127, + 526, + 161 + ], + "blocks": [ + { + "bbox": [ + 491, + 127, + 526, + 161 + ], + "lines": [ + { + "bbox": [ + 491, + 127, + 526, + 161 + ], + "spans": [ + { + "bbox": [ + 491, + 127, + 526, + 161 + ], + "type": "image", + "image_path": "36f1a76d08c9302bf58ab2e161b52e9e87a3cce888471e1f2b586db192f1cbeb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 355, + 185, + 419, + 241 + ], + "blocks": [ + { + "bbox": [ + 406, + 176, + 480, + 184 + ], + "lines": [ + { + "bbox": [ + 406, + 176, + 480, + 184 + ], + "spans": [ + { + "bbox": [ + 406, + 176, + 480, + 184 + ], + "type": "text", + "content": "w/o monocentric lens" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 185, + 419, + 241 + ], + "lines": [ + { + "bbox": [ + 355, + 185, + 419, + 241 + ], + "spans": [ + { + "bbox": [ + 355, + 185, + 419, + 241 + ], + "type": "image", + "image_path": "011618e35a7c6f8b9d23e0ca58837101d5f693a6e5b266310edfbf05dc51cfe2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 242, + 393, + 251 + ], + "lines": [ + { + "bbox": [ + 382, + 242, + 393, + 251 + ], + "spans": [ + { + "bbox": [ + 382, + 242, + 393, + 251 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 258, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 301 + ], + "type": "text", + "content": "Figure 13. Results of 3D Shape Estimation. The second and the third row compares \"with\" and \"w/o\" the monocentric lens. (a) Room-light appearance. (b) Teleidoscopic image. (c) Reconstructed 3D shapes." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 430, + 201, + 476, + 224 + ], + "blocks": [ + { + "bbox": [ + 430, + 201, + 476, + 224 + ], + "lines": [ + { + "bbox": [ + 430, + 201, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 430, + 201, + 476, + 224 + ], + "type": "image", + "image_path": "c88bbca32efa14ac1fbdf8f46b9490602bbb756aed8546a90df57fd3cdcf9592.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 476, + 201, + 526, + 226 + ], + "blocks": [ + { + "bbox": [ + 427, + 232, + 491, + 241 + ], + "lines": [ + { + "bbox": [ + 427, + 232, + 491, + 241 + ], + "spans": [ + { + "bbox": [ + 427, + 232, + 491, + 241 + ], + "type": "text", + "content": "8,264 pts recovered" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 201, + 526, + 226 + ], + "lines": [ + { + "bbox": [ + 476, + 201, + 526, + 226 + ], + "spans": [ + { + "bbox": [ + 476, + 201, + 526, + 226 + ], + "type": "image", + "image_path": "444d6283fdb898dd71ccf9d0bff27a2e3e7237613ae2a3cb5e603ab77c458349.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 242, + 486, + 251 + ], + "lines": [ + { + "bbox": [ + 476, + 242, + 486, + 251 + ], + "spans": [ + { + "bbox": [ + 476, + 242, + 486, + 251 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 308, + 545, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 343 + ], + "type": "text", + "content": "of triangulated 3D points using the ground-truth correspondences as shown in the figure (40,639 vs. 8,264). See the supplementary material for more results." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 502 + ], + "type": "text", + "content": "Limitations The major limitations of our method are twofold. The monocentric lens has a relatively large spherical aberration, and the optical resolution of the entire system is limited by such aberrations. One possible solution is to introduce nested concentric glass spheres with different refractive indices for the monocentric lens [6, 35]. The other limitation is the trade-off between the number of virtual viewpoints and the effective resolution of each view because the use of a kaleidoscope requires zooming-out to capture the reflected images. While larger number of the cameras contributes to robustify the triangulation, it inevitably reduces the resolution of the reconstruction as the number of pixels per view decreases." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 507, + 378, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 507, + 378, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 378, + 519 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 524, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 668 + ], + "type": "text", + "content": "This paper proposed the concept of image-based 3D measurement by a teleidoscopic imaging system for microscale objects. By combining a monocentric lens with a kaleidoscope, our system realized a closeup and surrounding multi-view capture of the target. Our contributions are twofold. We proposed a ray-pixel camera model that realized an efficient numerical 3D-2D projection computation, and also proposed a calibration method for the teleidoscopic imaging system. Our future work includes aberration and diffraction blur handling such as the inverse filter proposed in [5], and also refraction/attenuation modeling in the case of capturing objects in water [18, 19]." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 678, + 408, + 690 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 678, + 408, + 690 + ], + "spans": [ + { + "bbox": [ + 306, + 678, + 408, + 690 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "This work is in part supported by JSPS KAKENHI JP22K17914." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20820" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Rajat Aggarwal, Amrisha Vohra, and Anoop M. Namboodiri. Panoramic stereo videos with a single camera. In CVPR, pages 3755-3763, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "type": "text", + "content": "[2] Amit Agrawal and Srikumar Ramalingam. Single image calibration of multi-axial imaging systems. In CVPR, pages 1399–1406, 2013. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "type": "text", + "content": "[3] Byeongjoo Ahn, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Kaleidoscopic structured light. ACM TOG, 40(6):1-15, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 286, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 234 + ], + "type": "text", + "content": "[4] Michael Chen, Zachary F. Phillips, and Laura Waller. Quantitative differential phase contrast (dpc) microscopy with computational aberration correction. Optics Express, (25):32888-32899. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 257 + ], + "type": "text", + "content": "[5] O. S. Cossairt, D. Miau, and S. K. Nayar. Gigapixel computational imaging. In ICCP, pages 1-8, 2011. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "text", + "content": "[6] Donald G. Dansereau, Glenn Schuster, Joseph Ford, and Gordon Wetzstein. A wide-field-of-view monocentric light field camera. In CVPR, 2017. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 291, + 286, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 313 + ], + "type": "text", + "content": "[7] Michael De Zeeuw and Aswin C Sankaranarayanan. Widebaseline light fields using ellipsoidal mirrors. 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 315, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 286, + 346 + ], + "type": "text", + "content": "[8] Keith Forbes, Fred Nicolls, Gerhard De Jager, and Anthon Voigt. Shape-from-silhouette with two mirrors and an uncalibrated camera. In ECCV, 2006. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "type": "text", + "content": "[9] Brian J. Ford. Single lens: The story of the simple microscope. Journal of the History of Biology, 19(2):320-321, 1986. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 380, + 286, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 380, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 286, + 424 + ], + "type": "text", + "content": "[10] Joshua Gluckman, Shree K. Nayar, and Keith J. Thoresz. Real-time omnidirectional and panoramic stereo. In Proceedings of the 1998 DARPA Image Understanding Workshop, pages 299-303. Morgan Kaufmann, 1998. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 426, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 286, + 456 + ], + "type": "text", + "content": "[11] Michael D. Grossberg and Shree K. Nayar. The raxel imaging model and ray-based calibration. *IJCV*, 61(2):119–137, 2005. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 459, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 286, + 501 + ], + "type": "text", + "content": "[12] Mohit Gupta, Amit Agrawal, Ashok Veeraraghavan, and Srinivasa G. Narasimhan. Structured light 3d scanning in the presence of global illumination. In CVPR, pages 713-720, 2011. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 502, + 286, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 525 + ], + "type": "text", + "content": "[13] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2000. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 525, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 590 + ], + "type": "text", + "content": "[14] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE TPAMI, 41(1):190-204, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 624 + ], + "type": "text", + "content": "[15] Srinivasa Karthik and Ambasamudram N. Rajagopalan. Underwater microscopic shape from focus. In ICPR, pages 2107-2112, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 625, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 679 + ], + "type": "text", + "content": "[16] Ryo Kawahara, Shohei Nobuhara, and Takashi Matsuyama. A pixel-wise varifocal camera model for efficient forward projection and linear extrinsic calibration of underwater cameras with flat housings. In Proc. of ICCV 2013 Underwater Vision Workshop, pages 819-824, 2013. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[17] Gurunandan Krishnan and Shree K. Nayar. Towards a true spherical camera. In SPIE, volume 7240, 724002, pages 1-13, 2009. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Meng-Yu Jennifer Kuo, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Non-rigid shape from water. IEEE TPAMI, 43(7):2220-2232, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[19] Meng-Yu Jennifer Kuo, Satoshi Murai, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Surface normals and shape from water. IEEE TPAMI, 44(12):9150-9162, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "text", + "content": "[20] Douglas Lanman, Daniel Crispell, and Gabriel Taubin. Surround structured lightning: 3-d scanning with orthographic illumination. In CVIU, pages 1107-1117, 2009. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 232 + ], + "type": "text", + "content": "[21] A. Manakov, J. Restrepo, O. Klehm, R. Hegedus, E. Eisemann, H. P. Seidel, and I. Ihrke. A reconfigurable camera add-on for high dynamic range, multispectral, polarization, and light-field imaging. ACM TOG, 32(4):47-1, 2013. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 234, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 255 + ], + "type": "text", + "content": "[22] P. Miraldo and H. Araujo. Calibration of smooth camera models. IEEE TPAMI, 35(9):2091-2103, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 258, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 545, + 289 + ], + "type": "text", + "content": "[23] Daniel Moreno and Gabriel Taubin. Simple, accurate, and robust projector-camera calibration. In 3DPVT, pages 464-471, 2012. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 292, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 545, + 335 + ], + "type": "text", + "content": "[24] Andrew D. Mullen, Tali Treibitz, Paul LD Roberts, Emily L. A. Kelly, Rael Horwitz, Jennifer E. Smith, and Jules S. Jaffe. Underwater microscopy for in situ studies of benthic ecosystems. Nature Communications, 7, 2016/07 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 337, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 370 + ], + "type": "text", + "content": "[25] Shree K. Nayar. Omnidirectional video camera. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 235-241, 1997. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 373, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 437 + ], + "type": "text", + "content": "[26] Markus Niederost, Jana Niederost, and Jiri Scucka. Automatic 3d reconstruction and visualization of microscopic objects from a monoscopic multifocus image sequence. International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XXXIV-5/W10:5-10, 01 2002. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "text", + "content": "[27] Ryoichiro Nishi, Takahito Aoto, Norihiko Kawai, Tomokazu Sato, Yasuhiro Mukaigawa, and Naokazu Yokoya. *Ultrashallow dof imaging using faced paraboloidal mirrors*. In ACCV, pages 115–128, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 485, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 545, + 528 + ], + "type": "text", + "content": "[28] Ilya Reshetouski and Ivo Ihrke. Mirrors in Computer Graphics, Computer Vision and Time-of-Flight Imaging, pages 77–104. Springer Berlin Heidelberg, Berlin, Heidelberg, 2013." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 531, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 563 + ], + "type": "text", + "content": "[29] Ilya Reshetouski, Alkhazur Manakov, Hans-Peter Seidel, and Ivo Ihrke. Three-dimensional kaleidoscopic imaging. In CVPR, pages 353-360, 2011. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 565, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 597 + ], + "type": "text", + "content": "[30] J. Starck, A. Maki, S. Nobuhara, A. Hilton, and T. Matsuyama. The multiple-camera 3-d production studio. IEEE TCSVT, 19(6):856-869, 2009. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 600, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 622 + ], + "type": "text", + "content": "[31] Peter Sturm. Multi-view geometry for general camera models. In CVPR, volume 1, pages 206-212, 2005. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[32] Seiichi Tagawa, Yasuhiro Mukaigawa, Jaewon Kim, Ramesh Raskar, Yasuyuki Matsushita, and Yasushi Yagi. Hemispherical confocal imaging. IPSJ Transactions on Computer Vision and Applications, 3:222-235, 2011. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[33] Kosuke Takahashi, Akihiro Miyata, Shohei Nobuhara, and Takashi Matsuyama. A linear extrinsic calibration of kaleidoscopic imaging system from single 3d point. In CVPR, 2017. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20821" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 262 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[34] Lei Tian, Jingyan Wang, and Laura Waller. 3d differential phase-contrast microscopy with computational illumination using an led array. Optics letters, 39:1326-9, 2014. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 148 + ], + "type": "text", + "content": "[35] Eric J Tremblay, Daniel L Marks, David J Brady, and Joseph E Ford. Design and scaling of monocentric multiscale imagers. Applied Optics, 51(20):4691-4702, 2012. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 194 + ], + "type": "text", + "content": "[36] Borislav Trifonov, Derek Bradley, and Wolfgang Heidrich. Tomographic reconstruction of transparent objects. In Proc. of Eurographics Conference on Rendering Techniques, pages 51-60, 2006. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 228 + ], + "type": "text", + "content": "[37] David Uhlig and Michael Heizmann. A calibration method for the generalized imaging model with uncertain calibration target coordinates. In ACCV. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 262 + ], + "type": "text", + "content": "[38] Frank Verbiest, Marc Proesmans, and Luc Van Gool. Modeling the effects of windshield refraction for camera calibration. In ECCV, pages 397-412, 2020. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20822" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_content_list.json b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7c5922543a9fac82e34ef84a6fbcff6d464af7e4 --- /dev/null +++ b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "Tell Me What Happened: Unifying Text-guided Video Completion via Multimodal Masked Video Generation", + "text_level": 1, + "bbox": [ + 153, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsu-Jui Fu $^{1}$ , Licheng Yu $^{2}$ , Ning Zhang $^{2}$ , Cheng-Yang Fu $^{2}$ , Jong-Chyi Su $^{3}$ , William Yang Wang $^{1}$ , Sean Bell $^{2}$ $^{1}$ UC Santa Barbara $^{2}$ Meta $^{3}$ NEC Laboratories America", + "bbox": [ + 251, + 202, + 723, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tsu-juifu, william}@cs.ucsb.edu jcsu@nec-labs.com {lichengyu, ningzhang, chengyangfu, seanbell}@meta.com", + "bbox": [ + 246, + 258, + 723, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating a video given the first several static frames is challenging as it anticipates reasonable future frames with temporal coherence. Besides video prediction, the ability to rewind from the last frame or infilling between the head and tail is also crucial, but they have rarely been explored for video completion. Since there could be different outcomes from the hints of just a few frames, a system that can follow natural language to perform video completion may significantly improve controllability. Inspired by this, we introduce a novel task, text-guided video completion (TVC), which requests the model to generate a video from partial frames guided by an instruction. We then propose Multimodal Masked Video Generation (MMVG) to address this TVC task. During training, MMVG discretizes the video frames into visual tokens and masks most of them to perform video completion from any time point. At inference time, a single MMVG model can address all 3 cases of TVC, including video prediction, rewind, and infilling, by applying corresponding masking conditions. We evaluate MMVG in various video scenarios, including egocentric, animation, and gaming. Extensive experimental results indicate that MMVG is effective in generating high-quality visual appearances with text guidance for TVC.", + "bbox": [ + 73, + 358, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative video modeling [15, 70, 84] has made great progress, which first succeeds in unconditional video generation [40, 64]. More recently, video prediction [28, 36, 47] has been trying the controllable setting, which anticipates the future by completing a video from the past frames or a static starting image [37, 97]. However, video prediction may produce various outcomes, which makes it difficult to meet human expectations. For the example in Fig. 1(a), the game agent can keep jumping to the right or move back and turn left. The limited guidance from only the first frame is", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71221f8c7f3394e88f91f99ec3b42f2d8cf77fdeed9c37ac836e7776eccb86ad.jpg", + "image_caption": [ + "(a) Video Prediction is Uncertain" + ], + "image_footnote": [], + "bbox": [ + 511, + 324, + 879, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Prediction put bowl in saucepan", + "bbox": [ + 503, + 465, + 599, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Rewind Barney is trying to sing in a Living room.", + "bbox": [ + 501, + 513, + 602, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Infilling Runs to Left. Then collects a coin and a gem.", + "bbox": [ + 504, + 566, + 599, + 612 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40b1d76980540e6b924dcec212f21bccbdcc37e75f605ffc9788a581bbea9fc1.jpg", + "image_caption": [ + "(b) Text-guided Video Completion (TVC)", + "Figure 1. The introduced text-guided video completion (TVC) task. (a) Video prediction may have different outcomes without text guidance. (b) TVC performs video completion from the first frame (prediction), the last frame (rewind), or both (infilling), guided by the textual description." + ], + "image_footnote": [], + "bbox": [ + 612, + 458, + 890, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "insufficient to tell the intention. For humans, language is the most straightforward way of communication. If a system can follow an instruction to accomplish video completion, it will significantly improve its controllability and make a vast application impact. On the other hand, compared with video prediction, video rewind and infilling have been rarely studied [39, 79], but they are also crucial. Breaking the limitation of chronological guidance should make the visual guidance more flexible, leading to a general video completion.", + "bbox": [ + 496, + 717, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We thus introduce a novel task, text-guided video completion (TVC), where the partial frames and a given instruction jointly guide the video generation. As illustrated in", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10681", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fig. 1(b), we consider three scenarios of video completion: prediction from the first frame, rewind from the last frame, and infilling between the head and tail. The missing (to-be-completed) event should follow the textual instruction. Compared to generating content from scratch [43, 87], TVC requests models to understand the given visual and textual guidance before generation, which better mimics how human imagines after seeing and listening in our daily lives.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle TVC, we present Multimodal Masked Video Generation (MMVG) to perform video completion. Specifically, we represent the video frames as discrete visual tokens by temporal-aware VQGAN [54, 76]. One key challenge is to deal with the video frames that are not presented in chronological (e.g., the last frame for rewind). Different from autoregressive models [23, 86] that only condition on the previous frames, MMVG carries out video completion in an encoder-decoder manner. Specifically, we propose a masking strategy that masks different parts of the video and feeds them as the input to the multimodal encoder with the instruction. As shown in Fig. 2, we allow MMVG to consider the visual hints from different time points, and the decoder learns to produce the full target video. By varying the masking conditions (including the cases of only the first or last frame being accessible), a single MMVG can address all TVC tasks, including video prediction, rewind, and infilling. Moreover, learning the recovery from partial frames also empowers MMVG with a strong temporal coherence, contributing to better generative video quality.", + "bbox": [ + 75, + 213, + 470, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider videos in diverse scenarios for the TVC evaluation. There are Kitchen [13], Flintstones [26], and MUGEN [29] corresponding to the egocentric, animation, and gaming scenes. The model should generate videos such as performing kitchen activities in the first-person view, making characters act the assigned behavior, or imitating an agent playing game. All should be guided with the first/last (or both) frame(s) and controlled through the given human instructions. We also compare MMVG with previous methods [23,57,94,96] on UCF-101 [69] and BAIR [18] for the classic video generation/prediction tasks.", + "bbox": [ + 75, + 517, + 468, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results demonstrate that instruction is necessary to make video completion controllable, MMVG can address all three TVC tasks, and our proposed masking strategy enhances the temporal modeling, which further benefits general video generation/prediction. In summary, our contributions are three-fold:", + "bbox": [ + 75, + 685, + 468, + 776 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce TVC to generate a video from partial frames and control the temporal dynamics via natural language, where our video completion includes 3 cases: prediction, rewind, and infilling.", + "- We propose MMVG with an effective masking strategy to address all TVC tasks through a single training.", + "- Extensive experiments show that our MMVG can handle various types of video completion as well as video gener" + ], + "bbox": [ + 76, + 779, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ation/prediction. We believe TVC can become a new topic in vision-and-language research.", + "bbox": [ + 511, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 137, + 640, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Generation/Prediction. Video generation aims to synthesize diverse videos from latent inputs [1, 74, 80]. Various generative modelings have shown promising results, including generative adversarial networks (GAN) [10, 24, 72, 96], autoregressive transformers [23, 77, 94], and denoising diffusion models [16, 33, 34]. Upon that, video prediction [2, 3, 25, 57, 85], which considers past frames to anticipate future observations, should maintain temporal dynamics from static images. Though the overall idea is also to complete a video from partial frames, other tasks, such as rewind and infilling [39, 79, 90], are not extensively explored. In this paper, we introduce TVC to comprehensively investigate the ability of video completion and make it more maneuverable via textual description.", + "bbox": [ + 496, + 162, + 890, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image/Video Genration. Generating visual content from language [9, 50, 71] has a vast application value in creative visual design. Previous works rely on the GANs framework [49] to produce images [19, 20, 22, 55, 61, 92] or videos [43, 46, 52], conditioned on text. With large-scale datasets [4, 65, 66, 83], recent pre-trained models can generate high-quality natural images from open-domain description through discrete visual tokens [12, 17, 54, 59, 76, 95] or the diffusion process [51, 58, 62, 63]. Leveraging such techniques further extends to generate vivid videos [32, 35, 68, 78, 86, 87]. However, those methods that depend on autoregressive generation can only be guided chronologically [27, 38]. Besides, video diffusion models require a deterministic video length, which cannot consider diverse temporal durations. In contrast, MMVG can perform video completion in different lengths from arbitrary time points and address all TVC tasks just with a single training.", + "bbox": [ + 496, + 378, + 890, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-guided Video-to-Video. Video inpainting [7, 41, 91], segmentation reconstruction [81, 82], or video style transfer [8, 14, 88] can be seen as a particular case of video-to-video synthesis (V2V). Even if text-guided V2V [5, 21, 93] can be controlled by language, it is still conditioned on a full video, where the temporal dynamics are usually provided. Different from that, TVC requires to regain the missing event from just partial guidance. It is more challenging since the model has to capture what happened from the instruction, maintain the temporal coherence among limited frames, and produce a complete video.", + "bbox": [ + 496, + 638, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Text-guided Video Completion (TVC)", + "text_level": 1, + "bbox": [ + 500, + 820, + 831, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Task Definition", + "text_level": 1, + "bbox": [ + 500, + 845, + 653, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We study the text-guided video completion (TVC) task to perform video completion from the first frame (prediction),", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10682", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ffed10af35ea387c77927a4c6216594d6774eeed7c65d4770f57a4ba0c64d3fe.jpg", + "image_caption": [ + "Figure 2. An overview of our Multimodal Masked Video Generation (MMVG). We present temporal-aware VQGAN (T-VQ) for discrete visual representation. MMVG considers the instruction $\\mathcal{X}$ and partial frames of the video $\\nu$ from diverse time points through masking, learning to generate the complete $\\nu$ . In this way, a single trained MMVG can perform all prediction, rewind, and infilling tasks." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 895, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the last frame (rewind), or the head and tail (infilling), conditioned on the textual instruction. During training, we have pairs of videos $\\mathcal{V}$ and corresponding instructions $\\mathcal{X}$ . Specifically, $\\mathcal{V}$ consists of $N$ frames as $\\{v_{1}, v_{2}, \\ldots, v_{N}\\}$ . Our goal is to train a unified model that generates the complete $\\mathcal{V}$ given the partial frames from arbitrary time points and $\\mathcal{X}$ .", + "bbox": [ + 75, + 417, + 470, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multimodal Masked Video Generation", + "text_level": 1, + "bbox": [ + 76, + 523, + 410, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. An overview of our Multimodal Masked Video Generation (MMVG) is illustrated in Fig. 2. To model the video along with language, we propose temporal-aware VQGAN to represent a frame as visual tokens, which converts it into the same discrete space as the words. We present an effective masking strategy that masks different video parts for video completion learning. Those missing fragments are replaced by the unique [SPAN] tokens, and we consider the visual guidance from diverse time points. The multimodal encoder consumes the text and the partial-missing video, and the decoder learns to produce the complete video from arbitrary guided frames. By varying the masking conditions, MMVG learns to utilize the [SPAN] token and unifies all TVC tasks during the training.", + "bbox": [ + 75, + 547, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal-aware Discrete Visual Tokens. VQ-VAE [76] has shown promising capability in representing data as discrete tokens. VQGAN [54] further models the prior distribution of the latent space via a transformer with the GAN training. If VQGAN is directly applied onto videos, it will ignore the inner temporal coherence and treat each frame as an independent image, resulting in an unsmooth video reconstruction. Though TATS [23] attempts to handle this by making $k$ consecutive frames altogether during VQ, it has", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to pre-define the constant $k$ before training. Such constraint forbids it from representing a frame at any timestamp.", + "bbox": [ + 496, + 417, + 890, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address it with flexibility, we propose temporal-aware VQGAN (T-VQ) to inject the temporal relationship into the latent representation. We first follow VQGAN to learn the target visual tokens $z_{i}$ by reconstructing a video frame $v_{i}$ :", + "bbox": [ + 496, + 449, + 890, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {i} = \\mathbf {q} (\\mathsf {E n c} ^ {\\mathsf {Q}} (v _ {i}) \\mid C), \\\\ \\hat {v} _ {i} = \\operatorname {D e c} ^ {0} (\\mathrm {z} _ {i}), \\\\ \\mathcal {L} _ {\\mathrm {V Q}} = \\underbrace {\\left| \\left| \\hat {v _ {i}} - v _ {i} \\right| \\right| _ {1}} _ {\\text {r e c o n s t r c u t i o n}} + \\underbrace {\\left| \\left| \\operatorname {s g} [ \\operatorname {E n c} ^ {Q} (v _ {i}) ] - C _ {z _ {i}} \\right| \\right| _ {2} ^ {2}} _ {\\text {c o d e b o o k}} \\tag {1} \\\\ + \\underbrace {\\beta | | \\mathrm {s g} [ C _ {\\mathrm {z} _ {i}} ] - \\mathrm {E n c} ^ {\\mathrm {Q}} (v _ {i}) | | _ {2} ^ {2}} _ {\\text {c o m m i t}} + \\underbrace {| | \\mathcal {F} (\\hat {v} _ {i}) - \\mathcal {F} (v _ {i}) | | _ {1}} _ {\\text {m a t c h i n g}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 516, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{Enc}^{\\mathrm{Q}}$ and $\\mathrm{Dec}^{\\mathrm{Q}}$ are the VQ encoder and decoder. The discrete latent code $z_{i}$ is acquired from the quantization operation q [54], which adopts nearest neighbor search by the trainable codebook $C$ . We apply the straight-through estimator over the stop-gradient operation sg and use $\\beta$ as 0.25 [76]. We also append VGG [67] features matching $\\mathcal{F}$ to stabilize the VQ loss $\\mathcal{L}_{\\mathrm{VQ}}$ [23]. The adversarial training between the frame quality loss $\\mathcal{L}_G$ and discrimination loss $\\mathcal{L}_D$ are further calculated from the discriminator $D$ :", + "bbox": [ + 496, + 642, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {G} = \\log (1 - D (\\hat {v} _ {i})), \\\\ \\mathcal {L} _ {D} = \\log (1 - D (\\hat {v} _ {i})) + \\log (D (v _ {i})). \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 786, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To inject the temporal relationship into $z$ , T-VQ is trained with the introduced contrastive temporal reasoning:", + "bbox": [ + 500, + 828, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no _ {i} = \\mathbf {F C} ^ {\\mathrm {T}} \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 866, + 890, + 891 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T}} = \\operatorname {B C E L o s s} \\left(o _ {i}, 0 \\text {i f} i > j \\text {e l s e} 1\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 888, + 818, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $j$ is a random frame from the same video. $\\mathrm{FC^T}$ is the MLP classifier, and BCELoss is the binary cross-entropy for before/after. Learning the temporal order from $\\mathcal{L}_{\\mathrm{T}}$ , $\\mathbf{Z}$ facilitates an implicit temporal coherence, leading to smooth video modeling. Moreover, since $\\mathbf{Z}$ represents a single image, it is flexible to support frames at arbitrary timestamps.", + "bbox": [ + 75, + 89, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation from Masked Video. We propose the masking strategy $\\mathcal{M}$ to obtain the masked videos $\\overline{\\nu}$ from diverse time points. $\\mathcal{M}$ masks out most video frames with the probability $p$ and replaces each fragment as a unique [SPAN] token. For example, $\\mathcal{M}$ reserves the third and the fifth frame, and masks all the others over a video length of 5:", + "bbox": [ + 75, + 185, + 470, + 275 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {V}}} \\colon \\{[ S ], v _ {3}, [ S ], v _ {5} \\} = \\mathcal {M} (\\mathcal {V} \\mid p). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 286, + 468, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to recover the missing part from $\\overline{\\mathcal{V}}$ and perform video completion, guided by the instruction $\\mathcal{X}$ . To model between the vision and language modalities, we apply our Enc $^{\\mathrm{Q}}$ over $\\overline{\\mathcal{V}}$ for the discrete visual tokens $\\{[\\mathrm{S}],\\mathrm{z}_3,[\\mathrm{S}],\\mathrm{z}_5\\}$ . We also tokenize the text $\\mathcal{X}$ into word tokens $\\{w_i\\}_{i = 1}^L$ with the CLIP tokenizer [56], where $L$ is the length of $\\mathcal{X}$ . As in the same discrete space, MMVG can achieve cross-modal fusion by the multimodal encoder (Enc $^{\\mathrm{M}}$ ) through the self-attention mechanism as the transformer [77]:", + "bbox": [ + 75, + 316, + 470, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {i} ^ {w}, f _ {j} ^ {v} = \\mathrm {L P} ^ {w} \\left(w _ {i}\\right), \\mathrm {L P} ^ {v} \\left(\\mathrm {z} _ {j}\\right) \\\\ \\{1 \\} = \\sum_ {i = 1} ^ {n} M (f _ {i}) w _ {i} \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 463, + 468, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{h \\} = \\operatorname {E n c} ^ {\\mathrm {M}} ([ \\{f ^ {w} \\}, \\{f ^ {v} \\} ]),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 483, + 377, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where it obtains the features $f$ by the linear projection (LP), and $h$ is the hidden encoding features. We can also regard LP as the video/language embedder, which extracts the preliminary visual/linguistic features.", + "bbox": [ + 75, + 512, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After encoding the language hint and the partial-missing video from Enc $^{\\text{M}}$ , our video decoder (Dec $^{\\text{M}}$ ) learns to produce all frames for comprising the complete video. Dec $^{\\text{M}}$ follows the vanilla autoregressive decoder, which first conducts self-attention over the past generated tokens and then predicts the discrete visual tokens as the video frame, conditioned on the encoded features $h$ :", + "bbox": [ + 75, + 573, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t} = \\operatorname {D e c} ^ {\\mathrm {M}} (\\left\\{\\hat {z} _ {1}, \\dots , \\hat {z} _ {t - 1} \\right\\} \\mid \\{h \\}),\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 686, + 387, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\underset {N} {\\operatorname {C E L o s s}} \\left(\\hat {z} _ {t}, z _ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 708, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M}} = \\sum_ {t = 1} ^ {N} \\mathcal {L} _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 729, + 253, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t$ is the ground-truth tokens of the frame $v_t$ in the original $\\mathcal{V}$ . We calculate the video decoding loss $\\mathcal{L}_{\\mathrm{M}}$ by the cross-entropy (CELoss) to learn video generation as classification. Our $\\mathrm{Dec}^{\\mathrm{M}}$ is built upon VideoSwin [45], which has shown a strong visual perception on various video understanding tasks. The 3D-shifted windows [44] consider different levels of spatial-temporal attention, and each window models blocks of video patches across $T'$ consecutive", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Multimodal Masked Video Generation" + ], + "code_body": "1: while TRAIN_T-VQ do \n2: $\\mathcal{V}\\gets$ sample video \n3: $z_{i} = q(\\mathrm{Enc}^{Q}(v_{i})\\mid C)$ \n4: $\\hat{v}_i = \\mathrm{Dec}^{\\mathrm{Q}}(z_i)$ \n5: $o_i = \\mathrm{FC}^{\\mathrm{T}}(z_i,z_j)$ ▷ randomly sampled frame $j$ \n6: $\\mathcal{L}_{\\mathrm{VQ}},\\mathcal{L}_G\\gets$ reconstruction, frame quality loss $\\triangleright$ Eq. 1&2 \n7: $\\mathcal{L}_{\\mathrm{T}}\\gets$ temporal ordering loss $\\triangleright$ Eq. 3 \n8: Update T-VQ by minimizing $\\mathcal{L}_{\\mathrm{VQ}} + \\mathcal{L}_G + \\mathcal{L}_{\\mathrm{T}}$ \n9: $\\mathcal{L}_D\\gets$ discrimination loss $\\triangleright$ Eq. 2 \n10: Update D by maximizing $\\mathcal{L}_D$ \n11: end while \n12: \n13: while TRAIN_MMVG do \n14: $\\mathcal{V},\\mathcal{X},p\\gets$ sample video/instruction/probability \n15: $\\overline{\\mathcal{V}}:\\{v_a,[S],v_b,\\ldots\\} = \\mathcal{M}(\\mathcal{V}\\mid p)$ $\\triangleright$ diverse guided frames \n16: $\\{\\mathbf{z}_a,[S],\\mathbf{z}_b,\\ldots\\},\\{w\\} = \\mathrm{Enc}^{\\mathrm{Q}}(\\overline{\\mathcal{V}}),$ Tokenizer(X) \n17: $\\{h\\} = \\mathrm{Enc}^{\\mathrm{M}}([\\{w\\},\\{z_a,[S],z_b,\\ldots\\}])$ \n18: for $t\\gets 1$ to $N$ do \n19: $\\hat{z}_t = \\mathrm{Dec}^{\\mathrm{M}}(\\{z_1,\\dots,z_{t-1}\\}\\mid\\{h\\})$ $\\triangleright$ teacher-forcing \n20: $\\mathcal{L}_t\\gets$ video decoding loss $\\triangleright$ Eq. 6 \n21: end for \n22: $\\hat{\\nu} = \\mathrm{Dec}^{\\mathrm{Q}}(\\{\\hat{z}_{t=1}^N\\})$ \n23: $\\mathcal{L}_{\\mathrm{M}} = \\sum_{t=1}^{N}\\mathcal{L}_{t}$ \n24: Update MMVG by minimizing $\\mathcal{L}_{\\mathrm{M}}$ \n25: $p\\gets$ update masking probability $\\triangleright$ Eq. 8 \n26: end while", + "bbox": [ + 500, + 107, + 890, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frames. To ensure the same dimension for video generation in $\\mathrm{Dec}^{\\mathrm{M}}$ , we remove the temporal down-sampling layer. In the end, we can utilize $\\mathrm{Dec}^{\\mathrm{Q}}$ to reconstruct all the frames as our completed videos $\\hat{\\mathcal{V}}$ :", + "bbox": [ + 496, + 494, + 890, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {V}} = \\operatorname {D e c} ^ {\\mathrm {Q}} \\left(\\left\\{\\hat {z} \\right\\} _ {t = 1} ^ {N}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 564, + 890, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By varying the masking conditions through $\\mathcal{M}$ , MMVG learns how to complete a video from partial frames $\\overline{\\nu}$ at arbitrary time points with the text, which overcomes the limitation of chronological guidance. To make $\\mathcal{M}$ more effective, we apply an adaptive probability $p$ instead of random sampling every time. Each video $\\mathcal{V}$ keeps its own $p$ , and all frames are equally initialized in the beginning. Based on the prediction error, we adjust the masking probability $p_t$ of the $t$ -th frame:", + "bbox": [ + 496, + 594, + 890, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} = p _ {t} + \\alpha \\left(\\left(\\frac {\\mathcal {L} _ {t}}{\\mathcal {L} _ {M}} \\sum p\\right) - p _ {t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 736, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is the adjusting rate. A larger video decoding loss $\\mathcal{L}_t$ indicates that the $t$ -th frame is more difficult to recover. MMVG learns more from those challenging cases and can bring better generative quality for video completion.", + "bbox": [ + 496, + 776, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unifying TVC during Inference. After training with text and partial-missing video, MMVG learns to perform video completion over [SPAN] tokens. Then for inference, Enc $^{\\text{M}}$ takes the following as its input to support different tasks:", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10684", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/89cb21800a614ffb7186a8866114b97b41a1dbcfa173611dd84bbede6ae29dcc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTextTVPredictionTVRewindTVInfilling
KitchenFlintstonesMUGENKitchenFlintstonesMUGENKitchenFlintstonesMUGEN
FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑
FILM [60]✘------------250.256.1352.751.4538.85.9
VideoMAE [73]✘328.947.6317.555.6548.77.0365.948.2335.555.9545.27.1246.954.7211.560.6494.97.8
TATS [23]✘106.964.4127.560.3376.57.1107.762.7127.660.2350.87.271.572.7119.566.7328.28.4
MMVGU✘105.663.3124.860.5374.57.2109.862.6124.359.7356.47.071.573.4121.866.3328.47.8
MMVGS✘103.864.5123.860.8369.47.3105.963.6123.860.5347.87.268.573.6118.567.9324.38.4
TATS [23]✓87.266.3115.970.690.167.989.863.3116.370.489.868.757.477.695.878.258.973.6
MMVGU✓80.268.4108.272.984.870.283.266.9113.271.693.168.459.877.892.878.359.273.2
MMVGS✓75.668.8106.373.783.371.179.768.1107.272.988.770.056.078.191.679.657.274.1
", + "bbox": [ + 80, + 88, + 893, + 247 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TVPrediction: $\\left[\\left\\{w\\right\\},\\left\\{z_{1},[SPAN]\\right\\}\\right]$", + "- TVRewind: $\\left[\\left\\{\\boldsymbol{w}\\right\\},\\left[\\left[\\mathrm{SPAN}\\right],\\mathbf{z}_{N}\\right]\\right]$", + "TVInfilling: $\\left[\\left\\{w\\right\\} ,\\left\\{\\mathbf{z}_1,\\left[\\mathrm{SPAN}\\right],\\mathbf{z}_N\\right\\}\\right]$" + ], + "bbox": [ + 76, + 291, + 341, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, a single trained MMVG can unify all TVC tasks without the specific downstream fine-tuning.", + "bbox": [ + 75, + 338, + 467, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning of MMVG", + "text_level": 1, + "bbox": [ + 76, + 375, + 269, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algo. 1 illustrates the learning process of our proposed MMVG for TVC. We first train T-VQ over video frames for discrete visual tokens with contrastive temporal reasoning. Specifically, we minimize the VQ reconstruction loss $\\mathcal{L}_{\\mathrm{VQ}}$ and frame quality loss $\\mathcal{L}_G$ along with our temporal ordering loss $\\mathcal{L}_{\\mathrm{T}}$ to optimize T-VQ. At the same time, we also update the discriminator $D$ via the standard adversarial training by maximizing the discrimination loss $\\mathcal{L}_D$ . For video completion, the masking strategy $\\mathcal{M}$ masks the video frames with the probability $p$ and then acquires guided frames from diverse time points. MMVG regards text and partial-missing video by Enc for cross-modal fusion, and Dec further predicts the visual tokens of frames autoregressively as the complete video. As a sequential generation process, we apply the teacher-forcing trick. Instead of our predicted $\\hat{z}$ , the ground-truth $z$ from the previous timestamp is fed to stabilize the training. Each video decoding loss $\\mathcal{L}_t$ at timestamp $t$ is summed up as $\\mathcal{L}_{\\mathrm{M}}$ to optimize MMVG. According to $\\mathcal{L}_t$ , we update $p$ for effective masking probability. The entire optimization object can be summarized as two phases:", + "bbox": [ + 75, + 397, + 472, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {T - V Q}: \\min _ {\\mathrm {E n c} ^ {\\mathrm {Q}}, \\mathrm {D e c} ^ {\\mathrm {Q}}, C, \\mathrm {F C} ^ {\\mathrm {T}}} \\max _ {D} \\mathcal {L} _ {\\mathrm {V Q}} + \\mathcal {L} _ {G} + \\mathcal {L} _ {D} + \\mathcal {L} _ {\\mathrm {T}} \\\\ \\text{MMVG:}\\min_{\\substack{\\text{Enc}^{\\text{M}},\\text{Dec}^{\\text{M}}}}\\mathcal{L}_{\\text{M}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 705, + 434, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 786, + 266, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. As a new task, we consider diverse video scenes with natural instructions for TVC. Kitchen [13] records 22K egocentric videos about kitchen activity, which have different lengths (4-16 frames) with narrations. Flintstones [26] contains 25K animation videos (15 frames) from The Flintstones, where each video description includes the characters", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/031f12fab0b218d434a6cb4c1c78a8713d8aacb8d93d8a79b76ddc4eefccdecb.jpg", + "table_caption": [ + "Table 1. Results of TVC, including prediction, rewind, and infilling, on Kitchen [13], Flintstones [26], and MUGEN [29]. TATS [23] requires specific training to support different tasks. We further train the unified MMVG $^{\\text{U}}$ for each specific task as MMVG $^{\\text{S}}$ ." + ], + "table_footnote": [], + "table_body": "
DatasetTrain / Val#Frame#WordFPS
Kitchen [13]16,695 / 5,8048.32.86
Flintstones [26]22,666 / 2,5181516.55
MUGEN [29]362,239 / 12,8481620.65
", + "bbox": [ + 540, + 289, + 849, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. The statistics of our used datsets to evaluate TVC.", + "bbox": [ + 521, + 359, + 867, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and their behavior. MUGEN [29] is built from agents playing CoinRun [11], which consists of 375K gaming videos (16 frames) with detailed text annotations. All videos in these three datasets are resized into $128 \\times 128$ . An overview is shown in Table 2 and Fig. 1(b). Since MMVG can unify various tasks, we also evaluate video generation/prediction on widely-used UCF-101 [69] and BAIR [18], video infilling on UCF-101 following RaMViD [39], and text-to-video generation on MSRVTT [89].", + "bbox": [ + 496, + 388, + 890, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We apply the following metrics to evaluate TVC results: 1) FVD [75] computes the video features [6] distance to the ground truth; 2) RCS [86] is the relative visual-text similarity to the instruction, compared to the ground-truth video. We fine-tune the CLIP model [56] on each dataset and adapt it to the video scene for a more precise alignment. Apart from automatic metrics, we also conduct a human evaluation from aspects of video quality, instruction relevance, and ground-truth similarity. We sample 75 TVP results for each task and adopt MTurk1 to rank over baselines and our MMVG. To avoid the potential ranking bias, we hire 3 MTurkers for each sample of prediction, rewind, and infilling tasks.", + "bbox": [ + 496, + 527, + 892, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Detail. T-VQ contains ResBlocks [31] as the visual auto-encoder (Enc $^{\\text{Q}}$ and Dec $^{\\text{Q}}$ ). The discriminator $D$ follows a similar architecture to Enc $^{\\text{Q}}$ . For the vector quantization, we use a patch size 16, where a 128x128 video frame transforms into 8x8 discrete visual tokens. There are 1024 vocabularies in the codebook $C$ , and the hidden embedding size is 256. We adopt batch size 32 with a learning rate of 4.5e-6 to optimize T-VQ by Adam [42]. MMVG is built in an encoder-decoder manner, where Enc $^{\\text{M}}$ is a trans", + "bbox": [ + 496, + 727, + 893, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Amazon MTurk: https://www.mturk.com. Our studies have been cleared by the human subject committee as an IRB-exempt protocol.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26bdefd1ba60f8306286c71e21c386b97f60830a3e3647c152ae09d6de27a468.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodKitchenFlintstonesMUGEN
VideoDiff [34]138.6206.4410.7
MCVD [79]119.9183.8400.2
TATS [23]115.5157.5386.4
MMVG109.1127.6368.6
", + "bbox": [ + 140, + 88, + 405, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "former with 24 layers, 16 attention heads, and hidden embedding size 1024. $\\mathrm{Dec}^{\\mathrm{M}}$ adopts a similar setting with temporal window size 3 in VideoSwin. The initial sample rate $p$ of the masking strategy $\\mathcal{M}$ is 0.9 with an adjusting rate $\\alpha$ as 0.1. We optimize MMVG through the mixed precision [48] with batch size 4 by Adam. The learning rate is also 4.5e-6. All experiments are implemented in PyTorch [53] and done on 8 NVIDIA A100 GPUs.", + "bbox": [ + 75, + 200, + 468, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 330, + 215, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the results of all text-guided prediction, rewind, and infilling for TVC. VideoMAE [73] is built upon MAE [30] and reconstructs the missing video cubes, which performs TVC by masking all video frames except the first or the last (or both). TATS [23], the SOTA on video generation, also produces videos as discrete visual tokens. Since TATS can only consider the past through the autoregressive transformer, it requires specific training for each task. We have MMVG $^{\\text{U}}$ as the unified model that can support all TVC tasks with a single training and MMVG $^{\\text{S}}$ to further train for each prediction, rewind, and infilling. We treat TATS as our main baseline $^{2}$ and study the importance of guided text.", + "bbox": [ + 75, + 354, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TVPrediction. VideoMAE attempts to produce all frames simultaneously, which is difficult to maintain video temporal consistency, resulting in a high 328.9 FVD on Kitchen. TATS is inherently designed for prediction as it generates the frames one after one. However, our unified MMVGU performs better than TATS on all datasets (e.g., lower 105.6 and 124.8 FVD on Kitchen and Flintstones). These results support that learning from diverse time points will not hurt the prediction from the past. In contrast, our masking strategy can bring superior temporal coherence. MMVGS further improves itself through training prediction as completion from the head. However, there are too many possible outcomes from just the beginning, where the predicted results may not meet the expectation (e.g., a high 370 FVD on MUGEN). The instruction as guidance makes it related to the expected ground-truth result. We can let MUGEN run, jump, or collect coins as the textual descriptions to achieve more controllability, leading to a noticeable improvement (e.g., a lower 84.8 FVD by MMVGU). The higher 70.2 RCS also shows that our MMVG can produce MUGEN videos", + "bbox": [ + 75, + 537, + 470, + 840 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f99134fe919406babff284fa6354c8e8b17b2542bc5ce050e73fc41295e5ff09.jpg", + "table_caption": [ + "Table 3. FVD results of video generation on our TVC datasets." + ], + "table_footnote": [], + "table_body": "
MethodUCF-101
IS↑FVD↓
VideoGPT [94]24.7-
DIGAN [96]32.7577
VideoDiff [34]57.0-
TATS [23]57.6420
MMVG58.3395
", + "bbox": [ + 531, + 88, + 692, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/13db1d95cea444e2f6977f895e93de00da5a67dd55e412ba9142f097880e873f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBAIR
FVD↓
VideoGPT [94]103.3
MaskViT [25]93.6
MCVD [79]89.5
TATS [23]88.6
MMVG85.2
", + "bbox": [ + 745, + 88, + 880, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Results of video prediction on BAIR [18].", + "bbox": [ + 728, + 204, + 895, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results of video generation on UCF-101 [69].", + "bbox": [ + 498, + 205, + 725, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that confirm with the instruction. Although the model may try to imagine the animation or the kitchen activity, the language hint can provide a clear goal to anticipate. Likewise, MMVG $^{\\text{U}}$ with text surpasses TATS, even though it is not designed for prediction only. The specific trained MMVG $^{\\text{S}}$ benefits the unified model for further improvement.", + "bbox": [ + 496, + 243, + 890, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TVRewind. Rewind from the last allows the model to imagine what happened along with a suitable opening. In addition, the objects may not display on the last frame (e.g., the spoons and forks for \"close drawer\"), which makes it more challenging to complete. Similar to prediction, VideoMAE cannot have feasible rewind results. Language is still essential to remind the past and establish an adequate beginning, where we can find a significant performance gap between with and without text (e.g., 90 vs. 350 FVD on MUGEN). Our unified MMVG $^{\\text{U}}$ achieves comparable results to TATS and even outperforms on Kitchen and Flintstones (e.g., higher 66.9 and 71.6 RCS). With the learning of completion from partial frames, autoregressive model can still accomplish video rewind without specific training. If following TATS design to train MMVG $^{\\text{U}}$ for rewind, MMVGS gains more improvement and utterly surpasses it.", + "bbox": [ + 496, + 335, + 892, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TVInfilling. We consider the additional FILM [60] for infilling, which performs video interpolation with in-between motion. Despite synthesizing intermediate frames between the first and the last, the visual dynamics are changing too rapidly to handle, resulting in a higher FVD. With guidance from the head and tail, we find a noticeable improvement even without instruction (e.g., lower FVDs on Kitchen), which is helpful in temporal video modeling. To capture the expected missing event, we still require the language hint for more controllability. Our unified MMVG $^{\\text{U}}$ achieves comparable performance to TATS again, which is specifically trained for the infilling task. It shows that completion from partial frames at different time points still helps, and MMVG $^{\\text{S}}$ further outperforms on TVInfilling.", + "bbox": [ + 496, + 580, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Additional Study", + "text_level": 1, + "bbox": [ + 500, + 801, + 668, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Generation/Prediction. We further evaluate the classic video generation and prediction tasks. Table 3 shows FVD scores of unconditional video generation on our TVC datasets. Note that only videos but no texts are used in these experiments. Both VideoDiff [34] and MCVD [79] are built", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Since we cannot receive feasible results after training diffusion methods for our TVC, we evaluate unconditional video generation in Sec. 4.3. We use this repo (https://github.com/lucidrains/video-diffusion-pytorch) as VideoDiff and the official repo as MCVD.", + "bbox": [ + 75, + 849, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10686", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16ba09493c37618a7ef190d4d39597c36d3958197395b371c661ea2eece24d4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodUCF-101 (FVD↓)
K=+1+2+5±1±2
RaMViD [39]349.7300.6260.5215.4162.5
MMVG316.3258.5194.6183.2120.3
", + "bbox": [ + 138, + 88, + 405, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "upon denoising diffusion [33], where MCVD also considers different partial frames during training. The results first indicate that the vanilla token-based method is superior to the diffusion-based model (TATS vs. VideoDiff) for video generation. In addition, MMVG, with the masking strategy that learns the visual guidance from diverse time points, further boosts the performance (the lowest 127.6, 109.1, and 368.6 FVD on Flintstones, Kitchen, and MUGEN, respectively).", + "bbox": [ + 75, + 195, + 467, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate MMVG on UCF-101 [69], which is challenging to generate natural human videos. Table 4 supports that our MMVG can produce videos with higher visual similarity (a higher 58.3 IS) and temporal alignment (a lower 395 FVD) to the ground truth. For video prediction, we apply the widely-used BAIR [18] in Table 5, where the model has to anticipate how a robot pushes objects from the given first frame. MMVG again surpasses TATS with the lowest 85.2 FVD. Although both generation and prediction are generating video frames chronologically, the ability to recover arbitrary missing frames for video completion empowers MMVG with a stronger temporal coherence, leading to better generative video quality.", + "bbox": [ + 75, + 316, + 467, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Infilling. We follow RaMViD [39] to evaluate video infilling on UCF-101. We consider various guidance settings $K$ in Table 6. For example, $K = +1$ means given the first frame, and $K = \\pm 2$ should provide the first and last two frames. For prediction, MMVG outperforms RaMViD on all $K$ , and the performance gap gets even larger when more guided frames are accessible (e.g., 33.4 on $K = +1$ and 65.9 on $K = +5$ ). A similar result can be found for infilling, where MMVG can make the lowest 120.3 FVD on $K = \\pm 2$ . Despite having a similar masking strategy, it shows that generating frames one after one still brings superior results. MMVG allows autoregressive models to condition on visual hints from any time point, which produces more similar videos to the ground truth when infilling between the head and tail.", + "bbox": [ + 75, + 518, + 467, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-to-Video Generation. Being a multimodal generative model, MMVG supports text-to-video generation. To compare with those large-scale methods, we pre-train MMVG using WebVid [4], which contains 2.5M text-video pairs. We adopt the masking strategy to treat the pre-training as video completion. MMVG outperforms GODIVA [86] and NUWA [87] without access the MSRVTT [89] data in Table 7. Surprisingly, MMVG can generate videos that are more related to the texts (a higher 0.2644 CLIP-S [86]) thanCogVideo [35], even though using twice less data. This result encourages the effectiveness of completion from partial", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d827b65f5d15a938e25b3d6b6b721a6e45ffd561617d0fe3e678773bfb74ef83.jpg", + "table_caption": [ + "Table 6. Results of video prediction and infilling on UCF-101." + ], + "table_footnote": [], + "table_body": "
MethodPre-trainingZero-shotMSRVTT
FID↓CLIP-S↑
GODIVA [86]136MX-0.2402
NUWA [87]3.9MX47.70.2439
CogVideo [35]5.4M✓23.60.2631
Make-A-Video [68]20M✓13.20.3049
MMVG2.5M✓23.40.2644
TATS [23]-X63.20.2326
MMVG-X60.60.2385
", + "bbox": [ + 531, + 88, + 857, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0bdb589d447fc28bd8ba5ae729f86c7d0853e568b824d9e6b7e5f14a79d94ea0.jpg", + "table_caption": [ + "Table 7. Results of text-to-video generation on MSRVTT [89]. We gray out methods that use significantly more pre-training data." + ], + "table_footnote": [], + "table_body": "
MethodTextKitchenFlintstonesMUGEN
Q.T.GTQ.T.GTQ.T.GT
MMVG✘1.991.811.821.731.661.622.031.561.55
TATS [23]✓1.972.072.032.072.122.171.942.112.19
MMVG✓2.042.122.152.202.222.212.032.332.26
", + "bbox": [ + 501, + 275, + 890, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Human evaluation for TVP with aspects of video quality (Q.), instruction relevance (T.), and ground-truth similarity (GT).", + "bbox": [ + 500, + 367, + 890, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames. For a fair comparison without additional data, we directly train on MSRVTT. MMVG still outperforms TATS, which shows that text-to-video generation can be improved through learning from video completion as well.", + "bbox": [ + 498, + 409, + 890, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Evaluation. We study the video quality (Q.), the relevance to the instructions (T.), and the similarity to the ground-truth video (GT) of the produced videos from the human aspect. The results in Table 8 are calculated as the mean ranking score (from 1 to 3, the higher is better) of each method for TVPrediction. MMVG without text even generates higher quality videos than TATS with text on Kitchen and MUGEN, where completion from partial frames benefits the temporal coherence of generative video modeling. While, the lowest ground-truth similarity illustrates that language guidance is crucial for controllability. With instruction, MMVG anticipates the future as the text (the highest T.) and generates videos that meet the ground truth (the highest GT), achieving the best overall performance.", + "bbox": [ + 496, + 473, + 890, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 3 illustrates the keyframes of the generated examples of three datasets. Thanks to the learning of completion from partial frames at diverse time points, a single trained MMVG can support all TVC tasks. For prediction, MMVG makes Fred \"turn his head\" or MUGEN \"jumps over the gear\" from the guided text. MMVG further recovers the missing spoons and forks for \"close drawer\" from the last frame in a more challenging rewind scenario. MMVG infills the missing event described in language such as \"stand up for dancing\", \"walk across the kitchen\", \"jump onto the stage\" from the head and tail.", + "bbox": [ + 496, + 688, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the same visual guidance, MMVG can lead TVC results using different texts, achieving controllable video completion. For example, we can let MUGEN \"jump down", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10687", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25f55c810660193d84f63031116c2a08db8790412d864d211dcf0dae800a38a1.jpg", + "image_caption": [ + "Figure 3. Qualitative examples of TVC on Kitchen [13], Flintstones [26], and MUGEN [29]. We also illustrate video generation on UCF-101 [69], video prediction on BAIR [18], and text-to-video prediction on WebVid [4]." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ground\" or \"land on a face\", starting from the same beginning. We can also control the behavior as \"keep walking\" or \"jump over a gear\" to recover the missing middle event. Furthermore, MMVG also carries out unconditional video generation with smooth temporal coherence. We can use language to produce natural dynamics in diverse scenes, such as \"sunset on the sea\", \"green sea turtle swims\", or a close look of \"cut chicken\". The presented videos indicate that our method not only unifies all TVC tasks but also performs the classical video generation/text-to-video well.", + "bbox": [ + 75, + 704, + 472, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 703, + 620, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a novel task of text-guided video completion (TVC) that performs video completion from the first, last, or both frame(s) controlled by language. We present Multimodal Masked Video Generation (MMVG) with an effective masking strategy to learn the visual guidance from any time point. By varying the masking conditions, MMVG addresses all prediction, rewind, and infilling tasks within one model. Experiments on various video scenes show that our MMVG effectively addresses TVC as well as generative video modeling. We believe TVC can help advance a new field toward vision-and-language research.", + "bbox": [ + 496, + 728, + 893, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10688", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards High Resolution Video Generation with Progressive Growing of Sliced Wasserstein GANs. In arXiv:1810.02419, 2018. 2", + "[2] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H. Campbell, and Sergey Levine. Stochastic Variational Video Prediction. In International Conference for Learning Representations (ICLR), 2018. 2", + "[3] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. FitVid: Overfitting in Pixel-Level Video Prediction. In arXiv:2106.13195, 2021. 2", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 7, 8", + "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2LIVE: Text-Driven Layered Image and Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2", + "[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5", + "[7] Ya-Liang Chang, Zhe Liu Yu, and Winston Hsu. VOR-Net: Spatio-temporally Consistent Video Inpainting for Object Removal. In Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2", + "[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent Online Video Style Transfer. In International Conference on Computer Vision (ICCV), 2017. 2", + "[9] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Jonathan Warrell, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy Mitra, and Philip Torr. ImageSpirit: Verbal Guided Image Parsing. In Conference on Human Factors in Computing Systems (CHI), 2013. 2", + "[10] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial Video Generation on Complex Datasets. In arXiv:1907.06571, 2019. 2", + "[11] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying Generalization in Reinforcement Learning. In International Conference on Machine Learning (ICML), 2019. 5", + "[12] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance. In European Conference on Computer Vision (ECCV), 2022. 2", + "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling Egocentric Vision: The EPIC-KITCHENS Dataset. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8", + "[14] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary Video Style" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transfer via Multi-Channel Correlation. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[15] Emily Denton and Rob Fergus. Stochastic Video Generation with a Learned Prior. In International Conference on Machine Learning (ICML), 2018. 1", + "[16] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In arXiv:2105.05233, 2021. 2", + "[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering Text-to-Image Generation via Transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2", + "[18] Frederik Ebert, Chelsea Finn, Alex X. Lee, and Sergey Levine. Self-Supervised Visual Planning with Temporal Skip Connections. In Conference on Robot Learning (CoRL), 2017. 2, 5, 6, 7, 8", + "[19] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W.Taylor. Tell, Draw, and Repeat: Generating and Modifying Images Based on Continual Linguistic Instruction. In International Conference on Computer Vision (ICCV), 2019. 2", + "[20] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. SSCR: Iterative Language-Based Image Editing via Self-Supervised Counterfactual Reasoning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 2", + "[21] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. M $^3$ L: Language-based Video Editing via Multi-Modal Multi-Level Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[22] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-Driven Artistic Style Transfer. In European Conference on Computer Vision (ECCV), 2022. 2", + "[23] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer. In European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7", + "[24] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2", + "[25] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martín, and Li Fei-Fei. MaskViT: Masked Visual Pre-Training for Video Prediction. In arXiv:2206.11894, 2022. 2, 6", + "[26] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine This! Scripts to Compositions to Videos. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8", + "[27] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10689", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zekun Hao, Xun Huang, and Serge Belongie. Controllable Video Generation with Sparse Trajectories. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1", + "[29] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Qiyuan Hu, and Devi Parikh. MUGEN: A Playground for Video-Audio-Text Multimodal Understanding and GENeration. In European Conference on Computer Vision (ECCV), 2022. 2, 5, 8", + "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6", + "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In International Conference on Computer Vision (ICCV), 2015. 5", + "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In arXiv:2210.02303, 2022. 2", + "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 7", + "[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In arXiv:2204.03458, 2022. 2, 6", + "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers. In arXiv:2205.15868, 2022. 2, 7", + "[36] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Learning to Decompose and Disentangle Representations for Video Prediction. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 1", + "[37] Qiyang Hu, Adrian Waelchli, Tiziano Portenier, Matthias Zwicker, and Paolo Favaro. Video Synthesis from a Single Image and Motion Stroke. In arXiv:1812.01874, 2018. 1", + "[38] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable Image-to-Video Generation with Text Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[39] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion Models for Video Prediction and Infilling. In arXiv:2206.07696, 2022. 1, 2, 5, 7", + "[40] Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video Pixel Networks. In International Conference on Machine Learning (ICML), 2017. 1", + "[41] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[42] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference for Learning Representations (ICLR), 2015. 5", + "[43] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video Generation From Text. In AAAI Conference on Artificial Intelligence (AAAI), 2018. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision (ICCV), 2021. 4", + "[45] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4", + "[46] Tanya Marwah, Gaurav Mittal, and Vineeth N. Balasubramanian. Attentive Semantic Video Generation using Captions. In International Conference on Computer Vision (ICCV), 2017. 2", + "[47] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep Multi-Scale Video Prediction Beyond Mean Square Error. In International Conference on Learning Representations (ICLR), 2018. 1", + "[48] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference for Learning Representations (ICLR), 2018. 6", + "[49] Mehdi Mirza and Simon Osindero. Conditional Generative Adversarial Nets. In arXiv:1411.1784, 2014. 2", + "[50] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In International Conference on Machine Learning (ICML), 2022. 2", + "[52] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To Create What You Tell: Generating Videos from Captions. In International Conference on Multimedia (MM), 2017. 2", + "[53] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In International Conference on Learning Representations Workshop (ICLRW), 2017. 6", + "[54] Björn Ommer Patrick Esser, Robin Rombach. Taming Transformers for High-Resolution Image Synthesis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[55] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. MirrorGAN: Learning Text-to-image Generation by Redescription. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Interna-" + ], + "bbox": [ + 501, + 90, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10690", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on Machine Learning (ICML), 2021. 4, 5", + "[57] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent Video Transformer. In arXiv:2006.10704, 2020. 2", + "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. In arXiv:2204.06125, 2022. 2", + "[59] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. In International Conference on Machine Learning (ICML), 2021. 2", + "[60] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. FILM: Frame Interpolation for Large Motion. In European Conference on Computer Vision (ECCV), 2022. 5, 6", + "[61] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. In International Conference on Machine Learning (ICML), 2016. 2", + "[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[63] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In arXiv:2205.11487, 2022. 2", + "[64] Masaki Saito and Shunta Saito. TGANv2: Efficient Training of Large Models for Video Generation with Multiple Subsampling Layers. In arXiv:1811.09245, 2018. 1", + "[65] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In arXiv:2111.02114, 2021. 2", + "[66] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Annual Meetings of the Association for Computational Linguistics (ACL), 2018. 2", + "[67] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations (ICLR), 2015. 3", + "[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. In arXiv:2209.14792, 2022. 2, 7", + "[69] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. In arXiv:1212.0402, 2012. 2, 5, 6, 7, 8", + "[70] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised Learning of Video Representations using" + ], + "bbox": [ + 78, + 92, + 467, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "LSTMs. In International Conference on Machine Learning (ICML), 2015. 1", + "[71] Fuwen Tan, Song Feng, and Vicente Ordonez. Text2Scene: Generating Compositional Scenes from Textual Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[72] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A Good Image Generator Is What You Need for High-Resolution Video Synthesis. In International Conference for Learning Representations (ICLR), 2021. 2", + "[73] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5, 6", + "[74] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing Motion and Content for Video Generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[75] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, and Sylvain Gelly Marcin Michalski. Towards Accurate Generative Models of Video: A New Metric & Challenges. In International Conference on Learning Representations Workshop (ICLRW), 2019. 5", + "[76] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural Discrete Representation Learning. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 3", + "[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 4", + "[78] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation From Open Domain Textual Description. In arXiv:2210.02399, 2022. 2", + "[79] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 6", + "[80] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating Videos with Scene Dynamics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 2", + "[81] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2", + "[82] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2", + "[83] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. VATEX: A Large-Scale," + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10691", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "High-Quality Multilingual Dataset for Video-and-Language Research. In International Conference on Computer Vision (ICCV), 2019. 2", + "[84] Dirk Weissenborn, Oscar Tackstrom, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference on Learning Representations (ICLR), 2020. 1", + "[85] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference for Learning Representations (ICLR), 2020. 2", + "[86] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating Open-Domain Videos from nAtural Descriptions. In arXiv:2104.14806, 2021. 2, 5, 7", + "[87] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. NUWA: Visual Synthesis Pretraining for Neural visUal World creAtion. In European Conference on Computer Vision (ECCV), 2022. 2, 7", + "[88] Xide Xia, Tianfan Xue, Wei-Sheng Lai, Zheng Sun, Abby Chang, Brian Kulis, and Jiawen Chen. Real-Time Localized Photorealistic Video Style Transfer. In Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7", + "[90] Qiangeng Xu, Hanwang Zhang, Weiyue Wang, Peter N. Belhumeur, and Ulrich Neumann. Stochastic Dynamics for Video Infilling. In Winter Conference on Applications of Computer Vision (WACV), 2020. 2", + "[91] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep Flow-Guided Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[92] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[93] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally Consistent Semantic Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2", + "[94] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers. In arXiv:2104.10157, 2021. 2, 6", + "[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. In arXiv:2206.10789, 2022. 2", + "[96] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks. In International Conference for Learning Representations (ICLR), 2022. 2, 6", + "[97] Jiangning Zhang, Chao Xu, Liang Liu, Mengmeng Wang, Xia Wu, Yong liu, and Yunliang Jiang. DTVNet: Dynamic" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Time-lapse Video Generation via Single Still Image. In European Conference on Computer Vision (ECCV), 2020. 1", + "bbox": [ + 532, + 90, + 890, + 119 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10692", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_model.json b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dae5705d9cac51ea88d1a95b8e8215f83dc4f0d6 --- /dev/null +++ b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_model.json @@ -0,0 +1,2721 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.819, + 0.175 + ], + "angle": 0, + "content": "Tell Me What Happened: Unifying Text-guided Video Completion via Multimodal Masked Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.203, + 0.724, + 0.256 + ], + "angle": 0, + "content": "Tsu-Jui Fu\\(^{1}\\), Licheng Yu\\(^{2}\\), Ning Zhang\\(^{2}\\), Cheng-Yang Fu\\(^{2}\\), Jong-Chyi Su\\(^{3}\\), William Yang Wang\\(^{1}\\), Sean Bell\\(^{2}\\) \n\\(^{1}\\)UC Santa Barbara \\(^{2}\\)Meta \\(^{3}\\)NEC Laboratories America" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.259, + 0.724, + 0.293 + ], + "angle": 0, + "content": "{tsu-juifu, william}@cs.ucsb.edu jcsu@nec-labs.com {lichengyu, ningzhang, chengyangfu, seanbell}@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.359, + 0.474, + 0.707 + ], + "angle": 0, + "content": "Generating a video given the first several static frames is challenging as it anticipates reasonable future frames with temporal coherence. Besides video prediction, the ability to rewind from the last frame or infilling between the head and tail is also crucial, but they have rarely been explored for video completion. Since there could be different outcomes from the hints of just a few frames, a system that can follow natural language to perform video completion may significantly improve controllability. Inspired by this, we introduce a novel task, text-guided video completion (TVC), which requests the model to generate a video from partial frames guided by an instruction. We then propose Multimodal Masked Video Generation (MMVG) to address this TVC task. During training, MMVG discretizes the video frames into visual tokens and masks most of them to perform video completion from any time point. At inference time, a single MMVG model can address all 3 cases of TVC, including video prediction, rewind, and infilling, by applying corresponding masking conditions. We evaluate MMVG in various video scenarios, including egocentric, animation, and gaming. Extensive experimental results indicate that MMVG is effective in generating high-quality visual appearances with text guidance for TVC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Generative video modeling [15, 70, 84] has made great progress, which first succeeds in unconditional video generation [40, 64]. More recently, video prediction [28, 36, 47] has been trying the controllable setting, which anticipates the future by completing a video from the past frames or a static starting image [37, 97]. However, video prediction may produce various outcomes, which makes it difficult to meet human expectations. For the example in Fig. 1(a), the game agent can keep jumping to the right or move back and turn left. The limited guidance from only the first frame is" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.325, + 0.88, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.437, + 0.786, + 0.449 + ], + "angle": 0, + "content": "(a) Video Prediction is Uncertain" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.467, + 0.601, + 0.504 + ], + "angle": 0, + "content": "Video Prediction put bowl in saucepan" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.515, + 0.603, + 0.561 + ], + "angle": 0, + "content": "Video Rewind Barney is trying to sing in a Living room." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.567, + 0.6, + 0.613 + ], + "angle": 0, + "content": "Video Infilling Runs to Left. Then collects a coin and a gem." + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.459, + 0.892, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.619, + 0.807, + 0.632 + ], + "angle": 0, + "content": "(b) Text-guided Video Completion (TVC)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.635, + 0.894, + 0.704 + ], + "angle": 0, + "content": "Figure 1. The introduced text-guided video completion (TVC) task. (a) Video prediction may have different outcomes without text guidance. (b) TVC performs video completion from the first frame (prediction), the last frame (rewind), or both (infilling), guided by the textual description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.854 + ], + "angle": 0, + "content": "insufficient to tell the intention. For humans, language is the most straightforward way of communication. If a system can follow an instruction to accomplish video completion, it will significantly improve its controllability and make a vast application impact. On the other hand, compared with video prediction, video rewind and infilling have been rarely studied [39, 79], but they are also crucial. Breaking the limitation of chronological guidance should make the visual guidance more flexible, leading to a general video completion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We thus introduce a novel task, text-guided video completion (TVC), where the partial frames and a given instruction jointly guide the video generation. As illustrated in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Fig. 1(b), we consider three scenarios of video completion: prediction from the first frame, rewind from the last frame, and infilling between the head and tail. The missing (to-be-completed) event should follow the textual instruction. Compared to generating content from scratch [43, 87], TVC requests models to understand the given visual and textual guidance before generation, which better mimics how human imagines after seeing and listening in our daily lives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.471, + 0.517 + ], + "angle": 0, + "content": "To tackle TVC, we present Multimodal Masked Video Generation (MMVG) to perform video completion. Specifically, we represent the video frames as discrete visual tokens by temporal-aware VQGAN [54, 76]. One key challenge is to deal with the video frames that are not presented in chronological (e.g., the last frame for rewind). Different from autoregressive models [23, 86] that only condition on the previous frames, MMVG carries out video completion in an encoder-decoder manner. Specifically, we propose a masking strategy that masks different parts of the video and feeds them as the input to the multimodal encoder with the instruction. As shown in Fig. 2, we allow MMVG to consider the visual hints from different time points, and the decoder learns to produce the full target video. By varying the masking conditions (including the cases of only the first or last frame being accessible), a single MMVG can address all TVC tasks, including video prediction, rewind, and infilling. Moreover, learning the recovery from partial frames also empowers MMVG with a strong temporal coherence, contributing to better generative video quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.684 + ], + "angle": 0, + "content": "We consider videos in diverse scenarios for the TVC evaluation. There are Kitchen [13], Flintstones [26], and MUGEN [29] corresponding to the egocentric, animation, and gaming scenes. The model should generate videos such as performing kitchen activities in the first-person view, making characters act the assigned behavior, or imitating an agent playing game. All should be guided with the first/last (or both) frame(s) and controlled through the given human instructions. We also compare MMVG with previous methods [23,57,94,96] on UCF-101 [69] and BAIR [18] for the classic video generation/prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Experimental results demonstrate that instruction is necessary to make video completion controllable, MMVG can address all three TVC tasks, and our proposed masking strategy enhances the temporal modeling, which further benefits general video generation/prediction. In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.78, + 0.468, + 0.839 + ], + "angle": 0, + "content": "- We introduce TVC to generate a video from partial frames and control the temporal dynamics via natural language, where our video completion includes 3 cases: prediction, rewind, and infilling." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.468, + 0.87 + ], + "angle": 0, + "content": "- We propose MMVG with an effective masking strategy to address all TVC tasks through a single training." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Extensive experiments show that our MMVG can handle various types of video completion as well as video gener" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "ation/prediction. We believe TVC can become a new topic in vision-and-language research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.138, + 0.642, + 0.154 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.164, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Video Generation/Prediction. Video generation aims to synthesize diverse videos from latent inputs [1, 74, 80]. Various generative modelings have shown promising results, including generative adversarial networks (GAN) [10, 24, 72, 96], autoregressive transformers [23, 77, 94], and denoising diffusion models [16, 33, 34]. Upon that, video prediction [2, 3, 25, 57, 85], which considers past frames to anticipate future observations, should maintain temporal dynamics from static images. Though the overall idea is also to complete a video from partial frames, other tasks, such as rewind and infilling [39, 79, 90], are not extensively explored. In this paper, we introduce TVC to comprehensively investigate the ability of video completion and make it more maneuverable via textual description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Text-to-Image/Video Genration. Generating visual content from language [9, 50, 71] has a vast application value in creative visual design. Previous works rely on the GANs framework [49] to produce images [19, 20, 22, 55, 61, 92] or videos [43, 46, 52], conditioned on text. With large-scale datasets [4, 65, 66, 83], recent pre-trained models can generate high-quality natural images from open-domain description through discrete visual tokens [12, 17, 54, 59, 76, 95] or the diffusion process [51, 58, 62, 63]. Leveraging such techniques further extends to generate vivid videos [32, 35, 68, 78, 86, 87]. However, those methods that depend on autoregressive generation can only be guided chronologically [27, 38]. Besides, video diffusion models require a deterministic video length, which cannot consider diverse temporal durations. In contrast, MMVG can perform video completion in different lengths from arbitrary time points and address all TVC tasks just with a single training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Text-guided Video-to-Video. Video inpainting [7, 41, 91], segmentation reconstruction [81, 82], or video style transfer [8, 14, 88] can be seen as a particular case of video-to-video synthesis (V2V). Even if text-guided V2V [5, 21, 93] can be controlled by language, it is still conditioned on a full video, where the temporal dynamics are usually provided. Different from that, TVC requires to regain the missing event from just partial guidance. It is more challenging since the model has to capture what happened from the instruction, maintain the temporal coherence among limited frames, and produce a complete video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.833, + 0.838 + ], + "angle": 0, + "content": "3. Text-guided Video Completion (TVC)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.655, + 0.862 + ], + "angle": 0, + "content": "3.1. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We study the text-guided video completion (TVC) task to perform video completion from the first frame (prediction)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.896, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Figure 2. An overview of our Multimodal Masked Video Generation (MMVG). We present temporal-aware VQGAN (T-VQ) for discrete visual representation. MMVG considers the instruction \\(\\mathcal{X}\\) and partial frames of the video \\(\\nu\\) from diverse time points through masking, learning to generate the complete \\(\\nu\\). In this way, a single trained MMVG can perform all prediction, rewind, and infilling tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.471, + 0.51 + ], + "angle": 0, + "content": "the last frame (rewind), or the head and tail (infilling), conditioned on the textual instruction. During training, we have pairs of videos \\(\\mathcal{V}\\) and corresponding instructions \\(\\mathcal{X}\\). Specifically, \\(\\mathcal{V}\\) consists of \\(N\\) frames as \\(\\{v_{1}, v_{2}, \\ldots, v_{N}\\}\\). Our goal is to train a unified model that generates the complete \\(\\mathcal{V}\\) given the partial frames from arbitrary time points and \\(\\mathcal{X}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.524, + 0.411, + 0.539 + ], + "angle": 0, + "content": "3.2. Multimodal Masked Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.761 + ], + "angle": 0, + "content": "Overview. An overview of our Multimodal Masked Video Generation (MMVG) is illustrated in Fig. 2. To model the video along with language, we propose temporal-aware VQGAN to represent a frame as visual tokens, which converts it into the same discrete space as the words. We present an effective masking strategy that masks different video parts for video completion learning. Those missing fragments are replaced by the unique [SPAN] tokens, and we consider the visual guidance from diverse time points. The multimodal encoder consumes the text and the partial-missing video, and the decoder learns to produce the complete video from arbitrary guided frames. By varying the masking conditions, MMVG learns to utilize the [SPAN] token and unifies all TVC tasks during the training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Temporal-aware Discrete Visual Tokens. VQ-VAE [76] has shown promising capability in representing data as discrete tokens. VQGAN [54] further models the prior distribution of the latent space via a transformer with the GAN training. If VQGAN is directly applied onto videos, it will ignore the inner temporal coherence and treat each frame as an independent image, resulting in an unsmooth video reconstruction. Though TATS [23] attempts to handle this by making \\( k \\) consecutive frames altogether during VQ, it has" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.45 + ], + "angle": 0, + "content": "to pre-define the constant \\( k \\) before training. Such constraint forbids it from representing a frame at any timestamp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.51 + ], + "angle": 0, + "content": "To address it with flexibility, we propose temporal-aware VQGAN (T-VQ) to inject the temporal relationship into the latent representation. We first follow VQGAN to learn the target visual tokens \\( z_{i} \\) by reconstructing a video frame \\( v_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.517, + 0.892, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {i} = \\mathbf {q} (\\mathsf {E n c} ^ {\\mathsf {Q}} (v _ {i}) \\mid C), \\\\ \\hat {v} _ {i} = \\operatorname {D e c} ^ {0} (\\mathrm {z} _ {i}), \\\\ \\mathcal {L} _ {\\mathrm {V Q}} = \\underbrace {\\left| \\left| \\hat {v _ {i}} - v _ {i} \\right| \\right| _ {1}} _ {\\text {r e c o n s t r c u t i o n}} + \\underbrace {\\left| \\left| \\operatorname {s g} [ \\operatorname {E n c} ^ {Q} (v _ {i}) ] - C _ {z _ {i}} \\right| \\right| _ {2} ^ {2}} _ {\\text {c o d e b o o k}} \\tag {1} \\\\ + \\underbrace {\\beta | | \\mathrm {s g} [ C _ {\\mathrm {z} _ {i}} ] - \\mathrm {E n c} ^ {\\mathrm {Q}} (v _ {i}) | | _ {2} ^ {2}} _ {\\text {c o m m i t}} + \\underbrace {| | \\mathcal {F} (\\hat {v} _ {i}) - \\mathcal {F} (v _ {i}) | | _ {1}} _ {\\text {m a t c h i n g}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathrm{Enc}^{\\mathrm{Q}}\\) and \\(\\mathrm{Dec}^{\\mathrm{Q}}\\) are the VQ encoder and decoder. The discrete latent code \\(z_{i}\\) is acquired from the quantization operation q [54], which adopts nearest neighbor search by the trainable codebook \\(C\\). We apply the straight-through estimator over the stop-gradient operation sg and use \\(\\beta\\) as 0.25 [76]. We also append VGG [67] features matching \\(\\mathcal{F}\\) to stabilize the VQ loss \\(\\mathcal{L}_{\\mathrm{VQ}}\\) [23]. The adversarial training between the frame quality loss \\(\\mathcal{L}_G\\) and discrimination loss \\(\\mathcal{L}_D\\) are further calculated from the discriminator \\(D\\):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.787, + 0.892, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {G} = \\log (1 - D (\\hat {v} _ {i})), \\\\ \\mathcal {L} _ {D} = \\log (1 - D (\\hat {v} _ {i})) + \\log (D (v _ {i})). \\end{array} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.829, + 0.892, + 0.86 + ], + "angle": 0, + "content": "To inject the temporal relationship into \\( z \\), T-VQ is trained with the introduced contrastive temporal reasoning:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.867, + 0.892, + 0.892 + ], + "angle": 0, + "content": "\\[\no _ {i} = \\mathbf {F C} ^ {\\mathrm {T}} \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.89, + 0.82, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T}} = \\operatorname {B C E L o s s} \\left(o _ {i}, 0 \\text {i f} i > j \\text {e l s e} 1\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.183 + ], + "angle": 0, + "content": "where \\( j \\) is a random frame from the same video. \\( \\mathrm{FC^T} \\) is the MLP classifier, and BCELoss is the binary cross-entropy for before/after. Learning the temporal order from \\( \\mathcal{L}_{\\mathrm{T}} \\), \\( \\mathbf{Z} \\) facilitates an implicit temporal coherence, leading to smooth video modeling. Moreover, since \\( \\mathbf{Z} \\) represents a single image, it is flexible to support frames at arbitrary timestamps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.471, + 0.276 + ], + "angle": 0, + "content": "Generation from Masked Video. We propose the masking strategy \\(\\mathcal{M}\\) to obtain the masked videos \\(\\overline{\\nu}\\) from diverse time points. \\(\\mathcal{M}\\) masks out most video frames with the probability \\(p\\) and replaces each fragment as a unique [SPAN] token. For example, \\(\\mathcal{M}\\) reserves the third and the fifth frame, and masks all the others over a video length of 5:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.287, + 0.47, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {V}}} \\colon \\{[ S ], v _ {3}, [ S ], v _ {5} \\} = \\mathcal {M} (\\mathcal {V} \\mid p). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.471, + 0.453 + ], + "angle": 0, + "content": "Our goal is to recover the missing part from \\(\\overline{\\mathcal{V}}\\) and perform video completion, guided by the instruction \\(\\mathcal{X}\\). To model between the vision and language modalities, we apply our Enc\\(^{\\mathrm{Q}}\\) over \\(\\overline{\\mathcal{V}}\\) for the discrete visual tokens \\(\\{[\\mathrm{S}],\\mathrm{z}_3,[\\mathrm{S}],\\mathrm{z}_5\\}\\). We also tokenize the text \\(\\mathcal{X}\\) into word tokens \\(\\{w_i\\}_{i = 1}^L\\) with the CLIP tokenizer [56], where \\(L\\) is the length of \\(\\mathcal{X}\\). As in the same discrete space, MMVG can achieve cross-modal fusion by the multimodal encoder (Enc\\(^{\\mathrm{M}}\\)) through the self-attention mechanism as the transformer [77]:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.464, + 0.469, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {i} ^ {w}, f _ {j} ^ {v} = \\mathrm {L P} ^ {w} \\left(w _ {i}\\right), \\mathrm {L P} ^ {v} \\left(\\mathrm {z} _ {j}\\right) \\\\ \\{1 \\} = \\sum_ {i = 1} ^ {n} M (f _ {i}) w _ {i} \\end{array} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.484, + 0.379, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\{h \\} = \\operatorname {E n c} ^ {\\mathrm {M}} ([ \\{f ^ {w} \\}, \\{f ^ {v} \\} ]),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.469, + 0.573 + ], + "angle": 0, + "content": "where it obtains the features \\( f \\) by the linear projection (LP), and \\( h \\) is the hidden encoding features. We can also regard LP as the video/language embedder, which extracts the preliminary visual/linguistic features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.469, + 0.678 + ], + "angle": 0, + "content": "After encoding the language hint and the partial-missing video from Enc\\(^{\\text{M}}\\), our video decoder (Dec\\(^{\\text{M}}\\)) learns to produce all frames for comprising the complete video. Dec\\(^{\\text{M}}\\) follows the vanilla autoregressive decoder, which first conducts self-attention over the past generated tokens and then predicts the discrete visual tokens as the video frame, conditioned on the encoded features \\( h \\):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.688, + 0.388, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t} = \\operatorname {D e c} ^ {\\mathrm {M}} (\\left\\{\\hat {z} _ {1}, \\dots , \\hat {z} _ {t - 1} \\right\\} \\mid \\{h \\}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.709, + 0.469, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\underset {N} {\\operatorname {C E L o s s}} \\left(\\hat {z} _ {t}, z _ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.73, + 0.254, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M}} = \\sum_ {t = 1} ^ {N} \\mathcal {L} _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\( \\mathbf{z}_t \\) is the ground-truth tokens of the frame \\( v_t \\) in the original \\( \\mathcal{V} \\). We calculate the video decoding loss \\( \\mathcal{L}_{\\mathrm{M}} \\) by the cross-entropy (CELoss) to learn video generation as classification. Our \\( \\mathrm{Dec}^{\\mathrm{M}} \\) is built upon VideoSwin [45], which has shown a strong visual perception on various video understanding tasks. The 3D-shifted windows [44] consider different levels of spatial-temporal attention, and each window models blocks of video patches across \\( T' \\) consecutive" + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.091, + 0.847, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Multimodal Masked Video Generation" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.108, + 0.892, + 0.47 + ], + "angle": 0, + "content": "1: while TRAIN_T-VQ do \n2: \\(\\mathcal{V}\\gets\\) sample video \n3: \\(z_{i} = q(\\mathrm{Enc}^{Q}(v_{i})\\mid C)\\) \n4: \\(\\hat{v}_i = \\mathrm{Dec}^{\\mathrm{Q}}(z_i)\\) \n5: \\(o_i = \\mathrm{FC}^{\\mathrm{T}}(z_i,z_j)\\) ▷ randomly sampled frame \\(j\\) \n6: \\(\\mathcal{L}_{\\mathrm{VQ}},\\mathcal{L}_G\\gets\\) reconstruction, frame quality loss \\(\\triangleright\\) Eq. 1&2 \n7: \\(\\mathcal{L}_{\\mathrm{T}}\\gets\\) temporal ordering loss \\(\\triangleright\\) Eq. 3 \n8: Update T-VQ by minimizing \\(\\mathcal{L}_{\\mathrm{VQ}} + \\mathcal{L}_G + \\mathcal{L}_{\\mathrm{T}}\\) \n9: \\(\\mathcal{L}_D\\gets\\) discrimination loss \\(\\triangleright\\) Eq. 2 \n10: Update D by maximizing \\(\\mathcal{L}_D\\) \n11: end while \n12: \n13: while TRAIN_MMVG do \n14: \\(\\mathcal{V},\\mathcal{X},p\\gets\\) sample video/instruction/probability \n15: \\(\\overline{\\mathcal{V}}:\\{v_a,[S],v_b,\\ldots\\} = \\mathcal{M}(\\mathcal{V}\\mid p)\\) \\(\\triangleright\\) diverse guided frames \n16: \\(\\{\\mathbf{z}_a,[S],\\mathbf{z}_b,\\ldots\\},\\{w\\} = \\mathrm{Enc}^{\\mathrm{Q}}(\\overline{\\mathcal{V}}),\\) Tokenizer(X) \n17: \\(\\{h\\} = \\mathrm{Enc}^{\\mathrm{M}}([\\{w\\},\\{z_a,[S],z_b,\\ldots\\}])\\) \n18: for \\(t\\gets 1\\) to \\(N\\) do \n19: \\(\\hat{z}_t = \\mathrm{Dec}^{\\mathrm{M}}(\\{z_1,\\dots,z_{t-1}\\}\\mid\\{h\\})\\) \\(\\triangleright\\) teacher-forcing \n20: \\(\\mathcal{L}_t\\gets\\) video decoding loss \\(\\triangleright\\) Eq. 6 \n21: end for \n22: \\(\\hat{\\nu} = \\mathrm{Dec}^{\\mathrm{Q}}(\\{\\hat{z}_{t=1}^N\\})\\) \n23: \\(\\mathcal{L}_{\\mathrm{M}} = \\sum_{t=1}^{N}\\mathcal{L}_{t}\\) \n24: Update MMVG by minimizing \\(\\mathcal{L}_{\\mathrm{M}}\\) \n25: \\(p\\gets\\) update masking probability \\(\\triangleright\\) Eq. 8 \n26: end while" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.556 + ], + "angle": 0, + "content": "frames. To ensure the same dimension for video generation in \\(\\mathrm{Dec}^{\\mathrm{M}}\\), we remove the temporal down-sampling layer. In the end, we can utilize \\(\\mathrm{Dec}^{\\mathrm{Q}}\\) to reconstruct all the frames as our completed videos \\(\\hat{\\mathcal{V}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.565, + 0.892, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {V}} = \\operatorname {D e c} ^ {\\mathrm {Q}} \\left(\\left\\{\\hat {z} \\right\\} _ {t = 1} ^ {N}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.729 + ], + "angle": 0, + "content": "By varying the masking conditions through \\(\\mathcal{M}\\), MMVG learns how to complete a video from partial frames \\(\\overline{\\nu}\\) at arbitrary time points with the text, which overcomes the limitation of chronological guidance. To make \\(\\mathcal{M}\\) more effective, we apply an adaptive probability \\(p\\) instead of random sampling every time. Each video \\(\\mathcal{V}\\) keeps its own \\(p\\), and all frames are equally initialized in the beginning. Based on the prediction error, we adjust the masking probability \\(p_t\\) of the \\(t\\)-th frame:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.737, + 0.892, + 0.769 + ], + "angle": 0, + "content": "\\[\np _ {t} = p _ {t} + \\alpha \\left(\\left(\\frac {\\mathcal {L} _ {t}}{\\mathcal {L} _ {M}} \\sum p\\right) - p _ {t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.892, + 0.838 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the adjusting rate. A larger video decoding loss \\(\\mathcal{L}_t\\) indicates that the \\(t\\)-th frame is more difficult to recover. MMVG learns more from those challenging cases and can bring better generative quality for video completion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unifying TVC during Inference. After training with text and partial-missing video, MMVG learns to perform video completion over [SPAN] tokens. Then for inference, Enc\\(^{\\text{M}}\\) takes the following as its input to support different tasks:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10684" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.248 + ], + "angle": 0, + "content": "
MethodTextTVPredictionTVRewindTVInfilling
KitchenFlintstonesMUGENKitchenFlintstonesMUGENKitchenFlintstonesMUGEN
FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑
FILM [60]✘------------250.256.1352.751.4538.85.9
VideoMAE [73]✘328.947.6317.555.6548.77.0365.948.2335.555.9545.27.1246.954.7211.560.6494.97.8
TATS [23]✘106.964.4127.560.3376.57.1107.762.7127.660.2350.87.271.572.7119.566.7328.28.4
MMVGU✘105.663.3124.860.5374.57.2109.862.6124.359.7356.47.071.573.4121.866.3328.47.8
MMVGS✘103.864.5123.860.8369.47.3105.963.6123.860.5347.87.268.573.6118.567.9324.38.4
TATS [23]✓87.266.3115.970.690.167.989.863.3116.370.489.868.757.477.695.878.258.973.6
MMVGU✓80.268.4108.272.984.870.283.266.9113.271.693.168.459.877.892.878.359.273.2
MMVGS✓75.668.8106.373.783.371.179.768.1107.272.988.770.056.078.191.679.657.274.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Table 1. Results of TVC, including prediction, rewind, and infilling, on Kitchen [13], Flintstones [26], and MUGEN [29]. TATS [23] requires specific training to support different tasks. We further train the unified MMVG\\(^{\\text{U}}\\) for each specific task as MMVG\\(^{\\text{S}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.292, + 0.329, + 0.308 + ], + "angle": 0, + "content": "- TVPrediction: \\(\\left[\\left\\{w\\right\\},\\left\\{z_{1},[SPAN]\\right\\}\\right]\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.308, + 0.316, + 0.323 + ], + "angle": 0, + "content": "- TVRewind: \\(\\left[\\left\\{\\boldsymbol{w}\\right\\},\\left[\\left[\\mathrm{SPAN}\\right],\\mathbf{z}_{N}\\right]\\right]\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.323, + 0.343, + 0.338 + ], + "angle": 0, + "content": "TVInfilling: \\(\\left[\\left\\{w\\right\\} ,\\left\\{\\mathbf{z}_1,\\left[\\mathrm{SPAN}\\right],\\mathbf{z}_N\\right\\}\\right]\\)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.292, + 0.343, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.468, + 0.369 + ], + "angle": 0, + "content": "In this way, a single trained MMVG can unify all TVC tasks without the specific downstream fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.27, + 0.391 + ], + "angle": 0, + "content": "3.3. Learning of MMVG" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.701 + ], + "angle": 0, + "content": "Algo. 1 illustrates the learning process of our proposed MMVG for TVC. We first train T-VQ over video frames for discrete visual tokens with contrastive temporal reasoning. Specifically, we minimize the VQ reconstruction loss \\(\\mathcal{L}_{\\mathrm{VQ}}\\) and frame quality loss \\(\\mathcal{L}_G\\) along with our temporal ordering loss \\(\\mathcal{L}_{\\mathrm{T}}\\) to optimize T-VQ. At the same time, we also update the discriminator \\(D\\) via the standard adversarial training by maximizing the discrimination loss \\(\\mathcal{L}_D\\). For video completion, the masking strategy \\(\\mathcal{M}\\) masks the video frames with the probability \\(p\\) and then acquires guided frames from diverse time points. MMVG regards text and partial-missing video by Enc for cross-modal fusion, and Dec further predicts the visual tokens of frames autoregressively as the complete video. As a sequential generation process, we apply the teacher-forcing trick. Instead of our predicted \\(\\hat{z}\\), the ground-truth \\(z\\) from the previous timestamp is fed to stabilize the training. Each video decoding loss \\(\\mathcal{L}_t\\) at timestamp \\(t\\) is summed up as \\(\\mathcal{L}_{\\mathrm{M}}\\) to optimize MMVG. According to \\(\\mathcal{L}_t\\), we update \\(p\\) for effective masking probability. The entire optimization object can be summarized as two phases:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.706, + 0.435, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {T - V Q}: \\min _ {\\mathrm {E n c} ^ {\\mathrm {Q}}, \\mathrm {D e c} ^ {\\mathrm {Q}}, C, \\mathrm {F C} ^ {\\mathrm {T}}} \\max _ {D} \\mathcal {L} _ {\\mathrm {V Q}} + \\mathcal {L} _ {G} + \\mathcal {L} _ {D} + \\mathcal {L} _ {\\mathrm {T}} \\\\ \\text{MMVG:}\\min_{\\substack{\\text{Enc}^{\\text{M}},\\text{Dec}^{\\text{M}}}}\\mathcal{L}_{\\text{M}} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.21, + 0.78 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.267, + 0.804 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Datasets. As a new task, we consider diverse video scenes with natural instructions for TVC. Kitchen [13] records 22K egocentric videos about kitchen activity, which have different lengths (4-16 frames) with narrations. Flintstones [26] contains 25K animation videos (15 frames) from The Flintstones, where each video description includes the characters" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.29, + 0.851, + 0.359 + ], + "angle": 0, + "content": "
DatasetTrain / Val#Frame#WordFPS
Kitchen [13]16,695 / 5,8048.32.86
Flintstones [26]22,666 / 2,5181516.55
MUGEN [29]362,239 / 12,8481620.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.361, + 0.868, + 0.374 + ], + "angle": 0, + "content": "Table 2. The statistics of our used datsets to evaluate TVC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.892, + 0.526 + ], + "angle": 0, + "content": "and their behavior. MUGEN [29] is built from agents playing CoinRun [11], which consists of 375K gaming videos (16 frames) with detailed text annotations. All videos in these three datasets are resized into \\(128 \\times 128\\). An overview is shown in Table 2 and Fig. 1(b). Since MMVG can unify various tasks, we also evaluate video generation/prediction on widely-used UCF-101 [69] and BAIR [18], video infilling on UCF-101 following RaMViD [39], and text-to-video generation on MSRVTT [89]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.893, + 0.726 + ], + "angle": 0, + "content": "Evaluation Metrics. We apply the following metrics to evaluate TVC results: 1) FVD [75] computes the video features [6] distance to the ground truth; 2) RCS [86] is the relative visual-text similarity to the instruction, compared to the ground-truth video. We fine-tune the CLIP model [56] on each dataset and adapt it to the video scene for a more precise alignment. Apart from automatic metrics, we also conduct a human evaluation from aspects of video quality, instruction relevance, and ground-truth similarity. We sample 75 TVP results for each task and adopt MTurk1 to rank over baselines and our MMVG. To avoid the potential ranking bias, we hire 3 MTurkers for each sample of prediction, rewind, and infilling tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.894, + 0.865 + ], + "angle": 0, + "content": "Implementation Detail. T-VQ contains ResBlocks [31] as the visual auto-encoder (Enc\\(^{\\text{Q}}\\) and Dec\\(^{\\text{Q}}\\)). The discriminator \\(D\\) follows a similar architecture to Enc\\(^{\\text{Q}}\\). For the vector quantization, we use a patch size 16, where a 128x128 video frame transforms into 8x8 discrete visual tokens. There are 1024 vocabularies in the codebook \\(C\\), and the hidden embedding size is 256. We adopt batch size 32 with a learning rate of 4.5e-6 to optimize T-VQ by Adam [42]. MMVG is built in an encoder-decoder manner, where Enc\\(^{\\text{M}}\\) is a trans" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Amazon MTurk: https://www.mturk.com. Our studies have been cleared by the human subject committee as an IRB-exempt protocol." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10685" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.089, + 0.406, + 0.171 + ], + "angle": 0, + "content": "
MethodKitchenFlintstonesMUGEN
VideoDiff [34]138.6206.4410.7
MCVD [79]119.9183.8400.2
TATS [23]115.5157.5386.4
MMVG109.1127.6368.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.173, + 0.462, + 0.187 + ], + "angle": 0, + "content": "Table 3. FVD results of video generation on our TVC datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.201, + 0.47, + 0.321 + ], + "angle": 0, + "content": "former with 24 layers, 16 attention heads, and hidden embedding size 1024. \\(\\mathrm{Dec}^{\\mathrm{M}}\\) adopts a similar setting with temporal window size 3 in VideoSwin. The initial sample rate \\(p\\) of the masking strategy \\(\\mathcal{M}\\) is 0.9 with an adjusting rate \\(\\alpha\\) as 0.1. We optimize MMVG through the mixed precision [48] with batch size 4 by Adam. The learning rate is also 4.5e-6. All experiments are implemented in PyTorch [53] and done on 8 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.216, + 0.346 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Table 1 shows the results of all text-guided prediction, rewind, and infilling for TVC. VideoMAE [73] is built upon MAE [30] and reconstructs the missing video cubes, which performs TVC by masking all video frames except the first or the last (or both). TATS [23], the SOTA on video generation, also produces videos as discrete visual tokens. Since TATS can only consider the past through the autoregressive transformer, it requires specific training for each task. We have MMVG\\(^{\\text{U}}\\) as the unified model that can support all TVC tasks with a single training and MMVG\\(^{\\text{S}}\\) to further train for each prediction, rewind, and infilling. We treat TATS as our main baseline\\(^{2}\\) and study the importance of guided text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.842 + ], + "angle": 0, + "content": "TVPrediction. VideoMAE attempts to produce all frames simultaneously, which is difficult to maintain video temporal consistency, resulting in a high 328.9 FVD on Kitchen. TATS is inherently designed for prediction as it generates the frames one after one. However, our unified MMVGU performs better than TATS on all datasets (e.g., lower 105.6 and 124.8 FVD on Kitchen and Flintstones). These results support that learning from diverse time points will not hurt the prediction from the past. In contrast, our masking strategy can bring superior temporal coherence. MMVGS further improves itself through training prediction as completion from the head. However, there are too many possible outcomes from just the beginning, where the predicted results may not meet the expectation (e.g., a high 370 FVD on MUGEN). The instruction as guidance makes it related to the expected ground-truth result. We can let MUGEN run, jump, or collect coins as the textual descriptions to achieve more controllability, leading to a noticeable improvement (e.g., a lower 84.8 FVD by MMVGU). The higher 70.2 RCS also shows that our MMVG can produce MUGEN videos" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.089, + 0.693, + 0.203 + ], + "angle": 0, + "content": "
MethodUCF-101
IS↑FVD↓
VideoGPT [94]24.7-
DIGAN [96]32.7577
VideoDiff [34]57.0-
TATS [23]57.6420
MMVG58.3395
" + }, + { + "type": "table", + "bbox": [ + 0.746, + 0.089, + 0.882, + 0.203 + ], + "angle": 0, + "content": "
MethodBAIR
FVD↓
VideoGPT [94]103.3
MaskViT [25]93.6
MCVD [79]89.5
TATS [23]88.6
MMVG85.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.73, + 0.205, + 0.897, + 0.233 + ], + "angle": 0, + "content": "Table 5. Results of video prediction on BAIR [18]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.206, + 0.726, + 0.233 + ], + "angle": 0, + "content": "Table 4. Results of video generation on UCF-101 [69]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.334 + ], + "angle": 0, + "content": "that confirm with the instruction. Although the model may try to imagine the animation or the kitchen activity, the language hint can provide a clear goal to anticipate. Likewise, MMVG\\(^{\\text{U}}\\) with text surpasses TATS, even though it is not designed for prediction only. The specific trained MMVG\\(^{\\text{S}}\\) benefits the unified model for further improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.893, + 0.579 + ], + "angle": 0, + "content": "TVRewind. Rewind from the last allows the model to imagine what happened along with a suitable opening. In addition, the objects may not display on the last frame (e.g., the spoons and forks for \"close drawer\"), which makes it more challenging to complete. Similar to prediction, VideoMAE cannot have feasible rewind results. Language is still essential to remind the past and establish an adequate beginning, where we can find a significant performance gap between with and without text (e.g., 90 vs. 350 FVD on MUGEN). Our unified MMVG\\(^{\\text{U}}\\) achieves comparable results to TATS and even outperforms on Kitchen and Flintstones (e.g., higher 66.9 and 71.6 RCS). With the learning of completion from partial frames, autoregressive model can still accomplish video rewind without specific training. If following TATS design to train MMVG\\(^{\\text{U}}\\) for rewind, MMVGS gains more improvement and utterly surpasses it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.793 + ], + "angle": 0, + "content": "TVInfilling. We consider the additional FILM [60] for infilling, which performs video interpolation with in-between motion. Despite synthesizing intermediate frames between the first and the last, the visual dynamics are changing too rapidly to handle, resulting in a higher FVD. With guidance from the head and tail, we find a noticeable improvement even without instruction (e.g., lower FVDs on Kitchen), which is helpful in temporal video modeling. To capture the expected missing event, we still require the language hint for more controllability. Our unified MMVG\\(^{\\text{U}}\\) achieves comparable performance to TATS again, which is specifically trained for the infilling task. It shows that completion from partial frames at different time points still helps, and MMVG\\(^{\\text{S}}\\) further outperforms on TVInfilling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.669, + 0.818 + ], + "angle": 0, + "content": "4.3. Additional Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Video Generation/Prediction. We further evaluate the classic video generation and prediction tasks. Table 3 shows FVD scores of unconditional video generation on our TVC datasets. Note that only videos but no texts are used in these experiments. Both VideoDiff [34] and MCVD [79] are built" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2Since we cannot receive feasible results after training diffusion methods for our TVC, we evaluate unconditional video generation in Sec. 4.3. We use this repo (https://github.com/lucidrains/video-diffusion-pytorch) as VideoDiff and the official repo as MCVD." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10686" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.139, + 0.089, + 0.406, + 0.166 + ], + "angle": 0, + "content": "
MethodUCF-101 (FVD↓)
K=+1+2+5±1±2
RaMViD [39]349.7300.6260.5215.4162.5
MMVG316.3258.5194.6183.2120.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.166, + 0.462, + 0.18 + ], + "angle": 0, + "content": "Table 6. Results of video prediction and infilling on UCF-101." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.196, + 0.468, + 0.317 + ], + "angle": 0, + "content": "upon denoising diffusion [33], where MCVD also considers different partial frames during training. The results first indicate that the vanilla token-based method is superior to the diffusion-based model (TATS vs. VideoDiff) for video generation. In addition, MMVG, with the masking strategy that learns the visual guidance from diverse time points, further boosts the performance (the lowest 127.6, 109.1, and 368.6 FVD on Flintstones, Kitchen, and MUGEN, respectively)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.468, + 0.515 + ], + "angle": 0, + "content": "We also evaluate MMVG on UCF-101 [69], which is challenging to generate natural human videos. Table 4 supports that our MMVG can produce videos with higher visual similarity (a higher 58.3 IS) and temporal alignment (a lower 395 FVD) to the ground truth. For video prediction, we apply the widely-used BAIR [18] in Table 5, where the model has to anticipate how a robot pushes objects from the given first frame. MMVG again surpasses TATS with the lowest 85.2 FVD. Although both generation and prediction are generating video frames chronologically, the ability to recover arbitrary missing frames for video completion empowers MMVG with a stronger temporal coherence, leading to better generative video quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.731 + ], + "angle": 0, + "content": "Video Infilling. We follow RaMViD [39] to evaluate video infilling on UCF-101. We consider various guidance settings \\( K \\) in Table 6. For example, \\( K = +1 \\) means given the first frame, and \\( K = \\pm 2 \\) should provide the first and last two frames. For prediction, MMVG outperforms RaMViD on all \\( K \\), and the performance gap gets even larger when more guided frames are accessible (e.g., 33.4 on \\( K = +1 \\) and 65.9 on \\( K = +5 \\)). A similar result can be found for infilling, where MMVG can make the lowest 120.3 FVD on \\( K = \\pm 2 \\). Despite having a similar masking strategy, it shows that generating frames one after one still brings superior results. MMVG allows autoregressive models to condition on visual hints from any time point, which produces more similar videos to the ground truth when infilling between the head and tail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Text-to-Video Generation. Being a multimodal generative model, MMVG supports text-to-video generation. To compare with those large-scale methods, we pre-train MMVG using WebVid [4], which contains 2.5M text-video pairs. We adopt the masking strategy to treat the pre-training as video completion. MMVG outperforms GODIVA [86] and NUWA [87] without access the MSRVTT [89] data in Table 7. Surprisingly, MMVG can generate videos that are more related to the texts (a higher 0.2644 CLIP-S [86]) thanCogVideo [35], even though using twice less data. This result encourages the effectiveness of completion from partial" + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.089, + 0.858, + 0.235 + ], + "angle": 0, + "content": "
MethodPre-trainingZero-shotMSRVTT
FID↓CLIP-S↑
GODIVA [86]136MX-0.2402
NUWA [87]3.9MX47.70.2439
CogVideo [35]5.4M✓23.60.2631
Make-A-Video [68]20M✓13.20.3049
MMVG2.5M✓23.40.2644
TATS [23]-X63.20.2326
MMVG-X60.60.2385
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.238, + 0.892, + 0.267 + ], + "angle": 0, + "content": "Table 7. Results of text-to-video generation on MSRVTT [89]. We gray out methods that use significantly more pre-training data." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.276, + 0.892, + 0.365 + ], + "angle": 0, + "content": "
MethodTextKitchenFlintstonesMUGEN
Q.T.GTQ.T.GTQ.T.GT
MMVG✘1.991.811.821.731.661.622.031.561.55
TATS [23]✓1.972.072.032.072.122.171.942.112.19
MMVG✓2.042.122.152.202.222.212.032.332.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.368, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Table 8. Human evaluation for TVP with aspects of video quality (Q.), instruction relevance (T.), and ground-truth similarity (GT)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.471 + ], + "angle": 0, + "content": "frames. For a fair comparison without additional data, we directly train on MSRVTT. MMVG still outperforms TATS, which shows that text-to-video generation can be improved through learning from video completion as well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Human Evaluation. We study the video quality (Q.), the relevance to the instructions (T.), and the similarity to the ground-truth video (GT) of the produced videos from the human aspect. The results in Table 8 are calculated as the mean ranking score (from 1 to 3, the higher is better) of each method for TVPrediction. MMVG without text even generates higher quality videos than TATS with text on Kitchen and MUGEN, where completion from partial frames benefits the temporal coherence of generative video modeling. While, the lowest ground-truth similarity illustrates that language guidance is crucial for controllability. With instruction, MMVG anticipates the future as the text (the highest T.) and generates videos that meet the ground truth (the highest GT), achieving the best overall performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 3 illustrates the keyframes of the generated examples of three datasets. Thanks to the learning of completion from partial frames at diverse time points, a single trained MMVG can support all TVC tasks. For prediction, MMVG makes Fred \"turn his head\" or MUGEN \"jumps over the gear\" from the guided text. MMVG further recovers the missing spoons and forks for \"close drawer\" from the last frame in a more challenging rewind scenario. MMVG infills the missing event described in language such as \"stand up for dancing\", \"walk across the kitchen\", \"jump onto the stage\" from the head and tail." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "From the same visual guidance, MMVG can lead TVC results using different texts, achieving controllable video completion. For example, we can let MUGEN \"jump down" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.668, + 0.893, + 0.697 + ], + "angle": 0, + "content": "Figure 3. Qualitative examples of TVC on Kitchen [13], Flintstones [26], and MUGEN [29]. We also illustrate video generation on UCF-101 [69], video prediction on BAIR [18], and text-to-video prediction on WebVid [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.858 + ], + "angle": 0, + "content": "the ground\" or \"land on a face\", starting from the same beginning. We can also control the behavior as \"keep walking\" or \"jump over a gear\" to recover the missing middle event. Furthermore, MMVG also carries out unconditional video generation with smooth temporal coherence. We can use language to produce natural dynamics in diverse scenes, such as \"sunset on the sea\", \"green sea turtle swims\", or a close look of \"cut chicken\". The presented videos indicate that our method not only unifies all TVC tasks but also performs the classical video generation/text-to-video well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.704, + 0.621, + 0.72 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.895, + 0.897 + ], + "angle": 0, + "content": "We introduce a novel task of text-guided video completion (TVC) that performs video completion from the first, last, or both frame(s) controlled by language. We present Multimodal Masked Video Generation (MMVG) with an effective masking strategy to learn the visual guidance from any time point. By varying the masking conditions, MMVG addresses all prediction, rewind, and infilling tasks within one model. Experiments on various video scenes show that our MMVG effectively addresses TVC as well as generative video modeling. We believe TVC can help advance a new field toward vision-and-language research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10688" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards High Resolution Video Generation with Progressive Growing of Sliced Wasserstein GANs. In arXiv:1810.02419, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H. Campbell, and Sergey Levine. Stochastic Variational Video Prediction. In International Conference for Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. FitVid: Overfitting in Pixel-Level Video Prediction. In arXiv:2106.13195, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.34 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2LIVE: Text-Driven Layered Image and Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[7] Ya-Liang Chang, Zhe Liu Yu, and Winston Hsu. VOR-Net: Spatio-temporally Consistent Video Inpainting for Object Removal. In Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent Online Video Style Transfer. In International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[9] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Jonathan Warrell, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy Mitra, and Philip Torr. ImageSpirit: Verbal Guided Image Parsing. In Conference on Human Factors in Computing Systems (CHI), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[10] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial Video Generation on Complex Datasets. In arXiv:1907.06571, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[11] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying Generalization in Reinforcement Learning. In International Conference on Machine Learning (ICML), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[12] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.79, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling Egocentric Vision: The EPIC-KITCHENS Dataset. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[14] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary Video Style" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Transfer via Multi-Channel Correlation. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[15] Emily Denton and Rob Fergus. Stochastic Video Generation with a Learned Prior. In International Conference on Machine Learning (ICML), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In arXiv:2105.05233, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering Text-to-Image Generation via Transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[18] Frederik Ebert, Chelsea Finn, Alex X. Lee, and Sergey Levine. Self-Supervised Visual Planning with Temporal Skip Connections. In Conference on Robot Learning (CoRL), 2017. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[19] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W.Taylor. Tell, Draw, and Repeat: Generating and Modifying Images Based on Continual Linguistic Instruction. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[20] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. SSCR: Iterative Language-Based Image Editing via Self-Supervised Counterfactual Reasoning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. M\\(^3\\)L: Language-based Video Editing via Multi-Modal Multi-Level Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-Driven Artistic Style Transfer. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[23] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer. In European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[24] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martín, and Li Fei-Fei. MaskViT: Masked Visual Pre-Training for Video Prediction. In arXiv:2206.11894, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[26] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine This! Scripts to Compositions to Videos. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[27] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10689" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[28] Zekun Hao, Xun Huang, and Serge Belongie. Controllable Video Generation with Sparse Trajectories. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[29] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Qiyuan Hu, and Devi Parikh. MUGEN: A Playground for Video-Audio-Text Multimodal Understanding and GENeration. In European Conference on Computer Vision (ECCV), 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In International Conference on Computer Vision (ICCV), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In arXiv:2204.03458, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers. In arXiv:2205.15868, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[36] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Learning to Decompose and Disentangle Representations for Video Prediction. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[37] Qiyang Hu, Adrian Waelchli, Tiziano Portenier, Matthias Zwicker, and Paolo Favaro. Video Synthesis from a Single Image and Motion Stroke. In arXiv:1812.01874, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[38] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable Image-to-Video Generation with Text Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[39] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion Models for Video Prediction and Infilling. In arXiv:2206.07696, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.776 + ], + "angle": 0, + "content": "[40] Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video Pixel Networks. In International Conference on Machine Learning (ICML), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.817 + ], + "angle": 0, + "content": "[41] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[42] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference for Learning Representations (ICLR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[43] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video Generation From Text. In AAAI Conference on Artificial Intelligence (AAAI), 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision (ICCV), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[45] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[46] Tanya Marwah, Gaurav Mittal, and Vineeth N. Balasubramanian. Attentive Semantic Video Generation using Captions. In International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[47] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep Multi-Scale Video Prediction Beyond Mean Square Error. In International Conference on Learning Representations (ICLR), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[48] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference for Learning Representations (ICLR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.433 + ], + "angle": 0, + "content": "[49] Mehdi Mirza and Simon Osindero. Conditional Generative Adversarial Nets. In arXiv:1411.1784, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[50] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.893, + 0.587 + ], + "angle": 0, + "content": "[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In International Conference on Machine Learning (ICML), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.893, + 0.644 + ], + "angle": 0, + "content": "[52] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To Create What You Tell: Generating Videos from Captions. In International Conference on Multimedia (MM), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[53] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In International Conference on Learning Representations Workshop (ICLRW), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[54] Björn Ommer Patrick Esser, Robin Rombach. Taming Transformers for High-Resolution Image Synthesis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[55] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. MirrorGAN: Learning Text-to-image Generation by Redescription. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Interna-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10690" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.118 + ], + "angle": 0, + "content": "tional Conference on Machine Learning (ICML), 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.468, + 0.162 + ], + "angle": 0, + "content": "[57] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent Video Transformer. In arXiv:2006.10704, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. In arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[59] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.468, + 0.316 + ], + "angle": 0, + "content": "[60] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. FILM: Frame Interpolation for Large Motion. In European Conference on Computer Vision (ECCV), 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.468, + 0.372 + ], + "angle": 0, + "content": "[61] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. In International Conference on Machine Learning (ICML), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.468, + 0.428 + ], + "angle": 0, + "content": "[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.429, + 0.468, + 0.525 + ], + "angle": 0, + "content": "[63] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.567 + ], + "angle": 0, + "content": "[64] Masaki Saito and Shunta Saito. TGANv2: Efficient Training of Large Models for Video Generation with Multiple Subsampling Layers. In arXiv:1811.09245, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.468, + 0.636 + ], + "angle": 0, + "content": "[65] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.468, + 0.705 + ], + "angle": 0, + "content": "[66] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Annual Meetings of the Association for Computational Linguistics (ACL), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.761 + ], + "angle": 0, + "content": "[67] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations (ICLR), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. In arXiv:2209.14792, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.873 + ], + "angle": 0, + "content": "[69] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. In arXiv:1212.0402, 2012. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.903 + ], + "angle": 0, + "content": "[70] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised Learning of Video Representations using" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "LSTMs. In International Conference on Machine Learning (ICML), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[71] Fuwen Tan, Song Feng, and Vicente Ordonez. Text2Scene: Generating Compositional Scenes from Textual Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[72] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A Good Image Generator Is What You Need for High-Resolution Video Synthesis. In International Conference for Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[73] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[74] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing Motion and Content for Video Generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[75] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, and Sylvain Gelly Marcin Michalski. Towards Accurate Generative Models of Video: A New Metric & Challenges. In International Conference on Learning Representations Workshop (ICLRW), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[76] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural Discrete Representation Learning. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[78] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation From Open Domain Textual Description. In arXiv:2210.02399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[79] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[80] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating Videos with Scene Dynamics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[81] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[82] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[83] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. VATEX: A Large-Scale," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10691" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "High-Quality Multilingual Dataset for Video-and-Language Research. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[84] Dirk Weissenborn, Oscar Tackstrom, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference on Learning Representations (ICLR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[85] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference for Learning Representations (ICLR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[86] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating Open-Domain Videos from nAtural Descriptions. In arXiv:2104.14806, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[87] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. NUWA: Visual Synthesis Pretraining for Neural visUal World creAtion. In European Conference on Computer Vision (ECCV), 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[88] Xide Xia, Tianfan Xue, Wei-Sheng Lai, Zheng Sun, Abby Chang, Brian Kulis, and Jiawen Chen. Real-Time Localized Photorealistic Video Style Transfer. In Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.504 + ], + "angle": 0, + "content": "[90] Qiangeng Xu, Hanwang Zhang, Weiyue Wang, Peter N. Belhumeur, and Ulrich Neumann. Stochastic Dynamics for Video Infilling. In Winter Conference on Applications of Computer Vision (WACV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[91] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep Flow-Guided Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[92] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[93] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally Consistent Semantic Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[94] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers. In arXiv:2104.10157, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. In arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[96] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks. In International Conference for Learning Representations (ICLR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[97] Jiangning Zhang, Chao Xu, Liang Liu, Mengmeng Wang, Xia Wu, Yong liu, and Yunliang Jiang. DTVNet: Dynamic" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Time-lapse Video Generation via Single Still Image. In European Conference on Computer Vision (ECCV), 2020. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10692" + } + ] +] \ No newline at end of file diff --git a/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_origin.pdf b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bbbf94ef16c32f6ecc0a5168cf5df1a1b61d6641 --- /dev/null +++ b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/0b8e3b80-d861-4dd7-999a-b659b22d0266_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8c0a70dc35d7448e17a7ee42d9053601d9562244d7134a67daa34939914143f +size 2231778 diff --git a/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/full.md b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b99b44b5264454fcc77d682a83d09afdcb34fe02 --- /dev/null +++ b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/full.md @@ -0,0 +1,386 @@ +# Tell Me What Happened: Unifying Text-guided Video Completion via Multimodal Masked Video Generation + +Tsu-Jui Fu $^{1}$ , Licheng Yu $^{2}$ , Ning Zhang $^{2}$ , Cheng-Yang Fu $^{2}$ , Jong-Chyi Su $^{3}$ , William Yang Wang $^{1}$ , Sean Bell $^{2}$ $^{1}$ UC Santa Barbara $^{2}$ Meta $^{3}$ NEC Laboratories America + +{tsu-juifu, william}@cs.ucsb.edu jcsu@nec-labs.com {lichengyu, ningzhang, chengyangfu, seanbell}@meta.com + +# Abstract + +Generating a video given the first several static frames is challenging as it anticipates reasonable future frames with temporal coherence. Besides video prediction, the ability to rewind from the last frame or infilling between the head and tail is also crucial, but they have rarely been explored for video completion. Since there could be different outcomes from the hints of just a few frames, a system that can follow natural language to perform video completion may significantly improve controllability. Inspired by this, we introduce a novel task, text-guided video completion (TVC), which requests the model to generate a video from partial frames guided by an instruction. We then propose Multimodal Masked Video Generation (MMVG) to address this TVC task. During training, MMVG discretizes the video frames into visual tokens and masks most of them to perform video completion from any time point. At inference time, a single MMVG model can address all 3 cases of TVC, including video prediction, rewind, and infilling, by applying corresponding masking conditions. We evaluate MMVG in various video scenarios, including egocentric, animation, and gaming. Extensive experimental results indicate that MMVG is effective in generating high-quality visual appearances with text guidance for TVC. + +# 1. Introduction + +Generative video modeling [15, 70, 84] has made great progress, which first succeeds in unconditional video generation [40, 64]. More recently, video prediction [28, 36, 47] has been trying the controllable setting, which anticipates the future by completing a video from the past frames or a static starting image [37, 97]. However, video prediction may produce various outcomes, which makes it difficult to meet human expectations. For the example in Fig. 1(a), the game agent can keep jumping to the right or move back and turn left. The limited guidance from only the first frame is + +![](images/71221f8c7f3394e88f91f99ec3b42f2d8cf77fdeed9c37ac836e7776eccb86ad.jpg) +(a) Video Prediction is Uncertain + +Video Prediction put bowl in saucepan + +Video Rewind Barney is trying to sing in a Living room. + +Video Infilling Runs to Left. Then collects a coin and a gem. + +![](images/40b1d76980540e6b924dcec212f21bccbdcc37e75f605ffc9788a581bbea9fc1.jpg) +(b) Text-guided Video Completion (TVC) +Figure 1. The introduced text-guided video completion (TVC) task. (a) Video prediction may have different outcomes without text guidance. (b) TVC performs video completion from the first frame (prediction), the last frame (rewind), or both (infilling), guided by the textual description. + +insufficient to tell the intention. For humans, language is the most straightforward way of communication. If a system can follow an instruction to accomplish video completion, it will significantly improve its controllability and make a vast application impact. On the other hand, compared with video prediction, video rewind and infilling have been rarely studied [39, 79], but they are also crucial. Breaking the limitation of chronological guidance should make the visual guidance more flexible, leading to a general video completion. + +We thus introduce a novel task, text-guided video completion (TVC), where the partial frames and a given instruction jointly guide the video generation. As illustrated in + +Fig. 1(b), we consider three scenarios of video completion: prediction from the first frame, rewind from the last frame, and infilling between the head and tail. The missing (to-be-completed) event should follow the textual instruction. Compared to generating content from scratch [43, 87], TVC requests models to understand the given visual and textual guidance before generation, which better mimics how human imagines after seeing and listening in our daily lives. + +To tackle TVC, we present Multimodal Masked Video Generation (MMVG) to perform video completion. Specifically, we represent the video frames as discrete visual tokens by temporal-aware VQGAN [54, 76]. One key challenge is to deal with the video frames that are not presented in chronological (e.g., the last frame for rewind). Different from autoregressive models [23, 86] that only condition on the previous frames, MMVG carries out video completion in an encoder-decoder manner. Specifically, we propose a masking strategy that masks different parts of the video and feeds them as the input to the multimodal encoder with the instruction. As shown in Fig. 2, we allow MMVG to consider the visual hints from different time points, and the decoder learns to produce the full target video. By varying the masking conditions (including the cases of only the first or last frame being accessible), a single MMVG can address all TVC tasks, including video prediction, rewind, and infilling. Moreover, learning the recovery from partial frames also empowers MMVG with a strong temporal coherence, contributing to better generative video quality. + +We consider videos in diverse scenarios for the TVC evaluation. There are Kitchen [13], Flintstones [26], and MUGEN [29] corresponding to the egocentric, animation, and gaming scenes. The model should generate videos such as performing kitchen activities in the first-person view, making characters act the assigned behavior, or imitating an agent playing game. All should be guided with the first/last (or both) frame(s) and controlled through the given human instructions. We also compare MMVG with previous methods [23,57,94,96] on UCF-101 [69] and BAIR [18] for the classic video generation/prediction tasks. + +Experimental results demonstrate that instruction is necessary to make video completion controllable, MMVG can address all three TVC tasks, and our proposed masking strategy enhances the temporal modeling, which further benefits general video generation/prediction. In summary, our contributions are three-fold: + +- We introduce TVC to generate a video from partial frames and control the temporal dynamics via natural language, where our video completion includes 3 cases: prediction, rewind, and infilling. +- We propose MMVG with an effective masking strategy to address all TVC tasks through a single training. +- Extensive experiments show that our MMVG can handle various types of video completion as well as video gener + +ation/prediction. We believe TVC can become a new topic in vision-and-language research. + +# 2. Related Work + +Video Generation/Prediction. Video generation aims to synthesize diverse videos from latent inputs [1, 74, 80]. Various generative modelings have shown promising results, including generative adversarial networks (GAN) [10, 24, 72, 96], autoregressive transformers [23, 77, 94], and denoising diffusion models [16, 33, 34]. Upon that, video prediction [2, 3, 25, 57, 85], which considers past frames to anticipate future observations, should maintain temporal dynamics from static images. Though the overall idea is also to complete a video from partial frames, other tasks, such as rewind and infilling [39, 79, 90], are not extensively explored. In this paper, we introduce TVC to comprehensively investigate the ability of video completion and make it more maneuverable via textual description. + +Text-to-Image/Video Genration. Generating visual content from language [9, 50, 71] has a vast application value in creative visual design. Previous works rely on the GANs framework [49] to produce images [19, 20, 22, 55, 61, 92] or videos [43, 46, 52], conditioned on text. With large-scale datasets [4, 65, 66, 83], recent pre-trained models can generate high-quality natural images from open-domain description through discrete visual tokens [12, 17, 54, 59, 76, 95] or the diffusion process [51, 58, 62, 63]. Leveraging such techniques further extends to generate vivid videos [32, 35, 68, 78, 86, 87]. However, those methods that depend on autoregressive generation can only be guided chronologically [27, 38]. Besides, video diffusion models require a deterministic video length, which cannot consider diverse temporal durations. In contrast, MMVG can perform video completion in different lengths from arbitrary time points and address all TVC tasks just with a single training. + +Text-guided Video-to-Video. Video inpainting [7, 41, 91], segmentation reconstruction [81, 82], or video style transfer [8, 14, 88] can be seen as a particular case of video-to-video synthesis (V2V). Even if text-guided V2V [5, 21, 93] can be controlled by language, it is still conditioned on a full video, where the temporal dynamics are usually provided. Different from that, TVC requires to regain the missing event from just partial guidance. It is more challenging since the model has to capture what happened from the instruction, maintain the temporal coherence among limited frames, and produce a complete video. + +# 3. Text-guided Video Completion (TVC) + +# 3.1. Task Definition + +We study the text-guided video completion (TVC) task to perform video completion from the first frame (prediction), + +![](images/ffed10af35ea387c77927a4c6216594d6774eeed7c65d4770f57a4ba0c64d3fe.jpg) +Figure 2. An overview of our Multimodal Masked Video Generation (MMVG). We present temporal-aware VQGAN (T-VQ) for discrete visual representation. MMVG considers the instruction $\mathcal{X}$ and partial frames of the video $\nu$ from diverse time points through masking, learning to generate the complete $\nu$ . In this way, a single trained MMVG can perform all prediction, rewind, and infilling tasks. + +the last frame (rewind), or the head and tail (infilling), conditioned on the textual instruction. During training, we have pairs of videos $\mathcal{V}$ and corresponding instructions $\mathcal{X}$ . Specifically, $\mathcal{V}$ consists of $N$ frames as $\{v_{1}, v_{2}, \ldots, v_{N}\}$ . Our goal is to train a unified model that generates the complete $\mathcal{V}$ given the partial frames from arbitrary time points and $\mathcal{X}$ . + +# 3.2. Multimodal Masked Video Generation + +Overview. An overview of our Multimodal Masked Video Generation (MMVG) is illustrated in Fig. 2. To model the video along with language, we propose temporal-aware VQGAN to represent a frame as visual tokens, which converts it into the same discrete space as the words. We present an effective masking strategy that masks different video parts for video completion learning. Those missing fragments are replaced by the unique [SPAN] tokens, and we consider the visual guidance from diverse time points. The multimodal encoder consumes the text and the partial-missing video, and the decoder learns to produce the complete video from arbitrary guided frames. By varying the masking conditions, MMVG learns to utilize the [SPAN] token and unifies all TVC tasks during the training. + +Temporal-aware Discrete Visual Tokens. VQ-VAE [76] has shown promising capability in representing data as discrete tokens. VQGAN [54] further models the prior distribution of the latent space via a transformer with the GAN training. If VQGAN is directly applied onto videos, it will ignore the inner temporal coherence and treat each frame as an independent image, resulting in an unsmooth video reconstruction. Though TATS [23] attempts to handle this by making $k$ consecutive frames altogether during VQ, it has + +to pre-define the constant $k$ before training. Such constraint forbids it from representing a frame at any timestamp. + +To address it with flexibility, we propose temporal-aware VQGAN (T-VQ) to inject the temporal relationship into the latent representation. We first follow VQGAN to learn the target visual tokens $z_{i}$ by reconstructing a video frame $v_{i}$ : + +$$ +\begin{array}{l} \mathbf {z} _ {i} = \mathbf {q} (\mathsf {E n c} ^ {\mathsf {Q}} (v _ {i}) \mid C), \\ \hat {v} _ {i} = \operatorname {D e c} ^ {0} (\mathrm {z} _ {i}), \\ \mathcal {L} _ {\mathrm {V Q}} = \underbrace {\left| \left| \hat {v _ {i}} - v _ {i} \right| \right| _ {1}} _ {\text {r e c o n s t r c u t i o n}} + \underbrace {\left| \left| \operatorname {s g} [ \operatorname {E n c} ^ {Q} (v _ {i}) ] - C _ {z _ {i}} \right| \right| _ {2} ^ {2}} _ {\text {c o d e b o o k}} \tag {1} \\ + \underbrace {\beta | | \mathrm {s g} [ C _ {\mathrm {z} _ {i}} ] - \mathrm {E n c} ^ {\mathrm {Q}} (v _ {i}) | | _ {2} ^ {2}} _ {\text {c o m m i t}} + \underbrace {| | \mathcal {F} (\hat {v} _ {i}) - \mathcal {F} (v _ {i}) | | _ {1}} _ {\text {m a t c h i n g}}, \\ \end{array} +$$ + +where $\mathrm{Enc}^{\mathrm{Q}}$ and $\mathrm{Dec}^{\mathrm{Q}}$ are the VQ encoder and decoder. The discrete latent code $z_{i}$ is acquired from the quantization operation q [54], which adopts nearest neighbor search by the trainable codebook $C$ . We apply the straight-through estimator over the stop-gradient operation sg and use $\beta$ as 0.25 [76]. We also append VGG [67] features matching $\mathcal{F}$ to stabilize the VQ loss $\mathcal{L}_{\mathrm{VQ}}$ [23]. The adversarial training between the frame quality loss $\mathcal{L}_G$ and discrimination loss $\mathcal{L}_D$ are further calculated from the discriminator $D$ : + +$$ +\begin{array}{l} \mathcal {L} _ {G} = \log (1 - D (\hat {v} _ {i})), \\ \mathcal {L} _ {D} = \log (1 - D (\hat {v} _ {i})) + \log (D (v _ {i})). \end{array} \tag {2} +$$ + +To inject the temporal relationship into $z$ , T-VQ is trained with the introduced contrastive temporal reasoning: + +$$ +o _ {i} = \mathbf {F C} ^ {\mathrm {T}} \left(\mathbf {z} _ {i}, \mathbf {z} _ {j}\right), \tag {3} +$$ + +$$ +\mathcal {L} _ {\mathrm {T}} = \operatorname {B C E L o s s} \left(o _ {i}, 0 \text {i f} i > j \text {e l s e} 1\right), +$$ + +where $j$ is a random frame from the same video. $\mathrm{FC^T}$ is the MLP classifier, and BCELoss is the binary cross-entropy for before/after. Learning the temporal order from $\mathcal{L}_{\mathrm{T}}$ , $\mathbf{Z}$ facilitates an implicit temporal coherence, leading to smooth video modeling. Moreover, since $\mathbf{Z}$ represents a single image, it is flexible to support frames at arbitrary timestamps. + +Generation from Masked Video. We propose the masking strategy $\mathcal{M}$ to obtain the masked videos $\overline{\nu}$ from diverse time points. $\mathcal{M}$ masks out most video frames with the probability $p$ and replaces each fragment as a unique [SPAN] token. For example, $\mathcal{M}$ reserves the third and the fifth frame, and masks all the others over a video length of 5: + +$$ +\overline {{\mathcal {V}}} \colon \{[ S ], v _ {3}, [ S ], v _ {5} \} = \mathcal {M} (\mathcal {V} \mid p). \tag {4} +$$ + +Our goal is to recover the missing part from $\overline{\mathcal{V}}$ and perform video completion, guided by the instruction $\mathcal{X}$ . To model between the vision and language modalities, we apply our Enc $^{\mathrm{Q}}$ over $\overline{\mathcal{V}}$ for the discrete visual tokens $\{[\mathrm{S}],\mathrm{z}_3,[\mathrm{S}],\mathrm{z}_5\}$ . We also tokenize the text $\mathcal{X}$ into word tokens $\{w_i\}_{i = 1}^L$ with the CLIP tokenizer [56], where $L$ is the length of $\mathcal{X}$ . As in the same discrete space, MMVG can achieve cross-modal fusion by the multimodal encoder (Enc $^{\mathrm{M}}$ ) through the self-attention mechanism as the transformer [77]: + +$$ +\begin{array}{l} f _ {i} ^ {w}, f _ {j} ^ {v} = \mathrm {L P} ^ {w} \left(w _ {i}\right), \mathrm {L P} ^ {v} \left(\mathrm {z} _ {j}\right) \\ \{1 \} = \sum_ {i = 1} ^ {n} M (f _ {i}) w _ {i} \end{array} \tag {5} +$$ + +$$ +\{h \} = \operatorname {E n c} ^ {\mathrm {M}} ([ \{f ^ {w} \}, \{f ^ {v} \} ]), +$$ + +where it obtains the features $f$ by the linear projection (LP), and $h$ is the hidden encoding features. We can also regard LP as the video/language embedder, which extracts the preliminary visual/linguistic features. + +After encoding the language hint and the partial-missing video from Enc $^{\text{M}}$ , our video decoder (Dec $^{\text{M}}$ ) learns to produce all frames for comprising the complete video. Dec $^{\text{M}}$ follows the vanilla autoregressive decoder, which first conducts self-attention over the past generated tokens and then predicts the discrete visual tokens as the video frame, conditioned on the encoded features $h$ : + +$$ +\hat {z} _ {t} = \operatorname {D e c} ^ {\mathrm {M}} (\left\{\hat {z} _ {1}, \dots , \hat {z} _ {t - 1} \right\} \mid \{h \}), +$$ + +$$ +\mathcal {L} _ {t} = \underset {N} {\operatorname {C E L o s s}} \left(\hat {z} _ {t}, z _ {t}\right), \tag {6} +$$ + +$$ +\mathcal {L} _ {\mathrm {M}} = \sum_ {t = 1} ^ {N} \mathcal {L} _ {t}, +$$ + +where $\mathbf{z}_t$ is the ground-truth tokens of the frame $v_t$ in the original $\mathcal{V}$ . We calculate the video decoding loss $\mathcal{L}_{\mathrm{M}}$ by the cross-entropy (CELoss) to learn video generation as classification. Our $\mathrm{Dec}^{\mathrm{M}}$ is built upon VideoSwin [45], which has shown a strong visual perception on various video understanding tasks. The 3D-shifted windows [44] consider different levels of spatial-temporal attention, and each window models blocks of video patches across $T'$ consecutive + +Algorithm 1 Multimodal Masked Video Generation +1: while TRAIN_T-VQ do +2: $\mathcal{V}\gets$ sample video +3: $z_{i} = q(\mathrm{Enc}^{Q}(v_{i})\mid C)$ +4: $\hat{v}_i = \mathrm{Dec}^{\mathrm{Q}}(z_i)$ +5: $o_i = \mathrm{FC}^{\mathrm{T}}(z_i,z_j)$ ▷ randomly sampled frame $j$ +6: $\mathcal{L}_{\mathrm{VQ}},\mathcal{L}_G\gets$ reconstruction, frame quality loss $\triangleright$ Eq. 1&2 +7: $\mathcal{L}_{\mathrm{T}}\gets$ temporal ordering loss $\triangleright$ Eq. 3 +8: Update T-VQ by minimizing $\mathcal{L}_{\mathrm{VQ}} + \mathcal{L}_G + \mathcal{L}_{\mathrm{T}}$ +9: $\mathcal{L}_D\gets$ discrimination loss $\triangleright$ Eq. 2 +10: Update D by maximizing $\mathcal{L}_D$ +11: end while +12: +13: while TRAIN_MMVG do +14: $\mathcal{V},\mathcal{X},p\gets$ sample video/instruction/probability +15: $\overline{\mathcal{V}}:\{v_a,[S],v_b,\ldots\} = \mathcal{M}(\mathcal{V}\mid p)$ $\triangleright$ diverse guided frames +16: $\{\mathbf{z}_a,[S],\mathbf{z}_b,\ldots\},\{w\} = \mathrm{Enc}^{\mathrm{Q}}(\overline{\mathcal{V}}),$ Tokenizer(X) +17: $\{h\} = \mathrm{Enc}^{\mathrm{M}}([\{w\},\{z_a,[S],z_b,\ldots\}])$ +18: for $t\gets 1$ to $N$ do +19: $\hat{z}_t = \mathrm{Dec}^{\mathrm{M}}(\{z_1,\dots,z_{t-1}\}\mid\{h\})$ $\triangleright$ teacher-forcing +20: $\mathcal{L}_t\gets$ video decoding loss $\triangleright$ Eq. 6 +21: end for +22: $\hat{\nu} = \mathrm{Dec}^{\mathrm{Q}}(\{\hat{z}_{t=1}^N\})$ +23: $\mathcal{L}_{\mathrm{M}} = \sum_{t=1}^{N}\mathcal{L}_{t}$ +24: Update MMVG by minimizing $\mathcal{L}_{\mathrm{M}}$ +25: $p\gets$ update masking probability $\triangleright$ Eq. 8 +26: end while + +frames. To ensure the same dimension for video generation in $\mathrm{Dec}^{\mathrm{M}}$ , we remove the temporal down-sampling layer. In the end, we can utilize $\mathrm{Dec}^{\mathrm{Q}}$ to reconstruct all the frames as our completed videos $\hat{\mathcal{V}}$ : + +$$ +\hat {\mathcal {V}} = \operatorname {D e c} ^ {\mathrm {Q}} \left(\left\{\hat {z} \right\} _ {t = 1} ^ {N}\right). \tag {7} +$$ + +By varying the masking conditions through $\mathcal{M}$ , MMVG learns how to complete a video from partial frames $\overline{\nu}$ at arbitrary time points with the text, which overcomes the limitation of chronological guidance. To make $\mathcal{M}$ more effective, we apply an adaptive probability $p$ instead of random sampling every time. Each video $\mathcal{V}$ keeps its own $p$ , and all frames are equally initialized in the beginning. Based on the prediction error, we adjust the masking probability $p_t$ of the $t$ -th frame: + +$$ +p _ {t} = p _ {t} + \alpha \left(\left(\frac {\mathcal {L} _ {t}}{\mathcal {L} _ {M}} \sum p\right) - p _ {t}\right), \tag {8} +$$ + +where $\alpha$ is the adjusting rate. A larger video decoding loss $\mathcal{L}_t$ indicates that the $t$ -th frame is more difficult to recover. MMVG learns more from those challenging cases and can bring better generative quality for video completion. + +Unifying TVC during Inference. After training with text and partial-missing video, MMVG learns to perform video completion over [SPAN] tokens. Then for inference, Enc $^{\text{M}}$ takes the following as its input to support different tasks: + +
MethodTextTVPredictionTVRewindTVInfilling
KitchenFlintstonesMUGENKitchenFlintstonesMUGENKitchenFlintstonesMUGEN
FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑
FILM [60]✘------------250.256.1352.751.4538.85.9
VideoMAE [73]✘328.947.6317.555.6548.77.0365.948.2335.555.9545.27.1246.954.7211.560.6494.97.8
TATS [23]✘106.964.4127.560.3376.57.1107.762.7127.660.2350.87.271.572.7119.566.7328.28.4
MMVGU✘105.663.3124.860.5374.57.2109.862.6124.359.7356.47.071.573.4121.866.3328.47.8
MMVGS✘103.864.5123.860.8369.47.3105.963.6123.860.5347.87.268.573.6118.567.9324.38.4
TATS [23]✓87.266.3115.970.690.167.989.863.3116.370.489.868.757.477.695.878.258.973.6
MMVGU✓80.268.4108.272.984.870.283.266.9113.271.693.168.459.877.892.878.359.273.2
MMVGS✓75.668.8106.373.783.371.179.768.1107.272.988.770.056.078.191.679.657.274.1
+ +- TVPrediction: $\left[\left\{w\right\},\left\{z_{1},[SPAN]\right\}\right]$ +- TVRewind: $\left[\left\{\boldsymbol{w}\right\},\left[\left[\mathrm{SPAN}\right],\mathbf{z}_{N}\right]\right]$ +TVInfilling: $\left[\left\{w\right\} ,\left\{\mathbf{z}_1,\left[\mathrm{SPAN}\right],\mathbf{z}_N\right\}\right]$ + +In this way, a single trained MMVG can unify all TVC tasks without the specific downstream fine-tuning. + +# 3.3. Learning of MMVG + +Algo. 1 illustrates the learning process of our proposed MMVG for TVC. We first train T-VQ over video frames for discrete visual tokens with contrastive temporal reasoning. Specifically, we minimize the VQ reconstruction loss $\mathcal{L}_{\mathrm{VQ}}$ and frame quality loss $\mathcal{L}_G$ along with our temporal ordering loss $\mathcal{L}_{\mathrm{T}}$ to optimize T-VQ. At the same time, we also update the discriminator $D$ via the standard adversarial training by maximizing the discrimination loss $\mathcal{L}_D$ . For video completion, the masking strategy $\mathcal{M}$ masks the video frames with the probability $p$ and then acquires guided frames from diverse time points. MMVG regards text and partial-missing video by Enc for cross-modal fusion, and Dec further predicts the visual tokens of frames autoregressively as the complete video. As a sequential generation process, we apply the teacher-forcing trick. Instead of our predicted $\hat{z}$ , the ground-truth $z$ from the previous timestamp is fed to stabilize the training. Each video decoding loss $\mathcal{L}_t$ at timestamp $t$ is summed up as $\mathcal{L}_{\mathrm{M}}$ to optimize MMVG. According to $\mathcal{L}_t$ , we update $p$ for effective masking probability. The entire optimization object can be summarized as two phases: + +$$ +\begin{array}{l} \mathrm {T - V Q}: \min _ {\mathrm {E n c} ^ {\mathrm {Q}}, \mathrm {D e c} ^ {\mathrm {Q}}, C, \mathrm {F C} ^ {\mathrm {T}}} \max _ {D} \mathcal {L} _ {\mathrm {V Q}} + \mathcal {L} _ {G} + \mathcal {L} _ {D} + \mathcal {L} _ {\mathrm {T}} \\ \text{MMVG:}\min_{\substack{\text{Enc}^{\text{M}},\text{Dec}^{\text{M}}}}\mathcal{L}_{\text{M}} \\ \end{array} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. As a new task, we consider diverse video scenes with natural instructions for TVC. Kitchen [13] records 22K egocentric videos about kitchen activity, which have different lengths (4-16 frames) with narrations. Flintstones [26] contains 25K animation videos (15 frames) from The Flintstones, where each video description includes the characters + +Table 1. Results of TVC, including prediction, rewind, and infilling, on Kitchen [13], Flintstones [26], and MUGEN [29]. TATS [23] requires specific training to support different tasks. We further train the unified MMVG $^{\text{U}}$ for each specific task as MMVG $^{\text{S}}$ . + +
DatasetTrain / Val#Frame#WordFPS
Kitchen [13]16,695 / 5,8048.32.86
Flintstones [26]22,666 / 2,5181516.55
MUGEN [29]362,239 / 12,8481620.65
+ +Table 2. The statistics of our used datsets to evaluate TVC. + +and their behavior. MUGEN [29] is built from agents playing CoinRun [11], which consists of 375K gaming videos (16 frames) with detailed text annotations. All videos in these three datasets are resized into $128 \times 128$ . An overview is shown in Table 2 and Fig. 1(b). Since MMVG can unify various tasks, we also evaluate video generation/prediction on widely-used UCF-101 [69] and BAIR [18], video infilling on UCF-101 following RaMViD [39], and text-to-video generation on MSRVTT [89]. + +Evaluation Metrics. We apply the following metrics to evaluate TVC results: 1) FVD [75] computes the video features [6] distance to the ground truth; 2) RCS [86] is the relative visual-text similarity to the instruction, compared to the ground-truth video. We fine-tune the CLIP model [56] on each dataset and adapt it to the video scene for a more precise alignment. Apart from automatic metrics, we also conduct a human evaluation from aspects of video quality, instruction relevance, and ground-truth similarity. We sample 75 TVP results for each task and adopt MTurk1 to rank over baselines and our MMVG. To avoid the potential ranking bias, we hire 3 MTurkers for each sample of prediction, rewind, and infilling tasks. + +Implementation Detail. T-VQ contains ResBlocks [31] as the visual auto-encoder (Enc $^{\text{Q}}$ and Dec $^{\text{Q}}$ ). The discriminator $D$ follows a similar architecture to Enc $^{\text{Q}}$ . For the vector quantization, we use a patch size 16, where a 128x128 video frame transforms into 8x8 discrete visual tokens. There are 1024 vocabularies in the codebook $C$ , and the hidden embedding size is 256. We adopt batch size 32 with a learning rate of 4.5e-6 to optimize T-VQ by Adam [42]. MMVG is built in an encoder-decoder manner, where Enc $^{\text{M}}$ is a trans + +
MethodKitchenFlintstonesMUGEN
VideoDiff [34]138.6206.4410.7
MCVD [79]119.9183.8400.2
TATS [23]115.5157.5386.4
MMVG109.1127.6368.6
+ +former with 24 layers, 16 attention heads, and hidden embedding size 1024. $\mathrm{Dec}^{\mathrm{M}}$ adopts a similar setting with temporal window size 3 in VideoSwin. The initial sample rate $p$ of the masking strategy $\mathcal{M}$ is 0.9 with an adjusting rate $\alpha$ as 0.1. We optimize MMVG through the mixed precision [48] with batch size 4 by Adam. The learning rate is also 4.5e-6. All experiments are implemented in PyTorch [53] and done on 8 NVIDIA A100 GPUs. + +# 4.2. Main Results + +Table 1 shows the results of all text-guided prediction, rewind, and infilling for TVC. VideoMAE [73] is built upon MAE [30] and reconstructs the missing video cubes, which performs TVC by masking all video frames except the first or the last (or both). TATS [23], the SOTA on video generation, also produces videos as discrete visual tokens. Since TATS can only consider the past through the autoregressive transformer, it requires specific training for each task. We have MMVG $^{\text{U}}$ as the unified model that can support all TVC tasks with a single training and MMVG $^{\text{S}}$ to further train for each prediction, rewind, and infilling. We treat TATS as our main baseline $^{2}$ and study the importance of guided text. + +TVPrediction. VideoMAE attempts to produce all frames simultaneously, which is difficult to maintain video temporal consistency, resulting in a high 328.9 FVD on Kitchen. TATS is inherently designed for prediction as it generates the frames one after one. However, our unified MMVGU performs better than TATS on all datasets (e.g., lower 105.6 and 124.8 FVD on Kitchen and Flintstones). These results support that learning from diverse time points will not hurt the prediction from the past. In contrast, our masking strategy can bring superior temporal coherence. MMVGS further improves itself through training prediction as completion from the head. However, there are too many possible outcomes from just the beginning, where the predicted results may not meet the expectation (e.g., a high 370 FVD on MUGEN). The instruction as guidance makes it related to the expected ground-truth result. We can let MUGEN run, jump, or collect coins as the textual descriptions to achieve more controllability, leading to a noticeable improvement (e.g., a lower 84.8 FVD by MMVGU). The higher 70.2 RCS also shows that our MMVG can produce MUGEN videos + +Table 3. FVD results of video generation on our TVC datasets. + +
MethodUCF-101
IS↑FVD↓
VideoGPT [94]24.7-
DIGAN [96]32.7577
VideoDiff [34]57.0-
TATS [23]57.6420
MMVG58.3395
+ +
MethodBAIR
FVD↓
VideoGPT [94]103.3
MaskViT [25]93.6
MCVD [79]89.5
TATS [23]88.6
MMVG85.2
+ +Table 5. Results of video prediction on BAIR [18]. + +Table 4. Results of video generation on UCF-101 [69]. + +that confirm with the instruction. Although the model may try to imagine the animation or the kitchen activity, the language hint can provide a clear goal to anticipate. Likewise, MMVG $^{\text{U}}$ with text surpasses TATS, even though it is not designed for prediction only. The specific trained MMVG $^{\text{S}}$ benefits the unified model for further improvement. + +TVRewind. Rewind from the last allows the model to imagine what happened along with a suitable opening. In addition, the objects may not display on the last frame (e.g., the spoons and forks for "close drawer"), which makes it more challenging to complete. Similar to prediction, VideoMAE cannot have feasible rewind results. Language is still essential to remind the past and establish an adequate beginning, where we can find a significant performance gap between with and without text (e.g., 90 vs. 350 FVD on MUGEN). Our unified MMVG $^{\text{U}}$ achieves comparable results to TATS and even outperforms on Kitchen and Flintstones (e.g., higher 66.9 and 71.6 RCS). With the learning of completion from partial frames, autoregressive model can still accomplish video rewind without specific training. If following TATS design to train MMVG $^{\text{U}}$ for rewind, MMVGS gains more improvement and utterly surpasses it. + +TVInfilling. We consider the additional FILM [60] for infilling, which performs video interpolation with in-between motion. Despite synthesizing intermediate frames between the first and the last, the visual dynamics are changing too rapidly to handle, resulting in a higher FVD. With guidance from the head and tail, we find a noticeable improvement even without instruction (e.g., lower FVDs on Kitchen), which is helpful in temporal video modeling. To capture the expected missing event, we still require the language hint for more controllability. Our unified MMVG $^{\text{U}}$ achieves comparable performance to TATS again, which is specifically trained for the infilling task. It shows that completion from partial frames at different time points still helps, and MMVG $^{\text{S}}$ further outperforms on TVInfilling. + +# 4.3. Additional Study + +Video Generation/Prediction. We further evaluate the classic video generation and prediction tasks. Table 3 shows FVD scores of unconditional video generation on our TVC datasets. Note that only videos but no texts are used in these experiments. Both VideoDiff [34] and MCVD [79] are built + +
MethodUCF-101 (FVD↓)
K=+1+2+5±1±2
RaMViD [39]349.7300.6260.5215.4162.5
MMVG316.3258.5194.6183.2120.3
+ +upon denoising diffusion [33], where MCVD also considers different partial frames during training. The results first indicate that the vanilla token-based method is superior to the diffusion-based model (TATS vs. VideoDiff) for video generation. In addition, MMVG, with the masking strategy that learns the visual guidance from diverse time points, further boosts the performance (the lowest 127.6, 109.1, and 368.6 FVD on Flintstones, Kitchen, and MUGEN, respectively). + +We also evaluate MMVG on UCF-101 [69], which is challenging to generate natural human videos. Table 4 supports that our MMVG can produce videos with higher visual similarity (a higher 58.3 IS) and temporal alignment (a lower 395 FVD) to the ground truth. For video prediction, we apply the widely-used BAIR [18] in Table 5, where the model has to anticipate how a robot pushes objects from the given first frame. MMVG again surpasses TATS with the lowest 85.2 FVD. Although both generation and prediction are generating video frames chronologically, the ability to recover arbitrary missing frames for video completion empowers MMVG with a stronger temporal coherence, leading to better generative video quality. + +Video Infilling. We follow RaMViD [39] to evaluate video infilling on UCF-101. We consider various guidance settings $K$ in Table 6. For example, $K = +1$ means given the first frame, and $K = \pm 2$ should provide the first and last two frames. For prediction, MMVG outperforms RaMViD on all $K$ , and the performance gap gets even larger when more guided frames are accessible (e.g., 33.4 on $K = +1$ and 65.9 on $K = +5$ ). A similar result can be found for infilling, where MMVG can make the lowest 120.3 FVD on $K = \pm 2$ . Despite having a similar masking strategy, it shows that generating frames one after one still brings superior results. MMVG allows autoregressive models to condition on visual hints from any time point, which produces more similar videos to the ground truth when infilling between the head and tail. + +Text-to-Video Generation. Being a multimodal generative model, MMVG supports text-to-video generation. To compare with those large-scale methods, we pre-train MMVG using WebVid [4], which contains 2.5M text-video pairs. We adopt the masking strategy to treat the pre-training as video completion. MMVG outperforms GODIVA [86] and NUWA [87] without access the MSRVTT [89] data in Table 7. Surprisingly, MMVG can generate videos that are more related to the texts (a higher 0.2644 CLIP-S [86]) thanCogVideo [35], even though using twice less data. This result encourages the effectiveness of completion from partial + +Table 6. Results of video prediction and infilling on UCF-101. + +
MethodPre-trainingZero-shotMSRVTT
FID↓CLIP-S↑
GODIVA [86]136MX-0.2402
NUWA [87]3.9MX47.70.2439
CogVideo [35]5.4M✓23.60.2631
Make-A-Video [68]20M✓13.20.3049
MMVG2.5M✓23.40.2644
TATS [23]-X63.20.2326
MMVG-X60.60.2385
+ +Table 7. Results of text-to-video generation on MSRVTT [89]. We gray out methods that use significantly more pre-training data. + +
MethodTextKitchenFlintstonesMUGEN
Q.T.GTQ.T.GTQ.T.GT
MMVG✘1.991.811.821.731.661.622.031.561.55
TATS [23]✓1.972.072.032.072.122.171.942.112.19
MMVG✓2.042.122.152.202.222.212.032.332.26
+ +Table 8. Human evaluation for TVP with aspects of video quality (Q.), instruction relevance (T.), and ground-truth similarity (GT). + +frames. For a fair comparison without additional data, we directly train on MSRVTT. MMVG still outperforms TATS, which shows that text-to-video generation can be improved through learning from video completion as well. + +Human Evaluation. We study the video quality (Q.), the relevance to the instructions (T.), and the similarity to the ground-truth video (GT) of the produced videos from the human aspect. The results in Table 8 are calculated as the mean ranking score (from 1 to 3, the higher is better) of each method for TVPrediction. MMVG without text even generates higher quality videos than TATS with text on Kitchen and MUGEN, where completion from partial frames benefits the temporal coherence of generative video modeling. While, the lowest ground-truth similarity illustrates that language guidance is crucial for controllability. With instruction, MMVG anticipates the future as the text (the highest T.) and generates videos that meet the ground truth (the highest GT), achieving the best overall performance. + +Qualitative Results. Fig. 3 illustrates the keyframes of the generated examples of three datasets. Thanks to the learning of completion from partial frames at diverse time points, a single trained MMVG can support all TVC tasks. For prediction, MMVG makes Fred "turn his head" or MUGEN "jumps over the gear" from the guided text. MMVG further recovers the missing spoons and forks for "close drawer" from the last frame in a more challenging rewind scenario. MMVG infills the missing event described in language such as "stand up for dancing", "walk across the kitchen", "jump onto the stage" from the head and tail. + +From the same visual guidance, MMVG can lead TVC results using different texts, achieving controllable video completion. For example, we can let MUGEN "jump down + +![](images/25f55c810660193d84f63031116c2a08db8790412d864d211dcf0dae800a38a1.jpg) +Figure 3. Qualitative examples of TVC on Kitchen [13], Flintstones [26], and MUGEN [29]. We also illustrate video generation on UCF-101 [69], video prediction on BAIR [18], and text-to-video prediction on WebVid [4]. + +the ground" or "land on a face", starting from the same beginning. We can also control the behavior as "keep walking" or "jump over a gear" to recover the missing middle event. Furthermore, MMVG also carries out unconditional video generation with smooth temporal coherence. We can use language to produce natural dynamics in diverse scenes, such as "sunset on the sea", "green sea turtle swims", or a close look of "cut chicken". The presented videos indicate that our method not only unifies all TVC tasks but also performs the classical video generation/text-to-video well. + +# 5. Conclusion + +We introduce a novel task of text-guided video completion (TVC) that performs video completion from the first, last, or both frame(s) controlled by language. We present Multimodal Masked Video Generation (MMVG) with an effective masking strategy to learn the visual guidance from any time point. By varying the masking conditions, MMVG addresses all prediction, rewind, and infilling tasks within one model. Experiments on various video scenes show that our MMVG effectively addresses TVC as well as generative video modeling. We believe TVC can help advance a new field toward vision-and-language research. + +# References + +[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards High Resolution Video Generation with Progressive Growing of Sliced Wasserstein GANs. In arXiv:1810.02419, 2018. 2 +[2] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H. Campbell, and Sergey Levine. Stochastic Variational Video Prediction. In International Conference for Learning Representations (ICLR), 2018. 2 +[3] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. FitVid: Overfitting in Pixel-Level Video Prediction. In arXiv:2106.13195, 2021. 2 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 7, 8 +[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2LIVE: Text-Driven Layered Image and Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2 +[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5 +[7] Ya-Liang Chang, Zhe Liu Yu, and Winston Hsu. VOR-Net: Spatio-temporally Consistent Video Inpainting for Object Removal. In Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2 +[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent Online Video Style Transfer. In International Conference on Computer Vision (ICCV), 2017. 2 +[9] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Jonathan Warrell, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy Mitra, and Philip Torr. ImageSpirit: Verbal Guided Image Parsing. In Conference on Human Factors in Computing Systems (CHI), 2013. 2 +[10] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial Video Generation on Complex Datasets. In arXiv:1907.06571, 2019. 2 +[11] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying Generalization in Reinforcement Learning. In International Conference on Machine Learning (ICML), 2019. 5 +[12] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance. In European Conference on Computer Vision (ECCV), 2022. 2 +[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling Egocentric Vision: The EPIC-KITCHENS Dataset. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8 +[14] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary Video Style + +Transfer via Multi-Channel Correlation. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[15] Emily Denton and Rob Fergus. Stochastic Video Generation with a Learned Prior. In International Conference on Machine Learning (ICML), 2018. 1 +[16] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In arXiv:2105.05233, 2021. 2 +[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering Text-to-Image Generation via Transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2 +[18] Frederik Ebert, Chelsea Finn, Alex X. Lee, and Sergey Levine. Self-Supervised Visual Planning with Temporal Skip Connections. In Conference on Robot Learning (CoRL), 2017. 2, 5, 6, 7, 8 +[19] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W.Taylor. Tell, Draw, and Repeat: Generating and Modifying Images Based on Continual Linguistic Instruction. In International Conference on Computer Vision (ICCV), 2019. 2 +[20] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. SSCR: Iterative Language-Based Image Editing via Self-Supervised Counterfactual Reasoning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 2 +[21] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. M $^3$ L: Language-based Video Editing via Multi-Modal Multi-Level Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[22] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-Driven Artistic Style Transfer. In European Conference on Computer Vision (ECCV), 2022. 2 +[23] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer. In European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7 +[24] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2 +[25] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martín, and Li Fei-Fei. MaskViT: Masked Visual Pre-Training for Video Prediction. In arXiv:2206.11894, 2022. 2, 6 +[26] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine This! Scripts to Compositions to Videos. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8 +[27] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 + +[28] Zekun Hao, Xun Huang, and Serge Belongie. Controllable Video Generation with Sparse Trajectories. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[29] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Qiyuan Hu, and Devi Parikh. MUGEN: A Playground for Video-Audio-Text Multimodal Understanding and GENeration. In European Conference on Computer Vision (ECCV), 2022. 2, 5, 8 +[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6 +[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In International Conference on Computer Vision (ICCV), 2015. 5 +[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In arXiv:2210.02303, 2022. 2 +[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 7 +[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In arXiv:2204.03458, 2022. 2, 6 +[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers. In arXiv:2205.15868, 2022. 2, 7 +[36] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Learning to Decompose and Disentangle Representations for Video Prediction. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 1 +[37] Qiyang Hu, Adrian Waelchli, Tiziano Portenier, Matthias Zwicker, and Paolo Favaro. Video Synthesis from a Single Image and Motion Stroke. In arXiv:1812.01874, 2018. 1 +[38] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable Image-to-Video Generation with Text Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[39] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion Models for Video Prediction and Infilling. In arXiv:2206.07696, 2022. 1, 2, 5, 7 +[40] Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video Pixel Networks. In International Conference on Machine Learning (ICML), 2017. 1 +[41] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[42] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference for Learning Representations (ICLR), 2015. 5 +[43] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video Generation From Text. In AAAI Conference on Artificial Intelligence (AAAI), 2018. 2 + +[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision (ICCV), 2021. 4 +[45] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4 +[46] Tanya Marwah, Gaurav Mittal, and Vineeth N. Balasubramanian. Attentive Semantic Video Generation using Captions. In International Conference on Computer Vision (ICCV), 2017. 2 +[47] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep Multi-Scale Video Prediction Beyond Mean Square Error. In International Conference on Learning Representations (ICLR), 2018. 1 +[48] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference for Learning Representations (ICLR), 2018. 6 +[49] Mehdi Mirza and Simon Osindero. Conditional Generative Adversarial Nets. In arXiv:1411.1784, 2014. 2 +[50] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In International Conference on Machine Learning (ICML), 2022. 2 +[52] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To Create What You Tell: Generating Videos from Captions. In International Conference on Multimedia (MM), 2017. 2 +[53] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In International Conference on Learning Representations Workshop (ICLRW), 2017. 6 +[54] Björn Ommer Patrick Esser, Robin Rombach. Taming Transformers for High-Resolution Image Synthesis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[55] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. MirrorGAN: Learning Text-to-image Generation by Redescription. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Interna- + +tional Conference on Machine Learning (ICML), 2021. 4, 5 +[57] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent Video Transformer. In arXiv:2006.10704, 2020. 2 +[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. In arXiv:2204.06125, 2022. 2 +[59] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. In International Conference on Machine Learning (ICML), 2021. 2 +[60] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. FILM: Frame Interpolation for Large Motion. In European Conference on Computer Vision (ECCV), 2022. 5, 6 +[61] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. In International Conference on Machine Learning (ICML), 2016. 2 +[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[63] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In arXiv:2205.11487, 2022. 2 +[64] Masaki Saito and Shunta Saito. TGANv2: Efficient Training of Large Models for Video Generation with Multiple Subsampling Layers. In arXiv:1811.09245, 2018. 1 +[65] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In arXiv:2111.02114, 2021. 2 +[66] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Annual Meetings of the Association for Computational Linguistics (ACL), 2018. 2 +[67] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations (ICLR), 2015. 3 +[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. In arXiv:2209.14792, 2022. 2, 7 +[69] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. In arXiv:1212.0402, 2012. 2, 5, 6, 7, 8 +[70] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised Learning of Video Representations using + +LSTMs. In International Conference on Machine Learning (ICML), 2015. 1 +[71] Fuwen Tan, Song Feng, and Vicente Ordonez. Text2Scene: Generating Compositional Scenes from Textual Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[72] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A Good Image Generator Is What You Need for High-Resolution Video Synthesis. In International Conference for Learning Representations (ICLR), 2021. 2 +[73] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5, 6 +[74] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing Motion and Content for Video Generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[75] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, and Sylvain Gelly Marcin Michalski. Towards Accurate Generative Models of Video: A New Metric & Challenges. In International Conference on Learning Representations Workshop (ICLRW), 2019. 5 +[76] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural Discrete Representation Learning. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 3 +[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 4 +[78] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation From Open Domain Textual Description. In arXiv:2210.02399, 2022. 2 +[79] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 6 +[80] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating Videos with Scene Dynamics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 2 +[81] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2 +[82] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2 +[83] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. VATEX: A Large-Scale, + +High-Quality Multilingual Dataset for Video-and-Language Research. In International Conference on Computer Vision (ICCV), 2019. 2 +[84] Dirk Weissenborn, Oscar Tackstrom, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference on Learning Representations (ICLR), 2020. 1 +[85] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference for Learning Representations (ICLR), 2020. 2 +[86] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating Open-Domain Videos from nAtural Descriptions. In arXiv:2104.14806, 2021. 2, 5, 7 +[87] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. NUWA: Visual Synthesis Pretraining for Neural visUal World creAtion. In European Conference on Computer Vision (ECCV), 2022. 2, 7 +[88] Xide Xia, Tianfan Xue, Wei-Sheng Lai, Zheng Sun, Abby Chang, Brian Kulis, and Jiawen Chen. Real-Time Localized Photorealistic Video Style Transfer. In Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7 +[90] Qiangeng Xu, Hanwang Zhang, Weiyue Wang, Peter N. Belhumeur, and Ulrich Neumann. Stochastic Dynamics for Video Infilling. In Winter Conference on Applications of Computer Vision (WACV), 2020. 2 +[91] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep Flow-Guided Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[92] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[93] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally Consistent Semantic Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2 +[94] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers. In arXiv:2104.10157, 2021. 2, 6 +[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. In arXiv:2206.10789, 2022. 2 +[96] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks. In International Conference for Learning Representations (ICLR), 2022. 2, 6 +[97] Jiangning Zhang, Chao Xu, Liang Liu, Mengmeng Wang, Xia Wu, Yong liu, and Yunliang Jiang. DTVNet: Dynamic + +Time-lapse Video Generation via Single Still Image. In European Conference on Computer Vision (ECCV), 2020. 1 \ No newline at end of file diff --git a/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/images.zip b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8765483fbd3a8ec954fcc931c468c299527eee82 --- /dev/null +++ b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd655c30d602a1f17adc3ff285979339980b574e0016157aa441945e4509718e +size 710911 diff --git a/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/layout.json b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d72a8867dd7088d17feb9290da3e95b96078fdeb --- /dev/null +++ b/2023/Tell Me What Happened_ Unifying Text-Guided Video Completion via Multimodal Masked Video Generation/layout.json @@ -0,0 +1,11340 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 501, + 138 + ], + "type": "text", + "content": "Tell Me What Happened: Unifying Text-guided Video Completion via Multimodal Masked Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "spans": [ + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": "Tsu-Jui Fu" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": ", Licheng Yu" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": ", Ning Zhang" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": ", Cheng-Yang Fu" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": ", Jong-Chyi Su" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": ", William Yang Wang" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": ", Sean Bell" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": "UC Santa Barbara " + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": "Meta " + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 160, + 443, + 202 + ], + "type": "text", + "content": "NEC Laboratories America" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 205, + 443, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 205, + 443, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 205, + 443, + 232 + ], + "type": "text", + "content": "{tsu-juifu, william}@cs.ucsb.edu jcsu@nec-labs.com {lichengyu, ningzhang, chengyangfu, seanbell}@meta.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 284, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 284, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 45, + 284, + 290, + 559 + ], + "type": "text", + "content": "Generating a video given the first several static frames is challenging as it anticipates reasonable future frames with temporal coherence. Besides video prediction, the ability to rewind from the last frame or infilling between the head and tail is also crucial, but they have rarely been explored for video completion. Since there could be different outcomes from the hints of just a few frames, a system that can follow natural language to perform video completion may significantly improve controllability. Inspired by this, we introduce a novel task, text-guided video completion (TVC), which requests the model to generate a video from partial frames guided by an instruction. We then propose Multimodal Masked Video Generation (MMVG) to address this TVC task. During training, MMVG discretizes the video frames into visual tokens and masks most of them to perform video completion from any time point. At inference time, a single MMVG model can address all 3 cases of TVC, including video prediction, rewind, and infilling, by applying corresponding masking conditions. We evaluate MMVG in various video scenarios, including egocentric, animation, and gaming. Extensive experimental results indicate that MMVG is effective in generating high-quality visual appearances with text guidance for TVC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Generative video modeling [15, 70, 84] has made great progress, which first succeeds in unconditional video generation [40, 64]. More recently, video prediction [28, 36, 47] has been trying the controllable setting, which anticipates the future by completing a video from the past frames or a static starting image [37, 97]. However, video prediction may produce various outcomes, which makes it difficult to meet human expectations. For the example in Fig. 1(a), the game agent can keep jumping to the right or move back and turn left. The limited guidance from only the first frame is" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 257, + 538, + 342 + ], + "blocks": [ + { + "bbox": [ + 313, + 257, + 538, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 538, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 538, + 342 + ], + "type": "image", + "image_path": "71221f8c7f3394e88f91f99ec3b42f2d8cf77fdeed9c37ac836e7776eccb86ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 346, + 481, + 355 + ], + "lines": [ + { + "bbox": [ + 370, + 346, + 481, + 355 + ], + "spans": [ + { + "bbox": [ + 370, + 346, + 481, + 355 + ], + "type": "text", + "content": "(a) Video Prediction is Uncertain" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 369, + 367, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 369, + 367, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 369, + 367, + 399 + ], + "type": "text", + "content": "Video Prediction put bowl in saucepan" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 407, + 369, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 369, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 369, + 444 + ], + "type": "text", + "content": "Video Rewind Barney is trying to sing in a Living room." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 449, + 367, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 449, + 367, + 485 + ], + "spans": [ + { + "bbox": [ + 309, + 449, + 367, + 485 + ], + "type": "text", + "content": "Video Infilling Runs to Left. Then collects a coin and a gem." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 375, + 363, + 545, + 486 + ], + "blocks": [ + { + "bbox": [ + 375, + 363, + 545, + 486 + ], + "lines": [ + { + "bbox": [ + 375, + 363, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 375, + 363, + 545, + 486 + ], + "type": "image", + "image_path": "40b1d76980540e6b924dcec212f21bccbdcc37e75f605ffc9788a581bbea9fc1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 490, + 493, + 500 + ], + "lines": [ + { + "bbox": [ + 358, + 490, + 493, + 500 + ], + "spans": [ + { + "bbox": [ + 358, + 490, + 493, + 500 + ], + "type": "text", + "content": "(b) Text-guided Video Completion (TVC)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 502, + 547, + 557 + ], + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 557 + ], + "type": "text", + "content": "Figure 1. The introduced text-guided video completion (TVC) task. (a) Video prediction may have different outcomes without text guidance. (b) TVC performs video completion from the first frame (prediction), the last frame (rewind), or both (infilling), guided by the textual description." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 676 + ], + "type": "text", + "content": "insufficient to tell the intention. For humans, language is the most straightforward way of communication. If a system can follow an instruction to accomplish video completion, it will significantly improve its controllability and make a vast application impact. On the other hand, compared with video prediction, video rewind and infilling have been rarely studied [39, 79], but they are also crucial. Breaking the limitation of chronological guidance should make the visual guidance more flexible, leading to a general video completion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "We thus introduce a novel task, text-guided video completion (TVC), where the partial frames and a given instruction jointly guide the video generation. As illustrated in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10681" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "Fig. 1(b), we consider three scenarios of video completion: prediction from the first frame, rewind from the last frame, and infilling between the head and tail. The missing (to-be-completed) event should follow the textual instruction. Compared to generating content from scratch [43, 87], TVC requests models to understand the given visual and textual guidance before generation, which better mimics how human imagines after seeing and listening in our daily lives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 288, + 409 + ], + "type": "text", + "content": "To tackle TVC, we present Multimodal Masked Video Generation (MMVG) to perform video completion. Specifically, we represent the video frames as discrete visual tokens by temporal-aware VQGAN [54, 76]. One key challenge is to deal with the video frames that are not presented in chronological (e.g., the last frame for rewind). Different from autoregressive models [23, 86] that only condition on the previous frames, MMVG carries out video completion in an encoder-decoder manner. Specifically, we propose a masking strategy that masks different parts of the video and feeds them as the input to the multimodal encoder with the instruction. As shown in Fig. 2, we allow MMVG to consider the visual hints from different time points, and the decoder learns to produce the full target video. By varying the masking conditions (including the cases of only the first or last frame being accessible), a single MMVG can address all TVC tasks, including video prediction, rewind, and infilling. Moreover, learning the recovery from partial frames also empowers MMVG with a strong temporal coherence, contributing to better generative video quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 541 + ], + "type": "text", + "content": "We consider videos in diverse scenarios for the TVC evaluation. There are Kitchen [13], Flintstones [26], and MUGEN [29] corresponding to the egocentric, animation, and gaming scenes. The model should generate videos such as performing kitchen activities in the first-person view, making characters act the assigned behavior, or imitating an agent playing game. All should be guided with the first/last (or both) frame(s) and controlled through the given human instructions. We also compare MMVG with previous methods [23,57,94,96] on UCF-101 [69] and BAIR [18] for the classic video generation/prediction tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": "Experimental results demonstrate that instruction is necessary to make video completion controllable, MMVG can address all three TVC tasks, and our proposed masking strategy enhances the temporal modeling, which further benefits general video generation/prediction. In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 617, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 617, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 286, + 664 + ], + "type": "text", + "content": "- We introduce TVC to generate a video from partial frames and control the temporal dynamics via natural language, where our video completion includes 3 cases: prediction, rewind, and infilling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 666, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 286, + 689 + ], + "type": "text", + "content": "- We propose MMVG with an effective masking strategy to address all TVC tasks through a single training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "- Extensive experiments show that our MMVG can handle various types of video completion as well as video gener" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 96 + ], + "type": "text", + "content": "ation/prediction. We believe TVC can become a new topic in vision-and-language research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 109, + 392, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 109, + 392, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 109, + 392, + 121 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 129, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 297 + ], + "type": "text", + "content": "Video Generation/Prediction. Video generation aims to synthesize diverse videos from latent inputs [1, 74, 80]. Various generative modelings have shown promising results, including generative adversarial networks (GAN) [10, 24, 72, 96], autoregressive transformers [23, 77, 94], and denoising diffusion models [16, 33, 34]. Upon that, video prediction [2, 3, 25, 57, 85], which considers past frames to anticipate future observations, should maintain temporal dynamics from static images. Though the overall idea is also to complete a video from partial frames, other tasks, such as rewind and infilling [39, 79, 90], are not extensively explored. In this paper, we introduce TVC to comprehensively investigate the ability of video completion and make it more maneuverable via textual description." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 503 + ], + "type": "text", + "content": "Text-to-Image/Video Genration. Generating visual content from language [9, 50, 71] has a vast application value in creative visual design. Previous works rely on the GANs framework [49] to produce images [19, 20, 22, 55, 61, 92] or videos [43, 46, 52], conditioned on text. With large-scale datasets [4, 65, 66, 83], recent pre-trained models can generate high-quality natural images from open-domain description through discrete visual tokens [12, 17, 54, 59, 76, 95] or the diffusion process [51, 58, 62, 63]. Leveraging such techniques further extends to generate vivid videos [32, 35, 68, 78, 86, 87]. However, those methods that depend on autoregressive generation can only be guided chronologically [27, 38]. Besides, video diffusion models require a deterministic video length, which cannot consider diverse temporal durations. In contrast, MMVG can perform video completion in different lengths from arbitrary time points and address all TVC tasks just with a single training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": "Text-guided Video-to-Video. Video inpainting [7, 41, 91], segmentation reconstruction [81, 82], or video style transfer [8, 14, 88] can be seen as a particular case of video-to-video synthesis (V2V). Even if text-guided V2V [5, 21, 93] can be controlled by language, it is still conditioned on a full video, where the temporal dynamics are usually provided. Different from that, TVC requires to regain the missing event from just partial guidance. It is more challenging since the model has to capture what happened from the instruction, maintain the temporal coherence among limited frames, and produce a complete video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 650, + 509, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 509, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 509, + 663 + ], + "type": "text", + "content": "3. Text-guided Video Completion (TVC)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 670, + 400, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 400, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 400, + 682 + ], + "type": "text", + "content": "3.1. Task Definition" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "We study the text-guided video completion (TVC) task to perform video completion from the first frame (prediction)," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10682" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 548, + 289 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 548, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 548, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 548, + 289 + ], + "type": "image", + "image_path": "ffed10af35ea387c77927a4c6216594d6774eeed7c65d4770f57a4ba0c64d3fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "text", + "content": "Figure 2. An overview of our Multimodal Masked Video Generation (MMVG). We present temporal-aware VQGAN (T-VQ) for discrete visual representation. MMVG considers the instruction " + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "text", + "content": " and partial frames of the video " + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "text", + "content": " from diverse time points through masking, learning to generate the complete " + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 290, + 547, + 324 + ], + "type": "text", + "content": ". In this way, a single trained MMVG can perform all prediction, rewind, and infilling tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": "the last frame (rewind), or the head and tail (infilling), conditioned on the textual instruction. During training, we have pairs of videos " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": " and corresponding instructions " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": " frames as " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\{v_{1}, v_{2}, \\ldots, v_{N}\\}" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": ". Our goal is to train a unified model that generates the complete " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": " given the partial frames from arbitrary time points and " + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 331, + 288, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 251, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 251, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 251, + 426 + ], + "type": "text", + "content": "3.2. Multimodal Masked Video Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 602 + ], + "type": "text", + "content": "Overview. An overview of our Multimodal Masked Video Generation (MMVG) is illustrated in Fig. 2. To model the video along with language, we propose temporal-aware VQGAN to represent a frame as visual tokens, which converts it into the same discrete space as the words. We present an effective masking strategy that masks different video parts for video completion learning. Those missing fragments are replaced by the unique [SPAN] tokens, and we consider the visual guidance from diverse time points. The multimodal encoder consumes the text and the partial-missing video, and the decoder learns to produce the complete video from arbitrary guided frames. By varying the masking conditions, MMVG learns to utilize the [SPAN] token and unifies all TVC tasks during the training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Temporal-aware Discrete Visual Tokens. VQ-VAE [76] has shown promising capability in representing data as discrete tokens. VQGAN [54] further models the prior distribution of the latent space via a transformer with the GAN training. If VQGAN is directly applied onto videos, it will ignore the inner temporal coherence and treat each frame as an independent image, resulting in an unsmooth video reconstruction. Though TATS [23] attempts to handle this by making " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " consecutive frames altogether during VQ, it has" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 331, + 545, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 356 + ], + "type": "text", + "content": "to pre-define the constant " + }, + { + "bbox": [ + 304, + 331, + 545, + 356 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 331, + 545, + 356 + ], + "type": "text", + "content": " before training. Such constraint forbids it from representing a frame at any timestamp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "type": "text", + "content": "To address it with flexibility, we propose temporal-aware VQGAN (T-VQ) to inject the temporal relationship into the latent representation. We first follow VQGAN to learn the target visual tokens " + }, + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "type": "text", + "content": " by reconstructing a video frame " + }, + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 356, + 545, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 409, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 409, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 321, + 409, + 545, + 502 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {i} = \\mathbf {q} (\\mathsf {E n c} ^ {\\mathsf {Q}} (v _ {i}) \\mid C), \\\\ \\hat {v} _ {i} = \\operatorname {D e c} ^ {0} (\\mathrm {z} _ {i}), \\\\ \\mathcal {L} _ {\\mathrm {V Q}} = \\underbrace {\\left| \\left| \\hat {v _ {i}} - v _ {i} \\right| \\right| _ {1}} _ {\\text {r e c o n s t r c u t i o n}} + \\underbrace {\\left| \\left| \\operatorname {s g} [ \\operatorname {E n c} ^ {Q} (v _ {i}) ] - C _ {z _ {i}} \\right| \\right| _ {2} ^ {2}} _ {\\text {c o d e b o o k}} \\tag {1} \\\\ + \\underbrace {\\beta | | \\mathrm {s g} [ C _ {\\mathrm {z} _ {i}} ] - \\mathrm {E n c} ^ {\\mathrm {Q}} (v _ {i}) | | _ {2} ^ {2}} _ {\\text {c o m m i t}} + \\underbrace {| | \\mathcal {F} (\\hat {v} _ {i}) - \\mathcal {F} (v _ {i}) | | _ {1}} _ {\\text {m a t c h i n g}}, \\\\ \\end{array}", + "image_path": "805a6d423fa35a7210d090c8000734132353e4f0b0c7816a8449649ffd76c879.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}^{\\mathrm{Q}}" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}^{\\mathrm{Q}}" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " are the VQ encoder and decoder. The discrete latent code " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " is acquired from the quantization operation q [54], which adopts nearest neighbor search by the trainable codebook " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": ". We apply the straight-through estimator over the stop-gradient operation sg and use " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " as 0.25 [76]. We also append VGG [67] features matching " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " to stabilize the VQ loss " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}}" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " [23]. The adversarial training between the frame quality loss " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " and discrimination loss " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": " are further calculated from the discriminator " + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 509, + 545, + 617 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 623, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 623, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 349, + 623, + 545, + 651 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {G} = \\log (1 - D (\\hat {v} _ {i})), \\\\ \\mathcal {L} _ {D} = \\log (1 - D (\\hat {v} _ {i})) + \\log (D (v _ {i})). \\end{array} \\tag {2}", + "image_path": "9f2b915fd9066a4682359d6c162567f2f25469b71b6128100dbf8cd3896ac899.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "text", + "content": "To inject the temporal relationship into " + }, + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 306, + 656, + 545, + 681 + ], + "type": "text", + "content": ", T-VQ is trained with the introduced contrastive temporal reasoning:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 686, + 545, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 686, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 351, + 686, + 545, + 706 + ], + "type": "interline_equation", + "content": "o _ {i} = \\mathbf {F C} ^ {\\mathrm {T}} \\left(\\mathbf {z} _ {i}, \\mathbf {z} _ {j}\\right), \\tag {3}", + "image_path": "92a33a2d76c1bc06f68364147c8db6f6461a9b958c56ce38d42de8fb0c317c7e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 704, + 501, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 704, + 501, + 715 + ], + "spans": [ + { + "bbox": [ + 351, + 704, + 501, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T}} = \\operatorname {B C E L o s s} \\left(o _ {i}, 0 \\text {i f} i > j \\text {e l s e} 1\\right),", + "image_path": "5cea151b3ad52710c27a3f8ca8a6cbb47d32ab9443bb1d4445939856a618d593.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "content": " is a random frame from the same video. " + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{FC^T}" + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "content": " is the MLP classifier, and BCELoss is the binary cross-entropy for before/after. Learning the temporal order from " + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "content": " facilitates an implicit temporal coherence, leading to smooth video modeling. Moreover, since " + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 71, + 287, + 144 + ], + "type": "text", + "content": " represents a single image, it is flexible to support frames at arbitrary timestamps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "content": "Generation from Masked Video. We propose the masking strategy " + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "content": " to obtain the masked videos " + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "inline_equation", + "content": "\\overline{\\nu}" + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "content": " from diverse time points. " + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "content": " masks out most video frames with the probability " + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "content": " and replaces each fragment as a unique [SPAN] token. For example, " + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 147, + 288, + 218 + ], + "type": "text", + "content": " reserves the third and the fifth frame, and masks all the others over a video length of 5:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 227, + 287, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 227, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 97, + 227, + 287, + 242 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {V}}} \\colon \\{[ S ], v _ {3}, [ S ], v _ {5} \\} = \\mathcal {M} (\\mathcal {V} \\mid p). \\tag {4}", + "image_path": "3225e57a9d191752ecd86801f0068aaf1de1a4a1e3c031bb0bb968f5f3cbbbf6.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": "Our goal is to recover the missing part from " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": " and perform video completion, guided by the instruction " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": ". To model between the vision and language modalities, we apply our Enc" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "^{\\mathrm{Q}}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{V}}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": " for the discrete visual tokens " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{S}],\\mathrm{z}_3,[\\mathrm{S}],\\mathrm{z}_5\\}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": ". We also tokenize the text " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": " into word tokens " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\{w_i\\}_{i = 1}^L" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": " with the CLIP tokenizer [56], where " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": " is the length of " + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": ". As in the same discrete space, MMVG can achieve cross-modal fusion by the multimodal encoder (Enc" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "inline_equation", + "content": "^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 251, + 288, + 358 + ], + "type": "text", + "content": ") through the self-attention mechanism as the transformer [77]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 367, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 367, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 102, + 367, + 287, + 388 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {i} ^ {w}, f _ {j} ^ {v} = \\mathrm {L P} ^ {w} \\left(w _ {i}\\right), \\mathrm {L P} ^ {v} \\left(\\mathrm {z} _ {j}\\right) \\\\ \\{1 \\} = \\sum_ {i = 1} ^ {n} M (f _ {i}) w _ {i} \\end{array} \\tag {5}", + "image_path": "43dffc0546cc5e88f1be3fabd516ea9a19cfb295ae167838c63159724e2378df.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 383, + 231, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 383, + 231, + 397 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 231, + 397 + ], + "type": "interline_equation", + "content": "\\{h \\} = \\operatorname {E n c} ^ {\\mathrm {M}} ([ \\{f ^ {w} \\}, \\{f ^ {v} \\} ]),", + "image_path": "a2b4a7809ee3d477376774bc5d1a50c4b6ceddf0cb08a904e554032a2208247b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "type": "text", + "content": "where it obtains the features " + }, + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "type": "text", + "content": " by the linear projection (LP), and " + }, + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 406, + 287, + 453 + ], + "type": "text", + "content": " is the hidden encoding features. We can also regard LP as the video/language embedder, which extracts the preliminary visual/linguistic features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "text", + "content": "After encoding the language hint and the partial-missing video from Enc" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "inline_equation", + "content": "^{\\text{M}}" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "text", + "content": ", our video decoder (Dec" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "inline_equation", + "content": "^{\\text{M}}" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "text", + "content": ") learns to produce all frames for comprising the complete video. Dec" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "inline_equation", + "content": "^{\\text{M}}" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "text", + "content": " follows the vanilla autoregressive decoder, which first conducts self-attention over the past generated tokens and then predicts the discrete visual tokens as the video frame, conditioned on the encoded features " + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 454, + 287, + 536 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 101, + 544, + 237, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 544, + 237, + 559 + ], + "spans": [ + { + "bbox": [ + 101, + 544, + 237, + 559 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t} = \\operatorname {D e c} ^ {\\mathrm {M}} (\\left\\{\\hat {z} _ {1}, \\dots , \\hat {z} _ {t - 1} \\right\\} \\mid \\{h \\}),", + "image_path": "eb286b5c267547d086de2980e3b7df45ff4f971e2c86d273a2fa43b13fd08063.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 101, + 561, + 287, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 561, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 101, + 561, + 287, + 582 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\underset {N} {\\operatorname {C E L o s s}} \\left(\\hat {z} _ {t}, z _ {t}\\right), \\tag {6}", + "image_path": "ee325a7db8d4730ecac69389499ca6297a1d44e6f9c45552e44457ec92897f37.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 578, + 155, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 578, + 155, + 609 + ], + "spans": [ + { + "bbox": [ + 97, + 578, + 155, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M}} = \\sum_ {t = 1} ^ {N} \\mathcal {L} _ {t},", + "image_path": "d9d2fe68f0b3660c5ad3cd8aa218108f313cb68ab8e8b571cae3b83ccee3e9e5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is the ground-truth tokens of the frame " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " in the original " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". We calculate the video decoding loss " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " by the cross-entropy (CELoss) to learn video generation as classification. Our " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " is built upon VideoSwin [45], which has shown a strong visual perception on various video understanding tasks. The 3D-shifted windows [44] consider different levels of spatial-temporal attention, and each window models blocks of video patches across " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " consecutive" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 306, + 85, + 545, + 372 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 518, + 83 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 518, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 518, + 83 + ], + "type": "text", + "content": "Algorithm 1 Multimodal Masked Video Generation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": "1: while TRAIN_T-VQ do \n2: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " sample video \n3: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "z_{i} = q(\\mathrm{Enc}^{Q}(v_{i})\\mid C)" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\hat{v}_i = \\mathrm{Dec}^{\\mathrm{Q}}(z_i)" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "o_i = \\mathrm{FC}^{\\mathrm{T}}(z_i,z_j)" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " ▷ randomly sampled frame " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}},\\mathcal{L}_G\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " reconstruction, frame quality loss " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " Eq. 1&2 \n7: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " temporal ordering loss " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " Eq. 3 \n8: Update T-VQ by minimizing " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}} + \\mathcal{L}_G + \\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " discrimination loss " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " Eq. 2 \n10: Update D by maximizing " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n11: end while \n12: \n13: while TRAIN_MMVG do \n14: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{V},\\mathcal{X},p\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " sample video/instruction/probability \n15: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{V}}:\\{v_a,[S],v_b,\\ldots\\} = \\mathcal{M}(\\mathcal{V}\\mid p)" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " diverse guided frames \n16: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_a,[S],\\mathbf{z}_b,\\ldots\\},\\{w\\} = \\mathrm{Enc}^{\\mathrm{Q}}(\\overline{\\mathcal{V}})," + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " Tokenizer(X) \n17: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\{h\\} = \\mathrm{Enc}^{\\mathrm{M}}([\\{w\\},\\{z_a,[S],z_b,\\ldots\\}])" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n18: for " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "t\\gets 1" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " do \n19: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\hat{z}_t = \\mathrm{Dec}^{\\mathrm{M}}(\\{z_1,\\dots,z_{t-1}\\}\\mid\\{h\\})" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " teacher-forcing \n20: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " video decoding loss " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " Eq. 6 \n21: end for \n22: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\hat{\\nu} = \\mathrm{Dec}^{\\mathrm{Q}}(\\{\\hat{z}_{t=1}^N\\})" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n23: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}} = \\sum_{t=1}^{N}\\mathcal{L}_{t}" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n24: Update MMVG by minimizing " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}}" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " \n25: " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "p\\gets" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " update masking probability " + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 85, + 545, + 372 + ], + "type": "text", + "content": " Eq. 8 \n26: end while" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "text", + "content": "frames. To ensure the same dimension for video generation in " + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}^{\\mathrm{M}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "text", + "content": ", we remove the temporal down-sampling layer. In the end, we can utilize " + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}^{\\mathrm{Q}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "text", + "content": " to reconstruct all the frames as our completed videos " + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 304, + 392, + 545, + 440 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 447, + 545, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 447, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 383, + 447, + 545, + 462 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {V}} = \\operatorname {D e c} ^ {\\mathrm {Q}} \\left(\\left\\{\\hat {z} \\right\\} _ {t = 1} ^ {N}\\right). \\tag {7}", + "image_path": "9becec6bcbbc2fe52a3502d5d6e6181d8c53a4e2d54b6d7fc4e547e3ba97a0c9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": "By varying the masking conditions through " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": ", MMVG learns how to complete a video from partial frames " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\overline{\\nu}" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": " at arbitrary time points with the text, which overcomes the limitation of chronological guidance. To make " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": " more effective, we apply an adaptive probability " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": " instead of random sampling every time. Each video " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": " keeps its own " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": ", and all frames are equally initialized in the beginning. Based on the prediction error, we adjust the masking probability " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 471, + 545, + 577 + ], + "type": "text", + "content": "-th frame:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 583, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 583, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 359, + 583, + 545, + 609 + ], + "type": "interline_equation", + "content": "p _ {t} = p _ {t} + \\alpha \\left(\\left(\\frac {\\mathcal {L} _ {t}}{\\mathcal {L} _ {M}} \\sum p\\right) - p _ {t}\\right), \\tag {8}", + "image_path": "b61a296445ce33792b6d41e2d3dec2d982900765006b11b4542225d2933f4d5c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " is the adjusting rate. A larger video decoding loss " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " indicates that the " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": "-th frame is more difficult to recover. MMVG learns more from those challenging cases and can bring better generative quality for video completion." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Unifying TVC during Inference. After training with text and partial-missing video, MMVG learns to perform video completion over [SPAN] tokens. Then for inference, Enc" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{\\text{M}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " takes the following as its input to support different tasks:" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10684" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 196 + ], + "type": "table", + "html": "
MethodTextTVPredictionTVRewindTVInfilling
KitchenFlintstonesMUGENKitchenFlintstonesMUGENKitchenFlintstonesMUGEN
FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑FVD↓RCS↑
FILM [60]✘------------250.256.1352.751.4538.85.9
VideoMAE [73]✘328.947.6317.555.6548.77.0365.948.2335.555.9545.27.1246.954.7211.560.6494.97.8
TATS [23]✘106.964.4127.560.3376.57.1107.762.7127.660.2350.87.271.572.7119.566.7328.28.4
MMVGU✘105.663.3124.860.5374.57.2109.862.6124.359.7356.47.071.573.4121.866.3328.47.8
MMVGS✘103.864.5123.860.8369.47.3105.963.6123.860.5347.87.268.573.6118.567.9324.38.4
TATS [23]✓87.266.3115.970.690.167.989.863.3116.370.489.868.757.477.695.878.258.973.6
MMVGU✓80.268.4108.272.984.870.283.266.9113.271.693.168.459.877.892.878.359.273.2
MMVGS✓75.668.8106.373.783.371.179.768.1107.272.988.770.056.078.191.679.657.274.1
", + "image_path": "89cb21800a614ffb7186a8866114b97b41a1dbcfa173611dd84bbede6ae29dcc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 231, + 209, + 267 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 231, + 201, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 201, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 201, + 243 + ], + "type": "text", + "content": "- TVPrediction: " + }, + { + "bbox": [ + 47, + 231, + 201, + 243 + ], + "type": "inline_equation", + "content": "\\left[\\left\\{w\\right\\},\\left\\{z_{1},[SPAN]\\right\\}\\right]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 243, + 193, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 193, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 193, + 255 + ], + "type": "text", + "content": "- TVRewind: " + }, + { + "bbox": [ + 47, + 243, + 193, + 255 + ], + "type": "inline_equation", + "content": "\\left[\\left\\{\\boldsymbol{w}\\right\\},\\left[\\left[\\mathrm{SPAN}\\right],\\mathbf{z}_{N}\\right]\\right]" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 255, + 209, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 209, + 267 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 209, + 267 + ], + "type": "text", + "content": "TVInfilling: " + }, + { + "bbox": [ + 47, + 255, + 209, + 267 + ], + "type": "inline_equation", + "content": "\\left[\\left\\{w\\right\\} ,\\left\\{\\mathbf{z}_1,\\left[\\mathrm{SPAN}\\right],\\mathbf{z}_N\\right\\}\\right]" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 268, + 286, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 286, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 286, + 292 + ], + "type": "text", + "content": "In this way, a single trained MMVG can unify all TVC tasks without the specific downstream fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 297, + 165, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 165, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 165, + 309 + ], + "type": "text", + "content": "3.3. Learning of MMVG" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": "Algo. 1 illustrates the learning process of our proposed MMVG for TVC. We first train T-VQ over video frames for discrete visual tokens with contrastive temporal reasoning. Specifically, we minimize the VQ reconstruction loss " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}}" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " and frame quality loss " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_G" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " along with our temporal ordering loss " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " to optimize T-VQ. At the same time, we also update the discriminator " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " via the standard adversarial training by maximizing the discrimination loss " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": ". For video completion, the masking strategy " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " masks the video frames with the probability " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " and then acquires guided frames from diverse time points. MMVG regards text and partial-missing video by Enc for cross-modal fusion, and Dec further predicts the visual tokens of frames autoregressively as the complete video. As a sequential generation process, we apply the teacher-forcing trick. Instead of our predicted " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": ", the ground-truth " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " from the previous timestamp is fed to stabilize the training. Each video decoding loss " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " at timestamp " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " is summed up as " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " to optimize MMVG. According to " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": ", we update " + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": " for effective masking probability. The entire optimization object can be summarized as two phases:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 266, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 266, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 266, + 598 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {T - V Q}: \\min _ {\\mathrm {E n c} ^ {\\mathrm {Q}}, \\mathrm {D e c} ^ {\\mathrm {Q}}, C, \\mathrm {F C} ^ {\\mathrm {T}}} \\max _ {D} \\mathcal {L} _ {\\mathrm {V Q}} + \\mathcal {L} _ {G} + \\mathcal {L} _ {D} + \\mathcal {L} _ {\\mathrm {T}} \\\\ \\text{MMVG:}\\min_{\\substack{\\text{Enc}^{\\text{M}},\\text{Dec}^{\\text{M}}}}\\mathcal{L}_{\\text{M}} \\\\ \\end{array}", + "image_path": "6ce067d4c266e01c2308ae95a2e7eb9c2e16049018bea0a095b3b85aa5b1ca17.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 623, + 163, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 163, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 163, + 636 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Datasets. As a new task, we consider diverse video scenes with natural instructions for TVC. Kitchen [13] records 22K egocentric videos about kitchen activity, which have different lengths (4-16 frames) with narrations. Flintstones [26] contains 25K animation videos (15 frames) from The Flintstones, where each video description includes the characters" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 331, + 229, + 520, + 284 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "type": "text", + "content": "Table 1. Results of TVC, including prediction, rewind, and infilling, on Kitchen [13], Flintstones [26], and MUGEN [29]. TATS [23] requires specific training to support different tasks. We further train the unified MMVG" + }, + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "type": "inline_equation", + "content": "^{\\text{U}}" + }, + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "type": "text", + "content": " for each specific task as MMVG" + }, + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "type": "inline_equation", + "content": "^{\\text{S}}" + }, + { + "bbox": [ + 46, + 198, + 547, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 229, + 520, + 284 + ], + "lines": [ + { + "bbox": [ + 331, + 229, + 520, + 284 + ], + "spans": [ + { + "bbox": [ + 331, + 229, + 520, + 284 + ], + "type": "table", + "html": "
DatasetTrain / Val#Frame#WordFPS
Kitchen [13]16,695 / 5,8048.32.86
Flintstones [26]22,666 / 2,5181516.55
MUGEN [29]362,239 / 12,8481620.65
", + "image_path": "031f12fab0b218d434a6cb4c1c78a8713d8aacb8d93d8a79b76ddc4eefccdecb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 285, + 531, + 296 + ], + "lines": [ + { + "bbox": [ + 319, + 285, + 531, + 296 + ], + "spans": [ + { + "bbox": [ + 319, + 285, + 531, + 296 + ], + "type": "text", + "content": "Table 2. The statistics of our used datsets to evaluate TVC." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "type": "text", + "content": "and their behavior. MUGEN [29] is built from agents playing CoinRun [11], which consists of 375K gaming videos (16 frames) with detailed text annotations. All videos in these three datasets are resized into " + }, + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 308, + 545, + 416 + ], + "type": "text", + "content": ". An overview is shown in Table 2 and Fig. 1(b). Since MMVG can unify various tasks, we also evaluate video generation/prediction on widely-used UCF-101 [69] and BAIR [18], video infilling on UCF-101 following RaMViD [39], and text-to-video generation on MSRVTT [89]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 418, + 546, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 574 + ], + "type": "text", + "content": "Evaluation Metrics. We apply the following metrics to evaluate TVC results: 1) FVD [75] computes the video features [6] distance to the ground truth; 2) RCS [86] is the relative visual-text similarity to the instruction, compared to the ground-truth video. We fine-tune the CLIP model [56] on each dataset and adapt it to the video scene for a more precise alignment. Apart from automatic metrics, we also conduct a human evaluation from aspects of video quality, instruction relevance, and ground-truth similarity. We sample 75 TVP results for each task and adopt MTurk1 to rank over baselines and our MMVG. To avoid the potential ranking bias, we hire 3 MTurkers for each sample of prediction, rewind, and infilling tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": "Implementation Detail. T-VQ contains ResBlocks [31] as the visual auto-encoder (Enc" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "inline_equation", + "content": "^{\\text{Q}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": " and Dec" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "inline_equation", + "content": "^{\\text{Q}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": "). The discriminator " + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": " follows a similar architecture to Enc" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "inline_equation", + "content": "^{\\text{Q}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": ". For the vector quantization, we use a patch size 16, where a 128x128 video frame transforms into 8x8 discrete visual tokens. There are 1024 vocabularies in the codebook " + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": ", and the hidden embedding size is 256. We adopt batch size 32 with a learning rate of 4.5e-6 to optimize T-VQ by Adam [42]. MMVG is built in an encoder-decoder manner, where Enc" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "inline_equation", + "content": "^{\\text{M}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": " is a trans" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "Amazon MTurk: https://www.mturk.com. Our studies have been cleared by the human subject committee as an IRB-exempt protocol." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10685" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 70, + 248, + 135 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 248, + 135 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 248, + 135 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 248, + 135 + ], + "type": "table", + "html": "
MethodKitchenFlintstonesMUGEN
VideoDiff [34]138.6206.4410.7
MCVD [79]119.9183.8400.2
TATS [23]115.5157.5386.4
MMVG109.1127.6368.6
", + "image_path": "26bdefd1ba60f8306286c71e21c386b97f60830a3e3647c152ae09d6de27a468.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "text", + "content": "former with 24 layers, 16 attention heads, and hidden embedding size 1024. " + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "text", + "content": " adopts a similar setting with temporal window size 3 in VideoSwin. The initial sample rate " + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "text", + "content": " of the masking strategy " + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "text", + "content": " is 0.9 with an adjusting rate " + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 159, + 287, + 254 + ], + "type": "text", + "content": " as 0.1. We optimize MMVG through the mixed precision [48] with batch size 4 by Adam. The learning rate is also 4.5e-6. All experiments are implemented in PyTorch [53] and done on 8 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 262, + 132, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 132, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 132, + 274 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": "Table 1 shows the results of all text-guided prediction, rewind, and infilling for TVC. VideoMAE [73] is built upon MAE [30] and reconstructs the missing video cubes, which performs TVC by masking all video frames except the first or the last (or both). TATS [23], the SOTA on video generation, also produces videos as discrete visual tokens. Since TATS can only consider the past through the autoregressive transformer, it requires specific training for each task. We have MMVG" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "^{\\text{U}}" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": " as the unified model that can support all TVC tasks with a single training and MMVG" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "^{\\text{S}}" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": " to further train for each prediction, rewind, and infilling. We treat TATS as our main baseline" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 46, + 281, + 287, + 425 + ], + "type": "text", + "content": " and study the importance of guided text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 666 + ], + "type": "text", + "content": "TVPrediction. VideoMAE attempts to produce all frames simultaneously, which is difficult to maintain video temporal consistency, resulting in a high 328.9 FVD on Kitchen. TATS is inherently designed for prediction as it generates the frames one after one. However, our unified MMVGU performs better than TATS on all datasets (e.g., lower 105.6 and 124.8 FVD on Kitchen and Flintstones). These results support that learning from diverse time points will not hurt the prediction from the past. In contrast, our masking strategy can bring superior temporal coherence. MMVGS further improves itself through training prediction as completion from the head. However, there are too many possible outcomes from just the beginning, where the predicted results may not meet the expectation (e.g., a high 370 FVD on MUGEN). The instruction as guidance makes it related to the expected ground-truth result. We can let MUGEN run, jump, or collect coins as the textual descriptions to achieve more controllability, leading to a noticeable improvement (e.g., a lower 84.8 FVD by MMVGU). The higher 70.2 RCS also shows that our MMVG can produce MUGEN videos" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 325, + 70, + 424, + 160 + ], + "blocks": [ + { + "bbox": [ + 51, + 137, + 282, + 148 + ], + "lines": [ + { + "bbox": [ + 51, + 137, + 282, + 148 + ], + "spans": [ + { + "bbox": [ + 51, + 137, + 282, + 148 + ], + "type": "text", + "content": "Table 3. FVD results of video generation on our TVC datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 70, + 424, + 160 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 424, + 160 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 424, + 160 + ], + "type": "table", + "html": "
MethodUCF-101
IS↑FVD↓
VideoGPT [94]24.7-
DIGAN [96]32.7577
VideoDiff [34]57.0-
TATS [23]57.6420
MMVG58.3395
", + "image_path": "f99134fe919406babff284fa6354c8e8b17b2542bc5ce050e73fc41295e5ff09.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 456, + 70, + 539, + 160 + ], + "blocks": [ + { + "bbox": [ + 456, + 70, + 539, + 160 + ], + "lines": [ + { + "bbox": [ + 456, + 70, + 539, + 160 + ], + "spans": [ + { + "bbox": [ + 456, + 70, + 539, + 160 + ], + "type": "table", + "html": "
MethodBAIR
FVD↓
VideoGPT [94]103.3
MaskViT [25]93.6
MCVD [79]89.5
TATS [23]88.6
MMVG85.2
", + "image_path": "13db1d95cea444e2f6977f895e93de00da5a67dd55e412ba9142f097880e873f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 446, + 162, + 548, + 184 + ], + "lines": [ + { + "bbox": [ + 446, + 162, + 548, + 184 + ], + "spans": [ + { + "bbox": [ + 446, + 162, + 548, + 184 + ], + "type": "text", + "content": "Table 5. Results of video prediction on BAIR [18]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 163, + 444, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 163, + 444, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 444, + 184 + ], + "type": "text", + "content": "Table 4. Results of video generation on UCF-101 [69]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "type": "text", + "content": "that confirm with the instruction. Although the model may try to imagine the animation or the kitchen activity, the language hint can provide a clear goal to anticipate. Likewise, MMVG" + }, + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "type": "inline_equation", + "content": "^{\\text{U}}" + }, + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "type": "text", + "content": " with text surpasses TATS, even though it is not designed for prediction only. The specific trained MMVG" + }, + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "type": "inline_equation", + "content": "^{\\text{S}}" + }, + { + "bbox": [ + 304, + 193, + 545, + 264 + ], + "type": "text", + "content": " benefits the unified model for further improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "type": "text", + "content": "TVRewind. Rewind from the last allows the model to imagine what happened along with a suitable opening. In addition, the objects may not display on the last frame (e.g., the spoons and forks for \"close drawer\"), which makes it more challenging to complete. Similar to prediction, VideoMAE cannot have feasible rewind results. Language is still essential to remind the past and establish an adequate beginning, where we can find a significant performance gap between with and without text (e.g., 90 vs. 350 FVD on MUGEN). Our unified MMVG" + }, + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "type": "inline_equation", + "content": "^{\\text{U}}" + }, + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "type": "text", + "content": " achieves comparable results to TATS and even outperforms on Kitchen and Flintstones (e.g., higher 66.9 and 71.6 RCS). With the learning of completion from partial frames, autoregressive model can still accomplish video rewind without specific training. If following TATS design to train MMVG" + }, + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "type": "inline_equation", + "content": "^{\\text{U}}" + }, + { + "bbox": [ + 304, + 266, + 546, + 458 + ], + "type": "text", + "content": " for rewind, MMVGS gains more improvement and utterly surpasses it." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "type": "text", + "content": "TVInfilling. We consider the additional FILM [60] for infilling, which performs video interpolation with in-between motion. Despite synthesizing intermediate frames between the first and the last, the visual dynamics are changing too rapidly to handle, resulting in a higher FVD. With guidance from the head and tail, we find a noticeable improvement even without instruction (e.g., lower FVDs on Kitchen), which is helpful in temporal video modeling. To capture the expected missing event, we still require the language hint for more controllability. Our unified MMVG" + }, + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "type": "inline_equation", + "content": "^{\\text{U}}" + }, + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "type": "text", + "content": " achieves comparable performance to TATS again, which is specifically trained for the infilling task. It shows that completion from partial frames at different time points still helps, and MMVG" + }, + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "type": "inline_equation", + "content": "^{\\text{S}}" + }, + { + "bbox": [ + 304, + 460, + 545, + 628 + ], + "type": "text", + "content": " further outperforms on TVInfilling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 635, + 409, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 409, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 409, + 647 + ], + "type": "text", + "content": "4.3. Additional Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Video Generation/Prediction. We further evaluate the classic video generation and prediction tasks. Table 3 shows FVD scores of unconditional video generation on our TVC datasets. Note that only videos but no texts are used in these experiments. Both VideoDiff [34] and MCVD [79] are built" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 286, + 713 + ], + "type": "text", + "content": "2Since we cannot receive feasible results after training diffusion methods for our TVC, we evaluate unconditional video generation in Sec. 4.3. We use this repo (https://github.com/lucidrains/video-diffusion-pytorch) as VideoDiff and the official repo as MCVD." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10686" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 248, + 131 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 248, + 131 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 248, + 131 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 248, + 131 + ], + "type": "table", + "html": "
MethodUCF-101 (FVD↓)
K=+1+2+5±1±2
RaMViD [39]349.7300.6260.5215.4162.5
MMVG316.3258.5194.6183.2120.3
", + "image_path": "16ba09493c37618a7ef190d4d39597c36d3958197395b371c661ea2eece24d4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 155, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 155, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 286, + 251 + ], + "type": "text", + "content": "upon denoising diffusion [33], where MCVD also considers different partial frames during training. The results first indicate that the vanilla token-based method is superior to the diffusion-based model (TATS vs. VideoDiff) for video generation. In addition, MMVG, with the masking strategy that learns the visual guidance from diverse time points, further boosts the performance (the lowest 127.6, 109.1, and 368.6 FVD on Flintstones, Kitchen, and MUGEN, respectively)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 407 + ], + "type": "text", + "content": "We also evaluate MMVG on UCF-101 [69], which is challenging to generate natural human videos. Table 4 supports that our MMVG can produce videos with higher visual similarity (a higher 58.3 IS) and temporal alignment (a lower 395 FVD) to the ground truth. For video prediction, we apply the widely-used BAIR [18] in Table 5, where the model has to anticipate how a robot pushes objects from the given first frame. MMVG again surpasses TATS with the lowest 85.2 FVD. Although both generation and prediction are generating video frames chronologically, the ability to recover arbitrary missing frames for video completion empowers MMVG with a stronger temporal coherence, leading to better generative video quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": "Video Infilling. We follow RaMViD [39] to evaluate video infilling on UCF-101. We consider various guidance settings " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": " in Table 6. For example, " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K = +1" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": " means given the first frame, and " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K = \\pm 2" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": " should provide the first and last two frames. For prediction, MMVG outperforms RaMViD on all " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": ", and the performance gap gets even larger when more guided frames are accessible (e.g., 33.4 on " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K = +1" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": " and 65.9 on " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K = +5" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": "). A similar result can be found for infilling, where MMVG can make the lowest 120.3 FVD on " + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "inline_equation", + "content": "K = \\pm 2" + }, + { + "bbox": [ + 46, + 411, + 286, + 578 + ], + "type": "text", + "content": ". Despite having a similar masking strategy, it shows that generating frames one after one still brings superior results. MMVG allows autoregressive models to condition on visual hints from any time point, which produces more similar videos to the ground truth when infilling between the head and tail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Text-to-Video Generation. Being a multimodal generative model, MMVG supports text-to-video generation. To compare with those large-scale methods, we pre-train MMVG using WebVid [4], which contains 2.5M text-video pairs. We adopt the masking strategy to treat the pre-training as video completion. MMVG outperforms GODIVA [86] and NUWA [87] without access the MSRVTT [89] data in Table 7. Surprisingly, MMVG can generate videos that are more related to the texts (a higher 0.2644 CLIP-S [86]) thanCogVideo [35], even though using twice less data. This result encourages the effectiveness of completion from partial" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 325, + 70, + 525, + 186 + ], + "blocks": [ + { + "bbox": [ + 51, + 131, + 282, + 142 + ], + "lines": [ + { + "bbox": [ + 51, + 131, + 282, + 142 + ], + "spans": [ + { + "bbox": [ + 51, + 131, + 282, + 142 + ], + "type": "text", + "content": "Table 6. Results of video prediction and infilling on UCF-101." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 70, + 525, + 186 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 525, + 186 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 525, + 186 + ], + "type": "table", + "html": "
MethodPre-trainingZero-shotMSRVTT
FID↓CLIP-S↑
GODIVA [86]136MX-0.2402
NUWA [87]3.9MX47.70.2439
CogVideo [35]5.4M✓23.60.2631
Make-A-Video [68]20M✓13.20.3049
MMVG2.5M✓23.40.2644
TATS [23]-X63.20.2326
MMVG-X60.60.2385
", + "image_path": "d827b65f5d15a938e25b3d6b6b721a6e45ffd561617d0fe3e678773bfb74ef83.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 218, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 306, + 188, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 306, + 188, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 545, + 211 + ], + "type": "text", + "content": "Table 7. Results of text-to-video generation on MSRVTT [89]. We gray out methods that use significantly more pre-training data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 218, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 289 + ], + "type": "table", + "html": "
MethodTextKitchenFlintstonesMUGEN
Q.T.GTQ.T.GTQ.T.GT
MMVG✘1.991.811.821.731.661.622.031.561.55
TATS [23]✓1.972.072.032.072.122.171.942.112.19
MMVG✓2.042.122.152.202.222.212.032.332.26
", + "image_path": "0bdb589d447fc28bd8ba5ae729f86c7d0853e568b824d9e6b7e5f14a79d94ea0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 291, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 291, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 291, + 545, + 314 + ], + "type": "text", + "content": "Table 8. Human evaluation for TVP with aspects of video quality (Q.), instruction relevance (T.), and ground-truth similarity (GT)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 324, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 373 + ], + "type": "text", + "content": "frames. For a fair comparison without additional data, we directly train on MSRVTT. MMVG still outperforms TATS, which shows that text-to-video generation can be improved through learning from video completion as well." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 375, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 543 + ], + "type": "text", + "content": "Human Evaluation. We study the video quality (Q.), the relevance to the instructions (T.), and the similarity to the ground-truth video (GT) of the produced videos from the human aspect. The results in Table 8 are calculated as the mean ranking score (from 1 to 3, the higher is better) of each method for TVPrediction. MMVG without text even generates higher quality videos than TATS with text on Kitchen and MUGEN, where completion from partial frames benefits the temporal coherence of generative video modeling. While, the lowest ground-truth similarity illustrates that language guidance is crucial for controllability. With instruction, MMVG anticipates the future as the text (the highest T.) and generates videos that meet the ground truth (the highest GT), achieving the best overall performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": "Qualitative Results. Fig. 3 illustrates the keyframes of the generated examples of three datasets. Thanks to the learning of completion from partial frames at diverse time points, a single trained MMVG can support all TVC tasks. For prediction, MMVG makes Fred \"turn his head\" or MUGEN \"jumps over the gear\" from the guided text. MMVG further recovers the missing spoons and forks for \"close drawer\" from the last frame in a more challenging rewind scenario. MMVG infills the missing event described in language such as \"stand up for dancing\", \"walk across the kitchen\", \"jump onto the stage\" from the head and tail." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "From the same visual guidance, MMVG can lead TVC results using different texts, achieving controllable video completion. For example, we can let MUGEN \"jump down" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10687" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 527 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 527 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 527 + ], + "type": "image", + "image_path": "25f55c810660193d84f63031116c2a08db8790412d864d211dcf0dae800a38a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 529, + 546, + 552 + ], + "lines": [ + { + "bbox": [ + 47, + 529, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 546, + 552 + ], + "type": "text", + "content": "Figure 3. Qualitative examples of TVC on Kitchen [13], Flintstones [26], and MUGEN [29]. We also illustrate video generation on UCF-101 [69], video prediction on BAIR [18], and text-to-video prediction on WebVid [4]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 558, + 289, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 679 + ], + "type": "text", + "content": "the ground\" or \"land on a face\", starting from the same beginning. We can also control the behavior as \"keep walking\" or \"jump over a gear\" to recover the missing middle event. Furthermore, MMVG also carries out unconditional video generation with smooth temporal coherence. We can use language to produce natural dynamics in diverse scenes, such as \"sunset on the sea\", \"green sea turtle swims\", or a close look of \"cut chicken\". The presented videos indicate that our method not only unifies all TVC tasks but also performs the classical video generation/text-to-video well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 557, + 380, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 380, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 380, + 570 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 577, + 547, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 547, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 547, + 710 + ], + "type": "text", + "content": "We introduce a novel task of text-guided video completion (TVC) that performs video completion from the first, last, or both frame(s) controlled by language. We present Multimodal Masked Video Generation (MMVG) with an effective masking strategy to learn the visual guidance from any time point. By varying the masking conditions, MMVG addresses all prediction, rewind, and infilling tasks within one model. Experiments on various video scenes show that our MMVG effectively addresses TVC as well as generative video modeling. We believe TVC can help advance a new field toward vision-and-language research." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10688" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards High Resolution Video Generation with Progressive Growing of Sliced Wasserstein GANs. In arXiv:1810.02419, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "text", + "content": "[2] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H. Campbell, and Sergey Levine. Stochastic Variational Video Prediction. In International Conference for Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "text", + "content": "[3] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. FitVid: Overfitting in Pixel-Level Video Prediction. In arXiv:2106.13195, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 269 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 312 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2LIVE: Text-Driven Layered Image and Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "text", + "content": "[7] Ya-Liang Chang, Zhe Liu Yu, and Winston Hsu. VOR-Net: Spatio-temporally Consistent Video Inpainting for Object Removal. In Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 435 + ], + "type": "text", + "content": "[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent Online Video Style Transfer. In International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "text", + "content": "[9] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Jonathan Warrell, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy Mitra, and Philip Torr. ImageSpirit: Verbal Guided Image Parsing. In Conference on Human Factors in Computing Systems (CHI), 2013. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "text", + "content": "[10] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial Video Generation on Complex Datasets. In arXiv:1907.06571, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[11] Karl Cobbe, Oleg Klimov, Chris Hesse, Taehoon Kim, and John Schulman. Quantifying Generalization in Reinforcement Learning. In International Conference on Machine Learning (ICML), 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "text", + "content": "[12] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. VQGAN-CLIP: Open Domain Image Generation and Editing with Natural Language Guidance. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 690 + ], + "type": "text", + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling Egocentric Vision: The EPIC-KITCHENS Dataset. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[14] Yingying Deng, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, and Changsheng Xu. Arbitrary Video Style" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Transfer via Multi-Channel Correlation. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[15] Emily Denton and Rob Fergus. Stochastic Video Generation with a Learned Prior. In International Conference on Machine Learning (ICML), 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 150 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alex Nichol. Diffusion Models Beat GANs on Image Synthesis. In arXiv:2105.05233, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering Text-to-Image Generation via Transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 206, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 250 + ], + "type": "text", + "content": "[18] Frederik Ebert, Chelsea Finn, Alex X. Lee, and Sergey Levine. Self-Supervised Visual Planning with Temporal Skip Connections. In Conference on Robot Learning (CoRL), 2017. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "type": "text", + "content": "[19] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W.Taylor. Tell, Draw, and Repeat: Generating and Modifying Images Based on Continual Linguistic Instruction. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "text", + "content": "[20] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. SSCR: Iterative Language-Based Image Editing via Self-Supervised Counterfactual Reasoning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[21] Tsu-Jui Fu, Xin Eric Wang, Scott Grafton, Miguel Eckstein, and William Yang Wang. M" + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "L: Language-based Video Editing via Multi-Modal Multi-Level Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[22] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-Driven Artistic Style Transfer. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[23] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer. In European Conference on Computer Vision (ECCV), 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "text", + "content": "[24] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "text", + "content": "[25] Agrim Gupta, Stephen Tian, Yunzhi Zhang, Jiajun Wu, Roberto Martin-Martín, and Li Fei-Fei. MaskViT: Masked Visual Pre-Training for Video Prediction. In arXiv:2206.11894, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[26] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine This! Scripts to Compositions to Videos. In European Conference on Computer Vision (ECCV), 2018. 2, 5, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 715 + ], + "type": "text", + "content": "[27] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10689" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[28] Zekun Hao, Xun Huang, and Serge Belongie. Controllable Video Generation with Sparse Trajectories. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "text", + "content": "[29] Thomas Hayes, Songyang Zhang, Xi Yin, Guan Pang, Sasha Sheng, Harry Yang, Songwei Ge, Qiyuan Hu, and Devi Parikh. MUGEN: A Playground for Video-Audio-Text Multimodal Understanding and GENeration. In European Conference on Computer Vision (ECCV), 2022. 2, 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[30] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked Autoencoders Are Scalable Vision Learners. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[31] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In International Conference on Computer Vision (ICCV), 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 288, + 304 + ], + "type": "text", + "content": "[32] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. In arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 337 + ], + "type": "text", + "content": "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[34] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models. In arXiv:2204.03458, 2022. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers. In arXiv:2205.15868, 2022. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[36] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Learning to Decompose and Disentangle Representations for Video Prediction. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 459, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 492 + ], + "type": "text", + "content": "[37] Qiyang Hu, Adrian Waelchli, Tiziano Portenier, Matthias Zwicker, and Paolo Favaro. Video Synthesis from a Single Image and Motion Stroke. In arXiv:1812.01874, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[38] Yaosi Hu, Chong Luo, and Zhenzhong Chen. Make It Move: Controllable Image-to-Video Generation with Text Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "type": "text", + "content": "[39] Tobias Hoppe, Arash Mehrjou, Stefan Bauer, Didrik Nielsen, and Andrea Dittadi. Diffusion Models for Video Prediction and Infilling. In arXiv:2206.07696, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 614 + ], + "type": "text", + "content": "[40] Nal Kalchbrenner, Aaron van den Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video Pixel Networks. In International Conference on Machine Learning (ICML), 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 647 + ], + "type": "text", + "content": "[41] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[42] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference for Learning Representations (ICLR), 2015. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[43] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video Generation From Text. In AAAI Conference on Artificial Intelligence (AAAI), 2018. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "text", + "content": "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision (ICCV), 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[45] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "type": "text", + "content": "[46] Tanya Marwah, Gaurav Mittal, and Vineeth N. Balasubramanian. Attentive Semantic Video Generation using Captions. In International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "type": "text", + "content": "[47] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep Multi-Scale Video Prediction Beyond Mean Square Error. In International Conference on Learning Representations (ICLR), 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 319 + ], + "type": "text", + "content": "[48] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference for Learning Representations (ICLR), 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 320, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 342 + ], + "type": "text", + "content": "[49] Mehdi Mirza and Simon Osindero. Conditional Generative Adversarial Nets. In arXiv:1411.1784, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 398 + ], + "type": "text", + "content": "[50] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 546, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 464 + ], + "type": "text", + "content": "[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In International Conference on Machine Learning (ICML), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 467, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 546, + 510 + ], + "type": "text", + "content": "[52] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To Create What You Tell: Generating Videos from Captions. In International Conference on Multimedia (MM), 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 512, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 546, + 567 + ], + "type": "text", + "content": "[53] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in PyTorch. In International Conference on Learning Representations Workshop (ICLRW), 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 546, + 611 + ], + "type": "text", + "content": "[54] Björn Ommer Patrick Esser, Robin Rombach. Taming Transformers for High-Resolution Image Synthesis. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 657 + ], + "type": "text", + "content": "[55] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. MirrorGAN: Learning Text-to-image Generation by Redescription. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Interna-" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10690" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 93 + ], + "type": "text", + "content": "tional Conference on Machine Learning (ICML), 2021. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 286, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 286, + 128 + ], + "type": "text", + "content": "[57] Ruslan Rakhimov, Denis Volkhonskiy, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. Latent Video Transformer. In arXiv:2006.10704, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 286, + 161 + ], + "type": "text", + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. In arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 286, + 205 + ], + "type": "text", + "content": "[59] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 286, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 286, + 250 + ], + "type": "text", + "content": "[60] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. FILM: Frame Interpolation for Large Motion. In European Conference on Computer Vision (ECCV), 2022. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 251, + 286, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 286, + 294 + ], + "type": "text", + "content": "[61] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative Adversarial Text to Image Synthesis. In International Conference on Machine Learning (ICML), 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 286, + 338 + ], + "type": "text", + "content": "[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 339, + 286, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 286, + 415 + ], + "type": "text", + "content": "[63] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 286, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 449 + ], + "type": "text", + "content": "[64] Masaki Saito and Shunta Saito. TGANv2: Efficient Training of Large Models for Video Generation with Multiple Subsampling Layers. In arXiv:1811.09245, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 286, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 286, + 503 + ], + "type": "text", + "content": "[65] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. In arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 286, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 286, + 558 + ], + "type": "text", + "content": "[66] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In Annual Meetings of the Association for Computational Linguistics (ACL), 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 286, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 602 + ], + "type": "text", + "content": "[67] Karen Simonyan and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. In International Conference on Learning Representations (ICLR), 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 657 + ], + "type": "text", + "content": "[68] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-Video Generation without Text-Video Data. In arXiv:2209.14792, 2022. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 691 + ], + "type": "text", + "content": "[69] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild. In arXiv:1212.0402, 2012. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 286, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 715 + ], + "type": "text", + "content": "[70] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhutdinov. Unsupervised Learning of Video Representations using" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "LSTMs. In International Conference on Machine Learning (ICML), 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[71] Fuwen Tan, Song Feng, and Vicente Ordonez. Text2Scene: Generating Compositional Scenes from Textual Descriptions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[72] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N. Metaxas, and Sergey Tulyakov. A Good Image Generator Is What You Need for High-Resolution Video Synthesis. In International Conference for Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[73] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "text", + "content": "[74] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing Motion and Content for Video Generation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 354 + ], + "type": "text", + "content": "[75] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, and Sylvain Gelly Marcin Michalski. Towards Accurate Generative Models of Video: A New Metric & Challenges. In International Conference on Learning Representations Workshop (ICLRW), 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[76] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural Discrete Representation Learning. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 453 + ], + "type": "text", + "content": "[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention Is All You Need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "text", + "content": "[78] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable Length Video Generation From Open Domain Textual Description. In arXiv:2210.02399, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 510, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 566 + ], + "type": "text", + "content": "[79] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 601 + ], + "type": "text", + "content": "[80] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating Videos with Scene Dynamics. In Conference on Neural Information Processing Systems (NeurIPS), 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[81] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 646, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 690 + ], + "type": "text", + "content": "[82] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-Video Synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[83] Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. VATEX: A Large-Scale," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10691" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "High-Quality Multilingual Dataset for Video-and-Language Research. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[84] Dirk Weissenborn, Oscar Tackstrom, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference on Learning Representations (ICLR), 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[85] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling Autoregressive Video Models. In International Conference for Learning Representations (ICLR), 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[86] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. GODIVA: Generating Open-Domain Videos from nAtural Descriptions. In arXiv:2104.14806, 2021. 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "text", + "content": "[87] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. NUWA: Visual Synthesis Pretraining for Neural visUal World creAtion. In European Conference on Computer Vision (ECCV), 2022. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[88] Xide Xia, Tianfan Xue, Wei-Sheng Lai, Zheng Sun, Abby Chang, Brian Kulis, and Jiawen Chen. Real-Time Localized Photorealistic Video Style Transfer. In Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "text", + "content": "[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A Large Video Description Dataset for Bridging Video and Language. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[90] Qiangeng Xu, Hanwang Zhang, Weiyue Wang, Peter N. Belhumeur, and Ulrich Neumann. Stochastic Dynamics for Video Infilling. In Winter Conference on Applications of Computer Vision (WACV), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "type": "text", + "content": "[91] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep Flow-Guided Video Inpainting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "text", + "content": "[92] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks. In Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 522 + ], + "type": "text", + "content": "[93] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally Consistent Semantic Video Editing. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 556 + ], + "type": "text", + "content": "[94] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers. In arXiv:2104.10157, 2021. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 633 + ], + "type": "text", + "content": "[95] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. In arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[96] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks. In International Conference for Learning Representations (ICLR), 2022. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[97] Jiangning Zhang, Chao Xu, Liang Liu, Mengmeng Wang, Xia Wu, Yong liu, and Yunliang Jiang. DTVNet: Dynamic" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "Time-lapse Video Generation via Single Still Image. In European Conference on Computer Vision (ECCV), 2020. 1" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10692" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_content_list.json b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f27320239d585971b681bdf83a051f5766043357 --- /dev/null +++ b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_content_list.json @@ -0,0 +1,1909 @@ +[ + { + "type": "text", + "text": "TempSAL - Uncovering Temporal Information for Deep Saliency Prediction", + "text_level": 1, + "bbox": [ + 102, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bahar Aydemir, Ludo Hoffstetter, Tong Zhang, Mathieu Salzmann, Sabine Süssstrunk \nSchool of Computer and Communication Sciences, EPFL, Switzerland", + "bbox": [ + 148, + 180, + 818, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bahir/aydemir, tong.zhang, mathieu.salzmann, sabine.susstrunk}@epfl.ch", + "bbox": [ + 163, + 219, + 790, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/307199b203a9151932e8b29750a4934c05f1f6976b3e556d1ac0d7db25e64408.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 81, + 270, + 274, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e15e40f345eb4b7fad635135f663abdfdbc88a11d76ea3ed497b8de67ec5b896.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 268, + 374, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/760dc8cce4cc441bcc3251a3efe1f44fa0df592a9af200b8c3ff8d8a505de97c.jpg", + "image_caption": [ + "$\\mathcal{T}_1 = 0 - 1s$" + ], + "image_footnote": [], + "bbox": [ + 284, + 325, + 374, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/322905c8f2176207b7b1e982390b7749d7d3c7950cd96ce07140a8975b26d2f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 268, + 468, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a574130edd1b16f31495a2035538b3ecbe3b4119b7f0c9ad428335419109cea9.jpg", + "image_caption": [ + "$\\mathcal{T}_2 = 1 - 2s$", + "Figure 1. An example of how human attention evolves over time when observing a single image. Top row: Temporal and image saliency ground truth from the SALICON dataset [20]. Bottom row: Our temporal and image saliency predictions. Each temporal saliency map $\\mathcal{T}_i$ , $i \\in \\{1, \\dots, 5\\}$ represents one second of observation time. Note that in $\\mathcal{T}_1$ , the chef is salient, while in $\\mathcal{T}_2$ and $\\mathcal{T}_3$ , the food on the barbecue becomes the most salient region in this scene. We can predict the temporal saliency maps for each interval separately, or combine them to create a single, refined image saliency map for the entire observation period." + ], + "image_footnote": [], + "bbox": [ + 375, + 325, + 468, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2c271bcedad4b8e69371aaaac4523ff87dcf3d9b534b545c989479dfe12f5d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 268, + 566, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f77badc9d6fc58a890254a7584666f696dd04e65abce8b360651d5f929dfb65.jpg", + "image_caption": [ + "$\\mathcal{T}_3 = 2 - 3s$" + ], + "image_footnote": [], + "bbox": [ + 470, + 325, + 566, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55ab983def07961ecc49365fe9b8a5a98d1054c71fcc284cdfeca0438010f8f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 268, + 665, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d05853d6cade2062efa0a306798bc71c23afbc0eee60418eaf2dedd517b692f6.jpg", + "image_caption": [ + "$\\mathcal{T}_4 = 3 - 4s$" + ], + "image_footnote": [], + "bbox": [ + 568, + 325, + 665, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19db42a72cd02b4933528fca4d9cb92cedee2fee606ad0ceeac9de5851ff626f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 268, + 764, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e868de0d03bd3fa78ce3014e1244435e5fb70ded7272c6835c8a66c6dde1205f.jpg", + "image_caption": [ + "$\\mathcal{T}_{5} = 4 - 5s$" + ], + "image_footnote": [], + "bbox": [ + 668, + 325, + 764, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26934dac085f6443b31ee8d659c89b76270b622bf4ac1875d24c0aeb8276bcce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 268, + 880, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4eff5060045e80a783b3c018b1891ef6b211dea6e8315fb2ee105f12e4060b2b.jpg", + "image_caption": [ + "0-5s" + ], + "image_footnote": [], + "bbox": [ + 785, + 325, + 880, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 500, + 313, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep saliency prediction algorithms complement the object recognition features, they typically rely on additional information such as scene context, semantic relationships, gaze direction, and object dissimilarity. However, none of these models consider the temporal nature of gaze shifts during image observation. We introduce a novel saliency prediction model that learns to output saliency maps in sequential time intervals by exploiting human temporal attention patterns. Our approach locally modulates the saliency predictions by combining the learned temporal maps. Our experiments show that our method outperforms the state-of-the-art models, including a multi-duration saliency model, on the SALICON benchmark and CodeCharts1k dataset. Our code is publicly available on GitHub1.", + "bbox": [ + 73, + 532, + 472, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 773, + 207, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans have developed attention mechanisms that allow them to selectively focus on the important parts of a scene. Saliency prediction algorithms aim to computationally detect these regions that stand out relative to their surroundings. These predictions have numerous applications", + "bbox": [ + 75, + 800, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in image compression [37], image enhancement [51], image retargeting [1], rendering [43], and segmentation [28].", + "bbox": [ + 500, + 501, + 890, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the seminal work of Itti et al. [18], many have developed solutions using both handcrafted features [7] and deep ones [9, 17, 26, 34, 46, 48]. Nowadays, employing deep neural networks is preferred in saliency prediction as they outperform bottom-up models. These methods typically depend on pre-trained object recognition networks to extract features from the input image [31]. In addition to these features, scene context [47], object co-occurrence [50], and dissimilarity [2] have been exploited to improve the saliency prediction. However, while these approaches model the scene context and objects, they fail to consider that humans dynamically observe scenes [49]. In neuroscience, the inhibition of return paradigm states that a suppression mechanism reduces visual attention towards recently attended objects [39] and encourages selective attention to novel regions. Motivated by this principle, we develop a saliency prediction model that incorporates temporal information.", + "bbox": [ + 496, + 534, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fosco et al. [14] also exploit temporal information in saliency prediction, but they consider snapshots containing observations up to 0.5, 3, and 5 seconds, thus not leveraging saliency trajectory but rather saliency accumulation. By contrast, here, we model consecutive time slices, connecting our approach more directly with the human gaze and thus opening the door to automated visual appeal assess-", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 1, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://ivrl.github.io/Tempsal/", + "bbox": [ + 94, + 887, + 346, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6461", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment in applications such as website design [32], advertisement [36] and infographics [13].", + "bbox": [ + 76, + 90, + 467, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we show that when viewing images, human attention yields temporally evolving patterns, and we introduce a network capable of exploiting this temporal information for saliency prediction. Specifically, our model learns time-specific predictions and is able to combine them with a conventional image saliency map to obtain a temporally modulated image saliency prediction.", + "bbox": [ + 75, + 121, + 467, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on the SALICON [20] and CodeCharts1k [14] datasets which contain temporal information, unlike the other popular saliency datasets such as CAT2000 [3] and MIT1003 [23]. By showing the benefits of estimating temporal saliency we hope to encourage the community to publish the temporal information of their data along with the final saliency maps. Note that existing works typically collect saliency data by conducting psychophysical experiments [3, 14, 20, 23], and the attention data recorded during these experiments already includes temporal information. Therefore, no further experiments are required.", + "bbox": [ + 75, + 227, + 467, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As evidenced by our experiments, using temporal information boosts the accuracy of the baseline network, enabling us to consistently outperform the state-of-the-art models on the SALICON saliency benchmark. Moreover, in the CodeCharts1k dataset we outperform the multiduration model [14] in two out of three metrics.", + "bbox": [ + 76, + 407, + 467, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 96, + 500, + 387, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We evidence the presence of temporally evolving patterns in human attention.", + "- We show that temporal information in the form of a saliency trajectory improves saliency prediction in natural images, providing an investigation of the SALICON dataset for temporal attention shifts.", + "- We introduce a novel, saliency prediction model, called TempSAL, capable of simultaneously predicting conventional image saliency and temporal saliency trajectories.", + "- We propose a spatiotemporal mixing module that learns time-dependent patterns from temporal saliency maps. Our approach outperforms the state-of-the-art image saliency models that either do not consider temporal information or encode it in a cumulative manner." + ], + "bbox": [ + 96, + 523, + 467, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 792, + 212, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Saliency prediction for natural images", + "text_level": 1, + "bbox": [ + 76, + 816, + 405, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early saliency prediction methods were biologically inspired and bottom-up. In particular, Itti et al. used color, intensity, and orientation contrast [18]. Goferman employed global and local contrast as contextual cues [15]. Judd et", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "al. [23] further incorporated mid-level and high-level semantic features, using horizon, face, person, and car detectors. Later, Vig et al. [46] showed that deep neural networks can be applied to saliency prediction. Yet, saliency prediction lacks the large-scale annotated datasets that are available for image classification tasks [11], which prevents training robust models. To overcome this, Kummerer et al. [25] showed that using pre-trained object recognition networks significantly improves saliency predictions. Subsequent state-of-the-art models such as EML-Net [19], DeepGaze2 [26], and SALICON [17] similarly use pretrained convolutional neural network (VGG [42]) encoders. Recent works utilize additional sources of information such as scene context [24], external knowledge [50] and object dissimilarity [2] to improve saliency prediction. Yet, none of these methods take into account the temporal evolution of the human gaze, which occurs even when the image stimuli are static. In our work, we make use of these temporal patterns as an additional source of information to boost conventional image saliency prediction.", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-duration saliency", + "text_level": 1, + "bbox": [ + 500, + 404, + 718, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing image saliency ground truth maps include all fixations made throughout the observation period. Aggregating all these fixations that have different timestamps into a single ground truth map results in the loss of temporal information. Representing the fixations as scanpaths retains the temporal clues by encoding the change of gaze of an individual over time. However, merging numerous scanpaths is challenging [16]. Fosco et al. [14] proposed multiduration saliency to characterize the attention of a group of individuals while taking into account time-dependent attention shifts. The temporal maps they rely on, however, encode the attention distribution of many observers across overlapping time periods of increasing durations. While this is a convenient way of capturing a population's attention patterns, it does not reflect the saliency trajectory over time. Similarly, [35] use order of fixations as sequential metadata for deep supervision but they do not model the evolution of attention through time. In our work, we model multiduration saliency to analyze underlying attention patterns by using mutually exclusive time slices. We provide this temporal information to our spatiotemporal mixing module to refine the initial image saliency prediction with temporal information. Moreover, this lets us predict temporal saliency maps for each second of attention.", + "bbox": [ + 496, + 428, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Temporal saliency data analysis", + "text_level": 1, + "bbox": [ + 500, + 806, + 790, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 830, + 601, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SALICON [20] is the largest human attention dataset on natural images. It was created via a crowdsourced mouse tracking experiment, which was shown to be similar to eye", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tracking [20] and widely used in the saliency prediction literature. SALICON consists of 10000 training, 5000 validation and 5000 test images from the MS-COCO dataset [30]. The SALICON dataset provides saliency maps, fixations, and gaze points for each image and observer. A gaze point is a raw data point recorded by a tracking device. It describes the spatial coordinates of the eye/mouse on the associated stimuli at a given timestamp. Conversely, fixations describe the coordinates of the long pause when the eyes are fixated on an image detail. Following common practice in eye tracking experiments, Jiang et al. [20] grouped spatially and temporally close gaze points to create fixations. Since the fixations were created by grouping multiple gaze points, they do not have an associated timestamp. To address this, SALICON-MD [14] assumes that the fixations are uniformly distributed across the total viewing time. We provide a finer approximation for recovering the fixations' timestamps, by minimizing the spatial and temporal distance between a fixation and the nearest gaze point. We refer the reader to the supplementary material for the details of this approximation process.", + "bbox": [ + 76, + 90, + 472, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also assess the performance and generalization ability of our method using CodeCharts1k [14], which is the first saliency dataset to report multiple viewing durations. The dataset consists of 1000 images and saliency maps collected via crowdsourcing which correspond to 0.5, 3, and 5 seconds of viewing. We present the results in Section 5.5.", + "bbox": [ + 76, + 407, + 470, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Temporal patterns in the dataset", + "text_level": 1, + "bbox": [ + 76, + 508, + 364, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we examine how temporality evolves during human visual attention. To observe the evolution of attention over time, we slice the data into five slices, one for each second of observation. In particular, we inspect the dissimilarity between slices, the agreement with the average, and the distribution of fixations in the time-saliency space.", + "bbox": [ + 76, + 532, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Average maps. Viewing patterns in image saliency experiments tend to show a concentration towards the image center [44], which is known as the photographer's bias or center bias. We observe a similar spatial bias for each temporal slice, shown in Figure 2, where we plot the average heat maps for each temporal slice. Note that the gaze tends to converge to the center of the image as time passes. This means that the observers revisit the previously seen important center regions [49].", + "bbox": [ + 76, + 650, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ffe0ee55273b1b5895d1b17fe04b06004642875b8d8a7bc7f10cc4297d21eb86.jpg", + "image_caption": [ + "$\\mathcal{T}_1 = 0 - 1s$", + "Figure 2. Average heat maps for each one-second interval. Note that a center bias occurs, similar to image saliency prediction's average ground truth maps." + ], + "image_footnote": [], + "bbox": [ + 81, + 792, + 158, + 839 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cdbc9e400162f880d3485a1fd4bd135e827c763a7689b971c09434206e2b0767.jpg", + "image_caption": [ + "$\\mathcal{T}_2 = 1 - 2s$" + ], + "image_footnote": [], + "bbox": [ + 158, + 792, + 233, + 838 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/762975aa094de7f177e9b1017bc5539b11eebe7adb60d6b9ec970289e87a2221.jpg", + "image_caption": [ + "$\\mathcal{T}_3 = 2 - 3s$" + ], + "image_footnote": [], + "bbox": [ + 235, + 792, + 310, + 838 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4a217d9d80a4e7dcbd0a9fe997d92e831bf068e747e38f212c188ea185301e7a.jpg", + "image_caption": [ + "$\\mathcal{T}_4 = 3 - 4s$" + ], + "image_footnote": [], + "bbox": [ + 312, + 792, + 387, + 838 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5db29a0a754eeedf751a251b47e9bc98f09ff3f614646f3e11f789fec6159343.jpg", + "image_caption": [ + "$\\mathcal{T}_{5} = 4 - 5s$" + ], + "image_footnote": [], + "bbox": [ + 388, + 792, + 464, + 838 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/145916055125368e337b23624ad2948a8c83f85b16b77eca11585c5f66b8b4d1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
T1T2T3T4T5
T11.000.700.540.500.54
T20.701.000.730.660.65
T30.540.731.000.750.70
T40.500.660.751.000.73
T50.540.650.700.731.00
Average0.660.750.740.730.72
", + "bbox": [ + 501, + 87, + 874, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Correlation scores (CC) of the temporal slices with each other in a single image, averaged over all images. All slices show more similarity to their direct temporal neighbors. The last row shows the average similarity of a slice with the other slices, $\\mathcal{T}_1$ being the most dissimilar one.", + "bbox": [ + 496, + 213, + 893, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We plot the differences of the consecutive average temporal slices in Figure 3 to illustrate attention shifts. Light blue indicates the regions with reduced attention, whereas (light) red indicates increased attention. We observe that attention shifts from left to right, with a subsequent dispersion from the center towards the corners. Then, attention increases at the center of the image, slightly skewed to the left. Interestingly, the trend (especially in $\\mathcal{T}_2 - \\mathcal{T}_1$ ) coincides with the western left-to-right reading direction [8].", + "bbox": [ + 496, + 297, + 890, + 434 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b8fd29b00402c164408e1dba5f14ef592c81129913940cb05b6a5f0eb43daa82.jpg", + "image_caption": [ + "$\\mathcal{T}_2 - \\mathcal{T}_1$" + ], + "image_footnote": [], + "bbox": [ + 503, + 449, + 596, + 503 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/24396633b3c0ea0ddd207908eb52b265e40709d2293359c1a7128cd687539492.jpg", + "image_caption": [ + "$\\mathcal{T}_3 - \\mathcal{T}_2$", + "Figure 3. Differences of the consecutive average temporal slices shown in Fig. 2. Red indicates regions of increased attention whereas blue indicates decreased attention." + ], + "image_footnote": [], + "bbox": [ + 601, + 450, + 694, + 503 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/143c619e6ee97624f626f8c2ff6cbbebd8dbb400af7b5cd29361428c02205bfe.jpg", + "image_caption": [ + "$\\mathcal{T}_4 - \\mathcal{T}_3$" + ], + "image_footnote": [], + "bbox": [ + 697, + 450, + 792, + 503 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8c35bff611d8b8e0d6037d7d107a197151daecc13b054a5897428ac6b90f9d6.jpg", + "image_caption": [ + "$\\mathcal{T}_{5} - \\mathcal{T}_{4}$" + ], + "image_footnote": [], + "bbox": [ + 797, + 450, + 888, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inter-slice similarity across time. We expect temporal saliency slices to be more similar to their closer-in-time slices than to the ones further away since human attention is continuous over time. Table 1 contains the correlation coefficients between each pair of saliency slices in a single image, averaged over all images. We calculate the correlation coefficient between slices $\\mathcal{T}_j$ and $\\mathcal{T}_k$ as", + "bbox": [ + 496, + 590, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C C} \\left(\\mathcal {T} _ {j}, \\mathcal {T} _ {k}\\right) = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\operatorname {C C} \\left(\\mathcal {T} _ {i j}, \\mathcal {T} _ {i k}\\right), \\quad j, k \\in \\{1, \\dots , 5 \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 710, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the total number of images, and $\\mathcal{T}_{ij}$ and $\\mathcal{T}_{ik}$ denote the $j^{th}$ and $k^{th}$ slice of the $i^{th}$ image.", + "bbox": [ + 496, + 763, + 888, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By calculating t-test scores on the pairwise comparisons, we observe that all of the pairwise differences except $\\mathcal{T}_1$ , $\\mathcal{T}_3$ and $\\mathcal{T}_1$ , $\\mathcal{T}_5$ are statistically significant ( $p < 0.01$ ). Thus, the attention residuals between different time intervals in one image are significantly different. We provide more details in the supplementary material.", + "bbox": [ + 496, + 795, + 890, + 886 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6463", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c c c c c c} & \\mathcal {T} _ {1} & \\mathcal {T} _ {2} & \\mathcal {T} _ {3} & \\mathcal {T} _ {4} & \\mathcal {T} _ {5} \\\\ \\text {C C} & \\hline 0. 5 7 4 & \\hline 0. 4 3 3 & \\hline 0. 4 3 1 & \\hline 0. 4 2 6 & \\hline 0. 4 4 7 \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 87, + 439, + 125 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Correlation scores (CC) of each time slice in a single image with the average maps presented in Figure 2, averaged over all images. The similarity between slices across images decreases with time, with the exception of the last slice.", + "bbox": [ + 75, + 137, + 468, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intra-slice similarity across images. We also investigate the deviation of each slice from its respective average slices. The average slices are depicted in Figure 2. Table 2 shows the deviation of a slice from the average time slices per image. We compute CC scores between a single slice and the corresponding average slice as", + "bbox": [ + 75, + 210, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C C} \\left(\\mathcal {T} _ {j}, A _ {j}\\right) = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\operatorname {C C} \\left(\\mathcal {T} _ {i j}, A _ {j}\\right), \\quad j \\in \\{1, \\dots , 5 \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 313, + 468, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{j}$ denotes the $j^{th}$ average slice.", + "bbox": [ + 76, + 367, + 339, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We average the scores across all images. Higher values of CC indicate more agreement with the average whereas lower values of CC indicate more deviation from the average. Note that the similarity with the average across images decreases with time, except for the last slice. This can be explained by the more prominent center bias in $\\mathcal{T}_5$ as seen in Figure 2. $\\mathcal{T}_1$ has the least deviation from the average by a significant margin ( $p \\ll 0.01$ ). This shows that humans tend to look at similar places at first, and then their attention scatters around to less important regions.", + "bbox": [ + 75, + 383, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Early and late fixations versus saliency. Lastly, we investigate the relationship between the fixation timestamps and their respective saliency values. We assign a saliency value to each fixation as the normalized pixel value in the corresponding saliency map. We plot the histogram of the number of fixations with their saliency and timestamp values in Figure 4. The fixation time stamps range from 0 to $5000\\mathrm{ms}$ and the saliency values range from 0 to 1. Late fixations tend to have lower saliency values than earlier fixations, as indicated by the darker color towards the bottom right corner. That is, the first region we glance at in an image is more important (salient) than the following regions [7, 18].", + "bbox": [ + 75, + 550, + 468, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 761, + 210, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Temporal slices", + "text_level": 1, + "bbox": [ + 76, + 785, + 232, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to recover fixation timestamps to train models with this temporal information. We extract temporal slices by grouping the fixations in several time intervals and, following common practice [5], blurring with a Gaussian kernel. We break down the fixations into time slices with two time-slicing (grouping) alternatives, namely equal duration", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/38c5d35b9376831b01f7c2dbd658ed646856b95d9e2fb3b228b4a79a0ab66c94.jpg", + "image_caption": [ + "Figure 4. Number of fixations with their respective saliency values and timestamps. Lighter colors indicate a higher number of occurrences while darker areas denote fewer occurrences. We see that late fixations tend to be less salient, which can be seen as the decrease in the number of salient fixations along the arrow. The most salient fixations appear at approximately 1s." + ], + "image_footnote": [], + "bbox": [ + 509, + 95, + 895, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and equal distribution. The equal duration model outperforms the equal distribution one and is easier to interpret; we present a comparison of these two alternatives as an ablation study in the supplementary material.", + "bbox": [ + 498, + 402, + 892, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Temporal saliency model", + "text_level": 1, + "bbox": [ + 500, + 470, + 725, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us now introduce our framework that exploits temporal human attention information. Our model is depicted in Figure 5. We extract image features using a pre-trained object recognition encoder [33]. Then, we decode these features by a temporal slice decoder to obtain one saliency map per time slice. These temporal saliency slices are useful in automated visual appeal assessment in applications such as website design [32], advertisement [36] and infographics [13]. In parallel, we decode the same image features into an initial image saliency prediction. Finally, we combine the temporal slices and the image saliency predictions in the spatiotemporal mixing module to produce a final image saliency map. We describe each component in detail in the following sections.", + "bbox": [ + 498, + 494, + 892, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Image encoder and saliency decoders", + "text_level": 1, + "bbox": [ + 500, + 724, + 810, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the previous saliency prediction architectures [26, 31, 40], we first encode the input image with a pretrained image classification network, in our case PNASNet-5 [33]. We extract encoded features at various levels for multi-level integration, similar to a U-Net structure [41]. Formally, we denote the image encoder as", + "bbox": [ + 498, + 748, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} (\\mathcal {I}) = \\left[ \\mathcal {E} _ {i} \\right], \\quad i \\in \\{1, \\dots , 5 \\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 845, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ is the input image, and $\\mathcal{E}_i$ the $i^{\\mathrm{th}}$ encoder block. The output of $\\mathcal{E}(\\cdot)$ therefore is a 5D vector. The early en", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6464", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/abb002f7c2ec13981456acd36eccae123401d7651d21feb97d8dbd2fc00ffcf7.jpg", + "image_caption": [ + "Figure 5. Overview of the proposed architecture. We encode image features into encoder blocks consisting of multi-level image features. We then pass these blocks to the temporal saliency decoder (shown in orange) to decode them into temporal saliency predictions, which are saliency maps in sequential time intervals. In parallel, the image saliency decoder (shown in green) decodes the encoder blocks into an image saliency prediction. We then combine the temporal saliency maps, the image saliency map, and the encoder blocks in the spatiotemporal mixing module (shown in pink). (Best viewed in color.)" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "coder blocks extract low-level features, such as edges, color, and contrast, while the later blocks encode high-level semantics. We pass these blocks to the temporal saliency decoder, the image saliency decoder, and the spatiotemporal mixing module.", + "bbox": [ + 75, + 391, + 468, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our temporal slice decoder, namely $\\mathcal{D}_T$ , processes the encoder blocks with four 3x3 convolution layers followed by ReLU functions, integrating one encoder block after each convolution. Later, two 3x3 convolution layers with a ReLU in-between and a sigmoid function at the end produce $n$ temporal saliency maps. Formally, we write the temporal saliency decoder as", + "bbox": [ + 75, + 467, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {T} \\big (\\mathcal {E} (\\mathcal {I}) \\big) = [ \\mathcal {T} _ {n} ] =: \\mathcal {T}, \\quad n \\in \\{1, \\dots , 5 \\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 585, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}_n$ denotes the $n^{\\mathrm{th}}$ temporal saliency slice. Through this branch of the network, a single image input produces $n$ temporal saliency slices. We use this component to provide temporal predictions to the spatiotemporal mixing module.", + "bbox": [ + 75, + 614, + 468, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our image saliency decoder, namely $\\mathcal{D}_S$ , has the same structure as $\\mathcal{D}_T$ , with the exception of the number of output channels. This component produces a single map $S_I$ , which corresponds to the conventional image saliency map of the input image. As such, we can write", + "bbox": [ + 75, + 676, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {I} = \\mathcal {D} _ {S} (\\mathcal {E} (\\mathcal {I})). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 763, + 468, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use this module to provide cumulative saliency information to the spatiotemporal mixing module.", + "bbox": [ + 76, + 794, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Spatiotemporal Mixing Module", + "text_level": 1, + "bbox": [ + 76, + 845, + 352, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To incorporate temporal information into the saliency prediction, we introduce a module that combines temporal and", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "spatial saliency maps. Our design is inspired by the feature pyramid networks [29], which have the benefit of integrating features from multiple levels of the encoder, thus capturing low-level cues (such as color contrast, brightness, and edges) and high-level ones (such as semantics and scene context). Such cues have been shown to be critical for saliency estimation [15, 24, 38]. This module takes temporal saliency predictions, the initial image saliency prediction, and the encoded image feature blocks as input. We write this as", + "bbox": [ + 496, + 391, + 893, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {R} = \\operatorname {S M M} \\left(\\mathcal {E} (\\mathcal {I}), \\mathcal {T}, \\mathcal {S} _ {I}\\right), n \\in \\{1, \\dots , 5 \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 549, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{R}$ denotes the final, refined image saliency map.", + "bbox": [ + 498, + 571, + 870, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The module architecture is shown in Figure 6. It takes the last two encoder blocks $[\\mathcal{E}_5, \\mathcal{E}_4]$ and passes them through a $3 \\times 3$ convolution. We then concatenate the other encoder blocks with the image saliency and temporal saliency maps passing through $3 \\times 3$ convolution, ReLU, and linear upsampling to keep the spatial dimensions consistent. We add the static and temporal saliency maps to the encoded image features at each block to prevent the information in these maps from vanishing. In the last step, we only add the saliency maps to output a final refined saliency map $S_R$ . The final design of the module (i.e., combining E4 and E5) is based on empirical performance. This module eliminates the need for optimizing a weight parameter between the spatial and temporal maps. It can also modulate the maps within the spatial range of convolutions, which allows the selection of different regions from different maps.", + "bbox": [ + 496, + 588, + 892, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3 Loss Functions", + "text_level": 1, + "bbox": [ + 500, + 845, + 660, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train our network, we use the Kullback-Leibler divergence (KL) [45] and the Correlation Coefficient (CC)", + "bbox": [ + 498, + 869, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6465", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f5ed3128e6c7678728e74b86e95a39992ba6aecaf587da806bdcfa0220bbf434.jpg", + "image_caption": [ + "Figure 6. The spatiotemporal mixing module combines temporal saliency predictions with the conventional image saliency prediction by multi-level image feature integration. $S_{I}$ denotes the predicted image saliency map, $T_{1,\\dots,n}$ the temporal saliency predictions for $n$ frames, and $\\mathcal{E}_{1,\\dots,5}$ the encoder blocks. This multi-level integration scheme provides information from earlier layers of the network to the next blocks in this module. $S_{R}$ denotes the temporally refined image saliency map output." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 486, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[21] between the predicted and ground truth saliency maps. First, we train the temporal branch using", + "bbox": [ + 76, + 369, + 467, + 401 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} (\\mathcal {I}) = \\lambda_ {1} * \\operatorname {C C} \\left(G T _ {n}, \\mathcal {T} _ {n}\\right) + \\beta_ {1} * \\operatorname {K L} \\left(G T _ {n}, \\mathcal {T} _ {n}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 410, + 468, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $GT_{n}$ denotes the temporal ground truth for the $n^{th}$ slice. We then freeze the weights in this component and train the spatiotemporal mixing module using", + "bbox": [ + 76, + 436, + 468, + 484 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} (\\mathcal {I}) = \\lambda_ {2} * \\operatorname {C C} (G T, \\mathcal {S} _ {R}) + \\beta_ {2} * \\operatorname {K L} (G T, \\mathcal {S} _ {R}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 493, + 468, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $GT$ is the image saliency ground truth for image $I$ .", + "bbox": [ + 76, + 520, + 460, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 547, + 312, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 573, + 267, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a batch size of 32 and an initial learning rate of 1e-4, reduced by a factor of ten every two epochs. We train the temporal branch first and then freeze the weights. We found that 10 epochs of training on SALICON was sufficient. For SALICON, we used the provided test, train, and validation splits.", + "bbox": [ + 76, + 595, + 470, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Metrics", + "text_level": 1, + "bbox": [ + 76, + 696, + 173, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the obtained saliency predictions according to the following standard metrics used by the community.", + "bbox": [ + 76, + 719, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Area Under the Curve (AUC) [5]: Saliency prediction can be interpreted as classifying fixation vs non-fixation points. The area under the ROC curve shows the trade-off between true positives (TP) and false positives (FP). A higher AUC score indicates fewer FPs. While AUC computes the TP and FP rates using all the ground truth fixation points, sAUC [4] samples FP points from ground truth fixations of other observers, compensating for center bias in natural images by taking for inter- and intra-observer variability into account. Normalized Scanpath Saliency (NSS) [38]: This metric", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compares the predicted saliency values at the ground truth fixation points to the average predicted saliency. An NSS score of one indicates that the predicted saliency values at the ground truth fixation points are one standard deviation above the average.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Kullback - Leibler Divergence (KL) [45]: The KL measures the cumulative distance between the predicted and the ground truth saliency maps. A KL score close to zero indicates a better approximation of the ground truth saliency map by the predicted one.", + "bbox": [ + 496, + 166, + 890, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pearson's correlation coefficient (CC) [21]: This metric measures the linear relationship between the predicted and ground truth saliency maps. It ranges from -1 to 1. A CC score close to one indicates a strong linear correlation between the two maps.", + "bbox": [ + 496, + 242, + 890, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarity (SIM) score [22]: The similarity score sums the minimum value between the predicted and the ground truth saliency maps over all pixels. A similarity score of 1 indicates a perfect prediction since both of the maps are probability distributions summing to 1.", + "bbox": [ + 496, + 316, + 890, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Information Gain (IG) score [27]: The information gain is an information-theoretic metric that measures the difference in average log-likelihood between the predicted saliency map and center-bias prior.", + "bbox": [ + 496, + 393, + 890, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 500, + 469, + 692, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the state-of-the-art models, namely SAM-Resnet [10], MSI-Net, GazeGAN, MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14], in image saliency prediction. Our model outperforms these methods in five out of seven metrics, showing the benefit of incorporating temporal information. Moreover, our model outperforms the only other multi-duration saliency model [14] in image saliency prediction by a significant margin. Furthermore, we compare our model with this multi-duration model and a multi-duration baseline trained on individual temporal slices. Our model improves the saliency prediction in two durations consisting of 0.5 and 3 seconds in two out of three metrics and in all three metrics in the five-second duration.", + "bbox": [ + 496, + 494, + 890, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 500, + 722, + 857, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate the performance of our model on image saliency prediction on the SALICON benchmark [20]. The ground truth of SALICON's test set is exclusively hosted on the CodaLab website2. Table 3 shows the comparison of standard evaluation metrics for different state-of-the-art saliency models alongside our model TempSAL. TempSAL outperforms all the baselines in almost all metrics. When it does not, it still yields competitive results.", + "bbox": [ + 496, + 748, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://competitions.codalab.org/competitions/17136", + "bbox": [ + 514, + 886, + 803, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6466", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/19ac404a43052f0b3a7a2fef7ab6ae360e8431ec88ada98a582ca5d7b4fc6eed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMDAUC ↑CC ↑KL ↓SAUC ↑IG ↑NSS ↑SIM ↑
SAM-Resnet [10]X0.8650.8990.6100.7410.5381.9900.793
MSI-Net [24]X0.8650.8990.3070.7360.7931.9310.784
GazeGAN [6]X0.8640.8790.3760.7360.7201.8990.773
SimpleNet [40]X0.8690.9070.2010.7430.8801.9600.793
MDNSal [40]X0.8650.8990.2210.7360.8631.9350.790
UNISAL [12]X0.8640.8790.3540.7390.7801.9520.775
DeepGaze IIE [31]X0.8690.8720.2850.7670.7661.9960.733
MD-SEM [14]✓0.8640.8680.5680.7460.6602.0580.774
TempSAL✓0.8690.9110.1950.7450.8961.9670.800
", + "bbox": [ + 125, + 85, + 848, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7494abdb5497e4291b13bd49af0bc91041b06d3cb78051630639ceb238765dca.jpg", + "table_caption": [ + "Table 3. Evaluation results on the SALICON (LSUN 2017) test benchmark. We compare our model with the state-of-the-art saliency prediction models, namely SAM-Resnet [10], MSI-Net [24], GazeGAN [6], MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14]. The results in bold show the best performance. Our method outperforms the state-of-the-art on conventional image saliency in five metrics. The MD column denotes the ability of the models to predict multi-duration saliency. Our model outperforms the only other multi-duration saliency model by a significant margin on five out of seven metrics." + ], + "table_footnote": [], + "table_body": "
ModelSlice 1 (0-500 ms)Slice 2 (0-3000 ms)Slice 3 (0-5000 ms)Average
CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑
SAM-MD [14]0.8050.3703.1810.7380.4692.5410.7150.5352.4950.7530.4582.739
MD-SEM [14]0.8160.3513.3740.7450.4522.6940.7340.4872.6770.7650.4302.915
TempSAL0.8190.4963.4220.7520.5122.7030.8220.4713.3370.7970.4933.154
", + "bbox": [ + 86, + 340, + 883, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Comparison with the multi-duration method", + "text_level": 1, + "bbox": [ + 75, + 529, + 455, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare our method with the only other multiduration model [14], we modify our network to output three temporal slices. We train our network on a three slice SALICON multi-duration dataset first and then fine-tune it on the CodeCharts1k dataset [14] using the given training and validation splits. We report the results of the comparison in Table 4.", + "bbox": [ + 75, + 551, + 470, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 665, + 261, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 7, we compare the temporal and image saliency maps obtained with our method with the ground truth from SALICON [20]. Our model learns time-specific predictions and is able to combine such predictions into a conventional image saliency map. We provide additional qualitative results in the supplementary material.", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.7. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 786, + 238, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the effect of different components in our model. We also provide a comparison with a multi-duration baseline model on the temporal SALICON dataset. Moreover, we provide an ablation study on the effect of the number of time slices and the two temporal slicing methods in the supplementary material.", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the SMM module: We evaluate the effect of the spatiotemporal mixing module (SMM) and the image saliency decoder in Table 5. The first model consists of the image encoder and temporal saliency decoder only. We take the average of the temporal slices to measure its performance by comparing it with the ground truth image saliency map. In the second row, we add the image saliency decoder to our model. Similarly, we take the average of the predicted maps $\\mathcal{T}_n$ and $S_I$ . Lastly, we add the spatiotemporal mixing module, which effectively modulates these predicted maps and combines them into a final image saliency map $S_R$ .", + "bbox": [ + 496, + 529, + 893, + 696 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21f80f5109139943c355c69bfe55cb6c6c7566fd5d3f8f733b8a177df09bd9ff.jpg", + "table_caption": [ + "Table 4. Results of our model, MD-SEM [14], and the baseline SAM-MD [14] across different durations on the CodeCharts1k dataset [14]. Our model improves the saliency prediction in the first two intervals, consisting of $500~\\mathrm{ms}$ and $3000~\\mathrm{ms}$ observations, in two out of three metrics. In this comparison, the time slices are cumulative, not mutually exclusive. Although our model benefits from non-overlapping temporal slices, it also performs well with cumulative time slices, particularly in the last slice corresponding to the image saliency with an observation duration of $5000~\\mathrm{ms}$ ." + ], + "table_footnote": [], + "table_body": "
ModelCC ↑KL ↓NSS ↑SIM ↑
DT(E(I))0.8520.2431.9730.754
+DS(E(I))0.8570.2521.9430.760
+SMM0.9060.1981.9300.798
", + "bbox": [ + 519, + 704, + 872, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results of ablation studies on the temporal SALICON validation dataset. The first row denotes the model with only the temporal saliency decoder. In the second row, the model has both the temporal and image saliency decoders. The last row denotes the performance with the spatiotemporal mixing module (SMM). As evidenced by the improved accuracy metrics, the SMM effectively modulates the spatial and temporal saliency maps to refine the initial image saliency prediction.", + "bbox": [ + 496, + 785, + 893, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6467", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28652b45f66112ae8ea863ee92661c48428b9089e513e89dc83ac92f351ad510.jpg", + "image_caption": [ + "Figure 7. The first row shows the input image with the ground truth saliency overlaid. The second row shows the ground truth temporal saliency and image saliency. We show our temporal saliency and image saliency predictions on the third row. Red-yellow maps are temporal saliency maps for one-second intervals. Black and white maps are image saliency maps for the whole observation duration. Our approach captures the attention shifts in sequential temporal maps. Moreover, our model is able to produce accurate image saliency predictions which are close to the ground truth maps." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 867, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with a temporal baseline model: The performance of our TempSAL model with five temporal slices is provided in Table 6. Note that each saliency slice contains five times fewer samples than the original image saliency map. Therefore, individual slices contain more variation compared to conventional accumulated maps. As a baseline to our model, we compute the performance of an architecture composed of five replicated SimpleNet models (5xSimpleNet) [40], each trained on one saliency slice. This baseline model uses an unshared encoder and decoder for each slice, while we share the decoder among slices. Therefore, we do not benefit from increased model capacity. We observe an accuracy decline in the baseline model, which confirms the increased discrepancy in the data.", + "bbox": [ + 75, + 575, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 800, + 194, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a saliency prediction method that can learn time-specific predictions and is also able to exploit temporal information to improve overall image saliency prediction. In particular, we show that the temporally evolving patterns in human attention play an important", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99a3ebc885f68051023d4200285511e902c8b4ed08b66790682efe014cf9b720.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TimeBaselineTempSAL
CC↑KL↓NSS↑SIM↑CC↑KL↓NSS↑SIM↑
T10.8980.2112.4360.7780.8990.2142.4530.782
T20.8700.2192.1590.7650.8770.2152.2110.776
T30.8400.2471.8400.7530.8430.2471.8780.758
T40.8200.2731.7290.7430.8250.2641.7400.749
T50.8110.2751.6460.7380.8130.2761.6540.743
Average0.8480.2451.9620.7560.8520.2431.9870.761
", + "bbox": [ + 504, + 561, + 887, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Results of the baseline model (left) and our TempSAL model (right) across different time slices. In 18 out of 20 comparisons, our model consistently outperforms the baseline. Note that both models perform best in the first slice, in which the intra-slice agreement is more prominent than in the other slices, as mentioned in Section 3.2.", + "bbox": [ + 498, + 676, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "role in saliency prediction in natural images. This is evidenced by our experiments that demonstrate our TempSAL method outperforming the state-of-the-art, including a multi-duration method exploiting cumulative temporal saliency maps.", + "bbox": [ + 496, + 773, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the Swiss National Science Foundation via the Sinergia grant CRSII5-180359.", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6468", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta and Sabine Süsstrunk. Saliency detection for content-aware image resizing. In IEEE International Conference on Image Processing (ICIP), pages 1005-1008, 2009. 1", + "[2] Bahar Aydemir, Deblina Bhattacharjee, Seungryong Kim, Tong Zhang, Mathieu Salzmann, and Sabine Süsstrunk. Modeling object dissimilarity for deep saliency prediction. Transactions on Machine Learning Research (TMLR), 2022. 1, 2", + "[3] Ali Borji and Laurent Itti. CAT2000: A large scale fixation dataset for boosting saliency research. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2015. 2", + "[4] Ali Borji, Dicky N. Sihite, and Laurent Itti. Quantitative analysis of human-model agreement in visual saliency modeling: A comparative study. IEEE Transactions on Image Processing (TIP), 22(1):55-69, 2013. 6", + "[5] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Fredo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41(3):740, 2019. 4, 6", + "[6] Zhaohui Che, Ali Borji, Guangtao Zhai, Xiongkuo Min, Guodong Guo, and Patrick Le Callet. Gazegan: A generative adversarial saliency model based on invariance analysis of human gaze during scene free viewing. ArXiv, abs/1905.06803, 2019. 7", + "[7] Ming-Ming Cheng, Niloy J. Mitra, Xiaolei Huang, Philip H. S. Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 37(3):569-582, 2015. 1, 4", + "[8] Hannah Faye Chua, Julie E. Boland, and Richard E. Nisbett. Cultural variation in eye movements during scene perception. Proceedings of the National Academy of Sciences, 102(35):12629-12633, 2005. 3", + "[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. A deep multi-level network for saliency prediction. In IEEE International Conference on Pattern Recognition (ICPR), pages 3488-3493, 2016. 1", + "[10] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an LSTM-based saliency attentive model. IEEE Transactions on Image Processing (TIP), 27(10):5142-5154, 2018. 6, 7", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2", + "[12] Richard Droste, Jianbo Jiao, and J. Alison Noble. Unified Image and Video Saliency Modeling. In European Conference on Computer Vision (ECCV), 2020. 7", + "[13] Camilo Fosco, Vincent Casser, Amish Kumar Bedi, Peter O'Donovan, Aaron Hertzmann, and Zoya Bylinskii. Predicting visual importance across graphic design types. In ACM Symposium on User Interface Software and Technology (UIST), page 249-260, 2020. 2, 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Camilo Fosco, Anelise Newman, Pat Sukhum, Yun Bin Zhang, Nanxuan Zhao, Aude Oliva, and Zoya Bylinskii. How much time do you have? modeling multi-duration saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4482, 2020. 1, 2, 3, 6, 7", + "[15] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 34(10):1915-1926, 2011. 2, 5", + "[16] Joseph H. Goldberg and Jonathan I. Helfman. Visual scanpath representation. In Symposium on Eye-Tracking Research & Applications (ETRA), pages 203–210, 2010. 2", + "[17] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. SALICON: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In IEEE International Conference on Computer Vision (ICCV), pages 262-270, 2015. 1, 2", + "[18] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 20(11):1254-1259, 1998. 1, 2, 4", + "[19] Sen Jia and Neil D. B. Bruce. EML-NET: An expandable Multi-Layer NETwork for saliency prediction. Image and Vision Computing, 95:103887, 2020. 2", + "[20] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. SALICON: Saliency in context. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 2, 3, 6, 7", + "[21] Timothee Jost, Nabil Ouerhani, Roman von Wartburg, René Muri, and Heinz Hügli. Assessing the contribution of color in visual attention. Computer Vision and Image Understanding, 100(1-2):107-123, 2005. 6", + "[22] Tilke Judd, Fredo Durand, and Antonio Torralba. A benchmark of computational models of saliency to predict human fixations. MIT Technical Report, 2012. 6", + "[23] Tilke Judd, Krista Ehinger, Fredo Durand, and Antonio Torralba. Learning to predict where humans look. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2009. 2", + "[24] Alexander Kroner, Mario Senden, Kurt Driessens, and Rainer Goebel. Contextual encoder-decoder network for visual saliency prediction. Neural Networks, 129:261 - 270, 2020. 2, 5, 7", + "[25] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep Gaze I: Boosting saliency prediction with feature maps trained on ImageNet. In International Conference on Learning Representations (ICLR) Workshops, 2015. 2", + "[26] Matthias Kümmerer, Tom Wallis, and Matthias Bethge. DeepGaze II: Predicting fixations from deep features over time and tasks. Journal of Vision (JOV), 17(10):1147, 2017. 1, 2, 4", + "[27] Matthias Kümmerer, Thomas S. A. Wallis, and Matthias Bethge. Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences (PNAS), 112(52):16054-16059, 2015. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6469", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Qingshan Li, Yue Zhou, and Jie Yang. Saliency based image segmentation. In IEEE International Conference on Multimedia Technology, pages 5068-5071, 2011. 1", + "[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 5", + "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 3", + "[31] Akis Linardos, Matthias Kümmerer, Ori Press, and Matthias Bethge. Deepgaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12919-12928, 2021. 1, 4, 6, 7", + "[32] Gitte Lindgaard, Gary Fernandes, Cathy Dudek, and J. Brown. Attention web designers: You have 50 milliseconds to make a good first impression! Behaviour & Information Technology, 25(2):115-126, 2006. 2, 4", + "[33] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In European Conference on Computer Vision (ECCV), pages 19-34, 2018. 4", + "[34] Nian Liu and Junwei Han. A deep spatial contextual long-term recurrent convolutional network for saliency detection. IEEE Transactions on Image Processing (TIP), 27(7):3264-3274, 2018. 1", + "[35] Sandeep Mishra and Oindrila Saha. RecSal : Deep recursive supervision for visual saliency prediction. In British Machine Vision Conference (BMVC), 2020. 2", + "[36] Johanna Palcu, Jennifer Sudkamp, and Arnd Florack. Judgments at gaze value: Gaze cuing in banner advertisements, its effect on attention allocation and product judgments. Frontiers in Psychology, 8, 2017. 2, 4", + "[37] Yash Patel, Srikar Appalaraju, and R. Manmatha. Saliency driven perceptual image compression. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 227-236, 2021. 1", + "[38] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397-2416, 2005. 5, 6", + "[39] Michael Posner, Robert Rafal, Lisa Choatec, and Jonathan Vaughan. Inhibition of return: Neural basis and function. Cognitive Neuropsychology, Vol. 2:211 - 228, 09 1985. 1", + "[40] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In International Conference on Intelligent Robots and Systems (IROS), 2020. 4, 6, 7, 8", + "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015. 4" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations (ICLR), 2015. 2", + "[43] Markus Steinberger, Bernhard Kainz, Stefan Hauswiesner, Rostislav Khlebnikov, Denis Kalkofen, and Dieter Schmalstieg. Ray prioritization using stylization and visual saliency. Computers & Graphics, 36(6):673-684, 2012. 1", + "[44] Po-He Tseng, Ran Carmi, Ian G. M. Cameron, Douglas P. Munoz, and Laurent Itti. Quantifying center bias of observers in free viewing of dynamic natural scenes. Journal of Vision (JOV), 9(7):4-4, 07 2009. 3", + "[45] Mathukumalli Vidyasagar. Kullback-leibler divergence rate between probability distributions on sets of different cardinalities. In IEEE Conference on Decision and Control (CDC), pages 948-953, 2010. 5, 6", + "[46] Eleonora Vig, Michael Dorr, and David Cox. Large-scale optimization of hierarchical features for saliency prediction in natural images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2798-2805, 2014. 1, 2", + "[47] W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, and R. Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(06):3239-3259, 2022. 1", + "[48] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. IEEE Transactions on Multimedia, 22(8):2163-2176, 2020. 1", + "[49] Alfred L. Yarbus. Eye Movements and Vision, volume 2. Plenum Press, 1967. 1, 3", + "[50] Yifeng Zhang, Ming Jiang, and Qi Zhao. Saliency prediction with external knowledge. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 484-493, 2021. 1, 2", + "[51] Jufeng Zhao, Yueting Chen, Huajun Feng, Zhihai Xu, and Qi Li. Fast image enhancement using multi-scale saliency extraction in infrared imagery. Optik, 125(15):4039-4042, 2014. 1" + ], + "bbox": [ + 501, + 92, + 890, + 641 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6470", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_model.json b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c710cea0c8035674423704e0527e3353b04f4d1a --- /dev/null +++ b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_model.json @@ -0,0 +1,2596 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "TempSAL - Uncovering Temporal Information for Deep Saliency Prediction" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.181, + 0.82, + 0.216 + ], + "angle": 0, + "content": "Bahar Aydemir, Ludo Hoffstetter, Tong Zhang, Mathieu Salzmann, Sabine Süssstrunk \nSchool of Computer and Communication Sciences, EPFL, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.22, + 0.792, + 0.234 + ], + "angle": 0, + "content": "{bahir/aydemir, tong.zhang, mathieu.salzmann, sabine.susstrunk}@epfl.ch" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.271, + 0.276, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.386, + 0.216, + 0.398 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.27, + 0.375, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.327, + 0.375, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.385, + 0.37, + 0.397 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_1 = 0 - 1s\\)" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.27, + 0.47, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.327, + 0.47, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.385, + 0.466, + 0.397 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_2 = 1 - 2s\\)" + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.27, + 0.568, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.327, + 0.568, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.385, + 0.563, + 0.397 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_3 = 2 - 3s\\)" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.27, + 0.666, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.327, + 0.666, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.385, + 0.66, + 0.397 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_4 = 3 - 4s\\)" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.27, + 0.766, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.327, + 0.766, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.385, + 0.757, + 0.397 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_{5} = 4 - 5s\\)" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.27, + 0.882, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.327, + 0.882, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.385, + 0.857, + 0.397 + ], + "angle": 0, + "content": "0-5s" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Figure 1. An example of how human attention evolves over time when observing a single image. Top row: Temporal and image saliency ground truth from the SALICON dataset [20]. Bottom row: Our temporal and image saliency predictions. Each temporal saliency map \\(\\mathcal{T}_i\\), \\(i \\in \\{1, \\dots, 5\\}\\) represents one second of observation time. Note that in \\(\\mathcal{T}_1\\), the chef is salient, while in \\(\\mathcal{T}_2\\) and \\(\\mathcal{T}_3\\), the food on the barbecue becomes the most salient region in this scene. We can predict the temporal saliency maps for each interval separately, or combine them to create a single, refined image saliency map for the entire observation period." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.5, + 0.314, + 0.516 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.533, + 0.473, + 0.745 + ], + "angle": 0, + "content": "Deep saliency prediction algorithms complement the object recognition features, they typically rely on additional information such as scene context, semantic relationships, gaze direction, and object dissimilarity. However, none of these models consider the temporal nature of gaze shifts during image observation. We introduce a novel saliency prediction model that learns to output saliency maps in sequential time intervals by exploiting human temporal attention patterns. Our approach locally modulates the saliency predictions by combining the learned temporal maps. Our experiments show that our method outperforms the state-of-the-art models, including a multi-duration saliency model, on the SALICON benchmark and CodeCharts1k dataset. Our code is publicly available on GitHub1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.775, + 0.208, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Humans have developed attention mechanisms that allow them to selectively focus on the important parts of a scene. Saliency prediction algorithms aim to computationally detect these regions that stand out relative to their surroundings. These predictions have numerous applications" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.502, + 0.891, + 0.533 + ], + "angle": 0, + "content": "in image compression [37], image enhancement [51], image retargeting [1], rendering [43], and segmentation [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Since the seminal work of Itti et al. [18], many have developed solutions using both handcrafted features [7] and deep ones [9, 17, 26, 34, 46, 48]. Nowadays, employing deep neural networks is preferred in saliency prediction as they outperform bottom-up models. These methods typically depend on pre-trained object recognition networks to extract features from the input image [31]. In addition to these features, scene context [47], object co-occurrence [50], and dissimilarity [2] have been exploited to improve the saliency prediction. However, while these approaches model the scene context and objects, they fail to consider that humans dynamically observe scenes [49]. In neuroscience, the inhibition of return paradigm states that a suppression mechanism reduces visual attention towards recently attended objects [39] and encourages selective attention to novel regions. Motivated by this principle, we develop a saliency prediction model that incorporates temporal information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Fosco et al. [14] also exploit temporal information in saliency prediction, but they consider snapshots containing observations up to 0.5, 3, and 5 seconds, thus not leveraging saliency trajectory but rather saliency accumulation. By contrast, here, we model consecutive time slices, connecting our approach more directly with the human gaze and thus opening the door to automated visual appeal assess-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.348, + 0.902 + ], + "angle": 0, + "content": "1https://ivrl.github.io/Tempsal/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6461" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "ment in applications such as website design [32], advertisement [36] and infographics [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.227 + ], + "angle": 0, + "content": "To achieve this, we show that when viewing images, human attention yields temporally evolving patterns, and we introduce a network capable of exploiting this temporal information for saliency prediction. Specifically, our model learns time-specific predictions and is able to combine them with a conventional image saliency map to obtain a temporally modulated image saliency prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.408 + ], + "angle": 0, + "content": "We evaluate our method on the SALICON [20] and CodeCharts1k [14] datasets which contain temporal information, unlike the other popular saliency datasets such as CAT2000 [3] and MIT1003 [23]. By showing the benefits of estimating temporal saliency we hope to encourage the community to publish the temporal information of their data along with the final saliency maps. Note that existing works typically collect saliency data by conducting psychophysical experiments [3, 14, 20, 23], and the attention data recorded during these experiments already includes temporal information. Therefore, no further experiments are required." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.468, + 0.498 + ], + "angle": 0, + "content": "As evidenced by our experiments, using temporal information boosts the accuracy of the baseline network, enabling us to consistently outperform the state-of-the-art models on the SALICON saliency benchmark. Moreover, in the CodeCharts1k dataset we outperform the multiduration model [14] in two out of three metrics." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.5, + 0.388, + 0.513 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.524, + 0.468, + 0.552 + ], + "angle": 0, + "content": "- We evidence the presence of temporally evolving patterns in human attention." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.564, + 0.468, + 0.623 + ], + "angle": 0, + "content": "- We show that temporal information in the form of a saliency trajectory improves saliency prediction in natural images, providing an investigation of the SALICON dataset for temporal attention shifts." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.634, + 0.468, + 0.693 + ], + "angle": 0, + "content": "- We introduce a novel, saliency prediction model, called TempSAL, capable of simultaneously predicting conventional image saliency and temporal saliency trajectories." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.704, + 0.468, + 0.779 + ], + "angle": 0, + "content": "- We propose a spatiotemporal mixing module that learns time-dependent patterns from temporal saliency maps. Our approach outperforms the state-of-the-art image saliency models that either do not consider temporal information or encode it in a cumulative manner." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.524, + 0.468, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.793, + 0.214, + 0.807 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.406, + 0.833 + ], + "angle": 0, + "content": "2.1. Saliency prediction for natural images" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Early saliency prediction methods were biologically inspired and bottom-up. In particular, Itti et al. used color, intensity, and orientation contrast [18]. Goferman employed global and local contrast as contextual cues [15]. Judd et" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": "al. [23] further incorporated mid-level and high-level semantic features, using horizon, face, person, and car detectors. Later, Vig et al. [46] showed that deep neural networks can be applied to saliency prediction. Yet, saliency prediction lacks the large-scale annotated datasets that are available for image classification tasks [11], which prevents training robust models. To overcome this, Kummerer et al. [25] showed that using pre-trained object recognition networks significantly improves saliency predictions. Subsequent state-of-the-art models such as EML-Net [19], DeepGaze2 [26], and SALICON [17] similarly use pretrained convolutional neural network (VGG [42]) encoders. Recent works utilize additional sources of information such as scene context [24], external knowledge [50] and object dissimilarity [2] to improve saliency prediction. Yet, none of these methods take into account the temporal evolution of the human gaze, which occurs even when the image stimuli are static. In our work, we make use of these temporal patterns as an additional source of information to boost conventional image saliency prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.719, + 0.421 + ], + "angle": 0, + "content": "2.2. Multi-duration saliency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Existing image saliency ground truth maps include all fixations made throughout the observation period. Aggregating all these fixations that have different timestamps into a single ground truth map results in the loss of temporal information. Representing the fixations as scanpaths retains the temporal clues by encoding the change of gaze of an individual over time. However, merging numerous scanpaths is challenging [16]. Fosco et al. [14] proposed multiduration saliency to characterize the attention of a group of individuals while taking into account time-dependent attention shifts. The temporal maps they rely on, however, encode the attention distribution of many observers across overlapping time periods of increasing durations. While this is a convenient way of capturing a population's attention patterns, it does not reflect the saliency trajectory over time. Similarly, [35] use order of fixations as sequential metadata for deep supervision but they do not model the evolution of attention through time. In our work, we model multiduration saliency to analyze underlying attention patterns by using mutually exclusive time slices. We provide this temporal information to our spatiotemporal mixing module to refine the initial image saliency prediction with temporal information. Moreover, this lets us predict temporal saliency maps for each second of attention." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.791, + 0.823 + ], + "angle": 0, + "content": "3. Temporal saliency data analysis" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.602, + 0.846 + ], + "angle": 0, + "content": "3.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SALICON [20] is the largest human attention dataset on natural images. It was created via a crowdsourced mouse tracking experiment, which was shown to be similar to eye" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6462" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.408 + ], + "angle": 0, + "content": "tracking [20] and widely used in the saliency prediction literature. SALICON consists of 10000 training, 5000 validation and 5000 test images from the MS-COCO dataset [30]. The SALICON dataset provides saliency maps, fixations, and gaze points for each image and observer. A gaze point is a raw data point recorded by a tracking device. It describes the spatial coordinates of the eye/mouse on the associated stimuli at a given timestamp. Conversely, fixations describe the coordinates of the long pause when the eyes are fixated on an image detail. Following common practice in eye tracking experiments, Jiang et al. [20] grouped spatially and temporally close gaze points to create fixations. Since the fixations were created by grouping multiple gaze points, they do not have an associated timestamp. To address this, SALICON-MD [14] assumes that the fixations are uniformly distributed across the total viewing time. We provide a finer approximation for recovering the fixations' timestamps, by minimizing the spatial and temporal distance between a fixation and the nearest gaze point. We refer the reader to the supplementary material for the details of this approximation process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.472, + 0.501 + ], + "angle": 0, + "content": "We also assess the performance and generalization ability of our method using CodeCharts1k [14], which is the first saliency dataset to report multiple viewing durations. The dataset consists of 1000 images and saliency maps collected via crowdsourcing which correspond to 0.5, 3, and 5 seconds of viewing. We present the results in Section 5.5." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.51, + 0.366, + 0.526 + ], + "angle": 0, + "content": "3.2. Temporal patterns in the dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.47, + 0.64 + ], + "angle": 0, + "content": "In this section, we examine how temporality evolves during human visual attention. To observe the evolution of attention over time, we slice the data into five slices, one for each second of observation. In particular, we inspect the dissimilarity between slices, the agreement with the average, and the distribution of fixations in the time-saliency space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.651, + 0.47, + 0.788 + ], + "angle": 0, + "content": "Average maps. Viewing patterns in image saliency experiments tend to show a concentration towards the image center [44], which is known as the photographer's bias or center bias. We observe a similar spatial bias for each temporal slice, shown in Figure 2, where we plot the average heat maps for each temporal slice. Note that the gaze tends to converge to the center of the image as time passes. This means that the observers revisit the previously seen important center regions [49]." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.793, + 0.159, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.84, + 0.154, + 0.851 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_1 = 0 - 1s\\)" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.794, + 0.235, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.84, + 0.23, + 0.851 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_2 = 1 - 2s\\)" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.794, + 0.312, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.84, + 0.305, + 0.851 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_3 = 2 - 3s\\)" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.794, + 0.388, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.84, + 0.381, + 0.851 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_4 = 3 - 4s\\)" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.794, + 0.465, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.84, + 0.457, + 0.851 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_{5} = 4 - 5s\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.855, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Figure 2. Average heat maps for each one-second interval. Note that a center bias occurs, similar to image saliency prediction's average ground truth maps." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.088, + 0.875, + 0.203 + ], + "angle": 0, + "content": "
T1T2T3T4T5
T11.000.700.540.500.54
T20.701.000.730.660.65
T30.540.731.000.750.70
T40.500.660.751.000.73
T50.540.650.700.731.00
Average0.660.750.740.730.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.214, + 0.894, + 0.285 + ], + "angle": 0, + "content": "Table 1. Correlation scores (CC) of the temporal slices with each other in a single image, averaged over all images. All slices show more similarity to their direct temporal neighbors. The last row shows the average similarity of a slice with the other slices, \\(\\mathcal{T}_1\\) being the most dissimilar one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.435 + ], + "angle": 0, + "content": "We plot the differences of the consecutive average temporal slices in Figure 3 to illustrate attention shifts. Light blue indicates the regions with reduced attention, whereas (light) red indicates increased attention. We observe that attention shifts from left to right, with a subsequent dispersion from the center towards the corners. Then, attention increases at the center of the image, slightly skewed to the left. Interestingly, the trend (especially in \\(\\mathcal{T}_2 - \\mathcal{T}_1\\)) coincides with the western left-to-right reading direction [8]." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.45, + 0.598, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.506, + 0.585, + 0.52 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_2 - \\mathcal{T}_1\\)" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.451, + 0.695, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.506, + 0.677, + 0.52 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_3 - \\mathcal{T}_2\\)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.451, + 0.793, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.506, + 0.776, + 0.52 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_4 - \\mathcal{T}_3\\)" + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.451, + 0.889, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.506, + 0.875, + 0.52 + ], + "angle": 0, + "content": "\\(\\mathcal{T}_{5} - \\mathcal{T}_{4}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.533, + 0.894, + 0.575 + ], + "angle": 0, + "content": "Figure 3. Differences of the consecutive average temporal slices shown in Fig. 2. Red indicates regions of increased attention whereas blue indicates decreased attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Inter-slice similarity across time. We expect temporal saliency slices to be more similar to their closer-in-time slices than to the ones further away since human attention is continuous over time. Table 1 contains the correlation coefficients between each pair of saliency slices in a single image, averaged over all images. We calculate the correlation coefficient between slices \\(\\mathcal{T}_j\\) and \\(\\mathcal{T}_k\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.711, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C C} \\left(\\mathcal {T} _ {j}, \\mathcal {T} _ {k}\\right) = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\operatorname {C C} \\left(\\mathcal {T} _ {i j}, \\mathcal {T} _ {i k}\\right), \\quad j, k \\in \\{1, \\dots , 5 \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.89, + 0.796 + ], + "angle": 0, + "content": "where \\(N\\) is the total number of images, and \\(\\mathcal{T}_{ij}\\) and \\(\\mathcal{T}_{ik}\\) denote the \\(j^{th}\\) and \\(k^{th}\\) slice of the \\(i^{th}\\) image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.887 + ], + "angle": 0, + "content": "By calculating t-test scores on the pairwise comparisons, we observe that all of the pairwise differences except \\(\\mathcal{T}_1\\), \\(\\mathcal{T}_3\\) and \\(\\mathcal{T}_1\\), \\(\\mathcal{T}_5\\) are statistically significant (\\(p < 0.01\\)). Thus, the attention residuals between different time intervals in one image are significantly different. We provide more details in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6463" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.09, + 0.088, + 0.44, + 0.126 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c c c c c c} & \\mathcal {T} _ {1} & \\mathcal {T} _ {2} & \\mathcal {T} _ {3} & \\mathcal {T} _ {4} & \\mathcal {T} _ {5} \\\\ \\text {C C} & \\hline 0. 5 7 4 & \\hline 0. 4 3 3 & \\hline 0. 4 3 1 & \\hline 0. 4 2 6 & \\hline 0. 4 4 7 \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.194 + ], + "angle": 0, + "content": "Table 2. Correlation scores (CC) of each time slice in a single image with the average maps presented in Figure 2, averaged over all images. The similarity between slices across images decreases with time, with the exception of the last slice." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.469, + 0.303 + ], + "angle": 0, + "content": "Intra-slice similarity across images. We also investigate the deviation of each slice from its respective average slices. The average slices are depicted in Figure 2. Table 2 shows the deviation of a slice from the average time slices per image. We compute CC scores between a single slice and the corresponding average slice as" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.314, + 0.469, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C C} \\left(\\mathcal {T} _ {j}, A _ {j}\\right) = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\operatorname {C C} \\left(\\mathcal {T} _ {i j}, A _ {j}\\right), \\quad j \\in \\{1, \\dots , 5 \\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.341, + 0.384 + ], + "angle": 0, + "content": "where \\(A_{j}\\) denotes the \\(j^{th}\\) average slice." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.469, + 0.536 + ], + "angle": 0, + "content": "We average the scores across all images. Higher values of CC indicate more agreement with the average whereas lower values of CC indicate more deviation from the average. Note that the similarity with the average across images decreases with time, except for the last slice. This can be explained by the more prominent center bias in \\(\\mathcal{T}_5\\) as seen in Figure 2. \\(\\mathcal{T}_1\\) has the least deviation from the average by a significant margin (\\(p \\ll 0.01\\)). This shows that humans tend to look at similar places at first, and then their attention scatters around to less important regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.469, + 0.747 + ], + "angle": 0, + "content": "Early and late fixations versus saliency. Lastly, we investigate the relationship between the fixation timestamps and their respective saliency values. We assign a saliency value to each fixation as the normalized pixel value in the corresponding saliency map. We plot the histogram of the number of fixations with their saliency and timestamp values in Figure 4. The fixation time stamps range from 0 to \\(5000\\mathrm{ms}\\) and the saliency values range from 0 to 1. Late fixations tend to have lower saliency values than earlier fixations, as indicated by the darker color towards the bottom right corner. That is, the first region we glance at in an image is more important (salient) than the following regions [7, 18]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.212, + 0.779 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.233, + 0.803 + ], + "angle": 0, + "content": "4.1. Temporal slices" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We aim to recover fixation timestamps to train models with this temporal information. We extract temporal slices by grouping the fixations in several time intervals and, following common practice [5], blurring with a Gaussian kernel. We break down the fixations into time slices with two time-slicing (grouping) alternatives, namely equal duration" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.097, + 0.897, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.303, + 0.894, + 0.387 + ], + "angle": 0, + "content": "Figure 4. Number of fixations with their respective saliency values and timestamps. Lighter colors indicate a higher number of occurrences while darker areas denote fewer occurrences. We see that late fixations tend to be less salient, which can be seen as the decrease in the number of salient fixations along the arrow. The most salient fixations appear at approximately 1s." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.893, + 0.466 + ], + "angle": 0, + "content": "and equal distribution. The equal duration model outperforms the equal distribution one and is easier to interpret; we present a comparison of these two alternatives as an ablation study in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.727, + 0.489 + ], + "angle": 0, + "content": "4.2. Temporal saliency model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.496, + 0.893, + 0.707 + ], + "angle": 0, + "content": "Let us now introduce our framework that exploits temporal human attention information. Our model is depicted in Figure 5. We extract image features using a pre-trained object recognition encoder [33]. Then, we decode these features by a temporal slice decoder to obtain one saliency map per time slice. These temporal saliency slices are useful in automated visual appeal assessment in applications such as website design [32], advertisement [36] and infographics [13]. In parallel, we decode the same image features into an initial image saliency prediction. Finally, we combine the temporal slices and the image saliency predictions in the spatiotemporal mixing module to produce a final image saliency map. We describe each component in detail in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.812, + 0.74 + ], + "angle": 0, + "content": "4.2.1 Image encoder and saliency decoders" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.749, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Following the previous saliency prediction architectures [26, 31, 40], we first encode the input image with a pretrained image classification network, in our case PNASNet-5 [33]. We extract encoded features at various levels for multi-level integration, similar to a U-Net structure [41]. Formally, we denote the image encoder as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.847, + 0.891, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} (\\mathcal {I}) = \\left[ \\mathcal {E} _ {i} \\right], \\quad i \\in \\{1, \\dots , 5 \\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) is the input image, and \\(\\mathcal{E}_i\\) the \\(i^{\\mathrm{th}}\\) encoder block. The output of \\(\\mathcal{E}(\\cdot)\\) therefore is a 5D vector. The early en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6464" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.896, + 0.373 + ], + "angle": 0, + "content": "Figure 5. Overview of the proposed architecture. We encode image features into encoder blocks consisting of multi-level image features. We then pass these blocks to the temporal saliency decoder (shown in orange) to decode them into temporal saliency predictions, which are saliency maps in sequential time intervals. In parallel, the image saliency decoder (shown in green) decodes the encoder blocks into an image saliency prediction. We then combine the temporal saliency maps, the image saliency map, and the encoder blocks in the spatiotemporal mixing module (shown in pink). (Best viewed in color.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.47, + 0.468 + ], + "angle": 0, + "content": "coder blocks extract low-level features, such as edges, color, and contrast, while the later blocks encode high-level semantics. We pass these blocks to the temporal saliency decoder, the image saliency decoder, and the spatiotemporal mixing module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.575 + ], + "angle": 0, + "content": "Our temporal slice decoder, namely \\(\\mathcal{D}_T\\), processes the encoder blocks with four 3x3 convolution layers followed by ReLU functions, integrating one encoder block after each convolution. Later, two 3x3 convolution layers with a ReLU in-between and a sigmoid function at the end produce \\(n\\) temporal saliency maps. Formally, we write the temporal saliency decoder as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.587, + 0.47, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {T} \\big (\\mathcal {E} (\\mathcal {I}) \\big) = [ \\mathcal {T} _ {n} ] =: \\mathcal {T}, \\quad n \\in \\{1, \\dots , 5 \\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}_n\\) denotes the \\(n^{\\mathrm{th}}\\) temporal saliency slice. Through this branch of the network, a single image input produces \\(n\\) temporal saliency slices. We use this component to provide temporal predictions to the spatiotemporal mixing module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.754 + ], + "angle": 0, + "content": "Our image saliency decoder, namely \\(\\mathcal{D}_S\\), has the same structure as \\(\\mathcal{D}_T\\), with the exception of the number of output channels. This component produces a single map \\(S_I\\), which corresponds to the conventional image saliency map of the input image. As such, we can write" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.765, + 0.47, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {I} = \\mathcal {D} _ {S} (\\mathcal {E} (\\mathcal {I})). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.827 + ], + "angle": 0, + "content": "We use this module to provide cumulative saliency information to the spatiotemporal mixing module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.353, + 0.863 + ], + "angle": 0, + "content": "4.2.2 Spatiotemporal Mixing Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To incorporate temporal information into the saliency prediction, we introduce a module that combines temporal and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.895, + 0.543 + ], + "angle": 0, + "content": "spatial saliency maps. Our design is inspired by the feature pyramid networks [29], which have the benefit of integrating features from multiple levels of the encoder, thus capturing low-level cues (such as color contrast, brightness, and edges) and high-level ones (such as semantics and scene context). Such cues have been shown to be critical for saliency estimation [15, 24, 38]. This module takes temporal saliency predictions, the initial image saliency prediction, and the encoded image feature blocks as input. We write this as" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.55, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {R} = \\operatorname {S M M} \\left(\\mathcal {E} (\\mathcal {I}), \\mathcal {T}, \\mathcal {S} _ {I}\\right), n \\in \\{1, \\dots , 5 \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.871, + 0.588 + ], + "angle": 0, + "content": "where \\(S_{R}\\) denotes the final, refined image saliency map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.893, + 0.83 + ], + "angle": 0, + "content": "The module architecture is shown in Figure 6. It takes the last two encoder blocks \\([\\mathcal{E}_5, \\mathcal{E}_4]\\) and passes them through a \\(3 \\times 3\\) convolution. We then concatenate the other encoder blocks with the image saliency and temporal saliency maps passing through \\(3 \\times 3\\) convolution, ReLU, and linear upsampling to keep the spatial dimensions consistent. We add the static and temporal saliency maps to the encoded image features at each block to prevent the information in these maps from vanishing. In the last step, we only add the saliency maps to output a final refined saliency map \\(S_R\\). The final design of the module (i.e., combining E4 and E5) is based on empirical performance. This module eliminates the need for optimizing a weight parameter between the spatial and temporal maps. It can also modulate the maps within the spatial range of convolutions, which allows the selection of different regions from different maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.661, + 0.862 + ], + "angle": 0, + "content": "4.2.3 Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "To train our network, we use the Kullback-Leibler divergence (KL) [45] and the Correlation Coefficient (CC)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6465" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.487, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.473, + 0.346 + ], + "angle": 0, + "content": "Figure 6. The spatiotemporal mixing module combines temporal saliency predictions with the conventional image saliency prediction by multi-level image feature integration. \\( S_{I} \\) denotes the predicted image saliency map, \\( T_{1,\\dots,n} \\) the temporal saliency predictions for \\( n \\) frames, and \\( \\mathcal{E}_{1,\\dots,5} \\) the encoder blocks. This multi-level integration scheme provides information from earlier layers of the network to the next blocks in this module. \\( S_{R} \\) denotes the temporally refined image saliency map output." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.37, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[21] between the predicted and ground truth saliency maps. First, we train the temporal branch using" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.411, + 0.47, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} (\\mathcal {I}) = \\lambda_ {1} * \\operatorname {C C} \\left(G T _ {n}, \\mathcal {T} _ {n}\\right) + \\beta_ {1} * \\operatorname {K L} \\left(G T _ {n}, \\mathcal {T} _ {n}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.47, + 0.485 + ], + "angle": 0, + "content": "where \\(GT_{n}\\) denotes the temporal ground truth for the \\(n^{th}\\) slice. We then freeze the weights in this component and train the spatiotemporal mixing module using" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.494, + 0.47, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} (\\mathcal {I}) = \\lambda_ {2} * \\operatorname {C C} (G T, \\mathcal {S} _ {R}) + \\beta_ {2} * \\operatorname {K L} (G T, \\mathcal {S} _ {R}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.521, + 0.462, + 0.537 + ], + "angle": 0, + "content": "where \\(GT\\) is the image saliency ground truth for image \\(I\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.549, + 0.313, + 0.566 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.574, + 0.268, + 0.591 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.597, + 0.471, + 0.689 + ], + "angle": 0, + "content": "We use a batch size of 32 and an initial learning rate of 1e-4, reduced by a factor of ten every two epochs. We train the temporal branch first and then freeze the weights. We found that 10 epochs of training on SALICON was sufficient. For SALICON, we used the provided test, train, and validation splits." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.174, + 0.71 + ], + "angle": 0, + "content": "5.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.749 + ], + "angle": 0, + "content": "We evaluate the obtained saliency predictions according to the following standard metrics used by the community." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Area Under the Curve (AUC) [5]: Saliency prediction can be interpreted as classifying fixation vs non-fixation points. The area under the ROC curve shows the trade-off between true positives (TP) and false positives (FP). A higher AUC score indicates fewer FPs. While AUC computes the TP and FP rates using all the ground truth fixation points, sAUC [4] samples FP points from ground truth fixations of other observers, compensating for center bias in natural images by taking for inter- and intra-observer variability into account. Normalized Scanpath Saliency (NSS) [38]: This metric" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "compares the predicted saliency values at the ground truth fixation points to the average predicted saliency. An NSS score of one indicates that the predicted saliency values at the ground truth fixation points are one standard deviation above the average." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.892, + 0.242 + ], + "angle": 0, + "content": "Kullback - Leibler Divergence (KL) [45]: The KL measures the cumulative distance between the predicted and the ground truth saliency maps. A KL score close to zero indicates a better approximation of the ground truth saliency map by the predicted one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Pearson's correlation coefficient (CC) [21]: This metric measures the linear relationship between the predicted and ground truth saliency maps. It ranges from -1 to 1. A CC score close to one indicates a strong linear correlation between the two maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Similarity (SIM) score [22]: The similarity score sums the minimum value between the predicted and the ground truth saliency maps over all pixels. A similarity score of 1 indicates a perfect prediction since both of the maps are probability distributions summing to 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Information Gain (IG) score [27]: The information gain is an information-theoretic metric that measures the difference in average log-likelihood between the predicted saliency map and center-bias prior." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.694, + 0.485 + ], + "angle": 0, + "content": "5.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.706 + ], + "angle": 0, + "content": "We compare our method with the state-of-the-art models, namely SAM-Resnet [10], MSI-Net, GazeGAN, MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14], in image saliency prediction. Our model outperforms these methods in five out of seven metrics, showing the benefit of incorporating temporal information. Moreover, our model outperforms the only other multi-duration saliency model [14] in image saliency prediction by a significant margin. Furthermore, we compare our model with this multi-duration model and a multi-duration baseline trained on individual temporal slices. Our model improves the saliency prediction in two durations consisting of 0.5 and 3 seconds in two out of three metrics and in all three metrics in the five-second duration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.723, + 0.859, + 0.739 + ], + "angle": 0, + "content": "5.4. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We first evaluate the performance of our model on image saliency prediction on the SALICON benchmark [20]. The ground truth of SALICON's test set is exclusively hosted on the CodaLab website2. Table 3 shows the comparison of standard evaluation metrics for different state-of-the-art saliency models alongside our model TempSAL. TempSAL outperforms all the baselines in almost all metrics. When it does not, it still yields competitive results." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.804, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://competitions.codalab.org/competitions/17136" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6466" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.087, + 0.849, + 0.25 + ], + "angle": 0, + "content": "
ModelMDAUC ↑CC ↑KL ↓SAUC ↑IG ↑NSS ↑SIM ↑
SAM-Resnet [10]X0.8650.8990.6100.7410.5381.9900.793
MSI-Net [24]X0.8650.8990.3070.7360.7931.9310.784
GazeGAN [6]X0.8640.8790.3760.7360.7201.8990.773
SimpleNet [40]X0.8690.9070.2010.7430.8801.9600.793
MDNSal [40]X0.8650.8990.2210.7360.8631.9350.790
UNISAL [12]X0.8640.8790.3540.7390.7801.9520.775
DeepGaze IIE [31]X0.8690.8720.2850.7670.7661.9960.733
MD-SEM [14]✓0.8640.8680.5680.7460.6602.0580.774
TempSAL✓0.8690.9110.1950.7450.8961.9670.800
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.259, + 0.895, + 0.331 + ], + "angle": 0, + "content": "Table 3. Evaluation results on the SALICON (LSUN 2017) test benchmark. We compare our model with the state-of-the-art saliency prediction models, namely SAM-Resnet [10], MSI-Net [24], GazeGAN [6], MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14]. The results in bold show the best performance. Our method outperforms the state-of-the-art on conventional image saliency in five metrics. The MD column denotes the ability of the models to predict multi-duration saliency. Our model outperforms the only other multi-duration saliency model by a significant margin on five out of seven metrics." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.342, + 0.885, + 0.429 + ], + "angle": 0, + "content": "
ModelSlice 1 (0-500 ms)Slice 2 (0-3000 ms)Slice 3 (0-5000 ms)Average
CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑
SAM-MD [14]0.8050.3703.1810.7380.4692.5410.7150.5352.4950.7530.4582.739
MD-SEM [14]0.8160.3513.3740.7450.4522.6940.7340.4872.6770.7650.4302.915
TempSAL0.8190.4963.4220.7520.5122.7030.8220.4713.3370.7970.4933.154
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.438, + 0.895, + 0.507 + ], + "angle": 0, + "content": "Table 4. Results of our model, MD-SEM [14], and the baseline SAM-MD [14] across different durations on the CodeCharts1k dataset [14]. Our model improves the saliency prediction in the first two intervals, consisting of \\(500~\\mathrm{ms}\\) and \\(3000~\\mathrm{ms}\\) observations, in two out of three metrics. In this comparison, the time slices are cumulative, not mutually exclusive. Although our model benefits from non-overlapping temporal slices, it also performs well with cumulative time slices, particularly in the last slice corresponding to the image saliency with an observation duration of \\(5000~\\mathrm{ms}\\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.53, + 0.457, + 0.546 + ], + "angle": 0, + "content": "5.5. Comparison with the multi-duration method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.471, + 0.658 + ], + "angle": 0, + "content": "To compare our method with the only other multiduration model [14], we modify our network to output three temporal slices. We train our network on a three slice SALICON multi-duration dataset first and then fine-tune it on the CodeCharts1k dataset [14] using the given training and validation splits. We report the results of the comparison in Table 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.262, + 0.683 + ], + "angle": 0, + "content": "5.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.78 + ], + "angle": 0, + "content": "In Figure 7, we compare the temporal and image saliency maps obtained with our method with the ground truth from SALICON [20]. Our model learns time-specific predictions and is able to combine such predictions into a conventional image saliency map. We provide additional qualitative results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.24, + 0.803 + ], + "angle": 0, + "content": "5.7. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we investigate the effect of different components in our model. We also provide a comparison with a multi-duration baseline model on the temporal SALICON dataset. Moreover, we provide an ablation study on the effect of the number of time slices and the two temporal slicing methods in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.698 + ], + "angle": 0, + "content": "Effect of the SMM module: We evaluate the effect of the spatiotemporal mixing module (SMM) and the image saliency decoder in Table 5. The first model consists of the image encoder and temporal saliency decoder only. We take the average of the temporal slices to measure its performance by comparing it with the ground truth image saliency map. In the second row, we add the image saliency decoder to our model. Similarly, we take the average of the predicted maps \\(\\mathcal{T}_n\\) and \\(S_I\\). Lastly, we add the spatiotemporal mixing module, which effectively modulates these predicted maps and combines them into a final image saliency map \\(S_R\\)." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.705, + 0.873, + 0.774 + ], + "angle": 0, + "content": "
ModelCC ↑KL ↓NSS ↑SIM ↑
DT(E(I))0.8520.2431.9730.754
+DS(E(I))0.8570.2521.9430.760
+SMM0.9060.1981.9300.798
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.895, + 0.898 + ], + "angle": 0, + "content": "Table 5. Results of ablation studies on the temporal SALICON validation dataset. The first row denotes the model with only the temporal saliency decoder. In the second row, the model has both the temporal and image saliency decoders. The last row denotes the performance with the spatiotemporal mixing module (SMM). As evidenced by the improved accuracy metrics, the SMM effectively modulates the spatial and temporal saliency maps to refine the initial image saliency prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6467" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.868, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.49, + 0.893, + 0.56 + ], + "angle": 0, + "content": "Figure 7. The first row shows the input image with the ground truth saliency overlaid. The second row shows the ground truth temporal saliency and image saliency. We show our temporal saliency and image saliency predictions on the third row. Red-yellow maps are temporal saliency maps for one-second intervals. Black and white maps are image saliency maps for the whole observation duration. Our approach captures the attention shifts in sequential temporal maps. Moreover, our model is able to produce accurate image saliency predictions which are close to the ground truth maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Comparison with a temporal baseline model: The performance of our TempSAL model with five temporal slices is provided in Table 6. Note that each saliency slice contains five times fewer samples than the original image saliency map. Therefore, individual slices contain more variation compared to conventional accumulated maps. As a baseline to our model, we compute the performance of an architecture composed of five replicated SimpleNet models (5xSimpleNet) [40], each trained on one saliency slice. This baseline model uses an unshared encoder and decoder for each slice, while we share the decoder among slices. Therefore, we do not benefit from increased model capacity. We observe an accuracy decline in the baseline model, which confirms the increased discrepancy in the data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.196, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We present a saliency prediction method that can learn time-specific predictions and is also able to exploit temporal information to improve overall image saliency prediction. In particular, we show that the temporally evolving patterns in human attention play an important" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.563, + 0.888, + 0.671 + ], + "angle": 0, + "content": "
TimeBaselineTempSAL
CC↑KL↓NSS↑SIM↑CC↑KL↓NSS↑SIM↑
T10.8980.2112.4360.7780.8990.2142.4530.782
T20.8700.2192.1590.7650.8770.2152.2110.776
T30.8400.2471.8400.7530.8430.2471.8780.758
T40.8200.2731.7290.7430.8250.2641.7400.749
T50.8110.2751.6460.7380.8130.2761.6540.743
Average0.8480.2451.9620.7560.8520.2431.9870.761
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.677, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Table 6. Results of the baseline model (left) and our TempSAL model (right) across different time slices. In 18 out of 20 comparisons, our model consistently outperforms the baseline. Note that both models perform best in the first slice, in which the intra-slice agreement is more prominent than in the other slices, as mentioned in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.774, + 0.892, + 0.85 + ], + "angle": 0, + "content": "role in saliency prediction in natural images. This is evidenced by our experiments that demonstrate our TempSAL method outperforming the state-of-the-art, including a multi-duration method exploiting cumulative temporal saliency maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the Swiss National Science Foundation via the Sinergia grant CRSII5-180359." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6468" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta and Sabine Süsstrunk. Saliency detection for content-aware image resizing. In IEEE International Conference on Image Processing (ICIP), pages 1005-1008, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Bahar Aydemir, Deblina Bhattacharjee, Seungryong Kim, Tong Zhang, Mathieu Salzmann, and Sabine Süsstrunk. Modeling object dissimilarity for deep saliency prediction. Transactions on Machine Learning Research (TMLR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Ali Borji and Laurent Itti. CAT2000: A large scale fixation dataset for boosting saliency research. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Ali Borji, Dicky N. Sihite, and Laurent Itti. Quantitative analysis of human-model agreement in visual saliency modeling: A comparative study. IEEE Transactions on Image Processing (TIP), 22(1):55-69, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[5] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Fredo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41(3):740, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Zhaohui Che, Ali Borji, Guangtao Zhai, Xiongkuo Min, Guodong Guo, and Patrick Le Callet. Gazegan: A generative adversarial saliency model based on invariance analysis of human gaze during scene free viewing. ArXiv, abs/1905.06803, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Ming-Ming Cheng, Niloy J. Mitra, Xiaolei Huang, Philip H. S. Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 37(3):569-582, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[8] Hannah Faye Chua, Julie E. Boland, and Richard E. Nisbett. Cultural variation in eye movements during scene perception. Proceedings of the National Academy of Sciences, 102(35):12629-12633, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.672 + ], + "angle": 0, + "content": "[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. A deep multi-level network for saliency prediction. In IEEE International Conference on Pattern Recognition (ICPR), pages 3488-3493, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[10] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an LSTM-based saliency attentive model. IEEE Transactions on Image Processing (TIP), 27(10):5142-5154, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.731, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Richard Droste, Jianbo Jiao, and J. Alison Noble. Unified Image and Video Saliency Modeling. In European Conference on Computer Vision (ECCV), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Camilo Fosco, Vincent Casser, Amish Kumar Bedi, Peter O'Donovan, Aaron Hertzmann, and Zoya Bylinskii. Predicting visual importance across graphic design types. In ACM Symposium on User Interface Software and Technology (UIST), page 249-260, 2020. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[14] Camilo Fosco, Anelise Newman, Pat Sukhum, Yun Bin Zhang, Nanxuan Zhao, Aude Oliva, and Zoya Bylinskii. How much time do you have? modeling multi-duration saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4482, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[15] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 34(10):1915-1926, 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.893, + 0.278 + ], + "angle": 0, + "content": "[16] Joseph H. Goldberg and Jonathan I. Helfman. Visual scanpath representation. In Symposium on Eye-Tracking Research & Applications (ETRA), pages 203–210, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.894, + 0.349 + ], + "angle": 0, + "content": "[17] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. SALICON: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In IEEE International Conference on Computer Vision (ICCV), pages 262-270, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.352, + 0.893, + 0.407 + ], + "angle": 0, + "content": "[18] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 20(11):1254-1259, 1998. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.409, + 0.893, + 0.452 + ], + "angle": 0, + "content": "[19] Sen Jia and Neil D. B. Bruce. EML-NET: An expandable Multi-Layer NETwork for saliency prediction. Image and Vision Computing, 95:103887, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[20] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. SALICON: Saliency in context. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.567 + ], + "angle": 0, + "content": "[21] Timothee Jost, Nabil Ouerhani, Roman von Wartburg, René Muri, and Heinz Hügli. Assessing the contribution of color in visual attention. Computer Vision and Image Understanding, 100(1-2):107-123, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.893, + 0.611 + ], + "angle": 0, + "content": "[22] Tilke Judd, Fredo Durand, and Antonio Torralba. A benchmark of computational models of saliency to predict human fixations. MIT Technical Report, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.613, + 0.893, + 0.668 + ], + "angle": 0, + "content": "[23] Tilke Judd, Krista Ehinger, Fredo Durand, and Antonio Torralba. Learning to predict where humans look. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.893, + 0.727 + ], + "angle": 0, + "content": "[24] Alexander Kroner, Mario Senden, Kurt Driessens, and Rainer Goebel. Contextual encoder-decoder network for visual saliency prediction. Neural Networks, 129:261 - 270, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.893, + 0.785 + ], + "angle": 0, + "content": "[25] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep Gaze I: Boosting saliency prediction with feature maps trained on ImageNet. In International Conference on Learning Representations (ICLR) Workshops, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.787, + 0.893, + 0.842 + ], + "angle": 0, + "content": "[26] Matthias Kümmerer, Tom Wallis, and Matthias Bethge. DeepGaze II: Predicting fixations from deep features over time and tasks. Journal of Vision (JOV), 17(10):1147, 2017. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Matthias Kümmerer, Thomas S. A. Wallis, and Matthias Bethge. Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences (PNAS), 112(52):16054-16059, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6469" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[28] Qingshan Li, Yue Zhou, and Jie Yang. Saliency based image segmentation. In IEEE International Conference on Multimedia Technology, pages 5068-5071, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.47, + 0.349 + ], + "angle": 0, + "content": "[31] Akis Linardos, Matthias Kümmerer, Ori Press, and Matthias Bethge. Deepgaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12919-12928, 2021. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.352, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[32] Gitte Lindgaard, Gary Fernandes, Cathy Dudek, and J. Brown. Attention web designers: You have 50 milliseconds to make a good first impression! Behaviour & Information Technology, 25(2):115-126, 2006. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[33] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In European Conference on Computer Vision (ECCV), pages 19-34, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[34] Nian Liu and Junwei Han. A deep spatial contextual long-term recurrent convolutional network for saliency detection. IEEE Transactions on Image Processing (TIP), 27(7):3264-3274, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.469, + 0.581 + ], + "angle": 0, + "content": "[35] Sandeep Mishra and Oindrila Saha. RecSal : Deep recursive supervision for visual saliency prediction. In British Machine Vision Conference (BMVC), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.469, + 0.639 + ], + "angle": 0, + "content": "[36] Johanna Palcu, Jennifer Sudkamp, and Arnd Florack. Judgments at gaze value: Gaze cuing in banner advertisements, its effect on attention allocation and product judgments. Frontiers in Psychology, 8, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[37] Yash Patel, Srikar Appalaraju, and R. Manmatha. Saliency driven perceptual image compression. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 227-236, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[38] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397-2416, 2005. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[39] Michael Posner, Robert Rafal, Lisa Choatec, and Jonathan Vaughan. Inhibition of return: Neural basis and function. Cognitive Neuropsychology, Vol. 2:211 - 228, 09 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[40] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In International Conference on Intelligent Robots and Systems (IROS), 2020. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations (ICLR), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[43] Markus Steinberger, Bernhard Kainz, Stefan Hauswiesner, Rostislav Khlebnikov, Denis Kalkofen, and Dieter Schmalstieg. Ray prioritization using stylization and visual saliency. Computers & Graphics, 36(6):673-684, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[44] Po-He Tseng, Ran Carmi, Ian G. M. Cameron, Douglas P. Munoz, and Laurent Itti. Quantifying center bias of observers in free viewing of dynamic natural scenes. Journal of Vision (JOV), 9(7):4-4, 07 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[45] Mathukumalli Vidyasagar. Kullback-leibler divergence rate between probability distributions on sets of different cardinalities. In IEEE Conference on Decision and Control (CDC), pages 948-953, 2010. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[46] Eleonora Vig, Michael Dorr, and David Cox. Large-scale optimization of hierarchical features for saliency prediction in natural images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2798-2805, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[47] W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, and R. Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(06):3239-3259, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[48] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. IEEE Transactions on Multimedia, 22(8):2163-2176, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[49] Alfred L. Yarbus. Eye Movements and Vision, volume 2. Plenum Press, 1967. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[50] Yifeng Zhang, Ming Jiang, and Qi Zhao. Saliency prediction with external knowledge. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 484-493, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[51] Jufeng Zhao, Yueting Chen, Huajun Feng, Zhihai Xu, and Qi Li. Fast image enhancement using multi-scale saliency extraction in infrared imagery. Optik, 125(15):4039-4042, 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "6470" + } + ] +] \ No newline at end of file diff --git a/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_origin.pdf b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..27eb7cfa7a7b33284862ff429f1d0005f2155d38 --- /dev/null +++ b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/a96ae848-89c1-42c8-883e-3ee6e955a530_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80eabd46223b5a5691348c2dc22839d9abd293d05f2f7fdf327e01a1eeb9d445 +size 1205376 diff --git a/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/full.md b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9611fe67d84bbcbc1afb30423a1b0bea35fb941c --- /dev/null +++ b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/full.md @@ -0,0 +1,371 @@ +# TempSAL - Uncovering Temporal Information for Deep Saliency Prediction + +Bahar Aydemir, Ludo Hoffstetter, Tong Zhang, Mathieu Salzmann, Sabine Süssstrunk +School of Computer and Communication Sciences, EPFL, Switzerland + +{bahir/aydemir, tong.zhang, mathieu.salzmann, sabine.susstrunk}@epfl.ch + +![](images/307199b203a9151932e8b29750a4934c05f1f6976b3e556d1ac0d7db25e64408.jpg) +Input image + +![](images/e15e40f345eb4b7fad635135f663abdfdbc88a11d76ea3ed497b8de67ec5b896.jpg) + +![](images/760dc8cce4cc441bcc3251a3efe1f44fa0df592a9af200b8c3ff8d8a505de97c.jpg) +$\mathcal{T}_1 = 0 - 1s$ + +![](images/322905c8f2176207b7b1e982390b7749d7d3c7950cd96ce07140a8975b26d2f3.jpg) + +![](images/a574130edd1b16f31495a2035538b3ecbe3b4119b7f0c9ad428335419109cea9.jpg) +$\mathcal{T}_2 = 1 - 2s$ +Figure 1. An example of how human attention evolves over time when observing a single image. Top row: Temporal and image saliency ground truth from the SALICON dataset [20]. Bottom row: Our temporal and image saliency predictions. Each temporal saliency map $\mathcal{T}_i$ , $i \in \{1, \dots, 5\}$ represents one second of observation time. Note that in $\mathcal{T}_1$ , the chef is salient, while in $\mathcal{T}_2$ and $\mathcal{T}_3$ , the food on the barbecue becomes the most salient region in this scene. We can predict the temporal saliency maps for each interval separately, or combine them to create a single, refined image saliency map for the entire observation period. + +![](images/a2c271bcedad4b8e69371aaaac4523ff87dcf3d9b534b545c989479dfe12f5d8.jpg) + +![](images/7f77badc9d6fc58a890254a7584666f696dd04e65abce8b360651d5f929dfb65.jpg) +$\mathcal{T}_3 = 2 - 3s$ + +![](images/55ab983def07961ecc49365fe9b8a5a98d1054c71fcc284cdfeca0438010f8f1.jpg) + +![](images/d05853d6cade2062efa0a306798bc71c23afbc0eee60418eaf2dedd517b692f6.jpg) +$\mathcal{T}_4 = 3 - 4s$ + +![](images/19db42a72cd02b4933528fca4d9cb92cedee2fee606ad0ceeac9de5851ff626f.jpg) + +![](images/e868de0d03bd3fa78ce3014e1244435e5fb70ded7272c6835c8a66c6dde1205f.jpg) +$\mathcal{T}_{5} = 4 - 5s$ + +![](images/26934dac085f6443b31ee8d659c89b76270b622bf4ac1875d24c0aeb8276bcce.jpg) + +![](images/4eff5060045e80a783b3c018b1891ef6b211dea6e8315fb2ee105f12e4060b2b.jpg) +0-5s + +# Abstract + +Deep saliency prediction algorithms complement the object recognition features, they typically rely on additional information such as scene context, semantic relationships, gaze direction, and object dissimilarity. However, none of these models consider the temporal nature of gaze shifts during image observation. We introduce a novel saliency prediction model that learns to output saliency maps in sequential time intervals by exploiting human temporal attention patterns. Our approach locally modulates the saliency predictions by combining the learned temporal maps. Our experiments show that our method outperforms the state-of-the-art models, including a multi-duration saliency model, on the SALICON benchmark and CodeCharts1k dataset. Our code is publicly available on GitHub1. + +# 1. Introduction + +Humans have developed attention mechanisms that allow them to selectively focus on the important parts of a scene. Saliency prediction algorithms aim to computationally detect these regions that stand out relative to their surroundings. These predictions have numerous applications + +in image compression [37], image enhancement [51], image retargeting [1], rendering [43], and segmentation [28]. + +Since the seminal work of Itti et al. [18], many have developed solutions using both handcrafted features [7] and deep ones [9, 17, 26, 34, 46, 48]. Nowadays, employing deep neural networks is preferred in saliency prediction as they outperform bottom-up models. These methods typically depend on pre-trained object recognition networks to extract features from the input image [31]. In addition to these features, scene context [47], object co-occurrence [50], and dissimilarity [2] have been exploited to improve the saliency prediction. However, while these approaches model the scene context and objects, they fail to consider that humans dynamically observe scenes [49]. In neuroscience, the inhibition of return paradigm states that a suppression mechanism reduces visual attention towards recently attended objects [39] and encourages selective attention to novel regions. Motivated by this principle, we develop a saliency prediction model that incorporates temporal information. + +Fosco et al. [14] also exploit temporal information in saliency prediction, but they consider snapshots containing observations up to 0.5, 3, and 5 seconds, thus not leveraging saliency trajectory but rather saliency accumulation. By contrast, here, we model consecutive time slices, connecting our approach more directly with the human gaze and thus opening the door to automated visual appeal assess- + +ment in applications such as website design [32], advertisement [36] and infographics [13]. + +To achieve this, we show that when viewing images, human attention yields temporally evolving patterns, and we introduce a network capable of exploiting this temporal information for saliency prediction. Specifically, our model learns time-specific predictions and is able to combine them with a conventional image saliency map to obtain a temporally modulated image saliency prediction. + +We evaluate our method on the SALICON [20] and CodeCharts1k [14] datasets which contain temporal information, unlike the other popular saliency datasets such as CAT2000 [3] and MIT1003 [23]. By showing the benefits of estimating temporal saliency we hope to encourage the community to publish the temporal information of their data along with the final saliency maps. Note that existing works typically collect saliency data by conducting psychophysical experiments [3, 14, 20, 23], and the attention data recorded during these experiments already includes temporal information. Therefore, no further experiments are required. + +As evidenced by our experiments, using temporal information boosts the accuracy of the baseline network, enabling us to consistently outperform the state-of-the-art models on the SALICON saliency benchmark. Moreover, in the CodeCharts1k dataset we outperform the multiduration model [14] in two out of three metrics. + +We summarize our contributions as follows: + +- We evidence the presence of temporally evolving patterns in human attention. +- We show that temporal information in the form of a saliency trajectory improves saliency prediction in natural images, providing an investigation of the SALICON dataset for temporal attention shifts. +- We introduce a novel, saliency prediction model, called TempSAL, capable of simultaneously predicting conventional image saliency and temporal saliency trajectories. +- We propose a spatiotemporal mixing module that learns time-dependent patterns from temporal saliency maps. Our approach outperforms the state-of-the-art image saliency models that either do not consider temporal information or encode it in a cumulative manner. + +# 2. Related work + +# 2.1. Saliency prediction for natural images + +Early saliency prediction methods were biologically inspired and bottom-up. In particular, Itti et al. used color, intensity, and orientation contrast [18]. Goferman employed global and local contrast as contextual cues [15]. Judd et + +al. [23] further incorporated mid-level and high-level semantic features, using horizon, face, person, and car detectors. Later, Vig et al. [46] showed that deep neural networks can be applied to saliency prediction. Yet, saliency prediction lacks the large-scale annotated datasets that are available for image classification tasks [11], which prevents training robust models. To overcome this, Kummerer et al. [25] showed that using pre-trained object recognition networks significantly improves saliency predictions. Subsequent state-of-the-art models such as EML-Net [19], DeepGaze2 [26], and SALICON [17] similarly use pretrained convolutional neural network (VGG [42]) encoders. Recent works utilize additional sources of information such as scene context [24], external knowledge [50] and object dissimilarity [2] to improve saliency prediction. Yet, none of these methods take into account the temporal evolution of the human gaze, which occurs even when the image stimuli are static. In our work, we make use of these temporal patterns as an additional source of information to boost conventional image saliency prediction. + +# 2.2. Multi-duration saliency + +Existing image saliency ground truth maps include all fixations made throughout the observation period. Aggregating all these fixations that have different timestamps into a single ground truth map results in the loss of temporal information. Representing the fixations as scanpaths retains the temporal clues by encoding the change of gaze of an individual over time. However, merging numerous scanpaths is challenging [16]. Fosco et al. [14] proposed multiduration saliency to characterize the attention of a group of individuals while taking into account time-dependent attention shifts. The temporal maps they rely on, however, encode the attention distribution of many observers across overlapping time periods of increasing durations. While this is a convenient way of capturing a population's attention patterns, it does not reflect the saliency trajectory over time. Similarly, [35] use order of fixations as sequential metadata for deep supervision but they do not model the evolution of attention through time. In our work, we model multiduration saliency to analyze underlying attention patterns by using mutually exclusive time slices. We provide this temporal information to our spatiotemporal mixing module to refine the initial image saliency prediction with temporal information. Moreover, this lets us predict temporal saliency maps for each second of attention. + +# 3. Temporal saliency data analysis + +# 3.1. Datasets + +SALICON [20] is the largest human attention dataset on natural images. It was created via a crowdsourced mouse tracking experiment, which was shown to be similar to eye + +tracking [20] and widely used in the saliency prediction literature. SALICON consists of 10000 training, 5000 validation and 5000 test images from the MS-COCO dataset [30]. The SALICON dataset provides saliency maps, fixations, and gaze points for each image and observer. A gaze point is a raw data point recorded by a tracking device. It describes the spatial coordinates of the eye/mouse on the associated stimuli at a given timestamp. Conversely, fixations describe the coordinates of the long pause when the eyes are fixated on an image detail. Following common practice in eye tracking experiments, Jiang et al. [20] grouped spatially and temporally close gaze points to create fixations. Since the fixations were created by grouping multiple gaze points, they do not have an associated timestamp. To address this, SALICON-MD [14] assumes that the fixations are uniformly distributed across the total viewing time. We provide a finer approximation for recovering the fixations' timestamps, by minimizing the spatial and temporal distance between a fixation and the nearest gaze point. We refer the reader to the supplementary material for the details of this approximation process. + +We also assess the performance and generalization ability of our method using CodeCharts1k [14], which is the first saliency dataset to report multiple viewing durations. The dataset consists of 1000 images and saliency maps collected via crowdsourcing which correspond to 0.5, 3, and 5 seconds of viewing. We present the results in Section 5.5. + +# 3.2. Temporal patterns in the dataset + +In this section, we examine how temporality evolves during human visual attention. To observe the evolution of attention over time, we slice the data into five slices, one for each second of observation. In particular, we inspect the dissimilarity between slices, the agreement with the average, and the distribution of fixations in the time-saliency space. + +Average maps. Viewing patterns in image saliency experiments tend to show a concentration towards the image center [44], which is known as the photographer's bias or center bias. We observe a similar spatial bias for each temporal slice, shown in Figure 2, where we plot the average heat maps for each temporal slice. Note that the gaze tends to converge to the center of the image as time passes. This means that the observers revisit the previously seen important center regions [49]. + +![](images/ffe0ee55273b1b5895d1b17fe04b06004642875b8d8a7bc7f10cc4297d21eb86.jpg) +$\mathcal{T}_1 = 0 - 1s$ +Figure 2. Average heat maps for each one-second interval. Note that a center bias occurs, similar to image saliency prediction's average ground truth maps. + +![](images/cdbc9e400162f880d3485a1fd4bd135e827c763a7689b971c09434206e2b0767.jpg) +$\mathcal{T}_2 = 1 - 2s$ + +![](images/762975aa094de7f177e9b1017bc5539b11eebe7adb60d6b9ec970289e87a2221.jpg) +$\mathcal{T}_3 = 2 - 3s$ + +![](images/4a217d9d80a4e7dcbd0a9fe997d92e831bf068e747e38f212c188ea185301e7a.jpg) +$\mathcal{T}_4 = 3 - 4s$ + +![](images/5db29a0a754eeedf751a251b47e9bc98f09ff3f614646f3e11f789fec6159343.jpg) +$\mathcal{T}_{5} = 4 - 5s$ + +
T1T2T3T4T5
T11.000.700.540.500.54
T20.701.000.730.660.65
T30.540.731.000.750.70
T40.500.660.751.000.73
T50.540.650.700.731.00
Average0.660.750.740.730.72
+ +Table 1. Correlation scores (CC) of the temporal slices with each other in a single image, averaged over all images. All slices show more similarity to their direct temporal neighbors. The last row shows the average similarity of a slice with the other slices, $\mathcal{T}_1$ being the most dissimilar one. + +We plot the differences of the consecutive average temporal slices in Figure 3 to illustrate attention shifts. Light blue indicates the regions with reduced attention, whereas (light) red indicates increased attention. We observe that attention shifts from left to right, with a subsequent dispersion from the center towards the corners. Then, attention increases at the center of the image, slightly skewed to the left. Interestingly, the trend (especially in $\mathcal{T}_2 - \mathcal{T}_1$ ) coincides with the western left-to-right reading direction [8]. + +![](images/b8fd29b00402c164408e1dba5f14ef592c81129913940cb05b6a5f0eb43daa82.jpg) +$\mathcal{T}_2 - \mathcal{T}_1$ + +![](images/24396633b3c0ea0ddd207908eb52b265e40709d2293359c1a7128cd687539492.jpg) +$\mathcal{T}_3 - \mathcal{T}_2$ +Figure 3. Differences of the consecutive average temporal slices shown in Fig. 2. Red indicates regions of increased attention whereas blue indicates decreased attention. + +![](images/143c619e6ee97624f626f8c2ff6cbbebd8dbb400af7b5cd29361428c02205bfe.jpg) +$\mathcal{T}_4 - \mathcal{T}_3$ + +![](images/f8c35bff611d8b8e0d6037d7d107a197151daecc13b054a5897428ac6b90f9d6.jpg) +$\mathcal{T}_{5} - \mathcal{T}_{4}$ + +Inter-slice similarity across time. We expect temporal saliency slices to be more similar to their closer-in-time slices than to the ones further away since human attention is continuous over time. Table 1 contains the correlation coefficients between each pair of saliency slices in a single image, averaged over all images. We calculate the correlation coefficient between slices $\mathcal{T}_j$ and $\mathcal{T}_k$ as + +$$ +\operatorname {C C} \left(\mathcal {T} _ {j}, \mathcal {T} _ {k}\right) = \frac {1}{N} \sum_ {i = 0} ^ {N} \operatorname {C C} \left(\mathcal {T} _ {i j}, \mathcal {T} _ {i k}\right), \quad j, k \in \{1, \dots , 5 \}, \tag {1} +$$ + +where $N$ is the total number of images, and $\mathcal{T}_{ij}$ and $\mathcal{T}_{ik}$ denote the $j^{th}$ and $k^{th}$ slice of the $i^{th}$ image. + +By calculating t-test scores on the pairwise comparisons, we observe that all of the pairwise differences except $\mathcal{T}_1$ , $\mathcal{T}_3$ and $\mathcal{T}_1$ , $\mathcal{T}_5$ are statistically significant ( $p < 0.01$ ). Thus, the attention residuals between different time intervals in one image are significantly different. We provide more details in the supplementary material. + +$$ +\begin{array}{c c c c c c} & \mathcal {T} _ {1} & \mathcal {T} _ {2} & \mathcal {T} _ {3} & \mathcal {T} _ {4} & \mathcal {T} _ {5} \\ \text {C C} & \hline 0. 5 7 4 & \hline 0. 4 3 3 & \hline 0. 4 3 1 & \hline 0. 4 2 6 & \hline 0. 4 4 7 \end{array} +$$ + +Table 2. Correlation scores (CC) of each time slice in a single image with the average maps presented in Figure 2, averaged over all images. The similarity between slices across images decreases with time, with the exception of the last slice. + +Intra-slice similarity across images. We also investigate the deviation of each slice from its respective average slices. The average slices are depicted in Figure 2. Table 2 shows the deviation of a slice from the average time slices per image. We compute CC scores between a single slice and the corresponding average slice as + +$$ +\operatorname {C C} \left(\mathcal {T} _ {j}, A _ {j}\right) = \frac {1}{N} \sum_ {i = 0} ^ {N} \operatorname {C C} \left(\mathcal {T} _ {i j}, A _ {j}\right), \quad j \in \{1, \dots , 5 \}, \tag {2} +$$ + +where $A_{j}$ denotes the $j^{th}$ average slice. + +We average the scores across all images. Higher values of CC indicate more agreement with the average whereas lower values of CC indicate more deviation from the average. Note that the similarity with the average across images decreases with time, except for the last slice. This can be explained by the more prominent center bias in $\mathcal{T}_5$ as seen in Figure 2. $\mathcal{T}_1$ has the least deviation from the average by a significant margin ( $p \ll 0.01$ ). This shows that humans tend to look at similar places at first, and then their attention scatters around to less important regions. + +Early and late fixations versus saliency. Lastly, we investigate the relationship between the fixation timestamps and their respective saliency values. We assign a saliency value to each fixation as the normalized pixel value in the corresponding saliency map. We plot the histogram of the number of fixations with their saliency and timestamp values in Figure 4. The fixation time stamps range from 0 to $5000\mathrm{ms}$ and the saliency values range from 0 to 1. Late fixations tend to have lower saliency values than earlier fixations, as indicated by the darker color towards the bottom right corner. That is, the first region we glance at in an image is more important (salient) than the following regions [7, 18]. + +# 4. Methodology + +# 4.1. Temporal slices + +We aim to recover fixation timestamps to train models with this temporal information. We extract temporal slices by grouping the fixations in several time intervals and, following common practice [5], blurring with a Gaussian kernel. We break down the fixations into time slices with two time-slicing (grouping) alternatives, namely equal duration + +![](images/38c5d35b9376831b01f7c2dbd658ed646856b95d9e2fb3b228b4a79a0ab66c94.jpg) +Figure 4. Number of fixations with their respective saliency values and timestamps. Lighter colors indicate a higher number of occurrences while darker areas denote fewer occurrences. We see that late fixations tend to be less salient, which can be seen as the decrease in the number of salient fixations along the arrow. The most salient fixations appear at approximately 1s. + +and equal distribution. The equal duration model outperforms the equal distribution one and is easier to interpret; we present a comparison of these two alternatives as an ablation study in the supplementary material. + +# 4.2. Temporal saliency model + +Let us now introduce our framework that exploits temporal human attention information. Our model is depicted in Figure 5. We extract image features using a pre-trained object recognition encoder [33]. Then, we decode these features by a temporal slice decoder to obtain one saliency map per time slice. These temporal saliency slices are useful in automated visual appeal assessment in applications such as website design [32], advertisement [36] and infographics [13]. In parallel, we decode the same image features into an initial image saliency prediction. Finally, we combine the temporal slices and the image saliency predictions in the spatiotemporal mixing module to produce a final image saliency map. We describe each component in detail in the following sections. + +# 4.2.1 Image encoder and saliency decoders + +Following the previous saliency prediction architectures [26, 31, 40], we first encode the input image with a pretrained image classification network, in our case PNASNet-5 [33]. We extract encoded features at various levels for multi-level integration, similar to a U-Net structure [41]. Formally, we denote the image encoder as + +$$ +\mathcal {E} (\mathcal {I}) = \left[ \mathcal {E} _ {i} \right], \quad i \in \{1, \dots , 5 \}, \tag {3} +$$ + +where $\mathcal{I}$ is the input image, and $\mathcal{E}_i$ the $i^{\mathrm{th}}$ encoder block. The output of $\mathcal{E}(\cdot)$ therefore is a 5D vector. The early en + +![](images/abb002f7c2ec13981456acd36eccae123401d7651d21feb97d8dbd2fc00ffcf7.jpg) +Figure 5. Overview of the proposed architecture. We encode image features into encoder blocks consisting of multi-level image features. We then pass these blocks to the temporal saliency decoder (shown in orange) to decode them into temporal saliency predictions, which are saliency maps in sequential time intervals. In parallel, the image saliency decoder (shown in green) decodes the encoder blocks into an image saliency prediction. We then combine the temporal saliency maps, the image saliency map, and the encoder blocks in the spatiotemporal mixing module (shown in pink). (Best viewed in color.) + +coder blocks extract low-level features, such as edges, color, and contrast, while the later blocks encode high-level semantics. We pass these blocks to the temporal saliency decoder, the image saliency decoder, and the spatiotemporal mixing module. + +Our temporal slice decoder, namely $\mathcal{D}_T$ , processes the encoder blocks with four 3x3 convolution layers followed by ReLU functions, integrating one encoder block after each convolution. Later, two 3x3 convolution layers with a ReLU in-between and a sigmoid function at the end produce $n$ temporal saliency maps. Formally, we write the temporal saliency decoder as + +$$ +\mathcal {D} _ {T} \big (\mathcal {E} (\mathcal {I}) \big) = [ \mathcal {T} _ {n} ] =: \mathcal {T}, \quad n \in \{1, \dots , 5 \} \tag {4} +$$ + +where $\mathcal{T}_n$ denotes the $n^{\mathrm{th}}$ temporal saliency slice. Through this branch of the network, a single image input produces $n$ temporal saliency slices. We use this component to provide temporal predictions to the spatiotemporal mixing module. + +Our image saliency decoder, namely $\mathcal{D}_S$ , has the same structure as $\mathcal{D}_T$ , with the exception of the number of output channels. This component produces a single map $S_I$ , which corresponds to the conventional image saliency map of the input image. As such, we can write + +$$ +\mathcal {S} _ {I} = \mathcal {D} _ {S} (\mathcal {E} (\mathcal {I})). \tag {5} +$$ + +We use this module to provide cumulative saliency information to the spatiotemporal mixing module. + +# 4.2.2 Spatiotemporal Mixing Module + +To incorporate temporal information into the saliency prediction, we introduce a module that combines temporal and + +spatial saliency maps. Our design is inspired by the feature pyramid networks [29], which have the benefit of integrating features from multiple levels of the encoder, thus capturing low-level cues (such as color contrast, brightness, and edges) and high-level ones (such as semantics and scene context). Such cues have been shown to be critical for saliency estimation [15, 24, 38]. This module takes temporal saliency predictions, the initial image saliency prediction, and the encoded image feature blocks as input. We write this as + +$$ +\mathcal {S} _ {R} = \operatorname {S M M} \left(\mathcal {E} (\mathcal {I}), \mathcal {T}, \mathcal {S} _ {I}\right), n \in \{1, \dots , 5 \}, \tag {6} +$$ + +where $S_{R}$ denotes the final, refined image saliency map. + +The module architecture is shown in Figure 6. It takes the last two encoder blocks $[\mathcal{E}_5, \mathcal{E}_4]$ and passes them through a $3 \times 3$ convolution. We then concatenate the other encoder blocks with the image saliency and temporal saliency maps passing through $3 \times 3$ convolution, ReLU, and linear upsampling to keep the spatial dimensions consistent. We add the static and temporal saliency maps to the encoded image features at each block to prevent the information in these maps from vanishing. In the last step, we only add the saliency maps to output a final refined saliency map $S_R$ . The final design of the module (i.e., combining E4 and E5) is based on empirical performance. This module eliminates the need for optimizing a weight parameter between the spatial and temporal maps. It can also modulate the maps within the spatial range of convolutions, which allows the selection of different regions from different maps. + +# 4.2.3 Loss Functions + +To train our network, we use the Kullback-Leibler divergence (KL) [45] and the Correlation Coefficient (CC) + +![](images/f5ed3128e6c7678728e74b86e95a39992ba6aecaf587da806bdcfa0220bbf434.jpg) +Figure 6. The spatiotemporal mixing module combines temporal saliency predictions with the conventional image saliency prediction by multi-level image feature integration. $S_{I}$ denotes the predicted image saliency map, $T_{1,\dots,n}$ the temporal saliency predictions for $n$ frames, and $\mathcal{E}_{1,\dots,5}$ the encoder blocks. This multi-level integration scheme provides information from earlier layers of the network to the next blocks in this module. $S_{R}$ denotes the temporally refined image saliency map output. + +[21] between the predicted and ground truth saliency maps. First, we train the temporal branch using + +$$ +\mathcal {L} _ {1} (\mathcal {I}) = \lambda_ {1} * \operatorname {C C} \left(G T _ {n}, \mathcal {T} _ {n}\right) + \beta_ {1} * \operatorname {K L} \left(G T _ {n}, \mathcal {T} _ {n}\right), \tag {7} +$$ + +where $GT_{n}$ denotes the temporal ground truth for the $n^{th}$ slice. We then freeze the weights in this component and train the spatiotemporal mixing module using + +$$ +\mathcal {L} _ {2} (\mathcal {I}) = \lambda_ {2} * \operatorname {C C} (G T, \mathcal {S} _ {R}) + \beta_ {2} * \operatorname {K L} (G T, \mathcal {S} _ {R}), \tag {8} +$$ + +where $GT$ is the image saliency ground truth for image $I$ . + +# 5. Experiments and Results + +# 5.1. Experimental Setup + +We use a batch size of 32 and an initial learning rate of 1e-4, reduced by a factor of ten every two epochs. We train the temporal branch first and then freeze the weights. We found that 10 epochs of training on SALICON was sufficient. For SALICON, we used the provided test, train, and validation splits. + +# 5.2. Metrics + +We evaluate the obtained saliency predictions according to the following standard metrics used by the community. + +Area Under the Curve (AUC) [5]: Saliency prediction can be interpreted as classifying fixation vs non-fixation points. The area under the ROC curve shows the trade-off between true positives (TP) and false positives (FP). A higher AUC score indicates fewer FPs. While AUC computes the TP and FP rates using all the ground truth fixation points, sAUC [4] samples FP points from ground truth fixations of other observers, compensating for center bias in natural images by taking for inter- and intra-observer variability into account. Normalized Scanpath Saliency (NSS) [38]: This metric + +compares the predicted saliency values at the ground truth fixation points to the average predicted saliency. An NSS score of one indicates that the predicted saliency values at the ground truth fixation points are one standard deviation above the average. + +Kullback - Leibler Divergence (KL) [45]: The KL measures the cumulative distance between the predicted and the ground truth saliency maps. A KL score close to zero indicates a better approximation of the ground truth saliency map by the predicted one. + +Pearson's correlation coefficient (CC) [21]: This metric measures the linear relationship between the predicted and ground truth saliency maps. It ranges from -1 to 1. A CC score close to one indicates a strong linear correlation between the two maps. + +Similarity (SIM) score [22]: The similarity score sums the minimum value between the predicted and the ground truth saliency maps over all pixels. A similarity score of 1 indicates a perfect prediction since both of the maps are probability distributions summing to 1. + +Information Gain (IG) score [27]: The information gain is an information-theoretic metric that measures the difference in average log-likelihood between the predicted saliency map and center-bias prior. + +# 5.3.Quantitative Results + +We compare our method with the state-of-the-art models, namely SAM-Resnet [10], MSI-Net, GazeGAN, MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14], in image saliency prediction. Our model outperforms these methods in five out of seven metrics, showing the benefit of incorporating temporal information. Moreover, our model outperforms the only other multi-duration saliency model [14] in image saliency prediction by a significant margin. Furthermore, we compare our model with this multi-duration model and a multi-duration baseline trained on individual temporal slices. Our model improves the saliency prediction in two durations consisting of 0.5 and 3 seconds in two out of three metrics and in all three metrics in the five-second duration. + +# 5.4. Comparison with state-of-the-art methods + +We first evaluate the performance of our model on image saliency prediction on the SALICON benchmark [20]. The ground truth of SALICON's test set is exclusively hosted on the CodaLab website2. Table 3 shows the comparison of standard evaluation metrics for different state-of-the-art saliency models alongside our model TempSAL. TempSAL outperforms all the baselines in almost all metrics. When it does not, it still yields competitive results. + +
ModelMDAUC ↑CC ↑KL ↓SAUC ↑IG ↑NSS ↑SIM ↑
SAM-Resnet [10]X0.8650.8990.6100.7410.5381.9900.793
MSI-Net [24]X0.8650.8990.3070.7360.7931.9310.784
GazeGAN [6]X0.8640.8790.3760.7360.7201.8990.773
SimpleNet [40]X0.8690.9070.2010.7430.8801.9600.793
MDNSal [40]X0.8650.8990.2210.7360.8631.9350.790
UNISAL [12]X0.8640.8790.3540.7390.7801.9520.775
DeepGaze IIE [31]X0.8690.8720.2850.7670.7661.9960.733
MD-SEM [14]✓0.8640.8680.5680.7460.6602.0580.774
TempSAL✓0.8690.9110.1950.7450.8961.9670.800
+ +Table 3. Evaluation results on the SALICON (LSUN 2017) test benchmark. We compare our model with the state-of-the-art saliency prediction models, namely SAM-Resnet [10], MSI-Net [24], GazeGAN [6], MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14]. The results in bold show the best performance. Our method outperforms the state-of-the-art on conventional image saliency in five metrics. The MD column denotes the ability of the models to predict multi-duration saliency. Our model outperforms the only other multi-duration saliency model by a significant margin on five out of seven metrics. + +
ModelSlice 1 (0-500 ms)Slice 2 (0-3000 ms)Slice 3 (0-5000 ms)Average
CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑
SAM-MD [14]0.8050.3703.1810.7380.4692.5410.7150.5352.4950.7530.4582.739
MD-SEM [14]0.8160.3513.3740.7450.4522.6940.7340.4872.6770.7650.4302.915
TempSAL0.8190.4963.4220.7520.5122.7030.8220.4713.3370.7970.4933.154
+ +# 5.5. Comparison with the multi-duration method + +To compare our method with the only other multiduration model [14], we modify our network to output three temporal slices. We train our network on a three slice SALICON multi-duration dataset first and then fine-tune it on the CodeCharts1k dataset [14] using the given training and validation splits. We report the results of the comparison in Table 4. + +# 5.6. Qualitative Results + +In Figure 7, we compare the temporal and image saliency maps obtained with our method with the ground truth from SALICON [20]. Our model learns time-specific predictions and is able to combine such predictions into a conventional image saliency map. We provide additional qualitative results in the supplementary material. + +# 5.7. Ablation studies + +In this section, we investigate the effect of different components in our model. We also provide a comparison with a multi-duration baseline model on the temporal SALICON dataset. Moreover, we provide an ablation study on the effect of the number of time slices and the two temporal slicing methods in the supplementary material. + +Effect of the SMM module: We evaluate the effect of the spatiotemporal mixing module (SMM) and the image saliency decoder in Table 5. The first model consists of the image encoder and temporal saliency decoder only. We take the average of the temporal slices to measure its performance by comparing it with the ground truth image saliency map. In the second row, we add the image saliency decoder to our model. Similarly, we take the average of the predicted maps $\mathcal{T}_n$ and $S_I$ . Lastly, we add the spatiotemporal mixing module, which effectively modulates these predicted maps and combines them into a final image saliency map $S_R$ . + +Table 4. Results of our model, MD-SEM [14], and the baseline SAM-MD [14] across different durations on the CodeCharts1k dataset [14]. Our model improves the saliency prediction in the first two intervals, consisting of $500~\mathrm{ms}$ and $3000~\mathrm{ms}$ observations, in two out of three metrics. In this comparison, the time slices are cumulative, not mutually exclusive. Although our model benefits from non-overlapping temporal slices, it also performs well with cumulative time slices, particularly in the last slice corresponding to the image saliency with an observation duration of $5000~\mathrm{ms}$ . + +
ModelCC ↑KL ↓NSS ↑SIM ↑
DT(E(I))0.8520.2431.9730.754
+DS(E(I))0.8570.2521.9430.760
+SMM0.9060.1981.9300.798
+ +Table 5. Results of ablation studies on the temporal SALICON validation dataset. The first row denotes the model with only the temporal saliency decoder. In the second row, the model has both the temporal and image saliency decoders. The last row denotes the performance with the spatiotemporal mixing module (SMM). As evidenced by the improved accuracy metrics, the SMM effectively modulates the spatial and temporal saliency maps to refine the initial image saliency prediction. + +![](images/28652b45f66112ae8ea863ee92661c48428b9089e513e89dc83ac92f351ad510.jpg) +Figure 7. The first row shows the input image with the ground truth saliency overlaid. The second row shows the ground truth temporal saliency and image saliency. We show our temporal saliency and image saliency predictions on the third row. Red-yellow maps are temporal saliency maps for one-second intervals. Black and white maps are image saliency maps for the whole observation duration. Our approach captures the attention shifts in sequential temporal maps. Moreover, our model is able to produce accurate image saliency predictions which are close to the ground truth maps. + +Comparison with a temporal baseline model: The performance of our TempSAL model with five temporal slices is provided in Table 6. Note that each saliency slice contains five times fewer samples than the original image saliency map. Therefore, individual slices contain more variation compared to conventional accumulated maps. As a baseline to our model, we compute the performance of an architecture composed of five replicated SimpleNet models (5xSimpleNet) [40], each trained on one saliency slice. This baseline model uses an unshared encoder and decoder for each slice, while we share the decoder among slices. Therefore, we do not benefit from increased model capacity. We observe an accuracy decline in the baseline model, which confirms the increased discrepancy in the data. + +# 6. Conclusion + +We present a saliency prediction method that can learn time-specific predictions and is also able to exploit temporal information to improve overall image saliency prediction. In particular, we show that the temporally evolving patterns in human attention play an important + +
TimeBaselineTempSAL
CC↑KL↓NSS↑SIM↑CC↑KL↓NSS↑SIM↑
T10.8980.2112.4360.7780.8990.2142.4530.782
T20.8700.2192.1590.7650.8770.2152.2110.776
T30.8400.2471.8400.7530.8430.2471.8780.758
T40.8200.2731.7290.7430.8250.2641.7400.749
T50.8110.2751.6460.7380.8130.2761.6540.743
Average0.8480.2451.9620.7560.8520.2431.9870.761
+ +Table 6. Results of the baseline model (left) and our TempSAL model (right) across different time slices. In 18 out of 20 comparisons, our model consistently outperforms the baseline. Note that both models perform best in the first slice, in which the intra-slice agreement is more prominent than in the other slices, as mentioned in Section 3.2. + +role in saliency prediction in natural images. This is evidenced by our experiments that demonstrate our TempSAL method outperforming the state-of-the-art, including a multi-duration method exploiting cumulative temporal saliency maps. + +Acknowledgement. This work was supported in part by the Swiss National Science Foundation via the Sinergia grant CRSII5-180359. + +# References + +[1] Radhakrishna Achanta and Sabine Süsstrunk. Saliency detection for content-aware image resizing. In IEEE International Conference on Image Processing (ICIP), pages 1005-1008, 2009. 1 +[2] Bahar Aydemir, Deblina Bhattacharjee, Seungryong Kim, Tong Zhang, Mathieu Salzmann, and Sabine Süsstrunk. Modeling object dissimilarity for deep saliency prediction. Transactions on Machine Learning Research (TMLR), 2022. 1, 2 +[3] Ali Borji and Laurent Itti. CAT2000: A large scale fixation dataset for boosting saliency research. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2015. 2 +[4] Ali Borji, Dicky N. Sihite, and Laurent Itti. Quantitative analysis of human-model agreement in visual saliency modeling: A comparative study. IEEE Transactions on Image Processing (TIP), 22(1):55-69, 2013. 6 +[5] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Fredo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41(3):740, 2019. 4, 6 +[6] Zhaohui Che, Ali Borji, Guangtao Zhai, Xiongkuo Min, Guodong Guo, and Patrick Le Callet. Gazegan: A generative adversarial saliency model based on invariance analysis of human gaze during scene free viewing. ArXiv, abs/1905.06803, 2019. 7 +[7] Ming-Ming Cheng, Niloy J. Mitra, Xiaolei Huang, Philip H. S. Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 37(3):569-582, 2015. 1, 4 +[8] Hannah Faye Chua, Julie E. Boland, and Richard E. Nisbett. Cultural variation in eye movements during scene perception. Proceedings of the National Academy of Sciences, 102(35):12629-12633, 2005. 3 +[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. A deep multi-level network for saliency prediction. In IEEE International Conference on Pattern Recognition (ICPR), pages 3488-3493, 2016. 1 +[10] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an LSTM-based saliency attentive model. IEEE Transactions on Image Processing (TIP), 27(10):5142-5154, 2018. 6, 7 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2 +[12] Richard Droste, Jianbo Jiao, and J. Alison Noble. Unified Image and Video Saliency Modeling. In European Conference on Computer Vision (ECCV), 2020. 7 +[13] Camilo Fosco, Vincent Casser, Amish Kumar Bedi, Peter O'Donovan, Aaron Hertzmann, and Zoya Bylinskii. Predicting visual importance across graphic design types. In ACM Symposium on User Interface Software and Technology (UIST), page 249-260, 2020. 2, 4 + +[14] Camilo Fosco, Anelise Newman, Pat Sukhum, Yun Bin Zhang, Nanxuan Zhao, Aude Oliva, and Zoya Bylinskii. How much time do you have? modeling multi-duration saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4482, 2020. 1, 2, 3, 6, 7 +[15] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 34(10):1915-1926, 2011. 2, 5 +[16] Joseph H. Goldberg and Jonathan I. Helfman. Visual scanpath representation. In Symposium on Eye-Tracking Research & Applications (ETRA), pages 203–210, 2010. 2 +[17] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. SALICON: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In IEEE International Conference on Computer Vision (ICCV), pages 262-270, 2015. 1, 2 +[18] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 20(11):1254-1259, 1998. 1, 2, 4 +[19] Sen Jia and Neil D. B. Bruce. EML-NET: An expandable Multi-Layer NETwork for saliency prediction. Image and Vision Computing, 95:103887, 2020. 2 +[20] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. SALICON: Saliency in context. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 2, 3, 6, 7 +[21] Timothee Jost, Nabil Ouerhani, Roman von Wartburg, René Muri, and Heinz Hügli. Assessing the contribution of color in visual attention. Computer Vision and Image Understanding, 100(1-2):107-123, 2005. 6 +[22] Tilke Judd, Fredo Durand, and Antonio Torralba. A benchmark of computational models of saliency to predict human fixations. MIT Technical Report, 2012. 6 +[23] Tilke Judd, Krista Ehinger, Fredo Durand, and Antonio Torralba. Learning to predict where humans look. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2009. 2 +[24] Alexander Kroner, Mario Senden, Kurt Driessens, and Rainer Goebel. Contextual encoder-decoder network for visual saliency prediction. Neural Networks, 129:261 - 270, 2020. 2, 5, 7 +[25] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep Gaze I: Boosting saliency prediction with feature maps trained on ImageNet. In International Conference on Learning Representations (ICLR) Workshops, 2015. 2 +[26] Matthias Kümmerer, Tom Wallis, and Matthias Bethge. DeepGaze II: Predicting fixations from deep features over time and tasks. Journal of Vision (JOV), 17(10):1147, 2017. 1, 2, 4 +[27] Matthias Kümmerer, Thomas S. A. Wallis, and Matthias Bethge. Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences (PNAS), 112(52):16054-16059, 2015. 6 + +[28] Qingshan Li, Yue Zhou, and Jie Yang. Saliency based image segmentation. In IEEE International Conference on Multimedia Technology, pages 5068-5071, 2011. 1 +[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 5 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 3 +[31] Akis Linardos, Matthias Kümmerer, Ori Press, and Matthias Bethge. Deepgaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12919-12928, 2021. 1, 4, 6, 7 +[32] Gitte Lindgaard, Gary Fernandes, Cathy Dudek, and J. Brown. Attention web designers: You have 50 milliseconds to make a good first impression! Behaviour & Information Technology, 25(2):115-126, 2006. 2, 4 +[33] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In European Conference on Computer Vision (ECCV), pages 19-34, 2018. 4 +[34] Nian Liu and Junwei Han. A deep spatial contextual long-term recurrent convolutional network for saliency detection. IEEE Transactions on Image Processing (TIP), 27(7):3264-3274, 2018. 1 +[35] Sandeep Mishra and Oindrila Saha. RecSal : Deep recursive supervision for visual saliency prediction. In British Machine Vision Conference (BMVC), 2020. 2 +[36] Johanna Palcu, Jennifer Sudkamp, and Arnd Florack. Judgments at gaze value: Gaze cuing in banner advertisements, its effect on attention allocation and product judgments. Frontiers in Psychology, 8, 2017. 2, 4 +[37] Yash Patel, Srikar Appalaraju, and R. Manmatha. Saliency driven perceptual image compression. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 227-236, 2021. 1 +[38] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397-2416, 2005. 5, 6 +[39] Michael Posner, Robert Rafal, Lisa Choatec, and Jonathan Vaughan. Inhibition of return: Neural basis and function. Cognitive Neuropsychology, Vol. 2:211 - 228, 09 1985. 1 +[40] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In International Conference on Intelligent Robots and Systems (IROS), 2020. 4, 6, 7, 8 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015. 4 + +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations (ICLR), 2015. 2 +[43] Markus Steinberger, Bernhard Kainz, Stefan Hauswiesner, Rostislav Khlebnikov, Denis Kalkofen, and Dieter Schmalstieg. Ray prioritization using stylization and visual saliency. Computers & Graphics, 36(6):673-684, 2012. 1 +[44] Po-He Tseng, Ran Carmi, Ian G. M. Cameron, Douglas P. Munoz, and Laurent Itti. Quantifying center bias of observers in free viewing of dynamic natural scenes. Journal of Vision (JOV), 9(7):4-4, 07 2009. 3 +[45] Mathukumalli Vidyasagar. Kullback-leibler divergence rate between probability distributions on sets of different cardinalities. In IEEE Conference on Decision and Control (CDC), pages 948-953, 2010. 5, 6 +[46] Eleonora Vig, Michael Dorr, and David Cox. Large-scale optimization of hierarchical features for saliency prediction in natural images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2798-2805, 2014. 1, 2 +[47] W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, and R. Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(06):3239-3259, 2022. 1 +[48] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. IEEE Transactions on Multimedia, 22(8):2163-2176, 2020. 1 +[49] Alfred L. Yarbus. Eye Movements and Vision, volume 2. Plenum Press, 1967. 1, 3 +[50] Yifeng Zhang, Ming Jiang, and Qi Zhao. Saliency prediction with external knowledge. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 484-493, 2021. 1, 2 +[51] Jufeng Zhao, Yueting Chen, Huajun Feng, Zhihai Xu, and Qi Li. Fast image enhancement using multi-scale saliency extraction in infrared imagery. Optik, 125(15):4039-4042, 2014. 1 \ No newline at end of file diff --git a/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/images.zip b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d00dea9778e93fc64fc20a0aa0cc17650df6027a --- /dev/null +++ b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3e2e321f80392393853e8ebecbd390880010daa9f7b19b04b5a1ff4225ffc08 +size 521025 diff --git a/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/layout.json b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eeadb00154540c78327cb18d9ee52d3483ec50ef --- /dev/null +++ b/2023/TempSAL - Uncovering Temporal Information for Deep Saliency Prediction/layout.json @@ -0,0 +1,9299 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "text", + "content": "TempSAL - Uncovering Temporal Information for Deep Saliency Prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 143, + 501, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 143, + 501, + 171 + ], + "spans": [ + { + "bbox": [ + 91, + 143, + 501, + 171 + ], + "type": "text", + "content": "Bahar Aydemir, Ludo Hoffstetter, Tong Zhang, Mathieu Salzmann, Sabine Süssstrunk \nSchool of Computer and Communication Sciences, EPFL, Switzerland" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 174, + 484, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 174, + 484, + 185 + ], + "spans": [ + { + "bbox": [ + 100, + 174, + 484, + 185 + ], + "type": "text", + "content": "{bahir/aydemir, tong.zhang, mathieu.salzmann, sabine.susstrunk}@epfl.ch" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 214, + 168, + 304 + ], + "blocks": [ + { + "bbox": [ + 50, + 214, + 168, + 304 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 168, + 304 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 168, + 304 + ], + "type": "image", + "image_path": "307199b203a9151932e8b29750a4934c05f1f6976b3e556d1ac0d7db25e64408.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 305, + 132, + 315 + ], + "lines": [ + { + "bbox": [ + 85, + 305, + 132, + 315 + ], + "spans": [ + { + "bbox": [ + 85, + 305, + 132, + 315 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 173, + 213, + 229, + 258 + ], + "blocks": [ + { + "bbox": [ + 173, + 213, + 229, + 258 + ], + "lines": [ + { + "bbox": [ + 173, + 213, + 229, + 258 + ], + "spans": [ + { + "bbox": [ + 173, + 213, + 229, + 258 + ], + "type": "image", + "image_path": "e15e40f345eb4b7fad635135f663abdfdbc88a11d76ea3ed497b8de67ec5b896.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 174, + 258, + 229, + 303 + ], + "blocks": [ + { + "bbox": [ + 174, + 258, + 229, + 303 + ], + "lines": [ + { + "bbox": [ + 174, + 258, + 229, + 303 + ], + "spans": [ + { + "bbox": [ + 174, + 258, + 229, + 303 + ], + "type": "image", + "image_path": "760dc8cce4cc441bcc3251a3efe1f44fa0df592a9af200b8c3ff8d8a505de97c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 304, + 226, + 314 + ], + "lines": [ + { + "bbox": [ + 178, + 304, + 226, + 314 + ], + "spans": [ + { + "bbox": [ + 178, + 304, + 226, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1 = 0 - 1s" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 230, + 213, + 287, + 258 + ], + "blocks": [ + { + "bbox": [ + 230, + 213, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 230, + 213, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 230, + 213, + 287, + 258 + ], + "type": "image", + "image_path": "322905c8f2176207b7b1e982390b7749d7d3c7950cd96ce07140a8975b26d2f3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 230, + 258, + 287, + 304 + ], + "blocks": [ + { + "bbox": [ + 230, + 258, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 230, + 258, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 230, + 258, + 287, + 304 + ], + "type": "image", + "image_path": "a574130edd1b16f31495a2035538b3ecbe3b4119b7f0c9ad428335419109cea9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 304, + 285, + 314 + ], + "lines": [ + { + "bbox": [ + 236, + 304, + 285, + 314 + ], + "spans": [ + { + "bbox": [ + 236, + 304, + 285, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2 = 1 - 2s" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "text", + "content": "Figure 1. An example of how human attention evolves over time when observing a single image. Top row: Temporal and image saliency ground truth from the SALICON dataset [20]. Bottom row: Our temporal and image saliency predictions. Each temporal saliency map " + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, 5\\}" + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "text", + "content": " represents one second of observation time. Note that in " + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "text", + "content": ", the chef is salient, while in " + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2" + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_3" + }, + { + "bbox": [ + 46, + 325, + 545, + 380 + ], + "type": "text", + "content": ", the food on the barbecue becomes the most salient region in this scene. We can predict the temporal saliency maps for each interval separately, or combine them to create a single, refined image saliency map for the entire observation period." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 288, + 213, + 347, + 258 + ], + "blocks": [ + { + "bbox": [ + 288, + 213, + 347, + 258 + ], + "lines": [ + { + "bbox": [ + 288, + 213, + 347, + 258 + ], + "spans": [ + { + "bbox": [ + 288, + 213, + 347, + 258 + ], + "type": "image", + "image_path": "a2c271bcedad4b8e69371aaaac4523ff87dcf3d9b534b545c989479dfe12f5d8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 288, + 258, + 347, + 304 + ], + "blocks": [ + { + "bbox": [ + 288, + 258, + 347, + 304 + ], + "lines": [ + { + "bbox": [ + 288, + 258, + 347, + 304 + ], + "spans": [ + { + "bbox": [ + 288, + 258, + 347, + 304 + ], + "type": "image", + "image_path": "7f77badc9d6fc58a890254a7584666f696dd04e65abce8b360651d5f929dfb65.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 304, + 344, + 314 + ], + "lines": [ + { + "bbox": [ + 296, + 304, + 344, + 314 + ], + "spans": [ + { + "bbox": [ + 296, + 304, + 344, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_3 = 2 - 3s" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 348, + 213, + 407, + 258 + ], + "blocks": [ + { + "bbox": [ + 348, + 213, + 407, + 258 + ], + "lines": [ + { + "bbox": [ + 348, + 213, + 407, + 258 + ], + "spans": [ + { + "bbox": [ + 348, + 213, + 407, + 258 + ], + "type": "image", + "image_path": "55ab983def07961ecc49365fe9b8a5a98d1054c71fcc284cdfeca0438010f8f1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 348, + 258, + 407, + 304 + ], + "blocks": [ + { + "bbox": [ + 348, + 258, + 407, + 304 + ], + "lines": [ + { + "bbox": [ + 348, + 258, + 407, + 304 + ], + "spans": [ + { + "bbox": [ + 348, + 258, + 407, + 304 + ], + "type": "image", + "image_path": "d05853d6cade2062efa0a306798bc71c23afbc0eee60418eaf2dedd517b692f6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 304, + 403, + 314 + ], + "lines": [ + { + "bbox": [ + 354, + 304, + 403, + 314 + ], + "spans": [ + { + "bbox": [ + 354, + 304, + 403, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_4 = 3 - 4s" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 408, + 213, + 468, + 258 + ], + "blocks": [ + { + "bbox": [ + 408, + 213, + 468, + 258 + ], + "lines": [ + { + "bbox": [ + 408, + 213, + 468, + 258 + ], + "spans": [ + { + "bbox": [ + 408, + 213, + 468, + 258 + ], + "type": "image", + "image_path": "19db42a72cd02b4933528fca4d9cb92cedee2fee606ad0ceeac9de5851ff626f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 409, + 258, + 468, + 304 + ], + "blocks": [ + { + "bbox": [ + 409, + 258, + 468, + 304 + ], + "lines": [ + { + "bbox": [ + 409, + 258, + 468, + 304 + ], + "spans": [ + { + "bbox": [ + 409, + 258, + 468, + 304 + ], + "type": "image", + "image_path": "e868de0d03bd3fa78ce3014e1244435e5fb70ded7272c6835c8a66c6dde1205f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 304, + 463, + 314 + ], + "lines": [ + { + "bbox": [ + 414, + 304, + 463, + 314 + ], + "spans": [ + { + "bbox": [ + 414, + 304, + 463, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{5} = 4 - 5s" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 480, + 213, + 539, + 258 + ], + "blocks": [ + { + "bbox": [ + 480, + 213, + 539, + 258 + ], + "lines": [ + { + "bbox": [ + 480, + 213, + 539, + 258 + ], + "spans": [ + { + "bbox": [ + 480, + 213, + 539, + 258 + ], + "type": "image", + "image_path": "26934dac085f6443b31ee8d659c89b76270b622bf4ac1875d24c0aeb8276bcce.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 481, + 258, + 539, + 303 + ], + "blocks": [ + { + "bbox": [ + 481, + 258, + 539, + 303 + ], + "lines": [ + { + "bbox": [ + 481, + 258, + 539, + 303 + ], + "spans": [ + { + "bbox": [ + 481, + 258, + 539, + 303 + ], + "type": "image", + "image_path": "4eff5060045e80a783b3c018b1891ef6b211dea6e8315fb2ee105f12e4060b2b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 304, + 524, + 314 + ], + "lines": [ + { + "bbox": [ + 496, + 304, + 524, + 314 + ], + "spans": [ + { + "bbox": [ + 496, + 304, + 524, + 314 + ], + "type": "text", + "content": "0-5s" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 143, + 396, + 192, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 396, + 192, + 408 + ], + "spans": [ + { + "bbox": [ + 143, + 396, + 192, + 408 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 45, + 422, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 422, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 45, + 422, + 289, + 590 + ], + "type": "text", + "content": "Deep saliency prediction algorithms complement the object recognition features, they typically rely on additional information such as scene context, semantic relationships, gaze direction, and object dissimilarity. However, none of these models consider the temporal nature of gaze shifts during image observation. We introduce a novel saliency prediction model that learns to output saliency maps in sequential time intervals by exploiting human temporal attention patterns. Our approach locally modulates the saliency predictions by combining the learned temporal maps. Our experiments show that our method outperforms the state-of-the-art models, including a multi-duration saliency model, on the SALICON benchmark and CodeCharts1k dataset. Our code is publicly available on GitHub1." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 613, + 127, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 127, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 127, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 634, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 287, + 694 + ], + "type": "text", + "content": "Humans have developed attention mechanisms that allow them to selectively focus on the important parts of a scene. Saliency prediction algorithms aim to computationally detect these regions that stand out relative to their surroundings. These predictions have numerous applications" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 397, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 397, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 397, + 545, + 422 + ], + "type": "text", + "content": "in image compression [37], image enhancement [51], image retargeting [1], rendering [43], and segmentation [28]." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 423, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 628 + ], + "type": "text", + "content": "Since the seminal work of Itti et al. [18], many have developed solutions using both handcrafted features [7] and deep ones [9, 17, 26, 34, 46, 48]. Nowadays, employing deep neural networks is preferred in saliency prediction as they outperform bottom-up models. These methods typically depend on pre-trained object recognition networks to extract features from the input image [31]. In addition to these features, scene context [47], object co-occurrence [50], and dissimilarity [2] have been exploited to improve the saliency prediction. However, while these approaches model the scene context and objects, they fail to consider that humans dynamically observe scenes [49]. In neuroscience, the inhibition of return paradigm states that a suppression mechanism reduces visual attention towards recently attended objects [39] and encourages selective attention to novel regions. Motivated by this principle, we develop a saliency prediction model that incorporates temporal information." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "Fosco et al. [14] also exploit temporal information in saliency prediction, but they consider snapshots containing observations up to 0.5, 3, and 5 seconds, thus not leveraging saliency trajectory but rather saliency accumulation. By contrast, here, we model consecutive time slices, connecting our approach more directly with the human gaze and thus opening the door to automated visual appeal assess-" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 703, + 212, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 212, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 212, + 714 + ], + "type": "text", + "content": "1https://ivrl.github.io/Tempsal/" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6461" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "type": "text", + "content": "ment in applications such as website design [32], advertisement [36] and infographics [13]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 179 + ], + "type": "text", + "content": "To achieve this, we show that when viewing images, human attention yields temporally evolving patterns, and we introduce a network capable of exploiting this temporal information for saliency prediction. Specifically, our model learns time-specific predictions and is able to combine them with a conventional image saliency map to obtain a temporally modulated image saliency prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 323 + ], + "type": "text", + "content": "We evaluate our method on the SALICON [20] and CodeCharts1k [14] datasets which contain temporal information, unlike the other popular saliency datasets such as CAT2000 [3] and MIT1003 [23]. By showing the benefits of estimating temporal saliency we hope to encourage the community to publish the temporal information of their data along with the final saliency maps. Note that existing works typically collect saliency data by conducting psychophysical experiments [3, 14, 20, 23], and the attention data recorded during these experiments already includes temporal information. Therefore, no further experiments are required." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 286, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 286, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 286, + 394 + ], + "type": "text", + "content": "As evidenced by our experiments, using temporal information boosts the accuracy of the baseline network, enabling us to consistently outperform the state-of-the-art models on the SALICON saliency benchmark. Moreover, in the CodeCharts1k dataset we outperform the multiduration model [14] in two out of three metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 396, + 237, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 396, + 237, + 406 + ], + "spans": [ + { + "bbox": [ + 59, + 396, + 237, + 406 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 415, + 286, + 616 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 59, + 415, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 415, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 59, + 415, + 286, + 437 + ], + "type": "text", + "content": "- We evidence the presence of temporally evolving patterns in human attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 446, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 446, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 59, + 446, + 286, + 493 + ], + "type": "text", + "content": "- We show that temporal information in the form of a saliency trajectory improves saliency prediction in natural images, providing an investigation of the SALICON dataset for temporal attention shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 502, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 502, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 59, + 502, + 286, + 548 + ], + "type": "text", + "content": "- We introduce a novel, saliency prediction model, called TempSAL, capable of simultaneously predicting conventional image saliency and temporal saliency trajectories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 557, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 557, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 59, + 557, + 286, + 616 + ], + "type": "text", + "content": "- We propose a spatiotemporal mixing module that learns time-dependent patterns from temporal saliency maps. Our approach outperforms the state-of-the-art image saliency models that either do not consider temporal information or encode it in a cumulative manner." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 628, + 130, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 130, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 130, + 639 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 248, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 248, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 248, + 659 + ], + "type": "text", + "content": "2.1. Saliency prediction for natural images" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Early saliency prediction methods were biologically inspired and bottom-up. In particular, Itti et al. used color, intensity, and orientation contrast [18]. Goferman employed global and local contrast as contextual cues [15]. Judd et" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "al. [23] further incorporated mid-level and high-level semantic features, using horizon, face, person, and car detectors. Later, Vig et al. [46] showed that deep neural networks can be applied to saliency prediction. Yet, saliency prediction lacks the large-scale annotated datasets that are available for image classification tasks [11], which prevents training robust models. To overcome this, Kummerer et al. [25] showed that using pre-trained object recognition networks significantly improves saliency predictions. Subsequent state-of-the-art models such as EML-Net [19], DeepGaze2 [26], and SALICON [17] similarly use pretrained convolutional neural network (VGG [42]) encoders. Recent works utilize additional sources of information such as scene context [24], external knowledge [50] and object dissimilarity [2] to improve saliency prediction. Yet, none of these methods take into account the temporal evolution of the human gaze, which occurs even when the image stimuli are static. In our work, we make use of these temporal patterns as an additional source of information to boost conventional image saliency prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 320, + 440, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 440, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 440, + 333 + ], + "type": "text", + "content": "2.2. Multi-duration saliency" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 339, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 626 + ], + "type": "text", + "content": "Existing image saliency ground truth maps include all fixations made throughout the observation period. Aggregating all these fixations that have different timestamps into a single ground truth map results in the loss of temporal information. Representing the fixations as scanpaths retains the temporal clues by encoding the change of gaze of an individual over time. However, merging numerous scanpaths is challenging [16]. Fosco et al. [14] proposed multiduration saliency to characterize the attention of a group of individuals while taking into account time-dependent attention shifts. The temporal maps they rely on, however, encode the attention distribution of many observers across overlapping time periods of increasing durations. While this is a convenient way of capturing a population's attention patterns, it does not reflect the saliency trajectory over time. Similarly, [35] use order of fixations as sequential metadata for deep supervision but they do not model the evolution of attention through time. In our work, we model multiduration saliency to analyze underlying attention patterns by using mutually exclusive time slices. We provide this temporal information to our spatiotemporal mixing module to refine the initial image saliency prediction with temporal information. Moreover, this lets us predict temporal saliency maps for each second of attention." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 639, + 484, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 484, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 484, + 651 + ], + "type": "text", + "content": "3. Temporal saliency data analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 658, + 368, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 368, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 368, + 670 + ], + "type": "text", + "content": "3.1. Datasets" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "SALICON [20] is the largest human attention dataset on natural images. It was created via a crowdsourced mouse tracking experiment, which was shown to be similar to eye" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6462" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "content": "tracking [20] and widely used in the saliency prediction literature. SALICON consists of 10000 training, 5000 validation and 5000 test images from the MS-COCO dataset [30]. The SALICON dataset provides saliency maps, fixations, and gaze points for each image and observer. A gaze point is a raw data point recorded by a tracking device. It describes the spatial coordinates of the eye/mouse on the associated stimuli at a given timestamp. Conversely, fixations describe the coordinates of the long pause when the eyes are fixated on an image detail. Following common practice in eye tracking experiments, Jiang et al. [20] grouped spatially and temporally close gaze points to create fixations. Since the fixations were created by grouping multiple gaze points, they do not have an associated timestamp. To address this, SALICON-MD [14] assumes that the fixations are uniformly distributed across the total viewing time. We provide a finer approximation for recovering the fixations' timestamps, by minimizing the spatial and temporal distance between a fixation and the nearest gaze point. We refer the reader to the supplementary material for the details of this approximation process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 323, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 288, + 396 + ], + "type": "text", + "content": "We also assess the performance and generalization ability of our method using CodeCharts1k [14], which is the first saliency dataset to report multiple viewing durations. The dataset consists of 1000 images and saliency maps collected via crowdsourcing which correspond to 0.5, 3, and 5 seconds of viewing. We present the results in Section 5.5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 403, + 223, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 223, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 223, + 416 + ], + "type": "text", + "content": "3.2. Temporal patterns in the dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 422, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 506 + ], + "type": "text", + "content": "In this section, we examine how temporality evolves during human visual attention. To observe the evolution of attention over time, we slice the data into five slices, one for each second of observation. In particular, we inspect the dissimilarity between slices, the agreement with the average, and the distribution of fixations in the time-saliency space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 624 + ], + "type": "text", + "content": "Average maps. Viewing patterns in image saliency experiments tend to show a concentration towards the image center [44], which is known as the photographer's bias or center bias. We observe a similar spatial bias for each temporal slice, shown in Figure 2, where we plot the average heat maps for each temporal slice. Note that the gaze tends to converge to the center of the image as time passes. This means that the observers revisit the previously seen important center regions [49]." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 628, + 97, + 665 + ], + "blocks": [ + { + "bbox": [ + 50, + 628, + 97, + 665 + ], + "lines": [ + { + "bbox": [ + 50, + 628, + 97, + 665 + ], + "spans": [ + { + "bbox": [ + 50, + 628, + 97, + 665 + ], + "type": "image", + "image_path": "ffe0ee55273b1b5895d1b17fe04b06004642875b8d8a7bc7f10cc4297d21eb86.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 665, + 94, + 673 + ], + "lines": [ + { + "bbox": [ + 55, + 665, + 94, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 94, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1 = 0 - 1s" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 677, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 710 + ], + "type": "text", + "content": "Figure 2. Average heat maps for each one-second interval. Note that a center bias occurs, similar to image saliency prediction's average ground truth maps." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 97, + 628, + 143, + 664 + ], + "blocks": [ + { + "bbox": [ + 97, + 628, + 143, + 664 + ], + "lines": [ + { + "bbox": [ + 97, + 628, + 143, + 664 + ], + "spans": [ + { + "bbox": [ + 97, + 628, + 143, + 664 + ], + "type": "image", + "image_path": "cdbc9e400162f880d3485a1fd4bd135e827c763a7689b971c09434206e2b0767.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 665, + 140, + 673 + ], + "lines": [ + { + "bbox": [ + 99, + 665, + 140, + 673 + ], + "spans": [ + { + "bbox": [ + 99, + 665, + 140, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2 = 1 - 2s" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 144, + 628, + 190, + 664 + ], + "blocks": [ + { + "bbox": [ + 144, + 628, + 190, + 664 + ], + "lines": [ + { + "bbox": [ + 144, + 628, + 190, + 664 + ], + "spans": [ + { + "bbox": [ + 144, + 628, + 190, + 664 + ], + "type": "image", + "image_path": "762975aa094de7f177e9b1017bc5539b11eebe7adb60d6b9ec970289e87a2221.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 665, + 186, + 673 + ], + "lines": [ + { + "bbox": [ + 146, + 665, + 186, + 673 + ], + "spans": [ + { + "bbox": [ + 146, + 665, + 186, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_3 = 2 - 3s" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 191, + 628, + 237, + 664 + ], + "blocks": [ + { + "bbox": [ + 191, + 628, + 237, + 664 + ], + "lines": [ + { + "bbox": [ + 191, + 628, + 237, + 664 + ], + "spans": [ + { + "bbox": [ + 191, + 628, + 237, + 664 + ], + "type": "image", + "image_path": "4a217d9d80a4e7dcbd0a9fe997d92e831bf068e747e38f212c188ea185301e7a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 665, + 233, + 673 + ], + "lines": [ + { + "bbox": [ + 193, + 665, + 233, + 673 + ], + "spans": [ + { + "bbox": [ + 193, + 665, + 233, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_4 = 3 - 4s" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 238, + 628, + 284, + 664 + ], + "blocks": [ + { + "bbox": [ + 238, + 628, + 284, + 664 + ], + "lines": [ + { + "bbox": [ + 238, + 628, + 284, + 664 + ], + "spans": [ + { + "bbox": [ + 238, + 628, + 284, + 664 + ], + "type": "image", + "image_path": "5db29a0a754eeedf751a251b47e9bc98f09ff3f614646f3e11f789fec6159343.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 665, + 279, + 673 + ], + "lines": [ + { + "bbox": [ + 241, + 665, + 279, + 673 + ], + "spans": [ + { + "bbox": [ + 241, + 665, + 279, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{5} = 4 - 5s" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 307, + 69, + 535, + 160 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 535, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 535, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 535, + 160 + ], + "type": "table", + "html": "
T1T2T3T4T5
T11.000.700.540.500.54
T20.701.000.730.660.65
T30.540.731.000.750.70
T40.500.660.751.000.73
T50.540.650.700.731.00
Average0.660.750.740.730.72
", + "image_path": "145916055125368e337b23624ad2948a8c83f85b16b77eca11585c5f66b8b4d1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 169, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 547, + 225 + ], + "type": "text", + "content": "Table 1. Correlation scores (CC) of the temporal slices with each other in a single image, averaged over all images. All slices show more similarity to their direct temporal neighbors. The last row shows the average similarity of a slice with the other slices, " + }, + { + "bbox": [ + 304, + 169, + 547, + 225 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 304, + 169, + 547, + 225 + ], + "type": "text", + "content": " being the most dissimilar one." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": "We plot the differences of the consecutive average temporal slices in Figure 3 to illustrate attention shifts. Light blue indicates the regions with reduced attention, whereas (light) red indicates increased attention. We observe that attention shifts from left to right, with a subsequent dispersion from the center towards the corners. Then, attention increases at the center of the image, slightly skewed to the left. Interestingly, the trend (especially in " + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2 - \\mathcal{T}_1" + }, + { + "bbox": [ + 304, + 236, + 545, + 344 + ], + "type": "text", + "content": ") coincides with the western left-to-right reading direction [8]." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 308, + 356, + 365, + 399 + ], + "blocks": [ + { + "bbox": [ + 308, + 356, + 365, + 399 + ], + "lines": [ + { + "bbox": [ + 308, + 356, + 365, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 356, + 365, + 399 + ], + "type": "image", + "image_path": "b8fd29b00402c164408e1dba5f14ef592c81129913940cb05b6a5f0eb43daa82.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 400, + 358, + 411 + ], + "lines": [ + { + "bbox": [ + 318, + 400, + 358, + 411 + ], + "spans": [ + { + "bbox": [ + 318, + 400, + 358, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2 - \\mathcal{T}_1" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 368, + 357, + 425, + 399 + ], + "blocks": [ + { + "bbox": [ + 368, + 357, + 425, + 399 + ], + "lines": [ + { + "bbox": [ + 368, + 357, + 425, + 399 + ], + "spans": [ + { + "bbox": [ + 368, + 357, + 425, + 399 + ], + "type": "image", + "image_path": "24396633b3c0ea0ddd207908eb52b265e40709d2293359c1a7128cd687539492.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 400, + 414, + 411 + ], + "lines": [ + { + "bbox": [ + 374, + 400, + 414, + 411 + ], + "spans": [ + { + "bbox": [ + 374, + 400, + 414, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_3 - \\mathcal{T}_2" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 422, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 422, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 547, + 455 + ], + "type": "text", + "content": "Figure 3. Differences of the consecutive average temporal slices shown in Fig. 2. Red indicates regions of increased attention whereas blue indicates decreased attention." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 427, + 357, + 485, + 399 + ], + "blocks": [ + { + "bbox": [ + 427, + 357, + 485, + 399 + ], + "lines": [ + { + "bbox": [ + 427, + 357, + 485, + 399 + ], + "spans": [ + { + "bbox": [ + 427, + 357, + 485, + 399 + ], + "type": "image", + "image_path": "143c619e6ee97624f626f8c2ff6cbbebd8dbb400af7b5cd29361428c02205bfe.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 400, + 474, + 411 + ], + "lines": [ + { + "bbox": [ + 435, + 400, + 474, + 411 + ], + "spans": [ + { + "bbox": [ + 435, + 400, + 474, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_4 - \\mathcal{T}_3" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 488, + 357, + 544, + 399 + ], + "blocks": [ + { + "bbox": [ + 488, + 357, + 544, + 399 + ], + "lines": [ + { + "bbox": [ + 488, + 357, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 488, + 357, + 544, + 399 + ], + "type": "image", + "image_path": "f8c35bff611d8b8e0d6037d7d107a197151daecc13b054a5897428ac6b90f9d6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 400, + 535, + 411 + ], + "lines": [ + { + "bbox": [ + 495, + 400, + 535, + 411 + ], + "spans": [ + { + "bbox": [ + 495, + 400, + 535, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{5} - \\mathcal{T}_{4}" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "type": "text", + "content": "Inter-slice similarity across time. We expect temporal saliency slices to be more similar to their closer-in-time slices than to the ones further away since human attention is continuous over time. Table 1 contains the correlation coefficients between each pair of saliency slices in a single image, averaged over all images. We calculate the correlation coefficient between slices " + }, + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_j" + }, + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_k" + }, + { + "bbox": [ + 304, + 468, + 545, + 554 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 310, + 563, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 563, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 310, + 563, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\operatorname {C C} \\left(\\mathcal {T} _ {j}, \\mathcal {T} _ {k}\\right) = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\operatorname {C C} \\left(\\mathcal {T} _ {i j}, \\mathcal {T} _ {i k}\\right), \\quad j, k \\in \\{1, \\dots , 5 \\}, \\tag {1}", + "image_path": "9e0ac634f09cd6696546d49a7deeca8916f8e730a6f392294e76dc97ac60d762.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": " is the total number of images, and " + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{ij}" + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{ik}" + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": " slice of the " + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 605, + 544, + 630 + ], + "type": "text", + "content": " image." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "content": "By calculating t-test scores on the pairwise comparisons, we observe that all of the pairwise differences except " + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_3" + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_5" + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "content": " are statistically significant (" + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "inline_equation", + "content": "p < 0.01" + }, + { + "bbox": [ + 304, + 630, + 545, + 702 + ], + "type": "text", + "content": "). Thus, the attention residuals between different time intervals in one image are significantly different. We provide more details in the supplementary material." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6463" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 69, + 269, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 69, + 269, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 269, + 99 + ], + "type": "interline_equation", + "content": "\\begin{array}{c c c c c c} & \\mathcal {T} _ {1} & \\mathcal {T} _ {2} & \\mathcal {T} _ {3} & \\mathcal {T} _ {4} & \\mathcal {T} _ {5} \\\\ \\text {C C} & \\hline 0. 5 7 4 & \\hline 0. 4 3 3 & \\hline 0. 4 3 1 & \\hline 0. 4 2 6 & \\hline 0. 4 4 7 \\end{array}", + "image_path": "793e6839f0eb4cb4ea8afb26a928875df2716edb4e3e40daf658bae9a5ad042a.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 153 + ], + "type": "text", + "content": "Table 2. Correlation scores (CC) of each time slice in a single image with the average maps presented in Figure 2, averaged over all images. The similarity between slices across images decreases with time, with the exception of the last slice." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 239 + ], + "type": "text", + "content": "Intra-slice similarity across images. We also investigate the deviation of each slice from its respective average slices. The average slices are depicted in Figure 2. Table 2 shows the deviation of a slice from the average time slices per image. We compute CC scores between a single slice and the corresponding average slice as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 248, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 248, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\operatorname {C C} \\left(\\mathcal {T} _ {j}, A _ {j}\\right) = \\frac {1}{N} \\sum_ {i = 0} ^ {N} \\operatorname {C C} \\left(\\mathcal {T} _ {i j}, A _ {j}\\right), \\quad j \\in \\{1, \\dots , 5 \\}, \\tag {2}", + "image_path": "c5e8143e821970d91b8dbd2a75bb56bc1fc11340a089bc25f4687dd8c06953bf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 47, + 291, + 208, + 304 + ], + "type": "text", + "content": " average slice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "content": "We average the scores across all images. Higher values of CC indicate more agreement with the average whereas lower values of CC indicate more deviation from the average. Note that the similarity with the average across images decreases with time, except for the last slice. This can be explained by the more prominent center bias in " + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_5" + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "content": " as seen in Figure 2. " + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "content": " has the least deviation from the average by a significant margin (" + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "inline_equation", + "content": "p \\ll 0.01" + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "content": "). This shows that humans tend to look at similar places at first, and then their attention scatters around to less important regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 436, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 591 + ], + "type": "text", + "content": "Early and late fixations versus saliency. Lastly, we investigate the relationship between the fixation timestamps and their respective saliency values. We assign a saliency value to each fixation as the normalized pixel value in the corresponding saliency map. We plot the histogram of the number of fixations with their saliency and timestamp values in Figure 4. The fixation time stamps range from 0 to " + }, + { + "bbox": [ + 46, + 436, + 287, + 591 + ], + "type": "inline_equation", + "content": "5000\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 436, + 287, + 591 + ], + "type": "text", + "content": " and the saliency values range from 0 to 1. Late fixations tend to have lower saliency values than earlier fixations, as indicated by the darker color towards the bottom right corner. That is, the first region we glance at in an image is more important (salient) than the following regions [7, 18]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 603, + 129, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 129, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 129, + 616 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 622, + 142, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 142, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 142, + 635 + ], + "type": "text", + "content": "4.1. Temporal slices" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "We aim to recover fixation timestamps to train models with this temporal information. We extract temporal slices by grouping the fixations in several time intervals and, following common practice [5], blurring with a Gaussian kernel. We break down the fixations into time slices with two time-slicing (grouping) alternatives, namely equal duration" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 548, + 228 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 548, + 228 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 548, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 548, + 228 + ], + "type": "image", + "image_path": "38c5d35b9376831b01f7c2dbd658ed646856b95d9e2fb3b228b4a79a0ab66c94.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 239, + 547, + 306 + ], + "lines": [ + { + "bbox": [ + 305, + 239, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 547, + 306 + ], + "type": "text", + "content": "Figure 4. Number of fixations with their respective saliency values and timestamps. Lighter colors indicate a higher number of occurrences while darker areas denote fewer occurrences. We see that late fixations tend to be less salient, which can be seen as the decrease in the number of salient fixations along the arrow. The most salient fixations appear at approximately 1s." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 319, + 546, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 546, + 369 + ], + "type": "text", + "content": "and equal distribution. The equal duration model outperforms the equal distribution one and is easier to interpret; we present a comparison of these two alternatives as an ablation study in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 373, + 444, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 444, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 444, + 387 + ], + "type": "text", + "content": "4.2. Temporal saliency model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 392, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 392, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 546, + 559 + ], + "type": "text", + "content": "Let us now introduce our framework that exploits temporal human attention information. Our model is depicted in Figure 5. We extract image features using a pre-trained object recognition encoder [33]. Then, we decode these features by a temporal slice decoder to obtain one saliency map per time slice. These temporal saliency slices are useful in automated visual appeal assessment in applications such as website design [32], advertisement [36] and infographics [13]. In parallel, we decode the same image features into an initial image saliency prediction. Finally, we combine the temporal slices and the image saliency predictions in the spatiotemporal mixing module to produce a final image saliency map. We describe each component in detail in the following sections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 574, + 496, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 496, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 496, + 586 + ], + "type": "text", + "content": "4.2.1 Image encoder and saliency decoders" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 593, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 593, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 593, + 545, + 664 + ], + "type": "text", + "content": "Following the previous saliency prediction architectures [26, 31, 40], we first encode the input image with a pretrained image classification network, in our case PNASNet-5 [33]. We extract encoded features at various levels for multi-level integration, similar to a U-Net structure [41]. Formally, we denote the image encoder as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 670, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 362, + 670, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {E} (\\mathcal {I}) = \\left[ \\mathcal {E} _ {i} \\right], \\quad i \\in \\{1, \\dots , 5 \\}, \\tag {3}", + "image_path": "b3fc25c0424d55275e36e8db31a8e82cb904df27aebd384b07a01ff18e77fa69.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the input image, and " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " encoder block. The output of " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " therefore is a 5D vector. The early en" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6464" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 225 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 225 + ], + "type": "image", + "image_path": "abb002f7c2ec13981456acd36eccae123401d7651d21feb97d8dbd2fc00ffcf7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 548, + 295 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 548, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 548, + 295 + ], + "type": "text", + "content": "Figure 5. Overview of the proposed architecture. We encode image features into encoder blocks consisting of multi-level image features. We then pass these blocks to the temporal saliency decoder (shown in orange) to decode them into temporal saliency predictions, which are saliency maps in sequential time intervals. In parallel, the image saliency decoder (shown in green) decodes the encoder blocks into an image saliency prediction. We then combine the temporal saliency maps, the image saliency map, and the encoder blocks in the spatiotemporal mixing module (shown in pink). (Best viewed in color.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 310, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 370 + ], + "type": "text", + "content": "coder blocks extract low-level features, such as edges, color, and contrast, while the later blocks encode high-level semantics. We pass these blocks to the temporal saliency decoder, the image saliency decoder, and the spatiotemporal mixing module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "type": "text", + "content": "Our temporal slice decoder, namely " + }, + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T" + }, + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "type": "text", + "content": ", processes the encoder blocks with four 3x3 convolution layers followed by ReLU functions, integrating one encoder block after each convolution. Later, two 3x3 convolution layers with a ReLU in-between and a sigmoid function at the end produce " + }, + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 370, + 287, + 455 + ], + "type": "text", + "content": " temporal saliency maps. Formally, we write the temporal saliency decoder as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 464, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 464, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 79, + 464, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {T} \\big (\\mathcal {E} (\\mathcal {I}) \\big) = [ \\mathcal {T} _ {n} ] =: \\mathcal {T}, \\quad n \\in \\{1, \\dots , 5 \\} \\tag {4}", + "image_path": "e99a7cb42beaf332a03897517f5477534b964cfe93510bf3c48ca62c053da110.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_n" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "n^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " temporal saliency slice. Through this branch of the network, a single image input produces " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " temporal saliency slices. We use this component to provide temporal predictions to the spatiotemporal mixing module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "text", + "content": "Our image saliency decoder, namely " + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "text", + "content": ", has the same structure as " + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_T" + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "text", + "content": ", with the exception of the number of output channels. This component produces a single map " + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "inline_equation", + "content": "S_I" + }, + { + "bbox": [ + 46, + 536, + 287, + 597 + ], + "type": "text", + "content": ", which corresponds to the conventional image saliency map of the input image. As such, we can write" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 605, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 287, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {I} = \\mathcal {D} _ {S} (\\mathcal {E} (\\mathcal {I})). \\tag {5}", + "image_path": "9d33f8da15e8a2b3e57739c36c8107c99cd8f31c8eb7600a60eb39c879b85567.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 629, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 654 + ], + "type": "text", + "content": "We use this module to provide cumulative saliency information to the spatiotemporal mixing module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 216, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 216, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 216, + 683 + ], + "type": "text", + "content": "4.2.2 Spatiotemporal Mixing Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "To incorporate temporal information into the saliency prediction, we introduce a module that combines temporal and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 310, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 430 + ], + "type": "text", + "content": "spatial saliency maps. Our design is inspired by the feature pyramid networks [29], which have the benefit of integrating features from multiple levels of the encoder, thus capturing low-level cues (such as color contrast, brightness, and edges) and high-level ones (such as semantics and scene context). Such cues have been shown to be critical for saliency estimation [15, 24, 38]. This module takes temporal saliency predictions, the initial image saliency prediction, and the encoded image feature blocks as input. We write this as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 435, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 435, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 335, + 435, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {R} = \\operatorname {S M M} \\left(\\mathcal {E} (\\mathcal {I}), \\mathcal {T}, \\mathcal {S} _ {I}\\right), n \\in \\{1, \\dots , 5 \\}, \\tag {6}", + "image_path": "38f224f96148e621915b25966cca1025b8f4f4092d54b024ae125afe81aecf38.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 453, + 533, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 533, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 533, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 453, + 533, + 465 + ], + "type": "inline_equation", + "content": "S_{R}" + }, + { + "bbox": [ + 305, + 453, + 533, + 465 + ], + "type": "text", + "content": " denotes the final, refined image saliency map." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "text", + "content": "The module architecture is shown in Figure 6. It takes the last two encoder blocks " + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "inline_equation", + "content": "[\\mathcal{E}_5, \\mathcal{E}_4]" + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "text", + "content": " and passes them through a " + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "text", + "content": " convolution. We then concatenate the other encoder blocks with the image saliency and temporal saliency maps passing through " + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "text", + "content": " convolution, ReLU, and linear upsampling to keep the spatial dimensions consistent. We add the static and temporal saliency maps to the encoded image features at each block to prevent the information in these maps from vanishing. In the last step, we only add the saliency maps to output a final refined saliency map " + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "inline_equation", + "content": "S_R" + }, + { + "bbox": [ + 304, + 466, + 546, + 657 + ], + "type": "text", + "content": ". The final design of the module (i.e., combining E4 and E5) is based on empirical performance. This module eliminates the need for optimizing a weight parameter between the spatial and temporal maps. It can also modulate the maps within the spatial range of convolutions, which allows the selection of different regions from different maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 404, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 404, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 404, + 682 + ], + "type": "text", + "content": "4.2.3 Loss Functions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": "To train our network, we use the Kullback-Leibler divergence (KL) [45] and the Correlation Coefficient (CC)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6465" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 298, + 177 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 298, + 177 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 298, + 177 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 298, + 177 + ], + "type": "image", + "image_path": "f5ed3128e6c7678728e74b86e95a39992ba6aecaf587da806bdcfa0220bbf434.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "text", + "content": "Figure 6. The spatiotemporal mixing module combines temporal saliency predictions with the conventional image saliency prediction by multi-level image feature integration. " + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "inline_equation", + "content": "S_{I}" + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "text", + "content": " denotes the predicted image saliency map, " + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "inline_equation", + "content": "T_{1,\\dots,n}" + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "text", + "content": " the temporal saliency predictions for " + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "text", + "content": " frames, and " + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{1,\\dots,5}" + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "text", + "content": " the encoder blocks. This multi-level integration scheme provides information from earlier layers of the network to the next blocks in this module. " + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "inline_equation", + "content": "S_{R}" + }, + { + "bbox": [ + 46, + 184, + 289, + 274 + ], + "type": "text", + "content": " denotes the temporally refined image saliency map output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 286, + 318 + ], + "type": "text", + "content": "[21] between the predicted and ground truth saliency maps. First, we train the temporal branch using" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 325, + 287, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 287, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} (\\mathcal {I}) = \\lambda_ {1} * \\operatorname {C C} \\left(G T _ {n}, \\mathcal {T} _ {n}\\right) + \\beta_ {1} * \\operatorname {K L} \\left(G T _ {n}, \\mathcal {T} _ {n}\\right), \\tag {7}", + "image_path": "d5db1d0814fc4fbe4456193223bd5d6904a1e2105a10a1027a1498147db0c6af.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "type": "inline_equation", + "content": "GT_{n}" + }, + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "type": "text", + "content": " denotes the temporal ground truth for the " + }, + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "type": "inline_equation", + "content": "n^{th}" + }, + { + "bbox": [ + 47, + 346, + 287, + 384 + ], + "type": "text", + "content": " slice. We then freeze the weights in this component and train the spatiotemporal mixing module using" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 391, + 287, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 391, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 59, + 391, + 287, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} (\\mathcal {I}) = \\lambda_ {2} * \\operatorname {C C} (G T, \\mathcal {S} _ {R}) + \\beta_ {2} * \\operatorname {K L} (G T, \\mathcal {S} _ {R}), \\tag {8}", + "image_path": "9070ffa2c55f2b8b04eaa6542b49dc3f5ca5961d88d994a93b95832d9fe5b16c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "type": "text", + "content": " is the image saliency ground truth for image " + }, + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 412, + 282, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 434, + 191, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 191, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 191, + 448 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 454, + 164, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 164, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 164, + 468 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 472, + 288, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 288, + 545 + ], + "type": "text", + "content": "We use a batch size of 32 and an initial learning rate of 1e-4, reduced by a factor of ten every two epochs. We train the temporal branch first and then freeze the weights. We found that 10 epochs of training on SALICON was sufficient. For SALICON, we used the provided test, train, and validation splits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 552, + 106, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 106, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 106, + 562 + ], + "type": "text", + "content": "5.2. Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 570, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 593 + ], + "type": "text", + "content": "We evaluate the obtained saliency predictions according to the following standard metrics used by the community." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Area Under the Curve (AUC) [5]: Saliency prediction can be interpreted as classifying fixation vs non-fixation points. The area under the ROC curve shows the trade-off between true positives (TP) and false positives (FP). A higher AUC score indicates fewer FPs. While AUC computes the TP and FP rates using all the ground truth fixation points, sAUC [4] samples FP points from ground truth fixations of other observers, compensating for center bias in natural images by taking for inter- and intra-observer variability into account. Normalized Scanpath Saliency (NSS) [38]: This metric" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "compares the predicted saliency values at the ground truth fixation points to the average predicted saliency. An NSS score of one indicates that the predicted saliency values at the ground truth fixation points are one standard deviation above the average." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 132, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 545, + 191 + ], + "type": "text", + "content": "Kullback - Leibler Divergence (KL) [45]: The KL measures the cumulative distance between the predicted and the ground truth saliency maps. A KL score close to zero indicates a better approximation of the ground truth saliency map by the predicted one." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 251 + ], + "type": "text", + "content": "Pearson's correlation coefficient (CC) [21]: This metric measures the linear relationship between the predicted and ground truth saliency maps. It ranges from -1 to 1. A CC score close to one indicates a strong linear correlation between the two maps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 251, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 311 + ], + "type": "text", + "content": "Similarity (SIM) score [22]: The similarity score sums the minimum value between the predicted and the ground truth saliency maps over all pixels. A similarity score of 1 indicates a perfect prediction since both of the maps are probability distributions summing to 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 312, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 360 + ], + "type": "text", + "content": "Information Gain (IG) score [27]: The information gain is an information-theoretic metric that measures the difference in average log-likelihood between the predicted saliency map and center-bias prior." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 372, + 424, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 424, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 424, + 384 + ], + "type": "text", + "content": "5.3.Quantitative Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 392, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 559 + ], + "type": "text", + "content": "We compare our method with the state-of-the-art models, namely SAM-Resnet [10], MSI-Net, GazeGAN, MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14], in image saliency prediction. Our model outperforms these methods in five out of seven metrics, showing the benefit of incorporating temporal information. Moreover, our model outperforms the only other multi-duration saliency model [14] in image saliency prediction by a significant margin. Furthermore, we compare our model with this multi-duration model and a multi-duration baseline trained on individual temporal slices. Our model improves the saliency prediction in two durations consisting of 0.5 and 3 seconds in two out of three metrics and in all three metrics in the five-second duration." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 572, + 525, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 572, + 525, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 572, + 525, + 585 + ], + "type": "text", + "content": "5.4. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": "We first evaluate the performance of our model on image saliency prediction on the SALICON benchmark [20]. The ground truth of SALICON's test set is exclusively hosted on the CodaLab website2. Table 3 shows the comparison of standard evaluation metrics for different state-of-the-art saliency models alongside our model TempSAL. TempSAL outperforms all the baselines in almost all metrics. When it does not, it still yields competitive results." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 492, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 492, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 492, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 315, + 702, + 492, + 713 + ], + "type": "text", + "content": "https://competitions.codalab.org/competitions/17136" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6466" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 68, + 519, + 198 + ], + "blocks": [ + { + "bbox": [ + 77, + 68, + 519, + 198 + ], + "lines": [ + { + "bbox": [ + 77, + 68, + 519, + 198 + ], + "spans": [ + { + "bbox": [ + 77, + 68, + 519, + 198 + ], + "type": "table", + "html": "
ModelMDAUC ↑CC ↑KL ↓SAUC ↑IG ↑NSS ↑SIM ↑
SAM-Resnet [10]X0.8650.8990.6100.7410.5381.9900.793
MSI-Net [24]X0.8650.8990.3070.7360.7931.9310.784
GazeGAN [6]X0.8640.8790.3760.7360.7201.8990.773
SimpleNet [40]X0.8690.9070.2010.7430.8801.9600.793
MDNSal [40]X0.8650.8990.2210.7360.8631.9350.790
UNISAL [12]X0.8640.8790.3540.7390.7801.9520.775
DeepGaze IIE [31]X0.8690.8720.2850.7670.7661.9960.733
MD-SEM [14]✓0.8640.8680.5680.7460.6602.0580.774
TempSAL✓0.8690.9110.1950.7450.8961.9670.800
", + "image_path": "19ac404a43052f0b3a7a2fef7ab6ae360e8431ec88ada98a582ca5d7b4fc6eed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 270, + 541, + 339 + ], + "blocks": [ + { + "bbox": [ + 45, + 205, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 45, + 205, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 45, + 205, + 547, + 262 + ], + "type": "text", + "content": "Table 3. Evaluation results on the SALICON (LSUN 2017) test benchmark. We compare our model with the state-of-the-art saliency prediction models, namely SAM-Resnet [10], MSI-Net [24], GazeGAN [6], MDNSal [40], SimpleNet [40], DeepGaze IIE [31], and MD-SEM [14]. The results in bold show the best performance. Our method outperforms the state-of-the-art on conventional image saliency in five metrics. The MD column denotes the ability of the models to predict multi-duration saliency. Our model outperforms the only other multi-duration saliency model by a significant margin on five out of seven metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 270, + 541, + 339 + ], + "lines": [ + { + "bbox": [ + 53, + 270, + 541, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 541, + 339 + ], + "type": "table", + "html": "
ModelSlice 1 (0-500 ms)Slice 2 (0-3000 ms)Slice 3 (0-5000 ms)Average
CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑CC↑KL↓NSS↑
SAM-MD [14]0.8050.3703.1810.7380.4692.5410.7150.5352.4950.7530.4582.739
MD-SEM [14]0.8160.3513.3740.7450.4522.6940.7340.4872.6770.7650.4302.915
TempSAL0.8190.4963.4220.7520.5122.7030.8220.4713.3370.7970.4933.154
", + "image_path": "7494abdb5497e4291b13bd49af0bc91041b06d3cb78051630639ceb238765dca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 279, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 279, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 279, + 432 + ], + "type": "text", + "content": "5.5. Comparison with the multi-duration method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 521 + ], + "type": "text", + "content": "To compare our method with the only other multiduration model [14], we modify our network to output three temporal slices. We train our network on a three slice SALICON multi-duration dataset first and then fine-tune it on the CodeCharts1k dataset [14] using the given training and validation splits. We report the results of the comparison in Table 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 160, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 160, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 160, + 540 + ], + "type": "text", + "content": "5.6. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "In Figure 7, we compare the temporal and image saliency maps obtained with our method with the ground truth from SALICON [20]. Our model learns time-specific predictions and is able to combine such predictions into a conventional image saliency map. We provide additional qualitative results in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 146, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 146, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 146, + 635 + ], + "type": "text", + "content": "5.7. Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "In this section, we investigate the effect of different components in our model. We also provide a comparison with a multi-duration baseline model on the temporal SALICON dataset. Moreover, we provide an ablation study on the effect of the number of time slices and the two temporal slicing methods in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "text", + "content": "Effect of the SMM module: We evaluate the effect of the spatiotemporal mixing module (SMM) and the image saliency decoder in Table 5. The first model consists of the image encoder and temporal saliency decoder only. We take the average of the temporal slices to measure its performance by comparing it with the ground truth image saliency map. In the second row, we add the image saliency decoder to our model. Similarly, we take the average of the predicted maps " + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_n" + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "inline_equation", + "content": "S_I" + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "text", + "content": ". Lastly, we add the spatiotemporal mixing module, which effectively modulates these predicted maps and combines them into a final image saliency map " + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "inline_equation", + "content": "S_R" + }, + { + "bbox": [ + 304, + 419, + 547, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 558, + 534, + 613 + ], + "blocks": [ + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "lines": [ + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "text", + "content": "Table 4. Results of our model, MD-SEM [14], and the baseline SAM-MD [14] across different durations on the CodeCharts1k dataset [14]. Our model improves the saliency prediction in the first two intervals, consisting of " + }, + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "inline_equation", + "content": "500~\\mathrm{ms}" + }, + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "inline_equation", + "content": "3000~\\mathrm{ms}" + }, + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "text", + "content": " observations, in two out of three metrics. In this comparison, the time slices are cumulative, not mutually exclusive. Although our model benefits from non-overlapping temporal slices, it also performs well with cumulative time slices, particularly in the last slice corresponding to the image saliency with an observation duration of " + }, + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "inline_equation", + "content": "5000~\\mathrm{ms}" + }, + { + "bbox": [ + 45, + 346, + 547, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 558, + 534, + 613 + ], + "lines": [ + { + "bbox": [ + 318, + 558, + 534, + 613 + ], + "spans": [ + { + "bbox": [ + 318, + 558, + 534, + 613 + ], + "type": "table", + "html": "
ModelCC ↑KL ↓NSS ↑SIM ↑
DT(E(I))0.8520.2431.9730.754
+DS(E(I))0.8570.2521.9430.760
+SMM0.9060.1981.9300.798
", + "image_path": "21f80f5109139943c355c69bfe55cb6c6c7566fd5d3f8f733b8a177df09bd9ff.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 622, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 547, + 711 + ], + "type": "text", + "content": "Table 5. Results of ablation studies on the temporal SALICON validation dataset. The first row denotes the model with only the temporal saliency decoder. In the second row, the model has both the temporal and image saliency decoders. The last row denotes the performance with the spatiotemporal mixing module (SMM). As evidenced by the improved accuracy metrics, the SMM effectively modulates the spatial and temporal saliency maps to refine the initial image saliency prediction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6467" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 531, + 381 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 531, + 381 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 531, + 381 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 531, + 381 + ], + "type": "image", + "image_path": "28652b45f66112ae8ea863ee92661c48428b9089e513e89dc83ac92f351ad510.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 388, + 546, + 443 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 546, + 443 + ], + "type": "text", + "content": "Figure 7. The first row shows the input image with the ground truth saliency overlaid. The second row shows the ground truth temporal saliency and image saliency. We show our temporal saliency and image saliency predictions on the third row. Red-yellow maps are temporal saliency maps for one-second intervals. Black and white maps are image saliency maps for the whole observation duration. Our approach captures the attention shifts in sequential temporal maps. Moreover, our model is able to produce accurate image saliency predictions which are close to the ground truth maps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "content": "Comparison with a temporal baseline model: The performance of our TempSAL model with five temporal slices is provided in Table 6. Note that each saliency slice contains five times fewer samples than the original image saliency map. Therefore, individual slices contain more variation compared to conventional accumulated maps. As a baseline to our model, we compute the performance of an architecture composed of five replicated SimpleNet models (5xSimpleNet) [40], each trained on one saliency slice. This baseline model uses an unshared encoder and decoder for each slice, while we share the decoder among slices. Therefore, we do not benefit from increased model capacity. We observe an accuracy decline in the baseline model, which confirms the increased discrepancy in the data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 634, + 119, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 119, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 119, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "We present a saliency prediction method that can learn time-specific predictions and is also able to exploit temporal information to improve overall image saliency prediction. In particular, we show that the temporally evolving patterns in human attention play an important" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 445, + 543, + 531 + ], + "blocks": [ + { + "bbox": [ + 309, + 445, + 543, + 531 + ], + "lines": [ + { + "bbox": [ + 309, + 445, + 543, + 531 + ], + "spans": [ + { + "bbox": [ + 309, + 445, + 543, + 531 + ], + "type": "table", + "html": "
TimeBaselineTempSAL
CC↑KL↓NSS↑SIM↑CC↑KL↓NSS↑SIM↑
T10.8980.2112.4360.7780.8990.2142.4530.782
T20.8700.2192.1590.7650.8770.2152.2110.776
T30.8400.2471.8400.7530.8430.2471.8780.758
T40.8200.2731.7290.7430.8250.2641.7400.749
T50.8110.2751.6460.7380.8130.2761.6540.743
Average0.8480.2451.9620.7560.8520.2431.9870.761
", + "image_path": "99a3ebc885f68051023d4200285511e902c8b4ed08b66790682efe014cf9b720.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 536, + 545, + 601 + ], + "lines": [ + { + "bbox": [ + 305, + 536, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 536, + 545, + 601 + ], + "type": "text", + "content": "Table 6. Results of the baseline model (left) and our TempSAL model (right) across different time slices. In 18 out of 20 comparisons, our model consistently outperforms the baseline. Note that both models perform best in the first slice, in which the intra-slice agreement is more prominent than in the other slices, as mentioned in Section 3.2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 673 + ], + "type": "text", + "content": "role in saliency prediction in natural images. This is evidenced by our experiments that demonstrate our TempSAL method outperforming the state-of-the-art, including a multi-duration method exploiting cumulative temporal saliency maps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the Swiss National Science Foundation via the Sinergia grant CRSII5-180359." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6468" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta and Sabine Süsstrunk. Saliency detection for content-aware image resizing. In IEEE International Conference on Image Processing (ICIP), pages 1005-1008, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Bahar Aydemir, Deblina Bhattacharjee, Seungryong Kim, Tong Zhang, Mathieu Salzmann, and Sabine Süsstrunk. Modeling object dissimilarity for deep saliency prediction. Transactions on Machine Learning Research (TMLR), 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Ali Borji and Laurent Itti. CAT2000: A large scale fixation dataset for boosting saliency research. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Ali Borji, Dicky N. Sihite, and Laurent Itti. Quantitative analysis of human-model agreement in visual saliency modeling: A comparative study. IEEE Transactions on Image Processing (TIP), 22(1):55-69, 2013. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 338 + ], + "type": "text", + "content": "[5] Zoya Bylinskii, Tilke Judd, Aude Oliva, Antonio Torralba, and Fredo Durand. What do different evaluation metrics tell us about saliency models? IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41(3):740, 2019. 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 396 + ], + "type": "text", + "content": "[6] Zhaohui Che, Ali Borji, Guangtao Zhai, Xiongkuo Min, Guodong Guo, and Patrick Le Callet. Gazegan: A generative adversarial saliency model based on invariance analysis of human gaze during scene free viewing. ArXiv, abs/1905.06803, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[7] Ming-Ming Cheng, Niloy J. Mitra, Xiaolei Huang, Philip H. S. Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 37(3):569-582, 2015. 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[8] Hannah Faye Chua, Julie E. Boland, and Richard E. Nisbett. Cultural variation in eye movements during scene perception. Proceedings of the National Academy of Sciences, 102(35):12629-12633, 2005. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "text", + "content": "[9] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. A deep multi-level network for saliency prediction. In IEEE International Conference on Pattern Recognition (ICPR), pages 3488-3493, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "text", + "content": "[10] Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, and Rita Cucchiara. Predicting human eye fixations via an LSTM-based saliency attentive model. IEEE Transactions on Image Processing (TIP), 27(10):5142-5154, 2018. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[12] Richard Droste, Jianbo Jiao, and J. Alison Noble. Unified Image and Video Saliency Modeling. In European Conference on Computer Vision (ECCV), 2020. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Camilo Fosco, Vincent Casser, Amish Kumar Bedi, Peter O'Donovan, Aaron Hertzmann, and Zoya Bylinskii. Predicting visual importance across graphic design types. In ACM Symposium on User Interface Software and Technology (UIST), page 249-260, 2020. 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "text", + "content": "[14] Camilo Fosco, Anelise Newman, Pat Sukhum, Yun Bin Zhang, Nanxuan Zhao, Aude Oliva, and Zoya Bylinskii. How much time do you have? modeling multi-duration saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4482, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 185 + ], + "type": "text", + "content": "[15] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 34(10):1915-1926, 2011. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 546, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 546, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 546, + 220 + ], + "type": "text", + "content": "[16] Joseph H. Goldberg and Jonathan I. Helfman. Visual scanpath representation. In Symposium on Eye-Tracking Research & Applications (ETRA), pages 203–210, 2010. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 222, + 547, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 547, + 276 + ], + "type": "text", + "content": "[17] Xun Huang, Chengyao Shen, Xavier Boix, and Qi Zhao. SALICON: Reducing the semantic gap in saliency prediction by adapting deep neural networks. In IEEE International Conference on Computer Vision (ICCV), pages 262-270, 2015. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 278, + 546, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 278, + 546, + 322 + ], + "spans": [ + { + "bbox": [ + 307, + 278, + 546, + 322 + ], + "type": "text", + "content": "[18] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 20(11):1254-1259, 1998. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 323, + 546, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 323, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 546, + 357 + ], + "type": "text", + "content": "[19] Sen Jia and Neil D. B. Bruce. EML-NET: An expandable Multi-Layer NETwork for saliency prediction. Image and Vision Computing, 95:103887, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 358, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 546, + 403 + ], + "type": "text", + "content": "[20] Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. SALICON: Saliency in context. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 546, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 449 + ], + "type": "text", + "content": "[21] Timothee Jost, Nabil Ouerhani, Roman von Wartburg, René Muri, and Heinz Hügli. Assessing the contribution of color in visual attention. Computer Vision and Image Understanding, 100(1-2):107-123, 2005. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 450, + 546, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 546, + 483 + ], + "type": "text", + "content": "[22] Tilke Judd, Fredo Durand, and Antonio Torralba. A benchmark of computational models of saliency to predict human fixations. MIT Technical Report, 2012. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 485, + 546, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 485, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 485, + 546, + 529 + ], + "type": "text", + "content": "[23] Tilke Judd, Krista Ehinger, Fredo Durand, and Antonio Torralba. Learning to predict where humans look. In IEEE International Conference on Computer Vision (ICCV). IEEE, 2009. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 531, + 546, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 546, + 575 + ], + "type": "text", + "content": "[24] Alexander Kroner, Mario Senden, Kurt Driessens, and Rainer Goebel. Contextual encoder-decoder network for visual saliency prediction. Neural Networks, 129:261 - 270, 2020. 2, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 577, + 546, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 546, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 546, + 621 + ], + "type": "text", + "content": "[25] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep Gaze I: Boosting saliency prediction with feature maps trained on ImageNet. In International Conference on Learning Representations (ICLR) Workshops, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 623, + 546, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 623, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 623, + 546, + 666 + ], + "type": "text", + "content": "[26] Matthias Kümmerer, Tom Wallis, and Matthias Bethge. DeepGaze II: Predicting fixations from deep features over time and tasks. Journal of Vision (JOV), 17(10):1147, 2017. 1, 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "text", + "content": "[27] Matthias Kümmerer, Thomas S. A. Wallis, and Matthias Bethge. Information-theoretic model comparison unifies saliency metrics. Proceedings of the National Academy of Sciences (PNAS), 112(52):16054-16059, 2015. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6469" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[28] Qingshan Li, Yue Zhou, and Jie Yang. Saliency based image segmentation. In IEEE International Conference on Multimedia Technology, pages 5068-5071, 2011. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 288, + 162 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2117-2125, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 219 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 287, + 276 + ], + "type": "text", + "content": "[31] Akis Linardos, Matthias Kümmerer, Ori Press, and Matthias Bethge. Deepgaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12919-12928, 2021. 1, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 278, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 278, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 278, + 287, + 322 + ], + "type": "text", + "content": "[32] Gitte Lindgaard, Gary Fernandes, Cathy Dudek, and J. Brown. Attention web designers: You have 50 milliseconds to make a good first impression! Behaviour & Information Technology, 25(2):115-126, 2006. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 324, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 379 + ], + "type": "text", + "content": "[33] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In European Conference on Computer Vision (ECCV), pages 19-34, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 381, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 287, + 424 + ], + "type": "text", + "content": "[34] Nian Liu and Junwei Han. A deep spatial contextual long-term recurrent convolutional network for saliency detection. IEEE Transactions on Image Processing (TIP), 27(7):3264-3274, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 427, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 287, + 460 + ], + "type": "text", + "content": "[35] Sandeep Mishra and Oindrila Saha. RecSal : Deep recursive supervision for visual saliency prediction. In British Machine Vision Conference (BMVC), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 462, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 506 + ], + "type": "text", + "content": "[36] Johanna Palcu, Jennifer Sudkamp, and Arnd Florack. Judgments at gaze value: Gaze cuing in banner advertisements, its effect on attention allocation and product judgments. Frontiers in Psychology, 8, 2017. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 508, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 287, + 551 + ], + "type": "text", + "content": "[37] Yash Patel, Srikar Appalaraju, and R. Manmatha. Saliency driven perceptual image compression. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 227-236, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 554, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 586 + ], + "type": "text", + "content": "[38] Robert J. Peters, Asha Iyer, Laurent Itti, and Christof Koch. Components of bottom-up gaze allocation in natural images. Vision Research, 45(18):2397-2416, 2005. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[39] Michael Posner, Robert Rafal, Lisa Choatec, and Jonathan Vaughan. Inhibition of return: Neural basis and function. Cognitive Neuropsychology, Vol. 2:211 - 228, 09 1985. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[40] Navyasri Reddy, Samyak Jain, Pradeep Yarlagadda, and Vineet Gandhi. Tidying deep saliency prediction architectures. In International Conference on Intelligent Robots and Systems (IROS), 2020. 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 508 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. International Conference on Learning Representations (ICLR), 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[43] Markus Steinberger, Bernhard Kainz, Stefan Hauswiesner, Rostislav Khlebnikov, Denis Kalkofen, and Dieter Schmalstieg. Ray prioritization using stylization and visual saliency. Computers & Graphics, 36(6):673-684, 2012. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[44] Po-He Tseng, Ran Carmi, Ian G. M. Cameron, Douglas P. Munoz, and Laurent Itti. Quantifying center bias of observers in free viewing of dynamic natural scenes. Journal of Vision (JOV), 9(7):4-4, 07 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[45] Mathukumalli Vidyasagar. Kullback-leibler divergence rate between probability distributions on sets of different cardinalities. In IEEE Conference on Decision and Control (CDC), pages 948-953, 2010. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[46] Eleonora Vig, Michael Dorr, and David Cox. Large-scale optimization of hierarchical features for saliency prediction in natural images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2798-2805, 2014. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[47] W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, and R. Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(06):3239-3259, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 396 + ], + "type": "text", + "content": "[48] Sheng Yang, Guosheng Lin, Qiuping Jiang, and Weisi Lin. A dilated inception network for visual saliency prediction. IEEE Transactions on Multimedia, 22(8):2163-2176, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 398, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 419 + ], + "type": "text", + "content": "[49] Alfred L. Yarbus. Eye Movements and Vision, volume 2. Plenum Press, 1967. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[50] Yifeng Zhang, Ming Jiang, and Qi Zhao. Saliency prediction with external knowledge. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 484-493, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 508 + ], + "type": "text", + "content": "[51] Jufeng Zhao, Yueting Chen, Huajun Feng, Zhihai Xu, and Qi Li. Fast image enhancement using multi-scale saliency extraction in infrared imagery. Optik, 125(15):4039-4042, 2014. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6470" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_content_list.json b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..be928a7e52421e874b84f1c3fd3b23fd638c7f54 --- /dev/null +++ b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "Temporal Attention Unit: Towards Efficient Spatiotemporal Predictive Learning", + "text_level": 1, + "bbox": [ + 145, + 127, + 823, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Tan*, Zhangyang Gao*, Lirong Wu, Yongjie Xu, Jun Xia, Siyuan Li, Stan Z. Li† \nAI Lab, Research Center for Industries of the Future, Westlake University", + "bbox": [ + 148, + 175, + 828, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tancheng,gaozhangyang,wulirong,xuyongjie,xiajun, lisiyuan,stan.zq.li}@westlake.edu.cn", + "bbox": [ + 107, + 214, + 857, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 263, + 313, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatiotemporal predictive learning aims to generate future frames by learning from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame features and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recurrent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unparallelizable architectures. To parallelize the temporal module, we propose the Temporal Attention Unit (TAU), which decomposes temporal attention into intra-frame statistical attention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regularization to take inter-frame variations into account. Extensive experiments demonstrate that the proposed method enables the derived model to achieve competitive performance on various spatiotemporal prediction benchmarks.", + "bbox": [ + 75, + 295, + 473, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 612, + 209, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The last decade has witnessed revolutionary advances in deep learning across various supervised learning tasks such as image classification [34,52,87], object detection [73,75], computational biology [21, 22, 43, 83, 84], and etc. Despite significant breakthroughs in supervised learning, which relies on large-scale labeled datasets, the potential of unsupervised learning remains largely untapped. Self-supervised learning that designs pretext tasks to produce labels derived from the data itself is recognized as a subset of unsupervised learning. In the context of self-supervised learning, contrastive self-supervised learning [7,8,28,33,85,86,92,104] predicts the noise contrastive estimation from predefined positive or negative pairs, and masked self-supervised learning [17, 32, 45, 51, 57, 102] predicts the masked patches from the visible patches. Unlike these image-level self", + "bbox": [ + 75, + 637, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "supervised learning, spatiotemporal predictive learning that predicts future frames from past frames at the video-level [6, 19, 27, 46, 60, 63].", + "bbox": [ + 500, + 263, + 890, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5c293469b6f5e74b1e6033206984d84ea94bb1d9abdb072ddb8dcd75b091018.jpg", + "image_caption": [ + "(a) General framework" + ], + "image_footnote": [], + "bbox": [ + 509, + 321, + 614, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc32ec9ad6395a290747fab8c4f9d3486a90910666608caf856e2c2e0e4ed548.jpg", + "image_caption": [ + "(b) ConvLSTM" + ], + "image_footnote": [], + "bbox": [ + 651, + 321, + 751, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb3ddf32ff7706408aded2e7f89c9197dafd68e94b22f48dd6334c661798693a.jpg", + "image_caption": [ + "(c) E3D-LSTM" + ], + "image_footnote": [], + "bbox": [ + 792, + 321, + 890, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ace6ca9948b95a6cfff099dc3d0a384e61dd6b699131d9afd0472a6cccc7c3a4.jpg", + "image_caption": [ + "(d) PhyDNet" + ], + "image_footnote": [], + "bbox": [ + 511, + 493, + 614, + 627 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3a4ea7103d5b6672904d6f2093cf30ca3cf09d98405a5a41ad14688cba1a865.jpg", + "image_caption": [ + "(e) CrevNet" + ], + "image_footnote": [], + "bbox": [ + 651, + 494, + 751, + 627 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76b0760d4c0fd277c48ba60c4c1d8a34100c53c2f95af3c0a8a1e7700057ce87.jpg", + "image_caption": [ + "(f) Ours", + "Figure 1. Comparison of model architectures among common spatiotemporal predictive learning methods. Note that we denote 2D convolutional neural networks as Conv while 3D Conv means 3D convolutional neural networks." + ], + "image_footnote": [], + "bbox": [ + 792, + 493, + 890, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate spatiotemporal predictive learning can benefit broad practical applications in climate change [74, 77], human motion forecasting [91, 107], traffic flow prediction [18, 97], and representation learning [39, 71]. The significance of spatiotemporal predictive learning primarily lies in its potential of exploring both spatial correlation and temporal evolution in the physical world. Moreover, the self-supervised nature of spatiotemporal predictive learning aligns well with human learning styles without a large amount of labeled data. Massive videos can provide a rich source of visual information, enabling spatiotemporal pre", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18770", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dictive learning to serve as a generative pre-training strategy [35, 60] for feature representation learning towards diverse downstream visual supervised tasks.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the existing methods [2, 3, 9, 11, 12, 19, 24, 29, 41, 44, 65, 78, 79, 82, 88, 89, 93–96, 98, 100, 103] in spatiotemporal predictive learning employ hybrid architectures of convolutional neural networks and recurrent units in which spatial correlation and time evolution can be learned, respectively. Inspired by the success of long short-term memory (LSTM) [36] in sequential modeling, ConvLSTM [78] is a seminal work on the topic of spatiotemporal predictive learning that extends fully connected LSTM to convolutional LSTM towards accurate precipitation nowcasting. PredRNN [95] is an admirable work that proposes Spatiotemporal LSTM (ST-LSTM) units to model spatial appearances and temporal variations in a unified memory pool. This work provides insights on designing typical recurrent units for spatiotemporal predictive learning and inspires a series of subsequent works [4, 93, 96, 98]. E3D-LSTM [94] integrates 3D convolutional neural networks into recurrent units towards good representations with both short-term frame dependencies and long-term high-level relations. PhyDNet [29] introduces a two-branch architecture that involves physical-based PhyCells and ConvLSTMs for performing partial differential equation constraints in latent space. CrevNet [103] proposes an invertible two-way autoencoder based on flow [14, 15] and a conditionally reversible architecture for spatiotemporal predictive learning. As shown in Figure 1 (a), we present a general framework consisting of the spatial encoder/decoder and the middle temporal module, abstracted from these methods. Though these spatiotemporal predictive learning methods have different temporal modules and spatial encoders/decoders, they basically share a similar framework.", + "bbox": [ + 75, + 140, + 472, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the general framework, we argue that the temporal module plays an essential role in spatiotemporal predictive learning. While the common choice of the temporal module is the recurrent-based units, we explore a novel parallelizable attention module named Temporal Attention Unit (TAU) to capture time evolution. The proposed temporal attention is decomposed into intra-frame statistical attention and inter-frame dynamical attention. Furthermore, we argue that the mean square error loss only focuses on intra-frame differences, and we propose a differential divergence regularization that also cares about the inter-frame variations. Keeping the spatial encoder and decoder as simple as 2D convolutional neural networks, we deliberately implement our proposed TAU modules and surprisingly find the derived model achieves competitive performance as those recurrent-based models. This observation provides a new perspective to improve spatiotemporal predictive learning by parallelizable attention networks instead of commonly used recurrent units.", + "bbox": [ + 75, + 613, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on various datasets with different experimental settings: (1) Standard spatiotemporal predictive learning. Quantitative results on various datasets demonstrate our proposed method can achieve competitive performance on standard spatiotemporal predictive learning. (2) Generalization ability. To verify the generalization ability, we train our model on KITTI and test it on the Caltech Pedestrian dataset with different domains. (3) Predict future frames with flexible lengths. We tackle the long-length frames by feeding the predicted frames as the input and find the performance is consistently well. Through the superior performance in the above three experimental settings, we demonstrate that our proposed model can provide a novel manner that learns temporal dependencies without recurrent units.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 498, + 333, + 645, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Self-supervised learning", + "text_level": 1, + "bbox": [ + 500, + 357, + 720, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning has been well developed and applied in various fields [5, 55, 56, 109, 110]. Learning from massive data enables tremendous progress in supervised learning. By designing pretext tasks and generating labels from the data itself, self-supervised learning obtains supervisory signals. The model learns valuable representations by solving pretext tasks that leverage the underlying structure of the data. Early works on visual self-supervised learning design pretext tasks such as colorization [108], inpainting [69], rotation [26], jigsaw [66]. Contrastive self-supervised learning [7, 8, 28, 33, 92, 104] is a dominant manner in visual self-supervised learning that aims at a pretext task of grouping similar samples closer and diverse samples away from each other. However, contrastive self-supervised learning is limited by making pairs by multiple images, which affects its ability on small-scale datasets. Masked self-supervised learning [17, 32, 45, 51, 57, 102], which predicts the masked patches from the visible ones, is another research direction. Although masked pretraining has great success in natural language processing, its applications in visual tasks are challenging. Spatiotemporal predictive learning is another promising branch of self-supervised learning that focuses on video-level information and predicts future frames conditioned on past frames.", + "bbox": [ + 496, + 381, + 893, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the above image-level methods, spatiotemporal predictive learning focus on video-level information and predicts future frames conditioned on past frames. By learning the intrinsic motion dynamics, the model is enabled to easily decouple the foreground and background.", + "bbox": [ + 496, + 744, + 893, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Spatiotemporal predictive learning", + "text_level": 1, + "bbox": [ + 500, + 830, + 805, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recurrent models have achieved remarkable advances in spatiotemporal predictive learning. Inspired by recurrent neural networks, VideoModeling [62] adopts language", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18771", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modeling and quantizes the image patches into a large dictionary for recurrent units. ConvLSTM [78] leverages convolutional neural networks to model the LSTM architecture. PredNet [61] persistently predicts future video frames using deep recurrent convolutional neural networks with bottom-up and top-down connections. PredRNN [95] proposes a Spatiotemporal LSTM unit that extracts and memorizes spatial and temporal representations simultaneously, and its following work $\\mathrm{PredRNN + + }$ [93] further introduces gradient highway unit and Casual LSTM to adaptively capture temporal dependencies. E3D-LSTM [94] designs e-detic memory transition in recurrent convolutional units. Conv-TT-LSTM [82] employs a higher-order ConvLSTM to predict by combining convolutional features across time. MotionRNN [100] focuses on motion trends and transient variations. LMC-Memory [50] introduces a long-term motion context memory using memory alignment learning. PredRNN-v2 [96] extends PredRNN by leveraging a memory decoupling loss and curriculum learning strategy.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of using recurrent-based methods that are computationally expensive for spatiotemporal predictive learning, we introduce TAU, a model that uses visual attention mechanism to parallelize the temporal evolution without the recurrent structure. There are prior arts that have some similarities with our proposed model. PredCNN [101] and TrajectoryCNN [54] implement pure convolutional neural networks as the temporal module. SimVP [23] is a seminal work that applies blocks of Inception modules with a UNet architecture to learn the temporal evolution. Though their temporal modules are parallelizable, we argue that convolutions alone cannot capture long-term dependencies. Moreover, SimVP provides a simple baseline with minor complex attachment but a large space for further improvements. In general, SimVP first downsamples video sequences to reduce the computation, then uses Inception-UNet to learn essential spatiotemporal relationships, and upsamples the representations to predict future frames. Our work aims to replace the pivotal Inception-UNet with efficient attention modules that promote prediction performance. In this paper, we employ a simple yet effective attention mechanism to enable the temporal module not only to be parallelizable but also to capture long-term time evolution.", + "bbox": [ + 75, + 378, + 472, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 741, + 174, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 767, + 217, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formally define the spatiotemporal predictive learning problem as follows. Given a video sequence $\\mathcal{X}^{t,T} = \\{\\mathbf{x}^i\\}_{t - T + 1}^t$ at time $t$ with the past $T$ frames, we aim to predict the subsequent $T^{\\prime}$ frames $\\mathcal{Y}^{t + 1,T^{\\prime}} = \\{\\mathbf{x}^{i}\\}_{t + 1}^{t + 1 + T^{\\prime}}$ from time $t + 1$ , where $\\mathbf{x}_i\\in \\mathbb{R}^{C\\times H\\times W}$ is usually an image with channels $C$ , height $H$ , and width $W$ . In practice, we represent the video sequences as tensors, i.e., $\\mathcal{X}^{t,T}\\in$", + "bbox": [ + 75, + 790, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {R} ^ {T \\times C \\times H \\times W} \\text {a n d} \\mathcal {Y} ^ {t + 1, T ^ {\\prime}} \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C \\times H \\times W}.\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 88, + 794, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model with learnable parameters $\\Theta$ learns a mapping $\\mathcal{F}_{\\Theta}:\\mathcal{X}^{t,T}\\mapsto \\mathcal{Y}^{t + 1,T^{\\prime}}$ by exploring both spatial and temporal dependencies. In our case, the mapping $\\mathcal{F}_{\\Theta}$ is a neural network model trained to minimize the difference between the predicted future frames and the ground-truth future frames. The optimal parameters $\\Theta^{*}$ are:", + "bbox": [ + 496, + 106, + 893, + 198 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} \\left(\\mathcal {F} _ {\\Theta} \\left(\\mathcal {X} ^ {t, T}\\right), \\mathcal {Y} ^ {t + 1, T ^ {\\prime}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 208, + 890, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is a loss function that evaluates such differences.", + "bbox": [ + 498, + 243, + 880, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 268, + 609, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We illustrate the overview model in Figure 2 using the input Moving MNIST [81] data as an example.", + "bbox": [ + 498, + 292, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd3a7222ad8ab4804a5e072385ff3488a7ef65ca766d859d2da2b2e2f284ffbd.jpg", + "image_caption": [ + "Figure 2. The overview architecture of our proposed model." + ], + "image_footnote": [], + "bbox": [ + 519, + 338, + 870, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Striving for simplicity, the model follows the general framework in Figure 1, while the spatial encoder consists of four vanilla 2D convolutional layers, and the spatial decoder consists of four 2D transposed convolutional layers ('Conv2d' and 'ConvTranspose2d' in PyTorch respectively). We add a residual connection from the first convolutional layer to the last transposed convolutional layer for preserving the spatial-dependent features. Stacks of TAU modules are in the middle of the spatial encoder and decoder to extract temporal-dependent features. Though our model is simple, it can efficiently learn both spatial-dependent and temporal-dependent features without recurrent architectures.", + "bbox": [ + 496, + 549, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Temporal Attention Unit", + "text_level": 1, + "bbox": [ + 500, + 756, + 725, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose a batch of input video tensors $\\mathcal{B} \\in \\mathbb{R}^{B \\times T \\times C \\times H \\times W}$ with the number of videos $B = |\\mathcal{B}|$ is given. In the spatial encoder and decoder, we reshape the sequential input data $B \\times T \\times C \\times H \\times W$ as $(B \\times T) \\times C \\times H \\times W$ so that only spatial correlations are taken into account. In the temporal module, we reshape the feature $B \\times T \\times C \\times H \\times W$ as $B \\times (T \\times C) \\times H \\times W$ so that frames are arranged in order on the channel dimension.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18772", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We decompose the temporal attention into the intra-frame statical attention and the inter-frame dynamical attention, as shown in Figure 3. Inspired by the recent progress of vision Transformers (ViTs) [17,57] and large kernel convolutions [13, 30, 58], we propose to employ small kernel depth-wise convolutions (DW Conv), depth-wise convolutions with dilations (DW-D Conv), and $1 \\times 1$ convolutions to model the large kernel convolutions. Through the obtained large receptive field on intra-frames, the statical attention is able to capture long-range dependencies. However, the statical attention alone is not enough for learning temporal evolutions along the timeline. Thus, we employ the dynamical attention that learns the attention weights of channels in a squeeze-and-excitation manner [38]. The final attention is the product of dynamical attention and statical attention.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a61ec68faafa29d95fe56d8da580d234f0fbb960421c6c20f0b20b4824330584.jpg", + "image_caption": [ + "Figure 3. The intra-frame statical attention and the inter-frame dynamical attention." + ], + "image_footnote": [], + "bbox": [ + 81, + 330, + 472, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show the detailed scheme of our model in Figure 4. The proposed TAU module can be formally expressed as:", + "bbox": [ + 76, + 566, + 467, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S A} = \\operatorname {C o n v} _ {1 \\times 1} (\\mathrm {D W - D C o n v} (\\mathrm {D W C o n v} (H))),\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 611, + 428, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D A} = \\mathrm {F C} (\\operatorname {A v g P o o l} (H)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 630, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {\\prime} = (\\mathrm {S A} \\otimes \\mathrm {D A}) \\odot H,\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 648, + 300, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H \\in \\mathbb{R}^{B \\times (T \\times C') \\times H \\times W}$ is the hidden feature that will be fed into the TAU module, SA $\\in \\mathbb{R}^{B \\times (T \\times C') \\times H \\times W}$ , DA $\\in \\mathbb{R}^{B \\times (T \\times C') \\times 1 \\times 1}$ denote the statistical and dynamical attention, FC and AvgPool are fully connected layers and the average pooling. We represent the Kronecker product by $\\otimes$ and the Hadamard product by $\\odot$ .", + "bbox": [ + 75, + 679, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Differential Divergence Regularization", + "text_level": 1, + "bbox": [ + 76, + 784, + 410, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the prediction of our model, we further propose a differential divergence regularization that forces the model to learn the differences between consecutive frames and be aware of the inherent variation.", + "bbox": [ + 75, + 809, + 468, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the predicted frames $\\widehat{\\mathcal{Y}} = \\mathcal{F}_{\\Theta}(\\mathcal{X}) \\in \\mathbb{R}^{T' \\times C \\times H \\times W}$ and its corresponding ground-truth frames", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d583ca0e0b0e36fbf267a83ae0f4fcce757dd20b4d820959f029804b5c5c8add.jpg", + "image_caption": [ + "Figure 4. The detailed schema of our model." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 870, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{V}$ , we first calculate their forward difference $\\Delta \\widehat{\\mathcal{Y}}, \\Delta \\mathcal{Y} \\in \\mathbb{R}^{(T' - 1) \\times C \\times H \\times W}$ , where:", + "bbox": [ + 498, + 280, + 890, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta \\widehat {\\mathcal {Y}} _ {i} = \\widehat {\\mathcal {Y}} _ {i + 1} - \\widehat {\\mathcal {Y}} _ {i}, \\tag {3} \\\\ \\Delta \\mathcal {Y} _ {i} = \\mathcal {Y} _ {i + 1} - \\mathcal {Y} _ {i}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 324, + 890, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we transform the differences into probabilities by the softmax function on the channel, height, and width dimension and obtain $\\sigma (\\Delta \\widehat{\\mathcal{Y}}),\\sigma (\\Delta \\mathcal{Y})$ , where:", + "bbox": [ + 498, + 377, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sigma (\\Delta \\widehat {\\mathcal {Y}}) _ {i, j, k, l} = \\frac {\\exp \\left(\\Delta \\widehat {\\mathcal {Y}} _ {i , j , k , l} / \\tau\\right)}{\\sum_ {j ^ {\\prime} = 1} ^ {C} \\sum_ {k ^ {\\prime} = 1} ^ {H} \\sum_ {l ^ {\\prime} = 1} ^ {W} \\exp \\left(\\Delta \\widehat {\\mathcal {Y}} _ {i , j ^ {\\prime} , k ^ {\\prime} , l ^ {\\prime}} / \\tau\\right)}, \\\\ \\sigma \\left(\\Delta \\mathcal {Y}\\right) _ {i, j, k, l} = \\frac {\\exp \\left(\\Delta \\mathcal {Y} _ {i , j , k , l}\\right) / \\tau}{\\sum_ {j ^ {\\prime} = 1} ^ {C} \\sum_ {k ^ {\\prime} = 1} ^ {H} \\sum_ {l ^ {\\prime} = 1} ^ {W} \\exp \\left(\\Delta \\mathcal {Y} _ {i , j ^ {\\prime} , k ^ {\\prime} , l ^ {\\prime}} / \\tau\\right)}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 436, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\tau$ represents the temperature parameter which we empirically set as 0.1 to sharpen the probability distribution. Through the competition mechanism of the softmax function [10, 70, 99], the high difference frames are penalized.", + "bbox": [ + 498, + 527, + 890, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, the differential divergence regularization $\\mathcal{L}_{reg}$ is defined as the Kullback-Leibler divergence between the probability distributions $\\sigma (\\Delta \\widehat{\\mathcal{V}})$ and $\\sigma (\\Delta \\mathcal{V})$ :", + "bbox": [ + 498, + 589, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {r e g} (\\widehat {\\mathcal {Y}}, \\mathcal {Y}) = D _ {K L} (\\sigma (\\Delta \\widehat {\\mathcal {Y}}) \\mid \\mid \\sigma (\\Delta \\mathcal {Y})) \\\\ = \\sum_ {i = 1} ^ {T ^ {\\prime} - 1} \\sigma \\left(\\Delta \\hat {\\mathcal {Y}} _ {i}\\right) \\log \\frac {\\sigma \\left(\\Delta \\hat {\\mathcal {Y}} _ {i}\\right)}{\\sigma \\left(\\Delta \\mathcal {Y} _ {i}\\right)}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 647, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model is trained end-to-end in a fully unsupervised manner, and the overall objective function consists of the mean square error loss and the differential divergence regularization weighted by a constant $\\alpha$ :", + "bbox": [ + 498, + 726, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {T ^ {\\prime}} \\| \\widehat {\\mathcal {Y}} - \\mathcal {Y} \\| ^ {2} + \\alpha \\mathcal {L} _ {r e g} (\\widehat {\\mathcal {Y}}, \\mathcal {Y}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 800, + 890, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first term focuses on intra-frame-level differences, and the second regularization term focuses on inter-frame-level variations.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18773", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present experiments that demonstrate the effectiveness of our proposed method. The experiments are conducted on various datasets with different settings to validate our proposed model from three aspects:", + "bbox": [ + 75, + 114, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Standard spatiotemporal predictive learning (Section 4.2). We recognize the prediction problem of the same number of input and output frames as the standard spatiotemporal predictive learning. We evaluate the performance on standard spatiotemporal predictive learning and compare our model with state-of-the-art methods with Moving MNIST [81] and TaxiBJ [106] datasets.", + "- Generalization ability across different datasets (Section 4.3). Generalizing the learned knowledge to other domains is a challenge in unsupervised learning. We investigate the such ability of our method by training the model on the KITTI [25] dataset and evaluating it on the Caltech Pedestrian [16] dataset.", + "- Predicting frames with flexible lengths (Section 4.4). One of the advantages of recurrent units is that they can easily handle flexible-length frames like the KTH dataset [76]. Our work tackles the long-length frame prediction by imitating recurrent units that feed predicted frames as the input and recursively produce long-term predictions." + ], + "bbox": [ + 94, + 185, + 468, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 76, + 532, + 274, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We quantitatively evaluate our model on the following datasets for both synthetic and real-world scenarios:", + "bbox": [ + 76, + 556, + 468, + 587 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Moving MNIST [81] is a synthetic dataset consisting of two digits independently moving within the $64 \\times 64$ grid and bouncing off the boundary. It is a standard benchmark in spatiotemporal predictive learning.", + "- TaxiBJ contains the trajectory data in Beijing collected from taxicab GPS with two channels, i.e., inflow or outflow defined in [106]. Following the previous works [29, 98], we normalize the data into $[0, 1]$ .", + "- KTH [76] contains 25 individuals performing six types of actions. Following [90, 94], we use person 1-16 for training and 17-25 for testing. Models are trained to predict the next 20 or 40 frames from the previous 10 observations.", + "- Caltech Pedestrian is a driving dataset focusing on detecting pedestrians. It consists of approximately 10 hours of $640 \\times 480$ videos taken from vehicles driving through regular traffic in an urban environment. We follow the same protocol of PredNet [61] and" + ], + "bbox": [ + 94, + 595, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CrevNet [103] for pre-processing, training, and evaluation.", + "bbox": [ + 529, + 90, + 888, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize the statistics of the above datasets in Table 1, including the number of training samples $N_{train}$ and the number of testing samples $N_{test}$ .", + "bbox": [ + 498, + 133, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9caaa2808c30e822c694898d4930b62158aa22c34e2d5614b9d9d9a227ff73fc.jpg", + "table_caption": [ + "Table 1. The statistics of datasets. The training or testing set has $N_{train}$ or $N_{test}$ samples, composed by $T$ or $T'$ images with the shape $(C, H, W)$ ." + ], + "table_footnote": [], + "table_body": "
NtrainNtest(C,H,W)TT'
MMNIST1000010000(1,64,64)1010
TaxiBJ196271334(2,32,32)44
KTH52003167(1,128,128)1020 or 40
Caltech20421983(3,128,160)101
", + "bbox": [ + 501, + 247, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Measurement Following [29, 103], we employ Mean Squared Error (MSE), Mean Absolute Error (MAE), Structure Similarity Index Measure (SSIM), and Peak Signal to Noise Ratio (PSNR) to evaluate the quality of predictions. MSE and MAE estimate the absolute pixel-wise errors, SSIM measures the similarity of structural information within the spatial neighborhoods, and PSNR is an expression for the ratio between the maximum possible power of a signal and the power of distorted noise.", + "bbox": [ + 498, + 378, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details We implement the proposed method with the Pytorch framework and conduct experiments on a single NVIDIA-V100 GPU. The model is trained with a mini-batch of 16 video sequences while the AdamW optimizer is utilized with a learning rate of 0.01 and a weight decay of 0.05.", + "bbox": [ + 498, + 536, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Standard spatiotemporal predictive learning", + "text_level": 1, + "bbox": [ + 500, + 638, + 877, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Moving MNIST", + "text_level": 1, + "bbox": [ + 500, + 662, + 663, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This dataset is a standard benchmark in spatiotemporal predictive learning. We evaluate our proposed method against strong recent baselines, including competitive recurrent architectures: ConvLSTM [78], PredRNN [95], PredRNN++ [93], MIM [98], LMC [50], E3D-LSTM [94], Conv-TT-LSTM [82], and CrevNet [103]. We also compare the method DDPAE [37] that is specifically designed for this dataset. The quantitative results are reported in Table 2, and qualitative visualizations of the predicted results are shown in Figure 5.", + "bbox": [ + 496, + 688, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed method significantly outperforms all the baselines above under three different metrics. The performance gain is large with respect to state-of-the-art recurrent methods.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18774", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fbe5d313999c722cd256d67932b735dd8ce1ed7c7423bb70822b0b58353fb7bc.jpg", + "image_caption": [ + "Figure 5. Qualitative visualization of predicted results on Moving MNIST dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 893, + 334 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bbe974a303c38236ec0c4328e6239dc6f1ad62182749be19fa6206df00a741ad.jpg", + "table_caption": [ + "Table 2. Quantitative results of different methods on the Moving MNIST dataset (10 → 10 frames)." + ], + "table_footnote": [], + "table_body": "
MethodMoving MNIST
MSE↓MAE↓SSIM↑
ConvLSTM [78]103.3182.90.707
VPN [44]64.1-0.870
PredRNN [95]56.8126.10.867
PredRNN++ [93]46.5106.80.898
MIM [98]44.2101.10.910
LMC [50]41.5-0.924
E3D-LSTM [94]41.387.20.910
Conv-TT-LSTM [82]53.0-0.915
DDPAE [37]38.990.70.922
PhyDNet [29]24.470.30.947
SimVP [23]23.868.90.948
Crevnet [103]22.3-0.949
Ours19.860.30.957
", + "bbox": [ + 81, + 435, + 460, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 TaxiBJ", + "text_level": 1, + "bbox": [ + 76, + 707, + 181, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed model on a complicated real-world dataset, TaxiBJ [106]. Driven by human consciousness, the complex real-world traffic flows requires modeling transport phenomena and traffic diffusion for prediction. Due to the spatiotemporal nature of the traffic forecasting task, we straightforwardly implement our model for it.", + "bbox": [ + 75, + 732, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative visualizations of the predicted results are shown in Figure 6, and the quantitative results are reported in Table 3. Though the given frames are quite different from the future frames, our model can still accurately produce reliable frames. The difference between target frames", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/199de6d2bec999db32824ca989a31c8f958426e0847ec35051c8eef29d2791aa.jpg", + "image_caption": [ + "Figure 6. Qualitative visualization of predicted results on TaxiBJ dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + ], + "image_footnote": [], + "bbox": [ + 509, + 398, + 893, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and predicted frames is mainly located in central spots, but the overall trend is approximating the ground truth. It can also be observed that the quantitative results are consistently well, which demonstrates the practical application value in real-world scenarios.", + "bbox": [ + 496, + 738, + 893, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Generalization across different datasets", + "text_level": 1, + "bbox": [ + 500, + 829, + 839, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The generalization ability is one of the fundamental problems in artificial intelligence technology. Traditional supervised learning suffers from its poor generalization of", + "bbox": [ + 498, + 854, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18775", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ad642e9e4c5ef1047fb5c9fa5f9da974ca08334640ba131940ed28ead18a8ce0.jpg", + "table_caption": [ + "Table 3. Quantitative results of different methods on the TaxiBJ dataset (4 → 4 frames)." + ], + "table_footnote": [], + "table_body": "
MethodTaxiBJ
MSE × 100↓MAE↓SSIM↑
ConvLSTM [78]48.517.70.978
PredRNN [95]46.417.10.971
PredRNN++ [93]44.816.90.977
MIM [98]42.916.60.971
E3D-LSTM [94]43.216.90.979
PhyDNet [29]41.916.20.982
SimVP [23]41.416.20.982
Ours34.415.60.983
", + "bbox": [ + 78, + 127, + 468, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labeled datasets with different domains. Self-supervised learning aims to learn robust representations based on unlabeled data and evaluates the generalization ability of the learned model. While contrastive self-supervised learning and masked self-supervised learning in visual tasks usually evaluate such generalization ability by downstream tasks, we evaluate it by the prediction results across different datasets in spatiotemporal predictive learning.", + "bbox": [ + 75, + 324, + 470, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43795de1a2785063eb18e369c9dfc71db7ba3927c12e26cea33c646cd5a2d33b.jpg", + "image_caption": [ + "Figure 7. Qualitative visualization of predicted results on Caltech dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + ], + "image_footnote": [], + "bbox": [ + 76, + 464, + 472, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the previous works [53, 61, 103], we train the model using the raw video sequences from the KITTI dataset and evaluate the model by Caltech Pedestrian dataset that is made to match the frame rate and image size $(128 \\times 160)$ of the KTTI dataset. The final prediction is made for the next frame after a 10-frame warm-up.", + "bbox": [ + 75, + 792, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the qualitative visualization in Figure 7 and re", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "port the quantitative results in Table 4. Note that some of the baseline results are copied from [68]. It can be seen that our proposed method achieves state-of-the-art performance under both SSIM and PSNR metrics in this generalization evaluation task. Moreover, our model shows significantly robust predictions in terms of the variation of illumination and lane line, suggesting its practical applications in autonomous vehicles.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d07cf74ba0339100eb9e3575787b2b40098187db3ba0d35a91d799c5a8dd5fe.jpg", + "table_caption": [ + "Table 4. Quantitative results of different methods on the Caltech Pedestrian dataset (10 → 1 frame)." + ], + "table_footnote": [], + "table_body": "
MethodCaltech Pedestrian
SSIM↑PSNR↑
BeyondMSE [64]0.847-
MCnet [90]0.879-
DVF [59]0.89726.2
Dual-GAN [53]0.899-
CtrlGen [31]0.90026.5
PredNet [61]0.90527.6
ContextVP [4]0.92128.7
GAN-VGG [80]0.916-
G-VGG [80]0.917-
SDC-Net [72]0.918-
rCycleGan [47]0.91929.2
DPG [20]0.92328.2
G-MAE [80]0.923-
GAN-MAE [80]0.923-
CrevNet [103]0.92529.3
STMFANet [41]0.92729.1
SimVP [23]0.94033.1
Ours0.94633.7
", + "bbox": [ + 503, + 258, + 885, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Predicting frames with flexible lengths", + "text_level": 1, + "bbox": [ + 500, + 589, + 830, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Though recurrent units are adept at handling flexible-length frames, our model can also easily tackle such problems by imitating recurrent units that feed predicted frames as the input and recursively produce predictions. For this KTH dataset, our model is trained to predict the next 20 or 40 frames from the given 10 observations. Moreover, this dataset contains six types of human actions (walking, jogging, running, boxing, hand waving and hand clapping) performed several times by 25 subjects in four different scenarios: outdoors, outdoors with scale variations, outdoors with different clothes and indoors. The difficulty of this human motion prediction task not only lies in its flexible lengths of predicted frames but also in its complex dynamics involving the randomness of human consciousness. Following [95], we use the Peak Signal Noise Ratio (PSNR) and the Structural Similarity Index Measure (SSIM) as evaluation metrics to measure the framewise prediction quality from the perceptive view. The detailed quantitative results are shown in Table 5.", + "bbox": [ + 496, + 613, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18776", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7d616f49d984fe51fb7a7bb6836a378c006b4cf3392835d477b74832ab6ef48.jpg", + "table_caption": [ + "Table 5. Quantitative results of different methods on the KTH dataset (10 → 20 and 10 → 40 frames)." + ], + "table_footnote": [], + "table_body": "
MethodKTH (10 → 20)KTH (10 → 40)
SSIM↑PSNR↑SSIM↑PSNR↑
MCnet [90]0.80425.950.7323.89
ConvLSTM [78]0.71223.580.63922.85
SAVP [48]0.74625.380.70123.97
VPN [44]0.74623.76--
DFN [40]0.79427.260.65223.01
fRNN [67]0.77126.120.67823.77
Znet [105]0.81727.58--
SV2Pv [1]0.83827.790.78926.12
PredRNN [95]0.83927.550.70324.16
VarNet [42]0.84328.480.73925.37
SAVP-VAE [48]0.85227.770.81126.18
PredRNN++ [93]0.86528.470.74125.21
MSNET [49]0.87627.08--
E3d-LSTM [94]0.87929.310.81027.24
STMFANet [41]0.89329.850.85127.56
SimVP [23]0.90533.720.88632.93
Ours0.91134.130.89733.01
", + "bbox": [ + 80, + 127, + 460, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Empirical Running Time", + "text_level": 1, + "bbox": [ + 76, + 446, + 307, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TAU benefits from the parallelizable computation architecture, which leads to fast convergence and high training speed. We evaluate our efficiency by measuring the running time against state-of-the-art spatiotemporal predictive learning methods.", + "bbox": [ + 75, + 469, + 468, + 546 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/732999e3311e35bd1909c6417231d248f792a2d454ae85fc70f22e594ffad014.jpg", + "image_caption": [ + "Figure 8. The learning curve comparison between state-of-the-art methods and ours (evaluated by MSE). The red dotted line denotes MSE 35.0, and only the first 50 epochs are shown." + ], + "image_footnote": [], + "bbox": [ + 106, + 553, + 441, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The experiments are conducted on a single Tesla V100 GPU, and the batch size is set as 16. For the Moving MNIST dataset, CrevNet [103] needs about 30 minutes per epoch, and PhyDNet [29] needs about 7 minutes per epoch. Our model only requires 2.5 minutes per epoch. Furthermore, our method is able to converge at a rapid rate. As shown in Figure 8, on the Moving MNIST dataset, our model can achieve MSE 35.0 with only 50 epochs, while CrevNet and PhyDNet are far from this performance.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Computational Cost and Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 90, + 846, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the performance and computational cost with state-of-the-art methods in the first several rows in Table 6. Our model achieves superior results with better performance and much lower Flops. We also conduct ablation studies and summarize the results in Table 6. It can be seen that replacing TAU with the same number of convolutional blocks with vanilla $3 \\times 3$ convolutions (Conv Baseline) significantly degrades the performance. Training our model without differential divergence regularization (w/o DDR) will also weaken the prediction results. Both the TAU module and differential divergence regularization are useful. We also find that SA and DA of the TAU module play important roles in better performance.", + "bbox": [ + 496, + 113, + 890, + 309 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/19292d7295734c7b61002acdc24ec72fcf0ff200ef5090fd2ff5f31ac532330f.jpg", + "table_caption": [ + "Table 6. Ablation study of our proposed method." + ], + "table_footnote": [], + "table_body": "
MethodM-MNISTTaxiBJ (×100)Flops(G)
PredRNN56.846.4115.94
PredRNN++46.444.8171.73
MIM44.242.9179.17
E3DLSTM41.343.2298.85
SimVP23.841.419.43
Conv Baseline58.943.56.11
Ours w/o SA23.240.815.30
Ours w/o DA22.438.415.96
Ours w/o DDR21.137.715.96
Ours19.834.415.96
", + "bbox": [ + 504, + 347, + 879, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 547, + 619, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a general framework of spatiotemporal predictive learning and propose an attention-based temporal module to replace the common-used recurrent units. By decomposing the temporal module into the intra-frame statistical attention and the inter-frame dynamical attention, our proposed TAU module can achieve competitive performance across various experimental settings and datasets. Moreover, a novel differential divergence regularization is proposed to overcome the drawback of MSE loss that only considers the intra-frame error. In summary, our work highlights the importance of both intra-frame and inter-frame variations that enable the model to capture long-term relations and provide a new paradigm of efficient spatiotemporal predictive learning.", + "bbox": [ + 496, + 573, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (2022ZD0115100), the National Natural Science Foundation of China (U21A20427), the Competitive Research Fund (WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18777", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. In International Conference on Learning Representations, 2018. 8", + "[2] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. Fitvid: Overfitting in pixel-level video prediction. arXiv preprint arXiv:2106.13195, 2021. 2", + "[3] Sarthak Bhagat, Shagun Uppal, Zhuyun Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In European Conference on Computer Vision, pages 102-117. Springer, 2020. 2", + "[4] Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava, and Petros Koumoutsakos. Contextv: Fully context-aware video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 753-769, 2018. 2, 7", + "[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2", + "[6] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7608-7617, 2019. 1", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2", + "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in Neural Information Processing Systems, 33:22243-22255, 2020. 1, 2", + "[9] Yu Cheng, Bo Wang, Bo Yang, and Robby T Tan. Graph and temporal convolutional networks for 3d multi-person pose estimation in monocular videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 4, page 12, 2021. 2", + "[10] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. In International Conference on Learning Representations, 2020. 4", + "[11] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In International Conference on Machine Learning, pages 1174-1183. PMLR, 2018. 2", + "[12] Emily L Denton et al. Unsupervised learning of disentangled representations from video. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[13] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. arXiv preprint arXiv:2203.06717, 2022.4" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. In International Conference on Learning Representations Workshop, 2015. 2", + "[15] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 2", + "[16] Piotr Dólar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark. In 2009 IEEE conference on computer vision and pattern recognition, pages 304-311. IEEE, 2009. 5", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 4", + "[18] Shen Fang, Qi Zhang, Gaofeng Meng, Shiming Xiang, and Chunhong Pan. GSTnet: Global spatial-temporal network for traffic flow prediction. In IJCAI, pages 2286-2293, 2019. 1", + "[19] Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through video prediction. Advances in Neural Information Processing Systems, 29, 2016. 1, 2", + "[20] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7", + "[21] Zhangyang Gao, Cheng Tan, Stan Li, et al. Alphadesign: A graph protein design method and benchmark on alphafolddb. arXiv preprint arXiv:2202.01079, 2022. 1", + "[22] Zhangyang Gao, Cheng Tan, and Stan Z Li. Pifold: Toward effective and efficient protein inverse folding. arXiv preprint arXiv:2209.12643, 2022. 1", + "[23] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z. Li. Simvp: Simpler yet better video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3170-3180, June 2022. 3, 6, 7, 8", + "[24] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, pages 1243-1252. PMLR, 2017. 2", + "[25] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5", + "[26] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2", + "[27] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthew, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative varia" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18778", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019. 1", + "[28] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2", + "[29] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 2, 5, 6, 7, 8", + "[30] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 4", + "[31] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018. 7", + "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2", + "[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 1, 2", + "[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1", + "[35] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. arXiv preprint arXiv:2102.12627, 2021.2", + "[36] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2", + "[37] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li F Fei-Fei, and Juan Carlos Niebles. Learning to decompose and disentangle representations for video prediction. Advances in Neural Information Processing Systems, 31, 2018. 5, 6", + "[38] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4", + "[39] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision, pages 425-442. Springer, 2020. 1", + "[40] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 8" + ], + "bbox": [ + 86, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 2, 7, 8", + "[42] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 8", + "[43] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. Nature, 596(7873):583-589, 2021. 1", + "[44] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 2, 6, 8", + "[45] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1, 2", + "[46] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 1", + "[47] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 7", + "[48] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 8", + "[49] Jungbeom Lee, Jangho Lee, Sungmin Lee, and Sungroh Yoon. Mutual suppression network for video prediction using disentangled features. arXiv preprint arXiv:1804.04810, 2018. 8", + "[50] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 3, 5, 6", + "[51] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, 2020. 1, 2", + "[52] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1" + ], + "bbox": [ + 509, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18779", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Xiaodan Liang, Lisa Lee, Wei Dai, and Eric P Xing. Dual motion gan for future-flow embedded video prediction. In proceedings of the IEEE international conference on computer vision, pages 1744-1752, 2017. 7", + "[54] Xiaoli Liu, Jianqin Yin, Jin Liu, Pengxiang Ding, Jun Liu, and Huaping Liu. TrajectoryCNN: a new spatio-temporal feature learning network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2133-2146, 2020. 3", + "[55] Zicheng Liu, Siyuan Li, Ge Wang, Cheng Tan, Lirong Wu, and Stan Z Li. Decoupled mixup for data-efficient learning. arXiv preprint arXiv:2203.10761, 2022. 2", + "[56] Zicheng Liu, Siyuan Li, Di Wu, Zihan Liu, Zhiyuan Chen, Lirong Wu, and Stan Z Li. Automix: Unveiling the power of mixup for stronger classifiers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 441-458. Springer, 2022. 2", + "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4", + "[58] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 4", + "[59] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 7", + "[60] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning, pages 4114-4124. PMLR, 2019. 1, 2", + "[61] William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. In International Conference on Learning Representations, 2017. 3, 5, 7", + "[62] Arthur Szlam Marc'Aurelio Ranzato, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR, abs/1412.6604, 2, 2014. 2", + "[63] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019. 1", + "[64] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square error. arXiv preprint arXiv:1511.05440, 2015. 7", + "[65] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[66] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles." + ], + "bbox": [ + 86, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In European conference on computer vision, pages 69-84. Springer, 2016. 2", + "[67] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 8", + "[68] Sergiu Oprea, Pablo Martinez-Gonzalez, Alberto Garcia-Garcia, John Alejandro Castro-Vargas, Sergio Orts-Escalano, Jose Garcia-Rodriguez, and Antonis Argyros. A review on deep learning techniques for video prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7", + "[69] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2", + "[70] Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. Random feature attention. In International Conference on Learning Representations, 2020. 4", + "[71] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 1", + "[72] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 718-733, 2018. 7", + "[73] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016. 1", + "[74] Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, et al. Deep learning and process understanding for data-driven earth system science. Nature, 566(7743):195-204, 2019. 1", + "[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28:91-99, 2015. 1", + "[76] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 3, pages 32-36. IEEE, 2004. 5", + "[77] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 1", + "[78] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18780", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 2, 3, 5, 6, 7, 8", + "[79] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[80] Osamu Shouno. Photo-realistic video prediction on natural videos of largely changing frames. arXiv preprint arXiv:2003.08635, 2020. 7", + "[81] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 3, 5", + "[82] Jiahao Su, Wonmin Byeon, Jean Kossaifi, Furong Huang, Jan Kautz, and Anima Anandkumar. Convolutional tensortrain lstm for spatio-temporal learning. Advances in Neural Information Processing Systems, 33:13714-13726, 2020. 2, 3, 5, 6", + "[83] Cheng Tan, Zhangyang Gao, and Stan Z Li. Rfold: Towards simple yet effectiverna secondary structure prediction.arXiv preprint arXiv:2212.14041, 2022.1", + "[84] Cheng Tan, Zhangyang Gao, and Stan Z Li. Target-aware molecular graph generation. arXiv preprint arXiv:2202.04829, 2022.1", + "[85] Cheng Tan, Zhangyang Gao, Lirong Wu, Siyuan Li, and Stan Z Li. Hyperspherical consistency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7244-7255, 2022. 1", + "[86] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 1", + "[87] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1", + "[88] Ruben Villegas, Dumitru Erhan, Honglak Lee, et al. Hierarchical long-term video prediction without supervision. In International Conference on Machine Learning, pages 6038-6046. PMLR, 2018. 2", + "[89] Ruben Villegas, *Arkanath Pathak*, Harini Kannan, *Dumitru Erhan*, Quoc V Le, and Honglak Lee. *High fidelity video prediction with large stochastic recurrent neural networks*. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[90] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. In International Conference on Learning Representations, 2017. 5, 7, 8", + "[91] Pichao Wang, Wanqing Li, Philip Ogunbona, Jun Wan, and Sergio Escalera. Rgb-d-based human motion recognition with deep learning: A survey. Computer Vision and Image Understanding, 171:118-139, 2018. 1", + "[92] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uni" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2", + "[93] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrmn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 2, 3, 5, 6, 7, 8", + "[94] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 2, 3, 5, 6, 7, 8", + "[95] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in Neural Information Processing Systems, 30, 2017. 2, 3, 5, 6, 7, 8", + "[96] Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jianmin Wang, Philip S Yu, and Mingsheng Long. Predrnn: A recurrent neural network for spatiotemporal predictive learning. arXiv preprint arXiv:2103.09504, 2021. 2, 3", + "[97] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154–9162, 2019. 1", + "[98] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 2, 5, 6, 7", + "[99] Haixu Wu, Jialong Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Flowformer: Linearizing transformers with conservation flows. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 24226-24242. PMLR, 17-23 Jul 2022. 4", + "[100] Haixu Wu, Zhiyu Yao, Jianmin Wang, and Mingsheng Long. Motionrnn: A flexible model for video prediction with spacetime-varying motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15435-15444, 2021. 2, 3", + "[101] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 3", + "[102] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[103] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. In International Conference on Learning Representations, 2019. 2, 5, 6, 7, 8", + "[104] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18781", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "redundancy reduction. arXiv preprint arXiv:2103.03230, 2021. 1, 2", + "[105] Jianjin Zhang, Yunbo Wang, Mingsheng Long, Wang Jian-min, and S Yu Philip. Z-order recurrent neural networks for video prediction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 230-235. IEEE, 2019. 8", + "[106] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 5, 6", + "[107] Liang Zhang, Guangming Zhu, Peiyi Shen, Juan Song, Syed Afaq Shah, and Mohammed Bennamoun. Learning spatiotemporal features using 3dCNN and convolutional LSTM for gesture recognition. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 3120-3128, 2017. 1", + "[108] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016. 2", + "[109] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 2", + "[110] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Li. Using context-to-vector with graph retrofitting to improve word embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8154-8163, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 513 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "18782", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_model.json b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94c21037ae6eab705d41aa5c1be320e440919ad7 --- /dev/null +++ b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_model.json @@ -0,0 +1,3053 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.128, + 0.825, + 0.149 + ], + "angle": 0, + "content": "Temporal Attention Unit: Towards Efficient Spatiotemporal Predictive Learning" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.176, + 0.829, + 0.213 + ], + "angle": 0, + "content": "Cheng Tan*, Zhangyang Gao*, Lirong Wu, Yongjie Xu, Jun Xia, Siyuan Li, Stan Z. Li† \nAI Lab, Research Center for Industries of the Future, Westlake University" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.215, + 0.858, + 0.23 + ], + "angle": 0, + "content": "{tancheng,gaozhangyang,wulirong,xuyongjie,xiajun, lisiyuan,stan.zq.li}@westlake.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.264, + 0.314, + 0.28 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.474, + 0.584 + ], + "angle": 0, + "content": "Spatiotemporal predictive learning aims to generate future frames by learning from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame features and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recurrent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unparallelizable architectures. To parallelize the temporal module, we propose the Temporal Attention Unit (TAU), which decomposes temporal attention into intra-frame statistical attention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regularization to take inter-frame variations into account. Extensive experiments demonstrate that the proposed method enables the derived model to achieve competitive performance on various spatiotemporal prediction benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.613, + 0.21, + 0.629 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.866 + ], + "angle": 0, + "content": "The last decade has witnessed revolutionary advances in deep learning across various supervised learning tasks such as image classification [34,52,87], object detection [73,75], computational biology [21, 22, 43, 83, 84], and etc. Despite significant breakthroughs in supervised learning, which relies on large-scale labeled datasets, the potential of unsupervised learning remains largely untapped. Self-supervised learning that designs pretext tasks to produce labels derived from the data itself is recognized as a subset of unsupervised learning. In the context of self-supervised learning, contrastive self-supervised learning [7,8,28,33,85,86,92,104] predicts the noise contrastive estimation from predefined positive or negative pairs, and masked self-supervised learning [17, 32, 45, 51, 57, 102] predicts the masked patches from the visible patches. Unlike these image-level self" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.265, + 0.892, + 0.311 + ], + "angle": 0, + "content": "supervised learning, spatiotemporal predictive learning that predicts future frames from past frames at the video-level [6, 19, 27, 46, 60, 63]." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.323, + 0.615, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.468, + 0.637, + 0.482 + ], + "angle": 0, + "content": "(a) General framework" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.323, + 0.752, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.468, + 0.746, + 0.481 + ], + "angle": 0, + "content": "(b) ConvLSTM" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.323, + 0.891, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.469, + 0.886, + 0.482 + ], + "angle": 0, + "content": "(c) E3D-LSTM" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.494, + 0.615, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.633, + 0.596, + 0.647 + ], + "angle": 0, + "content": "(d) PhyDNet" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.495, + 0.752, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.634, + 0.734, + 0.647 + ], + "angle": 0, + "content": "(e) CrevNet" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.494, + 0.891, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.634, + 0.864, + 0.647 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.663, + 0.894, + 0.719 + ], + "angle": 0, + "content": "Figure 1. Comparison of model architectures among common spatiotemporal predictive learning methods. Note that we denote 2D convolutional neural networks as Conv while 3D Conv means 3D convolutional neural networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Accurate spatiotemporal predictive learning can benefit broad practical applications in climate change [74, 77], human motion forecasting [91, 107], traffic flow prediction [18, 97], and representation learning [39, 71]. The significance of spatiotemporal predictive learning primarily lies in its potential of exploring both spatial correlation and temporal evolution in the physical world. Moreover, the self-supervised nature of spatiotemporal predictive learning aligns well with human learning styles without a large amount of labeled data. Massive videos can provide a rich source of visual information, enabling spatiotemporal pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "dictive learning to serve as a generative pre-training strategy [35, 60] for feature representation learning towards diverse downstream visual supervised tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.141, + 0.473, + 0.612 + ], + "angle": 0, + "content": "Most of the existing methods [2, 3, 9, 11, 12, 19, 24, 29, 41, 44, 65, 78, 79, 82, 88, 89, 93–96, 98, 100, 103] in spatiotemporal predictive learning employ hybrid architectures of convolutional neural networks and recurrent units in which spatial correlation and time evolution can be learned, respectively. Inspired by the success of long short-term memory (LSTM) [36] in sequential modeling, ConvLSTM [78] is a seminal work on the topic of spatiotemporal predictive learning that extends fully connected LSTM to convolutional LSTM towards accurate precipitation nowcasting. PredRNN [95] is an admirable work that proposes Spatiotemporal LSTM (ST-LSTM) units to model spatial appearances and temporal variations in a unified memory pool. This work provides insights on designing typical recurrent units for spatiotemporal predictive learning and inspires a series of subsequent works [4, 93, 96, 98]. E3D-LSTM [94] integrates 3D convolutional neural networks into recurrent units towards good representations with both short-term frame dependencies and long-term high-level relations. PhyDNet [29] introduces a two-branch architecture that involves physical-based PhyCells and ConvLSTMs for performing partial differential equation constraints in latent space. CrevNet [103] proposes an invertible two-way autoencoder based on flow [14, 15] and a conditionally reversible architecture for spatiotemporal predictive learning. As shown in Figure 1 (a), we present a general framework consisting of the spatial encoder/decoder and the middle temporal module, abstracted from these methods. Though these spatiotemporal predictive learning methods have different temporal modules and spatial encoders/decoders, they basically share a similar framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Based on the general framework, we argue that the temporal module plays an essential role in spatiotemporal predictive learning. While the common choice of the temporal module is the recurrent-based units, we explore a novel parallelizable attention module named Temporal Attention Unit (TAU) to capture time evolution. The proposed temporal attention is decomposed into intra-frame statistical attention and inter-frame dynamical attention. Furthermore, we argue that the mean square error loss only focuses on intra-frame differences, and we propose a differential divergence regularization that also cares about the inter-frame variations. Keeping the spatial encoder and decoder as simple as 2D convolutional neural networks, we deliberately implement our proposed TAU modules and surprisingly find the derived model achieves competitive performance as those recurrent-based models. This observation provides a new perspective to improve spatiotemporal predictive learning by parallelizable attention networks instead of commonly used recurrent units." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "We conduct experiments on various datasets with different experimental settings: (1) Standard spatiotemporal predictive learning. Quantitative results on various datasets demonstrate our proposed method can achieve competitive performance on standard spatiotemporal predictive learning. (2) Generalization ability. To verify the generalization ability, we train our model on KITTI and test it on the Caltech Pedestrian dataset with different domains. (3) Predict future frames with flexible lengths. We tackle the long-length frames by feeding the predicted frames as the input and find the performance is consistently well. Through the superior performance in the above three experimental settings, we demonstrate that our proposed model can provide a novel manner that learns temporal dependencies without recurrent units." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.334, + 0.646, + 0.349 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.722, + 0.375 + ], + "angle": 0, + "content": "2.1. Self-supervised learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.744 + ], + "angle": 0, + "content": "Deep learning has been well developed and applied in various fields [5, 55, 56, 109, 110]. Learning from massive data enables tremendous progress in supervised learning. By designing pretext tasks and generating labels from the data itself, self-supervised learning obtains supervisory signals. The model learns valuable representations by solving pretext tasks that leverage the underlying structure of the data. Early works on visual self-supervised learning design pretext tasks such as colorization [108], inpainting [69], rotation [26], jigsaw [66]. Contrastive self-supervised learning [7, 8, 28, 33, 92, 104] is a dominant manner in visual self-supervised learning that aims at a pretext task of grouping similar samples closer and diverse samples away from each other. However, contrastive self-supervised learning is limited by making pairs by multiple images, which affects its ability on small-scale datasets. Masked self-supervised learning [17, 32, 45, 51, 57, 102], which predicts the masked patches from the visible ones, is another research direction. Although masked pretraining has great success in natural language processing, its applications in visual tasks are challenging. Spatiotemporal predictive learning is another promising branch of self-supervised learning that focuses on video-level information and predicts future frames conditioned on past frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.894, + 0.822 + ], + "angle": 0, + "content": "In contrast to the above image-level methods, spatiotemporal predictive learning focus on video-level information and predicts future frames conditioned on past frames. By learning the intrinsic motion dynamics, the model is enabled to easily decouple the foreground and background." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.807, + 0.849 + ], + "angle": 0, + "content": "2.2. Spatiotemporal predictive learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recurrent models have achieved remarkable advances in spatiotemporal predictive learning. Inspired by recurrent neural networks, VideoModeling [62] adopts language" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18771" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.379 + ], + "angle": 0, + "content": "modeling and quantizes the image patches into a large dictionary for recurrent units. ConvLSTM [78] leverages convolutional neural networks to model the LSTM architecture. PredNet [61] persistently predicts future video frames using deep recurrent convolutional neural networks with bottom-up and top-down connections. PredRNN [95] proposes a Spatiotemporal LSTM unit that extracts and memorizes spatial and temporal representations simultaneously, and its following work \\(\\mathrm{PredRNN + + }\\) [93] further introduces gradient highway unit and Casual LSTM to adaptively capture temporal dependencies. E3D-LSTM [94] designs e-detic memory transition in recurrent convolutional units. Conv-TT-LSTM [82] employs a higher-order ConvLSTM to predict by combining convolutional features across time. MotionRNN [100] focuses on motion trends and transient variations. LMC-Memory [50] introduces a long-term motion context memory using memory alignment learning. PredRNN-v2 [96] extends PredRNN by leveraging a memory decoupling loss and curriculum learning strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.473, + 0.728 + ], + "angle": 0, + "content": "Instead of using recurrent-based methods that are computationally expensive for spatiotemporal predictive learning, we introduce TAU, a model that uses visual attention mechanism to parallelize the temporal evolution without the recurrent structure. There are prior arts that have some similarities with our proposed model. PredCNN [101] and TrajectoryCNN [54] implement pure convolutional neural networks as the temporal module. SimVP [23] is a seminal work that applies blocks of Inception modules with a UNet architecture to learn the temporal evolution. Though their temporal modules are parallelizable, we argue that convolutions alone cannot capture long-term dependencies. Moreover, SimVP provides a simple baseline with minor complex attachment but a large space for further improvements. In general, SimVP first downsamples video sequences to reduce the computation, then uses Inception-UNet to learn essential spatiotemporal relationships, and upsamples the representations to predict future frames. Our work aims to replace the pivotal Inception-UNet with efficient attention modules that promote prediction performance. In this paper, we employ a simple yet effective attention mechanism to enable the temporal module not only to be parallelizable but also to capture long-term time evolution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.176, + 0.758 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.218, + 0.782 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We formally define the spatiotemporal predictive learning problem as follows. Given a video sequence \\(\\mathcal{X}^{t,T} = \\{\\mathbf{x}^i\\}_{t - T + 1}^t\\) at time \\(t\\) with the past \\(T\\) frames, we aim to predict the subsequent \\(T^{\\prime}\\) frames \\(\\mathcal{Y}^{t + 1,T^{\\prime}} = \\{\\mathbf{x}^{i}\\}_{t + 1}^{t + 1 + T^{\\prime}}\\) from time \\(t + 1\\), where \\(\\mathbf{x}_i\\in \\mathbb{R}^{C\\times H\\times W}\\) is usually an image with channels \\(C\\), height \\(H\\), and width \\(W\\). In practice, we represent the video sequences as tensors, i.e., \\(\\mathcal{X}^{t,T}\\in\\)" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.089, + 0.795, + 0.107 + ], + "angle": 0, + "content": "\\[\n\\mathbb {R} ^ {T \\times C \\times H \\times W} \\text {a n d} \\mathcal {Y} ^ {t + 1, T ^ {\\prime}} \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C \\times H \\times W}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.894, + 0.199 + ], + "angle": 0, + "content": "The model with learnable parameters \\(\\Theta\\) learns a mapping \\(\\mathcal{F}_{\\Theta}:\\mathcal{X}^{t,T}\\mapsto \\mathcal{Y}^{t + 1,T^{\\prime}}\\) by exploring both spatial and temporal dependencies. In our case, the mapping \\(\\mathcal{F}_{\\Theta}\\) is a neural network model trained to minimize the difference between the predicted future frames and the ground-truth future frames. The optimal parameters \\(\\Theta^{*}\\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.209, + 0.892, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} \\left(\\mathcal {F} _ {\\Theta} \\left(\\mathcal {X} ^ {t, T}\\right), \\mathcal {Y} ^ {t + 1, T ^ {\\prime}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.244, + 0.881, + 0.259 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is a loss function that evaluates such differences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.27, + 0.611, + 0.285 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.325 + ], + "angle": 0, + "content": "We illustrate the overview model in Figure 2 using the input Moving MNIST [81] data as an example." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.339, + 0.871, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.519, + 0.874, + 0.533 + ], + "angle": 0, + "content": "Figure 2. The overview architecture of our proposed model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Striving for simplicity, the model follows the general framework in Figure 1, while the spatial encoder consists of four vanilla 2D convolutional layers, and the spatial decoder consists of four 2D transposed convolutional layers ('Conv2d' and 'ConvTranspose2d' in PyTorch respectively). We add a residual connection from the first convolutional layer to the last transposed convolutional layer for preserving the spatial-dependent features. Stacks of TAU modules are in the middle of the spatial encoder and decoder to extract temporal-dependent features. Though our model is simple, it can efficiently learn both spatial-dependent and temporal-dependent features without recurrent architectures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.727, + 0.772 + ], + "angle": 0, + "content": "3.3. Temporal Attention Unit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Suppose a batch of input video tensors \\(\\mathcal{B} \\in \\mathbb{R}^{B \\times T \\times C \\times H \\times W}\\) with the number of videos \\(B = |\\mathcal{B}|\\) is given. In the spatial encoder and decoder, we reshape the sequential input data \\(B \\times T \\times C \\times H \\times W\\) as \\((B \\times T) \\times C \\times H \\times W\\) so that only spatial correlations are taken into account. In the temporal module, we reshape the feature \\(B \\times T \\times C \\times H \\times W\\) as \\(B \\times (T \\times C) \\times H \\times W\\) so that frames are arranged in order on the channel dimension." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18772" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "We decompose the temporal attention into the intra-frame statical attention and the inter-frame dynamical attention, as shown in Figure 3. Inspired by the recent progress of vision Transformers (ViTs) [17,57] and large kernel convolutions [13, 30, 58], we propose to employ small kernel depth-wise convolutions (DW Conv), depth-wise convolutions with dilations (DW-D Conv), and \\(1 \\times 1\\) convolutions to model the large kernel convolutions. Through the obtained large receptive field on intra-frames, the statical attention is able to capture long-range dependencies. However, the statical attention alone is not enough for learning temporal evolutions along the timeline. Thus, we employ the dynamical attention that learns the attention weights of channels in a squeeze-and-excitation manner [38]. The final attention is the product of dynamical attention and statical attention." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.331, + 0.473, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Figure 3. The intra-frame statical attention and the inter-frame dynamical attention." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.568, + 0.468, + 0.6 + ], + "angle": 0, + "content": "We show the detailed scheme of our model in Figure 4. The proposed TAU module can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.612, + 0.429, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S A} = \\operatorname {C o n v} _ {1 \\times 1} (\\mathrm {D W - D C o n v} (\\mathrm {D W C o n v} (H))),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.631, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D A} = \\mathrm {F C} (\\operatorname {A v g P o o l} (H)), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.65, + 0.301, + 0.667 + ], + "angle": 0, + "content": "\\[\nH ^ {\\prime} = (\\mathrm {S A} \\otimes \\mathrm {D A}) \\odot H,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.775 + ], + "angle": 0, + "content": "where \\(H \\in \\mathbb{R}^{B \\times (T \\times C') \\times H \\times W}\\) is the hidden feature that will be fed into the TAU module, SA \\(\\in \\mathbb{R}^{B \\times (T \\times C') \\times H \\times W}\\), DA \\(\\in \\mathbb{R}^{B \\times (T \\times C') \\times 1 \\times 1}\\) denote the statistical and dynamical attention, FC and AvgPool are fully connected layers and the average pooling. We represent the Kronecker product by \\(\\otimes\\) and the Hadamard product by \\(\\odot\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.411, + 0.802 + ], + "angle": 0, + "content": "3.4. Differential Divergence Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.869 + ], + "angle": 0, + "content": "To improve the prediction of our model, we further propose a differential divergence regularization that forces the model to learn the differences between consecutive frames and be aware of the inherent variation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Given the predicted frames \\(\\widehat{\\mathcal{Y}} = \\mathcal{F}_{\\Theta}(\\mathcal{X}) \\in \\mathbb{R}^{T' \\times C \\times H \\times W}\\) and its corresponding ground-truth frames" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.872, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.24, + 0.828, + 0.254 + ], + "angle": 0, + "content": "Figure 4. The detailed schema of our model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.281, + 0.892, + 0.313 + ], + "angle": 0, + "content": "\\(\\mathcal{V}\\), we first calculate their forward difference \\(\\Delta \\widehat{\\mathcal{Y}}, \\Delta \\mathcal{Y} \\in \\mathbb{R}^{(T' - 1) \\times C \\times H \\times W}\\), where:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.325, + 0.891, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta \\widehat {\\mathcal {Y}} _ {i} = \\widehat {\\mathcal {Y}} _ {i + 1} - \\widehat {\\mathcal {Y}} _ {i}, \\tag {3} \\\\ \\Delta \\mathcal {Y} _ {i} = \\mathcal {Y} _ {i + 1} - \\mathcal {Y} _ {i}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Then, we transform the differences into probabilities by the softmax function on the channel, height, and width dimension and obtain \\(\\sigma (\\Delta \\widehat{\\mathcal{Y}}),\\sigma (\\Delta \\mathcal{Y})\\) , where:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.437, + 0.891, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sigma (\\Delta \\widehat {\\mathcal {Y}}) _ {i, j, k, l} = \\frac {\\exp \\left(\\Delta \\widehat {\\mathcal {Y}} _ {i , j , k , l} / \\tau\\right)}{\\sum_ {j ^ {\\prime} = 1} ^ {C} \\sum_ {k ^ {\\prime} = 1} ^ {H} \\sum_ {l ^ {\\prime} = 1} ^ {W} \\exp \\left(\\Delta \\widehat {\\mathcal {Y}} _ {i , j ^ {\\prime} , k ^ {\\prime} , l ^ {\\prime}} / \\tau\\right)}, \\\\ \\sigma \\left(\\Delta \\mathcal {Y}\\right) _ {i, j, k, l} = \\frac {\\exp \\left(\\Delta \\mathcal {Y} _ {i , j , k , l}\\right) / \\tau}{\\sum_ {j ^ {\\prime} = 1} ^ {C} \\sum_ {k ^ {\\prime} = 1} ^ {H} \\sum_ {l ^ {\\prime} = 1} ^ {W} \\exp \\left(\\Delta \\mathcal {Y} _ {i , j ^ {\\prime} , k ^ {\\prime} , l ^ {\\prime}} / \\tau\\right)}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.892, + 0.589 + ], + "angle": 0, + "content": "and \\(\\tau\\) represents the temperature parameter which we empirically set as 0.1 to sharpen the probability distribution. Through the competition mechanism of the softmax function [10, 70, 99], the high difference frames are penalized." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Thus, the differential divergence regularization \\(\\mathcal{L}_{reg}\\) is defined as the Kullback-Leibler divergence between the probability distributions \\(\\sigma (\\Delta \\widehat{\\mathcal{V}})\\) and \\(\\sigma (\\Delta \\mathcal{V})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.648, + 0.891, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {r e g} (\\widehat {\\mathcal {Y}}, \\mathcal {Y}) = D _ {K L} (\\sigma (\\Delta \\widehat {\\mathcal {Y}}) \\mid \\mid \\sigma (\\Delta \\mathcal {Y})) \\\\ = \\sum_ {i = 1} ^ {T ^ {\\prime} - 1} \\sigma \\left(\\Delta \\hat {\\mathcal {Y}} _ {i}\\right) \\log \\frac {\\sigma \\left(\\Delta \\hat {\\mathcal {Y}} _ {i}\\right)}{\\sigma \\left(\\Delta \\mathcal {Y} _ {i}\\right)}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.727, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Our model is trained end-to-end in a fully unsupervised manner, and the overall objective function consists of the mean square error loss and the differential divergence regularization weighted by a constant \\(\\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {T ^ {\\prime}} \\| \\widehat {\\mathcal {Y}} - \\mathcal {Y} \\| ^ {2} + \\alpha \\mathcal {L} _ {r e g} (\\widehat {\\mathcal {Y}}, \\mathcal {Y}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where the first term focuses on intra-frame-level differences, and the second regularization term focuses on inter-frame-level variations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18773" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.178 + ], + "angle": 0, + "content": "In this section, we present experiments that demonstrate the effectiveness of our proposed method. The experiments are conducted on various datasets with different settings to validate our proposed model from three aspects:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.186, + 0.47, + 0.307 + ], + "angle": 0, + "content": "- Standard spatiotemporal predictive learning (Section 4.2). We recognize the prediction problem of the same number of input and output frames as the standard spatiotemporal predictive learning. We evaluate the performance on standard spatiotemporal predictive learning and compare our model with state-of-the-art methods with Moving MNIST [81] and TaxiBJ [106] datasets." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.317, + 0.47, + 0.409 + ], + "angle": 0, + "content": "- Generalization ability across different datasets (Section 4.3). Generalizing the learned knowledge to other domains is a challenge in unsupervised learning. We investigate the such ability of our method by training the model on the KITTI [25] dataset and evaluating it on the Caltech Pedestrian [16] dataset." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.418, + 0.47, + 0.526 + ], + "angle": 0, + "content": "- Predicting frames with flexible lengths (Section 4.4). One of the advantages of recurrent units is that they can easily handle flexible-length frames like the KTH dataset [76]. Our work tackles the long-length frame prediction by imitating recurrent units that feed predicted frames as the input and recursively produce long-term predictions." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.186, + 0.47, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.534, + 0.275, + 0.551 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.557, + 0.469, + 0.588 + ], + "angle": 0, + "content": "Datasets We quantitatively evaluate our model on the following datasets for both synthetic and real-world scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.597, + 0.469, + 0.659 + ], + "angle": 0, + "content": "- Moving MNIST [81] is a synthetic dataset consisting of two digits independently moving within the \\(64 \\times 64\\) grid and bouncing off the boundary. It is a standard benchmark in spatiotemporal predictive learning." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.668, + 0.469, + 0.73 + ], + "angle": 0, + "content": "- TaxiBJ contains the trajectory data in Beijing collected from taxicab GPS with two channels, i.e., inflow or outflow defined in [106]. Following the previous works [29, 98], we normalize the data into \\([0, 1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.739, + 0.469, + 0.815 + ], + "angle": 0, + "content": "- KTH [76] contains 25 individuals performing six types of actions. Following [90, 94], we use person 1-16 for training and 17-25 for testing. Models are trained to predict the next 20 or 40 frames from the previous 10 observations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- Caltech Pedestrian is a driving dataset focusing on detecting pedestrians. It consists of approximately 10 hours of \\(640 \\times 480\\) videos taken from vehicles driving through regular traffic in an urban environment. We follow the same protocol of PredNet [61] and" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.597, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "CrevNet [103] for pre-processing, training, and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.134, + 0.892, + 0.181 + ], + "angle": 0, + "content": "We summarize the statistics of the above datasets in Table 1, including the number of training samples \\(N_{train}\\) and the number of testing samples \\(N_{test}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.194, + 0.892, + 0.237 + ], + "angle": 0, + "content": "Table 1. The statistics of datasets. The training or testing set has \\( N_{train} \\) or \\( N_{test} \\) samples, composed by \\( T \\) or \\( T' \\) images with the shape \\( (C, H, W) \\)." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.34 + ], + "angle": 0, + "content": "
NtrainNtest(C,H,W)TT'
MMNIST1000010000(1,64,64)1010
TaxiBJ196271334(2,32,32)44
KTH52003167(1,128,128)1020 or 40
Caltech20421983(3,128,160)101
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Measurement Following [29, 103], we employ Mean Squared Error (MSE), Mean Absolute Error (MAE), Structure Similarity Index Measure (SSIM), and Peak Signal to Noise Ratio (PSNR) to evaluate the quality of predictions. MSE and MAE estimate the absolute pixel-wise errors, SSIM measures the similarity of structural information within the spatial neighborhoods, and PSNR is an expression for the ratio between the maximum possible power of a signal and the power of distorted noise." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Implementation details We implement the proposed method with the Pytorch framework and conduct experiments on a single NVIDIA-V100 GPU. The model is trained with a mini-batch of 16 video sequences while the AdamW optimizer is utilized with a learning rate of 0.01 and a weight decay of 0.05." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.878, + 0.657 + ], + "angle": 0, + "content": "4.2. Standard spatiotemporal predictive learning" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.665, + 0.679 + ], + "angle": 0, + "content": "4.2.1 Moving MNIST" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.839 + ], + "angle": 0, + "content": "This dataset is a standard benchmark in spatiotemporal predictive learning. We evaluate our proposed method against strong recent baselines, including competitive recurrent architectures: ConvLSTM [78], PredRNN [95], PredRNN++ [93], MIM [98], LMC [50], E3D-LSTM [94], Conv-TT-LSTM [82], and CrevNet [103]. We also compare the method DDPAE [37] that is specifically designed for this dataset. The quantitative results are reported in Table 2, and qualitative visualizations of the predicted results are shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Our proposed method significantly outperforms all the baselines above under three different metrics. The performance gain is large with respect to state-of-the-art recurrent methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18774" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.895, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.344, + 0.895, + 0.374 + ], + "angle": 0, + "content": "Figure 5. Qualitative visualization of predicted results on Moving MNIST dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.397, + 0.471, + 0.426 + ], + "angle": 0, + "content": "Table 2. Quantitative results of different methods on the Moving MNIST dataset (10 → 10 frames)." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.436, + 0.462, + 0.681 + ], + "angle": 0, + "content": "
MethodMoving MNIST
MSE↓MAE↓SSIM↑
ConvLSTM [78]103.3182.90.707
VPN [44]64.1-0.870
PredRNN [95]56.8126.10.867
PredRNN++ [93]46.5106.80.898
MIM [98]44.2101.10.910
LMC [50]41.5-0.924
E3D-LSTM [94]41.387.20.910
Conv-TT-LSTM [82]53.0-0.915
DDPAE [37]38.990.70.922
PhyDNet [29]24.470.30.947
SimVP [23]23.868.90.948
Crevnet [103]22.3-0.949
Ours19.860.30.957
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.182, + 0.722 + ], + "angle": 0, + "content": "4.2.2 TaxiBJ" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.824 + ], + "angle": 0, + "content": "We evaluate our proposed model on a complicated real-world dataset, TaxiBJ [106]. Driven by human consciousness, the complex real-world traffic flows requires modeling transport phenomena and traffic diffusion for prediction. Due to the spatiotemporal nature of the traffic forecasting task, we straightforwardly implement our model for it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The qualitative visualizations of the predicted results are shown in Figure 6, and the quantitative results are reported in Table 3. Though the given frames are quite different from the future frames, our model can still accurately produce reliable frames. The difference between target frames" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.399, + 0.895, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.666, + 0.895, + 0.708 + ], + "angle": 0, + "content": "Figure 6. Qualitative visualization of predicted results on TaxiBJ dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.894, + 0.815 + ], + "angle": 0, + "content": "and predicted frames is mainly located in central spots, but the overall trend is approximating the ground truth. It can also be observed that the quantitative results are consistently well, which demonstrates the practical application value in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.84, + 0.846 + ], + "angle": 0, + "content": "4.3. Generalization across different datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "The generalization ability is one of the fundamental problems in artificial intelligence technology. Traditional supervised learning suffers from its poor generalization of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18775" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 3. Quantitative results of different methods on the TaxiBJ dataset (4 → 4 frames)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.128, + 0.47, + 0.297 + ], + "angle": 0, + "content": "
MethodTaxiBJ
MSE × 100↓MAE↓SSIM↑
ConvLSTM [78]48.517.70.978
PredRNN [95]46.417.10.971
PredRNN++ [93]44.816.90.977
MIM [98]42.916.60.971
E3D-LSTM [94]43.216.90.979
PhyDNet [29]41.916.20.982
SimVP [23]41.416.20.982
Ours34.415.60.983
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.471, + 0.448 + ], + "angle": 0, + "content": "labeled datasets with different domains. Self-supervised learning aims to learn robust representations based on unlabeled data and evaluates the generalization ability of the learned model. While contrastive self-supervised learning and masked self-supervised learning in visual tasks usually evaluate such generalization ability by downstream tasks, we evaluate it by the prediction results across different datasets in spatiotemporal predictive learning." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.465, + 0.473, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Figure 7. Qualitative visualization of predicted results on Caltech dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Following the previous works [53, 61, 103], we train the model using the raw video sequences from the KITTI dataset and evaluate the model by Caltech Pedestrian dataset that is made to match the frame rate and image size \\((128 \\times 160)\\) of the KTTI dataset. The final prediction is made for the next frame after a 10-frame warm-up." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We show the qualitative visualization in Figure 7 and re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "port the quantitative results in Table 4. Note that some of the baseline results are copied from [68]. It can be seen that our proposed method achieves state-of-the-art performance under both SSIM and PSNR metrics in this generalization evaluation task. Moreover, our model shows significantly robust predictions in terms of the variation of illumination and lane line, suggesting its practical applications in autonomous vehicles." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 4. Quantitative results of different methods on the Caltech Pedestrian dataset (10 → 1 frame)." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.26, + 0.887, + 0.58 + ], + "angle": 0, + "content": "
MethodCaltech Pedestrian
SSIM↑PSNR↑
BeyondMSE [64]0.847-
MCnet [90]0.879-
DVF [59]0.89726.2
Dual-GAN [53]0.899-
CtrlGen [31]0.90026.5
PredNet [61]0.90527.6
ContextVP [4]0.92128.7
GAN-VGG [80]0.916-
G-VGG [80]0.917-
SDC-Net [72]0.918-
rCycleGan [47]0.91929.2
DPG [20]0.92328.2
G-MAE [80]0.923-
GAN-MAE [80]0.923-
CrevNet [103]0.92529.3
STMFANet [41]0.92729.1
SimVP [23]0.94033.1
Ours0.94633.7
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.831, + 0.607 + ], + "angle": 0, + "content": "4.4. Predicting frames with flexible lengths" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Though recurrent units are adept at handling flexible-length frames, our model can also easily tackle such problems by imitating recurrent units that feed predicted frames as the input and recursively produce predictions. For this KTH dataset, our model is trained to predict the next 20 or 40 frames from the given 10 observations. Moreover, this dataset contains six types of human actions (walking, jogging, running, boxing, hand waving and hand clapping) performed several times by 25 subjects in four different scenarios: outdoors, outdoors with scale variations, outdoors with different clothes and indoors. The difficulty of this human motion prediction task not only lies in its flexible lengths of predicted frames but also in its complex dynamics involving the randomness of human consciousness. Following [95], we use the Peak Signal Noise Ratio (PSNR) and the Structural Similarity Index Measure (SSIM) as evaluation metrics to measure the framewise prediction quality from the perceptive view. The detailed quantitative results are shown in Table 5." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18776" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 5. Quantitative results of different methods on the KTH dataset (10 → 20 and 10 → 40 frames)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.462, + 0.439 + ], + "angle": 0, + "content": "
MethodKTH (10 → 20)KTH (10 → 40)
SSIM↑PSNR↑SSIM↑PSNR↑
MCnet [90]0.80425.950.7323.89
ConvLSTM [78]0.71223.580.63922.85
SAVP [48]0.74625.380.70123.97
VPN [44]0.74623.76--
DFN [40]0.79427.260.65223.01
fRNN [67]0.77126.120.67823.77
Znet [105]0.81727.58--
SV2Pv [1]0.83827.790.78926.12
PredRNN [95]0.83927.550.70324.16
VarNet [42]0.84328.480.73925.37
SAVP-VAE [48]0.85227.770.81126.18
PredRNN++ [93]0.86528.470.74125.21
MSNET [49]0.87627.08--
E3d-LSTM [94]0.87929.310.81027.24
STMFANet [41]0.89329.850.85127.56
SimVP [23]0.90533.720.88632.93
Ours0.91134.130.89733.01
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.447, + 0.308, + 0.464 + ], + "angle": 0, + "content": "4.5. Empirical Running Time" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.547 + ], + "angle": 0, + "content": "TAU benefits from the parallelizable computation architecture, which leads to fast convergence and high training speed. We evaluate our efficiency by measuring the running time against state-of-the-art spatiotemporal predictive learning methods." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.554, + 0.442, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.757 + ], + "angle": 0, + "content": "Figure 8. The learning curve comparison between state-of-the-art methods and ours (evaluated by MSE). The red dotted line denotes MSE 35.0, and only the first 50 epochs are shown." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The experiments are conducted on a single Tesla V100 GPU, and the batch size is set as 16. For the Moving MNIST dataset, CrevNet [103] needs about 30 minutes per epoch, and PhyDNet [29] needs about 7 minutes per epoch. Our model only requires 2.5 minutes per epoch. Furthermore, our method is able to converge at a rapid rate. As shown in Figure 8, on the Moving MNIST dataset, our model can achieve MSE 35.0 with only 50 epochs, while CrevNet and PhyDNet are far from this performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.847, + 0.108 + ], + "angle": 0, + "content": "4.6. Computational Cost and Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.31 + ], + "angle": 0, + "content": "We compare the performance and computational cost with state-of-the-art methods in the first several rows in Table 6. Our model achieves superior results with better performance and much lower Flops. We also conduct ablation studies and summarize the results in Table 6. It can be seen that replacing TAU with the same number of convolutional blocks with vanilla \\(3 \\times 3\\) convolutions (Conv Baseline) significantly degrades the performance. Training our model without differential divergence regularization (w/o DDR) will also weaken the prediction results. Both the TAU module and differential divergence regularization are useful. We also find that SA and DA of the TAU module play important roles in better performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.322, + 0.842, + 0.337 + ], + "angle": 0, + "content": "Table 6. Ablation study of our proposed method." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.348, + 0.88, + 0.538 + ], + "angle": 0, + "content": "
MethodM-MNISTTaxiBJ (×100)Flops(G)
PredRNN56.846.4115.94
PredRNN++46.444.8171.73
MIM44.242.9179.17
E3DLSTM41.343.2298.85
SimVP23.841.419.43
Conv Baseline58.943.56.11
Ours w/o SA23.240.815.30
Ours w/o DA22.438.415.96
Ours w/o DDR21.137.715.96
Ours19.834.415.96
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.548, + 0.62, + 0.564 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.786 + ], + "angle": 0, + "content": "In this paper, we present a general framework of spatiotemporal predictive learning and propose an attention-based temporal module to replace the common-used recurrent units. By decomposing the temporal module into the intra-frame statistical attention and the inter-frame dynamical attention, our proposed TAU module can achieve competitive performance across various experimental settings and datasets. Moreover, a novel differential divergence regularization is proposed to overcome the drawback of MSE loss that only considers the intra-frame error. In summary, our work highlights the importance of both intra-frame and inter-frame variations that enable the model to capture long-term relations and provide a new paradigm of efficient spatiotemporal predictive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement. We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (2022ZD0115100), the National Natural Science Foundation of China (U21A20427), the Competitive Research Fund (WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18777" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.469, + 0.17 + ], + "angle": 0, + "content": "[1] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. In International Conference on Learning Representations, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.172, + 0.469, + 0.226 + ], + "angle": 0, + "content": "[2] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. Fitvid: Overfitting in pixel-level video prediction. arXiv preprint arXiv:2106.13195, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.228, + 0.469, + 0.297 + ], + "angle": 0, + "content": "[3] Sarthak Bhagat, Shagun Uppal, Zhuyun Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In European Conference on Computer Vision, pages 102-117. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.298, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[4] Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava, and Petros Koumoutsakos. Contextv: Fully context-aware video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 753-769, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.424, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[6] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7608-7617, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.481, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.537, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in Neural Information Processing Systems, 33:22243-22255, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.607, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[9] Yu Cheng, Bo Wang, Bo Yang, and Robby T Tan. Graph and temporal convolutional networks for 3d multi-person pose estimation in monocular videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 4, page 12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.677, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[10] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. In International Conference on Learning Representations, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[11] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In International Conference on Machine Learning, pages 1174-1183. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[12] Emily L Denton et al. Unsupervised learning of disentangled representations from video. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. arXiv preprint arXiv:2203.06717, 2022.4" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. In International Conference on Learning Representations Workshop, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.15, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[15] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Piotr Dólar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark. In 2009 IEEE conference on computer vision and pattern recognition, pages 304-311. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.25, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[18] Shen Fang, Qi Zhang, Gaofeng Meng, Shiming Xiang, and Chunhong Pan. GSTnet: Global spatial-temporal network for traffic flow prediction. In IJCAI, pages 2286-2293, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[19] Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through video prediction. Advances in Neural Information Processing Systems, 29, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.448, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[20] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Zhangyang Gao, Cheng Tan, Stan Li, et al. Alphadesign: A graph protein design method and benchmark on alphafolddb. arXiv preprint arXiv:2202.01079, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[22] Zhangyang Gao, Cheng Tan, and Stan Z Li. Pifold: Toward effective and efficient protein inverse folding. arXiv preprint arXiv:2209.12643, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.605, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[23] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z. Li. Simvp: Simpler yet better video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3170-3180, June 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[24] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, pages 1243-1252. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.733, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[25] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthew, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative varia" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18778" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "tional inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.121, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[28] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.261, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[29] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.33, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[30] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.372, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[31] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.428, + 0.47, + 0.495 + ], + "angle": 0, + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.497, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.567, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.623, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[35] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. arXiv preprint arXiv:2102.12627, 2021.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.665, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[36] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.693, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[37] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li F Fei-Fei, and Juan Carlos Niebles. Learning to decompose and disentangle representations for video prediction. Advances in Neural Information Processing Systems, 31, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.749, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[38] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[39] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision, pages 425-442. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 8" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[41] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.178, + 0.894, + 0.243 + ], + "angle": 0, + "content": "[42] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.247, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[43] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. Nature, 596(7873):583-589, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.317, + 0.894, + 0.383 + ], + "angle": 0, + "content": "[44] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.386, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[45] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.442, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[46] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.512, + 0.894, + 0.566 + ], + "angle": 0, + "content": "[47] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.568, + 0.893, + 0.62 + ], + "angle": 0, + "content": "[48] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.623, + 0.893, + 0.677 + ], + "angle": 0, + "content": "[49] Jungbeom Lee, Jangho Lee, Sungmin Lee, and Sungroh Yoon. Mutual suppression network for video prediction using disentangled features. arXiv preprint arXiv:1804.04810, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.679, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[50] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.749, + 0.893, + 0.844 + ], + "angle": 0, + "content": "[51] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[52] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18779" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Xiaodan Liang, Lisa Lee, Wei Dai, and Eric P Xing. Dual motion gan for future-flow embedded video prediction. In proceedings of the IEEE international conference on computer vision, pages 1744-1752, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[54] Xiaoli Liu, Jianqin Yin, Jin Liu, Pengxiang Ding, Jun Liu, and Huaping Liu. TrajectoryCNN: a new spatio-temporal feature learning network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2133-2146, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[55] Zicheng Liu, Siyuan Li, Ge Wang, Cheng Tan, Lirong Wu, and Stan Z Li. Decoupled mixup for data-efficient learning. arXiv preprint arXiv:2203.10761, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[56] Zicheng Liu, Siyuan Li, Di Wu, Zihan Liu, Zhiyuan Chen, Lirong Wu, and Stan Z Li. Automix: Unveiling the power of mixup for stronger classifiers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 441-458. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.414, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[58] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[59] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[60] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning, pages 4114-4124. PMLR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[61] William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. In International Conference on Learning Representations, 2017. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.651, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[62] Arthur Szlam Marc'Aurelio Ranzato, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR, abs/1412.6604, 2, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.706, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[63] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.762, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[64] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square error. arXiv preprint arXiv:1511.05440, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.804, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[65] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[66] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In European conference on computer vision, pages 69-84. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[67] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[68] Sergiu Oprea, Pablo Martinez-Gonzalez, Alberto Garcia-Garcia, John Alejandro Castro-Vargas, Sergio Orts-Escalano, Jose Garcia-Rodriguez, and Antonis Argyros. A review on deep learning techniques for video prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[69] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[70] Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. Random feature attention. In International Conference on Learning Representations, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.391, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[71] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[72] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 718-733, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.533, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[73] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.604, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[74] Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, et al. Deep learning and process understanding for data-driven earth system science. Nature, 566(7743):195-204, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.675, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28:91-99, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.732, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[76] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 3, pages 32-36. IEEE, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[77] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[78] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18780" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.469, + 0.134 + ], + "angle": 0, + "content": "network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[79] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.207, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[80] Osamu Shouno. Photo-realistic video prediction on natural videos of largely changing frames. arXiv preprint arXiv:2003.08635, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.249, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[81] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.306, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[82] Jiahao Su, Wonmin Byeon, Jean Kossaifi, Furong Huang, Jan Kautz, and Anima Anandkumar. Convolutional tensortrain lstm for spatio-temporal learning. Advances in Neural Information Processing Systems, 33:13714-13726, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.376, + 0.47, + 0.418 + ], + "angle": 0, + "content": "[83] Cheng Tan, Zhangyang Gao, and Stan Z Li. Rfold: Towards simple yet effectiverna secondary structure prediction.arXiv preprint arXiv:2212.14041, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.42, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[84] Cheng Tan, Zhangyang Gao, and Stan Z Li. Target-aware molecular graph generation. arXiv preprint arXiv:2202.04829, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.462, + 0.47, + 0.518 + ], + "angle": 0, + "content": "[85] Cheng Tan, Zhangyang Gao, Lirong Wu, Siyuan Li, and Stan Z Li. Hyperspherical consistency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7244-7255, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.519, + 0.47, + 0.575 + ], + "angle": 0, + "content": "[86] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.577, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[87] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.632, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[88] Ruben Villegas, Dumitru Erhan, Honglak Lee, et al. Hierarchical long-term video prediction without supervision. In International Conference on Machine Learning, pages 6038-6046. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[89] Ruben Villegas, *Arkanath Pathak*, Harini Kannan, *Dumitru Erhan*, Quoc V Le, and Honglak Lee. *High fidelity video prediction with large stochastic recurrent neural networks*. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[90] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. In International Conference on Learning Representations, 2017. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[91] Pichao Wang, Wanqing Li, Philip Ogunbona, Jun Wan, and Sergio Escalera. Rgb-d-based human motion recognition with deep learning: A survey. Computer Vision and Image Understanding, 171:118-139, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[92] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uni" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "formity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[93] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrmn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[94] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[95] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in Neural Information Processing Systems, 30, 2017. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.321, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[96] Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jianmin Wang, Philip S Yu, and Mingsheng Long. Predrnn: A recurrent neural network for spatiotemporal predictive learning. arXiv preprint arXiv:2103.09504, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.378, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[97] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154–9162, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.462, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[98] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.548, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[99] Haixu Wu, Jialong Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Flowformer: Linearizing transformers with conservation flows. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 24226-24242. PMLR, 17-23 Jul 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[100] Haixu Wu, Zhiyu Yao, Jianmin Wang, and Mingsheng Long. Motionrnn: A flexible model for video prediction with spacetime-varying motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15435-15444, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[101] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[102] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[103] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. In International Conference on Learning Representations, 2019. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[104] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "redundancy reduction. arXiv preprint arXiv:2103.03230, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[105] Jianjin Zhang, Yunbo Wang, Mingsheng Long, Wang Jian-min, and S Yu Philip. Z-order recurrent neural networks for video prediction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 230-235. IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[106] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[107] Liang Zhang, Guangming Zhu, Peiyi Shen, Juan Song, Syed Afaq Shah, and Mohammed Bennamoun. Learning spatiotemporal features using 3dCNN and convolutional LSTM for gesture recognition. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 3120-3128, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[108] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.376, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[109] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.433, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[110] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Li. Using context-to-vector with graph retrofitting to improve word embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8154-8163, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18782" + } + ] +] \ No newline at end of file diff --git a/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_origin.pdf b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f950164059efd4aa9231254b368474cebd737b3 --- /dev/null +++ b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/d3109e90-c15d-418e-bbe6-8d0dca603ccc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4f3d9fb1ca605330b3591691a6d4dc873a5f3ab9d24f2922bd270d634c5dedb +size 1282102 diff --git a/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/full.md b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9015d2cfcce158b2c2dc05a06872952446df4844 --- /dev/null +++ b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/full.md @@ -0,0 +1,385 @@ +# Temporal Attention Unit: Towards Efficient Spatiotemporal Predictive Learning + +Cheng Tan*, Zhangyang Gao*, Lirong Wu, Yongjie Xu, Jun Xia, Siyuan Li, Stan Z. Li† +AI Lab, Research Center for Industries of the Future, Westlake University + +{tancheng,gaozhangyang,wulirong,xuyongjie,xiajun, lisiyuan,stan.zq.li}@westlake.edu.cn + +# Abstract + +Spatiotemporal predictive learning aims to generate future frames by learning from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame features and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recurrent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unparallelizable architectures. To parallelize the temporal module, we propose the Temporal Attention Unit (TAU), which decomposes temporal attention into intra-frame statistical attention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regularization to take inter-frame variations into account. Extensive experiments demonstrate that the proposed method enables the derived model to achieve competitive performance on various spatiotemporal prediction benchmarks. + +# 1. Introduction + +The last decade has witnessed revolutionary advances in deep learning across various supervised learning tasks such as image classification [34,52,87], object detection [73,75], computational biology [21, 22, 43, 83, 84], and etc. Despite significant breakthroughs in supervised learning, which relies on large-scale labeled datasets, the potential of unsupervised learning remains largely untapped. Self-supervised learning that designs pretext tasks to produce labels derived from the data itself is recognized as a subset of unsupervised learning. In the context of self-supervised learning, contrastive self-supervised learning [7,8,28,33,85,86,92,104] predicts the noise contrastive estimation from predefined positive or negative pairs, and masked self-supervised learning [17, 32, 45, 51, 57, 102] predicts the masked patches from the visible patches. Unlike these image-level self + +supervised learning, spatiotemporal predictive learning that predicts future frames from past frames at the video-level [6, 19, 27, 46, 60, 63]. + +![](images/f5c293469b6f5e74b1e6033206984d84ea94bb1d9abdb072ddb8dcd75b091018.jpg) +(a) General framework + +![](images/fc32ec9ad6395a290747fab8c4f9d3486a90910666608caf856e2c2e0e4ed548.jpg) +(b) ConvLSTM + +![](images/cb3ddf32ff7706408aded2e7f89c9197dafd68e94b22f48dd6334c661798693a.jpg) +(c) E3D-LSTM + +![](images/ace6ca9948b95a6cfff099dc3d0a384e61dd6b699131d9afd0472a6cccc7c3a4.jpg) +(d) PhyDNet + +![](images/f3a4ea7103d5b6672904d6f2093cf30ca3cf09d98405a5a41ad14688cba1a865.jpg) +(e) CrevNet + +![](images/76b0760d4c0fd277c48ba60c4c1d8a34100c53c2f95af3c0a8a1e7700057ce87.jpg) +(f) Ours +Figure 1. Comparison of model architectures among common spatiotemporal predictive learning methods. Note that we denote 2D convolutional neural networks as Conv while 3D Conv means 3D convolutional neural networks. + +Accurate spatiotemporal predictive learning can benefit broad practical applications in climate change [74, 77], human motion forecasting [91, 107], traffic flow prediction [18, 97], and representation learning [39, 71]. The significance of spatiotemporal predictive learning primarily lies in its potential of exploring both spatial correlation and temporal evolution in the physical world. Moreover, the self-supervised nature of spatiotemporal predictive learning aligns well with human learning styles without a large amount of labeled data. Massive videos can provide a rich source of visual information, enabling spatiotemporal pre + +dictive learning to serve as a generative pre-training strategy [35, 60] for feature representation learning towards diverse downstream visual supervised tasks. + +Most of the existing methods [2, 3, 9, 11, 12, 19, 24, 29, 41, 44, 65, 78, 79, 82, 88, 89, 93–96, 98, 100, 103] in spatiotemporal predictive learning employ hybrid architectures of convolutional neural networks and recurrent units in which spatial correlation and time evolution can be learned, respectively. Inspired by the success of long short-term memory (LSTM) [36] in sequential modeling, ConvLSTM [78] is a seminal work on the topic of spatiotemporal predictive learning that extends fully connected LSTM to convolutional LSTM towards accurate precipitation nowcasting. PredRNN [95] is an admirable work that proposes Spatiotemporal LSTM (ST-LSTM) units to model spatial appearances and temporal variations in a unified memory pool. This work provides insights on designing typical recurrent units for spatiotemporal predictive learning and inspires a series of subsequent works [4, 93, 96, 98]. E3D-LSTM [94] integrates 3D convolutional neural networks into recurrent units towards good representations with both short-term frame dependencies and long-term high-level relations. PhyDNet [29] introduces a two-branch architecture that involves physical-based PhyCells and ConvLSTMs for performing partial differential equation constraints in latent space. CrevNet [103] proposes an invertible two-way autoencoder based on flow [14, 15] and a conditionally reversible architecture for spatiotemporal predictive learning. As shown in Figure 1 (a), we present a general framework consisting of the spatial encoder/decoder and the middle temporal module, abstracted from these methods. Though these spatiotemporal predictive learning methods have different temporal modules and spatial encoders/decoders, they basically share a similar framework. + +Based on the general framework, we argue that the temporal module plays an essential role in spatiotemporal predictive learning. While the common choice of the temporal module is the recurrent-based units, we explore a novel parallelizable attention module named Temporal Attention Unit (TAU) to capture time evolution. The proposed temporal attention is decomposed into intra-frame statistical attention and inter-frame dynamical attention. Furthermore, we argue that the mean square error loss only focuses on intra-frame differences, and we propose a differential divergence regularization that also cares about the inter-frame variations. Keeping the spatial encoder and decoder as simple as 2D convolutional neural networks, we deliberately implement our proposed TAU modules and surprisingly find the derived model achieves competitive performance as those recurrent-based models. This observation provides a new perspective to improve spatiotemporal predictive learning by parallelizable attention networks instead of commonly used recurrent units. + +We conduct experiments on various datasets with different experimental settings: (1) Standard spatiotemporal predictive learning. Quantitative results on various datasets demonstrate our proposed method can achieve competitive performance on standard spatiotemporal predictive learning. (2) Generalization ability. To verify the generalization ability, we train our model on KITTI and test it on the Caltech Pedestrian dataset with different domains. (3) Predict future frames with flexible lengths. We tackle the long-length frames by feeding the predicted frames as the input and find the performance is consistently well. Through the superior performance in the above three experimental settings, we demonstrate that our proposed model can provide a novel manner that learns temporal dependencies without recurrent units. + +# 2. Related works + +# 2.1. Self-supervised learning + +Deep learning has been well developed and applied in various fields [5, 55, 56, 109, 110]. Learning from massive data enables tremendous progress in supervised learning. By designing pretext tasks and generating labels from the data itself, self-supervised learning obtains supervisory signals. The model learns valuable representations by solving pretext tasks that leverage the underlying structure of the data. Early works on visual self-supervised learning design pretext tasks such as colorization [108], inpainting [69], rotation [26], jigsaw [66]. Contrastive self-supervised learning [7, 8, 28, 33, 92, 104] is a dominant manner in visual self-supervised learning that aims at a pretext task of grouping similar samples closer and diverse samples away from each other. However, contrastive self-supervised learning is limited by making pairs by multiple images, which affects its ability on small-scale datasets. Masked self-supervised learning [17, 32, 45, 51, 57, 102], which predicts the masked patches from the visible ones, is another research direction. Although masked pretraining has great success in natural language processing, its applications in visual tasks are challenging. Spatiotemporal predictive learning is another promising branch of self-supervised learning that focuses on video-level information and predicts future frames conditioned on past frames. + +In contrast to the above image-level methods, spatiotemporal predictive learning focus on video-level information and predicts future frames conditioned on past frames. By learning the intrinsic motion dynamics, the model is enabled to easily decouple the foreground and background. + +# 2.2. Spatiotemporal predictive learning + +Recurrent models have achieved remarkable advances in spatiotemporal predictive learning. Inspired by recurrent neural networks, VideoModeling [62] adopts language + +modeling and quantizes the image patches into a large dictionary for recurrent units. ConvLSTM [78] leverages convolutional neural networks to model the LSTM architecture. PredNet [61] persistently predicts future video frames using deep recurrent convolutional neural networks with bottom-up and top-down connections. PredRNN [95] proposes a Spatiotemporal LSTM unit that extracts and memorizes spatial and temporal representations simultaneously, and its following work $\mathrm{PredRNN + + }$ [93] further introduces gradient highway unit and Casual LSTM to adaptively capture temporal dependencies. E3D-LSTM [94] designs e-detic memory transition in recurrent convolutional units. Conv-TT-LSTM [82] employs a higher-order ConvLSTM to predict by combining convolutional features across time. MotionRNN [100] focuses on motion trends and transient variations. LMC-Memory [50] introduces a long-term motion context memory using memory alignment learning. PredRNN-v2 [96] extends PredRNN by leveraging a memory decoupling loss and curriculum learning strategy. + +Instead of using recurrent-based methods that are computationally expensive for spatiotemporal predictive learning, we introduce TAU, a model that uses visual attention mechanism to parallelize the temporal evolution without the recurrent structure. There are prior arts that have some similarities with our proposed model. PredCNN [101] and TrajectoryCNN [54] implement pure convolutional neural networks as the temporal module. SimVP [23] is a seminal work that applies blocks of Inception modules with a UNet architecture to learn the temporal evolution. Though their temporal modules are parallelizable, we argue that convolutions alone cannot capture long-term dependencies. Moreover, SimVP provides a simple baseline with minor complex attachment but a large space for further improvements. In general, SimVP first downsamples video sequences to reduce the computation, then uses Inception-UNet to learn essential spatiotemporal relationships, and upsamples the representations to predict future frames. Our work aims to replace the pivotal Inception-UNet with efficient attention modules that promote prediction performance. In this paper, we employ a simple yet effective attention mechanism to enable the temporal module not only to be parallelizable but also to capture long-term time evolution. + +# 3. Methods + +# 3.1. Preliminaries + +We formally define the spatiotemporal predictive learning problem as follows. Given a video sequence $\mathcal{X}^{t,T} = \{\mathbf{x}^i\}_{t - T + 1}^t$ at time $t$ with the past $T$ frames, we aim to predict the subsequent $T^{\prime}$ frames $\mathcal{Y}^{t + 1,T^{\prime}} = \{\mathbf{x}^{i}\}_{t + 1}^{t + 1 + T^{\prime}}$ from time $t + 1$ , where $\mathbf{x}_i\in \mathbb{R}^{C\times H\times W}$ is usually an image with channels $C$ , height $H$ , and width $W$ . In practice, we represent the video sequences as tensors, i.e., $\mathcal{X}^{t,T}\in$ + +$$ +\mathbb {R} ^ {T \times C \times H \times W} \text {a n d} \mathcal {Y} ^ {t + 1, T ^ {\prime}} \in \mathbb {R} ^ {T ^ {\prime} \times C \times H \times W}. +$$ + +The model with learnable parameters $\Theta$ learns a mapping $\mathcal{F}_{\Theta}:\mathcal{X}^{t,T}\mapsto \mathcal{Y}^{t + 1,T^{\prime}}$ by exploring both spatial and temporal dependencies. In our case, the mapping $\mathcal{F}_{\Theta}$ is a neural network model trained to minimize the difference between the predicted future frames and the ground-truth future frames. The optimal parameters $\Theta^{*}$ are: + +$$ +\Theta^ {*} = \arg \min _ {\Theta} \mathcal {L} \left(\mathcal {F} _ {\Theta} \left(\mathcal {X} ^ {t, T}\right), \mathcal {Y} ^ {t + 1, T ^ {\prime}}\right), \tag {1} +$$ + +where $\mathcal{L}$ is a loss function that evaluates such differences. + +# 3.2. Overview + +We illustrate the overview model in Figure 2 using the input Moving MNIST [81] data as an example. + +![](images/fd3a7222ad8ab4804a5e072385ff3488a7ef65ca766d859d2da2b2e2f284ffbd.jpg) +Figure 2. The overview architecture of our proposed model. + +Striving for simplicity, the model follows the general framework in Figure 1, while the spatial encoder consists of four vanilla 2D convolutional layers, and the spatial decoder consists of four 2D transposed convolutional layers ('Conv2d' and 'ConvTranspose2d' in PyTorch respectively). We add a residual connection from the first convolutional layer to the last transposed convolutional layer for preserving the spatial-dependent features. Stacks of TAU modules are in the middle of the spatial encoder and decoder to extract temporal-dependent features. Though our model is simple, it can efficiently learn both spatial-dependent and temporal-dependent features without recurrent architectures. + +# 3.3. Temporal Attention Unit + +Suppose a batch of input video tensors $\mathcal{B} \in \mathbb{R}^{B \times T \times C \times H \times W}$ with the number of videos $B = |\mathcal{B}|$ is given. In the spatial encoder and decoder, we reshape the sequential input data $B \times T \times C \times H \times W$ as $(B \times T) \times C \times H \times W$ so that only spatial correlations are taken into account. In the temporal module, we reshape the feature $B \times T \times C \times H \times W$ as $B \times (T \times C) \times H \times W$ so that frames are arranged in order on the channel dimension. + +We decompose the temporal attention into the intra-frame statical attention and the inter-frame dynamical attention, as shown in Figure 3. Inspired by the recent progress of vision Transformers (ViTs) [17,57] and large kernel convolutions [13, 30, 58], we propose to employ small kernel depth-wise convolutions (DW Conv), depth-wise convolutions with dilations (DW-D Conv), and $1 \times 1$ convolutions to model the large kernel convolutions. Through the obtained large receptive field on intra-frames, the statical attention is able to capture long-range dependencies. However, the statical attention alone is not enough for learning temporal evolutions along the timeline. Thus, we employ the dynamical attention that learns the attention weights of channels in a squeeze-and-excitation manner [38]. The final attention is the product of dynamical attention and statical attention. + +![](images/a61ec68faafa29d95fe56d8da580d234f0fbb960421c6c20f0b20b4824330584.jpg) +Figure 3. The intra-frame statical attention and the inter-frame dynamical attention. + +We show the detailed scheme of our model in Figure 4. The proposed TAU module can be formally expressed as: + +$$ +\mathrm {S A} = \operatorname {C o n v} _ {1 \times 1} (\mathrm {D W - D C o n v} (\mathrm {D W C o n v} (H))), +$$ + +$$ +\mathrm {D A} = \mathrm {F C} (\operatorname {A v g P o o l} (H)), \tag {2} +$$ + +$$ +H ^ {\prime} = (\mathrm {S A} \otimes \mathrm {D A}) \odot H, +$$ + +where $H \in \mathbb{R}^{B \times (T \times C') \times H \times W}$ is the hidden feature that will be fed into the TAU module, SA $\in \mathbb{R}^{B \times (T \times C') \times H \times W}$ , DA $\in \mathbb{R}^{B \times (T \times C') \times 1 \times 1}$ denote the statistical and dynamical attention, FC and AvgPool are fully connected layers and the average pooling. We represent the Kronecker product by $\otimes$ and the Hadamard product by $\odot$ . + +# 3.4. Differential Divergence Regularization + +To improve the prediction of our model, we further propose a differential divergence regularization that forces the model to learn the differences between consecutive frames and be aware of the inherent variation. + +Given the predicted frames $\widehat{\mathcal{Y}} = \mathcal{F}_{\Theta}(\mathcal{X}) \in \mathbb{R}^{T' \times C \times H \times W}$ and its corresponding ground-truth frames + +![](images/d583ca0e0b0e36fbf267a83ae0f4fcce757dd20b4d820959f029804b5c5c8add.jpg) +Figure 4. The detailed schema of our model. + +$\mathcal{V}$ , we first calculate their forward difference $\Delta \widehat{\mathcal{Y}}, \Delta \mathcal{Y} \in \mathbb{R}^{(T' - 1) \times C \times H \times W}$ , where: + +$$ +\begin{array}{l} \Delta \widehat {\mathcal {Y}} _ {i} = \widehat {\mathcal {Y}} _ {i + 1} - \widehat {\mathcal {Y}} _ {i}, \tag {3} \\ \Delta \mathcal {Y} _ {i} = \mathcal {Y} _ {i + 1} - \mathcal {Y} _ {i}. \\ \end{array} +$$ + +Then, we transform the differences into probabilities by the softmax function on the channel, height, and width dimension and obtain $\sigma (\Delta \widehat{\mathcal{Y}}),\sigma (\Delta \mathcal{Y})$ , where: + +$$ +\begin{array}{l} \sigma (\Delta \widehat {\mathcal {Y}}) _ {i, j, k, l} = \frac {\exp \left(\Delta \widehat {\mathcal {Y}} _ {i , j , k , l} / \tau\right)}{\sum_ {j ^ {\prime} = 1} ^ {C} \sum_ {k ^ {\prime} = 1} ^ {H} \sum_ {l ^ {\prime} = 1} ^ {W} \exp \left(\Delta \widehat {\mathcal {Y}} _ {i , j ^ {\prime} , k ^ {\prime} , l ^ {\prime}} / \tau\right)}, \\ \sigma \left(\Delta \mathcal {Y}\right) _ {i, j, k, l} = \frac {\exp \left(\Delta \mathcal {Y} _ {i , j , k , l}\right) / \tau}{\sum_ {j ^ {\prime} = 1} ^ {C} \sum_ {k ^ {\prime} = 1} ^ {H} \sum_ {l ^ {\prime} = 1} ^ {W} \exp \left(\Delta \mathcal {Y} _ {i , j ^ {\prime} , k ^ {\prime} , l ^ {\prime}} / \tau\right)}, \tag {4} \\ \end{array} +$$ + +and $\tau$ represents the temperature parameter which we empirically set as 0.1 to sharpen the probability distribution. Through the competition mechanism of the softmax function [10, 70, 99], the high difference frames are penalized. + +Thus, the differential divergence regularization $\mathcal{L}_{reg}$ is defined as the Kullback-Leibler divergence between the probability distributions $\sigma (\Delta \widehat{\mathcal{V}})$ and $\sigma (\Delta \mathcal{V})$ : + +$$ +\begin{array}{l} \mathcal {L} _ {r e g} (\widehat {\mathcal {Y}}, \mathcal {Y}) = D _ {K L} (\sigma (\Delta \widehat {\mathcal {Y}}) \mid \mid \sigma (\Delta \mathcal {Y})) \\ = \sum_ {i = 1} ^ {T ^ {\prime} - 1} \sigma \left(\Delta \hat {\mathcal {Y}} _ {i}\right) \log \frac {\sigma \left(\Delta \hat {\mathcal {Y}} _ {i}\right)}{\sigma \left(\Delta \mathcal {Y} _ {i}\right)}. \tag {5} \\ \end{array} +$$ + +Our model is trained end-to-end in a fully unsupervised manner, and the overall objective function consists of the mean square error loss and the differential divergence regularization weighted by a constant $\alpha$ : + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {T ^ {\prime}} \| \widehat {\mathcal {Y}} - \mathcal {Y} \| ^ {2} + \alpha \mathcal {L} _ {r e g} (\widehat {\mathcal {Y}}, \mathcal {Y}), \tag {6} +$$ + +where the first term focuses on intra-frame-level differences, and the second regularization term focuses on inter-frame-level variations. + +# 4. Experiments + +In this section, we present experiments that demonstrate the effectiveness of our proposed method. The experiments are conducted on various datasets with different settings to validate our proposed model from three aspects: + +- Standard spatiotemporal predictive learning (Section 4.2). We recognize the prediction problem of the same number of input and output frames as the standard spatiotemporal predictive learning. We evaluate the performance on standard spatiotemporal predictive learning and compare our model with state-of-the-art methods with Moving MNIST [81] and TaxiBJ [106] datasets. +- Generalization ability across different datasets (Section 4.3). Generalizing the learned knowledge to other domains is a challenge in unsupervised learning. We investigate the such ability of our method by training the model on the KITTI [25] dataset and evaluating it on the Caltech Pedestrian [16] dataset. +- Predicting frames with flexible lengths (Section 4.4). One of the advantages of recurrent units is that they can easily handle flexible-length frames like the KTH dataset [76]. Our work tackles the long-length frame prediction by imitating recurrent units that feed predicted frames as the input and recursively produce long-term predictions. + +# 4.1. Experimental Setups + +Datasets We quantitatively evaluate our model on the following datasets for both synthetic and real-world scenarios: + +- Moving MNIST [81] is a synthetic dataset consisting of two digits independently moving within the $64 \times 64$ grid and bouncing off the boundary. It is a standard benchmark in spatiotemporal predictive learning. +- TaxiBJ contains the trajectory data in Beijing collected from taxicab GPS with two channels, i.e., inflow or outflow defined in [106]. Following the previous works [29, 98], we normalize the data into $[0, 1]$ . +- KTH [76] contains 25 individuals performing six types of actions. Following [90, 94], we use person 1-16 for training and 17-25 for testing. Models are trained to predict the next 20 or 40 frames from the previous 10 observations. +- Caltech Pedestrian is a driving dataset focusing on detecting pedestrians. It consists of approximately 10 hours of $640 \times 480$ videos taken from vehicles driving through regular traffic in an urban environment. We follow the same protocol of PredNet [61] and + +CrevNet [103] for pre-processing, training, and evaluation. + +We summarize the statistics of the above datasets in Table 1, including the number of training samples $N_{train}$ and the number of testing samples $N_{test}$ . + +Table 1. The statistics of datasets. The training or testing set has $N_{train}$ or $N_{test}$ samples, composed by $T$ or $T'$ images with the shape $(C, H, W)$ . + +
NtrainNtest(C,H,W)TT'
MMNIST1000010000(1,64,64)1010
TaxiBJ196271334(2,32,32)44
KTH52003167(1,128,128)1020 or 40
Caltech20421983(3,128,160)101
+ +Measurement Following [29, 103], we employ Mean Squared Error (MSE), Mean Absolute Error (MAE), Structure Similarity Index Measure (SSIM), and Peak Signal to Noise Ratio (PSNR) to evaluate the quality of predictions. MSE and MAE estimate the absolute pixel-wise errors, SSIM measures the similarity of structural information within the spatial neighborhoods, and PSNR is an expression for the ratio between the maximum possible power of a signal and the power of distorted noise. + +Implementation details We implement the proposed method with the Pytorch framework and conduct experiments on a single NVIDIA-V100 GPU. The model is trained with a mini-batch of 16 video sequences while the AdamW optimizer is utilized with a learning rate of 0.01 and a weight decay of 0.05. + +# 4.2. Standard spatiotemporal predictive learning + +# 4.2.1 Moving MNIST + +This dataset is a standard benchmark in spatiotemporal predictive learning. We evaluate our proposed method against strong recent baselines, including competitive recurrent architectures: ConvLSTM [78], PredRNN [95], PredRNN++ [93], MIM [98], LMC [50], E3D-LSTM [94], Conv-TT-LSTM [82], and CrevNet [103]. We also compare the method DDPAE [37] that is specifically designed for this dataset. The quantitative results are reported in Table 2, and qualitative visualizations of the predicted results are shown in Figure 5. + +Our proposed method significantly outperforms all the baselines above under three different metrics. The performance gain is large with respect to state-of-the-art recurrent methods. + +![](images/fbe5d313999c722cd256d67932b735dd8ce1ed7c7423bb70822b0b58353fb7bc.jpg) +Figure 5. Qualitative visualization of predicted results on Moving MNIST dataset. The differences between the ground truth and the predicted frames are visualized in the last row. + +Table 2. Quantitative results of different methods on the Moving MNIST dataset (10 → 10 frames). + +
MethodMoving MNIST
MSE↓MAE↓SSIM↑
ConvLSTM [78]103.3182.90.707
VPN [44]64.1-0.870
PredRNN [95]56.8126.10.867
PredRNN++ [93]46.5106.80.898
MIM [98]44.2101.10.910
LMC [50]41.5-0.924
E3D-LSTM [94]41.387.20.910
Conv-TT-LSTM [82]53.0-0.915
DDPAE [37]38.990.70.922
PhyDNet [29]24.470.30.947
SimVP [23]23.868.90.948
Crevnet [103]22.3-0.949
Ours19.860.30.957
+ +# 4.2.2 TaxiBJ + +We evaluate our proposed model on a complicated real-world dataset, TaxiBJ [106]. Driven by human consciousness, the complex real-world traffic flows requires modeling transport phenomena and traffic diffusion for prediction. Due to the spatiotemporal nature of the traffic forecasting task, we straightforwardly implement our model for it. + +The qualitative visualizations of the predicted results are shown in Figure 6, and the quantitative results are reported in Table 3. Though the given frames are quite different from the future frames, our model can still accurately produce reliable frames. The difference between target frames + +![](images/199de6d2bec999db32824ca989a31c8f958426e0847ec35051c8eef29d2791aa.jpg) +Figure 6. Qualitative visualization of predicted results on TaxiBJ dataset. The differences between the ground truth and the predicted frames are visualized in the last row. + +and predicted frames is mainly located in central spots, but the overall trend is approximating the ground truth. It can also be observed that the quantitative results are consistently well, which demonstrates the practical application value in real-world scenarios. + +# 4.3. Generalization across different datasets + +The generalization ability is one of the fundamental problems in artificial intelligence technology. Traditional supervised learning suffers from its poor generalization of + +Table 3. Quantitative results of different methods on the TaxiBJ dataset (4 → 4 frames). + +
MethodTaxiBJ
MSE × 100↓MAE↓SSIM↑
ConvLSTM [78]48.517.70.978
PredRNN [95]46.417.10.971
PredRNN++ [93]44.816.90.977
MIM [98]42.916.60.971
E3D-LSTM [94]43.216.90.979
PhyDNet [29]41.916.20.982
SimVP [23]41.416.20.982
Ours34.415.60.983
+ +labeled datasets with different domains. Self-supervised learning aims to learn robust representations based on unlabeled data and evaluates the generalization ability of the learned model. While contrastive self-supervised learning and masked self-supervised learning in visual tasks usually evaluate such generalization ability by downstream tasks, we evaluate it by the prediction results across different datasets in spatiotemporal predictive learning. + +![](images/43795de1a2785063eb18e369c9dfc71db7ba3927c12e26cea33c646cd5a2d33b.jpg) +Figure 7. Qualitative visualization of predicted results on Caltech dataset. The differences between the ground truth and the predicted frames are visualized in the last row. + +Following the previous works [53, 61, 103], we train the model using the raw video sequences from the KITTI dataset and evaluate the model by Caltech Pedestrian dataset that is made to match the frame rate and image size $(128 \times 160)$ of the KTTI dataset. The final prediction is made for the next frame after a 10-frame warm-up. + +We show the qualitative visualization in Figure 7 and re + +port the quantitative results in Table 4. Note that some of the baseline results are copied from [68]. It can be seen that our proposed method achieves state-of-the-art performance under both SSIM and PSNR metrics in this generalization evaluation task. Moreover, our model shows significantly robust predictions in terms of the variation of illumination and lane line, suggesting its practical applications in autonomous vehicles. + +Table 4. Quantitative results of different methods on the Caltech Pedestrian dataset (10 → 1 frame). + +
MethodCaltech Pedestrian
SSIM↑PSNR↑
BeyondMSE [64]0.847-
MCnet [90]0.879-
DVF [59]0.89726.2
Dual-GAN [53]0.899-
CtrlGen [31]0.90026.5
PredNet [61]0.90527.6
ContextVP [4]0.92128.7
GAN-VGG [80]0.916-
G-VGG [80]0.917-
SDC-Net [72]0.918-
rCycleGan [47]0.91929.2
DPG [20]0.92328.2
G-MAE [80]0.923-
GAN-MAE [80]0.923-
CrevNet [103]0.92529.3
STMFANet [41]0.92729.1
SimVP [23]0.94033.1
Ours0.94633.7
+ +# 4.4. Predicting frames with flexible lengths + +Though recurrent units are adept at handling flexible-length frames, our model can also easily tackle such problems by imitating recurrent units that feed predicted frames as the input and recursively produce predictions. For this KTH dataset, our model is trained to predict the next 20 or 40 frames from the given 10 observations. Moreover, this dataset contains six types of human actions (walking, jogging, running, boxing, hand waving and hand clapping) performed several times by 25 subjects in four different scenarios: outdoors, outdoors with scale variations, outdoors with different clothes and indoors. The difficulty of this human motion prediction task not only lies in its flexible lengths of predicted frames but also in its complex dynamics involving the randomness of human consciousness. Following [95], we use the Peak Signal Noise Ratio (PSNR) and the Structural Similarity Index Measure (SSIM) as evaluation metrics to measure the framewise prediction quality from the perceptive view. The detailed quantitative results are shown in Table 5. + +Table 5. Quantitative results of different methods on the KTH dataset (10 → 20 and 10 → 40 frames). + +
MethodKTH (10 → 20)KTH (10 → 40)
SSIM↑PSNR↑SSIM↑PSNR↑
MCnet [90]0.80425.950.7323.89
ConvLSTM [78]0.71223.580.63922.85
SAVP [48]0.74625.380.70123.97
VPN [44]0.74623.76--
DFN [40]0.79427.260.65223.01
fRNN [67]0.77126.120.67823.77
Znet [105]0.81727.58--
SV2Pv [1]0.83827.790.78926.12
PredRNN [95]0.83927.550.70324.16
VarNet [42]0.84328.480.73925.37
SAVP-VAE [48]0.85227.770.81126.18
PredRNN++ [93]0.86528.470.74125.21
MSNET [49]0.87627.08--
E3d-LSTM [94]0.87929.310.81027.24
STMFANet [41]0.89329.850.85127.56
SimVP [23]0.90533.720.88632.93
Ours0.91134.130.89733.01
+ +# 4.5. Empirical Running Time + +TAU benefits from the parallelizable computation architecture, which leads to fast convergence and high training speed. We evaluate our efficiency by measuring the running time against state-of-the-art spatiotemporal predictive learning methods. + +![](images/732999e3311e35bd1909c6417231d248f792a2d454ae85fc70f22e594ffad014.jpg) +Figure 8. The learning curve comparison between state-of-the-art methods and ours (evaluated by MSE). The red dotted line denotes MSE 35.0, and only the first 50 epochs are shown. + +The experiments are conducted on a single Tesla V100 GPU, and the batch size is set as 16. For the Moving MNIST dataset, CrevNet [103] needs about 30 minutes per epoch, and PhyDNet [29] needs about 7 minutes per epoch. Our model only requires 2.5 minutes per epoch. Furthermore, our method is able to converge at a rapid rate. As shown in Figure 8, on the Moving MNIST dataset, our model can achieve MSE 35.0 with only 50 epochs, while CrevNet and PhyDNet are far from this performance. + +# 4.6. Computational Cost and Ablation Study + +We compare the performance and computational cost with state-of-the-art methods in the first several rows in Table 6. Our model achieves superior results with better performance and much lower Flops. We also conduct ablation studies and summarize the results in Table 6. It can be seen that replacing TAU with the same number of convolutional blocks with vanilla $3 \times 3$ convolutions (Conv Baseline) significantly degrades the performance. Training our model without differential divergence regularization (w/o DDR) will also weaken the prediction results. Both the TAU module and differential divergence regularization are useful. We also find that SA and DA of the TAU module play important roles in better performance. + +Table 6. Ablation study of our proposed method. + +
MethodM-MNISTTaxiBJ (×100)Flops(G)
PredRNN56.846.4115.94
PredRNN++46.444.8171.73
MIM44.242.9179.17
E3DLSTM41.343.2298.85
SimVP23.841.419.43
Conv Baseline58.943.56.11
Ours w/o SA23.240.815.30
Ours w/o DA22.438.415.96
Ours w/o DDR21.137.715.96
Ours19.834.415.96
+ +# 5. Conclusion + +In this paper, we present a general framework of spatiotemporal predictive learning and propose an attention-based temporal module to replace the common-used recurrent units. By decomposing the temporal module into the intra-frame statistical attention and the inter-frame dynamical attention, our proposed TAU module can achieve competitive performance across various experimental settings and datasets. Moreover, a novel differential divergence regularization is proposed to overcome the drawback of MSE loss that only considers the intra-frame error. In summary, our work highlights the importance of both intra-frame and inter-frame variations that enable the model to capture long-term relations and provide a new paradigm of efficient spatiotemporal predictive learning. + +Acknowledgement. We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (2022ZD0115100), the National Natural Science Foundation of China (U21A20427), the Competitive Research Fund (WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering + +# References + +[1] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. In International Conference on Learning Representations, 2018. 8 +[2] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. Fitvid: Overfitting in pixel-level video prediction. arXiv preprint arXiv:2106.13195, 2021. 2 +[3] Sarthak Bhagat, Shagun Uppal, Zhuyun Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In European Conference on Computer Vision, pages 102-117. Springer, 2020. 2 +[4] Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava, and Petros Koumoutsakos. Contextv: Fully context-aware video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 753-769, 2018. 2, 7 +[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2 +[6] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7608-7617, 2019. 1 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2 +[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in Neural Information Processing Systems, 33:22243-22255, 2020. 1, 2 +[9] Yu Cheng, Bo Wang, Bo Yang, and Robby T Tan. Graph and temporal convolutional networks for 3d multi-person pose estimation in monocular videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 4, page 12, 2021. 2 +[10] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. In International Conference on Learning Representations, 2020. 4 +[11] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In International Conference on Machine Learning, pages 1174-1183. PMLR, 2018. 2 +[12] Emily L Denton et al. Unsupervised learning of disentangled representations from video. Advances in Neural Information Processing Systems, 30, 2017. 2 +[13] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. arXiv preprint arXiv:2203.06717, 2022.4 + +[14] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. In International Conference on Learning Representations Workshop, 2015. 2 +[15] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 2 +[16] Piotr Dólar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark. In 2009 IEEE conference on computer vision and pattern recognition, pages 304-311. IEEE, 2009. 5 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 4 +[18] Shen Fang, Qi Zhang, Gaofeng Meng, Shiming Xiang, and Chunhong Pan. GSTnet: Global spatial-temporal network for traffic flow prediction. In IJCAI, pages 2286-2293, 2019. 1 +[19] Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through video prediction. Advances in Neural Information Processing Systems, 29, 2016. 1, 2 +[20] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7 +[21] Zhangyang Gao, Cheng Tan, Stan Li, et al. Alphadesign: A graph protein design method and benchmark on alphafolddb. arXiv preprint arXiv:2202.01079, 2022. 1 +[22] Zhangyang Gao, Cheng Tan, and Stan Z Li. Pifold: Toward effective and efficient protein inverse folding. arXiv preprint arXiv:2209.12643, 2022. 1 +[23] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z. Li. Simvp: Simpler yet better video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3170-3180, June 2022. 3, 6, 7, 8 +[24] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, pages 1243-1252. PMLR, 2017. 2 +[25] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5 +[26] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2 +[27] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthew, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative varia + +tional inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019. 1 +[28] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2 +[29] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 2, 5, 6, 7, 8 +[30] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 4 +[31] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018. 7 +[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2 +[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 1, 2 +[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1 +[35] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. arXiv preprint arXiv:2102.12627, 2021.2 +[36] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2 +[37] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li F Fei-Fei, and Juan Carlos Niebles. Learning to decompose and disentangle representations for video prediction. Advances in Neural Information Processing Systems, 31, 2018. 5, 6 +[38] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4 +[39] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision, pages 425-442. Springer, 2020. 1 +[40] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 8 + +[41] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 2, 7, 8 +[42] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 8 +[43] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. Nature, 596(7873):583-589, 2021. 1 +[44] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 2, 6, 8 +[45] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1, 2 +[46] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 1 +[47] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 7 +[48] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 8 +[49] Jungbeom Lee, Jangho Lee, Sungmin Lee, and Sungroh Yoon. Mutual suppression network for video prediction using disentangled features. arXiv preprint arXiv:1804.04810, 2018. 8 +[50] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 3, 5, 6 +[51] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, 2020. 1, 2 +[52] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1 + +[53] Xiaodan Liang, Lisa Lee, Wei Dai, and Eric P Xing. Dual motion gan for future-flow embedded video prediction. In proceedings of the IEEE international conference on computer vision, pages 1744-1752, 2017. 7 +[54] Xiaoli Liu, Jianqin Yin, Jin Liu, Pengxiang Ding, Jun Liu, and Huaping Liu. TrajectoryCNN: a new spatio-temporal feature learning network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2133-2146, 2020. 3 +[55] Zicheng Liu, Siyuan Li, Ge Wang, Cheng Tan, Lirong Wu, and Stan Z Li. Decoupled mixup for data-efficient learning. arXiv preprint arXiv:2203.10761, 2022. 2 +[56] Zicheng Liu, Siyuan Li, Di Wu, Zihan Liu, Zhiyuan Chen, Lirong Wu, and Stan Z Li. Automix: Unveiling the power of mixup for stronger classifiers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 441-458. Springer, 2022. 2 +[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4 +[58] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 4 +[59] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 7 +[60] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning, pages 4114-4124. PMLR, 2019. 1, 2 +[61] William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. In International Conference on Learning Representations, 2017. 3, 5, 7 +[62] Arthur Szlam Marc'Aurelio Ranzato, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR, abs/1412.6604, 2, 2014. 2 +[63] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019. 1 +[64] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square error. arXiv preprint arXiv:1511.05440, 2015. 7 +[65] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 2 +[66] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. + +In European conference on computer vision, pages 69-84. Springer, 2016. 2 +[67] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 8 +[68] Sergiu Oprea, Pablo Martinez-Gonzalez, Alberto Garcia-Garcia, John Alejandro Castro-Vargas, Sergio Orts-Escalano, Jose Garcia-Rodriguez, and Antonis Argyros. A review on deep learning techniques for video prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7 +[69] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2 +[70] Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. Random feature attention. In International Conference on Learning Representations, 2020. 4 +[71] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 1 +[72] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 718-733, 2018. 7 +[73] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016. 1 +[74] Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, et al. Deep learning and process understanding for data-driven earth system science. Nature, 566(7743):195-204, 2019. 1 +[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28:91-99, 2015. 1 +[76] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 3, pages 32-36. IEEE, 2004. 5 +[77] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 1 +[78] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM + +network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 2, 3, 5, 6, 7, 8 +[79] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in Neural Information Processing Systems, 30, 2017. 2 +[80] Osamu Shouno. Photo-realistic video prediction on natural videos of largely changing frames. arXiv preprint arXiv:2003.08635, 2020. 7 +[81] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 3, 5 +[82] Jiahao Su, Wonmin Byeon, Jean Kossaifi, Furong Huang, Jan Kautz, and Anima Anandkumar. Convolutional tensortrain lstm for spatio-temporal learning. Advances in Neural Information Processing Systems, 33:13714-13726, 2020. 2, 3, 5, 6 +[83] Cheng Tan, Zhangyang Gao, and Stan Z Li. Rfold: Towards simple yet effectiverna secondary structure prediction.arXiv preprint arXiv:2212.14041, 2022.1 +[84] Cheng Tan, Zhangyang Gao, and Stan Z Li. Target-aware molecular graph generation. arXiv preprint arXiv:2202.04829, 2022.1 +[85] Cheng Tan, Zhangyang Gao, Lirong Wu, Siyuan Li, and Stan Z Li. Hyperspherical consistency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7244-7255, 2022. 1 +[86] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 1 +[87] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1 +[88] Ruben Villegas, Dumitru Erhan, Honglak Lee, et al. Hierarchical long-term video prediction without supervision. In International Conference on Machine Learning, pages 6038-6046. PMLR, 2018. 2 +[89] Ruben Villegas, *Arkanath Pathak*, Harini Kannan, *Dumitru Erhan*, Quoc V Le, and Honglak Lee. *High fidelity video prediction with large stochastic recurrent neural networks*. Advances in Neural Information Processing Systems, 32, 2019. 2 +[90] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. In International Conference on Learning Representations, 2017. 5, 7, 8 +[91] Pichao Wang, Wanqing Li, Philip Ogunbona, Jun Wan, and Sergio Escalera. Rgb-d-based human motion recognition with deep learning: A survey. Computer Vision and Image Understanding, 171:118-139, 2018. 1 +[92] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uni + +formity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2 +[93] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrmn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 2, 3, 5, 6, 7, 8 +[94] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 2, 3, 5, 6, 7, 8 +[95] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in Neural Information Processing Systems, 30, 2017. 2, 3, 5, 6, 7, 8 +[96] Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jianmin Wang, Philip S Yu, and Mingsheng Long. Predrnn: A recurrent neural network for spatiotemporal predictive learning. arXiv preprint arXiv:2103.09504, 2021. 2, 3 +[97] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154–9162, 2019. 1 +[98] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 2, 5, 6, 7 +[99] Haixu Wu, Jialong Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Flowformer: Linearizing transformers with conservation flows. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 24226-24242. PMLR, 17-23 Jul 2022. 4 +[100] Haixu Wu, Zhiyu Yao, Jianmin Wang, and Mingsheng Long. Motionrnn: A flexible model for video prediction with spacetime-varying motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15435-15444, 2021. 2, 3 +[101] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 3 +[102] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[103] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. In International Conference on Learning Representations, 2019. 2, 5, 6, 7, 8 +[104] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via + +redundancy reduction. arXiv preprint arXiv:2103.03230, 2021. 1, 2 +[105] Jianjin Zhang, Yunbo Wang, Mingsheng Long, Wang Jian-min, and S Yu Philip. Z-order recurrent neural networks for video prediction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 230-235. IEEE, 2019. 8 +[106] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 5, 6 +[107] Liang Zhang, Guangming Zhu, Peiyi Shen, Juan Song, Syed Afaq Shah, and Mohammed Bennamoun. Learning spatiotemporal features using 3dCNN and convolutional LSTM for gesture recognition. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 3120-3128, 2017. 1 +[108] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016. 2 +[109] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 2 +[110] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Li. Using context-to-vector with graph retrofitting to improve word embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8154-8163, 2022. 2 \ No newline at end of file diff --git a/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/images.zip b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c393d61f089d2149f63143d6d2ea7cf5afd789e8 --- /dev/null +++ b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12650c78c7a255026a6ffcc9cdf0c4d4d50ed942d89e7558044c630906d4c7f3 +size 618790 diff --git a/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/layout.json b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1c798f63def84a843a21015d7983189e42a7a50c --- /dev/null +++ b/2023/Temporal Attention Unit_ Towards Efficient Spatiotemporal Predictive Learning/layout.json @@ -0,0 +1,10205 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 101, + 504, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 101, + 504, + 118 + ], + "spans": [ + { + "bbox": [ + 89, + 101, + 504, + 118 + ], + "type": "text", + "content": "Temporal Attention Unit: Towards Efficient Spatiotemporal Predictive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 139, + 507, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 139, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 91, + 139, + 507, + 168 + ], + "type": "text", + "content": "Cheng Tan*, Zhangyang Gao*, Lirong Wu, Yongjie Xu, Jun Xia, Siyuan Li, Stan Z. Li† \nAI Lab, Research Center for Industries of the Future, Westlake University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 170, + 525, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 170, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 66, + 170, + 525, + 182 + ], + "type": "text", + "content": "{tancheng,gaozhangyang,wulirong,xuyongjie,xiajun, lisiyuan,stan.zq.li}@westlake.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 209, + 192, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 209, + 192, + 221 + ], + "spans": [ + { + "bbox": [ + 143, + 209, + 192, + 221 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 234, + 290, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 290, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 290, + 462 + ], + "type": "text", + "content": "Spatiotemporal predictive learning aims to generate future frames by learning from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame features and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recurrent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unparallelizable architectures. To parallelize the temporal module, we propose the Temporal Attention Unit (TAU), which decomposes temporal attention into intra-frame statistical attention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regularization to take inter-frame variations into account. Extensive experiments demonstrate that the proposed method enables the derived model to achieve competitive performance on various spatiotemporal prediction benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 485, + 128, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 128, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 128, + 498 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 505, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 685 + ], + "type": "text", + "content": "The last decade has witnessed revolutionary advances in deep learning across various supervised learning tasks such as image classification [34,52,87], object detection [73,75], computational biology [21, 22, 43, 83, 84], and etc. Despite significant breakthroughs in supervised learning, which relies on large-scale labeled datasets, the potential of unsupervised learning remains largely untapped. Self-supervised learning that designs pretext tasks to produce labels derived from the data itself is recognized as a subset of unsupervised learning. In the context of self-supervised learning, contrastive self-supervised learning [7,8,28,33,85,86,92,104] predicts the noise contrastive estimation from predefined positive or negative pairs, and masked self-supervised learning [17, 32, 45, 51, 57, 102] predicts the masked patches from the visible patches. Unlike these image-level self" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 209, + 545, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 545, + 246 + ], + "type": "text", + "content": "supervised learning, spatiotemporal predictive learning that predicts future frames from past frames at the video-level [6, 19, 27, 46, 60, 63]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 255, + 376, + 364 + ], + "blocks": [ + { + "bbox": [ + 312, + 255, + 376, + 364 + ], + "lines": [ + { + "bbox": [ + 312, + 255, + 376, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 255, + 376, + 364 + ], + "type": "image", + "image_path": "f5c293469b6f5e74b1e6033206984d84ea94bb1d9abdb072ddb8dcd75b091018.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 370, + 389, + 381 + ], + "lines": [ + { + "bbox": [ + 312, + 370, + 389, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 370, + 389, + 381 + ], + "type": "text", + "content": "(a) General framework" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 399, + 255, + 460, + 364 + ], + "blocks": [ + { + "bbox": [ + 399, + 255, + 460, + 364 + ], + "lines": [ + { + "bbox": [ + 399, + 255, + 460, + 364 + ], + "spans": [ + { + "bbox": [ + 399, + 255, + 460, + 364 + ], + "type": "image", + "image_path": "fc32ec9ad6395a290747fab8c4f9d3486a90910666608caf856e2c2e0e4ed548.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 370, + 456, + 380 + ], + "lines": [ + { + "bbox": [ + 402, + 370, + 456, + 380 + ], + "spans": [ + { + "bbox": [ + 402, + 370, + 456, + 380 + ], + "type": "text", + "content": "(b) ConvLSTM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 485, + 255, + 545, + 364 + ], + "blocks": [ + { + "bbox": [ + 485, + 255, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 485, + 255, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 485, + 255, + 545, + 364 + ], + "type": "image", + "image_path": "cb3ddf32ff7706408aded2e7f89c9197dafd68e94b22f48dd6334c661798693a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 371, + 542, + 381 + ], + "lines": [ + { + "bbox": [ + 488, + 371, + 542, + 381 + ], + "spans": [ + { + "bbox": [ + 488, + 371, + 542, + 381 + ], + "type": "text", + "content": "(c) E3D-LSTM" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 313, + 391, + 376, + 497 + ], + "blocks": [ + { + "bbox": [ + 313, + 391, + 376, + 497 + ], + "lines": [ + { + "bbox": [ + 313, + 391, + 376, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 376, + 497 + ], + "type": "image", + "image_path": "ace6ca9948b95a6cfff099dc3d0a384e61dd6b699131d9afd0472a6cccc7c3a4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 501, + 364, + 512 + ], + "lines": [ + { + "bbox": [ + 320, + 501, + 364, + 512 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 364, + 512 + ], + "type": "text", + "content": "(d) PhyDNet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 399, + 392, + 460, + 497 + ], + "blocks": [ + { + "bbox": [ + 399, + 392, + 460, + 497 + ], + "lines": [ + { + "bbox": [ + 399, + 392, + 460, + 497 + ], + "spans": [ + { + "bbox": [ + 399, + 392, + 460, + 497 + ], + "type": "image", + "image_path": "f3a4ea7103d5b6672904d6f2093cf30ca3cf09d98405a5a41ad14688cba1a865.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 502, + 449, + 512 + ], + "lines": [ + { + "bbox": [ + 408, + 502, + 449, + 512 + ], + "spans": [ + { + "bbox": [ + 408, + 502, + 449, + 512 + ], + "type": "text", + "content": "(e) CrevNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 485, + 391, + 545, + 497 + ], + "blocks": [ + { + "bbox": [ + 485, + 391, + 545, + 497 + ], + "lines": [ + { + "bbox": [ + 485, + 391, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 485, + 391, + 545, + 497 + ], + "type": "image", + "image_path": "76b0760d4c0fd277c48ba60c4c1d8a34100c53c2f95af3c0a8a1e7700057ce87.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 502, + 528, + 512 + ], + "lines": [ + { + "bbox": [ + 500, + 502, + 528, + 512 + ], + "spans": [ + { + "bbox": [ + 500, + 502, + 528, + 512 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 525, + 547, + 569 + ], + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 569 + ], + "type": "text", + "content": "Figure 1. Comparison of model architectures among common spatiotemporal predictive learning methods. Note that we denote 2D convolutional neural networks as Conv while 3D Conv means 3D convolutional neural networks." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Accurate spatiotemporal predictive learning can benefit broad practical applications in climate change [74, 77], human motion forecasting [91, 107], traffic flow prediction [18, 97], and representation learning [39, 71]. The significance of spatiotemporal predictive learning primarily lies in its potential of exploring both spatial correlation and temporal evolution in the physical world. Moreover, the self-supervised nature of spatiotemporal predictive learning aligns well with human learning styles without a large amount of labeled data. Massive videos can provide a rich source of visual information, enabling spatiotemporal pre" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18770" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "dictive learning to serve as a generative pre-training strategy [35, 60] for feature representation learning towards diverse downstream visual supervised tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 111, + 289, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 111, + 289, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 111, + 289, + 484 + ], + "type": "text", + "content": "Most of the existing methods [2, 3, 9, 11, 12, 19, 24, 29, 41, 44, 65, 78, 79, 82, 88, 89, 93–96, 98, 100, 103] in spatiotemporal predictive learning employ hybrid architectures of convolutional neural networks and recurrent units in which spatial correlation and time evolution can be learned, respectively. Inspired by the success of long short-term memory (LSTM) [36] in sequential modeling, ConvLSTM [78] is a seminal work on the topic of spatiotemporal predictive learning that extends fully connected LSTM to convolutional LSTM towards accurate precipitation nowcasting. PredRNN [95] is an admirable work that proposes Spatiotemporal LSTM (ST-LSTM) units to model spatial appearances and temporal variations in a unified memory pool. This work provides insights on designing typical recurrent units for spatiotemporal predictive learning and inspires a series of subsequent works [4, 93, 96, 98]. E3D-LSTM [94] integrates 3D convolutional neural networks into recurrent units towards good representations with both short-term frame dependencies and long-term high-level relations. PhyDNet [29] introduces a two-branch architecture that involves physical-based PhyCells and ConvLSTMs for performing partial differential equation constraints in latent space. CrevNet [103] proposes an invertible two-way autoencoder based on flow [14, 15] and a conditionally reversible architecture for spatiotemporal predictive learning. As shown in Figure 1 (a), we present a general framework consisting of the spatial encoder/decoder and the middle temporal module, abstracted from these methods. Though these spatiotemporal predictive learning methods have different temporal modules and spatial encoders/decoders, they basically share a similar framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "content": "Based on the general framework, we argue that the temporal module plays an essential role in spatiotemporal predictive learning. While the common choice of the temporal module is the recurrent-based units, we explore a novel parallelizable attention module named Temporal Attention Unit (TAU) to capture time evolution. The proposed temporal attention is decomposed into intra-frame statistical attention and inter-frame dynamical attention. Furthermore, we argue that the mean square error loss only focuses on intra-frame differences, and we propose a differential divergence regularization that also cares about the inter-frame variations. Keeping the spatial encoder and decoder as simple as 2D convolutional neural networks, we deliberately implement our proposed TAU modules and surprisingly find the derived model achieves competitive performance as those recurrent-based models. This observation provides a new perspective to improve spatiotemporal predictive learning by parallelizable attention networks instead of commonly used recurrent units." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "We conduct experiments on various datasets with different experimental settings: (1) Standard spatiotemporal predictive learning. Quantitative results on various datasets demonstrate our proposed method can achieve competitive performance on standard spatiotemporal predictive learning. (2) Generalization ability. To verify the generalization ability, we train our model on KITTI and test it on the Caltech Pedestrian dataset with different domains. (3) Predict future frames with flexible lengths. We tackle the long-length frames by feeding the predicted frames as the input and find the performance is consistently well. Through the superior performance in the above three experimental settings, we demonstrate that our proposed model can provide a novel manner that learns temporal dependencies without recurrent units." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 264, + 395, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 395, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 395, + 276 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 283, + 441, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 441, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 441, + 297 + ], + "type": "text", + "content": "2.1. Self-supervised learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 302, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 589 + ], + "type": "text", + "content": "Deep learning has been well developed and applied in various fields [5, 55, 56, 109, 110]. Learning from massive data enables tremendous progress in supervised learning. By designing pretext tasks and generating labels from the data itself, self-supervised learning obtains supervisory signals. The model learns valuable representations by solving pretext tasks that leverage the underlying structure of the data. Early works on visual self-supervised learning design pretext tasks such as colorization [108], inpainting [69], rotation [26], jigsaw [66]. Contrastive self-supervised learning [7, 8, 28, 33, 92, 104] is a dominant manner in visual self-supervised learning that aims at a pretext task of grouping similar samples closer and diverse samples away from each other. However, contrastive self-supervised learning is limited by making pairs by multiple images, which affects its ability on small-scale datasets. Masked self-supervised learning [17, 32, 45, 51, 57, 102], which predicts the masked patches from the visible ones, is another research direction. Although masked pretraining has great success in natural language processing, its applications in visual tasks are challenging. Spatiotemporal predictive learning is another promising branch of self-supervised learning that focuses on video-level information and predicts future frames conditioned on past frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 590, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 651 + ], + "type": "text", + "content": "In contrast to the above image-level methods, spatiotemporal predictive learning focus on video-level information and predicts future frames conditioned on past frames. By learning the intrinsic motion dynamics, the model is enabled to easily decouple the foreground and background." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 493, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 493, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 493, + 672 + ], + "type": "text", + "content": "2.2. Spatiotemporal predictive learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Recurrent models have achieved remarkable advances in spatiotemporal predictive learning. Inspired by recurrent neural networks, VideoModeling [62] adopts language" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18771" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "modeling and quantizes the image patches into a large dictionary for recurrent units. ConvLSTM [78] leverages convolutional neural networks to model the LSTM architecture. PredNet [61] persistently predicts future video frames using deep recurrent convolutional neural networks with bottom-up and top-down connections. PredRNN [95] proposes a Spatiotemporal LSTM unit that extracts and memorizes spatial and temporal representations simultaneously, and its following work " + }, + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{PredRNN + + }" + }, + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": " [93] further introduces gradient highway unit and Casual LSTM to adaptively capture temporal dependencies. E3D-LSTM [94] designs e-detic memory transition in recurrent convolutional units. Conv-TT-LSTM [82] employs a higher-order ConvLSTM to predict by combining convolutional features across time. MotionRNN [100] focuses on motion trends and transient variations. LMC-Memory [50] introduces a long-term motion context memory using memory alignment learning. PredRNN-v2 [96] extends PredRNN by leveraging a memory decoupling loss and curriculum learning strategy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 576 + ], + "type": "text", + "content": "Instead of using recurrent-based methods that are computationally expensive for spatiotemporal predictive learning, we introduce TAU, a model that uses visual attention mechanism to parallelize the temporal evolution without the recurrent structure. There are prior arts that have some similarities with our proposed model. PredCNN [101] and TrajectoryCNN [54] implement pure convolutional neural networks as the temporal module. SimVP [23] is a seminal work that applies blocks of Inception modules with a UNet architecture to learn the temporal evolution. Though their temporal modules are parallelizable, we argue that convolutions alone cannot capture long-term dependencies. Moreover, SimVP provides a simple baseline with minor complex attachment but a large space for further improvements. In general, SimVP first downsamples video sequences to reduce the computation, then uses Inception-UNet to learn essential spatiotemporal relationships, and upsamples the representations to predict future frames. Our work aims to replace the pivotal Inception-UNet with efficient attention modules that promote prediction performance. In this paper, we employ a simple yet effective attention mechanism to enable the temporal module not only to be parallelizable but also to capture long-term time evolution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 587, + 107, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 107, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 107, + 600 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 608, + 133, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 133, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 133, + 619 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": "We formally define the spatiotemporal predictive learning problem as follows. Given a video sequence " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{t,T} = \\{\\mathbf{x}^i\\}_{t - T + 1}^t" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " with the past " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " frames, we aim to predict the subsequent " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "T^{\\prime}" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{t + 1,T^{\\prime}} = \\{\\mathbf{x}^{i}\\}_{t + 1}^{t + 1 + T^{\\prime}}" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " from time " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " is usually an image with channels " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": ", and width " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": ". In practice, we represent the video sequences as tensors, i.e., " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{t,T}\\in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 70, + 486, + 84 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 486, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 486, + 84 + ], + "type": "interline_equation", + "content": "\\mathbb {R} ^ {T \\times C \\times H \\times W} \\text {a n d} \\mathcal {Y} ^ {t + 1, T ^ {\\prime}} \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C \\times H \\times W}.", + "image_path": "895971fad7fe8169c67e455e67bcd140620bcb1f784fa2990ba67dda140999ea.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "text", + "content": "The model with learnable parameters " + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "text", + "content": " learns a mapping " + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Theta}:\\mathcal{X}^{t,T}\\mapsto \\mathcal{Y}^{t + 1,T^{\\prime}}" + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "text", + "content": " by exploring both spatial and temporal dependencies. In our case, the mapping " + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Theta}" + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "text", + "content": " is a neural network model trained to minimize the difference between the predicted future frames and the ground-truth future frames. The optimal parameters " + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "inline_equation", + "content": "\\Theta^{*}" + }, + { + "bbox": [ + 304, + 84, + 547, + 157 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 346, + 165, + 545, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 165, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 346, + 165, + 545, + 185 + ], + "type": "interline_equation", + "content": "\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} \\left(\\mathcal {F} _ {\\Theta} \\left(\\mathcal {X} ^ {t, T}\\right), \\mathcal {Y} ^ {t + 1, T ^ {\\prime}}\\right), \\tag {1}", + "image_path": "ebfec01950a6831fdeb4a002dfa12c2033d990249999ff15099debe0748fb2e1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 193, + 539, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 193, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 539, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 193, + 539, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 193, + 539, + 205 + ], + "type": "text", + "content": " is a loss function that evaluates such differences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 213, + 373, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 373, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 373, + 225 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 232, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 257 + ], + "type": "text", + "content": "We illustrate the overview model in Figure 2 using the input Moving MNIST [81] data as an example." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 268, + 533, + 398 + ], + "blocks": [ + { + "bbox": [ + 318, + 268, + 533, + 398 + ], + "lines": [ + { + "bbox": [ + 318, + 268, + 533, + 398 + ], + "spans": [ + { + "bbox": [ + 318, + 268, + 533, + 398 + ], + "type": "image", + "image_path": "fd3a7222ad8ab4804a5e072385ff3488a7ef65ca766d859d2da2b2e2f284ffbd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 411, + 534, + 422 + ], + "lines": [ + { + "bbox": [ + 316, + 411, + 534, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 534, + 422 + ], + "type": "text", + "content": "Figure 2. The overview architecture of our proposed model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 435, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 590 + ], + "type": "text", + "content": "Striving for simplicity, the model follows the general framework in Figure 1, while the spatial encoder consists of four vanilla 2D convolutional layers, and the spatial decoder consists of four 2D transposed convolutional layers ('Conv2d' and 'ConvTranspose2d' in PyTorch respectively). We add a residual connection from the first convolutional layer to the last transposed convolutional layer for preserving the spatial-dependent features. Stacks of TAU modules are in the middle of the spatial encoder and decoder to extract temporal-dependent features. Though our model is simple, it can efficiently learn both spatial-dependent and temporal-dependent features without recurrent architectures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 599, + 444, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 444, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 444, + 611 + ], + "type": "text", + "content": "3.3. Temporal Attention Unit" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Suppose a batch of input video tensors " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\in \\mathbb{R}^{B \\times T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " with the number of videos " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "B = |\\mathcal{B}|" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " is given. In the spatial encoder and decoder, we reshape the sequential input data " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "B \\times T \\times C \\times H \\times W" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "(B \\times T) \\times C \\times H \\times W" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " so that only spatial correlations are taken into account. In the temporal module, we reshape the feature " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "B \\times T \\times C \\times H \\times W" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "B \\times (T \\times C) \\times H \\times W" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " so that frames are arranged in order on the channel dimension." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18772" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "We decompose the temporal attention into the intra-frame statical attention and the inter-frame dynamical attention, as shown in Figure 3. Inspired by the recent progress of vision Transformers (ViTs) [17,57] and large kernel convolutions [13, 30, 58], we propose to employ small kernel depth-wise convolutions (DW Conv), depth-wise convolutions with dilations (DW-D Conv), and " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " convolutions to model the large kernel convolutions. Through the obtained large receptive field on intra-frames, the statical attention is able to capture long-range dependencies. However, the statical attention alone is not enough for learning temporal evolutions along the timeline. Thus, we employ the dynamical attention that learns the attention weights of channels in a squeeze-and-excitation manner [38]. The final attention is the product of dynamical attention and statical attention." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 262, + 289, + 405 + ], + "blocks": [ + { + "bbox": [ + 50, + 262, + 289, + 405 + ], + "lines": [ + { + "bbox": [ + 50, + 262, + 289, + 405 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 289, + 405 + ], + "type": "image", + "image_path": "a61ec68faafa29d95fe56d8da580d234f0fbb960421c6c20f0b20b4824330584.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 413, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 437 + ], + "type": "text", + "content": "Figure 3. The intra-frame statical attention and the inter-frame dynamical attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 449, + 286, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 286, + 475 + ], + "type": "text", + "content": "We show the detailed scheme of our model in Figure 4. The proposed TAU module can be formally expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 484, + 262, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 484, + 262, + 498 + ], + "spans": [ + { + "bbox": [ + 71, + 484, + 262, + 498 + ], + "type": "interline_equation", + "content": "\\mathrm {S A} = \\operatorname {C o n v} _ {1 \\times 1} (\\mathrm {D W - D C o n v} (\\mathrm {D W C o n v} (H))),", + "image_path": "8f3ceeeaa90177f86a2e5edb8d2c97b13725d64d7902ce827833dc13b71b9242.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 499, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 499, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 71, + 499, + 287, + 512 + ], + "type": "interline_equation", + "content": "\\mathrm {D A} = \\mathrm {F C} (\\operatorname {A v g P o o l} (H)), \\tag {2}", + "image_path": "9ff7d3a689ca2c27b1195e2607df90ed5cf8378de947b1ac81d6238d7530edda.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 514, + 184, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 514, + 184, + 528 + ], + "spans": [ + { + "bbox": [ + 74, + 514, + 184, + 528 + ], + "type": "interline_equation", + "content": "H ^ {\\prime} = (\\mathrm {S A} \\otimes \\mathrm {D A}) \\odot H,", + "image_path": "803970c8abf029d5753d7bee0b952fd35ee60c7ab7432a18663501ef1ec87a19.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^{B \\times (T \\times C') \\times H \\times W}" + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "content": " is the hidden feature that will be fed into the TAU module, SA " + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{B \\times (T \\times C') \\times H \\times W}" + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "content": ", DA " + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{B \\times (T \\times C') \\times 1 \\times 1}" + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "content": " denote the statistical and dynamical attention, FC and AvgPool are fully connected layers and the average pooling. We represent the Kronecker product by " + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "content": " and the Hadamard product by " + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 538, + 287, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 621, + 251, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 251, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 251, + 635 + ], + "type": "text", + "content": "3.4. Differential Divergence Regularization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 688 + ], + "type": "text", + "content": "To improve the prediction of our model, we further propose a differential divergence regularization that forces the model to learn the differences between consecutive frames and be aware of the inherent variation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Given the predicted frames " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{Y}} = \\mathcal{F}_{\\Theta}(\\mathcal{X}) \\in \\mathbb{R}^{T' \\times C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " and its corresponding ground-truth frames" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 533, + 182 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 533, + 182 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 533, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 533, + 182 + ], + "type": "image", + "image_path": "d583ca0e0b0e36fbf267a83ae0f4fcce757dd20b4d820959f029804b5c5c8add.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 190, + 506, + 201 + ], + "lines": [ + { + "bbox": [ + 343, + 190, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 343, + 190, + 506, + 201 + ], + "type": "text", + "content": "Figure 4. The detailed schema of our model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 222, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 222, + 545, + 247 + ], + "type": "text", + "content": ", we first calculate their forward difference " + }, + { + "bbox": [ + 305, + 222, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\Delta \\widehat{\\mathcal{Y}}, \\Delta \\mathcal{Y} \\in \\mathbb{R}^{(T' - 1) \\times C \\times H \\times W}" + }, + { + "bbox": [ + 305, + 222, + 545, + 247 + ], + "type": "text", + "content": ", where:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 257, + 545, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 257, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 386, + 257, + 545, + 287 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta \\widehat {\\mathcal {Y}} _ {i} = \\widehat {\\mathcal {Y}} _ {i + 1} - \\widehat {\\mathcal {Y}} _ {i}, \\tag {3} \\\\ \\Delta \\mathcal {Y} _ {i} = \\mathcal {Y} _ {i + 1} - \\mathcal {Y} _ {i}. \\\\ \\end{array}", + "image_path": "925c9812cfb9c07de869a5ed0b2dad1385163a08d8088a76b335de676339dc00.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "content": "Then, we transform the differences into probabilities by the softmax function on the channel, height, and width dimension and obtain " + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\sigma (\\Delta \\widehat{\\mathcal{Y}}),\\sigma (\\Delta \\mathcal{Y})" + }, + { + "bbox": [ + 305, + 299, + 545, + 335 + ], + "type": "text", + "content": " , where:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 346, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sigma (\\Delta \\widehat {\\mathcal {Y}}) _ {i, j, k, l} = \\frac {\\exp \\left(\\Delta \\widehat {\\mathcal {Y}} _ {i , j , k , l} / \\tau\\right)}{\\sum_ {j ^ {\\prime} = 1} ^ {C} \\sum_ {k ^ {\\prime} = 1} ^ {H} \\sum_ {l ^ {\\prime} = 1} ^ {W} \\exp \\left(\\Delta \\widehat {\\mathcal {Y}} _ {i , j ^ {\\prime} , k ^ {\\prime} , l ^ {\\prime}} / \\tau\\right)}, \\\\ \\sigma \\left(\\Delta \\mathcal {Y}\\right) _ {i, j, k, l} = \\frac {\\exp \\left(\\Delta \\mathcal {Y} _ {i , j , k , l}\\right) / \\tau}{\\sum_ {j ^ {\\prime} = 1} ^ {C} \\sum_ {k ^ {\\prime} = 1} ^ {H} \\sum_ {l ^ {\\prime} = 1} ^ {W} \\exp \\left(\\Delta \\mathcal {Y} _ {i , j ^ {\\prime} , k ^ {\\prime} , l ^ {\\prime}} / \\tau\\right)}, \\tag {4} \\\\ \\end{array}", + "image_path": "7cd848ad7bf949b167c43ad82c2794939e430112e0041789a11090c351039545.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 418, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 466 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 305, + 418, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 418, + 545, + 466 + ], + "type": "text", + "content": " represents the temperature parameter which we empirically set as 0.1 to sharpen the probability distribution. Through the competition mechanism of the softmax function [10, 70, 99], the high difference frames are penalized." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "text", + "content": "Thus, the differential divergence regularization " + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "text", + "content": " is defined as the Kullback-Leibler divergence between the probability distributions " + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\sigma (\\Delta \\widehat{\\mathcal{V}})" + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\sigma (\\Delta \\mathcal{V})" + }, + { + "bbox": [ + 305, + 467, + 545, + 503 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 341, + 513, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 513, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 341, + 513, + 545, + 564 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {r e g} (\\widehat {\\mathcal {Y}}, \\mathcal {Y}) = D _ {K L} (\\sigma (\\Delta \\widehat {\\mathcal {Y}}) \\mid \\mid \\sigma (\\Delta \\mathcal {Y})) \\\\ = \\sum_ {i = 1} ^ {T ^ {\\prime} - 1} \\sigma \\left(\\Delta \\hat {\\mathcal {Y}} _ {i}\\right) \\log \\frac {\\sigma \\left(\\Delta \\hat {\\mathcal {Y}} _ {i}\\right)}{\\sigma \\left(\\Delta \\mathcal {Y} _ {i}\\right)}. \\tag {5} \\\\ \\end{array}", + "image_path": "b71edd50bc111da6ecd883aabba35000eff97cbd660c6cc63d91b36abdb6620e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 575, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 545, + 624 + ], + "type": "text", + "content": "Our model is trained end-to-end in a fully unsupervised manner, and the overall objective function consists of the mean square error loss and the differential divergence regularization weighted by a constant " + }, + { + "bbox": [ + 305, + 575, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 575, + 545, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 634, + 545, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 351, + 634, + 545, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {T ^ {\\prime}} \\| \\widehat {\\mathcal {Y}} - \\mathcal {Y} \\| ^ {2} + \\alpha \\mathcal {L} _ {r e g} (\\widehat {\\mathcal {Y}}, \\mathcal {Y}), \\tag {6}", + "image_path": "05a00b8093060e9225ccb249f633fcceded55d6e7029949da2f88846bc47fb69.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "where the first term focuses on intra-frame-level differences, and the second regularization term focuses on inter-frame-level variations." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18773" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "type": "text", + "content": "In this section, we present experiments that demonstrate the effectiveness of our proposed method. The experiments are conducted on various datasets with different settings to validate our proposed model from three aspects:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 147, + 287, + 416 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 147, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 147, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 147, + 287, + 243 + ], + "type": "text", + "content": "- Standard spatiotemporal predictive learning (Section 4.2). We recognize the prediction problem of the same number of input and output frames as the standard spatiotemporal predictive learning. We evaluate the performance on standard spatiotemporal predictive learning and compare our model with state-of-the-art methods with Moving MNIST [81] and TaxiBJ [106] datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 251, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 251, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 287, + 323 + ], + "type": "text", + "content": "- Generalization ability across different datasets (Section 4.3). Generalizing the learned knowledge to other domains is a challenge in unsupervised learning. We investigate the such ability of our method by training the model on the KITTI [25] dataset and evaluating it on the Caltech Pedestrian [16] dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 331, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 331, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 287, + 416 + ], + "type": "text", + "content": "- Predicting frames with flexible lengths (Section 4.4). One of the advantages of recurrent units is that they can easily handle flexible-length frames like the KTH dataset [76]. Our work tackles the long-length frame prediction by imitating recurrent units that feed predicted frames as the input and recursively produce long-term predictions." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 422, + 168, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 168, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 168, + 436 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 441, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 465 + ], + "type": "text", + "content": "Datasets We quantitatively evaluate our model on the following datasets for both synthetic and real-world scenarios:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 472, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 58, + 472, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 472, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 58, + 472, + 287, + 521 + ], + "type": "text", + "content": "- Moving MNIST [81] is a synthetic dataset consisting of two digits independently moving within the " + }, + { + "bbox": [ + 58, + 472, + 287, + 521 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 58, + 472, + 287, + 521 + ], + "type": "text", + "content": " grid and bouncing off the boundary. It is a standard benchmark in spatiotemporal predictive learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 529, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 529, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 58, + 529, + 287, + 578 + ], + "type": "text", + "content": "- TaxiBJ contains the trajectory data in Beijing collected from taxicab GPS with two channels, i.e., inflow or outflow defined in [106]. Following the previous works [29, 98], we normalize the data into " + }, + { + "bbox": [ + 58, + 529, + 287, + 578 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 58, + 529, + 287, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 585, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 585, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 58, + 585, + 287, + 645 + ], + "type": "text", + "content": "- KTH [76] contains 25 individuals performing six types of actions. Following [90, 94], we use person 1-16 for training and 17-25 for testing. Models are trained to predict the next 20 or 40 frames from the previous 10 observations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 653, + 287, + 713 + ], + "type": "text", + "content": "- Caltech Pedestrian is a driving dataset focusing on detecting pedestrians. It consists of approximately 10 hours of " + }, + { + "bbox": [ + 58, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 58, + 653, + 287, + 713 + ], + "type": "text", + "content": " videos taken from vehicles driving through regular traffic in an urban environment. We follow the same protocol of PredNet [61] and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 544, + 95 + ], + "type": "text", + "content": "CrevNet [103] for pre-processing, training, and evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "type": "text", + "content": "We summarize the statistics of the above datasets in Table 1, including the number of training samples " + }, + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "type": "inline_equation", + "content": "N_{train}" + }, + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "type": "text", + "content": " and the number of testing samples " + }, + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "type": "inline_equation", + "content": "N_{test}" + }, + { + "bbox": [ + 305, + 106, + 545, + 143 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 307, + 196, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "lines": [ + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "text", + "content": "Table 1. The statistics of datasets. The training or testing set has " + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "inline_equation", + "content": "N_{train}" + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "inline_equation", + "content": "N_{test}" + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "text", + "content": " samples, composed by " + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "text", + "content": " images with the shape " + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "inline_equation", + "content": "(C, H, W)" + }, + { + "bbox": [ + 305, + 153, + 545, + 187 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 196, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 269 + ], + "type": "table", + "html": "
NtrainNtest(C,H,W)TT'
MMNIST1000010000(1,64,64)1010
TaxiBJ196271334(2,32,32)44
KTH52003167(1,128,128)1020 or 40
Caltech20421983(3,128,160)101
", + "image_path": "9caaa2808c30e822c694898d4930b62158aa22c34e2d5614b9d9d9a227ff73fc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 300, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 407 + ], + "type": "text", + "content": "Measurement Following [29, 103], we employ Mean Squared Error (MSE), Mean Absolute Error (MAE), Structure Similarity Index Measure (SSIM), and Peak Signal to Noise Ratio (PSNR) to evaluate the quality of predictions. MSE and MAE estimate the absolute pixel-wise errors, SSIM measures the similarity of structural information within the spatial neighborhoods, and PSNR is an expression for the ratio between the maximum possible power of a signal and the power of distorted noise." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 425, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 497 + ], + "type": "text", + "content": "Implementation details We implement the proposed method with the Pytorch framework and conduct experiments on a single NVIDIA-V100 GPU. The model is trained with a mini-batch of 16 video sequences while the AdamW optimizer is utilized with a learning rate of 0.01 and a weight decay of 0.05." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 506, + 537, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 537, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 537, + 520 + ], + "type": "text", + "content": "4.2. Standard spatiotemporal predictive learning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 525, + 406, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 406, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 406, + 537 + ], + "type": "text", + "content": "4.2.1 Moving MNIST" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 545, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 664 + ], + "type": "text", + "content": "This dataset is a standard benchmark in spatiotemporal predictive learning. We evaluate our proposed method against strong recent baselines, including competitive recurrent architectures: ConvLSTM [78], PredRNN [95], PredRNN++ [93], MIM [98], LMC [50], E3D-LSTM [94], Conv-TT-LSTM [82], and CrevNet [103]. We also compare the method DDPAE [37] that is specifically designed for this dataset. The quantitative results are reported in Table 2, and qualitative visualizations of the predicted results are shown in Figure 5." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Our proposed method significantly outperforms all the baselines above under three different metrics. The performance gain is large with respect to state-of-the-art recurrent methods." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18774" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 547, + 265 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 547, + 265 + ], + "type": "image", + "image_path": "fbe5d313999c722cd256d67932b735dd8ce1ed7c7423bb70822b0b58353fb7bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 272, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 272, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 547, + 296 + ], + "type": "text", + "content": "Figure 5. Qualitative visualization of predicted results on Moving MNIST dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 345, + 282, + 539 + ], + "blocks": [ + { + "bbox": [ + 47, + 314, + 288, + 337 + ], + "lines": [ + { + "bbox": [ + 47, + 314, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 288, + 337 + ], + "type": "text", + "content": "Table 2. Quantitative results of different methods on the Moving MNIST dataset (10 → 10 frames)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 345, + 282, + 539 + ], + "lines": [ + { + "bbox": [ + 50, + 345, + 282, + 539 + ], + "spans": [ + { + "bbox": [ + 50, + 345, + 282, + 539 + ], + "type": "table", + "html": "
MethodMoving MNIST
MSE↓MAE↓SSIM↑
ConvLSTM [78]103.3182.90.707
VPN [44]64.1-0.870
PredRNN [95]56.8126.10.867
PredRNN++ [93]46.5106.80.898
MIM [98]44.2101.10.910
LMC [50]41.5-0.924
E3D-LSTM [94]41.387.20.910
Conv-TT-LSTM [82]53.0-0.915
DDPAE [37]38.990.70.922
PhyDNet [29]24.470.30.947
SimVP [23]23.868.90.948
Crevnet [103]22.3-0.949
Ours19.860.30.957
", + "image_path": "bbe974a303c38236ec0c4328e6239dc6f1ad62182749be19fa6206df00a741ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 560, + 111, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 111, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 111, + 571 + ], + "type": "text", + "content": "4.2.2 TaxiBJ" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 652 + ], + "type": "text", + "content": "We evaluate our proposed model on a complicated real-world dataset, TaxiBJ [106]. Driven by human consciousness, the complex real-world traffic flows requires modeling transport phenomena and traffic diffusion for prediction. Due to the spatiotemporal nature of the traffic forecasting task, we straightforwardly implement our model for it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "The qualitative visualizations of the predicted results are shown in Figure 6, and the quantitative results are reported in Table 3. Though the given frames are quite different from the future frames, our model can still accurately produce reliable frames. The difference between target frames" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 316, + 547, + 521 + ], + "blocks": [ + { + "bbox": [ + 312, + 316, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 312, + 316, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 547, + 521 + ], + "type": "image", + "image_path": "199de6d2bec999db32824ca989a31c8f958426e0847ec35051c8eef29d2791aa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 527, + 547, + 560 + ], + "lines": [ + { + "bbox": [ + 305, + 527, + 547, + 560 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 547, + 560 + ], + "type": "text", + "content": "Figure 6. Qualitative visualization of predicted results on TaxiBJ dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 585, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 547, + 645 + ], + "type": "text", + "content": "and predicted frames is mainly located in central spots, but the overall trend is approximating the ground truth. It can also be observed that the quantitative results are consistently well, which demonstrates the practical application value in real-world scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 657, + 514, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 514, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 514, + 670 + ], + "type": "text", + "content": "4.3. Generalization across different datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 714 + ], + "type": "text", + "content": "The generalization ability is one of the fundamental problems in artificial intelligence technology. Traditional supervised learning suffers from its poor generalization of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18775" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 101, + 287, + 235 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 3. Quantitative results of different methods on the TaxiBJ dataset (4 → 4 frames)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 101, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 48, + 101, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 101, + 287, + 235 + ], + "type": "table", + "html": "
MethodTaxiBJ
MSE × 100↓MAE↓SSIM↑
ConvLSTM [78]48.517.70.978
PredRNN [95]46.417.10.971
PredRNN++ [93]44.816.90.977
MIM [98]42.916.60.971
E3D-LSTM [94]43.216.90.979
PhyDNet [29]41.916.20.982
SimVP [23]41.416.20.982
Ours34.415.60.983
", + "image_path": "ad642e9e4c5ef1047fb5c9fa5f9da974ca08334640ba131940ed28ead18a8ce0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 257, + 288, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 354 + ], + "type": "text", + "content": "labeled datasets with different domains. Self-supervised learning aims to learn robust representations based on unlabeled data and evaluates the generalization ability of the learned model. While contrastive self-supervised learning and masked self-supervised learning in visual tasks usually evaluate such generalization ability by downstream tasks, we evaluate it by the prediction results across different datasets in spatiotemporal predictive learning." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 368, + 289, + 575 + ], + "blocks": [ + { + "bbox": [ + 47, + 368, + 289, + 575 + ], + "lines": [ + { + "bbox": [ + 47, + 368, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 289, + 575 + ], + "type": "image", + "image_path": "43795de1a2785063eb18e369c9dfc71db7ba3927c12e26cea33c646cd5a2d33b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 581, + 287, + 615 + ], + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 615 + ], + "type": "text", + "content": "Figure 7. Qualitative visualization of predicted results on Caltech dataset. The differences between the ground truth and the predicted frames are visualized in the last row." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 628, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 700 + ], + "type": "text", + "content": "Following the previous works [53, 61, 103], we train the model using the raw video sequences from the KITTI dataset and evaluate the model by Caltech Pedestrian dataset that is made to match the frame rate and image size " + }, + { + "bbox": [ + 46, + 628, + 287, + 700 + ], + "type": "inline_equation", + "content": "(128 \\times 160)" + }, + { + "bbox": [ + 46, + 628, + 287, + 700 + ], + "type": "text", + "content": " of the KTTI dataset. The final prediction is made for the next frame after a 10-frame warm-up." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "We show the qualitative visualization in Figure 7 and re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "port the quantitative results in Table 4. Note that some of the baseline results are copied from [68]. It can be seen that our proposed method achieves state-of-the-art performance under both SSIM and PSNR metrics in this generalization evaluation task. Moreover, our model shows significantly robust predictions in terms of the variation of illumination and lane line, suggesting its practical applications in autonomous vehicles." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 205, + 542, + 459 + ], + "blocks": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 196 + ], + "type": "text", + "content": "Table 4. Quantitative results of different methods on the Caltech Pedestrian dataset (10 → 1 frame)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 205, + 542, + 459 + ], + "lines": [ + { + "bbox": [ + 308, + 205, + 542, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 542, + 459 + ], + "type": "table", + "html": "
MethodCaltech Pedestrian
SSIM↑PSNR↑
BeyondMSE [64]0.847-
MCnet [90]0.879-
DVF [59]0.89726.2
Dual-GAN [53]0.899-
CtrlGen [31]0.90026.5
PredNet [61]0.90527.6
ContextVP [4]0.92128.7
GAN-VGG [80]0.916-
G-VGG [80]0.917-
SDC-Net [72]0.918-
rCycleGan [47]0.91929.2
DPG [20]0.92328.2
G-MAE [80]0.923-
GAN-MAE [80]0.923-
CrevNet [103]0.92529.3
STMFANet [41]0.92729.1
SimVP [23]0.94033.1
Ours0.94633.7
", + "image_path": "7d07cf74ba0339100eb9e3575787b2b40098187db3ba0d35a91d799c5a8dd5fe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 467, + 508, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 508, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 508, + 480 + ], + "type": "text", + "content": "4.4. Predicting frames with flexible lengths" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 712 + ], + "type": "text", + "content": "Though recurrent units are adept at handling flexible-length frames, our model can also easily tackle such problems by imitating recurrent units that feed predicted frames as the input and recursively produce predictions. For this KTH dataset, our model is trained to predict the next 20 or 40 frames from the given 10 observations. Moreover, this dataset contains six types of human actions (walking, jogging, running, boxing, hand waving and hand clapping) performed several times by 25 subjects in four different scenarios: outdoors, outdoors with scale variations, outdoors with different clothes and indoors. The difficulty of this human motion prediction task not only lies in its flexible lengths of predicted frames but also in its complex dynamics involving the randomness of human consciousness. Following [95], we use the Peak Signal Noise Ratio (PSNR) and the Structural Similarity Index Measure (SSIM) as evaluation metrics to measure the framewise prediction quality from the perceptive view. The detailed quantitative results are shown in Table 5." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18776" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 282, + 347 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 5. Quantitative results of different methods on the KTH dataset (10 → 20 and 10 → 40 frames)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 282, + 347 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 282, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 282, + 347 + ], + "type": "table", + "html": "
MethodKTH (10 → 20)KTH (10 → 40)
SSIM↑PSNR↑SSIM↑PSNR↑
MCnet [90]0.80425.950.7323.89
ConvLSTM [78]0.71223.580.63922.85
SAVP [48]0.74625.380.70123.97
VPN [44]0.74623.76--
DFN [40]0.79427.260.65223.01
fRNN [67]0.77126.120.67823.77
Znet [105]0.81727.58--
SV2Pv [1]0.83827.790.78926.12
PredRNN [95]0.83927.550.70324.16
VarNet [42]0.84328.480.73925.37
SAVP-VAE [48]0.85227.770.81126.18
PredRNN++ [93]0.86528.470.74125.21
MSNET [49]0.87627.08--
E3d-LSTM [94]0.87929.310.81027.24
STMFANet [41]0.89329.850.85127.56
SimVP [23]0.90533.720.88632.93
Ours0.91134.130.89733.01
", + "image_path": "f7d616f49d984fe51fb7a7bb6836a378c006b4cf3392835d477b74832ab6ef48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 354, + 188, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 188, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 188, + 367 + ], + "type": "text", + "content": "4.5. Empirical Running Time" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "text", + "content": "TAU benefits from the parallelizable computation architecture, which leads to fast convergence and high training speed. We evaluate our efficiency by measuring the running time against state-of-the-art spatiotemporal predictive learning methods." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 438, + 270, + 555 + ], + "blocks": [ + { + "bbox": [ + 65, + 438, + 270, + 555 + ], + "lines": [ + { + "bbox": [ + 65, + 438, + 270, + 555 + ], + "spans": [ + { + "bbox": [ + 65, + 438, + 270, + 555 + ], + "type": "image", + "image_path": "732999e3311e35bd1909c6417231d248f792a2d454ae85fc70f22e594ffad014.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 566, + 287, + 599 + ], + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 599 + ], + "type": "text", + "content": "Figure 8. The learning curve comparison between state-of-the-art methods and ours (evaluated by MSE). The red dotted line denotes MSE 35.0, and only the first 50 epochs are shown." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "The experiments are conducted on a single Tesla V100 GPU, and the batch size is set as 16. For the Moving MNIST dataset, CrevNet [103] needs about 30 minutes per epoch, and PhyDNet [29] needs about 7 minutes per epoch. Our model only requires 2.5 minutes per epoch. Furthermore, our method is able to converge at a rapid rate. As shown in Figure 8, on the Moving MNIST dataset, our model can achieve MSE 35.0 with only 50 epochs, while CrevNet and PhyDNet are far from this performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 518, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 518, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 518, + 85 + ], + "type": "text", + "content": "4.6. Computational Cost and Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "type": "text", + "content": "We compare the performance and computational cost with state-of-the-art methods in the first several rows in Table 6. Our model achieves superior results with better performance and much lower Flops. We also conduct ablation studies and summarize the results in Table 6. It can be seen that replacing TAU with the same number of convolutional blocks with vanilla " + }, + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 90, + 545, + 245 + ], + "type": "text", + "content": " convolutions (Conv Baseline) significantly degrades the performance. Training our model without differential divergence regularization (w/o DDR) will also weaken the prediction results. Both the TAU module and differential divergence regularization are useful. We also find that SA and DA of the TAU module play important roles in better performance." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 275, + 538, + 426 + ], + "blocks": [ + { + "bbox": [ + 336, + 255, + 515, + 266 + ], + "lines": [ + { + "bbox": [ + 336, + 255, + 515, + 266 + ], + "spans": [ + { + "bbox": [ + 336, + 255, + 515, + 266 + ], + "type": "text", + "content": "Table 6. Ablation study of our proposed method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 275, + 538, + 426 + ], + "lines": [ + { + "bbox": [ + 309, + 275, + 538, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 275, + 538, + 426 + ], + "type": "table", + "html": "
MethodM-MNISTTaxiBJ (×100)Flops(G)
PredRNN56.846.4115.94
PredRNN++46.444.8171.73
MIM44.242.9179.17
E3DLSTM41.343.2298.85
SimVP23.841.419.43
Conv Baseline58.943.56.11
Ours w/o SA23.240.815.30
Ours w/o DA22.438.415.96
Ours w/o DDR21.137.715.96
Ours19.834.415.96
", + "image_path": "19292d7295734c7b61002acdc24ec72fcf0ff200ef5090fd2ff5f31ac532330f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 434, + 379, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 379, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 379, + 446 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 454, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 622 + ], + "type": "text", + "content": "In this paper, we present a general framework of spatiotemporal predictive learning and propose an attention-based temporal module to replace the common-used recurrent units. By decomposing the temporal module into the intra-frame statistical attention and the inter-frame dynamical attention, our proposed TAU module can achieve competitive performance across various experimental settings and datasets. Moreover, a novel differential divergence regularization is proposed to overcome the drawback of MSE loss that only considers the intra-frame error. In summary, our work highlights the importance of both intra-frame and inter-frame variations that enable the model to capture long-term relations and provide a new paradigm of efficient spatiotemporal predictive learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "Acknowledgement. We thank the anonymous reviewers for their constructive and helpful reviews. This work was supported by the National Key R&D Program of China (2022ZD0115100), the National Natural Science Foundation of China (U21A20427), the Competitive Research Fund (WU2022A009) from the Westlake Center for Synthetic Biology and Integrated Bioengineering" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18777" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Mohammad Babaeizadeh, Chelsea Finn, Dumitru Erhan, Roy H Campbell, and Sergey Levine. Stochastic variational video prediction. In International Conference on Learning Representations, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 136, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 136, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 136, + 287, + 178 + ], + "type": "text", + "content": "[2] Mohammad Babaeizadeh, Mohammad Taghi Saffar, Suraj Nair, Sergey Levine, Chelsea Finn, and Dumitru Erhan. Fitvid: Overfitting in pixel-level video prediction. arXiv preprint arXiv:2106.13195, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 180, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 180, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 287, + 235 + ], + "type": "text", + "content": "[3] Sarthak Bhagat, Shagun Uppal, Zhuyun Yin, and Nengli Lim. Disentangling multiple features in video sequences using gaussian processes in variational autoencoders. In European Conference on Computer Vision, pages 102-117. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 236, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 236, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 236, + 287, + 289 + ], + "type": "text", + "content": "[4] Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava, and Petros Koumoutsakos. Contextv: Fully context-aware video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 753-769, 2018. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 58, + 291, + 287, + 334 + ], + "type": "text", + "content": "[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 335, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 335, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 58, + 335, + 287, + 379 + ], + "type": "text", + "content": "[6] Lluis Castrejon, Nicolas Ballas, and Aaron Courville. Improved conditional vrnns for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7608-7617, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 380, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 380, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 58, + 380, + 287, + 424 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 425, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 425, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 58, + 425, + 287, + 479 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in Neural Information Processing Systems, 33:22243-22255, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 58, + 480, + 287, + 534 + ], + "type": "text", + "content": "[9] Yu Cheng, Bo Wang, Bo Yang, and Robby T Tan. Graph and temporal convolutional networks for 3d multi-person pose estimation in monocular videos. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 4, page 12, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 536, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 601 + ], + "type": "text", + "content": "[10] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. In International Conference on Learning Representations, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 635 + ], + "type": "text", + "content": "[11] Emily Denton and Rob Fergus. Stochastic video generation with a learned prior. In International Conference on Machine Learning, pages 1174-1183. PMLR, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 668 + ], + "type": "text", + "content": "[12] Emily L Denton et al. Unsupervised learning of disentangled representations from video. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. arXiv preprint arXiv:2203.06717, 2022.4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. In International Conference on Learning Representations Workshop, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 151 + ], + "type": "text", + "content": "[15] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "type": "text", + "content": "[16] Piotr Dólar, Christian Wojek, Bernt Schiele, and Pietro Perona. Pedestrian detection: A benchmark. In 2009 IEEE conference on computer vision and pattern recognition, pages 304-311. IEEE, 2009. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 262 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 307 + ], + "type": "text", + "content": "[18] Shen Fang, Qi Zhang, Gaofeng Meng, Shiming Xiang, and Chunhong Pan. GSTnet: Global spatial-temporal network for traffic flow prediction. In IJCAI, pages 2286-2293, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 545, + 352 + ], + "type": "text", + "content": "[19] Chelsea Finn, Ian Goodfellow, and Sergey Levine. Unsupervised learning for physical interaction through video prediction. Advances in Neural Information Processing Systems, 29, 2016. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 408 + ], + "type": "text", + "content": "[20] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "type": "text", + "content": "[21] Zhangyang Gao, Cheng Tan, Stan Li, et al. Alphadesign: A graph protein design method and benchmark on alphafolddb. arXiv preprint arXiv:2202.01079, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "type": "text", + "content": "[22] Zhangyang Gao, Cheng Tan, and Stan Z Li. Pifold: Toward effective and efficient protein inverse folding. arXiv preprint arXiv:2209.12643, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 479, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 545, + 532 + ], + "type": "text", + "content": "[23] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z. Li. Simvp: Simpler yet better video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3170-3180, June 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 535, + 545, + 578 + ], + "type": "text", + "content": "[24] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. In International Conference on Machine Learning, pages 1243-1252. PMLR, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 580, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 580, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 312, + 580, + 545, + 622 + ], + "type": "text", + "content": "[25] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 667 + ], + "type": "text", + "content": "[26] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthew, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative varia" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18778" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 95 + ], + "type": "text", + "content": "tional inference. In International Conference on Machine Learning, pages 2424-2433. PMLR, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 204 + ], + "type": "text", + "content": "[28] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and Michal Valko. Bootstrap your own latent - a new approach to self-supervised learning. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21271-21284. Curran Associates, Inc., 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 287, + 259 + ], + "type": "text", + "content": "[29] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 293 + ], + "type": "text", + "content": "[30] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 294, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 337 + ], + "type": "text", + "content": "[31] Zekun Hao, Xun Huang, and Serge Belongie. Controllable video generation with sparse trajectories. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7854-7863, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 287, + 392 + ], + "type": "text", + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 393, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 287, + 446 + ], + "type": "text", + "content": "[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 287, + 491 + ], + "type": "text", + "content": "[34] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 493, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 493, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 493, + 287, + 524 + ], + "type": "text", + "content": "[35] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. arXiv preprint arXiv:2102.12627, 2021.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 526, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 287, + 547 + ], + "type": "text", + "content": "[36] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 591 + ], + "type": "text", + "content": "[37] Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li F Fei-Fei, and Juan Carlos Niebles. Learning to decompose and disentangle representations for video prediction. Advances in Neural Information Processing Systems, 31, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 593, + 287, + 635 + ], + "type": "text", + "content": "[38] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 679 + ], + "type": "text", + "content": "[39] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision, pages 425-442. Springer, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "text", + "content": "[40] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 139 + ], + "type": "text", + "content": "[41] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 2, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 140, + 547, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 547, + 192 + ], + "type": "text", + "content": "[42] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 195, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 547, + 248 + ], + "type": "text", + "content": "[43] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. Nature, 596(7873):583-589, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 251, + 547, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 547, + 303 + ], + "type": "text", + "content": "[44] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 2, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 305, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 547, + 348 + ], + "type": "text", + "content": "[45] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 350, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 350, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 350, + 547, + 403 + ], + "type": "text", + "content": "[46] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 405, + 547, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 547, + 448 + ], + "type": "text", + "content": "[47] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 449, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 546, + 491 + ], + "type": "text", + "content": "[48] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 493, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 493, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 493, + 546, + 536 + ], + "type": "text", + "content": "[49] Jungbeom Lee, Jangho Lee, Sungmin Lee, and Sungroh Yoon. Mutual suppression network for video prediction using disentangled features. arXiv preprint arXiv:1804.04810, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 537, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 546, + 591 + ], + "type": "text", + "content": "[50] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 593, + 546, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 546, + 668 + ], + "type": "text", + "content": "[51] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 546, + 713 + ], + "type": "text", + "content": "[52] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Efficient multi-order gated aggregation network. arXiv preprint arXiv:2211.03295, 2022. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18779" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Xiaodan Liang, Lisa Lee, Wei Dai, and Eric P Xing. Dual motion gan for future-flow embedded video prediction. In proceedings of the IEEE international conference on computer vision, pages 1744-1752, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 172 + ], + "type": "text", + "content": "[54] Xiaoli Liu, Jianqin Yin, Jin Liu, Pengxiang Ding, Jun Liu, and Huaping Liu. TrajectoryCNN: a new spatio-temporal feature learning network for human motion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2133-2146, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 173, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 287, + 205 + ], + "type": "text", + "content": "[55] Zicheng Liu, Siyuan Li, Ge Wang, Cheng Tan, Lirong Wu, and Stan Z Li. Decoupled mixup for data-efficient learning. arXiv preprint arXiv:2203.10761, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 206, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 287, + 270 + ], + "type": "text", + "content": "[56] Zicheng Liu, Siyuan Li, Di Wu, Zihan Liu, Zhiyuan Chen, Lirong Wu, and Stan Z Li. Automix: Unveiling the power of mixup for stronger classifiers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIV, pages 441-458. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 326 + ], + "type": "text", + "content": "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 287, + 358 + ], + "type": "text", + "content": "[58] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 360, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 287, + 403 + ], + "type": "text", + "content": "[59] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE International Conference on Computer Vision, pages 4463-4471, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 404, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 468 + ], + "type": "text", + "content": "[60] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning, pages 4114-4124. PMLR, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 287, + 514 + ], + "type": "text", + "content": "[61] William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video prediction and unsupervised learning. In International Conference on Learning Representations, 2017. 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 515, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 515, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 287, + 557 + ], + "type": "text", + "content": "[62] Arthur Szlam Marc'Aurelio Ranzato, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. CoRR, abs/1412.6604, 2, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 602 + ], + "type": "text", + "content": "[63] Emile Mathieu, Tom Rainforth, Nana Siddharth, and Yee Whye Teh. Disentangling disentanglement in variational autoencoders. In International Conference on Machine Learning, pages 4402-4412. PMLR, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "text", + "content": "[64] Michael Mathieu, Camille Couprie, and Yann LeCun. Deep multi-scale video prediction beyond mean square error. arXiv preprint arXiv:1511.05440, 2015. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 689 + ], + "type": "text", + "content": "[65] Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[66] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "In European conference on computer vision, pages 69-84. Springer, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 140 + ], + "type": "text", + "content": "[67] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 312, + 141, + 545, + 206 + ], + "type": "text", + "content": "[68] Sergiu Oprea, Pablo Martinez-Gonzalez, Alberto Garcia-Garcia, John Alejandro Castro-Vargas, Sergio Orts-Escalano, Jose Garcia-Rodriguez, and Antonis Argyros. A review on deep learning techniques for video prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 262 + ], + "type": "text", + "content": "[69] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 308 + ], + "type": "text", + "content": "[70] Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. Random feature attention. In International Conference on Learning Representations, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 309, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 545, + 364 + ], + "type": "text", + "content": "[71] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 545, + 420 + ], + "type": "text", + "content": "[72] Fitsum A Reda, Guilin Liu, Kevin J Shih, Robert Kirby, Jon Barker, David Tarjan, Andrew Tao, and Bryan Catanzaro. Sdc-net: Video prediction using spatially-displaced convolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 718-733, 2018. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 545, + 475 + ], + "type": "text", + "content": "[73] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 478, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 478, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 312, + 478, + 545, + 531 + ], + "type": "text", + "content": "[74] Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, et al. Deep learning and process understanding for data-driven earth system science. Nature, 566(7743):195-204, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 545, + 578 + ], + "type": "text", + "content": "[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28:91-99, 2015. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 579, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 579, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 579, + 545, + 632 + ], + "type": "text", + "content": "[76] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., volume 3, pages 32-36. IEEE, 2004. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "text", + "content": "[77] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[78] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18780" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "type": "text", + "content": "network: A machine learning approach for precipitation nowcasting. Advances in Neural Information Processing Systems, 28, 2015. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 287, + 162 + ], + "type": "text", + "content": "[79] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 163, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 163, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 288, + 195 + ], + "type": "text", + "content": "[80] Osamu Shouno. Photo-realistic video prediction on natural videos of largely changing frames. arXiv preprint arXiv:2003.08635, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 287, + 241 + ], + "type": "text", + "content": "[81] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 242, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 288, + 296 + ], + "type": "text", + "content": "[82] Jiahao Su, Wonmin Byeon, Jean Kossaifi, Furong Huang, Jan Kautz, and Anima Anandkumar. Convolutional tensortrain lstm for spatio-temporal learning. Advances in Neural Information Processing Systems, 33:13714-13726, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 331 + ], + "type": "text", + "content": "[83] Cheng Tan, Zhangyang Gao, and Stan Z Li. Rfold: Towards simple yet effectiverna secondary structure prediction.arXiv preprint arXiv:2212.14041, 2022.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 332, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 288, + 364 + ], + "type": "text", + "content": "[84] Cheng Tan, Zhangyang Gao, and Stan Z Li. Target-aware molecular graph generation. arXiv preprint arXiv:2202.04829, 2022.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 365, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 287, + 410 + ], + "type": "text", + "content": "[85] Cheng Tan, Zhangyang Gao, Lirong Wu, Siyuan Li, and Stan Z Li. Hyperspherical consistency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7244-7255, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 455 + ], + "type": "text", + "content": "[86] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 287, + 498 + ], + "type": "text", + "content": "[87] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 500, + 287, + 544 + ], + "type": "text", + "content": "[88] Ruben Villegas, Dumitru Erhan, Honglak Lee, et al. Hierarchical long-term video prediction without supervision. In International Conference on Machine Learning, pages 6038-6046. PMLR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 287, + 599 + ], + "type": "text", + "content": "[89] Ruben Villegas, *Arkanath Pathak*, Harini Kannan, *Dumitru Erhan*, Quoc V Le, and Honglak Lee. *High fidelity video prediction with large stochastic recurrent neural networks*. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 287, + 645 + ], + "type": "text", + "content": "[90] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. In International Conference on Learning Representations, 2017. 5, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "text", + "content": "[91] Pichao Wang, Wanqing Li, Philip Ogunbona, Jun Wan, and Sergio Escalera. Rgb-d-based human motion recognition with deep learning: A survey. Computer Vision and Image Understanding, 171:118-139, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[92] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uni" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "formity on the hypersphere. In International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 96, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 151 + ], + "type": "text", + "content": "[93] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrmn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 196 + ], + "type": "text", + "content": "[94] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 251 + ], + "type": "text", + "content": "[95] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in Neural Information Processing Systems, 30, 2017. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 545, + 297 + ], + "type": "text", + "content": "[96] Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jianmin Wang, Philip S Yu, and Mingsheng Long. Predrnn: A recurrent neural network for spatiotemporal predictive learning. arXiv preprint arXiv:2103.09504, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 299, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 299, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 299, + 545, + 364 + ], + "type": "text", + "content": "[97] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154–9162, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 365, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 545, + 431 + ], + "type": "text", + "content": "[98] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 434, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 545, + 498 + ], + "type": "text", + "content": "[99] Haixu Wu, Jialong Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Flowformer: Linearizing transformers with conservation flows. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 24226-24242. PMLR, 17-23 Jul 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[100] Haixu Wu, Zhiyu Yao, Jianmin Wang, and Mingsheng Long. Motionrnn: A flexible model for video prediction with spacetime-varying motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15435-15444, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "text", + "content": "[101] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 643 + ], + "type": "text", + "content": "[102] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "text", + "content": "[103] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. In International Conference on Learning Representations, 2019. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[104] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18781" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "text", + "content": "redundancy reduction. arXiv preprint arXiv:2103.03230, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[105] Jianjin Zhang, Yunbo Wang, Mingsheng Long, Wang Jian-min, and S Yu Philip. Z-order recurrent neural networks for video prediction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 230-235. IEEE, 2019. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 195 + ], + "type": "text", + "content": "[106] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 287, + 262 + ], + "type": "text", + "content": "[107] Liang Zhang, Guangming Zhu, Peiyi Shen, Juan Song, Syed Afaq Shah, and Mohammed Bennamoun. Learning spatiotemporal features using 3dCNN and convolutional LSTM for gesture recognition. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 3120-3128, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 287, + 297 + ], + "type": "text", + "content": "[108] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pages 649-666. Springer, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 287, + 342 + ], + "type": "text", + "content": "[109] Jiangbin Zheng, Yidong Chen, Chong Wu, Xiaodong Shi, and Suhail Muhammad Kamal. Enhancing neural sign language translation by highlighting the facial expression information. Neurocomputing, 464:462-472, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 342, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 342, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 342, + 287, + 407 + ], + "type": "text", + "content": "[110] Jiangbin Zheng, Yile Wang, Ge Wang, Jun Xia, Yufei Huang, Guojiang Zhao, Yue Zhang, and Stan Li. Using context-to-vector with graph retrofitting to improve word embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8154-8163, 2022. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18782" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_content_list.json b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..37f7ff57d2649294444ce83c6cd4271532dea239 --- /dev/null +++ b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_content_list.json @@ -0,0 +1,1569 @@ +[ + { + "type": "text", + "text": "Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving", + "text_level": 1, + "bbox": [ + 194, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lucas Nunes*", + "bbox": [ + 248, + 198, + 362, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xieyuanli Chen*", + "bbox": [ + 256, + 215, + 390, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Louis Wiesmann*", + "bbox": [ + 392, + 198, + 535, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jens Behley*", + "bbox": [ + 419, + 215, + 524, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rodrigo Marcuzzi*", + "bbox": [ + 563, + 198, + 718, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cyrill Stachniss\\*,\\*", + "bbox": [ + 555, + 215, + 710, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*University of Bonn, *University of Oxford, ‡Lamarr Institute", + "bbox": [ + 299, + 234, + 669, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{firstname.lastname}@igg.uni-bonn.de", + "bbox": [ + 323, + 253, + 640, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 292, + 313, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic perception is a core building block in autonomous driving, since it provides information about the drivable space and location of other traffic participants. For learning-based perception, often a large amount of diverse training data is necessary to achieve high performance. Data labeling is usually a bottleneck for developing such methods, especially for dense prediction tasks, e.g., semantic segmentation or panoptic segmentation. For 3D LiDAR data, the annotation process demands even more effort than for images. Especially in autonomous driving, point clouds are sparse, and objects appearance depends on its distance from the sensor, making it harder to acquire large amounts of labeled training data. This paper aims at taking an alternative path proposing a self-supervised representation learning method for 3D LiDAR data. Our approach exploits the vehicle motion to match objects across time viewed in different scans. We then train a model to maximize the point-wise feature similarities from points of the associated object in different scans, which enables to learn a consistent representation across time. The experimental results show that our approach performs better than previous state-of-the-art self-supervised representation learning methods when fine-tuning to different downstream tasks. We furthermore show that with only $10\\%$ of labeled data, a network pre-trained with our approach can achieve better performance than the same network trained from scratch with all labels for semantic segmentation on SemanticKITTI.", + "bbox": [ + 75, + 314, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic perception is essential for safe interaction between autonomous vehicles and their surrounding. For learning-based perception, a massive amount of training data is usually required for training high-performance models. However, the data annotation is the bottleneck of collecting such large training sets, especially for dense prediction tasks, such as semantic segmentation [60, 80], instance", + "bbox": [ + 75, + 766, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a475f1413dde086649b200520421f2454fdc05c1f21b7c8d2a8d298cbddb7d0f.jpg", + "image_caption": [ + "Figure 1. Our pre-training (TARL) can reduce the amount of necessary labels during fine-tuning on SemanticKITTI [4, 22]. Our method requires only $10\\%$ of labels to surpass the network trained from scratch with the full dataset for semantic segmentation. For panoptic segmentation, our method requires only half of the labels." + ], + "image_footnote": [], + "bbox": [ + 501, + 292, + 696, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f093a366a4e33cb73fc9af0e7482fcf8d2953e25f8684f8e04d54414f1b860f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 292, + 890, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segmentation [46, 73], and panoptic segmentation [45, 79], which require fine-grained labels. In the context of autonomous driving, recent approaches exploit 3D LiDAR data [20, 54, 55, 61], where the data annotation process is more complex than on 2D RGB images due to the sparsity of the point cloud and object appearance varying with its distance to the sensor.", + "bbox": [ + 496, + 566, + 890, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent self-supervised representation learning methods [10, 11, 13-15, 25, 27, 76] tackle the lack of annotated data with a pre-training step requiring no labels. Those methods use data augmentation to generate different views from one data sample and train the network to learn an embedding space able to have similar representations for the generated augmented views. Other approaches [52, 63, 65, 69, 78] propose optimizing the pixel embedding space to learn a dense representation suited to be fine-tuned to more fine-grained tasks. For 3D point cloud data, recent approaches [21, 42, 49, 56] focus on synthetic point clouds of single objects to learn a robust representation for object classification. Other approaches [31, 36, 50, 67, 74, 75] target real-world data representation, such as LiDAR or RGB-D data, but fewer target autonomous driving scenarios.", + "bbox": [ + 496, + 672, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code: https://github.com/PRBonn/TARL", + "bbox": [ + 94, + 886, + 377, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5217", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel temporal association representation learning (TARL) method for 3D LiDAR data designed for autonomous driving data. We exploit the vehicle motion to extract different views of the same objects across time. Then, we compute point-wise features from the objects in the point cloud and use a Transformer encoder [2] as a projection head to learn a temporal association from the object representation, embedding the objects dynamics. We conduct extensive experiments to compare our approach with state-of-the-art methods and show that our approach surpasses previous self-supervised pretraining methods [50, 67, 75] when fine-tuning to different downstream tasks and datasets [4, 8, 22, 64]. In summary, our key contributions are:", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel self-supervised pre-training for 3D LiDAR data able to learn a robust and temporally-consistent representation by clustering together points from the same object viewed at different points in time.", + "- We achieve better performance than previous state-of-the-art methods on different downstream tasks, i.e., semantic segmentation, panoptic segmentation, and object detection.", + "- With our pre-training, we require only $10\\%$ of labels to surpass the network trained from scratch for semantic segmentation using the whole training set on SemanticKITTI (see Fig. 1).", + "- Our self-supervised pre-training produces representations more suited for transfer learning than supervised pre-training, achieving better performance when finetuning to a different dataset." + ], + "bbox": [ + 94, + 316, + 467, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 587, + 218, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised representation learning aims at initializing weights of a deep neural network for a downstream task by learning suitable representations without labeled data. Early approaches [18, 26, 28, 33, 48, 70] use so-called pretext tasks to learn useful representations, such as solving jigsaw puzzles [28, 33, 48] or reconstructing masked data [18, 26, 70]. More recently, contrastive learning [13, 14, 27] has drawn significant attention in the vision and robotics research community. Those contrastive methods employ random data augmentation over one image sample to generate different views of it. Then, the network is trained to learn an embedding space able to discriminate between the augmented samples from the same image and the augmented views from other images. Recent approaches exploit this discriminative pretext task and propose different ways to learn a self-supervised data representation, e.g., via non-contrastive methods [15, 25], redundancy reduction [76], or online clustering [10]. Other works [52, 63, 65, 69, 78] approach this in a pixel-wise man", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ner, learning dense representations applied to more fine-grained downstream tasks.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer encoders [2] were initially proposed in the natural language processing community achieving state-of-the-art performance in the field. Such architectures were then quickly adopted by the computer vision community [1, 30, 66, 72, 77]. With the attention mechanism, the Transformer can learn the global context from the image instead of just local neighboring context as in convolutional models [1]. With the growing interest in Transformers, some self-supervised representation learning methods were applied to Transformers [11, 59, 68] to evaluate the capacity of such models to learn meaningful representations.", + "bbox": [ + 496, + 123, + 890, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR point cloud perception brings further challenges compared to the image domain. Previous works [46, 47] deal with 3D scans by projecting the data to a 2D range image and then processing it with 2D convolutional methods. More recently, different convolution architectures were proposed to deal with 3D data, such as point kernels [54, 55, 61], sparse convolutions [20, 23, 24], or graph representations [37, 40, 62]. With those different architectures, it becomes possible to train a model to learn from 3D data and deal with different tasks in this domain, such as semantic segmentation [60,80], panoptic segmentation [3, 29, 44], object detection [38, 57, 58], or moving object segmentation [6, 16]. With increasing interest in Transformers, some works [12, 19, 34, 70] proposed ways of dealing with point cloud data in such architectures. In this domain, a key challenge lies in the number of points collected by LiDAR sensors together with the high memory requirements from the Transformer attention mechanism.", + "bbox": [ + 496, + 291, + 892, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D self-supervised representation learning aims at learning robust representations from point cloud data. Most of those methods focus on point clouds of single objects [21, 42,49,56]. Other methods [31,36,50,67,74,75] target dense point-wise representations for real-world point cloud scans. Xie et al. [67] propose learning a representation via point-to-point contrastive learning through matching corresponding points in two augmented point clouds. Similarly, Hou et al. [31] learn both point-wise correspondence and the region where the point is located in the point cloud. Zhang et al. [75] propose global representation learning through scan-to-scan discrimination. Chen et al. [74] synthetically add dynamic objects in the scene to embed temporal and dynamic understanding in the learned representation. Other methods propose discriminating regions of interest to extract more semantic information. Nunes et al. [50] extract coarse segments from point clouds and use contrastive loss to distinguish between objects segments. Similarly, Yin et al. [36] extract proposal regions to be discriminated, biasing the pre-training to object detection.", + "bbox": [ + 496, + 566, + 893, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the performance improvement brought by those 3D self-supervised representation learning methods, there is", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5218", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03c5c449c0b5312b5ae1b6082ef0e68336b71dac4d67ed8df76dcb0dbfb99322.jpg", + "image_caption": [ + "Figure 2. We aggregate $n$ point clouds and extract segments $S$ from it by removing the ground and clustering the remaining points. Then we sample two point clouds $\\mathcal{P}^{t_1}$ and $\\mathcal{P}^{t_2}$ , and apply random augmentations $\\mathcal{T}^{t_1}$ and $\\mathcal{T}^{t_2}$ to them. Next, we compute point-wise features for each scan and list its segments $S^{t_1}$ and $S^{t_2}$ with their corresponding point features. For $S^{t_2}$ we compute the mean feature vector for each segment and compute the segments target embeddings $\\overline{s}^{t_2}$ with a projection head. For $S^{t_1}$ we use a point-wise projector followed by a predictor to identify for each segment point features $s_{m,p}^{t_1}$ the corresponding target embedding $\\overline{s}_m^{t_2}$ by minimizing the loss $\\mathcal{L}^{t_1\\to t_2}$ ." + ], + "image_footnote": [], + "bbox": [ + 135, + 89, + 851, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "still a performance gap compared to the image domain due to the lack of data augmentations for point clouds. Similar to previous works [36,50], our approach focuses on learning representations over regions of interest. However, unlike previous methods, we extract natural augmented views of the objects in the scans collected at different points in time while the vehicle drives in the environment. Then, the network is trained to learn a common representation for those temporal views of one object. This enables our method to learn a more robust semantic representation by learning temporal consistent features and real-world dynamics.", + "bbox": [ + 75, + 349, + 472, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 525, + 169, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a new self-supervised representation learning method for LiDAR data obtained in the context of autonomous driving. Our approach requires only unlabeled point clouds and the corresponding scan pose. Pose information is usually readily available by means of GPS/IMU, odometry approaches [32, 53], or SLAM systems [5,17], thus, this is not a limitation in practice. We use a siamese network scheme with an online updated network and a momentum updated network (see Fig. 2). The whole pipeline consists of extracting objects as coarse segments viewed at different times over an interval of scans. Then, point-wise features are computed with the backbone and a Transformer encoder as a projector. Finally, we perform an implicit clustering to put points from different views of the same object together. In the following subsections, we explain the individual steps of our method thoroughly.", + "bbox": [ + 75, + 551, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal objects views", + "text_level": 1, + "bbox": [ + 76, + 801, + 292, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of only using data augmentation to generate artificial views of one object, we exploit the vehicle motion to extract real-world object segments viewed from different perspectives. Given the vehicle motion and the properties of LiDAR sensors, one object can have different ap", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9d6ea86d1691e11ba8e8aea2a9dde9b8082115d8af87298f2cf3c4bf4d41e3f2.jpg", + "image_caption": [ + "Figure 3. We aggregate $n$ point clouds from different times and extract coarse segments from the aggregated point cloud in an unsupervised manner. We keep the point indices mapping to extract views at different times from one object as augmented pairs." + ], + "image_footnote": [], + "bbox": [ + 521, + 354, + 870, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pearances depending on its position relative to the sensor. Such changing appearance is usually a problem when fine-tuning the model to data collected with different sensors [7, 35, 39, 71]. However, we exploit these properties in our favor to extract natural augmented views of one object. Previous works [31, 36, 50, 67, 74, 75] relied only on data augmentation to generate pairs of one scan, such as rotation, translation and scaling. Instead, we extract object views collected from different perspectives by the LiDAR sensor while the vehicle navigates through the environment. We then apply augmentations over the associated scans captured at a different time and train the network to learn a common representation for the object views.", + "bbox": [ + 496, + 609, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define a point cloud at time $t$ as $\\mathcal{P}^t = \\{\\pmb {p}_1^t,\\dots ,\\pmb {p}_R^t\\}$ as a set of 3D points $\\pmb {p}_r\\in \\mathbb{R}^3$ . To extract coarse object segments in an individual scan, we first separate $\\mathcal{P}^t$ into ground $\\mathcal{G}^t$ and non-ground points $\\hat{\\mathcal{P}}^t$ in an unsupervised manner using the method proposed by Lim et al. [41], where $\\mathcal{P}^t = \\hat{\\mathcal{P}}^t\\cup \\mathcal{G}^t$ and $\\hat{\\mathcal{P}}^t\\cap \\mathcal{G}^t = \\emptyset$ .Af", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5219", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a7dab86986ef2f1609875e06913fafd6596bd7a787b255eb8a5459a0bb303670.jpg", + "image_caption": [ + "Figure 4. Batch example, for an interval of $n = 6$ point clouds $\\mathcal{P}$ . We divide a sequence of $N$ scans into batches $\\mathcal{B}_b$ where we sample $t_1$ from the beginning and $t_2$ from the end of the interval. Then, the next batch starts in the middle of the previous interval." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 450, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "terwards, we cluster the non-ground points $\\hat{\\mathcal{P}}^t$ into $M$ segments $\\mathcal{S}^t = \\{S_1^t,\\dots,S_M^t\\}$ using HDBSCAN [9], where $\\hat{\\mathcal{P}}^t = \\bigcup_{m = 1}^{M}\\mathcal{S}_m^t$", + "bbox": [ + 75, + 253, + 467, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To map different segment views at different times, we define an interval of $n$ scans from which the views of the object will be extracted. Those scans are transformed to a common global coordinate frame to be then aggregated. The global scan poses can be easily acquired with GPS/IMU, SLAM systems [5, 17], or Li-DAR odometry [32, 53]. The aggregated point cloud $\\mathcal{P}$ is given by $\\mathcal{P} = \\{\\mathcal{P}^{t + 1},\\mathcal{P}^{t + 2},\\dots,\\mathcal{P}^{t + n}\\}$ . Similarly, we can aggregate the individual ground segmentation labels $\\mathcal{G} = \\{\\mathcal{G}^{t + 1},\\mathcal{G}^{t + 2},\\dots,\\mathcal{G}^{t + n}\\}$ and get the aggregated non-ground points $\\hat{\\mathcal{P}} = \\{\\hat{\\mathcal{P}}^{t + 1},\\hat{\\mathcal{P}}^{t + 2},\\dots,\\hat{\\mathcal{P}}^{t + n}\\}$ . As in the individual scan case, we can cluster $\\hat{\\mathcal{P}}$ to get the $M$ segments $S = \\{S_{1},\\dots,S_{M}\\}$ . By keeping the point index mapping from the aggregated point cloud to the individual $n$ scans, we can identify the $n$ segments of the same object viewed at different times as $S_{m} = \\{S_{m}^{t + 1},\\dots,S_{m}^{t + n}\\}$ . We then list the temporal views from each of the $M$ segments as $S_{1:M} = \\{S_1^{t + 1},\\dots,S_1^{t + n},\\dots,S_M^{t + 1},\\dots,S_M^{t + n}\\}$ .", + "bbox": [ + 75, + 303, + 467, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To properly segment views of both static and moving objects, we assume a LiDAR at the commonly used frequency of $10\\mathrm{Hz}$ , where the scan overlap is enough to cluster together also moving objects, as depicted in Fig. 3. We show further examples in the supplementary material. With this procedure, our augmented pairs are views of the same object at different times from static and moving objects. By maximizing the similarity between the object views at different times, the network needs to learn a more general representation consistent across time and embed the object dynamics. In the supplementary material, we provide an experimental comparison between the pre-training using the temporal object views as augmented pairs and with augmented pairs generated only with data augmentation to validate the use of temporal views.", + "bbox": [ + 75, + 577, + 467, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal batch", + "text_level": 1, + "bbox": [ + 76, + 815, + 235, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During pre-training, we need to sample a subset of $n$ consecutive scans from the training sequence to be aggregated and extract the temporal object views as explained in Sec. 3.1. This sample will be used during pre-training", + "bbox": [ + 75, + 839, + 467, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to learn a temporal consistent representation for the segmented objects. For a sequence of $N$ point clouds, we divide it into batches $\\mathcal{B}$ with intervals of $n$ scans. During the pre-training forward pass, we sample two scans at different times, $t_1$ and $t_2$ . Since the goal of this temporal sampling is to have different views of the same object, we enforce this by sampling $t_1$ from the beginning and $t_2$ from the end of this $n$ scans interval such that $t_1 < \\frac{n}{3}$ and $t_2 \\geq 2\\frac{n}{3}$ . In this case, the scans between $\\frac{n}{3} \\leq t < 2\\frac{n}{3}$ would never be sampled. To also process those scans, the next batch starts at this unseen interval. Therefore, for a batch sample $\\mathcal{B}_b$ the $n$ scans from the sequence of $N$ point clouds are $\\mathcal{B}_b = \\{\\mathcal{P}^k | b\\frac{n}{3} \\leq k < b\\frac{n}{3} + n\\}$ .", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Figure 4, we show how the batches would look like in an example with $n = 6$ scans interval. Even though we do not sample $t$ from $\\frac{n}{3} \\leq t < 2\\frac{n}{3}$ , we still need to aggregate and cluster these scans to correctly segment dynamic objects and match their correspondences between the scans.", + "bbox": [ + 496, + 287, + 890, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this sampling scheme, we guarantee that the objects views will be distant in time and still all the data will be seen during pre-training.", + "bbox": [ + 496, + 364, + 890, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Implicit clustering", + "text_level": 1, + "bbox": [ + 500, + 421, + 678, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With our coarse object segments $S$ , we have the prior that points from the same segment should be semantically similar. Given that we rely on a set of object segments, one straightforward way to distinguish the learned embeddings is to cluster together point features from the same object. Recent methods propose an online clustering scheme to separate samples around a fixed number of prototype learnable cluster centers [10]. We aim at clustering point features from the same object close together but in a more straightforward way with an implicit clustering.", + "bbox": [ + 496, + 445, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the temporal sampled pair $\\mathcal{P}^{t_1}$ and $\\mathcal{P}^{t_2}$ from a batch sample $\\mathcal{B}_b$ , we compute point-wise features $\\mathcal{F}^{t_1}$ and $\\mathcal{F}^{t_2}$ with the backbone. As the target embedding, we list from $\\mathcal{F}^{t_2}$ the set of $M$ segments $\\mathcal{S}^{t_2}$ . We compute for each segment a mean representation from its point-wise features, and project this embedding with a self-attention Transformer encoder as a projection head to get the $M$ target mean feature vectors $\\overline{\\boldsymbol{s}}^{t_2} \\in \\mathbb{R}^{M \\times D}$ where $D$ corresponds to the feature dimension.", + "bbox": [ + 496, + 597, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The segments $S^{t_1}$ are listed from the point-wise features $\\mathcal{F}^{t_1}$ . For $S^{t_1}$ , we do not compute mean embeddings but keep the features at point level, using the Transformer encoder to compute point-wise intra-class correspondences. To deal with the attention mechanism large memory requirement, we set a maximum number of $P$ points per segment, which are randomly sampled. After sampling $P$ points for each segment, we input the segments $S^{t_1}$ pointwise features to the Transformer projection head, followed by another self-attention Transformer encoder as a predictor. We then get for each segment the $P$ point-wise feature", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5220", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e51d1ef9df694a066e4299e8c7d050aaace530c05635b5c3bef02f6b7496c6c2.jpg", + "image_caption": [ + "Figure 5. Diagram of the implicit clustering scheme. We pool the segments $S^{t_1}$ and $S^{t_2}$ from the backbone features. For $t_2$ , we compute a mean representation for each segment and project it with the Transformer projection head. For $t_1$ , we project the pointwise features and then use the Transformer predictor to predict the corresponding mean representation from $t_2$ , clustering the point-wise features close to its mean target representation." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 449, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vectors $s^{t_1} \\in \\mathbb{R}^{M \\times P \\times D}$ .", + "bbox": [ + 76, + 362, + 243, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the segment target mean representations $\\overline{s}^{t_2}$ and the predicted point-wise feature vectors $s^{t_1}$ for each segment, we compute the loss to minimize the differences between the point features and the corresponding segment mean representation. For each point $p$ from a segment $m$ , we compute the temperature-scaled cosine similarity $\\delta_{m,p,k}^{t_1\\to t_2}$ between the normalized point-wise embedding $s_{m,p}^{t_1}\\in \\mathbb{R}^D$ and the mean representation $\\overline{s}_k^{t_2}\\in \\mathbb{R}^D$ from a segment $k$ as:", + "bbox": [ + 75, + 378, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {m, p, k} ^ {t _ {1} \\rightarrow t _ {2}} = \\frac {\\left(\\boldsymbol {s} _ {m , p} ^ {t _ {1}}\\right) ^ {\\top} \\bar {\\boldsymbol {s}} _ {k} ^ {t _ {2}}}{\\tau}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 517, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we use the cross-entropy loss to maximize the similarity between each point $p$ from a segment $m$ and the corresponding target segment mean representation as follows:", + "bbox": [ + 76, + 558, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {m} ^ {t _ {1} \\rightarrow t _ {2}} = - \\sum_ {p = 1} ^ {P} \\log \\left(\\frac {\\exp \\left(\\delta_ {m , p , m} ^ {t _ {1} \\rightarrow t _ {2}}\\right)}{\\sum_ {k} ^ {M} \\exp \\left(\\delta_ {m , p , k} ^ {t _ {1} \\rightarrow t _ {2}}\\right)}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 616, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compute this loss for all the $M$ segments and sum it to get the loss for predicting segments from $t_2$ with the points of the segments from $t_1$ :", + "bbox": [ + 76, + 666, + 468, + 713 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {t _ {1} \\rightarrow t _ {2}} = \\sum_ {m = 1} ^ {M} l _ {m} ^ {t _ {1} \\rightarrow t _ {2}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 724, + 468, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this formulation, our task is to predict for each point in the segment at time $t_1$ the corresponding segment at $t_2$ . Since the target for all the $P$ point embeddings $s_{m,p}^{t_1}$ from a segment $m$ is the same mean segment representation $\\overline{s}_m^{t_2}$ , the loss will push all the points from a segment to converge to a mean representation while separating from other segments, implicitly clustering together points from the same object as shown in Fig. 5.", + "bbox": [ + 75, + 777, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we repeat the forward pass swapping $t_1$ and $t_2$ to have a symmetric representation, learning both the correspondences from $t_1 \\rightarrow t_2$ and from $t_2 \\rightarrow t_1$ leading to:", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T A R L} = \\mathcal {L} ^ {t _ {1} \\rightarrow t _ {2}} + \\mathcal {L} ^ {t _ {2} \\rightarrow t _ {1}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 150, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the next section we give an intuition behind our choice of projection head, and how it can help the model to learn a more fine-grained representation during pre-training.", + "bbox": [ + 498, + 172, + 890, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Transformer projection", + "text_level": 1, + "bbox": [ + 500, + 226, + 718, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised representation learning methods [13, 14, 27] typically add a non-linear projection head to the backbone to project the embeddings to the target feature space. The idea is that the backbone should learn general features, and this projection head will overfit to the pre-training pretext task and later be discarded during fine-tuning. Our projection scheme follows more recent methods [15, 25], which shows that an asymmetric network can improve the learned representation. Besides, with the prior that points from one segment should be similar, we want to focus on the pointwise relationship. Therefore, we replace the non-linear projection head with a self-attention Transformer encoder.", + "bbox": [ + 496, + 250, + 890, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our siamese network scheme, the online updated network computes point-wise features. In this case, the queries, keys, and values for our Transformer projector will be point-wise features. We aim to learn intra-class features with the attention mechanism and identify correspondences between the points from one segment. In contrast, the momentum updated network computes a mean feature vector for each segment. In this case, the attention mechanism should look for similarities and differences within the different objects segments in the scene. By using the self-attention over points and segments, we aim at learning at the same time point-wise correspondences while also learning the differences between the different segments.", + "bbox": [ + 496, + 431, + 892, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the Transformer projector, the pre-training can lead the learned features towards a more fine-grained representation, where intra-class point-wise features are matched with the individual segment features. The supplementary material provides ablations comparing the Transformer with the commonly used non-linear projection head.", + "bbox": [ + 496, + 627, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Pre-training overview", + "text_level": 1, + "bbox": [ + 500, + 726, + 702, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secs. 3.1 to 3.4 explain the individual steps of our approach. This section summarizes the entire pre-training pipeline, putting together the modules explained in the previous subsections.", + "bbox": [ + 496, + 750, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our pipeline shown in Fig. 2 fetches a batch sample $\\mathcal{B}_b$ of $n$ sequential scans and applies a transformation to the scans $\\mathcal{P}^t$ with the corresponding poses to have all points in a common coordinate frame. We use an unsupervised ground segmentation approach to remove the ground $\\mathcal{G}$ . Next, we cluster the remaining points $\\hat{\\mathcal{P}}$ , defining coarse", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5221", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0016703c8be4c70bfe0a50b1d47b8bf1361de9fdd40e7da83ee4cad8e25dba0f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScribbles0.1%1%10%50%100%
Scratch54.9629.3542.7753.9658.2759.03
PointContrast [67]54.5232.6344.6258.6859.9861.45
DepthContrast [75]55.9031.6648.0557.1160.9961.14
SegContrast [50]56.7032.7544.8356.3160.4561.02
TARL (Ours)57.2538.5951.4260.3461.4261.47
", + "bbox": [ + 78, + 88, + 470, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "segments $S$ of the objects in the scene. We sample two random scans at different times $\\mathcal{P}^{t_1}$ and $\\mathcal{P}^{t_2}$ from $\\mathcal{B}_b$ , in which we apply random augmentations $\\mathcal{T}^{t_1}$ and $\\mathcal{T}^{t_2}$ . We compute point-wise features $\\mathcal{F}^{t_1}$ and $\\mathcal{F}^{t_2}$ from the sampled point clouds with the backbone, and list the segments $S^{t_1}$ and $S^{t_2}$ with its corresponding point-wise features. We calculate a mean embedding $\\overline{\\mathcal{S}}^{t_2}$ for each segment $m$ in $\\mathcal{P}^{t_2}$ and compute the target representation $\\overline{s}^{t_2}$ for each segment with a Transformer projection head. For the segments $S^{t_1}$ , we project the point-wise features with another Transformer projector followed by a Transformer encoder as a predictor to get the point-wise features $s^{t_1}$ . Lastly, we compute the loss to predict the corresponding segment representation at $t_2$ for each point in $t_1$ . We repeat the process changing $t_1$ and $t_2$ to symmetrically match points from scan $t_2$ to the corresponding segment at $t_1$ .", + "bbox": [ + 75, + 252, + 472, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 511, + 209, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use the SemanticKITTI dataset [4,22] for our pre-training, an autonomous driving LiDAR data benchmark with point-wise annotations. During pre-training, we use only the raw point clouds from the training sequences and the given poses to extract the temporal objects views. For fine-tuning, we use SemanticKITTI with the full scans annotations and the scribbles annotations [51]. Besides SemanticKITTI, we also use nuScenes [8,64] for fine-tuning, to evaluate our approach in different autonomous driving datasets collected with different LiDARs and sensor setups, reporting the results on the validation sets.", + "bbox": [ + 75, + 537, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model architecture. For our evaluation, we use a MinkUNet [20] as the backbone, which voxelizes the point clouds and uses sparse convolutions to extract features from the point cloud. The backbone receives the point cloud coordinates and intensity as input, and the output dimension feature has dimension $D = 96$ . We use a multi-head self-attention encoder for the projection head with one layer and eight attention heads. For the predictor, we use the same multi-head self-attention encoder architecture. Further details on the backbone used in our experiments are provided in the supplementary material.", + "bbox": [ + 75, + 704, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training. For pre-training, we use the AdamW optimizer [43] with a learning rate of $2 \\cdot 10^{-4}$ and decay of $10^{-4}$ ,", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/48b91ffa59d4b9cb5c0b26eefe54f3d61662287cd5720ea6a6aa59127355b895.jpg", + "table_caption": [ + "Table 1. Results (mIoU) of the models pre-trained on SemanticKITTI fine-tuned to semantic segmentation on SemanticKITTI with different percentage of labels and scribbles." + ], + "table_footnote": [], + "table_body": "
MethodMiniFull
Scratch26.9466.03
Supervised pre-training38.3967.35
PointContrast [67]31.9267.31
DepthContrast [75]27.8164.70
SegContrast [50]31.2767.70
TARL (Ours)39.3668.26
", + "bbox": [ + 542, + 88, + 846, + 214 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7b14a9015fd9ff46fb3043636ff7cf4fdca64264f14eca06aada0a6b3b039ea9.jpg", + "table_caption": [ + "Table 2. Results (mIoU) of the self-supervised and supervised pretrained models on SemanticKITTI fine-tuned to semantic segmentation on nuScenes mini and full training sets." + ], + "table_footnote": [], + "table_body": "
MethodKITTInuScenes
Frozen backbone4.105.62
PointContrast [67]25.5319.70
DepthContrast [75]13.879.65
SegContrast [50]27.4223.75
TARL (Ours)35.9026.08
", + "bbox": [ + 540, + 275, + 846, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Linear evaluation results (mIoU) on SemanticKITTI and nuScenes datasets for semantic segmentation.", + "bbox": [ + 498, + 387, + 890, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "training for 200 epochs with batch size 8 using a single NVIDIA RTX A6000. We use $n = 12$ for the scans to be aggregated. We provide experiments with different numbers of scans in the supplementary material to support this choice. The voxel resolution is set to $0.05m$ for the input point clouds, from which we sample a maximum of 40,000 points per point cloud. During the segment pooling, we limit it to a total of $M = 50$ segments with a maximum of $P = 300$ points per segment to avoid memory overflow. We use $\\tau = 0.1$ to compute $\\delta$ in Eq. (1) and momentum of 0.999 to update the momentum network according to the online network weights. For the baselines, we use their official repositories with their default parameters, also training them for 200 epochs and sampling 40,000 points per scan.", + "bbox": [ + 496, + 429, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fine-tuning. To evaluate our method and compare it with the baselines, we fine-tune the models to three downstream tasks: semantic segmentation, panoptic segmentation, and object detection. For semantic segmentation, we use the pipeline provided in SegContrast [50] with the AdamW optimizer [43] and a learning rate of 0.001. For panoptic segmentation, we use the baseline evaluated by Hong et al. [29], where a 3D backbone network is used together with a semantic and an instance heads, followed by a clustering post-processing to identify the instances. We use the pre-trained MinkUNet model as the 3D backbone. We use a learning rate of 0.2 with batch size 8, training for 50 epochs. For object detection, we use the same toolbox and hyperparameters used in DepthContrast [75] with the PartA2 detector [57]. For all the tasks, we evaluate the method with the same model trained from scratch (without pre-training) and with the model after pre-training with", + "bbox": [ + 496, + 643, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5222", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/601e889d300fec87ff89183f76f5acb61ebe8fcbe6d57662b6f844e16e1ff9f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
0.1%1%10%50%100%
MethodPQIoUPQIoUPQIoUPQIoUPQIoU
Scratch4.7611.1322.7230.8447.2053.5355.3261.9455.4059.75
PointContrast [67]5.8611.5127.3732.4947.5754.6354.2159.4855.8561.49
DepthContrast [75]7.6513.5627.3132.3046.8551.2754.5559.6056.1560.81
SegContrast [50]7.5814.4626.1432.8547.0253.4755.3860.0456.7361.96
TARL (Ours)10.2617.0129.2434.7151.2757.5956.1062.3656.5762.05
", + "bbox": [ + 158, + 88, + 803, + 212 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b355f88648c8eee1f6c7168677f48770d8afc826628fd9d2e00d59e33d1aa465.jpg", + "table_caption": [ + "Table 4. Results (PQ and IoU) when fine-tuning the pre-trained models to panoptic segmentation with different percentage of labels on SemanticKITTI." + ], + "table_footnote": [], + "table_body": "
MiniFull
PQIoUPQIoU
Scratch23.7823.9652.9858.17
Supervised pre-training24.7723.6053.1958.05
PointContrast [67]26.5825.4651.0656.39
DepthContrast [75]28.6627.3051.5157.06
SegContrast [50]28.8426.7952.3157.24
TARL (Ours)32.2230.7353.2659.14
", + "bbox": [ + 93, + 257, + 447, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results (PQ and IoU) of the self-supervised and supervised pre-trained models on SemanticKITTI fine-tuned to panoramic segmentation on nuScenes mini and full training sets.", + "bbox": [ + 75, + 409, + 468, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the different self-supervised methods. The supplementary material provides detailed information for each downstream task training procedure.", + "bbox": [ + 75, + 458, + 468, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Semantic segmentation", + "text_level": 1, + "bbox": [ + 76, + 513, + 290, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this evaluation, we want to measure the information about semantics learned by the network during pretraining. We fine-tune it to semantic segmentation on SemanticKITTI and nuScenes after pre-training on SemanticKITTI. For SemanticKITTI, we use the percentage scans subsets used in SegContrast [50], with $0.1\\%$ , $1\\%$ , $10\\%$ , $50\\%$ , and $100\\%$ of the labeled scans. We also evaluate on the scribbles annotations [51], which labels just a subset of points for each scan in the SemanticKITTI training sequences. We report the mean intersection-over-union (mIoU) of the models on the validation set, while fine-tuned on the aforementioned subset of labels from the training set.", + "bbox": [ + 75, + 537, + 468, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 shows the results from our approach compared to the baselines. As can be seen, overall the pre-training methods boost the performance compared to the network without pre-training on both scribbles annotations and the subsets of labeled scans. However, our approach surpasses previous state-of-the-art methods in all the different subsets, with a bigger gap when fewer labeled scans are used to fine-tune the model. We can also notice that, after pre-training with our method, only $10\\%$ of labeled scans were necessary for the model to surpass the performance of the network trained from scratch with $100\\%$ of the labeled scans. This result suggests that our method can effectively reduce the amount", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of labeled data necessary for the semantic segmentation task by around ten times. In the supplementary material, we also provide the IoU for each individual semantic class.", + "bbox": [ + 496, + 250, + 890, + 294 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization. In this experiment, we evaluate the generalization of the learned features. We use the network pretrained on SemanticKITTI and fine-tune it on nuScenes. We use nuScenes full training set and the mini training subset to fine-tune the network, and evaluate on the full validation set. In this evaluation, we also compare the results of the fully supervised semantic segmentation pre-training on SemanticKITTI fine-tuned on nuScenes. As shown in Tab. 2, even though all methods can improve the performance of the model trained on nuScenes, our approach can surpass previous self-supervised methods by around $8\\%$ when training with fewer labels. Besides, compared with supervised pre-training of the network on SemanticKITTI, our approach achieves better performance on both full and mini training sets. These results suggest that our method can replace supervised pre-training for LiDAR data, since it achieved better performance when fine-tuning to a different dataset, collected with a different sensor setup.", + "bbox": [ + 496, + 294, + 892, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Linear evaluation. In this experiment, the pre-trained backbone is frozen, training only a linear head on top of it to evaluate how descriptive the self-supervised learned representation is without fine-tuning it to the target task. We perform this evaluation on SemanticKITTI and nuScenes datasets. In Tab. 3, the randomly initialized network (frozen backbone) achieves low performance, suggesting that the features extracted by the backbone play the main role in achieving semantic segmentation. All the pre-training methods improve the performance compared to the network without pre-training. However, our approach has a clear performance gap compared to the baselines, especially when training to the same dataset used for pre-training. These results suggest that the representation learned by our method can embed more semantic information, even though not using labels, achieving higher performance than previous state-of-the-art methods on both datasets.", + "bbox": [ + 496, + 566, + 892, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Panoptic segmentation", + "text_level": 1, + "bbox": [ + 500, + 832, + 710, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we fine-tune the pre-trained models for panoptic segmentation to also evaluate the instance-level features of our learned representation. Same as for semantic", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5223", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/edf95e05107f63f162dab780ebcbb6a8085aefa9343afb106795e64e9080b457.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CarPedestrianCyclist
EasyModerateHardEasyModerateHardEasyModerateHard
Scratch91.7582.1380.1667.6563.1357.7391.7474.9470.34
PointContrast [67]91.5581.7079.8867.3362.3256.5592.3375.2070.66
DepthContrast [75]91.9882.0480.0369.1363.9757.7192.6873.6770.10
SegContrast [50]92.0882.1080.1868.7263.7757.3691.2674.7070.28
TARL (Ours)92.0982.1580.1068.5064.0058.7492.6175.4470.97
", + "bbox": [ + 151, + 88, + 816, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Results (AP50) when fine-tuning the pre-trained models to object detection on KITTI full training set for car, pedestrian and cyclist classes on the easy, moderate and hard difficult levels.", + "bbox": [ + 75, + 219, + 893, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "segmentation, we fine-tuned the model to SemanticKITTI using the same percentage subsets and nuScenes using the full and the mini training sets, reporting both the mIoU and the panoptic quality (PQ) on the validation sets.", + "bbox": [ + 75, + 253, + 468, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 shows that our method is consistently better than previous self-supervised pre-training approaches. As the amount of training samples increases, the differences diminish. However, our method has a clear performance gap compared to the baselines when trained with fewer labels.", + "bbox": [ + 75, + 316, + 468, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization. To evaluate the generalization of our learned features on panoptic segmentation, we use the network pre-trained on SemantiKITTI to fine-tune it on nuScenes full and mini training sets, evaluating it on the full validation set. Tab. 5 shows that the representation learned by our method was more suited to be transferred to a different dataset. As in semantic segmentation, our method achieved better performance than previous self-supervised approaches and supervised pre-training on SemanticKITTI. These results agree with the semantic segmentation evaluation, suggesting that our method is better suited for transfer learning than the supervised pre-training.", + "bbox": [ + 75, + 393, + 470, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With these experiments, we can validate that our learned representations can extract not only semantic knowledge from the data but also instance-level information, surpassing previous methods in IoU and PQ metrics. In the supplementary material, we provide further results, reporting also the segmentation and recognition quality.", + "bbox": [ + 75, + 575, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Object detection", + "text_level": 1, + "bbox": [ + 76, + 678, + 238, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a third downstream task, we compare the methods fine-tuned for object detection to evaluate how general is the representation learned by our method compared with previous state-of-the-art methods. In this evaluation, we fine-tune the model with the whole training set of the KITTI dataset [22] and report the average precision $(\\mathrm{AP}_{50})$ with 40 recall positions for car, pedestrian, and cyclist classes on the three difficulty levels.", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 6, we can compare the baselines performance with our approach. For the car class, the pre-training methods achieve a marginal improvement compared to the network trained from scratch. For pedestrian and cyclist, the improvements brought by the pre-training are more substan", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tial. On those classes, our method achieves the best performance on moderate and hard difficulties, showing that our approach can also boost the performance when fine-tuning the model on object detection. In our supplementary material, we provide further experiments with the different training subsets used by Zhang et al. [75].", + "bbox": [ + 496, + 253, + 890, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Together with the evaluations of the other downstream tasks, these results suggest that our method can learn a general representation. Our method achieved better performance than previous approaches on all three downstream tasks, showing that our approach learns a robust representation, not biased to a specific task but showing generalizable representations suitable for different tasks and datasets.", + "bbox": [ + 496, + 345, + 890, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 460, + 617, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a self-supervised representation learning method for 3D LiDAR data evaluated in the context of autonomous driving. We exploit the vehicle motion to extract different views of objects across time to learn a temporally-consistent representation. We use a Transformer encoder as a projection head and implicitly clustered points from the same object together while discriminating between different objects. Our experiments show that our method achieves better performance than previous state-of-the-art approaches on different downstream tasks. Besides, our approach could reduce the amount of necessary labeled data to only $10\\%$ when fine-tuning for semantic segmentation on SemanticKITTI. These results show that our proposed approach could learn a general LiDAR point cloud representation, embedding semantic information by matching objects representation viewed at different times. In addition, our method achieved better performance than supervised pre-training when transferring the learned representation to a dataset collected with different LiDAR sensors, suggesting that our approach could replace supervised pretraining in the LiDAR data domain.", + "bbox": [ + 496, + 484, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Despite encouraging results, there is space for further improvement. For the unsupervised segment extraction, the clustering parameters have to be tuned depending on the LiDAR used to collect the data. Also, we rely on high-frequency data to properly segment dynamic objects. Otherwise, temporal segments could be wrongly matched.", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5224", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proc. of the Int. Conf. on Learning Representations (ICLR), 2021. 2", + "[2] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 2", + "[3] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyril Stachniss, Jens Behley, and Laura Leal-Taixe. 4D Panoptic LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2, 6", + "[5] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS), 2018. 3, 4", + "[6] Benedikt Mersch, Xieyuanli Chen, Ignacio Vizzo, Lucas Nunes, Jens Behley, and Cyril Stachniss. Receding Moving Object Segmentation in 3D LiDAR Data Using Sparse 4D Convolutions. IEEE Robotics and Automation Letters (RA-L), 7(3):7503-7510, 2022. 2", + "[7] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised Domain Adaptation for LiDAR Panoptic Segmentation. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3", + "[8] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6", + "[9] Ricardo J. G. B. Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Advances in Knowledge Discovery and Data Mining, 2013. 3", + "[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 4", + "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2", + "[12] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "PTTR: Relational 3D Point Cloud Object Tracking With Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In Proc. of the Int. Conf. on Machine Learning (ICML), 2020. 1, 2, 5", + "[14] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint, arxiv:2003.04297, 2020. 1, 2, 5", + "[15] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5", + "[16] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jirgen Gall, Jens Behley, and Cyril Stachniss. Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data. IEEE Robotics and Automation Letters (RA-L), 6:6529-6536, 2021. 2", + "[17] Xieyuanli Chen, Andres Milioto, Emanuele Palazzolo, Philippe Giguere, Jens Behley, and Cyril Stachniss. SuMa++: Efficient LiDAR-based Semantic SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 3, 4", + "[18] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[19] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PatchFormer: An Efficient Point Transformer With Patch Attention. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6", + "[21] Fayao Liu, Guosheng Lin, and Chuan-Sheng Foo. Point discriminative learning for unsupervised representation learning on 3d point clouds. arXiv preprint, arXiv:2108.02104, 2021. 1, 2", + "[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 1, 2, 6, 8", + "[23] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[24] Benjamin Graham and Laurens van der Maaten. Submanifold Sparse Convolutional Networks. arXiv preprint, arxiv:1706.01307, 2017. 2", + "[25] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and" + ], + "bbox": [ + 503, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5225", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 5", + "[26] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J. Kusner. Unsupervised Point Cloud Pre-Training via Occlusion Completion. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2", + "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5", + "[28] Hiroharu Kato and Tatsuya Harada. Image Reconstruction from Bag-of-Visual-Words. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2", + "[29] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. LiDAR-Based Panoptic Segmentation via Dynamic Shifting Network. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6", + "[30] Hongxu Yin, Arash Vahdat, Jose Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-ViT: Adaptive Tokens for Efficient Vision Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[31] Ji Hou, Benjamin Graham, Matthias Niessner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding With Contrastive Scene Contexts. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3", + "[32] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP - Simple, Accurate, and Robust Registration If Done the Right Way. arXiv preprint arXiv:2209.15397, 2022. 3, 4", + "[33] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[34] Ishan Misra, Rohit Girdharm, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2", + "[35] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. ST3D: Self-Training for Unsupervised Domain Adaptation on 3D Object Detection. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[36] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposalcontrast: Unsupervised pre-training for lidar-based 3d object detection. Proc. of the Europ. Conf. on Computer Vision (ECCV),2022.1,2,3", + "[37] Kai Fischer, Martin Simon, Florian Oelsner, Stefan Milz, Horst-Michael Gross, and Patrick Maeder. StickyPillars: Robust and Efficient Feature Matching on Point Clouds Using Graph Neural Networks. In Proc. of the IEEE/CVF" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[38] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. PointPillars: Fast Encoders for Object Detection From Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[39] Ferdinand Langer, Andres Milioto, Alexandre Haag, Jens Behley, and Cyril Stachniss. Domain Transfer for Semantic Segmentation of LiDAR Data using Deep Neural Networks. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 3", + "[40] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph Attention Convolution for Point Cloud Semantic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[41] Hyungtae Lim, Sungwon Hwang, and Hyun Myung. ERASOR: Egocentric Ratio of Pseudo Occupancy-Based Dynamic Object Removal for Static 3D Point Cloud Map Building. IEEE Robotics and Automation Letters (RA-L), 6(2):2272-2279, 2021. 3", + "[42] Ling Zhang and Zhigang Zhu. Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks. In Proc. of the Intl. Conf. on 3D Vision (3DV), 2019. 1, 2", + "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2019. 6", + "[44] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Contrastive Instance Association for 4D Panoptic Segmentation using Sequences of 3D LiDAR Scans. IEEE Robotics and Automation Letters (RA-L), 7(2):1550-1557, 2022. 2", + "[45] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. LiDAR Panoptic Segmentation for Autonomous Driving. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 1", + "[46] Andres Milioto, Leonard Mandtler, and Cyril Stachniss. Fast Instance and Semantic Segmentation Exploiting Local Connectivity, Metric Learning, and One-Shot Detection for Robotics. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 1, 2", + "[47] Andres Milioto and Cyril Stachniss. Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics using CNNs. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 2", + "[48] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-BERT: Vision-Language Pre-Training on Fashion Domain. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[49] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5226", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. SegContrast: 3D Point Cloud Feature Representation Learning through Self-supervised Segment Discrimination. IEEE Robotics and Automation Letters (RA-L), 2022. 1, 2, 3, 6, 7, 8", + "[51] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 7", + "[52] Pedro O. Pinheiro, Amjad Almahairi, Ryan Y. Benmalek, Florian Golemo, and Aaron Courville. Unsupervised learning of dense visual representations. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), pages 4489-4500, 2020. 1, 2", + "[53] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP Real-Time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3, 4", + "[54] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2", + "[55] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2", + "[56] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Proc. of the Europ. Conf. on Computer Vision (ECCV), pages 626-642, 2020. 1, 2", + "[57] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network. IEEE Trans. on Pattern Analalysis and Machine Intelligence (TPAMI), 43:2647-2664, 2021. 2, 6", + "[58] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointR-CNN: 3D Object Proposal Generation and Detection From Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[59] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-Level Representation Learning for Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[60] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2", + "[61] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPCov: Flexible and Deformable Convolution for Point Clouds. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2", + "[62] Weijing Shi and Ragunathan (Raj) Rajkumar. Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring Cross-Image Pixel Contrast for Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2", + "[64] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters (RA-L), 7(2):3795-3802, 2022. 2, 6", + "[65] Xiangyun Zhao, Raviteja Vemulapalli, Philip Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive Learning for Label Efficient Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2", + "[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[67] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 6, 7, 8", + "[68] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2", + "[69] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense Contrastive Learning for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[70] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[71] Li Yi, Boqing Gong, and Thomas Funkhouser. Complete & Label: A Domain Adaptation Approach to Semantic Segmentation of LiDAR Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[72] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. MPViT: Multi-Path Vision Transformer for Dense Prediction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[73] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust Instance Segmentation Through Reasoning About Multi-Object Occlusion. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[74] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. Proc. of the Europ. Conf. on Computer Vision (ECCV), 2022. 1, 2, 3", + "[75] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-Supervised Pretraining of 3D Features on any" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5227", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Point-Cloud. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2, 3, 6, 7, 8", + "[76] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In Proc. of the Int. Conf. on Machine Learning (ICML), 2021. 1, 2", + "[77] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2", + "[78] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring Set Similarity for Dense Self-Supervised Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[79] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-Free LiDAR Point Cloud Panoptic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[80] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 458 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5228", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_model.json b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99e6502c688204c5c33010562bd48a287175160b --- /dev/null +++ b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_model.json @@ -0,0 +1,2479 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.199, + 0.363, + 0.215 + ], + "angle": 0, + "content": "Lucas Nunes*" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.217, + 0.392, + 0.234 + ], + "angle": 0, + "content": "Xieyuanli Chen*" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.199, + 0.536, + 0.216 + ], + "angle": 0, + "content": "Louis Wiesmann*" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.217, + 0.526, + 0.234 + ], + "angle": 0, + "content": "Jens Behley*" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.199, + 0.719, + 0.216 + ], + "angle": 0, + "content": "Rodrigo Marcuzzi*" + }, + { + "type": "text", + "bbox": [ + 0.557, + 0.217, + 0.712, + 0.234 + ], + "angle": 0, + "content": "Cyrill Stachniss\\*,\\*" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.236, + 0.671, + 0.252 + ], + "angle": 0, + "content": "*University of Bonn, *University of Oxford, ‡Lamarr Institute" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.255, + 0.642, + 0.27 + ], + "angle": 0, + "content": "{firstname.lastname}@igg.uni-bonn.de" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.294, + 0.314, + 0.309 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Semantic perception is a core building block in autonomous driving, since it provides information about the drivable space and location of other traffic participants. For learning-based perception, often a large amount of diverse training data is necessary to achieve high performance. Data labeling is usually a bottleneck for developing such methods, especially for dense prediction tasks, e.g., semantic segmentation or panoptic segmentation. For 3D LiDAR data, the annotation process demands even more effort than for images. Especially in autonomous driving, point clouds are sparse, and objects appearance depends on its distance from the sensor, making it harder to acquire large amounts of labeled training data. This paper aims at taking an alternative path proposing a self-supervised representation learning method for 3D LiDAR data. Our approach exploits the vehicle motion to match objects across time viewed in different scans. We then train a model to maximize the point-wise feature similarities from points of the associated object in different scans, which enables to learn a consistent representation across time. The experimental results show that our approach performs better than previous state-of-the-art self-supervised representation learning methods when fine-tuning to different downstream tasks. We furthermore show that with only \\(10\\%\\) of labeled data, a network pre-trained with our approach can achieve better performance than the same network trained from scratch with all labels for semantic segmentation on SemanticKITTI." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.741, + 0.21, + 0.757 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Semantic perception is essential for safe interaction between autonomous vehicles and their surrounding. For learning-based perception, a massive amount of training data is usually required for training high-performance models. However, the data annotation is the bottleneck of collecting such large training sets, especially for dense prediction tasks, such as semantic segmentation [60, 80], instance" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.294, + 0.697, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.294, + 0.891, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.489, + 0.895, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Our pre-training (TARL) can reduce the amount of necessary labels during fine-tuning on SemanticKITTI [4, 22]. Our method requires only \\(10\\%\\) of labels to surpass the network trained from scratch with the full dataset for semantic segmentation. For panoptic segmentation, our method requires only half of the labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.672 + ], + "angle": 0, + "content": "segmentation [46, 73], and panoptic segmentation [45, 79], which require fine-grained labels. In the context of autonomous driving, recent approaches exploit 3D LiDAR data [20, 54, 55, 61], where the data annotation process is more complex than on 2D RGB images due to the sparsity of the point cloud and object appearance varying with its distance to the sensor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Recent self-supervised representation learning methods [10, 11, 13-15, 25, 27, 76] tackle the lack of annotated data with a pre-training step requiring no labels. Those methods use data augmentation to generate different views from one data sample and train the network to learn an embedding space able to have similar representations for the generated augmented views. Other approaches [52, 63, 65, 69, 78] propose optimizing the pixel embedding space to learn a dense representation suited to be fine-tuned to more fine-grained tasks. For 3D point cloud data, recent approaches [21, 42, 49, 56] focus on synthetic point clouds of single objects to learn a robust representation for object classification. Other approaches [31, 36, 50, 67, 74, 75] target real-world data representation, such as LiDAR or RGB-D data, but fewer target autonomous driving scenarios." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.378, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code: https://github.com/PRBonn/TARL" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5217" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "In this paper, we propose a novel temporal association representation learning (TARL) method for 3D LiDAR data designed for autonomous driving data. We exploit the vehicle motion to extract different views of the same objects across time. Then, we compute point-wise features from the objects in the point cloud and use a Transformer encoder [2] as a projection head to learn a temporal association from the object representation, embedding the objects dynamics. We conduct extensive experiments to compare our approach with state-of-the-art methods and show that our approach surpasses previous self-supervised pretraining methods [50, 67, 75] when fine-tuning to different downstream tasks and datasets [4, 8, 22, 64]. In summary, our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.318, + 0.468, + 0.378 + ], + "angle": 0, + "content": "- We propose a novel self-supervised pre-training for 3D LiDAR data able to learn a robust and temporally-consistent representation by clustering together points from the same object viewed at different points in time." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.383, + 0.468, + 0.442 + ], + "angle": 0, + "content": "- We achieve better performance than previous state-of-the-art methods on different downstream tasks, i.e., semantic segmentation, panoptic segmentation, and object detection." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.447, + 0.468, + 0.506 + ], + "angle": 0, + "content": "- With our pre-training, we require only \\(10\\%\\) of labels to surpass the network trained from scratch for semantic segmentation using the whole training set on SemanticKITTI (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.512, + 0.468, + 0.571 + ], + "angle": 0, + "content": "- Our self-supervised pre-training produces representations more suited for transfer learning than supervised pre-training, achieving better performance when finetuning to a different dataset." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.318, + 0.468, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.22, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Self-supervised representation learning aims at initializing weights of a deep neural network for a downstream task by learning suitable representations without labeled data. Early approaches [18, 26, 28, 33, 48, 70] use so-called pretext tasks to learn useful representations, such as solving jigsaw puzzles [28, 33, 48] or reconstructing masked data [18, 26, 70]. More recently, contrastive learning [13, 14, 27] has drawn significant attention in the vision and robotics research community. Those contrastive methods employ random data augmentation over one image sample to generate different views of it. Then, the network is trained to learn an embedding space able to discriminate between the augmented samples from the same image and the augmented views from other images. Recent approaches exploit this discriminative pretext task and propose different ways to learn a self-supervised data representation, e.g., via non-contrastive methods [15, 25], redundancy reduction [76], or online clustering [10]. Other works [52, 63, 65, 69, 78] approach this in a pixel-wise man" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ner, learning dense representations applied to more fine-grained downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Transformer encoders [2] were initially proposed in the natural language processing community achieving state-of-the-art performance in the field. Such architectures were then quickly adopted by the computer vision community [1, 30, 66, 72, 77]. With the attention mechanism, the Transformer can learn the global context from the image instead of just local neighboring context as in convolutional models [1]. With the growing interest in Transformers, some self-supervised representation learning methods were applied to Transformers [11, 59, 68] to evaluate the capacity of such models to learn meaningful representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.893, + 0.563 + ], + "angle": 0, + "content": "LiDAR point cloud perception brings further challenges compared to the image domain. Previous works [46, 47] deal with 3D scans by projecting the data to a 2D range image and then processing it with 2D convolutional methods. More recently, different convolution architectures were proposed to deal with 3D data, such as point kernels [54, 55, 61], sparse convolutions [20, 23, 24], or graph representations [37, 40, 62]. With those different architectures, it becomes possible to train a model to learn from 3D data and deal with different tasks in this domain, such as semantic segmentation [60,80], panoptic segmentation [3, 29, 44], object detection [38, 57, 58], or moving object segmentation [6, 16]. With increasing interest in Transformers, some works [12, 19, 34, 70] proposed ways of dealing with point cloud data in such architectures. In this domain, a key challenge lies in the number of points collected by LiDAR sensors together with the high memory requirements from the Transformer attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.868 + ], + "angle": 0, + "content": "3D self-supervised representation learning aims at learning robust representations from point cloud data. Most of those methods focus on point clouds of single objects [21, 42,49,56]. Other methods [31,36,50,67,74,75] target dense point-wise representations for real-world point cloud scans. Xie et al. [67] propose learning a representation via point-to-point contrastive learning through matching corresponding points in two augmented point clouds. Similarly, Hou et al. [31] learn both point-wise correspondence and the region where the point is located in the point cloud. Zhang et al. [75] propose global representation learning through scan-to-scan discrimination. Chen et al. [74] synthetically add dynamic objects in the scene to embed temporal and dynamic understanding in the learned representation. Other methods propose discriminating regions of interest to extract more semantic information. Nunes et al. [50] extract coarse segments from point clouds and use contrastive loss to distinguish between objects segments. Similarly, Yin et al. [36] extract proposal regions to be discriminated, biasing the pre-training to object detection." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Despite the performance improvement brought by those 3D self-supervised representation learning methods, there is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.09, + 0.852, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 2. We aggregate \\( n \\) point clouds and extract segments \\( S \\) from it by removing the ground and clustering the remaining points. Then we sample two point clouds \\( \\mathcal{P}^{t_1} \\) and \\( \\mathcal{P}^{t_2} \\), and apply random augmentations \\( \\mathcal{T}^{t_1} \\) and \\( \\mathcal{T}^{t_2} \\) to them. Next, we compute point-wise features for each scan and list its segments \\( S^{t_1} \\) and \\( S^{t_2} \\) with their corresponding point features. For \\( S^{t_2} \\) we compute the mean feature vector for each segment and compute the segments target embeddings \\( \\overline{s}^{t_2} \\) with a projection head. For \\( S^{t_1} \\) we use a point-wise projector followed by a predictor to identify for each segment point features \\( s_{m,p}^{t_1} \\) the corresponding target embedding \\( \\overline{s}_m^{t_2} \\) by minimizing the loss \\( \\mathcal{L}^{t_1\\to t_2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.518 + ], + "angle": 0, + "content": "still a performance gap compared to the image domain due to the lack of data augmentations for point clouds. Similar to previous works [36,50], our approach focuses on learning representations over regions of interest. However, unlike previous methods, we extract natural augmented views of the objects in the scans collected at different points in time while the vehicle drives in the environment. Then, the network is trained to learn a common representation for those temporal views of one object. This enables our method to learn a more robust semantic representation by learning temporal consistent features and real-world dynamics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.17, + 0.541 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.794 + ], + "angle": 0, + "content": "In this paper, we propose a new self-supervised representation learning method for LiDAR data obtained in the context of autonomous driving. Our approach requires only unlabeled point clouds and the corresponding scan pose. Pose information is usually readily available by means of GPS/IMU, odometry approaches [32, 53], or SLAM systems [5,17], thus, this is not a limitation in practice. We use a siamese network scheme with an online updated network and a momentum updated network (see Fig. 2). The whole pipeline consists of extracting objects as coarse segments viewed at different times over an interval of scans. Then, point-wise features are computed with the backbone and a Transformer encoder as a projector. Finally, we perform an implicit clustering to put points from different views of the same object together. In the following subsections, we explain the individual steps of our method thoroughly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.294, + 0.818 + ], + "angle": 0, + "content": "3.1. Temporal objects views" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Instead of only using data augmentation to generate artificial views of one object, we exploit the vehicle motion to extract real-world object segments viewed from different perspectives. Given the vehicle motion and the properties of LiDAR sensors, one object can have different ap" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.356, + 0.871, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.54, + 0.894, + 0.598 + ], + "angle": 0, + "content": "Figure 3. We aggregate \\( n \\) point clouds from different times and extract coarse segments from the aggregated point cloud in an unsupervised manner. We keep the point indices mapping to extract views at different times from one object as augmented pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.807 + ], + "angle": 0, + "content": "pearances depending on its position relative to the sensor. Such changing appearance is usually a problem when fine-tuning the model to data collected with different sensors [7, 35, 39, 71]. However, we exploit these properties in our favor to extract natural augmented views of one object. Previous works [31, 36, 50, 67, 74, 75] relied only on data augmentation to generate pairs of one scan, such as rotation, translation and scaling. Instead, we extract object views collected from different perspectives by the LiDAR sensor while the vehicle navigates through the environment. We then apply augmentations over the associated scans captured at a different time and train the network to learn a common representation for the object views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We define a point cloud at time \\(t\\) as \\(\\mathcal{P}^t = \\{\\pmb {p}_1^t,\\dots ,\\pmb {p}_R^t\\}\\) as a set of 3D points \\(\\pmb {p}_r\\in \\mathbb{R}^3\\) . To extract coarse object segments in an individual scan, we first separate \\(\\mathcal{P}^t\\) into ground \\(\\mathcal{G}^t\\) and non-ground points \\(\\hat{\\mathcal{P}}^t\\) in an unsupervised manner using the method proposed by Lim et al. [41], where \\(\\mathcal{P}^t = \\hat{\\mathcal{P}}^t\\cup \\mathcal{G}^t\\) and \\(\\hat{\\mathcal{P}}^t\\cap \\mathcal{G}^t = \\emptyset\\) .Af" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5219" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.451, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.184, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Figure 4. Batch example, for an interval of \\( n = 6 \\) point clouds \\( \\mathcal{P} \\). We divide a sequence of \\( N \\) scans into batches \\( \\mathcal{B}_b \\) where we sample \\( t_1 \\) from the beginning and \\( t_2 \\) from the end of the interval. Then, the next batch starts in the middle of the previous interval." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.468, + 0.304 + ], + "angle": 0, + "content": "terwards, we cluster the non-ground points \\(\\hat{\\mathcal{P}}^t\\) into \\(M\\) segments \\(\\mathcal{S}^t = \\{S_1^t,\\dots,S_M^t\\}\\) using HDBSCAN [9], where \\(\\hat{\\mathcal{P}}^t = \\bigcup_{m = 1}^{M}\\mathcal{S}_m^t\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.468, + 0.576 + ], + "angle": 0, + "content": "To map different segment views at different times, we define an interval of \\(n\\) scans from which the views of the object will be extracted. Those scans are transformed to a common global coordinate frame to be then aggregated. The global scan poses can be easily acquired with GPS/IMU, SLAM systems [5, 17], or Li-DAR odometry [32, 53]. The aggregated point cloud \\(\\mathcal{P}\\) is given by \\(\\mathcal{P} = \\{\\mathcal{P}^{t + 1},\\mathcal{P}^{t + 2},\\dots,\\mathcal{P}^{t + n}\\}\\). Similarly, we can aggregate the individual ground segmentation labels \\(\\mathcal{G} = \\{\\mathcal{G}^{t + 1},\\mathcal{G}^{t + 2},\\dots,\\mathcal{G}^{t + n}\\}\\) and get the aggregated non-ground points \\(\\hat{\\mathcal{P}} = \\{\\hat{\\mathcal{P}}^{t + 1},\\hat{\\mathcal{P}}^{t + 2},\\dots,\\hat{\\mathcal{P}}^{t + n}\\}\\). As in the individual scan case, we can cluster \\(\\hat{\\mathcal{P}}\\) to get the \\(M\\) segments \\(S = \\{S_{1},\\dots,S_{M}\\}\\). By keeping the point index mapping from the aggregated point cloud to the individual \\(n\\) scans, we can identify the \\(n\\) segments of the same object viewed at different times as \\(S_{m} = \\{S_{m}^{t + 1},\\dots,S_{m}^{t + n}\\}\\). We then list the temporal views from each of the \\(M\\) segments as \\(S_{1:M} = \\{S_1^{t + 1},\\dots,S_1^{t + n},\\dots,S_M^{t + 1},\\dots,S_M^{t + n}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.468, + 0.803 + ], + "angle": 0, + "content": "To properly segment views of both static and moving objects, we assume a LiDAR at the commonly used frequency of \\(10\\mathrm{Hz}\\), where the scan overlap is enough to cluster together also moving objects, as depicted in Fig. 3. We show further examples in the supplementary material. With this procedure, our augmented pairs are views of the same object at different times from static and moving objects. By maximizing the similarity between the object views at different times, the network needs to learn a more general representation consistent across time and embed the object dynamics. In the supplementary material, we provide an experimental comparison between the pre-training using the temporal object views as augmented pairs and with augmented pairs generated only with data augmentation to validate the use of temporal views." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.236, + 0.832 + ], + "angle": 0, + "content": "3.2. Temporal batch" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "During pre-training, we need to sample a subset of \\( n \\) consecutive scans from the training sequence to be aggregated and extract the temporal object views as explained in Sec. 3.1. This sample will be used during pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "to learn a temporal consistent representation for the segmented objects. For a sequence of \\(N\\) point clouds, we divide it into batches \\(\\mathcal{B}\\) with intervals of \\(n\\) scans. During the pre-training forward pass, we sample two scans at different times, \\(t_1\\) and \\(t_2\\). Since the goal of this temporal sampling is to have different views of the same object, we enforce this by sampling \\(t_1\\) from the beginning and \\(t_2\\) from the end of this \\(n\\) scans interval such that \\(t_1 < \\frac{n}{3}\\) and \\(t_2 \\geq 2\\frac{n}{3}\\). In this case, the scans between \\(\\frac{n}{3} \\leq t < 2\\frac{n}{3}\\) would never be sampled. To also process those scans, the next batch starts at this unseen interval. Therefore, for a batch sample \\(\\mathcal{B}_b\\) the \\(n\\) scans from the sequence of \\(N\\) point clouds are \\(\\mathcal{B}_b = \\{\\mathcal{P}^k | b\\frac{n}{3} \\leq k < b\\frac{n}{3} + n\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.364 + ], + "angle": 0, + "content": "In Figure 4, we show how the batches would look like in an example with \\( n = 6 \\) scans interval. Even though we do not sample \\( t \\) from \\( \\frac{n}{3} \\leq t < 2\\frac{n}{3} \\), we still need to aggregate and cluster these scans to correctly segment dynamic objects and match their correspondences between the scans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.411 + ], + "angle": 0, + "content": "With this sampling scheme, we guarantee that the objects views will be distant in time and still all the data will be seen during pre-training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.679, + 0.438 + ], + "angle": 0, + "content": "3.3. Implicit clustering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.597 + ], + "angle": 0, + "content": "With our coarse object segments \\( S \\), we have the prior that points from the same segment should be semantically similar. Given that we rely on a set of object segments, one straightforward way to distinguish the learned embeddings is to cluster together point features from the same object. Recent methods propose an online clustering scheme to separate samples around a fixed number of prototype learnable cluster centers [10]. We aim at clustering point features from the same object close together but in a more straightforward way with an implicit clustering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Given the temporal sampled pair \\(\\mathcal{P}^{t_1}\\) and \\(\\mathcal{P}^{t_2}\\) from a batch sample \\(\\mathcal{B}_b\\), we compute point-wise features \\(\\mathcal{F}^{t_1}\\) and \\(\\mathcal{F}^{t_2}\\) with the backbone. As the target embedding, we list from \\(\\mathcal{F}^{t_2}\\) the set of \\(M\\) segments \\(\\mathcal{S}^{t_2}\\). We compute for each segment a mean representation from its point-wise features, and project this embedding with a self-attention Transformer encoder as a projection head to get the \\(M\\) target mean feature vectors \\(\\overline{\\boldsymbol{s}}^{t_2} \\in \\mathbb{R}^{M \\times D}\\) where \\(D\\) corresponds to the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The segments \\( S^{t_1} \\) are listed from the point-wise features \\( \\mathcal{F}^{t_1} \\). For \\( S^{t_1} \\), we do not compute mean embeddings but keep the features at point level, using the Transformer encoder to compute point-wise intra-class correspondences. To deal with the attention mechanism large memory requirement, we set a maximum number of \\( P \\) points per segment, which are randomly sampled. After sampling \\( P \\) points for each segment, we input the segments \\( S^{t_1} \\) pointwise features to the Transformer projection head, followed by another self-attention Transformer encoder as a predictor. We then get for each segment the \\( P \\) point-wise feature" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5220" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.45, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Figure 5. Diagram of the implicit clustering scheme. We pool the segments \\( S^{t_1} \\) and \\( S^{t_2} \\) from the backbone features. For \\( t_2 \\), we compute a mean representation for each segment and project it with the Transformer projection head. For \\( t_1 \\), we project the pointwise features and then use the Transformer predictor to predict the corresponding mean representation from \\( t_2 \\), clustering the point-wise features close to its mean target representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.363, + 0.245, + 0.379 + ], + "angle": 0, + "content": "vectors \\(s^{t_1} \\in \\mathbb{R}^{M \\times P \\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.508 + ], + "angle": 0, + "content": "With the segment target mean representations \\(\\overline{s}^{t_2}\\) and the predicted point-wise feature vectors \\(s^{t_1}\\) for each segment, we compute the loss to minimize the differences between the point features and the corresponding segment mean representation. For each point \\(p\\) from a segment \\(m\\), we compute the temperature-scaled cosine similarity \\(\\delta_{m,p,k}^{t_1\\to t_2}\\) between the normalized point-wise embedding \\(s_{m,p}^{t_1}\\in \\mathbb{R}^D\\) and the mean representation \\(\\overline{s}_k^{t_2}\\in \\mathbb{R}^D\\) from a segment \\(k\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.518, + 0.47, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\delta_ {m, p, k} ^ {t _ {1} \\rightarrow t _ {2}} = \\frac {\\left(\\boldsymbol {s} _ {m , p} ^ {t _ {1}}\\right) ^ {\\top} \\bar {\\boldsymbol {s}} _ {k} ^ {t _ {2}}}{\\tau}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.559, + 0.47, + 0.605 + ], + "angle": 0, + "content": "Next, we use the cross-entropy loss to maximize the similarity between each point \\( p \\) from a segment \\( m \\) and the corresponding target segment mean representation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.617, + 0.469, + 0.66 + ], + "angle": 0, + "content": "\\[\nl _ {m} ^ {t _ {1} \\rightarrow t _ {2}} = - \\sum_ {p = 1} ^ {P} \\log \\left(\\frac {\\exp \\left(\\delta_ {m , p , m} ^ {t _ {1} \\rightarrow t _ {2}}\\right)}{\\sum_ {k} ^ {M} \\exp \\left(\\delta_ {m , p , k} ^ {t _ {1} \\rightarrow t _ {2}}\\right)}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.667, + 0.47, + 0.714 + ], + "angle": 0, + "content": "We compute this loss for all the \\(M\\) segments and sum it to get the loss for predicting segments from \\(t_2\\) with the points of the segments from \\(t_1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.725, + 0.469, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {t _ {1} \\rightarrow t _ {2}} = \\sum_ {m = 1} ^ {M} l _ {m} ^ {t _ {1} \\rightarrow t _ {2}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With this formulation, our task is to predict for each point in the segment at time \\( t_1 \\) the corresponding segment at \\( t_2 \\). Since the target for all the \\( P \\) point embeddings \\( s_{m,p}^{t_1} \\) from a segment \\( m \\) is the same mean segment representation \\( \\overline{s}_m^{t_2} \\), the loss will push all the points from a segment to converge to a mean representation while separating from other segments, implicitly clustering together points from the same object as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "Lastly, we repeat the forward pass swapping \\( t_1 \\) and \\( t_2 \\) to have a symmetric representation, learning both the correspondences from \\( t_1 \\rightarrow t_2 \\) and from \\( t_2 \\rightarrow t_1 \\) leading to:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.151, + 0.892, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T A R L} = \\mathcal {L} ^ {t _ {1} \\rightarrow t _ {2}} + \\mathcal {L} ^ {t _ {2} \\rightarrow t _ {1}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.173, + 0.892, + 0.22 + ], + "angle": 0, + "content": "In the next section we give an intuition behind our choice of projection head, and how it can help the model to learn a more fine-grained representation during pre-training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.227, + 0.719, + 0.243 + ], + "angle": 0, + "content": "3.4. Transformer projection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Self-supervised representation learning methods [13, 14, 27] typically add a non-linear projection head to the backbone to project the embeddings to the target feature space. The idea is that the backbone should learn general features, and this projection head will overfit to the pre-training pretext task and later be discarded during fine-tuning. Our projection scheme follows more recent methods [15, 25], which shows that an asymmetric network can improve the learned representation. Besides, with the prior that points from one segment should be similar, we want to focus on the pointwise relationship. Therefore, we replace the non-linear projection head with a self-attention Transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.627 + ], + "angle": 0, + "content": "In our siamese network scheme, the online updated network computes point-wise features. In this case, the queries, keys, and values for our Transformer projector will be point-wise features. We aim to learn intra-class features with the attention mechanism and identify correspondences between the points from one segment. In contrast, the momentum updated network computes a mean feature vector for each segment. In this case, the attention mechanism should look for similarities and differences within the different objects segments in the scene. By using the self-attention over points and segments, we aim at learning at the same time point-wise correspondences while also learning the differences between the different segments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.719 + ], + "angle": 0, + "content": "With the Transformer projector, the pre-training can lead the learned features towards a more fine-grained representation, where intra-class point-wise features are matched with the individual segment features. The supplementary material provides ablations comparing the Transformer with the commonly used non-linear projection head." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.704, + 0.743 + ], + "angle": 0, + "content": "3.5. Pre-training overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Secs. 3.1 to 3.4 explain the individual steps of our approach. This section summarizes the entire pre-training pipeline, putting together the modules explained in the previous subsections." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our pipeline shown in Fig. 2 fetches a batch sample \\(\\mathcal{B}_b\\) of \\(n\\) sequential scans and applies a transformation to the scans \\(\\mathcal{P}^t\\) with the corresponding poses to have all points in a common coordinate frame. We use an unsupervised ground segmentation approach to remove the ground \\(\\mathcal{G}\\). Next, we cluster the remaining points \\(\\hat{\\mathcal{P}}\\), defining coarse" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5221" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.196 + ], + "angle": 0, + "content": "
MethodScribbles0.1%1%10%50%100%
Scratch54.9629.3542.7753.9658.2759.03
PointContrast [67]54.5232.6344.6258.6859.9861.45
DepthContrast [75]55.9031.6648.0557.1160.9961.14
SegContrast [50]56.7032.7544.8356.3160.4561.02
TARL (Ours)57.2538.5951.4260.3461.4261.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.201, + 0.47, + 0.245 + ], + "angle": 0, + "content": "Table 1. Results (mIoU) of the models pre-trained on SemanticKITTI fine-tuned to semantic segmentation on SemanticKITTI with different percentage of labels and scribbles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.473, + 0.499 + ], + "angle": 0, + "content": "segments \\(S\\) of the objects in the scene. We sample two random scans at different times \\(\\mathcal{P}^{t_1}\\) and \\(\\mathcal{P}^{t_2}\\) from \\(\\mathcal{B}_b\\), in which we apply random augmentations \\(\\mathcal{T}^{t_1}\\) and \\(\\mathcal{T}^{t_2}\\). We compute point-wise features \\(\\mathcal{F}^{t_1}\\) and \\(\\mathcal{F}^{t_2}\\) from the sampled point clouds with the backbone, and list the segments \\(S^{t_1}\\) and \\(S^{t_2}\\) with its corresponding point-wise features. We calculate a mean embedding \\(\\overline{\\mathcal{S}}^{t_2}\\) for each segment \\(m\\) in \\(\\mathcal{P}^{t_2}\\) and compute the target representation \\(\\overline{s}^{t_2}\\) for each segment with a Transformer projection head. For the segments \\(S^{t_1}\\), we project the point-wise features with another Transformer projector followed by a Transformer encoder as a predictor to get the point-wise features \\(s^{t_1}\\). Lastly, we compute the loss to predict the corresponding segment representation at \\(t_2\\) for each point in \\(t_1\\). We repeat the process changing \\(t_1\\) and \\(t_2\\) to symmetrically match points from scan \\(t_2\\) to the corresponding segment at \\(t_1\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.21, + 0.53 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Datasets. We use the SemanticKITTI dataset [4,22] for our pre-training, an autonomous driving LiDAR data benchmark with point-wise annotations. During pre-training, we use only the raw point clouds from the training sequences and the given poses to extract the temporal objects views. For fine-tuning, we use SemanticKITTI with the full scans annotations and the scribbles annotations [51]. Besides SemanticKITTI, we also use nuScenes [8,64] for fine-tuning, to evaluate our approach in different autonomous driving datasets collected with different LiDARs and sensor setups, reporting the results on the validation sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Model architecture. For our evaluation, we use a MinkUNet [20] as the backbone, which voxelizes the point clouds and uses sparse convolutions to extract features from the point cloud. The backbone receives the point cloud coordinates and intensity as input, and the output dimension feature has dimension \\( D = 96 \\). We use a multi-head self-attention encoder for the projection head with one layer and eight attention heads. For the predictor, we use the same multi-head self-attention encoder architecture. Further details on the backbone used in our experiments are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Pre-training. For pre-training, we use the AdamW optimizer [43] with a learning rate of \\(2 \\cdot 10^{-4}\\) and decay of \\(10^{-4}\\)," + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.089, + 0.848, + 0.215 + ], + "angle": 0, + "content": "
MethodMiniFull
Scratch26.9466.03
Supervised pre-training38.3967.35
PointContrast [67]31.9267.31
DepthContrast [75]27.8164.70
SegContrast [50]31.2767.70
TARL (Ours)39.3668.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.22, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 2. Results (mIoU) of the self-supervised and supervised pretrained models on SemanticKITTI fine-tuned to semantic segmentation on nuScenes mini and full training sets." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.276, + 0.848, + 0.383 + ], + "angle": 0, + "content": "
MethodKITTInuScenes
Frozen backbone4.105.62
PointContrast [67]25.5319.70
DepthContrast [75]13.879.65
SegContrast [50]27.4223.75
TARL (Ours)35.9026.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.388, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Table 3. Linear evaluation results (mIoU) on SemanticKITTI and nuScenes datasets for semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.643 + ], + "angle": 0, + "content": "training for 200 epochs with batch size 8 using a single NVIDIA RTX A6000. We use \\( n = 12 \\) for the scans to be aggregated. We provide experiments with different numbers of scans in the supplementary material to support this choice. The voxel resolution is set to \\( 0.05m \\) for the input point clouds, from which we sample a maximum of 40,000 points per point cloud. During the segment pooling, we limit it to a total of \\( M = 50 \\) segments with a maximum of \\( P = 300 \\) points per segment to avoid memory overflow. We use \\( \\tau = 0.1 \\) to compute \\( \\delta \\) in Eq. (1) and momentum of 0.999 to update the momentum network according to the online network weights. For the baselines, we use their official repositories with their default parameters, also training them for 200 epochs and sampling 40,000 points per scan." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Fine-tuning. To evaluate our method and compare it with the baselines, we fine-tune the models to three downstream tasks: semantic segmentation, panoptic segmentation, and object detection. For semantic segmentation, we use the pipeline provided in SegContrast [50] with the AdamW optimizer [43] and a learning rate of 0.001. For panoptic segmentation, we use the baseline evaluated by Hong et al. [29], where a 3D backbone network is used together with a semantic and an instance heads, followed by a clustering post-processing to identify the instances. We use the pre-trained MinkUNet model as the 3D backbone. We use a learning rate of 0.2 with batch size 8, training for 50 epochs. For object detection, we use the same toolbox and hyperparameters used in DepthContrast [75] with the PartA2 detector [57]. For all the tasks, we evaluate the method with the same model trained from scratch (without pre-training) and with the model after pre-training with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5222" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.805, + 0.213 + ], + "angle": 0, + "content": "
0.1%1%10%50%100%
MethodPQIoUPQIoUPQIoUPQIoUPQIoU
Scratch4.7611.1322.7230.8447.2053.5355.3261.9455.4059.75
PointContrast [67]5.8611.5127.3732.4947.5754.6354.2159.4855.8561.49
DepthContrast [75]7.6513.5627.3132.3046.8551.2754.5559.6056.1560.81
SegContrast [50]7.5814.4626.1432.8547.0253.4755.3860.0456.7361.96
TARL (Ours)10.2617.0129.2434.7151.2757.5956.1062.3656.5762.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.22, + 0.893, + 0.248 + ], + "angle": 0, + "content": "Table 4. Results (PQ and IoU) when fine-tuning the pre-trained models to panoptic segmentation with different percentage of labels on SemanticKITTI." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.258, + 0.449, + 0.403 + ], + "angle": 0, + "content": "
MiniFull
PQIoUPQIoU
Scratch23.7823.9652.9858.17
Supervised pre-training24.7723.6053.1958.05
PointContrast [67]26.5825.4651.0656.39
DepthContrast [75]28.6627.3051.5157.06
SegContrast [50]28.8426.7952.3157.24
TARL (Ours)32.2230.7353.2659.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.452 + ], + "angle": 0, + "content": "Table 5. Results (PQ and IoU) of the self-supervised and supervised pre-trained models on SemanticKITTI fine-tuned to panoramic segmentation on nuScenes mini and full training sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.505 + ], + "angle": 0, + "content": "the different self-supervised methods. The supplementary material provides detailed information for each downstream task training procedure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.292, + 0.531 + ], + "angle": 0, + "content": "4.1. Semantic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.719 + ], + "angle": 0, + "content": "In this evaluation, we want to measure the information about semantics learned by the network during pretraining. We fine-tune it to semantic segmentation on SemanticKITTI and nuScenes after pre-training on SemanticKITTI. For SemanticKITTI, we use the percentage scans subsets used in SegContrast [50], with \\(0.1\\%\\), \\(1\\%\\), \\(10\\%\\), \\(50\\%\\), and \\(100\\%\\) of the labeled scans. We also evaluate on the scribbles annotations [51], which labels just a subset of points for each scan in the SemanticKITTI training sequences. We report the mean intersection-over-union (mIoU) of the models on the validation set, while fine-tuned on the aforementioned subset of labels from the training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Tab. 1 shows the results from our approach compared to the baselines. As can be seen, overall the pre-training methods boost the performance compared to the network without pre-training on both scribbles annotations and the subsets of labeled scans. However, our approach surpasses previous state-of-the-art methods in all the different subsets, with a bigger gap when fewer labeled scans are used to fine-tune the model. We can also notice that, after pre-training with our method, only \\(10\\%\\) of labeled scans were necessary for the model to surpass the performance of the network trained from scratch with \\(100\\%\\) of the labeled scans. This result suggests that our method can effectively reduce the amount" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.295 + ], + "angle": 0, + "content": "of labeled data necessary for the semantic segmentation task by around ten times. In the supplementary material, we also provide the IoU for each individual semantic class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Generalization. In this experiment, we evaluate the generalization of the learned features. We use the network pretrained on SemanticKITTI and fine-tune it on nuScenes. We use nuScenes full training set and the mini training subset to fine-tune the network, and evaluate on the full validation set. In this evaluation, we also compare the results of the fully supervised semantic segmentation pre-training on SemanticKITTI fine-tuned on nuScenes. As shown in Tab. 2, even though all methods can improve the performance of the model trained on nuScenes, our approach can surpass previous self-supervised methods by around \\(8\\%\\) when training with fewer labels. Besides, compared with supervised pre-training of the network on SemanticKITTI, our approach achieves better performance on both full and mini training sets. These results suggest that our method can replace supervised pre-training for LiDAR data, since it achieved better performance when fine-tuning to a different dataset, collected with a different sensor setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Linear evaluation. In this experiment, the pre-trained backbone is frozen, training only a linear head on top of it to evaluate how descriptive the self-supervised learned representation is without fine-tuning it to the target task. We perform this evaluation on SemanticKITTI and nuScenes datasets. In Tab. 3, the randomly initialized network (frozen backbone) achieves low performance, suggesting that the features extracted by the backbone play the main role in achieving semantic segmentation. All the pre-training methods improve the performance compared to the network without pre-training. However, our approach has a clear performance gap compared to the baselines, especially when training to the same dataset used for pre-training. These results suggest that the representation learned by our method can embed more semantic information, even though not using labels, achieving higher performance than previous state-of-the-art methods on both datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.712, + 0.849 + ], + "angle": 0, + "content": "4.2. Panoptic segmentation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this experiment, we fine-tune the pre-trained models for panoptic segmentation to also evaluate the instance-level features of our learned representation. Same as for semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5223" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.817, + 0.215 + ], + "angle": 0, + "content": "
CarPedestrianCyclist
EasyModerateHardEasyModerateHardEasyModerateHard
Scratch91.7582.1380.1667.6563.1357.7391.7474.9470.34
PointContrast [67]91.5581.7079.8867.3362.3256.5592.3375.2070.66
DepthContrast [75]91.9882.0480.0369.1363.9757.7192.6873.6770.10
SegContrast [50]92.0882.1080.1868.7263.7757.3691.2674.7070.28
TARL (Ours)92.0982.1580.1068.5064.0058.7492.6175.4470.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.22, + 0.894, + 0.248 + ], + "angle": 0, + "content": "Table 6. Results (AP50) when fine-tuning the pre-trained models to object detection on KITTI full training set for car, pedestrian and cyclist classes on the easy, moderate and hard difficult levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.317 + ], + "angle": 0, + "content": "segmentation, we fine-tuned the model to SemanticKITTI using the same percentage subsets and nuScenes using the full and the mini training sets, reporting both the mIoU and the panoptic quality (PQ) on the validation sets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Tab. 4 shows that our method is consistently better than previous self-supervised pre-training approaches. As the amount of training samples increases, the differences diminish. However, our method has a clear performance gap compared to the baselines when trained with fewer labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.471, + 0.575 + ], + "angle": 0, + "content": "Generalization. To evaluate the generalization of our learned features on panoptic segmentation, we use the network pre-trained on SemantiKITTI to fine-tune it on nuScenes full and mini training sets, evaluating it on the full validation set. Tab. 5 shows that the representation learned by our method was more suited to be transferred to a different dataset. As in semantic segmentation, our method achieved better performance than previous self-supervised approaches and supervised pre-training on SemanticKITTI. These results agree with the semantic segmentation evaluation, suggesting that our method is better suited for transfer learning than the supervised pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.668 + ], + "angle": 0, + "content": "With these experiments, we can validate that our learned representations can extract not only semantic knowledge from the data but also instance-level information, surpassing previous methods in IoU and PQ metrics. In the supplementary material, we provide further results, reporting also the segmentation and recognition quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.24, + 0.696 + ], + "angle": 0, + "content": "4.3. Object detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.825 + ], + "angle": 0, + "content": "As a third downstream task, we compare the methods fine-tuned for object detection to evaluate how general is the representation learned by our method compared with previous state-of-the-art methods. In this evaluation, we fine-tune the model with the whole training set of the KITTI dataset [22] and report the average precision \\((\\mathrm{AP}_{50})\\) with 40 recall positions for car, pedestrian, and cyclist classes on the three difficulty levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In Tab. 6, we can compare the baselines performance with our approach. For the car class, the pre-training methods achieve a marginal improvement compared to the network trained from scratch. For pedestrian and cyclist, the improvements brought by the pre-training are more substan" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.346 + ], + "angle": 0, + "content": "tial. On those classes, our method achieves the best performance on moderate and hard difficulties, showing that our approach can also boost the performance when fine-tuning the model on object detection. In our supplementary material, we provide further experiments with the different training subsets used by Zhang et al. [75]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.452 + ], + "angle": 0, + "content": "Together with the evaluations of the other downstream tasks, these results suggest that our method can learn a general representation. Our method achieved better performance than previous approaches on all three downstream tasks, showing that our approach learns a robust representation, not biased to a specific task but showing generalizable representations suitable for different tasks and datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.619, + 0.477 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.803 + ], + "angle": 0, + "content": "In this paper, we propose a self-supervised representation learning method for 3D LiDAR data evaluated in the context of autonomous driving. We exploit the vehicle motion to extract different views of objects across time to learn a temporally-consistent representation. We use a Transformer encoder as a projection head and implicitly clustered points from the same object together while discriminating between different objects. Our experiments show that our method achieves better performance than previous state-of-the-art approaches on different downstream tasks. Besides, our approach could reduce the amount of necessary labeled data to only \\(10\\%\\) when fine-tuning for semantic segmentation on SemanticKITTI. These results show that our proposed approach could learn a general LiDAR point cloud representation, embedding semantic information by matching objects representation viewed at different times. In addition, our method achieved better performance than supervised pre-training when transferring the learned representation to a dataset collected with different LiDAR sensors, suggesting that our approach could replace supervised pretraining in the LiDAR data domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Limitations. Despite encouraging results, there is space for further improvement. For the unsupervised segment extraction, the clustering parameters have to be tuned depending on the LiDAR used to collect the data. Also, we rely on high-frequency data to properly segment dynamic objects. Otherwise, temporal segments could be wrongly matched." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5224" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proc. of the Int. Conf. on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.282 + ], + "angle": 0, + "content": "[2] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.472, + 0.352 + ], + "angle": 0, + "content": "[3] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyril Stachniss, Jens Behley, and Laura Leal-Taixe. 4D Panoptic LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.353, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.423, + 0.472, + 0.465 + ], + "angle": 0, + "content": "[5] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS), 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[6] Benedikt Mersch, Xieyuanli Chen, Ignacio Vizzo, Lucas Nunes, Jens Behley, and Cyril Stachniss. Receding Moving Object Segmentation in 3D LiDAR Data Using Sparse 4D Convolutions. IEEE Robotics and Automation Letters (RA-L), 7(3):7503-7510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.472, + 0.591 + ], + "angle": 0, + "content": "[7] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised Domain Adaptation for LiDAR Panoptic Segmentation. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.472, + 0.675 + ], + "angle": 0, + "content": "[8] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[9] Ricardo J. G. B. Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Advances in Knowledge Discovery and Data Mining, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "PTTR: Relational 3D Point Cloud Object Tracking With Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In Proc. of the Int. Conf. on Machine Learning (ICML), 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint, arxiv:2003.04297, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[15] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jirgen Gall, Jens Behley, and Cyril Stachniss. Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data. IEEE Robotics and Automation Letters (RA-L), 6:6529-6536, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[17] Xieyuanli Chen, Andres Milioto, Emanuele Palazzolo, Philippe Giguere, Jens Behley, and Cyril Stachniss. SuMa++: Efficient LiDAR-based Semantic SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[19] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PatchFormer: An Efficient Point Transformer With Patch Attention. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[21] Fayao Liu, Guosheng Lin, and Chuan-Sheng Foo. Point discriminative learning for unsupervised representation learning on 3d point clouds. arXiv preprint, arXiv:2108.02104, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[24] Benjamin Graham and Laurens van der Maaten. Submanifold Sparse Convolutional Networks. arXiv preprint, arxiv:1706.01307, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[25] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5225" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[26] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J. Kusner. Unsupervised Point Cloud Pre-Training via Occlusion Completion. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[28] Hiroharu Kato and Tatsuya Harada. Image Reconstruction from Bag-of-Visual-Words. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[29] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. LiDAR-Based Panoptic Segmentation via Dynamic Shifting Network. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[30] Hongxu Yin, Arash Vahdat, Jose Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-ViT: Adaptive Tokens for Efficient Vision Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[31] Ji Hou, Benjamin Graham, Matthias Niessner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding With Contrastive Scene Contexts. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[32] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP - Simple, Accurate, and Robust Registration If Done the Right Way. arXiv preprint arXiv:2209.15397, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[33] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[34] Ishan Misra, Rohit Girdharm, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[35] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. ST3D: Self-Training for Unsupervised Domain Adaptation on 3D Object Detection. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[36] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposalcontrast: Unsupervised pre-training for lidar-based 3d object detection. Proc. of the Europ. Conf. on Computer Vision (ECCV),2022.1,2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Kai Fischer, Martin Simon, Florian Oelsner, Stefan Milz, Horst-Michael Gross, and Patrick Maeder. StickyPillars: Robust and Efficient Feature Matching on Point Clouds Using Graph Neural Networks. In Proc. of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[38] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. PointPillars: Fast Encoders for Object Detection From Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[39] Ferdinand Langer, Andres Milioto, Alexandre Haag, Jens Behley, and Cyril Stachniss. Domain Transfer for Semantic Segmentation of LiDAR Data using Deep Neural Networks. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[40] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph Attention Convolution for Point Cloud Semantic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[41] Hyungtae Lim, Sungwon Hwang, and Hyun Myung. ERASOR: Egocentric Ratio of Pseudo Occupancy-Based Dynamic Object Removal for Static 3D Point Cloud Map Building. IEEE Robotics and Automation Letters (RA-L), 6(2):2272-2279, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[42] Ling Zhang and Zhigang Zhu. Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks. In Proc. of the Intl. Conf. on 3D Vision (3DV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[44] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Contrastive Instance Association for 4D Panoptic Segmentation using Sequences of 3D LiDAR Scans. IEEE Robotics and Automation Letters (RA-L), 7(2):1550-1557, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[45] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. LiDAR Panoptic Segmentation for Autonomous Driving. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[46] Andres Milioto, Leonard Mandtler, and Cyril Stachniss. Fast Instance and Semantic Segmentation Exploiting Local Connectivity, Metric Learning, and One-Shot Detection for Robotics. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[47] Andres Milioto and Cyril Stachniss. Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics using CNNs. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[48] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-BERT: Vision-Language Pre-Training on Fashion Domain. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5226" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[50] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. SegContrast: 3D Point Cloud Feature Representation Learning through Self-supervised Segment Discrimination. IEEE Robotics and Automation Letters (RA-L), 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[51] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[52] Pedro O. Pinheiro, Amjad Almahairi, Ryan Y. Benmalek, Florian Golemo, and Aaron Courville. Unsupervised learning of dense visual representations. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), pages 4489-4500, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[53] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP Real-Time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[54] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[55] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[56] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Proc. of the Europ. Conf. on Computer Vision (ECCV), pages 626-642, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[57] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network. IEEE Trans. on Pattern Analalysis and Machine Intelligence (TPAMI), 43:2647-2664, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[58] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointR-CNN: 3D Object Proposal Generation and Detection From Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[59] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-Level Representation Learning for Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[60] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[61] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPCov: Flexible and Deformable Convolution for Point Clouds. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[62] Weijing Shi and Ragunathan (Raj) Rajkumar. Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[63] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring Cross-Image Pixel Contrast for Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[64] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters (RA-L), 7(2):3795-3802, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[65] Xiangyun Zhao, Raviteja Vemulapalli, Philip Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive Learning for Label Efficient Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[67] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[68] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[69] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense Contrastive Learning for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[70] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[71] Li Yi, Boqing Gong, and Thomas Funkhouser. Complete & Label: A Domain Adaptation Approach to Semantic Segmentation of LiDAR Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[72] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. MPViT: Multi-Path Vision Transformer for Dense Prediction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[73] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust Instance Segmentation Through Reasoning About Multi-Object Occlusion. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[74] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. Proc. of the Europ. Conf. on Computer Vision (ECCV), 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[75] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-Supervised Pretraining of 3D Features on any" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5227" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Point-Cloud. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[76] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In Proc. of the Int. Conf. on Machine Learning (ICML), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[77] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[78] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring Set Similarity for Dense Self-Supervised Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[79] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-Free LiDAR Point Cloud Panoptic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[80] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5228" + } + ] +] \ No newline at end of file diff --git a/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_origin.pdf b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e7b7fa90338b60bd3b6c5ec049847ede55fc0205 --- /dev/null +++ b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/4a98930e-4a72-4866-a1b4-e6692bb0ab17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb26d60d5186804c64562ca115cb246d31fe47ccfb4a6d8b01d7a956e2856271 +size 2953845 diff --git a/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/full.md b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6954ee7a44efd002478cdb85330e5ba6657c0f7a --- /dev/null +++ b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/full.md @@ -0,0 +1,322 @@ +# Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving + +Lucas Nunes* + +Xieyuanli Chen* + +Louis Wiesmann* + +Jens Behley* + +Rodrigo Marcuzzi* + +Cyrill Stachniss\*,\* + +*University of Bonn, *University of Oxford, ‡Lamarr Institute + +{firstname.lastname}@igg.uni-bonn.de + +# Abstract + +Semantic perception is a core building block in autonomous driving, since it provides information about the drivable space and location of other traffic participants. For learning-based perception, often a large amount of diverse training data is necessary to achieve high performance. Data labeling is usually a bottleneck for developing such methods, especially for dense prediction tasks, e.g., semantic segmentation or panoptic segmentation. For 3D LiDAR data, the annotation process demands even more effort than for images. Especially in autonomous driving, point clouds are sparse, and objects appearance depends on its distance from the sensor, making it harder to acquire large amounts of labeled training data. This paper aims at taking an alternative path proposing a self-supervised representation learning method for 3D LiDAR data. Our approach exploits the vehicle motion to match objects across time viewed in different scans. We then train a model to maximize the point-wise feature similarities from points of the associated object in different scans, which enables to learn a consistent representation across time. The experimental results show that our approach performs better than previous state-of-the-art self-supervised representation learning methods when fine-tuning to different downstream tasks. We furthermore show that with only $10\%$ of labeled data, a network pre-trained with our approach can achieve better performance than the same network trained from scratch with all labels for semantic segmentation on SemanticKITTI. + +# 1. Introduction + +Semantic perception is essential for safe interaction between autonomous vehicles and their surrounding. For learning-based perception, a massive amount of training data is usually required for training high-performance models. However, the data annotation is the bottleneck of collecting such large training sets, especially for dense prediction tasks, such as semantic segmentation [60, 80], instance + +![](images/a475f1413dde086649b200520421f2454fdc05c1f21b7c8d2a8d298cbddb7d0f.jpg) +Figure 1. Our pre-training (TARL) can reduce the amount of necessary labels during fine-tuning on SemanticKITTI [4, 22]. Our method requires only $10\%$ of labels to surpass the network trained from scratch with the full dataset for semantic segmentation. For panoptic segmentation, our method requires only half of the labels. + +![](images/0f093a366a4e33cb73fc9af0e7482fcf8d2953e25f8684f8e04d54414f1b860f.jpg) + +segmentation [46, 73], and panoptic segmentation [45, 79], which require fine-grained labels. In the context of autonomous driving, recent approaches exploit 3D LiDAR data [20, 54, 55, 61], where the data annotation process is more complex than on 2D RGB images due to the sparsity of the point cloud and object appearance varying with its distance to the sensor. + +Recent self-supervised representation learning methods [10, 11, 13-15, 25, 27, 76] tackle the lack of annotated data with a pre-training step requiring no labels. Those methods use data augmentation to generate different views from one data sample and train the network to learn an embedding space able to have similar representations for the generated augmented views. Other approaches [52, 63, 65, 69, 78] propose optimizing the pixel embedding space to learn a dense representation suited to be fine-tuned to more fine-grained tasks. For 3D point cloud data, recent approaches [21, 42, 49, 56] focus on synthetic point clouds of single objects to learn a robust representation for object classification. Other approaches [31, 36, 50, 67, 74, 75] target real-world data representation, such as LiDAR or RGB-D data, but fewer target autonomous driving scenarios. + +In this paper, we propose a novel temporal association representation learning (TARL) method for 3D LiDAR data designed for autonomous driving data. We exploit the vehicle motion to extract different views of the same objects across time. Then, we compute point-wise features from the objects in the point cloud and use a Transformer encoder [2] as a projection head to learn a temporal association from the object representation, embedding the objects dynamics. We conduct extensive experiments to compare our approach with state-of-the-art methods and show that our approach surpasses previous self-supervised pretraining methods [50, 67, 75] when fine-tuning to different downstream tasks and datasets [4, 8, 22, 64]. In summary, our key contributions are: + +- We propose a novel self-supervised pre-training for 3D LiDAR data able to learn a robust and temporally-consistent representation by clustering together points from the same object viewed at different points in time. +- We achieve better performance than previous state-of-the-art methods on different downstream tasks, i.e., semantic segmentation, panoptic segmentation, and object detection. +- With our pre-training, we require only $10\%$ of labels to surpass the network trained from scratch for semantic segmentation using the whole training set on SemanticKITTI (see Fig. 1). +- Our self-supervised pre-training produces representations more suited for transfer learning than supervised pre-training, achieving better performance when finetuning to a different dataset. + +# 2. Related Work + +Self-supervised representation learning aims at initializing weights of a deep neural network for a downstream task by learning suitable representations without labeled data. Early approaches [18, 26, 28, 33, 48, 70] use so-called pretext tasks to learn useful representations, such as solving jigsaw puzzles [28, 33, 48] or reconstructing masked data [18, 26, 70]. More recently, contrastive learning [13, 14, 27] has drawn significant attention in the vision and robotics research community. Those contrastive methods employ random data augmentation over one image sample to generate different views of it. Then, the network is trained to learn an embedding space able to discriminate between the augmented samples from the same image and the augmented views from other images. Recent approaches exploit this discriminative pretext task and propose different ways to learn a self-supervised data representation, e.g., via non-contrastive methods [15, 25], redundancy reduction [76], or online clustering [10]. Other works [52, 63, 65, 69, 78] approach this in a pixel-wise man + +ner, learning dense representations applied to more fine-grained downstream tasks. + +Transformer encoders [2] were initially proposed in the natural language processing community achieving state-of-the-art performance in the field. Such architectures were then quickly adopted by the computer vision community [1, 30, 66, 72, 77]. With the attention mechanism, the Transformer can learn the global context from the image instead of just local neighboring context as in convolutional models [1]. With the growing interest in Transformers, some self-supervised representation learning methods were applied to Transformers [11, 59, 68] to evaluate the capacity of such models to learn meaningful representations. + +LiDAR point cloud perception brings further challenges compared to the image domain. Previous works [46, 47] deal with 3D scans by projecting the data to a 2D range image and then processing it with 2D convolutional methods. More recently, different convolution architectures were proposed to deal with 3D data, such as point kernels [54, 55, 61], sparse convolutions [20, 23, 24], or graph representations [37, 40, 62]. With those different architectures, it becomes possible to train a model to learn from 3D data and deal with different tasks in this domain, such as semantic segmentation [60,80], panoptic segmentation [3, 29, 44], object detection [38, 57, 58], or moving object segmentation [6, 16]. With increasing interest in Transformers, some works [12, 19, 34, 70] proposed ways of dealing with point cloud data in such architectures. In this domain, a key challenge lies in the number of points collected by LiDAR sensors together with the high memory requirements from the Transformer attention mechanism. + +3D self-supervised representation learning aims at learning robust representations from point cloud data. Most of those methods focus on point clouds of single objects [21, 42,49,56]. Other methods [31,36,50,67,74,75] target dense point-wise representations for real-world point cloud scans. Xie et al. [67] propose learning a representation via point-to-point contrastive learning through matching corresponding points in two augmented point clouds. Similarly, Hou et al. [31] learn both point-wise correspondence and the region where the point is located in the point cloud. Zhang et al. [75] propose global representation learning through scan-to-scan discrimination. Chen et al. [74] synthetically add dynamic objects in the scene to embed temporal and dynamic understanding in the learned representation. Other methods propose discriminating regions of interest to extract more semantic information. Nunes et al. [50] extract coarse segments from point clouds and use contrastive loss to distinguish between objects segments. Similarly, Yin et al. [36] extract proposal regions to be discriminated, biasing the pre-training to object detection. + +Despite the performance improvement brought by those 3D self-supervised representation learning methods, there is + +![](images/03c5c449c0b5312b5ae1b6082ef0e68336b71dac4d67ed8df76dcb0dbfb99322.jpg) +Figure 2. We aggregate $n$ point clouds and extract segments $S$ from it by removing the ground and clustering the remaining points. Then we sample two point clouds $\mathcal{P}^{t_1}$ and $\mathcal{P}^{t_2}$ , and apply random augmentations $\mathcal{T}^{t_1}$ and $\mathcal{T}^{t_2}$ to them. Next, we compute point-wise features for each scan and list its segments $S^{t_1}$ and $S^{t_2}$ with their corresponding point features. For $S^{t_2}$ we compute the mean feature vector for each segment and compute the segments target embeddings $\overline{s}^{t_2}$ with a projection head. For $S^{t_1}$ we use a point-wise projector followed by a predictor to identify for each segment point features $s_{m,p}^{t_1}$ the corresponding target embedding $\overline{s}_m^{t_2}$ by minimizing the loss $\mathcal{L}^{t_1\to t_2}$ . + +still a performance gap compared to the image domain due to the lack of data augmentations for point clouds. Similar to previous works [36,50], our approach focuses on learning representations over regions of interest. However, unlike previous methods, we extract natural augmented views of the objects in the scans collected at different points in time while the vehicle drives in the environment. Then, the network is trained to learn a common representation for those temporal views of one object. This enables our method to learn a more robust semantic representation by learning temporal consistent features and real-world dynamics. + +# 3. Method + +In this paper, we propose a new self-supervised representation learning method for LiDAR data obtained in the context of autonomous driving. Our approach requires only unlabeled point clouds and the corresponding scan pose. Pose information is usually readily available by means of GPS/IMU, odometry approaches [32, 53], or SLAM systems [5,17], thus, this is not a limitation in practice. We use a siamese network scheme with an online updated network and a momentum updated network (see Fig. 2). The whole pipeline consists of extracting objects as coarse segments viewed at different times over an interval of scans. Then, point-wise features are computed with the backbone and a Transformer encoder as a projector. Finally, we perform an implicit clustering to put points from different views of the same object together. In the following subsections, we explain the individual steps of our method thoroughly. + +# 3.1. Temporal objects views + +Instead of only using data augmentation to generate artificial views of one object, we exploit the vehicle motion to extract real-world object segments viewed from different perspectives. Given the vehicle motion and the properties of LiDAR sensors, one object can have different ap + +![](images/9d6ea86d1691e11ba8e8aea2a9dde9b8082115d8af87298f2cf3c4bf4d41e3f2.jpg) +Figure 3. We aggregate $n$ point clouds from different times and extract coarse segments from the aggregated point cloud in an unsupervised manner. We keep the point indices mapping to extract views at different times from one object as augmented pairs. + +pearances depending on its position relative to the sensor. Such changing appearance is usually a problem when fine-tuning the model to data collected with different sensors [7, 35, 39, 71]. However, we exploit these properties in our favor to extract natural augmented views of one object. Previous works [31, 36, 50, 67, 74, 75] relied only on data augmentation to generate pairs of one scan, such as rotation, translation and scaling. Instead, we extract object views collected from different perspectives by the LiDAR sensor while the vehicle navigates through the environment. We then apply augmentations over the associated scans captured at a different time and train the network to learn a common representation for the object views. + +We define a point cloud at time $t$ as $\mathcal{P}^t = \{\pmb {p}_1^t,\dots ,\pmb {p}_R^t\}$ as a set of 3D points $\pmb {p}_r\in \mathbb{R}^3$ . To extract coarse object segments in an individual scan, we first separate $\mathcal{P}^t$ into ground $\mathcal{G}^t$ and non-ground points $\hat{\mathcal{P}}^t$ in an unsupervised manner using the method proposed by Lim et al. [41], where $\mathcal{P}^t = \hat{\mathcal{P}}^t\cup \mathcal{G}^t$ and $\hat{\mathcal{P}}^t\cap \mathcal{G}^t = \emptyset$ .Af + +![](images/a7dab86986ef2f1609875e06913fafd6596bd7a787b255eb8a5459a0bb303670.jpg) +Figure 4. Batch example, for an interval of $n = 6$ point clouds $\mathcal{P}$ . We divide a sequence of $N$ scans into batches $\mathcal{B}_b$ where we sample $t_1$ from the beginning and $t_2$ from the end of the interval. Then, the next batch starts in the middle of the previous interval. + +terwards, we cluster the non-ground points $\hat{\mathcal{P}}^t$ into $M$ segments $\mathcal{S}^t = \{S_1^t,\dots,S_M^t\}$ using HDBSCAN [9], where $\hat{\mathcal{P}}^t = \bigcup_{m = 1}^{M}\mathcal{S}_m^t$ + +To map different segment views at different times, we define an interval of $n$ scans from which the views of the object will be extracted. Those scans are transformed to a common global coordinate frame to be then aggregated. The global scan poses can be easily acquired with GPS/IMU, SLAM systems [5, 17], or Li-DAR odometry [32, 53]. The aggregated point cloud $\mathcal{P}$ is given by $\mathcal{P} = \{\mathcal{P}^{t + 1},\mathcal{P}^{t + 2},\dots,\mathcal{P}^{t + n}\}$ . Similarly, we can aggregate the individual ground segmentation labels $\mathcal{G} = \{\mathcal{G}^{t + 1},\mathcal{G}^{t + 2},\dots,\mathcal{G}^{t + n}\}$ and get the aggregated non-ground points $\hat{\mathcal{P}} = \{\hat{\mathcal{P}}^{t + 1},\hat{\mathcal{P}}^{t + 2},\dots,\hat{\mathcal{P}}^{t + n}\}$ . As in the individual scan case, we can cluster $\hat{\mathcal{P}}$ to get the $M$ segments $S = \{S_{1},\dots,S_{M}\}$ . By keeping the point index mapping from the aggregated point cloud to the individual $n$ scans, we can identify the $n$ segments of the same object viewed at different times as $S_{m} = \{S_{m}^{t + 1},\dots,S_{m}^{t + n}\}$ . We then list the temporal views from each of the $M$ segments as $S_{1:M} = \{S_1^{t + 1},\dots,S_1^{t + n},\dots,S_M^{t + 1},\dots,S_M^{t + n}\}$ . + +To properly segment views of both static and moving objects, we assume a LiDAR at the commonly used frequency of $10\mathrm{Hz}$ , where the scan overlap is enough to cluster together also moving objects, as depicted in Fig. 3. We show further examples in the supplementary material. With this procedure, our augmented pairs are views of the same object at different times from static and moving objects. By maximizing the similarity between the object views at different times, the network needs to learn a more general representation consistent across time and embed the object dynamics. In the supplementary material, we provide an experimental comparison between the pre-training using the temporal object views as augmented pairs and with augmented pairs generated only with data augmentation to validate the use of temporal views. + +# 3.2. Temporal batch + +During pre-training, we need to sample a subset of $n$ consecutive scans from the training sequence to be aggregated and extract the temporal object views as explained in Sec. 3.1. This sample will be used during pre-training + +to learn a temporal consistent representation for the segmented objects. For a sequence of $N$ point clouds, we divide it into batches $\mathcal{B}$ with intervals of $n$ scans. During the pre-training forward pass, we sample two scans at different times, $t_1$ and $t_2$ . Since the goal of this temporal sampling is to have different views of the same object, we enforce this by sampling $t_1$ from the beginning and $t_2$ from the end of this $n$ scans interval such that $t_1 < \frac{n}{3}$ and $t_2 \geq 2\frac{n}{3}$ . In this case, the scans between $\frac{n}{3} \leq t < 2\frac{n}{3}$ would never be sampled. To also process those scans, the next batch starts at this unseen interval. Therefore, for a batch sample $\mathcal{B}_b$ the $n$ scans from the sequence of $N$ point clouds are $\mathcal{B}_b = \{\mathcal{P}^k | b\frac{n}{3} \leq k < b\frac{n}{3} + n\}$ . + +In Figure 4, we show how the batches would look like in an example with $n = 6$ scans interval. Even though we do not sample $t$ from $\frac{n}{3} \leq t < 2\frac{n}{3}$ , we still need to aggregate and cluster these scans to correctly segment dynamic objects and match their correspondences between the scans. + +With this sampling scheme, we guarantee that the objects views will be distant in time and still all the data will be seen during pre-training. + +# 3.3. Implicit clustering + +With our coarse object segments $S$ , we have the prior that points from the same segment should be semantically similar. Given that we rely on a set of object segments, one straightforward way to distinguish the learned embeddings is to cluster together point features from the same object. Recent methods propose an online clustering scheme to separate samples around a fixed number of prototype learnable cluster centers [10]. We aim at clustering point features from the same object close together but in a more straightforward way with an implicit clustering. + +Given the temporal sampled pair $\mathcal{P}^{t_1}$ and $\mathcal{P}^{t_2}$ from a batch sample $\mathcal{B}_b$ , we compute point-wise features $\mathcal{F}^{t_1}$ and $\mathcal{F}^{t_2}$ with the backbone. As the target embedding, we list from $\mathcal{F}^{t_2}$ the set of $M$ segments $\mathcal{S}^{t_2}$ . We compute for each segment a mean representation from its point-wise features, and project this embedding with a self-attention Transformer encoder as a projection head to get the $M$ target mean feature vectors $\overline{\boldsymbol{s}}^{t_2} \in \mathbb{R}^{M \times D}$ where $D$ corresponds to the feature dimension. + +The segments $S^{t_1}$ are listed from the point-wise features $\mathcal{F}^{t_1}$ . For $S^{t_1}$ , we do not compute mean embeddings but keep the features at point level, using the Transformer encoder to compute point-wise intra-class correspondences. To deal with the attention mechanism large memory requirement, we set a maximum number of $P$ points per segment, which are randomly sampled. After sampling $P$ points for each segment, we input the segments $S^{t_1}$ pointwise features to the Transformer projection head, followed by another self-attention Transformer encoder as a predictor. We then get for each segment the $P$ point-wise feature + +![](images/e51d1ef9df694a066e4299e8c7d050aaace530c05635b5c3bef02f6b7496c6c2.jpg) +Figure 5. Diagram of the implicit clustering scheme. We pool the segments $S^{t_1}$ and $S^{t_2}$ from the backbone features. For $t_2$ , we compute a mean representation for each segment and project it with the Transformer projection head. For $t_1$ , we project the pointwise features and then use the Transformer predictor to predict the corresponding mean representation from $t_2$ , clustering the point-wise features close to its mean target representation. + +vectors $s^{t_1} \in \mathbb{R}^{M \times P \times D}$ . + +With the segment target mean representations $\overline{s}^{t_2}$ and the predicted point-wise feature vectors $s^{t_1}$ for each segment, we compute the loss to minimize the differences between the point features and the corresponding segment mean representation. For each point $p$ from a segment $m$ , we compute the temperature-scaled cosine similarity $\delta_{m,p,k}^{t_1\to t_2}$ between the normalized point-wise embedding $s_{m,p}^{t_1}\in \mathbb{R}^D$ and the mean representation $\overline{s}_k^{t_2}\in \mathbb{R}^D$ from a segment $k$ as: + +$$ +\delta_ {m, p, k} ^ {t _ {1} \rightarrow t _ {2}} = \frac {\left(\boldsymbol {s} _ {m , p} ^ {t _ {1}}\right) ^ {\top} \bar {\boldsymbol {s}} _ {k} ^ {t _ {2}}}{\tau}. \tag {1} +$$ + +Next, we use the cross-entropy loss to maximize the similarity between each point $p$ from a segment $m$ and the corresponding target segment mean representation as follows: + +$$ +l _ {m} ^ {t _ {1} \rightarrow t _ {2}} = - \sum_ {p = 1} ^ {P} \log \left(\frac {\exp \left(\delta_ {m , p , m} ^ {t _ {1} \rightarrow t _ {2}}\right)}{\sum_ {k} ^ {M} \exp \left(\delta_ {m , p , k} ^ {t _ {1} \rightarrow t _ {2}}\right)}\right). \tag {2} +$$ + +We compute this loss for all the $M$ segments and sum it to get the loss for predicting segments from $t_2$ with the points of the segments from $t_1$ : + +$$ +\mathcal {L} ^ {t _ {1} \rightarrow t _ {2}} = \sum_ {m = 1} ^ {M} l _ {m} ^ {t _ {1} \rightarrow t _ {2}}. \tag {3} +$$ + +With this formulation, our task is to predict for each point in the segment at time $t_1$ the corresponding segment at $t_2$ . Since the target for all the $P$ point embeddings $s_{m,p}^{t_1}$ from a segment $m$ is the same mean segment representation $\overline{s}_m^{t_2}$ , the loss will push all the points from a segment to converge to a mean representation while separating from other segments, implicitly clustering together points from the same object as shown in Fig. 5. + +Lastly, we repeat the forward pass swapping $t_1$ and $t_2$ to have a symmetric representation, learning both the correspondences from $t_1 \rightarrow t_2$ and from $t_2 \rightarrow t_1$ leading to: + +$$ +\mathcal {L} _ {T A R L} = \mathcal {L} ^ {t _ {1} \rightarrow t _ {2}} + \mathcal {L} ^ {t _ {2} \rightarrow t _ {1}}. \tag {4} +$$ + +In the next section we give an intuition behind our choice of projection head, and how it can help the model to learn a more fine-grained representation during pre-training. + +# 3.4. Transformer projection + +Self-supervised representation learning methods [13, 14, 27] typically add a non-linear projection head to the backbone to project the embeddings to the target feature space. The idea is that the backbone should learn general features, and this projection head will overfit to the pre-training pretext task and later be discarded during fine-tuning. Our projection scheme follows more recent methods [15, 25], which shows that an asymmetric network can improve the learned representation. Besides, with the prior that points from one segment should be similar, we want to focus on the pointwise relationship. Therefore, we replace the non-linear projection head with a self-attention Transformer encoder. + +In our siamese network scheme, the online updated network computes point-wise features. In this case, the queries, keys, and values for our Transformer projector will be point-wise features. We aim to learn intra-class features with the attention mechanism and identify correspondences between the points from one segment. In contrast, the momentum updated network computes a mean feature vector for each segment. In this case, the attention mechanism should look for similarities and differences within the different objects segments in the scene. By using the self-attention over points and segments, we aim at learning at the same time point-wise correspondences while also learning the differences between the different segments. + +With the Transformer projector, the pre-training can lead the learned features towards a more fine-grained representation, where intra-class point-wise features are matched with the individual segment features. The supplementary material provides ablations comparing the Transformer with the commonly used non-linear projection head. + +# 3.5. Pre-training overview + +Secs. 3.1 to 3.4 explain the individual steps of our approach. This section summarizes the entire pre-training pipeline, putting together the modules explained in the previous subsections. + +Our pipeline shown in Fig. 2 fetches a batch sample $\mathcal{B}_b$ of $n$ sequential scans and applies a transformation to the scans $\mathcal{P}^t$ with the corresponding poses to have all points in a common coordinate frame. We use an unsupervised ground segmentation approach to remove the ground $\mathcal{G}$ . Next, we cluster the remaining points $\hat{\mathcal{P}}$ , defining coarse + +
MethodScribbles0.1%1%10%50%100%
Scratch54.9629.3542.7753.9658.2759.03
PointContrast [67]54.5232.6344.6258.6859.9861.45
DepthContrast [75]55.9031.6648.0557.1160.9961.14
SegContrast [50]56.7032.7544.8356.3160.4561.02
TARL (Ours)57.2538.5951.4260.3461.4261.47
+ +segments $S$ of the objects in the scene. We sample two random scans at different times $\mathcal{P}^{t_1}$ and $\mathcal{P}^{t_2}$ from $\mathcal{B}_b$ , in which we apply random augmentations $\mathcal{T}^{t_1}$ and $\mathcal{T}^{t_2}$ . We compute point-wise features $\mathcal{F}^{t_1}$ and $\mathcal{F}^{t_2}$ from the sampled point clouds with the backbone, and list the segments $S^{t_1}$ and $S^{t_2}$ with its corresponding point-wise features. We calculate a mean embedding $\overline{\mathcal{S}}^{t_2}$ for each segment $m$ in $\mathcal{P}^{t_2}$ and compute the target representation $\overline{s}^{t_2}$ for each segment with a Transformer projection head. For the segments $S^{t_1}$ , we project the point-wise features with another Transformer projector followed by a Transformer encoder as a predictor to get the point-wise features $s^{t_1}$ . Lastly, we compute the loss to predict the corresponding segment representation at $t_2$ for each point in $t_1$ . We repeat the process changing $t_1$ and $t_2$ to symmetrically match points from scan $t_2$ to the corresponding segment at $t_1$ . + +# 4. Experiments + +Datasets. We use the SemanticKITTI dataset [4,22] for our pre-training, an autonomous driving LiDAR data benchmark with point-wise annotations. During pre-training, we use only the raw point clouds from the training sequences and the given poses to extract the temporal objects views. For fine-tuning, we use SemanticKITTI with the full scans annotations and the scribbles annotations [51]. Besides SemanticKITTI, we also use nuScenes [8,64] for fine-tuning, to evaluate our approach in different autonomous driving datasets collected with different LiDARs and sensor setups, reporting the results on the validation sets. + +Model architecture. For our evaluation, we use a MinkUNet [20] as the backbone, which voxelizes the point clouds and uses sparse convolutions to extract features from the point cloud. The backbone receives the point cloud coordinates and intensity as input, and the output dimension feature has dimension $D = 96$ . We use a multi-head self-attention encoder for the projection head with one layer and eight attention heads. For the predictor, we use the same multi-head self-attention encoder architecture. Further details on the backbone used in our experiments are provided in the supplementary material. + +Pre-training. For pre-training, we use the AdamW optimizer [43] with a learning rate of $2 \cdot 10^{-4}$ and decay of $10^{-4}$ , + +Table 1. Results (mIoU) of the models pre-trained on SemanticKITTI fine-tuned to semantic segmentation on SemanticKITTI with different percentage of labels and scribbles. + +
MethodMiniFull
Scratch26.9466.03
Supervised pre-training38.3967.35
PointContrast [67]31.9267.31
DepthContrast [75]27.8164.70
SegContrast [50]31.2767.70
TARL (Ours)39.3668.26
+ +Table 2. Results (mIoU) of the self-supervised and supervised pretrained models on SemanticKITTI fine-tuned to semantic segmentation on nuScenes mini and full training sets. + +
MethodKITTInuScenes
Frozen backbone4.105.62
PointContrast [67]25.5319.70
DepthContrast [75]13.879.65
SegContrast [50]27.4223.75
TARL (Ours)35.9026.08
+ +Table 3. Linear evaluation results (mIoU) on SemanticKITTI and nuScenes datasets for semantic segmentation. + +training for 200 epochs with batch size 8 using a single NVIDIA RTX A6000. We use $n = 12$ for the scans to be aggregated. We provide experiments with different numbers of scans in the supplementary material to support this choice. The voxel resolution is set to $0.05m$ for the input point clouds, from which we sample a maximum of 40,000 points per point cloud. During the segment pooling, we limit it to a total of $M = 50$ segments with a maximum of $P = 300$ points per segment to avoid memory overflow. We use $\tau = 0.1$ to compute $\delta$ in Eq. (1) and momentum of 0.999 to update the momentum network according to the online network weights. For the baselines, we use their official repositories with their default parameters, also training them for 200 epochs and sampling 40,000 points per scan. + +Fine-tuning. To evaluate our method and compare it with the baselines, we fine-tune the models to three downstream tasks: semantic segmentation, panoptic segmentation, and object detection. For semantic segmentation, we use the pipeline provided in SegContrast [50] with the AdamW optimizer [43] and a learning rate of 0.001. For panoptic segmentation, we use the baseline evaluated by Hong et al. [29], where a 3D backbone network is used together with a semantic and an instance heads, followed by a clustering post-processing to identify the instances. We use the pre-trained MinkUNet model as the 3D backbone. We use a learning rate of 0.2 with batch size 8, training for 50 epochs. For object detection, we use the same toolbox and hyperparameters used in DepthContrast [75] with the PartA2 detector [57]. For all the tasks, we evaluate the method with the same model trained from scratch (without pre-training) and with the model after pre-training with + +
0.1%1%10%50%100%
MethodPQIoUPQIoUPQIoUPQIoUPQIoU
Scratch4.7611.1322.7230.8447.2053.5355.3261.9455.4059.75
PointContrast [67]5.8611.5127.3732.4947.5754.6354.2159.4855.8561.49
DepthContrast [75]7.6513.5627.3132.3046.8551.2754.5559.6056.1560.81
SegContrast [50]7.5814.4626.1432.8547.0253.4755.3860.0456.7361.96
TARL (Ours)10.2617.0129.2434.7151.2757.5956.1062.3656.5762.05
+ +Table 4. Results (PQ and IoU) when fine-tuning the pre-trained models to panoptic segmentation with different percentage of labels on SemanticKITTI. + +
MiniFull
PQIoUPQIoU
Scratch23.7823.9652.9858.17
Supervised pre-training24.7723.6053.1958.05
PointContrast [67]26.5825.4651.0656.39
DepthContrast [75]28.6627.3051.5157.06
SegContrast [50]28.8426.7952.3157.24
TARL (Ours)32.2230.7353.2659.14
+ +Table 5. Results (PQ and IoU) of the self-supervised and supervised pre-trained models on SemanticKITTI fine-tuned to panoramic segmentation on nuScenes mini and full training sets. + +the different self-supervised methods. The supplementary material provides detailed information for each downstream task training procedure. + +# 4.1. Semantic segmentation + +In this evaluation, we want to measure the information about semantics learned by the network during pretraining. We fine-tune it to semantic segmentation on SemanticKITTI and nuScenes after pre-training on SemanticKITTI. For SemanticKITTI, we use the percentage scans subsets used in SegContrast [50], with $0.1\%$ , $1\%$ , $10\%$ , $50\%$ , and $100\%$ of the labeled scans. We also evaluate on the scribbles annotations [51], which labels just a subset of points for each scan in the SemanticKITTI training sequences. We report the mean intersection-over-union (mIoU) of the models on the validation set, while fine-tuned on the aforementioned subset of labels from the training set. + +Tab. 1 shows the results from our approach compared to the baselines. As can be seen, overall the pre-training methods boost the performance compared to the network without pre-training on both scribbles annotations and the subsets of labeled scans. However, our approach surpasses previous state-of-the-art methods in all the different subsets, with a bigger gap when fewer labeled scans are used to fine-tune the model. We can also notice that, after pre-training with our method, only $10\%$ of labeled scans were necessary for the model to surpass the performance of the network trained from scratch with $100\%$ of the labeled scans. This result suggests that our method can effectively reduce the amount + +of labeled data necessary for the semantic segmentation task by around ten times. In the supplementary material, we also provide the IoU for each individual semantic class. + +Generalization. In this experiment, we evaluate the generalization of the learned features. We use the network pretrained on SemanticKITTI and fine-tune it on nuScenes. We use nuScenes full training set and the mini training subset to fine-tune the network, and evaluate on the full validation set. In this evaluation, we also compare the results of the fully supervised semantic segmentation pre-training on SemanticKITTI fine-tuned on nuScenes. As shown in Tab. 2, even though all methods can improve the performance of the model trained on nuScenes, our approach can surpass previous self-supervised methods by around $8\%$ when training with fewer labels. Besides, compared with supervised pre-training of the network on SemanticKITTI, our approach achieves better performance on both full and mini training sets. These results suggest that our method can replace supervised pre-training for LiDAR data, since it achieved better performance when fine-tuning to a different dataset, collected with a different sensor setup. + +Linear evaluation. In this experiment, the pre-trained backbone is frozen, training only a linear head on top of it to evaluate how descriptive the self-supervised learned representation is without fine-tuning it to the target task. We perform this evaluation on SemanticKITTI and nuScenes datasets. In Tab. 3, the randomly initialized network (frozen backbone) achieves low performance, suggesting that the features extracted by the backbone play the main role in achieving semantic segmentation. All the pre-training methods improve the performance compared to the network without pre-training. However, our approach has a clear performance gap compared to the baselines, especially when training to the same dataset used for pre-training. These results suggest that the representation learned by our method can embed more semantic information, even though not using labels, achieving higher performance than previous state-of-the-art methods on both datasets. + +# 4.2. Panoptic segmentation + +In this experiment, we fine-tune the pre-trained models for panoptic segmentation to also evaluate the instance-level features of our learned representation. Same as for semantic + +
CarPedestrianCyclist
EasyModerateHardEasyModerateHardEasyModerateHard
Scratch91.7582.1380.1667.6563.1357.7391.7474.9470.34
PointContrast [67]91.5581.7079.8867.3362.3256.5592.3375.2070.66
DepthContrast [75]91.9882.0480.0369.1363.9757.7192.6873.6770.10
SegContrast [50]92.0882.1080.1868.7263.7757.3691.2674.7070.28
TARL (Ours)92.0982.1580.1068.5064.0058.7492.6175.4470.97
+ +Table 6. Results (AP50) when fine-tuning the pre-trained models to object detection on KITTI full training set for car, pedestrian and cyclist classes on the easy, moderate and hard difficult levels. + +segmentation, we fine-tuned the model to SemanticKITTI using the same percentage subsets and nuScenes using the full and the mini training sets, reporting both the mIoU and the panoptic quality (PQ) on the validation sets. + +Tab. 4 shows that our method is consistently better than previous self-supervised pre-training approaches. As the amount of training samples increases, the differences diminish. However, our method has a clear performance gap compared to the baselines when trained with fewer labels. + +Generalization. To evaluate the generalization of our learned features on panoptic segmentation, we use the network pre-trained on SemantiKITTI to fine-tune it on nuScenes full and mini training sets, evaluating it on the full validation set. Tab. 5 shows that the representation learned by our method was more suited to be transferred to a different dataset. As in semantic segmentation, our method achieved better performance than previous self-supervised approaches and supervised pre-training on SemanticKITTI. These results agree with the semantic segmentation evaluation, suggesting that our method is better suited for transfer learning than the supervised pre-training. + +With these experiments, we can validate that our learned representations can extract not only semantic knowledge from the data but also instance-level information, surpassing previous methods in IoU and PQ metrics. In the supplementary material, we provide further results, reporting also the segmentation and recognition quality. + +# 4.3. Object detection + +As a third downstream task, we compare the methods fine-tuned for object detection to evaluate how general is the representation learned by our method compared with previous state-of-the-art methods. In this evaluation, we fine-tune the model with the whole training set of the KITTI dataset [22] and report the average precision $(\mathrm{AP}_{50})$ with 40 recall positions for car, pedestrian, and cyclist classes on the three difficulty levels. + +In Tab. 6, we can compare the baselines performance with our approach. For the car class, the pre-training methods achieve a marginal improvement compared to the network trained from scratch. For pedestrian and cyclist, the improvements brought by the pre-training are more substan + +tial. On those classes, our method achieves the best performance on moderate and hard difficulties, showing that our approach can also boost the performance when fine-tuning the model on object detection. In our supplementary material, we provide further experiments with the different training subsets used by Zhang et al. [75]. + +Together with the evaluations of the other downstream tasks, these results suggest that our method can learn a general representation. Our method achieved better performance than previous approaches on all three downstream tasks, showing that our approach learns a robust representation, not biased to a specific task but showing generalizable representations suitable for different tasks and datasets. + +# 5. Conclusion + +In this paper, we propose a self-supervised representation learning method for 3D LiDAR data evaluated in the context of autonomous driving. We exploit the vehicle motion to extract different views of objects across time to learn a temporally-consistent representation. We use a Transformer encoder as a projection head and implicitly clustered points from the same object together while discriminating between different objects. Our experiments show that our method achieves better performance than previous state-of-the-art approaches on different downstream tasks. Besides, our approach could reduce the amount of necessary labeled data to only $10\%$ when fine-tuning for semantic segmentation on SemanticKITTI. These results show that our proposed approach could learn a general LiDAR point cloud representation, embedding semantic information by matching objects representation viewed at different times. In addition, our method achieved better performance than supervised pre-training when transferring the learned representation to a dataset collected with different LiDAR sensors, suggesting that our approach could replace supervised pretraining in the LiDAR data domain. + +Limitations. Despite encouraging results, there is space for further improvement. For the unsupervised segment extraction, the clustering parameters have to be tuned depending on the LiDAR used to collect the data. Also, we rely on high-frequency data to properly segment dynamic objects. Otherwise, temporal segments could be wrongly matched. + +# References + +[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proc. of the Int. Conf. on Learning Representations (ICLR), 2021. 2 +[2] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 2 +[3] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyril Stachniss, Jens Behley, and Laura Leal-Taixe. 4D Panoptic LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2, 6 +[5] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS), 2018. 3, 4 +[6] Benedikt Mersch, Xieyuanli Chen, Ignacio Vizzo, Lucas Nunes, Jens Behley, and Cyril Stachniss. Receding Moving Object Segmentation in 3D LiDAR Data Using Sparse 4D Convolutions. IEEE Robotics and Automation Letters (RA-L), 7(3):7503-7510, 2022. 2 +[7] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised Domain Adaptation for LiDAR Panoptic Segmentation. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3 +[8] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6 +[9] Ricardo J. G. B. Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Advances in Knowledge Discovery and Data Mining, 2013. 3 +[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 4 +[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2 +[12] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. + +PTTR: Relational 3D Point Cloud Object Tracking With Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In Proc. of the Int. Conf. on Machine Learning (ICML), 2020. 1, 2, 5 +[14] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint, arxiv:2003.04297, 2020. 1, 2, 5 +[15] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5 +[16] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jirgen Gall, Jens Behley, and Cyril Stachniss. Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data. IEEE Robotics and Automation Letters (RA-L), 6:6529-6536, 2021. 2 +[17] Xieyuanli Chen, Andres Milioto, Emanuele Palazzolo, Philippe Giguere, Jens Behley, and Cyril Stachniss. SuMa++: Efficient LiDAR-based Semantic SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 3, 4 +[18] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[19] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PatchFormer: An Efficient Point Transformer With Patch Attention. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6 +[21] Fayao Liu, Guosheng Lin, and Chuan-Sheng Foo. Point discriminative learning for unsupervised representation learning on 3d point clouds. arXiv preprint, arXiv:2108.02104, 2021. 1, 2 +[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 1, 2, 6, 8 +[23] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[24] Benjamin Graham and Laurens van der Maaten. Submanifold Sparse Convolutional Networks. arXiv preprint, arxiv:1706.01307, 2017. 2 +[25] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and + +Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 5 +[26] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J. Kusner. Unsupervised Point Cloud Pre-Training via Occlusion Completion. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2 +[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5 +[28] Hiroharu Kato and Tatsuya Harada. Image Reconstruction from Bag-of-Visual-Words. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2 +[29] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. LiDAR-Based Panoptic Segmentation via Dynamic Shifting Network. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6 +[30] Hongxu Yin, Arash Vahdat, Jose Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-ViT: Adaptive Tokens for Efficient Vision Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[31] Ji Hou, Benjamin Graham, Matthias Niessner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding With Contrastive Scene Contexts. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3 +[32] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP - Simple, Accurate, and Robust Registration If Done the Right Way. arXiv preprint arXiv:2209.15397, 2022. 3, 4 +[33] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[34] Ishan Misra, Rohit Girdharm, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2 +[35] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. ST3D: Self-Training for Unsupervised Domain Adaptation on 3D Object Detection. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[36] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposalcontrast: Unsupervised pre-training for lidar-based 3d object detection. Proc. of the Europ. Conf. on Computer Vision (ECCV),2022.1,2,3 +[37] Kai Fischer, Martin Simon, Florian Oelsner, Stefan Milz, Horst-Michael Gross, and Patrick Maeder. StickyPillars: Robust and Efficient Feature Matching on Point Clouds Using Graph Neural Networks. In Proc. of the IEEE/CVF + +Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[38] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. PointPillars: Fast Encoders for Object Detection From Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[39] Ferdinand Langer, Andres Milioto, Alexandre Haag, Jens Behley, and Cyril Stachniss. Domain Transfer for Semantic Segmentation of LiDAR Data using Deep Neural Networks. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 3 +[40] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph Attention Convolution for Point Cloud Semantic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[41] Hyungtae Lim, Sungwon Hwang, and Hyun Myung. ERASOR: Egocentric Ratio of Pseudo Occupancy-Based Dynamic Object Removal for Static 3D Point Cloud Map Building. IEEE Robotics and Automation Letters (RA-L), 6(2):2272-2279, 2021. 3 +[42] Ling Zhang and Zhigang Zhu. Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks. In Proc. of the Intl. Conf. on 3D Vision (3DV), 2019. 1, 2 +[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2019. 6 +[44] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Contrastive Instance Association for 4D Panoptic Segmentation using Sequences of 3D LiDAR Scans. IEEE Robotics and Automation Letters (RA-L), 7(2):1550-1557, 2022. 2 +[45] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. LiDAR Panoptic Segmentation for Autonomous Driving. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 1 +[46] Andres Milioto, Leonard Mandtler, and Cyril Stachniss. Fast Instance and Semantic Segmentation Exploiting Local Connectivity, Metric Learning, and One-Shot Detection for Robotics. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 1, 2 +[47] Andres Milioto and Cyril Stachniss. Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics using CNNs. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 2 +[48] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-BERT: Vision-Language Pre-Training on Fashion Domain. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[49] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 + +[50] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. SegContrast: 3D Point Cloud Feature Representation Learning through Self-supervised Segment Discrimination. IEEE Robotics and Automation Letters (RA-L), 2022. 1, 2, 3, 6, 7, 8 +[51] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 7 +[52] Pedro O. Pinheiro, Amjad Almahairi, Ryan Y. Benmalek, Florian Golemo, and Aaron Courville. Unsupervised learning of dense visual representations. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), pages 4489-4500, 2020. 1, 2 +[53] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP Real-Time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3, 4 +[54] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2 +[55] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2 +[56] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Proc. of the Europ. Conf. on Computer Vision (ECCV), pages 626-642, 2020. 1, 2 +[57] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network. IEEE Trans. on Pattern Analalysis and Machine Intelligence (TPAMI), 43:2647-2664, 2021. 2, 6 +[58] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointR-CNN: 3D Object Proposal Generation and Detection From Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[59] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-Level Representation Learning for Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[60] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2 +[61] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPCov: Flexible and Deformable Convolution for Point Clouds. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2 +[62] Weijing Shi and Ragunathan (Raj) Rajkumar. Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2 + +[63] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring Cross-Image Pixel Contrast for Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2 +[64] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters (RA-L), 7(2):3795-3802, 2022. 2, 6 +[65] Xiangyun Zhao, Raviteja Vemulapalli, Philip Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive Learning for Label Efficient Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2 +[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[67] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 6, 7, 8 +[68] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2 +[69] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense Contrastive Learning for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[70] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[71] Li Yi, Boqing Gong, and Thomas Funkhouser. Complete & Label: A Domain Adaptation Approach to Semantic Segmentation of LiDAR Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[72] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. MPViT: Multi-Path Vision Transformer for Dense Prediction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[73] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust Instance Segmentation Through Reasoning About Multi-Object Occlusion. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[74] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. Proc. of the Europ. Conf. on Computer Vision (ECCV), 2022. 1, 2, 3 +[75] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-Supervised Pretraining of 3D Features on any + +Point-Cloud. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2, 3, 6, 7, 8 +[76] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In Proc. of the Int. Conf. on Machine Learning (ICML), 2021. 1, 2 +[77] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2 +[78] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring Set Similarity for Dense Self-Supervised Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[79] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-Free LiDAR Point Cloud Panoptic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[80] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 \ No newline at end of file diff --git a/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/images.zip b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ea77c4881eb61f39c60953757d0cef3bc1c0eac --- /dev/null +++ b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25979710d8d3955e7fa81f2fd1ba30b8f2cd758d19df170cf53617f935dc0fde +size 427413 diff --git a/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/layout.json b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..487bb0ff0bf13679f9caff383cf15865d29b3b81 --- /dev/null +++ b/2023/Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving/layout.json @@ -0,0 +1,10759 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "type": "text", + "content": "Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 157, + 222, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 157, + 222, + 170 + ], + "spans": [ + { + "bbox": [ + 152, + 157, + 222, + 170 + ], + "type": "text", + "content": "Lucas Nunes*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 171, + 239, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 171, + 239, + 185 + ], + "spans": [ + { + "bbox": [ + 157, + 171, + 239, + 185 + ], + "type": "text", + "content": "Xieyuanli Chen*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 157, + 328, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 157, + 328, + 171 + ], + "spans": [ + { + "bbox": [ + 240, + 157, + 328, + 171 + ], + "type": "text", + "content": "Louis Wiesmann*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 257, + 171, + 321, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 171, + 321, + 185 + ], + "spans": [ + { + "bbox": [ + 257, + 171, + 321, + 185 + ], + "type": "text", + "content": "Jens Behley*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 345, + 157, + 440, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 157, + 440, + 171 + ], + "spans": [ + { + "bbox": [ + 345, + 157, + 440, + 171 + ], + "type": "text", + "content": "Rodrigo Marcuzzi*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 340, + 171, + 435, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 171, + 435, + 185 + ], + "spans": [ + { + "bbox": [ + 340, + 171, + 435, + 185 + ], + "type": "text", + "content": "Cyrill Stachniss\\*,\\*" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 183, + 186, + 410, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 186, + 410, + 199 + ], + "spans": [ + { + "bbox": [ + 183, + 186, + 410, + 199 + ], + "type": "text", + "content": "*University of Bonn, *University of Oxford, ‡Lamarr Institute" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 198, + 201, + 392, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 201, + 392, + 213 + ], + "spans": [ + { + "bbox": [ + 198, + 201, + 392, + 213 + ], + "type": "text", + "content": "{firstname.lastname}@igg.uni-bonn.de" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 232, + 192, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 232, + 192, + 244 + ], + "spans": [ + { + "bbox": [ + 143, + 232, + 192, + 244 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 249, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 290, + 574 + ], + "type": "text", + "content": "Semantic perception is a core building block in autonomous driving, since it provides information about the drivable space and location of other traffic participants. For learning-based perception, often a large amount of diverse training data is necessary to achieve high performance. Data labeling is usually a bottleneck for developing such methods, especially for dense prediction tasks, e.g., semantic segmentation or panoptic segmentation. For 3D LiDAR data, the annotation process demands even more effort than for images. Especially in autonomous driving, point clouds are sparse, and objects appearance depends on its distance from the sensor, making it harder to acquire large amounts of labeled training data. This paper aims at taking an alternative path proposing a self-supervised representation learning method for 3D LiDAR data. Our approach exploits the vehicle motion to match objects across time viewed in different scans. We then train a model to maximize the point-wise feature similarities from points of the associated object in different scans, which enables to learn a consistent representation across time. The experimental results show that our approach performs better than previous state-of-the-art self-supervised representation learning methods when fine-tuning to different downstream tasks. We furthermore show that with only " + }, + { + "bbox": [ + 46, + 249, + 290, + 574 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 249, + 290, + 574 + ], + "type": "text", + "content": " of labeled data, a network pre-trained with our approach can achieve better performance than the same network trained from scratch with all labels for semantic segmentation on SemanticKITTI." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 586, + 128, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 599 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 607, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 693 + ], + "type": "text", + "content": "Semantic perception is essential for safe interaction between autonomous vehicles and their surrounding. For learning-based perception, a massive amount of training data is usually required for training high-performance models. However, the data annotation is the bottleneck of collecting such large training sets, especially for dense prediction tasks, such as semantic segmentation [60, 80], instance" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 426, + 377 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 426, + 377 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 426, + 377 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 426, + 377 + ], + "type": "image", + "image_path": "a475f1413dde086649b200520421f2454fdc05c1f21b7c8d2a8d298cbddb7d0f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 387, + 547, + 444 + ], + "lines": [ + { + "bbox": [ + 304, + 387, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 547, + 444 + ], + "type": "text", + "content": "Figure 1. Our pre-training (TARL) can reduce the amount of necessary labels during fine-tuning on SemanticKITTI [4, 22]. Our method requires only " + }, + { + "bbox": [ + 304, + 387, + 547, + 444 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 387, + 547, + 444 + ], + "type": "text", + "content": " of labels to surpass the network trained from scratch with the full dataset for semantic segmentation. For panoptic segmentation, our method requires only half of the labels." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 427, + 232, + 545, + 377 + ], + "blocks": [ + { + "bbox": [ + 427, + 232, + 545, + 377 + ], + "lines": [ + { + "bbox": [ + 427, + 232, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 427, + 232, + 545, + 377 + ], + "type": "image", + "image_path": "0f093a366a4e33cb73fc9af0e7482fcf8d2953e25f8684f8e04d54414f1b860f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 449, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 532 + ], + "type": "text", + "content": "segmentation [46, 73], and panoptic segmentation [45, 79], which require fine-grained labels. In the context of autonomous driving, recent approaches exploit 3D LiDAR data [20, 54, 55, 61], where the data annotation process is more complex than on 2D RGB images due to the sparsity of the point cloud and object appearance varying with its distance to the sensor." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "content": "Recent self-supervised representation learning methods [10, 11, 13-15, 25, 27, 76] tackle the lack of annotated data with a pre-training step requiring no labels. Those methods use data augmentation to generate different views from one data sample and train the network to learn an embedding space able to have similar representations for the generated augmented views. Other approaches [52, 63, 65, 69, 78] propose optimizing the pixel embedding space to learn a dense representation suited to be fine-tuned to more fine-grained tasks. For 3D point cloud data, recent approaches [21, 42, 49, 56] focus on synthetic point clouds of single objects to learn a robust representation for object classification. Other approaches [31, 36, 50, 67, 74, 75] target real-world data representation, such as LiDAR or RGB-D data, but fewer target autonomous driving scenarios." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "text", + "content": "Code: https://github.com/PRBonn/TARL" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5217" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "In this paper, we propose a novel temporal association representation learning (TARL) method for 3D LiDAR data designed for autonomous driving data. We exploit the vehicle motion to extract different views of the same objects across time. Then, we compute point-wise features from the objects in the point cloud and use a Transformer encoder [2] as a projection head to learn a temporal association from the object representation, embedding the objects dynamics. We conduct extensive experiments to compare our approach with state-of-the-art methods and show that our approach surpasses previous self-supervised pretraining methods [50, 67, 75] when fine-tuning to different downstream tasks and datasets [4, 8, 22, 64]. In summary, our key contributions are:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 251, + 286, + 452 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 251, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 251, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 286, + 299 + ], + "type": "text", + "content": "- We propose a novel self-supervised pre-training for 3D LiDAR data able to learn a robust and temporally-consistent representation by clustering together points from the same object viewed at different points in time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 303, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 303, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 286, + 350 + ], + "type": "text", + "content": "- We achieve better performance than previous state-of-the-art methods on different downstream tasks, i.e., semantic segmentation, panoptic segmentation, and object detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 354, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 354, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 286, + 400 + ], + "type": "text", + "content": "- With our pre-training, we require only " + }, + { + "bbox": [ + 58, + 354, + 286, + 400 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 58, + 354, + 286, + 400 + ], + "type": "text", + "content": " of labels to surpass the network trained from scratch for semantic segmentation using the whole training set on SemanticKITTI (see Fig. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 405, + 286, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 405, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 58, + 405, + 286, + 452 + ], + "type": "text", + "content": "- Our self-supervised pre-training produces representations more suited for transfer learning than supervised pre-training, achieving better performance when finetuning to a different dataset." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 465, + 134, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 134, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 134, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Self-supervised representation learning aims at initializing weights of a deep neural network for a downstream task by learning suitable representations without labeled data. Early approaches [18, 26, 28, 33, 48, 70] use so-called pretext tasks to learn useful representations, such as solving jigsaw puzzles [28, 33, 48] or reconstructing masked data [18, 26, 70]. More recently, contrastive learning [13, 14, 27] has drawn significant attention in the vision and robotics research community. Those contrastive methods employ random data augmentation over one image sample to generate different views of it. Then, the network is trained to learn an embedding space able to discriminate between the augmented samples from the same image and the augmented views from other images. Recent approaches exploit this discriminative pretext task and propose different ways to learn a self-supervised data representation, e.g., via non-contrastive methods [15, 25], redundancy reduction [76], or online clustering [10]. Other works [52, 63, 65, 69, 78] approach this in a pixel-wise man" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "ner, learning dense representations applied to more fine-grained downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 98, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 545, + 229 + ], + "type": "text", + "content": "Transformer encoders [2] were initially proposed in the natural language processing community achieving state-of-the-art performance in the field. Such architectures were then quickly adopted by the computer vision community [1, 30, 66, 72, 77]. With the attention mechanism, the Transformer can learn the global context from the image instead of just local neighboring context as in convolutional models [1]. With the growing interest in Transformers, some self-supervised representation learning methods were applied to Transformers [11, 59, 68] to evaluate the capacity of such models to learn meaningful representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 231, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 546, + 445 + ], + "type": "text", + "content": "LiDAR point cloud perception brings further challenges compared to the image domain. Previous works [46, 47] deal with 3D scans by projecting the data to a 2D range image and then processing it with 2D convolutional methods. More recently, different convolution architectures were proposed to deal with 3D data, such as point kernels [54, 55, 61], sparse convolutions [20, 23, 24], or graph representations [37, 40, 62]. With those different architectures, it becomes possible to train a model to learn from 3D data and deal with different tasks in this domain, such as semantic segmentation [60,80], panoptic segmentation [3, 29, 44], object detection [38, 57, 58], or moving object segmentation [6, 16]. With increasing interest in Transformers, some works [12, 19, 34, 70] proposed ways of dealing with point cloud data in such architectures. In this domain, a key challenge lies in the number of points collected by LiDAR sensors together with the high memory requirements from the Transformer attention mechanism." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 687 + ], + "type": "text", + "content": "3D self-supervised representation learning aims at learning robust representations from point cloud data. Most of those methods focus on point clouds of single objects [21, 42,49,56]. Other methods [31,36,50,67,74,75] target dense point-wise representations for real-world point cloud scans. Xie et al. [67] propose learning a representation via point-to-point contrastive learning through matching corresponding points in two augmented point clouds. Similarly, Hou et al. [31] learn both point-wise correspondence and the region where the point is located in the point cloud. Zhang et al. [75] propose global representation learning through scan-to-scan discrimination. Chen et al. [74] synthetically add dynamic objects in the scene to embed temporal and dynamic understanding in the learned representation. Other methods propose discriminating regions of interest to extract more semantic information. Nunes et al. [50] extract coarse segments from point clouds and use contrastive loss to distinguish between objects segments. Similarly, Yin et al. [36] extract proposal regions to be discriminated, biasing the pre-training to object detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Despite the performance improvement brought by those 3D self-supervised representation learning methods, there is" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5218" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 71, + 521, + 211 + ], + "blocks": [ + { + "bbox": [ + 83, + 71, + 521, + 211 + ], + "lines": [ + { + "bbox": [ + 83, + 71, + 521, + 211 + ], + "spans": [ + { + "bbox": [ + 83, + 71, + 521, + 211 + ], + "type": "image", + "image_path": "03c5c449c0b5312b5ae1b6082ef0e68336b71dac4d67ed8df76dcb0dbfb99322.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": "Figure 2. We aggregate " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " point clouds and extract segments " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " from it by removing the ground and clustering the remaining points. Then we sample two point clouds " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_1}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": ", and apply random augmentations " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{t_1}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " to them. Next, we compute point-wise features for each scan and list its segments " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "S^{t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " with their corresponding point features. For " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "S^{t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " we compute the mean feature vector for each segment and compute the segments target embeddings " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\overline{s}^{t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " with a projection head. For " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " we use a point-wise projector followed by a predictor to identify for each segment point features " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "s_{m,p}^{t_1}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " the corresponding target embedding " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\overline{s}_m^{t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": " by minimizing the loss " + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{t_1\\to t_2}" + }, + { + "bbox": [ + 46, + 216, + 547, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 410 + ], + "type": "text", + "content": "still a performance gap compared to the image domain due to the lack of data augmentations for point clouds. Similar to previous works [36,50], our approach focuses on learning representations over regions of interest. However, unlike previous methods, we extract natural augmented views of the objects in the scans collected at different points in time while the vehicle drives in the environment. Then, the network is trained to learn a common representation for those temporal views of one object. This enables our method to learn a more robust semantic representation by learning temporal consistent features and real-world dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 416, + 104, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 104, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 104, + 428 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 628 + ], + "type": "text", + "content": "In this paper, we propose a new self-supervised representation learning method for LiDAR data obtained in the context of autonomous driving. Our approach requires only unlabeled point clouds and the corresponding scan pose. Pose information is usually readily available by means of GPS/IMU, odometry approaches [32, 53], or SLAM systems [5,17], thus, this is not a limitation in practice. We use a siamese network scheme with an online updated network and a momentum updated network (see Fig. 2). The whole pipeline consists of extracting objects as coarse segments viewed at different times over an interval of scans. Then, point-wise features are computed with the backbone and a Transformer encoder as a projector. Finally, we perform an implicit clustering to put points from different views of the same object together. In the following subsections, we explain the individual steps of our method thoroughly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 179, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 179, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 179, + 647 + ], + "type": "text", + "content": "3.1. Temporal objects views" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "Instead of only using data augmentation to generate artificial views of one object, we exploit the vehicle motion to extract real-world object segments viewed from different perspectives. Given the vehicle motion and the properties of LiDAR sensors, one object can have different ap" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 281, + 533, + 422 + ], + "blocks": [ + { + "bbox": [ + 319, + 281, + 533, + 422 + ], + "lines": [ + { + "bbox": [ + 319, + 281, + 533, + 422 + ], + "spans": [ + { + "bbox": [ + 319, + 281, + 533, + 422 + ], + "type": "image", + "image_path": "9d6ea86d1691e11ba8e8aea2a9dde9b8082115d8af87298f2cf3c4bf4d41e3f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 427, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 427, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 547, + 473 + ], + "type": "text", + "content": "Figure 3. We aggregate " + }, + { + "bbox": [ + 305, + 427, + 547, + 473 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 427, + 547, + 473 + ], + "type": "text", + "content": " point clouds from different times and extract coarse segments from the aggregated point cloud in an unsupervised manner. We keep the point indices mapping to extract views at different times from one object as augmented pairs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 483, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 639 + ], + "type": "text", + "content": "pearances depending on its position relative to the sensor. Such changing appearance is usually a problem when fine-tuning the model to data collected with different sensors [7, 35, 39, 71]. However, we exploit these properties in our favor to extract natural augmented views of one object. Previous works [31, 36, 50, 67, 74, 75] relied only on data augmentation to generate pairs of one scan, such as rotation, translation and scaling. Instead, we extract object views collected from different perspectives by the LiDAR sensor while the vehicle navigates through the environment. We then apply augmentations over the associated scans captured at a different time and train the network to learn a common representation for the object views." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "We define a point cloud at time " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t = \\{\\pmb {p}_1^t,\\dots ,\\pmb {p}_R^t\\}" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " as a set of 3D points " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {p}_r\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " . To extract coarse object segments in an individual scan, we first separate " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " into ground " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^t" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " and non-ground points " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^t" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " in an unsupervised manner using the method proposed by Lim et al. [41], where " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t = \\hat{\\mathcal{P}}^t\\cup \\mathcal{G}^t" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^t\\cap \\mathcal{G}^t = \\emptyset" + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": " .Af" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5219" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 276, + 137 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 276, + 137 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 276, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 276, + 137 + ], + "type": "image", + "image_path": "a7dab86986ef2f1609875e06913fafd6596bd7a787b255eb8a5459a0bb303670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": "Figure 4. Batch example, for an interval of " + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "inline_equation", + "content": "n = 6" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": " point clouds " + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": ". We divide a sequence of " + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": " scans into batches " + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_b" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": " where we sample " + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": " from the beginning and " + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 145, + 287, + 191 + ], + "type": "text", + "content": " from the end of the interval. Then, the next batch starts in the middle of the previous interval." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "text", + "content": "terwards, we cluster the non-ground points " + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^t" + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "text", + "content": " segments " + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^t = \\{S_1^t,\\dots,S_M^t\\}" + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "text", + "content": " using HDBSCAN [9], where " + }, + { + "bbox": [ + 46, + 201, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}^t = \\bigcup_{m = 1}^{M}\\mathcal{S}_m^t" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": "To map different segment views at different times, we define an interval of " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " scans from which the views of the object will be extracted. Those scans are transformed to a common global coordinate frame to be then aggregated. The global scan poses can be easily acquired with GPS/IMU, SLAM systems [5, 17], or Li-DAR odometry [32, 53]. The aggregated point cloud " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathcal{P}^{t + 1},\\mathcal{P}^{t + 2},\\dots,\\mathcal{P}^{t + n}\\}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": ". Similarly, we can aggregate the individual ground segmentation labels " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{G}^{t + 1},\\mathcal{G}^{t + 2},\\dots,\\mathcal{G}^{t + n}\\}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " and get the aggregated non-ground points " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}} = \\{\\hat{\\mathcal{P}}^{t + 1},\\hat{\\mathcal{P}}^{t + 2},\\dots,\\hat{\\mathcal{P}}^{t + n}\\}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": ". As in the individual scan case, we can cluster " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " to get the " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " segments " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "S = \\{S_{1},\\dots,S_{M}\\}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": ". By keeping the point index mapping from the aggregated point cloud to the individual " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " scans, we can identify the " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " segments of the same object viewed at different times as " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "S_{m} = \\{S_{m}^{t + 1},\\dots,S_{m}^{t + n}\\}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": ". We then list the temporal views from each of the " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": " segments as " + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "inline_equation", + "content": "S_{1:M} = \\{S_1^{t + 1},\\dots,S_1^{t + n},\\dots,S_M^{t + 1},\\dots,S_M^{t + n}\\}" + }, + { + "bbox": [ + 46, + 240, + 286, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 286, + 635 + ], + "type": "text", + "content": "To properly segment views of both static and moving objects, we assume a LiDAR at the commonly used frequency of " + }, + { + "bbox": [ + 46, + 457, + 286, + 635 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 457, + 286, + 635 + ], + "type": "text", + "content": ", where the scan overlap is enough to cluster together also moving objects, as depicted in Fig. 3. We show further examples in the supplementary material. With this procedure, our augmented pairs are views of the same object at different times from static and moving objects. By maximizing the similarity between the object views at different times, the network needs to learn a more general representation consistent across time and embed the object dynamics. In the supplementary material, we provide an experimental comparison between the pre-training using the temporal object views as augmented pairs and with augmented pairs generated only with data augmentation to validate the use of temporal views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 646, + 144, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 144, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 144, + 658 + ], + "type": "text", + "content": "3.2. Temporal batch" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "During pre-training, we need to sample a subset of " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " consecutive scans from the training sequence to be aggregated and extract the temporal object views as explained in Sec. 3.1. This sample will be used during pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "to learn a temporal consistent representation for the segmented objects. For a sequence of " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " point clouds, we divide it into batches " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " with intervals of " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " scans. During the pre-training forward pass, we sample two scans at different times, " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": ". Since the goal of this temporal sampling is to have different views of the same object, we enforce this by sampling " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " from the beginning and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " from the end of this " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " scans interval such that " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "t_1 < \\frac{n}{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "t_2 \\geq 2\\frac{n}{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": ". In this case, the scans between " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\frac{n}{3} \\leq t < 2\\frac{n}{3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " would never be sampled. To also process those scans, the next batch starts at this unseen interval. Therefore, for a batch sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_b" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " scans from the sequence of " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " point clouds are " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_b = \\{\\mathcal{P}^k | b\\frac{n}{3} \\leq k < b\\frac{n}{3} + n\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "content": "In Figure 4, we show how the batches would look like in an example with " + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "inline_equation", + "content": "n = 6" + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "content": " scans interval. Even though we do not sample " + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\frac{n}{3} \\leq t < 2\\frac{n}{3}" + }, + { + "bbox": [ + 304, + 228, + 545, + 288 + ], + "type": "text", + "content": ", we still need to aggregate and cluster these scans to correctly segment dynamic objects and match their correspondences between the scans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 325 + ], + "type": "text", + "content": "With this sampling scheme, we guarantee that the objects views will be distant in time and still all the data will be seen during pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 334, + 415, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 415, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 415, + 346 + ], + "type": "text", + "content": "3.3. Implicit clustering" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 353, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 472 + ], + "type": "text", + "content": "With our coarse object segments " + }, + { + "bbox": [ + 304, + 353, + 545, + 472 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 353, + 545, + 472 + ], + "type": "text", + "content": ", we have the prior that points from the same segment should be semantically similar. Given that we rely on a set of object segments, one straightforward way to distinguish the learned embeddings is to cluster together point features from the same object. Recent methods propose an online clustering scheme to separate samples around a fixed number of prototype learnable cluster centers [10]. We aim at clustering point features from the same object close together but in a more straightforward way with an implicit clustering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": "Given the temporal sampled pair " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_1}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_2}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " from a batch sample " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_b" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": ", we compute point-wise features " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{t_1}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{t_2}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " with the backbone. As the target embedding, we list from " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{t_2}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " the set of " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " segments " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^{t_2}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": ". We compute for each segment a mean representation from its point-wise features, and project this embedding with a self-attention Transformer encoder as a projection head to get the " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " target mean feature vectors " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\overline{\\boldsymbol{s}}^{t_2} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": " corresponds to the feature dimension." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The segments " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " are listed from the point-wise features " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{t_1}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", we do not compute mean embeddings but keep the features at point level, using the Transformer encoder to compute point-wise intra-class correspondences. To deal with the attention mechanism large memory requirement, we set a maximum number of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " points per segment, which are randomly sampled. After sampling " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " points for each segment, we input the segments " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " pointwise features to the Transformer projection head, followed by another self-attention Transformer encoder as a predictor. We then get for each segment the " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " point-wise feature" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5220" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 275, + 190 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 275, + 190 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 275, + 190 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 275, + 190 + ], + "type": "image", + "image_path": "e51d1ef9df694a066e4299e8c7d050aaace530c05635b5c3bef02f6b7496c6c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "text", + "content": "Figure 5. Diagram of the implicit clustering scheme. We pool the segments " + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "inline_equation", + "content": "S^{t_2}" + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "text", + "content": " from the backbone features. For " + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "text", + "content": ", we compute a mean representation for each segment and project it with the Transformer projection head. For " + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "text", + "content": ", we project the pointwise features and then use the Transformer predictor to predict the corresponding mean representation from " + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 198, + 289, + 275 + ], + "type": "text", + "content": ", clustering the point-wise features close to its mean target representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 287, + 149, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 149, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 149, + 300 + ], + "type": "text", + "content": "vectors " + }, + { + "bbox": [ + 47, + 287, + 149, + 300 + ], + "type": "inline_equation", + "content": "s^{t_1} \\in \\mathbb{R}^{M \\times P \\times D}" + }, + { + "bbox": [ + 47, + 287, + 149, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": "With the segment target mean representations " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\overline{s}^{t_2}" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " and the predicted point-wise feature vectors " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "s^{t_1}" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " for each segment, we compute the loss to minimize the differences between the point features and the corresponding segment mean representation. For each point " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " from a segment " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": ", we compute the temperature-scaled cosine similarity " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\delta_{m,p,k}^{t_1\\to t_2}" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " between the normalized point-wise embedding " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "s_{m,p}^{t_1}\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " and the mean representation " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\overline{s}_k^{t_2}\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " from a segment " + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 300, + 287, + 402 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 410, + 287, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 410, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 119, + 410, + 287, + 437 + ], + "type": "interline_equation", + "content": "\\delta_ {m, p, k} ^ {t _ {1} \\rightarrow t _ {2}} = \\frac {\\left(\\boldsymbol {s} _ {m , p} ^ {t _ {1}}\\right) ^ {\\top} \\bar {\\boldsymbol {s}} _ {k} ^ {t _ {2}}}{\\tau}. \\tag {1}", + "image_path": "178312001bb1856aea9c7669b18ac89ac0def71bbdacdbcd7a4d903886a82f47.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "type": "text", + "content": "Next, we use the cross-entropy loss to maximize the similarity between each point " + }, + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "type": "text", + "content": " from a segment " + }, + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 442, + 287, + 479 + ], + "type": "text", + "content": " and the corresponding target segment mean representation as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 488, + 287, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 488, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 78, + 488, + 287, + 522 + ], + "type": "interline_equation", + "content": "l _ {m} ^ {t _ {1} \\rightarrow t _ {2}} = - \\sum_ {p = 1} ^ {P} \\log \\left(\\frac {\\exp \\left(\\delta_ {m , p , m} ^ {t _ {1} \\rightarrow t _ {2}}\\right)}{\\sum_ {k} ^ {M} \\exp \\left(\\delta_ {m , p , k} ^ {t _ {1} \\rightarrow t _ {2}}\\right)}\\right). \\tag {2}", + "image_path": "20d6539084dfe94ed7aacedd22400f1e03d0f97a5ca085e8fa014a26b0c91a73.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "text", + "content": "We compute this loss for all the " + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "text", + "content": " segments and sum it to get the loss for predicting segments from " + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "text", + "content": " with the points of the segments from " + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 47, + 528, + 287, + 565 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 574, + 287, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 574, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 121, + 574, + 287, + 607 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {t _ {1} \\rightarrow t _ {2}} = \\sum_ {m = 1} ^ {M} l _ {m} ^ {t _ {1} \\rightarrow t _ {2}}. \\tag {3}", + "image_path": "6b477cb66d3bd871fb4c6e8ae2cc75e6d7564c9bdf8760c1800ff4c2ead0654b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": "With this formulation, our task is to predict for each point in the segment at time " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " the corresponding segment at " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": ". Since the target for all the " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " point embeddings " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "s_{m,p}^{t_1}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " from a segment " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is the same mean segment representation " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\overline{s}_m^{t_2}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": ", the loss will push all the points from a segment to converge to a mean representation while separating from other segments, implicitly clustering together points from the same object as shown in Fig. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "Lastly, we repeat the forward pass swapping " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " to have a symmetric representation, learning both the correspondences from " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "t_1 \\rightarrow t_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " and from " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "t_2 \\rightarrow t_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " leading to:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 119, + 545, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 119, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 365, + 119, + 545, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T A R L} = \\mathcal {L} ^ {t _ {1} \\rightarrow t _ {2}} + \\mathcal {L} ^ {t _ {2} \\rightarrow t _ {1}}. \\tag {4}", + "image_path": "123e31255c9e3ad9f7b69a39d658873f1dc820fc5fb52ba11199f58f4d25f8ef.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 174 + ], + "type": "text", + "content": "In the next section we give an intuition behind our choice of projection head, and how it can help the model to learn a more fine-grained representation during pre-training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 179, + 440, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 179, + 440, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 179, + 440, + 192 + ], + "type": "text", + "content": "3.4. Transformer projection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 198, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 341 + ], + "type": "text", + "content": "Self-supervised representation learning methods [13, 14, 27] typically add a non-linear projection head to the backbone to project the embeddings to the target feature space. The idea is that the backbone should learn general features, and this projection head will overfit to the pre-training pretext task and later be discarded during fine-tuning. Our projection scheme follows more recent methods [15, 25], which shows that an asymmetric network can improve the learned representation. Besides, with the prior that points from one segment should be similar, we want to focus on the pointwise relationship. Therefore, we replace the non-linear projection head with a self-attention Transformer encoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 342, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 496 + ], + "type": "text", + "content": "In our siamese network scheme, the online updated network computes point-wise features. In this case, the queries, keys, and values for our Transformer projector will be point-wise features. We aim to learn intra-class features with the attention mechanism and identify correspondences between the points from one segment. In contrast, the momentum updated network computes a mean feature vector for each segment. In this case, the attention mechanism should look for similarities and differences within the different objects segments in the scene. By using the self-attention over points and segments, we aim at learning at the same time point-wise correspondences while also learning the differences between the different segments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 497, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 569 + ], + "type": "text", + "content": "With the Transformer projector, the pre-training can lead the learned features towards a more fine-grained representation, where intra-class point-wise features are matched with the individual segment features. The supplementary material provides ablations comparing the Transformer with the commonly used non-linear projection head." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 575, + 430, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 430, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 430, + 588 + ], + "type": "text", + "content": "3.5. Pre-training overview" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 640 + ], + "type": "text", + "content": "Secs. 3.1 to 3.4 explain the individual steps of our approach. This section summarizes the entire pre-training pipeline, putting together the modules explained in the previous subsections." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Our pipeline shown in Fig. 2 fetches a batch sample " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_b" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " sequential scans and applies a transformation to the scans " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " with the corresponding poses to have all points in a common coordinate frame. We use an unsupervised ground segmentation approach to remove the ground " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". Next, we cluster the remaining points " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ", defining coarse" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5221" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 155 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 155 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 155 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 155 + ], + "type": "table", + "html": "
MethodScribbles0.1%1%10%50%100%
Scratch54.9629.3542.7753.9658.2759.03
PointContrast [67]54.5232.6344.6258.6859.9861.45
DepthContrast [75]55.9031.6648.0557.1160.9961.14
SegContrast [50]56.7032.7544.8356.3160.4561.02
TARL (Ours)57.2538.5951.4260.3461.4261.47
", + "image_path": "0016703c8be4c70bfe0a50b1d47b8bf1361de9fdd40e7da83ee4cad8e25dba0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": "segments " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " of the objects in the scene. We sample two random scans at different times " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_1}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_b" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": ", in which we apply random augmentations " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{t_1}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": ". We compute point-wise features " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{t_1}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " from the sampled point clouds with the backbone, and list the segments " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "S^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " with its corresponding point-wise features. We calculate a mean embedding " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{S}}^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " for each segment " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " and compute the target representation " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "\\overline{s}^{t_2}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " for each segment with a Transformer projection head. For the segments " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "S^{t_1}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": ", we project the point-wise features with another Transformer projector followed by a Transformer encoder as a predictor to get the point-wise features " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "s^{t_1}" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": ". Lastly, we compute the loss to predict the corresponding segment representation at " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " for each point in " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": ". We repeat the process changing " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " to symmetrically match points from scan " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": " to the corresponding segment at " + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 200, + 289, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 405, + 128, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 128, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 128, + 419 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 557 + ], + "type": "text", + "content": "Datasets. We use the SemanticKITTI dataset [4,22] for our pre-training, an autonomous driving LiDAR data benchmark with point-wise annotations. During pre-training, we use only the raw point clouds from the training sequences and the given poses to extract the temporal objects views. For fine-tuning, we use SemanticKITTI with the full scans annotations and the scribbles annotations [51]. Besides SemanticKITTI, we also use nuScenes [8,64] for fine-tuning, to evaluate our approach in different autonomous driving datasets collected with different LiDARs and sensor setups, reporting the results on the validation sets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "content": "Model architecture. For our evaluation, we use a MinkUNet [20] as the backbone, which voxelizes the point clouds and uses sparse convolutions to extract features from the point cloud. The backbone receives the point cloud coordinates and intensity as input, and the output dimension feature has dimension " + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "inline_equation", + "content": "D = 96" + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "content": ". We use a multi-head self-attention encoder for the projection head with one layer and eight attention heads. For the predictor, we use the same multi-head self-attention encoder architecture. Further details on the backbone used in our experiments are provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Pre-training. For pre-training, we use the AdamW optimizer [43] with a learning rate of " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "2 \\cdot 10^{-4}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " and decay of " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 332, + 70, + 518, + 170 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 194 + ], + "type": "text", + "content": "Table 1. Results (mIoU) of the models pre-trained on SemanticKITTI fine-tuned to semantic segmentation on SemanticKITTI with different percentage of labels and scribbles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 70, + 518, + 170 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 518, + 170 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 518, + 170 + ], + "type": "table", + "html": "
MethodMiniFull
Scratch26.9466.03
Supervised pre-training38.3967.35
PointContrast [67]31.9267.31
DepthContrast [75]27.8164.70
SegContrast [50]31.2767.70
TARL (Ours)39.3668.26
", + "image_path": "48b91ffa59d4b9cb5c0b26eefe54f3d61662287cd5720ea6a6aa59127355b895.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 331, + 218, + 518, + 303 + ], + "blocks": [ + { + "bbox": [ + 305, + 174, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 305, + 174, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 174, + 545, + 208 + ], + "type": "text", + "content": "Table 2. Results (mIoU) of the self-supervised and supervised pretrained models on SemanticKITTI fine-tuned to semantic segmentation on nuScenes mini and full training sets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 218, + 518, + 303 + ], + "lines": [ + { + "bbox": [ + 331, + 218, + 518, + 303 + ], + "spans": [ + { + "bbox": [ + 331, + 218, + 518, + 303 + ], + "type": "table", + "html": "
MethodKITTInuScenes
Frozen backbone4.105.62
PointContrast [67]25.5319.70
DepthContrast [75]13.879.65
SegContrast [50]27.4223.75
TARL (Ours)35.9026.08
", + "image_path": "7b14a9015fd9ff46fb3043636ff7cf4fdca64264f14eca06aada0a6b3b039ea9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "type": "text", + "content": "Table 3. Linear evaluation results (mIoU) on SemanticKITTI and nuScenes datasets for semantic segmentation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": "training for 200 epochs with batch size 8 using a single NVIDIA RTX A6000. We use " + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "inline_equation", + "content": "n = 12" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": " for the scans to be aggregated. We provide experiments with different numbers of scans in the supplementary material to support this choice. The voxel resolution is set to " + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "inline_equation", + "content": "0.05m" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": " for the input point clouds, from which we sample a maximum of 40,000 points per point cloud. During the segment pooling, we limit it to a total of " + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "inline_equation", + "content": "M = 50" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": " segments with a maximum of " + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "inline_equation", + "content": "P = 300" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": " points per segment to avoid memory overflow. We use " + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 340, + 545, + 509 + ], + "type": "text", + "content": " in Eq. (1) and momentum of 0.999 to update the momentum network according to the online network weights. For the baselines, we use their official repositories with their default parameters, also training them for 200 epochs and sampling 40,000 points per scan." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 715 + ], + "type": "text", + "content": "Fine-tuning. To evaluate our method and compare it with the baselines, we fine-tune the models to three downstream tasks: semantic segmentation, panoptic segmentation, and object detection. For semantic segmentation, we use the pipeline provided in SegContrast [50] with the AdamW optimizer [43] and a learning rate of 0.001. For panoptic segmentation, we use the baseline evaluated by Hong et al. [29], where a 3D backbone network is used together with a semantic and an instance heads, followed by a clustering post-processing to identify the instances. We use the pre-trained MinkUNet model as the 3D backbone. We use a learning rate of 0.2 with batch size 8, training for 50 epochs. For object detection, we use the same toolbox and hyperparameters used in DepthContrast [75] with the PartA2 detector [57]. For all the tasks, we evaluate the method with the same model trained from scratch (without pre-training) and with the model after pre-training with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5222" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 492, + 168 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 492, + 168 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 492, + 168 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 492, + 168 + ], + "type": "table", + "html": "
0.1%1%10%50%100%
MethodPQIoUPQIoUPQIoUPQIoUPQIoU
Scratch4.7611.1322.7230.8447.2053.5355.3261.9455.4059.75
PointContrast [67]5.8611.5127.3732.4947.5754.6354.2159.4855.8561.49
DepthContrast [75]7.6513.5627.3132.3046.8551.2754.5559.6056.1560.81
SegContrast [50]7.5814.4626.1432.8547.0253.4755.3860.0456.7361.96
TARL (Ours)10.2617.0129.2434.7151.2757.5956.1062.3656.5762.05
", + "image_path": "601e889d300fec87ff89183f76f5acb61ebe8fcbe6d57662b6f844e16e1ff9f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 204, + 274, + 319 + ], + "blocks": [ + { + "bbox": [ + 46, + 174, + 546, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 546, + 196 + ], + "type": "text", + "content": "Table 4. Results (PQ and IoU) when fine-tuning the pre-trained models to panoptic segmentation with different percentage of labels on SemanticKITTI." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 204, + 274, + 319 + ], + "lines": [ + { + "bbox": [ + 57, + 204, + 274, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 204, + 274, + 319 + ], + "type": "table", + "html": "
MiniFull
PQIoUPQIoU
Scratch23.7823.9652.9858.17
Supervised pre-training24.7723.6053.1958.05
PointContrast [67]26.5825.4651.0656.39
DepthContrast [75]28.6627.3051.5157.06
SegContrast [50]28.8426.7952.3157.24
TARL (Ours)32.2230.7353.2659.14
", + "image_path": "b355f88648c8eee1f6c7168677f48770d8afc826628fd9d2e00d59e33d1aa465.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 357 + ], + "type": "text", + "content": "Table 5. Results (PQ and IoU) of the self-supervised and supervised pre-trained models on SemanticKITTI fine-tuned to panoramic segmentation on nuScenes mini and full training sets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 363, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 399 + ], + "type": "text", + "content": "the different self-supervised methods. The supplementary material provides detailed information for each downstream task training procedure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 178, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 178, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 178, + 420 + ], + "type": "text", + "content": "4.1. Semantic segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "content": "In this evaluation, we want to measure the information about semantics learned by the network during pretraining. We fine-tune it to semantic segmentation on SemanticKITTI and nuScenes after pre-training on SemanticKITTI. For SemanticKITTI, we use the percentage scans subsets used in SegContrast [50], with " + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 569 + ], + "type": "text", + "content": " of the labeled scans. We also evaluate on the scribbles annotations [51], which labels just a subset of points for each scan in the SemanticKITTI training sequences. We report the mean intersection-over-union (mIoU) of the models on the validation set, while fine-tuned on the aforementioned subset of labels from the training set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Tab. 1 shows the results from our approach compared to the baselines. As can be seen, overall the pre-training methods boost the performance compared to the network without pre-training on both scribbles annotations and the subsets of labeled scans. However, our approach surpasses previous state-of-the-art methods in all the different subsets, with a bigger gap when fewer labeled scans are used to fine-tune the model. We can also notice that, after pre-training with our method, only " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " of labeled scans were necessary for the model to surpass the performance of the network trained from scratch with " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " of the labeled scans. This result suggests that our method can effectively reduce the amount" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 198, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 233 + ], + "type": "text", + "content": "of labeled data necessary for the semantic segmentation task by around ten times. In the supplementary material, we also provide the IoU for each individual semantic class." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 233, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 449 + ], + "type": "text", + "content": "Generalization. In this experiment, we evaluate the generalization of the learned features. We use the network pretrained on SemanticKITTI and fine-tune it on nuScenes. We use nuScenes full training set and the mini training subset to fine-tune the network, and evaluate on the full validation set. In this evaluation, we also compare the results of the fully supervised semantic segmentation pre-training on SemanticKITTI fine-tuned on nuScenes. As shown in Tab. 2, even though all methods can improve the performance of the model trained on nuScenes, our approach can surpass previous self-supervised methods by around " + }, + { + "bbox": [ + 304, + 233, + 546, + 449 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 233, + 546, + 449 + ], + "type": "text", + "content": " when training with fewer labels. Besides, compared with supervised pre-training of the network on SemanticKITTI, our approach achieves better performance on both full and mini training sets. These results suggest that our method can replace supervised pre-training for LiDAR data, since it achieved better performance when fine-tuning to a different dataset, collected with a different sensor setup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "type": "text", + "content": "Linear evaluation. In this experiment, the pre-trained backbone is frozen, training only a linear head on top of it to evaluate how descriptive the self-supervised learned representation is without fine-tuning it to the target task. We perform this evaluation on SemanticKITTI and nuScenes datasets. In Tab. 3, the randomly initialized network (frozen backbone) achieves low performance, suggesting that the features extracted by the backbone play the main role in achieving semantic segmentation. All the pre-training methods improve the performance compared to the network without pre-training. However, our approach has a clear performance gap compared to the baselines, especially when training to the same dataset used for pre-training. These results suggest that the representation learned by our method can embed more semantic information, even though not using labels, achieving higher performance than previous state-of-the-art methods on both datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 659, + 435, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 435, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 435, + 672 + ], + "type": "text", + "content": "4.2. Panoptic segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "In this experiment, we fine-tune the pre-trained models for panoptic segmentation to also evaluate the instance-level features of our learned representation. Same as for semantic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5223" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 500, + 170 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 500, + 170 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 500, + 170 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 500, + 170 + ], + "type": "table", + "html": "
CarPedestrianCyclist
EasyModerateHardEasyModerateHardEasyModerateHard
Scratch91.7582.1380.1667.6563.1357.7391.7474.9470.34
PointContrast [67]91.5581.7079.8867.3362.3256.5592.3375.2070.66
DepthContrast [75]91.9882.0480.0369.1363.9757.7192.6873.6770.10
SegContrast [50]92.0882.1080.1868.7263.7757.3691.2674.7070.28
TARL (Ours)92.0982.1580.1068.5064.0058.7492.6175.4470.97
", + "image_path": "edf95e05107f63f162dab780ebcbb6a8085aefa9343afb106795e64e9080b457.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": "Table 6. Results (AP50) when fine-tuning the pre-trained models to object detection on KITTI full training set for car, pedestrian and cyclist classes on the easy, moderate and hard difficult levels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 201, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 251 + ], + "type": "text", + "content": "segmentation, we fine-tuned the model to SemanticKITTI using the same percentage subsets and nuScenes using the full and the mini training sets, reporting both the mIoU and the panoptic quality (PQ) on the validation sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 310 + ], + "type": "text", + "content": "Tab. 4 shows that our method is consistently better than previous self-supervised pre-training approaches. As the amount of training samples increases, the differences diminish. However, our method has a clear performance gap compared to the baselines when trained with fewer labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 312, + 288, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 455 + ], + "type": "text", + "content": "Generalization. To evaluate the generalization of our learned features on panoptic segmentation, we use the network pre-trained on SemantiKITTI to fine-tune it on nuScenes full and mini training sets, evaluating it on the full validation set. Tab. 5 shows that the representation learned by our method was more suited to be transferred to a different dataset. As in semantic segmentation, our method achieved better performance than previous self-supervised approaches and supervised pre-training on SemanticKITTI. These results agree with the semantic segmentation evaluation, suggesting that our method is better suited for transfer learning than the supervised pre-training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 529 + ], + "type": "text", + "content": "With these experiments, we can validate that our learned representations can extract not only semantic knowledge from the data but also instance-level information, surpassing previous methods in IoU and PQ metrics. In the supplementary material, we provide further results, reporting also the segmentation and recognition quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 537, + 146, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 146, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 146, + 551 + ], + "type": "text", + "content": "4.3. Object detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "As a third downstream task, we compare the methods fine-tuned for object detection to evaluate how general is the representation learned by our method compared with previous state-of-the-art methods. In this evaluation, we fine-tune the model with the whole training set of the KITTI dataset [22] and report the average precision " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{50})" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": " with 40 recall positions for car, pedestrian, and cyclist classes on the three difficulty levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "In Tab. 6, we can compare the baselines performance with our approach. For the car class, the pre-training methods achieve a marginal improvement compared to the network trained from scratch. For pedestrian and cyclist, the improvements brought by the pre-training are more substan" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 201, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 274 + ], + "type": "text", + "content": "tial. On those classes, our method achieves the best performance on moderate and hard difficulties, showing that our approach can also boost the performance when fine-tuning the model on object detection. In our supplementary material, we provide further experiments with the different training subsets used by Zhang et al. [75]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 357 + ], + "type": "text", + "content": "Together with the evaluations of the other downstream tasks, these results suggest that our method can learn a general representation. Our method achieved better performance than previous approaches on all three downstream tasks, showing that our approach learns a robust representation, not biased to a specific task but showing generalizable representations suitable for different tasks and datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 365, + 378, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 378, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 378, + 377 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 635 + ], + "type": "text", + "content": "In this paper, we propose a self-supervised representation learning method for 3D LiDAR data evaluated in the context of autonomous driving. We exploit the vehicle motion to extract different views of objects across time to learn a temporally-consistent representation. We use a Transformer encoder as a projection head and implicitly clustered points from the same object together while discriminating between different objects. Our experiments show that our method achieves better performance than previous state-of-the-art approaches on different downstream tasks. Besides, our approach could reduce the amount of necessary labeled data to only " + }, + { + "bbox": [ + 304, + 384, + 545, + 635 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 635 + ], + "type": "text", + "content": " when fine-tuning for semantic segmentation on SemanticKITTI. These results show that our proposed approach could learn a general LiDAR point cloud representation, embedding semantic information by matching objects representation viewed at different times. In addition, our method achieved better performance than supervised pre-training when transferring the learned representation to a dataset collected with different LiDAR sensors, suggesting that our approach could replace supervised pretraining in the LiDAR data domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "Limitations. Despite encouraging results, there is space for further improvement. For the unsupervised segment extraction, the clustering parameters have to be tuned depending on the LiDAR used to collect the data. Also, we rely on high-frequency data to properly segment dynamic objects. Otherwise, temporal segments could be wrongly matched." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5224" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 167 + ], + "type": "text", + "content": "[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. Proc. of the Int. Conf. on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "text", + "content": "[2] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "type": "text", + "content": "[3] Mehmet Aygun, Aljosa Osep, Mark Weber, Maxim Maximov, Cyril Stachniss, Jens Behley, and Laura Leal-Taixe. 4D Panoptic LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 279, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 279, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 279, + 288, + 334 + ], + "type": "text", + "content": "[4] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[5] Jens Behley and Cyril Stachniss. Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments. In Proc. of Robotics: Science and Systems (RSS), 2018. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[6] Benedikt Mersch, Xieyuanli Chen, Ignacio Vizzo, Lucas Nunes, Jens Behley, and Cyril Stachniss. Receding Moving Object Segmentation in 3D LiDAR Data Using Sparse 4D Convolutions. IEEE Robotics and Automation Letters (RA-L), 7(3):7503-7510, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[7] Borna Bešić, Nikhil Gosala, Daniele Cattaneo, and Abhinav Valada. Unsupervised Domain Adaptation for LiDAR Panoptic Segmentation. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 534 + ], + "type": "text", + "content": "[8] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A Multimodal Dataset for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 535, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 288, + 579 + ], + "type": "text", + "content": "[9] Ricardo J. G. B. Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Advances in Knowledge Discovery and Data Mining, 2013. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "text", + "content": "[10] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments. Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[12] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "PTTR: Relational 3D Point Cloud Object Tracking With Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 106, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 150 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In Proc. of the Int. Conf. on Machine Learning (ICML), 2020. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "type": "text", + "content": "[14] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive Learning. arXiv preprint, arxiv:2003.04297, 2020. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[15] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "text", + "content": "[16] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jirgen Gall, Jens Behley, and Cyril Stachniss. Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data. IEEE Robotics and Automation Letters (RA-L), 6:6529-6536, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "text", + "content": "[17] Xieyuanli Chen, Andres Milioto, Emanuele Palazzolo, Philippe Giguere, Jens Behley, and Cyril Stachniss. SuMa++: Efficient LiDAR-based Semantic SLAM. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2019. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[18] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked Feature Prediction for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 437 + ], + "type": "text", + "content": "[19] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. PatchFormer: An Efficient Point Transformer With Patch Attention. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[20] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[21] Fayao Liu, Guosheng Lin, and Chuan-Sheng Foo. Point discriminative learning for unsupervised representation learning on 3d point clouds. arXiv preprint, arXiv:2108.02104, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 580 + ], + "type": "text", + "content": "[22] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. 1, 2, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 581, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 634 + ], + "type": "text", + "content": "[23] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 668 + ], + "type": "text", + "content": "[24] Benjamin Graham and Laurens van der Maaten. Submanifold Sparse Convolutional Networks. arXiv preprint, arxiv:1706.01307, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 714 + ], + "type": "text", + "content": "[25] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, koray kavukcuoglu, Remi Munos, and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5225" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Michal Valko. Bootstrap Your Own Latent - A New Approach to Self-Supervised Learning. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[26] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J. Kusner. Unsupervised Point Cloud Pre-Training via Occlusion Completion. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 215 + ], + "type": "text", + "content": "[27] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum Contrast for Unsupervised Visual Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "text", + "content": "[28] Hiroharu Kato and Tatsuya Harada. Image Reconstruction from Bag-of-Visual-Words. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "text", + "content": "[29] Fangzhou Hong, Hui Zhou, Xinge Zhu, Hongsheng Li, and Ziwei Liu. LiDAR-Based Panoptic Segmentation via Dynamic Shifting Network. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "text", + "content": "[30] Hongxu Yin, Arash Vahdat, Jose Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-ViT: Adaptive Tokens for Efficient Vision Transformer. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "text", + "content": "[31] Ji Hou, Benjamin Graham, Matthias Niessner, and Saining Xie. Exploring Data-Efficient 3D Scene Understanding With Contrastive Scene Contexts. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 469 + ], + "type": "text", + "content": "[32] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP - Simple, Accurate, and Robust Registration If Done the Right Way. arXiv preprint arXiv:2209.15397, 2022. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[33] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 558 + ], + "type": "text", + "content": "[34] Ishan Misra, Rohit Girdharm, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[35] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. ST3D: Self-Training for Unsupervised Domain Adaptation on 3D Object Detection. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[36] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposalcontrast: Unsupervised pre-training for lidar-based 3d object detection. Proc. of the Europ. Conf. on Computer Vision (ECCV),2022.1,2,3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[37] Kai Fischer, Martin Simon, Florian Oelsner, Stefan Milz, Horst-Michael Gross, and Patrick Maeder. StickyPillars: Robust and Efficient Feature Matching on Point Clouds Using Graph Neural Networks. In Proc. of the IEEE/CVF" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[38] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. PointPillars: Fast Encoders for Object Detection From Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[39] Ferdinand Langer, Andres Milioto, Alexandre Haag, Jens Behley, and Cyril Stachniss. Domain Transfer for Semantic Segmentation of LiDAR Data using Deep Neural Networks. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[40] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph Attention Convolution for Point Cloud Semantic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[41] Hyungtae Lim, Sungwon Hwang, and Hyun Myung. ERASOR: Egocentric Ratio of Pseudo Occupancy-Based Dynamic Object Removal for Static 3D Point Cloud Map Building. IEEE Robotics and Automation Letters (RA-L), 6(2):2272-2279, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[42] Ling Zhang and Zhigang Zhu. Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks. In Proc. of the Intl. Conf. on 3D Vision (3DV), 2019. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "text", + "content": "[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 389, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 545, + 443 + ], + "type": "text", + "content": "[44] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Contrastive Instance Association for 4D Panoptic Segmentation using Sequences of 3D LiDAR Scans. IEEE Robotics and Automation Letters (RA-L), 7(2):1550-1557, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "type": "text", + "content": "[45] Andres Milioto, Jens Behley, Chris McCool, and Cyril Stachniss. LiDAR Panoptic Segmentation for Autonomous Driving. In Proc. of the IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "text", + "content": "[46] Andres Milioto, Leonard Mandtler, and Cyril Stachniss. Fast Instance and Semantic Segmentation Exploiting Local Connectivity, Metric Learning, and One-Shot Detection for Robotics. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "text", + "content": "[47] Andres Milioto and Cyril Stachniss. Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics using CNNs. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 591, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 646 + ], + "type": "text", + "content": "[48] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-BERT: Vision-Language Pre-Training on Fashion Domain. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[49] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. CrossPoint: Self-Supervised Cross-Modal Contrastive Learning for 3D Point Cloud Understanding. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5226" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[50] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. SegContrast: 3D Point Cloud Feature Representation Learning through Self-supervised Segment Discrimination. IEEE Robotics and Automation Letters (RA-L), 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[51] Ozan Unal, Dengxin Dai, and Luc Van Gool. Scribblesupervised lidar semantic segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[52] Pedro O. Pinheiro, Amjad Almahairi, Ryan Y. Benmalek, Florian Golemo, and Aaron Courville. Unsupervised learning of dense visual representations. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), pages 4489-4500, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "text", + "content": "[53] Pierre Dellenbach, Jean-Emmanuel Deschaud, Bastien Jacquet, and François Goulette. CT-ICP Real-Time Elastic LiDAR Odometry with Loop Closure. In Proc. of the IEEE Intl. Conf. on Robotics & Automation (ICRA), 2022. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[54] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "type": "text", + "content": "[55] Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point-Net++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Proc. of the Conference on Neural Information Processing Systems (NeurIPS), 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[56] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Proc. of the Europ. Conf. on Computer Vision (ECCV), pages 626-642, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "text", + "content": "[57] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From Points to Parts: 3D Object Detection from Point Cloud with Part-aware and Part-aggregation Network. IEEE Trans. on Pattern Analalysis and Machine Intelligence (TPAMI), 43:2647-2664, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[58] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. PointR-CNN: 3D Object Proposal Generation and Detection From Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 566 + ], + "type": "text", + "content": "[59] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-Level Representation Learning for Self-Supervised Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[60] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching Efficient 3D Architectures with Sparse Point-Voxel Convolution. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[61] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. KPCov: Flexible and Deformable Convolution for Point Clouds. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[62] Weijing Shi and Ragunathan (Raj) Rajkumar. Point-GNN: Graph Neural Network for 3D Object Detection in a Point Cloud. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[63] Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring Cross-Image Pixel Contrast for Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[64] Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. IEEE Robotics and Automation Letters (RA-L), 7(2):3795-3802, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[65] Xiangyun Zhao, Raviteja Vemulapalli, Philip Mansfield, Boqing Gong, Bradley Green, Lior Shapira, and Ying Wu. Contrastive Learning for Label Efficient Semantic Segmentation. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[66] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling Vision Transformers. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 336 + ], + "type": "text", + "content": "[67] Saining Xie, Jiatao Gu, Demi Guo, Charles R. Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised Pretraining for 3D Point Cloud Understanding. In Proc. of the Europ. Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[68] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "text", + "content": "[69] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense Contrastive Learning for Self-Supervised Visual Pre-Training. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[70] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-Training 3D Point Cloud Transformers With Masked Point Modeling. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 493, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 546 + ], + "type": "text", + "content": "[71] Li Yi, Boqing Gong, and Thomas Funkhouser. Complete & Label: A Domain Adaptation Approach to Semantic Segmentation of LiDAR Point Clouds. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[72] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. MPViT: Multi-Path Vision Transformer for Dense Prediction. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 645 + ], + "type": "text", + "content": "[73] Xiaoding Yuan, Adam Kortylewski, Yihong Sun, and Alan Yuille. Robust Instance Segmentation Through Reasoning About Multi-Object Occlusion. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[74] Yujin Chen, Matthias Nießner, and Angela Dai. 4dcontrast: Contrastive learning with dynamic correspondences for 3d scene understanding. Proc. of the Europ. Conf. on Computer Vision (ECCV), 2022. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[75] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-Supervised Pretraining of 3D Features on any" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5227" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 363 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "Point-Cloud. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[76] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In Proc. of the Int. Conf. on Machine Learning (ICML), 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[77] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proc. of the IEEE/CVF Intl. Conf. on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 261 + ], + "type": "text", + "content": "[78] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring Set Similarity for Dense Self-Supervised Representation Learning. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 307 + ], + "type": "text", + "content": "[79] Zixiang Zhou, Yang Zhang, and Hassan Foroosh. Panoptic-PolarNet: Proposal-Free LiDAR Point Cloud Panoptic Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 363 + ], + "type": "text", + "content": "[80] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation. In Proc. of the IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5228" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_content_list.json b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d24cd0dead4fb51067c2d167992e567fd670bbed --- /dev/null +++ b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_content_list.json @@ -0,0 +1,1592 @@ +[ + { + "type": "text", + "text": "Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suntheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, Nahiup Kang Samsung Advanced Institute of Technology (SAIT)", + "bbox": [ + 102, + 180, + 864, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sh2019.park, minjungs.son, swan.jang, ychun.ahn, jiyeon31.kim, nahyup.kang}@samsung.com", + "bbox": [ + 93, + 218, + 870, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c9bfc4c36b86ff5abc585719ac7ba8dd5aec02a845824cc465d8911ab948ee0.jpg", + "image_caption": [ + "Figure 1. We propose simple yet effective feature interpolation methods for training neural radiance fields of dynamic scenes based on temporal interpolation. We provide two different feature vector representations, neural representation (top) and grid representation (bottom), both of which are the concatenation of static feature vectors and temporally-interpolated dynamic feature vectors. The neural representation exhibits high-quality rendering performance with small-sized models while the grid representation shows competitive rendering results with astonishingly fast training speed." + ], + "image_footnote": [], + "bbox": [ + 117, + 284, + 861, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 626, + 313, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal interpolation often plays a crucial role to learn meaningful representations in dynamic scenes. In this paper, we propose a novel method to train spatiotemporal neural radiance fields of dynamic scenes based on temporal interpolation of feature vectors. Two feature interpolation methods are suggested depending on underlying representations, neural networks or grids. In the neural representation, we extract features from space-time inputs via multiple neural network modules and interpolate them based on time frames. The proposed multi-level feature interpolation network effectively captures features of both short-term and long-term time ranges. In the grid representation, space-time features are learned via four-dimensional hash grids, which remarkably reduces training time. The grid representation shows more than $100 \\times$ faster training speed than the previous neural-net-based methods while maintaining", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the rendering quality. Concatenating static and dynamic features and adding a simple smoothness term further improve the performance of our proposed models. Despite the simplicity of the model architectures, our method achieved state-of-the-art performance both in rendering quality for the neural representation and in training speed for the grid representation.", + "bbox": [ + 496, + 628, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 767, + 632, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D reconstruction and photo-realistic rendering have been long-lasting problems in the fields of computer vision and graphics. Along with the advancements of deep learning, differentiable rendering [11, 14] or neural rendering, has emerged to bridge the gap between the two problems. Recently proposed Neural Radiance Field (NeRF) [18] has finally unleashed the era of neural rendering. Using NeRF,", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4212", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "it is able to reconstruct accurate 3D structures from multiple 2D images and to synthesize photo-realistic images from unseen viewpoints. Tiny neural networks are sufficient to save and retrieve complex 3D scenes, which can be trained in a self-supervised manner given 2D images and camera parameters.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, as our world typically involves dynamic changes, it is crucial to reconstruct the captured scene through 4D spatiotemporal space. Since it is often not possible to capture scenes at different viewpoints simultaneously, reconstructing dynamic scenes from images is inherently an under-constrained problem. While NeRF was originally designed to deal with only static scenes, there have been a few approaches in the literature that extend NeRF to dynamic scenes [13,22,23,25], which are so called as dynamic NeRFs. Inspired by the non-rigid structure-from-motion algorithms [1, 4] that reconstruct 3D structures of deformable objects, most previous works solved the under-constrained setting by estimating scene deformations [21,22,25] or 3D scene flows [5,9,12] for each frame.", + "bbox": [ + 75, + 184, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, since the parameters of deformation estimation modules are jointly optimized with NeRF network simultaneously, it is questionable that the modules can accurately estimate deformations or scene flows in accordance with its design principles. In many dynamic scenes, it is challenging to resolve the ambiguities whether a point was newly appeared, moved, or changed its color. It is expected that those ambiguities and deformation estimation can be separately solved within a single network, but in practice, it is hard to let the network implicitly learn the separation, especially for general dynamic scenes without any prior knowledge of deformation.", + "bbox": [ + 75, + 397, + 468, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, grid representations in NeRF training [7, 19, 24] have grabbed a lot of attentions mainly due to its fast training speed. Simple trilinear interpolation is enough to fill in the 3D space between grid points. While the representation can be directly adopted to dynamic NeRFs together with the warping estimation module [6], it still requires additional neural networks that affects training and rendering speed.", + "bbox": [ + 75, + 580, + 468, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the aforementioned analyses, we present a simple yet effective architecture for training dynamic NeRFs. The key idea in this paper is to apply feature interpolation to the temporal domain instead of using warping functions or 3D scene flows. While the feature interpolation in 2D or 3D space has been thoroughly studied, to the best of our knowledge, feature interpolation method in temporal domain for dynamic NeRF has not been proposed yet. We propose two multi-level feature interpolation methods depending on feature representation which is either neural nets or hash grids [19]. Overview of the two representations, namely the neural representation and the grid representation, are illustrated in Fig. 1. In addition,", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "noting that 3D shapes deform smoothly over time in dynamic scenes, we additionally introduced a simple smoothness term that encourages feature similarity between the adjacent frames. We let the neural networks or the feature grids to learn meaningful representations implicitly without imposing any constraints or restrictions but the smoothness regularizer, which grants the flexibility to deal with various types of deformations. Extensive experiments on both synthetic and real-world datasets validate the effectiveness of the proposed method. We summarized the main contributions of the proposed method as follows:", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a simple yet effective feature extraction network that interpolates two feature vectors along the temporal axis. The proposed interpolation scheme outperforms existing methods without having a deformation or flow estimation module.", + "- We integrate temporal interpolation into hash-grid representation [19], which remarkably accelerates training speed more than $100 \\times$ faster compared to the neural network models.", + "- We propose a smoothness regularizer which effectively improves the overall performance of dynamic NeRFs." + ], + "bbox": [ + 517, + 268, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 467, + 640, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic NeRF. There have been various attempts to extend NeRF to dynamic scenes. Existing methods can be categorized into three types: warping-based, flow-based, and direct inference from space-time inputs.", + "bbox": [ + 496, + 493, + 890, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Warping-based methods learn how 3D structure of the scene is deformed. 3D radiance field for each frame is warped to single or multiple canonical frames using the estimated deformation. Deformation is parameterized as 3D translation [23], rotation and translation via angle-axis representation [21, 22], or weighted translation [25].", + "bbox": [ + 496, + 553, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, flow-based methods estimate the correspondence of 3d points in consecutive frames. Neural scene flow fields [13] estimates 3D scene flow between the radiance fields of two time stamps. Xian et al. [27] suggested irradiance fields which spans over 4D space-time fields with a few constraints derived from prior knowledge.", + "bbox": [ + 496, + 643, + 890, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although warping-based and flow-based approaches showed successful results, they have a few shortcomings. For instance, warping-based methods cannot deal with topological variations. Since those methods warp every input frame to a single canonical scene, it is hard to correctly represent newly appeared or disappeared 3D radiance fields in the canonical scene. HyperNeRF [22] learns hyperspace which represents multidimensional canonical shape bases. Neural scene flow fields [13] solve the problem by introducing occlusion masks which are used to estimate the regions where scene flows are not applicable. However, the masks", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4213", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and the flow fields should be learned simultaneously during the training, which makes the training procedure complicated and imposes dependence on extra information such as monocular depth or optical flow estimation.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without estimating shape deformations or scene flows, DyNeRF [12] used a simple neural network to train NeRFs of dynamic scenes, although the experiments are only conducted on synchronized multi-view videos. It utilizes 3D position and time frame as inputs, and predicts color and density through a series of fully connected neural networks. Our neural representation applies temporal interpolation to intermediate features which enhances representation power for dynamic features while keeping the simple structure.", + "bbox": [ + 75, + 151, + 470, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Category-specific NeRFs are one of the popular research directions for training dynamic objects such as humans [8, 26] or animals [28]. They are based on parametric models or use additional inputs such as pose information. Our method focuses on training NeRFs of general dynamic scenes without any prior knowledge about scenes or objects. Nevertheless, temporal interpolation proposed in this paper can be easily applied to the dynamic NeRFs with templates. Grid representations in NeRF. One of the major drawbacks of the original NeRF [18] is slow training speed. A few concurrent works are proposed to speed up NeRF training. Among them, grid-based representation such as Plenoxels [7], direct voxel grid optimization [24] show superior performance in terms of training time, where it takes only a few minutes to learn plausible radiance fields. Combining grid representation with a hash function further improves the efficiency and training speed of the feature grids [19]. Fang et al. [6] firstly applied a voxel grid representation to train dynamic NeRFs and achieved much faster training speed compared to the neural-network-based methods. Guo et al. [10] also proposed a deformable voxel grid method for fast training. Unlike aforementioned works, our grid representation does not estimate warping or deformation. Instead, it directly estimates color and density using the features obtained from 4D hash grids, which decreases the computational burden and enables faster training.", + "bbox": [ + 75, + 287, + 470, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Temporal Interpolation for Dynamic NeRF", + "text_level": 1, + "bbox": [ + 76, + 691, + 460, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 717, + 217, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given 3D position $\\mathbf{x} \\in \\mathbb{R}^3$ and 2D viewing direction $\\mathbf{d} \\in \\mathbb{R}^2$ , NeRF [18] aims to estimate volume density $\\sigma \\in \\mathbb{R}$ and emitted RGB color $\\mathbf{c} \\in \\mathbb{R}^3$ using neural networks. We formulate dynamic NeRF as a direct extension from 3D to 4D by adding time frame $t \\in \\mathbb{R}$ to inputs, i.e.", + "bbox": [ + 75, + 739, + 467, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(\\mathbf {c}, \\sigma) = f (\\mathbf {v} (\\mathbf {x}, t), \\mathbf {d}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 828, + 468, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$f$ is implemented as fully connected neural networks and a space-time feature representation $\\mathbf{v}$ can be MLP-based neural nets or explicit grid values as explained in Sec. 3.2.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the camera with origin $\\mathbf{o}$ and ray direction $\\mathbf{d}$ , the color of camera ray $\\mathbf{r}(u) = \\mathbf{o} + u\\mathbf{d}$ at time frame $t$ is", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} (\\mathbf {r}, t) = \\int_ {u _ {n}} ^ {u _ {f}} U (u, t) \\sigma (\\mathbf {v}) \\mathbf {c} (\\mathbf {v}, \\mathbf {d}) d u, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 130, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u_{n}$ and $u_{f}$ denote the bounds of the scene volume, $\\mathbf{v}$ is an abbreviation of $\\mathbf{v}(\\mathbf{r}(u), t)$ and $U(u, t) = \\exp(-\\int_{u_n}^u \\sigma(\\mathbf{v}(\\mathbf{r}(s), t)) ds)$ is the accumulated transmittance from $u_{n}$ to $u$ . Then, the RGB color loss $L_{c}$ is defined to minimize the $l_{2}$ -loss between the estimated colors $\\hat{C}(\\mathbf{r}, t)$ and the ground truth colors $C(\\mathbf{r}, t)$ over all rays $\\mathbf{r}$ in camera views $R$ and time frames $t \\in T$ as follows:", + "bbox": [ + 496, + 172, + 890, + 277 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c} = \\sum_ {\\mathbf {r} \\in R, t \\in T} | | \\hat {C} (\\mathbf {r}, t) - C (\\mathbf {r}, t) | | _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 286, + 890, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Space-Time Feature Representation", + "text_level": 1, + "bbox": [ + 500, + 329, + 810, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contribution of our paper lies on the novel feature representation method for dynamic NeRFs. We define the feature vector $\\mathbf{v}$ , which is fed into the neural network to determine a color and a volume density, as the concatenation of static and dynamic feature vectors, i.e.,", + "bbox": [ + 496, + 353, + 890, + 429 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} (\\mathbf {x}, t) = \\left[ \\mathbf {v} _ {s} (\\mathbf {x}), \\mathbf {v} _ {d} (\\mathbf {x}, t) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 439, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\cdot, \\cdot]$ means concatenation operator of two vectors. The static feature vector $\\mathbf{v}_s$ only depends on the 3D position $\\mathbf{x}$ . As most dynamic scenes also contain static regions such as background, $\\mathbf{v}_s$ is designed to learn the static features that are consistent across time. Meanwhile, $\\mathbf{v}_d$ learns dynamic features which may vary across time. We propose two novel feature representation methods in which temporal interpolation is applied. First, the neural representation, which is essentially a series of neural networks combined with linear interpolation, is suggested in Sec. 3.2.1. The model trained using the neural representation is able to render high-quality space-time view synthesis with small-sized neural networks ( $\\sim 20\\mathrm{MB}$ ). Second, the grid representation, which is an temporal extension of recently proposed voxel grid representations [19], is explained in Sec. 3.2.2. Dynamic NeRFs can be trained in less than 5 minutes with the proposed grid representation.", + "bbox": [ + 496, + 465, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Neural Representation", + "text_level": 1, + "bbox": [ + 500, + 739, + 710, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the neural representation, features that fed into the template NeRFs are determined by a series of neural nets. In other words, both static and dynamic feature extractor are formulated as multi-layer perceptrons (MLPs). First, the whole time frame is divided into equally spaced time slots. For each time slot, two MLPs are assigned. Then the whole feature vector is interpolated from the assigned MLPs. Overall feature extraction procedure is illustrated in Fig. 2.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4214", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a88608fb7e439d02a916731ce76434e718f64fef4450211653bda669a61126a2.jpg", + "image_caption": [ + "Figure 2. Illustration of the arrangements of MLPs (left) and the feature extraction (right) in the neural representation. $\\phi_s$ and $\\phi_{ij}$ are small MLPs that extract static and dynamic features respectively. The static features $\\mathbf{v}_s$ and the dynamic feature $\\mathbf{v}_d$ are concatenated and fed into the template NeRF. Detailed structures of the template NeRF can be found in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 465, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/23ce09fdbe5b05a7ff892ae461ec6de8d67f636ebbf12472f387191cd5653cb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 859, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, let $\\mathbf{v}_d(\\mathbf{x},\\mathbf{z}_t)$ be the feature vector for a 3D point $\\mathbf{x}$ at time $t$ , which will be fed to the template NeRF. Here, we used $\\mathbf{z}_t$ as an embedding vector for an input time frame $t$ [12, 16, 22]. For equally spaced $n$ time slots, there are $n + 1$ keyframes $t_i = \\frac{i}{n} (i = 0,1,2,\\dots,n)$ . An MLP $\\phi_{i}$ is assigned to each keyframe $t_i$ which is responsible for two adjacent time slots $[t_{i - 1},t_i]$ and $[t_i,t_{i + 1}]$ . For inputs with time $t\\in [t_i,t_{i + 1}]$ , $\\mathbf{x}$ and $\\mathbf{z}_t$ are fed into $\\phi_{i}$ and $\\phi_{i + 1}$ . Then, the outputs from two MLPs are interpolated as follows:", + "bbox": [ + 75, + 318, + 472, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) = \\Delta t \\cdot \\phi_ {i} (\\mathbf {x}, \\mathbf {z} _ {t}) + (1 - \\Delta t) \\cdot \\phi_ {i + 1} (\\mathbf {x}, \\mathbf {z} _ {t}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 467, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta t = (t_{i + 1} - t) / (t_{i + 1} - t_i)$ .", + "bbox": [ + 76, + 496, + 316, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The purpose of this interpolation is to efficiently learn the features between keyframes in a scalable manner. Thanks to the temporal interpolation, we can learn the features of the continuous time range $[t_i, t_{i+1}]$ by enforcing the MLPs $\\phi_i$ and $\\phi_{i+1}$ responsible for that time range. While the static feature $\\mathbf{v}_s$ represents the features across the whole timeline, $\\mathbf{v}_d$ represents the features that are focused more on dynamic regions. In addition, since each MLP for dynamic feature is responsible only for two adjacent time slots, it is able to make each $\\phi_i$ learn features that are specific to a certain period of time.", + "bbox": [ + 75, + 512, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To exploit the features of both long-term and short-term, there can be multiple levels of dynamic features which have different number of keyframes. For multi-level dynamic feature extractor with level $l$ , each level contains different number of keyframes $n_1, n_2, \\dots, n_l$ . Let $\\mathbf{v}_{di}$ denote the dynamic feature of $i$ -th level, then the output dynamic feature is the concatenation of features from all levels, i.e.,", + "bbox": [ + 75, + 678, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) = \\left[ \\mathbf {v} _ {d 1} (\\mathbf {x}, \\mathbf {z} _ {t}), \\mathbf {v} _ {d 2} (\\mathbf {x}, \\mathbf {z} _ {t}), \\dots , \\mathbf {v} _ {d l} (\\mathbf {x}, \\mathbf {z} _ {t}) \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 796, + 468, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we used the settings $l = 2$ , $n_1 = 5$ , $n_2 = 20$ otherwise stated. The dimensions of feature vectors are determined as 128 for $\\mathbf{v}_s$ , and 64 for $\\mathbf{v}_0$ and $\\mathbf{v}_1$ . MLP with one hidden layer whose hidden size is the same as its output dimension is used for feature extraction of both $\\mathbf{v}_s$ and $\\mathbf{v}_{di}$ .", + "bbox": [ + 76, + 824, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The concatenated feature vectors are fed to the template NeRF which outputs volume density $\\sigma$ and emitted color c. We used the same structure as used in HyperNeRF [22] for the template NeRF of the neural representation.", + "bbox": [ + 496, + 318, + 893, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Grid Representation", + "text_level": 1, + "bbox": [ + 498, + 397, + 699, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, InstantNGP [19] suggested a novel multi-level grid representation with hash tables. We adopt the hash grid representation from [19] and extend it for fast dynamic NeRF training. Similar to the neural representation in Sec. 3.2.1, the feature vector from the proposed hash grid contains static and dynamic feature vectors. The static feature $\\mathbf{v}_s$ is analogous to the one used in [19]. On the other hand, the dynamic feature $\\mathbf{v}_d$ comes from the 4D hash grids.", + "bbox": [ + 496, + 420, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, to extract static and dynamic feature vectors whose dimensions are $m_{s}$ and $m_{d}$ respectively, a hash table of size $H$ that contains $(m_{s} + m_{d})$ dimension feature vectors is constructed. The hash function of $d$ -dimensional vector, $h_d(\\mathbf{x})$ , is defined as", + "bbox": [ + 496, + 542, + 890, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {d} (\\mathbf {x}) = \\bigoplus_ {i = 1} ^ {d} \\left(x _ {i} P _ {i}\\right) \\mod H, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 628, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{i}$ is a large prime number and $\\bigoplus$ is an XOR operator. Then, the feature vector $\\mathbf{v}$ is retrieved by concatenating the outputs of 3D and 4D hash functions:", + "bbox": [ + 496, + 679, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} (\\mathbf {x}, t) = \\left[ h _ {3} (\\mathbf {x}), h _ {4} (\\mathbf {x}, t) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 734, + 890, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The 3D and 4D grids are constructed as the multi-level hash grids proposed in [19]. We applied different scaling factors for 3D space and time frame since the number of frames for training sequences are usually much smaller than the finest 3D grid resolutions.", + "bbox": [ + 496, + 762, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Smoothness Regularization", + "text_level": 1, + "bbox": [ + 500, + 847, + 745, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As our dynamic world smoothly changes, it is reasonable to impose a smoothness term to adjacent time frames.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4215", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/309b7c6d5564d7c61da017fb3f3b8e61da15c0b3c49fff352e6d05e406c0202b.jpg", + "image_caption": [ + "Figure 3. Illustration of grid representation with hash tables at a certain level. A static feature vector of a 3D grid point is extracted via 3D hash function, while dynamic feature vector for 4D grid point is extracted via 4D hash. The static feature vector $\\mathbf{v}_s$ is determined as trilinear interpolation of 8 vectors of grid points, while the dynamic feature vector $\\mathbf{v}_d$ is calculated as quadrilinear interpolation of 16 vectors. In the figure, we only depicted feature retrieval process of two grid points for illustration purpose." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 851, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that the smoothness term is only applied to the input feature space, not to the estimated outputs such as RGB color or density. For the neural representation, we provided a simple smoothness term which is calculated as,", + "bbox": [ + 75, + 335, + 470, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} (\\mathbf {x}, t) = \\left\\| \\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) - \\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t + 1}) \\right\\| _ {2} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 406, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This regularization term has two advantages. First, it reflects the intuition that the observed point $\\mathbf{x}$ at time $t$ will be stationary if there is no observation for $\\mathbf{x}$ at time $t + 1$ . Second, by propagating additional gradients to the feature networks or the hash grids, the smoothness term acts as a regularizer that stabilizes the training.", + "bbox": [ + 75, + 434, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the grid representation, we impose a smoothness term to the grid points that are temporally adjacent:", + "bbox": [ + 75, + 523, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} (\\mathbf {x}, t) = \\frac {1}{n _ {f} ^ {2}} \\| h _ {4} (\\mathbf {x}, t _ {a}) - h _ {4} (\\mathbf {x}, t _ {b}) \\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 563, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_f$ is the number of frames in the training sequence and $t_a, t_b$ are two adjacent grid points that satisfies $t_a \\leq t \\leq t_b$ . In fact, Eq. (10) can be obtained from Eq. (9) in the grid representation, which indicates that imposing smoothness term to adjacent time frames is equivalent to add smoothness to two temporally adjacent grids with constant multiplier. Detailed derivation of this relationship is provided in the supplementary materials.", + "bbox": [ + 75, + 606, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The smoothness term is added to the loss function with a weight $\\lambda$ , and the total loss is minimized together during the training. Accordingly, the total loss of our dynamic neural radiance fields is given by", + "bbox": [ + 75, + 727, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {c} + \\lambda L _ {s}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 799, + 468, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that the smoothness term is especially powerful in the static background regions that appeared only in a small fraction of frames. In Fig. 4, we show examples of the rendered images which are trained with and without the smoothness term. It can be clearly seen that the boxed", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9fa34a09239bbae0bd572ed78019774be582e9f92744321a79540e65d4114698.jpg", + "image_caption": [ + "(a) w/o smooth term" + ], + "image_footnote": [], + "bbox": [ + 506, + 333, + 624, + 494 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/235e6988d1d7dd2add057a5bdd31f54f7e0e970a615fb90457c02ce72f629703.jpg", + "image_caption": [ + "(b) w/ smooth term" + ], + "image_footnote": [], + "bbox": [ + 635, + 333, + 751, + 493 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd4b726b530f6b0778fc7b3df8b52e1ec9598fc3a0d2d6f008260fe58acb3528.jpg", + "image_caption": [ + "(c) GT", + "Figure 4. Effectiveness of the smoothness term. It can be observed that the model trained with the smoothness term accurately renders the corner part which appears in only few frames within the whole sequence (blue box) and removes spurious artifacts (red box)." + ], + "image_footnote": [], + "bbox": [ + 763, + 333, + 880, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "regions of the rendered image shows much plausible when the model is trained with the smoothness term.", + "bbox": [ + 498, + 595, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 635, + 718, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the neural representation, we adopted the template NeRF architecture of HyperNeRF [22] for density and color estimation. The network is an 8-layered MLP with hidden size of 256. The dimension of the embedding vector $\\mathbf{z}_t$ is set to 8. The smoothness weight $\\lambda$ is set to 0.01 or 0.001 depending on the characteristics of the datasets used in the experiments. We set $\\lambda$ to a large value in the sequences that the viewpoint change is significant in a short period of time although performance variations depending on the value of $\\lambda$ is not significant.", + "bbox": [ + 496, + 659, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the grid representation, $\\lambda$ is set to $1e^{-4}$ . We applied the smoothness loss only for the finest two levels of the temporal dimension since applying it to every level slows training speed without much improvements in performance. After feature vectors are extracted from the hash grids, we fed them to a 3-layered MLP with hidden size of 128 to estimate", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4216", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "volume density followed by one additional layer for RGB color estimation similar to [19]. We set $m_{s} = 2$ , $m_{d} = 6$ and $H = 2^{19}$ for the grid representation experiments otherwise stated. The 3D and 4D hash grids are composed of 12 levels. For the spatial dimension, we set the base resolution to 8, and the scaling factor is set to 1.45. The base resolution of the temporal dimension is 2, which is multiplied by 1.4 in every other level. Detailed hyperparameter settings and network architectures can be found in the supplementary materials.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 257, + 282, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 282, + 179, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the superiority of our method, we conducted extensive experiments on various datasets. We used three publicly available datasets that are used for dynamic NeRFs. For all experiments, we trained our models for each sequence individually, and then per-scene results are averaged and reported in this section.", + "bbox": [ + 75, + 305, + 468, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "D-NeRF Dataset [23]. The dataset consists of synthetically rendered images of moving and deforming 3D objects. For each frame, the image is rendered via a synthetic camera of random rotation. There are eight scenes in total, and each scene contains 50-200 training images with 20 test views.", + "bbox": [ + 75, + 396, + 468, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HyperNeRF Dataset [22]. The dataset contains video sequences taken from mobile phones. There are four sequences in vrig-dataset which are taken using the camera rig with two phones vertically aligned. In addition, there are six sequences in interp-dataset which are taken from a single camera in order to estimate the image in the middle of two consecutive training images.", + "bbox": [ + 75, + 472, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DyNeRF Dataset [12]. The dataset consists of videos obtained from a capture system that consists of 21 GoPro Black Hero cameras. The cameras are located at a fixed position, and all video frames are synchronized to build the multi-view video dataset.", + "bbox": [ + 75, + 579, + 468, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Neural Representation", + "text_level": 1, + "bbox": [ + 76, + 665, + 287, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we reported the performance of the proposed neural representation models. First, the experimental results on D-NeRF dataset are shown in Tab. 1. Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS) [29] are used as evaluation metrics following the previous works. We also reported the average metric (AVG) proposed in [3] which aggregates three metrics to a single value. Our method with the neural representation (Ours-NN) achieves state-of-the-art results on all evaluation metrics. It is worth noting that the smoothness term dramatically improves overall performance.", + "bbox": [ + 75, + 688, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we evaluated our method on HyperNeRF dataset [22] and compared with existing methods, which is shown in Tab. 2. Here, we reported PSNR and multi", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4b54537e35643441f1b2f3c89aeb296f635b373396e1108cb66e739a17615d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓AVG ↓
NeRF [18]19.000.870.180.09
T-NeRF [23]29.500.950.080.03
D-NeRF [23]30.430.950.070.02
NDVG-full [10]27.840.8620.0410.029
TiNeuVox-B [6]32.670.9710.0410.016
Ours-NN (w/o smooth)30.180.9630.0380.019
Ours-NN (w/ smooth)32.730.9740.0330.014
", + "bbox": [ + 506, + 87, + 883, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Experimental results on D-NeRF [23] datasets.", + "bbox": [ + 529, + 203, + 861, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scale SSIM (MS-SSIM), and excluded LPIPS metric since its value cannot reliably reproduced [6]. Our method shows second-best results on both vrig and interp datasets. While flow-based method [13] suffers from interpolating motions between two consecutive frames, our method, which implicitly learns intermediate feature representation of inbetween frames, achieves competitive performance with warping-based methods [21, 22].", + "bbox": [ + 496, + 239, + 890, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results for the neural representation are shown in Fig. 5. It is clearly observed that our method captures fine details compared to D-NeRF [23]. We highlighted the regions that show notable differences with colored boxes and placed zoomed-in images of the regions next to the rendered images. In addition, we also qualitatively compared our method with HyperNeRF [22] on their datasets. When the warping estimation module of [22] does not correctly estimate warping parameters, HyperNeRF produces implausible results. It can be observed that the head and the body of the chicken toy is not properly interlocked and the position of the hand peeling banana is incorrect. On the other hand, our method accurately recovers 3D geometry of those scenes. Thus, without using the physically meaningful warping estimation module in neural networks, the proposed temporal interpolation and the smoothness regularizer provides simple yet effective way to learn complex deformations.", + "bbox": [ + 496, + 362, + 890, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lastly, performance evaluation on DyNeRF dataset [12] is presented in Tab. 3. We adopt PSNR, LPIPS, and FLIP [2] for evaluation metrics to compare with previous works. Our method achieves best PSNR while ranked second in LPIPS and FLIP. However, those two metrics are also better than DyNeRF† [12] which does not use importance sampling strategy in [12]. Since we do not use any sampling strategy during training, it can be concluded that our feature interpolation method is superior to the network architecture of [12]. Notably, our method outperforms DyNeRF with smaller network size (20MB) than the DyNeRF models (28MB).", + "bbox": [ + 496, + 635, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Grid Representation", + "text_level": 1, + "bbox": [ + 500, + 830, + 692, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used D-NeRF datasets to evaluate the performance of the proposed grid representation. Since this representation is mainly intended for fast training, we report the re", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4217", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd18c36bc3f6ae14344bccf0c60f783edb6644c192d6101835ff1bb4fc154e96.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our method (Ours-NN) on of D-NeRF [23] and HyperNeRF [22] datasets. The regions that show significant difference with compared methods are scaled up (red and green boxes)." + ], + "image_footnote": [], + "bbox": [ + 86, + 85, + 893, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8874f3b1b5f04a5782eb8d839688a0b4ed6443af9c6bf25d750719eb44dd7c29.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
vriginterp
PSNR↑MS-SSIM↑PSNR↑MS-SSIM↑
NeRF [18]20.130.74522.270.804
NV [15]16.850.57126.050.911
NSFF [13]26.330.91625.800.883
Nerfies [21]22.230.80328.470.939
HyperNeRF [22]22.380.81429.000.945
Ours-NN24.350.86728.670.940
", + "bbox": [ + 84, + 340, + 460, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e67ae5c34aeb0004155ca1c7e88c916e15b155050243739d422b70500e6b840.jpg", + "table_caption": [ + "Table 2. Experimental results on HyperNeRF [22] datasets." + ], + "table_footnote": [], + "table_body": "
PSNR↑LPIPS ↓FLIP ↓
MVS19.120.25990.2542
NeuralVolumes [15]22.800.29510.2049
LLFF [17]23.240.23460.1867
NeRF-T [12]28.450.1000.1415
DyNeRF† [12]28.500.09850.1455
DyNeRF [12]29.580.08320.1347
Ours-NN29.880.09600.1413
", + "bbox": [ + 106, + 481, + 439, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults in a short period time ( $\\sim 8$ minutes) and compare the results with the concurrent works [6, 10], both of which are based on voxel grid representation and showed the fastest training speed for dynamic NeRFs so far. We also examined the performance of the original implementation of InstantNGP [19] as a baseline with no temporal extension. All of the grid representation models in the experiments are trained on a single RTX 3090 GPU for fair comparison with [6, 10].", + "bbox": [ + 75, + 642, + 467, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results are shown with elapsed training time in Tab. 4. The grid representation demonstrates much faster training speed even compared to the recent voxel-grid based methods [6, 10]. In comparison with neural network models [23], our method exceeds SSIM and LPIPS of [23] in 5 minutes which have arithmetically 240 times faster training speed. By taking the benefits of fully fused neural network [20] as well as the efficient hash grid repre", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f58f82202ffb2df305e1cc42c46dcf5fba478ad3c402d5738b963c8da91d9086.jpg", + "table_caption": [ + "Table 3. Experimental results on DyNeRF [12] datasets." + ], + "table_footnote": [], + "table_body": "
Train timePSNRSSIMLPIPSAVG
InstantNGP [19]5 min20.280.8880.1460.077
D-NeRF [23]20 hours30.430.950.070.02
NDVG-full [10]35 min27.840.8620.0410.029
TinueVox-B [6]30 min32.670.9710.0410.016
NDVG-half [10]23 min27.150.8570.0480.033
TinueVox-S [6]8 min30.750.9560.0670.023
Ours-grid1 min26.770.9330.1070.039
Ours-grid5 min29.730.9610.0630.024
Ours-grid8 min29.840.9620.0620.023
", + "bbox": [ + 504, + 340, + 885, + 470 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/909dc8d9942fc307d5141842b08d46a6ad8817011fb290fc1e832d2b82d29bb8.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of training time and performance on D-NeRF datasets using the grid representation." + ], + "table_footnote": [], + "table_body": "
D-NeRF datasetHyperNeRF vrig
PSNRSSIMLPIPSPSNRMS-SSIM
NN (dynamic only)30.910.9630.04324.070.861
NN (dynamic+static)32.730.9740.03324.350.866
Grid (dynamic only)29.080.9480.07622.110.755
Grid (dynamic+static)29.840.9620.06222.980.802
", + "bbox": [ + 504, + 520, + 880, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study on the effectiveness of static features.", + "bbox": [ + 511, + 604, + 877, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sentation, our method quickly learns meaningful features in dynamic scenes and generates acceptable novel view synthesis results with just one minute of training. Our method shows superior SSIM, LPIPS and slightly inferior PSNR to TinueVox-S [6], which indicates clearer and more detailed rendering results, when trained for the same period of time.", + "bbox": [ + 496, + 641, + 890, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We depict training progress of the grid representation qualitatively in Fig. 6. We also compared the results of TinueVox-S [6] which are trained for 8 minutes. After one minute of training, the grid representation model produces blurry images but accurately render dynamic regions compared to InstantNGP [19]. After 5 minutes, the model successfully renders sharp images that are similar to the ground truth images. Notably, our model exhibits more sharp and clear results than TinueVox-S despite shorter training time, which results in better SSIM and LPIPS..", + "bbox": [ + 496, + 733, + 890, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we compared the performance of the grid repre", + "bbox": [ + 517, + 885, + 888, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4218", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52cd234492271ee32b58b128e4eca79d5dadbdc3930297731d6885878f44447a.jpg", + "image_caption": [ + "Figure 6. Qualitative results of the grid representation. While InstantNGP [19] is not able to reconstruct dynamic regions, our method produces acceptable rendering results in one minute and performs superior to TiNueVox-S [6] with shorter training time." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 872, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2be207e18b621c54cd9a068a2f8376b3ee623e1be515aa151e0e62be6e107db3.jpg", + "image_caption": [ + "Figure 7. Performance comparison with different feature dimension settings. Using smaller feature dimensions results in faster training but quicker saturation of model performance." + ], + "image_footnote": [], + "bbox": [ + 83, + 339, + 254, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d76829c607b7fcee9142d93845602e78203c5aee593bedc1beeb3eb81a7893e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 339, + 450, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sentation by varying the dimension of static and dynamic feature vectors, $m_{s}$ and $m_{d}$ . PSNR and LPIPS on test images are measured per minute and illustrated in Fig. 7. We trained the grid representation models in three different settings, $m_{s} = 2, m_{d} = 6$ , $m_{s} = 2, m_{d} = 2$ , $m_{s} = 4, m_{d} = 4$ . When smaller feature dimension is used, the training speed is faster, so PSNR increases fast in the early stage of training. However, the performance also saturates faster, and LPIPS is inferior to the other settings. The model with $m_{s} = 2, m_{d} = 6$ significantly outperformed the others in terms of LPIPS.", + "bbox": [ + 75, + 511, + 468, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Effectiveness of the Static Features", + "text_level": 1, + "bbox": [ + 76, + 689, + 379, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of the static features, we conducted an ablation study on the static features and illustrated the results in Tab. 5. For both the neural representation (NN) and the grid representation (Grid), the models with static features performs superior to the ones using only dynamic features in all metrics. Additional ablation studies and qualitative results can be found in the supplementary materials.", + "bbox": [ + 75, + 713, + 468, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Failure Cases", + "text_level": 1, + "bbox": [ + 76, + 845, + 218, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the proposed feature interpolation is able to learn meaningful spatiotemporal features in most cases,", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19bb8c6b5be247b5e0af71038c78a774d2b58e7cbeabe3eca0e87d0f6a4b7386.jpg", + "image_caption": [ + "Figure 8. Examples of failure cases." + ], + "image_footnote": [], + "bbox": [ + 501, + 334, + 890, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "there are a few failure cases as presented in Fig. 8. For instance, our method has difficulty in recovering 3D structures when small objects in a sequence rapidly move (Fig. 8 left) or when there exist dynamic regions that are not observed in the training sequence (Fig. 8 right).", + "bbox": [ + 498, + 473, + 892, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 569, + 617, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective feature interpolation method for training dynamic NeRFs. Both the neural representation and the grid representation showed impressive performance, and the smoothness term applied to the intermediate feature vectors further improves the performance. Since these methods are unrelated to the existing methods of modeling deformations or estimating scene flows, we believe that the proposed method suggests a new direction of training dynamic NeRFs.", + "bbox": [ + 496, + 597, + 890, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While the neural representation model shows high-quality rendering results owing to the representation power of neural networks, it requires hours of training and seconds of rendering which impose barriers to real-time applications. On the other hand, the grid representation is able to render dynamic scenes in less than a second after a few minutes of training, which makes it more practical for real-world applications. Both representations are mutually complementary, and investigating hybrid representations that take advantages of both representations would be an interesting research direction.", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4219", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 2", + "[2] Pontus Andersson, Jim Nilsson, Tomas Akenine-Möller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 6", + "[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 2, pages 690-696. IEEE, 2000. 2", + "[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2", + "[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 2, 3, 6, 7, 8", + "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3", + "[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2", + "[10] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3, 6, 7", + "[11] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 1", + "[12] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 4, 6, 7", + "[13] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6, 7", + "[14] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 1", + "[15] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019. 7", + "[16] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 4", + "[17] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 7", + "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 6, 7", + "[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 3, 4, 6, 7, 8", + "[20] Thomas Müller, Fabrice Rousselle, Jan Novák, and Alexander Keller. Real-time neural radiance caching for path tracing. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 7", + "[21] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 6, 7", + "[22] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 4,5,6,7", + "[23] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2, 6, 7", + "[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4220", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3", + "[25] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Nonrigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2", + "[26] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 3", + "[27] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2", + "[28] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 3", + "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + ], + "bbox": [ + 78, + 90, + 468, + 484 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4221", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_model.json b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff8e2a8bef425016fae682b0eae54ed0d477d171 --- /dev/null +++ b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.181, + 0.865, + 0.217 + ], + "angle": 0, + "content": "Suntheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, Nahiup Kang Samsung Advanced Institute of Technology (SAIT)" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.219, + 0.871, + 0.235 + ], + "angle": 0, + "content": "{sh2019.park, minjungs.son, swan.jang, ychun.ahn, jiyeon31.kim, nahyup.kang}@samsung.com" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.285, + 0.862, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.542, + 0.893, + 0.613 + ], + "angle": 0, + "content": "Figure 1. We propose simple yet effective feature interpolation methods for training neural radiance fields of dynamic scenes based on temporal interpolation. We provide two different feature vector representations, neural representation (top) and grid representation (bottom), both of which are the concatenation of static feature vectors and temporally-interpolated dynamic feature vectors. The neural representation exhibits high-quality rendering performance with small-sized models while the grid representation shows competitive rendering results with astonishingly fast training speed." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.627, + 0.314, + 0.643 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Temporal interpolation often plays a crucial role to learn meaningful representations in dynamic scenes. In this paper, we propose a novel method to train spatiotemporal neural radiance fields of dynamic scenes based on temporal interpolation of feature vectors. Two feature interpolation methods are suggested depending on underlying representations, neural networks or grids. In the neural representation, we extract features from space-time inputs via multiple neural network modules and interpolate them based on time frames. The proposed multi-level feature interpolation network effectively captures features of both short-term and long-term time ranges. In the grid representation, space-time features are learned via four-dimensional hash grids, which remarkably reduces training time. The grid representation shows more than \\(100 \\times\\) faster training speed than the previous neural-net-based methods while maintaining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.734 + ], + "angle": 0, + "content": "the rendering quality. Concatenating static and dynamic features and adding a simple smoothness term further improve the performance of our proposed models. Despite the simplicity of the model architectures, our method achieved state-of-the-art performance both in rendering quality for the neural representation and in training speed for the grid representation." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.768, + 0.633, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3D reconstruction and photo-realistic rendering have been long-lasting problems in the fields of computer vision and graphics. Along with the advancements of deep learning, differentiable rendering [11, 14] or neural rendering, has emerged to bridge the gap between the two problems. Recently proposed Neural Radiance Field (NeRF) [18] has finally unleashed the era of neural rendering. Using NeRF," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4212" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "it is able to reconstruct accurate 3D structures from multiple 2D images and to synthesize photo-realistic images from unseen viewpoints. Tiny neural networks are sufficient to save and retrieve complex 3D scenes, which can be trained in a self-supervised manner given 2D images and camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Meanwhile, as our world typically involves dynamic changes, it is crucial to reconstruct the captured scene through 4D spatiotemporal space. Since it is often not possible to capture scenes at different viewpoints simultaneously, reconstructing dynamic scenes from images is inherently an under-constrained problem. While NeRF was originally designed to deal with only static scenes, there have been a few approaches in the literature that extend NeRF to dynamic scenes [13,22,23,25], which are so called as dynamic NeRFs. Inspired by the non-rigid structure-from-motion algorithms [1, 4] that reconstruct 3D structures of deformable objects, most previous works solved the under-constrained setting by estimating scene deformations [21,22,25] or 3D scene flows [5,9,12] for each frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.579 + ], + "angle": 0, + "content": "However, since the parameters of deformation estimation modules are jointly optimized with NeRF network simultaneously, it is questionable that the modules can accurately estimate deformations or scene flows in accordance with its design principles. In many dynamic scenes, it is challenging to resolve the ambiguities whether a point was newly appeared, moved, or changed its color. It is expected that those ambiguities and deformation estimation can be separately solved within a single network, but in practice, it is hard to let the network implicitly learn the separation, especially for general dynamic scenes without any prior knowledge of deformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.703 + ], + "angle": 0, + "content": "On the other hand, grid representations in NeRF training [7, 19, 24] have grabbed a lot of attentions mainly due to its fast training speed. Simple trilinear interpolation is enough to fill in the 3D space between grid points. While the representation can be directly adopted to dynamic NeRFs together with the warping estimation module [6], it still requires additional neural networks that affects training and rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Motivated by the aforementioned analyses, we present a simple yet effective architecture for training dynamic NeRFs. The key idea in this paper is to apply feature interpolation to the temporal domain instead of using warping functions or 3D scene flows. While the feature interpolation in 2D or 3D space has been thoroughly studied, to the best of our knowledge, feature interpolation method in temporal domain for dynamic NeRF has not been proposed yet. We propose two multi-level feature interpolation methods depending on feature representation which is either neural nets or hash grids [19]. Overview of the two representations, namely the neural representation and the grid representation, are illustrated in Fig. 1. In addition," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "noting that 3D shapes deform smoothly over time in dynamic scenes, we additionally introduced a simple smoothness term that encourages feature similarity between the adjacent frames. We let the neural networks or the feature grids to learn meaningful representations implicitly without imposing any constraints or restrictions but the smoothness regularizer, which grants the flexibility to deal with various types of deformations. Extensive experiments on both synthetic and real-world datasets validate the effectiveness of the proposed method. We summarized the main contributions of the proposed method as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.27, + 0.892, + 0.345 + ], + "angle": 0, + "content": "- We propose a simple yet effective feature extraction network that interpolates two feature vectors along the temporal axis. The proposed interpolation scheme outperforms existing methods without having a deformation or flow estimation module." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.355, + 0.892, + 0.415 + ], + "angle": 0, + "content": "- We integrate temporal interpolation into hash-grid representation [19], which remarkably accelerates training speed more than \\(100 \\times\\) faster compared to the neural network models." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.425, + 0.892, + 0.456 + ], + "angle": 0, + "content": "- We propose a smoothness regularizer which effectively improves the overall performance of dynamic NeRFs." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.27, + 0.892, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.642, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Dynamic NeRF. There have been various attempts to extend NeRF to dynamic scenes. Existing methods can be categorized into three types: warping-based, flow-based, and direct inference from space-time inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.645 + ], + "angle": 0, + "content": "Warping-based methods learn how 3D structure of the scene is deformed. 3D radiance field for each frame is warped to single or multiple canonical frames using the estimated deformation. Deformation is parameterized as 3D translation [23], rotation and translation via angle-axis representation [21, 22], or weighted translation [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.735 + ], + "angle": 0, + "content": "On the other hand, flow-based methods estimate the correspondence of 3d points in consecutive frames. Neural scene flow fields [13] estimates 3D scene flow between the radiance fields of two time stamps. Xian et al. [27] suggested irradiance fields which spans over 4D space-time fields with a few constraints derived from prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although warping-based and flow-based approaches showed successful results, they have a few shortcomings. For instance, warping-based methods cannot deal with topological variations. Since those methods warp every input frame to a single canonical scene, it is hard to correctly represent newly appeared or disappeared 3D radiance fields in the canonical scene. HyperNeRF [22] learns hyperspace which represents multidimensional canonical shape bases. Neural scene flow fields [13] solve the problem by introducing occlusion masks which are used to estimate the regions where scene flows are not applicable. However, the masks" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4213" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "and the flow fields should be learned simultaneously during the training, which makes the training procedure complicated and imposes dependence on extra information such as monocular depth or optical flow estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.471, + 0.287 + ], + "angle": 0, + "content": "Without estimating shape deformations or scene flows, DyNeRF [12] used a simple neural network to train NeRFs of dynamic scenes, although the experiments are only conducted on synchronized multi-view videos. It utilizes 3D position and time frame as inputs, and predicts color and density through a series of fully connected neural networks. Our neural representation applies temporal interpolation to intermediate features which enhances representation power for dynamic features while keeping the simple structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.471, + 0.681 + ], + "angle": 0, + "content": "Category-specific NeRFs are one of the popular research directions for training dynamic objects such as humans [8, 26] or animals [28]. They are based on parametric models or use additional inputs such as pose information. Our method focuses on training NeRFs of general dynamic scenes without any prior knowledge about scenes or objects. Nevertheless, temporal interpolation proposed in this paper can be easily applied to the dynamic NeRFs with templates. Grid representations in NeRF. One of the major drawbacks of the original NeRF [18] is slow training speed. A few concurrent works are proposed to speed up NeRF training. Among them, grid-based representation such as Plenoxels [7], direct voxel grid optimization [24] show superior performance in terms of training time, where it takes only a few minutes to learn plausible radiance fields. Combining grid representation with a hash function further improves the efficiency and training speed of the feature grids [19]. Fang et al. [6] firstly applied a voxel grid representation to train dynamic NeRFs and achieved much faster training speed compared to the neural-network-based methods. Guo et al. [10] also proposed a deformable voxel grid method for fast training. Unlike aforementioned works, our grid representation does not estimate warping or deformation. Instead, it directly estimates color and density using the features obtained from 4D hash grids, which decreases the computational burden and enables faster training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.462, + 0.712 + ], + "angle": 0, + "content": "3. Temporal Interpolation for Dynamic NeRF" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.718, + 0.218, + 0.733 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.468, + 0.817 + ], + "angle": 0, + "content": "Given 3D position \\(\\mathbf{x} \\in \\mathbb{R}^3\\) and 2D viewing direction \\(\\mathbf{d} \\in \\mathbb{R}^2\\), NeRF [18] aims to estimate volume density \\(\\sigma \\in \\mathbb{R}\\) and emitted RGB color \\(\\mathbf{c} \\in \\mathbb{R}^3\\) using neural networks. We formulate dynamic NeRF as a direct extension from 3D to 4D by adding time frame \\(t \\in \\mathbb{R}\\) to inputs, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.829, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n(\\mathbf {c}, \\sigma) = f (\\mathbf {v} (\\mathbf {x}, t), \\mathbf {d}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(f\\) is implemented as fully connected neural networks and a space-time feature representation \\(\\mathbf{v}\\) can be MLP-based neural nets or explicit grid values as explained in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "From the camera with origin \\(\\mathbf{o}\\) and ray direction \\(\\mathbf{d}\\), the color of camera ray \\(\\mathbf{r}(u) = \\mathbf{o} + u\\mathbf{d}\\) at time frame \\(t\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.131, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} (\\mathbf {r}, t) = \\int_ {u _ {n}} ^ {u _ {f}} U (u, t) \\sigma (\\mathbf {v}) \\mathbf {c} (\\mathbf {v}, \\mathbf {d}) d u, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.174, + 0.892, + 0.279 + ], + "angle": 0, + "content": "where \\( u_{n} \\) and \\( u_{f} \\) denote the bounds of the scene volume, \\( \\mathbf{v} \\) is an abbreviation of \\( \\mathbf{v}(\\mathbf{r}(u), t) \\) and \\( U(u, t) = \\exp(-\\int_{u_n}^u \\sigma(\\mathbf{v}(\\mathbf{r}(s), t)) ds) \\) is the accumulated transmittance from \\( u_{n} \\) to \\( u \\). Then, the RGB color loss \\( L_{c} \\) is defined to minimize the \\( l_{2} \\)-loss between the estimated colors \\( \\hat{C}(\\mathbf{r}, t) \\) and the ground truth colors \\( C(\\mathbf{r}, t) \\) over all rays \\( \\mathbf{r} \\) in camera views \\( R \\) and time frames \\( t \\in T \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.287, + 0.892, + 0.32 + ], + "angle": 0, + "content": "\\[\nL _ {c} = \\sum_ {\\mathbf {r} \\in R, t \\in T} | | \\hat {C} (\\mathbf {r}, t) - C (\\mathbf {r}, t) | | _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.811, + 0.347 + ], + "angle": 0, + "content": "3.2. Space-Time Feature Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.43 + ], + "angle": 0, + "content": "The main contribution of our paper lies on the novel feature representation method for dynamic NeRFs. We define the feature vector \\(\\mathbf{v}\\), which is fed into the neural network to determine a color and a volume density, as the concatenation of static and dynamic feature vectors, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.44, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} (\\mathbf {x}, t) = \\left[ \\mathbf {v} _ {s} (\\mathbf {x}), \\mathbf {v} _ {d} (\\mathbf {x}, t) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.724 + ], + "angle": 0, + "content": "where \\([\\cdot, \\cdot]\\) means concatenation operator of two vectors. The static feature vector \\(\\mathbf{v}_s\\) only depends on the 3D position \\(\\mathbf{x}\\). As most dynamic scenes also contain static regions such as background, \\(\\mathbf{v}_s\\) is designed to learn the static features that are consistent across time. Meanwhile, \\(\\mathbf{v}_d\\) learns dynamic features which may vary across time. We propose two novel feature representation methods in which temporal interpolation is applied. First, the neural representation, which is essentially a series of neural networks combined with linear interpolation, is suggested in Sec. 3.2.1. The model trained using the neural representation is able to render high-quality space-time view synthesis with small-sized neural networks (\\(\\sim 20\\mathrm{MB}\\)). Second, the grid representation, which is an temporal extension of recently proposed voxel grid representations [19], is explained in Sec. 3.2.2. Dynamic NeRFs can be trained in less than 5 minutes with the proposed grid representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.712, + 0.757 + ], + "angle": 0, + "content": "3.2.1 Neural Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the neural representation, features that fed into the template NeRFs are determined by a series of neural nets. In other words, both static and dynamic feature extractor are formulated as multi-layer perceptrons (MLPs). First, the whole time frame is divided into equally spaced time slots. For each time slot, two MLPs are assigned. Then the whole feature vector is interpolated from the assigned MLPs. Overall feature extraction procedure is illustrated in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4214" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.466, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.092, + 0.861, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 2. Illustration of the arrangements of MLPs (left) and the feature extraction (right) in the neural representation. \\(\\phi_s\\) and \\(\\phi_{ij}\\) are small MLPs that extract static and dynamic features respectively. The static features \\(\\mathbf{v}_s\\) and the dynamic feature \\(\\mathbf{v}_d\\) are concatenated and fed into the template NeRF. Detailed structures of the template NeRF can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Concretely, let \\(\\mathbf{v}_d(\\mathbf{x},\\mathbf{z}_t)\\) be the feature vector for a 3D point \\(\\mathbf{x}\\) at time \\(t\\), which will be fed to the template NeRF. Here, we used \\(\\mathbf{z}_t\\) as an embedding vector for an input time frame \\(t\\) [12, 16, 22]. For equally spaced \\(n\\) time slots, there are \\(n + 1\\) keyframes \\(t_i = \\frac{i}{n} (i = 0,1,2,\\dots,n)\\). An MLP \\(\\phi_{i}\\) is assigned to each keyframe \\(t_i\\) which is responsible for two adjacent time slots \\([t_{i - 1},t_i]\\) and \\([t_i,t_{i + 1}]\\). For inputs with time \\(t\\in [t_i,t_{i + 1}]\\), \\(\\mathbf{x}\\) and \\(\\mathbf{z}_t\\) are fed into \\(\\phi_{i}\\) and \\(\\phi_{i + 1}\\). Then, the outputs from two MLPs are interpolated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.468, + 0.47, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) = \\Delta t \\cdot \\phi_ {i} (\\mathbf {x}, \\mathbf {z} _ {t}) + (1 - \\Delta t) \\cdot \\phi_ {i + 1} (\\mathbf {x}, \\mathbf {z} _ {t}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.318, + 0.512 + ], + "angle": 0, + "content": "where \\(\\Delta t = (t_{i + 1} - t) / (t_{i + 1} - t_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.47, + 0.678 + ], + "angle": 0, + "content": "The purpose of this interpolation is to efficiently learn the features between keyframes in a scalable manner. Thanks to the temporal interpolation, we can learn the features of the continuous time range \\([t_i, t_{i+1}]\\) by enforcing the MLPs \\(\\phi_i\\) and \\(\\phi_{i+1}\\) responsible for that time range. While the static feature \\(\\mathbf{v}_s\\) represents the features across the whole timeline, \\(\\mathbf{v}_d\\) represents the features that are focused more on dynamic regions. In addition, since each MLP for dynamic feature is responsible only for two adjacent time slots, it is able to make each \\(\\phi_i\\) learn features that are specific to a certain period of time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.784 + ], + "angle": 0, + "content": "To exploit the features of both long-term and short-term, there can be multiple levels of dynamic features which have different number of keyframes. For multi-level dynamic feature extractor with level \\( l \\), each level contains different number of keyframes \\( n_1, n_2, \\dots, n_l \\). Let \\( \\mathbf{v}_{di} \\) denote the dynamic feature of \\( i \\)-th level, then the output dynamic feature is the concatenation of features from all levels, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.797, + 0.47, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) = \\left[ \\mathbf {v} _ {d 1} (\\mathbf {x}, \\mathbf {z} _ {t}), \\mathbf {v} _ {d 2} (\\mathbf {x}, \\mathbf {z} _ {t}), \\dots , \\mathbf {v} _ {d l} (\\mathbf {x}, \\mathbf {z} _ {t}) \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this paper, we used the settings \\( l = 2 \\), \\( n_1 = 5 \\), \\( n_2 = 20 \\) otherwise stated. The dimensions of feature vectors are determined as 128 for \\( \\mathbf{v}_s \\), and 64 for \\( \\mathbf{v}_0 \\) and \\( \\mathbf{v}_1 \\). MLP with one hidden layer whose hidden size is the same as its output dimension is used for feature extraction of both \\( \\mathbf{v}_s \\) and \\( \\mathbf{v}_{di} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.381 + ], + "angle": 0, + "content": "The concatenated feature vectors are fed to the template NeRF which outputs volume density \\(\\sigma\\) and emitted color c. We used the same structure as used in HyperNeRF [22] for the template NeRF of the neural representation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.398, + 0.7, + 0.414 + ], + "angle": 0, + "content": "3.2.2 Grid Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Recently, InstantNGP [19] suggested a novel multi-level grid representation with hash tables. We adopt the hash grid representation from [19] and extend it for fast dynamic NeRF training. Similar to the neural representation in Sec. 3.2.1, the feature vector from the proposed hash grid contains static and dynamic feature vectors. The static feature \\(\\mathbf{v}_s\\) is analogous to the one used in [19]. On the other hand, the dynamic feature \\(\\mathbf{v}_d\\) comes from the 4D hash grids." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Concretely, to extract static and dynamic feature vectors whose dimensions are \\( m_{s} \\) and \\( m_{d} \\) respectively, a hash table of size \\( H \\) that contains \\( (m_{s} + m_{d}) \\) dimension feature vectors is constructed. The hash function of \\( d \\)-dimensional vector, \\( h_d(\\mathbf{x}) \\), is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.629, + 0.892, + 0.67 + ], + "angle": 0, + "content": "\\[\nh _ {d} (\\mathbf {x}) = \\bigoplus_ {i = 1} ^ {d} \\left(x _ {i} P _ {i}\\right) \\mod H, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.725 + ], + "angle": 0, + "content": "where \\(P_{i}\\) is a large prime number and \\(\\bigoplus\\) is an XOR operator. Then, the feature vector \\(\\mathbf{v}\\) is retrieved by concatenating the outputs of 3D and 4D hash functions:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.736, + 0.892, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} (\\mathbf {x}, t) = \\left[ h _ {3} (\\mathbf {x}), h _ {4} (\\mathbf {x}, t) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.839 + ], + "angle": 0, + "content": "The 3D and 4D grids are constructed as the multi-level hash grids proposed in [19]. We applied different scaling factors for 3D space and time frame since the number of frames for training sequences are usually much smaller than the finest 3D grid resolutions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.746, + 0.864 + ], + "angle": 0, + "content": "3.3. Smoothness Regularization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As our dynamic world smoothly changes, it is reasonable to impose a smoothness term to adjacent time frames." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4215" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.092, + 0.852, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Illustration of grid representation with hash tables at a certain level. A static feature vector of a 3D grid point is extracted via 3D hash function, while dynamic feature vector for 4D grid point is extracted via 4D hash. The static feature vector \\(\\mathbf{v}_s\\) is determined as trilinear interpolation of 8 vectors of grid points, while the dynamic feature vector \\(\\mathbf{v}_d\\) is calculated as quadrilinear interpolation of 16 vectors. In the figure, we only depicted feature retrieval process of two grid points for illustration purpose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.471, + 0.398 + ], + "angle": 0, + "content": "We note that the smoothness term is only applied to the input feature space, not to the estimated outputs such as RGB color or density. For the neural representation, we provided a simple smoothness term which is calculated as," + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.407, + 0.47, + 0.426 + ], + "angle": 0, + "content": "\\[\nL _ {s} (\\mathbf {x}, t) = \\left\\| \\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) - \\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t + 1}) \\right\\| _ {2} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.525 + ], + "angle": 0, + "content": "This regularization term has two advantages. First, it reflects the intuition that the observed point \\(\\mathbf{x}\\) at time \\(t\\) will be stationary if there is no observation for \\(\\mathbf{x}\\) at time \\(t + 1\\). Second, by propagating additional gradients to the feature networks or the hash grids, the smoothness term acts as a regularizer that stabilizes the training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.556 + ], + "angle": 0, + "content": "For the grid representation, we impose a smoothness term to the grid points that are temporally adjacent:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.564, + 0.47, + 0.599 + ], + "angle": 0, + "content": "\\[\nL _ {s} (\\mathbf {x}, t) = \\frac {1}{n _ {f} ^ {2}} \\| h _ {4} (\\mathbf {x}, t _ {a}) - h _ {4} (\\mathbf {x}, t _ {b}) \\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.728 + ], + "angle": 0, + "content": "where \\( n_f \\) is the number of frames in the training sequence and \\( t_a, t_b \\) are two adjacent grid points that satisfies \\( t_a \\leq t \\leq t_b \\). In fact, Eq. (10) can be obtained from Eq. (9) in the grid representation, which indicates that imposing smoothness term to adjacent time frames is equivalent to add smoothness to two temporally adjacent grids with constant multiplier. Detailed derivation of this relationship is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.47, + 0.789 + ], + "angle": 0, + "content": "The smoothness term is added to the loss function with a weight \\(\\lambda\\), and the total loss is minimized together during the training. Accordingly, the total loss of our dynamic neural radiance fields is given by" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.8, + 0.469, + 0.815 + ], + "angle": 0, + "content": "\\[\nL = L _ {c} + \\lambda L _ {s}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We observe that the smoothness term is especially powerful in the static background regions that appeared only in a small fraction of frames. In Fig. 4, we show examples of the rendered images which are trained with and without the smoothness term. It can be clearly seen that the boxed" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.334, + 0.625, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.497, + 0.622, + 0.507 + ], + "angle": 0, + "content": "(a) w/o smooth term" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.334, + 0.753, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.496, + 0.748, + 0.507 + ], + "angle": 0, + "content": "(b) w/ smooth term" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.334, + 0.882, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.496, + 0.845, + 0.507 + ], + "angle": 0, + "content": "(c) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.52, + 0.894, + 0.577 + ], + "angle": 0, + "content": "Figure 4. Effectiveness of the smoothness term. It can be observed that the model trained with the smoothness term accurately renders the corner part which appears in only few frames within the whole sequence (blue box) and removes spurious artifacts (red box)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.627 + ], + "angle": 0, + "content": "regions of the rendered image shows much plausible when the model is trained with the smoothness term." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.719, + 0.652 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.81 + ], + "angle": 0, + "content": "In the neural representation, we adopted the template NeRF architecture of HyperNeRF [22] for density and color estimation. The network is an 8-layered MLP with hidden size of 256. The dimension of the embedding vector \\(\\mathbf{z}_t\\) is set to 8. The smoothness weight \\(\\lambda\\) is set to 0.01 or 0.001 depending on the characteristics of the datasets used in the experiments. We set \\(\\lambda\\) to a large value in the sequences that the viewpoint change is significant in a short period of time although performance variations depending on the value of \\(\\lambda\\) is not significant." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the grid representation, \\(\\lambda\\) is set to \\(1e^{-4}\\). We applied the smoothness loss only for the finest two levels of the temporal dimension since applying it to every level slows training speed without much improvements in performance. After feature vectors are extracted from the hash grids, we fed them to a 3-layered MLP with hidden size of 128 to estimate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4216" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "volume density followed by one additional layer for RGB color estimation similar to [19]. We set \\( m_{s} = 2 \\), \\( m_{d} = 6 \\) and \\( H = 2^{19} \\) for the grid representation experiments otherwise stated. The 3D and 4D hash grids are composed of 12 levels. For the spatial dimension, we set the base resolution to 8, and the scaling factor is set to 1.45. The base resolution of the temporal dimension is 2, which is multiplied by 1.4 in every other level. Detailed hyperparameter settings and network architectures can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.258, + 0.283, + 0.274 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.283, + 0.18, + 0.297 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.469, + 0.397 + ], + "angle": 0, + "content": "To validate the superiority of our method, we conducted extensive experiments on various datasets. We used three publicly available datasets that are used for dynamic NeRFs. For all experiments, we trained our models for each sequence individually, and then per-scene results are averaged and reported in this section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.473 + ], + "angle": 0, + "content": "D-NeRF Dataset [23]. The dataset consists of synthetically rendered images of moving and deforming 3D objects. For each frame, the image is rendered via a synthetic camera of random rotation. There are eight scenes in total, and each scene contains 50-200 training images with 20 test views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.579 + ], + "angle": 0, + "content": "HyperNeRF Dataset [22]. The dataset contains video sequences taken from mobile phones. There are four sequences in vrig-dataset which are taken using the camera rig with two phones vertically aligned. In addition, there are six sequences in interp-dataset which are taken from a single camera in order to estimate the image in the middle of two consecutive training images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.654 + ], + "angle": 0, + "content": "DyNeRF Dataset [12]. The dataset consists of videos obtained from a capture system that consists of 21 GoPro Black Hero cameras. The cameras are located at a fixed position, and all video frames are synchronized to build the multi-view video dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.288, + 0.682 + ], + "angle": 0, + "content": "4.2. Neural Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.855 + ], + "angle": 0, + "content": "In this section, we reported the performance of the proposed neural representation models. First, the experimental results on D-NeRF dataset are shown in Tab. 1. Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS) [29] are used as evaluation metrics following the previous works. We also reported the average metric (AVG) proposed in [3] which aggregates three metrics to a single value. Our method with the neural representation (Ours-NN) achieves state-of-the-art results on all evaluation metrics. It is worth noting that the smoothness term dramatically improves overall performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Next, we evaluated our method on HyperNeRF dataset [22] and compared with existing methods, which is shown in Tab. 2. Here, we reported PSNR and multi" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.088, + 0.885, + 0.194 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓AVG ↓
NeRF [18]19.000.870.180.09
T-NeRF [23]29.500.950.080.03
D-NeRF [23]30.430.950.070.02
NDVG-full [10]27.840.8620.0410.029
TiNeuVox-B [6]32.670.9710.0410.016
Ours-NN (w/o smooth)30.180.9630.0380.019
Ours-NN (w/ smooth)32.730.9740.0330.014
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.204, + 0.862, + 0.218 + ], + "angle": 0, + "content": "Table 1. Experimental results on D-NeRF [23] datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.241, + 0.892, + 0.362 + ], + "angle": 0, + "content": "scale SSIM (MS-SSIM), and excluded LPIPS metric since its value cannot reliably reproduced [6]. Our method shows second-best results on both vrig and interp datasets. While flow-based method [13] suffers from interpolating motions between two consecutive frames, our method, which implicitly learns intermediate feature representation of inbetween frames, achieves competitive performance with warping-based methods [21, 22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Qualitative results for the neural representation are shown in Fig. 5. It is clearly observed that our method captures fine details compared to D-NeRF [23]. We highlighted the regions that show notable differences with colored boxes and placed zoomed-in images of the regions next to the rendered images. In addition, we also qualitatively compared our method with HyperNeRF [22] on their datasets. When the warping estimation module of [22] does not correctly estimate warping parameters, HyperNeRF produces implausible results. It can be observed that the head and the body of the chicken toy is not properly interlocked and the position of the hand peeling banana is incorrect. On the other hand, our method accurately recovers 3D geometry of those scenes. Thus, without using the physically meaningful warping estimation module in neural networks, the proposed temporal interpolation and the smoothness regularizer provides simple yet effective way to learn complex deformations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Lastly, performance evaluation on DyNeRF dataset [12] is presented in Tab. 3. We adopt PSNR, LPIPS, and FLIP [2] for evaluation metrics to compare with previous works. Our method achieves best PSNR while ranked second in LPIPS and FLIP. However, those two metrics are also better than DyNeRF† [12] which does not use importance sampling strategy in [12]. Since we do not use any sampling strategy during training, it can be concluded that our feature interpolation method is superior to the network architecture of [12]. Notably, our method outperforms DyNeRF with smaller network size (20MB) than the DyNeRF models (28MB)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.694, + 0.847 + ], + "angle": 0, + "content": "4.3. Grid Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We used D-NeRF datasets to evaluate the performance of the proposed grid representation. Since this representation is mainly intended for fast training, we report the re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.086, + 0.895, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our method (Ours-NN) on of D-NeRF [23] and HyperNeRF [22] datasets. The regions that show significant difference with compared methods are scaled up (red and green boxes)." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.342, + 0.462, + 0.442 + ], + "angle": 0, + "content": "
vriginterp
PSNR↑MS-SSIM↑PSNR↑MS-SSIM↑
NeRF [18]20.130.74522.270.804
NV [15]16.850.57126.050.911
NSFF [13]26.330.91625.800.883
Nerfies [21]22.230.80328.470.939
HyperNeRF [22]22.380.81429.000.945
Ours-NN24.350.86728.670.940
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.453, + 0.449, + 0.467 + ], + "angle": 0, + "content": "Table 2. Experimental results on HyperNeRF [22] datasets." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.482, + 0.441, + 0.597 + ], + "angle": 0, + "content": "
PSNR↑LPIPS ↓FLIP ↓
MVS19.120.25990.2542
NeuralVolumes [15]22.800.29510.2049
LLFF [17]23.240.23460.1867
NeRF-T [12]28.450.1000.1415
DyNeRF† [12]28.500.09850.1455
DyNeRF [12]29.580.08320.1347
Ours-NN29.880.09600.1413
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.608, + 0.441, + 0.622 + ], + "angle": 0, + "content": "Table 3. Experimental results on DyNeRF [12] datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.779 + ], + "angle": 0, + "content": "sults in a short period time (\\(\\sim 8\\) minutes) and compare the results with the concurrent works [6, 10], both of which are based on voxel grid representation and showed the fastest training speed for dynamic NeRFs so far. We also examined the performance of the original implementation of InstantNGP [19] as a baseline with no temporal extension. All of the grid representation models in the experiments are trained on a single RTX 3090 GPU for fair comparison with [6, 10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The quantitative results are shown with elapsed training time in Tab. 4. The grid representation demonstrates much faster training speed even compared to the recent voxel-grid based methods [6, 10]. In comparison with neural network models [23], our method exceeds SSIM and LPIPS of [23] in 5 minutes which have arithmetically 240 times faster training speed. By taking the benefits of fully fused neural network [20] as well as the efficient hash grid repre" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.341, + 0.886, + 0.472 + ], + "angle": 0, + "content": "
Train timePSNRSSIMLPIPSAVG
InstantNGP [19]5 min20.280.8880.1460.077
D-NeRF [23]20 hours30.430.950.070.02
NDVG-full [10]35 min27.840.8620.0410.029
TinueVox-B [6]30 min32.670.9710.0410.016
NDVG-half [10]23 min27.150.8570.0480.033
TinueVox-S [6]8 min30.750.9560.0670.023
Ours-grid1 min26.770.9330.1070.039
Ours-grid5 min29.730.9610.0630.024
Ours-grid8 min29.840.9620.0620.023
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.483, + 0.89, + 0.511 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of training time and performance on D-NeRF datasets using the grid representation." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.521, + 0.882, + 0.595 + ], + "angle": 0, + "content": "
D-NeRF datasetHyperNeRF vrig
PSNRSSIMLPIPSPSNRMS-SSIM
NN (dynamic only)30.910.9630.04324.070.861
NN (dynamic+static)32.730.9740.03324.350.866
Grid (dynamic only)29.080.9480.07622.110.755
Grid (dynamic+static)29.840.9620.06222.980.802
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.605, + 0.878, + 0.618 + ], + "angle": 0, + "content": "Table 5. Ablation study on the effectiveness of static features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.733 + ], + "angle": 0, + "content": "sentation, our method quickly learns meaningful features in dynamic scenes and generates acceptable novel view synthesis results with just one minute of training. Our method shows superior SSIM, LPIPS and slightly inferior PSNR to TinueVox-S [6], which indicates clearer and more detailed rendering results, when trained for the same period of time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.883 + ], + "angle": 0, + "content": "We depict training progress of the grid representation qualitatively in Fig. 6. We also compared the results of TinueVox-S [6] which are trained for 8 minutes. After one minute of training, the grid representation model produces blurry images but accurately render dynamic regions compared to InstantNGP [19]. After 5 minutes, the model successfully renders sharp images that are similar to the ground truth images. Notably, our model exhibits more sharp and clear results than TinueVox-S despite shorter training time, which results in better SSIM and LPIPS.." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.89, + 0.901 + ], + "angle": 0, + "content": "Finally, we compared the performance of the grid repre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.874, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.893, + 0.313 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of the grid representation. While InstantNGP [19] is not able to reconstruct dynamic regions, our method produces acceptable rendering results in one minute and performs superior to TiNueVox-S [6] with shorter training time." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.34, + 0.255, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.34, + 0.451, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.492 + ], + "angle": 0, + "content": "Figure 7. Performance comparison with different feature dimension settings. Using smaller feature dimensions results in faster training but quicker saturation of model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.678 + ], + "angle": 0, + "content": "sentation by varying the dimension of static and dynamic feature vectors, \\( m_{s} \\) and \\( m_{d} \\). PSNR and LPIPS on test images are measured per minute and illustrated in Fig. 7. We trained the grid representation models in three different settings, \\( m_{s} = 2, m_{d} = 6 \\), \\( m_{s} = 2, m_{d} = 2 \\), \\( m_{s} = 4, m_{d} = 4 \\). When smaller feature dimension is used, the training speed is faster, so PSNR increases fast in the early stage of training. However, the performance also saturates faster, and LPIPS is inferior to the other settings. The model with \\( m_{s} = 2, m_{d} = 6 \\) significantly outperformed the others in terms of LPIPS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.69, + 0.38, + 0.705 + ], + "angle": 0, + "content": "4.4. Effectiveness of the Static Features" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.469, + 0.835 + ], + "angle": 0, + "content": "To validate the effectiveness of the static features, we conducted an ablation study on the static features and illustrated the results in Tab. 5. For both the neural representation (NN) and the grid representation (Grid), the models with static features performs superior to the ones using only dynamic features in all metrics. Additional ablation studies and qualitative results can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.219, + 0.862 + ], + "angle": 0, + "content": "4.5. Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Although the proposed feature interpolation is able to learn meaningful spatiotemporal features in most cases," + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.335, + 0.891, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.436, + 0.804, + 0.451 + ], + "angle": 0, + "content": "Figure 8. Examples of failure cases." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.893, + 0.552 + ], + "angle": 0, + "content": "there are a few failure cases as presented in Fig. 8. For instance, our method has difficulty in recovering 3D structures when small objects in a sequence rapidly move (Fig. 8 left) or when there exist dynamic regions that are not observed in the training sequence (Fig. 8 right)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.57, + 0.619, + 0.585 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.734 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective feature interpolation method for training dynamic NeRFs. Both the neural representation and the grid representation showed impressive performance, and the smoothness term applied to the intermediate feature vectors further improves the performance. Since these methods are unrelated to the existing methods of modeling deformations or estimating scene flows, we believe that the proposed method suggests a new direction of training dynamic NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "While the neural representation model shows high-quality rendering results owing to the representation power of neural networks, it requires hours of training and seconds of rendering which impose barriers to real-time applications. On the other hand, the grid representation is able to render dynamic scenes in less than a second after a few minutes of training, which makes it more practical for real-world applications. Both representations are mutually complementary, and investigating hybrid representations that take advantages of both representations would be an interesting research direction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4219" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Pontus Andersson, Jim Nilsson, Tomas Akenine-Möller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 2, pages 690-696. IEEE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.591 + ], + "angle": 0, + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[10] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[11] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[13] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[14] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[16] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[17] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[20] Thomas Müller, Fabrice Rousselle, Jan Novák, and Alexander Keller. Real-time neural radiance caching for path tracing. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[21] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[22] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 4,5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[23] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[25] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Nonrigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[26] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[27] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.375, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[28] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4221" + } + ] +] \ No newline at end of file diff --git a/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_origin.pdf b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dab314c691ae331bb6ad51fe1fdea45c66d24adf --- /dev/null +++ b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/22a9031d-e022-4324-b56f-e7ab749525c7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab2e995bfd147eadbd0635be64d429b026d31220ed15f6411c8090911e4a4bf0 +size 4767132 diff --git a/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/full.md b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1c2dcbef690e15e0345458ee08a047586edf6b0c --- /dev/null +++ b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/full.md @@ -0,0 +1,308 @@ +# Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields + +Suntheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, Nahiup Kang Samsung Advanced Institute of Technology (SAIT) + +{sh2019.park, minjungs.son, swan.jang, ychun.ahn, jiyeon31.kim, nahyup.kang}@samsung.com + +![](images/7c9bfc4c36b86ff5abc585719ac7ba8dd5aec02a845824cc465d8911ab948ee0.jpg) +Figure 1. We propose simple yet effective feature interpolation methods for training neural radiance fields of dynamic scenes based on temporal interpolation. We provide two different feature vector representations, neural representation (top) and grid representation (bottom), both of which are the concatenation of static feature vectors and temporally-interpolated dynamic feature vectors. The neural representation exhibits high-quality rendering performance with small-sized models while the grid representation shows competitive rendering results with astonishingly fast training speed. + +# Abstract + +Temporal interpolation often plays a crucial role to learn meaningful representations in dynamic scenes. In this paper, we propose a novel method to train spatiotemporal neural radiance fields of dynamic scenes based on temporal interpolation of feature vectors. Two feature interpolation methods are suggested depending on underlying representations, neural networks or grids. In the neural representation, we extract features from space-time inputs via multiple neural network modules and interpolate them based on time frames. The proposed multi-level feature interpolation network effectively captures features of both short-term and long-term time ranges. In the grid representation, space-time features are learned via four-dimensional hash grids, which remarkably reduces training time. The grid representation shows more than $100 \times$ faster training speed than the previous neural-net-based methods while maintaining + +the rendering quality. Concatenating static and dynamic features and adding a simple smoothness term further improve the performance of our proposed models. Despite the simplicity of the model architectures, our method achieved state-of-the-art performance both in rendering quality for the neural representation and in training speed for the grid representation. + +# 1. Introduction + +3D reconstruction and photo-realistic rendering have been long-lasting problems in the fields of computer vision and graphics. Along with the advancements of deep learning, differentiable rendering [11, 14] or neural rendering, has emerged to bridge the gap between the two problems. Recently proposed Neural Radiance Field (NeRF) [18] has finally unleashed the era of neural rendering. Using NeRF, + +it is able to reconstruct accurate 3D structures from multiple 2D images and to synthesize photo-realistic images from unseen viewpoints. Tiny neural networks are sufficient to save and retrieve complex 3D scenes, which can be trained in a self-supervised manner given 2D images and camera parameters. + +Meanwhile, as our world typically involves dynamic changes, it is crucial to reconstruct the captured scene through 4D spatiotemporal space. Since it is often not possible to capture scenes at different viewpoints simultaneously, reconstructing dynamic scenes from images is inherently an under-constrained problem. While NeRF was originally designed to deal with only static scenes, there have been a few approaches in the literature that extend NeRF to dynamic scenes [13,22,23,25], which are so called as dynamic NeRFs. Inspired by the non-rigid structure-from-motion algorithms [1, 4] that reconstruct 3D structures of deformable objects, most previous works solved the under-constrained setting by estimating scene deformations [21,22,25] or 3D scene flows [5,9,12] for each frame. + +However, since the parameters of deformation estimation modules are jointly optimized with NeRF network simultaneously, it is questionable that the modules can accurately estimate deformations or scene flows in accordance with its design principles. In many dynamic scenes, it is challenging to resolve the ambiguities whether a point was newly appeared, moved, or changed its color. It is expected that those ambiguities and deformation estimation can be separately solved within a single network, but in practice, it is hard to let the network implicitly learn the separation, especially for general dynamic scenes without any prior knowledge of deformation. + +On the other hand, grid representations in NeRF training [7, 19, 24] have grabbed a lot of attentions mainly due to its fast training speed. Simple trilinear interpolation is enough to fill in the 3D space between grid points. While the representation can be directly adopted to dynamic NeRFs together with the warping estimation module [6], it still requires additional neural networks that affects training and rendering speed. + +Motivated by the aforementioned analyses, we present a simple yet effective architecture for training dynamic NeRFs. The key idea in this paper is to apply feature interpolation to the temporal domain instead of using warping functions or 3D scene flows. While the feature interpolation in 2D or 3D space has been thoroughly studied, to the best of our knowledge, feature interpolation method in temporal domain for dynamic NeRF has not been proposed yet. We propose two multi-level feature interpolation methods depending on feature representation which is either neural nets or hash grids [19]. Overview of the two representations, namely the neural representation and the grid representation, are illustrated in Fig. 1. In addition, + +noting that 3D shapes deform smoothly over time in dynamic scenes, we additionally introduced a simple smoothness term that encourages feature similarity between the adjacent frames. We let the neural networks or the feature grids to learn meaningful representations implicitly without imposing any constraints or restrictions but the smoothness regularizer, which grants the flexibility to deal with various types of deformations. Extensive experiments on both synthetic and real-world datasets validate the effectiveness of the proposed method. We summarized the main contributions of the proposed method as follows: + +- We propose a simple yet effective feature extraction network that interpolates two feature vectors along the temporal axis. The proposed interpolation scheme outperforms existing methods without having a deformation or flow estimation module. +- We integrate temporal interpolation into hash-grid representation [19], which remarkably accelerates training speed more than $100 \times$ faster compared to the neural network models. +- We propose a smoothness regularizer which effectively improves the overall performance of dynamic NeRFs. + +# 2. Related Work + +Dynamic NeRF. There have been various attempts to extend NeRF to dynamic scenes. Existing methods can be categorized into three types: warping-based, flow-based, and direct inference from space-time inputs. + +Warping-based methods learn how 3D structure of the scene is deformed. 3D radiance field for each frame is warped to single or multiple canonical frames using the estimated deformation. Deformation is parameterized as 3D translation [23], rotation and translation via angle-axis representation [21, 22], or weighted translation [25]. + +On the other hand, flow-based methods estimate the correspondence of 3d points in consecutive frames. Neural scene flow fields [13] estimates 3D scene flow between the radiance fields of two time stamps. Xian et al. [27] suggested irradiance fields which spans over 4D space-time fields with a few constraints derived from prior knowledge. + +Although warping-based and flow-based approaches showed successful results, they have a few shortcomings. For instance, warping-based methods cannot deal with topological variations. Since those methods warp every input frame to a single canonical scene, it is hard to correctly represent newly appeared or disappeared 3D radiance fields in the canonical scene. HyperNeRF [22] learns hyperspace which represents multidimensional canonical shape bases. Neural scene flow fields [13] solve the problem by introducing occlusion masks which are used to estimate the regions where scene flows are not applicable. However, the masks + +and the flow fields should be learned simultaneously during the training, which makes the training procedure complicated and imposes dependence on extra information such as monocular depth or optical flow estimation. + +Without estimating shape deformations or scene flows, DyNeRF [12] used a simple neural network to train NeRFs of dynamic scenes, although the experiments are only conducted on synchronized multi-view videos. It utilizes 3D position and time frame as inputs, and predicts color and density through a series of fully connected neural networks. Our neural representation applies temporal interpolation to intermediate features which enhances representation power for dynamic features while keeping the simple structure. + +Category-specific NeRFs are one of the popular research directions for training dynamic objects such as humans [8, 26] or animals [28]. They are based on parametric models or use additional inputs such as pose information. Our method focuses on training NeRFs of general dynamic scenes without any prior knowledge about scenes or objects. Nevertheless, temporal interpolation proposed in this paper can be easily applied to the dynamic NeRFs with templates. Grid representations in NeRF. One of the major drawbacks of the original NeRF [18] is slow training speed. A few concurrent works are proposed to speed up NeRF training. Among them, grid-based representation such as Plenoxels [7], direct voxel grid optimization [24] show superior performance in terms of training time, where it takes only a few minutes to learn plausible radiance fields. Combining grid representation with a hash function further improves the efficiency and training speed of the feature grids [19]. Fang et al. [6] firstly applied a voxel grid representation to train dynamic NeRFs and achieved much faster training speed compared to the neural-network-based methods. Guo et al. [10] also proposed a deformable voxel grid method for fast training. Unlike aforementioned works, our grid representation does not estimate warping or deformation. Instead, it directly estimates color and density using the features obtained from 4D hash grids, which decreases the computational burden and enables faster training. + +# 3. Temporal Interpolation for Dynamic NeRF + +# 3.1. Preliminaries + +Given 3D position $\mathbf{x} \in \mathbb{R}^3$ and 2D viewing direction $\mathbf{d} \in \mathbb{R}^2$ , NeRF [18] aims to estimate volume density $\sigma \in \mathbb{R}$ and emitted RGB color $\mathbf{c} \in \mathbb{R}^3$ using neural networks. We formulate dynamic NeRF as a direct extension from 3D to 4D by adding time frame $t \in \mathbb{R}$ to inputs, i.e. + +$$ +(\mathbf {c}, \sigma) = f (\mathbf {v} (\mathbf {x}, t), \mathbf {d}). \tag {1} +$$ + +$f$ is implemented as fully connected neural networks and a space-time feature representation $\mathbf{v}$ can be MLP-based neural nets or explicit grid values as explained in Sec. 3.2. + +From the camera with origin $\mathbf{o}$ and ray direction $\mathbf{d}$ , the color of camera ray $\mathbf{r}(u) = \mathbf{o} + u\mathbf{d}$ at time frame $t$ is + +$$ +\mathbf {C} (\mathbf {r}, t) = \int_ {u _ {n}} ^ {u _ {f}} U (u, t) \sigma (\mathbf {v}) \mathbf {c} (\mathbf {v}, \mathbf {d}) d u, \tag {2} +$$ + +where $u_{n}$ and $u_{f}$ denote the bounds of the scene volume, $\mathbf{v}$ is an abbreviation of $\mathbf{v}(\mathbf{r}(u), t)$ and $U(u, t) = \exp(-\int_{u_n}^u \sigma(\mathbf{v}(\mathbf{r}(s), t)) ds)$ is the accumulated transmittance from $u_{n}$ to $u$ . Then, the RGB color loss $L_{c}$ is defined to minimize the $l_{2}$ -loss between the estimated colors $\hat{C}(\mathbf{r}, t)$ and the ground truth colors $C(\mathbf{r}, t)$ over all rays $\mathbf{r}$ in camera views $R$ and time frames $t \in T$ as follows: + +$$ +L _ {c} = \sum_ {\mathbf {r} \in R, t \in T} | | \hat {C} (\mathbf {r}, t) - C (\mathbf {r}, t) | | _ {2} ^ {2}. \tag {3} +$$ + +# 3.2. Space-Time Feature Representation + +The main contribution of our paper lies on the novel feature representation method for dynamic NeRFs. We define the feature vector $\mathbf{v}$ , which is fed into the neural network to determine a color and a volume density, as the concatenation of static and dynamic feature vectors, i.e., + +$$ +\mathbf {v} (\mathbf {x}, t) = \left[ \mathbf {v} _ {s} (\mathbf {x}), \mathbf {v} _ {d} (\mathbf {x}, t) \right], \tag {4} +$$ + +where $[\cdot, \cdot]$ means concatenation operator of two vectors. The static feature vector $\mathbf{v}_s$ only depends on the 3D position $\mathbf{x}$ . As most dynamic scenes also contain static regions such as background, $\mathbf{v}_s$ is designed to learn the static features that are consistent across time. Meanwhile, $\mathbf{v}_d$ learns dynamic features which may vary across time. We propose two novel feature representation methods in which temporal interpolation is applied. First, the neural representation, which is essentially a series of neural networks combined with linear interpolation, is suggested in Sec. 3.2.1. The model trained using the neural representation is able to render high-quality space-time view synthesis with small-sized neural networks ( $\sim 20\mathrm{MB}$ ). Second, the grid representation, which is an temporal extension of recently proposed voxel grid representations [19], is explained in Sec. 3.2.2. Dynamic NeRFs can be trained in less than 5 minutes with the proposed grid representation. + +# 3.2.1 Neural Representation + +In the neural representation, features that fed into the template NeRFs are determined by a series of neural nets. In other words, both static and dynamic feature extractor are formulated as multi-layer perceptrons (MLPs). First, the whole time frame is divided into equally spaced time slots. For each time slot, two MLPs are assigned. Then the whole feature vector is interpolated from the assigned MLPs. Overall feature extraction procedure is illustrated in Fig. 2. + +![](images/a88608fb7e439d02a916731ce76434e718f64fef4450211653bda669a61126a2.jpg) +Figure 2. Illustration of the arrangements of MLPs (left) and the feature extraction (right) in the neural representation. $\phi_s$ and $\phi_{ij}$ are small MLPs that extract static and dynamic features respectively. The static features $\mathbf{v}_s$ and the dynamic feature $\mathbf{v}_d$ are concatenated and fed into the template NeRF. Detailed structures of the template NeRF can be found in the supplementary materials. + +![](images/23ce09fdbe5b05a7ff892ae461ec6de8d67f636ebbf12472f387191cd5653cb1.jpg) + +Concretely, let $\mathbf{v}_d(\mathbf{x},\mathbf{z}_t)$ be the feature vector for a 3D point $\mathbf{x}$ at time $t$ , which will be fed to the template NeRF. Here, we used $\mathbf{z}_t$ as an embedding vector for an input time frame $t$ [12, 16, 22]. For equally spaced $n$ time slots, there are $n + 1$ keyframes $t_i = \frac{i}{n} (i = 0,1,2,\dots,n)$ . An MLP $\phi_{i}$ is assigned to each keyframe $t_i$ which is responsible for two adjacent time slots $[t_{i - 1},t_i]$ and $[t_i,t_{i + 1}]$ . For inputs with time $t\in [t_i,t_{i + 1}]$ , $\mathbf{x}$ and $\mathbf{z}_t$ are fed into $\phi_{i}$ and $\phi_{i + 1}$ . Then, the outputs from two MLPs are interpolated as follows: + +$$ +\mathbf {v} _ {d} (\mathbf {x}, \mathbf {z} _ {t}) = \Delta t \cdot \phi_ {i} (\mathbf {x}, \mathbf {z} _ {t}) + (1 - \Delta t) \cdot \phi_ {i + 1} (\mathbf {x}, \mathbf {z} _ {t}) \tag {5} +$$ + +where $\Delta t = (t_{i + 1} - t) / (t_{i + 1} - t_i)$ . + +The purpose of this interpolation is to efficiently learn the features between keyframes in a scalable manner. Thanks to the temporal interpolation, we can learn the features of the continuous time range $[t_i, t_{i+1}]$ by enforcing the MLPs $\phi_i$ and $\phi_{i+1}$ responsible for that time range. While the static feature $\mathbf{v}_s$ represents the features across the whole timeline, $\mathbf{v}_d$ represents the features that are focused more on dynamic regions. In addition, since each MLP for dynamic feature is responsible only for two adjacent time slots, it is able to make each $\phi_i$ learn features that are specific to a certain period of time. + +To exploit the features of both long-term and short-term, there can be multiple levels of dynamic features which have different number of keyframes. For multi-level dynamic feature extractor with level $l$ , each level contains different number of keyframes $n_1, n_2, \dots, n_l$ . Let $\mathbf{v}_{di}$ denote the dynamic feature of $i$ -th level, then the output dynamic feature is the concatenation of features from all levels, i.e., + +$$ +\mathbf {v} _ {d} (\mathbf {x}, \mathbf {z} _ {t}) = \left[ \mathbf {v} _ {d 1} (\mathbf {x}, \mathbf {z} _ {t}), \mathbf {v} _ {d 2} (\mathbf {x}, \mathbf {z} _ {t}), \dots , \mathbf {v} _ {d l} (\mathbf {x}, \mathbf {z} _ {t}) \right]. \tag {6} +$$ + +In this paper, we used the settings $l = 2$ , $n_1 = 5$ , $n_2 = 20$ otherwise stated. The dimensions of feature vectors are determined as 128 for $\mathbf{v}_s$ , and 64 for $\mathbf{v}_0$ and $\mathbf{v}_1$ . MLP with one hidden layer whose hidden size is the same as its output dimension is used for feature extraction of both $\mathbf{v}_s$ and $\mathbf{v}_{di}$ . + +The concatenated feature vectors are fed to the template NeRF which outputs volume density $\sigma$ and emitted color c. We used the same structure as used in HyperNeRF [22] for the template NeRF of the neural representation. + +# 3.2.2 Grid Representation + +Recently, InstantNGP [19] suggested a novel multi-level grid representation with hash tables. We adopt the hash grid representation from [19] and extend it for fast dynamic NeRF training. Similar to the neural representation in Sec. 3.2.1, the feature vector from the proposed hash grid contains static and dynamic feature vectors. The static feature $\mathbf{v}_s$ is analogous to the one used in [19]. On the other hand, the dynamic feature $\mathbf{v}_d$ comes from the 4D hash grids. + +Concretely, to extract static and dynamic feature vectors whose dimensions are $m_{s}$ and $m_{d}$ respectively, a hash table of size $H$ that contains $(m_{s} + m_{d})$ dimension feature vectors is constructed. The hash function of $d$ -dimensional vector, $h_d(\mathbf{x})$ , is defined as + +$$ +h _ {d} (\mathbf {x}) = \bigoplus_ {i = 1} ^ {d} \left(x _ {i} P _ {i}\right) \mod H, \tag {7} +$$ + +where $P_{i}$ is a large prime number and $\bigoplus$ is an XOR operator. Then, the feature vector $\mathbf{v}$ is retrieved by concatenating the outputs of 3D and 4D hash functions: + +$$ +\mathbf {v} (\mathbf {x}, t) = \left[ h _ {3} (\mathbf {x}), h _ {4} (\mathbf {x}, t) \right]. \tag {8} +$$ + +The 3D and 4D grids are constructed as the multi-level hash grids proposed in [19]. We applied different scaling factors for 3D space and time frame since the number of frames for training sequences are usually much smaller than the finest 3D grid resolutions. + +# 3.3. Smoothness Regularization + +As our dynamic world smoothly changes, it is reasonable to impose a smoothness term to adjacent time frames. + +![](images/309b7c6d5564d7c61da017fb3f3b8e61da15c0b3c49fff352e6d05e406c0202b.jpg) +Figure 3. Illustration of grid representation with hash tables at a certain level. A static feature vector of a 3D grid point is extracted via 3D hash function, while dynamic feature vector for 4D grid point is extracted via 4D hash. The static feature vector $\mathbf{v}_s$ is determined as trilinear interpolation of 8 vectors of grid points, while the dynamic feature vector $\mathbf{v}_d$ is calculated as quadrilinear interpolation of 16 vectors. In the figure, we only depicted feature retrieval process of two grid points for illustration purpose. + +We note that the smoothness term is only applied to the input feature space, not to the estimated outputs such as RGB color or density. For the neural representation, we provided a simple smoothness term which is calculated as, + +$$ +L _ {s} (\mathbf {x}, t) = \left\| \mathbf {v} _ {d} (\mathbf {x}, \mathbf {z} _ {t}) - \mathbf {v} _ {d} (\mathbf {x}, \mathbf {z} _ {t + 1}) \right\| _ {2} ^ {2}. \tag {9} +$$ + +This regularization term has two advantages. First, it reflects the intuition that the observed point $\mathbf{x}$ at time $t$ will be stationary if there is no observation for $\mathbf{x}$ at time $t + 1$ . Second, by propagating additional gradients to the feature networks or the hash grids, the smoothness term acts as a regularizer that stabilizes the training. + +For the grid representation, we impose a smoothness term to the grid points that are temporally adjacent: + +$$ +L _ {s} (\mathbf {x}, t) = \frac {1}{n _ {f} ^ {2}} \| h _ {4} (\mathbf {x}, t _ {a}) - h _ {4} (\mathbf {x}, t _ {b}) \| _ {2} ^ {2}, \tag {10} +$$ + +where $n_f$ is the number of frames in the training sequence and $t_a, t_b$ are two adjacent grid points that satisfies $t_a \leq t \leq t_b$ . In fact, Eq. (10) can be obtained from Eq. (9) in the grid representation, which indicates that imposing smoothness term to adjacent time frames is equivalent to add smoothness to two temporally adjacent grids with constant multiplier. Detailed derivation of this relationship is provided in the supplementary materials. + +The smoothness term is added to the loss function with a weight $\lambda$ , and the total loss is minimized together during the training. Accordingly, the total loss of our dynamic neural radiance fields is given by + +$$ +L = L _ {c} + \lambda L _ {s}. \tag {11} +$$ + +We observe that the smoothness term is especially powerful in the static background regions that appeared only in a small fraction of frames. In Fig. 4, we show examples of the rendered images which are trained with and without the smoothness term. It can be clearly seen that the boxed + +![](images/9fa34a09239bbae0bd572ed78019774be582e9f92744321a79540e65d4114698.jpg) +(a) w/o smooth term + +![](images/235e6988d1d7dd2add057a5bdd31f54f7e0e970a615fb90457c02ce72f629703.jpg) +(b) w/ smooth term + +![](images/bd4b726b530f6b0778fc7b3df8b52e1ec9598fc3a0d2d6f008260fe58acb3528.jpg) +(c) GT +Figure 4. Effectiveness of the smoothness term. It can be observed that the model trained with the smoothness term accurately renders the corner part which appears in only few frames within the whole sequence (blue box) and removes spurious artifacts (red box). + +regions of the rendered image shows much plausible when the model is trained with the smoothness term. + +# 3.4. Implementation Details + +In the neural representation, we adopted the template NeRF architecture of HyperNeRF [22] for density and color estimation. The network is an 8-layered MLP with hidden size of 256. The dimension of the embedding vector $\mathbf{z}_t$ is set to 8. The smoothness weight $\lambda$ is set to 0.01 or 0.001 depending on the characteristics of the datasets used in the experiments. We set $\lambda$ to a large value in the sequences that the viewpoint change is significant in a short period of time although performance variations depending on the value of $\lambda$ is not significant. + +In the grid representation, $\lambda$ is set to $1e^{-4}$ . We applied the smoothness loss only for the finest two levels of the temporal dimension since applying it to every level slows training speed without much improvements in performance. After feature vectors are extracted from the hash grids, we fed them to a 3-layered MLP with hidden size of 128 to estimate + +volume density followed by one additional layer for RGB color estimation similar to [19]. We set $m_{s} = 2$ , $m_{d} = 6$ and $H = 2^{19}$ for the grid representation experiments otherwise stated. The 3D and 4D hash grids are composed of 12 levels. For the spatial dimension, we set the base resolution to 8, and the scaling factor is set to 1.45. The base resolution of the temporal dimension is 2, which is multiplied by 1.4 in every other level. Detailed hyperparameter settings and network architectures can be found in the supplementary materials. + +# 4. Experimental Results + +# 4.1. Datasets + +To validate the superiority of our method, we conducted extensive experiments on various datasets. We used three publicly available datasets that are used for dynamic NeRFs. For all experiments, we trained our models for each sequence individually, and then per-scene results are averaged and reported in this section. + +D-NeRF Dataset [23]. The dataset consists of synthetically rendered images of moving and deforming 3D objects. For each frame, the image is rendered via a synthetic camera of random rotation. There are eight scenes in total, and each scene contains 50-200 training images with 20 test views. + +HyperNeRF Dataset [22]. The dataset contains video sequences taken from mobile phones. There are four sequences in vrig-dataset which are taken using the camera rig with two phones vertically aligned. In addition, there are six sequences in interp-dataset which are taken from a single camera in order to estimate the image in the middle of two consecutive training images. + +DyNeRF Dataset [12]. The dataset consists of videos obtained from a capture system that consists of 21 GoPro Black Hero cameras. The cameras are located at a fixed position, and all video frames are synchronized to build the multi-view video dataset. + +# 4.2. Neural Representation + +In this section, we reported the performance of the proposed neural representation models. First, the experimental results on D-NeRF dataset are shown in Tab. 1. Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS) [29] are used as evaluation metrics following the previous works. We also reported the average metric (AVG) proposed in [3] which aggregates three metrics to a single value. Our method with the neural representation (Ours-NN) achieves state-of-the-art results on all evaluation metrics. It is worth noting that the smoothness term dramatically improves overall performance. + +Next, we evaluated our method on HyperNeRF dataset [22] and compared with existing methods, which is shown in Tab. 2. Here, we reported PSNR and multi + +
PSNR ↑SSIM ↑LPIPS ↓AVG ↓
NeRF [18]19.000.870.180.09
T-NeRF [23]29.500.950.080.03
D-NeRF [23]30.430.950.070.02
NDVG-full [10]27.840.8620.0410.029
TiNeuVox-B [6]32.670.9710.0410.016
Ours-NN (w/o smooth)30.180.9630.0380.019
Ours-NN (w/ smooth)32.730.9740.0330.014
+ +Table 1. Experimental results on D-NeRF [23] datasets. + +scale SSIM (MS-SSIM), and excluded LPIPS metric since its value cannot reliably reproduced [6]. Our method shows second-best results on both vrig and interp datasets. While flow-based method [13] suffers from interpolating motions between two consecutive frames, our method, which implicitly learns intermediate feature representation of inbetween frames, achieves competitive performance with warping-based methods [21, 22]. + +Qualitative results for the neural representation are shown in Fig. 5. It is clearly observed that our method captures fine details compared to D-NeRF [23]. We highlighted the regions that show notable differences with colored boxes and placed zoomed-in images of the regions next to the rendered images. In addition, we also qualitatively compared our method with HyperNeRF [22] on their datasets. When the warping estimation module of [22] does not correctly estimate warping parameters, HyperNeRF produces implausible results. It can be observed that the head and the body of the chicken toy is not properly interlocked and the position of the hand peeling banana is incorrect. On the other hand, our method accurately recovers 3D geometry of those scenes. Thus, without using the physically meaningful warping estimation module in neural networks, the proposed temporal interpolation and the smoothness regularizer provides simple yet effective way to learn complex deformations. + +Lastly, performance evaluation on DyNeRF dataset [12] is presented in Tab. 3. We adopt PSNR, LPIPS, and FLIP [2] for evaluation metrics to compare with previous works. Our method achieves best PSNR while ranked second in LPIPS and FLIP. However, those two metrics are also better than DyNeRF† [12] which does not use importance sampling strategy in [12]. Since we do not use any sampling strategy during training, it can be concluded that our feature interpolation method is superior to the network architecture of [12]. Notably, our method outperforms DyNeRF with smaller network size (20MB) than the DyNeRF models (28MB). + +# 4.3. Grid Representation + +We used D-NeRF datasets to evaluate the performance of the proposed grid representation. Since this representation is mainly intended for fast training, we report the re + +![](images/cd18c36bc3f6ae14344bccf0c60f783edb6644c192d6101835ff1bb4fc154e96.jpg) +Figure 5. Qualitative results of our method (Ours-NN) on of D-NeRF [23] and HyperNeRF [22] datasets. The regions that show significant difference with compared methods are scaled up (red and green boxes). + +
vriginterp
PSNR↑MS-SSIM↑PSNR↑MS-SSIM↑
NeRF [18]20.130.74522.270.804
NV [15]16.850.57126.050.911
NSFF [13]26.330.91625.800.883
Nerfies [21]22.230.80328.470.939
HyperNeRF [22]22.380.81429.000.945
Ours-NN24.350.86728.670.940
+ +Table 2. Experimental results on HyperNeRF [22] datasets. + +
PSNR↑LPIPS ↓FLIP ↓
MVS19.120.25990.2542
NeuralVolumes [15]22.800.29510.2049
LLFF [17]23.240.23460.1867
NeRF-T [12]28.450.1000.1415
DyNeRF† [12]28.500.09850.1455
DyNeRF [12]29.580.08320.1347
Ours-NN29.880.09600.1413
+ +sults in a short period time ( $\sim 8$ minutes) and compare the results with the concurrent works [6, 10], both of which are based on voxel grid representation and showed the fastest training speed for dynamic NeRFs so far. We also examined the performance of the original implementation of InstantNGP [19] as a baseline with no temporal extension. All of the grid representation models in the experiments are trained on a single RTX 3090 GPU for fair comparison with [6, 10]. + +The quantitative results are shown with elapsed training time in Tab. 4. The grid representation demonstrates much faster training speed even compared to the recent voxel-grid based methods [6, 10]. In comparison with neural network models [23], our method exceeds SSIM and LPIPS of [23] in 5 minutes which have arithmetically 240 times faster training speed. By taking the benefits of fully fused neural network [20] as well as the efficient hash grid repre + +Table 3. Experimental results on DyNeRF [12] datasets. + +
Train timePSNRSSIMLPIPSAVG
InstantNGP [19]5 min20.280.8880.1460.077
D-NeRF [23]20 hours30.430.950.070.02
NDVG-full [10]35 min27.840.8620.0410.029
TinueVox-B [6]30 min32.670.9710.0410.016
NDVG-half [10]23 min27.150.8570.0480.033
TinueVox-S [6]8 min30.750.9560.0670.023
Ours-grid1 min26.770.9330.1070.039
Ours-grid5 min29.730.9610.0630.024
Ours-grid8 min29.840.9620.0620.023
+ +Table 4. Quantitative comparison of training time and performance on D-NeRF datasets using the grid representation. + +
D-NeRF datasetHyperNeRF vrig
PSNRSSIMLPIPSPSNRMS-SSIM
NN (dynamic only)30.910.9630.04324.070.861
NN (dynamic+static)32.730.9740.03324.350.866
Grid (dynamic only)29.080.9480.07622.110.755
Grid (dynamic+static)29.840.9620.06222.980.802
+ +Table 5. Ablation study on the effectiveness of static features. + +sentation, our method quickly learns meaningful features in dynamic scenes and generates acceptable novel view synthesis results with just one minute of training. Our method shows superior SSIM, LPIPS and slightly inferior PSNR to TinueVox-S [6], which indicates clearer and more detailed rendering results, when trained for the same period of time. + +We depict training progress of the grid representation qualitatively in Fig. 6. We also compared the results of TinueVox-S [6] which are trained for 8 minutes. After one minute of training, the grid representation model produces blurry images but accurately render dynamic regions compared to InstantNGP [19]. After 5 minutes, the model successfully renders sharp images that are similar to the ground truth images. Notably, our model exhibits more sharp and clear results than TinueVox-S despite shorter training time, which results in better SSIM and LPIPS.. + +Finally, we compared the performance of the grid repre + +![](images/52cd234492271ee32b58b128e4eca79d5dadbdc3930297731d6885878f44447a.jpg) +Figure 6. Qualitative results of the grid representation. While InstantNGP [19] is not able to reconstruct dynamic regions, our method produces acceptable rendering results in one minute and performs superior to TiNueVox-S [6] with shorter training time. + +![](images/2be207e18b621c54cd9a068a2f8376b3ee623e1be515aa151e0e62be6e107db3.jpg) +Figure 7. Performance comparison with different feature dimension settings. Using smaller feature dimensions results in faster training but quicker saturation of model performance. + +![](images/8d76829c607b7fcee9142d93845602e78203c5aee593bedc1beeb3eb81a7893e.jpg) + +sentation by varying the dimension of static and dynamic feature vectors, $m_{s}$ and $m_{d}$ . PSNR and LPIPS on test images are measured per minute and illustrated in Fig. 7. We trained the grid representation models in three different settings, $m_{s} = 2, m_{d} = 6$ , $m_{s} = 2, m_{d} = 2$ , $m_{s} = 4, m_{d} = 4$ . When smaller feature dimension is used, the training speed is faster, so PSNR increases fast in the early stage of training. However, the performance also saturates faster, and LPIPS is inferior to the other settings. The model with $m_{s} = 2, m_{d} = 6$ significantly outperformed the others in terms of LPIPS. + +# 4.4. Effectiveness of the Static Features + +To validate the effectiveness of the static features, we conducted an ablation study on the static features and illustrated the results in Tab. 5. For both the neural representation (NN) and the grid representation (Grid), the models with static features performs superior to the ones using only dynamic features in all metrics. Additional ablation studies and qualitative results can be found in the supplementary materials. + +# 4.5. Failure Cases + +Although the proposed feature interpolation is able to learn meaningful spatiotemporal features in most cases, + +![](images/19bb8c6b5be247b5e0af71038c78a774d2b58e7cbeabe3eca0e87d0f6a4b7386.jpg) +Figure 8. Examples of failure cases. + +there are a few failure cases as presented in Fig. 8. For instance, our method has difficulty in recovering 3D structures when small objects in a sequence rapidly move (Fig. 8 left) or when there exist dynamic regions that are not observed in the training sequence (Fig. 8 right). + +# 5. Conclusion + +In this paper, we propose a simple yet effective feature interpolation method for training dynamic NeRFs. Both the neural representation and the grid representation showed impressive performance, and the smoothness term applied to the intermediate feature vectors further improves the performance. Since these methods are unrelated to the existing methods of modeling deformations or estimating scene flows, we believe that the proposed method suggests a new direction of training dynamic NeRFs. + +While the neural representation model shows high-quality rendering results owing to the representation power of neural networks, it requires hours of training and seconds of rendering which impose barriers to real-time applications. On the other hand, the grid representation is able to render dynamic scenes in less than a second after a few minutes of training, which makes it more practical for real-world applications. Both representations are mutually complementary, and investigating hybrid representations that take advantages of both representations would be an interesting research direction. + +# References + +[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 2 +[2] Pontus Andersson, Jim Nilsson, Tomas Akenine-Möller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 6 +[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 2, pages 690-696. IEEE, 2000. 2 +[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2 +[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 2, 3, 6, 7, 8 +[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3 +[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2 +[10] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3, 6, 7 +[11] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 1 +[12] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 4, 6, 7 +[13] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6, 7 +[14] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 1 +[15] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019. 7 +[16] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 4 +[17] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 7 +[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 6, 7 +[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 3, 4, 6, 7, 8 +[20] Thomas Müller, Fabrice Rousselle, Jan Novák, and Alexander Keller. Real-time neural radiance caching for path tracing. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 7 +[21] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 6, 7 +[22] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 4,5,6,7 +[23] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2, 6, 7 +[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields + +reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3 +[25] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Nonrigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2 +[26] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 3 +[27] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2 +[28] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 3 +[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 \ No newline at end of file diff --git a/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/images.zip b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01e2b9f1f9fc738d3c31b2aee0a5b031e1906e50 --- /dev/null +++ b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca85455a97ba2ba6d2c1cfc44926c11afa3f065310b4769b3e589e0f2640ffbd +size 557055 diff --git a/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/layout.json b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18b0a501f3ba015b321d9be0d616cb957289444f --- /dev/null +++ b/2023/Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields/layout.json @@ -0,0 +1,8430 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 143, + 529, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 143, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 143, + 529, + 171 + ], + "type": "text", + "content": "Suntheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, Nahiup Kang Samsung Advanced Institute of Technology (SAIT)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 173, + 533, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 173, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 173, + 533, + 186 + ], + "type": "text", + "content": "{sh2019.park, minjungs.son, swan.jang, ychun.ahn, jiyeon31.kim, nahyup.kang}@samsung.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 225, + 527, + 417 + ], + "blocks": [ + { + "bbox": [ + 72, + 225, + 527, + 417 + ], + "lines": [ + { + "bbox": [ + 72, + 225, + 527, + 417 + ], + "spans": [ + { + "bbox": [ + 72, + 225, + 527, + 417 + ], + "type": "image", + "image_path": "7c9bfc4c36b86ff5abc585719ac7ba8dd5aec02a845824cc465d8911ab948ee0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 429, + 546, + 485 + ], + "lines": [ + { + "bbox": [ + 45, + 429, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 45, + 429, + 546, + 485 + ], + "type": "text", + "content": "Figure 1. We propose simple yet effective feature interpolation methods for training neural radiance fields of dynamic scenes based on temporal interpolation. We provide two different feature vector representations, neural representation (top) and grid representation (bottom), both of which are the concatenation of static feature vectors and temporally-interpolated dynamic feature vectors. The neural representation exhibits high-quality rendering performance with small-sized models while the grid representation shows competitive rendering results with astonishingly fast training speed." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 496, + 192, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 496, + 192, + 509 + ], + "spans": [ + { + "bbox": [ + 143, + 496, + 192, + 509 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "Temporal interpolation often plays a crucial role to learn meaningful representations in dynamic scenes. In this paper, we propose a novel method to train spatiotemporal neural radiance fields of dynamic scenes based on temporal interpolation of feature vectors. Two feature interpolation methods are suggested depending on underlying representations, neural networks or grids. In the neural representation, we extract features from space-time inputs via multiple neural network modules and interpolate them based on time frames. The proposed multi-level feature interpolation network effectively captures features of both short-term and long-term time ranges. In the grid representation, space-time features are learned via four-dimensional hash grids, which remarkably reduces training time. The grid representation shows more than " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " faster training speed than the previous neural-net-based methods while maintaining" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 581 + ], + "type": "text", + "content": "the rendering quality. Concatenating static and dynamic features and adding a simple smoothness term further improve the performance of our proposed models. Despite the simplicity of the model architectures, our method achieved state-of-the-art performance both in rendering quality for the neural representation and in training speed for the grid representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 608, + 387, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 608, + 387, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 608, + 387, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "3D reconstruction and photo-realistic rendering have been long-lasting problems in the fields of computer vision and graphics. Along with the advancements of deep learning, differentiable rendering [11, 14] or neural rendering, has emerged to bridge the gap between the two problems. Recently proposed Neural Radiance Field (NeRF) [18] has finally unleashed the era of neural rendering. Using NeRF," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4212" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "it is able to reconstruct accurate 3D structures from multiple 2D images and to synthesize photo-realistic images from unseen viewpoints. Tiny neural networks are sufficient to save and retrieve complex 3D scenes, which can be trained in a self-supervised manner given 2D images and camera parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 314 + ], + "type": "text", + "content": "Meanwhile, as our world typically involves dynamic changes, it is crucial to reconstruct the captured scene through 4D spatiotemporal space. Since it is often not possible to capture scenes at different viewpoints simultaneously, reconstructing dynamic scenes from images is inherently an under-constrained problem. While NeRF was originally designed to deal with only static scenes, there have been a few approaches in the literature that extend NeRF to dynamic scenes [13,22,23,25], which are so called as dynamic NeRFs. Inspired by the non-rigid structure-from-motion algorithms [1, 4] that reconstruct 3D structures of deformable objects, most previous works solved the under-constrained setting by estimating scene deformations [21,22,25] or 3D scene flows [5,9,12] for each frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 458 + ], + "type": "text", + "content": "However, since the parameters of deformation estimation modules are jointly optimized with NeRF network simultaneously, it is questionable that the modules can accurately estimate deformations or scene flows in accordance with its design principles. In many dynamic scenes, it is challenging to resolve the ambiguities whether a point was newly appeared, moved, or changed its color. It is expected that those ambiguities and deformation estimation can be separately solved within a single network, but in practice, it is hard to let the network implicitly learn the separation, especially for general dynamic scenes without any prior knowledge of deformation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "On the other hand, grid representations in NeRF training [7, 19, 24] have grabbed a lot of attentions mainly due to its fast training speed. Simple trilinear interpolation is enough to fill in the 3D space between grid points. While the representation can be directly adopted to dynamic NeRFs together with the warping estimation module [6], it still requires additional neural networks that affects training and rendering speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Motivated by the aforementioned analyses, we present a simple yet effective architecture for training dynamic NeRFs. The key idea in this paper is to apply feature interpolation to the temporal domain instead of using warping functions or 3D scene flows. While the feature interpolation in 2D or 3D space has been thoroughly studied, to the best of our knowledge, feature interpolation method in temporal domain for dynamic NeRF has not been proposed yet. We propose two multi-level feature interpolation methods depending on feature representation which is either neural nets or hash grids [19]. Overview of the two representations, namely the neural representation and the grid representation, are illustrated in Fig. 1. In addition," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "noting that 3D shapes deform smoothly over time in dynamic scenes, we additionally introduced a simple smoothness term that encourages feature similarity between the adjacent frames. We let the neural networks or the feature grids to learn meaningful representations implicitly without imposing any constraints or restrictions but the smoothness regularizer, which grants the flexibility to deal with various types of deformations. Extensive experiments on both synthetic and real-world datasets validate the effectiveness of the proposed method. We summarized the main contributions of the proposed method as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 213, + 545, + 361 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 213, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 213, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 545, + 273 + ], + "type": "text", + "content": "- We propose a simple yet effective feature extraction network that interpolates two feature vectors along the temporal axis. The proposed interpolation scheme outperforms existing methods without having a deformation or flow estimation module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 281, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 281, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 317, + 281, + 545, + 328 + ], + "type": "text", + "content": "- We integrate temporal interpolation into hash-grid representation [19], which remarkably accelerates training speed more than " + }, + { + "bbox": [ + 317, + 281, + 545, + 328 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 317, + 281, + 545, + 328 + ], + "type": "text", + "content": " faster compared to the neural network models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 336, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 336, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 317, + 336, + 545, + 361 + ], + "type": "text", + "content": "- We propose a smoothness regularizer which effectively improves the overall performance of dynamic NeRFs." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 391, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 438 + ], + "type": "text", + "content": "Dynamic NeRF. There have been various attempts to extend NeRF to dynamic scenes. Existing methods can be categorized into three types: warping-based, flow-based, and direct inference from space-time inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": "Warping-based methods learn how 3D structure of the scene is deformed. 3D radiance field for each frame is warped to single or multiple canonical frames using the estimated deformation. Deformation is parameterized as 3D translation [23], rotation and translation via angle-axis representation [21, 22], or weighted translation [25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": "On the other hand, flow-based methods estimate the correspondence of 3d points in consecutive frames. Neural scene flow fields [13] estimates 3D scene flow between the radiance fields of two time stamps. Xian et al. [27] suggested irradiance fields which spans over 4D space-time fields with a few constraints derived from prior knowledge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Although warping-based and flow-based approaches showed successful results, they have a few shortcomings. For instance, warping-based methods cannot deal with topological variations. Since those methods warp every input frame to a single canonical scene, it is hard to correctly represent newly appeared or disappeared 3D radiance fields in the canonical scene. HyperNeRF [22] learns hyperspace which represents multidimensional canonical shape bases. Neural scene flow fields [13] solve the problem by introducing occlusion masks which are used to estimate the regions where scene flows are not applicable. However, the masks" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4213" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "and the flow fields should be learned simultaneously during the training, which makes the training procedure complicated and imposes dependence on extra information such as monocular depth or optical flow estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 288, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 227 + ], + "type": "text", + "content": "Without estimating shape deformations or scene flows, DyNeRF [12] used a simple neural network to train NeRFs of dynamic scenes, although the experiments are only conducted on synchronized multi-view videos. It utilizes 3D position and time frame as inputs, and predicts color and density through a series of fully connected neural networks. Our neural representation applies temporal interpolation to intermediate features which enhances representation power for dynamic features while keeping the simple structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 539 + ], + "type": "text", + "content": "Category-specific NeRFs are one of the popular research directions for training dynamic objects such as humans [8, 26] or animals [28]. They are based on parametric models or use additional inputs such as pose information. Our method focuses on training NeRFs of general dynamic scenes without any prior knowledge about scenes or objects. Nevertheless, temporal interpolation proposed in this paper can be easily applied to the dynamic NeRFs with templates. Grid representations in NeRF. One of the major drawbacks of the original NeRF [18] is slow training speed. A few concurrent works are proposed to speed up NeRF training. Among them, grid-based representation such as Plenoxels [7], direct voxel grid optimization [24] show superior performance in terms of training time, where it takes only a few minutes to learn plausible radiance fields. Combining grid representation with a hash function further improves the efficiency and training speed of the feature grids [19]. Fang et al. [6] firstly applied a voxel grid representation to train dynamic NeRFs and achieved much faster training speed compared to the neural-network-based methods. Guo et al. [10] also proposed a deformable voxel grid method for fast training. Unlike aforementioned works, our grid representation does not estimate warping or deformation. Instead, it directly estimates color and density using the features obtained from 4D hash grids, which decreases the computational burden and enables faster training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 548, + 282, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 282, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 282, + 563 + ], + "type": "text", + "content": "3. Temporal Interpolation for Dynamic NeRF" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 568, + 133, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 133, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 133, + 580 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": "Given 3D position " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": " and 2D viewing direction " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": ", NeRF [18] aims to estimate volume density " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": " and emitted RGB color " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": " using neural networks. We formulate dynamic NeRF as a direct extension from 3D to 4D by adding time frame " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": " to inputs, i.e." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 656, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 656, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 120, + 656, + 287, + 669 + ], + "type": "interline_equation", + "content": "(\\mathbf {c}, \\sigma) = f (\\mathbf {v} (\\mathbf {x}, t), \\mathbf {d}). \\tag {1}", + "image_path": "26ced66b0b39368512cb338dee3c5f811b651ec1960cf2e7742c45f84fa0dbdc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " is implemented as fully connected neural networks and a space-time feature representation " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " can be MLP-based neural nets or explicit grid values as explained in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "From the camera with origin " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " and ray direction " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ", the color of camera ray " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(u) = \\mathbf{o} + u\\mathbf{d}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " at time frame " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 103, + 545, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 103, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 345, + 103, + 545, + 131 + ], + "type": "interline_equation", + "content": "\\mathbf {C} (\\mathbf {r}, t) = \\int_ {u _ {n}} ^ {u _ {f}} U (u, t) \\sigma (\\mathbf {v}) \\mathbf {c} (\\mathbf {v}, \\mathbf {d}) d u, \\tag {2}", + "image_path": "1e07eb531fc5f96a931da3197fa10daa18af988f10838b41a9bf9b34990e8b7c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "u_{n}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "u_{f}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " denote the bounds of the scene volume, " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " is an abbreviation of " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{v}(\\mathbf{r}(u), t)" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "U(u, t) = \\exp(-\\int_{u_n}^u \\sigma(\\mathbf{v}(\\mathbf{r}(s), t)) ds)" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " is the accumulated transmittance from " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "u_{n}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": ". Then, the RGB color loss " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " is defined to minimize the " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": "-loss between the estimated colors " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r}, t)" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " and the ground truth colors " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r}, t)" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " over all rays " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " in camera views " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " and time frames " + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 304, + 137, + 545, + 220 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 227, + 545, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 227, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 352, + 227, + 545, + 253 + ], + "type": "interline_equation", + "content": "L _ {c} = \\sum_ {\\mathbf {r} \\in R, t \\in T} | | \\hat {C} (\\mathbf {r}, t) - C (\\mathbf {r}, t) | | _ {2} ^ {2}. \\tag {3}", + "image_path": "c04246da55948241181ca0e43b456918e01270cf9534896f9169114f6ba725d4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 261, + 496, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 496, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 496, + 274 + ], + "type": "text", + "content": "3.2. Space-Time Feature Representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 280, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 340 + ], + "type": "text", + "content": "The main contribution of our paper lies on the novel feature representation method for dynamic NeRFs. We define the feature vector " + }, + { + "bbox": [ + 304, + 280, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 280, + 545, + 340 + ], + "type": "text", + "content": ", which is fed into the neural network to determine a color and a volume density, as the concatenation of static and dynamic feature vectors, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 348, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 348, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 368, + 348, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\mathbf {v} (\\mathbf {x}, t) = \\left[ \\mathbf {v} _ {s} (\\mathbf {x}), \\mathbf {v} _ {d} (\\mathbf {x}, t) \\right], \\tag {4}", + "image_path": "217420d64c28355dba10857f65f01cf91fd6b6099ad2e4305dc65b6fc63496f1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": " means concatenation operator of two vectors. The static feature vector " + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": " only depends on the 3D position " + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": ". As most dynamic scenes also contain static regions such as background, " + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": " is designed to learn the static features that are consistent across time. Meanwhile, " + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_d" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": " learns dynamic features which may vary across time. We propose two novel feature representation methods in which temporal interpolation is applied. First, the neural representation, which is essentially a series of neural networks combined with linear interpolation, is suggested in Sec. 3.2.1. The model trained using the neural representation is able to render high-quality space-time view synthesis with small-sized neural networks (" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{MB}" + }, + { + "bbox": [ + 304, + 369, + 545, + 573 + ], + "type": "text", + "content": "). Second, the grid representation, which is an temporal extension of recently proposed voxel grid representations [19], is explained in Sec. 3.2.2. Dynamic NeRFs can be trained in less than 5 minutes with the proposed grid representation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 586, + 435, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 435, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 435, + 599 + ], + "type": "text", + "content": "3.2.1 Neural Representation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In the neural representation, features that fed into the template NeRFs are determined by a series of neural nets. In other words, both static and dynamic feature extractor are formulated as multi-layer perceptrons (MLPs). First, the whole time frame is divided into equally spaced time slots. For each time slot, two MLPs are assigned. Then the whole feature vector is interpolated from the assigned MLPs. Overall feature extraction procedure is illustrated in Fig. 2." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4214" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 285, + 190 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 285, + 190 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 285, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 285, + 190 + ], + "type": "image", + "image_path": "a88608fb7e439d02a916731ce76434e718f64fef4450211653bda669a61126a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "Figure 2. Illustration of the arrangements of MLPs (left) and the feature extraction (right) in the neural representation. " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\phi_s" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\phi_{ij}" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " are small MLPs that extract static and dynamic features respectively. The static features " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " and the dynamic feature " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_d" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " are concatenated and fed into the template NeRF. Detailed structures of the template NeRF can be found in the supplementary materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 526, + 190 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 526, + 190 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 526, + 190 + ], + "type": "image", + "image_path": "23ce09fdbe5b05a7ff892ae461ec6de8d67f636ebbf12472f387191cd5653cb1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": "Concretely, let " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_d(\\mathbf{x},\\mathbf{z}_t)" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " be the feature vector for a 3D point " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": ", which will be fed to the template NeRF. Here, we used " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " as an embedding vector for an input time frame " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " [12, 16, 22]. For equally spaced " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " time slots, there are " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " keyframes " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "t_i = \\frac{i}{n} (i = 0,1,2,\\dots,n)" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": ". An MLP " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " is assigned to each keyframe " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " which is responsible for two adjacent time slots " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "[t_{i - 1},t_i]" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "[t_i,t_{i + 1}]" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": ". For inputs with time " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "t\\in [t_i,t_{i + 1}]" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\phi_{i + 1}" + }, + { + "bbox": [ + 46, + 252, + 289, + 361 + ], + "type": "text", + "content": ". Then, the outputs from two MLPs are interpolated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 287, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 287, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) = \\Delta t \\cdot \\phi_ {i} (\\mathbf {x}, \\mathbf {z} _ {t}) + (1 - \\Delta t) \\cdot \\phi_ {i + 1} (\\mathbf {x}, \\mathbf {z} _ {t}) \\tag {5}", + "image_path": "fbc6f545f304a910d91a908341ecdbb8841a0daef23450b332008638d97d939b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 393, + 194, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 194, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 194, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 393, + 194, + 405 + ], + "type": "inline_equation", + "content": "\\Delta t = (t_{i + 1} - t) / (t_{i + 1} - t_i)" + }, + { + "bbox": [ + 47, + 393, + 194, + 405 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": "The purpose of this interpolation is to efficiently learn the features between keyframes in a scalable manner. Thanks to the temporal interpolation, we can learn the features of the continuous time range " + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "inline_equation", + "content": "[t_i, t_{i+1}]" + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": " by enforcing the MLPs " + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\phi_{i+1}" + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": " responsible for that time range. While the static feature " + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": " represents the features across the whole timeline, " + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_d" + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": " represents the features that are focused more on dynamic regions. In addition, since each MLP for dynamic feature is responsible only for two adjacent time slots, it is able to make each " + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\phi_i" + }, + { + "bbox": [ + 46, + 406, + 287, + 536 + ], + "type": "text", + "content": " learn features that are specific to a certain period of time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "text", + "content": "To exploit the features of both long-term and short-term, there can be multiple levels of dynamic features which have different number of keyframes. For multi-level dynamic feature extractor with level " + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "text", + "content": ", each level contains different number of keyframes " + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "inline_equation", + "content": "n_1, n_2, \\dots, n_l" + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{di}" + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "text", + "content": " denote the dynamic feature of " + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 537, + 287, + 620 + ], + "type": "text", + "content": "-th level, then the output dynamic feature is the concatenation of features from all levels, i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 631, + 287, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 631, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 54, + 631, + 287, + 644 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) = \\left[ \\mathbf {v} _ {d 1} (\\mathbf {x}, \\mathbf {z} _ {t}), \\mathbf {v} _ {d 2} (\\mathbf {x}, \\mathbf {z} _ {t}), \\dots , \\mathbf {v} _ {d l} (\\mathbf {x}, \\mathbf {z} _ {t}) \\right]. \\tag {6}", + "image_path": "379e3e02ae97ea313655249646bf27683d0e8c1f2007a29c44d7b4af321de0ed.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": "In this paper, we used the settings " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "l = 2" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "n_1 = 5" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "n_2 = 20" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": " otherwise stated. The dimensions of feature vectors are determined as 128 for " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": ", and 64 for " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_1" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": ". MLP with one hidden layer whose hidden size is the same as its output dimension is used for feature extraction of both " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{di}" + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 252, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 301 + ], + "type": "text", + "content": "The concatenated feature vectors are fed to the template NeRF which outputs volume density " + }, + { + "bbox": [ + 304, + 252, + 547, + 301 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 252, + 547, + 301 + ], + "type": "text", + "content": " and emitted color c. We used the same structure as used in HyperNeRF [22] for the template NeRF of the neural representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 315, + 428, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 428, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 428, + 327 + ], + "type": "text", + "content": "3.2.2 Grid Representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "type": "text", + "content": "Recently, InstantNGP [19] suggested a novel multi-level grid representation with hash tables. We adopt the hash grid representation from [19] and extend it for fast dynamic NeRF training. Similar to the neural representation in Sec. 3.2.1, the feature vector from the proposed hash grid contains static and dynamic feature vectors. The static feature " + }, + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "type": "text", + "content": " is analogous to the one used in [19]. On the other hand, the dynamic feature " + }, + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_d" + }, + { + "bbox": [ + 304, + 333, + 545, + 430 + ], + "type": "text", + "content": " comes from the 4D hash grids." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": "Concretely, to extract static and dynamic feature vectors whose dimensions are " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "m_{s}" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "m_{d}" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " respectively, a hash table of size " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " that contains " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "(m_{s} + m_{d})" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": " dimension feature vectors is constructed. The hash function of " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": "-dimensional vector, " + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "inline_equation", + "content": "h_d(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": ", is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 498, + 545, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 498, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 362, + 498, + 545, + 530 + ], + "type": "interline_equation", + "content": "h _ {d} (\\mathbf {x}) = \\bigoplus_ {i = 1} ^ {d} \\left(x _ {i} P _ {i}\\right) \\mod H, \\tag {7}", + "image_path": "3eceed2540302476cc296aa385633db6e8a5e735d881218456cc20423043ae30.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "text", + "content": " is a large prime number and " + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\bigoplus" + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "text", + "content": " is an XOR operator. Then, the feature vector " + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 538, + 545, + 574 + ], + "type": "text", + "content": " is retrieved by concatenating the outputs of 3D and 4D hash functions:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 369, + 582, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 582, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 369, + 582, + 545, + 597 + ], + "type": "interline_equation", + "content": "\\mathbf {v} (\\mathbf {x}, t) = \\left[ h _ {3} (\\mathbf {x}), h _ {4} (\\mathbf {x}, t) \\right]. \\tag {8}", + "image_path": "c05da5491ea274d9fc798f4e21973ab48242dc1a2d29a6c3d62afe4f6721ab30.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 604, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 664 + ], + "type": "text", + "content": "The 3D and 4D grids are constructed as the multi-level hash grids proposed in [19]. We applied different scaling factors for 3D space and time frame since the number of frames for training sequences are usually much smaller than the finest 3D grid resolutions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 456, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 456, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 456, + 684 + ], + "type": "text", + "content": "3.3. Smoothness Regularization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "As our dynamic world smoothly changes, it is reasonable to impose a smoothness term to adjacent time frames." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4215" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 521, + 188 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 521, + 188 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 521, + 188 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 521, + 188 + ], + "type": "image", + "image_path": "309b7c6d5564d7c61da017fb3f3b8e61da15c0b3c49fff352e6d05e406c0202b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": "Figure 3. Illustration of grid representation with hash tables at a certain level. A static feature vector of a 3D grid point is extracted via 3D hash function, while dynamic feature vector for 4D grid point is extracted via 4D hash. The static feature vector " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " is determined as trilinear interpolation of 8 vectors of grid points, while the dynamic feature vector " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_d" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " is calculated as quadrilinear interpolation of 16 vectors. In the figure, we only depicted feature retrieval process of two grid points for illustration purpose." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 288, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 315 + ], + "type": "text", + "content": "We note that the smoothness term is only applied to the input feature space, not to the estimated outputs such as RGB color or density. For the neural representation, we provided a simple smoothness term which is calculated as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 322, + 287, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 322, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 85, + 322, + 287, + 337 + ], + "type": "interline_equation", + "content": "L _ {s} (\\mathbf {x}, t) = \\left\\| \\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t}) - \\mathbf {v} _ {d} (\\mathbf {x}, \\mathbf {z} _ {t + 1}) \\right\\| _ {2} ^ {2}. \\tag {9}", + "image_path": "8733509ced9962676303c06649ce8bd24c7301f5d8ed54821a99d72e80ce7b94.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": "This regularization term has two advantages. First, it reflects the intuition that the observed point " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": " will be stationary if there is no observation for " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 344, + 287, + 415 + ], + "type": "text", + "content": ". Second, by propagating additional gradients to the feature networks or the hash grids, the smoothness term acts as a regularizer that stabilizes the training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 440 + ], + "type": "text", + "content": "For the grid representation, we impose a smoothness term to the grid points that are temporally adjacent:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 446, + 287, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 446, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 85, + 446, + 287, + 474 + ], + "type": "interline_equation", + "content": "L _ {s} (\\mathbf {x}, t) = \\frac {1}{n _ {f} ^ {2}} \\| h _ {4} (\\mathbf {x}, t _ {a}) - h _ {4} (\\mathbf {x}, t _ {b}) \\| _ {2} ^ {2}, \\tag {10}", + "image_path": "307e569c241740103a3eb443219e5cf423b0cb942332a4ddb163c3a89cb6f01c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "text", + "content": " is the number of frames in the training sequence and " + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "inline_equation", + "content": "t_a, t_b" + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "text", + "content": " are two adjacent grid points that satisfies " + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "inline_equation", + "content": "t_a \\leq t \\leq t_b" + }, + { + "bbox": [ + 46, + 480, + 287, + 576 + ], + "type": "text", + "content": ". In fact, Eq. (10) can be obtained from Eq. (9) in the grid representation, which indicates that imposing smoothness term to adjacent time frames is equivalent to add smoothness to two temporally adjacent grids with constant multiplier. Detailed derivation of this relationship is provided in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": "The smoothness term is added to the loss function with a weight " + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 576, + 287, + 624 + ], + "type": "text", + "content": ", and the total loss is minimized together during the training. Accordingly, the total loss of our dynamic neural radiance fields is given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 135, + 633, + 287, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 633, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 135, + 633, + 287, + 645 + ], + "type": "interline_equation", + "content": "L = L _ {c} + \\lambda L _ {s}. \\tag {11}", + "image_path": "60c4183689f370fdcd0264769adab18480d70ded5059f4b434b64c46fb3232f9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "We observe that the smoothness term is especially powerful in the static background regions that appeared only in a small fraction of frames. In Fig. 4, we show examples of the rendered images which are trained with and without the smoothness term. It can be clearly seen that the boxed" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 264, + 382, + 392 + ], + "blocks": [ + { + "bbox": [ + 310, + 264, + 382, + 392 + ], + "lines": [ + { + "bbox": [ + 310, + 264, + 382, + 392 + ], + "spans": [ + { + "bbox": [ + 310, + 264, + 382, + 392 + ], + "type": "image", + "image_path": "9fa34a09239bbae0bd572ed78019774be582e9f92744321a79540e65d4114698.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 393, + 380, + 401 + ], + "lines": [ + { + "bbox": [ + 315, + 393, + 380, + 401 + ], + "spans": [ + { + "bbox": [ + 315, + 393, + 380, + 401 + ], + "type": "text", + "content": "(a) w/o smooth term" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 389, + 264, + 460, + 391 + ], + "blocks": [ + { + "bbox": [ + 389, + 264, + 460, + 391 + ], + "lines": [ + { + "bbox": [ + 389, + 264, + 460, + 391 + ], + "spans": [ + { + "bbox": [ + 389, + 264, + 460, + 391 + ], + "type": "image", + "image_path": "235e6988d1d7dd2add057a5bdd31f54f7e0e970a615fb90457c02ce72f629703.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 392, + 457, + 401 + ], + "lines": [ + { + "bbox": [ + 395, + 392, + 457, + 401 + ], + "spans": [ + { + "bbox": [ + 395, + 392, + 457, + 401 + ], + "type": "text", + "content": "(b) w/ smooth term" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 467, + 264, + 539, + 391 + ], + "blocks": [ + { + "bbox": [ + 467, + 264, + 539, + 391 + ], + "lines": [ + { + "bbox": [ + 467, + 264, + 539, + 391 + ], + "spans": [ + { + "bbox": [ + 467, + 264, + 539, + 391 + ], + "type": "image", + "image_path": "bd4b726b530f6b0778fc7b3df8b52e1ec9598fc3a0d2d6f008260fe58acb3528.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 392, + 517, + 401 + ], + "lines": [ + { + "bbox": [ + 493, + 392, + 517, + 401 + ], + "spans": [ + { + "bbox": [ + 493, + 392, + 517, + 401 + ], + "type": "text", + "content": "(c) GT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 411, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 456 + ], + "type": "text", + "content": "Figure 4. Effectiveness of the smoothness term. It can be observed that the model trained with the smoothness term accurately renders the corner part which appears in only few frames within the whole sequence (blue box) and removes spurious artifacts (red box)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 472, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 496 + ], + "type": "text", + "content": "regions of the rendered image shows much plausible when the model is trained with the smoothness term." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 503, + 440, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 440, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 440, + 516 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": "In the neural representation, we adopted the template NeRF architecture of HyperNeRF [22] for density and color estimation. The network is an 8-layered MLP with hidden size of 256. The dimension of the embedding vector " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": " is set to 8. The smoothness weight " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": " is set to 0.01 or 0.001 depending on the characteristics of the datasets used in the experiments. We set " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": " to a large value in the sequences that the viewpoint change is significant in a short period of time although performance variations depending on the value of " + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 522, + 545, + 641 + ], + "type": "text", + "content": " is not significant." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "In the grid representation, " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ". We applied the smoothness loss only for the finest two levels of the temporal dimension since applying it to every level slows training speed without much improvements in performance. After feature vectors are extracted from the hash grids, we fed them to a 3-layered MLP with hidden size of 128 to estimate" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4216" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "volume density followed by one additional layer for RGB color estimation similar to [19]. We set " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "m_{s} = 2" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "m_{d} = 6" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "inline_equation", + "content": "H = 2^{19}" + }, + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": " for the grid representation experiments otherwise stated. The 3D and 4D hash grids are composed of 12 levels. For the spatial dimension, we set the base resolution to 8, and the scaling factor is set to 1.45. The base resolution of the temporal dimension is 2, which is multiplied by 1.4 in every other level. Detailed hyperparameter settings and network architectures can be found in the supplementary materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 204, + 173, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 173, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 173, + 217 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 224, + 110, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 110, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 110, + 235 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 242, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 314 + ], + "type": "text", + "content": "To validate the superiority of our method, we conducted extensive experiments on various datasets. We used three publicly available datasets that are used for dynamic NeRFs. For all experiments, we trained our models for each sequence individually, and then per-scene results are averaged and reported in this section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 314, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 374 + ], + "type": "text", + "content": "D-NeRF Dataset [23]. The dataset consists of synthetically rendered images of moving and deforming 3D objects. For each frame, the image is rendered via a synthetic camera of random rotation. There are eight scenes in total, and each scene contains 50-200 training images with 20 test views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 374, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 458 + ], + "type": "text", + "content": "HyperNeRF Dataset [22]. The dataset contains video sequences taken from mobile phones. There are four sequences in vrig-dataset which are taken using the camera rig with two phones vertically aligned. In addition, there are six sequences in interp-dataset which are taken from a single camera in order to estimate the image in the middle of two consecutive training images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 459, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 517 + ], + "type": "text", + "content": "DyNeRF Dataset [12]. The dataset consists of videos obtained from a capture system that consists of 21 GoPro Black Hero cameras. The cameras are located at a fixed position, and all video frames are synchronized to build the multi-view video dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 527, + 176, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 176, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 176, + 540 + ], + "type": "text", + "content": "4.2. Neural Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": "In this section, we reported the performance of the proposed neural representation models. First, the experimental results on D-NeRF dataset are shown in Tab. 1. Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS) [29] are used as evaluation metrics following the previous works. We also reported the average metric (AVG) proposed in [3] which aggregates three metrics to a single value. Our method with the neural representation (Ours-NN) achieves state-of-the-art results on all evaluation metrics. It is worth noting that the smoothness term dramatically improves overall performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Next, we evaluated our method on HyperNeRF dataset [22] and compared with existing methods, which is shown in Tab. 2. Here, we reported PSNR and multi" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 69, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 541, + 153 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓AVG ↓
NeRF [18]19.000.870.180.09
T-NeRF [23]29.500.950.080.03
D-NeRF [23]30.430.950.070.02
NDVG-full [10]27.840.8620.0410.029
TiNeuVox-B [6]32.670.9710.0410.016
Ours-NN (w/o smooth)30.180.9630.0380.019
Ours-NN (w/ smooth)32.730.9740.0330.014
", + "image_path": "a4b54537e35643441f1b2f3c89aeb296f635b373396e1108cb66e739a17615d9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 161, + 527, + 172 + ], + "lines": [ + { + "bbox": [ + 324, + 161, + 527, + 172 + ], + "spans": [ + { + "bbox": [ + 324, + 161, + 527, + 172 + ], + "type": "text", + "content": "Table 1. Experimental results on D-NeRF [23] datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 190, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 286 + ], + "type": "text", + "content": "scale SSIM (MS-SSIM), and excluded LPIPS metric since its value cannot reliably reproduced [6]. Our method shows second-best results on both vrig and interp datasets. While flow-based method [13] suffers from interpolating motions between two consecutive frames, our method, which implicitly learns intermediate feature representation of inbetween frames, achieves competitive performance with warping-based methods [21, 22]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 287, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 502 + ], + "type": "text", + "content": "Qualitative results for the neural representation are shown in Fig. 5. It is clearly observed that our method captures fine details compared to D-NeRF [23]. We highlighted the regions that show notable differences with colored boxes and placed zoomed-in images of the regions next to the rendered images. In addition, we also qualitatively compared our method with HyperNeRF [22] on their datasets. When the warping estimation module of [22] does not correctly estimate warping parameters, HyperNeRF produces implausible results. It can be observed that the head and the body of the chicken toy is not properly interlocked and the position of the hand peeling banana is incorrect. On the other hand, our method accurately recovers 3D geometry of those scenes. Thus, without using the physically meaningful warping estimation module in neural networks, the proposed temporal interpolation and the smoothness regularizer provides simple yet effective way to learn complex deformations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 647 + ], + "type": "text", + "content": "Lastly, performance evaluation on DyNeRF dataset [12] is presented in Tab. 3. We adopt PSNR, LPIPS, and FLIP [2] for evaluation metrics to compare with previous works. Our method achieves best PSNR while ranked second in LPIPS and FLIP. However, those two metrics are also better than DyNeRF† [12] which does not use importance sampling strategy in [12]. Since we do not use any sampling strategy during training, it can be concluded that our feature interpolation method is superior to the network architecture of [12]. Notably, our method outperforms DyNeRF with smaller network size (20MB) than the DyNeRF models (28MB)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 658, + 424, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 424, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 424, + 670 + ], + "type": "text", + "content": "4.3. Grid Representation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We used D-NeRF datasets to evaluate the performance of the proposed grid representation. Since this representation is mainly intended for fast training, we report the re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4217" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 68, + 547, + 223 + ], + "blocks": [ + { + "bbox": [ + 53, + 68, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 53, + 68, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 68, + 547, + 223 + ], + "type": "image", + "image_path": "cd18c36bc3f6ae14344bccf0c60f783edb6644c192d6101835ff1bb4fc154e96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 546, + 251 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our method (Ours-NN) on of D-NeRF [23] and HyperNeRF [22] datasets. The regions that show significant difference with compared methods are scaled up (red and green boxes)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 270, + 282, + 350 + ], + "blocks": [ + { + "bbox": [ + 52, + 270, + 282, + 350 + ], + "lines": [ + { + "bbox": [ + 52, + 270, + 282, + 350 + ], + "spans": [ + { + "bbox": [ + 52, + 270, + 282, + 350 + ], + "type": "table", + "html": "
vriginterp
PSNR↑MS-SSIM↑PSNR↑MS-SSIM↑
NeRF [18]20.130.74522.270.804
NV [15]16.850.57126.050.911
NSFF [13]26.330.91625.800.883
Nerfies [21]22.230.80328.470.939
HyperNeRF [22]22.380.81429.000.945
Ours-NN24.350.86728.670.940
", + "image_path": "8874f3b1b5f04a5782eb8d839688a0b4ed6443af9c6bf25d750719eb44dd7c29.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 65, + 381, + 269, + 472 + ], + "blocks": [ + { + "bbox": [ + 59, + 358, + 274, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 358, + 274, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 358, + 274, + 369 + ], + "type": "text", + "content": "Table 2. Experimental results on HyperNeRF [22] datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 381, + 269, + 472 + ], + "lines": [ + { + "bbox": [ + 65, + 381, + 269, + 472 + ], + "spans": [ + { + "bbox": [ + 65, + 381, + 269, + 472 + ], + "type": "table", + "html": "
PSNR↑LPIPS ↓FLIP ↓
MVS19.120.25990.2542
NeuralVolumes [15]22.800.29510.2049
LLFF [17]23.240.23460.1867
NeRF-T [12]28.450.1000.1415
DyNeRF† [12]28.500.09850.1455
DyNeRF [12]29.580.08320.1347
Ours-NN29.880.09600.1413
", + "image_path": "7e67ae5c34aeb0004155ca1c7e88c916e15b155050243739d422b70500e6b840.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 616 + ], + "type": "text", + "content": "sults in a short period time (" + }, + { + "bbox": [ + 46, + 509, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\sim 8" + }, + { + "bbox": [ + 46, + 509, + 286, + 616 + ], + "type": "text", + "content": " minutes) and compare the results with the concurrent works [6, 10], both of which are based on voxel grid representation and showed the fastest training speed for dynamic NeRFs so far. We also examined the performance of the original implementation of InstantNGP [19] as a baseline with no temporal extension. All of the grid representation models in the experiments are trained on a single RTX 3090 GPU for fair comparison with [6, 10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "The quantitative results are shown with elapsed training time in Tab. 4. The grid representation demonstrates much faster training speed even compared to the recent voxel-grid based methods [6, 10]. In comparison with neural network models [23], our method exceeds SSIM and LPIPS of [23] in 5 minutes which have arithmetically 240 times faster training speed. By taking the benefits of fully fused neural network [20] as well as the efficient hash grid repre" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 270, + 542, + 373 + ], + "blocks": [ + { + "bbox": [ + 64, + 481, + 269, + 492 + ], + "lines": [ + { + "bbox": [ + 64, + 481, + 269, + 492 + ], + "spans": [ + { + "bbox": [ + 64, + 481, + 269, + 492 + ], + "type": "text", + "content": "Table 3. Experimental results on DyNeRF [12] datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 270, + 542, + 373 + ], + "lines": [ + { + "bbox": [ + 309, + 270, + 542, + 373 + ], + "spans": [ + { + "bbox": [ + 309, + 270, + 542, + 373 + ], + "type": "table", + "html": "
Train timePSNRSSIMLPIPSAVG
InstantNGP [19]5 min20.280.8880.1460.077
D-NeRF [23]20 hours30.430.950.070.02
NDVG-full [10]35 min27.840.8620.0410.029
TinueVox-B [6]30 min32.670.9710.0410.016
NDVG-half [10]23 min27.150.8570.0480.033
TinueVox-S [6]8 min30.750.9560.0670.023
Ours-grid1 min26.770.9330.1070.039
Ours-grid5 min29.730.9610.0630.024
Ours-grid8 min29.840.9620.0620.023
", + "image_path": "f58f82202ffb2df305e1cc42c46dcf5fba478ad3c402d5738b963c8da91d9086.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 412, + 539, + 471 + ], + "blocks": [ + { + "bbox": [ + 306, + 382, + 544, + 404 + ], + "lines": [ + { + "bbox": [ + 306, + 382, + 544, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 544, + 404 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of training time and performance on D-NeRF datasets using the grid representation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 412, + 539, + 471 + ], + "lines": [ + { + "bbox": [ + 309, + 412, + 539, + 471 + ], + "spans": [ + { + "bbox": [ + 309, + 412, + 539, + 471 + ], + "type": "table", + "html": "
D-NeRF datasetHyperNeRF vrig
PSNRSSIMLPIPSPSNRMS-SSIM
NN (dynamic only)30.910.9630.04324.070.861
NN (dynamic+static)32.730.9740.03324.350.866
Grid (dynamic only)29.080.9480.07622.110.755
Grid (dynamic+static)29.840.9620.06222.980.802
", + "image_path": "909dc8d9942fc307d5141842b08d46a6ad8817011fb290fc1e832d2b82d29bb8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 537, + 489 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 537, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 537, + 489 + ], + "type": "text", + "content": "Table 5. Ablation study on the effectiveness of static features." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 508, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 580 + ], + "type": "text", + "content": "sentation, our method quickly learns meaningful features in dynamic scenes and generates acceptable novel view synthesis results with just one minute of training. Our method shows superior SSIM, LPIPS and slightly inferior PSNR to TinueVox-S [6], which indicates clearer and more detailed rendering results, when trained for the same period of time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 581, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 699 + ], + "type": "text", + "content": "We depict training progress of the grid representation qualitatively in Fig. 6. We also compared the results of TinueVox-S [6] which are trained for 8 minutes. After one minute of training, the grid representation model produces blurry images but accurately render dynamic regions compared to InstantNGP [19]. After 5 minutes, the model successfully renders sharp images that are similar to the ground truth images. Notably, our model exhibits more sharp and clear results than TinueVox-S despite shorter training time, which results in better SSIM and LPIPS.." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 544, + 713 + ], + "type": "text", + "content": "Finally, we compared the performance of the grid repre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4218" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 534, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 534, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 534, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 534, + 220 + ], + "type": "image", + "image_path": "52cd234492271ee32b58b128e4eca79d5dadbdc3930297731d6885878f44447a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 546, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 546, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 546, + 247 + ], + "type": "text", + "content": "Figure 6. Qualitative results of the grid representation. While InstantNGP [19] is not able to reconstruct dynamic regions, our method produces acceptable rendering results in one minute and performs superior to TiNueVox-S [6] with shorter training time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 269, + 156, + 347 + ], + "blocks": [ + { + "bbox": [ + 51, + 269, + 156, + 347 + ], + "lines": [ + { + "bbox": [ + 51, + 269, + 156, + 347 + ], + "spans": [ + { + "bbox": [ + 51, + 269, + 156, + 347 + ], + "type": "image", + "image_path": "2be207e18b621c54cd9a068a2f8376b3ee623e1be515aa151e0e62be6e107db3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 354, + 288, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 389 + ], + "type": "text", + "content": "Figure 7. Performance comparison with different feature dimension settings. Using smaller feature dimensions results in faster training but quicker saturation of model performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 269, + 276, + 346 + ], + "blocks": [ + { + "bbox": [ + 170, + 269, + 276, + 346 + ], + "lines": [ + { + "bbox": [ + 170, + 269, + 276, + 346 + ], + "spans": [ + { + "bbox": [ + 170, + 269, + 276, + 346 + ], + "type": "image", + "image_path": "8d76829c607b7fcee9142d93845602e78203c5aee593bedc1beeb3eb81a7893e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": "sentation by varying the dimension of static and dynamic feature vectors, " + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "inline_equation", + "content": "m_{s}" + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "inline_equation", + "content": "m_{d}" + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": ". PSNR and LPIPS on test images are measured per minute and illustrated in Fig. 7. We trained the grid representation models in three different settings, " + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "inline_equation", + "content": "m_{s} = 2, m_{d} = 6" + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "inline_equation", + "content": "m_{s} = 2, m_{d} = 2" + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "inline_equation", + "content": "m_{s} = 4, m_{d} = 4" + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": ". When smaller feature dimension is used, the training speed is faster, so PSNR increases fast in the early stage of training. However, the performance also saturates faster, and LPIPS is inferior to the other settings. The model with " + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "inline_equation", + "content": "m_{s} = 2, m_{d} = 6" + }, + { + "bbox": [ + 46, + 405, + 287, + 536 + ], + "type": "text", + "content": " significantly outperformed the others in terms of LPIPS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 546, + 232, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 232, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 232, + 558 + ], + "type": "text", + "content": "4.4. Effectiveness of the Static Features" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 565, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 661 + ], + "type": "text", + "content": "To validate the effectiveness of the static features, we conducted an ablation study on the static features and illustrated the results in Tab. 5. For both the neural representation (NN) and the grid representation (Grid), the models with static features performs superior to the ones using only dynamic features in all metrics. Additional ablation studies and qualitative results can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 134, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 134, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 134, + 682 + ], + "type": "text", + "content": "4.5. Failure Cases" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Although the proposed feature interpolation is able to learn meaningful spatiotemporal features in most cases," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 265, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 307, + 265, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 338 + ], + "type": "image", + "image_path": "19bb8c6b5be247b5e0af71038c78a774d2b58e7cbeabe3eca0e87d0f6a4b7386.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 345, + 492, + 357 + ], + "lines": [ + { + "bbox": [ + 359, + 345, + 492, + 357 + ], + "spans": [ + { + "bbox": [ + 359, + 345, + 492, + 357 + ], + "type": "text", + "content": "Figure 8. Examples of failure cases." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 375, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 546, + 437 + ], + "type": "text", + "content": "there are a few failure cases as presented in Fig. 8. For instance, our method has difficulty in recovering 3D structures when small objects in a sequence rapidly move (Fig. 8 left) or when there exist dynamic regions that are not observed in the training sequence (Fig. 8 right)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 451, + 378, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 451, + 378, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 451, + 378, + 463 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 581 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective feature interpolation method for training dynamic NeRFs. Both the neural representation and the grid representation showed impressive performance, and the smoothness term applied to the intermediate feature vectors further improves the performance. Since these methods are unrelated to the existing methods of modeling deformations or estimating scene flows, we believe that the proposed method suggests a new direction of training dynamic NeRFs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "While the neural representation model shows high-quality rendering results owing to the representation power of neural networks, it requires hours of training and seconds of rendering which impose barriers to real-time applications. On the other hand, the grid representation is able to render dynamic scenes in less than a second after a few minutes of training, which makes it more practical for real-world applications. Both representations are mutually complementary, and investigating hybrid representations that take advantages of both representations would be an interesting research direction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4219" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Pontus Andersson, Jim Nilsson, Tomas Akenine-Möller, Magnus Oskarsson, Kalle Åström, and Mark D. Fairchild. FLIP: A Difference Evaluator for Alternating Images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(2):15:1-15:23, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 255 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 2, pages 690-696. IEEE, 2000. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 368 + ], + "type": "text", + "content": "[5] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 412 + ], + "type": "text", + "content": "[6] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. arXiv preprint arXiv:2205.15285, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[7] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "text", + "content": "[8] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 568 + ], + "type": "text", + "content": "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[10] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2022. 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[11] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3907-3916, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[12] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "type": "text", + "content": "[13] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[14] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[15] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "type": "text", + "content": "[16] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[17] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[19] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[20] Thomas Müller, Fabrice Rousselle, Jan Novák, and Alexander Keller. Real-time neural radiance caching for path tracing. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "text", + "content": "[21] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 570, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 634 + ], + "type": "text", + "content": "[22] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 4,5,6,7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[23] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[24] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4220" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 384 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "text", + "content": "[25] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Nonrigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 239 + ], + "type": "text", + "content": "[26] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "type": "text", + "content": "[27] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[28] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 353, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 384 + ], + "type": "text", + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4221" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_content_list.json b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b3a7640d728085c2fbccbfe615ba90e67d3f58e6 --- /dev/null +++ b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_content_list.json @@ -0,0 +1,1645 @@ +[ + { + "type": "text", + "text": "Temporally Consistent Online Depth Estimation Using Point-Based Fusion", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numair Khan", + "bbox": [ + 120, + 181, + 233, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "numairkhan@meta.com", + "bbox": [ + 91, + 202, + 263, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eric Penner", + "bbox": [ + 321, + 181, + 419, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "epenner@meta.com", + "bbox": [ + 297, + 202, + 442, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Douglas Lanman", + "bbox": [ + 511, + 181, + 650, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "douglas.lanman@meta.com", + "bbox": [ + 478, + 202, + 684, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Xiao", + "bbox": [ + 761, + 181, + 836, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lei.xiao@meta.com", + "bbox": [ + 720, + 202, + 872, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth estimation is an important step in many computer vision problems such as 3D reconstruction, novel view synthesis, and computational photography. Most existing work focuses on depth estimation from single frames. When applied to videos, the result lacks temporal consistency, showing flickering and swimming artifacts. In this paper we aim to estimate temporally consistent depth maps of video streams in an online setting. This is a difficult problem as future frames are not available and the method must choose between enforcing consistency and correcting errors from previous estimations. The presence of dynamic objects further complicates the problem. We propose to address these challenges by using a global point cloud that is dynamically updated each frame, along with a learned fusion approach in image space. Our approach encourages consistency while simultaneously allowing updates to handle errors and dynamic objects. Qualitative and quantitative results show that our method achieves state-of-the-art quality for consistent video depth estimation.", + "bbox": [ + 73, + 282, + 473, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 601, + 209, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth reconstruction is a long-standing, fundamental problem in computer vision. For decades the most popular depth estimation techniques were based on stereo matching [37] or structure-from-motion [36]. However, more recently the best results have come from learning-based approaches [7]. As the overall reconstruction quality has improved, focus has shifted to new areas such as monocular estimation [21,32,33], edge quality [58], and temporal consistency [20, 26]. The latter is particularly important for video applications in computational photography and virtual reality [3, 52] as inconsistent depth can cause objectionable flickering and swimming artifacts.", + "bbox": [ + 75, + 628, + 468, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consistent video depth estimation, however, remains a difficult problem as even the best method will suffer from unpredictable errors and imperfections based on scene content, especially in textureless and specular regions. This difficulty is aggravated by many of the aforementioned applications requiring online reconstruction: future frames are", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c59f991eb56e9e396582eac6048e123bef54ac9a771812d77f882e3f43acb06.jpg", + "image_caption": [ + "Figure 1. Comparing monocular depth estimation on the ScanNet dataset across four frames. Clockwise from Top-left: Input RGB image, Ranftl et al.'s DPT [32], our result with a DPT backbone, RGB-D sensor ground truth." + ], + "image_footnote": [], + "bbox": [ + 509, + 250, + 893, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "not known beforehand and temporal consistency must be balanced with error-correction. Furthermore, the presence of dynamic objects — which are inherently inconsistent — adds an additional layer of complication. Luo et al. [26] address these problems by assuming all frames are known beforehand and fine-tuning their method for each input video. Other approaches seek an online solution by encoding consistency in network weights either through a training loss [21], by using recurrent architectures [8, 30, 56], or by conditioning it on the input [24]. Each method, however, ultimately relies on the raw output of a neural network being consistent, which is difficult to achieve due to camera noise and aliasing in convolutional networks [14, 44, 57].", + "bbox": [ + 496, + 547, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose the use of a global point cloud to encourage temporal consistency in online video depth estimation. We demonstrate how to tackle the twin problems of handling dynamic objects and updating a static — and potentially erroneous — point cloud when future frames are not known. With quantitative and qualitative results, we show our approach significantly improves the temporal consistency of both stereo and monocular depth estimation without sacrificing spatial quality. In summary, our contributions are as follows:", + "bbox": [ + 496, + 750, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9119", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5161cd180a0fc18d6e040a09699a090c5380cf88fb2a5d2031501e2e71a431d5.jpg", + "image_caption": [ + "Figure 2. We generate temporally consistent depth maps for each RGB video frame $t$ by fusing the projected depth from a prior point cloud $d_{\\mathrm{p}}^{t}$ with the estimated depth $d^{t}$ . This is done by temporally fusing $d_{\\mathrm{p}}^{t}$ to update dynamic regions, followed by spatial fusion with $d^{t}$ based on confidence estimates of accuracy. The final result $d_{\\mathrm{o}}^{t}$ is used to update the point cloud for the next frame." + ], + "image_footnote": [], + "bbox": [ + 138, + 85, + 834, + 224 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose point cloud-based fusion for temporally consistent video depth estimation.", + "2. We present a three-stage approach to encourage consistency in online settings, while simultaneously allowing updates to improve the accuracy of reconstruction and handle dynamic scenes.", + "3. We present an image-space approach to dynamics estimation and depth fusion that is lightweight and has low runtime overhead." + ], + "bbox": [ + 89, + 290, + 467, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 449, + 218, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth Estimation: The recovery of geometry from color images is an historically important problem in computer vision. In the last decade, deep learning has increasingly been applied to different stages of classic algorithms including geometric [4, 23, 39, 53], photometric [54] and multi-view stereo [12], as well as structure-from-motion (SfM) [50]. More recently, approaches based entirely on learnt priors have allowed high-quality depth estimation from single images [21, 32, 33, 45]. Traditionally, the focus of these methods has been on single frames, with most suffering from temporal artifacts on videos.", + "bbox": [ + 75, + 474, + 468, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Consistency: A number of works have recently demonstrated the advantage of using temporal information for improving depth accuracy and consistency [24, 25, 30]. Long et al. [25] propose temporal coherence by fusing the cost volumes of neighboring frames with a transformer. Similarly, Liu et al.'s [24] method accumulates depth probability over consecutive frames using Bayesian filtering. However, the final outcome in both cases is generated by a refinement network which does not have any explicit temporal constraints. Luo et al. [26] fine-tune a monocular depth network at inference-time using constraints from optical flow and structure-from-motion. Kopf et al. [17] build on this to refine camera poses too. Both methods run offline with the inference-time fine-tuning taking around twenty minutes.", + "bbox": [ + 75, + 643, + 468, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Luo et al. and Kopf et al. are special instances of a more general approach to enforcing consistency in video", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks using an optical flow-based warping loss during training [2, 18, 20]. Used in isolation, however, this approach fails to generalize well to unseen data, often requiring fine-tuning for each new dataset. But it does have the advantage of not relying on camera poses or external structures. Lai et al. [18] and Cao et al. [2] pair a warping loss with a convolutional recurrent module which allows the network to learn temporal affinities more effectively.", + "bbox": [ + 496, + 290, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In fact, the use of recurrent structures is another way of enforcing temporal consistency in video tasks [8, 30, 56]. Zhang et al. [56] propagate spatio-temporal depth information across frames using a novel convolutional LSTM module which is trained using an adversarial loss. While this loss does not have any explicit temporal constraint like the warping loss, it is nonetheless shown to improve depth consistency. Patil et al.'s [30] work on depth prediction and completion demonstrates that recurrent modules enforce consistency and improve accuracy even without special loss functions. Convolutional LSTMs, however, can be difficult to train due to their large memory requirements.", + "bbox": [ + 496, + 410, + 892, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth Fusion: Fusion methods [5, 29, 48, 49] achieve scene reconstruction by blending weighted signed distance field (SDF) volumes for each frame. The use of SDF volumes makes them scale poorly to large scenes and high resolutions. Keller et al. [15] and Lefloch et al. [19] propose point clouds as an alternative to SDFs. Traditional fusion methods, however, are not directly applicable to depth reconstruction as their focus is geometric reconstruction over multiple frames. Completeness, in the sense of reconstructing each visible point in a frame, is not a requirement.", + "bbox": [ + 496, + 592, + 890, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Consistent Video Depth Estimation", + "text_level": 1, + "bbox": [ + 500, + 753, + 816, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a sequence of RGB images $c^t$ with known poses, our goal is to estimate a depth map $d^t$ for each $c^t$ , $t \\in \\{0,1,2,\\ldots\\}$ such that the geometric representation of all scene points visible at time $t$ is accurate and consistent from $d^{t-1}$ to $d^t$ . Furthermore, we want to solve this problem online; that is, at any time instant $t$ future frames $d^i$ , $i > t$ are not known. Accuracy subsumes consistency since a 100% accurate reconstruction is also 100% consistent, but", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9120", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3b10e74cb48eeae057957e958ae240b78ca3c18030061875bcddfc4691f7c172.jpg", + "image_caption": [ + "Figure 3. The temporal fusion stage generates a 2D mask $\\alpha$ for scene regions that change due to motion in frame $t$ . This mask is used to update the reprojected point cloud depth $d_{\\mathrm{p}}^{t}$ with dynamic changes from the observed depth $d^{t}$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 449, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not vice versa. As a result, the majority of past depth estimation methods have focused on optimizing accuracy of reconstruction [4,23,39,53]. In practice, however, a perfectly accurate reconstruction is hardly ever achievable such that most of these methods suffer from inconsistency.", + "bbox": [ + 75, + 281, + 468, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A simple way of enforcing temporal consistency is to use the known camera pose to elevate a depth map $d^t$ to a 3D point cloud and then reproject it into future frames. However, such a reprojection-based approach suffers from the following shortcomings: it fails to handle dynamic scenes, it creates holes in disoccluded regions, and it will propagate any errors in $d^t$ to all subsequent frames.", + "bbox": [ + 75, + 357, + 468, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed approach seeks to exploit the advantages of a global point cloud while addressing the above-mentioned shortcomings. We do this in three steps: (i) The temporal fusion step (Sec. 3.1) identifies and updates regions of the scene that changed from frame $t - 1$ to $t$ by blending $d^{t}$ with prior depth from the global point cloud in image space. (ii) Spatial fusion (Sec. 3.2) recovers spatial details and corrects any errors in the blended output depth map from the first stage. (iii) Finally, we update the global point cloud to incorporate changes from frame $t$ (Sec. 3.3).", + "bbox": [ + 75, + 463, + 468, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal Fusion", + "text_level": 1, + "bbox": [ + 76, + 622, + 243, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{P} \\in \\mathbb{R}^3$ be a prior 3D point cloud representation of a scene at time $t$ generated using the depth input from previous frames. For each $\\mathbf{x} \\in \\mathcal{P}$ let $\\varsigma(\\mathbf{x}) \\in \\mathbb{R}^3$ represent the RGB color, and $\\rho(\\mathbf{x}) \\in \\mathbb{R}$ represent a confidence estimate of the point. Given the camera pose $\\mathbf{T}_t$ and intrinsics matrix $\\mathbf{K}$ , we define $d_{\\mathrm{p}}^t$ , $c_{\\mathrm{p}}^t$ and $w_{\\mathrm{p}}^t$ as the $H \\times W$ depth, color and confidence projection, respectively, of $\\mathcal{P}$ .", + "bbox": [ + 75, + 643, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the temporal fusion stage is to generate a 2D mask $\\alpha \\in [0,1]^{H\\times W}$ for the dynamic regions in $d_{\\mathrm{p}}^t$ . Specifically, we want $\\alpha$ to identify visible points that change position from frame $t - 1$ to $t$ solely due to motion, and not because of camera movement or uncertainty in the depth estimation. We then compute a fused depth map $d_{\\mathrm{f}}^t$ with the dynamic parts updated from $d^t$ while preserving consistent depths from prior frames in the static regions:", + "bbox": [ + 75, + 751, + 468, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {f}} ^ {t} = \\alpha d ^ {t} + (1 - \\alpha) d _ {\\mathrm {p}} ^ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 878, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dc6f55ed986b3ac6399514deb0191d3a79a5c6b5043b755dc1fe44fcb82e7cb.jpg", + "image_caption": [ + "Figure 4. The spatial fusion stage uses the observed depth $d^t$ to improve static regions in the temporally fused depth $d_{\\mathrm{f}}^{t}$ based on confidence in the accuracy of both depths. The confidence is estimated as the learnt uncertainty in the depth prediction model." + ], + "image_footnote": [], + "bbox": [ + 522, + 87, + 879, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc07a5a8645fd7c9a9ded58164c8f401be92de416b2f22aa32870fc34ce20c1a.jpg", + "image_caption": [ + "Figure 5. Left to right: The input $c^t$ , the dynamic change mask $\\alpha$ (Eq. 1), the confidence estimate $\\beta$ (Eq. 4), and the confidence estimate $\\gamma$ (Eq. 3). Original image © Blender Foundation | www.sintel.org" + ], + "image_footnote": [], + "bbox": [ + 501, + 308, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The blending mask $\\alpha$ may be computed as the residual between consecutive frames warped using rigid flow. We observe, however, that a hand-crafted threshold for the residual does not generalize to all scenes, and is prone to failure in the presence of noise and depth artifacts. Thus, we use $\\alpha = \\Theta(d^{t}, d_{\\mathrm{p}}^{t}, c^{t}, c_{\\mathrm{p}}^{t})$ instead, where $\\Theta(\\cdot)$ is a convolutional neural network (Fig. 3). We want $\\alpha$ to have a strong bias towards the prior depth map in static regions. This is encouraged by using the ground truth — rather than the generated — prior point cloud $\\mathcal{P}$ to render $c_{\\mathrm{p}}^{t}$ and $d_{\\mathrm{p}}^{t}$ during training. Using an L1 loss on the blended output $d_{\\mathrm{f}}^{t}$ then ensures that $\\alpha = 1$ minimizes the loss in scene regions that do not change from $t - 1$ to $t$ . As observed earlier, training $\\Theta(\\cdot)$ to directly output $d_{\\mathrm{f}}^{t}$ does not guarantee consistency due to CNN aliasing and rotational invariance.", + "bbox": [ + 496, + 412, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the very first frame $t = 0$ , we use $d_{\\mathrm{p}}^{t} = d^{t}$ . Subsequently, we render the point cloud $\\mathcal{P}$ by splatting each $\\mathbf{x} \\in \\mathcal{P}$ , along with color $\\varsigma(\\mathbf{x})$ and confidence $\\rho(\\mathbf{x})$ , to the nearest rounded pixel with a Z-buffer check. While this introduces aliasing artifacts, it is much faster than surfel-based methods [31]. We ameliorate the aliasing through supersampling, and we use Rosenthal and Linsens's method [35] to fill holes and remove background points.", + "bbox": [ + 496, + 638, + 890, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spatial Fusion", + "text_level": 1, + "bbox": [ + 500, + 771, + 645, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the ground truth to generate $d_{\\mathrm{p}}^{t}$ allows for easily training the temporal fusion stage to maintain consistency in static regions. However, the ground truth is not available at inference-time so that the prior point cloud $\\mathcal{P}$ and consequently $d_{\\mathrm{p}}^{t}$ is likely to have errors from imperfect depth estimation in previous frames. We propose to address this problem in the spatial fusion stage.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9121", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal here is to assign a pixel-wise weight to $d^{t}$ and $d_{\\mathrm{f}}^{t}$ based on a confidence estimate of the accuracy of each depth. This is then used to fuse the two depths in image space to generate the final output $d_{\\mathrm{o}}^{t}$ :", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {o}} ^ {t} = \\frac {\\beta d _ {\\mathrm {f}} ^ {t} + \\gamma d ^ {t}}{\\beta + \\gamma} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 162, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We estimate the confidence weights $\\beta, \\gamma \\in \\mathbb{R}^{H \\times W}$ using a neural network $\\Phi(\\cdot)$ (Fig. 4). Note that while formulated similar to Eq. 1, the weights in Eq. 2 serve a different purpose and, hence, require a different training procedure for $\\Phi(\\cdot)$ . In particular, we represent the weights as the input-dependent aleatoric uncertainty [16] which can be learnt in a self-supervised manner. The confidence $\\gamma$ for $d^t$ is then inversely related to the uncertainty as,", + "bbox": [ + 76, + 207, + 468, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\exp \\left[ - \\Phi \\left(d ^ {t}, c ^ {t}\\right) \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 340, + 468, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The confidence $\\beta$ for $d_{\\mathrm{f}}^{t}$ , in addition to the aleatoric uncertainty $\\Phi (d_{\\mathrm{f}}^{t})$ , also incorporates the prior confidence $w_{\\mathrm{p}}^{t}$ weighted by the temporal mask $\\alpha$ :", + "bbox": [ + 76, + 371, + 468, + 416 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta = \\alpha w _ {\\mathrm {p}} ^ {t} \\exp \\left[ - \\Phi \\left(d _ {\\mathrm {f}} ^ {t}, c ^ {t}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 428, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We box-filter $w_{\\mathrm{p}}^{t}$ before computing $\\beta$ to suppress any high-frequency artifacts from the splattering process. Note that $\\Phi(\\cdot)$ can model uncertainty in $d_{\\mathrm{f}}^{t}$ resulting from errors in temporal filtering, so that Eq. 2 is not the same as $\\alpha$ -blending $d^{t}$ with $w_{\\mathrm{p}}^{t}\\exp \\left[-\\Phi (d_{\\mathrm{p}}^{t},c^{t})\\right]d_{\\mathrm{p}}^{t}$ .", + "bbox": [ + 76, + 458, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Global Point Cloud Update", + "text_level": 1, + "bbox": [ + 76, + 544, + 321, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a final step, we update the global point cloud $\\mathcal{P} \\subseteq \\mathbb{R}^3$ to reflect changes in the scene from frame $t$ . This includes updating the confidence $\\rho(\\mathbf{x})$ , and the color $\\varsigma(\\mathbf{x})$ of all points $\\mathbf{x} \\in \\mathcal{P}$ observed at $t$ . This is similar to the TSDF volume integration step in traditional fusion methods [13, 29]. However, here we have a discrete point cloud representation of the scene which cannot be directly averaged like a TSDF. Therefore, achieving a similar effect involves: (i) associating points in $\\mathcal{P}$ with pixels in $d_{\\mathrm{o}}^{t}$ , and updating the position, confidence and color of points for which a correspondence is found; (ii) adding pixels from $d^{t}$ for which no association is found as newly observed points into $\\mathcal{P}$ ; and (iii) removing points with low confidence from $\\mathcal{P}$ .", + "bbox": [ + 76, + 566, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $z^t \\in \\mathbb{R}^{H \\times W \\times 3}$ represent the inverse projection of depth map $d^t$ to 3D. Updates to each $\\mathbf{x} \\in \\mathcal{P}$ are computed in continuous image space to minimize quantization artifacts from using a point cloud. Thus, we update a given $\\mathbf{x}$ by sampling $z^t$ and the color image $c^t$ at the projected 2D location $\\hat{\\mathbf{x}}$ of $\\mathbf{x}$ on the image plane. We denote the sampled values as $c^t[\\hat{\\mathbf{x}}] \\in \\mathbb{R}^3$ and $z^t[\\hat{\\mathbf{x}}] \\in \\mathbb{R}^3$ , respectively, using square brackets to represent bilinear sampling of the image at the indexed location. We only update unoccluded points", + "bbox": [ + 76, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "identified by $\\alpha[\\hat{\\mathbf{x}}] < 0.5$ , where $\\alpha$ is from Eq. 1. The updated $\\mathbf{x}$ , color $\\varsigma(\\mathbf{x})$ and confidence $\\rho(\\mathbf{x})$ are computed as,", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} = \\frac {\\beta [ \\hat {\\mathbf {x}} ] \\cdot \\mathbf {x} + \\gamma [ \\hat {\\mathbf {x}} ] \\cdot z ^ {t} [ \\hat {\\mathbf {x}} ]}{\\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ]} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 138, + 890, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varsigma (\\mathbf {x}) = \\frac {\\beta [ \\hat {\\mathbf {x}} ] \\cdot \\varsigma (\\mathbf {x}) + \\gamma [ \\hat {\\mathbf {x}} ] \\cdot c ^ {t} [ \\hat {\\mathbf {x}} ]}{\\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ]} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 175, + 890, + 208 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho (\\mathbf {x}) = \\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 210, + 890, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ , $\\gamma \\in \\mathbb{R}^{H\\times W}$ are the fusion weights from Eqs. 3 and 4. For any $\\mathbf{x}$ outside the field-of-view, or which is occluded, we set $\\rho (\\mathbf{x}) = \\rho (\\mathbf{x}) - 1$ . We identify newly observed points as pixel locations in $z^t$ for which $\\alpha \\geq 0.5$ , and add them to $\\mathcal{P}$ . Finally, we remove all points $\\mathbf{x}\\in \\mathcal{P}$ for which $\\rho (\\mathbf{x}) < \\epsilon$ . We find $\\epsilon = 3\\times 10^{-2}$ to work well empirically.", + "bbox": [ + 498, + 234, + 890, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Training Procedure", + "text_level": 1, + "bbox": [ + 500, + 338, + 687, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the temporal fusion network $\\Theta(\\cdot)$ using samples $(d^{t}, c^{t}, d_{\\mathrm{p}}^{t}, c_{\\mathrm{p}}^{t})$ where $(c^{t}, d^{t})$ are the color image and estimated depth of the current frame $t$ , and $(d_{\\mathrm{p}}^{t}, c_{\\mathrm{p}}^{t})$ are generated by reproducing the ground truth depth and color from frame $t - k$ to $t$ , for some arbitrary $k$ between -7 and 7. We use future frames ( $k < 0$ ) only during training as a data augmentation technique — during inference, our method only sees the current frame $t$ and the global point cloud.", + "bbox": [ + 498, + 364, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Theta (\\cdot)$ consists of two sets of three residual layers followed by a four-layer U-Net [34]. The residual layers extract features from the concatenated depths $\\langle d_{\\mathrm{p}}^{t},d^{t}\\rangle$ and concatenated color images $\\langle c_{\\mathrm{p}}^{t},c^{t}\\rangle$ which are then fed into the U-Net to generate the blending mask $\\alpha$ (Fig. 3). This network architecture is much simpler and has fewer parameters than either Mask-RCNN [11] or optical flow methods, both of which have been previously used [22, 26, 55] to mask dynamic elements of a scene (Table 5). We use ReLU activations on all but the last layer of the U-Net, which has a sigmoid. As batch normalization adds a depth-dependent bias [48], we use layer normalization instead.", + "bbox": [ + 496, + 484, + 890, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motivation behind using the ground truth rather than the estimated depth for $d_{\\mathrm{p}}^{t}$ is to encourage a strong bias towards the prior depth $(\\alpha = 0)$ in static regions. This ensures $\\Theta(\\cdot)$ only identifies changes due to motion. The reprojected ground truth remains optimal in regions that do not change from frame $t - k$ to $t$ . Thus, an L1 loss on $d_{\\mathrm{f}}^{t}$ encourages $\\alpha = 0$ in these parts. For dynamic objects, the prior depth may or may not be optimal based on the quality of the estimated depth, and the network learns a suitable blending weight. We found a binary cross entropy (BCE) loss on $\\alpha$ improves convergence and pushes it closer to binary:", + "bbox": [ + 496, + 666, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\mathrm {B C E}} = \\mathrm {B C E} \\left(\\alpha , \\mathbb {1} \\left(\\| d ^ {t} - g ^ {t} \\| _ {1} < \\| d _ {\\mathrm {p}} ^ {t} - g ^ {t} \\| _ {1}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 838, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function of the pixels satisfying the inequality and $g^{t}$ is the ground truth depth. We also", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9122", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/fa7bf8f492d3a835cddf178ff6815ffd373e09b6c60b30f189be138b246eecfa.jpg", + "table_caption": [], + "table_footnote": [ + "†Method uses temporal data structure. ‡ Method has temporal consistency constraints." + ], + "table_body": "
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MiDaS [33]0.6110.6750.2110.2330.4130.5970.2793.0130.6100.7980.8961
WSVD [45]‡0.6130.7040.2880.3560.4490.6540.3603.7750.4660.7050.8252
DPT [32]0.4240.4930.3200.4170.4820.5390.2242.6780.6860.8640.9321
Ours-DPT†‡0.2550.2950.4890.4700.5590.4740.1972.4000.7100.8850.9421
ScanNet Dataset
ST-CLSTM [56]‡0.0870.0870.3100.1340.4000.0760.2890.7510.7720.9320.9835
TCMD [20]‡0.0310.0310.6280.1660.5480.0830.3000.8110.7400.9270.9811
N-RGB→D [24]†‡0.0430.0420.5570.1690.4820.0880.2850.6740.8140.9520.9831
Demon [43]0.1600.1590.1430.1070.2710.0980.2280.6510.8390.9510.9822
DeepV2D [40]0.0380.0380.4830.1430.4950.0580.1530.5020.9570.9950.9992
DPT [32]0.0330.0330.5400.1410.5360.0550.2130.5820.9710.9960.9991
Ours-DPT†‡0.0110.0100.8630.1790.6390.0520.2100.5760.9740.9960.9991
", + "bbox": [ + 76, + 88, + 893, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparing our method using a DPT backbone to SOTA monocular depth methods. $N$ is the number of frames required at inference. Our method achieves the best results overall with significantly higher consistency, and similar or better accuracy.", + "bbox": [ + 75, + 368, + 893, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "include an L1 loss $\\mathrm{L_G}$ on the depth gradients, and a VGG loss on the fused depth. We found the latter helps capture fine detail better than an L1 loss alone. Thus, we supervise the temporal fusion network with loss:", + "bbox": [ + 75, + 421, + 468, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\Theta} = \\left\\| d _ {\\mathrm {f}} ^ {t} - g ^ {t} \\right\\| _ {1} + \\mathrm {L} _ {\\mathrm {B C E}} + \\mathrm {L} _ {\\mathrm {G}} + \\mathrm {V G G} \\left(d _ {\\mathrm {f}} ^ {t}, g ^ {t}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 487, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We weigh the terms as 10, 0.1, 0.05 and 0.05, respectively. We implement the depth fusion network $\\Phi(\\cdot)$ as a four-layer U-Net with instance normalization, and ReLU activations on all layers. We train the network using a self-supervised loss on the uncertainty $s^t = \\Phi(d^t, c^t)$ of the input depth:", + "bbox": [ + 75, + 510, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\Phi} = \\frac {1}{m} \\left\\| \\right. \\exp \\left(- s ^ {t}\\right) \\mid d ^ {t} - g ^ {t} \\left. \\right| + \\lambda_ {s} s ^ {t} \\left\\| \\right. _ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 595, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m$ is the number of pixels and the second term is a regularizer to prevent the trivial $s^t = \\infty$ . We use $\\lambda_s = 0.03$ . As Kendall et al. [16] show, this loss term minimizes the log variance of a Gaussian likelihood model for the uncertainty.", + "bbox": [ + 75, + 630, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train both networks on the FlyingThings3D dataset [27]. This consists of rendered sequences of objects moving along random 3D trajectories, and has ground truth depth and camera poses. For fine-tuning we render 50 60-frame sequences from a custom Blender scene with human-like motion in an indoor setting. We estimate $d^{t}$ with RAFT-Stereo [23] and use the inverse depth for training.", + "bbox": [ + 75, + 689, + 468, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 832, + 294, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both networks are implemented in PyTorch and trained on six Nvidia Tesla V100 GPUs. The temporal fusion network is trained for five epochs on $500 \\times 500$ crops from the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/827895ef8ab15ec6a29c5d68b71735e32957e69be3c365a767d451294f75b57f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 417, + 629, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8d501c75ed02f815b23c5748ba42744ba17616b429e0bd8a1c095b8f2a6f7fc.jpg", + "image_caption": [ + "COLMAP MVS", + "Figure 6. MVS and fusion methods like NeuralRecon [38] suffer from holes in the output making them ill-suited to AR applications." + ], + "image_footnote": [], + "bbox": [ + 629, + 417, + 756, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/382439c940b7df1644d1daedba32de6769c443bc372d66bd3eb1de440ba284ba.jpg", + "image_caption": [ + "NeuralRecon" + ], + "image_footnote": [], + "bbox": [ + 759, + 417, + 887, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FlyingThings3D dataset in batches of 18. We use a learning rate of $5 \\times 10^{-4}$ and a weight decay of $1 \\times 10^{-4}$ . The spatial fusion network is trained on the depth output of the temporal filtering stage $d_{\\mathrm{f}}^{t}$ using the FlyingThings3D dataset with a similar batch and input size, and a learning rate of $1 \\times 10^{-4}$ . We fine-tune both networks on $320 \\times 320$ crops of our Blender data for 14 epochs at a learning rate of $1 \\times 10^{-6}$ .", + "bbox": [ + 498, + 537, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Testing Datasets", + "text_level": 1, + "bbox": [ + 500, + 660, + 660, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MPI Sintel: The MPI Sintel dataset [1, 51] consists of long synthetic stereo sequences with large motion and depth range. We evaluate both stereo and monocular depth estimation on the final pass of the 22 \"training scenes\" which have ground truth optical flow, pose, and depth. Given the extremely large motion and depth of some scenes, we follow Teed et al. [42] and restrict our evaluation to pixels with flow $\\leq 250$ pixels and depth $\\leq 30$ meters.", + "bbox": [ + 496, + 686, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScanNet: The ScanNet dataset [9] consists of video sequences of static indoor scenes captured with an RGB-D sensor and annotated with camera poses. We evaluate the monocular variant of our method on 20 60-frame sequences from different scenes. ScanNet does not have ground truth optical flow, so we estimate it using RAFT [41].", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9123", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6db6ba1ff2b0d6ad0b7cfe2c79daf3a8e6b98f879a190499ec0c9443dc2eb851.jpg", + "table_caption": [ + "MPI Sintel Dataset" + ], + "table_footnote": [], + "table_body": "
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MC [21]0.5530.6080.2330.4230.4020.6190.3713.7710.4620.6900.8201
CVD-MC [26]0.2260.2450.4020.5150.4600.5890.3383.0780.4930.7020.819All
-59.1%-59.7%+72.5%+21.7%+14.4%-4.85%-8.89%-18.4%+6.71%+1.74%-0.12%
Ours-MC0.3770.3990.3600.4430.4230.5830.3613.7110.4660.6960.8231
-31.8%-34.4%+54.5%+4.73%+5.22%-5.82%-2.70%-1.59%+0.87%+0.87%+0.37%
", + "bbox": [ + 78, + 111, + 897, + 219 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8500ed0da8e894bd564a38ca6b6b716c78e01b825375e27d7136af65bd7e632b.jpg", + "table_caption": [ + "Table 2. We compare our approach to CVD using the Mannequin Challenge (MC) backbone for both methods. The percentage change in each metric over the baseline (MC) is listed in color. Our method realizes a large portion of the gains of CVD, even though CVD fine-tunes the backbone network for each input ( $\\sim$ 20 mins) and requires all frames beforehand. Our method is trained once and runs online." + ], + "table_footnote": [], + "table_body": "
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑
RAFT-Stereo [23]0.4090.5070.6150.4240.6420.7950.23931.260.9150.9440.961
Ours-RAFT-S.0.2680.3400.7010.4920.6740.7450.22721.860.9150.9430.959
-34.5%-32.9%+14.0%+16.0%+4.98%-6.29%-5.02%-30.1%+0.00%-0.11%-0.21%
HITNet [39]0.2250.2640.5990.4410.6410.3410.08712.320.8780.9150.941
Ours-HITNet0.1580.1970.6750.5060.6750.3160.0868.9000.8780.9160.943
-29.8%-25.4%+12.7%+14.7%+5.30%-7.33%-1.15%-27.8%+0.00%+0.11%+0.21%
", + "bbox": [ + 84, + 282, + 885, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Evaluating two SOTA stereo methods when used with our point cloud-based fusion approach. The percentage change is listed in color below our result for each metric. Our method leads to significant improvement in temporal consistency and preserves spatial quality.", + "bbox": [ + 75, + 435, + 892, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Baselines", + "text_level": 1, + "bbox": [ + 76, + 479, + 184, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our choice of baselines is motivated by the use cases we envision for an online method. These include low-latency AR/VR applications [3,52] requiring temporal consistency, completeness (i.e. no holes), and the handling of dynamic objects. Hence, we evaluate stereo depth methods [23,39] on the MPI Sintel dataset, and monocular methods [20,24,32,33,45,56] on both the MPI Sintel and ScanNet datasets. Additionally, we evaluate learned structure-from-motion methods [40,43]. We do not compare with traditional multi-view stereo or TSDF fusion methods as they suffer incompleteness, and fail in dynamic scenes (Fig. 6).", + "bbox": [ + 75, + 503, + 468, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each baseline, we use pre-trained models provided by the authors and evaluate it only on the dataset it has originally been tested on — with the only exception being models trained on the NYU-V2 dataset [28]. We evaluate these models on ScanNet instead (NYU-V2 is a similar sensor-captured RGB-D dataset, but does not have ground truth poses). We use the efficient Hybrid variant of Ranftl et al.'s DPT [32] in all experiments, and the five-layer 3D-GAN version of Zhang et al.'s ST-CLSTM [56].", + "bbox": [ + 75, + 670, + 470, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 816, + 261, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given a sequence of estimated depth maps $d^0, d^1, \\ldots, d^n$ along with color images $c^0, c^1, \\ldots, c^n$ and ground truth $g^0, g^1, \\ldots, g^n$ , we evaluate depth estimation quality both temporally and spatially. A common metric for tempo", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ral consistency of video uses the optical flow $O_{t \\to t + 1}$ to evaluate the change in a pixel's depth from frame $t$ to $t + 1$ [2,18,46]. Let $d_w^{t + 1}$ be the estimated depth $d^{t + 1}$ backward warped into frame $t$ using $O_{t \\to t + 1}$ . Then the optical flow-based warping error (OPW) is defined as:", + "bbox": [ + 496, + 481, + 892, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {O P W} = \\frac {1}{n - 1} \\sum_ {t} \\| M ^ {t} \\left(d _ {w} ^ {t + 1} - d ^ {t}\\right) \\| _ {1} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 566, + 890, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M^t = \\exp(-\\kappa (c_w^{t+1} - c^t))$ is a pixel-wise occlusion mask. Intuitively, OPW is the average meter change in the depth of a point across frames. It is expected to be higher for the outdoor scenes in MPI-Sintel than the indoor ScanNet dataset. Li et al. [20] present a variant of this error known as the relative temporal consistency (RTC) metric, which is shown to align with humans' perception of consistency:", + "bbox": [ + 496, + 614, + 893, + 723 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R T C} = \\frac {1}{m n} \\sum_ {t} \\left\\| \\mathbb {1} \\left(M ^ {t} \\max \\left(\\frac {d _ {w} ^ {t + 1}}{d _ {t}}, \\frac {d _ {t}}{d _ {w} ^ {t + 1}}\\right) < \\tau\\right) \\right\\| _ {0} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 744, + 890, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m$ is the number of pixels, $\\mathbb{1}(\\cdot)$ is the indicator function of the set of pixels satisfying the enclosed inequality, and the 0-norm measures the number of non-zero pixel values. We set $\\kappa = 50$ and $\\tau = 1.01$ for our evaluation.", + "bbox": [ + 496, + 794, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A disadvantage of OPW and RTC is that both metrics are optimized by the degenerate case $d^{t} = k \\forall t \\in \\{0,1,\\dots,n\\}$ for any constant $k$ . Hence we augment OPW and RTC with", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9124", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4269a440ac7f4920fbc0739ccba6c0230fc6300c5f25e98135dee76442dd47c8.jpg", + "image_caption": [ + "Figure 7. The improvement in consistency from our method, visualized as the variation in an epipolar plane image (EPI). An EPI represents a slice of the video in the $x - t$ dimensions for a fixed $y$ (indicated in green, red and blue). We use a high-gradient color-map to better visualize changes in the EPI. Our method reduces high-frequency oscillation along $t$ indicating better frame-to-frame consistency, and preserves the spatial $(x - y)$ quality of depth. We omit the axis labels in subsequent figures. Original images © Blender Foundation | www.sintel.org" + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 890, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the temporal change consistency (TCC) and the temporal motion consistency(TCM) metrics of Zhang et al. [56]:", + "bbox": [ + 76, + 411, + 470, + 443 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T C C} = \\frac {1}{n - 1} \\sum_ {t} \\operatorname {S S I M} \\left(\\left| d ^ {t} - d ^ {t + 1} \\right|, \\left| g ^ {t} - g ^ {t + 1} \\right|\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 452, + 468, + 487 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T C M} = \\frac {1}{n - 1} \\sum_ {t} \\operatorname {S S I M} \\left(O \\left(d ^ {t}, d ^ {t + 1}\\right), O \\left(g ^ {t}, g ^ {t + 1}\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 489, + 468, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where SSIM is the structural self-similarity [47], and $O$ is the optical flow $O_{t\\rightarrow t + 1}$ . In addition, we propose the self-consistency (SC) metric which uses the estimated depth $d^t$ to compute $d_w^{t + 1}$ instead of the optical flow. Like OPW, it measures the average meter change in depth frame-to-frame. Finally, as in Long et al. [25] we measure the standard deviation of the L1 error SD(L1) across all frames.", + "bbox": [ + 75, + 535, + 468, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the spatial error we use the standard metrics: relative absolute error (RAE); root mean squared error (RMS); and the percentage of bad pixels $\\delta_1$ , $\\delta_2$ , $\\delta_3$ at thresholds 1.25, $1.25^2$ and $1.25^3$ , respectively [56]. We align the affine-invariant monocular methods for evaluation by using the ground truth to calculate a scale and shift factor [33].", + "bbox": [ + 75, + 641, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Results", + "text_level": 1, + "bbox": [ + 76, + 739, + 171, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents the quantitative evaluation of our method and the monocular baselines, with a qualitative comparison in Fig. 8. Across both datasets, our depth has a $50\\%$ lower OPW error than the next best method. Intuitively, this means for an average scene point, the frame-to-frame meter difference in depth is less than half that of the other method. Furthermore, the structural similarity of the temporal change in our depths and ground truth (TCC) is $15\\%$ higher, indicating that dynamic regions are not simply", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/70d5e44ab865d05ea2036838a0042edb4e2ee3ec47dcf6ade3c8a5dbb9d41df0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TCM↑TCC↑OPW↓RAE↓
No Temporal Fusion-11.4%-12.1%+27.7%+65.7%
No Spatial Fusion-7.44%-9.06%+39.6%+17.1%
No Global PC-7.45%-9.57%+39.5%+13.0%
", + "bbox": [ + 501, + 407, + 890, + 486 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e2d8f45647b2c32d65aaf5e845de2fcd154a6b7aa4a92da8a4463333a480caf.jpg", + "table_caption": [ + "Table 4. Ablating our pipeline to evaluate its contribution to the final result. We list the percentage change with respect to Ours-DPT in Table 1 on the dynamic MPI Sintel dataset. For TCM and TCC higher is better, and vice versa for OPW and RAE." + ], + "table_footnote": [], + "table_body": "
Params. (M)GMACsRuntime (ms)
RAFT (20 iters.)5.26309.586.2
Mask-RCNN44.2134.642.4
Our Θ(·)4.4535.047.20
Our Φ(·)4.4437.145.10
", + "bbox": [ + 500, + 561, + 893, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Run-time and complexity evaluation of our two networks $\\Theta (\\cdot)$ and $\\Phi (\\cdot)$ with a $512\\times 512$ image on an NVIDIA GeForce RTX 3080 GPU. We use the default 20 iterations for RAFT.", + "bbox": [ + 498, + 665, + 892, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "averaged across frames. A similar trend is seen in the improved performance of the stereo methods in Table 3 with an improvement in OPW and TCC of around $30\\%$ and $5\\%$ respectively. Fig. 7 provides a qualitative visualization of the stereo results. The improvement from our method is compared to Luo et al.'s CVD [26] in Table 2. CVD runs offline, fine-tuning the backbone network for $\\sim 20$ minutes at inference-time. As such, it represents the high bar for temporal consistency. Our method achieves a good portion of the gains of CVD while being generalizable and online.", + "bbox": [ + 496, + 718, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 lists the size and runtime of our two networks, comparing $\\Theta (\\cdot)$ to RAFT [41] and Mask-RCNN [11] — the", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9125", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ace5f4dbfef425fb84a7c156ef1bdee4982e679ae45df1646f7308c88d2cbe46.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison of the baseline monocular methods, and our approach with a DPT [32] backbone. We visualize consistency as variation in the cyan slice along the temporal dimension. Our method is more consistent, and captures finer spatial details." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 885, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "latter having been used for dynamics filtering [20, 26, 55]. Table 4 show how ablating each stage affects the performance of our method. For No Temporal Fusion we set $\\alpha = 0$ in Eq. 1; No Spatial Fusion uses $d_{\\mathrm{o}}^{t} = d_{\\mathrm{f}}^{t}$ ; and No Global PC does not use a global point cloud, using the reprojected points from frame $t - 1$ only. The temporal fusion stage has the strongest effect on consistency and accuracy. Setting $\\alpha = 0$ effectively averages the frames via Eq. 2 yielding temporally smooth — albeit erroneous — output. Hence, the drop in TCC but the lower OPW error as the latter is minimized when depth variance across frames is low.", + "bbox": [ + 75, + 631, + 472, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 814, + 194, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a method for temporally consistent video depth estimation in online settings. Our approach uses a global point cloud, and learnt image-space fusion to encourage consistency and preserve accuracy. While we assumed", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "camera poses and the scale of monocular depth was known, these can also be computed using a point cloud alignment algorithm such as Iterative Closest Point (ICP) [13, 29], or any visual SLAM system [10]. In the interest of generality, our method does not assume any particular depth-estimation algorithm. But the results may be improved by conditioning the estimation of $d^{t}$ on the reprojected depth $d_{\\mathrm{p}}^{t}$ . Furthermore, while our approach does not utilize the second view in stereo settings, it may be possible to exploit left-right view consistency in addition to temporal consistency [6]. We leave these directions for future work.", + "bbox": [ + 496, + 631, + 893, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 825, + 668, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Julia Majors and Thu Nguyen-Phuoc for proofreading the draft. Zhao Dong and Zhaoyang Lv provided useful insights on the technical aspects of the project.", + "bbox": [ + 498, + 854, + 892, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9126", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 5", + "[2] Yuanzhouhan Cao, Yidong Li, Haokui Zhang, Chao Ren, and Yifan Liu. Learning structure affinity for video depth estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 190-198, 2021. 2, 6", + "[3] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 1, 6", + "[4] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33, 2020. 2, 3", + "[5] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021. 2", + "[6] Steven D Cochran and Gerard Medioni. 3-d surface description from binocular stereo. IEEE Transactions on Pattern Analysis & Machine Intelligence, 14(10):981-994, 1992. 8", + "[7] Middlebury College. Middlebury stereo benchmark. https://vision.middley.edu/stereo/eval3/.1", + "[8] Arun CS Kumar, Suchendra M Bhandarkar, and Mukta Prasad. Depthnet: A recurrent neural network architecture for monocular depth prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 283-291, 2018. 1, 2", + "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5", + "[10] Jorge Fuentes-Pacheco, José Ruiz-Ascencio, and Juan Manuel Rendon-Mancha. Visual simultaneous localization and mapping: a survey. Artificial intelligence review, 43(1):55–81, 2015. 8", + "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4, 7", + "[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[13] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 4, 8", + "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 1", + "[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2", + "[16] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4, 5", + "[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018, 2, 6", + "[19] Damien Lefloch, Tim Weyrich, and Andreas Kolb. Anisotropic point-based fusion. In Proceedings of International Conference on Information Fusion (FUSION), pages 1-9. ISIF, July 2015. 2", + "[20] Siyuan Li, Yue Luo, Ye Zhu, Xun Zhao, Yu Li, and Ying Shan. Enforcing temporal consistency in video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 8", + "[21] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 1, 2, 6", + "[22] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. arXiv preprint arXiv:2111.09337, 2021. 4", + "[23] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In International Conference on 3D Vision (3DV), 2021. 2, 3, 5, 6", + "[24] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural $\\mathrm{rgb} \\rightarrow \\mathrm{d}$ sensing: Depth and uncertainty from a video camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986-10995, 2019. 1, 2, 5, 6", + "[25] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8258-8267, June 2021. 2, 7", + "[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 2, 4, 6, 7, 8" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9127", + "bbox": [ + 480, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 5", + "[28] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6", + "[29] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2, 4, 8", + "[30] Vaishakh Patil, Wouter Van Gansbeke, Dengxin Dai, and Luc Van Gool. Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters, 5(4):6813-6820, 2020. 1, 2", + "[31] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3", + "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1, 2, 5, 6, 8", + "[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5, 6, 7", + "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4", + "[35] Paul Rosenthal and Lars Linsen. Image-space point cloud rendering. In Proceedings of Computer Graphics International, pages 136-143, 2008. 3", + "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1", + "[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 5", + "[39] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 2, 3, 6", + "[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 5, 6", + "[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 5, 7", + "[42] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8375–8384, 2021. 5", + "[43] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 5, 6", + "[44] Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin, Rob Romijnders, Nicolas Le Roux, and Ross Goroshin. Impact of aliasing on generalization in deep convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10529-10538, 2021. 1", + "[45] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 2, 5, 6", + "[46] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. arXiv preprint arXiv:2208.00380, 2022. 6", + "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[48] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 4", + "[49] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Neuralfusion: Online depth fusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3162-3172, June 2021. 2", + "[50] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In European conference on computer vision, pages 230–247. Springer, 2020. 2", + "[51] J. Wulff, D. J. Butler, G. B. Stanley, and M. J. Black. Lessons and insights from creating a synthetic optical flow benchmark. In A. Fusiello et al. (Eds.), editor, ECCV Workshop on Unsolved Problems in Optical Flow and Stereo Estimation, Part II, LNCS 7584, pages 168-177. Springer-Verlag, Oct. 2012. 5", + "[52] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpass through: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 1, 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9128", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 2, 3", + "[54] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2", + "[55] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 4, 8", + "[56] Haokui Zhang, Chunhua Shen, Ying Li, Yuzhouhan Cao, Yu Liu, and Youliang Yan. Exploiting temporal consistency for real-time video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1, 2, 5, 6, 7", + "[57] Richard Zhang. Making convolutional networks shift-invariant again. In ICML, 2019. 1", + "[58] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13116-13125, 2020. 1" + ], + "bbox": [ + 78, + 90, + 470, + 455 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9129", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_model.json b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..16645c03a5860f0a02033dd4f1f6124d903ad2ef --- /dev/null +++ b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Temporally Consistent Online Depth Estimation Using Point-Based Fusion" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.182, + 0.234, + 0.198 + ], + "angle": 0, + "content": "Numair Khan" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.203, + 0.264, + 0.216 + ], + "angle": 0, + "content": "numairkhan@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.182, + 0.42, + 0.197 + ], + "angle": 0, + "content": "Eric Penner" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.203, + 0.443, + 0.216 + ], + "angle": 0, + "content": "epenner@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.651, + 0.199 + ], + "angle": 0, + "content": "Douglas Lanman" + }, + { + "type": "text", + "bbox": [ + 0.479, + 0.203, + 0.686, + 0.216 + ], + "angle": 0, + "content": "douglas.lanman@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.182, + 0.837, + 0.197 + ], + "angle": 0, + "content": "Lei Xiao" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.203, + 0.874, + 0.216 + ], + "angle": 0, + "content": "lei.xiao@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.284, + 0.474, + 0.572 + ], + "angle": 0, + "content": "Depth estimation is an important step in many computer vision problems such as 3D reconstruction, novel view synthesis, and computational photography. Most existing work focuses on depth estimation from single frames. When applied to videos, the result lacks temporal consistency, showing flickering and swimming artifacts. In this paper we aim to estimate temporally consistent depth maps of video streams in an online setting. This is a difficult problem as future frames are not available and the method must choose between enforcing consistency and correcting errors from previous estimations. The presence of dynamic objects further complicates the problem. We propose to address these challenges by using a global point cloud that is dynamically updated each frame, along with a learned fusion approach in image space. Our approach encourages consistency while simultaneously allowing updates to handle errors and dynamic objects. Qualitative and quantitative results show that our method achieves state-of-the-art quality for consistent video depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.602, + 0.21, + 0.618 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Depth reconstruction is a long-standing, fundamental problem in computer vision. For decades the most popular depth estimation techniques were based on stereo matching [37] or structure-from-motion [36]. However, more recently the best results have come from learning-based approaches [7]. As the overall reconstruction quality has improved, focus has shifted to new areas such as monocular estimation [21,32,33], edge quality [58], and temporal consistency [20, 26]. The latter is particularly important for video applications in computational photography and virtual reality [3, 52] as inconsistent depth can cause objectionable flickering and swimming artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Consistent video depth estimation, however, remains a difficult problem as even the best method will suffer from unpredictable errors and imperfections based on scene content, especially in textureless and specular regions. This difficulty is aggravated by many of the aforementioned applications requiring online reconstruction: future frames are" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.251, + 0.895, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.471, + 0.895, + 0.528 + ], + "angle": 0, + "content": "Figure 1. Comparing monocular depth estimation on the ScanNet dataset across four frames. Clockwise from Top-left: Input RGB image, Ranftl et al.'s DPT [32], our result with a DPT backbone, RGB-D sensor ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.746 + ], + "angle": 0, + "content": "not known beforehand and temporal consistency must be balanced with error-correction. Furthermore, the presence of dynamic objects — which are inherently inconsistent — adds an additional layer of complication. Luo et al. [26] address these problems by assuming all frames are known beforehand and fine-tuning their method for each input video. Other approaches seek an online solution by encoding consistency in network weights either through a training loss [21], by using recurrent architectures [8, 30, 56], or by conditioning it on the input [24]. Each method, however, ultimately relies on the raw output of a neural network being consistent, which is difficult to achieve due to camera noise and aliasing in convolutional networks [14, 44, 57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.896, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose the use of a global point cloud to encourage temporal consistency in online video depth estimation. We demonstrate how to tackle the twin problems of handling dynamic objects and updating a static — and potentially erroneous — point cloud when future frames are not known. With quantitative and qualitative results, we show our approach significantly improves the temporal consistency of both stereo and monocular depth estimation without sacrificing spatial quality. In summary, our contributions are as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9119" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.087, + 0.835, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 2. We generate temporally consistent depth maps for each RGB video frame \\( t \\) by fusing the projected depth from a prior point cloud \\( d_{\\mathrm{p}}^{t} \\) with the estimated depth \\( d^{t} \\). This is done by temporally fusing \\( d_{\\mathrm{p}}^{t} \\) to update dynamic regions, followed by spatial fusion with \\( d^{t} \\) based on confidence estimates of accuracy. The final result \\( d_{\\mathrm{o}}^{t} \\) is used to update the point cloud for the next frame." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.291, + 0.468, + 0.32 + ], + "angle": 0, + "content": "1. We propose point cloud-based fusion for temporally consistent video depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.468, + 0.385 + ], + "angle": 0, + "content": "2. We present a three-stage approach to encourage consistency in online settings, while simultaneously allowing updates to improve the accuracy of reconstruction and handle dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.468, + 0.434 + ], + "angle": 0, + "content": "3. We present an image-space approach to dynamics estimation and depth fusion that is lightweight and has low runtime overhead." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.291, + 0.468, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.22, + 0.466 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Depth Estimation: The recovery of geometry from color images is an historically important problem in computer vision. In the last decade, deep learning has increasingly been applied to different stages of classic algorithms including geometric [4, 23, 39, 53], photometric [54] and multi-view stereo [12], as well as structure-from-motion (SfM) [50]. More recently, approaches based entirely on learnt priors have allowed high-quality depth estimation from single images [21, 32, 33, 45]. Traditionally, the focus of these methods has been on single frames, with most suffering from temporal artifacts on videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Temporal Consistency: A number of works have recently demonstrated the advantage of using temporal information for improving depth accuracy and consistency [24, 25, 30]. Long et al. [25] propose temporal coherence by fusing the cost volumes of neighboring frames with a transformer. Similarly, Liu et al.'s [24] method accumulates depth probability over consecutive frames using Bayesian filtering. However, the final outcome in both cases is generated by a refinement network which does not have any explicit temporal constraints. Luo et al. [26] fine-tune a monocular depth network at inference-time using constraints from optical flow and structure-from-motion. Kopf et al. [17] build on this to refine camera poses too. Both methods run offline with the inference-time fine-tuning taking around twenty minutes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Luo et al. and Kopf et al. are special instances of a more general approach to enforcing consistency in video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.41 + ], + "angle": 0, + "content": "tasks using an optical flow-based warping loss during training [2, 18, 20]. Used in isolation, however, this approach fails to generalize well to unseen data, often requiring fine-tuning for each new dataset. But it does have the advantage of not relying on camera poses or external structures. Lai et al. [18] and Cao et al. [2] pair a warping loss with a convolutional recurrent module which allows the network to learn temporal affinities more effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.893, + 0.592 + ], + "angle": 0, + "content": "In fact, the use of recurrent structures is another way of enforcing temporal consistency in video tasks [8, 30, 56]. Zhang et al. [56] propagate spatio-temporal depth information across frames using a novel convolutional LSTM module which is trained using an adversarial loss. While this loss does not have any explicit temporal constraint like the warping loss, it is nonetheless shown to improve depth consistency. Patil et al.'s [30] work on depth prediction and completion demonstrates that recurrent modules enforce consistency and improve accuracy even without special loss functions. Convolutional LSTMs, however, can be difficult to train due to their large memory requirements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Depth Fusion: Fusion methods [5, 29, 48, 49] achieve scene reconstruction by blending weighted signed distance field (SDF) volumes for each frame. The use of SDF volumes makes them scale poorly to large scenes and high resolutions. Keller et al. [15] and Lefloch et al. [19] propose point clouds as an alternative to SDFs. Traditional fusion methods, however, are not directly applicable to depth reconstruction as their focus is geometric reconstruction over multiple frames. Completeness, in the sense of reconstructing each visible point in a frame, is not a requirement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.818, + 0.772 + ], + "angle": 0, + "content": "3. Consistent Video Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a sequence of RGB images \\( c^t \\) with known poses, our goal is to estimate a depth map \\( d^t \\) for each \\( c^t \\), \\( t \\in \\{0,1,2,\\ldots\\} \\) such that the geometric representation of all scene points visible at time \\( t \\) is accurate and consistent from \\( d^{t-1} \\) to \\( d^t \\). Furthermore, we want to solve this problem online; that is, at any time instant \\( t \\) future frames \\( d^i \\), \\( i > t \\) are not known. Accuracy subsumes consistency since a 100% accurate reconstruction is also 100% consistent, but" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9120" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.45, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 3. The temporal fusion stage generates a 2D mask \\(\\alpha\\) for scene regions that change due to motion in frame \\(t\\). This mask is used to update the reprojected point cloud depth \\(d_{\\mathrm{p}}^{t}\\) with dynamic changes from the observed depth \\(d^{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.469, + 0.357 + ], + "angle": 0, + "content": "not vice versa. As a result, the majority of past depth estimation methods have focused on optimizing accuracy of reconstruction [4,23,39,53]. In practice, however, a perfectly accurate reconstruction is hardly ever achievable such that most of these methods suffer from inconsistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.463 + ], + "angle": 0, + "content": "A simple way of enforcing temporal consistency is to use the known camera pose to elevate a depth map \\(d^t\\) to a 3D point cloud and then reproject it into future frames. However, such a reprojection-based approach suffers from the following shortcomings: it fails to handle dynamic scenes, it creates holes in disoccluded regions, and it will propagate any errors in \\(d^t\\) to all subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.469, + 0.614 + ], + "angle": 0, + "content": "Our proposed approach seeks to exploit the advantages of a global point cloud while addressing the above-mentioned shortcomings. We do this in three steps: (i) The temporal fusion step (Sec. 3.1) identifies and updates regions of the scene that changed from frame \\( t - 1 \\) to \\( t \\) by blending \\( d^{t} \\) with prior depth from the global point cloud in image space. (ii) Spatial fusion (Sec. 3.2) recovers spatial details and corrects any errors in the blended output depth map from the first stage. (iii) Finally, we update the global point cloud to incorporate changes from frame \\( t \\) (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.623, + 0.245, + 0.639 + ], + "angle": 0, + "content": "3.1. Temporal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.751 + ], + "angle": 0, + "content": "Let \\(\\mathcal{P} \\in \\mathbb{R}^3\\) be a prior 3D point cloud representation of a scene at time \\(t\\) generated using the depth input from previous frames. For each \\(\\mathbf{x} \\in \\mathcal{P}\\) let \\(\\varsigma(\\mathbf{x}) \\in \\mathbb{R}^3\\) represent the RGB color, and \\(\\rho(\\mathbf{x}) \\in \\mathbb{R}\\) represent a confidence estimate of the point. Given the camera pose \\(\\mathbf{T}_t\\) and intrinsics matrix \\(\\mathbf{K}\\), we define \\(d_{\\mathrm{p}}^t\\), \\(c_{\\mathrm{p}}^t\\) and \\(w_{\\mathrm{p}}^t\\) as the \\(H \\times W\\) depth, color and confidence projection, respectively, of \\(\\mathcal{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.752, + 0.469, + 0.872 + ], + "angle": 0, + "content": "The goal of the temporal fusion stage is to generate a 2D mask \\(\\alpha \\in [0,1]^{H\\times W}\\) for the dynamic regions in \\(d_{\\mathrm{p}}^t\\). Specifically, we want \\(\\alpha\\) to identify visible points that change position from frame \\(t - 1\\) to \\(t\\) solely due to motion, and not because of camera movement or uncertainty in the depth estimation. We then compute a fused depth map \\(d_{\\mathrm{f}}^t\\) with the dynamic parts updated from \\(d^t\\) while preserving consistent depths from prior frames in the static regions:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.88, + 0.469, + 0.898 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {f}} ^ {t} = \\alpha d ^ {t} + (1 - \\alpha) d _ {\\mathrm {p}} ^ {t} \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.088, + 0.88, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.238, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 4. The spatial fusion stage uses the observed depth \\(d^t\\) to improve static regions in the temporally fused depth \\(d_{\\mathrm{f}}^{t}\\) based on confidence in the accuracy of both depths. The confidence is estimated as the learnt uncertainty in the depth prediction model." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.309, + 0.892, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.362, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Figure 5. Left to right: The input \\( c^t \\), the dynamic change mask \\( \\alpha \\) (Eq. 1), the confidence estimate \\( \\beta \\) (Eq. 4), and the confidence estimate \\( \\gamma \\) (Eq. 3). Original image © Blender Foundation | www.sintel.org" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.639 + ], + "angle": 0, + "content": "The blending mask \\(\\alpha\\) may be computed as the residual between consecutive frames warped using rigid flow. We observe, however, that a hand-crafted threshold for the residual does not generalize to all scenes, and is prone to failure in the presence of noise and depth artifacts. Thus, we use \\(\\alpha = \\Theta(d^{t}, d_{\\mathrm{p}}^{t}, c^{t}, c_{\\mathrm{p}}^{t})\\) instead, where \\(\\Theta(\\cdot)\\) is a convolutional neural network (Fig. 3). We want \\(\\alpha\\) to have a strong bias towards the prior depth map in static regions. This is encouraged by using the ground truth — rather than the generated — prior point cloud \\(\\mathcal{P}\\) to render \\(c_{\\mathrm{p}}^{t}\\) and \\(d_{\\mathrm{p}}^{t}\\) during training. Using an L1 loss on the blended output \\(d_{\\mathrm{f}}^{t}\\) then ensures that \\(\\alpha = 1\\) minimizes the loss in scene regions that do not change from \\(t - 1\\) to \\(t\\). As observed earlier, training \\(\\Theta(\\cdot)\\) to directly output \\(d_{\\mathrm{f}}^{t}\\) does not guarantee consistency due to CNN aliasing and rotational invariance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.761 + ], + "angle": 0, + "content": "For the very first frame \\( t = 0 \\), we use \\( d_{\\mathrm{p}}^{t} = d^{t} \\). Subsequently, we render the point cloud \\( \\mathcal{P} \\) by splatting each \\( \\mathbf{x} \\in \\mathcal{P} \\), along with color \\( \\varsigma(\\mathbf{x}) \\) and confidence \\( \\rho(\\mathbf{x}) \\), to the nearest rounded pixel with a Z-buffer check. While this introduces aliasing artifacts, it is much faster than surfel-based methods [31]. We ameliorate the aliasing through supersampling, and we use Rosenthal and Linsens's method [35] to fill holes and remove background points." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.646, + 0.787 + ], + "angle": 0, + "content": "3.2. Spatial Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Using the ground truth to generate \\( d_{\\mathrm{p}}^{t} \\) allows for easily training the temporal fusion stage to maintain consistency in static regions. However, the ground truth is not available at inference-time so that the prior point cloud \\( \\mathcal{P} \\) and consequently \\( d_{\\mathrm{p}}^{t} \\) is likely to have errors from imperfect depth estimation in previous frames. We propose to address this problem in the spatial fusion stage." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9121" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "The goal here is to assign a pixel-wise weight to \\( d^{t} \\) and \\( d_{\\mathrm{f}}^{t} \\) based on a confidence estimate of the accuracy of each depth. This is then used to fuse the two depths in image space to generate the final output \\( d_{\\mathrm{o}}^{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.164, + 0.469, + 0.197 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {o}} ^ {t} = \\frac {\\beta d _ {\\mathrm {f}} ^ {t} + \\gamma d ^ {t}}{\\beta + \\gamma} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.208, + 0.469, + 0.33 + ], + "angle": 0, + "content": "We estimate the confidence weights \\(\\beta, \\gamma \\in \\mathbb{R}^{H \\times W}\\) using a neural network \\(\\Phi(\\cdot)\\) (Fig. 4). Note that while formulated similar to Eq. 1, the weights in Eq. 2 serve a different purpose and, hence, require a different training procedure for \\(\\Phi(\\cdot)\\). In particular, we represent the weights as the input-dependent aleatoric uncertainty [16] which can be learnt in a self-supervised manner. The confidence \\(\\gamma\\) for \\(d^t\\) is then inversely related to the uncertainty as," + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.342, + 0.469, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\exp \\left[ - \\Phi \\left(d ^ {t}, c ^ {t}\\right) \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.372, + 0.469, + 0.417 + ], + "angle": 0, + "content": "The confidence \\(\\beta\\) for \\(d_{\\mathrm{f}}^{t}\\), in addition to the aleatoric uncertainty \\(\\Phi (d_{\\mathrm{f}}^{t})\\), also incorporates the prior confidence \\(w_{\\mathrm{p}}^{t}\\) weighted by the temporal mask \\(\\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.429, + 0.469, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\beta = \\alpha w _ {\\mathrm {p}} ^ {t} \\exp \\left[ - \\Phi \\left(d _ {\\mathrm {f}} ^ {t}, c ^ {t}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.469, + 0.537 + ], + "angle": 0, + "content": "We box-filter \\( w_{\\mathrm{p}}^{t} \\) before computing \\( \\beta \\) to suppress any high-frequency artifacts from the splattering process. Note that \\( \\Phi(\\cdot) \\) can model uncertainty in \\( d_{\\mathrm{f}}^{t} \\) resulting from errors in temporal filtering, so that Eq. 2 is not the same as \\( \\alpha \\)-blending \\( d^{t} \\) with \\( w_{\\mathrm{p}}^{t}\\exp \\left[-\\Phi (d_{\\mathrm{p}}^{t},c^{t})\\right]d_{\\mathrm{p}}^{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.545, + 0.323, + 0.56 + ], + "angle": 0, + "content": "3.3. Global Point Cloud Update" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.568, + 0.469, + 0.763 + ], + "angle": 0, + "content": "In a final step, we update the global point cloud \\(\\mathcal{P} \\subseteq \\mathbb{R}^3\\) to reflect changes in the scene from frame \\(t\\). This includes updating the confidence \\(\\rho(\\mathbf{x})\\), and the color \\(\\varsigma(\\mathbf{x})\\) of all points \\(\\mathbf{x} \\in \\mathcal{P}\\) observed at \\(t\\). This is similar to the TSDF volume integration step in traditional fusion methods [13, 29]. However, here we have a discrete point cloud representation of the scene which cannot be directly averaged like a TSDF. Therefore, achieving a similar effect involves: (i) associating points in \\(\\mathcal{P}\\) with pixels in \\(d_{\\mathrm{o}}^{t}\\), and updating the position, confidence and color of points for which a correspondence is found; (ii) adding pixels from \\(d^{t}\\) for which no association is found as newly observed points into \\(\\mathcal{P}\\); and (iii) removing points with low confidence from \\(\\mathcal{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.764, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Let \\( z^t \\in \\mathbb{R}^{H \\times W \\times 3} \\) represent the inverse projection of depth map \\( d^t \\) to 3D. Updates to each \\( \\mathbf{x} \\in \\mathcal{P} \\) are computed in continuous image space to minimize quantization artifacts from using a point cloud. Thus, we update a given \\( \\mathbf{x} \\) by sampling \\( z^t \\) and the color image \\( c^t \\) at the projected 2D location \\( \\hat{\\mathbf{x}} \\) of \\( \\mathbf{x} \\) on the image plane. We denote the sampled values as \\( c^t[\\hat{\\mathbf{x}}] \\in \\mathbb{R}^3 \\) and \\( z^t[\\hat{\\mathbf{x}}] \\in \\mathbb{R}^3 \\), respectively, using square brackets to represent bilinear sampling of the image at the indexed location. We only update unoccluded points" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "identified by \\(\\alpha[\\hat{\\mathbf{x}}] < 0.5\\), where \\(\\alpha\\) is from Eq. 1. The updated \\(\\mathbf{x}\\), color \\(\\varsigma(\\mathbf{x})\\) and confidence \\(\\rho(\\mathbf{x})\\) are computed as," + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.14, + 0.891, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} = \\frac {\\beta [ \\hat {\\mathbf {x}} ] \\cdot \\mathbf {x} + \\gamma [ \\hat {\\mathbf {x}} ] \\cdot z ^ {t} [ \\hat {\\mathbf {x}} ]}{\\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ]} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.176, + 0.891, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\varsigma (\\mathbf {x}) = \\frac {\\beta [ \\hat {\\mathbf {x}} ] \\cdot \\varsigma (\\mathbf {x}) + \\gamma [ \\hat {\\mathbf {x}} ] \\cdot c ^ {t} [ \\hat {\\mathbf {x}} ]}{\\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ]} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.212, + 0.891, + 0.229 + ], + "angle": 0, + "content": "\\[\n\\rho (\\mathbf {x}) = \\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.236, + 0.892, + 0.328 + ], + "angle": 0, + "content": "where \\(\\beta\\), \\(\\gamma \\in \\mathbb{R}^{H\\times W}\\) are the fusion weights from Eqs. 3 and 4. For any \\(\\mathbf{x}\\) outside the field-of-view, or which is occluded, we set \\(\\rho (\\mathbf{x}) = \\rho (\\mathbf{x}) - 1\\). We identify newly observed points as pixel locations in \\(z^t\\) for which \\(\\alpha \\geq 0.5\\), and add them to \\(\\mathcal{P}\\). Finally, we remove all points \\(\\mathbf{x}\\in \\mathcal{P}\\) for which \\(\\rho (\\mathbf{x}) < \\epsilon\\). We find \\(\\epsilon = 3\\times 10^{-2}\\) to work well empirically." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.339, + 0.689, + 0.357 + ], + "angle": 0, + "content": "4. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.485 + ], + "angle": 0, + "content": "We train the temporal fusion network \\(\\Theta(\\cdot)\\) using samples \\((d^{t}, c^{t}, d_{\\mathrm{p}}^{t}, c_{\\mathrm{p}}^{t})\\) where \\((c^{t}, d^{t})\\) are the color image and estimated depth of the current frame \\(t\\), and \\((d_{\\mathrm{p}}^{t}, c_{\\mathrm{p}}^{t})\\) are generated by reproducing the ground truth depth and color from frame \\(t - k\\) to \\(t\\), for some arbitrary \\(k\\) between -7 and 7. We use future frames (\\(k < 0\\)) only during training as a data augmentation technique — during inference, our method only sees the current frame \\(t\\) and the global point cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.666 + ], + "angle": 0, + "content": "\\(\\Theta (\\cdot)\\) consists of two sets of three residual layers followed by a four-layer U-Net [34]. The residual layers extract features from the concatenated depths \\(\\langle d_{\\mathrm{p}}^{t},d^{t}\\rangle\\) and concatenated color images \\(\\langle c_{\\mathrm{p}}^{t},c^{t}\\rangle\\) which are then fed into the U-Net to generate the blending mask \\(\\alpha\\) (Fig. 3). This network architecture is much simpler and has fewer parameters than either Mask-RCNN [11] or optical flow methods, both of which have been previously used [22, 26, 55] to mask dynamic elements of a scene (Table 5). We use ReLU activations on all but the last layer of the U-Net, which has a sigmoid. As batch normalization adds a depth-dependent bias [48], we use layer normalization instead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.833 + ], + "angle": 0, + "content": "The motivation behind using the ground truth rather than the estimated depth for \\( d_{\\mathrm{p}}^{t} \\) is to encourage a strong bias towards the prior depth \\( (\\alpha = 0) \\) in static regions. This ensures \\( \\Theta(\\cdot) \\) only identifies changes due to motion. The reprojected ground truth remains optimal in regions that do not change from frame \\( t - k \\) to \\( t \\). Thus, an L1 loss on \\( d_{\\mathrm{f}}^{t} \\) encourages \\( \\alpha = 0 \\) in these parts. For dynamic objects, the prior depth may or may not be optimal based on the quality of the estimated depth, and the network learns a suitable blending weight. We found a binary cross entropy (BCE) loss on \\( \\alpha \\) improves convergence and pushes it closer to binary:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.839, + 0.891, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\mathrm {B C E}} = \\mathrm {B C E} \\left(\\alpha , \\mathbb {1} \\left(\\| d ^ {t} - g ^ {t} \\| _ {1} < \\| d _ {\\mathrm {p}} ^ {t} - g ^ {t} \\| _ {1}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function of the pixels satisfying the inequality and \\(g^{t}\\) is the ground truth depth. We also" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9122" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.35 + ], + "angle": 0, + "content": "
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MiDaS [33]0.6110.6750.2110.2330.4130.5970.2793.0130.6100.7980.8961
WSVD [45]‡0.6130.7040.2880.3560.4490.6540.3603.7750.4660.7050.8252
DPT [32]0.4240.4930.3200.4170.4820.5390.2242.6780.6860.8640.9321
Ours-DPT†‡0.2550.2950.4890.4700.5590.4740.1972.4000.7100.8850.9421
ScanNet Dataset
ST-CLSTM [56]‡0.0870.0870.3100.1340.4000.0760.2890.7510.7720.9320.9835
TCMD [20]‡0.0310.0310.6280.1660.5480.0830.3000.8110.7400.9270.9811
N-RGB→D [24]†‡0.0430.0420.5570.1690.4820.0880.2850.6740.8140.9520.9831
Demon [43]0.1600.1590.1430.1070.2710.0980.2280.6510.8390.9510.9822
DeepV2D [40]0.0380.0380.4830.1430.4950.0580.1530.5020.9570.9950.9992
DPT [32]0.0330.0330.5400.1410.5360.0550.2130.5820.9710.9960.9991
Ours-DPT†‡0.0110.0100.8630.1790.6390.0520.2100.5760.9740.9960.9991
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.428, + 0.351, + 0.882, + 0.365 + ], + "angle": 0, + "content": "†Method uses temporal data structure. ‡ Method has temporal consistency constraints." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.369, + 0.895, + 0.397 + ], + "angle": 0, + "content": "Table 1. Comparing our method using a DPT backbone to SOTA monocular depth methods. \\(N\\) is the number of frames required at inference. Our method achieves the best results overall with significantly higher consistency, and similar or better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.482 + ], + "angle": 0, + "content": "include an L1 loss \\( \\mathrm{L_G} \\) on the depth gradients, and a VGG loss on the fused depth. We found the latter helps capture fine detail better than an L1 loss alone. Thus, we supervise the temporal fusion network with loss:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.488, + 0.47, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\Theta} = \\left\\| d _ {\\mathrm {f}} ^ {t} - g ^ {t} \\right\\| _ {1} + \\mathrm {L} _ {\\mathrm {B C E}} + \\mathrm {L} _ {\\mathrm {G}} + \\mathrm {V G G} \\left(d _ {\\mathrm {f}} ^ {t}, g ^ {t}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.587 + ], + "angle": 0, + "content": "We weigh the terms as 10, 0.1, 0.05 and 0.05, respectively. We implement the depth fusion network \\(\\Phi(\\cdot)\\) as a four-layer U-Net with instance normalization, and ReLU activations on all layers. We train the network using a self-supervised loss on the uncertainty \\(s^t = \\Phi(d^t, c^t)\\) of the input depth:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.597, + 0.469, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\Phi} = \\frac {1}{m} \\left\\| \\right. \\exp \\left(- s ^ {t}\\right) \\mid d ^ {t} - g ^ {t} \\left. \\right| + \\lambda_ {s} s ^ {t} \\left\\| \\right. _ {1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.469, + 0.689 + ], + "angle": 0, + "content": "where \\( m \\) is the number of pixels and the second term is a regularizer to prevent the trivial \\( s^t = \\infty \\). We use \\( \\lambda_s = 0.03 \\). As Kendall et al. [16] show, this loss term minimizes the log variance of a Gaussian likelihood model for the uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.797 + ], + "angle": 0, + "content": "We train both networks on the FlyingThings3D dataset [27]. This consists of rendered sequences of objects moving along random 3D trajectories, and has ground truth depth and camera poses. For fine-tuning we render 50 60-frame sequences from a custom Blender scene with human-like motion in an indoor setting. We estimate \\(d^{t}\\) with RAFT-Stereo [23] and use the inverse depth for training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.808, + 0.21, + 0.825 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.295, + 0.849 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Both networks are implemented in PyTorch and trained on six Nvidia Tesla V100 GPUs. The temporal fusion network is trained for five epochs on \\(500 \\times 500\\) crops from the" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.419, + 0.63, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.419, + 0.758, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.469, + 0.744, + 0.48 + ], + "angle": 0, + "content": "COLMAP MVS" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.419, + 0.888, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.469, + 0.865, + 0.48 + ], + "angle": 0, + "content": "NeuralRecon" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.49, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Figure 6. MVS and fusion methods like NeuralRecon [38] suffer from holes in the output making them ill-suited to AR applications." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.539, + 0.892, + 0.644 + ], + "angle": 0, + "content": "FlyingThings3D dataset in batches of 18. We use a learning rate of \\(5 \\times 10^{-4}\\) and a weight decay of \\(1 \\times 10^{-4}\\). The spatial fusion network is trained on the depth output of the temporal filtering stage \\(d_{\\mathrm{f}}^{t}\\) using the FlyingThings3D dataset with a similar batch and input size, and a learning rate of \\(1 \\times 10^{-4}\\). We fine-tune both networks on \\(320 \\times 320\\) crops of our Blender data for 14 epochs at a learning rate of \\(1 \\times 10^{-6}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.661, + 0.677 + ], + "angle": 0, + "content": "5.2. Testing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.808 + ], + "angle": 0, + "content": "MPI Sintel: The MPI Sintel dataset [1, 51] consists of long synthetic stereo sequences with large motion and depth range. We evaluate both stereo and monocular depth estimation on the final pass of the 22 \"training scenes\" which have ground truth optical flow, pose, and depth. Given the extremely large motion and depth of some scenes, we follow Teed et al. [42] and restrict our evaluation to pixels with flow \\(\\leq 250\\) pixels and depth \\(\\leq 30\\) meters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ScanNet: The ScanNet dataset [9] consists of video sequences of static indoor scenes captured with an RGB-D sensor and annotated with camera poses. We evaluate the monocular variant of our method on 20 60-frame sequences from different scenes. ScanNet does not have ground truth optical flow, so we estimate it using RAFT [41]." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9123" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.456, + 0.096, + 0.584, + 0.109 + ], + "angle": 0, + "content": "MPI Sintel Dataset" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.112, + 0.898, + 0.22 + ], + "angle": 0, + "content": "
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MC [21]0.5530.6080.2330.4230.4020.6190.3713.7710.4620.6900.8201
CVD-MC [26]0.2260.2450.4020.5150.4600.5890.3383.0780.4930.7020.819All
-59.1%-59.7%+72.5%+21.7%+14.4%-4.85%-8.89%-18.4%+6.71%+1.74%-0.12%
Ours-MC0.3770.3990.3600.4430.4230.5830.3613.7110.4660.6960.8231
-31.8%-34.4%+54.5%+4.73%+5.22%-5.82%-2.70%-1.59%+0.87%+0.87%+0.37%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.229, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Table 2. We compare our approach to CVD using the Mannequin Challenge (MC) backbone for both methods. The percentage change in each metric over the baseline (MC) is listed in color. Our method realizes a large portion of the gains of CVD, even though CVD fine-tunes the backbone network for each input (\\(\\sim\\)20 mins) and requires all frames beforehand. Our method is trained once and runs online." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.283, + 0.887, + 0.427 + ], + "angle": 0, + "content": "
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑
RAFT-Stereo [23]0.4090.5070.6150.4240.6420.7950.23931.260.9150.9440.961
Ours-RAFT-S.0.2680.3400.7010.4920.6740.7450.22721.860.9150.9430.959
-34.5%-32.9%+14.0%+16.0%+4.98%-6.29%-5.02%-30.1%+0.00%-0.11%-0.21%
HITNet [39]0.2250.2640.5990.4410.6410.3410.08712.320.8780.9150.941
Ours-HITNet0.1580.1970.6750.5060.6750.3160.0868.9000.8780.9160.943
-29.8%-25.4%+12.7%+14.7%+5.30%-7.33%-1.15%-27.8%+0.00%+0.11%+0.21%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.436, + 0.893, + 0.465 + ], + "angle": 0, + "content": "Table 3. Evaluating two SOTA stereo methods when used with our point cloud-based fusion approach. The percentage change is listed in color below our result for each metric. Our method leads to significant improvement in temporal consistency and preserves spatial quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.481, + 0.186, + 0.496 + ], + "angle": 0, + "content": "5.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Our choice of baselines is motivated by the use cases we envision for an online method. These include low-latency AR/VR applications [3,52] requiring temporal consistency, completeness (i.e. no holes), and the handling of dynamic objects. Hence, we evaluate stereo depth methods [23,39] on the MPI Sintel dataset, and monocular methods [20,24,32,33,45,56] on both the MPI Sintel and ScanNet datasets. Additionally, we evaluate learned structure-from-motion methods [40,43]. We do not compare with traditional multi-view stereo or TSDF fusion methods as they suffer incompleteness, and fail in dynamic scenes (Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.471, + 0.807 + ], + "angle": 0, + "content": "For each baseline, we use pre-trained models provided by the authors and evaluate it only on the dataset it has originally been tested on — with the only exception being models trained on the NYU-V2 dataset [28]. We evaluate these models on ScanNet instead (NYU-V2 is a similar sensor-captured RGB-D dataset, but does not have ground truth poses). We use the efficient Hybrid variant of Ranftl et al.'s DPT [32] in all experiments, and the five-layer 3D-GAN version of Zhang et al.'s ST-CLSTM [56]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.262, + 0.832 + ], + "angle": 0, + "content": "5.4. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given a sequence of estimated depth maps \\(d^0, d^1, \\ldots, d^n\\) along with color images \\(c^0, c^1, \\ldots, c^n\\) and ground truth \\(g^0, g^1, \\ldots, g^n\\), we evaluate depth estimation quality both temporally and spatially. A common metric for tempo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.893, + 0.558 + ], + "angle": 0, + "content": "ral consistency of video uses the optical flow \\( O_{t \\to t + 1} \\) to evaluate the change in a pixel's depth from frame \\( t \\) to \\( t + 1 \\) [2,18,46]. Let \\( d_w^{t + 1} \\) be the estimated depth \\( d^{t + 1} \\) backward warped into frame \\( t \\) using \\( O_{t \\to t + 1} \\). Then the optical flow-based warping error (OPW) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.568, + 0.892, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathrm {O P W} = \\frac {1}{n - 1} \\sum_ {t} \\| M ^ {t} \\left(d _ {w} ^ {t + 1} - d ^ {t}\\right) \\| _ {1} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.895, + 0.724 + ], + "angle": 0, + "content": "where \\( M^t = \\exp(-\\kappa (c_w^{t+1} - c^t)) \\) is a pixel-wise occlusion mask. Intuitively, OPW is the average meter change in the depth of a point across frames. It is expected to be higher for the outdoor scenes in MPI-Sintel than the indoor ScanNet dataset. Li et al. [20] present a variant of this error known as the relative temporal consistency (RTC) metric, which is shown to align with humans' perception of consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.745, + 0.892, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R T C} = \\frac {1}{m n} \\sum_ {t} \\left\\| \\mathbb {1} \\left(M ^ {t} \\max \\left(\\frac {d _ {w} ^ {t + 1}}{d _ {t}}, \\frac {d _ {t}}{d _ {w} ^ {t + 1}}\\right) < \\tau\\right) \\right\\| _ {0} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\( m \\) is the number of pixels, \\( \\mathbb{1}(\\cdot) \\) is the indicator function of the set of pixels satisfying the enclosed inequality, and the 0-norm measures the number of non-zero pixel values. We set \\( \\kappa = 50 \\) and \\( \\tau = 1.01 \\) for our evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "A disadvantage of OPW and RTC is that both metrics are optimized by the degenerate case \\( d^{t} = k \\forall t \\in \\{0,1,\\dots,n\\} \\) for any constant \\( k \\). Hence we augment OPW and RTC with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9124" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.087, + 0.891, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.398 + ], + "angle": 0, + "content": "Figure 7. The improvement in consistency from our method, visualized as the variation in an epipolar plane image (EPI). An EPI represents a slice of the video in the \\(x - t\\) dimensions for a fixed \\(y\\) (indicated in green, red and blue). We use a high-gradient color-map to better visualize changes in the EPI. Our method reduces high-frequency oscillation along \\(t\\) indicating better frame-to-frame consistency, and preserves the spatial \\((x - y)\\) quality of depth. We omit the axis labels in subsequent figures. Original images © Blender Foundation | www.sintel.org" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.471, + 0.444 + ], + "angle": 0, + "content": "the temporal change consistency (TCC) and the temporal motion consistency(TCM) metrics of Zhang et al. [56]:" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.453, + 0.469, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T C C} = \\frac {1}{n - 1} \\sum_ {t} \\operatorname {S S I M} \\left(\\left| d ^ {t} - d ^ {t + 1} \\right|, \\left| g ^ {t} - g ^ {t + 1} \\right|\\right) \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.49, + 0.469, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T C M} = \\frac {1}{n - 1} \\sum_ {t} \\operatorname {S S I M} \\left(O \\left(d ^ {t}, d ^ {t + 1}\\right), O \\left(g ^ {t}, g ^ {t + 1}\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.469, + 0.641 + ], + "angle": 0, + "content": "where SSIM is the structural self-similarity [47], and \\(O\\) is the optical flow \\(O_{t\\rightarrow t + 1}\\). In addition, we propose the self-consistency (SC) metric which uses the estimated depth \\(d^t\\) to compute \\(d_w^{t + 1}\\) instead of the optical flow. Like OPW, it measures the average meter change in depth frame-to-frame. Finally, as in Long et al. [25] we measure the standard deviation of the L1 error SD(L1) across all frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.734 + ], + "angle": 0, + "content": "For the spatial error we use the standard metrics: relative absolute error (RAE); root mean squared error (RMS); and the percentage of bad pixels \\(\\delta_1\\), \\(\\delta_2\\), \\(\\delta_3\\) at thresholds 1.25, \\(1.25^2\\) and \\(1.25^3\\), respectively [56]. We align the affine-invariant monocular methods for evaluation by using the ground truth to calculate a scale and shift factor [33]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.172, + 0.756 + ], + "angle": 0, + "content": "5.5. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 1 presents the quantitative evaluation of our method and the monocular baselines, with a qualitative comparison in Fig. 8. Across both datasets, our depth has a \\(50\\%\\) lower OPW error than the next best method. Intuitively, this means for an average scene point, the frame-to-frame meter difference in depth is less than half that of the other method. Furthermore, the structural similarity of the temporal change in our depths and ground truth (TCC) is \\(15\\%\\) higher, indicating that dynamic regions are not simply" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.409, + 0.891, + 0.487 + ], + "angle": 0, + "content": "
TCM↑TCC↑OPW↓RAE↓
No Temporal Fusion-11.4%-12.1%+27.7%+65.7%
No Spatial Fusion-7.44%-9.06%+39.6%+17.1%
No Global PC-7.45%-9.57%+39.5%+13.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.497, + 0.893, + 0.553 + ], + "angle": 0, + "content": "Table 4. Ablating our pipeline to evaluate its contribution to the final result. We list the percentage change with respect to Ours-DPT in Table 1 on the dynamic MPI Sintel dataset. For TCM and TCC higher is better, and vice versa for OPW and RAE." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.563, + 0.895, + 0.656 + ], + "angle": 0, + "content": "
Params. (M)GMACsRuntime (ms)
RAFT (20 iters.)5.26309.586.2
Mask-RCNN44.2134.642.4
Our Θ(·)4.4535.047.20
Our Φ(·)4.4437.145.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.666, + 0.893, + 0.709 + ], + "angle": 0, + "content": "Table 5. Run-time and complexity evaluation of our two networks \\(\\Theta (\\cdot)\\) and \\(\\Phi (\\cdot)\\) with a \\(512\\times 512\\) image on an NVIDIA GeForce RTX 3080 GPU. We use the default 20 iterations for RAFT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.87 + ], + "angle": 0, + "content": "averaged across frames. A similar trend is seen in the improved performance of the stereo methods in Table 3 with an improvement in OPW and TCC of around \\(30\\%\\) and \\(5\\%\\) respectively. Fig. 7 provides a qualitative visualization of the stereo results. The improvement from our method is compared to Luo et al.'s CVD [26] in Table 2. CVD runs offline, fine-tuning the backbone network for \\(\\sim 20\\) minutes at inference-time. As such, it represents the high bar for temporal consistency. Our method achieves a good portion of the gains of CVD while being generalizable and online." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Table 5 lists the size and runtime of our two networks, comparing \\(\\Theta (\\cdot)\\) to RAFT [41] and Mask-RCNN [11] — the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9125" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.088, + 0.887, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.588, + 0.893, + 0.618 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison of the baseline monocular methods, and our approach with a DPT [32] backbone. We visualize consistency as variation in the cyan slice along the temporal dimension. Our method is more consistent, and captures finer spatial details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.473, + 0.799 + ], + "angle": 0, + "content": "latter having been used for dynamics filtering [20, 26, 55]. Table 4 show how ablating each stage affects the performance of our method. For No Temporal Fusion we set \\(\\alpha = 0\\) in Eq. 1; No Spatial Fusion uses \\(d_{\\mathrm{o}}^{t} = d_{\\mathrm{f}}^{t}\\); and No Global PC does not use a global point cloud, using the reprojected points from frame \\(t - 1\\) only. The temporal fusion stage has the strongest effect on consistency and accuracy. Setting \\(\\alpha = 0\\) effectively averages the frames via Eq. 2 yielding temporally smooth — albeit erroneous — output. Hence, the drop in TCC but the lower OPW error as the latter is minimized when depth variance across frames is low." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.196, + 0.83 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We presented a method for temporally consistent video depth estimation in online settings. Our approach uses a global point cloud, and learnt image-space fusion to encourage consistency and preserve accuracy. While we assumed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.895, + 0.799 + ], + "angle": 0, + "content": "camera poses and the scale of monocular depth was known, these can also be computed using a point cloud alignment algorithm such as Iterative Closest Point (ICP) [13, 29], or any visual SLAM system [10]. In the interest of generality, our method does not assume any particular depth-estimation algorithm. But the results may be improved by conditioning the estimation of \\( d^{t} \\) on the reprojected depth \\( d_{\\mathrm{p}}^{t} \\). Furthermore, while our approach does not utilize the second view in stereo settings, it may be possible to exploit left-right view consistency in addition to temporal consistency [6]. We leave these directions for future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.826, + 0.669, + 0.843 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "We thank Julia Majors and Thu Nguyen-Phuoc for proofreading the draft. Zhao Dong and Zhaoyang Lv provided useful insights on the technical aspects of the project." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9126" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Yuanzhouhan Cao, Yidong Li, Haokui Zhang, Chao Ren, and Yifan Liu. Learning structure affinity for video depth estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 190-198, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[4] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[5] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Steven D Cochran and Gerard Medioni. 3-d surface description from binocular stereo. IEEE Transactions on Pattern Analysis & Machine Intelligence, 14(10):981-994, 1992. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Middlebury College. Middlebury stereo benchmark. https://vision.middley.edu/stereo/eval3/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Arun CS Kumar, Suchendra M Bhandarkar, and Mukta Prasad. Depthnet: A recurrent neural network architecture for monocular depth prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 283-291, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[10] Jorge Fuentes-Pacheco, José Ruiz-Ascencio, and Juan Manuel Rendon-Mancha. Visual simultaneous localization and mapping: a survey. Artificial intelligence review, 43(1):55–81, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[16] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[19] Damien Lefloch, Tim Weyrich, and Andreas Kolb. Anisotropic point-based fusion. In Proceedings of International Conference on Information Fusion (FUSION), pages 1-9. ISIF, July 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[20] Siyuan Li, Yue Luo, Ye Zhu, Xun Zhao, Yu Li, and Ying Shan. Enforcing temporal consistency in video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[22] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. arXiv preprint arXiv:2111.09337, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[23] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In International Conference on 3D Vision (3DV), 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[24] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural \\(\\mathrm{rgb} \\rightarrow \\mathrm{d}\\) sensing: Depth and uncertainty from a video camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986-10995, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[25] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8258-8267, June 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 2, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9127" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[27] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[28] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[29] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[30] Vaishakh Patil, Wouter Van Gansbeke, Dengxin Dai, and Luc Van Gool. Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters, 5(4):6813-6820, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.43 + ], + "angle": 0, + "content": "[31] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[35] Paul Rosenthal and Lars Linsen. Image-space point cloud rendering. In Proceedings of Computer Graphics International, pages 136-143, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[42] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8375–8384, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[43] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[44] Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin, Rob Romijnders, Nicolas Le Roux, and Ross Goroshin. Impact of aliasing on generalization in deep convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10529-10538, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[45] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[46] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. arXiv preprint arXiv:2208.00380, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[48] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[49] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Neuralfusion: Online depth fusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3162-3172, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[50] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In European conference on computer vision, pages 230–247. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[51] J. Wulff, D. J. Butler, G. B. Stanley, and M. J. Black. Lessons and insights from creating a synthetic optical flow benchmark. In A. Fusiello et al. (Eds.), editor, ECCV Workshop on Unsolved Problems in Optical Flow and Stereo Estimation, Part II, LNCS 7584, pages 168-177. Springer-Verlag, Oct. 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpass through: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9128" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[54] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[55] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[56] Haokui Zhang, Chunhua Shen, Ying Li, Yuzhouhan Cao, Yu Liu, and Youliang Yan. Exploiting temporal consistency for real-time video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.36, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[57] Richard Zhang. Making convolutional networks shift-invariant again. In ICML, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[58] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13116-13125, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9129" + } + ] +] \ No newline at end of file diff --git a/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_origin.pdf b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d48331bfc26c2ce507b533e463b98a8212ebc9de --- /dev/null +++ b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/5039f27e-eabc-410c-b1b5-8a065a9cead0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90fa41643e441914b921e0546cad75d075aebc70769ef7d2a9189e684620502a +size 6850036 diff --git a/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/full.md b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc1f27038386b33cd59b84bf6213a91753206fe9 --- /dev/null +++ b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/full.md @@ -0,0 +1,344 @@ +# Temporally Consistent Online Depth Estimation Using Point-Based Fusion + +Numair Khan + +numairkhan@meta.com + +Eric Penner + +epenner@meta.com + +Douglas Lanman + +douglas.lanman@meta.com + +Lei Xiao + +lei.xiao@meta.com + +# Abstract + +Depth estimation is an important step in many computer vision problems such as 3D reconstruction, novel view synthesis, and computational photography. Most existing work focuses on depth estimation from single frames. When applied to videos, the result lacks temporal consistency, showing flickering and swimming artifacts. In this paper we aim to estimate temporally consistent depth maps of video streams in an online setting. This is a difficult problem as future frames are not available and the method must choose between enforcing consistency and correcting errors from previous estimations. The presence of dynamic objects further complicates the problem. We propose to address these challenges by using a global point cloud that is dynamically updated each frame, along with a learned fusion approach in image space. Our approach encourages consistency while simultaneously allowing updates to handle errors and dynamic objects. Qualitative and quantitative results show that our method achieves state-of-the-art quality for consistent video depth estimation. + +# 1. Introduction + +Depth reconstruction is a long-standing, fundamental problem in computer vision. For decades the most popular depth estimation techniques were based on stereo matching [37] or structure-from-motion [36]. However, more recently the best results have come from learning-based approaches [7]. As the overall reconstruction quality has improved, focus has shifted to new areas such as monocular estimation [21,32,33], edge quality [58], and temporal consistency [20, 26]. The latter is particularly important for video applications in computational photography and virtual reality [3, 52] as inconsistent depth can cause objectionable flickering and swimming artifacts. + +Consistent video depth estimation, however, remains a difficult problem as even the best method will suffer from unpredictable errors and imperfections based on scene content, especially in textureless and specular regions. This difficulty is aggravated by many of the aforementioned applications requiring online reconstruction: future frames are + +![](images/4c59f991eb56e9e396582eac6048e123bef54ac9a771812d77f882e3f43acb06.jpg) +Figure 1. Comparing monocular depth estimation on the ScanNet dataset across four frames. Clockwise from Top-left: Input RGB image, Ranftl et al.'s DPT [32], our result with a DPT backbone, RGB-D sensor ground truth. + +not known beforehand and temporal consistency must be balanced with error-correction. Furthermore, the presence of dynamic objects — which are inherently inconsistent — adds an additional layer of complication. Luo et al. [26] address these problems by assuming all frames are known beforehand and fine-tuning their method for each input video. Other approaches seek an online solution by encoding consistency in network weights either through a training loss [21], by using recurrent architectures [8, 30, 56], or by conditioning it on the input [24]. Each method, however, ultimately relies on the raw output of a neural network being consistent, which is difficult to achieve due to camera noise and aliasing in convolutional networks [14, 44, 57]. + +In this work, we propose the use of a global point cloud to encourage temporal consistency in online video depth estimation. We demonstrate how to tackle the twin problems of handling dynamic objects and updating a static — and potentially erroneous — point cloud when future frames are not known. With quantitative and qualitative results, we show our approach significantly improves the temporal consistency of both stereo and monocular depth estimation without sacrificing spatial quality. In summary, our contributions are as follows: + +![](images/5161cd180a0fc18d6e040a09699a090c5380cf88fb2a5d2031501e2e71a431d5.jpg) +Figure 2. We generate temporally consistent depth maps for each RGB video frame $t$ by fusing the projected depth from a prior point cloud $d_{\mathrm{p}}^{t}$ with the estimated depth $d^{t}$ . This is done by temporally fusing $d_{\mathrm{p}}^{t}$ to update dynamic regions, followed by spatial fusion with $d^{t}$ based on confidence estimates of accuracy. The final result $d_{\mathrm{o}}^{t}$ is used to update the point cloud for the next frame. + +1. We propose point cloud-based fusion for temporally consistent video depth estimation. +2. We present a three-stage approach to encourage consistency in online settings, while simultaneously allowing updates to improve the accuracy of reconstruction and handle dynamic scenes. +3. We present an image-space approach to dynamics estimation and depth fusion that is lightweight and has low runtime overhead. + +# 2. Related Work + +Depth Estimation: The recovery of geometry from color images is an historically important problem in computer vision. In the last decade, deep learning has increasingly been applied to different stages of classic algorithms including geometric [4, 23, 39, 53], photometric [54] and multi-view stereo [12], as well as structure-from-motion (SfM) [50]. More recently, approaches based entirely on learnt priors have allowed high-quality depth estimation from single images [21, 32, 33, 45]. Traditionally, the focus of these methods has been on single frames, with most suffering from temporal artifacts on videos. + +Temporal Consistency: A number of works have recently demonstrated the advantage of using temporal information for improving depth accuracy and consistency [24, 25, 30]. Long et al. [25] propose temporal coherence by fusing the cost volumes of neighboring frames with a transformer. Similarly, Liu et al.'s [24] method accumulates depth probability over consecutive frames using Bayesian filtering. However, the final outcome in both cases is generated by a refinement network which does not have any explicit temporal constraints. Luo et al. [26] fine-tune a monocular depth network at inference-time using constraints from optical flow and structure-from-motion. Kopf et al. [17] build on this to refine camera poses too. Both methods run offline with the inference-time fine-tuning taking around twenty minutes. + +Luo et al. and Kopf et al. are special instances of a more general approach to enforcing consistency in video + +tasks using an optical flow-based warping loss during training [2, 18, 20]. Used in isolation, however, this approach fails to generalize well to unseen data, often requiring fine-tuning for each new dataset. But it does have the advantage of not relying on camera poses or external structures. Lai et al. [18] and Cao et al. [2] pair a warping loss with a convolutional recurrent module which allows the network to learn temporal affinities more effectively. + +In fact, the use of recurrent structures is another way of enforcing temporal consistency in video tasks [8, 30, 56]. Zhang et al. [56] propagate spatio-temporal depth information across frames using a novel convolutional LSTM module which is trained using an adversarial loss. While this loss does not have any explicit temporal constraint like the warping loss, it is nonetheless shown to improve depth consistency. Patil et al.'s [30] work on depth prediction and completion demonstrates that recurrent modules enforce consistency and improve accuracy even without special loss functions. Convolutional LSTMs, however, can be difficult to train due to their large memory requirements. + +Depth Fusion: Fusion methods [5, 29, 48, 49] achieve scene reconstruction by blending weighted signed distance field (SDF) volumes for each frame. The use of SDF volumes makes them scale poorly to large scenes and high resolutions. Keller et al. [15] and Lefloch et al. [19] propose point clouds as an alternative to SDFs. Traditional fusion methods, however, are not directly applicable to depth reconstruction as their focus is geometric reconstruction over multiple frames. Completeness, in the sense of reconstructing each visible point in a frame, is not a requirement. + +# 3. Consistent Video Depth Estimation + +Given a sequence of RGB images $c^t$ with known poses, our goal is to estimate a depth map $d^t$ for each $c^t$ , $t \in \{0,1,2,\ldots\}$ such that the geometric representation of all scene points visible at time $t$ is accurate and consistent from $d^{t-1}$ to $d^t$ . Furthermore, we want to solve this problem online; that is, at any time instant $t$ future frames $d^i$ , $i > t$ are not known. Accuracy subsumes consistency since a 100% accurate reconstruction is also 100% consistent, but + +![](images/3b10e74cb48eeae057957e958ae240b78ca3c18030061875bcddfc4691f7c172.jpg) +Figure 3. The temporal fusion stage generates a 2D mask $\alpha$ for scene regions that change due to motion in frame $t$ . This mask is used to update the reprojected point cloud depth $d_{\mathrm{p}}^{t}$ with dynamic changes from the observed depth $d^{t}$ . + +not vice versa. As a result, the majority of past depth estimation methods have focused on optimizing accuracy of reconstruction [4,23,39,53]. In practice, however, a perfectly accurate reconstruction is hardly ever achievable such that most of these methods suffer from inconsistency. + +A simple way of enforcing temporal consistency is to use the known camera pose to elevate a depth map $d^t$ to a 3D point cloud and then reproject it into future frames. However, such a reprojection-based approach suffers from the following shortcomings: it fails to handle dynamic scenes, it creates holes in disoccluded regions, and it will propagate any errors in $d^t$ to all subsequent frames. + +Our proposed approach seeks to exploit the advantages of a global point cloud while addressing the above-mentioned shortcomings. We do this in three steps: (i) The temporal fusion step (Sec. 3.1) identifies and updates regions of the scene that changed from frame $t - 1$ to $t$ by blending $d^{t}$ with prior depth from the global point cloud in image space. (ii) Spatial fusion (Sec. 3.2) recovers spatial details and corrects any errors in the blended output depth map from the first stage. (iii) Finally, we update the global point cloud to incorporate changes from frame $t$ (Sec. 3.3). + +# 3.1. Temporal Fusion + +Let $\mathcal{P} \in \mathbb{R}^3$ be a prior 3D point cloud representation of a scene at time $t$ generated using the depth input from previous frames. For each $\mathbf{x} \in \mathcal{P}$ let $\varsigma(\mathbf{x}) \in \mathbb{R}^3$ represent the RGB color, and $\rho(\mathbf{x}) \in \mathbb{R}$ represent a confidence estimate of the point. Given the camera pose $\mathbf{T}_t$ and intrinsics matrix $\mathbf{K}$ , we define $d_{\mathrm{p}}^t$ , $c_{\mathrm{p}}^t$ and $w_{\mathrm{p}}^t$ as the $H \times W$ depth, color and confidence projection, respectively, of $\mathcal{P}$ . + +The goal of the temporal fusion stage is to generate a 2D mask $\alpha \in [0,1]^{H\times W}$ for the dynamic regions in $d_{\mathrm{p}}^t$ . Specifically, we want $\alpha$ to identify visible points that change position from frame $t - 1$ to $t$ solely due to motion, and not because of camera movement or uncertainty in the depth estimation. We then compute a fused depth map $d_{\mathrm{f}}^t$ with the dynamic parts updated from $d^t$ while preserving consistent depths from prior frames in the static regions: + +$$ +d _ {\mathrm {f}} ^ {t} = \alpha d ^ {t} + (1 - \alpha) d _ {\mathrm {p}} ^ {t} \tag {1} +$$ + +![](images/2dc6f55ed986b3ac6399514deb0191d3a79a5c6b5043b755dc1fe44fcb82e7cb.jpg) +Figure 4. The spatial fusion stage uses the observed depth $d^t$ to improve static regions in the temporally fused depth $d_{\mathrm{f}}^{t}$ based on confidence in the accuracy of both depths. The confidence is estimated as the learnt uncertainty in the depth prediction model. + +![](images/dc07a5a8645fd7c9a9ded58164c8f401be92de416b2f22aa32870fc34ce20c1a.jpg) +Figure 5. Left to right: The input $c^t$ , the dynamic change mask $\alpha$ (Eq. 1), the confidence estimate $\beta$ (Eq. 4), and the confidence estimate $\gamma$ (Eq. 3). Original image © Blender Foundation | www.sintel.org + +The blending mask $\alpha$ may be computed as the residual between consecutive frames warped using rigid flow. We observe, however, that a hand-crafted threshold for the residual does not generalize to all scenes, and is prone to failure in the presence of noise and depth artifacts. Thus, we use $\alpha = \Theta(d^{t}, d_{\mathrm{p}}^{t}, c^{t}, c_{\mathrm{p}}^{t})$ instead, where $\Theta(\cdot)$ is a convolutional neural network (Fig. 3). We want $\alpha$ to have a strong bias towards the prior depth map in static regions. This is encouraged by using the ground truth — rather than the generated — prior point cloud $\mathcal{P}$ to render $c_{\mathrm{p}}^{t}$ and $d_{\mathrm{p}}^{t}$ during training. Using an L1 loss on the blended output $d_{\mathrm{f}}^{t}$ then ensures that $\alpha = 1$ minimizes the loss in scene regions that do not change from $t - 1$ to $t$ . As observed earlier, training $\Theta(\cdot)$ to directly output $d_{\mathrm{f}}^{t}$ does not guarantee consistency due to CNN aliasing and rotational invariance. + +For the very first frame $t = 0$ , we use $d_{\mathrm{p}}^{t} = d^{t}$ . Subsequently, we render the point cloud $\mathcal{P}$ by splatting each $\mathbf{x} \in \mathcal{P}$ , along with color $\varsigma(\mathbf{x})$ and confidence $\rho(\mathbf{x})$ , to the nearest rounded pixel with a Z-buffer check. While this introduces aliasing artifacts, it is much faster than surfel-based methods [31]. We ameliorate the aliasing through supersampling, and we use Rosenthal and Linsens's method [35] to fill holes and remove background points. + +# 3.2. Spatial Fusion + +Using the ground truth to generate $d_{\mathrm{p}}^{t}$ allows for easily training the temporal fusion stage to maintain consistency in static regions. However, the ground truth is not available at inference-time so that the prior point cloud $\mathcal{P}$ and consequently $d_{\mathrm{p}}^{t}$ is likely to have errors from imperfect depth estimation in previous frames. We propose to address this problem in the spatial fusion stage. + +The goal here is to assign a pixel-wise weight to $d^{t}$ and $d_{\mathrm{f}}^{t}$ based on a confidence estimate of the accuracy of each depth. This is then used to fuse the two depths in image space to generate the final output $d_{\mathrm{o}}^{t}$ : + +$$ +d _ {\mathrm {o}} ^ {t} = \frac {\beta d _ {\mathrm {f}} ^ {t} + \gamma d ^ {t}}{\beta + \gamma} \tag {2} +$$ + +We estimate the confidence weights $\beta, \gamma \in \mathbb{R}^{H \times W}$ using a neural network $\Phi(\cdot)$ (Fig. 4). Note that while formulated similar to Eq. 1, the weights in Eq. 2 serve a different purpose and, hence, require a different training procedure for $\Phi(\cdot)$ . In particular, we represent the weights as the input-dependent aleatoric uncertainty [16] which can be learnt in a self-supervised manner. The confidence $\gamma$ for $d^t$ is then inversely related to the uncertainty as, + +$$ +\gamma = \exp \left[ - \Phi \left(d ^ {t}, c ^ {t}\right) \right] \tag {3} +$$ + +The confidence $\beta$ for $d_{\mathrm{f}}^{t}$ , in addition to the aleatoric uncertainty $\Phi (d_{\mathrm{f}}^{t})$ , also incorporates the prior confidence $w_{\mathrm{p}}^{t}$ weighted by the temporal mask $\alpha$ : + +$$ +\beta = \alpha w _ {\mathrm {p}} ^ {t} \exp \left[ - \Phi \left(d _ {\mathrm {f}} ^ {t}, c ^ {t}\right) \right] \tag {4} +$$ + +We box-filter $w_{\mathrm{p}}^{t}$ before computing $\beta$ to suppress any high-frequency artifacts from the splattering process. Note that $\Phi(\cdot)$ can model uncertainty in $d_{\mathrm{f}}^{t}$ resulting from errors in temporal filtering, so that Eq. 2 is not the same as $\alpha$ -blending $d^{t}$ with $w_{\mathrm{p}}^{t}\exp \left[-\Phi (d_{\mathrm{p}}^{t},c^{t})\right]d_{\mathrm{p}}^{t}$ . + +# 3.3. Global Point Cloud Update + +In a final step, we update the global point cloud $\mathcal{P} \subseteq \mathbb{R}^3$ to reflect changes in the scene from frame $t$ . This includes updating the confidence $\rho(\mathbf{x})$ , and the color $\varsigma(\mathbf{x})$ of all points $\mathbf{x} \in \mathcal{P}$ observed at $t$ . This is similar to the TSDF volume integration step in traditional fusion methods [13, 29]. However, here we have a discrete point cloud representation of the scene which cannot be directly averaged like a TSDF. Therefore, achieving a similar effect involves: (i) associating points in $\mathcal{P}$ with pixels in $d_{\mathrm{o}}^{t}$ , and updating the position, confidence and color of points for which a correspondence is found; (ii) adding pixels from $d^{t}$ for which no association is found as newly observed points into $\mathcal{P}$ ; and (iii) removing points with low confidence from $\mathcal{P}$ . + +Let $z^t \in \mathbb{R}^{H \times W \times 3}$ represent the inverse projection of depth map $d^t$ to 3D. Updates to each $\mathbf{x} \in \mathcal{P}$ are computed in continuous image space to minimize quantization artifacts from using a point cloud. Thus, we update a given $\mathbf{x}$ by sampling $z^t$ and the color image $c^t$ at the projected 2D location $\hat{\mathbf{x}}$ of $\mathbf{x}$ on the image plane. We denote the sampled values as $c^t[\hat{\mathbf{x}}] \in \mathbb{R}^3$ and $z^t[\hat{\mathbf{x}}] \in \mathbb{R}^3$ , respectively, using square brackets to represent bilinear sampling of the image at the indexed location. We only update unoccluded points + +identified by $\alpha[\hat{\mathbf{x}}] < 0.5$ , where $\alpha$ is from Eq. 1. The updated $\mathbf{x}$ , color $\varsigma(\mathbf{x})$ and confidence $\rho(\mathbf{x})$ are computed as, + +$$ +\mathbf {x} = \frac {\beta [ \hat {\mathbf {x}} ] \cdot \mathbf {x} + \gamma [ \hat {\mathbf {x}} ] \cdot z ^ {t} [ \hat {\mathbf {x}} ]}{\beta [ \hat {\mathbf {x}} ] + \gamma [ \hat {\mathbf {x}} ]} \tag {5} +$$ + +$$ +\varsigma (\mathbf {x}) = \frac {\beta [ \hat {\mathbf {x}} ] \cdot \varsigma (\mathbf {x}) + \gamma [ \hat {\mathbf {x}} ] \cdot c ^ {t} [ \hat {\mathbf {x}} ]}{\beta [ \hat {\mathbf {x}} ] + \gamma [ \hat {\mathbf {x}} ]} \tag {6} +$$ + +$$ +\rho (\mathbf {x}) = \beta [ \hat {\mathbf {x}} ] + \gamma [ \hat {\mathbf {x}} ] \tag {7} +$$ + +where $\beta$ , $\gamma \in \mathbb{R}^{H\times W}$ are the fusion weights from Eqs. 3 and 4. For any $\mathbf{x}$ outside the field-of-view, or which is occluded, we set $\rho (\mathbf{x}) = \rho (\mathbf{x}) - 1$ . We identify newly observed points as pixel locations in $z^t$ for which $\alpha \geq 0.5$ , and add them to $\mathcal{P}$ . Finally, we remove all points $\mathbf{x}\in \mathcal{P}$ for which $\rho (\mathbf{x}) < \epsilon$ . We find $\epsilon = 3\times 10^{-2}$ to work well empirically. + +# 4. Training Procedure + +We train the temporal fusion network $\Theta(\cdot)$ using samples $(d^{t}, c^{t}, d_{\mathrm{p}}^{t}, c_{\mathrm{p}}^{t})$ where $(c^{t}, d^{t})$ are the color image and estimated depth of the current frame $t$ , and $(d_{\mathrm{p}}^{t}, c_{\mathrm{p}}^{t})$ are generated by reproducing the ground truth depth and color from frame $t - k$ to $t$ , for some arbitrary $k$ between -7 and 7. We use future frames ( $k < 0$ ) only during training as a data augmentation technique — during inference, our method only sees the current frame $t$ and the global point cloud. + +$\Theta (\cdot)$ consists of two sets of three residual layers followed by a four-layer U-Net [34]. The residual layers extract features from the concatenated depths $\langle d_{\mathrm{p}}^{t},d^{t}\rangle$ and concatenated color images $\langle c_{\mathrm{p}}^{t},c^{t}\rangle$ which are then fed into the U-Net to generate the blending mask $\alpha$ (Fig. 3). This network architecture is much simpler and has fewer parameters than either Mask-RCNN [11] or optical flow methods, both of which have been previously used [22, 26, 55] to mask dynamic elements of a scene (Table 5). We use ReLU activations on all but the last layer of the U-Net, which has a sigmoid. As batch normalization adds a depth-dependent bias [48], we use layer normalization instead. + +The motivation behind using the ground truth rather than the estimated depth for $d_{\mathrm{p}}^{t}$ is to encourage a strong bias towards the prior depth $(\alpha = 0)$ in static regions. This ensures $\Theta(\cdot)$ only identifies changes due to motion. The reprojected ground truth remains optimal in regions that do not change from frame $t - k$ to $t$ . Thus, an L1 loss on $d_{\mathrm{f}}^{t}$ encourages $\alpha = 0$ in these parts. For dynamic objects, the prior depth may or may not be optimal based on the quality of the estimated depth, and the network learns a suitable blending weight. We found a binary cross entropy (BCE) loss on $\alpha$ improves convergence and pushes it closer to binary: + +$$ +\mathrm {L} _ {\mathrm {B C E}} = \mathrm {B C E} \left(\alpha , \mathbb {1} \left(\| d ^ {t} - g ^ {t} \| _ {1} < \| d _ {\mathrm {p}} ^ {t} - g ^ {t} \| _ {1}\right)\right) \tag {8} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function of the pixels satisfying the inequality and $g^{t}$ is the ground truth depth. We also + +
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MiDaS [33]0.6110.6750.2110.2330.4130.5970.2793.0130.6100.7980.8961
WSVD [45]‡0.6130.7040.2880.3560.4490.6540.3603.7750.4660.7050.8252
DPT [32]0.4240.4930.3200.4170.4820.5390.2242.6780.6860.8640.9321
Ours-DPT†‡0.2550.2950.4890.4700.5590.4740.1972.4000.7100.8850.9421
ScanNet Dataset
ST-CLSTM [56]‡0.0870.0870.3100.1340.4000.0760.2890.7510.7720.9320.9835
TCMD [20]‡0.0310.0310.6280.1660.5480.0830.3000.8110.7400.9270.9811
N-RGB→D [24]†‡0.0430.0420.5570.1690.4820.0880.2850.6740.8140.9520.9831
Demon [43]0.1600.1590.1430.1070.2710.0980.2280.6510.8390.9510.9822
DeepV2D [40]0.0380.0380.4830.1430.4950.0580.1530.5020.9570.9950.9992
DPT [32]0.0330.0330.5400.1410.5360.0550.2130.5820.9710.9960.9991
Ours-DPT†‡0.0110.0100.8630.1790.6390.0520.2100.5760.9740.9960.9991
+ +†Method uses temporal data structure. ‡ Method has temporal consistency constraints. + +Table 1. Comparing our method using a DPT backbone to SOTA monocular depth methods. $N$ is the number of frames required at inference. Our method achieves the best results overall with significantly higher consistency, and similar or better accuracy. + +include an L1 loss $\mathrm{L_G}$ on the depth gradients, and a VGG loss on the fused depth. We found the latter helps capture fine detail better than an L1 loss alone. Thus, we supervise the temporal fusion network with loss: + +$$ +\mathrm {L} _ {\Theta} = \left\| d _ {\mathrm {f}} ^ {t} - g ^ {t} \right\| _ {1} + \mathrm {L} _ {\mathrm {B C E}} + \mathrm {L} _ {\mathrm {G}} + \mathrm {V G G} \left(d _ {\mathrm {f}} ^ {t}, g ^ {t}\right) \tag {9} +$$ + +We weigh the terms as 10, 0.1, 0.05 and 0.05, respectively. We implement the depth fusion network $\Phi(\cdot)$ as a four-layer U-Net with instance normalization, and ReLU activations on all layers. We train the network using a self-supervised loss on the uncertainty $s^t = \Phi(d^t, c^t)$ of the input depth: + +$$ +\mathrm {L} _ {\Phi} = \frac {1}{m} \left\| \right. \exp \left(- s ^ {t}\right) \mid d ^ {t} - g ^ {t} \left. \right| + \lambda_ {s} s ^ {t} \left\| \right. _ {1} \tag {10} +$$ + +where $m$ is the number of pixels and the second term is a regularizer to prevent the trivial $s^t = \infty$ . We use $\lambda_s = 0.03$ . As Kendall et al. [16] show, this loss term minimizes the log variance of a Gaussian likelihood model for the uncertainty. + +We train both networks on the FlyingThings3D dataset [27]. This consists of rendered sequences of objects moving along random 3D trajectories, and has ground truth depth and camera poses. For fine-tuning we render 50 60-frame sequences from a custom Blender scene with human-like motion in an indoor setting. We estimate $d^{t}$ with RAFT-Stereo [23] and use the inverse depth for training. + +# 5. Experiments + +# 5.1. Implementation Details + +Both networks are implemented in PyTorch and trained on six Nvidia Tesla V100 GPUs. The temporal fusion network is trained for five epochs on $500 \times 500$ crops from the + +![](images/827895ef8ab15ec6a29c5d68b71735e32957e69be3c365a767d451294f75b57f.jpg) + +![](images/f8d501c75ed02f815b23c5748ba42744ba17616b429e0bd8a1c095b8f2a6f7fc.jpg) +COLMAP MVS +Figure 6. MVS and fusion methods like NeuralRecon [38] suffer from holes in the output making them ill-suited to AR applications. + +![](images/382439c940b7df1644d1daedba32de6769c443bc372d66bd3eb1de440ba284ba.jpg) +NeuralRecon + +FlyingThings3D dataset in batches of 18. We use a learning rate of $5 \times 10^{-4}$ and a weight decay of $1 \times 10^{-4}$ . The spatial fusion network is trained on the depth output of the temporal filtering stage $d_{\mathrm{f}}^{t}$ using the FlyingThings3D dataset with a similar batch and input size, and a learning rate of $1 \times 10^{-4}$ . We fine-tune both networks on $320 \times 320$ crops of our Blender data for 14 epochs at a learning rate of $1 \times 10^{-6}$ . + +# 5.2. Testing Datasets + +MPI Sintel: The MPI Sintel dataset [1, 51] consists of long synthetic stereo sequences with large motion and depth range. We evaluate both stereo and monocular depth estimation on the final pass of the 22 "training scenes" which have ground truth optical flow, pose, and depth. Given the extremely large motion and depth of some scenes, we follow Teed et al. [42] and restrict our evaluation to pixels with flow $\leq 250$ pixels and depth $\leq 30$ meters. + +ScanNet: The ScanNet dataset [9] consists of video sequences of static indoor scenes captured with an RGB-D sensor and annotated with camera poses. We evaluate the monocular variant of our method on 20 60-frame sequences from different scenes. ScanNet does not have ground truth optical flow, so we estimate it using RAFT [41]. + +MPI Sintel Dataset + +
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MC [21]0.5530.6080.2330.4230.4020.6190.3713.7710.4620.6900.8201
CVD-MC [26]0.2260.2450.4020.5150.4600.5890.3383.0780.4930.7020.819All
-59.1%-59.7%+72.5%+21.7%+14.4%-4.85%-8.89%-18.4%+6.71%+1.74%-0.12%
Ours-MC0.3770.3990.3600.4430.4230.5830.3613.7110.4660.6960.8231
-31.8%-34.4%+54.5%+4.73%+5.22%-5.82%-2.70%-1.59%+0.87%+0.87%+0.37%
+ +Table 2. We compare our approach to CVD using the Mannequin Challenge (MC) backbone for both methods. The percentage change in each metric over the baseline (MC) is listed in color. Our method realizes a large portion of the gains of CVD, even though CVD fine-tunes the backbone network for each input ( $\sim$ 20 mins) and requires all frames beforehand. Our method is trained once and runs online. + +
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑
RAFT-Stereo [23]0.4090.5070.6150.4240.6420.7950.23931.260.9150.9440.961
Ours-RAFT-S.0.2680.3400.7010.4920.6740.7450.22721.860.9150.9430.959
-34.5%-32.9%+14.0%+16.0%+4.98%-6.29%-5.02%-30.1%+0.00%-0.11%-0.21%
HITNet [39]0.2250.2640.5990.4410.6410.3410.08712.320.8780.9150.941
Ours-HITNet0.1580.1970.6750.5060.6750.3160.0868.9000.8780.9160.943
-29.8%-25.4%+12.7%+14.7%+5.30%-7.33%-1.15%-27.8%+0.00%+0.11%+0.21%
+ +Table 3. Evaluating two SOTA stereo methods when used with our point cloud-based fusion approach. The percentage change is listed in color below our result for each metric. Our method leads to significant improvement in temporal consistency and preserves spatial quality. + +# 5.3. Baselines + +Our choice of baselines is motivated by the use cases we envision for an online method. These include low-latency AR/VR applications [3,52] requiring temporal consistency, completeness (i.e. no holes), and the handling of dynamic objects. Hence, we evaluate stereo depth methods [23,39] on the MPI Sintel dataset, and monocular methods [20,24,32,33,45,56] on both the MPI Sintel and ScanNet datasets. Additionally, we evaluate learned structure-from-motion methods [40,43]. We do not compare with traditional multi-view stereo or TSDF fusion methods as they suffer incompleteness, and fail in dynamic scenes (Fig. 6). + +For each baseline, we use pre-trained models provided by the authors and evaluate it only on the dataset it has originally been tested on — with the only exception being models trained on the NYU-V2 dataset [28]. We evaluate these models on ScanNet instead (NYU-V2 is a similar sensor-captured RGB-D dataset, but does not have ground truth poses). We use the efficient Hybrid variant of Ranftl et al.'s DPT [32] in all experiments, and the five-layer 3D-GAN version of Zhang et al.'s ST-CLSTM [56]. + +# 5.4. Evaluation Metrics + +Given a sequence of estimated depth maps $d^0, d^1, \ldots, d^n$ along with color images $c^0, c^1, \ldots, c^n$ and ground truth $g^0, g^1, \ldots, g^n$ , we evaluate depth estimation quality both temporally and spatially. A common metric for tempo + +ral consistency of video uses the optical flow $O_{t \to t + 1}$ to evaluate the change in a pixel's depth from frame $t$ to $t + 1$ [2,18,46]. Let $d_w^{t + 1}$ be the estimated depth $d^{t + 1}$ backward warped into frame $t$ using $O_{t \to t + 1}$ . Then the optical flow-based warping error (OPW) is defined as: + +$$ +\mathrm {O P W} = \frac {1}{n - 1} \sum_ {t} \| M ^ {t} \left(d _ {w} ^ {t + 1} - d ^ {t}\right) \| _ {1} \tag {11} +$$ + +where $M^t = \exp(-\kappa (c_w^{t+1} - c^t))$ is a pixel-wise occlusion mask. Intuitively, OPW is the average meter change in the depth of a point across frames. It is expected to be higher for the outdoor scenes in MPI-Sintel than the indoor ScanNet dataset. Li et al. [20] present a variant of this error known as the relative temporal consistency (RTC) metric, which is shown to align with humans' perception of consistency: + +$$ +\mathrm {R T C} = \frac {1}{m n} \sum_ {t} \left\| \mathbb {1} \left(M ^ {t} \max \left(\frac {d _ {w} ^ {t + 1}}{d _ {t}}, \frac {d _ {t}}{d _ {w} ^ {t + 1}}\right) < \tau\right) \right\| _ {0} \tag {12} +$$ + +where $m$ is the number of pixels, $\mathbb{1}(\cdot)$ is the indicator function of the set of pixels satisfying the enclosed inequality, and the 0-norm measures the number of non-zero pixel values. We set $\kappa = 50$ and $\tau = 1.01$ for our evaluation. + +A disadvantage of OPW and RTC is that both metrics are optimized by the degenerate case $d^{t} = k \forall t \in \{0,1,\dots,n\}$ for any constant $k$ . Hence we augment OPW and RTC with + +![](images/4269a440ac7f4920fbc0739ccba6c0230fc6300c5f25e98135dee76442dd47c8.jpg) +Figure 7. The improvement in consistency from our method, visualized as the variation in an epipolar plane image (EPI). An EPI represents a slice of the video in the $x - t$ dimensions for a fixed $y$ (indicated in green, red and blue). We use a high-gradient color-map to better visualize changes in the EPI. Our method reduces high-frequency oscillation along $t$ indicating better frame-to-frame consistency, and preserves the spatial $(x - y)$ quality of depth. We omit the axis labels in subsequent figures. Original images © Blender Foundation | www.sintel.org + +the temporal change consistency (TCC) and the temporal motion consistency(TCM) metrics of Zhang et al. [56]: + +$$ +\mathrm {T C C} = \frac {1}{n - 1} \sum_ {t} \operatorname {S S I M} \left(\left| d ^ {t} - d ^ {t + 1} \right|, \left| g ^ {t} - g ^ {t + 1} \right|\right) \tag {13} +$$ + +$$ +\mathrm {T C M} = \frac {1}{n - 1} \sum_ {t} \operatorname {S S I M} \left(O \left(d ^ {t}, d ^ {t + 1}\right), O \left(g ^ {t}, g ^ {t + 1}\right)\right) \tag {14} +$$ + +where SSIM is the structural self-similarity [47], and $O$ is the optical flow $O_{t\rightarrow t + 1}$ . In addition, we propose the self-consistency (SC) metric which uses the estimated depth $d^t$ to compute $d_w^{t + 1}$ instead of the optical flow. Like OPW, it measures the average meter change in depth frame-to-frame. Finally, as in Long et al. [25] we measure the standard deviation of the L1 error SD(L1) across all frames. + +For the spatial error we use the standard metrics: relative absolute error (RAE); root mean squared error (RMS); and the percentage of bad pixels $\delta_1$ , $\delta_2$ , $\delta_3$ at thresholds 1.25, $1.25^2$ and $1.25^3$ , respectively [56]. We align the affine-invariant monocular methods for evaluation by using the ground truth to calculate a scale and shift factor [33]. + +# 5.5. Results + +Table 1 presents the quantitative evaluation of our method and the monocular baselines, with a qualitative comparison in Fig. 8. Across both datasets, our depth has a $50\%$ lower OPW error than the next best method. Intuitively, this means for an average scene point, the frame-to-frame meter difference in depth is less than half that of the other method. Furthermore, the structural similarity of the temporal change in our depths and ground truth (TCC) is $15\%$ higher, indicating that dynamic regions are not simply + +
TCM↑TCC↑OPW↓RAE↓
No Temporal Fusion-11.4%-12.1%+27.7%+65.7%
No Spatial Fusion-7.44%-9.06%+39.6%+17.1%
No Global PC-7.45%-9.57%+39.5%+13.0%
+ +Table 4. Ablating our pipeline to evaluate its contribution to the final result. We list the percentage change with respect to Ours-DPT in Table 1 on the dynamic MPI Sintel dataset. For TCM and TCC higher is better, and vice versa for OPW and RAE. + +
Params. (M)GMACsRuntime (ms)
RAFT (20 iters.)5.26309.586.2
Mask-RCNN44.2134.642.4
Our Θ(·)4.4535.047.20
Our Φ(·)4.4437.145.10
+ +Table 5. Run-time and complexity evaluation of our two networks $\Theta (\cdot)$ and $\Phi (\cdot)$ with a $512\times 512$ image on an NVIDIA GeForce RTX 3080 GPU. We use the default 20 iterations for RAFT. + +averaged across frames. A similar trend is seen in the improved performance of the stereo methods in Table 3 with an improvement in OPW and TCC of around $30\%$ and $5\%$ respectively. Fig. 7 provides a qualitative visualization of the stereo results. The improvement from our method is compared to Luo et al.'s CVD [26] in Table 2. CVD runs offline, fine-tuning the backbone network for $\sim 20$ minutes at inference-time. As such, it represents the high bar for temporal consistency. Our method achieves a good portion of the gains of CVD while being generalizable and online. + +Table 5 lists the size and runtime of our two networks, comparing $\Theta (\cdot)$ to RAFT [41] and Mask-RCNN [11] — the + +![](images/ace5f4dbfef425fb84a7c156ef1bdee4982e679ae45df1646f7308c88d2cbe46.jpg) +Figure 8. Qualitative comparison of the baseline monocular methods, and our approach with a DPT [32] backbone. We visualize consistency as variation in the cyan slice along the temporal dimension. Our method is more consistent, and captures finer spatial details. + +latter having been used for dynamics filtering [20, 26, 55]. Table 4 show how ablating each stage affects the performance of our method. For No Temporal Fusion we set $\alpha = 0$ in Eq. 1; No Spatial Fusion uses $d_{\mathrm{o}}^{t} = d_{\mathrm{f}}^{t}$ ; and No Global PC does not use a global point cloud, using the reprojected points from frame $t - 1$ only. The temporal fusion stage has the strongest effect on consistency and accuracy. Setting $\alpha = 0$ effectively averages the frames via Eq. 2 yielding temporally smooth — albeit erroneous — output. Hence, the drop in TCC but the lower OPW error as the latter is minimized when depth variance across frames is low. + +# 6. Conclusion + +We presented a method for temporally consistent video depth estimation in online settings. Our approach uses a global point cloud, and learnt image-space fusion to encourage consistency and preserve accuracy. While we assumed + +camera poses and the scale of monocular depth was known, these can also be computed using a point cloud alignment algorithm such as Iterative Closest Point (ICP) [13, 29], or any visual SLAM system [10]. In the interest of generality, our method does not assume any particular depth-estimation algorithm. But the results may be improved by conditioning the estimation of $d^{t}$ on the reprojected depth $d_{\mathrm{p}}^{t}$ . Furthermore, while our approach does not utilize the second view in stereo settings, it may be possible to exploit left-right view consistency in addition to temporal consistency [6]. We leave these directions for future work. + +# Acknowledgements + +We thank Julia Majors and Thu Nguyen-Phuoc for proofreading the draft. Zhao Dong and Zhaoyang Lv provided useful insights on the technical aspects of the project. + +# References + +[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 5 +[2] Yuanzhouhan Cao, Yidong Li, Haokui Zhang, Chao Ren, and Yifan Liu. Learning structure affinity for video depth estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 190-198, 2021. 2, 6 +[3] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 1, 6 +[4] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33, 2020. 2, 3 +[5] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021. 2 +[6] Steven D Cochran and Gerard Medioni. 3-d surface description from binocular stereo. IEEE Transactions on Pattern Analysis & Machine Intelligence, 14(10):981-994, 1992. 8 +[7] Middlebury College. Middlebury stereo benchmark. https://vision.middley.edu/stereo/eval3/.1 +[8] Arun CS Kumar, Suchendra M Bhandarkar, and Mukta Prasad. Depthnet: A recurrent neural network architecture for monocular depth prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 283-291, 2018. 1, 2 +[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5 +[10] Jorge Fuentes-Pacheco, José Ruiz-Ascencio, and Juan Manuel Rendon-Mancha. Visual simultaneous localization and mapping: a survey. Artificial intelligence review, 43(1):55–81, 2015. 8 +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4, 7 +[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[13] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the + +24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 4, 8 +[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 1 +[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2 +[16] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4, 5 +[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018, 2, 6 +[19] Damien Lefloch, Tim Weyrich, and Andreas Kolb. Anisotropic point-based fusion. In Proceedings of International Conference on Information Fusion (FUSION), pages 1-9. ISIF, July 2015. 2 +[20] Siyuan Li, Yue Luo, Ye Zhu, Xun Zhao, Yu Li, and Ying Shan. Enforcing temporal consistency in video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 8 +[21] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 1, 2, 6 +[22] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. arXiv preprint arXiv:2111.09337, 2021. 4 +[23] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In International Conference on 3D Vision (3DV), 2021. 2, 3, 5, 6 +[24] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural $\mathrm{rgb} \rightarrow \mathrm{d}$ sensing: Depth and uncertainty from a video camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986-10995, 2019. 1, 2, 5, 6 +[25] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8258-8267, June 2021. 2, 7 +[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 2, 4, 6, 7, 8 + +[27] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 5 +[28] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6 +[29] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2, 4, 8 +[30] Vaishakh Patil, Wouter Van Gansbeke, Dengxin Dai, and Luc Van Gool. Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters, 5(4):6813-6820, 2020. 1, 2 +[31] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3 +[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1, 2, 5, 6, 8 +[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5, 6, 7 +[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4 +[35] Paul Rosenthal and Lars Linsen. Image-space point cloud rendering. In Proceedings of Computer Graphics International, pages 136-143, 2008. 3 +[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1 +[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 5 +[39] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 2, 3, 6 +[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 5, 6 +[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 5, 7 +[42] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8375–8384, 2021. 5 +[43] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 5, 6 +[44] Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin, Rob Romijnders, Nicolas Le Roux, and Ross Goroshin. Impact of aliasing on generalization in deep convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10529-10538, 2021. 1 +[45] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 2, 5, 6 +[46] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. arXiv preprint arXiv:2208.00380, 2022. 6 +[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[48] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 4 +[49] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Neuralfusion: Online depth fusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3162-3172, June 2021. 2 +[50] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In European conference on computer vision, pages 230–247. Springer, 2020. 2 +[51] J. Wulff, D. J. Butler, G. B. Stanley, and M. J. Black. Lessons and insights from creating a synthetic optical flow benchmark. In A. Fusiello et al. (Eds.), editor, ECCV Workshop on Unsolved Problems in Optical Flow and Stereo Estimation, Part II, LNCS 7584, pages 168-177. Springer-Verlag, Oct. 2012. 5 +[52] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpass through: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 1, 6 + +[53] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 2, 3 +[54] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2 +[55] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 4, 8 +[56] Haokui Zhang, Chunhua Shen, Ying Li, Yuzhouhan Cao, Yu Liu, and Youliang Yan. Exploiting temporal consistency for real-time video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1, 2, 5, 6, 7 +[57] Richard Zhang. Making convolutional networks shift-invariant again. In ICML, 2019. 1 +[58] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13116-13125, 2020. 1 \ No newline at end of file diff --git a/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/images.zip b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b469a334ad6ee2a3f122eafa5da8a9da13dcc292 --- /dev/null +++ b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa84f45d1989218a008a72720aa47c515e04f6526fa07b0fc809a490d0b1803f +size 893151 diff --git a/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/layout.json b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6c7e18f8e00c1f30af6167cabc9186e97623beb3 --- /dev/null +++ b/2023/Temporally Consistent Online Depth Estimation Using Point-Based Fusion/layout.json @@ -0,0 +1,11505 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "Temporally Consistent Online Depth Estimation Using Point-Based Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 144, + 143, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 144, + 143, + 156 + ], + "spans": [ + { + "bbox": [ + 74, + 144, + 143, + 156 + ], + "type": "text", + "content": "Numair Khan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 160, + 161, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 161, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 161, + 171 + ], + "type": "text", + "content": "numairkhan@meta.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 144, + 257, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 144, + 257, + 156 + ], + "spans": [ + { + "bbox": [ + 197, + 144, + 257, + 156 + ], + "type": "text", + "content": "Eric Penner" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 160, + 271, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 160, + 271, + 171 + ], + "spans": [ + { + "bbox": [ + 182, + 160, + 271, + 171 + ], + "type": "text", + "content": "epenner@meta.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 398, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 398, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 398, + 157 + ], + "type": "text", + "content": "Douglas Lanman" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 160, + 419, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 160, + 419, + 171 + ], + "spans": [ + { + "bbox": [ + 293, + 160, + 419, + 171 + ], + "type": "text", + "content": "douglas.lanman@meta.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 466, + 144, + 512, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 144, + 512, + 156 + ], + "spans": [ + { + "bbox": [ + 466, + 144, + 512, + 156 + ], + "type": "text", + "content": "Lei Xiao" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 441, + 160, + 534, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 160, + 534, + 171 + ], + "spans": [ + { + "bbox": [ + 441, + 160, + 534, + 171 + ], + "type": "text", + "content": "lei.xiao@meta.com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 224, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 224, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 45, + 224, + 290, + 453 + ], + "type": "text", + "content": "Depth estimation is an important step in many computer vision problems such as 3D reconstruction, novel view synthesis, and computational photography. Most existing work focuses on depth estimation from single frames. When applied to videos, the result lacks temporal consistency, showing flickering and swimming artifacts. In this paper we aim to estimate temporally consistent depth maps of video streams in an online setting. This is a difficult problem as future frames are not available and the method must choose between enforcing consistency and correcting errors from previous estimations. The presence of dynamic objects further complicates the problem. We propose to address these challenges by using a global point cloud that is dynamically updated each frame, along with a learned fusion approach in image space. Our approach encourages consistency while simultaneously allowing updates to handle errors and dynamic objects. Qualitative and quantitative results show that our method achieves state-of-the-art quality for consistent video depth estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 476, + 128, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 128, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 128, + 489 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "content": "Depth reconstruction is a long-standing, fundamental problem in computer vision. For decades the most popular depth estimation techniques were based on stereo matching [37] or structure-from-motion [36]. However, more recently the best results have come from learning-based approaches [7]. As the overall reconstruction quality has improved, focus has shifted to new areas such as monocular estimation [21,32,33], edge quality [58], and temporal consistency [20, 26]. The latter is particularly important for video applications in computational photography and virtual reality [3, 52] as inconsistent depth can cause objectionable flickering and swimming artifacts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "Consistent video depth estimation, however, remains a difficult problem as even the best method will suffer from unpredictable errors and imperfections based on scene content, especially in textureless and specular regions. This difficulty is aggravated by many of the aforementioned applications requiring online reconstruction: future frames are" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 198, + 547, + 365 + ], + "blocks": [ + { + "bbox": [ + 312, + 198, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 312, + 198, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 547, + 365 + ], + "type": "image", + "image_path": "4c59f991eb56e9e396582eac6048e123bef54ac9a771812d77f882e3f43acb06.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 373, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 305, + 373, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 547, + 418 + ], + "type": "text", + "content": "Figure 1. Comparing monocular depth estimation on the ScanNet dataset across four frames. Clockwise from Top-left: Input RGB image, Ranftl et al.'s DPT [32], our result with a DPT backbone, RGB-D sensor ground truth." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 590 + ], + "type": "text", + "content": "not known beforehand and temporal consistency must be balanced with error-correction. Furthermore, the presence of dynamic objects — which are inherently inconsistent — adds an additional layer of complication. Luo et al. [26] address these problems by assuming all frames are known beforehand and fine-tuning their method for each input video. Other approaches seek an online solution by encoding consistency in network weights either through a training loss [21], by using recurrent architectures [8, 30, 56], or by conditioning it on the input [24]. Each method, however, ultimately relies on the raw output of a neural network being consistent, which is difficult to achieve due to camera noise and aliasing in convolutional networks [14, 44, 57]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 713 + ], + "type": "text", + "content": "In this work, we propose the use of a global point cloud to encourage temporal consistency in online video depth estimation. We demonstrate how to tackle the twin problems of handling dynamic objects and updating a static — and potentially erroneous — point cloud when future frames are not known. With quantitative and qualitative results, we show our approach significantly improves the temporal consistency of both stereo and monocular depth estimation without sacrificing spatial quality. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9119" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 68, + 511, + 178 + ], + "blocks": [ + { + "bbox": [ + 85, + 68, + 511, + 178 + ], + "lines": [ + { + "bbox": [ + 85, + 68, + 511, + 178 + ], + "spans": [ + { + "bbox": [ + 85, + 68, + 511, + 178 + ], + "type": "image", + "image_path": "5161cd180a0fc18d6e040a09699a090c5380cf88fb2a5d2031501e2e71a431d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": "Figure 2. We generate temporally consistent depth maps for each RGB video frame " + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": " by fusing the projected depth from a prior point cloud " + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": " with the estimated depth " + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": ". This is done by temporally fusing " + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": " to update dynamic regions, followed by spatial fusion with " + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": " based on confidence estimates of accuracy. The final result " + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{o}}^{t}" + }, + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": " is used to update the point cloud for the next frame." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 286, + 343 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 230, + 286, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 286, + 253 + ], + "type": "text", + "content": "1. We propose point cloud-based fusion for temporally consistent video depth estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 286, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 286, + 304 + ], + "type": "text", + "content": "2. We present a three-stage approach to encourage consistency in online settings, while simultaneously allowing updates to improve the accuracy of reconstruction and handle dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 286, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 286, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 286, + 343 + ], + "type": "text", + "content": "3. We present an image-space approach to dynamics estimation and depth fusion that is lightweight and has low runtime overhead." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 356, + 134, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 134, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 134, + 369 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": "Depth Estimation: The recovery of geometry from color images is an historically important problem in computer vision. In the last decade, deep learning has increasingly been applied to different stages of classic algorithms including geometric [4, 23, 39, 53], photometric [54] and multi-view stereo [12], as well as structure-from-motion (SfM) [50]. More recently, approaches based entirely on learnt priors have allowed high-quality depth estimation from single images [21, 32, 33, 45]. Traditionally, the focus of these methods has been on single frames, with most suffering from temporal artifacts on videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": "Temporal Consistency: A number of works have recently demonstrated the advantage of using temporal information for improving depth accuracy and consistency [24, 25, 30]. Long et al. [25] propose temporal coherence by fusing the cost volumes of neighboring frames with a transformer. Similarly, Liu et al.'s [24] method accumulates depth probability over consecutive frames using Bayesian filtering. However, the final outcome in both cases is generated by a refinement network which does not have any explicit temporal constraints. Luo et al. [26] fine-tune a monocular depth network at inference-time using constraints from optical flow and structure-from-motion. Kopf et al. [17] build on this to refine camera poses too. Both methods run offline with the inference-time fine-tuning taking around twenty minutes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Luo et al. and Kopf et al. are special instances of a more general approach to enforcing consistency in video" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 230, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 324 + ], + "type": "text", + "content": "tasks using an optical flow-based warping loss during training [2, 18, 20]. Used in isolation, however, this approach fails to generalize well to unseen data, often requiring fine-tuning for each new dataset. But it does have the advantage of not relying on camera poses or external structures. Lai et al. [18] and Cao et al. [2] pair a warping loss with a convolutional recurrent module which allows the network to learn temporal affinities more effectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 325, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 546, + 468 + ], + "type": "text", + "content": "In fact, the use of recurrent structures is another way of enforcing temporal consistency in video tasks [8, 30, 56]. Zhang et al. [56] propagate spatio-temporal depth information across frames using a novel convolutional LSTM module which is trained using an adversarial loss. While this loss does not have any explicit temporal constraint like the warping loss, it is nonetheless shown to improve depth consistency. Patil et al.'s [30] work on depth prediction and completion demonstrates that recurrent modules enforce consistency and improve accuracy even without special loss functions. Convolutional LSTMs, however, can be difficult to train due to their large memory requirements." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "type": "text", + "content": "Depth Fusion: Fusion methods [5, 29, 48, 49] achieve scene reconstruction by blending weighted signed distance field (SDF) volumes for each frame. The use of SDF volumes makes them scale poorly to large scenes and high resolutions. Keller et al. [15] and Lefloch et al. [19] propose point clouds as an alternative to SDFs. Traditional fusion methods, however, are not directly applicable to depth reconstruction as their focus is geometric reconstruction over multiple frames. Completeness, in the sense of reconstructing each visible point in a frame, is not a requirement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 597, + 500, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 500, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 500, + 611 + ], + "type": "text", + "content": "3. Consistent Video Depth Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Given a sequence of RGB images " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "c^t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " with known poses, our goal is to estimate a depth map " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "c^t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "t \\in \\{0,1,2,\\ldots\\}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " such that the geometric representation of all scene points visible at time " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " is accurate and consistent from " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "d^{t-1}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". Furthermore, we want to solve this problem online; that is, at any time instant " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " future frames " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "d^i" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "i > t" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " are not known. Accuracy subsumes consistency since a 100% accurate reconstruction is also 100% consistent, but" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9120" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 275, + 162 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 275, + 162 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 275, + 162 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 275, + 162 + ], + "type": "image", + "image_path": "3b10e74cb48eeae057957e958ae240b78ca3c18030061875bcddfc4691f7c172.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": "Figure 3. The temporal fusion stage generates a 2D mask " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": " for scene regions that change due to motion in frame " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": ". This mask is used to update the reprojected point cloud depth " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": " with dynamic changes from the observed depth " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 223, + 287, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 282 + ], + "type": "text", + "content": "not vice versa. As a result, the majority of past depth estimation methods have focused on optimizing accuracy of reconstruction [4,23,39,53]. In practice, however, a perfectly accurate reconstruction is hardly ever achievable such that most of these methods suffer from inconsistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "type": "text", + "content": "A simple way of enforcing temporal consistency is to use the known camera pose to elevate a depth map " + }, + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "type": "text", + "content": " to a 3D point cloud and then reproject it into future frames. However, such a reprojection-based approach suffers from the following shortcomings: it fails to handle dynamic scenes, it creates holes in disoccluded regions, and it will propagate any errors in " + }, + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 46, + 283, + 287, + 366 + ], + "type": "text", + "content": " to all subsequent frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "text", + "content": "Our proposed approach seeks to exploit the advantages of a global point cloud while addressing the above-mentioned shortcomings. We do this in three steps: (i) The temporal fusion step (Sec. 3.1) identifies and updates regions of the scene that changed from frame " + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "text", + "content": " by blending " + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "text", + "content": " with prior depth from the global point cloud in image space. (ii) Spatial fusion (Sec. 3.2) recovers spatial details and corrects any errors in the blended output depth map from the first stage. (iii) Finally, we update the global point cloud to incorporate changes from frame " + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 367, + 287, + 486 + ], + "type": "text", + "content": " (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 493, + 149, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 149, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 149, + 506 + ], + "type": "text", + "content": "3.1. Temporal Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " be a prior 3D point cloud representation of a scene at time " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " generated using the depth input from previous frames. For each " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{P}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " let " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\varsigma(\\mathbf{x}) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " represent the RGB color, and " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\rho(\\mathbf{x}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " represent a confidence estimate of the point. Given the camera pose " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_t" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " and intrinsics matrix " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^t" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{p}}^t" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{p}}^t" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": " depth, color and confidence projection, respectively, of " + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 510, + 287, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": "The goal of the temporal fusion stage is to generate a 2D mask " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]^{H\\times W}" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": " for the dynamic regions in " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^t" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": ". Specifically, we want " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": " to identify visible points that change position from frame " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": " solely due to motion, and not because of camera movement or uncertainty in the depth estimation. We then compute a fused depth map " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^t" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": " with the dynamic parts updated from " + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 46, + 595, + 287, + 690 + ], + "type": "text", + "content": " while preserving consistent depths from prior frames in the static regions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 696, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 696, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 118, + 696, + 287, + 711 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {f}} ^ {t} = \\alpha d ^ {t} + (1 - \\alpha) d _ {\\mathrm {p}} ^ {t} \\tag {1}", + "image_path": "15fe7c78b2ee6e055f456574d2f81dfbb32583754c13f097a49fc2044bcfb3cd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 69, + 538, + 180 + ], + "blocks": [ + { + "bbox": [ + 320, + 69, + 538, + 180 + ], + "lines": [ + { + "bbox": [ + 320, + 69, + 538, + 180 + ], + "spans": [ + { + "bbox": [ + 320, + 69, + 538, + 180 + ], + "type": "image", + "image_path": "2dc6f55ed986b3ac6399514deb0191d3a79a5c6b5043b755dc1fe44fcb82e7cb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "type": "text", + "content": "Figure 4. The spatial fusion stage uses the observed depth " + }, + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "type": "text", + "content": " to improve static regions in the temporally fused depth " + }, + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 305, + 188, + 545, + 232 + ], + "type": "text", + "content": " based on confidence in the accuracy of both depths. The confidence is estimated as the learnt uncertainty in the depth prediction model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 244, + 545, + 278 + ], + "blocks": [ + { + "bbox": [ + 307, + 244, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 278 + ], + "type": "image", + "image_path": "dc07a5a8645fd7c9a9ded58164c8f401be92de416b2f22aa32870fc34ce20c1a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "lines": [ + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "text", + "content": "Figure 5. Left to right: The input " + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "inline_equation", + "content": "c^t" + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "text", + "content": ", the dynamic change mask " + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "text", + "content": " (Eq. 1), the confidence estimate " + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "text", + "content": " (Eq. 4), and the confidence estimate " + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 286, + 545, + 317 + ], + "type": "text", + "content": " (Eq. 3). Original image © Blender Foundation | www.sintel.org" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": "The blending mask " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " may be computed as the residual between consecutive frames warped using rigid flow. We observe, however, that a hand-crafted threshold for the residual does not generalize to all scenes, and is prone to failure in the presence of noise and depth artifacts. Thus, we use " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha = \\Theta(d^{t}, d_{\\mathrm{p}}^{t}, c^{t}, c_{\\mathrm{p}}^{t})" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " instead, where " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " is a convolutional neural network (Fig. 3). We want " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " to have a strong bias towards the prior depth map in static regions. This is encouraged by using the ground truth — rather than the generated — prior point cloud " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " to render " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " during training. Using an L1 loss on the blended output " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " then ensures that " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " minimizes the loss in scene regions that do not change from " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": ". As observed earlier, training " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " to directly output " + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 304, + 327, + 545, + 506 + ], + "type": "text", + "content": " does not guarantee consistency due to CNN aliasing and rotational invariance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": "For the very first frame " + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t} = d^{t}" + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": ". Subsequently, we render the point cloud " + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": " by splatting each " + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{P}" + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": ", along with color " + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\varsigma(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": " and confidence " + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\rho(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 506, + 545, + 602 + ], + "type": "text", + "content": ", to the nearest rounded pixel with a Z-buffer check. While this introduces aliasing artifacts, it is much faster than surfel-based methods [31]. We ameliorate the aliasing through supersampling, and we use Rosenthal and Linsens's method [35] to fill holes and remove background points." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 611, + 395, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 395, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 395, + 623 + ], + "type": "text", + "content": "3.2. Spatial Fusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Using the ground truth to generate " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " allows for easily training the temporal fusion stage to maintain consistency in static regions. However, the ground truth is not available at inference-time so that the prior point cloud " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and consequently " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is likely to have errors from imperfect depth estimation in previous frames. We propose to address this problem in the spatial fusion stage." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9121" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "The goal here is to assign a pixel-wise weight to " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": " based on a confidence estimate of the accuracy of each depth. This is then used to fuse the two depths in image space to generate the final output " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{o}}^{t}" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 129, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 129, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 130, + 129, + 287, + 156 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {o}} ^ {t} = \\frac {\\beta d _ {\\mathrm {f}} ^ {t} + \\gamma d ^ {t}}{\\beta + \\gamma} \\tag {2}", + "image_path": "db264e10f0fd45d4fc0905ef77208bf551a8834d73f01cb43f9a6aad8f684472.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "content": "We estimate the confidence weights " + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "content": " using a neural network " + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "content": " (Fig. 4). Note that while formulated similar to Eq. 1, the weights in Eq. 2 serve a different purpose and, hence, require a different training procedure for " + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "content": ". In particular, we represent the weights as the input-dependent aleatoric uncertainty [16] which can be learnt in a self-supervised manner. The confidence " + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 47, + 164, + 287, + 261 + ], + "type": "text", + "content": " is then inversely related to the uncertainty as," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 270, + 287, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 270, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 120, + 270, + 287, + 285 + ], + "type": "interline_equation", + "content": "\\gamma = \\exp \\left[ - \\Phi \\left(d ^ {t}, c ^ {t}\\right) \\right] \\tag {3}", + "image_path": "cac6a4a538ce4122d19cd865f18fb89ef6d9b4ceef518716f482eac3528c7754.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "content": "The confidence " + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "content": ", in addition to the aleatoric uncertainty " + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\Phi (d_{\\mathrm{f}}^{t})" + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "content": ", also incorporates the prior confidence " + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "content": " weighted by the temporal mask " + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 294, + 287, + 330 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 339, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 339, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 110, + 339, + 287, + 354 + ], + "type": "interline_equation", + "content": "\\beta = \\alpha w _ {\\mathrm {p}} ^ {t} \\exp \\left[ - \\Phi \\left(d _ {\\mathrm {f}} ^ {t}, c ^ {t}\\right) \\right] \\tag {4}", + "image_path": "452c51f1856183565b3cf1836befaf93de5522e53682e23383d78a8e84706982.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": "We box-filter " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": " before computing " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": " to suppress any high-frequency artifacts from the splattering process. Note that " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": " can model uncertainty in " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": " resulting from errors in temporal filtering, so that Eq. 2 is not the same as " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": "-blending " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{p}}^{t}\\exp \\left[-\\Phi (d_{\\mathrm{p}}^{t},c^{t})\\right]d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 47, + 363, + 287, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 431, + 197, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 197, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 197, + 443 + ], + "type": "text", + "content": "3.3. Global Point Cloud Update" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": "In a final step, we update the global point cloud " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\subseteq \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": " to reflect changes in the scene from frame " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": ". This includes updating the confidence " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\rho(\\mathbf{x})" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": ", and the color " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\varsigma(\\mathbf{x})" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": " of all points " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{P}" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": " observed at " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": ". This is similar to the TSDF volume integration step in traditional fusion methods [13, 29]. However, here we have a discrete point cloud representation of the scene which cannot be directly averaged like a TSDF. Therefore, achieving a similar effect involves: (i) associating points in " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": " with pixels in " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{o}}^{t}" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": ", and updating the position, confidence and color of points for which a correspondence is found; (ii) adding pixels from " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": " for which no association is found as newly observed points into " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": "; and (iii) removing points with low confidence from " + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z^t \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " represent the inverse projection of depth map " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " to 3D. Updates to each " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{P}" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " are computed in continuous image space to minimize quantization artifacts from using a point cloud. Thus, we update a given " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " by sampling " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z^t" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " and the color image " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "c^t" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " at the projected 2D location " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " on the image plane. We denote the sampled values as " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "c^t[\\hat{\\mathbf{x}}] \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "z^t[\\hat{\\mathbf{x}}] \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 605, + 287, + 713 + ], + "type": "text", + "content": ", respectively, using square brackets to represent bilinear sampling of the image at the indexed location. We only update unoccluded points" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "identified by " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\alpha[\\hat{\\mathbf{x}}] < 0.5" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " is from Eq. 1. The updated " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\varsigma(\\mathbf{x})" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " and confidence " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\rho(\\mathbf{x})" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " are computed as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 110, + 545, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 110, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 373, + 110, + 545, + 137 + ], + "type": "interline_equation", + "content": "\\mathbf {x} = \\frac {\\beta [ \\hat {\\mathbf {x}} ] \\cdot \\mathbf {x} + \\gamma [ \\hat {\\mathbf {x}} ] \\cdot z ^ {t} [ \\hat {\\mathbf {x}} ]}{\\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ]} \\tag {5}", + "image_path": "b96572579261376733aa943de4e51d6ab4b078ed42f2811c68911e3b80300365.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 139, + 545, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 139, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 361, + 139, + 545, + 165 + ], + "type": "interline_equation", + "content": "\\varsigma (\\mathbf {x}) = \\frac {\\beta [ \\hat {\\mathbf {x}} ] \\cdot \\varsigma (\\mathbf {x}) + \\gamma [ \\hat {\\mathbf {x}} ] \\cdot c ^ {t} [ \\hat {\\mathbf {x}} ]}{\\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ]} \\tag {6}", + "image_path": "5195cf2f451df3eb0a0f27c36a10e44c8ccd52b5f4bc819096c38b202c495a87.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 360, + 167, + 545, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 167, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 360, + 167, + 545, + 181 + ], + "type": "interline_equation", + "content": "\\rho (\\mathbf {x}) = \\beta [ \\hat {\\mathbf {x}} ] + \\gamma [ \\hat {\\mathbf {x}} ] \\tag {7}", + "image_path": "82bc7fd6406e33c36655825bfab023729686cc52644ee6bd8a8dcddd08005182.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": " are the fusion weights from Eqs. 3 and 4. For any " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": " outside the field-of-view, or which is occluded, we set " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\rho (\\mathbf{x}) = \\rho (\\mathbf{x}) - 1" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": ". We identify newly observed points as pixel locations in " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "z^t" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\alpha \\geq 0.5" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": ", and add them to " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": ". Finally, we remove all points " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathcal{P}" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\rho (\\mathbf{x}) < \\epsilon" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": ". We find " + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "inline_equation", + "content": "\\epsilon = 3\\times 10^{-2}" + }, + { + "bbox": [ + 305, + 186, + 545, + 259 + ], + "type": "text", + "content": " to work well empirically." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 268, + 421, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 421, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 421, + 282 + ], + "type": "text", + "content": "4. Training Procedure" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": "We train the temporal fusion network " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " using samples " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "(d^{t}, c^{t}, d_{\\mathrm{p}}^{t}, c_{\\mathrm{p}}^{t})" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "(c^{t}, d^{t})" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " are the color image and estimated depth of the current frame " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{p}}^{t}, c_{\\mathrm{p}}^{t})" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " are generated by reproducing the ground truth depth and color from frame " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": ", for some arbitrary " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " between -7 and 7. We use future frames (" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "k < 0" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": ") only during training as a data augmentation technique — during inference, our method only sees the current frame " + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 289, + 545, + 384 + ], + "type": "text", + "content": " and the global point cloud." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\Theta (\\cdot)" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " consists of two sets of three residual layers followed by a four-layer U-Net [34]. The residual layers extract features from the concatenated depths " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\langle d_{\\mathrm{p}}^{t},d^{t}\\rangle" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " and concatenated color images " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\langle c_{\\mathrm{p}}^{t},c^{t}\\rangle" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " which are then fed into the U-Net to generate the blending mask " + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 384, + 545, + 527 + ], + "type": "text", + "content": " (Fig. 3). This network architecture is much simpler and has fewer parameters than either Mask-RCNN [11] or optical flow methods, both of which have been previously used [22, 26, 55] to mask dynamic elements of a scene (Table 5). We use ReLU activations on all but the last layer of the U-Net, which has a sigmoid. As batch normalization adds a depth-dependent bias [48], we use layer normalization instead." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": "The motivation behind using the ground truth rather than the estimated depth for " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " is to encourage a strong bias towards the prior depth " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "(\\alpha = 0)" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " in static regions. This ensures " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " only identifies changes due to motion. The reprojected ground truth remains optimal in regions that do not change from frame " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": ". Thus, an L1 loss on " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " encourages " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " in these parts. For dynamic objects, the prior depth may or may not be optimal based on the quality of the estimated depth, and the network learns a suitable blending weight. We found a binary cross entropy (BCE) loss on " + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 528, + 545, + 659 + ], + "type": "text", + "content": " improves convergence and pushes it closer to binary:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 664, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 664, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 318, + 664, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\mathrm {B C E}} = \\mathrm {B C E} \\left(\\alpha , \\mathbb {1} \\left(\\| d ^ {t} - g ^ {t} \\| _ {1} < \\| d _ {\\mathrm {p}} ^ {t} - g ^ {t} \\| _ {1}\\right)\\right) \\tag {8}", + "image_path": "1980cf54ebe19ddd3a6c71650141303e4246872cad2ae866c33309c3a05ed236.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the indicator function of the pixels satisfying the inequality and " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "g^{t}" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the ground truth depth. We also" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9122" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 277 + ], + "type": "table", + "html": "
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MiDaS [33]0.6110.6750.2110.2330.4130.5970.2793.0130.6100.7980.8961
WSVD [45]‡0.6130.7040.2880.3560.4490.6540.3603.7750.4660.7050.8252
DPT [32]0.4240.4930.3200.4170.4820.5390.2242.6780.6860.8640.9321
Ours-DPT†‡0.2550.2950.4890.4700.5590.4740.1972.4000.7100.8850.9421
ScanNet Dataset
ST-CLSTM [56]‡0.0870.0870.3100.1340.4000.0760.2890.7510.7720.9320.9835
TCMD [20]‡0.0310.0310.6280.1660.5480.0830.3000.8110.7400.9270.9811
N-RGB→D [24]†‡0.0430.0420.5570.1690.4820.0880.2850.6740.8140.9520.9831
Demon [43]0.1600.1590.1430.1070.2710.0980.2280.6510.8390.9510.9822
DeepV2D [40]0.0380.0380.4830.1430.4950.0580.1530.5020.9570.9950.9992
DPT [32]0.0330.0330.5400.1410.5360.0550.2130.5820.9710.9960.9991
Ours-DPT†‡0.0110.0100.8630.1790.6390.0520.2100.5760.9740.9960.9991
", + "image_path": "fa7bf8f492d3a835cddf178ff6815ffd373e09b6c60b30f189be138b246eecfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 261, + 277, + 539, + 289 + ], + "lines": [ + { + "bbox": [ + 261, + 277, + 539, + 289 + ], + "spans": [ + { + "bbox": [ + 261, + 277, + 539, + 289 + ], + "type": "text", + "content": "†Method uses temporal data structure. ‡ Method has temporal consistency constraints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 292, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 547, + 314 + ], + "type": "text", + "content": "Table 1. Comparing our method using a DPT backbone to SOTA monocular depth methods. " + }, + { + "bbox": [ + 46, + 292, + 547, + 314 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 292, + 547, + 314 + ], + "type": "text", + "content": " is the number of frames required at inference. Our method achieves the best results overall with significantly higher consistency, and similar or better accuracy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 334, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 381 + ], + "type": "text", + "content": "include an L1 loss " + }, + { + "bbox": [ + 46, + 334, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{L_G}" + }, + { + "bbox": [ + 46, + 334, + 287, + 381 + ], + "type": "text", + "content": " on the depth gradients, and a VGG loss on the fused depth. We found the latter helps capture fine detail better than an L1 loss alone. Thus, we supervise the temporal fusion network with loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 386, + 287, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 287, + 400 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\Theta} = \\left\\| d _ {\\mathrm {f}} ^ {t} - g ^ {t} \\right\\| _ {1} + \\mathrm {L} _ {\\mathrm {B C E}} + \\mathrm {L} _ {\\mathrm {G}} + \\mathrm {V G G} \\left(d _ {\\mathrm {f}} ^ {t}, g ^ {t}\\right) \\tag {9}", + "image_path": "ed63fce1af51dbea784594b455c8eb621a8f387e21a6cf00f7f45fa21ee46ab4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "type": "text", + "content": "We weigh the terms as 10, 0.1, 0.05 and 0.05, respectively. We implement the depth fusion network " + }, + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "type": "text", + "content": " as a four-layer U-Net with instance normalization, and ReLU activations on all layers. We train the network using a self-supervised loss on the uncertainty " + }, + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "type": "inline_equation", + "content": "s^t = \\Phi(d^t, c^t)" + }, + { + "bbox": [ + 46, + 404, + 287, + 464 + ], + "type": "text", + "content": " of the input depth:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 472, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 472, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 82, + 472, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\Phi} = \\frac {1}{m} \\left\\| \\right. \\exp \\left(- s ^ {t}\\right) \\mid d ^ {t} - g ^ {t} \\left. \\right| + \\lambda_ {s} s ^ {t} \\left\\| \\right. _ {1} \\tag {10}", + "image_path": "0d3368a36486a2a7853e4fea0df7b3a0a161c7d357119f1902789c969575e62d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "text", + "content": " is the number of pixels and the second term is a regularizer to prevent the trivial " + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "inline_equation", + "content": "s^t = \\infty" + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\lambda_s = 0.03" + }, + { + "bbox": [ + 46, + 499, + 287, + 545 + ], + "type": "text", + "content": ". As Kendall et al. [16] show, this loss term minimizes the log variance of a Gaussian likelihood model for the uncertainty." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 546, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 631 + ], + "type": "text", + "content": "We train both networks on the FlyingThings3D dataset [27]. This consists of rendered sequences of objects moving along random 3D trajectories, and has ground truth depth and camera poses. For fine-tuning we render 50 60-frame sequences from a custom Blender scene with human-like motion in an indoor setting. We estimate " + }, + { + "bbox": [ + 46, + 546, + 287, + 631 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 46, + 546, + 287, + 631 + ], + "type": "text", + "content": " with RAFT-Stereo [23] and use the inverse depth for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 639, + 128, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 653 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 659, + 180, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 180, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 180, + 672 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Both networks are implemented in PyTorch and trained on six Nvidia Tesla V100 GPUs. The temporal fusion network is trained for five epochs on " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "500 \\times 500" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " crops from the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 331, + 385, + 370 + ], + "blocks": [ + { + "bbox": [ + 307, + 331, + 385, + 370 + ], + "lines": [ + { + "bbox": [ + 307, + 331, + 385, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 385, + 370 + ], + "type": "image", + "image_path": "827895ef8ab15ec6a29c5d68b71735e32957e69be3c365a767d451294f75b57f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 385, + 331, + 463, + 370 + ], + "blocks": [ + { + "bbox": [ + 385, + 331, + 463, + 370 + ], + "lines": [ + { + "bbox": [ + 385, + 331, + 463, + 370 + ], + "spans": [ + { + "bbox": [ + 385, + 331, + 463, + 370 + ], + "type": "image", + "image_path": "f8d501c75ed02f815b23c5748ba42744ba17616b429e0bd8a1c095b8f2a6f7fc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 371, + 455, + 380 + ], + "lines": [ + { + "bbox": [ + 397, + 371, + 455, + 380 + ], + "spans": [ + { + "bbox": [ + 397, + 371, + 455, + 380 + ], + "type": "text", + "content": "COLMAP MVS" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 545, + 411 + ], + "type": "text", + "content": "Figure 6. MVS and fusion methods like NeuralRecon [38] suffer from holes in the output making them ill-suited to AR applications." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 465, + 331, + 543, + 370 + ], + "blocks": [ + { + "bbox": [ + 465, + 331, + 543, + 370 + ], + "lines": [ + { + "bbox": [ + 465, + 331, + 543, + 370 + ], + "spans": [ + { + "bbox": [ + 465, + 331, + 543, + 370 + ], + "type": "image", + "image_path": "382439c940b7df1644d1daedba32de6769c443bc372d66bd3eb1de440ba284ba.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 371, + 529, + 380 + ], + "lines": [ + { + "bbox": [ + 476, + 371, + 529, + 380 + ], + "spans": [ + { + "bbox": [ + 476, + 371, + 529, + 380 + ], + "type": "text", + "content": "NeuralRecon" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": "FlyingThings3D dataset in batches of 18. We use a learning rate of " + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": ". The spatial fusion network is trained on the depth output of the temporal filtering stage " + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": " using the FlyingThings3D dataset with a similar batch and input size, and a learning rate of " + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": ". We fine-tune both networks on " + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": " crops of our Blender data for 14 epochs at a learning rate of " + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 305, + 426, + 545, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 523, + 404, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 404, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 404, + 536 + ], + "type": "text", + "content": "5.2. Testing Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "content": "MPI Sintel: The MPI Sintel dataset [1, 51] consists of long synthetic stereo sequences with large motion and depth range. We evaluate both stereo and monocular depth estimation on the final pass of the 22 \"training scenes\" which have ground truth optical flow, pose, and depth. Given the extremely large motion and depth of some scenes, we follow Teed et al. [42] and restrict our evaluation to pixels with flow " + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\leq 250" + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "content": " pixels and depth " + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\leq 30" + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "content": " meters." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "ScanNet: The ScanNet dataset [9] consists of video sequences of static indoor scenes captured with an RGB-D sensor and annotated with camera poses. We evaluate the monocular variant of our method on 20 60-frame sequences from different scenes. ScanNet does not have ground truth optical flow, so we estimate it using RAFT [41]." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9123" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 88, + 549, + 174 + ], + "blocks": [ + { + "bbox": [ + 279, + 76, + 357, + 86 + ], + "lines": [ + { + "bbox": [ + 279, + 76, + 357, + 86 + ], + "spans": [ + { + "bbox": [ + 279, + 76, + 357, + 86 + ], + "type": "text", + "content": "MPI Sintel Dataset" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 88, + 549, + 174 + ], + "lines": [ + { + "bbox": [ + 48, + 88, + 549, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 88, + 549, + 174 + ], + "type": "table", + "html": "
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑N
MC [21]0.5530.6080.2330.4230.4020.6190.3713.7710.4620.6900.8201
CVD-MC [26]0.2260.2450.4020.5150.4600.5890.3383.0780.4930.7020.819All
-59.1%-59.7%+72.5%+21.7%+14.4%-4.85%-8.89%-18.4%+6.71%+1.74%-0.12%
Ours-MC0.3770.3990.3600.4430.4230.5830.3613.7110.4660.6960.8231
-31.8%-34.4%+54.5%+4.73%+5.22%-5.82%-2.70%-1.59%+0.87%+0.87%+0.37%
", + "image_path": "6db6ba1ff2b0d6ad0b7cfe2c79daf3a8e6b98f879a190499ec0c9443dc2eb851.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 224, + 542, + 338 + ], + "blocks": [ + { + "bbox": [ + 46, + 181, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 547, + 214 + ], + "type": "text", + "content": "Table 2. We compare our approach to CVD using the Mannequin Challenge (MC) backbone for both methods. The percentage change in each metric over the baseline (MC) is listed in color. Our method realizes a large portion of the gains of CVD, even though CVD fine-tunes the backbone network for each input (" + }, + { + "bbox": [ + 46, + 181, + 547, + 214 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 46, + 181, + 547, + 214 + ], + "type": "text", + "content": "20 mins) and requires all frames beforehand. Our method is trained once and runs online." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 224, + 542, + 338 + ], + "lines": [ + { + "bbox": [ + 52, + 224, + 542, + 338 + ], + "spans": [ + { + "bbox": [ + 52, + 224, + 542, + 338 + ], + "type": "table", + "html": "
MPI Sintel Dataset
OPW↓SC↓RTC↑TCM↑TCC↑SD(L1)↓RAE↓RMS↓δ1↑δ2↑δ3↑
RAFT-Stereo [23]0.4090.5070.6150.4240.6420.7950.23931.260.9150.9440.961
Ours-RAFT-S.0.2680.3400.7010.4920.6740.7450.22721.860.9150.9430.959
-34.5%-32.9%+14.0%+16.0%+4.98%-6.29%-5.02%-30.1%+0.00%-0.11%-0.21%
HITNet [39]0.2250.2640.5990.4410.6410.3410.08712.320.8780.9150.941
Ours-HITNet0.1580.1970.6750.5060.6750.3160.0868.9000.8780.9160.943
-29.8%-25.4%+12.7%+14.7%+5.30%-7.33%-1.15%-27.8%+0.00%+0.11%+0.21%
", + "image_path": "8500ed0da8e894bd564a38ca6b6b716c78e01b825375e27d7136af65bd7e632b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 345, + 546, + 368 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 546, + 368 + ], + "type": "text", + "content": "Table 3. Evaluating two SOTA stereo methods when used with our point cloud-based fusion approach. The percentage change is listed in color below our result for each metric. Our method leads to significant improvement in temporal consistency and preserves spatial quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 380, + 113, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 113, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 113, + 392 + ], + "type": "text", + "content": "5.3. Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 531 + ], + "type": "text", + "content": "Our choice of baselines is motivated by the use cases we envision for an online method. These include low-latency AR/VR applications [3,52] requiring temporal consistency, completeness (i.e. no holes), and the handling of dynamic objects. Hence, we evaluate stereo depth methods [23,39] on the MPI Sintel dataset, and monocular methods [20,24,32,33,45,56] on both the MPI Sintel and ScanNet datasets. Additionally, we evaluate learned structure-from-motion methods [40,43]. We do not compare with traditional multi-view stereo or TSDF fusion methods as they suffer incompleteness, and fail in dynamic scenes (Fig. 6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 531, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 288, + 639 + ], + "type": "text", + "content": "For each baseline, we use pre-trained models provided by the authors and evaluate it only on the dataset it has originally been tested on — with the only exception being models trained on the NYU-V2 dataset [28]. We evaluate these models on ScanNet instead (NYU-V2 is a similar sensor-captured RGB-D dataset, but does not have ground truth poses). We use the efficient Hybrid variant of Ranftl et al.'s DPT [32] in all experiments, and the five-layer 3D-GAN version of Zhang et al.'s ST-CLSTM [56]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 160, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 160, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 160, + 658 + ], + "type": "text", + "content": "5.4. Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Given a sequence of estimated depth maps " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "d^0, d^1, \\ldots, d^n" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": " along with color images " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "c^0, c^1, \\ldots, c^n" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "g^0, g^1, \\ldots, g^n" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": ", we evaluate depth estimation quality both temporally and spatially. A common metric for tempo" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": "ral consistency of video uses the optical flow " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "O_{t \\to t + 1}" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " to evaluate the change in a pixel's depth from frame " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " [2,18,46]. Let " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "d_w^{t + 1}" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " be the estimated depth " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "d^{t + 1}" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " backward warped into frame " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "inline_equation", + "content": "O_{t \\to t + 1}" + }, + { + "bbox": [ + 304, + 381, + 546, + 441 + ], + "type": "text", + "content": ". Then the optical flow-based warping error (OPW) is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 449, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 449, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 345, + 449, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\mathrm {O P W} = \\frac {1}{n - 1} \\sum_ {t} \\| M ^ {t} \\left(d _ {w} ^ {t + 1} - d ^ {t}\\right) \\| _ {1} \\tag {11}", + "image_path": "568d12d62ce5f3e333600e196e6876e291584c751f5ad0021d581e969275492b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 547, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 547, + 573 + ], + "type": "inline_equation", + "content": "M^t = \\exp(-\\kappa (c_w^{t+1} - c^t))" + }, + { + "bbox": [ + 304, + 487, + 547, + 573 + ], + "type": "text", + "content": " is a pixel-wise occlusion mask. Intuitively, OPW is the average meter change in the depth of a point across frames. It is expected to be higher for the outdoor scenes in MPI-Sintel than the indoor ScanNet dataset. Li et al. [20] present a variant of this error known as the relative temporal consistency (RTC) metric, which is shown to align with humans' perception of consistency:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 590, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 590, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 310, + 590, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\mathrm {R T C} = \\frac {1}{m n} \\sum_ {t} \\left\\| \\mathbb {1} \\left(M ^ {t} \\max \\left(\\frac {d _ {w} ^ {t + 1}}{d _ {t}}, \\frac {d _ {t}}{d _ {w} ^ {t + 1}}\\right) < \\tau\\right) \\right\\| _ {0} \\tag {12}", + "image_path": "2e4cb3b16349b06fe0c3669eba4a37d254ad814bdde7cc99733f70f0426188cb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "content": " is the number of pixels, " + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "content": " is the indicator function of the set of pixels satisfying the enclosed inequality, and the 0-norm measures the number of non-zero pixel values. We set " + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\kappa = 50" + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\tau = 1.01" + }, + { + "bbox": [ + 304, + 629, + 545, + 676 + ], + "type": "text", + "content": " for our evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "A disadvantage of OPW and RTC is that both metrics are optimized by the degenerate case " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "d^{t} = k \\forall t \\in \\{0,1,\\dots,n\\}" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " for any constant " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": ". Hence we augment OPW and RTC with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9124" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 545, + 268 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 545, + 268 + ], + "type": "image", + "image_path": "4269a440ac7f4920fbc0739ccba6c0230fc6300c5f25e98135dee76442dd47c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "text", + "content": "Figure 7. The improvement in consistency from our method, visualized as the variation in an epipolar plane image (EPI). An EPI represents a slice of the video in the " + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "inline_equation", + "content": "x - t" + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "text", + "content": " dimensions for a fixed " + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "text", + "content": " (indicated in green, red and blue). We use a high-gradient color-map to better visualize changes in the EPI. Our method reduces high-frequency oscillation along " + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "text", + "content": " indicating better frame-to-frame consistency, and preserves the spatial " + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "inline_equation", + "content": "(x - y)" + }, + { + "bbox": [ + 46, + 274, + 547, + 315 + ], + "type": "text", + "content": " quality of depth. We omit the axis labels in subsequent figures. Original images © Blender Foundation | www.sintel.org" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 326, + 288, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 288, + 351 + ], + "type": "text", + "content": "the temporal change consistency (TCC) and the temporal motion consistency(TCM) metrics of Zhang et al. [56]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 358, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 358, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 287, + 386 + ], + "type": "interline_equation", + "content": "\\mathrm {T C C} = \\frac {1}{n - 1} \\sum_ {t} \\operatorname {S S I M} \\left(\\left| d ^ {t} - d ^ {t + 1} \\right|, \\left| g ^ {t} - g ^ {t + 1} \\right|\\right) \\tag {13}", + "image_path": "57e0ba2613f90d20135109d2f01c69dab8c3058eb0dcc4c48bed938fcefd2137.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 388, + 287, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 388, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 52, + 388, + 287, + 416 + ], + "type": "interline_equation", + "content": "\\mathrm {T C M} = \\frac {1}{n - 1} \\sum_ {t} \\operatorname {S S I M} \\left(O \\left(d ^ {t}, d ^ {t + 1}\\right), O \\left(g ^ {t}, g ^ {t + 1}\\right)\\right) \\tag {14}", + "image_path": "c924a221666556be55007b04d0d729391eb3e4eb2ebfd11555e5ff4316f16706.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "text", + "content": "where SSIM is the structural self-similarity [47], and " + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "text", + "content": " is the optical flow " + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "inline_equation", + "content": "O_{t\\rightarrow t + 1}" + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "text", + "content": ". In addition, we propose the self-consistency (SC) metric which uses the estimated depth " + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "inline_equation", + "content": "d^t" + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "inline_equation", + "content": "d_w^{t + 1}" + }, + { + "bbox": [ + 46, + 424, + 287, + 507 + ], + "type": "text", + "content": " instead of the optical flow. Like OPW, it measures the average meter change in depth frame-to-frame. Finally, as in Long et al. [25] we measure the standard deviation of the L1 error SD(L1) across all frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": "For the spatial error we use the standard metrics: relative absolute error (RAE); root mean squared error (RMS); and the percentage of bad pixels " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\delta_1" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\delta_2" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\delta_3" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": " at thresholds 1.25, " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "1.25^2" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "1.25^3" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": ", respectively [56]. We align the affine-invariant monocular methods for evaluation by using the ground truth to calculate a scale and shift factor [33]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 105, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 105, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 105, + 598 + ], + "type": "text", + "content": "5.5. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Table 1 presents the quantitative evaluation of our method and the monocular baselines, with a qualitative comparison in Fig. 8. Across both datasets, our depth has a " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " lower OPW error than the next best method. Intuitively, this means for an average scene point, the frame-to-frame meter difference in depth is less than half that of the other method. Furthermore, the structural similarity of the temporal change in our depths and ground truth (TCC) is " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " higher, indicating that dynamic regions are not simply" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 323, + 545, + 385 + ], + "blocks": [ + { + "bbox": [ + 307, + 323, + 545, + 385 + ], + "lines": [ + { + "bbox": [ + 307, + 323, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 323, + 545, + 385 + ], + "type": "table", + "html": "
TCM↑TCC↑OPW↓RAE↓
No Temporal Fusion-11.4%-12.1%+27.7%+65.7%
No Spatial Fusion-7.44%-9.06%+39.6%+17.1%
No Global PC-7.45%-9.57%+39.5%+13.0%
", + "image_path": "70d5e44ab865d05ea2036838a0042edb4e2ee3ec47dcf6ade3c8a5dbb9d41df0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 445, + 547, + 519 + ], + "blocks": [ + { + "bbox": [ + 305, + 393, + 546, + 437 + ], + "lines": [ + { + "bbox": [ + 305, + 393, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 546, + 437 + ], + "type": "text", + "content": "Table 4. Ablating our pipeline to evaluate its contribution to the final result. We list the percentage change with respect to Ours-DPT in Table 1 on the dynamic MPI Sintel dataset. For TCM and TCC higher is better, and vice versa for OPW and RAE." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 445, + 547, + 519 + ], + "lines": [ + { + "bbox": [ + 306, + 445, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 547, + 519 + ], + "type": "table", + "html": "
Params. (M)GMACsRuntime (ms)
RAFT (20 iters.)5.26309.586.2
Mask-RCNN44.2134.642.4
Our Θ(·)4.4535.047.20
Our Φ(·)4.4437.145.10
", + "image_path": "5e2d8f45647b2c32d65aaf5e845de2fcd154a6b7aa4a92da8a4463333a480caf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "lines": [ + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "text", + "content": "Table 5. Run-time and complexity evaluation of our two networks " + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "inline_equation", + "content": "\\Theta (\\cdot)" + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "inline_equation", + "content": "\\Phi (\\cdot)" + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 305, + 527, + 546, + 561 + ], + "type": "text", + "content": " image on an NVIDIA GeForce RTX 3080 GPU. We use the default 20 iterations for RAFT." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": "averaged across frames. A similar trend is seen in the improved performance of the stereo methods in Table 3 with an improvement in OPW and TCC of around " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " respectively. Fig. 7 provides a qualitative visualization of the stereo results. The improvement from our method is compared to Luo et al.'s CVD [26] in Table 2. CVD runs offline, fine-tuning the backbone network for " + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\sim 20" + }, + { + "bbox": [ + 304, + 569, + 545, + 689 + ], + "type": "text", + "content": " minutes at inference-time. As such, it represents the high bar for temporal consistency. Our method achieves a good portion of the gains of CVD while being generalizable and online." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Table 5 lists the size and runtime of our two networks, comparing " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Theta (\\cdot)" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " to RAFT [41] and Mask-RCNN [11] — the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9125" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 542, + 457 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 542, + 457 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 542, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 542, + 457 + ], + "type": "image", + "image_path": "ace5f4dbfef425fb84a7c156ef1bdee4982e679ae45df1646f7308c88d2cbe46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 465, + 546, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 465, + 546, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 546, + 489 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison of the baseline monocular methods, and our approach with a DPT [32] backbone. We visualize consistency as variation in the cyan slice along the temporal dimension. Our method is more consistent, and captures finer spatial details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "text", + "content": "latter having been used for dynamics filtering [20, 26, 55]. Table 4 show how ablating each stage affects the performance of our method. For No Temporal Fusion we set " + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "text", + "content": " in Eq. 1; No Spatial Fusion uses " + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{o}}^{t} = d_{\\mathrm{f}}^{t}" + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "text", + "content": "; and No Global PC does not use a global point cloud, using the reprojected points from frame " + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "text", + "content": " only. The temporal fusion stage has the strongest effect on consistency and accuracy. Setting " + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 46, + 500, + 289, + 632 + ], + "type": "text", + "content": " effectively averages the frames via Eq. 2 yielding temporally smooth — albeit erroneous — output. Hence, the drop in TCC but the lower OPW error as the latter is minimized when depth variance across frames is low." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We presented a method for temporally consistent video depth estimation in online settings. Our approach uses a global point cloud, and learnt image-space fusion to encourage consistency and preserve accuracy. While we assumed" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "type": "text", + "content": "camera poses and the scale of monocular depth was known, these can also be computed using a point cloud alignment algorithm such as Iterative Closest Point (ICP) [13, 29], or any visual SLAM system [10]. In the interest of generality, our method does not assume any particular depth-estimation algorithm. But the results may be improved by conditioning the estimation of " + }, + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "type": "inline_equation", + "content": "d^{t}" + }, + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "type": "text", + "content": " on the reprojected depth " + }, + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{p}}^{t}" + }, + { + "bbox": [ + 304, + 500, + 547, + 632 + ], + "type": "text", + "content": ". Furthermore, while our approach does not utilize the second view in stereo settings, it may be possible to exploit left-right view consistency in addition to temporal consistency [6]. We leave these directions for future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 654, + 409, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 409, + 667 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 409, + 667 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": "We thank Julia Majors and Thu Nguyen-Phuoc for proofreading the draft. Zhao Dong and Zhaoyang Lv provided useful insights on the technical aspects of the project." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9126" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Yuanzhouhan Cao, Yidong Li, Haokui Zhang, Chao Ren, and Yifan Liu. Learning structure affinity for video depth estimation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 190-198, 2021. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 301 + ], + "type": "text", + "content": "[4] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in Neural Information Processing Systems, 33, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "type": "text", + "content": "[5] Jaesung Choe, Sunghoon Im, Francois Rameau, Minjun Kang, and In So Kweon. Volumefusion: Deep depth fusion for 3d scene reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16086-16095, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "text", + "content": "[6] Steven D Cochran and Gerard Medioni. 3-d surface description from binocular stereo. IEEE Transactions on Pattern Analysis & Machine Intelligence, 14(10):981-994, 1992. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 413 + ], + "type": "text", + "content": "[7] Middlebury College. Middlebury stereo benchmark. https://vision.middley.edu/stereo/eval3/.1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[8] Arun CS Kumar, Suchendra M Bhandarkar, and Mukta Prasad. Depthnet: A recurrent neural network architecture for monocular depth prediction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 283-291, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 523 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 568 + ], + "type": "text", + "content": "[10] Jorge Fuentes-Pacheco, José Ruiz-Ascencio, and Juan Manuel Rendon-Mancha. Visual simultaneous localization and mapping: a survey. Artificial intelligence review, 43(1):55–81, 2015. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 612 + ], + "type": "text", + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "text", + "content": "[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[13] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 4, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Proc. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "text", + "content": "[15] Maik Keller, Damien Lefloch, Martin Lambers, Shahram Izadi, Tim Weyrich, and Andreas Kolb. Real-time 3d reconstruction in dynamic scenes using point-based fusion. In 2013 International Conference on 3D Vision-3DV 2013, pages 1-8. IEEE, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "text", + "content": "[16] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 250 + ], + "type": "text", + "content": "[17] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 304 + ], + "type": "text", + "content": "[18] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 350 + ], + "type": "text", + "content": "[19] Damien Lefloch, Tim Weyrich, and Andreas Kolb. Anisotropic point-based fusion. In Proceedings of International Conference on Information Fusion (FUSION), pages 1-9. ISIF, July 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[20] Siyuan Li, Yue Luo, Ye Zhu, Xun Zhao, Yu Li, and Ying Shan. Enforcing temporal consistency in video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2021. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "text", + "content": "[21] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[22] Zhaoshuo Li, Wei Ye, Dilin Wang, Francis X Creighton, Russell H Taylor, Ganesh Venkatesh, and Mathias Unberath. Temporally consistent online depth estimation in dynamic scenes. arXiv preprint arXiv:2111.09337, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "text", + "content": "[23] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In International Conference on 3D Vision (3DV), 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[24] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural " + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{rgb} \\rightarrow \\mathrm{d}" + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": " sensing: Depth and uncertainty from a video camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986-10995, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[25] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8258-8267, June 2021. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 316, + 757 + ], + "type": "text", + "content": "9127" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[27] N. Mayer, E. Ilg, P. Häusser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2016. arXiv:1512.02134. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 173 + ], + "type": "text", + "content": "[28] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[29] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 2, 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 284 + ], + "type": "text", + "content": "[30] Vaishakh Patil, Wouter Van Gansbeke, Dengxin Dai, and Luc Van Gool. Don't forget the past: Recurrent depth estimation from monocular video. IEEE Robotics and Automation Letters, 5(4):6813-6820, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "type": "text", + "content": "[31] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "text", + "content": "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[33] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 498 + ], + "type": "text", + "content": "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[35] Paul Rosenthal and Lars Linsen. Image-space point cloud rendering. In Proceedings of Computer Graphics International, pages 136-143, 2008. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "text", + "content": "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[38] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[39] Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. Hitnet: Hierarchical iterative tile refinement network for real-time stereo matching. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 14362-14372, 2021. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 162 + ], + "type": "text", + "content": "[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[42] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8375–8384, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 271 + ], + "type": "text", + "content": "[43] Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038-5047, 2017. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[44] Cristina Vasconcelos, Hugo Larochelle, Vincent Dumoulin, Rob Romijnders, Nicolas Le Roux, and Ross Goroshin. Impact of aliasing on generalization in deep convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10529-10538, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "text", + "content": "[45] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[46] Yiran Wang, Zhiyu Pan, Xingyi Li, Zhiguo Cao, Ke Xian, and Jianming Zhang. Less is more: Consistent video depth estimation with masked frames modeling. arXiv preprint arXiv:2208.00380, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[48] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Routedfusion: Learning real-time depth map fusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[49] Silvan Weder, Johannes L. Schonberger, Marc Pollefeys, and Martin R. Oswald. Neuralfusion: Online depth fusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3162-3172, June 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[50] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xi-angyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In European conference on computer vision, pages 230–247. Springer, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 604, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 667 + ], + "type": "text", + "content": "[51] J. Wulff, D. J. Butler, G. B. Stanley, and M. J. Black. Lessons and insights from creating a synthetic optical flow benchmark. In A. Fusiello et al. (Eds.), editor, ECCV Workshop on Unsolved Problems in Optical Flow and Stereo Estimation, Part II, LNCS 7584, pages 168-177. Springer-Verlag, Oct. 2012. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[52] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpass through: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 1, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9128" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 361 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12981-12990, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[54] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 229 + ], + "type": "text", + "content": "[55] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5336-5345, 2020. 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "type": "text", + "content": "[56] Haokui Zhang, Chunhua Shen, Ying Li, Yuzhouhan Cao, Yu Liu, and Youliang Yan. Exploiting temporal consistency for real-time video depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 308 + ], + "type": "text", + "content": "[57] Richard Zhang. Making convolutional networks shift-invariant again. In ICML, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[58] Shengjie Zhu, Garrick Brazil, and Xiaoming Liu. The edge of depth: Explicit constraints between segmentation and depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13116-13125, 2020. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9129" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_content_list.json b/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..720ed78f4f311dfaed869aef9ecffe910477e5cc --- /dev/null +++ b/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "TensoIR: Tensorial Inverse Rendering", + "text_level": 1, + "bbox": [ + 292, + 130, + 679, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haian Jin\\*1 Isabella Liu\\*2 Peijia $\\mathrm{Xu}^3$ Xiaoshuai Zhang2 Songfang Han2 Sai Bi4 Xiaowei Zhou' Zexiang $\\mathrm{Xu}^{\\dagger_4}$ Hao Su", + "bbox": [ + 145, + 178, + 823, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University 2 UC San Diego 3 Kingstar Technology Inc. 4 Adobe Research", + "bbox": [ + 114, + 234, + 849, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose TensoIR, a novel inverse rendering approach based on tensor factorization and neural fields. Unlike previous works that use purely MLP-based neural fields, thus suffering from low capacity and high computation costs, we extend TensoRF, a state-of-the-art approach for radiance field modeling, to estimate scene geometry, surface reflectance, and environment illumination from multiview images captured under unknown lighting conditions. Our approach jointly achieves radiance field reconstruction and physically-based model estimation, leading to photorealistic novel view synthesis and relighting results. Benefiting from the efficiency and extensibility of the TensoRF-based representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under single or multiple unknown lighting conditions. The low-rank tensor representation allows us to not only achieve fast and compact reconstruction but also better exploit shared information under an arbitrary number of capturing lighting conditions. We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging synthetic and real-world scenes.", + "bbox": [ + 73, + 335, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 699, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse rendering is a long-standing problem in computer vision and graphics, aiming to reconstruct physical attributes (like shape and materials) of a 3D scene from captured images and thereby supporting many downstream applications such as novel view synthesis, relighting and material editing. This problem is inherently challenging and ill-posed, especially when the input images are captured in the wild under unknown illumination. Recent works [6,7,28,41] address this problem by learning neural field representations in the form of multi-layer perceptrons", + "bbox": [ + 75, + 724, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/131986b44b1a280cdb70bc2d86761bc81c8ca91ae8d8670c6fc7f6559c4755b3.jpg", + "image_caption": [ + "Figure 1. Given multi-view captured images of a real scene (a), our approach - TensoIR - is able to achieve high-quality shape and material reconstruction with high-frequency details (b). This allows us to render the scene under novel lighting and viewpoints (c), and also change its material properties (d)." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 890, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MLP) similar to NeRF [22]. However, pure MLP-based methods usually suffer from low capacity and high computational costs, greatly limiting the accuracy and efficiency of inverse rendering.", + "bbox": [ + 496, + 594, + 892, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a novel inverse rendering framework that is efficient and accurate. Instead of purely using MLPs, we build upon the recent TensoRF [11] scene representation, which achieves fast, compact, and state-of-the-art quality on radiance field reconstruction for novel view synthesis. Our tensor factorization-based inverse rendering framework can simultaneously estimate scene geometry, materials, and illumination from multi-view images captured under unknown lighting conditions. Benefiting from the efficiency and extensibility of the TensoRF representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under a single or multiple unknown lighting conditions.", + "bbox": [ + 496, + 656, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Similar to TensoRF, our approach models a scene as a neural voxel feature grid, factorized as multiple low-rank", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Equal advisory.", + "bbox": [ + 99, + 886, + 310, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "165", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tensor components. We apply multiple small MLPs on the same feature grid and regress volume density, view-dependent color, normal, and material properties, to model the scene geometry and appearance. This allows us to simultaneously achieve both radiance field rendering – using density and view-dependent color, as done in NeRF [22] – and physically-based rendering – using density, normal and material properties, as done in inverse rendering methods [3, 20]. We supervise both renderings with the captured images to jointly reconstruct all scene components. In essence, we reconstruct a scene using both a radiance field and a physically-based model to reproduce the scene's appearance. While inverse rendering is our focus and primarily enabled by the physically-based model, modeling the radiance field is crucial for the success of the reconstruction (see Fig. 3), in significantly facilitating the volume density reconstruction and effectively regularizing the same tensor features shared by the physically-based model. Despite that previous works [41] similarly reconstruct NeRFs in inverse rendering, their radiance field is pre-computed and fixed in the subsequent inverse rendering stage; in contrast, our radiance field is jointly reconstructed and also benefits the physically-based rendering model estimation during optimization, leading to much higher quality. Besides, our radiance field rendering can also be directly used to provide accurate indirect illumination for the physically-based rendering, further benefiting the inverse rendering process.", + "bbox": [ + 76, + 90, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accounting for indirect illumination and shadowing is a critical challenge in inverse rendering. This is especially challenging for volume rendering, since it requires sampling a lot of secondary rays and computing the integrals along the rays by performing ray marching. Limited by the high-cost MLP evaluation, previous NeRF-based methods and SDF-based methods either simply ignore secondary effects [6, 7, 39], or avoid online computation by approximating these effects in extra distilled MLPs [41, 42], requiring expensive pre-computation and leading to degradation in accuracy. In contrast, owing to our efficient tensor-factorized representation, we are able to explicitly compute the ray integrals online for accurate visibility and indirect lighting with the radiance field rendering using low-cost second-bounce ray marching. Consequently, our approach enables higher accuracy in modeling these secondary effects, which is crucial in achieving high-quality scene reconstruction (see Tab. 2).", + "bbox": [ + 76, + 503, + 472, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, the flexibility and efficiency of our tensor-factorized representation allows us to perform inverse rendering from multiple unknown lighting conditions with limited GPU memory. Multi-light capture is known to be beneficial for inverse rendering tasks by providing useful photometric cues and reducing ambiguities in material estimation, thus being commonly used [13,15,18]. However, since each lighting condition corresponds to a separate radiance", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "field, this can lead to extremely high computational costs if reconstructing multiple purely MLP-based NeRFs like previous works [28, 41, 42]. Instead, we propose to reconstruct radiance fields under multi-light in a joint manner as a factorized tensor. Extending from the original TensorRF representation that is a 4D tensor, we add an additional dimension corresponding to different lighting conditions, yielding a 5D tensor. Specifically, we add an additional vector factor (whose length equals the number of lights) per tensor component to explain the appearance variations under different lighting conditions, and we store this 5D tensor by a small number of bases whose outer-product reconstructs the 5D tensor. When multi-light capture is available, our framework can effectively utilize the additional photometric cues in the data, leading to better reconstruction quality than a single-light setting (see Tab. 1).", + "bbox": [ + 496, + 90, + 893, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and reflectance of a complex real scene captured under unknown natural illumination, enabling photo-realistic rendering under novel lighting conditions and additional applications like material editing. We evaluate our framework extensively on both synthetic and real data. Our approach outperforms previous inverse rendering methods [41, 42] by a large margin qualitatively and quantitatively on challenging synthetic scenes, achieving state-of-the-art quality in scene reconstruction – for both geometry and material properties – and rendering – for both novel view synthesis and relighting. Owing to our efficient tensorial representation and joint reconstruction scheme, our approach also leads to a much faster reconstruction speed than previous neural field-based reconstruction methods while achieving superior quality. In summary,", + "bbox": [ + 496, + 333, + 893, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel tensor factorization-based inverse rendering approach that jointly achieves physically-based rendering model estimation and radiance field reconstruction, leading to state-of-the-art scene reconstruction results;", + "- Our framework includes an efficient online visibility and indirect lighting computation technique, providing accurate second-bounce shading effects;", + "- We enable efficient multi-light reconstruction by modeling an additional lighting dimension in the factorized tensorial representation." + ], + "bbox": [ + 517, + 580, + 890, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 775, + 650, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural scene representations. Neural representations [1, 4, 21, 22, 30, 34, 35], as a promising alternative to traditional representations (like meshes, volumes, and point clouds), have been revolutionizing 3D content generation and modeling. Compared to traditional representations, such neural representations are more flexible and can more", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "166", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22b22337385c57a5fedf6367b1ac1da7d3c2cee834ce11c9b639d90ca8e86789.jpg", + "image_caption": [ + "Figure 2. Overview. We propose a novel inverse rendering approach to reconstruct scene geometry, materials, and unknown natural illumination (as an environment map) from captured images. We reconstruct a scene as a novel representation (Sec. 3.2) that uses factorized tensors and multiple MLPs to regress volume density $\\sigma$ , view-dependent color $c$ , normals $\\mathbf{n}$ , and material properties (i.e. BRDF parameters) $\\beta$ , enabling both radiance field rendering and physically-based rendering (Sec. 3.1). In particular, we march a camera ray from camera origin $\\mathbf{o}$ in viewing direction $\\mathbf{d}$ , sample points $\\mathbf{x}_j$ along the ray, and apply radiance field rendering using the density and view-dependent colors regressed from our representation (Eqn. 1). We also use the volume rendering weights to determine the surface point $\\hat{\\mathbf{x}}$ on the ray (Eqn. 2), at which we perform physically based rendering using the normals and material properties (Eqn. 3). We compute accurate visibility $V$ and indirect lighting $L_{\\mathrm{ind}}$ using radiance field rendering by marching secondary rays from the surface point $\\hat{\\mathbf{x}}$ along sampled incoming light direction $\\omega_i$ (Sec. 3.3), enabling accurate physically-based rendering. We supervise both the radiance field rendering $C_{\\mathrm{RF}}$ and physically-based rendering $C_{\\mathrm{PB}}$ with the captured images in a per-scene optimization for joint scene reconstruction (Sec. 3.5)." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 869, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "accurately reproduce the geometry and appearance of real-world scenes. In particular, NeRF and many following neural field representations [11, 22, 23, 33, 36] have been proposed and applied to enable high-fidelity rendering results on novel view synthesis and many other applications [9, 10, 19, 25, 37]. Originally neural fields [22] are modelled in the form of MLPs, which have limited capacity and high computational costs. Recently, many works [11, 23, 29] introduce more efficient neural scene representations that combine neural feature maps/volumes with light-weight MLPs to reduce the computational cost and accelerate the training and rendering. In this work, we adapt the efficient tensor-factorized neural representation, TensoRF [11], to achieve accurate and efficient inverse rendering.", + "bbox": [ + 75, + 438, + 472, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inverse rendering. Abundant works [5,15-17,24,32] have been proposed to infer the geometry and material properties of real-world objects from image collections. They typically represent the scene geometry using triangle meshes that are known or can be pre-reconstructed with depth sensors or multi-view stereo techniques [27]. To reduce the ambiguities in the inverse rendering, they often require controlled lighting conditions [5, 24], or make use of learnt domain-specific priors [2,5,14,20]. In this work, we jointly estimate the geometry, materials and lighting from images captured under unknown lighting conditions with neural field representation that is more efficient and robust. While neural representations have recently been used for inverse rendering tasks, they [3,6,7,28,38,39,41,42] are limited by the usage of computation-intensive MLPs. This inefficiency adds extra burden on inverse rendering when computing", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "secondary shading effects (like shadows), which requires to extensively sample secondary rays. Therefore, previous methods often ignore secondary effects [6, 7, 39], consider collocated flash lighting [3, 4, 38], or take extra costs to distill these effects into additional MLP networks [28, 41, 42]. In contrast, we base our model on the advanced TensoRF representation, utilizing factorized tensors, to achieve fast reconstruction. Our TensoRF-based approach supports fast density and radiance evaluation, enabling efficient online computation for secondary effects; this leads to more accurate shadow and indirect lighting modeling during reconstruction, further benefiting our reconstruction. Moreover, in contrast to previous methods [28, 39, 41, 42] that can only handle captures under a single lighting condition, our model is easily extended to support multi-light capture by modeling an additional lighting dimension in the tensor factors.", + "bbox": [ + 496, + 438, + 893, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 694, + 591, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our tensor factorization-based inverse rendering framework, shown in Fig. 2, which reconstructs scene geometry, materials, and illumination from multi-view input images under unknown lighting. We leverage rendering with both neural radiance fields and physically-based light transport to model and reproduce scene appearance (Sec. 3.1). We introduce a novel TensoRF-based scene representation that allows for both rendering methods in scene reconstruction (Sec. 3.2). Our representation not only enables accurate and efficient computation for visibility and indirect lighting (Sec. 3.3) but also supports capturing under multiple unknown lighting", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "167", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "conditions (Sec. 3.4). We simultaneously estimate all scene components in a joint optimization framework with rendering losses and regularization terms (Sec. 3.5).", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Rendering", + "text_level": 1, + "bbox": [ + 76, + 146, + 196, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply ray marching to achieve differentiable radiance field rendering as done in NeRF [22]. We further determine the expected surface intersection point for each ray using the volume density, and perform physically-based rendering at the surface point with the predicted scene properties.", + "bbox": [ + 76, + 170, + 468, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Radiance field rendering. Given a camera ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ from ray origin $\\mathbf{o}$ in direction $\\mathbf{d}$ , radiance field rendering samples $N$ points on the ray and compute the pixel color as", + "bbox": [ + 76, + 255, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\mathrm {R F}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {j = 1} ^ {N} T _ {j} \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right) c _ {j} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 311, + 468, + 363 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {j} = \\exp \\left(- \\sum_ {q = 1} ^ {j - 1} \\sigma_ {q} \\delta_ {q}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 359, + 346, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{j},\\delta_{j},c_{j}$ and $T_{j}$ are volume density, step size, viewdependent radiance color, and volume transmittance at each sampled point $\\mathbf{r}(t_j)$ .", + "bbox": [ + 76, + 407, + 468, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physically-based rendering. We apply a physically-based parametric BRDF model [8] $f_{r}$ and perform physically-based rendering using predicted geometry and material properties at surface points on the camera rays. Similar to previous methods [22, 41], these surface points $\\hat{\\mathbf{x}}$ are naturally determined using the volume rendering weights and sampled points from Eqn. 1:", + "bbox": [ + 76, + 462, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\sum_ {j = 1} ^ {N} w _ {j} \\mathbf {r} \\left(t _ {j}\\right), \\quad w _ {j} = T _ {j} \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 580, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We leverage the classic surface rendering equation to compute a physically-based shading color with an integral over the upper hemisphere $\\Omega$ at each surface point:", + "bbox": [ + 76, + 633, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\mathrm {P B}} (\\hat {\\mathbf {x}}, \\mathbf {d}) = \\int_ {\\Omega} L _ {\\mathrm {i}} (\\hat {\\mathbf {x}}, \\boldsymbol {\\omega} _ {i}) f _ {r} (\\hat {\\mathbf {x}}, \\boldsymbol {\\omega} _ {i}, \\mathbf {d}, \\beta) (\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 691, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{i}}(\\hat{\\mathbf{x}},\\omega_{i})$ is the incident illumination coming from direction $\\boldsymbol{\\omega}_{i}$ . $\\beta$ and $\\mathbf{n}$ represent the spatially-varying material parameters (albedo and roughness) of the BRDF and the surface normal at $\\hat{\\mathbf{x}}$ .", + "bbox": [ + 76, + 733, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In theory, accurately evaluating this integral in Eqn. 3 requires extensively sampling the lighting direction and computing the incident lighting $L_{\\mathrm{i}}(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i)$ to account for shadowing and indirect illumination. It requires additional ray marching along each lighting direction, which is known to be extremely computation-expensive. Previous NeRF-based methods often simplify it by only considering direct", + "bbox": [ + 76, + 795, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "illumination [6, 7] or using extra auxiliary MLPs to make approximations and avoid full ray marching [28, 41]. Instead, we evaluate the integral more accurately by marching secondary rays online and computing the incident lighting $L_{\\mathrm{i}}(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i)$ with accurate shadowing and indirect illumination (see Sec. 3.3). This is made possible by our novel efficient TensorRF-based scene representation.", + "bbox": [ + 498, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. TensoRF-Based Representation", + "text_level": 1, + "bbox": [ + 500, + 205, + 777, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now introduce our tensor factorization-based scene representation that simultaneously models volume density $\\sigma$ , view-dependent color $c$ , shading normal $\\mathbf{n}$ , and material properties $\\beta$ (including diffuse albedo and specular roughness of the Disney BRDF [8]). Our method can model the scene under both single or multiple lighting conditions. We first introduce the single-light setting in this section and will discuss the multi-light extension in Sec. 3.4.", + "bbox": [ + 498, + 228, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At a high-level, radiance, geometry, and material properties can all be represented as a 3D field. We can use a feature volume and voxel feature decoding functions to extract information at any point of the 3D field. To compress the space of the feature volume and also regularize the learning process, TensoRF [11] proposed to use a low-rank factorized tensor as the feature volume. In this work, we adopt the Vector-Matrix factorization proposed by TensoRF. In particular, we use two separate VM-factorized tensors, i.e. feature grids $\\mathcal{G}_{\\sigma}$ and $\\mathcal{G}_a$ , to model volume density and appearance, respectively. The appearance tensor $\\mathcal{G}_a$ is followed by multiple light-weight MLP decoders to regress various appearance properties.", + "bbox": [ + 498, + 349, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Density tensor. The density tensor $\\mathcal{G}_{\\sigma}$ is 3D and directly expresses volume density without any network. Specifically, the VM-factorized density tensor is expressed by the sum of vector-matrix outer products.", + "bbox": [ + 498, + 554, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {G} _ {\\sigma} = \\sum_ {k} \\mathbf {v} _ {\\sigma , k} ^ {X} \\circ \\mathbf {M} _ {\\sigma , k} ^ {Y Z} + \\mathbf {v} _ {\\sigma , k} ^ {Y} \\circ \\mathbf {M} _ {\\sigma , k} ^ {X Z} + \\mathbf {v} _ {\\sigma , k} ^ {Z} \\circ \\mathbf {M} _ {\\sigma , k} ^ {X Y} \\\\ = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {\\sigma , k} ^ {m} \\circ \\mathbf {M} _ {\\sigma , k} ^ {\\tilde {m}} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 622, + 890, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{v}_{\\sigma ,k}^{m}$ and $\\mathbf{M}_{\\sigma ,k}^{\\tilde{m}}$ represent the $k^{\\mathrm{th}}$ vector and matrix factors of their corresponding spatial axes $m$ ; for simplicity, $\\tilde{m}$ denotes the two axes orthogonal to $m$ (e.g. $\\tilde{X} = YZ$ ).", + "bbox": [ + 500, + 698, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance tensor. The appearance tensor $\\mathcal{G}_a$ is 4D, modeled by similar vector-matrix spatial factors and additional feature basis vectors $\\mathbf{b}$ , expressing a multi-channel voxel feature grid:", + "bbox": [ + 498, + 753, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {a} = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {a, k} ^ {m} \\circ \\mathbf {M} _ {a, k} ^ {\\tilde {m}} \\circ \\mathbf {b} _ {k} ^ {m} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 821, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Representing scene properties. We obtain density $\\sigma$ directly by linearly interpolating $\\mathcal{G}_{\\sigma}$ , and apply multiple MLP", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "168", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b1eb416dabd1b59ac2785762c28e758d091534ab08ac89ab3bfd1730c9b7e87.jpg", + "image_caption": [ + "Figure 3. We compare normal and albedo reconstruction results between our joint reconstruction model and an ablated model without radiance field rendering during reconstruction. Radiance field reconstruction is crucial for us to achieve good reconstruction with a clean background and reasonable scene geometry." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "networks to decode appearance properties from interpolated appearance features $\\mathcal{G}_a$ . This includes a radiance network $\\mathcal{D}_c$ , a shading normal network $\\mathcal{D}_{\\mathbf{n}}$ , and a material network $\\mathcal{D}_{\\beta}$ . Overall, our scene representation is expressed by", + "bbox": [ + 76, + 299, + 470, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\mathbf {x}} = \\mathcal {G} _ {\\sigma} (\\mathbf {x}), \\quad a _ {\\mathbf {x}} = \\mathcal {G} _ {a} (\\mathbf {x})\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 371, + 367, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} c _ {\\mathbf {x}}, \\mathbf {n} _ {\\mathbf {x}}, \\beta_ {\\mathbf {x}} = \\mathcal {D} _ {c} (a _ {\\mathbf {x}}), \\mathcal {D} _ {\\mathbf {n}} (a _ {\\mathbf {x}}), \\mathcal {D} _ {\\boldsymbol {\\beta}} (a _ {\\mathbf {x}}) \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 382, + 468, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{x}$ denotes an arbitrary 3D location, $\\sigma_{\\mathbf{x}} = \\mathcal{G}_{\\sigma}(\\mathbf{x})$ and $a_{\\mathbf{x}} = \\mathcal{G}_a(\\mathbf{x})$ are the linearly interpolated density and multichannel appearance features, computed by linearly interpolating the spatial vector and matrix factors (please refer to the TensoRF paper [11] for the details of feature computation and interpolation).", + "bbox": [ + 75, + 417, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, volume density $\\sigma$ and shading normals $\\mathbf{n}$ both express scene geometry, which describes global shapes and high-frequency geometric details, respectively. View-dependent color $c$ and physical shading properties (normal $\\mathbf{n}$ and material parameters $\\beta$ ) duplicatively model the scene appearance, determining the colors in the scene. We bind shading normal with volume density using a regularization term (see details in Sec. 3.5), correlating the scene geometry estimation and appearance reasoning.", + "bbox": [ + 75, + 508, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In essence, our TensoRF-based scene representation provides scene geometry and appearance properties that are required for both radiance field rendering (Eqn. 1) and physically-based rendering (Eqn. 3) described in Sec. 3.1. This represents a scene as a radiance field and a physically-based rendering model jointly. We let the two sub-representations share the same neural features in our tensors, allowing their learning processes to benefit each other. While the physically-based model is our main focus and achieves inverse rendering, modeling the radiance field can facilitate the volume density reconstruction and also regularize the appearance features to be meaningful since it has a shorter gradient path. As a result, modeling the radiance field is crucial and necessary for us to achieve high-quality physically-based scene reconstruction as shown in Fig. 3. In addition, the radiance field can naturally provide indirect illumination for physically-based rendering, enabling more", + "bbox": [ + 75, + 643, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d58a0398b89ce63d3de788a2231f9954035b0352afb67426ec82477548c7572.jpg", + "image_caption": [ + "a) Image" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 586, + 142 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/52cfdba5cbc7f95753745483028bb1b7bc010d52b01d5f374d921fafbcabe9bf.jpg", + "image_caption": [ + "b) Indirect lighting" + ], + "image_footnote": [], + "bbox": [ + 586, + 90, + 683, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a234fdc6c2f5b0bee2dfb7bdf4195f62a45ce3b57971839a22942e5d119089d.jpg", + "image_caption": [ + "c) Lighting visibility under two directional lights", + "Figure 4. We show our computed indirect illumination (b) of the full rendered image (a) and our lighting visibility (c) under two different directional lights." + ], + "image_footnote": [], + "bbox": [ + 687, + 90, + 887, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "accurate physical model reconstruction.", + "bbox": [ + 500, + 222, + 763, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Illumination and Visibility", + "text_level": 1, + "bbox": [ + 500, + 244, + 738, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our TensoRF-based scene representation is highly efficient for optimization and evaluation. Especially, our volume density can be computed by simple tensor interpolation without using any MLPs, leading to highly efficient computation of volume transmittance and rendering weights (Eqn. 1,2). This allows us to compute accurate incident illumination $L_{\\mathrm{i}}$ (as defined in Eqn. 3) that accounts for secondary shading effects with ray marching as shown in Fig. 4. In particular, the incident illumination is computed by", + "bbox": [ + 496, + 268, + 890, + 419 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i} (\\hat {\\mathbf {x}}, \\omega_ {i}) = V (\\hat {\\mathbf {x}}, \\omega_ {i}) L _ {\\mathrm {d}} (\\omega_ {i}) + L _ {\\mathrm {i n d}} (\\hat {\\mathbf {x}}, \\omega_ {i}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 430, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V$ is the light visibility function, $L_{\\mathrm{d}}$ is the direct illumination, and $L_{\\mathrm{ind}}$ is the indirect illumination.", + "bbox": [ + 498, + 458, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Direct illumination. We assume the unknown natural environment to be distant from the captured object and represent the global illumination as an environment map, parameterized by a mixture of spherical Gaussians (SG), which represents the direct illumination. Note that in contrast to previous methods [39,42] that use SG for computing the integral of the rendering equation with a closed-form approximation, we use SG only for its compact parameterized representation and compute the integral numerically by sampling secondary rays and performing ray marching, leading to more accurate lighting visibility and indirect lighting.", + "bbox": [ + 496, + 497, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visibility and indirect illumination. We make use of the jointly-trained radiance fields to model secondary shading effects. More specifically, the indirect illumination arriving at $\\hat{\\mathbf{x}}$ from $\\omega_{i}$ is inherently explained by the radiance color $C_{\\mathrm{RF}}$ along the ray $\\mathbf{r}_i(t) = \\hat{\\mathbf{x}} +t\\omega_i$ . The visibility function is exactly modeled by the transmittance function in volume rendering. Therefore, the light visibility and indirect illumination term in Eqn. 7 can be calculated as:", + "bbox": [ + 496, + 672, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV (\\hat {\\mathbf {x}}, \\omega_ {i}) = T (\\hat {\\mathbf {x}}, \\omega_ {i}), \\quad L _ {\\text {i n d}} (\\hat {\\mathbf {x}}, \\omega_ {i}) = C _ {\\mathrm {R F}} (\\hat {\\mathbf {x}}, \\omega_ {i}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 804, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $T(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i)$ represents the volume transmittance of the final point sampled on the ray $\\mathbf{r}_i(t) = \\hat{\\mathbf{x}} + t\\boldsymbol{\\omega}_i$ .", + "bbox": [ + 498, + 830, + 890, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second-bounce ray marching. To make our physically-based rendering $C_{\\mathrm{PB}}$ accurate, we compute the rendering", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "169", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "integral (Eqn. 3) via Monte Carlo integration by marching multiple rays from each surface point $\\hat{\\mathbf{x}}$ with stratified sampling when training. For each ray, we obtain direct illumination from the SGs and compute visibility $T(\\hat{\\mathbf{x}},\\omega_{i})$ and indirect illumination $C_{\\mathrm{RF}}(\\hat{\\mathbf{x}},\\omega_{i})$ using Eqn. 1 directly. This second-bounce ray marching is known to be expensive for previous NeRF-based methods, which either requires extremely high computational resources (128 TPUs) [28] or utilizes long (several days of) offline pre-computation [41]; both require training extra MLPs, which takes excessive computational costs and is unable to achieve high accuracy. We instead perform second-bounce ray marching online, achieving higher accuracy (see Fig. 6), which is affordable, thanks to our highly efficient tensor factorization-based representation.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Multi-Light Representation", + "text_level": 1, + "bbox": [ + 76, + 327, + 326, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have discussed our scene representation under a single unknown lighting condition in Sec. 3.2. Our method can be easily extended to support capturing under multiple unknown lighting conditions, owing to the generality and extensibility of tensor factorization. We achieve this by adding an additional lighting dimension with vector factors $\\mathbf{e}$ - where the length of each vector $\\mathbf{e}$ equals the number of lighting conditions - into the factorized appearance tensor $\\mathcal{G}_a$ , leading to a 5D tensor that expresses scene appearance under different lighting conditions. We denote the 5D multi-light appearance tensor as $\\mathcal{G}_a^{5\\mathrm{D}}$ , represented by the factorization:", + "bbox": [ + 75, + 351, + 472, + 532 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {a} ^ {\\mathrm {5 D}} = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {a, k} ^ {m} \\circ \\mathbf {M} _ {a, k} ^ {\\tilde {m}} \\circ \\mathbf {e} _ {k} ^ {m} \\circ \\mathbf {b} _ {k} ^ {m} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 542, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that, it is only the view-dependent colors that require being modeled separately under different lighting, while the physical scene properties - including volume density, normals, and material properties - are inherently shared across multiple lighting conditions. Therefore, we decode view-dependent color using the neural features per light and decode other shading properties using the mean features along the lighting dimension:", + "bbox": [ + 75, + 585, + 468, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {a} _ {\\mathbf {x}} = \\frac {\\sum_ {l} a _ {\\mathbf {x} , l}}{P}, \\quad a _ {\\mathbf {x}, l} = \\mathcal {G} _ {a} ^ {5 D} (\\mathbf {x}, l) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 718, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {\\mathbf {x}}, \\mathbf {n} _ {\\mathbf {x}}, \\beta_ {\\mathbf {x}} = \\mathcal {D} _ {c} (a _ {\\mathbf {x}, l}), \\mathcal {D} _ {\\mathbf {n}} (\\bar {a} _ {\\mathbf {x}}), \\mathcal {D} _ {\\beta} (\\bar {a} _ {\\mathbf {x}})\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 752, + 406, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $l$ is the light index, $P$ is the number of lighting conditions, and $\\bar{a}_{\\mathbf{x}}$ is the mean feature.", + "bbox": [ + 75, + 779, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By simply adding additional vectors in the factorized appearance tensor, our method allows us to efficiently reconstruct and query the radiance under different illumination conditions, thereby still allowing for indirect illumination computation in the multi-light setting with the method discussed in Sec. 3.3. The multi-light input can provide useful", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "photometric cues and reduce ambiguity in material prediction, and therefore leads to more accurate reconstruction of geometry and material estimation (see Tab. 1).", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Joint Reconstruction and Training Losses", + "text_level": 1, + "bbox": [ + 498, + 147, + 856, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For both single- and multi-light settings, we jointly reconstruct the scene geometry and appearance properties that are modeled by our tensor factorization-based representation, through an end-to-end per-scene optimization.", + "bbox": [ + 496, + 171, + 892, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rendering losses. We supervise the rendered colors from both the radiance field rendering $C_{\\mathrm{RF}}$ and the physically-based rendering $C_{\\mathrm{PB}}$ with the ground-truth colors $C_{\\mathrm{gt}}$ from the captured images and include two rendering loss terms:", + "bbox": [ + 496, + 241, + 892, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {R F}} = \\left\\| C _ {\\mathrm {R F}} - C _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2}, \\quad \\ell_ {\\mathrm {P B}} = \\left\\| C _ {\\mathrm {P B}} - C _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 314, + 892, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normal regularization. Our shading normals $\\mathbf{n}$ are regressed from an MLP to express high-frequency surface variations and used in physically-based rendering. However, since the entire system is highly ill-conditioned, only supervising the network output with rendering loss is prone to overfitting issues and produces incorrect normal predictions. On the other hand, many previous NeRF-based method [6, 26, 28, 41] use the negative direction of volume density gradients $\\mathbf{n}_{\\sigma} = -\\frac{\\nabla_{\\mathbf{x}}\\sigma}{\\|\\nabla_{\\mathbf{x}}\\sigma\\|}$ as normals for shading computation. Such derived normals are better aligned with the surface but can be noisy and lack fine details. Inspired by the rent work Ref-NeRF [31], we correlate our shading normal with the density-derived normal using a loss", + "bbox": [ + 496, + 344, + 893, + 542 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathbf {n}} = \\sum_ {j} w _ {j} \\| \\mathbf {n} _ {j} - \\mathbf {n} _ {\\sigma , j} \\| _ {2} ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 554, + 890, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This regularization term back-propagates gradients to both the appearance tensor (through the normal network) and the density tensor, further correlating our entire scene geometry and appearance reasoning, leading to better normal reconstruction. Note that while Ref-NeRF uses the same normal regularization, their normals are used in MLP-based shading computation; in contrast, our normals are used in a physically-based rendering process, allowing for more meaningful geometry reasoning and leading to more accurate normal reconstruction.", + "bbox": [ + 496, + 589, + 892, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Final loss. Following TensoRF, we apply additional $\\ell_1$ -regularization on all tensor factors, denoted as $\\ell_{\\mathcal{G}}$ . We also apply a smoothness regularization term on BRDF parameters $\\beta$ to enhance their spatial consistency. A loss term penalizing back-facing normals is used in addition. Please refer to the supplementary materials for the details of these regularization terms. We apply all these losses to supervise and regularize our whole system to jointly reconstruct the scene, optimizing our scene representation (with all tensor factors and MLPs), as well as the SG parameters of the en", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "170", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfc68463364d3421dcd711bef49ab7f671c19740f5bbe032c1a88adffc1f9a9a.jpg", + "image_caption": [ + "Normal", + "Albedo", + "Env. Map", + "Figure 5. Visual comparison against baseline methods. Our method produces inverse rendering results of higher quality with more detailed normals and more accurate albedo, thus leading to more photo-realistic relighting results." + ], + "image_footnote": [], + "bbox": [ + 125, + 65, + 888, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6be5eb91f094618ede79603ee7beebb66c237f95bd46630debcb3af68d7fa566.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNormal MAE ↓AlbedoNovel View SynthesisRelightingAverage Runtime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor6.31425.1250.9400.10924.6790.9220.12023.3830.9080.131>100 hrs
InvRender5.07427.3410.9330.10027.3670.9340.08923.9730.9010.10115 hrs
Ours in 25 minutes4.87628.2100.9470.09132.3500.9640.06127.4310.9350.09425 mins
Ours4.10029.2750.9500.08535.0880.9760.04028.5800.9440.0815 hrs
Ours w/ rotated multi-light3.60229.6720.9560.07934.3950.9740.04328.9550.9490.0775 hrs
Ours w/ general multi-light3.55129.3260.9510.08434.2230.9730.04529.0080.9470.0785 hrs
", + "bbox": [ + 80, + 387, + 893, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons on the synthetic dataset. Our (single-light) results have significantly outperformed the baseline methods by producing more accurate normal and albedo, thus achieving more realistic novel view synthesis and relighting results. Our method can further take images captured under different lighting conditions, and boost the performance in inverse rendering. (We scale each RGB channel of all albedo results by a global scalar, as done in NeRFactor [41]. For a fair comparison, all novel view synthesis results are generated with physically-based rendering, though our radiance field rendering has better quality.)", + "bbox": [ + 75, + 483, + 890, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vironment map, with a final loss", + "bbox": [ + 76, + 558, + 292, + 571 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\ell = \\alpha_ {\\mathrm {R F}} \\ell_ {\\mathrm {R F}} + \\alpha_ {\\mathrm {P B}} \\ell_ {\\mathrm {P B}} + \\alpha_ {\\beta} \\ell_ {\\beta} + \\alpha_ {\\mathbf {n}} \\ell_ {\\mathbf {n}} + \\alpha_ {\\mathbf {d}} \\ell_ {\\mathbf {d}} + \\alpha_ {\\mathcal {G}} \\ell_ {\\mathcal {G}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 585, + 468, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 614, + 209, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now evaluate our model on various challenging scenes. We make comparisons against previous state-of-the-art methods and also present an ablation study to verify the effectiveness of our design choices.", + "bbox": [ + 75, + 640, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We perform experiments on four complex synthetic scenes, including three blender scenes from [22] and one from the Stanford 3D scanning repository [12]. We rerender these scenes to obtain their ground-truth images, as well as BRDF parameters and normal maps. We also render two types of multi-light data: rotated multi-light and general multi-light. We also perform experiments on the original NeRF-synthetic dataset and 4 captured real data. Please refer to the supplementary for experimental results of those extra data, more details about our datasets, and more explanation about our multi-light setting.", + "bbox": [ + 75, + 709, + 468, + 876 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with previous methods. We compare our", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model with previous state-of-the-art neural field-based inverse rendering methods, NeRFactor [41] and InvRender [42], on these scenes using images captured under a single unknown lighting condition. We also compare with our model trained under multi-light settings. Table 1 shows the accuracy of the estimated albedo, normal, and relighting results using metrics including PSNR, SSIM and LPIPS [40] averaged over the four scenes. We can see that our single-light method (using the same input as baselines) outperforms both previous methods significantly on all metrics, demonstrating the superiority of our method. We also include a visual comparison in Fig. 5, showing that our method predicts more accurate albedo and normal that are closer to the ground truth compared to baseline methods, thus generating more realistic relighting results. In particular, our results produce normals that are of much higher quality and more faithfully reflect the geometry variations, while the baseline methods produce over-smooth results that lack shape details. Although NeRFactor's albedo result on the lego scene looks closer to the ground truth than our results. We claim that this is because that NeRFactor uses a high-weight BRDF smoothness loss, which dam", + "bbox": [ + 496, + 558, + 890, + 890 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "171", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9bc328725e2aab2b7aed3a16410bf380357916d8d0d3f8c4a17b93d10a6a4a07.jpg", + "image_caption": [ + "NeRFactor" + ], + "image_footnote": [], + "bbox": [ + 83, + 84, + 181, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc550f7267b1e4434e90f6597acdd028c039d528480ec21c86d3d66e5143dcbf.jpg", + "image_caption": [ + "InvRender" + ], + "image_footnote": [], + "bbox": [ + 181, + 83, + 274, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a2ee602f1c2fa211fc064bdf18a265f9be2a078de1e3e0578eb9710ec7ca18b.jpg", + "image_caption": [ + "Ours", + "Figure 6. We compare our relighting results with previous methods. Note that our approach recovers more accurate shadows thanks to our second-bounce ray marching." + ], + "image_footnote": [], + "bbox": [ + 277, + 83, + 370, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8665f30e03f7319af2a70d67d09193e38cf8aaf645b3db29315e95c0537b6d3.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 372, + 83, + 465, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages its reconstruction quality of other components. In the supplementary, we showcase that we can achieve a similar albedo result by increasing the BRDF smoothness weight. In particular, NeRFactor approximates the visibility function in a distilled MLP from a pre-trained NeRF and completely ignores the indirect illumination. While InvRender considers both visibility and indirect illumination in their pre-computation, its spherical Gaussian-based shading computation can only achieve limited accuracy. Both methods do inverse rendering in a second stage after pretraining NeRFs. In contrast, our tensor-factorized representation achieves a single-stage joint reconstruction and efficiently performs explicit second-bounce ray marching for more accurate shadowing (see Fig. 6) and indirect lighting, thus leading to significantly higher reconstruction quality.", + "bbox": [ + 75, + 220, + 467, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meanwhile, owing to our more efficient representation, our high reconstruction quality is achieved with the fastest reconstruction speed, as reflected by the reconstruction time. NeRF-based method NeRFactor takes days to compute because of visibility pre-computation. InvRender is faster than NeRFactor, due to its SDF-based rendering [36] and SG-based closed-form rendering integral computation; however, its SDF-based reconstruction fails on challenging scenes like the Lego shown in Fig. 5 and the SG-based integration leads to inaccurate secondary shading effects as mentioned. On the other hand, our approach leverages ray marching-based radiance rendering in the reconstruction, robustly producing high-quality results on all testing scenes with accurate secondary effects, while still being faster than InvRender. In fact, while our method takes 5 hours to finish its full training for its best performance, it can achieve good quality in a much shorter training period (25 minutes). As shown in Tab. 1 and Fig. 7, our approach can achieve high-quality geometry reconstruction in only 25 minutes and outperform previous methods that trained for tens of or even hundreds of hours.", + "bbox": [ + 75, + 450, + 467, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-light results. In addition, our framework can also effectively leverage the additional input in a multi-light capture and further boost the accuracy of the inverse rendering performance without adding additional computation costs, as shown in Tab. 1, while the MLP-based baseline methods cannot be trivially extended to support such setups in an efficient manner. We also found that multi-light settings can greatly improve geometry reconstruction and help solving", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/55283105783f38426984a3011641ce36a392fc18afd680ca32c3b461df551e14.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNormal MAE ↓Albedo PSNR ↑NVS PSNR ↑
w/o visibility4.71627.26534.703
w/o indirect illum.4.18629.00334.910
w/ indirect illum. + visibility4.10029.27535.088
", + "bbox": [ + 504, + 89, + 888, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Our full method models more physically-accurate light transport by accounting for lighting visibility and indirect illuminations, thus achieving much better accuracy in inverse rendering.", + "bbox": [ + 500, + 136, + 890, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c92480335c778171f47c7a47d0171ade181590b3bd119bb0d8d622253708edeb.jpg", + "image_caption": [ + "NeRFactor $(>100\\mathrm{hrs})$", + "Figure 7. We compare geometry reconstruction results of our model taking only 25 minutes of optimization, with previous methods, taking 15 and $>100$ hours. Our approach even recovers more high-frequency details with substantially less reconstruction time." + ], + "image_footnote": [], + "bbox": [ + 506, + 183, + 607, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebe63a888fc657481ca6c8dc52ba52f29708978f13ac7c72e9ed3d2091567978.jpg", + "image_caption": [ + "InvRender (15 hrs)" + ], + "image_footnote": [], + "bbox": [ + 607, + 183, + 697, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2260a2848ccbb38377b69d8767ca1b15a63cf24f3555ed7765c8485b5832dbc.jpg", + "image_caption": [ + "Ours (25 mins)" + ], + "image_footnote": [], + "bbox": [ + 697, + 183, + 792, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10eddaf8a4e80fde80c8911f775ad181be81d749ed2d4ef789c6b24c17fc3898.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 792, + 183, + 887, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the color ambiguity between lighting and materials. Please refer the supplementary for more results and analysis.", + "bbox": [ + 500, + 354, + 888, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Indirect illumination and visibility. Our tensor-factorized representation allows us to explicit sample secondary rays in an efficient way to account for computing lighting visibility and indirect illumination. As shown in Tab. 2, without including these terms, the model cannot accurately represent the secondary shading effects and tend to bake them in the albedo or normal, resulting in lower quality. Nonetheless, note that these ablated models of our method in fact already achieves superior quality compared to previous methods (that pre-computes NeRFs and do inverse rendering disjointly) shown in Tab. 1, demonstrating the effectiveness of our joint reconstruction framework. Our full method further achieves more accurate reconstruction because of more accurate light transport modeling.", + "bbox": [ + 498, + 397, + 890, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 630, + 617, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel inverse rendering approach that achieves efficient and high-quality scene reconstruction from multi-view images under unknown illumination. Our approach models a scene as both a radiance field and a physically-based model with density, normals, lighting, and material properties. By jointly reconstructing both models, we achieve high-quality geometry and reflectance reconstruction, enabling photo-realistic rendering and relighting. Owing to the efficiency and generality of the tensor factorized representation, our framework allows for accurate computation for shadowing and indirect lighting effects, and also flexibly supports capturing under an arbitrary number of lighting conditions. We demonstrate that our approach is able to achieve state-of-the-art inverse rendering results, outperforming previous neural methods in terms of both reconstruction quality and efficiency.", + "bbox": [ + 498, + 659, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "172", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 2", + "[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2015. 3", + "[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2, 3", + "[4] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In European Conference on Computer Vision, pages 294–311. Springer, 2020. 2, 3", + "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5960-5969, 2020. 3", + "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2, 3, 4, 6", + "[7] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 1, 2, 3, 4", + "[8] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, volume 2012, pages 1-7. vol. 2012, 2012. 4", + "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 3", + "[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3", + "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1, 3, 4, 5", + "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 7", + "[13] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2", + "[14] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics, 33(6):193, 2014. 3", + "[15] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying brdfs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2, 3", + "[16] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):548-554, 2008. 3", + "[17] Jason Lawrence, Szymon Rusinkiewicz, and Ravi Ramamoorthi. Efficient brdf importance sampling using a factored representation. ACM Transactions on Graphics (ToG), 23(3):496-505, 2004. 3", + "[18] Hendrik PA Lensch, Jochen Lang, Asla M Sá, and Hans-Peter Seidel. Planned sampling of spatially varying brdfs. In Computer graphics forum, volume 22, pages 473-482. Wiley Online Library, 2003. 2", + "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3", + "[20] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. In SIGGRAPH Asia 2018, page 269. ACM, 2018. 2, 3", + "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2", + "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 4, 7", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 3", + "[24] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. In SIGGRAPH Asia 2018, page 267. ACM, 2018. 3", + "[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3", + "[26] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "173", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 3", + "[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2, 3, 4, 6", + "[29] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 3", + "[30] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2", + "[31] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 6", + "[32] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics, 35(6):187, 2016. 3", + "[33] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3", + "[34] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics, 38(4):76, 2019. 2", + "[35] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics, 37(4):126, 2018. 2", + "[36] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 3, 8", + "[37] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 3", + "[38] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 3", + "[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2, 3, 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 3, 4, 6, 7", + "[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 3, 5, 7" + ], + "bbox": [ + 501, + 92, + 890, + 301 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "174", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_model.json b/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1f94a3bd90475f5baa94002e3c6b2ce561010e49 --- /dev/null +++ b/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.293, + 0.131, + 0.68, + 0.154 + ], + "angle": 0, + "content": "TensoIR: Tensorial Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.179, + 0.825, + 0.221 + ], + "angle": 0, + "content": "Haian Jin\\*1 Isabella Liu\\*2 Peijia \\(\\mathrm{Xu}^3\\) Xiaoshuai Zhang2 Songfang Han2 Sai Bi4 Xiaowei Zhou' Zexiang \\(\\mathrm{Xu}^{\\dagger_4}\\) Hao Su" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.236, + 0.851, + 0.258 + ], + "angle": 0, + "content": "1 Zhejiang University 2 UC San Diego 3 Kingstar Technology Inc. 4 Adobe Research" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.337, + 0.474, + 0.671 + ], + "angle": 0, + "content": "We propose TensoIR, a novel inverse rendering approach based on tensor factorization and neural fields. Unlike previous works that use purely MLP-based neural fields, thus suffering from low capacity and high computation costs, we extend TensoRF, a state-of-the-art approach for radiance field modeling, to estimate scene geometry, surface reflectance, and environment illumination from multiview images captured under unknown lighting conditions. Our approach jointly achieves radiance field reconstruction and physically-based model estimation, leading to photorealistic novel view synthesis and relighting results. Benefiting from the efficiency and extensibility of the TensoRF-based representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under single or multiple unknown lighting conditions. The low-rank tensor representation allows us to not only achieve fast and compact reconstruction but also better exploit shared information under an arbitrary number of capturing lighting conditions. We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging synthetic and real-world scenes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.7, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Inverse rendering is a long-standing problem in computer vision and graphics, aiming to reconstruct physical attributes (like shape and materials) of a 3D scene from captured images and thereby supporting many downstream applications such as novel view synthesis, relighting and material editing. This problem is inherently challenging and ill-posed, especially when the input images are captured in the wild under unknown illumination. Recent works [6,7,28,41] address this problem by learning neural field representations in the form of multi-layer perceptrons" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.289, + 0.891, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.498, + 0.894, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Given multi-view captured images of a real scene (a), our approach - TensoIR - is able to achieve high-quality shape and material reconstruction with high-frequency details (b). This allows us to render the scene under novel lighting and viewpoints (c), and also change its material properties (d)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.656 + ], + "angle": 0, + "content": "MLP) similar to NeRF [22]. However, pure MLP-based methods usually suffer from low capacity and high computational costs, greatly limiting the accuracy and efficiency of inverse rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.869 + ], + "angle": 0, + "content": "In this work, we propose a novel inverse rendering framework that is efficient and accurate. Instead of purely using MLPs, we build upon the recent TensoRF [11] scene representation, which achieves fast, compact, and state-of-the-art quality on radiance field reconstruction for novel view synthesis. Our tensor factorization-based inverse rendering framework can simultaneously estimate scene geometry, materials, and illumination from multi-view images captured under unknown lighting conditions. Benefiting from the efficiency and extensibility of the TensoRF representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under a single or multiple unknown lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Similar to TensoRF, our approach models a scene as a neural voxel feature grid, factorized as multiple low-rank" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.311, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Equal advisory." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.501 + ], + "angle": 0, + "content": "tensor components. We apply multiple small MLPs on the same feature grid and regress volume density, view-dependent color, normal, and material properties, to model the scene geometry and appearance. This allows us to simultaneously achieve both radiance field rendering – using density and view-dependent color, as done in NeRF [22] – and physically-based rendering – using density, normal and material properties, as done in inverse rendering methods [3, 20]. We supervise both renderings with the captured images to jointly reconstruct all scene components. In essence, we reconstruct a scene using both a radiance field and a physically-based model to reproduce the scene's appearance. While inverse rendering is our focus and primarily enabled by the physically-based model, modeling the radiance field is crucial for the success of the reconstruction (see Fig. 3), in significantly facilitating the volume density reconstruction and effectively regularizing the same tensor features shared by the physically-based model. Despite that previous works [41] similarly reconstruct NeRFs in inverse rendering, their radiance field is pre-computed and fixed in the subsequent inverse rendering stage; in contrast, our radiance field is jointly reconstructed and also benefits the physically-based rendering model estimation during optimization, leading to much higher quality. Besides, our radiance field rendering can also be directly used to provide accurate indirect illumination for the physically-based rendering, further benefiting the inverse rendering process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.504, + 0.473, + 0.777 + ], + "angle": 0, + "content": "Accounting for indirect illumination and shadowing is a critical challenge in inverse rendering. This is especially challenging for volume rendering, since it requires sampling a lot of secondary rays and computing the integrals along the rays by performing ray marching. Limited by the high-cost MLP evaluation, previous NeRF-based methods and SDF-based methods either simply ignore secondary effects [6, 7, 39], or avoid online computation by approximating these effects in extra distilled MLPs [41, 42], requiring expensive pre-computation and leading to degradation in accuracy. In contrast, owing to our efficient tensor-factorized representation, we are able to explicitly compute the ray integrals online for accurate visibility and indirect lighting with the radiance field rendering using low-cost second-bounce ray marching. Consequently, our approach enables higher accuracy in modeling these secondary effects, which is crucial in achieving high-quality scene reconstruction (see Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In addition, the flexibility and efficiency of our tensor-factorized representation allows us to perform inverse rendering from multiple unknown lighting conditions with limited GPU memory. Multi-light capture is known to be beneficial for inverse rendering tasks by providing useful photometric cues and reducing ambiguities in material estimation, thus being commonly used [13,15,18]. However, since each lighting condition corresponds to a separate radiance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.333 + ], + "angle": 0, + "content": "field, this can lead to extremely high computational costs if reconstructing multiple purely MLP-based NeRFs like previous works [28, 41, 42]. Instead, we propose to reconstruct radiance fields under multi-light in a joint manner as a factorized tensor. Extending from the original TensorRF representation that is a 4D tensor, we add an additional dimension corresponding to different lighting conditions, yielding a 5D tensor. Specifically, we add an additional vector factor (whose length equals the number of lights) per tensor component to explain the appearance variations under different lighting conditions, and we store this 5D tensor by a small number of bases whose outer-product reconstructs the 5D tensor. When multi-light capture is available, our framework can effectively utilize the additional photometric cues in the data, leading to better reconstruction quality than a single-light setting (see Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.895, + 0.576 + ], + "angle": 0, + "content": "As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and reflectance of a complex real scene captured under unknown natural illumination, enabling photo-realistic rendering under novel lighting conditions and additional applications like material editing. We evaluate our framework extensively on both synthetic and real data. Our approach outperforms previous inverse rendering methods [41, 42] by a large margin qualitatively and quantitatively on challenging synthetic scenes, achieving state-of-the-art quality in scene reconstruction – for both geometry and material properties – and rendering – for both novel view synthesis and relighting. Owing to our efficient tensorial representation and joint reconstruction scheme, our approach also leads to a much faster reconstruction speed than previous neural field-based reconstruction methods while achieving superior quality. In summary," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.581, + 0.892, + 0.656 + ], + "angle": 0, + "content": "- We propose a novel tensor factorization-based inverse rendering approach that jointly achieves physically-based rendering model estimation and radiance field reconstruction, leading to state-of-the-art scene reconstruction results;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.659, + 0.892, + 0.704 + ], + "angle": 0, + "content": "- Our framework includes an efficient online visibility and indirect lighting computation technique, providing accurate second-bounce shading effects;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.708, + 0.892, + 0.753 + ], + "angle": 0, + "content": "- We enable efficient multi-light reconstruction by modeling an additional lighting dimension in the factorized tensorial representation." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.581, + 0.892, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.776, + 0.651, + 0.792 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural scene representations. Neural representations [1, 4, 21, 22, 30, 34, 35], as a promising alternative to traditional representations (like meshes, volumes, and point clouds), have been revolutionizing 3D content generation and modeling. Compared to traditional representations, such neural representations are more flexible and can more" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "166" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.87, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.283, + 0.896, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Overview. We propose a novel inverse rendering approach to reconstruct scene geometry, materials, and unknown natural illumination (as an environment map) from captured images. We reconstruct a scene as a novel representation (Sec. 3.2) that uses factorized tensors and multiple MLPs to regress volume density \\(\\sigma\\), view-dependent color \\(c\\), normals \\(\\mathbf{n}\\), and material properties (i.e. BRDF parameters) \\(\\beta\\), enabling both radiance field rendering and physically-based rendering (Sec. 3.1). In particular, we march a camera ray from camera origin \\(\\mathbf{o}\\) in viewing direction \\(\\mathbf{d}\\), sample points \\(\\mathbf{x}_j\\) along the ray, and apply radiance field rendering using the density and view-dependent colors regressed from our representation (Eqn. 1). We also use the volume rendering weights to determine the surface point \\(\\hat{\\mathbf{x}}\\) on the ray (Eqn. 2), at which we perform physically based rendering using the normals and material properties (Eqn. 3). We compute accurate visibility \\(V\\) and indirect lighting \\(L_{\\mathrm{ind}}\\) using radiance field rendering by marching secondary rays from the surface point \\(\\hat{\\mathbf{x}}\\) along sampled incoming light direction \\(\\omega_i\\) (Sec. 3.3), enabling accurate physically-based rendering. We supervise both the radiance field rendering \\(C_{\\mathrm{RF}}\\) and physically-based rendering \\(C_{\\mathrm{PB}}\\) with the captured images in a per-scene optimization for joint scene reconstruction (Sec. 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.651 + ], + "angle": 0, + "content": "accurately reproduce the geometry and appearance of real-world scenes. In particular, NeRF and many following neural field representations [11, 22, 23, 33, 36] have been proposed and applied to enable high-fidelity rendering results on novel view synthesis and many other applications [9, 10, 19, 25, 37]. Originally neural fields [22] are modelled in the form of MLPs, which have limited capacity and high computational costs. Recently, many works [11, 23, 29] introduce more efficient neural scene representations that combine neural feature maps/volumes with light-weight MLPs to reduce the computational cost and accelerate the training and rendering. In this work, we adapt the efficient tensor-factorized neural representation, TensoRF [11], to achieve accurate and efficient inverse rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Inverse rendering. Abundant works [5,15-17,24,32] have been proposed to infer the geometry and material properties of real-world objects from image collections. They typically represent the scene geometry using triangle meshes that are known or can be pre-reconstructed with depth sensors or multi-view stereo techniques [27]. To reduce the ambiguities in the inverse rendering, they often require controlled lighting conditions [5, 24], or make use of learnt domain-specific priors [2,5,14,20]. In this work, we jointly estimate the geometry, materials and lighting from images captured under unknown lighting conditions with neural field representation that is more efficient and robust. While neural representations have recently been used for inverse rendering tasks, they [3,6,7,28,38,39,41,42] are limited by the usage of computation-intensive MLPs. This inefficiency adds extra burden on inverse rendering when computing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.683 + ], + "angle": 0, + "content": "secondary shading effects (like shadows), which requires to extensively sample secondary rays. Therefore, previous methods often ignore secondary effects [6, 7, 39], consider collocated flash lighting [3, 4, 38], or take extra costs to distill these effects into additional MLP networks [28, 41, 42]. In contrast, we base our model on the advanced TensoRF representation, utilizing factorized tensors, to achieve fast reconstruction. Our TensoRF-based approach supports fast density and radiance evaluation, enabling efficient online computation for secondary effects; this leads to more accurate shadow and indirect lighting modeling during reconstruction, further benefiting our reconstruction. Moreover, in contrast to previous methods [28, 39, 41, 42] that can only handle captures under a single lighting condition, our model is easily extended to support multi-light capture by modeling an additional lighting dimension in the tensor factors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.593, + 0.71 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we present our tensor factorization-based inverse rendering framework, shown in Fig. 2, which reconstructs scene geometry, materials, and illumination from multi-view input images under unknown lighting. We leverage rendering with both neural radiance fields and physically-based light transport to model and reproduce scene appearance (Sec. 3.1). We introduce a novel TensoRF-based scene representation that allows for both rendering methods in scene reconstruction (Sec. 3.2). Our representation not only enables accurate and efficient computation for visibility and indirect lighting (Sec. 3.3) but also supports capturing under multiple unknown lighting" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "conditions (Sec. 3.4). We simultaneously estimate all scene components in a joint optimization framework with rendering losses and regularization terms (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.147, + 0.197, + 0.163 + ], + "angle": 0, + "content": "3.1. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.171, + 0.469, + 0.248 + ], + "angle": 0, + "content": "We apply ray marching to achieve differentiable radiance field rendering as done in NeRF [22]. We further determine the expected surface intersection point for each ray using the volume density, and perform physically-based rendering at the surface point with the predicted scene properties." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.256, + 0.469, + 0.302 + ], + "angle": 0, + "content": "Radiance field rendering. Given a camera ray \\(\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\) from ray origin \\(\\mathbf{o}\\) in direction \\(\\mathbf{d}\\), radiance field rendering samples \\(N\\) points on the ray and compute the pixel color as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.313, + 0.469, + 0.364 + ], + "angle": 0, + "content": "\\[\nC _ {\\mathrm {R F}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {j = 1} ^ {N} T _ {j} \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right) c _ {j} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.36, + 0.348, + 0.401 + ], + "angle": 0, + "content": "\\[\nT _ {j} = \\exp \\left(- \\sum_ {q = 1} ^ {j - 1} \\sigma_ {q} \\delta_ {q}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.469, + 0.455 + ], + "angle": 0, + "content": "where \\(\\sigma_{j},\\delta_{j},c_{j}\\) and \\(T_{j}\\) are volume density, step size, viewdependent radiance color, and volume transmittance at each sampled point \\(\\mathbf{r}(t_j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.463, + 0.469, + 0.569 + ], + "angle": 0, + "content": "Physically-based rendering. We apply a physically-based parametric BRDF model [8] \\(f_{r}\\) and perform physically-based rendering using predicted geometry and material properties at surface points on the camera rays. Similar to previous methods [22, 41], these surface points \\(\\hat{\\mathbf{x}}\\) are naturally determined using the volume rendering weights and sampled points from Eqn. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.581, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\sum_ {j = 1} ^ {N} w _ {j} \\mathbf {r} \\left(t _ {j}\\right), \\quad w _ {j} = T _ {j} \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.635, + 0.469, + 0.68 + ], + "angle": 0, + "content": "We leverage the classic surface rendering equation to compute a physically-based shading color with an integral over the upper hemisphere \\(\\Omega\\) at each surface point:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.692, + 0.469, + 0.723 + ], + "angle": 0, + "content": "\\[\nC _ {\\mathrm {P B}} (\\hat {\\mathbf {x}}, \\mathbf {d}) = \\int_ {\\Omega} L _ {\\mathrm {i}} (\\hat {\\mathbf {x}}, \\boldsymbol {\\omega} _ {i}) f _ {r} (\\hat {\\mathbf {x}}, \\boldsymbol {\\omega} _ {i}, \\mathbf {d}, \\beta) (\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.469, + 0.794 + ], + "angle": 0, + "content": "where \\(L_{\\mathrm{i}}(\\hat{\\mathbf{x}},\\omega_{i})\\) is the incident illumination coming from direction \\(\\boldsymbol{\\omega}_{i}\\). \\(\\beta\\) and \\(\\mathbf{n}\\) represent the spatially-varying material parameters (albedo and roughness) of the BRDF and the surface normal at \\(\\hat{\\mathbf{x}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In theory, accurately evaluating this integral in Eqn. 3 requires extensively sampling the lighting direction and computing the incident lighting \\( L_{\\mathrm{i}}(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i) \\) to account for shadowing and indirect illumination. It requires additional ray marching along each lighting direction, which is known to be extremely computation-expensive. Previous NeRF-based methods often simplify it by only considering direct" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "illumination [6, 7] or using extra auxiliary MLPs to make approximations and avoid full ray marching [28, 41]. Instead, we evaluate the integral more accurately by marching secondary rays online and computing the incident lighting \\( L_{\\mathrm{i}}(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i) \\) with accurate shadowing and indirect illumination (see Sec. 3.3). This is made possible by our novel efficient TensorRF-based scene representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.206, + 0.779, + 0.221 + ], + "angle": 0, + "content": "3.2. TensoRF-Based Representation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.229, + 0.892, + 0.349 + ], + "angle": 0, + "content": "We now introduce our tensor factorization-based scene representation that simultaneously models volume density \\(\\sigma\\), view-dependent color \\(c\\), shading normal \\(\\mathbf{n}\\), and material properties \\(\\beta\\) (including diffuse albedo and specular roughness of the Disney BRDF [8]). Our method can model the scene under both single or multiple lighting conditions. We first introduce the single-light setting in this section and will discuss the multi-light extension in Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.35, + 0.892, + 0.547 + ], + "angle": 0, + "content": "At a high-level, radiance, geometry, and material properties can all be represented as a 3D field. We can use a feature volume and voxel feature decoding functions to extract information at any point of the 3D field. To compress the space of the feature volume and also regularize the learning process, TensoRF [11] proposed to use a low-rank factorized tensor as the feature volume. In this work, we adopt the Vector-Matrix factorization proposed by TensoRF. In particular, we use two separate VM-factorized tensors, i.e. feature grids \\(\\mathcal{G}_{\\sigma}\\) and \\(\\mathcal{G}_a\\), to model volume density and appearance, respectively. The appearance tensor \\(\\mathcal{G}_a\\) is followed by multiple light-weight MLP decoders to regress various appearance properties." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.615 + ], + "angle": 0, + "content": "Density tensor. The density tensor \\(\\mathcal{G}_{\\sigma}\\) is 3D and directly expresses volume density without any network. Specifically, the VM-factorized density tensor is expressed by the sum of vector-matrix outer products." + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.623, + 0.891, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {G} _ {\\sigma} = \\sum_ {k} \\mathbf {v} _ {\\sigma , k} ^ {X} \\circ \\mathbf {M} _ {\\sigma , k} ^ {Y Z} + \\mathbf {v} _ {\\sigma , k} ^ {Y} \\circ \\mathbf {M} _ {\\sigma , k} ^ {X Z} + \\mathbf {v} _ {\\sigma , k} ^ {Z} \\circ \\mathbf {M} _ {\\sigma , k} ^ {X Y} \\\\ = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {\\sigma , k} ^ {m} \\circ \\mathbf {M} _ {\\sigma , k} ^ {\\tilde {m}} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.699, + 0.891, + 0.745 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{v}_{\\sigma ,k}^{m}\\) and \\(\\mathbf{M}_{\\sigma ,k}^{\\tilde{m}}\\) represent the \\(k^{\\mathrm{th}}\\) vector and matrix factors of their corresponding spatial axes \\(m\\); for simplicity, \\(\\tilde{m}\\) denotes the two axes orthogonal to \\(m\\) (e.g. \\(\\tilde{X} = YZ\\))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.754, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Appearance tensor. The appearance tensor \\(\\mathcal{G}_a\\) is 4D, modeled by similar vector-matrix spatial factors and additional feature basis vectors \\(\\mathbf{b}\\), expressing a multi-channel voxel feature grid:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.823, + 0.891, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {a} = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {a, k} ^ {m} \\circ \\mathbf {M} _ {a, k} ^ {\\tilde {m}} \\circ \\mathbf {b} _ {k} ^ {m} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Representing scene properties. We obtain density \\(\\sigma\\) directly by linearly interpolating \\(\\mathcal{G}_{\\sigma}\\), and apply multiple MLP" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.293 + ], + "angle": 0, + "content": "Figure 3. We compare normal and albedo reconstruction results between our joint reconstruction model and an ablated model without radiance field rendering during reconstruction. Radiance field reconstruction is crucial for us to achieve good reconstruction with a clean background and reasonable scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.3, + 0.471, + 0.362 + ], + "angle": 0, + "content": "networks to decode appearance properties from interpolated appearance features \\(\\mathcal{G}_a\\). This includes a radiance network \\(\\mathcal{D}_c\\), a shading normal network \\(\\mathcal{D}_{\\mathbf{n}}\\), and a material network \\(\\mathcal{D}_{\\beta}\\). Overall, our scene representation is expressed by" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.372, + 0.369, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\mathbf {x}} = \\mathcal {G} _ {\\sigma} (\\mathbf {x}), \\quad a _ {\\mathbf {x}} = \\mathcal {G} _ {a} (\\mathbf {x})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.383, + 0.469, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} c _ {\\mathbf {x}}, \\mathbf {n} _ {\\mathbf {x}}, \\beta_ {\\mathbf {x}} = \\mathcal {D} _ {c} (a _ {\\mathbf {x}}), \\mathcal {D} _ {\\mathbf {n}} (a _ {\\mathbf {x}}), \\mathcal {D} _ {\\boldsymbol {\\beta}} (a _ {\\mathbf {x}}) \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.508 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{x}\\) denotes an arbitrary 3D location, \\(\\sigma_{\\mathbf{x}} = \\mathcal{G}_{\\sigma}(\\mathbf{x})\\) and \\(a_{\\mathbf{x}} = \\mathcal{G}_a(\\mathbf{x})\\) are the linearly interpolated density and multichannel appearance features, computed by linearly interpolating the spatial vector and matrix factors (please refer to the TensoRF paper [11] for the details of feature computation and interpolation)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.644 + ], + "angle": 0, + "content": "In particular, volume density \\(\\sigma\\) and shading normals \\(\\mathbf{n}\\) both express scene geometry, which describes global shapes and high-frequency geometric details, respectively. View-dependent color \\(c\\) and physical shading properties (normal \\(\\mathbf{n}\\) and material parameters \\(\\beta\\)) duplicatively model the scene appearance, determining the colors in the scene. We bind shading normal with volume density using a regularization term (see details in Sec. 3.5), correlating the scene geometry estimation and appearance reasoning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In essence, our TensoRF-based scene representation provides scene geometry and appearance properties that are required for both radiance field rendering (Eqn. 1) and physically-based rendering (Eqn. 3) described in Sec. 3.1. This represents a scene as a radiance field and a physically-based rendering model jointly. We let the two sub-representations share the same neural features in our tensors, allowing their learning processes to benefit each other. While the physically-based model is our main focus and achieves inverse rendering, modeling the radiance field can facilitate the volume density reconstruction and also regularize the appearance features to be meaningful since it has a shorter gradient path. As a result, modeling the radiance field is crucial and necessary for us to achieve high-quality physically-based scene reconstruction as shown in Fig. 3. In addition, the radiance field can naturally provide indirect illumination for physically-based rendering, enabling more" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.587, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.149, + 0.553, + 0.158 + ], + "angle": 0, + "content": "a) Image" + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.091, + 0.684, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.149, + 0.669, + 0.158 + ], + "angle": 0, + "content": "b) Indirect lighting" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.091, + 0.888, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.149, + 0.878, + 0.158 + ], + "angle": 0, + "content": "c) Lighting visibility under two directional lights" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.169, + 0.892, + 0.211 + ], + "angle": 0, + "content": "Figure 4. We show our computed indirect illumination (b) of the full rendered image (a) and our lighting visibility (c) under two different directional lights." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.223, + 0.764, + 0.237 + ], + "angle": 0, + "content": "accurate physical model reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.246, + 0.74, + 0.263 + ], + "angle": 0, + "content": "3.3. Illumination and Visibility" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Our TensoRF-based scene representation is highly efficient for optimization and evaluation. Especially, our volume density can be computed by simple tensor interpolation without using any MLPs, leading to highly efficient computation of volume transmittance and rendering weights (Eqn. 1,2). This allows us to compute accurate incident illumination \\( L_{\\mathrm{i}} \\) (as defined in Eqn. 3) that accounts for secondary shading effects with ray marching as shown in Fig. 4. In particular, the incident illumination is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.431, + 0.892, + 0.449 + ], + "angle": 0, + "content": "\\[\nL _ {i} (\\hat {\\mathbf {x}}, \\omega_ {i}) = V (\\hat {\\mathbf {x}}, \\omega_ {i}) L _ {\\mathrm {d}} (\\omega_ {i}) + L _ {\\mathrm {i n d}} (\\hat {\\mathbf {x}}, \\omega_ {i}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.891, + 0.489 + ], + "angle": 0, + "content": "where \\(V\\) is the light visibility function, \\(L_{\\mathrm{d}}\\) is the direct illumination, and \\(L_{\\mathrm{ind}}\\) is the indirect illumination." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Direct illumination. We assume the unknown natural environment to be distant from the captured object and represent the global illumination as an environment map, parameterized by a mixture of spherical Gaussians (SG), which represents the direct illumination. Note that in contrast to previous methods [39,42] that use SG for computing the integral of the rendering equation with a closed-form approximation, we use SG only for its compact parameterized representation and compute the integral numerically by sampling secondary rays and performing ray marching, leading to more accurate lighting visibility and indirect lighting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Visibility and indirect illumination. We make use of the jointly-trained radiance fields to model secondary shading effects. More specifically, the indirect illumination arriving at \\(\\hat{\\mathbf{x}}\\) from \\(\\omega_{i}\\) is inherently explained by the radiance color \\(C_{\\mathrm{RF}}\\) along the ray \\(\\mathbf{r}_i(t) = \\hat{\\mathbf{x}} +t\\omega_i\\). The visibility function is exactly modeled by the transmittance function in volume rendering. Therefore, the light visibility and indirect illumination term in Eqn. 7 can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.805, + 0.892, + 0.822 + ], + "angle": 0, + "content": "\\[\nV (\\hat {\\mathbf {x}}, \\omega_ {i}) = T (\\hat {\\mathbf {x}}, \\omega_ {i}), \\quad L _ {\\text {i n d}} (\\hat {\\mathbf {x}}, \\omega_ {i}) = C _ {\\mathrm {R F}} (\\hat {\\mathbf {x}}, \\omega_ {i}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.831, + 0.891, + 0.863 + ], + "angle": 0, + "content": "Here \\( T(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i) \\) represents the volume transmittance of the final point sampled on the ray \\( \\mathbf{r}_i(t) = \\hat{\\mathbf{x}} + t\\boldsymbol{\\omega}_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Second-bounce ray marching. To make our physically-based rendering \\( C_{\\mathrm{PB}} \\) accurate, we compute the rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "integral (Eqn. 3) via Monte Carlo integration by marching multiple rays from each surface point \\(\\hat{\\mathbf{x}}\\) with stratified sampling when training. For each ray, we obtain direct illumination from the SGs and compute visibility \\(T(\\hat{\\mathbf{x}},\\omega_{i})\\) and indirect illumination \\(C_{\\mathrm{RF}}(\\hat{\\mathbf{x}},\\omega_{i})\\) using Eqn. 1 directly. This second-bounce ray marching is known to be expensive for previous NeRF-based methods, which either requires extremely high computational resources (128 TPUs) [28] or utilizes long (several days of) offline pre-computation [41]; both require training extra MLPs, which takes excessive computational costs and is unable to achieve high accuracy. We instead perform second-bounce ray marching online, achieving higher accuracy (see Fig. 6), which is affordable, thanks to our highly efficient tensor factorization-based representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.328, + 0.327, + 0.346 + ], + "angle": 0, + "content": "3.4. Multi-Light Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.534 + ], + "angle": 0, + "content": "We have discussed our scene representation under a single unknown lighting condition in Sec. 3.2. Our method can be easily extended to support capturing under multiple unknown lighting conditions, owing to the generality and extensibility of tensor factorization. We achieve this by adding an additional lighting dimension with vector factors \\(\\mathbf{e}\\) - where the length of each vector \\(\\mathbf{e}\\) equals the number of lighting conditions - into the factorized appearance tensor \\(\\mathcal{G}_a\\), leading to a 5D tensor that expresses scene appearance under different lighting conditions. We denote the 5D multi-light appearance tensor as \\(\\mathcal{G}_a^{5\\mathrm{D}}\\), represented by the factorization:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.543, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {a} ^ {\\mathrm {5 D}} = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {a, k} ^ {m} \\circ \\mathbf {M} _ {a, k} ^ {\\tilde {m}} \\circ \\mathbf {e} _ {k} ^ {m} \\circ \\mathbf {b} _ {k} ^ {m} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Note that, it is only the view-dependent colors that require being modeled separately under different lighting, while the physical scene properties - including volume density, normals, and material properties - are inherently shared across multiple lighting conditions. Therefore, we decode view-dependent color using the neural features per light and decode other shading properties using the mean features along the lighting dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.719, + 0.469, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\bar {a} _ {\\mathbf {x}} = \\frac {\\sum_ {l} a _ {\\mathbf {x} , l}}{P}, \\quad a _ {\\mathbf {x}, l} = \\mathcal {G} _ {a} ^ {5 D} (\\mathbf {x}, l) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.753, + 0.408, + 0.769 + ], + "angle": 0, + "content": "\\[\nc _ {\\mathbf {x}}, \\mathbf {n} _ {\\mathbf {x}}, \\beta_ {\\mathbf {x}} = \\mathcal {D} _ {c} (a _ {\\mathbf {x}, l}), \\mathcal {D} _ {\\mathbf {n}} (\\bar {a} _ {\\mathbf {x}}), \\mathcal {D} _ {\\beta} (\\bar {a} _ {\\mathbf {x}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.81 + ], + "angle": 0, + "content": "where \\(l\\) is the light index, \\(P\\) is the number of lighting conditions, and \\(\\bar{a}_{\\mathbf{x}}\\) is the mean feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "By simply adding additional vectors in the factorized appearance tensor, our method allows us to efficiently reconstruct and query the radiance under different illumination conditions, thereby still allowing for indirect illumination computation in the multi-light setting with the method discussed in Sec. 3.3. The multi-light input can provide useful" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "photometric cues and reduce ambiguity in material prediction, and therefore leads to more accurate reconstruction of geometry and material estimation (see Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.148, + 0.857, + 0.166 + ], + "angle": 0, + "content": "3.5. Joint Reconstruction and Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.893, + 0.233 + ], + "angle": 0, + "content": "For both single- and multi-light settings, we jointly reconstruct the scene geometry and appearance properties that are modeled by our tensor factorization-based representation, through an end-to-end per-scene optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Rendering losses. We supervise the rendered colors from both the radiance field rendering \\( C_{\\mathrm{RF}} \\) and the physically-based rendering \\( C_{\\mathrm{PB}} \\) with the ground-truth colors \\( C_{\\mathrm{gt}} \\) from the captured images and include two rendering loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.315, + 0.893, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {R F}} = \\left\\| C _ {\\mathrm {R F}} - C _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2}, \\quad \\ell_ {\\mathrm {P B}} = \\left\\| C _ {\\mathrm {P B}} - C _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.895, + 0.543 + ], + "angle": 0, + "content": "Normal regularization. Our shading normals \\(\\mathbf{n}\\) are regressed from an MLP to express high-frequency surface variations and used in physically-based rendering. However, since the entire system is highly ill-conditioned, only supervising the network output with rendering loss is prone to overfitting issues and produces incorrect normal predictions. On the other hand, many previous NeRF-based method [6, 26, 28, 41] use the negative direction of volume density gradients \\(\\mathbf{n}_{\\sigma} = -\\frac{\\nabla_{\\mathbf{x}}\\sigma}{\\|\\nabla_{\\mathbf{x}}\\sigma\\|}\\) as normals for shading computation. Such derived normals are better aligned with the surface but can be noisy and lack fine details. Inspired by the rent work Ref-NeRF [31], we correlate our shading normal with the density-derived normal using a loss" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.555, + 0.892, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathbf {n}} = \\sum_ {j} w _ {j} \\| \\mathbf {n} _ {j} - \\mathbf {n} _ {\\sigma , j} \\| _ {2} ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.893, + 0.74 + ], + "angle": 0, + "content": "This regularization term back-propagates gradients to both the appearance tensor (through the normal network) and the density tensor, further correlating our entire scene geometry and appearance reasoning, leading to better normal reconstruction. Note that while Ref-NeRF uses the same normal regularization, their normals are used in MLP-based shading computation; in contrast, our normals are used in a physically-based rendering process, allowing for more meaningful geometry reasoning and leading to more accurate normal reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Final loss. Following TensoRF, we apply additional \\(\\ell_1\\)-regularization on all tensor factors, denoted as \\(\\ell_{\\mathcal{G}}\\). We also apply a smoothness regularization term on BRDF parameters \\(\\beta\\) to enhance their spatial consistency. A loss term penalizing back-facing normals is used in addition. Please refer to the supplementary materials for the details of these regularization terms. We apply all these losses to supervise and regularize our whole system to jointly reconstruct the scene, optimizing our scene representation (with all tensor factors and MLPs), as well as the SG parameters of the en" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "170" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.113, + 0.105, + 0.159 + ], + "angle": 270, + "content": "Normal" + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.22, + 0.105, + 0.262 + ], + "angle": 270, + "content": "Albedo" + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.302, + 0.105, + 0.355 + ], + "angle": 270, + "content": "Env. Map" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.066, + 0.89, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.36, + 0.893, + 0.387 + ], + "angle": 0, + "content": "Figure 5. Visual comparison against baseline methods. Our method produces inverse rendering results of higher quality with more detailed normals and more accurate albedo, thus leading to more photo-realistic relighting results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.388, + 0.894, + 0.481 + ], + "angle": 0, + "content": "
MethodNormal MAE ↓AlbedoNovel View SynthesisRelightingAverage Runtime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor6.31425.1250.9400.10924.6790.9220.12023.3830.9080.131>100 hrs
InvRender5.07427.3410.9330.10027.3670.9340.08923.9730.9010.10115 hrs
Ours in 25 minutes4.87628.2100.9470.09132.3500.9640.06127.4310.9350.09425 mins
Ours4.10029.2750.9500.08535.0880.9760.04028.5800.9440.0815 hrs
Ours w/ rotated multi-light3.60229.6720.9560.07934.3950.9740.04328.9550.9490.0775 hrs
Ours w/ general multi-light3.55129.3260.9510.08434.2230.9730.04529.0080.9470.0785 hrs
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.484, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the synthetic dataset. Our (single-light) results have significantly outperformed the baseline methods by producing more accurate normal and albedo, thus achieving more realistic novel view synthesis and relighting results. Our method can further take images captured under different lighting conditions, and boost the performance in inverse rendering. (We scale each RGB channel of all albedo results by a global scalar, as done in NeRFactor [41]. For a fair comparison, all novel view synthesis results are generated with physically-based rendering, though our radiance field rendering has better quality.)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.559, + 0.294, + 0.573 + ], + "angle": 0, + "content": "vironment map, with a final loss" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.586, + 0.47, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\ell = \\alpha_ {\\mathrm {R F}} \\ell_ {\\mathrm {R F}} + \\alpha_ {\\mathrm {P B}} \\ell_ {\\mathrm {P B}} + \\alpha_ {\\beta} \\ell_ {\\beta} + \\alpha_ {\\mathbf {n}} \\ell_ {\\mathbf {n}} + \\alpha_ {\\mathbf {d}} \\ell_ {\\mathbf {d}} + \\alpha_ {\\mathcal {G}} \\ell_ {\\mathcal {G}} \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.21, + 0.633 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.702 + ], + "angle": 0, + "content": "We now evaluate our model on various challenging scenes. We make comparisons against previous state-of-the-art methods and also present an ablation study to verify the effectiveness of our design choices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Datasets. We perform experiments on four complex synthetic scenes, including three blender scenes from [22] and one from the Stanford 3D scanning repository [12]. We rerender these scenes to obtain their ground-truth images, as well as BRDF parameters and normal maps. We also render two types of multi-light data: rotated multi-light and general multi-light. We also perform experiments on the original NeRF-synthetic dataset and 4 captured real data. Please refer to the supplementary for experimental results of those extra data, more details about our datasets, and more explanation about our multi-light setting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparisons with previous methods. We compare our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.892, + 0.891 + ], + "angle": 0, + "content": "model with previous state-of-the-art neural field-based inverse rendering methods, NeRFactor [41] and InvRender [42], on these scenes using images captured under a single unknown lighting condition. We also compare with our model trained under multi-light settings. Table 1 shows the accuracy of the estimated albedo, normal, and relighting results using metrics including PSNR, SSIM and LPIPS [40] averaged over the four scenes. We can see that our single-light method (using the same input as baselines) outperforms both previous methods significantly on all metrics, demonstrating the superiority of our method. We also include a visual comparison in Fig. 5, showing that our method predicts more accurate albedo and normal that are closer to the ground truth compared to baseline methods, thus generating more realistic relighting results. In particular, our results produce normals that are of much higher quality and more faithfully reflect the geometry variations, while the baseline methods produce over-smooth results that lack shape details. Although NeRFactor's albedo result on the lego scene looks closer to the ground truth than our results. We claim that this is because that NeRFactor uses a high-weight BRDF smoothness loss, which dam" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "171" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.085, + 0.182, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.161, + 0.15, + 0.168 + ], + "angle": 0, + "content": "NeRFactor" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.084, + 0.276, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.161, + 0.248, + 0.168 + ], + "angle": 0, + "content": "InvRender" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.084, + 0.371, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.161, + 0.33, + 0.168 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.084, + 0.467, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.161, + 0.439, + 0.167 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.171, + 0.468, + 0.211 + ], + "angle": 0, + "content": "Figure 6. We compare our relighting results with previous methods. Note that our approach recovers more accurate shadows thanks to our second-bounce ray marching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.468, + 0.447 + ], + "angle": 0, + "content": "ages its reconstruction quality of other components. In the supplementary, we showcase that we can achieve a similar albedo result by increasing the BRDF smoothness weight. In particular, NeRFactor approximates the visibility function in a distilled MLP from a pre-trained NeRF and completely ignores the indirect illumination. While InvRender considers both visibility and indirect illumination in their pre-computation, its spherical Gaussian-based shading computation can only achieve limited accuracy. Both methods do inverse rendering in a second stage after pretraining NeRFs. In contrast, our tensor-factorized representation achieves a single-stage joint reconstruction and efficiently performs explicit second-bounce ray marching for more accurate shadowing (see Fig. 6) and indirect lighting, thus leading to significantly higher reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.468, + 0.767 + ], + "angle": 0, + "content": "Meanwhile, owing to our more efficient representation, our high reconstruction quality is achieved with the fastest reconstruction speed, as reflected by the reconstruction time. NeRF-based method NeRFactor takes days to compute because of visibility pre-computation. InvRender is faster than NeRFactor, due to its SDF-based rendering [36] and SG-based closed-form rendering integral computation; however, its SDF-based reconstruction fails on challenging scenes like the Lego shown in Fig. 5 and the SG-based integration leads to inaccurate secondary shading effects as mentioned. On the other hand, our approach leverages ray marching-based radiance rendering in the reconstruction, robustly producing high-quality results on all testing scenes with accurate secondary effects, while still being faster than InvRender. In fact, while our method takes 5 hours to finish its full training for its best performance, it can achieve good quality in a much shorter training period (25 minutes). As shown in Tab. 1 and Fig. 7, our approach can achieve high-quality geometry reconstruction in only 25 minutes and outperform previous methods that trained for tens of or even hundreds of hours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Multi-light results. In addition, our framework can also effectively leverage the additional input in a multi-light capture and further boost the accuracy of the inverse rendering performance without adding additional computation costs, as shown in Tab. 1, while the MLP-based baseline methods cannot be trivially extended to support such setups in an efficient manner. We also found that multi-light settings can greatly improve geometry reconstruction and help solving" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.89, + 0.136 + ], + "angle": 0, + "content": "
MethodNormal MAE ↓Albedo PSNR ↑NVS PSNR ↑
w/o visibility4.71627.26534.703
w/o indirect illum.4.18629.00334.910
w/ indirect illum. + visibility4.10029.27535.088
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Table 2. Our full method models more physically-accurate light transport by accounting for lighting visibility and indirect illuminations, thus achieving much better accuracy in inverse rendering." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.184, + 0.609, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.278, + 0.613, + 0.286 + ], + "angle": 0, + "content": "NeRFactor \\((>100\\mathrm{hrs})\\)" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.184, + 0.699, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.278, + 0.708, + 0.286 + ], + "angle": 0, + "content": "InvRender (15 hrs)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.184, + 0.794, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.278, + 0.791, + 0.286 + ], + "angle": 0, + "content": "Ours (25 mins)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.184, + 0.888, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.278, + 0.878, + 0.286 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.289, + 0.891, + 0.344 + ], + "angle": 0, + "content": "Figure 7. We compare geometry reconstruction results of our model taking only 25 minutes of optimization, with previous methods, taking 15 and \\(>100\\) hours. Our approach even recovers more high-frequency details with substantially less reconstruction time." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.355, + 0.89, + 0.386 + ], + "angle": 0, + "content": "the color ambiguity between lighting and materials. Please refer the supplementary for more results and analysis." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Indirect illumination and visibility. Our tensor-factorized representation allows us to explicit sample secondary rays in an efficient way to account for computing lighting visibility and indirect illumination. As shown in Tab. 2, without including these terms, the model cannot accurately represent the secondary shading effects and tend to bake them in the albedo or normal, resulting in lower quality. Nonetheless, note that these ablated models of our method in fact already achieves superior quality compared to previous methods (that pre-computes NeRFs and do inverse rendering disjointly) shown in Tab. 1, demonstrating the effectiveness of our joint reconstruction framework. Our full method further achieves more accurate reconstruction because of more accurate light transport modeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.631, + 0.618, + 0.646 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present a novel inverse rendering approach that achieves efficient and high-quality scene reconstruction from multi-view images under unknown illumination. Our approach models a scene as both a radiance field and a physically-based model with density, normals, lighting, and material properties. By jointly reconstructing both models, we achieve high-quality geometry and reflectance reconstruction, enabling photo-realistic rendering and relighting. Owing to the efficiency and generality of the tensor factorized representation, our framework allows for accurate computation for shadowing and indirect lighting effects, and also flexibly supports capturing under an arbitrary number of lighting conditions. We demonstrate that our approach is able to achieve state-of-the-art inverse rendering results, outperforming previous neural methods in terms of both reconstruction quality and efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "172" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.296 + ], + "angle": 0, + "content": "[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In European Conference on Computer Vision, pages 294–311. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5960-5969, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.521 + ], + "angle": 0, + "content": "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[7] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.471, + 0.633 + ], + "angle": 0, + "content": "[8] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, volume 2012, pages 1-7. vol. 2012, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics, 33(6):193, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying brdfs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):548-554, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Jason Lawrence, Szymon Rusinkiewicz, and Ravi Ramamoorthi. Efficient brdf importance sampling using a factored representation. ACM Transactions on Graphics (ToG), 23(3):496-505, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Hendrik PA Lensch, Jochen Lang, Asla M Sá, and Hans-Peter Seidel. Planned sampling of spatially varying brdfs. In Computer graphics forum, volume 22, pages 473-482. Wiley Online Library, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[20] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. In SIGGRAPH Asia 2018, page 269. ACM, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.538, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[24] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. In SIGGRAPH Asia 2018, page 267. ACM, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "173" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.231 + ], + "angle": 0, + "content": "[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[29] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[30] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[31] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[32] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics, 35(6):187, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[33] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[34] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics, 38(4):76, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[35] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics, 37(4):126, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.624, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[36] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[37] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[38] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "174" + } + ] +] \ No newline at end of file diff --git a/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_origin.pdf b/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2dab3cc210bd17e50b2e932e0ed3fad7db086a65 --- /dev/null +++ b/2023/TensoIR_ Tensorial Inverse Rendering/911491ec-c0e6-468f-9b5f-cd8676ecd8ce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0185e1088de01c71b4c83ec11e16f39e2537d27b4970ea7bfdff5ea7f365fb0 +size 6470743 diff --git a/2023/TensoIR_ Tensorial Inverse Rendering/full.md b/2023/TensoIR_ Tensorial Inverse Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66b23546dc4679a09468886472ba926d9195eca3 --- /dev/null +++ b/2023/TensoIR_ Tensorial Inverse Rendering/full.md @@ -0,0 +1,329 @@ +# TensoIR: Tensorial Inverse Rendering + +Haian Jin\*1 Isabella Liu\*2 Peijia $\mathrm{Xu}^3$ Xiaoshuai Zhang2 Songfang Han2 Sai Bi4 Xiaowei Zhou' Zexiang $\mathrm{Xu}^{\dagger_4}$ Hao Su + +1 Zhejiang University 2 UC San Diego 3 Kingstar Technology Inc. 4 Adobe Research + +# Abstract + +We propose TensoIR, a novel inverse rendering approach based on tensor factorization and neural fields. Unlike previous works that use purely MLP-based neural fields, thus suffering from low capacity and high computation costs, we extend TensoRF, a state-of-the-art approach for radiance field modeling, to estimate scene geometry, surface reflectance, and environment illumination from multiview images captured under unknown lighting conditions. Our approach jointly achieves radiance field reconstruction and physically-based model estimation, leading to photorealistic novel view synthesis and relighting results. Benefiting from the efficiency and extensibility of the TensoRF-based representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under single or multiple unknown lighting conditions. The low-rank tensor representation allows us to not only achieve fast and compact reconstruction but also better exploit shared information under an arbitrary number of capturing lighting conditions. We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging synthetic and real-world scenes. + +# 1. Introduction + +Inverse rendering is a long-standing problem in computer vision and graphics, aiming to reconstruct physical attributes (like shape and materials) of a 3D scene from captured images and thereby supporting many downstream applications such as novel view synthesis, relighting and material editing. This problem is inherently challenging and ill-posed, especially when the input images are captured in the wild under unknown illumination. Recent works [6,7,28,41] address this problem by learning neural field representations in the form of multi-layer perceptrons + +![](images/131986b44b1a280cdb70bc2d86761bc81c8ca91ae8d8670c6fc7f6559c4755b3.jpg) +Figure 1. Given multi-view captured images of a real scene (a), our approach - TensoIR - is able to achieve high-quality shape and material reconstruction with high-frequency details (b). This allows us to render the scene under novel lighting and viewpoints (c), and also change its material properties (d). + +MLP) similar to NeRF [22]. However, pure MLP-based methods usually suffer from low capacity and high computational costs, greatly limiting the accuracy and efficiency of inverse rendering. + +In this work, we propose a novel inverse rendering framework that is efficient and accurate. Instead of purely using MLPs, we build upon the recent TensoRF [11] scene representation, which achieves fast, compact, and state-of-the-art quality on radiance field reconstruction for novel view synthesis. Our tensor factorization-based inverse rendering framework can simultaneously estimate scene geometry, materials, and illumination from multi-view images captured under unknown lighting conditions. Benefiting from the efficiency and extensibility of the TensoRF representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under a single or multiple unknown lighting conditions. + +Similar to TensoRF, our approach models a scene as a neural voxel feature grid, factorized as multiple low-rank + +tensor components. We apply multiple small MLPs on the same feature grid and regress volume density, view-dependent color, normal, and material properties, to model the scene geometry and appearance. This allows us to simultaneously achieve both radiance field rendering – using density and view-dependent color, as done in NeRF [22] – and physically-based rendering – using density, normal and material properties, as done in inverse rendering methods [3, 20]. We supervise both renderings with the captured images to jointly reconstruct all scene components. In essence, we reconstruct a scene using both a radiance field and a physically-based model to reproduce the scene's appearance. While inverse rendering is our focus and primarily enabled by the physically-based model, modeling the radiance field is crucial for the success of the reconstruction (see Fig. 3), in significantly facilitating the volume density reconstruction and effectively regularizing the same tensor features shared by the physically-based model. Despite that previous works [41] similarly reconstruct NeRFs in inverse rendering, their radiance field is pre-computed and fixed in the subsequent inverse rendering stage; in contrast, our radiance field is jointly reconstructed and also benefits the physically-based rendering model estimation during optimization, leading to much higher quality. Besides, our radiance field rendering can also be directly used to provide accurate indirect illumination for the physically-based rendering, further benefiting the inverse rendering process. + +Accounting for indirect illumination and shadowing is a critical challenge in inverse rendering. This is especially challenging for volume rendering, since it requires sampling a lot of secondary rays and computing the integrals along the rays by performing ray marching. Limited by the high-cost MLP evaluation, previous NeRF-based methods and SDF-based methods either simply ignore secondary effects [6, 7, 39], or avoid online computation by approximating these effects in extra distilled MLPs [41, 42], requiring expensive pre-computation and leading to degradation in accuracy. In contrast, owing to our efficient tensor-factorized representation, we are able to explicitly compute the ray integrals online for accurate visibility and indirect lighting with the radiance field rendering using low-cost second-bounce ray marching. Consequently, our approach enables higher accuracy in modeling these secondary effects, which is crucial in achieving high-quality scene reconstruction (see Tab. 2). + +In addition, the flexibility and efficiency of our tensor-factorized representation allows us to perform inverse rendering from multiple unknown lighting conditions with limited GPU memory. Multi-light capture is known to be beneficial for inverse rendering tasks by providing useful photometric cues and reducing ambiguities in material estimation, thus being commonly used [13,15,18]. However, since each lighting condition corresponds to a separate radiance + +field, this can lead to extremely high computational costs if reconstructing multiple purely MLP-based NeRFs like previous works [28, 41, 42]. Instead, we propose to reconstruct radiance fields under multi-light in a joint manner as a factorized tensor. Extending from the original TensorRF representation that is a 4D tensor, we add an additional dimension corresponding to different lighting conditions, yielding a 5D tensor. Specifically, we add an additional vector factor (whose length equals the number of lights) per tensor component to explain the appearance variations under different lighting conditions, and we store this 5D tensor by a small number of bases whose outer-product reconstructs the 5D tensor. When multi-light capture is available, our framework can effectively utilize the additional photometric cues in the data, leading to better reconstruction quality than a single-light setting (see Tab. 1). + +As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and reflectance of a complex real scene captured under unknown natural illumination, enabling photo-realistic rendering under novel lighting conditions and additional applications like material editing. We evaluate our framework extensively on both synthetic and real data. Our approach outperforms previous inverse rendering methods [41, 42] by a large margin qualitatively and quantitatively on challenging synthetic scenes, achieving state-of-the-art quality in scene reconstruction – for both geometry and material properties – and rendering – for both novel view synthesis and relighting. Owing to our efficient tensorial representation and joint reconstruction scheme, our approach also leads to a much faster reconstruction speed than previous neural field-based reconstruction methods while achieving superior quality. In summary, + +- We propose a novel tensor factorization-based inverse rendering approach that jointly achieves physically-based rendering model estimation and radiance field reconstruction, leading to state-of-the-art scene reconstruction results; +- Our framework includes an efficient online visibility and indirect lighting computation technique, providing accurate second-bounce shading effects; +- We enable efficient multi-light reconstruction by modeling an additional lighting dimension in the factorized tensorial representation. + +# 2. Related Works + +Neural scene representations. Neural representations [1, 4, 21, 22, 30, 34, 35], as a promising alternative to traditional representations (like meshes, volumes, and point clouds), have been revolutionizing 3D content generation and modeling. Compared to traditional representations, such neural representations are more flexible and can more + +![](images/22b22337385c57a5fedf6367b1ac1da7d3c2cee834ce11c9b639d90ca8e86789.jpg) +Figure 2. Overview. We propose a novel inverse rendering approach to reconstruct scene geometry, materials, and unknown natural illumination (as an environment map) from captured images. We reconstruct a scene as a novel representation (Sec. 3.2) that uses factorized tensors and multiple MLPs to regress volume density $\sigma$ , view-dependent color $c$ , normals $\mathbf{n}$ , and material properties (i.e. BRDF parameters) $\beta$ , enabling both radiance field rendering and physically-based rendering (Sec. 3.1). In particular, we march a camera ray from camera origin $\mathbf{o}$ in viewing direction $\mathbf{d}$ , sample points $\mathbf{x}_j$ along the ray, and apply radiance field rendering using the density and view-dependent colors regressed from our representation (Eqn. 1). We also use the volume rendering weights to determine the surface point $\hat{\mathbf{x}}$ on the ray (Eqn. 2), at which we perform physically based rendering using the normals and material properties (Eqn. 3). We compute accurate visibility $V$ and indirect lighting $L_{\mathrm{ind}}$ using radiance field rendering by marching secondary rays from the surface point $\hat{\mathbf{x}}$ along sampled incoming light direction $\omega_i$ (Sec. 3.3), enabling accurate physically-based rendering. We supervise both the radiance field rendering $C_{\mathrm{RF}}$ and physically-based rendering $C_{\mathrm{PB}}$ with the captured images in a per-scene optimization for joint scene reconstruction (Sec. 3.5). + +accurately reproduce the geometry and appearance of real-world scenes. In particular, NeRF and many following neural field representations [11, 22, 23, 33, 36] have been proposed and applied to enable high-fidelity rendering results on novel view synthesis and many other applications [9, 10, 19, 25, 37]. Originally neural fields [22] are modelled in the form of MLPs, which have limited capacity and high computational costs. Recently, many works [11, 23, 29] introduce more efficient neural scene representations that combine neural feature maps/volumes with light-weight MLPs to reduce the computational cost and accelerate the training and rendering. In this work, we adapt the efficient tensor-factorized neural representation, TensoRF [11], to achieve accurate and efficient inverse rendering. + +Inverse rendering. Abundant works [5,15-17,24,32] have been proposed to infer the geometry and material properties of real-world objects from image collections. They typically represent the scene geometry using triangle meshes that are known or can be pre-reconstructed with depth sensors or multi-view stereo techniques [27]. To reduce the ambiguities in the inverse rendering, they often require controlled lighting conditions [5, 24], or make use of learnt domain-specific priors [2,5,14,20]. In this work, we jointly estimate the geometry, materials and lighting from images captured under unknown lighting conditions with neural field representation that is more efficient and robust. While neural representations have recently been used for inverse rendering tasks, they [3,6,7,28,38,39,41,42] are limited by the usage of computation-intensive MLPs. This inefficiency adds extra burden on inverse rendering when computing + +secondary shading effects (like shadows), which requires to extensively sample secondary rays. Therefore, previous methods often ignore secondary effects [6, 7, 39], consider collocated flash lighting [3, 4, 38], or take extra costs to distill these effects into additional MLP networks [28, 41, 42]. In contrast, we base our model on the advanced TensoRF representation, utilizing factorized tensors, to achieve fast reconstruction. Our TensoRF-based approach supports fast density and radiance evaluation, enabling efficient online computation for secondary effects; this leads to more accurate shadow and indirect lighting modeling during reconstruction, further benefiting our reconstruction. Moreover, in contrast to previous methods [28, 39, 41, 42] that can only handle captures under a single lighting condition, our model is easily extended to support multi-light capture by modeling an additional lighting dimension in the tensor factors. + +# 3. Method + +In this section, we present our tensor factorization-based inverse rendering framework, shown in Fig. 2, which reconstructs scene geometry, materials, and illumination from multi-view input images under unknown lighting. We leverage rendering with both neural radiance fields and physically-based light transport to model and reproduce scene appearance (Sec. 3.1). We introduce a novel TensoRF-based scene representation that allows for both rendering methods in scene reconstruction (Sec. 3.2). Our representation not only enables accurate and efficient computation for visibility and indirect lighting (Sec. 3.3) but also supports capturing under multiple unknown lighting + +conditions (Sec. 3.4). We simultaneously estimate all scene components in a joint optimization framework with rendering losses and regularization terms (Sec. 3.5). + +# 3.1. Rendering + +We apply ray marching to achieve differentiable radiance field rendering as done in NeRF [22]. We further determine the expected surface intersection point for each ray using the volume density, and perform physically-based rendering at the surface point with the predicted scene properties. + +Radiance field rendering. Given a camera ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ from ray origin $\mathbf{o}$ in direction $\mathbf{d}$ , radiance field rendering samples $N$ points on the ray and compute the pixel color as + +$$ +C _ {\mathrm {R F}} (\mathbf {o}, \mathbf {d}) = \sum_ {j = 1} ^ {N} T _ {j} \left(1 - \exp \left(- \sigma_ {j} \delta_ {j}\right)\right) c _ {j} \tag {1} +$$ + +$$ +T _ {j} = \exp \left(- \sum_ {q = 1} ^ {j - 1} \sigma_ {q} \delta_ {q}\right) +$$ + +where $\sigma_{j},\delta_{j},c_{j}$ and $T_{j}$ are volume density, step size, viewdependent radiance color, and volume transmittance at each sampled point $\mathbf{r}(t_j)$ . + +Physically-based rendering. We apply a physically-based parametric BRDF model [8] $f_{r}$ and perform physically-based rendering using predicted geometry and material properties at surface points on the camera rays. Similar to previous methods [22, 41], these surface points $\hat{\mathbf{x}}$ are naturally determined using the volume rendering weights and sampled points from Eqn. 1: + +$$ +\hat {\mathbf {x}} = \sum_ {j = 1} ^ {N} w _ {j} \mathbf {r} \left(t _ {j}\right), \quad w _ {j} = T _ {j} \left(1 - \exp \left(- \sigma_ {j} \delta_ {j}\right)\right) \tag {2} +$$ + +We leverage the classic surface rendering equation to compute a physically-based shading color with an integral over the upper hemisphere $\Omega$ at each surface point: + +$$ +C _ {\mathrm {P B}} (\hat {\mathbf {x}}, \mathbf {d}) = \int_ {\Omega} L _ {\mathrm {i}} (\hat {\mathbf {x}}, \boldsymbol {\omega} _ {i}) f _ {r} (\hat {\mathbf {x}}, \boldsymbol {\omega} _ {i}, \mathbf {d}, \beta) (\boldsymbol {\omega} _ {i} \cdot \mathbf {n}) \mathrm {d} \boldsymbol {\omega} _ {i} \tag {3} +$$ + +where $L_{\mathrm{i}}(\hat{\mathbf{x}},\omega_{i})$ is the incident illumination coming from direction $\boldsymbol{\omega}_{i}$ . $\beta$ and $\mathbf{n}$ represent the spatially-varying material parameters (albedo and roughness) of the BRDF and the surface normal at $\hat{\mathbf{x}}$ . + +In theory, accurately evaluating this integral in Eqn. 3 requires extensively sampling the lighting direction and computing the incident lighting $L_{\mathrm{i}}(\hat{\mathbf{x}}, \boldsymbol{\omega}_i)$ to account for shadowing and indirect illumination. It requires additional ray marching along each lighting direction, which is known to be extremely computation-expensive. Previous NeRF-based methods often simplify it by only considering direct + +illumination [6, 7] or using extra auxiliary MLPs to make approximations and avoid full ray marching [28, 41]. Instead, we evaluate the integral more accurately by marching secondary rays online and computing the incident lighting $L_{\mathrm{i}}(\hat{\mathbf{x}}, \boldsymbol{\omega}_i)$ with accurate shadowing and indirect illumination (see Sec. 3.3). This is made possible by our novel efficient TensorRF-based scene representation. + +# 3.2. TensoRF-Based Representation + +We now introduce our tensor factorization-based scene representation that simultaneously models volume density $\sigma$ , view-dependent color $c$ , shading normal $\mathbf{n}$ , and material properties $\beta$ (including diffuse albedo and specular roughness of the Disney BRDF [8]). Our method can model the scene under both single or multiple lighting conditions. We first introduce the single-light setting in this section and will discuss the multi-light extension in Sec. 3.4. + +At a high-level, radiance, geometry, and material properties can all be represented as a 3D field. We can use a feature volume and voxel feature decoding functions to extract information at any point of the 3D field. To compress the space of the feature volume and also regularize the learning process, TensoRF [11] proposed to use a low-rank factorized tensor as the feature volume. In this work, we adopt the Vector-Matrix factorization proposed by TensoRF. In particular, we use two separate VM-factorized tensors, i.e. feature grids $\mathcal{G}_{\sigma}$ and $\mathcal{G}_a$ , to model volume density and appearance, respectively. The appearance tensor $\mathcal{G}_a$ is followed by multiple light-weight MLP decoders to regress various appearance properties. + +Density tensor. The density tensor $\mathcal{G}_{\sigma}$ is 3D and directly expresses volume density without any network. Specifically, the VM-factorized density tensor is expressed by the sum of vector-matrix outer products. + +$$ +\begin{array}{l} \mathcal {G} _ {\sigma} = \sum_ {k} \mathbf {v} _ {\sigma , k} ^ {X} \circ \mathbf {M} _ {\sigma , k} ^ {Y Z} + \mathbf {v} _ {\sigma , k} ^ {Y} \circ \mathbf {M} _ {\sigma , k} ^ {X Z} + \mathbf {v} _ {\sigma , k} ^ {Z} \circ \mathbf {M} _ {\sigma , k} ^ {X Y} \\ = \sum_ {k} \sum_ {m \in X Y Z} \mathbf {v} _ {\sigma , k} ^ {m} \circ \mathbf {M} _ {\sigma , k} ^ {\tilde {m}} \tag {4} \\ \end{array} +$$ + +Here, $\mathbf{v}_{\sigma ,k}^{m}$ and $\mathbf{M}_{\sigma ,k}^{\tilde{m}}$ represent the $k^{\mathrm{th}}$ vector and matrix factors of their corresponding spatial axes $m$ ; for simplicity, $\tilde{m}$ denotes the two axes orthogonal to $m$ (e.g. $\tilde{X} = YZ$ ). + +Appearance tensor. The appearance tensor $\mathcal{G}_a$ is 4D, modeled by similar vector-matrix spatial factors and additional feature basis vectors $\mathbf{b}$ , expressing a multi-channel voxel feature grid: + +$$ +\mathcal {G} _ {a} = \sum_ {k} \sum_ {m \in X Y Z} \mathbf {v} _ {a, k} ^ {m} \circ \mathbf {M} _ {a, k} ^ {\tilde {m}} \circ \mathbf {b} _ {k} ^ {m} \tag {5} +$$ + +Representing scene properties. We obtain density $\sigma$ directly by linearly interpolating $\mathcal{G}_{\sigma}$ , and apply multiple MLP + +![](images/8b1eb416dabd1b59ac2785762c28e758d091534ab08ac89ab3bfd1730c9b7e87.jpg) +Figure 3. We compare normal and albedo reconstruction results between our joint reconstruction model and an ablated model without radiance field rendering during reconstruction. Radiance field reconstruction is crucial for us to achieve good reconstruction with a clean background and reasonable scene geometry. + +networks to decode appearance properties from interpolated appearance features $\mathcal{G}_a$ . This includes a radiance network $\mathcal{D}_c$ , a shading normal network $\mathcal{D}_{\mathbf{n}}$ , and a material network $\mathcal{D}_{\beta}$ . Overall, our scene representation is expressed by + +$$ +\sigma_ {\mathbf {x}} = \mathcal {G} _ {\sigma} (\mathbf {x}), \quad a _ {\mathbf {x}} = \mathcal {G} _ {a} (\mathbf {x}) +$$ + +$$ +\begin{array}{l} c _ {\mathbf {x}}, \mathbf {n} _ {\mathbf {x}}, \beta_ {\mathbf {x}} = \mathcal {D} _ {c} (a _ {\mathbf {x}}), \mathcal {D} _ {\mathbf {n}} (a _ {\mathbf {x}}), \mathcal {D} _ {\boldsymbol {\beta}} (a _ {\mathbf {x}}) \end{array} \tag {6} +$$ + +Here, $\mathbf{x}$ denotes an arbitrary 3D location, $\sigma_{\mathbf{x}} = \mathcal{G}_{\sigma}(\mathbf{x})$ and $a_{\mathbf{x}} = \mathcal{G}_a(\mathbf{x})$ are the linearly interpolated density and multichannel appearance features, computed by linearly interpolating the spatial vector and matrix factors (please refer to the TensoRF paper [11] for the details of feature computation and interpolation). + +In particular, volume density $\sigma$ and shading normals $\mathbf{n}$ both express scene geometry, which describes global shapes and high-frequency geometric details, respectively. View-dependent color $c$ and physical shading properties (normal $\mathbf{n}$ and material parameters $\beta$ ) duplicatively model the scene appearance, determining the colors in the scene. We bind shading normal with volume density using a regularization term (see details in Sec. 3.5), correlating the scene geometry estimation and appearance reasoning. + +In essence, our TensoRF-based scene representation provides scene geometry and appearance properties that are required for both radiance field rendering (Eqn. 1) and physically-based rendering (Eqn. 3) described in Sec. 3.1. This represents a scene as a radiance field and a physically-based rendering model jointly. We let the two sub-representations share the same neural features in our tensors, allowing their learning processes to benefit each other. While the physically-based model is our main focus and achieves inverse rendering, modeling the radiance field can facilitate the volume density reconstruction and also regularize the appearance features to be meaningful since it has a shorter gradient path. As a result, modeling the radiance field is crucial and necessary for us to achieve high-quality physically-based scene reconstruction as shown in Fig. 3. In addition, the radiance field can naturally provide indirect illumination for physically-based rendering, enabling more + +![](images/7d58a0398b89ce63d3de788a2231f9954035b0352afb67426ec82477548c7572.jpg) +a) Image + +![](images/52cfdba5cbc7f95753745483028bb1b7bc010d52b01d5f374d921fafbcabe9bf.jpg) +b) Indirect lighting + +![](images/7a234fdc6c2f5b0bee2dfb7bdf4195f62a45ce3b57971839a22942e5d119089d.jpg) +c) Lighting visibility under two directional lights +Figure 4. We show our computed indirect illumination (b) of the full rendered image (a) and our lighting visibility (c) under two different directional lights. + +accurate physical model reconstruction. + +# 3.3. Illumination and Visibility + +Our TensoRF-based scene representation is highly efficient for optimization and evaluation. Especially, our volume density can be computed by simple tensor interpolation without using any MLPs, leading to highly efficient computation of volume transmittance and rendering weights (Eqn. 1,2). This allows us to compute accurate incident illumination $L_{\mathrm{i}}$ (as defined in Eqn. 3) that accounts for secondary shading effects with ray marching as shown in Fig. 4. In particular, the incident illumination is computed by + +$$ +L _ {i} (\hat {\mathbf {x}}, \omega_ {i}) = V (\hat {\mathbf {x}}, \omega_ {i}) L _ {\mathrm {d}} (\omega_ {i}) + L _ {\mathrm {i n d}} (\hat {\mathbf {x}}, \omega_ {i}) \tag {7} +$$ + +where $V$ is the light visibility function, $L_{\mathrm{d}}$ is the direct illumination, and $L_{\mathrm{ind}}$ is the indirect illumination. + +Direct illumination. We assume the unknown natural environment to be distant from the captured object and represent the global illumination as an environment map, parameterized by a mixture of spherical Gaussians (SG), which represents the direct illumination. Note that in contrast to previous methods [39,42] that use SG for computing the integral of the rendering equation with a closed-form approximation, we use SG only for its compact parameterized representation and compute the integral numerically by sampling secondary rays and performing ray marching, leading to more accurate lighting visibility and indirect lighting. + +Visibility and indirect illumination. We make use of the jointly-trained radiance fields to model secondary shading effects. More specifically, the indirect illumination arriving at $\hat{\mathbf{x}}$ from $\omega_{i}$ is inherently explained by the radiance color $C_{\mathrm{RF}}$ along the ray $\mathbf{r}_i(t) = \hat{\mathbf{x}} +t\omega_i$ . The visibility function is exactly modeled by the transmittance function in volume rendering. Therefore, the light visibility and indirect illumination term in Eqn. 7 can be calculated as: + +$$ +V (\hat {\mathbf {x}}, \omega_ {i}) = T (\hat {\mathbf {x}}, \omega_ {i}), \quad L _ {\text {i n d}} (\hat {\mathbf {x}}, \omega_ {i}) = C _ {\mathrm {R F}} (\hat {\mathbf {x}}, \omega_ {i}) \tag {8} +$$ + +Here $T(\hat{\mathbf{x}}, \boldsymbol{\omega}_i)$ represents the volume transmittance of the final point sampled on the ray $\mathbf{r}_i(t) = \hat{\mathbf{x}} + t\boldsymbol{\omega}_i$ . + +Second-bounce ray marching. To make our physically-based rendering $C_{\mathrm{PB}}$ accurate, we compute the rendering + +integral (Eqn. 3) via Monte Carlo integration by marching multiple rays from each surface point $\hat{\mathbf{x}}$ with stratified sampling when training. For each ray, we obtain direct illumination from the SGs and compute visibility $T(\hat{\mathbf{x}},\omega_{i})$ and indirect illumination $C_{\mathrm{RF}}(\hat{\mathbf{x}},\omega_{i})$ using Eqn. 1 directly. This second-bounce ray marching is known to be expensive for previous NeRF-based methods, which either requires extremely high computational resources (128 TPUs) [28] or utilizes long (several days of) offline pre-computation [41]; both require training extra MLPs, which takes excessive computational costs and is unable to achieve high accuracy. We instead perform second-bounce ray marching online, achieving higher accuracy (see Fig. 6), which is affordable, thanks to our highly efficient tensor factorization-based representation. + +# 3.4. Multi-Light Representation + +We have discussed our scene representation under a single unknown lighting condition in Sec. 3.2. Our method can be easily extended to support capturing under multiple unknown lighting conditions, owing to the generality and extensibility of tensor factorization. We achieve this by adding an additional lighting dimension with vector factors $\mathbf{e}$ - where the length of each vector $\mathbf{e}$ equals the number of lighting conditions - into the factorized appearance tensor $\mathcal{G}_a$ , leading to a 5D tensor that expresses scene appearance under different lighting conditions. We denote the 5D multi-light appearance tensor as $\mathcal{G}_a^{5\mathrm{D}}$ , represented by the factorization: + +$$ +\mathcal {G} _ {a} ^ {\mathrm {5 D}} = \sum_ {k} \sum_ {m \in X Y Z} \mathbf {v} _ {a, k} ^ {m} \circ \mathbf {M} _ {a, k} ^ {\tilde {m}} \circ \mathbf {e} _ {k} ^ {m} \circ \mathbf {b} _ {k} ^ {m} \tag {9} +$$ + +Note that, it is only the view-dependent colors that require being modeled separately under different lighting, while the physical scene properties - including volume density, normals, and material properties - are inherently shared across multiple lighting conditions. Therefore, we decode view-dependent color using the neural features per light and decode other shading properties using the mean features along the lighting dimension: + +$$ +\bar {a} _ {\mathbf {x}} = \frac {\sum_ {l} a _ {\mathbf {x} , l}}{P}, \quad a _ {\mathbf {x}, l} = \mathcal {G} _ {a} ^ {5 D} (\mathbf {x}, l) \tag {10} +$$ + +$$ +c _ {\mathbf {x}}, \mathbf {n} _ {\mathbf {x}}, \beta_ {\mathbf {x}} = \mathcal {D} _ {c} (a _ {\mathbf {x}, l}), \mathcal {D} _ {\mathbf {n}} (\bar {a} _ {\mathbf {x}}), \mathcal {D} _ {\beta} (\bar {a} _ {\mathbf {x}}) +$$ + +where $l$ is the light index, $P$ is the number of lighting conditions, and $\bar{a}_{\mathbf{x}}$ is the mean feature. + +By simply adding additional vectors in the factorized appearance tensor, our method allows us to efficiently reconstruct and query the radiance under different illumination conditions, thereby still allowing for indirect illumination computation in the multi-light setting with the method discussed in Sec. 3.3. The multi-light input can provide useful + +photometric cues and reduce ambiguity in material prediction, and therefore leads to more accurate reconstruction of geometry and material estimation (see Tab. 1). + +# 3.5. Joint Reconstruction and Training Losses + +For both single- and multi-light settings, we jointly reconstruct the scene geometry and appearance properties that are modeled by our tensor factorization-based representation, through an end-to-end per-scene optimization. + +Rendering losses. We supervise the rendered colors from both the radiance field rendering $C_{\mathrm{RF}}$ and the physically-based rendering $C_{\mathrm{PB}}$ with the ground-truth colors $C_{\mathrm{gt}}$ from the captured images and include two rendering loss terms: + +$$ +\ell_ {\mathrm {R F}} = \left\| C _ {\mathrm {R F}} - C _ {\mathrm {g t}} \right\| _ {2} ^ {2}, \quad \ell_ {\mathrm {P B}} = \left\| C _ {\mathrm {P B}} - C _ {\mathrm {g t}} \right\| _ {2} ^ {2} \tag {11} +$$ + +Normal regularization. Our shading normals $\mathbf{n}$ are regressed from an MLP to express high-frequency surface variations and used in physically-based rendering. However, since the entire system is highly ill-conditioned, only supervising the network output with rendering loss is prone to overfitting issues and produces incorrect normal predictions. On the other hand, many previous NeRF-based method [6, 26, 28, 41] use the negative direction of volume density gradients $\mathbf{n}_{\sigma} = -\frac{\nabla_{\mathbf{x}}\sigma}{\|\nabla_{\mathbf{x}}\sigma\|}$ as normals for shading computation. Such derived normals are better aligned with the surface but can be noisy and lack fine details. Inspired by the rent work Ref-NeRF [31], we correlate our shading normal with the density-derived normal using a loss + +$$ +\ell_ {\mathbf {n}} = \sum_ {j} w _ {j} \| \mathbf {n} _ {j} - \mathbf {n} _ {\sigma , j} \| _ {2} ^ {2} \tag {12} +$$ + +This regularization term back-propagates gradients to both the appearance tensor (through the normal network) and the density tensor, further correlating our entire scene geometry and appearance reasoning, leading to better normal reconstruction. Note that while Ref-NeRF uses the same normal regularization, their normals are used in MLP-based shading computation; in contrast, our normals are used in a physically-based rendering process, allowing for more meaningful geometry reasoning and leading to more accurate normal reconstruction. + +Final loss. Following TensoRF, we apply additional $\ell_1$ -regularization on all tensor factors, denoted as $\ell_{\mathcal{G}}$ . We also apply a smoothness regularization term on BRDF parameters $\beta$ to enhance their spatial consistency. A loss term penalizing back-facing normals is used in addition. Please refer to the supplementary materials for the details of these regularization terms. We apply all these losses to supervise and regularize our whole system to jointly reconstruct the scene, optimizing our scene representation (with all tensor factors and MLPs), as well as the SG parameters of the en + +![](images/dfc68463364d3421dcd711bef49ab7f671c19740f5bbe032c1a88adffc1f9a9a.jpg) +Normal +Albedo +Env. Map +Figure 5. Visual comparison against baseline methods. Our method produces inverse rendering results of higher quality with more detailed normals and more accurate albedo, thus leading to more photo-realistic relighting results. + +
MethodNormal MAE ↓AlbedoNovel View SynthesisRelightingAverage Runtime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor6.31425.1250.9400.10924.6790.9220.12023.3830.9080.131>100 hrs
InvRender5.07427.3410.9330.10027.3670.9340.08923.9730.9010.10115 hrs
Ours in 25 minutes4.87628.2100.9470.09132.3500.9640.06127.4310.9350.09425 mins
Ours4.10029.2750.9500.08535.0880.9760.04028.5800.9440.0815 hrs
Ours w/ rotated multi-light3.60229.6720.9560.07934.3950.9740.04328.9550.9490.0775 hrs
Ours w/ general multi-light3.55129.3260.9510.08434.2230.9730.04529.0080.9470.0785 hrs
+ +Table 1. Quantitative comparisons on the synthetic dataset. Our (single-light) results have significantly outperformed the baseline methods by producing more accurate normal and albedo, thus achieving more realistic novel view synthesis and relighting results. Our method can further take images captured under different lighting conditions, and boost the performance in inverse rendering. (We scale each RGB channel of all albedo results by a global scalar, as done in NeRFactor [41]. For a fair comparison, all novel view synthesis results are generated with physically-based rendering, though our radiance field rendering has better quality.) + +vironment map, with a final loss + +$$ +\ell = \alpha_ {\mathrm {R F}} \ell_ {\mathrm {R F}} + \alpha_ {\mathrm {P B}} \ell_ {\mathrm {P B}} + \alpha_ {\beta} \ell_ {\beta} + \alpha_ {\mathbf {n}} \ell_ {\mathbf {n}} + \alpha_ {\mathbf {d}} \ell_ {\mathbf {d}} + \alpha_ {\mathcal {G}} \ell_ {\mathcal {G}} \tag {13} +$$ + +# 4. Experiments + +We now evaluate our model on various challenging scenes. We make comparisons against previous state-of-the-art methods and also present an ablation study to verify the effectiveness of our design choices. + +Datasets. We perform experiments on four complex synthetic scenes, including three blender scenes from [22] and one from the Stanford 3D scanning repository [12]. We rerender these scenes to obtain their ground-truth images, as well as BRDF parameters and normal maps. We also render two types of multi-light data: rotated multi-light and general multi-light. We also perform experiments on the original NeRF-synthetic dataset and 4 captured real data. Please refer to the supplementary for experimental results of those extra data, more details about our datasets, and more explanation about our multi-light setting. + +Comparisons with previous methods. We compare our + +model with previous state-of-the-art neural field-based inverse rendering methods, NeRFactor [41] and InvRender [42], on these scenes using images captured under a single unknown lighting condition. We also compare with our model trained under multi-light settings. Table 1 shows the accuracy of the estimated albedo, normal, and relighting results using metrics including PSNR, SSIM and LPIPS [40] averaged over the four scenes. We can see that our single-light method (using the same input as baselines) outperforms both previous methods significantly on all metrics, demonstrating the superiority of our method. We also include a visual comparison in Fig. 5, showing that our method predicts more accurate albedo and normal that are closer to the ground truth compared to baseline methods, thus generating more realistic relighting results. In particular, our results produce normals that are of much higher quality and more faithfully reflect the geometry variations, while the baseline methods produce over-smooth results that lack shape details. Although NeRFactor's albedo result on the lego scene looks closer to the ground truth than our results. We claim that this is because that NeRFactor uses a high-weight BRDF smoothness loss, which dam + +![](images/9bc328725e2aab2b7aed3a16410bf380357916d8d0d3f8c4a17b93d10a6a4a07.jpg) +NeRFactor + +![](images/cc550f7267b1e4434e90f6597acdd028c039d528480ec21c86d3d66e5143dcbf.jpg) +InvRender + +![](images/0a2ee602f1c2fa211fc064bdf18a265f9be2a078de1e3e0578eb9710ec7ca18b.jpg) +Ours +Figure 6. We compare our relighting results with previous methods. Note that our approach recovers more accurate shadows thanks to our second-bounce ray marching. + +![](images/c8665f30e03f7319af2a70d67d09193e38cf8aaf645b3db29315e95c0537b6d3.jpg) +Ground truth + +ages its reconstruction quality of other components. In the supplementary, we showcase that we can achieve a similar albedo result by increasing the BRDF smoothness weight. In particular, NeRFactor approximates the visibility function in a distilled MLP from a pre-trained NeRF and completely ignores the indirect illumination. While InvRender considers both visibility and indirect illumination in their pre-computation, its spherical Gaussian-based shading computation can only achieve limited accuracy. Both methods do inverse rendering in a second stage after pretraining NeRFs. In contrast, our tensor-factorized representation achieves a single-stage joint reconstruction and efficiently performs explicit second-bounce ray marching for more accurate shadowing (see Fig. 6) and indirect lighting, thus leading to significantly higher reconstruction quality. + +Meanwhile, owing to our more efficient representation, our high reconstruction quality is achieved with the fastest reconstruction speed, as reflected by the reconstruction time. NeRF-based method NeRFactor takes days to compute because of visibility pre-computation. InvRender is faster than NeRFactor, due to its SDF-based rendering [36] and SG-based closed-form rendering integral computation; however, its SDF-based reconstruction fails on challenging scenes like the Lego shown in Fig. 5 and the SG-based integration leads to inaccurate secondary shading effects as mentioned. On the other hand, our approach leverages ray marching-based radiance rendering in the reconstruction, robustly producing high-quality results on all testing scenes with accurate secondary effects, while still being faster than InvRender. In fact, while our method takes 5 hours to finish its full training for its best performance, it can achieve good quality in a much shorter training period (25 minutes). As shown in Tab. 1 and Fig. 7, our approach can achieve high-quality geometry reconstruction in only 25 minutes and outperform previous methods that trained for tens of or even hundreds of hours. + +Multi-light results. In addition, our framework can also effectively leverage the additional input in a multi-light capture and further boost the accuracy of the inverse rendering performance without adding additional computation costs, as shown in Tab. 1, while the MLP-based baseline methods cannot be trivially extended to support such setups in an efficient manner. We also found that multi-light settings can greatly improve geometry reconstruction and help solving + +
MethodNormal MAE ↓Albedo PSNR ↑NVS PSNR ↑
w/o visibility4.71627.26534.703
w/o indirect illum.4.18629.00334.910
w/ indirect illum. + visibility4.10029.27535.088
+ +Table 2. Our full method models more physically-accurate light transport by accounting for lighting visibility and indirect illuminations, thus achieving much better accuracy in inverse rendering. + +![](images/c92480335c778171f47c7a47d0171ade181590b3bd119bb0d8d622253708edeb.jpg) +NeRFactor $(>100\mathrm{hrs})$ +Figure 7. We compare geometry reconstruction results of our model taking only 25 minutes of optimization, with previous methods, taking 15 and $>100$ hours. Our approach even recovers more high-frequency details with substantially less reconstruction time. + +![](images/ebe63a888fc657481ca6c8dc52ba52f29708978f13ac7c72e9ed3d2091567978.jpg) +InvRender (15 hrs) + +![](images/a2260a2848ccbb38377b69d8767ca1b15a63cf24f3555ed7765c8485b5832dbc.jpg) +Ours (25 mins) + +![](images/10eddaf8a4e80fde80c8911f775ad181be81d749ed2d4ef789c6b24c17fc3898.jpg) +Ground truth + +the color ambiguity between lighting and materials. Please refer the supplementary for more results and analysis. + +Indirect illumination and visibility. Our tensor-factorized representation allows us to explicit sample secondary rays in an efficient way to account for computing lighting visibility and indirect illumination. As shown in Tab. 2, without including these terms, the model cannot accurately represent the secondary shading effects and tend to bake them in the albedo or normal, resulting in lower quality. Nonetheless, note that these ablated models of our method in fact already achieves superior quality compared to previous methods (that pre-computes NeRFs and do inverse rendering disjointly) shown in Tab. 1, demonstrating the effectiveness of our joint reconstruction framework. Our full method further achieves more accurate reconstruction because of more accurate light transport modeling. + +# 5. Conclusion + +We present a novel inverse rendering approach that achieves efficient and high-quality scene reconstruction from multi-view images under unknown illumination. Our approach models a scene as both a radiance field and a physically-based model with density, normals, lighting, and material properties. By jointly reconstructing both models, we achieve high-quality geometry and reflectance reconstruction, enabling photo-realistic rendering and relighting. Owing to the efficiency and generality of the tensor factorized representation, our framework allows for accurate computation for shadowing and indirect lighting effects, and also flexibly supports capturing under an arbitrary number of lighting conditions. We demonstrate that our approach is able to achieve state-of-the-art inverse rendering results, outperforming previous neural methods in terms of both reconstruction quality and efficiency. + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 2 +[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2015. 3 +[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2, 3 +[4] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In European Conference on Computer Vision, pages 294–311. Springer, 2020. 2, 3 +[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5960-5969, 2020. 3 +[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2, 3, 4, 6 +[7] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 1, 2, 3, 4 +[8] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, volume 2012, pages 1-7. vol. 2012, 2012. 4 +[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 3 +[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3 +[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1, 3, 4, 5 +[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 7 +[13] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the + +reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2 +[14] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics, 33(6):193, 2014. 3 +[15] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying brdfs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2, 3 +[16] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):548-554, 2008. 3 +[17] Jason Lawrence, Szymon Rusinkiewicz, and Ravi Ramamoorthi. Efficient brdf importance sampling using a factored representation. ACM Transactions on Graphics (ToG), 23(3):496-505, 2004. 3 +[18] Hendrik PA Lensch, Jochen Lang, Asla M Sá, and Hans-Peter Seidel. Planned sampling of spatially varying brdfs. In Computer graphics forum, volume 22, pages 473-482. Wiley Online Library, 2003. 2 +[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3 +[20] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. In SIGGRAPH Asia 2018, page 269. ACM, 2018. 2, 3 +[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2 +[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 4, 7 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 3 +[24] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. In SIGGRAPH Asia 2018, page 267. ACM, 2018. 3 +[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3 +[26] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 6 + +[27] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 3 +[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2, 3, 4, 6 +[29] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 3 +[30] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2 +[31] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 6 +[32] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics, 35(6):187, 2016. 3 +[33] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3 +[34] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics, 38(4):76, 2019. 2 +[35] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics, 37(4):126, 2018. 2 +[36] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 3, 8 +[37] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 3 +[38] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 3 +[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2, 3, 5 + +[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 3, 4, 6, 7 +[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 3, 5, 7 \ No newline at end of file diff --git a/2023/TensoIR_ Tensorial Inverse Rendering/images.zip b/2023/TensoIR_ Tensorial Inverse Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..adbb040b24b5e6c0839c7cfe2b472c5efa33c3bb --- /dev/null +++ b/2023/TensoIR_ Tensorial Inverse Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f1e3b02e4498c752b41dace248db025b03efaa5be8a342fa8edd6efd886096c +size 531272 diff --git a/2023/TensoIR_ Tensorial Inverse Rendering/layout.json b/2023/TensoIR_ Tensorial Inverse Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f5463d5605dfc7100744d3075467cc3ad7e1828 --- /dev/null +++ b/2023/TensoIR_ Tensorial Inverse Rendering/layout.json @@ -0,0 +1,8667 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 179, + 103, + 416, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 103, + 416, + 121 + ], + "spans": [ + { + "bbox": [ + 179, + 103, + 416, + 121 + ], + "type": "text", + "content": "TensoIR: Tensorial Inverse Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "type": "text", + "content": "Haian Jin\\*1 Isabella Liu\\*2 Peijia " + }, + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^3" + }, + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "type": "text", + "content": " Xiaoshuai Zhang2 Songfang Han2 Sai Bi4 Xiaowei Zhou' Zexiang " + }, + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{\\dagger_4}" + }, + { + "bbox": [ + 89, + 141, + 504, + 175 + ], + "type": "text", + "content": " Hao Su" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 186, + 520, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 186, + 520, + 204 + ], + "spans": [ + { + "bbox": [ + 70, + 186, + 520, + 204 + ], + "type": "text", + "content": "1 Zhejiang University 2 UC San Diego 3 Kingstar Technology Inc. 4 Adobe Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 266, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 266, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 290, + 531 + ], + "type": "text", + "content": "We propose TensoIR, a novel inverse rendering approach based on tensor factorization and neural fields. Unlike previous works that use purely MLP-based neural fields, thus suffering from low capacity and high computation costs, we extend TensoRF, a state-of-the-art approach for radiance field modeling, to estimate scene geometry, surface reflectance, and environment illumination from multiview images captured under unknown lighting conditions. Our approach jointly achieves radiance field reconstruction and physically-based model estimation, leading to photorealistic novel view synthesis and relighting results. Benefiting from the efficiency and extensibility of the TensoRF-based representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under single or multiple unknown lighting conditions. The low-rank tensor representation allows us to not only achieve fast and compact reconstruction but also better exploit shared information under an arbitrary number of capturing lighting conditions. We demonstrate the superiority of our method to baseline methods qualitatively and quantitatively on various challenging synthetic and real-world scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 694 + ], + "type": "text", + "content": "Inverse rendering is a long-standing problem in computer vision and graphics, aiming to reconstruct physical attributes (like shape and materials) of a 3D scene from captured images and thereby supporting many downstream applications such as novel view synthesis, relighting and material editing. This problem is inherently challenging and ill-posed, especially when the input images are captured in the wild under unknown illumination. Recent works [6,7,28,41] address this problem by learning neural field representations in the form of multi-layer perceptrons" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 545, + 387 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 387 + ], + "type": "image", + "image_path": "131986b44b1a280cdb70bc2d86761bc81c8ca91ae8d8670c6fc7f6559c4755b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 394, + 547, + 450 + ], + "lines": [ + { + "bbox": [ + 305, + 394, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 547, + 450 + ], + "type": "text", + "content": "Figure 1. Given multi-view captured images of a real scene (a), our approach - TensoIR - is able to achieve high-quality shape and material reconstruction with high-frequency details (b). This allows us to render the scene under novel lighting and viewpoints (c), and also change its material properties (d)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 471, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 519 + ], + "type": "text", + "content": "MLP) similar to NeRF [22]. However, pure MLP-based methods usually suffer from low capacity and high computational costs, greatly limiting the accuracy and efficiency of inverse rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 688 + ], + "type": "text", + "content": "In this work, we propose a novel inverse rendering framework that is efficient and accurate. Instead of purely using MLPs, we build upon the recent TensoRF [11] scene representation, which achieves fast, compact, and state-of-the-art quality on radiance field reconstruction for novel view synthesis. Our tensor factorization-based inverse rendering framework can simultaneously estimate scene geometry, materials, and illumination from multi-view images captured under unknown lighting conditions. Benefiting from the efficiency and extensibility of the TensoRF representation, our method can accurately model secondary shading effects (like shadows and indirect lighting) and generally support input images captured under a single or multiple unknown lighting conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Similar to TensoRF, our approach models a scene as a neural voxel feature grid, factorized as multiple low-rank" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 190, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 190, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 190, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Equal advisory." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "165" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": "tensor components. We apply multiple small MLPs on the same feature grid and regress volume density, view-dependent color, normal, and material properties, to model the scene geometry and appearance. This allows us to simultaneously achieve both radiance field rendering – using density and view-dependent color, as done in NeRF [22] – and physically-based rendering – using density, normal and material properties, as done in inverse rendering methods [3, 20]. We supervise both renderings with the captured images to jointly reconstruct all scene components. In essence, we reconstruct a scene using both a radiance field and a physically-based model to reproduce the scene's appearance. While inverse rendering is our focus and primarily enabled by the physically-based model, modeling the radiance field is crucial for the success of the reconstruction (see Fig. 3), in significantly facilitating the volume density reconstruction and effectively regularizing the same tensor features shared by the physically-based model. Despite that previous works [41] similarly reconstruct NeRFs in inverse rendering, their radiance field is pre-computed and fixed in the subsequent inverse rendering stage; in contrast, our radiance field is jointly reconstructed and also benefits the physically-based rendering model estimation during optimization, leading to much higher quality. Besides, our radiance field rendering can also be directly used to provide accurate indirect illumination for the physically-based rendering, further benefiting the inverse rendering process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 399, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 289, + 615 + ], + "type": "text", + "content": "Accounting for indirect illumination and shadowing is a critical challenge in inverse rendering. This is especially challenging for volume rendering, since it requires sampling a lot of secondary rays and computing the integrals along the rays by performing ray marching. Limited by the high-cost MLP evaluation, previous NeRF-based methods and SDF-based methods either simply ignore secondary effects [6, 7, 39], or avoid online computation by approximating these effects in extra distilled MLPs [41, 42], requiring expensive pre-computation and leading to degradation in accuracy. In contrast, owing to our efficient tensor-factorized representation, we are able to explicitly compute the ray integrals online for accurate visibility and indirect lighting with the radiance field rendering using low-cost second-bounce ray marching. Consequently, our approach enables higher accuracy in modeling these secondary effects, which is crucial in achieving high-quality scene reconstruction (see Tab. 2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "In addition, the flexibility and efficiency of our tensor-factorized representation allows us to perform inverse rendering from multiple unknown lighting conditions with limited GPU memory. Multi-light capture is known to be beneficial for inverse rendering tasks by providing useful photometric cues and reducing ambiguities in material estimation, thus being commonly used [13,15,18]. However, since each lighting condition corresponds to a separate radiance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": "field, this can lead to extremely high computational costs if reconstructing multiple purely MLP-based NeRFs like previous works [28, 41, 42]. Instead, we propose to reconstruct radiance fields under multi-light in a joint manner as a factorized tensor. Extending from the original TensorRF representation that is a 4D tensor, we add an additional dimension corresponding to different lighting conditions, yielding a 5D tensor. Specifically, we add an additional vector factor (whose length equals the number of lights) per tensor component to explain the appearance variations under different lighting conditions, and we store this 5D tensor by a small number of bases whose outer-product reconstructs the 5D tensor. When multi-light capture is available, our framework can effectively utilize the additional photometric cues in the data, leading to better reconstruction quality than a single-light setting (see Tab. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 264, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 456 + ], + "type": "text", + "content": "As shown in Fig. 1, our approach can reconstruct high-fidelity geometry and reflectance of a complex real scene captured under unknown natural illumination, enabling photo-realistic rendering under novel lighting conditions and additional applications like material editing. We evaluate our framework extensively on both synthetic and real data. Our approach outperforms previous inverse rendering methods [41, 42] by a large margin qualitatively and quantitatively on challenging synthetic scenes, achieving state-of-the-art quality in scene reconstruction – for both geometry and material properties – and rendering – for both novel view synthesis and relighting. Owing to our efficient tensorial representation and joint reconstruction scheme, our approach also leads to a much faster reconstruction speed than previous neural field-based reconstruction methods while achieving superior quality. In summary," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 460, + 545, + 596 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 460, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 545, + 519 + ], + "type": "text", + "content": "- We propose a novel tensor factorization-based inverse rendering approach that jointly achieves physically-based rendering model estimation and radiance field reconstruction, leading to state-of-the-art scene reconstruction results;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 521, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 521, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 521, + 545, + 557 + ], + "type": "text", + "content": "- Our framework includes an efficient online visibility and indirect lighting computation technique, providing accurate second-bounce shading effects;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 560, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 560, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 317, + 560, + 545, + 596 + ], + "type": "text", + "content": "- We enable efficient multi-light reconstruction by modeling an additional lighting dimension in the factorized tensorial representation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 614, + 398, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 398, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 398, + 627 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Neural scene representations. Neural representations [1, 4, 21, 22, 30, 34, 35], as a promising alternative to traditional representations (like meshes, volumes, and point clouds), have been revolutionizing 3D content generation and modeling. Compared to traditional representations, such neural representations are more flexible and can more" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "166" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 532, + 222 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 532, + 222 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 532, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 532, + 222 + ], + "type": "image", + "image_path": "22b22337385c57a5fedf6367b1ac1da7d3c2cee834ce11c9b639d90ca8e86789.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": "Figure 2. Overview. We propose a novel inverse rendering approach to reconstruct scene geometry, materials, and unknown natural illumination (as an environment map) from captured images. We reconstruct a scene as a novel representation (Sec. 3.2) that uses factorized tensors and multiple MLPs to regress volume density " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": ", view-dependent color " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": ", normals " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": ", and material properties (i.e. BRDF parameters) " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": ", enabling both radiance field rendering and physically-based rendering (Sec. 3.1). In particular, we march a camera ray from camera origin " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " in viewing direction " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": ", sample points " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " along the ray, and apply radiance field rendering using the density and view-dependent colors regressed from our representation (Eqn. 1). We also use the volume rendering weights to determine the surface point " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " on the ray (Eqn. 2), at which we perform physically based rendering using the normals and material properties (Eqn. 3). We compute accurate visibility " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " and indirect lighting " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " using radiance field rendering by marching secondary rays from the surface point " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " along sampled incoming light direction " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " (Sec. 3.3), enabling accurate physically-based rendering. We supervise both the radiance field rendering " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{RF}}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " and physically-based rendering " + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{PB}}" + }, + { + "bbox": [ + 46, + 224, + 548, + 336 + ], + "type": "text", + "content": " with the captured images in a per-scene optimization for joint scene reconstruction (Sec. 3.5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 289, + 515 + ], + "type": "text", + "content": "accurately reproduce the geometry and appearance of real-world scenes. In particular, NeRF and many following neural field representations [11, 22, 23, 33, 36] have been proposed and applied to enable high-fidelity rendering results on novel view synthesis and many other applications [9, 10, 19, 25, 37]. Originally neural fields [22] are modelled in the form of MLPs, which have limited capacity and high computational costs. Recently, many works [11, 23, 29] introduce more efficient neural scene representations that combine neural feature maps/volumes with light-weight MLPs to reduce the computational cost and accelerate the training and rendering. In this work, we adapt the efficient tensor-factorized neural representation, TensoRF [11], to achieve accurate and efficient inverse rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Inverse rendering. Abundant works [5,15-17,24,32] have been proposed to infer the geometry and material properties of real-world objects from image collections. They typically represent the scene geometry using triangle meshes that are known or can be pre-reconstructed with depth sensors or multi-view stereo techniques [27]. To reduce the ambiguities in the inverse rendering, they often require controlled lighting conditions [5, 24], or make use of learnt domain-specific priors [2,5,14,20]. In this work, we jointly estimate the geometry, materials and lighting from images captured under unknown lighting conditions with neural field representation that is more efficient and robust. While neural representations have recently been used for inverse rendering tasks, they [3,6,7,28,38,39,41,42] are limited by the usage of computation-intensive MLPs. This inefficiency adds extra burden on inverse rendering when computing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 540 + ], + "type": "text", + "content": "secondary shading effects (like shadows), which requires to extensively sample secondary rays. Therefore, previous methods often ignore secondary effects [6, 7, 39], consider collocated flash lighting [3, 4, 38], or take extra costs to distill these effects into additional MLP networks [28, 41, 42]. In contrast, we base our model on the advanced TensoRF representation, utilizing factorized tensors, to achieve fast reconstruction. Our TensoRF-based approach supports fast density and radiance evaluation, enabling efficient online computation for secondary effects; this leads to more accurate shadow and indirect lighting modeling during reconstruction, further benefiting our reconstruction. Moreover, in contrast to previous methods [28, 39, 41, 42] that can only handle captures under a single lighting condition, our model is easily extended to support multi-light capture by modeling an additional lighting dimension in the tensor factors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 550, + 362, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 362, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 362, + 562 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "In this section, we present our tensor factorization-based inverse rendering framework, shown in Fig. 2, which reconstructs scene geometry, materials, and illumination from multi-view input images under unknown lighting. We leverage rendering with both neural radiance fields and physically-based light transport to model and reproduce scene appearance (Sec. 3.1). We introduce a novel TensoRF-based scene representation that allows for both rendering methods in scene reconstruction (Sec. 3.2). Our representation not only enables accurate and efficient computation for visibility and indirect lighting (Sec. 3.3) but also supports capturing under multiple unknown lighting" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "167" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "conditions (Sec. 3.4). We simultaneously estimate all scene components in a joint optimization framework with rendering losses and regularization terms (Sec. 3.5)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 116, + 120, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 120, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 120, + 129 + ], + "type": "text", + "content": "3.1. Rendering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 135, + 287, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 135, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 135, + 287, + 196 + ], + "type": "text", + "content": "We apply ray marching to achieve differentiable radiance field rendering as done in NeRF [22]. We further determine the expected surface intersection point for each ray using the volume density, and perform physically-based rendering at the surface point with the predicted scene properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": "Radiance field rendering. Given a camera ray " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": " from ray origin " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": " in direction " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": ", radiance field rendering samples " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": " points on the ray and compute the pixel color as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 247, + 287, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 247, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 83, + 247, + 287, + 288 + ], + "type": "interline_equation", + "content": "C _ {\\mathrm {R F}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {j = 1} ^ {N} T _ {j} \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right) c _ {j} \\tag {1}", + "image_path": "30bd81a290c688fe00281aaf62e3da410f0705b8096b1776bafbd323a48871e8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 285, + 212, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 285, + 212, + 317 + ], + "spans": [ + { + "bbox": [ + 121, + 285, + 212, + 317 + ], + "type": "interline_equation", + "content": "T _ {j} = \\exp \\left(- \\sum_ {q = 1} ^ {j - 1} \\sigma_ {q} \\delta_ {q}\\right)", + "image_path": "13af3147353f03f5f1a516dd4af6390e858e2f851390b88225766b09cd72c363.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\sigma_{j},\\delta_{j},c_{j}" + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "content": " are volume density, step size, viewdependent radiance color, and volume transmittance at each sampled point " + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t_j)" + }, + { + "bbox": [ + 47, + 323, + 287, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "content": "Physically-based rendering. We apply a physically-based parametric BRDF model [8] " + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "content": " and perform physically-based rendering using predicted geometry and material properties at surface points on the camera rays. Similar to previous methods [22, 41], these surface points " + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 47, + 366, + 287, + 450 + ], + "type": "text", + "content": " are naturally determined using the volume rendering weights and sampled points from Eqn. 1:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 460, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 460, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 64, + 460, + 287, + 493 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\sum_ {j = 1} ^ {N} w _ {j} \\mathbf {r} \\left(t _ {j}\\right), \\quad w _ {j} = T _ {j} \\left(1 - \\exp \\left(- \\sigma_ {j} \\delta_ {j}\\right)\\right) \\tag {2}", + "image_path": "773285094af2a8f3effdc86c10e3c9fc3d7b7c8dadf4e99c625636d5b6f9944c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 502, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 538 + ], + "type": "text", + "content": "We leverage the classic surface rendering equation to compute a physically-based shading color with an integral over the upper hemisphere " + }, + { + "bbox": [ + 47, + 502, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 47, + 502, + 287, + 538 + ], + "type": "text", + "content": " at each surface point:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 548, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 548, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 52, + 548, + 287, + 572 + ], + "type": "interline_equation", + "content": "C _ {\\mathrm {P B}} (\\hat {\\mathbf {x}}, \\mathbf {d}) = \\int_ {\\Omega} L _ {\\mathrm {i}} (\\hat {\\mathbf {x}}, \\boldsymbol {\\omega} _ {i}) f _ {r} (\\hat {\\mathbf {x}}, \\boldsymbol {\\omega} _ {i}, \\mathbf {d}, \\beta) (\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {i} \\tag {3}", + "image_path": "709d1981733cfa5ce35b49b3dabbbbbf25b34e8ba70d02dd76c903d6de8e5937.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}(\\hat{\\mathbf{x}},\\omega_{i})" + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": " is the incident illumination coming from direction " + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}_{i}" + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": " represent the spatially-varying material parameters (albedo and roughness) of the BRDF and the surface normal at " + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 47, + 581, + 287, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": "In theory, accurately evaluating this integral in Eqn. 3 requires extensively sampling the lighting direction and computing the incident lighting " + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i)" + }, + { + "bbox": [ + 47, + 630, + 287, + 713 + ], + "type": "text", + "content": " to account for shadowing and indirect illumination. It requires additional ray marching along each lighting direction, which is known to be extremely computation-expensive. Previous NeRF-based methods often simplify it by only considering direct" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": "illumination [6, 7] or using extra auxiliary MLPs to make approximations and avoid full ray marching [28, 41]. Instead, we evaluate the integral more accurately by marching secondary rays online and computing the incident lighting " + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i)" + }, + { + "bbox": [ + 305, + 72, + 545, + 156 + ], + "type": "text", + "content": " with accurate shadowing and indirect illumination (see Sec. 3.3). This is made possible by our novel efficient TensorRF-based scene representation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 163, + 476, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 163, + 476, + 175 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 476, + 175 + ], + "type": "text", + "content": "3.2. TensoRF-Based Representation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "text", + "content": "We now introduce our tensor factorization-based scene representation that simultaneously models volume density " + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "text", + "content": ", view-dependent color " + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "text", + "content": ", shading normal " + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "text", + "content": ", and material properties " + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 181, + 545, + 276 + ], + "type": "text", + "content": " (including diffuse albedo and specular roughness of the Disney BRDF [8]). Our method can model the scene under both single or multiple lighting conditions. We first introduce the single-light setting in this section and will discuss the multi-light extension in Sec. 3.4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "text", + "content": "At a high-level, radiance, geometry, and material properties can all be represented as a 3D field. We can use a feature volume and voxel feature decoding functions to extract information at any point of the 3D field. To compress the space of the feature volume and also regularize the learning process, TensoRF [11] proposed to use a low-rank factorized tensor as the feature volume. In this work, we adopt the Vector-Matrix factorization proposed by TensoRF. In particular, we use two separate VM-factorized tensors, i.e. feature grids " + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\sigma}" + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_a" + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "text", + "content": ", to model volume density and appearance, respectively. The appearance tensor " + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_a" + }, + { + "bbox": [ + 305, + 277, + 545, + 433 + ], + "type": "text", + "content": " is followed by multiple light-weight MLP decoders to regress various appearance properties." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 439, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 487 + ], + "type": "text", + "content": "Density tensor. The density tensor " + }, + { + "bbox": [ + 305, + 439, + 545, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\sigma}" + }, + { + "bbox": [ + 305, + 439, + 545, + 487 + ], + "type": "text", + "content": " is 3D and directly expresses volume density without any network. Specifically, the VM-factorized density tensor is expressed by the sum of vector-matrix outer products." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 493, + 545, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 545, + 546 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {G} _ {\\sigma} = \\sum_ {k} \\mathbf {v} _ {\\sigma , k} ^ {X} \\circ \\mathbf {M} _ {\\sigma , k} ^ {Y Z} + \\mathbf {v} _ {\\sigma , k} ^ {Y} \\circ \\mathbf {M} _ {\\sigma , k} ^ {X Z} + \\mathbf {v} _ {\\sigma , k} ^ {Z} \\circ \\mathbf {M} _ {\\sigma , k} ^ {X Y} \\\\ = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {\\sigma , k} ^ {m} \\circ \\mathbf {M} _ {\\sigma , k} ^ {\\tilde {m}} \\tag {4} \\\\ \\end{array}", + "image_path": "d1bcb9388d0eed17ac54311198beb7d6c800e22a4c924e205ee08903bc7d8276.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\sigma ,k}^{m}" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\sigma ,k}^{\\tilde{m}}" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": " vector and matrix factors of their corresponding spatial axes " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": "; for simplicity, " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": " denotes the two axes orthogonal to " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": " (e.g. " + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{X} = YZ" + }, + { + "bbox": [ + 306, + 553, + 545, + 590 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "type": "text", + "content": "Appearance tensor. The appearance tensor " + }, + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_a" + }, + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "type": "text", + "content": " is 4D, modeled by similar vector-matrix spatial factors and additional feature basis vectors " + }, + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 305, + 597, + 545, + 644 + ], + "type": "text", + "content": ", expressing a multi-channel voxel feature grid:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 350, + 651, + 545, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 651, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 350, + 651, + 545, + 676 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {a} = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {a, k} ^ {m} \\circ \\mathbf {M} _ {a, k} ^ {\\tilde {m}} \\circ \\mathbf {b} _ {k} ^ {m} \\tag {5}", + "image_path": "101459f590e9b66380772e05b50a1cc038073222e3eb31200638d2bf7c53e1e9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Representing scene properties. We obtain density " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " directly by linearly interpolating " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\sigma}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": ", and apply multiple MLP" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "168" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 171 + ], + "type": "image", + "image_path": "8b1eb416dabd1b59ac2785762c28e758d091534ab08ac89ab3bfd1730c9b7e87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 232 + ], + "type": "text", + "content": "Figure 3. We compare normal and albedo reconstruction results between our joint reconstruction model and an ablated model without radiance field rendering during reconstruction. Radiance field reconstruction is crucial for us to achieve good reconstruction with a clean background and reasonable scene geometry." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "text", + "content": "networks to decode appearance properties from interpolated appearance features " + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_a" + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "text", + "content": ". This includes a radiance network " + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "text", + "content": ", a shading normal network " + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathbf{n}}" + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "text", + "content": ", and a material network " + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\beta}" + }, + { + "bbox": [ + 47, + 237, + 288, + 286 + ], + "type": "text", + "content": ". Overall, our scene representation is expressed by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 294, + 225, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 294, + 225, + 308 + ], + "spans": [ + { + "bbox": [ + 110, + 294, + 225, + 308 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\mathbf {x}} = \\mathcal {G} _ {\\sigma} (\\mathbf {x}), \\quad a _ {\\mathbf {x}} = \\mathcal {G} _ {a} (\\mathbf {x})", + "image_path": "8c13295e95e1aeb373cbbb4f187f2a616aa2aa4c6f537946d4ff9e3460b6b3f6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 303, + 287, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 303, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 88, + 303, + 287, + 322 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} c _ {\\mathbf {x}}, \\mathbf {n} _ {\\mathbf {x}}, \\beta_ {\\mathbf {x}} = \\mathcal {D} _ {c} (a _ {\\mathbf {x}}), \\mathcal {D} _ {\\mathbf {n}} (a _ {\\mathbf {x}}), \\mathcal {D} _ {\\boldsymbol {\\beta}} (a _ {\\mathbf {x}}) \\end{array} \\tag {6}", + "image_path": "ea2ee219fbb5e77a6b8a2f0bceef93f8a55818fe859fb114e5f6e329271aa26b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "content": " denotes an arbitrary 3D location, " + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{x}} = \\mathcal{G}_{\\sigma}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "inline_equation", + "content": "a_{\\mathbf{x}} = \\mathcal{G}_a(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "content": " are the linearly interpolated density and multichannel appearance features, computed by linearly interpolating the spatial vector and matrix factors (please refer to the TensoRF paper [11] for the details of feature computation and interpolation)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "content": "In particular, volume density " + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "content": " and shading normals " + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "content": " both express scene geometry, which describes global shapes and high-frequency geometric details, respectively. View-dependent color " + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "content": " and physical shading properties (normal " + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "content": " and material parameters " + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 403, + 287, + 510 + ], + "type": "text", + "content": ") duplicatively model the scene appearance, determining the colors in the scene. We bind shading normal with volume density using a regularization term (see details in Sec. 3.5), correlating the scene geometry estimation and appearance reasoning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 714 + ], + "type": "text", + "content": "In essence, our TensoRF-based scene representation provides scene geometry and appearance properties that are required for both radiance field rendering (Eqn. 1) and physically-based rendering (Eqn. 3) described in Sec. 3.1. This represents a scene as a radiance field and a physically-based rendering model jointly. We let the two sub-representations share the same neural features in our tensors, allowing their learning processes to benefit each other. While the physically-based model is our main focus and achieves inverse rendering, modeling the radiance field can facilitate the volume density reconstruction and also regularize the appearance features to be meaningful since it has a shorter gradient path. As a result, modeling the radiance field is crucial and necessary for us to achieve high-quality physically-based scene reconstruction as shown in Fig. 3. In addition, the radiance field can naturally provide indirect illumination for physically-based rendering, enabling more" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 359, + 113 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 359, + 113 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 359, + 113 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 359, + 113 + ], + "type": "image", + "image_path": "7d58a0398b89ce63d3de788a2231f9954035b0352afb67426ec82477548c7572.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 118, + 338, + 125 + ], + "lines": [ + { + "bbox": [ + 316, + 118, + 338, + 125 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 338, + 125 + ], + "type": "text", + "content": "a) Image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 418, + 116 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 418, + 116 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 418, + 116 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 418, + 116 + ], + "type": "image", + "image_path": "52cfdba5cbc7f95753745483028bb1b7bc010d52b01d5f374d921fafbcabe9bf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 118, + 409, + 125 + ], + "lines": [ + { + "bbox": [ + 365, + 118, + 409, + 125 + ], + "spans": [ + { + "bbox": [ + 365, + 118, + 409, + 125 + ], + "type": "text", + "content": "b) Indirect lighting" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 543, + 116 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 543, + 116 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 543, + 116 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 543, + 116 + ], + "type": "image", + "image_path": "7a234fdc6c2f5b0bee2dfb7bdf4195f62a45ce3b57971839a22942e5d119089d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 118, + 537, + 125 + ], + "lines": [ + { + "bbox": [ + 427, + 118, + 537, + 125 + ], + "spans": [ + { + "bbox": [ + 427, + 118, + 537, + 125 + ], + "type": "text", + "content": "c) Lighting visibility under two directional lights" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 133, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 167 + ], + "type": "text", + "content": "Figure 4. We show our computed indirect illumination (b) of the full rendered image (a) and our lighting visibility (c) under two different directional lights." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 176, + 467, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 467, + 187 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 467, + 187 + ], + "type": "text", + "content": "accurate physical model reconstruction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 194, + 452, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 194, + 452, + 208 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 452, + 208 + ], + "type": "text", + "content": "3.3. Illumination and Visibility" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 213, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 332 + ], + "type": "text", + "content": "Our TensoRF-based scene representation is highly efficient for optimization and evaluation. Especially, our volume density can be computed by simple tensor interpolation without using any MLPs, leading to highly efficient computation of volume transmittance and rendering weights (Eqn. 1,2). This allows us to compute accurate incident illumination " + }, + { + "bbox": [ + 304, + 213, + 545, + 332 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 213, + 545, + 332 + ], + "type": "text", + "content": " (as defined in Eqn. 3) that accounts for secondary shading effects with ray marching as shown in Fig. 4. In particular, the incident illumination is computed by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 336, + 341, + 545, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 341, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 336, + 341, + 545, + 355 + ], + "type": "interline_equation", + "content": "L _ {i} (\\hat {\\mathbf {x}}, \\omega_ {i}) = V (\\hat {\\mathbf {x}}, \\omega_ {i}) L _ {\\mathrm {d}} (\\omega_ {i}) + L _ {\\mathrm {i n d}} (\\hat {\\mathbf {x}}, \\omega_ {i}) \\tag {7}", + "image_path": "6797356ac8e873ee9f96586e8b455ea6687496679b8a3ad586cf3d3e8d29d0f4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": " is the light visibility function, " + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{d}}" + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": " is the direct illumination, and " + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ind}}" + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": " is the indirect illumination." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 394, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 525 + ], + "type": "text", + "content": "Direct illumination. We assume the unknown natural environment to be distant from the captured object and represent the global illumination as an environment map, parameterized by a mixture of spherical Gaussians (SG), which represents the direct illumination. Note that in contrast to previous methods [39,42] that use SG for computing the integral of the rendering equation with a closed-form approximation, we use SG only for its compact parameterized representation and compute the integral numerically by sampling secondary rays and performing ray marching, leading to more accurate lighting visibility and indirect lighting." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": "Visibility and indirect illumination. We make use of the jointly-trained radiance fields to model secondary shading effects. More specifically, the indirect illumination arriving at " + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": " is inherently explained by the radiance color " + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{RF}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": " along the ray " + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i(t) = \\hat{\\mathbf{x}} +t\\omega_i" + }, + { + "bbox": [ + 304, + 533, + 545, + 628 + ], + "type": "text", + "content": ". The visibility function is exactly modeled by the transmittance function in volume rendering. Therefore, the light visibility and indirect illumination term in Eqn. 7 can be calculated as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 637, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 637, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 315, + 637, + 545, + 651 + ], + "type": "interline_equation", + "content": "V (\\hat {\\mathbf {x}}, \\omega_ {i}) = T (\\hat {\\mathbf {x}}, \\omega_ {i}), \\quad L _ {\\text {i n d}} (\\hat {\\mathbf {x}}, \\omega_ {i}) = C _ {\\mathrm {R F}} (\\hat {\\mathbf {x}}, \\omega_ {i}) \\tag {8}", + "image_path": "0f0e5fb24d7f5580c2581d7ffadd0bc532bd08b51c05a70fbccdacc008092a03.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "inline_equation", + "content": "T(\\hat{\\mathbf{x}}, \\boldsymbol{\\omega}_i)" + }, + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "text", + "content": " represents the volume transmittance of the final point sampled on the ray " + }, + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i(t) = \\hat{\\mathbf{x}} + t\\boldsymbol{\\omega}_i" + }, + { + "bbox": [ + 305, + 658, + 545, + 683 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Second-bounce ray marching. To make our physically-based rendering " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{PB}}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " accurate, we compute the rendering" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "169" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "integral (Eqn. 3) via Monte Carlo integration by marching multiple rays from each surface point " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " with stratified sampling when training. For each ray, we obtain direct illumination from the SGs and compute visibility " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "T(\\hat{\\mathbf{x}},\\omega_{i})" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " and indirect illumination " + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{RF}}(\\hat{\\mathbf{x}},\\omega_{i})" + }, + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": " using Eqn. 1 directly. This second-bounce ray marching is known to be expensive for previous NeRF-based methods, which either requires extremely high computational resources (128 TPUs) [28] or utilizes long (several days of) offline pre-computation [41]; both require training extra MLPs, which takes excessive computational costs and is unable to achieve high accuracy. We instead perform second-bounce ray marching online, achieving higher accuracy (see Fig. 6), which is affordable, thanks to our highly efficient tensor factorization-based representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 259, + 200, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 200, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 200, + 274 + ], + "type": "text", + "content": "3.4. Multi-Light Representation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "text", + "content": "We have discussed our scene representation under a single unknown lighting condition in Sec. 3.2. Our method can be easily extended to support capturing under multiple unknown lighting conditions, owing to the generality and extensibility of tensor factorization. We achieve this by adding an additional lighting dimension with vector factors " + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "text", + "content": " - where the length of each vector " + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "text", + "content": " equals the number of lighting conditions - into the factorized appearance tensor " + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_a" + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "text", + "content": ", leading to a 5D tensor that expresses scene appearance under different lighting conditions. We denote the 5D multi-light appearance tensor as " + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_a^{5\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 278, + 289, + 422 + ], + "type": "text", + "content": ", represented by the factorization:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 430, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 430, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 78, + 430, + 287, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {a} ^ {\\mathrm {5 D}} = \\sum_ {k} \\sum_ {m \\in X Y Z} \\mathbf {v} _ {a, k} ^ {m} \\circ \\mathbf {M} _ {a, k} ^ {\\tilde {m}} \\circ \\mathbf {e} _ {k} ^ {m} \\circ \\mathbf {b} _ {k} ^ {m} \\tag {9}", + "image_path": "ff8f032bc0e15dd7053f8eb23a17833eecb8261890b58eba31df4b907ed6c2ef.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 464, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 561 + ], + "type": "text", + "content": "Note that, it is only the view-dependent colors that require being modeled separately under different lighting, while the physical scene properties - including volume density, normals, and material properties - are inherently shared across multiple lighting conditions. Therefore, we decode view-dependent color using the neural features per light and decode other shading properties using the mean features along the lighting dimension:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 569, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 569, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 96, + 569, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\bar {a} _ {\\mathbf {x}} = \\frac {\\sum_ {l} a _ {\\mathbf {x} , l}}{P}, \\quad a _ {\\mathbf {x}, l} = \\mathcal {G} _ {a} ^ {5 D} (\\mathbf {x}, l) \\tag {10}", + "image_path": "a34b8ae016d7a25d320d512c93003afa30ab3579f6ac8d52c52e55b04b4e6788.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 596, + 249, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 596, + 249, + 609 + ], + "spans": [ + { + "bbox": [ + 85, + 596, + 249, + 609 + ], + "type": "interline_equation", + "content": "c _ {\\mathbf {x}}, \\mathbf {n} _ {\\mathbf {x}}, \\beta_ {\\mathbf {x}} = \\mathcal {D} _ {c} (a _ {\\mathbf {x}, l}), \\mathcal {D} _ {\\mathbf {n}} (\\bar {a} _ {\\mathbf {x}}), \\mathcal {D} _ {\\beta} (\\bar {a} _ {\\mathbf {x}})", + "image_path": "8f15cfa633010a45f85ffc4996f5d976c8cf30b71ef913b3b41268e65b7d5c6a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "content": " is the light index, " + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "content": " is the number of lighting conditions, and " + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\bar{a}_{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 641 + ], + "type": "text", + "content": " is the mean feature." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "By simply adding additional vectors in the factorized appearance tensor, our method allows us to efficiently reconstruct and query the radiance under different illumination conditions, thereby still allowing for indirect illumination computation in the multi-light setting with the method discussed in Sec. 3.3. The multi-light input can provide useful" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "photometric cues and reduce ambiguity in material prediction, and therefore leads to more accurate reconstruction of geometry and material estimation (see Tab. 1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 117, + 524, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 117, + 524, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 117, + 524, + 131 + ], + "type": "text", + "content": "3.5. Joint Reconstruction and Training Losses" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 136, + 546, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 546, + 184 + ], + "type": "text", + "content": "For both single- and multi-light settings, we jointly reconstruct the scene geometry and appearance properties that are modeled by our tensor factorization-based representation, through an end-to-end per-scene optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "text", + "content": "Rendering losses. We supervise the rendered colors from both the radiance field rendering " + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{RF}}" + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "text", + "content": " and the physically-based rendering " + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{PB}}" + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "text", + "content": " with the ground-truth colors " + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 191, + 546, + 240 + ], + "type": "text", + "content": " from the captured images and include two rendering loss terms:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 249, + 546, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 249, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 326, + 249, + 546, + 264 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {R F}} = \\left\\| C _ {\\mathrm {R F}} - C _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2}, \\quad \\ell_ {\\mathrm {P B}} = \\left\\| C _ {\\mathrm {P B}} - C _ {\\mathrm {g t}} \\right\\| _ {2} ^ {2} \\tag {11}", + "image_path": "03bb68bd56a1d379cd3b838c7c35327320dc1be53c7c6691066d9f2ab78c0744.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "type": "text", + "content": "Normal regularization. Our shading normals " + }, + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "type": "text", + "content": " are regressed from an MLP to express high-frequency surface variations and used in physically-based rendering. However, since the entire system is highly ill-conditioned, only supervising the network output with rendering loss is prone to overfitting issues and produces incorrect normal predictions. On the other hand, many previous NeRF-based method [6, 26, 28, 41] use the negative direction of volume density gradients " + }, + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\sigma} = -\\frac{\\nabla_{\\mathbf{x}}\\sigma}{\\|\\nabla_{\\mathbf{x}}\\sigma\\|}" + }, + { + "bbox": [ + 304, + 273, + 547, + 430 + ], + "type": "text", + "content": " as normals for shading computation. Such derived normals are better aligned with the surface but can be noisy and lack fine details. Inspired by the rent work Ref-NeRF [31], we correlate our shading normal with the density-derived normal using a loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 371, + 439, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 439, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 371, + 439, + 545, + 465 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathbf {n}} = \\sum_ {j} w _ {j} \\| \\mathbf {n} _ {j} - \\mathbf {n} _ {\\sigma , j} \\| _ {2} ^ {2} \\tag {12}", + "image_path": "f4b592aa09177572dfa01bc29dcd357673259541c61757a391913ee8b2bf5c7d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 467, + 546, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 546, + 586 + ], + "type": "text", + "content": "This regularization term back-propagates gradients to both the appearance tensor (through the normal network) and the density tensor, further correlating our entire scene geometry and appearance reasoning, leading to better normal reconstruction. Note that while Ref-NeRF uses the same normal regularization, their normals are used in MLP-based shading computation; in contrast, our normals are used in a physically-based rendering process, allowing for more meaningful geometry reasoning and leading to more accurate normal reconstruction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Final loss. Following TensoRF, we apply additional " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "-regularization on all tensor factors, denoted as " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathcal{G}}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ". We also apply a smoothness regularization term on BRDF parameters " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " to enhance their spatial consistency. A loss term penalizing back-facing normals is used in addition. Please refer to the supplementary materials for the details of these regularization terms. We apply all these losses to supervise and regularize our whole system to jointly reconstruct the scene, optimizing our scene representation (with all tensor factors and MLPs), as well as the SG parameters of the en" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "170" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 52, + 544, + 281 + ], + "blocks": [ + { + "bbox": [ + 52, + 89, + 64, + 125 + ], + "lines": [ + { + "bbox": [ + 52, + 89, + 64, + 125 + ], + "spans": [ + { + "bbox": [ + 52, + 89, + 64, + 125 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 174, + 64, + 207 + ], + "lines": [ + { + "bbox": [ + 53, + 174, + 64, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 64, + 207 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 1, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 239, + 64, + 281 + ], + "lines": [ + { + "bbox": [ + 53, + 239, + 64, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 64, + 281 + ], + "type": "text", + "content": "Env. Map" + } + ] + } + ], + "index": 2, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 52, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 77, + 52, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 77, + 52, + 544, + 281 + ], + "type": "image", + "image_path": "dfc68463364d3421dcd711bef49ab7f671c19740f5bbe032c1a88adffc1f9a9a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 285, + 546, + 306 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 546, + 306 + ], + "type": "text", + "content": "Figure 5. Visual comparison against baseline methods. Our method produces inverse rendering results of higher quality with more detailed normals and more accurate albedo, thus leading to more photo-realistic relighting results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 307, + 547, + 380 + ], + "blocks": [ + { + "bbox": [ + 49, + 307, + 547, + 380 + ], + "lines": [ + { + "bbox": [ + 49, + 307, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 307, + 547, + 380 + ], + "type": "table", + "html": "
MethodNormal MAE ↓AlbedoNovel View SynthesisRelightingAverage Runtime
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRFactor6.31425.1250.9400.10924.6790.9220.12023.3830.9080.131>100 hrs
InvRender5.07427.3410.9330.10027.3670.9340.08923.9730.9010.10115 hrs
Ours in 25 minutes4.87628.2100.9470.09132.3500.9640.06127.4310.9350.09425 mins
Ours4.10029.2750.9500.08535.0880.9760.04028.5800.9440.0815 hrs
Ours w/ rotated multi-light3.60229.6720.9560.07934.3950.9740.04328.9550.9490.0775 hrs
Ours w/ general multi-light3.55129.3260.9510.08434.2230.9730.04529.0080.9470.0785 hrs
", + "image_path": "6be5eb91f094618ede79603ee7beebb66c237f95bd46630debcb3af68d7fa566.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 383, + 545, + 438 + ], + "lines": [ + { + "bbox": [ + 46, + 383, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 545, + 438 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the synthetic dataset. Our (single-light) results have significantly outperformed the baseline methods by producing more accurate normal and albedo, thus achieving more realistic novel view synthesis and relighting results. Our method can further take images captured under different lighting conditions, and boost the performance in inverse rendering. (We scale each RGB channel of all albedo results by a global scalar, as done in NeRFactor [41]. For a fair comparison, all novel view synthesis results are generated with physically-based rendering, though our radiance field rendering has better quality.)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 442, + 179, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 179, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 179, + 453 + ], + "type": "text", + "content": "vironment map, with a final loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 464, + 287, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 464, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 52, + 464, + 287, + 477 + ], + "type": "interline_equation", + "content": "\\ell = \\alpha_ {\\mathrm {R F}} \\ell_ {\\mathrm {R F}} + \\alpha_ {\\mathrm {P B}} \\ell_ {\\mathrm {P B}} + \\alpha_ {\\beta} \\ell_ {\\beta} + \\alpha_ {\\mathbf {n}} \\ell_ {\\mathbf {n}} + \\alpha_ {\\mathbf {d}} \\ell_ {\\mathbf {d}} + \\alpha_ {\\mathcal {G}} \\ell_ {\\mathcal {G}} \\tag {13}", + "image_path": "cf2d68c85c310a187d17a94d263bdab79882acff87c08e3a88d925bdf6e1e078.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 487, + 128, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 128, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 128, + 501 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "type": "text", + "content": "We now evaluate our model on various challenging scenes. We make comparisons against previous state-of-the-art methods and also present an ablation study to verify the effectiveness of our design choices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 562, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 694 + ], + "type": "text", + "content": "Datasets. We perform experiments on four complex synthetic scenes, including three blender scenes from [22] and one from the Stanford 3D scanning repository [12]. We rerender these scenes to obtain their ground-truth images, as well as BRDF parameters and normal maps. We also render two types of multi-light data: rotated multi-light and general multi-light. We also perform experiments on the original NeRF-synthetic dataset and 4 captured real data. Please refer to the supplementary for experimental results of those extra data, more details about our datasets, and more explanation about our multi-light setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Comparisons with previous methods. We compare our" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 442, + 545, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 545, + 705 + ], + "type": "text", + "content": "model with previous state-of-the-art neural field-based inverse rendering methods, NeRFactor [41] and InvRender [42], on these scenes using images captured under a single unknown lighting condition. We also compare with our model trained under multi-light settings. Table 1 shows the accuracy of the estimated albedo, normal, and relighting results using metrics including PSNR, SSIM and LPIPS [40] averaged over the four scenes. We can see that our single-light method (using the same input as baselines) outperforms both previous methods significantly on all metrics, demonstrating the superiority of our method. We also include a visual comparison in Fig. 5, showing that our method predicts more accurate albedo and normal that are closer to the ground truth compared to baseline methods, thus generating more realistic relighting results. In particular, our results produce normals that are of much higher quality and more faithfully reflect the geometry variations, while the baseline methods produce over-smooth results that lack shape details. Although NeRFactor's albedo result on the lego scene looks closer to the ground truth than our results. We claim that this is because that NeRFactor uses a high-weight BRDF smoothness loss, which dam" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "171" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 67, + 111, + 125 + ], + "blocks": [ + { + "bbox": [ + 51, + 67, + 111, + 125 + ], + "lines": [ + { + "bbox": [ + 51, + 67, + 111, + 125 + ], + "spans": [ + { + "bbox": [ + 51, + 67, + 111, + 125 + ], + "type": "image", + "image_path": "9bc328725e2aab2b7aed3a16410bf380357916d8d0d3f8c4a17b93d10a6a4a07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 127, + 91, + 133 + ], + "lines": [ + { + "bbox": [ + 69, + 127, + 91, + 133 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 91, + 133 + ], + "type": "text", + "content": "NeRFactor" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 111, + 66, + 168, + 125 + ], + "blocks": [ + { + "bbox": [ + 111, + 66, + 168, + 125 + ], + "lines": [ + { + "bbox": [ + 111, + 66, + 168, + 125 + ], + "spans": [ + { + "bbox": [ + 111, + 66, + 168, + 125 + ], + "type": "image", + "image_path": "cc550f7267b1e4434e90f6597acdd028c039d528480ec21c86d3d66e5143dcbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 127, + 151, + 133 + ], + "lines": [ + { + "bbox": [ + 129, + 127, + 151, + 133 + ], + "spans": [ + { + "bbox": [ + 129, + 127, + 151, + 133 + ], + "type": "text", + "content": "InvRender" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 66, + 227, + 125 + ], + "blocks": [ + { + "bbox": [ + 170, + 66, + 227, + 125 + ], + "lines": [ + { + "bbox": [ + 170, + 66, + 227, + 125 + ], + "spans": [ + { + "bbox": [ + 170, + 66, + 227, + 125 + ], + "type": "image", + "image_path": "0a2ee602f1c2fa211fc064bdf18a265f9be2a078de1e3e0578eb9710ec7ca18b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 127, + 201, + 133 + ], + "lines": [ + { + "bbox": [ + 190, + 127, + 201, + 133 + ], + "spans": [ + { + "bbox": [ + 190, + 127, + 201, + 133 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 135, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 135, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 135, + 286, + 167 + ], + "type": "text", + "content": "Figure 6. We compare our relighting results with previous methods. Note that our approach recovers more accurate shadows thanks to our second-bounce ray marching." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 66, + 285, + 125 + ], + "blocks": [ + { + "bbox": [ + 228, + 66, + 285, + 125 + ], + "lines": [ + { + "bbox": [ + 228, + 66, + 285, + 125 + ], + "spans": [ + { + "bbox": [ + 228, + 66, + 285, + 125 + ], + "type": "image", + "image_path": "c8665f30e03f7319af2a70d67d09193e38cf8aaf645b3db29315e95c0537b6d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 127, + 268, + 132 + ], + "lines": [ + { + "bbox": [ + 240, + 127, + 268, + 132 + ], + "spans": [ + { + "bbox": [ + 240, + 127, + 268, + 132 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 175, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 286, + 354 + ], + "type": "text", + "content": "ages its reconstruction quality of other components. In the supplementary, we showcase that we can achieve a similar albedo result by increasing the BRDF smoothness weight. In particular, NeRFactor approximates the visibility function in a distilled MLP from a pre-trained NeRF and completely ignores the indirect illumination. While InvRender considers both visibility and indirect illumination in their pre-computation, its spherical Gaussian-based shading computation can only achieve limited accuracy. Both methods do inverse rendering in a second stage after pretraining NeRFs. In contrast, our tensor-factorized representation achieves a single-stage joint reconstruction and efficiently performs explicit second-bounce ray marching for more accurate shadowing (see Fig. 6) and indirect lighting, thus leading to significantly higher reconstruction quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 357, + 286, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 286, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 286, + 607 + ], + "type": "text", + "content": "Meanwhile, owing to our more efficient representation, our high reconstruction quality is achieved with the fastest reconstruction speed, as reflected by the reconstruction time. NeRF-based method NeRFactor takes days to compute because of visibility pre-computation. InvRender is faster than NeRFactor, due to its SDF-based rendering [36] and SG-based closed-form rendering integral computation; however, its SDF-based reconstruction fails on challenging scenes like the Lego shown in Fig. 5 and the SG-based integration leads to inaccurate secondary shading effects as mentioned. On the other hand, our approach leverages ray marching-based radiance rendering in the reconstruction, robustly producing high-quality results on all testing scenes with accurate secondary effects, while still being faster than InvRender. In fact, while our method takes 5 hours to finish its full training for its best performance, it can achieve good quality in a much shorter training period (25 minutes). As shown in Tab. 1 and Fig. 7, our approach can achieve high-quality geometry reconstruction in only 25 minutes and outperform previous methods that trained for tens of or even hundreds of hours." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Multi-light results. In addition, our framework can also effectively leverage the additional input in a multi-light capture and further boost the accuracy of the inverse rendering performance without adding additional computation costs, as shown in Tab. 1, while the MLP-based baseline methods cannot be trivially extended to support such setups in an efficient manner. We also found that multi-light settings can greatly improve geometry reconstruction and help solving" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 544, + 107 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 544, + 107 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 544, + 107 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 544, + 107 + ], + "type": "table", + "html": "
MethodNormal MAE ↓Albedo PSNR ↑NVS PSNR ↑
w/o visibility4.71627.26534.703
w/o indirect illum.4.18629.00334.910
w/ indirect illum. + visibility4.10029.27535.088
", + "image_path": "55283105783f38426984a3011641ce36a392fc18afd680ca32c3b461df551e14.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 108, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 306, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 545, + 140 + ], + "type": "text", + "content": "Table 2. Our full method models more physically-accurate light transport by accounting for lighting visibility and indirect illuminations, thus achieving much better accuracy in inverse rendering." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 145, + 372, + 218 + ], + "blocks": [ + { + "bbox": [ + 310, + 145, + 372, + 218 + ], + "lines": [ + { + "bbox": [ + 310, + 145, + 372, + 218 + ], + "spans": [ + { + "bbox": [ + 310, + 145, + 372, + 218 + ], + "type": "image", + "image_path": "c92480335c778171f47c7a47d0171ade181590b3bd119bb0d8d622253708edeb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 220, + 375, + 226 + ], + "lines": [ + { + "bbox": [ + 320, + 220, + 375, + 226 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 375, + 226 + ], + "type": "text", + "content": "NeRFactor " + }, + { + "bbox": [ + 320, + 220, + 375, + 226 + ], + "type": "inline_equation", + "content": "(>100\\mathrm{hrs})" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 228, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 272 + ], + "type": "text", + "content": "Figure 7. We compare geometry reconstruction results of our model taking only 25 minutes of optimization, with previous methods, taking 15 and " + }, + { + "bbox": [ + 306, + 228, + 545, + 272 + ], + "type": "inline_equation", + "content": ">100" + }, + { + "bbox": [ + 306, + 228, + 545, + 272 + ], + "type": "text", + "content": " hours. Our approach even recovers more high-frequency details with substantially less reconstruction time." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 372, + 145, + 427, + 217 + ], + "blocks": [ + { + "bbox": [ + 372, + 145, + 427, + 217 + ], + "lines": [ + { + "bbox": [ + 372, + 145, + 427, + 217 + ], + "spans": [ + { + "bbox": [ + 372, + 145, + 427, + 217 + ], + "type": "image", + "image_path": "ebe63a888fc657481ca6c8dc52ba52f29708978f13ac7c72e9ed3d2091567978.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 220, + 433, + 226 + ], + "lines": [ + { + "bbox": [ + 381, + 220, + 433, + 226 + ], + "spans": [ + { + "bbox": [ + 381, + 220, + 433, + 226 + ], + "type": "text", + "content": "InvRender (15 hrs)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 145, + 485, + 217 + ], + "blocks": [ + { + "bbox": [ + 427, + 145, + 485, + 217 + ], + "lines": [ + { + "bbox": [ + 427, + 145, + 485, + 217 + ], + "spans": [ + { + "bbox": [ + 427, + 145, + 485, + 217 + ], + "type": "image", + "image_path": "a2260a2848ccbb38377b69d8767ca1b15a63cf24f3555ed7765c8485b5832dbc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 220, + 484, + 226 + ], + "lines": [ + { + "bbox": [ + 441, + 220, + 484, + 226 + ], + "spans": [ + { + "bbox": [ + 441, + 220, + 484, + 226 + ], + "type": "text", + "content": "Ours (25 mins)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 485, + 145, + 543, + 217 + ], + "blocks": [ + { + "bbox": [ + 485, + 145, + 543, + 217 + ], + "lines": [ + { + "bbox": [ + 485, + 145, + 543, + 217 + ], + "spans": [ + { + "bbox": [ + 485, + 145, + 543, + 217 + ], + "type": "image", + "image_path": "10eddaf8a4e80fde80c8911f775ad181be81d749ed2d4ef789c6b24c17fc3898.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 220, + 537, + 226 + ], + "lines": [ + { + "bbox": [ + 499, + 220, + 537, + 226 + ], + "spans": [ + { + "bbox": [ + 499, + 220, + 537, + 226 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 281, + 544, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 544, + 305 + ], + "type": "text", + "content": "the color ambiguity between lighting and materials. Please refer the supplementary for more results and analysis." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 315, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 482 + ], + "type": "text", + "content": "Indirect illumination and visibility. Our tensor-factorized representation allows us to explicit sample secondary rays in an efficient way to account for computing lighting visibility and indirect illumination. As shown in Tab. 2, without including these terms, the model cannot accurately represent the secondary shading effects and tend to bake them in the albedo or normal, resulting in lower quality. Nonetheless, note that these ablated models of our method in fact already achieves superior quality compared to previous methods (that pre-computes NeRFs and do inverse rendering disjointly) shown in Tab. 1, demonstrating the effectiveness of our joint reconstruction framework. Our full method further achieves more accurate reconstruction because of more accurate light transport modeling." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 499, + 378, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 499, + 378, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 499, + 378, + 511 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 522, + 545, + 713 + ], + "type": "text", + "content": "We present a novel inverse rendering approach that achieves efficient and high-quality scene reconstruction from multi-view images under unknown illumination. Our approach models a scene as both a radiance field and a physically-based model with density, normals, lighting, and material properties. By jointly reconstructing both models, we achieve high-quality geometry and reflectance reconstruction, enabling photo-realistic rendering and relighting. Owing to the efficiency and generality of the tensor factorized representation, our framework allows for accurate computation for shadowing and indirect lighting effects, and also flexibly supports capturing under an arbitrary number of lighting conditions. We demonstrate that our approach is able to achieve state-of-the-art inverse rendering results, outperforming previous neural methods in terms of both reconstruction quality and efficiency." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "172" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 234 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 234 + ], + "type": "text", + "content": "[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 300 + ], + "type": "text", + "content": "[4] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In European Conference on Computer Vision, pages 294–311. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 356 + ], + "type": "text", + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5960-5969, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 412 + ], + "type": "text", + "content": "[6] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 467 + ], + "type": "text", + "content": "[7] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 288, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 501 + ], + "type": "text", + "content": "[8] Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. In ACM SIGGRAPH, volume 2012, pages 1-7. vol. 2012, 2012. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 288, + 555 + ], + "type": "text", + "content": "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[10] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[11] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 1, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Paul Debevec, Tim Hawkins, Chris Tchou, Haarm-Pieter Duiker, Westley Sarokin, and Mark Sagar. Acquiring the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "reflectance field of a human face. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 145-156, 2000. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[14] Yue Dong, Guojun Chen, Pieter Peers, Jiawan Zhang, and Xin Tong. Appearance-from-motion: Recovering spatially varying surface reflectance under unknown lighting. ACM Transactions on Graphics, 33(6):193, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "text", + "content": "[15] Dan B Goldman, Brian Curless, Aaron Hertzmann, and Steven M Seitz. Shape and spatially-varying brdfs from photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(6):1060-1071, 2009. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "text", + "content": "[16] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(3):548-554, 2008. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[17] Jason Lawrence, Szymon Rusinkiewicz, and Ravi Ramamoorthi. Efficient brdf importance sampling using a factored representation. ACM Transactions on Graphics (ToG), 23(3):496-505, 2004. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 545, + 316 + ], + "type": "text", + "content": "[18] Hendrik PA Lensch, Jochen Lang, Asla M Sá, and Hans-Peter Seidel. Planned sampling of spatially varying brdfs. In Computer graphics forum, volume 22, pages 473-482. Wiley Online Library, 2003. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "text", + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 371, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 424 + ], + "type": "text", + "content": "[20] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. In SIGGRAPH Asia 2018, page 269. ACM, 2018. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "text", + "content": "[24] Giljoo Nam, Joo Ho Lee, Diego Gutierrez, and Min H Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. In SIGGRAPH Asia 2018, page 267. ACM, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[26] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision (ECCV), 2022. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "173" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 182 + ], + "type": "text", + "content": "[28] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 215 + ], + "type": "text", + "content": "[29] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "text", + "content": "[30] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[31] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "text", + "content": "[32] Rui Xia, Yue Dong, Pieter Peers, and Xin Tong. Recovering shape and spatially-varying surface reflectance under unknown illumination. ACM Transactions on Graphics, 35(6):187, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[33] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438–5448, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 447 + ], + "type": "text", + "content": "[34] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics, 38(4):76, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[35] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics, 37(4):126, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 494, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 494, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 494, + 287, + 546 + ], + "type": "text", + "content": "[36] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "text", + "content": "[37] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. Arf: Artistic radiance fields. In European Conference on Computer Vision, pages 717-733. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "type": "text", + "content": "[38] Kai Zhang, Fujun Luan, Zhengqi Li, and Noah Snavely. Iron: Inverse rendering by optimizing neural sdfs and materials from photometric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5565-5574, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[39] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2, 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 239 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[41] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 239 + ], + "type": "text", + "content": "[42] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18643-18652, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "174" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_content_list.json b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b10d0adffaaa75321e6ef35bd90c2cc2565f5c7 --- /dev/null +++ b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Tensor4D : Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering", + "text_level": 1, + "bbox": [ + 119, + 128, + 851, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruizhi Shao $^{1}$ , Zerong Zheng $^{1,2}$ , Hanzhang Tu $^{1}$ , Boning Liu $^{1}$ , Hongwen Zhang $^{1}$ , Yebin Liu $^{1}$ $^{1}$ Department of Automation, Tsinghua University $^{2}$ NNKosmos Technology", + "bbox": [ + 122, + 200, + 843, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e25a3aca3dd57916fda3911e09fe6f70a32d45c183c4ac8866fcb0881f82aeba.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 84, + 280, + 228, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92b7b4a3a63a35de1019598d4318a137121a9c585bd60fabda35cfb0ad77b65d.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 230, + 280, + 362, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c86ca73fc314a3ea32b333eaadcea2be0763d818d6c12218673aab1a3db2e52c.jpg", + "image_caption": [ + "(c)", + "Figure 1. Given 4 sparse static RGB camera views of a dynamic scene (a), our proposed Tensor4D decomposition enables multiview reconstruction to achieve fine-grained geometry reconstruction even on human fingers (b) and temporal-consistent novel view synthesis on a 3D holographic display (c,d,e). The 4 cameras are settled on four corners of the display. The proposed method demonstrates low-cost, portable and highly immersive telepresence experience." + ], + "image_footnote": [], + "bbox": [ + 364, + 281, + 531, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/953d43cd16790621adae1b3dc19a758e9df275f158279bb9f728d8dfda42ff5f.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 532, + 281, + 702, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f413ef76e9da69b86fe19d50875e1650a112f2de45461c536d419a6f44692ff0.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 704, + 281, + 883, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 479, + 313, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Tensor4D, an efficient yet effective approach to dynamic scene modeling. The key of our solution is an efficient 4D tensor decomposition method so that the dynamic scene can be directly represented as a 4D spatiotemporal tensor. To tackle the accompanying memory issue, we decompose the 4D tensor hierarchically by projecting it first into three time-aware volumes and then nine compact feature planes. In this way, spatial information over time can be simultaneously captured in a compact and memory-efficient manner. When applying Tensor4D for dynamic scene reconstruction and rendering, we further factorize the 4D fields to different scales in the sense that structural motions and dynamic detailed changes can be learned from coarse to fine. The effectiveness of our method is validated on both synthetic and real-world scenes. Extensive experiments show that our method is able to achieve high-quality dynamic reconstruction and rendering from sparse-view camera rigs or even a monocular camera. The code and dataset will be released at https://github.com/DSaurus/Tensor4D.", + "bbox": [ + 73, + 512, + 473, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High quality reconstruction and Photo-realistic rendering of a dynamic scene from a set of input images is nec", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "essary for many applications such as AR/VR, 3D content production and entertainment. Traditional methods use classical mesh-based representation to reconstruct the dynamic scenes, which, unfortunately, are prone to produce reconstruction errors and rendering artifacts when the scene contains thin structures, specular surfaces and topological changes [9, 19, 23, 26, 49].", + "bbox": [ + 496, + 479, + 892, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in neural rendering approaches, which learn scene representations in the form of neural radiance fields (NeRF), have shown impressive novel view synthesis of general static scenes given only multi-view images [32]. They are immediately extended to dynamic scenes: some methods (e.g., NeRF-T) consider time as an additional input dimension to NeRF representation [56,62], while other methods (e.g., D-NeRF) disentangle a dynamic scene into a canonical radiance field and a dynamic motion field [11,28,37,40,55]. Either way, learning a 4D function is one of the main cornerstones. Unfortunately, directly using MLP to fit such a function often suffers from high time and computation cost, i.e., dozens of hours on high-end GPUs.", + "bbox": [ + 496, + 585, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In fact, the aforementioned limitation also exists in conventional NeRF-based methods for static scenes, and researchers have proposed to use discrete data structures like voxel grids [65] or triplanes [7] to accelerate NeRF training and rendering. However, these techniques are difficult to be extended to dynamic domains as introducing an additional time dimension will exponentially increase memory footprint, hindering them from modeling high-quality ap", + "bbox": [ + 496, + 781, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16632", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pearance details.", + "bbox": [ + 78, + 90, + 189, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we pursue a dynamic scene representation that also utilizes explicit feature grids to accelerate network training while avoiding huge memory consumption when introducing an additional time dimension. To this end, we bypass the construction of a high resolution 4D tensor; instead, we propose to model a 4D field using hierarchical tri-projection decomposition. Our decomposition method extends the tri-projection in EG3D [7]. It firstly project a full 4D field into three time-aware volumes, each of which is then further decomposed into three feature planes. In this way, we model the 4D field using only nine 2D feature planes, and we empirically find that although being highly compact, such a representation is powerful enough to represent dynamic scenes containing complex motions. Moreover, the usage of explicit data structure also allows us to design a coarse-to-fine strategy to further improve the performance of our method.", + "bbox": [ + 80, + 107, + 467, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By utilizing and factorizing an explicit 4D tensor, our method enables both efficient reconstruction and compact representation of dynamic scenes. Besides, the decomposition scheme also introduces implicit constraints on the representation since only low-rank tensors can be approximated by a small number of lower-dimensional components. Such a constraint can serve as an inherent regularization when the input observation is limited, e.g., under sparse and fixed cameras setting or even monocular inputs. In this paper, we first apply our method for sparse-view dynamic reconstruction by adopting our Tensor4D decomposition to time-conditioned radiance fields in \"NeRF-T\". In addition, our decomposition method can also be used for single-view dynamic reconstruction. This is achieved through decomposing both the 4D dynamic motion field and the canonical radiance field in \"D-NeRF\". With proper regularization, our system enables efficient and high-quality reconstruction of dynamic objects under both camera settings.", + "bbox": [ + 80, + 363, + 467, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 648, + 215, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiview Reconstruction and Rendering. There are various ways to capture and reconstruct the 3D dynamic scenes, involving methods based on silhouette [24, 54], stereo [30, 46], flow [17, 44, 70], segmentation [41, 42], and photometric [3, 20]. With RGBD cameras, real-time solutions like DynamicFusion [34] estimates the non-rigid deformations of a dynamic scene and integrates depth frames to reconstruct the geometry model in the canonical space. This method is later extended for telepresence and holographic communication [10, 25, 66]. However, these systems heavily rely on depth sensors to obtain accurate geometry. In contrast, our method can reconstruct and render dynamic scenes using sparse-view RGB cameras.", + "bbox": [ + 80, + 674, + 467, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the past few years, neural implicit representations underwent rapid development and have been applied for multi-", + "bbox": [ + 80, + 869, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "view reconstruction and rendering of static scenes. Some methods represent the geometry as the zero level-set of a neural network, and use differentiable surface rendering to optimize the network weights [35,47,64]. Given dense observation of an object, these methods are able to accurately recover its surface. NeRF [32], on the other hand, uses volume rendering to optimize the scene representation. Its simplicity and impressive results inspires a lot of following works, including in-the-wild reconstruction [31, 52], lighting and material estimation [4,48,67], generation [1,6,18, 22,43,53], human rendering [45,51,57,68,69] and so on. More recently, several methods unify surface and volumetric rendering, enabling accurate geometry reconstruction and high-quality novel view synthesis [36,59,63]. Compared to these static scene representations, we aim to enable free view synthesis of dynamic scenes from extremely sparse cameras.", + "bbox": [ + 503, + 90, + 888, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF for Dynamic Scenes. Modeling scenes in 4D domain with time dimension included is a direct solution to extend NeRF for dynamic domains. Typical approaches include VideoNeRF [62], NeRFlow [11], DyNeRF [27], and DCT-NeRF [56]. Specifically, VideoNeRF [62] directly learns a spatiotemporal irradiance field from a single video and uses depth estimation to address the shape-motion ambiguities in monocular inputs, while NeRFlow [11] and DCT-NeRF [56] use point trajectory to regularize the network optimization. To deal with the limitation of topology-change modeling in deformation field, Park et al. [38] represent HyperNeRF which can lift NeRF to higher dimensions.", + "bbox": [ + 503, + 345, + 888, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic scenes can also be rendered by deforming the radiance field in the canonical space. For example, Nerfies [37] optimizes an additional continuous deformation field by warping each observed point into a canonical 5D NeRF. D-Nerf [40] and NR-Nerf [55] follow a similar framework, but take only monocular videos as training data. In addition, DeVRF [28] uses voxel-based representation instead of MLPs to model the 3D canonical space and the 4D deformation field. Using parametric body templates as the semantic prior, methods like Neural Body [39] and HumanNeRF [60] enable photo-realistic novel view synthesis of complex human performance. The discrepancy between the practical capture process and the existing experimental protocols for monocular videos has been shown in [16].", + "bbox": [ + 503, + 527, + 888, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF Acceleration. Numerous works emerge with the purpose of speeding up static NeRF using explicit data structures including feature maps, voxels and tensors. For instance, DVGO [50] achieves fast convergence through an explicit representation of a density voxel grid and a feature voxel grid. With the sparse voxel octree structure, NSVF [29] accelerates the novel view reconstruction by discarding the empty voxels in a coarse-to-fine manner. Similarly, Plenoxels [14] and PlenOctree [65] model a scene through a hierarchical 3D grid with spherical harmonics, which can realize an optimization with two orders of magnitude faster", + "bbox": [ + 503, + 736, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16633", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9901884b7286d67e7be05168283f994b5080c4338c3bbd4d85da1ed82e944656.jpg", + "image_caption": [ + "Figure 2. Illustration of our hierarchical tri-projection decomposition method. For a neural 4D field $f(x,y,z,t)$ , we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumes, which are then further projected onto nine 2D planes." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 460, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "than NeRF. In DIVeR [61], ray marching only finds a fixed number of hits on the voxel grid to accelerate volumetric rendering. Moreover, hashing encoding [33] and tensor decomposition [8] are also used as compact representations for NeRF acceleration.", + "bbox": [ + 75, + 388, + 468, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For dynamic scene modeling, DeVRF [28] enables fast non-rigid neural rendering with both 3D volumetric and 4D voxel field. In addition, V4D [15] introduce an effective conditional positional encoding for 4D data to realize fast novel view synthesis. TiNeuVox [12] represents scenes with identifiable explicit data structures and accelerates radiance fields modeling, while Wang et al. extended PlenOctrees [65] into free-view video rendering [58]. However, the low-res volumetric design of these works hinders the capacity of rendering high-quality images. In addition, there are also several concurrent works [5, 13, 21] adopt 6-plane decomposition for dynamic scenes. Compared with these methods, our hierarchical decomposition is more efficient to capture time variations and represent dynamic scenes.", + "bbox": [ + 75, + 465, + 470, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 690, + 168, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on prior work for spatio-temporal representations and tri-projection decomposition (Sec. 3.1), we propose a hierarchical tri-projection decomposition method (Sec. 3.2) and a coarse-to-fine strategy (Sec. 3.3), which allow us to learn a 4D field at a modest cost of training time and GPU memory.", + "bbox": [ + 75, + 715, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 816, + 207, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatio-temporal 4D NeRF representation. To represent dynamic objects using neural radiance fields, a naive way is to condition the original neural radiance fields on the times-stamp [56, 62], which we term as NeRF-T. Mathematically,", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF-T can be formulated as:", + "bbox": [ + 500, + 90, + 702, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f (x, y, z, t) = (\\mathbf {f}, \\sigma), \\tag {1} \\\\ g (\\mathbf {f}, \\theta , \\phi) = \\mathbf {c}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 116, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ is a dynamic implicit field that produces a high-dimensional feature $\\mathbf{f}$ and a density value $\\sigma$ for a point at position $(x, y, z)$ and time instance $t$ , and $g$ is a function that takes into account the viewing direction $(\\theta, \\phi)$ to predict the final RGB color.", + "bbox": [ + 496, + 162, + 890, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve better disentanglement of shape and motion, some methods like D-NeRF [40] propose deformable neural radiance field which adopts a canonical 3D representation with the 4D flow fields:", + "bbox": [ + 496, + 238, + 890, + 297 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f (x, y, z, t) = (\\hat {x}, \\hat {y}, \\hat {z}), \\tag {2} \\\\ g (\\hat {x}, \\hat {y}, \\hat {z}, \\theta , \\phi) = (\\mathbf {c}, \\sigma), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 309, + 890, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g$ is the radiance field in canonical configuration and $f$ is a scene flow field representing the mapping between the scene at time instant $t$ and the canonical space.", + "bbox": [ + 496, + 356, + 890, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the above formulation one can easily observe that both NeRF-T and D-NeRF rely on modeling a 4D field $f(x,y,z,t)$ , i.e., the dynamic implicit field in NeRF-T and the flow field in D-NeRF. Existing methods mainly adopt MLPs to fit these 4D fields. Such an implicit neural representation does not have an explicit structure and requires extensive computation time for both training and rendering. Tri-projection Decomposition. The tri-projection decomposition is widely used in recent work including EG3D [7] and TensorRF [8] in order to accelerate the training and rendering process in MLP-based NeRF frameworks. Such decomposition factorizes a $n$ -dimensional tensor $V_{h}$ into three lower-dimensional $((n - 1)$ -D) tensors $V_{l}^{i} \\{i = 1,2,3\\}$ by projecting $V_{h}$ along the $x$ , $y$ and $z$ -axis respectively. For example, triplane-based decomposition proposed by EG3D projects the 3D tensor into three 2D feature planes. Compared to voxel-based representations, triplane representation effectively reduces the memory footprint and improves the performance of 3D generation and reconstruction.", + "bbox": [ + 496, + 401, + 892, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hierarchical Tri-projection Decomposition", + "text_level": 1, + "bbox": [ + 500, + 696, + 862, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to design a dynamic scene representation with explicit feature grids to accelerate network training and volumetric rendering. However, directly constructing a 4D tensor costs a huge amount of memory and is unacceptable for the purpose of high-resolution rendering. Therefore, we propose a hierarchical triprojection decomposition to factorize the 4D tensor into several compact features, which reduces memory consumption by a large margin while preserving the capability of fitting 4D fields.", + "bbox": [ + 496, + 720, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, for a neural 4D field $f(x,y,z,t)$ , we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumetric tensors via tri-", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16634", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "projection decomposition:", + "bbox": [ + 76, + 92, + 253, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Pi_ {3} (f (x, y, z, t)) = \\left\\{f _ {z} (x, y, t), f _ {y} (x, z, t), f _ {x} (y, z, t) \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 113, + 467, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Pi_3$ denotes the projection operator. To further lower space complexity and enable high-resolution representation, we decompose each feature volume into three feature planes as:", + "bbox": [ + 76, + 143, + 468, + 204 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {3} (f _ {z}) = \\left\\{f _ {z t} (x, y), f _ {z y} (x, t), f _ {z x} (y, t) \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 210, + 408, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {3} (f _ {y}) = \\left\\{f _ {y t} (x, z), f _ {y z} (x, t), f _ {y x} (z, t) \\right\\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 229, + 468, + 247 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {3} (f _ {x}) = \\left\\{f _ {x t} (y, z), f _ {x z} (y, t), f _ {x y} (z, t) \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 250, + 408, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_3$ denotes the volume-to-plane tri-projection. In this way, we compactly represent a 4D field using 9 planes. Given any spatio-temporal coordinate $(x,y,z,t)$ , we can efficiently query its value in the 4D field by projecting it onto the planes and retrieving the corresponding value via bilinear interpolation. Fig 2 is an illustration of our decomposition method. Our hierarchical triprojection decomposition reduces the space complexity from $\\mathcal{O}(n^4)$ to $\\mathcal{O}(n^2)$ with $n$ being the spatial resolution of the grid, significantly lowering memory footprint without sacrificing representation power.", + "bbox": [ + 75, + 271, + 468, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differences from 6-plane decomposition. Compared with 6-plane decomposition [5, 13, 21] which pairs the time dimension with only one spatial dimension $(xt,yt,zt)$ , our method first decomposes the 4D tensor in the spatial domain to obtain three time-aware volumes and then 9 decomposed planes. Note that the three volumes are decomposed independently, making the 9 planes different from each other. Specifically, when using a linear layer $g = \\{\\}$ to decompose the volume $f_{z}(x,y,t) = \\{.,f_{zy}(x,t),\\}.$ and the volume $f_{y}(x,z,t) = \\{.,f_{yz}(x,t),\\}.$ , the planes $f_{zy}(x,t)$ and $f_{yz}(x,t)$ behave differently since they are optimized independently to capture time variations in the volumes $f_{z}(x,y,t)$ and $f_{y}(x,z,t)$ , respectively. In this way, the 9 planes can leverage time-aware information in all possible combinations of the time and spatial dimensions $(xt,yt,zt,xyt,yzt,xzt,xyzt)$ . This property enables our representation to capture various dynamic information hierarchically at different levels of spatial dimensions. Therefore, our method is more efficient in representing dynamic scenes, where variations in time dimension are often intense, complex, and long-range.", + "bbox": [ + 75, + 436, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our decomposition supports various types of 4D fields including the time-conditioned radiance field in NeRF-T and the 4D flow field in D-NeRF. In this paper, we present a dynamic radiance field decomposition method (Sec. 4.1) for multiview dynamic reconstruction as well as a 4D flow decomposition method (Sec. 4.2) for single-view setting.", + "bbox": [ + 75, + 750, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Coarse-to-fine Strategy", + "text_level": 1, + "bbox": [ + 76, + 847, + 290, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further improve the efficiency of our 4D decomposition, we propose an optional coarse-to-fine strategy to fac", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "torize the 4D fields into different scales in different training phases. In coarse level, we adopt low-resolution feature planes $(128 \\times 128)$ to decompose the 4D fields, which improves the robustness of the training process and achieves fast convergence. After coarse level training, we additionally use high-resolution feature planes $(512 \\times 512)$ for 4D decomposition to represent dynamic details and achieve high-quality rendering. Specifically, we factorize the 4D fields into different scales:", + "bbox": [ + 496, + 90, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (x, y, z, t) = \\left\\{\\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right), \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {H R}\\right)\\right) \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 236, + 890, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f^{LR}$ and $f^{HR}$ are the coarse-level and fine-level components of $f(x,y,z,t)$ , respectively. In the coarse level, the decomposed feature planes is low-resolution to represent coarse 3D structures and 4D dynamic changes. In the fine level, we adopt the high-resolution feature planes in each element to decompose the 4D fields, which focuses more on recovering dynamic details.", + "bbox": [ + 496, + 266, + 890, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Tensor4D for Dynamic Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 385, + 846, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we apply our 4D tensor decomposition into the task of dynamic reconstruction with two inputs:", + "bbox": [ + 498, + 411, + 890, + 440 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Dynamic reconstruction under sparse and fixed cameras. For this setting, we factorize NeRF-T instead of D-NeRF using our proposed 4D decomposition, which we found more efficient and flexible to represent topologically-varying objects (Sec. 4.1).", + "2) Dynamic reconstruction using monocular camera. For this setting, we separately decompose the 4D flow fields and the 3D canonical representation since it ensures appearance consistency across different frames and achieves more robust performance (Sec. 4.2)." + ], + "bbox": [ + 496, + 440, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Multi-view Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 599, + 740, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our dynamic reconstruction system under sparse multi-view setting. The goal of our system is 1) efficient and high-quality dynamic reconstruction with low memory and time cost, and 2) robust reconstruction under sparse and fixed camera setting.", + "bbox": [ + 496, + 623, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we adopt our 4D decomposition (Eq. 5) to factorize the NeRF-T representation with the coarse-to-fine strategy. Specifically, we can obtain nine low-resolution feature planes $T^{LR}$ and nine high-resolution feature planes $T^{HR}$ to represent the 4D NeRF-T fields after our 4D decomposition. Then for volume rendering, when sampling a point $p$ at $(x,y,z,t)$ in the ray $r$ with direction of $(\\theta ,\\phi)$ , we first query the feature $F_{p}$ of $p$ in both the LR and the HR feature planes. Take the LR features for example:", + "bbox": [ + 496, + 696, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {p} ^ {L R} (x, y, z, t) = \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right) \\\\ = \\oplus \\left(\\pi_ {3} \\left(f _ {z} ^ {L R}\\right), \\pi_ {3} \\left(f _ {y} ^ {L R}\\right), \\pi_ {3} \\left(f _ {x} ^ {L R}\\right)\\right) \\tag {6} \\\\ = \\oplus (\\oplus (T _ {z t} ^ {L R} (x, y), T _ {z y} ^ {L R} (x, t), T _ {z x} ^ {L R} (y, t)), \\dots), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 842, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16635", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a3bdc139b699f7c24c0f5bae50f399896add823a58f9586da5d036be18534439.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 436, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0f8f3c10739fd181a473aaa5cbad2e9abf03bffb760a0cfae6ea6412220e2f4.jpg", + "image_caption": [ + "(b)", + "Figure 3. The framework of Tensor4D for multi-view and monocular reconstruction. a). Tensor4D for multi-view reconstruction. The 4D NeRF-T fields are separately factorized by the LR and HR feature planes. b). Tensor4D for monocular reconstruction. The 4D flow fields are factorized by the LR feature plane for better disentanglement of shape and motion. The 3D canonical representation is factorized by three LR and HR feature planes." + ], + "image_footnote": [], + "bbox": [ + 439, + 87, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we concatenate all features queried in the nine LR planes as the final LR feature of the sampling point $p$ . Then we concatenate the LR and HR features with the positional encoding of $(x, y, z, t)$ into the geometry MLP $E_{g}$ to obtain the density $\\sigma$ and high-dimension feature $\\mathbf{f}$ :", + "bbox": [ + 75, + 324, + 470, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {g} \\left(F _ {p} ^ {L R}, F _ {p} ^ {H R}, \\gamma (x, y, z, t)\\right) = (\\mathbf {f}, \\sigma), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 404, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma (\\cdot)$ is the positional encoding function. Next, we concatenate the high-dimension feature $\\mathbf{f}$ with the positional encoding of $(\\theta ,\\phi)$ and feed it into the color MLP:", + "bbox": [ + 76, + 430, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {c} (\\mathbf {f}, \\gamma (\\theta , \\phi)) = c \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 482, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we can render the images through volume rendering and adopt color loss to train our decomposed feature planes $T$ and neural network $E_{c}$ and $E_{g}$ :", + "bbox": [ + 76, + 506, + 468, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c} = \\left\\| C _ {i, j} - C _ {i, j} ^ {*} \\right\\| \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 558, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With our coarse-to-fine design, our method can recover high-fidelity dynamic details effectively and efficiently.", + "bbox": [ + 76, + 582, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve robust dynamic reconstruction under sparse multi-view setting, we further adopt regularization for all decomposed feature planes $T$ :", + "bbox": [ + 76, + 612, + 468, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r} = \\sum_ {T} \\sum_ {i, j} \\sqrt {\\left(T _ {i + 1 , j} - T _ {i , j}\\right) ^ {2} + \\left(T _ {i , j + 1} - T _ {i , j}\\right) ^ {2}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 662, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{r}$ is the TV loss for each feature plane $T$ to keep their sparsity. To regularize the geometry, we also introduce surface constraint into volume rendering. Specifically, we adopt the SDF as the base geometry representation and follow NeuS [59] to render the SDF field. Then we add surface constraint loss to enforce a smooth surface:", + "bbox": [ + 76, + 712, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e} = \\left\\| \\left\\| \\nabla s (x, y, z, t) \\right\\| _ {2} - 1 \\right\\| _ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 809, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final training loss $L_{m}$ is the regularization loss $L_{r}, L_{e}$ and the color loss $L_{c}$ :", + "bbox": [ + 76, + 833, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m} = \\lambda_ {r} L _ {r} + \\lambda_ {e} L _ {e} + \\lambda_ {c} L _ {c} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 872, + 468, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Monocular Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 324, + 741, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from the sparse view setting, we adopt 4D decomposition for D-NeRF in monocular capture cases. This is because monocular setting is much more ill-posed than sparse-view inputs, and the explicit disentanglement of appearance and motions can guarantee the consistency across different frames. Since D-NeRF represents the dynamic objects with the 4D flow fields and a 3D canonical representation, we separately factorize these two fields. First, for the 4D flow fields, we only adopt coarse level decomposition and factorize it into low-resolution feature planes:", + "bbox": [ + 496, + 347, + 890, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x, y, z, t) = \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 505, + 890, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our coarse decomposition focuses more on the coarse and rigid motion, which improves the robustness of flow estimation and can achieve better disentanglement of shape and motion. Then for the 3D canonical representation, we adopt both coarse and fine level decomposition:", + "bbox": [ + 496, + 532, + 893, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh (\\hat {x}, \\hat {y}, \\hat {z}) = \\pi_ {3} \\left(h ^ {H R}\\right) + \\pi_ {3} \\left(h ^ {L R}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 616, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we can obtain 9 flow feature planes $T_{f}$ for 4D flow fields and 6 canonical feature planes $T_{h}$ for 3D canonical representation. The volume rendering in monocular cases is similar to multi-view cases. For a sampling point $p$ , we first obtain the point flow feature $F_{f}$ using Eq. 6 with nine flow planes $T_{f}$ . Then we adopt the flow MLP $E_{f}$ to predict the movement of the point:", + "bbox": [ + 496, + 643, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {f} \\left(F _ {p}, \\gamma (x, y, z, t)\\right) = (\\hat {x}, \\hat {y}, \\hat {z}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 758, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we obtain the point canonical feature $F_{c}$ by querying the canonical feature planes $T_{h}$ . Take the LR feature $F_{c}^{LR}$ for example:", + "bbox": [ + 496, + 785, + 890, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c} ^ {L R} = \\pi_ {3} \\left(h ^ {L} R\\right) = \\oplus \\left(T _ {h _ {z}} ^ {L R} (x, y), T _ {h _ {y}} ^ {L R} (x, z), T _ {h _ {z}} ^ {L R} (x, y)\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 838, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we feed canonical feature $F_{c}^{LR}$ and the positional encoding of $(\\hat{x},\\hat{y},\\hat{z})$ into the geometry MLP $E_{g}$ to obtain", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16636", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/772e007bb5bd11c0beaecf189bd005bec28fcbb2b10d1cb0d75a3666dd5c8522.jpg", + "image_caption": [ + "Figure 4. Example results of our method. Space and time novel view rendering results from sparse-view fixed cameras. The top three results are from four front view cameras and the bottom is from 12 circular cameras." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 325, + 496 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1a23377c08547516a08abb86d7e4c73f70ee98eaf1641b064c6ee641c65a4e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 88, + 870, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "high-dimension feature $\\mathbf{f}$ and density $\\sigma$", + "bbox": [ + 76, + 544, + 343, + 559 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {g} \\left(F _ {c} ^ {L R}, \\gamma (\\hat {x}, \\hat {y}, \\hat {z})\\right) = (\\mathbf {f}, \\sigma). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 569, + 468, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we adopt Eq. 8 to predict color $c$ for volume rendering and Eq. 9 for training color loss $L_{c}$ . We also add feature regularization loss $L_{r}$ in Eq. 10 and surface constraint loss $L_{e}$ in Eq. 11. The total training loss $L_{s}$ is:", + "bbox": [ + 76, + 598, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\lambda_ {c} L _ {c} + \\lambda_ {r} L _ {r} + \\lambda_ {e} L _ {e}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 670, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 710, + 200, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To evaluate the performance of our methods for multiview inputs, we build a sparse-view capture system with 6 forward-facing RGB cameras mounted on the borders of a $32^{\\prime \\prime}$ Looking Glass 3D holographic display [2]. All cameras are synchronized and calibrated. Using this system, we capture multiple sequences of various challenging human motions, including dancing, thumbing up, waving hands, wearing hats and manipulating bags. We use 4 of them for reconstruction and rendering in all our experiments, while leaving the other two for quantitative evaluation. We also use three $360^{\\circ}$ multiview full body sequences", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "captured with 12 evenly spaced cameras on a camera ring for qualitative evaluation. For monocular evaluation, we use the synthetic dataset provided by D-NeRF [40] and select 3 scenes (\"lego\", \"standup\" and \"jumpingjacks\") from this dataset, with the numbers of training frames ranging from 50 to 200. More details about data collection and preprocessing can be found in the Supp.Mat..", + "bbox": [ + 496, + 544, + 890, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We mainly compare our method against the following state-of-the-art baselines that are most related to our work: D-NeRF [40], NeRF-T, TiNeuVox [12] and NeuS-T. Here, NeRF-T is our extension of vanilla NeRF [32] by introducing an additional time input, and NeuS-T is extended from NeuS [59] similarly. Among these baselines, D-NeRF and TiNeuVox represent the dynamic scenes through deforming a canonical one, while NeRF-T and NeuS-T directly learn a time-conditioned 4D radiance fields. TiNeuVox uses explicit voxel grids to accelerate network training, while others purely use MLPs to model the scene.", + "bbox": [ + 496, + 650, + 892, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Results and Comparison", + "text_level": 1, + "bbox": [ + 500, + 830, + 723, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. We train our model for each individual sequences, and present some example results for novel view synthesis in Fig. 4 and the Supp.Mat.. The results", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16637", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a14f9d797e8a62129d4ecb3955caf376d3547580a7a52147704965d4b2174ccf.jpg", + "image_caption": [ + "Figure 5. Comparison on monocular synthetic dataset against D-NeRF [40] and TiNeuVox [12]." + ], + "image_footnote": [], + "bbox": [ + 89, + 83, + 879, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acf46aa699b1d62650b438d808f1141b69a1759a158d82ff834a2149bc844d31.jpg", + "image_caption": [ + "Figure 6. Comparison on sparse-view real-world dataset against D-Nerf [40], TiNeuVox [12] and Neus-T [59]. Four of camera views from our datasets are used for reconstruction and rendering." + ], + "image_footnote": [], + "bbox": [ + 89, + 338, + 879, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cover various body motions, clothing styles and accessories. As shown in Fig. 4, our method can render high-quality images for dynamic scenes and faithfully recover appearance details like the thin finger motions, semi-transparent silk, hand-object interaction, face expressions and cloth wrinkles. See our Supp Video for better visualization.", + "bbox": [ + 75, + 638, + 468, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on monocular dynamic dataset. We first compare our method with the baselines on monocular synthetic dataset. Qualitative results are presented in Fig. 5. Compared to other methods, ours recovers more appearance details and generate less artifacts. The numeric results in Tab. 1 also prove that our method outperforms state-of-the-art methods in terms of rendering quality and accuracy.", + "bbox": [ + 75, + 732, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on sparse view dataset. We then evaluate the performance of different methods for novel view synthesis given four camera views from our collected six camera-view real-world datasets. The remained two views are used", + "bbox": [ + 75, + 840, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for quantitative evaluation. Results reported in Fig. 6 and Tab. 2 show again that our method performs better in accurate and high-quality appearance detail synthesis.", + "bbox": [ + 498, + 638, + 890, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons of training efficiency. We compare the training time and model size for memory in Tabs. 3 and 4. Our method requires significantly less training time compared to the original D-NeRF, NeRF-T, and Neus-T in monocular and multi-view scenarios. Compared to Ours-NeRF-T and Ours-D-NeRF, the longer training time observed in our method is due to the additional computation for geometry smoothness terms, which reduces artifacts and enhances rendering quality. Compared to TiNeuVox, our method has a reduction in memory consumption and a clear improvement in rendering quality since Tensor4D enables the efficient decomposition of 4D fields at higher spatial resolutions (512 $^3$ vs. 256 $^3$ ).", + "bbox": [ + 496, + 705, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16638", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c1d9bdd2152000a12f33af7d4b05440870eea11cc500f1594f6c781fc9f200b.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on monocular synthetic datasets." + ], + "table_footnote": [], + "table_body": "
MethodLegoStandupJumpingjacks
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]7.83e-321.260.8691.72e-14.63e-434.380.9892.18e-25.46e-433.370.9874.84e-2
NeRF-T3.89e-324.320.9041.55e-16.82e-431.440.9682.36e-25.99e-432.270.9795.37e-2
TiNeuVox [12]3.15e-325.140.9248.37e-22.90e-436.180.9862.02e-23.89e-434.760.9833.33e-2
Ours-NeRF-T3.10e-325.130.9221.24e-15.92e-432.380.9772.29e-25.62e-432.730.9805.16e-2
Ours-D-NeRF4.61e-323.370.8901.19e-12.82e-435.930.9812.07e-24.22e-434.100.9823.41e-2
Ours2.26e-326.710.9533.49e-22.50e-436.320.9831.74e-23.91e-434.430.9823.18e-2
", + "bbox": [ + 80, + 102, + 890, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab063f6a2197923e964f5c65749a7bb40d478b96a63fe5692b04391f834d0fd7.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on the six camera-view real-world datasets." + ], + "table_footnote": [], + "table_body": "
MethodSequence1-thzEarphoneSequence3-yxd
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]3.13e-325.150.9100.1858.16e-320.920.8540.2564.84e-323.220.9370.147
TiNeuVox [12]5.72e-322.790.8320.2091.49e-218.550.7070.3198.18e-321.190.8160.233
Neus-T [59]4.95e-323.070.8870.1303.75e-324.220.8770.1842.20e-326.590.9450.099
Ours-NeRF-T5.85e-322.460.8420.1911.27e-219.130.8380.2185.02e-323.060.9140.141
Ours-D-NeRF4.74e-323.270.8640.1768.17e-321.070.8830.1984.06e-323.980.9260.130
Ours1.53e-328.270.9420.0843.20e-325.000.9030.1531.31e-328.830.9620.072
", + "bbox": [ + 80, + 218, + 890, + 318 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a140ee71bad878dacdb7ae0facf4357796904ededcb64351203527c1bfd224e.jpg", + "table_caption": [ + "Table 3. Training time and memory report in monocular cases." + ], + "table_footnote": [], + "table_body": "
MethodLego
PSNRTimeIterations#Params
TiNeuVox25.1434min50k102M
D-NeRF21.2645h800k4.8M
Ours-D-NeRF23.3795min50k20M
NeRF-T24.3238h800k4.4M
Ours-NeRF-T25.1373min50k43M
Ours(6-planes)26.34135min50k17M
Ours26.71144min50k20M
", + "bbox": [ + 96, + 345, + 444, + 458 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/50b32a2236fab83c008260b44917dfffd0b4d93a5aae3738794ad5469e4edc47.jpg", + "table_caption": [ + "Table 4. Training time and memory report in multi-view cases." + ], + "table_footnote": [], + "table_body": "
MethodSequence1-thz
PSNRTimeIterations#Params
TiNeuVox22.7930min50k102M
D-NeRF25.1537h800k4.8M
Neus-T23.0745h800k5.0M
Ours-NeRF-T22.4668min50k43M
Ours(6-planes)27.86105min50k32M
Ours28.27117min50K43M
", + "bbox": [ + 96, + 478, + 444, + 577 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/73ec3938bcbbfa8ebd575483f56155acca6518b9dfdd3876ec533bf33059a11a.jpg", + "table_caption": [ + "Table 5. Ablation study of hierarchical decomposition and smoothness terms." + ], + "table_footnote": [], + "table_body": "
MethodLegoStandup
PSNRSSIMPSNRSSIM
Ours(w/o regular)26.490.94635.910.977
Ours(6-planes)26.440.94435.790.978
Ours26.710.95336.320.983
Sequence-thzSequence-earphone
PSNRSSIMPSNRSSIM
Ours(w/o regular)27.920.93224.750.885
Ours(6-planes)27.740.93424.390.889
Ours28.270.94225.000.903
", + "bbox": [ + 89, + 609, + 450, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 742, + 225, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regularization. We quantitatively ablate the regularization terms in our method. We implement two strong baselines \"Ours-NeRF-T\" and \"Ours-D-NeRF\", in which we directly apply our Tensor4D decomposition for NeRF-T and D-NeRF without the regularization terms. The results are reported in Tab. 1 and Tab. 2. Benefiting from our 4D decomposition, they achieve better performance than the original NeRF-T and D-NeRF. However, their rendering quality is still worse than our full method with regularization. In", + "bbox": [ + 75, + 766, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "addition, We ablate for TV regular smoothness terms (Ours vs. Ours(w/o regular)) and the results are reported in Tab. 5, which further validates the effectiveness of our smoothness terms for rendering quality enhancement.", + "bbox": [ + 498, + 330, + 890, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hierarchical Decomposition We quantitatively ablate our hierarchical decomposition with 6-plane decomposition. As shown in Tab 5, our method achieves superior performance in both multi-view and monocular cases, which validates the effectiveness of our hierarchical decomposition.", + "bbox": [ + 498, + 391, + 890, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 479, + 612, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Since our method needs to decompose the 4D fields to several 2D feature planes, a pre-set bounding box of the scene is necessary. Therefore, it is difficult for our method to reconstruct backgrounds or objects which are out of the bounding box. Another limitation is our strong regularization terms. Though these terms benefit the robustness of our reconstruction under sparse views, they also limit our ability to handle challenging cases such as fluid and fog.", + "bbox": [ + 496, + 506, + 890, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We presented Tensor4D, a new method for learning high-quality neural representation for dynamic scenes from sparse-view videos or even a monocular video. To capture the spatio-temporal information in a compact and memory-efficient manner, we propose to develop a novel hierarchical tri-projection decomposition method that models a 4D tensor with nine 2D feature planes. With proper design of training losses and regularization, our method provides an efficient yet effective solution to model the radiance fields of dynamic scenes. We believe our work can inspire future research towards low-cost, portable and immersive telepresence systems.", + "bbox": [ + 496, + 627, + 890, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107 and No.61827805 and the Beijing Science and Technology Project (Z211100004021006).", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16639", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] 2", + "[2] The looking glass holographic display. https://lookingglassfactory.com/.6", + "[3] Naveed Ahmed, Christian Theobalt, Petar Dobrev, Hans-Peter Seidel, and Sebastian Thrun. Robust fusion of dynamic shape and normal capture for high-quality reconstruction of time-varying geometry. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2", + "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3, 4", + "[6] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In arXiv, 2020. 2", + "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 3", + "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 3", + "[9] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. 34(4), jul 2015. 1", + "[10] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6), nov 2017. 2", + "[11] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE International Conference on Computer Vision, pages 14304-14314, 2021. 1, 2", + "[12] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In arXiv preprint arXiv:2205.15285, 2022. 3, 6, 7, 8", + "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 4", + "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5491–5500, 2022. 2" + ], + "bbox": [ + 78, + 116, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4d: Voxel for 4d novel view synthesis. In arXiv preprint arXiv:2205.14332, 2022. 3", + "[16] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In In Neural Information Processing Systems (NeurIPS), 2022. 2", + "[17] Paulo FU Gotardo, Tomas Simon, Yaser Sheikh, and Iain Matthews. Photogeometric scene flow for high-detail dynamic 3d reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 846-854, 2015. 2", + "[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2", + "[19] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Trans. Graph., 38(6), nov 2019. 1", + "[20] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Dynamic shape capture using multi-view photometric stereo. In ACM SIGGRAPH Asia 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery. 2", + "[21] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 3, 4", + "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2", + "[23] T. Kanade, P. Rander, and P.J. Narayanan. Virtualized reality: constructing virtual worlds from real scenes. IEEE MultiMedia, 4(1):34-47, 1997. 1", + "[24] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3-d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2", + "[25] Jason Lawrence, Danb Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Trans. Graph., 40(6), dec 2021. 2", + "[26] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Trans. Graph., 31(1), feb 2012. 1", + "[27] Tianye Li, Mira Slavcheva, Michael Zollhofer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16640", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2", + "[28] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In arXiv preprint, arXiv:2205.15723, 2022. 1, 2, 3", + "[29] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2", + "[30] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2", + "[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2", + "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 6", + "[33] Thomas Muller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. 41(4), jul 2022. 3", + "[34] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2", + "[35] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision (ICCV), 2021. 2", + "[37] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2", + "[38] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higherdimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2", + "[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Albert Pumarola, Enric Corona, and and Francesc Moreno-Noguer Gerard Pons-Moll. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 3, 6, 7, 8", + "[41] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2", + "[42] Chris Russell, Rui Yu, and Lourdes Agapito. Video pop-up: Monocular 3d reconstruction of dynamic scenes. In Computer Vision - ECCV 2014, pages 583-598, Cham, 2014. Springer International Publishing. 2", + "[43] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 20154-20166. Curran Associates, Inc., 2020. 2", + "[44] Ruizhi Shao, Liliang Chen, Zerong Zheng, Hongwen Zhang, Yuxiang Zhang, Han Huang, Yandong Guo, and Yebin Liu. Floren: Real-time high-quality human performance rendering via appearance flow using sparse rgb cameras. In SIGGRAPH Asia 2022 Conference Papers, pages 1-10, 2022. 2", + "[45] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yanpei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In CVPR, 2022. 2", + "[46] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. Diffustereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 702-720. Springer, 2022. 2", + "[47] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2", + "[48] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2", + "[49] Jonathan Starck and Adrian Hilton. Surface capture for performance-based animation. IEEE Computer Graphics and Applications, 27(3):21-31, 2007. 1", + "[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2", + "[52] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16641", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 2", + "[53] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022. 2", + "[54] Aparna Taneja, Luca Ballan, and Marc Pollefeys. Modeling dynamic scenes recorded with freely moving cameras. Computer Vision - ACCV 2010, page 613-626, 2011. 2", + "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE International Conference on Computer Vision, pages 12959-12970, 2021. 1, 2", + "[56] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. In arXiv preprint, arXiv:2105.05994, 2021. 1, 2, 3", + "[57] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. IButter: Neural Interactive Bullet Time Generator for Human Free-Viewpoint Rendering, page 4641–4650. Association for Computing Machinery, New York, NY, USA, 2021. 2", + "[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 3", + "[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 5, 6, 7, 8", + "[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 2", + "[61] Liwen Wu, Jae Yong Lee, Yu-Xiong Wang Anand Bhattachad, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 3", + "[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 1, 2, 3", + "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 2", + "[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and ap" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pearance. Advances in Neural Information Processing Systems, 33, 2020. 2", + "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2, 3", + "[66] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5746-5756, 2021. 2", + "[67] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 40(6), dec 2021. 2", + "[68] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7743-7753, June 2022. 2", + "[69] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 2", + "[70] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2" + ], + "bbox": [ + 503, + 92, + 890, + 551 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16642", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_model.json b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cc84c99a913c693417aac0722682be7e793e8360 --- /dev/null +++ b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.852, + 0.175 + ], + "angle": 0, + "content": "Tensor4D : Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.202, + 0.844, + 0.24 + ], + "angle": 0, + "content": "Ruizhi Shao\\(^{1}\\), Zerong Zheng\\(^{1,2}\\), Hanzhang Tu\\(^{1}\\), Boning Liu\\(^{1}\\), Hongwen Zhang\\(^{1}\\), Yebin Liu\\(^{1}\\) \n\\(^{1}\\)Department of Automation, Tsinghua University \\(^{2}\\)NNKosmos Technology" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.281, + 0.229, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.396, + 0.166, + 0.407 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.281, + 0.363, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.396, + 0.305, + 0.408 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.282, + 0.532, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.396, + 0.457, + 0.408 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.282, + 0.703, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.396, + 0.627, + 0.408 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.282, + 0.884, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.396, + 0.802, + 0.408 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.412, + 0.893, + 0.467 + ], + "angle": 0, + "content": "Figure 1. Given 4 sparse static RGB camera views of a dynamic scene (a), our proposed Tensor4D decomposition enables multiview reconstruction to achieve fine-grained geometry reconstruction even on human fingers (b) and temporal-consistent novel view synthesis on a 3D holographic display (c,d,e). The 4 cameras are settled on four corners of the display. The proposed method demonstrates low-cost, portable and highly immersive telepresence experience." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.48, + 0.314, + 0.495 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.513, + 0.474, + 0.812 + ], + "angle": 0, + "content": "We present Tensor4D, an efficient yet effective approach to dynamic scene modeling. The key of our solution is an efficient 4D tensor decomposition method so that the dynamic scene can be directly represented as a 4D spatiotemporal tensor. To tackle the accompanying memory issue, we decompose the 4D tensor hierarchically by projecting it first into three time-aware volumes and then nine compact feature planes. In this way, spatial information over time can be simultaneously captured in a compact and memory-efficient manner. When applying Tensor4D for dynamic scene reconstruction and rendering, we further factorize the 4D fields to different scales in the sense that structural motions and dynamic detailed changes can be learned from coarse to fine. The effectiveness of our method is validated on both synthetic and real-world scenes. Extensive experiments show that our method is able to achieve high-quality dynamic reconstruction and rendering from sparse-view camera rigs or even a monocular camera. The code and dataset will be released at https://github.com/DSaurus/Tensor4D." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "High quality reconstruction and Photo-realistic rendering of a dynamic scene from a set of input images is nec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.893, + 0.586 + ], + "angle": 0, + "content": "essary for many applications such as AR/VR, 3D content production and entertainment. Traditional methods use classical mesh-based representation to reconstruct the dynamic scenes, which, unfortunately, are prone to produce reconstruction errors and rendering artifacts when the scene contains thin structures, specular surfaces and topological changes [9, 19, 23, 26, 49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Recent advances in neural rendering approaches, which learn scene representations in the form of neural radiance fields (NeRF), have shown impressive novel view synthesis of general static scenes given only multi-view images [32]. They are immediately extended to dynamic scenes: some methods (e.g., NeRF-T) consider time as an additional input dimension to NeRF representation [56,62], while other methods (e.g., D-NeRF) disentangle a dynamic scene into a canonical radiance field and a dynamic motion field [11,28,37,40,55]. Either way, learning a 4D function is one of the main cornerstones. Unfortunately, directly using MLP to fit such a function often suffers from high time and computation cost, i.e., dozens of hours on high-end GPUs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In fact, the aforementioned limitation also exists in conventional NeRF-based methods for static scenes, and researchers have proposed to use discrete data structures like voxel grids [65] or triplanes [7] to accelerate NeRF training and rendering. However, these techniques are difficult to be extended to dynamic domains as introducing an additional time dimension will exponentially increase memory footprint, hindering them from modeling high-quality ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16632" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.19, + 0.106 + ], + "angle": 0, + "content": "pearance details." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.108, + 0.468, + 0.36 + ], + "angle": 0, + "content": "In this work, we pursue a dynamic scene representation that also utilizes explicit feature grids to accelerate network training while avoiding huge memory consumption when introducing an additional time dimension. To this end, we bypass the construction of a high resolution 4D tensor; instead, we propose to model a 4D field using hierarchical tri-projection decomposition. Our decomposition method extends the tri-projection in EG3D [7]. It firstly project a full 4D field into three time-aware volumes, each of which is then further decomposed into three feature planes. In this way, we model the 4D field using only nine 2D feature planes, and we empirically find that although being highly compact, such a representation is powerful enough to represent dynamic scenes containing complex motions. Moreover, the usage of explicit data structure also allows us to design a coarse-to-fine strategy to further improve the performance of our method." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.364, + 0.468, + 0.631 + ], + "angle": 0, + "content": "By utilizing and factorizing an explicit 4D tensor, our method enables both efficient reconstruction and compact representation of dynamic scenes. Besides, the decomposition scheme also introduces implicit constraints on the representation since only low-rank tensors can be approximated by a small number of lower-dimensional components. Such a constraint can serve as an inherent regularization when the input observation is limited, e.g., under sparse and fixed cameras setting or even monocular inputs. In this paper, we first apply our method for sparse-view dynamic reconstruction by adopting our Tensor4D decomposition to time-conditioned radiance fields in \"NeRF-T\". In addition, our decomposition method can also be used for single-view dynamic reconstruction. This is achieved through decomposing both the 4D dynamic motion field and the canonical radiance field in \"D-NeRF\". With proper regularization, our system enables efficient and high-quality reconstruction of dynamic objects under both camera settings." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.649, + 0.217, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.675, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Multiview Reconstruction and Rendering. There are various ways to capture and reconstruct the 3D dynamic scenes, involving methods based on silhouette [24, 54], stereo [30, 46], flow [17, 44, 70], segmentation [41, 42], and photometric [3, 20]. With RGBD cameras, real-time solutions like DynamicFusion [34] estimates the non-rigid deformations of a dynamic scene and integrates depth frames to reconstruct the geometry model in the canonical space. This method is later extended for telepresence and holographic communication [10, 25, 66]. However, these systems heavily rely on depth sensors to obtain accurate geometry. In contrast, our method can reconstruct and render dynamic scenes using sparse-view RGB cameras." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In the past few years, neural implicit representations underwent rapid development and have been applied for multi-" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.345 + ], + "angle": 0, + "content": "view reconstruction and rendering of static scenes. Some methods represent the geometry as the zero level-set of a neural network, and use differentiable surface rendering to optimize the network weights [35,47,64]. Given dense observation of an object, these methods are able to accurately recover its surface. NeRF [32], on the other hand, uses volume rendering to optimize the scene representation. Its simplicity and impressive results inspires a lot of following works, including in-the-wild reconstruction [31, 52], lighting and material estimation [4,48,67], generation [1,6,18, 22,43,53], human rendering [45,51,57,68,69] and so on. More recently, several methods unify surface and volumetric rendering, enabling accurate geometry reconstruction and high-quality novel view synthesis [36,59,63]. Compared to these static scene representations, we aim to enable free view synthesis of dynamic scenes from extremely sparse cameras." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.347, + 0.89, + 0.526 + ], + "angle": 0, + "content": "NeRF for Dynamic Scenes. Modeling scenes in 4D domain with time dimension included is a direct solution to extend NeRF for dynamic domains. Typical approaches include VideoNeRF [62], NeRFlow [11], DyNeRF [27], and DCT-NeRF [56]. Specifically, VideoNeRF [62] directly learns a spatiotemporal irradiance field from a single video and uses depth estimation to address the shape-motion ambiguities in monocular inputs, while NeRFlow [11] and DCT-NeRF [56] use point trajectory to regularize the network optimization. To deal with the limitation of topology-change modeling in deformation field, Park et al. [38] represent HyperNeRF which can lift NeRF to higher dimensions." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.528, + 0.89, + 0.735 + ], + "angle": 0, + "content": "Dynamic scenes can also be rendered by deforming the radiance field in the canonical space. For example, Nerfies [37] optimizes an additional continuous deformation field by warping each observed point into a canonical 5D NeRF. D-Nerf [40] and NR-Nerf [55] follow a similar framework, but take only monocular videos as training data. In addition, DeVRF [28] uses voxel-based representation instead of MLPs to model the 3D canonical space and the 4D deformation field. Using parametric body templates as the semantic prior, methods like Neural Body [39] and HumanNeRF [60] enable photo-realistic novel view synthesis of complex human performance. The discrepancy between the practical capture process and the existing experimental protocols for monocular videos has been shown in [16]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.737, + 0.89, + 0.9 + ], + "angle": 0, + "content": "NeRF Acceleration. Numerous works emerge with the purpose of speeding up static NeRF using explicit data structures including feature maps, voxels and tensors. For instance, DVGO [50] achieves fast convergence through an explicit representation of a density voxel grid and a feature voxel grid. With the sparse voxel octree structure, NSVF [29] accelerates the novel view reconstruction by discarding the empty voxels in a coarse-to-fine manner. Similarly, Plenoxels [14] and PlenOctree [65] model a scene through a hierarchical 3D grid with spherical harmonics, which can realize an optimization with two orders of magnitude faster" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16633" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.462, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Illustration of our hierarchical tri-projection decomposition method. For a neural 4D field \\( f(x,y,z,t) \\), we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumes, which are then further projected onto nine 2D planes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.464 + ], + "angle": 0, + "content": "than NeRF. In DIVeR [61], ray marching only finds a fixed number of hits on the voxel grid to accelerate volumetric rendering. Moreover, hashing encoding [33] and tensor decomposition [8] are also used as compact representations for NeRF acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.471, + 0.676 + ], + "angle": 0, + "content": "For dynamic scene modeling, DeVRF [28] enables fast non-rigid neural rendering with both 3D volumetric and 4D voxel field. In addition, V4D [15] introduce an effective conditional positional encoding for 4D data to realize fast novel view synthesis. TiNeuVox [12] represents scenes with identifiable explicit data structures and accelerates radiance fields modeling, while Wang et al. extended PlenOctrees [65] into free-view video rendering [58]. However, the low-res volumetric design of these works hinders the capacity of rendering high-quality images. In addition, there are also several concurrent works [5, 13, 21] adopt 6-plane decomposition for dynamic scenes. Compared with these methods, our hierarchical decomposition is more efficient to capture time variations and represent dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.169, + 0.706 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Building on prior work for spatio-temporal representations and tri-projection decomposition (Sec. 3.1), we propose a hierarchical tri-projection decomposition method (Sec. 3.2) and a coarse-to-fine strategy (Sec. 3.3), which allow us to learn a 4D field at a modest cost of training time and GPU memory." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.208, + 0.833 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Spatio-temporal 4D NeRF representation. To represent dynamic objects using neural radiance fields, a naive way is to condition the original neural radiance fields on the times-stamp [56, 62], which we term as NeRF-T. Mathematically," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.704, + 0.106 + ], + "angle": 0, + "content": "NeRF-T can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.117, + 0.891, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f (x, y, z, t) = (\\mathbf {f}, \\sigma), \\tag {1} \\\\ g (\\mathbf {f}, \\theta , \\phi) = \\mathbf {c}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.164, + 0.892, + 0.237 + ], + "angle": 0, + "content": "where \\( f \\) is a dynamic implicit field that produces a high-dimensional feature \\( \\mathbf{f} \\) and a density value \\( \\sigma \\) for a point at position \\( (x, y, z) \\) and time instance \\( t \\), and \\( g \\) is a function that takes into account the viewing direction \\( (\\theta, \\phi) \\) to predict the final RGB color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.239, + 0.892, + 0.299 + ], + "angle": 0, + "content": "To achieve better disentanglement of shape and motion, some methods like D-NeRF [40] propose deformable neural radiance field which adopts a canonical 3D representation with the 4D flow fields:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.31, + 0.891, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f (x, y, z, t) = (\\hat {x}, \\hat {y}, \\hat {z}), \\tag {2} \\\\ g (\\hat {x}, \\hat {y}, \\hat {z}, \\theta , \\phi) = (\\mathbf {c}, \\sigma), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.401 + ], + "angle": 0, + "content": "where \\( g \\) is the radiance field in canonical configuration and \\( f \\) is a scene flow field representing the mapping between the scene at time instant \\( t \\) and the canonical space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.687 + ], + "angle": 0, + "content": "From the above formulation one can easily observe that both NeRF-T and D-NeRF rely on modeling a 4D field \\( f(x,y,z,t) \\), i.e., the dynamic implicit field in NeRF-T and the flow field in D-NeRF. Existing methods mainly adopt MLPs to fit these 4D fields. Such an implicit neural representation does not have an explicit structure and requires extensive computation time for both training and rendering. Tri-projection Decomposition. The tri-projection decomposition is widely used in recent work including EG3D [7] and TensorRF [8] in order to accelerate the training and rendering process in MLP-based NeRF frameworks. Such decomposition factorizes a \\( n \\)-dimensional tensor \\( V_{h} \\) into three lower-dimensional \\( ((n - 1) \\)-D) tensors \\( V_{l}^{i} \\{i = 1,2,3\\} \\) by projecting \\( V_{h} \\) along the \\( x \\), \\( y \\) and \\( z \\)-axis respectively. For example, triplane-based decomposition proposed by EG3D projects the 3D tensor into three 2D feature planes. Compared to voxel-based representations, triplane representation effectively reduces the memory footprint and improves the performance of 3D generation and reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.864, + 0.714 + ], + "angle": 0, + "content": "3.2. Hierarchical Tri-projection Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Our goal is to design a dynamic scene representation with explicit feature grids to accelerate network training and volumetric rendering. However, directly constructing a 4D tensor costs a huge amount of memory and is unacceptable for the purpose of high-resolution rendering. Therefore, we propose a hierarchical triprojection decomposition to factorize the 4D tensor into several compact features, which reduces memory consumption by a large margin while preserving the capability of fitting 4D fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Specifically, for a neural 4D field \\( f(x,y,z,t) \\), we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumetric tensors via tri-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.093, + 0.254, + 0.107 + ], + "angle": 0, + "content": "projection decomposition:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.114, + 0.468, + 0.144 + ], + "angle": 0, + "content": "\\[\n\\Pi_ {3} (f (x, y, z, t)) = \\left\\{f _ {z} (x, y, t), f _ {y} (x, z, t), f _ {x} (y, z, t) \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.145, + 0.47, + 0.205 + ], + "angle": 0, + "content": "where \\(\\Pi_3\\) denotes the projection operator. To further lower space complexity and enable high-resolution representation, we decompose each feature volume into three feature planes as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.212, + 0.41, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\pi_ {3} (f _ {z}) = \\left\\{f _ {z t} (x, y), f _ {z y} (x, t), f _ {z x} (y, t) \\right\\}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.231, + 0.469, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\pi_ {3} (f _ {y}) = \\left\\{f _ {y t} (x, z), f _ {y z} (x, t), f _ {y x} (z, t) \\right\\} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.25, + 0.41, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\pi_ {3} (f _ {x}) = \\left\\{f _ {x t} (y, z), f _ {x z} (y, t), f _ {x y} (z, t) \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.469, + 0.436 + ], + "angle": 0, + "content": "where \\(\\pi_3\\) denotes the volume-to-plane tri-projection. In this way, we compactly represent a 4D field using 9 planes. Given any spatio-temporal coordinate \\((x,y,z,t)\\), we can efficiently query its value in the 4D field by projecting it onto the planes and retrieving the corresponding value via bilinear interpolation. Fig 2 is an illustration of our decomposition method. Our hierarchical triprojection decomposition reduces the space complexity from \\(\\mathcal{O}(n^4)\\) to \\(\\mathcal{O}(n^2)\\) with \\(n\\) being the spatial resolution of the grid, significantly lowering memory footprint without sacrificing representation power." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Differences from 6-plane decomposition. Compared with 6-plane decomposition [5, 13, 21] which pairs the time dimension with only one spatial dimension \\((xt,yt,zt)\\), our method first decomposes the 4D tensor in the spatial domain to obtain three time-aware volumes and then 9 decomposed planes. Note that the three volumes are decomposed independently, making the 9 planes different from each other. Specifically, when using a linear layer \\(g = \\{\\}\\) to decompose the volume \\(f_{z}(x,y,t) = \\{.,f_{zy}(x,t),\\}.\\) and the volume \\(f_{y}(x,z,t) = \\{.,f_{yz}(x,t),\\}.\\), the planes \\(f_{zy}(x,t)\\) and \\(f_{yz}(x,t)\\) behave differently since they are optimized independently to capture time variations in the volumes \\(f_{z}(x,y,t)\\) and \\(f_{y}(x,z,t)\\), respectively. In this way, the 9 planes can leverage time-aware information in all possible combinations of the time and spatial dimensions \\((xt,yt,zt,xyt,yzt,xzt,xyzt)\\). This property enables our representation to capture various dynamic information hierarchically at different levels of spatial dimensions. Therefore, our method is more efficient in representing dynamic scenes, where variations in time dimension are often intense, complex, and long-range." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.751, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Our decomposition supports various types of 4D fields including the time-conditioned radiance field in NeRF-T and the 4D flow field in D-NeRF. In this paper, we present a dynamic radiance field decomposition method (Sec. 4.1) for multiview dynamic reconstruction as well as a 4D flow decomposition method (Sec. 4.2) for single-view setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.292, + 0.865 + ], + "angle": 0, + "content": "3.3. Coarse-to-fine Strategy" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "To further improve the efficiency of our 4D decomposition, we propose an optional coarse-to-fine strategy to fac" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.226 + ], + "angle": 0, + "content": "torize the 4D fields into different scales in different training phases. In coarse level, we adopt low-resolution feature planes \\((128 \\times 128)\\) to decompose the 4D fields, which improves the robustness of the training process and achieves fast convergence. After coarse level training, we additionally use high-resolution feature planes \\((512 \\times 512)\\) for 4D decomposition to represent dynamic details and achieve high-quality rendering. Specifically, we factorize the 4D fields into different scales:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.237, + 0.892, + 0.256 + ], + "angle": 0, + "content": "\\[\nf (x, y, z, t) = \\left\\{\\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right), \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {H R}\\right)\\right) \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.372 + ], + "angle": 0, + "content": "where \\( f^{LR} \\) and \\( f^{HR} \\) are the coarse-level and fine-level components of \\( f(x,y,z,t) \\), respectively. In the coarse level, the decomposed feature planes is low-resolution to represent coarse 3D structures and 4D dynamic changes. In the fine level, we adopt the high-resolution feature planes in each element to decompose the 4D fields, which focuses more on recovering dynamic details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.386, + 0.847, + 0.403 + ], + "angle": 0, + "content": "4. Tensor4D for Dynamic Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.891, + 0.441 + ], + "angle": 0, + "content": "As shown in Fig. 3, we apply our 4D tensor decomposition into the task of dynamic reconstruction with two inputs:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.516 + ], + "angle": 0, + "content": "1) Dynamic reconstruction under sparse and fixed cameras. For this setting, we factorize NeRF-T instead of D-NeRF using our proposed 4D decomposition, which we found more efficient and flexible to represent topologically-varying objects (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.59 + ], + "angle": 0, + "content": "2) Dynamic reconstruction using monocular camera. For this setting, we separately decompose the 4D flow fields and the 3D canonical representation since it ensures appearance consistency across different frames and achieves more robust performance (Sec. 4.2)." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.6, + 0.741, + 0.615 + ], + "angle": 0, + "content": "4.1. Multi-view Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.698 + ], + "angle": 0, + "content": "In this section, we present our dynamic reconstruction system under sparse multi-view setting. The goal of our system is 1) efficient and high-quality dynamic reconstruction with low memory and time cost, and 2) robust reconstruction under sparse and fixed camera setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.833 + ], + "angle": 0, + "content": "To this end, we adopt our 4D decomposition (Eq. 5) to factorize the NeRF-T representation with the coarse-to-fine strategy. Specifically, we can obtain nine low-resolution feature planes \\( T^{LR} \\) and nine high-resolution feature planes \\( T^{HR} \\) to represent the 4D NeRF-T fields after our 4D decomposition. Then for volume rendering, when sampling a point \\( p \\) at \\( (x,y,z,t) \\) in the ray \\( r \\) with direction of \\( (\\theta ,\\phi) \\), we first query the feature \\( F_{p} \\) of \\( p \\) in both the LR and the HR feature planes. Take the LR features for example:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.843, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {p} ^ {L R} (x, y, z, t) = \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right) \\\\ = \\oplus \\left(\\pi_ {3} \\left(f _ {z} ^ {L R}\\right), \\pi_ {3} \\left(f _ {y} ^ {L R}\\right), \\pi_ {3} \\left(f _ {x} ^ {L R}\\right)\\right) \\tag {6} \\\\ = \\oplus (\\oplus (T _ {z t} ^ {L R} (x, y), T _ {z y} ^ {L R} (x, t), T _ {z x} ^ {L R} (y, t)), \\dots), \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16635" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.437, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.237, + 0.264, + 0.251 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.088, + 0.891, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.236, + 0.678, + 0.251 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 3. The framework of Tensor4D for multi-view and monocular reconstruction. a). Tensor4D for multi-view reconstruction. The 4D NeRF-T fields are separately factorized by the LR and HR feature planes. b). Tensor4D for monocular reconstruction. The 4D flow fields are factorized by the LR feature plane for better disentanglement of shape and motion. The 3D canonical representation is factorized by three LR and HR feature planes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.471, + 0.4 + ], + "angle": 0, + "content": "where we concatenate all features queried in the nine LR planes as the final LR feature of the sampling point \\( p \\). Then we concatenate the LR and HR features with the positional encoding of \\( (x, y, z, t) \\) into the geometry MLP \\( E_{g} \\) to obtain the density \\( \\sigma \\) and high-dimension feature \\( \\mathbf{f} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.405, + 0.47, + 0.427 + ], + "angle": 0, + "content": "\\[\nE _ {g} \\left(F _ {p} ^ {L R}, F _ {p} ^ {H R}, \\gamma (x, y, z, t)\\right) = (\\mathbf {f}, \\sigma), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.47, + 0.476 + ], + "angle": 0, + "content": "where \\(\\gamma (\\cdot)\\) is the positional encoding function. Next, we concatenate the high-dimension feature \\(\\mathbf{f}\\) with the positional encoding of \\((\\theta ,\\phi)\\) and feed it into the color MLP:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.483, + 0.469, + 0.5 + ], + "angle": 0, + "content": "\\[\nE _ {c} (\\mathbf {f}, \\gamma (\\theta , \\phi)) = c \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.507, + 0.47, + 0.553 + ], + "angle": 0, + "content": "In this way, we can render the images through volume rendering and adopt color loss to train our decomposed feature planes \\( T \\) and neural network \\( E_{c} \\) and \\( E_{g} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.559, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\nL _ {c} = \\left\\| C _ {i, j} - C _ {i, j} ^ {*} \\right\\| \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.583, + 0.469, + 0.613 + ], + "angle": 0, + "content": "With our coarse-to-fine design, our method can recover high-fidelity dynamic details effectively and efficiently." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.613, + 0.47, + 0.657 + ], + "angle": 0, + "content": "To achieve robust dynamic reconstruction under sparse multi-view setting, we further adopt regularization for all decomposed feature planes \\( T \\):" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.664, + 0.469, + 0.712 + ], + "angle": 0, + "content": "\\[\nL _ {r} = \\sum_ {T} \\sum_ {i, j} \\sqrt {\\left(T _ {i + 1 , j} - T _ {i , j}\\right) ^ {2} + \\left(T _ {i , j + 1} - T _ {i , j}\\right) ^ {2}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.713, + 0.47, + 0.803 + ], + "angle": 0, + "content": "where \\( L_{r} \\) is the TV loss for each feature plane \\( T \\) to keep their sparsity. To regularize the geometry, we also introduce surface constraint into volume rendering. Specifically, we adopt the SDF as the base geometry representation and follow NeuS [59] to render the SDF field. Then we add surface constraint loss to enforce a smooth surface:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.81, + 0.469, + 0.827 + ], + "angle": 0, + "content": "\\[\nL _ {e} = \\left\\| \\left\\| \\nabla s (x, y, z, t) \\right\\| _ {2} - 1 \\right\\| _ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.834, + 0.469, + 0.863 + ], + "angle": 0, + "content": "The final training loss \\( L_{m} \\) is the regularization loss \\( L_{r}, L_{e} \\) and the color loss \\( L_{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.873, + 0.469, + 0.889 + ], + "angle": 0, + "content": "\\[\nL _ {m} = \\lambda_ {r} L _ {r} + \\lambda_ {e} L _ {e} + \\lambda_ {c} L _ {c} \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.325, + 0.743, + 0.339 + ], + "angle": 0, + "content": "4.2. Monocular Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.497 + ], + "angle": 0, + "content": "Different from the sparse view setting, we adopt 4D decomposition for D-NeRF in monocular capture cases. This is because monocular setting is much more ill-posed than sparse-view inputs, and the explicit disentanglement of appearance and motions can guarantee the consistency across different frames. Since D-NeRF represents the dynamic objects with the 4D flow fields and a 3D canonical representation, we separately factorize these two fields. First, for the 4D flow fields, we only adopt coarse level decomposition and factorize it into low-resolution feature planes:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.506, + 0.892, + 0.525 + ], + "angle": 0, + "content": "\\[\nf (x, y, z, t) = \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.894, + 0.608 + ], + "angle": 0, + "content": "Our coarse decomposition focuses more on the coarse and rigid motion, which improves the robustness of flow estimation and can achieve better disentanglement of shape and motion. Then for the 3D canonical representation, we adopt both coarse and fine level decomposition:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.617, + 0.892, + 0.635 + ], + "angle": 0, + "content": "\\[\nh (\\hat {x}, \\hat {y}, \\hat {z}) = \\pi_ {3} \\left(h ^ {H R}\\right) + \\pi_ {3} \\left(h ^ {L R}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Therefore, we can obtain 9 flow feature planes \\( T_{f} \\) for 4D flow fields and 6 canonical feature planes \\( T_{h} \\) for 3D canonical representation. The volume rendering in monocular cases is similar to multi-view cases. For a sampling point \\( p \\), we first obtain the point flow feature \\( F_{f} \\) using Eq. 6 with nine flow planes \\( T_{f} \\). Then we adopt the flow MLP \\( E_{f} \\) to predict the movement of the point:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.759, + 0.891, + 0.777 + ], + "angle": 0, + "content": "\\[\nE _ {f} \\left(F _ {p}, \\gamma (x, y, z, t)\\right) = (\\hat {x}, \\hat {y}, \\hat {z}). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.786, + 0.892, + 0.831 + ], + "angle": 0, + "content": "Then we obtain the point canonical feature \\( F_{c} \\) by querying the canonical feature planes \\( T_{h} \\). Take the LR feature \\( F_{c}^{LR} \\) for example:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.871 + ], + "angle": 0, + "content": "\\[\nF _ {c} ^ {L R} = \\pi_ {3} \\left(h ^ {L} R\\right) = \\oplus \\left(T _ {h _ {z}} ^ {L R} (x, y), T _ {h _ {y}} ^ {L R} (x, z), T _ {h _ {z}} ^ {L R} (x, y)\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Next, we feed canonical feature \\( F_{c}^{LR} \\) and the positional encoding of \\( (\\hat{x},\\hat{y},\\hat{z}) \\) into the geometry MLP \\( E_{g} \\) to obtain" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.326, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.871, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.501, + 0.893, + 0.529 + ], + "angle": 0, + "content": "Figure 4. Example results of our method. Space and time novel view rendering results from sparse-view fixed cameras. The top three results are from four front view cameras and the bottom is from 12 circular cameras." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.344, + 0.56 + ], + "angle": 0, + "content": "high-dimension feature \\(\\mathbf{f}\\) and density \\(\\sigma\\)" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.57, + 0.47, + 0.589 + ], + "angle": 0, + "content": "\\[\nE _ {g} \\left(F _ {c} ^ {L R}, \\gamma (\\hat {x}, \\hat {y}, \\hat {z})\\right) = (\\mathbf {f}, \\sigma). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Finally, we adopt Eq. 8 to predict color \\( c \\) for volume rendering and Eq. 9 for training color loss \\( L_{c} \\). We also add feature regularization loss \\( L_{r} \\) in Eq. 10 and surface constraint loss \\( L_{e} \\) in Eq. 11. The total training loss \\( L_{s} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.671, + 0.469, + 0.687 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\lambda_ {c} L _ {c} + \\lambda_ {r} L _ {r} + \\lambda_ {e} L _ {e}. \\tag {18}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.711, + 0.201, + 0.728 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Dataset. To evaluate the performance of our methods for multiview inputs, we build a sparse-view capture system with 6 forward-facing RGB cameras mounted on the borders of a \\(32^{\\prime \\prime}\\) Looking Glass 3D holographic display [2]. All cameras are synchronized and calibrated. Using this system, we capture multiple sequences of various challenging human motions, including dancing, thumbing up, waving hands, wearing hats and manipulating bags. We use 4 of them for reconstruction and rendering in all our experiments, while leaving the other two for quantitative evaluation. We also use three \\(360^{\\circ}\\) multiview full body sequences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.65 + ], + "angle": 0, + "content": "captured with 12 evenly spaced cameras on a camera ring for qualitative evaluation. For monocular evaluation, we use the synthetic dataset provided by D-NeRF [40] and select 3 scenes (\"lego\", \"standup\" and \"jumpingjacks\") from this dataset, with the numbers of training frames ranging from 50 to 200. More details about data collection and preprocessing can be found in the Supp.Mat.." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.893, + 0.816 + ], + "angle": 0, + "content": "Baselines. We mainly compare our method against the following state-of-the-art baselines that are most related to our work: D-NeRF [40], NeRF-T, TiNeuVox [12] and NeuS-T. Here, NeRF-T is our extension of vanilla NeRF [32] by introducing an additional time input, and NeuS-T is extended from NeuS [59] similarly. Among these baselines, D-NeRF and TiNeuVox represent the dynamic scenes through deforming a canonical one, while NeRF-T and NeuS-T directly learn a time-conditioned 4D radiance fields. TiNeuVox uses explicit voxel grids to accelerate network training, while others purely use MLPs to model the scene." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.725, + 0.848 + ], + "angle": 0, + "content": "5.1. Results and Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. We train our model for each individual sequences, and present some example results for novel view synthesis in Fig. 4 and the Supp.Mat.. The results" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16637" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.084, + 0.88, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.318, + 0.779, + 0.332 + ], + "angle": 0, + "content": "Figure 5. Comparison on monocular synthetic dataset against D-NeRF [40] and TiNeuVox [12]." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.339, + 0.88, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.585, + 0.893, + 0.613 + ], + "angle": 0, + "content": "Figure 6. Comparison on sparse-view real-world dataset against D-Nerf [40], TiNeuVox [12] and Neus-T [59]. Four of camera views from our datasets are used for reconstruction and rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.729 + ], + "angle": 0, + "content": "cover various body motions, clothing styles and accessories. As shown in Fig. 4, our method can render high-quality images for dynamic scenes and faithfully recover appearance details like the thin finger motions, semi-transparent silk, hand-object interaction, face expressions and cloth wrinkles. See our Supp Video for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Comparisons on monocular dynamic dataset. We first compare our method with the baselines on monocular synthetic dataset. Qualitative results are presented in Fig. 5. Compared to other methods, ours recovers more appearance details and generate less artifacts. The numeric results in Tab. 1 also prove that our method outperforms state-of-the-art methods in terms of rendering quality and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Comparisons on sparse view dataset. We then evaluate the performance of different methods for novel view synthesis given four camera views from our collected six camera-view real-world datasets. The remained two views are used" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.64, + 0.892, + 0.684 + ], + "angle": 0, + "content": "for quantitative evaluation. Results reported in Fig. 6 and Tab. 2 show again that our method performs better in accurate and high-quality appearance detail synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Comparisons of training efficiency. We compare the training time and model size for memory in Tabs. 3 and 4. Our method requires significantly less training time compared to the original D-NeRF, NeRF-T, and Neus-T in monocular and multi-view scenarios. Compared to Ours-NeRF-T and Ours-D-NeRF, the longer training time observed in our method is due to the additional computation for geometry smoothness terms, which reduces artifacts and enhances rendering quality. Compared to TiNeuVox, our method has a reduction in memory consumption and a clear improvement in rendering quality since Tensor4D enables the efficient decomposition of 4D fields at higher spatial resolutions (512\\(^3\\) vs. 256\\(^3\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16638" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.284, + 0.09, + 0.684, + 0.103 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on monocular synthetic datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.103, + 0.891, + 0.202 + ], + "angle": 0, + "content": "
MethodLegoStandupJumpingjacks
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]7.83e-321.260.8691.72e-14.63e-434.380.9892.18e-25.46e-433.370.9874.84e-2
NeRF-T3.89e-324.320.9041.55e-16.82e-431.440.9682.36e-25.99e-432.270.9795.37e-2
TiNeuVox [12]3.15e-325.140.9248.37e-22.90e-436.180.9862.02e-23.89e-434.760.9833.33e-2
Ours-NeRF-T3.10e-325.130.9221.24e-15.92e-432.380.9772.29e-25.62e-432.730.9805.16e-2
Ours-D-NeRF4.61e-323.370.8901.19e-12.82e-435.930.9812.07e-24.22e-434.100.9823.41e-2
Ours2.26e-326.710.9533.49e-22.50e-436.320.9831.74e-23.91e-434.430.9823.18e-2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.206, + 0.715, + 0.219 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on the six camera-view real-world datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.219, + 0.891, + 0.319 + ], + "angle": 0, + "content": "
MethodSequence1-thzEarphoneSequence3-yxd
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]3.13e-325.150.9100.1858.16e-320.920.8540.2564.84e-323.220.9370.147
TiNeuVox [12]5.72e-322.790.8320.2091.49e-218.550.7070.3198.18e-321.190.8160.233
Neus-T [59]4.95e-323.070.8870.1303.75e-324.220.8770.1842.20e-326.590.9450.099
Ours-NeRF-T5.85e-322.460.8420.1911.27e-219.130.8380.2185.02e-323.060.9140.141
Ours-D-NeRF4.74e-323.270.8640.1768.17e-321.070.8830.1984.06e-323.980.9260.130
Ours1.53e-328.270.9420.0843.20e-325.000.9030.1531.31e-328.830.9620.072
" + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.33, + 0.458, + 0.344 + ], + "angle": 0, + "content": "Table 3. Training time and memory report in monocular cases." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.346, + 0.445, + 0.459 + ], + "angle": 0, + "content": "
MethodLego
PSNRTimeIterations#Params
TiNeuVox25.1434min50k102M
D-NeRF21.2645h800k4.8M
Ours-D-NeRF23.3795min50k20M
NeRF-T24.3238h800k4.4M
Ours-NeRF-T25.1373min50k43M
Ours(6-planes)26.34135min50k17M
Ours26.71144min50k20M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.463, + 0.458, + 0.476 + ], + "angle": 0, + "content": "Table 4. Training time and memory report in multi-view cases." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.479, + 0.445, + 0.578 + ], + "angle": 0, + "content": "
MethodSequence1-thz
PSNRTimeIterations#Params
TiNeuVox22.7930min50k102M
D-NeRF25.1537h800k4.8M
Neus-T23.0745h800k5.0M
Ours-NeRF-T22.4668min50k43M
Ours(6-planes)27.86105min50k32M
Ours28.27117min50K43M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.581, + 0.47, + 0.607 + ], + "angle": 0, + "content": "Table 5. Ablation study of hierarchical decomposition and smoothness terms." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.611, + 0.451, + 0.735 + ], + "angle": 0, + "content": "
MethodLegoStandup
PSNRSSIMPSNRSSIM
Ours(w/o regular)26.490.94635.910.977
Ours(6-planes)26.440.94435.790.978
Ours26.710.95336.320.983
Sequence-thzSequence-earphone
PSNRSSIMPSNRSSIM
Ours(w/o regular)27.920.93224.750.885
Ours(6-planes)27.740.93424.390.889
Ours28.270.94225.000.903
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.743, + 0.227, + 0.759 + ], + "angle": 0, + "content": "5.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Regularization. We quantitatively ablate the regularization terms in our method. We implement two strong baselines \"Ours-NeRF-T\" and \"Ours-D-NeRF\", in which we directly apply our Tensor4D decomposition for NeRF-T and D-NeRF without the regularization terms. The results are reported in Tab. 1 and Tab. 2. Benefiting from our 4D decomposition, they achieve better performance than the original NeRF-T and D-NeRF. However, their rendering quality is still worse than our full method with regularization. In" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.392 + ], + "angle": 0, + "content": "addition, We ablate for TV regular smoothness terms (Ours vs. Ours(w/o regular)) and the results are reported in Tab. 5, which further validates the effectiveness of our smoothness terms for rendering quality enhancement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.392, + 0.892, + 0.467 + ], + "angle": 0, + "content": "Hierarchical Decomposition We quantitatively ablate our hierarchical decomposition with 6-plane decomposition. As shown in Tab 5, our method achieves superior performance in both multi-view and monocular cases, which validates the effectiveness of our hierarchical decomposition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.613, + 0.497 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Limitations. Since our method needs to decompose the 4D fields to several 2D feature planes, a pre-set bounding box of the scene is necessary. Therefore, it is difficult for our method to reconstruct backgrounds or objects which are out of the bounding box. Another limitation is our strong regularization terms. Though these terms benefit the robustness of our reconstruction under sparse views, they also limit our ability to handle challenging cases such as fluid and fog." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Conclusion. We presented Tensor4D, a new method for learning high-quality neural representation for dynamic scenes from sparse-view videos or even a monocular video. To capture the spatio-temporal information in a compact and memory-efficient manner, we propose to develop a novel hierarchical tri-projection decomposition method that models a 4D tensor with nine 2D feature planes. With proper design of training losses and regularization, our method provides an efficient yet effective solution to model the radiance fields of dynamic scenes. We believe our work can inspire future research towards low-cost, portable and immersive telepresence systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107 and No.61827805 and the Beijing Science and Technology Project (Z211100004021006)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16639" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.125, + 0.129 + ], + "angle": 0, + "content": "[1] 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.133, + 0.468, + 0.159 + ], + "angle": 0, + "content": "[2] The looking glass holographic display. https://lookingglassfactory.com/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.163, + 0.468, + 0.231 + ], + "angle": 0, + "content": "[3] Naveed Ahmed, Christian Theobalt, Petar Dobrev, Hans-Peter Seidel, and Sebastian Thrun. Robust fusion of dynamic shape and normal capture for high-quality reconstruction of time-varying geometry. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.468, + 0.299 + ], + "angle": 0, + "content": "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.348, + 0.468, + 0.401 + ], + "angle": 0, + "content": "[6] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.405, + 0.468, + 0.472 + ], + "angle": 0, + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.468, + 0.517 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.519, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[9] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. 34(4), jul 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.468, + 0.643 + ], + "angle": 0, + "content": "[10] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6), nov 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[11] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE International Conference on Computer Vision, pages 14304-14314, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[12] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In arXiv preprint arXiv:2205.15285, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5491–5500, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4d: Voxel for 4d novel view synthesis. In arXiv preprint arXiv:2205.14332, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In In Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[17] Paulo FU Gotardo, Tomas Simon, Yaser Sheikh, and Iain Matthews. Photogeometric scene flow for high-detail dynamic 3d reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 846-854, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[19] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Trans. Graph., 38(6), nov 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[20] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Dynamic shape capture using multi-view photometric stereo. In ACM SIGGRAPH Asia 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.596, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[23] T. Kanade, P. Rander, and P.J. Narayanan. Virtualized reality: constructing virtual worlds from real scenes. IEEE MultiMedia, 4(1):34-47, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[24] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3-d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] Jason Lawrence, Danb Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Trans. Graph., 40(6), dec 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Trans. Graph., 31(1), feb 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Tianye Li, Mira Slavcheva, Michael Zollhofer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16640" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[28] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In arXiv preprint, arXiv:2205.15723, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[29] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[30] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[33] Thomas Muller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. 41(4), jul 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[34] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[35] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[37] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[38] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higherdimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[40] Albert Pumarola, Enric Corona, and and Francesc Moreno-Noguer Gerard Pons-Moll. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[41] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[42] Chris Russell, Rui Yu, and Lourdes Agapito. Video pop-up: Monocular 3d reconstruction of dynamic scenes. In Computer Vision - ECCV 2014, pages 583-598, Cham, 2014. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[43] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 20154-20166. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[44] Ruizhi Shao, Liliang Chen, Zerong Zheng, Hongwen Zhang, Yuxiang Zhang, Han Huang, Yandong Guo, and Yebin Liu. Floren: Real-time high-quality human performance rendering via appearance flow using sparse rgb cameras. In SIGGRAPH Asia 2022 Conference Papers, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[45] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yanpei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[46] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. Diffustereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 702-720. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[47] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[48] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[49] Jonathan Starck and Adrian Hilton. Surface capture for performance-based animation. IEEE Computer Graphics and Applications, 27(3):21-31, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16641" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[53] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[54] Aparna Taneja, Luca Ballan, and Marc Pollefeys. Modeling dynamic scenes recorded with freely moving cameras. Computer Vision - ACCV 2010, page 613-626, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE International Conference on Computer Vision, pages 12959-12970, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[56] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. In arXiv preprint, arXiv:2105.05994, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[57] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. IButter: Neural Interactive Bullet Time Generator for Human Free-Viewpoint Rendering, page 4641–4650. Association for Computing Machinery, New York, NY, USA, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.429, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.652, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[61] Liwen Wu, Jae Yong Lee, Yu-Xiong Wang Anand Bhattachad, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and ap" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "pearance. Advances in Neural Information Processing Systems, 33, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[66] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5746-5756, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[67] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 40(6), dec 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[68] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7743-7753, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[69] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[70] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16642" + } + ] +] \ No newline at end of file diff --git a/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_origin.pdf b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..73f15d5f233bd6a2dd58948d9ff5106fc635ae08 --- /dev/null +++ b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/eeed7500-cf04-43cc-8f38-007884849085_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9378bf82b9d7fe4e1b02d9182f2a2c23c50b87f44e45af940bd0c70166dd5ac3 +size 6915976 diff --git a/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/full.md b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a1ea52ce73810327422b721839118b4ef9c69f34 --- /dev/null +++ b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/full.md @@ -0,0 +1,382 @@ +# Tensor4D : Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering + +Ruizhi Shao $^{1}$ , Zerong Zheng $^{1,2}$ , Hanzhang Tu $^{1}$ , Boning Liu $^{1}$ , Hongwen Zhang $^{1}$ , Yebin Liu $^{1}$ $^{1}$ Department of Automation, Tsinghua University $^{2}$ NNKosmos Technology + +![](images/e25a3aca3dd57916fda3911e09fe6f70a32d45c183c4ac8866fcb0881f82aeba.jpg) +(a) + +![](images/92b7b4a3a63a35de1019598d4318a137121a9c585bd60fabda35cfb0ad77b65d.jpg) +(b) + +![](images/c86ca73fc314a3ea32b333eaadcea2be0763d818d6c12218673aab1a3db2e52c.jpg) +(c) +Figure 1. Given 4 sparse static RGB camera views of a dynamic scene (a), our proposed Tensor4D decomposition enables multiview reconstruction to achieve fine-grained geometry reconstruction even on human fingers (b) and temporal-consistent novel view synthesis on a 3D holographic display (c,d,e). The 4 cameras are settled on four corners of the display. The proposed method demonstrates low-cost, portable and highly immersive telepresence experience. + +![](images/953d43cd16790621adae1b3dc19a758e9df275f158279bb9f728d8dfda42ff5f.jpg) +(d) + +![](images/f413ef76e9da69b86fe19d50875e1650a112f2de45461c536d419a6f44692ff0.jpg) +(e) + +# Abstract + +We present Tensor4D, an efficient yet effective approach to dynamic scene modeling. The key of our solution is an efficient 4D tensor decomposition method so that the dynamic scene can be directly represented as a 4D spatiotemporal tensor. To tackle the accompanying memory issue, we decompose the 4D tensor hierarchically by projecting it first into three time-aware volumes and then nine compact feature planes. In this way, spatial information over time can be simultaneously captured in a compact and memory-efficient manner. When applying Tensor4D for dynamic scene reconstruction and rendering, we further factorize the 4D fields to different scales in the sense that structural motions and dynamic detailed changes can be learned from coarse to fine. The effectiveness of our method is validated on both synthetic and real-world scenes. Extensive experiments show that our method is able to achieve high-quality dynamic reconstruction and rendering from sparse-view camera rigs or even a monocular camera. The code and dataset will be released at https://github.com/DSaurus/Tensor4D. + +# 1. Introduction + +High quality reconstruction and Photo-realistic rendering of a dynamic scene from a set of input images is nec + +essary for many applications such as AR/VR, 3D content production and entertainment. Traditional methods use classical mesh-based representation to reconstruct the dynamic scenes, which, unfortunately, are prone to produce reconstruction errors and rendering artifacts when the scene contains thin structures, specular surfaces and topological changes [9, 19, 23, 26, 49]. + +Recent advances in neural rendering approaches, which learn scene representations in the form of neural radiance fields (NeRF), have shown impressive novel view synthesis of general static scenes given only multi-view images [32]. They are immediately extended to dynamic scenes: some methods (e.g., NeRF-T) consider time as an additional input dimension to NeRF representation [56,62], while other methods (e.g., D-NeRF) disentangle a dynamic scene into a canonical radiance field and a dynamic motion field [11,28,37,40,55]. Either way, learning a 4D function is one of the main cornerstones. Unfortunately, directly using MLP to fit such a function often suffers from high time and computation cost, i.e., dozens of hours on high-end GPUs. + +In fact, the aforementioned limitation also exists in conventional NeRF-based methods for static scenes, and researchers have proposed to use discrete data structures like voxel grids [65] or triplanes [7] to accelerate NeRF training and rendering. However, these techniques are difficult to be extended to dynamic domains as introducing an additional time dimension will exponentially increase memory footprint, hindering them from modeling high-quality ap + +pearance details. + +In this work, we pursue a dynamic scene representation that also utilizes explicit feature grids to accelerate network training while avoiding huge memory consumption when introducing an additional time dimension. To this end, we bypass the construction of a high resolution 4D tensor; instead, we propose to model a 4D field using hierarchical tri-projection decomposition. Our decomposition method extends the tri-projection in EG3D [7]. It firstly project a full 4D field into three time-aware volumes, each of which is then further decomposed into three feature planes. In this way, we model the 4D field using only nine 2D feature planes, and we empirically find that although being highly compact, such a representation is powerful enough to represent dynamic scenes containing complex motions. Moreover, the usage of explicit data structure also allows us to design a coarse-to-fine strategy to further improve the performance of our method. + +By utilizing and factorizing an explicit 4D tensor, our method enables both efficient reconstruction and compact representation of dynamic scenes. Besides, the decomposition scheme also introduces implicit constraints on the representation since only low-rank tensors can be approximated by a small number of lower-dimensional components. Such a constraint can serve as an inherent regularization when the input observation is limited, e.g., under sparse and fixed cameras setting or even monocular inputs. In this paper, we first apply our method for sparse-view dynamic reconstruction by adopting our Tensor4D decomposition to time-conditioned radiance fields in "NeRF-T". In addition, our decomposition method can also be used for single-view dynamic reconstruction. This is achieved through decomposing both the 4D dynamic motion field and the canonical radiance field in "D-NeRF". With proper regularization, our system enables efficient and high-quality reconstruction of dynamic objects under both camera settings. + +# 2. Related Work + +Multiview Reconstruction and Rendering. There are various ways to capture and reconstruct the 3D dynamic scenes, involving methods based on silhouette [24, 54], stereo [30, 46], flow [17, 44, 70], segmentation [41, 42], and photometric [3, 20]. With RGBD cameras, real-time solutions like DynamicFusion [34] estimates the non-rigid deformations of a dynamic scene and integrates depth frames to reconstruct the geometry model in the canonical space. This method is later extended for telepresence and holographic communication [10, 25, 66]. However, these systems heavily rely on depth sensors to obtain accurate geometry. In contrast, our method can reconstruct and render dynamic scenes using sparse-view RGB cameras. + +In the past few years, neural implicit representations underwent rapid development and have been applied for multi- + +view reconstruction and rendering of static scenes. Some methods represent the geometry as the zero level-set of a neural network, and use differentiable surface rendering to optimize the network weights [35,47,64]. Given dense observation of an object, these methods are able to accurately recover its surface. NeRF [32], on the other hand, uses volume rendering to optimize the scene representation. Its simplicity and impressive results inspires a lot of following works, including in-the-wild reconstruction [31, 52], lighting and material estimation [4,48,67], generation [1,6,18, 22,43,53], human rendering [45,51,57,68,69] and so on. More recently, several methods unify surface and volumetric rendering, enabling accurate geometry reconstruction and high-quality novel view synthesis [36,59,63]. Compared to these static scene representations, we aim to enable free view synthesis of dynamic scenes from extremely sparse cameras. + +NeRF for Dynamic Scenes. Modeling scenes in 4D domain with time dimension included is a direct solution to extend NeRF for dynamic domains. Typical approaches include VideoNeRF [62], NeRFlow [11], DyNeRF [27], and DCT-NeRF [56]. Specifically, VideoNeRF [62] directly learns a spatiotemporal irradiance field from a single video and uses depth estimation to address the shape-motion ambiguities in monocular inputs, while NeRFlow [11] and DCT-NeRF [56] use point trajectory to regularize the network optimization. To deal with the limitation of topology-change modeling in deformation field, Park et al. [38] represent HyperNeRF which can lift NeRF to higher dimensions. + +Dynamic scenes can also be rendered by deforming the radiance field in the canonical space. For example, Nerfies [37] optimizes an additional continuous deformation field by warping each observed point into a canonical 5D NeRF. D-Nerf [40] and NR-Nerf [55] follow a similar framework, but take only monocular videos as training data. In addition, DeVRF [28] uses voxel-based representation instead of MLPs to model the 3D canonical space and the 4D deformation field. Using parametric body templates as the semantic prior, methods like Neural Body [39] and HumanNeRF [60] enable photo-realistic novel view synthesis of complex human performance. The discrepancy between the practical capture process and the existing experimental protocols for monocular videos has been shown in [16]. + +NeRF Acceleration. Numerous works emerge with the purpose of speeding up static NeRF using explicit data structures including feature maps, voxels and tensors. For instance, DVGO [50] achieves fast convergence through an explicit representation of a density voxel grid and a feature voxel grid. With the sparse voxel octree structure, NSVF [29] accelerates the novel view reconstruction by discarding the empty voxels in a coarse-to-fine manner. Similarly, Plenoxels [14] and PlenOctree [65] model a scene through a hierarchical 3D grid with spherical harmonics, which can realize an optimization with two orders of magnitude faster + +![](images/9901884b7286d67e7be05168283f994b5080c4338c3bbd4d85da1ed82e944656.jpg) +Figure 2. Illustration of our hierarchical tri-projection decomposition method. For a neural 4D field $f(x,y,z,t)$ , we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumes, which are then further projected onto nine 2D planes. + +than NeRF. In DIVeR [61], ray marching only finds a fixed number of hits on the voxel grid to accelerate volumetric rendering. Moreover, hashing encoding [33] and tensor decomposition [8] are also used as compact representations for NeRF acceleration. + +For dynamic scene modeling, DeVRF [28] enables fast non-rigid neural rendering with both 3D volumetric and 4D voxel field. In addition, V4D [15] introduce an effective conditional positional encoding for 4D data to realize fast novel view synthesis. TiNeuVox [12] represents scenes with identifiable explicit data structures and accelerates radiance fields modeling, while Wang et al. extended PlenOctrees [65] into free-view video rendering [58]. However, the low-res volumetric design of these works hinders the capacity of rendering high-quality images. In addition, there are also several concurrent works [5, 13, 21] adopt 6-plane decomposition for dynamic scenes. Compared with these methods, our hierarchical decomposition is more efficient to capture time variations and represent dynamic scenes. + +# 3. Method + +Building on prior work for spatio-temporal representations and tri-projection decomposition (Sec. 3.1), we propose a hierarchical tri-projection decomposition method (Sec. 3.2) and a coarse-to-fine strategy (Sec. 3.3), which allow us to learn a 4D field at a modest cost of training time and GPU memory. + +# 3.1. Preliminary + +Spatio-temporal 4D NeRF representation. To represent dynamic objects using neural radiance fields, a naive way is to condition the original neural radiance fields on the times-stamp [56, 62], which we term as NeRF-T. Mathematically, + +NeRF-T can be formulated as: + +$$ +\begin{array}{l} f (x, y, z, t) = (\mathbf {f}, \sigma), \tag {1} \\ g (\mathbf {f}, \theta , \phi) = \mathbf {c}, \\ \end{array} +$$ + +where $f$ is a dynamic implicit field that produces a high-dimensional feature $\mathbf{f}$ and a density value $\sigma$ for a point at position $(x, y, z)$ and time instance $t$ , and $g$ is a function that takes into account the viewing direction $(\theta, \phi)$ to predict the final RGB color. + +To achieve better disentanglement of shape and motion, some methods like D-NeRF [40] propose deformable neural radiance field which adopts a canonical 3D representation with the 4D flow fields: + +$$ +\begin{array}{l} f (x, y, z, t) = (\hat {x}, \hat {y}, \hat {z}), \tag {2} \\ g (\hat {x}, \hat {y}, \hat {z}, \theta , \phi) = (\mathbf {c}, \sigma), \\ \end{array} +$$ + +where $g$ is the radiance field in canonical configuration and $f$ is a scene flow field representing the mapping between the scene at time instant $t$ and the canonical space. + +From the above formulation one can easily observe that both NeRF-T and D-NeRF rely on modeling a 4D field $f(x,y,z,t)$ , i.e., the dynamic implicit field in NeRF-T and the flow field in D-NeRF. Existing methods mainly adopt MLPs to fit these 4D fields. Such an implicit neural representation does not have an explicit structure and requires extensive computation time for both training and rendering. Tri-projection Decomposition. The tri-projection decomposition is widely used in recent work including EG3D [7] and TensorRF [8] in order to accelerate the training and rendering process in MLP-based NeRF frameworks. Such decomposition factorizes a $n$ -dimensional tensor $V_{h}$ into three lower-dimensional $((n - 1)$ -D) tensors $V_{l}^{i} \{i = 1,2,3\}$ by projecting $V_{h}$ along the $x$ , $y$ and $z$ -axis respectively. For example, triplane-based decomposition proposed by EG3D projects the 3D tensor into three 2D feature planes. Compared to voxel-based representations, triplane representation effectively reduces the memory footprint and improves the performance of 3D generation and reconstruction. + +# 3.2. Hierarchical Tri-projection Decomposition + +Our goal is to design a dynamic scene representation with explicit feature grids to accelerate network training and volumetric rendering. However, directly constructing a 4D tensor costs a huge amount of memory and is unacceptable for the purpose of high-resolution rendering. Therefore, we propose a hierarchical triprojection decomposition to factorize the 4D tensor into several compact features, which reduces memory consumption by a large margin while preserving the capability of fitting 4D fields. + +Specifically, for a neural 4D field $f(x,y,z,t)$ , we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumetric tensors via tri- + +projection decomposition: + +$$ +\Pi_ {3} (f (x, y, z, t)) = \left\{f _ {z} (x, y, t), f _ {y} (x, z, t), f _ {x} (y, z, t) \right\}, \tag {3} +$$ + +where $\Pi_3$ denotes the projection operator. To further lower space complexity and enable high-resolution representation, we decompose each feature volume into three feature planes as: + +$$ +\pi_ {3} (f _ {z}) = \left\{f _ {z t} (x, y), f _ {z y} (x, t), f _ {z x} (y, t) \right\} +$$ + +$$ +\pi_ {3} (f _ {y}) = \left\{f _ {y t} (x, z), f _ {y z} (x, t), f _ {y x} (z, t) \right\} \tag {4} +$$ + +$$ +\pi_ {3} (f _ {x}) = \left\{f _ {x t} (y, z), f _ {x z} (y, t), f _ {x y} (z, t) \right\} +$$ + +where $\pi_3$ denotes the volume-to-plane tri-projection. In this way, we compactly represent a 4D field using 9 planes. Given any spatio-temporal coordinate $(x,y,z,t)$ , we can efficiently query its value in the 4D field by projecting it onto the planes and retrieving the corresponding value via bilinear interpolation. Fig 2 is an illustration of our decomposition method. Our hierarchical triprojection decomposition reduces the space complexity from $\mathcal{O}(n^4)$ to $\mathcal{O}(n^2)$ with $n$ being the spatial resolution of the grid, significantly lowering memory footprint without sacrificing representation power. + +Differences from 6-plane decomposition. Compared with 6-plane decomposition [5, 13, 21] which pairs the time dimension with only one spatial dimension $(xt,yt,zt)$ , our method first decomposes the 4D tensor in the spatial domain to obtain three time-aware volumes and then 9 decomposed planes. Note that the three volumes are decomposed independently, making the 9 planes different from each other. Specifically, when using a linear layer $g = \{\}$ to decompose the volume $f_{z}(x,y,t) = \{.,f_{zy}(x,t),\}.$ and the volume $f_{y}(x,z,t) = \{.,f_{yz}(x,t),\}.$ , the planes $f_{zy}(x,t)$ and $f_{yz}(x,t)$ behave differently since they are optimized independently to capture time variations in the volumes $f_{z}(x,y,t)$ and $f_{y}(x,z,t)$ , respectively. In this way, the 9 planes can leverage time-aware information in all possible combinations of the time and spatial dimensions $(xt,yt,zt,xyt,yzt,xzt,xyzt)$ . This property enables our representation to capture various dynamic information hierarchically at different levels of spatial dimensions. Therefore, our method is more efficient in representing dynamic scenes, where variations in time dimension are often intense, complex, and long-range. + +Our decomposition supports various types of 4D fields including the time-conditioned radiance field in NeRF-T and the 4D flow field in D-NeRF. In this paper, we present a dynamic radiance field decomposition method (Sec. 4.1) for multiview dynamic reconstruction as well as a 4D flow decomposition method (Sec. 4.2) for single-view setting. + +# 3.3. Coarse-to-fine Strategy + +To further improve the efficiency of our 4D decomposition, we propose an optional coarse-to-fine strategy to fac + +torize the 4D fields into different scales in different training phases. In coarse level, we adopt low-resolution feature planes $(128 \times 128)$ to decompose the 4D fields, which improves the robustness of the training process and achieves fast convergence. After coarse level training, we additionally use high-resolution feature planes $(512 \times 512)$ for 4D decomposition to represent dynamic details and achieve high-quality rendering. Specifically, we factorize the 4D fields into different scales: + +$$ +f (x, y, z, t) = \left\{\pi_ {3} \left(\Pi_ {3} \left(f ^ {L R}\right)\right), \pi_ {3} \left(\Pi_ {3} \left(f ^ {H R}\right)\right) \right\}, \tag {5} +$$ + +where $f^{LR}$ and $f^{HR}$ are the coarse-level and fine-level components of $f(x,y,z,t)$ , respectively. In the coarse level, the decomposed feature planes is low-resolution to represent coarse 3D structures and 4D dynamic changes. In the fine level, we adopt the high-resolution feature planes in each element to decompose the 4D fields, which focuses more on recovering dynamic details. + +# 4. Tensor4D for Dynamic Reconstruction + +As shown in Fig. 3, we apply our 4D tensor decomposition into the task of dynamic reconstruction with two inputs: + +1) Dynamic reconstruction under sparse and fixed cameras. For this setting, we factorize NeRF-T instead of D-NeRF using our proposed 4D decomposition, which we found more efficient and flexible to represent topologically-varying objects (Sec. 4.1). +2) Dynamic reconstruction using monocular camera. For this setting, we separately decompose the 4D flow fields and the 3D canonical representation since it ensures appearance consistency across different frames and achieves more robust performance (Sec. 4.2). + +# 4.1. Multi-view Reconstruction + +In this section, we present our dynamic reconstruction system under sparse multi-view setting. The goal of our system is 1) efficient and high-quality dynamic reconstruction with low memory and time cost, and 2) robust reconstruction under sparse and fixed camera setting. + +To this end, we adopt our 4D decomposition (Eq. 5) to factorize the NeRF-T representation with the coarse-to-fine strategy. Specifically, we can obtain nine low-resolution feature planes $T^{LR}$ and nine high-resolution feature planes $T^{HR}$ to represent the 4D NeRF-T fields after our 4D decomposition. Then for volume rendering, when sampling a point $p$ at $(x,y,z,t)$ in the ray $r$ with direction of $(\theta ,\phi)$ , we first query the feature $F_{p}$ of $p$ in both the LR and the HR feature planes. Take the LR features for example: + +$$ +\begin{array}{l} F _ {p} ^ {L R} (x, y, z, t) = \pi_ {3} \left(\Pi_ {3} \left(f ^ {L R}\right)\right) \\ = \oplus \left(\pi_ {3} \left(f _ {z} ^ {L R}\right), \pi_ {3} \left(f _ {y} ^ {L R}\right), \pi_ {3} \left(f _ {x} ^ {L R}\right)\right) \tag {6} \\ = \oplus (\oplus (T _ {z t} ^ {L R} (x, y), T _ {z y} ^ {L R} (x, t), T _ {z x} ^ {L R} (y, t)), \dots), \\ \end{array} +$$ + +![](images/a3bdc139b699f7c24c0f5bae50f399896add823a58f9586da5d036be18534439.jpg) +(a) + +![](images/d0f8f3c10739fd181a473aaa5cbad2e9abf03bffb760a0cfae6ea6412220e2f4.jpg) +(b) +Figure 3. The framework of Tensor4D for multi-view and monocular reconstruction. a). Tensor4D for multi-view reconstruction. The 4D NeRF-T fields are separately factorized by the LR and HR feature planes. b). Tensor4D for monocular reconstruction. The 4D flow fields are factorized by the LR feature plane for better disentanglement of shape and motion. The 3D canonical representation is factorized by three LR and HR feature planes. + +where we concatenate all features queried in the nine LR planes as the final LR feature of the sampling point $p$ . Then we concatenate the LR and HR features with the positional encoding of $(x, y, z, t)$ into the geometry MLP $E_{g}$ to obtain the density $\sigma$ and high-dimension feature $\mathbf{f}$ : + +$$ +E _ {g} \left(F _ {p} ^ {L R}, F _ {p} ^ {H R}, \gamma (x, y, z, t)\right) = (\mathbf {f}, \sigma), \tag {7} +$$ + +where $\gamma (\cdot)$ is the positional encoding function. Next, we concatenate the high-dimension feature $\mathbf{f}$ with the positional encoding of $(\theta ,\phi)$ and feed it into the color MLP: + +$$ +E _ {c} (\mathbf {f}, \gamma (\theta , \phi)) = c \tag {8} +$$ + +In this way, we can render the images through volume rendering and adopt color loss to train our decomposed feature planes $T$ and neural network $E_{c}$ and $E_{g}$ : + +$$ +L _ {c} = \left\| C _ {i, j} - C _ {i, j} ^ {*} \right\| \tag {9} +$$ + +With our coarse-to-fine design, our method can recover high-fidelity dynamic details effectively and efficiently. + +To achieve robust dynamic reconstruction under sparse multi-view setting, we further adopt regularization for all decomposed feature planes $T$ : + +$$ +L _ {r} = \sum_ {T} \sum_ {i, j} \sqrt {\left(T _ {i + 1 , j} - T _ {i , j}\right) ^ {2} + \left(T _ {i , j + 1} - T _ {i , j}\right) ^ {2}}, \tag {10} +$$ + +where $L_{r}$ is the TV loss for each feature plane $T$ to keep their sparsity. To regularize the geometry, we also introduce surface constraint into volume rendering. Specifically, we adopt the SDF as the base geometry representation and follow NeuS [59] to render the SDF field. Then we add surface constraint loss to enforce a smooth surface: + +$$ +L _ {e} = \left\| \left\| \nabla s (x, y, z, t) \right\| _ {2} - 1 \right\| _ {2} \tag {11} +$$ + +The final training loss $L_{m}$ is the regularization loss $L_{r}, L_{e}$ and the color loss $L_{c}$ : + +$$ +L _ {m} = \lambda_ {r} L _ {r} + \lambda_ {e} L _ {e} + \lambda_ {c} L _ {c} \tag {12} +$$ + +# 4.2. Monocular Reconstruction + +Different from the sparse view setting, we adopt 4D decomposition for D-NeRF in monocular capture cases. This is because monocular setting is much more ill-posed than sparse-view inputs, and the explicit disentanglement of appearance and motions can guarantee the consistency across different frames. Since D-NeRF represents the dynamic objects with the 4D flow fields and a 3D canonical representation, we separately factorize these two fields. First, for the 4D flow fields, we only adopt coarse level decomposition and factorize it into low-resolution feature planes: + +$$ +f (x, y, z, t) = \pi_ {3} \left(\Pi_ {3} \left(f ^ {L R}\right)\right). \tag {13} +$$ + +Our coarse decomposition focuses more on the coarse and rigid motion, which improves the robustness of flow estimation and can achieve better disentanglement of shape and motion. Then for the 3D canonical representation, we adopt both coarse and fine level decomposition: + +$$ +h (\hat {x}, \hat {y}, \hat {z}) = \pi_ {3} \left(h ^ {H R}\right) + \pi_ {3} \left(h ^ {L R}\right). \tag {14} +$$ + +Therefore, we can obtain 9 flow feature planes $T_{f}$ for 4D flow fields and 6 canonical feature planes $T_{h}$ for 3D canonical representation. The volume rendering in monocular cases is similar to multi-view cases. For a sampling point $p$ , we first obtain the point flow feature $F_{f}$ using Eq. 6 with nine flow planes $T_{f}$ . Then we adopt the flow MLP $E_{f}$ to predict the movement of the point: + +$$ +E _ {f} \left(F _ {p}, \gamma (x, y, z, t)\right) = (\hat {x}, \hat {y}, \hat {z}). \tag {15} +$$ + +Then we obtain the point canonical feature $F_{c}$ by querying the canonical feature planes $T_{h}$ . Take the LR feature $F_{c}^{LR}$ for example: + +$$ +F _ {c} ^ {L R} = \pi_ {3} \left(h ^ {L} R\right) = \oplus \left(T _ {h _ {z}} ^ {L R} (x, y), T _ {h _ {y}} ^ {L R} (x, z), T _ {h _ {z}} ^ {L R} (x, y)\right). \tag {16} +$$ + +Next, we feed canonical feature $F_{c}^{LR}$ and the positional encoding of $(\hat{x},\hat{y},\hat{z})$ into the geometry MLP $E_{g}$ to obtain + +![](images/772e007bb5bd11c0beaecf189bd005bec28fcbb2b10d1cb0d75a3666dd5c8522.jpg) +Figure 4. Example results of our method. Space and time novel view rendering results from sparse-view fixed cameras. The top three results are from four front view cameras and the bottom is from 12 circular cameras. + +![](images/d1a23377c08547516a08abb86d7e4c73f70ee98eaf1641b064c6ee641c65a4e6.jpg) + +high-dimension feature $\mathbf{f}$ and density $\sigma$ + +$$ +E _ {g} \left(F _ {c} ^ {L R}, \gamma (\hat {x}, \hat {y}, \hat {z})\right) = (\mathbf {f}, \sigma). \tag {17} +$$ + +Finally, we adopt Eq. 8 to predict color $c$ for volume rendering and Eq. 9 for training color loss $L_{c}$ . We also add feature regularization loss $L_{r}$ in Eq. 10 and surface constraint loss $L_{e}$ in Eq. 11. The total training loss $L_{s}$ is: + +$$ +L _ {s} = \lambda_ {c} L _ {c} + \lambda_ {r} L _ {r} + \lambda_ {e} L _ {e}. \tag {18} +$$ + +# 5. Experiment + +Dataset. To evaluate the performance of our methods for multiview inputs, we build a sparse-view capture system with 6 forward-facing RGB cameras mounted on the borders of a $32^{\prime \prime}$ Looking Glass 3D holographic display [2]. All cameras are synchronized and calibrated. Using this system, we capture multiple sequences of various challenging human motions, including dancing, thumbing up, waving hands, wearing hats and manipulating bags. We use 4 of them for reconstruction and rendering in all our experiments, while leaving the other two for quantitative evaluation. We also use three $360^{\circ}$ multiview full body sequences + +captured with 12 evenly spaced cameras on a camera ring for qualitative evaluation. For monocular evaluation, we use the synthetic dataset provided by D-NeRF [40] and select 3 scenes ("lego", "standup" and "jumpingjacks") from this dataset, with the numbers of training frames ranging from 50 to 200. More details about data collection and preprocessing can be found in the Supp.Mat.. + +Baselines. We mainly compare our method against the following state-of-the-art baselines that are most related to our work: D-NeRF [40], NeRF-T, TiNeuVox [12] and NeuS-T. Here, NeRF-T is our extension of vanilla NeRF [32] by introducing an additional time input, and NeuS-T is extended from NeuS [59] similarly. Among these baselines, D-NeRF and TiNeuVox represent the dynamic scenes through deforming a canonical one, while NeRF-T and NeuS-T directly learn a time-conditioned 4D radiance fields. TiNeuVox uses explicit voxel grids to accelerate network training, while others purely use MLPs to model the scene. + +# 5.1. Results and Comparison + +Qualitative Results. We train our model for each individual sequences, and present some example results for novel view synthesis in Fig. 4 and the Supp.Mat.. The results + +![](images/a14f9d797e8a62129d4ecb3955caf376d3547580a7a52147704965d4b2174ccf.jpg) +Figure 5. Comparison on monocular synthetic dataset against D-NeRF [40] and TiNeuVox [12]. + +![](images/acf46aa699b1d62650b438d808f1141b69a1759a158d82ff834a2149bc844d31.jpg) +Figure 6. Comparison on sparse-view real-world dataset against D-Nerf [40], TiNeuVox [12] and Neus-T [59]. Four of camera views from our datasets are used for reconstruction and rendering. + +cover various body motions, clothing styles and accessories. As shown in Fig. 4, our method can render high-quality images for dynamic scenes and faithfully recover appearance details like the thin finger motions, semi-transparent silk, hand-object interaction, face expressions and cloth wrinkles. See our Supp Video for better visualization. + +Comparisons on monocular dynamic dataset. We first compare our method with the baselines on monocular synthetic dataset. Qualitative results are presented in Fig. 5. Compared to other methods, ours recovers more appearance details and generate less artifacts. The numeric results in Tab. 1 also prove that our method outperforms state-of-the-art methods in terms of rendering quality and accuracy. + +Comparisons on sparse view dataset. We then evaluate the performance of different methods for novel view synthesis given four camera views from our collected six camera-view real-world datasets. The remained two views are used + +for quantitative evaluation. Results reported in Fig. 6 and Tab. 2 show again that our method performs better in accurate and high-quality appearance detail synthesis. + +Comparisons of training efficiency. We compare the training time and model size for memory in Tabs. 3 and 4. Our method requires significantly less training time compared to the original D-NeRF, NeRF-T, and Neus-T in monocular and multi-view scenarios. Compared to Ours-NeRF-T and Ours-D-NeRF, the longer training time observed in our method is due to the additional computation for geometry smoothness terms, which reduces artifacts and enhances rendering quality. Compared to TiNeuVox, our method has a reduction in memory consumption and a clear improvement in rendering quality since Tensor4D enables the efficient decomposition of 4D fields at higher spatial resolutions (512 $^3$ vs. 256 $^3$ ). + +Table 1. Quantitative comparisons on monocular synthetic datasets. + +
MethodLegoStandupJumpingjacks
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]7.83e-321.260.8691.72e-14.63e-434.380.9892.18e-25.46e-433.370.9874.84e-2
NeRF-T3.89e-324.320.9041.55e-16.82e-431.440.9682.36e-25.99e-432.270.9795.37e-2
TiNeuVox [12]3.15e-325.140.9248.37e-22.90e-436.180.9862.02e-23.89e-434.760.9833.33e-2
Ours-NeRF-T3.10e-325.130.9221.24e-15.92e-432.380.9772.29e-25.62e-432.730.9805.16e-2
Ours-D-NeRF4.61e-323.370.8901.19e-12.82e-435.930.9812.07e-24.22e-434.100.9823.41e-2
Ours2.26e-326.710.9533.49e-22.50e-436.320.9831.74e-23.91e-434.430.9823.18e-2
+ +Table 2. Quantitative comparisons on the six camera-view real-world datasets. + +
MethodSequence1-thzEarphoneSequence3-yxd
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]3.13e-325.150.9100.1858.16e-320.920.8540.2564.84e-323.220.9370.147
TiNeuVox [12]5.72e-322.790.8320.2091.49e-218.550.7070.3198.18e-321.190.8160.233
Neus-T [59]4.95e-323.070.8870.1303.75e-324.220.8770.1842.20e-326.590.9450.099
Ours-NeRF-T5.85e-322.460.8420.1911.27e-219.130.8380.2185.02e-323.060.9140.141
Ours-D-NeRF4.74e-323.270.8640.1768.17e-321.070.8830.1984.06e-323.980.9260.130
Ours1.53e-328.270.9420.0843.20e-325.000.9030.1531.31e-328.830.9620.072
+ +Table 3. Training time and memory report in monocular cases. + +
MethodLego
PSNRTimeIterations#Params
TiNeuVox25.1434min50k102M
D-NeRF21.2645h800k4.8M
Ours-D-NeRF23.3795min50k20M
NeRF-T24.3238h800k4.4M
Ours-NeRF-T25.1373min50k43M
Ours(6-planes)26.34135min50k17M
Ours26.71144min50k20M
+ +Table 4. Training time and memory report in multi-view cases. + +
MethodSequence1-thz
PSNRTimeIterations#Params
TiNeuVox22.7930min50k102M
D-NeRF25.1537h800k4.8M
Neus-T23.0745h800k5.0M
Ours-NeRF-T22.4668min50k43M
Ours(6-planes)27.86105min50k32M
Ours28.27117min50K43M
+ +Table 5. Ablation study of hierarchical decomposition and smoothness terms. + +
MethodLegoStandup
PSNRSSIMPSNRSSIM
Ours(w/o regular)26.490.94635.910.977
Ours(6-planes)26.440.94435.790.978
Ours26.710.95336.320.983
Sequence-thzSequence-earphone
PSNRSSIMPSNRSSIM
Ours(w/o regular)27.920.93224.750.885
Ours(6-planes)27.740.93424.390.889
Ours28.270.94225.000.903
+ +# 5.2. Ablation study + +Regularization. We quantitatively ablate the regularization terms in our method. We implement two strong baselines "Ours-NeRF-T" and "Ours-D-NeRF", in which we directly apply our Tensor4D decomposition for NeRF-T and D-NeRF without the regularization terms. The results are reported in Tab. 1 and Tab. 2. Benefiting from our 4D decomposition, they achieve better performance than the original NeRF-T and D-NeRF. However, their rendering quality is still worse than our full method with regularization. In + +addition, We ablate for TV regular smoothness terms (Ours vs. Ours(w/o regular)) and the results are reported in Tab. 5, which further validates the effectiveness of our smoothness terms for rendering quality enhancement. + +Hierarchical Decomposition We quantitatively ablate our hierarchical decomposition with 6-plane decomposition. As shown in Tab 5, our method achieves superior performance in both multi-view and monocular cases, which validates the effectiveness of our hierarchical decomposition. + +# 6. Discussion + +Limitations. Since our method needs to decompose the 4D fields to several 2D feature planes, a pre-set bounding box of the scene is necessary. Therefore, it is difficult for our method to reconstruct backgrounds or objects which are out of the bounding box. Another limitation is our strong regularization terms. Though these terms benefit the robustness of our reconstruction under sparse views, they also limit our ability to handle challenging cases such as fluid and fog. + +Conclusion. We presented Tensor4D, a new method for learning high-quality neural representation for dynamic scenes from sparse-view videos or even a monocular video. To capture the spatio-temporal information in a compact and memory-efficient manner, we propose to develop a novel hierarchical tri-projection decomposition method that models a 4D tensor with nine 2D feature planes. With proper design of training losses and regularization, our method provides an efficient yet effective solution to model the radiance fields of dynamic scenes. We believe our work can inspire future research towards low-cost, portable and immersive telepresence systems. + +Acknowledgements. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107 and No.61827805 and the Beijing Science and Technology Project (Z211100004021006). + +# References + +[1] 2 +[2] The looking glass holographic display. https://lookingglassfactory.com/.6 +[3] Naveed Ahmed, Christian Theobalt, Petar Dobrev, Hans-Peter Seidel, and Sebastian Thrun. Robust fusion of dynamic shape and normal capture for high-quality reconstruction of time-varying geometry. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2 +[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3, 4 +[6] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In arXiv, 2020. 2 +[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 3 +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 3 +[9] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. 34(4), jul 2015. 1 +[10] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6), nov 2017. 2 +[11] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE International Conference on Computer Vision, pages 14304-14314, 2021. 1, 2 +[12] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In arXiv preprint arXiv:2205.15285, 2022. 3, 6, 7, 8 +[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 4 +[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5491–5500, 2022. 2 + +[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4d: Voxel for 4d novel view synthesis. In arXiv preprint arXiv:2205.14332, 2022. 3 +[16] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In In Neural Information Processing Systems (NeurIPS), 2022. 2 +[17] Paulo FU Gotardo, Tomas Simon, Yaser Sheikh, and Iain Matthews. Photogeometric scene flow for high-detail dynamic 3d reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 846-854, 2015. 2 +[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2 +[19] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Trans. Graph., 38(6), nov 2019. 1 +[20] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Dynamic shape capture using multi-view photometric stereo. In ACM SIGGRAPH Asia 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery. 2 +[21] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 3, 4 +[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2 +[23] T. Kanade, P. Rander, and P.J. Narayanan. Virtualized reality: constructing virtual worlds from real scenes. IEEE MultiMedia, 4(1):34-47, 1997. 1 +[24] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3-d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2 +[25] Jason Lawrence, Danb Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Trans. Graph., 40(6), dec 2021. 2 +[26] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Trans. Graph., 31(1), feb 2012. 1 +[27] Tianye Li, Mira Slavcheva, Michael Zollhofer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view + +video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2 +[28] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In arXiv preprint, arXiv:2205.15723, 2022. 1, 2, 3 +[29] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2 +[30] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2 +[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2 +[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 6 +[33] Thomas Muller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. 41(4), jul 2022. 3 +[34] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2 +[35] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision (ICCV), 2021. 2 +[37] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2 +[38] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higherdimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2 +[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2 + +[40] Albert Pumarola, Enric Corona, and and Francesc Moreno-Noguer Gerard Pons-Moll. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 3, 6, 7, 8 +[41] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2 +[42] Chris Russell, Rui Yu, and Lourdes Agapito. Video pop-up: Monocular 3d reconstruction of dynamic scenes. In Computer Vision - ECCV 2014, pages 583-598, Cham, 2014. Springer International Publishing. 2 +[43] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 20154-20166. Curran Associates, Inc., 2020. 2 +[44] Ruizhi Shao, Liliang Chen, Zerong Zheng, Hongwen Zhang, Yuxiang Zhang, Han Huang, Yandong Guo, and Yebin Liu. Floren: Real-time high-quality human performance rendering via appearance flow using sparse rgb cameras. In SIGGRAPH Asia 2022 Conference Papers, pages 1-10, 2022. 2 +[45] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yanpei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In CVPR, 2022. 2 +[46] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. Diffustereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 702-720. Springer, 2022. 2 +[47] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2 +[48] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2 +[49] Jonathan Starck and Adrian Hilton. Surface capture for performance-based animation. IEEE Computer Graphics and Applications, 27(3):21-31, 2007. 1 +[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2 +[52] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural + +3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 2 +[53] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022. 2 +[54] Aparna Taneja, Luca Ballan, and Marc Pollefeys. Modeling dynamic scenes recorded with freely moving cameras. Computer Vision - ACCV 2010, page 613-626, 2011. 2 +[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE International Conference on Computer Vision, pages 12959-12970, 2021. 1, 2 +[56] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. In arXiv preprint, arXiv:2105.05994, 2021. 1, 2, 3 +[57] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. IButter: Neural Interactive Bullet Time Generator for Human Free-Viewpoint Rendering, page 4641–4650. Association for Computing Machinery, New York, NY, USA, 2021. 2 +[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 3 +[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 5, 6, 7, 8 +[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 2 +[61] Liwen Wu, Jae Yong Lee, Yu-Xiong Wang Anand Bhattachad, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 3 +[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 1, 2, 3 +[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 2 +[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and ap + +pearance. Advances in Neural Information Processing Systems, 33, 2020. 2 +[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2, 3 +[66] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5746-5756, 2021. 2 +[67] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 40(6), dec 2021. 2 +[68] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7743-7753, June 2022. 2 +[69] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 2 +[70] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2 \ No newline at end of file diff --git a/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/images.zip b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1f9ee62aa875345197b728f38e3946615a08a18a --- /dev/null +++ b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a38905c85bdd7b204fd690bb8c77214dcfb3b1e6e510fc7deba94dc2923ae4ca +size 754837 diff --git a/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/layout.json b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..acba6426cfd4fd19ba9657b850405e7018ffe6b5 --- /dev/null +++ b/2023/Tensor4D_ Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering/layout.json @@ -0,0 +1,9799 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 521, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 521, + 138 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 521, + 138 + ], + "type": "text", + "content": "Tensor4D : Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "spans": [ + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": "Ruizhi Shao" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": ", Zerong Zheng" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": ", Hanzhang Tu" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": ", Boning Liu" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": ", Hongwen Zhang" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": ", Yebin Liu" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University " + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 159, + 516, + 190 + ], + "type": "text", + "content": "NNKosmos Technology" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 222, + 140, + 310 + ], + "blocks": [ + { + "bbox": [ + 52, + 222, + 140, + 310 + ], + "lines": [ + { + "bbox": [ + 52, + 222, + 140, + 310 + ], + "spans": [ + { + "bbox": [ + 52, + 222, + 140, + 310 + ], + "type": "image", + "image_path": "e25a3aca3dd57916fda3911e09fe6f70a32d45c183c4ac8866fcb0881f82aeba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 313, + 101, + 322 + ], + "lines": [ + { + "bbox": [ + 90, + 313, + 101, + 322 + ], + "spans": [ + { + "bbox": [ + 90, + 313, + 101, + 322 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 141, + 222, + 222, + 310 + ], + "blocks": [ + { + "bbox": [ + 141, + 222, + 222, + 310 + ], + "lines": [ + { + "bbox": [ + 141, + 222, + 222, + 310 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 222, + 310 + ], + "type": "image", + "image_path": "92b7b4a3a63a35de1019598d4318a137121a9c585bd60fabda35cfb0ad77b65d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 313, + 186, + 323 + ], + "lines": [ + { + "bbox": [ + 175, + 313, + 186, + 323 + ], + "spans": [ + { + "bbox": [ + 175, + 313, + 186, + 323 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 223, + 325, + 309 + ], + "blocks": [ + { + "bbox": [ + 223, + 223, + 325, + 309 + ], + "lines": [ + { + "bbox": [ + 223, + 223, + 325, + 309 + ], + "spans": [ + { + "bbox": [ + 223, + 223, + 325, + 309 + ], + "type": "image", + "image_path": "c86ca73fc314a3ea32b333eaadcea2be0763d818d6c12218673aab1a3db2e52c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 313, + 279, + 323 + ], + "lines": [ + { + "bbox": [ + 269, + 313, + 279, + 323 + ], + "spans": [ + { + "bbox": [ + 269, + 313, + 279, + 323 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 326, + 546, + 369 + ], + "lines": [ + { + "bbox": [ + 45, + 326, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 45, + 326, + 546, + 369 + ], + "type": "text", + "content": "Figure 1. Given 4 sparse static RGB camera views of a dynamic scene (a), our proposed Tensor4D decomposition enables multiview reconstruction to achieve fine-grained geometry reconstruction even on human fingers (b) and temporal-consistent novel view synthesis on a 3D holographic display (c,d,e). The 4 cameras are settled on four corners of the display. The proposed method demonstrates low-cost, portable and highly immersive telepresence experience." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 223, + 430, + 309 + ], + "blocks": [ + { + "bbox": [ + 326, + 223, + 430, + 309 + ], + "lines": [ + { + "bbox": [ + 326, + 223, + 430, + 309 + ], + "spans": [ + { + "bbox": [ + 326, + 223, + 430, + 309 + ], + "type": "image", + "image_path": "953d43cd16790621adae1b3dc19a758e9df275f158279bb9f728d8dfda42ff5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 313, + 383, + 323 + ], + "lines": [ + { + "bbox": [ + 372, + 313, + 383, + 323 + ], + "spans": [ + { + "bbox": [ + 372, + 313, + 383, + 323 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 431, + 223, + 541, + 309 + ], + "blocks": [ + { + "bbox": [ + 431, + 223, + 541, + 309 + ], + "lines": [ + { + "bbox": [ + 431, + 223, + 541, + 309 + ], + "spans": [ + { + "bbox": [ + 431, + 223, + 541, + 309 + ], + "type": "image", + "image_path": "f413ef76e9da69b86fe19d50875e1650a112f2de45461c536d419a6f44692ff0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 313, + 490, + 323 + ], + "lines": [ + { + "bbox": [ + 480, + 313, + 490, + 323 + ], + "spans": [ + { + "bbox": [ + 480, + 313, + 490, + 323 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 380, + 192, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 380, + 192, + 392 + ], + "spans": [ + { + "bbox": [ + 143, + 380, + 192, + 392 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 406, + 290, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 406, + 290, + 643 + ], + "spans": [ + { + "bbox": [ + 45, + 406, + 290, + 643 + ], + "type": "text", + "content": "We present Tensor4D, an efficient yet effective approach to dynamic scene modeling. The key of our solution is an efficient 4D tensor decomposition method so that the dynamic scene can be directly represented as a 4D spatiotemporal tensor. To tackle the accompanying memory issue, we decompose the 4D tensor hierarchically by projecting it first into three time-aware volumes and then nine compact feature planes. In this way, spatial information over time can be simultaneously captured in a compact and memory-efficient manner. When applying Tensor4D for dynamic scene reconstruction and rendering, we further factorize the 4D fields to different scales in the sense that structural motions and dynamic detailed changes can be learned from coarse to fine. The effectiveness of our method is validated on both synthetic and real-world scenes. Extensive experiments show that our method is able to achieve high-quality dynamic reconstruction and rendering from sparse-view camera rigs or even a monocular camera. The code and dataset will be released at https://github.com/DSaurus/Tensor4D." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "High quality reconstruction and Photo-realistic rendering of a dynamic scene from a set of input images is nec" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 380, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 464 + ], + "type": "text", + "content": "essary for many applications such as AR/VR, 3D content production and entertainment. Traditional methods use classical mesh-based representation to reconstruct the dynamic scenes, which, unfortunately, are prone to produce reconstruction errors and rendering artifacts when the scene contains thin structures, specular surfaces and topological changes [9, 19, 23, 26, 49]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 464, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 617 + ], + "type": "text", + "content": "Recent advances in neural rendering approaches, which learn scene representations in the form of neural radiance fields (NeRF), have shown impressive novel view synthesis of general static scenes given only multi-view images [32]. They are immediately extended to dynamic scenes: some methods (e.g., NeRF-T) consider time as an additional input dimension to NeRF representation [56,62], while other methods (e.g., D-NeRF) disentangle a dynamic scene into a canonical radiance field and a dynamic motion field [11,28,37,40,55]. Either way, learning a 4D function is one of the main cornerstones. Unfortunately, directly using MLP to fit such a function often suffers from high time and computation cost, i.e., dozens of hours on high-end GPUs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 619, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 547, + 714 + ], + "type": "text", + "content": "In fact, the aforementioned limitation also exists in conventional NeRF-based methods for static scenes, and researchers have proposed to use discrete data structures like voxel grids [65] or triplanes [7] to accelerate NeRF training and rendering. However, these techniques are difficult to be extended to dynamic domains as introducing an additional time dimension will exponentially increase memory footprint, hindering them from modeling high-quality ap" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16632" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 116, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 116, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 116, + 83 + ], + "type": "text", + "content": "pearance details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 85, + 286, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 85, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 85, + 286, + 285 + ], + "type": "text", + "content": "In this work, we pursue a dynamic scene representation that also utilizes explicit feature grids to accelerate network training while avoiding huge memory consumption when introducing an additional time dimension. To this end, we bypass the construction of a high resolution 4D tensor; instead, we propose to model a 4D field using hierarchical tri-projection decomposition. Our decomposition method extends the tri-projection in EG3D [7]. It firstly project a full 4D field into three time-aware volumes, each of which is then further decomposed into three feature planes. In this way, we model the 4D field using only nine 2D feature planes, and we empirically find that although being highly compact, such a representation is powerful enough to represent dynamic scenes containing complex motions. Moreover, the usage of explicit data structure also allows us to design a coarse-to-fine strategy to further improve the performance of our method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 288, + 286, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 288, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 49, + 288, + 286, + 499 + ], + "type": "text", + "content": "By utilizing and factorizing an explicit 4D tensor, our method enables both efficient reconstruction and compact representation of dynamic scenes. Besides, the decomposition scheme also introduces implicit constraints on the representation since only low-rank tensors can be approximated by a small number of lower-dimensional components. Such a constraint can serve as an inherent regularization when the input observation is limited, e.g., under sparse and fixed cameras setting or even monocular inputs. In this paper, we first apply our method for sparse-view dynamic reconstruction by adopting our Tensor4D decomposition to time-conditioned radiance fields in \"NeRF-T\". In addition, our decomposition method can also be used for single-view dynamic reconstruction. This is achieved through decomposing both the 4D dynamic motion field and the canonical radiance field in \"D-NeRF\". With proper regularization, our system enables efficient and high-quality reconstruction of dynamic objects under both camera settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 514, + 132, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 132, + 526 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 132, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 534, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 286, + 688 + ], + "type": "text", + "content": "Multiview Reconstruction and Rendering. There are various ways to capture and reconstruct the 3D dynamic scenes, involving methods based on silhouette [24, 54], stereo [30, 46], flow [17, 44, 70], segmentation [41, 42], and photometric [3, 20]. With RGBD cameras, real-time solutions like DynamicFusion [34] estimates the non-rigid deformations of a dynamic scene and integrates depth frames to reconstruct the geometry model in the canonical space. This method is later extended for telepresence and holographic communication [10, 25, 66]. However, these systems heavily rely on depth sensors to obtain accurate geometry. In contrast, our method can reconstruct and render dynamic scenes using sparse-view RGB cameras." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 689, + 286, + 712 + ], + "type": "text", + "content": "In the past few years, neural implicit representations underwent rapid development and have been applied for multi-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 273 + ], + "type": "text", + "content": "view reconstruction and rendering of static scenes. Some methods represent the geometry as the zero level-set of a neural network, and use differentiable surface rendering to optimize the network weights [35,47,64]. Given dense observation of an object, these methods are able to accurately recover its surface. NeRF [32], on the other hand, uses volume rendering to optimize the scene representation. Its simplicity and impressive results inspires a lot of following works, including in-the-wild reconstruction [31, 52], lighting and material estimation [4,48,67], generation [1,6,18, 22,43,53], human rendering [45,51,57,68,69] and so on. More recently, several methods unify surface and volumetric rendering, enabling accurate geometry reconstruction and high-quality novel view synthesis [36,59,63]. Compared to these static scene representations, we aim to enable free view synthesis of dynamic scenes from extremely sparse cameras." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 274, + 544, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 544, + 416 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 544, + 416 + ], + "type": "text", + "content": "NeRF for Dynamic Scenes. Modeling scenes in 4D domain with time dimension included is a direct solution to extend NeRF for dynamic domains. Typical approaches include VideoNeRF [62], NeRFlow [11], DyNeRF [27], and DCT-NeRF [56]. Specifically, VideoNeRF [62] directly learns a spatiotemporal irradiance field from a single video and uses depth estimation to address the shape-motion ambiguities in monocular inputs, while NeRFlow [11] and DCT-NeRF [56] use point trajectory to regularize the network optimization. To deal with the limitation of topology-change modeling in deformation field, Park et al. [38] represent HyperNeRF which can lift NeRF to higher dimensions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 418, + 544, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 544, + 582 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 544, + 582 + ], + "type": "text", + "content": "Dynamic scenes can also be rendered by deforming the radiance field in the canonical space. For example, Nerfies [37] optimizes an additional continuous deformation field by warping each observed point into a canonical 5D NeRF. D-Nerf [40] and NR-Nerf [55] follow a similar framework, but take only monocular videos as training data. In addition, DeVRF [28] uses voxel-based representation instead of MLPs to model the 3D canonical space and the 4D deformation field. Using parametric body templates as the semantic prior, methods like Neural Body [39] and HumanNeRF [60] enable photo-realistic novel view synthesis of complex human performance. The discrepancy between the practical capture process and the existing experimental protocols for monocular videos has been shown in [16]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 583, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 583, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 583, + 544, + 712 + ], + "type": "text", + "content": "NeRF Acceleration. Numerous works emerge with the purpose of speeding up static NeRF using explicit data structures including feature maps, voxels and tensors. For instance, DVGO [50] achieves fast convergence through an explicit representation of a density voxel grid and a feature voxel grid. With the sparse voxel octree structure, NSVF [29] accelerates the novel view reconstruction by discarding the empty voxels in a coarse-to-fine manner. Similarly, Plenoxels [14] and PlenOctree [65] model a scene through a hierarchical 3D grid with spherical harmonics, which can realize an optimization with two orders of magnitude faster" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "16633" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 282, + 238 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 282, + 238 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 282, + 238 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 282, + 238 + ], + "type": "image", + "image_path": "9901884b7286d67e7be05168283f994b5080c4338c3bbd4d85da1ed82e944656.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "type": "text", + "content": "Figure 2. Illustration of our hierarchical tri-projection decomposition method. For a neural 4D field " + }, + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "type": "inline_equation", + "content": "f(x,y,z,t)" + }, + { + "bbox": [ + 46, + 240, + 287, + 297 + ], + "type": "text", + "content": ", we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumes, which are then further projected onto nine 2D planes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 367 + ], + "type": "text", + "content": "than NeRF. In DIVeR [61], ray marching only finds a fixed number of hits on the voxel grid to accelerate volumetric rendering. Moreover, hashing encoding [33] and tensor decomposition [8] are also used as compact representations for NeRF acceleration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 288, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 535 + ], + "type": "text", + "content": "For dynamic scene modeling, DeVRF [28] enables fast non-rigid neural rendering with both 3D volumetric and 4D voxel field. In addition, V4D [15] introduce an effective conditional positional encoding for 4D data to realize fast novel view synthesis. TiNeuVox [12] represents scenes with identifiable explicit data structures and accelerates radiance fields modeling, while Wang et al. extended PlenOctrees [65] into free-view video rendering [58]. However, the low-res volumetric design of these works hinders the capacity of rendering high-quality images. In addition, there are also several concurrent works [5, 13, 21] adopt 6-plane decomposition for dynamic scenes. Compared with these methods, our hierarchical decomposition is more efficient to capture time variations and represent dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 547, + 103, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 103, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 103, + 559 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 639 + ], + "type": "text", + "content": "Building on prior work for spatio-temporal representations and tri-projection decomposition (Sec. 3.1), we propose a hierarchical tri-projection decomposition method (Sec. 3.2) and a coarse-to-fine strategy (Sec. 3.3), which allow us to learn a 4D field at a modest cost of training time and GPU memory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "Spatio-temporal 4D NeRF representation. To represent dynamic objects using neural radiance fields, a naive way is to condition the original neural radiance fields on the times-stamp [56, 62], which we term as NeRF-T. Mathematically," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 430, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 430, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 430, + 83 + ], + "type": "text", + "content": "NeRF-T can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 92, + 545, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 92, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 381, + 92, + 545, + 119 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f (x, y, z, t) = (\\mathbf {f}, \\sigma), \\tag {1} \\\\ g (\\mathbf {f}, \\theta , \\phi) = \\mathbf {c}, \\\\ \\end{array}", + "image_path": "59e1acb5a34e03a61d8bf6adb64f38497ef1bb303cc34c993dd62884476b1d3f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": " is a dynamic implicit field that produces a high-dimensional feature " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": " and a density value " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": " for a point at position " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": " and time instance " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": " is a function that takes into account the viewing direction " + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 304, + 129, + 545, + 187 + ], + "type": "text", + "content": " to predict the final RGB color." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 189, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 545, + 236 + ], + "type": "text", + "content": "To achieve better disentanglement of shape and motion, some methods like D-NeRF [40] propose deformable neural radiance field which adopts a canonical 3D representation with the 4D flow fields:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 245, + 545, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 245, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 372, + 245, + 545, + 272 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f (x, y, z, t) = (\\hat {x}, \\hat {y}, \\hat {z}), \\tag {2} \\\\ g (\\hat {x}, \\hat {y}, \\hat {z}, \\theta , \\phi) = (\\mathbf {c}, \\sigma), \\\\ \\end{array}", + "image_path": "1175881f71b7ac88fdba33d6760b975d159d01a6bfa21b71d0910b6ca938bf0b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "text", + "content": " is the radiance field in canonical configuration and " + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "text", + "content": " is a scene flow field representing the mapping between the scene at time instant " + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 282, + 545, + 317 + ], + "type": "text", + "content": " and the canonical space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": "From the above formulation one can easily observe that both NeRF-T and D-NeRF rely on modeling a 4D field " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "f(x,y,z,t)" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": ", i.e., the dynamic implicit field in NeRF-T and the flow field in D-NeRF. Existing methods mainly adopt MLPs to fit these 4D fields. Such an implicit neural representation does not have an explicit structure and requires extensive computation time for both training and rendering. Tri-projection Decomposition. The tri-projection decomposition is widely used in recent work including EG3D [7] and TensorRF [8] in order to accelerate the training and rendering process in MLP-based NeRF frameworks. Such decomposition factorizes a " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": "-dimensional tensor " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "V_{h}" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": " into three lower-dimensional " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "((n - 1)" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": "-D) tensors " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "V_{l}^{i} \\{i = 1,2,3\\}" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": " by projecting " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "V_{h}" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 318, + 546, + 544 + ], + "type": "text", + "content": "-axis respectively. For example, triplane-based decomposition proposed by EG3D projects the 3D tensor into three 2D feature planes. Compared to voxel-based representations, triplane representation effectively reduces the memory footprint and improves the performance of 3D generation and reconstruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 552, + 528, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 528, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 528, + 565 + ], + "type": "text", + "content": "3.2. Hierarchical Tri-projection Decomposition" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 571, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 677 + ], + "type": "text", + "content": "Our goal is to design a dynamic scene representation with explicit feature grids to accelerate network training and volumetric rendering. However, directly constructing a 4D tensor costs a huge amount of memory and is unacceptable for the purpose of high-resolution rendering. Therefore, we propose a hierarchical triprojection decomposition to factorize the 4D tensor into several compact features, which reduces memory consumption by a large margin while preserving the capability of fitting 4D fields." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Specifically, for a neural 4D field " + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "f(x,y,z,t)" + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": ", we first decompose the 3D space part from 4D spatio-temporal tensor into three time-aware volumetric tensors via tri-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16634" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 155, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 155, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 155, + 84 + ], + "type": "text", + "content": "projection decomposition:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 90, + 286, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 90, + 286, + 114 + ], + "spans": [ + { + "bbox": [ + 52, + 90, + 286, + 114 + ], + "type": "interline_equation", + "content": "\\Pi_ {3} (f (x, y, z, t)) = \\left\\{f _ {z} (x, y, t), f _ {y} (x, z, t), f _ {x} (y, z, t) \\right\\}, \\tag {3}", + "image_path": "96f1fd320ca57caed8dab5a278e1d627fd6d59877422e37a94ef56d8b52e593b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 114, + 287, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 162 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 114, + 287, + 162 + ], + "type": "inline_equation", + "content": "\\Pi_3" + }, + { + "bbox": [ + 47, + 114, + 287, + 162 + ], + "type": "text", + "content": " denotes the projection operator. To further lower space complexity and enable high-resolution representation, we decompose each feature volume into three feature planes as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 167, + 250, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 167, + 250, + 180 + ], + "spans": [ + { + "bbox": [ + 83, + 167, + 250, + 180 + ], + "type": "interline_equation", + "content": "\\pi_ {3} (f _ {z}) = \\left\\{f _ {z t} (x, y), f _ {z y} (x, t), f _ {z x} (y, t) \\right\\}", + "image_path": "011855e49a2a6b144609490a66b2b63ab16d8b5ae784042b313b219f7c76ee65.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 182, + 287, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 182, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 83, + 182, + 287, + 196 + ], + "type": "interline_equation", + "content": "\\pi_ {3} (f _ {y}) = \\left\\{f _ {y t} (x, z), f _ {y z} (x, t), f _ {y x} (z, t) \\right\\} \\tag {4}", + "image_path": "48352e0373089acc31782f586f47a7dd539e656e66addebf8b1dbff8b05c1b7d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 198, + 250, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 198, + 250, + 211 + ], + "spans": [ + { + "bbox": [ + 83, + 198, + 250, + 211 + ], + "type": "interline_equation", + "content": "\\pi_ {3} (f _ {x}) = \\left\\{f _ {x t} (y, z), f _ {x z} (y, t), f _ {x y} (z, t) \\right\\}", + "image_path": "6ecb2439ae8657e25737a43b126585419c7b0961e08df9230710eaf6beda90a2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "inline_equation", + "content": "\\pi_3" + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "content": " denotes the volume-to-plane tri-projection. In this way, we compactly represent a 4D field using 9 planes. Given any spatio-temporal coordinate " + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "inline_equation", + "content": "(x,y,z,t)" + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "content": ", we can efficiently query its value in the 4D field by projecting it onto the planes and retrieving the corresponding value via bilinear interpolation. Fig 2 is an illustration of our decomposition method. Our hierarchical triprojection decomposition reduces the space complexity from " + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^4)" + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n^2)" + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 215, + 287, + 345 + ], + "type": "text", + "content": " being the spatial resolution of the grid, significantly lowering memory footprint without sacrificing representation power." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": "Differences from 6-plane decomposition. Compared with 6-plane decomposition [5, 13, 21] which pairs the time dimension with only one spatial dimension " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "(xt,yt,zt)" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": ", our method first decomposes the 4D tensor in the spatial domain to obtain three time-aware volumes and then 9 decomposed planes. Note that the three volumes are decomposed independently, making the 9 planes different from each other. Specifically, when using a linear layer " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "g = \\{\\}" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": " to decompose the volume " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{z}(x,y,t) = \\{.,f_{zy}(x,t),\\}." + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": " and the volume " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{y}(x,z,t) = \\{.,f_{yz}(x,t),\\}." + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": ", the planes " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{zy}(x,t)" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{yz}(x,t)" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": " behave differently since they are optimized independently to capture time variations in the volumes " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{z}(x,y,t)" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "f_{y}(x,z,t)" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": ", respectively. In this way, the 9 planes can leverage time-aware information in all possible combinations of the time and spatial dimensions " + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "inline_equation", + "content": "(xt,yt,zt,xyt,yzt,xzt,xyzt)" + }, + { + "bbox": [ + 46, + 346, + 287, + 594 + ], + "type": "text", + "content": ". This property enables our representation to capture various dynamic information hierarchically at different levels of spatial dimensions. Therefore, our method is more efficient in representing dynamic scenes, where variations in time dimension are often intense, complex, and long-range." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 665 + ], + "type": "text", + "content": "Our decomposition supports various types of 4D fields including the time-conditioned radiance field in NeRF-T and the 4D flow field in D-NeRF. In this paper, we present a dynamic radiance field decomposition method (Sec. 4.1) for multiview dynamic reconstruction as well as a 4D flow decomposition method (Sec. 4.2) for single-view setting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 178, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 178, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 178, + 685 + ], + "type": "text", + "content": "3.3. Coarse-to-fine Strategy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "To further improve the efficiency of our 4D decomposition, we propose an optional coarse-to-fine strategy to fac" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "text", + "content": "torize the 4D fields into different scales in different training phases. In coarse level, we adopt low-resolution feature planes " + }, + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "text", + "content": " to decompose the 4D fields, which improves the robustness of the training process and achieves fast convergence. After coarse level training, we additionally use high-resolution feature planes " + }, + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "inline_equation", + "content": "(512 \\times 512)" + }, + { + "bbox": [ + 304, + 72, + 545, + 178 + ], + "type": "text", + "content": " for 4D decomposition to represent dynamic details and achieve high-quality rendering. Specifically, we factorize the 4D fields into different scales:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 187, + 545, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 187, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 321, + 187, + 545, + 202 + ], + "type": "interline_equation", + "content": "f (x, y, z, t) = \\left\\{\\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right), \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {H R}\\right)\\right) \\right\\}, \\tag {5}", + "image_path": "73e3b2c59b7e99165fe82c3101abf7a3ed4dc0c7c9b16f84ceeb80e804c30a80.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "inline_equation", + "content": "f^{LR}" + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "inline_equation", + "content": "f^{HR}" + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "text", + "content": " are the coarse-level and fine-level components of " + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "inline_equation", + "content": "f(x,y,z,t)" + }, + { + "bbox": [ + 304, + 211, + 545, + 294 + ], + "type": "text", + "content": ", respectively. In the coarse level, the decomposed feature planes is low-resolution to represent coarse 3D structures and 4D dynamic changes. In the fine level, we adopt the high-resolution feature planes in each element to decompose the 4D fields, which focuses more on recovering dynamic details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 305, + 518, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 305, + 518, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 305, + 518, + 319 + ], + "type": "text", + "content": "4. Tensor4D for Dynamic Reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 326, + 545, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 349 + ], + "type": "text", + "content": "As shown in Fig. 3, we apply our 4D tensor decomposition into the task of dynamic reconstruction with two inputs:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 349, + 545, + 467 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 408 + ], + "type": "text", + "content": "1) Dynamic reconstruction under sparse and fixed cameras. For this setting, we factorize NeRF-T instead of D-NeRF using our proposed 4D decomposition, which we found more efficient and flexible to represent topologically-varying objects (Sec. 4.1)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 408, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 467 + ], + "type": "text", + "content": "2) Dynamic reconstruction using monocular camera. For this setting, we separately decompose the 4D flow fields and the 3D canonical representation since it ensures appearance consistency across different frames and achieves more robust performance (Sec. 4.2)." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 475, + 453, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 453, + 487 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 453, + 487 + ], + "type": "text", + "content": "4.1. Multi-view Reconstruction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 494, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 552 + ], + "type": "text", + "content": "In this section, we present our dynamic reconstruction system under sparse multi-view setting. The goal of our system is 1) efficient and high-quality dynamic reconstruction with low memory and time cost, and 2) robust reconstruction under sparse and fixed camera setting." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": "To this end, we adopt our 4D decomposition (Eq. 5) to factorize the NeRF-T representation with the coarse-to-fine strategy. Specifically, we can obtain nine low-resolution feature planes " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "T^{LR}" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " and nine high-resolution feature planes " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "T^{HR}" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " to represent the 4D NeRF-T fields after our 4D decomposition. Then for volume rendering, when sampling a point " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "(x,y,z,t)" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " in the ray " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " with direction of " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": ", we first query the feature " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 552, + 545, + 659 + ], + "type": "text", + "content": " in both the LR and the HR feature planes. Take the LR features for example:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 667, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 667, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 667, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {p} ^ {L R} (x, y, z, t) = \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right) \\\\ = \\oplus \\left(\\pi_ {3} \\left(f _ {z} ^ {L R}\\right), \\pi_ {3} \\left(f _ {y} ^ {L R}\\right), \\pi_ {3} \\left(f _ {x} ^ {L R}\\right)\\right) \\tag {6} \\\\ = \\oplus (\\oplus (T _ {z t} ^ {L R} (x, y), T _ {z y} ^ {L R} (x, t), T _ {z x} ^ {L R} (y, t)), \\dots), \\\\ \\end{array}", + "image_path": "9b050af13358212b5d8fc6689e44485a16ee2ff8518115986bb66dffdf3bb98c.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16635" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 267, + 185 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 267, + 185 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 267, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 267, + 185 + ], + "type": "image", + "image_path": "a3bdc139b699f7c24c0f5bae50f399896add823a58f9586da5d036be18534439.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 187, + 161, + 198 + ], + "lines": [ + { + "bbox": [ + 148, + 187, + 161, + 198 + ], + "spans": [ + { + "bbox": [ + 148, + 187, + 161, + 198 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 269, + 69, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 269, + 69, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 269, + 69, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 269, + 69, + 545, + 185 + ], + "type": "image", + "image_path": "d0f8f3c10739fd181a473aaa5cbad2e9abf03bffb760a0cfae6ea6412220e2f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 186, + 414, + 198 + ], + "lines": [ + { + "bbox": [ + 401, + 186, + 414, + 198 + ], + "spans": [ + { + "bbox": [ + 401, + 186, + 414, + 198 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 547, + 251 + ], + "type": "text", + "content": "Figure 3. The framework of Tensor4D for multi-view and monocular reconstruction. a). Tensor4D for multi-view reconstruction. The 4D NeRF-T fields are separately factorized by the LR and HR feature planes. b). Tensor4D for monocular reconstruction. The 4D flow fields are factorized by the LR feature plane for better disentanglement of shape and motion. The 3D canonical representation is factorized by three LR and HR feature planes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "content": "where we concatenate all features queried in the nine LR planes as the final LR feature of the sampling point " + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "content": ". Then we concatenate the LR and HR features with the positional encoding of " + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "inline_equation", + "content": "(x, y, z, t)" + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "content": " into the geometry MLP " + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "inline_equation", + "content": "E_{g}" + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "content": " to obtain the density " + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "content": " and high-dimension feature " + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 257, + 288, + 316 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 320, + 287, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 320, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 88, + 320, + 287, + 338 + ], + "type": "interline_equation", + "content": "E _ {g} \\left(F _ {p} ^ {L R}, F _ {p} ^ {H R}, \\gamma (x, y, z, t)\\right) = (\\mathbf {f}, \\sigma), \\tag {7}", + "image_path": "ab8c2e00f01568066a82b2249b99c34b9479dd39737db0fdac91679009e7260b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "text", + "content": " is the positional encoding function. Next, we concatenate the high-dimension feature " + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "text", + "content": " with the positional encoding of " + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 47, + 341, + 287, + 376 + ], + "type": "text", + "content": " and feed it into the color MLP:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 128, + 382, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 382, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 128, + 382, + 287, + 396 + ], + "type": "interline_equation", + "content": "E _ {c} (\\mathbf {f}, \\gamma (\\theta , \\phi)) = c \\tag {8}", + "image_path": "00062587b15156b72b2e731ea005161bacdcf217b9ce5d86dd49550002130e30.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": "In this way, we can render the images through volume rendering and adopt color loss to train our decomposed feature planes " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": " and neural network " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "inline_equation", + "content": "E_{g}" + }, + { + "bbox": [ + 47, + 401, + 287, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 125, + 442, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 442, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 125, + 442, + 287, + 456 + ], + "type": "interline_equation", + "content": "L _ {c} = \\left\\| C _ {i, j} - C _ {i, j} ^ {*} \\right\\| \\tag {9}", + "image_path": "029adc3ca26b856819e15d5dff80fc518db7971e60c962fbb7cdb91585543af0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 461, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 461, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 461, + 287, + 485 + ], + "type": "text", + "content": "With our coarse-to-fine design, our method can recover high-fidelity dynamic details effectively and efficiently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 485, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 287, + 520 + ], + "type": "text", + "content": "To achieve robust dynamic reconstruction under sparse multi-view setting, we further adopt regularization for all decomposed feature planes " + }, + { + "bbox": [ + 47, + 485, + 287, + 520 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 485, + 287, + 520 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 525, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 525, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 59, + 525, + 287, + 563 + ], + "type": "interline_equation", + "content": "L _ {r} = \\sum_ {T} \\sum_ {i, j} \\sqrt {\\left(T _ {i + 1 , j} - T _ {i , j}\\right) ^ {2} + \\left(T _ {i , j + 1} - T _ {i , j}\\right) ^ {2}}, \\tag {10}", + "image_path": "72f364c1d7e54c07346b4036820917f6b3348bbb1b66aaebd0e7f55707d4bf51.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "type": "text", + "content": " is the TV loss for each feature plane " + }, + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 564, + 287, + 635 + ], + "type": "text", + "content": " to keep their sparsity. To regularize the geometry, we also introduce surface constraint into volume rendering. Specifically, we adopt the SDF as the base geometry representation and follow NeuS [59] to render the SDF field. Then we add surface constraint loss to enforce a smooth surface:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 641, + 287, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 641, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 287, + 654 + ], + "type": "interline_equation", + "content": "L _ {e} = \\left\\| \\left\\| \\nabla s (x, y, z, t) \\right\\| _ {2} - 1 \\right\\| _ {2} \\tag {11}", + "image_path": "a0cb7e2b9fc82a864ebba1ae77084b8fe35ef3066627f488bcd4362b39812b4c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "text", + "content": "The final training loss " + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "text", + "content": " is the regularization loss " + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "inline_equation", + "content": "L_{r}, L_{e}" + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "text", + "content": " and the color loss " + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 47, + 660, + 287, + 683 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 691, + 287, + 704 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 691, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 691, + 287, + 704 + ], + "type": "interline_equation", + "content": "L _ {m} = \\lambda_ {r} L _ {r} + \\lambda_ {e} L _ {e} + \\lambda_ {c} L _ {c} \\tag {12}", + "image_path": "d31d96d75d4da85e58757dae6192be0336ddf5a0763037d024def07ba0c5005d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 257, + 454, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 454, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 454, + 268 + ], + "type": "text", + "content": "4.2. Monocular Reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 275, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 393 + ], + "type": "text", + "content": "Different from the sparse view setting, we adopt 4D decomposition for D-NeRF in monocular capture cases. This is because monocular setting is much more ill-posed than sparse-view inputs, and the explicit disentanglement of appearance and motions can guarantee the consistency across different frames. Since D-NeRF represents the dynamic objects with the 4D flow fields and a 3D canonical representation, we separately factorize these two fields. First, for the 4D flow fields, we only adopt coarse level decomposition and factorize it into low-resolution feature planes:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 365, + 400, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 400, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 365, + 400, + 545, + 415 + ], + "type": "interline_equation", + "content": "f (x, y, z, t) = \\pi_ {3} \\left(\\Pi_ {3} \\left(f ^ {L R}\\right)\\right). \\tag {13}", + "image_path": "d24f445c6c70a3a90cab9b0abf9f3baa368f75fa9cad0f8ed5274e75eb5c0fdd.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 422, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 481 + ], + "type": "text", + "content": "Our coarse decomposition focuses more on the coarse and rigid motion, which improves the robustness of flow estimation and can achieve better disentanglement of shape and motion. Then for the 3D canonical representation, we adopt both coarse and fine level decomposition:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 353, + 488, + 545, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 488, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 353, + 488, + 545, + 502 + ], + "type": "interline_equation", + "content": "h (\\hat {x}, \\hat {y}, \\hat {z}) = \\pi_ {3} \\left(h ^ {H R}\\right) + \\pi_ {3} \\left(h ^ {L R}\\right). \\tag {14}", + "image_path": "6ce05bb496aec8d3782b3b0a9a1f4b8f3f8e70ca26d7ad383e5e7f16e472ebe7.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": "Therefore, we can obtain 9 flow feature planes " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": " for 4D flow fields and 6 canonical feature planes " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "T_{h}" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": " for 3D canonical representation. The volume rendering in monocular cases is similar to multi-view cases. For a sampling point " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": ", we first obtain the point flow feature " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "F_{f}" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": " using Eq. 6 with nine flow planes " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": ". Then we adopt the flow MLP " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "E_{f}" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": " to predict the movement of the point:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 359, + 601, + 545, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 601, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 359, + 601, + 545, + 615 + ], + "type": "interline_equation", + "content": "E _ {f} \\left(F _ {p}, \\gamma (x, y, z, t)\\right) = (\\hat {x}, \\hat {y}, \\hat {z}). \\tag {15}", + "image_path": "924ef781f13ae271f1703646466de1c02e98343455cb73b135815245745ffc04.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "text", + "content": "Then we obtain the point canonical feature " + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "text", + "content": " by querying the canonical feature planes " + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "inline_equation", + "content": "T_{h}" + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "text", + "content": ". Take the LR feature " + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "inline_equation", + "content": "F_{c}^{LR}" + }, + { + "bbox": [ + 304, + 622, + 545, + 658 + ], + "type": "text", + "content": " for example:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 664, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 545, + 689 + ], + "type": "interline_equation", + "content": "F _ {c} ^ {L R} = \\pi_ {3} \\left(h ^ {L} R\\right) = \\oplus \\left(T _ {h _ {z}} ^ {L R} (x, y), T _ {h _ {y}} ^ {L R} (x, z), T _ {h _ {z}} ^ {L R} (x, y)\\right). \\tag {16}", + "image_path": "1e75d8c0d401a064e22884c80e7e7e41b74dd21d38815a366421d372857d3d15.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "Next, we feed canonical feature " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "F_{c}^{LR}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " and the positional encoding of " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "(\\hat{x},\\hat{y},\\hat{z})" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " into the geometry MLP " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "E_{g}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " to obtain" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16636" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 199, + 393 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 199, + 393 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 199, + 393 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 199, + 393 + ], + "type": "image", + "image_path": "772e007bb5bd11c0beaecf189bd005bec28fcbb2b10d1cb0d75a3666dd5c8522.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 546, + 418 + ], + "type": "text", + "content": "Figure 4. Example results of our method. Space and time novel view rendering results from sparse-view fixed cameras. The top three results are from four front view cameras and the bottom is from 12 circular cameras." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 533, + 394 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 533, + 394 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 533, + 394 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 533, + 394 + ], + "type": "image", + "image_path": "d1a23377c08547516a08abb86d7e4c73f70ee98eaf1641b064c6ee641c65a4e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 431, + 210, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 210, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 210, + 443 + ], + "type": "text", + "content": "high-dimension feature " + }, + { + "bbox": [ + 47, + 431, + 210, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 47, + 431, + 210, + 443 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 47, + 431, + 210, + 443 + ], + "type": "inline_equation", + "content": "\\sigma" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 451, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 451, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 287, + 466 + ], + "type": "interline_equation", + "content": "E _ {g} \\left(F _ {c} ^ {L R}, \\gamma (\\hat {x}, \\hat {y}, \\hat {z})\\right) = (\\mathbf {f}, \\sigma). \\tag {17}", + "image_path": "97ad3916206e991c6fed0fbface2995c63e649c3b286152743f765c436bf2b53.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "content": "Finally, we adopt Eq. 8 to predict color " + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "content": " for volume rendering and Eq. 9 for training color loss " + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "inline_equation", + "content": "L_{c}" + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "content": ". We also add feature regularization loss " + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "content": " in Eq. 10 and surface constraint loss " + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "content": " in Eq. 11. The total training loss " + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 47, + 474, + 287, + 522 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 531, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 531, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 108, + 531, + 287, + 544 + ], + "type": "interline_equation", + "content": "L _ {s} = \\lambda_ {c} L _ {c} + \\lambda_ {r} L _ {r} + \\lambda_ {e} L _ {e}. \\tag {18}", + "image_path": "e450a326a5f5e77bd2bf81258f760819dc3d7cc49e10c7661e34d79356b00ab5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 123, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 123, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 123, + 576 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Dataset. To evaluate the performance of our methods for multiview inputs, we build a sparse-view capture system with 6 forward-facing RGB cameras mounted on the borders of a " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "32^{\\prime \\prime}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " Looking Glass 3D holographic display [2]. All cameras are synchronized and calibrated. Using this system, we capture multiple sequences of various challenging human motions, including dancing, thumbing up, waving hands, wearing hats and manipulating bags. We use 4 of them for reconstruction and rendering in all our experiments, while leaving the other two for quantitative evaluation. We also use three " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " multiview full body sequences" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 514 + ], + "type": "text", + "content": "captured with 12 evenly spaced cameras on a camera ring for qualitative evaluation. For monocular evaluation, we use the synthetic dataset provided by D-NeRF [40] and select 3 scenes (\"lego\", \"standup\" and \"jumpingjacks\") from this dataset, with the numbers of training frames ranging from 50 to 200. More details about data collection and preprocessing can be found in the Supp.Mat.." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 515, + 546, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 546, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 546, + 646 + ], + "type": "text", + "content": "Baselines. We mainly compare our method against the following state-of-the-art baselines that are most related to our work: D-NeRF [40], NeRF-T, TiNeuVox [12] and NeuS-T. Here, NeRF-T is our extension of vanilla NeRF [32] by introducing an additional time input, and NeuS-T is extended from NeuS [59] similarly. Among these baselines, D-NeRF and TiNeuVox represent the dynamic scenes through deforming a canonical one, while NeRF-T and NeuS-T directly learn a time-conditioned 4D radiance fields. TiNeuVox uses explicit voxel grids to accelerate network training, while others purely use MLPs to model the scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 443, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 443, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 443, + 671 + ], + "type": "text", + "content": "5.1. Results and Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Qualitative Results. We train our model for each individual sequences, and present some example results for novel view synthesis in Fig. 4 and the Supp.Mat.. The results" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16637" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 66, + 538, + 250 + ], + "blocks": [ + { + "bbox": [ + 55, + 66, + 538, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 66, + 538, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 66, + 538, + 250 + ], + "type": "image", + "image_path": "a14f9d797e8a62129d4ecb3955caf376d3547580a7a52147704965d4b2174ccf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 251, + 476, + 262 + ], + "lines": [ + { + "bbox": [ + 116, + 251, + 476, + 262 + ], + "spans": [ + { + "bbox": [ + 116, + 251, + 476, + 262 + ], + "type": "text", + "content": "Figure 5. Comparison on monocular synthetic dataset against D-NeRF [40] and TiNeuVox [12]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 268, + 538, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 268, + 538, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 538, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 538, + 460 + ], + "type": "image", + "image_path": "acf46aa699b1d62650b438d808f1141b69a1759a158d82ff834a2149bc844d31.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 463, + 546, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 463, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 546, + 485 + ], + "type": "text", + "content": "Figure 6. Comparison on sparse-view real-world dataset against D-Nerf [40], TiNeuVox [12] and Neus-T [59]. Four of camera views from our datasets are used for reconstruction and rendering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 506, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 577 + ], + "type": "text", + "content": "cover various body motions, clothing styles and accessories. As shown in Fig. 4, our method can render high-quality images for dynamic scenes and faithfully recover appearance details like the thin finger motions, semi-transparent silk, hand-object interaction, face expressions and cloth wrinkles. See our Supp Video for better visualization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": "Comparisons on monocular dynamic dataset. We first compare our method with the baselines on monocular synthetic dataset. Qualitative results are presented in Fig. 5. Compared to other methods, ours recovers more appearance details and generate less artifacts. The numeric results in Tab. 1 also prove that our method outperforms state-of-the-art methods in terms of rendering quality and accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": "Comparisons on sparse view dataset. We then evaluate the performance of different methods for novel view synthesis given four camera views from our collected six camera-view real-world datasets. The remained two views are used" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 506, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 545, + 541 + ], + "type": "text", + "content": "for quantitative evaluation. Results reported in Fig. 6 and Tab. 2 show again that our method performs better in accurate and high-quality appearance detail synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": "Comparisons of training efficiency. We compare the training time and model size for memory in Tabs. 3 and 4. Our method requires significantly less training time compared to the original D-NeRF, NeRF-T, and Neus-T in monocular and multi-view scenarios. Compared to Ours-NeRF-T and Ours-D-NeRF, the longer training time observed in our method is due to the additional computation for geometry smoothness terms, which reduces artifacts and enhances rendering quality. Compared to TiNeuVox, our method has a reduction in memory consumption and a clear improvement in rendering quality since Tensor4D enables the efficient decomposition of 4D fields at higher spatial resolutions (512" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": " vs. 256" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 559, + 546, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16638" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 81, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 418, + 81 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 418, + 81 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 418, + 81 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on monocular synthetic datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 81, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 81, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 81, + 545, + 159 + ], + "type": "table", + "html": "
MethodLegoStandupJumpingjacks
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]7.83e-321.260.8691.72e-14.63e-434.380.9892.18e-25.46e-433.370.9874.84e-2
NeRF-T3.89e-324.320.9041.55e-16.82e-431.440.9682.36e-25.99e-432.270.9795.37e-2
TiNeuVox [12]3.15e-325.140.9248.37e-22.90e-436.180.9862.02e-23.89e-434.760.9833.33e-2
Ours-NeRF-T3.10e-325.130.9221.24e-15.92e-432.380.9772.29e-25.62e-432.730.9805.16e-2
Ours-D-NeRF4.61e-323.370.8901.19e-12.82e-435.930.9812.07e-24.22e-434.100.9823.41e-2
Ours2.26e-326.710.9533.49e-22.50e-436.320.9831.74e-23.91e-434.430.9823.18e-2
", + "image_path": "6c1d9bdd2152000a12f33af7d4b05440870eea11cc500f1594f6c781fc9f200b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 173, + 545, + 252 + ], + "blocks": [ + { + "bbox": [ + 155, + 163, + 437, + 173 + ], + "lines": [ + { + "bbox": [ + 155, + 163, + 437, + 173 + ], + "spans": [ + { + "bbox": [ + 155, + 163, + 437, + 173 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on the six camera-view real-world datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 173, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 49, + 173, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 545, + 252 + ], + "type": "table", + "html": "
MethodSequence1-thzEarphoneSequence3-yxd
MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓MSE↓PSNR↑SSIM↑LPIPS↓
D-NeRF [40]3.13e-325.150.9100.1858.16e-320.920.8540.2564.84e-323.220.9370.147
TiNeuVox [12]5.72e-322.790.8320.2091.49e-218.550.7070.3198.18e-321.190.8160.233
Neus-T [59]4.95e-323.070.8870.1303.75e-324.220.8770.1842.20e-326.590.9450.099
Ours-NeRF-T5.85e-322.460.8420.1911.27e-219.130.8380.2185.02e-323.060.9140.141
Ours-D-NeRF4.74e-323.270.8640.1768.17e-321.070.8830.1984.06e-323.980.9260.130
Ours1.53e-328.270.9420.0843.20e-325.000.9030.1531.31e-328.830.9620.072
", + "image_path": "ab063f6a2197923e964f5c65749a7bb40d478b96a63fe5692b04391f834d0fd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 59, + 274, + 272, + 363 + ], + "blocks": [ + { + "bbox": [ + 53, + 261, + 280, + 272 + ], + "lines": [ + { + "bbox": [ + 53, + 261, + 280, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 280, + 272 + ], + "type": "text", + "content": "Table 3. Training time and memory report in monocular cases." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 274, + 272, + 363 + ], + "lines": [ + { + "bbox": [ + 59, + 274, + 272, + 363 + ], + "spans": [ + { + "bbox": [ + 59, + 274, + 272, + 363 + ], + "type": "table", + "html": "
MethodLego
PSNRTimeIterations#Params
TiNeuVox25.1434min50k102M
D-NeRF21.2645h800k4.8M
Ours-D-NeRF23.3795min50k20M
NeRF-T24.3238h800k4.4M
Ours-NeRF-T25.1373min50k43M
Ours(6-planes)26.34135min50k17M
Ours26.71144min50k20M
", + "image_path": "7a140ee71bad878dacdb7ae0facf4357796904ededcb64351203527c1bfd224e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 59, + 379, + 272, + 457 + ], + "blocks": [ + { + "bbox": [ + 53, + 366, + 280, + 376 + ], + "lines": [ + { + "bbox": [ + 53, + 366, + 280, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 280, + 376 + ], + "type": "text", + "content": "Table 4. Training time and memory report in multi-view cases." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 379, + 272, + 457 + ], + "lines": [ + { + "bbox": [ + 59, + 379, + 272, + 457 + ], + "spans": [ + { + "bbox": [ + 59, + 379, + 272, + 457 + ], + "type": "table", + "html": "
MethodSequence1-thz
PSNRTimeIterations#Params
TiNeuVox22.7930min50k102M
D-NeRF25.1537h800k4.8M
Neus-T23.0745h800k5.0M
Ours-NeRF-T22.4668min50k43M
Ours(6-planes)27.86105min50k32M
Ours28.27117min50K43M
", + "image_path": "50b32a2236fab83c008260b44917dfffd0b4d93a5aae3738794ad5469e4edc47.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 55, + 483, + 276, + 582 + ], + "blocks": [ + { + "bbox": [ + 47, + 460, + 287, + 480 + ], + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 480 + ], + "type": "text", + "content": "Table 5. Ablation study of hierarchical decomposition and smoothness terms." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 483, + 276, + 582 + ], + "lines": [ + { + "bbox": [ + 55, + 483, + 276, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 276, + 582 + ], + "type": "table", + "html": "
MethodLegoStandup
PSNRSSIMPSNRSSIM
Ours(w/o regular)26.490.94635.910.977
Ours(6-planes)26.440.94435.790.978
Ours26.710.95336.320.983
Sequence-thzSequence-earphone
PSNRSSIMPSNRSSIM
Ours(w/o regular)27.920.93224.750.885
Ours(6-planes)27.740.93424.390.889
Ours28.270.94225.000.903
", + "image_path": "73ec3938bcbbfa8ebd575483f56155acca6518b9dfdd3876ec533bf33059a11a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 588, + 138, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 138, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 138, + 601 + ], + "type": "text", + "content": "5.2. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 607, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 713 + ], + "type": "text", + "content": "Regularization. We quantitatively ablate the regularization terms in our method. We implement two strong baselines \"Ours-NeRF-T\" and \"Ours-D-NeRF\", in which we directly apply our Tensor4D decomposition for NeRF-T and D-NeRF without the regularization terms. The results are reported in Tab. 1 and Tab. 2. Benefiting from our 4D decomposition, they achieve better performance than the original NeRF-T and D-NeRF. However, their rendering quality is still worse than our full method with regularization. In" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 262, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 310 + ], + "type": "text", + "content": "addition, We ablate for TV regular smoothness terms (Ours vs. Ours(w/o regular)) and the results are reported in Tab. 5, which further validates the effectiveness of our smoothness terms for rendering quality enhancement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 310, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 545, + 369 + ], + "type": "text", + "content": "Hierarchical Decomposition We quantitatively ablate our hierarchical decomposition with 6-plane decomposition. As shown in Tab 5, our method achieves superior performance in both multi-view and monocular cases, which validates the effectiveness of our hierarchical decomposition." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 380, + 375, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 375, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 375, + 393 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 401, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 496 + ], + "type": "text", + "content": "Limitations. Since our method needs to decompose the 4D fields to several 2D feature planes, a pre-set bounding box of the scene is necessary. Therefore, it is difficult for our method to reconstruct backgrounds or objects which are out of the bounding box. Another limitation is our strong regularization terms. Though these terms benefit the robustness of our reconstruction under sparse views, they also limit our ability to handle challenging cases such as fluid and fog." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 497, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 638 + ], + "type": "text", + "content": "Conclusion. We presented Tensor4D, a new method for learning high-quality neural representation for dynamic scenes from sparse-view videos or even a monocular video. To capture the spatio-temporal information in a compact and memory-efficient manner, we propose to develop a novel hierarchical tri-projection decomposition method that models a 4D tensor with nine 2D feature planes. With proper design of training losses and regularization, our method provides an efficient yet effective solution to model the radiance fields of dynamic scenes. We believe our work can inspire future research towards low-cost, portable and immersive telepresence systems." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This paper is supported by National Key R&D Program of China (2022YFF0902200), the NSFC project No.62125107 and No.61827805 and the Beijing Science and Technology Project (Z211100004021006)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16639" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 92, + 76, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 76, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 76, + 102 + ], + "type": "text", + "content": "[1] 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 105, + 286, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 105, + 286, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 105, + 286, + 125 + ], + "type": "text", + "content": "[2] The looking glass holographic display. https://lookingglassfactory.com/.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 129, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 286, + 182 + ], + "type": "text", + "content": "[3] Naveed Ahmed, Christian Theobalt, Petar Dobrev, Hans-Peter Seidel, and Sebastian Thrun. Robust fusion of dynamic shape and normal capture for high-quality reconstruction of time-varying geometry. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 184, + 286, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 286, + 236 + ], + "type": "text", + "content": "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 240, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 286, + 272 + ], + "type": "text", + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. arXiv preprint arXiv:2301.09632, 2023. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 275, + 286, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 275, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 275, + 286, + 317 + ], + "type": "text", + "content": "[6] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In arXiv, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 320, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 320, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 320, + 286, + 373 + ], + "type": "text", + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 286, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 286, + 409 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 411, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 286, + 453 + ], + "type": "text", + "content": "[9] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. 34(4), jul 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 286, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 509 + ], + "type": "text", + "content": "[10] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6), nov 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 566 + ], + "type": "text", + "content": "[11] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the IEEE International Conference on Computer Vision, pages 14304-14314, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 611 + ], + "type": "text", + "content": "[12] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In arXiv preprint arXiv:2205.15285, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 656 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. arXiv preprint arXiv:2301.10241, 2023. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance fields without neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5491–5500, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4d: Voxel for 4d novel view synthesis. In arXiv preprint arXiv:2205.14332, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[16] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In In Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 204 + ], + "type": "text", + "content": "[17] Paulo FU Gotardo, Tomas Simon, Yaser Sheikh, and Iain Matthews. Photogeometric scene flow for high-detail dynamic 3d reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 846-854, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "text", + "content": "[18] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 251, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 337 + ], + "type": "text", + "content": "[19] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, Danhang Tang, Anastasia Tkach, Adarsh Kowdle, Emily Cooper, Ming-song Dou, Sean Fanello, Graham Fyffe, Christoph Rhemann, Jonathan Taylor, Paul Debevec, and Shahram Izadi. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Trans. Graph., 38(6), nov 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[20] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Dynamic shape capture using multi-view photometric stereo. In ACM SIGGRAPH Asia 2009 Papers, New York, NY, USA, 2009. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 425 + ], + "type": "text", + "content": "[21] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "text", + "content": "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949-12958, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 503 + ], + "type": "text", + "content": "[23] T. Kanade, P. Rander, and P.J. Narayanan. Virtualized reality: constructing virtual worlds from real scenes. IEEE MultiMedia, 4(1):34-47, 1997. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 505, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 545, + 548 + ], + "type": "text", + "content": "[24] Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. Outdoor dynamic 3-d scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1611-1622, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 624 + ], + "type": "text", + "content": "[25] Jason Lawrence, Danb Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Trans. Graph., 40(6), dec 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[26] Hao Li, Linjie Luo, Daniel Vlasic, Pieter Peers, Jovan Popovic, Mark Pauly, and Szymon Rusinkiewicz. Temporally coherent completion of dynamic shapes. ACM Trans. Graph., 31(1), feb 2012. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Tianye Li, Mira Slavcheva, Michael Zollhofer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16640" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[28] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. In arXiv preprint, arXiv:2205.15723, 2022. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[29] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "text", + "content": "[30] Zhaoyang Lv, Kihwan Kim, Alejandro Troccoli, Deqing Sun, James M. Rehg, and Jan Kautz. Learning rigidity in dynamic scenes with a moving camera for 3d motion field estimation. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 484–501, Cham, 2018. Springer International Publishing. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "text", + "content": "[33] Thomas Muller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. 41(4), jul 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 389, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 443 + ], + "type": "text", + "content": "[34] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 343-352, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "text", + "content": "[35] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "text", + "content": "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[37] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "type": "text", + "content": "[38] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higherdimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[40] Albert Pumarola, Enric Corona, and and Francesc Moreno-Noguer Gerard Pons-Moll. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[41] René Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4058-4066, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[42] Chris Russell, Rui Yu, and Lourdes Agapito. Video pop-up: Monocular 3d reconstruction of dynamic scenes. In Computer Vision - ECCV 2014, pages 583-598, Cham, 2014. Springer International Publishing. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "text", + "content": "[43] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 20154-20166. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[44] Ruizhi Shao, Liliang Chen, Zerong Zheng, Hongwen Zhang, Yuxiang Zhang, Han Huang, Yandong Guo, and Yebin Liu. Floren: Real-time high-quality human performance rendering via appearance flow using sparse rgb cameras. In SIGGRAPH Asia 2022 Conference Papers, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 386 + ], + "type": "text", + "content": "[45] Ruizhi Shao, Hongwen Zhang, He Zhang, Mingjia Chen, Yanpei Cao, Tao Yu, and Yebin Liu. Doublefield: Bridging the neural surface and radiance fields for high-fidelity human reconstruction and rendering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "text", + "content": "[46] Ruizhi Shao, Zerong Zheng, Hongwen Zhang, Jingxiang Sun, and Yebin Liu. Diffustereo: High quality human reconstruction via diffusion-based stereo using sparse cameras. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 702-720. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[47] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Advances in Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 500, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 543 + ], + "type": "text", + "content": "[48] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 545, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 577 + ], + "type": "text", + "content": "[49] Jonathan Starck and Adrian Hilton. Surface capture for performance-based animation. IEEE Computer Graphics and Applications, 27(3):21-31, 2007. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 580, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 632 + ], + "type": "text", + "content": "[50] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[52] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. Neural" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16641" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "3D reconstruction in the wild. In SIGGRAPH Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[53] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "type": "text", + "content": "[54] Aparna Taneja, Luca Ballan, and Marc Pollefeys. Modeling dynamic scenes recorded with freely moving cameras. Computer Vision - ACCV 2010, page 613-626, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 249 + ], + "type": "text", + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE International Conference on Computer Vision, pages 12959-12970, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 282 + ], + "type": "text", + "content": "[56] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. In arXiv preprint, arXiv:2105.05994, 2021. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 338 + ], + "type": "text", + "content": "[57] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. IButter: Neural Interactive Bullet Time Generator for Human Free-Viewpoint Rendering, page 4641–4650. Association for Computing Machinery, New York, NY, USA, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 339, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 287, + 403 + ], + "type": "text", + "content": "[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "text", + "content": "[59] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "type": "text", + "content": "[60] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16210-16220, June 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 516, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 516, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 516, + 287, + 579 + ], + "type": "text", + "content": "[61] Liwen Wu, Jae Yong Lee, Yu-Xiong Wang Anand Bhattachad, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16200-16209, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 634 + ], + "type": "text", + "content": "[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "type": "text", + "content": "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[64] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and ap" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 437 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "pearance. Advances in Neural Information Processing Systems, 33, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE International Conference on Computer Vision, pages 5752-5761, 2021. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[66] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5746-5756, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 207, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 258 + ], + "type": "text", + "content": "[67] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Trans. Graph., 40(6), dec 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 325 + ], + "type": "text", + "content": "[68] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7743-7753, June 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[69] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[70] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16642" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_content_list.json b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa2c203e76b8bb10eec02430306f645f0baa4730 --- /dev/null +++ b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_content_list.json @@ -0,0 +1,1779 @@ +[ + { + "type": "text", + "text": "Test Time Adaptation with Regularized Loss for Weakly Supervised Salient Object Detection", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Olga Veksler \nUniversity of Waterloo, Canada \noveksler@uwaterloo.ca", + "bbox": [ + 359, + 203, + 612, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is well known that CNNs tend to overfit to the training data. Test-time adaptation is an extreme approach to deal with overfitting: given a test image, the aim is to adapt the trained model to that image. Indeed nothing can be closer to the test data than the test image itself. The main difficulty of test-time adaptation is that the ground truth is not available. Thus test-time adaptation, while intriguing, applies to only a few scenarios where one can design an effective loss function that does not require ground truth. We propose the first approach for test-time Salient Object Detection (SOD) in the context of weak supervision. Our approach is based on a so-called regularized loss function, which can be used for training CNN when pixel precise ground truth is unavailable. Regularized loss tends to have lower values for the more likely object segments, and thus it can be used to fine-tune an already trained CNN to a given test image, adapting to images unseen during training. We develop a regularized loss function particularly suitable for test-time adaptation and show that our approach significantly outperforms prior work for weakly supervised SOD.", + "bbox": [ + 75, + 324, + 472, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 660, + 209, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A well known problem with CNNs is that they tend to overfit to the training data. One approach to deal with overfitting is test-time adaptation [30]. A model is trained on the training data, and then fine-tuned for a few epochs during test time on a given test image. The main difficulty of test-time adaptation is that there is no ground truth for the test image. Thus test-time adaptation has been previously used for only a few applications [11, 16, 22, 30, 45] where a suitable loss function can be designed without ground truth.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose the first approach for test-time Salient Object Detection (SOD). The goal of SOD is to find image regions that attract human attention. Convolutional Neural Networks (CNNs) [14, 15] brought a significant breakthrough for SOD [18, 28, 46, 47]. Traditional training of", + "bbox": [ + 73, + 824, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CNNs for SOD requires pixel precise ground truth, but obtaining such annotations is a substantial effort. Therefore, there is an increased interest in semi-supervised [20,23] and weakly supervised SOD [17, 25, 33, 40, 41, 44]. Weak supervision requires less annotation effort, compared to semi-supervision. In this paper, we assume image level supervision, the weakest supervision type, where one provides only images containing salient objects, with no other annotation.", + "bbox": [ + 496, + 292, + 890, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most image level weakly supervised SOD approaches [17, 25, 40, 41, 44] are based on noisy pseudo labels, constructed from unsupervised SOD methods. These approaches are hard to modify for test-time adaptation, as at test time, one has only one image to create noisy pseudo-ground truth samples, but combating noise requires diverse samples.", + "bbox": [ + 496, + 414, + 890, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our test-time adaptation is based on the approach in [33]. Different from the other image level supervised SOD methods, the approach in [33] does not require pseudo labels. They design a regularized loss1 to use for CNN training without pixel precise annotations. Regularized loss models class-independent properties of object shapes and is likely to have lower values for segmentations that separate the object from the background. This makes regularized loss approach particularly suitable for test time adaptation.", + "bbox": [ + 496, + 520, + 890, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We design a regularized loss function tailored for test time adaptation. The main problem with regularized loss in [33] is that it may result in a trivial empty solution if hyperparameters of the loss function are not chosen correctly. When training on a large dataset, an occasional empty result is not a big problem, but when training on a single test image, an empty result is disastrous, catastrophically worsening the performance. Thus we must develop a hyperparameter setting method that avoids empty solutions. However, in the absence of ground truth, setting hyperparameter weights is not trivial. We propose a method for setting hyperparameters specific for each test image such that an empty solution is avoided. This method is crucial for obtaining good results", + "bbox": [ + 496, + 656, + 890, + 852 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1In the context of CNNs, regularization is a term often used to refer to the norm regularization on the network weights [8]. Here, regularized loss refers to the loss function on the output of CNN.", + "bbox": [ + 498, + 862, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7360", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd45db1c1cc850ef9d44998be988f8b357dca97054df8e422f50af75122bac10.jpg", + "image_caption": [ + "test image" + ], + "image_footnote": [], + "bbox": [ + 174, + 89, + 321, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3454a93a7dcf85bd04077bfdab49857fd57f70659ed184de3ca23bbacc3cbb4c.jpg", + "image_caption": [ + "ground truth" + ], + "image_footnote": [], + "bbox": [ + 331, + 90, + 478, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/72799618e0e9319c9e71d76c3ae4423885b6dff5557b91b9882a0f35aad62c2a.jpg", + "image_caption": [ + "base CNN result" + ], + "image_footnote": [], + "bbox": [ + 490, + 90, + 637, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f90606f858ae72d52a2879cf303d734598d61d03148baa73ce85d615b81b2a4d.jpg", + "image_caption": [ + "base CNN result + dense CRF post-processing" + ], + "image_footnote": [], + "bbox": [ + 645, + 90, + 794, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b0b88a0b64a087b2a441f5b7717551c4257267d6eec52095a5c3cfa222ab672d.jpg", + "image_caption": [ + "test-time dataset" + ], + "image_footnote": [], + "bbox": [ + 174, + 186, + 320, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/98db31ad872e83f89fc2ed2fac9bd070d05caecc7582527ea9c68431b4b60497.jpg", + "image_caption": [ + "test-time adaptation with regularized loss" + ], + "image_footnote": [], + "bbox": [ + 321, + 199, + 400, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb35cfe430594d7b2868b9cda7110f66fe324b2e1346eb30d4ee3a7eda771049.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 199, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5f32a82d05e8556e4345a6bda6b9b2701ffb6c6a1c261e5a022d26c0b0f7ba7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 199, + 640, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b5452838b5756a4c971af85f1688d2236ec00fc2a37dd109b6d87d42764d5753.jpg", + "image_caption": [ + "adapted CNN result", + "Figure 1. Overview of our approach. Top: test image, ground truth, output of base CNN on the test image, and the result of dense CRF [13] post-processing of the base CNN result. Bottom: test-time dataset, CNN output on the test image during several test time epochs, and the final result of the adapted CNN." + ], + "image_footnote": [], + "bbox": [ + 648, + 189, + 794, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "during test-time adaptation.", + "bbox": [ + 75, + 371, + 261, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 is an overview of our approach. First we train CNN for SOD in weakly supervised setting with image level labels using [33]. The result is called the base CNN. The top row of Fig. 1 shows a test image, ground truth, and the result of the base CNN. Given a test image, we create a small training dataset by augmentation, see Fig. 1, bottom, left. Then we fine tune the base CNN on the small dataset, using our version of regularized loss, Fig. 1, bottom, middle. The resulting CNN is called the adapted CNN.", + "bbox": [ + 75, + 386, + 468, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The result of base CNN (Fig. 1, top, third image) has many erroneous regions. Sometimes dense CRF [13] is used for post processing to improve performance. We apply dense CRF to the output of base CNN (Fig. 1, top, right). Dense CRF removes small spurious regions but is unable to remove the large erroneous regions as these are salient with high confidence according to the base CNN.", + "bbox": [ + 75, + 522, + 468, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, our approach is able to produce much better results, Fig. 1, bottom, right. This is because the base CNN has a high value of regularized loss for this test image. As the fine tuning proceeds, CNN is forced to find alternative segmentations with better loss values, resulting in increasingly better segmentations. Unlike CRF post processing, our adapted CNN is able to remove the large erroneous regions, as their high confidence values according to the base CNN are ignored, and these regions give higher values of regularized loss. Both dense CRF and our approach use CRF models. However, as opposed to post-processing, we use CRF for CNN supervision, enabling CNN to learn a better segmentation.", + "bbox": [ + 75, + 628, + 468, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments on the standard benchmarks show that test time adaptation significantly improves performance, achieving the new state-of-art in image level weakly supervised SOD.", + "bbox": [ + 75, + 824, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper is organized as follows: Sec. 2 is related", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work, Sec. 3 explains the approach in [33], Sec. 4 describes our approach, and Sec. 5 presents experiments.", + "bbox": [ + 498, + 369, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 417, + 640, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly Supervised Salient Object Detection", + "text_level": 1, + "bbox": [ + 500, + 443, + 812, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most prior work on image level weak supervision consists of two stages. In the first stage, pseudo labels for SOD task are generated, and in the second stage, CNN is trained on the pseudo labels.", + "bbox": [ + 498, + 460, + 890, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One of the first methods to use image level weak supervision for SOD is in [34]. Their method generates pseudo labels by training CNN for object category prediction. In their setting, image level supervision consists of a large set of 200 object classes categories from Image net. After training for object class prediction, they extract foreground heat maps capturing potential object regions. These generalize to unseen categories, and provide initial pseudo labels. Next, they have a self-training stage which alternates between estimating pseudo labels and training CNN on them. For more accurate pseudo-labels, the predicted pseudo-labels are refined with dense CRF [13].", + "bbox": [ + 496, + 521, + 892, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Later image level SOD approaches [17,25,27,40,41,44] use one or more conventional weak (i.e. unsupervised) saliency methods [4,9,10,19] for pseudo labels. In [40], they use multiple weak saliency methods and fuse their results in pseudo-ground truth. During the fusion process, based on the difficulty of training data, they estimate the corresponding confidence measure and utilize it for training with their pseudo ground truth to better handle label noise.", + "bbox": [ + 496, + 702, + 890, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In [44] they develop a noise modeling module, which enables them to deal with noisy saliency maps obtained from multiple weak saliency methods in a probabilistic way. They later extend their method to rely only on a single weak saliency method [41].", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7361", + "bbox": [ + 482, + 924, + 513, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In [17], they approach the noise in pseudo labels by exploiting dense CRF [13]. Initial pseudo labels are generated by a weak saliency method. Then they alternate applying dense CRF and CNN training. Dense CRF corrects the noisy labels through spatial consistency and structure preservation, while CNN is trained on the current version of pseudo labels to update and improve the current model.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [25], they develop a curriculum framework to incrementally refine the pseudo-labels. Instead of directly using the saliency maps produced by weak saliency methods for training, they first train CNNs to generate pseudo labels for each weak saliency method. This improves the pseudo labels as CNNs are forced to learn representations across a broad set of images. Then pseudo labels are further refined via an iterative self-supervision technique.", + "bbox": [ + 75, + 198, + 467, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [27] they utilize multiple pseudo labels to achieve robustness to noise. They design a multi-filter directive CNN which uses multiple directive filters and a multi-guidance loss to integrate multiple pseudo labels.", + "bbox": [ + 76, + 319, + 467, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [42] they propose a noise-aware encoder-decoder to deal with noisy pseudo labels. Their method has a saliency predictor that maps input images to clean saliency maps, and a noise generator. The model that represents noisy labels is a sum of these two models. They train their model to simultaneously infer the corresponding latent vector of each noisy label and the saliency predictor.", + "bbox": [ + 75, + 380, + 467, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All approaches above develop elaborate techniques to deal with noisy pseudo labels. In contrast to the above approaches, in [33] they develop a method for image level supervised SOD which does not rely on pseudo labels from conventional weak saliency methods. The main tool in [33] is regularized loss, based on sparse CRF [2]. The approach in [33] uses off-the-shelf Unet [29] architecture, consists of a single stage, has an intuitive and easy to interpret loss function, and outperforms other image level supervised SOD methods. We review [33] in detail in Sec 3.", + "bbox": [ + 75, + 488, + 467, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to image level weakly supervised SOD methods, there are approaches that use additional sources of supervision. They do not necessarily rely on conventional weak saliency methods. In [39], they use additional weak sources, such as captions, etc. In [38,43] they use scribbles, and in [21] they use boxes, both much stronger forms of weak supervision. We achieve comparable or better results without additional weak sources.", + "bbox": [ + 75, + 640, + 467, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularized Loss", + "text_level": 1, + "bbox": [ + 76, + 762, + 202, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularized loss for weak supervision was first used in [31], for semantic segmentation supervised with scribbles. It was subsequently generalized in [24, 32].", + "bbox": [ + 75, + 777, + 467, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test Time Adaptation", + "text_level": 1, + "bbox": [ + 76, + 824, + 233, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The idea of test time adaptation has been used before. For example, in the context of traditional computer vision, it was used in [30] for online updates in visual tracking and in [11] for adapting detectors from images to videos. In the", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "context of deep learning, it was used in [45] for person re-identification, in [16] for image inpainting, and in [22] for enforcing consistency constraints in depth estimation. In all these methods, the loss to perform test time adaptation is application specific and is unrelated to our approach.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Regularized Loss for SOD", + "text_level": 1, + "bbox": [ + 498, + 180, + 743, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we review the approach in [33], which is designed for image level weakly supervised segmentation for datasets with a single object class. It naturally applies to SOD, since there is only one class, the salient object. The main idea is to design a regularized loss which incorporates the likely properties of generic objects and to use this loss to train CNN instead of cross entropy on ground truth.", + "bbox": [ + 496, + 205, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularized loss is applied to the output of CNN, which is the same size as the input image. Let $x$ denote CNN output, and $x_{p}$ denote the output for pixel $p$ . The last layer of CNN is sigmoid, so that $x_{p} \\in [0,1]$ . Background corresponds to 0 and salient object to 1.", + "bbox": [ + 496, + 311, + 890, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularized loss in [33] consists of a weighted combination of several components. The most important component is sparse CRF loss [2].", + "bbox": [ + 496, + 386, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\operatorname {c r f}} (\\boldsymbol {x}) = \\frac {1}{| \\mathcal {P} |} \\sum_ {(p, q) \\in \\mathcal {N}} e ^ {- \\frac {| | C _ {p} - C _ {q} | | ^ {2}}{2 \\sigma^ {2}}} \\cdot | x _ {p} - x _ {q} |, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 441, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ is the set of all pixels in the image, $\\mathcal{N}$ is the set of neighboring pixel pairs on a 4-connected grid, and $C_p \\in \\mathbb{R}^3$ is the color of $p$ . Parameter $\\sigma$ controls the edge strength and is set to $\\sigma = 0.15$ in all experiments.", + "bbox": [ + 496, + 492, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Usually CNN produces a sharp distribution, i.e. most $x_{p}$ are close either to 0 or 1. Thus if two neighboring pixels are not assigned to the same class, there is a penalty which depends on the strength of the edge between them. Sparse CRF loss is low when segment boundary aligns with image intensity edges. Salient objects are likely to cause image intensity edges, and thus sparse CRF is the main driving tool for salient object discovery. The lowest value of sparse-CRF is zero, achieved at trivial solutions: empty, everything classified as object, or full, everything classified as background. Thus one cannot train with sparse CRF alone.", + "bbox": [ + 496, + 553, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another component of regularized loss is batch volumetric loss. It encourages the average object size in a batch of images to be half of the image size. This loss is useful to prevent both empty and full trivial solutions. Averaging over a batch makes the loss less strict: some batch objects can be significantly smaller or larger than half of the image.", + "bbox": [ + 496, + 719, + 890, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The next component of regularized loss is minimum volume loss. It encourages the object to be of at least a certain size. It is more likely to be valid in practice, compared to volumetric batch loss, as objects can be expected to be at least a certain minimum size. The last two components of regularized loss are border and center losses. The border", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7362", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "loss encourage the image border to be the background, and the center loss encourages image center be the object. This is a realistic assumption for many, but not all, images containing a salient object.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training CNN with regularized loss is difficult and tends to get stuck in a bad local minimum. In [33] they devise a two-stage strategy for training. For our approach, we start with base CNN trained with the method developed in [33].", + "bbox": [ + 75, + 151, + 468, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our test-time adaptation, we only use sparse CRF loss in Eq. (1), minimum volume loss, and border loss. We now describe minimum volume and border loss in detail. Let $\\bar{\\pmb{x}} = \\frac{1}{|\\mathcal{P}|}\\sum_{p\\in \\mathcal{P}}x_p$ , i.e. the normalized object size. Minimum volume loss $L_{\\mathrm{m}}$ penalizes segmentations if the normalized object size is less than $\\mathrm{obj}_{min}$", + "bbox": [ + 75, + 212, + 468, + 305 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {m}} (\\boldsymbol {x}) = R e L U ^ {2} \\left(\\operatorname {o b j} _ {\\min } - \\bar {\\boldsymbol {x}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 314, + 468, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [33], we set $\\mathrm{obj}_{\\min} = 0.15$", + "bbox": [ + 76, + 342, + 333, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathcal{B}$ be the set of pixels on the image border of width $w = 3$ . The border loss $L_{\\mathrm{b}}(\\pmb{x})$ is", + "bbox": [ + 76, + 358, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {b}} (\\boldsymbol {x}) = \\left(\\frac {1}{| \\mathcal {B} |} \\sum_ {p \\in \\mathcal {B}} x _ {p}\\right) ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 398, + 468, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Test Time Adaptation", + "text_level": 1, + "bbox": [ + 76, + 445, + 284, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe our approach. We start with an overview in Sec. 4.1. We describe our regularized loss in Sec. 4.2, and our hyperparameter setting method in Sec. 4.3.", + "bbox": [ + 75, + 470, + 468, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 525, + 187, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of our approach is in Fig. 1. Given a test image $I$ , we first construct a small training dataset $D(I)$ from $I$ by data augmentation. There are various forms of augmentation. However, since we are training without pixel precise ground truth, we need to avoid augmentations which make the salient object more difficult to detect in an image. For example, large crops can remove the salient object. Similarly, adding random image noise changes the distribution of intensity edges, and sparse CRF in Eq. (1) may become less useful for object discovery. We use small random vertical and horizontal image shifts, and random additive shift to color channels (while keeping the shift the same for all image pixels). We found random vertical and horizontal shifts to be the most effective.", + "bbox": [ + 75, + 549, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we take the base CNN trained for SOD task in image level weakly supervised setting, according to the method proposed in [33], and train it on $D(I)$ for a small number of epochs using the regularized loss approach described in Sec. 4.3. The resulting CNN is called the adapted CNN. Training for a small number of iterations ensures that the adapted CNN retains most of the information useful for", + "bbox": [ + 75, + 760, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SOD learned from the large training dataset, as well as ensuring efficiency. In practice, we train for 10 epochs.", + "bbox": [ + 500, + 90, + 888, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Regularized Loss", + "text_level": 1, + "bbox": [ + 500, + 131, + 668, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe our loss function. For test-time adaptation, we only use some of the regularized loss components from [33]. We do not use batch volumetric loss as it encourages the salient segment to be half of the image size and is far from realistic for most test images. We also do not use center loss. It assumes that the image center belongs to the salient object and is likely not true for many specific test images we wish to adapt CNN for.", + "bbox": [ + 496, + 155, + 890, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our regularized loss, we use sparse CRF, minimum volume, and border losses, Eq. (1, 2, 3). Sparse CRF is the most important loss component for discovering salient objects, as it favours segments that align with image intensity edges. Minimum volume loss is necessary to prevent collapse to an empty solution favoured by sparse CRF. Border loss encourages image border to be assigned to the background, and is a realistic assumption for most test images. The complete regularized loss for test time adaptation is", + "bbox": [ + 496, + 276, + 890, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\operatorname {r e g}} (\\boldsymbol {x}) = \\lambda_ {\\operatorname {c r f}} L _ {\\operatorname {c r f}} (\\boldsymbol {x}) + \\lambda_ {m} L _ {\\mathrm {m}} (\\boldsymbol {x}) + \\lambda_ {\\mathrm {b}} L _ {\\mathrm {b}} (\\boldsymbol {x}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 440, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Hyperparameter Setting Method", + "text_level": 1, + "bbox": [ + 500, + 468, + 787, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider the loss components in Eq. (4). To get the best performance during test-time adaptation, it is important to choose a good setting of their relative weights. Each test image has its own best setting, however, we cannot find this setting as the ground truth is not available. Thus we must develop a hyperparameter setting method that does not rely on ground truth.", + "bbox": [ + 496, + 491, + 890, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Out of three components in Eq. (4), the border loss is the least important as it contributes the least to the salient object discovery. Indeed, if our only knowledge about a segmentation is that the image border is assigned to the background, then this gives us almost no information to judge whether this segmentation corresponds to the salient object. Therefore, we set $\\lambda_{b}$ to a low value of 1. Out of the two remaining components, sparse CRF is more important for object discovery. Since objects tend to cause intensity edges in an image, a segment that aligns to image edges is more likely to correspond to a salient object, compared to a segment that just obeys the minimum size constraints. Thus sparse CRF weight $\\lambda_{\\mathrm{crf}}$ should be much higher compared to the minimum volume weight $\\lambda_{m}$ . At the same time, the minimum volume weight should be sufficiently large to prevent the collapse to a trivial empty solution. We can compute this sufficiently large weight from the solution obtained from our base CNN.", + "bbox": [ + 496, + 598, + 890, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now explain how to set $\\lambda_{\\mathrm{m}}$ relative to $\\lambda_{\\mathrm{crf}}$ so that a trivial solution is avoided. Let $x^{i}$ be the output of the base", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2The width of the border can be adjusted relative to the image size, but we train on fixed size images, $256\\times 256$", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7363", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CNN on input image $I$ . Assuming that $\\pmb{x}^i$ is not empty, we set the weight of minimum volume loss large enough to make the loss on $\\pmb{x}^i$ smaller than the loss of an empty solution. Then switching to an empty solution from the initial $\\pmb{x}^i$ is too costly.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results produced by the base CNN tend to be sharp, that is most pixels have values close to either 0 or 1. Therefore we will assume that the trivial empty solution corresponds to all pixels having value 0 in the output. Let $\\pmb{x}^0$ denote such a trivial empty solution, i.e. $x_p^0 = 0$ for all pixels $p$ . Suppose we wish to set sparse-CRF weight to $\\lambda_{\\mathrm{crf}}$ . We need to set $\\lambda_m$ to a value that makes collapse to an empty solution too costly. For this, we need $\\lambda_m$ that satisfies", + "bbox": [ + 75, + 167, + 467, + 286 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\operatorname {r e g}} \\left(\\boldsymbol {x} ^ {i}\\right) = L _ {\\operatorname {r e g}} \\left(\\boldsymbol {x} ^ {0}\\right) + \\epsilon , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 292, + 467, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon > 0$ is a small constant, set in practice to 0.1.", + "bbox": [ + 76, + 316, + 429, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Plugging $\\pmb{x}^0$ in Eqs.(1, 2, 3) we obtain", + "bbox": [ + 96, + 332, + 352, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {b}} \\left(\\boldsymbol {x} ^ {0}\\right) = 0, \\quad L _ {\\mathrm {c r f}} \\left(\\boldsymbol {x} ^ {0}\\right) = 0, \\quad L _ {\\mathrm {m}} \\left(\\boldsymbol {x} ^ {0}\\right) = \\operatorname {o b j} _ {\\min } ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 352, + 468, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Plugging Eq. (6) into Eq. (5) and solving for $\\lambda_{m}$ , we get our formula for setting $\\lambda_{m}$", + "bbox": [ + 76, + 377, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {m} = \\frac {\\lambda_ {\\mathrm {c r f}} L _ {\\mathrm {c r f}} \\left(\\boldsymbol {x} ^ {i}\\right) + L _ {\\mathrm {b}} \\left(\\boldsymbol {x} ^ {i}\\right) - \\epsilon}{\\operatorname {o b j} _ {\\min } ^ {2} - L _ {\\mathrm {m}} \\left(\\boldsymbol {x} ^ {i}\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 412, + 468, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. (7) applies when the output $\\boldsymbol{x}^i$ of the base CNN is not a trivial empty solution. If $\\boldsymbol{x}^i$ is an empty solution, it means that the base CNN fails to extract a salient object. In this case, we set both CRF and minimum volume loss to be of equal weight, $\\lambda_{\\mathrm{crf}} = \\lambda_m$ . This reflects the fact that extracting a salient object from the current image is difficult and the penalty for not satisfying the minim object size is increased. We do not set $\\lambda_{\\mathrm{m}}$ higher than $\\lambda_{\\mathrm{crf}}$ as some images may not contain a salient object, in which case an empty solution is appropriate. In addition, while setting $\\lambda_{\\mathrm{m}}$ high enough would result in some non-empty solution, most likely it would be an erroneous solution, as minimum volume loss is less useful for extracting object segments and, therefore, should not have a weight higher than that of sparse CRF. In practice, we set $\\lambda_{\\mathrm{crf}} = 10^3$ and compute $\\lambda_{\\mathrm{m}}$ for each test image using Eq. (7).", + "bbox": [ + 75, + 454, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 1 we experimentally evaluate the effectiveness of our method for setting $\\lambda_{\\mathrm{m}}$ . We randomly select 500 images from DUTO [37] dataset. We test our adaptive $\\lambda_{\\mathrm{m}}$ for each image, and the same fixed setting of $\\lambda_{\\mathrm{m}}$ for all test images, in a range from 1 to $10^{3}$ , see Table 1. For this experiment, we use a fixed $\\lambda_{\\mathrm{crf}} = 10^{3}$ . The performance metric is $F_{\\beta}$ . The adaptive $\\lambda_{\\mathrm{m}}$ performs better than any fixed choice of $\\lambda_{\\mathrm{m}}$ . As expected, the performance degrades significantly for the larger values of $\\lambda_{\\mathrm{m}}$ as minimum volume loss starts to play a bigger role in the regularized loss function, however, it is less discriminating of salient objects. For smaller $\\lambda_{\\mathrm{m}}$ , the performance suffers as the results on some images collapse to an empty solution.", + "bbox": [ + 75, + 696, + 468, + 892 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 89, + 705, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our implementation is in Pytorch using RTX 2080 GPU. For our base CNN, we use image level weakly supervised SOD approach [33]. We use Unet [29] architecture with ResNeXt [35] fixed features pretrained onImagenet [5] and train on $256 \\times 256$ images. Unlike [33], we train on larger images and on a larger training dataset, thus getting better results for the base CNN than those reported in [33].", + "bbox": [ + 496, + 114, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the base CNN on the training set of DUTS [34], which contains 10,553 images. We then evaluate our test-time CNN adaptation approach on the test set of DUTS (5019 images), DUT-OMRON [37] (5168 images), ECSSD [36] (1000 images), THUR [3] (6233 images), PascalS [19] (850 images), and HKU-IS [18] (4447 images).", + "bbox": [ + 496, + 220, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For test time adaptation, given a test image $I$ , we construct a test time adaptation dataset $D(I)$ that has 16 images. These images are constructed from $I$ by shifting to the left or right with probability 0.5, and, in addition, shifting up or down with probability 0.5. Each shift is chosen uniformly from the range [0, 15]. This creates a dataset of images similar to $I$ , and it is unlikely that the salient object is removed out of the sample by these small shifts. We use Adam optimizer [12], a fixed learning rate 0.001, and batch size equal to the size of the test time adaptation dataset, set to 16 images. We train for 10 epochs. The images are scaled to resolution $256 \\times 256$ . Test time adaptation takes 2.8 seconds per image.", + "bbox": [ + 496, + 311, + 892, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the standard metrics widely adopted for SOD. Our first metric is $F_{\\beta}$ score [1], with $\\beta^2 = 0.3$ . We also use mae [26], defined as the average absolute per pixel difference between the predicted saliency and ground truth. In addition, we use two newer metrics, namely $S_{\\alpha}$ [6], and $E_{s}$ [7]. $S_{\\alpha}$ evaluates the structural information for saliency map and region-aware and object-aware structural similarity between saliency maps and ground truth. $E_{s}$ is based on unification of global and local information.", + "bbox": [ + 496, + 508, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. SOD results", + "text_level": 1, + "bbox": [ + 500, + 650, + 629, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 2, we report the results of our approach and compare to prior image level weakly supervised SOD methods. Our results with fixed hyperparameter settings in Eq. (4) are in column \"ours\" (with $\\lambda_{\\mathrm{m}} = 20$ , which is the best fixed setting according to Table 1). The results with our method for hyperparameter setting for each image individually, as described in Sec. 4.3, are in column 'ours + h'. Except one case, all metrics for our approach with hyperparameter setting method are improved, often significantly. This shows that our hyperparameter setting method is crucial for obtaining a significantly improved performance.", + "bbox": [ + 496, + 674, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dense CRF [13] is often used to post-process CNN results. Therefore, we also post-process base CNN results with dense CRF for comparison. The performance of the base CNN is in column [33], and its performance after post-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7364", + "bbox": [ + 482, + 925, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/54a76dd05b3812b467845c7b58f15620abe54006bb9f5446c2073e5de3002611.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
λm151020501002005001000adaptive
Fβ.741.738.752.755.742.732.728.718.703.765
", + "bbox": [ + 223, + 88, + 743, + 122 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b6c5b71e35bffb17cbda81a77c6d8e99c9939a043072a7c73053f45eaa97f380.jpg", + "table_caption": [ + "Table 1. Performance in terms of $F_{\\beta}$ score of fine tuning with a fixed value of $\\lambda_{m}$ vs. our adaptive $\\lambda_{\\mathrm{m}}$ computation according to Eq. (7). For all experiments, $\\lambda_{\\mathrm{crf}} = 10^{3}$ ." + ], + "table_footnote": [], + "table_body": "
Metric[34][17][39][44][42][27][33][33]+dCRFoursours+h
DUTS\\( F_{\\beta} \\)↑.654.614.684.725.747.710.753.763.771.795
mae↓.100.116.091.075.060.076.056.055.054.052
\\( E_s \\)↑.795.772.814.853.859.839.833.842.855.863
\\( S_{\\alpha} \\)↑.748.697.759.813.827.775.791.798.800.817
ECSSD\\( F_{\\beta} \\)↑.823.797.609.810.852.854.868.873.882.911
mae↓.104.110.109.092.071.084.063.060.057.046
\\( E_s \\)↑.869.853.673.836.883.885.884.886.885.909
\\( S_{\\alpha} \\)↑.811.802.756.846.860.834.845.855.859.883
DUTO\\( F_{\\beta} \\)↑.603.622.609.597.701.646.697.710.721.752
mae↓.109.101.109.103.070.087.074.069.066.057
\\( E_s \\)↑.768.776.673.712.816.803.821.827.833.842
\\( S_{\\alpha} \\)↑.725.752.756.733.791.742.763.771.782.798
PascalS\\( F_{\\beta} \\)↑.715.693.713--.751.796.801.801.807
mae↓.139.149.135--.115.083.082.082.084
\\( E_s \\)↑.791.772.790--.817.827.827.826.828
\\( S_{\\alpha} \\)↑.744.717.768--.770.791.793.792.796
HKUIS\\( F_{\\beta} \\)↑--.814.820.878.851.866.871.883.899
mae↓--.084.065.043.059.044.040.039.037
\\( E_s \\)↑--.895.858.919.921.912.925.926.931
\\( S_{\\alpha} \\)↑--.818.860.890.846.857.866.871.879
THUR\\( F_{\\beta} \\)↑---.694.719-.731.749.753.760
mae↓---.086.070-.073.069.067.064
\\( E_s \\)↑---.807.838-.840.842.843.844
\\( S_{\\alpha} \\)↑---.804.810-.794.798.806.819
", + "bbox": [ + 183, + 176, + 785, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison to SOD methods that also use image level weak supervision using metrics $F_{\\beta}$ , $mae$ , $E_s$ , and $S_{\\alpha}$ . We use $\\uparrow$ to indicate that the higher score is better, and $\\downarrow$ to indicate the lower score is better. The best result is in bold, and the second best is underlined.", + "bbox": [ + 75, + 571, + 892, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "processing with dense CRF is in column [33] + dCRF. Our approach is significantly better than [33] with dense CRF post-processing. This illustrates that it is important to include CRF during training in the loss function, rather than use CRF for just post-processing.", + "bbox": [ + 75, + 626, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our test time adaptation improves the performance of [33] across all metrics, except one case. For easier datasets (ECSSD, HKUIS), the performance gain is not as large as the base CNN already performs well, but for harder to fit datasets, the performance gain is significant.", + "bbox": [ + 75, + 703, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing to other SOD methods, our method is better in almost all cases except $S_{\\alpha}$ metric for HKUIS and DUTS. For most datasets, our approach significantly improves $F_{\\beta}$ metric over the second best method, notably by roughly $7\\%$ for DUTO and DUTS datasets.", + "bbox": [ + 75, + 779, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3, we compare our method to weakly supervised SOD methods that use stronger form of weak supervision. In [21], they use bounding box supervision, and in [38, 43],", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "they use scribbles. Our method is better than the bounding box method [43] in all cases. Our method is almost always better than [43], and better in about half of the cases compared to [38]. If Fig. 2, we show qualitative comparisons of our method to the scribble supervised methods of [43] and [38]. For this illustration, we chose examples where all methods work relatively well, in terms of the standard metrics. However, our results have better detail preservation qualities, especially when the salient object has sharp edges on the boundary. Observe the detail preservation in the chair, in the last row. This is because CRF based loss function is adapted to the particular test image, driving the segmentation boundary to align with image edges.", + "bbox": [ + 496, + 626, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Experiments", + "text_level": 1, + "bbox": [ + 500, + 832, + 705, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameters in Regularized Loss", + "text_level": 1, + "bbox": [ + 500, + 854, + 769, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first discuss and perform ablation for hyperparameters of our regularized loss function, Eq. (4). The most im", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7365", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78b9d39f91795bfd64eb022f95c9e80c17165fa925997d614be26b6145817a1b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DUTSECSSDDUTOPascalSHKUISTHUR
boxes [21]Fβ↑.736.860.686-.853-
mae ↓.079.072.081-.058-
Es ↑.831.889.810-.897-
Sα ↑.789.858.776-.852-
scribbles [43]Fβ↑.747.865.715.788.858.718
mae ↓.062.061.068.140.047.077
Es ↑.865.908.835.798.923.873
Sα ↑------
scribbles [38]Fβ↑.823.900.758.823.896.755
mae ↓.049.049.060.078.038.069
Es ↑.890.908.862.847.938.843
Sα ↑------
oursFβ↑.795.911.752.807.899.760
mae ↓.052.046.057.082.037.064
Es ↑.863.909.842.828.931.844
Sα ↑.817.883.798.796.879.819
", + "bbox": [ + 209, + 87, + 763, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef7b8bb40b2ddef51973af2c6b910fd95b82b4db13b11e69bc9db2725b21f246.jpg", + "image_caption": [ + "Figure 2. Qualitative comparison. From left to right: input image, ground truth, results of [43], [38], and our results, respectively. Note that [43], [38] both use scribble supervision, a much stronger supervision form compared to what we use." + ], + "image_footnote": [], + "bbox": [ + 81, + 431, + 475, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "important parameter to set is $\\lambda_{m}$ in relation to $\\lambda_{\\mathrm{crf}}$ , as the correct relative settings of these two parameters ensures there is no collapse to a trivial solution. For this, we have de", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f88d256151a6f0ecc12559cec9b3e8f1f540f581bbf255283e1cfd27c2d816c4.jpg", + "table_caption": [ + "Table 3. Comparison to prior SOD methods that use stronger forms of weak supervision using metrics $F_{\\beta}$ , $mae$ , $E_s$ , and $S_{\\alpha}$ . In [21], they use bounding box supervision, and in [38,43], they use scribbles. We use $\\uparrow$ to indicate that the higher score is better, and $\\downarrow$ to indicate the lower score is better. The best result is in bold, and second best is underlined." + ], + "table_footnote": [], + "table_body": "
with border losswithout border loss
DUTS.795.789
ECSSD.911.910
DUTO.752.748
PascalS.807.800
HKUIS.899.892
THUR.760.755
", + "bbox": [ + 527, + 428, + 864, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance of our approach with border loss and without border loss. Performance metric is $F_{\\beta}$ score.", + "bbox": [ + 500, + 553, + 890, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "veloped an effective automatic approach, Sec. 4.3. Since $\\lambda_{m}$ is computed in relation to $\\lambda_{\\mathrm{crf}}$ , the setting of $\\lambda_{\\mathrm{crf}}$ can be arbitrary (up to the precision that can be handled by the software), $\\lambda_{\\mathrm{m}}$ is adjusted accordingly: for larger $\\lambda_{\\mathrm{crf}}$ , $\\lambda_{\\mathrm{m}}$ will be automatically set larger according to Eq. (7).", + "bbox": [ + 498, + 609, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The setting of $\\lambda_{b}$ , as long as it is much smaller than $\\lambda_{\\mathrm{crf}}$ , is not important. Parameter $\\lambda_{\\mathrm{b}}$ controls the border loss that encourages border pixels to be assigned to the background. While this loss is helpful, it is the least important regularized loss component. To show that this loss is the least important, we perform an experiment where we remove border loss from regularized loss ( $\\lambda_{b} = 0$ ). The results are summarized in Table 4. Without border loss, the results worsen only a little. On the other hand, omitting either CRF or minimum volume losses from our loss function in Eq. (4) results in a catastrophic performance collapse.", + "bbox": [ + 496, + 686, + 892, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Eq. (1), we set $\\sigma = 0.15$ as in [33]. Since the training images are normalized, this setting works well across all datasets. We also experimented with other values of $\\sigma$ , see", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7366", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4378e20d888e4e69d3ff6a0fbb37cfe811b11213ed7fb7f2cf3a5ac475b1d4fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
σ = 0.1σ = 0.15σ = 0.20
DUTS.789.795.797
ECSSD.894.911.915
DUTO.751.752.743
PascalS.803.807.803
HKUIS.902.899.895
THUR.763.760.757
", + "bbox": [ + 135, + 88, + 411, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3e503d688d11c80a3d23ccdf16fc2e1c56ce8a8d6a139dafa037ca51c1f41ecc.jpg", + "table_caption": [ + "Table 5. Performance of our approach for different $\\sigma$ values. Performance metric is ${F}_{\\beta }$ score." + ], + "table_footnote": [], + "table_body": "
|D(I)|248163264
Fβ↑.879.887.890.910.915.913
mae ↓.058.055.049.046.045.046
", + "bbox": [ + 104, + 246, + 442, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc89a302da96f30bdefbd43c7251e82546f2ab41960fbb6e17c5979e3f5766bf.jpg", + "table_caption": [ + "Table 6. Different choice of test time dataset size for ECSSD dataset." + ], + "table_footnote": [], + "table_body": "
# epoch2510203050
Fβ↑.873.883.910.915.915.912
mae ↓.062.053.047.045.046.047
", + "bbox": [ + 98, + 344, + 444, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The performance is stable for a range of $\\sigma$ values.", + "bbox": [ + 76, + 457, + 462, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Other Hyperparameters", + "text_level": 1, + "bbox": [ + 76, + 483, + 302, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We now perform ablation experiments for various hyperparameter settings (other than those in the regularized loss function) of our approach. We perform experiments on ECSSD dataset, and evaluate the sensitivity of our approach to the size of the test time adaptation dataset $D(I)$ , the number of epoch performed for test time adaptation, and different data augmentation strategies. For all the experiments, the performance metric is $F_{\\beta}$ and $mae$ .", + "bbox": [ + 75, + 506, + 467, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 6, we show how performance depends on varying the size of the test time dataset $D(I)$ . Our standard setting for the size of $D(I)$ is 16. Decreasing it to 8 slightly degrades the performance, increasing it to 64 results in almost no change, but the running time is twice longer.", + "bbox": [ + 75, + 628, + 467, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 7, we show how performance depends on increasing the number of epoch for test time adaptation. Our standard setting for the number of epoch is 10. Decreasing it to 2 noticeably degrades the performance, increasing it to 50 results in only a slight improvement but significantly increases the running time.", + "bbox": [ + 75, + 704, + 467, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 8, we show the effect of different augmentation strategies. Our main experiments use small random shifts in the vertical and horizontal directions. We also tested adding random Gaussian noise to images, a standard data augmentation technique, but it significantly degrades the performance. Adding Gaussian noise decreases the effectiveness of our sparse CRF loss as neighboring pixels become less", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/624e0365c1aac42d391b4c65353ca9574b8a9eda3b446523c81a02a34320fca7.jpg", + "table_caption": [ + "Table 7. Different number of test time adaptation iterations for ECSSD dataset." + ], + "table_footnote": [], + "table_body": "
typeshiftsGauss noisecolor perturbationshifts+color
Fβ↑.911.854.897.909
mae↓.046.060.050.049
", + "bbox": [ + 500, + 88, + 923, + 133 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b4667ab6f34772c872e7628f91d648afe8a30225f7d0662e24511d2f083705c.jpg", + "table_caption": [ + "Table 8. Different augmentation strategies when fine tuning on ECSSD dataset." + ], + "table_footnote": [], + "table_body": "
DUTSECSSDDUTOPascalSHKUISTHUR
% improved79.383.785.381.276.775.0
mean Fβ increase.048.053.064.17.037.036
mean Fβ decrease.021.009.011.08.019.015
", + "bbox": [ + 500, + 191, + 942, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Percentage of images for which performance improves as the result of test-time adaptation. For the improved (and degraded) images, we also show the mean increase (and the decrease) in $F_{\\beta}$ score.", + "bbox": [ + 500, + 257, + 890, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "strongly connected. We also tested random color perturbation where all pixels change the value of a color channel by the same random value. This works much better than Gaussian noise, but not as well as random shifts. Shifts and color perturbation work as well as shifts alone.", + "bbox": [ + 498, + 345, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Additional Experiments", + "text_level": 1, + "bbox": [ + 500, + 439, + 720, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During test-time adaptation, the majority, but not all images, improve. In Table 9, we show the percentage of images that improve, in terms of their $F_{\\beta}$ score, for each of the test datasets. In addition, we show the mean improvement of $F_{\\beta}$ score, averaged over the images that improve, and the mean decline of $F_{\\beta}$ score, averaged over the images that decline in performance. As can be seen from the table, for all dataset, over $75\\%$ of images improve during test-time adaptation. Also, the average improvement is significantly larger than the average decline of the results. Thus overall, test-time adaptation significantly improves the performance. In the supplementary materials, we show the most improved and the most degraded images during test time adaptation.", + "bbox": [ + 498, + 465, + 890, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 690, + 625, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the first approach to test time adaptation for SOD with image level weak supervision. We develop a regularized loss function particularly effective for test time adaptation, and a method to find an appropriate setting of the hyperparameters in our loss function. Our approach uses a standard CNN architecture and a loss function that is intuitive and simple to interpret. We achieve a new state of the art in image level supervised SOD, and a better or competitive result with the prior work that uses stronger forms of weak supervision. The drawback of our approach is that the running time for testing is more expensive, since the base CNN needs to be fine-tuned for each test image.", + "bbox": [ + 498, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7367", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5", + "[2] Yuri Boykov, Olga Veksler, and Ramin Zabih. A variable window approach to early vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1283-1294, December 1998. An earlier version of this work appeared in CVPR '97. 3", + "[3] Ming-Ming Cheng, NiloyJ. Mitra, Xiaolei Huang, and ShiMin Hu. Salientshape: group saliency in image collections. The Visual Computer, 30(4):443-453, 2014. 5", + "[4] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2", + "[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5", + "[6] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 5", + "[7] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 698-704, 2018. 5", + "[8] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT Press, 2016. 1", + "[9] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2", + "[10] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2", + "[11] Vidit Jain and Erik Learned-Miller. Online domain adaptation of a pre-trained cascade of classifiers. In CVPR 2011, pages 577-584. IEEE, 2011. 1, 3", + "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[13] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In Neural Information Processing Systems, pages 109-117, 2011. 2, 3, 5", + "[14] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Neural Information Processing Systems, 2012. 1", + "[15] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Comput., 1(4):541-551, December 1989. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021. 1, 3", + "[17] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 3, 6", + "[18] G. Li and Y. Yu. Deep contrast learning for salient object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 478-487, June 2016. 1, 5", + "[19] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 280-287, 2014. 2, 5", + "[20] Jiawei Liu, Jing Zhang, and Nick Barnes. Semi-supervised salient object detection with effective confidence estimation. arXiv preprint arXiv:2112.14019, 2021. 1", + "[21] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE Transactions on Image Processing, 30:4423-4435, 2021. 3, 6, 7", + "[22] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 3", + "[23] Yunqiu Lv, Bowen Liu, Jing Zhang, Yuchao Dai, Aixuan Li, and Tong Zhang. Semi-supervised active salient object detection. Pattern Recognition, 123:108364, 2022. 1", + "[24] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6608-6618, October 2021. 3", + "[25] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. In Advances in Neural Information Processing Systems, pages 204-214, 2019. 1, 2, 3", + "[26] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In Conference on Computer Vision and Pattern Recognition, pages 733-740. IEEE, 2012. 5", + "[27] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 2, 3, 6", + "[28] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 1", + "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 3, 5", + "[30] David A Ross, Jongwoo Lim, Ruei-Sung Lin, and Ming-Hsuan Yang. Incremental learning for robust visual track" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7368", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. International journal of computer vision, 77(1):125-141, 2008. 1, 3", + "[31] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised CNN segmentation. In Conference on Computer Vision and Pattern Recognition, pages 1818-1827, 2018. 3", + "[32] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised CNN segmentation. In European Conference on Computer Vision, pages 524-540, 2018. 3", + "[33] Olga Veksler. Regularized loss for weakly supervised single class semantic segmentation. In European Conference on Computer Vision, pages 348-365. Springer, 2020. 1, 2, 3, 4, 5, 6, 7", + "[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5, 6", + "[35] Saining Xie, Ross B. Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5", + "[36] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 1155-1162. IEEE, 2013. 5", + "[37] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Conference on Computer Vision and Pattern Recognition, pages 3166-3173. IEEE, 2013. 5", + "[38] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3234-3242. AAAI Press, 2021. 3, 6, 7", + "[39] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 6074-6083, 2019. 3, 6", + "[40] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2", + "[41] Jing Zhang, Yuchao Dai, Tong Zhang, Mehrtash Harandi, Nick Barnes, and Richard Hartley. Learning saliency from single noisy labelling: A robust model fitting perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2866-2873, 2021. 1, 2", + "[42] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European" + ], + "bbox": [ + 78, + 92, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision, pages 349-366. Springer, 2020. 3, 6", + "[43] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In Conference on Computer Vision and Pattern Recognition, pages 12546-12555, 2020. 3, 6, 7", + "[44] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9029-9038, 2018. 1, 2, 6", + "[45] Shun Zhang, Jia-Bin Huang, Jongwoo Lim, Yihong Gong, Jinjun Wang, Narendra Ahuja, and Ming-Hsuan Yang. Tracking persons-of-interest via unsupervised representation adaptation. International Journal of Computer Vision, 128(1):96-120, 2020. 1, 3", + "[46] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3085-3094, 2019. 1", + "[47] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. 2020. 1" + ], + "bbox": [ + 501, + 92, + 890, + 416 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7369", + "bbox": [ + 482, + 926, + 514, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_model.json b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8ff43488b50ab5167df1cf9ef9dc2d21de63465f --- /dev/null +++ b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.178 + ], + "angle": 0, + "content": "Test Time Adaptation with Regularized Loss for Weakly Supervised Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.204, + 0.613, + 0.257 + ], + "angle": 0, + "content": "Olga Veksler \nUniversity of Waterloo, Canada \noveksler@uwaterloo.ca" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.473, + 0.628 + ], + "angle": 0, + "content": "It is well known that CNNs tend to overfit to the training data. Test-time adaptation is an extreme approach to deal with overfitting: given a test image, the aim is to adapt the trained model to that image. Indeed nothing can be closer to the test data than the test image itself. The main difficulty of test-time adaptation is that the ground truth is not available. Thus test-time adaptation, while intriguing, applies to only a few scenarios where one can design an effective loss function that does not require ground truth. We propose the first approach for test-time Salient Object Detection (SOD) in the context of weak supervision. Our approach is based on a so-called regularized loss function, which can be used for training CNN when pixel precise ground truth is unavailable. Regularized loss tends to have lower values for the more likely object segments, and thus it can be used to fine-tune an already trained CNN to a given test image, adapting to images unseen during training. We develop a regularized loss function particularly suitable for test-time adaptation and show that our approach significantly outperforms prior work for weakly supervised SOD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.21, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.824 + ], + "angle": 0, + "content": "A well known problem with CNNs is that they tend to overfit to the training data. One approach to deal with overfitting is test-time adaptation [30]. A model is trained on the training data, and then fine-tuned for a few epochs during test time on a given test image. The main difficulty of test-time adaptation is that there is no ground truth for the test image. Thus test-time adaptation has been previously used for only a few applications [11, 16, 22, 30, 45] where a suitable loss function can be designed without ground truth." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We propose the first approach for test-time Salient Object Detection (SOD). The goal of SOD is to find image regions that attract human attention. Convolutional Neural Networks (CNNs) [14, 15] brought a significant breakthrough for SOD [18, 28, 46, 47]. Traditional training of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.414 + ], + "angle": 0, + "content": "CNNs for SOD requires pixel precise ground truth, but obtaining such annotations is a substantial effort. Therefore, there is an increased interest in semi-supervised [20,23] and weakly supervised SOD [17, 25, 33, 40, 41, 44]. Weak supervision requires less annotation effort, compared to semi-supervision. In this paper, we assume image level supervision, the weakest supervision type, where one provides only images containing salient objects, with no other annotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Most image level weakly supervised SOD approaches [17, 25, 40, 41, 44] are based on noisy pseudo labels, constructed from unsupervised SOD methods. These approaches are hard to modify for test-time adaptation, as at test time, one has only one image to create noisy pseudo-ground truth samples, but combating noise requires diverse samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Our test-time adaptation is based on the approach in [33]. Different from the other image level supervised SOD methods, the approach in [33] does not require pseudo labels. They design a regularized loss1 to use for CNN training without pixel precise annotations. Regularized loss models class-independent properties of object shapes and is likely to have lower values for segmentations that separate the object from the background. This makes regularized loss approach particularly suitable for test time adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.853 + ], + "angle": 0, + "content": "We design a regularized loss function tailored for test time adaptation. The main problem with regularized loss in [33] is that it may result in a trivial empty solution if hyperparameters of the loss function are not chosen correctly. When training on a large dataset, an occasional empty result is not a big problem, but when training on a single test image, an empty result is disastrous, catastrophically worsening the performance. Thus we must develop a hyperparameter setting method that avoids empty solutions. However, in the absence of ground truth, setting hyperparameter weights is not trivial. We propose a method for setting hyperparameters specific for each test image such that an empty solution is avoided. This method is crucial for obtaining good results" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1In the context of CNNs, regularization is a term often used to refer to the norm regularization on the network weights [8]. Here, regularized loss refers to the loss function on the output of CNN." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.09, + 0.323, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.168, + 0.274, + 0.177 + ], + "angle": 0, + "content": "test image" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.091, + 0.48, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.168, + 0.442, + 0.177 + ], + "angle": 0, + "content": "ground truth" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.091, + 0.638, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.168, + 0.608, + 0.176 + ], + "angle": 0, + "content": "base CNN result" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.795, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.169, + 0.795, + 0.188 + ], + "angle": 0, + "content": "base CNN result + dense CRF post-processing" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.188, + 0.321, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.266, + 0.311, + 0.276 + ], + "angle": 0, + "content": "test-time dataset" + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.2, + 0.401, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.25, + 0.595, + 0.26 + ], + "angle": 0, + "content": "test-time adaptation with regularized loss" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.2, + 0.481, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.2, + 0.641, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.19, + 0.795, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.267, + 0.777, + 0.277 + ], + "angle": 0, + "content": "adapted CNN result" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Figure 1. Overview of our approach. Top: test image, ground truth, output of base CNN on the test image, and the result of dense CRF [13] post-processing of the base CNN result. Bottom: test-time dataset, CNN output on the test image during several test time epochs, and the final result of the adapted CNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.262, + 0.386 + ], + "angle": 0, + "content": "during test-time adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Fig. 1 is an overview of our approach. First we train CNN for SOD in weakly supervised setting with image level labels using [33]. The result is called the base CNN. The top row of Fig. 1 shows a test image, ground truth, and the result of the base CNN. Given a test image, we create a small training dataset by augmentation, see Fig. 1, bottom, left. Then we fine tune the base CNN on the small dataset, using our version of regularized loss, Fig. 1, bottom, middle. The resulting CNN is called the adapted CNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.628 + ], + "angle": 0, + "content": "The result of base CNN (Fig. 1, top, third image) has many erroneous regions. Sometimes dense CRF [13] is used for post processing to improve performance. We apply dense CRF to the output of base CNN (Fig. 1, top, right). Dense CRF removes small spurious regions but is unable to remove the large erroneous regions as these are salient with high confidence according to the base CNN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.824 + ], + "angle": 0, + "content": "In contrast, our approach is able to produce much better results, Fig. 1, bottom, right. This is because the base CNN has a high value of regularized loss for this test image. As the fine tuning proceeds, CNN is forced to find alternative segmentations with better loss values, resulting in increasingly better segmentations. Unlike CRF post processing, our adapted CNN is able to remove the large erroneous regions, as their high confidence values according to the base CNN are ignored, and these regions give higher values of regularized loss. Both dense CRF and our approach use CRF models. However, as opposed to post-processing, we use CRF for CNN supervision, enabling CNN to learn a better segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Our experiments on the standard benchmarks show that test time adaptation significantly improves performance, achieving the new state-of-art in image level weakly supervised SOD." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This paper is organized as follows: Sec. 2 is related" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.402 + ], + "angle": 0, + "content": "work, Sec. 3 explains the approach in [33], Sec. 4 describes our approach, and Sec. 5 presents experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.418, + 0.642, + 0.433 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.444, + 0.813, + 0.459 + ], + "angle": 0, + "content": "Weakly Supervised Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Most prior work on image level weak supervision consists of two stages. In the first stage, pseudo labels for SOD task are generated, and in the second stage, CNN is trained on the pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.893, + 0.702 + ], + "angle": 0, + "content": "One of the first methods to use image level weak supervision for SOD is in [34]. Their method generates pseudo labels by training CNN for object category prediction. In their setting, image level supervision consists of a large set of 200 object classes categories from Image net. After training for object class prediction, they extract foreground heat maps capturing potential object regions. These generalize to unseen categories, and provide initial pseudo labels. Next, they have a self-training stage which alternates between estimating pseudo labels and training CNN on them. For more accurate pseudo-labels, the predicted pseudo-labels are refined with dense CRF [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Later image level SOD approaches [17,25,27,40,41,44] use one or more conventional weak (i.e. unsupervised) saliency methods [4,9,10,19] for pseudo labels. In [40], they use multiple weak saliency methods and fuse their results in pseudo-ground truth. During the fusion process, based on the difficulty of training data, they estimate the corresponding confidence measure and utilize it for training with their pseudo ground truth to better handle label noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In [44] they develop a noise modeling module, which enables them to deal with noisy saliency maps obtained from multiple weak saliency methods in a probabilistic way. They later extend their method to rely only on a single weak saliency method [41]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.514, + 0.937 + ], + "angle": 0, + "content": "7361" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "In [17], they approach the noise in pseudo labels by exploiting dense CRF [13]. Initial pseudo labels are generated by a weak saliency method. Then they alternate applying dense CRF and CNN training. Dense CRF corrects the noisy labels through spatial consistency and structure preservation, while CNN is trained on the current version of pseudo labels to update and improve the current model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.319 + ], + "angle": 0, + "content": "In [25], they develop a curriculum framework to incrementally refine the pseudo-labels. Instead of directly using the saliency maps produced by weak saliency methods for training, they first train CNNs to generate pseudo labels for each weak saliency method. This improves the pseudo labels as CNNs are forced to learn representations across a broad set of images. Then pseudo labels are further refined via an iterative self-supervision technique." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.468, + 0.381 + ], + "angle": 0, + "content": "In [27] they utilize multiple pseudo labels to achieve robustness to noise. They design a multi-filter directive CNN which uses multiple directive filters and a multi-guidance loss to integrate multiple pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.468, + 0.488 + ], + "angle": 0, + "content": "In [42] they propose a noise-aware encoder-decoder to deal with noisy pseudo labels. Their method has a saliency predictor that maps input images to clean saliency maps, and a noise generator. The model that represents noisy labels is a sum of these two models. They train their model to simultaneously infer the corresponding latent vector of each noisy label and the saliency predictor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.468, + 0.639 + ], + "angle": 0, + "content": "All approaches above develop elaborate techniques to deal with noisy pseudo labels. In contrast to the above approaches, in [33] they develop a method for image level supervised SOD which does not rely on pseudo labels from conventional weak saliency methods. The main tool in [33] is regularized loss, based on sparse CRF [2]. The approach in [33] uses off-the-shelf Unet [29] architecture, consists of a single stage, has an intuitive and easy to interpret loss function, and outperforms other image level supervised SOD methods. We review [33] in detail in Sec 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.468, + 0.76 + ], + "angle": 0, + "content": "In addition to image level weakly supervised SOD methods, there are approaches that use additional sources of supervision. They do not necessarily rely on conventional weak saliency methods. In [39], they use additional weak sources, such as captions, etc. In [38,43] they use scribbles, and in [21] they use boxes, both much stronger forms of weak supervision. We achieve comparable or better results without additional weak sources." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.203, + 0.777 + ], + "angle": 0, + "content": "Regularized Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Regularized loss for weak supervision was first used in [31], for semantic segmentation supervised with scribbles. It was subsequently generalized in [24, 32]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.825, + 0.235, + 0.839 + ], + "angle": 0, + "content": "Test Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The idea of test time adaptation has been used before. For example, in the context of traditional computer vision, it was used in [30] for online updates in visual tracking and in [11] for adapting detectors from images to videos. In the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "context of deep learning, it was used in [45] for person re-identification, in [16] for image inpainting, and in [22] for enforcing consistency constraints in depth estimation. In all these methods, the loss to perform test time adaptation is application specific and is unrelated to our approach." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.181, + 0.744, + 0.198 + ], + "angle": 0, + "content": "3. Regularized Loss for SOD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.312 + ], + "angle": 0, + "content": "In this section, we review the approach in [33], which is designed for image level weakly supervised segmentation for datasets with a single object class. It naturally applies to SOD, since there is only one class, the salient object. The main idea is to design a regularized loss which incorporates the likely properties of generic objects and to use this loss to train CNN instead of cross entropy on ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Regularized loss is applied to the output of CNN, which is the same size as the input image. Let \\( x \\) denote CNN output, and \\( x_{p} \\) denote the output for pixel \\( p \\). The last layer of CNN is sigmoid, so that \\( x_{p} \\in [0,1] \\). Background corresponds to 0 and salient object to 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.433 + ], + "angle": 0, + "content": "Regularized loss in [33] consists of a weighted combination of several components. The most important component is sparse CRF loss [2]." + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.443, + 0.892, + 0.483 + ], + "angle": 0, + "content": "\\[\nL _ {\\operatorname {c r f}} (\\boldsymbol {x}) = \\frac {1}{| \\mathcal {P} |} \\sum_ {(p, q) \\in \\mathcal {N}} e ^ {- \\frac {| | C _ {p} - C _ {q} | | ^ {2}}{2 \\sigma^ {2}}} \\cdot | x _ {p} - x _ {q} |, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.554 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) is the set of all pixels in the image, \\(\\mathcal{N}\\) is the set of neighboring pixel pairs on a 4-connected grid, and \\(C_p \\in \\mathbb{R}^3\\) is the color of \\(p\\). Parameter \\(\\sigma\\) controls the edge strength and is set to \\(\\sigma = 0.15\\) in all experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Usually CNN produces a sharp distribution, i.e. most \\( x_{p} \\) are close either to 0 or 1. Thus if two neighboring pixels are not assigned to the same class, there is a penalty which depends on the strength of the edge between them. Sparse CRF loss is low when segment boundary aligns with image intensity edges. Salient objects are likely to cause image intensity edges, and thus sparse CRF is the main driving tool for salient object discovery. The lowest value of sparse-CRF is zero, achieved at trivial solutions: empty, everything classified as object, or full, everything classified as background. Thus one cannot train with sparse CRF alone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Another component of regularized loss is batch volumetric loss. It encourages the average object size in a batch of images to be half of the image size. This loss is useful to prevent both empty and full trivial solutions. Averaging over a batch makes the loss less strict: some batch objects can be significantly smaller or larger than half of the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The next component of regularized loss is minimum volume loss. It encourages the object to be of at least a certain size. It is more likely to be valid in practice, compared to volumetric batch loss, as objects can be expected to be at least a certain minimum size. The last two components of regularized loss are border and center losses. The border" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7362" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "loss encourage the image border to be the background, and the center loss encourages image center be the object. This is a realistic assumption for many, but not all, images containing a salient object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Training CNN with regularized loss is difficult and tends to get stuck in a bad local minimum. In [33] they devise a two-stage strategy for training. For our approach, we start with base CNN trained with the method developed in [33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.469, + 0.306 + ], + "angle": 0, + "content": "In our test-time adaptation, we only use sparse CRF loss in Eq. (1), minimum volume loss, and border loss. We now describe minimum volume and border loss in detail. Let \\(\\bar{\\pmb{x}} = \\frac{1}{|\\mathcal{P}|}\\sum_{p\\in \\mathcal{P}}x_p\\), i.e. the normalized object size. Minimum volume loss \\(L_{\\mathrm{m}}\\) penalizes segmentations if the normalized object size is less than \\(\\mathrm{obj}_{min}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.315, + 0.469, + 0.332 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {m}} (\\boldsymbol {x}) = R e L U ^ {2} \\left(\\operatorname {o b j} _ {\\min } - \\bar {\\boldsymbol {x}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.343, + 0.334, + 0.359 + ], + "angle": 0, + "content": "Following [33], we set \\(\\mathrm{obj}_{\\min} = 0.15\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.469, + 0.389 + ], + "angle": 0, + "content": "Let \\(\\mathcal{B}\\) be the set of pixels on the image border of width \\(w = 3\\). The border loss \\(L_{\\mathrm{b}}(\\pmb{x})\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.399, + 0.469, + 0.435 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {b}} (\\boldsymbol {x}) = \\left(\\frac {1}{| \\mathcal {B} |} \\sum_ {p \\in \\mathcal {B}} x _ {p}\\right) ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.446, + 0.285, + 0.464 + ], + "angle": 0, + "content": "4. Test Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.469, + 0.517 + ], + "angle": 0, + "content": "We now describe our approach. We start with an overview in Sec. 4.1. We describe our regularized loss in Sec. 4.2, and our hyperparameter setting method in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.188, + 0.54 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.469, + 0.76 + ], + "angle": 0, + "content": "The overview of our approach is in Fig. 1. Given a test image \\( I \\), we first construct a small training dataset \\( D(I) \\) from \\( I \\) by data augmentation. There are various forms of augmentation. However, since we are training without pixel precise ground truth, we need to avoid augmentations which make the salient object more difficult to detect in an image. For example, large crops can remove the salient object. Similarly, adding random image noise changes the distribution of intensity edges, and sparse CRF in Eq. (1) may become less useful for object discovery. We use small random vertical and horizontal image shifts, and random additive shift to color channels (while keeping the shift the same for all image pixels). We found random vertical and horizontal shifts to be the most effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.469, + 0.866 + ], + "angle": 0, + "content": "Next, we take the base CNN trained for SOD task in image level weakly supervised setting, according to the method proposed in [33], and train it on \\( D(I) \\) for a small number of epochs using the regularized loss approach described in Sec. 4.3. The resulting CNN is called the adapted CNN. Training for a small number of iterations ensures that the adapted CNN retains most of the information useful for" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "SOD learned from the large training dataset, as well as ensuring efficiency. In practice, we train for 10 epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.132, + 0.669, + 0.148 + ], + "angle": 0, + "content": "4.2. Regularized Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.276 + ], + "angle": 0, + "content": "We now describe our loss function. For test-time adaptation, we only use some of the regularized loss components from [33]. We do not use batch volumetric loss as it encourages the salient segment to be half of the image size and is far from realistic for most test images. We also do not use center loss. It assumes that the image center belongs to the salient object and is likely not true for many specific test images we wish to adapt CNN for." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.413 + ], + "angle": 0, + "content": "In our regularized loss, we use sparse CRF, minimum volume, and border losses, Eq. (1, 2, 3). Sparse CRF is the most important loss component for discovering salient objects, as it favours segments that align with image intensity edges. Minimum volume loss is necessary to prevent collapse to an empty solution favoured by sparse CRF. Border loss encourages image border to be assigned to the background, and is a realistic assumption for most test images. The complete regularized loss for test time adaptation is" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.441, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\[\nL _ {\\operatorname {r e g}} (\\boldsymbol {x}) = \\lambda_ {\\operatorname {c r f}} L _ {\\operatorname {c r f}} (\\boldsymbol {x}) + \\lambda_ {m} L _ {\\mathrm {m}} (\\boldsymbol {x}) + \\lambda_ {\\mathrm {b}} L _ {\\mathrm {b}} (\\boldsymbol {x}). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.788, + 0.485 + ], + "angle": 0, + "content": "4.3. Hyperparameter Setting Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Consider the loss components in Eq. (4). To get the best performance during test-time adaptation, it is important to choose a good setting of their relative weights. Each test image has its own best setting, however, we cannot find this setting as the ground truth is not available. Thus we must develop a hyperparameter setting method that does not rely on ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Out of three components in Eq. (4), the border loss is the least important as it contributes the least to the salient object discovery. Indeed, if our only knowledge about a segmentation is that the image border is assigned to the background, then this gives us almost no information to judge whether this segmentation corresponds to the salient object. Therefore, we set \\(\\lambda_{b}\\) to a low value of 1. Out of the two remaining components, sparse CRF is more important for object discovery. Since objects tend to cause intensity edges in an image, a segment that aligns to image edges is more likely to correspond to a salient object, compared to a segment that just obeys the minimum size constraints. Thus sparse CRF weight \\(\\lambda_{\\mathrm{crf}}\\) should be much higher compared to the minimum volume weight \\(\\lambda_{m}\\). At the same time, the minimum volume weight should be sufficiently large to prevent the collapse to a trivial empty solution. We can compute this sufficiently large weight from the solution obtained from our base CNN." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We now explain how to set \\(\\lambda_{\\mathrm{m}}\\) relative to \\(\\lambda_{\\mathrm{crf}}\\) so that a trivial solution is avoided. Let \\(x^{i}\\) be the output of the base" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2The width of the border can be adjusted relative to the image size, but we train on fixed size images, \\(256\\times 256\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7363" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.166 + ], + "angle": 0, + "content": "CNN on input image \\(I\\). Assuming that \\(\\pmb{x}^i\\) is not empty, we set the weight of minimum volume loss large enough to make the loss on \\(\\pmb{x}^i\\) smaller than the loss of an empty solution. Then switching to an empty solution from the initial \\(\\pmb{x}^i\\) is too costly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.287 + ], + "angle": 0, + "content": "The results produced by the base CNN tend to be sharp, that is most pixels have values close to either 0 or 1. Therefore we will assume that the trivial empty solution corresponds to all pixels having value 0 in the output. Let \\( \\pmb{x}^0 \\) denote such a trivial empty solution, i.e. \\( x_p^0 = 0 \\) for all pixels \\( p \\). Suppose we wish to set sparse-CRF weight to \\( \\lambda_{\\mathrm{crf}} \\). We need to set \\( \\lambda_m \\) to a value that makes collapse to an empty solution too costly. For this, we need \\( \\lambda_m \\) that satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.294, + 0.468, + 0.312 + ], + "angle": 0, + "content": "\\[\nL _ {\\operatorname {r e g}} \\left(\\boldsymbol {x} ^ {i}\\right) = L _ {\\operatorname {r e g}} \\left(\\boldsymbol {x} ^ {0}\\right) + \\epsilon , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.43, + 0.332 + ], + "angle": 0, + "content": "where \\(\\epsilon > 0\\) is a small constant, set in practice to 0.1." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.333, + 0.353, + 0.348 + ], + "angle": 0, + "content": "Plugging \\( \\pmb{x}^0 \\) in Eqs.(1, 2, 3) we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.353, + 0.469, + 0.372 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {b}} \\left(\\boldsymbol {x} ^ {0}\\right) = 0, \\quad L _ {\\mathrm {c r f}} \\left(\\boldsymbol {x} ^ {0}\\right) = 0, \\quad L _ {\\mathrm {m}} \\left(\\boldsymbol {x} ^ {0}\\right) = \\operatorname {o b j} _ {\\min } ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.378, + 0.469, + 0.408 + ], + "angle": 0, + "content": "Plugging Eq. (6) into Eq. (5) and solving for \\(\\lambda_{m}\\), we get our formula for setting \\(\\lambda_{m}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.414, + 0.469, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {m} = \\frac {\\lambda_ {\\mathrm {c r f}} L _ {\\mathrm {c r f}} \\left(\\boldsymbol {x} ^ {i}\\right) + L _ {\\mathrm {b}} \\left(\\boldsymbol {x} ^ {i}\\right) - \\epsilon}{\\operatorname {o b j} _ {\\min } ^ {2} - L _ {\\mathrm {m}} \\left(\\boldsymbol {x} ^ {i}\\right)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.469, + 0.697 + ], + "angle": 0, + "content": "Eq. (7) applies when the output \\( \\boldsymbol{x}^i \\) of the base CNN is not a trivial empty solution. If \\( \\boldsymbol{x}^i \\) is an empty solution, it means that the base CNN fails to extract a salient object. In this case, we set both CRF and minimum volume loss to be of equal weight, \\( \\lambda_{\\mathrm{crf}} = \\lambda_m \\). This reflects the fact that extracting a salient object from the current image is difficult and the penalty for not satisfying the minim object size is increased. We do not set \\( \\lambda_{\\mathrm{m}} \\) higher than \\( \\lambda_{\\mathrm{crf}} \\) as some images may not contain a salient object, in which case an empty solution is appropriate. In addition, while setting \\( \\lambda_{\\mathrm{m}} \\) high enough would result in some non-empty solution, most likely it would be an erroneous solution, as minimum volume loss is less useful for extracting object segments and, therefore, should not have a weight higher than that of sparse CRF. In practice, we set \\( \\lambda_{\\mathrm{crf}} = 10^3 \\) and compute \\( \\lambda_{\\mathrm{m}} \\) for each test image using Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.469, + 0.893 + ], + "angle": 0, + "content": "In Table 1 we experimentally evaluate the effectiveness of our method for setting \\(\\lambda_{\\mathrm{m}}\\). We randomly select 500 images from DUTO [37] dataset. We test our adaptive \\(\\lambda_{\\mathrm{m}}\\) for each image, and the same fixed setting of \\(\\lambda_{\\mathrm{m}}\\) for all test images, in a range from 1 to \\(10^{3}\\), see Table 1. For this experiment, we use a fixed \\(\\lambda_{\\mathrm{crf}} = 10^{3}\\). The performance metric is \\(F_{\\beta}\\). The adaptive \\(\\lambda_{\\mathrm{m}}\\) performs better than any fixed choice of \\(\\lambda_{\\mathrm{m}}\\). As expected, the performance degrades significantly for the larger values of \\(\\lambda_{\\mathrm{m}}\\) as minimum volume loss starts to play a bigger role in the regularized loss function, however, it is less discriminating of salient objects. For smaller \\(\\lambda_{\\mathrm{m}}\\), the performance suffers as the results on some images collapse to an empty solution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.706, + 0.108 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.221 + ], + "angle": 0, + "content": "Our implementation is in Pytorch using RTX 2080 GPU. For our base CNN, we use image level weakly supervised SOD approach [33]. We use Unet [29] architecture with ResNeXt [35] fixed features pretrained onImagenet [5] and train on \\(256 \\times 256\\) images. Unlike [33], we train on larger images and on a larger training dataset, thus getting better results for the base CNN than those reported in [33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.892, + 0.312 + ], + "angle": 0, + "content": "We train the base CNN on the training set of DUTS [34], which contains 10,553 images. We then evaluate our test-time CNN adaptation approach on the test set of DUTS (5019 images), DUT-OMRON [37] (5168 images), ECSSD [36] (1000 images), THUR [3] (6233 images), PascalS [19] (850 images), and HKU-IS [18] (4447 images)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.893, + 0.508 + ], + "angle": 0, + "content": "For test time adaptation, given a test image \\( I \\), we construct a test time adaptation dataset \\( D(I) \\) that has 16 images. These images are constructed from \\( I \\) by shifting to the left or right with probability 0.5, and, in addition, shifting up or down with probability 0.5. Each shift is chosen uniformly from the range [0, 15]. This creates a dataset of images similar to \\( I \\), and it is unlikely that the salient object is removed out of the sample by these small shifts. We use Adam optimizer [12], a fixed learning rate 0.001, and batch size equal to the size of the test time adaptation dataset, set to 16 images. We train for 10 epochs. The images are scaled to resolution \\( 256 \\times 256 \\). Test time adaptation takes 2.8 seconds per image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.644 + ], + "angle": 0, + "content": "We use the standard metrics widely adopted for SOD. Our first metric is \\( F_{\\beta} \\) score [1], with \\( \\beta^2 = 0.3 \\). We also use mae [26], defined as the average absolute per pixel difference between the predicted saliency and ground truth. In addition, we use two newer metrics, namely \\( S_{\\alpha} \\) [6], and \\( E_{s} \\) [7]. \\( S_{\\alpha} \\) evaluates the structural information for saliency map and region-aware and object-aware structural similarity between saliency maps and ground truth. \\( E_{s} \\) is based on unification of global and local information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.63, + 0.667 + ], + "angle": 0, + "content": "5.1. SOD results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In Table 2, we report the results of our approach and compare to prior image level weakly supervised SOD methods. Our results with fixed hyperparameter settings in Eq. (4) are in column \"ours\" (with \\(\\lambda_{\\mathrm{m}} = 20\\), which is the best fixed setting according to Table 1). The results with our method for hyperparameter setting for each image individually, as described in Sec. 4.3, are in column 'ours + h'. Except one case, all metrics for our approach with hyperparameter setting method are improved, often significantly. This shows that our hyperparameter setting method is crucial for obtaining a significantly improved performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dense CRF [13] is often used to post-process CNN results. Therefore, we also post-process base CNN results with dense CRF for comparison. The performance of the base CNN is in column [33], and its performance after post-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7364" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.225, + 0.089, + 0.744, + 0.123 + ], + "angle": 0, + "content": "
λm151020501002005001000adaptive
Fβ.741.738.752.755.742.732.728.718.703.765
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.133, + 0.892, + 0.163 + ], + "angle": 0, + "content": "Table 1. Performance in terms of \\(F_{\\beta}\\) score of fine tuning with a fixed value of \\(\\lambda_{m}\\) vs. our adaptive \\(\\lambda_{\\mathrm{m}}\\) computation according to Eq. (7). For all experiments, \\(\\lambda_{\\mathrm{crf}} = 10^{3}\\)." + }, + { + "type": "table", + "bbox": [ + 0.184, + 0.177, + 0.787, + 0.561 + ], + "angle": 0, + "content": "
Metric[34][17][39][44][42][27][33][33]+dCRFoursours+h
DUTS\\( F_{\\beta} \\)↑.654.614.684.725.747.710.753.763.771.795
mae↓.100.116.091.075.060.076.056.055.054.052
\\( E_s \\)↑.795.772.814.853.859.839.833.842.855.863
\\( S_{\\alpha} \\)↑.748.697.759.813.827.775.791.798.800.817
ECSSD\\( F_{\\beta} \\)↑.823.797.609.810.852.854.868.873.882.911
mae↓.104.110.109.092.071.084.063.060.057.046
\\( E_s \\)↑.869.853.673.836.883.885.884.886.885.909
\\( S_{\\alpha} \\)↑.811.802.756.846.860.834.845.855.859.883
DUTO\\( F_{\\beta} \\)↑.603.622.609.597.701.646.697.710.721.752
mae↓.109.101.109.103.070.087.074.069.066.057
\\( E_s \\)↑.768.776.673.712.816.803.821.827.833.842
\\( S_{\\alpha} \\)↑.725.752.756.733.791.742.763.771.782.798
PascalS\\( F_{\\beta} \\)↑.715.693.713--.751.796.801.801.807
mae↓.139.149.135--.115.083.082.082.084
\\( E_s \\)↑.791.772.790--.817.827.827.826.828
\\( S_{\\alpha} \\)↑.744.717.768--.770.791.793.792.796
HKUIS\\( F_{\\beta} \\)↑--.814.820.878.851.866.871.883.899
mae↓--.084.065.043.059.044.040.039.037
\\( E_s \\)↑--.895.858.919.921.912.925.926.931
\\( S_{\\alpha} \\)↑--.818.860.890.846.857.866.871.879
THUR\\( F_{\\beta} \\)↑---.694.719-.731.749.753.760
mae↓---.086.070-.073.069.067.064
\\( E_s \\)↑---.807.838-.840.842.843.844
\\( S_{\\alpha} \\)↑---.804.810-.794.798.806.819
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.572, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Table 2. Comparison to SOD methods that also use image level weak supervision using metrics \\(F_{\\beta}\\), \\(mae\\), \\(E_s\\), and \\(S_{\\alpha}\\). We use \\(\\uparrow\\) to indicate that the higher score is better, and \\(\\downarrow\\) to indicate the lower score is better. The best result is in bold, and the second best is underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.703 + ], + "angle": 0, + "content": "processing with dense CRF is in column [33] + dCRF. Our approach is significantly better than [33] with dense CRF post-processing. This illustrates that it is important to include CRF during training in the loss function, rather than use CRF for just post-processing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Our test time adaptation improves the performance of [33] across all metrics, except one case. For easier datasets (ECSSD, HKUIS), the performance gain is not as large as the base CNN already performs well, but for harder to fit datasets, the performance gain is significant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Comparing to other SOD methods, our method is better in almost all cases except \\(S_{\\alpha}\\) metric for HKUIS and DUTS. For most datasets, our approach significantly improves \\(F_{\\beta}\\) metric over the second best method, notably by roughly \\(7\\%\\) for DUTO and DUTS datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In Table 3, we compare our method to weakly supervised SOD methods that use stronger form of weak supervision. In [21], they use bounding box supervision, and in [38, 43]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.825 + ], + "angle": 0, + "content": "they use scribbles. Our method is better than the bounding box method [43] in all cases. Our method is almost always better than [43], and better in about half of the cases compared to [38]. If Fig. 2, we show qualitative comparisons of our method to the scribble supervised methods of [43] and [38]. For this illustration, we chose examples where all methods work relatively well, in terms of the standard metrics. However, our results have better detail preservation qualities, especially when the salient object has sharp edges on the boundary. Observe the detail preservation in the chair, in the last row. This is because CRF based loss function is adapted to the particular test image, driving the segmentation boundary to align with image edges." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.707, + 0.849 + ], + "angle": 0, + "content": "5.2. Ablation Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.856, + 0.771, + 0.871 + ], + "angle": 0, + "content": "Hyperparameters in Regularized Loss" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We first discuss and perform ablation for hyperparameters of our regularized loss function, Eq. (4). The most im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7365" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.21, + 0.088, + 0.764, + 0.354 + ], + "angle": 0, + "content": "
DUTSECSSDDUTOPascalSHKUISTHUR
boxes [21]Fβ↑.736.860.686-.853-
mae ↓.079.072.081-.058-
Es ↑.831.889.810-.897-
Sα ↑.789.858.776-.852-
scribbles [43]Fβ↑.747.865.715.788.858.718
mae ↓.062.061.068.140.047.077
Es ↑.865.908.835.798.923.873
Sα ↑------
scribbles [38]Fβ↑.823.900.758.823.896.755
mae ↓.049.049.060.078.038.069
Es ↑.890.908.862.847.938.843
Sα ↑------
oursFβ↑.795.911.752.807.899.760
mae ↓.052.046.057.082.037.064
Es ↑.863.909.842.828.931.844
Sα ↑.817.883.798.796.879.819
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.363, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Table 3. Comparison to prior SOD methods that use stronger forms of weak supervision using metrics \\(F_{\\beta}\\), \\(mae\\), \\(E_s\\), and \\(S_{\\alpha}\\). In [21], they use bounding box supervision, and in [38,43], they use scribbles. We use \\(\\uparrow\\) to indicate that the higher score is better, and \\(\\downarrow\\) to indicate the lower score is better. The best result is in bold, and second best is underlined." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.433, + 0.477, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.756, + 0.471, + 0.813 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparison. From left to right: input image, ground truth, results of [43], [38], and our results, respectively. Note that [43], [38] both use scribble supervision, a much stronger supervision form compared to what we use." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "important parameter to set is \\(\\lambda_{m}\\) in relation to \\(\\lambda_{\\mathrm{crf}}\\), as the correct relative settings of these two parameters ensures there is no collapse to a trivial solution. For this, we have de" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.429, + 0.866, + 0.543 + ], + "angle": 0, + "content": "
with border losswithout border loss
DUTS.795.789
ECSSD.911.910
DUTO.752.748
PascalS.807.800
HKUIS.899.892
THUR.760.755
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.554, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Table 4. Performance of our approach with border loss and without border loss. Performance metric is \\( F_{\\beta} \\) score." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.611, + 0.892, + 0.687 + ], + "angle": 0, + "content": "veloped an effective automatic approach, Sec. 4.3. Since \\(\\lambda_{m}\\) is computed in relation to \\(\\lambda_{\\mathrm{crf}}\\), the setting of \\(\\lambda_{\\mathrm{crf}}\\) can be arbitrary (up to the precision that can be handled by the software), \\(\\lambda_{\\mathrm{m}}\\) is adjusted accordingly: for larger \\(\\lambda_{\\mathrm{crf}}\\), \\(\\lambda_{\\mathrm{m}}\\) will be automatically set larger according to Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.893, + 0.854 + ], + "angle": 0, + "content": "The setting of \\(\\lambda_{b}\\), as long as it is much smaller than \\(\\lambda_{\\mathrm{crf}}\\), is not important. Parameter \\(\\lambda_{\\mathrm{b}}\\) controls the border loss that encourages border pixels to be assigned to the background. While this loss is helpful, it is the least important regularized loss component. To show that this loss is the least important, we perform an experiment where we remove border loss from regularized loss (\\(\\lambda_{b} = 0\\)). The results are summarized in Table 4. Without border loss, the results worsen only a little. On the other hand, omitting either CRF or minimum volume losses from our loss function in Eq. (4) results in a catastrophic performance collapse." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In Eq. (1), we set \\(\\sigma = 0.15\\) as in [33]. Since the training images are normalized, this setting works well across all datasets. We also experimented with other values of \\(\\sigma\\), see" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7366" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.413, + 0.194 + ], + "angle": 0, + "content": "
σ = 0.1σ = 0.15σ = 0.20
DUTS.789.795.797
ECSSD.894.911.915
DUTO.751.752.743
PascalS.803.807.803
HKUIS.902.899.895
THUR.763.760.757
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.204, + 0.47, + 0.233 + ], + "angle": 0, + "content": "Table 5. Performance of our approach for different \\( \\sigma \\) values. Performance metric is \\( {F}_{\\beta } \\) score." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.247, + 0.443, + 0.294 + ], + "angle": 0, + "content": "
|D(I)|248163264
Fβ↑.879.887.890.910.915.913
mae ↓.058.055.049.046.045.046
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.303, + 0.468, + 0.331 + ], + "angle": 0, + "content": "Table 6. Different choice of test time dataset size for ECSSD dataset." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.345, + 0.446, + 0.393 + ], + "angle": 0, + "content": "
# epoch2510203050
Fβ↑.873.883.910.915.915.912
mae ↓.062.053.047.045.046.047
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.403, + 0.468, + 0.43 + ], + "angle": 0, + "content": "Table 7. Different number of test time adaptation iterations for ECSSD dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.463, + 0.473 + ], + "angle": 0, + "content": "Table 5. The performance is stable for a range of \\( \\sigma \\) values." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.303, + 0.501 + ], + "angle": 0, + "content": "5.3. Other Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.468, + 0.628 + ], + "angle": 0, + "content": "We now perform ablation experiments for various hyperparameter settings (other than those in the regularized loss function) of our approach. We perform experiments on ECSSD dataset, and evaluate the sensitivity of our approach to the size of the test time adaptation dataset \\( D(I) \\), the number of epoch performed for test time adaptation, and different data augmentation strategies. For all the experiments, the performance metric is \\( F_{\\beta} \\) and \\( mae \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.704 + ], + "angle": 0, + "content": "In Table 6, we show how performance depends on varying the size of the test time dataset \\( D(I) \\). Our standard setting for the size of \\( D(I) \\) is 16. Decreasing it to 8 slightly degrades the performance, increasing it to 64 results in almost no change, but the running time is twice longer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.795 + ], + "angle": 0, + "content": "In Table 7, we show how performance depends on increasing the number of epoch for test time adaptation. Our standard setting for the number of epoch is 10. Decreasing it to 2 noticeably degrades the performance, increasing it to 50 results in only a slight improvement but significantly increases the running time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In Table 8, we show the effect of different augmentation strategies. Our main experiments use small random shifts in the vertical and horizontal directions. We also tested adding random Gaussian noise to images, a standard data augmentation technique, but it significantly degrades the performance. Adding Gaussian noise decreases the effectiveness of our sparse CRF loss as neighboring pixels become less" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.924, + 0.135 + ], + "angle": 0, + "content": "
typeshiftsGauss noisecolor perturbationshifts+color
Fβ↑.911.854.897.909
mae↓.046.060.050.049
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.145, + 0.892, + 0.172 + ], + "angle": 0, + "content": "Table 8. Different augmentation strategies when fine tuning on ECSSD dataset." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.193, + 0.943, + 0.247 + ], + "angle": 0, + "content": "
DUTSECSSDDUTOPascalSHKUISTHUR
% improved79.383.785.381.276.775.0
mean Fβ increase.048.053.064.17.037.036
mean Fβ decrease.021.009.011.08.019.015
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.258, + 0.892, + 0.312 + ], + "angle": 0, + "content": "Table 9. Percentage of images for which performance improves as the result of test-time adaptation. For the improved (and degraded) images, we also show the mean increase (and the decrease) in \\( F_{\\beta} \\) score." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.346, + 0.892, + 0.422 + ], + "angle": 0, + "content": "strongly connected. We also tested random color perturbation where all pixels change the value of a color channel by the same random value. This works much better than Gaussian noise, but not as well as random shifts. Shifts and color perturbation work as well as shifts alone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.722, + 0.457 + ], + "angle": 0, + "content": "5.4. Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.467, + 0.892, + 0.664 + ], + "angle": 0, + "content": "During test-time adaptation, the majority, but not all images, improve. In Table 9, we show the percentage of images that improve, in terms of their \\( F_{\\beta} \\) score, for each of the test datasets. In addition, we show the mean improvement of \\( F_{\\beta} \\) score, averaged over the images that improve, and the mean decline of \\( F_{\\beta} \\) score, averaged over the images that decline in performance. As can be seen from the table, for all dataset, over \\( 75\\% \\) of images improve during test-time adaptation. Also, the average improvement is significantly larger than the average decline of the results. Thus overall, test-time adaptation significantly improves the performance. In the supplementary materials, we show the most improved and the most degraded images during test time adaptation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.627, + 0.707 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose the first approach to test time adaptation for SOD with image level weak supervision. We develop a regularized loss function particularly effective for test time adaptation, and a method to find an appropriate setting of the hyperparameters in our loss function. Our approach uses a standard CNN architecture and a loss function that is intuitive and simple to interpret. We achieve a new state of the art in image level supervised SOD, and a better or competitive result with the prior work that uses stronger forms of weak supervision. The drawback of our approach is that the running time for testing is more expensive, since the base CNN needs to be fine-tuned for each test image." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7367" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Yuri Boykov, Olga Veksler, and Ramin Zabih. A variable window approach to early vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1283-1294, December 1998. An earlier version of this work appeared in CVPR '97. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.47, + 0.285 + ], + "angle": 0, + "content": "[3] Ming-Ming Cheng, NiloyJ. Mitra, Xiaolei Huang, and ShiMin Hu. Salientshape: group saliency in image collections. The Visual Computer, 30(4):443-453, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[6] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[7] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 698-704, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[8] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT Press, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.598 + ], + "angle": 0, + "content": "[9] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[10] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[11] Vidit Jain and Erik Learned-Miller. Online domain adaptation of a pre-trained cascade of classifiers. In CVPR 2011, pages 577-584. IEEE, 2011. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[13] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In Neural Information Processing Systems, pages 109-117, 2011. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[14] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Neural Information Processing Systems, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Comput., 1(4):541-551, December 1989. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[16] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] G. Li and Y. Yu. Deep contrast learning for salient object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 478-487, June 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[19] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 280-287, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[20] Jiawei Liu, Jing Zhang, and Nick Barnes. Semi-supervised salient object detection with effective confidence estimation. arXiv preprint arXiv:2112.14019, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[21] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE Transactions on Image Processing, 30:4423-4435, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[22] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[23] Yunqiu Lv, Bowen Liu, Jing Zhang, Yuchao Dai, Aixuan Li, and Tong Zhang. Semi-supervised active salient object detection. Pattern Recognition, 123:108364, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[24] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6608-6618, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[25] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. In Advances in Neural Information Processing Systems, pages 204-214, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[26] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In Conference on Computer Vision and Pattern Recognition, pages 733-740. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[27] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[28] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] David A Ross, Jongwoo Lim, Ruei-Sung Lin, and Ming-Hsuan Yang. Incremental learning for robust visual track" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7368" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ing. International journal of computer vision, 77(1):125-141, 2008. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[31] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised CNN segmentation. In Conference on Computer Vision and Pattern Recognition, pages 1818-1827, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[32] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised CNN segmentation. In European Conference on Computer Vision, pages 524-540, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[33] Olga Veksler. Regularized loss for weakly supervised single class semantic segmentation. In European Conference on Computer Vision, pages 348-365. Springer, 2020. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.434 + ], + "angle": 0, + "content": "[35] Saining Xie, Ross B. Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.47, + 0.476 + ], + "angle": 0, + "content": "[36] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 1155-1162. IEEE, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.47, + 0.533 + ], + "angle": 0, + "content": "[37] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Conference on Computer Vision and Pattern Recognition, pages 3166-3173. IEEE, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[38] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3234-3242. AAAI Press, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[39] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 6074-6083, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[40] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[41] Jing Zhang, Yuchao Dai, Tong Zhang, Mehrtash Harandi, Nick Barnes, and Richard Hartley. Learning saliency from single noisy labelling: A robust model fitting perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2866-2873, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[42] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "conference on computer vision, pages 349-366. Springer, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[43] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In Conference on Computer Vision and Pattern Recognition, pages 12546-12555, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[44] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9029-9038, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[45] Shun Zhang, Jia-Bin Huang, Jongwoo Lim, Yihong Gong, Jinjun Wang, Narendra Ahuja, and Ming-Hsuan Yang. Tracking persons-of-interest via unsupervised representation adaptation. International Journal of Computer Vision, 128(1):96-120, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[46] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3085-3094, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[47] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.927, + 0.516, + 0.938 + ], + "angle": 0, + "content": "7369" + } + ] +] \ No newline at end of file diff --git a/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_origin.pdf b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ccf2f513e11bea3076572381eded1d30de7455c --- /dev/null +++ b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/2441aa83-753b-4443-ba38-3993bb6d5100_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0af1cb9a6a367f15adf863eb74ed0220a33bc027a50b6ce0420e8ad8d959320 +size 789379 diff --git a/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/full.md b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c2ee9090421025aed12497e0d2e34dea31b48fcd --- /dev/null +++ b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/full.md @@ -0,0 +1,347 @@ +# Test Time Adaptation with Regularized Loss for Weakly Supervised Salient Object Detection + +Olga Veksler +University of Waterloo, Canada +oveksler@uwaterloo.ca + +# Abstract + +It is well known that CNNs tend to overfit to the training data. Test-time adaptation is an extreme approach to deal with overfitting: given a test image, the aim is to adapt the trained model to that image. Indeed nothing can be closer to the test data than the test image itself. The main difficulty of test-time adaptation is that the ground truth is not available. Thus test-time adaptation, while intriguing, applies to only a few scenarios where one can design an effective loss function that does not require ground truth. We propose the first approach for test-time Salient Object Detection (SOD) in the context of weak supervision. Our approach is based on a so-called regularized loss function, which can be used for training CNN when pixel precise ground truth is unavailable. Regularized loss tends to have lower values for the more likely object segments, and thus it can be used to fine-tune an already trained CNN to a given test image, adapting to images unseen during training. We develop a regularized loss function particularly suitable for test-time adaptation and show that our approach significantly outperforms prior work for weakly supervised SOD. + +# 1. Introduction + +A well known problem with CNNs is that they tend to overfit to the training data. One approach to deal with overfitting is test-time adaptation [30]. A model is trained on the training data, and then fine-tuned for a few epochs during test time on a given test image. The main difficulty of test-time adaptation is that there is no ground truth for the test image. Thus test-time adaptation has been previously used for only a few applications [11, 16, 22, 30, 45] where a suitable loss function can be designed without ground truth. + +We propose the first approach for test-time Salient Object Detection (SOD). The goal of SOD is to find image regions that attract human attention. Convolutional Neural Networks (CNNs) [14, 15] brought a significant breakthrough for SOD [18, 28, 46, 47]. Traditional training of + +CNNs for SOD requires pixel precise ground truth, but obtaining such annotations is a substantial effort. Therefore, there is an increased interest in semi-supervised [20,23] and weakly supervised SOD [17, 25, 33, 40, 41, 44]. Weak supervision requires less annotation effort, compared to semi-supervision. In this paper, we assume image level supervision, the weakest supervision type, where one provides only images containing salient objects, with no other annotation. + +Most image level weakly supervised SOD approaches [17, 25, 40, 41, 44] are based on noisy pseudo labels, constructed from unsupervised SOD methods. These approaches are hard to modify for test-time adaptation, as at test time, one has only one image to create noisy pseudo-ground truth samples, but combating noise requires diverse samples. + +Our test-time adaptation is based on the approach in [33]. Different from the other image level supervised SOD methods, the approach in [33] does not require pseudo labels. They design a regularized loss1 to use for CNN training without pixel precise annotations. Regularized loss models class-independent properties of object shapes and is likely to have lower values for segmentations that separate the object from the background. This makes regularized loss approach particularly suitable for test time adaptation. + +We design a regularized loss function tailored for test time adaptation. The main problem with regularized loss in [33] is that it may result in a trivial empty solution if hyperparameters of the loss function are not chosen correctly. When training on a large dataset, an occasional empty result is not a big problem, but when training on a single test image, an empty result is disastrous, catastrophically worsening the performance. Thus we must develop a hyperparameter setting method that avoids empty solutions. However, in the absence of ground truth, setting hyperparameter weights is not trivial. We propose a method for setting hyperparameters specific for each test image such that an empty solution is avoided. This method is crucial for obtaining good results + +![](images/bd45db1c1cc850ef9d44998be988f8b357dca97054df8e422f50af75122bac10.jpg) +test image + +![](images/3454a93a7dcf85bd04077bfdab49857fd57f70659ed184de3ca23bbacc3cbb4c.jpg) +ground truth + +![](images/72799618e0e9319c9e71d76c3ae4423885b6dff5557b91b9882a0f35aad62c2a.jpg) +base CNN result + +![](images/f90606f858ae72d52a2879cf303d734598d61d03148baa73ce85d615b81b2a4d.jpg) +base CNN result + dense CRF post-processing + +![](images/b0b88a0b64a087b2a441f5b7717551c4257267d6eec52095a5c3cfa222ab672d.jpg) +test-time dataset + +![](images/98db31ad872e83f89fc2ed2fac9bd070d05caecc7582527ea9c68431b4b60497.jpg) +test-time adaptation with regularized loss + +![](images/eb35cfe430594d7b2868b9cda7110f66fe324b2e1346eb30d4ee3a7eda771049.jpg) + +![](images/e5f32a82d05e8556e4345a6bda6b9b2701ffb6c6a1c261e5a022d26c0b0f7ba7.jpg) + +![](images/b5452838b5756a4c971af85f1688d2236ec00fc2a37dd109b6d87d42764d5753.jpg) +adapted CNN result +Figure 1. Overview of our approach. Top: test image, ground truth, output of base CNN on the test image, and the result of dense CRF [13] post-processing of the base CNN result. Bottom: test-time dataset, CNN output on the test image during several test time epochs, and the final result of the adapted CNN. + +during test-time adaptation. + +Fig. 1 is an overview of our approach. First we train CNN for SOD in weakly supervised setting with image level labels using [33]. The result is called the base CNN. The top row of Fig. 1 shows a test image, ground truth, and the result of the base CNN. Given a test image, we create a small training dataset by augmentation, see Fig. 1, bottom, left. Then we fine tune the base CNN on the small dataset, using our version of regularized loss, Fig. 1, bottom, middle. The resulting CNN is called the adapted CNN. + +The result of base CNN (Fig. 1, top, third image) has many erroneous regions. Sometimes dense CRF [13] is used for post processing to improve performance. We apply dense CRF to the output of base CNN (Fig. 1, top, right). Dense CRF removes small spurious regions but is unable to remove the large erroneous regions as these are salient with high confidence according to the base CNN. + +In contrast, our approach is able to produce much better results, Fig. 1, bottom, right. This is because the base CNN has a high value of regularized loss for this test image. As the fine tuning proceeds, CNN is forced to find alternative segmentations with better loss values, resulting in increasingly better segmentations. Unlike CRF post processing, our adapted CNN is able to remove the large erroneous regions, as their high confidence values according to the base CNN are ignored, and these regions give higher values of regularized loss. Both dense CRF and our approach use CRF models. However, as opposed to post-processing, we use CRF for CNN supervision, enabling CNN to learn a better segmentation. + +Our experiments on the standard benchmarks show that test time adaptation significantly improves performance, achieving the new state-of-art in image level weakly supervised SOD. + +This paper is organized as follows: Sec. 2 is related + +work, Sec. 3 explains the approach in [33], Sec. 4 describes our approach, and Sec. 5 presents experiments. + +# 2. Related Work + +# Weakly Supervised Salient Object Detection + +Most prior work on image level weak supervision consists of two stages. In the first stage, pseudo labels for SOD task are generated, and in the second stage, CNN is trained on the pseudo labels. + +One of the first methods to use image level weak supervision for SOD is in [34]. Their method generates pseudo labels by training CNN for object category prediction. In their setting, image level supervision consists of a large set of 200 object classes categories from Image net. After training for object class prediction, they extract foreground heat maps capturing potential object regions. These generalize to unseen categories, and provide initial pseudo labels. Next, they have a self-training stage which alternates between estimating pseudo labels and training CNN on them. For more accurate pseudo-labels, the predicted pseudo-labels are refined with dense CRF [13]. + +Later image level SOD approaches [17,25,27,40,41,44] use one or more conventional weak (i.e. unsupervised) saliency methods [4,9,10,19] for pseudo labels. In [40], they use multiple weak saliency methods and fuse their results in pseudo-ground truth. During the fusion process, based on the difficulty of training data, they estimate the corresponding confidence measure and utilize it for training with their pseudo ground truth to better handle label noise. + +In [44] they develop a noise modeling module, which enables them to deal with noisy saliency maps obtained from multiple weak saliency methods in a probabilistic way. They later extend their method to rely only on a single weak saliency method [41]. + +In [17], they approach the noise in pseudo labels by exploiting dense CRF [13]. Initial pseudo labels are generated by a weak saliency method. Then they alternate applying dense CRF and CNN training. Dense CRF corrects the noisy labels through spatial consistency and structure preservation, while CNN is trained on the current version of pseudo labels to update and improve the current model. + +In [25], they develop a curriculum framework to incrementally refine the pseudo-labels. Instead of directly using the saliency maps produced by weak saliency methods for training, they first train CNNs to generate pseudo labels for each weak saliency method. This improves the pseudo labels as CNNs are forced to learn representations across a broad set of images. Then pseudo labels are further refined via an iterative self-supervision technique. + +In [27] they utilize multiple pseudo labels to achieve robustness to noise. They design a multi-filter directive CNN which uses multiple directive filters and a multi-guidance loss to integrate multiple pseudo labels. + +In [42] they propose a noise-aware encoder-decoder to deal with noisy pseudo labels. Their method has a saliency predictor that maps input images to clean saliency maps, and a noise generator. The model that represents noisy labels is a sum of these two models. They train their model to simultaneously infer the corresponding latent vector of each noisy label and the saliency predictor. + +All approaches above develop elaborate techniques to deal with noisy pseudo labels. In contrast to the above approaches, in [33] they develop a method for image level supervised SOD which does not rely on pseudo labels from conventional weak saliency methods. The main tool in [33] is regularized loss, based on sparse CRF [2]. The approach in [33] uses off-the-shelf Unet [29] architecture, consists of a single stage, has an intuitive and easy to interpret loss function, and outperforms other image level supervised SOD methods. We review [33] in detail in Sec 3. + +In addition to image level weakly supervised SOD methods, there are approaches that use additional sources of supervision. They do not necessarily rely on conventional weak saliency methods. In [39], they use additional weak sources, such as captions, etc. In [38,43] they use scribbles, and in [21] they use boxes, both much stronger forms of weak supervision. We achieve comparable or better results without additional weak sources. + +# Regularized Loss + +Regularized loss for weak supervision was first used in [31], for semantic segmentation supervised with scribbles. It was subsequently generalized in [24, 32]. + +# Test Time Adaptation + +The idea of test time adaptation has been used before. For example, in the context of traditional computer vision, it was used in [30] for online updates in visual tracking and in [11] for adapting detectors from images to videos. In the + +context of deep learning, it was used in [45] for person re-identification, in [16] for image inpainting, and in [22] for enforcing consistency constraints in depth estimation. In all these methods, the loss to perform test time adaptation is application specific and is unrelated to our approach. + +# 3. Regularized Loss for SOD + +In this section, we review the approach in [33], which is designed for image level weakly supervised segmentation for datasets with a single object class. It naturally applies to SOD, since there is only one class, the salient object. The main idea is to design a regularized loss which incorporates the likely properties of generic objects and to use this loss to train CNN instead of cross entropy on ground truth. + +Regularized loss is applied to the output of CNN, which is the same size as the input image. Let $x$ denote CNN output, and $x_{p}$ denote the output for pixel $p$ . The last layer of CNN is sigmoid, so that $x_{p} \in [0,1]$ . Background corresponds to 0 and salient object to 1. + +Regularized loss in [33] consists of a weighted combination of several components. The most important component is sparse CRF loss [2]. + +$$ +L _ {\operatorname {c r f}} (\boldsymbol {x}) = \frac {1}{| \mathcal {P} |} \sum_ {(p, q) \in \mathcal {N}} e ^ {- \frac {| | C _ {p} - C _ {q} | | ^ {2}}{2 \sigma^ {2}}} \cdot | x _ {p} - x _ {q} |, \tag {1} +$$ + +where $\mathcal{P}$ is the set of all pixels in the image, $\mathcal{N}$ is the set of neighboring pixel pairs on a 4-connected grid, and $C_p \in \mathbb{R}^3$ is the color of $p$ . Parameter $\sigma$ controls the edge strength and is set to $\sigma = 0.15$ in all experiments. + +Usually CNN produces a sharp distribution, i.e. most $x_{p}$ are close either to 0 or 1. Thus if two neighboring pixels are not assigned to the same class, there is a penalty which depends on the strength of the edge between them. Sparse CRF loss is low when segment boundary aligns with image intensity edges. Salient objects are likely to cause image intensity edges, and thus sparse CRF is the main driving tool for salient object discovery. The lowest value of sparse-CRF is zero, achieved at trivial solutions: empty, everything classified as object, or full, everything classified as background. Thus one cannot train with sparse CRF alone. + +Another component of regularized loss is batch volumetric loss. It encourages the average object size in a batch of images to be half of the image size. This loss is useful to prevent both empty and full trivial solutions. Averaging over a batch makes the loss less strict: some batch objects can be significantly smaller or larger than half of the image. + +The next component of regularized loss is minimum volume loss. It encourages the object to be of at least a certain size. It is more likely to be valid in practice, compared to volumetric batch loss, as objects can be expected to be at least a certain minimum size. The last two components of regularized loss are border and center losses. The border + +loss encourage the image border to be the background, and the center loss encourages image center be the object. This is a realistic assumption for many, but not all, images containing a salient object. + +Training CNN with regularized loss is difficult and tends to get stuck in a bad local minimum. In [33] they devise a two-stage strategy for training. For our approach, we start with base CNN trained with the method developed in [33]. + +In our test-time adaptation, we only use sparse CRF loss in Eq. (1), minimum volume loss, and border loss. We now describe minimum volume and border loss in detail. Let $\bar{\pmb{x}} = \frac{1}{|\mathcal{P}|}\sum_{p\in \mathcal{P}}x_p$ , i.e. the normalized object size. Minimum volume loss $L_{\mathrm{m}}$ penalizes segmentations if the normalized object size is less than $\mathrm{obj}_{min}$ + +$$ +L _ {\mathrm {m}} (\boldsymbol {x}) = R e L U ^ {2} \left(\operatorname {o b j} _ {\min } - \bar {\boldsymbol {x}}\right). \tag {2} +$$ + +Following [33], we set $\mathrm{obj}_{\min} = 0.15$ + +Let $\mathcal{B}$ be the set of pixels on the image border of width $w = 3$ . The border loss $L_{\mathrm{b}}(\pmb{x})$ is + +$$ +L _ {\mathrm {b}} (\boldsymbol {x}) = \left(\frac {1}{| \mathcal {B} |} \sum_ {p \in \mathcal {B}} x _ {p}\right) ^ {2}. \tag {3} +$$ + +# 4. Test Time Adaptation + +We now describe our approach. We start with an overview in Sec. 4.1. We describe our regularized loss in Sec. 4.2, and our hyperparameter setting method in Sec. 4.3. + +# 4.1. Overview + +The overview of our approach is in Fig. 1. Given a test image $I$ , we first construct a small training dataset $D(I)$ from $I$ by data augmentation. There are various forms of augmentation. However, since we are training without pixel precise ground truth, we need to avoid augmentations which make the salient object more difficult to detect in an image. For example, large crops can remove the salient object. Similarly, adding random image noise changes the distribution of intensity edges, and sparse CRF in Eq. (1) may become less useful for object discovery. We use small random vertical and horizontal image shifts, and random additive shift to color channels (while keeping the shift the same for all image pixels). We found random vertical and horizontal shifts to be the most effective. + +Next, we take the base CNN trained for SOD task in image level weakly supervised setting, according to the method proposed in [33], and train it on $D(I)$ for a small number of epochs using the regularized loss approach described in Sec. 4.3. The resulting CNN is called the adapted CNN. Training for a small number of iterations ensures that the adapted CNN retains most of the information useful for + +SOD learned from the large training dataset, as well as ensuring efficiency. In practice, we train for 10 epochs. + +# 4.2. Regularized Loss + +We now describe our loss function. For test-time adaptation, we only use some of the regularized loss components from [33]. We do not use batch volumetric loss as it encourages the salient segment to be half of the image size and is far from realistic for most test images. We also do not use center loss. It assumes that the image center belongs to the salient object and is likely not true for many specific test images we wish to adapt CNN for. + +In our regularized loss, we use sparse CRF, minimum volume, and border losses, Eq. (1, 2, 3). Sparse CRF is the most important loss component for discovering salient objects, as it favours segments that align with image intensity edges. Minimum volume loss is necessary to prevent collapse to an empty solution favoured by sparse CRF. Border loss encourages image border to be assigned to the background, and is a realistic assumption for most test images. The complete regularized loss for test time adaptation is + +$$ +L _ {\operatorname {r e g}} (\boldsymbol {x}) = \lambda_ {\operatorname {c r f}} L _ {\operatorname {c r f}} (\boldsymbol {x}) + \lambda_ {m} L _ {\mathrm {m}} (\boldsymbol {x}) + \lambda_ {\mathrm {b}} L _ {\mathrm {b}} (\boldsymbol {x}). \tag {4} +$$ + +# 4.3. Hyperparameter Setting Method + +Consider the loss components in Eq. (4). To get the best performance during test-time adaptation, it is important to choose a good setting of their relative weights. Each test image has its own best setting, however, we cannot find this setting as the ground truth is not available. Thus we must develop a hyperparameter setting method that does not rely on ground truth. + +Out of three components in Eq. (4), the border loss is the least important as it contributes the least to the salient object discovery. Indeed, if our only knowledge about a segmentation is that the image border is assigned to the background, then this gives us almost no information to judge whether this segmentation corresponds to the salient object. Therefore, we set $\lambda_{b}$ to a low value of 1. Out of the two remaining components, sparse CRF is more important for object discovery. Since objects tend to cause intensity edges in an image, a segment that aligns to image edges is more likely to correspond to a salient object, compared to a segment that just obeys the minimum size constraints. Thus sparse CRF weight $\lambda_{\mathrm{crf}}$ should be much higher compared to the minimum volume weight $\lambda_{m}$ . At the same time, the minimum volume weight should be sufficiently large to prevent the collapse to a trivial empty solution. We can compute this sufficiently large weight from the solution obtained from our base CNN. + +We now explain how to set $\lambda_{\mathrm{m}}$ relative to $\lambda_{\mathrm{crf}}$ so that a trivial solution is avoided. Let $x^{i}$ be the output of the base + +CNN on input image $I$ . Assuming that $\pmb{x}^i$ is not empty, we set the weight of minimum volume loss large enough to make the loss on $\pmb{x}^i$ smaller than the loss of an empty solution. Then switching to an empty solution from the initial $\pmb{x}^i$ is too costly. + +The results produced by the base CNN tend to be sharp, that is most pixels have values close to either 0 or 1. Therefore we will assume that the trivial empty solution corresponds to all pixels having value 0 in the output. Let $\pmb{x}^0$ denote such a trivial empty solution, i.e. $x_p^0 = 0$ for all pixels $p$ . Suppose we wish to set sparse-CRF weight to $\lambda_{\mathrm{crf}}$ . We need to set $\lambda_m$ to a value that makes collapse to an empty solution too costly. For this, we need $\lambda_m$ that satisfies + +$$ +L _ {\operatorname {r e g}} \left(\boldsymbol {x} ^ {i}\right) = L _ {\operatorname {r e g}} \left(\boldsymbol {x} ^ {0}\right) + \epsilon , \tag {5} +$$ + +where $\epsilon > 0$ is a small constant, set in practice to 0.1. + +Plugging $\pmb{x}^0$ in Eqs.(1, 2, 3) we obtain + +$$ +L _ {\mathrm {b}} \left(\boldsymbol {x} ^ {0}\right) = 0, \quad L _ {\mathrm {c r f}} \left(\boldsymbol {x} ^ {0}\right) = 0, \quad L _ {\mathrm {m}} \left(\boldsymbol {x} ^ {0}\right) = \operatorname {o b j} _ {\min } ^ {2}. \tag {6} +$$ + +Plugging Eq. (6) into Eq. (5) and solving for $\lambda_{m}$ , we get our formula for setting $\lambda_{m}$ + +$$ +\lambda_ {m} = \frac {\lambda_ {\mathrm {c r f}} L _ {\mathrm {c r f}} \left(\boldsymbol {x} ^ {i}\right) + L _ {\mathrm {b}} \left(\boldsymbol {x} ^ {i}\right) - \epsilon}{\operatorname {o b j} _ {\min } ^ {2} - L _ {\mathrm {m}} \left(\boldsymbol {x} ^ {i}\right)}. \tag {7} +$$ + +Eq. (7) applies when the output $\boldsymbol{x}^i$ of the base CNN is not a trivial empty solution. If $\boldsymbol{x}^i$ is an empty solution, it means that the base CNN fails to extract a salient object. In this case, we set both CRF and minimum volume loss to be of equal weight, $\lambda_{\mathrm{crf}} = \lambda_m$ . This reflects the fact that extracting a salient object from the current image is difficult and the penalty for not satisfying the minim object size is increased. We do not set $\lambda_{\mathrm{m}}$ higher than $\lambda_{\mathrm{crf}}$ as some images may not contain a salient object, in which case an empty solution is appropriate. In addition, while setting $\lambda_{\mathrm{m}}$ high enough would result in some non-empty solution, most likely it would be an erroneous solution, as minimum volume loss is less useful for extracting object segments and, therefore, should not have a weight higher than that of sparse CRF. In practice, we set $\lambda_{\mathrm{crf}} = 10^3$ and compute $\lambda_{\mathrm{m}}$ for each test image using Eq. (7). + +In Table 1 we experimentally evaluate the effectiveness of our method for setting $\lambda_{\mathrm{m}}$ . We randomly select 500 images from DUTO [37] dataset. We test our adaptive $\lambda_{\mathrm{m}}$ for each image, and the same fixed setting of $\lambda_{\mathrm{m}}$ for all test images, in a range from 1 to $10^{3}$ , see Table 1. For this experiment, we use a fixed $\lambda_{\mathrm{crf}} = 10^{3}$ . The performance metric is $F_{\beta}$ . The adaptive $\lambda_{\mathrm{m}}$ performs better than any fixed choice of $\lambda_{\mathrm{m}}$ . As expected, the performance degrades significantly for the larger values of $\lambda_{\mathrm{m}}$ as minimum volume loss starts to play a bigger role in the regularized loss function, however, it is less discriminating of salient objects. For smaller $\lambda_{\mathrm{m}}$ , the performance suffers as the results on some images collapse to an empty solution. + +# 5. Experimental Results + +Our implementation is in Pytorch using RTX 2080 GPU. For our base CNN, we use image level weakly supervised SOD approach [33]. We use Unet [29] architecture with ResNeXt [35] fixed features pretrained onImagenet [5] and train on $256 \times 256$ images. Unlike [33], we train on larger images and on a larger training dataset, thus getting better results for the base CNN than those reported in [33]. + +We train the base CNN on the training set of DUTS [34], which contains 10,553 images. We then evaluate our test-time CNN adaptation approach on the test set of DUTS (5019 images), DUT-OMRON [37] (5168 images), ECSSD [36] (1000 images), THUR [3] (6233 images), PascalS [19] (850 images), and HKU-IS [18] (4447 images). + +For test time adaptation, given a test image $I$ , we construct a test time adaptation dataset $D(I)$ that has 16 images. These images are constructed from $I$ by shifting to the left or right with probability 0.5, and, in addition, shifting up or down with probability 0.5. Each shift is chosen uniformly from the range [0, 15]. This creates a dataset of images similar to $I$ , and it is unlikely that the salient object is removed out of the sample by these small shifts. We use Adam optimizer [12], a fixed learning rate 0.001, and batch size equal to the size of the test time adaptation dataset, set to 16 images. We train for 10 epochs. The images are scaled to resolution $256 \times 256$ . Test time adaptation takes 2.8 seconds per image. + +We use the standard metrics widely adopted for SOD. Our first metric is $F_{\beta}$ score [1], with $\beta^2 = 0.3$ . We also use mae [26], defined as the average absolute per pixel difference between the predicted saliency and ground truth. In addition, we use two newer metrics, namely $S_{\alpha}$ [6], and $E_{s}$ [7]. $S_{\alpha}$ evaluates the structural information for saliency map and region-aware and object-aware structural similarity between saliency maps and ground truth. $E_{s}$ is based on unification of global and local information. + +# 5.1. SOD results + +In Table 2, we report the results of our approach and compare to prior image level weakly supervised SOD methods. Our results with fixed hyperparameter settings in Eq. (4) are in column "ours" (with $\lambda_{\mathrm{m}} = 20$ , which is the best fixed setting according to Table 1). The results with our method for hyperparameter setting for each image individually, as described in Sec. 4.3, are in column 'ours + h'. Except one case, all metrics for our approach with hyperparameter setting method are improved, often significantly. This shows that our hyperparameter setting method is crucial for obtaining a significantly improved performance. + +Dense CRF [13] is often used to post-process CNN results. Therefore, we also post-process base CNN results with dense CRF for comparison. The performance of the base CNN is in column [33], and its performance after post- + +
λm151020501002005001000adaptive
Fβ.741.738.752.755.742.732.728.718.703.765
+ +Table 1. Performance in terms of $F_{\beta}$ score of fine tuning with a fixed value of $\lambda_{m}$ vs. our adaptive $\lambda_{\mathrm{m}}$ computation according to Eq. (7). For all experiments, $\lambda_{\mathrm{crf}} = 10^{3}$ . + +
Metric[34][17][39][44][42][27][33][33]+dCRFoursours+h
DUTS\( F_{\beta} \)↑.654.614.684.725.747.710.753.763.771.795
mae↓.100.116.091.075.060.076.056.055.054.052
\( E_s \)↑.795.772.814.853.859.839.833.842.855.863
\( S_{\alpha} \)↑.748.697.759.813.827.775.791.798.800.817
ECSSD\( F_{\beta} \)↑.823.797.609.810.852.854.868.873.882.911
mae↓.104.110.109.092.071.084.063.060.057.046
\( E_s \)↑.869.853.673.836.883.885.884.886.885.909
\( S_{\alpha} \)↑.811.802.756.846.860.834.845.855.859.883
DUTO\( F_{\beta} \)↑.603.622.609.597.701.646.697.710.721.752
mae↓.109.101.109.103.070.087.074.069.066.057
\( E_s \)↑.768.776.673.712.816.803.821.827.833.842
\( S_{\alpha} \)↑.725.752.756.733.791.742.763.771.782.798
PascalS\( F_{\beta} \)↑.715.693.713--.751.796.801.801.807
mae↓.139.149.135--.115.083.082.082.084
\( E_s \)↑.791.772.790--.817.827.827.826.828
\( S_{\alpha} \)↑.744.717.768--.770.791.793.792.796
HKUIS\( F_{\beta} \)↑--.814.820.878.851.866.871.883.899
mae↓--.084.065.043.059.044.040.039.037
\( E_s \)↑--.895.858.919.921.912.925.926.931
\( S_{\alpha} \)↑--.818.860.890.846.857.866.871.879
THUR\( F_{\beta} \)↑---.694.719-.731.749.753.760
mae↓---.086.070-.073.069.067.064
\( E_s \)↑---.807.838-.840.842.843.844
\( S_{\alpha} \)↑---.804.810-.794.798.806.819
+ +Table 2. Comparison to SOD methods that also use image level weak supervision using metrics $F_{\beta}$ , $mae$ , $E_s$ , and $S_{\alpha}$ . We use $\uparrow$ to indicate that the higher score is better, and $\downarrow$ to indicate the lower score is better. The best result is in bold, and the second best is underlined. + +processing with dense CRF is in column [33] + dCRF. Our approach is significantly better than [33] with dense CRF post-processing. This illustrates that it is important to include CRF during training in the loss function, rather than use CRF for just post-processing. + +Our test time adaptation improves the performance of [33] across all metrics, except one case. For easier datasets (ECSSD, HKUIS), the performance gain is not as large as the base CNN already performs well, but for harder to fit datasets, the performance gain is significant. + +Comparing to other SOD methods, our method is better in almost all cases except $S_{\alpha}$ metric for HKUIS and DUTS. For most datasets, our approach significantly improves $F_{\beta}$ metric over the second best method, notably by roughly $7\%$ for DUTO and DUTS datasets. + +In Table 3, we compare our method to weakly supervised SOD methods that use stronger form of weak supervision. In [21], they use bounding box supervision, and in [38, 43], + +they use scribbles. Our method is better than the bounding box method [43] in all cases. Our method is almost always better than [43], and better in about half of the cases compared to [38]. If Fig. 2, we show qualitative comparisons of our method to the scribble supervised methods of [43] and [38]. For this illustration, we chose examples where all methods work relatively well, in terms of the standard metrics. However, our results have better detail preservation qualities, especially when the salient object has sharp edges on the boundary. Observe the detail preservation in the chair, in the last row. This is because CRF based loss function is adapted to the particular test image, driving the segmentation boundary to align with image edges. + +# 5.2. Ablation Experiments + +# Hyperparameters in Regularized Loss + +We first discuss and perform ablation for hyperparameters of our regularized loss function, Eq. (4). The most im + +
DUTSECSSDDUTOPascalSHKUISTHUR
boxes [21]Fβ↑.736.860.686-.853-
mae ↓.079.072.081-.058-
Es ↑.831.889.810-.897-
Sα ↑.789.858.776-.852-
scribbles [43]Fβ↑.747.865.715.788.858.718
mae ↓.062.061.068.140.047.077
Es ↑.865.908.835.798.923.873
Sα ↑------
scribbles [38]Fβ↑.823.900.758.823.896.755
mae ↓.049.049.060.078.038.069
Es ↑.890.908.862.847.938.843
Sα ↑------
oursFβ↑.795.911.752.807.899.760
mae ↓.052.046.057.082.037.064
Es ↑.863.909.842.828.931.844
Sα ↑.817.883.798.796.879.819
+ +![](images/ef7b8bb40b2ddef51973af2c6b910fd95b82b4db13b11e69bc9db2725b21f246.jpg) +Figure 2. Qualitative comparison. From left to right: input image, ground truth, results of [43], [38], and our results, respectively. Note that [43], [38] both use scribble supervision, a much stronger supervision form compared to what we use. + +important parameter to set is $\lambda_{m}$ in relation to $\lambda_{\mathrm{crf}}$ , as the correct relative settings of these two parameters ensures there is no collapse to a trivial solution. For this, we have de + +Table 3. Comparison to prior SOD methods that use stronger forms of weak supervision using metrics $F_{\beta}$ , $mae$ , $E_s$ , and $S_{\alpha}$ . In [21], they use bounding box supervision, and in [38,43], they use scribbles. We use $\uparrow$ to indicate that the higher score is better, and $\downarrow$ to indicate the lower score is better. The best result is in bold, and second best is underlined. + +
with border losswithout border loss
DUTS.795.789
ECSSD.911.910
DUTO.752.748
PascalS.807.800
HKUIS.899.892
THUR.760.755
+ +Table 4. Performance of our approach with border loss and without border loss. Performance metric is $F_{\beta}$ score. + +veloped an effective automatic approach, Sec. 4.3. Since $\lambda_{m}$ is computed in relation to $\lambda_{\mathrm{crf}}$ , the setting of $\lambda_{\mathrm{crf}}$ can be arbitrary (up to the precision that can be handled by the software), $\lambda_{\mathrm{m}}$ is adjusted accordingly: for larger $\lambda_{\mathrm{crf}}$ , $\lambda_{\mathrm{m}}$ will be automatically set larger according to Eq. (7). + +The setting of $\lambda_{b}$ , as long as it is much smaller than $\lambda_{\mathrm{crf}}$ , is not important. Parameter $\lambda_{\mathrm{b}}$ controls the border loss that encourages border pixels to be assigned to the background. While this loss is helpful, it is the least important regularized loss component. To show that this loss is the least important, we perform an experiment where we remove border loss from regularized loss ( $\lambda_{b} = 0$ ). The results are summarized in Table 4. Without border loss, the results worsen only a little. On the other hand, omitting either CRF or minimum volume losses from our loss function in Eq. (4) results in a catastrophic performance collapse. + +In Eq. (1), we set $\sigma = 0.15$ as in [33]. Since the training images are normalized, this setting works well across all datasets. We also experimented with other values of $\sigma$ , see + +
σ = 0.1σ = 0.15σ = 0.20
DUTS.789.795.797
ECSSD.894.911.915
DUTO.751.752.743
PascalS.803.807.803
HKUIS.902.899.895
THUR.763.760.757
+ +Table 5. Performance of our approach for different $\sigma$ values. Performance metric is ${F}_{\beta }$ score. + +
|D(I)|248163264
Fβ↑.879.887.890.910.915.913
mae ↓.058.055.049.046.045.046
+ +Table 6. Different choice of test time dataset size for ECSSD dataset. + +
# epoch2510203050
Fβ↑.873.883.910.915.915.912
mae ↓.062.053.047.045.046.047
+ +Table 5. The performance is stable for a range of $\sigma$ values. + +# 5.3. Other Hyperparameters + +We now perform ablation experiments for various hyperparameter settings (other than those in the regularized loss function) of our approach. We perform experiments on ECSSD dataset, and evaluate the sensitivity of our approach to the size of the test time adaptation dataset $D(I)$ , the number of epoch performed for test time adaptation, and different data augmentation strategies. For all the experiments, the performance metric is $F_{\beta}$ and $mae$ . + +In Table 6, we show how performance depends on varying the size of the test time dataset $D(I)$ . Our standard setting for the size of $D(I)$ is 16. Decreasing it to 8 slightly degrades the performance, increasing it to 64 results in almost no change, but the running time is twice longer. + +In Table 7, we show how performance depends on increasing the number of epoch for test time adaptation. Our standard setting for the number of epoch is 10. Decreasing it to 2 noticeably degrades the performance, increasing it to 50 results in only a slight improvement but significantly increases the running time. + +In Table 8, we show the effect of different augmentation strategies. Our main experiments use small random shifts in the vertical and horizontal directions. We also tested adding random Gaussian noise to images, a standard data augmentation technique, but it significantly degrades the performance. Adding Gaussian noise decreases the effectiveness of our sparse CRF loss as neighboring pixels become less + +Table 7. Different number of test time adaptation iterations for ECSSD dataset. + +
typeshiftsGauss noisecolor perturbationshifts+color
Fβ↑.911.854.897.909
mae↓.046.060.050.049
+ +Table 8. Different augmentation strategies when fine tuning on ECSSD dataset. + +
DUTSECSSDDUTOPascalSHKUISTHUR
% improved79.383.785.381.276.775.0
mean Fβ increase.048.053.064.17.037.036
mean Fβ decrease.021.009.011.08.019.015
+ +Table 9. Percentage of images for which performance improves as the result of test-time adaptation. For the improved (and degraded) images, we also show the mean increase (and the decrease) in $F_{\beta}$ score. + +strongly connected. We also tested random color perturbation where all pixels change the value of a color channel by the same random value. This works much better than Gaussian noise, but not as well as random shifts. Shifts and color perturbation work as well as shifts alone. + +# 5.4. Additional Experiments + +During test-time adaptation, the majority, but not all images, improve. In Table 9, we show the percentage of images that improve, in terms of their $F_{\beta}$ score, for each of the test datasets. In addition, we show the mean improvement of $F_{\beta}$ score, averaged over the images that improve, and the mean decline of $F_{\beta}$ score, averaged over the images that decline in performance. As can be seen from the table, for all dataset, over $75\%$ of images improve during test-time adaptation. Also, the average improvement is significantly larger than the average decline of the results. Thus overall, test-time adaptation significantly improves the performance. In the supplementary materials, we show the most improved and the most degraded images during test time adaptation. + +# 6. Conclusions + +We propose the first approach to test time adaptation for SOD with image level weak supervision. We develop a regularized loss function particularly effective for test time adaptation, and a method to find an appropriate setting of the hyperparameters in our loss function. Our approach uses a standard CNN architecture and a loss function that is intuitive and simple to interpret. We achieve a new state of the art in image level supervised SOD, and a better or competitive result with the prior work that uses stronger forms of weak supervision. The drawback of our approach is that the running time for testing is more expensive, since the base CNN needs to be fine-tuned for each test image. + +# References + +[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5 +[2] Yuri Boykov, Olga Veksler, and Ramin Zabih. A variable window approach to early vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1283-1294, December 1998. An earlier version of this work appeared in CVPR '97. 3 +[3] Ming-Ming Cheng, NiloyJ. Mitra, Xiaolei Huang, and ShiMin Hu. Salientshape: group saliency in image collections. The Visual Computer, 30(4):443-453, 2014. 5 +[4] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2 +[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5 +[6] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 5 +[7] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 698-704, 2018. 5 +[8] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT Press, 2016. 1 +[9] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2 +[10] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2 +[11] Vidit Jain and Erik Learned-Miller. Online domain adaptation of a pre-trained cascade of classifiers. In CVPR 2011, pages 577-584. IEEE, 2011. 1, 3 +[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[13] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In Neural Information Processing Systems, pages 109-117, 2011. 2, 3, 5 +[14] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Neural Information Processing Systems, 2012. 1 +[15] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Comput., 1(4):541-551, December 1989. 1 + +[16] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021. 1, 3 +[17] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 3, 6 +[18] G. Li and Y. Yu. Deep contrast learning for salient object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 478-487, June 2016. 1, 5 +[19] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 280-287, 2014. 2, 5 +[20] Jiawei Liu, Jing Zhang, and Nick Barnes. Semi-supervised salient object detection with effective confidence estimation. arXiv preprint arXiv:2112.14019, 2021. 1 +[21] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE Transactions on Image Processing, 30:4423-4435, 2021. 3, 6, 7 +[22] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 3 +[23] Yunqiu Lv, Bowen Liu, Jing Zhang, Yuchao Dai, Aixuan Li, and Tong Zhang. Semi-supervised active salient object detection. Pattern Recognition, 123:108364, 2022. 1 +[24] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6608-6618, October 2021. 3 +[25] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. In Advances in Neural Information Processing Systems, pages 204-214, 2019. 1, 2, 3 +[26] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In Conference on Computer Vision and Pattern Recognition, pages 733-740. IEEE, 2012. 5 +[27] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 2, 3, 6 +[28] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 1 +[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 3, 5 +[30] David A Ross, Jongwoo Lim, Ruei-Sung Lin, and Ming-Hsuan Yang. Incremental learning for robust visual track + +ing. International journal of computer vision, 77(1):125-141, 2008. 1, 3 +[31] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised CNN segmentation. In Conference on Computer Vision and Pattern Recognition, pages 1818-1827, 2018. 3 +[32] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised CNN segmentation. In European Conference on Computer Vision, pages 524-540, 2018. 3 +[33] Olga Veksler. Regularized loss for weakly supervised single class semantic segmentation. In European Conference on Computer Vision, pages 348-365. Springer, 2020. 1, 2, 3, 4, 5, 6, 7 +[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5, 6 +[35] Saining Xie, Ross B. Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5 +[36] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 1155-1162. IEEE, 2013. 5 +[37] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Conference on Computer Vision and Pattern Recognition, pages 3166-3173. IEEE, 2013. 5 +[38] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3234-3242. AAAI Press, 2021. 3, 6, 7 +[39] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 6074-6083, 2019. 3, 6 +[40] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2 +[41] Jing Zhang, Yuchao Dai, Tong Zhang, Mehrtash Harandi, Nick Barnes, and Richard Hartley. Learning saliency from single noisy labelling: A robust model fitting perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2866-2873, 2021. 1, 2 +[42] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European + +conference on computer vision, pages 349-366. Springer, 2020. 3, 6 +[43] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In Conference on Computer Vision and Pattern Recognition, pages 12546-12555, 2020. 3, 6, 7 +[44] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9029-9038, 2018. 1, 2, 6 +[45] Shun Zhang, Jia-Bin Huang, Jongwoo Lim, Yihong Gong, Jinjun Wang, Narendra Ahuja, and Ming-Hsuan Yang. Tracking persons-of-interest via unsupervised representation adaptation. International Journal of Computer Vision, 128(1):96-120, 2020. 1, 3 +[46] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3085-3094, 2019. 1 +[47] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. 2020. 1 \ No newline at end of file diff --git a/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/images.zip b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da81d3117bbfb7380013a258851f76fdd1683830 --- /dev/null +++ b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69e6faa7052162f4285fc5642bc12b9f9a2353c354e4ea9d00354b901720b20e +size 461505 diff --git a/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/layout.json b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..883ee6e7dca402d64bef7db0d2d187bd4d90f3d4 --- /dev/null +++ b/2023/Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection/layout.json @@ -0,0 +1,9627 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 140 + ], + "type": "text", + "content": "Test Time Adaptation with Regularized Loss for Weakly Supervised Salient Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 161, + 375, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 161, + 375, + 203 + ], + "spans": [ + { + "bbox": [ + 220, + 161, + 375, + 203 + ], + "type": "text", + "content": "Olga Veksler \nUniversity of Waterloo, Canada \noveksler@uwaterloo.ca" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 257, + 289, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 289, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 289, + 497 + ], + "type": "text", + "content": "It is well known that CNNs tend to overfit to the training data. Test-time adaptation is an extreme approach to deal with overfitting: given a test image, the aim is to adapt the trained model to that image. Indeed nothing can be closer to the test data than the test image itself. The main difficulty of test-time adaptation is that the ground truth is not available. Thus test-time adaptation, while intriguing, applies to only a few scenarios where one can design an effective loss function that does not require ground truth. We propose the first approach for test-time Salient Object Detection (SOD) in the context of weak supervision. Our approach is based on a so-called regularized loss function, which can be used for training CNN when pixel precise ground truth is unavailable. Regularized loss tends to have lower values for the more likely object segments, and thus it can be used to fine-tune an already trained CNN to a given test image, adapting to images unseen during training. We develop a regularized loss function particularly suitable for test-time adaptation and show that our approach significantly outperforms prior work for weakly supervised SOD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "A well known problem with CNNs is that they tend to overfit to the training data. One approach to deal with overfitting is test-time adaptation [30]. A model is trained on the training data, and then fine-tuned for a few epochs during test time on a given test image. The main difficulty of test-time adaptation is that there is no ground truth for the test image. Thus test-time adaptation has been previously used for only a few applications [11, 16, 22, 30, 45] where a suitable loss function can be designed without ground truth." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 45, + 653, + 287, + 714 + ], + "type": "text", + "content": "We propose the first approach for test-time Salient Object Detection (SOD). The goal of SOD is to find image regions that attract human attention. Convolutional Neural Networks (CNNs) [14, 15] brought a significant breakthrough for SOD [18, 28, 46, 47]. Traditional training of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 327 + ], + "type": "text", + "content": "CNNs for SOD requires pixel precise ground truth, but obtaining such annotations is a substantial effort. Therefore, there is an increased interest in semi-supervised [20,23] and weakly supervised SOD [17, 25, 33, 40, 41, 44]. Weak supervision requires less annotation effort, compared to semi-supervision. In this paper, we assume image level supervision, the weakest supervision type, where one provides only images containing salient objects, with no other annotation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 328, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 411 + ], + "type": "text", + "content": "Most image level weakly supervised SOD approaches [17, 25, 40, 41, 44] are based on noisy pseudo labels, constructed from unsupervised SOD methods. These approaches are hard to modify for test-time adaptation, as at test time, one has only one image to create noisy pseudo-ground truth samples, but combating noise requires diverse samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 412, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 519 + ], + "type": "text", + "content": "Our test-time adaptation is based on the approach in [33]. Different from the other image level supervised SOD methods, the approach in [33] does not require pseudo labels. They design a regularized loss1 to use for CNN training without pixel precise annotations. Regularized loss models class-independent properties of object shapes and is likely to have lower values for segmentations that separate the object from the background. This makes regularized loss approach particularly suitable for test time adaptation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": "We design a regularized loss function tailored for test time adaptation. The main problem with regularized loss in [33] is that it may result in a trivial empty solution if hyperparameters of the loss function are not chosen correctly. When training on a large dataset, an occasional empty result is not a big problem, but when training on a single test image, an empty result is disastrous, catastrophically worsening the performance. Thus we must develop a hyperparameter setting method that avoids empty solutions. However, in the absence of ground truth, setting hyperparameter weights is not trivial. We propose a method for setting hyperparameters specific for each test image such that an empty solution is avoided. This method is crucial for obtaining good results" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": "1In the context of CNNs, regularization is a term often used to refer to the norm regularization on the network weights [8]. Here, regularized loss refers to the loss function on the output of CNN." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7360" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 71, + 197, + 132 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 197, + 132 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 197, + 132 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 197, + 132 + ], + "type": "image", + "image_path": "bd45db1c1cc850ef9d44998be988f8b357dca97054df8e422f50af75122bac10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 133, + 167, + 140 + ], + "lines": [ + { + "bbox": [ + 132, + 133, + 167, + 140 + ], + "spans": [ + { + "bbox": [ + 132, + 133, + 167, + 140 + ], + "type": "text", + "content": "test image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 293, + 132 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 293, + 132 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 293, + 132 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 293, + 132 + ], + "type": "image", + "image_path": "3454a93a7dcf85bd04077bfdab49857fd57f70659ed184de3ca23bbacc3cbb4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 133, + 270, + 140 + ], + "lines": [ + { + "bbox": [ + 227, + 133, + 270, + 140 + ], + "spans": [ + { + "bbox": [ + 227, + 133, + 270, + 140 + ], + "type": "text", + "content": "ground truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 300, + 72, + 390, + 132 + ], + "blocks": [ + { + "bbox": [ + 300, + 72, + 390, + 132 + ], + "lines": [ + { + "bbox": [ + 300, + 72, + 390, + 132 + ], + "spans": [ + { + "bbox": [ + 300, + 72, + 390, + 132 + ], + "type": "image", + "image_path": "72799618e0e9319c9e71d76c3ae4423885b6dff5557b91b9882a0f35aad62c2a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 133, + 372, + 139 + ], + "lines": [ + { + "bbox": [ + 317, + 133, + 372, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 372, + 139 + ], + "type": "text", + "content": "base CNN result" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 486, + 132 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 486, + 132 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 486, + 132 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 486, + 132 + ], + "type": "image", + "image_path": "f90606f858ae72d52a2879cf303d734598d61d03148baa73ce85d615b81b2a4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 133, + 486, + 148 + ], + "lines": [ + { + "bbox": [ + 391, + 133, + 486, + 148 + ], + "spans": [ + { + "bbox": [ + 391, + 133, + 486, + 148 + ], + "type": "text", + "content": "base CNN result + dense CRF post-processing" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 107, + 148, + 196, + 207 + ], + "blocks": [ + { + "bbox": [ + 107, + 148, + 196, + 207 + ], + "lines": [ + { + "bbox": [ + 107, + 148, + 196, + 207 + ], + "spans": [ + { + "bbox": [ + 107, + 148, + 196, + 207 + ], + "type": "image", + "image_path": "b0b88a0b64a087b2a441f5b7717551c4257267d6eec52095a5c3cfa222ab672d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 210, + 190, + 218 + ], + "lines": [ + { + "bbox": [ + 132, + 210, + 190, + 218 + ], + "spans": [ + { + "bbox": [ + 132, + 210, + 190, + 218 + ], + "type": "text", + "content": "test-time dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 197, + 158, + 245, + 191 + ], + "blocks": [ + { + "bbox": [ + 197, + 158, + 245, + 191 + ], + "lines": [ + { + "bbox": [ + 197, + 158, + 245, + 191 + ], + "spans": [ + { + "bbox": [ + 197, + 158, + 245, + 191 + ], + "type": "image", + "image_path": "98db31ad872e83f89fc2ed2fac9bd070d05caecc7582527ea9c68431b4b60497.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 198, + 364, + 205 + ], + "lines": [ + { + "bbox": [ + 225, + 198, + 364, + 205 + ], + "spans": [ + { + "bbox": [ + 225, + 198, + 364, + 205 + ], + "type": "text", + "content": "test-time adaptation with regularized loss" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 246, + 158, + 294, + 191 + ], + "blocks": [ + { + "bbox": [ + 246, + 158, + 294, + 191 + ], + "lines": [ + { + "bbox": [ + 246, + 158, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 246, + 158, + 294, + 191 + ], + "type": "image", + "image_path": "eb35cfe430594d7b2868b9cda7110f66fe324b2e1346eb30d4ee3a7eda771049.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 294, + 158, + 392, + 191 + ], + "blocks": [ + { + "bbox": [ + 294, + 158, + 392, + 191 + ], + "lines": [ + { + "bbox": [ + 294, + 158, + 392, + 191 + ], + "spans": [ + { + "bbox": [ + 294, + 158, + 392, + 191 + ], + "type": "image", + "image_path": "e5f32a82d05e8556e4345a6bda6b9b2701ffb6c6a1c261e5a022d26c0b0f7ba7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 397, + 150, + 486, + 210 + ], + "blocks": [ + { + "bbox": [ + 397, + 150, + 486, + 210 + ], + "lines": [ + { + "bbox": [ + 397, + 150, + 486, + 210 + ], + "spans": [ + { + "bbox": [ + 397, + 150, + 486, + 210 + ], + "type": "image", + "image_path": "b5452838b5756a4c971af85f1688d2236ec00fc2a37dd109b6d87d42764d5753.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 211, + 475, + 219 + ], + "lines": [ + { + "bbox": [ + 410, + 211, + 475, + 219 + ], + "spans": [ + { + "bbox": [ + 410, + 211, + 475, + 219 + ], + "type": "text", + "content": "adapted CNN result" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "type": "text", + "content": "Figure 1. Overview of our approach. Top: test image, ground truth, output of base CNN on the test image, and the result of dense CRF [13] post-processing of the base CNN result. Bottom: test-time dataset, CNN output on the test image during several test time epochs, and the final result of the adapted CNN." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 294, + 160, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 160, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 160, + 305 + ], + "type": "text", + "content": "during test-time adaptation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 413 + ], + "type": "text", + "content": "Fig. 1 is an overview of our approach. First we train CNN for SOD in weakly supervised setting with image level labels using [33]. The result is called the base CNN. The top row of Fig. 1 shows a test image, ground truth, and the result of the base CNN. Given a test image, we create a small training dataset by augmentation, see Fig. 1, bottom, left. Then we fine tune the base CNN on the small dataset, using our version of regularized loss, Fig. 1, bottom, middle. The resulting CNN is called the adapted CNN." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "content": "The result of base CNN (Fig. 1, top, third image) has many erroneous regions. Sometimes dense CRF [13] is used for post processing to improve performance. We apply dense CRF to the output of base CNN (Fig. 1, top, right). Dense CRF removes small spurious regions but is unable to remove the large erroneous regions as these are salient with high confidence according to the base CNN." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 498, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 652 + ], + "type": "text", + "content": "In contrast, our approach is able to produce much better results, Fig. 1, bottom, right. This is because the base CNN has a high value of regularized loss for this test image. As the fine tuning proceeds, CNN is forced to find alternative segmentations with better loss values, resulting in increasingly better segmentations. Unlike CRF post processing, our adapted CNN is able to remove the large erroneous regions, as their high confidence values according to the base CNN are ignored, and these regions give higher values of regularized loss. Both dense CRF and our approach use CRF models. However, as opposed to post-processing, we use CRF for CNN supervision, enabling CNN to learn a better segmentation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "content": "Our experiments on the standard benchmarks show that test time adaptation significantly improves performance, achieving the new state-of-art in image level weakly supervised SOD." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "This paper is organized as follows: Sec. 2 is related" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 293, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 318 + ], + "type": "text", + "content": "work, Sec. 3 explains the approach in [33], Sec. 4 describes our approach, and Sec. 5 presents experiments." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 331, + 392, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 392, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 392, + 342 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 351, + 497, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 497, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 497, + 363 + ], + "type": "text", + "content": "Weakly Supervised Salient Object Detection" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 365, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 412 + ], + "type": "text", + "content": "Most prior work on image level weak supervision consists of two stages. In the first stage, pseudo labels for SOD task are generated, and in the second stage, CNN is trained on the pseudo labels." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 413, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 546, + 555 + ], + "type": "text", + "content": "One of the first methods to use image level weak supervision for SOD is in [34]. Their method generates pseudo labels by training CNN for object category prediction. In their setting, image level supervision consists of a large set of 200 object classes categories from Image net. After training for object class prediction, they extract foreground heat maps capturing potential object regions. These generalize to unseen categories, and provide initial pseudo labels. Next, they have a self-training stage which alternates between estimating pseudo labels and training CNN on them. For more accurate pseudo-labels, the predicted pseudo-labels are refined with dense CRF [13]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "type": "text", + "content": "Later image level SOD approaches [17,25,27,40,41,44] use one or more conventional weak (i.e. unsupervised) saliency methods [4,9,10,19] for pseudo labels. In [40], they use multiple weak saliency methods and fuse their results in pseudo-ground truth. During the fusion process, based on the difficulty of training data, they estimate the corresponding confidence measure and utilize it for training with their pseudo ground truth to better handle label noise." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "In [44] they develop a noise modeling module, which enables them to deal with noisy saliency maps obtained from multiple weak saliency methods in a probabilistic way. They later extend their method to rely only on a single weak saliency method [41]." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "type": "text", + "content": "7361" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "In [17], they approach the noise in pseudo labels by exploiting dense CRF [13]. Initial pseudo labels are generated by a weak saliency method. Then they alternate applying dense CRF and CNN training. Dense CRF corrects the noisy labels through spatial consistency and structure preservation, while CNN is trained on the current version of pseudo labels to update and improve the current model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 252 + ], + "type": "text", + "content": "In [25], they develop a curriculum framework to incrementally refine the pseudo-labels. Instead of directly using the saliency maps produced by weak saliency methods for training, they first train CNNs to generate pseudo labels for each weak saliency method. This improves the pseudo labels as CNNs are forced to learn representations across a broad set of images. Then pseudo labels are further refined via an iterative self-supervision technique." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 286, + 301 + ], + "type": "text", + "content": "In [27] they utilize multiple pseudo labels to achieve robustness to noise. They design a multi-filter directive CNN which uses multiple directive filters and a multi-guidance loss to integrate multiple pseudo labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 301, + 286, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 286, + 386 + ], + "type": "text", + "content": "In [42] they propose a noise-aware encoder-decoder to deal with noisy pseudo labels. Their method has a saliency predictor that maps input images to clean saliency maps, and a noise generator. The model that represents noisy labels is a sum of these two models. They train their model to simultaneously infer the corresponding latent vector of each noisy label and the saliency predictor." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 387, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 286, + 506 + ], + "type": "text", + "content": "All approaches above develop elaborate techniques to deal with noisy pseudo labels. In contrast to the above approaches, in [33] they develop a method for image level supervised SOD which does not rely on pseudo labels from conventional weak saliency methods. The main tool in [33] is regularized loss, based on sparse CRF [2]. The approach in [33] uses off-the-shelf Unet [29] architecture, consists of a single stage, has an intuitive and easy to interpret loss function, and outperforms other image level supervised SOD methods. We review [33] in detail in Sec 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 286, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 286, + 601 + ], + "type": "text", + "content": "In addition to image level weakly supervised SOD methods, there are approaches that use additional sources of supervision. They do not necessarily rely on conventional weak saliency methods. In [39], they use additional weak sources, such as captions, etc. In [38,43] they use scribbles, and in [21] they use boxes, both much stronger forms of weak supervision. We achieve comparable or better results without additional weak sources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 124, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 124, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 124, + 615 + ], + "type": "text", + "content": "Regularized Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 652 + ], + "type": "text", + "content": "Regularized loss for weak supervision was first used in [31], for semantic segmentation supervised with scribbles. It was subsequently generalized in [24, 32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 143, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 143, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 143, + 664 + ], + "type": "text", + "content": "Test Time Adaptation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "The idea of test time adaptation has been used before. For example, in the context of traditional computer vision, it was used in [30] for online updates in visual tracking and in [11] for adapting detectors from images to videos. In the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "context of deep learning, it was used in [45] for person re-identification, in [16] for image inpainting, and in [22] for enforcing consistency constraints in depth estimation. In all these methods, the loss to perform test time adaptation is application specific and is unrelated to our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 143, + 455, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 143, + 455, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 455, + 156 + ], + "type": "text", + "content": "3. Regularized Loss for SOD" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 163, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 247 + ], + "type": "text", + "content": "In this section, we review the approach in [33], which is designed for image level weakly supervised segmentation for datasets with a single object class. It naturally applies to SOD, since there is only one class, the salient object. The main idea is to design a regularized loss which incorporates the likely properties of generic objects and to use this loss to train CNN instead of cross entropy on ground truth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": "Regularized loss is applied to the output of CNN, which is the same size as the input image. Let " + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": " denote CNN output, and " + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": " denote the output for pixel " + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": ". The last layer of CNN is sigmoid, so that " + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "inline_equation", + "content": "x_{p} \\in [0,1]" + }, + { + "bbox": [ + 304, + 247, + 545, + 306 + ], + "type": "text", + "content": ". Background corresponds to 0 and salient object to 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "content": "Regularized loss in [33] consists of a weighted combination of several components. The most important component is sparse CRF loss [2]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 329, + 350, + 545, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 329, + 350, + 545, + 382 + ], + "type": "interline_equation", + "content": "L _ {\\operatorname {c r f}} (\\boldsymbol {x}) = \\frac {1}{| \\mathcal {P} |} \\sum_ {(p, q) \\in \\mathcal {N}} e ^ {- \\frac {| | C _ {p} - C _ {q} | | ^ {2}}{2 \\sigma^ {2}}} \\cdot | x _ {p} - x _ {q} |, \\tag {1}", + "image_path": "821fb37fce93afa1a1f122ee00452144cc44ec2a6cf3ecd2725bade9358635d9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": " is the set of all pixels in the image, " + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": " is the set of neighboring pixel pairs on a 4-connected grid, and " + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "inline_equation", + "content": "C_p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": " is the color of " + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": ". Parameter " + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": " controls the edge strength and is set to " + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\sigma = 0.15" + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": " in all experiments." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "type": "text", + "content": "Usually CNN produces a sharp distribution, i.e. most " + }, + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 304, + 438, + 545, + 569 + ], + "type": "text", + "content": " are close either to 0 or 1. Thus if two neighboring pixels are not assigned to the same class, there is a penalty which depends on the strength of the edge between them. Sparse CRF loss is low when segment boundary aligns with image intensity edges. Salient objects are likely to cause image intensity edges, and thus sparse CRF is the main driving tool for salient object discovery. The lowest value of sparse-CRF is zero, achieved at trivial solutions: empty, everything classified as object, or full, everything classified as background. Thus one cannot train with sparse CRF alone." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 642 + ], + "type": "text", + "content": "Another component of regularized loss is batch volumetric loss. It encourages the average object size in a batch of images to be half of the image size. This loss is useful to prevent both empty and full trivial solutions. Averaging over a batch makes the loss less strict: some batch objects can be significantly smaller or larger than half of the image." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "The next component of regularized loss is minimum volume loss. It encourages the object to be of at least a certain size. It is more likely to be valid in practice, compared to volumetric batch loss, as objects can be expected to be at least a certain minimum size. The last two components of regularized loss are border and center losses. The border" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "7362" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "loss encourage the image border to be the background, and the center loss encourages image center be the object. This is a realistic assumption for many, but not all, images containing a salient object." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 167 + ], + "type": "text", + "content": "Training CNN with regularized loss is difficult and tends to get stuck in a bad local minimum. In [33] they devise a two-stage strategy for training. For our approach, we start with base CNN trained with the method developed in [33]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "text", + "content": "In our test-time adaptation, we only use sparse CRF loss in Eq. (1), minimum volume loss, and border loss. We now describe minimum volume and border loss in detail. Let " + }, + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}} = \\frac{1}{|\\mathcal{P}|}\\sum_{p\\in \\mathcal{P}}x_p" + }, + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "text", + "content": ", i.e. the normalized object size. Minimum volume loss " + }, + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "text", + "content": " penalizes segmentations if the normalized object size is less than " + }, + { + "bbox": [ + 46, + 168, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\mathrm{obj}_{min}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 249, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 249, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 100, + 249, + 287, + 262 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {m}} (\\boldsymbol {x}) = R e L U ^ {2} \\left(\\operatorname {o b j} _ {\\min } - \\bar {\\boldsymbol {x}}\\right). \\tag {2}", + "image_path": "fffcd3563fccce6c18ce4541e8eb4d717a09a3adefe8177564ae239d2f14f9be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 271, + 204, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 204, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 204, + 284 + ], + "type": "text", + "content": "Following [33], we set " + }, + { + "bbox": [ + 47, + 271, + 204, + 284 + ], + "type": "inline_equation", + "content": "\\mathrm{obj}_{\\min} = 0.15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": " be the set of pixels on the image border of width " + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "inline_equation", + "content": "w = 3" + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": ". The border loss " + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{b}}(\\pmb{x})" + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 316, + 287, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 316, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 116, + 316, + 287, + 344 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {b}} (\\boldsymbol {x}) = \\left(\\frac {1}{| \\mathcal {B} |} \\sum_ {p \\in \\mathcal {B}} x _ {p}\\right) ^ {2}. \\tag {3}", + "image_path": "dd49693dd1f184df9a7ae442f35208f4dcad23145fc3461b521c4e16ab87b527.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 353, + 174, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 174, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 174, + 367 + ], + "type": "text", + "content": "4. Test Time Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 373, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 409 + ], + "type": "text", + "content": "We now describe our approach. We start with an overview in Sec. 4.1. We describe our regularized loss in Sec. 4.2, and our hyperparameter setting method in Sec. 4.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 416, + 115, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 115, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 115, + 427 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "text", + "content": "The overview of our approach is in Fig. 1. Given a test image " + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "text", + "content": ", we first construct a small training dataset " + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 435, + 287, + 601 + ], + "type": "text", + "content": " by data augmentation. There are various forms of augmentation. However, since we are training without pixel precise ground truth, we need to avoid augmentations which make the salient object more difficult to detect in an image. For example, large crops can remove the salient object. Similarly, adding random image noise changes the distribution of intensity edges, and sparse CRF in Eq. (1) may become less useful for object discovery. We use small random vertical and horizontal image shifts, and random additive shift to color channels (while keeping the shift the same for all image pixels). We found random vertical and horizontal shifts to be the most effective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 602, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 685 + ], + "type": "text", + "content": "Next, we take the base CNN trained for SOD task in image level weakly supervised setting, according to the method proposed in [33], and train it on " + }, + { + "bbox": [ + 46, + 602, + 287, + 685 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 46, + 602, + 287, + 685 + ], + "type": "text", + "content": " for a small number of epochs using the regularized loss approach described in Sec. 4.3. The resulting CNN is called the adapted CNN. Training for a small number of iterations ensures that the adapted CNN retains most of the information useful for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 544, + 96 + ], + "type": "text", + "content": "SOD learned from the large training dataset, as well as ensuring efficiency. In practice, we train for 10 epochs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 104, + 409, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 104, + 409, + 117 + ], + "spans": [ + { + "bbox": [ + 306, + 104, + 409, + 117 + ], + "type": "text", + "content": "4.2. Regularized Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 123, + 545, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 218 + ], + "type": "text", + "content": "We now describe our loss function. For test-time adaptation, we only use some of the regularized loss components from [33]. We do not use batch volumetric loss as it encourages the salient segment to be half of the image size and is far from realistic for most test images. We also do not use center loss. It assumes that the image center belongs to the salient object and is likely not true for many specific test images we wish to adapt CNN for." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 219, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 327 + ], + "type": "text", + "content": "In our regularized loss, we use sparse CRF, minimum volume, and border losses, Eq. (1, 2, 3). Sparse CRF is the most important loss component for discovering salient objects, as it favours segments that align with image intensity edges. Minimum volume loss is necessary to prevent collapse to an empty solution favoured by sparse CRF. Border loss encourages image border to be assigned to the background, and is a realistic assumption for most test images. The complete regularized loss for test time adaptation is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 329, + 349, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 349, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 329, + 349, + 545, + 361 + ], + "type": "interline_equation", + "content": "L _ {\\operatorname {r e g}} (\\boldsymbol {x}) = \\lambda_ {\\operatorname {c r f}} L _ {\\operatorname {c r f}} (\\boldsymbol {x}) + \\lambda_ {m} L _ {\\mathrm {m}} (\\boldsymbol {x}) + \\lambda_ {\\mathrm {b}} L _ {\\mathrm {b}} (\\boldsymbol {x}). \\tag {4}", + "image_path": "f3614fc1527910b5e2c20bc44b8230f3d6c4843c2b2458c823ac536af9858ada.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 371, + 482, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 482, + 384 + ], + "type": "text", + "content": "4.3. Hyperparameter Setting Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 473 + ], + "type": "text", + "content": "Consider the loss components in Eq. (4). To get the best performance during test-time adaptation, it is important to choose a good setting of their relative weights. Each test image has its own best setting, however, we cannot find this setting as the ground truth is not available. Thus we must develop a hyperparameter setting method that does not rely on ground truth." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "text", + "content": "Out of three components in Eq. (4), the border loss is the least important as it contributes the least to the salient object discovery. Indeed, if our only knowledge about a segmentation is that the image border is assigned to the background, then this gives us almost no information to judge whether this segmentation corresponds to the salient object. Therefore, we set " + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\lambda_{b}" + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "text", + "content": " to a low value of 1. Out of the two remaining components, sparse CRF is more important for object discovery. Since objects tend to cause intensity edges in an image, a segment that aligns to image edges is more likely to correspond to a salient object, compared to a segment that just obeys the minimum size constraints. Thus sparse CRF weight " + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "text", + "content": " should be much higher compared to the minimum volume weight " + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 304, + 474, + 545, + 688 + ], + "type": "text", + "content": ". At the same time, the minimum volume weight should be sufficiently large to prevent the collapse to a trivial empty solution. We can compute this sufficiently large weight from the solution obtained from our base CNN." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We now explain how to set " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " relative to " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " so that a trivial solution is avoided. Let " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " be the output of the base" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "2The width of the border can be adjusted relative to the image size, but we train on fixed size images, " + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "inline_equation", + "content": "256\\times 256" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "7363" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "CNN on input image " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": ". Assuming that " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\pmb{x}^i" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " is not empty, we set the weight of minimum volume loss large enough to make the loss on " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\pmb{x}^i" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " smaller than the loss of an empty solution. Then switching to an empty solution from the initial " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\pmb{x}^i" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " is too costly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": "The results produced by the base CNN tend to be sharp, that is most pixels have values close to either 0 or 1. Therefore we will assume that the trivial empty solution corresponds to all pixels having value 0 in the output. Let " + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\pmb{x}^0" + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": " denote such a trivial empty solution, i.e. " + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "inline_equation", + "content": "x_p^0 = 0" + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": " for all pixels " + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": ". Suppose we wish to set sparse-CRF weight to " + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": ". We need to set " + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\lambda_m" + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": " to a value that makes collapse to an empty solution too costly. For this, we need " + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "inline_equation", + "content": "\\lambda_m" + }, + { + "bbox": [ + 46, + 133, + 286, + 227 + ], + "type": "text", + "content": " that satisfies" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 232, + 286, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 232, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 286, + 247 + ], + "type": "interline_equation", + "content": "L _ {\\operatorname {r e g}} \\left(\\boldsymbol {x} ^ {i}\\right) = L _ {\\operatorname {r e g}} \\left(\\boldsymbol {x} ^ {0}\\right) + \\epsilon , \\tag {5}", + "image_path": "b98c527dfc899cd15f2e5e17b57eb5d7541ab1f3ac74d89b954ef97b2c6d1b6f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 251, + 263, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 263, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 263, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 251, + 263, + 262 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 47, + 251, + 263, + 262 + ], + "type": "text", + "content": " is a small constant, set in practice to 0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 263, + 216, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 263, + 216, + 275 + ], + "spans": [ + { + "bbox": [ + 59, + 263, + 216, + 275 + ], + "type": "text", + "content": "Plugging " + }, + { + "bbox": [ + 59, + 263, + 216, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{x}^0" + }, + { + "bbox": [ + 59, + 263, + 216, + 275 + ], + "type": "text", + "content": " in Eqs.(1, 2, 3) we obtain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 279, + 287, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 279, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 58, + 279, + 287, + 294 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {b}} \\left(\\boldsymbol {x} ^ {0}\\right) = 0, \\quad L _ {\\mathrm {c r f}} \\left(\\boldsymbol {x} ^ {0}\\right) = 0, \\quad L _ {\\mathrm {m}} \\left(\\boldsymbol {x} ^ {0}\\right) = \\operatorname {o b j} _ {\\min } ^ {2}. \\tag {6}", + "image_path": "786a2d85774d799f968573a5550e229121a9ad9ad8b3753a5216cb204af023a4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 299, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 287, + 323 + ], + "type": "text", + "content": "Plugging Eq. (6) into Eq. (5) and solving for " + }, + { + "bbox": [ + 47, + 299, + 287, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 47, + 299, + 287, + 323 + ], + "type": "text", + "content": ", we get our formula for setting " + }, + { + "bbox": [ + 47, + 299, + 287, + 323 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 327, + 287, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 327, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 97, + 327, + 287, + 355 + ], + "type": "interline_equation", + "content": "\\lambda_ {m} = \\frac {\\lambda_ {\\mathrm {c r f}} L _ {\\mathrm {c r f}} \\left(\\boldsymbol {x} ^ {i}\\right) + L _ {\\mathrm {b}} \\left(\\boldsymbol {x} ^ {i}\\right) - \\epsilon}{\\operatorname {o b j} _ {\\min } ^ {2} - L _ {\\mathrm {m}} \\left(\\boldsymbol {x} ^ {i}\\right)}. \\tag {7}", + "image_path": "0b28bee3453600c84d31a177ab3c3a13d830e813077832d70bcd90b34a79ad7a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": "Eq. (7) applies when the output " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^i" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " of the base CNN is not a trivial empty solution. If " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^i" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " is an empty solution, it means that the base CNN fails to extract a salient object. In this case, we set both CRF and minimum volume loss to be of equal weight, " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}} = \\lambda_m" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": ". This reflects the fact that extracting a salient object from the current image is difficult and the penalty for not satisfying the minim object size is increased. We do not set " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " higher than " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " as some images may not contain a salient object, in which case an empty solution is appropriate. In addition, while setting " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " high enough would result in some non-empty solution, most likely it would be an erroneous solution, as minimum volume loss is less useful for extracting object segments and, therefore, should not have a weight higher than that of sparse CRF. In practice, we set " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}} = 10^3" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " and compute " + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 360, + 287, + 552 + ], + "type": "text", + "content": " for each test image using Eq. (7)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": "In Table 1 we experimentally evaluate the effectiveness of our method for setting " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": ". We randomly select 500 images from DUTO [37] dataset. We test our adaptive " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": " for each image, and the same fixed setting of " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": " for all test images, in a range from 1 to " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": ", see Table 1. For this experiment, we use a fixed " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}} = 10^{3}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": ". The performance metric is " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": ". The adaptive " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": " performs better than any fixed choice of " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": ". As expected, the performance degrades significantly for the larger values of " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": " as minimum volume loss starts to play a bigger role in the regularized loss function, however, it is less discriminating of salient objects. For smaller " + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 552, + 287, + 707 + ], + "type": "text", + "content": ", the performance suffers as the results on some images collapse to an empty solution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 432, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 432, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 432, + 85 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": "Our implementation is in Pytorch using RTX 2080 GPU. For our base CNN, we use image level weakly supervised SOD approach [33]. We use Unet [29] architecture with ResNeXt [35] fixed features pretrained onImagenet [5] and train on " + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": " images. Unlike [33], we train on larger images and on a larger training dataset, thus getting better results for the base CNN than those reported in [33]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": "We train the base CNN on the training set of DUTS [34], which contains 10,553 images. We then evaluate our test-time CNN adaptation approach on the test set of DUTS (5019 images), DUT-OMRON [37] (5168 images), ECSSD [36] (1000 images), THUR [3] (6233 images), PascalS [19] (850 images), and HKU-IS [18] (4447 images)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "content": "For test time adaptation, given a test image " + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "content": ", we construct a test time adaptation dataset " + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "content": " that has 16 images. These images are constructed from " + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "content": " by shifting to the left or right with probability 0.5, and, in addition, shifting up or down with probability 0.5. Each shift is chosen uniformly from the range [0, 15]. This creates a dataset of images similar to " + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "content": ", and it is unlikely that the salient object is removed out of the sample by these small shifts. We use Adam optimizer [12], a fixed learning rate 0.001, and batch size equal to the size of the test time adaptation dataset, set to 16 images. We train for 10 epochs. The images are scaled to resolution " + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 247, + 546, + 402 + ], + "type": "text", + "content": ". Test time adaptation takes 2.8 seconds per image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": "We use the standard metrics widely adopted for SOD. Our first metric is " + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": " score [1], with " + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\beta^2 = 0.3" + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": ". We also use mae [26], defined as the average absolute per pixel difference between the predicted saliency and ground truth. In addition, we use two newer metrics, namely " + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": " [6], and " + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": " [7]. " + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": " evaluates the structural information for saliency map and region-aware and object-aware structural similarity between saliency maps and ground truth. " + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 304, + 403, + 545, + 510 + ], + "type": "text", + "content": " is based on unification of global and local information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 515, + 385, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 385, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 385, + 528 + ], + "type": "text", + "content": "5.1. SOD results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "content": "In Table 2, we report the results of our approach and compare to prior image level weakly supervised SOD methods. Our results with fixed hyperparameter settings in Eq. (4) are in column \"ours\" (with " + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}} = 20" + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "content": ", which is the best fixed setting according to Table 1). The results with our method for hyperparameter setting for each image individually, as described in Sec. 4.3, are in column 'ours + h'. Except one case, all metrics for our approach with hyperparameter setting method are improved, often significantly. This shows that our hyperparameter setting method is crucial for obtaining a significantly improved performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Dense CRF [13] is often used to post-process CNN results. Therefore, we also post-process base CNN results with dense CRF for comparison. The performance of the base CNN is in column [33], and its performance after post-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "text", + "content": "7364" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 70, + 455, + 97 + ], + "blocks": [ + { + "bbox": [ + 137, + 70, + 455, + 97 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 455, + 97 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 455, + 97 + ], + "type": "table", + "html": "
λm151020501002005001000adaptive
Fβ.741.738.752.755.742.732.728.718.703.765
", + "image_path": "54a76dd05b3812b467845c7b58f15620abe54006bb9f5446c2073e5de3002611.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 112, + 140, + 481, + 444 + ], + "blocks": [ + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "text", + "content": "Table 1. Performance in terms of " + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "text", + "content": " score of fine tuning with a fixed value of " + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "text", + "content": " vs. our adaptive " + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "text", + "content": " computation according to Eq. (7). For all experiments, " + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}} = 10^{3}" + }, + { + "bbox": [ + 46, + 105, + 545, + 129 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 112, + 140, + 481, + 444 + ], + "lines": [ + { + "bbox": [ + 112, + 140, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 112, + 140, + 481, + 444 + ], + "type": "table", + "html": "
Metric[34][17][39][44][42][27][33][33]+dCRFoursours+h
DUTS\\( F_{\\beta} \\)↑.654.614.684.725.747.710.753.763.771.795
mae↓.100.116.091.075.060.076.056.055.054.052
\\( E_s \\)↑.795.772.814.853.859.839.833.842.855.863
\\( S_{\\alpha} \\)↑.748.697.759.813.827.775.791.798.800.817
ECSSD\\( F_{\\beta} \\)↑.823.797.609.810.852.854.868.873.882.911
mae↓.104.110.109.092.071.084.063.060.057.046
\\( E_s \\)↑.869.853.673.836.883.885.884.886.885.909
\\( S_{\\alpha} \\)↑.811.802.756.846.860.834.845.855.859.883
DUTO\\( F_{\\beta} \\)↑.603.622.609.597.701.646.697.710.721.752
mae↓.109.101.109.103.070.087.074.069.066.057
\\( E_s \\)↑.768.776.673.712.816.803.821.827.833.842
\\( S_{\\alpha} \\)↑.725.752.756.733.791.742.763.771.782.798
PascalS\\( F_{\\beta} \\)↑.715.693.713--.751.796.801.801.807
mae↓.139.149.135--.115.083.082.082.084
\\( E_s \\)↑.791.772.790--.817.827.827.826.828
\\( S_{\\alpha} \\)↑.744.717.768--.770.791.793.792.796
HKUIS\\( F_{\\beta} \\)↑--.814.820.878.851.866.871.883.899
mae↓--.084.065.043.059.044.040.039.037
\\( E_s \\)↑--.895.858.919.921.912.925.926.931
\\( S_{\\alpha} \\)↑--.818.860.890.846.857.866.871.879
THUR\\( F_{\\beta} \\)↑---.694.719-.731.749.753.760
mae↓---.086.070-.073.069.067.064
\\( E_s \\)↑---.807.838-.840.842.843.844
\\( S_{\\alpha} \\)↑---.804.810-.794.798.806.819
", + "image_path": "b6c5b71e35bffb17cbda81a77c6d8e99c9939a043072a7c73053f45eaa97f380.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": "Table 2. Comparison to SOD methods that also use image level weak supervision using metrics " + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "inline_equation", + "content": "mae" + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "inline_equation", + "content": "E_s" + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": " to indicate that the higher score is better, and " + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 453, + 546, + 477 + ], + "type": "text", + "content": " to indicate the lower score is better. The best result is in bold, and the second best is underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 496, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 556 + ], + "type": "text", + "content": "processing with dense CRF is in column [33] + dCRF. Our approach is significantly better than [33] with dense CRF post-processing. This illustrates that it is important to include CRF during training in the loss function, rather than use CRF for just post-processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 616 + ], + "type": "text", + "content": "Our test time adaptation improves the performance of [33] across all metrics, except one case. For easier datasets (ECSSD, HKUIS), the performance gain is not as large as the base CNN already performs well, but for harder to fit datasets, the performance gain is significant." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": "Comparing to other SOD methods, our method is better in almost all cases except " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " metric for HKUIS and DUTS. For most datasets, our approach significantly improves " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " metric over the second best method, notably by roughly " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " for DUTO and DUTS datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "In Table 3, we compare our method to weakly supervised SOD methods that use stronger form of weak supervision. In [21], they use bounding box supervision, and in [38, 43]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 496, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 653 + ], + "type": "text", + "content": "they use scribbles. Our method is better than the bounding box method [43] in all cases. Our method is almost always better than [43], and better in about half of the cases compared to [38]. If Fig. 2, we show qualitative comparisons of our method to the scribble supervised methods of [43] and [38]. For this illustration, we chose examples where all methods work relatively well, in terms of the standard metrics. However, our results have better detail preservation qualities, especially when the salient object has sharp edges on the boundary. Observe the detail preservation in the chair, in the last row. This is because CRF based loss function is adapted to the particular test image, driving the segmentation boundary to align with image edges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 659, + 432, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 432, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 432, + 672 + ], + "type": "text", + "content": "5.2. Ablation Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 677, + 471, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 471, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 471, + 689 + ], + "type": "text", + "content": "Hyperparameters in Regularized Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "We first discuss and perform ablation for hyperparameters of our regularized loss function, Eq. (4). The most im" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "7365" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 128, + 69, + 467, + 280 + ], + "blocks": [ + { + "bbox": [ + 128, + 69, + 467, + 280 + ], + "lines": [ + { + "bbox": [ + 128, + 69, + 467, + 280 + ], + "spans": [ + { + "bbox": [ + 128, + 69, + 467, + 280 + ], + "type": "table", + "html": "
DUTSECSSDDUTOPascalSHKUISTHUR
boxes [21]Fβ↑.736.860.686-.853-
mae ↓.079.072.081-.058-
Es ↑.831.889.810-.897-
Sα ↑.789.858.776-.852-
scribbles [43]Fβ↑.747.865.715.788.858.718
mae ↓.062.061.068.140.047.077
Es ↑.865.908.835.798.923.873
Sα ↑------
scribbles [38]Fβ↑.823.900.758.823.896.755
mae ↓.049.049.060.078.038.069
Es ↑.890.908.862.847.938.843
Sα ↑------
oursFβ↑.795.911.752.807.899.760
mae ↓.052.046.057.082.037.064
Es ↑.863.909.842.828.931.844
Sα ↑.817.883.798.796.879.819
", + "image_path": "78b9d39f91795bfd64eb022f95c9e80c17165fa925997d614be26b6145817a1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 342, + 291, + 588 + ], + "blocks": [ + { + "bbox": [ + 50, + 342, + 291, + 588 + ], + "lines": [ + { + "bbox": [ + 50, + 342, + 291, + 588 + ], + "spans": [ + { + "bbox": [ + 50, + 342, + 291, + 588 + ], + "type": "image", + "image_path": "ef7b8bb40b2ddef51973af2c6b910fd95b82b4db13b11e69bc9db2725b21f246.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 598, + 288, + 643 + ], + "lines": [ + { + "bbox": [ + 46, + 598, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 288, + 643 + ], + "type": "text", + "content": "Figure 2. Qualitative comparison. From left to right: input image, ground truth, results of [43], [38], and our results, respectively. Note that [43], [38] both use scribble supervision, a much stronger supervision form compared to what we use." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "important parameter to set is " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " in relation to " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": ", as the correct relative settings of these two parameters ensures there is no collapse to a trivial solution. For this, we have de" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 323, + 339, + 529, + 430 + ], + "blocks": [ + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": "Table 3. Comparison to prior SOD methods that use stronger forms of weak supervision using metrics " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "mae" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "E_s" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": ". In [21], they use bounding box supervision, and in [38,43], they use scribbles. We use " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": " to indicate that the higher score is better, and " + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 287, + 547, + 321 + ], + "type": "text", + "content": " to indicate the lower score is better. The best result is in bold, and second best is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 339, + 529, + 430 + ], + "lines": [ + { + "bbox": [ + 323, + 339, + 529, + 430 + ], + "spans": [ + { + "bbox": [ + 323, + 339, + 529, + 430 + ], + "type": "table", + "html": "
with border losswithout border loss
DUTS.795.789
ECSSD.911.910
DUTO.752.748
PascalS.807.800
HKUIS.899.892
THUR.760.755
", + "image_path": "f88d256151a6f0ecc12559cec9b3e8f1f540f581bbf255283e1cfd27c2d816c4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "lines": [ + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "type": "text", + "content": "Table 4. Performance of our approach with border loss and without border loss. Performance metric is " + }, + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 306, + 438, + 545, + 461 + ], + "type": "text", + "content": " score." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": "veloped an effective automatic approach, Sec. 4.3. Since " + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": " is computed in relation to " + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": ", the setting of " + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": " can be arbitrary (up to the precision that can be handled by the software), " + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": " is adjusted accordingly: for larger " + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{m}}" + }, + { + "bbox": [ + 305, + 483, + 545, + 544 + ], + "type": "text", + "content": " will be automatically set larger according to Eq. (7)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "text", + "content": "The setting of " + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{b}" + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "text", + "content": ", as long as it is much smaller than " + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{crf}}" + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "text", + "content": ", is not important. Parameter " + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{b}}" + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "text", + "content": " controls the border loss that encourages border pixels to be assigned to the background. While this loss is helpful, it is the least important regularized loss component. To show that this loss is the least important, we perform an experiment where we remove border loss from regularized loss (" + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{b} = 0" + }, + { + "bbox": [ + 304, + 544, + 546, + 676 + ], + "type": "text", + "content": "). The results are summarized in Table 4. Without border loss, the results worsen only a little. On the other hand, omitting either CRF or minimum volume losses from our loss function in Eq. (4) results in a catastrophic performance collapse." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In Eq. (1), we set " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma = 0.15" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": " as in [33]. Since the training images are normalized, this setting works well across all datasets. We also experimented with other values of " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": ", see" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "7366" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 252, + 153 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 252, + 153 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 252, + 153 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 252, + 153 + ], + "type": "table", + "html": "
σ = 0.1σ = 0.15σ = 0.20
DUTS.789.795.797
ECSSD.894.911.915
DUTO.751.752.743
PascalS.803.807.803
HKUIS.902.899.895
THUR.763.760.757
", + "image_path": "4378e20d888e4e69d3ff6a0fbb37cfe811b11213ed7fb7f2cf3a5ac475b1d4fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 195, + 271, + 232 + ], + "blocks": [ + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "lines": [ + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "type": "text", + "content": "Table 5. Performance of our approach for different " + }, + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "type": "text", + "content": " values. Performance metric is " + }, + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "type": "inline_equation", + "content": "{F}_{\\beta }" + }, + { + "bbox": [ + 47, + 161, + 287, + 184 + ], + "type": "text", + "content": " score." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 195, + 271, + 232 + ], + "lines": [ + { + "bbox": [ + 64, + 195, + 271, + 232 + ], + "spans": [ + { + "bbox": [ + 64, + 195, + 271, + 232 + ], + "type": "table", + "html": "
|D(I)|248163264
Fβ↑.879.887.890.910.915.913
mae ↓.058.055.049.046.045.046
", + "image_path": "3e503d688d11c80a3d23ccdf16fc2e1c56ce8a8d6a139dafa037ca51c1f41ecc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 60, + 273, + 272, + 311 + ], + "blocks": [ + { + "bbox": [ + 47, + 239, + 286, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 239, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 286, + 262 + ], + "type": "text", + "content": "Table 6. Different choice of test time dataset size for ECSSD dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 273, + 272, + 311 + ], + "lines": [ + { + "bbox": [ + 60, + 273, + 272, + 311 + ], + "spans": [ + { + "bbox": [ + 60, + 273, + 272, + 311 + ], + "type": "table", + "html": "
# epoch2510203050
Fβ↑.873.883.910.915.915.912
mae ↓.062.053.047.045.046.047
", + "image_path": "bc89a302da96f30bdefbd43c7251e82546f2ab41960fbb6e17c5979e3f5766bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 362, + 283, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 283, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 283, + 374 + ], + "type": "text", + "content": "Table 5. The performance is stable for a range of " + }, + { + "bbox": [ + 47, + 362, + 283, + 374 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 362, + 283, + 374 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 383, + 185, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 185, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 185, + 396 + ], + "type": "text", + "content": "5.3. Other Hyperparameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "text", + "content": "We now perform ablation experiments for various hyperparameter settings (other than those in the regularized loss function) of our approach. We perform experiments on ECSSD dataset, and evaluate the sensitivity of our approach to the size of the test time adaptation dataset " + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "text", + "content": ", the number of epoch performed for test time adaptation, and different data augmentation strategies. For all the experiments, the performance metric is " + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "inline_equation", + "content": "mae" + }, + { + "bbox": [ + 46, + 401, + 286, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "type": "text", + "content": "In Table 6, we show how performance depends on varying the size of the test time dataset " + }, + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "type": "text", + "content": ". Our standard setting for the size of " + }, + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 46, + 498, + 286, + 557 + ], + "type": "text", + "content": " is 16. Decreasing it to 8 slightly degrades the performance, increasing it to 64 results in almost no change, but the running time is twice longer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 286, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 629 + ], + "type": "text", + "content": "In Table 7, we show how performance depends on increasing the number of epoch for test time adaptation. Our standard setting for the number of epoch is 10. Decreasing it to 2 noticeably degrades the performance, increasing it to 50 results in only a slight improvement but significantly increases the running time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "In Table 8, we show the effect of different augmentation strategies. Our main experiments use small random shifts in the vertical and horizontal directions. We also tested adding random Gaussian noise to images, a standard data augmentation technique, but it significantly degrades the performance. Adding Gaussian noise decreases the effectiveness of our sparse CRF loss as neighboring pixels become less" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 565, + 106 + ], + "blocks": [ + { + "bbox": [ + 47, + 319, + 286, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 319, + 286, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 286, + 340 + ], + "type": "text", + "content": "Table 7. Different number of test time adaptation iterations for ECSSD dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 565, + 106 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 565, + 106 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 565, + 106 + ], + "type": "table", + "html": "
typeshiftsGauss noisecolor perturbationshifts+color
Fβ↑.911.854.897.909
mae↓.046.060.050.049
", + "image_path": "624e0365c1aac42d391b4c65353ca9574b8a9eda3b446523c81a02a34320fca7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 306, + 152, + 577, + 195 + ], + "blocks": [ + { + "bbox": [ + 306, + 114, + 545, + 136 + ], + "lines": [ + { + "bbox": [ + 306, + 114, + 545, + 136 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 545, + 136 + ], + "type": "text", + "content": "Table 8. Different augmentation strategies when fine tuning on ECSSD dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 152, + 577, + 195 + ], + "lines": [ + { + "bbox": [ + 306, + 152, + 577, + 195 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 577, + 195 + ], + "type": "table", + "html": "
DUTSECSSDDUTOPascalSHKUISTHUR
% improved79.383.785.381.276.775.0
mean Fβ increase.048.053.064.17.037.036
mean Fβ decrease.021.009.011.08.019.015
", + "image_path": "7b4667ab6f34772c872e7628f91d648afe8a30225f7d0662e24511d2f083705c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 204, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 306, + 204, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 545, + 247 + ], + "type": "text", + "content": "Table 9. Percentage of images for which performance improves as the result of test-time adaptation. For the improved (and degraded) images, we also show the mean increase (and the decrease) in " + }, + { + "bbox": [ + 306, + 204, + 545, + 247 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 306, + 204, + 545, + 247 + ], + "type": "text", + "content": " score." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 274, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 334 + ], + "type": "text", + "content": "strongly connected. We also tested random color perturbation where all pixels change the value of a color channel by the same random value. This works much better than Gaussian noise, but not as well as random shifts. Shifts and color perturbation work as well as shifts alone." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 348, + 441, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 441, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 441, + 361 + ], + "type": "text", + "content": "5.4. Additional Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "text", + "content": "During test-time adaptation, the majority, but not all images, improve. In Table 9, we show the percentage of images that improve, in terms of their " + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "text", + "content": " score, for each of the test datasets. In addition, we show the mean improvement of " + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "text", + "content": " score, averaged over the images that improve, and the mean decline of " + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "text", + "content": " score, averaged over the images that decline in performance. As can be seen from the table, for all dataset, over " + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 305, + 369, + 545, + 525 + ], + "type": "text", + "content": " of images improve during test-time adaptation. Also, the average improvement is significantly larger than the average decline of the results. Thus overall, test-time adaptation significantly improves the performance. In the supplementary materials, we show the most improved and the most degraded images during test time adaptation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 547, + 383, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 383, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 383, + 559 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 713 + ], + "type": "text", + "content": "We propose the first approach to test time adaptation for SOD with image level weak supervision. We develop a regularized loss function particularly effective for test time adaptation, and a method to find an appropriate setting of the hyperparameters in our loss function. Our approach uses a standard CNN architecture and a loss function that is intuitive and simple to interpret. We achieve a new state of the art in image level supervised SOD, and a better or competitive result with the prior work that uses stronger forms of weak supervision. The drawback of our approach is that the running time for testing is more expensive, since the base CNN needs to be fine-tuned for each test image." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "7367" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 190 + ], + "type": "text", + "content": "[2] Yuri Boykov, Olga Veksler, and Ramin Zabih. A variable window approach to early vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(12):1283-1294, December 1998. An earlier version of this work appeared in CVPR '97. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 225 + ], + "type": "text", + "content": "[3] Ming-Ming Cheng, NiloyJ. Mitra, Xiaolei Huang, and ShiMin Hu. Salientshape: group saliency in image collections. The Visual Computer, 30(4):443-453, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "text", + "content": "[4] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 304 + ], + "type": "text", + "content": "[5] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[6] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 407 + ], + "type": "text", + "content": "[7] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 698-704, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 429 + ], + "type": "text", + "content": "[8] I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT Press, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 473 + ], + "type": "text", + "content": "[9] Xiaodi Hou and Liqing Zhang. Saliency detection: A spectral residual approach. In 2007 IEEE Conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "text", + "content": "[10] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "text", + "content": "[11] Vidit Jain and Erik Learned-Miller. Online domain adaptation of a pre-trained cascade of classifiers. In CVPR 2011, pages 577-584. IEEE, 2011. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[13] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. In Neural Information Processing Systems, pages 109-117, 2011. 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[14] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Neural Information Processing Systems, 2012. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[15] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. Backpropagation applied to handwritten zip code recognition. Neural Comput., 1(4):541-551, December 1989. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[17] Guanbin Li, Yuan Xie, and Liang Lin. Weakly supervised salient object detection using image labels. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[18] G. Li and Y. Yu. Deep contrast learning for salient object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 478-487, June 2016. 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[19] Yin Li, Xiaodi Hou, Christof Koch, James M Rehg, and Alan L Yuille. The secrets of salient object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 280-287, 2014. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 263 + ], + "type": "text", + "content": "[20] Jiawei Liu, Jing Zhang, and Nick Barnes. Semi-supervised salient object detection with effective confidence estimation. arXiv preprint arXiv:2112.14019, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[21] Yuxuan Liu, Pengjie Wang, Ying Cao, Zijian Liang, and Rynson WH Lau. Weakly-supervised salient object detection with saliency bounding boxes. IEEE Transactions on Image Processing, 30:4423-4435, 2021. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 342 + ], + "type": "text", + "content": "[22] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "text", + "content": "[23] Yunqiu Lv, Bowen Liu, Jing Zhang, Yuchao Dai, Aixuan Li, and Tong Zhang. Semi-supervised active salient object detection. Pattern Recognition, 123:108364, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 378, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 422 + ], + "type": "text", + "content": "[24] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6608-6618, October 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 423, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 487 + ], + "type": "text", + "content": "[25] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. In Advances in Neural Information Processing Systems, pages 204-214, 2019. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 490, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 543 + ], + "type": "text", + "content": "[26] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In Conference on Computer Vision and Pattern Recognition, pages 733-740. IEEE, 2012. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "text", + "content": "[27] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[28] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 3, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[30] David A Ross, Jongwoo Lim, Ruei-Sung Lin, and Ming-Hsuan Yang. Incremental learning for robust visual track" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "7368" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "ing. International journal of computer vision, 77(1):125-141, 2008. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[31] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised CNN segmentation. In Conference on Computer Vision and Pattern Recognition, pages 1818-1827, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[32] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised CNN segmentation. In European Conference on Computer Vision, pages 524-540, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[33] Olga Veksler. Regularized loss for weakly supervised single class semantic segmentation. In European Conference on Computer Vision, pages 348-365. Springer, 2020. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[34] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 343 + ], + "type": "text", + "content": "[35] Saining Xie, Ross B. Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 376 + ], + "type": "text", + "content": "[36] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 1155-1162. IEEE, 2013. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 422 + ], + "type": "text", + "content": "[37] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Conference on Computer Vision and Pattern Recognition, pages 3166-3173. IEEE, 2013. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 521 + ], + "type": "text", + "content": "[38] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2021, pages 3234-3242. AAAI Press, 2021. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[39] Yu Zeng, Yunzhi Zhuge, Hutchuan Lu, Lihe Zhang, Mingyang Qian, and Yizhou Yu. Multi-source weak supervision for saliency detection. In Conference on Computer Vision and Pattern Recognition, pages 6074-6083, 2019. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 624 + ], + "type": "text", + "content": "[40] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[41] Jing Zhang, Yuchao Dai, Tong Zhang, Mehrtash Harandi, Nick Barnes, and Richard Hartley. Learning saliency from single noisy labelling: A robust model fitting perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2866-2873, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[42] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 330 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "conference on computer vision, pages 349-366. Springer, 2020. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[43] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via scribble annotations. In Conference on Computer Vision and Pattern Recognition, pages 12546-12555, 2020. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[44] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9029-9038, 2018. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[45] Shun Zhang, Jia-Bin Huang, Jongwoo Lim, Yihong Gong, Jinjun Wang, Narendra Ahuja, and Ming-Hsuan Yang. Tracking persons-of-interest via unsupervised representation adaptation. International Journal of Computer Vision, 128(1):96-120, 2020. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[46] Ting Zhao and Xiangqian Wu. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3085-3094, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 330 + ], + "type": "text", + "content": "[47] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. 2020. 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 742 + ], + "type": "text", + "content": "7369" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_content_list.json b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..004c7422f3679b7f4e3b5afd1714fb169ac0bdea --- /dev/null +++ b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_content_list.json @@ -0,0 +1,1872 @@ +[ + { + "type": "text", + "text": "Test of Time: Instilling Video-Language Models with a Sense of Time", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Piyush Bagad", + "bbox": [ + 163, + 181, + 279, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Amsterdam", + "bbox": [ + 137, + 200, + 307, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Makarand Tapaswi", + "bbox": [ + 408, + 181, + 560, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IIIT Hyderabad", + "bbox": [ + 431, + 200, + 537, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cees G.M. Snoek", + "bbox": [ + 674, + 181, + 818, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Amsterdam", + "bbox": [ + 661, + 200, + 831, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bpiyush.github.io/testoftime-website", + "bbox": [ + 349, + 219, + 619, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modelling and understanding time remains a challenge in contemporary video understanding models. With language emerging as a key driver towards powerful generalization, it is imperative for foundational video-language models to have a sense of time. In this paper, we consider a specific aspect of temporal understanding: consistency of time order as elicited by before/after relations. We establish that seven existing video-language models struggle to understand even such simple temporal relations. We then question whether it is feasible to equip these foundational models with temporal awareness without re-training them from scratch. Towards this, we propose a temporal adaptation recipe on top of one such model, VideoCLIP, based on post-pretraining on a small amount of video-text data. We conduct a zero-shot evaluation of the adapted models on six datasets for three downstream tasks which require varying degrees of time awareness. We observe encouraging performance gains especially when the task needs higher time awareness. Our work serves as a first step towards probing and instilling a sense of time in existing video-language models without the need for data and compute-intensive training from scratch.", + "bbox": [ + 75, + 300, + 473, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 651, + 209, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised pretraining at scale on multimodal web corpora tied with powerful architectures [107] has led to foundational models [12] for images [2, 49, 59, 83, 84] and videos [2, 6, 26, 109, 119, 126]. These models have enabled remarkable improvements on a plethora of downstream video-language tasks such as video-text retrieval, video question-answering, and action recognition. Given the cost and difficulty of video annotations, even for a small amount of downstream data, such foundational models are emerging as the de-facto backbone for zero-shot [119, 122, 127] and few-shot generalization [2]. However, it remains unclear if these video-language models capture essential properties of a video beyond what can be learned from static images, most notably: time.", + "bbox": [ + 75, + 674, + 468, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many before us have shown that existing video-language", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0547fbbf6ef66282f5fcf4e972eacb39c83d3355c7d50989704ec08ef25296db.jpg", + "image_caption": [ + "1." + ], + "image_footnote": [], + "bbox": [ + 531, + 268, + 612, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f93c9201089595b0e6f3dc220b735703fcb6f6f7abfd03c99e9340eb033c094.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 270, + 700, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2b8ca251bc4b75da58650cb76cfb9473621692e19b795ee3e859abd8df6b267.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 270, + 789, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/448db9de872f0dbabe74cb5e24df2aa24362cab688f77f3bbbb708efca7cfe3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 273, + 877, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b85dd3be0e0a4f8b3dfdb23fc5784ab19f471e1f361b8615af310f2c2f4936c.jpg", + "image_caption": [ + "2." + ], + "image_footnote": [], + "bbox": [ + 529, + 323, + 612, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed74bac97aba135b1c3be4dd919b1069c3acd91772eb876ab5e20e007483f5c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 323, + 700, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/baba012e51a4f48e4bfb214edace9dc46cf7c58b5d4e4285d4b9bbf2224f1721.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 321, + 789, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bea8a049582657baeb8d9eeffda3dc108efd817149214d60d920e479d1cf7e77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 321, + 877, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1c0853e3f209fca343a837a022dbe64f9b4f2675c37e7ec83d1dcf51ad3ee7c.jpg", + "image_caption": [ + "3." + ], + "image_footnote": [], + "bbox": [ + 529, + 373, + 612, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7573eb9b7773f08653a5476096a0e2536069043db96368106b8de73408b349b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 373, + 700, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4b5f7e459b996314e387c808d1cf50f7c00ef6483813a8b8a8f477f987373ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 373, + 789, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac9aa9f03961a12e1b69245c1882606afd83c7a3fa18d17c3e0b15a127d7727b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 373, + 877, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47bf84d000abbcaf973fc77fa37a1151a1b671dabad57d3b2f4babe2bff5250c.jpg", + "image_caption": [ + "4.", + "A.", + "C.", + "The baby eats food after it looks into the camera", + "Figure 1. Can you match the correct video-text pairs? Understanding the time order of events across video and language is necessary to be able to solve this task. See footnote on next page for answers." + ], + "image_footnote": [], + "bbox": [ + 529, + 422, + 612, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4eb3058bf121d59246e5f101f4b33c6ab5c75f13f1834e381acdf2a44daf585.jpg", + "image_caption": [ + "Dog runs away before it brings a ball to the man" + ], + "image_footnote": [], + "bbox": [ + 616, + 422, + 700, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f44f942b293c338a0cbded02f28ecc265b49c3e17a2ad80f9722c6b762ffbbe.jpg", + "image_caption": [ + "B.", + "The dog brings a a ball to the man before it runs away", + "D.", + "The baby looks into the camera after it eats food" + ], + "image_footnote": [], + "bbox": [ + 704, + 422, + 789, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccc38d16e1c011d3ebf940f0488db10055b7d95938d7918501954b9fedb1a695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 422, + 877, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models [6, 57, 66, 119] can achieve impressive performance on several video benchmarks [22, 41, 120] without reliably encoding time [13, 56, 59]. For example, Buch et al. [13] show that a model that uses a single (carefully selected) frame often outperforms recent video-language models [57, 119] on standard video benchmarks such as MSR-VTT [120]. Lei et al. [56] report similar findings with a single-frame pretraining approach. These findings hint at a lack of time awareness in video models. However, it remains unclear if these findings are caused, indeed, by the lack of time in video models or whether the benchmarks themselves do not mandate time awareness. Furthermore, there is no clear definition of what it means for a model to be time aware. In this paper, we strive to shed light on all these factors of time awareness in video-language models.", + "bbox": [ + 496, + 612, + 892, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a first step, we consider a simple notion of understanding time, i.e., understanding temporal relations such as before and after [4]. Consider the task presented in Fig. 1. A time invariant model shall be able to associate (A) with (1)", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2503", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "or (2) and (B) with (3) or (4) based on static frames alone. But to distinguish between (1) and (2), one needs to be able to understand time order and connect it across video and language1. Thus, the first question we ask in Section 3: do the representations learnt by foundational video-language models encode this sense of time? To reliably attribute lack of time awareness to models and not existing benchmarks, we design our own synthetic dataset to probe models for this sense of time. We create video-language pairs that show a sequence of two events. Then, we alter the order of events either in the text or the video and check if models can connect the order in video and language. We find that existing video-language models indeed struggle to associate the time order across video and language.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of these findings, the second question we ask in Section 4 is: can we adapt a video-language model, without expensive re-training from scratch, to instill this sense of time? Towards this, we take inspiration from literature on understanding time in natural language, where there has been much work on developing time aware language models [20, 36, 37, 130, 131]. Our objective is to instill time awareness in a video-language model without having to pretrain from scratch. To do that, we propose TACT: Temporal Adaptation by Consistent Time-ordering based on two key components: (i) we artificially create samples that provide temporal signal, for example, by flipping the order of events in the video or the text, and (ii) we introduce a modified contrastive loss to learn time order consistency based on these samples. Instead of training from scratch, we adapt an existing video-language model, VideoCLIP [61], using the paradigm of post-pretraining on a small amount of videotext data [66, 121]. We demonstrate the effectiveness of TACT in connecting the time order in video and language on four diverse real datasets in Section 5.", + "bbox": [ + 75, + 301, + 472, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, in line with the original motive of video-language models for zero-shot generalization, we evaluate in Section 6 our TACT-adapted model for three sets of tasks on six downstream datasets which require a varying degree of time awareness. On tasks that need higher time awareness, with the appropriate choice of adaptation dataset, TACT outperforms a strong baseline that is based on post-pretraining on canonical clip-text pairs without consideration of time-order.", + "bbox": [ + 75, + 604, + 472, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (i) We show that existing video-language models struggle to associate time order in video and language through controlled experiments on synthetic data and several evaluations on real datasets. (ii) Based on VideoCLIP [119], we propose TACT, a method for temporal adaptation using this time order consistency without having to pretrain from scratch. (iii) We demonstrate improved zero-shot generalizability of TACT-adapted models on tasks that require higher time awareness.", + "bbox": [ + 75, + 741, + 472, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 785, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly discuss recent advances in video-language models followed by their consideration of time.", + "bbox": [ + 500, + 116, + 890, + 147 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Foundational video-language models. Large-scale datasets, self-supervision, and the advent of Transformers [107] have led to the emergence of powerful encoders for images [21,39,103], videos [5,11,24,104,117], language models [19,64,69,86] and even universal encoders [32,46]. These encoders form the basis for several vision-language foundational models. Popular image-language models such as CLIP [83] and ALIGN [49] are trained on massive datasets by using web images and alt-text. Similarly, videolanguage models are catching up and can be categorised into two broad directions: (i) adapting image-language models for videos [8, 23, 50, 51, 63, 66, 71, 110, 112, 121], and (ii) pure video-based models that are learned using large video-text datasets [3,7,27-29,31,58,62,65,67,68,95,119]. Recently, a new paradigm of post-pretraining has emerged where an existing image- or video-language model goes through another stage of self-supervised pretraining on a small amount of video data before it is evaluated on downstream tasks [66, 121]. This is promising as it circumvents the prohibitive cost of pretraining on large datasets from scratch. In [66], the post-pretraining uses time-invariant mean-pooling, while [121] strives to bridge the domain gap between image captions and video subtitles. In contrast, our proposed temporal adaptation involves post-pretraining of VideoCLIP [119] with a small amount of data that instills the model to learn the time-order of events in a video.", + "bbox": [ + 496, + 152, + 893, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Time in vision. Time separates videos from static images or an unordered set of frames. While modeling time remains a challenge, it also presents a natural source of supervision that has been exploited for self-supervised learning. For example, as a proxy signal by posing pretext tasks involving spatio-temporal jigsaw [1, 44, 53], video speed [10, 17, 48, 94, 111, 124], arrow of time [78, 80, 114], frame/clip ordering [25, 70, 90, 97, 118], video continuity [61], or tracking [45, 108, 113]. Several works have also used contrastive learning to obtain spatio-temporal representations by (i) contrasting temporally augmented versions of a clip [47, 77, 81], or (ii) encouraging consistency between local and global temporal contexts [9, 18, 85, 123]. Nevertheless, it remains unclear if the learnt representations actually encode time reliably. Time-aware features have also been explored for specific downstream tasks such as action recognition [30, 100, 101]. There has been some very recent work on evaluating self-supervised video representations [87, 99] on their temporal recognition ability instead of only relying on time as a guidance for training.", + "bbox": [ + 496, + 551, + 893, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the same spirit, a related direction pursues evaluation and benchmarking of time awareness in video datasets [88], models [13, 14, 30, 56, 89, 125] or both [43, 92]. Huang et", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Answers: (A)-(2), (B)-(1), (C)-(4), (D)-(3).", + "bbox": [ + 94, + 886, + 331, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2504", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "al. [43] measure the effect of motion on temporal action recognition to find that only a subset of classes in UCF-101 and Kinetics-400 require motion information. Ghodrati et al. [30] propose new tasks to evaluate temporal asymmetry, continuity and causality in video models. Our work derives inspiration from these but applies more generally to video-language models as language provides a basis for open-world generalization.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Time in language. Time has also been extensively studied in the natural language literature. Early works identified temporal structures in language such as temporal prepositions and quantifiers [4, 79]. More recent literature focuses on tasks such as extracting temporal relations [35, 72-74], as well as temporal reasoning [36, 37, 82, 130, 131]. For example, Han et al. [36, 37] and Zhou et al. [131] pretrain language models specifically to focus on understanding temporal relations such as before, after, during, etc. The emergence of large language models has also spurred an increased interest in developing benchmarks to test for time awareness in these models [20, 75, 76, 102, 106, 129]. For example, Ning et al. [75] propose a new benchmark of reading comprehension with questions involving before/after relations. Since temporal relations in language are grounded in the video, we draw inspiration from [36, 37, 131] and aim to instill time awareness in video-language models.", + "bbox": [ + 75, + 220, + 472, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Time in video-language models appears implicitly in tasks like video-text alignment [38, 98] and temporal grounding [42, 60]. In this work, we focus on self-supervised video-language models that can generalize to a variety of tasks rather than models designed for a specific task, e.g., temporal grounding. Some recent works have shown the under-utilisation of time in classic video-text benchmarks such as MSR-VTT [120], YouCook [132], ActivityNet [22], and DiDeMo [41]. For example, [13, 56, 57] discover that on several benchmarks, using only one or few frames or clips achieves competitive performance. Adaptations of the popular CLIP architecture for videos (e.g., CLIP4Clip [66]) show that weighted pooling of frames already achieves impressive performance on retrieval benchmarks.", + "bbox": [ + 75, + 488, + 472, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These raise some key questions: do existing videolanguage models truly understand time in the sense of correctly associating order of events in language and video? If not, can we adapt them to instill time awareness? Our work addresses these questions. There has been some work in using time-order across video and language as a source of self-supervision. Specifically, concurrent to our work, both Sun et al. [96] and Cao et al. [15] propose fine-grained temporal alignment between video and text as the pretraining objective. Different from these works, we consider the notion of time-order and we aim to adapt a given videolanguage model using post-pretraining, which circumvents the need for a new round of compute-intense pretraining.", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d9515c00a7205a5547f167d10e5e105eb89b5511d2539566856bab8d7dfb87f1.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed task to evaluate time-order consistency across synthetic video-language pairs having before/after relations. We also define a control task to check if the synthetic videos are considered out-of-distribution by the model." + ], + "image_footnote": [], + "bbox": [ + 506, + 95, + 710, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/11780cf09827d603373cf3543228a8e0569fd9158c10836b95ef3a12274c5a67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 98, + 864, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Do Video-Language Models Sense Time?", + "text_level": 1, + "bbox": [ + 498, + 299, + 864, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Probing video-language models for temporal understanding is an open direction of research. In this work, we consider a specific sense of temporal understanding: consistency in the order of events in a video with the associated textual description. For example, consider a text description: A red circle appears before a yellow circle. This imposes an order constraint on the video stream to have the event red circle appears happen before the event yellow circle appears. Can existing video-language models connect time-order in text with that in video? To answer this, we design an experiment with synthetic data.", + "bbox": [ + 496, + 325, + 890, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic dataset. We construct simple videos that comprise of a pair of events such as the ones mentioned above. We generate $N = 180$ video-language pairs as a combination of $C = 6$ colors, $S = 3$ shapes, and $|\\tau| = 2$ temporal relations: before and after. The corresponding caption describes the order of events connected with a before/after temporal relation. We call this caption as an attractor since it is consistent with the time-ordering in the video. Likewise, we construct a distractor where we flip the order of event descriptions while retaining the temporal relation. An example pair is illustrated in Fig. 2 (left). Ideally, a time aware video-language model should be able to associate the video with the temporally consistent text, or vice versa. We refer to this task as time-order consistency check. To rule out the possibility that synthetic videos are out-of-distribution, we also perform the same experiment with canonical clips with a single event and the text describes that same event as shown in Fig. 2 (right). We refer to this as the control task.", + "bbox": [ + 496, + 496, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Choice of models. We consider recent video-language models, broadly categorized into three groups: (i) image-language models like CLIP [83] that are adapted to videos [23,66,128], (ii) pure video-language models trained on a contrastive learning objective [6,16,119], and (iii) pure video-language models trained on a masking objective [29].", + "bbox": [ + 496, + 773, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Findings. We evaluate video-to-text and text-to-video retrieval on both time-order consistency and control tasks.", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2505", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/9775656550ca29f10ef4c0d57fe5d78f7cdcba4fb278dbcbe2abf3c610dfdf19.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Results on synthetic control ( ⌒ ) and time-order consistency ( ⌒ ) task as described in Fig. 2. Existing video-language models show random performance on our time-order task." + ], + "table_body": "
ParadigmMethodVideo-to-TextText-to-Video
◎◎◎◎
Chance-50.050.050.050.0
Image-Language adapted to videoCLIP4Clip [66]49.451.150.049.4
CLIP2Video [23]100.047.897.852.3
CenterCLIP [128]91.746.197.251.1
Video-Language ContrastiveVideoCLIP [119]87.151.166.748.3
Frozen in Time [6]97.849.4100.050.6
VindLU [16]98.052.0100.051.1
Video-Language MaskingBridgeFormer [29]100.051.197.242.2
", + "bbox": [ + 78, + 88, + 472, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Tab. 1, we observe that while most video-language models perform well on the control task, all of them struggle and perform on par with random chance on the temporal task. This gap in performance deserves attention given the importance of time in videos. Note that while synthetic data allows for controlled experiments, we also expand this evaluation to real video datasets in the following section.", + "bbox": [ + 75, + 332, + 468, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Adaptation by Consistent Time-Ordering", + "text_level": 1, + "bbox": [ + 76, + 454, + 447, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe a post-pretraining recipe to instill a video-language model with a sense of time. We propose TACT: Temporal Adaptation by Consistency of Time-order, that is based on two key components: (i) we artificially create samples that provide temporal signals, e.g., by flipping the order of events; and (ii) we introduce a modified contrastive loss to learn temporal consistency based on these samples. We start by defining the notation and then describe the key components of our adaptation recipe.", + "bbox": [ + 75, + 479, + 468, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminaries. Let $\\mathcal{V}$ be the space of video clips and $\\mathcal{T}$ be the space of text clips. Consider two non-overlapping video clips $v_{i}, v_{j} \\in \\mathcal{V}$ . Let $\\zeta_{i}, \\zeta_{j} \\in \\mathcal{T}$ be their respective captions. Let $\\tau$ be a temporal relation, $\\tau \\in \\{\\text{before, after}\\}$ . Then, we denote a stitched and time-order consistent clip as $(u_{ij}, t_{ij})$ , where $u_{ij} := [v_{i}; v_{j}]$ , $t_{ij} := [\\zeta_{i}; \\tau; \\zeta_{j}]$ , and $[\\cdot; \\cdot]$ denotes concatenation. Note that depending on $\\tau$ , the order of $v_{i}$ and $v_{j}$ may need to change in $u_{ij}$ . For brevity, we drop the subscripts and refer to the stitched pair as $(u, t)$ unless stated otherwise.", + "bbox": [ + 75, + 622, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Time-order reversal. The classic contrastive learning paradigm for video-language models aligns components of a video clip $v_{i}$ with its text counterpart $\\zeta_{i}$ and contrasts against other texts $\\zeta_{j}$ that usually describe a completely different clip. This makes such models ignore the finer details of temporal understanding as it is easier to contrast the negatives by simply focusing on the objects or the scene. This is evident from simple bag-of-word-like methods that", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/166102ea621382b55c43696766ac1217cc91881b037ffa89b9afd5f188d2163d.jpg", + "image_caption": [ + "Figure 3. Overview of TACT. Along with the usual contrastive loss, where negatives come from other samples in the batch, we make use of time-order reversal within the same sample and cross samples to generate additional negatives for both video and text. We also extend the contrastive loss to time-order reversed video and text corresponding to reverse consistency $\\mathcal{L}_r$ ." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 866, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are shown to work well for contrastive learning, both on the visual (e.g., CLIP4Clip [66]) and textual (e.g., MILNCE [67]) modalities. We hypothesize that unless there are negatives in a contrastive setup that contain the same scenes and objects, models need not learn a sense of time. Thus, we present a simple strategy to generate negatives that force the learning process to focus on the temporal order.", + "bbox": [ + 496, + 478, + 890, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define a time-order reversal function $\\mathbb{T}$ that operates on the stitched video clip or text description and temporally swaps its constituents:", + "bbox": [ + 498, + 583, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {T} (u) = \\mathbb {T} ([ v _ {i}; v _ {j} ]) := [ v _ {j}; v _ {i} ], \\quad \\text {a n d} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 637, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {T} (t) = \\mathbb {T} ([ \\zeta_ {i}; \\tau ; \\zeta_ {j} ]) := [ \\zeta_ {j}; \\tau ; \\zeta_ {i} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 655, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An illustration of $\\mathbb{T}$ is shown in Fig. 3. Note that $\\mathbb{T}$ does not reverse the actual video, i.e., time does not flow backwards, but only changes the order in which events happen in the stitched clips. Our objective is to train a model that is able to distinguish between the original pair $(u,t)$ and time-reversed versions $(u,\\mathbb{T}(t))$ , and $(\\mathbb{T}(u),t)$ .", + "bbox": [ + 496, + 683, + 890, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss function. We assume access to an existing pre-trained video-language model with a visual encoder $f_{\\theta}$ and text encoder $g_{\\phi}$ . We obtain the video encoding $\\mathbf{z}_u \\coloneqq f_{\\theta}(u) \\in \\mathbb{R}^d$ and the text encoding $\\mathbf{z}_t \\coloneqq g_{\\phi}(t) \\in \\mathbb{R}^d$ . Our goal is to adapt $\\Theta = \\{\\theta, \\phi\\}$ via post-pretraining such that the resulting model is time aware while maintaining its original performance on tasks such as retrieval. As we aim to use a small amount of data, we only update some parameters of", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2506", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the model $(\\Theta)$ , such as the last few layers.", + "bbox": [ + 76, + 90, + 354, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now introduce our recipe for temporal adaptation based on the InfoNCE loss [105] to learn time-order sensitive video-text correspondence. For a positive (or time-order consistent) video-text pair $(u, t)$ , we first define a forward loss where the stitched pair is in its original time-order.", + "bbox": [ + 75, + 106, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\sum_ {(u, t) \\in \\mathcal {B}} \\left(\\operatorname {T N C E} \\left(\\mathbf {z} _ {u}, \\mathbf {z} _ {t}\\right) + \\operatorname {T N C E} \\left(\\mathbf {z} _ {t}, \\mathbf {z} _ {u}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 186, + 467, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where TNCE is the Noise Contrastive Estimation (NCE) loss for temporal adaptation, defined as:", + "bbox": [ + 76, + 224, + 467, + 255 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T N C E} \\left(\\mathbf {z} _ {u}, \\mathbf {z} _ {t}\\right) := - \\log \\frac {\\exp \\left(\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {t}\\right)}{\\sum_ {t ^ {\\prime} \\in \\mathcal {B} _ {t}} \\exp \\left(\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {t ^ {\\prime}}\\right) + \\mathcal {C} ^ {\\text {t i m e}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 260, + 467, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is the batch of $(u,t)$ pairs and $\\mathcal{B}_t$ specifically refers to other stitched text captions in the batch. $\\mathcal{C}^{\\mathrm{time}}$ accumulates negatives defined using time-order reversal as:", + "bbox": [ + 76, + 296, + 467, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} ^ {\\text {t i m e}} = \\alpha_ {\\text {s a m e}} \\exp (\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {\\mathbb {T} (t)}) + \\alpha_ {\\text {c r o s s}} \\sum_ {t ^ {\\prime} \\in \\mathcal {B} _ {t} \\backslash \\{t \\}} \\exp (\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {\\mathbb {T} \\left(t ^ {\\prime}\\right)}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 345, + 467, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{\\mathrm{same}}$ controls the effect of contrasting text from the same sample but with reversed text time-order, i.e., $\\mathbb{T}(t)$ , and $\\alpha_{\\mathrm{cross}}$ encourages the model to contrast between reversed versions of other text captions, i.e., $\\mathbb{T}(t')$ . Note that when both $\\alpha_{\\mathrm{same}}$ and $\\alpha_{\\mathrm{cross}}$ are 0, we revert back to the standard NCE formulation, albeit on stitched pairs. While Eq. (4) corresponds to the video-text loss $\\mathrm{TNCE}(\\mathbf{z}_u,\\mathbf{z}_t)$ , the text-video loss $\\mathrm{TNCE}(\\mathbf{z}_t,\\mathbf{z}_u)$ is defined symmetrically. Furthermore, we also apply a reverse loss $\\mathcal{L}_r$ to bring time-order reversed versions of both the video and the text together. Note that as we consider $(u,t)$ as a positive pair, $(\\mathbb{T}(u),\\mathbb{T}(t))$ also form a positive pair,", + "bbox": [ + 76, + 381, + 467, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\sum_ {\\left(\\mathbb {T} (u), \\mathbb {T} (t)\\right) \\in \\mathcal {B}} \\left(\\operatorname {T N C E} \\left(\\mathbf {z} _ {\\mathbb {T} (u)}, \\mathbf {z} _ {\\mathbb {T} (t)}\\right) + \\operatorname {T N C E} \\left(\\mathbf {z} _ {\\mathbb {T} (t)}, \\mathbf {z} _ {\\mathbb {T} (u)}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 568, + 470, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the TNCE term operates on time-reversed clips and $\\mathcal{C}^{\\mathrm{time}}$ contrasts $(\\mathbb{T}(u),\\mathbb{T}(t))$ with un-reversed text clips in the batch $(\\mathbb{T}(u),t)$ . The overall loss function is defined as:", + "bbox": [ + 76, + 614, + 467, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {f} + \\beta \\mathcal {L} _ {r}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 667, + 467, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depending on the choice of loss coefficients $\\alpha_{\\mathrm{same}}, \\alpha_{\\mathrm{cross}}, \\beta \\in \\{0,1\\}$ , we can vary properties of the adapted model. For example, setting $\\alpha_{\\mathrm{same}} = 1$ encourages high sensitivity to time-order reversal. As we will see empirically, the loss coefficients also provide the flexibility to adapt the model to various datasets.", + "bbox": [ + 76, + 689, + 467, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We illustrate this temporal extension of the contrastive loss in Fig. 3 (best seen in colour). $\\mathbb{T}$ illustrates the time order reversal function. The top half corresponds to $\\mathcal{L}_f$ while the bottom half visualizes $\\mathcal{L}_r$ . In particular, the top-left quadrant alone corresponds to the standard contrastive loss on stitched pairs. While the green diagonal terms are positive pairs, the red diagonal terms are the strongest drivers for instilling temporal understanding in the model.", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e47de4e73308fcdeb1dccf000a151003764f08b6eee353f09930b6d0d019777f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTrainValidationTestEgoLength
NvNcNvNcNvNc(s)
TEMPO3,90428,4274111,0003961,000X30
ActivityNet7,44095,474453906456912X120
Charades5,26299,9285001,0005001,000X30
Charades-Ego2,679155,3065001,000210420✓31
", + "bbox": [ + 504, + 89, + 890, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Statistics of datasets we consider for temporal adaptation. $N_{\\mathcal{V}}$ is the number of unique videos and $N_{\\mathcal{C}}$ is the number of stitched clips. Based on $N_{\\mathcal{V}}$ , TEMPO and Charades-Ego are smaller as compared to ActivityNet and Charades.", + "bbox": [ + 498, + 200, + 890, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments: TACT Ablations", + "text_level": 1, + "bbox": [ + 500, + 282, + 779, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base model. We demonstrate the effectiveness of TACT as an adaptation recipe on top of VideoCLIP [119] owing to its simple architecture, contrastive objective, and use of pre-computed S3D [116] features that enable faster experimentation and allow encoding a long temporal context ( $\\sim 32$ secs). We initialize $\\Theta$ from the model pretrained on HowTo100M [68] and post-pretrain on multiple datasets.", + "bbox": [ + 498, + 311, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. One of our key objectives is to post-pretrain on a small amount of data in contrast to massive pretraining datasets such as WebVid2M [7] or HowTo100M [68]. We consider dense video captioning datasets that offer sufficient diversity in terms of size, backgrounds, clip durations, viewpoints and activities. Specifically, we experiment with: (i) TEMPO [40]: the subset of stitched diverse third-person videos from DiDeMo [41] with text descriptions for fixed 5s segments that contain before/after relations; (ii) ActivityNet Captions [55]: a dense video captioning dataset with human-centric actions; (iii) Charades [93]: a scripted indoor daily human activities video dataset; and (iv) Charades-Ego [91]: similar to Charades, scripted human activities from the egocentric viewpoint. To construct stitched clips, we randomly sample any two non-overlapping clip-text pairs in the video. Since we require stitched clips instead of raw videos, we create new splits for each dataset (see Tab. 2).", + "bbox": [ + 496, + 422, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate the post-pretrained model using two sets of metrics: (i) standard retrieval metrics, recall $R@1$ , $R@5$ , $R@10$ and median-rank evaluated on stitched video-text clips; and (ii) time-order consistency, i.e., the fraction of videos for which the model correctly associates text that is time order consistent with the video:", + "bbox": [ + 498, + 700, + 890, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\text {t i m e}} := \\frac {1}{| \\mathcal {D} |} \\sum_ {(u, t) \\in \\mathcal {D}} \\mathbb {I} [ d (\\mathbf {z} _ {u}, \\mathbf {z} _ {t}) < d (\\mathbf {z} _ {u}, \\mathbf {z} _ {\\mathbb {T} (t)}) ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 800, + 890, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(u,t)$ are time-order consistent pairs, $\\mathcal{D}$ is the dataset, and $d(\\cdot ,\\cdot)$ is a distance metric based on cosine similarity.", + "bbox": [ + 498, + 849, + 888, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-pretraining details. We freeze the word embeddings", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2507", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and layers 1 to 5 for both the video and text encoders in VideoCLIP. For adaptation, we use the Adam optimizer [54] with learning rate $5e^{-6}$ , batch size 32 trained on a single node with 4 GeForce GTX 1080 GPUs. On TEMPO, we train for 60 epochs while on the other datasets, we train for 10 epochs and pick the checkpoint that maximizes the geometric mean of $R@1$ and $A_{\\mathrm{time}}$ on the respective validation set. A typical adaptation run takes about 1-3 hours.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on the test set. Results in Tab. 3 show that TACT\\* with optimal loss coefficients outperforms TACT† (all 0 loss coefficients) and the zero-shot baseline (no postpretraining similar to the synthetic data experiment), both on the retrieval and time-order consistency tasks. This indicates the robustness of the adaptation.", + "bbox": [ + 75, + 219, + 468, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of loss coefficients. Choosing appropriate values for loss coefficients $\\Theta_l \\coloneqq \\{\\alpha_{\\mathrm{same}}, \\alpha_{\\mathrm{cross}}, \\beta\\}$ allows the model to learn various aspects and adapt using different datasets. On each dataset, we vary $\\Theta_l \\in \\{0,1\\}^3$ and find the best configuration based on the GeometricMean $(R@1, \\max(A_{\\mathrm{time}} - 50, 0))$ on the validation sets. The above metric ensures the geometric mean is not overpowered by $R@1$ . The results are shown in Tab. 4. As $\\alpha_{\\mathrm{same}}$ is directly responsible for discriminating between original and time-reversed orders, unsurprisingly, setting it to 1 is necessary to achieve the best results on $A_{\\mathrm{time}}$ for all the datasets. For TEMPO and Charades-Ego, using all loss components (all 1) provides the best results, whereas $\\alpha_{\\mathrm{cross}} = 1$ and $\\beta = 0$ achieves a better trade-off for ActivityNet and Charades. Choosing $\\beta = 1$ leads to an improve", + "bbox": [ + 75, + 316, + 472, + 545 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24c1cfe1554551d57cd0fad26ded5e51496f93812b07f6db51ae33274824e12c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodRetrievalTime-order
R@1↑MedR ↓Atime ↑
TEMPOZero-shot3.749.048.1
TACT†7.713.046.5
TACT*9.39.066.5
ActivityNetZero-shot1.144.049.6
TACT†5.834.059.7
TACT*5.835.085.7
CharadesZero-shot1.3170.047.1
TACT†5.338.573.5
TACT*5.735.077.0
Charades-EgoZero-shot1.664.053.7
TACT†6.435.060.1
TACT*10.128.568.2
", + "bbox": [ + 91, + 570, + 455, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Results for the best TACT model on test sets. $\\mathrm{TACT}^{\\star}$ has optimal loss coefficients and $\\mathrm{TACT}^{\\dagger}$ is a baseline with all coefficients 0. On time order, TACT generalizes well with $\\mathrm{TACT}^{\\star}$ outperforming the baselines. On retrieval, for TEMPO and Charades-Ego, $\\mathrm{TACT}^{\\star}$ outperforms the baseline as their optimal models have $\\beta = 1$ which helps retrieval with a small amount of data.", + "bbox": [ + 75, + 811, + 468, + 896 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0ad7ac203b57eb66af53635addd83ee6d31039c0ecd77dd1f9031cf8dc486c5.jpg", + "image_caption": [ + "Figure 4. Time-distance between stitched clips in datasets for temporal adaptation $(\\Delta_{\\mathrm{time}})$ . TEMPO has stitched clips close to each other while those in Charades-Ego are farthest apart suggesting a correlation between $\\Delta_{\\mathrm{time}}$ and difficulty of temporal adaptation." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 887, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ment in retrieval performance for TEMPO and Charades-Ego but leads to a decline for ActivityNet and Charades. We attribute this to the number of unique videos in the train set for these datasets. As ActivityNet and Charades have more videos than TEMPO or Charades-Ego (see train $N_{\\mathcal{V}}$ Tab. 2) additional positives introduced by setting $\\beta = 1$ are not necessary and in fact, hurt performance. Finally, we note that carefully setting the value of $\\alpha_{\\mathrm{same}}$ provides a convenient trade-off between spatial and temporal understanding. Please refer to the supplement for detailed experiments.", + "bbox": [ + 496, + 258, + 890, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What makes temporal adaptation hard? We observe a large gap in $A_{\\mathrm{time}}$ between TEMPO and ActivityNet. We hypothesize that the distance (in seconds) between the two clips ( $\\Delta_{\\mathrm{time}}$ ) in a stitched clip is strongly correlated with the difficulty of adaptation. Intuitively, it is easier to infer the time-order consistency for a stitched clip $u$ with text $t$ that has distant constituent clips $v_i, v_j$ since the objects and scene could be vastly different. In contrast, it is harder to discern the correct time-order when the constituent clips are closer in time. Fig. 4 shows distribution of $\\Delta_{\\mathrm{time}}$ for each dataset. Indeed, $\\Delta_{\\mathrm{time}}$ in ActivityNet (58.8s) is much higher than that in TEMPO (6.4s) making the task harder on TEMPO. To further test our hypothesis, we conduct a controlled experiment where we gradually vary the distribution of $\\Delta_{\\mathrm{time}}$ for Charades-Ego to match it to that of TEMPO. We find a strong correlation ( $\\rho = 0.92$ ) between $\\Delta_{\\mathrm{time}}$ and hardness of adaptation. More details are in the supplement.", + "bbox": [ + 496, + 415, + 892, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments: Downstream Evaluation", + "text_level": 1, + "bbox": [ + 500, + 686, + 846, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The goal of video-language foundation models is to generalize in a zero- or few-shot manner to a diverse range of downstream tasks. We evaluate TACT models on three sets of downstream tasks that need low-to-high time awareness.", + "bbox": [ + 496, + 713, + 890, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline: Standard post-pretraining. Comparing our temporally adapted models with pretrained VideoCLIP is not fair since adapted models see data beyond the pretraining phase. In addition to the zero-shot comparison, we compare against a baseline model that is trained for standard video-text retrieval on the same datasets as temporal adaptation. Instead of using stitched clips, we use simple canonical pairs, i.e., $(v_{i},\\zeta_{i})$ instead of $(u_{ij},t_{ij})$ .", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2508", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c1792805d00aeba217706a514c757e950fdf5b95ffba5f542c038e12ef2191af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Loss coefficientsTEMPOActivityNetCharadesCharades-Ego
α_sameα_crossβR@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑
Chance0.1500.050.00.1453.050.00.1500.050.00.1500.050.0
0008.314.049.46.430.057.35.742.071.52.944.064.6
0018.214.049.55.627.047.04.258.075.13.241.565.2
0108.215.049.36.129.078.85.245.078.93.438.064.5
0118.114.049.55.827.048.34.258.075.13.141.067.0
1006.420.060.65.928.079.16.138.076.33.242.066.1
1016.524.062.95.626.063.14.951.078.03.339.070.7
1105.924.059.76.029.086.36.643.077.83.740.567.9
1117.517.062.55.727.063.85.151.077.73.838.568.3
", + "bbox": [ + 78, + 88, + 893, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88c28340b4223d9fd0b9ac0aa171790f630340f36804ad87921a0ade34088e3c.jpg", + "table_caption": [ + "Table 4. Impact of loss coefficients for TACT post-pretraining on validation sets of various datasets. For each dataset, the corresponding color-marked row denotes the best configuration based on the geometric mean of $R@1$ and $A_{\\mathrm{time}}$ . TACT is able to connect time-order in video and language while maintaining its retrieval capabilities across several datasets." + ], + "table_footnote": [], + "table_body": "
AdaptationI. Text-to-Video RetrievalII. Temporal VQAIII. Action-to-Video Retrieval
DatasetMethodMSR-VTTYouCookIINext-QA (ATP)AGQASSv2SSv2 (events)Temporal
R@1R@5R@10R@1R@5R@10AccuracyAccuracymAPmAPmAP
-Chance0.10.51.00.10.51.020.050.00.62.02.0
-None10.623.429.918.245.559.923.449.93.46.413.0
TEMPOBaseline12.029.337.321.548.261.825.050.83.96.815.9
TACT12.826.535.720.445.158.727.657.14.27.716.1
ActivityNetBaseline15.734.444.915.638.851.423.750.73.77.016.0
TACT13.829.639.616.036.949.825.450.33.57.216.2
CharadesBaseline12.325.833.621.548.661.726.050.54.17.113.7
TACT11.725.232.422.449.162.425.254.84.37.814.1
Charades-EgoBaseline13.127.534.519.447.160.824.349.74.06.914.7
TACT11.124.630.621.948.261.925.658.43.96.913.5
", + "bbox": [ + 76, + 329, + 890, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results on downstream zero-shot evaluation with tasks requiring increasing time awareness from I to III. None corresponds to direct evaluation of the VideoCLIP model on the downstream dataset. Green denotes an improvement for the TACT adapted model w.r.t. the baseline, red denotes a deterioration. As we move from tasks that need low to high time awareness, the effectiveness of TACT increases. See Sec. 6 for a more detailed discussion. The table is best viewed on screen in colour.", + "bbox": [ + 75, + 575, + 893, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluating TACT adapted models on synthetic data. On the video-to-text variant, TACT adapted on TEMPO achieves $64.4\\%$ , ActivityNet $52.5\\%$ , Charades $65.0\\%$ , Charades-Ego $85.6\\%$ . This is usually higher than the performance that non-adapted models achieve in Tab. 1. This highlights that TACT models learn useful signals to match time-order in video and language.", + "bbox": [ + 75, + 657, + 470, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "I. Text-to-video retrieval. We consider two widely used benchmarks: MSR-VTT [120] and YouCookII [132] and adopt standard retrieval metrics. Recent work has identified a bias for spatial understanding in these datasets [8, 13, 43, 56, 59, 66]. Thus, we consider this class of tasks as requiring low time awareness. As shown in Tab. 5 set I, on MSR-VTT [120], we observe that TACT is worse (marked in red) or at par with the baseline across datasets. This", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aligns well with findings in [13, 56] that these benchmarks do not need time awareness. On YouCookII [132], TACT models based on Charades(-Ego) outperform the baseline (marked in green). We believe this is a consequence of a lower domain shift between YouCookII and Charades.", + "bbox": [ + 496, + 657, + 890, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "II. Temporal video QA. Next, we use subsets of recently released multiple-choice video question answering benchmarks: Next-QA [115] and AGQA [34]. The idea is to check if we can probe models for temporal understanding by asking questions with temporal language. Buch et al. [13] identify a subset of Next-QA, dubbed as ATP-hard², with a higher concentration of temporally challenging data. For AGQA, we pick a subset of $\\sim 6k$ questions that explicitly have a question with before/after relation. We consider", + "bbox": [ + 496, + 739, + 893, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Available here: github.com/StanfordVL/atp-video-language", + "bbox": [ + 514, + 886, + 872, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2509", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "these benchmarks as requiring a moderate-high level of time awareness and AGQA in particular is also close to our adaptation task. We use accuracy as the standard metric. We observe (see Tab. 5 set II) that indeed TACT almost always outperforms (marked in green) baselines on both Next-QA and AGQA. TEMPO-adapted TACT seems to generalize particularly well on both benchmarks. Likewise, Charades-adapted TACT does well on AGQA since AGQA is also based on the Charades videos accounting for reduced domain shift. We affirm that temporal adaptation is useful, especially when the downstream tasks require it.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "III. Action-to-video retrieval. Finally, we consider action recognition benchmarks such as Something-Something (SSv2) [33] and Temporal [88]. SSv2 was designed to capture richer temporal information [33, 56]. We follow Lei et al. [56], who propose the template-retrieval task that encourages temporal modelling and use their evaluation split3 containing $C = 174$ actions and $K = 12$ videos per class. Interestingly, different actions in SSv2 require differing levels of time awareness. We create a subset SSv2 (events) with $C_{\\text{events}} = 49$ actions that have at least two verbs in the label as the occurrence of multiple verbs is indicative of multiple events occurring in sequence. Finally, we also evaluate against the Temporal benchmark [88], a combination of 50 action classes from SSv2 [33] and Kinetics-400 [52] for which temporal information is deemed to be essential for recognition. Similar to text-to-video retrieval, we use the action class as a text query and obtain a ranking over all videos. Different from the retrieval setup, since a single query has multiple correct answers (up to $K = 12$ videos), we report mAP as the metric for these benchmarks. This task set needs high time awareness. Furthermore, unlike QA tasks in II, there is a shift in several (uncontrolled) factors as we move from temporal adaptation task to these tasks. From Tab. 5, we observe that TEMPO and Charades-adapted models generalize well across set III benchmarks. ActivityNet-adapted TACT underperforms on SSv2 but outperforms the baseline on strongly temporal actions in SSv2 (events) and Temporal. Finally, TACT adapted on Charades-Ego is at-par or slightly worse than the baseline on SSv2 variants, and also on Temporal, perhaps due to the shift from egocentric to third-person videos. Overall, despite SSv2 and Temporal requiring high time awareness, TACT models show promising zero-shot generalization with the right choice of the adaptation dataset.", + "bbox": [ + 75, + 261, + 472, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 76, + 787, + 326, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to other temporal prompts. The time-order of events in language can be described using various sentence structures. Although we train video-language models using before/after relations, it is natural to ask if", + "bbox": [ + 75, + 818, + 470, + 878 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61536b1950e6d38904c1f501b46a3b61bd72c9c087aac9deb41cbf0b23d8ebd8.jpg", + "image_caption": [ + "Figure 5. Models trained by TACT with before/after relations generalize to a new kind of prompt such as First, ..., then ... indicating learning of the underlying true time-order of events." + ], + "image_footnote": [], + "bbox": [ + 531, + 97, + 862, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model still correctly associates time-order for a different prompt such as First, ..., then, ... To systematically test this, we gather event pairs $E_1$ , $E_2$ ( $E_1$ occurs before $E_2$ in the video) for each sample in the validation set and stitch them using three prompts as follows: (i) $E_1$ before $E_2$ , (ii) $E_2$ after $E_1$ , (iii) First $E_1$ , then $E_2$ . As shown in Fig. 5, TACT-adapted models generalize well to a new prompt (iii). This substantiates the learning of time-order of events rather than merely learning the order of words in the sentence.", + "bbox": [ + 496, + 324, + 893, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While we present a promising way of instilling time in video-language models, our work is limited to the VideoCLIP [119] pretrained model. Our initial experiments with Frozen in Time [6] were not as promising, perhaps because it uses a much shorter temporal context (4 frames). Please see the supplement for results on more pretrained models. Furthermore, we consider a specific definition of time awareness derived from temporal relations like before/after. It is natural to ask if this can be extended to more general notions of temporality, e.g., as defined by Allen [4]. Finally, there can always be more downstream tasks considered such as (spatio-)temporal localization.", + "bbox": [ + 496, + 465, + 893, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. Given the essence of time in video-language models, we present a simple experiment based on synthetic data to test for time awareness. We find that existing models lack a sense of time defined in terms of consistency of order of events in video and language. To fill this gap, building upon VideoCLIP [119], we present TACT, a recipe to instill this sense of time in video-language models. Finally, we analyze the zero-shot generalizability of TACT-adapted models to a diverse set of tasks. We hope that this work provokes further probing and instilling time awareness in video-language models, and also inspires other adaptations of foundational models to solve various challenging tasks.", + "bbox": [ + 496, + 652, + 895, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We acknowledge support from ELLIS Amsterdam and the AMS Scholarship to Piyush. We thank Dr Dennis Koelma for their help with compute infrastructure and Dr Hazel Doughty for useful discussions.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3Available here: github.com/jayleicn/singularity", + "bbox": [ + 94, + 886, + 392, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2510", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Winter Conference on Applications of Computer Vision (WACV), pages 179-189. IEEE, 2019. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1", + "[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. ArXiv, abs/2006.16228, 2020. 2", + "[4] James F. Allen. Towards a general theory of action and time. In Artificial Intelligence, 1984. 1, 3, 8", + "[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 2", + "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 4, 8", + "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 5", + "[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval, 2022. 2, 7", + "[9] Nadine Behrmann, Mohsen Fayyaz, Juergen Gall, and Mehdi Noroozi. Long short view feature decomposition via contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9224-9233, 2021. 2", + "[10] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9922–9931, 2020. 2", + "[11] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021. 2", + "[12] Rishi Bommasani et al. On the opportunities and risks of foundation models. ArXiv, abs/2108.07258, 2021. 1", + "[13] Shyamal Buch, Cristobal Eyzaguirre, Adrien Gaidon, Jia-jun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting" + ], + "bbox": [ + 86, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the \"Video\" in Video-Language Understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7", + "[14] Petr Byvshev, Pascal Mettes, and Yu Xiao. Are 3d convolutional networks inherently biased towards appearance? Computer Vision and Image Understanding, 220:103437, 2022. 2", + "[15] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision (ECCV), 2022. 3", + "[16] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051, 2022. 3, 4", + "[17] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised spatio-temporal representation learning using variable playback speed prediction. IEEE Access, 9:79562-79571, 2021. 2", + "[18] Ishan Rajendra Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TcIr: Temporal contrastive learning for video representation. Computer Vision and Image Understanding (CVIU), 219:103406, 2022. 2", + "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2", + "[20] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 3", + "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 2", + "[22] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 1, 3", + "[23] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 3, 4", + "[24] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 2", + "[25] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Conference on Computer" + ], + "bbox": [ + 509, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2511", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition (CVPR), pages 3636-3645, 2017. 2", + "[26] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 1", + "[27] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 2", + "[28] Valentin Gabeur, Chen Sun, Alahari Karteek, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision (ECCV), 2020. 2", + "[29] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 16167-16176, June 2022. 2, 3, 4", + "[30] Amir Ghodrati, Efstratios Gavves, and Cees G. M. Snoek. Video time: Properties, encoders and evaluation. In British Machine Vision Conference (BMVC), 2018. 2, 3", + "[31] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. ArXiv, abs/2011.00597, 2020. 2", + "[32] Rohit Girdhar, Mannat Singh, Nikhil Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. Conference on Computer Vision and Pattern Recognition (CVPR), pages 16081-16091, 2022. 2", + "[33] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 8", + "[34] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7", + "[35] Rujun Han, Qiang Ning, and Nanyun Peng. Joint event and temporal relation extraction with shared representations and structured prediction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3", + "[36] Rujun Han, Xiang Ren, and Nanyun Peng. Deer: A data efficient language model for event temporal reasoning. ArXiv, abs/2012.15283, 2020. 2, 3", + "[37] Rujun Han, Xiang Ren, and Nanyun Peng. Econet: Effective continual pretraining of language models for event temporal reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 2, 3", + "[38] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2906, 2022. 3" + ], + "bbox": [ + 86, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In _Conference on Computer Vision and Pattern Recognition (CVPR), CVPR '16, pages 770–778, June 2016._2", + "[40] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), pages 1380-1390, 2018. 5", + "[41] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. International Conference on Computer Vision (ICCV), pages 5804-5813, 2017. 1, 3, 5", + "[42] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), 2018. 3", + "[43] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7366–7375, 2018. 2, 3, 7", + "[44] Yuqi Huo, Mingyu Ding, Haoyu Lu, Zhiwu Lu, Tao Xiang, Ji-Rong Wen, Ziyuan Huang, Jianwen Jiang, Shiwei Zhang, Mingqian Tang, Songfang Huang, and Ping Luo. Self-supervised video representation learning with constrained spatiotemporal jigsaw. In International Joint Conference on Artificial Intelligence (IJCAI), 2021. 2", + "[45] A. Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. ArXiv, abs/2006.14613, 2020. 2", + "[46] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver: General perception with iterative attention. In International Conference on Machine Learning (ICML), 2021. 2", + "[47] S. Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9950-9960, 2021. 2", + "[48] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), pages 425-442, 2020. 2", + "[49] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2", + "[50] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2", + "[51] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2512", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8", + "[53] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Association for the Advancement of Artificial Intelligence (AAAI), pages 8545-8552, 2019. 2", + "[54] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6", + "[55] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-Captioning Events in Videos. In International Conference on Computer Vision (ICCV), 2017. 5", + "[56] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning, 2022. 1, 2, 3, 7, 8", + "[57] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. Conference on Computer Vision and Pattern Recognition (CVPR), pages 7327-7337, 2021. 1, 3", + "[58] Gen Li, Nan Duan, Yuejian Fang, Daxin Jiang, and Ming Zhou. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Association for the Advancement of Artificial Intelligence (AAAI), 2020. 2", + "[59] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 1, 7", + "[60] Juncheng Li, Junlin Xie, Long Qian, Linchao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal grounding with structured variational cross-graph correspondence learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 3022-3031, 2022. 3", + "[61] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. ArXiv, abs/2112.05883, 2022. 2", + "[62] Kevin Lin, Alex Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z. Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Ego-centric video-language pretraining. ArXiv, abs/2206.01670, 2022. 2", + "[63] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Y. Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. ArXiv, abs/2208.03550, 2022. 2", + "[64] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2", + "[66] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2, 3, 4, 7", + "[67] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. Conference on Computer Vision and Pattern Recognition (CVPR), pages 9876-9886, 2020. 2, 4", + "[68] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In International Conference on Computer Vision (ICCV), 2019. 2, 5", + "[69] Tomas Mikolov, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In International Conference on Learning Representations (ICLR), 2013. 2", + "[70] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), pages 527-544, 2016. 2", + "[71] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. European Conference on Computer Vision (ECCV), 2022. 2", + "[72] Qiang Ning, Zhili Feng, and Dan Roth. A structured learning approach to temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2017. 3", + "[73] Qiang Ning, Zhili Feng, Hao Wu, and Dan Roth. Joint reasoning for temporal and causal relations. In Association of Computational Linguistics (ACL), 2018. 3", + "[74] Qiang Ning, Sanjay Subramanian, and Dan Roth. An improved neural baseline for temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3", + "[75] Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. Torque: A reading comprehension dataset of temporal ordering questions. ArXiv, abs/2005.00242, 2020. 3", + "[76] Qiang Ning, Ben Zhou, Zhili Feng, Haoruo Peng, and Dan Roth. Cogcomptime: A tool for understanding time in natural language. ArXiv, abs/1906.04940, 2018. 3", + "[77] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. Conference on Computer Vision and Pattern Recognition (CVPR), pages 11200-11209, 2021. 2", + "[78] Lyndsey C. Pickup, Zheng Pan, Donglai Wei, YiChang Shih, Changshui Zhang, Andrew Zisserman, Bernhard" + ], + "bbox": [ + 509, + 90, + 893, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2513", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Scholkopf, and William T. Freeman. Seeing the arrow of time. In *Conference on Computer Vision and Pattern Recognition* (CVPR), pages 2043-2050, 2014. 2", + "[79] Ianthe Pratt and Nissim Francez. Temporal prepositions and temporal generalized quantifiers. Linguistics and Philosophy, 24:187-222, 2001. 3", + "[80] Will Price and Dima Damen. Retro-actions: Learning 'close' by time-reversing 'open' videos. International Conference on Computer Vision Workshops (ICCVW), pages 1371-1380, 2019. 2", + "[81] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, H. Wang, Serge J. Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6960-6970, 2021. 2", + "[82] Lianhui Qin, Aditya Gupta, Shyam Upadhyay, Luheng He, Yejin Choi, and Manaal Faruqui. Timedial: Temporal commonsense reasoning in dialog. In Association of Computational Linguistics (ACL), 2021. 3", + "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3", + "[84] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1", + "[85] Adrià Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Patraucean, Florent Altch'e, Michael Valko, Jean-Bastien Grill, Aäron van den Oord, and Andrew Zisserman. Broaden your views for self-supervised video learning. International Conference on Computer Vision (ICCV), pages 1235–1245, 2021. 2", + "[86] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108, 2019. 2", + "[87] Madeline Chantry Schiappa, Yogesh Singh Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ArXiv, abs/2207.00419, 2022. 2", + "[88] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. CoRR, abs/1907.08340, 2019. 2, 8", + "[89] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? International Conference on Computer Vision (ICCV), pages 9641-9649, 2021. 2", + "[90] Vivek Sharma, Makarand Tapaswi, and Rainer Stiefelhagen. Deep Multimodal Feature Encoding for Video Ordering. In International Conference on Computer Vision Workshops (ICCVW), 2019. 2", + "[91] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint" + ], + "bbox": [ + 86, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "modeling of first and third-person videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7396-7404, 2018. 5", + "[92] Gunnar A. Sigurdsson, Olga Russakovsky, and Abhinav Kumar Gupta. What actions are needed for understanding human actions in videos? 2017 IEEE International Conference on Computer Vision (ICCV), pages 2156-2165, 2017. 2", + "[93] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, Cham, 2016. 5", + "[94] Ankit Singh, Omprakash Chakraborty, Ashutosh Varshney, Rameswar Panda, Rogério Schmidt Feris, Kate Saenko, and Abir Das. Semi-supervised action recognition with temporal contrastive learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 10384-10394, 2021. 2", + "[95] Chen Sun, Austin Myers, Carl Vondrick, Kevin P. Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. International Conference on Computer Vision (ICCV), pages 7463-7472, 2019. 2", + "[96] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[97] Tomoyuki Suzuki, Takahiro Itazuri, Kensho Hara, and Hirokatsu Kataoka. Learning spatiotemporal 3d convolution with video order self-supervision. In European Conference on Computer Vision Workshops (ECCV), pages 0–0, 2018. 2", + "[98] Makarand Tapaswi, Martin Bauml, and Rainer Stiefelhagen. Book2movie: Aligning video scenes with book chapters. In CVPR, June 2015. 3", + "[99] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees G. M. Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 2", + "[100] Fida Mohammad Thoker, Hazel Doughty, and Cees Snoek. Tubelet-contrastive self-supervision for video-efficient generalization, 2023. 2", + "[101] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Skeleton-contrastive 3d action representation learning. ACM MM, 2021. 2", + "[102] Shivin Thukral, Kunal Kukreja, and Christian Kavouras. Probing language models for understanding of temporal expressions. In BLACKBOXNLP, 2021. 3", + "[103] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. arXiv preprint arXiv:2105.01601, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2514", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[104] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6450-6459, 2018. 2", + "[105] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. ArXiv, abs/1807.03748, 2018. 5", + "[106] Siddharth Vashishtha, Adam Poliak, Yash Kumar Lal, Benjamin Van Durme, and Aaron Steven White. Temporal reasoning in natural language inference. In FINDINGS, 2020. 3", + "[107] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017. 1, 2", + "[108] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2563-2572, 2021. 2", + "[109] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 1", + "[110] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 2", + "[111] Jiangliu Wang, Jianbo Jiao, and Yun-Hui Liu. Self-supervised video representation learning by pace prediction. In European Conference on Computer Vision (ECCV), pages 504-521, 2020. 2", + "[112] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 2", + "[113] X. Wang, A. Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2561-2571, 2019. 2", + "[114] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 8052-8060, 2018. 2", + "[115] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7", + "[116] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In European Conference on Computer Vision (ECCV), pages 318–335, Cham, 2018. 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[117] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin P. Murphy. Rethinking spatiotemporal feature learning for video understanding. ArXiv, abs/1712.04851, 2017. 2", + "[118] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10334-10343, 2019. 2", + "[119] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing (EMNLP), pages 6787-6800, 2021. 1, 2, 3, 4, 5, 8", + "[120] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1, 3, 7", + "[121] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Rui Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. ArXiv, abs/2209.06430, 2022. 2", + "[122] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1", + "[123] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Back to the future: Cycle encoding prediction for self-supervised contrastive video representation learning. ArXiv, abs/2010.07217, 2020. 2", + "[124] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6548-6557, 2020. 2", + "[125] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. Clevrer: Collision events for video representation and reasoning. ArXiv, abs/1910.01442, 2020. 2", + "[126] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel C. F. Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. ArXiv, abs/2111.11432, 2021. 1", + "[127] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 1", + "[128] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In SIGIR, 2022. 3, 4" + ], + "bbox": [ + 501, + 90, + 892, + 898 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "2515", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[129] Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. \"going on a vacation\" takes longer than \"going for a walk\": A study of temporal commonsense understanding. *ArXiv*, abs/1909.03065, 2019. 3", + "[130] Ben Zhou, Qiang Ning, Daniel Khashabi, and Dan Roth. Temporal common sense acquisition with minimal supervision. ArXiv, abs/2005.04304, 2020. 2, 3", + "[131] Ben Zhou, Kyle Richardson, Qiang Ning, Tushar Khot, Ashish Sabharwal, and Dan Roth. Temporal reasoning on implicit events from distant supervision. ArXiv, abs/2010.12753, 2021. 2, 3", + "[132] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence (AAAI), pages 7590-7598, 2018. 3, 7" + ], + "bbox": [ + 78, + 90, + 468, + 304 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "2516", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_model.json b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f798b002638611b436d21110f3acaf3ef8089839 --- /dev/null +++ b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_model.json @@ -0,0 +1,3341 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Test of Time: Instilling Video-Language Models with a Sense of Time" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.182, + 0.28, + 0.2 + ], + "angle": 0, + "content": "Piyush Bagad" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.202, + 0.308, + 0.216 + ], + "angle": 0, + "content": "University of Amsterdam" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.182, + 0.562, + 0.2 + ], + "angle": 0, + "content": "Makarand Tapaswi" + }, + { + "type": "text", + "bbox": [ + 0.432, + 0.202, + 0.538, + 0.216 + ], + "angle": 0, + "content": "IIIT Hyderabad" + }, + { + "type": "text", + "bbox": [ + 0.675, + 0.182, + 0.819, + 0.198 + ], + "angle": 0, + "content": "Cees G.M. Snoek" + }, + { + "type": "text", + "bbox": [ + 0.663, + 0.201, + 0.832, + 0.216 + ], + "angle": 0, + "content": "University of Amsterdam" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.22, + 0.62, + 0.233 + ], + "angle": 0, + "content": "bpiyush.github.io/testoftime-website" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.633 + ], + "angle": 0, + "content": "Modelling and understanding time remains a challenge in contemporary video understanding models. With language emerging as a key driver towards powerful generalization, it is imperative for foundational video-language models to have a sense of time. In this paper, we consider a specific aspect of temporal understanding: consistency of time order as elicited by before/after relations. We establish that seven existing video-language models struggle to understand even such simple temporal relations. We then question whether it is feasible to equip these foundational models with temporal awareness without re-training them from scratch. Towards this, we propose a temporal adaptation recipe on top of one such model, VideoCLIP, based on post-pretraining on a small amount of video-text data. We conduct a zero-shot evaluation of the adapted models on six datasets for three downstream tasks which require varying degrees of time awareness. We observe encouraging performance gains especially when the task needs higher time awareness. Our work serves as a first step towards probing and instilling a sense of time in existing video-language models without the need for data and compute-intensive training from scratch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.652, + 0.21, + 0.668 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Self-supervised pretraining at scale on multimodal web corpora tied with powerful architectures [107] has led to foundational models [12] for images [2, 49, 59, 83, 84] and videos [2, 6, 26, 109, 119, 126]. These models have enabled remarkable improvements on a plethora of downstream video-language tasks such as video-text retrieval, video question-answering, and action recognition. Given the cost and difficulty of video annotations, even for a small amount of downstream data, such foundational models are emerging as the de-facto backbone for zero-shot [119, 122, 127] and few-shot generalization [2]. However, it remains unclear if these video-language models capture essential properties of a video beyond what can be learned from static images, most notably: time." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Many before us have shown that existing video-language" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.291, + 0.528, + 0.304 + ], + "angle": 0, + "content": "1." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.27, + 0.614, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.271, + 0.702, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.271, + 0.79, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.275, + 0.878, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.343, + 0.529, + 0.355 + ], + "angle": 0, + "content": "2." + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.324, + 0.614, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.324, + 0.702, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.323, + 0.79, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.323, + 0.878, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.393, + 0.529, + 0.405 + ], + "angle": 0, + "content": "3." + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.374, + 0.613, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.374, + 0.702, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.374, + 0.79, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.374, + 0.878, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.444, + 0.528, + 0.455 + ], + "angle": 0, + "content": "4." + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.424, + 0.614, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.424, + 0.702, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.424, + 0.79, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.424, + 0.878, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.483, + 0.531, + 0.498 + ], + "angle": 0, + "content": "A." + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.481, + 0.664, + 0.504 + ], + "angle": 0, + "content": "Dog runs away before it brings a ball to the man" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.531, + 0.522 + ], + "angle": 0, + "content": "C." + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.508, + 0.676, + 0.53 + ], + "angle": 0, + "content": "The baby eats food after it looks into the camera" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.482, + 0.724, + 0.498 + ], + "angle": 0, + "content": "B." + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.482, + 0.871, + 0.505 + ], + "angle": 0, + "content": "The dog brings a a ball to the man before it runs away" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.514, + 0.724, + 0.525 + ], + "angle": 0, + "content": "D." + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.513, + 0.853, + 0.535 + ], + "angle": 0, + "content": "The baby looks into the camera after it eats food" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.543, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Figure 1. Can you match the correct video-text pairs? Understanding the time order of events across video and language is necessary to be able to solve this task. See footnote on next page for answers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.84 + ], + "angle": 0, + "content": "models [6, 57, 66, 119] can achieve impressive performance on several video benchmarks [22, 41, 120] without reliably encoding time [13, 56, 59]. For example, Buch et al. [13] show that a model that uses a single (carefully selected) frame often outperforms recent video-language models [57, 119] on standard video benchmarks such as MSR-VTT [120]. Lei et al. [56] report similar findings with a single-frame pretraining approach. These findings hint at a lack of time awareness in video models. However, it remains unclear if these findings are caused, indeed, by the lack of time in video models or whether the benchmarks themselves do not mandate time awareness. Furthermore, there is no clear definition of what it means for a model to be time aware. In this paper, we strive to shed light on all these factors of time awareness in video-language models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "As a first step, we consider a simple notion of understanding time, i.e., understanding temporal relations such as before and after [4]. Consider the task presented in Fig. 1. A time invariant model shall be able to associate (A) with (1)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "or (2) and (B) with (3) or (4) based on static frames alone. But to distinguish between (1) and (2), one needs to be able to understand time order and connect it across video and language1. Thus, the first question we ask in Section 3: do the representations learnt by foundational video-language models encode this sense of time? To reliably attribute lack of time awareness to models and not existing benchmarks, we design our own synthetic dataset to probe models for this sense of time. We create video-language pairs that show a sequence of two events. Then, we alter the order of events either in the text or the video and check if models can connect the order in video and language. We find that existing video-language models indeed struggle to associate the time order across video and language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.604 + ], + "angle": 0, + "content": "In light of these findings, the second question we ask in Section 4 is: can we adapt a video-language model, without expensive re-training from scratch, to instill this sense of time? Towards this, we take inspiration from literature on understanding time in natural language, where there has been much work on developing time aware language models [20, 36, 37, 130, 131]. Our objective is to instill time awareness in a video-language model without having to pretrain from scratch. To do that, we propose TACT: Temporal Adaptation by Consistent Time-ordering based on two key components: (i) we artificially create samples that provide temporal signal, for example, by flipping the order of events in the video or the text, and (ii) we introduce a modified contrastive loss to learn time order consistency based on these samples. Instead of training from scratch, we adapt an existing video-language model, VideoCLIP [61], using the paradigm of post-pretraining on a small amount of videotext data [66, 121]. We demonstrate the effectiveness of TACT in connecting the time order in video and language on four diverse real datasets in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.473, + 0.74 + ], + "angle": 0, + "content": "Finally, in line with the original motive of video-language models for zero-shot generalization, we evaluate in Section 6 our TACT-adapted model for three sets of tasks on six downstream datasets which require a varying degree of time awareness. On tasks that need higher time awareness, with the appropriate choice of adaptation dataset, TACT outperforms a strong baseline that is based on post-pretraining on canonical clip-text pairs without consideration of time-order." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.473, + 0.879 + ], + "angle": 0, + "content": "In summary, our contributions are: (i) We show that existing video-language models struggle to associate time order in video and language through controlled experiments on synthetic data and several evaluations on real datasets. (ii) Based on VideoCLIP [119], we propose TACT, a method for temporal adaptation using this time order consistency without having to pretrain from scratch. (iii) We demonstrate improved zero-shot generalizability of TACT-adapted models on tasks that require higher time awareness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.787, + 0.108 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.117, + 0.892, + 0.148 + ], + "angle": 0, + "content": "We briefly discuss recent advances in video-language models followed by their consideration of time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.895, + 0.546 + ], + "angle": 0, + "content": "Foundational video-language models. Large-scale datasets, self-supervision, and the advent of Transformers [107] have led to the emergence of powerful encoders for images [21,39,103], videos [5,11,24,104,117], language models [19,64,69,86] and even universal encoders [32,46]. These encoders form the basis for several vision-language foundational models. Popular image-language models such as CLIP [83] and ALIGN [49] are trained on massive datasets by using web images and alt-text. Similarly, videolanguage models are catching up and can be categorised into two broad directions: (i) adapting image-language models for videos [8, 23, 50, 51, 63, 66, 71, 110, 112, 121], and (ii) pure video-based models that are learned using large video-text datasets [3,7,27-29,31,58,62,65,67,68,95,119]. Recently, a new paradigm of post-pretraining has emerged where an existing image- or video-language model goes through another stage of self-supervised pretraining on a small amount of video data before it is evaluated on downstream tasks [66, 121]. This is promising as it circumvents the prohibitive cost of pretraining on large datasets from scratch. In [66], the post-pretraining uses time-invariant mean-pooling, while [121] strives to bridge the domain gap between image captions and video subtitles. In contrast, our proposed temporal adaptation involves post-pretraining of VideoCLIP [119] with a small amount of data that instills the model to learn the time-order of events in a video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Time in vision. Time separates videos from static images or an unordered set of frames. While modeling time remains a challenge, it also presents a natural source of supervision that has been exploited for self-supervised learning. For example, as a proxy signal by posing pretext tasks involving spatio-temporal jigsaw [1, 44, 53], video speed [10, 17, 48, 94, 111, 124], arrow of time [78, 80, 114], frame/clip ordering [25, 70, 90, 97, 118], video continuity [61], or tracking [45, 108, 113]. Several works have also used contrastive learning to obtain spatio-temporal representations by (i) contrasting temporally augmented versions of a clip [47, 77, 81], or (ii) encouraging consistency between local and global temporal contexts [9, 18, 85, 123]. Nevertheless, it remains unclear if the learnt representations actually encode time reliably. Time-aware features have also been explored for specific downstream tasks such as action recognition [30, 100, 101]. There has been some very recent work on evaluating self-supervised video representations [87, 99] on their temporal recognition ability instead of only relying on time as a guidance for training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In the same spirit, a related direction pursues evaluation and benchmarking of time awareness in video datasets [88], models [13, 14, 30, 56, 89, 125] or both [43, 92]. Huang et" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.332, + 0.902 + ], + "angle": 0, + "content": "1 Answers: (A)-(2), (B)-(1), (C)-(4), (D)-(3)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "2504" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "al. [43] measure the effect of motion on temporal action recognition to find that only a subset of classes in UCF-101 and Kinetics-400 require motion information. Ghodrati et al. [30] propose new tasks to evaluate temporal asymmetry, continuity and causality in video models. Our work derives inspiration from these but applies more generally to video-language models as language provides a basis for open-world generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.473, + 0.48 + ], + "angle": 0, + "content": "Time in language. Time has also been extensively studied in the natural language literature. Early works identified temporal structures in language such as temporal prepositions and quantifiers [4, 79]. More recent literature focuses on tasks such as extracting temporal relations [35, 72-74], as well as temporal reasoning [36, 37, 82, 130, 131]. For example, Han et al. [36, 37] and Zhou et al. [131] pretrain language models specifically to focus on understanding temporal relations such as before, after, during, etc. The emergence of large language models has also spurred an increased interest in developing benchmarks to test for time awareness in these models [20, 75, 76, 102, 106, 129]. For example, Ning et al. [75] propose a new benchmark of reading comprehension with questions involving before/after relations. Since temporal relations in language are grounded in the video, we draw inspiration from [36, 37, 131] and aim to instill time awareness in video-language models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.473, + 0.701 + ], + "angle": 0, + "content": "Time in video-language models appears implicitly in tasks like video-text alignment [38, 98] and temporal grounding [42, 60]. In this work, we focus on self-supervised video-language models that can generalize to a variety of tasks rather than models designed for a specific task, e.g., temporal grounding. Some recent works have shown the under-utilisation of time in classic video-text benchmarks such as MSR-VTT [120], YouCook [132], ActivityNet [22], and DiDeMo [41]. For example, [13, 56, 57] discover that on several benchmarks, using only one or few frames or clips achieves competitive performance. Adaptations of the popular CLIP architecture for videos (e.g., CLIP4Clip [66]) show that weighted pooling of frames already achieves impressive performance on retrieval benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "These raise some key questions: do existing videolanguage models truly understand time in the sense of correctly associating order of events in language and video? If not, can we adapt them to instill time awareness? Our work addresses these questions. There has been some work in using time-order across video and language as a source of self-supervision. Specifically, concurrent to our work, both Sun et al. [96] and Cao et al. [15] propose fine-grained temporal alignment between video and text as the pretraining objective. Different from these works, we consider the notion of time-order and we aim to adapt a given videolanguage model using post-pretraining, which circumvents the need for a new round of compute-intense pretraining." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.096, + 0.712, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.099, + 0.865, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.277 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed task to evaluate time-order consistency across synthetic video-language pairs having before/after relations. We also define a control task to check if the synthetic videos are considered out-of-distribution by the model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.3, + 0.866, + 0.318 + ], + "angle": 0, + "content": "3. Do Video-Language Models Sense Time?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Probing video-language models for temporal understanding is an open direction of research. In this work, we consider a specific sense of temporal understanding: consistency in the order of events in a video with the associated textual description. For example, consider a text description: A red circle appears before a yellow circle. This imposes an order constraint on the video stream to have the event red circle appears happen before the event yellow circle appears. Can existing video-language models connect time-order in text with that in video? To answer this, we design an experiment with synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Synthetic dataset. We construct simple videos that comprise of a pair of events such as the ones mentioned above. We generate \\( N = 180 \\) video-language pairs as a combination of \\( C = 6 \\) colors, \\( S = 3 \\) shapes, and \\( |\\tau| = 2 \\) temporal relations: before and after. The corresponding caption describes the order of events connected with a before/after temporal relation. We call this caption as an attractor since it is consistent with the time-ordering in the video. Likewise, we construct a distractor where we flip the order of event descriptions while retaining the temporal relation. An example pair is illustrated in Fig. 2 (left). Ideally, a time aware video-language model should be able to associate the video with the temporally consistent text, or vice versa. We refer to this task as time-order consistency check. To rule out the possibility that synthetic videos are out-of-distribution, we also perform the same experiment with canonical clips with a single event and the text describes that same event as shown in Fig. 2 (right). We refer to this as the control task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Choice of models. We consider recent video-language models, broadly categorized into three groups: (i) image-language models like CLIP [83] that are adapted to videos [23,66,128], (ii) pure video-language models trained on a contrastive learning objective [6,16,119], and (iii) pure video-language models trained on a masking objective [29]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Findings. We evaluate video-to-text and text-to-video retrieval on both time-order consistency and control tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2505" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.254 + ], + "angle": 0, + "content": "
ParadigmMethodVideo-to-TextText-to-Video
◎◎◎◎
Chance-50.050.050.050.0
Image-Language adapted to videoCLIP4Clip [66]49.451.150.049.4
CLIP2Video [23]100.047.897.852.3
CenterCLIP [128]91.746.197.251.1
Video-Language ContrastiveVideoCLIP [119]87.151.166.748.3
Frozen in Time [6]97.849.4100.050.6
VindLU [16]98.052.0100.051.1
Video-Language MaskingBridgeFormer [29]100.051.197.242.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Table 1. Results on synthetic control ( ⌒ ) and time-order consistency ( ⌒ ) task as described in Fig. 2. Existing video-language models show random performance on our time-order task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.47, + 0.439 + ], + "angle": 0, + "content": "From Tab. 1, we observe that while most video-language models perform well on the control task, all of them struggle and perform on par with random chance on the temporal task. This gap in performance deserves attention given the importance of time in videos. Note that while synthetic data allows for controlled experiments, we also expand this evaluation to real video datasets in the following section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.449, + 0.473 + ], + "angle": 0, + "content": "4. Adaptation by Consistent Time-Ordering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.618 + ], + "angle": 0, + "content": "We describe a post-pretraining recipe to instill a video-language model with a sense of time. We propose TACT: Temporal Adaptation by Consistency of Time-order, that is based on two key components: (i) we artificially create samples that provide temporal signals, e.g., by flipping the order of events; and (ii) we introduce a modified contrastive loss to learn temporal consistency based on these samples. We start by defining the notation and then describe the key components of our adaptation recipe." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Preliminaries. Let \\(\\mathcal{V}\\) be the space of video clips and \\(\\mathcal{T}\\) be the space of text clips. Consider two non-overlapping video clips \\(v_{i}, v_{j} \\in \\mathcal{V}\\). Let \\(\\zeta_{i}, \\zeta_{j} \\in \\mathcal{T}\\) be their respective captions. Let \\(\\tau\\) be a temporal relation, \\(\\tau \\in \\{\\text{before, after}\\}\\). Then, we denote a stitched and time-order consistent clip as \\((u_{ij}, t_{ij})\\), where \\(u_{ij} := [v_{i}; v_{j}]\\), \\(t_{ij} := [\\zeta_{i}; \\tau; \\zeta_{j}]\\), and \\([\\cdot; \\cdot]\\) denotes concatenation. Note that depending on \\(\\tau\\), the order of \\(v_{i}\\) and \\(v_{j}\\) may need to change in \\(u_{ij}\\). For brevity, we drop the subscripts and refer to the stitched pair as \\((u, t)\\) unless stated otherwise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Time-order reversal. The classic contrastive learning paradigm for video-language models aligns components of a video clip \\( v_{i} \\) with its text counterpart \\( \\zeta_{i} \\) and contrasts against other texts \\( \\zeta_{j} \\) that usually describe a completely different clip. This makes such models ignore the finer details of temporal understanding as it is easier to contrast the negatives by simply focusing on the objects or the scene. This is evident from simple bag-of-word-like methods that" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.089, + 0.867, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.369, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Figure 3. Overview of TACT. Along with the usual contrastive loss, where negatives come from other samples in the batch, we make use of time-order reversal within the same sample and cross samples to generate additional negatives for both video and text. We also extend the contrastive loss to time-order reversed video and text corresponding to reverse consistency \\(\\mathcal{L}_r\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.584 + ], + "angle": 0, + "content": "are shown to work well for contrastive learning, both on the visual (e.g., CLIP4Clip [66]) and textual (e.g., MILNCE [67]) modalities. We hypothesize that unless there are negatives in a contrastive setup that contain the same scenes and objects, models need not learn a sense of time. Thus, we present a simple strategy to generate negatives that force the learning process to focus on the temporal order." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.63 + ], + "angle": 0, + "content": "We define a time-order reversal function \\(\\mathbb{T}\\) that operates on the stitched video clip or text description and temporally swaps its constituents:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.638, + 0.891, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathbb {T} (u) = \\mathbb {T} ([ v _ {i}; v _ {j} ]) := [ v _ {j}; v _ {i} ], \\quad \\text {a n d} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.656, + 0.891, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathbb {T} (t) = \\mathbb {T} ([ \\zeta_ {i}; \\tau ; \\zeta_ {j} ]) := [ \\zeta_ {j}; \\tau ; \\zeta_ {i} ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.776 + ], + "angle": 0, + "content": "An illustration of \\(\\mathbb{T}\\) is shown in Fig. 3. Note that \\(\\mathbb{T}\\) does not reverse the actual video, i.e., time does not flow backwards, but only changes the order in which events happen in the stitched clips. Our objective is to train a model that is able to distinguish between the original pair \\((u,t)\\) and time-reversed versions \\((u,\\mathbb{T}(t))\\), and \\((\\mathbb{T}(u),t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Loss function. We assume access to an existing pre-trained video-language model with a visual encoder \\( f_{\\theta} \\) and text encoder \\( g_{\\phi} \\). We obtain the video encoding \\( \\mathbf{z}_u \\coloneqq f_{\\theta}(u) \\in \\mathbb{R}^d \\) and the text encoding \\( \\mathbf{z}_t \\coloneqq g_{\\phi}(t) \\in \\mathbb{R}^d \\). Our goal is to adapt \\( \\Theta = \\{\\theta, \\phi\\} \\) via post-pretraining such that the resulting model is time aware while maintaining its original performance on tasks such as retrieval. As we aim to use a small amount of data, we only update some parameters of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2506" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.356, + 0.106 + ], + "angle": 0, + "content": "the model \\((\\Theta)\\), such as the last few layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.183 + ], + "angle": 0, + "content": "We now introduce our recipe for temporal adaptation based on the InfoNCE loss [105] to learn time-order sensitive video-text correspondence. For a positive (or time-order consistent) video-text pair \\((u, t)\\), we first define a forward loss where the stitched pair is in its original time-order." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.187, + 0.468, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\sum_ {(u, t) \\in \\mathcal {B}} \\left(\\operatorname {T N C E} \\left(\\mathbf {z} _ {u}, \\mathbf {z} _ {t}\\right) + \\operatorname {T N C E} \\left(\\mathbf {z} _ {t}, \\mathbf {z} _ {u}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.226, + 0.468, + 0.256 + ], + "angle": 0, + "content": "where TNCE is the Noise Contrastive Estimation (NCE) loss for temporal adaptation, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.261, + 0.468, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T N C E} \\left(\\mathbf {z} _ {u}, \\mathbf {z} _ {t}\\right) := - \\log \\frac {\\exp \\left(\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {t}\\right)}{\\sum_ {t ^ {\\prime} \\in \\mathcal {B} _ {t}} \\exp \\left(\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {t ^ {\\prime}}\\right) + \\mathcal {C} ^ {\\text {t i m e}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.468, + 0.342 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is the batch of \\((u,t)\\) pairs and \\(\\mathcal{B}_t\\) specifically refers to other stitched text captions in the batch. \\(\\mathcal{C}^{\\mathrm{time}}\\) accumulates negatives defined using time-order reversal as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.347, + 0.468, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} ^ {\\text {t i m e}} = \\alpha_ {\\text {s a m e}} \\exp (\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {\\mathbb {T} (t)}) + \\alpha_ {\\text {c r o s s}} \\sum_ {t ^ {\\prime} \\in \\mathcal {B} _ {t} \\backslash \\{t \\}} \\exp (\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {\\mathbb {T} \\left(t ^ {\\prime}\\right)}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.382, + 0.468, + 0.563 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\mathrm{same}}\\) controls the effect of contrasting text from the same sample but with reversed text time-order, i.e., \\(\\mathbb{T}(t)\\), and \\(\\alpha_{\\mathrm{cross}}\\) encourages the model to contrast between reversed versions of other text captions, i.e., \\(\\mathbb{T}(t')\\). Note that when both \\(\\alpha_{\\mathrm{same}}\\) and \\(\\alpha_{\\mathrm{cross}}\\) are 0, we revert back to the standard NCE formulation, albeit on stitched pairs. While Eq. (4) corresponds to the video-text loss \\(\\mathrm{TNCE}(\\mathbf{z}_u,\\mathbf{z}_t)\\), the text-video loss \\(\\mathrm{TNCE}(\\mathbf{z}_t,\\mathbf{z}_u)\\) is defined symmetrically. Furthermore, we also apply a reverse loss \\(\\mathcal{L}_r\\) to bring time-order reversed versions of both the video and the text together. Note that as we consider \\((u,t)\\) as a positive pair, \\((\\mathbb{T}(u),\\mathbb{T}(t))\\) also form a positive pair," + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.569, + 0.472, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\sum_ {\\left(\\mathbb {T} (u), \\mathbb {T} (t)\\right) \\in \\mathcal {B}} \\left(\\operatorname {T N C E} \\left(\\mathbf {z} _ {\\mathbb {T} (u)}, \\mathbf {z} _ {\\mathbb {T} (t)}\\right) + \\operatorname {T N C E} \\left(\\mathbf {z} _ {\\mathbb {T} (t)}, \\mathbf {z} _ {\\mathbb {T} (u)}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.616, + 0.468, + 0.661 + ], + "angle": 0, + "content": "Here, the TNCE term operates on time-reversed clips and \\(\\mathcal{C}^{\\mathrm{time}}\\) contrasts \\((\\mathbb{T}(u),\\mathbb{T}(t))\\) with un-reversed text clips in the batch \\((\\mathbb{T}(u),t)\\). The overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.668, + 0.468, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {f} + \\beta \\mathcal {L} _ {r}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.69, + 0.468, + 0.779 + ], + "angle": 0, + "content": "Depending on the choice of loss coefficients \\(\\alpha_{\\mathrm{same}}, \\alpha_{\\mathrm{cross}}, \\beta \\in \\{0,1\\}\\), we can vary properties of the adapted model. For example, setting \\(\\alpha_{\\mathrm{same}} = 1\\) encourages high sensitivity to time-order reversal. As we will see empirically, the loss coefficients also provide the flexibility to adapt the model to various datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We illustrate this temporal extension of the contrastive loss in Fig. 3 (best seen in colour). \\(\\mathbb{T}\\) illustrates the time order reversal function. The top half corresponds to \\(\\mathcal{L}_f\\) while the bottom half visualizes \\(\\mathcal{L}_r\\). In particular, the top-left quadrant alone corresponds to the standard contrastive loss on stitched pairs. While the green diagonal terms are positive pairs, the red diagonal terms are the strongest drivers for instilling temporal understanding in the model." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.196 + ], + "angle": 0, + "content": "
DatasetTrainValidationTestEgoLength
NvNcNvNcNvNc(s)
TEMPO3,90428,4274111,0003961,000X30
ActivityNet7,44095,474453906456912X120
Charades5,26299,9285001,0005001,000X30
Charades-Ego2,679155,3065001,000210420✓31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.201, + 0.892, + 0.257 + ], + "angle": 0, + "content": "Table 2. Statistics of datasets we consider for temporal adaptation. \\( N_{\\mathcal{V}} \\) is the number of unique videos and \\( N_{\\mathcal{C}} \\) is the number of stitched clips. Based on \\( N_{\\mathcal{V}} \\), TEMPO and Charades-Ego are smaller as compared to ActivityNet and Charades." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.283, + 0.78, + 0.3 + ], + "angle": 0, + "content": "5. Experiments: TACT Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.419 + ], + "angle": 0, + "content": "Base model. We demonstrate the effectiveness of TACT as an adaptation recipe on top of VideoCLIP [119] owing to its simple architecture, contrastive objective, and use of pre-computed S3D [116] features that enable faster experimentation and allow encoding a long temporal context (\\(\\sim 32\\) secs). We initialize \\(\\Theta\\) from the model pretrained on HowTo100M [68] and post-pretrain on multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Datasets. One of our key objectives is to post-pretrain on a small amount of data in contrast to massive pretraining datasets such as WebVid2M [7] or HowTo100M [68]. We consider dense video captioning datasets that offer sufficient diversity in terms of size, backgrounds, clip durations, viewpoints and activities. Specifically, we experiment with: (i) TEMPO [40]: the subset of stitched diverse third-person videos from DiDeMo [41] with text descriptions for fixed 5s segments that contain before/after relations; (ii) ActivityNet Captions [55]: a dense video captioning dataset with human-centric actions; (iii) Charades [93]: a scripted indoor daily human activities video dataset; and (iv) Charades-Ego [91]: similar to Charades, scripted human activities from the egocentric viewpoint. To construct stitched clips, we randomly sample any two non-overlapping clip-text pairs in the video. Since we require stitched clips instead of raw videos, we create new splits for each dataset (see Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.702, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate the post-pretrained model using two sets of metrics: (i) standard retrieval metrics, recall \\( R@1 \\), \\( R@5 \\), \\( R@10 \\) and median-rank evaluated on stitched video-text clips; and (ii) time-order consistency, i.e., the fraction of videos for which the model correctly associates text that is time order consistent with the video:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.801, + 0.891, + 0.838 + ], + "angle": 0, + "content": "\\[\nA _ {\\text {t i m e}} := \\frac {1}{| \\mathcal {D} |} \\sum_ {(u, t) \\in \\mathcal {D}} \\mathbb {I} [ d (\\mathbf {z} _ {u}, \\mathbf {z} _ {t}) < d (\\mathbf {z} _ {u}, \\mathbf {z} _ {\\mathbb {T} (t)}) ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.851, + 0.89, + 0.881 + ], + "angle": 0, + "content": "where \\((u,t)\\) are time-order consistent pairs, \\(\\mathcal{D}\\) is the dataset, and \\(d(\\cdot ,\\cdot)\\) is a distance metric based on cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Post-pretraining details. We freeze the word embeddings" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2507" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "and layers 1 to 5 for both the video and text encoders in VideoCLIP. For adaptation, we use the Adam optimizer [54] with learning rate \\(5e^{-6}\\), batch size 32 trained on a single node with 4 GeForce GTX 1080 GPUs. On TEMPO, we train for 60 epochs while on the other datasets, we train for 10 epochs and pick the checkpoint that maximizes the geometric mean of \\(R@1\\) and \\(A_{\\mathrm{time}}\\) on the respective validation set. A typical adaptation run takes about 1-3 hours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.22, + 0.47, + 0.312 + ], + "angle": 0, + "content": "Evaluation on the test set. Results in Tab. 3 show that TACT\\* with optimal loss coefficients outperforms TACT† (all 0 loss coefficients) and the zero-shot baseline (no postpretraining similar to the synthetic data experiment), both on the retrieval and time-order consistency tasks. This indicates the robustness of the adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.546 + ], + "angle": 0, + "content": "Impact of loss coefficients. Choosing appropriate values for loss coefficients \\(\\Theta_l \\coloneqq \\{\\alpha_{\\mathrm{same}}, \\alpha_{\\mathrm{cross}}, \\beta\\}\\) allows the model to learn various aspects and adapt using different datasets. On each dataset, we vary \\(\\Theta_l \\in \\{0,1\\}^3\\) and find the best configuration based on the GeometricMean \\((R@1, \\max(A_{\\mathrm{time}} - 50, 0))\\) on the validation sets. The above metric ensures the geometric mean is not overpowered by \\(R@1\\). The results are shown in Tab. 4. As \\(\\alpha_{\\mathrm{same}}\\) is directly responsible for discriminating between original and time-reversed orders, unsurprisingly, setting it to 1 is necessary to achieve the best results on \\(A_{\\mathrm{time}}\\) for all the datasets. For TEMPO and Charades-Ego, using all loss components (all 1) provides the best results, whereas \\(\\alpha_{\\mathrm{cross}} = 1\\) and \\(\\beta = 0\\) achieves a better trade-off for ActivityNet and Charades. Choosing \\(\\beta = 1\\) leads to an improve" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.571, + 0.457, + 0.808 + ], + "angle": 0, + "content": "
DatasetMethodRetrievalTime-order
R@1↑MedR ↓Atime ↑
TEMPOZero-shot3.749.048.1
TACT†7.713.046.5
TACT*9.39.066.5
ActivityNetZero-shot1.144.049.6
TACT†5.834.059.7
TACT*5.835.085.7
CharadesZero-shot1.3170.047.1
TACT†5.338.573.5
TACT*5.735.077.0
Charades-EgoZero-shot1.664.053.7
TACT†6.435.060.1
TACT*10.128.568.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Table 3. Results for the best TACT model on test sets. \\(\\mathrm{TACT}^{\\star}\\) has optimal loss coefficients and \\(\\mathrm{TACT}^{\\dagger}\\) is a baseline with all coefficients 0. On time order, TACT generalizes well with \\(\\mathrm{TACT}^{\\star}\\) outperforming the baselines. On retrieval, for TEMPO and Charades-Ego, \\(\\mathrm{TACT}^{\\star}\\) outperforms the baseline as their optimal models have \\(\\beta = 1\\) which helps retrieval with a small amount of data." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.888, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.176, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Figure 4. Time-distance between stitched clips in datasets for temporal adaptation \\((\\Delta_{\\mathrm{time}})\\). TEMPO has stitched clips close to each other while those in Charades-Ego are farthest apart suggesting a correlation between \\(\\Delta_{\\mathrm{time}}\\) and difficulty of temporal adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.412 + ], + "angle": 0, + "content": "ment in retrieval performance for TEMPO and Charades-Ego but leads to a decline for ActivityNet and Charades. We attribute this to the number of unique videos in the train set for these datasets. As ActivityNet and Charades have more videos than TEMPO or Charades-Ego (see train \\( N_{\\mathcal{V}} \\) Tab. 2) additional positives introduced by setting \\( \\beta = 1 \\) are not necessary and in fact, hurt performance. Finally, we note that carefully setting the value of \\( \\alpha_{\\mathrm{same}} \\) provides a convenient trade-off between spatial and temporal understanding. Please refer to the supplement for detailed experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.675 + ], + "angle": 0, + "content": "What makes temporal adaptation hard? We observe a large gap in \\( A_{\\mathrm{time}} \\) between TEMPO and ActivityNet. We hypothesize that the distance (in seconds) between the two clips (\\( \\Delta_{\\mathrm{time}} \\)) in a stitched clip is strongly correlated with the difficulty of adaptation. Intuitively, it is easier to infer the time-order consistency for a stitched clip \\( u \\) with text \\( t \\) that has distant constituent clips \\( v_i, v_j \\) since the objects and scene could be vastly different. In contrast, it is harder to discern the correct time-order when the constituent clips are closer in time. Fig. 4 shows distribution of \\( \\Delta_{\\mathrm{time}} \\) for each dataset. Indeed, \\( \\Delta_{\\mathrm{time}} \\) in ActivityNet (58.8s) is much higher than that in TEMPO (6.4s) making the task harder on TEMPO. To further test our hypothesis, we conduct a controlled experiment where we gradually vary the distribution of \\( \\Delta_{\\mathrm{time}} \\) for Charades-Ego to match it to that of TEMPO. We find a strong correlation (\\( \\rho = 0.92 \\)) between \\( \\Delta_{\\mathrm{time}} \\) and hardness of adaptation. More details are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.847, + 0.705 + ], + "angle": 0, + "content": "6. Experiments: Downstream Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.774 + ], + "angle": 0, + "content": "The goal of video-language foundation models is to generalize in a zero- or few-shot manner to a diverse range of downstream tasks. We evaluate TACT models on three sets of downstream tasks that need low-to-high time awareness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Baseline: Standard post-pretraining. Comparing our temporally adapted models with pretrained VideoCLIP is not fair since adapted models see data beyond the pretraining phase. In addition to the zero-shot comparison, we compare against a baseline model that is trained for standard video-text retrieval on the same datasets as temporal adaptation. Instead of using stitched clips, we use simple canonical pairs, i.e., \\((v_{i},\\zeta_{i})\\) instead of \\((u_{ij},t_{ij})\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2508" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.27 + ], + "angle": 0, + "content": "
Loss coefficientsTEMPOActivityNetCharadesCharades-Ego
α_sameα_crossβR@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑
Chance0.1500.050.00.1453.050.00.1500.050.00.1500.050.0
0008.314.049.46.430.057.35.742.071.52.944.064.6
0018.214.049.55.627.047.04.258.075.13.241.565.2
0108.215.049.36.129.078.85.245.078.93.438.064.5
0118.114.049.55.827.048.34.258.075.13.141.067.0
1006.420.060.65.928.079.16.138.076.33.242.066.1
1016.524.062.95.626.063.14.951.078.03.339.070.7
1105.924.059.76.029.086.36.643.077.83.740.567.9
1117.517.062.55.727.063.85.151.077.73.838.568.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.276, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Table 4. Impact of loss coefficients for TACT post-pretraining on validation sets of various datasets. For each dataset, the corresponding color-marked row denotes the best configuration based on the geometric mean of \\( R@1 \\) and \\( A_{\\mathrm{time}} \\). TACT is able to connect time-order in video and language while maintaining its retrieval capabilities across several datasets." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.33, + 0.892, + 0.572 + ], + "angle": 0, + "content": "
AdaptationI. Text-to-Video RetrievalII. Temporal VQAIII. Action-to-Video Retrieval
DatasetMethodMSR-VTTYouCookIINext-QA (ATP)AGQASSv2SSv2 (events)Temporal
R@1R@5R@10R@1R@5R@10AccuracyAccuracymAPmAPmAP
-Chance0.10.51.00.10.51.020.050.00.62.02.0
-None10.623.429.918.245.559.923.449.93.46.413.0
TEMPOBaseline12.029.337.321.548.261.825.050.83.96.815.9
TACT12.826.535.720.445.158.727.657.14.27.716.1
ActivityNetBaseline15.734.444.915.638.851.423.750.73.77.016.0
TACT13.829.639.616.036.949.825.450.33.57.216.2
CharadesBaseline12.325.833.621.548.661.726.050.54.17.113.7
TACT11.725.232.422.449.162.425.254.84.37.814.1
Charades-EgoBaseline13.127.534.519.447.160.824.349.74.06.914.7
TACT11.124.630.621.948.261.925.658.43.96.913.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.577, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Table 5. Results on downstream zero-shot evaluation with tasks requiring increasing time awareness from I to III. None corresponds to direct evaluation of the VideoCLIP model on the downstream dataset. Green denotes an improvement for the TACT adapted model w.r.t. the baseline, red denotes a deterioration. As we move from tasks that need low to high time awareness, the effectiveness of TACT increases. See Sec. 6 for a more detailed discussion. The table is best viewed on screen in colour." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.766 + ], + "angle": 0, + "content": "Evaluating TACT adapted models on synthetic data. On the video-to-text variant, TACT adapted on TEMPO achieves \\(64.4\\%\\), ActivityNet \\(52.5\\%\\), Charades \\(65.0\\%\\), Charades-Ego \\(85.6\\%\\). This is usually higher than the performance that non-adapted models achieve in Tab. 1. This highlights that TACT models learn useful signals to match time-order in video and language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.902 + ], + "angle": 0, + "content": "I. Text-to-video retrieval. We consider two widely used benchmarks: MSR-VTT [120] and YouCookII [132] and adopt standard retrieval metrics. Recent work has identified a bias for spatial understanding in these datasets [8, 13, 43, 56, 59, 66]. Thus, we consider this class of tasks as requiring low time awareness. As shown in Tab. 5 set I, on MSR-VTT [120], we observe that TACT is worse (marked in red) or at par with the baseline across datasets. This" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.735 + ], + "angle": 0, + "content": "aligns well with findings in [13, 56] that these benchmarks do not need time awareness. On YouCookII [132], TACT models based on Charades(-Ego) outperform the baseline (marked in green). We believe this is a consequence of a lower domain shift between YouCookII and Charades." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.895, + 0.879 + ], + "angle": 0, + "content": "II. Temporal video QA. Next, we use subsets of recently released multiple-choice video question answering benchmarks: Next-QA [115] and AGQA [34]. The idea is to check if we can probe models for temporal understanding by asking questions with temporal language. Buch et al. [13] identify a subset of Next-QA, dubbed as ATP-hard², with a higher concentration of temporally challenging data. For AGQA, we pick a subset of \\(\\sim 6k\\) questions that explicitly have a question with before/after relation. We consider" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.874, + 0.902 + ], + "angle": 0, + "content": "2Available here: github.com/StanfordVL/atp-video-language" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2509" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "these benchmarks as requiring a moderate-high level of time awareness and AGQA in particular is also close to our adaptation task. We use accuracy as the standard metric. We observe (see Tab. 5 set II) that indeed TACT almost always outperforms (marked in green) baselines on both Next-QA and AGQA. TEMPO-adapted TACT seems to generalize particularly well on both benchmarks. Likewise, Charades-adapted TACT does well on AGQA since AGQA is also based on the Charades videos accounting for reduced domain shift. We affirm that temporal adaptation is useful, especially when the downstream tasks require it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.779 + ], + "angle": 0, + "content": "III. Action-to-video retrieval. Finally, we consider action recognition benchmarks such as Something-Something (SSv2) [33] and Temporal [88]. SSv2 was designed to capture richer temporal information [33, 56]. We follow Lei et al. [56], who propose the template-retrieval task that encourages temporal modelling and use their evaluation split3 containing \\( C = 174 \\) actions and \\( K = 12 \\) videos per class. Interestingly, different actions in SSv2 require differing levels of time awareness. We create a subset SSv2 (events) with \\( C_{\\text{events}} = 49 \\) actions that have at least two verbs in the label as the occurrence of multiple verbs is indicative of multiple events occurring in sequence. Finally, we also evaluate against the Temporal benchmark [88], a combination of 50 action classes from SSv2 [33] and Kinetics-400 [52] for which temporal information is deemed to be essential for recognition. Similar to text-to-video retrieval, we use the action class as a text query and obtain a ranking over all videos. Different from the retrieval setup, since a single query has multiple correct answers (up to \\( K = 12 \\) videos), we report mAP as the metric for these benchmarks. This task set needs high time awareness. Furthermore, unlike QA tasks in II, there is a shift in several (uncontrolled) factors as we move from temporal adaptation task to these tasks. From Tab. 5, we observe that TEMPO and Charades-adapted models generalize well across set III benchmarks. ActivityNet-adapted TACT underperforms on SSv2 but outperforms the baseline on strongly temporal actions in SSv2 (events) and Temporal. Finally, TACT adapted on Charades-Ego is at-par or slightly worse than the baseline on SSv2 variants, and also on Temporal, perhaps due to the shift from egocentric to third-person videos. Overall, despite SSv2 and Temporal requiring high time awareness, TACT models show promising zero-shot generalization with the right choice of the adaptation dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.789, + 0.327, + 0.805 + ], + "angle": 0, + "content": "7. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.819, + 0.472, + 0.88 + ], + "angle": 0, + "content": "Generalization to other temporal prompts. The time-order of events in language can be described using various sentence structures. Although we train video-language models using before/after relations, it is natural to ask if" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.098, + 0.864, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Figure 5. Models trained by TACT with before/after relations generalize to a new kind of prompt such as First, ..., then ... indicating learning of the underlying true time-order of events." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.462 + ], + "angle": 0, + "content": "the model still correctly associates time-order for a different prompt such as First, ..., then, ... To systematically test this, we gather event pairs \\( E_1 \\), \\( E_2 \\) (\\( E_1 \\) occurs before \\( E_2 \\) in the video) for each sample in the validation set and stitch them using three prompts as follows: (i) \\( E_1 \\) before \\( E_2 \\), (ii) \\( E_2 \\) after \\( E_1 \\), (iii) First \\( E_1 \\), then \\( E_2 \\). As shown in Fig. 5, TACT-adapted models generalize well to a new prompt (iii). This substantiates the learning of time-order of events rather than merely learning the order of words in the sentence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.895, + 0.649 + ], + "angle": 0, + "content": "Limitations. While we present a promising way of instilling time in video-language models, our work is limited to the VideoCLIP [119] pretrained model. Our initial experiments with Frozen in Time [6] were not as promising, perhaps because it uses a much shorter temporal context (4 frames). Please see the supplement for results on more pretrained models. Furthermore, we consider a specific definition of time awareness derived from temporal relations like before/after. It is natural to ask if this can be extended to more general notions of temporality, e.g., as defined by Allen [4]. Finally, there can always be more downstream tasks considered such as (spatio-)temporal localization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.896, + 0.837 + ], + "angle": 0, + "content": "Conclusion. Given the essence of time in video-language models, we present a simple experiment based on synthetic data to test for time awareness. We find that existing models lack a sense of time defined in terms of consistency of order of events in video and language. To fill this gap, building upon VideoCLIP [119], we present TACT, a recipe to instill this sense of time in video-language models. Finally, we analyze the zero-shot generalizability of TACT-adapted models to a diverse set of tasks. We hope that this work provokes further probing and instilling time awareness in video-language models, and also inspires other adaptations of foundational models to solve various challenging tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. We acknowledge support from ELLIS Amsterdam and the AMS Scholarship to Piyush. We thank Dr Dennis Koelma for their help with compute infrastructure and Dr Hazel Doughty for useful discussions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.393, + 0.901 + ], + "angle": 0, + "content": "3Available here: github.com/jayleicn/singularity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2510" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Winter Conference on Applications of Computer Vision (WACV), pages 179-189. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.187, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.328, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. ArXiv, abs/2006.16228, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.399, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[4] James F. Allen. Towards a general theory of action and time. In Artificial Intelligence, 1984. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.429, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.486, + 0.47, + 0.541 + ], + "angle": 0, + "content": "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.543, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.601, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.644, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[9] Nadine Behrmann, Mohsen Fayyaz, Juergen Gall, and Mehdi Noroozi. Long short view feature decomposition via contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9224-9233, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.716, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[10] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9922–9931, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.786, + 0.47, + 0.841 + ], + "angle": 0, + "content": "[11] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.843, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Rishi Bommasani et al. On the opportunities and risks of foundation models. ArXiv, abs/2108.07258, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Shyamal Buch, Cristobal Eyzaguirre, Adrien Gaidon, Jia-jun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "the \"Video\" in Video-Language Understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[14] Petr Byvshev, Pascal Mettes, and Yu Xiao. Are 3d convolutional networks inherently biased towards appearance? Computer Vision and Image Understanding, 220:103437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.193, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[15] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.25, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[16] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.307, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[17] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised spatio-temporal representation learning using variable playback speed prediction. IEEE Access, 9:79562-79571, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.364, + 0.893, + 0.419 + ], + "angle": 0, + "content": "[18] Ishan Rajendra Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TcIr: Temporal contrastive learning for video representation. Computer Vision and Image Understanding (CVIU), 219:103406, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.421, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.519, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[20] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.59, + 0.893, + 0.686 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.688, + 0.893, + 0.758 + ], + "angle": 0, + "content": "[22] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.76, + 0.893, + 0.801 + ], + "angle": 0, + "content": "[23] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.803, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[24] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[25] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2511" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition (CVPR), pages 3636-3645, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[26] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[27] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.236, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[28] Valentin Gabeur, Chen Sun, Alahari Karteek, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.292, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[29] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 16167-16176, June 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.363, + 0.468, + 0.404 + ], + "angle": 0, + "content": "[30] Amir Ghodrati, Efstratios Gavves, and Cees G. M. Snoek. Video time: Properties, encoders and evaluation. In British Machine Vision Conference (BMVC), 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.407, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[31] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. ArXiv, abs/2011.00597, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.463, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[32] Rohit Girdhar, Mannat Singh, Nikhil Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. Conference on Computer Vision and Pattern Recognition (CVPR), pages 16081-16091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.534, + 0.468, + 0.63 + ], + "angle": 0, + "content": "[33] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.633, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[34] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.468, + 0.744 + ], + "angle": 0, + "content": "[35] Rujun Han, Qiang Ning, and Nanyun Peng. Joint event and temporal relation extraction with shared representations and structured prediction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[36] Rujun Han, Xiang Ren, and Nanyun Peng. Deer: A data efficient language model for event temporal reasoning. ArXiv, abs/2012.15283, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.789, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[37] Rujun Han, Xiang Ren, and Nanyun Peng. Econet: Effective continual pretraining of language models for event temporal reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[38] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2906, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In _Conference on Computer Vision and Pattern Recognition (CVPR), CVPR '16, pages 770–778, June 2016._2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[40] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), pages 1380-1390, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.219, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[41] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. International Conference on Computer Vision (ICCV), pages 5804-5813, 2017. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.288, + 0.892, + 0.354 + ], + "angle": 0, + "content": "[42] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.357, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[43] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7366–7375, 2018. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.455, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[44] Yuqi Huo, Mingyu Ding, Haoyu Lu, Zhiwu Lu, Tao Xiang, Ji-Rong Wen, Ziyuan Huang, Jianwen Jiang, Shiwei Zhang, Mingqian Tang, Songfang Huang, and Ping Luo. Self-supervised video representation learning with constrained spatiotemporal jigsaw. In International Joint Conference on Artificial Intelligence (IJCAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[45] A. Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. ArXiv, abs/2006.14613, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[46] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver: General perception with iterative attention. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[47] S. Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9950-9960, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[48] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), pages 425-442, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.735, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[49] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[50] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2512" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.162, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[53] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Association for the Advancement of Artificial Intelligence (AAAI), pages 8545-8552, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[54] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[55] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-Captioning Events in Videos. In International Conference on Computer Vision (ICCV), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[56] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[57] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. Conference on Computer Vision and Pattern Recognition (CVPR), pages 7327-7337, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.428, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[58] Gen Li, Nan Duan, Yuejian Fang, Daxin Jiang, and Ming Zhou. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Association for the Advancement of Artificial Intelligence (AAAI), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.484, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[59] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.553, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[60] Juncheng Li, Junlin Xie, Long Qian, Linchao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal grounding with structured variational cross-graph correspondence learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 3022-3031, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.636, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[61] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. ArXiv, abs/2112.05883, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[62] Kevin Lin, Alex Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z. Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Ego-centric video-language pretraining. ArXiv, abs/2206.01670, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.775, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[63] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Y. Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. ArXiv, abs/2208.03550, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.831, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[64] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[65] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.162, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[66] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.218, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[67] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. Conference on Computer Vision and Pattern Recognition (CVPR), pages 9876-9886, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.288, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[68] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In International Conference on Computer Vision (ICCV), 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.371, + 0.894, + 0.427 + ], + "angle": 0, + "content": "[69] Tomas Mikolov, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In International Conference on Learning Representations (ICLR), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.427, + 0.894, + 0.483 + ], + "angle": 0, + "content": "[70] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), pages 527-544, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.483, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[71] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.552, + 0.894, + 0.607 + ], + "angle": 0, + "content": "[72] Qiang Ning, Zhili Feng, and Dan Roth. A structured learning approach to temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.608, + 0.894, + 0.65 + ], + "angle": 0, + "content": "[73] Qiang Ning, Zhili Feng, Hao Wu, and Dan Roth. Joint reasoning for temporal and causal relations. In Association of Computational Linguistics (ACL), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.65, + 0.894, + 0.705 + ], + "angle": 0, + "content": "[74] Qiang Ning, Sanjay Subramanian, and Dan Roth. An improved neural baseline for temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.705, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[75] Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. Torque: A reading comprehension dataset of temporal ordering questions. ArXiv, abs/2005.00242, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.761, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[76] Qiang Ning, Ben Zhou, Zhili Feng, Haoruo Peng, and Dan Roth. Cogcomptime: A tool for understanding time in natural language. ArXiv, abs/1906.04940, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.803, + 0.894, + 0.873 + ], + "angle": 0, + "content": "[77] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. Conference on Computer Vision and Pattern Recognition (CVPR), pages 11200-11209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[78] Lyndsey C. Pickup, Zheng Pan, Donglai Wei, YiChang Shih, Changshui Zhang, Andrew Zisserman, Bernhard" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2513" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Scholkopf, and William T. Freeman. Seeing the arrow of time. In *Conference on Computer Vision and Pattern Recognition* (CVPR), pages 2043-2050, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.136, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[79] Ianthe Pratt and Nissim Francez. Temporal prepositions and temporal generalized quantifiers. Linguistics and Philosophy, 24:187-222, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[80] Will Price and Dima Damen. Retro-actions: Learning 'close' by time-reversing 'open' videos. International Conference on Computer Vision Workshops (ICCVW), pages 1371-1380, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.236, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[81] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, H. Wang, Serge J. Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6960-6970, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.307, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[82] Lianhui Qin, Aditya Gupta, Shyam Upadhyay, Luheng He, Yejin Choi, and Manaal Faruqui. Timedial: Temporal commonsense reasoning in dialog. In Association of Computational Linguistics (ACL), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.363, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.448, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[84] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.505, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[85] Adrià Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Patraucean, Florent Altch'e, Michael Valko, Jean-Bastien Grill, Aäron van den Oord, and Andrew Zisserman. Broaden your views for self-supervised video learning. International Conference on Computer Vision (ICCV), pages 1235–1245, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.603, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[86] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.647, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[87] Madeline Chantry Schiappa, Yogesh Singh Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ArXiv, abs/2207.00419, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[88] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. CoRR, abs/1907.08340, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[89] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? International Conference on Computer Vision (ICCV), pages 9641-9649, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[90] Vivek Sharma, Makarand Tapaswi, and Rainer Stiefelhagen. Deep Multimodal Feature Encoding for Video Ordering. In International Conference on Computer Vision Workshops (ICCVW), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[91] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "modeling of first and third-person videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7396-7404, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[92] Gunnar A. Sigurdsson, Olga Russakovsky, and Abhinav Kumar Gupta. What actions are needed for understanding human actions in videos? 2017 IEEE International Conference on Computer Vision (ICCV), pages 2156-2165, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.209, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[93] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, Cham, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.281, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[94] Ankit Singh, Omprakash Chakraborty, Ashutosh Varshney, Rameswar Panda, Rogério Schmidt Feris, Kate Saenko, and Abir Das. Semi-supervised action recognition with temporal contrastive learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 10384-10394, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.367, + 0.892, + 0.434 + ], + "angle": 0, + "content": "[95] Chen Sun, Austin Myers, Carl Vondrick, Kevin P. Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. International Conference on Computer Vision (ICCV), pages 7463-7472, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.439, + 0.892, + 0.507 + ], + "angle": 0, + "content": "[96] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[97] Tomoyuki Suzuki, Takahiro Itazuri, Kensho Hara, and Hirokatsu Kataoka. Learning spatiotemporal 3d convolution with video order self-supervision. In European Conference on Computer Vision Workshops (ECCV), pages 0–0, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.582, + 0.892, + 0.623 + ], + "angle": 0, + "content": "[98] Makarand Tapaswi, Martin Bauml, and Rainer Stiefelhagen. Book2movie: Aligning video scenes with book chapters. In CVPR, June 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.627, + 0.892, + 0.682 + ], + "angle": 0, + "content": "[99] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees G. M. Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[100] Fida Mohammad Thoker, Hazel Doughty, and Cees Snoek. Tubelet-contrastive self-supervision for video-efficient generalization, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[101] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Skeleton-contrastive 3d action representation learning. ACM MM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[102] Shivin Thukral, Kunal Kukreja, and Christian Kavouras. Probing language models for understanding of temporal expressions. In BLACKBOXNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[103] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. arXiv preprint arXiv:2105.01601, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2514" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[104] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6450-6459, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[105] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. ArXiv, abs/1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[106] Siddharth Vashishtha, Adam Poliak, Yash Kumar Lal, Benjamin Van Durme, and Aaron Steven White. Temporal reasoning in natural language inference. In FINDINGS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[107] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[108] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2563-2572, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[109] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[110] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[111] Jiangliu Wang, Jianbo Jiao, and Yun-Hui Liu. Self-supervised video representation learning by pace prediction. In European Conference on Computer Vision (ECCV), pages 504-521, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[112] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[113] X. Wang, A. Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2561-2571, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[114] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 8052-8060, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[115] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[116] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In European Conference on Computer Vision (ECCV), pages 318–335, Cham, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[117] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin P. Murphy. Rethinking spatiotemporal feature learning for video understanding. ArXiv, abs/1712.04851, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[118] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10334-10343, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.304 + ], + "angle": 0, + "content": "[119] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing (EMNLP), pages 6787-6800, 2021. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[120] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.893, + 0.418 + ], + "angle": 0, + "content": "[121] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Rui Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. ArXiv, abs/2209.06430, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.893, + 0.474 + ], + "angle": 0, + "content": "[122] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.893, + 0.532 + ], + "angle": 0, + "content": "[123] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Back to the future: Cycle encoding prediction for self-supervised contrastive video representation learning. ArXiv, abs/2010.07217, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.893, + 0.602 + ], + "angle": 0, + "content": "[124] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6548-6557, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.66 + ], + "angle": 0, + "content": "[125] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. Clevrer: Collision events for video representation and reasoning. ArXiv, abs/1910.01442, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[126] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel C. F. Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. ArXiv, abs/2111.11432, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[127] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[128] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In SIGIR, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2515" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[129] Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. \"going on a vacation\" takes longer than \"going for a walk\": A study of temporal commonsense understanding. *ArXiv*, abs/1909.03065, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[130] Ben Zhou, Qiang Ning, Daniel Khashabi, and Dan Roth. Temporal common sense acquisition with minimal supervision. ArXiv, abs/2005.04304, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[131] Ben Zhou, Kyle Richardson, Qiang Ning, Tushar Khot, Ashish Sabharwal, and Dan Roth. Temporal reasoning on implicit events from distant supervision. ArXiv, abs/2010.12753, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[132] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence (AAAI), pages 7590-7598, 2018. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2516" + } + ] +] \ No newline at end of file diff --git a/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_origin.pdf b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d4f86dc63d0828a7bb306cce2c83e7a2388c7450 --- /dev/null +++ b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/f3a89bc6-be03-43ea-932f-9df8dc46f9ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f813f4949ee11eb64b5d14edf57faf5edaafac79209e12d7cf251a4b9b8c1d2b +size 1786247 diff --git a/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/full.md b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ea779ca70b94e970ff572de384d14e0e68f905b --- /dev/null +++ b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/full.md @@ -0,0 +1,411 @@ +# Test of Time: Instilling Video-Language Models with a Sense of Time + +Piyush Bagad + +University of Amsterdam + +Makarand Tapaswi + +IIIT Hyderabad + +Cees G.M. Snoek + +University of Amsterdam + +bpiyush.github.io/testoftime-website + +# Abstract + +Modelling and understanding time remains a challenge in contemporary video understanding models. With language emerging as a key driver towards powerful generalization, it is imperative for foundational video-language models to have a sense of time. In this paper, we consider a specific aspect of temporal understanding: consistency of time order as elicited by before/after relations. We establish that seven existing video-language models struggle to understand even such simple temporal relations. We then question whether it is feasible to equip these foundational models with temporal awareness without re-training them from scratch. Towards this, we propose a temporal adaptation recipe on top of one such model, VideoCLIP, based on post-pretraining on a small amount of video-text data. We conduct a zero-shot evaluation of the adapted models on six datasets for three downstream tasks which require varying degrees of time awareness. We observe encouraging performance gains especially when the task needs higher time awareness. Our work serves as a first step towards probing and instilling a sense of time in existing video-language models without the need for data and compute-intensive training from scratch. + +# 1. Introduction + +Self-supervised pretraining at scale on multimodal web corpora tied with powerful architectures [107] has led to foundational models [12] for images [2, 49, 59, 83, 84] and videos [2, 6, 26, 109, 119, 126]. These models have enabled remarkable improvements on a plethora of downstream video-language tasks such as video-text retrieval, video question-answering, and action recognition. Given the cost and difficulty of video annotations, even for a small amount of downstream data, such foundational models are emerging as the de-facto backbone for zero-shot [119, 122, 127] and few-shot generalization [2]. However, it remains unclear if these video-language models capture essential properties of a video beyond what can be learned from static images, most notably: time. + +Many before us have shown that existing video-language + +![](images/0547fbbf6ef66282f5fcf4e972eacb39c83d3355c7d50989704ec08ef25296db.jpg) +1. + +![](images/5f93c9201089595b0e6f3dc220b735703fcb6f6f7abfd03c99e9340eb033c094.jpg) + +![](images/e2b8ca251bc4b75da58650cb76cfb9473621692e19b795ee3e859abd8df6b267.jpg) + +![](images/448db9de872f0dbabe74cb5e24df2aa24362cab688f77f3bbbb708efca7cfe3e.jpg) + +![](images/6b85dd3be0e0a4f8b3dfdb23fc5784ab19f471e1f361b8615af310f2c2f4936c.jpg) +2. + +![](images/ed74bac97aba135b1c3be4dd919b1069c3acd91772eb876ab5e20e007483f5c5.jpg) + +![](images/baba012e51a4f48e4bfb214edace9dc46cf7c58b5d4e4285d4b9bbf2224f1721.jpg) + +![](images/bea8a049582657baeb8d9eeffda3dc108efd817149214d60d920e479d1cf7e77.jpg) + +![](images/a1c0853e3f209fca343a837a022dbe64f9b4f2675c37e7ec83d1dcf51ad3ee7c.jpg) +3. + +![](images/7573eb9b7773f08653a5476096a0e2536069043db96368106b8de73408b349b2.jpg) + +![](images/c4b5f7e459b996314e387c808d1cf50f7c00ef6483813a8b8a8f477f987373ce.jpg) + +![](images/ac9aa9f03961a12e1b69245c1882606afd83c7a3fa18d17c3e0b15a127d7727b.jpg) + +![](images/47bf84d000abbcaf973fc77fa37a1151a1b671dabad57d3b2f4babe2bff5250c.jpg) +4. +A. +C. +The baby eats food after it looks into the camera +Figure 1. Can you match the correct video-text pairs? Understanding the time order of events across video and language is necessary to be able to solve this task. See footnote on next page for answers. + +![](images/e4eb3058bf121d59246e5f101f4b33c6ab5c75f13f1834e381acdf2a44daf585.jpg) +Dog runs away before it brings a ball to the man + +![](images/2f44f942b293c338a0cbded02f28ecc265b49c3e17a2ad80f9722c6b762ffbbe.jpg) +B. +The dog brings a a ball to the man before it runs away +D. +The baby looks into the camera after it eats food + +![](images/ccc38d16e1c011d3ebf940f0488db10055b7d95938d7918501954b9fedb1a695.jpg) + +models [6, 57, 66, 119] can achieve impressive performance on several video benchmarks [22, 41, 120] without reliably encoding time [13, 56, 59]. For example, Buch et al. [13] show that a model that uses a single (carefully selected) frame often outperforms recent video-language models [57, 119] on standard video benchmarks such as MSR-VTT [120]. Lei et al. [56] report similar findings with a single-frame pretraining approach. These findings hint at a lack of time awareness in video models. However, it remains unclear if these findings are caused, indeed, by the lack of time in video models or whether the benchmarks themselves do not mandate time awareness. Furthermore, there is no clear definition of what it means for a model to be time aware. In this paper, we strive to shed light on all these factors of time awareness in video-language models. + +As a first step, we consider a simple notion of understanding time, i.e., understanding temporal relations such as before and after [4]. Consider the task presented in Fig. 1. A time invariant model shall be able to associate (A) with (1) + +or (2) and (B) with (3) or (4) based on static frames alone. But to distinguish between (1) and (2), one needs to be able to understand time order and connect it across video and language1. Thus, the first question we ask in Section 3: do the representations learnt by foundational video-language models encode this sense of time? To reliably attribute lack of time awareness to models and not existing benchmarks, we design our own synthetic dataset to probe models for this sense of time. We create video-language pairs that show a sequence of two events. Then, we alter the order of events either in the text or the video and check if models can connect the order in video and language. We find that existing video-language models indeed struggle to associate the time order across video and language. + +In light of these findings, the second question we ask in Section 4 is: can we adapt a video-language model, without expensive re-training from scratch, to instill this sense of time? Towards this, we take inspiration from literature on understanding time in natural language, where there has been much work on developing time aware language models [20, 36, 37, 130, 131]. Our objective is to instill time awareness in a video-language model without having to pretrain from scratch. To do that, we propose TACT: Temporal Adaptation by Consistent Time-ordering based on two key components: (i) we artificially create samples that provide temporal signal, for example, by flipping the order of events in the video or the text, and (ii) we introduce a modified contrastive loss to learn time order consistency based on these samples. Instead of training from scratch, we adapt an existing video-language model, VideoCLIP [61], using the paradigm of post-pretraining on a small amount of videotext data [66, 121]. We demonstrate the effectiveness of TACT in connecting the time order in video and language on four diverse real datasets in Section 5. + +Finally, in line with the original motive of video-language models for zero-shot generalization, we evaluate in Section 6 our TACT-adapted model for three sets of tasks on six downstream datasets which require a varying degree of time awareness. On tasks that need higher time awareness, with the appropriate choice of adaptation dataset, TACT outperforms a strong baseline that is based on post-pretraining on canonical clip-text pairs without consideration of time-order. + +In summary, our contributions are: (i) We show that existing video-language models struggle to associate time order in video and language through controlled experiments on synthetic data and several evaluations on real datasets. (ii) Based on VideoCLIP [119], we propose TACT, a method for temporal adaptation using this time order consistency without having to pretrain from scratch. (iii) We demonstrate improved zero-shot generalizability of TACT-adapted models on tasks that require higher time awareness. + +# 2. Background and Related Work + +We briefly discuss recent advances in video-language models followed by their consideration of time. + +Foundational video-language models. Large-scale datasets, self-supervision, and the advent of Transformers [107] have led to the emergence of powerful encoders for images [21,39,103], videos [5,11,24,104,117], language models [19,64,69,86] and even universal encoders [32,46]. These encoders form the basis for several vision-language foundational models. Popular image-language models such as CLIP [83] and ALIGN [49] are trained on massive datasets by using web images and alt-text. Similarly, videolanguage models are catching up and can be categorised into two broad directions: (i) adapting image-language models for videos [8, 23, 50, 51, 63, 66, 71, 110, 112, 121], and (ii) pure video-based models that are learned using large video-text datasets [3,7,27-29,31,58,62,65,67,68,95,119]. Recently, a new paradigm of post-pretraining has emerged where an existing image- or video-language model goes through another stage of self-supervised pretraining on a small amount of video data before it is evaluated on downstream tasks [66, 121]. This is promising as it circumvents the prohibitive cost of pretraining on large datasets from scratch. In [66], the post-pretraining uses time-invariant mean-pooling, while [121] strives to bridge the domain gap between image captions and video subtitles. In contrast, our proposed temporal adaptation involves post-pretraining of VideoCLIP [119] with a small amount of data that instills the model to learn the time-order of events in a video. + +Time in vision. Time separates videos from static images or an unordered set of frames. While modeling time remains a challenge, it also presents a natural source of supervision that has been exploited for self-supervised learning. For example, as a proxy signal by posing pretext tasks involving spatio-temporal jigsaw [1, 44, 53], video speed [10, 17, 48, 94, 111, 124], arrow of time [78, 80, 114], frame/clip ordering [25, 70, 90, 97, 118], video continuity [61], or tracking [45, 108, 113]. Several works have also used contrastive learning to obtain spatio-temporal representations by (i) contrasting temporally augmented versions of a clip [47, 77, 81], or (ii) encouraging consistency between local and global temporal contexts [9, 18, 85, 123]. Nevertheless, it remains unclear if the learnt representations actually encode time reliably. Time-aware features have also been explored for specific downstream tasks such as action recognition [30, 100, 101]. There has been some very recent work on evaluating self-supervised video representations [87, 99] on their temporal recognition ability instead of only relying on time as a guidance for training. + +In the same spirit, a related direction pursues evaluation and benchmarking of time awareness in video datasets [88], models [13, 14, 30, 56, 89, 125] or both [43, 92]. Huang et + +al. [43] measure the effect of motion on temporal action recognition to find that only a subset of classes in UCF-101 and Kinetics-400 require motion information. Ghodrati et al. [30] propose new tasks to evaluate temporal asymmetry, continuity and causality in video models. Our work derives inspiration from these but applies more generally to video-language models as language provides a basis for open-world generalization. + +Time in language. Time has also been extensively studied in the natural language literature. Early works identified temporal structures in language such as temporal prepositions and quantifiers [4, 79]. More recent literature focuses on tasks such as extracting temporal relations [35, 72-74], as well as temporal reasoning [36, 37, 82, 130, 131]. For example, Han et al. [36, 37] and Zhou et al. [131] pretrain language models specifically to focus on understanding temporal relations such as before, after, during, etc. The emergence of large language models has also spurred an increased interest in developing benchmarks to test for time awareness in these models [20, 75, 76, 102, 106, 129]. For example, Ning et al. [75] propose a new benchmark of reading comprehension with questions involving before/after relations. Since temporal relations in language are grounded in the video, we draw inspiration from [36, 37, 131] and aim to instill time awareness in video-language models. + +Time in video-language models appears implicitly in tasks like video-text alignment [38, 98] and temporal grounding [42, 60]. In this work, we focus on self-supervised video-language models that can generalize to a variety of tasks rather than models designed for a specific task, e.g., temporal grounding. Some recent works have shown the under-utilisation of time in classic video-text benchmarks such as MSR-VTT [120], YouCook [132], ActivityNet [22], and DiDeMo [41]. For example, [13, 56, 57] discover that on several benchmarks, using only one or few frames or clips achieves competitive performance. Adaptations of the popular CLIP architecture for videos (e.g., CLIP4Clip [66]) show that weighted pooling of frames already achieves impressive performance on retrieval benchmarks. + +These raise some key questions: do existing videolanguage models truly understand time in the sense of correctly associating order of events in language and video? If not, can we adapt them to instill time awareness? Our work addresses these questions. There has been some work in using time-order across video and language as a source of self-supervision. Specifically, concurrent to our work, both Sun et al. [96] and Cao et al. [15] propose fine-grained temporal alignment between video and text as the pretraining objective. Different from these works, we consider the notion of time-order and we aim to adapt a given videolanguage model using post-pretraining, which circumvents the need for a new round of compute-intense pretraining. + +![](images/d9515c00a7205a5547f167d10e5e105eb89b5511d2539566856bab8d7dfb87f1.jpg) +Figure 2. Overview of the proposed task to evaluate time-order consistency across synthetic video-language pairs having before/after relations. We also define a control task to check if the synthetic videos are considered out-of-distribution by the model. + +![](images/11780cf09827d603373cf3543228a8e0569fd9158c10836b95ef3a12274c5a67.jpg) + +# 3. Do Video-Language Models Sense Time? + +Probing video-language models for temporal understanding is an open direction of research. In this work, we consider a specific sense of temporal understanding: consistency in the order of events in a video with the associated textual description. For example, consider a text description: A red circle appears before a yellow circle. This imposes an order constraint on the video stream to have the event red circle appears happen before the event yellow circle appears. Can existing video-language models connect time-order in text with that in video? To answer this, we design an experiment with synthetic data. + +Synthetic dataset. We construct simple videos that comprise of a pair of events such as the ones mentioned above. We generate $N = 180$ video-language pairs as a combination of $C = 6$ colors, $S = 3$ shapes, and $|\tau| = 2$ temporal relations: before and after. The corresponding caption describes the order of events connected with a before/after temporal relation. We call this caption as an attractor since it is consistent with the time-ordering in the video. Likewise, we construct a distractor where we flip the order of event descriptions while retaining the temporal relation. An example pair is illustrated in Fig. 2 (left). Ideally, a time aware video-language model should be able to associate the video with the temporally consistent text, or vice versa. We refer to this task as time-order consistency check. To rule out the possibility that synthetic videos are out-of-distribution, we also perform the same experiment with canonical clips with a single event and the text describes that same event as shown in Fig. 2 (right). We refer to this as the control task. + +Choice of models. We consider recent video-language models, broadly categorized into three groups: (i) image-language models like CLIP [83] that are adapted to videos [23,66,128], (ii) pure video-language models trained on a contrastive learning objective [6,16,119], and (iii) pure video-language models trained on a masking objective [29]. + +Findings. We evaluate video-to-text and text-to-video retrieval on both time-order consistency and control tasks. + +
ParadigmMethodVideo-to-TextText-to-Video
◎◎◎◎
Chance-50.050.050.050.0
Image-Language adapted to videoCLIP4Clip [66]49.451.150.049.4
CLIP2Video [23]100.047.897.852.3
CenterCLIP [128]91.746.197.251.1
Video-Language ContrastiveVideoCLIP [119]87.151.166.748.3
Frozen in Time [6]97.849.4100.050.6
VindLU [16]98.052.0100.051.1
Video-Language MaskingBridgeFormer [29]100.051.197.242.2
+ +Table 1. Results on synthetic control ( ⌒ ) and time-order consistency ( ⌒ ) task as described in Fig. 2. Existing video-language models show random performance on our time-order task. + +From Tab. 1, we observe that while most video-language models perform well on the control task, all of them struggle and perform on par with random chance on the temporal task. This gap in performance deserves attention given the importance of time in videos. Note that while synthetic data allows for controlled experiments, we also expand this evaluation to real video datasets in the following section. + +# 4. Adaptation by Consistent Time-Ordering + +We describe a post-pretraining recipe to instill a video-language model with a sense of time. We propose TACT: Temporal Adaptation by Consistency of Time-order, that is based on two key components: (i) we artificially create samples that provide temporal signals, e.g., by flipping the order of events; and (ii) we introduce a modified contrastive loss to learn temporal consistency based on these samples. We start by defining the notation and then describe the key components of our adaptation recipe. + +Preliminaries. Let $\mathcal{V}$ be the space of video clips and $\mathcal{T}$ be the space of text clips. Consider two non-overlapping video clips $v_{i}, v_{j} \in \mathcal{V}$ . Let $\zeta_{i}, \zeta_{j} \in \mathcal{T}$ be their respective captions. Let $\tau$ be a temporal relation, $\tau \in \{\text{before, after}\}$ . Then, we denote a stitched and time-order consistent clip as $(u_{ij}, t_{ij})$ , where $u_{ij} := [v_{i}; v_{j}]$ , $t_{ij} := [\zeta_{i}; \tau; \zeta_{j}]$ , and $[\cdot; \cdot]$ denotes concatenation. Note that depending on $\tau$ , the order of $v_{i}$ and $v_{j}$ may need to change in $u_{ij}$ . For brevity, we drop the subscripts and refer to the stitched pair as $(u, t)$ unless stated otherwise. + +Time-order reversal. The classic contrastive learning paradigm for video-language models aligns components of a video clip $v_{i}$ with its text counterpart $\zeta_{i}$ and contrasts against other texts $\zeta_{j}$ that usually describe a completely different clip. This makes such models ignore the finer details of temporal understanding as it is easier to contrast the negatives by simply focusing on the objects or the scene. This is evident from simple bag-of-word-like methods that + +![](images/166102ea621382b55c43696766ac1217cc91881b037ffa89b9afd5f188d2163d.jpg) +Figure 3. Overview of TACT. Along with the usual contrastive loss, where negatives come from other samples in the batch, we make use of time-order reversal within the same sample and cross samples to generate additional negatives for both video and text. We also extend the contrastive loss to time-order reversed video and text corresponding to reverse consistency $\mathcal{L}_r$ . + +are shown to work well for contrastive learning, both on the visual (e.g., CLIP4Clip [66]) and textual (e.g., MILNCE [67]) modalities. We hypothesize that unless there are negatives in a contrastive setup that contain the same scenes and objects, models need not learn a sense of time. Thus, we present a simple strategy to generate negatives that force the learning process to focus on the temporal order. + +We define a time-order reversal function $\mathbb{T}$ that operates on the stitched video clip or text description and temporally swaps its constituents: + +$$ +\mathbb {T} (u) = \mathbb {T} ([ v _ {i}; v _ {j} ]) := [ v _ {j}; v _ {i} ], \quad \text {a n d} \tag {1} +$$ + +$$ +\mathbb {T} (t) = \mathbb {T} ([ \zeta_ {i}; \tau ; \zeta_ {j} ]) := [ \zeta_ {j}; \tau ; \zeta_ {i} ]. \tag {2} +$$ + +An illustration of $\mathbb{T}$ is shown in Fig. 3. Note that $\mathbb{T}$ does not reverse the actual video, i.e., time does not flow backwards, but only changes the order in which events happen in the stitched clips. Our objective is to train a model that is able to distinguish between the original pair $(u,t)$ and time-reversed versions $(u,\mathbb{T}(t))$ , and $(\mathbb{T}(u),t)$ . + +Loss function. We assume access to an existing pre-trained video-language model with a visual encoder $f_{\theta}$ and text encoder $g_{\phi}$ . We obtain the video encoding $\mathbf{z}_u \coloneqq f_{\theta}(u) \in \mathbb{R}^d$ and the text encoding $\mathbf{z}_t \coloneqq g_{\phi}(t) \in \mathbb{R}^d$ . Our goal is to adapt $\Theta = \{\theta, \phi\}$ via post-pretraining such that the resulting model is time aware while maintaining its original performance on tasks such as retrieval. As we aim to use a small amount of data, we only update some parameters of + +the model $(\Theta)$ , such as the last few layers. + +We now introduce our recipe for temporal adaptation based on the InfoNCE loss [105] to learn time-order sensitive video-text correspondence. For a positive (or time-order consistent) video-text pair $(u, t)$ , we first define a forward loss where the stitched pair is in its original time-order. + +$$ +\mathcal {L} _ {f} = \sum_ {(u, t) \in \mathcal {B}} \left(\operatorname {T N C E} \left(\mathbf {z} _ {u}, \mathbf {z} _ {t}\right) + \operatorname {T N C E} \left(\mathbf {z} _ {t}, \mathbf {z} _ {u}\right)\right), \tag {3} +$$ + +where TNCE is the Noise Contrastive Estimation (NCE) loss for temporal adaptation, defined as: + +$$ +\operatorname {T N C E} \left(\mathbf {z} _ {u}, \mathbf {z} _ {t}\right) := - \log \frac {\exp \left(\mathbf {z} _ {u} \cdot \mathbf {z} _ {t}\right)}{\sum_ {t ^ {\prime} \in \mathcal {B} _ {t}} \exp \left(\mathbf {z} _ {u} \cdot \mathbf {z} _ {t ^ {\prime}}\right) + \mathcal {C} ^ {\text {t i m e}}}, \tag {4} +$$ + +where $\mathcal{B}$ is the batch of $(u,t)$ pairs and $\mathcal{B}_t$ specifically refers to other stitched text captions in the batch. $\mathcal{C}^{\mathrm{time}}$ accumulates negatives defined using time-order reversal as: + +$$ +\mathcal {C} ^ {\text {t i m e}} = \alpha_ {\text {s a m e}} \exp (\mathbf {z} _ {u} \cdot \mathbf {z} _ {\mathbb {T} (t)}) + \alpha_ {\text {c r o s s}} \sum_ {t ^ {\prime} \in \mathcal {B} _ {t} \backslash \{t \}} \exp (\mathbf {z} _ {u} \cdot \mathbf {z} _ {\mathbb {T} \left(t ^ {\prime}\right)}), \tag {5} +$$ + +where $\alpha_{\mathrm{same}}$ controls the effect of contrasting text from the same sample but with reversed text time-order, i.e., $\mathbb{T}(t)$ , and $\alpha_{\mathrm{cross}}$ encourages the model to contrast between reversed versions of other text captions, i.e., $\mathbb{T}(t')$ . Note that when both $\alpha_{\mathrm{same}}$ and $\alpha_{\mathrm{cross}}$ are 0, we revert back to the standard NCE formulation, albeit on stitched pairs. While Eq. (4) corresponds to the video-text loss $\mathrm{TNCE}(\mathbf{z}_u,\mathbf{z}_t)$ , the text-video loss $\mathrm{TNCE}(\mathbf{z}_t,\mathbf{z}_u)$ is defined symmetrically. Furthermore, we also apply a reverse loss $\mathcal{L}_r$ to bring time-order reversed versions of both the video and the text together. Note that as we consider $(u,t)$ as a positive pair, $(\mathbb{T}(u),\mathbb{T}(t))$ also form a positive pair, + +$$ +\mathcal {L} _ {r} = \sum_ {\left(\mathbb {T} (u), \mathbb {T} (t)\right) \in \mathcal {B}} \left(\operatorname {T N C E} \left(\mathbf {z} _ {\mathbb {T} (u)}, \mathbf {z} _ {\mathbb {T} (t)}\right) + \operatorname {T N C E} \left(\mathbf {z} _ {\mathbb {T} (t)}, \mathbf {z} _ {\mathbb {T} (u)}\right)\right). \tag {6} +$$ + +Here, the TNCE term operates on time-reversed clips and $\mathcal{C}^{\mathrm{time}}$ contrasts $(\mathbb{T}(u),\mathbb{T}(t))$ with un-reversed text clips in the batch $(\mathbb{T}(u),t)$ . The overall loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {f} + \beta \mathcal {L} _ {r}. \tag {7} +$$ + +Depending on the choice of loss coefficients $\alpha_{\mathrm{same}}, \alpha_{\mathrm{cross}}, \beta \in \{0,1\}$ , we can vary properties of the adapted model. For example, setting $\alpha_{\mathrm{same}} = 1$ encourages high sensitivity to time-order reversal. As we will see empirically, the loss coefficients also provide the flexibility to adapt the model to various datasets. + +We illustrate this temporal extension of the contrastive loss in Fig. 3 (best seen in colour). $\mathbb{T}$ illustrates the time order reversal function. The top half corresponds to $\mathcal{L}_f$ while the bottom half visualizes $\mathcal{L}_r$ . In particular, the top-left quadrant alone corresponds to the standard contrastive loss on stitched pairs. While the green diagonal terms are positive pairs, the red diagonal terms are the strongest drivers for instilling temporal understanding in the model. + +
DatasetTrainValidationTestEgoLength
NvNcNvNcNvNc(s)
TEMPO3,90428,4274111,0003961,000X30
ActivityNet7,44095,474453906456912X120
Charades5,26299,9285001,0005001,000X30
Charades-Ego2,679155,3065001,000210420✓31
+ +Table 2. Statistics of datasets we consider for temporal adaptation. $N_{\mathcal{V}}$ is the number of unique videos and $N_{\mathcal{C}}$ is the number of stitched clips. Based on $N_{\mathcal{V}}$ , TEMPO and Charades-Ego are smaller as compared to ActivityNet and Charades. + +# 5. Experiments: TACT Ablations + +Base model. We demonstrate the effectiveness of TACT as an adaptation recipe on top of VideoCLIP [119] owing to its simple architecture, contrastive objective, and use of pre-computed S3D [116] features that enable faster experimentation and allow encoding a long temporal context ( $\sim 32$ secs). We initialize $\Theta$ from the model pretrained on HowTo100M [68] and post-pretrain on multiple datasets. + +Datasets. One of our key objectives is to post-pretrain on a small amount of data in contrast to massive pretraining datasets such as WebVid2M [7] or HowTo100M [68]. We consider dense video captioning datasets that offer sufficient diversity in terms of size, backgrounds, clip durations, viewpoints and activities. Specifically, we experiment with: (i) TEMPO [40]: the subset of stitched diverse third-person videos from DiDeMo [41] with text descriptions for fixed 5s segments that contain before/after relations; (ii) ActivityNet Captions [55]: a dense video captioning dataset with human-centric actions; (iii) Charades [93]: a scripted indoor daily human activities video dataset; and (iv) Charades-Ego [91]: similar to Charades, scripted human activities from the egocentric viewpoint. To construct stitched clips, we randomly sample any two non-overlapping clip-text pairs in the video. Since we require stitched clips instead of raw videos, we create new splits for each dataset (see Tab. 2). + +Evaluation metrics. We evaluate the post-pretrained model using two sets of metrics: (i) standard retrieval metrics, recall $R@1$ , $R@5$ , $R@10$ and median-rank evaluated on stitched video-text clips; and (ii) time-order consistency, i.e., the fraction of videos for which the model correctly associates text that is time order consistent with the video: + +$$ +A _ {\text {t i m e}} := \frac {1}{| \mathcal {D} |} \sum_ {(u, t) \in \mathcal {D}} \mathbb {I} [ d (\mathbf {z} _ {u}, \mathbf {z} _ {t}) < d (\mathbf {z} _ {u}, \mathbf {z} _ {\mathbb {T} (t)}) ], \tag {8} +$$ + +where $(u,t)$ are time-order consistent pairs, $\mathcal{D}$ is the dataset, and $d(\cdot ,\cdot)$ is a distance metric based on cosine similarity. + +Post-pretraining details. We freeze the word embeddings + +and layers 1 to 5 for both the video and text encoders in VideoCLIP. For adaptation, we use the Adam optimizer [54] with learning rate $5e^{-6}$ , batch size 32 trained on a single node with 4 GeForce GTX 1080 GPUs. On TEMPO, we train for 60 epochs while on the other datasets, we train for 10 epochs and pick the checkpoint that maximizes the geometric mean of $R@1$ and $A_{\mathrm{time}}$ on the respective validation set. A typical adaptation run takes about 1-3 hours. + +Evaluation on the test set. Results in Tab. 3 show that TACT\* with optimal loss coefficients outperforms TACT† (all 0 loss coefficients) and the zero-shot baseline (no postpretraining similar to the synthetic data experiment), both on the retrieval and time-order consistency tasks. This indicates the robustness of the adaptation. + +Impact of loss coefficients. Choosing appropriate values for loss coefficients $\Theta_l \coloneqq \{\alpha_{\mathrm{same}}, \alpha_{\mathrm{cross}}, \beta\}$ allows the model to learn various aspects and adapt using different datasets. On each dataset, we vary $\Theta_l \in \{0,1\}^3$ and find the best configuration based on the GeometricMean $(R@1, \max(A_{\mathrm{time}} - 50, 0))$ on the validation sets. The above metric ensures the geometric mean is not overpowered by $R@1$ . The results are shown in Tab. 4. As $\alpha_{\mathrm{same}}$ is directly responsible for discriminating between original and time-reversed orders, unsurprisingly, setting it to 1 is necessary to achieve the best results on $A_{\mathrm{time}}$ for all the datasets. For TEMPO and Charades-Ego, using all loss components (all 1) provides the best results, whereas $\alpha_{\mathrm{cross}} = 1$ and $\beta = 0$ achieves a better trade-off for ActivityNet and Charades. Choosing $\beta = 1$ leads to an improve + +
DatasetMethodRetrievalTime-order
R@1↑MedR ↓Atime ↑
TEMPOZero-shot3.749.048.1
TACT†7.713.046.5
TACT*9.39.066.5
ActivityNetZero-shot1.144.049.6
TACT†5.834.059.7
TACT*5.835.085.7
CharadesZero-shot1.3170.047.1
TACT†5.338.573.5
TACT*5.735.077.0
Charades-EgoZero-shot1.664.053.7
TACT†6.435.060.1
TACT*10.128.568.2
+ +Table 3. Results for the best TACT model on test sets. $\mathrm{TACT}^{\star}$ has optimal loss coefficients and $\mathrm{TACT}^{\dagger}$ is a baseline with all coefficients 0. On time order, TACT generalizes well with $\mathrm{TACT}^{\star}$ outperforming the baselines. On retrieval, for TEMPO and Charades-Ego, $\mathrm{TACT}^{\star}$ outperforms the baseline as their optimal models have $\beta = 1$ which helps retrieval with a small amount of data. + +![](images/d0ad7ac203b57eb66af53635addd83ee6d31039c0ecd77dd1f9031cf8dc486c5.jpg) +Figure 4. Time-distance between stitched clips in datasets for temporal adaptation $(\Delta_{\mathrm{time}})$ . TEMPO has stitched clips close to each other while those in Charades-Ego are farthest apart suggesting a correlation between $\Delta_{\mathrm{time}}$ and difficulty of temporal adaptation. + +ment in retrieval performance for TEMPO and Charades-Ego but leads to a decline for ActivityNet and Charades. We attribute this to the number of unique videos in the train set for these datasets. As ActivityNet and Charades have more videos than TEMPO or Charades-Ego (see train $N_{\mathcal{V}}$ Tab. 2) additional positives introduced by setting $\beta = 1$ are not necessary and in fact, hurt performance. Finally, we note that carefully setting the value of $\alpha_{\mathrm{same}}$ provides a convenient trade-off between spatial and temporal understanding. Please refer to the supplement for detailed experiments. + +What makes temporal adaptation hard? We observe a large gap in $A_{\mathrm{time}}$ between TEMPO and ActivityNet. We hypothesize that the distance (in seconds) between the two clips ( $\Delta_{\mathrm{time}}$ ) in a stitched clip is strongly correlated with the difficulty of adaptation. Intuitively, it is easier to infer the time-order consistency for a stitched clip $u$ with text $t$ that has distant constituent clips $v_i, v_j$ since the objects and scene could be vastly different. In contrast, it is harder to discern the correct time-order when the constituent clips are closer in time. Fig. 4 shows distribution of $\Delta_{\mathrm{time}}$ for each dataset. Indeed, $\Delta_{\mathrm{time}}$ in ActivityNet (58.8s) is much higher than that in TEMPO (6.4s) making the task harder on TEMPO. To further test our hypothesis, we conduct a controlled experiment where we gradually vary the distribution of $\Delta_{\mathrm{time}}$ for Charades-Ego to match it to that of TEMPO. We find a strong correlation ( $\rho = 0.92$ ) between $\Delta_{\mathrm{time}}$ and hardness of adaptation. More details are in the supplement. + +# 6. Experiments: Downstream Evaluation + +The goal of video-language foundation models is to generalize in a zero- or few-shot manner to a diverse range of downstream tasks. We evaluate TACT models on three sets of downstream tasks that need low-to-high time awareness. + +Baseline: Standard post-pretraining. Comparing our temporally adapted models with pretrained VideoCLIP is not fair since adapted models see data beyond the pretraining phase. In addition to the zero-shot comparison, we compare against a baseline model that is trained for standard video-text retrieval on the same datasets as temporal adaptation. Instead of using stitched clips, we use simple canonical pairs, i.e., $(v_{i},\zeta_{i})$ instead of $(u_{ij},t_{ij})$ . + +
Loss coefficientsTEMPOActivityNetCharadesCharades-Ego
α_sameα_crossβR@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑
Chance0.1500.050.00.1453.050.00.1500.050.00.1500.050.0
0008.314.049.46.430.057.35.742.071.52.944.064.6
0018.214.049.55.627.047.04.258.075.13.241.565.2
0108.215.049.36.129.078.85.245.078.93.438.064.5
0118.114.049.55.827.048.34.258.075.13.141.067.0
1006.420.060.65.928.079.16.138.076.33.242.066.1
1016.524.062.95.626.063.14.951.078.03.339.070.7
1105.924.059.76.029.086.36.643.077.83.740.567.9
1117.517.062.55.727.063.85.151.077.73.838.568.3
+ +Table 4. Impact of loss coefficients for TACT post-pretraining on validation sets of various datasets. For each dataset, the corresponding color-marked row denotes the best configuration based on the geometric mean of $R@1$ and $A_{\mathrm{time}}$ . TACT is able to connect time-order in video and language while maintaining its retrieval capabilities across several datasets. + +
AdaptationI. Text-to-Video RetrievalII. Temporal VQAIII. Action-to-Video Retrieval
DatasetMethodMSR-VTTYouCookIINext-QA (ATP)AGQASSv2SSv2 (events)Temporal
R@1R@5R@10R@1R@5R@10AccuracyAccuracymAPmAPmAP
-Chance0.10.51.00.10.51.020.050.00.62.02.0
-None10.623.429.918.245.559.923.449.93.46.413.0
TEMPOBaseline12.029.337.321.548.261.825.050.83.96.815.9
TACT12.826.535.720.445.158.727.657.14.27.716.1
ActivityNetBaseline15.734.444.915.638.851.423.750.73.77.016.0
TACT13.829.639.616.036.949.825.450.33.57.216.2
CharadesBaseline12.325.833.621.548.661.726.050.54.17.113.7
TACT11.725.232.422.449.162.425.254.84.37.814.1
Charades-EgoBaseline13.127.534.519.447.160.824.349.74.06.914.7
TACT11.124.630.621.948.261.925.658.43.96.913.5
+ +Table 5. Results on downstream zero-shot evaluation with tasks requiring increasing time awareness from I to III. None corresponds to direct evaluation of the VideoCLIP model on the downstream dataset. Green denotes an improvement for the TACT adapted model w.r.t. the baseline, red denotes a deterioration. As we move from tasks that need low to high time awareness, the effectiveness of TACT increases. See Sec. 6 for a more detailed discussion. The table is best viewed on screen in colour. + +Evaluating TACT adapted models on synthetic data. On the video-to-text variant, TACT adapted on TEMPO achieves $64.4\%$ , ActivityNet $52.5\%$ , Charades $65.0\%$ , Charades-Ego $85.6\%$ . This is usually higher than the performance that non-adapted models achieve in Tab. 1. This highlights that TACT models learn useful signals to match time-order in video and language. + +I. Text-to-video retrieval. We consider two widely used benchmarks: MSR-VTT [120] and YouCookII [132] and adopt standard retrieval metrics. Recent work has identified a bias for spatial understanding in these datasets [8, 13, 43, 56, 59, 66]. Thus, we consider this class of tasks as requiring low time awareness. As shown in Tab. 5 set I, on MSR-VTT [120], we observe that TACT is worse (marked in red) or at par with the baseline across datasets. This + +aligns well with findings in [13, 56] that these benchmarks do not need time awareness. On YouCookII [132], TACT models based on Charades(-Ego) outperform the baseline (marked in green). We believe this is a consequence of a lower domain shift between YouCookII and Charades. + +II. Temporal video QA. Next, we use subsets of recently released multiple-choice video question answering benchmarks: Next-QA [115] and AGQA [34]. The idea is to check if we can probe models for temporal understanding by asking questions with temporal language. Buch et al. [13] identify a subset of Next-QA, dubbed as ATP-hard², with a higher concentration of temporally challenging data. For AGQA, we pick a subset of $\sim 6k$ questions that explicitly have a question with before/after relation. We consider + +these benchmarks as requiring a moderate-high level of time awareness and AGQA in particular is also close to our adaptation task. We use accuracy as the standard metric. We observe (see Tab. 5 set II) that indeed TACT almost always outperforms (marked in green) baselines on both Next-QA and AGQA. TEMPO-adapted TACT seems to generalize particularly well on both benchmarks. Likewise, Charades-adapted TACT does well on AGQA since AGQA is also based on the Charades videos accounting for reduced domain shift. We affirm that temporal adaptation is useful, especially when the downstream tasks require it. + +III. Action-to-video retrieval. Finally, we consider action recognition benchmarks such as Something-Something (SSv2) [33] and Temporal [88]. SSv2 was designed to capture richer temporal information [33, 56]. We follow Lei et al. [56], who propose the template-retrieval task that encourages temporal modelling and use their evaluation split3 containing $C = 174$ actions and $K = 12$ videos per class. Interestingly, different actions in SSv2 require differing levels of time awareness. We create a subset SSv2 (events) with $C_{\text{events}} = 49$ actions that have at least two verbs in the label as the occurrence of multiple verbs is indicative of multiple events occurring in sequence. Finally, we also evaluate against the Temporal benchmark [88], a combination of 50 action classes from SSv2 [33] and Kinetics-400 [52] for which temporal information is deemed to be essential for recognition. Similar to text-to-video retrieval, we use the action class as a text query and obtain a ranking over all videos. Different from the retrieval setup, since a single query has multiple correct answers (up to $K = 12$ videos), we report mAP as the metric for these benchmarks. This task set needs high time awareness. Furthermore, unlike QA tasks in II, there is a shift in several (uncontrolled) factors as we move from temporal adaptation task to these tasks. From Tab. 5, we observe that TEMPO and Charades-adapted models generalize well across set III benchmarks. ActivityNet-adapted TACT underperforms on SSv2 but outperforms the baseline on strongly temporal actions in SSv2 (events) and Temporal. Finally, TACT adapted on Charades-Ego is at-par or slightly worse than the baseline on SSv2 variants, and also on Temporal, perhaps due to the shift from egocentric to third-person videos. Overall, despite SSv2 and Temporal requiring high time awareness, TACT models show promising zero-shot generalization with the right choice of the adaptation dataset. + +# 7. Discussion and Conclusion + +Generalization to other temporal prompts. The time-order of events in language can be described using various sentence structures. Although we train video-language models using before/after relations, it is natural to ask if + +![](images/61536b1950e6d38904c1f501b46a3b61bd72c9c087aac9deb41cbf0b23d8ebd8.jpg) +Figure 5. Models trained by TACT with before/after relations generalize to a new kind of prompt such as First, ..., then ... indicating learning of the underlying true time-order of events. + +the model still correctly associates time-order for a different prompt such as First, ..., then, ... To systematically test this, we gather event pairs $E_1$ , $E_2$ ( $E_1$ occurs before $E_2$ in the video) for each sample in the validation set and stitch them using three prompts as follows: (i) $E_1$ before $E_2$ , (ii) $E_2$ after $E_1$ , (iii) First $E_1$ , then $E_2$ . As shown in Fig. 5, TACT-adapted models generalize well to a new prompt (iii). This substantiates the learning of time-order of events rather than merely learning the order of words in the sentence. + +Limitations. While we present a promising way of instilling time in video-language models, our work is limited to the VideoCLIP [119] pretrained model. Our initial experiments with Frozen in Time [6] were not as promising, perhaps because it uses a much shorter temporal context (4 frames). Please see the supplement for results on more pretrained models. Furthermore, we consider a specific definition of time awareness derived from temporal relations like before/after. It is natural to ask if this can be extended to more general notions of temporality, e.g., as defined by Allen [4]. Finally, there can always be more downstream tasks considered such as (spatio-)temporal localization. + +Conclusion. Given the essence of time in video-language models, we present a simple experiment based on synthetic data to test for time awareness. We find that existing models lack a sense of time defined in terms of consistency of order of events in video and language. To fill this gap, building upon VideoCLIP [119], we present TACT, a recipe to instill this sense of time in video-language models. Finally, we analyze the zero-shot generalizability of TACT-adapted models to a diverse set of tasks. We hope that this work provokes further probing and instilling time awareness in video-language models, and also inspires other adaptations of foundational models to solve various challenging tasks. + +Acknowledgements. We acknowledge support from ELLIS Amsterdam and the AMS Scholarship to Piyush. We thank Dr Dennis Koelma for their help with compute infrastructure and Dr Hazel Doughty for useful discussions. + +# References + +[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Winter Conference on Applications of Computer Vision (WACV), pages 179-189. IEEE, 2019. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1 +[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. ArXiv, abs/2006.16228, 2020. 2 +[4] James F. Allen. Towards a general theory of action and time. In Artificial Intelligence, 1984. 1, 3, 8 +[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 2 +[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 4, 8 +[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 5 +[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval, 2022. 2, 7 +[9] Nadine Behrmann, Mohsen Fayyaz, Juergen Gall, and Mehdi Noroozi. Long short view feature decomposition via contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9224-9233, 2021. 2 +[10] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9922–9931, 2020. 2 +[11] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021. 2 +[12] Rishi Bommasani et al. On the opportunities and risks of foundation models. ArXiv, abs/2108.07258, 2021. 1 +[13] Shyamal Buch, Cristobal Eyzaguirre, Adrien Gaidon, Jia-jun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting + +the "Video" in Video-Language Understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7 +[14] Petr Byvshev, Pascal Mettes, and Yu Xiao. Are 3d convolutional networks inherently biased towards appearance? Computer Vision and Image Understanding, 220:103437, 2022. 2 +[15] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision (ECCV), 2022. 3 +[16] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051, 2022. 3, 4 +[17] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised spatio-temporal representation learning using variable playback speed prediction. IEEE Access, 9:79562-79571, 2021. 2 +[18] Ishan Rajendra Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TcIr: Temporal contrastive learning for video representation. Computer Vision and Image Understanding (CVIU), 219:103406, 2022. 2 +[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2 +[20] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 3 +[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 2 +[22] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 1, 3 +[23] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 3, 4 +[24] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 2 +[25] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Conference on Computer + +Vision and Pattern Recognition (CVPR), pages 3636-3645, 2017. 2 +[26] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 1 +[27] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 2 +[28] Valentin Gabeur, Chen Sun, Alahari Karteek, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision (ECCV), 2020. 2 +[29] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 16167-16176, June 2022. 2, 3, 4 +[30] Amir Ghodrati, Efstratios Gavves, and Cees G. M. Snoek. Video time: Properties, encoders and evaluation. In British Machine Vision Conference (BMVC), 2018. 2, 3 +[31] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. ArXiv, abs/2011.00597, 2020. 2 +[32] Rohit Girdhar, Mannat Singh, Nikhil Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. Conference on Computer Vision and Pattern Recognition (CVPR), pages 16081-16091, 2022. 2 +[33] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The "something something" video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 8 +[34] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7 +[35] Rujun Han, Qiang Ning, and Nanyun Peng. Joint event and temporal relation extraction with shared representations and structured prediction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3 +[36] Rujun Han, Xiang Ren, and Nanyun Peng. Deer: A data efficient language model for event temporal reasoning. ArXiv, abs/2012.15283, 2020. 2, 3 +[37] Rujun Han, Xiang Ren, and Nanyun Peng. Econet: Effective continual pretraining of language models for event temporal reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 2, 3 +[38] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2906, 2022. 3 + +[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In _Conference on Computer Vision and Pattern Recognition (CVPR), CVPR '16, pages 770–778, June 2016._2 +[40] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), pages 1380-1390, 2018. 5 +[41] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. International Conference on Computer Vision (ICCV), pages 5804-5813, 2017. 1, 3, 5 +[42] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), 2018. 3 +[43] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7366–7375, 2018. 2, 3, 7 +[44] Yuqi Huo, Mingyu Ding, Haoyu Lu, Zhiwu Lu, Tao Xiang, Ji-Rong Wen, Ziyuan Huang, Jianwen Jiang, Shiwei Zhang, Mingqian Tang, Songfang Huang, and Ping Luo. Self-supervised video representation learning with constrained spatiotemporal jigsaw. In International Joint Conference on Artificial Intelligence (IJCAI), 2021. 2 +[45] A. Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. ArXiv, abs/2006.14613, 2020. 2 +[46] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver: General perception with iterative attention. In International Conference on Machine Learning (ICML), 2021. 2 +[47] S. Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9950-9960, 2021. 2 +[48] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), pages 425-442, 2020. 2 +[49] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2 +[50] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2 +[51] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2 + +[52] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8 +[53] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Association for the Advancement of Artificial Intelligence (AAAI), pages 8545-8552, 2019. 2 +[54] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6 +[55] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-Captioning Events in Videos. In International Conference on Computer Vision (ICCV), 2017. 5 +[56] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning, 2022. 1, 2, 3, 7, 8 +[57] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. Conference on Computer Vision and Pattern Recognition (CVPR), pages 7327-7337, 2021. 1, 3 +[58] Gen Li, Nan Duan, Yuejian Fang, Daxin Jiang, and Ming Zhou. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Association for the Advancement of Artificial Intelligence (AAAI), 2020. 2 +[59] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 1, 7 +[60] Juncheng Li, Junlin Xie, Long Qian, Linchao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal grounding with structured variational cross-graph correspondence learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 3022-3031, 2022. 3 +[61] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. ArXiv, abs/2112.05883, 2022. 2 +[62] Kevin Lin, Alex Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z. Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Ego-centric video-language pretraining. ArXiv, abs/2206.01670, 2022. 2 +[63] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Y. Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. ArXiv, abs/2208.03550, 2022. 2 +[64] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2 + +[65] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2 +[66] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2, 3, 4, 7 +[67] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. Conference on Computer Vision and Pattern Recognition (CVPR), pages 9876-9886, 2020. 2, 4 +[68] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In International Conference on Computer Vision (ICCV), 2019. 2, 5 +[69] Tomas Mikolov, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In International Conference on Learning Representations (ICLR), 2013. 2 +[70] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), pages 527-544, 2016. 2 +[71] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. European Conference on Computer Vision (ECCV), 2022. 2 +[72] Qiang Ning, Zhili Feng, and Dan Roth. A structured learning approach to temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2017. 3 +[73] Qiang Ning, Zhili Feng, Hao Wu, and Dan Roth. Joint reasoning for temporal and causal relations. In Association of Computational Linguistics (ACL), 2018. 3 +[74] Qiang Ning, Sanjay Subramanian, and Dan Roth. An improved neural baseline for temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3 +[75] Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. Torque: A reading comprehension dataset of temporal ordering questions. ArXiv, abs/2005.00242, 2020. 3 +[76] Qiang Ning, Ben Zhou, Zhili Feng, Haoruo Peng, and Dan Roth. Cogcomptime: A tool for understanding time in natural language. ArXiv, abs/1906.04940, 2018. 3 +[77] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. Conference on Computer Vision and Pattern Recognition (CVPR), pages 11200-11209, 2021. 2 +[78] Lyndsey C. Pickup, Zheng Pan, Donglai Wei, YiChang Shih, Changshui Zhang, Andrew Zisserman, Bernhard + +Scholkopf, and William T. Freeman. Seeing the arrow of time. In *Conference on Computer Vision and Pattern Recognition* (CVPR), pages 2043-2050, 2014. 2 +[79] Ianthe Pratt and Nissim Francez. Temporal prepositions and temporal generalized quantifiers. Linguistics and Philosophy, 24:187-222, 2001. 3 +[80] Will Price and Dima Damen. Retro-actions: Learning 'close' by time-reversing 'open' videos. International Conference on Computer Vision Workshops (ICCVW), pages 1371-1380, 2019. 2 +[81] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, H. Wang, Serge J. Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6960-6970, 2021. 2 +[82] Lianhui Qin, Aditya Gupta, Shyam Upadhyay, Luheng He, Yejin Choi, and Manaal Faruqui. Timedial: Temporal commonsense reasoning in dialog. In Association of Computational Linguistics (ACL), 2021. 3 +[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3 +[84] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1 +[85] Adrià Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Patraucean, Florent Altch'e, Michael Valko, Jean-Bastien Grill, Aäron van den Oord, and Andrew Zisserman. Broaden your views for self-supervised video learning. International Conference on Computer Vision (ICCV), pages 1235–1245, 2021. 2 +[86] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108, 2019. 2 +[87] Madeline Chantry Schiappa, Yogesh Singh Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ArXiv, abs/2207.00419, 2022. 2 +[88] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. CoRR, abs/1907.08340, 2019. 2, 8 +[89] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? International Conference on Computer Vision (ICCV), pages 9641-9649, 2021. 2 +[90] Vivek Sharma, Makarand Tapaswi, and Rainer Stiefelhagen. Deep Multimodal Feature Encoding for Video Ordering. In International Conference on Computer Vision Workshops (ICCVW), 2019. 2 +[91] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint + +modeling of first and third-person videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7396-7404, 2018. 5 +[92] Gunnar A. Sigurdsson, Olga Russakovsky, and Abhinav Kumar Gupta. What actions are needed for understanding human actions in videos? 2017 IEEE International Conference on Computer Vision (ICCV), pages 2156-2165, 2017. 2 +[93] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, Cham, 2016. 5 +[94] Ankit Singh, Omprakash Chakraborty, Ashutosh Varshney, Rameswar Panda, Rogério Schmidt Feris, Kate Saenko, and Abir Das. Semi-supervised action recognition with temporal contrastive learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 10384-10394, 2021. 2 +[95] Chen Sun, Austin Myers, Carl Vondrick, Kevin P. Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. International Conference on Computer Vision (ICCV), pages 7463-7472, 2019. 2 +[96] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[97] Tomoyuki Suzuki, Takahiro Itazuri, Kensho Hara, and Hirokatsu Kataoka. Learning spatiotemporal 3d convolution with video order self-supervision. In European Conference on Computer Vision Workshops (ECCV), pages 0–0, 2018. 2 +[98] Makarand Tapaswi, Martin Bauml, and Rainer Stiefelhagen. Book2movie: Aligning video scenes with book chapters. In CVPR, June 2015. 3 +[99] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees G. M. Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 2 +[100] Fida Mohammad Thoker, Hazel Doughty, and Cees Snoek. Tubelet-contrastive self-supervision for video-efficient generalization, 2023. 2 +[101] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Skeleton-contrastive 3d action representation learning. ACM MM, 2021. 2 +[102] Shivin Thukral, Kunal Kukreja, and Christian Kavouras. Probing language models for understanding of temporal expressions. In BLACKBOXNLP, 2021. 3 +[103] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. arXiv preprint arXiv:2105.01601, 2021. 2 + +[104] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6450-6459, 2018. 2 +[105] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. ArXiv, abs/1807.03748, 2018. 5 +[106] Siddharth Vashishtha, Adam Poliak, Yash Kumar Lal, Benjamin Van Durme, and Aaron Steven White. Temporal reasoning in natural language inference. In FINDINGS, 2020. 3 +[107] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017. 1, 2 +[108] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2563-2572, 2021. 2 +[109] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 1 +[110] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 2 +[111] Jiangliu Wang, Jianbo Jiao, and Yun-Hui Liu. Self-supervised video representation learning by pace prediction. In European Conference on Computer Vision (ECCV), pages 504-521, 2020. 2 +[112] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 2 +[113] X. Wang, A. Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2561-2571, 2019. 2 +[114] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 8052-8060, 2018. 2 +[115] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7 +[116] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In European Conference on Computer Vision (ECCV), pages 318–335, Cham, 2018. 5 + +[117] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin P. Murphy. Rethinking spatiotemporal feature learning for video understanding. ArXiv, abs/1712.04851, 2017. 2 +[118] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10334-10343, 2019. 2 +[119] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing (EMNLP), pages 6787-6800, 2021. 1, 2, 3, 4, 5, 8 +[120] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1, 3, 7 +[121] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Rui Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. ArXiv, abs/2209.06430, 2022. 2 +[122] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1 +[123] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Back to the future: Cycle encoding prediction for self-supervised contrastive video representation learning. ArXiv, abs/2010.07217, 2020. 2 +[124] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6548-6557, 2020. 2 +[125] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. Clevrer: Collision events for video representation and reasoning. ArXiv, abs/1910.01442, 2020. 2 +[126] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel C. F. Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. ArXiv, abs/2111.11432, 2021. 1 +[127] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 1 +[128] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In SIGIR, 2022. 3, 4 + +[129] Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. "going on a vacation" takes longer than "going for a walk": A study of temporal commonsense understanding. *ArXiv*, abs/1909.03065, 2019. 3 +[130] Ben Zhou, Qiang Ning, Daniel Khashabi, and Dan Roth. Temporal common sense acquisition with minimal supervision. ArXiv, abs/2005.04304, 2020. 2, 3 +[131] Ben Zhou, Kyle Richardson, Qiang Ning, Tushar Khot, Ashish Sabharwal, and Dan Roth. Temporal reasoning on implicit events from distant supervision. ArXiv, abs/2010.12753, 2021. 2, 3 +[132] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence (AAAI), pages 7590-7598, 2018. 3, 7 \ No newline at end of file diff --git a/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/images.zip b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bc13216f9aaa828a91eb5301b16787141749ec5d --- /dev/null +++ b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e74e708d30797b66fcbdf6b1f9687a8094dd4cba025ba6578f1c413d3cae9526 +size 493075 diff --git a/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/layout.json b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79b3c5af866135703f442853688070bbfe4195bc --- /dev/null +++ b/2023/Test of Time_ Instilling Video-Language Models With a Sense of Time/layout.json @@ -0,0 +1,12709 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "Test of Time: Instilling Video-Language Models with a Sense of Time" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 144, + 171, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 144, + 171, + 158 + ], + "spans": [ + { + "bbox": [ + 100, + 144, + 171, + 158 + ], + "type": "text", + "content": "Piyush Bagad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 159, + 188, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 159, + 188, + 171 + ], + "spans": [ + { + "bbox": [ + 84, + 159, + 188, + 171 + ], + "type": "text", + "content": "University of Amsterdam" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 250, + 144, + 343, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 144, + 343, + 158 + ], + "spans": [ + { + "bbox": [ + 250, + 144, + 343, + 158 + ], + "type": "text", + "content": "Makarand Tapaswi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 264, + 159, + 329, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 159, + 329, + 171 + ], + "spans": [ + { + "bbox": [ + 264, + 159, + 329, + 171 + ], + "type": "text", + "content": "IIIT Hyderabad" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 413, + 144, + 501, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 144, + 501, + 156 + ], + "spans": [ + { + "bbox": [ + 413, + 144, + 501, + 156 + ], + "type": "text", + "content": "Cees G.M. Snoek" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 405, + 159, + 509, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 159, + 509, + 171 + ], + "spans": [ + { + "bbox": [ + 405, + 159, + 509, + 171 + ], + "type": "text", + "content": "University of Amsterdam" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 214, + 174, + 379, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 174, + 379, + 184 + ], + "spans": [ + { + "bbox": [ + 214, + 174, + 379, + 184 + ], + "type": "text", + "content": "bpiyush.github.io/testoftime-website" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 238, + 290, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 501 + ], + "type": "text", + "content": "Modelling and understanding time remains a challenge in contemporary video understanding models. With language emerging as a key driver towards powerful generalization, it is imperative for foundational video-language models to have a sense of time. In this paper, we consider a specific aspect of temporal understanding: consistency of time order as elicited by before/after relations. We establish that seven existing video-language models struggle to understand even such simple temporal relations. We then question whether it is feasible to equip these foundational models with temporal awareness without re-training them from scratch. Towards this, we propose a temporal adaptation recipe on top of one such model, VideoCLIP, based on post-pretraining on a small amount of video-text data. We conduct a zero-shot evaluation of the adapted models on six datasets for three downstream tasks which require varying degrees of time awareness. We observe encouraging performance gains especially when the task needs higher time awareness. Our work serves as a first step towards probing and instilling a sense of time in existing video-language models without the need for data and compute-intensive training from scratch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 128, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 534, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 701 + ], + "type": "text", + "content": "Self-supervised pretraining at scale on multimodal web corpora tied with powerful architectures [107] has led to foundational models [12] for images [2, 49, 59, 83, 84] and videos [2, 6, 26, 109, 119, 126]. These models have enabled remarkable improvements on a plethora of downstream video-language tasks such as video-text retrieval, video question-answering, and action recognition. Given the cost and difficulty of video annotations, even for a small amount of downstream data, such foundational models are emerging as the de-facto backbone for zero-shot [119, 122, 127] and few-shot generalization [2]. However, it remains unclear if these video-language models capture essential properties of a video beyond what can be learned from static images, most notably: time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Many before us have shown that existing video-language" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 325, + 213, + 375, + 255 + ], + "blocks": [ + { + "bbox": [ + 313, + 230, + 323, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 323, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 323, + 240 + ], + "type": "text", + "content": "1." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 213, + 375, + 255 + ], + "lines": [ + { + "bbox": [ + 325, + 213, + 375, + 255 + ], + "spans": [ + { + "bbox": [ + 325, + 213, + 375, + 255 + ], + "type": "image", + "image_path": "0547fbbf6ef66282f5fcf4e972eacb39c83d3355c7d50989704ec08ef25296db.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 377, + 214, + 429, + 255 + ], + "blocks": [ + { + "bbox": [ + 377, + 214, + 429, + 255 + ], + "lines": [ + { + "bbox": [ + 377, + 214, + 429, + 255 + ], + "spans": [ + { + "bbox": [ + 377, + 214, + 429, + 255 + ], + "type": "image", + "image_path": "5f93c9201089595b0e6f3dc220b735703fcb6f6f7abfd03c99e9340eb033c094.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 431, + 214, + 483, + 255 + ], + "blocks": [ + { + "bbox": [ + 431, + 214, + 483, + 255 + ], + "lines": [ + { + "bbox": [ + 431, + 214, + 483, + 255 + ], + "spans": [ + { + "bbox": [ + 431, + 214, + 483, + 255 + ], + "type": "image", + "image_path": "e2b8ca251bc4b75da58650cb76cfb9473621692e19b795ee3e859abd8df6b267.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 485, + 217, + 537, + 255 + ], + "blocks": [ + { + "bbox": [ + 485, + 217, + 537, + 255 + ], + "lines": [ + { + "bbox": [ + 485, + 217, + 537, + 255 + ], + "spans": [ + { + "bbox": [ + 485, + 217, + 537, + 255 + ], + "type": "image", + "image_path": "448db9de872f0dbabe74cb5e24df2aa24362cab688f77f3bbbb708efca7cfe3e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 324, + 256, + 375, + 296 + ], + "blocks": [ + { + "bbox": [ + 315, + 271, + 323, + 281 + ], + "lines": [ + { + "bbox": [ + 315, + 271, + 323, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 323, + 281 + ], + "type": "text", + "content": "2." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 256, + 375, + 296 + ], + "lines": [ + { + "bbox": [ + 324, + 256, + 375, + 296 + ], + "spans": [ + { + "bbox": [ + 324, + 256, + 375, + 296 + ], + "type": "image", + "image_path": "6b85dd3be0e0a4f8b3dfdb23fc5784ab19f471e1f361b8615af310f2c2f4936c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 377, + 256, + 429, + 296 + ], + "blocks": [ + { + "bbox": [ + 377, + 256, + 429, + 296 + ], + "lines": [ + { + "bbox": [ + 377, + 256, + 429, + 296 + ], + "spans": [ + { + "bbox": [ + 377, + 256, + 429, + 296 + ], + "type": "image", + "image_path": "ed74bac97aba135b1c3be4dd919b1069c3acd91772eb876ab5e20e007483f5c5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 431, + 255, + 483, + 296 + ], + "blocks": [ + { + "bbox": [ + 431, + 255, + 483, + 296 + ], + "lines": [ + { + "bbox": [ + 431, + 255, + 483, + 296 + ], + "spans": [ + { + "bbox": [ + 431, + 255, + 483, + 296 + ], + "type": "image", + "image_path": "baba012e51a4f48e4bfb214edace9dc46cf7c58b5d4e4285d4b9bbf2224f1721.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 485, + 255, + 537, + 296 + ], + "blocks": [ + { + "bbox": [ + 485, + 255, + 537, + 296 + ], + "lines": [ + { + "bbox": [ + 485, + 255, + 537, + 296 + ], + "spans": [ + { + "bbox": [ + 485, + 255, + 537, + 296 + ], + "type": "image", + "image_path": "bea8a049582657baeb8d9eeffda3dc108efd817149214d60d920e479d1cf7e77.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 324, + 296, + 375, + 335 + ], + "blocks": [ + { + "bbox": [ + 314, + 311, + 323, + 320 + ], + "lines": [ + { + "bbox": [ + 314, + 311, + 323, + 320 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 323, + 320 + ], + "type": "text", + "content": "3." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 296, + 375, + 335 + ], + "lines": [ + { + "bbox": [ + 324, + 296, + 375, + 335 + ], + "spans": [ + { + "bbox": [ + 324, + 296, + 375, + 335 + ], + "type": "image", + "image_path": "a1c0853e3f209fca343a837a022dbe64f9b4f2675c37e7ec83d1dcf51ad3ee7c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 377, + 296, + 429, + 335 + ], + "blocks": [ + { + "bbox": [ + 377, + 296, + 429, + 335 + ], + "lines": [ + { + "bbox": [ + 377, + 296, + 429, + 335 + ], + "spans": [ + { + "bbox": [ + 377, + 296, + 429, + 335 + ], + "type": "image", + "image_path": "7573eb9b7773f08653a5476096a0e2536069043db96368106b8de73408b349b2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 431, + 296, + 483, + 335 + ], + "blocks": [ + { + "bbox": [ + 431, + 296, + 483, + 335 + ], + "lines": [ + { + "bbox": [ + 431, + 296, + 483, + 335 + ], + "spans": [ + { + "bbox": [ + 431, + 296, + 483, + 335 + ], + "type": "image", + "image_path": "c4b5f7e459b996314e387c808d1cf50f7c00ef6483813a8b8a8f477f987373ce.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 485, + 296, + 537, + 335 + ], + "blocks": [ + { + "bbox": [ + 485, + 296, + 537, + 335 + ], + "lines": [ + { + "bbox": [ + 485, + 296, + 537, + 335 + ], + "spans": [ + { + "bbox": [ + 485, + 296, + 537, + 335 + ], + "type": "image", + "image_path": "ac9aa9f03961a12e1b69245c1882606afd83c7a3fa18d17c3e0b15a127d7727b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 324, + 335, + 375, + 376 + ], + "blocks": [ + { + "bbox": [ + 313, + 351, + 323, + 360 + ], + "lines": [ + { + "bbox": [ + 313, + 351, + 323, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 323, + 360 + ], + "type": "text", + "content": "4." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 335, + 375, + 376 + ], + "lines": [ + { + "bbox": [ + 324, + 335, + 375, + 376 + ], + "spans": [ + { + "bbox": [ + 324, + 335, + 375, + 376 + ], + "type": "image", + "image_path": "47bf84d000abbcaf973fc77fa37a1151a1b671dabad57d3b2f4babe2bff5250c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 382, + 324, + 394 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 324, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 324, + 394 + ], + "type": "text", + "content": "A." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 403, + 324, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 324, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 324, + 413 + ], + "type": "text", + "content": "C." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 402, + 413, + 419 + ], + "lines": [ + { + "bbox": [ + 325, + 402, + 413, + 419 + ], + "spans": [ + { + "bbox": [ + 325, + 402, + 413, + 419 + ], + "type": "text", + "content": "The baby eats food after it looks into the camera" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 430, + 545, + 463 + ], + "lines": [ + { + "bbox": [ + 305, + 430, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 545, + 463 + ], + "type": "text", + "content": "Figure 1. Can you match the correct video-text pairs? Understanding the time order of events across video and language is necessary to be able to solve this task. See footnote on next page for answers." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 377, + 335, + 429, + 376 + ], + "blocks": [ + { + "bbox": [ + 377, + 335, + 429, + 376 + ], + "lines": [ + { + "bbox": [ + 377, + 335, + 429, + 376 + ], + "spans": [ + { + "bbox": [ + 377, + 335, + 429, + 376 + ], + "type": "image", + "image_path": "e4eb3058bf121d59246e5f101f4b33c6ab5c75f13f1834e381acdf2a44daf585.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 380, + 406, + 399 + ], + "lines": [ + { + "bbox": [ + 325, + 380, + 406, + 399 + ], + "spans": [ + { + "bbox": [ + 325, + 380, + 406, + 399 + ], + "type": "text", + "content": "Dog runs away before it brings a ball to the man" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 431, + 335, + 483, + 376 + ], + "blocks": [ + { + "bbox": [ + 431, + 335, + 483, + 376 + ], + "lines": [ + { + "bbox": [ + 431, + 335, + 483, + 376 + ], + "spans": [ + { + "bbox": [ + 431, + 335, + 483, + 376 + ], + "type": "image", + "image_path": "2f44f942b293c338a0cbded02f28ecc265b49c3e17a2ad80f9722c6b762ffbbe.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 381, + 443, + 394 + ], + "lines": [ + { + "bbox": [ + 432, + 381, + 443, + 394 + ], + "spans": [ + { + "bbox": [ + 432, + 381, + 443, + 394 + ], + "type": "text", + "content": "B." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 381, + 533, + 399 + ], + "lines": [ + { + "bbox": [ + 444, + 381, + 533, + 399 + ], + "spans": [ + { + "bbox": [ + 444, + 381, + 533, + 399 + ], + "type": "text", + "content": "The dog brings a a ball to the man before it runs away" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 407, + 443, + 415 + ], + "lines": [ + { + "bbox": [ + 432, + 407, + 443, + 415 + ], + "spans": [ + { + "bbox": [ + 432, + 407, + 443, + 415 + ], + "type": "text", + "content": "D." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 406, + 522, + 423 + ], + "lines": [ + { + "bbox": [ + 444, + 406, + 522, + 423 + ], + "spans": [ + { + "bbox": [ + 444, + 406, + 522, + 423 + ], + "type": "text", + "content": "The baby looks into the camera after it eats food" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 485, + 335, + 537, + 376 + ], + "blocks": [ + { + "bbox": [ + 485, + 335, + 537, + 376 + ], + "lines": [ + { + "bbox": [ + 485, + 335, + 537, + 376 + ], + "spans": [ + { + "bbox": [ + 485, + 335, + 537, + 376 + ], + "type": "image", + "image_path": "ccc38d16e1c011d3ebf940f0488db10055b7d95938d7918501954b9fedb1a695.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 485, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 665 + ], + "type": "text", + "content": "models [6, 57, 66, 119] can achieve impressive performance on several video benchmarks [22, 41, 120] without reliably encoding time [13, 56, 59]. For example, Buch et al. [13] show that a model that uses a single (carefully selected) frame often outperforms recent video-language models [57, 119] on standard video benchmarks such as MSR-VTT [120]. Lei et al. [56] report similar findings with a single-frame pretraining approach. These findings hint at a lack of time awareness in video models. However, it remains unclear if these findings are caused, indeed, by the lack of time in video models or whether the benchmarks themselves do not mandate time awareness. Furthermore, there is no clear definition of what it means for a model to be time aware. In this paper, we strive to shed light on all these factors of time awareness in video-language models." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "As a first step, we consider a simple notion of understanding time, i.e., understanding temporal relations such as before and after [4]. Consider the task presented in Fig. 1. A time invariant model shall be able to associate (A) with (1)" + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2503" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "or (2) and (B) with (3) or (4) based on static frames alone. But to distinguish between (1) and (2), one needs to be able to understand time order and connect it across video and language1. Thus, the first question we ask in Section 3: do the representations learnt by foundational video-language models encode this sense of time? To reliably attribute lack of time awareness to models and not existing benchmarks, we design our own synthetic dataset to probe models for this sense of time. We create video-language pairs that show a sequence of two events. Then, we alter the order of events either in the text or the video and check if models can connect the order in video and language. We find that existing video-language models indeed struggle to associate the time order across video and language." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 478 + ], + "type": "text", + "content": "In light of these findings, the second question we ask in Section 4 is: can we adapt a video-language model, without expensive re-training from scratch, to instill this sense of time? Towards this, we take inspiration from literature on understanding time in natural language, where there has been much work on developing time aware language models [20, 36, 37, 130, 131]. Our objective is to instill time awareness in a video-language model without having to pretrain from scratch. To do that, we propose TACT: Temporal Adaptation by Consistent Time-ordering based on two key components: (i) we artificially create samples that provide temporal signal, for example, by flipping the order of events in the video or the text, and (ii) we introduce a modified contrastive loss to learn time order consistency based on these samples. Instead of training from scratch, we adapt an existing video-language model, VideoCLIP [61], using the paradigm of post-pretraining on a small amount of videotext data [66, 121]. We demonstrate the effectiveness of TACT in connecting the time order in video and language on four diverse real datasets in Section 5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 479, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 289, + 586 + ], + "type": "text", + "content": "Finally, in line with the original motive of video-language models for zero-shot generalization, we evaluate in Section 6 our TACT-adapted model for three sets of tasks on six downstream datasets which require a varying degree of time awareness. On tasks that need higher time awareness, with the appropriate choice of adaptation dataset, TACT outperforms a strong baseline that is based on post-pretraining on canonical clip-text pairs without consideration of time-order." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 587, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 289, + 696 + ], + "type": "text", + "content": "In summary, our contributions are: (i) We show that existing video-language models struggle to associate time order in video and language through controlled experiments on synthetic data and several evaluations on real datasets. (ii) Based on VideoCLIP [119], we propose TACT, a method for temporal adaptation using this time order consistency without having to pretrain from scratch. (iii) We demonstrate improved zero-shot generalizability of TACT-adapted models on tasks that require higher time awareness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 71, + 481, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 481, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 481, + 85 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 92, + 545, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 92, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 306, + 92, + 545, + 117 + ], + "type": "text", + "content": "We briefly discuss recent advances in video-language models followed by their consideration of time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 547, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 432 + ], + "type": "text", + "content": "Foundational video-language models. Large-scale datasets, self-supervision, and the advent of Transformers [107] have led to the emergence of powerful encoders for images [21,39,103], videos [5,11,24,104,117], language models [19,64,69,86] and even universal encoders [32,46]. These encoders form the basis for several vision-language foundational models. Popular image-language models such as CLIP [83] and ALIGN [49] are trained on massive datasets by using web images and alt-text. Similarly, videolanguage models are catching up and can be categorised into two broad directions: (i) adapting image-language models for videos [8, 23, 50, 51, 63, 66, 71, 110, 112, 121], and (ii) pure video-based models that are learned using large video-text datasets [3,7,27-29,31,58,62,65,67,68,95,119]. Recently, a new paradigm of post-pretraining has emerged where an existing image- or video-language model goes through another stage of self-supervised pretraining on a small amount of video data before it is evaluated on downstream tasks [66, 121]. This is promising as it circumvents the prohibitive cost of pretraining on large datasets from scratch. In [66], the post-pretraining uses time-invariant mean-pooling, while [121] strives to bridge the domain gap between image captions and video subtitles. In contrast, our proposed temporal adaptation involves post-pretraining of VideoCLIP [119] with a small amount of data that instills the model to learn the time-order of events in a video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 437, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 676 + ], + "type": "text", + "content": "Time in vision. Time separates videos from static images or an unordered set of frames. While modeling time remains a challenge, it also presents a natural source of supervision that has been exploited for self-supervised learning. For example, as a proxy signal by posing pretext tasks involving spatio-temporal jigsaw [1, 44, 53], video speed [10, 17, 48, 94, 111, 124], arrow of time [78, 80, 114], frame/clip ordering [25, 70, 90, 97, 118], video continuity [61], or tracking [45, 108, 113]. Several works have also used contrastive learning to obtain spatio-temporal representations by (i) contrasting temporally augmented versions of a clip [47, 77, 81], or (ii) encouraging consistency between local and global temporal contexts [9, 18, 85, 123]. Nevertheless, it remains unclear if the learnt representations actually encode time reliably. Time-aware features have also been explored for specific downstream tasks such as action recognition [30, 100, 101]. There has been some very recent work on evaluating self-supervised video representations [87, 99] on their temporal recognition ability instead of only relying on time as a guidance for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "In the same spirit, a related direction pursues evaluation and benchmarking of time awareness in video datasets [88], models [13, 14, 30, 56, 89, 125] or both [43, 92]. Huang et" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 203, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 203, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 203, + 714 + ], + "type": "text", + "content": "1 Answers: (A)-(2), (B)-(1), (C)-(4), (D)-(3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "2504" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "al. [43] measure the effect of motion on temporal action recognition to find that only a subset of classes in UCF-101 and Kinetics-400 require motion information. Ghodrati et al. [30] propose new tasks to evaluate temporal asymmetry, continuity and causality in video models. Our work derives inspiration from these but applies more generally to video-language models as language provides a basis for open-world generalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 175, + 289, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 380 + ], + "type": "text", + "content": "Time in language. Time has also been extensively studied in the natural language literature. Early works identified temporal structures in language such as temporal prepositions and quantifiers [4, 79]. More recent literature focuses on tasks such as extracting temporal relations [35, 72-74], as well as temporal reasoning [36, 37, 82, 130, 131]. For example, Han et al. [36, 37] and Zhou et al. [131] pretrain language models specifically to focus on understanding temporal relations such as before, after, during, etc. The emergence of large language models has also spurred an increased interest in developing benchmarks to test for time awareness in these models [20, 75, 76, 102, 106, 129]. For example, Ning et al. [75] propose a new benchmark of reading comprehension with questions involving before/after relations. Since temporal relations in language are grounded in the video, we draw inspiration from [36, 37, 131] and aim to instill time awareness in video-language models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 387, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 289, + 555 + ], + "type": "text", + "content": "Time in video-language models appears implicitly in tasks like video-text alignment [38, 98] and temporal grounding [42, 60]. In this work, we focus on self-supervised video-language models that can generalize to a variety of tasks rather than models designed for a specific task, e.g., temporal grounding. Some recent works have shown the under-utilisation of time in classic video-text benchmarks such as MSR-VTT [120], YouCook [132], ActivityNet [22], and DiDeMo [41]. For example, [13, 56, 57] discover that on several benchmarks, using only one or few frames or clips achieves competitive performance. Adaptations of the popular CLIP architecture for videos (e.g., CLIP4Clip [66]) show that weighted pooling of frames already achieves impressive performance on retrieval benchmarks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "These raise some key questions: do existing videolanguage models truly understand time in the sense of correctly associating order of events in language and video? If not, can we adapt them to instill time awareness? Our work addresses these questions. There has been some work in using time-order across video and language as a source of self-supervision. Specifically, concurrent to our work, both Sun et al. [96] and Cao et al. [15] propose fine-grained temporal alignment between video and text as the pretraining objective. Different from these works, we consider the notion of time-order and we aim to adapt a given videolanguage model using post-pretraining, which circumvents the need for a new round of compute-intense pretraining." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 435, + 167 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 435, + 167 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 435, + 167 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 435, + 167 + ], + "type": "image", + "image_path": "d9515c00a7205a5547f167d10e5e105eb89b5511d2539566856bab8d7dfb87f1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 173, + 547, + 219 + ], + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 219 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed task to evaluate time-order consistency across synthetic video-language pairs having before/after relations. We also define a control task to check if the synthetic videos are considered out-of-distribution by the model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 462, + 78, + 529, + 167 + ], + "blocks": [ + { + "bbox": [ + 462, + 78, + 529, + 167 + ], + "lines": [ + { + "bbox": [ + 462, + 78, + 529, + 167 + ], + "spans": [ + { + "bbox": [ + 462, + 78, + 529, + 167 + ], + "type": "image", + "image_path": "11780cf09827d603373cf3543228a8e0569fd9158c10836b95ef3a12274c5a67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 237, + 529, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 237, + 529, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 529, + 251 + ], + "type": "text", + "content": "3. Do Video-Language Models Sense Time?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 258, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 389 + ], + "type": "text", + "content": "Probing video-language models for temporal understanding is an open direction of research. In this work, we consider a specific sense of temporal understanding: consistency in the order of events in a video with the associated textual description. For example, consider a text description: A red circle appears before a yellow circle. This imposes an order constraint on the video stream to have the event red circle appears happen before the event yellow circle appears. Can existing video-language models connect time-order in text with that in video? To answer this, we design an experiment with synthetic data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "text", + "content": "Synthetic dataset. We construct simple videos that comprise of a pair of events such as the ones mentioned above. We generate " + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "inline_equation", + "content": "N = 180" + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "text", + "content": " video-language pairs as a combination of " + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "inline_equation", + "content": "C = 6" + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "text", + "content": " colors, " + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "inline_equation", + "content": "S = 3" + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "text", + "content": " shapes, and " + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "inline_equation", + "content": "|\\tau| = 2" + }, + { + "bbox": [ + 304, + 393, + 545, + 609 + ], + "type": "text", + "content": " temporal relations: before and after. The corresponding caption describes the order of events connected with a before/after temporal relation. We call this caption as an attractor since it is consistent with the time-ordering in the video. Likewise, we construct a distractor where we flip the order of event descriptions while retaining the temporal relation. An example pair is illustrated in Fig. 2 (left). Ideally, a time aware video-language model should be able to associate the video with the temporally consistent text, or vice versa. We refer to this task as time-order consistency check. To rule out the possibility that synthetic videos are out-of-distribution, we also perform the same experiment with canonical clips with a single event and the text describes that same event as shown in Fig. 2 (right). We refer to this as the control task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 613, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 686 + ], + "type": "text", + "content": "Choice of models. We consider recent video-language models, broadly categorized into three groups: (i) image-language models like CLIP [83] that are adapted to videos [23,66,128], (ii) pure video-language models trained on a contrastive learning objective [6,16,119], and (iii) pure video-language models trained on a masking objective [29]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Findings. We evaluate video-to-text and text-to-video retrieval on both time-order consistency and control tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2505" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 289, + 201 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 201 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 201 + ], + "type": "table", + "html": "
ParadigmMethodVideo-to-TextText-to-Video
◎◎◎◎
Chance-50.050.050.050.0
Image-Language adapted to videoCLIP4Clip [66]49.451.150.049.4
CLIP2Video [23]100.047.897.852.3
CenterCLIP [128]91.746.197.251.1
Video-Language ContrastiveVideoCLIP [119]87.151.166.748.3
Frozen in Time [6]97.849.4100.050.6
VindLU [16]98.052.0100.051.1
Video-Language MaskingBridgeFormer [29]100.051.197.242.2
", + "image_path": "9775656550ca29f10ef4c0d57fe5d78f7cdcba4fb278dbcbe2abf3c610dfdf19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 205, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 242 + ], + "type": "text", + "content": "Table 1. Results on synthetic control ( ⌒ ) and time-order consistency ( ⌒ ) task as described in Fig. 2. Existing video-language models show random performance on our time-order task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 347 + ], + "type": "text", + "content": "From Tab. 1, we observe that while most video-language models perform well on the control task, all of them struggle and perform on par with random chance on the temporal task. This gap in performance deserves attention given the importance of time in videos. Note that while synthetic data allows for controlled experiments, we also expand this evaluation to real video datasets in the following section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 274, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 274, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 274, + 374 + ], + "type": "text", + "content": "4. Adaptation by Consistent Time-Ordering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 380, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 489 + ], + "type": "text", + "content": "We describe a post-pretraining recipe to instill a video-language model with a sense of time. We propose TACT: Temporal Adaptation by Consistency of Time-order, that is based on two key components: (i) we artificially create samples that provide temporal signals, e.g., by flipping the order of events; and (ii) we introduce a modified contrastive loss to learn temporal consistency based on these samples. We start by defining the notation and then describe the key components of our adaptation recipe." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": "Preliminaries. Let " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " be the space of video clips and " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " be the space of text clips. Consider two non-overlapping video clips " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "v_{i}, v_{j} \\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\zeta_{i}, \\zeta_{j} \\in \\mathcal{T}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " be their respective captions. Let " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " be a temporal relation, " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\tau \\in \\{\\text{before, after}\\}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ". Then, we denote a stitched and time-order consistent clip as " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "(u_{ij}, t_{ij})" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "u_{ij} := [v_{i}; v_{j}]" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "t_{ij} := [\\zeta_{i}; \\tau; \\zeta_{j}]" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "[\\cdot; \\cdot]" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " denotes concatenation. Note that depending on " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ", the order of " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " may need to change in " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "u_{ij}" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": ". For brevity, we drop the subscripts and refer to the stitched pair as " + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "inline_equation", + "content": "(u, t)" + }, + { + "bbox": [ + 46, + 493, + 287, + 613 + ], + "type": "text", + "content": " unless stated otherwise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Time-order reversal. The classic contrastive learning paradigm for video-language models aligns components of a video clip " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " with its text counterpart " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\zeta_{i}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and contrasts against other texts " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\zeta_{j}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " that usually describe a completely different clip. This makes such models ignore the finer details of temporal understanding as it is easier to contrast the negatives by simply focusing on the objects or the scene. This is evident from simple bag-of-word-like methods that" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 530, + 284 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 530, + 284 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 530, + 284 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 530, + 284 + ], + "type": "image", + "image_path": "166102ea621382b55c43696766ac1217cc91881b037ffa89b9afd5f188d2163d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 292, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 292, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 545, + 358 + ], + "type": "text", + "content": "Figure 3. Overview of TACT. Along with the usual contrastive loss, where negatives come from other samples in the batch, we make use of time-order reversal within the same sample and cross samples to generate additional negatives for both video and text. We also extend the contrastive loss to time-order reversed video and text corresponding to reverse consistency " + }, + { + "bbox": [ + 305, + 292, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 305, + 292, + 545, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "content": "are shown to work well for contrastive learning, both on the visual (e.g., CLIP4Clip [66]) and textual (e.g., MILNCE [67]) modalities. We hypothesize that unless there are negatives in a contrastive setup that contain the same scenes and objects, models need not learn a sense of time. Thus, we present a simple strategy to generate negatives that force the learning process to focus on the temporal order." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "content": "We define a time-order reversal function " + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 305, + 462, + 545, + 498 + ], + "type": "text", + "content": " that operates on the stitched video clip or text description and temporally swaps its constituents:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 505, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 505, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 350, + 505, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\mathbb {T} (u) = \\mathbb {T} ([ v _ {i}; v _ {j} ]) := [ v _ {j}; v _ {i} ], \\quad \\text {a n d} \\tag {1}", + "image_path": "723ca35dcd546e6669c1ef1130ecb8ed462de1ee98508e06cc49715af5908bf7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 519, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 519, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 353, + 519, + 545, + 533 + ], + "type": "interline_equation", + "content": "\\mathbb {T} (t) = \\mathbb {T} ([ \\zeta_ {i}; \\tau ; \\zeta_ {j} ]) := [ \\zeta_ {j}; \\tau ; \\zeta_ {i} ]. \\tag {2}", + "image_path": "8b812a2d5f72c41d23527d5b079211f21cd60a11d314995a80ecfb45e91cdfc5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "content": "An illustration of " + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "content": " is shown in Fig. 3. Note that " + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "content": " does not reverse the actual video, i.e., time does not flow backwards, but only changes the order in which events happen in the stitched clips. Our objective is to train a model that is able to distinguish between the original pair " + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "inline_equation", + "content": "(u,t)" + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "content": " and time-reversed versions " + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "inline_equation", + "content": "(u,\\mathbb{T}(t))" + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "inline_equation", + "content": "(\\mathbb{T}(u),t)" + }, + { + "bbox": [ + 304, + 541, + 545, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Loss function. We assume access to an existing pre-trained video-language model with a visual encoder " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". We obtain the video encoding " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_u \\coloneqq f_{\\theta}(u) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " and the text encoding " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t \\coloneqq g_{\\phi}(t) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". Our goal is to adapt " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\theta, \\phi\\}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " via post-pretraining such that the resulting model is time aware while maintaining its original performance on tasks such as retrieval. As we aim to use a small amount of data, we only update some parameters of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2506" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "content": "the model " + }, + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "inline_equation", + "content": "(\\Theta)" + }, + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "content": ", such as the last few layers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "type": "text", + "content": "We now introduce our recipe for temporal adaptation based on the InfoNCE loss [105] to learn time-order sensitive video-text correspondence. For a positive (or time-order consistent) video-text pair " + }, + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "type": "inline_equation", + "content": "(u, t)" + }, + { + "bbox": [ + 46, + 84, + 286, + 144 + ], + "type": "text", + "content": ", we first define a forward loss where the stitched pair is in its original time-order." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 286, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 286, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\sum_ {(u, t) \\in \\mathcal {B}} \\left(\\operatorname {T N C E} \\left(\\mathbf {z} _ {u}, \\mathbf {z} _ {t}\\right) + \\operatorname {T N C E} \\left(\\mathbf {z} _ {t}, \\mathbf {z} _ {u}\\right)\\right), \\tag {3}", + "image_path": "112260acb2ebd9b73ebc63ad528e2d3bde3908adbb19f1037c6ae211119dc970.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 178, + 286, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 286, + 202 + ], + "type": "text", + "content": "where TNCE is the Noise Contrastive Estimation (NCE) loss for temporal adaptation, defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 286, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 286, + 232 + ], + "type": "interline_equation", + "content": "\\operatorname {T N C E} \\left(\\mathbf {z} _ {u}, \\mathbf {z} _ {t}\\right) := - \\log \\frac {\\exp \\left(\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {t}\\right)}{\\sum_ {t ^ {\\prime} \\in \\mathcal {B} _ {t}} \\exp \\left(\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {t ^ {\\prime}}\\right) + \\mathcal {C} ^ {\\text {t i m e}}}, \\tag {4}", + "image_path": "a0a5ce6a234344fc7dfa930b952f2a3b2be8ef9a14287db4c8aa6cec46c65bdc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "text", + "content": " is the batch of " + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "inline_equation", + "content": "(u,t)" + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "text", + "content": " pairs and " + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_t" + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "text", + "content": " specifically refers to other stitched text captions in the batch. " + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{time}}" + }, + { + "bbox": [ + 47, + 235, + 286, + 270 + ], + "type": "text", + "content": " accumulates negatives defined using time-order reversal as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 274, + 286, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 274, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 51, + 274, + 286, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {C} ^ {\\text {t i m e}} = \\alpha_ {\\text {s a m e}} \\exp (\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {\\mathbb {T} (t)}) + \\alpha_ {\\text {c r o s s}} \\sum_ {t ^ {\\prime} \\in \\mathcal {B} _ {t} \\backslash \\{t \\}} \\exp (\\mathbf {z} _ {u} \\cdot \\mathbf {z} _ {\\mathbb {T} \\left(t ^ {\\prime}\\right)}), \\tag {5}", + "image_path": "4237fb51334b87311ee6dd31b35c58547fa5fdd26606042997303ab4b011277a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{same}}" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " controls the effect of contrasting text from the same sample but with reversed text time-order, i.e., " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\mathbb{T}(t)" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{cross}}" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " encourages the model to contrast between reversed versions of other text captions, i.e., " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\mathbb{T}(t')" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": ". Note that when both " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{same}}" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{cross}}" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " are 0, we revert back to the standard NCE formulation, albeit on stitched pairs. While Eq. (4) corresponds to the video-text loss " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{TNCE}(\\mathbf{z}_u,\\mathbf{z}_t)" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": ", the text-video loss " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{TNCE}(\\mathbf{z}_t,\\mathbf{z}_u)" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " is defined symmetrically. Furthermore, we also apply a reverse loss " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " to bring time-order reversed versions of both the video and the text together. Note that as we consider " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "(u,t)" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " as a positive pair, " + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "inline_equation", + "content": "(\\mathbb{T}(u),\\mathbb{T}(t))" + }, + { + "bbox": [ + 47, + 302, + 286, + 445 + ], + "type": "text", + "content": " also form a positive pair," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 450, + 288, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 288, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\sum_ {\\left(\\mathbb {T} (u), \\mathbb {T} (t)\\right) \\in \\mathcal {B}} \\left(\\operatorname {T N C E} \\left(\\mathbf {z} _ {\\mathbb {T} (u)}, \\mathbf {z} _ {\\mathbb {T} (t)}\\right) + \\operatorname {T N C E} \\left(\\mathbf {z} _ {\\mathbb {T} (t)}, \\mathbf {z} _ {\\mathbb {T} (u)}\\right)\\right). \\tag {6}", + "image_path": "e0a2db3a652a5fdb92db4cb0494946101e256183ea54c270b8c7c3b718682d51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "text", + "content": "Here, the TNCE term operates on time-reversed clips and " + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{time}}" + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "text", + "content": " contrasts " + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "inline_equation", + "content": "(\\mathbb{T}(u),\\mathbb{T}(t))" + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "text", + "content": " with un-reversed text clips in the batch " + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "inline_equation", + "content": "(\\mathbb{T}(u),t)" + }, + { + "bbox": [ + 47, + 487, + 286, + 523 + ], + "type": "text", + "content": ". The overall loss function is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 529, + 286, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 529, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 132, + 529, + 286, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {f} + \\beta \\mathcal {L} _ {r}. \\tag {7}", + "image_path": "5473fa0f9cc340db00ac5bf941415065a3700416f7f3d4843f0a1ae7c306266e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "text", + "content": "Depending on the choice of loss coefficients " + }, + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{same}}, \\alpha_{\\mathrm{cross}}, \\beta \\in \\{0,1\\}" + }, + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "text", + "content": ", we can vary properties of the adapted model. For example, setting " + }, + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{same}} = 1" + }, + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "text", + "content": " encourages high sensitivity to time-order reversal. As we will see empirically, the loss coefficients also provide the flexibility to adapt the model to various datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "We illustrate this temporal extension of the contrastive loss in Fig. 3 (best seen in colour). " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " illustrates the time order reversal function. The top half corresponds to " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_f" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " while the bottom half visualizes " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": ". In particular, the top-left quadrant alone corresponds to the standard contrastive loss on stitched pairs. While the green diagonal terms are positive pairs, the red diagonal terms are the strongest drivers for instilling temporal understanding in the model." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 155 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 155 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 155 + ], + "type": "table", + "html": "
DatasetTrainValidationTestEgoLength
NvNcNvNcNvNc(s)
TEMPO3,90428,4274111,0003961,000X30
ActivityNet7,44095,474453906456912X120
Charades5,26299,9285001,0005001,000X30
Charades-Ego2,679155,3065001,000210420✓31
", + "image_path": "e47de4e73308fcdeb1dccf000a151003764f08b6eee353f09930b6d0d019777f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "lines": [ + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "text", + "content": "Table 2. Statistics of datasets we consider for temporal adaptation. " + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{V}}" + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "text", + "content": " is the number of unique videos and " + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{C}}" + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "text", + "content": " is the number of stitched clips. Based on " + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{V}}" + }, + { + "bbox": [ + 305, + 159, + 545, + 203 + ], + "type": "text", + "content": ", TEMPO and Charades-Ego are smaller as compared to ActivityNet and Charades." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 224, + 477, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 477, + 237 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 477, + 237 + ], + "type": "text", + "content": "5. Experiments: TACT Ablations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "type": "text", + "content": "Base model. We demonstrate the effectiveness of TACT as an adaptation recipe on top of VideoCLIP [119] owing to its simple architecture, contrastive objective, and use of pre-computed S3D [116] features that enable faster experimentation and allow encoding a long temporal context (" + }, + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\sim 32" + }, + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "type": "text", + "content": " secs). We initialize " + }, + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 305, + 247, + 545, + 331 + ], + "type": "text", + "content": " from the model pretrained on HowTo100M [68] and post-pretrain on multiple datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 335, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 552 + ], + "type": "text", + "content": "Datasets. One of our key objectives is to post-pretrain on a small amount of data in contrast to massive pretraining datasets such as WebVid2M [7] or HowTo100M [68]. We consider dense video captioning datasets that offer sufficient diversity in terms of size, backgrounds, clip durations, viewpoints and activities. Specifically, we experiment with: (i) TEMPO [40]: the subset of stitched diverse third-person videos from DiDeMo [41] with text descriptions for fixed 5s segments that contain before/after relations; (ii) ActivityNet Captions [55]: a dense video captioning dataset with human-centric actions; (iii) Charades [93]: a scripted indoor daily human activities video dataset; and (iv) Charades-Ego [91]: similar to Charades, scripted human activities from the egocentric viewpoint. To construct stitched clips, we randomly sample any two non-overlapping clip-text pairs in the video. Since we require stitched clips instead of raw videos, we create new splits for each dataset (see Tab. 2)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate the post-pretrained model using two sets of metrics: (i) standard retrieval metrics, recall " + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "inline_equation", + "content": "R@5" + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "inline_equation", + "content": "R@10" + }, + { + "bbox": [ + 305, + 555, + 545, + 626 + ], + "type": "text", + "content": " and median-rank evaluated on stitched video-text clips; and (ii) time-order consistency, i.e., the fraction of videos for which the model correctly associates text that is time order consistent with the video:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 634, + 545, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 634, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 321, + 634, + 545, + 663 + ], + "type": "interline_equation", + "content": "A _ {\\text {t i m e}} := \\frac {1}{| \\mathcal {D} |} \\sum_ {(u, t) \\in \\mathcal {D}} \\mathbb {I} [ d (\\mathbf {z} _ {u}, \\mathbf {z} _ {t}) < d (\\mathbf {z} _ {u}, \\mathbf {z} _ {\\mathbb {T} (t)}) ], \\tag {8}", + "image_path": "32afabbe6bba60e106c3d381846a7adb40ea0863bf7b37bd87f5647f37e9a598.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "inline_equation", + "content": "(u,t)" + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "text", + "content": " are time-order consistent pairs, " + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "text", + "content": " is the dataset, and " + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 305, + 673, + 544, + 697 + ], + "type": "text", + "content": " is a distance metric based on cosine similarity." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Post-pretraining details. We freeze the word embeddings" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2507" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "and layers 1 to 5 for both the video and text encoders in VideoCLIP. For adaptation, we use the Adam optimizer [54] with learning rate " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "5e^{-6}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", batch size 32 trained on a single node with 4 GeForce GTX 1080 GPUs. On TEMPO, we train for 60 epochs while on the other datasets, we train for 10 epochs and pick the checkpoint that maximizes the geometric mean of " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{time}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " on the respective validation set. A typical adaptation run takes about 1-3 hours." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 174, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 174, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 287, + 247 + ], + "type": "text", + "content": "Evaluation on the test set. Results in Tab. 3 show that TACT\\* with optimal loss coefficients outperforms TACT† (all 0 loss coefficients) and the zero-shot baseline (no postpretraining similar to the synthetic data experiment), both on the retrieval and time-order consistency tasks. This indicates the robustness of the adaptation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": "Impact of loss coefficients. Choosing appropriate values for loss coefficients " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\Theta_l \\coloneqq \\{\\alpha_{\\mathrm{same}}, \\alpha_{\\mathrm{cross}}, \\beta\\}" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " allows the model to learn various aspects and adapt using different datasets. On each dataset, we vary " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\Theta_l \\in \\{0,1\\}^3" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " and find the best configuration based on the GeometricMean " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "(R@1, \\max(A_{\\mathrm{time}} - 50, 0))" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " on the validation sets. The above metric ensures the geometric mean is not overpowered by " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": ". The results are shown in Tab. 4. As " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{same}}" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " is directly responsible for discriminating between original and time-reversed orders, unsurprisingly, setting it to 1 is necessary to achieve the best results on " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{time}}" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " for all the datasets. For TEMPO and Charades-Ego, using all loss components (all 1) provides the best results, whereas " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{cross}} = 1" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " achieves a better trade-off for ActivityNet and Charades. Choosing " + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 46, + 251, + 289, + 432 + ], + "type": "text", + "content": " leads to an improve" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 452, + 279, + 639 + ], + "blocks": [ + { + "bbox": [ + 56, + 452, + 279, + 639 + ], + "lines": [ + { + "bbox": [ + 56, + 452, + 279, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 279, + 639 + ], + "type": "table", + "html": "
DatasetMethodRetrievalTime-order
R@1↑MedR ↓Atime ↑
TEMPOZero-shot3.749.048.1
TACT†7.713.046.5
TACT*9.39.066.5
ActivityNetZero-shot1.144.049.6
TACT†5.834.059.7
TACT*5.835.085.7
CharadesZero-shot1.3170.047.1
TACT†5.338.573.5
TACT*5.735.077.0
Charades-EgoZero-shot1.664.053.7
TACT†6.435.060.1
TACT*10.128.568.2
", + "image_path": "24c1cfe1554551d57cd0fad26ded5e51496f93812b07f6db51ae33274824e12c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "text", + "content": "Table 3. Results for the best TACT model on test sets. " + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "inline_equation", + "content": "\\mathrm{TACT}^{\\star}" + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "text", + "content": " has optimal loss coefficients and " + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "inline_equation", + "content": "\\mathrm{TACT}^{\\dagger}" + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "text", + "content": " is a baseline with all coefficients 0. On time order, TACT generalizes well with " + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "inline_equation", + "content": "\\mathrm{TACT}^{\\star}" + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "text", + "content": " outperforming the baselines. On retrieval, for TEMPO and Charades-Ego, " + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "inline_equation", + "content": "\\mathrm{TACT}^{\\star}" + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "text", + "content": " outperforms the baseline as their optimal models have " + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 46, + 643, + 287, + 710 + ], + "type": "text", + "content": " which helps retrieval with a small amount of data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 543, + 135 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 543, + 135 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 543, + 135 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 543, + 135 + ], + "type": "image", + "image_path": "d0ad7ac203b57eb66af53635addd83ee6d31039c0ecd77dd1f9031cf8dc486c5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "type": "text", + "content": "Figure 4. Time-distance between stitched clips in datasets for temporal adaptation " + }, + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "type": "inline_equation", + "content": "(\\Delta_{\\mathrm{time}})" + }, + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "type": "text", + "content": ". TEMPO has stitched clips close to each other while those in Charades-Ego are farthest apart suggesting a correlation between " + }, + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 139, + 545, + 183 + ], + "type": "text", + "content": " and difficulty of temporal adaptation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "text", + "content": "ment in retrieval performance for TEMPO and Charades-Ego but leads to a decline for ActivityNet and Charades. We attribute this to the number of unique videos in the train set for these datasets. As ActivityNet and Charades have more videos than TEMPO or Charades-Ego (see train " + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{V}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "text", + "content": " Tab. 2) additional positives introduced by setting " + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "text", + "content": " are not necessary and in fact, hurt performance. Finally, we note that carefully setting the value of " + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{same}}" + }, + { + "bbox": [ + 304, + 205, + 545, + 326 + ], + "type": "text", + "content": " provides a convenient trade-off between spatial and temporal understanding. Please refer to the supplement for detailed experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": "What makes temporal adaptation hard? We observe a large gap in " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " between TEMPO and ActivityNet. We hypothesize that the distance (in seconds) between the two clips (" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": ") in a stitched clip is strongly correlated with the difficulty of adaptation. Intuitively, it is easier to infer the time-order consistency for a stitched clip " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " with text " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " that has distant constituent clips " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "v_i, v_j" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " since the objects and scene could be vastly different. In contrast, it is harder to discern the correct time-order when the constituent clips are closer in time. Fig. 4 shows distribution of " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " for each dataset. Indeed, " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " in ActivityNet (58.8s) is much higher than that in TEMPO (6.4s) making the task harder on TEMPO. To further test our hypothesis, we conduct a controlled experiment where we gradually vary the distribution of " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " for Charades-Ego to match it to that of TEMPO. We find a strong correlation (" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\rho = 0.92" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": ") between " + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 329, + 546, + 534 + ], + "type": "text", + "content": " and hardness of adaptation. More details are in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 544, + 518, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 518, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 518, + 558 + ], + "type": "text", + "content": "6. Experiments: Downstream Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 565, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 613 + ], + "type": "text", + "content": "The goal of video-language foundation models is to generalize in a zero- or few-shot manner to a diverse range of downstream tasks. We evaluate TACT models on three sets of downstream tasks that need low-to-high time awareness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Baseline: Standard post-pretraining. Comparing our temporally adapted models with pretrained VideoCLIP is not fair since adapted models see data beyond the pretraining phase. In addition to the zero-shot comparison, we compare against a baseline model that is trained for standard video-text retrieval on the same datasets as temporal adaptation. Instead of using stitched clips, we use simple canonical pairs, i.e., " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "(v_{i},\\zeta_{i})" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "(u_{ij},t_{ij})" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2508" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 213 + ], + "type": "table", + "html": "
Loss coefficientsTEMPOActivityNetCharadesCharades-Ego
α_sameα_crossβR@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑R@1↑MedR↓Atime↑
Chance0.1500.050.00.1453.050.00.1500.050.00.1500.050.0
0008.314.049.46.430.057.35.742.071.52.944.064.6
0018.214.049.55.627.047.04.258.075.13.241.565.2
0108.215.049.36.129.078.85.245.078.93.438.064.5
0118.114.049.55.827.048.34.258.075.13.141.067.0
1006.420.060.65.928.079.16.138.076.33.242.066.1
1016.524.062.95.626.063.14.951.078.03.339.070.7
1105.924.059.76.029.086.36.643.077.83.740.567.9
1117.517.062.55.727.063.85.151.077.73.838.568.3
", + "image_path": "c1792805d00aeba217706a514c757e950fdf5b95ffba5f542c038e12ef2191af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 261, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "type": "text", + "content": "Table 4. Impact of loss coefficients for TACT post-pretraining on validation sets of various datasets. For each dataset, the corresponding color-marked row denotes the best configuration based on the geometric mean of " + }, + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{time}}" + }, + { + "bbox": [ + 46, + 218, + 547, + 251 + ], + "type": "text", + "content": ". TACT is able to connect time-order in video and language while maintaining its retrieval capabilities across several datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 261, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 47, + 261, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 545, + 453 + ], + "type": "table", + "html": "
AdaptationI. Text-to-Video RetrievalII. Temporal VQAIII. Action-to-Video Retrieval
DatasetMethodMSR-VTTYouCookIINext-QA (ATP)AGQASSv2SSv2 (events)Temporal
R@1R@5R@10R@1R@5R@10AccuracyAccuracymAPmAPmAP
-Chance0.10.51.00.10.51.020.050.00.62.02.0
-None10.623.429.918.245.559.923.449.93.46.413.0
TEMPOBaseline12.029.337.321.548.261.825.050.83.96.815.9
TACT12.826.535.720.445.158.727.657.14.27.716.1
ActivityNetBaseline15.734.444.915.638.851.423.750.73.77.016.0
TACT13.829.639.616.036.949.825.450.33.57.216.2
CharadesBaseline12.325.833.621.548.661.726.050.54.17.113.7
TACT11.725.232.422.449.162.425.254.84.37.814.1
Charades-EgoBaseline13.127.534.519.447.160.824.349.74.06.914.7
TACT11.124.630.621.948.261.925.658.43.96.913.5
", + "image_path": "88c28340b4223d9fd0b9ac0aa171790f630340f36804ad87921a0ade34088e3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 456, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 46, + 456, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 547, + 502 + ], + "type": "text", + "content": "Table 5. Results on downstream zero-shot evaluation with tasks requiring increasing time awareness from I to III. None corresponds to direct evaluation of the VideoCLIP model on the downstream dataset. Green denotes an improvement for the TACT adapted model w.r.t. the baseline, red denotes a deterioration. As we move from tasks that need low to high time awareness, the effectiveness of TACT increases. See Sec. 6 for a more detailed discussion. The table is best viewed on screen in colour." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "text", + "content": "Evaluating TACT adapted models on synthetic data. On the video-to-text variant, TACT adapted on TEMPO achieves " + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "inline_equation", + "content": "64.4\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "text", + "content": ", ActivityNet " + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "inline_equation", + "content": "52.5\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "text", + "content": ", Charades " + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "inline_equation", + "content": "65.0\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "text", + "content": ", Charades-Ego " + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "inline_equation", + "content": "85.6\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 606 + ], + "type": "text", + "content": ". This is usually higher than the performance that non-adapted models achieve in Tab. 1. This highlights that TACT models learn useful signals to match time-order in video and language." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "I. Text-to-video retrieval. We consider two widely used benchmarks: MSR-VTT [120] and YouCookII [132] and adopt standard retrieval metrics. Recent work has identified a bias for spatial understanding in these datasets [8, 13, 43, 56, 59, 66]. Thus, we consider this class of tasks as requiring low time awareness. As shown in Tab. 5 set I, on MSR-VTT [120], we observe that TACT is worse (marked in red) or at par with the baseline across datasets. This" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 582 + ], + "type": "text", + "content": "aligns well with findings in [13, 56] that these benchmarks do not need time awareness. On YouCookII [132], TACT models based on Charades(-Ego) outperform the baseline (marked in green). We believe this is a consequence of a lower domain shift between YouCookII and Charades." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 586, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 696 + ], + "type": "text", + "content": "II. Temporal video QA. Next, we use subsets of recently released multiple-choice video question answering benchmarks: Next-QA [115] and AGQA [34]. The idea is to check if we can probe models for temporal understanding by asking questions with temporal language. Buch et al. [13] identify a subset of Next-QA, dubbed as ATP-hard², with a higher concentration of temporally challenging data. For AGQA, we pick a subset of " + }, + { + "bbox": [ + 304, + 586, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\sim 6k" + }, + { + "bbox": [ + 304, + 586, + 547, + 696 + ], + "type": "text", + "content": " questions that explicitly have a question with before/after relation. We consider" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 534, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 534, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 534, + 714 + ], + "type": "text", + "content": "2Available here: github.com/StanfordVL/atp-video-language" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2509" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "these benchmarks as requiring a moderate-high level of time awareness and AGQA in particular is also close to our adaptation task. We use accuracy as the standard metric. We observe (see Tab. 5 set II) that indeed TACT almost always outperforms (marked in green) baselines on both Next-QA and AGQA. TEMPO-adapted TACT seems to generalize particularly well on both benchmarks. Likewise, Charades-adapted TACT does well on AGQA since AGQA is also based on the Charades videos accounting for reduced domain shift. We affirm that temporal adaptation is useful, especially when the downstream tasks require it." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "text", + "content": "III. Action-to-video retrieval. Finally, we consider action recognition benchmarks such as Something-Something (SSv2) [33] and Temporal [88]. SSv2 was designed to capture richer temporal information [33, 56]. We follow Lei et al. [56], who propose the template-retrieval task that encourages temporal modelling and use their evaluation split3 containing " + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "inline_equation", + "content": "C = 174" + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "text", + "content": " actions and " + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "inline_equation", + "content": "K = 12" + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "text", + "content": " videos per class. Interestingly, different actions in SSv2 require differing levels of time awareness. We create a subset SSv2 (events) with " + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "inline_equation", + "content": "C_{\\text{events}} = 49" + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "text", + "content": " actions that have at least two verbs in the label as the occurrence of multiple verbs is indicative of multiple events occurring in sequence. Finally, we also evaluate against the Temporal benchmark [88], a combination of 50 action classes from SSv2 [33] and Kinetics-400 [52] for which temporal information is deemed to be essential for recognition. Similar to text-to-video retrieval, we use the action class as a text query and obtain a ranking over all videos. Different from the retrieval setup, since a single query has multiple correct answers (up to " + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "inline_equation", + "content": "K = 12" + }, + { + "bbox": [ + 46, + 207, + 289, + 616 + ], + "type": "text", + "content": " videos), we report mAP as the metric for these benchmarks. This task set needs high time awareness. Furthermore, unlike QA tasks in II, there is a shift in several (uncontrolled) factors as we move from temporal adaptation task to these tasks. From Tab. 5, we observe that TEMPO and Charades-adapted models generalize well across set III benchmarks. ActivityNet-adapted TACT underperforms on SSv2 but outperforms the baseline on strongly temporal actions in SSv2 (events) and Temporal. Finally, TACT adapted on Charades-Ego is at-par or slightly worse than the baseline on SSv2 variants, and also on Temporal, perhaps due to the shift from egocentric to third-person videos. Overall, despite SSv2 and Temporal requiring high time awareness, TACT models show promising zero-shot generalization with the right choice of the adaptation dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 624, + 200, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 200, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 200, + 637 + ], + "type": "text", + "content": "7. Discussion and Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 648, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 648, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 288, + 696 + ], + "type": "text", + "content": "Generalization to other temporal prompts. The time-order of events in language can be described using various sentence structures. Although we train video-language models using before/after relations, it is natural to ask if" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 325, + 77, + 528, + 198 + ], + "blocks": [ + { + "bbox": [ + 325, + 77, + 528, + 198 + ], + "lines": [ + { + "bbox": [ + 325, + 77, + 528, + 198 + ], + "spans": [ + { + "bbox": [ + 325, + 77, + 528, + 198 + ], + "type": "image", + "image_path": "61536b1950e6d38904c1f501b46a3b61bd72c9c087aac9deb41cbf0b23d8ebd8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 202, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 236 + ], + "type": "text", + "content": "Figure 5. Models trained by TACT with before/after relations generalize to a new kind of prompt such as First, ..., then ... indicating learning of the underlying true time-order of events." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": "the model still correctly associates time-order for a different prompt such as First, ..., then, ... To systematically test this, we gather event pairs " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_1" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_2" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_1" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": " occurs before " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_2" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": " in the video) for each sample in the validation set and stitch them using three prompts as follows: (i) " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_1" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": " before " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_2" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": ", (ii) " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_2" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_1" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": ", (iii) First " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_1" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "inline_equation", + "content": "E_2" + }, + { + "bbox": [ + 304, + 257, + 547, + 365 + ], + "type": "text", + "content": ". As shown in Fig. 5, TACT-adapted models generalize well to a new prompt (iii). This substantiates the learning of time-order of events rather than merely learning the order of words in the sentence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 514 + ], + "type": "text", + "content": "Limitations. While we present a promising way of instilling time in video-language models, our work is limited to the VideoCLIP [119] pretrained model. Our initial experiments with Frozen in Time [6] were not as promising, perhaps because it uses a much shorter temporal context (4 frames). Please see the supplement for results on more pretrained models. Furthermore, we consider a specific definition of time awareness derived from temporal relations like before/after. It is natural to ask if this can be extended to more general notions of temporality, e.g., as defined by Allen [4]. Finally, there can always be more downstream tasks considered such as (spatio-)temporal localization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 517, + 548, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 548, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 548, + 662 + ], + "type": "text", + "content": "Conclusion. Given the essence of time in video-language models, we present a simple experiment based on synthetic data to test for time awareness. We find that existing models lack a sense of time defined in terms of consistency of order of events in video and language. To fill this gap, building upon VideoCLIP [119], we present TACT, a recipe to instill this sense of time in video-language models. Finally, we analyze the zero-shot generalizability of TACT-adapted models to a diverse set of tasks. We hope that this work provokes further probing and instilling time awareness in video-language models, and also inspires other adaptations of foundational models to solve various challenging tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. We acknowledge support from ELLIS Amsterdam and the AMS Scholarship to Piyush. We thank Dr Dennis Koelma for their help with compute infrastructure and Dr Hazel Doughty for useful discussions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 240, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 240, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 240, + 713 + ], + "type": "text", + "content": "3Available here: github.com/jayleicn/singularity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2510" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Winter Conference on Applications of Computer Vision (WACV), pages 179-189. IEEE, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 288, + 257 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 259, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 287, + 313 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-supervised multimodal versatile networks. ArXiv, abs/2006.16228, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 316, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 316, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 57, + 316, + 287, + 337 + ], + "type": "text", + "content": "[4] James F. Allen. Towards a general theory of action and time. In Artificial Intelligence, 1984. 1, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 339, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 339, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 57, + 339, + 288, + 382 + ], + "type": "text", + "content": "[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 384, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 384, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 57, + 384, + 287, + 428 + ], + "type": "text", + "content": "[6] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 430, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 430, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 57, + 430, + 287, + 474 + ], + "type": "text", + "content": "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 475, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 287, + 507 + ], + "type": "text", + "content": "[8] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 510, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 287, + 563 + ], + "type": "text", + "content": "[9] Nadine Behrmann, Mohsen Fayyaz, Juergen Gall, and Mehdi Noroozi. Long short view feature decomposition via contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9224-9233, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 567, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 287, + 620 + ], + "type": "text", + "content": "[10] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9922–9931, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 622, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 622, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 622, + 287, + 666 + ], + "type": "text", + "content": "[11] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 667, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 667, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 667, + 287, + 689 + ], + "type": "text", + "content": "[12] Rishi Bommasani et al. On the opportunities and risks of foundation models. ArXiv, abs/2108.07258, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Shyamal Buch, Cristobal Eyzaguirre, Adrien Gaidon, Jia-jun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 546, + 106 + ], + "type": "text", + "content": "the \"Video\" in Video-Language Understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "text", + "content": "[14] Petr Byvshev, Pascal Mettes, and Yu Xiao. Are 3d convolutional networks inherently biased towards appearance? Computer Vision and Image Understanding, 220:103437, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 152, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 546, + 196 + ], + "type": "text", + "content": "[15] Meng Cao, Tianyu Yang, Junwu Weng, Can Zhang, Jue Wang, and Yuexian Zou. Locvtp: Video-text pre-training for temporal localization. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 198, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 547, + 241 + ], + "type": "text", + "content": "[16] Feng Cheng, Xizi Wang, Jie Lei, David Crandall, Mohit Bansal, and Gedas Bertasius. Vindlu: A recipe for effective video-and-language pretraining. arXiv preprint arXiv:2212.05051, 2022. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 243, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 546, + 285 + ], + "type": "text", + "content": "[17] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised spatio-temporal representation learning using variable playback speed prediction. IEEE Access, 9:79562-79571, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 288, + 546, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 288, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 288, + 546, + 331 + ], + "type": "text", + "content": "[18] Ishan Rajendra Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TcIr: Temporal contrastive learning for video representation. Computer Vision and Image Understanding (CVIU), 219:103406, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 333, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 333, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 333, + 547, + 409 + ], + "type": "text", + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 411, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 546, + 465 + ], + "type": "text", + "content": "[20] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 467, + 546, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 546, + 543 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 544, + 546, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 546, + 600 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 546, + 600 + ], + "type": "text", + "content": "[22] Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 961-970, 2015. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 601, + 546, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 601, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 312, + 601, + 546, + 634 + ], + "type": "text", + "content": "[23] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 635, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 546, + 678 + ], + "type": "text", + "content": "[24] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 680, + 546, + 714 + ], + "type": "text", + "content": "[25] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Conference on Computer" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2511" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition (CVPR), pages 3636-3645, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "type": "text", + "content": "[26] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 286, + 184 + ], + "type": "text", + "content": "[27] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. Violet: End-to-end video-language transformers with masked visual-token modeling. ArXiv, abs/2111.12681, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 286, + 228 + ], + "type": "text", + "content": "[28] Valentin Gabeur, Chen Sun, Alahari Karteek, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 286, + 285 + ], + "type": "text", + "content": "[29] Yuying Ge, Yixiao Ge, Xihui Liu, Dian Li, Ying Shan, Xiaohu Qie, and Ping Luo. Bridging video-text retrieval with multiple choice questions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 16167-16176, June 2022. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 287, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 287, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 287, + 286, + 319 + ], + "type": "text", + "content": "[30] Amir Ghodrati, Efstratios Gavves, and Cees G. M. Snoek. Video time: Properties, encoders and evaluation. In British Machine Vision Conference (BMVC), 2018. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 322, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 286, + 364 + ], + "type": "text", + "content": "[31] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. Coot: Cooperative hierarchical transformer for video-text representation learning. ArXiv, abs/2011.00597, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 366, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 366, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 286, + 420 + ], + "type": "text", + "content": "[32] Rohit Girdhar, Mannat Singh, Nikhil Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A single model for many visual modalities. Conference on Computer Vision and Pattern Recognition (CVPR), pages 16081-16091, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 422, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 286, + 498 + ], + "type": "text", + "content": "[33] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In International Conference on Computer Vision (ICCV), pages 5842-5850, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 501, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 286, + 544 + ], + "type": "text", + "content": "[34] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 545, + 286, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 286, + 589 + ], + "type": "text", + "content": "[35] Rujun Han, Qiang Ning, and Nanyun Peng. Joint event and temporal relation extraction with shared representations and structured prediction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 590, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 286, + 622 + ], + "type": "text", + "content": "[36] Rujun Han, Xiang Ren, and Nanyun Peng. Deer: A data efficient language model for event temporal reasoning. ArXiv, abs/2012.15283, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 624, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 286, + 667 + ], + "type": "text", + "content": "[37] Rujun Han, Xiang Ren, and Nanyun Peng. Econet: Effective continual pretraining of language models for event temporal reasoning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "type": "text", + "content": "[38] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2906, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In _Conference on Computer Vision and Pattern Recognition (CVPR), CVPR '16, pages 770–778, June 2016._2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 171 + ], + "type": "text", + "content": "[40] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), pages 1380-1390, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 173, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 225 + ], + "type": "text", + "content": "[41] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with natural language. International Conference on Computer Vision (ICCV), pages 5804-5813, 2017. 1, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 228, + 545, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 280 + ], + "type": "text", + "content": "[42] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan C. Russell. Localizing moments in video with temporal language. In Empirical Methods in Natural Language Processing (EMNLP), 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 282, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 282, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 545, + 357 + ], + "type": "text", + "content": "[43] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7366–7375, 2018. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 360, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 360, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 312, + 360, + 545, + 425 + ], + "type": "text", + "content": "[44] Yuqi Huo, Mingyu Ding, Haoyu Lu, Zhiwu Lu, Tao Xiang, Ji-Rong Wen, Ziyuan Huang, Jianwen Jiang, Shiwei Zhang, Mingqian Tang, Songfang Huang, and Ping Luo. Self-supervised video representation learning with constrained spatiotemporal jigsaw. In International Joint Conference on Artificial Intelligence (IJCAI), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 426, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 457 + ], + "type": "text", + "content": "[45] A. Jabri, Andrew Owens, and Alexei A. Efros. Spacetime correspondence as a contrastive random walk. ArXiv, abs/2006.14613, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 502 + ], + "type": "text", + "content": "[46] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver: General perception with iterative attention. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 545, + 536 + ], + "type": "text", + "content": "[47] S. Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. International Conference on Computer Vision (ICCV), pages 9950-9960, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 580 + ], + "type": "text", + "content": "[48] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), pages 425-442, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 582, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 545, + 645 + ], + "type": "text", + "content": "[49] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 679 + ], + "type": "text", + "content": "[50] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 712 + ], + "type": "text", + "content": "[51] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. ArXiv, abs/2112.04478, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2512" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 128, + 288, + 172 + ], + "type": "text", + "content": "[53] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Association for the Advancement of Artificial Intelligence (AAAI), pages 8545-8552, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "type": "text", + "content": "[54] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "text", + "content": "[55] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-Captioning Events in Videos. In International Conference on Computer Vision (ICCV), 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "type": "text", + "content": "[56] Jie Lei, Tamara L. Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 338 + ], + "type": "text", + "content": "[57] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. Conference on Computer Vision and Pattern Recognition (CVPR), pages 7327-7337, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 338, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 382 + ], + "type": "text", + "content": "[58] Gen Li, Nan Duan, Yuejian Fang, Daxin Jiang, and Ming Zhou. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Association for the Advancement of Artificial Intelligence (AAAI), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 288, + 437 + ], + "type": "text", + "content": "[59] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 437, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 503 + ], + "type": "text", + "content": "[60] Juncheng Li, Junlin Xie, Long Qian, Linchao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal grounding with structured variational cross-graph correspondence learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 3022-3031, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 503, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 288, + 548 + ], + "type": "text", + "content": "[61] Hanwen Liang, Niamul Quader, Zhixiang Chi, Lizhe Chen, Peng Dai, Juwei Lu, and Yang Wang. Self-supervised spatiotemporal representation learning by exploiting video continuity. ArXiv, abs/2112.05883, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 548, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 613 + ], + "type": "text", + "content": "[62] Kevin Lin, Alex Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z. Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, Chengfei Cai, Hongfa Wang, Dima Damen, Bernard Ghanem, Wei Liu, and Mike Zheng Shou. Ego-centric video-language pretraining. ArXiv, abs/2206.01670, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 613, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 288, + 658 + ], + "type": "text", + "content": "[63] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard de Melo, Xiaogang Wang, Jifeng Dai, Y. Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. ArXiv, abs/2208.03550, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 712 + ], + "type": "text", + "content": "[64] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ArXiv, abs/1907.11692, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 72, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 127 + ], + "type": "text", + "content": "[65] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 128, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 128, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 128, + 547, + 172 + ], + "type": "text", + "content": "[66] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 172, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 172, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 547, + 228 + ], + "type": "text", + "content": "[67] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. Conference on Computer Vision and Pattern Recognition (CVPR), pages 9876-9886, 2020. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 228, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 547, + 293 + ], + "type": "text", + "content": "[68] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In International Conference on Computer Vision (ICCV), 2019. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 293, + 547, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 547, + 338 + ], + "type": "text", + "content": "[69] Tomas Mikolov, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In International Conference on Learning Representations (ICLR), 2013. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 338, + 547, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 547, + 382 + ], + "type": "text", + "content": "[70] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), pages 527-544, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 382, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 382, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 382, + 547, + 437 + ], + "type": "text", + "content": "[71] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 437, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 547, + 480 + ], + "type": "text", + "content": "[72] Qiang Ning, Zhili Feng, and Dan Roth. A structured learning approach to temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 481, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 547, + 514 + ], + "type": "text", + "content": "[73] Qiang Ning, Zhili Feng, Hao Wu, and Dan Roth. Joint reasoning for temporal and causal relations. In Association of Computational Linguistics (ACL), 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 514, + 547, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 547, + 558 + ], + "type": "text", + "content": "[74] Qiang Ning, Sanjay Subramanian, and Dan Roth. An improved neural baseline for temporal relation extraction. In Empirical Methods in Natural Language Processing (EMNLP), 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 558, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 547, + 602 + ], + "type": "text", + "content": "[75] Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. Torque: A reading comprehension dataset of temporal ordering questions. ArXiv, abs/2005.00242, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 602, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 602, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 602, + 547, + 635 + ], + "type": "text", + "content": "[76] Qiang Ning, Ben Zhou, Zhili Feng, Haoruo Peng, and Dan Roth. Cogcomptime: A tool for understanding time in natural language. ArXiv, abs/1906.04940, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 635, + 547, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 691 + ], + "type": "text", + "content": "[77] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. Conference on Computer Vision and Pattern Recognition (CVPR), pages 11200-11209, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 547, + 714 + ], + "type": "text", + "content": "[78] Lyndsey C. Pickup, Zheng Pan, Donglai Wei, YiChang Shih, Changshui Zhang, Andrew Zisserman, Bernhard" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2513" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "text", + "content": "Scholkopf, and William T. Freeman. Seeing the arrow of time. In *Conference on Computer Vision and Pattern Recognition* (CVPR), pages 2043-2050, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 287, + 140 + ], + "type": "text", + "content": "[79] Ianthe Pratt and Nissim Francez. Temporal prepositions and temporal generalized quantifiers. Linguistics and Philosophy, 24:187-222, 2001. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 54, + 141, + 287, + 184 + ], + "type": "text", + "content": "[80] Will Price and Dima Damen. Retro-actions: Learning 'close' by time-reversing 'open' videos. International Conference on Computer Vision Workshops (ICCVW), pages 1371-1380, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 186, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 186, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 54, + 186, + 287, + 241 + ], + "type": "text", + "content": "[81] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, H. Wang, Serge J. Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6960-6970, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 287, + 285 + ], + "type": "text", + "content": "[82] Lianhui Qin, Aditya Gupta, Shyam Upadhyay, Luheng He, Yejin Choi, and Manaal Faruqui. Timedial: Temporal commonsense reasoning in dialog. In Association of Computational Linguistics (ACL), 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 287, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 287, + 353 + ], + "type": "text", + "content": "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 287, + 397 + ], + "type": "text", + "content": "[84] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 287, + 475 + ], + "type": "text", + "content": "[85] Adrià Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Patraucean, Florent Altch'e, Michael Valko, Jean-Bastien Grill, Aäron van den Oord, and Andrew Zisserman. Broaden your views for self-supervised video learning. International Conference on Computer Vision (ICCV), pages 1235–1245, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 287, + 510 + ], + "type": "text", + "content": "[86] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 544 + ], + "type": "text", + "content": "[87] Madeline Chantry Schiappa, Yogesh Singh Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ArXiv, abs/2207.00419, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 287, + 588 + ], + "type": "text", + "content": "[88] Laura Sevilla-Lara, Shengxin Zha, Zhicheng Yan, Vedanuj Goswami, Matt Feiszli, and Lorenzo Torresani. Only time can tell: Discovering temporal data for temporal modeling. CoRR, abs/1907.08340, 2019. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 644 + ], + "type": "text", + "content": "[89] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? International Conference on Computer Vision (ICCV), pages 9641-9649, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "text", + "content": "[90] Vivek Sharma, Makarand Tapaswi, and Rainer Stiefelhagen. Deep Multimodal Feature Encoding for Video Ordering. In International Conference on Computer Vision Workshops (ICCVW), 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "text", + "content": "[91] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Actor and observer: Joint" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 106 + ], + "type": "text", + "content": "modeling of first and third-person videos. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 7396-7404, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "type": "text", + "content": "[92] Gunnar A. Sigurdsson, Olga Russakovsky, and Abhinav Kumar Gupta. What actions are needed for understanding human actions in videos? 2017 IEEE International Conference on Computer Vision (ICCV), pages 2156-2165, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 165, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 165, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 312, + 165, + 545, + 220 + ], + "type": "text", + "content": "[93] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, Cham, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 222, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 222, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 222, + 545, + 287 + ], + "type": "text", + "content": "[94] Ankit Singh, Omprakash Chakraborty, Ashutosh Varshney, Rameswar Panda, Rogério Schmidt Feris, Kate Saenko, and Abir Das. Semi-supervised action recognition with temporal contrastive learning. Conference on Computer Vision and Pattern Recognition (CVPR), pages 10384-10394, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 290, + 545, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 290, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 312, + 290, + 545, + 343 + ], + "type": "text", + "content": "[95] Chen Sun, Austin Myers, Carl Vondrick, Kevin P. Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. International Conference on Computer Vision (ICCV), pages 7463-7472, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 347, + 545, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 347, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 312, + 347, + 545, + 401 + ], + "type": "text", + "content": "[96] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 458 + ], + "type": "text", + "content": "[97] Tomoyuki Suzuki, Takahiro Itazuri, Kensho Hara, and Hirokatsu Kataoka. Learning spatiotemporal 3d convolution with video order self-supervision. In European Conference on Computer Vision Workshops (ECCV), pages 0–0, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 460, + 545, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 493 + ], + "type": "text", + "content": "[98] Makarand Tapaswi, Martin Bauml, and Rainer Stiefelhagen. Book2movie: Aligning video scenes with book chapters. In CVPR, June 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 496, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 496, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 312, + 496, + 545, + 540 + ], + "type": "text", + "content": "[99] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees G. M. Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "type": "text", + "content": "[100] Fida Mohammad Thoker, Hazel Doughty, and Cees Snoek. Tubelet-contrastive self-supervision for video-efficient generalization, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "text", + "content": "[101] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Skeleton-contrastive 3d action representation learning. ACM MM, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[102] Shivin Thukral, Kunal Kukreja, and Christian Kavouras. Probing language models for understanding of temporal expressions. In BLACKBOXNLP, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[103] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. arXiv preprint arXiv:2105.01601, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2514" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[104] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. Conference on Computer Vision and Pattern Recognition (CVPR), pages 6450-6459, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "text", + "content": "[105] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. ArXiv, abs/1807.03748, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 204 + ], + "type": "text", + "content": "[106] Siddharth Vashishtha, Adam Poliak, Yash Kumar Lal, Benjamin Van Durme, and Aaron Steven White. Temporal reasoning in natural language inference. In FINDINGS, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "text", + "content": "[107] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "type": "text", + "content": "[108] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2563-2572, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "text", + "content": "[109] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 425 + ], + "type": "text", + "content": "[110] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnil: One foundation model for image-language and video-language tasks. ArXiv, abs/2209.07526, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[111] Jiangliu Wang, Jianbo Jiao, and Yun-Hui Liu. Self-supervised video representation learning by pace prediction. In European Conference on Computer Vision (ECCV), pages 504-521, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[112] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 569 + ], + "type": "text", + "content": "[113] X. Wang, A. Jabri, and Alexei A. Efros. Learning correspondence from the cycle-consistency of time. Conference on Computer Vision and Pattern Recognition (CVPR), pages 2561-2571, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "text", + "content": "[114] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 8052-8060, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[115] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 9777–9786, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[116] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In European Conference on Computer Vision (ECCV), pages 318–335, Cham, 2018. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[117] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin P. Murphy. Rethinking spatiotemporal feature learning for video understanding. ArXiv, abs/1712.04851, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "type": "text", + "content": "[118] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10334-10343, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 546, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 240 + ], + "type": "text", + "content": "[119] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing (EMNLP), pages 6787-6800, 2021. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "type": "text", + "content": "[120] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 546, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 546, + 331 + ], + "type": "text", + "content": "[121] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Rui Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. ArXiv, abs/2209.06430, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 546, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 546, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 546, + 375 + ], + "type": "text", + "content": "[122] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. arXiv preprint arXiv:2206.08155, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 376, + 546, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 546, + 421 + ], + "type": "text", + "content": "[123] Xinyu Yang, Majid Mirmehdi, and Tilo Burghardt. Back to the future: Cycle encoding prediction for self-supervised contrastive video representation learning. ArXiv, abs/2010.07217, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 546, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 546, + 476 + ], + "type": "text", + "content": "[124] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 6548-6557, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 546, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 522 + ], + "type": "text", + "content": "[125] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. Clevrer: Collision events for video representation and reasoning. ArXiv, abs/1910.01442, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 523, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 546, + 611 + ], + "type": "text", + "content": "[126] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel C. F. Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. ArXiv, abs/2111.11432, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 613, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 679 + ], + "type": "text", + "content": "[127] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 680, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 712 + ], + "type": "text", + "content": "[128] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. In SIGIR, 2022. 3, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2515" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 241 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[129] Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. \"going on a vacation\" takes longer than \"going for a walk\": A study of temporal commonsense understanding. *ArXiv*, abs/1909.03065, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 149 + ], + "type": "text", + "content": "[130] Ben Zhou, Qiang Ning, Daniel Khashabi, and Dan Roth. Temporal common sense acquisition with minimal supervision. ArXiv, abs/2005.04304, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[131] Ben Zhou, Kyle Richardson, Qiang Ning, Tushar Khot, Ashish Sabharwal, and Dan Roth. Temporal reasoning on implicit events from distant supervision. ArXiv, abs/2010.12753, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 241 + ], + "type": "text", + "content": "[132] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence (AAAI), pages 7590-7598, 2018. 3, 7" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "2516" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_content_list.json b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..72b46637d7c3693af4adfc530696af08a4e0e713 --- /dev/null +++ b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_content_list.json @@ -0,0 +1,1577 @@ +[ + { + "type": "text", + "text": "TexPose: Neural Texture Learning for Self-Supervised 6D Object Pose Estimation", + "text_level": 1, + "bbox": [ + 203, + 150, + 766, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanzhi Chen $^{1}$ Fabian Manhardt $^{2}$ Nassir Navab $^{1}$ Benjamin Busam $^{1,3}$", + "bbox": [ + 176, + 219, + 791, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich $^{2}$ Google Inc. $^{3}$ 3Dwe.ai", + "bbox": [ + 226, + 237, + 737, + 254 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hanzhi.chen@tum.de fabianmanhardt@google.com b.busam@tum.de", + "bbox": [ + 196, + 257, + 764, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 226, + 302, + 307, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce neural texture learning for 6D object pose estimation from synthetic data and a few unlabelled real images. Our major contribution is a novel learning scheme which removes the drawbacks of previous works, namely the strong dependency on co-modalities or additional refinement. These have been previously necessary to provide training signals for convergence. We formulate such a scheme as two sub-optimisation problems on texture learning and pose learning. We separately learn to predict realistic texture of objects from real image collections and learn pose estimation from pixel-perfect synthetic data. Combining these two capabilities allows then to synthesise photorealistic novel views to supervise the pose estimator with accurate geometry. To alleviate pose noise and segmentation imperfection present during the texture learning phase, we propose a surfel-based adversarial training loss together with texture regularisation from synthetic data. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen scenes. Remarkably, our scheme improves the adopted pose estimators substantially even when initialised with much inferior performance.", + "bbox": [ + 65, + 334, + 472, + 663 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 65, + 694, + 200, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For spatial interaction with objects, one needs an understanding of the translation and rotation of targets within 3D space. Inferring these 6D object pose parameters from a single RGB image is a core task for 3D computer vision. This visually retrieved information has a wide range of applications in AR/VR [13, 33], autonomous driving [12, 30, 57], and robotic manipulation [24, 51, 53]. Noteworthy, accuracy and runtime have both recently made a huge leap forward thanks to deep learning [19, 23, 31, 39, 40]. Unfortunately, most of these methods heavily rely on a massive amount of labelled data for supervision to learn precise models with", + "bbox": [ + 63, + 720, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strong generalisation capabilities [16, 39, 53, 59]. However, it is very labor-intensive and time consuming to generate accurate annotations for pose data [15, 53]. Meanwhile, this process also easily suffers from labelling errors as precise annotation in 3D is highly challenging [14]. Therefore, most of the benchmarks have only few hundreds images, which does not allow proper learning of large models. In fact, methods training with such low amount of real data tend to strongly overfit to the data domain and fail to generalise to new scenes [18].", + "bbox": [ + 497, + 304, + 904, + 447 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a consequence, many different approaches have been proposed in the literature to tackle this problem. The simplest solution is to employ cut-and-paste strategy to increase domain invariance [11,26]. Nonetheless, this requires highly accurate manual annotations and most models tend to overfit to the original domain. Another alternative is to rely on a large amount of synthetically generated data to prevent overfitting. Though this process is relatively cheap and fast, rendered images can exhibit drastic visual discrepancy in comparison with real images even when advanced physically-based renderers [9] are used. A handful of approaches try to close the domain gap via the use of generative adversarial networks to translate the synthetic images into the real domain [2,58]. Unfortunately, these methods do not achieve promising results as the generated images are still easily distinguishable from real imagery. Notably, very recently a new line of work that proposes to self-supervise the pose estimator on real data has emerged. After training in simulation they fine-tune on the new datasets [46, 47]. While these methods achieve impressive results that are even on par with fully supervised methods, they still suffer from significant performance drop when not leveraging additional supervisory signals such as depth data [46] or ground truth camera poses [25, 42]. In this work, we propose a novel way to conduct self-supervised 6D pose estimation that is free from any additional supervision sources and further yields state-of-the-art performance.", + "bbox": [ + 497, + 448, + 904, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The core idea of previous attempts [44,46,47,54] to adapt pretrained pose estimators to the real domain is to conduct render-and-compare using a differentiable renderer. With", + "bbox": [ + 497, + 834, + 904, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 119, + 36, + 189, + 73 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 35, + 800, + 76 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4841", + "bbox": [ + 468, + 891, + 500, + 903 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CAD models and poses given, a renderer can output several attributes of the current object pose (e.g. mask, colour, depth) which allows to refine the pose through iterative comparison between the renderings and observations. However, when relying on 2D visual contents (mask and colour), such strategies tend to fail due to measured silhouette imperfection, lack of textures, and domain discrepancies. This undesirable behaviour is also discussed in [47].", + "bbox": [ + 63, + 114, + 468, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, different from the previous attempts that heavily rely on render-and-compare for self-supervision [46, 47], we instead propose to regard realistic textures of the objects as an intermediate representation before conducting training for the pose estimator. Our approach is formulated as two interconnected sub-optimisation problems on texture learning and pose learning. In the core, we first learn realistic textures of objects from raw image collections, then synthesise training data to supervise the pose estimator with pixel-perfect labels and realistic appearance. The key challenge of our proposed scheme lies in capturing accurate texture under noise introduced by poses initialised by a pretrained pose estimator during supervision. To this end, in addition to leveraging synthetic data to establish geometry priors, we learn robust supervision through adversarial training by conditioning synthesised colours on local surface information. Furthermore, we establish regularisation from synthetic textures to compensate segmentation artefacts during a texture-learning phase. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen domains. Our method significantly improves even difficult objects with little variance in appearance and geometry through self-supervision. Impressively, Our approach demonstrates a robust self-improving ability for the employed pose estimators even when initialised with much inferior pose estimates than stronger baselines [46].", + "bbox": [ + 63, + 231, + 472, + 629 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarise, our main contributions are:", + "bbox": [ + 84, + 631, + 374, + 644 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We formulate a new learning scheme, TexPose, that decomposes self-supervision for 6D object pose into Texture learning and Pose learning.", + "- We propose a surfel-conditioned adversarial training loss and a synthetic texture regularisation term to handle pose errors and segmentation imperfection during texture learning, further delivering self-improving ability to the pose estimators.", + "- We show significant improvements over recent strong baselines with additional supervision signals. Our pose estimators demonstrate a substantial generalisation ability even on unseen scenes." + ], + "bbox": [ + 82, + 656, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 112, + 645, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model-based 6D Pose Estimation To retrieve the 6D pose, early methods use local or global features and search for key points correspondence on CAD models [1,6,7,28]. In recent years, learning-based methods dominate the field and solve the task using convolution neural networks (CNN) under the supervision from annotated data to extract deep features. There are two major approaches for pose estimation, in particular, correspondence-based and regression/classification-based approaches. Correspondence-based methods establish 2D-3D correspondences [37, 39, 40, 42, 59], prior to leveraging a variant of the RANSAC&PnP paradigm to solve for pose. Regression-based approaches, on the other hand, directly regress or classify the pose of the detected object. Initially these methods usually have shown a lower performance due to the existence of ambiguities [29] such as pose symmetries [53]. Therefore, methods like SSD-6D [19] discretise the rotation space to circumvent this issue. Recently, with better continuous representations for rotation [60], these methods gradually demonstrate high effectiveness [10, 48].", + "bbox": [ + 497, + 138, + 904, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Pose Learning Considering tremendous effort to collect large amount of annotations for 6D object pose [17, 50], several recent works have been proposed to explore the possibility of self-supervised pose learning using labeled synthetic data together with unlabelled real sensor data. [8] designed a novel labelling pipeline using a manipulator to generate reliable pose annotations for supervision. Self6D [47] proposed a self-supervision workflow by first pretraining a pose estimator with synthetic data, which was then adapted to the real world through a render-and-compare strategy by imposing consistencies between the rendered depth and sensed depth under the current pose estimate. CPS++ [32] used similar approach for categorical-level object pose estimation, and the shape is jointly deformed during optimisation. AAE [45] parameterise $SO(3)$ space using a latent embedding obtained from synthetic images, which is later employed for orientation retrieval. Both works demonstrate the strong requirement for depth data in order to allow for a reasonable performance either in training or testing. However, due to the uninformative appearance and the sim-to-real domain gap, render-and-compare strategies easily diverge when there is no depth data available as shown in [47]. Hence, consecutive works introduce other supervisory signals to prune the need for depth. DSC-PoseNet [54] employs a key point consistency regularisation for dual-scale images with labelled 2D bounding box. Socket al. [44] use photometric consistency from multiple views to refine the raw estimate, while ground-truth masks are required for supervision. Recently, Self6D++ [46] proposed to initialise the render-and-compare process with a powerful deep pose refiner [26] to guide the learning process of the pose esti", + "bbox": [ + 497, + 435, + 904, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4842", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mator, which currently yields state-of-the-art performance among all methods without manual annotations. Chen et al. [4] leverages a tailored heuristics for pseudo labelling under student-teacher learning scheme.", + "bbox": [ + 63, + 115, + 468, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Novel View Synthesis With a few images given, novel view synthesis aims to render scene content from unseen viewpoints, which can be regarded as a process to acquire \"textures\" of the scene. NOL [38] is proposed to address the difficulty of CAD model texturing for 6D pose estimation. Though requiring pose annotations, it demonstrates high potential to gain improvement for pose estimators with more synthesised real data in a semi-supervised fashion. Outside the scope of 6D pose, neural scene representation as implicit scene models recently get significant attention due to their ability for photorealistic novel view synthesis. NeRF [36] represent scenes using simple multi-layer perceptrons to encode the density and radiance for every 3D point. Assuming precise poses to transform views into a common reference frame, NeRF is able to capture volumes using an image reconstruction loss. NeRF-w [34] further extends NeRF to handle changing lighting and transient effects for image collections with latent embeddings. To release the rigid requirement of accurate poses, NeRF-- [52] jointly optimises camera parameters and radiance fields. iNeRF [56] performs a render-and-compare strategy using a trained NeRF for pose refinement. In contrast, BARF [27] demonstrates promising results when encountering imperfect pose initialisation by bundle-adjusting NeRF and camera poses with coarse-to-fine positional encoding scheme. GNeRF [35] formulates the learning of radiance fields and coarse pose estimates as a generative modelling process and then jointly refines both using photometric information. Recently, NeRF-Supervision [55] and NeRF-Pose [25] are proposed to leverage NeRF to learn dense descriptors or 3D representations of a given object under pose annotations when CAD model is inaccessible.", + "bbox": [ + 63, + 188, + 472, + 629 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 65, + 638, + 159, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to conduct per-instance 6D pose learning from monocular colour images with no manual labelling effort, namely self-supervised 6D pose estimation originally defined in [47]. Since the CAD model is known as a prior, we follow the two-stage approach of previous works [44, 46, 47, 54, 56], which first pretrain a pose estimator with labelled synthetic data, and consequently conduct self-training with unannotated real images captured from a specific real scene demanding a more precise deployment (e.g. robotic bin picking). We mainly introduce the self-training process in this section. Notably, in this tuning process, pose estimator tends to suffer from forgetting problem due to little variance in illumination and backgrounds of the fed training samples. In contrast, We will show the strength of our scheme to alleviate this issue in Section 4. After pretraining we acquire", + "bbox": [ + 63, + 663, + 468, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a pose estimator $f_{\\theta}$ able to initialise a coarse pose estimate $T$ and a segmentation mask $M$ for each real image $\\hat{I}$ . $M$ is usually used as a pseudo silhouette cue to refine the pose $T$ with render-and-compare strategy [5,22,47,56].", + "bbox": [ + 497, + 115, + 904, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first revisit the commonly used strategy of previous methods in this task, then introduce our new formulation requiring no additional supervisory sources clearly advantageous over previous attempts.", + "bbox": [ + 497, + 172, + 904, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting Self-Supervised 6D Pose Learning", + "text_level": 1, + "bbox": [ + 497, + 236, + 889, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Render-and-compare is a common approach for 6D pose self-supervision [44, 46, 47, 54, 56]. As shown in Eqn. 1, it utilises a differentiable renderer $\\mathcal{R}$ fed with an estimated pose to output a rendered mask, image, $[M_r, I_r] = \\mathcal{R}(T)$ . $E_{rend}$ enforces consistency between the rendered outputs $[M_r, I_r]$ and the observations $[M, \\hat{I}]$ . It aims to refine the pseudo pose labels for each training sample, and further supervise pose estimator $f_{\\theta}$ under $E_{pose}$ . Previous attempts are majorly focused on improving $E_{rend}$ to acquire more accurate pseudo pose annotations for better supervision.", + "bbox": [ + 497, + 258, + 904, + 401 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta , \\{T _ {i}, \\dots , T _ {N} \\}} \\sum_ {i} E _ {r e n d} (\\mathcal {R} (T _ {i}), [ M _ {i}, \\hat {I} _ {i} ]) \\\\ + \\sum_ {i} ^ {\\iota} E _ {p o s e} \\left(f _ {\\theta} \\left(\\hat {I} _ {i}\\right), T _ {i}\\right). \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 409, + 902, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When relying on 2D information $[M_r, I_r]$ , minimising $E_{rend}$ to improve $\\{T_1, \\ldots, T_N\\}$ can be extremely difficult due to uninformative shape and appearance and sim-to-real visual discrepancy (e.g., Duck in Fig. 1). To compensate for the lack of information, related methods introduce additional supervision with varying difficulty levels for the acquisition (e.g., depth [47], weak labels [44, 54], deep pose refiner [46], etc.). Deep pose refiner could be acquired with the least effort as it only requires synthetic pretraining as well. However, it can introduce a strong upper bound as illustrated in [46]. The pose refiner can even lead to divergence of the pose estimator and hence lacks robustness for geometrically and photometrically challenging objects.", + "bbox": [ + 497, + 478, + 904, + 663 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. TexPose: Learning Pose from Texture", + "text_level": 1, + "bbox": [ + 499, + 671, + 835, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the aforementioned limitations, we reformulate self-supervised 6D pose learning as two alternating optimisation problems for a texture learner $h_{\\xi}$ and a pose estimator $f_{\\theta}$ as defined in Eqn. 21 and depicted in Fig. 1 (Left).", + "bbox": [ + 497, + 693, + 904, + 751 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta , \\xi} \\sum_ {i} E _ {t e x} \\left(h _ {\\xi} \\circ f _ {\\theta} (\\hat {I} _ {i}), \\hat {I} _ {i}\\right) \\\\ + \\sum_ {j} ^ {i} E _ {\\text {p o s e}} \\left(f _ {\\theta} \\circ h _ {\\xi} \\left(\\hat {T} _ {j}\\right), \\hat {T} _ {j}\\right). \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 579, + 760, + 900, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of leveraging raw images as regression targets for poses, we insert object textures as an intermediate rep", + "bbox": [ + 499, + 828, + 904, + 858 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "${}^{1}$ Green indicates weight update and orange indicates weight freezing.", + "bbox": [ + 517, + 863, + 894, + 877 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4843", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/46b6b04ace8a19027146050f0825ef9a44b5b78d1e9e11de5b072c158d41798a.jpg", + "image_caption": [ + "Figure 1. Left: Schematic overview of TexPose. We fully explore the benefit of both synthetic and real dataset with perfect geometric labels and realistic appearance to synthesise informative train dataset for pose learning. Right: Details of the texture learner used in our scheme. We first conduct geometric pretraining to acquire geometry branch $h_{\\sigma}$ to ensure consistent coordinates system and geometry aligned with the CAD model. Then, we transfer and freeze $h_{\\sigma}$ to the real together with newly established radiance branch $h_c^s$ and $h_c^t$ as our texture learner. It is supervised by $E_{rec}$ , $E_{reg}$ and $E_{adv}$ capable of improving texture quality so as to synthesise more informative dataset for supervision." + ], + "image_footnote": [], + "bbox": [ + 110, + 119, + 440, + 305 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/481aa02d56454925e4a47a8c7d1ef8cd8a90ed729954fe0ef9032816adc10126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 445, + 119, + 855, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resentation before pose learning. Specifically, we start with optimising $E_{tex}$ by fixing the parameters of pose estimator $f_{\\theta}$ used to initialise poses $\\{T_1,\\dots ,T_N\\}$ for each raw image $\\{\\hat{I}_1,\\ldots ,\\hat{I}_N\\}$ , aiming to provide supervision for the realistic textures embedded by $h_\\xi$ . Afterwards, we freeze $h_\\xi$ to synthesise training samples $\\{I_1,\\dots ,I_{N'}\\}$ paired with perfect labels $\\{\\hat{T}_1,\\dots ,\\hat{T}_{N'}\\}$ . Gradients will thus be propagated through $f_{\\theta}$ to update its parameters during the minimisation for $E_{pose}$ over these train data. We typically chose $N' > N$ .", + "bbox": [ + 63, + 414, + 468, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this reformulation, we deliberately omit pose optimisation from 2D information usually sensitive to initialisation and prone to divergence, and bridge the domain gap directly through photorealistic synthesis. Our goal is then shifted to limit texture mapping error under the presence of pose noise so that the synthesised data can be highly informative with perfect geometric labels and as-accurate-as-possible realistic appearance. Intuitively, through multi-view supervision, it's possible to capture more accurate appearance with compensation for per-view pose noise. Our proposed strategy to further increase reliability of the synthesised data will be discussed later. Notably, different from previous attempts, no additional supervision is required in our design. The pose estimator is equipped with a self-improving ability through intermediate phase for texture learning.", + "bbox": [ + 63, + 542, + 470, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF preliminaries. We leverage neural radiance fields (NeRF) [36] to embed the texture information due to its simplicity and capability of photorealistic view synthesis2.", + "bbox": [ + 63, + 771, + 467, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF represents the object model with two MLPs to encode information for every queried point $\\mathbf{x}$ , namely the", + "bbox": [ + 63, + 815, + 467, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "geometry branch to predict volume density $\\sigma$ and geometric feature $\\gamma$ with $[\\sigma, \\gamma] = h_{\\sigma}(\\mathbf{x})$ and radiance branch for colour $\\mathbf{c}$ prediction with $\\mathbf{c} = h_c(\\mathbf{x}, \\gamma)$ (weight $\\xi$ is omitted). When rendering from a given viewpoint, $K$ samples $\\{x_k\\}_1^K$ are gathered along the ray $\\mathbf{r} = \\mathbf{o} + t\\mathbf{d}$ where $\\mathbf{o}, \\mathbf{d} \\in \\mathbb{R}^3$ represent the camera centre and the view direction respectively and $t \\in \\mathbb{R}$ is the distance. The colour $\\mathbf{C}(\\mathbf{r})$ will be estimated through numerical quadrature approximation with $w_k = \\exp(-\\delta_k \\sigma_k)$ and $\\alpha_k = \\prod_{j=1}^{k-1} w_j$ with $\\delta_k$ being the step size between adjacent points, and $\\mathbf{C}(\\mathbf{r}) = \\sum_{k=1}^{K} \\alpha_k (1 - w_k) \\mathbf{c_k}$ . Rendering for depth $D(\\mathbf{r})$ and mask $M(\\mathbf{r})$ is achieved by modifying integration element (see supp. mat. for details).", + "bbox": [ + 499, + 414, + 904, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric pretraining. Though geometry and radiance branch can be trained jointly as shown in NeRF and its variants [21,34,36], the pose for supervision in our scenario is highly noisy, and the coordinate system of NeRF can end up as an arbitrary one and further cause erroneous labels for synthesised data. We hence opt to pretrain geometry branch $h_{\\sigma}$ using pixel-perfect data rendered by the CAD model with accurate pose $\\hat{T}$ , depth $\\hat{D}$ and mask $\\hat{M}$ to supervise NeRF following Eqn. 4, where $E_{d}$ is scale-invariant loss from [20], $\\lambda_{m}$ and $\\lambda_{d}$ are weighting factors.", + "bbox": [ + 499, + 606, + 904, + 749 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {2 d} (\\mathbf {r}) = \\left\\| \\mathbf {C} (\\mathbf {r}) - \\hat {\\mathbf {C}} (\\mathbf {r}) \\right\\| _ {2} ^ {2} + \\lambda_ {m} \\left\\| M (\\mathbf {r}) - \\hat {M} (\\mathbf {r}) \\right\\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 768, + 902, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {p r e} (\\mathbf {r}) = E _ {2 d} (\\mathbf {r}) + \\lambda_ {d} E _ {d} (D (\\mathbf {r}), \\hat {D} (\\mathbf {r})). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 559, + 812, + 902, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, though our current goal is to establish a geometry branch with precise density output \"aligned\" with the CAD model, training for radiance branch is also included, though", + "bbox": [ + 499, + 834, + 904, + 877 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2Radiance, texture and appearance are discussed as interchangeable concepts in our paper.", + "bbox": [ + 63, + 851, + 467, + 877 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4844", + "bbox": [ + 468, + 891, + 500, + 903 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "it will be discarded afterwards. We find this greatly helps to extract meaningful geometric features fed into the radiance branch, and further eases the effort for realistic appearance transfer in the later texture learning stage. We quantitatively detail the necessity of this stage in supp. mat.", + "bbox": [ + 63, + 114, + 468, + 187 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Texture learning from noisy poses. Our current goal is to learn texture encoded by the radiance branch $h_c$ using raw images paired with poses initialised by the pose estimator $f_{\\theta}$ . Note in this stage, we only re-optimise parameters from $h_c$ with $h_{\\sigma}$ being fixed to capture accurate texture and meanwhile ensure no shift of coordinate system and geometry. Though jointly optimising textures and poses through photometric consistency has a potential to improve both aspects as shown in [21, 27, 56], we find it can easily fail when dealing with the textureless objects. Alternatively, we seek to supervise the texture learner with higher pose error tolerance leveraging stronger radiance branch and robust loss.", + "bbox": [ + 63, + 205, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from the vanilla NeRF used in geometric pretraining, we now employ two radiance branches, $h_c^s$ and $h_c^t$ , to translate the extracted geometric features into static radiance (accurate underlying textures) and transient radiance (texture outliers caused by pose errors) respectively. Apart from geometric features $\\gamma$ , we further condition it on view directions $\\mathbf{d}$ , lighting embedding $\\tau^a$ and transient embedding $\\tau^t$ [34] (see Fig. 1 Right). For each view, we use $\\tau^a$ to capture illumination variations and $\\tau^t$ to produce transient effect caused by pose noise. Patch-based training strategy [41] is used to preserve the spatial structures for robust loss supervision. An image reconstruction loss $E_{rec}$ is incorporated with uncertainty awareness to learn textures robustly. We refer to the supp. mat. for the details of the radiance branches and uncertainty-aware $E_{rec}$ .", + "bbox": [ + 63, + 376, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We noticed that supervision from $E_{rec}$ is still not adequate to capture accurate texture under pose noise. We further design a surfel-conditioned adversarial loss for sampled patches to tolerate texture shifts caused by pose errors. For each location on the object surface, we compute normalised object coordinate $S_{x}$ [49] and normal $S_{n}$ with current pose estimate $T$ to represent its local information. By conditioning the predicted patch colour $P$ on its local surface information $[S_{x}, S_{n}]$ , we employ an adversarial scheme:", + "bbox": [ + 63, + 589, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E _ {a d v} (P) = \\mathbb {E} _ {[ S _ {x}, S _ {n} ], P} (1 - \\log D ([ S _ {x}, S _ {n} ], P)) \\tag {5} \\\\ + \\mathbb {E} _ {[ S _ {x}, S _ {n} ], \\hat {P}} (\\log D ([ S _ {x}, S _ {n} ], M \\hat {P})). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 730, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also observed the imperfect predicted mask $M$ can cause background colour artefacts within the object. We therefore establish a regularisation from synthetic cues to compensate the boundary imperfections by generating synthetic patch $P_{syn}$ and mask $M_{syn}$ with current pose estimate $T$ , then applying erosion on the predicted mask $M$ to compute a padded mask with $M_{pad} = M_{syn}(1 - M)$ . The foreground", + "bbox": [ + 63, + 776, + 468, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "boundaries of the real images for supervision are therefore padded with the synthetic colours. This regularisation $E_{feat}$ is realised by minimising the MSE error of deep features extracted by a pretrained VGG19 network [43] to align the boundary features of the predicted patch $P$ with the synthetic patch $P_{syn}$ in a more abstract level. The second part of Eqn. 6 ensures that rendered patch region within the mask $M$ are not affected by the synthetic information.", + "bbox": [ + 497, + 114, + 904, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E _ {r e g} (P) = E _ {f e a t} \\left(P, M \\hat {P} + M _ {p a d} P _ {s y n}\\right) \\tag {6} \\\\ + \\lambda_ {f g} E _ {\\text {f e a t}} (M P + (1 - M) \\hat {P}, \\hat {P}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 239, + 900, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, the loss function $E_{tex}$ for the texture learner supervision is summarised as:", + "bbox": [ + 497, + 285, + 904, + 314 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {t e x} = E _ {r e c} + \\lambda_ {a d v} E _ {a d v} + \\lambda_ {r e g} E _ {r e g}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 569, + 326, + 900, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{adv}$ and $\\lambda_{reg}$ are weighting factors.", + "bbox": [ + 499, + 350, + 793, + 365 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose learning. The entire pipeline runs by first optimising $E_{tex}$ , and subsequent minimisation of $E_{pose}$ with the synthesised dataset as depicted in Fig. 1 (Left). As our scheme operates at data level and thus method-agnostic, $E_{pose}$ can be generalised across methods with different loss designs, e.g., direct pose or correspondence regression. Interestingly, we observed that the pose estimator can converge instantly with only one optimisation loop on $E_{tex}$ and $E_{pose}$ , while adding more loops, i.e., continue to optimise $E_{tex}$ with improved poses after $E_{pose}$ , only brings insignificant improvements. We attribute this to the pixel-perfect supervision from texture learner and its strong ability to mitigate pose errors.", + "bbox": [ + 497, + 381, + 904, + 552 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 499, + 564, + 635, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we first introduce the employed metrics and dataset before presenting our results for the task of 6D pose estimation, then provide detailed analysis of our proposed scheme in ablation study.", + "bbox": [ + 497, + 587, + 902, + 644 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 499, + 653, + 689, + 668 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In line with previous work, we present our results using the ADD(-S) metric, as proposed in [14]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. If the average distance is below $10\\%$ of the object diameter, the estimated pose will be considered correct. For symmetric objects, ADD score is modified to measure the distance to the nearest vertices, here referred to as ADD-S.", + "bbox": [ + 497, + 675, + 904, + 802 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Dataset", + "text_level": 1, + "bbox": [ + 499, + 812, + 596, + 826 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic BOP PBR. We pretrain our models using synthetic data from [9, 16], employing physically-based rendering (PBR) techniques. Although PBR can capture more", + "bbox": [ + 497, + 833, + 904, + 877 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4845", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/177bbbc8c78199e5df482a9ac1b0c2e08b0019eed85e60a21923fcf01dbb737e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 154, + 112, + 284, + 201 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fe9c29f3748510ec1d002e5500d9f1f6e988638dfd82c0f05f7755fc702f3c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 112, + 416, + 201 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7651a735986fa4e359e8a7e88ee5b665da02f1747dcbc86059d31231992a4451.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 112, + 551, + 201 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b8fa52066ebefc90d28ec1532337a763a35649d36376512213c664ac881c2d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 554, + 112, + 685, + 201 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df10ff94f8b2898c3b1c8438d586ef901187e9786d3a669d1614f4af5501b7cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 690, + 112, + 818, + 201 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/799cfbe78b9509ebbc881fda5832ef8212c7933cca2caaf17238c47e580bb3dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 154, + 204, + 284, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/013f191936a71f8586a7a541ca96034d2103e224b9063b692b272005ca7f10a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 204, + 416, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/507d4882d86a864f6af4a6ab2563c14743dd8941fe913d3f1485a2cc97d0b0c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 204, + 551, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cff58455bebe7043dfd341ab15c378f9300358d933553041d035ec68dc20c3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 554, + 204, + 685, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d50e06d3f4e19b83656249dfea0707f890f99414ef688097b7f55518a19d0ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 690, + 204, + 818, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5f7d70aab0d442004ec9992473263371d6dfff30755ae082ad4b672cb6ccfd9.jpg", + "image_caption": [ + "Figure 2. Qualitative results on 5 objects from LineMOD. Top: Renderings of the reconstructed CAD models which are part of LineMOD. Middle: Same objects rendered with our learnt textures. Bottom: Demonstration of our pose estimation quality. Thereby, the green, red, and blue boxes depict ground-truth pose as well as estimated pose before and after our self-supervision, respectively." + ], + "image_footnote": [], + "bbox": [ + 154, + 299, + 284, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5587bfdd5289abb4a6481f785d6863ecdf779df6898e1c51a3ab75641d7c430b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 299, + 418, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/329d0a5f5526d413c97073e5bcd88bd2c77ac2421491d16fcbf49d0dceefa47c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 299, + 551, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d30778e7ea971753d7effb6abc9e27da7a6b14e8f8c4ba5a57206b4899c92d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 554, + 299, + 685, + 387 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a136a2b687596b3587a7739b0a52cb2a913997ba4b3cceb0f2c4760b40565d33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 690, + 299, + 818, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "realistic lighting conditions, there still exhibits a large sim-to-real domain gap. Such discrepancies can strongly degrade the pose estimation performance when being deployed in the real world.", + "bbox": [ + 63, + 445, + 470, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LineMOD. [14] consists of 13 objects with watertight CAD models. Thereby, each object is accompanied by a small image sequence of around 1k frames, captured in cluttered environment with challenging lighting. Following previous works [40, 47], we employ $15\\%$ of the captured RGB images for self-supervision, discarding any manual annotations, and test our model on the remaining samples.", + "bbox": [ + 63, + 521, + 468, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Occluded LineMOD. [3] is an extension of one LineMOD sequence showing eight different objects from LineMOD undergoing severe occlusion. For fairness we adopt the same train/test split for self-supervision as [46].", + "bbox": [ + 63, + 640, + 468, + 697 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HomebrewedDB. [18] is a recently established dataset with various kinds of objects and poses. To test generalisability we test on the three objects that it shares with LineMOD [18].", + "bbox": [ + 63, + 716, + 468, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Performance Analysis", + "text_level": 1, + "bbox": [ + 63, + 783, + 275, + 798 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on LineMOD and Occluded LineMOD In Table 1, we demonstrate our results for LineMOD with respect to the ADD(-S) metric. Note that GDR-LB and GDR-UB correspond to the performance of our base pose estimators, respectively supervised with synthetic data and real", + "bbox": [ + 63, + 805, + 468, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data, reflecting the lower and upper performance bound. As one can easily see there is a significant performance gap between fully-supervised methods and approaches trained with synthetic data instead (c.f. DPODv2 [42], GDR [48]). Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities to achieve comparable results, as for example Self6D++ [46] with depth supervision. In contrast, our method is able to further close the domain gap whilst relying solely on RGB images, outperforming all other self-supervised approaches by at least $3\\%$ for RGB-D approaches and $6\\%$ for RGB-only methods. Noticeably, we are able to improve our lower bound performance from $77.4\\%$ to $91.7\\%$ , which is even slightly better than the performance of our upper bound model trained with full supervision.", + "bbox": [ + 497, + 445, + 904, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We would like to emphasise that our model is capable of providing significant performance boosts on highly challenging tiny and textureless objects (e.g., Ape, Duck, and Holep.) by improving the lower bound from $50.9\\%$ , $24.6\\%$ , $32.1\\%$ to $80.9\\%$ , $83.4\\%$ , $79.3\\%$ , respectively. These objects are typically hard to handle by other self-supervised methods without real supervision (c.f. results from Self6D++ [46], Sock et al. [44], etc). As the RGB variant of Self6D++ [46] leverages DeepIM to refine the output pose in order to generate pseudo pose annotations, this process imposes a strong upper bound on the pose estimator. In fact, the performance achieved by Self6D++ is clearly inferior to DeepIM for almost all objects. This problem is even more apparent when the pose refiner is worse than the pose estimator (c.f. Holep. results from GDR [48], DeepIM [26] and Self6D++ [46]).", + "bbox": [ + 497, + 663, + 905, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4846", + "bbox": [ + 468, + 891, + 502, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1328f0e9d7f3dc632cc0ec7719f2b7f5d618b255d206b67bb26e9ccdd3257dd5.jpg", + "table_caption": [ + "Table 1. Evaluation results on LineMOD dataset. \\*: objects with symmetry. \\*\\*: used as our pretrained pose estimator. $\\star$ : use no PBR renderings for synthetic pretraining. †: re-implemented version from [46], supervision source for Self6D++. ‡: with depth supervision" + ], + "table_footnote": [], + "table_body": "
SupervisionSynSyn + RealSyn + Self
Supervision signals required in real domain
Depth--------✓---✓-
Pose Refiner-----------✓✓-
Weak labels---------✓✓---
GT Pose labels-----✓✓✓------
MethodsAAE* [45]MHP* [29]DPODv2 [42]GDR-LB** [48]DeepIM† [26]DPODv2 [42]GDR-UB [48]SO-Pose [10]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape4.011.962.150.985.880.085.0-38.937.631.276.075.480.9
Benchwise20.966.288.399.493.199.799.8-75.278.683.091.694.999
Camera30.522.492.589.299.199.296.5-36.965.649.697.197.094.8
Can35.959.896.697.299.899.699.3-65.665.656.599.899.599.7
Cat17.926.986.179.998.795.193.0-57.952.557.985.686.692.6
Driller24.044.690.198.7100.098.9100.0-67.048.873.798.898.997.4
Duck4.98.354.824.661.979.565.3-19.635.131.356.568.383.4
Eggbox*81.055.798.681.193.599.699.9-99.089.296.091.099.094.9
Glue*45.554.695.481.293.399.898.1-94.164.563.492.296.193.4
Holep.17.615.527.041.932.172.373.4-16.241.538.835.441.979.3
Iron32.060.898.298.8100.099.486.9-77.980.961.999.599.499.8
Lamp60.5-91.098.999.196.399.6-68.270.764.797.498.998.3
Phone33.834.474.364.394.896.886.3-50.160.554.491.894.378.9
Average31.438.881.277.488.093.591.096.058.960.658.685.688.591.7
", + "bbox": [ + 67, + 152, + 905, + 368 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2c20daeb1c8c1947496d537190ba4ae3de63d17dfa4a556943ca42a8bc4099a.jpg", + "table_caption": [ + "Table 2. Evaluation results on Occluded LineMOD dataset. \\*: objects with symmetry. \\*\\*: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. $\\dagger$ :re-implemented version from [46], supervision source for Self6D++. $\\ddagger$ : with depth supervision." + ], + "table_footnote": [], + "table_body": "
SupervisionSynSyn + Self
MethodsGDR** [48]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape44.013.712.09.157.759.460.5
Can83.943.227.521.195.096.593.4
Cat49.118.712.026.052.660.856.1
Driller88.532.520.533.590.592.092.5
Duck15.014.423.012.226.730.655.5
Eggbox*33.957.825.139.445.051.146.0
Glue*75.054.327.037.087.188.682.8
Holep.34.022.035.020.423.538.546.5
Average52.932.122.824.859.864.766.7
", + "bbox": [ + 73, + 467, + 460, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the contrary, our method relies on texture optimisation to synthesise realistic training samples with precise pose annotations for supervision, making self-supervision more robust towards pose errors. We further validate our scheme with another pose estimator, DPODv2 [42] in supp. mat. As for our results on Occluded LineMOD in Table 2, we can make the same observations. We again report the best overall performance of all self-supervised methods, despite not relying on any elaborate design for occlusion-awareness as Self6D++ [46], with $66.7\\%$ vs $64.7\\%$ and $59.8\\%$ for Self6D++ [46] with and without depth data, respectively.", + "bbox": [ + 63, + 599, + 468, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on HomebrewDB and generalisation In this experiment, we want to study the generalisation ability of our method towards new scenes with changing illumination. From Table 3, we observe that after self-supervision on LineMOD, Self6D++ undergoes a clear performance drop compared to their respective pretrained models when testing on HomebrewDB (86.8% vs. 80.6%). This reveals that Self6D++'s self-supervision scheme is prone to forgetting.", + "bbox": [ + 63, + 763, + 470, + 878 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6bc7fe8c6004e583a48cfc6c1683a0a68579d2a53376b04921b1d9dd79f8cb5.jpg", + "table_caption": [ + "Table 3. Evaluation results on HomebrewedDB dataset. **: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. †: Pre-process testing images with the same intrinsics as LineMOD, different from [46] as they directly feed raw images without scaling. ‡: Use raw images for re-adaptation." + ], + "table_footnote": [], + "table_body": "
SupervisionSynSyn + Self
MethodGDR** † [48]Sock et al. [44]Self6D++ † [46]Ours †Ours ‡
Benchvise88.857.375.793.192.9
Driller92.846.689.494.894.2
Phone78.741.576.879.381.2
Average86.852.080.689.189.4
", + "bbox": [ + 504, + 467, + 899, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method can instead generalise to the new scene with even boosted performance (86.8% vs. 89.1%). We attribute this to the learnt realistic textures, which are used to uniformly synthesise photorealistic data that can further be augmented with random backgrounds to increase domain invariance [11,26]. Further, we investigate the adaptation ability of our method on these new scenes. To this end, we conduct self-supervision on raw images from HomebrewedDB. We first synthesise novels view with the pre-trained texture learner from LineMOD in an effort to adapt the pose estimator to the camera from HomebrewedDB. Afterwards, we re-train the texture learner with raw images from the new target scene. Our results again illustrate that our method outperforms all other baselines. Interestingly, we notice that the choice of adaptation domain (LineMOD or HomebrewedDB) has marginal effect on our final performance, suggesting that our method has potential to close the sim-to-real domain gap even trained for one scene.", + "bbox": [ + 497, + 547, + 905, + 802 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 499, + 812, + 658, + 827 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dealing with bad initialisation. We randomly sample 200 testing images from the Duck sequence of LineMOD (shown in Fig. 1). For the baseline render-and-compare", + "bbox": [ + 497, + 834, + 904, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4847", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c30065e74a9aeeb477ff190a737d0bfc3591f2e0e5bf5501a4bd462e1bc6663.jpg", + "image_caption": [ + "Figure 3. Ablation of the different schemes and the required amount of data for self-supervision. Left: Pose errors after every 10 iterations for the baseline scheme (pseudo pose) and our proposed pseudo-texture scheme. Right: Self-supervision with different amount of training data." + ], + "image_footnote": [], + "bbox": [ + 95, + 112, + 321, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/329c65a021266d7e8a6a177438467c0ea819bf03b84b7cf10854ead6586a674b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 329, + 112, + 542, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68cb7e21ea46b56466eb03d6e9a13153e720613db22a76d8983ee68a1b1f6bec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 112, + 870, + 254 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scheme (pseudo pose), we minimise $E_{\\text{rend}}$ in Eqn. 1 to refine the pose with the raw initialisation from the pose estimator. For our proposed scheme (pseudo texture), we repeatedly run an evaluation after every 10 weight update steps of the pose estimator. From Fig. 3 (left), we see that the baseline scheme diverges, while our proposed scheme significantly enhances the convergence properties thanks to the synthesised supervision from the texture learner.", + "bbox": [ + 63, + 312, + 468, + 426 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1ee3061e64bf7dc6325db00adf94ba64b094e7b425f139664663daf7d1238b1.jpg", + "table_caption": [ + "Table 4. Ablation study on partial objects from LineMOD dataset.*: objects with symmetry." + ], + "table_footnote": [], + "table_body": "
EadvEregApeCatDuckGlue*Holep.PhoneAverage
w/o self-sup.50.979.924.681.241.964.357.1
--76.390.365.982.359.665.373.3
✓-80.489.476.590.976.570.780.7
-✓82.187.878.496.477.176.283.0
✓✓80.992.683.493.479.378.984.7
", + "bbox": [ + 70, + 475, + 467, + 557 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of $E_{adv}$ and $E_{reg}$ . We study the individual components of our proposed training objective to guide texture learning under pose noise using the six worst performing objects. As shown in Table 4, even when training the texture learner with noisy poses, we are able to significantly improve the lower bound for most objects. We attribute this behaviour to the strong synthesis capabilities of NeRF [36]. Notably, for objects with richer textures or articulation (Glue, Phone), the baseline model without their supervision yields nearly no improvements. However, when respectively introducing $E_{adv}$ and $E_{reg}$ to the baseline model, we observe ca. $10\\%$ absolute improvement on ADD(-S) score. Further, combining both of them yields the best performance.", + "bbox": [ + 63, + 586, + 468, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of training data size. As the core idea of our method is to learn the object texture from multiple views, instead of optimising per-sample pose for supervision, we believe such formulation has a potential to allow for pose learning with much less data. To this end, we conduct self-supervision with an increasing dataset size taken from the", + "bbox": [ + 63, + 787, + 468, + 874 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "original training split. On the right of Fig. 3, we can see that the performance keeps steadily increasing when the amount of data is enlarged. Surprisingly, we observe that already with $5\\%$ of the training split (only 9 images), our method can bring a significant absolute improvement of ca. $18\\%$ , achieving around $75\\%$ of the optimal performance. This further confirms our method can even serve as a strong baseline for few-shot learning under extreme data scarcity.", + "bbox": [ + 497, + 312, + 904, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We refer interested readers to supp. mat for further detailed ablation experiments, e.g., analysis on optimisation design (Eqn. 1 and Eqn. 2) in Table 7.", + "bbox": [ + 497, + 426, + 904, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 480, + 621, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we formulate a new learning paradigm for the task of 6D pose estimation. Separating the task into texture and pose learning enables a novel self-supervision scheme to adapt pretrained 6D pose estimators to the real domain. Only a few unlabelled images are needed within this framework to generate pixel-perfect supervision. With this reformulation, we successfully bridge the gap between synthetic pre-training and real data usage through geometry-transfer from pixel-perfect in silico data onto data-induced appearance information. It successfully addresses the difficulty in optimising pose with raw 2D observations and further prunes the need for additional supervisions as guidance. Although initialised with much inferior accuracy, it improves the state-of-the-art self-supervision RGB method by an impressive $6.1\\%$ while even being superior to RGBD methods by $3.2\\%$ . We further demonstrates substantial generalisation improvements towards real-world scenarios.", + "bbox": [ + 497, + 505, + 902, + 745 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For limitations, though we require as little as nine images for texture learning, they shall have mild occlusions to provide sufficient appearance information of the objects. We also share current limitations with other model-based pose estimation methods and hence need a CAD model to ensure a precise object-centric neural representation. For future works, we strongly believe that our reformulation can impact self-supervision pipelines also for unseen objects with joint optimisation of both geometry and texture.", + "bbox": [ + 497, + 745, + 904, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4848", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 65, + 112, + 164, + 128 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: speeded up robust features. In Ales Leonardis, Horst Bischof, and Axel Pinz, editors, Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part I, volume 3951 of Lecture Notes in Computer Science, pages 404-417. Springer, 2006. 2", + "[2] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 95-104. IEEE Computer Society, 2017. 1", + "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In David J. Fleet, Tomás Pajdla, Bernt Schiele, and Tinne Tuyte-laars, editors, Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II, volume 8690 of Lecture Notes in Computer Science, pages 536-551. Springer, 2014. 6", + "[4] Kai Chen, Rui Cao, Stephen James, Yichuan Li, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Sim-to-real 6d object pose estimation via iterative self-training for robotic bin-picking. arXiv preprint arXiv:2204.07049, 2022. 3", + "[5] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI, volume 12371 of Lecture Notes in Computer Science, pages 139-156. Springer, 2020. 3", + "[6] Alvaro Collet, Manuel Martínez, and Siddhartha S. Srinivasa. The MOPED framework: Object recognition and pose estimation for manipulation. Int. J. Robotics Res., 30(10):1284-1306, 2011. 2", + "[7] Alvaro Collet and Siddhartha S. Srinivasa. Efficient multiview object recognition and full pose estimation. In IEEE International Conference on Robotics and Automation, ICRA 2010, Anchorage, Alaska, USA, 3-7 May 2010, pages 2050-2055. IEEE, 2010. 2", + "[8] Xinke Deng, Yu Xiang, Arsalan Mousavian, Clemens Eppner, Timothy Bretl, and Dieter Fox. Self-supervised 6d object pose estimation for robot manipulation. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 3665-3671. IEEE, 2020. 2", + "[9] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. CoRR, abs/1911.01911, 2019. 1, 5", + "[10] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting selfocclusion for direct 6d pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021," + ], + "bbox": [ + 67, + 137, + 470, + 876 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Montreal, QC, Canada, October 10-17, 2021, pages 12376-12385. IEEE, 2021. 2, 7", + "[11] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1310-1319. IEEE Computer Society, 2017. 1, 7", + "[12] Stefano Gasperini, Jan Haug, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, Benjamin Busam, and Federico Tombari. Certainnet: Sampling-free uncertainty estimation for object detection. IEEE Robotics Autom. Lett., 7(2):698-705, 2022. 1", + "[13] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 3193-3203. Computer Vision Foundation / IEEE, 2020. 1", + "[14] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Kyoung Mu Lee, Yasuyuki Matsushita, James M. Rehg, and Zhanyi Hu, editors, Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, volume 7724 of Lecture Notes in Computer Science, pages 548-562. Springer, 2012. 1, 5, 6", + "[15] Tomas Hodan, Pavel Haluza, Stepan Obdrzalek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 1", + "[16] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Adrien Bartoli and Andrea Fusiello, editors, Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12536 of Lecture Notes in Computer Science, pages 577-594. Springer, 2020. 1, 5", + "[17] HyunJun Jung, Patrick Ruhkamp, Guangyao Zhai, Nikolas Brasch, Yitong Li, Yannick Verdie, Jifei Song, Yiren Zhou, Anil Armagan, Slobodan Ilic, et al. Is my depth ground-truth good enough? hammer-highly accurate multimodal dataset for dense 3d scene regression. arXiv preprint arXiv:2205.04565, 2022. 2", + "[18] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 1, 6", + "[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy," + ], + "bbox": [ + 502, + 116, + 904, + 876 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4849", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "October 22-29, 2017, pages 1530-1538. IEEE Computer Society, 2017. 1, 2", + "[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 4", + "[21] Zhengfei Kuang, Kyle Olszewski, Mengei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 41(4), jul 2022. 4, 5", + "[22] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 3559-3568. Computer Vision Foundation / IEEE Computer Society, 2018. 3", + "[23] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi-object 6d pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII, volume 12362 of Lecture Notes in Computer Science, pages 574-591. Springer, 2020. 1", + "[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1654-1663. Computer Vision Foundation / IEEE, 2021. 1", + "[25] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022. 1, 3", + "[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis., 128(3):657-678, 2020. 1, 2, 6, 7", + "[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: bundle-adjusting neural radiance fields. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 5721-5731. IEEE, 2021. 3, 5", + "[28] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2", + "[29] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6841-6850, 2019. 2, 7", + "[30] Fabian Manhardt, Wadim Kehl, and Adrien Gaidon. ROI10D: monocular lifting of 2d detection to 6d pose and metric shape. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2069-2078. Computer Vision Foundation / IEEE, 2019. 1" + ], + "bbox": [ + 68, + 115, + 470, + 875 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in RGB. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIV, volume 11218 of Lecture Notes in Computer Science, pages 833-849. Springer, 2018. 1", + "[32] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning. arXiv preprint arXiv:2003.05848, 2020. 2", + "[33] Éric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph., 22(12):2633-2651, 2016. 1", + "[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7210-7219. Computer Vision Foundation / IEEE, 2021. 3, 4, 5", + "[35] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6331-6341. IEEE, 2021. 3", + "[36] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 8", + "[37] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 7667-7676. IEEE, 2019. 2", + "[38] Kiru Park, Timothy Patten, and Markus Vincze. Neural object learning for 6d pose estimation using a few cluttered images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV, volume 12349 of Lecture Notes in Computer Science, pages 656-673. Springer, 2020. 3", + "[39] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4561-4570. Computer Vision Foundation / IEEE, 2019. 1, 2", + "[40] Mahdi Rad and Vincent Lepetit. BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In IEEE International Conference on Computer Vision, ICCV 2017," + ], + "bbox": [ + 502, + 116, + 902, + 876 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4850", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Venice, Italy, October 22-29, 2017, pages 3848-3856. IEEE Computer Society, 2017. 1, 2, 6", + "[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 5", + "[42] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 6, 7", + "[43] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5", + "[44] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In Vitomir Struc and Francisco Gomez Fernandez, editors, 8th International Conference on 3D Vision, 3DV 2020, Virtual Event, Japan, November 25-28, 2020, pages 291-300. IEEE, 2020. 1, 2, 3, 6, 7", + "[45] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2, 7", + "[46] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 3, 6, 7", + "[47] Gu Wang, Fabian Manhardt, Jianzhuon Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 108-125. Springer, 2020. 1, 2, 3, 6, 7", + "[48] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16611-16621. Computer Vision Foundation / IEEE, 2021. 2, 6, 7", + "[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2642-2651. Computer Vision Foundation / IEEE, 2019. 5" + ], + "bbox": [ + 68, + 115, + 470, + 868 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Loranzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multimodal dataset for category-level object pose estimation with photometrically challenging objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[51] Pengyuan Wang, Fabian Manhardt, Luca Minciullo, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Demograsp: Few-shot learning for robotic grasping with human demonstration. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5733-5740. IEEE, 2021. 1", + "[52] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, abs/2102.07064, 2021. 3", + "[53] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 1, 2", + "[54] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 3907-3916. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 7", + "[55] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. NeRF-Supervision: Learning dense object descriptors from neural radiance fields. In IEEE Conference on Robotics and Automation (ICRA), 2022. 3", + "[56] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. iNeRF: Inverting neural radiance fields for pose estimation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 3, 5", + "[57] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of SDF shape priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 12221-12230. Computer Vision Foundation / IEEE, 2020. 1", + "[58] Sergey Zakharov, Wadim Kehl, and Slobodan Ilic. Deceptionnet: Network-driven domain randomization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 532-541. IEEE, 2019. 1", + "[59] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6d pose object detector and refiner. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 1941-1950. IEEE, 2019. 1, 2", + "[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conference on Computer Vision and" + ], + "bbox": [ + 502, + 116, + 904, + 870 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4851", + "bbox": [ + 468, + 891, + 499, + 903 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5745-5753. Computer Vision Foundation / IEEE, 2019. 2", + "bbox": [ + 100, + 116, + 468, + 154 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4852", + "bbox": [ + 468, + 891, + 500, + 902 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_model.json b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e0f86460334bde22a66a298b8ef0531a27cb139 --- /dev/null +++ b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_model.json @@ -0,0 +1,2237 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.12, + 0.038, + 0.191, + 0.074 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.036, + 0.8, + 0.077 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.152, + 0.768, + 0.194 + ], + "angle": 0, + "content": "TexPose: Neural Texture Learning for Self-Supervised 6D Object Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.22, + 0.792, + 0.238 + ], + "angle": 0, + "content": "Hanzhi Chen\\(^{1}\\) Fabian Manhardt\\(^{2}\\) Nassir Navab\\(^{1}\\) Benjamin Busam\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.238, + 0.739, + 0.255 + ], + "angle": 0, + "content": "\\(^{1}\\) Technical University of Munich \\(^{2}\\) Google Inc. \\(^{3}\\) 3Dwe.ai" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.258, + 0.766, + 0.271 + ], + "angle": 0, + "content": "hanzhi.chen@tum.de fabianmanhardt@google.com b.busam@tum.de" + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.304, + 0.308, + 0.319 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.067, + 0.336, + 0.473, + 0.664 + ], + "angle": 0, + "content": "In this paper, we introduce neural texture learning for 6D object pose estimation from synthetic data and a few unlabelled real images. Our major contribution is a novel learning scheme which removes the drawbacks of previous works, namely the strong dependency on co-modalities or additional refinement. These have been previously necessary to provide training signals for convergence. We formulate such a scheme as two sub-optimisation problems on texture learning and pose learning. We separately learn to predict realistic texture of objects from real image collections and learn pose estimation from pixel-perfect synthetic data. Combining these two capabilities allows then to synthesise photorealistic novel views to supervise the pose estimator with accurate geometry. To alleviate pose noise and segmentation imperfection present during the texture learning phase, we propose a surfel-based adversarial training loss together with texture regularisation from synthetic data. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen scenes. Remarkably, our scheme improves the adopted pose estimators substantially even when initialised with much inferior performance." + }, + { + "type": "title", + "bbox": [ + 0.067, + 0.695, + 0.201, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.721, + 0.471, + 0.879 + ], + "angle": 0, + "content": "For spatial interaction with objects, one needs an understanding of the translation and rotation of targets within 3D space. Inferring these 6D object pose parameters from a single RGB image is a core task for 3D computer vision. This visually retrieved information has a wide range of applications in AR/VR [13, 33], autonomous driving [12, 30, 57], and robotic manipulation [24, 51, 53]. Noteworthy, accuracy and runtime have both recently made a huge leap forward thanks to deep learning [19, 23, 31, 39, 40]. Unfortunately, most of these methods heavily rely on a massive amount of labelled data for supervision to learn precise models with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.905, + 0.448 + ], + "angle": 0, + "content": "strong generalisation capabilities [16, 39, 53, 59]. However, it is very labor-intensive and time consuming to generate accurate annotations for pose data [15, 53]. Meanwhile, this process also easily suffers from labelling errors as precise annotation in 3D is highly challenging [14]. Therefore, most of the benchmarks have only few hundreds images, which does not allow proper learning of large models. In fact, methods training with such low amount of real data tend to strongly overfit to the data domain and fail to generalise to new scenes [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.905, + 0.833 + ], + "angle": 0, + "content": "As a consequence, many different approaches have been proposed in the literature to tackle this problem. The simplest solution is to employ cut-and-paste strategy to increase domain invariance [11,26]. Nonetheless, this requires highly accurate manual annotations and most models tend to overfit to the original domain. Another alternative is to rely on a large amount of synthetically generated data to prevent overfitting. Though this process is relatively cheap and fast, rendered images can exhibit drastic visual discrepancy in comparison with real images even when advanced physically-based renderers [9] are used. A handful of approaches try to close the domain gap via the use of generative adversarial networks to translate the synthetic images into the real domain [2,58]. Unfortunately, these methods do not achieve promising results as the generated images are still easily distinguishable from real imagery. Notably, very recently a new line of work that proposes to self-supervise the pose estimator on real data has emerged. After training in simulation they fine-tune on the new datasets [46, 47]. While these methods achieve impressive results that are even on par with fully supervised methods, they still suffer from significant performance drop when not leveraging additional supervisory signals such as depth data [46] or ground truth camera poses [25, 42]. In this work, we propose a novel way to conduct self-supervised 6D pose estimation that is free from any additional supervision sources and further yields state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.835, + 0.905, + 0.879 + ], + "angle": 0, + "content": "The core idea of previous attempts [44,46,47,54] to adapt pretrained pose estimators to the real domain is to conduct render-and-compare using a differentiable renderer. With" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.904 + ], + "angle": 0, + "content": "4841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.115, + 0.47, + 0.23 + ], + "angle": 0, + "content": "CAD models and poses given, a renderer can output several attributes of the current object pose (e.g. mask, colour, depth) which allows to refine the pose through iterative comparison between the renderings and observations. However, when relying on 2D visual contents (mask and colour), such strategies tend to fail due to measured silhouette imperfection, lack of textures, and domain discrepancies. This undesirable behaviour is also discussed in [47]." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.232, + 0.473, + 0.63 + ], + "angle": 0, + "content": "Thus, different from the previous attempts that heavily rely on render-and-compare for self-supervision [46, 47], we instead propose to regard realistic textures of the objects as an intermediate representation before conducting training for the pose estimator. Our approach is formulated as two interconnected sub-optimisation problems on texture learning and pose learning. In the core, we first learn realistic textures of objects from raw image collections, then synthesise training data to supervise the pose estimator with pixel-perfect labels and realistic appearance. The key challenge of our proposed scheme lies in capturing accurate texture under noise introduced by poses initialised by a pretrained pose estimator during supervision. To this end, in addition to leveraging synthetic data to establish geometry priors, we learn robust supervision through adversarial training by conditioning synthesised colours on local surface information. Furthermore, we establish regularisation from synthetic textures to compensate segmentation artefacts during a texture-learning phase. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen domains. Our method significantly improves even difficult objects with little variance in appearance and geometry through self-supervision. Impressively, Our approach demonstrates a robust self-improving ability for the employed pose estimators even when initialised with much inferior pose estimates than stronger baselines [46]." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.632, + 0.375, + 0.646 + ], + "angle": 0, + "content": "To summarise, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.657, + 0.469, + 0.7 + ], + "angle": 0, + "content": "- We formulate a new learning scheme, TexPose, that decomposes self-supervision for 6D object pose into Texture learning and Pose learning." + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.719, + 0.469, + 0.79 + ], + "angle": 0, + "content": "- We propose a surfel-conditioned adversarial training loss and a synthetic texture regularisation term to handle pose errors and segmentation imperfection during texture learning, further delivering self-improving ability to the pose estimators." + }, + { + "type": "text", + "bbox": [ + 0.083, + 0.81, + 0.468, + 0.867 + ], + "angle": 0, + "content": "- We show significant improvements over recent strong baselines with additional supervision signals. Our pose estimators demonstrate a substantial generalisation ability even on unseen scenes." + }, + { + "type": "list", + "bbox": [ + 0.083, + 0.657, + 0.469, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.114, + 0.646, + 0.129 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.139, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Model-based 6D Pose Estimation To retrieve the 6D pose, early methods use local or global features and search for key points correspondence on CAD models [1,6,7,28]. In recent years, learning-based methods dominate the field and solve the task using convolution neural networks (CNN) under the supervision from annotated data to extract deep features. There are two major approaches for pose estimation, in particular, correspondence-based and regression/classification-based approaches. Correspondence-based methods establish 2D-3D correspondences [37, 39, 40, 42, 59], prior to leveraging a variant of the RANSAC&PnP paradigm to solve for pose. Regression-based approaches, on the other hand, directly regress or classify the pose of the detected object. Initially these methods usually have shown a lower performance due to the existence of ambiguities [29] such as pose symmetries [53]. Therefore, methods like SSD-6D [19] discretise the rotation space to circumvent this issue. Recently, with better continuous representations for rotation [60], these methods gradually demonstrate high effectiveness [10, 48]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.905, + 0.879 + ], + "angle": 0, + "content": "Self-Supervised Pose Learning Considering tremendous effort to collect large amount of annotations for 6D object pose [17, 50], several recent works have been proposed to explore the possibility of self-supervised pose learning using labeled synthetic data together with unlabelled real sensor data. [8] designed a novel labelling pipeline using a manipulator to generate reliable pose annotations for supervision. Self6D [47] proposed a self-supervision workflow by first pretraining a pose estimator with synthetic data, which was then adapted to the real world through a render-and-compare strategy by imposing consistencies between the rendered depth and sensed depth under the current pose estimate. CPS++ [32] used similar approach for categorical-level object pose estimation, and the shape is jointly deformed during optimisation. AAE [45] parameterise \\(SO(3)\\) space using a latent embedding obtained from synthetic images, which is later employed for orientation retrieval. Both works demonstrate the strong requirement for depth data in order to allow for a reasonable performance either in training or testing. However, due to the uninformative appearance and the sim-to-real domain gap, render-and-compare strategies easily diverge when there is no depth data available as shown in [47]. Hence, consecutive works introduce other supervisory signals to prune the need for depth. DSC-PoseNet [54] employs a key point consistency regularisation for dual-scale images with labelled 2D bounding box. Socket al. [44] use photometric consistency from multiple views to refine the raw estimate, while ground-truth masks are required for supervision. Recently, Self6D++ [46] proposed to initialise the render-and-compare process with a powerful deep pose refiner [26] to guide the learning process of the pose esti" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "4842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.116, + 0.47, + 0.173 + ], + "angle": 0, + "content": "mator, which currently yields state-of-the-art performance among all methods without manual annotations. Chen et al. [4] leverages a tailored heuristics for pseudo labelling under student-teacher learning scheme." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.189, + 0.473, + 0.63 + ], + "angle": 0, + "content": "Novel View Synthesis With a few images given, novel view synthesis aims to render scene content from unseen viewpoints, which can be regarded as a process to acquire \"textures\" of the scene. NOL [38] is proposed to address the difficulty of CAD model texturing for 6D pose estimation. Though requiring pose annotations, it demonstrates high potential to gain improvement for pose estimators with more synthesised real data in a semi-supervised fashion. Outside the scope of 6D pose, neural scene representation as implicit scene models recently get significant attention due to their ability for photorealistic novel view synthesis. NeRF [36] represent scenes using simple multi-layer perceptrons to encode the density and radiance for every 3D point. Assuming precise poses to transform views into a common reference frame, NeRF is able to capture volumes using an image reconstruction loss. NeRF-w [34] further extends NeRF to handle changing lighting and transient effects for image collections with latent embeddings. To release the rigid requirement of accurate poses, NeRF-- [52] jointly optimises camera parameters and radiance fields. iNeRF [56] performs a render-and-compare strategy using a trained NeRF for pose refinement. In contrast, BARF [27] demonstrates promising results when encountering imperfect pose initialisation by bundle-adjusting NeRF and camera poses with coarse-to-fine positional encoding scheme. GNeRF [35] formulates the learning of radiance fields and coarse pose estimates as a generative modelling process and then jointly refines both using photometric information. Recently, NeRF-Supervision [55] and NeRF-Pose [25] are proposed to leverage NeRF to learn dense descriptors or 3D representations of a given object under pose annotations when CAD model is inaccessible." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.64, + 0.16, + 0.655 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.664, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Our objective is to conduct per-instance 6D pose learning from monocular colour images with no manual labelling effort, namely self-supervised 6D pose estimation originally defined in [47]. Since the CAD model is known as a prior, we follow the two-stage approach of previous works [44, 46, 47, 54, 56], which first pretrain a pose estimator with labelled synthetic data, and consequently conduct self-training with unannotated real images captured from a specific real scene demanding a more precise deployment (e.g. robotic bin picking). We mainly introduce the self-training process in this section. Notably, in this tuning process, pose estimator tends to suffer from forgetting problem due to little variance in illumination and backgrounds of the fed training samples. In contrast, We will show the strength of our scheme to alleviate this issue in Section 4. After pretraining we acquire" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.905, + 0.173 + ], + "angle": 0, + "content": "a pose estimator \\( f_{\\theta} \\) able to initialise a coarse pose estimate \\( T \\) and a segmentation mask \\( M \\) for each real image \\( \\hat{I} \\). \\( M \\) is usually used as a pseudo silhouette cue to refine the pose \\( T \\) with render-and-compare strategy [5,22,47,56]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.173, + 0.905, + 0.23 + ], + "angle": 0, + "content": "In this section, we first revisit the commonly used strategy of previous methods in this task, then introduce our new formulation requiring no additional supervisory sources clearly advantageous over previous attempts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.237, + 0.89, + 0.254 + ], + "angle": 0, + "content": "3.1. Revisiting Self-Supervised 6D Pose Learning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.905, + 0.402 + ], + "angle": 0, + "content": "Render-and-compare is a common approach for 6D pose self-supervision [44, 46, 47, 54, 56]. As shown in Eqn. 1, it utilises a differentiable renderer \\(\\mathcal{R}\\) fed with an estimated pose to output a rendered mask, image, \\([M_r, I_r] = \\mathcal{R}(T)\\). \\(E_{rend}\\) enforces consistency between the rendered outputs \\([M_r, I_r]\\) and the observations \\([M, \\hat{I}]\\). It aims to refine the pseudo pose labels for each training sample, and further supervise pose estimator \\(f_{\\theta}\\) under \\(E_{pose}\\). Previous attempts are majorly focused on improving \\(E_{rend}\\) to acquire more accurate pseudo pose annotations for better supervision." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.41, + 0.903, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta , \\{T _ {i}, \\dots , T _ {N} \\}} \\sum_ {i} E _ {r e n d} (\\mathcal {R} (T _ {i}), [ M _ {i}, \\hat {I} _ {i} ]) \\\\ + \\sum_ {i} ^ {\\iota} E _ {p o s e} \\left(f _ {\\theta} \\left(\\hat {I} _ {i}\\right), T _ {i}\\right). \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.905, + 0.665 + ], + "angle": 0, + "content": "When relying on 2D information \\([M_r, I_r]\\), minimising \\(E_{rend}\\) to improve \\(\\{T_1, \\ldots, T_N\\}\\) can be extremely difficult due to uninformative shape and appearance and sim-to-real visual discrepancy (e.g., Duck in Fig. 1). To compensate for the lack of information, related methods introduce additional supervision with varying difficulty levels for the acquisition (e.g., depth [47], weak labels [44, 54], deep pose refiner [46], etc.). Deep pose refiner could be acquired with the least effort as it only requires synthetic pretraining as well. However, it can introduce a strong upper bound as illustrated in [46]. The pose refiner can even lead to divergence of the pose estimator and hence lacks robustness for geometrically and photometrically challenging objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.836, + 0.687 + ], + "angle": 0, + "content": "3.2. TexPose: Learning Pose from Texture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.694, + 0.905, + 0.752 + ], + "angle": 0, + "content": "To address the aforementioned limitations, we reformulate self-supervised 6D pose learning as two alternating optimisation problems for a texture learner \\( h_{\\xi} \\) and a pose estimator \\( f_{\\theta} \\) as defined in Eqn. 21 and depicted in Fig. 1 (Left)." + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.761, + 0.901, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta , \\xi} \\sum_ {i} E _ {t e x} \\left(h _ {\\xi} \\circ f _ {\\theta} (\\hat {I} _ {i}), \\hat {I} _ {i}\\right) \\\\ + \\sum_ {j} ^ {i} E _ {\\text {p o s e}} \\left(f _ {\\theta} \\circ h _ {\\xi} \\left(\\hat {T} _ {j}\\right), \\hat {T} _ {j}\\right). \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.829, + 0.905, + 0.859 + ], + "angle": 0, + "content": "Instead of leveraging raw images as regression targets for poses, we insert object textures as an intermediate rep" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.864, + 0.895, + 0.878 + ], + "angle": 0, + "content": "\\( {}^{1} \\) Green indicates weight update and orange indicates weight freezing." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "4843" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.121, + 0.442, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.12, + 0.856, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.064, + 0.324, + 0.906, + 0.391 + ], + "angle": 0, + "content": "Figure 1. Left: Schematic overview of TexPose. We fully explore the benefit of both synthetic and real dataset with perfect geometric labels and realistic appearance to synthesise informative train dataset for pose learning. Right: Details of the texture learner used in our scheme. We first conduct geometric pretraining to acquire geometry branch \\( h_{\\sigma} \\) to ensure consistent coordinates system and geometry aligned with the CAD model. Then, we transfer and freeze \\( h_{\\sigma} \\) to the real together with newly established radiance branch \\( h_c^s \\) and \\( h_c^t \\) as our texture learner. It is supervised by \\( E_{rec} \\), \\( E_{reg} \\) and \\( E_{adv} \\) capable of improving texture quality so as to synthesise more informative dataset for supervision." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.415, + 0.47, + 0.542 + ], + "angle": 0, + "content": "resentation before pose learning. Specifically, we start with optimising \\( E_{tex} \\) by fixing the parameters of pose estimator \\( f_{\\theta} \\) used to initialise poses \\( \\{T_1,\\dots ,T_N\\} \\) for each raw image \\( \\{\\hat{I}_1,\\ldots ,\\hat{I}_N\\} \\), aiming to provide supervision for the realistic textures embedded by \\( h_\\xi \\). Afterwards, we freeze \\( h_\\xi \\) to synthesise training samples \\( \\{I_1,\\dots ,I_{N'}\\} \\) paired with perfect labels \\( \\{\\hat{T}_1,\\dots ,\\hat{T}_{N'}\\} \\). Gradients will thus be propagated through \\( f_{\\theta} \\) to update its parameters during the minimisation for \\( E_{pose} \\) over these train data. We typically chose \\( N' > N \\)." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.543, + 0.471, + 0.756 + ], + "angle": 0, + "content": "With this reformulation, we deliberately omit pose optimisation from 2D information usually sensitive to initialisation and prone to divergence, and bridge the domain gap directly through photorealistic synthesis. Our goal is then shifted to limit texture mapping error under the presence of pose noise so that the synthesised data can be highly informative with perfect geometric labels and as-accurate-as-possible realistic appearance. Intuitively, through multi-view supervision, it's possible to capture more accurate appearance with compensation for per-view pose noise. Our proposed strategy to further increase reliability of the synthesised data will be discussed later. Notably, different from previous attempts, no additional supervision is required in our design. The pose estimator is equipped with a self-improving ability through intermediate phase for texture learning." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.773, + 0.468, + 0.816 + ], + "angle": 0, + "content": "NeRF preliminaries. We leverage neural radiance fields (NeRF) [36] to embed the texture information due to its simplicity and capability of photorealistic view synthesis2." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.816, + 0.468, + 0.845 + ], + "angle": 0, + "content": "NeRF represents the object model with two MLPs to encode information for every queried point \\(\\mathbf{x}\\), namely the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.415, + 0.905, + 0.591 + ], + "angle": 0, + "content": "geometry branch to predict volume density \\(\\sigma\\) and geometric feature \\(\\gamma\\) with \\([\\sigma, \\gamma] = h_{\\sigma}(\\mathbf{x})\\) and radiance branch for colour \\(\\mathbf{c}\\) prediction with \\(\\mathbf{c} = h_c(\\mathbf{x}, \\gamma)\\) (weight \\(\\xi\\) is omitted). When rendering from a given viewpoint, \\(K\\) samples \\(\\{x_k\\}_1^K\\) are gathered along the ray \\(\\mathbf{r} = \\mathbf{o} + t\\mathbf{d}\\) where \\(\\mathbf{o}, \\mathbf{d} \\in \\mathbb{R}^3\\) represent the camera centre and the view direction respectively and \\(t \\in \\mathbb{R}\\) is the distance. The colour \\(\\mathbf{C}(\\mathbf{r})\\) will be estimated through numerical quadrature approximation with \\(w_k = \\exp(-\\delta_k \\sigma_k)\\) and \\(\\alpha_k = \\prod_{j=1}^{k-1} w_j\\) with \\(\\delta_k\\) being the step size between adjacent points, and \\(\\mathbf{C}(\\mathbf{r}) = \\sum_{k=1}^{K} \\alpha_k (1 - w_k) \\mathbf{c_k}\\). Rendering for depth \\(D(\\mathbf{r})\\) and mask \\(M(\\mathbf{r})\\) is achieved by modifying integration element (see supp. mat. for details)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.607, + 0.905, + 0.75 + ], + "angle": 0, + "content": "Geometric pretraining. Though geometry and radiance branch can be trained jointly as shown in NeRF and its variants [21,34,36], the pose for supervision in our scenario is highly noisy, and the coordinate system of NeRF can end up as an arbitrary one and further cause erroneous labels for synthesised data. We hence opt to pretrain geometry branch \\( h_{\\sigma} \\) using pixel-perfect data rendered by the CAD model with accurate pose \\( \\hat{T} \\), depth \\( \\hat{D} \\) and mask \\( \\hat{M} \\) to supervise NeRF following Eqn. 4, where \\( E_{d} \\) is scale-invariant loss from [20], \\( \\lambda_{m} \\) and \\( \\lambda_{d} \\) are weighting factors." + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.769, + 0.903, + 0.798 + ], + "angle": 0, + "content": "\\[\nE _ {2 d} (\\mathbf {r}) = \\left\\| \\mathbf {C} (\\mathbf {r}) - \\hat {\\mathbf {C}} (\\mathbf {r}) \\right\\| _ {2} ^ {2} + \\lambda_ {m} \\left\\| M (\\mathbf {r}) - \\hat {M} (\\mathbf {r}) \\right\\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.813, + 0.903, + 0.831 + ], + "angle": 0, + "content": "\\[\nE _ {p r e} (\\mathbf {r}) = E _ {2 d} (\\mathbf {r}) + \\lambda_ {d} E _ {d} (D (\\mathbf {r}), \\hat {D} (\\mathbf {r})). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.835, + 0.905, + 0.878 + ], + "angle": 0, + "content": "Notably, though our current goal is to establish a geometry branch with precise density output \"aligned\" with the CAD model, training for radiance branch is also included, though" + }, + { + "type": "page_footnote", + "bbox": [ + 0.065, + 0.852, + 0.468, + 0.878 + ], + "angle": 0, + "content": "2Radiance, texture and appearance are discussed as interchangeable concepts in our paper." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.904 + ], + "angle": 0, + "content": "4844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.115, + 0.47, + 0.188 + ], + "angle": 0, + "content": "it will be discarded afterwards. We find this greatly helps to extract meaningful geometric features fed into the radiance branch, and further eases the effort for realistic appearance transfer in the later texture learning stage. We quantitatively detail the necessity of this stage in supp. mat." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.206, + 0.47, + 0.376 + ], + "angle": 0, + "content": "Texture learning from noisy poses. Our current goal is to learn texture encoded by the radiance branch \\( h_c \\) using raw images paired with poses initialised by the pose estimator \\( f_{\\theta} \\). Note in this stage, we only re-optimise parameters from \\( h_c \\) with \\( h_{\\sigma} \\) being fixed to capture accurate texture and meanwhile ensure no shift of coordinate system and geometry. Though jointly optimising textures and poses through photometric consistency has a potential to improve both aspects as shown in [21, 27, 56], we find it can easily fail when dealing with the textureless objects. Alternatively, we seek to supervise the texture learner with higher pose error tolerance leveraging stronger radiance branch and robust loss." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.377, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Different from the vanilla NeRF used in geometric pretraining, we now employ two radiance branches, \\( h_c^s \\) and \\( h_c^t \\), to translate the extracted geometric features into static radiance (accurate underlying textures) and transient radiance (texture outliers caused by pose errors) respectively. Apart from geometric features \\( \\gamma \\), we further condition it on view directions \\( \\mathbf{d} \\), lighting embedding \\( \\tau^a \\) and transient embedding \\( \\tau^t \\) [34] (see Fig. 1 Right). For each view, we use \\( \\tau^a \\) to capture illumination variations and \\( \\tau^t \\) to produce transient effect caused by pose noise. Patch-based training strategy [41] is used to preserve the spatial structures for robust loss supervision. An image reconstruction loss \\( E_{rec} \\) is incorporated with uncertainty awareness to learn textures robustly. We refer to the supp. mat. for the details of the radiance branches and uncertainty-aware \\( E_{rec} \\)." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.59, + 0.47, + 0.718 + ], + "angle": 0, + "content": "We noticed that supervision from \\( E_{rec} \\) is still not adequate to capture accurate texture under pose noise. We further design a surfel-conditioned adversarial loss for sampled patches to tolerate texture shifts caused by pose errors. For each location on the object surface, we compute normalised object coordinate \\( S_{x} \\) [49] and normal \\( S_{n} \\) with current pose estimate \\( T \\) to represent its local information. By conditioning the predicted patch colour \\( P \\) on its local surface information \\( [S_{x}, S_{n}] \\), we employ an adversarial scheme:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.731, + 0.47, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E _ {a d v} (P) = \\mathbb {E} _ {[ S _ {x}, S _ {n} ], P} (1 - \\log D ([ S _ {x}, S _ {n} ], P)) \\tag {5} \\\\ + \\mathbb {E} _ {[ S _ {x}, S _ {n} ], \\hat {P}} (\\log D ([ S _ {x}, S _ {n} ], M \\hat {P})). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.777, + 0.469, + 0.878 + ], + "angle": 0, + "content": "We also observed the imperfect predicted mask \\(M\\) can cause background colour artefacts within the object. We therefore establish a regularisation from synthetic cues to compensate the boundary imperfections by generating synthetic patch \\(P_{syn}\\) and mask \\(M_{syn}\\) with current pose estimate \\(T\\), then applying erosion on the predicted mask \\(M\\) to compute a padded mask with \\(M_{pad} = M_{syn}(1 - M)\\). The foreground" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.905, + 0.23 + ], + "angle": 0, + "content": "boundaries of the real images for supervision are therefore padded with the synthetic colours. This regularisation \\( E_{feat} \\) is realised by minimising the MSE error of deep features extracted by a pretrained VGG19 network [43] to align the boundary features of the predicted patch \\( P \\) with the synthetic patch \\( P_{syn} \\) in a more abstract level. The second part of Eqn. 6 ensures that rendered patch region within the mask \\( M \\) are not affected by the synthetic information." + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.241, + 0.901, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E _ {r e g} (P) = E _ {f e a t} \\left(P, M \\hat {P} + M _ {p a d} P _ {s y n}\\right) \\tag {6} \\\\ + \\lambda_ {f g} E _ {\\text {f e a t}} (M P + (1 - M) \\hat {P}, \\hat {P}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Formally, the loss function \\( E_{tex} \\) for the texture learner supervision is summarised as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.327, + 0.901, + 0.341 + ], + "angle": 0, + "content": "\\[\nE _ {t e x} = E _ {r e c} + \\lambda_ {a d v} E _ {a d v} + \\lambda_ {r e g} E _ {r e g}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.351, + 0.794, + 0.366 + ], + "angle": 0, + "content": "where \\(\\lambda_{adv}\\) and \\(\\lambda_{reg}\\) are weighting factors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.905, + 0.553 + ], + "angle": 0, + "content": "Pose learning. The entire pipeline runs by first optimising \\( E_{tex} \\), and subsequent minimisation of \\( E_{pose} \\) with the synthesised dataset as depicted in Fig. 1 (Left). As our scheme operates at data level and thus method-agnostic, \\( E_{pose} \\) can be generalised across methods with different loss designs, e.g., direct pose or correspondence regression. Interestingly, we observed that the pose estimator can converge instantly with only one optimisation loop on \\( E_{tex} \\) and \\( E_{pose} \\), while adding more loops, i.e., continue to optimise \\( E_{tex} \\) with improved poses after \\( E_{pose} \\), only brings insignificant improvements. We attribute this to the pixel-perfect supervision from texture learner and its strong ability to mitigate pose errors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.636, + 0.581 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.904, + 0.646 + ], + "angle": 0, + "content": "In this section we first introduce the employed metrics and dataset before presenting our results for the task of 6D pose estimation, then provide detailed analysis of our proposed scheme in ablation study." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.69, + 0.669 + ], + "angle": 0, + "content": "4.1. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.905, + 0.803 + ], + "angle": 0, + "content": "In line with previous work, we present our results using the ADD(-S) metric, as proposed in [14]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. If the average distance is below \\(10\\%\\) of the object diameter, the estimated pose will be considered correct. For symmetric objects, ADD score is modified to measure the distance to the nearest vertices, here referred to as ADD-S." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.598, + 0.827 + ], + "angle": 0, + "content": "4.2. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.834, + 0.905, + 0.878 + ], + "angle": 0, + "content": "Synthetic BOP PBR. We pretrain our models using synthetic data from [9, 16], employing physically-based rendering (PBR) techniques. Although PBR can capture more" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "4845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.155, + 0.113, + 0.285, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.113, + 0.418, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.113, + 0.552, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.113, + 0.686, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.113, + 0.82, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.205, + 0.285, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.205, + 0.418, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.205, + 0.552, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.205, + 0.686, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.205, + 0.819, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.3, + 0.285, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.3, + 0.419, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.3, + 0.552, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.3, + 0.686, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.3, + 0.819, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.064, + 0.398, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Figure 2. Qualitative results on 5 objects from LineMOD. Top: Renderings of the reconstructed CAD models which are part of LineMOD. Middle: Same objects rendered with our learnt textures. Bottom: Demonstration of our pose estimation quality. Thereby, the green, red, and blue boxes depict ground-truth pose as well as estimated pose before and after our self-supervision, respectively." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.446, + 0.471, + 0.503 + ], + "angle": 0, + "content": "realistic lighting conditions, there still exhibits a large sim-to-real domain gap. Such discrepancies can strongly degrade the pose estimation performance when being deployed in the real world." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.522, + 0.469, + 0.622 + ], + "angle": 0, + "content": "LineMOD. [14] consists of 13 objects with watertight CAD models. Thereby, each object is accompanied by a small image sequence of around 1k frames, captured in cluttered environment with challenging lighting. Following previous works [40, 47], we employ \\(15\\%\\) of the captured RGB images for self-supervision, discarding any manual annotations, and test our model on the remaining samples." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.641, + 0.469, + 0.698 + ], + "angle": 0, + "content": "Occluded LineMOD. [3] is an extension of one LineMOD sequence showing eight different objects from LineMOD undergoing severe occlusion. For fairness we adopt the same train/test split for self-supervision as [46]." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.717, + 0.47, + 0.773 + ], + "angle": 0, + "content": "HomebrewedDB. [18] is a recently established dataset with various kinds of objects and poses. To test generalisability we test on the three objects that it shares with LineMOD [18]." + }, + { + "type": "title", + "bbox": [ + 0.065, + 0.784, + 0.277, + 0.799 + ], + "angle": 0, + "content": "4.3. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.806, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Performance on LineMOD and Occluded LineMOD In Table 1, we demonstrate our results for LineMOD with respect to the ADD(-S) metric. Note that GDR-LB and GDR-UB correspond to the performance of our base pose estimators, respectively supervised with synthetic data and real" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.905, + 0.659 + ], + "angle": 0, + "content": "data, reflecting the lower and upper performance bound. As one can easily see there is a significant performance gap between fully-supervised methods and approaches trained with synthetic data instead (c.f. DPODv2 [42], GDR [48]). Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities to achieve comparable results, as for example Self6D++ [46] with depth supervision. In contrast, our method is able to further close the domain gap whilst relying solely on RGB images, outperforming all other self-supervised approaches by at least \\(3\\%\\) for RGB-D approaches and \\(6\\%\\) for RGB-only methods. Noticeably, we are able to improve our lower bound performance from \\(77.4\\%\\) to \\(91.7\\%\\), which is even slightly better than the performance of our upper bound model trained with full supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.906, + 0.877 + ], + "angle": 0, + "content": "We would like to emphasise that our model is capable of providing significant performance boosts on highly challenging tiny and textureless objects (e.g., Ape, Duck, and Holep.) by improving the lower bound from \\(50.9\\%\\), \\(24.6\\%\\), \\(32.1\\%\\) to \\(80.9\\%\\), \\(83.4\\%\\), \\(79.3\\%\\), respectively. These objects are typically hard to handle by other self-supervised methods without real supervision (c.f. results from Self6D++ [46], Sock et al. [44], etc). As the RGB variant of Self6D++ [46] leverages DeepIM to refine the output pose in order to generate pseudo pose annotations, this process imposes a strong upper bound on the pose estimator. In fact, the performance achieved by Self6D++ is clearly inferior to DeepIM for almost all objects. This problem is even more apparent when the pose refiner is worse than the pose estimator (c.f. Holep. results from GDR [48], DeepIM [26] and Self6D++ [46])." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.503, + 0.903 + ], + "angle": 0, + "content": "4846" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.065, + 0.114, + 0.904, + 0.141 + ], + "angle": 0, + "content": "Table 1. Evaluation results on LineMOD dataset. \\*: objects with symmetry. \\*\\*: used as our pretrained pose estimator. \\(\\star\\) : use no PBR renderings for synthetic pretraining. †: re-implemented version from [46], supervision source for Self6D++. ‡: with depth supervision" + }, + { + "type": "table", + "bbox": [ + 0.068, + 0.153, + 0.907, + 0.369 + ], + "angle": 0, + "content": "
SupervisionSynSyn + RealSyn + Self
Supervision signals required in real domain
Depth--------✓---✓-
Pose Refiner-----------✓✓-
Weak labels---------✓✓---
GT Pose labels-----✓✓✓------
MethodsAAE* [45]MHP* [29]DPODv2 [42]GDR-LB** [48]DeepIM† [26]DPODv2 [42]GDR-UB [48]SO-Pose [10]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape4.011.962.150.985.880.085.0-38.937.631.276.075.480.9
Benchwise20.966.288.399.493.199.799.8-75.278.683.091.694.999
Camera30.522.492.589.299.199.296.5-36.965.649.697.197.094.8
Can35.959.896.697.299.899.699.3-65.665.656.599.899.599.7
Cat17.926.986.179.998.795.193.0-57.952.557.985.686.692.6
Driller24.044.690.198.7100.098.9100.0-67.048.873.798.898.997.4
Duck4.98.354.824.661.979.565.3-19.635.131.356.568.383.4
Eggbox*81.055.798.681.193.599.699.9-99.089.296.091.099.094.9
Glue*45.554.695.481.293.399.898.1-94.164.563.492.296.193.4
Holep.17.615.527.041.932.172.373.4-16.241.538.835.441.979.3
Iron32.060.898.298.8100.099.486.9-77.980.961.999.599.499.8
Lamp60.5-91.098.999.196.399.6-68.270.764.797.498.998.3
Phone33.834.474.364.394.896.886.3-50.160.554.491.894.378.9
Average31.438.881.277.488.093.591.096.058.960.658.685.688.591.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.065, + 0.393, + 0.471, + 0.458 + ], + "angle": 0, + "content": "Table 2. Evaluation results on Occluded LineMOD dataset. \\*: objects with symmetry. \\*\\*: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. \\(\\dagger\\):re-implemented version from [46], supervision source for Self6D++. \\(\\ddagger\\): with depth supervision." + }, + { + "type": "table", + "bbox": [ + 0.074, + 0.469, + 0.462, + 0.581 + ], + "angle": 0, + "content": "
SupervisionSynSyn + Self
MethodsGDR** [48]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape44.013.712.09.157.759.460.5
Can83.943.227.521.195.096.593.4
Cat49.118.712.026.052.660.856.1
Driller88.532.520.533.590.592.092.5
Duck15.014.423.012.226.730.655.5
Eggbox*33.957.825.139.445.051.146.0
Glue*75.054.327.037.087.188.682.8
Holep.34.022.035.020.423.538.546.5
Average52.932.122.824.859.864.766.7
" + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.6, + 0.47, + 0.756 + ], + "angle": 0, + "content": "On the contrary, our method relies on texture optimisation to synthesise realistic training samples with precise pose annotations for supervision, making self-supervision more robust towards pose errors. We further validate our scheme with another pose estimator, DPODv2 [42] in supp. mat. As for our results on Occluded LineMOD in Table 2, we can make the same observations. We again report the best overall performance of all self-supervised methods, despite not relying on any elaborate design for occlusion-awareness as Self6D++ [46], with \\(66.7\\%\\) vs \\(64.7\\%\\) and \\(59.8\\%\\) for Self6D++ [46] with and without depth data, respectively." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.764, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Performance on HomebrewDB and generalisation In this experiment, we want to study the generalisation ability of our method towards new scenes with changing illumination. From Table 3, we observe that after self-supervision on LineMOD, Self6D++ undergoes a clear performance drop compared to their respective pretrained models when testing on HomebrewDB (86.8% vs. 80.6%). This reveals that Self6D++'s self-supervision scheme is prone to forgetting." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.393, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Table 3. Evaluation results on HomebrewedDB dataset. **: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. †: Pre-process testing images with the same intrinsics as LineMOD, different from [46] as they directly feed raw images without scaling. ‡: Use raw images for re-adaptation." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.469, + 0.9, + 0.53 + ], + "angle": 0, + "content": "
SupervisionSynSyn + Self
MethodGDR** † [48]Sock et al. [44]Self6D++ † [46]Ours †Ours ‡
Benchvise88.857.375.793.192.9
Driller92.846.689.494.894.2
Phone78.741.576.879.381.2
Average86.852.080.689.189.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.907, + 0.804 + ], + "angle": 0, + "content": "Our method can instead generalise to the new scene with even boosted performance (86.8% vs. 89.1%). We attribute this to the learnt realistic textures, which are used to uniformly synthesise photorealistic data that can further be augmented with random backgrounds to increase domain invariance [11,26]. Further, we investigate the adaptation ability of our method on these new scenes. To this end, we conduct self-supervision on raw images from HomebrewedDB. We first synthesise novels view with the pre-trained texture learner from LineMOD in an effort to adapt the pose estimator to the camera from HomebrewedDB. Afterwards, we re-train the texture learner with raw images from the new target scene. Our results again illustrate that our method outperforms all other baselines. Interestingly, we notice that the choice of adaptation domain (LineMOD or HomebrewedDB) has marginal effect on our final performance, suggesting that our method has potential to close the sim-to-real domain gap even trained for one scene." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.659, + 0.828 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.835, + 0.905, + 0.879 + ], + "angle": 0, + "content": "Dealing with bad initialisation. We randomly sample 200 testing images from the Duck sequence of LineMOD (shown in Fig. 1). For the baseline render-and-compare" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.501, + 0.903 + ], + "angle": 0, + "content": "4847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.113, + 0.322, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.113, + 0.543, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.114, + 0.871, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.064, + 0.265, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Ablation of the different schemes and the required amount of data for self-supervision. Left: Pose errors after every 10 iterations for the baseline scheme (pseudo pose) and our proposed pseudo-texture scheme. Right: Self-supervision with different amount of training data." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.313, + 0.469, + 0.427 + ], + "angle": 0, + "content": "scheme (pseudo pose), we minimise \\( E_{\\text{rend}} \\) in Eqn. 1 to refine the pose with the raw initialisation from the pose estimator. For our proposed scheme (pseudo texture), we repeatedly run an evaluation after every 10 weight update steps of the pose estimator. From Fig. 3 (left), we see that the baseline scheme diverges, while our proposed scheme significantly enhances the convergence properties thanks to the synthesised supervision from the texture learner." + }, + { + "type": "table_caption", + "bbox": [ + 0.065, + 0.439, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Table 4. Ablation study on partial objects from LineMOD dataset.*: objects with symmetry." + }, + { + "type": "table", + "bbox": [ + 0.071, + 0.476, + 0.468, + 0.558 + ], + "angle": 0, + "content": "
EadvEregApeCatDuckGlue*Holep.PhoneAverage
w/o self-sup.50.979.924.681.241.964.357.1
--76.390.365.982.359.665.373.3
✓-80.489.476.590.976.570.780.7
-✓82.187.878.496.477.176.283.0
✓✓80.992.683.493.479.378.984.7
" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.587, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Impact of \\( E_{adv} \\) and \\( E_{reg} \\). We study the individual components of our proposed training objective to guide texture learning under pose noise using the six worst performing objects. As shown in Table 4, even when training the texture learner with noisy poses, we are able to significantly improve the lower bound for most objects. We attribute this behaviour to the strong synthesis capabilities of NeRF [36]. Notably, for objects with richer textures or articulation (Glue, Phone), the baseline model without their supervision yields nearly no improvements. However, when respectively introducing \\( E_{adv} \\) and \\( E_{reg} \\) to the baseline model, we observe ca. \\( 10\\% \\) absolute improvement on ADD(-S) score. Further, combining both of them yields the best performance." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.788, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Impact of training data size. As the core idea of our method is to learn the object texture from multiple views, instead of optimising per-sample pose for supervision, we believe such formulation has a potential to allow for pose learning with much less data. To this end, we conduct self-supervision with an increasing dataset size taken from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.905, + 0.427 + ], + "angle": 0, + "content": "original training split. On the right of Fig. 3, we can see that the performance keeps steadily increasing when the amount of data is enlarged. Surprisingly, we observe that already with \\(5\\%\\) of the training split (only 9 images), our method can bring a significant absolute improvement of ca. \\(18\\%\\), achieving around \\(75\\%\\) of the optimal performance. This further confirms our method can even serve as a strong baseline for few-shot learning under extreme data scarcity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.905, + 0.47 + ], + "angle": 0, + "content": "We refer interested readers to supp. mat for further detailed ablation experiments, e.g., analysis on optimisation design (Eqn. 1 and Eqn. 2) in Table 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.623, + 0.497 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.904, + 0.747 + ], + "angle": 0, + "content": "In this paper, we formulate a new learning paradigm for the task of 6D pose estimation. Separating the task into texture and pose learning enables a novel self-supervision scheme to adapt pretrained 6D pose estimators to the real domain. Only a few unlabelled images are needed within this framework to generate pixel-perfect supervision. With this reformulation, we successfully bridge the gap between synthetic pre-training and real data usage through geometry-transfer from pixel-perfect in silico data onto data-induced appearance information. It successfully addresses the difficulty in optimising pose with raw 2D observations and further prunes the need for additional supervisions as guidance. Although initialised with much inferior accuracy, it improves the state-of-the-art self-supervision RGB method by an impressive \\(6.1\\%\\) while even being superior to RGBD methods by \\(3.2\\%\\). We further demonstrates substantial generalisation improvements towards real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.905, + 0.876 + ], + "angle": 0, + "content": "For limitations, though we require as little as nine images for texture learning, they shall have mild occlusions to provide sufficient appearance information of the objects. We also share current limitations with other model-based pose estimation methods and hence need a CAD model to ensure a precise object-centric neural representation. For future works, we strongly believe that our reformulation can impact self-supervision pipelines also for unseen objects with joint optimisation of both geometry and texture." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "4848" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.067, + 0.114, + 0.165, + 0.129 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.075, + 0.138, + 0.47, + 0.228 + ], + "angle": 0, + "content": "[1] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: speeded up robust features. In Ales Leonardis, Horst Bischof, and Axel Pinz, editors, Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part I, volume 3951 of Lecture Notes in Computer Science, pages 404-417. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.075, + 0.231, + 0.472, + 0.309 + ], + "angle": 0, + "content": "[2] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 95-104. IEEE Computer Society, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.31, + 0.471, + 0.414 + ], + "angle": 0, + "content": "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In David J. Fleet, Tomás Pajdla, Bernt Schiele, and Tinne Tuyte-laars, editors, Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II, volume 8690 of Lecture Notes in Computer Science, pages 536-551. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.415, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[4] Kai Chen, Rui Cao, Stephen James, Yichuan Li, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Sim-to-real 6d object pose estimation via iterative self-training for robotic bin-picking. arXiv preprint arXiv:2204.07049, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.468, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[5] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI, volume 12371 of Lecture Notes in Computer Science, pages 139-156. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.573, + 0.471, + 0.625 + ], + "angle": 0, + "content": "[6] Alvaro Collet, Manuel Martínez, and Siddhartha S. Srinivasa. The MOPED framework: Object recognition and pose estimation for manipulation. Int. J. Robotics Res., 30(10):1284-1306, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.627, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[7] Alvaro Collet and Siddhartha S. Srinivasa. Efficient multiview object recognition and full pose estimation. In IEEE International Conference on Robotics and Automation, ICRA 2010, Anchorage, Alaska, USA, 3-7 May 2010, pages 2050-2055. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.693, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[8] Xinke Deng, Yu Xiang, Arsalan Mousavian, Clemens Eppner, Timothy Bretl, and Dieter Fox. Self-supervised 6d object pose estimation for robot manipulation. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 3665-3671. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.772, + 0.471, + 0.823 + ], + "angle": 0, + "content": "[9] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. CoRR, abs/1911.01911, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.825, + 0.471, + 0.877 + ], + "angle": 0, + "content": "[10] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting selfocclusion for direct 6d pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021," + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.138, + 0.472, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.117, + 0.905, + 0.142 + ], + "angle": 0, + "content": "Montreal, QC, Canada, October 10-17, 2021, pages 12376-12385. IEEE, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.144, + 0.905, + 0.208 + ], + "angle": 0, + "content": "[11] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1310-1319. IEEE Computer Society, 2017. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.21, + 0.904, + 0.274 + ], + "angle": 0, + "content": "[12] Stefano Gasperini, Jan Haug, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, Benjamin Busam, and Federico Tombari. Certainnet: Sampling-free uncertainty estimation for object detection. IEEE Robotics Autom. Lett., 7(2):698-705, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.904, + 0.352 + ], + "angle": 0, + "content": "[13] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 3193-3203. Computer Vision Foundation / IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[14] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Kyoung Mu Lee, Yasuyuki Matsushita, James M. Rehg, and Zhanyi Hu, editors, Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, volume 7724 of Lecture Notes in Computer Science, pages 548-562. Springer, 2012. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[15] Tomas Hodan, Pavel Haluza, Stepan Obdrzalek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.904, + 0.667 + ], + "angle": 0, + "content": "[16] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Adrien Bartoli and Andrea Fusiello, editors, Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12536 of Lecture Notes in Computer Science, pages 577-594. Springer, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.668, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[17] HyunJun Jung, Patrick Ruhkamp, Guangyao Zhai, Nikolas Brasch, Yitong Li, Yannick Verdie, Jifei Song, Yiren Zhou, Anil Armagan, Slobodan Ilic, et al. Is my depth ground-truth good enough? hammer-highly accurate multimodal dataset for dense 3d scene regression. arXiv preprint arXiv:2205.04565, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.905, + 0.823 + ], + "angle": 0, + "content": "[18] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.825, + 0.905, + 0.877 + ], + "angle": 0, + "content": "[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.117, + 0.905, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.892, + 0.502, + 0.903 + ], + "angle": 0, + "content": "4849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.116, + 0.47, + 0.142 + ], + "angle": 0, + "content": "October 22-29, 2017, pages 1530-1538. IEEE Computer Society, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.144, + 0.471, + 0.195 + ], + "angle": 0, + "content": "[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.196, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[21] Zhengfei Kuang, Kyle Olszewski, Mengei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 41(4), jul 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.248, + 0.471, + 0.327 + ], + "angle": 0, + "content": "[22] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 3559-3568. Computer Vision Foundation / IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.327, + 0.471, + 0.43 + ], + "angle": 0, + "content": "[23] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi-object 6d pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII, volume 12362 of Lecture Notes in Computer Science, pages 574-591. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.431, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1654-1663. Computer Vision Foundation / IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.51, + 0.471, + 0.562 + ], + "angle": 0, + "content": "[25] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.563, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis., 128(3):657-678, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.602, + 0.471, + 0.667 + ], + "angle": 0, + "content": "[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: bundle-adjusting neural radiance fields. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 5721-5731. IEEE, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.668, + 0.471, + 0.72 + ], + "angle": 0, + "content": "[28] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.721, + 0.471, + 0.798 + ], + "angle": 0, + "content": "[29] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6841-6850, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.799, + 0.471, + 0.876 + ], + "angle": 0, + "content": "[30] Fabian Manhardt, Wadim Kehl, and Adrien Gaidon. ROI10D: monocular lifting of 2d detection to 6d pose and metric shape. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2069-2078. Computer Vision Foundation / IEEE, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.069, + 0.116, + 0.471, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.117, + 0.903, + 0.209 + ], + "angle": 0, + "content": "[31] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in RGB. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIV, volume 11218 of Lecture Notes in Computer Science, pages 833-849. Springer, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.21, + 0.904, + 0.274 + ], + "angle": 0, + "content": "[32] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning. arXiv preprint arXiv:2003.05848, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.903, + 0.313 + ], + "angle": 0, + "content": "[33] Éric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph., 22(12):2633-2651, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.314, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7210-7219. Computer Vision Foundation / IEEE, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.904, + 0.483 + ], + "angle": 0, + "content": "[35] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6331-6341. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.904, + 0.589 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.904, + 0.655 + ], + "angle": 0, + "content": "[37] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 7667-7676. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.655, + 0.904, + 0.746 + ], + "angle": 0, + "content": "[38] Kiru Park, Timothy Patten, and Markus Vincze. Neural object learning for 6d pose estimation using a few cluttered images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV, volume 12349 of Lecture Notes in Computer Science, pages 656-673. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.904, + 0.823 + ], + "angle": 0, + "content": "[39] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4561-4570. Computer Vision Foundation / IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.824, + 0.904, + 0.877 + ], + "angle": 0, + "content": "[40] Mahdi Rad and Vincent Lepetit. BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In IEEE International Conference on Computer Vision, ICCV 2017," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.117, + 0.904, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.893, + 0.502, + 0.903 + ], + "angle": 0, + "content": "4850" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "Venice, Italy, October 22-29, 2017, pages 3848-3856. IEEE Computer Society, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.145, + 0.47, + 0.235 + ], + "angle": 0, + "content": "[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.237, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[42] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.29, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[43] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.369, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[44] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In Vitomir Struc and Francisco Gomez Fernandez, editors, 8th International Conference on 3D Vision, 3DV 2020, Virtual Event, Japan, November 25-28, 2020, pages 291-300. IEEE, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.461, + 0.471, + 0.526 + ], + "angle": 0, + "content": "[45] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.528, + 0.471, + 0.591 + ], + "angle": 0, + "content": "[46] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.594, + 0.471, + 0.698 + ], + "angle": 0, + "content": "[47] Gu Wang, Fabian Manhardt, Jianzhuon Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 108-125. Springer, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.7, + 0.471, + 0.776 + ], + "angle": 0, + "content": "[48] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16611-16621. Computer Vision Foundation / IEEE, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.779, + 0.471, + 0.869 + ], + "angle": 0, + "content": "[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2642-2651. Computer Vision Foundation / IEEE, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.069, + 0.116, + 0.471, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.117, + 0.905, + 0.195 + ], + "angle": 0, + "content": "[50] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Loranzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multimodal dataset for category-level object pose estimation with photometrically challenging objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.197, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[51] Pengyuan Wang, Fabian Manhardt, Luca Minciullo, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Demograsp: Few-shot learning for robotic grasping with human demonstration. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5733-5740. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[52] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, abs/2102.07064, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.905, + 0.406 + ], + "angle": 0, + "content": "[53] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.408, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[54] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 3907-3916. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.487, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[55] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. NeRF-Supervision: Learning dense object descriptors from neural radiance fields. In IEEE Conference on Robotics and Automation (ICRA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.905, + 0.619 + ], + "angle": 0, + "content": "[56] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. iNeRF: Inverting neural radiance fields for pose estimation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.62, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[57] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of SDF shape priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 12221-12230. Computer Vision Foundation / IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.905, + 0.764 + ], + "angle": 0, + "content": "[58] Sergey Zakharov, Wadim Kehl, and Slobodan Ilic. Deceptionnet: Network-driven domain randomization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 532-541. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.766, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[59] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6d pose object detector and refiner. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 1941-1950. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conference on Computer Vision and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.117, + 0.905, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.893, + 0.5, + 0.904 + ], + "angle": 0, + "content": "4851" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.101, + 0.117, + 0.47, + 0.155 + ], + "angle": 0, + "content": "Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5745-5753. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.47, + 0.893, + 0.501, + 0.903 + ], + "angle": 0, + "content": "4852" + } + ] +] \ No newline at end of file diff --git a/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_origin.pdf b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..13f65a14063f163cee368c44262822f2dbb3f228 --- /dev/null +++ b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/3847ca6c-d45c-45c5-ba3b-fa7d3a58ce42_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0472c207326ad5b71f51bc02b0a0070cc019e937a442f59327d6c5d2bbf912e5 +size 662259 diff --git a/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/full.md b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35862f21f8832f69340748b1bf2d10db7ff86d14 --- /dev/null +++ b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/full.md @@ -0,0 +1,305 @@ +# TexPose: Neural Texture Learning for Self-Supervised 6D Object Pose Estimation + +Hanzhi Chen $^{1}$ Fabian Manhardt $^{2}$ Nassir Navab $^{1}$ Benjamin Busam $^{1,3}$ + +$^{1}$ Technical University of Munich $^{2}$ Google Inc. $^{3}$ 3Dwe.ai + +hanzhi.chen@tum.de fabianmanhardt@google.com b.busam@tum.de + +# Abstract + +In this paper, we introduce neural texture learning for 6D object pose estimation from synthetic data and a few unlabelled real images. Our major contribution is a novel learning scheme which removes the drawbacks of previous works, namely the strong dependency on co-modalities or additional refinement. These have been previously necessary to provide training signals for convergence. We formulate such a scheme as two sub-optimisation problems on texture learning and pose learning. We separately learn to predict realistic texture of objects from real image collections and learn pose estimation from pixel-perfect synthetic data. Combining these two capabilities allows then to synthesise photorealistic novel views to supervise the pose estimator with accurate geometry. To alleviate pose noise and segmentation imperfection present during the texture learning phase, we propose a surfel-based adversarial training loss together with texture regularisation from synthetic data. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen scenes. Remarkably, our scheme improves the adopted pose estimators substantially even when initialised with much inferior performance. + +# 1. Introduction + +For spatial interaction with objects, one needs an understanding of the translation and rotation of targets within 3D space. Inferring these 6D object pose parameters from a single RGB image is a core task for 3D computer vision. This visually retrieved information has a wide range of applications in AR/VR [13, 33], autonomous driving [12, 30, 57], and robotic manipulation [24, 51, 53]. Noteworthy, accuracy and runtime have both recently made a huge leap forward thanks to deep learning [19, 23, 31, 39, 40]. Unfortunately, most of these methods heavily rely on a massive amount of labelled data for supervision to learn precise models with + +strong generalisation capabilities [16, 39, 53, 59]. However, it is very labor-intensive and time consuming to generate accurate annotations for pose data [15, 53]. Meanwhile, this process also easily suffers from labelling errors as precise annotation in 3D is highly challenging [14]. Therefore, most of the benchmarks have only few hundreds images, which does not allow proper learning of large models. In fact, methods training with such low amount of real data tend to strongly overfit to the data domain and fail to generalise to new scenes [18]. + +As a consequence, many different approaches have been proposed in the literature to tackle this problem. The simplest solution is to employ cut-and-paste strategy to increase domain invariance [11,26]. Nonetheless, this requires highly accurate manual annotations and most models tend to overfit to the original domain. Another alternative is to rely on a large amount of synthetically generated data to prevent overfitting. Though this process is relatively cheap and fast, rendered images can exhibit drastic visual discrepancy in comparison with real images even when advanced physically-based renderers [9] are used. A handful of approaches try to close the domain gap via the use of generative adversarial networks to translate the synthetic images into the real domain [2,58]. Unfortunately, these methods do not achieve promising results as the generated images are still easily distinguishable from real imagery. Notably, very recently a new line of work that proposes to self-supervise the pose estimator on real data has emerged. After training in simulation they fine-tune on the new datasets [46, 47]. While these methods achieve impressive results that are even on par with fully supervised methods, they still suffer from significant performance drop when not leveraging additional supervisory signals such as depth data [46] or ground truth camera poses [25, 42]. In this work, we propose a novel way to conduct self-supervised 6D pose estimation that is free from any additional supervision sources and further yields state-of-the-art performance. + +The core idea of previous attempts [44,46,47,54] to adapt pretrained pose estimators to the real domain is to conduct render-and-compare using a differentiable renderer. With + +CAD models and poses given, a renderer can output several attributes of the current object pose (e.g. mask, colour, depth) which allows to refine the pose through iterative comparison between the renderings and observations. However, when relying on 2D visual contents (mask and colour), such strategies tend to fail due to measured silhouette imperfection, lack of textures, and domain discrepancies. This undesirable behaviour is also discussed in [47]. + +Thus, different from the previous attempts that heavily rely on render-and-compare for self-supervision [46, 47], we instead propose to regard realistic textures of the objects as an intermediate representation before conducting training for the pose estimator. Our approach is formulated as two interconnected sub-optimisation problems on texture learning and pose learning. In the core, we first learn realistic textures of objects from raw image collections, then synthesise training data to supervise the pose estimator with pixel-perfect labels and realistic appearance. The key challenge of our proposed scheme lies in capturing accurate texture under noise introduced by poses initialised by a pretrained pose estimator during supervision. To this end, in addition to leveraging synthetic data to establish geometry priors, we learn robust supervision through adversarial training by conditioning synthesised colours on local surface information. Furthermore, we establish regularisation from synthetic textures to compensate segmentation artefacts during a texture-learning phase. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen domains. Our method significantly improves even difficult objects with little variance in appearance and geometry through self-supervision. Impressively, Our approach demonstrates a robust self-improving ability for the employed pose estimators even when initialised with much inferior pose estimates than stronger baselines [46]. + +To summarise, our main contributions are: + +- We formulate a new learning scheme, TexPose, that decomposes self-supervision for 6D object pose into Texture learning and Pose learning. +- We propose a surfel-conditioned adversarial training loss and a synthetic texture regularisation term to handle pose errors and segmentation imperfection during texture learning, further delivering self-improving ability to the pose estimators. +- We show significant improvements over recent strong baselines with additional supervision signals. Our pose estimators demonstrate a substantial generalisation ability even on unseen scenes. + +# 2. Related Work + +Model-based 6D Pose Estimation To retrieve the 6D pose, early methods use local or global features and search for key points correspondence on CAD models [1,6,7,28]. In recent years, learning-based methods dominate the field and solve the task using convolution neural networks (CNN) under the supervision from annotated data to extract deep features. There are two major approaches for pose estimation, in particular, correspondence-based and regression/classification-based approaches. Correspondence-based methods establish 2D-3D correspondences [37, 39, 40, 42, 59], prior to leveraging a variant of the RANSAC&PnP paradigm to solve for pose. Regression-based approaches, on the other hand, directly regress or classify the pose of the detected object. Initially these methods usually have shown a lower performance due to the existence of ambiguities [29] such as pose symmetries [53]. Therefore, methods like SSD-6D [19] discretise the rotation space to circumvent this issue. Recently, with better continuous representations for rotation [60], these methods gradually demonstrate high effectiveness [10, 48]. + +Self-Supervised Pose Learning Considering tremendous effort to collect large amount of annotations for 6D object pose [17, 50], several recent works have been proposed to explore the possibility of self-supervised pose learning using labeled synthetic data together with unlabelled real sensor data. [8] designed a novel labelling pipeline using a manipulator to generate reliable pose annotations for supervision. Self6D [47] proposed a self-supervision workflow by first pretraining a pose estimator with synthetic data, which was then adapted to the real world through a render-and-compare strategy by imposing consistencies between the rendered depth and sensed depth under the current pose estimate. CPS++ [32] used similar approach for categorical-level object pose estimation, and the shape is jointly deformed during optimisation. AAE [45] parameterise $SO(3)$ space using a latent embedding obtained from synthetic images, which is later employed for orientation retrieval. Both works demonstrate the strong requirement for depth data in order to allow for a reasonable performance either in training or testing. However, due to the uninformative appearance and the sim-to-real domain gap, render-and-compare strategies easily diverge when there is no depth data available as shown in [47]. Hence, consecutive works introduce other supervisory signals to prune the need for depth. DSC-PoseNet [54] employs a key point consistency regularisation for dual-scale images with labelled 2D bounding box. Socket al. [44] use photometric consistency from multiple views to refine the raw estimate, while ground-truth masks are required for supervision. Recently, Self6D++ [46] proposed to initialise the render-and-compare process with a powerful deep pose refiner [26] to guide the learning process of the pose esti + +mator, which currently yields state-of-the-art performance among all methods without manual annotations. Chen et al. [4] leverages a tailored heuristics for pseudo labelling under student-teacher learning scheme. + +Novel View Synthesis With a few images given, novel view synthesis aims to render scene content from unseen viewpoints, which can be regarded as a process to acquire "textures" of the scene. NOL [38] is proposed to address the difficulty of CAD model texturing for 6D pose estimation. Though requiring pose annotations, it demonstrates high potential to gain improvement for pose estimators with more synthesised real data in a semi-supervised fashion. Outside the scope of 6D pose, neural scene representation as implicit scene models recently get significant attention due to their ability for photorealistic novel view synthesis. NeRF [36] represent scenes using simple multi-layer perceptrons to encode the density and radiance for every 3D point. Assuming precise poses to transform views into a common reference frame, NeRF is able to capture volumes using an image reconstruction loss. NeRF-w [34] further extends NeRF to handle changing lighting and transient effects for image collections with latent embeddings. To release the rigid requirement of accurate poses, NeRF-- [52] jointly optimises camera parameters and radiance fields. iNeRF [56] performs a render-and-compare strategy using a trained NeRF for pose refinement. In contrast, BARF [27] demonstrates promising results when encountering imperfect pose initialisation by bundle-adjusting NeRF and camera poses with coarse-to-fine positional encoding scheme. GNeRF [35] formulates the learning of radiance fields and coarse pose estimates as a generative modelling process and then jointly refines both using photometric information. Recently, NeRF-Supervision [55] and NeRF-Pose [25] are proposed to leverage NeRF to learn dense descriptors or 3D representations of a given object under pose annotations when CAD model is inaccessible. + +# 3. Method + +Our objective is to conduct per-instance 6D pose learning from monocular colour images with no manual labelling effort, namely self-supervised 6D pose estimation originally defined in [47]. Since the CAD model is known as a prior, we follow the two-stage approach of previous works [44, 46, 47, 54, 56], which first pretrain a pose estimator with labelled synthetic data, and consequently conduct self-training with unannotated real images captured from a specific real scene demanding a more precise deployment (e.g. robotic bin picking). We mainly introduce the self-training process in this section. Notably, in this tuning process, pose estimator tends to suffer from forgetting problem due to little variance in illumination and backgrounds of the fed training samples. In contrast, We will show the strength of our scheme to alleviate this issue in Section 4. After pretraining we acquire + +a pose estimator $f_{\theta}$ able to initialise a coarse pose estimate $T$ and a segmentation mask $M$ for each real image $\hat{I}$ . $M$ is usually used as a pseudo silhouette cue to refine the pose $T$ with render-and-compare strategy [5,22,47,56]. + +In this section, we first revisit the commonly used strategy of previous methods in this task, then introduce our new formulation requiring no additional supervisory sources clearly advantageous over previous attempts. + +# 3.1. Revisiting Self-Supervised 6D Pose Learning + +Render-and-compare is a common approach for 6D pose self-supervision [44, 46, 47, 54, 56]. As shown in Eqn. 1, it utilises a differentiable renderer $\mathcal{R}$ fed with an estimated pose to output a rendered mask, image, $[M_r, I_r] = \mathcal{R}(T)$ . $E_{rend}$ enforces consistency between the rendered outputs $[M_r, I_r]$ and the observations $[M, \hat{I}]$ . It aims to refine the pseudo pose labels for each training sample, and further supervise pose estimator $f_{\theta}$ under $E_{pose}$ . Previous attempts are majorly focused on improving $E_{rend}$ to acquire more accurate pseudo pose annotations for better supervision. + +$$ +\begin{array}{l} \min _ {\theta , \{T _ {i}, \dots , T _ {N} \}} \sum_ {i} E _ {r e n d} (\mathcal {R} (T _ {i}), [ M _ {i}, \hat {I} _ {i} ]) \\ + \sum_ {i} ^ {\iota} E _ {p o s e} \left(f _ {\theta} \left(\hat {I} _ {i}\right), T _ {i}\right). \tag {1} \\ \end{array} +$$ + +When relying on 2D information $[M_r, I_r]$ , minimising $E_{rend}$ to improve $\{T_1, \ldots, T_N\}$ can be extremely difficult due to uninformative shape and appearance and sim-to-real visual discrepancy (e.g., Duck in Fig. 1). To compensate for the lack of information, related methods introduce additional supervision with varying difficulty levels for the acquisition (e.g., depth [47], weak labels [44, 54], deep pose refiner [46], etc.). Deep pose refiner could be acquired with the least effort as it only requires synthetic pretraining as well. However, it can introduce a strong upper bound as illustrated in [46]. The pose refiner can even lead to divergence of the pose estimator and hence lacks robustness for geometrically and photometrically challenging objects. + +# 3.2. TexPose: Learning Pose from Texture + +To address the aforementioned limitations, we reformulate self-supervised 6D pose learning as two alternating optimisation problems for a texture learner $h_{\xi}$ and a pose estimator $f_{\theta}$ as defined in Eqn. 21 and depicted in Fig. 1 (Left). + +$$ +\begin{array}{l} \min _ {\theta , \xi} \sum_ {i} E _ {t e x} \left(h _ {\xi} \circ f _ {\theta} (\hat {I} _ {i}), \hat {I} _ {i}\right) \\ + \sum_ {j} ^ {i} E _ {\text {p o s e}} \left(f _ {\theta} \circ h _ {\xi} \left(\hat {T} _ {j}\right), \hat {T} _ {j}\right). \tag {2} \\ \end{array} +$$ + +Instead of leveraging raw images as regression targets for poses, we insert object textures as an intermediate rep + +![](images/46b6b04ace8a19027146050f0825ef9a44b5b78d1e9e11de5b072c158d41798a.jpg) +Figure 1. Left: Schematic overview of TexPose. We fully explore the benefit of both synthetic and real dataset with perfect geometric labels and realistic appearance to synthesise informative train dataset for pose learning. Right: Details of the texture learner used in our scheme. We first conduct geometric pretraining to acquire geometry branch $h_{\sigma}$ to ensure consistent coordinates system and geometry aligned with the CAD model. Then, we transfer and freeze $h_{\sigma}$ to the real together with newly established radiance branch $h_c^s$ and $h_c^t$ as our texture learner. It is supervised by $E_{rec}$ , $E_{reg}$ and $E_{adv}$ capable of improving texture quality so as to synthesise more informative dataset for supervision. + +![](images/481aa02d56454925e4a47a8c7d1ef8cd8a90ed729954fe0ef9032816adc10126.jpg) + +resentation before pose learning. Specifically, we start with optimising $E_{tex}$ by fixing the parameters of pose estimator $f_{\theta}$ used to initialise poses $\{T_1,\dots ,T_N\}$ for each raw image $\{\hat{I}_1,\ldots ,\hat{I}_N\}$ , aiming to provide supervision for the realistic textures embedded by $h_\xi$ . Afterwards, we freeze $h_\xi$ to synthesise training samples $\{I_1,\dots ,I_{N'}\}$ paired with perfect labels $\{\hat{T}_1,\dots ,\hat{T}_{N'}\}$ . Gradients will thus be propagated through $f_{\theta}$ to update its parameters during the minimisation for $E_{pose}$ over these train data. We typically chose $N' > N$ . + +With this reformulation, we deliberately omit pose optimisation from 2D information usually sensitive to initialisation and prone to divergence, and bridge the domain gap directly through photorealistic synthesis. Our goal is then shifted to limit texture mapping error under the presence of pose noise so that the synthesised data can be highly informative with perfect geometric labels and as-accurate-as-possible realistic appearance. Intuitively, through multi-view supervision, it's possible to capture more accurate appearance with compensation for per-view pose noise. Our proposed strategy to further increase reliability of the synthesised data will be discussed later. Notably, different from previous attempts, no additional supervision is required in our design. The pose estimator is equipped with a self-improving ability through intermediate phase for texture learning. + +NeRF preliminaries. We leverage neural radiance fields (NeRF) [36] to embed the texture information due to its simplicity and capability of photorealistic view synthesis2. + +NeRF represents the object model with two MLPs to encode information for every queried point $\mathbf{x}$ , namely the + +geometry branch to predict volume density $\sigma$ and geometric feature $\gamma$ with $[\sigma, \gamma] = h_{\sigma}(\mathbf{x})$ and radiance branch for colour $\mathbf{c}$ prediction with $\mathbf{c} = h_c(\mathbf{x}, \gamma)$ (weight $\xi$ is omitted). When rendering from a given viewpoint, $K$ samples $\{x_k\}_1^K$ are gathered along the ray $\mathbf{r} = \mathbf{o} + t\mathbf{d}$ where $\mathbf{o}, \mathbf{d} \in \mathbb{R}^3$ represent the camera centre and the view direction respectively and $t \in \mathbb{R}$ is the distance. The colour $\mathbf{C}(\mathbf{r})$ will be estimated through numerical quadrature approximation with $w_k = \exp(-\delta_k \sigma_k)$ and $\alpha_k = \prod_{j=1}^{k-1} w_j$ with $\delta_k$ being the step size between adjacent points, and $\mathbf{C}(\mathbf{r}) = \sum_{k=1}^{K} \alpha_k (1 - w_k) \mathbf{c_k}$ . Rendering for depth $D(\mathbf{r})$ and mask $M(\mathbf{r})$ is achieved by modifying integration element (see supp. mat. for details). + +Geometric pretraining. Though geometry and radiance branch can be trained jointly as shown in NeRF and its variants [21,34,36], the pose for supervision in our scenario is highly noisy, and the coordinate system of NeRF can end up as an arbitrary one and further cause erroneous labels for synthesised data. We hence opt to pretrain geometry branch $h_{\sigma}$ using pixel-perfect data rendered by the CAD model with accurate pose $\hat{T}$ , depth $\hat{D}$ and mask $\hat{M}$ to supervise NeRF following Eqn. 4, where $E_{d}$ is scale-invariant loss from [20], $\lambda_{m}$ and $\lambda_{d}$ are weighting factors. + +$$ +E _ {2 d} (\mathbf {r}) = \left\| \mathbf {C} (\mathbf {r}) - \hat {\mathbf {C}} (\mathbf {r}) \right\| _ {2} ^ {2} + \lambda_ {m} \left\| M (\mathbf {r}) - \hat {M} (\mathbf {r}) \right\| _ {2} ^ {2} \tag {3} +$$ + +$$ +E _ {p r e} (\mathbf {r}) = E _ {2 d} (\mathbf {r}) + \lambda_ {d} E _ {d} (D (\mathbf {r}), \hat {D} (\mathbf {r})). \tag {4} +$$ + +Notably, though our current goal is to establish a geometry branch with precise density output "aligned" with the CAD model, training for radiance branch is also included, though + +it will be discarded afterwards. We find this greatly helps to extract meaningful geometric features fed into the radiance branch, and further eases the effort for realistic appearance transfer in the later texture learning stage. We quantitatively detail the necessity of this stage in supp. mat. + +Texture learning from noisy poses. Our current goal is to learn texture encoded by the radiance branch $h_c$ using raw images paired with poses initialised by the pose estimator $f_{\theta}$ . Note in this stage, we only re-optimise parameters from $h_c$ with $h_{\sigma}$ being fixed to capture accurate texture and meanwhile ensure no shift of coordinate system and geometry. Though jointly optimising textures and poses through photometric consistency has a potential to improve both aspects as shown in [21, 27, 56], we find it can easily fail when dealing with the textureless objects. Alternatively, we seek to supervise the texture learner with higher pose error tolerance leveraging stronger radiance branch and robust loss. + +Different from the vanilla NeRF used in geometric pretraining, we now employ two radiance branches, $h_c^s$ and $h_c^t$ , to translate the extracted geometric features into static radiance (accurate underlying textures) and transient radiance (texture outliers caused by pose errors) respectively. Apart from geometric features $\gamma$ , we further condition it on view directions $\mathbf{d}$ , lighting embedding $\tau^a$ and transient embedding $\tau^t$ [34] (see Fig. 1 Right). For each view, we use $\tau^a$ to capture illumination variations and $\tau^t$ to produce transient effect caused by pose noise. Patch-based training strategy [41] is used to preserve the spatial structures for robust loss supervision. An image reconstruction loss $E_{rec}$ is incorporated with uncertainty awareness to learn textures robustly. We refer to the supp. mat. for the details of the radiance branches and uncertainty-aware $E_{rec}$ . + +We noticed that supervision from $E_{rec}$ is still not adequate to capture accurate texture under pose noise. We further design a surfel-conditioned adversarial loss for sampled patches to tolerate texture shifts caused by pose errors. For each location on the object surface, we compute normalised object coordinate $S_{x}$ [49] and normal $S_{n}$ with current pose estimate $T$ to represent its local information. By conditioning the predicted patch colour $P$ on its local surface information $[S_{x}, S_{n}]$ , we employ an adversarial scheme: + +$$ +\begin{array}{l} E _ {a d v} (P) = \mathbb {E} _ {[ S _ {x}, S _ {n} ], P} (1 - \log D ([ S _ {x}, S _ {n} ], P)) \tag {5} \\ + \mathbb {E} _ {[ S _ {x}, S _ {n} ], \hat {P}} (\log D ([ S _ {x}, S _ {n} ], M \hat {P})). \\ \end{array} +$$ + +We also observed the imperfect predicted mask $M$ can cause background colour artefacts within the object. We therefore establish a regularisation from synthetic cues to compensate the boundary imperfections by generating synthetic patch $P_{syn}$ and mask $M_{syn}$ with current pose estimate $T$ , then applying erosion on the predicted mask $M$ to compute a padded mask with $M_{pad} = M_{syn}(1 - M)$ . The foreground + +boundaries of the real images for supervision are therefore padded with the synthetic colours. This regularisation $E_{feat}$ is realised by minimising the MSE error of deep features extracted by a pretrained VGG19 network [43] to align the boundary features of the predicted patch $P$ with the synthetic patch $P_{syn}$ in a more abstract level. The second part of Eqn. 6 ensures that rendered patch region within the mask $M$ are not affected by the synthetic information. + +$$ +\begin{array}{l} E _ {r e g} (P) = E _ {f e a t} \left(P, M \hat {P} + M _ {p a d} P _ {s y n}\right) \tag {6} \\ + \lambda_ {f g} E _ {\text {f e a t}} (M P + (1 - M) \hat {P}, \hat {P}). \\ \end{array} +$$ + +Formally, the loss function $E_{tex}$ for the texture learner supervision is summarised as: + +$$ +E _ {t e x} = E _ {r e c} + \lambda_ {a d v} E _ {a d v} + \lambda_ {r e g} E _ {r e g}. \tag {7} +$$ + +where $\lambda_{adv}$ and $\lambda_{reg}$ are weighting factors. + +Pose learning. The entire pipeline runs by first optimising $E_{tex}$ , and subsequent minimisation of $E_{pose}$ with the synthesised dataset as depicted in Fig. 1 (Left). As our scheme operates at data level and thus method-agnostic, $E_{pose}$ can be generalised across methods with different loss designs, e.g., direct pose or correspondence regression. Interestingly, we observed that the pose estimator can converge instantly with only one optimisation loop on $E_{tex}$ and $E_{pose}$ , while adding more loops, i.e., continue to optimise $E_{tex}$ with improved poses after $E_{pose}$ , only brings insignificant improvements. We attribute this to the pixel-perfect supervision from texture learner and its strong ability to mitigate pose errors. + +# 4. Experiments + +In this section we first introduce the employed metrics and dataset before presenting our results for the task of 6D pose estimation, then provide detailed analysis of our proposed scheme in ablation study. + +# 4.1. Evaluation Metrics + +In line with previous work, we present our results using the ADD(-S) metric, as proposed in [14]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. If the average distance is below $10\%$ of the object diameter, the estimated pose will be considered correct. For symmetric objects, ADD score is modified to measure the distance to the nearest vertices, here referred to as ADD-S. + +# 4.2. Dataset + +Synthetic BOP PBR. We pretrain our models using synthetic data from [9, 16], employing physically-based rendering (PBR) techniques. Although PBR can capture more + +![](images/177bbbc8c78199e5df482a9ac1b0c2e08b0019eed85e60a21923fcf01dbb737e.jpg) + +![](images/9fe9c29f3748510ec1d002e5500d9f1f6e988638dfd82c0f05f7755fc702f3c0.jpg) + +![](images/7651a735986fa4e359e8a7e88ee5b665da02f1747dcbc86059d31231992a4451.jpg) + +![](images/0b8fa52066ebefc90d28ec1532337a763a35649d36376512213c664ac881c2d4.jpg) + +![](images/df10ff94f8b2898c3b1c8438d586ef901187e9786d3a669d1614f4af5501b7cd.jpg) + +![](images/799cfbe78b9509ebbc881fda5832ef8212c7933cca2caaf17238c47e580bb3dc.jpg) + +![](images/013f191936a71f8586a7a541ca96034d2103e224b9063b692b272005ca7f10a9.jpg) + +![](images/507d4882d86a864f6af4a6ab2563c14743dd8941fe913d3f1485a2cc97d0b0c5.jpg) + +![](images/9cff58455bebe7043dfd341ab15c378f9300358d933553041d035ec68dc20c3f.jpg) + +![](images/2d50e06d3f4e19b83656249dfea0707f890f99414ef688097b7f55518a19d0ea.jpg) + +![](images/b5f7d70aab0d442004ec9992473263371d6dfff30755ae082ad4b672cb6ccfd9.jpg) +Figure 2. Qualitative results on 5 objects from LineMOD. Top: Renderings of the reconstructed CAD models which are part of LineMOD. Middle: Same objects rendered with our learnt textures. Bottom: Demonstration of our pose estimation quality. Thereby, the green, red, and blue boxes depict ground-truth pose as well as estimated pose before and after our self-supervision, respectively. + +![](images/5587bfdd5289abb4a6481f785d6863ecdf779df6898e1c51a3ab75641d7c430b.jpg) + +![](images/329d0a5f5526d413c97073e5bcd88bd2c77ac2421491d16fcbf49d0dceefa47c.jpg) + +![](images/8d30778e7ea971753d7effb6abc9e27da7a6b14e8f8c4ba5a57206b4899c92d1.jpg) + +![](images/a136a2b687596b3587a7739b0a52cb2a913997ba4b3cceb0f2c4760b40565d33.jpg) + +realistic lighting conditions, there still exhibits a large sim-to-real domain gap. Such discrepancies can strongly degrade the pose estimation performance when being deployed in the real world. + +LineMOD. [14] consists of 13 objects with watertight CAD models. Thereby, each object is accompanied by a small image sequence of around 1k frames, captured in cluttered environment with challenging lighting. Following previous works [40, 47], we employ $15\%$ of the captured RGB images for self-supervision, discarding any manual annotations, and test our model on the remaining samples. + +Occluded LineMOD. [3] is an extension of one LineMOD sequence showing eight different objects from LineMOD undergoing severe occlusion. For fairness we adopt the same train/test split for self-supervision as [46]. + +HomebrewedDB. [18] is a recently established dataset with various kinds of objects and poses. To test generalisability we test on the three objects that it shares with LineMOD [18]. + +# 4.3. Performance Analysis + +Performance on LineMOD and Occluded LineMOD In Table 1, we demonstrate our results for LineMOD with respect to the ADD(-S) metric. Note that GDR-LB and GDR-UB correspond to the performance of our base pose estimators, respectively supervised with synthetic data and real + +data, reflecting the lower and upper performance bound. As one can easily see there is a significant performance gap between fully-supervised methods and approaches trained with synthetic data instead (c.f. DPODv2 [42], GDR [48]). Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities to achieve comparable results, as for example Self6D++ [46] with depth supervision. In contrast, our method is able to further close the domain gap whilst relying solely on RGB images, outperforming all other self-supervised approaches by at least $3\%$ for RGB-D approaches and $6\%$ for RGB-only methods. Noticeably, we are able to improve our lower bound performance from $77.4\%$ to $91.7\%$ , which is even slightly better than the performance of our upper bound model trained with full supervision. + +We would like to emphasise that our model is capable of providing significant performance boosts on highly challenging tiny and textureless objects (e.g., Ape, Duck, and Holep.) by improving the lower bound from $50.9\%$ , $24.6\%$ , $32.1\%$ to $80.9\%$ , $83.4\%$ , $79.3\%$ , respectively. These objects are typically hard to handle by other self-supervised methods without real supervision (c.f. results from Self6D++ [46], Sock et al. [44], etc). As the RGB variant of Self6D++ [46] leverages DeepIM to refine the output pose in order to generate pseudo pose annotations, this process imposes a strong upper bound on the pose estimator. In fact, the performance achieved by Self6D++ is clearly inferior to DeepIM for almost all objects. This problem is even more apparent when the pose refiner is worse than the pose estimator (c.f. Holep. results from GDR [48], DeepIM [26] and Self6D++ [46]). + +Table 1. Evaluation results on LineMOD dataset. \*: objects with symmetry. \*\*: used as our pretrained pose estimator. $\star$ : use no PBR renderings for synthetic pretraining. †: re-implemented version from [46], supervision source for Self6D++. ‡: with depth supervision + +
SupervisionSynSyn + RealSyn + Self
Supervision signals required in real domain
Depth--------✓---✓-
Pose Refiner-----------✓✓-
Weak labels---------✓✓---
GT Pose labels-----✓✓✓------
MethodsAAE* [45]MHP* [29]DPODv2 [42]GDR-LB** [48]DeepIM† [26]DPODv2 [42]GDR-UB [48]SO-Pose [10]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape4.011.962.150.985.880.085.0-38.937.631.276.075.480.9
Benchwise20.966.288.399.493.199.799.8-75.278.683.091.694.999
Camera30.522.492.589.299.199.296.5-36.965.649.697.197.094.8
Can35.959.896.697.299.899.699.3-65.665.656.599.899.599.7
Cat17.926.986.179.998.795.193.0-57.952.557.985.686.692.6
Driller24.044.690.198.7100.098.9100.0-67.048.873.798.898.997.4
Duck4.98.354.824.661.979.565.3-19.635.131.356.568.383.4
Eggbox*81.055.798.681.193.599.699.9-99.089.296.091.099.094.9
Glue*45.554.695.481.293.399.898.1-94.164.563.492.296.193.4
Holep.17.615.527.041.932.172.373.4-16.241.538.835.441.979.3
Iron32.060.898.298.8100.099.486.9-77.980.961.999.599.499.8
Lamp60.5-91.098.999.196.399.6-68.270.764.797.498.998.3
Phone33.834.474.364.394.896.886.3-50.160.554.491.894.378.9
Average31.438.881.277.488.093.591.096.058.960.658.685.688.591.7
+ +Table 2. Evaluation results on Occluded LineMOD dataset. \*: objects with symmetry. \*\*: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. $\dagger$ :re-implemented version from [46], supervision source for Self6D++. $\ddagger$ : with depth supervision. + +
SupervisionSynSyn + Self
MethodsGDR** [48]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape44.013.712.09.157.759.460.5
Can83.943.227.521.195.096.593.4
Cat49.118.712.026.052.660.856.1
Driller88.532.520.533.590.592.092.5
Duck15.014.423.012.226.730.655.5
Eggbox*33.957.825.139.445.051.146.0
Glue*75.054.327.037.087.188.682.8
Holep.34.022.035.020.423.538.546.5
Average52.932.122.824.859.864.766.7
+ +On the contrary, our method relies on texture optimisation to synthesise realistic training samples with precise pose annotations for supervision, making self-supervision more robust towards pose errors. We further validate our scheme with another pose estimator, DPODv2 [42] in supp. mat. As for our results on Occluded LineMOD in Table 2, we can make the same observations. We again report the best overall performance of all self-supervised methods, despite not relying on any elaborate design for occlusion-awareness as Self6D++ [46], with $66.7\%$ vs $64.7\%$ and $59.8\%$ for Self6D++ [46] with and without depth data, respectively. + +Performance on HomebrewDB and generalisation In this experiment, we want to study the generalisation ability of our method towards new scenes with changing illumination. From Table 3, we observe that after self-supervision on LineMOD, Self6D++ undergoes a clear performance drop compared to their respective pretrained models when testing on HomebrewDB (86.8% vs. 80.6%). This reveals that Self6D++'s self-supervision scheme is prone to forgetting. + +Table 3. Evaluation results on HomebrewedDB dataset. **: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. †: Pre-process testing images with the same intrinsics as LineMOD, different from [46] as they directly feed raw images without scaling. ‡: Use raw images for re-adaptation. + +
SupervisionSynSyn + Self
MethodGDR** † [48]Sock et al. [44]Self6D++ † [46]Ours †Ours ‡
Benchvise88.857.375.793.192.9
Driller92.846.689.494.894.2
Phone78.741.576.879.381.2
Average86.852.080.689.189.4
+ +Our method can instead generalise to the new scene with even boosted performance (86.8% vs. 89.1%). We attribute this to the learnt realistic textures, which are used to uniformly synthesise photorealistic data that can further be augmented with random backgrounds to increase domain invariance [11,26]. Further, we investigate the adaptation ability of our method on these new scenes. To this end, we conduct self-supervision on raw images from HomebrewedDB. We first synthesise novels view with the pre-trained texture learner from LineMOD in an effort to adapt the pose estimator to the camera from HomebrewedDB. Afterwards, we re-train the texture learner with raw images from the new target scene. Our results again illustrate that our method outperforms all other baselines. Interestingly, we notice that the choice of adaptation domain (LineMOD or HomebrewedDB) has marginal effect on our final performance, suggesting that our method has potential to close the sim-to-real domain gap even trained for one scene. + +# 4.4. Ablation Study + +Dealing with bad initialisation. We randomly sample 200 testing images from the Duck sequence of LineMOD (shown in Fig. 1). For the baseline render-and-compare + +![](images/7c30065e74a9aeeb477ff190a737d0bfc3591f2e0e5bf5501a4bd462e1bc6663.jpg) +Figure 3. Ablation of the different schemes and the required amount of data for self-supervision. Left: Pose errors after every 10 iterations for the baseline scheme (pseudo pose) and our proposed pseudo-texture scheme. Right: Self-supervision with different amount of training data. + +![](images/329c65a021266d7e8a6a177438467c0ea819bf03b84b7cf10854ead6586a674b.jpg) + +![](images/68cb7e21ea46b56466eb03d6e9a13153e720613db22a76d8983ee68a1b1f6bec.jpg) + +scheme (pseudo pose), we minimise $E_{\text{rend}}$ in Eqn. 1 to refine the pose with the raw initialisation from the pose estimator. For our proposed scheme (pseudo texture), we repeatedly run an evaluation after every 10 weight update steps of the pose estimator. From Fig. 3 (left), we see that the baseline scheme diverges, while our proposed scheme significantly enhances the convergence properties thanks to the synthesised supervision from the texture learner. + +Table 4. Ablation study on partial objects from LineMOD dataset.*: objects with symmetry. + +
EadvEregApeCatDuckGlue*Holep.PhoneAverage
w/o self-sup.50.979.924.681.241.964.357.1
--76.390.365.982.359.665.373.3
✓-80.489.476.590.976.570.780.7
-✓82.187.878.496.477.176.283.0
✓✓80.992.683.493.479.378.984.7
+ +Impact of $E_{adv}$ and $E_{reg}$ . We study the individual components of our proposed training objective to guide texture learning under pose noise using the six worst performing objects. As shown in Table 4, even when training the texture learner with noisy poses, we are able to significantly improve the lower bound for most objects. We attribute this behaviour to the strong synthesis capabilities of NeRF [36]. Notably, for objects with richer textures or articulation (Glue, Phone), the baseline model without their supervision yields nearly no improvements. However, when respectively introducing $E_{adv}$ and $E_{reg}$ to the baseline model, we observe ca. $10\%$ absolute improvement on ADD(-S) score. Further, combining both of them yields the best performance. + +Impact of training data size. As the core idea of our method is to learn the object texture from multiple views, instead of optimising per-sample pose for supervision, we believe such formulation has a potential to allow for pose learning with much less data. To this end, we conduct self-supervision with an increasing dataset size taken from the + +original training split. On the right of Fig. 3, we can see that the performance keeps steadily increasing when the amount of data is enlarged. Surprisingly, we observe that already with $5\%$ of the training split (only 9 images), our method can bring a significant absolute improvement of ca. $18\%$ , achieving around $75\%$ of the optimal performance. This further confirms our method can even serve as a strong baseline for few-shot learning under extreme data scarcity. + +We refer interested readers to supp. mat for further detailed ablation experiments, e.g., analysis on optimisation design (Eqn. 1 and Eqn. 2) in Table 7. + +# 5. Conclusion + +In this paper, we formulate a new learning paradigm for the task of 6D pose estimation. Separating the task into texture and pose learning enables a novel self-supervision scheme to adapt pretrained 6D pose estimators to the real domain. Only a few unlabelled images are needed within this framework to generate pixel-perfect supervision. With this reformulation, we successfully bridge the gap between synthetic pre-training and real data usage through geometry-transfer from pixel-perfect in silico data onto data-induced appearance information. It successfully addresses the difficulty in optimising pose with raw 2D observations and further prunes the need for additional supervisions as guidance. Although initialised with much inferior accuracy, it improves the state-of-the-art self-supervision RGB method by an impressive $6.1\%$ while even being superior to RGBD methods by $3.2\%$ . We further demonstrates substantial generalisation improvements towards real-world scenarios. + +For limitations, though we require as little as nine images for texture learning, they shall have mild occlusions to provide sufficient appearance information of the objects. We also share current limitations with other model-based pose estimation methods and hence need a CAD model to ensure a precise object-centric neural representation. For future works, we strongly believe that our reformulation can impact self-supervision pipelines also for unseen objects with joint optimisation of both geometry and texture. + +# References + +[1] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: speeded up robust features. In Ales Leonardis, Horst Bischof, and Axel Pinz, editors, Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part I, volume 3951 of Lecture Notes in Computer Science, pages 404-417. Springer, 2006. 2 +[2] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 95-104. IEEE Computer Society, 2017. 1 +[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In David J. Fleet, Tomás Pajdla, Bernt Schiele, and Tinne Tuyte-laars, editors, Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II, volume 8690 of Lecture Notes in Computer Science, pages 536-551. Springer, 2014. 6 +[4] Kai Chen, Rui Cao, Stephen James, Yichuan Li, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Sim-to-real 6d object pose estimation via iterative self-training for robotic bin-picking. arXiv preprint arXiv:2204.07049, 2022. 3 +[5] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI, volume 12371 of Lecture Notes in Computer Science, pages 139-156. Springer, 2020. 3 +[6] Alvaro Collet, Manuel Martínez, and Siddhartha S. Srinivasa. The MOPED framework: Object recognition and pose estimation for manipulation. Int. J. Robotics Res., 30(10):1284-1306, 2011. 2 +[7] Alvaro Collet and Siddhartha S. Srinivasa. Efficient multiview object recognition and full pose estimation. In IEEE International Conference on Robotics and Automation, ICRA 2010, Anchorage, Alaska, USA, 3-7 May 2010, pages 2050-2055. IEEE, 2010. 2 +[8] Xinke Deng, Yu Xiang, Arsalan Mousavian, Clemens Eppner, Timothy Bretl, and Dieter Fox. Self-supervised 6d object pose estimation for robot manipulation. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 3665-3671. IEEE, 2020. 2 +[9] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. CoRR, abs/1911.01911, 2019. 1, 5 +[10] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting selfocclusion for direct 6d pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, + +Montreal, QC, Canada, October 10-17, 2021, pages 12376-12385. IEEE, 2021. 2, 7 +[11] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1310-1319. IEEE Computer Society, 2017. 1, 7 +[12] Stefano Gasperini, Jan Haug, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, Benjamin Busam, and Federico Tombari. Certainnet: Sampling-free uncertainty estimation for object detection. IEEE Robotics Autom. Lett., 7(2):698-705, 2022. 1 +[13] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 3193-3203. Computer Vision Foundation / IEEE, 2020. 1 +[14] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Kyoung Mu Lee, Yasuyuki Matsushita, James M. Rehg, and Zhanyi Hu, editors, Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, volume 7724 of Lecture Notes in Computer Science, pages 548-562. Springer, 2012. 1, 5, 6 +[15] Tomas Hodan, Pavel Haluza, Stepan Obdrzalek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 1 +[16] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Adrien Bartoli and Andrea Fusiello, editors, Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12536 of Lecture Notes in Computer Science, pages 577-594. Springer, 2020. 1, 5 +[17] HyunJun Jung, Patrick Ruhkamp, Guangyao Zhai, Nikolas Brasch, Yitong Li, Yannick Verdie, Jifei Song, Yiren Zhou, Anil Armagan, Slobodan Ilic, et al. Is my depth ground-truth good enough? hammer-highly accurate multimodal dataset for dense 3d scene regression. arXiv preprint arXiv:2205.04565, 2022. 2 +[18] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 1, 6 +[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, + +October 22-29, 2017, pages 1530-1538. IEEE Computer Society, 2017. 1, 2 +[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 4 +[21] Zhengfei Kuang, Kyle Olszewski, Mengei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 41(4), jul 2022. 4, 5 +[22] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 3559-3568. Computer Vision Foundation / IEEE Computer Society, 2018. 3 +[23] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi-object 6d pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII, volume 12362 of Lecture Notes in Computer Science, pages 574-591. Springer, 2020. 1 +[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1654-1663. Computer Vision Foundation / IEEE, 2021. 1 +[25] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022. 1, 3 +[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis., 128(3):657-678, 2020. 1, 2, 6, 7 +[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: bundle-adjusting neural radiance fields. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 5721-5731. IEEE, 2021. 3, 5 +[28] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2 +[29] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6841-6850, 2019. 2, 7 +[30] Fabian Manhardt, Wadim Kehl, and Adrien Gaidon. ROI10D: monocular lifting of 2d detection to 6d pose and metric shape. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2069-2078. Computer Vision Foundation / IEEE, 2019. 1 + +[31] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in RGB. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIV, volume 11218 of Lecture Notes in Computer Science, pages 833-849. Springer, 2018. 1 +[32] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning. arXiv preprint arXiv:2003.05848, 2020. 2 +[33] Éric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph., 22(12):2633-2651, 2016. 1 +[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7210-7219. Computer Vision Foundation / IEEE, 2021. 3, 4, 5 +[35] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6331-6341. IEEE, 2021. 3 +[36] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 8 +[37] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 7667-7676. IEEE, 2019. 2 +[38] Kiru Park, Timothy Patten, and Markus Vincze. Neural object learning for 6d pose estimation using a few cluttered images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV, volume 12349 of Lecture Notes in Computer Science, pages 656-673. Springer, 2020. 3 +[39] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4561-4570. Computer Vision Foundation / IEEE, 2019. 1, 2 +[40] Mahdi Rad and Vincent Lepetit. BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In IEEE International Conference on Computer Vision, ICCV 2017, + +Venice, Italy, October 22-29, 2017, pages 3848-3856. IEEE Computer Society, 2017. 1, 2, 6 +[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 5 +[42] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 6, 7 +[43] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5 +[44] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In Vitomir Struc and Francisco Gomez Fernandez, editors, 8th International Conference on 3D Vision, 3DV 2020, Virtual Event, Japan, November 25-28, 2020, pages 291-300. IEEE, 2020. 1, 2, 3, 6, 7 +[45] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2, 7 +[46] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 3, 6, 7 +[47] Gu Wang, Fabian Manhardt, Jianzhuon Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 108-125. Springer, 2020. 1, 2, 3, 6, 7 +[48] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16611-16621. Computer Vision Foundation / IEEE, 2021. 2, 6, 7 +[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2642-2651. Computer Vision Foundation / IEEE, 2019. 5 + +[50] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Loranzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multimodal dataset for category-level object pose estimation with photometrically challenging objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[51] Pengyuan Wang, Fabian Manhardt, Luca Minciullo, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Demograsp: Few-shot learning for robotic grasping with human demonstration. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5733-5740. IEEE, 2021. 1 +[52] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, abs/2102.07064, 2021. 3 +[53] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 1, 2 +[54] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 3907-3916. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 7 +[55] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. NeRF-Supervision: Learning dense object descriptors from neural radiance fields. In IEEE Conference on Robotics and Automation (ICRA), 2022. 3 +[56] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. iNeRF: Inverting neural radiance fields for pose estimation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 3, 5 +[57] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of SDF shape priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 12221-12230. Computer Vision Foundation / IEEE, 2020. 1 +[58] Sergey Zakharov, Wadim Kehl, and Slobodan Ilic. Deceptionnet: Network-driven domain randomization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 532-541. IEEE, 2019. 1 +[59] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6d pose object detector and refiner. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 1941-1950. IEEE, 2019. 1, 2 +[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conference on Computer Vision and \ No newline at end of file diff --git a/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/images.zip b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..afb38e8fac2ebdaaec3b7b30ec7e184b6b799da2 --- /dev/null +++ b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc6845654f4eafb24e599fc0ca1d5fbd3cb5288f8bf34e572dd53853d5d984c4 +size 436193 diff --git a/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/layout.json b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4780b6bb0ed45645568002055b656544c364486c --- /dev/null +++ b/2023/TexPose_ Neural Texture Learning for Self-Supervised 6D Object Pose Estimation/layout.json @@ -0,0 +1,9726 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 127, + 456, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 127, + 456, + 163 + ], + "spans": [ + { + "bbox": [ + 121, + 127, + 456, + 163 + ], + "type": "text", + "content": "TexPose: Neural Texture Learning for Self-Supervised 6D Object Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "text", + "content": "Hanzhi Chen" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "text", + "content": " Fabian Manhardt" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "text", + "content": " Nassir Navab" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "text", + "content": " Benjamin Busam" + }, + { + "bbox": [ + 105, + 185, + 471, + 200 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "spans": [ + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "text", + "content": " Technical University of Munich " + }, + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "text", + "content": " Google Inc. " + }, + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 200, + 439, + 214 + ], + "type": "text", + "content": " 3Dwe.ai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 217, + 455, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 217, + 455, + 228 + ], + "spans": [ + { + "bbox": [ + 117, + 217, + 455, + 228 + ], + "type": "text", + "content": "hanzhi.chen@tum.de fabianmanhardt@google.com b.busam@tum.de" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 255, + 183, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 255, + 183, + 268 + ], + "spans": [ + { + "bbox": [ + 135, + 255, + 183, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 39, + 282, + 281, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 282, + 281, + 559 + ], + "spans": [ + { + "bbox": [ + 39, + 282, + 281, + 559 + ], + "type": "text", + "content": "In this paper, we introduce neural texture learning for 6D object pose estimation from synthetic data and a few unlabelled real images. Our major contribution is a novel learning scheme which removes the drawbacks of previous works, namely the strong dependency on co-modalities or additional refinement. These have been previously necessary to provide training signals for convergence. We formulate such a scheme as two sub-optimisation problems on texture learning and pose learning. We separately learn to predict realistic texture of objects from real image collections and learn pose estimation from pixel-perfect synthetic data. Combining these two capabilities allows then to synthesise photorealistic novel views to supervise the pose estimator with accurate geometry. To alleviate pose noise and segmentation imperfection present during the texture learning phase, we propose a surfel-based adversarial training loss together with texture regularisation from synthetic data. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen scenes. Remarkably, our scheme improves the adopted pose estimators substantially even when initialised with much inferior performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 39, + 585, + 119, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 585, + 119, + 597 + ], + "spans": [ + { + "bbox": [ + 39, + 585, + 119, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 38, + 607, + 280, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 607, + 280, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 607, + 280, + 740 + ], + "type": "text", + "content": "For spatial interaction with objects, one needs an understanding of the translation and rotation of targets within 3D space. Inferring these 6D object pose parameters from a single RGB image is a core task for 3D computer vision. This visually retrieved information has a wide range of applications in AR/VR [13, 33], autonomous driving [12, 30, 57], and robotic manipulation [24, 51, 53]. Noteworthy, accuracy and runtime have both recently made a huge leap forward thanks to deep learning [19, 23, 31, 39, 40]. Unfortunately, most of these methods heavily rely on a massive amount of labelled data for supervision to learn precise models with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 256, + 538, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 256, + 538, + 377 + ], + "spans": [ + { + "bbox": [ + 296, + 256, + 538, + 377 + ], + "type": "text", + "content": "strong generalisation capabilities [16, 39, 53, 59]. However, it is very labor-intensive and time consuming to generate accurate annotations for pose data [15, 53]. Meanwhile, this process also easily suffers from labelling errors as precise annotation in 3D is highly challenging [14]. Therefore, most of the benchmarks have only few hundreds images, which does not allow proper learning of large models. In fact, methods training with such low amount of real data tend to strongly overfit to the data domain and fail to generalise to new scenes [18]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 378, + 538, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 378, + 538, + 701 + ], + "spans": [ + { + "bbox": [ + 296, + 378, + 538, + 701 + ], + "type": "text", + "content": "As a consequence, many different approaches have been proposed in the literature to tackle this problem. The simplest solution is to employ cut-and-paste strategy to increase domain invariance [11,26]. Nonetheless, this requires highly accurate manual annotations and most models tend to overfit to the original domain. Another alternative is to rely on a large amount of synthetically generated data to prevent overfitting. Though this process is relatively cheap and fast, rendered images can exhibit drastic visual discrepancy in comparison with real images even when advanced physically-based renderers [9] are used. A handful of approaches try to close the domain gap via the use of generative adversarial networks to translate the synthetic images into the real domain [2,58]. Unfortunately, these methods do not achieve promising results as the generated images are still easily distinguishable from real imagery. Notably, very recently a new line of work that proposes to self-supervise the pose estimator on real data has emerged. After training in simulation they fine-tune on the new datasets [46, 47]. While these methods achieve impressive results that are even on par with fully supervised methods, they still suffer from significant performance drop when not leveraging additional supervisory signals such as depth data [46] or ground truth camera poses [25, 42]. In this work, we propose a novel way to conduct self-supervised 6D pose estimation that is free from any additional supervision sources and further yields state-of-the-art performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 703, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 703, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 703, + 538, + 740 + ], + "type": "text", + "content": "The core idea of previous attempts [44,46,47,54] to adapt pretrained pose estimators to the real domain is to conduct render-and-compare using a differentiable renderer. With" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "spans": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 30, + 476, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 30, + 476, + 64 + ], + "spans": [ + { + "bbox": [ + 146, + 30, + 476, + 64 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "text", + "content": "4841" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 96, + 279, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 96, + 279, + 193 + ], + "spans": [ + { + "bbox": [ + 38, + 96, + 279, + 193 + ], + "type": "text", + "content": "CAD models and poses given, a renderer can output several attributes of the current object pose (e.g. mask, colour, depth) which allows to refine the pose through iterative comparison between the renderings and observations. However, when relying on 2D visual contents (mask and colour), such strategies tend to fail due to measured silhouette imperfection, lack of textures, and domain discrepancies. This undesirable behaviour is also discussed in [47]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 195, + 281, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 195, + 281, + 530 + ], + "spans": [ + { + "bbox": [ + 38, + 195, + 281, + 530 + ], + "type": "text", + "content": "Thus, different from the previous attempts that heavily rely on render-and-compare for self-supervision [46, 47], we instead propose to regard realistic textures of the objects as an intermediate representation before conducting training for the pose estimator. Our approach is formulated as two interconnected sub-optimisation problems on texture learning and pose learning. In the core, we first learn realistic textures of objects from raw image collections, then synthesise training data to supervise the pose estimator with pixel-perfect labels and realistic appearance. The key challenge of our proposed scheme lies in capturing accurate texture under noise introduced by poses initialised by a pretrained pose estimator during supervision. To this end, in addition to leveraging synthetic data to establish geometry priors, we learn robust supervision through adversarial training by conditioning synthesised colours on local surface information. Furthermore, we establish regularisation from synthetic textures to compensate segmentation artefacts during a texture-learning phase. We demonstrate that the proposed approach significantly outperforms the recent state-of-the-art methods without ground-truth pose annotations and demonstrates substantial generalisation improvements towards unseen domains. Our method significantly improves even difficult objects with little variance in appearance and geometry through self-supervision. Impressively, Our approach demonstrates a robust self-improving ability for the employed pose estimators even when initialised with much inferior pose estimates than stronger baselines [46]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 532, + 223, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 532, + 223, + 543 + ], + "spans": [ + { + "bbox": [ + 50, + 532, + 223, + 543 + ], + "type": "text", + "content": "To summarise, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 553, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 49, + 553, + 279, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 553, + 279, + 589 + ], + "spans": [ + { + "bbox": [ + 49, + 553, + 279, + 589 + ], + "type": "text", + "content": "- We formulate a new learning scheme, TexPose, that decomposes self-supervision for 6D object pose into Texture learning and Pose learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 605, + 279, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 605, + 279, + 665 + ], + "spans": [ + { + "bbox": [ + 49, + 605, + 279, + 665 + ], + "type": "text", + "content": "- We propose a surfel-conditioned adversarial training loss and a synthetic texture regularisation term to handle pose errors and segmentation imperfection during texture learning, further delivering self-improving ability to the pose estimators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 682, + 278, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 682, + 278, + 730 + ], + "spans": [ + { + "bbox": [ + 49, + 682, + 278, + 730 + ], + "type": "text", + "content": "- We show significant improvements over recent strong baselines with additional supervision signals. Our pose estimators demonstrate a substantial generalisation ability even on unseen scenes." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 298, + 95, + 384, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 95, + 384, + 108 + ], + "spans": [ + { + "bbox": [ + 298, + 95, + 384, + 108 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 117, + 538, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 117, + 538, + 346 + ], + "spans": [ + { + "bbox": [ + 296, + 117, + 538, + 346 + ], + "type": "text", + "content": "Model-based 6D Pose Estimation To retrieve the 6D pose, early methods use local or global features and search for key points correspondence on CAD models [1,6,7,28]. In recent years, learning-based methods dominate the field and solve the task using convolution neural networks (CNN) under the supervision from annotated data to extract deep features. There are two major approaches for pose estimation, in particular, correspondence-based and regression/classification-based approaches. Correspondence-based methods establish 2D-3D correspondences [37, 39, 40, 42, 59], prior to leveraging a variant of the RANSAC&PnP paradigm to solve for pose. Regression-based approaches, on the other hand, directly regress or classify the pose of the detected object. Initially these methods usually have shown a lower performance due to the existence of ambiguities [29] such as pose symmetries [53]. Therefore, methods like SSD-6D [19] discretise the rotation space to circumvent this issue. Recently, with better continuous representations for rotation [60], these methods gradually demonstrate high effectiveness [10, 48]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 367, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 367, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 367, + 538, + 740 + ], + "type": "text", + "content": "Self-Supervised Pose Learning Considering tremendous effort to collect large amount of annotations for 6D object pose [17, 50], several recent works have been proposed to explore the possibility of self-supervised pose learning using labeled synthetic data together with unlabelled real sensor data. [8] designed a novel labelling pipeline using a manipulator to generate reliable pose annotations for supervision. Self6D [47] proposed a self-supervision workflow by first pretraining a pose estimator with synthetic data, which was then adapted to the real world through a render-and-compare strategy by imposing consistencies between the rendered depth and sensed depth under the current pose estimate. CPS++ [32] used similar approach for categorical-level object pose estimation, and the shape is jointly deformed during optimisation. AAE [45] parameterise " + }, + { + "bbox": [ + 296, + 367, + 538, + 740 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 296, + 367, + 538, + 740 + ], + "type": "text", + "content": " space using a latent embedding obtained from synthetic images, which is later employed for orientation retrieval. Both works demonstrate the strong requirement for depth data in order to allow for a reasonable performance either in training or testing. However, due to the uninformative appearance and the sim-to-real domain gap, render-and-compare strategies easily diverge when there is no depth data available as shown in [47]. Hence, consecutive works introduce other supervisory signals to prune the need for depth. DSC-PoseNet [54] employs a key point consistency regularisation for dual-scale images with labelled 2D bounding box. Socket al. [44] use photometric consistency from multiple views to refine the raw estimate, while ground-truth masks are required for supervision. Recently, Self6D++ [46] proposed to initialise the render-and-compare process with a powerful deep pose refiner [26] to guide the learning process of the pose esti" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4842" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 97, + 279, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 97, + 279, + 145 + ], + "spans": [ + { + "bbox": [ + 38, + 97, + 279, + 145 + ], + "type": "text", + "content": "mator, which currently yields state-of-the-art performance among all methods without manual annotations. Chen et al. [4] leverages a tailored heuristics for pseudo labelling under student-teacher learning scheme." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 159, + 281, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 159, + 281, + 530 + ], + "spans": [ + { + "bbox": [ + 38, + 159, + 281, + 530 + ], + "type": "text", + "content": "Novel View Synthesis With a few images given, novel view synthesis aims to render scene content from unseen viewpoints, which can be regarded as a process to acquire \"textures\" of the scene. NOL [38] is proposed to address the difficulty of CAD model texturing for 6D pose estimation. Though requiring pose annotations, it demonstrates high potential to gain improvement for pose estimators with more synthesised real data in a semi-supervised fashion. Outside the scope of 6D pose, neural scene representation as implicit scene models recently get significant attention due to their ability for photorealistic novel view synthesis. NeRF [36] represent scenes using simple multi-layer perceptrons to encode the density and radiance for every 3D point. Assuming precise poses to transform views into a common reference frame, NeRF is able to capture volumes using an image reconstruction loss. NeRF-w [34] further extends NeRF to handle changing lighting and transient effects for image collections with latent embeddings. To release the rigid requirement of accurate poses, NeRF-- [52] jointly optimises camera parameters and radiance fields. iNeRF [56] performs a render-and-compare strategy using a trained NeRF for pose refinement. In contrast, BARF [27] demonstrates promising results when encountering imperfect pose initialisation by bundle-adjusting NeRF and camera poses with coarse-to-fine positional encoding scheme. GNeRF [35] formulates the learning of radiance fields and coarse pose estimates as a generative modelling process and then jointly refines both using photometric information. Recently, NeRF-Supervision [55] and NeRF-Pose [25] are proposed to leverage NeRF to learn dense descriptors or 3D representations of a given object under pose annotations when CAD model is inaccessible." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 39, + 538, + 95, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 538, + 95, + 551 + ], + "spans": [ + { + "bbox": [ + 39, + 538, + 95, + 551 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 559, + 279, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 559, + 279, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 559, + 279, + 740 + ], + "type": "text", + "content": "Our objective is to conduct per-instance 6D pose learning from monocular colour images with no manual labelling effort, namely self-supervised 6D pose estimation originally defined in [47]. Since the CAD model is known as a prior, we follow the two-stage approach of previous works [44, 46, 47, 54, 56], which first pretrain a pose estimator with labelled synthetic data, and consequently conduct self-training with unannotated real images captured from a specific real scene demanding a more precise deployment (e.g. robotic bin picking). We mainly introduce the self-training process in this section. Notably, in this tuning process, pose estimator tends to suffer from forgetting problem due to little variance in illumination and backgrounds of the fed training samples. In contrast, We will show the strength of our scheme to alleviate this issue in Section 4. After pretraining we acquire" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "spans": [ + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": "a pose estimator " + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": " able to initialise a coarse pose estimate " + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": " and a segmentation mask " + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": " for each real image " + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": " is usually used as a pseudo silhouette cue to refine the pose " + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 296, + 97, + 538, + 145 + ], + "type": "text", + "content": " with render-and-compare strategy [5,22,47,56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 145, + 538, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 145, + 538, + 193 + ], + "spans": [ + { + "bbox": [ + 296, + 145, + 538, + 193 + ], + "type": "text", + "content": "In this section, we first revisit the commonly used strategy of previous methods in this task, then introduce our new formulation requiring no additional supervisory sources clearly advantageous over previous attempts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 199, + 529, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 199, + 529, + 213 + ], + "spans": [ + { + "bbox": [ + 296, + 199, + 529, + 213 + ], + "type": "text", + "content": "3.1. Revisiting Self-Supervised 6D Pose Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "spans": [ + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": "Render-and-compare is a common approach for 6D pose self-supervision [44, 46, 47, 54, 56]. As shown in Eqn. 1, it utilises a differentiable renderer " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": " fed with an estimated pose to output a rendered mask, image, " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "[M_r, I_r] = \\mathcal{R}(T)" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "E_{rend}" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": " enforces consistency between the rendered outputs " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "[M_r, I_r]" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": " and the observations " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "[M, \\hat{I}]" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": ". It aims to refine the pseudo pose labels for each training sample, and further supervise pose estimator " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "E_{pose}" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": ". Previous attempts are majorly focused on improving " + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "inline_equation", + "content": "E_{rend}" + }, + { + "bbox": [ + 296, + 218, + 538, + 338 + ], + "type": "text", + "content": " to acquire more accurate pseudo pose annotations for better supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 332, + 345, + 537, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 345, + 537, + 395 + ], + "spans": [ + { + "bbox": [ + 332, + 345, + 537, + 395 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta , \\{T _ {i}, \\dots , T _ {N} \\}} \\sum_ {i} E _ {r e n d} (\\mathcal {R} (T _ {i}), [ M _ {i}, \\hat {I} _ {i} ]) \\\\ + \\sum_ {i} ^ {\\iota} E _ {p o s e} \\left(f _ {\\theta} \\left(\\hat {I} _ {i}\\right), T _ {i}\\right). \\tag {1} \\\\ \\end{array}", + "image_path": "1ac7296c6d77f0026b61a3fa09f751600a2ea9da65de79d53cdd83b1932b73b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "spans": [ + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "text", + "content": "When relying on 2D information " + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "inline_equation", + "content": "[M_r, I_r]" + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "text", + "content": ", minimising " + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "inline_equation", + "content": "E_{rend}" + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "text", + "content": " to improve " + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "inline_equation", + "content": "\\{T_1, \\ldots, T_N\\}" + }, + { + "bbox": [ + 296, + 403, + 538, + 559 + ], + "type": "text", + "content": " can be extremely difficult due to uninformative shape and appearance and sim-to-real visual discrepancy (e.g., Duck in Fig. 1). To compensate for the lack of information, related methods introduce additional supervision with varying difficulty levels for the acquisition (e.g., depth [47], weak labels [44, 54], deep pose refiner [46], etc.). Deep pose refiner could be acquired with the least effort as it only requires synthetic pretraining as well. However, it can introduce a strong upper bound as illustrated in [46]. The pose refiner can even lead to divergence of the pose estimator and hence lacks robustness for geometrically and photometrically challenging objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 565, + 497, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 565, + 497, + 578 + ], + "spans": [ + { + "bbox": [ + 297, + 565, + 497, + 578 + ], + "type": "text", + "content": "3.2. TexPose: Learning Pose from Texture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "spans": [ + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "type": "text", + "content": "To address the aforementioned limitations, we reformulate self-supervised 6D pose learning as two alternating optimisation problems for a texture learner " + }, + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "type": "inline_equation", + "content": "h_{\\xi}" + }, + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "type": "text", + "content": " and a pose estimator " + }, + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 296, + 584, + 538, + 633 + ], + "type": "text", + "content": " as defined in Eqn. 21 and depicted in Fig. 1 (Left)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 640, + 536, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 640, + 536, + 694 + ], + "spans": [ + { + "bbox": [ + 345, + 640, + 536, + 694 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta , \\xi} \\sum_ {i} E _ {t e x} \\left(h _ {\\xi} \\circ f _ {\\theta} (\\hat {I} _ {i}), \\hat {I} _ {i}\\right) \\\\ + \\sum_ {j} ^ {i} E _ {\\text {p o s e}} \\left(f _ {\\theta} \\circ h _ {\\xi} \\left(\\hat {T} _ {j}\\right), \\hat {T} _ {j}\\right). \\tag {2} \\\\ \\end{array}", + "image_path": "bcaf6704aa0162181dc3de865ac429ea089b554781a221b433d4d8d55fb6ce63.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 698, + 538, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 698, + 538, + 723 + ], + "spans": [ + { + "bbox": [ + 297, + 698, + 538, + 723 + ], + "type": "text", + "content": "Instead of leveraging raw images as regression targets for poses, we insert object textures as an intermediate rep" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 727, + 532, + 739 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 727, + 532, + 739 + ], + "spans": [ + { + "bbox": [ + 308, + 727, + 532, + 739 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 308, + 727, + 532, + 739 + ], + "type": "text", + "content": " Green indicates weight update and orange indicates weight freezing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4843" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 101, + 262, + 257 + ], + "blocks": [ + { + "bbox": [ + 66, + 101, + 262, + 257 + ], + "lines": [ + { + "bbox": [ + 66, + 101, + 262, + 257 + ], + "spans": [ + { + "bbox": [ + 66, + 101, + 262, + 257 + ], + "type": "image", + "image_path": "46b6b04ace8a19027146050f0825ef9a44b5b78d1e9e11de5b072c158d41798a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "lines": [ + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "spans": [ + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": "Figure 1. Left: Schematic overview of TexPose. We fully explore the benefit of both synthetic and real dataset with perfect geometric labels and realistic appearance to synthesise informative train dataset for pose learning. Right: Details of the texture learner used in our scheme. We first conduct geometric pretraining to acquire geometry branch " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "h_{\\sigma}" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": " to ensure consistent coordinates system and geometry aligned with the CAD model. Then, we transfer and freeze " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "h_{\\sigma}" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": " to the real together with newly established radiance branch " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "h_c^s" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "h_c^t" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": " as our texture learner. It is supervised by " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "E_{rec}" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "E_{reg}" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "inline_equation", + "content": "E_{adv}" + }, + { + "bbox": [ + 38, + 272, + 539, + 329 + ], + "type": "text", + "content": " capable of improving texture quality so as to synthesise more informative dataset for supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 265, + 101, + 509, + 264 + ], + "blocks": [ + { + "bbox": [ + 265, + 101, + 509, + 264 + ], + "lines": [ + { + "bbox": [ + 265, + 101, + 509, + 264 + ], + "spans": [ + { + "bbox": [ + 265, + 101, + 509, + 264 + ], + "type": "image", + "image_path": "481aa02d56454925e4a47a8c7d1ef8cd8a90ed729954fe0ef9032816adc10126.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "spans": [ + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": "resentation before pose learning. Specifically, we start with optimising " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "E_{tex}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " by fixing the parameters of pose estimator " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " used to initialise poses " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "\\{T_1,\\dots ,T_N\\}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " for each raw image " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_1,\\ldots ,\\hat{I}_N\\}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": ", aiming to provide supervision for the realistic textures embedded by " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "h_\\xi" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": ". Afterwards, we freeze " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "h_\\xi" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " to synthesise training samples " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "\\{I_1,\\dots ,I_{N'}\\}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " paired with perfect labels " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "\\{\\hat{T}_1,\\dots ,\\hat{T}_{N'}\\}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": ". Gradients will thus be propagated through " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " to update its parameters during the minimisation for " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "E_{pose}" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": " over these train data. We typically chose " + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "inline_equation", + "content": "N' > N" + }, + { + "bbox": [ + 38, + 349, + 279, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 457, + 280, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 457, + 280, + 636 + ], + "spans": [ + { + "bbox": [ + 38, + 457, + 280, + 636 + ], + "type": "text", + "content": "With this reformulation, we deliberately omit pose optimisation from 2D information usually sensitive to initialisation and prone to divergence, and bridge the domain gap directly through photorealistic synthesis. Our goal is then shifted to limit texture mapping error under the presence of pose noise so that the synthesised data can be highly informative with perfect geometric labels and as-accurate-as-possible realistic appearance. Intuitively, through multi-view supervision, it's possible to capture more accurate appearance with compensation for per-view pose noise. Our proposed strategy to further increase reliability of the synthesised data will be discussed later. Notably, different from previous attempts, no additional supervision is required in our design. The pose estimator is equipped with a self-improving ability through intermediate phase for texture learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 650, + 278, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 650, + 278, + 687 + ], + "spans": [ + { + "bbox": [ + 38, + 650, + 278, + 687 + ], + "type": "text", + "content": "NeRF preliminaries. We leverage neural radiance fields (NeRF) [36] to embed the texture information due to its simplicity and capability of photorealistic view synthesis2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 687, + 278, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 687, + 278, + 711 + ], + "spans": [ + { + "bbox": [ + 38, + 687, + 278, + 711 + ], + "type": "text", + "content": "NeRF represents the object model with two MLPs to encode information for every queried point " + }, + { + "bbox": [ + 38, + 687, + 278, + 711 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 38, + 687, + 278, + 711 + ], + "type": "text", + "content": ", namely the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "spans": [ + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": "geometry branch to predict volume density " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " and geometric feature " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "[\\sigma, \\gamma] = h_{\\sigma}(\\mathbf{x})" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " and radiance branch for colour " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " prediction with " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = h_c(\\mathbf{x}, \\gamma)" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " (weight " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " is omitted). When rendering from a given viewpoint, " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\{x_k\\}_1^K" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " are gathered along the ray " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{o}, \\mathbf{d} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " represent the camera centre and the view direction respectively and " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " is the distance. The colour " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(\\mathbf{r})" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " will be estimated through numerical quadrature approximation with " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "w_k = \\exp(-\\delta_k \\sigma_k)" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\alpha_k = \\prod_{j=1}^{k-1} w_j" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\delta_k" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " being the step size between adjacent points, and " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{C}(\\mathbf{r}) = \\sum_{k=1}^{K} \\alpha_k (1 - w_k) \\mathbf{c_k}" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": ". Rendering for depth " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "D(\\mathbf{r})" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "inline_equation", + "content": "M(\\mathbf{r})" + }, + { + "bbox": [ + 297, + 349, + 538, + 497 + ], + "type": "text", + "content": " is achieved by modifying integration element (see supp. mat. for details)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "spans": [ + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": "Geometric pretraining. Though geometry and radiance branch can be trained jointly as shown in NeRF and its variants [21,34,36], the pose for supervision in our scenario is highly noisy, and the coordinate system of NeRF can end up as an arbitrary one and further cause erroneous labels for synthesised data. We hence opt to pretrain geometry branch " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "h_{\\sigma}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": " using pixel-perfect data rendered by the CAD model with accurate pose " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": ", depth " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": " to supervise NeRF following Eqn. 4, where " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "E_{d}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": " is scale-invariant loss from [20], " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "inline_equation", + "content": "\\lambda_{d}" + }, + { + "bbox": [ + 297, + 511, + 538, + 631 + ], + "type": "text", + "content": " are weighting factors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 647, + 537, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 647, + 537, + 671 + ], + "spans": [ + { + "bbox": [ + 303, + 647, + 537, + 671 + ], + "type": "interline_equation", + "content": "E _ {2 d} (\\mathbf {r}) = \\left\\| \\mathbf {C} (\\mathbf {r}) - \\hat {\\mathbf {C}} (\\mathbf {r}) \\right\\| _ {2} ^ {2} + \\lambda_ {m} \\left\\| M (\\mathbf {r}) - \\hat {M} (\\mathbf {r}) \\right\\| _ {2} ^ {2} \\tag {3}", + "image_path": "2e42f80a1f85d0445125d62d35de99d360907a85141a98048a7ec514bed60288.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 333, + 684, + 537, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 684, + 537, + 699 + ], + "spans": [ + { + "bbox": [ + 333, + 684, + 537, + 699 + ], + "type": "interline_equation", + "content": "E _ {p r e} (\\mathbf {r}) = E _ {2 d} (\\mathbf {r}) + \\lambda_ {d} E _ {d} (D (\\mathbf {r}), \\hat {D} (\\mathbf {r})). \\tag {4}", + "image_path": "e08220fb12fd1860a63a248d275fccd82b308a294691f3adcea1350768f6564f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 703, + 538, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 703, + 538, + 739 + ], + "spans": [ + { + "bbox": [ + 297, + 703, + 538, + 739 + ], + "type": "text", + "content": "Notably, though our current goal is to establish a geometry branch with precise density output \"aligned\" with the CAD model, training for radiance branch is also included, though" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 38, + 717, + 278, + 739 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 717, + 278, + 739 + ], + "spans": [ + { + "bbox": [ + 38, + 717, + 278, + 739 + ], + "type": "text", + "content": "2Radiance, texture and appearance are discussed as interchangeable concepts in our paper." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 761 + ], + "type": "text", + "content": "4844" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 96, + 279, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 96, + 279, + 158 + ], + "spans": [ + { + "bbox": [ + 38, + 96, + 279, + 158 + ], + "type": "text", + "content": "it will be discarded afterwards. We find this greatly helps to extract meaningful geometric features fed into the radiance branch, and further eases the effort for realistic appearance transfer in the later texture learning stage. We quantitatively detail the necessity of this stage in supp. mat." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "spans": [ + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "text", + "content": "Texture learning from noisy poses. Our current goal is to learn texture encoded by the radiance branch " + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "text", + "content": " using raw images paired with poses initialised by the pose estimator " + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "text", + "content": ". Note in this stage, we only re-optimise parameters from " + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "inline_equation", + "content": "h_{\\sigma}" + }, + { + "bbox": [ + 38, + 173, + 279, + 316 + ], + "type": "text", + "content": " being fixed to capture accurate texture and meanwhile ensure no shift of coordinate system and geometry. Though jointly optimising textures and poses through photometric consistency has a potential to improve both aspects as shown in [21, 27, 56], we find it can easily fail when dealing with the textureless objects. Alternatively, we seek to supervise the texture learner with higher pose error tolerance leveraging stronger radiance branch and robust loss." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "spans": [ + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": "Different from the vanilla NeRF used in geometric pretraining, we now employ two radiance branches, " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "h_c^s" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "h_c^t" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": ", to translate the extracted geometric features into static radiance (accurate underlying textures) and transient radiance (texture outliers caused by pose errors) respectively. Apart from geometric features " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": ", we further condition it on view directions " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": ", lighting embedding " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "\\tau^a" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": " and transient embedding " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "\\tau^t" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": " [34] (see Fig. 1 Right). For each view, we use " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "\\tau^a" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": " to capture illumination variations and " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "\\tau^t" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": " to produce transient effect caused by pose noise. Patch-based training strategy [41] is used to preserve the spatial structures for robust loss supervision. An image reconstruction loss " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "E_{rec}" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": " is incorporated with uncertainty awareness to learn textures robustly. We refer to the supp. mat. for the details of the radiance branches and uncertainty-aware " + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "inline_equation", + "content": "E_{rec}" + }, + { + "bbox": [ + 38, + 317, + 279, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "spans": [ + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": "We noticed that supervision from " + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "inline_equation", + "content": "E_{rec}" + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": " is still not adequate to capture accurate texture under pose noise. We further design a surfel-conditioned adversarial loss for sampled patches to tolerate texture shifts caused by pose errors. For each location on the object surface, we compute normalised object coordinate " + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": " [49] and normal " + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": " with current pose estimate " + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": " to represent its local information. By conditioning the predicted patch colour " + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": " on its local surface information " + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "inline_equation", + "content": "[S_{x}, S_{n}]" + }, + { + "bbox": [ + 38, + 496, + 279, + 604 + ], + "type": "text", + "content": ", we employ an adversarial scheme:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 615, + 279, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 615, + 279, + 646 + ], + "spans": [ + { + "bbox": [ + 64, + 615, + 279, + 646 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E _ {a d v} (P) = \\mathbb {E} _ {[ S _ {x}, S _ {n} ], P} (1 - \\log D ([ S _ {x}, S _ {n} ], P)) \\tag {5} \\\\ + \\mathbb {E} _ {[ S _ {x}, S _ {n} ], \\hat {P}} (\\log D ([ S _ {x}, S _ {n} ], M \\hat {P})). \\\\ \\end{array}", + "image_path": "29f40a389e19661524cb1bd7bde6d24e2c5bb6a5c5fc254b65af231885a081f8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "spans": [ + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": "We also observed the imperfect predicted mask " + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": " can cause background colour artefacts within the object. We therefore establish a regularisation from synthetic cues to compensate the boundary imperfections by generating synthetic patch " + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "inline_equation", + "content": "P_{syn}" + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "inline_equation", + "content": "M_{syn}" + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": " with current pose estimate " + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": ", then applying erosion on the predicted mask " + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": " to compute a padded mask with " + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "inline_equation", + "content": "M_{pad} = M_{syn}(1 - M)" + }, + { + "bbox": [ + 38, + 654, + 279, + 739 + ], + "type": "text", + "content": ". The foreground" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "spans": [ + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "text", + "content": "boundaries of the real images for supervision are therefore padded with the synthetic colours. This regularisation " + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "inline_equation", + "content": "E_{feat}" + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "text", + "content": " is realised by minimising the MSE error of deep features extracted by a pretrained VGG19 network [43] to align the boundary features of the predicted patch " + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "text", + "content": " with the synthetic patch " + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "inline_equation", + "content": "P_{syn}" + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "text", + "content": " in a more abstract level. The second part of Eqn. 6 ensures that rendered patch region within the mask " + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 296, + 96, + 538, + 193 + ], + "type": "text", + "content": " are not affected by the synthetic information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 202, + 536, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 202, + 536, + 233 + ], + "spans": [ + { + "bbox": [ + 326, + 202, + 536, + 233 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E _ {r e g} (P) = E _ {f e a t} \\left(P, M \\hat {P} + M _ {p a d} P _ {s y n}\\right) \\tag {6} \\\\ + \\lambda_ {f g} E _ {\\text {f e a t}} (M P + (1 - M) \\hat {P}, \\hat {P}). \\\\ \\end{array}", + "image_path": "42fd1f4b525cb24b67f4611a771f1ee5b9491ea24d9815c073585c86929666b3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 240, + 538, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 240, + 538, + 265 + ], + "spans": [ + { + "bbox": [ + 296, + 240, + 538, + 265 + ], + "type": "text", + "content": "Formally, the loss function " + }, + { + "bbox": [ + 296, + 240, + 538, + 265 + ], + "type": "inline_equation", + "content": "E_{tex}" + }, + { + "bbox": [ + 296, + 240, + 538, + 265 + ], + "type": "text", + "content": " for the texture learner supervision is summarised as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 275, + 536, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 275, + 536, + 287 + ], + "spans": [ + { + "bbox": [ + 339, + 275, + 536, + 287 + ], + "type": "interline_equation", + "content": "E _ {t e x} = E _ {r e c} + \\lambda_ {a d v} E _ {a d v} + \\lambda_ {r e g} E _ {r e g}. \\tag {7}", + "image_path": "de3b4376a3278f13efc4faa7093bde1621784be604038260a085885fb0fc5e3f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "spans": [ + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "type": "inline_equation", + "content": "\\lambda_{adv}" + }, + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 297, + 295, + 472, + 308 + ], + "type": "text", + "content": " are weighting factors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "spans": [ + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": "Pose learning. The entire pipeline runs by first optimising " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{tex}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": ", and subsequent minimisation of " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{pose}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": " with the synthesised dataset as depicted in Fig. 1 (Left). As our scheme operates at data level and thus method-agnostic, " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{pose}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": " can be generalised across methods with different loss designs, e.g., direct pose or correspondence regression. Interestingly, we observed that the pose estimator can converge instantly with only one optimisation loop on " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{tex}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{pose}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": ", while adding more loops, i.e., continue to optimise " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{tex}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": " with improved poses after " + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "inline_equation", + "content": "E_{pose}" + }, + { + "bbox": [ + 296, + 321, + 538, + 465 + ], + "type": "text", + "content": ", only brings insignificant improvements. We attribute this to the pixel-perfect supervision from texture learner and its strong ability to mitigate pose errors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 475, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 475, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 297, + 475, + 378, + 489 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 495, + 537, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 495, + 537, + 543 + ], + "spans": [ + { + "bbox": [ + 296, + 495, + 537, + 543 + ], + "type": "text", + "content": "In this section we first introduce the employed metrics and dataset before presenting our results for the task of 6D pose estimation, then provide detailed analysis of our proposed scheme in ablation study." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 550, + 410, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 550, + 410, + 563 + ], + "spans": [ + { + "bbox": [ + 297, + 550, + 410, + 563 + ], + "type": "text", + "content": "4.1. Evaluation Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 569, + 538, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 569, + 538, + 676 + ], + "spans": [ + { + "bbox": [ + 296, + 569, + 538, + 676 + ], + "type": "text", + "content": "In line with previous work, we present our results using the ADD(-S) metric, as proposed in [14]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. If the average distance is below " + }, + { + "bbox": [ + 296, + 569, + 538, + 676 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 296, + 569, + 538, + 676 + ], + "type": "text", + "content": " of the object diameter, the estimated pose will be considered correct. For symmetric objects, ADD score is modified to measure the distance to the nearest vertices, here referred to as ADD-S." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 684, + 355, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 684, + 355, + 696 + ], + "spans": [ + { + "bbox": [ + 297, + 684, + 355, + 696 + ], + "type": "text", + "content": "4.2. Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 702, + 538, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 702, + 538, + 739 + ], + "spans": [ + { + "bbox": [ + 296, + 702, + 538, + 739 + ], + "type": "text", + "content": "Synthetic BOP PBR. We pretrain our models using synthetic data from [9, 16], employing physically-based rendering (PBR) techniques. Although PBR can capture more" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4845" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 95, + 169, + 170 + ], + "blocks": [ + { + "bbox": [ + 92, + 95, + 169, + 170 + ], + "lines": [ + { + "bbox": [ + 92, + 95, + 169, + 170 + ], + "spans": [ + { + "bbox": [ + 92, + 95, + 169, + 170 + ], + "type": "image", + "image_path": "177bbbc8c78199e5df482a9ac1b0c2e08b0019eed85e60a21923fcf01dbb737e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 95, + 248, + 170 + ], + "blocks": [ + { + "bbox": [ + 171, + 95, + 248, + 170 + ], + "lines": [ + { + "bbox": [ + 171, + 95, + 248, + 170 + ], + "spans": [ + { + "bbox": [ + 171, + 95, + 248, + 170 + ], + "type": "image", + "image_path": "9fe9c29f3748510ec1d002e5500d9f1f6e988638dfd82c0f05f7755fc702f3c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 251, + 95, + 328, + 170 + ], + "blocks": [ + { + "bbox": [ + 251, + 95, + 328, + 170 + ], + "lines": [ + { + "bbox": [ + 251, + 95, + 328, + 170 + ], + "spans": [ + { + "bbox": [ + 251, + 95, + 328, + 170 + ], + "type": "image", + "image_path": "7651a735986fa4e359e8a7e88ee5b665da02f1747dcbc86059d31231992a4451.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 330, + 95, + 408, + 170 + ], + "blocks": [ + { + "bbox": [ + 330, + 95, + 408, + 170 + ], + "lines": [ + { + "bbox": [ + 330, + 95, + 408, + 170 + ], + "spans": [ + { + "bbox": [ + 330, + 95, + 408, + 170 + ], + "type": "image", + "image_path": "0b8fa52066ebefc90d28ec1532337a763a35649d36376512213c664ac881c2d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 411, + 95, + 487, + 170 + ], + "blocks": [ + { + "bbox": [ + 411, + 95, + 487, + 170 + ], + "lines": [ + { + "bbox": [ + 411, + 95, + 487, + 170 + ], + "spans": [ + { + "bbox": [ + 411, + 95, + 487, + 170 + ], + "type": "image", + "image_path": "df10ff94f8b2898c3b1c8438d586ef901187e9786d3a669d1614f4af5501b7cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 92, + 172, + 169, + 249 + ], + "blocks": [ + { + "bbox": [ + 92, + 172, + 169, + 249 + ], + "lines": [ + { + "bbox": [ + 92, + 172, + 169, + 249 + ], + "spans": [ + { + "bbox": [ + 92, + 172, + 169, + 249 + ], + "type": "image", + "image_path": "799cfbe78b9509ebbc881fda5832ef8212c7933cca2caaf17238c47e580bb3dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 172, + 248, + 250 + ], + "blocks": [ + { + "bbox": [ + 171, + 172, + 248, + 250 + ], + "lines": [ + { + "bbox": [ + 171, + 172, + 248, + 250 + ], + "spans": [ + { + "bbox": [ + 171, + 172, + 248, + 250 + ], + "type": "image", + "image_path": "013f191936a71f8586a7a541ca96034d2103e224b9063b692b272005ca7f10a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 172, + 328, + 250 + ], + "blocks": [ + { + "bbox": [ + 251, + 172, + 328, + 250 + ], + "lines": [ + { + "bbox": [ + 251, + 172, + 328, + 250 + ], + "spans": [ + { + "bbox": [ + 251, + 172, + 328, + 250 + ], + "type": "image", + "image_path": "507d4882d86a864f6af4a6ab2563c14743dd8941fe913d3f1485a2cc97d0b0c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 330, + 172, + 408, + 250 + ], + "blocks": [ + { + "bbox": [ + 330, + 172, + 408, + 250 + ], + "lines": [ + { + "bbox": [ + 330, + 172, + 408, + 250 + ], + "spans": [ + { + "bbox": [ + 330, + 172, + 408, + 250 + ], + "type": "image", + "image_path": "9cff58455bebe7043dfd341ab15c378f9300358d933553041d035ec68dc20c3f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 411, + 172, + 487, + 250 + ], + "blocks": [ + { + "bbox": [ + 411, + 172, + 487, + 250 + ], + "lines": [ + { + "bbox": [ + 411, + 172, + 487, + 250 + ], + "spans": [ + { + "bbox": [ + 411, + 172, + 487, + 250 + ], + "type": "image", + "image_path": "2d50e06d3f4e19b83656249dfea0707f890f99414ef688097b7f55518a19d0ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 92, + 252, + 169, + 326 + ], + "blocks": [ + { + "bbox": [ + 92, + 252, + 169, + 326 + ], + "lines": [ + { + "bbox": [ + 92, + 252, + 169, + 326 + ], + "spans": [ + { + "bbox": [ + 92, + 252, + 169, + 326 + ], + "type": "image", + "image_path": "b5f7d70aab0d442004ec9992473263371d6dfff30755ae082ad4b672cb6ccfd9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 38, + 335, + 538, + 368 + ], + "lines": [ + { + "bbox": [ + 38, + 335, + 538, + 368 + ], + "spans": [ + { + "bbox": [ + 38, + 335, + 538, + 368 + ], + "type": "text", + "content": "Figure 2. Qualitative results on 5 objects from LineMOD. Top: Renderings of the reconstructed CAD models which are part of LineMOD. Middle: Same objects rendered with our learnt textures. Bottom: Demonstration of our pose estimation quality. Thereby, the green, red, and blue boxes depict ground-truth pose as well as estimated pose before and after our self-supervision, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 171, + 252, + 249, + 326 + ], + "blocks": [ + { + "bbox": [ + 171, + 252, + 249, + 326 + ], + "lines": [ + { + "bbox": [ + 171, + 252, + 249, + 326 + ], + "spans": [ + { + "bbox": [ + 171, + 252, + 249, + 326 + ], + "type": "image", + "image_path": "5587bfdd5289abb4a6481f785d6863ecdf779df6898e1c51a3ab75641d7c430b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 251, + 252, + 328, + 326 + ], + "blocks": [ + { + "bbox": [ + 251, + 252, + 328, + 326 + ], + "lines": [ + { + "bbox": [ + 251, + 252, + 328, + 326 + ], + "spans": [ + { + "bbox": [ + 251, + 252, + 328, + 326 + ], + "type": "image", + "image_path": "329d0a5f5526d413c97073e5bcd88bd2c77ac2421491d16fcbf49d0dceefa47c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 330, + 252, + 408, + 326 + ], + "blocks": [ + { + "bbox": [ + 330, + 252, + 408, + 326 + ], + "lines": [ + { + "bbox": [ + 330, + 252, + 408, + 326 + ], + "spans": [ + { + "bbox": [ + 330, + 252, + 408, + 326 + ], + "type": "image", + "image_path": "8d30778e7ea971753d7effb6abc9e27da7a6b14e8f8c4ba5a57206b4899c92d1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 411, + 252, + 487, + 326 + ], + "blocks": [ + { + "bbox": [ + 411, + 252, + 487, + 326 + ], + "lines": [ + { + "bbox": [ + 411, + 252, + 487, + 326 + ], + "spans": [ + { + "bbox": [ + 411, + 252, + 487, + 326 + ], + "type": "image", + "image_path": "a136a2b687596b3587a7739b0a52cb2a913997ba4b3cceb0f2c4760b40565d33.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 38, + 375, + 280, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 375, + 280, + 423 + ], + "spans": [ + { + "bbox": [ + 38, + 375, + 280, + 423 + ], + "type": "text", + "content": "realistic lighting conditions, there still exhibits a large sim-to-real domain gap. Such discrepancies can strongly degrade the pose estimation performance when being deployed in the real world." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 38, + 439, + 279, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 439, + 279, + 523 + ], + "spans": [ + { + "bbox": [ + 38, + 439, + 279, + 523 + ], + "type": "text", + "content": "LineMOD. [14] consists of 13 objects with watertight CAD models. Thereby, each object is accompanied by a small image sequence of around 1k frames, captured in cluttered environment with challenging lighting. Following previous works [40, 47], we employ " + }, + { + "bbox": [ + 38, + 439, + 279, + 523 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 38, + 439, + 279, + 523 + ], + "type": "text", + "content": " of the captured RGB images for self-supervision, discarding any manual annotations, and test our model on the remaining samples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 38, + 539, + 279, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 539, + 279, + 587 + ], + "spans": [ + { + "bbox": [ + 38, + 539, + 279, + 587 + ], + "type": "text", + "content": "Occluded LineMOD. [3] is an extension of one LineMOD sequence showing eight different objects from LineMOD undergoing severe occlusion. For fairness we adopt the same train/test split for self-supervision as [46]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 38, + 603, + 279, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 603, + 279, + 650 + ], + "spans": [ + { + "bbox": [ + 38, + 603, + 279, + 650 + ], + "type": "text", + "content": "HomebrewedDB. [18] is a recently established dataset with various kinds of objects and poses. To test generalisability we test on the three objects that it shares with LineMOD [18]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 38, + 660, + 164, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 660, + 164, + 672 + ], + "spans": [ + { + "bbox": [ + 38, + 660, + 164, + 672 + ], + "type": "text", + "content": "4.3. Performance Analysis" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 38, + 678, + 279, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 678, + 279, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 678, + 279, + 738 + ], + "type": "text", + "content": "Performance on LineMOD and Occluded LineMOD In Table 1, we demonstrate our results for LineMOD with respect to the ADD(-S) metric. Note that GDR-LB and GDR-UB correspond to the performance of our base pose estimators, respectively supervised with synthetic data and real" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "spans": [ + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "text", + "content": "data, reflecting the lower and upper performance bound. As one can easily see there is a significant performance gap between fully-supervised methods and approaches trained with synthetic data instead (c.f. DPODv2 [42], GDR [48]). Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities to achieve comparable results, as for example Self6D++ [46] with depth supervision. In contrast, our method is able to further close the domain gap whilst relying solely on RGB images, outperforming all other self-supervised approaches by at least " + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "text", + "content": " for RGB-D approaches and " + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "text", + "content": " for RGB-only methods. Noticeably, we are able to improve our lower bound performance from " + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "inline_equation", + "content": "77.4\\%" + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "inline_equation", + "content": "91.7\\%" + }, + { + "bbox": [ + 296, + 375, + 538, + 554 + ], + "type": "text", + "content": ", which is even slightly better than the performance of our upper bound model trained with full supervision." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "spans": [ + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": "We would like to emphasise that our model is capable of providing significant performance boosts on highly challenging tiny and textureless objects (e.g., Ape, Duck, and Holep.) by improving the lower bound from " + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "inline_equation", + "content": "50.9\\%" + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "inline_equation", + "content": "24.6\\%" + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "inline_equation", + "content": "32.1\\%" + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "inline_equation", + "content": "83.4\\%" + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "inline_equation", + "content": "79.3\\%" + }, + { + "bbox": [ + 296, + 559, + 539, + 738 + ], + "type": "text", + "content": ", respectively. These objects are typically hard to handle by other self-supervised methods without real supervision (c.f. results from Self6D++ [46], Sock et al. [44], etc). As the RGB variant of Self6D++ [46] leverages DeepIM to refine the output pose in order to generate pseudo pose annotations, this process imposes a strong upper bound on the pose estimator. In fact, the performance achieved by Self6D++ is clearly inferior to DeepIM for almost all objects. This problem is even more apparent when the pose refiner is worse than the pose estimator (c.f. Holep. results from GDR [48], DeepIM [26] and Self6D++ [46])." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 299, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 299, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 299, + 760 + ], + "type": "text", + "content": "4846" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 40, + 128, + 539, + 310 + ], + "blocks": [ + { + "bbox": [ + 38, + 95, + 537, + 118 + ], + "lines": [ + { + "bbox": [ + 38, + 95, + 537, + 118 + ], + "spans": [ + { + "bbox": [ + 38, + 95, + 537, + 118 + ], + "type": "text", + "content": "Table 1. Evaluation results on LineMOD dataset. \\*: objects with symmetry. \\*\\*: used as our pretrained pose estimator. " + }, + { + "bbox": [ + 38, + 95, + 537, + 118 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 38, + 95, + 537, + 118 + ], + "type": "text", + "content": " : use no PBR renderings for synthetic pretraining. †: re-implemented version from [46], supervision source for Self6D++. ‡: with depth supervision" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 40, + 128, + 539, + 310 + ], + "lines": [ + { + "bbox": [ + 40, + 128, + 539, + 310 + ], + "spans": [ + { + "bbox": [ + 40, + 128, + 539, + 310 + ], + "type": "table", + "html": "
SupervisionSynSyn + RealSyn + Self
Supervision signals required in real domain
Depth--------✓---✓-
Pose Refiner-----------✓✓-
Weak labels---------✓✓---
GT Pose labels-----✓✓✓------
MethodsAAE* [45]MHP* [29]DPODv2 [42]GDR-LB** [48]DeepIM† [26]DPODv2 [42]GDR-UB [48]SO-Pose [10]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape4.011.962.150.985.880.085.0-38.937.631.276.075.480.9
Benchwise20.966.288.399.493.199.799.8-75.278.683.091.694.999
Camera30.522.492.589.299.199.296.5-36.965.649.697.197.094.8
Can35.959.896.697.299.899.699.3-65.665.656.599.899.599.7
Cat17.926.986.179.998.795.193.0-57.952.557.985.686.692.6
Driller24.044.690.198.7100.098.9100.0-67.048.873.798.898.997.4
Duck4.98.354.824.661.979.565.3-19.635.131.356.568.383.4
Eggbox*81.055.798.681.193.599.699.9-99.089.296.091.099.094.9
Glue*45.554.695.481.293.399.898.1-94.164.563.492.296.193.4
Holep.17.615.527.041.932.172.373.4-16.241.538.835.441.979.3
Iron32.060.898.298.8100.099.486.9-77.980.961.999.599.499.8
Lamp60.5-91.098.999.196.399.6-68.270.764.797.498.998.3
Phone33.834.474.364.394.896.886.3-50.160.554.491.894.378.9
Average31.438.881.277.488.093.591.096.058.960.658.685.688.591.7
", + "image_path": "1328f0e9d7f3dc632cc0ec7719f2b7f5d618b255d206b67bb26e9ccdd3257dd5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 44, + 394, + 274, + 489 + ], + "blocks": [ + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "lines": [ + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "spans": [ + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "type": "text", + "content": "Table 2. Evaluation results on Occluded LineMOD dataset. \\*: objects with symmetry. \\*\\*: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. " + }, + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "type": "text", + "content": ":re-implemented version from [46], supervision source for Self6D++. " + }, + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 38, + 330, + 280, + 385 + ], + "type": "text", + "content": ": with depth supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 44, + 394, + 274, + 489 + ], + "lines": [ + { + "bbox": [ + 44, + 394, + 274, + 489 + ], + "spans": [ + { + "bbox": [ + 44, + 394, + 274, + 489 + ], + "type": "table", + "html": "
SupervisionSynSyn + Self
MethodsGDR** [48]Self6D [47]Sock et al. [44]DSC [54]Self6D++ [46]Self6D++‡ [46]Ours
Ape44.013.712.09.157.759.460.5
Can83.943.227.521.195.096.593.4
Cat49.118.712.026.052.660.856.1
Driller88.532.520.533.590.592.092.5
Duck15.014.423.012.226.730.655.5
Eggbox*33.957.825.139.445.051.146.0
Glue*75.054.327.037.087.188.682.8
Holep.34.022.035.020.423.538.546.5
Average52.932.122.824.859.864.766.7
", + "image_path": "d2c20daeb1c8c1947496d537190ba4ae3de63d17dfa4a556943ca42a8bc4099a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "spans": [ + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "text", + "content": "On the contrary, our method relies on texture optimisation to synthesise realistic training samples with precise pose annotations for supervision, making self-supervision more robust towards pose errors. We further validate our scheme with another pose estimator, DPODv2 [42] in supp. mat. As for our results on Occluded LineMOD in Table 2, we can make the same observations. We again report the best overall performance of all self-supervised methods, despite not relying on any elaborate design for occlusion-awareness as Self6D++ [46], with " + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "inline_equation", + "content": "64.7\\%" + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "inline_equation", + "content": "59.8\\%" + }, + { + "bbox": [ + 38, + 505, + 279, + 636 + ], + "type": "text", + "content": " for Self6D++ [46] with and without depth data, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 643, + 280, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 643, + 280, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 643, + 280, + 740 + ], + "type": "text", + "content": "Performance on HomebrewDB and generalisation In this experiment, we want to study the generalisation ability of our method towards new scenes with changing illumination. From Table 3, we observe that after self-supervision on LineMOD, Self6D++ undergoes a clear performance drop compared to their respective pretrained models when testing on HomebrewDB (86.8% vs. 80.6%). This reveals that Self6D++'s self-supervision scheme is prone to forgetting." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 300, + 394, + 535, + 446 + ], + "blocks": [ + { + "bbox": [ + 296, + 330, + 539, + 386 + ], + "lines": [ + { + "bbox": [ + 296, + 330, + 539, + 386 + ], + "spans": [ + { + "bbox": [ + 296, + 330, + 539, + 386 + ], + "type": "text", + "content": "Table 3. Evaluation results on HomebrewedDB dataset. **: used as our pretrained pose estimator before self-supervision, consistent with Self6D++. †: Pre-process testing images with the same intrinsics as LineMOD, different from [46] as they directly feed raw images without scaling. ‡: Use raw images for re-adaptation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 300, + 394, + 535, + 446 + ], + "lines": [ + { + "bbox": [ + 300, + 394, + 535, + 446 + ], + "spans": [ + { + "bbox": [ + 300, + 394, + 535, + 446 + ], + "type": "table", + "html": "
SupervisionSynSyn + Self
MethodGDR** † [48]Sock et al. [44]Self6D++ † [46]Ours †Ours ‡
Benchvise88.857.375.793.192.9
Driller92.846.689.494.894.2
Phone78.741.576.879.381.2
Average86.852.080.689.189.4
", + "image_path": "c6bc7fe8c6004e583a48cfc6c1683a0a68579d2a53376b04921b1d9dd79f8cb5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 461, + 539, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 461, + 539, + 676 + ], + "spans": [ + { + "bbox": [ + 296, + 461, + 539, + 676 + ], + "type": "text", + "content": "Our method can instead generalise to the new scene with even boosted performance (86.8% vs. 89.1%). We attribute this to the learnt realistic textures, which are used to uniformly synthesise photorealistic data that can further be augmented with random backgrounds to increase domain invariance [11,26]. Further, we investigate the adaptation ability of our method on these new scenes. To this end, we conduct self-supervision on raw images from HomebrewedDB. We first synthesise novels view with the pre-trained texture learner from LineMOD in an effort to adapt the pose estimator to the camera from HomebrewedDB. Afterwards, we re-train the texture learner with raw images from the new target scene. Our results again illustrate that our method outperforms all other baselines. Interestingly, we notice that the choice of adaptation domain (LineMOD or HomebrewedDB) has marginal effect on our final performance, suggesting that our method has potential to close the sim-to-real domain gap even trained for one scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 684, + 392, + 697 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 684, + 392, + 697 + ], + "spans": [ + { + "bbox": [ + 297, + 684, + 392, + 697 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 703, + 538, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 703, + 538, + 740 + ], + "spans": [ + { + "bbox": [ + 296, + 703, + 538, + 740 + ], + "type": "text", + "content": "Dealing with bad initialisation. We randomly sample 200 testing images from the Duck sequence of LineMOD (shown in Fig. 1). For the baseline render-and-compare" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4847" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 95, + 191, + 213 + ], + "blocks": [ + { + "bbox": [ + 57, + 95, + 191, + 213 + ], + "lines": [ + { + "bbox": [ + 57, + 95, + 191, + 213 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 191, + 213 + ], + "type": "image", + "image_path": "7c30065e74a9aeeb477ff190a737d0bfc3591f2e0e5bf5501a4bd462e1bc6663.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 38, + 223, + 538, + 255 + ], + "lines": [ + { + "bbox": [ + 38, + 223, + 538, + 255 + ], + "spans": [ + { + "bbox": [ + 38, + 223, + 538, + 255 + ], + "type": "text", + "content": "Figure 3. Ablation of the different schemes and the required amount of data for self-supervision. Left: Pose errors after every 10 iterations for the baseline scheme (pseudo pose) and our proposed pseudo-texture scheme. Right: Self-supervision with different amount of training data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 196, + 95, + 323, + 213 + ], + "blocks": [ + { + "bbox": [ + 196, + 95, + 323, + 213 + ], + "lines": [ + { + "bbox": [ + 196, + 95, + 323, + 213 + ], + "spans": [ + { + "bbox": [ + 196, + 95, + 323, + 213 + ], + "type": "image", + "image_path": "329c65a021266d7e8a6a177438467c0ea819bf03b84b7cf10854ead6586a674b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 340, + 95, + 518, + 214 + ], + "blocks": [ + { + "bbox": [ + 340, + 95, + 518, + 214 + ], + "lines": [ + { + "bbox": [ + 340, + 95, + 518, + 214 + ], + "spans": [ + { + "bbox": [ + 340, + 95, + 518, + 214 + ], + "type": "image", + "image_path": "68cb7e21ea46b56466eb03d6e9a13153e720613db22a76d8983ee68a1b1f6bec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 263, + 279, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 263, + 279, + 359 + ], + "spans": [ + { + "bbox": [ + 38, + 263, + 279, + 359 + ], + "type": "text", + "content": "scheme (pseudo pose), we minimise " + }, + { + "bbox": [ + 38, + 263, + 279, + 359 + ], + "type": "inline_equation", + "content": "E_{\\text{rend}}" + }, + { + "bbox": [ + 38, + 263, + 279, + 359 + ], + "type": "text", + "content": " in Eqn. 1 to refine the pose with the raw initialisation from the pose estimator. For our proposed scheme (pseudo texture), we repeatedly run an evaluation after every 10 weight update steps of the pose estimator. From Fig. 3 (left), we see that the baseline scheme diverges, while our proposed scheme significantly enhances the convergence properties thanks to the synthesised supervision from the texture learner." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 42, + 400, + 278, + 469 + ], + "blocks": [ + { + "bbox": [ + 38, + 369, + 279, + 392 + ], + "lines": [ + { + "bbox": [ + 38, + 369, + 279, + 392 + ], + "spans": [ + { + "bbox": [ + 38, + 369, + 279, + 392 + ], + "type": "text", + "content": "Table 4. Ablation study on partial objects from LineMOD dataset.*: objects with symmetry." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 42, + 400, + 278, + 469 + ], + "lines": [ + { + "bbox": [ + 42, + 400, + 278, + 469 + ], + "spans": [ + { + "bbox": [ + 42, + 400, + 278, + 469 + ], + "type": "table", + "html": "
EadvEregApeCatDuckGlue*Holep.PhoneAverage
w/o self-sup.50.979.924.681.241.964.357.1
--76.390.365.982.359.665.373.3
✓-80.489.476.590.976.570.780.7
-✓82.187.878.496.477.176.283.0
✓✓80.992.683.493.479.378.984.7
", + "image_path": "f1ee3061e64bf7dc6325db00adf94ba64b094e7b425f139664663daf7d1238b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "spans": [ + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "content": "Impact of " + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "inline_equation", + "content": "E_{adv}" + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "inline_equation", + "content": "E_{reg}" + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "content": ". We study the individual components of our proposed training objective to guide texture learning under pose noise using the six worst performing objects. As shown in Table 4, even when training the texture learner with noisy poses, we are able to significantly improve the lower bound for most objects. We attribute this behaviour to the strong synthesis capabilities of NeRF [36]. Notably, for objects with richer textures or articulation (Glue, Phone), the baseline model without their supervision yields nearly no improvements. However, when respectively introducing " + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "inline_equation", + "content": "E_{adv}" + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "inline_equation", + "content": "E_{reg}" + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "content": " to the baseline model, we observe ca. " + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 38, + 494, + 279, + 650 + ], + "type": "text", + "content": " absolute improvement on ADD(-S) score. Further, combining both of them yields the best performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 663, + 279, + 736 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 663, + 279, + 736 + ], + "spans": [ + { + "bbox": [ + 38, + 663, + 279, + 736 + ], + "type": "text", + "content": "Impact of training data size. As the core idea of our method is to learn the object texture from multiple views, instead of optimising per-sample pose for supervision, we believe such formulation has a potential to allow for pose learning with much less data. To this end, we conduct self-supervision with an increasing dataset size taken from the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "spans": [ + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "text", + "content": "original training split. On the right of Fig. 3, we can see that the performance keeps steadily increasing when the amount of data is enlarged. Surprisingly, we observe that already with " + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "text", + "content": " of the training split (only 9 images), our method can bring a significant absolute improvement of ca. " + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "text", + "content": ", achieving around " + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 296, + 263, + 538, + 359 + ], + "type": "text", + "content": " of the optimal performance. This further confirms our method can even serve as a strong baseline for few-shot learning under extreme data scarcity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 359, + 538, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 359, + 538, + 395 + ], + "spans": [ + { + "bbox": [ + 296, + 359, + 538, + 395 + ], + "type": "text", + "content": "We refer interested readers to supp. mat for further detailed ablation experiments, e.g., analysis on optimisation design (Eqn. 1 and Eqn. 2) in Table 7." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 405, + 370, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 405, + 370, + 418 + ], + "spans": [ + { + "bbox": [ + 297, + 405, + 370, + 418 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "spans": [ + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "type": "text", + "content": "In this paper, we formulate a new learning paradigm for the task of 6D pose estimation. Separating the task into texture and pose learning enables a novel self-supervision scheme to adapt pretrained 6D pose estimators to the real domain. Only a few unlabelled images are needed within this framework to generate pixel-perfect supervision. With this reformulation, we successfully bridge the gap between synthetic pre-training and real data usage through geometry-transfer from pixel-perfect in silico data onto data-induced appearance information. It successfully addresses the difficulty in optimising pose with raw 2D observations and further prunes the need for additional supervisions as guidance. Although initialised with much inferior accuracy, it improves the state-of-the-art self-supervision RGB method by an impressive " + }, + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "type": "text", + "content": " while even being superior to RGBD methods by " + }, + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 296, + 426, + 537, + 628 + ], + "type": "text", + "content": ". We further demonstrates substantial generalisation improvements towards real-world scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 628, + 538, + 737 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 628, + 538, + 737 + ], + "spans": [ + { + "bbox": [ + 296, + 628, + 538, + 737 + ], + "type": "text", + "content": "For limitations, though we require as little as nine images for texture learning, they shall have mild occlusions to provide sufficient appearance information of the objects. We also share current limitations with other model-based pose estimation methods and hence need a CAD model to ensure a precise object-centric neural representation. For future works, we strongly believe that our reformulation can impact self-supervision pipelines also for unseen objects with joint optimisation of both geometry and texture." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4848" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "spans": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 40, + 116, + 280, + 738 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 44, + 116, + 279, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 116, + 279, + 191 + ], + "spans": [ + { + "bbox": [ + 44, + 116, + 279, + 191 + ], + "type": "text", + "content": "[1] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: speeded up robust features. In Ales Leonardis, Horst Bischof, and Axel Pinz, editors, Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part I, volume 3951 of Lecture Notes in Computer Science, pages 404-417. Springer, 2006. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 44, + 194, + 280, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 194, + 280, + 260 + ], + "spans": [ + { + "bbox": [ + 44, + 194, + 280, + 260 + ], + "type": "text", + "content": "[2] Konstantinos Bousmalis, Nathan Silberman, David Dohan, Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 95-104. IEEE Computer Society, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 261, + 280, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 261, + 280, + 348 + ], + "spans": [ + { + "bbox": [ + 45, + 261, + 280, + 348 + ], + "type": "text", + "content": "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In David J. Fleet, Tomás Pajdla, Bernt Schiele, and Tinne Tuyte-laars, editors, Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II, volume 8690 of Lecture Notes in Computer Science, pages 536-551. Springer, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 349, + 280, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 349, + 280, + 392 + ], + "spans": [ + { + "bbox": [ + 45, + 349, + 280, + 392 + ], + "type": "text", + "content": "[4] Kai Chen, Rui Cao, Stephen James, Yichuan Li, Yun-Hui Liu, Pieter Abbeel, and Qi Dou. Sim-to-real 6d object pose estimation via iterative self-training for robotic bin-picking. arXiv preprint arXiv:2204.07049, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 394, + 280, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 394, + 280, + 481 + ], + "spans": [ + { + "bbox": [ + 45, + 394, + 280, + 481 + ], + "type": "text", + "content": "[5] Xu Chen, Zijian Dong, Jie Song, Andreas Geiger, and Otmar Hilliges. Category level object pose estimation via neural analysis-by-synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI, volume 12371 of Lecture Notes in Computer Science, pages 139-156. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 482, + 280, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 482, + 280, + 526 + ], + "spans": [ + { + "bbox": [ + 45, + 482, + 280, + 526 + ], + "type": "text", + "content": "[6] Alvaro Collet, Manuel Martínez, and Siddhartha S. Srinivasa. The MOPED framework: Object recognition and pose estimation for manipulation. Int. J. Robotics Res., 30(10):1284-1306, 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 527, + 280, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 527, + 280, + 581 + ], + "spans": [ + { + "bbox": [ + 45, + 527, + 280, + 581 + ], + "type": "text", + "content": "[7] Alvaro Collet and Siddhartha S. Srinivasa. Efficient multiview object recognition and full pose estimation. In IEEE International Conference on Robotics and Automation, ICRA 2010, Anchorage, Alaska, USA, 3-7 May 2010, pages 2050-2055. IEEE, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 583, + 280, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 583, + 280, + 648 + ], + "spans": [ + { + "bbox": [ + 45, + 583, + 280, + 648 + ], + "type": "text", + "content": "[8] Xinke Deng, Yu Xiang, Arsalan Mousavian, Clemens Eppner, Timothy Bretl, and Dieter Fox. Self-supervised 6d object pose estimation for robot manipulation. In 2020 IEEE International Conference on Robotics and Automation, ICRA 2020, Paris, France, May 31 - August 31, 2020, pages 3665-3671. IEEE, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 650, + 280, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 650, + 280, + 692 + ], + "spans": [ + { + "bbox": [ + 45, + 650, + 280, + 692 + ], + "type": "text", + "content": "[9] Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer, Youssef Zidan, Dmitry Olefir, Mohamad Elbadrawy, Ahsan Lodhi, and Harinandan Katam. Blenderproc. CoRR, abs/1911.01911, 2019. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 40, + 694, + 280, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 694, + 280, + 738 + ], + "spans": [ + { + "bbox": [ + 40, + 694, + 280, + 738 + ], + "type": "text", + "content": "[10] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting selfocclusion for direct 6d pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 538, + 738 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 318, + 98, + 538, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 98, + 538, + 119 + ], + "spans": [ + { + "bbox": [ + 318, + 98, + 538, + 119 + ], + "type": "text", + "content": "Montreal, QC, Canada, October 10-17, 2021, pages 12376-12385. IEEE, 2021. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 121, + 538, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 121, + 538, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 121, + 538, + 175 + ], + "type": "text", + "content": "[11] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pages 1310-1319. IEEE Computer Society, 2017. 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 176, + 537, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 176, + 537, + 230 + ], + "spans": [ + { + "bbox": [ + 299, + 176, + 537, + 230 + ], + "type": "text", + "content": "[12] Stefano Gasperini, Jan Haug, Mohammad-Ali Nikouei Mahani, Alvaro Marcos-Ramiro, Nassir Navab, Benjamin Busam, and Federico Tombari. Certainnet: Sampling-free uncertainty estimation for object detection. IEEE Robotics Autom. Lett., 7(2):698-705, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 231, + 537, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 231, + 537, + 296 + ], + "spans": [ + { + "bbox": [ + 299, + 231, + 537, + 296 + ], + "type": "text", + "content": "[13] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 3193-3203. Computer Vision Foundation / IEEE, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 297, + 538, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 297, + 538, + 406 + ], + "spans": [ + { + "bbox": [ + 299, + 297, + 538, + 406 + ], + "type": "text", + "content": "[14] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary R. Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Kyoung Mu Lee, Yasuyuki Matsushita, James M. Rehg, and Zhanyi Hu, editors, Computer Vision - ACCV 2012 - 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part I, volume 7724 of Lecture Notes in Computer Science, pages 548-562. Springer, 2012. 1, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 408, + 538, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 408, + 538, + 483 + ], + "spans": [ + { + "bbox": [ + 299, + 408, + 538, + 483 + ], + "type": "text", + "content": "[15] Tomas Hodan, Pavel Haluza, Stepan Obdrzalek, Jiri Matas, Manolis I. A. Lourakis, and Xenophon Zabulis. T-LESS: an RGB-D dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision, WACV 2017, Santa Rosa, CA, USA, March 24-31, 2017, pages 880-888. IEEE Computer Society, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 484, + 537, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 484, + 537, + 561 + ], + "spans": [ + { + "bbox": [ + 299, + 484, + 537, + 561 + ], + "type": "text", + "content": "[16] Tomás Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP challenge 2020 on 6d object localization. In Adrien Bartoli and Andrea Fusiello, editors, Computer Vision - ECCV 2020 Workshops - Glasgow, UK, August 23-28, 2020, Proceedings, Part II, volume 12536 of Lecture Notes in Computer Science, pages 577-594. Springer, 2020. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 562, + 538, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 562, + 538, + 627 + ], + "spans": [ + { + "bbox": [ + 299, + 562, + 538, + 627 + ], + "type": "text", + "content": "[17] HyunJun Jung, Patrick Ruhkamp, Guangyao Zhai, Nikolas Brasch, Yitong Li, Yannick Verdie, Jifei Song, Yiren Zhou, Anil Armagan, Slobodan Ilic, et al. Is my depth ground-truth good enough? hammer-highly accurate multimodal dataset for dense 3d scene regression. arXiv preprint arXiv:2205.04565, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 628, + 538, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 628, + 538, + 692 + ], + "spans": [ + { + "bbox": [ + 299, + 628, + 538, + 692 + ], + "type": "text", + "content": "[18] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D dataset for 6d pose estimation of 3d objects. In 2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2767-2776. IEEE, 2019. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 694, + 538, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 694, + 538, + 738 + ], + "spans": [ + { + "bbox": [ + 299, + 694, + 538, + 738 + ], + "type": "text", + "content": "[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based 3d detection and 6d pose estimation great again. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy," + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4849" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 41, + 97, + 280, + 737 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 59, + 97, + 279, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 97, + 279, + 119 + ], + "spans": [ + { + "bbox": [ + 59, + 97, + 279, + 119 + ], + "type": "text", + "content": "October 22-29, 2017, pages 1530-1538. IEEE Computer Society, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 41, + 121, + 280, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 121, + 280, + 164 + ], + "spans": [ + { + "bbox": [ + 41, + 121, + 280, + 164 + ], + "type": "text", + "content": "[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 165, + 280, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 165, + 280, + 208 + ], + "spans": [ + { + "bbox": [ + 41, + 165, + 280, + 208 + ], + "type": "text", + "content": "[21] Zhengfei Kuang, Kyle Olszewski, Mengei Chai, Zeng Huang, Panos Achlioptas, and Sergey Tulyakov. Neroic: Neural rendering of objects from online image collections. ACM Trans. Graph., 41(4), jul 2022. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 208, + 280, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 208, + 280, + 275 + ], + "spans": [ + { + "bbox": [ + 41, + 208, + 280, + 275 + ], + "type": "text", + "content": "[22] Abhijit Kundu, Yin Li, and James M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 3559-3568. Computer Vision Foundation / IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 275, + 280, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 275, + 280, + 362 + ], + "spans": [ + { + "bbox": [ + 41, + 275, + 280, + 362 + ], + "type": "text", + "content": "[23] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypose: Consistent multi-view multi-object 6d pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII, volume 12362 of Lecture Notes in Computer Science, pages 574-591. Springer, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 41, + 362, + 280, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 362, + 280, + 428 + ], + "spans": [ + { + "bbox": [ + 41, + 362, + 280, + 428 + ], + "type": "text", + "content": "[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Single-view robot pose and joint angle estimation via render & compare. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1654-1663. Computer Vision Foundation / IEEE, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 41, + 429, + 280, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 429, + 280, + 473 + ], + "spans": [ + { + "bbox": [ + 41, + 429, + 280, + 473 + ], + "type": "text", + "content": "[25] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 41, + 474, + 280, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 474, + 280, + 506 + ], + "spans": [ + { + "bbox": [ + 41, + 474, + 280, + 506 + ], + "type": "text", + "content": "[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. Int. J. Comput. Vis., 128(3):657-678, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 506, + 280, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 506, + 280, + 561 + ], + "spans": [ + { + "bbox": [ + 41, + 506, + 280, + 561 + ], + "type": "text", + "content": "[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: bundle-adjusting neural radiance fields. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 5721-5731. IEEE, 2021. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 562, + 280, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 562, + 280, + 606 + ], + "spans": [ + { + "bbox": [ + 41, + 562, + 280, + 606 + ], + "type": "text", + "content": "[28] David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision, Kerkyra, Corfu, Greece, September 20-25, 1999, pages 1150-1157. IEEE Computer Society, 1999. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 41, + 607, + 280, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 607, + 280, + 671 + ], + "spans": [ + { + "bbox": [ + 41, + 607, + 280, + 671 + ], + "type": "text", + "content": "[29] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6841-6850, 2019. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 41, + 672, + 280, + 737 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 672, + 280, + 737 + ], + "spans": [ + { + "bbox": [ + 41, + 672, + 280, + 737 + ], + "type": "text", + "content": "[30] Fabian Manhardt, Wadim Kehl, and Adrien Gaidon. ROI10D: monocular lifting of 2d detection to 6d pose and metric shape. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2069-2078. Computer Vision Foundation / IEEE, 2019. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 537, + 738 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 299, + 98, + 537, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 98, + 537, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 98, + 537, + 175 + ], + "type": "text", + "content": "[31] Fabian Manhardt, Wadim Kehl, Nassir Navab, and Federico Tombari. Deep model-based 6d pose refinement in RGB. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XIV, volume 11218 of Lecture Notes in Computer Science, pages 833-849. Springer, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 176, + 537, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 176, + 537, + 230 + ], + "spans": [ + { + "bbox": [ + 299, + 176, + 537, + 230 + ], + "type": "text", + "content": "[32] Fabian Manhardt, Gu Wang, Benjamin Busam, Manuel Nickel, Sven Meier, Luca Minciullo, Xiangyang Ji, and Nassir Navab. Cps++: Improving class-level 6d pose and shape estimation from monocular images with self-supervised learning. arXiv preprint arXiv:2003.05848, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 230, + 537, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 230, + 537, + 263 + ], + "spans": [ + { + "bbox": [ + 299, + 230, + 537, + 263 + ], + "type": "text", + "content": "[33] Éric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. IEEE Trans. Vis. Comput. Graph., 22(12):2633-2651, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 264, + 537, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 264, + 537, + 340 + ], + "spans": [ + { + "bbox": [ + 299, + 264, + 537, + 340 + ], + "type": "text", + "content": "[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 7210-7219. Computer Vision Foundation / IEEE, 2021. 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 341, + 537, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 341, + 537, + 406 + ], + "spans": [ + { + "bbox": [ + 299, + 341, + 537, + 406 + ], + "type": "text", + "content": "[35] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6331-6341. IEEE, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 407, + 537, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 407, + 537, + 495 + ], + "spans": [ + { + "bbox": [ + 299, + 407, + 537, + 495 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 405-421. Springer, 2020. 3, 4, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 495, + 537, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 495, + 537, + 551 + ], + "spans": [ + { + "bbox": [ + 299, + 495, + 537, + 551 + ], + "type": "text", + "content": "[37] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 7667-7676. IEEE, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 551, + 537, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 551, + 537, + 628 + ], + "spans": [ + { + "bbox": [ + 299, + 551, + 537, + 628 + ], + "type": "text", + "content": "[38] Kiru Park, Timothy Patten, and Markus Vincze. Neural object learning for 6d pose estimation using a few cluttered images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV, volume 12349 of Lecture Notes in Computer Science, pages 656-673. Springer, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 628, + 537, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 628, + 537, + 692 + ], + "spans": [ + { + "bbox": [ + 299, + 628, + 537, + 692 + ], + "type": "text", + "content": "[39] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4561-4570. Computer Vision Foundation / IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 693, + 537, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 693, + 537, + 738 + ], + "spans": [ + { + "bbox": [ + 299, + 693, + 537, + 738 + ], + "type": "text", + "content": "[40] Mahdi Rad and Vincent Lepetit. BB8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In IEEE International Conference on Computer Vision, ICCV 2017," + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4850" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 41, + 97, + 280, + 731 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 59, + 97, + 278, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 97, + 278, + 120 + ], + "spans": [ + { + "bbox": [ + 59, + 97, + 278, + 120 + ], + "type": "text", + "content": "Venice, Italy, October 22-29, 2017, pages 3848-3856. IEEE Computer Society, 2017. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 41, + 122, + 279, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 122, + 279, + 197 + ], + "spans": [ + { + "bbox": [ + 41, + 122, + 279, + 197 + ], + "type": "text", + "content": "[41] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: generative radiance fields for 3d-aware image synthesis. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 199, + 280, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 199, + 280, + 242 + ], + "spans": [ + { + "bbox": [ + 41, + 199, + 280, + 242 + ], + "type": "text", + "content": "[42] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 244, + 280, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 244, + 280, + 309 + ], + "spans": [ + { + "bbox": [ + 41, + 244, + 280, + 309 + ], + "type": "text", + "content": "[43] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 310, + 280, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 310, + 280, + 387 + ], + "spans": [ + { + "bbox": [ + 41, + 310, + 280, + 387 + ], + "type": "text", + "content": "[44] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In Vitomir Struc and Francisco Gomez Fernandez, editors, 8th International Conference on 3D Vision, 3DV 2020, Virtual Event, Japan, November 25-28, 2020, pages 291-300. IEEE, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 41, + 388, + 280, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 388, + 280, + 442 + ], + "spans": [ + { + "bbox": [ + 41, + 388, + 280, + 442 + ], + "type": "text", + "content": "[45] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 41, + 444, + 280, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 444, + 280, + 497 + ], + "spans": [ + { + "bbox": [ + 41, + 444, + 280, + 497 + ], + "type": "text", + "content": "[46] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 41, + 500, + 280, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 500, + 280, + 587 + ], + "spans": [ + { + "bbox": [ + 41, + 500, + 280, + 587 + ], + "type": "text", + "content": "[47] Gu Wang, Fabian Manhardt, Jianzhuon Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, volume 12346 of Lecture Notes in Computer Science, pages 108-125. Springer, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 589, + 280, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 589, + 280, + 653 + ], + "spans": [ + { + "bbox": [ + 41, + 589, + 280, + 653 + ], + "type": "text", + "content": "[48] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16611-16621. Computer Vision Foundation / IEEE, 2021. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 655, + 280, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 655, + 280, + 731 + ], + "spans": [ + { + "bbox": [ + 41, + 655, + 280, + 731 + ], + "type": "text", + "content": "[49] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2642-2651. Computer Vision Foundation / IEEE, 2019. 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 538, + 733 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 299, + 98, + 538, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 98, + 538, + 164 + ], + "spans": [ + { + "bbox": [ + 299, + 98, + 538, + 164 + ], + "type": "text", + "content": "[50] Pengyuan Wang, HyunJun Jung, Yitong Li, Siyuan Shen, Rahul Parthasarathy Srikanth, Loranzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Phocal: A multimodal dataset for category-level object pose estimation with photometrically challenging objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 299, + 165, + 538, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 165, + 538, + 229 + ], + "spans": [ + { + "bbox": [ + 299, + 165, + 538, + 229 + ], + "type": "text", + "content": "[51] Pengyuan Wang, Fabian Manhardt, Luca Minciullo, Lorenzo Garattoni, Sven Meier, Nassir Navab, and Benjamin Busam. Demograsp: Few-shot learning for robotic grasping with human demonstration. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5733-5740. IEEE, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 231, + 538, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 231, + 538, + 264 + ], + "spans": [ + { + "bbox": [ + 299, + 231, + 538, + 264 + ], + "type": "text", + "content": "[52] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, abs/2102.07064, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 299, + 266, + 538, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 266, + 538, + 341 + ], + "spans": [ + { + "bbox": [ + 299, + 266, + 538, + 341 + ], + "type": "text", + "content": "[53] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 343, + 538, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 343, + 538, + 408 + ], + "spans": [ + { + "bbox": [ + 299, + 343, + 538, + 408 + ], + "type": "text", + "content": "[54] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 3907-3916. Computer Vision Foundation / IEEE, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 410, + 538, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 410, + 538, + 464 + ], + "spans": [ + { + "bbox": [ + 299, + 410, + 538, + 464 + ], + "type": "text", + "content": "[55] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Tsung-Yi Lin, Alberto Rodriguez, and Phillip Isola. NeRF-Supervision: Learning dense object descriptors from neural radiance fields. In IEEE Conference on Robotics and Automation (ICRA), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 466, + 538, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 466, + 538, + 521 + ], + "spans": [ + { + "bbox": [ + 299, + 466, + 538, + 521 + ], + "type": "text", + "content": "[56] Lin Yen-Chen, Pete Florence, Jonathan T. Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. iNeRF: Inverting neural radiance fields for pose estimation. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 522, + 538, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 522, + 538, + 587 + ], + "spans": [ + { + "bbox": [ + 299, + 522, + 538, + 587 + ], + "type": "text", + "content": "[57] Sergey Zakharov, Wadim Kehl, Arjun Bhargava, and Adrien Gaidon. Autolabeling 3d objects with differentiable rendering of SDF shape priors. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 12221-12230. Computer Vision Foundation / IEEE, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 589, + 538, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 589, + 538, + 643 + ], + "spans": [ + { + "bbox": [ + 299, + 589, + 538, + 643 + ], + "type": "text", + "content": "[58] Sergey Zakharov, Wadim Kehl, and Slobodan Ilic. Deceptionnet: Network-driven domain randomization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 532-541. IEEE, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 644, + 538, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 644, + 538, + 698 + ], + "spans": [ + { + "bbox": [ + 299, + 644, + 538, + 698 + ], + "type": "text", + "content": "[59] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6d pose object detector and refiner. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 1941-1950. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 700, + 538, + 733 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 700, + 538, + 733 + ], + "spans": [ + { + "bbox": [ + 299, + 700, + 538, + 733 + ], + "type": "text", + "content": "[60] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conference on Computer Vision and" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 751, + 297, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 297, + 761 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 297, + 761 + ], + "type": "text", + "content": "4851" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + }, + { + "para_blocks": [], + "discarded_blocks": [ + { + "bbox": [ + 60, + 98, + 279, + 130 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 98, + 279, + 130 + ], + "spans": [ + { + "bbox": [ + 60, + 98, + 279, + 130 + ], + "type": "text", + "content": "Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5745-5753. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "spans": [ + { + "bbox": [ + 279, + 751, + 298, + 760 + ], + "type": "text", + "content": "4852" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_content_list.json b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..af432b706726ba6ac5ca0f4e8f27502675ad861b --- /dev/null +++ b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_content_list.json @@ -0,0 +1,1472 @@ +[ + { + "type": "text", + "text": "Text with Knowledge Graph Augmented Transformer for Video Captioning", + "text_level": 1, + "bbox": [ + 106, + 128, + 867, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Gu $^{1,3}$ , Guang Chen $^{2}$ , Yufei Wang $^{2}$ , Libo Zhang $^{3,1*}$ , Tiejian Luo $^{1}$ , Longyin Wen $^{2}$ $^{1}$ University of Chinese Academy of Sciences, Beijing, China \n $^{2}$ ByteDance Inc., San Jose, USA", + "bbox": [ + 153, + 178, + 816, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Institute of Software Chinese Academy of Sciences, Beijing, China", + "bbox": [ + 214, + 233, + 754, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guxin21@mails.ucas.edu.cn, guang.chen@bytedance.com, yufei.wang@bytedance.com", + "bbox": [ + 142, + 253, + 826, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "libo@iscas.ac.cn, tjluo@ucas.ac.cn, longyin.wen@bytedance.com", + "bbox": [ + 210, + 272, + 754, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video captioning aims to describe the content of videos using natural language. Although significant progress has been made, there is still much room to improve the performance for real-world applications, mainly due to the long-tail words challenge. In this paper, we propose a text with knowledge graph augmented transformer (TextKG) for video captioning. Notably, TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is designed to absorb additional knowledge, which models the interactions between the additional knowledge, e.g., pre-built knowledge graph, and the built-in information of videos, e.g., the salient object regions, speech transcripts, and video captions, to mitigate the long-tail words challenge. Meanwhile, the internal stream is designed to exploit the multi-modality information in videos (e.g., the appearance of video frames, speech transcripts, and video captions) to ensure the quality of caption results. In addition, the cross attention mechanism is also used in between the two streams for sharing information. In this way, the two streams can help each other for more accurate results. Extensive experiments conducted on four challenging video captioning datasets, i.e., YouCookII, ActivityNet Captions, MSR-VTT, and MSVD, demonstrate that the proposed method performs favorably against the state-of-the-art methods. Specifically, the proposed TextKG method outperforms the best published results by improving $18.7\\%$ absolute CIDEr scores on the YouCookII dataset.", + "bbox": [ + 75, + 354, + 472, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video captioning aims to generate a complete and natural sentence to describe video content, which attracts much", + "bbox": [ + 76, + 820, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attention in recent years. Generally, most existing methods [21,38,41,58] require a large amount of paired video and description data for model training. Several datasets, such as YouCookII [69], and ActivityNet Captions [19] are constructed to promote the development of video captioning field. Meanwhile, some methods [29, 40, 48, 72] also use the large-scale narrated video dataset HowTo100M [30] to pretrain the captioning model to further improve the accuracy.", + "bbox": [ + 500, + 323, + 893, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although significant progress has been witnessed, it is still a challenge for video captioning methods to be applied in real applications, mainly due to the long-tail issues of words. Most existing methods [29, 40, 48, 72] attempt to design powerful neural networks, trained on the large-scale video-text datasets to make the network learn the relations between video appearances and descriptions. However, it is pretty tough for the networks to accurately predict the objects, properties, or behaviors that are infrequently or never appearing in training data. Some methods [14, 71] attempt to use knowledge graph to exploit the relations between objects for long-tail challenge in image or video captioning, which produces promising results.", + "bbox": [ + 496, + 460, + 893, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a text with knowledge graph augmented transformer (TextKG), which integrates additional knowledge in knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge. TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is used to absorb additional knowledge to help mitigate long-tail words challenge by modeling the interactions between the additional knowledge in pre-built knowledge graph, and the built-in information of videos, such as the salient object regions in each frame, speech transcripts, and video captions. Specifically, the information is first retrieved from the pre-built knowledge graphs based on the detected salient objects. After that, we combine the features of the retrieved information, the appearance features of detected salient objects, the features of speech transcripts and", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author:libo@iscas.ac.cn, Libo Zhang was supported by Youth Innovation Promotion Association, CAS (2020111). This work was done during internship at ByteDance Inc.", + "bbox": [ + 75, + 862, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18941", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "captions, then feed them into the external stream of TextKG to model the interactions. The internal stream is designed to exploit the multi-modality information in videos, such as the appearance of video frames, speech transcripts and video captions, which can ensure the quality of caption results. To share information between two streams, the cross attention mechanism is introduced. In this way, the two streams can obtain the required modal information from each other for generating more accurate results. The architecture of the proposed method is shown in Figure 1.", + "bbox": [ + 75, + 90, + 467, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several experiments conducted on four challenging datasets, i.e., YouCookII [69], ActivityNet Captions [19], MSR-VTT [56], and MSVD [3] demonstrate that the proposed method performs favorably against the state-of-the-art methods. Notably, our TextKG method outperforms the best published results by improving $18.7\\%$ and $3.2\\%$ absolute CIDEr scores in the paragraph-level evaluation mode on the YouCookII and Activity-Net Captions datasets.", + "bbox": [ + 75, + 242, + 467, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 376, + 217, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video captioning attracts much attention of researchers in recent years. The best practice has been achieved by attention-based methods, which attempts to associate visual components with sentences in videos. Some of them focus on designing powerful network architectures. VLM [55] and VideoBERT [43] take the visual and text modalities as input, and use a shared transformer to construct a task-agnostic video-language model. ViLBERT [28] processes visual and linguistic information separately with two parallel streams, and then use the attention mechanism to model the interactions between visual and language features. Instead of using the separate encoder-decoder architecture, MART [21] designs a shared encoder-decoder network and augments it with the memory module. ActBert [72] uses local regional features to learn better visual-language alignment. WLT [35] takes audio features as an additional input, and uses context fusion to generate multimodal features.", + "bbox": [ + 75, + 402, + 467, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, some methods [10,14,15,62,65,71] focus on exploiting prior knowledge to provide semantic correlations and constraints between objects for image or video captioning, producing promising results. ORG-TRL [64] uses the knowledge information in the language model (BERT) to provide candidate words for video captioning. In contrast, we propose a two-stream transformer for video captioning, with the internal stream used to exploit multi-modality information in videos, and the external stream used to model the interactions between the additional knowledge and the built-in information of videos. These two streams use the cross-attention mechanism to share information in different modalities for generating more accurate results.", + "bbox": [ + 75, + 659, + 467, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-and-language representation learning is a hot topic in recent years. ViLBERT [28], LXMERT [46], UNITER [6], UNIMO [25] and Unified-VL [68] learn", + "bbox": [ + 76, + 856, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the representations between image and text, while Univl [29], VideoBERT [43], ActBERT [72] and MV-GPT [40] learn the representations between videos and transcripts. Notably, most of these methods attempt to learn powerful vision-and-language representations by pre-training the models on the large-scale datasets, e.g., Howto100M [31] and WebVid-2M [1], and then finetune them on downstream tasks such as video captioning, video-text retrieval and visual question answering. In contrast, our TextKG method uses the speech transcripts as the text to model the visual and linguistic representations and integrate the additional knowledge in knowledge graph to mitigate long-tail words challenge in video captioning.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge graph in NLP. Knowledge graph is an useful tool to indicate the real-world entities and their relations, which provides rich structured knowledge facts for language modeling. Large-scale knowledge graphs are used to train knowledge enhanced language models for various natural language processing (NLP) tasks. CoLAKE [44] proposes to inject the knowledge context of an entity, and to jointly learn the contextualized representation for both language and knowledge by a unified structure. ERNIE [63] enhances BERT architecture to better integrate the knowledge information and textual information. KEPLER [52] not only improves language models by integrating factual knowledge, but also generates text-enhanced knowledge representation. JAKET [59] proposes a joint pre-training framework to model knowledge graph and language simultaneously. Inspired by CoLAKE, our method jointly learns the representations of vision, language and knowledge, and enhances the joint visual-language representations by retrieving relevant knowledge in knowledge graphs.", + "bbox": [ + 496, + 287, + 890, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 500, + 588, + 648, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As mentioned above, our method aims to integrate additional knowledge using knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge in the field of video captioning. We design a text with knowledge graph augmented transformer, i.e., a two-stream transformer formed by the external and internal streams. Both streams are constructed by $N$ sets of alternately stacked self-attention and cross-attention blocks. Specifically, we first use a detector to generate the salient object regions in each video frame, and use the automatic speech recognition (ASR) method [2] to generate the speech transcripts in videos. After that, the class labels of detected salient objects are used to retrieve the prior knowledge in knowledge graphs. Then, the appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior knowledge, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent words of the predicted captions. The overall architecture of the proposed TextKG is shown in Figure 1.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18942", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6ce9d9fb5365b87c9942daa62558c558037a08a6e7a75d9ff93e0c0f3387b6e9.jpg", + "image_caption": [ + "Figure 1. The architecture of our TextKG method, which is formed the external and internal streams. Each stream is stacked with $N$ sets of multi-modality self-attention and cross-attention modules. The cross attention modules are designed to align tokens in different modalities." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 872, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Two-Stream Transformer", + "text_level": 1, + "bbox": [ + 76, + 448, + 308, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As described above, our two-stream transformer is formed by the external stream and internal stream. Several self-attention and cross-attention blocks are interleaved to construct the two streams.", + "bbox": [ + 75, + 470, + 468, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\{y_{1}, y_{2}, \\dots, y_{l}\\}$ be the predicted captions, where $y_{i}$ is the index of the $i$ -th word in the dictionary, $\\{p_{1}, p_{2}, \\dots, p_{l}\\}$ to be the predicted probabilities of words in the dictionary, where $p_{i}$ is the probability of the $i$ -th word. The index of the $i$ -th word $y_{i}$ is computed as $y_{i} = \\arg \\max_{i} p_{i}$ . Meanwhile, let $z_{i}^{\\mathrm{ext}}$ and $z_{i}^{\\mathrm{int}}$ to be the output probabilities of the external and internal streams of the $i$ -th word. Thus, we have $p_{i} = \\omega_{1} z_{i}^{\\mathrm{ext}} + \\omega_{2} z_{i}^{\\mathrm{int}}$ , where $\\omega_{1}$ and $\\omega_{2}$ are the hyperparameters used to balance the outputs of the two-streams. In this way, $y_{i}$ is computed as", + "bbox": [ + 76, + 532, + 470, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} = \\arg \\max _ {i} \\left(\\omega_ {1} z _ {i} ^ {\\text {e x t}} + \\omega_ {2} z _ {i} ^ {\\text {i n t}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 691, + 468, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following sections, we will describe each module in our two-stream transformer in more details.", + "bbox": [ + 75, + 719, + 468, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modality Self-Attention Module. As shown in Figure 1, we use the self-attention module [49] in both the external and internal streams to model the interactions among multi-modality information. Specifically, for the external stream, the concatenation of feature embeddings of detected objects $\\mathcal{F}_{\\mathrm{r}}$ , retrieved prior knowledge $\\mathcal{F}_{\\mathrm{k}}$ , speech transcripts $\\mathcal{F}_{\\mathrm{s}}$ , and predicted video captions $\\mathcal{F}_{\\mathrm{c}}$ are fed into the self-attention module to model the interactions, i.e., $\\mathcal{X}_{\\mathrm{ext}} = \\mathrm{Concat}(\\mathcal{F}_{\\mathrm{r}}, \\mathcal{F}_{\\mathrm{k}}, \\mathcal{F}_{\\mathrm{s}}, \\mathcal{F}_{\\mathrm{c}})$ . Meanwhile, for the internal stream, the concatenation of feature embeddings of speech", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transcripts $\\mathcal{F}_{\\mathrm{s}}$ , predicted video captions $\\mathcal{F}_{\\mathrm{c}}$ , and video frame $\\mathcal{F}_{\\mathrm{v}}$ are fed into the self-attention model to exploit the interactions, i.e., $\\mathcal{X}_{\\mathrm{int}} = \\operatorname {Concat}(\\mathcal{F}_{\\mathrm{s}},\\mathcal{F}_{\\mathrm{c}},\\mathcal{F}_{\\mathrm{v}})$ . The self-attention module computes the interactions as follows,", + "bbox": [ + 498, + 448, + 890, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} + M\\right) V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 518, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is the feature dimension of queries $Q$ and keys $K$ . We have $Q = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{Q}}$ , $K = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{K}}$ , $V = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{V}}$ , where $W^{\\mathrm{Q}}$ , $W^{\\mathrm{K}}$ , and $W^{\\mathrm{V}}$ are learnable parameters, and $\\mathcal{X}_{(\\cdot)} \\in \\{\\mathcal{X}_{\\mathrm{ext}}, \\mathcal{X}_{\\mathrm{int}}\\}$ are the concatenated feature embeddings for the external and internal streams, respectively.", + "bbox": [ + 496, + 551, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [21], we introduce a mask matrix $M$ in attention function (2) of both the external and internal streams to prevent the model from seeing future words. That is, for the $i$ -th caption token, we set $M_{i,j} = 0$ for $j = 1,\\dots ,i$ , and set $M_{i,j} = -\\inf$ for $j > i$ . Meanwhile, for the external stream, we set $M_{i,j} = -\\inf$ to prevent the associations between irrelevant retrieved prior knowledge and the detected salient object regions.", + "bbox": [ + 496, + 628, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modality Cross-Attention Module. Besides self-attention module, we use the cross-attention module to align the interactions modeled by both the external and internal streams. Specifically, we compute the affinity matrix to guide the alignments between the modeled interactions by injecting the information. Similar to the self-attention module, we also introduce the mask matrix into the attention function to compute the cross-attention. Intuitively, the retrieved prior knowledge should not have direct effect on the predicted captions, but through the detected salient objects.", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18943", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, we set the corresponding elements to $-\\inf$ to prevent the retrieved prior knowledge directly influencing the predicted captions.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Optimization", + "text_level": 1, + "bbox": [ + 76, + 143, + 217, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross entropy loss function is used to guide the training of our TextKG method. Given the ground-truth indices of previous $(i - 1)$ words and the concatenated of two-streams $\\mathcal{X}_{\\mathrm{ext}}$ and $\\mathcal{X}_{\\mathrm{int}}$ , we can get the predictions of the current $i$ -th word. After that, the training loss of our method is computed as", + "bbox": [ + 75, + 167, + 467, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\sum_ {i = 1} ^ {l} \\left(\\lambda_ {1} \\log z _ {i} ^ {\\text {e x t}} + \\lambda_ {2} \\log z _ {i} ^ {\\text {i n t}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 266, + 468, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{i}^{\\mathrm{ext}}$ and $z_{i}^{\\mathrm{int}}$ are the $i$ -th output word of the external and internal streams, $\\lambda_{1}$ and $\\lambda_{2}$ are the preset parameters used to balance the two streams, and $l$ is the total length of predicted captions. The Adam optimization algorithm [17] with an initial learning rate of $1e - 4$ , $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ and $L2$ weight decay of 0.01 is used to train the model. The Warmup strategy is adopted in training phase by linearly increasing the learning rate from 0 to the initial learning rate for the first $10\\%$ of training epochs, and linearly decreasing to 0 for the remaining $90\\%$ of epochs.", + "bbox": [ + 75, + 295, + 468, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multimodal Tokens", + "text_level": 1, + "bbox": [ + 76, + 454, + 264, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent captions after tokenization. We will describe the feature extraction and tokenization processes of the aforementioned embeddings.", + "bbox": [ + 75, + 477, + 467, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video tokens encode the appearance and motion information of video frames. Specifically, the aligned appearance and motion features form the video tokens. In each second of the video, we sample 2 frames to extract features. Following [21], we use the \"Flatten-673\" layer in ResNet200 [12] to extract appearance features, and the \"global pool\" layer in BN-Inception [16] to extract the motion features on the YouCookII [69] and ActivityNet Captions [19] datasets. Meanwhile, for fair comparison, we use the InceptionResNetV2 [45] and C3D [11] models to extract appearance and motion features on the MSR-VTT [56] and MSVD [3] datasets, following the methods [58, 64, 66].", + "bbox": [ + 75, + 568, + 467, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region tokens are used to describe the visual information of the detected salient objects. We use the Faster R-CNN method [37] to extract visual features for detected object regions, which is pretrained on the Visual Genome dataset [20], and the $N_{\\mathrm{r}}$ detected objects with highest confidence score in each frame are reserved for video captioning.", + "bbox": [ + 75, + 750, + 467, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transcript tokens are used to encode the content of speech transcripts, which contains some key information in videos. The GloVe model [34] is used to extract features of each word in transcripts to form the transcript tokens.", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Caption tokens encode the information of predicted captions. Similar to transcript tokens, GloVe [34] is adopted to extract features of each word in the generated captions.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge tokens model the content of retrieved knowledge, which is important to mitigate the long-tail words challenge in video captioning. Our TextKG method uses a triple structure to encode the retrieved knowledge, formed by two object items and the relations between them, i.e., $(\\alpha_{\\mathrm{head}},\\alpha_{\\mathrm{tail}},\\rho_{\\mathrm{rel}})$ , where $\\alpha_{\\mathrm{head}}$ and $\\alpha_{\\mathrm{tail}}$ are the two nodes in knowledge graph, and $\\rho_{\\mathrm{rel}}$ indicates the edge encoding the relations between them. For example, $\\alpha_{\\mathrm{head}} = \"\\mathrm{knife}$ , $\\alpha_{\\mathrm{tail}} = \"\\mathrm{hard}$ , and $\\rho_{\\mathrm{rel}} = \"\\mathrm{has~property}$ . For each detected object $\\alpha_{\\mathrm{head}}$ , we aim to integrate prior context information for more accurate caption results. That is, we use the GloVe model to extract linguistic features of the node $\\alpha_{\\mathrm{tail}}$ and set the learnable embeddings for the edge $\\rho_{\\mathrm{rel}}$ . After that, we compute the summation of them to get the knowledge embeddings of the knowledge tokens.", + "bbox": [ + 498, + 137, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each detected object, we use its category name to retrieve knowledge (i.e., triples containing the category name of the detected object) from the knowledge graph. Notably, only a portion of the retrieved knowledge are highly related to the video. The rests are potentially harmful to the performance of caption generation as it may distract the training and inference and also increase computational cost. For example, to generate caption of a cooking tutorial video, the knowledge \"knife - used to -cut\" is more useful rather than \"knife - has property - hard\". Because cutting food may be an important procedure for food preparation. Thus, we design a primary rank mechanism to sort the retrieved knowledge based on their semantic similarity to the video. The cosine similarity between the embeddings of knowledge tokens and video transcripts is computed as the semantic similarity for ranking. The pretrained SBERT [36] model is used to extract feature embeddings of transcripts. Finally, $N_{\\mathrm{k}}$ knowledge items with the highest similarity scores are reserved for video captioning.", + "bbox": [ + 498, + 364, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Knowledge Graphs Construction", + "text_level": 1, + "bbox": [ + 500, + 664, + 790, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned above, we use the knowledge graph to retain the prior knowledge. The nodes in the graph could correspond to a noun, an adjective, a verb or an adverb, etc. The edges in the graph are used to encode the relations between different nodes. Apparently, it is important to include all key information of videos in knowledge graphs for accurate caption results. Thus, besides general knowledge graph covering most key knowledge in general scenarios, we also introduce the specific knowledge graph for each dataset covering the key knowledge in specific scenarios.", + "bbox": [ + 498, + 688, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General knowledge graph is designed to include most key information in general scenarios that we are interested, such as cooking and activity. Specifically, the general knowledge graph is exported from the public available giant knowledge", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18944", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "graph ConceptNet [42] by extracting key words1 in ConceptNet with the connected edges and neighboring nodes.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specific knowledge graph. Besides the general knowledge graph, we also construct the specific knowledge graph to cover key information in specific scenarios. We believe that the speech transcripts of videos contain most of the crucial information and use the speech transcripts as the source to construct the specific knowledge graph. We first use the automatic speech recognition (ASR) model [2] to convert the speech in all videos to transcripts. After that, we use the Stanford NLP library² to analyze the components in each sentence in transcripts to form a structured grammar tree. We collect the \"adjective and noun\", \"noun and noun\" and \"adverb and verb\" phrases from the grammar trees to construct the specific knowledge graph.", + "bbox": [ + 75, + 121, + 467, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 332, + 207, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 357, + 178, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "YoucookII contains 2,000 long untrimmed videos from 89 cooking recipes, including 1,333 videos for training, 457 videos for testing. The average length of videos is 5.3 minutes and each video is divided into 7.7 video clips on average with the manually labeled caption sentences.", + "bbox": [ + 75, + 380, + 467, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ActivityNet Captions is a large-scale challenging dataset for video captioning, with 10,024 videos for training, 4,926 for validation, and 5,044 for testing. The average length of each video is 2.1 minutes and is divided into 3.65 clips on average. Each clip is assigned with a manually labeled caption sentence. Since the testing set is not publicly available, following MART [21], we split the validation set into two subsets, i.e., ae-val with 2,460 videos for validation and aetest with 2,457 videos for testing.", + "bbox": [ + 75, + 455, + 467, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSR-VTT includes 10,000 video clips with 15 seconds on average. Each video clip contains 20 human annotated captions and the average length of each sentence is 9.28 words. Following [5,38,58], 6,513,497, and 2,990 videos are used for training, validation, and testing, respectively.", + "bbox": [ + 75, + 592, + 467, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSVD consists of 1970 video clips collected from YouTube and each clip is approximately 10 seconds long. Each clip is labeled with 40 sentences. Following [5,38,58], we split the dataset into 1,200,100,and 670 video clips for training, validation and testing, respectively.", + "bbox": [ + 75, + 667, + 467, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation metrics", + "text_level": 1, + "bbox": [ + 76, + 753, + 258, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compare the performance of the proposed method with other methods, five model-free automatic evaluation metrics are used, including BLEU@4 [33] (abbreviated to B) that are precision-based metric, METEOR [8] (abbreviated to M) that computes sentence-level scores, CIDEr [50]", + "bbox": [ + 75, + 777, + 467, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(abbreviated to C) that is consensus-based metric, ROUGE [26] (abbreviated to R) that uses longest common subsequence to compute the similarities between sentences, and Rep@4 [54] (abbreviated to Rep) that computes the redundancy score for every captions. Lower Rep indicates better model. The CIDEr score is used as the primary metric in evaluation. Notably, both micro-level and paragraph-level evaluation modes are used. The micro-level mode evaluates the predicted captions of each video clip independently, while the paragraph-level mode considers the relations between captions of different clips. Following [9, 21], the paragraph-level evaluation mode is used in the YouCookII and ActivityNet datasets, which contains multiple clips in each video.", + "bbox": [ + 500, + 90, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 316, + 715, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We sample every 2 frames per second in each video and use the pre-trained models to extract video features. For the YouCookII and ActivityNet Captions datasets, similar to [21], the ResNet-200 and BN-Inception models are used to extract the appearance and motion features, respectively. Following [58, 64, 66], the InceptionResNetV2 and C3D models are used to extract the appearance and motion features for the MSR-VTT and MSVD dataset, respectively. Meanwhile, the Faster R-CNN method using ResNet-101 as the backbone is adopted to detect object in each video frame, which is trained on the Visual Genome dataset [20]. Notably, only $N_{\\mathrm{r}} = 6$ classes of the detected objects with top scores are used to form region features in each frame.", + "bbox": [ + 496, + 340, + 890, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The caption sentences, speech transcripts, and retrieved knowledge in knowledge graph are split into individual words using the NLTK toolkit3, and the GloVe [34] algorithm is used to extract the word embeddings for each word to construct the transcript, caption, and knowledge tokens. The dimension of word embeddings is set to 300. The captions are truncated to 20 words and the maximum text length of transcripts is set to 300 words. For each detected object, we retrieve up to $N_{k} = 5$ pieces of knowledge items in knowledge graph depending on the relevant scores.", + "bbox": [ + 496, + 537, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, the proposed method is implemented using Pytorch. 3 blocks of multi-modality self-attention and cross-attention modules are used in the MSR-VTT [56] and MSVD [3] datasets, and 2 are used in the YouCookII [69] and ActivityNet Captions [19] datasets, to form the two-stream transformer, respectively. The feature dimension in each block is set to 768, and the number of heads in multi-head architecture is set to 12. The proposed method is trained using the Adam algorithm [17] with the initial learning rate of $1e - 4$ . The batch size in training phase is set to 6. The hyper-parameters $\\omega_{1},\\omega_{2},\\lambda_{1}$ and $\\lambda_{2}$ are set to 0.8, 0.2, 0.5 and 0.5, empirically.", + "bbox": [ + 496, + 691, + 890, + 871 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1The key words are exported from the speech transcripts and annotated ground-truth captions in the training sets of the caption datasets. 2https://stanfordnlp.github.io/CoreNLP/", + "bbox": [ + 76, + 862, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://github.com/nltk/nltk", + "bbox": [ + 517, + 886, + 669, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7f94f5089420798a9e46cb1f641c5af26bde2ba0880124b063499016cc0c688d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBMCRep
Van-Trans [70]7.615.732.37.8
Trans-XL [7]6.614.826.46.3
Trans-XLRG [22]6.614.725.96.0
MART [21]8.016.035.74.4
Van-Trans+COOT [9]11.119.855.65.7
COOT [9]11.319.957.26.7
TextKG14.022.175.92.8
Human---1.0
", + "bbox": [ + 78, + 88, + 470, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison to the State of the Art Methods", + "text_level": 1, + "bbox": [ + 76, + 282, + 449, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YouCookII dataset. To validate the effectiveness of the proposed method, we compare it to the state-of-the-art (SOTA) methods using the paragraph-level evaluation mode, reported in Table 1. As shown in Table 1, our TextKG method achieves the best results on the YouCookII dataset. Specifically, TextKG improves $18.7\\%$ absolute CIDEr score compared to the SOTA method COOT [9]. COOT [9] focuses on leveraging the hierarchy information and modeling the interactions between different levels granularity and different modalities, such as frames and words, clip and sentences, and videos and paragraphs, which is pre-trained on the large-scale HowTo100M [31] dataset. In contrast, our TextKG method aims to exploit additional knowledge in knowledge graph and multi-modality information in videos to improve the caption results.", + "bbox": [ + 75, + 306, + 468, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Beside the paragraph-level evaluation mode, we also report the evaluation results based on the micro-level mode on the YouCookII dataset in Table 2. As shown in Table 2, TextKG significantly outperforms other methods, including Univl [29] and AT [13], both of which use transcript information to enhance model performance by directly concatenating visual and transcript embedding. We argue that with the use of knowledge graphs, TextKG gives a better understanding of the content included in the speech transcripts and thus yields favorable results.", + "bbox": [ + 75, + 536, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ActivityNet Caption dataset. We also evaluate our TextKG method on the challenging ActivityNet Caption dataset in Table 3. As shown in Table 3, TextKG achieves the best results on 3 out of 4 metrics, i.e., BLEU, METOR, and CIDEr. The ActivityNet Caption dataset [19] contains a series of complex events with annotated caption sentences describing the events that occur. These events may occur over various periods of time and may co-occur. In contrast to existing methods [5, 38, 58, 66], our TextKG understands complex events more comprehensively with the help of exploiting information from transcripts and integrating the relevant knowledge in knowledge graph to enhance the common-sense information for video captioning.", + "bbox": [ + 75, + 686, + 468, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MSR-VTT dataset. The evaluation results on the MSR-", + "bbox": [ + 76, + 885, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9865f3ba52f68c5e4e28cb09a42691235fcd28f3c0eae356a791fe02b7c9657a.jpg", + "table_caption": [ + "Table 1. Evaluation results on the YouCookII val subset in the paragraph-level evaluation mode." + ], + "table_footnote": [], + "table_body": "
MethodInputBMRC
Masked Trans [70]V3.811.627.438
S3D [53]V3.29.526.131
VideoAsMT [18]V5.313.4--
SwinBERT [27]V915.637.3109
VideoBERT [43]V4.011.027.549
VideoBERT+S3D [43]V4.311.928.850
ActBERT [72]V5.414.330.665
AT [13]V+S9.017.836.7112
DPC [41]V+S2.818.1--
VALUE [24]V+S12.418.840.4130
Univl [29]V+S9.516.337.4115
MV-GPT [40]V+S13.317.635.5103
TextKGV+S11.718.440.2133
", + "bbox": [ + 504, + 89, + 890, + 294 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d72b6cdb96c3ea95823d5ba3249236f1495df23c344f11537e468fbc30bc0e5.jpg", + "table_caption": [ + "Table 2. Evaluation results on the YouCookII val subset in the micro-level evaluation mode. 'V' indicates the methods use video appearance information, and 'S' indicates the methods use the speech information." + ], + "table_footnote": [], + "table_body": "
MethodBMCRep
HSE [60]9.813.818.813.2
GVD [67]11.015.722.08.8
GVDsup [67]11.316.422.97.0
Van-Trans [70]9.815.622.27.8
Trans-XL [7]10.415.121.78.5
Trans-XLRG [22]10.214.820.48.9
MART [21]10.315.723.45.2
TextKG11.316.526.66.3
", + "bbox": [ + 501, + 375, + 893, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Evaluation results on the Activity-Net Captions ae-val subset in the paragraph-level evaluation mode.", + "bbox": [ + 498, + 526, + 892, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VTT dataset are reported in Table 4. For the fair comparison, we use the same visual features as the existing methods [5, 58, 64, 66]. In contrast to existing methods focusing on network architecture design to exploit visual information, our TextKG aims to exploit mutli-modality information in external knowledge graph and original videos, achieving the state-of-the-art performance on 3 out of 4 metrics, i.e., BLEU, METEOR, and CIDEr.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, we also compare the TextKG method to the SOTA methods focusing on multi-modality pretraining models in Table 5. As shown in Table 5, our TextKG method performs favorably against existing methods, such as DECEM [48], UniVL [29] and MV-GPT [40], that are pretrained on the large-scale datasets (e.g., Howto100M [31]), i.e., improve 0.7 absolute CIDEr score compared to the SOTA method LAVENDER [23].", + "bbox": [ + 496, + 688, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MSVD dataset. We also evaluate the proposed TextKG method on the MSVD dataset [3] that doesn't have speech and repoprt the results in Table 6. As shown in Table 6, our TextKG achieves the best results without speech transcripts by improving $1.2\\%$ and $3.8\\%$ CIDEr scores compared to the SOTA method HMN [58] and the JCRR method [14].", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18946", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/93b03b25980df951812cd658f0a66a306d241a1fe52e421370f314d396406a4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBMRC
OA-BTG [61]41.428.2-46.9
POS-CG [51]4228.261.648.7
MGSA [4]42.427.6-47.5
STG-KD [32]40.528.360.947.1
ORG-TRL [64]43.628.862.150.9
SGN [38]40.828.360.849.5
MGRMP [5]41.728.962.151.4
HMN [58]43.52962.751.5
TextKG43.729.662.452.4
", + "bbox": [ + 84, + 88, + 460, + 237 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ea450eb079dc50f1e8e6786efe961d5517f5c689a21714664bac70adb6bc5f9.jpg", + "table_caption": [ + "Table 4. Evaluation results on the MSR-VTT test subset in the micro-level evaluation mode." + ], + "table_footnote": [], + "table_body": "
MethodInputFeaturesBMRC
SWINBERT [27]VVidSwin45.430.664.155.9
CLIP4C [47]VCLIP46.130.763.757.7
CMVC [57]VCLIP48.231.364.858.7
LAVENDER [23]VVidSwin---60.1
DeCEM [48]V+SBERT45.229.764.752.3
UniVL [29]V+SS3D41.828.960.850.0
MV-GPT [40]V+SViViT48.9*38.7*64.060.0
TextKG (CLIP)V+SCLIP46.630.564.860.8
", + "bbox": [ + 84, + 282, + 460, + 417 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61c12f5bc46d00b274e3f2b5aae6009b132b41f57c917518c7b90abe8750a3c0.jpg", + "table_caption": [ + "Table 5. Comparison to the methods focusing on model pretraining on the MSR-VTT dataset in the micro-level evaluation mode. Notably, following CLIP4C, we use the pre-trained CLIP model on LAION-400M [39] to extract video appearance features. * The authors use a different library to compute BLEU and METOR. Thus, the results on BLEU and METOR are not directly comparable to other methods." + ], + "table_footnote": [], + "table_body": "
MethodBMRC
OA-BTG [61]56.936.2-90.6
POS-CG [51]52.534.171.388.7
MGSA [4]53.435-86.7
STG-KD [32]52.236.973.993.0
ORG-TRL [64]54.336.473.995.2
SGN [38]52.835.572.994.3
MGRMP [5]55.836.974.598.5
JCRR [14]57.036.8-96.8
HMN [58]59.237.775.1104.0
TextKG60.838.575.1105.2
", + "bbox": [ + 81, + 542, + 465, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exploiting knowledge graph to model the relations between objects, demonstrating the effectiveness of the knowledge graph usage in our method for video captioning.", + "bbox": [ + 76, + 772, + 467, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 830, + 225, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of different modules. To validate the effectiveness of different modules in TextKG, we conduct several experiments on the YouCookII dataset in Table 7. As", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b5d7130fab0d7a9e13da8534f77b15307d9353059c7e4e1b504749b89e62bd6c.jpg", + "table_caption": [ + "Table 6. Evaluation results on the MSVD test subset in the microlevel evaluation mode." + ], + "table_footnote": [], + "table_body": "
V-FR-FTextG-KGS-KGK-SBMCRep
✓7.415.732.14.1
✓✓9.517.745.95.2
✓✓✓✓9.717.848.94.2
✓✓✓✓9.718.048.53.3
✓✓✓✓✓9.617.749.85.5
✓✓13.021.262.52.7
✓✓✓13.922.171.32.9
✓✓✓✓✓13.722.073.52.0
✓✓✓✓✓13.521.974.82.1
✓✓✓✓✓13.721.872.02.8
✓✓✓✓✓✓14.022.175.92.8
", + "bbox": [ + 504, + 88, + 887, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Ablation study on the YouCookII val subset in the paragraph-level evaluation mode. \"V-F\" and \"R-F\" indicate the video and region features, \"Text\" indicates the speech transcripts features, \"G-KG\" and \"S-KG\" indicate general and specific knowledge graph, and \"K-S\" indicates the knowledge selection mechanism.", + "bbox": [ + 496, + 268, + 890, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shown in Table 7, the baseline method considers the information of video features and predicted captions, achieving 32.1 CIDEr score. After integrating the speech transcripts, the CIDEr score is improved to 62.5, which ensures the key information of videos is considered in generating video captions. After integrating the region features, TextKG achieves 71.3 CIDEr score. The CIDEr score can be further improved 4.6 CIDEr score to 75.9 by exploiting additional knowledge in knowledge graph. Notably, even without the information of speech transcripts, the knowledge graphs are still crucial for video captioning, i.e., improving 3.9 CIDEr score (49.8 vs. 45.9). Meanwhile, we also demonstrate the effectiveness of knowledge selection mechanism in Table 7. As shown in Table 7, without the knowledge selection mechanism, the CIDEr score significantly drops 3.9 to 72.0. We believe that the noisy knowledge is harmful to the accuracy of video captions.", + "bbox": [ + 496, + 364, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of transcript quality. To explore the effect of transcript quality on the knowledge graph, we add some noises to the speech transcripts (i.e., mask some amount of speech transcripts) and evaluate the performance of our models on the YouCookII dataset in Figure 3. We find that the method using both the general and specific knowledge graphs performs favorable against other methods with different degrees of masked speech transcripts. Meanwhile, as the amount of masked speech transcripts increasing, the accuracy of the method using specific knowledge graph drops sharply, while the accuracy of the method using general knowledge graph drops at a much slower pace. The aforementioned results demonstrate the importance of general knowledge graph for the video captioning task.", + "bbox": [ + 496, + 622, + 890, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 845, + 681, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the qualitative results of the proposed method on the YouCookII dataset in Figure 2, including the key", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae5f712b8f6985f028f0661f11d732f3f0808acc8f826d1b89a9aa421a3dd5b0.jpg", + "image_caption": [ + "Figure 2. Qualitative results of the proposed method on the YouCookII dataset. The information in the speech transcripts and knowledge graph is not considered in the baseline." + ], + "image_footnote": [], + "bbox": [ + 120, + 90, + 843, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e507a2235cb0683d733417660532c4da6f9607b49896c50c532d291e2a7622e.jpg", + "image_caption": [ + "Figure 3. Influence of the transcripts qualities. \"w/o KG\" indicates that the method does not use the general and specific knowledge graphs. \"with G-KG\" and \"with S-KG\" indicate that only the general and specific knowledge graphs are used, respectively. \"with G- and S-KG\" indicates that both the general and specific knowledge graphs are used. The CIDEr differences between the \"with G-KG\", \"with S-KG\", \"with G- and S-KG\" methods and the \"w/o KG\" method are shown at the bottom corner." + ], + "image_footnote": [], + "bbox": [ + 91, + 450, + 455, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frames, detected salient objects, speech transcripts, retrieved knowledge items, caption results predicted by our model, ground-truth captions, and key words distributions in the training set. As shown in Figure 2, we observe that the transcript contains key information in the video, e.g., egg, salt, pepper, com starch and flour, to ensure the qual", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity of the generated video captions. However, the model with only the speech transcripts fails to predict the long-tail phrases, such as sesame oil and black pepper. With the help of knowledge graph providing the relations among these phrases, e.g., sesame oil, black pepper and corn starch, our TextKG model can predict the phrases more accurately. This indicates that the knowledge graph is crucial to mitigate the long-tail word challenges in video captioning.", + "bbox": [ + 496, + 446, + 890, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 585, + 617, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a text with knowledge graph augmented transformer for video captioning, which aims to integrate external knowledge in knowledge graph and exploit the multi-modality information in video to mitigate long-tail words challenge. Extensive experiments conducted on four challenging datasets demonstrate the effectiveness of the proposed method.", + "bbox": [ + 496, + 613, + 890, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we plan to improve the proposed method in two directions, i.e., (1) optimizing the knowledge retrieve strategy by considering semantic context information of detected objects and corresponding actions in videos; (2) constructing multi-modal knowledge graph (e.g., the nodes in knowledge graph formed by text, speech, and images or videos) to improve the accuracy of video captioning.", + "bbox": [ + 496, + 720, + 892, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 845, + 596, + 861 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for", + "bbox": [ + 506, + 872, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18948", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1708-1718, 2021. 2", + "[2] William Chan, Navdeep Jaitly, Quoc V. Le, and Oriol Vinyals. Listen, attend and spell. 2015. 2, 5", + "[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Dekang Lin, Yuji Matsumoto, and Rada Mihalcea, editors, Annual Meeting of the Association for Computational Linguistics, 2011. 2, 4, 5, 6", + "[4] Shaoxiang Chen and Yu-Gang Jiang. Motion guided spatial attention for video captioning. In Association for the Advancement of Artificial Intelligence, pages 8191-8198, 2019. 7", + "[5] Shaoxiang Chen and Yu-Gang Jiang. Motion guided region message passing for video captioning. In IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 5, 6, 7", + "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. CoRR, 2019. 2", + "[7] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc Viet Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics, pages 2978-2988, 2019. 6", + "[8] Michael J. Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 376-380, 2014. 5", + "[9] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. COOT: cooperative hierarchical transformer for video-text representation learning. In Advances in Neural Information Processing Systems, 2020. 5, 6", + "[10] Xin Gu, Hanhua Ye, Guang Chen, Yufei Wang, Libo Zhang, and Longyin Wen. Dual-stream transformer for generic event boundary captioning. 2022. 2", + "[11] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 4", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4", + "[13] Jack Hessel, Bo Pang, Zhenhai Zhu, and Radu Soricut. A case study on combining ASR and visual features for generating instructional video captions. In Conference on Computational Natural Language Learning, pages 419-429, 2019. 6", + "[14] Jingyi Hou, Xinxiao Wu, Xiaoxun Zhang, Yayun Qi, Yunde Jia, and Jiebo Luo. Joint commonsense and relation reasoning for image and video captioning. In Association for the Advancement of Artificial Intelligence, pages 10973-10980, 2020. 1, 2, 6, 7", + "[15] Feicheng Huang, Zhixin Li, Shengjia Chen, Canlong Zhang, and Huifang Ma. Image captioning with internal and external" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "knowledge. In CIKM '20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020, pages 535-544. ACM, 2020. 2", + "[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, International Conference on Machine Learning, pages 448-456, 2015. 4", + "[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, International Conference on Learning Representations, 2015. 4, 5", + "[18] Bruno Korbar, Fabio Petroni, Rohit Girdhar, and Lorenzo Torresani. Video understanding as machine translation. CoRR, abs/2006.07203, 2020. 6", + "[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In IEEE/CVF International Conference on Computer Vision, pages 706–715, 2017. 1, 2, 4, 5, 6", + "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 5", + "[21] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 1, 2, 3, 4, 5, 6", + "[22] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 6", + "[23] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. LAVENDER: unifying video-language understanding as masked language modeling. 2022. 6, 7", + "[24] Linjie Li, Jie Lei, Zhe Gan, Licheng Yu, Yen-Chun Chen, Rohit Pillai, Yu Cheng, Luowei Zhou, Xin Wang, William Yang Wang, Tamara L. Berg, Mohit Bansal, Jingjing Liu, Lijuan Wang, and Zicheng Liu. VALUE: A multitask benchmark for video-and-language understanding evaluation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 6", + "[25] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: towards unified-modal understanding and generation via cross-modal contrastive learning. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors, ACL-IJCNLP, pages 2592-2607, 2021. 2", + "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, pages 74-81, 2004. 5" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18949", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17928-17937, 2022. 6, 7", + "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 13–23, 2019. 2", + "[29] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Xilin Chen, and Ming Zhou. Univilm: A unified video and language pre-training model for multimodal understanding and generation. CoRR, abs/2002.06353, 2020. 1, 2, 6, 7", + "[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1", + "[31] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2, 6", + "[32] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10876, 2020. 7", + "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5", + "[34] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing, pages 1532-1543, 2014. 4, 5", + "[35] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In IEEE/CVF International Conference on Computer Vision, pages 8907-8916, 2019. 2", + "[36] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 4", + "[37] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 1137-1149, 2017. 4", + "[38] Hobin Ryu, Sunghun Kang, Haeyong Kang, and Chang D. Yoo. Semantic grouping network for video captioning. In Association for the Advancement of Artificial Intelligence, pages 2514-2522, 2021. 1, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, 2021. 7", + "[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17938-17947, 2022. 1, 2, 6, 7", + "[41] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In Annual Meeting of the Association for Computational Linguistics, pages 6382-6391, 2019. 1, 6", + "[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Association for the Advancement of Artificial Intelligence, pages 4444-4451, 2017. 5", + "[43] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7463-7472, 2019. 2, 6", + "[44] Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuanjing Huang, and Zheng Zhang. Colake: Contextualized language and knowledge embedding. In International Conference on Computational Linguistics, pages 3660-3670, 2020. 2", + "[45] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Satinder Singh and Shaul Markovitch, editors, Association for the Advancement of Artificial Intelligence, 2017. 4", + "[46] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 5099-5110, 2019. 2", + "[47] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: CLIP for video caption. In ACM International Conference on Multimedia, pages 4858-4862, 2021. 7", + "[48] Zineng Tang, Jie Lei, and Mohit Bansal. December: Learning from noisy instructional videos via dense captions and entropy minimization. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 2415-2426, 2021. 1, 6, 7", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 3", + "[50] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 5" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18950", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Bairui Wang, Lin Ma, Wei Zhang, Wenhao Jiang, Jingwen Wang, and Wei Liu. Controllable video captioning with POS sequence guidance based on gated fusion network. In IEEE/CVF International Conference on Computer Vision, pages 2641-2650, 2019. 7", + "[52] Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. KEPLER: A unified model for knowledge embedding and pre-trained language representation. Transactions of the Association for Computational Linguistics, 9:176-194, 2021. 2", + "[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning for video understanding. CoRR, abs/1712.04851, 2017. 6", + "[54] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In European Conference on Computer Vision, pages 468-483, 2018. 5", + "[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. VLM: task-agnostic video-language model pre-training for video understanding. In Findings of the Association for Computational Linguistics, pages 4227–4239, 2021. 2", + "[56] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 2, 4, 5", + "[57] Bang Yang, Tong Zhang, and Yuexian Zou. CLIP meets video captioning: Concept-aware representation learning does matter. In Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Shenzhen, China, November 4-7, 2022, Proceedings, Part I, 2022. 7", + "[58] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. pages 17939-17948, 2022. 1, 4, 5, 6, 7", + "[59] Donghan Yu, Chenguang Zhu, Yiming Yang, and Michael Zeng. JAKET: joint pre-training of knowledge graph and language understanding. In Association for the Advancement of Artificial Intelligence, pages 11630-11638, 2022. 2", + "[60] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In European Conference on Computer Vision, pages 385-401, 2018. 6", + "[61] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8327-8336, 2019. 7", + "[62] Yu Zhang, Xinyu Shi, Siya Mi, and Xu Yang. Image captioning with transformer and knowledge graph. Pattern Recognition Letters, 143:43-49, 2021. 2", + "[63] Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. ERNIE: enhanced language representation with informative entities. In Annual Meeting of the Association for Computational Linguistics, pages 1441-1451, 2019. 2", + "[64] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph with teacher-recommended learning for video captioning." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13275-13285, 2020. 2, 4, 5, 6, 7", + "[65] Wentian Zhao, Yao Hu, Heda Wang, Xinxiao Wu, and Jiebo Luo. Boosting entity-aware image captioning with multimodal knowledge graph. CoRR, 2021. 2", + "[66] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13093-13102, 2020. 4, 5, 6", + "[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J. Corso, and Marcus Rohrbach. Grounded video description. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6578-6587, 2019. 6", + "[68] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pre-training for image captioning and VQA. In Association for the Advancement of Artificial Intelligence, pages 13041-13049, 2020. 2", + "[69] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence, pages 7590-7598, 2018. 1, 2, 4, 5", + "[70] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8739-8748, 2018. 6", + "[71] Yimin Zhou, Yiwei Sun, and Vasant G. Honavar. Improving image captioning by leveraging knowledge graphs. In IEEE Winter Conference on Applications of Computer Vision, pages 283-293, 2019. 1, 2", + "[72] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8743-8752, 2020. 1, 2, 6" + ], + "bbox": [ + 503, + 92, + 890, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18951", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_model.json b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7138e0b01c6954d0c9b492c29f0c5fb23aec7b64 --- /dev/null +++ b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.868, + 0.155 + ], + "angle": 0, + "content": "Text with Knowledge Graph Augmented Transformer for Video Captioning" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.179, + 0.817, + 0.233 + ], + "angle": 0, + "content": "Xin Gu\\(^{1,3}\\), Guang Chen\\(^{2}\\), Yufei Wang\\(^{2}\\), Libo Zhang\\(^{3,1*}\\), Tiejian Luo\\(^{1}\\), Longyin Wen\\(^{2}\\) \n\\(^{1}\\)University of Chinese Academy of Sciences, Beijing, China \n\\(^{2}\\)ByteDance Inc., San Jose, USA" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.234, + 0.756, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Institute of Software Chinese Academy of Sciences, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.255, + 0.828, + 0.269 + ], + "angle": 0, + "content": "guxin21@mails.ucas.edu.cn, guang.chen@bytedance.com, yufei.wang@bytedance.com" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.273, + 0.756, + 0.287 + ], + "angle": 0, + "content": "libo@iscas.ac.cn, tjluo@ucas.ac.cn, longyin.wen@bytedance.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Video captioning aims to describe the content of videos using natural language. Although significant progress has been made, there is still much room to improve the performance for real-world applications, mainly due to the long-tail words challenge. In this paper, we propose a text with knowledge graph augmented transformer (TextKG) for video captioning. Notably, TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is designed to absorb additional knowledge, which models the interactions between the additional knowledge, e.g., pre-built knowledge graph, and the built-in information of videos, e.g., the salient object regions, speech transcripts, and video captions, to mitigate the long-tail words challenge. Meanwhile, the internal stream is designed to exploit the multi-modality information in videos (e.g., the appearance of video frames, speech transcripts, and video captions) to ensure the quality of caption results. In addition, the cross attention mechanism is also used in between the two streams for sharing information. In this way, the two streams can help each other for more accurate results. Extensive experiments conducted on four challenging video captioning datasets, i.e., YouCookII, ActivityNet Captions, MSR-VTT, and MSVD, demonstrate that the proposed method performs favorably against the state-of-the-art methods. Specifically, the proposed TextKG method outperforms the best published results by improving \\(18.7\\%\\) absolute CIDEr scores on the YouCookII dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.794, + 0.21, + 0.81 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.821, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Video captioning aims to generate a complete and natural sentence to describe video content, which attracts much" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.324, + 0.895, + 0.459 + ], + "angle": 0, + "content": "attention in recent years. Generally, most existing methods [21,38,41,58] require a large amount of paired video and description data for model training. Several datasets, such as YouCookII [69], and ActivityNet Captions [19] are constructed to promote the development of video captioning field. Meanwhile, some methods [29, 40, 48, 72] also use the large-scale narrated video dataset HowTo100M [30] to pretrain the captioning model to further improve the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.657 + ], + "angle": 0, + "content": "Although significant progress has been witnessed, it is still a challenge for video captioning methods to be applied in real applications, mainly due to the long-tail issues of words. Most existing methods [29, 40, 48, 72] attempt to design powerful neural networks, trained on the large-scale video-text datasets to make the network learn the relations between video appearances and descriptions. However, it is pretty tough for the networks to accurately predict the objects, properties, or behaviors that are infrequently or never appearing in training data. Some methods [14, 71] attempt to use knowledge graph to exploit the relations between objects for long-tail challenge in image or video captioning, which produces promising results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.896, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present a text with knowledge graph augmented transformer (TextKG), which integrates additional knowledge in knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge. TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is used to absorb additional knowledge to help mitigate long-tail words challenge by modeling the interactions between the additional knowledge in pre-built knowledge graph, and the built-in information of videos, such as the salient object regions in each frame, speech transcripts, and video captions. Specifically, the information is first retrieved from the pre-built knowledge graphs based on the detected salient objects. After that, we combine the features of the retrieved information, the appearance features of detected salient objects, the features of speech transcripts and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author:libo@iscas.ac.cn, Libo Zhang was supported by Youth Innovation Promotion Association, CAS (2020111). This work was done during internship at ByteDance Inc." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18941" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.242 + ], + "angle": 0, + "content": "captions, then feed them into the external stream of TextKG to model the interactions. The internal stream is designed to exploit the multi-modality information in videos, such as the appearance of video frames, speech transcripts and video captions, which can ensure the quality of caption results. To share information between two streams, the cross attention mechanism is introduced. In this way, the two streams can obtain the required modal information from each other for generating more accurate results. The architecture of the proposed method is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.363 + ], + "angle": 0, + "content": "Several experiments conducted on four challenging datasets, i.e., YouCookII [69], ActivityNet Captions [19], MSR-VTT [56], and MSVD [3] demonstrate that the proposed method performs favorably against the state-of-the-art methods. Notably, our TextKG method outperforms the best published results by improving \\(18.7\\%\\) and \\(3.2\\%\\) absolute CIDEr scores in the paragraph-level evaluation mode on the YouCookII and Activity-Net Captions datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.377, + 0.218, + 0.392 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.468, + 0.658 + ], + "angle": 0, + "content": "Video captioning attracts much attention of researchers in recent years. The best practice has been achieved by attention-based methods, which attempts to associate visual components with sentences in videos. Some of them focus on designing powerful network architectures. VLM [55] and VideoBERT [43] take the visual and text modalities as input, and use a shared transformer to construct a task-agnostic video-language model. ViLBERT [28] processes visual and linguistic information separately with two parallel streams, and then use the attention mechanism to model the interactions between visual and language features. Instead of using the separate encoder-decoder architecture, MART [21] designs a shared encoder-decoder network and augments it with the memory module. ActBert [72] uses local regional features to learn better visual-language alignment. WLT [35] takes audio features as an additional input, and uses context fusion to generate multimodal features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.468, + 0.855 + ], + "angle": 0, + "content": "Meanwhile, some methods [10,14,15,62,65,71] focus on exploiting prior knowledge to provide semantic correlations and constraints between objects for image or video captioning, producing promising results. ORG-TRL [64] uses the knowledge information in the language model (BERT) to provide candidate words for video captioning. In contrast, we propose a two-stream transformer for video captioning, with the internal stream used to exploit multi-modality information in videos, and the external stream used to model the interactions between the additional knowledge and the built-in information of videos. These two streams use the cross-attention mechanism to share information in different modalities for generating more accurate results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.857, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Vision-and-language representation learning is a hot topic in recent years. ViLBERT [28], LXMERT [46], UNITER [6], UNIMO [25] and Unified-VL [68] learn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "the representations between image and text, while Univl [29], VideoBERT [43], ActBERT [72] and MV-GPT [40] learn the representations between videos and transcripts. Notably, most of these methods attempt to learn powerful vision-and-language representations by pre-training the models on the large-scale datasets, e.g., Howto100M [31] and WebVid-2M [1], and then finetune them on downstream tasks such as video captioning, video-text retrieval and visual question answering. In contrast, our TextKG method uses the speech transcripts as the text to model the visual and linguistic representations and integrate the additional knowledge in knowledge graph to mitigate long-tail words challenge in video captioning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Knowledge graph in NLP. Knowledge graph is an useful tool to indicate the real-world entities and their relations, which provides rich structured knowledge facts for language modeling. Large-scale knowledge graphs are used to train knowledge enhanced language models for various natural language processing (NLP) tasks. CoLAKE [44] proposes to inject the knowledge context of an entity, and to jointly learn the contextualized representation for both language and knowledge by a unified structure. ERNIE [63] enhances BERT architecture to better integrate the knowledge information and textual information. KEPLER [52] not only improves language models by integrating factual knowledge, but also generates text-enhanced knowledge representation. JAKET [59] proposes a joint pre-training framework to model knowledge graph and language simultaneously. Inspired by CoLAKE, our method jointly learns the representations of vision, language and knowledge, and enhances the joint visual-language representations by retrieving relevant knowledge in knowledge graphs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.649, + 0.606 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As mentioned above, our method aims to integrate additional knowledge using knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge in the field of video captioning. We design a text with knowledge graph augmented transformer, i.e., a two-stream transformer formed by the external and internal streams. Both streams are constructed by \\( N \\) sets of alternately stacked self-attention and cross-attention blocks. Specifically, we first use a detector to generate the salient object regions in each video frame, and use the automatic speech recognition (ASR) method [2] to generate the speech transcripts in videos. After that, the class labels of detected salient objects are used to retrieve the prior knowledge in knowledge graphs. Then, the appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior knowledge, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent words of the predicted captions. The overall architecture of the proposed TextKG is shown in Figure 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.873, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.405, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 1. The architecture of our TextKG method, which is formed the external and internal streams. Each stream is stacked with \\(N\\) sets of multi-modality self-attention and cross-attention modules. The cross attention modules are designed to align tokens in different modalities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.31, + 0.463 + ], + "angle": 0, + "content": "3.1. Two-Stream Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.532 + ], + "angle": 0, + "content": "As described above, our two-stream transformer is formed by the external stream and internal stream. Several self-attention and cross-attention blocks are interleaved to construct the two streams." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.471, + 0.685 + ], + "angle": 0, + "content": "Let \\(\\{y_{1}, y_{2}, \\dots, y_{l}\\}\\) be the predicted captions, where \\(y_{i}\\) is the index of the \\(i\\)-th word in the dictionary, \\(\\{p_{1}, p_{2}, \\dots, p_{l}\\}\\) to be the predicted probabilities of words in the dictionary, where \\(p_{i}\\) is the probability of the \\(i\\)-th word. The index of the \\(i\\)-th word \\(y_{i}\\) is computed as \\(y_{i} = \\arg \\max_{i} p_{i}\\). Meanwhile, let \\(z_{i}^{\\mathrm{ext}}\\) and \\(z_{i}^{\\mathrm{int}}\\) to be the output probabilities of the external and internal streams of the \\(i\\)-th word. Thus, we have \\(p_{i} = \\omega_{1} z_{i}^{\\mathrm{ext}} + \\omega_{2} z_{i}^{\\mathrm{int}}\\), where \\(\\omega_{1}\\) and \\(\\omega_{2}\\) are the hyperparameters used to balance the outputs of the two-streams. In this way, \\(y_{i}\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.693, + 0.47, + 0.712 + ], + "angle": 0, + "content": "\\[\ny _ {i} = \\arg \\max _ {i} \\left(\\omega_ {1} z _ {i} ^ {\\text {e x t}} + \\omega_ {2} z _ {i} ^ {\\text {i n t}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.748 + ], + "angle": 0, + "content": "In the following sections, we will describe each module in our two-stream transformer in more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Multi-modality Self-Attention Module. As shown in Figure 1, we use the self-attention module [49] in both the external and internal streams to model the interactions among multi-modality information. Specifically, for the external stream, the concatenation of feature embeddings of detected objects \\(\\mathcal{F}_{\\mathrm{r}}\\), retrieved prior knowledge \\(\\mathcal{F}_{\\mathrm{k}}\\), speech transcripts \\(\\mathcal{F}_{\\mathrm{s}}\\), and predicted video captions \\(\\mathcal{F}_{\\mathrm{c}}\\) are fed into the self-attention module to model the interactions, i.e., \\(\\mathcal{X}_{\\mathrm{ext}} = \\mathrm{Concat}(\\mathcal{F}_{\\mathrm{r}}, \\mathcal{F}_{\\mathrm{k}}, \\mathcal{F}_{\\mathrm{s}}, \\mathcal{F}_{\\mathrm{c}})\\). Meanwhile, for the internal stream, the concatenation of feature embeddings of speech" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.449, + 0.892, + 0.51 + ], + "angle": 0, + "content": "transcripts \\(\\mathcal{F}_{\\mathrm{s}}\\) , predicted video captions \\(\\mathcal{F}_{\\mathrm{c}}\\) , and video frame \\(\\mathcal{F}_{\\mathrm{v}}\\) are fed into the self-attention model to exploit the interactions, i.e., \\(\\mathcal{X}_{\\mathrm{int}} = \\operatorname {Concat}(\\mathcal{F}_{\\mathrm{s}},\\mathcal{F}_{\\mathrm{c}},\\mathcal{F}_{\\mathrm{v}})\\) . The self-attention module computes the interactions as follows," + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.52, + 0.892, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\Phi (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} + M\\right) V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.629 + ], + "angle": 0, + "content": "where \\(d\\) is the feature dimension of queries \\(Q\\) and keys \\(K\\). We have \\(Q = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{Q}}\\), \\(K = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{K}}\\), \\(V = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{V}}\\), where \\(W^{\\mathrm{Q}}\\), \\(W^{\\mathrm{K}}\\), and \\(W^{\\mathrm{V}}\\) are learnable parameters, and \\(\\mathcal{X}_{(\\cdot)} \\in \\{\\mathcal{X}_{\\mathrm{ext}}, \\mathcal{X}_{\\mathrm{int}}\\}\\) are the concatenated feature embeddings for the external and internal streams, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Following [21], we introduce a mask matrix \\( M \\) in attention function (2) of both the external and internal streams to prevent the model from seeing future words. That is, for the \\( i \\)-th caption token, we set \\( M_{i,j} = 0 \\) for \\( j = 1,\\dots ,i \\), and set \\( M_{i,j} = -\\inf \\) for \\( j > i \\). Meanwhile, for the external stream, we set \\( M_{i,j} = -\\inf \\) to prevent the associations between irrelevant retrieved prior knowledge and the detected salient object regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Multi-modality Cross-Attention Module. Besides self-attention module, we use the cross-attention module to align the interactions modeled by both the external and internal streams. Specifically, we compute the affinity matrix to guide the alignments between the modeled interactions by injecting the information. Similar to the self-attention module, we also introduce the mask matrix into the attention function to compute the cross-attention. Intuitively, the retrieved prior knowledge should not have direct effect on the predicted captions, but through the detected salient objects." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Thus, we set the corresponding elements to \\(-\\inf\\) to prevent the retrieved prior knowledge directly influencing the predicted captions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.145, + 0.218, + 0.162 + ], + "angle": 0, + "content": "3.2. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.468, + 0.259 + ], + "angle": 0, + "content": "The cross entropy loss function is used to guide the training of our TextKG method. Given the ground-truth indices of previous \\((i - 1)\\) words and the concatenated of two-streams \\(\\mathcal{X}_{\\mathrm{ext}}\\) and \\(\\mathcal{X}_{\\mathrm{int}}\\), we can get the predictions of the current \\(i\\)-th word. After that, the training loss of our method is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.267, + 0.469, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\sum_ {i = 1} ^ {l} \\left(\\lambda_ {1} \\log z _ {i} ^ {\\text {e x t}} + \\lambda_ {2} \\log z _ {i} ^ {\\text {i n t}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.448 + ], + "angle": 0, + "content": "where \\( z_{i}^{\\mathrm{ext}} \\) and \\( z_{i}^{\\mathrm{int}} \\) are the \\( i \\)-th output word of the external and internal streams, \\( \\lambda_{1} \\) and \\( \\lambda_{2} \\) are the preset parameters used to balance the two streams, and \\( l \\) is the total length of predicted captions. The Adam optimization algorithm [17] with an initial learning rate of \\( 1e - 4 \\), \\( \\beta_{1} = 0.9 \\), \\( \\beta_{2} = 0.999 \\) and \\( L2 \\) weight decay of 0.01 is used to train the model. The Warmup strategy is adopted in training phase by linearly increasing the learning rate from 0 to the initial learning rate for the first \\( 10\\% \\) of training epochs, and linearly decreasing to 0 for the remaining \\( 90\\% \\) of epochs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.265, + 0.47 + ], + "angle": 0, + "content": "3.3. Multimodal Tokens" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.569 + ], + "angle": 0, + "content": "The appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent captions after tokenization. We will describe the feature extraction and tokenization processes of the aforementioned embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.75 + ], + "angle": 0, + "content": "Video tokens encode the appearance and motion information of video frames. Specifically, the aligned appearance and motion features form the video tokens. In each second of the video, we sample 2 frames to extract features. Following [21], we use the \"Flatten-673\" layer in ResNet200 [12] to extract appearance features, and the \"global pool\" layer in BN-Inception [16] to extract the motion features on the YouCookII [69] and ActivityNet Captions [19] datasets. Meanwhile, for fair comparison, we use the InceptionResNetV2 [45] and C3D [11] models to extract appearance and motion features on the MSR-VTT [56] and MSVD [3] datasets, following the methods [58, 64, 66]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Region tokens are used to describe the visual information of the detected salient objects. We use the Faster R-CNN method [37] to extract visual features for detected object regions, which is pretrained on the Visual Genome dataset [20], and the \\(N_{\\mathrm{r}}\\) detected objects with highest confidence score in each frame are reserved for video captioning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Transcript tokens are used to encode the content of speech transcripts, which contains some key information in videos. The GloVe model [34] is used to extract features of each word in transcripts to form the transcript tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Caption tokens encode the information of predicted captions. Similar to transcript tokens, GloVe [34] is adopted to extract features of each word in the generated captions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Knowledge tokens model the content of retrieved knowledge, which is important to mitigate the long-tail words challenge in video captioning. Our TextKG method uses a triple structure to encode the retrieved knowledge, formed by two object items and the relations between them, i.e., \\((\\alpha_{\\mathrm{head}},\\alpha_{\\mathrm{tail}},\\rho_{\\mathrm{rel}})\\), where \\(\\alpha_{\\mathrm{head}}\\) and \\(\\alpha_{\\mathrm{tail}}\\) are the two nodes in knowledge graph, and \\(\\rho_{\\mathrm{rel}}\\) indicates the edge encoding the relations between them. For example, \\(\\alpha_{\\mathrm{head}} = \"\\mathrm{knife}\\), \\(\\alpha_{\\mathrm{tail}} = \"\\mathrm{hard}\\), and \\(\\rho_{\\mathrm{rel}} = \"\\mathrm{has~property}\\). For each detected object \\(\\alpha_{\\mathrm{head}}\\), we aim to integrate prior context information for more accurate caption results. That is, we use the GloVe model to extract linguistic features of the node \\(\\alpha_{\\mathrm{tail}}\\) and set the learnable embeddings for the edge \\(\\rho_{\\mathrm{rel}}\\). After that, we compute the summation of them to get the knowledge embeddings of the knowledge tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.366, + 0.892, + 0.653 + ], + "angle": 0, + "content": "For each detected object, we use its category name to retrieve knowledge (i.e., triples containing the category name of the detected object) from the knowledge graph. Notably, only a portion of the retrieved knowledge are highly related to the video. The rests are potentially harmful to the performance of caption generation as it may distract the training and inference and also increase computational cost. For example, to generate caption of a cooking tutorial video, the knowledge \"knife - used to -cut\" is more useful rather than \"knife - has property - hard\". Because cutting food may be an important procedure for food preparation. Thus, we design a primary rank mechanism to sort the retrieved knowledge based on their semantic similarity to the video. The cosine similarity between the embeddings of knowledge tokens and video transcripts is computed as the semantic similarity for ranking. The pretrained SBERT [36] model is used to extract feature embeddings of transcripts. Finally, \\( N_{\\mathrm{k}} \\) knowledge items with the highest similarity scores are reserved for video captioning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.791, + 0.681 + ], + "angle": 0, + "content": "3.4. Knowledge Graphs Construction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.892, + 0.84 + ], + "angle": 0, + "content": "As mentioned above, we use the knowledge graph to retain the prior knowledge. The nodes in the graph could correspond to a noun, an adjective, a verb or an adverb, etc. The edges in the graph are used to encode the relations between different nodes. Apparently, it is important to include all key information of videos in knowledge graphs for accurate caption results. Thus, besides general knowledge graph covering most key knowledge in general scenarios, we also introduce the specific knowledge graph for each dataset covering the key knowledge in specific scenarios." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "General knowledge graph is designed to include most key information in general scenarios that we are interested, such as cooking and activity. Specifically, the general knowledge graph is exported from the public available giant knowledge" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "graph ConceptNet [42] by extracting key words1 in ConceptNet with the connected edges and neighboring nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.318 + ], + "angle": 0, + "content": "Specific knowledge graph. Besides the general knowledge graph, we also construct the specific knowledge graph to cover key information in specific scenarios. We believe that the speech transcripts of videos contain most of the crucial information and use the speech transcripts as the source to construct the specific knowledge graph. We first use the automatic speech recognition (ASR) model [2] to convert the speech in all videos to transcripts. After that, we use the Stanford NLP library² to analyze the components in each sentence in transcripts to form a structured grammar tree. We collect the \"adjective and noun\", \"noun and noun\" and \"adverb and verb\" phrases from the grammar trees to construct the specific knowledge graph." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.333, + 0.209, + 0.349 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.358, + 0.179, + 0.372 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.468, + 0.456 + ], + "angle": 0, + "content": "YoucookII contains 2,000 long untrimmed videos from 89 cooking recipes, including 1,333 videos for training, 457 videos for testing. The average length of videos is 5.3 minutes and each video is divided into 7.7 video clips on average with the manually labeled caption sentences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.468, + 0.592 + ], + "angle": 0, + "content": "ActivityNet Captions is a large-scale challenging dataset for video captioning, with 10,024 videos for training, 4,926 for validation, and 5,044 for testing. The average length of each video is 2.1 minutes and is divided into 3.65 clips on average. Each clip is assigned with a manually labeled caption sentence. Since the testing set is not publicly available, following MART [21], we split the validation set into two subsets, i.e., ae-val with 2,460 videos for validation and aetest with 2,457 videos for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.468, + 0.668 + ], + "angle": 0, + "content": "MSR-VTT includes 10,000 video clips with 15 seconds on average. Each video clip contains 20 human annotated captions and the average length of each sentence is 9.28 words. Following [5,38,58], 6,513,497, and 2,990 videos are used for training, validation, and testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.468, + 0.744 + ], + "angle": 0, + "content": "MSVD consists of 1970 video clips collected from YouTube and each clip is approximately 10 seconds long. Each clip is labeled with 40 sentences. Following [5,38,58], we split the dataset into 1,200,100,and 670 video clips for training, validation and testing, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.259, + 0.769 + ], + "angle": 0, + "content": "4.2. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.468, + 0.853 + ], + "angle": 0, + "content": "To compare the performance of the proposed method with other methods, five model-free automatic evaluation metrics are used, including BLEU@4 [33] (abbreviated to B) that are precision-based metric, METEOR [8] (abbreviated to M) that computes sentence-level scores, CIDEr [50]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.302 + ], + "angle": 0, + "content": "(abbreviated to C) that is consensus-based metric, ROUGE [26] (abbreviated to R) that uses longest common subsequence to compute the similarities between sentences, and Rep@4 [54] (abbreviated to Rep) that computes the redundancy score for every captions. Lower Rep indicates better model. The CIDEr score is used as the primary metric in evaluation. Notably, both micro-level and paragraph-level evaluation modes are used. The micro-level mode evaluates the predicted captions of each video clip independently, while the paragraph-level mode considers the relations between captions of different clips. Following [9, 21], the paragraph-level evaluation mode is used in the YouCookII and ActivityNet datasets, which contains multiple clips in each video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.716, + 0.332 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.537 + ], + "angle": 0, + "content": "We sample every 2 frames per second in each video and use the pre-trained models to extract video features. For the YouCookII and ActivityNet Captions datasets, similar to [21], the ResNet-200 and BN-Inception models are used to extract the appearance and motion features, respectively. Following [58, 64, 66], the InceptionResNetV2 and C3D models are used to extract the appearance and motion features for the MSR-VTT and MSVD dataset, respectively. Meanwhile, the Faster R-CNN method using ResNet-101 as the backbone is adopted to detect object in each video frame, which is trained on the Visual Genome dataset [20]. Notably, only \\( N_{\\mathrm{r}} = 6 \\) classes of the detected objects with top scores are used to form region features in each frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The caption sentences, speech transcripts, and retrieved knowledge in knowledge graph are split into individual words using the NLTK toolkit3, and the GloVe [34] algorithm is used to extract the word embeddings for each word to construct the transcript, caption, and knowledge tokens. The dimension of word embeddings is set to 300. The captions are truncated to 20 words and the maximum text length of transcripts is set to 300 words. For each detected object, we retrieve up to \\( N_{k} = 5 \\) pieces of knowledge items in knowledge graph depending on the relevant scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.872 + ], + "angle": 0, + "content": "Notably, the proposed method is implemented using Pytorch. 3 blocks of multi-modality self-attention and cross-attention modules are used in the MSR-VTT [56] and MSVD [3] datasets, and 2 are used in the YouCookII [69] and ActivityNet Captions [19] datasets, to form the two-stream transformer, respectively. The feature dimension in each block is set to 768, and the number of heads in multi-head architecture is set to 12. The proposed method is trained using the Adam algorithm [17] with the initial learning rate of \\(1e - 4\\). The batch size in training phase is set to 6. The hyper-parameters \\(\\omega_{1},\\omega_{2},\\lambda_{1}\\) and \\(\\lambda_{2}\\) are set to 0.8, 0.2, 0.5 and 0.5, empirically." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.468, + 0.901 + ], + "angle": 0, + "content": "1The key words are exported from the speech transcripts and annotated ground-truth captions in the training sets of the caption datasets. 2https://stanfordnlp.github.io/CoreNLP/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.671, + 0.9 + ], + "angle": 0, + "content": "3https://github.com/nltk/nltk" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18945" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.232 + ], + "angle": 0, + "content": "
MethodBMCRep
Van-Trans [70]7.615.732.37.8
Trans-XL [7]6.614.826.46.3
Trans-XLRG [22]6.614.725.96.0
MART [21]8.016.035.74.4
Van-Trans+COOT [9]11.119.855.65.7
COOT [9]11.319.957.26.7
TextKG14.022.175.92.8
Human---1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.242, + 0.47, + 0.27 + ], + "angle": 0, + "content": "Table 1. Evaluation results on the YouCookII val subset in the paragraph-level evaluation mode." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.284, + 0.45, + 0.3 + ], + "angle": 0, + "content": "4.4. Comparison to the State of the Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.534 + ], + "angle": 0, + "content": "YouCookII dataset. To validate the effectiveness of the proposed method, we compare it to the state-of-the-art (SOTA) methods using the paragraph-level evaluation mode, reported in Table 1. As shown in Table 1, our TextKG method achieves the best results on the YouCookII dataset. Specifically, TextKG improves \\(18.7\\%\\) absolute CIDEr score compared to the SOTA method COOT [9]. COOT [9] focuses on leveraging the hierarchy information and modeling the interactions between different levels granularity and different modalities, such as frames and words, clip and sentences, and videos and paragraphs, which is pre-trained on the large-scale HowTo100M [31] dataset. In contrast, our TextKG method aims to exploit additional knowledge in knowledge graph and multi-modality information in videos to improve the caption results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Beside the paragraph-level evaluation mode, we also report the evaluation results based on the micro-level mode on the YouCookII dataset in Table 2. As shown in Table 2, TextKG significantly outperforms other methods, including Univl [29] and AT [13], both of which use transcript information to enhance model performance by directly concatenating visual and transcript embedding. We argue that with the use of knowledge graphs, TextKG gives a better understanding of the content included in the speech transcripts and thus yields favorable results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.884 + ], + "angle": 0, + "content": "ActivityNet Caption dataset. We also evaluate our TextKG method on the challenging ActivityNet Caption dataset in Table 3. As shown in Table 3, TextKG achieves the best results on 3 out of 4 metrics, i.e., BLEU, METOR, and CIDEr. The ActivityNet Caption dataset [19] contains a series of complex events with annotated caption sentences describing the events that occur. These events may occur over various periods of time and may co-occur. In contrast to existing methods [5, 38, 58, 66], our TextKG understands complex events more comprehensively with the help of exploiting information from transcripts and integrating the relevant knowledge in knowledge graph to enhance the common-sense information for video captioning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "MSR-VTT dataset. The evaluation results on the MSR-" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.295 + ], + "angle": 0, + "content": "
MethodInputBMRC
Masked Trans [70]V3.811.627.438
S3D [53]V3.29.526.131
VideoAsMT [18]V5.313.4--
SwinBERT [27]V915.637.3109
VideoBERT [43]V4.011.027.549
VideoBERT+S3D [43]V4.311.928.850
ActBERT [72]V5.414.330.665
AT [13]V+S9.017.836.7112
DPC [41]V+S2.818.1--
VALUE [24]V+S12.418.840.4130
Univl [29]V+S9.516.337.4115
MV-GPT [40]V+S13.317.635.5103
TextKGV+S11.718.440.2133
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.304, + 0.893, + 0.36 + ], + "angle": 0, + "content": "Table 2. Evaluation results on the YouCookII val subset in the micro-level evaluation mode. 'V' indicates the methods use video appearance information, and 'S' indicates the methods use the speech information." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.375, + 0.895, + 0.517 + ], + "angle": 0, + "content": "
MethodBMCRep
HSE [60]9.813.818.813.2
GVD [67]11.015.722.08.8
GVDsup [67]11.316.422.97.0
Van-Trans [70]9.815.622.27.8
Trans-XL [7]10.415.121.78.5
Trans-XLRG [22]10.214.820.48.9
MART [21]10.315.723.45.2
TextKG11.316.526.66.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.527, + 0.893, + 0.556 + ], + "angle": 0, + "content": "Table 3. Evaluation results on the Activity-Net Captions ae-val subset in the paragraph-level evaluation mode." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.688 + ], + "angle": 0, + "content": "VTT dataset are reported in Table 4. For the fair comparison, we use the same visual features as the existing methods [5, 58, 64, 66]. In contrast to existing methods focusing on network architecture design to exploit visual information, our TextKG aims to exploit mutli-modality information in external knowledge graph and original videos, achieving the state-of-the-art performance on 3 out of 4 metrics, i.e., BLEU, METEOR, and CIDEr." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Meanwhile, we also compare the TextKG method to the SOTA methods focusing on multi-modality pretraining models in Table 5. As shown in Table 5, our TextKG method performs favorably against existing methods, such as DECEM [48], UniVL [29] and MV-GPT [40], that are pretrained on the large-scale datasets (e.g., Howto100M [31]), i.e., improve 0.7 absolute CIDEr score compared to the SOTA method LAVENDER [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "MSVD dataset. We also evaluate the proposed TextKG method on the MSVD dataset [3] that doesn't have speech and repoprt the results in Table 6. As shown in Table 6, our TextKG achieves the best results without speech transcripts by improving \\(1.2\\%\\) and \\(3.8\\%\\) CIDEr scores compared to the SOTA method HMN [58] and the JCRR method [14]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18946" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.462, + 0.238 + ], + "angle": 0, + "content": "
MethodBMRC
OA-BTG [61]41.428.2-46.9
POS-CG [51]4228.261.648.7
MGSA [4]42.427.6-47.5
STG-KD [32]40.528.360.947.1
ORG-TRL [64]43.628.862.150.9
SGN [38]40.828.360.849.5
MGRMP [5]41.728.962.151.4
HMN [58]43.52962.751.5
TextKG43.729.662.452.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.25, + 0.468, + 0.276 + ], + "angle": 0, + "content": "Table 4. Evaluation results on the MSR-VTT test subset in the micro-level evaluation mode." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.284, + 0.462, + 0.418 + ], + "angle": 0, + "content": "
MethodInputFeaturesBMRC
SWINBERT [27]VVidSwin45.430.664.155.9
CLIP4C [47]VCLIP46.130.763.757.7
CMVC [57]VCLIP48.231.364.858.7
LAVENDER [23]VVidSwin---60.1
DeCEM [48]V+SBERT45.229.764.752.3
UniVL [29]V+SS3D41.828.960.850.0
MV-GPT [40]V+SViViT48.9*38.7*64.060.0
TextKG (CLIP)V+SCLIP46.630.564.860.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.429, + 0.468, + 0.526 + ], + "angle": 0, + "content": "Table 5. Comparison to the methods focusing on model pretraining on the MSR-VTT dataset in the micro-level evaluation mode. Notably, following CLIP4C, we use the pre-trained CLIP model on LAION-400M [39] to extract video appearance features. * The authors use a different library to compute BLEU and METOR. Thus, the results on BLEU and METOR are not directly comparable to other methods." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.543, + 0.467, + 0.705 + ], + "angle": 0, + "content": "
MethodBMRC
OA-BTG [61]56.936.2-90.6
POS-CG [51]52.534.171.388.7
MGSA [4]53.435-86.7
STG-KD [32]52.236.973.993.0
ORG-TRL [64]54.336.473.995.2
SGN [38]52.835.572.994.3
MGRMP [5]55.836.974.598.5
JCRR [14]57.036.8-96.8
HMN [58]59.237.775.1104.0
TextKG60.838.575.1105.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.717, + 0.468, + 0.743 + ], + "angle": 0, + "content": "Table 6. Evaluation results on the MSVD test subset in the microlevel evaluation mode." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.773, + 0.468, + 0.819 + ], + "angle": 0, + "content": "exploiting knowledge graph to model the relations between objects, demonstrating the effectiveness of the knowledge graph usage in our method for video captioning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.227, + 0.848 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Influence of different modules. To validate the effectiveness of different modules in TextKG, we conduct several experiments on the YouCookII dataset in Table 7. As" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.888, + 0.259 + ], + "angle": 0, + "content": "
V-FR-FTextG-KGS-KGK-SBMCRep
✓7.415.732.14.1
✓✓9.517.745.95.2
✓✓✓✓9.717.848.94.2
✓✓✓✓9.718.048.53.3
✓✓✓✓✓9.617.749.85.5
✓✓13.021.262.52.7
✓✓✓13.922.171.32.9
✓✓✓✓✓13.722.073.52.0
✓✓✓✓✓13.521.974.82.1
✓✓✓✓✓13.721.872.02.8
✓✓✓✓✓✓14.022.175.92.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Table 7. Ablation study on the YouCookII val subset in the paragraph-level evaluation mode. \"V-F\" and \"R-F\" indicate the video and region features, \"Text\" indicates the speech transcripts features, \"G-KG\" and \"S-KG\" indicate general and specific knowledge graph, and \"K-S\" indicates the knowledge selection mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.621 + ], + "angle": 0, + "content": "shown in Table 7, the baseline method considers the information of video features and predicted captions, achieving 32.1 CIDEr score. After integrating the speech transcripts, the CIDEr score is improved to 62.5, which ensures the key information of videos is considered in generating video captions. After integrating the region features, TextKG achieves 71.3 CIDEr score. The CIDEr score can be further improved 4.6 CIDEr score to 75.9 by exploiting additional knowledge in knowledge graph. Notably, even without the information of speech transcripts, the knowledge graphs are still crucial for video captioning, i.e., improving 3.9 CIDEr score (49.8 vs. 45.9). Meanwhile, we also demonstrate the effectiveness of knowledge selection mechanism in Table 7. As shown in Table 7, without the knowledge selection mechanism, the CIDEr score significantly drops 3.9 to 72.0. We believe that the noisy knowledge is harmful to the accuracy of video captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Influence of transcript quality. To explore the effect of transcript quality on the knowledge graph, we add some noises to the speech transcripts (i.e., mask some amount of speech transcripts) and evaluate the performance of our models on the YouCookII dataset in Figure 3. We find that the method using both the general and specific knowledge graphs performs favorable against other methods with different degrees of masked speech transcripts. Meanwhile, as the amount of masked speech transcripts increasing, the accuracy of the method using specific knowledge graph drops sharply, while the accuracy of the method using general knowledge graph drops at a much slower pace. The aforementioned results demonstrate the importance of general knowledge graph for the video captioning task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.682, + 0.862 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present the qualitative results of the proposed method on the YouCookII dataset in Figure 2, including the key" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.091, + 0.844, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.893, + 0.421 + ], + "angle": 0, + "content": "Figure 2. Qualitative results of the proposed method on the YouCookII dataset. The information in the speech transcripts and knowledge graph is not considered in the baseline." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.451, + 0.456, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Figure 3. Influence of the transcripts qualities. \"w/o KG\" indicates that the method does not use the general and specific knowledge graphs. \"with G-KG\" and \"with S-KG\" indicate that only the general and specific knowledge graphs are used, respectively. \"with G- and S-KG\" indicates that both the general and specific knowledge graphs are used. The CIDEr differences between the \"with G-KG\", \"with S-KG\", \"with G- and S-KG\" methods and the \"w/o KG\" method are shown at the bottom corner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "frames, detected salient objects, speech transcripts, retrieved knowledge items, caption results predicted by our model, ground-truth captions, and key words distributions in the training set. As shown in Figure 2, we observe that the transcript contains key information in the video, e.g., egg, salt, pepper, com starch and flour, to ensure the qual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.569 + ], + "angle": 0, + "content": "ity of the generated video captions. However, the model with only the speech transcripts fails to predict the long-tail phrases, such as sesame oil and black pepper. With the help of knowledge graph providing the relations among these phrases, e.g., sesame oil, black pepper and corn starch, our TextKG model can predict the phrases more accurately. This indicates that the knowledge graph is crucial to mitigate the long-tail word challenges in video captioning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.619, + 0.603 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.72 + ], + "angle": 0, + "content": "In this paper, we present a text with knowledge graph augmented transformer for video captioning, which aims to integrate external knowledge in knowledge graph and exploit the multi-modality information in video to mitigate long-tail words challenge. Extensive experiments conducted on four challenging datasets demonstrate the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.722, + 0.893, + 0.828 + ], + "angle": 0, + "content": "In the future, we plan to improve the proposed method in two directions, i.e., (1) optimizing the knowledge retrieve strategy by considering semantic context information of detected objects and corresponding actions in videos; (2) constructing multi-modal knowledge graph (e.g., the nodes in knowledge graph formed by text, speech, and images or videos) to improve the accuracy of video captioning." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.846, + 0.597, + 0.862 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.507, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18948" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1708-1718, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.149 + ], + "angle": 0, + "content": "[2] William Chan, Navdeep Jaitly, Quoc V. Le, and Oriol Vinyals. Listen, attend and spell. 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.151, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Dekang Lin, Yuji Matsumoto, and Rada Mihalcea, editors, Annual Meeting of the Association for Computational Linguistics, 2011. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.221, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[4] Shaoxiang Chen and Yu-Gang Jiang. Motion guided spatial attention for video captioning. In Association for the Advancement of Artificial Intelligence, pages 8191-8198, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[5] Shaoxiang Chen and Yu-Gang Jiang. Motion guided region message passing for video captioning. In IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. CoRR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.391, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[7] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc Viet Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics, pages 2978-2988, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[8] Michael J. Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 376-380, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[9] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. COOT: cooperative hierarchical transformer for video-text representation learning. In Advances in Neural Information Processing Systems, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[10] Xin Gu, Hanhua Ye, Guang Chen, Yufei Wang, Libo Zhang, and Longyin Wen. Dual-stream transformer for generic event boundary captioning. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[11] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[13] Jack Hessel, Bo Pang, Zhenhai Zhu, and Radu Soricut. A case study on combining ASR and visual features for generating instructional video captions. In Conference on Computational Natural Language Learning, pages 419-429, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[14] Jingyi Hou, Xinxiao Wu, Xiaoxun Zhang, Yayun Qi, Yunde Jia, and Jiebo Luo. Joint commonsense and relation reasoning for image and video captioning. In Association for the Advancement of Artificial Intelligence, pages 10973-10980, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Feicheng Huang, Zhixin Li, Shengjia Chen, Canlong Zhang, and Huifang Ma. Image captioning with internal and external" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "knowledge. In CIKM '20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020, pages 535-544. ACM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, International Conference on Machine Learning, pages 448-456, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, International Conference on Learning Representations, 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[18] Bruno Korbar, Fabio Petroni, Rohit Girdhar, and Lorenzo Torresani. Video understanding as machine translation. CoRR, abs/2006.07203, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In IEEE/CVF International Conference on Computer Vision, pages 706–715, 2017. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[21] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[22] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[23] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. LAVENDER: unifying video-language understanding as masked language modeling. 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Linjie Li, Jie Lei, Zhe Gan, Licheng Yu, Yen-Chun Chen, Rohit Pillai, Yu Cheng, Luowei Zhou, Xin Wang, William Yang Wang, Tamara L. Berg, Mohit Bansal, Jingjing Liu, Lijuan Wang, and Zicheng Liu. VALUE: A multitask benchmark for video-and-language understanding evaluation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[25] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: towards unified-modal understanding and generation via cross-modal contrastive learning. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors, ACL-IJCNLP, pages 2592-2607, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, pages 74-81, 2004. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18949" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17928-17937, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 13–23, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[29] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Xilin Chen, and Ming Zhou. Univilm: A unified video and language pre-training model for multimodal understanding and generation. CoRR, abs/2002.06353, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[31] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[32] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10876, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[34] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing, pages 1532-1543, 2014. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[35] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In IEEE/CVF International Conference on Computer Vision, pages 8907-8916, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[36] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[37] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 1137-1149, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[38] Hobin Ryu, Sunghun Kang, Haeyong Kang, and Chang D. Yoo. Semantic grouping network for video captioning. In Association for the Advancement of Artificial Intelligence, pages 2514-2522, 2021. 1, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.231 + ], + "angle": 0, + "content": "[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17938-17947, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[41] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In Annual Meeting of the Association for Computational Linguistics, pages 6382-6391, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Association for the Advancement of Artificial Intelligence, pages 4444-4451, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.425 + ], + "angle": 0, + "content": "[43] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7463-7472, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[44] Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuanjing Huang, and Zheng Zhang. Colake: Contextualized language and knowledge embedding. In International Conference on Computational Linguistics, pages 3660-3670, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.893, + 0.566 + ], + "angle": 0, + "content": "[45] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Satinder Singh and Shaul Markovitch, editors, Association for the Advancement of Artificial Intelligence, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.893, + 0.622 + ], + "angle": 0, + "content": "[46] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 5099-5110, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.893, + 0.677 + ], + "angle": 0, + "content": "[47] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: CLIP for video caption. In ACM International Conference on Multimedia, pages 4858-4862, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[48] Zineng Tang, Jie Lei, and Mohit Bansal. December: Learning from noisy instructional videos via dense captions and entropy minimization. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 2415-2426, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.893, + 0.844 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[50] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18950" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[51] Bairui Wang, Lin Ma, Wei Zhang, Wenhao Jiang, Jingwen Wang, and Wei Liu. Controllable video captioning with POS sequence guidance based on gated fusion network. In IEEE/CVF International Conference on Computer Vision, pages 2641-2650, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[52] Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. KEPLER: A unified model for knowledge embedding and pre-trained language representation. Transactions of the Association for Computational Linguistics, 9:176-194, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning for video understanding. CoRR, abs/1712.04851, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[54] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In European Conference on Computer Vision, pages 468-483, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. VLM: task-agnostic video-language model pre-training for video understanding. In Findings of the Association for Computational Linguistics, pages 4227–4239, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[56] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[57] Bang Yang, Tong Zhang, and Yuexian Zou. CLIP meets video captioning: Concept-aware representation learning does matter. In Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Shenzhen, China, November 4-7, 2022, Proceedings, Part I, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[58] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. pages 17939-17948, 2022. 1, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[59] Donghan Yu, Chenguang Zhu, Yiming Yang, and Michael Zeng. JAKET: joint pre-training of knowledge graph and language understanding. In Association for the Advancement of Artificial Intelligence, pages 11630-11638, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[60] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In European Conference on Computer Vision, pages 385-401, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[61] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8327-8336, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[62] Yu Zhang, Xinyu Shi, Siya Mi, and Xu Yang. Image captioning with transformer and knowledge graph. Pattern Recognition Letters, 143:43-49, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[63] Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. ERNIE: enhanced language representation with informative entities. In Annual Meeting of the Association for Computational Linguistics, pages 1441-1451, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[64] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph with teacher-recommended learning for video captioning." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13275-13285, 2020. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[65] Wentian Zhao, Yao Hu, Heda Wang, Xinxiao Wu, and Jiebo Luo. Boosting entity-aware image captioning with multimodal knowledge graph. CoRR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[66] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13093-13102, 2020. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J. Corso, and Marcus Rohrbach. Grounded video description. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6578-6587, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[68] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pre-training for image captioning and VQA. In Association for the Advancement of Artificial Intelligence, pages 13041-13049, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[69] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence, pages 7590-7598, 2018. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[70] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8739-8748, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[71] Yimin Zhou, Yiwei Sun, and Vasant G. Honavar. Improving image captioning by leveraging knowledge graphs. In IEEE Winter Conference on Applications of Computer Vision, pages 283-293, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[72] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8743-8752, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18951" + } + ] +] \ No newline at end of file diff --git a/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_origin.pdf b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f29bb0c03b076ccefb730b122a03c227e5695ad7 --- /dev/null +++ b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/4ae6afce-15c9-4a54-b4c2-e5743fb4d16e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14c710ca1fa3f7bf775fa6645abb198d0449c2c2c16aa22829b5cc2445b2ce49 +size 796988 diff --git a/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/full.md b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aa7dedb973924c4e3720cfcb8c4ba3564546ef86 --- /dev/null +++ b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/full.md @@ -0,0 +1,295 @@ +# Text with Knowledge Graph Augmented Transformer for Video Captioning + +Xin Gu $^{1,3}$ , Guang Chen $^{2}$ , Yufei Wang $^{2}$ , Libo Zhang $^{3,1*}$ , Tiejian Luo $^{1}$ , Longyin Wen $^{2}$ $^{1}$ University of Chinese Academy of Sciences, Beijing, China + $^{2}$ ByteDance Inc., San Jose, USA + +$^{3}$ Institute of Software Chinese Academy of Sciences, Beijing, China + +guxin21@mails.ucas.edu.cn, guang.chen@bytedance.com, yufei.wang@bytedance.com + +libo@iscas.ac.cn, tjluo@ucas.ac.cn, longyin.wen@bytedance.com + +# Abstract + +Video captioning aims to describe the content of videos using natural language. Although significant progress has been made, there is still much room to improve the performance for real-world applications, mainly due to the long-tail words challenge. In this paper, we propose a text with knowledge graph augmented transformer (TextKG) for video captioning. Notably, TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is designed to absorb additional knowledge, which models the interactions between the additional knowledge, e.g., pre-built knowledge graph, and the built-in information of videos, e.g., the salient object regions, speech transcripts, and video captions, to mitigate the long-tail words challenge. Meanwhile, the internal stream is designed to exploit the multi-modality information in videos (e.g., the appearance of video frames, speech transcripts, and video captions) to ensure the quality of caption results. In addition, the cross attention mechanism is also used in between the two streams for sharing information. In this way, the two streams can help each other for more accurate results. Extensive experiments conducted on four challenging video captioning datasets, i.e., YouCookII, ActivityNet Captions, MSR-VTT, and MSVD, demonstrate that the proposed method performs favorably against the state-of-the-art methods. Specifically, the proposed TextKG method outperforms the best published results by improving $18.7\%$ absolute CIDEr scores on the YouCookII dataset. + +# 1. Introduction + +Video captioning aims to generate a complete and natural sentence to describe video content, which attracts much + +attention in recent years. Generally, most existing methods [21,38,41,58] require a large amount of paired video and description data for model training. Several datasets, such as YouCookII [69], and ActivityNet Captions [19] are constructed to promote the development of video captioning field. Meanwhile, some methods [29, 40, 48, 72] also use the large-scale narrated video dataset HowTo100M [30] to pretrain the captioning model to further improve the accuracy. + +Although significant progress has been witnessed, it is still a challenge for video captioning methods to be applied in real applications, mainly due to the long-tail issues of words. Most existing methods [29, 40, 48, 72] attempt to design powerful neural networks, trained on the large-scale video-text datasets to make the network learn the relations between video appearances and descriptions. However, it is pretty tough for the networks to accurately predict the objects, properties, or behaviors that are infrequently or never appearing in training data. Some methods [14, 71] attempt to use knowledge graph to exploit the relations between objects for long-tail challenge in image or video captioning, which produces promising results. + +In this paper, we present a text with knowledge graph augmented transformer (TextKG), which integrates additional knowledge in knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge. TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is used to absorb additional knowledge to help mitigate long-tail words challenge by modeling the interactions between the additional knowledge in pre-built knowledge graph, and the built-in information of videos, such as the salient object regions in each frame, speech transcripts, and video captions. Specifically, the information is first retrieved from the pre-built knowledge graphs based on the detected salient objects. After that, we combine the features of the retrieved information, the appearance features of detected salient objects, the features of speech transcripts and + +captions, then feed them into the external stream of TextKG to model the interactions. The internal stream is designed to exploit the multi-modality information in videos, such as the appearance of video frames, speech transcripts and video captions, which can ensure the quality of caption results. To share information between two streams, the cross attention mechanism is introduced. In this way, the two streams can obtain the required modal information from each other for generating more accurate results. The architecture of the proposed method is shown in Figure 1. + +Several experiments conducted on four challenging datasets, i.e., YouCookII [69], ActivityNet Captions [19], MSR-VTT [56], and MSVD [3] demonstrate that the proposed method performs favorably against the state-of-the-art methods. Notably, our TextKG method outperforms the best published results by improving $18.7\%$ and $3.2\%$ absolute CIDEr scores in the paragraph-level evaluation mode on the YouCookII and Activity-Net Captions datasets. + +# 2. Related Work + +Video captioning attracts much attention of researchers in recent years. The best practice has been achieved by attention-based methods, which attempts to associate visual components with sentences in videos. Some of them focus on designing powerful network architectures. VLM [55] and VideoBERT [43] take the visual and text modalities as input, and use a shared transformer to construct a task-agnostic video-language model. ViLBERT [28] processes visual and linguistic information separately with two parallel streams, and then use the attention mechanism to model the interactions between visual and language features. Instead of using the separate encoder-decoder architecture, MART [21] designs a shared encoder-decoder network and augments it with the memory module. ActBert [72] uses local regional features to learn better visual-language alignment. WLT [35] takes audio features as an additional input, and uses context fusion to generate multimodal features. + +Meanwhile, some methods [10,14,15,62,65,71] focus on exploiting prior knowledge to provide semantic correlations and constraints between objects for image or video captioning, producing promising results. ORG-TRL [64] uses the knowledge information in the language model (BERT) to provide candidate words for video captioning. In contrast, we propose a two-stream transformer for video captioning, with the internal stream used to exploit multi-modality information in videos, and the external stream used to model the interactions between the additional knowledge and the built-in information of videos. These two streams use the cross-attention mechanism to share information in different modalities for generating more accurate results. + +Vision-and-language representation learning is a hot topic in recent years. ViLBERT [28], LXMERT [46], UNITER [6], UNIMO [25] and Unified-VL [68] learn + +the representations between image and text, while Univl [29], VideoBERT [43], ActBERT [72] and MV-GPT [40] learn the representations between videos and transcripts. Notably, most of these methods attempt to learn powerful vision-and-language representations by pre-training the models on the large-scale datasets, e.g., Howto100M [31] and WebVid-2M [1], and then finetune them on downstream tasks such as video captioning, video-text retrieval and visual question answering. In contrast, our TextKG method uses the speech transcripts as the text to model the visual and linguistic representations and integrate the additional knowledge in knowledge graph to mitigate long-tail words challenge in video captioning. + +Knowledge graph in NLP. Knowledge graph is an useful tool to indicate the real-world entities and their relations, which provides rich structured knowledge facts for language modeling. Large-scale knowledge graphs are used to train knowledge enhanced language models for various natural language processing (NLP) tasks. CoLAKE [44] proposes to inject the knowledge context of an entity, and to jointly learn the contextualized representation for both language and knowledge by a unified structure. ERNIE [63] enhances BERT architecture to better integrate the knowledge information and textual information. KEPLER [52] not only improves language models by integrating factual knowledge, but also generates text-enhanced knowledge representation. JAKET [59] proposes a joint pre-training framework to model knowledge graph and language simultaneously. Inspired by CoLAKE, our method jointly learns the representations of vision, language and knowledge, and enhances the joint visual-language representations by retrieving relevant knowledge in knowledge graphs. + +# 3. Our Approach + +As mentioned above, our method aims to integrate additional knowledge using knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge in the field of video captioning. We design a text with knowledge graph augmented transformer, i.e., a two-stream transformer formed by the external and internal streams. Both streams are constructed by $N$ sets of alternately stacked self-attention and cross-attention blocks. Specifically, we first use a detector to generate the salient object regions in each video frame, and use the automatic speech recognition (ASR) method [2] to generate the speech transcripts in videos. After that, the class labels of detected salient objects are used to retrieve the prior knowledge in knowledge graphs. Then, the appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior knowledge, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent words of the predicted captions. The overall architecture of the proposed TextKG is shown in Figure 1. + +![](images/6ce9d9fb5365b87c9942daa62558c558037a08a6e7a75d9ff93e0c0f3387b6e9.jpg) +Figure 1. The architecture of our TextKG method, which is formed the external and internal streams. Each stream is stacked with $N$ sets of multi-modality self-attention and cross-attention modules. The cross attention modules are designed to align tokens in different modalities. + +# 3.1. Two-Stream Transformer + +As described above, our two-stream transformer is formed by the external stream and internal stream. Several self-attention and cross-attention blocks are interleaved to construct the two streams. + +Let $\{y_{1}, y_{2}, \dots, y_{l}\}$ be the predicted captions, where $y_{i}$ is the index of the $i$ -th word in the dictionary, $\{p_{1}, p_{2}, \dots, p_{l}\}$ to be the predicted probabilities of words in the dictionary, where $p_{i}$ is the probability of the $i$ -th word. The index of the $i$ -th word $y_{i}$ is computed as $y_{i} = \arg \max_{i} p_{i}$ . Meanwhile, let $z_{i}^{\mathrm{ext}}$ and $z_{i}^{\mathrm{int}}$ to be the output probabilities of the external and internal streams of the $i$ -th word. Thus, we have $p_{i} = \omega_{1} z_{i}^{\mathrm{ext}} + \omega_{2} z_{i}^{\mathrm{int}}$ , where $\omega_{1}$ and $\omega_{2}$ are the hyperparameters used to balance the outputs of the two-streams. In this way, $y_{i}$ is computed as + +$$ +y _ {i} = \arg \max _ {i} \left(\omega_ {1} z _ {i} ^ {\text {e x t}} + \omega_ {2} z _ {i} ^ {\text {i n t}}\right). \tag {1} +$$ + +In the following sections, we will describe each module in our two-stream transformer in more details. + +Multi-modality Self-Attention Module. As shown in Figure 1, we use the self-attention module [49] in both the external and internal streams to model the interactions among multi-modality information. Specifically, for the external stream, the concatenation of feature embeddings of detected objects $\mathcal{F}_{\mathrm{r}}$ , retrieved prior knowledge $\mathcal{F}_{\mathrm{k}}$ , speech transcripts $\mathcal{F}_{\mathrm{s}}$ , and predicted video captions $\mathcal{F}_{\mathrm{c}}$ are fed into the self-attention module to model the interactions, i.e., $\mathcal{X}_{\mathrm{ext}} = \mathrm{Concat}(\mathcal{F}_{\mathrm{r}}, \mathcal{F}_{\mathrm{k}}, \mathcal{F}_{\mathrm{s}}, \mathcal{F}_{\mathrm{c}})$ . Meanwhile, for the internal stream, the concatenation of feature embeddings of speech + +transcripts $\mathcal{F}_{\mathrm{s}}$ , predicted video captions $\mathcal{F}_{\mathrm{c}}$ , and video frame $\mathcal{F}_{\mathrm{v}}$ are fed into the self-attention model to exploit the interactions, i.e., $\mathcal{X}_{\mathrm{int}} = \operatorname {Concat}(\mathcal{F}_{\mathrm{s}},\mathcal{F}_{\mathrm{c}},\mathcal{F}_{\mathrm{v}})$ . The self-attention module computes the interactions as follows, + +$$ +\Phi (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}} + M\right) V, \tag {2} +$$ + +where $d$ is the feature dimension of queries $Q$ and keys $K$ . We have $Q = \mathcal{X}_{(\cdot)}W^{\mathrm{Q}}$ , $K = \mathcal{X}_{(\cdot)}W^{\mathrm{K}}$ , $V = \mathcal{X}_{(\cdot)}W^{\mathrm{V}}$ , where $W^{\mathrm{Q}}$ , $W^{\mathrm{K}}$ , and $W^{\mathrm{V}}$ are learnable parameters, and $\mathcal{X}_{(\cdot)} \in \{\mathcal{X}_{\mathrm{ext}}, \mathcal{X}_{\mathrm{int}}\}$ are the concatenated feature embeddings for the external and internal streams, respectively. + +Following [21], we introduce a mask matrix $M$ in attention function (2) of both the external and internal streams to prevent the model from seeing future words. That is, for the $i$ -th caption token, we set $M_{i,j} = 0$ for $j = 1,\dots ,i$ , and set $M_{i,j} = -\inf$ for $j > i$ . Meanwhile, for the external stream, we set $M_{i,j} = -\inf$ to prevent the associations between irrelevant retrieved prior knowledge and the detected salient object regions. + +Multi-modality Cross-Attention Module. Besides self-attention module, we use the cross-attention module to align the interactions modeled by both the external and internal streams. Specifically, we compute the affinity matrix to guide the alignments between the modeled interactions by injecting the information. Similar to the self-attention module, we also introduce the mask matrix into the attention function to compute the cross-attention. Intuitively, the retrieved prior knowledge should not have direct effect on the predicted captions, but through the detected salient objects. + +Thus, we set the corresponding elements to $-\inf$ to prevent the retrieved prior knowledge directly influencing the predicted captions. + +# 3.2. Optimization + +The cross entropy loss function is used to guide the training of our TextKG method. Given the ground-truth indices of previous $(i - 1)$ words and the concatenated of two-streams $\mathcal{X}_{\mathrm{ext}}$ and $\mathcal{X}_{\mathrm{int}}$ , we can get the predictions of the current $i$ -th word. After that, the training loss of our method is computed as + +$$ +\mathcal {L} = - \sum_ {i = 1} ^ {l} \left(\lambda_ {1} \log z _ {i} ^ {\text {e x t}} + \lambda_ {2} \log z _ {i} ^ {\text {i n t}}\right), \tag {3} +$$ + +where $z_{i}^{\mathrm{ext}}$ and $z_{i}^{\mathrm{int}}$ are the $i$ -th output word of the external and internal streams, $\lambda_{1}$ and $\lambda_{2}$ are the preset parameters used to balance the two streams, and $l$ is the total length of predicted captions. The Adam optimization algorithm [17] with an initial learning rate of $1e - 4$ , $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ and $L2$ weight decay of 0.01 is used to train the model. The Warmup strategy is adopted in training phase by linearly increasing the learning rate from 0 to the initial learning rate for the first $10\%$ of training epochs, and linearly decreasing to 0 for the remaining $90\%$ of epochs. + +# 3.3. Multimodal Tokens + +The appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent captions after tokenization. We will describe the feature extraction and tokenization processes of the aforementioned embeddings. + +Video tokens encode the appearance and motion information of video frames. Specifically, the aligned appearance and motion features form the video tokens. In each second of the video, we sample 2 frames to extract features. Following [21], we use the "Flatten-673" layer in ResNet200 [12] to extract appearance features, and the "global pool" layer in BN-Inception [16] to extract the motion features on the YouCookII [69] and ActivityNet Captions [19] datasets. Meanwhile, for fair comparison, we use the InceptionResNetV2 [45] and C3D [11] models to extract appearance and motion features on the MSR-VTT [56] and MSVD [3] datasets, following the methods [58, 64, 66]. + +Region tokens are used to describe the visual information of the detected salient objects. We use the Faster R-CNN method [37] to extract visual features for detected object regions, which is pretrained on the Visual Genome dataset [20], and the $N_{\mathrm{r}}$ detected objects with highest confidence score in each frame are reserved for video captioning. + +Transcript tokens are used to encode the content of speech transcripts, which contains some key information in videos. The GloVe model [34] is used to extract features of each word in transcripts to form the transcript tokens. + +Caption tokens encode the information of predicted captions. Similar to transcript tokens, GloVe [34] is adopted to extract features of each word in the generated captions. + +Knowledge tokens model the content of retrieved knowledge, which is important to mitigate the long-tail words challenge in video captioning. Our TextKG method uses a triple structure to encode the retrieved knowledge, formed by two object items and the relations between them, i.e., $(\alpha_{\mathrm{head}},\alpha_{\mathrm{tail}},\rho_{\mathrm{rel}})$ , where $\alpha_{\mathrm{head}}$ and $\alpha_{\mathrm{tail}}$ are the two nodes in knowledge graph, and $\rho_{\mathrm{rel}}$ indicates the edge encoding the relations between them. For example, $\alpha_{\mathrm{head}} = "\mathrm{knife}$ , $\alpha_{\mathrm{tail}} = "\mathrm{hard}$ , and $\rho_{\mathrm{rel}} = "\mathrm{has~property}$ . For each detected object $\alpha_{\mathrm{head}}$ , we aim to integrate prior context information for more accurate caption results. That is, we use the GloVe model to extract linguistic features of the node $\alpha_{\mathrm{tail}}$ and set the learnable embeddings for the edge $\rho_{\mathrm{rel}}$ . After that, we compute the summation of them to get the knowledge embeddings of the knowledge tokens. + +For each detected object, we use its category name to retrieve knowledge (i.e., triples containing the category name of the detected object) from the knowledge graph. Notably, only a portion of the retrieved knowledge are highly related to the video. The rests are potentially harmful to the performance of caption generation as it may distract the training and inference and also increase computational cost. For example, to generate caption of a cooking tutorial video, the knowledge "knife - used to -cut" is more useful rather than "knife - has property - hard". Because cutting food may be an important procedure for food preparation. Thus, we design a primary rank mechanism to sort the retrieved knowledge based on their semantic similarity to the video. The cosine similarity between the embeddings of knowledge tokens and video transcripts is computed as the semantic similarity for ranking. The pretrained SBERT [36] model is used to extract feature embeddings of transcripts. Finally, $N_{\mathrm{k}}$ knowledge items with the highest similarity scores are reserved for video captioning. + +# 3.4. Knowledge Graphs Construction + +As mentioned above, we use the knowledge graph to retain the prior knowledge. The nodes in the graph could correspond to a noun, an adjective, a verb or an adverb, etc. The edges in the graph are used to encode the relations between different nodes. Apparently, it is important to include all key information of videos in knowledge graphs for accurate caption results. Thus, besides general knowledge graph covering most key knowledge in general scenarios, we also introduce the specific knowledge graph for each dataset covering the key knowledge in specific scenarios. + +General knowledge graph is designed to include most key information in general scenarios that we are interested, such as cooking and activity. Specifically, the general knowledge graph is exported from the public available giant knowledge + +graph ConceptNet [42] by extracting key words1 in ConceptNet with the connected edges and neighboring nodes. + +Specific knowledge graph. Besides the general knowledge graph, we also construct the specific knowledge graph to cover key information in specific scenarios. We believe that the speech transcripts of videos contain most of the crucial information and use the speech transcripts as the source to construct the specific knowledge graph. We first use the automatic speech recognition (ASR) model [2] to convert the speech in all videos to transcripts. After that, we use the Stanford NLP library² to analyze the components in each sentence in transcripts to form a structured grammar tree. We collect the "adjective and noun", "noun and noun" and "adverb and verb" phrases from the grammar trees to construct the specific knowledge graph. + +# 4. Experiments + +# 4.1. Datasets + +YoucookII contains 2,000 long untrimmed videos from 89 cooking recipes, including 1,333 videos for training, 457 videos for testing. The average length of videos is 5.3 minutes and each video is divided into 7.7 video clips on average with the manually labeled caption sentences. + +ActivityNet Captions is a large-scale challenging dataset for video captioning, with 10,024 videos for training, 4,926 for validation, and 5,044 for testing. The average length of each video is 2.1 minutes and is divided into 3.65 clips on average. Each clip is assigned with a manually labeled caption sentence. Since the testing set is not publicly available, following MART [21], we split the validation set into two subsets, i.e., ae-val with 2,460 videos for validation and aetest with 2,457 videos for testing. + +MSR-VTT includes 10,000 video clips with 15 seconds on average. Each video clip contains 20 human annotated captions and the average length of each sentence is 9.28 words. Following [5,38,58], 6,513,497, and 2,990 videos are used for training, validation, and testing, respectively. + +MSVD consists of 1970 video clips collected from YouTube and each clip is approximately 10 seconds long. Each clip is labeled with 40 sentences. Following [5,38,58], we split the dataset into 1,200,100,and 670 video clips for training, validation and testing, respectively. + +# 4.2. Evaluation metrics + +To compare the performance of the proposed method with other methods, five model-free automatic evaluation metrics are used, including BLEU@4 [33] (abbreviated to B) that are precision-based metric, METEOR [8] (abbreviated to M) that computes sentence-level scores, CIDEr [50] + +(abbreviated to C) that is consensus-based metric, ROUGE [26] (abbreviated to R) that uses longest common subsequence to compute the similarities between sentences, and Rep@4 [54] (abbreviated to Rep) that computes the redundancy score for every captions. Lower Rep indicates better model. The CIDEr score is used as the primary metric in evaluation. Notably, both micro-level and paragraph-level evaluation modes are used. The micro-level mode evaluates the predicted captions of each video clip independently, while the paragraph-level mode considers the relations between captions of different clips. Following [9, 21], the paragraph-level evaluation mode is used in the YouCookII and ActivityNet datasets, which contains multiple clips in each video. + +# 4.3. Implementation Details + +We sample every 2 frames per second in each video and use the pre-trained models to extract video features. For the YouCookII and ActivityNet Captions datasets, similar to [21], the ResNet-200 and BN-Inception models are used to extract the appearance and motion features, respectively. Following [58, 64, 66], the InceptionResNetV2 and C3D models are used to extract the appearance and motion features for the MSR-VTT and MSVD dataset, respectively. Meanwhile, the Faster R-CNN method using ResNet-101 as the backbone is adopted to detect object in each video frame, which is trained on the Visual Genome dataset [20]. Notably, only $N_{\mathrm{r}} = 6$ classes of the detected objects with top scores are used to form region features in each frame. + +The caption sentences, speech transcripts, and retrieved knowledge in knowledge graph are split into individual words using the NLTK toolkit3, and the GloVe [34] algorithm is used to extract the word embeddings for each word to construct the transcript, caption, and knowledge tokens. The dimension of word embeddings is set to 300. The captions are truncated to 20 words and the maximum text length of transcripts is set to 300 words. For each detected object, we retrieve up to $N_{k} = 5$ pieces of knowledge items in knowledge graph depending on the relevant scores. + +Notably, the proposed method is implemented using Pytorch. 3 blocks of multi-modality self-attention and cross-attention modules are used in the MSR-VTT [56] and MSVD [3] datasets, and 2 are used in the YouCookII [69] and ActivityNet Captions [19] datasets, to form the two-stream transformer, respectively. The feature dimension in each block is set to 768, and the number of heads in multi-head architecture is set to 12. The proposed method is trained using the Adam algorithm [17] with the initial learning rate of $1e - 4$ . The batch size in training phase is set to 6. The hyper-parameters $\omega_{1},\omega_{2},\lambda_{1}$ and $\lambda_{2}$ are set to 0.8, 0.2, 0.5 and 0.5, empirically. + +
MethodBMCRep
Van-Trans [70]7.615.732.37.8
Trans-XL [7]6.614.826.46.3
Trans-XLRG [22]6.614.725.96.0
MART [21]8.016.035.74.4
Van-Trans+COOT [9]11.119.855.65.7
COOT [9]11.319.957.26.7
TextKG14.022.175.92.8
Human---1.0
+ +# 4.4. Comparison to the State of the Art Methods + +YouCookII dataset. To validate the effectiveness of the proposed method, we compare it to the state-of-the-art (SOTA) methods using the paragraph-level evaluation mode, reported in Table 1. As shown in Table 1, our TextKG method achieves the best results on the YouCookII dataset. Specifically, TextKG improves $18.7\%$ absolute CIDEr score compared to the SOTA method COOT [9]. COOT [9] focuses on leveraging the hierarchy information and modeling the interactions between different levels granularity and different modalities, such as frames and words, clip and sentences, and videos and paragraphs, which is pre-trained on the large-scale HowTo100M [31] dataset. In contrast, our TextKG method aims to exploit additional knowledge in knowledge graph and multi-modality information in videos to improve the caption results. + +Beside the paragraph-level evaluation mode, we also report the evaluation results based on the micro-level mode on the YouCookII dataset in Table 2. As shown in Table 2, TextKG significantly outperforms other methods, including Univl [29] and AT [13], both of which use transcript information to enhance model performance by directly concatenating visual and transcript embedding. We argue that with the use of knowledge graphs, TextKG gives a better understanding of the content included in the speech transcripts and thus yields favorable results. + +ActivityNet Caption dataset. We also evaluate our TextKG method on the challenging ActivityNet Caption dataset in Table 3. As shown in Table 3, TextKG achieves the best results on 3 out of 4 metrics, i.e., BLEU, METOR, and CIDEr. The ActivityNet Caption dataset [19] contains a series of complex events with annotated caption sentences describing the events that occur. These events may occur over various periods of time and may co-occur. In contrast to existing methods [5, 38, 58, 66], our TextKG understands complex events more comprehensively with the help of exploiting information from transcripts and integrating the relevant knowledge in knowledge graph to enhance the common-sense information for video captioning. + +MSR-VTT dataset. The evaluation results on the MSR- + +Table 1. Evaluation results on the YouCookII val subset in the paragraph-level evaluation mode. + +
MethodInputBMRC
Masked Trans [70]V3.811.627.438
S3D [53]V3.29.526.131
VideoAsMT [18]V5.313.4--
SwinBERT [27]V915.637.3109
VideoBERT [43]V4.011.027.549
VideoBERT+S3D [43]V4.311.928.850
ActBERT [72]V5.414.330.665
AT [13]V+S9.017.836.7112
DPC [41]V+S2.818.1--
VALUE [24]V+S12.418.840.4130
Univl [29]V+S9.516.337.4115
MV-GPT [40]V+S13.317.635.5103
TextKGV+S11.718.440.2133
+ +Table 2. Evaluation results on the YouCookII val subset in the micro-level evaluation mode. 'V' indicates the methods use video appearance information, and 'S' indicates the methods use the speech information. + +
MethodBMCRep
HSE [60]9.813.818.813.2
GVD [67]11.015.722.08.8
GVDsup [67]11.316.422.97.0
Van-Trans [70]9.815.622.27.8
Trans-XL [7]10.415.121.78.5
Trans-XLRG [22]10.214.820.48.9
MART [21]10.315.723.45.2
TextKG11.316.526.66.3
+ +Table 3. Evaluation results on the Activity-Net Captions ae-val subset in the paragraph-level evaluation mode. + +VTT dataset are reported in Table 4. For the fair comparison, we use the same visual features as the existing methods [5, 58, 64, 66]. In contrast to existing methods focusing on network architecture design to exploit visual information, our TextKG aims to exploit mutli-modality information in external knowledge graph and original videos, achieving the state-of-the-art performance on 3 out of 4 metrics, i.e., BLEU, METEOR, and CIDEr. + +Meanwhile, we also compare the TextKG method to the SOTA methods focusing on multi-modality pretraining models in Table 5. As shown in Table 5, our TextKG method performs favorably against existing methods, such as DECEM [48], UniVL [29] and MV-GPT [40], that are pretrained on the large-scale datasets (e.g., Howto100M [31]), i.e., improve 0.7 absolute CIDEr score compared to the SOTA method LAVENDER [23]. + +MSVD dataset. We also evaluate the proposed TextKG method on the MSVD dataset [3] that doesn't have speech and repoprt the results in Table 6. As shown in Table 6, our TextKG achieves the best results without speech transcripts by improving $1.2\%$ and $3.8\%$ CIDEr scores compared to the SOTA method HMN [58] and the JCRR method [14]. + +
MethodBMRC
OA-BTG [61]41.428.2-46.9
POS-CG [51]4228.261.648.7
MGSA [4]42.427.6-47.5
STG-KD [32]40.528.360.947.1
ORG-TRL [64]43.628.862.150.9
SGN [38]40.828.360.849.5
MGRMP [5]41.728.962.151.4
HMN [58]43.52962.751.5
TextKG43.729.662.452.4
+ +Table 4. Evaluation results on the MSR-VTT test subset in the micro-level evaluation mode. + +
MethodInputFeaturesBMRC
SWINBERT [27]VVidSwin45.430.664.155.9
CLIP4C [47]VCLIP46.130.763.757.7
CMVC [57]VCLIP48.231.364.858.7
LAVENDER [23]VVidSwin---60.1
DeCEM [48]V+SBERT45.229.764.752.3
UniVL [29]V+SS3D41.828.960.850.0
MV-GPT [40]V+SViViT48.9*38.7*64.060.0
TextKG (CLIP)V+SCLIP46.630.564.860.8
+ +Table 5. Comparison to the methods focusing on model pretraining on the MSR-VTT dataset in the micro-level evaluation mode. Notably, following CLIP4C, we use the pre-trained CLIP model on LAION-400M [39] to extract video appearance features. * The authors use a different library to compute BLEU and METOR. Thus, the results on BLEU and METOR are not directly comparable to other methods. + +
MethodBMRC
OA-BTG [61]56.936.2-90.6
POS-CG [51]52.534.171.388.7
MGSA [4]53.435-86.7
STG-KD [32]52.236.973.993.0
ORG-TRL [64]54.336.473.995.2
SGN [38]52.835.572.994.3
MGRMP [5]55.836.974.598.5
JCRR [14]57.036.8-96.8
HMN [58]59.237.775.1104.0
TextKG60.838.575.1105.2
+ +exploiting knowledge graph to model the relations between objects, demonstrating the effectiveness of the knowledge graph usage in our method for video captioning. + +# 4.5. Ablation study + +Influence of different modules. To validate the effectiveness of different modules in TextKG, we conduct several experiments on the YouCookII dataset in Table 7. As + +Table 6. Evaluation results on the MSVD test subset in the microlevel evaluation mode. + +
V-FR-FTextG-KGS-KGK-SBMCRep
✓7.415.732.14.1
✓✓9.517.745.95.2
✓✓✓✓9.717.848.94.2
✓✓✓✓9.718.048.53.3
✓✓✓✓✓9.617.749.85.5
✓✓13.021.262.52.7
✓✓✓13.922.171.32.9
✓✓✓✓✓13.722.073.52.0
✓✓✓✓✓13.521.974.82.1
✓✓✓✓✓13.721.872.02.8
✓✓✓✓✓✓14.022.175.92.8
+ +Table 7. Ablation study on the YouCookII val subset in the paragraph-level evaluation mode. "V-F" and "R-F" indicate the video and region features, "Text" indicates the speech transcripts features, "G-KG" and "S-KG" indicate general and specific knowledge graph, and "K-S" indicates the knowledge selection mechanism. + +shown in Table 7, the baseline method considers the information of video features and predicted captions, achieving 32.1 CIDEr score. After integrating the speech transcripts, the CIDEr score is improved to 62.5, which ensures the key information of videos is considered in generating video captions. After integrating the region features, TextKG achieves 71.3 CIDEr score. The CIDEr score can be further improved 4.6 CIDEr score to 75.9 by exploiting additional knowledge in knowledge graph. Notably, even without the information of speech transcripts, the knowledge graphs are still crucial for video captioning, i.e., improving 3.9 CIDEr score (49.8 vs. 45.9). Meanwhile, we also demonstrate the effectiveness of knowledge selection mechanism in Table 7. As shown in Table 7, without the knowledge selection mechanism, the CIDEr score significantly drops 3.9 to 72.0. We believe that the noisy knowledge is harmful to the accuracy of video captions. + +Influence of transcript quality. To explore the effect of transcript quality on the knowledge graph, we add some noises to the speech transcripts (i.e., mask some amount of speech transcripts) and evaluate the performance of our models on the YouCookII dataset in Figure 3. We find that the method using both the general and specific knowledge graphs performs favorable against other methods with different degrees of masked speech transcripts. Meanwhile, as the amount of masked speech transcripts increasing, the accuracy of the method using specific knowledge graph drops sharply, while the accuracy of the method using general knowledge graph drops at a much slower pace. The aforementioned results demonstrate the importance of general knowledge graph for the video captioning task. + +# 4.6. Qualitative Results + +We present the qualitative results of the proposed method on the YouCookII dataset in Figure 2, including the key + +![](images/ae5f712b8f6985f028f0661f11d732f3f0808acc8f826d1b89a9aa421a3dd5b0.jpg) +Figure 2. Qualitative results of the proposed method on the YouCookII dataset. The information in the speech transcripts and knowledge graph is not considered in the baseline. + +![](images/1e507a2235cb0683d733417660532c4da6f9607b49896c50c532d291e2a7622e.jpg) +Figure 3. Influence of the transcripts qualities. "w/o KG" indicates that the method does not use the general and specific knowledge graphs. "with G-KG" and "with S-KG" indicate that only the general and specific knowledge graphs are used, respectively. "with G- and S-KG" indicates that both the general and specific knowledge graphs are used. The CIDEr differences between the "with G-KG", "with S-KG", "with G- and S-KG" methods and the "w/o KG" method are shown at the bottom corner. + +frames, detected salient objects, speech transcripts, retrieved knowledge items, caption results predicted by our model, ground-truth captions, and key words distributions in the training set. As shown in Figure 2, we observe that the transcript contains key information in the video, e.g., egg, salt, pepper, com starch and flour, to ensure the qual + +ity of the generated video captions. However, the model with only the speech transcripts fails to predict the long-tail phrases, such as sesame oil and black pepper. With the help of knowledge graph providing the relations among these phrases, e.g., sesame oil, black pepper and corn starch, our TextKG model can predict the phrases more accurately. This indicates that the knowledge graph is crucial to mitigate the long-tail word challenges in video captioning. + +# 5. Conclusion + +In this paper, we present a text with knowledge graph augmented transformer for video captioning, which aims to integrate external knowledge in knowledge graph and exploit the multi-modality information in video to mitigate long-tail words challenge. Extensive experiments conducted on four challenging datasets demonstrate the effectiveness of the proposed method. + +In the future, we plan to improve the proposed method in two directions, i.e., (1) optimizing the knowledge retrieve strategy by considering semantic context information of detected objects and corresponding actions in videos; (2) constructing multi-modal knowledge graph (e.g., the nodes in knowledge graph formed by text, speech, and images or videos) to improve the accuracy of video captioning. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for + +end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1708-1718, 2021. 2 +[2] William Chan, Navdeep Jaitly, Quoc V. Le, and Oriol Vinyals. Listen, attend and spell. 2015. 2, 5 +[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Dekang Lin, Yuji Matsumoto, and Rada Mihalcea, editors, Annual Meeting of the Association for Computational Linguistics, 2011. 2, 4, 5, 6 +[4] Shaoxiang Chen and Yu-Gang Jiang. Motion guided spatial attention for video captioning. In Association for the Advancement of Artificial Intelligence, pages 8191-8198, 2019. 7 +[5] Shaoxiang Chen and Yu-Gang Jiang. Motion guided region message passing for video captioning. In IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 5, 6, 7 +[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. CoRR, 2019. 2 +[7] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc Viet Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics, pages 2978-2988, 2019. 6 +[8] Michael J. Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 376-380, 2014. 5 +[9] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. COOT: cooperative hierarchical transformer for video-text representation learning. In Advances in Neural Information Processing Systems, 2020. 5, 6 +[10] Xin Gu, Hanhua Ye, Guang Chen, Yufei Wang, Libo Zhang, and Longyin Wen. Dual-stream transformer for generic event boundary captioning. 2022. 2 +[11] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 4 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4 +[13] Jack Hessel, Bo Pang, Zhenhai Zhu, and Radu Soricut. A case study on combining ASR and visual features for generating instructional video captions. In Conference on Computational Natural Language Learning, pages 419-429, 2019. 6 +[14] Jingyi Hou, Xinxiao Wu, Xiaoxun Zhang, Yayun Qi, Yunde Jia, and Jiebo Luo. Joint commonsense and relation reasoning for image and video captioning. In Association for the Advancement of Artificial Intelligence, pages 10973-10980, 2020. 1, 2, 6, 7 +[15] Feicheng Huang, Zhixin Li, Shengjia Chen, Canlong Zhang, and Huifang Ma. Image captioning with internal and external + +knowledge. In CIKM '20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020, pages 535-544. ACM, 2020. 2 +[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, International Conference on Machine Learning, pages 448-456, 2015. 4 +[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, International Conference on Learning Representations, 2015. 4, 5 +[18] Bruno Korbar, Fabio Petroni, Rohit Girdhar, and Lorenzo Torresani. Video understanding as machine translation. CoRR, abs/2006.07203, 2020. 6 +[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In IEEE/CVF International Conference on Computer Vision, pages 706–715, 2017. 1, 2, 4, 5, 6 +[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 5 +[21] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 1, 2, 3, 4, 5, 6 +[22] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 6 +[23] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. LAVENDER: unifying video-language understanding as masked language modeling. 2022. 6, 7 +[24] Linjie Li, Jie Lei, Zhe Gan, Licheng Yu, Yen-Chun Chen, Rohit Pillai, Yu Cheng, Luowei Zhou, Xin Wang, William Yang Wang, Tamara L. Berg, Mohit Bansal, Jingjing Liu, Lijuan Wang, and Zicheng Liu. VALUE: A multitask benchmark for video-and-language understanding evaluation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 6 +[25] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: towards unified-modal understanding and generation via cross-modal contrastive learning. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors, ACL-IJCNLP, pages 2592-2607, 2021. 2 +[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, pages 74-81, 2004. 5 + +[27] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17928-17937, 2022. 6, 7 +[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 13–23, 2019. 2 +[29] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Xilin Chen, and Ming Zhou. Univilm: A unified video and language pre-training model for multimodal understanding and generation. CoRR, abs/2002.06353, 2020. 1, 2, 6, 7 +[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1 +[31] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2, 6 +[32] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10876, 2020. 7 +[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5 +[34] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing, pages 1532-1543, 2014. 4, 5 +[35] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In IEEE/CVF International Conference on Computer Vision, pages 8907-8916, 2019. 2 +[36] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 4 +[37] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 1137-1149, 2017. 4 +[38] Hobin Ryu, Sunghun Kang, Haeyong Kang, and Chang D. Yoo. Semantic grouping network for video captioning. In Association for the Advancement of Artificial Intelligence, pages 2514-2522, 2021. 1, 5, 6, 7 + +[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, 2021. 7 +[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17938-17947, 2022. 1, 2, 6, 7 +[41] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In Annual Meeting of the Association for Computational Linguistics, pages 6382-6391, 2019. 1, 6 +[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Association for the Advancement of Artificial Intelligence, pages 4444-4451, 2017. 5 +[43] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7463-7472, 2019. 2, 6 +[44] Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuanjing Huang, and Zheng Zhang. Colake: Contextualized language and knowledge embedding. In International Conference on Computational Linguistics, pages 3660-3670, 2020. 2 +[45] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Satinder Singh and Shaul Markovitch, editors, Association for the Advancement of Artificial Intelligence, 2017. 4 +[46] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 5099-5110, 2019. 2 +[47] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: CLIP for video caption. In ACM International Conference on Multimedia, pages 4858-4862, 2021. 7 +[48] Zineng Tang, Jie Lei, and Mohit Bansal. December: Learning from noisy instructional videos via dense captions and entropy minimization. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 2415-2426, 2021. 1, 6, 7 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 3 +[50] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 5 + +[51] Bairui Wang, Lin Ma, Wei Zhang, Wenhao Jiang, Jingwen Wang, and Wei Liu. Controllable video captioning with POS sequence guidance based on gated fusion network. In IEEE/CVF International Conference on Computer Vision, pages 2641-2650, 2019. 7 +[52] Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. KEPLER: A unified model for knowledge embedding and pre-trained language representation. Transactions of the Association for Computational Linguistics, 9:176-194, 2021. 2 +[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning for video understanding. CoRR, abs/1712.04851, 2017. 6 +[54] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In European Conference on Computer Vision, pages 468-483, 2018. 5 +[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. VLM: task-agnostic video-language model pre-training for video understanding. In Findings of the Association for Computational Linguistics, pages 4227–4239, 2021. 2 +[56] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 2, 4, 5 +[57] Bang Yang, Tong Zhang, and Yuexian Zou. CLIP meets video captioning: Concept-aware representation learning does matter. In Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Shenzhen, China, November 4-7, 2022, Proceedings, Part I, 2022. 7 +[58] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. pages 17939-17948, 2022. 1, 4, 5, 6, 7 +[59] Donghan Yu, Chenguang Zhu, Yiming Yang, and Michael Zeng. JAKET: joint pre-training of knowledge graph and language understanding. In Association for the Advancement of Artificial Intelligence, pages 11630-11638, 2022. 2 +[60] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In European Conference on Computer Vision, pages 385-401, 2018. 6 +[61] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8327-8336, 2019. 7 +[62] Yu Zhang, Xinyu Shi, Siya Mi, and Xu Yang. Image captioning with transformer and knowledge graph. Pattern Recognition Letters, 143:43-49, 2021. 2 +[63] Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. ERNIE: enhanced language representation with informative entities. In Annual Meeting of the Association for Computational Linguistics, pages 1441-1451, 2019. 2 +[64] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph with teacher-recommended learning for video captioning. + +In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13275-13285, 2020. 2, 4, 5, 6, 7 +[65] Wentian Zhao, Yao Hu, Heda Wang, Xinxiao Wu, and Jiebo Luo. Boosting entity-aware image captioning with multimodal knowledge graph. CoRR, 2021. 2 +[66] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13093-13102, 2020. 4, 5, 6 +[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J. Corso, and Marcus Rohrbach. Grounded video description. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6578-6587, 2019. 6 +[68] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pre-training for image captioning and VQA. In Association for the Advancement of Artificial Intelligence, pages 13041-13049, 2020. 2 +[69] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence, pages 7590-7598, 2018. 1, 2, 4, 5 +[70] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8739-8748, 2018. 6 +[71] Yimin Zhou, Yiwei Sun, and Vasant G. Honavar. Improving image captioning by leveraging knowledge graphs. In IEEE Winter Conference on Applications of Computer Vision, pages 283-293, 2019. 1, 2 +[72] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8743-8752, 2020. 1, 2, 6 \ No newline at end of file diff --git a/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/images.zip b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0ecd926cb672594302927e888d2e01423cd9e04 --- /dev/null +++ b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32f5a09b804ca72718ce78783bbd4e8e31cecb4b6297913d53fe089c06ae980b +size 542334 diff --git a/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/layout.json b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b116955f54cc6f6a41fbfd4096c47738a3d58045 --- /dev/null +++ b/2023/Text With Knowledge Graph Augmented Transformer for Video Captioning/layout.json @@ -0,0 +1,8649 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 531, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 531, + 122 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 531, + 122 + ], + "type": "text", + "content": "Text with Knowledge Graph Augmented Transformer for Video Captioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "spans": [ + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": "Xin Gu" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": ", Guang Chen" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": ", Yufei Wang" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": ", Libo Zhang" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{3,1*}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": ", Tiejian Luo" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": ", Longyin Wen" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences, Beijing, China \n" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 141, + 500, + 184 + ], + "type": "text", + "content": "ByteDance Inc., San Jose, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 185, + 462, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 185, + 462, + 200 + ], + "spans": [ + { + "bbox": [ + 131, + 185, + 462, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 131, + 185, + 462, + 200 + ], + "type": "text", + "content": "Institute of Software Chinese Academy of Sciences, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 201, + 506, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 201, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 87, + 201, + 506, + 213 + ], + "type": "text", + "content": "guxin21@mails.ucas.edu.cn, guang.chen@bytedance.com, yufei.wang@bytedance.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 216, + 462, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 216, + 462, + 227 + ], + "spans": [ + { + "bbox": [ + 129, + 216, + 462, + 227 + ], + "type": "text", + "content": "libo@iscas.ac.cn, tjluo@ucas.ac.cn, longyin.wen@bytedance.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 281, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 605 + ], + "type": "text", + "content": "Video captioning aims to describe the content of videos using natural language. Although significant progress has been made, there is still much room to improve the performance for real-world applications, mainly due to the long-tail words challenge. In this paper, we propose a text with knowledge graph augmented transformer (TextKG) for video captioning. Notably, TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is designed to absorb additional knowledge, which models the interactions between the additional knowledge, e.g., pre-built knowledge graph, and the built-in information of videos, e.g., the salient object regions, speech transcripts, and video captions, to mitigate the long-tail words challenge. Meanwhile, the internal stream is designed to exploit the multi-modality information in videos (e.g., the appearance of video frames, speech transcripts, and video captions) to ensure the quality of caption results. In addition, the cross attention mechanism is also used in between the two streams for sharing information. In this way, the two streams can help each other for more accurate results. Extensive experiments conducted on four challenging video captioning datasets, i.e., YouCookII, ActivityNet Captions, MSR-VTT, and MSVD, demonstrate that the proposed method performs favorably against the state-of-the-art methods. Specifically, the proposed TextKG method outperforms the best published results by improving " + }, + { + "bbox": [ + 46, + 281, + 289, + 605 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 46, + 281, + 289, + 605 + ], + "type": "text", + "content": " absolute CIDEr scores on the YouCookII dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 650, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 287, + 673 + ], + "type": "text", + "content": "Video captioning aims to generate a complete and natural sentence to describe video content, which attracts much" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 256, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 547, + 363 + ], + "type": "text", + "content": "attention in recent years. Generally, most existing methods [21,38,41,58] require a large amount of paired video and description data for model training. Several datasets, such as YouCookII [69], and ActivityNet Captions [19] are constructed to promote the development of video captioning field. Meanwhile, some methods [29, 40, 48, 72] also use the large-scale narrated video dataset HowTo100M [30] to pretrain the captioning model to further improve the accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 520 + ], + "type": "text", + "content": "Although significant progress has been witnessed, it is still a challenge for video captioning methods to be applied in real applications, mainly due to the long-tail issues of words. Most existing methods [29, 40, 48, 72] attempt to design powerful neural networks, trained on the large-scale video-text datasets to make the network learn the relations between video appearances and descriptions. However, it is pretty tough for the networks to accurately predict the objects, properties, or behaviors that are infrequently or never appearing in training data. Some methods [14, 71] attempt to use knowledge graph to exploit the relations between objects for long-tail challenge in image or video captioning, which produces promising results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "In this paper, we present a text with knowledge graph augmented transformer (TextKG), which integrates additional knowledge in knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge. TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is used to absorb additional knowledge to help mitigate long-tail words challenge by modeling the interactions between the additional knowledge in pre-built knowledge graph, and the built-in information of videos, such as the salient object regions in each frame, speech transcripts, and video captions. Specifically, the information is first retrieved from the pre-built knowledge graphs based on the detected salient objects. After that, we combine the features of the retrieved information, the appearance features of detected salient objects, the features of speech transcripts and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 714 + ], + "type": "text", + "content": "*Corresponding author:libo@iscas.ac.cn, Libo Zhang was supported by Youth Innovation Promotion Association, CAS (2020111). This work was done during internship at ByteDance Inc." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18941" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 191 + ], + "type": "text", + "content": "captions, then feed them into the external stream of TextKG to model the interactions. The internal stream is designed to exploit the multi-modality information in videos, such as the appearance of video frames, speech transcripts and video captions, which can ensure the quality of caption results. To share information between two streams, the cross attention mechanism is introduced. In this way, the two streams can obtain the required modal information from each other for generating more accurate results. The architecture of the proposed method is shown in Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "type": "text", + "content": "Several experiments conducted on four challenging datasets, i.e., YouCookII [69], ActivityNet Captions [19], MSR-VTT [56], and MSVD [3] demonstrate that the proposed method performs favorably against the state-of-the-art methods. Notably, our TextKG method outperforms the best published results by improving " + }, + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 46, + 192, + 286, + 287 + ], + "type": "text", + "content": " absolute CIDEr scores in the paragraph-level evaluation mode on the YouCookII and Activity-Net Captions datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 298, + 133, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 133, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 133, + 310 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 286, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 521 + ], + "type": "text", + "content": "Video captioning attracts much attention of researchers in recent years. The best practice has been achieved by attention-based methods, which attempts to associate visual components with sentences in videos. Some of them focus on designing powerful network architectures. VLM [55] and VideoBERT [43] take the visual and text modalities as input, and use a shared transformer to construct a task-agnostic video-language model. ViLBERT [28] processes visual and linguistic information separately with two parallel streams, and then use the attention mechanism to model the interactions between visual and language features. Instead of using the separate encoder-decoder architecture, MART [21] designs a shared encoder-decoder network and augments it with the memory module. ActBert [72] uses local regional features to learn better visual-language alignment. WLT [35] takes audio features as an additional input, and uses context fusion to generate multimodal features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 286, + 677 + ], + "type": "text", + "content": "Meanwhile, some methods [10,14,15,62,65,71] focus on exploiting prior knowledge to provide semantic correlations and constraints between objects for image or video captioning, producing promising results. ORG-TRL [64] uses the knowledge information in the language model (BERT) to provide candidate words for video captioning. In contrast, we propose a two-stream transformer for video captioning, with the internal stream used to exploit multi-modality information in videos, and the external stream used to model the interactions between the additional knowledge and the built-in information of videos. These two streams use the cross-attention mechanism to share information in different modalities for generating more accurate results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 678, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 678, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 678, + 286, + 713 + ], + "type": "text", + "content": "Vision-and-language representation learning is a hot topic in recent years. ViLBERT [28], LXMERT [46], UNITER [6], UNIMO [25] and Unified-VL [68] learn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "the representations between image and text, while Univl [29], VideoBERT [43], ActBERT [72] and MV-GPT [40] learn the representations between videos and transcripts. Notably, most of these methods attempt to learn powerful vision-and-language representations by pre-training the models on the large-scale datasets, e.g., Howto100M [31] and WebVid-2M [1], and then finetune them on downstream tasks such as video captioning, video-text retrieval and visual question answering. In contrast, our TextKG method uses the speech transcripts as the text to model the visual and linguistic representations and integrate the additional knowledge in knowledge graph to mitigate long-tail words challenge in video captioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 455 + ], + "type": "text", + "content": "Knowledge graph in NLP. Knowledge graph is an useful tool to indicate the real-world entities and their relations, which provides rich structured knowledge facts for language modeling. Large-scale knowledge graphs are used to train knowledge enhanced language models for various natural language processing (NLP) tasks. CoLAKE [44] proposes to inject the knowledge context of an entity, and to jointly learn the contextualized representation for both language and knowledge by a unified structure. ERNIE [63] enhances BERT architecture to better integrate the knowledge information and textual information. KEPLER [52] not only improves language models by integrating factual knowledge, but also generates text-enhanced knowledge representation. JAKET [59] proposes a joint pre-training framework to model knowledge graph and language simultaneously. Inspired by CoLAKE, our method jointly learns the representations of vision, language and knowledge, and enhances the joint visual-language representations by retrieving relevant knowledge in knowledge graphs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 466, + 397, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 397, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 397, + 479 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "As mentioned above, our method aims to integrate additional knowledge using knowledge graph and exploits the multi-modality information in videos to mitigate the long-tail words challenge in the field of video captioning. We design a text with knowledge graph augmented transformer, i.e., a two-stream transformer formed by the external and internal streams. Both streams are constructed by " + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": " sets of alternately stacked self-attention and cross-attention blocks. Specifically, we first use a detector to generate the salient object regions in each video frame, and use the automatic speech recognition (ASR) method [2] to generate the speech transcripts in videos. After that, the class labels of detected salient objects are used to retrieve the prior knowledge in knowledge graphs. Then, the appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior knowledge, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent words of the predicted captions. The overall architecture of the proposed TextKG is shown in Figure 1." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18942" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 534, + 314 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 534, + 314 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 534, + 314 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 534, + 314 + ], + "type": "image", + "image_path": "6ce9d9fb5365b87c9942daa62558c558037a08a6e7a75d9ff93e0c0f3387b6e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 320, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 320, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 546, + 344 + ], + "type": "text", + "content": "Figure 1. The architecture of our TextKG method, which is formed the external and internal streams. Each stream is stacked with " + }, + { + "bbox": [ + 46, + 320, + 546, + 344 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 320, + 546, + 344 + ], + "type": "text", + "content": " sets of multi-modality self-attention and cross-attention modules. The cross attention modules are designed to align tokens in different modalities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 355, + 189, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 189, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 189, + 366 + ], + "type": "text", + "content": "3.1. Two-Stream Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 421 + ], + "type": "text", + "content": "As described above, our two-stream transformer is formed by the external stream and internal stream. Several self-attention and cross-attention blocks are interleaved to construct the two streams." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "\\{y_{1}, y_{2}, \\dots, y_{l}\\}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " be the predicted captions, where " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " is the index of the " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": "-th word in the dictionary, " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "\\{p_{1}, p_{2}, \\dots, p_{l}\\}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " to be the predicted probabilities of words in the dictionary, where " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " is the probability of the " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": "-th word. The index of the " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": "-th word " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "y_{i} = \\arg \\max_{i} p_{i}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": ". Meanwhile, let " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "z_{i}^{\\mathrm{ext}}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "z_{i}^{\\mathrm{int}}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " to be the output probabilities of the external and internal streams of the " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": "-th word. Thus, we have " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "p_{i} = \\omega_{1} z_{i}^{\\mathrm{ext}} + \\omega_{2} z_{i}^{\\mathrm{int}}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "\\omega_{1}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "\\omega_{2}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " are the hyperparameters used to balance the outputs of the two-streams. In this way, " + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 47, + 422, + 288, + 542 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 548, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 548, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 100, + 548, + 287, + 563 + ], + "type": "interline_equation", + "content": "y _ {i} = \\arg \\max _ {i} \\left(\\omega_ {1} z _ {i} ^ {\\text {e x t}} + \\omega_ {2} z _ {i} ^ {\\text {i n t}}\\right). \\tag {1}", + "image_path": "7f8d31bbaeba8a5954645361e0ab9f59fd4813d69546f825df1a10b063260122.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 592 + ], + "type": "text", + "content": "In the following sections, we will describe each module in our two-stream transformer in more details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Multi-modality Self-Attention Module. As shown in Figure 1, we use the self-attention module [49] in both the external and internal streams to model the interactions among multi-modality information. Specifically, for the external stream, the concatenation of feature embeddings of detected objects " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", retrieved prior knowledge " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{k}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", speech transcripts " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{s}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", and predicted video captions " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " are fed into the self-attention module to model the interactions, i.e., " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{ext}} = \\mathrm{Concat}(\\mathcal{F}_{\\mathrm{r}}, \\mathcal{F}_{\\mathrm{k}}, \\mathcal{F}_{\\mathrm{s}}, \\mathcal{F}_{\\mathrm{c}})" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". Meanwhile, for the internal stream, the concatenation of feature embeddings of speech" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "text", + "content": "transcripts " + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{s}}" + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "text", + "content": " , predicted video captions " + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{c}}" + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "text", + "content": " , and video frame " + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{v}}" + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "text", + "content": " are fed into the self-attention model to exploit the interactions, i.e., " + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{int}} = \\operatorname {Concat}(\\mathcal{F}_{\\mathrm{s}},\\mathcal{F}_{\\mathrm{c}},\\mathcal{F}_{\\mathrm{v}})" + }, + { + "bbox": [ + 305, + 355, + 545, + 403 + ], + "type": "text", + "content": " . The self-attention module computes the interactions as follows," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 346, + 411, + 545, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 411, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 346, + 411, + 545, + 431 + ], + "type": "interline_equation", + "content": "\\Phi (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} + M\\right) V, \\tag {2}", + "image_path": "7adfaf7ed232081262095aebe95d398a217ba621dca953b164b3c1d259c56037.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " is the feature dimension of queries " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ". We have " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "Q = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{Q}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "K = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{K}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "V = \\mathcal{X}_{(\\cdot)}W^{\\mathrm{V}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{Q}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{K}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{V}}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " are learnable parameters, and " + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{(\\cdot)} \\in \\{\\mathcal{X}_{\\mathrm{ext}}, \\mathcal{X}_{\\mathrm{int}}\\}" + }, + { + "bbox": [ + 304, + 437, + 545, + 498 + ], + "type": "text", + "content": " are the concatenated feature embeddings for the external and internal streams, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "Following [21], we introduce a mask matrix " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " in attention function (2) of both the external and internal streams to prevent the model from seeing future words. That is, for the " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "-th caption token, we set " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "M_{i,j} = 0" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "j = 1,\\dots ,i" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "M_{i,j} = -\\inf" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "j > i" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": ". Meanwhile, for the external stream, we set " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "M_{i,j} = -\\inf" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " to prevent the associations between irrelevant retrieved prior knowledge and the detected salient object regions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Multi-modality Cross-Attention Module. Besides self-attention module, we use the cross-attention module to align the interactions modeled by both the external and internal streams. Specifically, we compute the affinity matrix to guide the alignments between the modeled interactions by injecting the information. Similar to the self-attention module, we also introduce the mask matrix into the attention function to compute the cross-attention. Intuitively, the retrieved prior knowledge should not have direct effect on the predicted captions, but through the detected salient objects." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18943" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "Thus, we set the corresponding elements to " + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "-\\inf" + }, + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": " to prevent the retrieved prior knowledge directly influencing the predicted captions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 114, + 133, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 133, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 133, + 128 + ], + "type": "text", + "content": "3.2. Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": "The cross entropy loss function is used to guide the training of our TextKG method. Given the ground-truth indices of previous " + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": " words and the concatenated of two-streams " + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{ext}}" + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": ", we can get the predictions of the current " + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": "-th word. After that, the training loss of our method is computed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 211, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 211, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 86, + 211, + 287, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\sum_ {i = 1} ^ {l} \\left(\\lambda_ {1} \\log z _ {i} ^ {\\text {e x t}} + \\lambda_ {2} \\log z _ {i} ^ {\\text {i n t}}\\right), \\tag {3}", + "image_path": "8b0874736d7f12416b2ce6011074d731818403bcf40593f51ca1723e4c632628.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{i}^{\\mathrm{ext}}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{i}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": "-th output word of the external and internal streams, " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " are the preset parameters used to balance the two streams, and " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " is the total length of predicted captions. The Adam optimization algorithm [17] with an initial learning rate of " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " weight decay of 0.01 is used to train the model. The Warmup strategy is adopted in training phase by linearly increasing the learning rate from 0 to the initial learning rate for the first " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " of training epochs, and linearly decreasing to 0 for the remaining " + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 234, + 287, + 354 + ], + "type": "text", + "content": " of epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 360, + 162, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 162, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 162, + 372 + ], + "type": "text", + "content": "3.3. Multimodal Tokens" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 378, + 286, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 450 + ], + "type": "text", + "content": "The appearance embeddings of detected salient objects and video frames, and embeddings of retrieved prior, speech transcripts, and predicted captions are fed into the two-stream transformer to generate subsequent captions after tokenization. We will describe the feature extraction and tokenization processes of the aforementioned embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 450, + 286, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 594 + ], + "type": "text", + "content": "Video tokens encode the appearance and motion information of video frames. Specifically, the aligned appearance and motion features form the video tokens. In each second of the video, we sample 2 frames to extract features. Following [21], we use the \"Flatten-673\" layer in ResNet200 [12] to extract appearance features, and the \"global pool\" layer in BN-Inception [16] to extract the motion features on the YouCookII [69] and ActivityNet Captions [19] datasets. Meanwhile, for fair comparison, we use the InceptionResNetV2 [45] and C3D [11] models to extract appearance and motion features on the MSR-VTT [56] and MSVD [3] datasets, following the methods [58, 64, 66]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "content": "Region tokens are used to describe the visual information of the detected salient objects. We use the Faster R-CNN method [37] to extract visual features for detected object regions, which is pretrained on the Visual Genome dataset [20], and the " + }, + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 594, + 286, + 665 + ], + "type": "text", + "content": " detected objects with highest confidence score in each frame are reserved for video captioning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Transcript tokens are used to encode the content of speech transcripts, which contains some key information in videos. The GloVe model [34] is used to extract features of each word in transcripts to form the transcript tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "Caption tokens encode the information of predicted captions. Similar to transcript tokens, GloVe [34] is adopted to extract features of each word in the generated captions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": "Knowledge tokens model the content of retrieved knowledge, which is important to mitigate the long-tail words challenge in video captioning. Our TextKG method uses a triple structure to encode the retrieved knowledge, formed by two object items and the relations between them, i.e., " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "(\\alpha_{\\mathrm{head}},\\alpha_{\\mathrm{tail}},\\rho_{\\mathrm{rel}})" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{head}}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{tail}}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": " are the two nodes in knowledge graph, and " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{rel}}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": " indicates the edge encoding the relations between them. For example, " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{head}} = \"\\mathrm{knife}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{tail}} = \"\\mathrm{hard}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{rel}} = \"\\mathrm{has~property}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": ". For each detected object " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{head}}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": ", we aim to integrate prior context information for more accurate caption results. That is, we use the GloVe model to extract linguistic features of the node " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{tail}}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": " and set the learnable embeddings for the edge " + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathrm{rel}}" + }, + { + "bbox": [ + 305, + 109, + 545, + 289 + ], + "type": "text", + "content": ". After that, we compute the summation of them to get the knowledge embeddings of the knowledge tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 289, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 517 + ], + "type": "text", + "content": "For each detected object, we use its category name to retrieve knowledge (i.e., triples containing the category name of the detected object) from the knowledge graph. Notably, only a portion of the retrieved knowledge are highly related to the video. The rests are potentially harmful to the performance of caption generation as it may distract the training and inference and also increase computational cost. For example, to generate caption of a cooking tutorial video, the knowledge \"knife - used to -cut\" is more useful rather than \"knife - has property - hard\". Because cutting food may be an important procedure for food preparation. Thus, we design a primary rank mechanism to sort the retrieved knowledge based on their semantic similarity to the video. The cosine similarity between the embeddings of knowledge tokens and video transcripts is computed as the semantic similarity for ranking. The pretrained SBERT [36] model is used to extract feature embeddings of transcripts. Finally, " + }, + { + "bbox": [ + 305, + 289, + 545, + 517 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{k}}" + }, + { + "bbox": [ + 305, + 289, + 545, + 517 + ], + "type": "text", + "content": " knowledge items with the highest similarity scores are reserved for video captioning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 526, + 484, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 484, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 484, + 539 + ], + "type": "text", + "content": "3.4. Knowledge Graphs Construction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 545, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 665 + ], + "type": "text", + "content": "As mentioned above, we use the knowledge graph to retain the prior knowledge. The nodes in the graph could correspond to a noun, an adjective, a verb or an adverb, etc. The edges in the graph are used to encode the relations between different nodes. Apparently, it is important to include all key information of videos in knowledge graphs for accurate caption results. Thus, besides general knowledge graph covering most key knowledge in general scenarios, we also introduce the specific knowledge graph for each dataset covering the key knowledge in specific scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "General knowledge graph is designed to include most key information in general scenarios that we are interested, such as cooking and activity. Specifically, the general knowledge graph is exported from the public available giant knowledge" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18944" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "graph ConceptNet [42] by extracting key words1 in ConceptNet with the connected edges and neighboring nodes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "content": "Specific knowledge graph. Besides the general knowledge graph, we also construct the specific knowledge graph to cover key information in specific scenarios. We believe that the speech transcripts of videos contain most of the crucial information and use the speech transcripts as the source to construct the specific knowledge graph. We first use the automatic speech recognition (ASR) model [2] to convert the speech in all videos to transcripts. After that, we use the Stanford NLP library² to analyze the components in each sentence in transcripts to form a structured grammar tree. We collect the \"adjective and noun\", \"noun and noun\" and \"adverb and verb\" phrases from the grammar trees to construct the specific knowledge graph." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 263, + 127, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 127, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 127, + 276 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 283, + 109, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 109, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 109, + 294 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 301, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 286, + 361 + ], + "type": "text", + "content": "YoucookII contains 2,000 long untrimmed videos from 89 cooking recipes, including 1,333 videos for training, 457 videos for testing. The average length of videos is 5.3 minutes and each video is divided into 7.7 video clips on average with the manually labeled caption sentences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 286, + 468 + ], + "type": "text", + "content": "ActivityNet Captions is a large-scale challenging dataset for video captioning, with 10,024 videos for training, 4,926 for validation, and 5,044 for testing. The average length of each video is 2.1 minutes and is divided into 3.65 clips on average. Each clip is assigned with a manually labeled caption sentence. Since the testing set is not publicly available, following MART [21], we split the validation set into two subsets, i.e., ae-val with 2,460 videos for validation and aetest with 2,457 videos for testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 469, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 286, + 529 + ], + "type": "text", + "content": "MSR-VTT includes 10,000 video clips with 15 seconds on average. Each video clip contains 20 human annotated captions and the average length of each sentence is 9.28 words. Following [5,38,58], 6,513,497, and 2,990 videos are used for training, validation, and testing, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 529, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 589 + ], + "type": "text", + "content": "MSVD consists of 1970 video clips collected from YouTube and each clip is approximately 10 seconds long. Each clip is labeled with 40 sentences. Following [5,38,58], we split the dataset into 1,200,100,and 670 video clips for training, validation and testing, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 597, + 158, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 158, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 158, + 609 + ], + "type": "text", + "content": "4.2. Evaluation metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 616, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 675 + ], + "type": "text", + "content": "To compare the performance of the proposed method with other methods, five model-free automatic evaluation metrics are used, including BLEU@4 [33] (abbreviated to B) that are precision-based metric, METEOR [8] (abbreviated to M) that computes sentence-level scores, CIDEr [50]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 239 + ], + "type": "text", + "content": "(abbreviated to C) that is consensus-based metric, ROUGE [26] (abbreviated to R) that uses longest common subsequence to compute the similarities between sentences, and Rep@4 [54] (abbreviated to Rep) that computes the redundancy score for every captions. Lower Rep indicates better model. The CIDEr score is used as the primary metric in evaluation. Notably, both micro-level and paragraph-level evaluation modes are used. The micro-level mode evaluates the predicted captions of each video clip independently, while the paragraph-level mode considers the relations between captions of different clips. Following [9, 21], the paragraph-level evaluation mode is used in the YouCookII and ActivityNet datasets, which contains multiple clips in each video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 251, + 438, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 438, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 438, + 262 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 270, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 425 + ], + "type": "text", + "content": "We sample every 2 frames per second in each video and use the pre-trained models to extract video features. For the YouCookII and ActivityNet Captions datasets, similar to [21], the ResNet-200 and BN-Inception models are used to extract the appearance and motion features, respectively. Following [58, 64, 66], the InceptionResNetV2 and C3D models are used to extract the appearance and motion features for the MSR-VTT and MSVD dataset, respectively. Meanwhile, the Faster R-CNN method using ResNet-101 as the backbone is adopted to detect object in each video frame, which is trained on the Visual Genome dataset [20]. Notably, only " + }, + { + "bbox": [ + 304, + 270, + 545, + 425 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{r}} = 6" + }, + { + "bbox": [ + 304, + 270, + 545, + 425 + ], + "type": "text", + "content": " classes of the detected objects with top scores are used to form region features in each frame." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "content": "The caption sentences, speech transcripts, and retrieved knowledge in knowledge graph are split into individual words using the NLTK toolkit3, and the GloVe [34] algorithm is used to extract the word embeddings for each word to construct the transcript, caption, and knowledge tokens. The dimension of word embeddings is set to 300. The captions are truncated to 20 words and the maximum text length of transcripts is set to 300 words. For each detected object, we retrieve up to " + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "inline_equation", + "content": "N_{k} = 5" + }, + { + "bbox": [ + 304, + 426, + 545, + 545 + ], + "type": "text", + "content": " pieces of knowledge items in knowledge graph depending on the relevant scores." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "text", + "content": "Notably, the proposed method is implemented using Pytorch. 3 blocks of multi-modality self-attention and cross-attention modules are used in the MSR-VTT [56] and MSVD [3] datasets, and 2 are used in the YouCookII [69] and ActivityNet Captions [19] datasets, to form the two-stream transformer, respectively. The feature dimension in each block is set to 768, and the number of heads in multi-head architecture is set to 12. The proposed method is trained using the Adam algorithm [17] with the initial learning rate of " + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "text", + "content": ". The batch size in training phase is set to 6. The hyper-parameters " + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\omega_{1},\\omega_{2},\\lambda_{1}" + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 548, + 545, + 690 + ], + "type": "text", + "content": " are set to 0.8, 0.2, 0.5 and 0.5, empirically." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 286, + 713 + ], + "type": "text", + "content": "1The key words are exported from the speech transcripts and annotated ground-truth captions in the training sets of the caption datasets. 2https://stanfordnlp.github.io/CoreNLP/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 702, + 410, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 410, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 410, + 712 + ], + "type": "text", + "content": "3https://github.com/nltk/nltk" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18945" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 183 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 183 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 183 + ], + "type": "table", + "html": "
MethodBMCRep
Van-Trans [70]7.615.732.37.8
Trans-XL [7]6.614.826.46.3
Trans-XLRG [22]6.614.725.96.0
MART [21]8.016.035.74.4
Van-Trans+COOT [9]11.119.855.65.7
COOT [9]11.319.957.26.7
TextKG14.022.175.92.8
Human---1.0
", + "image_path": "7f94f5089420798a9e46cb1f641c5af26bde2ba0880124b063499016cc0c688d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 224, + 275, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 275, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 275, + 237 + ], + "type": "text", + "content": "4.4. Comparison to the State of the Art Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 422 + ], + "type": "text", + "content": "YouCookII dataset. To validate the effectiveness of the proposed method, we compare it to the state-of-the-art (SOTA) methods using the paragraph-level evaluation mode, reported in Table 1. As shown in Table 1, our TextKG method achieves the best results on the YouCookII dataset. Specifically, TextKG improves " + }, + { + "bbox": [ + 46, + 243, + 287, + 422 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 46, + 243, + 287, + 422 + ], + "type": "text", + "content": " absolute CIDEr score compared to the SOTA method COOT [9]. COOT [9] focuses on leveraging the hierarchy information and modeling the interactions between different levels granularity and different modalities, such as frames and words, clip and sentences, and videos and paragraphs, which is pre-trained on the large-scale HowTo100M [31] dataset. In contrast, our TextKG method aims to exploit additional knowledge in knowledge graph and multi-modality information in videos to improve the caption results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 544 + ], + "type": "text", + "content": "Beside the paragraph-level evaluation mode, we also report the evaluation results based on the micro-level mode on the YouCookII dataset in Table 2. As shown in Table 2, TextKG significantly outperforms other methods, including Univl [29] and AT [13], both of which use transcript information to enhance model performance by directly concatenating visual and transcript embedding. We argue that with the use of knowledge graphs, TextKG gives a better understanding of the content included in the speech transcripts and thus yields favorable results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 700 + ], + "type": "text", + "content": "ActivityNet Caption dataset. We also evaluate our TextKG method on the challenging ActivityNet Caption dataset in Table 3. As shown in Table 3, TextKG achieves the best results on 3 out of 4 metrics, i.e., BLEU, METOR, and CIDEr. The ActivityNet Caption dataset [19] contains a series of complex events with annotated caption sentences describing the events that occur. These events may occur over various periods of time and may co-occur. In contrast to existing methods [5, 38, 58, 66], our TextKG understands complex events more comprehensively with the help of exploiting information from transcripts and integrating the relevant knowledge in knowledge graph to enhance the common-sense information for video captioning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": "MSR-VTT dataset. The evaluation results on the MSR-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 233 + ], + "blocks": [ + { + "bbox": [ + 46, + 191, + 287, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 213 + ], + "type": "text", + "content": "Table 1. Evaluation results on the YouCookII val subset in the paragraph-level evaluation mode." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 233 + ], + "type": "table", + "html": "
MethodInputBMRC
Masked Trans [70]V3.811.627.438
S3D [53]V3.29.526.131
VideoAsMT [18]V5.313.4--
SwinBERT [27]V915.637.3109
VideoBERT [43]V4.011.027.549
VideoBERT+S3D [43]V4.311.928.850
ActBERT [72]V5.414.330.665
AT [13]V+S9.017.836.7112
DPC [41]V+S2.818.1--
VALUE [24]V+S12.418.840.4130
Univl [29]V+S9.516.337.4115
MV-GPT [40]V+S13.317.635.5103
TextKGV+S11.718.440.2133
", + "image_path": "9865f3ba52f68c5e4e28cb09a42691235fcd28f3c0eae356a791fe02b7c9657a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 297, + 547, + 409 + ], + "blocks": [ + { + "bbox": [ + 305, + 240, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 305, + 240, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 546, + 285 + ], + "type": "text", + "content": "Table 2. Evaluation results on the YouCookII val subset in the micro-level evaluation mode. 'V' indicates the methods use video appearance information, and 'S' indicates the methods use the speech information." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 297, + 547, + 409 + ], + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 409 + ], + "type": "table", + "html": "
MethodBMCRep
HSE [60]9.813.818.813.2
GVD [67]11.015.722.08.8
GVDsup [67]11.316.422.97.0
Van-Trans [70]9.815.622.27.8
Trans-XL [7]10.415.121.78.5
Trans-XLRG [22]10.214.820.48.9
MART [21]10.315.723.45.2
TextKG11.316.526.66.3
", + "image_path": "5d72b6cdb96c3ea95823d5ba3249236f1495df23c344f11537e468fbc30bc0e5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 417, + 546, + 440 + ], + "lines": [ + { + "bbox": [ + 305, + 417, + 546, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 546, + 440 + ], + "type": "text", + "content": "Table 3. Evaluation results on the Activity-Net Captions ae-val subset in the paragraph-level evaluation mode." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "VTT dataset are reported in Table 4. For the fair comparison, we use the same visual features as the existing methods [5, 58, 64, 66]. In contrast to existing methods focusing on network architecture design to exploit visual information, our TextKG aims to exploit mutli-modality information in external knowledge graph and original videos, achieving the state-of-the-art performance on 3 out of 4 metrics, i.e., BLEU, METEOR, and CIDEr." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "content": "Meanwhile, we also compare the TextKG method to the SOTA methods focusing on multi-modality pretraining models in Table 5. As shown in Table 5, our TextKG method performs favorably against existing methods, such as DECEM [48], UniVL [29] and MV-GPT [40], that are pretrained on the large-scale datasets (e.g., Howto100M [31]), i.e., improve 0.7 absolute CIDEr score compared to the SOTA method LAVENDER [23]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "MSVD dataset. We also evaluate the proposed TextKG method on the MSVD dataset [3] that doesn't have speech and repoprt the results in Table 6. As shown in Table 6, our TextKG achieves the best results without speech transcripts by improving " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": " CIDEr scores compared to the SOTA method HMN [58] and the JCRR method [14]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18946" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 282, + 188 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 188 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 188 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 188 + ], + "type": "table", + "html": "
MethodBMRC
OA-BTG [61]41.428.2-46.9
POS-CG [51]4228.261.648.7
MGSA [4]42.427.6-47.5
STG-KD [32]40.528.360.947.1
ORG-TRL [64]43.628.862.150.9
SGN [38]40.828.360.849.5
MGRMP [5]41.728.962.151.4
HMN [58]43.52962.751.5
TextKG43.729.662.452.4
", + "image_path": "93b03b25980df951812cd658f0a66a306d241a1fe52e421370f314d396406a4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 224, + 282, + 331 + ], + "blocks": [ + { + "bbox": [ + 47, + 198, + 286, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 286, + 218 + ], + "type": "text", + "content": "Table 4. Evaluation results on the MSR-VTT test subset in the micro-level evaluation mode." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 224, + 282, + 331 + ], + "lines": [ + { + "bbox": [ + 52, + 224, + 282, + 331 + ], + "spans": [ + { + "bbox": [ + 52, + 224, + 282, + 331 + ], + "type": "table", + "html": "
MethodInputFeaturesBMRC
SWINBERT [27]VVidSwin45.430.664.155.9
CLIP4C [47]VCLIP46.130.763.757.7
CMVC [57]VCLIP48.231.364.858.7
LAVENDER [23]VVidSwin---60.1
DeCEM [48]V+SBERT45.229.764.752.3
UniVL [29]V+SS3D41.828.960.850.0
MV-GPT [40]V+SViViT48.9*38.7*64.060.0
TextKG (CLIP)V+SCLIP46.630.564.860.8
", + "image_path": "2ea450eb079dc50f1e8e6786efe961d5517f5c689a21714664bac70adb6bc5f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 430, + 285, + 558 + ], + "blocks": [ + { + "bbox": [ + 46, + 339, + 286, + 416 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 286, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 286, + 416 + ], + "type": "text", + "content": "Table 5. Comparison to the methods focusing on model pretraining on the MSR-VTT dataset in the micro-level evaluation mode. Notably, following CLIP4C, we use the pre-trained CLIP model on LAION-400M [39] to extract video appearance features. * The authors use a different library to compute BLEU and METOR. Thus, the results on BLEU and METOR are not directly comparable to other methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 430, + 285, + 558 + ], + "lines": [ + { + "bbox": [ + 50, + 430, + 285, + 558 + ], + "spans": [ + { + "bbox": [ + 50, + 430, + 285, + 558 + ], + "type": "table", + "html": "
MethodBMRC
OA-BTG [61]56.936.2-90.6
POS-CG [51]52.534.171.388.7
MGSA [4]53.435-86.7
STG-KD [32]52.236.973.993.0
ORG-TRL [64]54.336.473.995.2
SGN [38]52.835.572.994.3
MGRMP [5]55.836.974.598.5
JCRR [14]57.036.8-96.8
HMN [58]59.237.775.1104.0
TextKG60.838.575.1105.2
", + "image_path": "61c12f5bc46d00b274e3f2b5aae6009b132b41f57c917518c7b90abe8750a3c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 612, + 286, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 612, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 612, + 286, + 648 + ], + "type": "text", + "content": "exploiting knowledge graph to model the relations between objects, demonstrating the effectiveness of the knowledge graph usage in our method for video captioning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 138, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 138, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 138, + 671 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Influence of different modules. To validate the effectiveness of different modules in TextKG, we conduct several experiments on the YouCookII dataset in Table 7. As" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 567, + 286, + 588 + ], + "lines": [ + { + "bbox": [ + 47, + 567, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 286, + 588 + ], + "type": "text", + "content": "Table 6. Evaluation results on the MSVD test subset in the microlevel evaluation mode." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 543, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 205 + ], + "type": "table", + "html": "
V-FR-FTextG-KGS-KGK-SBMCRep
✓7.415.732.14.1
✓✓9.517.745.95.2
✓✓✓✓9.717.848.94.2
✓✓✓✓9.718.048.53.3
✓✓✓✓✓9.617.749.85.5
✓✓13.021.262.52.7
✓✓✓13.922.171.32.9
✓✓✓✓✓13.722.073.52.0
✓✓✓✓✓13.521.974.82.1
✓✓✓✓✓13.721.872.02.8
✓✓✓✓✓✓14.022.175.92.8
", + "image_path": "b5d7130fab0d7a9e13da8534f77b15307d9353059c7e4e1b504749b89e62bd6c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 278 + ], + "type": "text", + "content": "Table 7. Ablation study on the YouCookII val subset in the paragraph-level evaluation mode. \"V-F\" and \"R-F\" indicate the video and region features, \"Text\" indicates the speech transcripts features, \"G-KG\" and \"S-KG\" indicate general and specific knowledge graph, and \"K-S\" indicates the knowledge selection mechanism." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 289, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 491 + ], + "type": "text", + "content": "shown in Table 7, the baseline method considers the information of video features and predicted captions, achieving 32.1 CIDEr score. After integrating the speech transcripts, the CIDEr score is improved to 62.5, which ensures the key information of videos is considered in generating video captions. After integrating the region features, TextKG achieves 71.3 CIDEr score. The CIDEr score can be further improved 4.6 CIDEr score to 75.9 by exploiting additional knowledge in knowledge graph. Notably, even without the information of speech transcripts, the knowledge graphs are still crucial for video captioning, i.e., improving 3.9 CIDEr score (49.8 vs. 45.9). Meanwhile, we also demonstrate the effectiveness of knowledge selection mechanism in Table 7. As shown in Table 7, without the knowledge selection mechanism, the CIDEr score significantly drops 3.9 to 72.0. We believe that the noisy knowledge is harmful to the accuracy of video captions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 493, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 660 + ], + "type": "text", + "content": "Influence of transcript quality. To explore the effect of transcript quality on the knowledge graph, we add some noises to the speech transcripts (i.e., mask some amount of speech transcripts) and evaluate the performance of our models on the YouCookII dataset in Figure 3. We find that the method using both the general and specific knowledge graphs performs favorable against other methods with different degrees of masked speech transcripts. Meanwhile, as the amount of masked speech transcripts increasing, the accuracy of the method using specific knowledge graph drops sharply, while the accuracy of the method using general knowledge graph drops at a much slower pace. The aforementioned results demonstrate the importance of general knowledge graph for the video captioning task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 670, + 417, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 417, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 417, + 682 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We present the qualitative results of the proposed method on the YouCookII dataset in Figure 2, including the key" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18947" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 72, + 516, + 300 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 516, + 300 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 516, + 300 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 516, + 300 + ], + "type": "image", + "image_path": "ae5f712b8f6985f028f0661f11d732f3f0808acc8f826d1b89a9aa421a3dd5b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 310, + 546, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 546, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 546, + 333 + ], + "type": "text", + "content": "Figure 2. Qualitative results of the proposed method on the YouCookII dataset. The information in the speech transcripts and knowledge graph is not considered in the baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 357, + 279, + 526 + ], + "blocks": [ + { + "bbox": [ + 56, + 357, + 279, + 526 + ], + "lines": [ + { + "bbox": [ + 56, + 357, + 279, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 279, + 526 + ], + "type": "image", + "image_path": "1e507a2235cb0683d733417660532c4da6f9607b49896c50c532d291e2a7622e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 538, + 287, + 625 + ], + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 625 + ], + "type": "text", + "content": "Figure 3. Influence of the transcripts qualities. \"w/o KG\" indicates that the method does not use the general and specific knowledge graphs. \"with G-KG\" and \"with S-KG\" indicate that only the general and specific knowledge graphs are used, respectively. \"with G- and S-KG\" indicates that both the general and specific knowledge graphs are used. The CIDEr differences between the \"with G-KG\", \"with S-KG\", \"with G- and S-KG\" methods and the \"w/o KG\" method are shown at the bottom corner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "frames, detected salient objects, speech transcripts, retrieved knowledge items, caption results predicted by our model, ground-truth captions, and key words distributions in the training set. As shown in Figure 2, we observe that the transcript contains key information in the video, e.g., egg, salt, pepper, com starch and flour, to ensure the qual" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "type": "text", + "content": "ity of the generated video captions. However, the model with only the speech transcripts fails to predict the long-tail phrases, such as sesame oil and black pepper. With the help of knowledge graph providing the relations among these phrases, e.g., sesame oil, black pepper and corn starch, our TextKG model can predict the phrases more accurately. This indicates that the knowledge graph is crucial to mitigate the long-tail word challenges in video captioning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 378, + 477 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "content": "In this paper, we present a text with knowledge graph augmented transformer for video captioning, which aims to integrate external knowledge in knowledge graph and exploit the multi-modality information in video to mitigate long-tail words challenge. Extensive experiments conducted on four challenging datasets demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 571, + 546, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 546, + 655 + ], + "type": "text", + "content": "In the future, we plan to improve the proposed method in two directions, i.e., (1) optimizing the knowledge retrieve strategy by considering semantic context information of detected objects and corresponding actions in videos; (2) constructing multi-modal knowledge graph (e.g., the nodes in knowledge graph formed by text, speech, and images or videos) to improve the accuracy of video captioning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 670, + 365, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 365, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 365, + 682 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 546, + 713 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18948" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "end-to-end retrieval. In IEEE/CVF International Conference on Computer Vision, pages 1708-1718, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 118 + ], + "type": "text", + "content": "[2] William Chan, Navdeep Jaitly, Quoc V. Le, and Oriol Vinyals. Listen, attend and spell. 2015. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 119, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 119, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 119, + 288, + 173 + ], + "type": "text", + "content": "[3] David L. Chen and William B. Dolan. Collecting highly parallel data for paraphrase evaluation. In Dekang Lin, Yuji Matsumoto, and Rada Mihalcea, editors, Annual Meeting of the Association for Computational Linguistics, 2011. 2, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 287, + 217 + ], + "type": "text", + "content": "[4] Shaoxiang Chen and Yu-Gang Jiang. Motion guided spatial attention for video captioning. In Association for the Advancement of Artificial Intelligence, pages 8191-8198, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 262 + ], + "type": "text", + "content": "[5] Shaoxiang Chen and Yu-Gang Jiang. Motion guided region message passing for video captioning. In IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 287, + 308 + ], + "type": "text", + "content": "[6] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. UNITER: learning universal image-text representations. CoRR, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 309, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 287, + 365 + ], + "type": "text", + "content": "[7] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc Viet Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Annual Meeting of the Association for Computational Linguistics, pages 2978-2988, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 287, + 409 + ], + "type": "text", + "content": "[8] Michael J. Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 376-380, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 287, + 454 + ], + "type": "text", + "content": "[9] Simon Ging, Mohammadreza Zolfaghari, Hamed Pirsiavash, and Thomas Brox. COOT: cooperative hierarchical transformer for video-text representation learning. In Advances in Neural Information Processing Systems, 2020. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[10] Xin Gu, Hanhua Ye, Guang Chen, Yufei Wang, Libo Zhang, and Longyin Wen. Dual-stream transformer for generic event boundary captioning. 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "text", + "content": "[11] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6546-6555, 2018. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[13] Jack Hessel, Bo Pang, Zhenhai Zhu, and Radu Soricut. A case study on combining ASR and visual features for generating instructional video captions. In Conference on Computational Natural Language Learning, pages 419-429, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[14] Jingyi Hou, Xinxiao Wu, Xiaoxun Zhang, Yayun Qi, Yunde Jia, and Jiebo Luo. Joint commonsense and relation reasoning for image and video captioning. In Association for the Advancement of Artificial Intelligence, pages 10973-10980, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[15] Feicheng Huang, Zhixin Li, Shengjia Chen, Canlong Zhang, and Huifang Ma. Image captioning with internal and external" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "text", + "content": "knowledge. In CIKM '20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020, pages 535-544. ACM, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, International Conference on Machine Learning, pages 448-456, 2015. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 172, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 215 + ], + "type": "text", + "content": "[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, International Conference on Learning Representations, 2015. 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 248 + ], + "type": "text", + "content": "[18] Bruno Korbar, Fabio Petroni, Rohit Girdhar, and Lorenzo Torresani. Video understanding as machine translation. CoRR, abs/2006.07203, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "text", + "content": "[19] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In IEEE/CVF International Conference on Computer Vision, pages 706–715, 2017. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "type": "text", + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123(1):32-73, 2017. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 415 + ], + "type": "text", + "content": "[21] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 470 + ], + "type": "text", + "content": "[22] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L. Berg, and Mohit Bansal. MART: memory-augmented recurrent transformer for coherent video paragraph captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2603-2614, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[23] Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, and Lijuan Wang. LAVENDER: unifying video-language understanding as masked language modeling. 2022. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 612 + ], + "type": "text", + "content": "[24] Linjie Li, Jie Lei, Zhe Gan, Licheng Yu, Yen-Chun Chen, Rohit Pillai, Yu Cheng, Luowei Zhou, Xin Wang, William Yang Wang, Tamara L. Berg, Mohit Bansal, Jingjing Liu, Lijuan Wang, and Zicheng Liu. VALUE: A multitask benchmark for video-and-language understanding evaluation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurIPS Datasets and Benchmarks 2021, December 2021, virtual, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 679 + ], + "type": "text", + "content": "[25] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. UNIMO: towards unified-modal understanding and generation via cross-modal contrastive learning. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors, ACL-IJCNLP, pages 2592-2607, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics, pages 74-81, 2004. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18949" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17928-17937, 2022. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 205 + ], + "type": "text", + "content": "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alché-Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 13–23, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 207, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 287, + 262 + ], + "type": "text", + "content": "[29] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Xilin Chen, and Ming Zhou. Univilm: A unified video and language pre-training model for multimodal understanding and generation. CoRR, abs/2002.06353, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 263, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 328 + ], + "type": "text", + "content": "[30] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[31] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "type": "text", + "content": "[32] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10876, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 454, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 498 + ], + "type": "text", + "content": "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Annual Meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 499, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 544 + ], + "type": "text", + "content": "[34] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing, pages 1532-1543, 2014. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 589 + ], + "type": "text", + "content": "[35] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In IEEE/CVF International Conference on Computer Vision, pages 8907-8916, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[36] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 3980-3990, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[37] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 1137-1149, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[38] Hobin Ryu, Sunghun Kang, Haeyong Kang, and Chang D. Yoo. Semantic grouping network for video captioning. In Association for the Advancement of Artificial Intelligence, pages 2514-2522, 2021. 1, 5, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. CoRR, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 182 + ], + "type": "text", + "content": "[40] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17938-17947, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 184, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 547, + 237 + ], + "type": "text", + "content": "[41] Botian Shi, Lei Ji, Yaobo Liang, Nan Duan, Peng Chen, Zhendong Niu, and Ming Zhou. Dense procedure captioning in narrated instructional videos. In Annual Meeting of the Association for Computational Linguistics, pages 6382-6391, 2019. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 239, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 546, + 282 + ], + "type": "text", + "content": "[42] Robyn Speer, Joshua Chin, and Catherine Havasi. Concept-net 5.5: An open multilingual graph of general knowledge. In Association for the Advancement of Artificial Intelligence, pages 4444-4451, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 283, + 546, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 336 + ], + "type": "text", + "content": "[43] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In IEEE/CVF International Conference on Computer Vision, pages 7463-7472, 2019. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 338, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 546, + 392 + ], + "type": "text", + "content": "[44] Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuanjing Huang, and Zheng Zhang. Colake: Contextualized language and knowledge embedding. In International Conference on Computational Linguistics, pages 3660-3670, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 394, + 546, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 546, + 448 + ], + "type": "text", + "content": "[45] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A. Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Satinder Singh and Shaul Markovitch, editors, Association for the Advancement of Artificial Intelligence, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 449, + 546, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 546, + 492 + ], + "type": "text", + "content": "[46] Hao Tan and Mohit Bansal. LXMERT: learning cross-modality encoder representations from transformers. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors, EMNLP-IJCNLP, pages 5099-5110, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "type": "text", + "content": "[47] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: CLIP for video caption. In ACM International Conference on Multimedia, pages 4858-4862, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "type": "text", + "content": "[48] Zineng Tang, Jie Lei, and Mohit Bansal. December: Learning from noisy instructional videos via dense captions and entropy minimization. In Conference of the North American Chapter of the Association for Computational Linguistics, pages 2415-2426, 2021. 1, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 593, + 546, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 546, + 668 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett, editors, Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "type": "text", + "content": "[50] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18950" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[51] Bairui Wang, Lin Ma, Wei Zhang, Wenhao Jiang, Jingwen Wang, and Wei Liu. Controllable video captioning with POS sequence guidance based on gated fusion network. In IEEE/CVF International Conference on Computer Vision, pages 2641-2650, 2019. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[52] Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. KEPLER: A unified model for knowledge embedding and pre-trained language representation. Transactions of the Association for Computational Linguistics, 9:176-194, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 217 + ], + "type": "text", + "content": "[53] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning for video understanding. CoRR, abs/1712.04851, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[54] Yilei Xiong, Bo Dai, and Dahua Lin. Move forward and tell: A progressive generator of video descriptions. In European Conference on Computer Vision, pages 468-483, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "type": "text", + "content": "[55] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. VLM: task-agnostic video-language model pre-training for video understanding. In Findings of the Association for Computational Linguistics, pages 4227–4239, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "type": "text", + "content": "[56] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5288-5296, 2016. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 420 + ], + "type": "text", + "content": "[57] Bang Yang, Tong Zhang, and Yuexian Zou. CLIP meets video captioning: Concept-aware representation learning does matter. In Pattern Recognition and Computer Vision - 5th Chinese Conference, PRCV 2022, Shenzhen, China, November 4-7, 2022, Proceedings, Part I, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[58] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. pages 17939-17948, 2022. 1, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 510 + ], + "type": "text", + "content": "[59] Donghan Yu, Chenguang Zhu, Yiming Yang, and Michael Zeng. JAKET: joint pre-training of knowledge graph and language understanding. In Association for the Advancement of Artificial Intelligence, pages 11630-11638, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "text", + "content": "[60] Bowen Zhang, Hexiang Hu, and Fei Sha. Cross-modal and hierarchical modeling of video and text. In European Conference on Computer Vision, pages 385-401, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[61] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8327-8336, 2019. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[62] Yu Zhang, Xinyu Shi, Siya Mi, and Xu Yang. Image captioning with transformer and knowledge graph. Pattern Recognition Letters, 143:43-49, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[63] Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. ERNIE: enhanced language representation with informative entities. In Annual Meeting of the Association for Computational Linguistics, pages 1441-1451, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[64] Ziqi Zhang, Yaya Shi, Chunfeng Yuan, Bing Li, Peijin Wang, Weiming Hu, and Zheng-Jun Zha. Object relational graph with teacher-recommended learning for video captioning." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 453 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13275-13285, 2020. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[65] Wentian Zhao, Yao Hu, Heda Wang, Xinxiao Wu, and Jiebo Luo. Boosting entity-aware image captioning with multimodal knowledge graph. CoRR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "text", + "content": "[66] Qi Zheng, Chaoyue Wang, and Dacheng Tao. Syntax-aware action targeting for video captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13093-13102, 2020. 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 218 + ], + "type": "text", + "content": "[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J. Corso, and Marcus Rohrbach. Grounded video description. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6578-6587, 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[68] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pre-training for image captioning and VQA. In Association for the Advancement of Artificial Intelligence, pages 13041-13049, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "text", + "content": "[69] Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In Association for the Advancement of Artificial Intelligence, pages 7590-7598, 2018. 1, 2, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "text", + "content": "[70] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8739-8748, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[71] Yimin Zhou, Yiwei Sun, and Vasant G. Honavar. Improving image captioning by leveraging knowledge graphs. In IEEE Winter Conference on Applications of Computer Vision, pages 283-293, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[72] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8743-8752, 2020. 1, 2, 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18951" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_content_list.json b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db7e065568dc39ab7fcbb166201e7bffa0bc6e80 --- /dev/null +++ b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "Text-guided Unsupervised Latent Transformation for Multi-attribute Image Manipulation", + "text_level": 1, + "bbox": [ + 137, + 128, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiwen Wei $^{1}$ , Zhen Xu $^{1}$ , Cheng Liu $^{2}$ , Si Wu $^{1,3*}$ , Zhiwen Yu $^{1}$ , and Hau San Wong $^{4}$", + "bbox": [ + 161, + 202, + 807, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1School of Computer Science and Engineering, South China University of Technology", + "bbox": [ + 145, + 220, + 823, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, Shantou University", + "bbox": [ + 267, + 239, + 700, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Peng Cheng Laboratory", + "bbox": [ + 385, + 256, + 583, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Computer Science, City University of Hong Kong", + "bbox": [ + 225, + 273, + 745, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{202021044777, csxuzhen}@mail.scut.edu.cn, cliu@stu.edu.cn, {cswusi, zhwyu}@scut.edu.cn, cshswong@cityu.edu.hk", + "bbox": [ + 98, + 294, + 867, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Great progress has been made in StyleGAN-based image editing. To associate with preset attributes, most existing approaches focus on supervised learning for semantically meaningful latent space traversal directions, and each manipulation step is typically determined for an individual attribute. To address this limitation, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model, which adaptively infers a single transformation step in the latent space of StyleGAN to simultaneously manipulate multiple attributes on a given input image. Specifically, we adopt a two-stage architecture for a latent mapping network to break down the transformation process into two manageable steps. Our network first learns a diverse set of semantic directions tailored to an input image, and later nonlinearly fuses the ones associated with the target attributes to infer a residual vector. The resulting tightly interlinked two-stage architecture delivers the flexibility to handle diverse attribute combinations. By leveraging the cross-modal text-image representation of CLIP, we can perform pseudo annotations based on the semantic similarity between preset attribute text descriptions and training images, and further jointly train an auxiliary attribute classifier with the latent mapping network to provide semantic guidance. We perform extensive experiments to demonstrate that the adopted strategies contribute to the superior performance of TUSLT.", + "bbox": [ + 76, + 393, + 472, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 818, + 209, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual attributes represent semantically meaningful features inherent in images, and attribute manipulation has ex", + "bbox": [ + 76, + 844, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/caadc3e34e247be75e033a195370d14e78152b20310c2ffb06ea525daab8656f.jpg", + "image_caption": [ + "Figure 1. Visually comparing TUSLT with StyleFlow (supervised) and StyleCLIP (text-driven) in precisely manipulating multiple attributes and preserving irrelevant attributes." + ], + "image_footnote": [], + "bbox": [ + 521, + 359, + 867, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perceived great improvements, due to the advent of Generative Adversarial Network [13] (GAN)-based generative models, e.g. StyleGAN [21, 22] and StarGAN [7, 8]. Recent works [15, 37, 43] have discovered that the latent space of StyleGAN possesses semantic disentanglement properties, enabling a variety of image editing operations via latent transformations.", + "bbox": [ + 496, + 627, + 890, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "StyleGAN-based methods for image attribute manipulation typically involve a large number of manual annotations or well-trained attribute classifiers. Furthermore, the discovered semantic latent directions are associated with individual attributes. The editing on a target attribute is carried out by moving the latent code of an input image along one of the directions. For $K$ target attributes, these models require $K$ transformation steps to handle the translation. As a result, they are not scalable to the increasing number of target attributes in multi-attribute transformation tasks. As shown in Figure 1, we test a state-of-the-art supervised", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19285", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/381eb510a2cae551df366df077fe0b8c552d8e3c5c58f68c9db239760106fa78.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed model, TUSLT, consisting of two learnable components: an auxiliary attribute classifier $A$ trained on the CLIP-based labeled data, and a latent mapping network $\\{\\Gamma, \\Phi\\}$ . $\\Gamma$ infers latent directions $\\{\\gamma_w^{(1)}, \\dots, \\gamma_w^{(K)}\\}$ for preset attributes, and $\\Phi$ transforms the target-related directions as indicated by mask $M_{\\Delta}$ into a residual vector $\\gamma_w^{\\Delta}$ , to which the initial latent code is added. Precise multi-attribute transfer is allowed by such a single transformation step, and the generator $G$ synthesizes a new image reflecting the target attributes under the guidance of $A$ and CLIP encoders." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 651, + 295 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50d62be75bf706a5d532f66b202e1df4b4dfb8640a0836bb9e726e47b18c9e37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 89, + 893, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model, StyleFlow [2], and find that multiple transformation steps lead to undesired deviation from the input image on irrelevant attributes. Compared to the state-of-the-art text-driven model, StyleCLIP [33], we can also achieve a better manipulation result by seeking a single latent transformation step for the task.", + "bbox": [ + 75, + 417, + 468, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model that supports simultaneous manipulation on multiple attributes. As shown in Figure 2, the key is to jointly learn a mapping network to infer the latent transformation and an auxiliary attribute classifier to assess manipulation quality. We employ the Contrastive Language-Image Pre-training (CLIP) model [34] to generate pseudo-labeled data by measuring the semantic similarities between attribute text descriptions and training images. Compared to CLIP, the jointly trained classifier extracts domain-specific information to better characterize the differences among attributes. This benefits the mapping network to seek more suitable transformations, such that the synthesized images reflect target attributes. Further, we adopt a two-stage architecture for the mapping network: the earlier stage employs a prediction subnetwork to infer a set of semantic directions, and the latter stage operates on the resulting directions and nonlinearly fuses the target-related ones. The intermediate semantic directions are associated with preset attributes and tailored for the input image. This design allows us to deal with a wide range of attribute combinations in a single transformation step. We perform extensive experiments and provide both qualitative and quantitative results in diverse multi-attribute transformation tasks, showing the", + "bbox": [ + 75, + 523, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "superiority of our model over the competing methods.", + "bbox": [ + 500, + 417, + 854, + 433 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this work are given as follows: (a) The existing image editing methods focus on discovering semantic latent directions associated with individual visual attributes, and a sequential manipulation process is thus needed for multi-attribute manipulation. In contrast, the proposed model infers a single step of latent space walk to simultaneously manipulate multiple attributes. (b) Benefiting from the cross-modal text-image representation of CLIP, we jointly train a latent mapping network with an auxiliary attribute classifier, which leads to more precise attribute rendering without requiring additional manual annotations. (c) Due to the two-stage nature, our latent mapping network breaks down the challenging multi-attribute manipulation task into sub-tasks: inferring diverse semantic directions and integrating the target-related ones into a single transformation vector. This design gives our model interpretability and flexibility in dealing with a variety of attribute combinations.", + "bbox": [ + 496, + 434, + 890, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 715, + 640, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Generic Image-to-image Translation", + "text_level": 1, + "bbox": [ + 500, + 741, + 815, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As one of the earliest image translation models, pix2pix [18] learnt a cross-domain mapping via conditional GAN [31]. In addition to an adversarial training loss, the consistency regularization between each input image and the corresponding ground truth was imposed in the training process. To alleviate the problem of pairwise training data, many unpaired image translation models have been developed. UNIT [29] trained two generators to approximate the joint distribution of images from different domains. Cy", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19286", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cle consistency regularization was adopted in CycleGAN [48] and discoGAN [23]. To efficiently learn the mappings among multiple domains, StarGAN [7, 9] was a unified framework in which a single generator was trained to translate an input image into different domains, conditioned on domain information.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Attribute Manipulation", + "text_level": 1, + "bbox": [ + 76, + 191, + 294, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FaderNet [26] used an encoder-decoder architecture and learnt attribute-aware latent representations in an unsupervised manner. To ensure correct manipulation on the desired attributes, an attribute classification constraint was incorporated in AttGAN [16]. Instead of handling all target attributes, STGAN [28] contained a selective transfer module to only encode the changed attributes, based on the d-difference between source and reference images. In IOAGAN [5], the relationship of preset attributes were leveraged via a Graph Convolutional Network (GCN) [25] to infer and inject an integrated embedding of attributes into a translation network.", + "bbox": [ + 75, + 215, + 468, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The latent spaces of GANs demonstrate promising semantic organization [10, 12, 19]. With the progress of high-fidelity GAN inversion [14, 36, 40], there are many works on exploring the latent space of a pre-trained StyleGAN generator for image manipulation [3, 41]. SeFa [39] disentangled semantics from StyleGAN by decomposing the matrix of generator weights. In addition, LEFS [47] applied s-parse representation learning to unsupervised semantic disentanglement. As a representative supervised method, InterFaceGAN [38] searched for semantic latent directions by solving a series of binary classification problems about preset attributes. Instead of latent directions, StyleSpace [43] explored style-associated channels with the help of a pretrained classifier. Different from the above linear latent transformation methods, StyleFlow [2] employed a conditional flow model to learn non-linear paths for attribute manipulation.", + "bbox": [ + 75, + 398, + 472, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Text-guided Image Editing", + "text_level": 1, + "bbox": [ + 76, + 665, + 318, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal representation learning on visual and language data brings about substantive progress in text-guided image synthesis. An early attempt was to train a GAN-based model [31], conditioned on text embedding [35]. Further, a stacked structure [46] and an attention-based regularization approach [45] were designed to improve synthesis quality. In [30, 32], image content from visual attributes were disentangled and associated with text descriptions without using manually annotated data. ManiGAN [27] aimed to learn text and image cross-modality representations, such that semantic regions were associated with the corresponding text via an affine transformation. TediGAN [44] aligned and projected text and image representations into the latent space of StyleGAN. Further, Style", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [33] integrated CLIP [34] with StyleGAN to leverage CLIP-based linguistic-visual semantic consistency regularization for better manipulation quality.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among the aforementioned models, StyleFlow [2] and StyleCLIP [33] are the most relevant to our work. StyleFlow is a supervised method, and the learnt semantic directions are associated with individual attributes. For multi-attribute manipulation, StyleFlow needs a sequential transformation process, in which the target attributes are edited one at a time. In contrast, the proposed model learns to infer a single transformation step to simultaneously manipulate multiple attributes in an unsupervised manner. StyleCLIP trains a dedicated mapping network for each text description, and thus has limited scalability in handling a variety of attribute combinations. Unlike StyleCLIP, our model uses a two-stage mapping network and jointly trains an auxiliary attribute classifier without manual annotations. This structure benefits the scalability and semantic accuracy in changing multiple attributes.", + "bbox": [ + 496, + 136, + 892, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 498, + 388, + 671, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to translate an input image into a new one reflecting the target attributes. Based on the linguistic-visual representation of CLIP, we can pseudo-label training images by measuring the semantic similarities between attribute text descriptions and training images. We believe that an auxiliary attribute classifier can capture the most discriminative information, which complements the role of CLIP encoders to a certain extent. Based on this, the constructed supervision is leveraged by jointly training a latent mapping network and the classifier. The former has a two-stage architecture to infer a diverse set of semantic latent directions followed by selectively integrating target-related ones, and the latter provides semantic guidance. As a result, multi-attribute manipulation can be precisely carried out in a single forward pass.", + "bbox": [ + 496, + 414, + 893, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Auxiliary Attribute Classifier", + "text_level": 1, + "bbox": [ + 498, + 647, + 761, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP consists of a text encoder $E^{txt}$ and an image encoder $E^{img}$ , and encodes both types of input into 512-D embedding vectors. Let $T = \\{t^{(1)},\\dots ,t^{(K)}\\}$ denote a set of text prompts, and $t^{(i)}$ describes the $i$ -th preset attribute. To identify the attributes reflected in images, we embed training images and $T$ in the shared embedding space, and measure the semantic similarity as", + "bbox": [ + 496, + 671, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {(i)} (x) = \\cos \\left(E ^ {t x t} \\left(t ^ {(i)}\\right), E ^ {i m g} (x)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 781, + 890, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ denotes the cosine distance between input vectors. $S^{(i)}(x)$ should be larger when $t^{(i)}$ and $x$ represent the same attribute. At this point, we pseudo-announce training images, and the corresponding label $\\tilde{y}$ is defined as", + "bbox": [ + 496, + 805, + 893, + 866 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} ^ {(i)} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S ^ {(i)} (x) > \\tau , \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 869, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19287", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{y}^{(i)}$ indicates whether image $x$ reflects attribute $i$ , and $\\tau$ is a threshold. Based on the constructed supervision, we train an auxiliary classifier $A$ to capture the visual characteristics of preset attributes. For multi-attribute recognition, the evaluation term is defined as", + "bbox": [ + 76, + 90, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {c l s} = \\mathbf {E} _ {x} \\left[ \\sum_ {i = 1} ^ {K} - \\tilde {y} ^ {(i)} \\log A ^ {(i)} (x) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 176, + 468, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A^{(i)}(\\cdot)$ represents the predicted probability of an input reflecting attribute $i$ . Different from $E^{img}$ that is trained on generic data, $A$ focuses more on the domain-specific data, and can be expected to learn the most discriminative features to identify preset attributes.", + "bbox": [ + 76, + 231, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Latent Mapping Network", + "text_level": 1, + "bbox": [ + 76, + 319, + 310, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The latent space $\\mathcal{W}$ of StyleGAN possesses semantic disentanglement properties, and the works [1, 36, 40] further extend to $\\mathcal{W}+$ for better inversion quality. For a given input image $x$ , we adopt the e4e model $E^{sty}$ [40] to predict its latent code $w = E^{sty}(x)$ in the $\\mathcal{W}+$ space.", + "bbox": [ + 76, + 343, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is challenging to directly infer the latent transformations for a variety of attribute combinations. To address this problem, we design a two-stage architecture for our latent mapping network. The first stage is based on a direction prediction subnetwork $\\Gamma$ that produces latent directions denoted by $\\Gamma(w) = \\{\\gamma_w^{(1)},\\dots,\\gamma_w^{(K)}\\}$ , where $\\gamma_w^{(i)}$ associates with the preset attribute $i$ , conditioned on $w$ . At the second stage, a fusion subnetwork $\\Phi$ operates on the produced directions. We define a binary vector $\\Delta \\in \\{0,1\\}^K$ to indicate target attributes. $\\Phi$ learns to integrate the directions as indicated by $\\Delta$ , and infer a residual vector $\\gamma_w^\\Delta$ defined as", + "bbox": [ + 75, + 419, + 468, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {w} ^ {\\Delta} = \\Phi (w, M _ {\\Delta} \\otimes \\Gamma (w)), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 599, + 468, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the mask $M_{\\Delta}$ is constructed by broadcasting $\\Delta$ across an array having the same dimensionality as the latent code, and $\\otimes$ denotes the Hadamard product. As a result, multi-attribute manipulation can be carried out by simply adding the initial latent code to the residual vector. The generator $G$ of StyleGAN is employed to decode the resulting latent vector as", + "bbox": [ + 75, + 630, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} ^ {\\Delta} = G \\left(w + \\alpha \\gamma_ {w} ^ {\\Delta}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 747, + 468, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ controls the manipulation strength. Although our mapping network stacks two stages, each stage has access to the latent code of the input image.", + "bbox": [ + 76, + 779, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that the intermediate directions are semantically meaningful, they are required to work directly on the manipulation process of individual attributes. In this case, $\\Phi$ should not produce any changes to a single direction as input, and the corresponding invariance loss is formulated", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as follows:", + "bbox": [ + 500, + 90, + 575, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {s n g l} = \\mathbf {E} _ {x} \\left[ \\sum_ {i = 1} ^ {K} \\| \\Phi (w, M _ {i} \\otimes \\Gamma (w)) - \\gamma_ {w} ^ {(i)} \\| _ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 111, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{i}$ denotes the mask for selecting $\\gamma_w^{(i)}$ only. This design allows the mapping network to flexibly learn transformations with respect to different attribute combinations.", + "bbox": [ + 498, + 162, + 890, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic Evaluation", + "text_level": 1, + "bbox": [ + 500, + 217, + 694, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target attribute identification. The synthesized image $\\hat{x}^{\\Delta}$ should properly reflect the target attributes as indicated by $\\Delta$ . We leverage both classifier $A$ and CLIP encoders $\\{E^{txt}, E^{img}\\}$ to impose a target attribute identification loss, and the formulation is expressed as follows:", + "bbox": [ + 498, + 239, + 890, + 315 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {t r g t} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\sum_ {i = 1} ^ {K} - \\Delta^ {(i)} \\log A ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) \\right. \\tag {7} \\\\ \\left. + \\sum_ {i = 1} ^ {K} \\Delta^ {(i)} \\big (1 - \\cos (E ^ {t x t} (t ^ {(i)}), E ^ {i m g} (\\hat {x} ^ {\\Delta})) \\big) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 323, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the above equation, the first term evaluates the predictions of $A$ , and the second term measures the semantic consistency between the manipulation result and attribute text descriptions in the embedding space of CLIP. By minimizing $\\mathcal{L}^{trgt}$ , our mapping network seeks suitable transformations such that the target attributes can be well reflected in both views of $A$ and CLIP. In this way, we significantly improve the semantic accuracy of the transformation.", + "bbox": [ + 496, + 415, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Non-target attribute preservation. Minimizing the above losses does not guarantee that the synthesized images properly preserve the content of the input images while at the same time changing only the part related to the target attributes. To ensure that irrelevant attributes are unchanged before and after transformation, we further impose an attribute-aware consistency regularization on the mapping model. Specifically, we define the attribute-aware representation as", + "bbox": [ + 496, + 537, + 890, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {A} ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) = \\nu^ {(i)} \\otimes f _ {A} \\left(\\hat {x} ^ {\\Delta}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 678, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{A}(\\cdot)$ denotes the attribute classifier features, and $\\nu^{(i)}$ represents the weight vector of the head associated with the $i$ -th attribute. With the help of $\\nu^{(i)}$ , we can suppress the less informative features and highlight the useful ones due to the reason that the $i$ -th head measures its emergence in images. By modulating the classifier features in this way, the resulting ones capture more specific information, and we thus measure the semantic consistency on non-target attributes as", + "bbox": [ + 496, + 704, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {p r s v} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\sum_ {i = 1} ^ {K} \\left(1 - \\Delta^ {(i)}\\right) \\left\\| f _ {A} ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) - f _ {A} ^ {(i)} (x) \\right\\| _ {2} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 845, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19288", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to the predictions of $A$ , we find that the features $\\{f_A^{(1)},\\ldots ,f_A^{(K)}\\}$ encode richer information on the attributes. Minimizing $\\mathcal{L}^{prsv}$ enforces our model to preserve the non-target characteristics of the input image while manipulating the target attributes faithfully.", + "bbox": [ + 75, + 90, + 472, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Optimization", + "text_level": 1, + "bbox": [ + 76, + 179, + 269, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By integrating the above aspects of classifier training, mapping regularization and evaluation, the optimization formulation of the learnable components is expressed as follows:", + "bbox": [ + 75, + 202, + 468, + 262 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {A} \\mathcal {L} ^ {c l s},\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 272, + 212, + 295 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\Gamma} \\mathcal {L} ^ {t r g t} + \\mathcal {L} ^ {p r s v} + \\lambda \\mathcal {L} ^ {l o c} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 297, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\Phi} \\mathcal {L} ^ {t r g t} + \\mathcal {L} ^ {p r s v} + \\lambda \\mathcal {L} ^ {l o c} + \\mu \\mathcal {L} ^ {s n g l},\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 323, + 403, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the local searching term $\\mathcal{L}^{loc}$ is defined as", + "bbox": [ + 76, + 359, + 398, + 376 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {l o c} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\max \\left(\\left\\| \\gamma_ {w} ^ {\\Delta} \\right\\| _ {2} - \\epsilon , 0\\right) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 386, + 468, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\epsilon$ is a margin. Minimizing $\\mathcal{L}^{loc}$ prevents the modified latent code from deviating too far from the initial one. In addition, $\\lambda$ and $\\mu$ in Eq.(10) are the weighting factors that control the relative importance of the regularization terms, compared to the semantic evaluation terms, respectively. We jointly train $\\{A,\\Gamma ,\\Phi \\}$ from scratch to flexibly transform input images by randomly specifying the target attributes. Note that we do not perform a dedicated optimization for individual attribute (combination). In the test stage, both $\\Gamma$ and $\\Phi$ are fixed, and attribute manipulation is performed in a single forward pass for any given data.", + "bbox": [ + 75, + 417, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 599, + 209, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the test datasets and evaluation setups. Next, we investigate the effect of the main components of TUSLT. We further compare our model against multiple leading methods by performing both user studies and quantitative evaluation. Note that all the experimental results are obtained by applying the model on unseen images during the training phase.", + "bbox": [ + 75, + 625, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 741, + 267, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and preset attributes CelebA-HQ [20] is a widely used benchmark for facial image editing, and contains 30k high-resolution face images of celebrities. We follow [33] to set 38 attributes on hair style, hair color, expression, gender, age and others. We also use the animal and anime face datasets, AFHQ-cats/dogs [9] and Danbooru AnimeFace [4], with large intra-domain differences to evaluate the proposed model in manipulating 4 and 6 preset attributes, respectively.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f45625aa6bbdaf6b92249dd6e48f117675c64d4912a6f28ab7fb6e2248d4ce5d.jpg", + "image_caption": [ + "Figure 3. Single-attribute transformation results of StyleCLIP and TUSLT." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture and hyperparameters. The proposed model consists of three pre-trained networks (CLIP text and image encoders $\\{E^{txt}, E^{img}\\}$ and an e4e encoder $E^{sty}$ ) and two learnable components (an attribute classifier $A$ and a latent mapping network $\\{\\Gamma, \\Phi\\}$ ). We adopt a ResNet-50 [6] for $A$ , and the network architectures of $\\Gamma$ and $\\Phi$ consist of 4 fully connected layers. The threshold $\\tau$ in Eq.(1), the coefficient $\\alpha$ in Eq.(5), the weighting factors $\\lambda$ and $\\mu$ in Eq.(10) and the margin $\\epsilon$ in Eq.(11) are set to 0.7, 0.1, 0.8, 1 and 0.01, respectively. The proposed model is trained using the Adam optimizer [24] with a learning rate of 0.5. There are 50k training iterations, and each batch contains 2 images. We empirically find that our model can converge to a good solution without heavy tuning.", + "bbox": [ + 498, + 321, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocols. For all competing methods, we use the open source codes or pre-trained models. We assess the diversity and manipulation quality of synthesized images using the Fréchet Inception Distance (FID) [17]. On the other hand, we measure the correctness of attribute manipulation by an independent attribute classifier [28], and report the Target Attribute Recognition Rate (TARR). To quantitatively assess the model performance in irrelevant attribute preservation, we use the metric of IDentity Distance (IDD) before and after transformation in the feature space of a well-trained face recognition network [11].", + "bbox": [ + 498, + 532, + 892, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Analysis of Main Components", + "text_level": 1, + "bbox": [ + 500, + 710, + 767, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantically meaningful directions. We first perform an experiment to demonstrate the effectiveness of the prediction subnetwork $\\Gamma$ in inferring diverse semantic directions associated with preset attributes. For individual attributes, the manipulation is performed along the direction-s produced by $\\Gamma$ . In Figure 3, we visually compare with the main competing method, StyleCLIP, and observe that TUSLT is able to render more precise semantics about the target attributes.", + "bbox": [ + 498, + 733, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Precise manipulation on multiple attributes. To assess the fusion subnetwork $\\Phi$ , we can feed the directions", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19289", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3291adb8232c594e4b6a839014593648048decc68af0e8e8698d9e9fa911f0fa.jpg", + "image_caption": [ + "Figure 4. Ablative experiment results. The combinations of target attributes are (upper) 'Bald +Beard +Happy' and (bottom) 'Happy +Makeup +Bushy-eyebrows'." + ], + "image_footnote": [], + "bbox": [ + 161, + 88, + 803, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffd4fd9530f13bf6b17c561f6fb225f0492760b3d770cff201dd12c1946fac1e.jpg", + "image_caption": [ + "Figure 5. Comparison between TUSLT and variants in terms of (left) TARR and (right) IDD." + ], + "image_footnote": [], + "bbox": [ + 78, + 325, + 468, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\gamma_{CLIP} = \\{\\gamma_{CLIP}^{(1)},\\dots,\\gamma_{CLIP}^{(K)}\\}$ learnt by StyleCLIP, and the resulting model is referred to as $\\gamma_{CLIP} + \\Phi$ . On the other hand, we build another variant TUSLT w/o $A^\\prime$ by disabling the attribute classifier. In addition to visualizing the synthesis results in Figure 4, we report the TARR and IDD scores of the variants in Figure 5. $\\gamma_{CLIP} + \\Phi$ achieves comparable performance with StyleCLIP in terms of TARR, while the IDD value of the former is significantly lower than that of the latter. This confirms the effectiveness of $\\Phi$ in identity preservation. The results of TUSLT w/o $A^\\prime$ are unsatisfactory due to the lack of domain-specific knowledge learnt from the target data. In contrast, benefiting from $A$ our full model exhibits more precise manipulation ability.", + "bbox": [ + 75, + 547, + 468, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Irrelevant attribute preservation. It is also an important aspect to preserve the visual characteristics of the input image apart from the target attributes before and after transformation. We construct the attribute-aware features by Eq.(8) and impose the corresponding regularization term $\\mathcal{L}^{prsv}$ defined in Eq.(9). We remove $\\mathcal{L}^{prsv}$ from the overall training loss to build a variant 'TUSLT w/o $\\mathcal{L}^{prsv}$ . In Figures 4-5, we can observe that removing $\\mathcal{L}^{prsv}$ leads to more significant changes on the target attributes (a higher TARR score), while at the same time increasing the identity", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b00f9769625cda79c7ef3d7aa9d3a62202d7b76a7be055a105be033d2ff1d93.jpg", + "image_caption": [ + "Figure 6. Multi-attribute manipulation results of TUSLT and competing methods. Note that we examine the common attributes which all the methods are able to manipulate." + ], + "image_footnote": [], + "bbox": [ + 501, + 327, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distance between input and transformed images. The results suggest that our strategy leads to a significant improvement in the maintenance of irrelevant attributes.", + "bbox": [ + 498, + 776, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Human Evaluation", + "text_level": 1, + "bbox": [ + 500, + 830, + 681, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform user studies to evaluate TUSLT and a number of main leading methods in multi-attribute manipulation tasks. For InterFaceGAN and StyleFlow, we perfor", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19290", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/620becf2732cbef0957e2d9a0e001a1b821bc5f2571eb184673527052373e61e.jpg", + "image_caption": [ + "Figure 7. Diverse image synthesis results of TUSLT and StyleCLIP." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 885, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea2d315285e71c1ca504ace87c1b2406d7382e7d9d40d64f00ed48c74b49ebd1.jpg", + "table_caption": [ + "Table 1. The result of ranking TUSLT and competing models in multi-attribute transfer tasks. MA, IAP and VR denote Manipulation Accuracy, Irrelevant Attribute Preservation and Visual Realism, respectively." + ], + "table_footnote": [], + "table_body": "
MethodMAIAPVR
TUSLT (Ours)1.251.311.44
StyleCLIP [33]2.512.442.81
StyleFlow [2]2.633.252.56
InterFaceGAN [38]3.693.313.38
TediGAN [44]4.944.694.81
", + "bbox": [ + 155, + 472, + 390, + 560 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/726b93cd365d0ebf21cf73e735fd0e046e01d9021686a6bf038524b8515e6d55.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of TUSLT and competing methods." + ], + "table_footnote": [], + "table_body": "
MethodFID (↓)IDD (↓)PSNR (↑)SSIM (↑)TARR (↑)
TUSLT (Ours)56.910.4524.920.7587.85
StyleCLIP [33]63.930.4819.240.7185.74
StyleFlow [2]61.530.4621.840.7385.10
InterFaceGAN [38]69.310.5117.920.6881.80
TediGAN [44]58.740.4920.610.6918.93
", + "bbox": [ + 78, + 603, + 467, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "m multi-step translations to manipulate multiple target attributes. There are 18 attribute combinations, and 10 questions for each one. For each question, given an input image and the target attributes, the options are the images synthesized by the competing methods, and the order is randomly shuffled. A total of 50 validated workers are instructed to rank the synthesized images in terms of manipulation accuracy, irrelevant attribute preservation and visual realism. Table 1 summarizes the average ranking values of the competing methods. TUSLT is able to produce the best transformation results in most cases. We also provide the representative synthesis results of the competing methods in Figure", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. We observe that TediGAN fails in most of the manipulation tasks. In contrast, TUSLT produces images with more precise manipulation results, compared to the main competing models, StyleFlow and StyleCLIP. We can also find that TUSLT is the only model which is able to successfully manipulate all four attributes in the last example. In Figure 7, we visually compare with StyleCLIP in more tasks, and find that TUSLT yields more natural and plausible images.", + "bbox": [ + 496, + 417, + 890, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 556, + 718, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further perform a quantitative comparison between the proposed approach and competing methods. Note that this experiment involves the common attribute combinations which all of the competing models are able to manipulate. In Table 2, we report the average quantitative comparison results of our TUSLT and the competing methods in terms of FID, TARR and IDD. In addition, we adopt the metrics of Peak Signal-to-Noise Ratio (PSNR) and Structure SIMilarity (SSIM) to measure low to mid-level similarity between the input and synthesized images. The manipulations are performed on 5 attribute combinations. TUSLT outperforms the competing models by a large margin in both manipulation quality and irrelevant attribute preservation. In particular, our model achieves the highest PSNR/TARR score of 24.92/87.85, which is higher than that of the second best method (StyleFlow 21.84/85.10) by about 3 points. We further perform a comparison with a CLIP-based image manipulation method, HairCLIP, which focuses on editing hair color and style. Some representative synthesized images are shown in Figure 8, and the corresponding quantitative assessment on 6 attribute combina", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19291", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8606a651478ccbad7a0f392797425513e4edc50201b2b2e394c035201dc503d0.jpg", + "image_caption": [ + "Figure 8. Visual comparison between TUSLT and HairCLIP in hair attribute manipulation." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/001fcb2adcd9e53452249fee873d7d7eabd4cace5e3f7c70af7435975b262668.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of TUSLT and HairCLIP." + ], + "table_footnote": [], + "table_body": "
MethodFID (↓)PSNR (↑)sCLIP (↑)
TUSLT (Ours)50.1027.140.32
HairCLIP [42]50.8326.310.30
", + "bbox": [ + 127, + 319, + 419, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef5ce49f1dc0ee7876b13e45e1691b7ee1cf9080d8fa6a2b55081b78399c74cb.jpg", + "image_caption": [ + "Figure 9. Synthesis results of TUSLT on AFHQ." + ], + "image_footnote": [], + "bbox": [ + 81, + 385, + 467, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions is reported in Table 3. To measure manipulation accuracy, we compute the cosine similarity $s_{CLIP}$ between the attribute text descriptions and synthesized images in the CLIP feature space. HairCLIP performs less satisfactorily in rendering hair color as described in the text prompts.", + "bbox": [ + 75, + 579, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. One can find that the results lead to the same conclusions as our user studies. This is mainly due to the joint training of an auxiliary attribute classifier with CLIP-based supervision, and thus precise semantics of preset attributes can be captured. Another reason is the adopted two-stage architecture for latent transformation. Fusing semantically meaningful latent directions simplify the task of multi-attribute manipulation. This allows the proposed model to flexibly transform the input image to reflect a variety of attribute combinations.", + "bbox": [ + 75, + 655, + 468, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Results on AFHQ and AnimeFace", + "text_level": 1, + "bbox": [ + 76, + 816, + 374, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also show the ability of the proposed model to manipulate multiple attributes on AFHQ-cats/dogs and AnimeFace. For AFHQ, we follow the setting of StyleCLIP to specify preset attribute text descriptions and train TUSLT", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/463808c463bf3da2dec84d32e12dc021b549830883c62b64bfa9287983055bc9.jpg", + "image_caption": [ + "Figure 10. Synthesis results of TUSLT on AnimeFace." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to capture the corresponding semantics. For AnimeFace, we use the same setting as the experiments on CelebA-HQ. The results shown in Figures 9-10 confirm again that our model is scalable to multi-attribute transformations on animal and anime facial images and lead to significant visual modification.", + "bbox": [ + 498, + 282, + 890, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 388, + 617, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a text-guided unsupervised multi-attribute manipulation model to edit images in a single latent transformation step. Benefiting from the cross-modal image and text representation of CLIP, we can jointly train an auxiliary attribute classifier and a latent mapping network for precise attribute manipulation. We suggest two main reasons why the proposed model is able to successfully manipulate multiple attributes on diverse input images, compared to the leading methods. First, the use of a two-stage architecture enables our model to take care of all preset individual attributes, and thus provide the flexibility to handle diverse attribute combinations. Second, in contrast to CLIP encoders, the classifier learns domain-specific features to identify p-reset attributes, and offers semantic guidance not only for manipulating target attributes, but for preserving irrelevant attributes of the input image. This work significantly increases the scalability of StyleGAN-based image attribute manipulation without causing any manual annotation cost.", + "bbox": [ + 496, + 414, + 892, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 700, + 658, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the China Scholarship Council, in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the Natural Science Foundation of Guangdong Province (Project No. 2022A1515011160).", + "bbox": [ + 496, + 726, + 890, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 847, + 596, + 862 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: how to embed images into the StyleGAN", + "bbox": [ + 506, + 872, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19292", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "latent space? In Proc. International Conference on Computer Vision, 2019. 4", + "[2] Rameen Abdal, Peihao Zhu, Niloy J. Mitra, and Peter Wonka. StyleFlow: attribute-conditioned exploration of StyleGAN-Generated images using conditional continuous normalizing flows. ACM Transactions on Graphics, 40(3):1-21, 2021. 2, 3, 7", + "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Transactions on Graphics, 40(4):1-12, 2021. 3", + "[4] Gwern Branwen Aaron Gokaslan Anonymous, the Danbooru community. Danbooru2018: A large-scale crowdsourced and tagged anime illustration dataset. https://www.gwern.net/Danbooru2018, January 2019. 5", + "[5] Binod Bhattacharai and Tae-Kyun Kim. Inducing optimal attribute representations for conditional GANs. In Proc. European Conference on Computer Vision, 2020. 3", + "[6] Yue Chen, Yalong Bai, Wei Zhang, and Tao Mei. Destruction and construction learning for fine-grained image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 5157-5166, 2019. 5", + "[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 1, 3", + "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1", + "[9] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 3, 5", + "[10] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 3", + "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: additive angular margin loss for deep face recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5", + "[12] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive image properties. In Proc. International Conference on Computer Vision, 2019. 3", + "[13] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 1", + "[14] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 3", + "[15] Erik Harkonen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: discovering interpretable GAN" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "controls. In Proc. Neural Information Processing Systems, 2020. 1", + "[16] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. AttGAN: facial attribute editing by only changing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 3", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, and Bernhard Nessler. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Proc. Neural Information Processing Systems, 2017. 5", + "[18] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[19] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"ssteerability\" of generative adversarial networks. In arXiv:1907.07171, 2019. 3", + "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Proc. International Conference on Learning Representation, 2018. 5", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1", + "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1", + "[23] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 3", + "[24] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5", + "[25] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proc. International Conference on Learning Representation, 2017. 3", + "[26] Guillaume Lample, Neil Zeghidour, Nicolas Usunier, Antoine Bordes, Ludovic Denoyer, and Marc'Aurelio Ranzato. Fader networks: manipulating images by sliding attributes. In Proc. Neural Information Processing Systems, 2017. 3", + "[27] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. ManiGAN: text-guided image manipulation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3", + "[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: a unified selective transfer network for arbitrary image attribute editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3, 5", + "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2", + "[30] Yahui Liu, Marco De Nadai, Deng Cai, Huayang Li, Xavier Alameda-Pineda, Nicu Sebe, and Bruno Lepri. Describe" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19293", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "what to change: a text-guided unsupervised image-to-image translation approach. In Proc. ACM International Conference on Multimedia, 2020. 3", + "[31] M. Mirza and S. Osindero. Conditional generative adversarial nets. In arXiv:1411.1784, 2014. 2, 3", + "[32] Seonghyeon Nam, Yunji Kim, and Seon Joo Kim. Text-adaptive generative adversarial networks: manipulating images with natural language. In Proc. Neural Information Processing Systems, 2018. 3", + "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: text-driven manipulation of StyleGAN imagery. In Proc. International Conference on Computer Vision, 2021. 2, 3, 5, 7", + "[34] Alec Redford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning, 2021. 2, 3", + "[35] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Proc. International Conference on Machine Learning, 2016. 3", + "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Deniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3, 4", + "[37] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1", + "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: interpreting the disentangled face representation learned by GANs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2004-2018, 2020. 3, 7", + "[39] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 3, 4", + "[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Cross-domain and disentangled face manipulation with 3D guidance. IEEE Transactions on Visualization and Computer Graphics (Early Access), pages 1-15, 2021. 3", + "[42] Tianyi Wei, Dongdong Chen, Wenbo Zhou, and Jing Liao. HairCLIP: design your hier by text and reference image. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2022. 8", + "[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: disentangled controls for StyleGAN image generation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: text-guided diverse face image generation and manipulation. In arXiv:2012.03308, 2020. 3, 7", + "[45] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: fine-grained text to image generation with attentional generative adversarial networks. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2018. 3", + "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: text to photo-realistic image synthesis with stacked generative adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3", + "[47] Yutong Zheng, Yu-Kai Huang, Ran Tao, Zhiqiang Shen, and Marios Savvides. Unsupervised disentanglement of linear-encoded facial semantics. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[48] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3" + ], + "bbox": [ + 501, + 92, + 890, + 388 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19294", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_model.json b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4971335537f97f44c0cef50636fc47f44f6b1de9 --- /dev/null +++ b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.834, + 0.178 + ], + "angle": 0, + "content": "Text-guided Unsupervised Latent Transformation for Multi-attribute Image Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.203, + 0.808, + 0.222 + ], + "angle": 0, + "content": "Xiwen Wei\\(^{1}\\), Zhen Xu\\(^{1}\\), Cheng Liu\\(^{2}\\), Si Wu\\(^{1,3*}\\), Zhiwen Yu\\(^{1}\\), and Hau San Wong\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.222, + 0.825, + 0.24 + ], + "angle": 0, + "content": "1School of Computer Science and Engineering, South China University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.24, + 0.702, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, Shantou University" + }, + { + "type": "text", + "bbox": [ + 0.387, + 0.257, + 0.584, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.275, + 0.746, + 0.293 + ], + "angle": 0, + "content": "\\(^{4}\\)Department of Computer Science, City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.295, + 0.868, + 0.327 + ], + "angle": 0, + "content": "{202021044777, csxuzhen}@mail.scut.edu.cn, cliu@stu.edu.cn, {cswusi, zhwyu}@scut.edu.cn, cshswong@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.473, + 0.788 + ], + "angle": 0, + "content": "Great progress has been made in StyleGAN-based image editing. To associate with preset attributes, most existing approaches focus on supervised learning for semantically meaningful latent space traversal directions, and each manipulation step is typically determined for an individual attribute. To address this limitation, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model, which adaptively infers a single transformation step in the latent space of StyleGAN to simultaneously manipulate multiple attributes on a given input image. Specifically, we adopt a two-stage architecture for a latent mapping network to break down the transformation process into two manageable steps. Our network first learns a diverse set of semantic directions tailored to an input image, and later nonlinearly fuses the ones associated with the target attributes to infer a residual vector. The resulting tightly interlinked two-stage architecture delivers the flexibility to handle diverse attribute combinations. By leveraging the cross-modal text-image representation of CLIP, we can perform pseudo annotations based on the semantic similarity between preset attribute text descriptions and training images, and further jointly train an auxiliary attribute classifier with the latent mapping network to provide semantic guidance. We perform extensive experiments to demonstrate that the adopted strategies contribute to the superior performance of TUSLT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.819, + 0.21, + 0.834 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.845, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Visual attributes represent semantically meaningful features inherent in images, and attribute manipulation has ex" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.361, + 0.868, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Visually comparing TUSLT with StyleFlow (supervised) and StyleCLIP (text-driven) in precisely manipulating multiple attributes and preserving irrelevant attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.732 + ], + "angle": 0, + "content": "perceived great improvements, due to the advent of Generative Adversarial Network [13] (GAN)-based generative models, e.g. StyleGAN [21, 22] and StarGAN [7, 8]. Recent works [15, 37, 43] have discovered that the latent space of StyleGAN possesses semantic disentanglement properties, enabling a variety of image editing operations via latent transformations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "StyleGAN-based methods for image attribute manipulation typically involve a large number of manual annotations or well-trained attribute classifiers. Furthermore, the discovered semantic latent directions are associated with individual attributes. The editing on a target attribute is carried out by moving the latent code of an input image along one of the directions. For \\( K \\) target attributes, these models require \\( K \\) transformation steps to handle the translation. As a result, they are not scalable to the increasing number of target attributes in multi-attribute transformation tasks. As shown in Figure 1, we test a state-of-the-art supervised" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19285" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.652, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.09, + 0.895, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.321, + 0.893, + 0.392 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed model, TUSLT, consisting of two learnable components: an auxiliary attribute classifier \\( A \\) trained on the CLIP-based labeled data, and a latent mapping network \\( \\{\\Gamma, \\Phi\\} \\). \\( \\Gamma \\) infers latent directions \\( \\{\\gamma_w^{(1)}, \\dots, \\gamma_w^{(K)}\\} \\) for preset attributes, and \\( \\Phi \\) transforms the target-related directions as indicated by mask \\( M_{\\Delta} \\) into a residual vector \\( \\gamma_w^{\\Delta} \\), to which the initial latent code is added. Precise multi-attribute transfer is allowed by such a single transformation step, and the generator \\( G \\) synthesizes a new image reflecting the target attributes under the guidance of \\( A \\) and CLIP encoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.509 + ], + "angle": 0, + "content": "model, StyleFlow [2], and find that multiple transformation steps lead to undesired deviation from the input image on irrelevant attributes. Compared to the state-of-the-art text-driven model, StyleCLIP [33], we can also achieve a better manipulation result by seeking a single latent transformation step for the task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.471, + 0.902 + ], + "angle": 0, + "content": "More specifically, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model that supports simultaneous manipulation on multiple attributes. As shown in Figure 2, the key is to jointly learn a mapping network to infer the latent transformation and an auxiliary attribute classifier to assess manipulation quality. We employ the Contrastive Language-Image Pre-training (CLIP) model [34] to generate pseudo-labeled data by measuring the semantic similarities between attribute text descriptions and training images. Compared to CLIP, the jointly trained classifier extracts domain-specific information to better characterize the differences among attributes. This benefits the mapping network to seek more suitable transformations, such that the synthesized images reflect target attributes. Further, we adopt a two-stage architecture for the mapping network: the earlier stage employs a prediction subnetwork to infer a set of semantic directions, and the latter stage operates on the resulting directions and nonlinearly fuses the target-related ones. The intermediate semantic directions are associated with preset attributes and tailored for the input image. This design allows us to deal with a wide range of attribute combinations in a single transformation step. We perform extensive experiments and provide both qualitative and quantitative results in diverse multi-attribute transformation tasks, showing the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.419, + 0.856, + 0.434 + ], + "angle": 0, + "content": "superiority of our model over the competing methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.705 + ], + "angle": 0, + "content": "In summary, the main contributions of this work are given as follows: (a) The existing image editing methods focus on discovering semantic latent directions associated with individual visual attributes, and a sequential manipulation process is thus needed for multi-attribute manipulation. In contrast, the proposed model infers a single step of latent space walk to simultaneously manipulate multiple attributes. (b) Benefiting from the cross-modal text-image representation of CLIP, we jointly train a latent mapping network with an auxiliary attribute classifier, which leads to more precise attribute rendering without requiring additional manual annotations. (c) Due to the two-stage nature, our latent mapping network breaks down the challenging multi-attribute manipulation task into sub-tasks: inferring diverse semantic directions and integrating the target-related ones into a single transformation vector. This design gives our model interpretability and flexibility in dealing with a variety of attribute combinations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.642, + 0.733 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.816, + 0.759 + ], + "angle": 0, + "content": "2.1. Generic Image-to-image Translation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "As one of the earliest image translation models, pix2pix [18] learnt a cross-domain mapping via conditional GAN [31]. In addition to an adversarial training loss, the consistency regularization between each input image and the corresponding ground truth was imposed in the training process. To alleviate the problem of pairwise training data, many unpaired image translation models have been developed. UNIT [29] trained two generators to approximate the joint distribution of images from different domains. Cy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "cle consistency regularization was adopted in CycleGAN [48] and discoGAN [23]. To efficiently learn the mappings among multiple domains, StarGAN [7, 9] was a unified framework in which a single generator was trained to translate an input image into different domains, conditioned on domain information." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.193, + 0.295, + 0.21 + ], + "angle": 0, + "content": "2.2. Attribute Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.398 + ], + "angle": 0, + "content": "FaderNet [26] used an encoder-decoder architecture and learnt attribute-aware latent representations in an unsupervised manner. To ensure correct manipulation on the desired attributes, an attribute classification constraint was incorporated in AttGAN [16]. Instead of handling all target attributes, STGAN [28] contained a selective transfer module to only encode the changed attributes, based on the d-difference between source and reference images. In IOAGAN [5], the relationship of preset attributes were leveraged via a Graph Convolutional Network (GCN) [25] to infer and inject an integrated embedding of attributes into a translation network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.655 + ], + "angle": 0, + "content": "The latent spaces of GANs demonstrate promising semantic organization [10, 12, 19]. With the progress of high-fidelity GAN inversion [14, 36, 40], there are many works on exploring the latent space of a pre-trained StyleGAN generator for image manipulation [3, 41]. SeFa [39] disentangled semantics from StyleGAN by decomposing the matrix of generator weights. In addition, LEFS [47] applied s-parse representation learning to unsupervised semantic disentanglement. As a representative supervised method, InterFaceGAN [38] searched for semantic latent directions by solving a series of binary classification problems about preset attributes. Instead of latent directions, StyleSpace [43] explored style-associated channels with the help of a pretrained classifier. Different from the above linear latent transformation methods, StyleFlow [2] employed a conditional flow model to learn non-linear paths for attribute manipulation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.32, + 0.683 + ], + "angle": 0, + "content": "2.3. Text-guided Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Cross-modal representation learning on visual and language data brings about substantive progress in text-guided image synthesis. An early attempt was to train a GAN-based model [31], conditioned on text embedding [35]. Further, a stacked structure [46] and an attention-based regularization approach [45] were designed to improve synthesis quality. In [30, 32], image content from visual attributes were disentangled and associated with text descriptions without using manually annotated data. ManiGAN [27] aimed to learn text and image cross-modality representations, such that semantic regions were associated with the corresponding text via an affine transformation. TediGAN [44] aligned and projected text and image representations into the latent space of StyleGAN. Further, Style" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "CLIP [33] integrated CLIP [34] with StyleGAN to leverage CLIP-based linguistic-visual semantic consistency regularization for better manipulation quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.379 + ], + "angle": 0, + "content": "Among the aforementioned models, StyleFlow [2] and StyleCLIP [33] are the most relevant to our work. StyleFlow is a supervised method, and the learnt semantic directions are associated with individual attributes. For multi-attribute manipulation, StyleFlow needs a sequential transformation process, in which the target attributes are edited one at a time. In contrast, the proposed model learns to infer a single transformation step to simultaneously manipulate multiple attributes in an unsupervised manner. StyleCLIP trains a dedicated mapping network for each text description, and thus has limited scalability in handling a variety of attribute combinations. Unlike StyleCLIP, our model uses a two-stage mapping network and jointly trains an auxiliary attribute classifier without manual annotations. This structure benefits the scalability and semantic accuracy in changing multiple attributes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.39, + 0.673, + 0.407 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.642 + ], + "angle": 0, + "content": "Our goal is to translate an input image into a new one reflecting the target attributes. Based on the linguistic-visual representation of CLIP, we can pseudo-label training images by measuring the semantic similarities between attribute text descriptions and training images. We believe that an auxiliary attribute classifier can capture the most discriminative information, which complements the role of CLIP encoders to a certain extent. Based on this, the constructed supervision is leveraged by jointly training a latent mapping network and the classifier. The former has a two-stage architecture to infer a diverse set of semantic latent directions followed by selectively integrating target-related ones, and the latter provides semantic guidance. As a result, multi-attribute manipulation can be precisely carried out in a single forward pass." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.648, + 0.762, + 0.665 + ], + "angle": 0, + "content": "3.1. Auxiliary Attribute Classifier" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.779 + ], + "angle": 0, + "content": "CLIP consists of a text encoder \\( E^{txt} \\) and an image encoder \\( E^{img} \\), and encodes both types of input into 512-D embedding vectors. Let \\( T = \\{t^{(1)},\\dots ,t^{(K)}\\} \\) denote a set of text prompts, and \\( t^{(i)} \\) describes the \\( i \\)-th preset attribute. To identify the attributes reflected in images, we embed training images and \\( T \\) in the shared embedding space, and measure the semantic similarity as" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.782, + 0.891, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {(i)} (x) = \\cos \\left(E ^ {t x t} \\left(t ^ {(i)}\\right), E ^ {i m g} (x)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.806, + 0.894, + 0.867 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) denotes the cosine distance between input vectors. \\(S^{(i)}(x)\\) should be larger when \\(t^{(i)}\\) and \\(x\\) represent the same attribute. At this point, we pseudo-announce training images, and the corresponding label \\(\\tilde{y}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.871, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} ^ {(i)} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S ^ {(i)} (x) > \\tau , \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19287" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.166 + ], + "angle": 0, + "content": "where \\(\\tilde{y}^{(i)}\\) indicates whether image \\(x\\) reflects attribute \\(i\\), and \\(\\tau\\) is a threshold. Based on the constructed supervision, we train an auxiliary classifier \\(A\\) to capture the visual characteristics of preset attributes. For multi-attribute recognition, the evaluation term is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {c l s} = \\mathbf {E} _ {x} \\left[ \\sum_ {i = 1} ^ {K} - \\tilde {y} ^ {(i)} \\log A ^ {(i)} (x) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.232, + 0.469, + 0.308 + ], + "angle": 0, + "content": "where \\( A^{(i)}(\\cdot) \\) represents the predicted probability of an input reflecting attribute \\( i \\). Different from \\( E^{img} \\) that is trained on generic data, \\( A \\) focuses more on the domain-specific data, and can be expected to learn the most discriminative features to identify preset attributes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.32, + 0.311, + 0.336 + ], + "angle": 0, + "content": "3.2. Latent Mapping Network" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.469, + 0.419 + ], + "angle": 0, + "content": "The latent space \\(\\mathcal{W}\\) of StyleGAN possesses semantic disentanglement properties, and the works [1, 36, 40] further extend to \\(\\mathcal{W}+\\) for better inversion quality. For a given input image \\(x\\), we adopt the e4e model \\(E^{sty}\\) [40] to predict its latent code \\(w = E^{sty}(x)\\) in the \\(\\mathcal{W}+\\) space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.469, + 0.589 + ], + "angle": 0, + "content": "It is challenging to directly infer the latent transformations for a variety of attribute combinations. To address this problem, we design a two-stage architecture for our latent mapping network. The first stage is based on a direction prediction subnetwork \\(\\Gamma\\) that produces latent directions denoted by \\(\\Gamma(w) = \\{\\gamma_w^{(1)},\\dots,\\gamma_w^{(K)}\\}\\), where \\(\\gamma_w^{(i)}\\) associates with the preset attribute \\(i\\), conditioned on \\(w\\). At the second stage, a fusion subnetwork \\(\\Phi\\) operates on the produced directions. We define a binary vector \\(\\Delta \\in \\{0,1\\}^K\\) to indicate target attributes. \\(\\Phi\\) learns to integrate the directions as indicated by \\(\\Delta\\), and infer a residual vector \\(\\gamma_w^\\Delta\\) defined as" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.601, + 0.469, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {w} ^ {\\Delta} = \\Phi (w, M _ {\\Delta} \\otimes \\Gamma (w)), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.469, + 0.734 + ], + "angle": 0, + "content": "where the mask \\( M_{\\Delta} \\) is constructed by broadcasting \\( \\Delta \\) across an array having the same dimensionality as the latent code, and \\( \\otimes \\) denotes the Hadamard product. As a result, multi-attribute manipulation can be carried out by simply adding the initial latent code to the residual vector. The generator \\( G \\) of StyleGAN is employed to decode the resulting latent vector as" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.748, + 0.469, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\hat {x} ^ {\\Delta} = G \\left(w + \\alpha \\gamma_ {w} ^ {\\Delta}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.825 + ], + "angle": 0, + "content": "where \\(\\alpha\\) controls the manipulation strength. Although our mapping network stacks two stages, each stage has access to the latent code of the input image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To ensure that the intermediate directions are semantically meaningful, they are required to work directly on the manipulation process of individual attributes. In this case, \\(\\Phi\\) should not produce any changes to a single direction as input, and the corresponding invariance loss is formulated" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.576, + 0.105 + ], + "angle": 0, + "content": "as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.112, + 0.892, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {s n g l} = \\mathbf {E} _ {x} \\left[ \\sum_ {i = 1} ^ {K} \\| \\Phi (w, M _ {i} \\otimes \\Gamma (w)) - \\gamma_ {w} ^ {(i)} \\| _ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.21 + ], + "angle": 0, + "content": "where \\(M_{i}\\) denotes the mask for selecting \\(\\gamma_w^{(i)}\\) only. This design allows the mapping network to flexibly learn transformations with respect to different attribute combinations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.218, + 0.696, + 0.232 + ], + "angle": 0, + "content": "3.3. Semantic Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.241, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Target attribute identification. The synthesized image \\(\\hat{x}^{\\Delta}\\) should properly reflect the target attributes as indicated by \\(\\Delta\\). We leverage both classifier \\(A\\) and CLIP encoders \\(\\{E^{txt}, E^{img}\\}\\) to impose a target attribute identification loss, and the formulation is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.324, + 0.892, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {t r g t} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\sum_ {i = 1} ^ {K} - \\Delta^ {(i)} \\log A ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) \\right. \\tag {7} \\\\ \\left. + \\sum_ {i = 1} ^ {K} \\Delta^ {(i)} \\big (1 - \\cos (E ^ {t x t} (t ^ {(i)}), E ^ {i m g} (\\hat {x} ^ {\\Delta})) \\big) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.536 + ], + "angle": 0, + "content": "In the above equation, the first term evaluates the predictions of \\( A \\), and the second term measures the semantic consistency between the manipulation result and attribute text descriptions in the embedding space of CLIP. By minimizing \\( \\mathcal{L}^{trgt} \\), our mapping network seeks suitable transformations such that the target attributes can be well reflected in both views of \\( A \\) and CLIP. In this way, we significantly improve the semantic accuracy of the transformation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Non-target attribute preservation. Minimizing the above losses does not guarantee that the synthesized images properly preserve the content of the input images while at the same time changing only the part related to the target attributes. To ensure that irrelevant attributes are unchanged before and after transformation, we further impose an attribute-aware consistency regularization on the mapping model. Specifically, we define the attribute-aware representation as" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.679, + 0.891, + 0.699 + ], + "angle": 0, + "content": "\\[\nf _ {A} ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) = \\nu^ {(i)} \\otimes f _ {A} \\left(\\hat {x} ^ {\\Delta}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.84 + ], + "angle": 0, + "content": "where \\( f_{A}(\\cdot) \\) denotes the attribute classifier features, and \\( \\nu^{(i)} \\) represents the weight vector of the head associated with the \\( i \\)-th attribute. With the help of \\( \\nu^{(i)} \\), we can suppress the less informative features and highlight the useful ones due to the reason that the \\( i \\)-th head measures its emergence in images. By modulating the classifier features in this way, the resulting ones capture more specific information, and we thus measure the semantic consistency on non-target attributes as" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.847, + 0.891, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {p r s v} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\sum_ {i = 1} ^ {K} \\left(1 - \\Delta^ {(i)}\\right) \\left\\| f _ {A} ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) - f _ {A} ^ {(i)} (x) \\right\\| _ {2} \\right]. \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19288" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.171 + ], + "angle": 0, + "content": "Compared to the predictions of \\(A\\), we find that the features \\(\\{f_A^{(1)},\\ldots ,f_A^{(K)}\\}\\) encode richer information on the attributes. Minimizing \\(\\mathcal{L}^{prsv}\\) enforces our model to preserve the non-target characteristics of the input image while manipulating the target attributes faithfully." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.18, + 0.27, + 0.196 + ], + "angle": 0, + "content": "3.4. Model Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.203, + 0.47, + 0.263 + ], + "angle": 0, + "content": "By integrating the above aspects of classifier training, mapping regularization and evaluation, the optimization formulation of the learnable components is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.273, + 0.214, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\min _ {A} \\mathcal {L} ^ {c l s},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.298, + 0.469, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\Gamma} \\mathcal {L} ^ {t r g t} + \\mathcal {L} ^ {p r s v} + \\lambda \\mathcal {L} ^ {l o c} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.324, + 0.405, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\Phi} \\mathcal {L} ^ {t r g t} + \\mathcal {L} ^ {p r s v} + \\lambda \\mathcal {L} ^ {l o c} + \\mu \\mathcal {L} ^ {s n g l},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.361, + 0.4, + 0.377 + ], + "angle": 0, + "content": "where the local searching term \\(\\mathcal{L}^{loc}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.387, + 0.469, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {l o c} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\max \\left(\\left\\| \\gamma_ {w} ^ {\\Delta} \\right\\| _ {2} - \\epsilon , 0\\right) \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.587 + ], + "angle": 0, + "content": "and \\(\\epsilon\\) is a margin. Minimizing \\(\\mathcal{L}^{loc}\\) prevents the modified latent code from deviating too far from the initial one. In addition, \\(\\lambda\\) and \\(\\mu\\) in Eq.(10) are the weighting factors that control the relative importance of the regularization terms, compared to the semantic evaluation terms, respectively. We jointly train \\(\\{A,\\Gamma ,\\Phi \\}\\) from scratch to flexibly transform input images by randomly specifying the target attributes. Note that we do not perform a dedicated optimization for individual attribute (combination). In the test stage, both \\(\\Gamma\\) and \\(\\Phi\\) are fixed, and attribute manipulation is performed in a single forward pass for any given data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.6, + 0.21, + 0.617 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.732 + ], + "angle": 0, + "content": "In this section, we first describe the test datasets and evaluation setups. Next, we investigate the effect of the main components of TUSLT. We further compare our model against multiple leading methods by performing both user studies and quantitative evaluation. Note that all the experimental results are obtained by applying the model on unseen images during the training phase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.268, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Datasets and preset attributes CelebA-HQ [20] is a widely used benchmark for facial image editing, and contains 30k high-resolution face images of celebrities. We follow [33] to set 38 attributes on hair style, hair color, expression, gender, age and others. We also use the animal and anime face datasets, AFHQ-cats/dogs [9] and Danbooru AnimeFace [4], with large intra-domain differences to evaluate the proposed model in manipulating 4 and 6 preset attributes, respectively." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.266, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Single-attribute transformation results of StyleCLIP and TUSLT." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.322, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Architecture and hyperparameters. The proposed model consists of three pre-trained networks (CLIP text and image encoders \\(\\{E^{txt}, E^{img}\\}\\) and an e4e encoder \\(E^{sty}\\)) and two learnable components (an attribute classifier \\(A\\) and a latent mapping network \\(\\{\\Gamma, \\Phi\\}\\)). We adopt a ResNet-50 [6] for \\(A\\), and the network architectures of \\(\\Gamma\\) and \\(\\Phi\\) consist of 4 fully connected layers. The threshold \\(\\tau\\) in Eq.(1), the coefficient \\(\\alpha\\) in Eq.(5), the weighting factors \\(\\lambda\\) and \\(\\mu\\) in Eq.(10) and the margin \\(\\epsilon\\) in Eq.(11) are set to 0.7, 0.1, 0.8, 1 and 0.01, respectively. The proposed model is trained using the Adam optimizer [24] with a learning rate of 0.5. There are 50k training iterations, and each batch contains 2 images. We empirically find that our model can converge to a good solution without heavy tuning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.534, + 0.893, + 0.701 + ], + "angle": 0, + "content": "Evaluation protocols. For all competing methods, we use the open source codes or pre-trained models. We assess the diversity and manipulation quality of synthesized images using the Fréchet Inception Distance (FID) [17]. On the other hand, we measure the correctness of attribute manipulation by an independent attribute classifier [28], and report the Target Attribute Recognition Rate (TARR). To quantitatively assess the model performance in irrelevant attribute preservation, we use the metric of IDentity Distance (IDD) before and after transformation in the feature space of a well-trained face recognition network [11]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.768, + 0.727 + ], + "angle": 0, + "content": "4.2. Analysis of Main Components" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.734, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Semantically meaningful directions. We first perform an experiment to demonstrate the effectiveness of the prediction subnetwork \\(\\Gamma\\) in inferring diverse semantic directions associated with preset attributes. For individual attributes, the manipulation is performed along the direction-s produced by \\(\\Gamma\\). In Figure 3, we visually compare with the main competing method, StyleCLIP, and observe that TUSLT is able to render more precise semantics about the target attributes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Precise manipulation on multiple attributes. To assess the fusion subnetwork \\(\\Phi\\), we can feed the directions" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.089, + 0.805, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Ablative experiment results. The combinations of target attributes are (upper) 'Bald +Beard +Happy' and (bottom) 'Happy +Makeup +Bushy-eyebrows'." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.326, + 0.47, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.52 + ], + "angle": 0, + "content": "Figure 5. Comparison between TUSLT and variants in terms of (left) TARR and (right) IDD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.748 + ], + "angle": 0, + "content": "\\(\\gamma_{CLIP} = \\{\\gamma_{CLIP}^{(1)},\\dots,\\gamma_{CLIP}^{(K)}\\}\\) learnt by StyleCLIP, and the resulting model is referred to as \\(\\gamma_{CLIP} + \\Phi\\) . On the other hand, we build another variant TUSLT w/o \\(A^\\prime\\) by disabling the attribute classifier. In addition to visualizing the synthesis results in Figure 4, we report the TARR and IDD scores of the variants in Figure 5. \\(\\gamma_{CLIP} + \\Phi\\) achieves comparable performance with StyleCLIP in terms of TARR, while the IDD value of the former is significantly lower than that of the latter. This confirms the effectiveness of \\(\\Phi\\) in identity preservation. The results of TUSLT w/o \\(A^\\prime\\) are unsatisfactory due to the lack of domain-specific knowledge learnt from the target data. In contrast, benefiting from \\(A\\) our full model exhibits more precise manipulation ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Irrelevant attribute preservation. It is also an important aspect to preserve the visual characteristics of the input image apart from the target attributes before and after transformation. We construct the attribute-aware features by Eq.(8) and impose the corresponding regularization term \\(\\mathcal{L}^{prsv}\\) defined in Eq.(9). We remove \\(\\mathcal{L}^{prsv}\\) from the overall training loss to build a variant 'TUSLT w/o \\(\\mathcal{L}^{prsv}\\). In Figures 4-5, we can observe that removing \\(\\mathcal{L}^{prsv}\\) leads to more significant changes on the target attributes (a higher TARR score), while at the same time increasing the identity" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.328, + 0.892, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.708, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Figure 6. Multi-attribute manipulation results of TUSLT and competing methods. Note that we examine the common attributes which all the methods are able to manipulate." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.822 + ], + "angle": 0, + "content": "distance between input and transformed images. The results suggest that our strategy leads to a significant improvement in the maintenance of irrelevant attributes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.683, + 0.848 + ], + "angle": 0, + "content": "4.3. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We perform user studies to evaluate TUSLT and a number of main leading methods in multi-attribute manipulation tasks. For InterFaceGAN and StyleFlow, we perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19290" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.887, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.377, + 0.685, + 0.391 + ], + "angle": 0, + "content": "Figure 7. Diverse image synthesis results of TUSLT and StyleCLIP." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.471 + ], + "angle": 0, + "content": "Table 1. The result of ranking TUSLT and competing models in multi-attribute transfer tasks. MA, IAP and VR denote Manipulation Accuracy, Irrelevant Attribute Preservation and Visual Realism, respectively." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.473, + 0.391, + 0.561 + ], + "angle": 0, + "content": "
MethodMAIAPVR
TUSLT (Ours)1.251.311.44
StyleCLIP [33]2.512.442.81
StyleFlow [2]2.633.252.56
InterFaceGAN [38]3.693.313.38
TediGAN [44]4.944.694.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.576, + 0.468, + 0.602 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of TUSLT and competing methods." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.604, + 0.468, + 0.693 + ], + "angle": 0, + "content": "
MethodFID (↓)IDD (↓)PSNR (↑)SSIM (↑)TARR (↑)
TUSLT (Ours)56.910.4524.920.7587.85
StyleCLIP [33]63.930.4819.240.7185.74
StyleFlow [2]61.530.4621.840.7385.10
InterFaceGAN [38]69.310.5117.920.6881.80
TediGAN [44]58.740.4920.610.6918.93
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "m multi-step translations to manipulate multiple target attributes. There are 18 attribute combinations, and 10 questions for each one. For each question, given an input image and the target attributes, the options are the images synthesized by the competing methods, and the order is randomly shuffled. A total of 50 validated workers are instructed to rank the synthesized images in terms of manipulation accuracy, irrelevant attribute preservation and visual realism. Table 1 summarizes the average ranking values of the competing methods. TUSLT is able to produce the best transformation results in most cases. We also provide the representative synthesis results of the competing methods in Figure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.54 + ], + "angle": 0, + "content": "6. We observe that TediGAN fails in most of the manipulation tasks. In contrast, TUSLT produces images with more precise manipulation results, compared to the main competing models, StyleFlow and StyleCLIP. We can also find that TUSLT is the only model which is able to successfully manipulate all four attributes in the last example. In Figure 7, we visually compare with StyleCLIP in more tasks, and find that TUSLT yields more natural and plausible images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.72, + 0.573 + ], + "angle": 0, + "content": "4.4.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further perform a quantitative comparison between the proposed approach and competing methods. Note that this experiment involves the common attribute combinations which all of the competing models are able to manipulate. In Table 2, we report the average quantitative comparison results of our TUSLT and the competing methods in terms of FID, TARR and IDD. In addition, we adopt the metrics of Peak Signal-to-Noise Ratio (PSNR) and Structure SIMilarity (SSIM) to measure low to mid-level similarity between the input and synthesized images. The manipulations are performed on 5 attribute combinations. TUSLT outperforms the competing models by a large margin in both manipulation quality and irrelevant attribute preservation. In particular, our model achieves the highest PSNR/TARR score of 24.92/87.85, which is higher than that of the second best method (StyleFlow 21.84/85.10) by about 3 points. We further perform a comparison with a CLIP-based image manipulation method, HairCLIP, which focuses on editing hair color and style. Some representative synthesized images are shown in Figure 8, and the corresponding quantitative assessment on 6 attribute combina" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19291" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.261, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Figure 8. Visual comparison between TUSLT and HairCLIP in hair attribute manipulation." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.305, + 0.449, + 0.318 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of TUSLT and HairCLIP." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.32, + 0.42, + 0.372 + ], + "angle": 0, + "content": "
MethodFID (↓)PSNR (↑)sCLIP (↑)
TUSLT (Ours)50.1027.140.32
HairCLIP [42]50.8326.310.30
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.386, + 0.468, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.539, + 0.419, + 0.553 + ], + "angle": 0, + "content": "Figure 9. Synthesis results of TUSLT on AFHQ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.655 + ], + "angle": 0, + "content": "tions is reported in Table 3. To measure manipulation accuracy, we compute the cosine similarity \\( s_{CLIP} \\) between the attribute text descriptions and synthesized images in the CLIP feature space. HairCLIP performs less satisfactorily in rendering hair color as described in the text prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.469, + 0.806 + ], + "angle": 0, + "content": "Discussion. One can find that the results lead to the same conclusions as our user studies. This is mainly due to the joint training of an auxiliary attribute classifier with CLIP-based supervision, and thus precise semantics of preset attributes can be captured. Another reason is the adopted two-stage architecture for latent transformation. Fusing semantically meaningful latent directions simplify the task of multi-attribute manipulation. This allows the proposed model to flexibly transform the input image to reflect a variety of attribute combinations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.375, + 0.833 + ], + "angle": 0, + "content": "4.5. Results on AFHQ and AnimeFace" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We also show the ability of the proposed model to manipulate multiple attributes on AFHQ-cats/dogs and AnimeFace. For AFHQ, we follow the setting of StyleCLIP to specify preset attribute text descriptions and train TUSLT" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.243, + 0.858, + 0.258 + ], + "angle": 0, + "content": "Figure 10. Synthesis results of TUSLT on AnimeFace." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.892, + 0.375 + ], + "angle": 0, + "content": "to capture the corresponding semantics. For AnimeFace, we use the same setting as the experiments on CelebA-HQ. The results shown in Figures 9-10 confirm again that our model is scalable to multi-attribute transformations on animal and anime facial images and lead to significant visual modification." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.39, + 0.619, + 0.405 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.893, + 0.688 + ], + "angle": 0, + "content": "We propose a text-guided unsupervised multi-attribute manipulation model to edit images in a single latent transformation step. Benefiting from the cross-modal image and text representation of CLIP, we can jointly train an auxiliary attribute classifier and a latent mapping network for precise attribute manipulation. We suggest two main reasons why the proposed model is able to successfully manipulate multiple attributes on diverse input images, compared to the leading methods. First, the use of a two-stage architecture enables our model to take care of all preset individual attributes, and thus provide the flexibility to handle diverse attribute combinations. Second, in contrast to CLIP encoders, the classifier learns domain-specific features to identify p-reset attributes, and offers semantic guidance not only for manipulating target attributes, but for preserving irrelevant attributes of the input image. This work significantly increases the scalability of StyleGAN-based image attribute manipulation without causing any manual annotation cost." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.702, + 0.659, + 0.719 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.833 + ], + "angle": 0, + "content": "This work was supported in part by the China Scholarship Council, in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the Natural Science Foundation of Guangdong Province (Project No. 2022A1515011160)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.848, + 0.597, + 0.863 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: how to embed images into the StyleGAN" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "latent space? In Proc. International Conference on Computer Vision, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J. Mitra, and Peter Wonka. StyleFlow: attribute-conditioned exploration of StyleGAN-Generated images using conditional continuous normalizing flows. ACM Transactions on Graphics, 40(3):1-21, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.192, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Transactions on Graphics, 40(4):1-12, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.249, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[4] Gwern Branwen Aaron Gokaslan Anonymous, the Danbooru community. Danbooru2018: A large-scale crowdsourced and tagged anime illustration dataset. https://www.gwern.net/Danbooru2018, January 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.307, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[5] Binod Bhattacharai and Tae-Kyun Kim. Inducing optimal attribute representations for conditional GANs. In Proc. European Conference on Computer Vision, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.349, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[6] Yue Chen, Yalong Bai, Wei Zhang, and Tao Mei. Destruction and construction learning for fine-grained image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 5157-5166, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.406, + 0.471, + 0.475 + ], + "angle": 0, + "content": "[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.477, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.533, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[9] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[10] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: additive angular margin loss for deep face recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[12] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive image properties. In Proc. International Conference on Computer Vision, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[13] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[14] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[15] Erik Harkonen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: discovering interpretable GAN" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "controls. In Proc. Neural Information Processing Systems, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. AttGAN: facial attribute editing by only changing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, and Bernhard Nessler. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Proc. Neural Information Processing Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[18] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[19] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"ssteerability\" of generative adversarial networks. In arXiv:1907.07171, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Proc. International Conference on Learning Representation, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[23] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[24] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[25] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proc. International Conference on Learning Representation, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[26] Guillaume Lample, Neil Zeghidour, Nicolas Usunier, Antoine Bordes, Ludovic Denoyer, and Marc'Aurelio Ranzato. Fader networks: manipulating images by sliding attributes. In Proc. Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[27] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. ManiGAN: text-guided image manipulation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: a unified selective transfer network for arbitrary image attribute editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[30] Yahui Liu, Marco De Nadai, Deng Cai, Huayang Li, Xavier Alameda-Pineda, Nicu Sebe, and Bruno Lepri. Describe" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "what to change: a text-guided unsupervised image-to-image translation approach. In Proc. ACM International Conference on Multimedia, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[31] M. Mirza and S. Osindero. Conditional generative adversarial nets. In arXiv:1411.1784, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.47, + 0.222 + ], + "angle": 0, + "content": "[32] Seonghyeon Nam, Yunji Kim, and Seon Joo Kim. Text-adaptive generative adversarial networks: manipulating images with natural language. In Proc. Neural Information Processing Systems, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.225, + 0.47, + 0.28 + ], + "angle": 0, + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: text-driven manipulation of StyleGAN imagery. In Proc. International Conference on Computer Vision, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.283, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[34] Alec Redford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[35] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Proc. International Conference on Machine Learning, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.47, + 0.495 + ], + "angle": 0, + "content": "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Deniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.47, + 0.553 + ], + "angle": 0, + "content": "[37] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.47, + 0.612 + ], + "angle": 0, + "content": "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: interpreting the disentangled face representation learned by GANs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2004-2018, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[39] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.712 + ], + "angle": 0, + "content": "[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Cross-domain and disentangled face manipulation with 3D guidance. IEEE Transactions on Visualization and Computer Graphics (Early Access), pages 1-15, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[42] Tianyi Wei, Dongdong Chen, Wenbo Zhou, and Jing Liao. HairCLIP: design your hier by text and reference image. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: disentangled controls for StyleGAN image generation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[44] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: text-guided diverse face image generation and manipulation. In arXiv:2012.03308, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[45] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: fine-grained text to image generation with attentional generative adversarial networks. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: text to photo-realistic image synthesis with stacked generative adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[47] Yutong Zheng, Yu-Kai Huang, Ran Tao, Zhiqiang Shen, and Marios Savvides. Unsupervised disentanglement of linear-encoded facial semantics. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[48] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19294" + } + ] +] \ No newline at end of file diff --git a/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_origin.pdf b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cac2e703a79b1484ab0e94b2b4d3bc817a45150b --- /dev/null +++ b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/4306b93e-f301-4712-887d-3f71b781b5b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeed693ea952997268f262cb80497f2418696a2854874ddf8897c3749861c2d0 +size 7535090 diff --git a/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/full.md b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1f7827506d940ad5d17a220caf6061153e2fbd1 --- /dev/null +++ b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/full.md @@ -0,0 +1,313 @@ +# Text-guided Unsupervised Latent Transformation for Multi-attribute Image Manipulation + +Xiwen Wei $^{1}$ , Zhen Xu $^{1}$ , Cheng Liu $^{2}$ , Si Wu $^{1,3*}$ , Zhiwen Yu $^{1}$ , and Hau San Wong $^{4}$ + +1School of Computer Science and Engineering, South China University of Technology + +$^{2}$ Department of Computer Science, Shantou University + +$^{3}$ Peng Cheng Laboratory + +$^{4}$ Department of Computer Science, City University of Hong Kong + +{202021044777, csxuzhen}@mail.scut.edu.cn, cliu@stu.edu.cn, {cswusi, zhwyu}@scut.edu.cn, cshswong@cityu.edu.hk + +# Abstract + +Great progress has been made in StyleGAN-based image editing. To associate with preset attributes, most existing approaches focus on supervised learning for semantically meaningful latent space traversal directions, and each manipulation step is typically determined for an individual attribute. To address this limitation, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model, which adaptively infers a single transformation step in the latent space of StyleGAN to simultaneously manipulate multiple attributes on a given input image. Specifically, we adopt a two-stage architecture for a latent mapping network to break down the transformation process into two manageable steps. Our network first learns a diverse set of semantic directions tailored to an input image, and later nonlinearly fuses the ones associated with the target attributes to infer a residual vector. The resulting tightly interlinked two-stage architecture delivers the flexibility to handle diverse attribute combinations. By leveraging the cross-modal text-image representation of CLIP, we can perform pseudo annotations based on the semantic similarity between preset attribute text descriptions and training images, and further jointly train an auxiliary attribute classifier with the latent mapping network to provide semantic guidance. We perform extensive experiments to demonstrate that the adopted strategies contribute to the superior performance of TUSLT. + +# 1. Introduction + +Visual attributes represent semantically meaningful features inherent in images, and attribute manipulation has ex + +![](images/caadc3e34e247be75e033a195370d14e78152b20310c2ffb06ea525daab8656f.jpg) +Figure 1. Visually comparing TUSLT with StyleFlow (supervised) and StyleCLIP (text-driven) in precisely manipulating multiple attributes and preserving irrelevant attributes. + +perceived great improvements, due to the advent of Generative Adversarial Network [13] (GAN)-based generative models, e.g. StyleGAN [21, 22] and StarGAN [7, 8]. Recent works [15, 37, 43] have discovered that the latent space of StyleGAN possesses semantic disentanglement properties, enabling a variety of image editing operations via latent transformations. + +StyleGAN-based methods for image attribute manipulation typically involve a large number of manual annotations or well-trained attribute classifiers. Furthermore, the discovered semantic latent directions are associated with individual attributes. The editing on a target attribute is carried out by moving the latent code of an input image along one of the directions. For $K$ target attributes, these models require $K$ transformation steps to handle the translation. As a result, they are not scalable to the increasing number of target attributes in multi-attribute transformation tasks. As shown in Figure 1, we test a state-of-the-art supervised + +![](images/381eb510a2cae551df366df077fe0b8c552d8e3c5c58f68c9db239760106fa78.jpg) +Figure 2. Overview of the proposed model, TUSLT, consisting of two learnable components: an auxiliary attribute classifier $A$ trained on the CLIP-based labeled data, and a latent mapping network $\{\Gamma, \Phi\}$ . $\Gamma$ infers latent directions $\{\gamma_w^{(1)}, \dots, \gamma_w^{(K)}\}$ for preset attributes, and $\Phi$ transforms the target-related directions as indicated by mask $M_{\Delta}$ into a residual vector $\gamma_w^{\Delta}$ , to which the initial latent code is added. Precise multi-attribute transfer is allowed by such a single transformation step, and the generator $G$ synthesizes a new image reflecting the target attributes under the guidance of $A$ and CLIP encoders. + +![](images/50d62be75bf706a5d532f66b202e1df4b4dfb8640a0836bb9e726e47b18c9e37.jpg) + +model, StyleFlow [2], and find that multiple transformation steps lead to undesired deviation from the input image on irrelevant attributes. Compared to the state-of-the-art text-driven model, StyleCLIP [33], we can also achieve a better manipulation result by seeking a single latent transformation step for the task. + +More specifically, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model that supports simultaneous manipulation on multiple attributes. As shown in Figure 2, the key is to jointly learn a mapping network to infer the latent transformation and an auxiliary attribute classifier to assess manipulation quality. We employ the Contrastive Language-Image Pre-training (CLIP) model [34] to generate pseudo-labeled data by measuring the semantic similarities between attribute text descriptions and training images. Compared to CLIP, the jointly trained classifier extracts domain-specific information to better characterize the differences among attributes. This benefits the mapping network to seek more suitable transformations, such that the synthesized images reflect target attributes. Further, we adopt a two-stage architecture for the mapping network: the earlier stage employs a prediction subnetwork to infer a set of semantic directions, and the latter stage operates on the resulting directions and nonlinearly fuses the target-related ones. The intermediate semantic directions are associated with preset attributes and tailored for the input image. This design allows us to deal with a wide range of attribute combinations in a single transformation step. We perform extensive experiments and provide both qualitative and quantitative results in diverse multi-attribute transformation tasks, showing the + +superiority of our model over the competing methods. + +In summary, the main contributions of this work are given as follows: (a) The existing image editing methods focus on discovering semantic latent directions associated with individual visual attributes, and a sequential manipulation process is thus needed for multi-attribute manipulation. In contrast, the proposed model infers a single step of latent space walk to simultaneously manipulate multiple attributes. (b) Benefiting from the cross-modal text-image representation of CLIP, we jointly train a latent mapping network with an auxiliary attribute classifier, which leads to more precise attribute rendering without requiring additional manual annotations. (c) Due to the two-stage nature, our latent mapping network breaks down the challenging multi-attribute manipulation task into sub-tasks: inferring diverse semantic directions and integrating the target-related ones into a single transformation vector. This design gives our model interpretability and flexibility in dealing with a variety of attribute combinations. + +# 2. Related Work + +# 2.1. Generic Image-to-image Translation + +As one of the earliest image translation models, pix2pix [18] learnt a cross-domain mapping via conditional GAN [31]. In addition to an adversarial training loss, the consistency regularization between each input image and the corresponding ground truth was imposed in the training process. To alleviate the problem of pairwise training data, many unpaired image translation models have been developed. UNIT [29] trained two generators to approximate the joint distribution of images from different domains. Cy + +cle consistency regularization was adopted in CycleGAN [48] and discoGAN [23]. To efficiently learn the mappings among multiple domains, StarGAN [7, 9] was a unified framework in which a single generator was trained to translate an input image into different domains, conditioned on domain information. + +# 2.2. Attribute Manipulation + +FaderNet [26] used an encoder-decoder architecture and learnt attribute-aware latent representations in an unsupervised manner. To ensure correct manipulation on the desired attributes, an attribute classification constraint was incorporated in AttGAN [16]. Instead of handling all target attributes, STGAN [28] contained a selective transfer module to only encode the changed attributes, based on the d-difference between source and reference images. In IOAGAN [5], the relationship of preset attributes were leveraged via a Graph Convolutional Network (GCN) [25] to infer and inject an integrated embedding of attributes into a translation network. + +The latent spaces of GANs demonstrate promising semantic organization [10, 12, 19]. With the progress of high-fidelity GAN inversion [14, 36, 40], there are many works on exploring the latent space of a pre-trained StyleGAN generator for image manipulation [3, 41]. SeFa [39] disentangled semantics from StyleGAN by decomposing the matrix of generator weights. In addition, LEFS [47] applied s-parse representation learning to unsupervised semantic disentanglement. As a representative supervised method, InterFaceGAN [38] searched for semantic latent directions by solving a series of binary classification problems about preset attributes. Instead of latent directions, StyleSpace [43] explored style-associated channels with the help of a pretrained classifier. Different from the above linear latent transformation methods, StyleFlow [2] employed a conditional flow model to learn non-linear paths for attribute manipulation. + +# 2.3. Text-guided Image Editing + +Cross-modal representation learning on visual and language data brings about substantive progress in text-guided image synthesis. An early attempt was to train a GAN-based model [31], conditioned on text embedding [35]. Further, a stacked structure [46] and an attention-based regularization approach [45] were designed to improve synthesis quality. In [30, 32], image content from visual attributes were disentangled and associated with text descriptions without using manually annotated data. ManiGAN [27] aimed to learn text and image cross-modality representations, such that semantic regions were associated with the corresponding text via an affine transformation. TediGAN [44] aligned and projected text and image representations into the latent space of StyleGAN. Further, Style + +CLIP [33] integrated CLIP [34] with StyleGAN to leverage CLIP-based linguistic-visual semantic consistency regularization for better manipulation quality. + +Among the aforementioned models, StyleFlow [2] and StyleCLIP [33] are the most relevant to our work. StyleFlow is a supervised method, and the learnt semantic directions are associated with individual attributes. For multi-attribute manipulation, StyleFlow needs a sequential transformation process, in which the target attributes are edited one at a time. In contrast, the proposed model learns to infer a single transformation step to simultaneously manipulate multiple attributes in an unsupervised manner. StyleCLIP trains a dedicated mapping network for each text description, and thus has limited scalability in handling a variety of attribute combinations. Unlike StyleCLIP, our model uses a two-stage mapping network and jointly trains an auxiliary attribute classifier without manual annotations. This structure benefits the scalability and semantic accuracy in changing multiple attributes. + +# 3. Proposed Method + +Our goal is to translate an input image into a new one reflecting the target attributes. Based on the linguistic-visual representation of CLIP, we can pseudo-label training images by measuring the semantic similarities between attribute text descriptions and training images. We believe that an auxiliary attribute classifier can capture the most discriminative information, which complements the role of CLIP encoders to a certain extent. Based on this, the constructed supervision is leveraged by jointly training a latent mapping network and the classifier. The former has a two-stage architecture to infer a diverse set of semantic latent directions followed by selectively integrating target-related ones, and the latter provides semantic guidance. As a result, multi-attribute manipulation can be precisely carried out in a single forward pass. + +# 3.1. Auxiliary Attribute Classifier + +CLIP consists of a text encoder $E^{txt}$ and an image encoder $E^{img}$ , and encodes both types of input into 512-D embedding vectors. Let $T = \{t^{(1)},\dots ,t^{(K)}\}$ denote a set of text prompts, and $t^{(i)}$ describes the $i$ -th preset attribute. To identify the attributes reflected in images, we embed training images and $T$ in the shared embedding space, and measure the semantic similarity as + +$$ +\mathcal {S} ^ {(i)} (x) = \cos \left(E ^ {t x t} \left(t ^ {(i)}\right), E ^ {i m g} (x)\right), \tag {1} +$$ + +where $\cos (\cdot ,\cdot)$ denotes the cosine distance between input vectors. $S^{(i)}(x)$ should be larger when $t^{(i)}$ and $x$ represent the same attribute. At this point, we pseudo-announce training images, and the corresponding label $\tilde{y}$ is defined as + +$$ +\tilde {y} ^ {(i)} = \left\{ \begin{array}{l l} 1, & \text {i f} S ^ {(i)} (x) > \tau , \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {2} +$$ + +where $\tilde{y}^{(i)}$ indicates whether image $x$ reflects attribute $i$ , and $\tau$ is a threshold. Based on the constructed supervision, we train an auxiliary classifier $A$ to capture the visual characteristics of preset attributes. For multi-attribute recognition, the evaluation term is defined as + +$$ +\mathcal {L} ^ {c l s} = \mathbf {E} _ {x} \left[ \sum_ {i = 1} ^ {K} - \tilde {y} ^ {(i)} \log A ^ {(i)} (x) \right], \tag {3} +$$ + +where $A^{(i)}(\cdot)$ represents the predicted probability of an input reflecting attribute $i$ . Different from $E^{img}$ that is trained on generic data, $A$ focuses more on the domain-specific data, and can be expected to learn the most discriminative features to identify preset attributes. + +# 3.2. Latent Mapping Network + +The latent space $\mathcal{W}$ of StyleGAN possesses semantic disentanglement properties, and the works [1, 36, 40] further extend to $\mathcal{W}+$ for better inversion quality. For a given input image $x$ , we adopt the e4e model $E^{sty}$ [40] to predict its latent code $w = E^{sty}(x)$ in the $\mathcal{W}+$ space. + +It is challenging to directly infer the latent transformations for a variety of attribute combinations. To address this problem, we design a two-stage architecture for our latent mapping network. The first stage is based on a direction prediction subnetwork $\Gamma$ that produces latent directions denoted by $\Gamma(w) = \{\gamma_w^{(1)},\dots,\gamma_w^{(K)}\}$ , where $\gamma_w^{(i)}$ associates with the preset attribute $i$ , conditioned on $w$ . At the second stage, a fusion subnetwork $\Phi$ operates on the produced directions. We define a binary vector $\Delta \in \{0,1\}^K$ to indicate target attributes. $\Phi$ learns to integrate the directions as indicated by $\Delta$ , and infer a residual vector $\gamma_w^\Delta$ defined as + +$$ +\gamma_ {w} ^ {\Delta} = \Phi (w, M _ {\Delta} \otimes \Gamma (w)), \tag {4} +$$ + +where the mask $M_{\Delta}$ is constructed by broadcasting $\Delta$ across an array having the same dimensionality as the latent code, and $\otimes$ denotes the Hadamard product. As a result, multi-attribute manipulation can be carried out by simply adding the initial latent code to the residual vector. The generator $G$ of StyleGAN is employed to decode the resulting latent vector as + +$$ +\hat {x} ^ {\Delta} = G \left(w + \alpha \gamma_ {w} ^ {\Delta}\right), \tag {5} +$$ + +where $\alpha$ controls the manipulation strength. Although our mapping network stacks two stages, each stage has access to the latent code of the input image. + +To ensure that the intermediate directions are semantically meaningful, they are required to work directly on the manipulation process of individual attributes. In this case, $\Phi$ should not produce any changes to a single direction as input, and the corresponding invariance loss is formulated + +as follows: + +$$ +\mathcal {L} ^ {s n g l} = \mathbf {E} _ {x} \left[ \sum_ {i = 1} ^ {K} \| \Phi (w, M _ {i} \otimes \Gamma (w)) - \gamma_ {w} ^ {(i)} \| _ {2} \right], \tag {6} +$$ + +where $M_{i}$ denotes the mask for selecting $\gamma_w^{(i)}$ only. This design allows the mapping network to flexibly learn transformations with respect to different attribute combinations. + +# 3.3. Semantic Evaluation + +Target attribute identification. The synthesized image $\hat{x}^{\Delta}$ should properly reflect the target attributes as indicated by $\Delta$ . We leverage both classifier $A$ and CLIP encoders $\{E^{txt}, E^{img}\}$ to impose a target attribute identification loss, and the formulation is expressed as follows: + +$$ +\begin{array}{l} \mathcal {L} ^ {t r g t} = \mathbf {E} _ {(x, \Delta)} \left[ \sum_ {i = 1} ^ {K} - \Delta^ {(i)} \log A ^ {(i)} \left(\hat {x} ^ {\Delta}\right) \right. \tag {7} \\ \left. + \sum_ {i = 1} ^ {K} \Delta^ {(i)} \big (1 - \cos (E ^ {t x t} (t ^ {(i)}), E ^ {i m g} (\hat {x} ^ {\Delta})) \big) \right]. \\ \end{array} +$$ + +In the above equation, the first term evaluates the predictions of $A$ , and the second term measures the semantic consistency between the manipulation result and attribute text descriptions in the embedding space of CLIP. By minimizing $\mathcal{L}^{trgt}$ , our mapping network seeks suitable transformations such that the target attributes can be well reflected in both views of $A$ and CLIP. In this way, we significantly improve the semantic accuracy of the transformation. + +Non-target attribute preservation. Minimizing the above losses does not guarantee that the synthesized images properly preserve the content of the input images while at the same time changing only the part related to the target attributes. To ensure that irrelevant attributes are unchanged before and after transformation, we further impose an attribute-aware consistency regularization on the mapping model. Specifically, we define the attribute-aware representation as + +$$ +f _ {A} ^ {(i)} \left(\hat {x} ^ {\Delta}\right) = \nu^ {(i)} \otimes f _ {A} \left(\hat {x} ^ {\Delta}\right), \tag {8} +$$ + +where $f_{A}(\cdot)$ denotes the attribute classifier features, and $\nu^{(i)}$ represents the weight vector of the head associated with the $i$ -th attribute. With the help of $\nu^{(i)}$ , we can suppress the less informative features and highlight the useful ones due to the reason that the $i$ -th head measures its emergence in images. By modulating the classifier features in this way, the resulting ones capture more specific information, and we thus measure the semantic consistency on non-target attributes as + +$$ +\mathcal {L} ^ {p r s v} = \mathbf {E} _ {(x, \Delta)} \left[ \sum_ {i = 1} ^ {K} \left(1 - \Delta^ {(i)}\right) \left\| f _ {A} ^ {(i)} \left(\hat {x} ^ {\Delta}\right) - f _ {A} ^ {(i)} (x) \right\| _ {2} \right]. \tag {9} +$$ + +Compared to the predictions of $A$ , we find that the features $\{f_A^{(1)},\ldots ,f_A^{(K)}\}$ encode richer information on the attributes. Minimizing $\mathcal{L}^{prsv}$ enforces our model to preserve the non-target characteristics of the input image while manipulating the target attributes faithfully. + +# 3.4. Model Optimization + +By integrating the above aspects of classifier training, mapping regularization and evaluation, the optimization formulation of the learnable components is expressed as follows: + +$$ +\min _ {A} \mathcal {L} ^ {c l s}, +$$ + +$$ +\min _ {\Gamma} \mathcal {L} ^ {t r g t} + \mathcal {L} ^ {p r s v} + \lambda \mathcal {L} ^ {l o c} \tag {10} +$$ + +$$ +\min _ {\Phi} \mathcal {L} ^ {t r g t} + \mathcal {L} ^ {p r s v} + \lambda \mathcal {L} ^ {l o c} + \mu \mathcal {L} ^ {s n g l}, +$$ + +where the local searching term $\mathcal{L}^{loc}$ is defined as + +$$ +\mathcal {L} ^ {l o c} = \mathbf {E} _ {(x, \Delta)} \left[ \max \left(\left\| \gamma_ {w} ^ {\Delta} \right\| _ {2} - \epsilon , 0\right) \right], \tag {11} +$$ + +and $\epsilon$ is a margin. Minimizing $\mathcal{L}^{loc}$ prevents the modified latent code from deviating too far from the initial one. In addition, $\lambda$ and $\mu$ in Eq.(10) are the weighting factors that control the relative importance of the regularization terms, compared to the semantic evaluation terms, respectively. We jointly train $\{A,\Gamma ,\Phi \}$ from scratch to flexibly transform input images by randomly specifying the target attributes. Note that we do not perform a dedicated optimization for individual attribute (combination). In the test stage, both $\Gamma$ and $\Phi$ are fixed, and attribute manipulation is performed in a single forward pass for any given data. + +# 4. Experiments + +In this section, we first describe the test datasets and evaluation setups. Next, we investigate the effect of the main components of TUSLT. We further compare our model against multiple leading methods by performing both user studies and quantitative evaluation. Note that all the experimental results are obtained by applying the model on unseen images during the training phase. + +# 4.1. Experimental Setup + +Datasets and preset attributes CelebA-HQ [20] is a widely used benchmark for facial image editing, and contains 30k high-resolution face images of celebrities. We follow [33] to set 38 attributes on hair style, hair color, expression, gender, age and others. We also use the animal and anime face datasets, AFHQ-cats/dogs [9] and Danbooru AnimeFace [4], with large intra-domain differences to evaluate the proposed model in manipulating 4 and 6 preset attributes, respectively. + +![](images/f45625aa6bbdaf6b92249dd6e48f117675c64d4912a6f28ab7fb6e2248d4ce5d.jpg) +Figure 3. Single-attribute transformation results of StyleCLIP and TUSLT. + +Architecture and hyperparameters. The proposed model consists of three pre-trained networks (CLIP text and image encoders $\{E^{txt}, E^{img}\}$ and an e4e encoder $E^{sty}$ ) and two learnable components (an attribute classifier $A$ and a latent mapping network $\{\Gamma, \Phi\}$ ). We adopt a ResNet-50 [6] for $A$ , and the network architectures of $\Gamma$ and $\Phi$ consist of 4 fully connected layers. The threshold $\tau$ in Eq.(1), the coefficient $\alpha$ in Eq.(5), the weighting factors $\lambda$ and $\mu$ in Eq.(10) and the margin $\epsilon$ in Eq.(11) are set to 0.7, 0.1, 0.8, 1 and 0.01, respectively. The proposed model is trained using the Adam optimizer [24] with a learning rate of 0.5. There are 50k training iterations, and each batch contains 2 images. We empirically find that our model can converge to a good solution without heavy tuning. + +Evaluation protocols. For all competing methods, we use the open source codes or pre-trained models. We assess the diversity and manipulation quality of synthesized images using the Fréchet Inception Distance (FID) [17]. On the other hand, we measure the correctness of attribute manipulation by an independent attribute classifier [28], and report the Target Attribute Recognition Rate (TARR). To quantitatively assess the model performance in irrelevant attribute preservation, we use the metric of IDentity Distance (IDD) before and after transformation in the feature space of a well-trained face recognition network [11]. + +# 4.2. Analysis of Main Components + +Semantically meaningful directions. We first perform an experiment to demonstrate the effectiveness of the prediction subnetwork $\Gamma$ in inferring diverse semantic directions associated with preset attributes. For individual attributes, the manipulation is performed along the direction-s produced by $\Gamma$ . In Figure 3, we visually compare with the main competing method, StyleCLIP, and observe that TUSLT is able to render more precise semantics about the target attributes. + +Precise manipulation on multiple attributes. To assess the fusion subnetwork $\Phi$ , we can feed the directions + +![](images/3291adb8232c594e4b6a839014593648048decc68af0e8e8698d9e9fa911f0fa.jpg) +Figure 4. Ablative experiment results. The combinations of target attributes are (upper) 'Bald +Beard +Happy' and (bottom) 'Happy +Makeup +Bushy-eyebrows'. + +![](images/ffd4fd9530f13bf6b17c561f6fb225f0492760b3d770cff201dd12c1946fac1e.jpg) +Figure 5. Comparison between TUSLT and variants in terms of (left) TARR and (right) IDD. + +$\gamma_{CLIP} = \{\gamma_{CLIP}^{(1)},\dots,\gamma_{CLIP}^{(K)}\}$ learnt by StyleCLIP, and the resulting model is referred to as $\gamma_{CLIP} + \Phi$ . On the other hand, we build another variant TUSLT w/o $A^\prime$ by disabling the attribute classifier. In addition to visualizing the synthesis results in Figure 4, we report the TARR and IDD scores of the variants in Figure 5. $\gamma_{CLIP} + \Phi$ achieves comparable performance with StyleCLIP in terms of TARR, while the IDD value of the former is significantly lower than that of the latter. This confirms the effectiveness of $\Phi$ in identity preservation. The results of TUSLT w/o $A^\prime$ are unsatisfactory due to the lack of domain-specific knowledge learnt from the target data. In contrast, benefiting from $A$ our full model exhibits more precise manipulation ability. + +Irrelevant attribute preservation. It is also an important aspect to preserve the visual characteristics of the input image apart from the target attributes before and after transformation. We construct the attribute-aware features by Eq.(8) and impose the corresponding regularization term $\mathcal{L}^{prsv}$ defined in Eq.(9). We remove $\mathcal{L}^{prsv}$ from the overall training loss to build a variant 'TUSLT w/o $\mathcal{L}^{prsv}$ . In Figures 4-5, we can observe that removing $\mathcal{L}^{prsv}$ leads to more significant changes on the target attributes (a higher TARR score), while at the same time increasing the identity + +![](images/4b00f9769625cda79c7ef3d7aa9d3a62202d7b76a7be055a105be033d2ff1d93.jpg) +Figure 6. Multi-attribute manipulation results of TUSLT and competing methods. Note that we examine the common attributes which all the methods are able to manipulate. + +distance between input and transformed images. The results suggest that our strategy leads to a significant improvement in the maintenance of irrelevant attributes. + +# 4.3. Human Evaluation + +We perform user studies to evaluate TUSLT and a number of main leading methods in multi-attribute manipulation tasks. For InterFaceGAN and StyleFlow, we perfor + +![](images/620becf2732cbef0957e2d9a0e001a1b821bc5f2571eb184673527052373e61e.jpg) +Figure 7. Diverse image synthesis results of TUSLT and StyleCLIP. + +Table 1. The result of ranking TUSLT and competing models in multi-attribute transfer tasks. MA, IAP and VR denote Manipulation Accuracy, Irrelevant Attribute Preservation and Visual Realism, respectively. + +
MethodMAIAPVR
TUSLT (Ours)1.251.311.44
StyleCLIP [33]2.512.442.81
StyleFlow [2]2.633.252.56
InterFaceGAN [38]3.693.313.38
TediGAN [44]4.944.694.81
+ +Table 2. Quantitative comparison of TUSLT and competing methods. + +
MethodFID (↓)IDD (↓)PSNR (↑)SSIM (↑)TARR (↑)
TUSLT (Ours)56.910.4524.920.7587.85
StyleCLIP [33]63.930.4819.240.7185.74
StyleFlow [2]61.530.4621.840.7385.10
InterFaceGAN [38]69.310.5117.920.6881.80
TediGAN [44]58.740.4920.610.6918.93
+ +m multi-step translations to manipulate multiple target attributes. There are 18 attribute combinations, and 10 questions for each one. For each question, given an input image and the target attributes, the options are the images synthesized by the competing methods, and the order is randomly shuffled. A total of 50 validated workers are instructed to rank the synthesized images in terms of manipulation accuracy, irrelevant attribute preservation and visual realism. Table 1 summarizes the average ranking values of the competing methods. TUSLT is able to produce the best transformation results in most cases. We also provide the representative synthesis results of the competing methods in Figure + +6. We observe that TediGAN fails in most of the manipulation tasks. In contrast, TUSLT produces images with more precise manipulation results, compared to the main competing models, StyleFlow and StyleCLIP. We can also find that TUSLT is the only model which is able to successfully manipulate all four attributes in the last example. In Figure 7, we visually compare with StyleCLIP in more tasks, and find that TUSLT yields more natural and plausible images. + +# 4.4.Quantitative Evaluation + +We further perform a quantitative comparison between the proposed approach and competing methods. Note that this experiment involves the common attribute combinations which all of the competing models are able to manipulate. In Table 2, we report the average quantitative comparison results of our TUSLT and the competing methods in terms of FID, TARR and IDD. In addition, we adopt the metrics of Peak Signal-to-Noise Ratio (PSNR) and Structure SIMilarity (SSIM) to measure low to mid-level similarity between the input and synthesized images. The manipulations are performed on 5 attribute combinations. TUSLT outperforms the competing models by a large margin in both manipulation quality and irrelevant attribute preservation. In particular, our model achieves the highest PSNR/TARR score of 24.92/87.85, which is higher than that of the second best method (StyleFlow 21.84/85.10) by about 3 points. We further perform a comparison with a CLIP-based image manipulation method, HairCLIP, which focuses on editing hair color and style. Some representative synthesized images are shown in Figure 8, and the corresponding quantitative assessment on 6 attribute combina + +![](images/8606a651478ccbad7a0f392797425513e4edc50201b2b2e394c035201dc503d0.jpg) +Figure 8. Visual comparison between TUSLT and HairCLIP in hair attribute manipulation. + +Table 3. Quantitative comparison of TUSLT and HairCLIP. + +
MethodFID (↓)PSNR (↑)sCLIP (↑)
TUSLT (Ours)50.1027.140.32
HairCLIP [42]50.8326.310.30
+ +![](images/ef5ce49f1dc0ee7876b13e45e1691b7ee1cf9080d8fa6a2b55081b78399c74cb.jpg) +Figure 9. Synthesis results of TUSLT on AFHQ. + +tions is reported in Table 3. To measure manipulation accuracy, we compute the cosine similarity $s_{CLIP}$ between the attribute text descriptions and synthesized images in the CLIP feature space. HairCLIP performs less satisfactorily in rendering hair color as described in the text prompts. + +Discussion. One can find that the results lead to the same conclusions as our user studies. This is mainly due to the joint training of an auxiliary attribute classifier with CLIP-based supervision, and thus precise semantics of preset attributes can be captured. Another reason is the adopted two-stage architecture for latent transformation. Fusing semantically meaningful latent directions simplify the task of multi-attribute manipulation. This allows the proposed model to flexibly transform the input image to reflect a variety of attribute combinations. + +# 4.5. Results on AFHQ and AnimeFace + +We also show the ability of the proposed model to manipulate multiple attributes on AFHQ-cats/dogs and AnimeFace. For AFHQ, we follow the setting of StyleCLIP to specify preset attribute text descriptions and train TUSLT + +![](images/463808c463bf3da2dec84d32e12dc021b549830883c62b64bfa9287983055bc9.jpg) +Figure 10. Synthesis results of TUSLT on AnimeFace. + +to capture the corresponding semantics. For AnimeFace, we use the same setting as the experiments on CelebA-HQ. The results shown in Figures 9-10 confirm again that our model is scalable to multi-attribute transformations on animal and anime facial images and lead to significant visual modification. + +# 5. Conclusion + +We propose a text-guided unsupervised multi-attribute manipulation model to edit images in a single latent transformation step. Benefiting from the cross-modal image and text representation of CLIP, we can jointly train an auxiliary attribute classifier and a latent mapping network for precise attribute manipulation. We suggest two main reasons why the proposed model is able to successfully manipulate multiple attributes on diverse input images, compared to the leading methods. First, the use of a two-stage architecture enables our model to take care of all preset individual attributes, and thus provide the flexibility to handle diverse attribute combinations. Second, in contrast to CLIP encoders, the classifier learns domain-specific features to identify p-reset attributes, and offers semantic guidance not only for manipulating target attributes, but for preserving irrelevant attributes of the input image. This work significantly increases the scalability of StyleGAN-based image attribute manipulation without causing any manual annotation cost. + +# Acknowledgments + +This work was supported in part by the China Scholarship Council, in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the Natural Science Foundation of Guangdong Province (Project No. 2022A1515011160). + +# References + +[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: how to embed images into the StyleGAN + +latent space? In Proc. International Conference on Computer Vision, 2019. 4 +[2] Rameen Abdal, Peihao Zhu, Niloy J. Mitra, and Peter Wonka. StyleFlow: attribute-conditioned exploration of StyleGAN-Generated images using conditional continuous normalizing flows. ACM Transactions on Graphics, 40(3):1-21, 2021. 2, 3, 7 +[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Transactions on Graphics, 40(4):1-12, 2021. 3 +[4] Gwern Branwen Aaron Gokaslan Anonymous, the Danbooru community. Danbooru2018: A large-scale crowdsourced and tagged anime illustration dataset. https://www.gwern.net/Danbooru2018, January 2019. 5 +[5] Binod Bhattacharai and Tae-Kyun Kim. Inducing optimal attribute representations for conditional GANs. In Proc. European Conference on Computer Vision, 2020. 3 +[6] Yue Chen, Yalong Bai, Wei Zhang, and Tao Mei. Destruction and construction learning for fine-grained image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 5157-5166, 2019. 5 +[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 1, 3 +[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1 +[9] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 3, 5 +[10] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 3 +[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: additive angular margin loss for deep face recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5 +[12] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive image properties. In Proc. International Conference on Computer Vision, 2019. 3 +[13] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 1 +[14] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 3 +[15] Erik Harkonen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: discovering interpretable GAN + +controls. In Proc. Neural Information Processing Systems, 2020. 1 +[16] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. AttGAN: facial attribute editing by only changing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 3 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, and Bernhard Nessler. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Proc. Neural Information Processing Systems, 2017. 5 +[18] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[19] Ali Jahanian, Lucy Chai, and Phillip Isola. On the "ssteerability" of generative adversarial networks. In arXiv:1907.07171, 2019. 3 +[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Proc. International Conference on Learning Representation, 2018. 5 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1 +[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1 +[23] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 3 +[24] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5 +[25] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proc. International Conference on Learning Representation, 2017. 3 +[26] Guillaume Lample, Neil Zeghidour, Nicolas Usunier, Antoine Bordes, Ludovic Denoyer, and Marc'Aurelio Ranzato. Fader networks: manipulating images by sliding attributes. In Proc. Neural Information Processing Systems, 2017. 3 +[27] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. ManiGAN: text-guided image manipulation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3 +[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: a unified selective transfer network for arbitrary image attribute editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3, 5 +[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2 +[30] Yahui Liu, Marco De Nadai, Deng Cai, Huayang Li, Xavier Alameda-Pineda, Nicu Sebe, and Bruno Lepri. Describe + +what to change: a text-guided unsupervised image-to-image translation approach. In Proc. ACM International Conference on Multimedia, 2020. 3 +[31] M. Mirza and S. Osindero. Conditional generative adversarial nets. In arXiv:1411.1784, 2014. 2, 3 +[32] Seonghyeon Nam, Yunji Kim, and Seon Joo Kim. Text-adaptive generative adversarial networks: manipulating images with natural language. In Proc. Neural Information Processing Systems, 2018. 3 +[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: text-driven manipulation of StyleGAN imagery. In Proc. International Conference on Computer Vision, 2021. 2, 3, 5, 7 +[34] Alec Redford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning, 2021. 2, 3 +[35] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Proc. International Conference on Machine Learning, 2016. 3 +[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Deniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3, 4 +[37] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1 +[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: interpreting the disentangled face representation learned by GANs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2004-2018, 2020. 3, 7 +[39] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3 +[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 3, 4 +[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Cross-domain and disentangled face manipulation with 3D guidance. IEEE Transactions on Visualization and Computer Graphics (Early Access), pages 1-15, 2021. 3 +[42] Tianyi Wei, Dongdong Chen, Wenbo Zhou, and Jing Liao. HairCLIP: design your hier by text and reference image. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2022. 8 +[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: disentangled controls for StyleGAN image generation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 3 + +[44] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: text-guided diverse face image generation and manipulation. In arXiv:2012.03308, 2020. 3, 7 +[45] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: fine-grained text to image generation with attentional generative adversarial networks. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2018. 3 +[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: text to photo-realistic image synthesis with stacked generative adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3 +[47] Yutong Zheng, Yu-Kai Huang, Ran Tao, Zhiqiang Shen, and Marios Savvides. Unsupervised disentanglement of linear-encoded facial semantics. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3 +[48] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3 \ No newline at end of file diff --git a/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/images.zip b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..68baf8f98210be35a14324ff97e6f42e46d83dfc --- /dev/null +++ b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17e8dc8bb2518bd4a04e26f1009570df4589c28361bcf46dd95c8e08418eca5e +size 713124 diff --git a/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/layout.json b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c0444a349a3cf6e9af9b0479e12c846fc1df2a2 --- /dev/null +++ b/2023/Text-Guided Unsupervised Latent Transformation for Multi-Attribute Image Manipulation/layout.json @@ -0,0 +1,8691 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 510, + 140 + ], + "type": "text", + "content": "Text-guided Unsupervised Latent Transformation for Multi-attribute Image Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "spans": [ + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "content": "Xiwen Wei" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "content": ", Zhen Xu" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "content": ", Cheng Liu" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "content": ", Si Wu" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "content": ", Zhiwen Yu" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "text", + "content": ", and Hau San Wong" + }, + { + "bbox": [ + 99, + 160, + 494, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 175, + 504, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 175, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 89, + 175, + 504, + 190 + ], + "type": "text", + "content": "1School of Computer Science and Engineering, South China University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 190, + 429, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 190, + 429, + 203 + ], + "spans": [ + { + "bbox": [ + 164, + 190, + 429, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 190, + 429, + 203 + ], + "type": "text", + "content": "Department of Computer Science, Shantou University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 203, + 357, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 203, + 357, + 217 + ], + "spans": [ + { + "bbox": [ + 236, + 203, + 357, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 236, + 203, + 357, + 217 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 217, + 456, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 217, + 456, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 217, + 456, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 217, + 456, + 232 + ], + "type": "text", + "content": "Department of Computer Science, City University of Hong Kong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 233, + 531, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 233, + 531, + 258 + ], + "spans": [ + { + "bbox": [ + 60, + 233, + 531, + 258 + ], + "type": "text", + "content": "{202021044777, csxuzhen}@mail.scut.edu.cn, cliu@stu.edu.cn, {cswusi, zhwyu}@scut.edu.cn, cshswong@cityu.edu.hk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 312, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 289, + 624 + ], + "type": "text", + "content": "Great progress has been made in StyleGAN-based image editing. To associate with preset attributes, most existing approaches focus on supervised learning for semantically meaningful latent space traversal directions, and each manipulation step is typically determined for an individual attribute. To address this limitation, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model, which adaptively infers a single transformation step in the latent space of StyleGAN to simultaneously manipulate multiple attributes on a given input image. Specifically, we adopt a two-stage architecture for a latent mapping network to break down the transformation process into two manageable steps. Our network first learns a diverse set of semantic directions tailored to an input image, and later nonlinearly fuses the ones associated with the target attributes to infer a residual vector. The resulting tightly interlinked two-stage architecture delivers the flexibility to handle diverse attribute combinations. By leveraging the cross-modal text-image representation of CLIP, we can perform pseudo annotations based on the semantic similarity between preset attribute text descriptions and training images, and further jointly train an auxiliary attribute classifier with the latent mapping network to provide semantic guidance. We perform extensive experiments to demonstrate that the adopted strategies contribute to the superior performance of TUSLT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 648, + 128, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 128, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 128, + 660 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "content": "Visual attributes represent semantically meaningful features inherent in images, and attribute manipulation has ex" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 285, + 531, + 430 + ], + "blocks": [ + { + "bbox": [ + 319, + 285, + 531, + 430 + ], + "lines": [ + { + "bbox": [ + 319, + 285, + 531, + 430 + ], + "spans": [ + { + "bbox": [ + 319, + 285, + 531, + 430 + ], + "type": "image", + "image_path": "caadc3e34e247be75e033a195370d14e78152b20310c2ffb06ea525daab8656f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 439, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 473 + ], + "type": "text", + "content": "Figure 1. Visually comparing TUSLT with StyleFlow (supervised) and StyleCLIP (text-driven) in precisely manipulating multiple attributes and preserving irrelevant attributes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 497, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 579 + ], + "type": "text", + "content": "perceived great improvements, due to the advent of Generative Adversarial Network [13] (GAN)-based generative models, e.g. StyleGAN [21, 22] and StarGAN [7, 8]. Recent works [15, 37, 43] have discovered that the latent space of StyleGAN possesses semantic disentanglement properties, enabling a variety of image editing operations via latent transformations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "StyleGAN-based methods for image attribute manipulation typically involve a large number of manual annotations or well-trained attribute classifiers. Furthermore, the discovered semantic latent directions are associated with individual attributes. The editing on a target attribute is carried out by moving the latent code of an input image along one of the directions. For " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " target attributes, these models require " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": " transformation steps to handle the translation. As a result, they are not scalable to the increasing number of target attributes in multi-attribute transformation tasks. As shown in Figure 1, we test a state-of-the-art supervised" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19285" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 399, + 234 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 399, + 234 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 399, + 234 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 399, + 234 + ], + "type": "image", + "image_path": "381eb510a2cae551df366df077fe0b8c552d8e3c5c58f68c9db239760106fa78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed model, TUSLT, consisting of two learnable components: an auxiliary attribute classifier " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " trained on the CLIP-based labeled data, and a latent mapping network " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\{\\Gamma, \\Phi\\}" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " infers latent directions " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\{\\gamma_w^{(1)}, \\dots, \\gamma_w^{(K)}\\}" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " for preset attributes, and " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " transforms the target-related directions as indicated by mask " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "M_{\\Delta}" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " into a residual vector " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "\\gamma_w^{\\Delta}" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": ", to which the initial latent code is added. Precise multi-attribute transfer is allowed by such a single transformation step, and the generator " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " synthesizes a new image reflecting the target attributes under the guidance of " + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 254, + 546, + 310 + ], + "type": "text", + "content": " and CLIP encoders." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 415, + 71, + 547, + 233 + ], + "blocks": [ + { + "bbox": [ + 415, + 71, + 547, + 233 + ], + "lines": [ + { + "bbox": [ + 415, + 71, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 415, + 71, + 547, + 233 + ], + "type": "image", + "image_path": "50d62be75bf706a5d532f66b202e1df4b4dfb8640a0836bb9e726e47b18c9e37.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": "model, StyleFlow [2], and find that multiple transformation steps lead to undesired deviation from the input image on irrelevant attributes. Compared to the state-of-the-art text-driven model, StyleCLIP [33], we can also achieve a better manipulation result by seeking a single latent transformation step for the task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 714 + ], + "type": "text", + "content": "More specifically, we propose a Text-guided Unsupervised StyleGAN Latent Transformation (TUSLT) model that supports simultaneous manipulation on multiple attributes. As shown in Figure 2, the key is to jointly learn a mapping network to infer the latent transformation and an auxiliary attribute classifier to assess manipulation quality. We employ the Contrastive Language-Image Pre-training (CLIP) model [34] to generate pseudo-labeled data by measuring the semantic similarities between attribute text descriptions and training images. Compared to CLIP, the jointly trained classifier extracts domain-specific information to better characterize the differences among attributes. This benefits the mapping network to seek more suitable transformations, such that the synthesized images reflect target attributes. Further, we adopt a two-stage architecture for the mapping network: the earlier stage employs a prediction subnetwork to infer a set of semantic directions, and the latter stage operates on the resulting directions and nonlinearly fuses the target-related ones. The intermediate semantic directions are associated with preset attributes and tailored for the input image. This design allows us to deal with a wide range of attribute combinations in a single transformation step. We perform extensive experiments and provide both qualitative and quantitative results in diverse multi-attribute transformation tasks, showing the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 331, + 523, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 523, + 343 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 523, + 343 + ], + "type": "text", + "content": "superiority of our model over the competing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 344, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 558 + ], + "type": "text", + "content": "In summary, the main contributions of this work are given as follows: (a) The existing image editing methods focus on discovering semantic latent directions associated with individual visual attributes, and a sequential manipulation process is thus needed for multi-attribute manipulation. In contrast, the proposed model infers a single step of latent space walk to simultaneously manipulate multiple attributes. (b) Benefiting from the cross-modal text-image representation of CLIP, we jointly train a latent mapping network with an auxiliary attribute classifier, which leads to more precise attribute rendering without requiring additional manual annotations. (c) Due to the two-stage nature, our latent mapping network breaks down the challenging multi-attribute manipulation task into sub-tasks: inferring diverse semantic directions and integrating the target-related ones into a single transformation vector. This design gives our model interpretability and flexibility in dealing with a variety of attribute combinations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 567, + 392, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 392, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 392, + 580 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 587, + 499, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 499, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 499, + 601 + ], + "type": "text", + "content": "2.1. Generic Image-to-image Translation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "As one of the earliest image translation models, pix2pix [18] learnt a cross-domain mapping via conditional GAN [31]. In addition to an adversarial training loss, the consistency regularization between each input image and the corresponding ground truth was imposed in the training process. To alleviate the problem of pairwise training data, many unpaired image translation models have been developed. UNIT [29] trained two generators to approximate the joint distribution of images from different domains. Cy" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19286" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "cle consistency regularization was adopted in CycleGAN [48] and discoGAN [23]. To efficiently learn the mappings among multiple domains, StarGAN [7, 9] was a unified framework in which a single generator was trained to translate an input image into different domains, conditioned on domain information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 180, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 180, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 180, + 166 + ], + "type": "text", + "content": "2.2. Attribute Manipulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 315 + ], + "type": "text", + "content": "FaderNet [26] used an encoder-decoder architecture and learnt attribute-aware latent representations in an unsupervised manner. To ensure correct manipulation on the desired attributes, an attribute classification constraint was incorporated in AttGAN [16]. Instead of handling all target attributes, STGAN [28] contained a selective transfer module to only encode the changed attributes, based on the d-difference between source and reference images. In IOAGAN [5], the relationship of preset attributes were leveraged via a Graph Convolutional Network (GCN) [25] to infer and inject an integrated embedding of attributes into a translation network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 518 + ], + "type": "text", + "content": "The latent spaces of GANs demonstrate promising semantic organization [10, 12, 19]. With the progress of high-fidelity GAN inversion [14, 36, 40], there are many works on exploring the latent space of a pre-trained StyleGAN generator for image manipulation [3, 41]. SeFa [39] disentangled semantics from StyleGAN by decomposing the matrix of generator weights. In addition, LEFS [47] applied s-parse representation learning to unsupervised semantic disentanglement. As a representative supervised method, InterFaceGAN [38] searched for semantic latent directions by solving a series of binary classification problems about preset attributes. Instead of latent directions, StyleSpace [43] explored style-associated channels with the help of a pretrained classifier. Different from the above linear latent transformation methods, StyleFlow [2] employed a conditional flow model to learn non-linear paths for attribute manipulation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 527, + 195, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 195, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 195, + 540 + ], + "type": "text", + "content": "2.3. Text-guided Image Editing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Cross-modal representation learning on visual and language data brings about substantive progress in text-guided image synthesis. An early attempt was to train a GAN-based model [31], conditioned on text embedding [35]. Further, a stacked structure [46] and an attention-based regularization approach [45] were designed to improve synthesis quality. In [30, 32], image content from visual attributes were disentangled and associated with text descriptions without using manually annotated data. ManiGAN [27] aimed to learn text and image cross-modality representations, such that semantic regions were associated with the corresponding text via an affine transformation. TediGAN [44] aligned and projected text and image representations into the latent space of StyleGAN. Further, Style" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "CLIP [33] integrated CLIP [34] with StyleGAN to leverage CLIP-based linguistic-visual semantic consistency regularization for better manipulation quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 108, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 300 + ], + "type": "text", + "content": "Among the aforementioned models, StyleFlow [2] and StyleCLIP [33] are the most relevant to our work. StyleFlow is a supervised method, and the learnt semantic directions are associated with individual attributes. For multi-attribute manipulation, StyleFlow needs a sequential transformation process, in which the target attributes are edited one at a time. In contrast, the proposed model learns to infer a single transformation step to simultaneously manipulate multiple attributes in an unsupervised manner. StyleCLIP trains a dedicated mapping network for each text description, and thus has limited scalability in handling a variety of attribute combinations. Unlike StyleCLIP, our model uses a two-stage mapping network and jointly trains an auxiliary attribute classifier without manual annotations. This structure benefits the scalability and semantic accuracy in changing multiple attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 308, + 411, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 411, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 411, + 322 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 508 + ], + "type": "text", + "content": "Our goal is to translate an input image into a new one reflecting the target attributes. Based on the linguistic-visual representation of CLIP, we can pseudo-label training images by measuring the semantic similarities between attribute text descriptions and training images. We believe that an auxiliary attribute classifier can capture the most discriminative information, which complements the role of CLIP encoders to a certain extent. Based on this, the constructed supervision is leveraged by jointly training a latent mapping network and the classifier. The former has a two-stage architecture to infer a diverse set of semantic latent directions followed by selectively integrating target-related ones, and the latter provides semantic guidance. As a result, multi-attribute manipulation can be precisely carried out in a single forward pass." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 513, + 466, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 466, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 466, + 526 + ], + "type": "text", + "content": "3.1. Auxiliary Attribute Classifier" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": "CLIP consists of a text encoder " + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "inline_equation", + "content": "E^{txt}" + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": " and an image encoder " + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "inline_equation", + "content": "E^{img}" + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": ", and encodes both types of input into 512-D embedding vectors. Let " + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "inline_equation", + "content": "T = \\{t^{(1)},\\dots ,t^{(K)}\\}" + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": " denote a set of text prompts, and " + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "inline_equation", + "content": "t^{(i)}" + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": " describes the " + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": "-th preset attribute. To identify the attributes reflected in images, we embed training images and " + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 532, + 545, + 616 + ], + "type": "text", + "content": " in the shared embedding space, and measure the semantic similarity as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 619, + 545, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 619, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 347, + 619, + 545, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {(i)} (x) = \\cos \\left(E ^ {t x t} \\left(t ^ {(i)}\\right), E ^ {i m g} (x)\\right), \\tag {1}", + "image_path": "3809d43d31fb66b26614de83da592bfa72c04d47a1a89ac3ebbb702ccaaf7e0a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "content": " denotes the cosine distance between input vectors. " + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "inline_equation", + "content": "S^{(i)}(x)" + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "content": " should be larger when " + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "inline_equation", + "content": "t^{(i)}" + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "content": " represent the same attribute. At this point, we pseudo-announce training images, and the corresponding label " + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 638, + 547, + 686 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 689, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 689, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 361, + 689, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\tilde {y} ^ {(i)} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S ^ {(i)} (x) > \\tau , \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}", + "image_path": "34ae801333cd58b79c5fc669c95980f2fa301b39e9e6451b24a15f0332e6ed02.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19287" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\tilde{y}^{(i)}" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " indicates whether image " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " reflects attribute " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " is a threshold. Based on the constructed supervision, we train an auxiliary classifier " + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 72, + 287, + 131 + ], + "type": "text", + "content": " to capture the visual characteristics of preset attributes. For multi-attribute recognition, the evaluation term is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 140, + 287, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 94, + 140, + 287, + 172 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {c l s} = \\mathbf {E} _ {x} \\left[ \\sum_ {i = 1} ^ {K} - \\tilde {y} ^ {(i)} \\log A ^ {(i)} (x) \\right], \\tag {3}", + "image_path": "492e2514f89c5f86881fdfd8f0dedf39e1144fde9a02a56e3787b5373f5afc18.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "inline_equation", + "content": "A^{(i)}(\\cdot)" + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "text", + "content": " represents the predicted probability of an input reflecting attribute " + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "text", + "content": ". Different from " + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "inline_equation", + "content": "E^{img}" + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "text", + "content": " that is trained on generic data, " + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 183, + 287, + 243 + ], + "type": "text", + "content": " focuses more on the domain-specific data, and can be expected to learn the most discriminative features to identify preset attributes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 190, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 190, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 190, + 266 + ], + "type": "text", + "content": "3.2. Latent Mapping Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": "The latent space " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " of StyleGAN possesses semantic disentanglement properties, and the works [1, 36, 40] further extend to " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " for better inversion quality. For a given input image " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": ", we adopt the e4e model " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "E^{sty}" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " [40] to predict its latent code " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "w = E^{sty}(x)" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 47, + 272, + 287, + 331 + ], + "type": "text", + "content": " space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": "It is challenging to directly infer the latent transformations for a variety of attribute combinations. To address this problem, we design a two-stage architecture for our latent mapping network. The first stage is based on a direction prediction subnetwork " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": " that produces latent directions denoted by " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\Gamma(w) = \\{\\gamma_w^{(1)},\\dots,\\gamma_w^{(K)}\\}" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\gamma_w^{(i)}" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": " associates with the preset attribute " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": ". At the second stage, a fusion subnetwork " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": " operates on the produced directions. We define a binary vector " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\{0,1\\}^K" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": " to indicate target attributes. " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": " learns to integrate the directions as indicated by " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": ", and infer a residual vector " + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "inline_equation", + "content": "\\gamma_w^\\Delta" + }, + { + "bbox": [ + 46, + 332, + 287, + 466 + ], + "type": "text", + "content": " defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 475, + 287, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 475, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 111, + 475, + 287, + 489 + ], + "type": "interline_equation", + "content": "\\gamma_ {w} ^ {\\Delta} = \\Phi (w, M _ {\\Delta} \\otimes \\Gamma (w)), \\tag {4}", + "image_path": "dc61caa3b0ef12bb340d37626a5909d6e525b8e94b8d8955822e6567b68f4843.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "text", + "content": "where the mask " + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "inline_equation", + "content": "M_{\\Delta}" + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "text", + "content": " is constructed by broadcasting " + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "text", + "content": " across an array having the same dimensionality as the latent code, and " + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "text", + "content": " denotes the Hadamard product. As a result, multi-attribute manipulation can be carried out by simply adding the initial latent code to the residual vector. The generator " + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 499, + 287, + 581 + ], + "type": "text", + "content": " of StyleGAN is employed to decode the resulting latent vector as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 592, + 287, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 592, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 123, + 592, + 287, + 607 + ], + "type": "interline_equation", + "content": "\\hat {x} ^ {\\Delta} = G \\left(w + \\alpha \\gamma_ {w} ^ {\\Delta}\\right), \\tag {5}", + "image_path": "4cf47785bbe2f904aad4b0c8a85d08f1d83f2b20af58236952949d6c0fa79665.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 617, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 617, + 287, + 653 + ], + "type": "text", + "content": " controls the manipulation strength. Although our mapping network stacks two stages, each stage has access to the latent code of the input image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "To ensure that the intermediate directions are semantically meaningful, they are required to work directly on the manipulation process of individual attributes. In this case, " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " should not produce any changes to a single direction as input, and the corresponding invariance loss is formulated" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 352, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 352, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 352, + 83 + ], + "type": "text", + "content": "as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 88, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 88, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 88, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {s n g l} = \\mathbf {E} _ {x} \\left[ \\sum_ {i = 1} ^ {K} \\| \\Phi (w, M _ {i} \\otimes \\Gamma (w)) - \\gamma_ {w} ^ {(i)} \\| _ {2} \\right], \\tag {6}", + "image_path": "d91640df1ecb352b78bd4e829d4deca5e9f6281c84a44d6b64e38c5a20cf1a9d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "type": "text", + "content": " denotes the mask for selecting " + }, + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "type": "inline_equation", + "content": "\\gamma_w^{(i)}" + }, + { + "bbox": [ + 305, + 129, + 545, + 166 + ], + "type": "text", + "content": " only. This design allows the mapping network to flexibly learn transformations with respect to different attribute combinations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 172, + 425, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 172, + 425, + 183 + ], + "spans": [ + { + "bbox": [ + 306, + 172, + 425, + 183 + ], + "type": "text", + "content": "3.3. Semantic Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "text", + "content": "Target attribute identification. The synthesized image " + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\hat{x}^{\\Delta}" + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "text", + "content": " should properly reflect the target attributes as indicated by " + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "text", + "content": ". We leverage both classifier " + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "text", + "content": " and CLIP encoders " + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "inline_equation", + "content": "\\{E^{txt}, E^{img}\\}" + }, + { + "bbox": [ + 305, + 190, + 545, + 250 + ], + "type": "text", + "content": " to impose a target attribute identification loss, and the formulation is expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 256, + 545, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 256, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 256, + 545, + 324 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {t r g t} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\sum_ {i = 1} ^ {K} - \\Delta^ {(i)} \\log A ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) \\right. \\tag {7} \\\\ \\left. + \\sum_ {i = 1} ^ {K} \\Delta^ {(i)} \\big (1 - \\cos (E ^ {t x t} (t ^ {(i)}), E ^ {i m g} (\\hat {x} ^ {\\Delta})) \\big) \\right]. \\\\ \\end{array}", + "image_path": "09aefda8baaeaff381d04cc2f9ecb476d86c73fd6ffb1ed92e43f4f7039bfe82.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "text", + "content": "In the above equation, the first term evaluates the predictions of " + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "text", + "content": ", and the second term measures the semantic consistency between the manipulation result and attribute text descriptions in the embedding space of CLIP. By minimizing " + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{trgt}" + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "text", + "content": ", our mapping network seeks suitable transformations such that the target attributes can be well reflected in both views of " + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 329, + 545, + 424 + ], + "type": "text", + "content": " and CLIP. In this way, we significantly improve the semantic accuracy of the transformation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 426, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 532 + ], + "type": "text", + "content": "Non-target attribute preservation. Minimizing the above losses does not guarantee that the synthesized images properly preserve the content of the input images while at the same time changing only the part related to the target attributes. To ensure that irrelevant attributes are unchanged before and after transformation, we further impose an attribute-aware consistency regularization on the mapping model. Specifically, we define the attribute-aware representation as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 368, + 537, + 545, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 537, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 368, + 537, + 545, + 553 + ], + "type": "interline_equation", + "content": "f _ {A} ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) = \\nu^ {(i)} \\otimes f _ {A} \\left(\\hat {x} ^ {\\Delta}\\right), \\tag {8}", + "image_path": "2d2009d34a8dcfbd54025a512b5ad9070fa6942dcf5724e00085f04de6f2b30d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "inline_equation", + "content": "f_{A}(\\cdot)" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": " denotes the attribute classifier features, and " + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\nu^{(i)}" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": " represents the weight vector of the head associated with the " + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": "-th attribute. With the help of " + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\nu^{(i)}" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": ", we can suppress the less informative features and highlight the useful ones due to the reason that the " + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": "-th head measures its emergence in images. By modulating the classifier features in this way, the resulting ones capture more specific information, and we thus measure the semantic consistency on non-target attributes as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 309, + 670, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 670, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {p r s v} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\sum_ {i = 1} ^ {K} \\left(1 - \\Delta^ {(i)}\\right) \\left\\| f _ {A} ^ {(i)} \\left(\\hat {x} ^ {\\Delta}\\right) - f _ {A} ^ {(i)} (x) \\right\\| _ {2} \\right]. \\tag {9}", + "image_path": "d83aa58db95720fd9ff3467b4c6945c67692aad96d70589a7190f03f5b5336f0.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19288" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "text", + "content": "Compared to the predictions of " + }, + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "text", + "content": ", we find that the features " + }, + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "inline_equation", + "content": "\\{f_A^{(1)},\\ldots ,f_A^{(K)}\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "text", + "content": " encode richer information on the attributes. Minimizing " + }, + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{prsv}" + }, + { + "bbox": [ + 46, + 72, + 289, + 135 + ], + "type": "text", + "content": " enforces our model to preserve the non-target characteristics of the input image while manipulating the target attributes faithfully." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 142, + 165, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 142, + 165, + 155 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 165, + 155 + ], + "type": "text", + "content": "3.4. Model Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 160, + 287, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 160, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 287, + 208 + ], + "type": "text", + "content": "By integrating the above aspects of classifier training, mapping regularization and evaluation, the optimization formulation of the learnable components is expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 216, + 130, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 216, + 130, + 234 + ], + "spans": [ + { + "bbox": [ + 88, + 216, + 130, + 234 + ], + "type": "interline_equation", + "content": "\\min _ {A} \\mathcal {L} ^ {c l s},", + "image_path": "881fe80ca95c4cf1ac57a2537bad585ce6b80995533156ddab21aa98766a2d39.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 236, + 287, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 236, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 88, + 236, + 287, + 255 + ], + "type": "interline_equation", + "content": "\\min _ {\\Gamma} \\mathcal {L} ^ {t r g t} + \\mathcal {L} ^ {p r s v} + \\lambda \\mathcal {L} ^ {l o c} \\tag {10}", + "image_path": "19bd75128f29af583cc56778537506e0821c004fe943862e69661999c866ac59.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 256, + 247, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 256, + 247, + 274 + ], + "spans": [ + { + "bbox": [ + 88, + 256, + 247, + 274 + ], + "type": "interline_equation", + "content": "\\min _ {\\Phi} \\mathcal {L} ^ {t r g t} + \\mathcal {L} ^ {p r s v} + \\lambda \\mathcal {L} ^ {l o c} + \\mu \\mathcal {L} ^ {s n g l},", + "image_path": "5d12d5b42670107ec39ccc53cd3cacaa6e1d078b391627d9b13e269ba1febb9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 285, + 244, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 244, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 244, + 298 + ], + "type": "text", + "content": "where the local searching term " + }, + { + "bbox": [ + 47, + 285, + 244, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{loc}" + }, + { + "bbox": [ + 47, + 285, + 244, + 298 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 306, + 287, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 306, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 90, + 306, + 287, + 322 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {l o c} = \\mathbf {E} _ {(x, \\Delta)} \\left[ \\max \\left(\\left\\| \\gamma_ {w} ^ {\\Delta} \\right\\| _ {2} - \\epsilon , 0\\right) \\right], \\tag {11}", + "image_path": "75d66029db20f3399d79e1b9f74094ff18f8209d4d87569b151d461f6a7addac.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " is a margin. Minimizing " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{loc}" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " prevents the modified latent code from deviating too far from the initial one. In addition, " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " in Eq.(10) are the weighting factors that control the relative importance of the regularization terms, compared to the semantic evaluation terms, respectively. We jointly train " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\{A,\\Gamma ,\\Phi \\}" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " from scratch to flexibly transform input images by randomly specifying the target attributes. Note that we do not perform a dedicated optimization for individual attribute (combination). In the test stage, both " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 331, + 287, + 464 + ], + "type": "text", + "content": " are fixed, and attribute manipulation is performed in a single forward pass for any given data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 475, + 128, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 128, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 128, + 488 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "type": "text", + "content": "In this section, we first describe the test datasets and evaluation setups. Next, we investigate the effect of the main components of TUSLT. We further compare our model against multiple leading methods by performing both user studies and quantitative evaluation. Note that all the experimental results are obtained by applying the model on unseen images during the training phase." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 587, + 164, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 164, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 164, + 601 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Datasets and preset attributes CelebA-HQ [20] is a widely used benchmark for facial image editing, and contains 30k high-resolution face images of celebrities. We follow [33] to set 38 attributes on hair style, hair color, expression, gender, age and others. We also use the animal and anime face datasets, AFHQ-cats/dogs [9] and Danbooru AnimeFace [4], with large intra-domain differences to evaluate the proposed model in manipulating 4 and 6 preset attributes, respectively." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 198 + ], + "type": "image", + "image_path": "f45625aa6bbdaf6b92249dd6e48f117675c64d4912a6f28ab7fb6e2248d4ce5d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 210, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 210, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 210, + 545, + 232 + ], + "type": "text", + "content": "Figure 3. Single-attribute transformation results of StyleCLIP and TUSLT." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": "Architecture and hyperparameters. The proposed model consists of three pre-trained networks (CLIP text and image encoders " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\{E^{txt}, E^{img}\\}" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " and an e4e encoder " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "E^{sty}" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": ") and two learnable components (an attribute classifier " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " and a latent mapping network " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\{\\Gamma, \\Phi\\}" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": "). We adopt a ResNet-50 [6] for " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": ", and the network architectures of " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " consist of 4 fully connected layers. The threshold " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " in Eq.(1), the coefficient " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " in Eq.(5), the weighting factors " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " in Eq.(10) and the margin " + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 255, + 545, + 422 + ], + "type": "text", + "content": " in Eq.(11) are set to 0.7, 0.1, 0.8, 1 and 0.01, respectively. The proposed model is trained using the Adam optimizer [24] with a learning rate of 0.5. There are 50k training iterations, and each batch contains 2 images. We empirically find that our model can converge to a good solution without heavy tuning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 422, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 546, + 555 + ], + "type": "text", + "content": "Evaluation protocols. For all competing methods, we use the open source codes or pre-trained models. We assess the diversity and manipulation quality of synthesized images using the Fréchet Inception Distance (FID) [17]. On the other hand, we measure the correctness of attribute manipulation by an independent attribute classifier [28], and report the Target Attribute Recognition Rate (TARR). To quantitatively assess the model performance in irrelevant attribute preservation, we use the metric of IDentity Distance (IDD) before and after transformation in the feature space of a well-trained face recognition network [11]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 563, + 470, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 470, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 470, + 575 + ], + "type": "text", + "content": "4.2. Analysis of Main Components" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "type": "text", + "content": "Semantically meaningful directions. We first perform an experiment to demonstrate the effectiveness of the prediction subnetwork " + }, + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "type": "text", + "content": " in inferring diverse semantic directions associated with preset attributes. For individual attributes, the manipulation is performed along the direction-s produced by " + }, + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 305, + 581, + 545, + 689 + ], + "type": "text", + "content": ". In Figure 3, we visually compare with the main competing method, StyleCLIP, and observe that TUSLT is able to render more precise semantics about the target attributes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Precise manipulation on multiple attributes. To assess the fusion subnetwork " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": ", we can feed the directions" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19289" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 70, + 492, + 216 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 492, + 216 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 492, + 216 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 492, + 216 + ], + "type": "image", + "image_path": "3291adb8232c594e4b6a839014593648048decc68af0e8e8698d9e9fa911f0fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 546, + 240 + ], + "type": "text", + "content": "Figure 4. Ablative experiment results. The combinations of target attributes are (upper) 'Bald +Beard +Happy' and (bottom) 'Happy +Makeup +Bushy-eyebrows'." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 258, + 287, + 377 + ], + "blocks": [ + { + "bbox": [ + 48, + 258, + 287, + 377 + ], + "lines": [ + { + "bbox": [ + 48, + 258, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 48, + 258, + 287, + 377 + ], + "type": "image", + "image_path": "ffd4fd9530f13bf6b17c561f6fb225f0492760b3d770cff201dd12c1946fac1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 389, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 411 + ], + "type": "text", + "content": "Figure 5. Comparison between TUSLT and variants in terms of (left) TARR and (right) IDD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\gamma_{CLIP} = \\{\\gamma_{CLIP}^{(1)},\\dots,\\gamma_{CLIP}^{(K)}\\}" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " learnt by StyleCLIP, and the resulting model is referred to as " + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\gamma_{CLIP} + \\Phi" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " . On the other hand, we build another variant TUSLT w/o " + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "A^\\prime" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " by disabling the attribute classifier. In addition to visualizing the synthesis results in Figure 4, we report the TARR and IDD scores of the variants in Figure 5. " + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\gamma_{CLIP} + \\Phi" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " achieves comparable performance with StyleCLIP in terms of TARR, while the IDD value of the former is significantly lower than that of the latter. This confirms the effectiveness of " + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " in identity preservation. The results of TUSLT w/o " + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "A^\\prime" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " are unsatisfactory due to the lack of domain-specific knowledge learnt from the target data. In contrast, benefiting from " + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 434, + 287, + 592 + ], + "type": "text", + "content": " our full model exhibits more precise manipulation ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Irrelevant attribute preservation. It is also an important aspect to preserve the visual characteristics of the input image apart from the target attributes before and after transformation. We construct the attribute-aware features by Eq.(8) and impose the corresponding regularization term " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{prsv}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " defined in Eq.(9). We remove " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{prsv}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " from the overall training loss to build a variant 'TUSLT w/o " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{prsv}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". In Figures 4-5, we can observe that removing " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{prsv}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " leads to more significant changes on the target attributes (a higher TARR score), while at the same time increasing the identity" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 259, + 545, + 555 + ], + "blocks": [ + { + "bbox": [ + 307, + 259, + 545, + 555 + ], + "lines": [ + { + "bbox": [ + 307, + 259, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 545, + 555 + ], + "type": "image", + "image_path": "4b00f9769625cda79c7ef3d7aa9d3a62202d7b76a7be055a105be033d2ff1d93.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 560, + 545, + 594 + ], + "lines": [ + { + "bbox": [ + 305, + 560, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 560, + 545, + 594 + ], + "type": "text", + "content": "Figure 6. Multi-attribute manipulation results of TUSLT and competing methods. Note that we examine the common attributes which all the methods are able to manipulate." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 615, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 651 + ], + "type": "text", + "content": "distance between input and transformed images. The results suggest that our strategy leads to a significant improvement in the maintenance of irrelevant attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 658, + 417, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 417, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 417, + 671 + ], + "type": "text", + "content": "4.3. Human Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "We perform user studies to evaluate TUSLT and a number of main leading methods in multi-attribute manipulation tasks. For InterFaceGAN and StyleFlow, we perfor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19290" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 542, + 281 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 542, + 281 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 542, + 281 + ], + "type": "image", + "image_path": "620becf2732cbef0957e2d9a0e001a1b821bc5f2571eb184673527052373e61e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 298, + 419, + 309 + ], + "lines": [ + { + "bbox": [ + 173, + 298, + 419, + 309 + ], + "spans": [ + { + "bbox": [ + 173, + 298, + 419, + 309 + ], + "type": "text", + "content": "Figure 7. Diverse image synthesis results of TUSLT and StyleCLIP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 95, + 374, + 239, + 444 + ], + "blocks": [ + { + "bbox": [ + 46, + 329, + 287, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 373 + ], + "type": "text", + "content": "Table 1. The result of ranking TUSLT and competing models in multi-attribute transfer tasks. MA, IAP and VR denote Manipulation Accuracy, Irrelevant Attribute Preservation and Visual Realism, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 374, + 239, + 444 + ], + "lines": [ + { + "bbox": [ + 95, + 374, + 239, + 444 + ], + "spans": [ + { + "bbox": [ + 95, + 374, + 239, + 444 + ], + "type": "table", + "html": "
MethodMAIAPVR
TUSLT (Ours)1.251.311.44
StyleCLIP [33]2.512.442.81
StyleFlow [2]2.633.252.56
InterFaceGAN [38]3.693.313.38
TediGAN [44]4.944.694.81
", + "image_path": "ea2d315285e71c1ca504ace87c1b2406d7382e7d9d40d64f00ed48c74b49ebd1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 478, + 286, + 548 + ], + "blocks": [ + { + "bbox": [ + 46, + 456, + 286, + 476 + ], + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 476 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of TUSLT and competing methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 478, + 286, + 548 + ], + "lines": [ + { + "bbox": [ + 48, + 478, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 286, + 548 + ], + "type": "table", + "html": "
MethodFID (↓)IDD (↓)PSNR (↑)SSIM (↑)TARR (↑)
TUSLT (Ours)56.910.4524.920.7587.85
StyleCLIP [33]63.930.4819.240.7185.74
StyleFlow [2]61.530.4621.840.7385.10
InterFaceGAN [38]69.310.5117.920.6881.80
TediGAN [44]58.740.4920.610.6918.93
", + "image_path": "726b93cd365d0ebf21cf73e735fd0e046e01d9021686a6bf038524b8515e6d55.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "m multi-step translations to manipulate multiple target attributes. There are 18 attribute combinations, and 10 questions for each one. For each question, given an input image and the target attributes, the options are the images synthesized by the competing methods, and the order is randomly shuffled. A total of 50 validated workers are instructed to rank the synthesized images in terms of manipulation accuracy, irrelevant attribute preservation and visual realism. Table 1 summarizes the average ranking values of the competing methods. TUSLT is able to produce the best transformation results in most cases. We also provide the representative synthesis results of the competing methods in Figure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 427 + ], + "type": "text", + "content": "6. We observe that TediGAN fails in most of the manipulation tasks. In contrast, TUSLT produces images with more precise manipulation results, compared to the main competing models, StyleFlow and StyleCLIP. We can also find that TUSLT is the only model which is able to successfully manipulate all four attributes in the last example. In Figure 7, we visually compare with StyleCLIP in more tasks, and find that TUSLT yields more natural and plausible images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 441, + 440, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 440, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 440, + 453 + ], + "type": "text", + "content": "4.4.Quantitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "We further perform a quantitative comparison between the proposed approach and competing methods. Note that this experiment involves the common attribute combinations which all of the competing models are able to manipulate. In Table 2, we report the average quantitative comparison results of our TUSLT and the competing methods in terms of FID, TARR and IDD. In addition, we adopt the metrics of Peak Signal-to-Noise Ratio (PSNR) and Structure SIMilarity (SSIM) to measure low to mid-level similarity between the input and synthesized images. The manipulations are performed on 5 attribute combinations. TUSLT outperforms the competing models by a large margin in both manipulation quality and irrelevant attribute preservation. In particular, our model achieves the highest PSNR/TARR score of 24.92/87.85, which is higher than that of the second best method (StyleFlow 21.84/85.10) by about 3 points. We further perform a comparison with a CLIP-based image manipulation method, HairCLIP, which focuses on editing hair color and style. Some representative synthesized images are shown in Figure 8, and the corresponding quantitative assessment on 6 attribute combina" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19291" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 199 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 199 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 199 + ], + "type": "image", + "image_path": "8606a651478ccbad7a0f392797425513e4edc50201b2b2e394c035201dc503d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 206, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 287, + 229 + ], + "type": "text", + "content": "Figure 8. Visual comparison between TUSLT and HairCLIP in hair attribute manipulation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 253, + 257, + 294 + ], + "blocks": [ + { + "bbox": [ + 59, + 241, + 274, + 251 + ], + "lines": [ + { + "bbox": [ + 59, + 241, + 274, + 251 + ], + "spans": [ + { + "bbox": [ + 59, + 241, + 274, + 251 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of TUSLT and HairCLIP." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 253, + 257, + 294 + ], + "lines": [ + { + "bbox": [ + 78, + 253, + 257, + 294 + ], + "spans": [ + { + "bbox": [ + 78, + 253, + 257, + 294 + ], + "type": "table", + "html": "
MethodFID (↓)PSNR (↑)sCLIP (↑)
TUSLT (Ours)50.1027.140.32
HairCLIP [42]50.8326.310.30
", + "image_path": "001fcb2adcd9e53452249fee873d7d7eabd4cace5e3f7c70af7435975b262668.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 305, + 286, + 419 + ], + "blocks": [ + { + "bbox": [ + 50, + 305, + 286, + 419 + ], + "lines": [ + { + "bbox": [ + 50, + 305, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 50, + 305, + 286, + 419 + ], + "type": "image", + "image_path": "ef5ce49f1dc0ee7876b13e45e1691b7ee1cf9080d8fa6a2b55081b78399c74cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 426, + 256, + 437 + ], + "lines": [ + { + "bbox": [ + 78, + 426, + 256, + 437 + ], + "spans": [ + { + "bbox": [ + 78, + 426, + 256, + 437 + ], + "type": "text", + "content": "Figure 9. Synthesis results of TUSLT on AFHQ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "type": "text", + "content": "tions is reported in Table 3. To measure manipulation accuracy, we compute the cosine similarity " + }, + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "type": "inline_equation", + "content": "s_{CLIP}" + }, + { + "bbox": [ + 46, + 459, + 287, + 518 + ], + "type": "text", + "content": " between the attribute text descriptions and synthesized images in the CLIP feature space. HairCLIP performs less satisfactorily in rendering hair color as described in the text prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 519, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 638 + ], + "type": "text", + "content": "Discussion. One can find that the results lead to the same conclusions as our user studies. This is mainly due to the joint training of an auxiliary attribute classifier with CLIP-based supervision, and thus precise semantics of preset attributes can be captured. Another reason is the adopted two-stage architecture for latent transformation. Fusing semantically meaningful latent directions simplify the task of multi-attribute manipulation. This allows the proposed model to flexibly transform the input image to reflect a variety of attribute combinations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 229, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 229, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 229, + 659 + ], + "type": "text", + "content": "4.5. Results on AFHQ and AnimeFace" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We also show the ability of the proposed model to manipulate multiple attributes on AFHQ-cats/dogs and AnimeFace. For AFHQ, we follow the setting of StyleCLIP to specify preset attribute text descriptions and train TUSLT" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 186 + ], + "type": "image", + "image_path": "463808c463bf3da2dec84d32e12dc021b549830883c62b64bfa9287983055bc9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 192, + 525, + 204 + ], + "lines": [ + { + "bbox": [ + 326, + 192, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 326, + 192, + 525, + 204 + ], + "type": "text", + "content": "Figure 10. Synthesis results of TUSLT on AnimeFace." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 224, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 297 + ], + "type": "text", + "content": "to capture the corresponding semantics. For AnimeFace, we use the same setting as the experiments on CelebA-HQ. The results shown in Figures 9-10 confirm again that our model is scalable to multi-attribute transformations on animal and anime facial images and lead to significant visual modification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 308, + 378, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 378, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 378, + 320 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 328, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 546, + 544 + ], + "type": "text", + "content": "We propose a text-guided unsupervised multi-attribute manipulation model to edit images in a single latent transformation step. Benefiting from the cross-modal image and text representation of CLIP, we can jointly train an auxiliary attribute classifier and a latent mapping network for precise attribute manipulation. We suggest two main reasons why the proposed model is able to successfully manipulate multiple attributes on diverse input images, compared to the leading methods. First, the use of a two-stage architecture enables our model to take care of all preset individual attributes, and thus provide the flexibility to handle diverse attribute combinations. Second, in contrast to CLIP encoders, the classifier learns domain-specific features to identify p-reset attributes, and offers semantic guidance not only for manipulating target attributes, but for preserving irrelevant attributes of the input image. This work significantly increases the scalability of StyleGAN-based image attribute manipulation without causing any manual annotation cost." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 555, + 403, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 555, + 403, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 555, + 403, + 569 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 575, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 659 + ], + "type": "text", + "content": "This work was supported in part by the China Scholarship Council, in part by the National Natural Science Foundation of China (Project No. 62072189), in part by the Research Grants Council of the Hong Kong Special Administration Region (Project No. CityU 11206622), and in part by the Natural Science Foundation of Guangdong Province (Project No. 2022A1515011160)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 671, + 365, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 365, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 365, + 683 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "text", + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2StyleGAN: how to embed images into the StyleGAN" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19292" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "latent space? In Proc. International Conference on Computer Vision, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 150 + ], + "type": "text", + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J. Mitra, and Peter Wonka. StyleFlow: attribute-conditioned exploration of StyleGAN-Generated images using conditional continuous normalizing flows. ACM Transactions on Graphics, 40(3):1-21, 2021. 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 195 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Transactions on Graphics, 40(4):1-12, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 241 + ], + "type": "text", + "content": "[4] Gwern Branwen Aaron Gokaslan Anonymous, the Danbooru community. Danbooru2018: A large-scale crowdsourced and tagged anime illustration dataset. https://www.gwern.net/Danbooru2018, January 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 288, + 275 + ], + "type": "text", + "content": "[5] Binod Bhattacharai and Tae-Kyun Kim. Inducing optimal attribute representations for conditional GANs. In Proc. European Conference on Computer Vision, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 319 + ], + "type": "text", + "content": "[6] Yue Chen, Yalong Bai, Wei Zhang, and Tao Mei. Destruction and construction learning for fine-grained image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, pages 5157-5166, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 321, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 321, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 321, + 288, + 376 + ], + "type": "text", + "content": "[7] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. StarGAN: unified generative adversarial networks for multi-domain image-to-image translation. In Proc. IEEE conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2018. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 377, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 377, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 53, + 377, + 288, + 421 + ], + "type": "text", + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 288, + 465 + ], + "type": "text", + "content": "[9] Yunjoy Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: diverse image synthesis for multiple domains. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, June 2020. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[10] Edo Collins, Raja Bala, Bob Price, and Sabine Susstrunk. Editing in style: uncovering the local semantics of GANs. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: additive angular margin loss for deep face recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 601 + ], + "type": "text", + "content": "[12] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Philip Isola. GANalyze: toward visual definitions of cognitive image properties. In Proc. International Conference on Computer Vision, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[13] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. Neural Information Processing Systems, 2014. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Jinjin Gu, Yujun Shen, and Bolei Zhou. Image processing using multi-code GAN prior. In Proceedings of IEEE conference on computer vision and pattern recognition, pages 3012-3021, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[15] Erik Harkonen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. GANSpace: discovering interpretable GAN" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "controls. In Proc. Neural Information Processing Systems, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[16] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. AttGAN: facial attribute editing by only changing what you want. IEEE Transactions on Image Processing, 28(11):5464-5478, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, and Bernhard Nessler. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Proc. Neural Information Processing Systems, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[18] Philip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[19] Ali Jahanian, Lucy Chai, and Phillip Isola. On the \"ssteerability\" of generative adversarial networks. In arXiv:1907.07171, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In Proc. International Conference on Learning Representation, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 399 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[23] Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jung Kwon Lee, and Jiwon Kim. Learning to discover cross-domain relations with generative adversarial networks. In Proc. International Conference on Machine Learning, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 477 + ], + "type": "text", + "content": "[24] Diederik P. Kingma and Jimmy Lei Ba. Adam: a method for stochastic optimization. In Proc. International Conference on Learning Representation, 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 511 + ], + "type": "text", + "content": "[25] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In Proc. International Conference on Learning Representation, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[26] Guillaume Lample, Neil Zeghidour, Nicolas Usunier, Antoine Bordes, Ludovic Denoyer, and Marc'Aurelio Ranzato. Fader networks: manipulating images by sliding attributes. In Proc. Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 556, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 601 + ], + "type": "text", + "content": "[27] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. ManiGAN: text-guided image manipulation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[28] Ming Liu, Yukang Ding, Min Xia, Xiao Liu, Errui Ding, Wangmeng Zuo, and Shilei Wen. STGAN: a unified selective transfer network for arbitrary image attribute editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 545, + 690 + ], + "type": "text", + "content": "[29] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In Proc. Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[30] Yahui Liu, Marco De Nadai, Deng Cai, Huayang Li, Xavier Alameda-Pineda, Nicu Sebe, and Bruno Lepri. Describe" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19293" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "what to change: a text-guided unsupervised image-to-image translation approach. In Proc. ACM International Conference on Multimedia, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 129 + ], + "type": "text", + "content": "[31] M. Mirza and S. Osindero. Conditional generative adversarial nets. In arXiv:1411.1784, 2014. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 132, + 287, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 132, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 132, + 287, + 175 + ], + "type": "text", + "content": "[32] Seonghyeon Nam, Yunji Kim, and Seon Joo Kim. Text-adaptive generative adversarial networks: manipulating images with natural language. In Proc. Neural Information Processing Systems, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 178, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 178, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 178, + 287, + 221 + ], + "type": "text", + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: text-driven manipulation of StyleGAN imagery. In Proc. International Conference on Computer Vision, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 224, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 224, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 287, + 289 + ], + "type": "text", + "content": "[34] Alec Redford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proc. International Conference on Machine Learning, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 292, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 335 + ], + "type": "text", + "content": "[35] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In Proc. International Conference on Machine Learning, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 392 + ], + "type": "text", + "content": "[36] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Deniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 437 + ], + "type": "text", + "content": "[37] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of GANs for semantic face editing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 440, + 287, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 484 + ], + "type": "text", + "content": "[38] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. InterFaceGAN: interpreting the disentangled face representation learned by GANs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2004-2018, 2020. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "type": "text", + "content": "[39] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in GANs. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[40] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 620 + ], + "type": "text", + "content": "[41] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Cross-domain and disentangled face manipulation with 3D guidance. IEEE Transactions on Visualization and Computer Graphics (Early Access), pages 1-15, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "text", + "content": "[42] Tianyi Wei, Dongdong Chen, Wenbo Zhou, and Jing Liao. HairCLIP: design your hier by text and reference image. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: disentangled controls for StyleGAN image generation. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 308 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[44] Weihao Xia, Yujiu Yang, Jing-Hao Xue, and Baoyuan Wu. TediGAN: text-guided diverse face image generation and manipulation. In arXiv:2012.03308, 2020. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[45] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: fine-grained text to image generation with attentional generative adversarial networks. In Proc. IEEE conference on Computer Vision and Pattern Recognition, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: text to photo-realistic image synthesis with stacked generative adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[47] Yutong Zheng, Yu-Kai Huang, Ran Tao, Zhiqiang Shen, and Marios Savvides. Unsupervised disentanglement of linear-encoded facial semantics. In Proc. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[48] Jun-Yan Zhu, Taesung Park, Philip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. International Conference on Computer Vision, 2017. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19294" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_content_list.json b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ce07920426d9f974099abde034fd3d9f0c2b52a --- /dev/null +++ b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Text-Visual Prompting for Efficient 2D Temporal Video Grounding", + "text_level": 1, + "bbox": [ + 147, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yimeng Zhang $^{1,2}$ , Xin Chen $^{2}$ , Jinghan Jia $^{1}$ , Sijia Liu $^{1}$ , Ke Ding $^{2}$ $^{1}$ Michigan State University, $^{2}$ Applied AI, Intel \n{zhan1853, jiajingh, liusiji5}@msu.edu, \n{xin.chen, ke.ding}@intel.com", + "bbox": [ + 230, + 179, + 740, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we study the problem of temporal video grounding (TVG), which aims to predict the starting/ending time points of moments described by a text sentence within a long untrimmed video. Benefiting from fine-grained 3D visual features, the TVG techniques have achieved remarkable progress in recent years. However, the high complexity of 3D convolutional neural networks (CNNs) makes extracting dense 3D visual features time-consuming, which calls for intensive memory and computing resources. Towards efficient TVG, we propose a novel text-visual prompting (TVP) framework, which incorporates optimized perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. In sharp contrast to 3D CNNs, we show that TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. Further, we propose a Temporal-Distance IoU (TDIoU) loss for efficient learning of TVG. Experiments on two benchmark datasets, Charades-STA and ActivityNet Captions datasets, empirically show that the proposed TVP significantly boosts the performance of 2D TVG (e.g., $9.79\\%$ improvement on Charades-STA and $30.77\\%$ improvement on ActivityNet Captions) and achieves $5 \\times$ inference acceleration over TVG using 3D visual features. Codes are available at Open. Intel.", + "bbox": [ + 75, + 316, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, we have witnessed great progress on temporal video grounding (TVG) [30, 74]. One key to this success comes from the fine-grained dense 3D visual features extracted by 3D convolutional neural networks (CNNs) (e.g., C3D [56] and I3D [3]) since TVG tasks demand spatial-temporal context to locate the temporal interval of the moments described by the text query. However, due to the high cost of the dense 3D feature extraction, most existing TVG models only take these 3D visual features ex", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7481fb8c16d19e42311863a71b265ed5a2287a76e309ce3d7b1024aa2f52b992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 287, + 888, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/779651af2c94ec20721ac8fb8fe00ac25d73520599819a185e8d4f1de6bfdfe9.jpg", + "image_caption": [ + "Figure 1. The architecture and performance comparison among TVG methods: a) 3D TVG methods [14, 16, 18, 34, 43, 60-62, 64, 67, 69, 71, 73], b) 2D TVG methods [1, 7], and c) TVP-based 2D TVG (Ours), d) overall performance comparison. Ours is the most efficient (least inference time) and achieves competitive performance compared to 3D TVG methods. In contrast to existing TVG methods, which utilize dense video features extracted by non-trainable offline 3D CNNs and textual features, our proposed framework utilizes a trainable 2D CNN as the vision encoder to extract features from sparsely-sampled video frames with a universal set of frame-aware visual prompts and adds text prompts in textual feature space for end-to-end regression-based modeling." + ], + "image_footnote": [], + "bbox": [ + 504, + 426, + 689, + 559 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84dd07cd3e3a775994638e3f31ae76ba1aadeebb6da5242c165c28bdef03b367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 424, + 885, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tracted by offline 3D CNNs as inputs instead of co-training during TVG model training.", + "bbox": [ + 496, + 717, + 890, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although models using 3D visual features (that we call '3D methods or models') outperform these using the 2D features (that we call '2D methods or models'), a unique advantage of 2D methods is that extracting 2D visual features can significantly reduce the cost in TVG tasks [14, 15, 30, 34, 61, 62, 69, 74, 75]. An efficient and lightweight solution with reasonable performance is also demanded in computer vision, NLP, and video-language tasks [19, 23, 38, 41, 68, 76-80]. As discussed above, the methods employing 3D video features are challenging to be", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14794", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "employed in practical applications. It thus has significant practical and economic value to develop compact 2D solutions for TVG tasks. In this paper, we ask:", + "bbox": [ + 76, + 90, + 470, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How to advance 2D TVG methods so as to achieve comparable results to 3D TVG methods?", + "bbox": [ + 102, + 156, + 442, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, we propose a novel text-visual prompting (TVP) framework for training TVG models using 2D visual features. As shown in Fig. 1, for existing 2D TVG and 3D TVG methods, they all utilize offline pretrained vision encoders and language encoders to perform feature extraction. In contrast, our proposed TVP framework is end-to-end trainable. Furthermore, benefiting from text-visual prompting and cross-modal pretraining on large-scale image-text datasets, our proposed framework could achieve comparable performance to 3D TVG methods with significant inference time acceleration.", + "bbox": [ + 75, + 204, + 468, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conventionally, TVG methods consist of three stages: ① extracting feature from visual and text inputs; ② multimodal feature fusion; ③ cross-modal modelling. In contrast to conventional methods, TVP incorporates optimized input perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. We apply trainable parameters in the textual features as text prompts and develop a universal set of frame-aware patterns as visual prompts. Specially, we sample a fixed number of frames from a video and optimize text prompts for the input query sentence and a set of visual prompts for frames with different temporal locations during training. During testing, the same set of optimized visual prompts and textual prompts are applied to all test-time videos. We refer readers to Fig. 2 for illustrations of visual prompts and text prompts introduced. To the best of our knowledge, our work makes the first attempt to utilize prompt learning to successfully improve the performance of regression-based TVG tasks using 2D visual features.", + "bbox": [ + 75, + 371, + 470, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to 3D CNNs, 2D CNNs lose spatiotemporal information of the video during feature extraction. Inspired by the success of transformers on the vision-language tasks [9, 22, 35, 44, 47, 54, 55] and the recent application of prompt learning to transformers in both vision and language domains [2, 25, 27, 32, 37, 40], we choose transformer as our base TVG model and propose to utilize prompts to compensate for the lack of spatiotemporal information in 2D visual features. Furthermore, we develop a Temporal-Distance IoU (TDIoU) loss for training our proposed framework. There are two aspects that distinguish our proposed framework from existing works. First, our proposed framework is designed to boost the performance of the regression-based TVG methods utilizing 2D CNNs as the vision encoder, not for transfer learning [2, 21, 26] Second, our proposed framework utilizes 2D CNN to extract visual features from", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f460a3358496b498f2d79bbe0b7fd554ad0d4c9645dbf7a2ffbb32976efa959.jpg", + "image_caption": [ + "(a) Text Prompts", + "Figure 2. Text-visual prompting illustration. (a) Text prompts are directly applied in the feature space. (b) A set of visual prompts are applied to video frames in order." + ], + "image_footnote": [], + "bbox": [ + 529, + 92, + 640, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cdd49bfac571a910a37322476ba91c6f4ee2329484bd92f839f088de39cca0aa.jpg", + "image_caption": [ + "(b) Frame-aware Visual Prompts" + ], + "image_footnote": [], + "bbox": [ + 651, + 92, + 864, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sparsely-sampled video frames, which requires less memory and is easier to be applied in practical applications compared to 3D methods [34,60-62,69,75], especially for long videos. Furthermore, thanks to the compact 2D CNN as the vision encoder, our proposed framework could implement the language encoder and visual encoder co-training for better multimodal feature fusion. In summary, the contributions of this work are unfolded below:", + "bbox": [ + 496, + 323, + 890, + 444 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an effective and efficient framework to train 2D TVG models, in which we leverage TVP (text-visual prompting) to improve the utility of sparse 2D visual features without resorting to costly 3D features. To the best of our knowledge, it is the first work to expand the application of prompt learning for resolving TVG problems. Our method outperforms all of 2D methods and achieves competitive performance to 3D TVG methods.", + "- Technology-wise, we integrate visual prompt with text prompt to co-improve the effectiveness of 2D visual features. On top of that, we propose TDIoU (temporal-distance IoU)-based prompt-model co-training method to obtain high-accuracy 2D TVG models.", + "- Experiment-wise, we show the empirical success of our proposal to boost the performance of 2D TVG on Charades-STA and ActivityNet Captions datasets, e.g., $9.79\\%$ improvement in Charades-STA, and $30.77\\%$ in ActivityNet-Captions together with $5 \\times$ inference time acceleration over 3D TVG methods." + ], + "bbox": [ + 517, + 454, + 890, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 799, + 640, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Temporal Grounding (TVG). The objective of the TVG is to predict the starting/ending time points of target moments within an untrimmed video, which is described by a text sentence. Early TVG solutions [7,14,20,39,62,64,70] mainly employ two-stage \"propose-and-rank\" pipeline: ①", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14795", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Propose: utilize sliding windows or proposal network to generate proposal candidates from the input video. ② Rank: the proposed candidates would be ranked according to the text query, and then the proposal with the highest ranking would be the final prediction decision. In contrast to proposal-based methods, regression-based methods [16, 67, 69] directly predict the starting/ending time points of the target moments without ranking massive proposal candidates. Thus, regression-based methods are much faster than proposal-based methods, which is one reason why our work focuses on the regression-based TVG. Furthermore, reinforcement learning (RL)-based methods formulate the TVG task as a sequence of decisions to make [18, 60]. In particular, they train an agent to control the movement of a window by shifting or scaling. During training, the agent would be rewarded or punished based on whether the window is close to the target moment after an adjustment.", + "bbox": [ + 76, + 90, + 472, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Action Detection (TAD). TAD aims to determine whether predefined actions occur in a video and to predict the corresponding time intervals during which these actions occur [12,13,48,53,56,59,63]. Different from TVG, the input of TAD is only a video. In other words, TAD only requires a semantic understanding of videos. Compared to TAD, TVG is more challenging since it requires a semantic understanding of both videos and natural languages. Furthermore, TVG needs to process the multimodal interaction between videos and natural languages.", + "bbox": [ + 76, + 364, + 472, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Prompting. Prompting has recently achieved great success in the domain of natural language processing [25, 32, 37, 40, 46, 49-52, 58]. Text prompting is a process that leverages a data-agnostic perturbation operation applied to text inputs or their embeddings to improve the performance of the downstream task. The simplest way is to construct an input context template originating from human contemplation [46, 49-51]. Although the manually-crafted context templates are simple and interpretable, they are typically not the optimal input prompts. To tackle this issue, other work has focused on searching the optimal prompting in the discrete input space [25, 52, 58] or in the language model's embedding space [32, 37, 40].", + "bbox": [ + 76, + 518, + 472, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Prompting. Inspired by the idea of prompt learning in NLP [37], visual prompting (VP) was first proposed by Bahng et. al. [2] to reprogram a source vision model (e.g., ImageNet-pretrained classifier) to accomplish downstream target tasks (e.g., CIFAR-10 image classification). VP shares almost the same idea with the model reprogramming technology in the vision domain [4-6, 11, 57, 65, 72, 81], which incorporates a universal input perturbation into testing data so as to improve a desired performance metric, e.g., target task accuracy, robustness, and fairness.", + "bbox": [ + 76, + 717, + 472, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Modal Prompting. Although visual prompting and text prompting have recently attracted much attention, they", + "bbox": [ + 76, + 869, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are under-explored in the multi-modal learning, especially on the temporal video grounding task. The existing works [2, 27, 66] mainly focus on integrating text and visual prompts with the CLIP (Contrastive Language-Image Pre-trained) model to improve downstream tasks with imagery data. The problem of multi-modal prompting in the video understanding task has not been studied. In this paper, we for the first time develop the text-visual prompting technique to improve the performance of temporal video grounding using 2D visual features.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 256, + 599, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin with the problem formulation of regression-based TVG. Then we demonstrate the design of TVP (text-visual prompts) and present the overview of our proposed TVP framework.", + "bbox": [ + 496, + 281, + 893, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 351, + 683, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{v} \\in \\mathbb{R}^{N_{\\mathrm{vid}} \\times C \\times H \\times W}$ be an untrimmed video consisting of a sequence of $N_{\\mathrm{vid}}$ video frames, and $\\mathbf{s} \\in \\mathbb{R}^{N_{\\mathrm{tex}}}$ be a text query consisting of a sequence of $N_{\\mathrm{tex}}$ language tokens. Here, the video-query pair $(\\mathbf{v}, \\mathbf{s})$ belongs to a video-language dataset $\\mathcal{D}$ . Given $\\mathbf{v}$ and $\\mathbf{s}$ , TVG aims to predict the time interval $\\hat{\\mathbf{T}} = (\\hat{t}_{\\mathrm{sta}}, \\hat{t}_{\\mathrm{end}})$ of the target video moments described by the query $\\mathbf{s}$ . The TVG model that fuses the vision-language modalities can be described as:", + "bbox": [ + 496, + 373, + 893, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {T}} = f \\left(g _ {\\mathrm {t e x}} (\\mathbf {s}), g _ {\\mathrm {v i d}} (\\mathbf {v})\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 503, + 890, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ denotes TVG model, and $g_{\\mathrm{vid}}$ and $g_{\\mathrm{tex}}$ represent vision encoder and language encoder, respectively.", + "bbox": [ + 496, + 532, + 893, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. TDIoU Loss Function", + "text_level": 1, + "bbox": [ + 500, + 573, + 699, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventionally, the TVG model can be learned by minimizing the temporal IoU loss $\\mathcal{L}_{\\mathrm{tIoU}}$ defined below:", + "bbox": [ + 500, + 595, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t I o U}} = \\left(1 - \\frac {\\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) \\bigcap \\mathbf {T}}{\\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) \\bigcup \\mathbf {T}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 637, + 890, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where for ease of notation let $\\theta$ denote all the trainable parameters involved in (1), and $\\mathbf{T} = (t_{\\mathrm{sta}}, t_{\\mathrm{end}})$ is the label (i.e., the ground-truth time interval) of the target moment associated with the input video-query pair $(\\mathbf{v}, \\mathbf{s})$ . The rationale behind (2) is to maximize the overlapping between the predicted time interval and its ground truth.", + "bbox": [ + 496, + 688, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, for non-overlapping cases, the temporal IoU loss $\\mathcal{L}_{\\mathrm{tIoU}}$ would encounter a gradient vanishing problem. Inspired by [82], we develop a novel TDIoU (Temporal-Distance IoU) loss for training our proposed TVG models by incorporating the normalized central time point distance and duration difference between the predicted video clips and the target video clips. We elaborate on the proposed loss below.", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14796", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad268c97dbc9581d78d148a1fb367ff2a6e09f3845d89ea4aa24db4a76da923d.jpg", + "image_caption": [ + "Figure 3. Overview of our proposed TVP (text-visual prompting) framework for 2D TVG (temporal video grounding). The whole process contains four phases: 1 Video frame preprocessing: uniformly sample frames from input video and apply a set of frame-aware visual prompts to the sampled frames in order; 2 Feature extraction: 2D CNN extracts features from sampled video frames with visual prompts, and the language encoder extracts textual features. In addition, the visual features would be spatially downsampled and temporally fused by max pooling and mean pooling, respectively. 3 Multimodal feature processing: after spatial downsampling and temporal fusion, the 2D visual features would be integrated into the prompted textual features. 4 Crossmodal fusion: the multimodal features would be processed by a 12-layer transformer encoder, and MLP would predict the starting/ending time points of the target moment." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 877, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distance Loss $\\mathcal{L}_{\\mathrm{dis}}$ . To avoid the gradient vanishing problem caused by the non-overlapping case, we involve distance loss $\\mathcal{L}_{\\mathrm{dis}}$ to directly minimize the normalized central time point distance. In addition, we add a threshold $\\alpha_{1}$ to prevent oscillation in the later training phase. The distance loss is then given by:", + "bbox": [ + 75, + 455, + 472, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i s}} = \\max \\left(\\frac {\\left| \\left(t _ {\\mathrm {s t a}} + t _ {\\mathrm {e n d}}\\right) / 2 - \\left(\\hat {t} _ {\\mathrm {s t a}} + \\hat {t} _ {\\mathrm {e n d}}\\right) / 2 \\right|}{\\left| \\hat {\\mathbf {T}} \\bigcup \\mathbf {T} \\right|}, \\alpha_ {1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 556, + 468, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where recall that $\\mathbf{T} = (t_{\\mathrm{sta}}, t_{\\mathrm{end}})$ , $\\hat{\\mathbf{T}}$ is predicted by the TVG model (1), and we choose $\\alpha_1 = 0.2$ in experiments.", + "bbox": [ + 75, + 614, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Duration Loss $\\mathcal{L}_{\\mathrm{dur}}$ . The introduction of distance loss $\\mathcal{L}_{\\mathrm{dis}}$ avoids the gradient vanishing problem but only considers the central time point distance. Yet, this may not be precise enough. For example, even if the central time points are completely overlapped, the duration of two video clips may not be identical. Inspired by the above, we propose the duration loss:", + "bbox": [ + 75, + 646, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d u r}} = \\max \\left(\\frac {| \\mathbf {T} - \\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) |}{| \\mathbf {T} |}, \\alpha_ {2}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 752, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{2}$ is the precision tolerance threshold and set by 0.4 in our experiments.", + "bbox": [ + 75, + 797, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the proposed Temporal-Distance IoU (TDIoU) loss is given by", + "bbox": [ + 76, + 828, + 468, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {t I o U}} + \\beta_ {1} \\mathcal {L} _ {\\mathrm {d i s}} + \\beta_ {2} \\mathcal {L} _ {\\mathrm {d u r}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 862, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{1} > 0$ and $\\beta_{2} > 0$ are regularization parameters.", + "bbox": [ + 76, + 885, + 447, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Text-Visual Prompt Design", + "text_level": 1, + "bbox": [ + 498, + 455, + 743, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the application of prompts on transformers [2, 21, 36, 37], we propose jointly text-visual prompting to boost the performance of our models, in which prompts are optimized perturbation patterns. To improve data processing efficiency, we uniformly sample video frames from the untrimmed video $\\mathbf{v}$ to obtain $\\mathbf{v}_{\\mathrm{sam}} \\in \\mathbb{R}^{N_{\\mathrm{sam}} \\times C \\times H \\times W}$ , where $N_{\\mathrm{sam}}$ is the number of sampled video frames. In addition, we introduce a set of frame-aware visual prompts $\\delta_{\\mathrm{vp}} \\in \\mathbb{R}^{N_{\\mathrm{sam}} \\times d_{\\mathrm{vp}}}$ in the pixel space of sampled video frames $\\mathbf{v}_{\\mathrm{sam}}$ , and introduce text prompts $\\delta_{\\mathrm{tp}} \\in \\mathbb{R}^{N_{\\mathrm{tp}} \\times d_{\\mathrm{tp}}}$ in the textual feature space. By incorporating video frame sampling and text-visual prompts into the TVG model (1), we obtain:", + "bbox": [ + 496, + 479, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\hat {t} _ {\\mathrm {s t a}}, \\hat {t} _ {\\mathrm {e n d}}\\right) = f \\left(\\delta_ {\\mathrm {t p}}, g _ {\\mathrm {t e x}} (\\mathbf {s}), g _ {\\mathrm {v i d}} \\left(\\mathbf {v} _ {\\mathrm {s a m}} + \\delta_ {\\mathrm {v p}}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 686, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a pre-trained 2D TVG model $f$ , the objective of text-visual prompting (TVP) is to learn a universal set of visual prompts $\\delta_{\\mathrm{vp}}$ and text prompts $\\delta_{\\mathrm{tp}}$ to be integrated into sampled video frames and textual features, respectively. Specially, a set of different visual prompts are applied to uniformly-sampled frames of one untrimmed video in order. During training, only the set of visual prompts and text prompts are updated through backpropagation. During finetuning, prompts are frozen, and the parameters of the TVG model and encoders are updated. During testing, the set of optimized visual prompts and the optimized text prompts are applied to all test-time video-query pairs.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14797", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Framework", + "text_level": 1, + "bbox": [ + 76, + 90, + 204, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the success of transformers in vision-language tasks, we choose ClipBERT [31] as the base model for 2D TVG. Extended from ClipBERT, the input of our regression-based TVG model would be describable sentences and uniformly sampled frames of one untrimmed video as shown in Fig. 3. Then, the predicted starting and ending time points of the target video clip would be model outputs. As described in Algorithm 1, there are four phases of our proposed TVP framework: Video frame preprocessing: We obtain sparsely-sampled frames $\\mathbf{v}_{\\mathrm{sam}}$ from one input untrimmed video $\\mathbf{v}$ , and apply universal frame-aware visual prompts $\\delta_{\\mathrm{vp}}$ on top of frames at the padding location. Feature extraction: 2D vision encoder (first 5 ConvBlock of ResNet-50) $g_{\\mathrm{vid}}$ and language encoder (a trainable word embedding layer) $g_{\\mathrm{tex}}$ would extract features from the prompted frames $\\mathbf{v}_{\\mathrm{sam}}'$ and textual inputs $\\mathbf{s}$ , respectively. Multimodal feature processing: Following the setting of Pixel-BERT [22], the 2D visual features $\\mathbf{Q}_{\\mathrm{vid}}$ are downsampled spatially by a $2 \\times 2$ max-pooling layer and fused temporally by a mean-pooling layer. Then, text prompts $\\delta_{\\mathrm{tp}}$ are integrated into textual features $\\mathbf{Q}_{\\mathrm{tex}}$ . In addition, trainable 2D visual position embeddings $\\mathbf{M}_{2\\mathrm{D}}$ and textual position embeddings $\\mathbf{M}_{\\mathrm{pos}}$ are applied to the processed 2D visual features $\\mathbf{Q}_{\\mathrm{vid}}'$ and prompted textual features $\\mathbf{Q}_{\\mathrm{tex}}'$ , respectively [10, 31]. Afterwards, the processed and position-encoded 2D visual features $\\mathbf{Q}_{\\mathrm{vid}}''$ are flattened and integrated into prompted and position-encoded textual features $\\mathbf{Q}_{\\mathrm{tex}}''$ . Moreover, type embeddings $\\mathbf{M}_{\\mathrm{type}}$ would be added to the integrated multimodal features $\\mathbf{Q}_{\\mathrm{all}}$ to indicate the source type of features. Crossmodal fusion: A 12-layer transformer [10] is utilized for crossmodal fusion on $\\mathbf{Q}_{\\mathrm{all}}$ , and then multilayer perceptron (MLP) ending with sigmoid function is used as the prediction head to process the last-layer crossmodal representation $\\mathbf{Q}_{\\mathrm{CM}}$ of the transformer for generating the predicted starting/ending time points $(\\hat{t}_{\\mathrm{sta}}, \\hat{t}_{\\mathrm{sta}})$ of the target moments described by the text query input.", + "bbox": [ + 76, + 114, + 472, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 688, + 209, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate the effectiveness of our proposed TVP framework on Charades-STA and ActivityNet Captions datasets.", + "bbox": [ + 75, + 714, + 468, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 770, + 253, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. The evaluations are implemented on two standard benchmark datasets for TVG task, Charades-STA [14] and ActivityNet Captions [28]. Tab. 1 summarizes the details of both datasets. Charades-STA dataset contains 6,672 videos and 16,124 text queries in total. The average length of videos is $30.6s$ , and the average length of text query is 7.2 words. The average length of moments cor", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Overview of TVP framework", + "text_level": 1, + "bbox": [ + 500, + 90, + 785, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Input: vision encoder $g_{\\mathrm{vid}}$ , language encoder $g_{\\mathrm{tex}}$ , position embeddings $\\mathbf{M}_{\\mathrm{pos}}$ , 2D position embeddings $\\mathbf{M}_{\\mathrm{2D}}$ , type embeddings $\\mathbf{M}_{\\mathrm{type}}$ , transformer $f$ , prediction head $MLP$ , visual prompts $\\delta_{\\mathrm{vp}}$ , text prompts $\\delta_{\\mathrm{tp}}$", + "Output: Predicted time interval $\\hat{\\mathbf{T}} = (\\hat{t}_{\\mathrm{sta}},\\hat{t}_{\\mathrm{end}})$", + "Phase ①: Video frame preprocessing" + ], + "bbox": [ + 500, + 109, + 890, + 202 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1: $\\mathbf{v}_{\\mathrm{sam}} \\gets$ uniformly sample video frames from an untrimmed video $\\mathbf{v}$", + "2: $\\mathbf{v}_{\\mathrm{sam}}^{\\prime}\\gets$ apply visual prompts $\\delta_{\\mathrm{vp}}$ to the sampled video frames $\\mathbf{v}_{\\mathrm{sam}}$" + ], + "bbox": [ + 509, + 203, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase ②: Feature Extraction", + "text_level": 1, + "bbox": [ + 560, + 263, + 764, + 276 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3: $\\mathbf{Q}_{\\mathrm{vid}} = g_{\\mathrm{vid}}(\\mathbf{v}_{\\mathrm{sam}}^{\\prime})\\gets$ extracting 2D visual features", + "4: $\\mathbf{Q}_{\\mathrm{tex}} = g_{\\mathrm{tex}}(\\mathbf{s})$ ← extracting textual features" + ], + "bbox": [ + 509, + 277, + 870, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase 8: Multimodal feature processing", + "text_level": 1, + "bbox": [ + 560, + 309, + 846, + 321 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "5: $\\mathbf{Q}_{\\mathrm{vld}}^{\\prime}\\gets$ apply spatial downsampling and temporal fusion to 2D visual features $\\mathbf{Q}_{\\mathrm{vld}}$", + "6: $\\mathbf{Q}_{\\mathrm{tex}}^{\\prime}\\gets$ apply text prompts $\\delta_{\\mathrm{tp}}$ to textual features $\\mathbf{Q}_{\\mathrm{tex}}$", + "7: $\\mathbf{Q}_{\\mathrm{vid}}^{\\prime \\prime} \\gets$ add 2D visual position embeddings $\\mathbf{M}_{2\\mathrm{D}}$ on the processed 2D visual features $\\mathbf{Q}_{\\mathrm{vid}}^{\\prime}$", + "8: $\\mathbf{Q}_{\\mathrm{tex}}^{\\prime \\prime}\\gets$ add position embeddings $\\mathbf{M}_{\\mathrm{pos}}$ to prompted textual features $\\mathbf{Q}_{\\mathrm{tex}}^{\\prime}$", + "9: $\\mathbf{Q}_{\\mathrm{all}} \\gets$ integrate the processed and position-encoded textual features $\\mathbf{Q}_{\\mathrm{tex}}^{\\prime \\prime}$ and the processed and position-encoded 2D visual features $\\mathbf{Q}_{\\mathrm{vid}}^{\\prime \\prime}$", + "10: $\\mathbf{Q}_{\\mathrm{all}} + \\mathbf{M}_{\\mathrm{type}} \\gets$ add type embeddings $\\mathbf{M}_{\\mathrm{type}}$ to the integrated multimodal features $\\mathbf{Q}_{\\mathrm{all}}$" + ], + "bbox": [ + 504, + 323, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phase 4: Crossmodal fusion", + "text_level": 1, + "bbox": [ + 560, + 505, + 763, + 517 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "11: $\\mathbf{Q}_{\\mathrm{CM}} = f(\\mathbf{Q}_{\\mathrm{all}} + \\mathbf{M}_{\\mathrm{type}}) \\gets$ implement crossmodal fusion through transformer $f$", + "12: $(\\hat{t}_{\\mathrm{sta}},\\hat{t}_{\\mathrm{end}}) = MLP(\\mathbf{Q}_{\\mathrm{CM}})\\gets$ prediction head generates the predicted time interval according to crossmodal representation $\\mathbf{Q}_{\\mathrm{CM}}$" + ], + "bbox": [ + 504, + 518, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/07cc29e142346abe17768bdda32c71a64dc5e919e7c924d8d80761f55465582d.jpg", + "table_caption": [ + "Table 1. Statistics of TVG benchmark datasets (Charades-STA and ActivityNet Captions datasets)." + ], + "table_footnote": [], + "table_body": "
DatasetCharades-STAActivityNet Captions
DomainIndoor ActivityIndoor/Outdoor Activity
# Videos6,67214,926
Avg. Video Length (second)30.6117.6
# Moments11,76771,953
Avg. Moment Length (second)8.137.1
Vocabulary Size1,30315,505
# Queries16,12471,953
Avg. Query Length (word)7.214.4
", + "bbox": [ + 504, + 651, + 887, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "responding to the text query is 8.1s. Following the same dataset split as [14] for fair comparisons, there are 12,408 video-query pairs for training and 3,720 pairs for testing. ActivityNet Captions dataset contains 14,926 videos and 71,953 text queries in total. The average length of videos is 117.6s, and the average length of text query is 14.4 words.", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14798", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The average length of moments corresponding to the text query is 37.1s. ActivityNet Captions dataset is split into training set, validation set, and testing set in a $2:1:1$ ratio. Since the testing set is withheld for competition, only a training set and two validation sets (val1 and val2) can be accessed publicly. For fair comparisons, we evaluate our proposed framework on val1.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our proposal with 15 baseline methods: ① Proposal-based: CTRL [14], MCN [1], SAP [7], BPNet [62], LPNet [61], QSPN [64], MAN [71]; ② Proposal-free: ABLR [67], DRN [69], CPNet [34], DEBUG [43], ExCL [16], VSLNet [73]; ③ Reinforcement learning: TSP-PRL [60], TripNet [18].", + "bbox": [ + 75, + 198, + 470, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Following [14], we adopt Acc(R@1, IoU=m) as the performance evaluation metric, which represents the percentage accuracy of top-1 predicted moments whose tIoU (temporal IoU) with the ground-truth moment is larger than $m$ . By convention, we consider the following tIoU threshold values $m = \\{0.3, 0.5, 0.7\\}$ .", + "bbox": [ + 75, + 291, + 470, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Crossmodal pretraining setup. Our 2D vision encoder (ResNet-50) is initialized with the weight from gridfeat [24], which can extract effective grid features from visual inputs. In addition, both the language encoder and 12-layer transformer are initialized with the BERT-base model weight [10], which are pretrained on English Wikipedia and BookCorpus [83]. Thanks to the compact 2D vision encoder, TVP (our proposal) is able to directly utilize image-text pairs for end-to-end training. Since the benefits of cross-modal pretraining has been demonstrated by [22, 44, 55], our base model is pretrained on two large-scale image-text datasets, which are Visual Genome Captions [29] and COCO Captions [8]. To be more specific, image-text matching [44, 55] and masked language modeling [10] are employed for cross-modal pretraining.", + "bbox": [ + 75, + 383, + 470, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation setup. For video inputs, we uniformly sample $N_{\\mathrm{sam}}$ frames from a video ( $N_{\\mathrm{sam}} = 48$ for Charades-STA and $N_{\\mathrm{sam}} = 64$ for ActivityNet Captions). In addition, all video frames are resized to have a maximum longer side of 448 with an original aspect ratio, and then the frames are zero-padded to $448 \\times 448$ . The default visual prompt sizes for both dataset are 96. The default text prompt sizes are 10 and 20 for Charades-STA and ActivityNet Captions, respectively. We utilize the first 5 ConvBlocks of ResNet-50 as the 2D vision encoder and a trainable embedding layer as the language encoder for both Charades-STA and ActivityNet Captions datasets. For text queries, all word tokens are maintained after lower-case conversion and tokenization. We use AdamW [42] for end-to-end model training, with $\\beta_{1} = 1.0$ , $\\beta_{2} = 0.1$ , $\\alpha_{1} = 0.2$ , $\\alpha_{2} = 0.4$ . Initial learning rates are $1e - 1$ and $5e - 7$ for prompt training and model finetuning, respectively. In addition, the learning rate linearly decays to 0 with the first $10\\%$ training step for warmup. Our experiments are imple", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89b08d07523e5aae6f26bc9866e7209574c377aaad4c68bf1c22a142c286abeb.jpg", + "table_caption": [ + "Table 2. Performance comparison of different thresholds $m$ on the Charades-STA dataset." + ], + "table_footnote": [], + "table_body": "
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D-23.638.89
ABLR [67]C3D-24.369.01
BPNet [62]C3D55.4638.2520.51
LPNet [61]C3D59.1440.9421.13
QSPN [64]C3D54.7035.6015.80
TSP-PRL [60]C3D-45.4524.75
TripNet [18]C3D54.6438.2916.07
DRN [69]C3D-45.4026.40
CPNet [34]C3D-40.3222.47
DEBUG [43]C3D54.9537.3917.92
ExCL [16]I3D61.5044.122.40
VSLNet [73]I3D64.3047.3130.19
MAN [71]I3D-46.5322.72
2D TVGMCN [1]VGG-17.468.01
SAP [7]VGG-27.4213.36
Ours
TVP-Based 2D TVGBase w/o promptsResNet61.2940.4319.89
Base + Visual Prompts65.3844.3120.22
Base + Text Prompts65.8143.4420.65
Base + Both Prompts65.9244.3921.51
", + "bbox": [ + 503, + 123, + 888, + 358 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06bd71bcaaea49013f64961dd0a2fc0f3baae18be52ed233b39c575dcbd73f04.jpg", + "table_caption": [ + "Table 3. Performance comparison of different thresholds $m$ on the ActivityNet Captions dataset." + ], + "table_footnote": [], + "table_body": "
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D28.7014.00-
BPNet [62]C3D59.9842.0724.69
LPNet [61]C3D64.2945.9225.39
QSPN [64]C3D45.3027.7013.60
TSP-PRL [60]C3D56.0238.83-
TripNet [18]C3D48.4232.1913.93
DRN [69]C3D-45.4524.36
CPNet [34]C3D-40.5621.63
ABLR [67]C3D55.6736.79-
DEBUG [43]C3D55.9139.72-
ExCL [16]C3D63.0043.6024.10
VSLNet [73]C3D63.1643.2226.16
Ours
TVP-Based 2D TVGBase w/o promptsResNet57.2040.1619.14
Base + Visual Prompts60.1243.3923.71
Base + Text Prompts60.4842.5824.39
Base + Both Prompts60.7143.4425.03
", + "bbox": [ + 503, + 405, + 888, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mented in PyTorch [45], and models and prompts are finetuned separately for 12 epochs with the mixed precision on 8 NVIDIA V100 GPUs.", + "bbox": [ + 498, + 632, + 893, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiment Results", + "text_level": 1, + "bbox": [ + 498, + 686, + 689, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of TVP on Charades-STA. The performance comparisons with SOTA methods on the Charades-STA dataset are summarized in Tab. 2. Our proposed TVP framework can achieve competitive performance at all IoU thresholds $m$ in the case of utilizing 2D visual features extracted by ResNet-50, and reach the highest score at $m = 0.3$ . Compared to the 2D TVG methods using VGG as the vision encoder, our proposed framework could achieve around $2.5 \\times$ and $2.7 \\times$ performance gain at thresholds 0.5 and 0.7, respectively. Furthermore, we can find that for our base model only one of visual prompts and text prompts can achieve up to $7.37\\%$ and $9.60\\%$ improvement", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14799", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f930cf36323498ab1ee69cd319022a82926fb4450ac7405b8a56d5e204390eb4.jpg", + "image_caption": [ + "(a) Charades-STA" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 272, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f9010f95a5a2247a77be25bc98fd2db502eaf9d81d99450cd3a28ae2a91b6da.jpg", + "image_caption": [ + "(b) ActivityNet Captions", + "Figure 4. Impact of sampled frame numbers." + ], + "image_footnote": [], + "bbox": [ + 272, + 89, + 464, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "at tIoU thresholds $m = 0.3$ and $m = 0.5$ . The combination of text and visual prompts can not only achieve $7.55\\%$ and $9.79\\%$ improvements at tIoU thresholds $m = 0.3$ and $m = 0.5$ , but also improve the performance by $8.14\\%$ at $m = 0.7$ . This demonstrates the effectiveness and necessity of the joint text-visual prompting.", + "bbox": [ + 75, + 321, + 468, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of TVP on ActivityNet Captions. We focus on the performance comparisons with 3D TVG methods on ActivityNet since there are no results of 2D TVG method reported on ActivityNet Captions. The results of multiple methods on ActivityNet Captions datasets are reported in Tab. 3. Even on this more challenging dataset, our proposed method still has achieved competitive performance compared to 3D TVG methods. Different from the performance of TVP on Charades-STA dataset, text prompts or visual prompts can achieve a significant performance boost on the base model over all IoU thresholds $m$ alone (5.73% at $m = 0.3$ , 8.04% at $m = 0.5$ , 27.43% at $m = 0.7$ ), and the text-visual prompt combination could further boost the performance (6.14% at $m = 0.3$ , 8.17% at $m = 0.5$ , 30.77% at $m = 0.7$ ). It is worth noting that the performance gap over $m = 0.7$ between 2D TVG methods and 3D TVG methods is narrowed significantly.", + "bbox": [ + 75, + 414, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, through the experimental results on Charades-STA and ActivityNet Captions datasets, we can find that our proposed TVP framework could achieve competitive performance overall tIoU thresholds on Charades-STA and ActivityNet Captions by improving the utility of sparse 2D visual features. Thanks to the lightweight 2D vision encoder, cotraining language encoder and vision encoder on large-scale image-text datasets can be performed, which benefits the base model to achieve good performance. Furthermore, the combination of text and visual prompts can achieve better results than any single kind of prompts on both datasets, which again proves the importance of cross-modal training.", + "bbox": [ + 75, + 672, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video frame sampling effect. Fig. 4 demonstrates the performance of base model with different number $N_{\\mathrm{sam}}$", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1097b193d2d816f4610a418a81bff2571d8a2b54ea3cddcf4e10c0c45f2faa75.jpg", + "table_caption": [ + "Table 4. The performance comparison of different visual prompt sizes on Charades-STA dataset." + ], + "table_footnote": [], + "table_body": "
Visual Prompt SizeAcc(R@1, IoU=m)Prompt + Frame
m=0.3m=0.5m=0.7
061.2940.4319.89
1661.2940.4320.00
3261.9439.7819.35
4863.6642.3720.00
7263.8743.6619.78
9665.3844.3120.22
12864.7343.6619.78
", + "bbox": [ + 504, + 123, + 887, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/97c465ff5b852083d1feaf2013a797705d1dd8054e12e453b90a6a9223322e53.jpg", + "table_caption": [ + "Table 5. The performance comparison of different text prompt sizes on Charades-STA dataset." + ], + "table_footnote": [], + "table_body": "
Text Prompt SizeAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
057.2040.1619.14
565.3841.9420.43
1065.8143.4420.65
1565.5943.2321.29
2064.9543.8721.51
2563.6642.8020.65
3064.4642.6320.51
", + "bbox": [ + 513, + 353, + 879, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of sampled video frames as visual inputs. For Charades dataset, the base model performance keeps increasing before $N_{\\mathrm{sam}}$ reaches 48, but when it exceeds 48, performance starts to degrade. This is because frequent background changes harm the performance of object re-identification in videos, which are noisy for object motion analysis [17].", + "bbox": [ + 496, + 517, + 890, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For ActivityNet Caption dataset, base model performance continues to improve even when sampled frame number $N_{\\mathrm{sam}}$ exceeds 48, due to the longer average video length in ActivityNet Captions dataset. Balancing the frame number and batch size for training, we choose $N_{\\mathrm{sam}} = 64$ for ActivityNet Captions.", + "bbox": [ + 496, + 609, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TVP performance vs. prompt size. As shown in Tab. 4, we can find that when visual prompts are small, they cannot bring changes to the base model, and when visual prompts are too large, the performance starts to decrease. This is because key information within video frames might be removed. However, the text prompts can bring significant performance boost even when the text prompt size is small as shown in Tab. 5, which is because the textual features has a smaller dimension compared to visual features, and also the text prompts are directly optimized in feature space during training.", + "bbox": [ + 496, + 702, + 890, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TVP performance vs. visual prompt operation. Visual prompt is first proposed by [2], where visual prompts are", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14800", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/97f634bbd4d952f565a1efa47b8cb8cb54418863533f55c20640de4ff5bc7b3f.jpg", + "table_caption": [ + "Table 6. The performance comparison of different visual prompt operations ('remove', 'add', 'replace') with fixed visual prompt size $p = 96$ on Charades-STA and ActivityNet Captions datasets." + ], + "table_footnote": [], + "table_body": "
OperationCharades-STAActivityNet Captions
R@1, IoU=mR@1, IoU=m
m=0.3m=0.5m=0.7m=0.3m=0.5m=0.7
Original61.2940.4319.8957.2040.1619.14
Remove61.2940.4320.057.2040.1619.14
Add61.0839.5720.2257.1540.1619.27
Replace65.3844.3120.2260.1243.3923.71
", + "bbox": [ + 81, + 135, + 464, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a82c52377ad36275e67254268d4a39d76b11b976001d259d05775ede79124c0f.jpg", + "image_caption": [ + "(a) Vison Encoders", + "Figure 5. Inference time comparison. (a) inference time comparison between 2D vision encoder (ResNet-50) and 3D vision encoder (C3D). (b) inference time comparison between the vision encoder and the other modules of the 2D TVG model, where the sampled frame number for our TVP framework is $1.2 \\times$ the length of the video in seconds." + ], + "image_footnote": [], + "bbox": [ + 81, + 247, + 267, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3623d7cc6fbc44db6248b6619f8e1b6fcc50e40ae5a1a400036e943477297a12.jpg", + "image_caption": [ + "(b) Modules in 2D TVG Model" + ], + "image_footnote": [], + "bbox": [ + 269, + 247, + 460, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "added to the image for transfer learning on classification tasks. In contrast, our proposed prompting framework is designed to compensate for the spatiotemporal information loss in 2D visual features. Due to the differences in the task, we try two different prompt operation strategies, 'replace' and 'add'. 'add' is to add the visual prompts to the pixel value of the video frame at the corresponding padding locations. 'replace' is to replace the pixel values of video frames with visual prompts at corresponding padding locations. 'remove' is in order to study the impact of removing the pixel values at the padding location. As shown in Tab. 6, 'add' or 'remove' prompt operations have limited effects on the base model. However, 'replace' does boost the base model performance.", + "bbox": [ + 75, + 537, + 467, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TVP achieves inference efficiency. As shown in Fig. 5, we can find that the inference time required for visual feature extraction accounts for more than half of the inference time of the whole model, while the inference time required for the 3D vision encoder is more than $5 \\times$ compared to the 2D vision encoder, and even more than the time required for the whole TVG model using 2D vision encoder, which fully demonstrates the feasibility of accelerating the overall inference speed by reducing the complexity of the vision encoder. Need to note that if there are", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/815dfebc92ff3a1e6aeadd548cf3008821ddf94512aa7f7398ed1a4898f72db4.jpg", + "table_caption": [ + "Table 7. The performance comparison of different loss designs on Charades-STA dataset." + ], + "table_footnote": [], + "table_body": "
Loss Function SelectionR@1, IoU=m
m=0.3m=0.5m=0.7
LtIoU55.0529.8911.82
LtIoU + Ldis60.6431.1816.77
LtIoU + Ldur59.7830.9716.34
LtIoU + Ldis + Ldur61.2940.4319.89
", + "bbox": [ + 524, + 123, + 867, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c51b7edc18b9108057885efbb83f8ceb6497e9223bbd4bf91132b1a0c4bdcb4d.jpg", + "image_caption": [ + "Figure 6. Loss landscape visualization in 2D plane: Finetuning w/o prompts (left) and using prompts (right); see [33] for implementation." + ], + "image_footnote": [], + "bbox": [ + 506, + 227, + 686, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c473b2558978f69d281dc660c61bf9dac140ef26ca04348f7cc3e71d31bac6cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 228, + 885, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "multiple model weights for different sampled frame number settings and model weights can be adopted adaptively for different lengths of videos, the inference speed for short videos should increase, and the prediction results for long videos will be further improved.", + "bbox": [ + 496, + 362, + 890, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies. Through Tab. 7, we can find that the addition of either distance loss $\\mathcal{L}_{\\mathrm{dis}}$ or duration loss $\\mathcal{L}_{\\mathrm{dur}}$ will result in a performance increase, but the combination of the two will result in a significant performance increase (11.34% at $m = 0.3$ , 35.26% at $m = 0.5$ , 68.27% at $m = 0.7$ ), especially over tIoU thresholds $m = 0.5$ and $m = 0.7$ . This demonstrates that distance loss $\\mathcal{L}_{\\mathrm{dis}}$ and duration loss $\\mathcal{L}_{\\mathrm{dur}}$ could provide more precise training guides compared to only using temporal IoU loss $\\mathcal{L}_{\\mathrm{tIoU}}$ . Furthermore, we posit that prompting may encode additional spatial-temporal supervision to help the model trainer to escape from bad local optima as shown in Fig. 6, where finetuning w/ prompts yields a flatter loss landscape than the one w/o prompts.", + "bbox": [ + 496, + 438, + 892, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose text-visual prompting to boost the performance of 2D TVG methods by compensating for the lack of spatiotemporal information in 2D visual features. In contrast to 3D TVG methods, TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. The effectiveness of our proposed TVP (text-visual prompting) framework has been demonstrated on two standard datasets, Charades-STA and ActivityNet. Our models outperform all 2D models significantly, and also achieve comparable performance to 3D models. What is more, we achieve over $5 \\times$ inference speedup over TVG methods of using 3D visual features.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 6", + "[2] H Bahng, A Jahanian, S Sankaranarayanan, and P Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, page 2022, 2022. 2, 3, 4, 7", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1", + "[4] Aochuan Chen, Peter Lorenz, Yuguang Yao, Pin-Yu Chen, and Sijia Liu. Visual prompting for adversarial robustness. arXiv preprint arXiv:2210.06284, 2022. 3", + "[5] Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. arXiv preprint arXiv:2211.11635, 2022. 3", + "[6] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. arXiv:2202.10629, 2022.3", + "[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8199-8206, 2019. 1, 2, 6", + "[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6", + "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 6", + "[11] Gamaleldin F Elsayed, Ian Goodfellow, et al. Adversarial reprogramming of neural networks. arXiv:1806.11146, 2018. 3", + "[12] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3", + "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3", + "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yanjun Gao, Lulu Liu, Jason Wang, Xin Chen, Huayan Wang, and Rui Zhang. Evoquer: Enhancing temporal grounding with video-pivoted backquery generation. arXiv preprint arXiv:2109.04600, 2021. 1", + "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 1, 3, 6", + "[17] Xinqian Gu, Hong Chang, Bingpeng Ma, Hongkai Zhang, and Xilin Chen. Appearance-preserving 3d convolution for video-based person re-identification. In ECCV, 2020. 7", + "[18] Meera Hahn, Asim Kadav, James M Rehg, and Hans Peter Graf. Tripping through time: Efficient localization of activities in videos. arXiv preprint arXiv:1904.09936, 2019. 1, 3, 6", + "[19] Xiaochen Han, Bo Wu, Zheng Shou, Xiao-Yang Liu, Yi-meng Zhang, and Linghe Kong. Tensor fista-net for real-time snapshot compressive imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10933-10940, 2020. 1", + "[20] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 2", + "[21] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4", + "[22] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. arXiv preprint arXiv:2004.00849, 2020. 2, 5, 6", + "[23] Jinghan Jia, Mingyi Hong, Yimeng Zhang, Mehmet Akçakaya, and Sijia Liu. On the robustness of deep learning-based mri reconstruction to image transformations. arXiv preprint arXiv:2211.04930, 2022. 1", + "[24] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10267-10276, 2020. 6", + "[25] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 2, 3", + "[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 2", + "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. arXiv preprint arXiv:2210.03117, 2022. 2, 3", + "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14802", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6", + "[30] Xiaohan Lan, Yitian Yuan, Xin Wang, Zhi Wang, and Wenwu Zhu. A survey on temporal sentence grounding in videos. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 2021. 1", + "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 5", + "[32] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3", + "[33] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 2018. 8", + "[34] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 1, 2, 6", + "[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2", + "[36] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4", + "[37] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4", + "[38] Yanyu Li, Pu Zhao, Geng Yuan, Xue Lin, Yanzhi Wang, and Xin Chen. Pruning-as-search: Efficient neural architecture search via channel pruning and structural reparameterization. arXiv preprint arXiv:2206.01198, 2022. 1", + "[39] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2", + "[40] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 2, 3", + "[41] Xiao-Yang Liu, Yimeng Zhang, Yukang Liao, and Ling Jiang. Dynamic updating of the knowledge base for a large-scale question answering system. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 2020. 1", + "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 1, 6", + "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2, 6", + "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 2019. 6", + "[46] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3", + "[47] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2", + "[48] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In proceedings of the IEEE International Conference on Computer Vision, 2017. 3", + "[49] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 3", + "[50] Timo Schick and Hinrich Schütze. Few-shot text generation with pattern-exploiting training. arXiv preprint arXiv:2012.11926, 2020. 3", + "[51] Timo Schick and Hinrich Schütze. It's not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118, 2020. 3", + "[52] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3", + "[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 2014. 3", + "[54] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019.2", + "[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2, 6", + "[56] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3", + "[57] Yun-Yun Tsai et al. Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. arXiv:2007.08714, 2020. 3" + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14803", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3", + "[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 3", + "[60] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12386–12393, 2020. 1, 2, 3, 6", + "[61] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. arXiv preprint arXiv:2109.10678, 2021. 1, 2, 6", + "[62] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 1, 2, 6", + "[63] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), 2018. 3", + "[64] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9062–9069, 2019. 1, 2, 6", + "[65] Chao-Han Huck Yang, Yun-Yun Tsai, et al. Voice2series: Reprogramming acoustic models for time series classification. In ICML. PMLR, 2021. 3", + "[66] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3", + "[67] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 1, 3, 6", + "[68] Ofir Zafrir, Ariel Larey, Guy Boudoukh, Haihao Shen, and Moshe Wasserblat. Prune once for all: Sparse pre-trained language models. arXiv preprint arXiv:2111.05754, 2021. 1", + "[69] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 1, 2, 3, 6", + "[70] Yawen Zeng, Da Cao, Xiaochi Wei, Meng Liu, Zhou Zhao, and Zheng Qin. Multi-modal relational graph for cross-modal video moment retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2215-2224, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 6", + "[72] Guanhua Zhang, Yihua Zhang, Yang Zhang, Wenqi Fan, Qing Li, Sijia Liu, and Shiyu Chang. Fairness reprogramming. arXiv preprint arXiv:2209.10222, 2022.3", + "[73] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 1, 6", + "[74] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071, 2022. 1", + "[75] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 1, 2", + "[76] Yimeng Zhang, Akshay Karkal Kamath, Qiucheng Wu, Zhiwen Fan, Wuyang Chen, Zhangyang Wang, Shiyu Chang, Sijia Liu, and Cong Hao. Data-model-circuit tri-design for ultra-light video intelligence on edge devices. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 745-750, 2023. 1", + "[77] Yimeng Zhang, Xiao-Yang Liu, Bo Wu, and Anwar Walid. Video synthesis via transform-based tensor neural network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2454-2462, 2020. 1", + "[78] Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jinfeng Yi, Mingyi Hong, Shiyu Chang, and Sijia Liu. How to robustify black-box ml models? a zeroth-order optimization perspective. arXiv preprint arXiv:2203.14195, 2022. 1", + "[79] Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Mingyi Hong, Yanzhi Wang, and Sijia Liu. Advancing model pruning via bi-level optimization. arXiv preprint arXiv:2210.04092, 2022. 1", + "[80] Yihua Zhang, Guanhua Zhang, Prashant Khanduri, Mingyi Hong, Shiyu Chang, and Sijia Liu. Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In International Conference on Machine Learning, pages 26693-26712. PMLR, 2022. 1", + "[81] Yang Zheng, Xiaoyi Feng, et al. Why adversarial reprogramming works, when it fails, and how to tell the difference. arXiv:2108.11673, 2021. 3", + "[82] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020. 3", + "[83] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19-27, 2015. 6" + ], + "bbox": [ + 501, + 92, + 890, + 896 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14804", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_model.json b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed240cdfdea8f10d743c6854201b7a2dd5873723 --- /dev/null +++ b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_model.json @@ -0,0 +1,2719 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.131, + 0.825, + 0.154 + ], + "angle": 0, + "content": "Text-Visual Prompting for Efficient 2D Temporal Video Grounding" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.741, + 0.254 + ], + "angle": 0, + "content": "Yimeng Zhang\\(^{1,2}\\), Xin Chen\\(^{2}\\), Jinghan Jia\\(^{1}\\), Sijia Liu\\(^{1}\\), Ke Ding\\(^{2}\\) \n\\(^{1}\\)Michigan State University, \\(^{2}\\)Applied AI, Intel \n{zhan1853, jiajingh, liusiji5}@msu.edu, \n{xin.chen, ke.ding}@intel.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.712 + ], + "angle": 0, + "content": "In this paper, we study the problem of temporal video grounding (TVG), which aims to predict the starting/ending time points of moments described by a text sentence within a long untrimmed video. Benefiting from fine-grained 3D visual features, the TVG techniques have achieved remarkable progress in recent years. However, the high complexity of 3D convolutional neural networks (CNNs) makes extracting dense 3D visual features time-consuming, which calls for intensive memory and computing resources. Towards efficient TVG, we propose a novel text-visual prompting (TVP) framework, which incorporates optimized perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. In sharp contrast to 3D CNNs, we show that TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. Further, we propose a Temporal-Distance IoU (TDIoU) loss for efficient learning of TVG. Experiments on two benchmark datasets, Charades-STA and ActivityNet Captions datasets, empirically show that the proposed TVP significantly boosts the performance of 2D TVG (e.g., \\(9.79\\%\\) improvement on Charades-STA and \\(30.77\\%\\) improvement on ActivityNet Captions) and achieves \\(5 \\times\\) inference acceleration over TVG using 3D visual features. Codes are available at Open. Intel." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In recent years, we have witnessed great progress on temporal video grounding (TVG) [30, 74]. One key to this success comes from the fine-grained dense 3D visual features extracted by 3D convolutional neural networks (CNNs) (e.g., C3D [56] and I3D [3]) since TVG tasks demand spatial-temporal context to locate the temporal interval of the moments described by the text query. However, due to the high cost of the dense 3D feature extraction, most existing TVG models only take these 3D visual features ex" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.288, + 0.889, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.428, + 0.691, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.425, + 0.886, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.573, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Figure 1. The architecture and performance comparison among TVG methods: a) 3D TVG methods [14, 16, 18, 34, 43, 60-62, 64, 67, 69, 71, 73], b) 2D TVG methods [1, 7], and c) TVP-based 2D TVG (Ours), d) overall performance comparison. Ours is the most efficient (least inference time) and achieves competitive performance compared to 3D TVG methods. In contrast to existing TVG methods, which utilize dense video features extracted by non-trainable offline 3D CNNs and textual features, our proposed framework utilizes a trainable 2D CNN as the vision encoder to extract features from sparsely-sampled video frames with a universal set of frame-aware visual prompts and adds text prompts in textual feature space for end-to-end regression-based modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.748 + ], + "angle": 0, + "content": "tracted by offline 3D CNNs as inputs instead of co-training during TVG model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Although models using 3D visual features (that we call '3D methods or models') outperform these using the 2D features (that we call '2D methods or models'), a unique advantage of 2D methods is that extracting 2D visual features can significantly reduce the cost in TVG tasks [14, 15, 30, 34, 61, 62, 69, 74, 75]. An efficient and lightweight solution with reasonable performance is also demanded in computer vision, NLP, and video-language tasks [19, 23, 38, 41, 68, 76-80]. As discussed above, the methods employing 3D video features are challenging to be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.138 + ], + "angle": 0, + "content": "employed in practical applications. It thus has significant practical and economic value to develop compact 2D solutions for TVG tasks. In this paper, we ask:" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.157, + 0.444, + 0.187 + ], + "angle": 0, + "content": "How to advance 2D TVG methods so as to achieve comparable results to 3D TVG methods?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.47, + 0.37 + ], + "angle": 0, + "content": "To address this problem, we propose a novel text-visual prompting (TVP) framework for training TVG models using 2D visual features. As shown in Fig. 1, for existing 2D TVG and 3D TVG methods, they all utilize offline pretrained vision encoders and language encoders to perform feature extraction. In contrast, our proposed TVP framework is end-to-end trainable. Furthermore, benefiting from text-visual prompting and cross-modal pretraining on large-scale image-text datasets, our proposed framework could achieve comparable performance to 3D TVG methods with significant inference time acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.471, + 0.657 + ], + "angle": 0, + "content": "Conventionally, TVG methods consist of three stages: ① extracting feature from visual and text inputs; ② multimodal feature fusion; ③ cross-modal modelling. In contrast to conventional methods, TVP incorporates optimized input perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. We apply trainable parameters in the textual features as text prompts and develop a universal set of frame-aware patterns as visual prompts. Specially, we sample a fixed number of frames from a video and optimize text prompts for the input query sentence and a set of visual prompts for frames with different temporal locations during training. During testing, the same set of optimized visual prompts and textual prompts are applied to all test-time videos. We refer readers to Fig. 2 for illustrations of visual prompts and text prompts introduced. To the best of our knowledge, our work makes the first attempt to utilize prompt learning to successfully improve the performance of regression-based TVG tasks using 2D visual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Compared to 3D CNNs, 2D CNNs lose spatiotemporal information of the video during feature extraction. Inspired by the success of transformers on the vision-language tasks [9, 22, 35, 44, 47, 54, 55] and the recent application of prompt learning to transformers in both vision and language domains [2, 25, 27, 32, 37, 40], we choose transformer as our base TVG model and propose to utilize prompts to compensate for the lack of spatiotemporal information in 2D visual features. Furthermore, we develop a Temporal-Distance IoU (TDIoU) loss for training our proposed framework. There are two aspects that distinguish our proposed framework from existing works. First, our proposed framework is designed to boost the performance of the regression-based TVG methods utilizing 2D CNNs as the vision encoder, not for transfer learning [2, 21, 26] Second, our proposed framework utilizes 2D CNN to extract visual features from" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.093, + 0.641, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.245, + 0.631, + 0.257 + ], + "angle": 0, + "content": "(a) Text Prompts" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.093, + 0.865, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.245, + 0.846, + 0.257 + ], + "angle": 0, + "content": "(b) Frame-aware Visual Prompts" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Figure 2. Text-visual prompting illustration. (a) Text prompts are directly applied in the feature space. (b) A set of visual prompts are applied to video frames in order." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.445 + ], + "angle": 0, + "content": "sparsely-sampled video frames, which requires less memory and is easier to be applied in practical applications compared to 3D methods [34,60-62,69,75], especially for long videos. Furthermore, thanks to the compact 2D CNN as the vision encoder, our proposed framework could implement the language encoder and visual encoder co-training for better multimodal feature fusion. In summary, the contributions of this work are unfolded below:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.455, + 0.892, + 0.59 + ], + "angle": 0, + "content": "- We propose an effective and efficient framework to train 2D TVG models, in which we leverage TVP (text-visual prompting) to improve the utility of sparse 2D visual features without resorting to costly 3D features. To the best of our knowledge, it is the first work to expand the application of prompt learning for resolving TVG problems. Our method outperforms all of 2D methods and achieves competitive performance to 3D TVG methods." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.601, + 0.892, + 0.676 + ], + "angle": 0, + "content": "- Technology-wise, we integrate visual prompt with text prompt to co-improve the effectiveness of 2D visual features. On top of that, we propose TDIoU (temporal-distance IoU)-based prompt-model co-training method to obtain high-accuracy 2D TVG models." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.686, + 0.892, + 0.775 + ], + "angle": 0, + "content": "- Experiment-wise, we show the empirical success of our proposal to boost the performance of 2D TVG on Charades-STA and ActivityNet Captions datasets, e.g., \\(9.79\\%\\) improvement in Charades-STA, and \\(30.77\\%\\) in ActivityNet-Captions together with \\(5 \\times\\) inference time acceleration over 3D TVG methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.455, + 0.892, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.641, + 0.816 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Video Temporal Grounding (TVG). The objective of the TVG is to predict the starting/ending time points of target moments within an untrimmed video, which is described by a text sentence. Early TVG solutions [7,14,20,39,62,64,70] mainly employ two-stage \"propose-and-rank\" pipeline: ①" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14795" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Propose: utilize sliding windows or proposal network to generate proposal candidates from the input video. ② Rank: the proposed candidates would be ranked according to the text query, and then the proposal with the highest ranking would be the final prediction decision. In contrast to proposal-based methods, regression-based methods [16, 67, 69] directly predict the starting/ending time points of the target moments without ranking massive proposal candidates. Thus, regression-based methods are much faster than proposal-based methods, which is one reason why our work focuses on the regression-based TVG. Furthermore, reinforcement learning (RL)-based methods formulate the TVG task as a sequence of decisions to make [18, 60]. In particular, they train an agent to control the movement of a window by shifting or scaling. During training, the agent would be rewarded or punished based on whether the window is close to the target moment after an adjustment." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.473, + 0.518 + ], + "angle": 0, + "content": "Temporal Action Detection (TAD). TAD aims to determine whether predefined actions occur in a video and to predict the corresponding time intervals during which these actions occur [12,13,48,53,56,59,63]. Different from TVG, the input of TAD is only a video. In other words, TAD only requires a semantic understanding of videos. Compared to TAD, TVG is more challenging since it requires a semantic understanding of both videos and natural languages. Furthermore, TVG needs to process the multimodal interaction between videos and natural languages." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.52, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Text Prompting. Prompting has recently achieved great success in the domain of natural language processing [25, 32, 37, 40, 46, 49-52, 58]. Text prompting is a process that leverages a data-agnostic perturbation operation applied to text inputs or their embeddings to improve the performance of the downstream task. The simplest way is to construct an input context template originating from human contemplation [46, 49-51]. Although the manually-crafted context templates are simple and interpretable, they are typically not the optimal input prompts. To tackle this issue, other work has focused on searching the optimal prompting in the discrete input space [25, 52, 58] or in the language model's embedding space [32, 37, 40]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.473, + 0.869 + ], + "angle": 0, + "content": "Visual Prompting. Inspired by the idea of prompt learning in NLP [37], visual prompting (VP) was first proposed by Bahng et. al. [2] to reprogram a source vision model (e.g., ImageNet-pretrained classifier) to accomplish downstream target tasks (e.g., CIFAR-10 image classification). VP shares almost the same idea with the model reprogramming technology in the vision domain [4-6, 11, 57, 65, 72, 81], which incorporates a universal input perturbation into testing data so as to improve a desired performance metric, e.g., target task accuracy, robustness, and fairness." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Multi-Modal Prompting. Although visual prompting and text prompting have recently attracted much attention, they" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "are under-explored in the multi-modal learning, especially on the temporal video grounding task. The existing works [2, 27, 66] mainly focus on integrating text and visual prompts with the CLIP (Contrastive Language-Image Pre-trained) model to improve downstream tasks with imagery data. The problem of multi-modal prompting in the video understanding task has not been studied. In this paper, we for the first time develop the text-visual prompting technique to improve the performance of temporal video grounding using 2D visual features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.6, + 0.273 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.895, + 0.343 + ], + "angle": 0, + "content": "In this section, we begin with the problem formulation of regression-based TVG. Then we demonstrate the design of TVP (text-visual prompts) and present the overview of our proposed TVP framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.352, + 0.684, + 0.367 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.895, + 0.497 + ], + "angle": 0, + "content": "Let \\(\\mathbf{v} \\in \\mathbb{R}^{N_{\\mathrm{vid}} \\times C \\times H \\times W}\\) be an untrimmed video consisting of a sequence of \\(N_{\\mathrm{vid}}\\) video frames, and \\(\\mathbf{s} \\in \\mathbb{R}^{N_{\\mathrm{tex}}}\\) be a text query consisting of a sequence of \\(N_{\\mathrm{tex}}\\) language tokens. Here, the video-query pair \\((\\mathbf{v}, \\mathbf{s})\\) belongs to a video-language dataset \\(\\mathcal{D}\\). Given \\(\\mathbf{v}\\) and \\(\\mathbf{s}\\), TVG aims to predict the time interval \\(\\hat{\\mathbf{T}} = (\\hat{t}_{\\mathrm{sta}}, \\hat{t}_{\\mathrm{end}})\\) of the target video moments described by the query \\(\\mathbf{s}\\). The TVG model that fuses the vision-language modalities can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.504, + 0.892, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {T}} = f \\left(g _ {\\mathrm {t e x}} (\\mathbf {s}), g _ {\\mathrm {v i d}} (\\mathbf {v})\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.894, + 0.565 + ], + "angle": 0, + "content": "where \\( f \\) denotes TVG model, and \\( g_{\\mathrm{vid}} \\) and \\( g_{\\mathrm{tex}} \\) represent vision encoder and language encoder, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.7, + 0.589 + ], + "angle": 0, + "content": "3.2. TDIoU Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.597, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Conventionally, the TVG model can be learned by minimizing the temporal IoU loss \\(\\mathcal{L}_{\\mathrm{tIoU}}\\) defined below:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.638, + 0.892, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t I o U}} = \\left(1 - \\frac {\\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) \\bigcap \\mathbf {T}}{\\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) \\bigcup \\mathbf {T}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.78 + ], + "angle": 0, + "content": "where for ease of notation let \\(\\theta\\) denote all the trainable parameters involved in (1), and \\(\\mathbf{T} = (t_{\\mathrm{sta}}, t_{\\mathrm{end}})\\) is the label (i.e., the ground-truth time interval) of the target moment associated with the input video-query pair \\((\\mathbf{v}, \\mathbf{s})\\). The rationale behind (2) is to maximize the overlapping between the predicted time interval and its ground truth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "However, for non-overlapping cases, the temporal IoU loss \\(\\mathcal{L}_{\\mathrm{tIoU}}\\) would encounter a gradient vanishing problem. Inspired by [82], we develop a novel TDIoU (Temporal-Distance IoU) loss for training our proposed TVG models by incorporating the normalized central time point distance and duration difference between the predicted video clips and the target video clips. We elaborate on the proposed loss below." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.086, + 0.878, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.896, + 0.431 + ], + "angle": 0, + "content": "Figure 3. Overview of our proposed TVP (text-visual prompting) framework for 2D TVG (temporal video grounding). The whole process contains four phases: 1 Video frame preprocessing: uniformly sample frames from input video and apply a set of frame-aware visual prompts to the sampled frames in order; 2 Feature extraction: 2D CNN extracts features from sampled video frames with visual prompts, and the language encoder extracts textual features. In addition, the visual features would be spatially downsampled and temporally fused by max pooling and mean pooling, respectively. 3 Multimodal feature processing: after spatial downsampling and temporal fusion, the 2D visual features would be integrated into the prompted textual features. 4 Crossmodal fusion: the multimodal features would be processed by a 12-layer transformer encoder, and MLP would predict the starting/ending time points of the target moment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.549 + ], + "angle": 0, + "content": "Distance Loss \\(\\mathcal{L}_{\\mathrm{dis}}\\). To avoid the gradient vanishing problem caused by the non-overlapping case, we involve distance loss \\(\\mathcal{L}_{\\mathrm{dis}}\\) to directly minimize the normalized central time point distance. In addition, we add a threshold \\(\\alpha_{1}\\) to prevent oscillation in the later training phase. The distance loss is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.558, + 0.47, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i s}} = \\max \\left(\\frac {\\left| \\left(t _ {\\mathrm {s t a}} + t _ {\\mathrm {e n d}}\\right) / 2 - \\left(\\hat {t} _ {\\mathrm {s t a}} + \\hat {t} _ {\\mathrm {e n d}}\\right) / 2 \\right|}{\\left| \\hat {\\mathbf {T}} \\bigcup \\mathbf {T} \\right|}, \\alpha_ {1}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.647 + ], + "angle": 0, + "content": "where recall that \\(\\mathbf{T} = (t_{\\mathrm{sta}}, t_{\\mathrm{end}})\\), \\(\\hat{\\mathbf{T}}\\) is predicted by the TVG model (1), and we choose \\(\\alpha_1 = 0.2\\) in experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Duration Loss \\(\\mathcal{L}_{\\mathrm{dur}}\\). The introduction of distance loss \\(\\mathcal{L}_{\\mathrm{dis}}\\) avoids the gradient vanishing problem but only considers the central time point distance. Yet, this may not be precise enough. For example, even if the central time points are completely overlapped, the duration of two video clips may not be identical. Inspired by the above, we propose the duration loss:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.753, + 0.47, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d u r}} = \\max \\left(\\frac {| \\mathbf {T} - \\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) |}{| \\mathbf {T} |}, \\alpha_ {2}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.799, + 0.47, + 0.828 + ], + "angle": 0, + "content": "where \\(\\alpha_{2}\\) is the precision tolerance threshold and set by 0.4 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.829, + 0.47, + 0.859 + ], + "angle": 0, + "content": "Finally, the proposed Temporal-Distance IoU (TDIoU) loss is given by" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.863, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {t I o U}} + \\beta_ {1} \\mathcal {L} _ {\\mathrm {d i s}} + \\beta_ {2} \\mathcal {L} _ {\\mathrm {d u r}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.449, + 0.902 + ], + "angle": 0, + "content": "where \\(\\beta_{1} > 0\\) and \\(\\beta_{2} > 0\\) are regularization parameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.457, + 0.745, + 0.473 + ], + "angle": 0, + "content": "3.3. Text-Visual Prompt Design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Inspired by the application of prompts on transformers [2, 21, 36, 37], we propose jointly text-visual prompting to boost the performance of our models, in which prompts are optimized perturbation patterns. To improve data processing efficiency, we uniformly sample video frames from the untrimmed video \\(\\mathbf{v}\\) to obtain \\(\\mathbf{v}_{\\mathrm{sam}} \\in \\mathbb{R}^{N_{\\mathrm{sam}} \\times C \\times H \\times W}\\), where \\(N_{\\mathrm{sam}}\\) is the number of sampled video frames. In addition, we introduce a set of frame-aware visual prompts \\(\\delta_{\\mathrm{vp}} \\in \\mathbb{R}^{N_{\\mathrm{sam}} \\times d_{\\mathrm{vp}}}\\) in the pixel space of sampled video frames \\(\\mathbf{v}_{\\mathrm{sam}}\\), and introduce text prompts \\(\\delta_{\\mathrm{tp}} \\in \\mathbb{R}^{N_{\\mathrm{tp}} \\times d_{\\mathrm{tp}}}\\) in the textual feature space. By incorporating video frame sampling and text-visual prompts into the TVG model (1), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.688, + 0.892, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\left(\\hat {t} _ {\\mathrm {s t a}}, \\hat {t} _ {\\mathrm {e n d}}\\right) = f \\left(\\delta_ {\\mathrm {t p}}, g _ {\\mathrm {t e x}} (\\mathbf {s}), g _ {\\mathrm {v i d}} \\left(\\mathbf {v} _ {\\mathrm {s a m}} + \\delta_ {\\mathrm {v p}}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Given a pre-trained 2D TVG model \\( f \\), the objective of text-visual prompting (TVP) is to learn a universal set of visual prompts \\( \\delta_{\\mathrm{vp}} \\) and text prompts \\( \\delta_{\\mathrm{tp}} \\) to be integrated into sampled video frames and textual features, respectively. Specially, a set of different visual prompts are applied to uniformly-sampled frames of one untrimmed video in order. During training, only the set of visual prompts and text prompts are updated through backpropagation. During finetuning, prompts are frozen, and the parameters of the TVG model and encoders are updated. During testing, the set of optimized visual prompts and the optimized text prompts are applied to all test-time video-query pairs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14797" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.205, + 0.106 + ], + "angle": 0, + "content": "3.4. Framework" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.473, + 0.675 + ], + "angle": 0, + "content": "Inspired by the success of transformers in vision-language tasks, we choose ClipBERT [31] as the base model for 2D TVG. Extended from ClipBERT, the input of our regression-based TVG model would be describable sentences and uniformly sampled frames of one untrimmed video as shown in Fig. 3. Then, the predicted starting and ending time points of the target video clip would be model outputs. As described in Algorithm 1, there are four phases of our proposed TVP framework: Video frame preprocessing: We obtain sparsely-sampled frames \\(\\mathbf{v}_{\\mathrm{sam}}\\) from one input untrimmed video \\(\\mathbf{v}\\), and apply universal frame-aware visual prompts \\(\\delta_{\\mathrm{vp}}\\) on top of frames at the padding location. Feature extraction: 2D vision encoder (first 5 ConvBlock of ResNet-50) \\(g_{\\mathrm{vid}}\\) and language encoder (a trainable word embedding layer) \\(g_{\\mathrm{tex}}\\) would extract features from the prompted frames \\(\\mathbf{v}_{\\mathrm{sam}}'\\) and textual inputs \\(\\mathbf{s}\\), respectively. Multimodal feature processing: Following the setting of Pixel-BERT [22], the 2D visual features \\(\\mathbf{Q}_{\\mathrm{vid}}\\) are downsampled spatially by a \\(2 \\times 2\\) max-pooling layer and fused temporally by a mean-pooling layer. Then, text prompts \\(\\delta_{\\mathrm{tp}}\\) are integrated into textual features \\(\\mathbf{Q}_{\\mathrm{tex}}\\). In addition, trainable 2D visual position embeddings \\(\\mathbf{M}_{2\\mathrm{D}}\\) and textual position embeddings \\(\\mathbf{M}_{\\mathrm{pos}}\\) are applied to the processed 2D visual features \\(\\mathbf{Q}_{\\mathrm{vid}}'\\) and prompted textual features \\(\\mathbf{Q}_{\\mathrm{tex}}'\\), respectively [10, 31]. Afterwards, the processed and position-encoded 2D visual features \\(\\mathbf{Q}_{\\mathrm{vid}}''\\) are flattened and integrated into prompted and position-encoded textual features \\(\\mathbf{Q}_{\\mathrm{tex}}''\\). Moreover, type embeddings \\(\\mathbf{M}_{\\mathrm{type}}\\) would be added to the integrated multimodal features \\(\\mathbf{Q}_{\\mathrm{all}}\\) to indicate the source type of features. Crossmodal fusion: A 12-layer transformer [10] is utilized for crossmodal fusion on \\(\\mathbf{Q}_{\\mathrm{all}}\\), and then multilayer perceptron (MLP) ending with sigmoid function is used as the prediction head to process the last-layer crossmodal representation \\(\\mathbf{Q}_{\\mathrm{CM}}\\) of the transformer for generating the predicted starting/ending time points \\((\\hat{t}_{\\mathrm{sta}}, \\hat{t}_{\\mathrm{sta}})\\) of the target moments described by the text query input." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.689, + 0.21, + 0.707 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.762 + ], + "angle": 0, + "content": "In this section, we demonstrate the effectiveness of our proposed TVP framework on Charades-STA and ActivityNet Captions datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.254, + 0.789 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Datasets. The evaluations are implemented on two standard benchmark datasets for TVG task, Charades-STA [14] and ActivityNet Captions [28]. Tab. 1 summarizes the details of both datasets. Charades-STA dataset contains 6,672 videos and 16,124 text queries in total. The average length of videos is \\(30.6s\\), and the average length of text query is 7.2 words. The average length of moments cor" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.787, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Overview of TVP framework" + }, + { + "type": "ref_text", + "bbox": [ + 0.501, + 0.111, + 0.892, + 0.173 + ], + "angle": 0, + "content": "Input: vision encoder \\( g_{\\mathrm{vid}} \\), language encoder \\( g_{\\mathrm{tex}} \\), position embeddings \\( \\mathbf{M}_{\\mathrm{pos}} \\), 2D position embeddings \\( \\mathbf{M}_{\\mathrm{2D}} \\), type embeddings \\( \\mathbf{M}_{\\mathrm{type}} \\), transformer \\( f \\), prediction head \\( MLP \\), visual prompts \\( \\delta_{\\mathrm{vp}} \\), text prompts \\( \\delta_{\\mathrm{tp}} \\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.501, + 0.173, + 0.826, + 0.188 + ], + "angle": 0, + "content": "Output: Predicted time interval \\(\\hat{\\mathbf{T}} = (\\hat{t}_{\\mathrm{sta}},\\hat{t}_{\\mathrm{end}})\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.563, + 0.189, + 0.82, + 0.203 + ], + "angle": 0, + "content": "Phase ①: Video frame preprocessing" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.111, + 0.892, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.204, + 0.892, + 0.232 + ], + "angle": 0, + "content": "1: \\(\\mathbf{v}_{\\mathrm{sam}} \\gets\\) uniformly sample video frames from an untrimmed video \\(\\mathbf{v}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.892, + 0.263 + ], + "angle": 0, + "content": "2: \\(\\mathbf{v}_{\\mathrm{sam}}^{\\prime}\\gets\\) apply visual prompts \\(\\delta_{\\mathrm{vp}}\\) to the sampled video frames \\(\\mathbf{v}_{\\mathrm{sam}}\\)" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.204, + 0.892, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.562, + 0.264, + 0.766, + 0.277 + ], + "angle": 0, + "content": "Phase ②: Feature Extraction" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.278, + 0.872, + 0.294 + ], + "angle": 0, + "content": "3: \\(\\mathbf{Q}_{\\mathrm{vid}} = g_{\\mathrm{vid}}(\\mathbf{v}_{\\mathrm{sam}}^{\\prime})\\gets\\) extracting 2D visual features" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.295, + 0.828, + 0.308 + ], + "angle": 0, + "content": "4: \\(\\mathbf{Q}_{\\mathrm{tex}} = g_{\\mathrm{tex}}(\\mathbf{s})\\) ← extracting textual features" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.278, + 0.872, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.562, + 0.31, + 0.847, + 0.323 + ], + "angle": 0, + "content": "Phase 8: Multimodal feature processing" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.324, + 0.891, + 0.353 + ], + "angle": 0, + "content": "5: \\(\\mathbf{Q}_{\\mathrm{vld}}^{\\prime}\\gets\\) apply spatial downsampling and temporal fusion to 2D visual features \\(\\mathbf{Q}_{\\mathrm{vld}}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.354, + 0.891, + 0.37 + ], + "angle": 0, + "content": "6: \\(\\mathbf{Q}_{\\mathrm{tex}}^{\\prime}\\gets\\) apply text prompts \\(\\delta_{\\mathrm{tp}}\\) to textual features \\(\\mathbf{Q}_{\\mathrm{tex}}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.37, + 0.891, + 0.399 + ], + "angle": 0, + "content": "7: \\(\\mathbf{Q}_{\\mathrm{vid}}^{\\prime \\prime} \\gets\\) add 2D visual position embeddings \\(\\mathbf{M}_{2\\mathrm{D}}\\) on the processed 2D visual features \\(\\mathbf{Q}_{\\mathrm{vid}}^{\\prime}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.4, + 0.891, + 0.429 + ], + "angle": 0, + "content": "8: \\(\\mathbf{Q}_{\\mathrm{tex}}^{\\prime \\prime}\\gets\\) add position embeddings \\(\\mathbf{M}_{\\mathrm{pos}}\\) to prompted textual features \\(\\mathbf{Q}_{\\mathrm{tex}}^{\\prime}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.43, + 0.892, + 0.475 + ], + "angle": 0, + "content": "9: \\( \\mathbf{Q}_{\\mathrm{all}} \\gets \\) integrate the processed and position-encoded textual features \\( \\mathbf{Q}_{\\mathrm{tex}}^{\\prime \\prime} \\) and the processed and position-encoded 2D visual features \\( \\mathbf{Q}_{\\mathrm{vid}}^{\\prime \\prime} \\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.506, + 0.476, + 0.891, + 0.505 + ], + "angle": 0, + "content": "10: \\(\\mathbf{Q}_{\\mathrm{all}} + \\mathbf{M}_{\\mathrm{type}} \\gets\\) add type embeddings \\(\\mathbf{M}_{\\mathrm{type}}\\) to the integrated multimodal features \\(\\mathbf{Q}_{\\mathrm{all}}\\)" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.324, + 0.892, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.562, + 0.506, + 0.764, + 0.518 + ], + "angle": 0, + "content": "Phase 4: Crossmodal fusion" + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.52, + 0.891, + 0.55 + ], + "angle": 0, + "content": "11: \\(\\mathbf{Q}_{\\mathrm{CM}} = f(\\mathbf{Q}_{\\mathrm{all}} + \\mathbf{M}_{\\mathrm{type}}) \\gets\\) implement crossmodal fusion through transformer \\(f\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.55, + 0.892, + 0.596 + ], + "angle": 0, + "content": "12: \\((\\hat{t}_{\\mathrm{sta}},\\hat{t}_{\\mathrm{end}}) = MLP(\\mathbf{Q}_{\\mathrm{CM}})\\gets\\) prediction head generates the predicted time interval according to crossmodal representation \\(\\mathbf{Q}_{\\mathrm{CM}}\\)" + }, + { + "type": "list", + "bbox": [ + 0.505, + 0.52, + 0.892, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.617, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Table 1. Statistics of TVG benchmark datasets (Charades-STA and ActivityNet Captions datasets)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.652, + 0.888, + 0.781 + ], + "angle": 0, + "content": "
DatasetCharades-STAActivityNet Captions
DomainIndoor ActivityIndoor/Outdoor Activity
# Videos6,67214,926
Avg. Video Length (second)30.6117.6
# Moments11,76771,953
Avg. Moment Length (second)8.137.1
Vocabulary Size1,30315,505
# Queries16,12471,953
Avg. Query Length (word)7.214.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "responding to the text query is 8.1s. Following the same dataset split as [14] for fair comparisons, there are 12,408 video-query pairs for training and 3,720 pairs for testing. ActivityNet Captions dataset contains 14,926 videos and 71,953 text queries in total. The average length of videos is 117.6s, and the average length of text query is 14.4 words." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14798" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "The average length of moments corresponding to the text query is 37.1s. ActivityNet Captions dataset is split into training set, validation set, and testing set in a \\(2:1:1\\) ratio. Since the testing set is withheld for competition, only a training set and two validation sets (val1 and val2) can be accessed publicly. For fair comparisons, we evaluate our proposed framework on val1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.472, + 0.291 + ], + "angle": 0, + "content": "Baselines. We compare our proposal with 15 baseline methods: ① Proposal-based: CTRL [14], MCN [1], SAP [7], BPNet [62], LPNet [61], QSPN [64], MAN [71]; ② Proposal-free: ABLR [67], DRN [69], CPNet [34], DEBUG [43], ExCL [16], VSLNet [73]; ③ Reinforcement learning: TSP-PRL [60], TripNet [18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.472, + 0.385 + ], + "angle": 0, + "content": "Evaluation metrics. Following [14], we adopt Acc(R@1, IoU=m) as the performance evaluation metric, which represents the percentage accuracy of top-1 predicted moments whose tIoU (temporal IoU) with the ground-truth moment is larger than \\( m \\). By convention, we consider the following tIoU threshold values \\( m = \\{0.3, 0.5, 0.7\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.472, + 0.614 + ], + "angle": 0, + "content": "Crossmodal pretraining setup. Our 2D vision encoder (ResNet-50) is initialized with the weight from gridfeat [24], which can extract effective grid features from visual inputs. In addition, both the language encoder and 12-layer transformer are initialized with the BERT-base model weight [10], which are pretrained on English Wikipedia and BookCorpus [83]. Thanks to the compact 2D vision encoder, TVP (our proposal) is able to directly utilize image-text pairs for end-to-end training. Since the benefits of cross-modal pretraining has been demonstrated by [22, 44, 55], our base model is pretrained on two large-scale image-text datasets, which are Visual Genome Captions [29] and COCO Captions [8]. To be more specific, image-text matching [44, 55] and masked language modeling [10] are employed for cross-modal pretraining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Implementation setup. For video inputs, we uniformly sample \\(N_{\\mathrm{sam}}\\) frames from a video (\\(N_{\\mathrm{sam}} = 48\\) for Charades-STA and \\(N_{\\mathrm{sam}} = 64\\) for ActivityNet Captions). In addition, all video frames are resized to have a maximum longer side of 448 with an original aspect ratio, and then the frames are zero-padded to \\(448 \\times 448\\). The default visual prompt sizes for both dataset are 96. The default text prompt sizes are 10 and 20 for Charades-STA and ActivityNet Captions, respectively. We utilize the first 5 ConvBlocks of ResNet-50 as the 2D vision encoder and a trainable embedding layer as the language encoder for both Charades-STA and ActivityNet Captions datasets. For text queries, all word tokens are maintained after lower-case conversion and tokenization. We use AdamW [42] for end-to-end model training, with \\(\\beta_{1} = 1.0\\), \\(\\beta_{2} = 0.1\\), \\(\\alpha_{1} = 0.2\\), \\(\\alpha_{2} = 0.4\\). Initial learning rates are \\(1e - 1\\) and \\(5e - 7\\) for prompt training and model finetuning, respectively. In addition, the learning rate linearly decays to 0 with the first \\(10\\%\\) training step for warmup. Our experiments are imple" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.114 + ], + "angle": 0, + "content": "Table 2. Performance comparison of different thresholds \\( m \\) on the Charades-STA dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.124, + 0.889, + 0.359 + ], + "angle": 0, + "content": "
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D-23.638.89
ABLR [67]C3D-24.369.01
BPNet [62]C3D55.4638.2520.51
LPNet [61]C3D59.1440.9421.13
QSPN [64]C3D54.7035.6015.80
TSP-PRL [60]C3D-45.4524.75
TripNet [18]C3D54.6438.2916.07
DRN [69]C3D-45.4026.40
CPNet [34]C3D-40.3222.47
DEBUG [43]C3D54.9537.3917.92
ExCL [16]I3D61.5044.122.40
VSLNet [73]I3D64.3047.3130.19
MAN [71]I3D-46.5322.72
2D TVGMCN [1]VGG-17.468.01
SAP [7]VGG-27.4213.36
Ours
TVP-Based 2D TVGBase w/o promptsResNet61.2940.4319.89
Base + Visual Prompts65.3844.3120.22
Base + Text Prompts65.8143.4420.65
Base + Both Prompts65.9244.3921.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Table 3. Performance comparison of different thresholds \\( m \\) on the ActivityNet Captions dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.406, + 0.89, + 0.608 + ], + "angle": 0, + "content": "
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D28.7014.00-
BPNet [62]C3D59.9842.0724.69
LPNet [61]C3D64.2945.9225.39
QSPN [64]C3D45.3027.7013.60
TSP-PRL [60]C3D56.0238.83-
TripNet [18]C3D48.4232.1913.93
DRN [69]C3D-45.4524.36
CPNet [34]C3D-40.5621.63
ABLR [67]C3D55.6736.79-
DEBUG [43]C3D55.9139.72-
ExCL [16]C3D63.0043.6024.10
VSLNet [73]C3D63.1643.2226.16
Ours
TVP-Based 2D TVGBase w/o promptsResNet57.2040.1619.14
Base + Visual Prompts60.1243.3923.71
Base + Text Prompts60.4842.5824.39
Base + Both Prompts60.7143.4425.03
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.633, + 0.894, + 0.679 + ], + "angle": 0, + "content": "mented in PyTorch [45], and models and prompts are finetuned separately for 12 epochs with the mixed precision on 8 NVIDIA V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.688, + 0.69, + 0.705 + ], + "angle": 0, + "content": "4.2. Experiment Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of TVP on Charades-STA. The performance comparisons with SOTA methods on the Charades-STA dataset are summarized in Tab. 2. Our proposed TVP framework can achieve competitive performance at all IoU thresholds \\( m \\) in the case of utilizing 2D visual features extracted by ResNet-50, and reach the highest score at \\( m = 0.3 \\). Compared to the 2D TVG methods using VGG as the vision encoder, our proposed framework could achieve around \\( 2.5 \\times \\) and \\( 2.7 \\times \\) performance gain at thresholds 0.5 and 0.7, respectively. Furthermore, we can find that for our base model only one of visual prompts and text prompts can achieve up to \\( 7.37\\% \\) and \\( 9.60\\% \\) improvement" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14799" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.273, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.271, + 0.227, + 0.283 + ], + "angle": 0, + "content": "(a) Charades-STA" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.09, + 0.465, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.271, + 0.436, + 0.284 + ], + "angle": 0, + "content": "(b) ActivityNet Captions" + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.295, + 0.393, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Impact of sampled frame numbers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.414 + ], + "angle": 0, + "content": "at tIoU thresholds \\( m = 0.3 \\) and \\( m = 0.5 \\). The combination of text and visual prompts can not only achieve \\( 7.55\\% \\) and \\( 9.79\\% \\) improvements at tIoU thresholds \\( m = 0.3 \\) and \\( m = 0.5 \\), but also improve the performance by \\( 8.14\\% \\) at \\( m = 0.7 \\). This demonstrates the effectiveness and necessity of the joint text-visual prompting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.672 + ], + "angle": 0, + "content": "Effectiveness of TVP on ActivityNet Captions. We focus on the performance comparisons with 3D TVG methods on ActivityNet since there are no results of 2D TVG method reported on ActivityNet Captions. The results of multiple methods on ActivityNet Captions datasets are reported in Tab. 3. Even on this more challenging dataset, our proposed method still has achieved competitive performance compared to 3D TVG methods. Different from the performance of TVP on Charades-STA dataset, text prompts or visual prompts can achieve a significant performance boost on the base model over all IoU thresholds \\( m \\) alone (5.73% at \\( m = 0.3 \\), 8.04% at \\( m = 0.5 \\), 27.43% at \\( m = 0.7 \\)), and the text-visual prompt combination could further boost the performance (6.14% at \\( m = 0.3 \\), 8.17% at \\( m = 0.5 \\), 30.77% at \\( m = 0.7 \\)). It is worth noting that the performance gap over \\( m = 0.7 \\) between 2D TVG methods and 3D TVG methods is narrowed significantly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.869 + ], + "angle": 0, + "content": "In summary, through the experimental results on Charades-STA and ActivityNet Captions datasets, we can find that our proposed TVP framework could achieve competitive performance overall tIoU thresholds on Charades-STA and ActivityNet Captions by improving the utility of sparse 2D visual features. Thanks to the lightweight 2D vision encoder, cotraining language encoder and vision encoder on large-scale image-text datasets can be performed, which benefits the base model to achieve good performance. Furthermore, the combination of text and visual prompts can achieve better results than any single kind of prompts on both datasets, which again proves the importance of cross-modal training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Video frame sampling effect. Fig. 4 demonstrates the performance of base model with different number \\( N_{\\mathrm{sam}} \\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.113 + ], + "angle": 0, + "content": "Table 4. The performance comparison of different visual prompt sizes on Charades-STA dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.124, + 0.888, + 0.309 + ], + "angle": 0, + "content": "
Visual Prompt SizeAcc(R@1, IoU=m)Prompt + Frame
m=0.3m=0.5m=0.7
061.2940.4319.89
1661.2940.4320.00
3261.9439.7819.35
4863.6642.3720.00
7263.8743.6619.78
9665.3844.3120.22
12864.7343.6619.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.325, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Table 5. The performance comparison of different text prompt sizes on Charades-STA dataset." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.354, + 0.88, + 0.501 + ], + "angle": 0, + "content": "
Text Prompt SizeAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
057.2040.1619.14
565.3841.9420.43
1065.8143.4420.65
1565.5943.2321.29
2064.9543.8721.51
2563.6642.8020.65
3064.4642.6320.51
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.61 + ], + "angle": 0, + "content": "of sampled video frames as visual inputs. For Charades dataset, the base model performance keeps increasing before \\( N_{\\mathrm{sam}} \\) reaches 48, but when it exceeds 48, performance starts to degrade. This is because frequent background changes harm the performance of object re-identification in videos, which are noisy for object motion analysis [17]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.702 + ], + "angle": 0, + "content": "For ActivityNet Caption dataset, base model performance continues to improve even when sampled frame number \\( N_{\\mathrm{sam}} \\) exceeds 48, due to the longer average video length in ActivityNet Captions dataset. Balancing the frame number and batch size for training, we choose \\( N_{\\mathrm{sam}} = 64 \\) for ActivityNet Captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.869 + ], + "angle": 0, + "content": "TVP performance vs. prompt size. As shown in Tab. 4, we can find that when visual prompts are small, they cannot bring changes to the base model, and when visual prompts are too large, the performance starts to decrease. This is because key information within video frames might be removed. However, the text prompts can bring significant performance boost even when the text prompt size is small as shown in Tab. 5, which is because the textual features has a smaller dimension compared to visual features, and also the text prompts are directly optimized in feature space during training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "TVP performance vs. visual prompt operation. Visual prompt is first proposed by [2], where visual prompts are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14800" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.126 + ], + "angle": 0, + "content": "Table 6. The performance comparison of different visual prompt operations ('remove', 'add', 'replace') with fixed visual prompt size \\( p = 96 \\) on Charades-STA and ActivityNet Captions datasets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.136, + 0.465, + 0.23 + ], + "angle": 0, + "content": "
OperationCharades-STAActivityNet Captions
R@1, IoU=mR@1, IoU=m
m=0.3m=0.5m=0.7m=0.3m=0.5m=0.7
Original61.2940.4319.8957.2040.1619.14
Remove61.2940.4320.057.2040.1619.14
Add61.0839.5720.2257.1540.1619.27
Replace65.3844.3120.2260.1243.3923.71
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.248, + 0.268, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.427, + 0.238, + 0.439 + ], + "angle": 0, + "content": "(a) Vison Encoders" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.248, + 0.462, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.427, + 0.462, + 0.438 + ], + "angle": 0, + "content": "(b) Modules in 2D TVG Model" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.47, + 0.511 + ], + "angle": 0, + "content": "Figure 5. Inference time comparison. (a) inference time comparison between 2D vision encoder (ResNet-50) and 3D vision encoder (C3D). (b) inference time comparison between the vision encoder and the other modules of the 2D TVG model, where the sampled frame number for our TVP framework is \\(1.2 \\times\\) the length of the video in seconds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.749 + ], + "angle": 0, + "content": "added to the image for transfer learning on classification tasks. In contrast, our proposed prompting framework is designed to compensate for the spatiotemporal information loss in 2D visual features. Due to the differences in the task, we try two different prompt operation strategies, 'replace' and 'add'. 'add' is to add the visual prompts to the pixel value of the video frame at the corresponding padding locations. 'replace' is to replace the pixel values of video frames with visual prompts at corresponding padding locations. 'remove' is in order to study the impact of removing the pixel values at the padding location. As shown in Tab. 6, 'add' or 'remove' prompt operations have limited effects on the base model. However, 'replace' does boost the base model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "TVP achieves inference efficiency. As shown in Fig. 5, we can find that the inference time required for visual feature extraction accounts for more than half of the inference time of the whole model, while the inference time required for the 3D vision encoder is more than \\(5 \\times\\) compared to the 2D vision encoder, and even more than the time required for the whole TVG model using 2D vision encoder, which fully demonstrates the feasibility of accelerating the overall inference speed by reducing the complexity of the vision encoder. Need to note that if there are" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.113 + ], + "angle": 0, + "content": "Table 7. The performance comparison of different loss designs on Charades-STA dataset." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.124, + 0.868, + 0.221 + ], + "angle": 0, + "content": "
Loss Function SelectionR@1, IoU=m
m=0.3m=0.5m=0.7
LtIoU55.0529.8911.82
LtIoU + Ldis60.6431.1816.77
LtIoU + Ldur59.7830.9716.34
LtIoU + Ldis + Ldur61.2940.4319.89
" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.228, + 0.687, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.229, + 0.887, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Figure 6. Loss landscape visualization in 2D plane: Finetuning w/o prompts (left) and using prompts (right); see [33] for implementation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.439 + ], + "angle": 0, + "content": "multiple model weights for different sampled frame number settings and model weights can be adopted adaptively for different lengths of videos, the inference speed for short videos should increase, and the prediction results for long videos will be further improved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.893, + 0.651 + ], + "angle": 0, + "content": "Ablation studies. Through Tab. 7, we can find that the addition of either distance loss \\(\\mathcal{L}_{\\mathrm{dis}}\\) or duration loss \\(\\mathcal{L}_{\\mathrm{dur}}\\) will result in a performance increase, but the combination of the two will result in a significant performance increase (11.34% at \\(m = 0.3\\), 35.26% at \\(m = 0.5\\), 68.27% at \\(m = 0.7\\)), especially over tIoU thresholds \\(m = 0.5\\) and \\(m = 0.7\\). This demonstrates that distance loss \\(\\mathcal{L}_{\\mathrm{dis}}\\) and duration loss \\(\\mathcal{L}_{\\mathrm{dur}}\\) could provide more precise training guides compared to only using temporal IoU loss \\(\\mathcal{L}_{\\mathrm{tIoU}}\\). Furthermore, we posit that prompting may encode additional spatial-temporal supervision to help the model trainer to escape from bad local optima as shown in Fig. 6, where finetuning w/ prompts yields a flatter loss landscape than the one w/o prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.619, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose text-visual prompting to boost the performance of 2D TVG methods by compensating for the lack of spatiotemporal information in 2D visual features. In contrast to 3D TVG methods, TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. The effectiveness of our proposed TVP (text-visual prompting) framework has been demonstrated on two standard datasets, Charades-STA and ActivityNet. Our models outperform all 2D models significantly, and also achieve comparable performance to 3D models. What is more, we achieve over \\(5 \\times\\) inference speedup over TVG methods of using 3D visual features." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14801" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] H Bahng, A Jahanian, S Sankaranarayanan, and P Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, page 2022, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Aochuan Chen, Peter Lorenz, Yuguang Yao, Pin-Yu Chen, and Sijia Liu. Visual prompting for adversarial robustness. arXiv preprint arXiv:2210.06284, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[5] Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. arXiv preprint arXiv:2211.11635, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[6] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. arXiv:2202.10629, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8199-8206, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.558 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[11] Gamaleldin F Elsayed, Ian Goodfellow, et al. Adversarial reprogramming of neural networks. arXiv:1806.11146, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[12] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Yanjun Gao, Lulu Liu, Jason Wang, Xin Chen, Huayan Wang, and Rui Zhang. Evoquer: Enhancing temporal grounding with video-pivoted backquery generation. arXiv preprint arXiv:2109.04600, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[17] Xinqian Gu, Hong Chang, Bingpeng Ma, Hongkai Zhang, and Xilin Chen. Appearance-preserving 3d convolution for video-based person re-identification. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[18] Meera Hahn, Asim Kadav, James M Rehg, and Hans Peter Graf. Tripping through time: Efficient localization of activities in videos. arXiv preprint arXiv:1904.09936, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[19] Xiaochen Han, Bo Wu, Zheng Shou, Xiao-Yang Liu, Yi-meng Zhang, and Linghe Kong. Tensor fista-net for real-time snapshot compressive imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10933-10940, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[20] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[21] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[22] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. arXiv preprint arXiv:2004.00849, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[23] Jinghan Jia, Mingyi Hong, Yimeng Zhang, Mehmet Akçakaya, and Sijia Liu. On the robustness of deep learning-based mri reconstruction to image transformations. arXiv preprint arXiv:2211.04930, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[24] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10267-10276, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[25] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. arXiv preprint arXiv:2210.03117, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[30] Xiaohan Lan, Yitian Yuan, Xin Wang, Zhi Wang, and Wenwu Zhu. A survey on temporal sentence grounding in videos. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[32] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[33] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[34] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.471, + 0.502 + ], + "angle": 0, + "content": "[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[36] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[37] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[38] Yanyu Li, Pu Zhao, Geng Yuan, Xue Lin, Yanzhi Wang, and Xin Chen. Pruning-as-search: Efficient neural architecture search via channel pruning and structural reparameterization. arXiv preprint arXiv:2206.01198, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[39] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.471, + 0.772 + ], + "angle": 0, + "content": "[40] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[41] Xiao-Yang Liu, Yimeng Zhang, Yukang Liao, and Ling Jiang. Dynamic updating of the knowledge base for a large-scale question answering system. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.288 + ], + "angle": 0, + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.893, + 0.343 + ], + "angle": 0, + "content": "[46] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.893, + 0.398 + ], + "angle": 0, + "content": "[47] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[48] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In proceedings of the IEEE International Conference on Computer Vision, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.893, + 0.497 + ], + "angle": 0, + "content": "[49] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[50] Timo Schick and Hinrich Schütze. Few-shot text generation with pattern-exploiting training. arXiv preprint arXiv:2012.11926, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[51] Timo Schick and Hinrich Schütze. It's not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.893, + 0.637 + ], + "angle": 0, + "content": "[52] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.893, + 0.678 + ], + "angle": 0, + "content": "[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.734 + ], + "angle": 0, + "content": "[54] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.844 + ], + "angle": 0, + "content": "[56] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.899 + ], + "angle": 0, + "content": "[57] Yun-Yun Tsai et al. Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. arXiv:2007.08714, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14803" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[58] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[60] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12386–12393, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[61] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. arXiv preprint arXiv:2109.10678, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[62] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[63] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[64] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9062–9069, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[65] Chao-Han Huck Yang, Yun-Yun Tsai, et al. Voice2series: Reprogramming acoustic models for time series classification. In ICML. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[66] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[67] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[68] Ofir Zafrir, Ariel Larey, Guy Boudoukh, Haihao Shen, and Moshe Wasserblat. Prune once for all: Sparse pre-trained language models. arXiv preprint arXiv:2111.05754, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[69] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[70] Yawen Zeng, Da Cao, Xiaochi Wei, Meng Liu, Zhou Zhao, and Zheng Qin. Multi-modal relational graph for cross-modal video moment retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2215-2224, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[71] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[72] Guanhua Zhang, Yihua Zhang, Yang Zhang, Wenqi Fan, Qing Li, Sijia Liu, and Shiyu Chang. Fairness reprogramming. arXiv preprint arXiv:2209.10222, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[73] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[74] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.315, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[75] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.384, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[76] Yimeng Zhang, Akshay Karkal Kamath, Qiucheng Wu, Zhiwen Fan, Wuyang Chen, Zhangyang Wang, Shiyu Chang, Sijia Liu, and Cong Hao. Data-model-circuit tri-design for ultra-light video intelligence on edge devices. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 745-750, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.468, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[77] Yimeng Zhang, Xiao-Yang Liu, Bo Wu, and Anwar Walid. Video synthesis via transform-based tensor neural network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2454-2462, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.523, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[78] Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jinfeng Yi, Mingyi Hong, Shiyu Chang, and Sijia Liu. How to robustify black-box ml models? a zeroth-order optimization perspective. arXiv preprint arXiv:2203.14195, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.579, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[79] Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Mingyi Hong, Yanzhi Wang, and Sijia Liu. Advancing model pruning via bi-level optimization. arXiv preprint arXiv:2210.04092, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[80] Yihua Zhang, Guanhua Zhang, Prashant Khanduri, Mingyi Hong, Shiyu Chang, and Sijia Liu. Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In International Conference on Machine Learning, pages 26693-26712. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[81] Yang Zheng, Xiaoyi Feng, et al. Why adversarial reprogramming works, when it fails, and how to tell the difference. arXiv:2108.11673, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[82] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.814, + 0.892, + 0.897 + ], + "angle": 0, + "content": "[83] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19-27, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14804" + } + ] +] \ No newline at end of file diff --git a/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_origin.pdf b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6be263c7d411bcde342cd7abdaf03737bdf56414 --- /dev/null +++ b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/86b06816-ef9f-43d9-b909-e5a92eff83a2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:562a37b9e36e3267bdcb08e8b8a16f2120c2e8f67aeb8eb56ac1c85bb46e290d +size 2008767 diff --git a/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/full.md b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..101fdf27771ce739442f9f07c79b7051b0621ee9 --- /dev/null +++ b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/full.md @@ -0,0 +1,351 @@ +# Text-Visual Prompting for Efficient 2D Temporal Video Grounding + +Yimeng Zhang $^{1,2}$ , Xin Chen $^{2}$ , Jinghan Jia $^{1}$ , Sijia Liu $^{1}$ , Ke Ding $^{2}$ $^{1}$ Michigan State University, $^{2}$ Applied AI, Intel +{zhan1853, jiajingh, liusiji5}@msu.edu, +{xin.chen, ke.ding}@intel.com + +# Abstract + +In this paper, we study the problem of temporal video grounding (TVG), which aims to predict the starting/ending time points of moments described by a text sentence within a long untrimmed video. Benefiting from fine-grained 3D visual features, the TVG techniques have achieved remarkable progress in recent years. However, the high complexity of 3D convolutional neural networks (CNNs) makes extracting dense 3D visual features time-consuming, which calls for intensive memory and computing resources. Towards efficient TVG, we propose a novel text-visual prompting (TVP) framework, which incorporates optimized perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. In sharp contrast to 3D CNNs, we show that TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. Further, we propose a Temporal-Distance IoU (TDIoU) loss for efficient learning of TVG. Experiments on two benchmark datasets, Charades-STA and ActivityNet Captions datasets, empirically show that the proposed TVP significantly boosts the performance of 2D TVG (e.g., $9.79\%$ improvement on Charades-STA and $30.77\%$ improvement on ActivityNet Captions) and achieves $5 \times$ inference acceleration over TVG using 3D visual features. Codes are available at Open. Intel. + +# 1. Introduction + +In recent years, we have witnessed great progress on temporal video grounding (TVG) [30, 74]. One key to this success comes from the fine-grained dense 3D visual features extracted by 3D convolutional neural networks (CNNs) (e.g., C3D [56] and I3D [3]) since TVG tasks demand spatial-temporal context to locate the temporal interval of the moments described by the text query. However, due to the high cost of the dense 3D feature extraction, most existing TVG models only take these 3D visual features ex + +![](images/7481fb8c16d19e42311863a71b265ed5a2287a76e309ce3d7b1024aa2f52b992.jpg) + +![](images/779651af2c94ec20721ac8fb8fe00ac25d73520599819a185e8d4f1de6bfdfe9.jpg) +Figure 1. The architecture and performance comparison among TVG methods: a) 3D TVG methods [14, 16, 18, 34, 43, 60-62, 64, 67, 69, 71, 73], b) 2D TVG methods [1, 7], and c) TVP-based 2D TVG (Ours), d) overall performance comparison. Ours is the most efficient (least inference time) and achieves competitive performance compared to 3D TVG methods. In contrast to existing TVG methods, which utilize dense video features extracted by non-trainable offline 3D CNNs and textual features, our proposed framework utilizes a trainable 2D CNN as the vision encoder to extract features from sparsely-sampled video frames with a universal set of frame-aware visual prompts and adds text prompts in textual feature space for end-to-end regression-based modeling. + +![](images/84dd07cd3e3a775994638e3f31ae76ba1aadeebb6da5242c165c28bdef03b367.jpg) + +tracted by offline 3D CNNs as inputs instead of co-training during TVG model training. + +Although models using 3D visual features (that we call '3D methods or models') outperform these using the 2D features (that we call '2D methods or models'), a unique advantage of 2D methods is that extracting 2D visual features can significantly reduce the cost in TVG tasks [14, 15, 30, 34, 61, 62, 69, 74, 75]. An efficient and lightweight solution with reasonable performance is also demanded in computer vision, NLP, and video-language tasks [19, 23, 38, 41, 68, 76-80]. As discussed above, the methods employing 3D video features are challenging to be + +employed in practical applications. It thus has significant practical and economic value to develop compact 2D solutions for TVG tasks. In this paper, we ask: + +How to advance 2D TVG methods so as to achieve comparable results to 3D TVG methods? + +To address this problem, we propose a novel text-visual prompting (TVP) framework for training TVG models using 2D visual features. As shown in Fig. 1, for existing 2D TVG and 3D TVG methods, they all utilize offline pretrained vision encoders and language encoders to perform feature extraction. In contrast, our proposed TVP framework is end-to-end trainable. Furthermore, benefiting from text-visual prompting and cross-modal pretraining on large-scale image-text datasets, our proposed framework could achieve comparable performance to 3D TVG methods with significant inference time acceleration. + +Conventionally, TVG methods consist of three stages: ① extracting feature from visual and text inputs; ② multimodal feature fusion; ③ cross-modal modelling. In contrast to conventional methods, TVP incorporates optimized input perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. We apply trainable parameters in the textual features as text prompts and develop a universal set of frame-aware patterns as visual prompts. Specially, we sample a fixed number of frames from a video and optimize text prompts for the input query sentence and a set of visual prompts for frames with different temporal locations during training. During testing, the same set of optimized visual prompts and textual prompts are applied to all test-time videos. We refer readers to Fig. 2 for illustrations of visual prompts and text prompts introduced. To the best of our knowledge, our work makes the first attempt to utilize prompt learning to successfully improve the performance of regression-based TVG tasks using 2D visual features. + +Compared to 3D CNNs, 2D CNNs lose spatiotemporal information of the video during feature extraction. Inspired by the success of transformers on the vision-language tasks [9, 22, 35, 44, 47, 54, 55] and the recent application of prompt learning to transformers in both vision and language domains [2, 25, 27, 32, 37, 40], we choose transformer as our base TVG model and propose to utilize prompts to compensate for the lack of spatiotemporal information in 2D visual features. Furthermore, we develop a Temporal-Distance IoU (TDIoU) loss for training our proposed framework. There are two aspects that distinguish our proposed framework from existing works. First, our proposed framework is designed to boost the performance of the regression-based TVG methods utilizing 2D CNNs as the vision encoder, not for transfer learning [2, 21, 26] Second, our proposed framework utilizes 2D CNN to extract visual features from + +![](images/8f460a3358496b498f2d79bbe0b7fd554ad0d4c9645dbf7a2ffbb32976efa959.jpg) +(a) Text Prompts +Figure 2. Text-visual prompting illustration. (a) Text prompts are directly applied in the feature space. (b) A set of visual prompts are applied to video frames in order. + +![](images/cdd49bfac571a910a37322476ba91c6f4ee2329484bd92f839f088de39cca0aa.jpg) +(b) Frame-aware Visual Prompts + +sparsely-sampled video frames, which requires less memory and is easier to be applied in practical applications compared to 3D methods [34,60-62,69,75], especially for long videos. Furthermore, thanks to the compact 2D CNN as the vision encoder, our proposed framework could implement the language encoder and visual encoder co-training for better multimodal feature fusion. In summary, the contributions of this work are unfolded below: + +- We propose an effective and efficient framework to train 2D TVG models, in which we leverage TVP (text-visual prompting) to improve the utility of sparse 2D visual features without resorting to costly 3D features. To the best of our knowledge, it is the first work to expand the application of prompt learning for resolving TVG problems. Our method outperforms all of 2D methods and achieves competitive performance to 3D TVG methods. +- Technology-wise, we integrate visual prompt with text prompt to co-improve the effectiveness of 2D visual features. On top of that, we propose TDIoU (temporal-distance IoU)-based prompt-model co-training method to obtain high-accuracy 2D TVG models. +- Experiment-wise, we show the empirical success of our proposal to boost the performance of 2D TVG on Charades-STA and ActivityNet Captions datasets, e.g., $9.79\%$ improvement in Charades-STA, and $30.77\%$ in ActivityNet-Captions together with $5 \times$ inference time acceleration over 3D TVG methods. + +# 2. Related Work + +Video Temporal Grounding (TVG). The objective of the TVG is to predict the starting/ending time points of target moments within an untrimmed video, which is described by a text sentence. Early TVG solutions [7,14,20,39,62,64,70] mainly employ two-stage "propose-and-rank" pipeline: ① + +Propose: utilize sliding windows or proposal network to generate proposal candidates from the input video. ② Rank: the proposed candidates would be ranked according to the text query, and then the proposal with the highest ranking would be the final prediction decision. In contrast to proposal-based methods, regression-based methods [16, 67, 69] directly predict the starting/ending time points of the target moments without ranking massive proposal candidates. Thus, regression-based methods are much faster than proposal-based methods, which is one reason why our work focuses on the regression-based TVG. Furthermore, reinforcement learning (RL)-based methods formulate the TVG task as a sequence of decisions to make [18, 60]. In particular, they train an agent to control the movement of a window by shifting or scaling. During training, the agent would be rewarded or punished based on whether the window is close to the target moment after an adjustment. + +Temporal Action Detection (TAD). TAD aims to determine whether predefined actions occur in a video and to predict the corresponding time intervals during which these actions occur [12,13,48,53,56,59,63]. Different from TVG, the input of TAD is only a video. In other words, TAD only requires a semantic understanding of videos. Compared to TAD, TVG is more challenging since it requires a semantic understanding of both videos and natural languages. Furthermore, TVG needs to process the multimodal interaction between videos and natural languages. + +Text Prompting. Prompting has recently achieved great success in the domain of natural language processing [25, 32, 37, 40, 46, 49-52, 58]. Text prompting is a process that leverages a data-agnostic perturbation operation applied to text inputs or their embeddings to improve the performance of the downstream task. The simplest way is to construct an input context template originating from human contemplation [46, 49-51]. Although the manually-crafted context templates are simple and interpretable, they are typically not the optimal input prompts. To tackle this issue, other work has focused on searching the optimal prompting in the discrete input space [25, 52, 58] or in the language model's embedding space [32, 37, 40]. + +Visual Prompting. Inspired by the idea of prompt learning in NLP [37], visual prompting (VP) was first proposed by Bahng et. al. [2] to reprogram a source vision model (e.g., ImageNet-pretrained classifier) to accomplish downstream target tasks (e.g., CIFAR-10 image classification). VP shares almost the same idea with the model reprogramming technology in the vision domain [4-6, 11, 57, 65, 72, 81], which incorporates a universal input perturbation into testing data so as to improve a desired performance metric, e.g., target task accuracy, robustness, and fairness. + +Multi-Modal Prompting. Although visual prompting and text prompting have recently attracted much attention, they + +are under-explored in the multi-modal learning, especially on the temporal video grounding task. The existing works [2, 27, 66] mainly focus on integrating text and visual prompts with the CLIP (Contrastive Language-Image Pre-trained) model to improve downstream tasks with imagery data. The problem of multi-modal prompting in the video understanding task has not been studied. In this paper, we for the first time develop the text-visual prompting technique to improve the performance of temporal video grounding using 2D visual features. + +# 3. Methods + +In this section, we begin with the problem formulation of regression-based TVG. Then we demonstrate the design of TVP (text-visual prompts) and present the overview of our proposed TVP framework. + +# 3.1. Problem Definition + +Let $\mathbf{v} \in \mathbb{R}^{N_{\mathrm{vid}} \times C \times H \times W}$ be an untrimmed video consisting of a sequence of $N_{\mathrm{vid}}$ video frames, and $\mathbf{s} \in \mathbb{R}^{N_{\mathrm{tex}}}$ be a text query consisting of a sequence of $N_{\mathrm{tex}}$ language tokens. Here, the video-query pair $(\mathbf{v}, \mathbf{s})$ belongs to a video-language dataset $\mathcal{D}$ . Given $\mathbf{v}$ and $\mathbf{s}$ , TVG aims to predict the time interval $\hat{\mathbf{T}} = (\hat{t}_{\mathrm{sta}}, \hat{t}_{\mathrm{end}})$ of the target video moments described by the query $\mathbf{s}$ . The TVG model that fuses the vision-language modalities can be described as: + +$$ +\hat {\mathbf {T}} = f \left(g _ {\mathrm {t e x}} (\mathbf {s}), g _ {\mathrm {v i d}} (\mathbf {v})\right), \tag {1} +$$ + +where $f$ denotes TVG model, and $g_{\mathrm{vid}}$ and $g_{\mathrm{tex}}$ represent vision encoder and language encoder, respectively. + +# 3.2. TDIoU Loss Function + +Conventionally, the TVG model can be learned by minimizing the temporal IoU loss $\mathcal{L}_{\mathrm{tIoU}}$ defined below: + +$$ +\mathcal {L} _ {\mathrm {t I o U}} = \left(1 - \frac {\hat {\mathbf {T}} (\boldsymbol {\theta}) \bigcap \mathbf {T}}{\hat {\mathbf {T}} (\boldsymbol {\theta}) \bigcup \mathbf {T}}\right), \tag {2} +$$ + +where for ease of notation let $\theta$ denote all the trainable parameters involved in (1), and $\mathbf{T} = (t_{\mathrm{sta}}, t_{\mathrm{end}})$ is the label (i.e., the ground-truth time interval) of the target moment associated with the input video-query pair $(\mathbf{v}, \mathbf{s})$ . The rationale behind (2) is to maximize the overlapping between the predicted time interval and its ground truth. + +However, for non-overlapping cases, the temporal IoU loss $\mathcal{L}_{\mathrm{tIoU}}$ would encounter a gradient vanishing problem. Inspired by [82], we develop a novel TDIoU (Temporal-Distance IoU) loss for training our proposed TVG models by incorporating the normalized central time point distance and duration difference between the predicted video clips and the target video clips. We elaborate on the proposed loss below. + +![](images/ad268c97dbc9581d78d148a1fb367ff2a6e09f3845d89ea4aa24db4a76da923d.jpg) +Figure 3. Overview of our proposed TVP (text-visual prompting) framework for 2D TVG (temporal video grounding). The whole process contains four phases: 1 Video frame preprocessing: uniformly sample frames from input video and apply a set of frame-aware visual prompts to the sampled frames in order; 2 Feature extraction: 2D CNN extracts features from sampled video frames with visual prompts, and the language encoder extracts textual features. In addition, the visual features would be spatially downsampled and temporally fused by max pooling and mean pooling, respectively. 3 Multimodal feature processing: after spatial downsampling and temporal fusion, the 2D visual features would be integrated into the prompted textual features. 4 Crossmodal fusion: the multimodal features would be processed by a 12-layer transformer encoder, and MLP would predict the starting/ending time points of the target moment. + +Distance Loss $\mathcal{L}_{\mathrm{dis}}$ . To avoid the gradient vanishing problem caused by the non-overlapping case, we involve distance loss $\mathcal{L}_{\mathrm{dis}}$ to directly minimize the normalized central time point distance. In addition, we add a threshold $\alpha_{1}$ to prevent oscillation in the later training phase. The distance loss is then given by: + +$$ +\mathcal {L} _ {\mathrm {d i s}} = \max \left(\frac {\left| \left(t _ {\mathrm {s t a}} + t _ {\mathrm {e n d}}\right) / 2 - \left(\hat {t} _ {\mathrm {s t a}} + \hat {t} _ {\mathrm {e n d}}\right) / 2 \right|}{\left| \hat {\mathbf {T}} \bigcup \mathbf {T} \right|}, \alpha_ {1}\right), \tag {3} +$$ + +where recall that $\mathbf{T} = (t_{\mathrm{sta}}, t_{\mathrm{end}})$ , $\hat{\mathbf{T}}$ is predicted by the TVG model (1), and we choose $\alpha_1 = 0.2$ in experiments. + +Duration Loss $\mathcal{L}_{\mathrm{dur}}$ . The introduction of distance loss $\mathcal{L}_{\mathrm{dis}}$ avoids the gradient vanishing problem but only considers the central time point distance. Yet, this may not be precise enough. For example, even if the central time points are completely overlapped, the duration of two video clips may not be identical. Inspired by the above, we propose the duration loss: + +$$ +\mathcal {L} _ {\mathrm {d u r}} = \max \left(\frac {| \mathbf {T} - \hat {\mathbf {T}} (\boldsymbol {\theta}) |}{| \mathbf {T} |}, \alpha_ {2}\right), \tag {4} +$$ + +where $\alpha_{2}$ is the precision tolerance threshold and set by 0.4 in our experiments. + +Finally, the proposed Temporal-Distance IoU (TDIoU) loss is given by + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {t I o U}} + \beta_ {1} \mathcal {L} _ {\mathrm {d i s}} + \beta_ {2} \mathcal {L} _ {\mathrm {d u r}}, \tag {5} +$$ + +where $\beta_{1} > 0$ and $\beta_{2} > 0$ are regularization parameters. + +# 3.3. Text-Visual Prompt Design + +Inspired by the application of prompts on transformers [2, 21, 36, 37], we propose jointly text-visual prompting to boost the performance of our models, in which prompts are optimized perturbation patterns. To improve data processing efficiency, we uniformly sample video frames from the untrimmed video $\mathbf{v}$ to obtain $\mathbf{v}_{\mathrm{sam}} \in \mathbb{R}^{N_{\mathrm{sam}} \times C \times H \times W}$ , where $N_{\mathrm{sam}}$ is the number of sampled video frames. In addition, we introduce a set of frame-aware visual prompts $\delta_{\mathrm{vp}} \in \mathbb{R}^{N_{\mathrm{sam}} \times d_{\mathrm{vp}}}$ in the pixel space of sampled video frames $\mathbf{v}_{\mathrm{sam}}$ , and introduce text prompts $\delta_{\mathrm{tp}} \in \mathbb{R}^{N_{\mathrm{tp}} \times d_{\mathrm{tp}}}$ in the textual feature space. By incorporating video frame sampling and text-visual prompts into the TVG model (1), we obtain: + +$$ +\left(\hat {t} _ {\mathrm {s t a}}, \hat {t} _ {\mathrm {e n d}}\right) = f \left(\delta_ {\mathrm {t p}}, g _ {\mathrm {t e x}} (\mathbf {s}), g _ {\mathrm {v i d}} \left(\mathbf {v} _ {\mathrm {s a m}} + \delta_ {\mathrm {v p}}\right)\right). \tag {6} +$$ + +Given a pre-trained 2D TVG model $f$ , the objective of text-visual prompting (TVP) is to learn a universal set of visual prompts $\delta_{\mathrm{vp}}$ and text prompts $\delta_{\mathrm{tp}}$ to be integrated into sampled video frames and textual features, respectively. Specially, a set of different visual prompts are applied to uniformly-sampled frames of one untrimmed video in order. During training, only the set of visual prompts and text prompts are updated through backpropagation. During finetuning, prompts are frozen, and the parameters of the TVG model and encoders are updated. During testing, the set of optimized visual prompts and the optimized text prompts are applied to all test-time video-query pairs. + +# 3.4. Framework + +Inspired by the success of transformers in vision-language tasks, we choose ClipBERT [31] as the base model for 2D TVG. Extended from ClipBERT, the input of our regression-based TVG model would be describable sentences and uniformly sampled frames of one untrimmed video as shown in Fig. 3. Then, the predicted starting and ending time points of the target video clip would be model outputs. As described in Algorithm 1, there are four phases of our proposed TVP framework: Video frame preprocessing: We obtain sparsely-sampled frames $\mathbf{v}_{\mathrm{sam}}$ from one input untrimmed video $\mathbf{v}$ , and apply universal frame-aware visual prompts $\delta_{\mathrm{vp}}$ on top of frames at the padding location. Feature extraction: 2D vision encoder (first 5 ConvBlock of ResNet-50) $g_{\mathrm{vid}}$ and language encoder (a trainable word embedding layer) $g_{\mathrm{tex}}$ would extract features from the prompted frames $\mathbf{v}_{\mathrm{sam}}'$ and textual inputs $\mathbf{s}$ , respectively. Multimodal feature processing: Following the setting of Pixel-BERT [22], the 2D visual features $\mathbf{Q}_{\mathrm{vid}}$ are downsampled spatially by a $2 \times 2$ max-pooling layer and fused temporally by a mean-pooling layer. Then, text prompts $\delta_{\mathrm{tp}}$ are integrated into textual features $\mathbf{Q}_{\mathrm{tex}}$ . In addition, trainable 2D visual position embeddings $\mathbf{M}_{2\mathrm{D}}$ and textual position embeddings $\mathbf{M}_{\mathrm{pos}}$ are applied to the processed 2D visual features $\mathbf{Q}_{\mathrm{vid}}'$ and prompted textual features $\mathbf{Q}_{\mathrm{tex}}'$ , respectively [10, 31]. Afterwards, the processed and position-encoded 2D visual features $\mathbf{Q}_{\mathrm{vid}}''$ are flattened and integrated into prompted and position-encoded textual features $\mathbf{Q}_{\mathrm{tex}}''$ . Moreover, type embeddings $\mathbf{M}_{\mathrm{type}}$ would be added to the integrated multimodal features $\mathbf{Q}_{\mathrm{all}}$ to indicate the source type of features. Crossmodal fusion: A 12-layer transformer [10] is utilized for crossmodal fusion on $\mathbf{Q}_{\mathrm{all}}$ , and then multilayer perceptron (MLP) ending with sigmoid function is used as the prediction head to process the last-layer crossmodal representation $\mathbf{Q}_{\mathrm{CM}}$ of the transformer for generating the predicted starting/ending time points $(\hat{t}_{\mathrm{sta}}, \hat{t}_{\mathrm{sta}})$ of the target moments described by the text query input. + +# 4. Experiments + +In this section, we demonstrate the effectiveness of our proposed TVP framework on Charades-STA and ActivityNet Captions datasets. + +# 4.1. Experiment Setup + +Datasets. The evaluations are implemented on two standard benchmark datasets for TVG task, Charades-STA [14] and ActivityNet Captions [28]. Tab. 1 summarizes the details of both datasets. Charades-STA dataset contains 6,672 videos and 16,124 text queries in total. The average length of videos is $30.6s$ , and the average length of text query is 7.2 words. The average length of moments cor + +# Algorithm 1 Overview of TVP framework + +Input: vision encoder $g_{\mathrm{vid}}$ , language encoder $g_{\mathrm{tex}}$ , position embeddings $\mathbf{M}_{\mathrm{pos}}$ , 2D position embeddings $\mathbf{M}_{\mathrm{2D}}$ , type embeddings $\mathbf{M}_{\mathrm{type}}$ , transformer $f$ , prediction head $MLP$ , visual prompts $\delta_{\mathrm{vp}}$ , text prompts $\delta_{\mathrm{tp}}$ +Output: Predicted time interval $\hat{\mathbf{T}} = (\hat{t}_{\mathrm{sta}},\hat{t}_{\mathrm{end}})$ +Phase ①: Video frame preprocessing + +1: $\mathbf{v}_{\mathrm{sam}} \gets$ uniformly sample video frames from an untrimmed video $\mathbf{v}$ +2: $\mathbf{v}_{\mathrm{sam}}^{\prime}\gets$ apply visual prompts $\delta_{\mathrm{vp}}$ to the sampled video frames $\mathbf{v}_{\mathrm{sam}}$ + +# Phase ②: Feature Extraction + +3: $\mathbf{Q}_{\mathrm{vid}} = g_{\mathrm{vid}}(\mathbf{v}_{\mathrm{sam}}^{\prime})\gets$ extracting 2D visual features +4: $\mathbf{Q}_{\mathrm{tex}} = g_{\mathrm{tex}}(\mathbf{s})$ ← extracting textual features + +# Phase 8: Multimodal feature processing + +5: $\mathbf{Q}_{\mathrm{vld}}^{\prime}\gets$ apply spatial downsampling and temporal fusion to 2D visual features $\mathbf{Q}_{\mathrm{vld}}$ +6: $\mathbf{Q}_{\mathrm{tex}}^{\prime}\gets$ apply text prompts $\delta_{\mathrm{tp}}$ to textual features $\mathbf{Q}_{\mathrm{tex}}$ +7: $\mathbf{Q}_{\mathrm{vid}}^{\prime \prime} \gets$ add 2D visual position embeddings $\mathbf{M}_{2\mathrm{D}}$ on the processed 2D visual features $\mathbf{Q}_{\mathrm{vid}}^{\prime}$ +8: $\mathbf{Q}_{\mathrm{tex}}^{\prime \prime}\gets$ add position embeddings $\mathbf{M}_{\mathrm{pos}}$ to prompted textual features $\mathbf{Q}_{\mathrm{tex}}^{\prime}$ +9: $\mathbf{Q}_{\mathrm{all}} \gets$ integrate the processed and position-encoded textual features $\mathbf{Q}_{\mathrm{tex}}^{\prime \prime}$ and the processed and position-encoded 2D visual features $\mathbf{Q}_{\mathrm{vid}}^{\prime \prime}$ +10: $\mathbf{Q}_{\mathrm{all}} + \mathbf{M}_{\mathrm{type}} \gets$ add type embeddings $\mathbf{M}_{\mathrm{type}}$ to the integrated multimodal features $\mathbf{Q}_{\mathrm{all}}$ + +# Phase 4: Crossmodal fusion + +11: $\mathbf{Q}_{\mathrm{CM}} = f(\mathbf{Q}_{\mathrm{all}} + \mathbf{M}_{\mathrm{type}}) \gets$ implement crossmodal fusion through transformer $f$ +12: $(\hat{t}_{\mathrm{sta}},\hat{t}_{\mathrm{end}}) = MLP(\mathbf{Q}_{\mathrm{CM}})\gets$ prediction head generates the predicted time interval according to crossmodal representation $\mathbf{Q}_{\mathrm{CM}}$ + +Table 1. Statistics of TVG benchmark datasets (Charades-STA and ActivityNet Captions datasets). + +
DatasetCharades-STAActivityNet Captions
DomainIndoor ActivityIndoor/Outdoor Activity
# Videos6,67214,926
Avg. Video Length (second)30.6117.6
# Moments11,76771,953
Avg. Moment Length (second)8.137.1
Vocabulary Size1,30315,505
# Queries16,12471,953
Avg. Query Length (word)7.214.4
+ +responding to the text query is 8.1s. Following the same dataset split as [14] for fair comparisons, there are 12,408 video-query pairs for training and 3,720 pairs for testing. ActivityNet Captions dataset contains 14,926 videos and 71,953 text queries in total. The average length of videos is 117.6s, and the average length of text query is 14.4 words. + +The average length of moments corresponding to the text query is 37.1s. ActivityNet Captions dataset is split into training set, validation set, and testing set in a $2:1:1$ ratio. Since the testing set is withheld for competition, only a training set and two validation sets (val1 and val2) can be accessed publicly. For fair comparisons, we evaluate our proposed framework on val1. + +Baselines. We compare our proposal with 15 baseline methods: ① Proposal-based: CTRL [14], MCN [1], SAP [7], BPNet [62], LPNet [61], QSPN [64], MAN [71]; ② Proposal-free: ABLR [67], DRN [69], CPNet [34], DEBUG [43], ExCL [16], VSLNet [73]; ③ Reinforcement learning: TSP-PRL [60], TripNet [18]. + +Evaluation metrics. Following [14], we adopt Acc(R@1, IoU=m) as the performance evaluation metric, which represents the percentage accuracy of top-1 predicted moments whose tIoU (temporal IoU) with the ground-truth moment is larger than $m$ . By convention, we consider the following tIoU threshold values $m = \{0.3, 0.5, 0.7\}$ . + +Crossmodal pretraining setup. Our 2D vision encoder (ResNet-50) is initialized with the weight from gridfeat [24], which can extract effective grid features from visual inputs. In addition, both the language encoder and 12-layer transformer are initialized with the BERT-base model weight [10], which are pretrained on English Wikipedia and BookCorpus [83]. Thanks to the compact 2D vision encoder, TVP (our proposal) is able to directly utilize image-text pairs for end-to-end training. Since the benefits of cross-modal pretraining has been demonstrated by [22, 44, 55], our base model is pretrained on two large-scale image-text datasets, which are Visual Genome Captions [29] and COCO Captions [8]. To be more specific, image-text matching [44, 55] and masked language modeling [10] are employed for cross-modal pretraining. + +Implementation setup. For video inputs, we uniformly sample $N_{\mathrm{sam}}$ frames from a video ( $N_{\mathrm{sam}} = 48$ for Charades-STA and $N_{\mathrm{sam}} = 64$ for ActivityNet Captions). In addition, all video frames are resized to have a maximum longer side of 448 with an original aspect ratio, and then the frames are zero-padded to $448 \times 448$ . The default visual prompt sizes for both dataset are 96. The default text prompt sizes are 10 and 20 for Charades-STA and ActivityNet Captions, respectively. We utilize the first 5 ConvBlocks of ResNet-50 as the 2D vision encoder and a trainable embedding layer as the language encoder for both Charades-STA and ActivityNet Captions datasets. For text queries, all word tokens are maintained after lower-case conversion and tokenization. We use AdamW [42] for end-to-end model training, with $\beta_{1} = 1.0$ , $\beta_{2} = 0.1$ , $\alpha_{1} = 0.2$ , $\alpha_{2} = 0.4$ . Initial learning rates are $1e - 1$ and $5e - 7$ for prompt training and model finetuning, respectively. In addition, the learning rate linearly decays to 0 with the first $10\%$ training step for warmup. Our experiments are imple + +Table 2. Performance comparison of different thresholds $m$ on the Charades-STA dataset. + +
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D-23.638.89
ABLR [67]C3D-24.369.01
BPNet [62]C3D55.4638.2520.51
LPNet [61]C3D59.1440.9421.13
QSPN [64]C3D54.7035.6015.80
TSP-PRL [60]C3D-45.4524.75
TripNet [18]C3D54.6438.2916.07
DRN [69]C3D-45.4026.40
CPNet [34]C3D-40.3222.47
DEBUG [43]C3D54.9537.3917.92
ExCL [16]I3D61.5044.122.40
VSLNet [73]I3D64.3047.3130.19
MAN [71]I3D-46.5322.72
2D TVGMCN [1]VGG-17.468.01
SAP [7]VGG-27.4213.36
Ours
TVP-Based 2D TVGBase w/o promptsResNet61.2940.4319.89
Base + Visual Prompts65.3844.3120.22
Base + Text Prompts65.8143.4420.65
Base + Both Prompts65.9244.3921.51
+ +Table 3. Performance comparison of different thresholds $m$ on the ActivityNet Captions dataset. + +
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D28.7014.00-
BPNet [62]C3D59.9842.0724.69
LPNet [61]C3D64.2945.9225.39
QSPN [64]C3D45.3027.7013.60
TSP-PRL [60]C3D56.0238.83-
TripNet [18]C3D48.4232.1913.93
DRN [69]C3D-45.4524.36
CPNet [34]C3D-40.5621.63
ABLR [67]C3D55.6736.79-
DEBUG [43]C3D55.9139.72-
ExCL [16]C3D63.0043.6024.10
VSLNet [73]C3D63.1643.2226.16
Ours
TVP-Based 2D TVGBase w/o promptsResNet57.2040.1619.14
Base + Visual Prompts60.1243.3923.71
Base + Text Prompts60.4842.5824.39
Base + Both Prompts60.7143.4425.03
+ +mented in PyTorch [45], and models and prompts are finetuned separately for 12 epochs with the mixed precision on 8 NVIDIA V100 GPUs. + +# 4.2. Experiment Results + +Effectiveness of TVP on Charades-STA. The performance comparisons with SOTA methods on the Charades-STA dataset are summarized in Tab. 2. Our proposed TVP framework can achieve competitive performance at all IoU thresholds $m$ in the case of utilizing 2D visual features extracted by ResNet-50, and reach the highest score at $m = 0.3$ . Compared to the 2D TVG methods using VGG as the vision encoder, our proposed framework could achieve around $2.5 \times$ and $2.7 \times$ performance gain at thresholds 0.5 and 0.7, respectively. Furthermore, we can find that for our base model only one of visual prompts and text prompts can achieve up to $7.37\%$ and $9.60\%$ improvement + +![](images/f930cf36323498ab1ee69cd319022a82926fb4450ac7405b8a56d5e204390eb4.jpg) +(a) Charades-STA + +![](images/7f9010f95a5a2247a77be25bc98fd2db502eaf9d81d99450cd3a28ae2a91b6da.jpg) +(b) ActivityNet Captions +Figure 4. Impact of sampled frame numbers. + +at tIoU thresholds $m = 0.3$ and $m = 0.5$ . The combination of text and visual prompts can not only achieve $7.55\%$ and $9.79\%$ improvements at tIoU thresholds $m = 0.3$ and $m = 0.5$ , but also improve the performance by $8.14\%$ at $m = 0.7$ . This demonstrates the effectiveness and necessity of the joint text-visual prompting. + +Effectiveness of TVP on ActivityNet Captions. We focus on the performance comparisons with 3D TVG methods on ActivityNet since there are no results of 2D TVG method reported on ActivityNet Captions. The results of multiple methods on ActivityNet Captions datasets are reported in Tab. 3. Even on this more challenging dataset, our proposed method still has achieved competitive performance compared to 3D TVG methods. Different from the performance of TVP on Charades-STA dataset, text prompts or visual prompts can achieve a significant performance boost on the base model over all IoU thresholds $m$ alone (5.73% at $m = 0.3$ , 8.04% at $m = 0.5$ , 27.43% at $m = 0.7$ ), and the text-visual prompt combination could further boost the performance (6.14% at $m = 0.3$ , 8.17% at $m = 0.5$ , 30.77% at $m = 0.7$ ). It is worth noting that the performance gap over $m = 0.7$ between 2D TVG methods and 3D TVG methods is narrowed significantly. + +In summary, through the experimental results on Charades-STA and ActivityNet Captions datasets, we can find that our proposed TVP framework could achieve competitive performance overall tIoU thresholds on Charades-STA and ActivityNet Captions by improving the utility of sparse 2D visual features. Thanks to the lightweight 2D vision encoder, cotraining language encoder and vision encoder on large-scale image-text datasets can be performed, which benefits the base model to achieve good performance. Furthermore, the combination of text and visual prompts can achieve better results than any single kind of prompts on both datasets, which again proves the importance of cross-modal training. + +Video frame sampling effect. Fig. 4 demonstrates the performance of base model with different number $N_{\mathrm{sam}}$ + +Table 4. The performance comparison of different visual prompt sizes on Charades-STA dataset. + +
Visual Prompt SizeAcc(R@1, IoU=m)Prompt + Frame
m=0.3m=0.5m=0.7
061.2940.4319.89
1661.2940.4320.00
3261.9439.7819.35
4863.6642.3720.00
7263.8743.6619.78
9665.3844.3120.22
12864.7343.6619.78
+ +Table 5. The performance comparison of different text prompt sizes on Charades-STA dataset. + +
Text Prompt SizeAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
057.2040.1619.14
565.3841.9420.43
1065.8143.4420.65
1565.5943.2321.29
2064.9543.8721.51
2563.6642.8020.65
3064.4642.6320.51
+ +of sampled video frames as visual inputs. For Charades dataset, the base model performance keeps increasing before $N_{\mathrm{sam}}$ reaches 48, but when it exceeds 48, performance starts to degrade. This is because frequent background changes harm the performance of object re-identification in videos, which are noisy for object motion analysis [17]. + +For ActivityNet Caption dataset, base model performance continues to improve even when sampled frame number $N_{\mathrm{sam}}$ exceeds 48, due to the longer average video length in ActivityNet Captions dataset. Balancing the frame number and batch size for training, we choose $N_{\mathrm{sam}} = 64$ for ActivityNet Captions. + +TVP performance vs. prompt size. As shown in Tab. 4, we can find that when visual prompts are small, they cannot bring changes to the base model, and when visual prompts are too large, the performance starts to decrease. This is because key information within video frames might be removed. However, the text prompts can bring significant performance boost even when the text prompt size is small as shown in Tab. 5, which is because the textual features has a smaller dimension compared to visual features, and also the text prompts are directly optimized in feature space during training. + +TVP performance vs. visual prompt operation. Visual prompt is first proposed by [2], where visual prompts are + +Table 6. The performance comparison of different visual prompt operations ('remove', 'add', 'replace') with fixed visual prompt size $p = 96$ on Charades-STA and ActivityNet Captions datasets. + +
OperationCharades-STAActivityNet Captions
R@1, IoU=mR@1, IoU=m
m=0.3m=0.5m=0.7m=0.3m=0.5m=0.7
Original61.2940.4319.8957.2040.1619.14
Remove61.2940.4320.057.2040.1619.14
Add61.0839.5720.2257.1540.1619.27
Replace65.3844.3120.2260.1243.3923.71
+ +![](images/a82c52377ad36275e67254268d4a39d76b11b976001d259d05775ede79124c0f.jpg) +(a) Vison Encoders +Figure 5. Inference time comparison. (a) inference time comparison between 2D vision encoder (ResNet-50) and 3D vision encoder (C3D). (b) inference time comparison between the vision encoder and the other modules of the 2D TVG model, where the sampled frame number for our TVP framework is $1.2 \times$ the length of the video in seconds. + +![](images/3623d7cc6fbc44db6248b6619f8e1b6fcc50e40ae5a1a400036e943477297a12.jpg) +(b) Modules in 2D TVG Model + +added to the image for transfer learning on classification tasks. In contrast, our proposed prompting framework is designed to compensate for the spatiotemporal information loss in 2D visual features. Due to the differences in the task, we try two different prompt operation strategies, 'replace' and 'add'. 'add' is to add the visual prompts to the pixel value of the video frame at the corresponding padding locations. 'replace' is to replace the pixel values of video frames with visual prompts at corresponding padding locations. 'remove' is in order to study the impact of removing the pixel values at the padding location. As shown in Tab. 6, 'add' or 'remove' prompt operations have limited effects on the base model. However, 'replace' does boost the base model performance. + +TVP achieves inference efficiency. As shown in Fig. 5, we can find that the inference time required for visual feature extraction accounts for more than half of the inference time of the whole model, while the inference time required for the 3D vision encoder is more than $5 \times$ compared to the 2D vision encoder, and even more than the time required for the whole TVG model using 2D vision encoder, which fully demonstrates the feasibility of accelerating the overall inference speed by reducing the complexity of the vision encoder. Need to note that if there are + +Table 7. The performance comparison of different loss designs on Charades-STA dataset. + +
Loss Function SelectionR@1, IoU=m
m=0.3m=0.5m=0.7
LtIoU55.0529.8911.82
LtIoU + Ldis60.6431.1816.77
LtIoU + Ldur59.7830.9716.34
LtIoU + Ldis + Ldur61.2940.4319.89
+ +![](images/c51b7edc18b9108057885efbb83f8ceb6497e9223bbd4bf91132b1a0c4bdcb4d.jpg) +Figure 6. Loss landscape visualization in 2D plane: Finetuning w/o prompts (left) and using prompts (right); see [33] for implementation. + +![](images/c473b2558978f69d281dc660c61bf9dac140ef26ca04348f7cc3e71d31bac6cd.jpg) + +multiple model weights for different sampled frame number settings and model weights can be adopted adaptively for different lengths of videos, the inference speed for short videos should increase, and the prediction results for long videos will be further improved. + +Ablation studies. Through Tab. 7, we can find that the addition of either distance loss $\mathcal{L}_{\mathrm{dis}}$ or duration loss $\mathcal{L}_{\mathrm{dur}}$ will result in a performance increase, but the combination of the two will result in a significant performance increase (11.34% at $m = 0.3$ , 35.26% at $m = 0.5$ , 68.27% at $m = 0.7$ ), especially over tIoU thresholds $m = 0.5$ and $m = 0.7$ . This demonstrates that distance loss $\mathcal{L}_{\mathrm{dis}}$ and duration loss $\mathcal{L}_{\mathrm{dur}}$ could provide more precise training guides compared to only using temporal IoU loss $\mathcal{L}_{\mathrm{tIoU}}$ . Furthermore, we posit that prompting may encode additional spatial-temporal supervision to help the model trainer to escape from bad local optima as shown in Fig. 6, where finetuning w/ prompts yields a flatter loss landscape than the one w/o prompts. + +# 5. Conclusion + +In this paper, we propose text-visual prompting to boost the performance of 2D TVG methods by compensating for the lack of spatiotemporal information in 2D visual features. In contrast to 3D TVG methods, TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. The effectiveness of our proposed TVP (text-visual prompting) framework has been demonstrated on two standard datasets, Charades-STA and ActivityNet. Our models outperform all 2D models significantly, and also achieve comparable performance to 3D models. What is more, we achieve over $5 \times$ inference speedup over TVG methods of using 3D visual features. + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 6 +[2] H Bahng, A Jahanian, S Sankaranarayanan, and P Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, page 2022, 2022. 2, 3, 4, 7 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1 +[4] Aochuan Chen, Peter Lorenz, Yuguang Yao, Pin-Yu Chen, and Sijia Liu. Visual prompting for adversarial robustness. arXiv preprint arXiv:2210.06284, 2022. 3 +[5] Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. arXiv preprint arXiv:2211.11635, 2022. 3 +[6] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. arXiv:2202.10629, 2022.3 +[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8199-8206, 2019. 1, 2, 6 +[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6 +[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 6 +[11] Gamaleldin F Elsayed, Ian Goodfellow, et al. Adversarial reprogramming of neural networks. arXiv:1806.11146, 2018. 3 +[12] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3 +[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3 +[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6 + +[15] Yanjun Gao, Lulu Liu, Jason Wang, Xin Chen, Huayan Wang, and Rui Zhang. Evoquer: Enhancing temporal grounding with video-pivoted backquery generation. arXiv preprint arXiv:2109.04600, 2021. 1 +[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 1, 3, 6 +[17] Xinqian Gu, Hong Chang, Bingpeng Ma, Hongkai Zhang, and Xilin Chen. Appearance-preserving 3d convolution for video-based person re-identification. In ECCV, 2020. 7 +[18] Meera Hahn, Asim Kadav, James M Rehg, and Hans Peter Graf. Tripping through time: Efficient localization of activities in videos. arXiv preprint arXiv:1904.09936, 2019. 1, 3, 6 +[19] Xiaochen Han, Bo Wu, Zheng Shou, Xiao-Yang Liu, Yi-meng Zhang, and Linghe Kong. Tensor fista-net for real-time snapshot compressive imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10933-10940, 2020. 1 +[20] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 2 +[21] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4 +[22] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. arXiv preprint arXiv:2004.00849, 2020. 2, 5, 6 +[23] Jinghan Jia, Mingyi Hong, Yimeng Zhang, Mehmet Akçakaya, and Sijia Liu. On the robustness of deep learning-based mri reconstruction to image transformations. arXiv preprint arXiv:2211.04930, 2022. 1 +[24] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10267-10276, 2020. 6 +[25] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 2, 3 +[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 2 +[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. arXiv preprint arXiv:2210.03117, 2022. 2, 3 +[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5 + +[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6 +[30] Xiaohan Lan, Yitian Yuan, Xin Wang, Zhi Wang, and Wenwu Zhu. A survey on temporal sentence grounding in videos. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 2021. 1 +[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 5 +[32] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3 +[33] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 2018. 8 +[34] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 1, 2, 6 +[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2 +[36] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4 +[37] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4 +[38] Yanyu Li, Pu Zhao, Geng Yuan, Xue Lin, Yanzhi Wang, and Xin Chen. Pruning-as-search: Efficient neural architecture search via channel pruning and structural reparameterization. arXiv preprint arXiv:2206.01198, 2022. 1 +[39] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2 +[40] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 2, 3 +[41] Xiao-Yang Liu, Yimeng Zhang, Yukang Liao, and Ling Jiang. Dynamic updating of the knowledge base for a large-scale question answering system. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 2020. 1 +[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural + +language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 1, 6 +[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2, 6 +[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 2019. 6 +[46] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3 +[47] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2 +[48] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In proceedings of the IEEE International Conference on Computer Vision, 2017. 3 +[49] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 3 +[50] Timo Schick and Hinrich Schütze. Few-shot text generation with pattern-exploiting training. arXiv preprint arXiv:2012.11926, 2020. 3 +[51] Timo Schick and Hinrich Schütze. It's not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118, 2020. 3 +[52] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3 +[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 2014. 3 +[54] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019.2 +[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2, 6 +[56] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3 +[57] Yun-Yun Tsai et al. Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. arXiv:2007.08714, 2020. 3 + +[58] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3 +[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 3 +[60] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12386–12393, 2020. 1, 2, 3, 6 +[61] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. arXiv preprint arXiv:2109.10678, 2021. 1, 2, 6 +[62] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 1, 2, 6 +[63] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), 2018. 3 +[64] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9062–9069, 2019. 1, 2, 6 +[65] Chao-Han Huck Yang, Yun-Yun Tsai, et al. Voice2series: Reprogramming acoustic models for time series classification. In ICML. PMLR, 2021. 3 +[66] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3 +[67] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 1, 3, 6 +[68] Ofir Zafrir, Ariel Larey, Guy Boudoukh, Haihao Shen, and Moshe Wasserblat. Prune once for all: Sparse pre-trained language models. arXiv preprint arXiv:2111.05754, 2021. 1 +[69] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 1, 2, 3, 6 +[70] Yawen Zeng, Da Cao, Xiaochi Wei, Meng Liu, Zhou Zhao, and Zheng Qin. Multi-modal relational graph for cross-modal video moment retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2215-2224, 2021. 2 + +[71] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 6 +[72] Guanhua Zhang, Yihua Zhang, Yang Zhang, Wenqi Fan, Qing Li, Sijia Liu, and Shiyu Chang. Fairness reprogramming. arXiv preprint arXiv:2209.10222, 2022.3 +[73] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 1, 6 +[74] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071, 2022. 1 +[75] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 1, 2 +[76] Yimeng Zhang, Akshay Karkal Kamath, Qiucheng Wu, Zhiwen Fan, Wuyang Chen, Zhangyang Wang, Shiyu Chang, Sijia Liu, and Cong Hao. Data-model-circuit tri-design for ultra-light video intelligence on edge devices. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 745-750, 2023. 1 +[77] Yimeng Zhang, Xiao-Yang Liu, Bo Wu, and Anwar Walid. Video synthesis via transform-based tensor neural network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2454-2462, 2020. 1 +[78] Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jinfeng Yi, Mingyi Hong, Shiyu Chang, and Sijia Liu. How to robustify black-box ml models? a zeroth-order optimization perspective. arXiv preprint arXiv:2203.14195, 2022. 1 +[79] Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Mingyi Hong, Yanzhi Wang, and Sijia Liu. Advancing model pruning via bi-level optimization. arXiv preprint arXiv:2210.04092, 2022. 1 +[80] Yihua Zhang, Guanhua Zhang, Prashant Khanduri, Mingyi Hong, Shiyu Chang, and Sijia Liu. Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In International Conference on Machine Learning, pages 26693-26712. PMLR, 2022. 1 +[81] Yang Zheng, Xiaoyi Feng, et al. Why adversarial reprogramming works, when it fails, and how to tell the difference. arXiv:2108.11673, 2021. 3 +[82] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020. 3 +[83] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19-27, 2015. 6 \ No newline at end of file diff --git a/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/images.zip b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..210ac393ab1967a459cbf41270146281e3f0ae13 --- /dev/null +++ b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e170ddd8069492e3535a9dae9512708c7ef1b1293470b754f05da3000e572ce +size 546365 diff --git a/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/layout.json b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18245de5a7016ec46d0465eab5503ca6efae085c --- /dev/null +++ b/2023/Text-Visual Prompting for Efficient 2D Temporal Video Grounding/layout.json @@ -0,0 +1,11445 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 90, + 103, + 504, + 121 + ], + "type": "text", + "content": "Text-Visual Prompting for Efficient 2D Temporal Video Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": "Yimeng Zhang" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": ", Xin Chen" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": ", Jinghan Jia" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": ", Sijia Liu" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": ", Ke Ding" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": "Michigan State University, " + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 142, + 453, + 201 + ], + "type": "text", + "content": "Applied AI, Intel \n{zhan1853, jiajingh, liusiji5}@msu.edu, \n{xin.chen, ke.ding}@intel.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "text", + "content": "In this paper, we study the problem of temporal video grounding (TVG), which aims to predict the starting/ending time points of moments described by a text sentence within a long untrimmed video. Benefiting from fine-grained 3D visual features, the TVG techniques have achieved remarkable progress in recent years. However, the high complexity of 3D convolutional neural networks (CNNs) makes extracting dense 3D visual features time-consuming, which calls for intensive memory and computing resources. Towards efficient TVG, we propose a novel text-visual prompting (TVP) framework, which incorporates optimized perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. In sharp contrast to 3D CNNs, we show that TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. Further, we propose a Temporal-Distance IoU (TDIoU) loss for efficient learning of TVG. Experiments on two benchmark datasets, Charades-STA and ActivityNet Captions datasets, empirically show that the proposed TVP significantly boosts the performance of 2D TVG (e.g., " + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "9.79\\%" + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "text", + "content": " improvement on Charades-STA and " + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "30.77\\%" + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "text", + "content": " improvement on ActivityNet Captions) and achieves " + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 251, + 290, + 563 + ], + "type": "text", + "content": " inference acceleration over TVG using 3D visual features. Codes are available at Open. Intel." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "In recent years, we have witnessed great progress on temporal video grounding (TVG) [30, 74]. One key to this success comes from the fine-grained dense 3D visual features extracted by 3D convolutional neural networks (CNNs) (e.g., C3D [56] and I3D [3]) since TVG tasks demand spatial-temporal context to locate the temporal interval of the moments described by the text query. However, due to the high cost of the dense 3D feature extraction, most existing TVG models only take these 3D visual features ex" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 228, + 544, + 336 + ], + "blocks": [ + { + "bbox": [ + 309, + 228, + 544, + 336 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 544, + 336 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 544, + 336 + ], + "type": "image", + "image_path": "7481fb8c16d19e42311863a71b265ed5a2287a76e309ce3d7b1024aa2f52b992.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 338, + 422, + 443 + ], + "blocks": [ + { + "bbox": [ + 309, + 338, + 422, + 443 + ], + "lines": [ + { + "bbox": [ + 309, + 338, + 422, + 443 + ], + "spans": [ + { + "bbox": [ + 309, + 338, + 422, + 443 + ], + "type": "image", + "image_path": "779651af2c94ec20721ac8fb8fe00ac25d73520599819a185e8d4f1de6bfdfe9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 453, + 545, + 559 + ], + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 559 + ], + "type": "text", + "content": "Figure 1. The architecture and performance comparison among TVG methods: a) 3D TVG methods [14, 16, 18, 34, 43, 60-62, 64, 67, 69, 71, 73], b) 2D TVG methods [1, 7], and c) TVP-based 2D TVG (Ours), d) overall performance comparison. Ours is the most efficient (least inference time) and achieves competitive performance compared to 3D TVG methods. In contrast to existing TVG methods, which utilize dense video features extracted by non-trainable offline 3D CNNs and textual features, our proposed framework utilizes a trainable 2D CNN as the vision encoder to extract features from sparsely-sampled video frames with a universal set of frame-aware visual prompts and adds text prompts in textual feature space for end-to-end regression-based modeling." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 336, + 542, + 443 + ], + "blocks": [ + { + "bbox": [ + 434, + 336, + 542, + 443 + ], + "lines": [ + { + "bbox": [ + 434, + 336, + 542, + 443 + ], + "spans": [ + { + "bbox": [ + 434, + 336, + 542, + 443 + ], + "type": "image", + "image_path": "84dd07cd3e3a775994638e3f31ae76ba1aadeebb6da5242c165c28bdef03b367.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 568, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 592 + ], + "type": "text", + "content": "tracted by offline 3D CNNs as inputs instead of co-training during TVG model training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Although models using 3D visual features (that we call '3D methods or models') outperform these using the 2D features (that we call '2D methods or models'), a unique advantage of 2D methods is that extracting 2D visual features can significantly reduce the cost in TVG tasks [14, 15, 30, 34, 61, 62, 69, 74, 75]. An efficient and lightweight solution with reasonable performance is also demanded in computer vision, NLP, and video-language tasks [19, 23, 38, 41, 68, 76-80]. As discussed above, the methods employing 3D video features are challenging to be" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14794" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": "employed in practical applications. It thus has significant practical and economic value to develop compact 2D solutions for TVG tasks. In this paper, we ask:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 124, + 271, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 124, + 271, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 124, + 271, + 148 + ], + "type": "text", + "content": "How to advance 2D TVG methods so as to achieve comparable results to 3D TVG methods?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 162, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 293 + ], + "type": "text", + "content": "To address this problem, we propose a novel text-visual prompting (TVP) framework for training TVG models using 2D visual features. As shown in Fig. 1, for existing 2D TVG and 3D TVG methods, they all utilize offline pretrained vision encoders and language encoders to perform feature extraction. In contrast, our proposed TVP framework is end-to-end trainable. Furthermore, benefiting from text-visual prompting and cross-modal pretraining on large-scale image-text datasets, our proposed framework could achieve comparable performance to 3D TVG methods with significant inference time acceleration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 288, + 520 + ], + "type": "text", + "content": "Conventionally, TVG methods consist of three stages: ① extracting feature from visual and text inputs; ② multimodal feature fusion; ③ cross-modal modelling. In contrast to conventional methods, TVP incorporates optimized input perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. We apply trainable parameters in the textual features as text prompts and develop a universal set of frame-aware patterns as visual prompts. Specially, we sample a fixed number of frames from a video and optimize text prompts for the input query sentence and a set of visual prompts for frames with different temporal locations during training. During testing, the same set of optimized visual prompts and textual prompts are applied to all test-time videos. We refer readers to Fig. 2 for illustrations of visual prompts and text prompts introduced. To the best of our knowledge, our work makes the first attempt to utilize prompt learning to successfully improve the performance of regression-based TVG tasks using 2D visual features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Compared to 3D CNNs, 2D CNNs lose spatiotemporal information of the video during feature extraction. Inspired by the success of transformers on the vision-language tasks [9, 22, 35, 44, 47, 54, 55] and the recent application of prompt learning to transformers in both vision and language domains [2, 25, 27, 32, 37, 40], we choose transformer as our base TVG model and propose to utilize prompts to compensate for the lack of spatiotemporal information in 2D visual features. Furthermore, we develop a Temporal-Distance IoU (TDIoU) loss for training our proposed framework. There are two aspects that distinguish our proposed framework from existing works. First, our proposed framework is designed to boost the performance of the regression-based TVG methods utilizing 2D CNNs as the vision encoder, not for transfer learning [2, 21, 26] Second, our proposed framework utilizes 2D CNN to extract visual features from" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 73, + 392, + 188 + ], + "blocks": [ + { + "bbox": [ + 324, + 73, + 392, + 188 + ], + "lines": [ + { + "bbox": [ + 324, + 73, + 392, + 188 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 392, + 188 + ], + "type": "image", + "image_path": "8f460a3358496b498f2d79bbe0b7fd554ad0d4c9645dbf7a2ffbb32976efa959.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 194, + 386, + 203 + ], + "lines": [ + { + "bbox": [ + 331, + 194, + 386, + 203 + ], + "spans": [ + { + "bbox": [ + 331, + 194, + 386, + 203 + ], + "type": "text", + "content": "(a) Text Prompts" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 212, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 241 + ], + "type": "text", + "content": "Figure 2. Text-visual prompting illustration. (a) Text prompts are directly applied in the feature space. (b) A set of visual prompts are applied to video frames in order." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 399, + 73, + 529, + 188 + ], + "blocks": [ + { + "bbox": [ + 399, + 73, + 529, + 188 + ], + "lines": [ + { + "bbox": [ + 399, + 73, + 529, + 188 + ], + "spans": [ + { + "bbox": [ + 399, + 73, + 529, + 188 + ], + "type": "image", + "image_path": "cdd49bfac571a910a37322476ba91c6f4ee2329484bd92f839f088de39cca0aa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 194, + 517, + 203 + ], + "lines": [ + { + "bbox": [ + 411, + 194, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 411, + 194, + 517, + 203 + ], + "type": "text", + "content": "(b) Frame-aware Visual Prompts" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 256, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 352 + ], + "type": "text", + "content": "sparsely-sampled video frames, which requires less memory and is easier to be applied in practical applications compared to 3D methods [34,60-62,69,75], especially for long videos. Furthermore, thanks to the compact 2D CNN as the vision encoder, our proposed framework could implement the language encoder and visual encoder co-training for better multimodal feature fusion. In summary, the contributions of this work are unfolded below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 360, + 545, + 613 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 317, + 360, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 545, + 467 + ], + "type": "text", + "content": "- We propose an effective and efficient framework to train 2D TVG models, in which we leverage TVP (text-visual prompting) to improve the utility of sparse 2D visual features without resorting to costly 3D features. To the best of our knowledge, it is the first work to expand the application of prompt learning for resolving TVG problems. Our method outperforms all of 2D methods and achieves competitive performance to 3D TVG methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 475, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 545, + 535 + ], + "type": "text", + "content": "- Technology-wise, we integrate visual prompt with text prompt to co-improve the effectiveness of 2D visual features. On top of that, we propose TDIoU (temporal-distance IoU)-based prompt-model co-training method to obtain high-accuracy 2D TVG models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "text", + "content": "- Experiment-wise, we show the empirical success of our proposal to boost the performance of 2D TVG on Charades-STA and ActivityNet Captions datasets, e.g., " + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "inline_equation", + "content": "9.79\\%" + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "text", + "content": " improvement in Charades-STA, and " + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "inline_equation", + "content": "30.77\\%" + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "text", + "content": " in ActivityNet-Captions together with " + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 317, + 543, + 545, + 613 + ], + "type": "text", + "content": " inference time acceleration over 3D TVG methods." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 633, + 392, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 392, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 392, + 646 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Video Temporal Grounding (TVG). The objective of the TVG is to predict the starting/ending time points of target moments within an untrimmed video, which is described by a text sentence. Early TVG solutions [7,14,20,39,62,64,70] mainly employ two-stage \"propose-and-rank\" pipeline: ①" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14795" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 288 + ], + "type": "text", + "content": "Propose: utilize sliding windows or proposal network to generate proposal candidates from the input video. ② Rank: the proposed candidates would be ranked according to the text query, and then the proposal with the highest ranking would be the final prediction decision. In contrast to proposal-based methods, regression-based methods [16, 67, 69] directly predict the starting/ending time points of the target moments without ranking massive proposal candidates. Thus, regression-based methods are much faster than proposal-based methods, which is one reason why our work focuses on the regression-based TVG. Furthermore, reinforcement learning (RL)-based methods formulate the TVG task as a sequence of decisions to make [18, 60]. In particular, they train an agent to control the movement of a window by shifting or scaling. During training, the agent would be rewarded or punished based on whether the window is close to the target moment after an adjustment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 289, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 289, + 410 + ], + "type": "text", + "content": "Temporal Action Detection (TAD). TAD aims to determine whether predefined actions occur in a video and to predict the corresponding time intervals during which these actions occur [12,13,48,53,56,59,63]. Different from TVG, the input of TAD is only a video. In other words, TAD only requires a semantic understanding of videos. Compared to TAD, TVG is more challenging since it requires a semantic understanding of both videos and natural languages. Furthermore, TVG needs to process the multimodal interaction between videos and natural languages." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 411, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 289, + 567 + ], + "type": "text", + "content": "Text Prompting. Prompting has recently achieved great success in the domain of natural language processing [25, 32, 37, 40, 46, 49-52, 58]. Text prompting is a process that leverages a data-agnostic perturbation operation applied to text inputs or their embeddings to improve the performance of the downstream task. The simplest way is to construct an input context template originating from human contemplation [46, 49-51]. Although the manually-crafted context templates are simple and interpretable, they are typically not the optimal input prompts. To tackle this issue, other work has focused on searching the optimal prompting in the discrete input space [25, 52, 58] or in the language model's embedding space [32, 37, 40]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 568, + 289, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 289, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 289, + 688 + ], + "type": "text", + "content": "Visual Prompting. Inspired by the idea of prompt learning in NLP [37], visual prompting (VP) was first proposed by Bahng et. al. [2] to reprogram a source vision model (e.g., ImageNet-pretrained classifier) to accomplish downstream target tasks (e.g., CIFAR-10 image classification). VP shares almost the same idea with the model reprogramming technology in the vision domain [4-6, 11, 57, 65, 72, 81], which incorporates a universal input perturbation into testing data so as to improve a desired performance metric, e.g., target task accuracy, robustness, and fairness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": "Multi-Modal Prompting. Although visual prompting and text prompting have recently attracted much attention, they" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "are under-explored in the multi-modal learning, especially on the temporal video grounding task. The existing works [2, 27, 66] mainly focus on integrating text and visual prompts with the CLIP (Contrastive Language-Image Pre-trained) model to improve downstream tasks with imagery data. The problem of multi-modal prompting in the video understanding task has not been studied. In this paper, we for the first time develop the text-visual prompting technique to improve the performance of temporal video grounding using 2D visual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 203, + 367, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 367, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 367, + 216 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 223, + 547, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 547, + 271 + ], + "type": "text", + "content": "In this section, we begin with the problem formulation of regression-based TVG. Then we demonstrate the design of TVP (text-visual prompts) and present the overview of our proposed TVP framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 278, + 418, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 278, + 418, + 290 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 418, + 290 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{N_{\\mathrm{vid}} \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " be an untrimmed video consisting of a sequence of " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{vid}}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " video frames, and " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{N_{\\mathrm{tex}}}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " be a text query consisting of a sequence of " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{tex}}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " language tokens. Here, the video-query pair " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}, \\mathbf{s})" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " belongs to a video-language dataset " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": ", TVG aims to predict the time interval " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}} = (\\hat{t}_{\\mathrm{sta}}, \\hat{t}_{\\mathrm{end}})" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": " of the target video moments described by the query " + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 296, + 547, + 393 + ], + "type": "text", + "content": ". The TVG model that fuses the vision-language modalities can be described as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 369, + 399, + 545, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 399, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 369, + 399, + 545, + 414 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {T}} = f \\left(g _ {\\mathrm {t e x}} (\\mathbf {s}), g _ {\\mathrm {v i d}} (\\mathbf {v})\\right), \\tag {1}", + "image_path": "b273c4d4034a0063b2cbdc67ebe203e1277160844d4a19f5e0e28bd948c4e9d5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "text", + "content": " denotes TVG model, and " + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{vid}}" + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tex}}" + }, + { + "bbox": [ + 304, + 422, + 547, + 447 + ], + "type": "text", + "content": " represent vision encoder and language encoder, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 454, + 428, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 428, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 428, + 466 + ], + "type": "text", + "content": "3.2. TDIoU Loss Function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 472, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 545, + 498 + ], + "type": "text", + "content": "Conventionally, the TVG model can be learned by minimizing the temporal IoU loss " + }, + { + "bbox": [ + 306, + 472, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tIoU}}" + }, + { + "bbox": [ + 306, + 472, + 545, + 498 + ], + "type": "text", + "content": " defined below:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 505, + 545, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 505, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 365, + 505, + 545, + 538 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t I o U}} = \\left(1 - \\frac {\\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) \\bigcap \\mathbf {T}}{\\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) \\bigcup \\mathbf {T}}\\right), \\tag {2}", + "image_path": "1329c93e6e47f88730bb989478fe1f42282ddf295320709e5594a6cd0916a52c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": "where for ease of notation let " + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": " denote all the trainable parameters involved in (1), and " + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = (t_{\\mathrm{sta}}, t_{\\mathrm{end}})" + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": " is the label (i.e., the ground-truth time interval) of the target moment associated with the input video-query pair " + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}, \\mathbf{s})" + }, + { + "bbox": [ + 304, + 545, + 547, + 617 + ], + "type": "text", + "content": ". The rationale behind (2) is to maximize the overlapping between the predicted time interval and its ground truth." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "However, for non-overlapping cases, the temporal IoU loss " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tIoU}}" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " would encounter a gradient vanishing problem. Inspired by [82], we develop a novel TDIoU (Temporal-Distance IoU) loss for training our proposed TVG models by incorporating the normalized central time point distance and duration difference between the predicted video clips and the target video clips. We elaborate on the proposed loss below." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14796" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 537, + 266 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 537, + 266 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 537, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 537, + 266 + ], + "type": "image", + "image_path": "ad268c97dbc9581d78d148a1fb367ff2a6e09f3845d89ea4aa24db4a76da923d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 548, + 341 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 548, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 548, + 341 + ], + "type": "text", + "content": "Figure 3. Overview of our proposed TVP (text-visual prompting) framework for 2D TVG (temporal video grounding). The whole process contains four phases: 1 Video frame preprocessing: uniformly sample frames from input video and apply a set of frame-aware visual prompts to the sampled frames in order; 2 Feature extraction: 2D CNN extracts features from sampled video frames with visual prompts, and the language encoder extracts textual features. In addition, the visual features would be spatially downsampled and temporally fused by max pooling and mean pooling, respectively. 3 Multimodal feature processing: after spatial downsampling and temporal fusion, the 2D visual features would be integrated into the prompted textual features. 4 Crossmodal fusion: the multimodal features would be processed by a 12-layer transformer encoder, and MLP would predict the starting/ending time points of the target moment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "content": "Distance Loss " + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "content": ". To avoid the gradient vanishing problem caused by the non-overlapping case, we involve distance loss " + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "content": " to directly minimize the normalized central time point distance. In addition, we add a threshold " + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 361, + 289, + 434 + ], + "type": "text", + "content": " to prevent oscillation in the later training phase. The distance loss is then given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 287, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 287, + 481 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i s}} = \\max \\left(\\frac {\\left| \\left(t _ {\\mathrm {s t a}} + t _ {\\mathrm {e n d}}\\right) / 2 - \\left(\\hat {t} _ {\\mathrm {s t a}} + \\hat {t} _ {\\mathrm {e n d}}\\right) / 2 \\right|}{\\left| \\hat {\\mathbf {T}} \\bigcup \\mathbf {T} \\right|}, \\alpha_ {1}\\right), \\tag {3}", + "image_path": "3ed5122fc9a5d16c3236b0af32148bd9d699f4d92dd3c247e2a625ed10cf2d3b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "text", + "content": "where recall that " + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = (t_{\\mathrm{sta}}, t_{\\mathrm{end}})" + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "text", + "content": " is predicted by the TVG model (1), and we choose " + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "inline_equation", + "content": "\\alpha_1 = 0.2" + }, + { + "bbox": [ + 46, + 487, + 287, + 512 + ], + "type": "text", + "content": " in experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "type": "text", + "content": "Duration Loss " + }, + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dur}}" + }, + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "type": "text", + "content": ". The introduction of distance loss " + }, + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 46, + 512, + 287, + 594 + ], + "type": "text", + "content": " avoids the gradient vanishing problem but only considers the central time point distance. Yet, this may not be precise enough. For example, even if the central time points are completely overlapped, the duration of two video clips may not be identical. Inspired by the above, we propose the duration loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 596, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 596, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 95, + 596, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d u r}} = \\max \\left(\\frac {| \\mathbf {T} - \\hat {\\mathbf {T}} (\\boldsymbol {\\theta}) |}{| \\mathbf {T} |}, \\alpha_ {2}\\right), \\tag {4}", + "image_path": "2e1f6d35b5237079d5c3446fec345a2df73defffd3a7a7508b7d1f70fc158ceb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 46, + 632, + 287, + 655 + ], + "type": "text", + "content": " is the precision tolerance threshold and set by 0.4 in our experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 656, + 287, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 287, + 680 + ], + "type": "text", + "content": "Finally, the proposed Temporal-Distance IoU (TDIoU) loss is given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 683, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 683, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 102, + 683, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {t I o U}} + \\beta_ {1} \\mathcal {L} _ {\\mathrm {d i s}} + \\beta_ {2} \\mathcal {L} _ {\\mathrm {d u r}}, \\tag {5}", + "image_path": "697d0c92d6032fdfa8647453dd210b99d7f1e8574e1c36cc5f4f5fad586a6e6d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1} > 0" + }, + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2} > 0" + }, + { + "bbox": [ + 47, + 701, + 274, + 714 + ], + "type": "text", + "content": " are regularization parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 361, + 455, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 455, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 455, + 374 + ], + "type": "text", + "content": "3.3. Text-Visual Prompt Design" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": "Inspired by the application of prompts on transformers [2, 21, 36, 37], we propose jointly text-visual prompting to boost the performance of our models, in which prompts are optimized perturbation patterns. To improve data processing efficiency, we uniformly sample video frames from the untrimmed video " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}} \\in \\mathbb{R}^{N_{\\mathrm{sam}} \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": " is the number of sampled video frames. In addition, we introduce a set of frame-aware visual prompts " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{vp}} \\in \\mathbb{R}^{N_{\\mathrm{sam}} \\times d_{\\mathrm{vp}}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": " in the pixel space of sampled video frames " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", and introduce text prompts " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{tp}} \\in \\mathbb{R}^{N_{\\mathrm{tp}} \\times d_{\\mathrm{tp}}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": " in the textual feature space. By incorporating video frame sampling and text-visual prompts into the TVG model (1), we obtain:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 544, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\left(\\hat {t} _ {\\mathrm {s t a}}, \\hat {t} _ {\\mathrm {e n d}}\\right) = f \\left(\\delta_ {\\mathrm {t p}}, g _ {\\mathrm {t e x}} (\\mathbf {s}), g _ {\\mathrm {v i d}} \\left(\\mathbf {v} _ {\\mathrm {s a m}} + \\delta_ {\\mathrm {v p}}\\right)\\right). \\tag {6}", + "image_path": "5effab6aadf5cbc4e0f68f28266506342037da8e6e1dac9da72c06dbe5a81d22.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Given a pre-trained 2D TVG model " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": ", the objective of text-visual prompting (TVP) is to learn a universal set of visual prompts " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{vp}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " and text prompts " + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{tp}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": " to be integrated into sampled video frames and textual features, respectively. Specially, a set of different visual prompts are applied to uniformly-sampled frames of one untrimmed video in order. During training, only the set of visual prompts and text prompts are updated through backpropagation. During finetuning, prompts are frozen, and the parameters of the TVG model and encoders are updated. During testing, the set of optimized visual prompts and the optimized text prompts are applied to all test-time video-query pairs." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14797" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 125, + 83 + ], + "type": "text", + "content": "3.4. Framework" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": "Inspired by the success of transformers in vision-language tasks, we choose ClipBERT [31] as the base model for 2D TVG. Extended from ClipBERT, the input of our regression-based TVG model would be describable sentences and uniformly sampled frames of one untrimmed video as shown in Fig. 3. Then, the predicted starting and ending time points of the target video clip would be model outputs. As described in Algorithm 1, there are four phases of our proposed TVP framework: Video frame preprocessing: We obtain sparsely-sampled frames " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " from one input untrimmed video " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": ", and apply universal frame-aware visual prompts " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{vp}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " on top of frames at the padding location. Feature extraction: 2D vision encoder (first 5 ConvBlock of ResNet-50) " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{vid}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " and language encoder (a trainable word embedding layer) " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tex}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " would extract features from the prompted frames " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}}'" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " and textual inputs " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": ", respectively. Multimodal feature processing: Following the setting of Pixel-BERT [22], the 2D visual features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " are downsampled spatially by a " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " max-pooling layer and fused temporally by a mean-pooling layer. Then, text prompts " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{tp}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " are integrated into textual features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": ". In addition, trainable 2D visual position embeddings " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " and textual position embeddings " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " are applied to the processed 2D visual features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}}'" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " and prompted textual features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}'" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": ", respectively [10, 31]. Afterwards, the processed and position-encoded 2D visual features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}}''" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " are flattened and integrated into prompted and position-encoded textual features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}''" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": ". Moreover, type embeddings " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{type}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " would be added to the integrated multimodal features " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{all}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " to indicate the source type of features. Crossmodal fusion: A 12-layer transformer [10] is utilized for crossmodal fusion on " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{all}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": ", and then multilayer perceptron (MLP) ending with sigmoid function is used as the prediction head to process the last-layer crossmodal representation " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{CM}}" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " of the transformer for generating the predicted starting/ending time points " + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "inline_equation", + "content": "(\\hat{t}_{\\mathrm{sta}}, \\hat{t}_{\\mathrm{sta}})" + }, + { + "bbox": [ + 47, + 91, + 289, + 534 + ], + "type": "text", + "content": " of the target moments described by the text query input." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 545, + 128, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 128, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 128, + 559 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 566, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 603 + ], + "type": "text", + "content": "In this section, we demonstrate the effectiveness of our proposed TVP framework on Charades-STA and ActivityNet Captions datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 155, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 155, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 155, + 624 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "Datasets. The evaluations are implemented on two standard benchmark datasets for TVG task, Charades-STA [14] and ActivityNet Captions [28]. Tab. 1 summarizes the details of both datasets. Charades-STA dataset contains 6,672 videos and 16,124 text queries in total. The average length of videos is " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "30.6s" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ", and the average length of text query is 7.2 words. The average length of moments cor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 481, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 481, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 481, + 84 + ], + "type": "text", + "content": "Algorithm 1 Overview of TVP framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 87, + 545, + 160 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": "Input: vision encoder " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{vid}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", language encoder " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{tex}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", position embeddings " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", 2D position embeddings " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{2D}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", type embeddings " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{type}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", transformer " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", prediction head " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "MLP" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", visual prompts " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{vp}}" + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "text", + "content": ", text prompts " + }, + { + "bbox": [ + 306, + 87, + 545, + 137 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{tp}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 137, + 505, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 137, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 505, + 148 + ], + "type": "text", + "content": "Output: Predicted time interval " + }, + { + "bbox": [ + 306, + 137, + 505, + 148 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}} = (\\hat{t}_{\\mathrm{sta}},\\hat{t}_{\\mathrm{end}})" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 149, + 501, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 149, + 501, + 160 + ], + "spans": [ + { + "bbox": [ + 344, + 149, + 501, + 160 + ], + "type": "text", + "content": "Phase ①: Video frame preprocessing" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 161, + 545, + 208 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 312, + 161, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 161, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 161, + 545, + 183 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 312, + 161, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}} \\gets" + }, + { + "bbox": [ + 312, + 161, + 545, + 183 + ], + "type": "text", + "content": " uniformly sample video frames from an untrimmed video " + }, + { + "bbox": [ + 312, + 161, + 545, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}}^{\\prime}\\gets" + }, + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "text", + "content": " apply visual prompts " + }, + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{vp}}" + }, + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "text", + "content": " to the sampled video frames " + }, + { + "bbox": [ + 312, + 184, + 545, + 208 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sam}}" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 343, + 209, + 468, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 209, + 468, + 219 + ], + "spans": [ + { + "bbox": [ + 343, + 209, + 468, + 219 + ], + "type": "text", + "content": "Phase ②: Feature Extraction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 220, + 533, + 243 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 312, + 220, + 533, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 533, + 232 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 312, + 220, + 533, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}} = g_{\\mathrm{vid}}(\\mathbf{v}_{\\mathrm{sam}}^{\\prime})\\gets" + }, + { + "bbox": [ + 312, + 220, + 533, + 232 + ], + "type": "text", + "content": " extracting 2D visual features" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 233, + 506, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 233, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 233, + 506, + 243 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 233, + 506, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}} = g_{\\mathrm{tex}}(\\mathbf{s})" + }, + { + "bbox": [ + 312, + 233, + 506, + 243 + ], + "type": "text", + "content": " ← extracting textual features" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 343, + 245, + 518, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 245, + 518, + 255 + ], + "spans": [ + { + "bbox": [ + 343, + 245, + 518, + 255 + ], + "type": "text", + "content": "Phase 8: Multimodal feature processing" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 309, + 256, + 545, + 399 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 312, + 256, + 545, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 256, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 545, + 279 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 256, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vld}}^{\\prime}\\gets" + }, + { + "bbox": [ + 312, + 256, + 545, + 279 + ], + "type": "text", + "content": " apply spatial downsampling and temporal fusion to 2D visual features " + }, + { + "bbox": [ + 312, + 256, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vld}}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}^{\\prime}\\gets" + }, + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "text", + "content": " apply text prompts " + }, + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{tp}}" + }, + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "text", + "content": " to textual features " + }, + { + "bbox": [ + 312, + 280, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}}^{\\prime \\prime} \\gets" + }, + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "text", + "content": " add 2D visual position embeddings " + }, + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "text", + "content": " on the processed 2D visual features " + }, + { + "bbox": [ + 312, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}}^{\\prime}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}^{\\prime \\prime}\\gets" + }, + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "text", + "content": " add position embeddings " + }, + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "text", + "content": " to prompted textual features " + }, + { + "bbox": [ + 312, + 316, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}^{\\prime}" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{all}} \\gets" + }, + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "text", + "content": " integrate the processed and position-encoded textual features " + }, + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{tex}}^{\\prime \\prime}" + }, + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "text", + "content": " and the processed and position-encoded 2D visual features " + }, + { + "bbox": [ + 312, + 340, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{vid}}^{\\prime \\prime}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{all}} + \\mathbf{M}_{\\mathrm{type}} \\gets" + }, + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "text", + "content": " add type embeddings " + }, + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{type}}" + }, + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "text", + "content": " to the integrated multimodal features " + }, + { + "bbox": [ + 309, + 376, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{all}}" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 343, + 400, + 467, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 400, + 467, + 410 + ], + "spans": [ + { + "bbox": [ + 343, + 400, + 467, + 410 + ], + "type": "text", + "content": "Phase 4: Crossmodal fusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 309, + 411, + 545, + 472 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 309, + 411, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 411, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 309, + 411, + 545, + 435 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 309, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{CM}} = f(\\mathbf{Q}_{\\mathrm{all}} + \\mathbf{M}_{\\mathrm{type}}) \\gets" + }, + { + "bbox": [ + 309, + 411, + 545, + 435 + ], + "type": "text", + "content": " implement crossmodal fusion through transformer " + }, + { + "bbox": [ + 309, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "f" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 309, + 435, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 435, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 309, + 435, + 545, + 472 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 309, + 435, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\hat{t}_{\\mathrm{sta}},\\hat{t}_{\\mathrm{end}}) = MLP(\\mathbf{Q}_{\\mathrm{CM}})\\gets" + }, + { + "bbox": [ + 309, + 435, + 545, + 472 + ], + "type": "text", + "content": " prediction head generates the predicted time interval according to crossmodal representation " + }, + { + "bbox": [ + 309, + 435, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{CM}}" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + }, + { + "type": "table", + "bbox": [ + 309, + 516, + 543, + 618 + ], + "blocks": [ + { + "bbox": [ + 306, + 488, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 306, + 488, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 545, + 510 + ], + "type": "text", + "content": "Table 1. Statistics of TVG benchmark datasets (Charades-STA and ActivityNet Captions datasets)." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 516, + 543, + 618 + ], + "lines": [ + { + "bbox": [ + 309, + 516, + 543, + 618 + ], + "spans": [ + { + "bbox": [ + 309, + 516, + 543, + 618 + ], + "type": "table", + "html": "
DatasetCharades-STAActivityNet Captions
DomainIndoor ActivityIndoor/Outdoor Activity
# Videos6,67214,926
Avg. Video Length (second)30.6117.6
# Moments11,76771,953
Avg. Moment Length (second)8.137.1
Vocabulary Size1,30315,505
# Queries16,12471,953
Avg. Query Length (word)7.214.4
", + "image_path": "07cc29e142346abe17768bdda32c71a64dc5e919e7c924d8d80761f55465582d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "responding to the text query is 8.1s. Following the same dataset split as [14] for fair comparisons, there are 12,408 video-query pairs for training and 3,720 pairs for testing. ActivityNet Captions dataset contains 14,926 videos and 71,953 text queries in total. The average length of videos is 117.6s, and the average length of text query is 14.4 words." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14798" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "The average length of moments corresponding to the text query is 37.1s. ActivityNet Captions dataset is split into training set, validation set, and testing set in a " + }, + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "inline_equation", + "content": "2:1:1" + }, + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": " ratio. Since the testing set is withheld for competition, only a training set and two validation sets (val1 and val2) can be accessed publicly. For fair comparisons, we evaluate our proposed framework on val1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 230 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 230 + ], + "type": "text", + "content": "Baselines. We compare our proposal with 15 baseline methods: ① Proposal-based: CTRL [14], MCN [1], SAP [7], BPNet [62], LPNet [61], QSPN [64], MAN [71]; ② Proposal-free: ABLR [67], DRN [69], CPNet [34], DEBUG [43], ExCL [16], VSLNet [73]; ③ Reinforcement learning: TSP-PRL [60], TripNet [18]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "type": "text", + "content": "Evaluation metrics. Following [14], we adopt Acc(R@1, IoU=m) as the performance evaluation metric, which represents the percentage accuracy of top-1 predicted moments whose tIoU (temporal IoU) with the ground-truth moment is larger than " + }, + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "type": "text", + "content": ". By convention, we consider the following tIoU threshold values " + }, + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "type": "inline_equation", + "content": "m = \\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 46, + 231, + 288, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 288, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 486 + ], + "type": "text", + "content": "Crossmodal pretraining setup. Our 2D vision encoder (ResNet-50) is initialized with the weight from gridfeat [24], which can extract effective grid features from visual inputs. In addition, both the language encoder and 12-layer transformer are initialized with the BERT-base model weight [10], which are pretrained on English Wikipedia and BookCorpus [83]. Thanks to the compact 2D vision encoder, TVP (our proposal) is able to directly utilize image-text pairs for end-to-end training. Since the benefits of cross-modal pretraining has been demonstrated by [22, 44, 55], our base model is pretrained on two large-scale image-text datasets, which are Visual Genome Captions [29] and COCO Captions [8]. To be more specific, image-text matching [44, 55] and masked language modeling [10] are employed for cross-modal pretraining." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Implementation setup. For video inputs, we uniformly sample " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}}" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " frames from a video (" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}} = 48" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " for Charades-STA and " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}} = 64" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " for ActivityNet Captions). In addition, all video frames are resized to have a maximum longer side of 448 with an original aspect ratio, and then the frames are zero-padded to " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ". The default visual prompt sizes for both dataset are 96. The default text prompt sizes are 10 and 20 for Charades-STA and ActivityNet Captions, respectively. We utilize the first 5 ConvBlocks of ResNet-50 as the 2D vision encoder and a trainable embedding layer as the language encoder for both Charades-STA and ActivityNet Captions datasets. For text queries, all word tokens are maintained after lower-case conversion and tokenization. We use AdamW [42] for end-to-end model training, with " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 1.0" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.1" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 0.2" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 0.4" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ". Initial learning rates are " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "1e - 1" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "5e - 7" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " for prompt training and model finetuning, respectively. In addition, the learning rate linearly decays to 0 with the first " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " training step for warmup. Our experiments are imple" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 98, + 544, + 284 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 90 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 90 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 90 + ], + "type": "text", + "content": "Table 2. Performance comparison of different thresholds " + }, + { + "bbox": [ + 305, + 70, + 547, + 90 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 70, + 547, + 90 + ], + "type": "text", + "content": " on the Charades-STA dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 98, + 544, + 284 + ], + "lines": [ + { + "bbox": [ + 308, + 98, + 544, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 544, + 284 + ], + "type": "table", + "html": "
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D-23.638.89
ABLR [67]C3D-24.369.01
BPNet [62]C3D55.4638.2520.51
LPNet [61]C3D59.1440.9421.13
QSPN [64]C3D54.7035.6015.80
TSP-PRL [60]C3D-45.4524.75
TripNet [18]C3D54.6438.2916.07
DRN [69]C3D-45.4026.40
CPNet [34]C3D-40.3222.47
DEBUG [43]C3D54.9537.3917.92
ExCL [16]I3D61.5044.122.40
VSLNet [73]I3D64.3047.3130.19
MAN [71]I3D-46.5322.72
2D TVGMCN [1]VGG-17.468.01
SAP [7]VGG-27.4213.36
Ours
TVP-Based 2D TVGBase w/o promptsResNet61.2940.4319.89
Base + Visual Prompts65.3844.3120.22
Base + Text Prompts65.8143.4420.65
Base + Both Prompts65.9244.3921.51
", + "image_path": "89b08d07523e5aae6f26bc9866e7209574c377aaad4c68bf1c22a142c286abeb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 321, + 544, + 481 + ], + "blocks": [ + { + "bbox": [ + 305, + 293, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 314 + ], + "type": "text", + "content": "Table 3. Performance comparison of different thresholds " + }, + { + "bbox": [ + 305, + 293, + 545, + 314 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 293, + 545, + 314 + ], + "type": "text", + "content": " on the ActivityNet Captions dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 321, + 544, + 481 + ], + "lines": [ + { + "bbox": [ + 308, + 321, + 544, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 544, + 481 + ], + "type": "table", + "html": "
TypeMethodVisual FeatureAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
3D TVGCTRL [14]C3D28.7014.00-
BPNet [62]C3D59.9842.0724.69
LPNet [61]C3D64.2945.9225.39
QSPN [64]C3D45.3027.7013.60
TSP-PRL [60]C3D56.0238.83-
TripNet [18]C3D48.4232.1913.93
DRN [69]C3D-45.4524.36
CPNet [34]C3D-40.5621.63
ABLR [67]C3D55.6736.79-
DEBUG [43]C3D55.9139.72-
ExCL [16]C3D63.0043.6024.10
VSLNet [73]C3D63.1643.2226.16
Ours
TVP-Based 2D TVGBase w/o promptsResNet57.2040.1619.14
Base + Visual Prompts60.1243.3923.71
Base + Text Prompts60.4842.5824.39
Base + Both Prompts60.7143.4425.03
", + "image_path": "06bd71bcaaea49013f64961dd0a2fc0f3baae18be52ed233b39c575dcbd73f04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 501, + 547, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 547, + 537 + ], + "type": "text", + "content": "mented in PyTorch [45], and models and prompts are finetuned separately for 12 epochs with the mixed precision on 8 NVIDIA V100 GPUs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 544, + 422, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 422, + 558 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 422, + 558 + ], + "type": "text", + "content": "4.2. Experiment Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Effectiveness of TVP on Charades-STA. The performance comparisons with SOTA methods on the Charades-STA dataset are summarized in Tab. 2. Our proposed TVP framework can achieve competitive performance at all IoU thresholds " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " in the case of utilizing 2D visual features extracted by ResNet-50, and reach the highest score at " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "m = 0.3" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ". Compared to the 2D TVG methods using VGG as the vision encoder, our proposed framework could achieve around " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.5 \\times" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.7 \\times" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " performance gain at thresholds 0.5 and 0.7, respectively. Furthermore, we can find that for our base model only one of visual prompts and text prompts can achieve up to " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "7.37\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "9.60\\%" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " improvement" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14799" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 167, + 212 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 167, + 212 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 167, + 212 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 167, + 212 + ], + "type": "image", + "image_path": "f930cf36323498ab1ee69cd319022a82926fb4450ac7405b8a56d5e204390eb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 214, + 138, + 224 + ], + "lines": [ + { + "bbox": [ + 80, + 214, + 138, + 224 + ], + "spans": [ + { + "bbox": [ + 80, + 214, + 138, + 224 + ], + "type": "text", + "content": "(a) Charades-STA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 71, + 284, + 213 + ], + "blocks": [ + { + "bbox": [ + 167, + 71, + 284, + 213 + ], + "lines": [ + { + "bbox": [ + 167, + 71, + 284, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 71, + 284, + 213 + ], + "type": "image", + "image_path": "7f9010f95a5a2247a77be25bc98fd2db502eaf9d81d99450cd3a28ae2a91b6da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 214, + 266, + 224 + ], + "lines": [ + { + "bbox": [ + 185, + 214, + 266, + 224 + ], + "spans": [ + { + "bbox": [ + 185, + 214, + 266, + 224 + ], + "type": "text", + "content": "(b) ActivityNet Captions" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 233, + 240, + 243 + ], + "lines": [ + { + "bbox": [ + 93, + 233, + 240, + 243 + ], + "spans": [ + { + "bbox": [ + 93, + 233, + 240, + 243 + ], + "type": "text", + "content": "Figure 4. Impact of sampled frame numbers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": "at tIoU thresholds " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "m = 0.3" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": ". The combination of text and visual prompts can not only achieve " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "7.55\\%" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "9.79\\%" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": " improvements at tIoU thresholds " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "m = 0.3" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": ", but also improve the performance by " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "8.14\\%" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "inline_equation", + "content": "m = 0.7" + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": ". This demonstrates the effectiveness and necessity of the joint text-visual prompting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": "Effectiveness of TVP on ActivityNet Captions. We focus on the performance comparisons with 3D TVG methods on ActivityNet since there are no results of 2D TVG method reported on ActivityNet Captions. The results of multiple methods on ActivityNet Captions datasets are reported in Tab. 3. Even on this more challenging dataset, our proposed method still has achieved competitive performance compared to 3D TVG methods. Different from the performance of TVP on Charades-STA dataset, text prompts or visual prompts can achieve a significant performance boost on the base model over all IoU thresholds " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": " alone (5.73% at " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.3" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": ", 8.04% at " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": ", 27.43% at " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.7" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": "), and the text-visual prompt combination could further boost the performance (6.14% at " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.3" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": ", 8.17% at " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": ", 30.77% at " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.7" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": "). It is worth noting that the performance gap over " + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "inline_equation", + "content": "m = 0.7" + }, + { + "bbox": [ + 46, + 328, + 287, + 532 + ], + "type": "text", + "content": " between 2D TVG methods and 3D TVG methods is narrowed significantly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "content": "In summary, through the experimental results on Charades-STA and ActivityNet Captions datasets, we can find that our proposed TVP framework could achieve competitive performance overall tIoU thresholds on Charades-STA and ActivityNet Captions by improving the utility of sparse 2D visual features. Thanks to the lightweight 2D vision encoder, cotraining language encoder and vision encoder on large-scale image-text datasets can be performed, which benefits the base model to achieve good performance. Furthermore, the combination of text and visual prompts can achieve better results than any single kind of prompts on both datasets, which again proves the importance of cross-modal training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Video frame sampling effect. Fig. 4 demonstrates the performance of base model with different number " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}}" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 98, + 543, + 244 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "type": "text", + "content": "Table 4. The performance comparison of different visual prompt sizes on Charades-STA dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 98, + 543, + 244 + ], + "lines": [ + { + "bbox": [ + 309, + 98, + 543, + 244 + ], + "spans": [ + { + "bbox": [ + 309, + 98, + 543, + 244 + ], + "type": "table", + "html": "
Visual Prompt SizeAcc(R@1, IoU=m)Prompt + Frame
m=0.3m=0.5m=0.7
061.2940.4319.89
1661.2940.4320.00
3261.9439.7819.35
4863.6642.3720.00
7263.8743.6619.78
9665.3844.3120.22
12864.7343.6619.78
", + "image_path": "1097b193d2d816f4610a418a81bff2571d8a2b54ea3cddcf4e10c0c45f2faa75.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 314, + 280, + 538, + 396 + ], + "blocks": [ + { + "bbox": [ + 305, + 257, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 277 + ], + "type": "text", + "content": "Table 5. The performance comparison of different text prompt sizes on Charades-STA dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 280, + 538, + 396 + ], + "lines": [ + { + "bbox": [ + 314, + 280, + 538, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 280, + 538, + 396 + ], + "type": "table", + "html": "
Text Prompt SizeAcc(R@1, IoU=m)
m=0.3m=0.5m=0.7
057.2040.1619.14
565.3841.9420.43
1065.8143.4420.65
1565.5943.2321.29
2064.9543.8721.51
2563.6642.8020.65
3064.4642.6320.51
", + "image_path": "97c465ff5b852083d1feaf2013a797705d1dd8054e12e453b90a6a9223322e53.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 410, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 483 + ], + "type": "text", + "content": "of sampled video frames as visual inputs. For Charades dataset, the base model performance keeps increasing before " + }, + { + "bbox": [ + 304, + 410, + 545, + 483 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}}" + }, + { + "bbox": [ + 304, + 410, + 545, + 483 + ], + "type": "text", + "content": " reaches 48, but when it exceeds 48, performance starts to degrade. This is because frequent background changes harm the performance of object re-identification in videos, which are noisy for object motion analysis [17]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": "For ActivityNet Caption dataset, base model performance continues to improve even when sampled frame number " + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}}" + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": " exceeds 48, due to the longer average video length in ActivityNet Captions dataset. Balancing the frame number and batch size for training, we choose " + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sam}} = 64" + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": " for ActivityNet Captions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 556, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 688 + ], + "type": "text", + "content": "TVP performance vs. prompt size. As shown in Tab. 4, we can find that when visual prompts are small, they cannot bring changes to the base model, and when visual prompts are too large, the performance starts to decrease. This is because key information within video frames might be removed. However, the text prompts can bring significant performance boost even when the text prompt size is small as shown in Tab. 5, which is because the textual features has a smaller dimension compared to visual features, and also the text prompts are directly optimized in feature space during training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "TVP performance vs. visual prompt operation. Visual prompt is first proposed by [2], where visual prompts are" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14800" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 107, + 284, + 182 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 99 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 99 + ], + "type": "text", + "content": "Table 6. The performance comparison of different visual prompt operations ('remove', 'add', 'replace') with fixed visual prompt size " + }, + { + "bbox": [ + 47, + 70, + 287, + 99 + ], + "type": "inline_equation", + "content": "p = 96" + }, + { + "bbox": [ + 47, + 70, + 287, + 99 + ], + "type": "text", + "content": " on Charades-STA and ActivityNet Captions datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 107, + 284, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 107, + 284, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 107, + 284, + 182 + ], + "type": "table", + "html": "
OperationCharades-STAActivityNet Captions
R@1, IoU=mR@1, IoU=m
m=0.3m=0.5m=0.7m=0.3m=0.5m=0.7
Original61.2940.4319.8957.2040.1619.14
Remove61.2940.4320.057.2040.1619.14
Add61.0839.5720.2257.1540.1619.27
Replace65.3844.3120.2260.1243.3923.71
", + "image_path": "97f634bbd4d952f565a1efa47b8cb8cb54418863533f55c20640de4ff5bc7b3f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 196, + 164, + 335 + ], + "blocks": [ + { + "bbox": [ + 50, + 196, + 164, + 335 + ], + "lines": [ + { + "bbox": [ + 50, + 196, + 164, + 335 + ], + "spans": [ + { + "bbox": [ + 50, + 196, + 164, + 335 + ], + "type": "image", + "image_path": "a82c52377ad36275e67254268d4a39d76b11b976001d259d05775ede79124c0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 338, + 145, + 347 + ], + "lines": [ + { + "bbox": [ + 88, + 338, + 145, + 347 + ], + "spans": [ + { + "bbox": [ + 88, + 338, + 145, + 347 + ], + "type": "text", + "content": "(a) Vison Encoders" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 356, + 287, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 404 + ], + "type": "text", + "content": "Figure 5. Inference time comparison. (a) inference time comparison between 2D vision encoder (ResNet-50) and 3D vision encoder (C3D). (b) inference time comparison between the vision encoder and the other modules of the 2D TVG model, where the sampled frame number for our TVP framework is " + }, + { + "bbox": [ + 46, + 356, + 287, + 404 + ], + "type": "inline_equation", + "content": "1.2 \\times" + }, + { + "bbox": [ + 46, + 356, + 287, + 404 + ], + "type": "text", + "content": " the length of the video in seconds." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 196, + 282, + 334 + ], + "blocks": [ + { + "bbox": [ + 165, + 196, + 282, + 334 + ], + "lines": [ + { + "bbox": [ + 165, + 196, + 282, + 334 + ], + "spans": [ + { + "bbox": [ + 165, + 196, + 282, + 334 + ], + "type": "image", + "image_path": "3623d7cc6fbc44db6248b6619f8e1b6fcc50e40ae5a1a400036e943477297a12.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 338, + 282, + 346 + ], + "lines": [ + { + "bbox": [ + 190, + 338, + 282, + 346 + ], + "spans": [ + { + "bbox": [ + 190, + 338, + 282, + 346 + ], + "type": "text", + "content": "(b) Modules in 2D TVG Model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 593 + ], + "type": "text", + "content": "added to the image for transfer learning on classification tasks. In contrast, our proposed prompting framework is designed to compensate for the spatiotemporal information loss in 2D visual features. Due to the differences in the task, we try two different prompt operation strategies, 'replace' and 'add'. 'add' is to add the visual prompts to the pixel value of the video frame at the corresponding padding locations. 'replace' is to replace the pixel values of video frames with visual prompts at corresponding padding locations. 'remove' is in order to study the impact of removing the pixel values at the padding location. As shown in Tab. 6, 'add' or 'remove' prompt operations have limited effects on the base model. However, 'replace' does boost the base model performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "TVP achieves inference efficiency. As shown in Fig. 5, we can find that the inference time required for visual feature extraction accounts for more than half of the inference time of the whole model, while the inference time required for the 3D vision encoder is more than " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " compared to the 2D vision encoder, and even more than the time required for the whole TVG model using 2D vision encoder, which fully demonstrates the feasibility of accelerating the overall inference speed by reducing the complexity of the vision encoder. Need to note that if there are" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 98, + 531, + 175 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 89 + ], + "type": "text", + "content": "Table 7. The performance comparison of different loss designs on Charades-STA dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 98, + 531, + 175 + ], + "lines": [ + { + "bbox": [ + 321, + 98, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 321, + 98, + 531, + 175 + ], + "type": "table", + "html": "
Loss Function SelectionR@1, IoU=m
m=0.3m=0.5m=0.7
LtIoU55.0529.8911.82
LtIoU + Ldis60.6431.1816.77
LtIoU + Ldur59.7830.9716.34
LtIoU + Ldis + Ldur61.2940.4319.89
", + "image_path": "815dfebc92ff3a1e6aeadd548cf3008821ddf94512aa7f7398ed1a4898f72db4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 180, + 420, + 259 + ], + "blocks": [ + { + "bbox": [ + 310, + 180, + 420, + 259 + ], + "lines": [ + { + "bbox": [ + 310, + 180, + 420, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 180, + 420, + 259 + ], + "type": "image", + "image_path": "c51b7edc18b9108057885efbb83f8ceb6497e9223bbd4bf91132b1a0c4bdcb4d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 261, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 281 + ], + "type": "text", + "content": "Figure 6. Loss landscape visualization in 2D plane: Finetuning w/o prompts (left) and using prompts (right); see [33] for implementation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 432, + 181, + 542, + 259 + ], + "blocks": [ + { + "bbox": [ + 432, + 181, + 542, + 259 + ], + "lines": [ + { + "bbox": [ + 432, + 181, + 542, + 259 + ], + "spans": [ + { + "bbox": [ + 432, + 181, + 542, + 259 + ], + "type": "image", + "image_path": "c473b2558978f69d281dc660c61bf9dac140ef26ca04348f7cc3e71d31bac6cd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 287, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 347 + ], + "type": "text", + "content": "multiple model weights for different sampled frame number settings and model weights can be adopted adaptively for different lengths of videos, the inference speed for short videos should increase, and the prediction results for long videos will be further improved." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": "Ablation studies. Through Tab. 7, we can find that the addition of either distance loss " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": " or duration loss " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dur}}" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": " will result in a performance increase, but the combination of the two will result in a significant performance increase (11.34% at " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "m = 0.3" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": ", 35.26% at " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": ", 68.27% at " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "m = 0.7" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": "), especially over tIoU thresholds " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "m = 0.5" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "m = 0.7" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": ". This demonstrates that distance loss " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": " and duration loss " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dur}}" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": " could provide more precise training guides compared to only using temporal IoU loss " + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tIoU}}" + }, + { + "bbox": [ + 304, + 347, + 546, + 515 + ], + "type": "text", + "content": ". Furthermore, we posit that prompting may encode additional spatial-temporal supervision to help the model trainer to escape from bad local optima as shown in Fig. 6, where finetuning w/ prompts yields a flatter loss landscape than the one w/o prompts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose text-visual prompting to boost the performance of 2D TVG methods by compensating for the lack of spatiotemporal information in 2D visual features. In contrast to 3D TVG methods, TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross-modal feature fusion using only low-complexity sparse 2D visual features. The effectiveness of our proposed TVP (text-visual prompting) framework has been demonstrated on two standard datasets, Charades-STA and ActivityNet. Our models outperform all 2D models significantly, and also achieve comparable performance to 3D models. What is more, we achieve over " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " inference speedup over TVG methods of using 3D visual features." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14801" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 191 + ], + "type": "text", + "content": "[2] H Bahng, A Jahanian, S Sankaranarayanan, and P Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, page 2022, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 270 + ], + "type": "text", + "content": "[4] Aochuan Chen, Peter Lorenz, Yuguang Yao, Pin-Yu Chen, and Sijia Liu. Visual prompting for adversarial robustness. arXiv preprint arXiv:2210.06284, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "type": "text", + "content": "[5] Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. arXiv preprint arXiv:2211.11635, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 349 + ], + "type": "text", + "content": "[6] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. arXiv:2202.10629, 2022.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 396 + ], + "type": "text", + "content": "[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8199-8206, 2019. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[8] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 496 + ], + "type": "text", + "content": "[9] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 542 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "text", + "content": "[11] Gamaleldin F Elsayed, Ian Goodfellow, et al. Adversarial reprogramming of neural networks. arXiv:1806.11146, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[12] Christoph Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Yanjun Gao, Lulu Liu, Jason Wang, Xin Chen, Huayan Wang, and Rui Zhang. Evoquer: Enhancing temporal grounding with video-pivoted backquery generation. arXiv preprint arXiv:2109.04600, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 161 + ], + "type": "text", + "content": "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[17] Xinqian Gu, Hong Chang, Bingpeng Ma, Hongkai Zhang, and Xilin Chen. Appearance-preserving 3d convolution for video-based person re-identification. In ECCV, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "text", + "content": "[18] Meera Hahn, Asim Kadav, James M Rehg, and Hans Peter Graf. Tripping through time: Efficient localization of activities in videos. arXiv preprint arXiv:1904.09936, 2019. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "type": "text", + "content": "[19] Xiaochen Han, Bo Wu, Zheng Shou, Xiao-Yang Liu, Yi-meng Zhang, and Linghe Kong. Tensor fista-net for real-time snapshot compressive imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10933-10940, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 341 + ], + "type": "text", + "content": "[20] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. arXiv preprint arXiv:1809.01337, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "text", + "content": "[21] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[22] Zhicheng Huang, Zhaoyang Zeng, Bei Liu, Dongmei Fu, and Jianlong Fu. Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. arXiv preprint arXiv:2004.00849, 2020. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 488 + ], + "type": "text", + "content": "[23] Jinghan Jia, Mingyi Hong, Yimeng Zhang, Mehmet Akçakaya, and Sijia Liu. On the robustness of deep learning-based mri reconstruction to image transformations. arXiv preprint arXiv:2211.04930, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 543 + ], + "type": "text", + "content": "[24] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik Learned-Miller, and Xinlei Chen. In defense of grid features for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10267-10276, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 588 + ], + "type": "text", + "content": "[25] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "text", + "content": "[26] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[27] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. arXiv preprint arXiv:2210.03117, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14802" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 184 + ], + "type": "text", + "content": "[30] Xiaohan Lan, Yitian Yuan, Xin Wang, Zhi Wang, and Wenwu Zhu. A survey on temporal sentence grounding in videos. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 274 + ], + "type": "text", + "content": "[32] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 308 + ], + "type": "text", + "content": "[33] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 353 + ], + "type": "text", + "content": "[34] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1902-1910, 2021. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 397 + ], + "type": "text", + "content": "[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 288, + 453 + ], + "type": "text", + "content": "[36] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 288, + 487 + ], + "type": "text", + "content": "[37] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 488, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 533 + ], + "type": "text", + "content": "[38] Yanyu Li, Pu Zhao, Geng Yuan, Xue Lin, Yanzhi Wang, and Xin Chen. Pruning-as-search: Efficient neural architecture search via channel pruning and structural reparameterization. arXiv preprint arXiv:2206.01198, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 578 + ], + "type": "text", + "content": "[39] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 611 + ], + "type": "text", + "content": "[40] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 612, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 288, + 667 + ], + "type": "text", + "content": "[41] Xiao-Yang Liu, Yimeng Zhang, Yukang Liao, and Ling Jiang. Dynamic updating of the knowledge base for a large-scale question answering system. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 691 + ], + "type": "text", + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 127 + ], + "type": "text", + "content": "language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 128, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 546, + 172 + ], + "type": "text", + "content": "[44] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 172, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 228 + ], + "type": "text", + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 546, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 271 + ], + "type": "text", + "content": "[46] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 271, + 546, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 546, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 546, + 315 + ], + "type": "text", + "content": "[47] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 546, + 360 + ], + "type": "text", + "content": "[48] Zhaofan Qiu, Ting Yao, and Tao Mei. Learning spatiotemporal representation with pseudo-3d residual networks. In proceedings of the IEEE International Conference on Computer Vision, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 360, + 546, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 546, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 546, + 393 + ], + "type": "text", + "content": "[49] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 393, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 546, + 426 + ], + "type": "text", + "content": "[50] Timo Schick and Hinrich Schütze. Few-shot text generation with pattern-exploiting training. arXiv preprint arXiv:2012.11926, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 426, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 459 + ], + "type": "text", + "content": "[51] Timo Schick and Hinrich Schütze. It's not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 459, + 546, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 546, + 504 + ], + "type": "text", + "content": "[52] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 504, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 546, + 536 + ], + "type": "text", + "content": "[53] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 2014. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 537, + 546, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 581 + ], + "type": "text", + "content": "[54] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019.2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 581, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 546, + 613 + ], + "type": "text", + "content": "[55] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 614, + 546, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 668 + ], + "type": "text", + "content": "[56] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "text", + "content": "[57] Yun-Yun Tsai et al. Transfer learning without knowing: Reprogramming black-box machine learning models with scarce data and limited resources. arXiv:2007.08714, 2020. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14803" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[58] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[60] Jie Wu, Guanbin Li, Si Liu, and Liang Lin. Tree-structured policy based progressive reinforcement learning for temporally language grounding in video. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12386–12393, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[61] Shaoning Xiao, Long Chen, Jian Shao, Yueting Zhuang, and Jun Xiao. Natural language video localization with learnable moment proposals. arXiv preprint arXiv:2109.10678, 2021. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[62] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2986-2994, 2021. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[63] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In Proceedings of the European conference on computer vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "text", + "content": "[64] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9062–9069, 2019. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "text", + "content": "[65] Chao-Han Huck Yang, Yun-Yun Tsai, et al. Voice2series: Reprogramming acoustic models for time series classification. In ICML. PMLR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[66] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[67] Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159-9166, 2019. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "text", + "content": "[68] Ofir Zafrir, Ariel Larey, Guy Boudoukh, Haihao Shen, and Moshe Wasserblat. Prune once for all: Sparse pre-trained language models. arXiv preprint arXiv:2111.05754, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[69] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[70] Yawen Zeng, Da Cao, Xiaochi Wei, Meng Liu, Zhou Zhao, and Zheng Qin. Multi-modal relational graph for cross-modal video moment retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2215-2224, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 710 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[71] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 139, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 545, + 171 + ], + "type": "text", + "content": "[72] Guanhua Zhang, Yihua Zhang, Yang Zhang, Wenqi Fan, Qing Li, Sijia Liu, and Shiyu Chang. Fairness reprogramming. arXiv preprint arXiv:2209.10222, 2022.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "text", + "content": "[73] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 205, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 247 + ], + "type": "text", + "content": "[74] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. The elements of temporal sentence grounding in videos: A survey and future directions. arXiv preprint arXiv:2201.08071, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 249, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 249, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 545, + 303 + ], + "type": "text", + "content": "[75] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12870-12877, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 304, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 369 + ], + "type": "text", + "content": "[76] Yimeng Zhang, Akshay Karkal Kamath, Qiucheng Wu, Zhiwen Fan, Wuyang Chen, Zhangyang Wang, Shiyu Chang, Sijia Liu, and Cong Hao. Data-model-circuit tri-design for ultra-light video intelligence on edge devices. In Proceedings of the 28th Asia and South Pacific Design Automation Conference, pages 745-750, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 370, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 370, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 370, + 545, + 413 + ], + "type": "text", + "content": "[77] Yimeng Zhang, Xiao-Yang Liu, Bo Wu, and Anwar Walid. Video synthesis via transform-based tensor neural network. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2454-2462, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 414, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 414, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 414, + 545, + 457 + ], + "type": "text", + "content": "[78] Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jinfeng Yi, Mingyi Hong, Shiyu Chang, and Sijia Liu. How to robustify black-box ml models? a zeroth-order optimization perspective. arXiv preprint arXiv:2203.14195, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 458, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 458, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 308, + 458, + 545, + 501 + ], + "type": "text", + "content": "[79] Yihua Zhang, Yuguang Yao, Parikshit Ram, Pu Zhao, Tianlong Chen, Mingyi Hong, Yanzhi Wang, and Sijia Liu. Advancing model pruning via bi-level optimization. arXiv preprint arXiv:2210.04092, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 555 + ], + "type": "text", + "content": "[80] Yihua Zhang, Guanhua Zhang, Prashant Khanduri, Mingyi Hong, Shiyu Chang, and Sijia Liu. Revisiting and advancing fast adversarial training through the lens of bi-level optimization. In International Conference on Machine Learning, pages 26693-26712. PMLR, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 589 + ], + "type": "text", + "content": "[81] Yang Zheng, Xiaoyi Feng, et al. Why adversarial reprogramming works, when it fails, and how to tell the difference. arXiv:2108.11673, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 643 + ], + "type": "text", + "content": "[82] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 644, + 545, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 644, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 308, + 644, + 545, + 710 + ], + "type": "text", + "content": "[83] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19-27, 2015. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14804" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_content_list.json b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd17e476995ef3891ca709580a2306821e246b9d --- /dev/null +++ b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_content_list.json @@ -0,0 +1,1330 @@ +[ + { + "type": "text", + "text": "Text2Scene: Text-driven Indoor Scene Stylization with Part-aware Details", + "text_level": 1, + "bbox": [ + 112, + 130, + 854, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inwoo Hwang $^{1}$ , Hyeonwoo Kim $^{1}$ , and Young Min Kim $^{1,2}$", + "bbox": [ + 254, + 179, + 712, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Electrical and Computer Engineering, Seoul National University", + "bbox": [ + 243, + 200, + 750, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Interdisciplinary Program in Artificial Intelligence and INMC, Seoul National University", + "bbox": [ + 217, + 218, + 751, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/673ec43ade49462211c91ca5548d95ad0d40a65ad68d21ffbaf9013e67d5e0c8.jpg", + "image_caption": [ + "Figure 1. Stylization results of diverse objects. Text2Scene creates more realistic and part-aware textures for various categories of 3D objects without any dedicated datasets for training, which can constitute high-quality virtual scenes." + ], + "image_footnote": [], + "bbox": [ + 81, + 268, + 890, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 574, + 313, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Text2Scene, a method to automatically create realistic textures for virtual scenes composed of multiple objects. Guided by a reference image and text descriptions, our pipeline adds detailed texture on labeled 3D geometries in the room such that the generated colors respect the hierarchical structure or semantic parts that are often composed of similar materials. Instead of applying flat stylization on the entire scene at a single step, we obtain weak semantic cues from geometric segmentation, which are further clarified by assigning initial colors to segmented parts. Then we add texture details for individual objects such that their projections on image space exhibit feature embedding aligned with the embedding of the input. The decomposition makes the entire pipeline tractable to a moderate amount of computation resources and memory. As our framework utilizes the existing resources of image and text embedding, it does not require dedicated datasets with high-quality textures designed by skillful artists. To the best of our knowledge, it is the first practical and scalable approach that can", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "create detailed and realistic textures of the desired style that maintain structural context for scenes with multiple objects.", + "bbox": [ + 500, + 575, + 890, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 633, + 632, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virtual spaces provide an immersive experience for metaverse, films, or games. With increasing demands for virtual environments, various applications seek practical methods to create realistic 3D scenes with high-quality textures. Currently, skillful artists need to manually create 3D assets and accompanying textures with careful parameterization, which is not scalable enough to account for the diverse content the industry is heading for. Scenes can also be populated with existing 3D database models or created with recent shape-generation approaches using data-driven methods [33, 54]. However, most of them lack texture information or are limited to simple coloring.", + "bbox": [ + 496, + 659, + 890, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To build realistic content, we need fine details containing the artistic nuances of styles that obey the implicit correlations with geometric shapes and semantic structure. Recent works provide methods to color a single object with the help", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1890", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d89b9b0ec45bb56b67daa99846fb8e3566136869c74ef85896fc6368e0d7d6a6.jpg", + "image_caption": [ + "Figure 2. Our scene stylization results. Given a target image $I_{t}$ and the style text $t_{s}$ , Text2Scene can produce the stylized results for the entire scene." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 529, + 377 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7197f56ead959c7bf1a5ce7820d79c6ba6eb7f658d9df5fb88d77e817f8e27fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 90, + 887, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of differentiable rendering [5], but often they are limited to single texture or blurred boundaries [6,21,27,29,52]. More importantly, the 3D objects are often textured in isolation, and only limited attempts exist to add visual appearances for large-scale scenes with multiple objects [14, 15, 19]. The biggest challenge is adding consistent style for an entire scene, but still accounting for the boundaries of different materials due to the functional and semantic relationship between parts, as observed within real-world scenes.", + "bbox": [ + 75, + 430, + 468, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed Text2Scene adds plausible texture details on 3D scenes without explicit part labels or large-scale data with complex texturing. We take inspiration from abundant 3D shape and image datasets and decompose the problem into sub-parts such that the entire scene can be processed with a commodity memory and computation. Given scenes of multiple objects of 3D mesh geometry, we separately handle walls and individual objects. Specifically, the stylization of walls is formulated as texture retrieval, and the objects are initialized with base colors. From the base color assignment, we can deduce the part-level relationship for stylization and further refine them in later stages, such that their rendered images are close to the input text within the joint embedding space of foundational models.", + "bbox": [ + 75, + 566, + 468, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our coarse-to-fine strategy keeps the problem tractable yet generates high-quality texture with clean part boundaries. We first create segments of input mesh such that the segment boundaries align with low-level geometric cues. Then we start with the simplified problem of assigning a color per segment. Interestingly, the prior obtained from large-scale image datasets assign similar colors for the parts with similar textures, reflecting the semantic context or", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "symmetry as shown in Figure 1. We add the detailed texture on individual objects as an additional perturbation on the assigned base colors by enforcing constraints on the image features of their projections. The additional perturbations are high-frequency neural fields added to the base color.", + "bbox": [ + 496, + 430, + 890, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, Text2Scene is a new method that", + "bbox": [ + 517, + 506, + 823, + 520 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- can easily generate realistic texture colors of the scene with the desired style provided by text or an image;", + "- can add detailed texture that respects the semantic part boundaries of individual objects; and", + "- can process the entire scene without a large amount of textured 3D scenes or an extensive memory footprint." + ], + "bbox": [ + 517, + 530, + 890, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We expect the proposed approach to enable everyday users to quickly populate virtual scenes of their choices, and enjoy the possibility of next-generation technology with high-quality visual renderings.", + "bbox": [ + 496, + 650, + 890, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 724, + 643, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Shape Understanding and Segmentation Our proposed method creates a realistic texture that abides by implicit rules of the material composition of different object parts. Different materials or textures are often assigned to different objects' functional parts, whose boundaries align with geometric feature lines. A handful of previous works find segments that constitute a 3D model with geometric information [7, 12, 20, 34], while others train 3D features to distinguish part labels provided in datasets [9, 30, 38]. Recently, PartGlot [24] suggested discovering part segments", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1891", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8ef56bbfb6fe35971948f03ef924d56f162c09788638e135d2c0947895410c02.jpg", + "image_caption": [ + "Figure 3. Overall pipeline of Text2Scene. Given a 3D scene $S$ with optional text description, we generate textures guided by a target image $I_{t}$ and the appearance style provided as a text $t_s$ . We first stylize the structure by texture retrieval, and each object is pre-processed for part discovery for stylization. Then, we assign base colors to object parts and add local details for each object with the designated LNSF." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from language references, which should contain functional information. However, the geometric or functional distinction does not always clarify the texture boundaries, with possible additional diversity originating from the designers' choice. The intricate rules are multi-modal distributions composed of a mixture of discrete part assignments and continuous texture details, whose results are contained within visual datasets of scenes.", + "bbox": [ + 75, + 396, + 472, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Fields and Texture Traditional texture atlas of 3D geometry is represented as a mapping from a planar image to a manifold embedded in 3D space and involves complex parameterization. With the increasing popularity of neural representation, TextureFields [31] represented texture as a mapping from 3D surface points to RGB color values without separate parameterization. It is deeply connected with graphics pipelines that adapt coordinate-based functions to depict SDF shapes [26, 32, 45] or novel-view synthesis using implicit volumes [28]. Various works show neural implicit representation is highly flexible and free from domain structure or resolution [2, 22, 25, 37, 49]. Recently, Text2Mesh [27] generated the deformation and coloring of input mesh with neural fields guided by the joint embedding of rendered image and input text. Interestingly, the generated mesh deformation and coloring also contain semantic part information. Our Text2Scene framework further extends the ability and observes distinct part boundaries, which could not be captured in previous works. We also increase the reality of resulting scenes with high-frequency details. We encode the input to the neural network with a high-order basis of intrinsic features [23] and apply the coarse-to-fine strategy as shown with geometric details of Yifan et al. [51].", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-Driven 3D Stylization Recently neural networks trained with large-scale images and texts demonstrated powerful performance in many tasks with their extensive representation power. Here we primarily focus on 3D stylization attempts using image features. CLIP [39] learns latent space with a large amount of image text pairs, and the additional text input allows semantic manipulation for various generative tasks, including images [8, 35, 40–42], videos [3, 17], motions [46, 47], and 3D assets [18, 21, 36, 43]. However, for text-driven 3D stylization, it is still difficult to define clean texture boundaries and correlation with instance-level subtleties [6, 27, 29] or focus only on the specific type, such as human [16, 53]. As another way to stylize a 3D scene [4, 50], Yeh et al. [50] matches the input image features with CAD input using differentiable material graphs [44]. However, the representation is inherently limited to a combination of material libraries and cannot handle non-homogeneous details such as painting. On the other hand, our approach can generate texture beyond the repetitive low-level patterns of input materials and introduce part-aware texture with fine details that maintain geometric and semantic consistency.", + "bbox": [ + 496, + 396, + 893, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 750, + 591, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We stylize a 3D indoor scene without sophisticated techniques or software tools for 3D modeling or texturing. The input 3D scene $\\mathcal{S} = \\{\\mathcal{W},\\mathcal{O}\\}$ is a set of structure components $\\mathcal{W}$ and a set of objects $\\mathcal{O}$ . The structural components $\\mathcal{W}$ are walls, ceilings, and floors, whereas the objects $\\mathcal{O} = \\{M_i\\}$ are the 3D mesh models $M_{i}$ . We assume that all the components have their corresponding class labels and optionally have text descriptions.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1892", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The desired color distribution is provided as a target image $I_{t}$ , which could be retrieved from the web. Also, a specific appearance style description $t_s$ is provided as input to enforce style consistency among a set of objects $\\mathcal{O}$ . The color distribution compares the color histogram of the input target image $I_{t}$ against the rendering of the current stylized scene $I$ . Specifically, the histogram loss $\\mathcal{L}_{\\mathrm{hist}}$ is defined as below", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h i s t}} (I, I _ {t}) = \\| H ^ {1 / 2} (I) - H ^ {1 / 2} (I _ {t}) \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 224, + 468, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H(\\cdot)$ indicates differentiable color histogram operator [1]. Also, we augment the losses derived from the joint embedding of text and images [39] by generating text descriptions $T$ of the context using semantic labels and comparing them against the rendered image $I$ . If we denote the pre-trained encoder for image and text as $E_{1}$ and $E_{2}$ , respectively, the CLIP similarity loss is defined as", + "bbox": [ + 75, + 256, + 472, + 363 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l i p}} (I, T) = 1 - \\operatorname {s i m} \\left(E _ {1} (I), E _ {2} (T)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 376, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $sim(\\mathbf{x},\\mathbf{y}) = \\frac{\\mathbf{x}^\\top\\mathbf{y}}{\\|\\mathbf{x}\\|_2\\|\\mathbf{y}\\|_2}$ is the cosine similarity.", + "bbox": [ + 76, + 406, + 447, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline is described in Fig. 3. We obtain the texture for the structure $\\mathcal{W}$ by texture retrieval, which is described in Sec. 3.1. The objects are stylized with additional decomposition to respect local part boundaries and, simultaneously, to practically handle multiple entities with details (Sec. 3.2).", + "bbox": [ + 75, + 439, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Structure Stylization", + "text_level": 1, + "bbox": [ + 76, + 542, + 272, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assign one coherent texture per structural element of $\\mathcal{W}$ . Compared to objects, the structural elements, such as walls or ceilings, are of simple planar geometry, and their textures are not heavily dependent on the relationships between different functional parts. For structural elements, it suffices to pick texture from the texture set of an existing material library of MATch [44], which contains homogeneous material. If we instead utilize visual features or CLIP embeddings for them, the resulting stylization exhibits undesired artifacts of various sizes instead of constant patterns, as shown in the supplementary.", + "bbox": [ + 75, + 566, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We randomly initialize the texture from the texture set and render the structure image $I_{s}$ , a bare room only containing the structural elements $\\mathcal{W}$ without objects. Then the materials are compared to the target image $I_{t}$ for the histogram losses of Equation (1). The additional text prompt $T_{s}$ is given as 'a structure of a room' to provide the context with $\\mathcal{L}_{\\mathrm{clip}}(I_s,T_s)$ . In summary, the texture of the structural element is retrieved to have the lowest score on the following criteria:", + "bbox": [ + 75, + 734, + 470, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h i s t}} \\left(I _ {s}, I _ {t}\\right) + \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {c l i p}} \\left(I _ {s}, T _ {s}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 885, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/abc061eab5ee066c13372ce6e6b555c85487678e8f03a3a6c41c4a3c90b0863b.jpg", + "image_caption": [ + "Figure 4. Overall pipeline of part discovery. We discover parts for object stylization from the super-segments of 3D object mesh. Given a 3D object mesh $M_{i}$ with segments $\\{s_{ik}^{l}\\}$ at $l^{\\mathrm{th}}$ iteration, we assign a color $c(s_{ik}^{l})$ per segment and generate a graph $\\mathcal{G}_i^l$ . The pair of neighboring nodes is merged if the distance between assigned colors $c(s_{ik}^{l})$ is within a threshold, then, move to $(l + 1)^{\\mathrm{th}}$ iteration." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 893, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Object Stylization", + "text_level": 1, + "bbox": [ + 500, + 297, + 674, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object stylization involves understanding the semantic structure hidden behind the mesh representation. As a preprocessing, we first subdivide individual objects $M_{i}$ in $\\mathcal{O}$ into part segments $\\{s_{ik}\\}$ as described in Sec. 3.2.1. Then the scene is stylized in two steps. First, we assign base colors into individual parts to minimize the style loss for the entire scene $S = \\{\\mathcal{W},\\mathcal{O}\\}$ (Sec. 3.2.2). Here we are optimizing for the discrete set of colors assigned to subdivided parts obtained from the pre-processing. Then the textures for individual objects are further optimized to generate fine details (Sec. 3.2.3).", + "bbox": [ + 496, + 321, + 893, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Part Discovery for Object Stylization", + "text_level": 1, + "bbox": [ + 500, + 512, + 810, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first decompose individual objects into parts such that each part is composed of the same material or texture. The distinctive part boundaries are critical in providing semantic consistency and, therefore, perceptual realism toward the scene. Similar to 3D part segmentation methods, we first find super-segments based on geometric features which provide the granularity to define textural parts. For a given 3D object mesh $M_{i}$ , the initial segments $\\{s_{ik}^{0}\\}$ are the decomposition applying the method by Katz et al. [20]. The decomposition is designed to be an over-segmentation of our aim. We generate a graph $\\mathcal{G}_i^0$ where each node is the segment and edges connect neighboring segments. Then we incrementally merge segments that belong to the same texture until convergence. Note that, our part discovery method operates robustly regardless of the initial composition, but we use [20] which preserves the original geometry and details.", + "bbox": [ + 496, + 537, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The challenge here is that, unlike semantic segmentation approaches, no large-scale public dataset exists that provides the ground truth for 'texture similarity' as the segmentation labels. We find a supervision signal from the large-scale pre-trained model, and create a simple text prompt $T_{i,c}$ using the class name, such as a bed or a chair. At $l^{\\mathrm{th}}$ iteration, we assign a color $c(s_{ik}^{l})$ to each segment $\\{s_{ik}^{l}\\}$ , which is optimized to minimize the distance", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1893", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between the rendered images $I_{M_i}$ of multiple viewpoints and the text $T_{i,c}$ in the joint embedding space of CLIP, or $\\mathcal{L}_{\\mathrm{clip}}(I_{M_i},T_{i,c})$ as defined in Equation (2). If the resulting colors assigned to two adjacent segments are similar, the two parts are likely to be the parts with the same texture source. Therefore we merge the two segments for the next iteration $\\{s_{ik}^{l + 1}\\}$ . In particular, we merge segments if the assigned color has a distance of less than a threshold $\\lambda_{th}$ in the CIE color space which is known to be related to human perception. Note that, while the initial color is gray for the assignment, merging segments happen with the optimized color. We repeat the process until the number of segments does not decrease anymore and empirically found that it usually converges within 2-3 steps. The overall pipeline for part discovery is also described in Fig. 4.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Part-level Base Color Assignment", + "text_level": 1, + "bbox": [ + 76, + 334, + 366, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the objects $M_{i} \\in \\mathcal{O}$ are decomposed into parts $\\{s_{ik}\\}$ , we assign a solid color per part, namely $c(s_{ik})$ . The base color assignment handles a low-dimensional optimization space with a coarse set of discrete parts but still observes the holistic distribution of the entire scene. Then the base color is combined with the output of designed neural style fields in Sec. 3.2.3, which generate high-frequency local texture.", + "bbox": [ + 75, + 358, + 468, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We optimize the base color using the combined loss as before, $\\mathcal{L}_{\\mathrm{color,scene}} + \\mathcal{L}_{\\mathrm{clip,scene}}$ . The color loss again considers the similarity with the target image $\\mathcal{L}_{\\mathrm{hist}}(I, I_t)$ . The scene being optimized $I$ is rendered with the stylized structure $\\mathcal{W}$ and the current estimates of the base colors. The clip loss considers both individual objects and the global scene and is calculated as the sum of object clip loss and global clip loss.", + "bbox": [ + 75, + 464, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l i p , s c e n e}} = \\lambda_ {2} \\cdot \\sum_ {i} \\mathcal {L} _ {\\text {c l i p}} \\left(I _ {M _ {i}}, T _ {i}\\right) + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {c l i p}} (I, T). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 589, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike Sec. 3.2.1, text description $T_{i}$ for object $M_{i}$ could be a simple text prompt using the class name, or a detailed text prompt based on user choice. We render individual objects from various angles $I_{M_i}$ and compare them with the text description $T_{i}$ . The embedding for the scene is also compared against the text embedding $T$ represents the type of scene, for example, 'a bedroom'. By jointly applying the loss, the base color $c(s_{ik})$ is selected as a representative color that harmonizes nicely with the global context.", + "bbox": [ + 75, + 625, + 470, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Detailed Stylization", + "text_level": 1, + "bbox": [ + 76, + 779, + 267, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The base color is combined with additional details regressed from a neural network to express the detailed local texture. We define local neural style field (LNSF) for each object, which generates the local textures added to the base color. The color for point $p$ is defined by the following equation,", + "bbox": [ + 75, + 801, + 467, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc \\left(s _ {p}\\right) + \\alpha \\cdot \\mathcal {F} _ {i} (\\gamma (p), \\phi (p), s _ {p}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train a LNSF $\\mathcal{F}_i$ per object, which outputs the color to be added to the base color $c(s_p)$ , where $s_p$ indicates the part segment id. The color range $\\alpha$ maintains the final color to be similar to the base color. $\\gamma(\\cdot)$ is the positional encoding of the $xyz$ coordinate to capture high-frequency details, and $\\phi(\\cdot)$ represents the coefficients of the eigenfunctions for the intrinsic geometry using the Laplace-Beltrami operator. Therefore the object-specific neural fields respect the part boundaries and local geometric details.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LNSF for each object are trained with additional style context $\\mathcal{L}_{\\mathrm{clip}}(I_{M_i},T_i^+)$ . The text prompt $T_{i}^{+}$ augments the object description of $T_{i}$ with appearance style description $t_s$ , such as 'minimal style' or 'Mid-Century style'. By enforcing the same appearance style description, we can weakly bind the styles of individual objects in the same scene while optimizing separately. We could also use $T_{i}^{+}$ instead of $T_{i}$ for optimizing the base color, and it shows slightly better results.", + "bbox": [ + 496, + 227, + 890, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Part-aware Geometric Deformation Even though our main focus is to add colors to objects with local part-aware details, we could also concurrently produce part-aware geometric deformation with the same architecture. We can slightly change the LNSF $\\mathcal{F}$ to have two branches of output that estimate color and displacement. For each point on the surface, we assign color by Eq. 5 and adjust displacement along the vertex normal direction. To learn the effective deformation, we add geometric loss $\\mathcal{L}_{\\mathrm{clip}}(I_{M_i}^{geo}, T_i^+)$ , where $I_{M_i}^{geo}$ is an image rendering textureless geometry as [27].", + "bbox": [ + 496, + 383, + 890, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering and Implementation Details To render each object, the individual objects $M_{i}$ are scaled to fit a unit box. The predicted color of each vertex allows the entire mesh to be differentiable rendering through interpolation using [5]. We render individual objects with random augmented backgrounds (white, black, random Gaussian, chess board), which helps the pipeline focus on the foreground object [16, 18]. Inspired by [36], since CLIP has a bias for the canonical pose, we augment the text prompt with the azimuth angle of view, namely 'front view of', 'side view of' or 'back view of'. Finally, in Sec 3.2.3, random perspective transformation and random crop boost learning local details. To render a scene, we randomly sampled from predefined 20 camera poses to evenly cover the entire scene $S$ .", + "bbox": [ + 496, + 556, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 784, + 633, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our stylization is first evaluated for individual objects in Sec. 4.1. We evaluate the quality of textured object meshes and also assess that our pipeline can discover part segments to assign realistic stylization. Then we demonstrate the stylization results of room-scale scenes with multiple objects in Sec. 4.2.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1894", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aa4b62ba5d2e5162222124dd0f8c2b35d9144a639b34d54e50ae5f6716ada7a9.jpg", + "image_caption": [ + "Figure 5. Visual comparison with baselines. Our method utilizes the found part information and generates detailed textures through the designed network in combination with a coarse-to-fine stylization scheme. As a result, it generates a globally harmonious 3D style that clearly classifies part segments. We can handle diverse text such as simple categories or styles, detailed descriptions, or text that includes different two object categories." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 883, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Object Stylization", + "text_level": 1, + "bbox": [ + 76, + 645, + 251, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show that our method can stylize various objects and produce realistic 3D assets to populate virtual scenes. We use object meshes of various types and sizes, collected from Turbo Squid [48], 3D-FUTURE [13], and Amazon Berkeley Objects [10]. For detailed stylization, we subdivide the object into an average of 119769 faces and 60437 vertices. For large general objects such as beds, sofas, tables, etc., we use class labels and the text input for the specific style as 'a [class label], [specific] style'. However, we need a detailed explanation for small objects that cannot be explained only by class labels, such as books. For these objects, we configure detailed text individually, for example, 'a book of Harry Potter and the Sorcerer's Stone'.", + "bbox": [ + 75, + 671, + 470, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 shows the renderings of stylized results. We render objects in Blender [11] with a fixed lighting setup.", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The input mesh and the text are also provided in the left column. Since our method utilizes the obtained part information and coarse-to-fine stylization scheme with a two-step approach, it creates a more realistic texture with clear boundaries for each part of the 3D assets. Text2Mesh [27] is another text-driven 3D stylization approach, which adds the RGB color and deformation fields on the vertices of mesh. While it augments the detailed variations on the input mesh, the deformation map can occasionally introduce undesired artifacts and the part boundaries are only roughly estimated. We also provide the results of an ablated version that directly uses LNSF without a two-step stylization scheme. Our part discovery module and coarse-to-fine stylization scheme play a critical role in producing realistic assets with high-quality stylization. Figure 1 contains more results on diverse objects.", + "bbox": [ + 496, + 645, + 890, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1895", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbbc634a8d973165b74fb6edf98a0bfc222b7fbe8a748189cbb36b7bb5f1b2c7.jpg", + "image_caption": [ + "Figure 6. Results of user study for object stylization" + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 444, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Because stylization is a subjective task, and no ground truth or metrics exist, we conduct a user study for quantitative evaluation. We ask 98 users to rate the quality of generated outputs on a scale of 1 (worst) to 5 (best) in response to the following questions: (Q1) 'How natural are the output results?', (Q2) 'How well does the output contain text information?', and (Q3) 'How well does the output reflect part information?' Figure 6 contains the mean and the standard deviation of the scores. Our method outperforms competing methods in all aspects. Therefore we conclude that our method generates a realistic texture that abides by input text description and semantic part information.", + "bbox": [ + 75, + 233, + 472, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12fe3abb18708b9695db216600524072029067d8b93f4d94ee2709be27ce5338.jpg", + "image_caption": [ + "Figure 7. With a slightly modified LNSF $\\mathcal{F}$ , we can simultaneously generate the texture with both color and displacement fields, and all of the generated style information respects the part information discovered." + ], + "image_footnote": [], + "bbox": [ + 89, + 428, + 459, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also provide a parallel comparison against Text2Mesh by incorporating deformation fields with our approach. Our original method preserves the original shape of the mesh, while Text2Mesh deforms vertices along the normal direction in addition to generating texture. We can make a similar version of our LNSF and produce additional deformation as described in Sec. 3.2.3. Figure 7 shows the deformed results from the source mesh compared to Text2Mesh. Since our approach explicitly considers the part information with the two-stage approach, our results with deformation fields also respect different semantic parts of the object.", + "bbox": [ + 75, + 667, + 470, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part Discovery As a side product of object stylization, we can discover different parts with distinguished boundaries", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ade3d4c4ae2d3e45c72074051f8a85e117a2a6262b612c50c3bde8156a84282d.jpg", + "image_caption": [ + "Figure 8. The robustness against initial super-segmentation. The super-segments of the top and the bottom rows are generated by [20] and [7], respectively (left). Starting from the initial color assignment (middle), our approach stably finds part decomposition to assign different base colors and therefore different texture information (right). The text a chair is used." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that can guide a realistic color assignment (Sec. 3.2.1). Our part discovery combines geometric super-segments and implicit clues from image-text embeddings. The initial super-segments guide the algorithm to follow geometric feature lines, and the final results stably find part information despite different initializations. Figure 8 shows examples initialized with Katz et al. [20] (top) and BSP-Net [7] (bottom). Even for input mesh with bad topology or different initialization, our iterative part discovery quickly finds visually coherent parts without any training with segmentation labels.", + "bbox": [ + 496, + 359, + 890, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Scene Stylization", + "text_level": 1, + "bbox": [ + 500, + 541, + 666, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now we demonstrate that our text2scene framework can quickly generate a realistic texture for a room with multiple objects. Recall that the 3D scene $S = \\{\\mathcal{W}, \\mathcal{O}\\}$ is composed of the structure components $\\mathcal{W}$ and a set of objects $\\mathcal{O}$ . We use the same objects as described in Sec. 4.1 and arrange them to constitute scenes. As there is no existing dataset composite of complete object meshes with labels, we built a total of four scenes: two bedrooms and two living rooms. Each scene contains an average of 20 objects of various sizes and classes. Additionally, we provide a target image $I_{t}$ for the color distribution, and a text prompt describing the desired style $t_{s}$ . While it can be daunting to define the desired style for the entire scene, images and texts can provide a simple way to deliver the information.", + "bbox": [ + 496, + 566, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 2 shows stylized results of the same geometry, but observing various target images and style prompts. The generated textures respect the semantic labels of various furniture and different parts and contain localized diverse details. This is in contrast to many prior stylization methods where the fine perturbations are spread throughout the scenes. We also show scenes with different types of rooms containing different objects, but stylized based on the same", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1896", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "target image in Fig. 9. Additional results of various input configurations are available in the supplementary material. Note that our target image $I_{t}$ does not need to restrict as an indoor image, and can be replaced such as natural photographs. And by changing random seeds, diverse results can be obtained from the same input conditions. Also, since we stylize the entire object, we can easily edit the 3D scene through object relocation. These results can also be found in the supplementary.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec2cf0e19f6b4d261fead6170adddba1ab3e8a60072081e13a764f68aff450ec.jpg", + "image_caption": [ + "Livingroom", + "Figure 9. Result for the different arrangement of objects. We used target image number 3 in Fig. 2 in both spaces." + ], + "image_footnote": [], + "bbox": [ + 98, + 239, + 274, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b341e11d77ac96fd8cf200f5c16f769cd739ac944e9159906297bcad4d3a28e.jpg", + "image_caption": [ + "Bedroom" + ], + "image_footnote": [], + "bbox": [ + 274, + 238, + 450, + 347 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66cc16f9fc544a2bfb8fbe7cb3d5e67a839ffcd6d22bbe0e73e5ad573b916924.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
-retrieval-hist,glo-detailOurs
(Q1): Realistic2.23(±0.48)3.68(±0.49)4.02(±0.49)4.22(±0.42)
(Q2): Color2.09(±0.48)2.63(±0.59)3.95(±0.47)3.86(±0.49)
", + "bbox": [ + 81, + 414, + 467, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Results of user study for scene stylization", + "bbox": [ + 122, + 462, + 423, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61b09640015035873ba64b4676b91d1078418a9363d8611416900b2f9f2955aa.jpg", + "image_caption": [ + "- retrieval", + "Figure 10. Ablation results for the scene. We used target image number 1 in Fig. 2 in all spaces." + ], + "image_footnote": [], + "bbox": [ + 78, + 488, + 176, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff8c3cad378d18f90b5d540424e739799571d994195a70c3200b62b0e76677f8.jpg", + "image_caption": [ + "-hist,glo" + ], + "image_footnote": [], + "bbox": [ + 176, + 488, + 272, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/476364c798fe33cf99a0894c70cbd6bdf9721e35f03b1455e6be1a4f66ad7349.jpg", + "image_caption": [ + "-detail" + ], + "image_footnote": [], + "bbox": [ + 274, + 488, + 370, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e2a5d164df5e0d1b2770ab7b9aeaf6bcdf063b601e87e9b44acf4a47ad02076.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 370, + 488, + 468, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also provide users' evaluation of the quality of our stylization in Table 1. Users evaluate the results by answering the following questions: (Q1) 'How realistic is the output result?' (Q2) 'How similar is the color distribution of the scene to the given images?' Users assess the overall quality of the outputs, and how well they match the target image. Since there are no previous works that use the same setting, we compare the results of ablated versions: -retrieval replaces the separate texture retrieval module (Sec. 3.1) with a stylization network of objects for structure components; -hist,glo removes the color loss and the global clip loss for the base color assignment (Sec. 3.2.2); and -detail removes the detailed stylization step for objects (Sec. 3.2.3). The responses of (Q1) indicate that Text2Scene generates high-quality scenes and each of the components plays a crucial role to achieve reality. The effect of texture retrieval is the most prominent. The color distribution results (Q2) indicate that the base color assignment is critical. The added local details are more important for the realism", + "bbox": [ + 75, + 613, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the results. Figure 10 shows exemplar images of the ablated versions used for the user study.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPU Cost and Scalability Text2Scene first assigns base colors to discovered parts for all objects in the scene, and generates details of objects individually. The cardinality for the base color assignment is only a few hundred and it does not require much memory and allows us to consider the whole scene while expanding the scale. The most memory-intensive process is the detail generation using a neural network, which only processes a single object at a time. Therefore the entire process can be trained only on a single 11 GB GPU, making it an accessible tool for casual users. In a single GPU, the base color allocation of the entire space takes 5 hours, and learning the details for each object takes 10 minutes.", + "bbox": [ + 496, + 138, + 890, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations While our approach results in scene stylization, the pipeline separately handles individual objects after the base color assignment. This is a practical choice for scalability but may lack an understanding of the context of the entire space. Instead, we rely on the text description to weakly bind the objects into a similar style. We can design the pipeline to receive an additional input with a texture map or a lightweight network and extend our model to better observe the holistic scene context within a limited GPU memory. Also, our pipeline requires a class label or optional text description per object, which can be further automated.", + "bbox": [ + 496, + 354, + 890, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 536, + 625, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Text2Scene, a novel framework to generate a texture for 3D scenes. Our hierarchical framework can handle a variety of objects, including highly detailed textures for objects such as book or paintings. By leveraging the representation power of pre-trained CLIP, the framework does not require any 3D datasets with texture or part annotation. Given the 3D mesh models and the class labels or text descriptions of objects, our framework easily produces stylized results by picking a target image and a simple text description. We hope that Text2Scene to facilitate the automatic interior recommendation or realistic virtual space generation.", + "bbox": [ + 496, + 561, + 890, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00208197) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub). Inwoo Hwang is supported by Hyundai Motor Chung Mong-Koo Foundation. Young Min Kim is the corresponding author.", + "bbox": [ + 496, + 760, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1897", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Affifi, Marcus A. Brubaker, and Michael S. Brown. Histogram: Controlling colors of gan-generated and real images via color histograms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 4", + "[2] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290–6301, June 2022. 3", + "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3", + "[4] Kang Chen, Kun Xu, Yizhou Yu, Tian-Yi Wang, and ShiMin Hu. Magic decorator: Automatic material suggestion for indoor digital scenes. ACM Trans. Graph., 34(6):232:1-232:11, Oct. 2015. 3", + "[5] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 5", + "[6] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3", + "[7] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bspnet: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 7", + "[8] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3", + "[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. Proceedings of International Conference on Computer Vision (ICCV), 2019. 2", + "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 6", + "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6", + "[12] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2", + "[13] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d fur" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "niture shape with texture. International Journal of Computer Vision, pages 1-25, 2021. 6", + "[14] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 2", + "[15] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2", + "[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3, 5", + "[17] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. CVPR, 2022. 3, 5", + "[19] Jaebong Jeong, Janghun Jo, Sunghyun Cho, and Jaesik Park. 3d scene painting via semantic image synthesis. In Proceedings of the International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[20] Sagi Katz and Ayellet Tal. Hierarchical mesh decomposition using fuzzy clustering and cuts. ACM Transactions on Graphics, 22, 06 2003. 2, 4, 7", + "[21] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, December 2022. 2, 3", + "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 3", + "[23] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, 2022. 3", + "[24] Juil Koo, Ian Huang, Panos Achlioptas, Leonidas J Guibas, and Minhyuk Sung. Partglot: Learning shape part segmentation from language reference games. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 3", + "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3", + "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. arXiv preprint arXiv:2112.03221, 2021. 2, 3, 5, 6" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1898", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3", + "[29] Shailesh Mishra and Jonathan Granskog. Clip-based neural neighbor style transfer for 3d assets. In arXiv, 2022. 2, 3", + "[30] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2", + "[31] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3", + "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation, 2019. 3", + "[33] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1", + "[34] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 3", + "[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3, 5", + "[37] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961, 2020. 3", + "[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 4", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3", + "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arxiv:2208.12242, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 3", + "[44] Liang Shi, Beichen Li, Milos Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: Differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6):1-15, Dec. 2020. 3, 4", + "[45] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions, 2020. 3", + "[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 3", + "[47] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3", + "[48] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2022.6", + "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2021. 3", + "[50] Yu-Ying Yeh, Zhengqin Li, Yannick Hold-Geoffroy, Rui Zhu, Zexiang Xu, Miloš Hasan, Kalyan Sunkavalli, and Manmohan Chandraker. Photoscene: Photorealistic material and lighting transfer for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18562–18571, June 2022. 3", + "[51] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2022. 3", + "[52] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 2", + "[53] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In ECCV, 2022. 3", + "[54] Dongsu Zhang, Changwoon Choi, Inbum Park, and Young Min Kim. Probabilistic implicit scene completion. In International Conference on Learning Representations, 2022. 1" + ], + "bbox": [ + 501, + 92, + 890, + 739 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1899", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_model.json b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f60864356421710a7771a13b3bd7c35f36c96db --- /dev/null +++ b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.856, + 0.152 + ], + "angle": 0, + "content": "Text2Scene: Text-driven Indoor Scene Stylization with Part-aware Details" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.18, + 0.714, + 0.2 + ], + "angle": 0, + "content": "Inwoo Hwang\\(^{1}\\), Hyeonwoo Kim\\(^{1}\\), and Young Min Kim\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.201, + 0.751, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Electrical and Computer Engineering, Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.219, + 0.752, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Interdisciplinary Program in Artificial Intelligence and INMC, Seoul National University" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.27, + 0.891, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.534, + 0.893, + 0.564 + ], + "angle": 0, + "content": "Figure 1. Stylization results of diverse objects. Text2Scene creates more realistic and part-aware textures for various categories of 3D objects without any dedicated datasets for training, which can constitute high-quality virtual scenes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.575, + 0.314, + 0.591 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We propose Text2Scene, a method to automatically create realistic textures for virtual scenes composed of multiple objects. Guided by a reference image and text descriptions, our pipeline adds detailed texture on labeled 3D geometries in the room such that the generated colors respect the hierarchical structure or semantic parts that are often composed of similar materials. Instead of applying flat stylization on the entire scene at a single step, we obtain weak semantic cues from geometric segmentation, which are further clarified by assigning initial colors to segmented parts. Then we add texture details for individual objects such that their projections on image space exhibit feature embedding aligned with the embedding of the input. The decomposition makes the entire pipeline tractable to a moderate amount of computation resources and memory. As our framework utilizes the existing resources of image and text embedding, it does not require dedicated datasets with high-quality textures designed by skillful artists. To the best of our knowledge, it is the first practical and scalable approach that can" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.576, + 0.892, + 0.607 + ], + "angle": 0, + "content": "create detailed and realistic textures of the desired style that maintain structural context for scenes with multiple objects." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.634, + 0.633, + 0.65 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Virtual spaces provide an immersive experience for metaverse, films, or games. With increasing demands for virtual environments, various applications seek practical methods to create realistic 3D scenes with high-quality textures. Currently, skillful artists need to manually create 3D assets and accompanying textures with careful parameterization, which is not scalable enough to account for the diverse content the industry is heading for. Scenes can also be populated with existing 3D database models or created with recent shape-generation approaches using data-driven methods [33, 54]. However, most of them lack texture information or are limited to simple coloring." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To build realistic content, we need fine details containing the artistic nuances of styles that obey the implicit correlations with geometric shapes and semantic structure. Recent works provide methods to color a single object with the help" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1890" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.531, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.888, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.893, + 0.41 + ], + "angle": 0, + "content": "Figure 2. Our scene stylization results. Given a target image \\( I_{t} \\) and the style text \\( t_{s} \\), Text2Scene can produce the stylized results for the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.567 + ], + "angle": 0, + "content": "of differentiable rendering [5], but often they are limited to single texture or blurred boundaries [6,21,27,29,52]. More importantly, the 3D objects are often textured in isolation, and only limited attempts exist to add visual appearances for large-scale scenes with multiple objects [14, 15, 19]. The biggest challenge is adding consistent style for an entire scene, but still accounting for the boundaries of different materials due to the functional and semantic relationship between parts, as observed within real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Our proposed Text2Scene adds plausible texture details on 3D scenes without explicit part labels or large-scale data with complex texturing. We take inspiration from abundant 3D shape and image datasets and decompose the problem into sub-parts such that the entire scene can be processed with a commodity memory and computation. Given scenes of multiple objects of 3D mesh geometry, we separately handle walls and individual objects. Specifically, the stylization of walls is formulated as texture retrieval, and the objects are initialized with base colors. From the base color assignment, we can deduce the part-level relationship for stylization and further refine them in later stages, such that their rendered images are close to the input text within the joint embedding space of foundational models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our coarse-to-fine strategy keeps the problem tractable yet generates high-quality texture with clean part boundaries. We first create segments of input mesh such that the segment boundaries align with low-level geometric cues. Then we start with the simplified problem of assigning a color per segment. Interestingly, the prior obtained from large-scale image datasets assign similar colors for the parts with similar textures, reflecting the semantic context or" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.506 + ], + "angle": 0, + "content": "symmetry as shown in Figure 1. We add the detailed texture on individual objects as an additional perturbation on the assigned base colors by enforcing constraints on the image features of their projections. The additional perturbations are high-frequency neural fields added to the base color." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.507, + 0.825, + 0.521 + ], + "angle": 0, + "content": "In summary, Text2Scene is a new method that" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.531, + 0.891, + 0.561 + ], + "angle": 0, + "content": "- can easily generate realistic texture colors of the scene with the desired style provided by text or an image;" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.571, + 0.891, + 0.601 + ], + "angle": 0, + "content": "- can add detailed texture that respects the semantic part boundaries of individual objects; and" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.611, + 0.892, + 0.641 + ], + "angle": 0, + "content": "- can process the entire scene without a large amount of textured 3D scenes or an extensive memory footprint." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.531, + 0.892, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.711 + ], + "angle": 0, + "content": "We expect the proposed approach to enable everyday users to quickly populate virtual scenes of their choices, and enjoy the possibility of next-generation technology with high-quality visual renderings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.645, + 0.74 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3D Shape Understanding and Segmentation Our proposed method creates a realistic texture that abides by implicit rules of the material composition of different object parts. Different materials or textures are often assigned to different objects' functional parts, whose boundaries align with geometric feature lines. A handful of previous works find segments that constitute a 3D model with geometric information [7, 12, 20, 34], while others train 3D features to distinguish part labels provided in datasets [9, 30, 38]. Recently, PartGlot [24] suggested discovering part segments" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1891" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.895, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Overall pipeline of Text2Scene. Given a 3D scene \\(S\\) with optional text description, we generate textures guided by a target image \\(I_{t}\\) and the appearance style provided as a text \\(t_s\\). We first stylize the structure by texture retrieval, and each object is pre-processed for part discovery for stylization. Then, we assign base colors to object parts and add local details for each object with the designated LNSF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.473, + 0.518 + ], + "angle": 0, + "content": "from language references, which should contain functional information. However, the geometric or functional distinction does not always clarify the texture boundaries, with possible additional diversity originating from the designers' choice. The intricate rules are multi-modal distributions composed of a mixture of discrete part assignments and continuous texture details, whose results are contained within visual datasets of scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Neural Fields and Texture Traditional texture atlas of 3D geometry is represented as a mapping from a planar image to a manifold embedded in 3D space and involves complex parameterization. With the increasing popularity of neural representation, TextureFields [31] represented texture as a mapping from 3D surface points to RGB color values without separate parameterization. It is deeply connected with graphics pipelines that adapt coordinate-based functions to depict SDF shapes [26, 32, 45] or novel-view synthesis using implicit volumes [28]. Various works show neural implicit representation is highly flexible and free from domain structure or resolution [2, 22, 25, 37, 49]. Recently, Text2Mesh [27] generated the deformation and coloring of input mesh with neural fields guided by the joint embedding of rendered image and input text. Interestingly, the generated mesh deformation and coloring also contain semantic part information. Our Text2Scene framework further extends the ability and observes distinct part boundaries, which could not be captured in previous works. We also increase the reality of resulting scenes with high-frequency details. We encode the input to the neural network with a high-order basis of intrinsic features [23] and apply the coarse-to-fine strategy as shown with geometric details of Yifan et al. [51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.895, + 0.73 + ], + "angle": 0, + "content": "Text-Driven 3D Stylization Recently neural networks trained with large-scale images and texts demonstrated powerful performance in many tasks with their extensive representation power. Here we primarily focus on 3D stylization attempts using image features. CLIP [39] learns latent space with a large amount of image text pairs, and the additional text input allows semantic manipulation for various generative tasks, including images [8, 35, 40–42], videos [3, 17], motions [46, 47], and 3D assets [18, 21, 36, 43]. However, for text-driven 3D stylization, it is still difficult to define clean texture boundaries and correlation with instance-level subtleties [6, 27, 29] or focus only on the specific type, such as human [16, 53]. As another way to stylize a 3D scene [4, 50], Yeh et al. [50] matches the input image features with CAD input using differentiable material graphs [44]. However, the representation is inherently limited to a combination of material libraries and cannot handle non-homogeneous details such as painting. On the other hand, our approach can generate texture beyond the repetitive low-level patterns of input materials and introduce part-aware texture with fine details that maintain geometric and semantic consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.751, + 0.593, + 0.768 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We stylize a 3D indoor scene without sophisticated techniques or software tools for 3D modeling or texturing. The input 3D scene \\( \\mathcal{S} = \\{\\mathcal{W},\\mathcal{O}\\} \\) is a set of structure components \\( \\mathcal{W} \\) and a set of objects \\( \\mathcal{O} \\). The structural components \\( \\mathcal{W} \\) are walls, ceilings, and floors, whereas the objects \\( \\mathcal{O} = \\{M_i\\} \\) are the 3D mesh models \\( M_{i} \\). We assume that all the components have their corresponding class labels and optionally have text descriptions." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1892" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "The desired color distribution is provided as a target image \\(I_{t}\\), which could be retrieved from the web. Also, a specific appearance style description \\(t_s\\) is provided as input to enforce style consistency among a set of objects \\(\\mathcal{O}\\). The color distribution compares the color histogram of the input target image \\(I_{t}\\) against the rendering of the current stylized scene \\(I\\). Specifically, the histogram loss \\(\\mathcal{L}_{\\mathrm{hist}}\\) is defined as below" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.225, + 0.47, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h i s t}} (I, I _ {t}) = \\| H ^ {1 / 2} (I) - H ^ {1 / 2} (I _ {t}) \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.473, + 0.364 + ], + "angle": 0, + "content": "where \\( H(\\cdot) \\) indicates differentiable color histogram operator [1]. Also, we augment the losses derived from the joint embedding of text and images [39] by generating text descriptions \\( T \\) of the context using semantic labels and comparing them against the rendered image \\( I \\). If we denote the pre-trained encoder for image and text as \\( E_{1} \\) and \\( E_{2} \\), respectively, the CLIP similarity loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.377, + 0.47, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l i p}} (I, T) = 1 - \\operatorname {s i m} \\left(E _ {1} (I), E _ {2} (T)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.407, + 0.449, + 0.441 + ], + "angle": 0, + "content": "where \\(sim(\\mathbf{x},\\mathbf{y}) = \\frac{\\mathbf{x}^\\top\\mathbf{y}}{\\|\\mathbf{x}\\|_2\\|\\mathbf{y}\\|_2}\\) is the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.531 + ], + "angle": 0, + "content": "The overall pipeline is described in Fig. 3. We obtain the texture for the structure \\(\\mathcal{W}\\) by texture retrieval, which is described in Sec. 3.1. The objects are stylized with additional decomposition to respect local part boundaries and, simultaneously, to practically handle multiple entities with details (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.274, + 0.559 + ], + "angle": 0, + "content": "3.1. Structure Stylization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.734 + ], + "angle": 0, + "content": "We assign one coherent texture per structural element of \\(\\mathcal{W}\\). Compared to objects, the structural elements, such as walls or ceilings, are of simple planar geometry, and their textures are not heavily dependent on the relationships between different functional parts. For structural elements, it suffices to pick texture from the texture set of an existing material library of MATch [44], which contains homogeneous material. If we instead utilize visual features or CLIP embeddings for them, the resulting stylization exhibits undesired artifacts of various sizes instead of constant patterns, as shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.871 + ], + "angle": 0, + "content": "We randomly initialize the texture from the texture set and render the structure image \\( I_{s} \\), a bare room only containing the structural elements \\( \\mathcal{W} \\) without objects. Then the materials are compared to the target image \\( I_{t} \\) for the histogram losses of Equation (1). The additional text prompt \\( T_{s} \\) is given as 'a structure of a room' to provide the context with \\( \\mathcal{L}_{\\mathrm{clip}}(I_s,T_s) \\). In summary, the texture of the structural element is retrieved to have the lowest score on the following criteria:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.886, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h i s t}} \\left(I _ {s}, I _ {t}\\right) + \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {c l i p}} \\left(I _ {s}, T _ {s}\\right). \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.895, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.185, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Figure 4. Overall pipeline of part discovery. We discover parts for object stylization from the super-segments of 3D object mesh. Given a 3D object mesh \\( M_{i} \\) with segments \\( \\{s_{ik}^{l}\\} \\) at \\( l^{\\mathrm{th}} \\) iteration, we assign a color \\( c(s_{ik}^{l}) \\) per segment and generate a graph \\( \\mathcal{G}_i^l \\). The pair of neighboring nodes is merged if the distance between assigned colors \\( c(s_{ik}^{l}) \\) is within a threshold, then, move to \\( (l + 1)^{\\mathrm{th}} \\) iteration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.676, + 0.316 + ], + "angle": 0, + "content": "3.2. Object Stylization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.49 + ], + "angle": 0, + "content": "Object stylization involves understanding the semantic structure hidden behind the mesh representation. As a preprocessing, we first subdivide individual objects \\( M_{i} \\) in \\( \\mathcal{O} \\) into part segments \\( \\{s_{ik}\\} \\) as described in Sec. 3.2.1. Then the scene is stylized in two steps. First, we assign base colors into individual parts to minimize the style loss for the entire scene \\( S = \\{\\mathcal{W},\\mathcal{O}\\} \\) (Sec. 3.2.2). Here we are optimizing for the discrete set of colors assigned to subdivided parts obtained from the pre-processing. Then the textures for individual objects are further optimized to generate fine details (Sec. 3.2.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.812, + 0.529 + ], + "angle": 0, + "content": "3.2.1 Part Discovery for Object Stylization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We first decompose individual objects into parts such that each part is composed of the same material or texture. The distinctive part boundaries are critical in providing semantic consistency and, therefore, perceptual realism toward the scene. Similar to 3D part segmentation methods, we first find super-segments based on geometric features which provide the granularity to define textural parts. For a given 3D object mesh \\( M_{i} \\), the initial segments \\( \\{s_{ik}^{0}\\} \\) are the decomposition applying the method by Katz et al. [20]. The decomposition is designed to be an over-segmentation of our aim. We generate a graph \\( \\mathcal{G}_i^0 \\) where each node is the segment and edges connect neighboring segments. Then we incrementally merge segments that belong to the same texture until convergence. Note that, our part discovery method operates robustly regardless of the initial composition, but we use [20] which preserves the original geometry and details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The challenge here is that, unlike semantic segmentation approaches, no large-scale public dataset exists that provides the ground truth for 'texture similarity' as the segmentation labels. We find a supervision signal from the large-scale pre-trained model, and create a simple text prompt \\( T_{i,c} \\) using the class name, such as a bed or a chair. At \\( l^{\\mathrm{th}} \\) iteration, we assign a color \\( c(s_{ik}^{l}) \\) to each segment \\( \\{s_{ik}^{l}\\} \\), which is optimized to minimize the distance" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1893" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "between the rendered images \\( I_{M_i} \\) of multiple viewpoints and the text \\( T_{i,c} \\) in the joint embedding space of CLIP, or \\( \\mathcal{L}_{\\mathrm{clip}}(I_{M_i},T_{i,c}) \\) as defined in Equation (2). If the resulting colors assigned to two adjacent segments are similar, the two parts are likely to be the parts with the same texture source. Therefore we merge the two segments for the next iteration \\( \\{s_{ik}^{l + 1}\\} \\). In particular, we merge segments if the assigned color has a distance of less than a threshold \\( \\lambda_{th} \\) in the CIE color space which is known to be related to human perception. Note that, while the initial color is gray for the assignment, merging segments happen with the optimized color. We repeat the process until the number of segments does not decrease anymore and empirically found that it usually converges within 2-3 steps. The overall pipeline for part discovery is also described in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.335, + 0.367, + 0.351 + ], + "angle": 0, + "content": "3.2.2 Part-level Base Color Assignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.465 + ], + "angle": 0, + "content": "After the objects \\( M_{i} \\in \\mathcal{O} \\) are decomposed into parts \\( \\{s_{ik}\\} \\), we assign a solid color per part, namely \\( c(s_{ik}) \\). The base color assignment handles a low-dimensional optimization space with a coarse set of discrete parts but still observes the holistic distribution of the entire scene. Then the base color is combined with the output of designed neural style fields in Sec. 3.2.3, which generate high-frequency local texture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.584 + ], + "angle": 0, + "content": "We optimize the base color using the combined loss as before, \\(\\mathcal{L}_{\\mathrm{color,scene}} + \\mathcal{L}_{\\mathrm{clip,scene}}\\). The color loss again considers the similarity with the target image \\(\\mathcal{L}_{\\mathrm{hist}}(I, I_t)\\). The scene being optimized \\(I\\) is rendered with the stylized structure \\(\\mathcal{W}\\) and the current estimates of the base colors. The clip loss considers both individual objects and the global scene and is calculated as the sum of object clip loss and global clip loss." + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.59, + 0.47, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l i p , s c e n e}} = \\lambda_ {2} \\cdot \\sum_ {i} \\mathcal {L} _ {\\text {c l i p}} \\left(I _ {M _ {i}}, T _ {i}\\right) + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {c l i p}} (I, T). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Unlike Sec. 3.2.1, text description \\( T_{i} \\) for object \\( M_{i} \\) could be a simple text prompt using the class name, or a detailed text prompt based on user choice. We render individual objects from various angles \\( I_{M_i} \\) and compare them with the text description \\( T_{i} \\). The embedding for the scene is also compared against the text embedding \\( T \\) represents the type of scene, for example, 'a bedroom'. By jointly applying the loss, the base color \\( c(s_{ik}) \\) is selected as a representative color that harmonizes nicely with the global context." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.78, + 0.268, + 0.795 + ], + "angle": 0, + "content": "3.2.3 Detailed Stylization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.468, + 0.879 + ], + "angle": 0, + "content": "The base color is combined with additional details regressed from a neural network to express the detailed local texture. We define local neural style field (LNSF) for each object, which generates the local textures added to the base color. The color for point \\( p \\) is defined by the following equation," + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nc \\left(s _ {p}\\right) + \\alpha \\cdot \\mathcal {F} _ {i} (\\gamma (p), \\phi (p), s _ {p}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "We train a LNSF \\(\\mathcal{F}_i\\) per object, which outputs the color to be added to the base color \\(c(s_p)\\), where \\(s_p\\) indicates the part segment id. The color range \\(\\alpha\\) maintains the final color to be similar to the base color. \\(\\gamma(\\cdot)\\) is the positional encoding of the \\(xyz\\) coordinate to capture high-frequency details, and \\(\\phi(\\cdot)\\) represents the coefficients of the eigenfunctions for the intrinsic geometry using the Laplace-Beltrami operator. Therefore the object-specific neural fields respect the part boundaries and local geometric details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.364 + ], + "angle": 0, + "content": "LNSF for each object are trained with additional style context \\(\\mathcal{L}_{\\mathrm{clip}}(I_{M_i},T_i^+)\\). The text prompt \\(T_{i}^{+}\\) augments the object description of \\(T_{i}\\) with appearance style description \\(t_s\\), such as 'minimal style' or 'Mid-Century style'. By enforcing the same appearance style description, we can weakly bind the styles of individual objects in the same scene while optimizing separately. We could also use \\(T_{i}^{+}\\) instead of \\(T_{i}\\) for optimizing the base color, and it shows slightly better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Part-aware Geometric Deformation Even though our main focus is to add colors to objects with local part-aware details, we could also concurrently produce part-aware geometric deformation with the same architecture. We can slightly change the LNSF \\(\\mathcal{F}\\) to have two branches of output that estimate color and displacement. For each point on the surface, we assign color by Eq. 5 and adjust displacement along the vertex normal direction. To learn the effective deformation, we add geometric loss \\(\\mathcal{L}_{\\mathrm{clip}}(I_{M_i}^{geo}, T_i^+)\\), where \\(I_{M_i}^{geo}\\) is an image rendering textureless geometry as [27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Rendering and Implementation Details To render each object, the individual objects \\( M_{i} \\) are scaled to fit a unit box. The predicted color of each vertex allows the entire mesh to be differentiable rendering through interpolation using [5]. We render individual objects with random augmented backgrounds (white, black, random Gaussian, chess board), which helps the pipeline focus on the foreground object [16, 18]. Inspired by [36], since CLIP has a bias for the canonical pose, we augment the text prompt with the azimuth angle of view, namely 'front view of', 'side view of' or 'back view of'. Finally, in Sec 3.2.3, random perspective transformation and random crop boost learning local details. To render a scene, we randomly sampled from predefined 20 camera poses to evenly cover the entire scene \\( S \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.634, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Our stylization is first evaluated for individual objects in Sec. 4.1. We evaluate the quality of textured object meshes and also assess that our pipeline can discover part segments to assign realistic stylization. Then we demonstrate the stylization results of room-scale scenes with multiple objects in Sec. 4.2." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1894" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.572, + 0.895, + 0.63 + ], + "angle": 0, + "content": "Figure 5. Visual comparison with baselines. Our method utilizes the found part information and generates detailed textures through the designed network in combination with a coarse-to-fine stylization scheme. As a result, it generates a globally harmonious 3D style that clearly classifies part segments. We can handle diverse text such as simple categories or styles, detailed descriptions, or text that includes different two object categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.253, + 0.663 + ], + "angle": 0, + "content": "4.1. Object Stylization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.471, + 0.867 + ], + "angle": 0, + "content": "We show that our method can stylize various objects and produce realistic 3D assets to populate virtual scenes. We use object meshes of various types and sizes, collected from Turbo Squid [48], 3D-FUTURE [13], and Amazon Berkeley Objects [10]. For detailed stylization, we subdivide the object into an average of 119769 faces and 60437 vertices. For large general objects such as beds, sofas, tables, etc., we use class labels and the text input for the specific style as 'a [class label], [specific] style'. However, we need a detailed explanation for small objects that cannot be explained only by class labels, such as books. For these objects, we configure detailed text individually, for example, 'a book of Harry Potter and the Sorcerer's Stone'." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Figure 5 shows the renderings of stylized results. We render objects in Blender [11] with a fixed lighting setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.89 + ], + "angle": 0, + "content": "The input mesh and the text are also provided in the left column. Since our method utilizes the obtained part information and coarse-to-fine stylization scheme with a two-step approach, it creates a more realistic texture with clear boundaries for each part of the 3D assets. Text2Mesh [27] is another text-driven 3D stylization approach, which adds the RGB color and deformation fields on the vertices of mesh. While it augments the detailed variations on the input mesh, the deformation map can occasionally introduce undesired artifacts and the part boundaries are only roughly estimated. We also provide the results of an ablated version that directly uses LNSF without a two-step stylization scheme. Our part discovery module and coarse-to-fine stylization scheme play a critical role in producing realistic assets with high-quality stylization. Figure 1 contains more results on diverse objects." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1895" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.445, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.211, + 0.429, + 0.226 + ], + "angle": 0, + "content": "Figure 6. Results of user study for object stylization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.473, + 0.417 + ], + "angle": 0, + "content": "Because stylization is a subjective task, and no ground truth or metrics exist, we conduct a user study for quantitative evaluation. We ask 98 users to rate the quality of generated outputs on a scale of 1 (worst) to 5 (best) in response to the following questions: (Q1) 'How natural are the output results?', (Q2) 'How well does the output contain text information?', and (Q3) 'How well does the output reflect part information?' Figure 6 contains the mean and the standard deviation of the scores. Our method outperforms competing methods in all aspects. Therefore we conclude that our method generates a realistic texture that abides by input text description and semantic part information." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.429, + 0.46, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Figure 7. With a slightly modified LNSF \\(\\mathcal{F}\\), we can simultaneously generate the texture with both color and displacement fields, and all of the generated style information respects the part information discovered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.471, + 0.851 + ], + "angle": 0, + "content": "We also provide a parallel comparison against Text2Mesh by incorporating deformation fields with our approach. Our original method preserves the original shape of the mesh, while Text2Mesh deforms vertices along the normal direction in addition to generating texture. We can make a similar version of our LNSF and produce additional deformation as described in Sec. 3.2.3. Figure 7 shows the deformed results from the source mesh compared to Text2Mesh. Since our approach explicitly considers the part information with the two-stage approach, our results with deformation fields also respect different semantic parts of the object." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Part Discovery As a side product of object stylization, we can discover different parts with distinguished boundaries" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.259, + 0.894, + 0.343 + ], + "angle": 0, + "content": "Figure 8. The robustness against initial super-segmentation. The super-segments of the top and the bottom rows are generated by [20] and [7], respectively (left). Starting from the initial color assignment (middle), our approach stably finds part decomposition to assign different base colors and therefore different texture information (right). The text a chair is used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.527 + ], + "angle": 0, + "content": "that can guide a realistic color assignment (Sec. 3.2.1). Our part discovery combines geometric super-segments and implicit clues from image-text embeddings. The initial super-segments guide the algorithm to follow geometric feature lines, and the final results stably find part information despite different initializations. Figure 8 shows examples initialized with Katz et al. [20] (top) and BSP-Net [7] (bottom). Even for input mesh with bad topology or different initialization, our iterative part discovery quickly finds visually coherent parts without any training with segmentation labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.667, + 0.558 + ], + "angle": 0, + "content": "4.2. Scene Stylization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Now we demonstrate that our text2scene framework can quickly generate a realistic texture for a room with multiple objects. Recall that the 3D scene \\( S = \\{\\mathcal{W}, \\mathcal{O}\\} \\) is composed of the structure components \\( \\mathcal{W} \\) and a set of objects \\( \\mathcal{O} \\). We use the same objects as described in Sec. 4.1 and arrange them to constitute scenes. As there is no existing dataset composite of complete object meshes with labels, we built a total of four scenes: two bedrooms and two living rooms. Each scene contains an average of 20 objects of various sizes and classes. Additionally, we provide a target image \\( I_{t} \\) for the color distribution, and a text prompt describing the desired style \\( t_{s} \\). While it can be daunting to define the desired style for the entire scene, images and texts can provide a simple way to deliver the information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 2 shows stylized results of the same geometry, but observing various target images and style prompts. The generated textures respect the semantic labels of various furniture and different parts and contain localized diverse details. This is in contrast to many prior stylization methods where the fine perturbations are spread throughout the scenes. We also show scenes with different types of rooms containing different objects, but stylized based on the same" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1896" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "target image in Fig. 9. Additional results of various input configurations are available in the supplementary material. Note that our target image \\( I_{t} \\) does not need to restrict as an indoor image, and can be replaced such as natural photographs. And by changing random seeds, diverse results can be obtained from the same input conditions. Also, since we stylize the entire object, we can easily edit the 3D scene through object relocation. These results can also be found in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.24, + 0.275, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.348, + 0.216, + 0.358 + ], + "angle": 0, + "content": "Livingroom" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.239, + 0.451, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.348, + 0.386, + 0.358 + ], + "angle": 0, + "content": "Bedroom" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.399 + ], + "angle": 0, + "content": "Figure 9. Result for the different arrangement of objects. We used target image number 3 in Fig. 2 in both spaces." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.415, + 0.468, + 0.453 + ], + "angle": 0, + "content": "
-retrieval-hist,glo-detailOurs
(Q1): Realistic2.23(±0.48)3.68(±0.49)4.02(±0.49)4.22(±0.42)
(Q2): Color2.09(±0.48)2.63(±0.59)3.95(±0.47)3.86(±0.49)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.463, + 0.424, + 0.477 + ], + "angle": 0, + "content": "Table 1. Results of user study for scene stylization" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.489, + 0.177, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.553, + 0.159, + 0.563 + ], + "angle": 0, + "content": "- retrieval" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.489, + 0.274, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.553, + 0.254, + 0.564 + ], + "angle": 0, + "content": "-hist,glo" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.489, + 0.371, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.553, + 0.345, + 0.563 + ], + "angle": 0, + "content": "-detail" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.489, + 0.469, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.553, + 0.435, + 0.563 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.576, + 0.469, + 0.604 + ], + "angle": 0, + "content": "Figure 10. Ablation results for the scene. We used target image number 1 in Fig. 2 in all spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We also provide users' evaluation of the quality of our stylization in Table 1. Users evaluate the results by answering the following questions: (Q1) 'How realistic is the output result?' (Q2) 'How similar is the color distribution of the scene to the given images?' Users assess the overall quality of the outputs, and how well they match the target image. Since there are no previous works that use the same setting, we compare the results of ablated versions: -retrieval replaces the separate texture retrieval module (Sec. 3.1) with a stylization network of objects for structure components; -hist,glo removes the color loss and the global clip loss for the base color assignment (Sec. 3.2.2); and -detail removes the detailed stylization step for objects (Sec. 3.2.3). The responses of (Q1) indicate that Text2Scene generates high-quality scenes and each of the components plays a crucial role to achieve reality. The effect of texture retrieval is the most prominent. The color distribution results (Q2) indicate that the base color assignment is critical. The added local details are more important for the realism" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "of the results. Figure 10 shows exemplar images of the ablated versions used for the user study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.892, + 0.336 + ], + "angle": 0, + "content": "GPU Cost and Scalability Text2Scene first assigns base colors to discovered parts for all objects in the scene, and generates details of objects individually. The cardinality for the base color assignment is only a few hundred and it does not require much memory and allows us to consider the whole scene while expanding the scale. The most memory-intensive process is the detail generation using a neural network, which only processes a single object at a time. Therefore the entire process can be trained only on a single 11 GB GPU, making it an accessible tool for casual users. In a single GPU, the base color allocation of the entire space takes 5 hours, and learning the details for each object takes 10 minutes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Limitations While our approach results in scene stylization, the pipeline separately handles individual objects after the base color assignment. This is a practical choice for scalability but may lack an understanding of the context of the entire space. Instead, we rely on the text description to weakly bind the objects into a similar style. We can design the pipeline to receive an additional input with a texture map or a lightweight network and extend our model to better observe the holistic scene context within a limited GPU memory. Also, our pipeline requires a class label or optional text description per object, which can be further automated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.627, + 0.552 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.743 + ], + "angle": 0, + "content": "We introduce Text2Scene, a novel framework to generate a texture for 3D scenes. Our hierarchical framework can handle a variety of objects, including highly detailed textures for objects such as book or paintings. By leveraging the representation power of pre-trained CLIP, the framework does not require any 3D datasets with texture or part annotation. Given the 3D mesh models and the class labels or text descriptions of objects, our framework easily produces stylized results by picking a target image and a simple text description. We hope that Text2Scene to facilitate the automatic interior recommendation or realistic virtual space generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00208197) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub). Inwoo Hwang is supported by Hyundai Motor Chung Mong-Koo Foundation. Young Min Kim is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1897" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Mahmoud Affifi, Marcus A. Brubaker, and Michael S. Brown. Histogram: Controlling colors of gan-generated and real images via color histograms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290–6301, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Kang Chen, Kun Xu, Yizhou Yu, Tian-Yi Wang, and ShiMin Hu. Magic decorator: Automatic material suggestion for indoor digital scenes. ACM Trans. Graph., 34(6):232:1-232:11, Oct. 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[5] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[7] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bspnet: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[8] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.672 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. Proceedings of International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[12] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d fur" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "niture shape with texture. International Journal of Computer Vision, pages 1-25, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[14] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[15] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[17] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[19] Jaebong Jeong, Janghun Jo, Sunghyun Cho, and Jaesik Park. 3d scene painting via semantic image synthesis. In Proceedings of the International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[20] Sagi Katz and Ayellet Tal. Hierarchical mesh decomposition using fuzzy clustering and cuts. ACM Transactions on Graphics, 22, 06 2003. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.547 + ], + "angle": 0, + "content": "[21] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, December 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[23] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[24] Juil Koo, Ian Huang, Panos Achlioptas, Leonidas J Guibas, and Minhyuk Sung. Partglot: Learning shape part segmentation from language reference games. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. arXiv preprint arXiv:2112.03221, 2021. 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1898" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[29] Shailesh Mishra and Jonathan Granskog. Clip-based neural neighbor style transfer for 3d assets. In arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[30] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[31] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[33] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[34] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[37] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.69, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.761, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.847, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arxiv:2208.12242, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[44] Liang Shi, Beichen Li, Milos Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: Differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6):1-15, Dec. 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[45] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[47] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[48] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[50] Yu-Ying Yeh, Zhengqin Li, Yannick Hold-Geoffroy, Rui Zhu, Zexiang Xu, Miloš Hasan, Kalyan Sunkavalli, and Manmohan Chandraker. Photoscene: Photorealistic material and lighting transfer for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18562–18571, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[51] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[52] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[53] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[54] Dongsu Zhang, Changwoon Choi, Inbum Park, and Young Min Kim. Probabilistic implicit scene completion. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1899" + } + ] +] \ No newline at end of file diff --git a/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_origin.pdf b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f5480d798577c365ea9bfefbb578432598627bd6 --- /dev/null +++ b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/3c9b1104-12d9-49d4-abdb-99527f418bd3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d71bca38d0b4bcc7a88037739835c34a27500f6e821cebf6b97617f7b6f5280 +size 4849087 diff --git a/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/full.md b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/full.md new file mode 100644 index 0000000000000000000000000000000000000000..69294a30d75acc4c8bbbe0687873a188348d087a --- /dev/null +++ b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/full.md @@ -0,0 +1,268 @@ +# Text2Scene: Text-driven Indoor Scene Stylization with Part-aware Details + +Inwoo Hwang $^{1}$ , Hyeonwoo Kim $^{1}$ , and Young Min Kim $^{1,2}$ + +$^{1}$ Department of Electrical and Computer Engineering, Seoul National University + +$^{2}$ Interdisciplinary Program in Artificial Intelligence and INMC, Seoul National University + +![](images/673ec43ade49462211c91ca5548d95ad0d40a65ad68d21ffbaf9013e67d5e0c8.jpg) +Figure 1. Stylization results of diverse objects. Text2Scene creates more realistic and part-aware textures for various categories of 3D objects without any dedicated datasets for training, which can constitute high-quality virtual scenes. + +# Abstract + +We propose Text2Scene, a method to automatically create realistic textures for virtual scenes composed of multiple objects. Guided by a reference image and text descriptions, our pipeline adds detailed texture on labeled 3D geometries in the room such that the generated colors respect the hierarchical structure or semantic parts that are often composed of similar materials. Instead of applying flat stylization on the entire scene at a single step, we obtain weak semantic cues from geometric segmentation, which are further clarified by assigning initial colors to segmented parts. Then we add texture details for individual objects such that their projections on image space exhibit feature embedding aligned with the embedding of the input. The decomposition makes the entire pipeline tractable to a moderate amount of computation resources and memory. As our framework utilizes the existing resources of image and text embedding, it does not require dedicated datasets with high-quality textures designed by skillful artists. To the best of our knowledge, it is the first practical and scalable approach that can + +create detailed and realistic textures of the desired style that maintain structural context for scenes with multiple objects. + +# 1. Introduction + +Virtual spaces provide an immersive experience for metaverse, films, or games. With increasing demands for virtual environments, various applications seek practical methods to create realistic 3D scenes with high-quality textures. Currently, skillful artists need to manually create 3D assets and accompanying textures with careful parameterization, which is not scalable enough to account for the diverse content the industry is heading for. Scenes can also be populated with existing 3D database models or created with recent shape-generation approaches using data-driven methods [33, 54]. However, most of them lack texture information or are limited to simple coloring. + +To build realistic content, we need fine details containing the artistic nuances of styles that obey the implicit correlations with geometric shapes and semantic structure. Recent works provide methods to color a single object with the help + +![](images/d89b9b0ec45bb56b67daa99846fb8e3566136869c74ef85896fc6368e0d7d6a6.jpg) +Figure 2. Our scene stylization results. Given a target image $I_{t}$ and the style text $t_{s}$ , Text2Scene can produce the stylized results for the entire scene. + +![](images/7197f56ead959c7bf1a5ce7820d79c6ba6eb7f658d9df5fb88d77e817f8e27fe.jpg) + +of differentiable rendering [5], but often they are limited to single texture or blurred boundaries [6,21,27,29,52]. More importantly, the 3D objects are often textured in isolation, and only limited attempts exist to add visual appearances for large-scale scenes with multiple objects [14, 15, 19]. The biggest challenge is adding consistent style for an entire scene, but still accounting for the boundaries of different materials due to the functional and semantic relationship between parts, as observed within real-world scenes. + +Our proposed Text2Scene adds plausible texture details on 3D scenes without explicit part labels or large-scale data with complex texturing. We take inspiration from abundant 3D shape and image datasets and decompose the problem into sub-parts such that the entire scene can be processed with a commodity memory and computation. Given scenes of multiple objects of 3D mesh geometry, we separately handle walls and individual objects. Specifically, the stylization of walls is formulated as texture retrieval, and the objects are initialized with base colors. From the base color assignment, we can deduce the part-level relationship for stylization and further refine them in later stages, such that their rendered images are close to the input text within the joint embedding space of foundational models. + +Our coarse-to-fine strategy keeps the problem tractable yet generates high-quality texture with clean part boundaries. We first create segments of input mesh such that the segment boundaries align with low-level geometric cues. Then we start with the simplified problem of assigning a color per segment. Interestingly, the prior obtained from large-scale image datasets assign similar colors for the parts with similar textures, reflecting the semantic context or + +symmetry as shown in Figure 1. We add the detailed texture on individual objects as an additional perturbation on the assigned base colors by enforcing constraints on the image features of their projections. The additional perturbations are high-frequency neural fields added to the base color. + +In summary, Text2Scene is a new method that + +- can easily generate realistic texture colors of the scene with the desired style provided by text or an image; +- can add detailed texture that respects the semantic part boundaries of individual objects; and +- can process the entire scene without a large amount of textured 3D scenes or an extensive memory footprint. + +We expect the proposed approach to enable everyday users to quickly populate virtual scenes of their choices, and enjoy the possibility of next-generation technology with high-quality visual renderings. + +# 2. Related works + +3D Shape Understanding and Segmentation Our proposed method creates a realistic texture that abides by implicit rules of the material composition of different object parts. Different materials or textures are often assigned to different objects' functional parts, whose boundaries align with geometric feature lines. A handful of previous works find segments that constitute a 3D model with geometric information [7, 12, 20, 34], while others train 3D features to distinguish part labels provided in datasets [9, 30, 38]. Recently, PartGlot [24] suggested discovering part segments + +![](images/8ef56bbfb6fe35971948f03ef924d56f162c09788638e135d2c0947895410c02.jpg) +Figure 3. Overall pipeline of Text2Scene. Given a 3D scene $S$ with optional text description, we generate textures guided by a target image $I_{t}$ and the appearance style provided as a text $t_s$ . We first stylize the structure by texture retrieval, and each object is pre-processed for part discovery for stylization. Then, we assign base colors to object parts and add local details for each object with the designated LNSF. + +from language references, which should contain functional information. However, the geometric or functional distinction does not always clarify the texture boundaries, with possible additional diversity originating from the designers' choice. The intricate rules are multi-modal distributions composed of a mixture of discrete part assignments and continuous texture details, whose results are contained within visual datasets of scenes. + +Neural Fields and Texture Traditional texture atlas of 3D geometry is represented as a mapping from a planar image to a manifold embedded in 3D space and involves complex parameterization. With the increasing popularity of neural representation, TextureFields [31] represented texture as a mapping from 3D surface points to RGB color values without separate parameterization. It is deeply connected with graphics pipelines that adapt coordinate-based functions to depict SDF shapes [26, 32, 45] or novel-view synthesis using implicit volumes [28]. Various works show neural implicit representation is highly flexible and free from domain structure or resolution [2, 22, 25, 37, 49]. Recently, Text2Mesh [27] generated the deformation and coloring of input mesh with neural fields guided by the joint embedding of rendered image and input text. Interestingly, the generated mesh deformation and coloring also contain semantic part information. Our Text2Scene framework further extends the ability and observes distinct part boundaries, which could not be captured in previous works. We also increase the reality of resulting scenes with high-frequency details. We encode the input to the neural network with a high-order basis of intrinsic features [23] and apply the coarse-to-fine strategy as shown with geometric details of Yifan et al. [51]. + +Text-Driven 3D Stylization Recently neural networks trained with large-scale images and texts demonstrated powerful performance in many tasks with their extensive representation power. Here we primarily focus on 3D stylization attempts using image features. CLIP [39] learns latent space with a large amount of image text pairs, and the additional text input allows semantic manipulation for various generative tasks, including images [8, 35, 40–42], videos [3, 17], motions [46, 47], and 3D assets [18, 21, 36, 43]. However, for text-driven 3D stylization, it is still difficult to define clean texture boundaries and correlation with instance-level subtleties [6, 27, 29] or focus only on the specific type, such as human [16, 53]. As another way to stylize a 3D scene [4, 50], Yeh et al. [50] matches the input image features with CAD input using differentiable material graphs [44]. However, the representation is inherently limited to a combination of material libraries and cannot handle non-homogeneous details such as painting. On the other hand, our approach can generate texture beyond the repetitive low-level patterns of input materials and introduce part-aware texture with fine details that maintain geometric and semantic consistency. + +# 3. Method + +We stylize a 3D indoor scene without sophisticated techniques or software tools for 3D modeling or texturing. The input 3D scene $\mathcal{S} = \{\mathcal{W},\mathcal{O}\}$ is a set of structure components $\mathcal{W}$ and a set of objects $\mathcal{O}$ . The structural components $\mathcal{W}$ are walls, ceilings, and floors, whereas the objects $\mathcal{O} = \{M_i\}$ are the 3D mesh models $M_{i}$ . We assume that all the components have their corresponding class labels and optionally have text descriptions. + +The desired color distribution is provided as a target image $I_{t}$ , which could be retrieved from the web. Also, a specific appearance style description $t_s$ is provided as input to enforce style consistency among a set of objects $\mathcal{O}$ . The color distribution compares the color histogram of the input target image $I_{t}$ against the rendering of the current stylized scene $I$ . Specifically, the histogram loss $\mathcal{L}_{\mathrm{hist}}$ is defined as below + +$$ +\mathcal {L} _ {\text {h i s t}} (I, I _ {t}) = \| H ^ {1 / 2} (I) - H ^ {1 / 2} (I _ {t}) \| _ {2}, \tag {1} +$$ + +where $H(\cdot)$ indicates differentiable color histogram operator [1]. Also, we augment the losses derived from the joint embedding of text and images [39] by generating text descriptions $T$ of the context using semantic labels and comparing them against the rendered image $I$ . If we denote the pre-trained encoder for image and text as $E_{1}$ and $E_{2}$ , respectively, the CLIP similarity loss is defined as + +$$ +\mathcal {L} _ {\mathrm {c l i p}} (I, T) = 1 - \operatorname {s i m} \left(E _ {1} (I), E _ {2} (T)\right), \tag {2} +$$ + +where $sim(\mathbf{x},\mathbf{y}) = \frac{\mathbf{x}^\top\mathbf{y}}{\|\mathbf{x}\|_2\|\mathbf{y}\|_2}$ is the cosine similarity. + +The overall pipeline is described in Fig. 3. We obtain the texture for the structure $\mathcal{W}$ by texture retrieval, which is described in Sec. 3.1. The objects are stylized with additional decomposition to respect local part boundaries and, simultaneously, to practically handle multiple entities with details (Sec. 3.2). + +# 3.1. Structure Stylization + +We assign one coherent texture per structural element of $\mathcal{W}$ . Compared to objects, the structural elements, such as walls or ceilings, are of simple planar geometry, and their textures are not heavily dependent on the relationships between different functional parts. For structural elements, it suffices to pick texture from the texture set of an existing material library of MATch [44], which contains homogeneous material. If we instead utilize visual features or CLIP embeddings for them, the resulting stylization exhibits undesired artifacts of various sizes instead of constant patterns, as shown in the supplementary. + +We randomly initialize the texture from the texture set and render the structure image $I_{s}$ , a bare room only containing the structural elements $\mathcal{W}$ without objects. Then the materials are compared to the target image $I_{t}$ for the histogram losses of Equation (1). The additional text prompt $T_{s}$ is given as 'a structure of a room' to provide the context with $\mathcal{L}_{\mathrm{clip}}(I_s,T_s)$ . In summary, the texture of the structural element is retrieved to have the lowest score on the following criteria: + +$$ +\mathcal {L} _ {\text {h i s t}} \left(I _ {s}, I _ {t}\right) + \lambda_ {1} \cdot \mathcal {L} _ {\text {c l i p}} \left(I _ {s}, T _ {s}\right). \tag {3} +$$ + +![](images/abc061eab5ee066c13372ce6e6b555c85487678e8f03a3a6c41c4a3c90b0863b.jpg) +Figure 4. Overall pipeline of part discovery. We discover parts for object stylization from the super-segments of 3D object mesh. Given a 3D object mesh $M_{i}$ with segments $\{s_{ik}^{l}\}$ at $l^{\mathrm{th}}$ iteration, we assign a color $c(s_{ik}^{l})$ per segment and generate a graph $\mathcal{G}_i^l$ . The pair of neighboring nodes is merged if the distance between assigned colors $c(s_{ik}^{l})$ is within a threshold, then, move to $(l + 1)^{\mathrm{th}}$ iteration. + +# 3.2. Object Stylization + +Object stylization involves understanding the semantic structure hidden behind the mesh representation. As a preprocessing, we first subdivide individual objects $M_{i}$ in $\mathcal{O}$ into part segments $\{s_{ik}\}$ as described in Sec. 3.2.1. Then the scene is stylized in two steps. First, we assign base colors into individual parts to minimize the style loss for the entire scene $S = \{\mathcal{W},\mathcal{O}\}$ (Sec. 3.2.2). Here we are optimizing for the discrete set of colors assigned to subdivided parts obtained from the pre-processing. Then the textures for individual objects are further optimized to generate fine details (Sec. 3.2.3). + +# 3.2.1 Part Discovery for Object Stylization + +We first decompose individual objects into parts such that each part is composed of the same material or texture. The distinctive part boundaries are critical in providing semantic consistency and, therefore, perceptual realism toward the scene. Similar to 3D part segmentation methods, we first find super-segments based on geometric features which provide the granularity to define textural parts. For a given 3D object mesh $M_{i}$ , the initial segments $\{s_{ik}^{0}\}$ are the decomposition applying the method by Katz et al. [20]. The decomposition is designed to be an over-segmentation of our aim. We generate a graph $\mathcal{G}_i^0$ where each node is the segment and edges connect neighboring segments. Then we incrementally merge segments that belong to the same texture until convergence. Note that, our part discovery method operates robustly regardless of the initial composition, but we use [20] which preserves the original geometry and details. + +The challenge here is that, unlike semantic segmentation approaches, no large-scale public dataset exists that provides the ground truth for 'texture similarity' as the segmentation labels. We find a supervision signal from the large-scale pre-trained model, and create a simple text prompt $T_{i,c}$ using the class name, such as a bed or a chair. At $l^{\mathrm{th}}$ iteration, we assign a color $c(s_{ik}^{l})$ to each segment $\{s_{ik}^{l}\}$ , which is optimized to minimize the distance + +between the rendered images $I_{M_i}$ of multiple viewpoints and the text $T_{i,c}$ in the joint embedding space of CLIP, or $\mathcal{L}_{\mathrm{clip}}(I_{M_i},T_{i,c})$ as defined in Equation (2). If the resulting colors assigned to two adjacent segments are similar, the two parts are likely to be the parts with the same texture source. Therefore we merge the two segments for the next iteration $\{s_{ik}^{l + 1}\}$ . In particular, we merge segments if the assigned color has a distance of less than a threshold $\lambda_{th}$ in the CIE color space which is known to be related to human perception. Note that, while the initial color is gray for the assignment, merging segments happen with the optimized color. We repeat the process until the number of segments does not decrease anymore and empirically found that it usually converges within 2-3 steps. The overall pipeline for part discovery is also described in Fig. 4. + +# 3.2.2 Part-level Base Color Assignment + +After the objects $M_{i} \in \mathcal{O}$ are decomposed into parts $\{s_{ik}\}$ , we assign a solid color per part, namely $c(s_{ik})$ . The base color assignment handles a low-dimensional optimization space with a coarse set of discrete parts but still observes the holistic distribution of the entire scene. Then the base color is combined with the output of designed neural style fields in Sec. 3.2.3, which generate high-frequency local texture. + +We optimize the base color using the combined loss as before, $\mathcal{L}_{\mathrm{color,scene}} + \mathcal{L}_{\mathrm{clip,scene}}$ . The color loss again considers the similarity with the target image $\mathcal{L}_{\mathrm{hist}}(I, I_t)$ . The scene being optimized $I$ is rendered with the stylized structure $\mathcal{W}$ and the current estimates of the base colors. The clip loss considers both individual objects and the global scene and is calculated as the sum of object clip loss and global clip loss. + +$$ +\mathcal {L} _ {\text {c l i p , s c e n e}} = \lambda_ {2} \cdot \sum_ {i} \mathcal {L} _ {\text {c l i p}} \left(I _ {M _ {i}}, T _ {i}\right) + \lambda_ {3} \cdot \mathcal {L} _ {\text {c l i p}} (I, T). \tag {4} +$$ + +Unlike Sec. 3.2.1, text description $T_{i}$ for object $M_{i}$ could be a simple text prompt using the class name, or a detailed text prompt based on user choice. We render individual objects from various angles $I_{M_i}$ and compare them with the text description $T_{i}$ . The embedding for the scene is also compared against the text embedding $T$ represents the type of scene, for example, 'a bedroom'. By jointly applying the loss, the base color $c(s_{ik})$ is selected as a representative color that harmonizes nicely with the global context. + +# 3.2.3 Detailed Stylization + +The base color is combined with additional details regressed from a neural network to express the detailed local texture. We define local neural style field (LNSF) for each object, which generates the local textures added to the base color. The color for point $p$ is defined by the following equation, + +$$ +c \left(s _ {p}\right) + \alpha \cdot \mathcal {F} _ {i} (\gamma (p), \phi (p), s _ {p}). \tag {5} +$$ + +We train a LNSF $\mathcal{F}_i$ per object, which outputs the color to be added to the base color $c(s_p)$ , where $s_p$ indicates the part segment id. The color range $\alpha$ maintains the final color to be similar to the base color. $\gamma(\cdot)$ is the positional encoding of the $xyz$ coordinate to capture high-frequency details, and $\phi(\cdot)$ represents the coefficients of the eigenfunctions for the intrinsic geometry using the Laplace-Beltrami operator. Therefore the object-specific neural fields respect the part boundaries and local geometric details. + +LNSF for each object are trained with additional style context $\mathcal{L}_{\mathrm{clip}}(I_{M_i},T_i^+)$ . The text prompt $T_{i}^{+}$ augments the object description of $T_{i}$ with appearance style description $t_s$ , such as 'minimal style' or 'Mid-Century style'. By enforcing the same appearance style description, we can weakly bind the styles of individual objects in the same scene while optimizing separately. We could also use $T_{i}^{+}$ instead of $T_{i}$ for optimizing the base color, and it shows slightly better results. + +Part-aware Geometric Deformation Even though our main focus is to add colors to objects with local part-aware details, we could also concurrently produce part-aware geometric deformation with the same architecture. We can slightly change the LNSF $\mathcal{F}$ to have two branches of output that estimate color and displacement. For each point on the surface, we assign color by Eq. 5 and adjust displacement along the vertex normal direction. To learn the effective deformation, we add geometric loss $\mathcal{L}_{\mathrm{clip}}(I_{M_i}^{geo}, T_i^+)$ , where $I_{M_i}^{geo}$ is an image rendering textureless geometry as [27]. + +Rendering and Implementation Details To render each object, the individual objects $M_{i}$ are scaled to fit a unit box. The predicted color of each vertex allows the entire mesh to be differentiable rendering through interpolation using [5]. We render individual objects with random augmented backgrounds (white, black, random Gaussian, chess board), which helps the pipeline focus on the foreground object [16, 18]. Inspired by [36], since CLIP has a bias for the canonical pose, we augment the text prompt with the azimuth angle of view, namely 'front view of', 'side view of' or 'back view of'. Finally, in Sec 3.2.3, random perspective transformation and random crop boost learning local details. To render a scene, we randomly sampled from predefined 20 camera poses to evenly cover the entire scene $S$ . + +# 4. Experiments + +Our stylization is first evaluated for individual objects in Sec. 4.1. We evaluate the quality of textured object meshes and also assess that our pipeline can discover part segments to assign realistic stylization. Then we demonstrate the stylization results of room-scale scenes with multiple objects in Sec. 4.2. + +![](images/aa4b62ba5d2e5162222124dd0f8c2b35d9144a639b34d54e50ae5f6716ada7a9.jpg) +Figure 5. Visual comparison with baselines. Our method utilizes the found part information and generates detailed textures through the designed network in combination with a coarse-to-fine stylization scheme. As a result, it generates a globally harmonious 3D style that clearly classifies part segments. We can handle diverse text such as simple categories or styles, detailed descriptions, or text that includes different two object categories. + +# 4.1. Object Stylization + +We show that our method can stylize various objects and produce realistic 3D assets to populate virtual scenes. We use object meshes of various types and sizes, collected from Turbo Squid [48], 3D-FUTURE [13], and Amazon Berkeley Objects [10]. For detailed stylization, we subdivide the object into an average of 119769 faces and 60437 vertices. For large general objects such as beds, sofas, tables, etc., we use class labels and the text input for the specific style as 'a [class label], [specific] style'. However, we need a detailed explanation for small objects that cannot be explained only by class labels, such as books. For these objects, we configure detailed text individually, for example, 'a book of Harry Potter and the Sorcerer's Stone'. + +Figure 5 shows the renderings of stylized results. We render objects in Blender [11] with a fixed lighting setup. + +The input mesh and the text are also provided in the left column. Since our method utilizes the obtained part information and coarse-to-fine stylization scheme with a two-step approach, it creates a more realistic texture with clear boundaries for each part of the 3D assets. Text2Mesh [27] is another text-driven 3D stylization approach, which adds the RGB color and deformation fields on the vertices of mesh. While it augments the detailed variations on the input mesh, the deformation map can occasionally introduce undesired artifacts and the part boundaries are only roughly estimated. We also provide the results of an ablated version that directly uses LNSF without a two-step stylization scheme. Our part discovery module and coarse-to-fine stylization scheme play a critical role in producing realistic assets with high-quality stylization. Figure 1 contains more results on diverse objects. + +![](images/cbbc634a8d973165b74fb6edf98a0bfc222b7fbe8a748189cbb36b7bb5f1b2c7.jpg) +Figure 6. Results of user study for object stylization + +Because stylization is a subjective task, and no ground truth or metrics exist, we conduct a user study for quantitative evaluation. We ask 98 users to rate the quality of generated outputs on a scale of 1 (worst) to 5 (best) in response to the following questions: (Q1) 'How natural are the output results?', (Q2) 'How well does the output contain text information?', and (Q3) 'How well does the output reflect part information?' Figure 6 contains the mean and the standard deviation of the scores. Our method outperforms competing methods in all aspects. Therefore we conclude that our method generates a realistic texture that abides by input text description and semantic part information. + +![](images/12fe3abb18708b9695db216600524072029067d8b93f4d94ee2709be27ce5338.jpg) +Figure 7. With a slightly modified LNSF $\mathcal{F}$ , we can simultaneously generate the texture with both color and displacement fields, and all of the generated style information respects the part information discovered. + +We also provide a parallel comparison against Text2Mesh by incorporating deformation fields with our approach. Our original method preserves the original shape of the mesh, while Text2Mesh deforms vertices along the normal direction in addition to generating texture. We can make a similar version of our LNSF and produce additional deformation as described in Sec. 3.2.3. Figure 7 shows the deformed results from the source mesh compared to Text2Mesh. Since our approach explicitly considers the part information with the two-stage approach, our results with deformation fields also respect different semantic parts of the object. + +Part Discovery As a side product of object stylization, we can discover different parts with distinguished boundaries + +![](images/ade3d4c4ae2d3e45c72074051f8a85e117a2a6262b612c50c3bde8156a84282d.jpg) +Figure 8. The robustness against initial super-segmentation. The super-segments of the top and the bottom rows are generated by [20] and [7], respectively (left). Starting from the initial color assignment (middle), our approach stably finds part decomposition to assign different base colors and therefore different texture information (right). The text a chair is used. + +that can guide a realistic color assignment (Sec. 3.2.1). Our part discovery combines geometric super-segments and implicit clues from image-text embeddings. The initial super-segments guide the algorithm to follow geometric feature lines, and the final results stably find part information despite different initializations. Figure 8 shows examples initialized with Katz et al. [20] (top) and BSP-Net [7] (bottom). Even for input mesh with bad topology or different initialization, our iterative part discovery quickly finds visually coherent parts without any training with segmentation labels. + +# 4.2. Scene Stylization + +Now we demonstrate that our text2scene framework can quickly generate a realistic texture for a room with multiple objects. Recall that the 3D scene $S = \{\mathcal{W}, \mathcal{O}\}$ is composed of the structure components $\mathcal{W}$ and a set of objects $\mathcal{O}$ . We use the same objects as described in Sec. 4.1 and arrange them to constitute scenes. As there is no existing dataset composite of complete object meshes with labels, we built a total of four scenes: two bedrooms and two living rooms. Each scene contains an average of 20 objects of various sizes and classes. Additionally, we provide a target image $I_{t}$ for the color distribution, and a text prompt describing the desired style $t_{s}$ . While it can be daunting to define the desired style for the entire scene, images and texts can provide a simple way to deliver the information. + +Figure 2 shows stylized results of the same geometry, but observing various target images and style prompts. The generated textures respect the semantic labels of various furniture and different parts and contain localized diverse details. This is in contrast to many prior stylization methods where the fine perturbations are spread throughout the scenes. We also show scenes with different types of rooms containing different objects, but stylized based on the same + +target image in Fig. 9. Additional results of various input configurations are available in the supplementary material. Note that our target image $I_{t}$ does not need to restrict as an indoor image, and can be replaced such as natural photographs. And by changing random seeds, diverse results can be obtained from the same input conditions. Also, since we stylize the entire object, we can easily edit the 3D scene through object relocation. These results can also be found in the supplementary. + +![](images/ec2cf0e19f6b4d261fead6170adddba1ab3e8a60072081e13a764f68aff450ec.jpg) +Livingroom +Figure 9. Result for the different arrangement of objects. We used target image number 3 in Fig. 2 in both spaces. + +![](images/8b341e11d77ac96fd8cf200f5c16f769cd739ac944e9159906297bcad4d3a28e.jpg) +Bedroom + +
-retrieval-hist,glo-detailOurs
(Q1): Realistic2.23(±0.48)3.68(±0.49)4.02(±0.49)4.22(±0.42)
(Q2): Color2.09(±0.48)2.63(±0.59)3.95(±0.47)3.86(±0.49)
+ +Table 1. Results of user study for scene stylization + +![](images/61b09640015035873ba64b4676b91d1078418a9363d8611416900b2f9f2955aa.jpg) +- retrieval +Figure 10. Ablation results for the scene. We used target image number 1 in Fig. 2 in all spaces. + +![](images/ff8c3cad378d18f90b5d540424e739799571d994195a70c3200b62b0e76677f8.jpg) +-hist,glo + +![](images/476364c798fe33cf99a0894c70cbd6bdf9721e35f03b1455e6be1a4f66ad7349.jpg) +-detail + +![](images/2e2a5d164df5e0d1b2770ab7b9aeaf6bcdf063b601e87e9b44acf4a47ad02076.jpg) +Ours + +We also provide users' evaluation of the quality of our stylization in Table 1. Users evaluate the results by answering the following questions: (Q1) 'How realistic is the output result?' (Q2) 'How similar is the color distribution of the scene to the given images?' Users assess the overall quality of the outputs, and how well they match the target image. Since there are no previous works that use the same setting, we compare the results of ablated versions: -retrieval replaces the separate texture retrieval module (Sec. 3.1) with a stylization network of objects for structure components; -hist,glo removes the color loss and the global clip loss for the base color assignment (Sec. 3.2.2); and -detail removes the detailed stylization step for objects (Sec. 3.2.3). The responses of (Q1) indicate that Text2Scene generates high-quality scenes and each of the components plays a crucial role to achieve reality. The effect of texture retrieval is the most prominent. The color distribution results (Q2) indicate that the base color assignment is critical. The added local details are more important for the realism + +of the results. Figure 10 shows exemplar images of the ablated versions used for the user study. + +GPU Cost and Scalability Text2Scene first assigns base colors to discovered parts for all objects in the scene, and generates details of objects individually. The cardinality for the base color assignment is only a few hundred and it does not require much memory and allows us to consider the whole scene while expanding the scale. The most memory-intensive process is the detail generation using a neural network, which only processes a single object at a time. Therefore the entire process can be trained only on a single 11 GB GPU, making it an accessible tool for casual users. In a single GPU, the base color allocation of the entire space takes 5 hours, and learning the details for each object takes 10 minutes. + +Limitations While our approach results in scene stylization, the pipeline separately handles individual objects after the base color assignment. This is a practical choice for scalability but may lack an understanding of the context of the entire space. Instead, we rely on the text description to weakly bind the objects into a similar style. We can design the pipeline to receive an additional input with a texture map or a lightweight network and extend our model to better observe the holistic scene context within a limited GPU memory. Also, our pipeline requires a class label or optional text description per object, which can be further automated. + +# 5. Conclusions + +We introduce Text2Scene, a novel framework to generate a texture for 3D scenes. Our hierarchical framework can handle a variety of objects, including highly detailed textures for objects such as book or paintings. By leveraging the representation power of pre-trained CLIP, the framework does not require any 3D datasets with texture or part annotation. Given the 3D mesh models and the class labels or text descriptions of objects, our framework easily produces stylized results by picking a target image and a simple text description. We hope that Text2Scene to facilitate the automatic interior recommendation or realistic virtual space generation. + +Acknowledgements This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00208197) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub). Inwoo Hwang is supported by Hyundai Motor Chung Mong-Koo Foundation. Young Min Kim is the corresponding author. + +# References + +[1] Mahmoud Affifi, Marcus A. Brubaker, and Michael S. Brown. Histogram: Controlling colors of gan-generated and real images via color histograms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 4 +[2] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290–6301, June 2022. 3 +[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3 +[4] Kang Chen, Kun Xu, Yizhou Yu, Tian-Yi Wang, and ShiMin Hu. Magic decorator: Automatic material suggestion for indoor digital scenes. ACM Trans. Graph., 34(6):232:1-232:11, Oct. 2015. 3 +[5] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 5 +[6] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3 +[7] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bspnet: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 7 +[8] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3 +[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. Proceedings of International Conference on Computer Vision (ICCV), 2019. 2 +[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 6 +[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6 +[12] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2 +[13] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d fur + +niture shape with texture. International Journal of Computer Vision, pages 1-25, 2021. 6 +[14] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 2 +[15] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2 +[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3, 5 +[17] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. CVPR, 2022. 3, 5 +[19] Jaebong Jeong, Janghun Jo, Sunghyun Cho, and Jaesik Park. 3d scene painting via semantic image synthesis. In Proceedings of the International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[20] Sagi Katz and Ayellet Tal. Hierarchical mesh decomposition using fuzzy clustering and cuts. ACM Transactions on Graphics, 22, 06 2003. 2, 4, 7 +[21] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, December 2022. 2, 3 +[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 3 +[23] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, 2022. 3 +[24] Juil Koo, Ian Huang, Panos Achlioptas, Leonidas J Guibas, and Minhyuk Sung. Partglot: Learning shape part segmentation from language reference games. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 3 +[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3 +[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. arXiv preprint arXiv:2112.03221, 2021. 2, 3, 5, 6 + +[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3 +[29] Shailesh Mishra and Jonathan Granskog. Clip-based neural neighbor style transfer for 3d assets. In arXiv, 2022. 2, 3 +[30] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2 +[31] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3 +[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation, 2019. 3 +[33] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1 +[34] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 3 +[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3, 5 +[37] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961, 2020. 3 +[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 4 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3 +[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arxiv:2208.12242, 2022. 3 + +[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 3 +[44] Liang Shi, Beichen Li, Milos Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: Differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6):1-15, Dec. 2020. 3, 4 +[45] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions, 2020. 3 +[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 3 +[47] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3 +[48] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2022.6 +[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2021. 3 +[50] Yu-Ying Yeh, Zhengqin Li, Yannick Hold-Geoffroy, Rui Zhu, Zexiang Xu, Miloš Hasan, Kalyan Sunkavalli, and Manmohan Chandraker. Photoscene: Photorealistic material and lighting transfer for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18562–18571, June 2022. 3 +[51] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2022. 3 +[52] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 2 +[53] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In ECCV, 2022. 3 +[54] Dongsu Zhang, Changwoon Choi, Inbum Park, and Young Min Kim. Probabilistic implicit scene completion. In International Conference on Learning Representations, 2022. 1 \ No newline at end of file diff --git a/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/images.zip b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..54aaa2d6350cdebc805ce2f9e20def21b663f2b3 --- /dev/null +++ b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:756f58858aaa00565c3898c0d2549111c599f0fb8e9eed04edf825875bc4d097 +size 698260 diff --git a/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/layout.json b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a0d4afd0005c21c15737a5435c5a685265c45605 --- /dev/null +++ b/2023/Text2Scene_ Text-Driven Indoor Scene Stylization With Part-Aware Details/layout.json @@ -0,0 +1,8102 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 523, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 523, + 120 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 523, + 120 + ], + "type": "text", + "content": "Text2Scene: Text-driven Indoor Scene Stylization with Part-aware Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "spans": [ + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "text", + "content": "Inwoo Hwang" + }, + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "text", + "content": ", Hyeonwoo Kim" + }, + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "text", + "content": ", and Young Min Kim" + }, + { + "bbox": [ + 156, + 142, + 436, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 159, + 459, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 159, + 459, + 171 + ], + "spans": [ + { + "bbox": [ + 149, + 159, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 159, + 459, + 171 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering, Seoul National University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 173, + 460, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 460, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 460, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 173, + 460, + 185 + ], + "type": "text", + "content": "Interdisciplinary Program in Artificial Intelligence and INMC, Seoul National University" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 213, + 545, + 414 + ], + "blocks": [ + { + "bbox": [ + 50, + 213, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 50, + 213, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 50, + 213, + 545, + 414 + ], + "type": "image", + "image_path": "673ec43ade49462211c91ca5548d95ad0d40a65ad68d21ffbaf9013e67d5e0c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 422, + 546, + 446 + ], + "lines": [ + { + "bbox": [ + 46, + 422, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 546, + 446 + ], + "type": "text", + "content": "Figure 1. Stylization results of diverse objects. Text2Scene creates more realistic and part-aware textures for various categories of 3D objects without any dedicated datasets for training, which can constitute high-quality virtual scenes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 455, + 192, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 455, + 192, + 468 + ], + "spans": [ + { + "bbox": [ + 143, + 455, + 192, + 468 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "We propose Text2Scene, a method to automatically create realistic textures for virtual scenes composed of multiple objects. Guided by a reference image and text descriptions, our pipeline adds detailed texture on labeled 3D geometries in the room such that the generated colors respect the hierarchical structure or semantic parts that are often composed of similar materials. Instead of applying flat stylization on the entire scene at a single step, we obtain weak semantic cues from geometric segmentation, which are further clarified by assigning initial colors to segmented parts. Then we add texture details for individual objects such that their projections on image space exhibit feature embedding aligned with the embedding of the input. The decomposition makes the entire pipeline tractable to a moderate amount of computation resources and memory. As our framework utilizes the existing resources of image and text embedding, it does not require dedicated datasets with high-quality textures designed by skillful artists. To the best of our knowledge, it is the first practical and scalable approach that can" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 456, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 545, + 480 + ], + "type": "text", + "content": "create detailed and realistic textures of the desired style that maintain structural context for scenes with multiple objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 502, + 387, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 387, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 387, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 665 + ], + "type": "text", + "content": "Virtual spaces provide an immersive experience for metaverse, films, or games. With increasing demands for virtual environments, various applications seek practical methods to create realistic 3D scenes with high-quality textures. Currently, skillful artists need to manually create 3D assets and accompanying textures with careful parameterization, which is not scalable enough to account for the diverse content the industry is heading for. Scenes can also be populated with existing 3D database models or created with recent shape-generation approaches using data-driven methods [33, 54]. However, most of them lack texture information or are limited to simple coloring." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "To build realistic content, we need fine details containing the artistic nuances of styles that obey the implicit correlations with geometric shapes and semantic structure. Recent works provide methods to color a single object with the help" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1890" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 324, + 299 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 324, + 299 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 324, + 299 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 324, + 299 + ], + "type": "image", + "image_path": "d89b9b0ec45bb56b67daa99846fb8e3566136869c74ef85896fc6368e0d7d6a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "type": "text", + "content": "Figure 2. Our scene stylization results. Given a target image " + }, + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "type": "text", + "content": " and the style text " + }, + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 46, + 303, + 546, + 324 + ], + "type": "text", + "content": ", Text2Scene can produce the stylized results for the entire scene." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 543, + 298 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 543, + 298 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 543, + 298 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 543, + 298 + ], + "type": "image", + "image_path": "7197f56ead959c7bf1a5ce7820d79c6ba6eb7f658d9df5fb88d77e817f8e27fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": "of differentiable rendering [5], but often they are limited to single texture or blurred boundaries [6,21,27,29,52]. More importantly, the 3D objects are often textured in isolation, and only limited attempts exist to add visual appearances for large-scale scenes with multiple objects [14, 15, 19]. The biggest challenge is adding consistent style for an entire scene, but still accounting for the boundaries of different materials due to the functional and semantic relationship between parts, as observed within real-world scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 617 + ], + "type": "text", + "content": "Our proposed Text2Scene adds plausible texture details on 3D scenes without explicit part labels or large-scale data with complex texturing. We take inspiration from abundant 3D shape and image datasets and decompose the problem into sub-parts such that the entire scene can be processed with a commodity memory and computation. Given scenes of multiple objects of 3D mesh geometry, we separately handle walls and individual objects. Specifically, the stylization of walls is formulated as texture retrieval, and the objects are initialized with base colors. From the base color assignment, we can deduce the part-level relationship for stylization and further refine them in later stages, such that their rendered images are close to the input text within the joint embedding space of foundational models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Our coarse-to-fine strategy keeps the problem tractable yet generates high-quality texture with clean part boundaries. We first create segments of input mesh such that the segment boundaries align with low-level geometric cues. Then we start with the simplified problem of assigning a color per segment. Interestingly, the prior obtained from large-scale image datasets assign similar colors for the parts with similar textures, reflecting the semantic context or" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 341, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 400 + ], + "type": "text", + "content": "symmetry as shown in Figure 1. We add the detailed texture on individual objects as an additional perturbation on the assigned base colors by enforcing constraints on the image features of their projections. The additional perturbations are high-frequency neural fields added to the base color." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 401, + 504, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 504, + 412 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 504, + 412 + ], + "type": "text", + "content": "In summary, Text2Scene is a new method that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 420, + 545, + 507 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 420, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 420, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 420, + 545, + 444 + ], + "type": "text", + "content": "- can easily generate realistic texture colors of the scene with the desired style provided by text or an image;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 452, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 452, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 452, + 545, + 475 + ], + "type": "text", + "content": "- can add detailed texture that respects the semantic part boundaries of individual objects; and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 483, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 545, + 507 + ], + "type": "text", + "content": "- can process the entire scene without a large amount of textured 3D scenes or an extensive memory footprint." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 515, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 563 + ], + "type": "text", + "content": "We expect the proposed approach to enable everyday users to quickly populate virtual scenes of their choices, and enjoy the possibility of next-generation technology with high-quality visual renderings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 574, + 394, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 394, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 394, + 586 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "3D Shape Understanding and Segmentation Our proposed method creates a realistic texture that abides by implicit rules of the material composition of different object parts. Different materials or textures are often assigned to different objects' functional parts, whose boundaries align with geometric feature lines. A handful of previous works find segments that constitute a 3D model with geometric information [7, 12, 20, 34], while others train 3D features to distinguish part labels provided in datasets [9, 30, 38]. Recently, PartGlot [24] suggested discovering part segments" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1891" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 261 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 261 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 261 + ], + "type": "image", + "image_path": "8ef56bbfb6fe35971948f03ef924d56f162c09788638e135d2c0947895410c02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": "Figure 3. Overall pipeline of Text2Scene. Given a 3D scene " + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": " with optional text description, we generate textures guided by a target image " + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": " and the appearance style provided as a text " + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 268, + 547, + 304 + ], + "type": "text", + "content": ". We first stylize the structure by texture retrieval, and each object is pre-processed for part discovery for stylization. Then, we assign base colors to object parts and add local details for each object with the designated LNSF." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 314, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 289, + 410 + ], + "type": "text", + "content": "from language references, which should contain functional information. However, the geometric or functional distinction does not always clarify the texture boundaries, with possible additional diversity originating from the designers' choice. The intricate rules are multi-modal distributions composed of a mixture of discrete part assignments and continuous texture details, whose results are contained within visual datasets of scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "Neural Fields and Texture Traditional texture atlas of 3D geometry is represented as a mapping from a planar image to a manifold embedded in 3D space and involves complex parameterization. With the increasing popularity of neural representation, TextureFields [31] represented texture as a mapping from 3D surface points to RGB color values without separate parameterization. It is deeply connected with graphics pipelines that adapt coordinate-based functions to depict SDF shapes [26, 32, 45] or novel-view synthesis using implicit volumes [28]. Various works show neural implicit representation is highly flexible and free from domain structure or resolution [2, 22, 25, 37, 49]. Recently, Text2Mesh [27] generated the deformation and coloring of input mesh with neural fields guided by the joint embedding of rendered image and input text. Interestingly, the generated mesh deformation and coloring also contain semantic part information. Our Text2Scene framework further extends the ability and observes distinct part boundaries, which could not be captured in previous works. We also increase the reality of resulting scenes with high-frequency details. We encode the input to the neural network with a high-order basis of intrinsic features [23] and apply the coarse-to-fine strategy as shown with geometric details of Yifan et al. [51]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 578 + ], + "type": "text", + "content": "Text-Driven 3D Stylization Recently neural networks trained with large-scale images and texts demonstrated powerful performance in many tasks with their extensive representation power. Here we primarily focus on 3D stylization attempts using image features. CLIP [39] learns latent space with a large amount of image text pairs, and the additional text input allows semantic manipulation for various generative tasks, including images [8, 35, 40–42], videos [3, 17], motions [46, 47], and 3D assets [18, 21, 36, 43]. However, for text-driven 3D stylization, it is still difficult to define clean texture boundaries and correlation with instance-level subtleties [6, 27, 29] or focus only on the specific type, such as human [16, 53]. As another way to stylize a 3D scene [4, 50], Yeh et al. [50] matches the input image features with CAD input using differentiable material graphs [44]. However, the representation is inherently limited to a combination of material libraries and cannot handle non-homogeneous details such as painting. On the other hand, our approach can generate texture beyond the repetitive low-level patterns of input materials and introduce part-aware texture with fine details that maintain geometric and semantic consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 594, + 362, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 362, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 362, + 608 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We stylize a 3D indoor scene without sophisticated techniques or software tools for 3D modeling or texturing. The input 3D scene " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\mathcal{W},\\mathcal{O}\\}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " is a set of structure components " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and a set of objects " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". The structural components " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " are walls, ceilings, and floors, whereas the objects " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{M_i\\}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " are the 3D mesh models " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". We assume that all the components have their corresponding class labels and optionally have text descriptions." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1892" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "The desired color distribution is provided as a target image " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ", which could be retrieved from the web. Also, a specific appearance style description " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " is provided as input to enforce style consistency among a set of objects " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ". The color distribution compares the color histogram of the input target image " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " against the rendering of the current stylized scene " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ". Specifically, the histogram loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hist}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " is defined as below" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 178, + 287, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 178, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 85, + 178, + 287, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h i s t}} (I, I _ {t}) = \\| H ^ {1 / 2} (I) - H ^ {1 / 2} (I _ {t}) \\| _ {2}, \\tag {1}", + "image_path": "f24a2fae95183354c526df0c084195149085105c72034975b557541f32774e4b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "content": " indicates differentiable color histogram operator [1]. Also, we augment the losses derived from the joint embedding of text and images [39] by generating text descriptions " + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "content": " of the context using semantic labels and comparing them against the rendered image " + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "content": ". If we denote the pre-trained encoder for image and text as " + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 46, + 203, + 289, + 288 + ], + "type": "text", + "content": ", respectively, the CLIP similarity loss is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 298, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 298, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 83, + 298, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l i p}} (I, T) = 1 - \\operatorname {s i m} \\left(E _ {1} (I), E _ {2} (T)\\right), \\tag {2}", + "image_path": "059d01aabbd8fdfb83b38189d0fc3f01319b851a00e6db82cc573e184049c43f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 322, + 274, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 274, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 274, + 349 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 322, + 274, + 349 + ], + "type": "inline_equation", + "content": "sim(\\mathbf{x},\\mathbf{y}) = \\frac{\\mathbf{x}^\\top\\mathbf{y}}{\\|\\mathbf{x}\\|_2\\|\\mathbf{y}\\|_2}" + }, + { + "bbox": [ + 47, + 322, + 274, + 349 + ], + "type": "text", + "content": " is the cosine similarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 420 + ], + "type": "text", + "content": "The overall pipeline is described in Fig. 3. We obtain the texture for the structure " + }, + { + "bbox": [ + 46, + 348, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 348, + 287, + 420 + ], + "type": "text", + "content": " by texture retrieval, which is described in Sec. 3.1. The objects are stylized with additional decomposition to respect local part boundaries and, simultaneously, to practically handle multiple entities with details (Sec. 3.2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 430, + 167, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 167, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 167, + 442 + ], + "type": "text", + "content": "3.1. Structure Stylization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": "We assign one coherent texture per structural element of " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": ". Compared to objects, the structural elements, such as walls or ceilings, are of simple planar geometry, and their textures are not heavily dependent on the relationships between different functional parts. For structural elements, it suffices to pick texture from the texture set of an existing material library of MATch [44], which contains homogeneous material. If we instead utilize visual features or CLIP embeddings for them, the resulting stylization exhibits undesired artifacts of various sizes instead of constant patterns, as shown in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "content": "We randomly initialize the texture from the texture set and render the structure image " + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "content": ", a bare room only containing the structural elements " + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "content": " without objects. Then the materials are compared to the target image " + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "content": " for the histogram losses of Equation (1). The additional text prompt " + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "inline_equation", + "content": "T_{s}" + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "content": " is given as 'a structure of a room' to provide the context with " + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{clip}}(I_s,T_s)" + }, + { + "bbox": [ + 46, + 582, + 288, + 689 + ], + "type": "text", + "content": ". In summary, the texture of the structural element is retrieved to have the lowest score on the following criteria:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 701, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 99, + 701, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h i s t}} \\left(I _ {s}, I _ {t}\\right) + \\lambda_ {1} \\cdot \\mathcal {L} _ {\\text {c l i p}} \\left(I _ {s}, T _ {s}\\right). \\tag {3}", + "image_path": "71a6f802311fec7940734a061497c76f982459770fbf95580689445da4454d0d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 547, + 140 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 547, + 140 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 547, + 140 + ], + "type": "image", + "image_path": "abc061eab5ee066c13372ce6e6b555c85487678e8f03a3a6c41c4a3c90b0863b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": "Figure 4. Overall pipeline of part discovery. We discover parts for object stylization from the super-segments of 3D object mesh. Given a 3D object mesh " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": " with segments " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "\\{s_{ik}^{l}\\}" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "l^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": " iteration, we assign a color " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "c(s_{ik}^{l})" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": " per segment and generate a graph " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i^l" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": ". The pair of neighboring nodes is merged if the distance between assigned colors " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "c(s_{ik}^{l})" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": " is within a threshold, then, move to " + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "inline_equation", + "content": "(l + 1)^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 146, + 547, + 224 + ], + "type": "text", + "content": " iteration." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 236, + 413, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 413, + 250 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 413, + 250 + ], + "type": "text", + "content": "3.2. Object Stylization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": "Object stylization involves understanding the semantic structure hidden behind the mesh representation. As a preprocessing, we first subdivide individual objects " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " into part segments " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\{s_{ik}\\}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " as described in Sec. 3.2.1. Then the scene is stylized in two steps. First, we assign base colors into individual parts to minimize the style loss for the entire scene " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "S = \\{\\mathcal{W},\\mathcal{O}\\}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " (Sec. 3.2.2). Here we are optimizing for the discrete set of colors assigned to subdivided parts obtained from the pre-processing. Then the textures for individual objects are further optimized to generate fine details (Sec. 3.2.3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 406, + 496, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 496, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 496, + 418 + ], + "type": "text", + "content": "3.2.1 Part Discovery for Object Stylization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "content": "We first decompose individual objects into parts such that each part is composed of the same material or texture. The distinctive part boundaries are critical in providing semantic consistency and, therefore, perceptual realism toward the scene. Similar to 3D part segmentation methods, we first find super-segments based on geometric features which provide the granularity to define textural parts. For a given 3D object mesh " + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "content": ", the initial segments " + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\{s_{ik}^{0}\\}" + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "content": " are the decomposition applying the method by Katz et al. [20]. The decomposition is designed to be an over-segmentation of our aim. We generate a graph " + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i^0" + }, + { + "bbox": [ + 304, + 426, + 545, + 616 + ], + "type": "text", + "content": " where each node is the segment and edges connect neighboring segments. Then we incrementally merge segments that belong to the same texture until convergence. Note that, our part discovery method operates robustly regardless of the initial composition, but we use [20] which preserves the original geometry and details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "The challenge here is that, unlike semantic segmentation approaches, no large-scale public dataset exists that provides the ground truth for 'texture similarity' as the segmentation labels. We find a supervision signal from the large-scale pre-trained model, and create a simple text prompt " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{i,c}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " using the class name, such as a bed or a chair. At " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "l^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " iteration, we assign a color " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "c(s_{ik}^{l})" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " to each segment " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{s_{ik}^{l}\\}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", which is optimized to minimize the distance" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1893" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "between the rendered images " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "I_{M_i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " of multiple viewpoints and the text " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "T_{i,c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " in the joint embedding space of CLIP, or " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{clip}}(I_{M_i},T_{i,c})" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " as defined in Equation (2). If the resulting colors assigned to two adjacent segments are similar, the two parts are likely to be the parts with the same texture source. Therefore we merge the two segments for the next iteration " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\{s_{ik}^{l + 1}\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": ". In particular, we merge segments if the assigned color has a distance of less than a threshold " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\lambda_{th}" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": " in the CIE color space which is known to be related to human perception. Note that, while the initial color is gray for the assignment, merging segments happen with the optimized color. We repeat the process until the number of segments does not decrease anymore and empirically found that it usually converges within 2-3 steps. The overall pipeline for part discovery is also described in Fig. 4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 265, + 224, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 224, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 224, + 277 + ], + "type": "text", + "content": "3.2.2 Part-level Base Color Assignment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "text", + "content": "After the objects " + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "inline_equation", + "content": "M_{i} \\in \\mathcal{O}" + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "text", + "content": " are decomposed into parts " + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\{s_{ik}\\}" + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "text", + "content": ", we assign a solid color per part, namely " + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "inline_equation", + "content": "c(s_{ik})" + }, + { + "bbox": [ + 46, + 284, + 287, + 368 + ], + "type": "text", + "content": ". The base color assignment handles a low-dimensional optimization space with a coarse set of discrete parts but still observes the holistic distribution of the entire scene. Then the base color is combined with the output of designed neural style fields in Sec. 3.2.3, which generate high-frequency local texture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "text", + "content": "We optimize the base color using the combined loss as before, " + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color,scene}} + \\mathcal{L}_{\\mathrm{clip,scene}}" + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "text", + "content": ". The color loss again considers the similarity with the target image " + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hist}}(I, I_t)" + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "text", + "content": ". The scene being optimized " + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "text", + "content": " is rendered with the stylized structure " + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 368, + 287, + 462 + ], + "type": "text", + "content": " and the current estimates of the base colors. The clip loss considers both individual objects and the global scene and is calculated as the sum of object clip loss and global clip loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 467, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 467, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 52, + 467, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l i p , s c e n e}} = \\lambda_ {2} \\cdot \\sum_ {i} \\mathcal {L} _ {\\text {c l i p}} \\left(I _ {M _ {i}}, T _ {i}\\right) + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\text {c l i p}} (I, T). \\tag {4}", + "image_path": "c0438ecf7eaeaac06d13476e5b71da5725dcf2443cabf5bce13003649c85ab0e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": "Unlike Sec. 3.2.1, text description " + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": " for object " + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": " could be a simple text prompt using the class name, or a detailed text prompt based on user choice. We render individual objects from various angles " + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "inline_equation", + "content": "I_{M_i}" + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": " and compare them with the text description " + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": ". The embedding for the scene is also compared against the text embedding " + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": " represents the type of scene, for example, 'a bedroom'. By jointly applying the loss, the base color " + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "inline_equation", + "content": "c(s_{ik})" + }, + { + "bbox": [ + 46, + 495, + 288, + 604 + ], + "type": "text", + "content": " is selected as a representative color that harmonizes nicely with the global context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 617, + 164, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 164, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 164, + 629 + ], + "type": "text", + "content": "3.2.3 Detailed Stylization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 635, + 286, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 286, + 696 + ], + "type": "text", + "content": "The base color is combined with additional details regressed from a neural network to express the detailed local texture. We define local neural style field (LNSF) for each object, which generates the local textures added to the base color. The color for point " + }, + { + "bbox": [ + 46, + 635, + 286, + 696 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 635, + 286, + 696 + ], + "type": "text", + "content": " is defined by the following equation," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 97, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "c \\left(s _ {p}\\right) + \\alpha \\cdot \\mathcal {F} _ {i} (\\gamma (p), \\phi (p), s _ {p}). \\tag {5}", + "image_path": "6a05786aaec5cbdecac67e1e85a14164dad02fcd2b957bbc9f20348b5dc39414.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "We train a LNSF " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " per object, which outputs the color to be added to the base color " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "c(s_p)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " indicates the part segment id. The color range " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " maintains the final color to be similar to the base color. " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\gamma(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " is the positional encoding of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " coordinate to capture high-frequency details, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " represents the coefficients of the eigenfunctions for the intrinsic geometry using the Laplace-Beltrami operator. Therefore the object-specific neural fields respect the part boundaries and local geometric details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": "LNSF for each object are trained with additional style context " + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{clip}}(I_{M_i},T_i^+)" + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": ". The text prompt " + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "inline_equation", + "content": "T_{i}^{+}" + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": " augments the object description of " + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": " with appearance style description " + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": ", such as 'minimal style' or 'Mid-Century style'. By enforcing the same appearance style description, we can weakly bind the styles of individual objects in the same scene while optimizing separately. We could also use " + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "inline_equation", + "content": "T_{i}^{+}" + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 180, + 545, + 288 + ], + "type": "text", + "content": " for optimizing the base color, and it shows slightly better results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "text", + "content": "Part-aware Geometric Deformation Even though our main focus is to add colors to objects with local part-aware details, we could also concurrently produce part-aware geometric deformation with the same architecture. We can slightly change the LNSF " + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "text", + "content": " to have two branches of output that estimate color and displacement. For each point on the surface, we assign color by Eq. 5 and adjust displacement along the vertex normal direction. To learn the effective deformation, we add geometric loss " + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{clip}}(I_{M_i}^{geo}, T_i^+)" + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "inline_equation", + "content": "I_{M_i}^{geo}" + }, + { + "bbox": [ + 304, + 304, + 545, + 426 + ], + "type": "text", + "content": " is an image rendering textureless geometry as [27]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "type": "text", + "content": "Rendering and Implementation Details To render each object, the individual objects " + }, + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "type": "text", + "content": " are scaled to fit a unit box. The predicted color of each vertex allows the entire mesh to be differentiable rendering through interpolation using [5]. We render individual objects with random augmented backgrounds (white, black, random Gaussian, chess board), which helps the pipeline focus on the foreground object [16, 18]. Inspired by [36], since CLIP has a bias for the canonical pose, we augment the text prompt with the azimuth angle of view, namely 'front view of', 'side view of' or 'back view of'. Finally, in Sec 3.2.3, random perspective transformation and random crop boost learning local details. To render a scene, we randomly sampled from predefined 20 camera poses to evenly cover the entire scene " + }, + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 441, + 545, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 388, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "Our stylization is first evaluated for individual objects in Sec. 4.1. We evaluate the quality of textured object meshes and also assess that our pipeline can discover part segments to assign realistic stylization. Then we demonstrate the stylization results of room-scale scenes with multiple objects in Sec. 4.2." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1894" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 541, + 445 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 445 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 445 + ], + "type": "image", + "image_path": "aa4b62ba5d2e5162222124dd0f8c2b35d9144a639b34d54e50ae5f6716ada7a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 453, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 547, + 498 + ], + "type": "text", + "content": "Figure 5. Visual comparison with baselines. Our method utilizes the found part information and generates detailed textures through the designed network in combination with a coarse-to-fine stylization scheme. As a result, it generates a globally harmonious 3D style that clearly classifies part segments. We can handle diverse text such as simple categories or styles, detailed descriptions, or text that includes different two object categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 511, + 154, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 154, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 154, + 525 + ], + "type": "text", + "content": "4.1. Object Stylization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 532, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 288, + 686 + ], + "type": "text", + "content": "We show that our method can stylize various objects and produce realistic 3D assets to populate virtual scenes. We use object meshes of various types and sizes, collected from Turbo Squid [48], 3D-FUTURE [13], and Amazon Berkeley Objects [10]. For detailed stylization, we subdivide the object into an average of 119769 faces and 60437 vertices. For large general objects such as beds, sofas, tables, etc., we use class labels and the text input for the specific style as 'a [class label], [specific] style'. However, we need a detailed explanation for small objects that cannot be explained only by class labels, such as books. For these objects, we configure detailed text individually, for example, 'a book of Harry Potter and the Sorcerer's Stone'." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Figure 5 shows the renderings of stylized results. We render objects in Blender [11] with a fixed lighting setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 511, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 704 + ], + "type": "text", + "content": "The input mesh and the text are also provided in the left column. Since our method utilizes the obtained part information and coarse-to-fine stylization scheme with a two-step approach, it creates a more realistic texture with clear boundaries for each part of the 3D assets. Text2Mesh [27] is another text-driven 3D stylization approach, which adds the RGB color and deformation fields on the vertices of mesh. While it augments the detailed variations on the input mesh, the deformation map can occasionally introduce undesired artifacts and the part boundaries are only roughly estimated. We also provide the results of an ablated version that directly uses LNSF without a two-step stylization scheme. Our part discovery module and coarse-to-fine stylization scheme play a critical role in producing realistic assets with high-quality stylization. Figure 1 contains more results on diverse objects." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1895" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 272, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 272, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 272, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 272, + 159 + ], + "type": "image", + "image_path": "cbbc634a8d973165b74fb6edf98a0bfc222b7fbe8a748189cbb36b7bb5f1b2c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 167, + 262, + 178 + ], + "lines": [ + { + "bbox": [ + 72, + 167, + 262, + 178 + ], + "spans": [ + { + "bbox": [ + 72, + 167, + 262, + 178 + ], + "type": "text", + "content": "Figure 6. Results of user study for object stylization" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 185, + 289, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 289, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 289, + 330 + ], + "type": "text", + "content": "Because stylization is a subjective task, and no ground truth or metrics exist, we conduct a user study for quantitative evaluation. We ask 98 users to rate the quality of generated outputs on a scale of 1 (worst) to 5 (best) in response to the following questions: (Q1) 'How natural are the output results?', (Q2) 'How well does the output contain text information?', and (Q3) 'How well does the output reflect part information?' Figure 6 contains the mean and the standard deviation of the scores. Our method outperforms competing methods in all aspects. Therefore we conclude that our method generates a realistic texture that abides by input text description and semantic part information." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 339, + 281, + 475 + ], + "blocks": [ + { + "bbox": [ + 55, + 339, + 281, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 339, + 281, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 281, + 475 + ], + "type": "image", + "image_path": "12fe3abb18708b9695db216600524072029067d8b93f4d94ee2709be27ce5338.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 483, + 287, + 527 + ], + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 527 + ], + "type": "text", + "content": "Figure 7. With a slightly modified LNSF " + }, + { + "bbox": [ + 46, + 483, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 483, + 287, + 527 + ], + "type": "text", + "content": ", we can simultaneously generate the texture with both color and displacement fields, and all of the generated style information respects the part information discovered." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 529, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 288, + 673 + ], + "type": "text", + "content": "We also provide a parallel comparison against Text2Mesh by incorporating deformation fields with our approach. Our original method preserves the original shape of the mesh, while Text2Mesh deforms vertices along the normal direction in addition to generating texture. We can make a similar version of our LNSF and produce additional deformation as described in Sec. 3.2.3. Figure 7 shows the deformed results from the source mesh compared to Text2Mesh. Since our approach explicitly considers the part information with the two-stage approach, our results with deformation fields also respect different semantic parts of the object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Part Discovery As a side product of object stylization, we can discover different parts with distinguished boundaries" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 198 + ], + "type": "image", + "image_path": "ade3d4c4ae2d3e45c72074051f8a85e117a2a6262b612c50c3bde8156a84282d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 547, + 271 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 271 + ], + "type": "text", + "content": "Figure 8. The robustness against initial super-segmentation. The super-segments of the top and the bottom rows are generated by [20] and [7], respectively (left). Starting from the initial color assignment (middle), our approach stably finds part decomposition to assign different base colors and therefore different texture information (right). The text a chair is used." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 285, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 417 + ], + "type": "text", + "content": "that can guide a realistic color assignment (Sec. 3.2.1). Our part discovery combines geometric super-segments and implicit clues from image-text embeddings. The initial super-segments guide the algorithm to follow geometric feature lines, and the final results stably find part information despite different initializations. Figure 8 shows examples initialized with Katz et al. [20] (top) and BSP-Net [7] (bottom). Even for input mesh with bad topology or different initialization, our iterative part discovery quickly finds visually coherent parts without any training with segmentation labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 429, + 408, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 408, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 408, + 441 + ], + "type": "text", + "content": "4.2. Scene Stylization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": "Now we demonstrate that our text2scene framework can quickly generate a realistic texture for a room with multiple objects. Recall that the 3D scene " + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "inline_equation", + "content": "S = \\{\\mathcal{W}, \\mathcal{O}\\}" + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": " is composed of the structure components " + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": " and a set of objects " + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": ". We use the same objects as described in Sec. 4.1 and arrange them to constitute scenes. As there is no existing dataset composite of complete object meshes with labels, we built a total of four scenes: two bedrooms and two living rooms. Each scene contains an average of 20 objects of various sizes and classes. Additionally, we provide a target image " + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": " for the color distribution, and a text prompt describing the desired style " + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": ". While it can be daunting to define the desired style for the entire scene, images and texts can provide a simple way to deliver the information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Figure 2 shows stylized results of the same geometry, but observing various target images and style prompts. The generated textures respect the semantic labels of various furniture and different parts and contain localized diverse details. This is in contrast to many prior stylization methods where the fine perturbations are spread throughout the scenes. We also show scenes with different types of rooms containing different objects, but stylized based on the same" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1896" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "target image in Fig. 9. Additional results of various input configurations are available in the supplementary material. Note that our target image " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " does not need to restrict as an indoor image, and can be replaced such as natural photographs. And by changing random seeds, diverse results can be obtained from the same input conditions. Also, since we stylize the entire object, we can easily edit the 3D scene through object relocation. These results can also be found in the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 190, + 168, + 275 + ], + "blocks": [ + { + "bbox": [ + 60, + 190, + 168, + 275 + ], + "lines": [ + { + "bbox": [ + 60, + 190, + 168, + 275 + ], + "spans": [ + { + "bbox": [ + 60, + 190, + 168, + 275 + ], + "type": "image", + "image_path": "ec2cf0e19f6b4d261fead6170adddba1ab3e8a60072081e13a764f68aff450ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 275, + 132, + 283 + ], + "lines": [ + { + "bbox": [ + 95, + 275, + 132, + 283 + ], + "spans": [ + { + "bbox": [ + 95, + 275, + 132, + 283 + ], + "type": "text", + "content": "Livingroom" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 293, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 316 + ], + "type": "text", + "content": "Figure 9. Result for the different arrangement of objects. We used target image number 3 in Fig. 2 in both spaces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 189, + 276, + 275 + ], + "blocks": [ + { + "bbox": [ + 168, + 189, + 276, + 275 + ], + "lines": [ + { + "bbox": [ + 168, + 189, + 276, + 275 + ], + "spans": [ + { + "bbox": [ + 168, + 189, + 276, + 275 + ], + "type": "image", + "image_path": "8b341e11d77ac96fd8cf200f5c16f769cd739ac944e9159906297bcad4d3a28e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 275, + 236, + 283 + ], + "lines": [ + { + "bbox": [ + 207, + 275, + 236, + 283 + ], + "spans": [ + { + "bbox": [ + 207, + 275, + 236, + 283 + ], + "type": "text", + "content": "Bedroom" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 328, + 286, + 358 + ], + "blocks": [ + { + "bbox": [ + 50, + 328, + 286, + 358 + ], + "lines": [ + { + "bbox": [ + 50, + 328, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 50, + 328, + 286, + 358 + ], + "type": "table", + "html": "
-retrieval-hist,glo-detailOurs
(Q1): Realistic2.23(±0.48)3.68(±0.49)4.02(±0.49)4.22(±0.42)
(Q2): Color2.09(±0.48)2.63(±0.59)3.95(±0.47)3.86(±0.49)
", + "image_path": "66cc16f9fc544a2bfb8fbe7cb3d5e67a839ffcd6d22bbe0e73e5ad573b916924.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 366, + 259, + 377 + ], + "lines": [ + { + "bbox": [ + 75, + 366, + 259, + 377 + ], + "spans": [ + { + "bbox": [ + 75, + 366, + 259, + 377 + ], + "type": "text", + "content": "Table 1. Results of user study for scene stylization" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 387, + 108, + 436 + ], + "blocks": [ + { + "bbox": [ + 48, + 387, + 108, + 436 + ], + "lines": [ + { + "bbox": [ + 48, + 387, + 108, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 108, + 436 + ], + "type": "image", + "image_path": "61b09640015035873ba64b4676b91d1078418a9363d8611416900b2f9f2955aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 437, + 97, + 445 + ], + "lines": [ + { + "bbox": [ + 59, + 437, + 97, + 445 + ], + "spans": [ + { + "bbox": [ + 59, + 437, + 97, + 445 + ], + "type": "text", + "content": "- retrieval" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 456, + 287, + 478 + ], + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 478 + ], + "type": "text", + "content": "Figure 10. Ablation results for the scene. We used target image number 1 in Fig. 2 in all spaces." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 108, + 387, + 167, + 436 + ], + "blocks": [ + { + "bbox": [ + 108, + 387, + 167, + 436 + ], + "lines": [ + { + "bbox": [ + 108, + 387, + 167, + 436 + ], + "spans": [ + { + "bbox": [ + 108, + 387, + 167, + 436 + ], + "type": "image", + "image_path": "ff8c3cad378d18f90b5d540424e739799571d994195a70c3200b62b0e76677f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 437, + 155, + 446 + ], + "lines": [ + { + "bbox": [ + 120, + 437, + 155, + 446 + ], + "spans": [ + { + "bbox": [ + 120, + 437, + 155, + 446 + ], + "type": "text", + "content": "-hist,glo" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 168, + 387, + 227, + 436 + ], + "blocks": [ + { + "bbox": [ + 168, + 387, + 227, + 436 + ], + "lines": [ + { + "bbox": [ + 168, + 387, + 227, + 436 + ], + "spans": [ + { + "bbox": [ + 168, + 387, + 227, + 436 + ], + "type": "image", + "image_path": "476364c798fe33cf99a0894c70cbd6bdf9721e35f03b1455e6be1a4f66ad7349.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 437, + 211, + 445 + ], + "lines": [ + { + "bbox": [ + 183, + 437, + 211, + 445 + ], + "spans": [ + { + "bbox": [ + 183, + 437, + 211, + 445 + ], + "type": "text", + "content": "-detail" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 227, + 387, + 287, + 436 + ], + "blocks": [ + { + "bbox": [ + 227, + 387, + 287, + 436 + ], + "lines": [ + { + "bbox": [ + 227, + 387, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 227, + 387, + 287, + 436 + ], + "type": "image", + "image_path": "2e2a5d164df5e0d1b2770ab7b9aeaf6bcdf063b601e87e9b44acf4a47ad02076.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 437, + 266, + 445 + ], + "lines": [ + { + "bbox": [ + 247, + 437, + 266, + 445 + ], + "spans": [ + { + "bbox": [ + 247, + 437, + 266, + 445 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 713 + ], + "type": "text", + "content": "We also provide users' evaluation of the quality of our stylization in Table 1. Users evaluate the results by answering the following questions: (Q1) 'How realistic is the output result?' (Q2) 'How similar is the color distribution of the scene to the given images?' Users assess the overall quality of the outputs, and how well they match the target image. Since there are no previous works that use the same setting, we compare the results of ablated versions: -retrieval replaces the separate texture retrieval module (Sec. 3.1) with a stylization network of objects for structure components; -hist,glo removes the color loss and the global clip loss for the base color assignment (Sec. 3.2.2); and -detail removes the detailed stylization step for objects (Sec. 3.2.3). The responses of (Q1) indicate that Text2Scene generates high-quality scenes and each of the components plays a crucial role to achieve reality. The effect of texture retrieval is the most prominent. The color distribution results (Q2) indicate that the base color assignment is critical. The added local details are more important for the realism" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "of the results. Figure 10 shows exemplar images of the ablated versions used for the user study." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 110, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 266 + ], + "type": "text", + "content": "GPU Cost and Scalability Text2Scene first assigns base colors to discovered parts for all objects in the scene, and generates details of objects individually. The cardinality for the base color assignment is only a few hundred and it does not require much memory and allows us to consider the whole scene while expanding the scale. The most memory-intensive process is the detail generation using a neural network, which only processes a single object at a time. Therefore the entire process can be trained only on a single 11 GB GPU, making it an accessible tool for casual users. In a single GPU, the base color allocation of the entire space takes 5 hours, and learning the details for each object takes 10 minutes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 281, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 414 + ], + "type": "text", + "content": "Limitations While our approach results in scene stylization, the pipeline separately handles individual objects after the base color assignment. This is a practical choice for scalability but may lack an understanding of the context of the entire space. Instead, we rely on the text description to weakly bind the objects into a similar style. We can design the pipeline to receive an additional input with a texture map or a lightweight network and extend our model to better observe the holistic scene context within a limited GPU memory. Also, our pipeline requires a class label or optional text description per object, which can be further automated." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 425, + 383, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 383, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 383, + 437 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 445, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 588 + ], + "type": "text", + "content": "We introduce Text2Scene, a novel framework to generate a texture for 3D scenes. Our hierarchical framework can handle a variety of objects, including highly detailed textures for objects such as book or paintings. By leveraging the representation power of pre-trained CLIP, the framework does not require any 3D datasets with texture or part annotation. Given the 3D mesh models and the class labels or text descriptions of objects, our framework easily produces stylized results by picking a target image and a simple text description. We hope that Text2Scene to facilitate the automatic interior recommendation or realistic virtual space generation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 602, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS-2023-00208197) and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub). Inwoo Hwang is supported by Hyundai Motor Chung Mong-Koo Foundation. Young Min Kim is the corresponding author." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1897" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Mahmoud Affifi, Marcus A. Brubaker, and Michael S. Brown. Histogram: Controlling colors of gan-generated and real images via color histograms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6290–6301, June 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In ECCV, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Kang Chen, Kun Xu, Yizhou Yu, Tian-Yi Wang, and ShiMin Hu. Magic decorator: Automatic material suggestion for indoor digital scenes. ACM Trans. Graph., 34(6):232:1-232:11, Oct. 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 361 + ], + "type": "text", + "content": "[5] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 287, + 407 + ], + "type": "text", + "content": "[6] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 408, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 453 + ], + "type": "text", + "content": "[7] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bspnet: Generating compact meshes via binary space partitioning. Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 486 + ], + "type": "text", + "content": "[8] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Kangxue Yin, Matthew Fisher, Siddhartha Chaudhuri, and Hao Zhang. Bae-net: Branched autoencoder for shape co-segmentation. Proceedings of International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 599 + ], + "type": "text", + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, Matthieu Guillaumin, and Jitendra Malik. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "text", + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d fur" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "niture shape with texture. International Journal of Computer Vision, pages 1-25, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[14] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. GANcraft: Unsupervised 3D Neural Rendering of Minecraft Worlds. In ICCV, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 130, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 184 + ], + "type": "text", + "content": "[15] Lukas Hollein, Justin Johnson, and Matthias Nießner. Stylemesh: Style transfer for indoor 3d scene reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6198-6208, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 230 + ], + "type": "text", + "content": "[16] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. ACM Transactions on Graphics (TOG), 41(4):1-19, 2022. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[17] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[19] Jaebong Jeong, Janghun Jo, Sunghyun Cho, and Jaesik Park. 3d scene painting via semantic image synthesis. In Proceedings of the International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 388 + ], + "type": "text", + "content": "[20] Sagi Katz and Ayellet Tal. Hierarchical mesh decomposition using fuzzy clustering and cuts. ACM Transactions on Graphics, 22, 06 2003. 2, 4, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 389, + 545, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 433 + ], + "type": "text", + "content": "[21] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. SIGGRAPH Asia 2022 Conference Papers, December 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 477 + ], + "type": "text", + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems, volume 35, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 479, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 511 + ], + "type": "text", + "content": "[23] Lukas Koestler, Daniel Grittner, Michael Moeller, Daniel Cremers, and Zorah Lähner. Intrinsic neural fields: Learning functions on manifolds. In ECCV, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 547, + 567 + ], + "type": "text", + "content": "[24] Juil Koo, Ian Huang, Panos Achlioptas, Leonidas J Guibas, and Minhyuk Sung. Partglot: Learning shape part segmentation from language reference games. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[25] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. arXiv preprint arXiv:2112.03221, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1898" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "type": "text", + "content": "[29] Shailesh Mishra and Jonathan Granskog. Clip-based neural neighbor style transfer for 3d assets. In arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 196 + ], + "type": "text", + "content": "[30] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 241 + ], + "type": "text", + "content": "[31] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings IEEE International Conf. on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 243, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 287, + 284 + ], + "type": "text", + "content": "[32] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 331 + ], + "type": "text", + "content": "[33] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 332, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 287, + 386 + ], + "type": "text", + "content": "[34] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 442 + ], + "type": "text", + "content": "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 445, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 475 + ], + "type": "text", + "content": "[36] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 478, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 510 + ], + "type": "text", + "content": "[37] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. arXiv preprint arXiv:2011.13961, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "type": "text", + "content": "[38] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 546, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 546, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 546, + 287, + 600 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 602, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 602, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 602, + 287, + 634 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 636, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 667 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arxiv:2208.12242, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 586 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[44] Liang Shi, Beichen Li, Milos Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: Differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6):1-15, Dec. 2020. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "text", + "content": "[45] Vincent Sitzmann, Julien N. P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 219, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 261 + ], + "type": "text", + "content": "[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "text", + "content": "[47] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 319 + ], + "type": "text", + "content": "[48] TurboSquid. Turbosquid 3d model repository. In https://www.turbosquid.com/, 2022.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "type": "text", + "content": "[49] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 354, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 420 + ], + "type": "text", + "content": "[50] Yu-Ying Yeh, Zhengqin Li, Yannick Hold-Geoffroy, Rui Zhu, Zexiang Xu, Miloš Hasan, Kalyan Sunkavalli, and Manmohan Chandraker. Photoscene: Photorealistic material and lighting transfer for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18562–18571, June 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 464 + ], + "type": "text", + "content": "[51] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 509 + ], + "type": "text", + "content": "[52] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 511, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 542 + ], + "type": "text", + "content": "[53] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In ECCV, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 586 + ], + "type": "text", + "content": "[54] Dongsu Zhang, Changwoon Choi, Inbum Park, and Young Min Kim. Probabilistic implicit scene completion. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1899" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_content_list.json b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5d1a70b32d858c695e949dc841f166b48402e226 --- /dev/null +++ b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_content_list.json @@ -0,0 +1,1573 @@ +[ + { + "type": "text", + "text": "Texts as Images in Prompt Tuning for Multi-Label Image Recognition", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixian Guo $^{1,2*}$ Bowen Dong $^{1}$ Zhilong Ji $^{2}$ Jinfeng Bai $^{2}$ Yiwen Guo $^{4}$ Wangmeng Zuo $^{1,2(\\text{图})}$", + "bbox": [ + 189, + 181, + 776, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Harbin Institute of Technology $^{2}$ Tomorrow Advancing Life $^{3}$ Pazhou Lab, Guangzhou $^{4}$ Independent Researcher", + "bbox": [ + 132, + 200, + 836, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zixian.guo@foxmail.com cndongsky@gmail.com zhilongji@hotmail.com", + "bbox": [ + 187, + 219, + 781, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jfbai.bit@gmail.com guoyiwen89@gmail.com WMzuo@hit.edu.cn", + "bbox": [ + 210, + 237, + 745, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt tuning has been employed as an efficient way to adapt large vision-language pre-trained models (e.g. CLIP) to various downstream tasks in data-limited or label-limited settings. Nonetheless, visual data (e.g., images) is by default prerequisite for learning prompts in existing methods. In this work, we advocate that the effectiveness of image-text contrastive learning in aligning the two modalities (for training CLIP) further makes it feasible to treat texts as images for prompt tuning and introduce TaI prompting. In contrast to the visual data, text descriptions are easy to collect, and their class labels can be directly derived. Particularly, we apply TaI prompting to multi-label image recognition, where sentences in the wild serve as alternatives to images for prompt tuning. Moreover, with TaI, double-grained prompt tuning (TaI-DPT) is further presented to extract both coarse-grained and fine-grained embeddings for enhancing the multi-label recognition performance. Experimental results show that our proposed TaI-DPT outperforms zero-shot CLIP by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE, while it can be combined with existing methods of prompting from images to improve recognition performance further. The code is released at https://github.com/guozix/Tai-DPT.", + "bbox": [ + 75, + 319, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent few years have witnessed rapid progress in large vision-language (VL) pre-trained models [1, 16, 19, 24, 33, 36] as well as their remarkable performance on downstream vision tasks. A VL pre-trained model generally involves data encoders, and it is becoming increasingly popular to exploit image-test contrastive loss [24] to align the embedding of images and texts into a shared space. When adapting to downstream tasks in data-limited or label-limited settings, it is often ineffective to fine-tune the entire model, due to its high complexity. Then, prompt tuning as a representative parameter-efficient learning paradigm has emerged as", + "bbox": [ + 75, + 707, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da7c14fde341874046b193f3d59da4fdd6a813b0af28615007eb5717bd026ab4.jpg", + "image_caption": [ + "Figure 1. A comparison between prompting from images and our text-as-image (TaI) prompting. (a) Prompting from images (e.g., [41]) uses labeled images of task categories to learn the text prompts. Instead, (b) our TaI prompting learn the prompts with easily-accessed text descriptions containing target categories. (c) After training, the learned prompts in (a) or (b) can be readily applied to test images." + ], + "image_footnote": [], + "bbox": [ + 508, + 289, + 879, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an efficient way to adapt VL models to downstream tasks.", + "bbox": [ + 500, + 609, + 882, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Albeit considerable achievements have been made, existing prompt tuning methods generally require visual data to learn prompts (as shown in Fig. 1(a)). For example, CoOp [41] learns from annotated images. CoCoOp [40] further introduces generalizable input-conditional prompts. DualCoOp [28] adapts CLIP to multi-label recognition tasks by training pairs of positive and negative prompts with partial-labeled images. Nonetheless, the performance of these prompting methods may be limited when it is infeasible to obtain sufficient image data or annotate the required images.", + "bbox": [ + 496, + 626, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we advocate treating Texts as Images for prompt tuning, i.e., TaI prompting. It is considered feasible as the image encoder and text encoder in many pre-trained VL models [16, 24] encode images and texts into a shared space. Given an image and its caption, the visual features produced by the image encoder will be close to the text feature of the caption produced by the text encoder. There-", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Zixian Guo was a research intern at TAL.", + "bbox": [ + 94, + 886, + 460, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2808", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fore, in addition to extracting visual features from images, it is also feasible to extract text features as alternatives form, for example, descriptive sentences and captions, for prompt tuning (see Fig. 1(b)). TaI prompting has several interesting properties and merits. Taking a downstream image recognition task as an example, given a set of object categories, one can easily crawl a large set of text descriptions that contain object names from these categories. Text descriptions are easily accessible in this way, and class labels can be directly derived from text descriptions, which means, in contrast to prompting from images, TaI prompting may suffer less from the data-limited and label-limited issues.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use multi-label image recognition [8, 9, 11, 20, 35] to verify the effectiveness of our TaI prompting in this paper. To begin with, we crawl the captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., Open Images [18]) to form the training set of text descriptions. For any specific multi-label recognition task, we adopt a noun filter to map the nouns in the text descriptions to the corresponding object categories, and then only keep the text descriptions that contain one or more classes of target objects. To better cope with multi-label classification, we introduce double-grained prompt tuning (i.e., TaI-DPT) which involves: (i) a set of global prompts to generate embeddings for classifying whole sentences or images, and (ii) a set of local prompts to extract embeddings for discriminating text tokens or image patches. Given a set of text descriptions, global and local prompts can be tuned by minimizing the ranking loss [14]. Note that, though these prompts are learned from text descriptions solely, they can be readily deployed to classify whole images as well as image patches during testing (see Fig. 1(c)). Experimental results show that, without using any labeled images, our TaI prompting surpasses zero-shot CLIP [24] by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE.", + "bbox": [ + 75, + 273, + 472, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, when images are also available during training, our TaI prompting can be combined with existing methods of prompting from images to improve its performance. In particular, given a few annotated images, our TaI-DPT can be integrated with CoOp as a prompt ensemble for improving classification accuracy. With partially labeled training data being provided, we may also combine TaI-DPT and DualCoOp [28] to improve multi-label recognition accuracy consistently. Extensive results verify the effectiveness of our TaI-DPT in comparison to state-of-the-art. To sum up, the contributions of this work include:", + "bbox": [ + 75, + 636, + 472, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose Texts as Images in prompt tuning (i.e., TaI prompting) to adapt VL pre-trained models to multi-label image recognition. Text descriptions are easily accessible and, in contrast to images, their class labels can be directly derived, making our TaI prompting very compelling in practice.", + "bbox": [ + 94, + 810, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present double-grained prompt tuning (i.e. TaI-DPT) to extract both coarse-grained and fine-grained embeddings for enhancing multi-label image recognition. Experiments on multiple benchmarks show that TaI-DPT achieves comparable multi-label recognition accuracy against state-of-the-arts.", + "- The prompts learned by TaI-DPT can be easily combined with existing methods of prompting from images in an off-the-shelf manner, further improving multi-label recognition performance." + ], + "bbox": [ + 517, + 90, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 248, + 640, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-Label Image Recognition", + "text_level": 1, + "bbox": [ + 500, + 273, + 777, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-label image recognition [3,7,8,12,15,21,32,35,37] aims to recognize all the object categories [11, 20] or concepts [9] in an input image. Various modules [6, 30] have been introduced to better represent the inter-class relationships, and modern classification losses [3, 14] have been used to make model learning easier.", + "bbox": [ + 496, + 296, + 890, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model the label dependencies, CNN-RNN [30] introduces recurrent neural networks, e.g., RNN and LSTM, to predict appeared classes in a sequential manner. [6,8,31,35] use graph convolution modules to learn the correlation between class labels. CHAMP [29] measures the severity of misclassification by building a domain-specific hierarchy tree according to the relation of categories, where each class is related to a tree node, to improve the robustness of the model. Albeit effective, these methods require a considerable number of labeled images to learn the category relationships sufficiently. While in data-limited or label-limited regimes, e.g., few-shot or partial-label data, it will be difficult for these models to learn well as expected. Specifically designed loss functions also struggle to obtain significant improvements when learning with limited data.", + "bbox": [ + 496, + 387, + 892, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Label Recognition from Few-shot Samples. To better exploit the small number of samples, LaSO [2] synthesizes samples by manipulating the features of paired training images. Different ways of manipulating label sets are used to train the model, resulting in generalizable discriminative features. [27] introduces a meta-learning framework for better learning of past tasks and generalization to new tasks, and leverages the number of labels as useful information for learning.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Label Recognition from Partial-label Data. Partial-label refers to the scenarios where some labels are unknown. [10] propose a normalized BCE loss to balance the proportion of known labels. [5] learns to complement unknown labels by utilizing within-image and cross-image semantic correlations. [23] blends the representation of training images and class proxies to compensate for the loss of information due to unknown labels.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Albeit significant progress has been made, it remains a challenging issue for learning multi-label image recognition", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2809", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/502d18cbf3bbd3f4cf9cf9bfa0f79de8324e2b9bbe0a5338d78a146061fb634d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 450, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/71fabcbc5e0d75b9ccb3a6fb219451d240d60858ebedce1c220902e79a68be4e.jpg", + "image_caption": [ + "(b)", + "Figure 2. Training and testing pipeline of our proposed Text-as-Image (TaI) prompting, where we use text descriptions instead of labeled images to train the prompts. (a) During training, we use two identical text encoders from pre-trained CLIP to extract the global & local class embeddings $(\\mathbf{G} \\& \\mathbf{L})$ and overall & sequential text embeddings $(\\mathbf{h} \\& \\mathbf{H})$ respectively from the prompts and text description. The corresponding cosine similarity $(\\mathbf{p} \\& \\mathbf{P})$ between the embeddings are guided by the derived pseudo labels with ranking loss. (b) During testing, we replace the input from text descriptions to images. The global and local class embeddings can discriminate target classes from global & local image features $(\\mathbf{f} \\& \\mathbf{F})$ . The final classification results are obtained by merging the scores of the two branches." + ], + "image_footnote": [], + "bbox": [ + 452, + 89, + 885, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in image-limited or label-limited regimes. Built upon VL pre-trained models, this paper suggests learning prompts from texts instead of images, thereby offering a novel yet complementary perspective for handling low-resource multi-label image recognition.", + "bbox": [ + 75, + 398, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prompt Tuning for Vision-Language Models", + "text_level": 1, + "bbox": [ + 76, + 484, + 452, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To transfer pre-trained knowledge to downstream tasks in data-limited settings, prompt tuning [13, 17, 34, 39, 41, 42] has become a popular parameter-efficient way to achieve the goal, due to its flexibility and ease of use. CoOp [41] learns the prompts by using (a few) annotated images of each class from the target dataset. CoCoOp [40] further proposes to improve CoOp [41] by formulating the prompts in an image-conditional way to maintain better generalization to unseen classes. To avoid overfitting, ProGrad [42] leverages predictions from zero-shot CLIP to regularize gradients in prompt learning process. TPT [26] suggests optimizing test-time prompts by promoting the consistency of augmented test images. ProDA [22] uses multiple pieces of prompts to estimate the distribution of classifier weights for better handling of varying visual features. DualCoOp [28] firstly adapts CLIP to multi-label image recognition with partially labeled data by learning pairs of positive and negative prompts for each class to ensure independent binary classification for each class.", + "bbox": [ + 75, + 507, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Albeit existing prompt tuning approaches have achieved significant improvements in downstream tasks, images as well as a portion of class labels are prerequisite to supervise the optimization of the learnable prompts. In this paper, we propose to treat texts as images in prompt tuning, which, compared to labeled images, are much easier to collect with existing caption datasets and modern search engines. Our", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proposed TaI-DPT surpasses zero-shot CLIP by a large margin and can be combined with the prompts learned by existing methods of prompting from images to further boost recognition performance.", + "bbox": [ + 498, + 398, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 474, + 671, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we present Text-as-Image prompting, i.e., TaI prompting, for adapting pre-trained VL models to multi-label image recognition. Our TaI prompting uses only easily-accessed free-form texts as training data to learn effective prompts for downstream multi-label recognition tasks. To begin with, We present an overview of TaI prompting in Sec. 3.1. Then, we introduce our preparation of training texts in Sec. 3.2. We further explain the design of the double-grained prompt tuning (i.e., TaI-DPT) in Sec. 3.3, and provide the loss function used to train the model in Sec. 3.4. Finally, we propose to ensemble prompts from TaI-DPT with prompts learned from images as a flexible integration to improve multi-label recognition performance.", + "bbox": [ + 496, + 501, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of Our Method", + "text_level": 1, + "bbox": [ + 500, + 710, + 728, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the design of our proposed TaI-DPT framework, including the training and testing phases. During training, we learn prompts with only supervision from texts. Two identical copies of the text encoder $\\mathrm{Enc}_{\\mathrm{T}}$ from the pre-trained CLIP are used to encode the prompts and text data, respectively. We introduce two sorts of trainable prompts (i.e., the global prompts and local prompts) to obtain global and regional class embeddings. A noun filtering strategy is used to generate classification pseudo-labels for each text description, which is applied to supervise the classification scores obtained by calculating the cosine sim", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2810", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "iliarity of class embeddings and text features. Only the parameters in prompts are optimized in the training phase, while the text encoders are both kept frozen. During testing, the class embeddings are obtained by encoding the two sets of learned prompts with the text encoder $\\mathrm{Enc}_{\\mathrm{T}}$ as in training, while the other input source changes from text descriptions to test images. Pre-trained image encoder $\\mathrm{Enc}_{\\mathrm{I}}$ from CLIP is used to extract global and dense features of each test image, then the cosine similarity are computed between features and class embeddings of the global and local prompts. The final classification result is obtained by fusing the global and local cosine scores. In the following, we explain the details of our proposed method.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Preparation of Text Descriptions", + "text_level": 1, + "bbox": [ + 76, + 295, + 362, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain sufficient category information from the language that helps in image recognition, we have to ensure that: 1) the collected texts should contain content-rich descriptions of a scene, and 2) the contents of all text descriptions need to cover the category set of the target dataset for comprehensive representations of all categories. With the aim of ensuring reproducibility, we use readily available captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., OpenImages [18]) as our language data source, while avoiding the workloads associated with randomly crawling texts from the Internet in this paper. Note that although each caption is paired with a corresponding image and human-annotated labels, we only use the captions, and no information from the pictures and labels is disclosed during training.", + "bbox": [ + 75, + 318, + 468, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a target multi-label recognition dataset $\\mathcal{X}$ that has a category set $S = \\{\\mathrm{s}_1, \\mathrm{s}_2, \\mathrm{s}_3, \\dots, \\mathrm{s}_C\\}$ , where $C$ denotes the number of categories and $\\mathrm{s}_i$ denotes particular class name like \"dog\", \"plane\", etc., we search for sentences that contain at least one class name $\\mathrm{s}_i$ in $S$ . Since multiple words or phrases usually exist to represent the same meaning for each class, searching solely for exact match of category names in texts may lead to many false negatives in the obtained pseudo ground-truth labels, which is harmful to prompt tuning. Towards tackling this issue, we introduce a noun filter to map nouns with similar meanings into the corresponding class label. Specifically, we construct a synonym dictionary $\\mathcal{D}$ by including common synonyms of each class name in the target dataset. If a word in a text description matches any synonym of a specific class name, it is considered to contain a description of that category. Several examples of synonyms are shown as follows:", + "bbox": [ + 75, + 559, + 470, + 795 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "{'dog', 'pup', 'puppy', 'doggy'} \n{'person', 'people', 'man', 'woman', 'human'} \n{'bicycle', 'bike', 'cycle'} \n{'car', 'taxi', 'automobile'} \n{'boat', 'raft', 'dinghy'} \n...", + "guess_lang": "txt", + "bbox": [ + 94, + 801, + 460, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More details of the dictionary $\\mathcal{D}$ are provided in the Suppl.", + "bbox": [ + 76, + 885, + 464, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we conduct noun filtration by the following steps. First, for each text description, we use the tokenizer and lemmatizer from NLTK [4] to recover the stem of each word in the sentences. Next, for all keywords in $\\mathcal{D}$ , which contains all synonyms of the category set $S$ , we search in our language data source for sentences that contains at least one class name. For the text descriptions that do not match any synonym of any class name, we simply drop them away to ensure each piece of data has at least one concerned label. Finally, for each retained text description, we convert the class names it contains into binary pseudo-ground-truth vectors by setting classes that appear as positive and other classes as negative, following the order of class labels in the target dataset $\\mathcal{X}$ .", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The word-level filtered labels may not be precisely correct since our searching strategy mentioned above is rather simple considering the diversity of free-form texts, where complex paraphrases and misspellings that widely exist in the corpus are not fully addressed. However, such a simple noun filtration can guarantee the reproducibility of this work, and our experiments also demonstrate that this simple and efficient data preparation leads to compelling results of multi-label recognition of our TaI prompting.", + "bbox": [ + 496, + 301, + 892, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Text-as-Image for Dual-grained Prompt Tuning", + "text_level": 1, + "bbox": [ + 498, + 446, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [41], a prompt is defined as:", + "bbox": [ + 517, + 469, + 777, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {i} = \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {v} _ {2}, \\boldsymbol {v} _ {3}, \\dots , \\boldsymbol {v} _ {M}, \\boldsymbol {s} _ {i} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 492, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i \\in \\{1, 2, \\dots, C\\}$ is the class index, $s_i$ denotes word embedding of the $i$ -th class name $s_i$ . For $j \\in \\{1, \\dots, M\\}$ , $v_j$ is a learnable word embedding whose dimension is the same as the dimension of normal word embeddings in the vocabulary. Just like in previous methods, e.g. CoOp [41], the prompts are learned by maximizing the probability of classifying each image into its ground-truth class:", + "bbox": [ + 496, + 516, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (y = i | \\boldsymbol {x}) = \\frac {\\exp \\left(\\left\\langle \\operatorname {E n c} _ {\\mathrm {T}} \\left(\\boldsymbol {t} _ {i}\\right) , \\operatorname {E n c} _ {\\mathrm {I}} (\\boldsymbol {x}) \\right\\rangle / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\left\\langle \\operatorname {E n c} _ {\\mathrm {T}} \\left(\\boldsymbol {t} _ {j}\\right) , \\operatorname {E n c} _ {\\mathrm {I}} (\\boldsymbol {x}) \\right\\rangle / \\tau\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 628, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ denotes the labeled training image and $\\langle \\cdot, \\cdot \\rangle$ calculates the cosine similarity.", + "bbox": [ + 496, + 672, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After large-scale image-text contrastive pre-training, text features have been well-aligned to the image features of the same semantic meanings. Thus, features of texts that describe a certain type of object also exhibit categorical discriminability. Therefore, based on the aligned VL representation, we advocate considering the feature of a piece of text description that describes a specific category, as an alternative to an image feature.", + "bbox": [ + 496, + 704, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from using the global sentence representation (i.e., the coarsest-grained text feature), we find that the sequential feature of word tokens from CLIP also possesses rich fine-grained information which is very similar to the region feature of dense image feature. In CLIP [24], cosine similarity", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2811", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88cf9e5f6a9ea7e84eb447ea37e472b63ce24d59a391cf18895f58773c023236.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 338, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/67a48dd6aa75a5946ce6d02959c4a313c72446431e6a2642ff4284badae7645a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 88, + 596, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/98b8d8e05a2fd435780dfac9dbcdce3af1365023f52c14438e582a47709ed25f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 88, + 885, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fdb9e86433bebd7832a4daf80317954a0149179108449c53364544a2258375b9.jpg", + "image_caption": [ + "Figure 3. Visualization of correlations $\\pmb{P}$ between the local class embedding $\\pmb{L}$ and sequential token feature from texts. Each class embedding clearly correlates to words that describe the corresponding class (shown in highlight regions) rather than the global $<\\mathrm{EOS}>$ token.", + "Figure 4. Visualization of correlations between the local class embedding $\\mathbf{L}$ and dense image feature. The learned class embeddings can focus on the location of the object effectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 310, + 467, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between global image features, obtained by visual attention pooling, and global text features, obtained by projecting the feature of the last $<\\mathrm{EOS}>$ token, are directly supervised with contrastive loss. In general, the global feature is sufficient for single-label classification because the target object usually is prominent in the picture. However, in multi-label recognition, the global feature is usually dominated by major objects, suppressing the recognition of non-significant objects concurrently existing in the image. Thus, it motivates us to explore fine-grained features and avoid the domination of the overly prominent object.", + "bbox": [ + 75, + 542, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To grasp the discriminability of global features as well as learn from fine-grained features, we propose double-grained prompt tuning (i.e., TaI-DPT) that uses two sets of prompts to handle global (i.e., the coarsest-grained level) and local (i.e., the fine-grained level) features, respectively, in two parallel branches. Formally, the double-grained prompt is defined as follows:", + "bbox": [ + 75, + 710, + 470, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {i} ^ {G} = \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {v} _ {2}, \\boldsymbol {v} _ {3}, \\dots , \\boldsymbol {v} _ {M}, \\boldsymbol {s} _ {i} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 823, + 468, + 847 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {t} _ {i} ^ {L} = \\left[ \\boldsymbol {v} _ {1} ^ {\\prime}, \\boldsymbol {v} _ {2} ^ {\\prime}, \\boldsymbol {v} _ {3} ^ {\\prime}, \\dots , \\boldsymbol {v} _ {M} ^ {\\prime}, \\boldsymbol {s} _ {i} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 845, + 372, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{v}_j$ and $\\pmb{v}_j^\\prime$ are learnable embeddings that are concatenated with word embedding $s_i$ of the $i$ -th class to obtain the", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "global prompt $t_i^G$ and local prompt $t_i^L$ , respectively. The sequences in Eq. (3) are fed to a copy of the text encoder EncT of CLIP to generate global and local class embeddings for each class, i.e. $G_{i} = \\mathrm{Enc}_{\\mathrm{T}}(t_{i}^{G})$ and $L_{i} = \\mathrm{Enc}_{\\mathrm{T}}(t_{i}^{L})$ , $G = \\{G_{i}\\}_{i=1}^{C}$ and $L = \\{L_{i}\\}_{i=1}^{C}$ are encouraged to be correlated with global and local features, respectively. Note that the proposed double-grained prompts are different from dual prompts [28], which include a pair of contrastive positive and negative prompts for each class (More discussion about the differences between our method and DualCoOp is provided in the Suppl).", + "bbox": [ + 496, + 311, + 892, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To preserve the fine-grained region features for the input image, we maintain the feature map before attention pooling layer of CLIP. As for the input text description, we preserve the sequential token features of the entire sentence instead of only the token features. So we have:", + "bbox": [ + 496, + 479, + 893, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{f, F \\right\\} = \\operatorname {E n c} _ {\\mathrm {I}} (x), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 564, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {h}, \\boldsymbol {H} \\right\\} = \\operatorname {E n c} _ {\\mathrm {T}} (\\boldsymbol {r}),\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 583, + 766, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{r}$ denotes a piece of training text description. $f, h \\in \\mathbb{R}^D$ are the extracted global image and text features. $\\pmb{F} \\in \\mathbb{R}^{N_1 \\times D}$ and $\\pmb{H} \\in \\mathbb{R}^{N_2 \\times D}$ are the flattened dense image features and sequential token features, respectively, where $N_1 = H \\times W$ denotes the flattened spatial dimension of visual feature and $N_2$ denotes the length of text tokens.", + "bbox": [ + 496, + 608, + 893, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the global and local similarities are computed by:", + "bbox": [ + 517, + 699, + 887, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {i} = \\left\\langle \\boldsymbol {u}, \\boldsymbol {G} _ {i} \\right\\rangle , \\boldsymbol {P} _ {i j} = \\left\\langle \\boldsymbol {U} _ {j}, \\boldsymbol {L} _ {i} \\right\\rangle \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 726, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{u}$ denotes either language feature $h$ in training or visual feature $f$ in testing, and $\\pmb{U}$ denotes $\\pmb{H}$ or $\\pmb{F}$ coordinately. Information in local branch $\\pmb{P}$ (visualized in Fig. 3 and Fig. 4) is aggregated in a spatially weighted manner:", + "bbox": [ + 496, + 755, + 893, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {i} ^ {\\prime} = \\sum_ {j = 1} ^ {N} \\frac {\\exp \\left(\\boldsymbol {P} _ {i j} / \\tau_ {s}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\boldsymbol {P} _ {i j} / \\tau_ {s}\\right)} \\cdot \\boldsymbol {P} _ {i j} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 824, + 890, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{s}$ accommodates the extent of focusing on a specific location. $\\pmb{p}_i$ and $\\pmb{p}_i^{\\prime}$ are optimized by the loss terms $\\mathcal{L}_{global}$", + "bbox": [ + 496, + 869, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c89f2e793150d87967aed74564435b54c080dab9e3f4e0253f7cad2359ce743.jpg", + "image_caption": [ + "Figure 5. Our learned double-grained prompt tuning is easy to combine with existing prompt tuning methods with ensemble." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 444, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_{local}$ , respectively, which we will discuss in Sec. 3.4. And in the testing phase, $\\pmb{p}$ and $\\pmb{p}^{\\prime}$ are combined to obtain the final classification score.", + "bbox": [ + 75, + 277, + 467, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The visualization in Fig. 3 and Fig. 4 show that the learned local class embedding $\\pmb{L}$ can focus on the specific location where corresponding classes appear, both in sentences and images, even if the fine-grained visual and language features are not explicitly supervised in the training of CLIP.", + "bbox": [ + 75, + 323, + 467, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Learning Objective", + "text_level": 1, + "bbox": [ + 76, + 426, + 261, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall learning objective is defined as $\\mathcal{L} = \\mathcal{L}_{global} + \\mathcal{L}_{local}$ , where $\\mathcal{L}_{global}$ and $\\mathcal{L}_{local}$ are loss terms for global text embedding and local text tokens, respectively. We adopt the ranking loss [14] to measure the discrepancy between classification scores and pseudo-ground-truth labels. Specifically, $\\mathcal{L}_{global}$ and $\\mathcal{L}_{local}$ are formulated as follows:", + "bbox": [ + 75, + 450, + 467, + 554 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {g l o b a l}} = \\sum_ {i \\in \\{c ^ {+} \\}} \\sum_ {j \\in \\{c ^ {-} \\}} \\max (0, m - p _ {i} + p _ {j}), \\tag {7} \\\\ \\mathcal {L} _ {l o c a l} = \\sum_ {i \\in \\{c ^ {+} \\}} \\sum_ {j \\in \\{c ^ {-} \\}} \\max (0, m - \\pmb {p} _ {i} ^ {\\prime} + \\pmb {p} _ {j} ^ {\\prime}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 561, + 467, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pmb{p}$ and $\\pmb{p}'$ are global and aggregated local similarities described in Sec. 3.3, $m$ is the margin controlling how much higher the similarity score with the positive classes is than with the negative classes. During training, we minimize the overall objective $\\mathcal{L}$ with frozen text encoders, by optimizing the global and local prompts.", + "bbox": [ + 75, + 627, + 467, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Binary cross-entropy loss and its variations such as asymmetric loss [25] have shown remarkable results in classification tasks [25, 28]. They are generally accompanied by a sigmoid function $\\sigma(x) = 1 / (1 + \\exp(-x))$ to convert model outputs to probabilities. Nevertheless, we found that directly optimizing the probability $\\sigma(p)$ leads to a performance gap between training texts and testing images. We attribute this phenomenon to the modality gap between vision and language. Thus we consider ranking loss to be a more flexible and suitable way of supervision in the presence of the modality gap. A comparison of results between different loss functions is provided in Suppl.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Incorporating with Prompting from Images", + "text_level": 1, + "bbox": [ + 500, + 90, + 870, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Though our TaI-DPT is very different from existing methods of prompting from images, it is also complementary to them. To show this, we utilize an off-the-shelf prompt ensemble strategy to combine our TaI-DPT with existing methods in this section. As illustrated in Fig. 5, using CoOp [41] as an example, we can simply combine the scores of CoOp [41] and that of our TaI-DPT in a weighted sum manner. In particular, our TaI-DPT can be integrated with CoOp [41] when a few annotated images are provided and integrated with DualCoOp [28] when partially labeled training data are available.", + "bbox": [ + 496, + 113, + 890, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We ensemble prompts by fusing the predicted scores, rather than averaging the class embeddings generated by different prompts, since the image encoder used in different methods may be different (e.g. we conduct our experiments with ResNet50, while DualCoOp uses ResNet101 for partial-label prompting). So ensembling with the classification score is more convenient. In Sec. 4.3, we also empirically show that our prompt ensemble strategy is effective in advancing multi-label recognition performance in the few-shot and partially labeled settings.", + "bbox": [ + 496, + 279, + 890, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 444, + 630, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 469, + 715, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture. We adopt CLIP ResNet-50 [24] as the visual encoder, and use the CLIP Transformer as the text encoder. During training, the parameters of the two encoders are kept frozen, and only learnable prompts are optimized.", + "bbox": [ + 496, + 492, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Learnable Prompts. Our learnable prompts are shared among classes of all datasets. Class-specific prompting [41] (i.e., an individual set of parameters for each category) has also been explored, but brings limited benefits. We initialize the value of each parameter with the Gaussian noise sampled from $\\mathcal{N}(0,0.02)$ . The length of both the global prompts and local prompts are set to $M = 16$ , while a longer sequence brings trivial improvements.", + "bbox": [ + 496, + 553, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To evaluate our TaI-DPT, we conduct the experiments on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9]. VOC2007 contains 20 common categories, and following [6,8,28], we form the training/test set based on the official trainval/test split (5,011 images/4,952 images). MS-COCO includes 80 categories with 82,081 training images and 40,504 testing images. NUS-WIDE includes 81 concepts with 161,789 images for training. We adopt its test set (107,859 images) to evaluate our method. For zero-shot experiments in Sec. 4.2, the training sets of the datasets are not used, and we use only text data to learn the prompts as mentioned in Sec. 3.2. Besides, for VOC2007 and MS-COCO, the language data sources are captions from MS-COCO. For NUS-WIDE, we introduce localized narratives from OpenImages [18], which have a", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/725a6d4e13d14b027f1b2c58fe29684ae1fc057481359f891b8aebc29dcd038e.jpg", + "image_caption": [ + "Figure 6. Comparison of different methods in few-shot multi-label recognition on VOC2007 and MS-COCO. Our zero-shot TaI-DPT can achieve comparable results with methods trained by 16-shot labeled image samples. And learned prompt ensemble proofs the complementarity between images and texts." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 344, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc669f823ee0ce155b7f9f1bd2ff8dabceb87e0c74ceba00b4f9e1aff0ba5787.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 90, + 599, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb79468ea7ab2db108586757738f813ff4e1b9141864298059bd984958c7985b.jpg", + "image_caption": [ + "Figure 7. Ablation experiment on the number of texts and performance of TaI prompting on VOC2007." + ], + "image_footnote": [], + "bbox": [ + 632, + 90, + 887, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aecca2847ceff10c7b96bc3591ee4089b79baa5d8543ec765c0a947bfbee95bc.jpg", + "table_caption": [ + "Table 1. Comparison with zero-shot methods on VOC2007, MS-COCO, and NUS-WIDE. Our proposed TaI-DPT outperforms CLIP [24] by a large margin on all datasets." + ], + "table_footnote": [], + "table_body": "
MethodDPTVOC2007MS-COCONUSWIDE
ZSCLIPX76.247.336.4
✓77.349.737.4
TaIX86.061.144.9
✓88.365.146.5
", + "bbox": [ + 78, + 357, + 468, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "broader range of content, to cover all the concepts in NUS-WIDE. In Sec. 4.3 and Sec. 4.4, for each dataset, the corresponding training data is used to conduct the experiments of partial-label and few-shot multi-label classification.", + "bbox": [ + 75, + 459, + 468, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Details. We adopt SGD optimizer, and the training epoch is set to 20 for all datasets. The learning rates for MS-COCO, VOC2007, and NUS-WIDE are empirically initialized with 1e-4, 1e-4, and 1e-3, and decay by the cosine annealing rule. For ranking loss, we choose $m = 1$ , and scale the $p$ and $p'$ by a factor of 4. $\\tau_{s}$ is set as 0.02 via validation.", + "bbox": [ + 75, + 522, + 468, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with Zero-Shot Methods", + "text_level": 1, + "bbox": [ + 76, + 647, + 398, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed TaI and DPT, we first compare it with the zero-shot CLIP (ZSCLIP). For fair comparison, we also introduce the DPT to ZSCLIP. Specifically, we adopt two identical default prompts \"a photo of a [CLASS]\" to separately deal with global and local features as DPT does. Table 1 lists the comparison results on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9] datasets. From the table, our TaI prompting surpasses ZSCLIP by a large margin of $9.8\\%$ , $13.8\\%$ , and $8.5\\%$ mAP on VOC2007, MS-COCO, and NUS-WIDE, respectively, showing the effectiveness of our TaI. Furthermore, after training with fine-grained token features extracted from texts, our proposed DPT demonstrates a more powerful capability of discriminating local features than the default hand-crafted prompts and single global prompts.", + "bbox": [ + 75, + 672, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c41d6a9e80ac4e3fee010fbd716e051d81069b25017646529a68fc133888de7e.jpg", + "table_caption": [ + "Table 2. Comparison with existing multi-label few-shot learning methods on MS-COCO. The evaluation is based on mAP for zero-shot, 1-shot and 5-shot with 16 novel classes." + ], + "table_footnote": [], + "table_body": "
Method0-shot1-shot5-shot
LaSO [2]-45.358.1
ML-FSL [27]-54.463.6
CoOp [41]40.2 (ZSCLIP)46.955.6
Tip-Adapter [38]40.2 (ZSCLIP)53.859.7
TaI-DPT59.2--
", + "bbox": [ + 504, + 354, + 879, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with Few-Shot Methods", + "text_level": 1, + "bbox": [ + 500, + 478, + 816, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further compare with multi-label few-shot learning methods to verify the effectiveness of our TaI-DPT. In contrast to the well-studied single-label few-shot classification problem, few works tackle the multi-label few-shot scenario. Existing methods [2, 27] often deploy models trained on seen classes to few-shot novel classes. In Table 2, we compare zero-shot TaI-DPT to few-shot methods on 16 novel classes (see [2] for details about data split). TaI-DPT is comparable to the methods trained on 5-shot samples.", + "bbox": [ + 496, + 503, + 890, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, we consider a new multi-label few-shot setting where all the classes are regarded as novel classes. We select 1, 2, 4, 8, and 16-shot samples for each category following the strategy in [2]. For fair comparison, we train CoOp [41] and our TaI in the same settings, and we also extend them with DPT for a more comprehensive comparison. For CoOp-DPT, we set two sets of learnable prompts, to deal with global and local features, respectively. The results are illustrated in Fig. 6. One can see that, even without any image data regarding novel classes, our TaI can achieve comparable results to CoOp trained on 16-shot. Similar trends with the MS-COCO dataset and the DPT setting support our observation that the discriminative feature of text data can be used as images for prompting. Moreover, benefiting from the flexibility of prompts, we can easily integrate our TaI-DPT with CoOp-DPT by ensembling as said in Sec. 4.3. As illustrated in Fig. 6, though CoOp-DPT has", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f585208bfc47bd7da28bae26bb4670baef9fbef3f77f3c872dc3352a30087b79.jpg", + "table_caption": [ + "Table 3. Results of integrating our TaI-DPT with partial-label multi-label recognition method based on pre-trained CLIP. Our approach further improves the frontier performance of DualCoOp [28]. * indicates the results based on our own reproduction." + ], + "table_footnote": [], + "table_body": "
DatasetsMethod10%20%30%40%50%60%70%80%90%Avg.
MS-COCOSARB [23]71.275.077.178.378.979.679.880.580.577.9
DualCoOp [28]78.780.981.782.082.582.782.883.083.181.9
DualCoOp*81.082.382.983.483.583.984.084.184.383.3
+TaI-DPT81.582.683.383.783.984.084.284.484.583.6
PascalVOC 2007SARB [23]83.588.690.791.491.992.292.692.892.990.7
DualCoOp [28]90.392.292.893.393.693.994.094.194.293.2
DualCoOp*91.493.893.894.394.694.794.894.994.994.1
+TaI-DPT93.394.694.894.995.195.095.195.395.594.8
NUS-WIDEDualCoOp*54.056.256.957.457.957.957.658.258.857.2
+TaI-DPT56.457.957.858.158.558.858.659.159.458.3
", + "bbox": [ + 93, + 121, + 867, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieved a high accuracy, combining our prompts learned with text data still brings further improvement on recognition performance. This also proves that texts and images are complementary to each other to some extent.", + "bbox": [ + 75, + 323, + 470, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Integration with Partially Labeled Methods", + "text_level": 1, + "bbox": [ + 76, + 397, + 447, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Partially-labeled data refers to the problem where merely some labels of each sample are known. Following [5], we created the partial-label data by randomly masking out labels of the fully annotated data. During inference, the model is evaluated on all categories.", + "bbox": [ + 75, + 421, + 468, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We reproduce DualCoOp [28] on partial-labeled VOC2007 and MS-COCO with the same experimental setting as reported (reproduced results are marked with *) and explore the enhancement brought by integration with our TaI-DPT. We also deploy DualCoOp on partial-labeled NUS-WIDE for a comprehensive comparison.", + "bbox": [ + 75, + 498, + 468, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are reported in Table 3. With no prior knowledge from pre-trained models, previous forefront method like SARB [23] struggles to learn from incomplete labels. DualCoOp, directly trained with entire image set, is well-learned to classify multi-label images. Even so, for PascalVOC and NUS-WIDE, by using texts from exogenous sources, i.e., COCO and OpenImages, our method can bring complementary enhancement to image prompting. While, for COCO, we used captions derived from its own images, making it more difficult to provide additional benefits. Analogous to VOC and NUS-WIDE, the COCO results can also be further advanced by introducing more external text data in future work.", + "bbox": [ + 75, + 590, + 470, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 800, + 230, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To thoroughly investigate the effect of each component, we conduct a series of ablation studies on the quantity of texts, training loss, different VL pre-trained models, ensemble weights, and texts v.s. images for prompting. The complete ablation experiments are shown in the Suppl.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantity of texts. Following the data preparation procedure in Sec. 3.2, we end up with a total number of 66087 pieces of text that contain descriptions for 20 categories of VOC2007. We test the performance of TaI-DPT with different numbers of randomly selected texts, and the results are shown in Fig. 7. When no collected texts are available, 80 templates of hand-crafted prompts from [24], like \"a cropped photo of a [CLASS]\", are used for training (all templates are shown in the Suppl), and each template sentence correlates with one positive label corresponding to the class name inserted in [CLASS]. The increasing number of texts gradually forms a complete description of target categories, and the relationship between classes is also better characterized, which results in ascending performance.", + "bbox": [ + 496, + 323, + 893, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 554, + 619, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new view of treating texts as images in prompt tuning (i.e. TaI), which learns the prompt from discriminative features of text descriptions. Compared to prior prompt tuning methods trained with images, our TaI benefits from the easy accessibility of scalable content-rich texts, which enables prompt tuning for vision tasks (e.g., multi-label image recognition) even without downstream image data. Double-grained prompting is further introduced to utilize both the global and fine-grained features for better multi-label recognition ability. Nonetheless, when few-shot image samples or partial-labeled images are available, our TaI-DPT can conveniently integrate with existing prompting methods. Experiments on multiple benchmarks show the validity of TaI-DPT.", + "bbox": [ + 496, + 582, + 893, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 813, + 660, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by National Key R&D Program of China under Grant No. 2020AAA0104500, and by the National Natural Science Foundation of China (NSFC) under Grant No. U19A2073.", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2815", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1", + "[2] Amit Alfassy, Leonid Karlinsky, Amit Aides, Joseph Shtok, Sivan Harary, Rogerio Feris, Raja Giryes, and Alex M Bronstein. Laso: Label-set operations networks for multi-label few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6548-6557, 2019. 2, 7", + "[3] Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification, 2021. 2", + "[4] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit.\" O'Reilly Media, Inc.\", 2009. 4", + "[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 339-346, 2022. 2, 8", + "[6] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 522-531, 2019. 2, 6", + "[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME 2019. 2", + "[8] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In CVPR, 2019. 2, 6", + "[9] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In Proceedings of the ACM international conference on image and video retrieval, pages 1-9,2009.2,6,7", + "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 2", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 6, 7", + "[12] Bin-Bin Gao and Hong-Yu Zhou. Multi-label image recognition with multi-class attentional regions. arXiv preprint arXiv:2007.01755, 2020. 2", + "[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 3", + "[14] Yunchao Gong, Yangqing Jia, Thomas Leung, Alexander Toshev, and Sergey Ioffe. Deep convolutional ranking for mul" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tilabel image annotation. arXiv preprint arXiv:1312.4894, 2013. 2, 6", + "[15] Shiyi He, Chang Xu, Tianyu Guo, Chao Xu, and Dacheng Tao. Reinforced multi-label image classification by exploring curriculum. In AAAI, 2018. 2", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1", + "[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 3", + "[18] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Shahab Kamali, Matteo Malloci, Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://storage.googleapis.com/openimages/web/index.html, 2017. 2, 4, 6", + "[19] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 4, 6, 7", + "[21] Luchen Liu, Sheng Guo, Weilin Huang, and Matthew Scott. Decoupling category-wise independence and relevance with self-attention for multi-label image classification. In ICASSP 2019, 05. 2", + "[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3", + "[23] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. arXiv preprint arXiv:2203.02172, 2022. 2, 8", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 6, 7, 8", + "[25] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2816", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. arXiv preprint arXiv:2209.07511, 2022. 3", + "[27] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. Meta-learning for multi-label few-shot classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3951-3960, 2022. 2, 7", + "[28] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. arXiv preprint arXiv:2206.09541, 2022. 1, 2, 3, 5, 6, 8", + "[29] Ashwin Vaswani, Gaurav Aggarwal, Praneeth Netrapalli, and Narayan G Hegde. All mistakes are not equal: Comprehensive hierarchy aware multi-label predictions (champ). arXiv preprint arXiv:2206.08653, 2022. 2", + "[30] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, 2016. 2", + "[31] Yangtao Wang, Yanzhao Xie, Yu Liu, Ke Zhou, and Xiaocui Li. Fast graph convolution network based multi-label image recognition via cross-modal fusion. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 1575-1584, 2020. 2", + "[32] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE transactions on pattern analysis and machine intelligence, 38, 2015. 2", + "[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1", + "[34] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3", + "[35] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, 2020. 2", + "[36] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1", + "[37] Junjie Zhang, Qi Wu, Chunhua Shen, Jian Zhang, and Jianfeng Lu. Multilabel image classification with regional latent semantic dependencies. IEEE Transactions on Multimedia, 20, 2018. 2", + "[38] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022. 7", + "[39] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3", + "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 4, 6, 7", + "[42] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 260 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2817", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_model.json b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f3196d3c5902c532f1bef377e4259888310fa1b1 --- /dev/null +++ b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "Texts as Images in Prompt Tuning for Multi-Label Image Recognition" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.183, + 0.777, + 0.199 + ], + "angle": 0, + "content": "Zixian Guo\\(^{1,2*}\\) Bowen Dong\\(^{1}\\) Zhilong Ji\\(^{2}\\) Jinfeng Bai\\(^{2}\\) Yiwen Guo\\(^{4}\\) Wangmeng Zuo\\(^{1,2(\\text{图})}\\)" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.201, + 0.838, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Harbin Institute of Technology \\(^{2}\\)Tomorrow Advancing Life \\(^{3}\\)Pazhou Lab, Guangzhou \\(^{4}\\)Independent Researcher" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.22, + 0.782, + 0.234 + ], + "angle": 0, + "content": "zixian.guo@foxmail.com cndongsky@gmail.com zhilongji@hotmail.com" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.238, + 0.746, + 0.251 + ], + "angle": 0, + "content": "jfbai.bit@gmail.com guoyiwen89@gmail.com WMzuo@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.669 + ], + "angle": 0, + "content": "Prompt tuning has been employed as an efficient way to adapt large vision-language pre-trained models (e.g. CLIP) to various downstream tasks in data-limited or label-limited settings. Nonetheless, visual data (e.g., images) is by default prerequisite for learning prompts in existing methods. In this work, we advocate that the effectiveness of image-text contrastive learning in aligning the two modalities (for training CLIP) further makes it feasible to treat texts as images for prompt tuning and introduce TaI prompting. In contrast to the visual data, text descriptions are easy to collect, and their class labels can be directly derived. Particularly, we apply TaI prompting to multi-label image recognition, where sentences in the wild serve as alternatives to images for prompt tuning. Moreover, with TaI, double-grained prompt tuning (TaI-DPT) is further presented to extract both coarse-grained and fine-grained embeddings for enhancing the multi-label recognition performance. Experimental results show that our proposed TaI-DPT outperforms zero-shot CLIP by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE, while it can be combined with existing methods of prompting from images to improve recognition performance further. The code is released at https://github.com/guozix/Tai-DPT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Recent few years have witnessed rapid progress in large vision-language (VL) pre-trained models [1, 16, 19, 24, 33, 36] as well as their remarkable performance on downstream vision tasks. A VL pre-trained model generally involves data encoders, and it is becoming increasingly popular to exploit image-test contrastive loss [24] to align the embedding of images and texts into a shared space. When adapting to downstream tasks in data-limited or label-limited settings, it is often ineffective to fine-tune the entire model, due to its high complexity. Then, prompt tuning as a representative parameter-efficient learning paradigm has emerged as" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.29, + 0.88, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.499, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1. A comparison between prompting from images and our text-as-image (TaI) prompting. (a) Prompting from images (e.g., [41]) uses labeled images of task categories to learn the text prompts. Instead, (b) our TaI prompting learn the prompts with easily-accessed text descriptions containing target categories. (c) After training, the learned prompts in (a) or (b) can be readily applied to test images." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.61, + 0.883, + 0.625 + ], + "angle": 0, + "content": "an efficient way to adapt VL models to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Albeit considerable achievements have been made, existing prompt tuning methods generally require visual data to learn prompts (as shown in Fig. 1(a)). For example, CoOp [41] learns from annotated images. CoCoOp [40] further introduces generalizable input-conditional prompts. DualCoOp [28] adapts CLIP to multi-label recognition tasks by training pairs of positive and negative prompts with partial-labeled images. Nonetheless, the performance of these prompting methods may be limited when it is infeasible to obtain sufficient image data or annotate the required images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we advocate treating Texts as Images for prompt tuning, i.e., TaI prompting. It is considered feasible as the image encoder and text encoder in many pre-trained VL models [16, 24] encode images and texts into a shared space. Given an image and its caption, the visual features produced by the image encoder will be close to the text feature of the caption produced by the text encoder. There-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.462, + 0.9 + ], + "angle": 0, + "content": "*This work was done when Zixian Guo was a research intern at TAL." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2808" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "fore, in addition to extracting visual features from images, it is also feasible to extract text features as alternatives form, for example, descriptive sentences and captions, for prompt tuning (see Fig. 1(b)). TaI prompting has several interesting properties and merits. Taking a downstream image recognition task as an example, given a set of object categories, one can easily crawl a large set of text descriptions that contain object names from these categories. Text descriptions are easily accessible in this way, and class labels can be directly derived from text descriptions, which means, in contrast to prompting from images, TaI prompting may suffer less from the data-limited and label-limited issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.635 + ], + "angle": 0, + "content": "We use multi-label image recognition [8, 9, 11, 20, 35] to verify the effectiveness of our TaI prompting in this paper. To begin with, we crawl the captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., Open Images [18]) to form the training set of text descriptions. For any specific multi-label recognition task, we adopt a noun filter to map the nouns in the text descriptions to the corresponding object categories, and then only keep the text descriptions that contain one or more classes of target objects. To better cope with multi-label classification, we introduce double-grained prompt tuning (i.e., TaI-DPT) which involves: (i) a set of global prompts to generate embeddings for classifying whole sentences or images, and (ii) a set of local prompts to extract embeddings for discriminating text tokens or image patches. Given a set of text descriptions, global and local prompts can be tuned by minimizing the ranking loss [14]. Note that, though these prompts are learned from text descriptions solely, they can be readily deployed to classify whole images as well as image patches during testing (see Fig. 1(c)). Experimental results show that, without using any labeled images, our TaI prompting surpasses zero-shot CLIP [24] by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.473, + 0.803 + ], + "angle": 0, + "content": "Moreover, when images are also available during training, our TaI prompting can be combined with existing methods of prompting from images to improve its performance. In particular, given a few annotated images, our TaI-DPT can be integrated with CoOp as a prompt ensemble for improving classification accuracy. With partially labeled training data being provided, we may also combine TaI-DPT and DualCoOp [28] to improve multi-label recognition accuracy consistently. Extensive results verify the effectiveness of our TaI-DPT in comparison to state-of-the-art. To sum up, the contributions of this work include:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.811, + 0.472, + 0.902 + ], + "angle": 0, + "content": "- We propose Texts as Images in prompt tuning (i.e., TaI prompting) to adapt VL pre-trained models to multi-label image recognition. Text descriptions are easily accessible and, in contrast to images, their class labels can be directly derived, making our TaI prompting very compelling in practice." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "- We present double-grained prompt tuning (i.e. TaI-DPT) to extract both coarse-grained and fine-grained embeddings for enhancing multi-label image recognition. Experiments on multiple benchmarks show that TaI-DPT achieves comparable multi-label recognition accuracy against state-of-the-arts." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.183, + 0.894, + 0.243 + ], + "angle": 0, + "content": "- The prompts learned by TaI-DPT can be easily combined with existing methods of prompting from images in an off-the-shelf manner, further improving multi-label recognition performance." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.894, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.249, + 0.642, + 0.265 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.274, + 0.779, + 0.29 + ], + "angle": 0, + "content": "2.1. Multi-Label Image Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Multi-label image recognition [3,7,8,12,15,21,32,35,37] aims to recognize all the object categories [11, 20] or concepts [9] in an input image. Various modules [6, 30] have been introduced to better represent the inter-class relationships, and modern classification losses [3, 14] have been used to make model learning easier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.893, + 0.614 + ], + "angle": 0, + "content": "To model the label dependencies, CNN-RNN [30] introduces recurrent neural networks, e.g., RNN and LSTM, to predict appeared classes in a sequential manner. [6,8,31,35] use graph convolution modules to learn the correlation between class labels. CHAMP [29] measures the severity of misclassification by building a domain-specific hierarchy tree according to the relation of categories, where each class is related to a tree node, to improve the robustness of the model. Albeit effective, these methods require a considerable number of labeled images to learn the category relationships sufficiently. While in data-limited or label-limited regimes, e.g., few-shot or partial-label data, it will be difficult for these models to learn well as expected. Specifically designed loss functions also struggle to obtain significant improvements when learning with limited data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Multi-Label Recognition from Few-shot Samples. To better exploit the small number of samples, LaSO [2] synthesizes samples by manipulating the features of paired training images. Different ways of manipulating label sets are used to train the model, resulting in generalizable discriminative features. [27] introduces a meta-learning framework for better learning of past tasks and generalization to new tasks, and leverages the number of labels as useful information for learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Multi-Label Recognition from Partial-label Data. Partial-label refers to the scenarios where some labels are unknown. [10] propose a normalized BCE loss to balance the proportion of known labels. [5] learns to complement unknown labels by utilizing within-image and cross-image semantic correlations. [23] blends the representation of training images and class proxies to compensate for the loss of information due to unknown labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Albeit significant progress has been made, it remains a challenging issue for learning multi-label image recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.452, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.293, + 0.276, + 0.303 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.454, + 0.09, + 0.887, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.292, + 0.68, + 0.303 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Training and testing pipeline of our proposed Text-as-Image (TaI) prompting, where we use text descriptions instead of labeled images to train the prompts. (a) During training, we use two identical text encoders from pre-trained CLIP to extract the global & local class embeddings \\((\\mathbf{G} \\& \\mathbf{L})\\) and overall & sequential text embeddings \\((\\mathbf{h} \\& \\mathbf{H})\\) respectively from the prompts and text description. The corresponding cosine similarity \\((\\mathbf{p} \\& \\mathbf{P})\\) between the embeddings are guided by the derived pseudo labels with ranking loss. (b) During testing, we replace the input from text descriptions to images. The global and local class embeddings can discriminate target classes from global & local image features \\((\\mathbf{f} \\& \\mathbf{F})\\). The final classification results are obtained by merging the scores of the two branches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.475 + ], + "angle": 0, + "content": "in image-limited or label-limited regimes. Built upon VL pre-trained models, this paper suggests learning prompts from texts instead of images, thereby offering a novel yet complementary perspective for handling low-resource multi-label image recognition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.453, + 0.501 + ], + "angle": 0, + "content": "2.2. Prompt Tuning for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.794 + ], + "angle": 0, + "content": "To transfer pre-trained knowledge to downstream tasks in data-limited settings, prompt tuning [13, 17, 34, 39, 41, 42] has become a popular parameter-efficient way to achieve the goal, due to its flexibility and ease of use. CoOp [41] learns the prompts by using (a few) annotated images of each class from the target dataset. CoCoOp [40] further proposes to improve CoOp [41] by formulating the prompts in an image-conditional way to maintain better generalization to unseen classes. To avoid overfitting, ProGrad [42] leverages predictions from zero-shot CLIP to regularize gradients in prompt learning process. TPT [26] suggests optimizing test-time prompts by promoting the consistency of augmented test images. ProDA [22] uses multiple pieces of prompts to estimate the distribution of classifier weights for better handling of varying visual features. DualCoOp [28] firstly adapts CLIP to multi-label image recognition with partially labeled data by learning pairs of positive and negative prompts for each class to ensure independent binary classification for each class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Albeit existing prompt tuning approaches have achieved significant improvements in downstream tasks, images as well as a portion of class labels are prerequisite to supervise the optimization of the learnable prompts. In this paper, we propose to treat texts as images in prompt tuning, which, compared to labeled images, are much easier to collect with existing caption datasets and modern search engines. Our" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.46 + ], + "angle": 0, + "content": "proposed TaI-DPT surpasses zero-shot CLIP by a large margin and can be combined with the prompts learned by existing methods of prompting from images to further boost recognition performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.673, + 0.494 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Here we present Text-as-Image prompting, i.e., TaI prompting, for adapting pre-trained VL models to multi-label image recognition. Our TaI prompting uses only easily-accessed free-form texts as training data to learn effective prompts for downstream multi-label recognition tasks. To begin with, We present an overview of TaI prompting in Sec. 3.1. Then, we introduce our preparation of training texts in Sec. 3.2. We further explain the design of the double-grained prompt tuning (i.e., TaI-DPT) in Sec. 3.3, and provide the loss function used to train the model in Sec. 3.4. Finally, we propose to ensemble prompts from TaI-DPT with prompts learned from images as a flexible integration to improve multi-label recognition performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.73, + 0.725 + ], + "angle": 0, + "content": "3.1. Overview of Our Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 illustrates the design of our proposed TaI-DPT framework, including the training and testing phases. During training, we learn prompts with only supervision from texts. Two identical copies of the text encoder \\(\\mathrm{Enc}_{\\mathrm{T}}\\) from the pre-trained CLIP are used to encode the prompts and text data, respectively. We introduce two sorts of trainable prompts (i.e., the global prompts and local prompts) to obtain global and regional class embeddings. A noun filtering strategy is used to generate classification pseudo-labels for each text description, which is applied to supervise the classification scores obtained by calculating the cosine sim" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2810" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "iliarity of class embeddings and text features. Only the parameters in prompts are optimized in the training phase, while the text encoders are both kept frozen. During testing, the class embeddings are obtained by encoding the two sets of learned prompts with the text encoder \\(\\mathrm{Enc}_{\\mathrm{T}}\\) as in training, while the other input source changes from text descriptions to test images. Pre-trained image encoder \\(\\mathrm{Enc}_{\\mathrm{I}}\\) from CLIP is used to extract global and dense features of each test image, then the cosine similarity are computed between features and class embeddings of the global and local prompts. The final classification result is obtained by fusing the global and local cosine scores. In the following, we explain the details of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.296, + 0.364, + 0.312 + ], + "angle": 0, + "content": "3.2. Preparation of Text Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.56 + ], + "angle": 0, + "content": "To obtain sufficient category information from the language that helps in image recognition, we have to ensure that: 1) the collected texts should contain content-rich descriptions of a scene, and 2) the contents of all text descriptions need to cover the category set of the target dataset for comprehensive representations of all categories. With the aim of ensuring reproducibility, we use readily available captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., OpenImages [18]) as our language data source, while avoiding the workloads associated with randomly crawling texts from the Internet in this paper. Note that although each caption is paired with a corresponding image and human-annotated labels, we only use the captions, and no information from the pictures and labels is disclosed during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.471, + 0.796 + ], + "angle": 0, + "content": "For a target multi-label recognition dataset \\(\\mathcal{X}\\) that has a category set \\(S = \\{\\mathrm{s}_1, \\mathrm{s}_2, \\mathrm{s}_3, \\dots, \\mathrm{s}_C\\}\\), where \\(C\\) denotes the number of categories and \\(\\mathrm{s}_i\\) denotes particular class name like \"dog\", \"plane\", etc., we search for sentences that contain at least one class name \\(\\mathrm{s}_i\\) in \\(S\\). Since multiple words or phrases usually exist to represent the same meaning for each class, searching solely for exact match of category names in texts may lead to many false negatives in the obtained pseudo ground-truth labels, which is harmful to prompt tuning. Towards tackling this issue, we introduce a noun filter to map nouns with similar meanings into the corresponding class label. Specifically, we construct a synonym dictionary \\(\\mathcal{D}\\) by including common synonyms of each class name in the target dataset. If a word in a text description matches any synonym of a specific class name, it is considered to contain a description of that category. Several examples of synonyms are shown as follows:" + }, + { + "type": "code", + "bbox": [ + 0.096, + 0.803, + 0.462, + 0.884 + ], + "angle": 0, + "content": "{'dog', 'pup', 'puppy', 'doggy'} \n{'person', 'people', 'man', 'woman', 'human'} \n{'bicycle', 'bike', 'cycle'} \n{'car', 'taxi', 'automobile'} \n{'boat', 'raft', 'dinghy'} \n..." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.465, + 0.902 + ], + "angle": 0, + "content": "More details of the dictionary \\(\\mathcal{D}\\) are provided in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Then we conduct noun filtration by the following steps. First, for each text description, we use the tokenizer and lemmatizer from NLTK [4] to recover the stem of each word in the sentences. Next, for all keywords in \\(\\mathcal{D}\\), which contains all synonyms of the category set \\(S\\), we search in our language data source for sentences that contains at least one class name. For the text descriptions that do not match any synonym of any class name, we simply drop them away to ensure each piece of data has at least one concerned label. Finally, for each retained text description, we convert the class names it contains into binary pseudo-ground-truth vectors by setting classes that appear as positive and other classes as negative, following the order of class labels in the target dataset \\(\\mathcal{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.893, + 0.44 + ], + "angle": 0, + "content": "The word-level filtered labels may not be precisely correct since our searching strategy mentioned above is rather simple considering the diversity of free-form texts, where complex paraphrases and misspellings that widely exist in the corpus are not fully addressed. However, such a simple noun filtration can guarantee the reproducibility of this work, and our experiments also demonstrate that this simple and efficient data preparation leads to compelling results of multi-label recognition of our TaI prompting." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.463 + ], + "angle": 0, + "content": "3.3. Text-as-Image for Dual-grained Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.47, + 0.778, + 0.485 + ], + "angle": 0, + "content": "Following [41], a prompt is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.493, + 0.892, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {i} = \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {v} _ {2}, \\boldsymbol {v} _ {3}, \\dots , \\boldsymbol {v} _ {M}, \\boldsymbol {s} _ {i} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.624 + ], + "angle": 0, + "content": "where \\( i \\in \\{1, 2, \\dots, C\\} \\) is the class index, \\( s_i \\) denotes word embedding of the \\( i \\)-th class name \\( s_i \\). For \\( j \\in \\{1, \\dots, M\\} \\), \\( v_j \\) is a learnable word embedding whose dimension is the same as the dimension of normal word embeddings in the vocabulary. Just like in previous methods, e.g. CoOp [41], the prompts are learned by maximizing the probability of classifying each image into its ground-truth class:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.63, + 0.892, + 0.668 + ], + "angle": 0, + "content": "\\[\np (y = i | \\boldsymbol {x}) = \\frac {\\exp \\left(\\left\\langle \\operatorname {E n c} _ {\\mathrm {T}} \\left(\\boldsymbol {t} _ {i}\\right) , \\operatorname {E n c} _ {\\mathrm {I}} (\\boldsymbol {x}) \\right\\rangle / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\left\\langle \\operatorname {E n c} _ {\\mathrm {T}} \\left(\\boldsymbol {t} _ {j}\\right) , \\operatorname {E n c} _ {\\mathrm {I}} (\\boldsymbol {x}) \\right\\rangle / \\tau\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.891, + 0.704 + ], + "angle": 0, + "content": "where \\( \\pmb{x} \\) denotes the labeled training image and \\( \\langle \\cdot, \\cdot \\rangle \\) calculates the cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.825 + ], + "angle": 0, + "content": "After large-scale image-text contrastive pre-training, text features have been well-aligned to the image features of the same semantic meanings. Thus, features of texts that describe a certain type of object also exhibit categorical discriminability. Therefore, based on the aligned VL representation, we advocate considering the feature of a piece of text description that describes a specific category, as an alternative to an image feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Apart from using the global sentence representation (i.e., the coarsest-grained text feature), we find that the sequential feature of word tokens from CLIP also possesses rich fine-grained information which is very similar to the region feature of dense image feature. In CLIP [24], cosine similarity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.339, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.089, + 0.597, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.089, + 0.887, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.893, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Visualization of correlations \\( \\pmb{P} \\) between the local class embedding \\( \\pmb{L} \\) and sequential token feature from texts. Each class embedding clearly correlates to words that describe the corresponding class (shown in highlight regions) rather than the global \\( <\\mathrm{EOS}> \\) token." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.311, + 0.468, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.486, + 0.471, + 0.528 + ], + "angle": 0, + "content": "Figure 4. Visualization of correlations between the local class embedding \\( \\mathbf{L} \\) and dense image feature. The learned class embeddings can focus on the location of the object effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.711 + ], + "angle": 0, + "content": "between global image features, obtained by visual attention pooling, and global text features, obtained by projecting the feature of the last \\(<\\mathrm{EOS}>\\) token, are directly supervised with contrastive loss. In general, the global feature is sufficient for single-label classification because the target object usually is prominent in the picture. However, in multi-label recognition, the global feature is usually dominated by major objects, suppressing the recognition of non-significant objects concurrently existing in the image. Thus, it motivates us to explore fine-grained features and avoid the domination of the overly prominent object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.471, + 0.817 + ], + "angle": 0, + "content": "To grasp the discriminability of global features as well as learn from fine-grained features, we propose double-grained prompt tuning (i.e., TaI-DPT) that uses two sets of prompts to handle global (i.e., the coarsest-grained level) and local (i.e., the fine-grained level) features, respectively, in two parallel branches. Formally, the double-grained prompt is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.824, + 0.47, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {i} ^ {G} = \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {v} _ {2}, \\boldsymbol {v} _ {3}, \\dots , \\boldsymbol {v} _ {M}, \\boldsymbol {s} _ {i} \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.847, + 0.373, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {t} _ {i} ^ {L} = \\left[ \\boldsymbol {v} _ {1} ^ {\\prime}, \\boldsymbol {v} _ {2} ^ {\\prime}, \\boldsymbol {v} _ {3} ^ {\\prime}, \\dots , \\boldsymbol {v} _ {M} ^ {\\prime}, \\boldsymbol {s} _ {i} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_j\\) and \\(\\pmb{v}_j^\\prime\\) are learnable embeddings that are concatenated with word embedding \\(s_i\\) of the \\(i\\)-th class to obtain the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.893, + 0.48 + ], + "angle": 0, + "content": "global prompt \\( t_i^G \\) and local prompt \\( t_i^L \\), respectively. The sequences in Eq. (3) are fed to a copy of the text encoder EncT of CLIP to generate global and local class embeddings for each class, i.e. \\( G_{i} = \\mathrm{Enc}_{\\mathrm{T}}(t_{i}^{G}) \\) and \\( L_{i} = \\mathrm{Enc}_{\\mathrm{T}}(t_{i}^{L}) \\), \\( G = \\{G_{i}\\}_{i=1}^{C} \\) and \\( L = \\{L_{i}\\}_{i=1}^{C} \\) are encouraged to be correlated with global and local features, respectively. Note that the proposed double-grained prompts are different from dual prompts [28], which include a pair of contrastive positive and negative prompts for each class (More discussion about the differences between our method and DualCoOp is provided in the Suppl)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.894, + 0.555 + ], + "angle": 0, + "content": "To preserve the fine-grained region features for the input image, we maintain the feature map before attention pooling layer of CLIP. As for the input text description, we preserve the sequential token features of the entire sentence instead of only the token features. So we have:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.565, + 0.892, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\left\\{f, F \\right\\} = \\operatorname {E n c} _ {\\mathrm {I}} (x), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.584, + 0.767, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {h}, \\boldsymbol {H} \\right\\} = \\operatorname {E n c} _ {\\mathrm {T}} (\\boldsymbol {r}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.894, + 0.7 + ], + "angle": 0, + "content": "where \\(\\pmb{r}\\) denotes a piece of training text description. \\(f, h \\in \\mathbb{R}^D\\) are the extracted global image and text features. \\(\\pmb{F} \\in \\mathbb{R}^{N_1 \\times D}\\) and \\(\\pmb{H} \\in \\mathbb{R}^{N_2 \\times D}\\) are the flattened dense image features and sequential token features, respectively, where \\(N_1 = H \\times W\\) denotes the flattened spatial dimension of visual feature and \\(N_2\\) denotes the length of text tokens." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.7, + 0.888, + 0.716 + ], + "angle": 0, + "content": "Then, the global and local similarities are computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.727, + 0.892, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {i} = \\left\\langle \\boldsymbol {u}, \\boldsymbol {G} _ {i} \\right\\rangle , \\boldsymbol {P} _ {i j} = \\left\\langle \\boldsymbol {U} _ {j}, \\boldsymbol {L} _ {i} \\right\\rangle \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.817 + ], + "angle": 0, + "content": "where \\(\\pmb{u}\\) denotes either language feature \\(h\\) in training or visual feature \\(f\\) in testing, and \\(\\pmb{U}\\) denotes \\(\\pmb{H}\\) or \\(\\pmb{F}\\) coordinately. Information in local branch \\(\\pmb{P}\\) (visualized in Fig. 3 and Fig. 4) is aggregated in a spatially weighted manner:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.825, + 0.892, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {i} ^ {\\prime} = \\sum_ {j = 1} ^ {N} \\frac {\\exp \\left(\\boldsymbol {P} _ {i j} / \\tau_ {s}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\boldsymbol {P} _ {i j} / \\tau_ {s}\\right)} \\cdot \\boldsymbol {P} _ {i j} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "where \\(\\tau_{s}\\) accommodates the extent of focusing on a specific location. \\(\\pmb{p}_i\\) and \\(\\pmb{p}_i^{\\prime}\\) are optimized by the loss terms \\(\\mathcal{L}_{global}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.445, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.263 + ], + "angle": 0, + "content": "Figure 5. Our learned double-grained prompt tuning is easy to combine with existing prompt tuning methods with ensemble." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.468, + 0.322 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_{local}\\), respectively, which we will discuss in Sec. 3.4. And in the testing phase, \\(\\pmb{p}\\) and \\(\\pmb{p}^{\\prime}\\) are combined to obtain the final classification score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.468, + 0.413 + ], + "angle": 0, + "content": "The visualization in Fig. 3 and Fig. 4 show that the learned local class embedding \\( \\pmb{L} \\) can focus on the specific location where corresponding classes appear, both in sentences and images, even if the fine-grained visual and language features are not explicitly supervised in the training of CLIP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.427, + 0.263, + 0.443 + ], + "angle": 0, + "content": "3.4. Learning Objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.468, + 0.555 + ], + "angle": 0, + "content": "The overall learning objective is defined as \\(\\mathcal{L} = \\mathcal{L}_{global} + \\mathcal{L}_{local}\\), where \\(\\mathcal{L}_{global}\\) and \\(\\mathcal{L}_{local}\\) are loss terms for global text embedding and local text tokens, respectively. We adopt the ranking loss [14] to measure the discrepancy between classification scores and pseudo-ground-truth labels. Specifically, \\(\\mathcal{L}_{global}\\) and \\(\\mathcal{L}_{local}\\) are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.562, + 0.468, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {g l o b a l}} = \\sum_ {i \\in \\{c ^ {+} \\}} \\sum_ {j \\in \\{c ^ {-} \\}} \\max (0, m - p _ {i} + p _ {j}), \\tag {7} \\\\ \\mathcal {L} _ {l o c a l} = \\sum_ {i \\in \\{c ^ {+} \\}} \\sum_ {j \\in \\{c ^ {-} \\}} \\max (0, m - \\pmb {p} _ {i} ^ {\\prime} + \\pmb {p} _ {j} ^ {\\prime}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.468, + 0.718 + ], + "angle": 0, + "content": "where \\( \\pmb{p} \\) and \\( \\pmb{p}' \\) are global and aggregated local similarities described in Sec. 3.3, \\( m \\) is the margin controlling how much higher the similarity score with the positive classes is than with the negative classes. During training, we minimize the overall objective \\( \\mathcal{L} \\) with frozen text encoders, by optimizing the global and local prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Binary cross-entropy loss and its variations such as asymmetric loss [25] have shown remarkable results in classification tasks [25, 28]. They are generally accompanied by a sigmoid function \\(\\sigma(x) = 1 / (1 + \\exp(-x))\\) to convert model outputs to probabilities. Nevertheless, we found that directly optimizing the probability \\(\\sigma(p)\\) leads to a performance gap between training texts and testing images. We attribute this phenomenon to the modality gap between vision and language. Thus we consider ranking loss to be a more flexible and suitable way of supervision in the presence of the modality gap. A comparison of results between different loss functions is provided in Suppl." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.871, + 0.108 + ], + "angle": 0, + "content": "3.5. Incorporating with Prompting from Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Though our TaI-DPT is very different from existing methods of prompting from images, it is also complementary to them. To show this, we utilize an off-the-shelf prompt ensemble strategy to combine our TaI-DPT with existing methods in this section. As illustrated in Fig. 5, using CoOp [41] as an example, we can simply combine the scores of CoOp [41] and that of our TaI-DPT in a weighted sum manner. In particular, our TaI-DPT can be integrated with CoOp [41] when a few annotated images are provided and integrated with DualCoOp [28] when partially labeled training data are available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.432 + ], + "angle": 0, + "content": "We ensemble prompts by fusing the predicted scores, rather than averaging the class embeddings generated by different prompts, since the image encoder used in different methods may be different (e.g. we conduct our experiments with ResNet50, while DualCoOp uses ResNet101 for partial-label prompting). So ensembling with the classification score is more convenient. In Sec. 4.3, we also empirically show that our prompt ensemble strategy is effective in advancing multi-label recognition performance in the few-shot and partially labeled settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.445, + 0.632, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.717, + 0.485 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Architecture. We adopt CLIP ResNet-50 [24] as the visual encoder, and use the CLIP Transformer as the text encoder. During training, the parameters of the two encoders are kept frozen, and only learnable prompts are optimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Learnable Prompts. Our learnable prompts are shared among classes of all datasets. Class-specific prompting [41] (i.e., an individual set of parameters for each category) has also been explored, but brings limited benefits. We initialize the value of each parameter with the Gaussian noise sampled from \\(\\mathcal{N}(0,0.02)\\). The length of both the global prompts and local prompts are set to \\(M = 16\\), while a longer sequence brings trivial improvements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. To evaluate our TaI-DPT, we conduct the experiments on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9]. VOC2007 contains 20 common categories, and following [6,8,28], we form the training/test set based on the official trainval/test split (5,011 images/4,952 images). MS-COCO includes 80 categories with 82,081 training images and 40,504 testing images. NUS-WIDE includes 81 concepts with 161,789 images for training. We adopt its test set (107,859 images) to evaluate our method. For zero-shot experiments in Sec. 4.2, the training sets of the datasets are not used, and we use only text data to learn the prompts as mentioned in Sec. 3.2. Besides, for VOC2007 and MS-COCO, the language data sources are captions from MS-COCO. For NUS-WIDE, we introduce localized narratives from OpenImages [18], which have a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2813" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.346, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.091, + 0.6, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.61, + 0.301 + ], + "angle": 0, + "content": "Figure 6. Comparison of different methods in few-shot multi-label recognition on VOC2007 and MS-COCO. Our zero-shot TaI-DPT can achieve comparable results with methods trained by 16-shot labeled image samples. And learned prompt ensemble proofs the complementarity between images and texts." + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.092, + 0.888, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.252, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Figure 7. Ablation experiment on the number of texts and performance of TaI prompting on VOC2007." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Table 1. Comparison with zero-shot methods on VOC2007, MS-COCO, and NUS-WIDE. Our proposed TaI-DPT outperforms CLIP [24] by a large margin on all datasets." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.445 + ], + "angle": 0, + "content": "
MethodDPTVOC2007MS-COCONUSWIDE
ZSCLIPX76.247.336.4
✓77.349.737.4
TaIX86.061.144.9
✓88.365.146.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.469, + 0.52 + ], + "angle": 0, + "content": "broader range of content, to cover all the concepts in NUS-WIDE. In Sec. 4.3 and Sec. 4.4, for each dataset, the corresponding training data is used to conduct the experiments of partial-label and few-shot multi-label classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.629 + ], + "angle": 0, + "content": "Training Details. We adopt SGD optimizer, and the training epoch is set to 20 for all datasets. The learning rates for MS-COCO, VOC2007, and NUS-WIDE are empirically initialized with 1e-4, 1e-4, and 1e-3, and decay by the cosine annealing rule. For ranking loss, we choose \\( m = 1 \\), and scale the \\( p \\) and \\( p' \\) by a factor of 4. \\( \\tau_{s} \\) is set as 0.02 via validation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.4, + 0.664 + ], + "angle": 0, + "content": "4.2. Comparison with Zero-Shot Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed TaI and DPT, we first compare it with the zero-shot CLIP (ZSCLIP). For fair comparison, we also introduce the DPT to ZSCLIP. Specifically, we adopt two identical default prompts \"a photo of a [CLASS]\" to separately deal with global and local features as DPT does. Table 1 lists the comparison results on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9] datasets. From the table, our TaI prompting surpasses ZSCLIP by a large margin of \\(9.8\\%\\), \\(13.8\\%\\), and \\(8.5\\%\\) mAP on VOC2007, MS-COCO, and NUS-WIDE, respectively, showing the effectiveness of our TaI. Furthermore, after training with fine-grained token features extracted from texts, our proposed DPT demonstrates a more powerful capability of discriminating local features than the default hand-crafted prompts and single global prompts." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.311, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Table 2. Comparison with existing multi-label few-shot learning methods on MS-COCO. The evaluation is based on mAP for zero-shot, 1-shot and 5-shot with 16 novel classes." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.355, + 0.88, + 0.462 + ], + "angle": 0, + "content": "
Method0-shot1-shot5-shot
LaSO [2]-45.358.1
ML-FSL [27]-54.463.6
CoOp [41]40.2 (ZSCLIP)46.955.6
Tip-Adapter [38]40.2 (ZSCLIP)53.859.7
TaI-DPT59.2--
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.479, + 0.818, + 0.494 + ], + "angle": 0, + "content": "4.3. Comparison with Few-Shot Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.641 + ], + "angle": 0, + "content": "We further compare with multi-label few-shot learning methods to verify the effectiveness of our TaI-DPT. In contrast to the well-studied single-label few-shot classification problem, few works tackle the multi-label few-shot scenario. Existing methods [2, 27] often deploy models trained on seen classes to few-shot novel classes. In Table 2, we compare zero-shot TaI-DPT to few-shot methods on 16 novel classes (see [2] for details about data split). TaI-DPT is comparable to the methods trained on 5-shot samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Besides, we consider a new multi-label few-shot setting where all the classes are regarded as novel classes. We select 1, 2, 4, 8, and 16-shot samples for each category following the strategy in [2]. For fair comparison, we train CoOp [41] and our TaI in the same settings, and we also extend them with DPT for a more comprehensive comparison. For CoOp-DPT, we set two sets of learnable prompts, to deal with global and local features, respectively. The results are illustrated in Fig. 6. One can see that, even without any image data regarding novel classes, our TaI can achieve comparable results to CoOp trained on 16-shot. Similar trends with the MS-COCO dataset and the DPT setting support our observation that the discriminative feature of text data can be used as images for prompting. Moreover, benefiting from the flexibility of prompts, we can easily integrate our TaI-DPT with CoOp-DPT by ensembling as said in Sec. 4.3. As illustrated in Fig. 6, though CoOp-DPT has" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2814" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 3. Results of integrating our TaI-DPT with partial-label multi-label recognition method based on pre-trained CLIP. Our approach further improves the frontier performance of DualCoOp [28]. * indicates the results based on our own reproduction." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.122, + 0.869, + 0.306 + ], + "angle": 0, + "content": "
DatasetsMethod10%20%30%40%50%60%70%80%90%Avg.
MS-COCOSARB [23]71.275.077.178.378.979.679.880.580.577.9
DualCoOp [28]78.780.981.782.082.582.782.883.083.181.9
DualCoOp*81.082.382.983.483.583.984.084.184.383.3
+TaI-DPT81.582.683.383.783.984.084.284.484.583.6
PascalVOC 2007SARB [23]83.588.690.791.491.992.292.692.892.990.7
DualCoOp [28]90.392.292.893.393.693.994.094.194.293.2
DualCoOp*91.493.893.894.394.694.794.894.994.994.1
+TaI-DPT93.394.694.894.995.195.095.195.395.594.8
NUS-WIDEDualCoOp*54.056.256.957.457.957.957.658.258.857.2
+TaI-DPT56.457.957.858.158.558.858.659.159.458.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.471, + 0.385 + ], + "angle": 0, + "content": "achieved a high accuracy, combining our prompts learned with text data still brings further improvement on recognition performance. This also proves that texts and images are complementary to each other to some extent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.398, + 0.449, + 0.414 + ], + "angle": 0, + "content": "4.4. Integration with Partially Labeled Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.497 + ], + "angle": 0, + "content": "Partially-labeled data refers to the problem where merely some labels of each sample are known. Following [5], we created the partial-label data by randomly masking out labels of the fully annotated data. During inference, the model is evaluated on all categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.47, + 0.59 + ], + "angle": 0, + "content": "We reproduce DualCoOp [28] on partial-labeled VOC2007 and MS-COCO with the same experimental setting as reported (reproduced results are marked with *) and explore the enhancement brought by integration with our TaI-DPT. We also deploy DualCoOp on partial-labeled NUS-WIDE for a comprehensive comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.471, + 0.788 + ], + "angle": 0, + "content": "The results are reported in Table 3. With no prior knowledge from pre-trained models, previous forefront method like SARB [23] struggles to learn from incomplete labels. DualCoOp, directly trained with entire image set, is well-learned to classify multi-label images. Even so, for PascalVOC and NUS-WIDE, by using texts from exogenous sources, i.e., COCO and OpenImages, our method can bring complementary enhancement to image prompting. While, for COCO, we used captions derived from its own images, making it more difficult to provide additional benefits. Analogous to VOC and NUS-WIDE, the COCO results can also be further advanced by introducing more external text data in future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.231, + 0.817 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To thoroughly investigate the effect of each component, we conduct a series of ablation studies on the quantity of texts, training loss, different VL pre-trained models, ensemble weights, and texts v.s. images for prompting. The complete ablation experiments are shown in the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Quantity of texts. Following the data preparation procedure in Sec. 3.2, we end up with a total number of 66087 pieces of text that contain descriptions for 20 categories of VOC2007. We test the performance of TaI-DPT with different numbers of randomly selected texts, and the results are shown in Fig. 7. When no collected texts are available, 80 templates of hand-crafted prompts from [24], like \"a cropped photo of a [CLASS]\", are used for training (all templates are shown in the Suppl), and each template sentence correlates with one positive label corresponding to the class name inserted in [CLASS]. The increasing number of texts gradually forms a complete description of target categories, and the relationship between classes is also better characterized, which results in ascending performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.62, + 0.571 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.794 + ], + "angle": 0, + "content": "In this paper, we propose a new view of treating texts as images in prompt tuning (i.e. TaI), which learns the prompt from discriminative features of text descriptions. Compared to prior prompt tuning methods trained with images, our TaI benefits from the easy accessibility of scalable content-rich texts, which enables prompt tuning for vision tasks (e.g., multi-label image recognition) even without downstream image data. Double-grained prompting is further introduced to utilize both the global and fine-grained features for better multi-label recognition ability. Nonetheless, when few-shot image samples or partial-labeled images are available, our TaI-DPT can conveniently integrate with existing prompting methods. Experiments on multiple benchmarks show the validity of TaI-DPT." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.661, + 0.831 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by National Key R&D Program of China under Grant No. 2020AAA0104500, and by the National Natural Science Foundation of China (NSFC) under Grant No. U19A2073." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2815" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Amit Alfassy, Leonid Karlinsky, Amit Aides, Joseph Shtok, Sivan Harary, Rogerio Feris, Raja Giryes, and Alex M Bronstein. Laso: Label-set operations networks for multi-label few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6548-6557, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[3] Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit.\" O'Reilly Media, Inc.\", 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 339-346, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 522-531, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[8] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[9] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In Proceedings of the ACM international conference on image and video retrieval, pages 1-9,2009.2,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[12] Bin-Bin Gao and Hong-Yu Zhou. Multi-label image recognition with multi-class attentional regions. arXiv preprint arXiv:2007.01755, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Yunchao Gong, Yangqing Jia, Thomas Leung, Alexander Toshev, and Sergey Ioffe. Deep convolutional ranking for mul" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tilabel image annotation. arXiv preprint arXiv:1312.4894, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[15] Shiyi He, Chang Xu, Tianyu Guo, Chao Xu, and Dacheng Tao. Reinforced multi-label image classification by exploring curriculum. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Shahab Kamali, Matteo Malloci, Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://storage.googleapis.com/openimages/web/index.html, 2017. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[19] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[21] Luchen Liu, Sheng Guo, Weilin Huang, and Matthew Scott. Decoupling category-wise independence and relevance with self-attention for multi-label image classification. In ICASSP 2019, 05. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[23] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. arXiv preprint arXiv:2203.02172, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2816" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[26] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. arXiv preprint arXiv:2209.07511, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[27] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. Meta-learning for multi-label few-shot classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3951-3960, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[28] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. arXiv preprint arXiv:2206.09541, 2022. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.468, + 0.335 + ], + "angle": 0, + "content": "[29] Ashwin Vaswani, Gaurav Aggarwal, Praneeth Netrapalli, and Narayan G Hegde. All mistakes are not equal: Comprehensive hierarchy aware multi-label predictions (champ). arXiv preprint arXiv:2206.08653, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.468, + 0.379 + ], + "angle": 0, + "content": "[30] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.468, + 0.451 + ], + "angle": 0, + "content": "[31] Yangtao Wang, Yanzhao Xie, Yu Liu, Ke Zhou, and Xiaocui Li. Fast graph convolution network based multi-label image recognition via cross-modal fusion. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 1575-1584, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[32] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE transactions on pattern analysis and machine intelligence, 38, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.468, + 0.581 + ], + "angle": 0, + "content": "[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.468, + 0.638 + ], + "angle": 0, + "content": "[34] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.642, + 0.468, + 0.683 + ], + "angle": 0, + "content": "[35] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.468, + 0.754 + ], + "angle": 0, + "content": "[36] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.468, + 0.811 + ], + "angle": 0, + "content": "[37] Junjie Zhang, Qi Wu, Chunhua Shen, Jian Zhang, and Jianfeng Lu. Multilabel image classification with regional latent semantic dependencies. IEEE Transactions on Multimedia, 20, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[38] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[39] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2817" + } + ] +] \ No newline at end of file diff --git a/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_origin.pdf b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..02cf9f0dba462c572957c84dc2cf668c33e4c075 --- /dev/null +++ b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/df1598eb-cae0-4ed0-bda2-3a1705488a58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a17d379fd88a6deb392eab89f68a6c3554bbe331a0a5711c87fe9d7aa6d360db +size 7695639 diff --git a/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/full.md b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e0026105a9d04f1cc0f1e26cd58194a9c3a9555 --- /dev/null +++ b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/full.md @@ -0,0 +1,313 @@ +# Texts as Images in Prompt Tuning for Multi-Label Image Recognition + +Zixian Guo $^{1,2*}$ Bowen Dong $^{1}$ Zhilong Ji $^{2}$ Jinfeng Bai $^{2}$ Yiwen Guo $^{4}$ Wangmeng Zuo $^{1,2(\text{图})}$ + +$^{1}$ Harbin Institute of Technology $^{2}$ Tomorrow Advancing Life $^{3}$ Pazhou Lab, Guangzhou $^{4}$ Independent Researcher + +zixian.guo@foxmail.com cndongsky@gmail.com zhilongji@hotmail.com + +jfbai.bit@gmail.com guoyiwen89@gmail.com WMzuo@hit.edu.cn + +# Abstract + +Prompt tuning has been employed as an efficient way to adapt large vision-language pre-trained models (e.g. CLIP) to various downstream tasks in data-limited or label-limited settings. Nonetheless, visual data (e.g., images) is by default prerequisite for learning prompts in existing methods. In this work, we advocate that the effectiveness of image-text contrastive learning in aligning the two modalities (for training CLIP) further makes it feasible to treat texts as images for prompt tuning and introduce TaI prompting. In contrast to the visual data, text descriptions are easy to collect, and their class labels can be directly derived. Particularly, we apply TaI prompting to multi-label image recognition, where sentences in the wild serve as alternatives to images for prompt tuning. Moreover, with TaI, double-grained prompt tuning (TaI-DPT) is further presented to extract both coarse-grained and fine-grained embeddings for enhancing the multi-label recognition performance. Experimental results show that our proposed TaI-DPT outperforms zero-shot CLIP by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE, while it can be combined with existing methods of prompting from images to improve recognition performance further. The code is released at https://github.com/guozix/Tai-DPT. + +# 1. Introduction + +Recent few years have witnessed rapid progress in large vision-language (VL) pre-trained models [1, 16, 19, 24, 33, 36] as well as their remarkable performance on downstream vision tasks. A VL pre-trained model generally involves data encoders, and it is becoming increasingly popular to exploit image-test contrastive loss [24] to align the embedding of images and texts into a shared space. When adapting to downstream tasks in data-limited or label-limited settings, it is often ineffective to fine-tune the entire model, due to its high complexity. Then, prompt tuning as a representative parameter-efficient learning paradigm has emerged as + +![](images/da7c14fde341874046b193f3d59da4fdd6a813b0af28615007eb5717bd026ab4.jpg) +Figure 1. A comparison between prompting from images and our text-as-image (TaI) prompting. (a) Prompting from images (e.g., [41]) uses labeled images of task categories to learn the text prompts. Instead, (b) our TaI prompting learn the prompts with easily-accessed text descriptions containing target categories. (c) After training, the learned prompts in (a) or (b) can be readily applied to test images. + +an efficient way to adapt VL models to downstream tasks. + +Albeit considerable achievements have been made, existing prompt tuning methods generally require visual data to learn prompts (as shown in Fig. 1(a)). For example, CoOp [41] learns from annotated images. CoCoOp [40] further introduces generalizable input-conditional prompts. DualCoOp [28] adapts CLIP to multi-label recognition tasks by training pairs of positive and negative prompts with partial-labeled images. Nonetheless, the performance of these prompting methods may be limited when it is infeasible to obtain sufficient image data or annotate the required images. + +In this paper, we advocate treating Texts as Images for prompt tuning, i.e., TaI prompting. It is considered feasible as the image encoder and text encoder in many pre-trained VL models [16, 24] encode images and texts into a shared space. Given an image and its caption, the visual features produced by the image encoder will be close to the text feature of the caption produced by the text encoder. There- + +fore, in addition to extracting visual features from images, it is also feasible to extract text features as alternatives form, for example, descriptive sentences and captions, for prompt tuning (see Fig. 1(b)). TaI prompting has several interesting properties and merits. Taking a downstream image recognition task as an example, given a set of object categories, one can easily crawl a large set of text descriptions that contain object names from these categories. Text descriptions are easily accessible in this way, and class labels can be directly derived from text descriptions, which means, in contrast to prompting from images, TaI prompting may suffer less from the data-limited and label-limited issues. + +We use multi-label image recognition [8, 9, 11, 20, 35] to verify the effectiveness of our TaI prompting in this paper. To begin with, we crawl the captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., Open Images [18]) to form the training set of text descriptions. For any specific multi-label recognition task, we adopt a noun filter to map the nouns in the text descriptions to the corresponding object categories, and then only keep the text descriptions that contain one or more classes of target objects. To better cope with multi-label classification, we introduce double-grained prompt tuning (i.e., TaI-DPT) which involves: (i) a set of global prompts to generate embeddings for classifying whole sentences or images, and (ii) a set of local prompts to extract embeddings for discriminating text tokens or image patches. Given a set of text descriptions, global and local prompts can be tuned by minimizing the ranking loss [14]. Note that, though these prompts are learned from text descriptions solely, they can be readily deployed to classify whole images as well as image patches during testing (see Fig. 1(c)). Experimental results show that, without using any labeled images, our TaI prompting surpasses zero-shot CLIP [24] by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE. + +Moreover, when images are also available during training, our TaI prompting can be combined with existing methods of prompting from images to improve its performance. In particular, given a few annotated images, our TaI-DPT can be integrated with CoOp as a prompt ensemble for improving classification accuracy. With partially labeled training data being provided, we may also combine TaI-DPT and DualCoOp [28] to improve multi-label recognition accuracy consistently. Extensive results verify the effectiveness of our TaI-DPT in comparison to state-of-the-art. To sum up, the contributions of this work include: + +- We propose Texts as Images in prompt tuning (i.e., TaI prompting) to adapt VL pre-trained models to multi-label image recognition. Text descriptions are easily accessible and, in contrast to images, their class labels can be directly derived, making our TaI prompting very compelling in practice. + +- We present double-grained prompt tuning (i.e. TaI-DPT) to extract both coarse-grained and fine-grained embeddings for enhancing multi-label image recognition. Experiments on multiple benchmarks show that TaI-DPT achieves comparable multi-label recognition accuracy against state-of-the-arts. +- The prompts learned by TaI-DPT can be easily combined with existing methods of prompting from images in an off-the-shelf manner, further improving multi-label recognition performance. + +# 2. Related Work + +# 2.1. Multi-Label Image Recognition + +Multi-label image recognition [3,7,8,12,15,21,32,35,37] aims to recognize all the object categories [11, 20] or concepts [9] in an input image. Various modules [6, 30] have been introduced to better represent the inter-class relationships, and modern classification losses [3, 14] have been used to make model learning easier. + +To model the label dependencies, CNN-RNN [30] introduces recurrent neural networks, e.g., RNN and LSTM, to predict appeared classes in a sequential manner. [6,8,31,35] use graph convolution modules to learn the correlation between class labels. CHAMP [29] measures the severity of misclassification by building a domain-specific hierarchy tree according to the relation of categories, where each class is related to a tree node, to improve the robustness of the model. Albeit effective, these methods require a considerable number of labeled images to learn the category relationships sufficiently. While in data-limited or label-limited regimes, e.g., few-shot or partial-label data, it will be difficult for these models to learn well as expected. Specifically designed loss functions also struggle to obtain significant improvements when learning with limited data. + +Multi-Label Recognition from Few-shot Samples. To better exploit the small number of samples, LaSO [2] synthesizes samples by manipulating the features of paired training images. Different ways of manipulating label sets are used to train the model, resulting in generalizable discriminative features. [27] introduces a meta-learning framework for better learning of past tasks and generalization to new tasks, and leverages the number of labels as useful information for learning. + +Multi-Label Recognition from Partial-label Data. Partial-label refers to the scenarios where some labels are unknown. [10] propose a normalized BCE loss to balance the proportion of known labels. [5] learns to complement unknown labels by utilizing within-image and cross-image semantic correlations. [23] blends the representation of training images and class proxies to compensate for the loss of information due to unknown labels. + +Albeit significant progress has been made, it remains a challenging issue for learning multi-label image recognition + +![](images/502d18cbf3bbd3f4cf9cf9bfa0f79de8324e2b9bbe0a5338d78a146061fb634d.jpg) +(a) + +![](images/71fabcbc5e0d75b9ccb3a6fb219451d240d60858ebedce1c220902e79a68be4e.jpg) +(b) +Figure 2. Training and testing pipeline of our proposed Text-as-Image (TaI) prompting, where we use text descriptions instead of labeled images to train the prompts. (a) During training, we use two identical text encoders from pre-trained CLIP to extract the global & local class embeddings $(\mathbf{G} \& \mathbf{L})$ and overall & sequential text embeddings $(\mathbf{h} \& \mathbf{H})$ respectively from the prompts and text description. The corresponding cosine similarity $(\mathbf{p} \& \mathbf{P})$ between the embeddings are guided by the derived pseudo labels with ranking loss. (b) During testing, we replace the input from text descriptions to images. The global and local class embeddings can discriminate target classes from global & local image features $(\mathbf{f} \& \mathbf{F})$ . The final classification results are obtained by merging the scores of the two branches. + +in image-limited or label-limited regimes. Built upon VL pre-trained models, this paper suggests learning prompts from texts instead of images, thereby offering a novel yet complementary perspective for handling low-resource multi-label image recognition. + +# 2.2. Prompt Tuning for Vision-Language Models + +To transfer pre-trained knowledge to downstream tasks in data-limited settings, prompt tuning [13, 17, 34, 39, 41, 42] has become a popular parameter-efficient way to achieve the goal, due to its flexibility and ease of use. CoOp [41] learns the prompts by using (a few) annotated images of each class from the target dataset. CoCoOp [40] further proposes to improve CoOp [41] by formulating the prompts in an image-conditional way to maintain better generalization to unseen classes. To avoid overfitting, ProGrad [42] leverages predictions from zero-shot CLIP to regularize gradients in prompt learning process. TPT [26] suggests optimizing test-time prompts by promoting the consistency of augmented test images. ProDA [22] uses multiple pieces of prompts to estimate the distribution of classifier weights for better handling of varying visual features. DualCoOp [28] firstly adapts CLIP to multi-label image recognition with partially labeled data by learning pairs of positive and negative prompts for each class to ensure independent binary classification for each class. + +Albeit existing prompt tuning approaches have achieved significant improvements in downstream tasks, images as well as a portion of class labels are prerequisite to supervise the optimization of the learnable prompts. In this paper, we propose to treat texts as images in prompt tuning, which, compared to labeled images, are much easier to collect with existing caption datasets and modern search engines. Our + +proposed TaI-DPT surpasses zero-shot CLIP by a large margin and can be combined with the prompts learned by existing methods of prompting from images to further boost recognition performance. + +# 3. Proposed Method + +Here we present Text-as-Image prompting, i.e., TaI prompting, for adapting pre-trained VL models to multi-label image recognition. Our TaI prompting uses only easily-accessed free-form texts as training data to learn effective prompts for downstream multi-label recognition tasks. To begin with, We present an overview of TaI prompting in Sec. 3.1. Then, we introduce our preparation of training texts in Sec. 3.2. We further explain the design of the double-grained prompt tuning (i.e., TaI-DPT) in Sec. 3.3, and provide the loss function used to train the model in Sec. 3.4. Finally, we propose to ensemble prompts from TaI-DPT with prompts learned from images as a flexible integration to improve multi-label recognition performance. + +# 3.1. Overview of Our Method + +Fig. 2 illustrates the design of our proposed TaI-DPT framework, including the training and testing phases. During training, we learn prompts with only supervision from texts. Two identical copies of the text encoder $\mathrm{Enc}_{\mathrm{T}}$ from the pre-trained CLIP are used to encode the prompts and text data, respectively. We introduce two sorts of trainable prompts (i.e., the global prompts and local prompts) to obtain global and regional class embeddings. A noun filtering strategy is used to generate classification pseudo-labels for each text description, which is applied to supervise the classification scores obtained by calculating the cosine sim + +iliarity of class embeddings and text features. Only the parameters in prompts are optimized in the training phase, while the text encoders are both kept frozen. During testing, the class embeddings are obtained by encoding the two sets of learned prompts with the text encoder $\mathrm{Enc}_{\mathrm{T}}$ as in training, while the other input source changes from text descriptions to test images. Pre-trained image encoder $\mathrm{Enc}_{\mathrm{I}}$ from CLIP is used to extract global and dense features of each test image, then the cosine similarity are computed between features and class embeddings of the global and local prompts. The final classification result is obtained by fusing the global and local cosine scores. In the following, we explain the details of our proposed method. + +# 3.2. Preparation of Text Descriptions + +To obtain sufficient category information from the language that helps in image recognition, we have to ensure that: 1) the collected texts should contain content-rich descriptions of a scene, and 2) the contents of all text descriptions need to cover the category set of the target dataset for comprehensive representations of all categories. With the aim of ensuring reproducibility, we use readily available captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., OpenImages [18]) as our language data source, while avoiding the workloads associated with randomly crawling texts from the Internet in this paper. Note that although each caption is paired with a corresponding image and human-annotated labels, we only use the captions, and no information from the pictures and labels is disclosed during training. + +For a target multi-label recognition dataset $\mathcal{X}$ that has a category set $S = \{\mathrm{s}_1, \mathrm{s}_2, \mathrm{s}_3, \dots, \mathrm{s}_C\}$ , where $C$ denotes the number of categories and $\mathrm{s}_i$ denotes particular class name like "dog", "plane", etc., we search for sentences that contain at least one class name $\mathrm{s}_i$ in $S$ . Since multiple words or phrases usually exist to represent the same meaning for each class, searching solely for exact match of category names in texts may lead to many false negatives in the obtained pseudo ground-truth labels, which is harmful to prompt tuning. Towards tackling this issue, we introduce a noun filter to map nouns with similar meanings into the corresponding class label. Specifically, we construct a synonym dictionary $\mathcal{D}$ by including common synonyms of each class name in the target dataset. If a word in a text description matches any synonym of a specific class name, it is considered to contain a description of that category. Several examples of synonyms are shown as follows: + +```txt +{'dog', 'pup', 'puppy', 'doggy'} +{'person', 'people', 'man', 'woman', 'human'} +{'bicycle', 'bike', 'cycle'} +{'car', 'taxi', 'automobile'} +{'boat', 'raft', 'dinghy'} +... +``` + +More details of the dictionary $\mathcal{D}$ are provided in the Suppl. + +Then we conduct noun filtration by the following steps. First, for each text description, we use the tokenizer and lemmatizer from NLTK [4] to recover the stem of each word in the sentences. Next, for all keywords in $\mathcal{D}$ , which contains all synonyms of the category set $S$ , we search in our language data source for sentences that contains at least one class name. For the text descriptions that do not match any synonym of any class name, we simply drop them away to ensure each piece of data has at least one concerned label. Finally, for each retained text description, we convert the class names it contains into binary pseudo-ground-truth vectors by setting classes that appear as positive and other classes as negative, following the order of class labels in the target dataset $\mathcal{X}$ . + +The word-level filtered labels may not be precisely correct since our searching strategy mentioned above is rather simple considering the diversity of free-form texts, where complex paraphrases and misspellings that widely exist in the corpus are not fully addressed. However, such a simple noun filtration can guarantee the reproducibility of this work, and our experiments also demonstrate that this simple and efficient data preparation leads to compelling results of multi-label recognition of our TaI prompting. + +# 3.3. Text-as-Image for Dual-grained Prompt Tuning + +Following [41], a prompt is defined as: + +$$ +\boldsymbol {t} _ {i} = \left[ \boldsymbol {v} _ {1}, \boldsymbol {v} _ {2}, \boldsymbol {v} _ {3}, \dots , \boldsymbol {v} _ {M}, \boldsymbol {s} _ {i} \right] \tag {1} +$$ + +where $i \in \{1, 2, \dots, C\}$ is the class index, $s_i$ denotes word embedding of the $i$ -th class name $s_i$ . For $j \in \{1, \dots, M\}$ , $v_j$ is a learnable word embedding whose dimension is the same as the dimension of normal word embeddings in the vocabulary. Just like in previous methods, e.g. CoOp [41], the prompts are learned by maximizing the probability of classifying each image into its ground-truth class: + +$$ +p (y = i | \boldsymbol {x}) = \frac {\exp \left(\left\langle \operatorname {E n c} _ {\mathrm {T}} \left(\boldsymbol {t} _ {i}\right) , \operatorname {E n c} _ {\mathrm {I}} (\boldsymbol {x}) \right\rangle / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\left\langle \operatorname {E n c} _ {\mathrm {T}} \left(\boldsymbol {t} _ {j}\right) , \operatorname {E n c} _ {\mathrm {I}} (\boldsymbol {x}) \right\rangle / \tau\right)} \tag {2} +$$ + +where $\pmb{x}$ denotes the labeled training image and $\langle \cdot, \cdot \rangle$ calculates the cosine similarity. + +After large-scale image-text contrastive pre-training, text features have been well-aligned to the image features of the same semantic meanings. Thus, features of texts that describe a certain type of object also exhibit categorical discriminability. Therefore, based on the aligned VL representation, we advocate considering the feature of a piece of text description that describes a specific category, as an alternative to an image feature. + +Apart from using the global sentence representation (i.e., the coarsest-grained text feature), we find that the sequential feature of word tokens from CLIP also possesses rich fine-grained information which is very similar to the region feature of dense image feature. In CLIP [24], cosine similarity + +![](images/88cf9e5f6a9ea7e84eb447ea37e472b63ce24d59a391cf18895f58773c023236.jpg) + +![](images/67a48dd6aa75a5946ce6d02959c4a313c72446431e6a2642ff4284badae7645a.jpg) + +![](images/98b8d8e05a2fd435780dfac9dbcdce3af1365023f52c14438e582a47709ed25f.jpg) + +![](images/fdb9e86433bebd7832a4daf80317954a0149179108449c53364544a2258375b9.jpg) +Figure 3. Visualization of correlations $\pmb{P}$ between the local class embedding $\pmb{L}$ and sequential token feature from texts. Each class embedding clearly correlates to words that describe the corresponding class (shown in highlight regions) rather than the global $<\mathrm{EOS}>$ token. +Figure 4. Visualization of correlations between the local class embedding $\mathbf{L}$ and dense image feature. The learned class embeddings can focus on the location of the object effectively. + +between global image features, obtained by visual attention pooling, and global text features, obtained by projecting the feature of the last $<\mathrm{EOS}>$ token, are directly supervised with contrastive loss. In general, the global feature is sufficient for single-label classification because the target object usually is prominent in the picture. However, in multi-label recognition, the global feature is usually dominated by major objects, suppressing the recognition of non-significant objects concurrently existing in the image. Thus, it motivates us to explore fine-grained features and avoid the domination of the overly prominent object. + +To grasp the discriminability of global features as well as learn from fine-grained features, we propose double-grained prompt tuning (i.e., TaI-DPT) that uses two sets of prompts to handle global (i.e., the coarsest-grained level) and local (i.e., the fine-grained level) features, respectively, in two parallel branches. Formally, the double-grained prompt is defined as follows: + +$$ +\boldsymbol {t} _ {i} ^ {G} = \left[ \boldsymbol {v} _ {1}, \boldsymbol {v} _ {2}, \boldsymbol {v} _ {3}, \dots , \boldsymbol {v} _ {M}, \boldsymbol {s} _ {i} \right], \tag {3} +$$ + +$$ +\boldsymbol {t} _ {i} ^ {L} = \left[ \boldsymbol {v} _ {1} ^ {\prime}, \boldsymbol {v} _ {2} ^ {\prime}, \boldsymbol {v} _ {3} ^ {\prime}, \dots , \boldsymbol {v} _ {M} ^ {\prime}, \boldsymbol {s} _ {i} \right], +$$ + +where $\pmb{v}_j$ and $\pmb{v}_j^\prime$ are learnable embeddings that are concatenated with word embedding $s_i$ of the $i$ -th class to obtain the + +global prompt $t_i^G$ and local prompt $t_i^L$ , respectively. The sequences in Eq. (3) are fed to a copy of the text encoder EncT of CLIP to generate global and local class embeddings for each class, i.e. $G_{i} = \mathrm{Enc}_{\mathrm{T}}(t_{i}^{G})$ and $L_{i} = \mathrm{Enc}_{\mathrm{T}}(t_{i}^{L})$ , $G = \{G_{i}\}_{i=1}^{C}$ and $L = \{L_{i}\}_{i=1}^{C}$ are encouraged to be correlated with global and local features, respectively. Note that the proposed double-grained prompts are different from dual prompts [28], which include a pair of contrastive positive and negative prompts for each class (More discussion about the differences between our method and DualCoOp is provided in the Suppl). + +To preserve the fine-grained region features for the input image, we maintain the feature map before attention pooling layer of CLIP. As for the input text description, we preserve the sequential token features of the entire sentence instead of only the token features. So we have: + +$$ +\left\{f, F \right\} = \operatorname {E n c} _ {\mathrm {I}} (x), \tag {4} +$$ + +$$ +\left\{\boldsymbol {h}, \boldsymbol {H} \right\} = \operatorname {E n c} _ {\mathrm {T}} (\boldsymbol {r}), +$$ + +where $\pmb{r}$ denotes a piece of training text description. $f, h \in \mathbb{R}^D$ are the extracted global image and text features. $\pmb{F} \in \mathbb{R}^{N_1 \times D}$ and $\pmb{H} \in \mathbb{R}^{N_2 \times D}$ are the flattened dense image features and sequential token features, respectively, where $N_1 = H \times W$ denotes the flattened spatial dimension of visual feature and $N_2$ denotes the length of text tokens. + +Then, the global and local similarities are computed by: + +$$ +\boldsymbol {p} _ {i} = \left\langle \boldsymbol {u}, \boldsymbol {G} _ {i} \right\rangle , \boldsymbol {P} _ {i j} = \left\langle \boldsymbol {U} _ {j}, \boldsymbol {L} _ {i} \right\rangle \tag {5} +$$ + +where $\pmb{u}$ denotes either language feature $h$ in training or visual feature $f$ in testing, and $\pmb{U}$ denotes $\pmb{H}$ or $\pmb{F}$ coordinately. Information in local branch $\pmb{P}$ (visualized in Fig. 3 and Fig. 4) is aggregated in a spatially weighted manner: + +$$ +\boldsymbol {p} _ {i} ^ {\prime} = \sum_ {j = 1} ^ {N} \frac {\exp \left(\boldsymbol {P} _ {i j} / \tau_ {s}\right)}{\sum_ {j = 1} ^ {N} \exp \left(\boldsymbol {P} _ {i j} / \tau_ {s}\right)} \cdot \boldsymbol {P} _ {i j} \tag {6} +$$ + +where $\tau_{s}$ accommodates the extent of focusing on a specific location. $\pmb{p}_i$ and $\pmb{p}_i^{\prime}$ are optimized by the loss terms $\mathcal{L}_{global}$ + +![](images/2c89f2e793150d87967aed74564435b54c080dab9e3f4e0253f7cad2359ce743.jpg) +Figure 5. Our learned double-grained prompt tuning is easy to combine with existing prompt tuning methods with ensemble. + +and $\mathcal{L}_{local}$ , respectively, which we will discuss in Sec. 3.4. And in the testing phase, $\pmb{p}$ and $\pmb{p}^{\prime}$ are combined to obtain the final classification score. + +The visualization in Fig. 3 and Fig. 4 show that the learned local class embedding $\pmb{L}$ can focus on the specific location where corresponding classes appear, both in sentences and images, even if the fine-grained visual and language features are not explicitly supervised in the training of CLIP. + +# 3.4. Learning Objective + +The overall learning objective is defined as $\mathcal{L} = \mathcal{L}_{global} + \mathcal{L}_{local}$ , where $\mathcal{L}_{global}$ and $\mathcal{L}_{local}$ are loss terms for global text embedding and local text tokens, respectively. We adopt the ranking loss [14] to measure the discrepancy between classification scores and pseudo-ground-truth labels. Specifically, $\mathcal{L}_{global}$ and $\mathcal{L}_{local}$ are formulated as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {g l o b a l}} = \sum_ {i \in \{c ^ {+} \}} \sum_ {j \in \{c ^ {-} \}} \max (0, m - p _ {i} + p _ {j}), \tag {7} \\ \mathcal {L} _ {l o c a l} = \sum_ {i \in \{c ^ {+} \}} \sum_ {j \in \{c ^ {-} \}} \max (0, m - \pmb {p} _ {i} ^ {\prime} + \pmb {p} _ {j} ^ {\prime}) \\ \end{array} +$$ + +where $\pmb{p}$ and $\pmb{p}'$ are global and aggregated local similarities described in Sec. 3.3, $m$ is the margin controlling how much higher the similarity score with the positive classes is than with the negative classes. During training, we minimize the overall objective $\mathcal{L}$ with frozen text encoders, by optimizing the global and local prompts. + +Binary cross-entropy loss and its variations such as asymmetric loss [25] have shown remarkable results in classification tasks [25, 28]. They are generally accompanied by a sigmoid function $\sigma(x) = 1 / (1 + \exp(-x))$ to convert model outputs to probabilities. Nevertheless, we found that directly optimizing the probability $\sigma(p)$ leads to a performance gap between training texts and testing images. We attribute this phenomenon to the modality gap between vision and language. Thus we consider ranking loss to be a more flexible and suitable way of supervision in the presence of the modality gap. A comparison of results between different loss functions is provided in Suppl. + +# 3.5. Incorporating with Prompting from Images + +Though our TaI-DPT is very different from existing methods of prompting from images, it is also complementary to them. To show this, we utilize an off-the-shelf prompt ensemble strategy to combine our TaI-DPT with existing methods in this section. As illustrated in Fig. 5, using CoOp [41] as an example, we can simply combine the scores of CoOp [41] and that of our TaI-DPT in a weighted sum manner. In particular, our TaI-DPT can be integrated with CoOp [41] when a few annotated images are provided and integrated with DualCoOp [28] when partially labeled training data are available. + +We ensemble prompts by fusing the predicted scores, rather than averaging the class embeddings generated by different prompts, since the image encoder used in different methods may be different (e.g. we conduct our experiments with ResNet50, while DualCoOp uses ResNet101 for partial-label prompting). So ensembling with the classification score is more convenient. In Sec. 4.3, we also empirically show that our prompt ensemble strategy is effective in advancing multi-label recognition performance in the few-shot and partially labeled settings. + +# 4. Experiments + +# 4.1. Implementation Details + +Architecture. We adopt CLIP ResNet-50 [24] as the visual encoder, and use the CLIP Transformer as the text encoder. During training, the parameters of the two encoders are kept frozen, and only learnable prompts are optimized. + +Learnable Prompts. Our learnable prompts are shared among classes of all datasets. Class-specific prompting [41] (i.e., an individual set of parameters for each category) has also been explored, but brings limited benefits. We initialize the value of each parameter with the Gaussian noise sampled from $\mathcal{N}(0,0.02)$ . The length of both the global prompts and local prompts are set to $M = 16$ , while a longer sequence brings trivial improvements. + +Datasets. To evaluate our TaI-DPT, we conduct the experiments on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9]. VOC2007 contains 20 common categories, and following [6,8,28], we form the training/test set based on the official trainval/test split (5,011 images/4,952 images). MS-COCO includes 80 categories with 82,081 training images and 40,504 testing images. NUS-WIDE includes 81 concepts with 161,789 images for training. We adopt its test set (107,859 images) to evaluate our method. For zero-shot experiments in Sec. 4.2, the training sets of the datasets are not used, and we use only text data to learn the prompts as mentioned in Sec. 3.2. Besides, for VOC2007 and MS-COCO, the language data sources are captions from MS-COCO. For NUS-WIDE, we introduce localized narratives from OpenImages [18], which have a + +![](images/725a6d4e13d14b027f1b2c58fe29684ae1fc057481359f891b8aebc29dcd038e.jpg) +Figure 6. Comparison of different methods in few-shot multi-label recognition on VOC2007 and MS-COCO. Our zero-shot TaI-DPT can achieve comparable results with methods trained by 16-shot labeled image samples. And learned prompt ensemble proofs the complementarity between images and texts. + +![](images/cc669f823ee0ce155b7f9f1bd2ff8dabceb87e0c74ceba00b4f9e1aff0ba5787.jpg) + +![](images/cb79468ea7ab2db108586757738f813ff4e1b9141864298059bd984958c7985b.jpg) +Figure 7. Ablation experiment on the number of texts and performance of TaI prompting on VOC2007. + +Table 1. Comparison with zero-shot methods on VOC2007, MS-COCO, and NUS-WIDE. Our proposed TaI-DPT outperforms CLIP [24] by a large margin on all datasets. + +
MethodDPTVOC2007MS-COCONUSWIDE
ZSCLIPX76.247.336.4
✓77.349.737.4
TaIX86.061.144.9
✓88.365.146.5
+ +broader range of content, to cover all the concepts in NUS-WIDE. In Sec. 4.3 and Sec. 4.4, for each dataset, the corresponding training data is used to conduct the experiments of partial-label and few-shot multi-label classification. + +Training Details. We adopt SGD optimizer, and the training epoch is set to 20 for all datasets. The learning rates for MS-COCO, VOC2007, and NUS-WIDE are empirically initialized with 1e-4, 1e-4, and 1e-3, and decay by the cosine annealing rule. For ranking loss, we choose $m = 1$ , and scale the $p$ and $p'$ by a factor of 4. $\tau_{s}$ is set as 0.02 via validation. + +# 4.2. Comparison with Zero-Shot Methods + +To demonstrate the effectiveness of our proposed TaI and DPT, we first compare it with the zero-shot CLIP (ZSCLIP). For fair comparison, we also introduce the DPT to ZSCLIP. Specifically, we adopt two identical default prompts "a photo of a [CLASS]" to separately deal with global and local features as DPT does. Table 1 lists the comparison results on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9] datasets. From the table, our TaI prompting surpasses ZSCLIP by a large margin of $9.8\%$ , $13.8\%$ , and $8.5\%$ mAP on VOC2007, MS-COCO, and NUS-WIDE, respectively, showing the effectiveness of our TaI. Furthermore, after training with fine-grained token features extracted from texts, our proposed DPT demonstrates a more powerful capability of discriminating local features than the default hand-crafted prompts and single global prompts. + +Table 2. Comparison with existing multi-label few-shot learning methods on MS-COCO. The evaluation is based on mAP for zero-shot, 1-shot and 5-shot with 16 novel classes. + +
Method0-shot1-shot5-shot
LaSO [2]-45.358.1
ML-FSL [27]-54.463.6
CoOp [41]40.2 (ZSCLIP)46.955.6
Tip-Adapter [38]40.2 (ZSCLIP)53.859.7
TaI-DPT59.2--
+ +# 4.3. Comparison with Few-Shot Methods + +We further compare with multi-label few-shot learning methods to verify the effectiveness of our TaI-DPT. In contrast to the well-studied single-label few-shot classification problem, few works tackle the multi-label few-shot scenario. Existing methods [2, 27] often deploy models trained on seen classes to few-shot novel classes. In Table 2, we compare zero-shot TaI-DPT to few-shot methods on 16 novel classes (see [2] for details about data split). TaI-DPT is comparable to the methods trained on 5-shot samples. + +Besides, we consider a new multi-label few-shot setting where all the classes are regarded as novel classes. We select 1, 2, 4, 8, and 16-shot samples for each category following the strategy in [2]. For fair comparison, we train CoOp [41] and our TaI in the same settings, and we also extend them with DPT for a more comprehensive comparison. For CoOp-DPT, we set two sets of learnable prompts, to deal with global and local features, respectively. The results are illustrated in Fig. 6. One can see that, even without any image data regarding novel classes, our TaI can achieve comparable results to CoOp trained on 16-shot. Similar trends with the MS-COCO dataset and the DPT setting support our observation that the discriminative feature of text data can be used as images for prompting. Moreover, benefiting from the flexibility of prompts, we can easily integrate our TaI-DPT with CoOp-DPT by ensembling as said in Sec. 4.3. As illustrated in Fig. 6, though CoOp-DPT has + +Table 3. Results of integrating our TaI-DPT with partial-label multi-label recognition method based on pre-trained CLIP. Our approach further improves the frontier performance of DualCoOp [28]. * indicates the results based on our own reproduction. + +
DatasetsMethod10%20%30%40%50%60%70%80%90%Avg.
MS-COCOSARB [23]71.275.077.178.378.979.679.880.580.577.9
DualCoOp [28]78.780.981.782.082.582.782.883.083.181.9
DualCoOp*81.082.382.983.483.583.984.084.184.383.3
+TaI-DPT81.582.683.383.783.984.084.284.484.583.6
PascalVOC 2007SARB [23]83.588.690.791.491.992.292.692.892.990.7
DualCoOp [28]90.392.292.893.393.693.994.094.194.293.2
DualCoOp*91.493.893.894.394.694.794.894.994.994.1
+TaI-DPT93.394.694.894.995.195.095.195.395.594.8
NUS-WIDEDualCoOp*54.056.256.957.457.957.957.658.258.857.2
+TaI-DPT56.457.957.858.158.558.858.659.159.458.3
+ +achieved a high accuracy, combining our prompts learned with text data still brings further improvement on recognition performance. This also proves that texts and images are complementary to each other to some extent. + +# 4.4. Integration with Partially Labeled Methods + +Partially-labeled data refers to the problem where merely some labels of each sample are known. Following [5], we created the partial-label data by randomly masking out labels of the fully annotated data. During inference, the model is evaluated on all categories. + +We reproduce DualCoOp [28] on partial-labeled VOC2007 and MS-COCO with the same experimental setting as reported (reproduced results are marked with *) and explore the enhancement brought by integration with our TaI-DPT. We also deploy DualCoOp on partial-labeled NUS-WIDE for a comprehensive comparison. + +The results are reported in Table 3. With no prior knowledge from pre-trained models, previous forefront method like SARB [23] struggles to learn from incomplete labels. DualCoOp, directly trained with entire image set, is well-learned to classify multi-label images. Even so, for PascalVOC and NUS-WIDE, by using texts from exogenous sources, i.e., COCO and OpenImages, our method can bring complementary enhancement to image prompting. While, for COCO, we used captions derived from its own images, making it more difficult to provide additional benefits. Analogous to VOC and NUS-WIDE, the COCO results can also be further advanced by introducing more external text data in future work. + +# 4.5. Ablation Study + +To thoroughly investigate the effect of each component, we conduct a series of ablation studies on the quantity of texts, training loss, different VL pre-trained models, ensemble weights, and texts v.s. images for prompting. The complete ablation experiments are shown in the Suppl. + +Quantity of texts. Following the data preparation procedure in Sec. 3.2, we end up with a total number of 66087 pieces of text that contain descriptions for 20 categories of VOC2007. We test the performance of TaI-DPT with different numbers of randomly selected texts, and the results are shown in Fig. 7. When no collected texts are available, 80 templates of hand-crafted prompts from [24], like "a cropped photo of a [CLASS]", are used for training (all templates are shown in the Suppl), and each template sentence correlates with one positive label corresponding to the class name inserted in [CLASS]. The increasing number of texts gradually forms a complete description of target categories, and the relationship between classes is also better characterized, which results in ascending performance. + +# 5. Conclusion + +In this paper, we propose a new view of treating texts as images in prompt tuning (i.e. TaI), which learns the prompt from discriminative features of text descriptions. Compared to prior prompt tuning methods trained with images, our TaI benefits from the easy accessibility of scalable content-rich texts, which enables prompt tuning for vision tasks (e.g., multi-label image recognition) even without downstream image data. Double-grained prompting is further introduced to utilize both the global and fine-grained features for better multi-label recognition ability. Nonetheless, when few-shot image samples or partial-labeled images are available, our TaI-DPT can conveniently integrate with existing prompting methods. Experiments on multiple benchmarks show the validity of TaI-DPT. + +# Acknowledgement + +This work was supported in part by National Key R&D Program of China under Grant No. 2020AAA0104500, and by the National Natural Science Foundation of China (NSFC) under Grant No. U19A2073. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1 +[2] Amit Alfassy, Leonid Karlinsky, Amit Aides, Joseph Shtok, Sivan Harary, Rogerio Feris, Raja Giryes, and Alex M Bronstein. Laso: Label-set operations networks for multi-label few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6548-6557, 2019. 2, 7 +[3] Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification, 2021. 2 +[4] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit." O'Reilly Media, Inc.", 2009. 4 +[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 339-346, 2022. 2, 8 +[6] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 522-531, 2019. 2, 6 +[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME 2019. 2 +[8] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In CVPR, 2019. 2, 6 +[9] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In Proceedings of the ACM international conference on image and video retrieval, pages 1-9,2009.2,6,7 +[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 2 +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 6, 7 +[12] Bin-Bin Gao and Hong-Yu Zhou. Multi-label image recognition with multi-class attentional regions. arXiv preprint arXiv:2007.01755, 2020. 2 +[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 3 +[14] Yunchao Gong, Yangqing Jia, Thomas Leung, Alexander Toshev, and Sergey Ioffe. Deep convolutional ranking for mul + +tilabel image annotation. arXiv preprint arXiv:1312.4894, 2013. 2, 6 +[15] Shiyi He, Chang Xu, Tianyu Guo, Chao Xu, and Dacheng Tao. Reinforced multi-label image classification by exploring curriculum. In AAAI, 2018. 2 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1 +[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 3 +[18] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Shahab Kamali, Matteo Malloci, Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://storage.googleapis.com/openimages/web/index.html, 2017. 2, 4, 6 +[19] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 4, 6, 7 +[21] Luchen Liu, Sheng Guo, Weilin Huang, and Matthew Scott. Decoupling category-wise independence and relevance with self-attention for multi-label image classification. In ICASSP 2019, 05. 2 +[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3 +[23] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. arXiv preprint arXiv:2203.02172, 2022. 2, 8 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 6, 7, 8 +[25] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6 + +[26] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. arXiv preprint arXiv:2209.07511, 2022. 3 +[27] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. Meta-learning for multi-label few-shot classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3951-3960, 2022. 2, 7 +[28] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. arXiv preprint arXiv:2206.09541, 2022. 1, 2, 3, 5, 6, 8 +[29] Ashwin Vaswani, Gaurav Aggarwal, Praneeth Netrapalli, and Narayan G Hegde. All mistakes are not equal: Comprehensive hierarchy aware multi-label predictions (champ). arXiv preprint arXiv:2206.08653, 2022. 2 +[30] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, 2016. 2 +[31] Yangtao Wang, Yanzhao Xie, Yu Liu, Ke Zhou, and Xiaocui Li. Fast graph convolution network based multi-label image recognition via cross-modal fusion. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 1575-1584, 2020. 2 +[32] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE transactions on pattern analysis and machine intelligence, 38, 2015. 2 +[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1 +[34] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3 +[35] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, 2020. 2 +[36] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1 +[37] Junjie Zhang, Qi Wu, Chunhua Shen, Jian Zhang, and Jianfeng Lu. Multilabel image classification with regional latent semantic dependencies. IEEE Transactions on Multimedia, 20, 2018. 2 +[38] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022. 7 +[39] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 3 + +[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3 +[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 4, 6, 7 +[42] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 3 \ No newline at end of file diff --git a/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/images.zip b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f9a0df456a59cb8ff17efe1f6080adf02299c94c --- /dev/null +++ b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:102c357205b053285f30896c8a3fa6279977dd5683350a7abd1de795a6f0b3b3 +size 520198 diff --git a/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/layout.json b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a9316c05e40885dc2c33f50ae99e79cb24a925ae --- /dev/null +++ b/2023/Texts as Images in Prompt Tuning for Multi-Label Image Recognition/layout.json @@ -0,0 +1,8196 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "text", + "content": "Texts as Images in Prompt Tuning for Multi-Label Image Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "spans": [ + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "content": "Zixian Guo" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "content": " Bowen Dong" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "content": " Zhilong Ji" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "content": " Jinfeng Bai" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "content": " Yiwen Guo" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "text", + "content": " Wangmeng Zuo" + }, + { + "bbox": [ + 116, + 144, + 475, + 157 + ], + "type": "inline_equation", + "content": "^{1,2(\\text{图})}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "text", + "content": "Harbin Institute of Technology " + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "text", + "content": "Tomorrow Advancing Life " + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "text", + "content": "Pazhou Lab, Guangzhou " + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 81, + 159, + 512, + 171 + ], + "type": "text", + "content": "Independent Researcher" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 174, + 478, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 174, + 478, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 174, + 478, + 185 + ], + "type": "text", + "content": "zixian.guo@foxmail.com cndongsky@gmail.com zhilongji@hotmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 188, + 456, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 188, + 456, + 198 + ], + "spans": [ + { + "bbox": [ + 129, + 188, + 456, + 198 + ], + "type": "text", + "content": "jfbai.bit@gmail.com guoyiwen89@gmail.com WMzuo@hit.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "type": "text", + "content": "Prompt tuning has been employed as an efficient way to adapt large vision-language pre-trained models (e.g. CLIP) to various downstream tasks in data-limited or label-limited settings. Nonetheless, visual data (e.g., images) is by default prerequisite for learning prompts in existing methods. In this work, we advocate that the effectiveness of image-text contrastive learning in aligning the two modalities (for training CLIP) further makes it feasible to treat texts as images for prompt tuning and introduce TaI prompting. In contrast to the visual data, text descriptions are easy to collect, and their class labels can be directly derived. Particularly, we apply TaI prompting to multi-label image recognition, where sentences in the wild serve as alternatives to images for prompt tuning. Moreover, with TaI, double-grained prompt tuning (TaI-DPT) is further presented to extract both coarse-grained and fine-grained embeddings for enhancing the multi-label recognition performance. Experimental results show that our proposed TaI-DPT outperforms zero-shot CLIP by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE, while it can be combined with existing methods of prompting from images to improve recognition performance further. The code is released at https://github.com/guozix/Tai-DPT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 560, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 693 + ], + "type": "text", + "content": "Recent few years have witnessed rapid progress in large vision-language (VL) pre-trained models [1, 16, 19, 24, 33, 36] as well as their remarkable performance on downstream vision tasks. A VL pre-trained model generally involves data encoders, and it is becoming increasingly popular to exploit image-test contrastive loss [24] to align the embedding of images and texts into a shared space. When adapting to downstream tasks in data-limited or label-limited settings, it is often ineffective to fine-tune the entire model, due to its high complexity. Then, prompt tuning as a representative parameter-efficient learning paradigm has emerged as" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 311, + 229, + 538, + 392 + ], + "blocks": [ + { + "bbox": [ + 311, + 229, + 538, + 392 + ], + "lines": [ + { + "bbox": [ + 311, + 229, + 538, + 392 + ], + "spans": [ + { + "bbox": [ + 311, + 229, + 538, + 392 + ], + "type": "image", + "image_path": "da7c14fde341874046b193f3d59da4fdd6a813b0af28615007eb5717bd026ab4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 395, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 304, + 395, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 547, + 473 + ], + "type": "text", + "content": "Figure 1. A comparison between prompting from images and our text-as-image (TaI) prompting. (a) Prompting from images (e.g., [41]) uses labeled images of task categories to learn the text prompts. Instead, (b) our TaI prompting learn the prompts with easily-accessed text descriptions containing target categories. (c) After training, the learned prompts in (a) or (b) can be readily applied to test images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 483, + 540, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 540, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 540, + 495 + ], + "type": "text", + "content": "an efficient way to adapt VL models to downstream tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "type": "text", + "content": "Albeit considerable achievements have been made, existing prompt tuning methods generally require visual data to learn prompts (as shown in Fig. 1(a)). For example, CoOp [41] learns from annotated images. CoCoOp [40] further introduces generalizable input-conditional prompts. DualCoOp [28] adapts CLIP to multi-label recognition tasks by training pairs of positive and negative prompts with partial-labeled images. Nonetheless, the performance of these prompting methods may be limited when it is infeasible to obtain sufficient image data or annotate the required images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we advocate treating Texts as Images for prompt tuning, i.e., TaI prompting. It is considered feasible as the image encoder and text encoder in many pre-trained VL models [16, 24] encode images and texts into a shared space. Given an image and its caption, the visual features produced by the image encoder will be close to the text feature of the caption produced by the text encoder. There-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 282, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 282, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 282, + 712 + ], + "type": "text", + "content": "*This work was done when Zixian Guo was a research intern at TAL." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2808" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "fore, in addition to extracting visual features from images, it is also feasible to extract text features as alternatives form, for example, descriptive sentences and captions, for prompt tuning (see Fig. 1(b)). TaI prompting has several interesting properties and merits. Taking a downstream image recognition task as an example, given a set of object categories, one can easily crawl a large set of text descriptions that contain object names from these categories. Text descriptions are easily accessible in this way, and class labels can be directly derived from text descriptions, which means, in contrast to prompting from images, TaI prompting may suffer less from the data-limited and label-limited issues." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 502 + ], + "type": "text", + "content": "We use multi-label image recognition [8, 9, 11, 20, 35] to verify the effectiveness of our TaI prompting in this paper. To begin with, we crawl the captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., Open Images [18]) to form the training set of text descriptions. For any specific multi-label recognition task, we adopt a noun filter to map the nouns in the text descriptions to the corresponding object categories, and then only keep the text descriptions that contain one or more classes of target objects. To better cope with multi-label classification, we introduce double-grained prompt tuning (i.e., TaI-DPT) which involves: (i) a set of global prompts to generate embeddings for classifying whole sentences or images, and (ii) a set of local prompts to extract embeddings for discriminating text tokens or image patches. Given a set of text descriptions, global and local prompts can be tuned by minimizing the ranking loss [14]. Note that, though these prompts are learned from text descriptions solely, they can be readily deployed to classify whole images as well as image patches during testing (see Fig. 1(c)). Experimental results show that, without using any labeled images, our TaI prompting surpasses zero-shot CLIP [24] by a large margin on multiple benchmarks, e.g., MS-COCO, VOC2007, and NUS-WIDE." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 504, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 289, + 635 + ], + "type": "text", + "content": "Moreover, when images are also available during training, our TaI prompting can be combined with existing methods of prompting from images to improve its performance. In particular, given a few annotated images, our TaI-DPT can be integrated with CoOp as a prompt ensemble for improving classification accuracy. With partially labeled training data being provided, we may also combine TaI-DPT and DualCoOp [28] to improve multi-label recognition accuracy consistently. Extensive results verify the effectiveness of our TaI-DPT in comparison to state-of-the-art. To sum up, the contributions of this work include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 642, + 288, + 714 + ], + "type": "text", + "content": "- We propose Texts as Images in prompt tuning (i.e., TaI prompting) to adapt VL pre-trained models to multi-label image recognition. Text descriptions are easily accessible and, in contrast to images, their class labels can be directly derived, making our TaI prompting very compelling in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 72, + 547, + 192 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 143 + ], + "type": "text", + "content": "- We present double-grained prompt tuning (i.e. TaI-DPT) to extract both coarse-grained and fine-grained embeddings for enhancing multi-label image recognition. Experiments on multiple benchmarks show that TaI-DPT achieves comparable multi-label recognition accuracy against state-of-the-arts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 144, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 144, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 547, + 192 + ], + "type": "text", + "content": "- The prompts learned by TaI-DPT can be easily combined with existing methods of prompting from images in an off-the-shelf manner, further improving multi-label recognition performance." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 197, + 392, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 197, + 392, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 197, + 392, + 209 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 217, + 476, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 476, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 476, + 229 + ], + "type": "text", + "content": "2.1. Multi-Label Image Recognition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 235, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 306 + ], + "type": "text", + "content": "Multi-label image recognition [3,7,8,12,15,21,32,35,37] aims to recognize all the object categories [11, 20] or concepts [9] in an input image. Various modules [6, 30] have been introduced to better represent the inter-class relationships, and modern classification losses [3, 14] have been used to make model learning easier." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 307, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 546, + 486 + ], + "type": "text", + "content": "To model the label dependencies, CNN-RNN [30] introduces recurrent neural networks, e.g., RNN and LSTM, to predict appeared classes in a sequential manner. [6,8,31,35] use graph convolution modules to learn the correlation between class labels. CHAMP [29] measures the severity of misclassification by building a domain-specific hierarchy tree according to the relation of categories, where each class is related to a tree node, to improve the robustness of the model. Albeit effective, these methods require a considerable number of labeled images to learn the category relationships sufficiently. While in data-limited or label-limited regimes, e.g., few-shot or partial-label data, it will be difficult for these models to learn well as expected. Specifically designed loss functions also struggle to obtain significant improvements when learning with limited data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "Multi-Label Recognition from Few-shot Samples. To better exploit the small number of samples, LaSO [2] synthesizes samples by manipulating the features of paired training images. Different ways of manipulating label sets are used to train the model, resulting in generalizable discriminative features. [27] introduces a meta-learning framework for better learning of past tasks and generalization to new tasks, and leverages the number of labels as useful information for learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Multi-Label Recognition from Partial-label Data. Partial-label refers to the scenarios where some labels are unknown. [10] propose a normalized BCE loss to balance the proportion of known labels. [5] learns to complement unknown labels by utilizing within-image and cross-image semantic correlations. [23] blends the representation of training images and class proxies to compensate for the loss of information due to unknown labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Albeit significant progress has been made, it remains a challenging issue for learning multi-label image recognition" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2809" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 276, + 231 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 276, + 231 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 276, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 276, + 231 + ], + "type": "image", + "image_path": "502d18cbf3bbd3f4cf9cf9bfa0f79de8324e2b9bbe0a5338d78a146061fb634d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 232, + 168, + 239 + ], + "lines": [ + { + "bbox": [ + 156, + 232, + 168, + 239 + ], + "spans": [ + { + "bbox": [ + 156, + 232, + 168, + 239 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 277, + 71, + 542, + 230 + ], + "blocks": [ + { + "bbox": [ + 277, + 71, + 542, + 230 + ], + "lines": [ + { + "bbox": [ + 277, + 71, + 542, + 230 + ], + "spans": [ + { + "bbox": [ + 277, + 71, + 542, + 230 + ], + "type": "image", + "image_path": "71fabcbc5e0d75b9ccb3a6fb219451d240d60858ebedce1c220902e79a68be4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 231, + 416, + 239 + ], + "lines": [ + { + "bbox": [ + 403, + 231, + 416, + 239 + ], + "spans": [ + { + "bbox": [ + 403, + 231, + 416, + 239 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "text", + "content": "Figure 2. Training and testing pipeline of our proposed Text-as-Image (TaI) prompting, where we use text descriptions instead of labeled images to train the prompts. (a) During training, we use two identical text encoders from pre-trained CLIP to extract the global & local class embeddings " + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbf{G} \\& \\mathbf{L})" + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "text", + "content": " and overall & sequential text embeddings " + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbf{h} \\& \\mathbf{H})" + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "text", + "content": " respectively from the prompts and text description. The corresponding cosine similarity " + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbf{p} \\& \\mathbf{P})" + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "text", + "content": " between the embeddings are guided by the derived pseudo labels with ranking loss. (b) During testing, we replace the input from text descriptions to images. The global and local class embeddings can discriminate target classes from global & local image features " + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbf{f} \\& \\mathbf{F})" + }, + { + "bbox": [ + 46, + 240, + 545, + 307 + ], + "type": "text", + "content": ". The final classification results are obtained by merging the scores of the two branches." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 376 + ], + "type": "text", + "content": "in image-limited or label-limited regimes. Built upon VL pre-trained models, this paper suggests learning prompts from texts instead of images, thereby offering a novel yet complementary perspective for handling low-resource multi-label image recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 384, + 277, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 277, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 277, + 396 + ], + "type": "text", + "content": "2.2. Prompt Tuning for Vision-Language Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 402, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 628 + ], + "type": "text", + "content": "To transfer pre-trained knowledge to downstream tasks in data-limited settings, prompt tuning [13, 17, 34, 39, 41, 42] has become a popular parameter-efficient way to achieve the goal, due to its flexibility and ease of use. CoOp [41] learns the prompts by using (a few) annotated images of each class from the target dataset. CoCoOp [40] further proposes to improve CoOp [41] by formulating the prompts in an image-conditional way to maintain better generalization to unseen classes. To avoid overfitting, ProGrad [42] leverages predictions from zero-shot CLIP to regularize gradients in prompt learning process. TPT [26] suggests optimizing test-time prompts by promoting the consistency of augmented test images. ProDA [22] uses multiple pieces of prompts to estimate the distribution of classifier weights for better handling of varying visual features. DualCoOp [28] firstly adapts CLIP to multi-label image recognition with partially labeled data by learning pairs of positive and negative prompts for each class to ensure independent binary classification for each class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Albeit existing prompt tuning approaches have achieved significant improvements in downstream tasks, images as well as a portion of class labels are prerequisite to supervise the optimization of the learnable prompts. In this paper, we propose to treat texts as images in prompt tuning, which, compared to labeled images, are much easier to collect with existing caption datasets and modern search engines. Our" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 316, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 364 + ], + "type": "text", + "content": "proposed TaI-DPT surpasses zero-shot CLIP by a large margin and can be combined with the prompts learned by existing methods of prompting from images to further boost recognition performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 376, + 411, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 411, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 411, + 391 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 397, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 554 + ], + "type": "text", + "content": "Here we present Text-as-Image prompting, i.e., TaI prompting, for adapting pre-trained VL models to multi-label image recognition. Our TaI prompting uses only easily-accessed free-form texts as training data to learn effective prompts for downstream multi-label recognition tasks. To begin with, We present an overview of TaI prompting in Sec. 3.1. Then, we introduce our preparation of training texts in Sec. 3.2. We further explain the design of the double-grained prompt tuning (i.e., TaI-DPT) in Sec. 3.3, and provide the loss function used to train the model in Sec. 3.4. Finally, we propose to ensemble prompts from TaI-DPT with prompts learned from images as a flexible integration to improve multi-label recognition performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 563, + 446, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 446, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 446, + 574 + ], + "type": "text", + "content": "3.1. Overview of Our Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Fig. 2 illustrates the design of our proposed TaI-DPT framework, including the training and testing phases. During training, we learn prompts with only supervision from texts. Two identical copies of the text encoder " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}_{\\mathrm{T}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " from the pre-trained CLIP are used to encode the prompts and text data, respectively. We introduce two sorts of trainable prompts (i.e., the global prompts and local prompts) to obtain global and regional class embeddings. A noun filtering strategy is used to generate classification pseudo-labels for each text description, which is applied to supervise the classification scores obtained by calculating the cosine sim" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2810" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "iliarity of class embeddings and text features. Only the parameters in prompts are optimized in the training phase, while the text encoders are both kept frozen. During testing, the class embeddings are obtained by encoding the two sets of learned prompts with the text encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}_{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " as in training, while the other input source changes from text descriptions to test images. Pre-trained image encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}_{\\mathrm{I}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " from CLIP is used to extract global and dense features of each test image, then the cosine similarity are computed between features and class embeddings of the global and local prompts. The final classification result is obtained by fusing the global and local cosine scores. In the following, we explain the details of our proposed method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 234, + 222, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 222, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 222, + 247 + ], + "type": "text", + "content": "3.2. Preparation of Text Descriptions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 252, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 443 + ], + "type": "text", + "content": "To obtain sufficient category information from the language that helps in image recognition, we have to ensure that: 1) the collected texts should contain content-rich descriptions of a scene, and 2) the contents of all text descriptions need to cover the category set of the target dataset for comprehensive representations of all categories. With the aim of ensuring reproducibility, we use readily available captions from the public image caption datasets (e.g., MS-COCO [20]) and localized narratives from object detection datasets (e.g., OpenImages [18]) as our language data source, while avoiding the workloads associated with randomly crawling texts from the Internet in this paper. Note that although each caption is paired with a corresponding image and human-annotated labels, we only use the captions, and no information from the pictures and labels is disclosed during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": "For a target multi-label recognition dataset " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": " that has a category set " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "S = \\{\\mathrm{s}_1, \\mathrm{s}_2, \\mathrm{s}_3, \\dots, \\mathrm{s}_C\\}" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": " denotes the number of categories and " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "\\mathrm{s}_i" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": " denotes particular class name like \"dog\", \"plane\", etc., we search for sentences that contain at least one class name " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "\\mathrm{s}_i" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": ". Since multiple words or phrases usually exist to represent the same meaning for each class, searching solely for exact match of category names in texts may lead to many false negatives in the obtained pseudo ground-truth labels, which is harmful to prompt tuning. Towards tackling this issue, we introduce a noun filter to map nouns with similar meanings into the corresponding class label. Specifically, we construct a synonym dictionary " + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 443, + 288, + 630 + ], + "type": "text", + "content": " by including common synonyms of each class name in the target dataset. If a word in a text description matches any synonym of a specific class name, it is considered to contain a description of that category. Several examples of synonyms are shown as follows:" + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 58, + 635, + 282, + 700 + ], + "blocks": [ + { + "bbox": [ + 58, + 635, + 282, + 700 + ], + "lines": [ + { + "bbox": [ + 58, + 635, + 282, + 700 + ], + "spans": [ + { + "bbox": [ + 58, + 635, + 282, + 700 + ], + "type": "text", + "content": "{'dog', 'pup', 'puppy', 'doggy'} \n{'person', 'people', 'man', 'woman', 'human'} \n{'bicycle', 'bike', 'cycle'} \n{'car', 'taxi', 'automobile'} \n{'boat', 'raft', 'dinghy'} \n..." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 47, + 701, + 284, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 284, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 284, + 714 + ], + "type": "text", + "content": "More details of the dictionary " + }, + { + "bbox": [ + 47, + 701, + 284, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 701, + 284, + 714 + ], + "type": "text", + "content": " are provided in the Suppl." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "Then we conduct noun filtration by the following steps. First, for each text description, we use the tokenizer and lemmatizer from NLTK [4] to recover the stem of each word in the sentences. Next, for all keywords in " + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": ", which contains all synonyms of the category set " + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": ", we search in our language data source for sentences that contains at least one class name. For the text descriptions that do not match any synonym of any class name, we simply drop them away to ensure each piece of data has at least one concerned label. Finally, for each retained text description, we convert the class names it contains into binary pseudo-ground-truth vectors by setting classes that appear as positive and other classes as negative, following the order of class labels in the target dataset " + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 239, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 546, + 348 + ], + "type": "text", + "content": "The word-level filtered labels may not be precisely correct since our searching strategy mentioned above is rather simple considering the diversity of free-form texts, where complex paraphrases and misspellings that widely exist in the corpus are not fully addressed. However, such a simple noun filtration can guarantee the reproducibility of this work, and our experiments also demonstrate that this simple and efficient data preparation leads to compelling results of multi-label recognition of our TaI prompting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 354, + 545, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 366 + ], + "type": "text", + "content": "3.3. Text-as-Image for Dual-grained Prompt Tuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 372, + 476, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 476, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 476, + 384 + ], + "type": "text", + "content": "Following [41], a prompt is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 390, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 390, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 367, + 390, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {i} = \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {v} _ {2}, \\boldsymbol {v} _ {3}, \\dots , \\boldsymbol {v} _ {M}, \\boldsymbol {s} _ {i} \\right] \\tag {1}", + "image_path": "ee35c6f5a53dac4611a7fddef0824930d4d90560e04e7a6de2170364c8b1a85b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "inline_equation", + "content": "i \\in \\{1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": " is the class index, " + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": " denotes word embedding of the " + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": "-th class name " + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 304, + 409, + 545, + 494 + ], + "type": "text", + "content": " is a learnable word embedding whose dimension is the same as the dimension of normal word embeddings in the vocabulary. Just like in previous methods, e.g. CoOp [41], the prompts are learned by maximizing the probability of classifying each image into its ground-truth class:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 498, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 498, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 498, + 545, + 529 + ], + "type": "interline_equation", + "content": "p (y = i | \\boldsymbol {x}) = \\frac {\\exp \\left(\\left\\langle \\operatorname {E n c} _ {\\mathrm {T}} \\left(\\boldsymbol {t} _ {i}\\right) , \\operatorname {E n c} _ {\\mathrm {I}} (\\boldsymbol {x}) \\right\\rangle / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\left\\langle \\operatorname {E n c} _ {\\mathrm {T}} \\left(\\boldsymbol {t} _ {j}\\right) , \\operatorname {E n c} _ {\\mathrm {I}} (\\boldsymbol {x}) \\right\\rangle / \\tau\\right)} \\tag {2}", + "image_path": "9200f49012250bb0bdd23a0108d5fe36fc80058ae2e97a962221ea3c04da7b09.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "content": " denotes the labeled training image and " + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 304, + 533, + 545, + 557 + ], + "type": "text", + "content": " calculates the cosine similarity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 653 + ], + "type": "text", + "content": "After large-scale image-text contrastive pre-training, text features have been well-aligned to the image features of the same semantic meanings. Thus, features of texts that describe a certain type of object also exhibit categorical discriminability. Therefore, based on the aligned VL representation, we advocate considering the feature of a piece of text description that describes a specific category, as an alternative to an image feature." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Apart from using the global sentence representation (i.e., the coarsest-grained text feature), we find that the sequential feature of word tokens from CLIP also possesses rich fine-grained information which is very similar to the region feature of dense image feature. In CLIP [24], cosine similarity" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2811" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 207, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 207, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 207, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 207, + 205 + ], + "type": "image", + "image_path": "88cf9e5f6a9ea7e84eb447ea37e472b63ce24d59a391cf18895f58773c023236.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 208, + 70, + 365, + 209 + ], + "blocks": [ + { + "bbox": [ + 208, + 70, + 365, + 209 + ], + "lines": [ + { + "bbox": [ + 208, + 70, + 365, + 209 + ], + "spans": [ + { + "bbox": [ + 208, + 70, + 365, + 209 + ], + "type": "image", + "image_path": "67a48dd6aa75a5946ce6d02959c4a313c72446431e6a2642ff4284badae7645a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 365, + 70, + 542, + 213 + ], + "blocks": [ + { + "bbox": [ + 365, + 70, + 542, + 213 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 542, + 213 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 542, + 213 + ], + "type": "image", + "image_path": "98b8d8e05a2fd435780dfac9dbcdce3af1365023f52c14438e582a47709ed25f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 246, + 286, + 381 + ], + "blocks": [ + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "text", + "content": "Figure 3. Visualization of correlations " + }, + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "text", + "content": " between the local class embedding " + }, + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "inline_equation", + "content": "\\pmb{L}" + }, + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "text", + "content": " and sequential token feature from texts. Each class embedding clearly correlates to words that describe the corresponding class (shown in highlight regions) rather than the global " + }, + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "inline_equation", + "content": "<\\mathrm{EOS}>" + }, + { + "bbox": [ + 46, + 219, + 546, + 243 + ], + "type": "text", + "content": " token." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 246, + 286, + 381 + ], + "lines": [ + { + "bbox": [ + 50, + 246, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 50, + 246, + 286, + 381 + ], + "type": "image", + "image_path": "fdb9e86433bebd7832a4daf80317954a0149179108449c53364544a2258375b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 384, + 288, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 418 + ], + "type": "text", + "content": "Figure 4. Visualization of correlations between the local class embedding " + }, + { + "bbox": [ + 46, + 384, + 288, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 46, + 384, + 288, + 418 + ], + "type": "text", + "content": " and dense image feature. The learned class embeddings can focus on the location of the object effectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 563 + ], + "type": "text", + "content": "between global image features, obtained by visual attention pooling, and global text features, obtained by projecting the feature of the last " + }, + { + "bbox": [ + 46, + 430, + 287, + 563 + ], + "type": "inline_equation", + "content": "<\\mathrm{EOS}>" + }, + { + "bbox": [ + 46, + 430, + 287, + 563 + ], + "type": "text", + "content": " token, are directly supervised with contrastive loss. In general, the global feature is sufficient for single-label classification because the target object usually is prominent in the picture. However, in multi-label recognition, the global feature is usually dominated by major objects, suppressing the recognition of non-significant objects concurrently existing in the image. Thus, it motivates us to explore fine-grained features and avoid the domination of the overly prominent object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 288, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 647 + ], + "type": "text", + "content": "To grasp the discriminability of global features as well as learn from fine-grained features, we propose double-grained prompt tuning (i.e., TaI-DPT) that uses two sets of prompts to handle global (i.e., the coarsest-grained level) and local (i.e., the fine-grained level) features, respectively, in two parallel branches. Formally, the double-grained prompt is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 652, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 652, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 287, + 671 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {i} ^ {G} = \\left[ \\boldsymbol {v} _ {1}, \\boldsymbol {v} _ {2}, \\boldsymbol {v} _ {3}, \\dots , \\boldsymbol {v} _ {M}, \\boldsymbol {s} _ {i} \\right], \\tag {3}", + "image_path": "7583987d1c4a6e0373c7e71279661eca970baa420ff27a6f78d37f65dd4cc22e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 670, + 228, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 670, + 228, + 684 + ], + "spans": [ + { + "bbox": [ + 107, + 670, + 228, + 684 + ], + "type": "interline_equation", + "content": "\\boldsymbol {t} _ {i} ^ {L} = \\left[ \\boldsymbol {v} _ {1} ^ {\\prime}, \\boldsymbol {v} _ {2} ^ {\\prime}, \\boldsymbol {v} _ {3} ^ {\\prime}, \\dots , \\boldsymbol {v} _ {M} ^ {\\prime}, \\boldsymbol {s} _ {i} \\right],", + "image_path": "1bc67385b21dd6e447995698fd34679aa9dafbf5b720768d542edc901fb1e493.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{v}_j" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{v}_j^\\prime" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " are learnable embeddings that are concatenated with word embedding " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "-th class to obtain the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": "global prompt " + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "inline_equation", + "content": "t_i^G" + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": " and local prompt " + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "inline_equation", + "content": "t_i^L" + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": ", respectively. The sequences in Eq. (3) are fed to a copy of the text encoder EncT of CLIP to generate global and local class embeddings for each class, i.e. " + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "inline_equation", + "content": "G_{i} = \\mathrm{Enc}_{\\mathrm{T}}(t_{i}^{G})" + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "inline_equation", + "content": "L_{i} = \\mathrm{Enc}_{\\mathrm{T}}(t_{i}^{L})" + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "inline_equation", + "content": "G = \\{G_{i}\\}_{i=1}^{C}" + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "inline_equation", + "content": "L = \\{L_{i}\\}_{i=1}^{C}" + }, + { + "bbox": [ + 304, + 247, + 546, + 380 + ], + "type": "text", + "content": " are encouraged to be correlated with global and local features, respectively. Note that the proposed double-grained prompts are different from dual prompts [28], which include a pair of contrastive positive and negative prompts for each class (More discussion about the differences between our method and DualCoOp is provided in the Suppl)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 547, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 439 + ], + "type": "text", + "content": "To preserve the fine-grained region features for the input image, we maintain the feature map before attention pooling layer of CLIP. As for the input text description, we preserve the sequential token features of the entire sentence instead of only the token features. So we have:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 447, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 447, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 384, + 447, + 545, + 465 + ], + "type": "interline_equation", + "content": "\\left\\{f, F \\right\\} = \\operatorname {E n c} _ {\\mathrm {I}} (x), \\tag {4}", + "image_path": "236f4ebde1be79480654879f529e5029001df7126ddfd19716838c752144fb89.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 462, + 469, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 462, + 469, + 475 + ], + "spans": [ + { + "bbox": [ + 384, + 462, + 469, + 475 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {h}, \\boldsymbol {H} \\right\\} = \\operatorname {E n c} _ {\\mathrm {T}} (\\boldsymbol {r}),", + "image_path": "fd0dfd4caa734b4eba752fb9ec1be992984c89bb63c5a81889eff977d00305ca.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": " denotes a piece of training text description. " + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "inline_equation", + "content": "f, h \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": " are the extracted global image and text features. " + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{N_1 \\times D}" + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "inline_equation", + "content": "\\pmb{H} \\in \\mathbb{R}^{N_2 \\times D}" + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": " are the flattened dense image features and sequential token features, respectively, where " + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "inline_equation", + "content": "N_1 = H \\times W" + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": " denotes the flattened spatial dimension of visual feature and " + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 304, + 482, + 547, + 554 + ], + "type": "text", + "content": " denotes the length of text tokens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 554, + 543, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 554, + 543, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 554, + 543, + 567 + ], + "type": "text", + "content": "Then, the global and local similarities are computed by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 575, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 575, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 361, + 575, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {i} = \\left\\langle \\boldsymbol {u}, \\boldsymbol {G} _ {i} \\right\\rangle , \\boldsymbol {P} _ {i j} = \\left\\langle \\boldsymbol {U} _ {j}, \\boldsymbol {L} _ {i} \\right\\rangle \\tag {5}", + "image_path": "2cb2654658c6a9e8b726a62f28155bf13b607aa01f9472085daa0d3ef0b62c47.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " denotes either language feature " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " in training or visual feature " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " in testing, and " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " coordinately. Information in local branch " + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 304, + 598, + 547, + 647 + ], + "type": "text", + "content": " (visualized in Fig. 3 and Fig. 4) is aggregated in a spatially weighted manner:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 346, + 653, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 653, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 346, + 653, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {i} ^ {\\prime} = \\sum_ {j = 1} ^ {N} \\frac {\\exp \\left(\\boldsymbol {P} _ {i j} / \\tau_ {s}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\boldsymbol {P} _ {i j} / \\tau_ {s}\\right)} \\cdot \\boldsymbol {P} _ {i j} \\tag {6}", + "image_path": "659489cb17d7ed5fe608a899b8bd81215fea535e0dd6ad60690d1ff8c38ea817.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "text", + "content": " accommodates the extent of focusing on a specific location. " + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^{\\prime}" + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "text", + "content": " are optimized by the loss terms " + }, + { + "bbox": [ + 304, + 689, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2812" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 272, + 183 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 272, + 183 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 272, + 183 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 272, + 183 + ], + "type": "image", + "image_path": "2c89f2e793150d87967aed74564435b54c080dab9e3f4e0253f7cad2359ce743.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 186, + 287, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 208 + ], + "type": "text", + "content": "Figure 5. Our learned double-grained prompt tuning is easy to combine with existing prompt tuning methods with ensemble." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "text", + "content": ", respectively, which we will discuss in Sec. 3.4. And in the testing phase, " + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{\\prime}" + }, + { + "bbox": [ + 46, + 220, + 286, + 255 + ], + "type": "text", + "content": " are combined to obtain the final classification score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 286, + 327 + ], + "type": "text", + "content": "The visualization in Fig. 3 and Fig. 4 show that the learned local class embedding " + }, + { + "bbox": [ + 46, + 256, + 286, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{L}" + }, + { + "bbox": [ + 46, + 256, + 286, + 327 + ], + "type": "text", + "content": " can focus on the specific location where corresponding classes appear, both in sentences and images, even if the fine-grained visual and language features are not explicitly supervised in the training of CLIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 338, + 160, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 160, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 160, + 350 + ], + "type": "text", + "content": "3.4. Learning Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "content": "The overall learning objective is defined as " + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{global} + \\mathcal{L}_{local}" + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "content": " are loss terms for global text embedding and local text tokens, respectively. We adopt the ranking loss [14] to measure the discrepancy between classification scores and pseudo-ground-truth labels. Specifically, " + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{global}" + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{local}" + }, + { + "bbox": [ + 46, + 357, + 286, + 439 + ], + "type": "text", + "content": " are formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 445, + 286, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 286, + 487 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {g l o b a l}} = \\sum_ {i \\in \\{c ^ {+} \\}} \\sum_ {j \\in \\{c ^ {-} \\}} \\max (0, m - p _ {i} + p _ {j}), \\tag {7} \\\\ \\mathcal {L} _ {l o c a l} = \\sum_ {i \\in \\{c ^ {+} \\}} \\sum_ {j \\in \\{c ^ {-} \\}} \\max (0, m - \\pmb {p} _ {i} ^ {\\prime} + \\pmb {p} _ {j} ^ {\\prime}) \\\\ \\end{array}", + "image_path": "72cb5e5f876ce8803eb4f341c3cd32cefb91dea21cf0f3153c793cd6c81f8c8b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{p}'" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " are global and aggregated local similarities described in Sec. 3.3, " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " is the margin controlling how much higher the similarity score with the positive classes is than with the negative classes. During training, we minimize the overall objective " + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 497, + 286, + 568 + ], + "type": "text", + "content": " with frozen text encoders, by optimizing the global and local prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Binary cross-entropy loss and its variations such as asymmetric loss [25] have shown remarkable results in classification tasks [25, 28]. They are generally accompanied by a sigmoid function " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\sigma(x) = 1 / (1 + \\exp(-x))" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " to convert model outputs to probabilities. Nevertheless, we found that directly optimizing the probability " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\sigma(p)" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " leads to a performance gap between training texts and testing images. We attribute this phenomenon to the modality gap between vision and language. Thus we consider ranking loss to be a more flexible and suitable way of supervision in the presence of the modality gap. A comparison of results between different loss functions is provided in Suppl." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 533, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 533, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 533, + 85 + ], + "type": "text", + "content": "3.5. Incorporating with Prompting from Images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 90, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 221 + ], + "type": "text", + "content": "Though our TaI-DPT is very different from existing methods of prompting from images, it is also complementary to them. To show this, we utilize an off-the-shelf prompt ensemble strategy to combine our TaI-DPT with existing methods in this section. As illustrated in Fig. 5, using CoOp [41] as an example, we can simply combine the scores of CoOp [41] and that of our TaI-DPT in a weighted sum manner. In particular, our TaI-DPT can be integrated with CoOp [41] when a few annotated images are provided and integrated with DualCoOp [28] when partially labeled training data are available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 221, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 342 + ], + "type": "text", + "content": "We ensemble prompts by fusing the predicted scores, rather than averaging the class embeddings generated by different prompts, since the image encoder used in different methods may be different (e.g. we conduct our experiments with ResNet50, while DualCoOp uses ResNet101 for partial-label prompting). So ensembling with the classification score is more convenient. In Sec. 4.3, we also empirically show that our prompt ensemble strategy is effective in advancing multi-label recognition performance in the few-shot and partially labeled settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 352, + 386, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 386, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 386, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 372, + 438, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 438, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 438, + 384 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "content": "Architecture. We adopt CLIP ResNet-50 [24] as the visual encoder, and use the CLIP Transformer as the text encoder. During training, the parameters of the two encoders are kept frozen, and only learnable prompts are optimized." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "text", + "content": "Learnable Prompts. Our learnable prompts are shared among classes of all datasets. Class-specific prompting [41] (i.e., an individual set of parameters for each category) has also been explored, but brings limited benefits. We initialize the value of each parameter with the Gaussian noise sampled from " + }, + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,0.02)" + }, + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "text", + "content": ". The length of both the global prompts and local prompts are set to " + }, + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "inline_equation", + "content": "M = 16" + }, + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "text", + "content": ", while a longer sequence brings trivial improvements." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Datasets. To evaluate our TaI-DPT, we conduct the experiments on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9]. VOC2007 contains 20 common categories, and following [6,8,28], we form the training/test set based on the official trainval/test split (5,011 images/4,952 images). MS-COCO includes 80 categories with 82,081 training images and 40,504 testing images. NUS-WIDE includes 81 concepts with 161,789 images for training. We adopt its test set (107,859 images) to evaluate our method. For zero-shot experiments in Sec. 4.2, the training sets of the datasets are not used, and we use only text data to learn the prompts as mentioned in Sec. 3.2. Besides, for VOC2007 and MS-COCO, the language data sources are captions from MS-COCO. For NUS-WIDE, we introduce localized narratives from OpenImages [18], which have a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2813" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 211, + 186 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 211, + 186 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 211, + 186 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 211, + 186 + ], + "type": "image", + "image_path": "725a6d4e13d14b027f1b2c58fe29684ae1fc057481359f891b8aebc29dcd038e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 373, + 238 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 373, + 238 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 373, + 238 + ], + "type": "text", + "content": "Figure 6. Comparison of different methods in few-shot multi-label recognition on VOC2007 and MS-COCO. Our zero-shot TaI-DPT can achieve comparable results with methods trained by 16-shot labeled image samples. And learned prompt ensemble proofs the complementarity between images and texts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 72, + 367, + 186 + ], + "blocks": [ + { + "bbox": [ + 211, + 72, + 367, + 186 + ], + "lines": [ + { + "bbox": [ + 211, + 72, + 367, + 186 + ], + "spans": [ + { + "bbox": [ + 211, + 72, + 367, + 186 + ], + "type": "image", + "image_path": "cc669f823ee0ce155b7f9f1bd2ff8dabceb87e0c74ceba00b4f9e1aff0ba5787.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 387, + 72, + 543, + 194 + ], + "blocks": [ + { + "bbox": [ + 387, + 72, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 387, + 72, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 387, + 72, + 543, + 194 + ], + "type": "image", + "image_path": "cb79468ea7ab2db108586757738f813ff4e1b9141864298059bd984958c7985b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 199, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 383, + 199, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 383, + 199, + 545, + 232 + ], + "type": "text", + "content": "Figure 7. Ablation experiment on the number of texts and performance of TaI prompting on VOC2007." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 283, + 287, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 246, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 280 + ], + "type": "text", + "content": "Table 1. Comparison with zero-shot methods on VOC2007, MS-COCO, and NUS-WIDE. Our proposed TaI-DPT outperforms CLIP [24] by a large margin on all datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 283, + 287, + 352 + ], + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 352 + ], + "type": "table", + "html": "
MethodDPTVOC2007MS-COCONUSWIDE
ZSCLIPX76.247.336.4
✓77.349.737.4
TaIX86.061.144.9
✓88.365.146.5
", + "image_path": "aecca2847ceff10c7b96bc3591ee4089b79baa5d8543ec765c0a947bfbee95bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 364, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 411 + ], + "type": "text", + "content": "broader range of content, to cover all the concepts in NUS-WIDE. In Sec. 4.3 and Sec. 4.4, for each dataset, the corresponding training data is used to conduct the experiments of partial-label and few-shot multi-label classification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "content": "Training Details. We adopt SGD optimizer, and the training epoch is set to 20 for all datasets. The learning rates for MS-COCO, VOC2007, and NUS-WIDE are empirically initialized with 1e-4, 1e-4, and 1e-3, and decay by the cosine annealing rule. For ranking loss, we choose " + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "content": ", and scale the " + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "content": " by a factor of 4. " + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 46, + 414, + 287, + 498 + ], + "type": "text", + "content": " is set as 0.02 via validation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 244, + 525 + ], + "type": "text", + "content": "4.2. Comparison with Zero-Shot Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed TaI and DPT, we first compare it with the zero-shot CLIP (ZSCLIP). For fair comparison, we also introduce the DPT to ZSCLIP. Specifically, we adopt two identical default prompts \"a photo of a [CLASS]\" to separately deal with global and local features as DPT does. Table 1 lists the comparison results on VOC2007 [11], MS-COCO [20], and NUS-WIDE [9] datasets. From the table, our TaI prompting surpasses ZSCLIP by a large margin of " + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "inline_equation", + "content": "13.8\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": " mAP on VOC2007, MS-COCO, and NUS-WIDE, respectively, showing the effectiveness of our TaI. Furthermore, after training with fine-grained token features extracted from texts, our proposed DPT demonstrates a more powerful capability of discriminating local features than the default hand-crafted prompts and single global prompts." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 281, + 538, + 365 + ], + "blocks": [ + { + "bbox": [ + 305, + 246, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 305, + 246, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 545, + 278 + ], + "type": "text", + "content": "Table 2. Comparison with existing multi-label few-shot learning methods on MS-COCO. The evaluation is based on mAP for zero-shot, 1-shot and 5-shot with 16 novel classes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 281, + 538, + 365 + ], + "lines": [ + { + "bbox": [ + 309, + 281, + 538, + 365 + ], + "spans": [ + { + "bbox": [ + 309, + 281, + 538, + 365 + ], + "type": "table", + "html": "
Method0-shot1-shot5-shot
LaSO [2]-45.358.1
ML-FSL [27]-54.463.6
CoOp [41]40.2 (ZSCLIP)46.955.6
Tip-Adapter [38]40.2 (ZSCLIP)53.859.7
TaI-DPT59.2--
", + "image_path": "c41d6a9e80ac4e3fee010fbd716e051d81069b25017646529a68fc133888de7e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 379, + 500, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 500, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 500, + 391 + ], + "type": "text", + "content": "4.3. Comparison with Few-Shot Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 399, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 507 + ], + "type": "text", + "content": "We further compare with multi-label few-shot learning methods to verify the effectiveness of our TaI-DPT. In contrast to the well-studied single-label few-shot classification problem, few works tackle the multi-label few-shot scenario. Existing methods [2, 27] often deploy models trained on seen classes to few-shot novel classes. In Table 2, we compare zero-shot TaI-DPT to few-shot methods on 16 novel classes (see [2] for details about data split). TaI-DPT is comparable to the methods trained on 5-shot samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Besides, we consider a new multi-label few-shot setting where all the classes are regarded as novel classes. We select 1, 2, 4, 8, and 16-shot samples for each category following the strategy in [2]. For fair comparison, we train CoOp [41] and our TaI in the same settings, and we also extend them with DPT for a more comprehensive comparison. For CoOp-DPT, we set two sets of learnable prompts, to deal with global and local features, respectively. The results are illustrated in Fig. 6. One can see that, even without any image data regarding novel classes, our TaI can achieve comparable results to CoOp trained on 16-shot. Similar trends with the MS-COCO dataset and the DPT setting support our observation that the discriminative feature of text data can be used as images for prompting. Moreover, benefiting from the flexibility of prompts, we can easily integrate our TaI-DPT with CoOp-DPT by ensembling as said in Sec. 4.3. As illustrated in Fig. 6, though CoOp-DPT has" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2814" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 96, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 3. Results of integrating our TaI-DPT with partial-label multi-label recognition method based on pre-trained CLIP. Our approach further improves the frontier performance of DualCoOp [28]. * indicates the results based on our own reproduction." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 96, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 96, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 531, + 242 + ], + "type": "table", + "html": "
DatasetsMethod10%20%30%40%50%60%70%80%90%Avg.
MS-COCOSARB [23]71.275.077.178.378.979.679.880.580.577.9
DualCoOp [28]78.780.981.782.082.582.782.883.083.181.9
DualCoOp*81.082.382.983.483.583.984.084.184.383.3
+TaI-DPT81.582.683.383.783.984.084.284.484.583.6
PascalVOC 2007SARB [23]83.588.690.791.491.992.292.692.892.990.7
DualCoOp [28]90.392.292.893.393.693.994.094.194.293.2
DualCoOp*91.493.893.894.394.694.794.894.994.994.1
+TaI-DPT93.394.694.894.995.195.095.195.395.594.8
NUS-WIDEDualCoOp*54.056.256.957.457.957.957.658.258.857.2
+TaI-DPT56.457.957.858.158.558.858.659.159.458.3
", + "image_path": "f585208bfc47bd7da28bae26bb4670baef9fbef3f77f3c872dc3352a30087b79.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 256, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 288, + 304 + ], + "type": "text", + "content": "achieved a high accuracy, combining our prompts learned with text data still brings further improvement on recognition performance. This also proves that texts and images are complementary to each other to some extent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 315, + 274, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 274, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 274, + 327 + ], + "type": "text", + "content": "4.4. Integration with Partially Labeled Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "content": "Partially-labeled data refers to the problem where merely some labels of each sample are known. Following [5], we created the partial-label data by randomly masking out labels of the fully annotated data. During inference, the model is evaluated on all categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 395, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 467 + ], + "type": "text", + "content": "We reproduce DualCoOp [28] on partial-labeled VOC2007 and MS-COCO with the same experimental setting as reported (reproduced results are marked with *) and explore the enhancement brought by integration with our TaI-DPT. We also deploy DualCoOp on partial-labeled NUS-WIDE for a comprehensive comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 624 + ], + "type": "text", + "content": "The results are reported in Table 3. With no prior knowledge from pre-trained models, previous forefront method like SARB [23] struggles to learn from incomplete labels. DualCoOp, directly trained with entire image set, is well-learned to classify multi-label images. Even so, for PascalVOC and NUS-WIDE, by using texts from exogenous sources, i.e., COCO and OpenImages, our method can bring complementary enhancement to image prompting. While, for COCO, we used captions derived from its own images, making it more difficult to provide additional benefits. Analogous to VOC and NUS-WIDE, the COCO results can also be further advanced by introducing more external text data in future work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "To thoroughly investigate the effect of each component, we conduct a series of ablation studies on the quantity of texts, training loss, different VL pre-trained models, ensemble weights, and texts v.s. images for prompting. The complete ablation experiments are shown in the Suppl." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 256, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 547, + 425 + ], + "type": "text", + "content": "Quantity of texts. Following the data preparation procedure in Sec. 3.2, we end up with a total number of 66087 pieces of text that contain descriptions for 20 categories of VOC2007. We test the performance of TaI-DPT with different numbers of randomly selected texts, and the results are shown in Fig. 7. When no collected texts are available, 80 templates of hand-crafted prompts from [24], like \"a cropped photo of a [CLASS]\", are used for training (all templates are shown in the Suppl), and each template sentence correlates with one positive label corresponding to the class name inserted in [CLASS]. The increasing number of texts gradually forms a complete description of target categories, and the relationship between classes is also better characterized, which results in ascending performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 439, + 379, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 379, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 379, + 452 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 628 + ], + "type": "text", + "content": "In this paper, we propose a new view of treating texts as images in prompt tuning (i.e. TaI), which learns the prompt from discriminative features of text descriptions. Compared to prior prompt tuning methods trained with images, our TaI benefits from the easy accessibility of scalable content-rich texts, which enables prompt tuning for vision tasks (e.g., multi-label image recognition) even without downstream image data. Double-grained prompting is further introduced to utilize both the global and fine-grained features for better multi-label recognition ability. Nonetheless, when few-shot image samples or partial-labeled images are available, our TaI-DPT can conveniently integrate with existing prompting methods. Experiments on multiple benchmarks show the validity of TaI-DPT." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 644, + 404, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 404, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 404, + 658 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "This work was supported in part by National Key R&D Program of China under Grant No. 2020AAA0104500, and by the National Natural Science Foundation of China (NSFC) under Grant No. U19A2073." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2815" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 212 + ], + "type": "text", + "content": "[2] Amit Alfassy, Leonid Karlinsky, Amit Aides, Joseph Shtok, Sivan Harary, Rogerio Feris, Raja Giryes, and Alex M Bronstein. Laso: Label-set operations networks for multi-label few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6548-6557, 2019. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "type": "text", + "content": "[3] Emanuel Ben-Baruch, Tal Ridnik, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "type": "text", + "content": "[4] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python: analyzing text with the natural language toolkit.\" O'Reilly Media, Inc.\", 2009. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 334 + ], + "type": "text", + "content": "[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 339-346, 2022. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "type": "text", + "content": "[6] Tianshui Chen, Muxin Xu, Xiaolu Hui, Hefeng Wu, and Liang Lin. Learning semantic-specific graph representation for multi-label image recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 522-531, 2019. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[7] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. In ICME 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "type": "text", + "content": "[8] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "text", + "content": "[9] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofsingapore.In Proceedings of the ACM international conference on image and video retrieval, pages 1-9,2009.2,6,7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "type": "text", + "content": "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[12] Bin-Bin Gao and Hong-Yu Zhou. Multi-label image recognition with multi-class attentional regions. arXiv preprint arXiv:2007.01755, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[13] Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning. arXiv preprint arXiv:2202.06687, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Yunchao Gong, Yangqing Jia, Thomas Leung, Alexander Toshev, and Sergey Ioffe. Deep convolutional ranking for mul" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tilabel image annotation. arXiv preprint arXiv:1312.4894, 2013. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 129 + ], + "type": "text", + "content": "[15] Shiyi He, Chang Xu, Tianyu Guo, Chao Xu, and Dacheng Tao. Reinforced multi-label image classification by exploring curriculum. In AAAI, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 129, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 194 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 348 + ], + "type": "text", + "content": "[18] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Shahab Kamali, Matteo Malloci, Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://storage.googleapis.com/openimages/web/index.html, 2017. 2, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[19] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 4, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[21] Luchen Liu, Sheng Guo, Weilin Huang, and Matthew Scott. Decoupling category-wise independence and relevance with self-attention for multi-label image classification. In ICASSP 2019, 05. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "text", + "content": "[22] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[23] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. arXiv preprint arXiv:2203.02172, 2022. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 658 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2816" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[26] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. arXiv preprint arXiv:2209.07511, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[27] Christian Simon, Piotr Koniusz, and Mehrtash Harandi. Meta-learning for multi-label few-shot classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3951-3960, 2022. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 219 + ], + "type": "text", + "content": "[28] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. arXiv preprint arXiv:2206.09541, 2022. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 286, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 286, + 265 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 286, + 265 + ], + "type": "text", + "content": "[29] Ashwin Vaswani, Gaurav Aggarwal, Praneeth Netrapalli, and Narayan G Hegde. All mistakes are not equal: Comprehensive hierarchy aware multi-label predictions (champ). arXiv preprint arXiv:2206.08653, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 268, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 286, + 300 + ], + "type": "text", + "content": "[30] Jiang Wang, Yi Yang, Junhua Mao, Zhiheng Huang, Chang Huang, and Wei Xu. Cnn-rnn: A unified framework for multi-label image classification. In CVPR, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 303, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 286, + 357 + ], + "type": "text", + "content": "[31] Yangtao Wang, Yanzhao Xie, Yu Liu, Ke Zhou, and Xiaocui Li. Fast graph convolution network based multi-label image recognition via cross-modal fusion. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 1575-1584, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 359, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 286, + 413 + ], + "type": "text", + "content": "[32] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE transactions on pattern analysis and machine intelligence, 38, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 416, + 286, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 286, + 460 + ], + "type": "text", + "content": "[33] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 462, + 286, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 286, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 286, + 505 + ], + "type": "text", + "content": "[34] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 508, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 508, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 508, + 286, + 540 + ], + "type": "text", + "content": "[35] Jin Ye, Junjun He, Xiaojiang Peng, Wenhao Wu, and Yu Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. In ECCV, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 543, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 286, + 597 + ], + "type": "text", + "content": "[36] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 286, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 286, + 642 + ], + "type": "text", + "content": "[37] Junjie Zhang, Qi Wu, Chunhua Shen, Jian Zhang, and Jianfeng Lu. Multilabel image classification with regional latent semantic dependencies. IEEE Transactions on Multimedia, 20, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 689 + ], + "type": "text", + "content": "[38] Renrui Zhang, Zhang Wei, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "text", + "content": "[39] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 206 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[40] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "text", + "content": "[42] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. arXiv preprint arXiv:2205.14865, 2022. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2817" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_content_list.json b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b83bf63a7496705e9dc7cc064693ed596a458279 --- /dev/null +++ b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_content_list.json @@ -0,0 +1,1709 @@ +[ + { + "type": "text", + "text": "Texture-guided Saliency Distilling for Unsupervised Salient Object Detection", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huajun Zhou $^{1}$ , Bo Qiao $^{1}$ , Lingxiao Yang $^{1}$ , Jianhuang Lai $^{1,2,3}$ , Xiaohua Xie $^{1,2,3*}$", + "bbox": [ + 171, + 179, + 802, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China", + "bbox": [ + 171, + 198, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Guangdong Province Key Laboratory of Information Security Technology, China", + "bbox": [ + 161, + 215, + 808, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 96, + 233, + 872, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 312, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Learning-based Unsupervised Salient Object Detection (USOD) mainly relies on the noisy saliency pseudo labels that have been generated from traditional handcraft methods or pre-trained networks. To cope with the noisy labels problem, a class of methods focus on only easy samples with reliable labels but ignore valuable knowledge in hard samples. In this paper, we propose a novel USOD method to mine rich and accurate saliency knowledge from both easy and hard samples. First, we propose a Confidence-aware Saliency Distilling (CSD) strategy that scores samples conditioned on samples' confidences, which guides the model to distill saliency knowledge from easy samples to hard samples progressively. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the boundaries of noisy labels by matching the textures around the predicted boundaries. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance. Code is available at www.github.com/moothes/A2S-v2.", + "bbox": [ + 75, + 318, + 473, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 650, + 209, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised Salient Object Detection (USOD) methods aim to correctly localize and precisely segment salient objects simultaneously without using manual annotations. Compared to the supervised methods, USOD methods can easily adapt to more practical scenarios (e.g., industrial or medical images) where a large number of labeled images may be very hard to collect. Moreover, USOD methods also can assist some related methods for other tasks, e.g., object recognition [16, 44] and object detection [11, 35]. However, diverse objects, complex backgrounds, and other challenges", + "bbox": [ + 75, + 676, + 468, + 828 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/177627ce4564029ce13663453f419d74dfe7eb917f32ecac11e5693484224b9a.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 506, + 284, + 599, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/268fbdf8b043a26f2a8b3e41c3f4b2e43a40379c07fc2c6b920fb47a1d606d7f.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 602, + 285, + 696, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccd16f9f38784400db3c1a5561208d0252c3ad673d1075a6c13c04839589409a.jpg", + "image_caption": [ + "(c) RBD" + ], + "image_footnote": [], + "bbox": [ + 697, + 285, + 792, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faae30c40b0e6bf5f9603ed1f81264770393f9e285721d78245b02b8e9ed4fc3.jpg", + "image_caption": [ + "(d) HS" + ], + "image_footnote": [], + "bbox": [ + 794, + 285, + 888, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59f89ce8ff5861b1f2e67b3c7b774e17d7937050e829b1dde0a52dcd66fa6d2e.jpg", + "image_caption": [ + "(e) USPS" + ], + "image_footnote": [], + "bbox": [ + 504, + 356, + 599, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18359c3002312e97af235bd60b4545e7b71182ff888a9670f909bab56e9c6123.jpg", + "image_caption": [ + "(f) A2S" + ], + "image_footnote": [], + "bbox": [ + 602, + 356, + 696, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecd4f700752bfb27c7755b432a22a01af5847be04bcbf8bb3af961fb482c5995.jpg", + "image_caption": [ + "(g) Activation", + "Figure 1. Saliency maps generated by USOD methods. Our result is generated based on the activation map of a deep network." + ], + "image_footnote": [], + "bbox": [ + 697, + 356, + 792, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27c0cf6292ecebe1168c2ac0070e99bb3231f34e9bfbd1349693c44a92980d06.jpg", + "image_caption": [ + "(h) Ours" + ], + "image_footnote": [], + "bbox": [ + 794, + 356, + 888, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing conditions bring severe challenges to USOD methods.", + "bbox": [ + 500, + 487, + 880, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most Deep Learning-based (DL-based) methods [21, 29, 32, 55, 63, 66] base on the saliency cues extracted by traditional SOD methods (Fig. 1-c and 1-d). These handcrafted features related cues are employed as pseudo labels to train deep networks under certain constraints, e.g., binary cross-entropy (BCE) loss. However, saliency cues by traditional methods usually shift away from target objects, especially in complex scenes. Moreover, conventional constraints, such as BCE loss, works well on fully-supervised SOD methods, but is suboptimal when fitting the noisy labels for unsupervised methods (Fig. 1-e). Recently, Zhou et al. [72] addressed the first issue by extracting saliency cues (Fig. 1-f) based on a unsupervisedly pre-trained network (Fig. 1-g) instead of using traditional methods. During training, they focus on learning reliable saliency knowledge from easy samples, but ignore latent knowledge in hard samples. The main reason is that hard samples may be wrongly-labeled and corrupt the fragile saliency knowledge learned in the early training phase. Therefore, to leverage hard samples, we argue that all samples should be employed in a meaningful order (i.e., from high reliable to low reliable), which is crucial for mining accurate knowledge from noisy labels. Trained by such a strategy, the network can mine valuable knowledge from hard examples without corrupting the knowledge learned from easy samples.", + "bbox": [ + 496, + 503, + 890, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep networks can learn to localize salient regions from", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. This project is supported by the Key-Area Research and Development Program of Guangdong Province (2019B010155003), the National Natural Science Foundation of China (U22A2095, 62072482, 62076258, 62206316), and the Guangdong NSF Project (2022A1515011254).", + "bbox": [ + 75, + 838, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7257", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noisy labels [72], but still struggle to find the precise boundaries of target objects. Generally, the appearance around saliency boundary has a similar texture as in saliency map. Therefore, matching the textures between different maps can serve as a guidance for producing reasonable saliency boundaries. We will demonstrate that above strategies are applicable to multimodal data besides RGB image, including depth map, thermal image, and optical flow.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above analysis, we propose a novel framework to tackle the Unsupervised Salient Object Detection (USOD) tasks. Specifically, two strategies are proposed to mine saliency knowledge from noisy saliency labels. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme that scores samples with noisy labels conditioned on samples' confidence. Then, our CSD guides the network to learn saliency knowledge from easy samples to more complex ones progressively by employing an adaptive loss conditioned on the training progress. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the saliency boundaries of noisy labels by matching the textures around the predicted boundaries. During training, the predicted saliency boundaries are shifting toward surrounding edges in the appearance space of the whole image. Finally, guided by above two mechanisms, our method can produce high-quality pseudo labels to train generalized saliency detectors. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art performance compared to existing USOD methods.", + "bbox": [ + 75, + 212, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our novel USOD method are:", + "bbox": [ + 96, + 530, + 467, + 544 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a Confidence-aware Saliency Distilling (CSD) to mine rich and accurate saliency knowledge from noisy labels, which breaks through the limitation that existing methods cannot utilize hard samples.", + "2. We propose a Boundary-aware Texture Matching (BTM) to refine the boundary of the predicted saliency maps by matching textures in different spaces.", + "3. Extensive experiments on RGB, RGB-D, RGB-T and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance." + ], + "bbox": [ + 89, + 556, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 744, + 225, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Supervised Salient Object Detection", + "text_level": 1, + "bbox": [ + 76, + 770, + 388, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Researchers have developed a large family of fully-supervised Salient Object Detection (SOD) algorithms [5,7, 15,19,30,39,54,58,67-69,71,73] in the past decades. Ronneberger et al. [41] proposed a U-shape structure that progressively upsamples and concatenates the smaller features to the larger ones. To ease the annotation burden, Zhang et al. [65] relabeled the DUTS-TR dataset [52] with scribbles", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and leveraged an edge detection model for boundary localization. Yu et al. [62] proposed a local coherence loss to find precise boundary based on scribbles annotations.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal SOD tasks aim at using other modality data to improve the SOD performance, such as depth map, thermal image and optical flow. Recently, abundant methods [4,8,14,17,40,45-48,57,74] were proposed. Most of these methods employ a two-stream encoder-decoder structure to aggregate multi-level information in multimodal data. To reduce the annotating cost, Zhao et al. [70] proposed a video SOD dataset with scribbles annotations to indicate the location of salient objects.", + "bbox": [ + 496, + 136, + 893, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although the above methods have achieved significant performance, they require numerous human annotations for training, which are expensive to collect.", + "bbox": [ + 496, + 272, + 893, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unsupervised Salient Object Detection", + "text_level": 1, + "bbox": [ + 500, + 325, + 831, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional SOD methods [22, 26, 60, 76] extracted saliency cues from images by modeling the correlation of hand-crafted features. Inspired by the centralization prior, Jiang et al. [22] considered the distances between boundary superpixels and non-boundary superpixels as saliency scores. Yan et al. [60] employed a tree-structure graphical model to compute the saliency results based on several over-segmented maps. The above methods fail to accurately localize salient objects because the global information in hand-crafted features are not representative enough.", + "bbox": [ + 496, + 349, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing DL-based USOD methods can be divided into two pipelines based on the method used to extract saliency cues from images. First, most USOD methods [29, 32, 55, 63, 64, 66] focused on refining the coarse saliency cues extracted by several traditional SOD methods. For example, Zhang et al. [63] weighted these saliency cues by combining both intra-image and inter-image fusion streams. Zhang et al. [66] designed a noise modeling module to deal with noises in these saliency cues. Nguyen et al. [32] used these saliency cues as labels to train multiple deep networks. Ji et al. [21] refined the saliency maps extracted by traditional SOD methods to produce more accurate saliency predictions. Second, to prevent the localization errors caused by traditional SOD methods, Zhou et al. [72] proposed a novel framework that converts the activation maps of a pre-trained network to high-quality pseudo labels.", + "bbox": [ + 496, + 500, + 890, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose two novel mechanisms to tackle USOD tasks and explore the potential of improving the quality of pseudo labels by using multimodal data.", + "bbox": [ + 496, + 742, + 890, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 500, + 800, + 648, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our method, we propose two novel strategies to mine accurate saliency knowledge based on the noisy activation maps generated by a deep network $\\Phi$ , as shown in Fig. 2.", + "bbox": [ + 496, + 825, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Activation map generation. Following previous work [72], we generate an activation map $Y$ for input image $X$", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7258", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9f3403048fd88e99f5dac79ba7f3639bd8d12656dd8267e399a241f37b53610.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed unsupervised framework. The network are guided by two mechanisms: (1) Conidence-aware Saliency Distilling mines rich and accurate saliency knowledge from noisy labels, and (2) Boundary-aware Texture Matching makes the boundary of the predicted saliency map more accurate. We pad some constant maps for the missing modalities." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using a deep network $\\Phi(X) = Y$ :", + "bbox": [ + 76, + 397, + 305, + 412 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {3}, E _ {4}, E _ {5} = \\operatorname {E n c o d e r} (X),\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 425, + 310, + 441 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {i} = S E \\left(E _ {i}\\right), i \\in \\{3, 4, 5 \\}, \\\\ H = S E \\left(\\text {一} (E _ {i}, E _ {i}, E _ {i})\\right) \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 444, + 468, + 468 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH = S E (\\text {c o n c a t} (F _ {3}, F _ {4}, F _ {5})),\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 463, + 385, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY = \\operatorname {i n v} (\\operatorname {S i g m o i d} (\\operatorname {s u m} (H - \\bar {H}))),\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 481, + 426, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{H}$ is the spatial mean of $H$ . Specifically, we employ the ResNet-50 [18] pre-trained by MoCo-v2 [6] as our encoder, which is trained without extra manual annotations. In our decoder, four SE blocks [20] integrate multiple encoder features into $H$ . After that, we set $H - \\bar{H}$ to ensure that (1) noisy labels are adaptive to input images when using a fixed threshold and (2) coexisting of positive and negative samples. Next, features are summed over the channel dimension to produce one-channel activation map. Finally, the Sigmoid function produces the final saliency scores, and the inv function identifies the regions with more corner pixels as background.", + "bbox": [ + 75, + 511, + 468, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training strategy. The activation maps produced by network $\\Phi$ perceive some discriminative regions in input images [72], but still are low-quality because of widespread noises. To improve the quality, we employ three strategies to train the network $\\Phi$ by:", + "bbox": [ + 75, + 693, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\Phi} = \\lambda_ {c} \\left(L _ {c s d} + \\hat {L} _ {c s d}\\right) + \\lambda_ {b} \\left(L _ {b t m} + \\hat {L} _ {b t m}\\right) + \\lambda_ {m} L _ {m s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 779, + 468, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where all $\\lambda$ are hyperparameters. $\\hat{L}$ is the loss for the predictions of a different scale. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme to excavate valuable saliency knowledge from simple examples to more complex ones. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to align the boundaries in", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "appearance and the predicted saliency maps. In addition, a multi-scale consistency loss $L_{ms}$ ensures our method to produce consistent predictions for multi-scale inputs.", + "bbox": [ + 498, + 397, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Confidence-aware Saliency Distilling", + "text_level": 1, + "bbox": [ + 498, + 453, + 816, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary challenge for USOD methods is how to localize salient regions. Existing DL-based USOD methods [32,63,72] are based on noisy saliency cues from traditional SOD methods [22,26,76,77] or pre-trained networks [6,18]. Following [72], we employ the activation maps produced by a pre-trained network as our saliency cues instead of the saliency predictions of traditional SOD methods. The main reasons are two-fold: (1) traditional methods require extra computation loads; (2) network can focus on mining saliency knowledge instead of fitting the inductive bias of traditional methods. We binarize the saliency cues as the initial labels using a fixed threshold of 0.5, while Eqn. 1 ensures that these labels remain adaptive to the input image.", + "bbox": [ + 496, + 477, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate high-quality pseudo labels, USPS [32] uses these initial labels to train deep networks like other fully-supervised methods. However, the simple training strategy in fully-supervised methods is suboptimal for unsupervised methods. Specifically, we define pixels with saliency scores close to 0.5 as hard examples because of their low confidences. In the noisy labels, hard examples are likely to be wrongly labeled. Therefore, it is difficult to learn robust saliency knowledge from these hard samples using traditional loss functions (e.g., BCE loss). To this end, A2S [72] learns reliable saliency knowledge from easy samples. However, for hard examples, the saliency knowledge hidden in noises is not fully explored. In summary, a customized strategy that organizes the samples in a more meaningful order is crucial for USOD methods.", + "bbox": [ + 496, + 674, + 892, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7259", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d5307d04d37199010b5171636846edb834e620c762c96ef3e77705d129c57532.jpg", + "image_caption": [ + "(a) $\\mathrm{L}_1$" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 478, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/45f3d08fa0bdf4f49ca0d8b86556e7ad8757057b1f3417f4417299ff59654473.jpg", + "image_caption": [ + "(b) $\\mathrm{L}_{\\mathrm{bce}}$" + ], + "image_footnote": [], + "bbox": [ + 486, + 89, + 869, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/952434bf3adb66ac367a13640254c1e139d3a85115fd9e637c9459873ab3dc4d.jpg", + "image_caption": [ + "(c) $\\mathrm{L_{adb}}$", + "Figure 3. The gradient landscapes of different loss functions." + ], + "image_footnote": [], + "bbox": [ + 96, + 196, + 480, + 295 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c51815631348ac1e794cfc859084468d3de4b8c23d62a5131b7184473e736585.jpg", + "image_caption": [ + "(d) $\\mathrm{L}_{\\mathrm{csd}}$ (Ours)" + ], + "image_footnote": [], + "bbox": [ + 486, + 196, + 869, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the above problem, we propose a Confidence-aware Saliency Distillation (CSD) scheme that scores samples with noisy labels conditional on their confidence and training progress. Concretely, easy samples contain reliable knowledge, and thus can assist our method to learn reliable saliency knowledge. On the contrary, the saliency knowledge in hard samples is hidden within noises, and may corrupt the fragile saliency patterns in the early stages of network training. Inspired by self-paced learning [2], we dynamically adjust the gradients for samples by introducing a factor $\\rho$ , which is linearly increasing from 0 to 1 as training proceeding. Our $L_{csd}$ loss can be formulated as:", + "bbox": [ + 75, + 339, + 473, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c s d} = - \\frac {1}{N} \\sum_ {i} ^ {N} | \\Phi (\\mathbf {x} _ {i}) - 0. 5 | ^ {(2 ^ {1 - \\rho})}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 529, + 468, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of pixels, and $\\Phi (\\mathbf{x}_i)$ is the saliency prediction of pixel $\\mathbf{x}_i$ . Notice that noisy label $\\mathbf{G}$ is omitted in Eqn. 3 because it is generated based on $\\Phi (\\mathbf{x}_i)$ . Thus, we can calculate the confidence score $|k_{i}|$ of pixel $\\mathbf{x}_i$ in a simpler way as: $|k_{i}| = |\\Phi (\\mathbf{x}_{i}) - 0.5|$ , where the constant value 0.5 is related to $\\bar{H}$ in Eqn. 1. The partial derivative of our $L_{csd}$ over $\\Phi (\\mathbf{x}_i)$ is:", + "bbox": [ + 75, + 579, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L _ {c s d}}{\\partial \\Phi (\\mathbf {x} _ {i})} = - \\operatorname {s i g n} \\left(k _ {i}\\right) 2 ^ {(1 - \\rho)} \\left| k _ {i} \\right| ^ {2 ^ {(1 - \\rho)} - 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 694, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{sign}(k_i) \\in \\{-1, 1\\}$ for negative and positive values, respectively.", + "bbox": [ + 75, + 734, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For an intuitive comparison, we draw the gradient landscapes of different losses in Fig. 3, including $L_{1}$ , $L_{bce}$ , $L_{adb}$ [72] and our $L_{csd}$ . The gradient of other losses are consistent throughout the training process, while our $L_{csd}$ varies with the training process. Specifically, in the beginning, our $L_{csd}$ assigns low gradients to hard samples to learn reliable saliency knowledge from easy samples. As training proceeding, the gradients of hard samples are increasing to mine more valuable saliency knowledge.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3aa6303ecb981a1b07e5e5d25e23d2a66ed516e7077aa2f6638749a8c102f825.jpg", + "image_caption": [ + "Figure 4. Appearance information in multimodal data." + ], + "image_footnote": [], + "bbox": [ + 501, + 337, + 893, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Boundary-aware Texture Matching", + "text_level": 1, + "bbox": [ + 500, + 517, + 808, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, the appearance of saliency boundary has a similar texture as in saliency prediction. Therefore, matching the textures between different maps can guide our method to produce reasonable saliency scores. This strategy is also applicable to other modalities, such as depth map, thermal image, and optical flow, as shown in Fig. 4. By aggregating the rich appearance information in multimodal data, we can provide a more generalized guidance for our USOD method.", + "bbox": [ + 496, + 542, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above analysis, we propose a Boundary-aware Texture Matching (BTM) strategy to align the saliency boundaries and image edges by matching their textures. First of all, similar to the LBP feature [33], we extract texture vectors from saliency predictions and appearance, respectively. For saliency predictions, the texture vector $T_{i}^{s} = [t_{i,1}^{s}, t_{i,2}^{s}, \\dots, t_{i,k^{2}}^{s}]$ for the $i$ -th pixel is computed by $t_{i,j}^{s} = |\\Phi(\\mathbf{x}_{i}) - \\Phi(\\mathbf{x}_{j})|, j \\in K_{i}$ , where $K_{i}$ is $k \\times k$ neighborhoods around pixel $\\mathbf{x}_{i}$ . For appearance information, it may be one of RGB image, optical flow, depth map or thermal image. To extract more distinctive features from multimodal data, we formulate the texture vector $T_{i}^{a} = [t_{i,1}^{a}, t_{i,2}^{a}, \\dots, t_{i,k^{2}}^{a}]$ as $t_{i,j}^{a} = \\exp(-\\alpha \\sum_{m} \\| \\mathbf{x}_{i}^{(m)} - \\mathbf{x}_{j}^{(m)} \\|^{2}), j \\in K_{i}$ , where $\\alpha$ is a hyperparameter and $\\mathbf{x}_{i}^{(m)}$", + "bbox": [ + 496, + 681, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7260", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/944cfc938bb9fe20d6c3cfdd5d8358203ecf4da8f25e5226c99b8a85994c9163.jpg", + "image_caption": [ + "Figure 5. Graphical illustration of the proposed $L_{btm}$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 455, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "means the appearance data of modal $m$ . It is noteworthy that a small difference between two pixels produces a small element in $T^s$ , but a large element in $T^a$ . Therefore, $T_i^s \\cdot (T_i^a)^T$ is defined as the matching penalty for considering pixel $\\mathbf{x}_i$ as saliency boundary, contrary to similarity. Finally, the formula of our $L_{btm}$ is:", + "bbox": [ + 75, + 287, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b t m} = \\frac {\\sum_ {i} b _ {i} T _ {i} ^ {s} \\cdot \\left(T _ {i} ^ {a}\\right) ^ {T}}{\\sum_ {i} b _ {i}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 383, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b_{i}$ is the binary boundary mask of saliency prediction.", + "bbox": [ + 76, + 426, + 467, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An intuitive example is shown in Fig. 5. For the predicted boundary pixels, we find some nearby pixels that differ significantly in saliency scores. In general, the appearances between boundary and these pixels are also different because they locate within salient objects and backgrounds, respectively. If not the case, we expect the saliency score of this boundary pixel to be close to those nearby pixels, so that the boundary will shift in the opposite direction. After an iterative training process, our method can align the predicted saliency boundaries with image edges.", + "bbox": [ + 75, + 441, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-scale Consistency", + "text_level": 1, + "bbox": [ + 76, + 601, + 295, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Salient objects are consistent in multi-scale inputs. Thus, we resize input images to a reference scale and encourage our method to produce consistent predictions by:", + "bbox": [ + 76, + 625, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m s} = \\sum_ {i} \\left| y _ {i} - r e s i z e \\left(\\hat {y} _ {i}\\right) \\right| ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 678, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{i}$ and $\\hat{y}_i$ are saliency predictions of different scales.", + "bbox": [ + 76, + 718, + 467, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training a Generalized Detector", + "text_level": 1, + "bbox": [ + 76, + 741, + 362, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous USOD methods [29,32,72], we generate pseudo labels $\\ddot{Y}$ based on saliency predictions $Y$ through $\\ddot{Y} = CRF(Y)$ and use them to train an extra saliency detector with the IOU loss. For RGB SOD, we use the same detector as the previous method [72]. For multimodal SOD tasks, we employ the MIDD [47] network as our detector. To ensure a fair comparison with existing SOD methods, we train the extra detector only using task-specific data and corresponding pseudo labels.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. All experiments are implemented on a single GTX 1080 Ti GPU. The batch size is 8 and input images are resized to $320^2$ . The reference scale is randomly selected from $(192^2, 256^2, 384^2, 448^2)$ . Only horizontal flipping is employed as our data augmentation strategy. We train our method for 20 epochs using the SGD optimizer with an initial learning rate of 0.1, which is decayed linearly. $\\alpha$ , $\\lambda_c$ , $\\lambda_b$ , and $\\lambda_m$ are set to 200, 1, 0.05 and 1, respectively. We train the extra detectors for 10 epochs using the SGD optimizer with a learning rate of 0.005.", + "bbox": [ + 496, + 114, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In our experiments, we follow the prevalent settings of different SOD tasks. Specifically, for RGB SOD, we use the training subsets of MSRA-B [31] or DUTS [52] to train our method, respectively. ECSSD [42], PASCALS [27], HKU-IS [25], DUTS-TE [52], DUT-O [61], and the testing subset of MSRA-B are employed for evaluation. For RGB-D SOD, we choose 2185 samples from NLPR [36] and NJUD [23] as the training set. RGBD135 [9], SIP [13] and the testing subsets of NJUD and NLPR are employed for evaluation. For RGB-T SOD, 2500 images in VT5000 [49] are for training, while VT1000 [50], VT821 [51] and the rest 2500 images in VT5000 are for testing. For video SOD, we choose the training splits of DAVIS [37] and DAVSOD [14] to train our method. Moreover, we randomly select 5 frames of each video in DAVSOD to avoid overfitting. The testing splits of DAVIS, DAVSOD, SegV2 [24] and FBMS [3] are employed for evaluation.", + "bbox": [ + 496, + 267, + 892, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We adopt three criteria for evaluation, including ave- $F_{\\beta}$ , Mean Absolute Error $(\\mathcal{M})$ and E-Measure $(E_{\\xi})$ [12]. Specifically, $F_{\\beta} = \\frac{(1 + \\beta^{2}) \\times \\text{Precision} \\times \\text{Recall}}{\\beta^{2} \\times \\text{Precision} + \\text{Recall}}$ , where $\\beta^{2}$ is set to 0.3 [1]. The ave- $F_{\\beta}$ is the $F_{\\beta}$ scores by setting the threshold as two times mean values. $\\mathcal{M}$ is the absolute error between predictions and ground truth. $E_{\\xi}$ measures the global statistics and local pixel matching information.", + "bbox": [ + 496, + 525, + 890, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results on RGB SOD", + "text_level": 1, + "bbox": [ + 498, + 648, + 700, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Tab. 1, we compare the proposed method with fully-supervised methods, MINet [34], LDF [56], and KRN [59], weakly-supervised methods, WSSA [65], MFNet [38], and SCW [62], and unsupervised methods, SBF [63], MNL [66], USPS [32], EDNS [64], A2S [72], DCFD [29], and SelfMask [43]. We list our results with different settings: (1) saliency results without extra detector and post-processing, denoted as \"Ours $_s1$ \"; (2) our full method as \"Ours\"; (3) training our method using multimodal data, while only DUTS-TR or MSRA-B datasets with pseudo labels are employed to train the extra detector (\"Ours $_{mm}$ \").", + "bbox": [ + 496, + 672, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Either training on MSRA-B or DUTS-TR, the proposed method achieves significant improvements compared to existing USOD methods. Moreover, our unsupervised method", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7261", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/57609085ff481549fdbc57373181b956526948d4c36d8c291049676eff579532.jpg", + "table_caption": [ + "Table 1. Experiment results on SOD benchmarks. \"Sup.\" indicates the supervised signals used to train SOD methods. \"F\", \"W\" and \"U\" mean fully-supervised, weakly-supervised and unsupervised, respectively. Best scores are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsYearSup.ECSSDMSRA-BDUT-OPASCAL-SDUTS-TEHKU-IS
\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta} \\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)
Trained on MSRA-B
SBF [63]2017U.812.878.087.867.929.058.611.771.106.711.795.131.627.785.105.805.895.074
MNL [66]2018U.874.906.069.881.932.053.683.821.076.792.846.091---.874.932.047
USPS [32]2019U.875.903.064.896.938.042.715.839.069.770.828.107.730.840.072.880.933.043
DCFD [29]2022U.880.900.064.903.938.041.731.838.064.773.830.105.744.832.068.887.926.044
A2S [72]2022U.888.911.064.902.941.041.719.841.069.790.838.106.750.860.065.887.937.042
Ours-U.902.923.056.912.948.036.731.851.065.803.848.099.767.871.061.891.939.041
Trained on DUTS-TR
MINet [34]2020F.924.953.033.903.948.038.756.873.055.842.899.064.828.917.037.908.961.028
LDF [56]2020F.930.951.034.902.944.037.773.881.052.853.903.062.855.929.034.914.960.028
KRN [59]2021F.931.951.032.911.950.036.793.893.050.851.894.068.865.934.033.920.961.027
WSSA [65]2020W.870.917.059.869.929.049.703.845.068.785.855.096.742.869.062.860.932.047
MFNet [38]2021W.844.889.084.872.923.059.621.784.098.756.824.115.693.832.079.839.919.058
SCW [62]2021W.900.931.049.898.940.040.758.862.060.827.879.080.823.890.049.896.943.038
EDNS [64]2020U.872.906.068.880.932.051.682.821.076.801.846.097.735.847.065.874.933.046
SelfMask [43]2022U.856.920.058.844.925.050.668.815.078.774.856.087.714.848.063.819.915.053
DCFD [29]2022U.888.915.059.888.930.045.710.837.070.795.860.090.764.855.064.889.935.042
Ourss1-U.847.912.057.849.925.054.622.773.111.763.840.093.676.814.082.819.914.053
Ours-U.916.938.044.904.944.039.745.863.061.830.882.074.810.901.047.902.947.037
Oursmm-U.913.942.043.899.945.039.737.855.064.831.890.073.803.899.048.898.946.037
", + "bbox": [ + 78, + 119, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77e557d8a7e4f54dd74aff4836cab19cf2837d1a98c239306df109f9670f068b.jpg", + "image_caption": [ + "Figure 6. Visual comparison between state-of-the-art SOD methods and ours." + ], + "image_footnote": [], + "bbox": [ + 78, + 470, + 890, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is competitive to recent weakly-supervised SCW [62] and fully-supervised methods KRN [59]. Furthermore, the results of “ $\\mathbf{Ours}_{s1}$ ” prove that our method extracts precise saliency knowledge from training samples, but somehow weaken when deploying on unseen images. In addition, the similar results of “ $\\mathbf{Ours}_{mm}$ ” and “ $\\mathbf{Ours}$ ” imply that extra multimodal data may not bring significant improvements to RGB SOD task. As proved in [53], the large-scale DUTS-TR dataset is saturated for training saliency detectors, even in the unsupervised case.", + "bbox": [ + 75, + 712, + 470, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A qualitative comparison is illustrated in Fig. 6. Overall, our saliency predictions are much more precise than other", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unsupervised methods. For example, in the first example, our method and two fully-supervised methods well segment target object with precise boundaries, while other methods fail to capture the tiny salient object.", + "bbox": [ + 496, + 712, + 890, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Multimodal SOD", + "text_level": 1, + "bbox": [ + 498, + 785, + 751, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct more experiments on multimodal tasks, including RGB-D, RGB-T and video SOD. Noted that we precompute the optical flow for each video frame as an extra modality. To verify the improvements brought by multimodal data, we build two sets for training: (1) Task-specific data (\"Ours\"). For example, for RGB-T SOD task, we only", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7262", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bffa937a8fae0bccead981a1e8d4cf159cd75fab408909995374457df3be89c4.jpg", + "table_caption": [ + "Table 2. Results on RGB-D SOD benchmarks. \"F\" and \"U\" mean fully-supervised and unsupervised, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsYearSup.RGBD135NJUDNLPRSIP
\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)
DSA2F [45]2021F.899.958.021.901.937.039.897.953.024---
SPNet [74]2021F.927.984.013---.903.959.021.893.931.043
CCFE [28]2022F.911.964.020.914.953.032.907.962.021.889.923.047
DSU [21]2022U.767.895.061.719.797.135.745.879.065.619.774.156
Ours-U.834.945.037.787.829.109.834.924.043.716.784.124
\\(Ours_{mm}\\)-U.877.946.029.862.908.060.852.931.034.873.925.051
", + "bbox": [ + 88, + 104, + 883, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1f67e1e26141457ff93c686f278d4ab370adc1ec197d1418800ae0d377d4dccc.jpg", + "table_caption": [ + "Table 3. Results on VSOD benchmarks. \"F\", \"W\" and \"U\" mean fully-supervised, weakly-supervised and unsupervised, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsYearSup.DAVSODDAVISSegV2FBMS
\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M} \\downarrow\\)
PCSA [17]2020F.556.749.077.794.922.022.789.931.019.783.868.042
TENet [40]2020F.595.773.067.821.941.017---.851.915.026
STVS [4]2021F.563.764.080.812.940.022.835.950.016.821.903.042
WSVSOD [70]2021W.492.710.103.731.900.036.711.909.031.736.840.084
Ours-U.534.747.084.751.913.042.751.914.033.732.794.100
\\(Ours_{mm}\\)-U.547.762.085.756.908.037.808.927.021.795.876.060
", + "bbox": [ + 81, + 250, + 888, + 367 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d6f45fdc655a127d50a089fcdfc0ff2658dd3d311562fe39a157f5d51e18a59.jpg", + "table_caption": [ + "Table 4. Results on RGB-T SOD benchmarks. \"F\" and \"U\" mean fully-supervised and unsupervised, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsYearSup.VT5000VT1000VT821
Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓
MIED [46]2020F.761.880.050.853.928.030.760.877.050
MIDD [47]2021F.801.899.043.882.942.027.805.898.045
APNet [75]2021F.821.918.035.885.951.021.818.912.034
CCFE [28]2022F.859.937.030.906.963.018.857.934.027
Ours-U.810.904.046.885.939.031.805.900.043
Oursmm-U.807.903.047.881.939.032.805.899.044
", + "bbox": [ + 104, + 395, + 859, + 511 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcd14aded3894013eba04c9f67dfa0275fb9b66b0f4ea6b6b496204a463c3c2b.jpg", + "table_caption": [ + "Table 5. Label quality on multimodal SOD datasets." + ], + "table_footnote": [], + "table_body": "
MethodsRGBRGB-DVSODRGB-T
Fβ↑M↓Fβ↑M↓Fβ↑M↓Fβ↑M↓
Ours.917.038.746.110.557.099.932.029
Oursrgb.917.040.804.068.613.089.920.030
Oursmm.918.040.826.062.637.079.923.029
", + "bbox": [ + 80, + 541, + 462, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "use 2500 RGB-thermal image pairs for training as prevalent RGB-T SOD methods [46,47,75]; (2) Multimodal data from the training sets of four SOD tasks (\"Ours $_{mm}$ \").", + "bbox": [ + 75, + 633, + 468, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our method reports state-of-the-art performance on all multimodal SOD tasks. Specifically, for the RGB-D SOD task in Tab. 2, our method surpasses the latest unsupervised method DSU [21]. For video and RGB-T SOD tasks in Tab. 3 and 4, to the best of our knowledge, our method is the first unsupervised method and is competitive to supervised methods. Furthermore, our method trained on multimodal data (\"Oursmm\") produce high-quality pseudo labels for all multimodal SOD datasets simultaneously and has achieved performance improvements on most tasks.", + "bbox": [ + 75, + 680, + 470, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 845, + 230, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Label quality comparison. In Tab. 5, we exhibit the scores of pseudo labels on multimodal SOD datasets. Ex", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cept for the above two training sets, we employ an additional set that collects all RGB images from four tasks to train our method, denoted as “ $\\mathbf{Ours}_{rgb}$ ”. Overall, “ $\\mathbf{Ours}_{mm}$ ” reports more generalized results on all datasets. Moreover, training on task-specific data (“Ours”) slightly surpasses the performance of “ $\\mathbf{Ours}_{mm}$ ” on the training sets of RGB and RGB-T tasks. Since our method is trained without ground truth, fitting a hybrid dataset may cause a slight performance drop on some subsets. In addition, the comparison between “ $\\mathbf{Ours}_{rgb}$ ” and “ $\\mathbf{Ours}_{mm}$ ” prove that multimodal data significantly improves the quality of the generated pseudo labels. For an intuitive comparison, we show some examples in Fig. 7. Trained with task-specific data, our method precisely localizes salient regions but fails to segment the whole objects. With more training data, the network perceives fine-grained concepts of target objects, resulting in more complete segmentations.", + "bbox": [ + 496, + 527, + 890, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of initial saliency cues. Our method is based on the activation maps of a pre-trained network [6, 18] instead of traditional methods [22,26,76] as other DL-based USOD methods [29,32,55,63,66]. We compare the performance of these two types of designs in Tab. 6. Overall, the activation map is worse before training, whereas better than traditional methods [22,26,76] after training. The activation maps are", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7263", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d682fdaa7ecd89d795a228c17a8adb0068f9c14ce1cc576ee58a8f9be5f3eb53.jpg", + "image_caption": [ + "Figure 7. Examples of the generated pseudo labels. The \"D/T/O\" means depth map, thermal image or optical flow." + ], + "image_footnote": [], + "bbox": [ + 76, + 90, + 470, + 236 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e2e6a6b80621a3dfad72362ac05eeb1461a6fcca6b43e1fe724495303b1ed899.jpg", + "table_caption": [ + "Table 6. Comparison between different saliency cues." + ], + "table_footnote": [], + "table_body": "
Saliency cuesBefore trainingAfter training
\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}\\downarrow\\)
MC [22].810.877.145.889.914.053
DSR [26].780.867.118.886.901.059
RBD [76].803.883.109.891.912.053
Ours.451.659.353.914.940.039
", + "bbox": [ + 78, + 308, + 467, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dynamic during training, such that the network strengthens the learned saliency knowledge incrementally. On the contrary, using traditional methods, the network learns from fixed labels throughout the training process, and thus fits the biased knowledge in those traditional methods. In addition, traditional methods introduce extra computation loads, while the pre-trained network is indispensable for all DL-based USOD methods.", + "bbox": [ + 75, + 425, + 467, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of loss functions. We conduct ablation studies on various combinations of loss functions in Tab. 7 and more details are introduced as follows.", + "bbox": [ + 76, + 547, + 468, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In ablation study A, our method receives continuous improvements by appending new losses ( $L_{ms}$ and $L_{btm}$ ) or replacing the original $L_{adb}$ to our $L_{csd}$ . This experiment proves that these three losses can assist the network mine more detailed saliency knowledge from different perspectives. More importantly, these losses are supplementary to each other, such that their combination can collaborate to produce high-quality pseudo labels for SOD datasets.", + "bbox": [ + 75, + 595, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In ablation study B, our $L_{csd}$ loss can provide precise saliency localization information and thus achieves the best performance among all competitors. Specifically, $L_{bce}$ and $L_{1}$ fail to mine detailed and reliable saliency knowledge from noisy saliency cues. $L_{adb}$ [72] improves the robustness of the learned saliency knowledge by focusing on easy samples, and thus surpasses the $L_{1}$ loss. $L_{spl}$ [2] also employs a dynamic schedule, but its binary weights filter out latent saliency knowledge in hard samples.", + "bbox": [ + 75, + 717, + 468, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In ablation study C, we compare our $L_{btm}$ with $L_{lsc}$ [62] and two variants: (1) using L1 distance for the texture features in appearance space, denoted as $L_{c1}$ ; (2) removing the", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f099e270fed56ec94845ecf1b0ab4fa63e40237951015c0c0dcfa8f87d65dac1.jpg", + "table_caption": [ + "Table 7. Ablation studies on loss functions." + ], + "table_footnote": [], + "table_body": "
TagLossFβ↑Eξ↑M↓
A1Ladb [72] (Baseline).882.915.071
A2Ladb [72] + Lms.891.921.066
A3Lcsd + Lms.908.937.050
A4Lcsd + Lbtm + Lms (Ours).917.945.038
B1Lbce + Lbtm + Lms.558.540.208
B2L1 + Lbtm + Lms.895.928.044
B3Ladb [72] + Lbtm + Lms.903.934.042
B4Lspl [2] + Lbtm + Lms.910.938.041
B5Lcsd + Lbtm + Lms (Ours).917.945.038
C1Lcsd + Lc1 + Lms.735.819.120
C2Lcsd + Llsc [62] + Lms.906.923.052
C3Lcsd + Lc2 + Lms.907.926.051
C4Lcsd + Lbtm + Lms (Ours).917.945.038
", + "bbox": [ + 501, + 104, + 880, + 321 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/167caabc64bf43724b073b64d8a7eae12805f26f2e68b28ef0a3f5b41733e96a.jpg", + "table_caption": [ + "Table 8. Label quality on DUTS-TR dataset." + ], + "table_footnote": [], + "table_body": "
Pre-trainingFβ↑Eξ↑M↓
Supervised.915.942.039
Unsupervised (Ours).917.945.038
", + "bbox": [ + 519, + 351, + 872, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "boundary masks in $L_{btm}$ , denoted as $L_{c2}$ . In general, our $L_{btm}$ outperforms other variants. Specifically, the L1 distance in $L_{c1}$ causes the texture features to be not distinctive enough, resulting in ambiguous saliency boundaries. $L_{lsc}$ focuses more on the adjacent pixels so that the boundaries are more susceptible to a limited region rather than a larger patch. Moreover, using $L_{c2}$ , edges within objects or backgrounds may corrupt the learned saliency knowledge.", + "bbox": [ + 496, + 422, + 890, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Supervised or unsupervised pre-training? Under the unsupervised setting, whether we can use a encoder with supervised pre-training is controversial. For the ImageNet dataset [10], the class labels of images usually indicate the category of the most salient object. It means that the supervised encoder receives extra saliency knowledge from these manual labels. Thus, for completely unsupervised SOD, we employ unsupervised MoCo-v2 to initialize our encoder. As listed in Tab. 8, the performance of our method when using supervised encoder is comparable.", + "bbox": [ + 496, + 545, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an Unsupervised Salient Object Detection (USOD) method guided by two novel mechanisms. First, we propose a Confidence-aware Saliency Distilling (CSD) to learn saliency knowledge from easy samples to hard ones with noisy labels progressively. Second, we propose a Boundary-aware Texture Matching (BTM) to make the location of saliency prediction more accurate. As a result, the proposed method produces high-quality pseudo labels to train saliency detectors. Experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method outperforms existing USOD methods.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7264", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5", + "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 4, 8", + "[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295. Springer, 2010. 5", + "[4] Chenglizhao Chen, Guotao Wang, Chong Peng, Yuming Fang, Dingwen Zhang, and Hong Qin. Exploring rich and efficient spatial temporal interactions for real-time video salient object detection. IEEE Transactions on Image Processing, 30:3995-4007, 2021. 2, 7", + "[5] Peijia Chen, Jianhuang Lai, Guangcong Wang, and Huajun Zhou. Confidence-guided adaptive gate and dual differential enhancement for video salient object detection. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2", + "[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3, 7", + "[7] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE Transactions on Image Processing, 30:431-443, 2020. 2", + "[8] Xiaolong Cheng, Xuan Zheng, Jialun Pei, He Tang, Zehua Lyu, and Chuanbo Chen. Depth-induced gap-reducing network for rgb-d salient object detection: An interaction, guidance and refinement approach. IEEE Transactions on Multimedia, 2022. 2", + "[9] Yupeng Cheng, Huazhu Fu, Xingxing Wei, Jiangjian Xiao, and Xiaochun Cao. Depth enhanced saliency detection method. In Proceedings of international conference on internet multimedia computing and service, pages 23-27, 2014. 5", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 8", + "[11] Wenhui Dao, Xian Sun, Xinwei Zheng, Fangzheng Dou, Hongqi Wang, and Kun Fu. Efficient saliency-based object detection in remote sensing images using deep belief networks. IEEE Geoscience and Remote Sensing Letters, 13(2):137-141, 2016. 1", + "[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In International Joint Conference on Artificial Intelligence (IJCAI), pages 698-704, 2018. http://dpfan.net/e-measure/.5", + "[13] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on neural networks and learning systems, 32(5):2075-2089, 2020. 5", + "[14] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8554-8564, 2019. 2, 5", + "[15] Mengyang Feng, Hutchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1623-1632, 2019. 2", + "[16] Carola Figueroa Flores, Abel Gonzalez-Garcia, Joost van de Weijer, and Bogdan Raducanu. Saliency for fine-grained object recognition in domains with scarce training data. Pattern Recognition, 94:62-73, 2019. 1", + "[17] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 10869–10876, 2020. 2, 7", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7", + "[19] Qibin Hou, Ming-Ming Cheng, Xiaowei Hu, Ali Borji, Zhuowen Tu, and Philip HS Torr. Deeply supervised salient object detection with short connections. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3203-3212, 2017. 2", + "[20] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3", + "[21] Wei Ji, Jingjing Li, Qi Bi, Chuan Guo, Jie Liu, and Li Cheng. Promoting saliency from depth: Deep unsupervised rgb-d saliency detection. arXiv preprint arXiv:2205.07179, 2022. 1, 2, 7", + "[22] Bowen Jiang, Lihe Zhang, Huchuan Lu, Chuan Yang, and Ming-Hsuan Yang. Saliency detection via absorbing markov chain. In Proceedings of the IEEE international conference on computer vision, pages 1665-1672, 2013. 2, 3, 7, 8", + "[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In 2014 IEEE international conference on image processing (ICIP), pages 1115-1119. IEEE, 2014. 5", + "[24] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5", + "[25] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, 2015. 5", + "[26] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 2976-2983, 2013. 2, 3, 7, 8" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7265", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, 2014. 5", + "[28] Guibiao Liao, Wei Gao, Ge Li, Junle Wang, and Sam Kwong. Cross-collaborative fusion-encoder network for robust rgb-thermal salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7", + "[29] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In Thirty-sixth AAAI conference on artificial intelligence, 2022. 1, 2, 5, 6, 7", + "[30] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3089-3098, 2018. 2", + "[31] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence, 33(2):353-367, 2010. 5", + "[32] Duc Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction with self-supervision. arXiv preprint arXiv:1909.13055, 2019. 1, 2, 3, 5, 6, 7", + "[33] Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine intelligence, 24(7):971-987, 2002. 4", + "[34] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9413-9422, 2020. 5, 6", + "[35] Prashant W Patil, Subrahmanyam Murala, Abhinav Dhall, and Sachin Chaudhary. Msednet: multi-scale deep saliency learning for moving object detection. In 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 1670-1675. IEEE, 2018. 1", + "[36] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In European conference on computer vision, pages 92-109. Springer, 2014. 5", + "[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5", + "[38] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 5, 6", + "[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2", + "[40] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In European Conference on Computer Vision, pages 212-228. Springer, 2020. 2, 7", + "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2", + "[42] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4):717-729, 2015. 5", + "[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6", + "[44] Ali Shokoufandeh, Ivan Marsic, and Sven J Dickinson. View-based object recognition using saliency maps. Image and Vision Computing, 17(5-6):445-460, 1999. 1", + "[45] Peng Sun, Wenhu Zhang, Huanyu Wang, Songyuan Li, and Xi Li. Deep rgb-d saliency detection with depth-sensitive attention and automatic multi-modal fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1407-1417, 2021. 2, 7", + "[46] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive encoder-decoder network for rgbt salient object detection. arXiv e-prints, pages arXiv-2005, 2020. 2, 7", + "[47] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE Transactions on Image Processing, 30:5678–5691, 2021. 2, 5, 7", + "[48] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 2", + "[49] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 5", + "[50] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE Transactions on Multimedia, 22(1):160-173, 2019. 5", + "[51] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. RGB-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In Chinese Conference on Image and Graphics Technologies, pages 359-369. Springer, 2018. 5", + "[52] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7266", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. arXiv preprint arXiv:1904.09146, 2019. 6", + "[54] Wenguan Wang, Shuyang Zhao, Jianbing Shen, Steven CH Hoi, and Ali Borji. Salient object detection with pyramid attention and salient edges. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1448-1457, 2019. 2", + "[55] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11727-11736, 2022. 1, 2, 7", + "[56] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13025-13034, 2020. 5, 6", + "[57] Hongfa Wen, Chenggang Yan, Xiaofei Zhou, Runmin Cong, Yaoqi Sun, Bolun Zheng, Jiyong Zhang, Yongjun Bao, and Guiguang Ding. Dynamic selective network for rgb-d salient object detection. IEEE Transactions on Image Processing, 30:9179-9192, 2021. 2", + "[58] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3907-3916, 2019. 2", + "[59] Binwei Xu, Haoran Liang, Ronghua Liang, and Peng Chen. Locate globally, segment locally: A progressive architecture with knowledge review network for salient object detection. In Proceedings. of the AAAI Conference On Artificial Intelligence, pages 1-9, 2021. 5, 6", + "[60] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 2", + "[61] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 5", + "[62] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). AAAI Palo Alto, CA, USA, 2021. 2, 5, 6, 8", + "[63] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2, 3, 5, 6, 7", + "[64] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European conference on computer vision, pages 349-366. Springer, 2020. 2, 5, 6", + "[65] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scribble annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12546-12555, 2020. 2, 5, 6", + "[66] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9029-9038, 2018. 1, 2, 5, 6, 7", + "[67] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 202-211, 2017. 2", + "[68] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In ICCV, 2017. 2", + "[69] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 714-722, 2018. 2", + "[70] Wangbo Zhao, Jing Zhang, Long Li, Nick Barnes, Nian Liu, and Junwei Han. Weakly supervised video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16826-16835, 2021. 2, 7", + "[71] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2", + "[72] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 1, 2, 3, 4, 5, 6, 8", + "[73] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9141-9150, 2020. 2", + "[74] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving rgb-d saliency detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4681-4691, 2021. 2, 7", + "[75] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Apnet: Adversarial learning assistance and perceived importance fusion network for all-day rgb-t salient object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 7", + "[76] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 2, 3, 7, 8", + "[77] Wenbin Zou and Nikos Komodakis. Harf: Hierarchy-associated rich features for salient object detection. In Proceedings of the IEEE international conference on computer vision, pages 406-414, 2015. 3" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7267", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_model.json b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3933ec39f3e8bad0a66f592af544e16d16d977a3 --- /dev/null +++ b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.871, + 0.152 + ], + "angle": 0, + "content": "Texture-guided Saliency Distilling for Unsupervised Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.18, + 0.803, + 0.199 + ], + "angle": 0, + "content": "Huajun Zhou\\(^{1}\\), Bo Qiao\\(^{1}\\), Lingxiao Yang\\(^{1}\\), Jianhuang Lai\\(^{1,2,3}\\), Xiaohua Xie\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.199, + 0.794, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.217, + 0.809, + 0.234 + ], + "angle": 0, + "content": "2Guangdong Province Key Laboratory of Information Security Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.234, + 0.873, + 0.252 + ], + "angle": 0, + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.313, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.606 + ], + "angle": 0, + "content": "Deep Learning-based Unsupervised Salient Object Detection (USOD) mainly relies on the noisy saliency pseudo labels that have been generated from traditional handcraft methods or pre-trained networks. To cope with the noisy labels problem, a class of methods focus on only easy samples with reliable labels but ignore valuable knowledge in hard samples. In this paper, we propose a novel USOD method to mine rich and accurate saliency knowledge from both easy and hard samples. First, we propose a Confidence-aware Saliency Distilling (CSD) strategy that scores samples conditioned on samples' confidences, which guides the model to distill saliency knowledge from easy samples to hard samples progressively. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the boundaries of noisy labels by matching the textures around the predicted boundaries. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance. Code is available at www.github.com/moothes/A2S-v2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.21, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Unsupervised Salient Object Detection (USOD) methods aim to correctly localize and precisely segment salient objects simultaneously without using manual annotations. Compared to the supervised methods, USOD methods can easily adapt to more practical scenarios (e.g., industrial or medical images) where a large number of labeled images may be very hard to collect. Moreover, USOD methods also can assist some related methods for other tasks, e.g., object recognition [16, 44] and object detection [11, 35]. However, diverse objects, complex backgrounds, and other challenges" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.285, + 0.601, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.345, + 0.578, + 0.356 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.286, + 0.697, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.345, + 0.666, + 0.356 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.286, + 0.793, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.345, + 0.767, + 0.356 + ], + "angle": 0, + "content": "(c) RBD" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.286, + 0.889, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.345, + 0.858, + 0.356 + ], + "angle": 0, + "content": "(d) HS" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.357, + 0.601, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.416, + 0.577, + 0.428 + ], + "angle": 0, + "content": "(e) USPS" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.357, + 0.697, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.416, + 0.669, + 0.428 + ], + "angle": 0, + "content": "(f) A2S" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.357, + 0.794, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.416, + 0.782, + 0.428 + ], + "angle": 0, + "content": "(g) Activation" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.357, + 0.889, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.416, + 0.862, + 0.428 + ], + "angle": 0, + "content": "(h) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.44, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Figure 1. Saliency maps generated by USOD methods. Our result is generated based on the activation map of a deep network." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.488, + 0.882, + 0.503 + ], + "angle": 0, + "content": "ing conditions bring severe challenges to USOD methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.883 + ], + "angle": 0, + "content": "Most Deep Learning-based (DL-based) methods [21, 29, 32, 55, 63, 66] base on the saliency cues extracted by traditional SOD methods (Fig. 1-c and 1-d). These handcrafted features related cues are employed as pseudo labels to train deep networks under certain constraints, e.g., binary cross-entropy (BCE) loss. However, saliency cues by traditional methods usually shift away from target objects, especially in complex scenes. Moreover, conventional constraints, such as BCE loss, works well on fully-supervised SOD methods, but is suboptimal when fitting the noisy labels for unsupervised methods (Fig. 1-e). Recently, Zhou et al. [72] addressed the first issue by extracting saliency cues (Fig. 1-f) based on a unsupervisedly pre-trained network (Fig. 1-g) instead of using traditional methods. During training, they focus on learning reliable saliency knowledge from easy samples, but ignore latent knowledge in hard samples. The main reason is that hard samples may be wrongly-labeled and corrupt the fragile saliency knowledge learned in the early training phase. Therefore, to leverage hard samples, we argue that all samples should be employed in a meaningful order (i.e., from high reliable to low reliable), which is crucial for mining accurate knowledge from noisy labels. Trained by such a strategy, the network can mine valuable knowledge from hard examples without corrupting the knowledge learned from easy samples." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Deep networks can learn to localize salient regions from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. This project is supported by the Key-Area Research and Development Program of Guangdong Province (2019B010155003), the National Natural Science Foundation of China (U22A2095, 62072482, 62076258, 62206316), and the Guangdong NSF Project (2022A1515011254)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "noisy labels [72], but still struggle to find the precise boundaries of target objects. Generally, the appearance around saliency boundary has a similar texture as in saliency map. Therefore, matching the textures between different maps can serve as a guidance for producing reasonable saliency boundaries. We will demonstrate that above strategies are applicable to multimodal data besides RGB image, including depth map, thermal image, and optical flow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.473, + 0.529 + ], + "angle": 0, + "content": "Based on the above analysis, we propose a novel framework to tackle the Unsupervised Salient Object Detection (USOD) tasks. Specifically, two strategies are proposed to mine saliency knowledge from noisy saliency labels. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme that scores samples with noisy labels conditioned on samples' confidence. Then, our CSD guides the network to learn saliency knowledge from easy samples to more complex ones progressively by employing an adaptive loss conditioned on the training progress. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the saliency boundaries of noisy labels by matching the textures around the predicted boundaries. During training, the predicted saliency boundaries are shifting toward surrounding edges in the appearance space of the whole image. Finally, guided by above two mechanisms, our method can produce high-quality pseudo labels to train generalized saliency detectors. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art performance compared to existing USOD methods." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.531, + 0.468, + 0.545 + ], + "angle": 0, + "content": "The main contributions of our novel USOD method are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.557, + 0.47, + 0.618 + ], + "angle": 0, + "content": "1. We propose a Confidence-aware Saliency Distilling (CSD) to mine rich and accurate saliency knowledge from noisy labels, which breaks through the limitation that existing methods cannot utilize hard samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.469, + 0.675 + ], + "angle": 0, + "content": "2. We propose a Boundary-aware Texture Matching (BTM) to refine the boundary of the predicted saliency maps by matching textures in different spaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.469, + 0.732 + ], + "angle": 0, + "content": "3. Extensive experiments on RGB, RGB-D, RGB-T and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.557, + 0.47, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.746, + 0.227, + 0.761 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.389, + 0.788 + ], + "angle": 0, + "content": "2.1. Supervised Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Researchers have developed a large family of fully-supervised Salient Object Detection (SOD) algorithms [5,7, 15,19,30,39,54,58,67-69,71,73] in the past decades. Ronneberger et al. [41] proposed a U-shape structure that progressively upsamples and concatenates the smaller features to the larger ones. To ease the annotation burden, Zhang et al. [65] relabeled the DUTS-TR dataset [52] with scribbles" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "and leveraged an edge detection model for boundary localization. Yu et al. [62] proposed a local coherence loss to find precise boundary based on scribbles annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.894, + 0.272 + ], + "angle": 0, + "content": "Multimodal SOD tasks aim at using other modality data to improve the SOD performance, such as depth map, thermal image and optical flow. Recently, abundant methods [4,8,14,17,40,45-48,57,74] were proposed. Most of these methods employ a two-stream encoder-decoder structure to aggregate multi-level information in multimodal data. To reduce the annotating cost, Zhao et al. [70] proposed a video SOD dataset with scribbles annotations to indicate the location of salient objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.894, + 0.318 + ], + "angle": 0, + "content": "Although the above methods have achieved significant performance, they require numerous human annotations for training, which are expensive to collect." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.327, + 0.833, + 0.343 + ], + "angle": 0, + "content": "2.2. Unsupervised Salient Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Traditional SOD methods [22, 26, 60, 76] extracted saliency cues from images by modeling the correlation of hand-crafted features. Inspired by the centralization prior, Jiang et al. [22] considered the distances between boundary superpixels and non-boundary superpixels as saliency scores. Yan et al. [60] employed a tree-structure graphical model to compute the saliency results based on several over-segmented maps. The above methods fail to accurately localize salient objects because the global information in hand-crafted features are not representative enough." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.742 + ], + "angle": 0, + "content": "Existing DL-based USOD methods can be divided into two pipelines based on the method used to extract saliency cues from images. First, most USOD methods [29, 32, 55, 63, 64, 66] focused on refining the coarse saliency cues extracted by several traditional SOD methods. For example, Zhang et al. [63] weighted these saliency cues by combining both intra-image and inter-image fusion streams. Zhang et al. [66] designed a noise modeling module to deal with noises in these saliency cues. Nguyen et al. [32] used these saliency cues as labels to train multiple deep networks. Ji et al. [21] refined the saliency maps extracted by traditional SOD methods to produce more accurate saliency predictions. Second, to prevent the localization errors caused by traditional SOD methods, Zhou et al. [72] proposed a novel framework that converts the activation maps of a pre-trained network to high-quality pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.788 + ], + "angle": 0, + "content": "In this paper, we propose two novel mechanisms to tackle USOD tasks and explore the potential of improving the quality of pseudo labels by using multimodal data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.649, + 0.817 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.871 + ], + "angle": 0, + "content": "In our method, we propose two novel strategies to mine accurate saliency knowledge based on the noisy activation maps generated by a deep network \\(\\Phi\\), as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Activation map generation. Following previous work [72], we generate an activation map \\( Y \\) for input image \\( X \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7258" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.894, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed unsupervised framework. The network are guided by two mechanisms: (1) Conidence-aware Saliency Distilling mines rich and accurate saliency knowledge from noisy labels, and (2) Boundary-aware Texture Matching makes the boundary of the predicted saliency map more accurate. We pad some constant maps for the missing modalities." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.398, + 0.306, + 0.414 + ], + "angle": 0, + "content": "using a deep network \\(\\Phi(X) = Y\\):" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.426, + 0.312, + 0.442 + ], + "angle": 0, + "content": "\\[\nE _ {3}, E _ {4}, E _ {5} = \\operatorname {E n c o d e r} (X),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.445, + 0.469, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {i} = S E \\left(E _ {i}\\right), i \\in \\{3, 4, 5 \\}, \\\\ H = S E \\left(\\text {一} (E _ {i}, E _ {i}, E _ {i})\\right) \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.464, + 0.386, + 0.48 + ], + "angle": 0, + "content": "\\[\nH = S E (\\text {c o n c a t} (F _ {3}, F _ {4}, F _ {5})),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.482, + 0.427, + 0.5 + ], + "angle": 0, + "content": "\\[\nY = \\operatorname {i n v} (\\operatorname {S i g m o i d} (\\operatorname {s u m} (H - \\bar {H}))),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.693 + ], + "angle": 0, + "content": "where \\(\\bar{H}\\) is the spatial mean of \\(H\\). Specifically, we employ the ResNet-50 [18] pre-trained by MoCo-v2 [6] as our encoder, which is trained without extra manual annotations. In our decoder, four SE blocks [20] integrate multiple encoder features into \\(H\\). After that, we set \\(H - \\bar{H}\\) to ensure that (1) noisy labels are adaptive to input images when using a fixed threshold and (2) coexisting of positive and negative samples. Next, features are summed over the channel dimension to produce one-channel activation map. Finally, the Sigmoid function produces the final saliency scores, and the inv function identifies the regions with more corner pixels as background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.47, + 0.77 + ], + "angle": 0, + "content": "Training strategy. The activation maps produced by network \\(\\Phi\\) perceive some discriminative regions in input images [72], but still are low-quality because of widespread noises. To improve the quality, we employ three strategies to train the network \\(\\Phi\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.78, + 0.469, + 0.799 + ], + "angle": 0, + "content": "\\[\nL _ {\\Phi} = \\lambda_ {c} \\left(L _ {c s d} + \\hat {L} _ {c s d}\\right) + \\lambda_ {b} \\left(L _ {b t m} + \\hat {L} _ {b t m}\\right) + \\lambda_ {m} L _ {m s}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where all \\(\\lambda\\) are hyperparameters. \\(\\hat{L}\\) is the loss for the predictions of a different scale. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme to excavate valuable saliency knowledge from simple examples to more complex ones. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to align the boundaries in" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.445 + ], + "angle": 0, + "content": "appearance and the predicted saliency maps. In addition, a multi-scale consistency loss \\( L_{ms} \\) ensures our method to produce consistent predictions for multi-scale inputs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.454, + 0.818, + 0.471 + ], + "angle": 0, + "content": "3.1. Confidence-aware Saliency Distilling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.674 + ], + "angle": 0, + "content": "The primary challenge for USOD methods is how to localize salient regions. Existing DL-based USOD methods [32,63,72] are based on noisy saliency cues from traditional SOD methods [22,26,76,77] or pre-trained networks [6,18]. Following [72], we employ the activation maps produced by a pre-trained network as our saliency cues instead of the saliency predictions of traditional SOD methods. The main reasons are two-fold: (1) traditional methods require extra computation loads; (2) network can focus on mining saliency knowledge instead of fitting the inductive bias of traditional methods. We binarize the saliency cues as the initial labels using a fixed threshold of 0.5, while Eqn. 1 ensures that these labels remain adaptive to the input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.9 + ], + "angle": 0, + "content": "To generate high-quality pseudo labels, USPS [32] uses these initial labels to train deep networks like other fully-supervised methods. However, the simple training strategy in fully-supervised methods is suboptimal for unsupervised methods. Specifically, we define pixels with saliency scores close to 0.5 as hard examples because of their low confidences. In the noisy labels, hard examples are likely to be wrongly labeled. Therefore, it is difficult to learn robust saliency knowledge from these hard samples using traditional loss functions (e.g., BCE loss). To this end, A2S [72] learns reliable saliency knowledge from easy samples. However, for hard examples, the saliency knowledge hidden in noises is not fully explored. In summary, a customized strategy that organizes the samples in a more meaningful order is crucial for USOD methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7259" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.48, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.184, + 0.302, + 0.195 + ], + "angle": 0, + "content": "(a) \\(\\mathrm{L}_1\\)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.09, + 0.87, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.184, + 0.697, + 0.195 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{L}_{\\mathrm{bce}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.197, + 0.482, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.295, + 0.308, + 0.305 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{L_{adb}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.197, + 0.87, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.294, + 0.714, + 0.305 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{L}_{\\mathrm{csd}}\\) (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.309, + 0.666, + 0.323 + ], + "angle": 0, + "content": "Figure 3. The gradient landscapes of different loss functions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.474, + 0.522 + ], + "angle": 0, + "content": "To address the above problem, we propose a Confidence-aware Saliency Distillation (CSD) scheme that scores samples with noisy labels conditional on their confidence and training progress. Concretely, easy samples contain reliable knowledge, and thus can assist our method to learn reliable saliency knowledge. On the contrary, the saliency knowledge in hard samples is hidden within noises, and may corrupt the fragile saliency patterns in the early stages of network training. Inspired by self-paced learning [2], we dynamically adjust the gradients for samples by introducing a factor \\(\\rho\\), which is linearly increasing from 0 to 1 as training proceeding. Our \\(L_{csd}\\) loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.53, + 0.47, + 0.571 + ], + "angle": 0, + "content": "\\[\nL _ {c s d} = - \\frac {1}{N} \\sum_ {i} ^ {N} | \\Phi (\\mathbf {x} _ {i}) - 0. 5 | ^ {(2 ^ {1 - \\rho})}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.687 + ], + "angle": 0, + "content": "where \\(N\\) is the number of pixels, and \\(\\Phi (\\mathbf{x}_i)\\) is the saliency prediction of pixel \\(\\mathbf{x}_i\\). Notice that noisy label \\(\\mathbf{G}\\) is omitted in Eqn. 3 because it is generated based on \\(\\Phi (\\mathbf{x}_i)\\). Thus, we can calculate the confidence score \\(|k_{i}|\\) of pixel \\(\\mathbf{x}_i\\) in a simpler way as: \\(|k_{i}| = |\\Phi (\\mathbf{x}_{i}) - 0.5|\\), where the constant value 0.5 is related to \\(\\bar{H}\\) in Eqn. 1. The partial derivative of our \\(L_{csd}\\) over \\(\\Phi (\\mathbf{x}_i)\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.695, + 0.47, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L _ {c s d}}{\\partial \\Phi (\\mathbf {x} _ {i})} = - \\operatorname {s i g n} \\left(k _ {i}\\right) 2 ^ {(1 - \\rho)} \\left| k _ {i} \\right| ^ {2 ^ {(1 - \\rho)} - 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.765 + ], + "angle": 0, + "content": "where \\( \\text{sign}(k_i) \\in \\{-1, 1\\} \\) for negative and positive values, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "For an intuitive comparison, we draw the gradient landscapes of different losses in Fig. 3, including \\( L_{1} \\), \\( L_{bce} \\), \\( L_{adb} \\) [72] and our \\( L_{csd} \\). The gradient of other losses are consistent throughout the training process, while our \\( L_{csd} \\) varies with the training process. Specifically, in the beginning, our \\( L_{csd} \\) assigns low gradients to hard samples to learn reliable saliency knowledge from easy samples. As training proceeding, the gradients of hard samples are increasing to mine more valuable saliency knowledge." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.338, + 0.895, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.48, + 0.858, + 0.494 + ], + "angle": 0, + "content": "Figure 4. Appearance information in multimodal data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.809, + 0.535 + ], + "angle": 0, + "content": "3.2. Boundary-aware Texture Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.679 + ], + "angle": 0, + "content": "In general, the appearance of saliency boundary has a similar texture as in saliency prediction. Therefore, matching the textures between different maps can guide our method to produce reasonable saliency scores. This strategy is also applicable to other modalities, such as depth map, thermal image, and optical flow, as shown in Fig. 4. By aggregating the rich appearance information in multimodal data, we can provide a more generalized guidance for our USOD method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.893, + 0.904 + ], + "angle": 0, + "content": "Based on the above analysis, we propose a Boundary-aware Texture Matching (BTM) strategy to align the saliency boundaries and image edges by matching their textures. First of all, similar to the LBP feature [33], we extract texture vectors from saliency predictions and appearance, respectively. For saliency predictions, the texture vector \\( T_{i}^{s} = [t_{i,1}^{s}, t_{i,2}^{s}, \\dots, t_{i,k^{2}}^{s}] \\) for the \\( i \\)-th pixel is computed by \\( t_{i,j}^{s} = |\\Phi(\\mathbf{x}_{i}) - \\Phi(\\mathbf{x}_{j})|, j \\in K_{i} \\), where \\( K_{i} \\) is \\( k \\times k \\) neighborhoods around pixel \\( \\mathbf{x}_{i} \\). For appearance information, it may be one of RGB image, optical flow, depth map or thermal image. To extract more distinctive features from multimodal data, we formulate the texture vector \\( T_{i}^{a} = [t_{i,1}^{a}, t_{i,2}^{a}, \\dots, t_{i,k^{2}}^{a}] \\) as \\( t_{i,j}^{a} = \\exp(-\\alpha \\sum_{m} \\| \\mathbf{x}_{i}^{(m)} - \\mathbf{x}_{j}^{(m)} \\|^{2}), j \\in K_{i} \\), where \\( \\alpha \\) is a hyperparameter and \\( \\mathbf{x}_{i}^{(m)} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7260" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.457, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.263, + 0.435, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Graphical illustration of the proposed \\(L_{btm}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.379 + ], + "angle": 0, + "content": "means the appearance data of modal \\( m \\). It is noteworthy that a small difference between two pixels produces a small element in \\( T^s \\), but a large element in \\( T^a \\). Therefore, \\( T_i^s \\cdot (T_i^a)^T \\) is defined as the matching penalty for considering pixel \\( \\mathbf{x}_i \\) as saliency boundary, contrary to similarity. Finally, the formula of our \\( L_{btm} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.385, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\nL _ {b t m} = \\frac {\\sum_ {i} b _ {i} T _ {i} ^ {s} \\cdot \\left(T _ {i} ^ {a}\\right) ^ {T}}{\\sum_ {i} b _ {i}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.468, + 0.442 + ], + "angle": 0, + "content": "where \\( b_{i} \\) is the binary boundary mask of saliency prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.469, + 0.593 + ], + "angle": 0, + "content": "An intuitive example is shown in Fig. 5. For the predicted boundary pixels, we find some nearby pixels that differ significantly in saliency scores. In general, the appearances between boundary and these pixels are also different because they locate within salient objects and backgrounds, respectively. If not the case, we expect the saliency score of this boundary pixel to be close to those nearby pixels, so that the boundary will shift in the opposite direction. After an iterative training process, our method can align the predicted saliency boundaries with image edges." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.296, + 0.618 + ], + "angle": 0, + "content": "3.3. Multi-scale Consistency" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.626, + 0.469, + 0.671 + ], + "angle": 0, + "content": "Salient objects are consistent in multi-scale inputs. Thus, we resize input images to a reference scale and encourage our method to produce consistent predictions by:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.679, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\nL _ {m s} = \\sum_ {i} \\left| y _ {i} - r e s i z e \\left(\\hat {y} _ {i}\\right) \\right| ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.719, + 0.468, + 0.734 + ], + "angle": 0, + "content": "where \\(y_{i}\\) and \\(\\hat{y}_i\\) are saliency predictions of different scales." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.363, + 0.758 + ], + "angle": 0, + "content": "3.4. Training a Generalized Detector" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Following previous USOD methods [29,32,72], we generate pseudo labels \\(\\ddot{Y}\\) based on saliency predictions \\(Y\\) through \\(\\ddot{Y} = CRF(Y)\\) and use them to train an extra saliency detector with the IOU loss. For RGB SOD, we use the same detector as the previous method [72]. For multimodal SOD tasks, we employ the MIDD [47] network as our detector. To ensure a fair comparison with existing SOD methods, we train the extra detector only using task-specific data and corresponding pseudo labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Implementation details. All experiments are implemented on a single GTX 1080 Ti GPU. The batch size is 8 and input images are resized to \\(320^2\\). The reference scale is randomly selected from \\((192^2, 256^2, 384^2, 448^2)\\). Only horizontal flipping is employed as our data augmentation strategy. We train our method for 20 epochs using the SGD optimizer with an initial learning rate of 0.1, which is decayed linearly. \\(\\alpha\\), \\(\\lambda_c\\), \\(\\lambda_b\\), and \\(\\lambda_m\\) are set to 200, 1, 0.05 and 1, respectively. We train the extra detectors for 10 epochs using the SGD optimizer with a learning rate of 0.005." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.893, + 0.525 + ], + "angle": 0, + "content": "Datasets. In our experiments, we follow the prevalent settings of different SOD tasks. Specifically, for RGB SOD, we use the training subsets of MSRA-B [31] or DUTS [52] to train our method, respectively. ECSSD [42], PASCALS [27], HKU-IS [25], DUTS-TE [52], DUT-O [61], and the testing subset of MSRA-B are employed for evaluation. For RGB-D SOD, we choose 2185 samples from NLPR [36] and NJUD [23] as the training set. RGBD135 [9], SIP [13] and the testing subsets of NJUD and NLPR are employed for evaluation. For RGB-T SOD, 2500 images in VT5000 [49] are for training, while VT1000 [50], VT821 [51] and the rest 2500 images in VT5000 are for testing. For video SOD, we choose the training splits of DAVIS [37] and DAVSOD [14] to train our method. Moreover, we randomly select 5 frames of each video in DAVSOD to avoid overfitting. The testing splits of DAVIS, DAVSOD, SegV2 [24] and FBMS [3] are employed for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Metrics. We adopt three criteria for evaluation, including ave-\\(F_{\\beta}\\), Mean Absolute Error \\((\\mathcal{M})\\) and E-Measure \\((E_{\\xi})\\) [12]. Specifically, \\(F_{\\beta} = \\frac{(1 + \\beta^{2}) \\times \\text{Precision} \\times \\text{Recall}}{\\beta^{2} \\times \\text{Precision} + \\text{Recall}}\\), where \\(\\beta^{2}\\) is set to 0.3 [1]. The ave-\\(F_{\\beta}\\) is the \\(F_{\\beta}\\) scores by setting the threshold as two times mean values. \\(\\mathcal{M}\\) is the absolute error between predictions and ground truth. \\(E_{\\xi}\\) measures the global statistics and local pixel matching information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.65, + 0.702, + 0.665 + ], + "angle": 0, + "content": "4.1. Results on RGB SOD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.854 + ], + "angle": 0, + "content": "As shown in Tab. 1, we compare the proposed method with fully-supervised methods, MINet [34], LDF [56], and KRN [59], weakly-supervised methods, WSSA [65], MFNet [38], and SCW [62], and unsupervised methods, SBF [63], MNL [66], USPS [32], EDNS [64], A2S [72], DCFD [29], and SelfMask [43]. We list our results with different settings: (1) saliency results without extra detector and post-processing, denoted as \"Ours\\(_s1\\)\"; (2) our full method as \"Ours\"; (3) training our method using multimodal data, while only DUTS-TR or MSRA-B datasets with pseudo labels are employed to train the extra detector (\"Ours\\(_{mm}\\)\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Either training on MSRA-B or DUTS-TR, the proposed method achieves significant improvements compared to existing USOD methods. Moreover, our unsupervised method" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7261" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Experiment results on SOD benchmarks. \"Sup.\" indicates the supervised signals used to train SOD methods. \"F\", \"W\" and \"U\" mean fully-supervised, weakly-supervised and unsupervised, respectively. Best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.12, + 0.892, + 0.461 + ], + "angle": 0, + "content": "
MethodsYearSup.ECSSDMSRA-BDUT-OPASCAL-SDUTS-TEHKU-IS
\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta} \\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)
Trained on MSRA-B
SBF [63]2017U.812.878.087.867.929.058.611.771.106.711.795.131.627.785.105.805.895.074
MNL [66]2018U.874.906.069.881.932.053.683.821.076.792.846.091---.874.932.047
USPS [32]2019U.875.903.064.896.938.042.715.839.069.770.828.107.730.840.072.880.933.043
DCFD [29]2022U.880.900.064.903.938.041.731.838.064.773.830.105.744.832.068.887.926.044
A2S [72]2022U.888.911.064.902.941.041.719.841.069.790.838.106.750.860.065.887.937.042
Ours-U.902.923.056.912.948.036.731.851.065.803.848.099.767.871.061.891.939.041
Trained on DUTS-TR
MINet [34]2020F.924.953.033.903.948.038.756.873.055.842.899.064.828.917.037.908.961.028
LDF [56]2020F.930.951.034.902.944.037.773.881.052.853.903.062.855.929.034.914.960.028
KRN [59]2021F.931.951.032.911.950.036.793.893.050.851.894.068.865.934.033.920.961.027
WSSA [65]2020W.870.917.059.869.929.049.703.845.068.785.855.096.742.869.062.860.932.047
MFNet [38]2021W.844.889.084.872.923.059.621.784.098.756.824.115.693.832.079.839.919.058
SCW [62]2021W.900.931.049.898.940.040.758.862.060.827.879.080.823.890.049.896.943.038
EDNS [64]2020U.872.906.068.880.932.051.682.821.076.801.846.097.735.847.065.874.933.046
SelfMask [43]2022U.856.920.058.844.925.050.668.815.078.774.856.087.714.848.063.819.915.053
DCFD [29]2022U.888.915.059.888.930.045.710.837.070.795.860.090.764.855.064.889.935.042
Ourss1-U.847.912.057.849.925.054.622.773.111.763.840.093.676.814.082.819.914.053
Ours-U.916.938.044.904.944.039.745.863.061.830.882.074.810.901.047.902.947.037
Oursmm-U.913.942.043.899.945.039.737.855.064.831.890.073.803.899.048.898.946.037
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.471, + 0.891, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.68, + 0.713, + 0.695 + ], + "angle": 0, + "content": "Figure 6. Visual comparison between state-of-the-art SOD methods and ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.471, + 0.864 + ], + "angle": 0, + "content": "is competitive to recent weakly-supervised SCW [62] and fully-supervised methods KRN [59]. Furthermore, the results of “\\(\\mathbf{Ours}_{s1}\\)” prove that our method extracts precise saliency knowledge from training samples, but somehow weaken when deploying on unseen images. In addition, the similar results of “\\(\\mathbf{Ours}_{mm}\\)” and “\\(\\mathbf{Ours}\\)” imply that extra multimodal data may not bring significant improvements to RGB SOD task. As proved in [53], the large-scale DUTS-TR dataset is saturated for training saliency detectors, even in the unsupervised case." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "A qualitative comparison is illustrated in Fig. 6. Overall, our saliency predictions are much more precise than other" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.773 + ], + "angle": 0, + "content": "unsupervised methods. For example, in the first example, our method and two fully-supervised methods well segment target object with precise boundaries, while other methods fail to capture the tiny salient object." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.786, + 0.753, + 0.801 + ], + "angle": 0, + "content": "4.2. Results on Multimodal SOD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct more experiments on multimodal tasks, including RGB-D, RGB-T and video SOD. Noted that we precompute the optical flow for each video frame as an extra modality. To verify the improvements brought by multimodal data, we build two sets for training: (1) Task-specific data (\"Ours\"). For example, for RGB-T SOD task, we only" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7262" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.089, + 0.822, + 0.104 + ], + "angle": 0, + "content": "Table 2. Results on RGB-D SOD benchmarks. \"F\" and \"U\" mean fully-supervised and unsupervised, respectively." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.106, + 0.885, + 0.222 + ], + "angle": 0, + "content": "
MethodsYearSup.RGBD135NJUDNLPRSIP
\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)
DSA2F [45]2021F.899.958.021.901.937.039.897.953.024---
SPNet [74]2021F.927.984.013---.903.959.021.893.931.043
CCFE [28]2022F.911.964.020.914.953.032.907.962.021.889.923.047
DSU [21]2022U.767.895.061.719.797.135.745.879.065.619.774.156
Ours-U.834.945.037.787.829.109.834.924.043.716.784.124
\\(Ours_{mm}\\)-U.877.946.029.862.908.060.852.931.034.873.925.051
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.235, + 0.879, + 0.249 + ], + "angle": 0, + "content": "Table 3. Results on VSOD benchmarks. \"F\", \"W\" and \"U\" mean fully-supervised, weakly-supervised and unsupervised, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.251, + 0.89, + 0.368 + ], + "angle": 0, + "content": "
MethodsYearSup.DAVSODDAVISSegV2FBMS
\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M} \\downarrow\\)
PCSA [17]2020F.556.749.077.794.922.022.789.931.019.783.868.042
TENet [40]2020F.595.773.067.821.941.017---.851.915.026
STVS [4]2021F.563.764.080.812.940.022.835.950.016.821.903.042
WSVSOD [70]2021W.492.710.103.731.900.036.711.909.031.736.840.084
Ours-U.534.747.084.751.913.042.751.914.033.732.794.100
\\(Ours_{mm}\\)-U.547.762.085.756.908.037.808.927.021.795.876.060
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.38, + 0.822, + 0.394 + ], + "angle": 0, + "content": "Table 4. Results on RGB-T SOD benchmarks. \"F\" and \"U\" mean fully-supervised and unsupervised, respectively." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.396, + 0.861, + 0.512 + ], + "angle": 0, + "content": "
MethodsYearSup.VT5000VT1000VT821
Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓
MIED [46]2020F.761.880.050.853.928.030.760.877.050
MIDD [47]2021F.801.899.043.882.942.027.805.898.045
APNet [75]2021F.821.918.035.885.951.021.818.912.034
CCFE [28]2022F.859.937.030.906.963.018.857.934.027
Ours-U.810.904.046.885.939.031.805.900.043
Oursmm-U.807.903.047.881.939.032.805.899.044
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.526, + 0.427, + 0.54 + ], + "angle": 0, + "content": "Table 5. Label quality on multimodal SOD datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.542, + 0.464, + 0.617 + ], + "angle": 0, + "content": "
MethodsRGBRGB-DVSODRGB-T
Fβ↑M↓Fβ↑M↓Fβ↑M↓Fβ↑M↓
Ours.917.038.746.110.557.099.932.029
Oursrgb.917.040.804.068.613.089.920.030
Oursmm.918.040.826.062.637.079.923.029
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.681 + ], + "angle": 0, + "content": "use 2500 RGB-thermal image pairs for training as prevalent RGB-T SOD methods [46,47,75]; (2) Multimodal data from the training sets of four SOD tasks (\"Ours\\(_{mm}\\)\")." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.471, + 0.834 + ], + "angle": 0, + "content": "Overall, our method reports state-of-the-art performance on all multimodal SOD tasks. Specifically, for the RGB-D SOD task in Tab. 2, our method surpasses the latest unsupervised method DSU [21]. For video and RGB-T SOD tasks in Tab. 3 and 4, to the best of our knowledge, our method is the first unsupervised method and is competitive to supervised methods. Furthermore, our method trained on multimodal data (\"Oursmm\") produce high-quality pseudo labels for all multimodal SOD datasets simultaneously and has achieved performance improvements on most tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.231, + 0.863 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Label quality comparison. In Tab. 5, we exhibit the scores of pseudo labels on multimodal SOD datasets. Ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.786 + ], + "angle": 0, + "content": "cept for the above two training sets, we employ an additional set that collects all RGB images from four tasks to train our method, denoted as “\\(\\mathbf{Ours}_{rgb}\\)”. Overall, “\\(\\mathbf{Ours}_{mm}\\)” reports more generalized results on all datasets. Moreover, training on task-specific data (“Ours”) slightly surpasses the performance of “\\(\\mathbf{Ours}_{mm}\\)” on the training sets of RGB and RGB-T tasks. Since our method is trained without ground truth, fitting a hybrid dataset may cause a slight performance drop on some subsets. In addition, the comparison between “\\(\\mathbf{Ours}_{rgb}\\)” and “\\(\\mathbf{Ours}_{mm}\\)” prove that multimodal data significantly improves the quality of the generated pseudo labels. For an intuitive comparison, we show some examples in Fig. 7. Trained with task-specific data, our method precisely localizes salient regions but fails to segment the whole objects. With more training data, the network perceives fine-grained concepts of target objects, resulting in more complete segmentations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Impact of initial saliency cues. Our method is based on the activation maps of a pre-trained network [6, 18] instead of traditional methods [22,26,76] as other DL-based USOD methods [29,32,55,63,66]. We compare the performance of these two types of designs in Tab. 6. Overall, the activation map is worse before training, whereas better than traditional methods [22,26,76] after training. The activation maps are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7263" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.091, + 0.472, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.246, + 0.47, + 0.276 + ], + "angle": 0, + "content": "Figure 7. Examples of the generated pseudo labels. The \"D/T/O\" means depth map, thermal image or optical flow." + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.293, + 0.432, + 0.307 + ], + "angle": 0, + "content": "Table 6. Comparison between different saliency cues." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.309, + 0.468, + 0.405 + ], + "angle": 0, + "content": "
Saliency cuesBefore trainingAfter training
\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}\\downarrow\\)
MC [22].810.877.145.889.914.053
DSR [26].780.867.118.886.901.059
RBD [76].803.883.109.891.912.053
Ours.451.659.353.914.940.039
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.468, + 0.546 + ], + "angle": 0, + "content": "dynamic during training, such that the network strengthens the learned saliency knowledge incrementally. On the contrary, using traditional methods, the network learns from fixed labels throughout the training process, and thus fits the biased knowledge in those traditional methods. In addition, traditional methods introduce extra computation loads, while the pre-trained network is indispensable for all DL-based USOD methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.548, + 0.469, + 0.593 + ], + "angle": 0, + "content": "Effectiveness of loss functions. We conduct ablation studies on various combinations of loss functions in Tab. 7 and more details are introduced as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.469, + 0.716 + ], + "angle": 0, + "content": "In ablation study A, our method receives continuous improvements by appending new losses (\\(L_{ms}\\) and \\(L_{btm}\\)) or replacing the original \\(L_{adb}\\) to our \\(L_{csd}\\). This experiment proves that these three losses can assist the network mine more detailed saliency knowledge from different perspectives. More importantly, these losses are supplementary to each other, such that their combination can collaborate to produce high-quality pseudo labels for SOD datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.853 + ], + "angle": 0, + "content": "In ablation study B, our \\( L_{csd} \\) loss can provide precise saliency localization information and thus achieves the best performance among all competitors. Specifically, \\( L_{bce} \\) and \\( L_{1} \\) fail to mine detailed and reliable saliency knowledge from noisy saliency cues. \\( L_{adb} \\) [72] improves the robustness of the learned saliency knowledge by focusing on easy samples, and thus surpasses the \\( L_{1} \\) loss. \\( L_{spl} \\) [2] also employs a dynamic schedule, but its binary weights filter out latent saliency knowledge in hard samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In ablation study C, we compare our \\(L_{btm}\\) with \\(L_{lsc}\\) [62] and two variants: (1) using L1 distance for the texture features in appearance space, denoted as \\(L_{c1}\\); (2) removing the" + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.09, + 0.825, + 0.102 + ], + "angle": 0, + "content": "Table 7. Ablation studies on loss functions." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.106, + 0.882, + 0.323 + ], + "angle": 0, + "content": "
TagLossFβ↑Eξ↑M↓
A1Ladb [72] (Baseline).882.915.071
A2Ladb [72] + Lms.891.921.066
A3Lcsd + Lms.908.937.050
A4Lcsd + Lbtm + Lms (Ours).917.945.038
B1Lbce + Lbtm + Lms.558.540.208
B2L1 + Lbtm + Lms.895.928.044
B3Ladb [72] + Lbtm + Lms.903.934.042
B4Lspl [2] + Lbtm + Lms.910.938.041
B5Lcsd + Lbtm + Lms (Ours).917.945.038
C1Lcsd + Lc1 + Lms.735.819.120
C2Lcsd + Llsc [62] + Lms.906.923.052
C3Lcsd + Lc2 + Lms.907.926.051
C4Lcsd + Lbtm + Lms (Ours).917.945.038
" + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.335, + 0.828, + 0.349 + ], + "angle": 0, + "content": "Table 8. Label quality on DUTS-TR dataset." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.352, + 0.873, + 0.402 + ], + "angle": 0, + "content": "
Pre-trainingFβ↑Eξ↑M↓
Supervised.915.942.039
Unsupervised (Ours).917.945.038
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.546 + ], + "angle": 0, + "content": "boundary masks in \\( L_{btm} \\), denoted as \\( L_{c2} \\). In general, our \\( L_{btm} \\) outperforms other variants. Specifically, the L1 distance in \\( L_{c1} \\) causes the texture features to be not distinctive enough, resulting in ambiguous saliency boundaries. \\( L_{lsc} \\) focuses more on the adjacent pixels so that the boundaries are more susceptible to a limited region rather than a larger patch. Moreover, using \\( L_{c2} \\), edges within objects or backgrounds may corrupt the learned saliency knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Supervised or unsupervised pre-training? Under the unsupervised setting, whether we can use a encoder with supervised pre-training is controversial. For the ImageNet dataset [10], the class labels of images usually indicate the category of the most salient object. It means that the supervised encoder receives extra saliency knowledge from these manual labels. Thus, for completely unsupervised SOD, we employ unsupervised MoCo-v2 to initialize our encoder. As listed in Tab. 8, the performance of our method when using supervised encoder is comparable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an Unsupervised Salient Object Detection (USOD) method guided by two novel mechanisms. First, we propose a Confidence-aware Saliency Distilling (CSD) to learn saliency knowledge from easy samples to hard ones with noisy labels progressively. Second, we propose a Boundary-aware Texture Matching (BTM) to make the location of saliency prediction more accurate. As a result, the proposed method produces high-quality pseudo labels to train saliency detectors. Experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method outperforms existing USOD methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7264" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.471, + 0.226 + ], + "angle": 0, + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295. Springer, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[4] Chenglizhao Chen, Guotao Wang, Chong Peng, Yuming Fang, Dingwen Zhang, and Hong Qin. Exploring rich and efficient spatial temporal interactions for real-time video salient object detection. IEEE Transactions on Image Processing, 30:3995-4007, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Peijia Chen, Jianhuang Lai, Guangcong Wang, and Huajun Zhou. Confidence-guided adaptive gate and dual differential enhancement for video salient object detection. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE Transactions on Image Processing, 30:431-443, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.591 + ], + "angle": 0, + "content": "[8] Xiaolong Cheng, Xuan Zheng, Jialun Pei, He Tang, Zehua Lyu, and Chuanbo Chen. Depth-induced gap-reducing network for rgb-d salient object detection: An interaction, guidance and refinement approach. IEEE Transactions on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Yupeng Cheng, Huazhu Fu, Xingxing Wei, Jiangjian Xiao, and Xiaochun Cao. Depth enhanced saliency detection method. In Proceedings of international conference on internet multimedia computing and service, pages 23-27, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[11] Wenhui Dao, Xian Sun, Xinwei Zheng, Fangzheng Dou, Hongqi Wang, and Kun Fu. Efficient saliency-based object detection in remote sensing images using deep belief networks. IEEE Geoscience and Remote Sensing Letters, 13(2):137-141, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In International Joint Conference on Artificial Intelligence (IJCAI), pages 698-704, 2018. http://dpfan.net/e-measure/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on neural networks and learning systems, 32(5):2075-2089, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[14] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8554-8564, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[15] Mengyang Feng, Hutchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1623-1632, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[16] Carola Figueroa Flores, Abel Gonzalez-Garcia, Joost van de Weijer, and Bogdan Raducanu. Saliency for fine-grained object recognition in domains with scarce training data. Pattern Recognition, 94:62-73, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[17] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 10869–10876, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.382, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[19] Qibin Hou, Ming-Ming Cheng, Xiaowei Hu, Ali Borji, Zhuowen Tu, and Philip HS Torr. Deeply supervised salient object detection with short connections. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3203-3212, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[20] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.556, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[21] Wei Ji, Jingjing Li, Qi Bi, Chuan Guo, Jie Liu, and Li Cheng. Promoting saliency from depth: Deep unsupervised rgb-d saliency detection. arXiv preprint arXiv:2205.07179, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.613, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[22] Bowen Jiang, Lihe Zhang, Huchuan Lu, Chuan Yang, and Ming-Hsuan Yang. Saliency detection via absorbing markov chain. In Proceedings of the IEEE international conference on computer vision, pages 1665-1672, 2013. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In 2014 IEEE international conference on image processing (ICIP), pages 1115-1119. IEEE, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[24] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[25] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 2976-2983, 2013. 2, 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7265" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[27] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[28] Guibiao Liao, Wei Gao, Ge Li, Junle Wang, and Sam Kwong. Cross-collaborative fusion-encoder network for robust rgb-thermal salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[29] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In Thirty-sixth AAAI conference on artificial intelligence, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.471, + 0.306 + ], + "angle": 0, + "content": "[30] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3089-3098, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[31] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence, 33(2):353-367, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[32] Duc Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction with self-supervision. arXiv preprint arXiv:1909.13055, 2019. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.504 + ], + "angle": 0, + "content": "[33] Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine intelligence, 24(7):971-987, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[34] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9413-9422, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[35] Prashant W Patil, Subrahmanyam Murala, Abhinav Dhall, and Sachin Chaudhary. Msednet: multi-scale deep saliency learning for moving object detection. In 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 1670-1675. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[36] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In European conference on computer vision, pages 92-109. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[38] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[40] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In European Conference on Computer Vision, pages 212-228. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[42] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4):717-729, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[44] Ali Shokoufandeh, Ivan Marsic, and Sven J Dickinson. View-based object recognition using saliency maps. Image and Vision Computing, 17(5-6):445-460, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[45] Peng Sun, Wenhu Zhang, Huanyu Wang, Songyuan Li, and Xi Li. Deep rgb-d saliency detection with depth-sensitive attention and automatic multi-modal fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1407-1417, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[46] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive encoder-decoder network for rgbt salient object detection. arXiv e-prints, pages arXiv-2005, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[47] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE Transactions on Image Processing, 30:5678–5691, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[48] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[49] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[50] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE Transactions on Multimedia, 22(1):160-173, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[51] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. RGB-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In Chinese Conference on Image and Graphics Technologies, pages 359-369. Springer, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7266" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. arXiv preprint arXiv:1904.09146, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[54] Wenguan Wang, Shuyang Zhao, Jianbing Shen, Steven CH Hoi, and Ali Borji. Salient object detection with pyramid attention and salient edges. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1448-1457, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[55] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11727-11736, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[56] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13025-13034, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.431 + ], + "angle": 0, + "content": "[57] Hongfa Wen, Chenggang Yan, Xiaofei Zhou, Runmin Cong, Yaoqi Sun, Bolun Zheng, Jiyong Zhang, Yongjun Bao, and Guiguang Ding. Dynamic selective network for rgb-d salient object detection. IEEE Transactions on Image Processing, 30:9179-9192, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[58] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3907-3916, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[59] Binwei Xu, Haoran Liang, Ronghua Liang, and Peng Chen. Locate globally, segment locally: A progressive architecture with knowledge review network for salient object detection. In Proceedings. of the AAAI Conference On Artificial Intelligence, pages 1-9, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[60] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[61] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[62] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). AAAI Palo Alto, CA, USA, 2021. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[63] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[64] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European conference on computer vision, pages 349-366. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[65] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "scribble annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12546-12555, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[66] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9029-9038, 2018. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[67] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 202-211, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[68] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[69] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 714-722, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[70] Wangbo Zhao, Jing Zhang, Long Li, Nick Barnes, Nian Liu, and Junwei Han. Weakly supervised video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16826-16835, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[71] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[72] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[73] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9141-9150, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[74] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving rgb-d saliency detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4681-4691, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.892, + 0.781 + ], + "angle": 0, + "content": "[75] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Apnet: Adversarial learning assistance and perceived importance fusion network for all-day rgb-t salient object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.782, + 0.892, + 0.838 + ], + "angle": 0, + "content": "[76] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.839, + 0.892, + 0.894 + ], + "angle": 0, + "content": "[77] Wenbin Zou and Nikos Komodakis. Harf: Hierarchy-associated rich features for salient object detection. In Proceedings of the IEEE international conference on computer vision, pages 406-414, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7267" + } + ] +] \ No newline at end of file diff --git a/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_origin.pdf b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ed89a67c8f8e85497f6f9e7f0470cd8b72063491 --- /dev/null +++ b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/099f5794-0ad0-4870-a20c-4c38e466b27c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65448fd4a8ce3a5690f0324c39f8e5a7dd2841d684b5033c42a15edfebe62d0c +size 2695089 diff --git a/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/full.md b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35c0b8459f69fbf6a6d5501b82a343cc404aab8e --- /dev/null +++ b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/full.md @@ -0,0 +1,368 @@ +# Texture-guided Saliency Distilling for Unsupervised Salient Object Detection + +Huajun Zhou $^{1}$ , Bo Qiao $^{1}$ , Lingxiao Yang $^{1}$ , Jianhuang Lai $^{1,2,3}$ , Xiaohua Xie $^{1,2,3*}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China + +2Guangdong Province Key Laboratory of Information Security Technology, China + +3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +# Abstract + +Deep Learning-based Unsupervised Salient Object Detection (USOD) mainly relies on the noisy saliency pseudo labels that have been generated from traditional handcraft methods or pre-trained networks. To cope with the noisy labels problem, a class of methods focus on only easy samples with reliable labels but ignore valuable knowledge in hard samples. In this paper, we propose a novel USOD method to mine rich and accurate saliency knowledge from both easy and hard samples. First, we propose a Confidence-aware Saliency Distilling (CSD) strategy that scores samples conditioned on samples' confidences, which guides the model to distill saliency knowledge from easy samples to hard samples progressively. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the boundaries of noisy labels by matching the textures around the predicted boundaries. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance. Code is available at www.github.com/moothes/A2S-v2. + +# 1. Introduction + +Unsupervised Salient Object Detection (USOD) methods aim to correctly localize and precisely segment salient objects simultaneously without using manual annotations. Compared to the supervised methods, USOD methods can easily adapt to more practical scenarios (e.g., industrial or medical images) where a large number of labeled images may be very hard to collect. Moreover, USOD methods also can assist some related methods for other tasks, e.g., object recognition [16, 44] and object detection [11, 35]. However, diverse objects, complex backgrounds, and other challenges + +![](images/177627ce4564029ce13663453f419d74dfe7eb917f32ecac11e5693484224b9a.jpg) +(a) Image + +![](images/268fbdf8b043a26f2a8b3e41c3f4b2e43a40379c07fc2c6b920fb47a1d606d7f.jpg) +(b) GT + +![](images/ccd16f9f38784400db3c1a5561208d0252c3ad673d1075a6c13c04839589409a.jpg) +(c) RBD + +![](images/faae30c40b0e6bf5f9603ed1f81264770393f9e285721d78245b02b8e9ed4fc3.jpg) +(d) HS + +![](images/59f89ce8ff5861b1f2e67b3c7b774e17d7937050e829b1dde0a52dcd66fa6d2e.jpg) +(e) USPS + +![](images/18359c3002312e97af235bd60b4545e7b71182ff888a9670f909bab56e9c6123.jpg) +(f) A2S + +![](images/ecd4f700752bfb27c7755b432a22a01af5847be04bcbf8bb3af961fb482c5995.jpg) +(g) Activation +Figure 1. Saliency maps generated by USOD methods. Our result is generated based on the activation map of a deep network. + +![](images/27c0cf6292ecebe1168c2ac0070e99bb3231f34e9bfbd1349693c44a92980d06.jpg) +(h) Ours + +ing conditions bring severe challenges to USOD methods. + +Most Deep Learning-based (DL-based) methods [21, 29, 32, 55, 63, 66] base on the saliency cues extracted by traditional SOD methods (Fig. 1-c and 1-d). These handcrafted features related cues are employed as pseudo labels to train deep networks under certain constraints, e.g., binary cross-entropy (BCE) loss. However, saliency cues by traditional methods usually shift away from target objects, especially in complex scenes. Moreover, conventional constraints, such as BCE loss, works well on fully-supervised SOD methods, but is suboptimal when fitting the noisy labels for unsupervised methods (Fig. 1-e). Recently, Zhou et al. [72] addressed the first issue by extracting saliency cues (Fig. 1-f) based on a unsupervisedly pre-trained network (Fig. 1-g) instead of using traditional methods. During training, they focus on learning reliable saliency knowledge from easy samples, but ignore latent knowledge in hard samples. The main reason is that hard samples may be wrongly-labeled and corrupt the fragile saliency knowledge learned in the early training phase. Therefore, to leverage hard samples, we argue that all samples should be employed in a meaningful order (i.e., from high reliable to low reliable), which is crucial for mining accurate knowledge from noisy labels. Trained by such a strategy, the network can mine valuable knowledge from hard examples without corrupting the knowledge learned from easy samples. + +Deep networks can learn to localize salient regions from + +noisy labels [72], but still struggle to find the precise boundaries of target objects. Generally, the appearance around saliency boundary has a similar texture as in saliency map. Therefore, matching the textures between different maps can serve as a guidance for producing reasonable saliency boundaries. We will demonstrate that above strategies are applicable to multimodal data besides RGB image, including depth map, thermal image, and optical flow. + +Based on the above analysis, we propose a novel framework to tackle the Unsupervised Salient Object Detection (USOD) tasks. Specifically, two strategies are proposed to mine saliency knowledge from noisy saliency labels. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme that scores samples with noisy labels conditioned on samples' confidence. Then, our CSD guides the network to learn saliency knowledge from easy samples to more complex ones progressively by employing an adaptive loss conditioned on the training progress. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the saliency boundaries of noisy labels by matching the textures around the predicted boundaries. During training, the predicted saliency boundaries are shifting toward surrounding edges in the appearance space of the whole image. Finally, guided by above two mechanisms, our method can produce high-quality pseudo labels to train generalized saliency detectors. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art performance compared to existing USOD methods. + +The main contributions of our novel USOD method are: + +1. We propose a Confidence-aware Saliency Distilling (CSD) to mine rich and accurate saliency knowledge from noisy labels, which breaks through the limitation that existing methods cannot utilize hard samples. +2. We propose a Boundary-aware Texture Matching (BTM) to refine the boundary of the predicted saliency maps by matching textures in different spaces. +3. Extensive experiments on RGB, RGB-D, RGB-T and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance. + +# 2. Related Works + +# 2.1. Supervised Salient Object Detection + +Researchers have developed a large family of fully-supervised Salient Object Detection (SOD) algorithms [5,7, 15,19,30,39,54,58,67-69,71,73] in the past decades. Ronneberger et al. [41] proposed a U-shape structure that progressively upsamples and concatenates the smaller features to the larger ones. To ease the annotation burden, Zhang et al. [65] relabeled the DUTS-TR dataset [52] with scribbles + +and leveraged an edge detection model for boundary localization. Yu et al. [62] proposed a local coherence loss to find precise boundary based on scribbles annotations. + +Multimodal SOD tasks aim at using other modality data to improve the SOD performance, such as depth map, thermal image and optical flow. Recently, abundant methods [4,8,14,17,40,45-48,57,74] were proposed. Most of these methods employ a two-stream encoder-decoder structure to aggregate multi-level information in multimodal data. To reduce the annotating cost, Zhao et al. [70] proposed a video SOD dataset with scribbles annotations to indicate the location of salient objects. + +Although the above methods have achieved significant performance, they require numerous human annotations for training, which are expensive to collect. + +# 2.2. Unsupervised Salient Object Detection + +Traditional SOD methods [22, 26, 60, 76] extracted saliency cues from images by modeling the correlation of hand-crafted features. Inspired by the centralization prior, Jiang et al. [22] considered the distances between boundary superpixels and non-boundary superpixels as saliency scores. Yan et al. [60] employed a tree-structure graphical model to compute the saliency results based on several over-segmented maps. The above methods fail to accurately localize salient objects because the global information in hand-crafted features are not representative enough. + +Existing DL-based USOD methods can be divided into two pipelines based on the method used to extract saliency cues from images. First, most USOD methods [29, 32, 55, 63, 64, 66] focused on refining the coarse saliency cues extracted by several traditional SOD methods. For example, Zhang et al. [63] weighted these saliency cues by combining both intra-image and inter-image fusion streams. Zhang et al. [66] designed a noise modeling module to deal with noises in these saliency cues. Nguyen et al. [32] used these saliency cues as labels to train multiple deep networks. Ji et al. [21] refined the saliency maps extracted by traditional SOD methods to produce more accurate saliency predictions. Second, to prevent the localization errors caused by traditional SOD methods, Zhou et al. [72] proposed a novel framework that converts the activation maps of a pre-trained network to high-quality pseudo labels. + +In this paper, we propose two novel mechanisms to tackle USOD tasks and explore the potential of improving the quality of pseudo labels by using multimodal data. + +# 3. Our Approach + +In our method, we propose two novel strategies to mine accurate saliency knowledge based on the noisy activation maps generated by a deep network $\Phi$ , as shown in Fig. 2. + +Activation map generation. Following previous work [72], we generate an activation map $Y$ for input image $X$ + +![](images/c9f3403048fd88e99f5dac79ba7f3639bd8d12656dd8267e399a241f37b53610.jpg) +Figure 2. Overview of the proposed unsupervised framework. The network are guided by two mechanisms: (1) Conidence-aware Saliency Distilling mines rich and accurate saliency knowledge from noisy labels, and (2) Boundary-aware Texture Matching makes the boundary of the predicted saliency map more accurate. We pad some constant maps for the missing modalities. + +using a deep network $\Phi(X) = Y$ : + +$$ +E _ {3}, E _ {4}, E _ {5} = \operatorname {E n c o d e r} (X), +$$ + +$$ +\begin{array}{l} F _ {i} = S E \left(E _ {i}\right), i \in \{3, 4, 5 \}, \\ H = S E \left(\text {一} (E _ {i}, E _ {i}, E _ {i})\right) \end{array} \tag {1} +$$ + +$$ +H = S E (\text {c o n c a t} (F _ {3}, F _ {4}, F _ {5})), +$$ + +$$ +Y = \operatorname {i n v} (\operatorname {S i g m o i d} (\operatorname {s u m} (H - \bar {H}))), +$$ + +where $\bar{H}$ is the spatial mean of $H$ . Specifically, we employ the ResNet-50 [18] pre-trained by MoCo-v2 [6] as our encoder, which is trained without extra manual annotations. In our decoder, four SE blocks [20] integrate multiple encoder features into $H$ . After that, we set $H - \bar{H}$ to ensure that (1) noisy labels are adaptive to input images when using a fixed threshold and (2) coexisting of positive and negative samples. Next, features are summed over the channel dimension to produce one-channel activation map. Finally, the Sigmoid function produces the final saliency scores, and the inv function identifies the regions with more corner pixels as background. + +Training strategy. The activation maps produced by network $\Phi$ perceive some discriminative regions in input images [72], but still are low-quality because of widespread noises. To improve the quality, we employ three strategies to train the network $\Phi$ by: + +$$ +L _ {\Phi} = \lambda_ {c} \left(L _ {c s d} + \hat {L} _ {c s d}\right) + \lambda_ {b} \left(L _ {b t m} + \hat {L} _ {b t m}\right) + \lambda_ {m} L _ {m s}, \tag {2} +$$ + +where all $\lambda$ are hyperparameters. $\hat{L}$ is the loss for the predictions of a different scale. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme to excavate valuable saliency knowledge from simple examples to more complex ones. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to align the boundaries in + +appearance and the predicted saliency maps. In addition, a multi-scale consistency loss $L_{ms}$ ensures our method to produce consistent predictions for multi-scale inputs. + +# 3.1. Confidence-aware Saliency Distilling + +The primary challenge for USOD methods is how to localize salient regions. Existing DL-based USOD methods [32,63,72] are based on noisy saliency cues from traditional SOD methods [22,26,76,77] or pre-trained networks [6,18]. Following [72], we employ the activation maps produced by a pre-trained network as our saliency cues instead of the saliency predictions of traditional SOD methods. The main reasons are two-fold: (1) traditional methods require extra computation loads; (2) network can focus on mining saliency knowledge instead of fitting the inductive bias of traditional methods. We binarize the saliency cues as the initial labels using a fixed threshold of 0.5, while Eqn. 1 ensures that these labels remain adaptive to the input image. + +To generate high-quality pseudo labels, USPS [32] uses these initial labels to train deep networks like other fully-supervised methods. However, the simple training strategy in fully-supervised methods is suboptimal for unsupervised methods. Specifically, we define pixels with saliency scores close to 0.5 as hard examples because of their low confidences. In the noisy labels, hard examples are likely to be wrongly labeled. Therefore, it is difficult to learn robust saliency knowledge from these hard samples using traditional loss functions (e.g., BCE loss). To this end, A2S [72] learns reliable saliency knowledge from easy samples. However, for hard examples, the saliency knowledge hidden in noises is not fully explored. In summary, a customized strategy that organizes the samples in a more meaningful order is crucial for USOD methods. + +![](images/d5307d04d37199010b5171636846edb834e620c762c96ef3e77705d129c57532.jpg) +(a) $\mathrm{L}_1$ + +![](images/45f3d08fa0bdf4f49ca0d8b86556e7ad8757057b1f3417f4417299ff59654473.jpg) +(b) $\mathrm{L}_{\mathrm{bce}}$ + +![](images/952434bf3adb66ac367a13640254c1e139d3a85115fd9e637c9459873ab3dc4d.jpg) +(c) $\mathrm{L_{adb}}$ +Figure 3. The gradient landscapes of different loss functions. + +![](images/c51815631348ac1e794cfc859084468d3de4b8c23d62a5131b7184473e736585.jpg) +(d) $\mathrm{L}_{\mathrm{csd}}$ (Ours) + +To address the above problem, we propose a Confidence-aware Saliency Distillation (CSD) scheme that scores samples with noisy labels conditional on their confidence and training progress. Concretely, easy samples contain reliable knowledge, and thus can assist our method to learn reliable saliency knowledge. On the contrary, the saliency knowledge in hard samples is hidden within noises, and may corrupt the fragile saliency patterns in the early stages of network training. Inspired by self-paced learning [2], we dynamically adjust the gradients for samples by introducing a factor $\rho$ , which is linearly increasing from 0 to 1 as training proceeding. Our $L_{csd}$ loss can be formulated as: + +$$ +L _ {c s d} = - \frac {1}{N} \sum_ {i} ^ {N} | \Phi (\mathbf {x} _ {i}) - 0. 5 | ^ {(2 ^ {1 - \rho})}, \tag {3} +$$ + +where $N$ is the number of pixels, and $\Phi (\mathbf{x}_i)$ is the saliency prediction of pixel $\mathbf{x}_i$ . Notice that noisy label $\mathbf{G}$ is omitted in Eqn. 3 because it is generated based on $\Phi (\mathbf{x}_i)$ . Thus, we can calculate the confidence score $|k_{i}|$ of pixel $\mathbf{x}_i$ in a simpler way as: $|k_{i}| = |\Phi (\mathbf{x}_{i}) - 0.5|$ , where the constant value 0.5 is related to $\bar{H}$ in Eqn. 1. The partial derivative of our $L_{csd}$ over $\Phi (\mathbf{x}_i)$ is: + +$$ +\frac {\partial L _ {c s d}}{\partial \Phi (\mathbf {x} _ {i})} = - \operatorname {s i g n} \left(k _ {i}\right) 2 ^ {(1 - \rho)} \left| k _ {i} \right| ^ {2 ^ {(1 - \rho)} - 1}, \tag {4} +$$ + +where $\text{sign}(k_i) \in \{-1, 1\}$ for negative and positive values, respectively. + +For an intuitive comparison, we draw the gradient landscapes of different losses in Fig. 3, including $L_{1}$ , $L_{bce}$ , $L_{adb}$ [72] and our $L_{csd}$ . The gradient of other losses are consistent throughout the training process, while our $L_{csd}$ varies with the training process. Specifically, in the beginning, our $L_{csd}$ assigns low gradients to hard samples to learn reliable saliency knowledge from easy samples. As training proceeding, the gradients of hard samples are increasing to mine more valuable saliency knowledge. + +![](images/3aa6303ecb981a1b07e5e5d25e23d2a66ed516e7077aa2f6638749a8c102f825.jpg) +Figure 4. Appearance information in multimodal data. + +# 3.2. Boundary-aware Texture Matching + +In general, the appearance of saliency boundary has a similar texture as in saliency prediction. Therefore, matching the textures between different maps can guide our method to produce reasonable saliency scores. This strategy is also applicable to other modalities, such as depth map, thermal image, and optical flow, as shown in Fig. 4. By aggregating the rich appearance information in multimodal data, we can provide a more generalized guidance for our USOD method. + +Based on the above analysis, we propose a Boundary-aware Texture Matching (BTM) strategy to align the saliency boundaries and image edges by matching their textures. First of all, similar to the LBP feature [33], we extract texture vectors from saliency predictions and appearance, respectively. For saliency predictions, the texture vector $T_{i}^{s} = [t_{i,1}^{s}, t_{i,2}^{s}, \dots, t_{i,k^{2}}^{s}]$ for the $i$ -th pixel is computed by $t_{i,j}^{s} = |\Phi(\mathbf{x}_{i}) - \Phi(\mathbf{x}_{j})|, j \in K_{i}$ , where $K_{i}$ is $k \times k$ neighborhoods around pixel $\mathbf{x}_{i}$ . For appearance information, it may be one of RGB image, optical flow, depth map or thermal image. To extract more distinctive features from multimodal data, we formulate the texture vector $T_{i}^{a} = [t_{i,1}^{a}, t_{i,2}^{a}, \dots, t_{i,k^{2}}^{a}]$ as $t_{i,j}^{a} = \exp(-\alpha \sum_{m} \| \mathbf{x}_{i}^{(m)} - \mathbf{x}_{j}^{(m)} \|^{2}), j \in K_{i}$ , where $\alpha$ is a hyperparameter and $\mathbf{x}_{i}^{(m)}$ + +![](images/944cfc938bb9fe20d6c3cfdd5d8358203ecf4da8f25e5226c99b8a85994c9163.jpg) +Figure 5. Graphical illustration of the proposed $L_{btm}$ . + +means the appearance data of modal $m$ . It is noteworthy that a small difference between two pixels produces a small element in $T^s$ , but a large element in $T^a$ . Therefore, $T_i^s \cdot (T_i^a)^T$ is defined as the matching penalty for considering pixel $\mathbf{x}_i$ as saliency boundary, contrary to similarity. Finally, the formula of our $L_{btm}$ is: + +$$ +L _ {b t m} = \frac {\sum_ {i} b _ {i} T _ {i} ^ {s} \cdot \left(T _ {i} ^ {a}\right) ^ {T}}{\sum_ {i} b _ {i}}, \tag {5} +$$ + +where $b_{i}$ is the binary boundary mask of saliency prediction. + +An intuitive example is shown in Fig. 5. For the predicted boundary pixels, we find some nearby pixels that differ significantly in saliency scores. In general, the appearances between boundary and these pixels are also different because they locate within salient objects and backgrounds, respectively. If not the case, we expect the saliency score of this boundary pixel to be close to those nearby pixels, so that the boundary will shift in the opposite direction. After an iterative training process, our method can align the predicted saliency boundaries with image edges. + +# 3.3. Multi-scale Consistency + +Salient objects are consistent in multi-scale inputs. Thus, we resize input images to a reference scale and encourage our method to produce consistent predictions by: + +$$ +L _ {m s} = \sum_ {i} \left| y _ {i} - r e s i z e \left(\hat {y} _ {i}\right) \right| ^ {2}, \tag {6} +$$ + +where $y_{i}$ and $\hat{y}_i$ are saliency predictions of different scales. + +# 3.4. Training a Generalized Detector + +Following previous USOD methods [29,32,72], we generate pseudo labels $\ddot{Y}$ based on saliency predictions $Y$ through $\ddot{Y} = CRF(Y)$ and use them to train an extra saliency detector with the IOU loss. For RGB SOD, we use the same detector as the previous method [72]. For multimodal SOD tasks, we employ the MIDD [47] network as our detector. To ensure a fair comparison with existing SOD methods, we train the extra detector only using task-specific data and corresponding pseudo labels. + +# 4. Experiments + +Implementation details. All experiments are implemented on a single GTX 1080 Ti GPU. The batch size is 8 and input images are resized to $320^2$ . The reference scale is randomly selected from $(192^2, 256^2, 384^2, 448^2)$ . Only horizontal flipping is employed as our data augmentation strategy. We train our method for 20 epochs using the SGD optimizer with an initial learning rate of 0.1, which is decayed linearly. $\alpha$ , $\lambda_c$ , $\lambda_b$ , and $\lambda_m$ are set to 200, 1, 0.05 and 1, respectively. We train the extra detectors for 10 epochs using the SGD optimizer with a learning rate of 0.005. + +Datasets. In our experiments, we follow the prevalent settings of different SOD tasks. Specifically, for RGB SOD, we use the training subsets of MSRA-B [31] or DUTS [52] to train our method, respectively. ECSSD [42], PASCALS [27], HKU-IS [25], DUTS-TE [52], DUT-O [61], and the testing subset of MSRA-B are employed for evaluation. For RGB-D SOD, we choose 2185 samples from NLPR [36] and NJUD [23] as the training set. RGBD135 [9], SIP [13] and the testing subsets of NJUD and NLPR are employed for evaluation. For RGB-T SOD, 2500 images in VT5000 [49] are for training, while VT1000 [50], VT821 [51] and the rest 2500 images in VT5000 are for testing. For video SOD, we choose the training splits of DAVIS [37] and DAVSOD [14] to train our method. Moreover, we randomly select 5 frames of each video in DAVSOD to avoid overfitting. The testing splits of DAVIS, DAVSOD, SegV2 [24] and FBMS [3] are employed for evaluation. + +Metrics. We adopt three criteria for evaluation, including ave- $F_{\beta}$ , Mean Absolute Error $(\mathcal{M})$ and E-Measure $(E_{\xi})$ [12]. Specifically, $F_{\beta} = \frac{(1 + \beta^{2}) \times \text{Precision} \times \text{Recall}}{\beta^{2} \times \text{Precision} + \text{Recall}}$ , where $\beta^{2}$ is set to 0.3 [1]. The ave- $F_{\beta}$ is the $F_{\beta}$ scores by setting the threshold as two times mean values. $\mathcal{M}$ is the absolute error between predictions and ground truth. $E_{\xi}$ measures the global statistics and local pixel matching information. + +# 4.1. Results on RGB SOD + +As shown in Tab. 1, we compare the proposed method with fully-supervised methods, MINet [34], LDF [56], and KRN [59], weakly-supervised methods, WSSA [65], MFNet [38], and SCW [62], and unsupervised methods, SBF [63], MNL [66], USPS [32], EDNS [64], A2S [72], DCFD [29], and SelfMask [43]. We list our results with different settings: (1) saliency results without extra detector and post-processing, denoted as "Ours $_s1$ "; (2) our full method as "Ours"; (3) training our method using multimodal data, while only DUTS-TR or MSRA-B datasets with pseudo labels are employed to train the extra detector ("Ours $_{mm}$ "). + +Either training on MSRA-B or DUTS-TR, the proposed method achieves significant improvements compared to existing USOD methods. Moreover, our unsupervised method + +Table 1. Experiment results on SOD benchmarks. "Sup." indicates the supervised signals used to train SOD methods. "F", "W" and "U" mean fully-supervised, weakly-supervised and unsupervised, respectively. Best scores are in bold. + +
MethodsYearSup.ECSSDMSRA-BDUT-OPASCAL-SDUTS-TEHKU-IS
\(F_{\beta}\uparrow\)\(E_{\xi}\uparrow\)\(\mathcal{M}\downarrow\)\(F_{\beta}\uparrow\)\(E_{\xi}\uparrow\)\(\mathcal{M}\downarrow\)\(F_{\beta}\uparrow\)\(E_{\xi}\uparrow\)\(\mathcal{M}\downarrow\)\(F_{\beta}\uparrow\)\(E_{\xi}\uparrow\)\(\mathcal{M}\downarrow\)\(F_{\beta} \uparrow\)\(E_{\xi}\uparrow\)\(\mathcal{M}\downarrow\)
Trained on MSRA-B
SBF [63]2017U.812.878.087.867.929.058.611.771.106.711.795.131.627.785.105.805.895.074
MNL [66]2018U.874.906.069.881.932.053.683.821.076.792.846.091---.874.932.047
USPS [32]2019U.875.903.064.896.938.042.715.839.069.770.828.107.730.840.072.880.933.043
DCFD [29]2022U.880.900.064.903.938.041.731.838.064.773.830.105.744.832.068.887.926.044
A2S [72]2022U.888.911.064.902.941.041.719.841.069.790.838.106.750.860.065.887.937.042
Ours-U.902.923.056.912.948.036.731.851.065.803.848.099.767.871.061.891.939.041
Trained on DUTS-TR
MINet [34]2020F.924.953.033.903.948.038.756.873.055.842.899.064.828.917.037.908.961.028
LDF [56]2020F.930.951.034.902.944.037.773.881.052.853.903.062.855.929.034.914.960.028
KRN [59]2021F.931.951.032.911.950.036.793.893.050.851.894.068.865.934.033.920.961.027
WSSA [65]2020W.870.917.059.869.929.049.703.845.068.785.855.096.742.869.062.860.932.047
MFNet [38]2021W.844.889.084.872.923.059.621.784.098.756.824.115.693.832.079.839.919.058
SCW [62]2021W.900.931.049.898.940.040.758.862.060.827.879.080.823.890.049.896.943.038
EDNS [64]2020U.872.906.068.880.932.051.682.821.076.801.846.097.735.847.065.874.933.046
SelfMask [43]2022U.856.920.058.844.925.050.668.815.078.774.856.087.714.848.063.819.915.053
DCFD [29]2022U.888.915.059.888.930.045.710.837.070.795.860.090.764.855.064.889.935.042
Ourss1-U.847.912.057.849.925.054.622.773.111.763.840.093.676.814.082.819.914.053
Ours-U.916.938.044.904.944.039.745.863.061.830.882.074.810.901.047.902.947.037
Oursmm-U.913.942.043.899.945.039.737.855.064.831.890.073.803.899.048.898.946.037
+ +![](images/77e557d8a7e4f54dd74aff4836cab19cf2837d1a98c239306df109f9670f068b.jpg) +Figure 6. Visual comparison between state-of-the-art SOD methods and ours. + +is competitive to recent weakly-supervised SCW [62] and fully-supervised methods KRN [59]. Furthermore, the results of “ $\mathbf{Ours}_{s1}$ ” prove that our method extracts precise saliency knowledge from training samples, but somehow weaken when deploying on unseen images. In addition, the similar results of “ $\mathbf{Ours}_{mm}$ ” and “ $\mathbf{Ours}$ ” imply that extra multimodal data may not bring significant improvements to RGB SOD task. As proved in [53], the large-scale DUTS-TR dataset is saturated for training saliency detectors, even in the unsupervised case. + +A qualitative comparison is illustrated in Fig. 6. Overall, our saliency predictions are much more precise than other + +unsupervised methods. For example, in the first example, our method and two fully-supervised methods well segment target object with precise boundaries, while other methods fail to capture the tiny salient object. + +# 4.2. Results on Multimodal SOD + +We conduct more experiments on multimodal tasks, including RGB-D, RGB-T and video SOD. Noted that we precompute the optical flow for each video frame as an extra modality. To verify the improvements brought by multimodal data, we build two sets for training: (1) Task-specific data ("Ours"). For example, for RGB-T SOD task, we only + +Table 2. Results on RGB-D SOD benchmarks. "F" and "U" mean fully-supervised and unsupervised, respectively. + +
MethodsYearSup.RGBD135NJUDNLPRSIP
\(F_{\beta}\)↑\(E_{\xi}\)↑\(\mathcal{M}_{\downarrow}\)\(F_{\beta}\)↑\(E_{\xi}\)↑\(\mathcal{M}_{\downarrow}\)\(F_{\beta}\)↑\(E_{\xi}\)↑\(\mathcal{M}_{\downarrow}\)\(F_{\beta}\)↑\(E_{\xi}\)↑\(\mathcal{M}_{\downarrow}\)
DSA2F [45]2021F.899.958.021.901.937.039.897.953.024---
SPNet [74]2021F.927.984.013---.903.959.021.893.931.043
CCFE [28]2022F.911.964.020.914.953.032.907.962.021.889.923.047
DSU [21]2022U.767.895.061.719.797.135.745.879.065.619.774.156
Ours-U.834.945.037.787.829.109.834.924.043.716.784.124
\(Ours_{mm}\)-U.877.946.029.862.908.060.852.931.034.873.925.051
+ +Table 3. Results on VSOD benchmarks. "F", "W" and "U" mean fully-supervised, weakly-supervised and unsupervised, respectively. + +
MethodsYearSup.DAVSODDAVISSegV2FBMS
\(F_{\beta \uparrow}\)\(E_{\xi \uparrow}\)\(\mathcal{M}_{\downarrow}\)\(F_{\beta \uparrow}\)\(E_{\xi \uparrow}\)\(\mathcal{M}_{\downarrow}\)\(F_{\beta \uparrow}\)\(E_{\xi \uparrow}\)\(\mathcal{M}_{\downarrow}\)\(F_{\beta \uparrow}\)\(E_{\xi \uparrow}\)\(\mathcal{M} \downarrow\)
PCSA [17]2020F.556.749.077.794.922.022.789.931.019.783.868.042
TENet [40]2020F.595.773.067.821.941.017---.851.915.026
STVS [4]2021F.563.764.080.812.940.022.835.950.016.821.903.042
WSVSOD [70]2021W.492.710.103.731.900.036.711.909.031.736.840.084
Ours-U.534.747.084.751.913.042.751.914.033.732.794.100
\(Ours_{mm}\)-U.547.762.085.756.908.037.808.927.021.795.876.060
+ +Table 4. Results on RGB-T SOD benchmarks. "F" and "U" mean fully-supervised and unsupervised, respectively. + +
MethodsYearSup.VT5000VT1000VT821
Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓
MIED [46]2020F.761.880.050.853.928.030.760.877.050
MIDD [47]2021F.801.899.043.882.942.027.805.898.045
APNet [75]2021F.821.918.035.885.951.021.818.912.034
CCFE [28]2022F.859.937.030.906.963.018.857.934.027
Ours-U.810.904.046.885.939.031.805.900.043
Oursmm-U.807.903.047.881.939.032.805.899.044
+ +Table 5. Label quality on multimodal SOD datasets. + +
MethodsRGBRGB-DVSODRGB-T
Fβ↑M↓Fβ↑M↓Fβ↑M↓Fβ↑M↓
Ours.917.038.746.110.557.099.932.029
Oursrgb.917.040.804.068.613.089.920.030
Oursmm.918.040.826.062.637.079.923.029
+ +use 2500 RGB-thermal image pairs for training as prevalent RGB-T SOD methods [46,47,75]; (2) Multimodal data from the training sets of four SOD tasks ("Ours $_{mm}$ "). + +Overall, our method reports state-of-the-art performance on all multimodal SOD tasks. Specifically, for the RGB-D SOD task in Tab. 2, our method surpasses the latest unsupervised method DSU [21]. For video and RGB-T SOD tasks in Tab. 3 and 4, to the best of our knowledge, our method is the first unsupervised method and is competitive to supervised methods. Furthermore, our method trained on multimodal data ("Oursmm") produce high-quality pseudo labels for all multimodal SOD datasets simultaneously and has achieved performance improvements on most tasks. + +# 4.3. Ablation Study + +Label quality comparison. In Tab. 5, we exhibit the scores of pseudo labels on multimodal SOD datasets. Ex + +cept for the above two training sets, we employ an additional set that collects all RGB images from four tasks to train our method, denoted as “ $\mathbf{Ours}_{rgb}$ ”. Overall, “ $\mathbf{Ours}_{mm}$ ” reports more generalized results on all datasets. Moreover, training on task-specific data (“Ours”) slightly surpasses the performance of “ $\mathbf{Ours}_{mm}$ ” on the training sets of RGB and RGB-T tasks. Since our method is trained without ground truth, fitting a hybrid dataset may cause a slight performance drop on some subsets. In addition, the comparison between “ $\mathbf{Ours}_{rgb}$ ” and “ $\mathbf{Ours}_{mm}$ ” prove that multimodal data significantly improves the quality of the generated pseudo labels. For an intuitive comparison, we show some examples in Fig. 7. Trained with task-specific data, our method precisely localizes salient regions but fails to segment the whole objects. With more training data, the network perceives fine-grained concepts of target objects, resulting in more complete segmentations. + +Impact of initial saliency cues. Our method is based on the activation maps of a pre-trained network [6, 18] instead of traditional methods [22,26,76] as other DL-based USOD methods [29,32,55,63,66]. We compare the performance of these two types of designs in Tab. 6. Overall, the activation map is worse before training, whereas better than traditional methods [22,26,76] after training. The activation maps are + +![](images/d682fdaa7ecd89d795a228c17a8adb0068f9c14ce1cc576ee58a8f9be5f3eb53.jpg) +Figure 7. Examples of the generated pseudo labels. The "D/T/O" means depth map, thermal image or optical flow. + +Table 6. Comparison between different saliency cues. + +
Saliency cuesBefore trainingAfter training
\(F_{\beta}\)↑\(E_{\xi}\)↑\(\mathcal{M}\downarrow\)\(F_{\beta}\)↑\(E_{\xi}\)↑\(\mathcal{M}\downarrow\)
MC [22].810.877.145.889.914.053
DSR [26].780.867.118.886.901.059
RBD [76].803.883.109.891.912.053
Ours.451.659.353.914.940.039
+ +dynamic during training, such that the network strengthens the learned saliency knowledge incrementally. On the contrary, using traditional methods, the network learns from fixed labels throughout the training process, and thus fits the biased knowledge in those traditional methods. In addition, traditional methods introduce extra computation loads, while the pre-trained network is indispensable for all DL-based USOD methods. + +Effectiveness of loss functions. We conduct ablation studies on various combinations of loss functions in Tab. 7 and more details are introduced as follows. + +In ablation study A, our method receives continuous improvements by appending new losses ( $L_{ms}$ and $L_{btm}$ ) or replacing the original $L_{adb}$ to our $L_{csd}$ . This experiment proves that these three losses can assist the network mine more detailed saliency knowledge from different perspectives. More importantly, these losses are supplementary to each other, such that their combination can collaborate to produce high-quality pseudo labels for SOD datasets. + +In ablation study B, our $L_{csd}$ loss can provide precise saliency localization information and thus achieves the best performance among all competitors. Specifically, $L_{bce}$ and $L_{1}$ fail to mine detailed and reliable saliency knowledge from noisy saliency cues. $L_{adb}$ [72] improves the robustness of the learned saliency knowledge by focusing on easy samples, and thus surpasses the $L_{1}$ loss. $L_{spl}$ [2] also employs a dynamic schedule, but its binary weights filter out latent saliency knowledge in hard samples. + +In ablation study C, we compare our $L_{btm}$ with $L_{lsc}$ [62] and two variants: (1) using L1 distance for the texture features in appearance space, denoted as $L_{c1}$ ; (2) removing the + +Table 7. Ablation studies on loss functions. + +
TagLossFβ↑Eξ↑M↓
A1Ladb [72] (Baseline).882.915.071
A2Ladb [72] + Lms.891.921.066
A3Lcsd + Lms.908.937.050
A4Lcsd + Lbtm + Lms (Ours).917.945.038
B1Lbce + Lbtm + Lms.558.540.208
B2L1 + Lbtm + Lms.895.928.044
B3Ladb [72] + Lbtm + Lms.903.934.042
B4Lspl [2] + Lbtm + Lms.910.938.041
B5Lcsd + Lbtm + Lms (Ours).917.945.038
C1Lcsd + Lc1 + Lms.735.819.120
C2Lcsd + Llsc [62] + Lms.906.923.052
C3Lcsd + Lc2 + Lms.907.926.051
C4Lcsd + Lbtm + Lms (Ours).917.945.038
+ +Table 8. Label quality on DUTS-TR dataset. + +
Pre-trainingFβ↑Eξ↑M↓
Supervised.915.942.039
Unsupervised (Ours).917.945.038
+ +boundary masks in $L_{btm}$ , denoted as $L_{c2}$ . In general, our $L_{btm}$ outperforms other variants. Specifically, the L1 distance in $L_{c1}$ causes the texture features to be not distinctive enough, resulting in ambiguous saliency boundaries. $L_{lsc}$ focuses more on the adjacent pixels so that the boundaries are more susceptible to a limited region rather than a larger patch. Moreover, using $L_{c2}$ , edges within objects or backgrounds may corrupt the learned saliency knowledge. + +Supervised or unsupervised pre-training? Under the unsupervised setting, whether we can use a encoder with supervised pre-training is controversial. For the ImageNet dataset [10], the class labels of images usually indicate the category of the most salient object. It means that the supervised encoder receives extra saliency knowledge from these manual labels. Thus, for completely unsupervised SOD, we employ unsupervised MoCo-v2 to initialize our encoder. As listed in Tab. 8, the performance of our method when using supervised encoder is comparable. + +# 5. Conclusion + +In this paper, we propose an Unsupervised Salient Object Detection (USOD) method guided by two novel mechanisms. First, we propose a Confidence-aware Saliency Distilling (CSD) to learn saliency knowledge from easy samples to hard ones with noisy labels progressively. Second, we propose a Boundary-aware Texture Matching (BTM) to make the location of saliency prediction more accurate. As a result, the proposed method produces high-quality pseudo labels to train saliency detectors. Experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method outperforms existing USOD methods. + +# References + +[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5 +[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 4, 8 +[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295. Springer, 2010. 5 +[4] Chenglizhao Chen, Guotao Wang, Chong Peng, Yuming Fang, Dingwen Zhang, and Hong Qin. Exploring rich and efficient spatial temporal interactions for real-time video salient object detection. IEEE Transactions on Image Processing, 30:3995-4007, 2021. 2, 7 +[5] Peijia Chen, Jianhuang Lai, Guangcong Wang, and Huajun Zhou. Confidence-guided adaptive gate and dual differential enhancement for video salient object detection. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2 +[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3, 7 +[7] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE Transactions on Image Processing, 30:431-443, 2020. 2 +[8] Xiaolong Cheng, Xuan Zheng, Jialun Pei, He Tang, Zehua Lyu, and Chuanbo Chen. Depth-induced gap-reducing network for rgb-d salient object detection: An interaction, guidance and refinement approach. IEEE Transactions on Multimedia, 2022. 2 +[9] Yupeng Cheng, Huazhu Fu, Xingxing Wei, Jiangjian Xiao, and Xiaochun Cao. Depth enhanced saliency detection method. In Proceedings of international conference on internet multimedia computing and service, pages 23-27, 2014. 5 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 8 +[11] Wenhui Dao, Xian Sun, Xinwei Zheng, Fangzheng Dou, Hongqi Wang, and Kun Fu. Efficient saliency-based object detection in remote sensing images using deep belief networks. IEEE Geoscience and Remote Sensing Letters, 13(2):137-141, 2016. 1 +[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In International Joint Conference on Artificial Intelligence (IJCAI), pages 698-704, 2018. http://dpfan.net/e-measure/.5 +[13] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks. + +IEEE Transactions on neural networks and learning systems, 32(5):2075-2089, 2020. 5 +[14] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8554-8564, 2019. 2, 5 +[15] Mengyang Feng, Hutchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1623-1632, 2019. 2 +[16] Carola Figueroa Flores, Abel Gonzalez-Garcia, Joost van de Weijer, and Bogdan Raducanu. Saliency for fine-grained object recognition in domains with scarce training data. Pattern Recognition, 94:62-73, 2019. 1 +[17] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 10869–10876, 2020. 2, 7 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7 +[19] Qibin Hou, Ming-Ming Cheng, Xiaowei Hu, Ali Borji, Zhuowen Tu, and Philip HS Torr. Deeply supervised salient object detection with short connections. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3203-3212, 2017. 2 +[20] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3 +[21] Wei Ji, Jingjing Li, Qi Bi, Chuan Guo, Jie Liu, and Li Cheng. Promoting saliency from depth: Deep unsupervised rgb-d saliency detection. arXiv preprint arXiv:2205.07179, 2022. 1, 2, 7 +[22] Bowen Jiang, Lihe Zhang, Huchuan Lu, Chuan Yang, and Ming-Hsuan Yang. Saliency detection via absorbing markov chain. In Proceedings of the IEEE international conference on computer vision, pages 1665-1672, 2013. 2, 3, 7, 8 +[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In 2014 IEEE international conference on image processing (ICIP), pages 1115-1119. IEEE, 2014. 5 +[24] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5 +[25] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, 2015. 5 +[26] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 2976-2983, 2013. 2, 3, 7, 8 + +[27] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, 2014. 5 +[28] Guibiao Liao, Wei Gao, Ge Li, Junle Wang, and Sam Kwong. Cross-collaborative fusion-encoder network for robust rgb-thermal salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7 +[29] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In Thirty-sixth AAAI conference on artificial intelligence, 2022. 1, 2, 5, 6, 7 +[30] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3089-3098, 2018. 2 +[31] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence, 33(2):353-367, 2010. 5 +[32] Duc Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction with self-supervision. arXiv preprint arXiv:1909.13055, 2019. 1, 2, 3, 5, 6, 7 +[33] Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine intelligence, 24(7):971-987, 2002. 4 +[34] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9413-9422, 2020. 5, 6 +[35] Prashant W Patil, Subrahmanyam Murala, Abhinav Dhall, and Sachin Chaudhary. Msednet: multi-scale deep saliency learning for moving object detection. In 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 1670-1675. IEEE, 2018. 1 +[36] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In European conference on computer vision, pages 92-109. Springer, 2014. 5 +[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5 +[38] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 5, 6 +[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE + +Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2 +[40] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In European Conference on Computer Vision, pages 212-228. Springer, 2020. 2, 7 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2 +[42] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4):717-729, 2015. 5 +[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6 +[44] Ali Shokoufandeh, Ivan Marsic, and Sven J Dickinson. View-based object recognition using saliency maps. Image and Vision Computing, 17(5-6):445-460, 1999. 1 +[45] Peng Sun, Wenhu Zhang, Huanyu Wang, Songyuan Li, and Xi Li. Deep rgb-d saliency detection with depth-sensitive attention and automatic multi-modal fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1407-1417, 2021. 2, 7 +[46] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive encoder-decoder network for rgbt salient object detection. arXiv e-prints, pages arXiv-2005, 2020. 2, 7 +[47] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE Transactions on Image Processing, 30:5678–5691, 2021. 2, 5, 7 +[48] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 2 +[49] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 5 +[50] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE Transactions on Multimedia, 22(1):160-173, 2019. 5 +[51] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. RGB-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In Chinese Conference on Image and Graphics Technologies, pages 359-369. Springer, 2018. 5 +[52] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5 + +[53] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. arXiv preprint arXiv:1904.09146, 2019. 6 +[54] Wenguan Wang, Shuyang Zhao, Jianbing Shen, Steven CH Hoi, and Ali Borji. Salient object detection with pyramid attention and salient edges. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1448-1457, 2019. 2 +[55] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11727-11736, 2022. 1, 2, 7 +[56] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13025-13034, 2020. 5, 6 +[57] Hongfa Wen, Chenggang Yan, Xiaofei Zhou, Runmin Cong, Yaoqi Sun, Bolun Zheng, Jiyong Zhang, Yongjun Bao, and Guiguang Ding. Dynamic selective network for rgb-d salient object detection. IEEE Transactions on Image Processing, 30:9179-9192, 2021. 2 +[58] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3907-3916, 2019. 2 +[59] Binwei Xu, Haoran Liang, Ronghua Liang, and Peng Chen. Locate globally, segment locally: A progressive architecture with knowledge review network for salient object detection. In Proceedings. of the AAAI Conference On Artificial Intelligence, pages 1-9, 2021. 5, 6 +[60] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 2 +[61] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 5 +[62] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). AAAI Palo Alto, CA, USA, 2021. 2, 5, 6, 8 +[63] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2, 3, 5, 6, 7 +[64] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European conference on computer vision, pages 349-366. Springer, 2020. 2, 5, 6 +[65] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via + +scribble annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12546-12555, 2020. 2, 5, 6 +[66] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9029-9038, 2018. 1, 2, 5, 6, 7 +[67] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 202-211, 2017. 2 +[68] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In ICCV, 2017. 2 +[69] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 714-722, 2018. 2 +[70] Wangbo Zhao, Jing Zhang, Long Li, Nick Barnes, Nian Liu, and Junwei Han. Weakly supervised video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16826-16835, 2021. 2, 7 +[71] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2 +[72] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 1, 2, 3, 4, 5, 6, 8 +[73] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9141-9150, 2020. 2 +[74] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving rgb-d saliency detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4681-4691, 2021. 2, 7 +[75] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Apnet: Adversarial learning assistance and perceived importance fusion network for all-day rgb-t salient object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 7 +[76] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 2, 3, 7, 8 +[77] Wenbin Zou and Nikos Komodakis. Harf: Hierarchy-associated rich features for salient object detection. In Proceedings of the IEEE international conference on computer vision, pages 406-414, 2015. 3 \ No newline at end of file diff --git a/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/images.zip b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..97c58e2e8987dc1d6a4fd5981bd58f5eb209826f --- /dev/null +++ b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:595e828197050e38aa16f21c4266410d3c5f0a6c9ea1fcdfb0854bcc8398860a +size 973883 diff --git a/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/layout.json b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd13c94e5924c0ed52846d6db4cbd508df8e0d27 --- /dev/null +++ b/2023/Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection/layout.json @@ -0,0 +1,10513 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 120 + ], + "type": "text", + "content": "Texture-guided Saliency Distilling for Unsupervised Salient Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "text", + "content": "Huajun Zhou" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "text", + "content": ", Bo Qiao" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "text", + "content": ", Lingxiao Yang" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "text", + "content": ", Jianhuang Lai" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "text", + "content": ", Xiaohua Xie" + }, + { + "bbox": [ + 105, + 142, + 491, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 157, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 157, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 485, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 157, + 485, + 171 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 171, + 495, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 171, + 495, + 185 + ], + "spans": [ + { + "bbox": [ + 99, + 171, + 495, + 185 + ], + "type": "text", + "content": "2Guangdong Province Key Laboratory of Information Security Technology, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 185, + 534, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 185, + 534, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 185, + 534, + 199 + ], + "type": "text", + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 191, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 290, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 479 + ], + "type": "text", + "content": "Deep Learning-based Unsupervised Salient Object Detection (USOD) mainly relies on the noisy saliency pseudo labels that have been generated from traditional handcraft methods or pre-trained networks. To cope with the noisy labels problem, a class of methods focus on only easy samples with reliable labels but ignore valuable knowledge in hard samples. In this paper, we propose a novel USOD method to mine rich and accurate saliency knowledge from both easy and hard samples. First, we propose a Confidence-aware Saliency Distilling (CSD) strategy that scores samples conditioned on samples' confidences, which guides the model to distill saliency knowledge from easy samples to hard samples progressively. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the boundaries of noisy labels by matching the textures around the predicted boundaries. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance. Code is available at www.github.com/moothes/A2S-v2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": "Unsupervised Salient Object Detection (USOD) methods aim to correctly localize and precisely segment salient objects simultaneously without using manual annotations. Compared to the supervised methods, USOD methods can easily adapt to more practical scenarios (e.g., industrial or medical images) where a large number of labeled images may be very hard to collect. Moreover, USOD methods also can assist some related methods for other tasks, e.g., object recognition [16, 44] and object detection [11, 35]. However, diverse objects, complex backgrounds, and other challenges" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 225, + 367, + 271 + ], + "blocks": [ + { + "bbox": [ + 310, + 225, + 367, + 271 + ], + "lines": [ + { + "bbox": [ + 310, + 225, + 367, + 271 + ], + "spans": [ + { + "bbox": [ + 310, + 225, + 367, + 271 + ], + "type": "image", + "image_path": "177627ce4564029ce13663453f419d74dfe7eb917f32ecac11e5693484224b9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 273, + 353, + 281 + ], + "lines": [ + { + "bbox": [ + 321, + 273, + 353, + 281 + ], + "spans": [ + { + "bbox": [ + 321, + 273, + 353, + 281 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 369, + 226, + 426, + 271 + ], + "blocks": [ + { + "bbox": [ + 369, + 226, + 426, + 271 + ], + "lines": [ + { + "bbox": [ + 369, + 226, + 426, + 271 + ], + "spans": [ + { + "bbox": [ + 369, + 226, + 426, + 271 + ], + "type": "image", + "image_path": "268fbdf8b043a26f2a8b3e41c3f4b2e43a40379c07fc2c6b920fb47a1d606d7f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 273, + 407, + 281 + ], + "lines": [ + { + "bbox": [ + 384, + 273, + 407, + 281 + ], + "spans": [ + { + "bbox": [ + 384, + 273, + 407, + 281 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 427, + 226, + 485, + 271 + ], + "blocks": [ + { + "bbox": [ + 427, + 226, + 485, + 271 + ], + "lines": [ + { + "bbox": [ + 427, + 226, + 485, + 271 + ], + "spans": [ + { + "bbox": [ + 427, + 226, + 485, + 271 + ], + "type": "image", + "image_path": "ccd16f9f38784400db3c1a5561208d0252c3ad673d1075a6c13c04839589409a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 273, + 469, + 281 + ], + "lines": [ + { + "bbox": [ + 441, + 273, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 441, + 273, + 469, + 281 + ], + "type": "text", + "content": "(c) RBD" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 486, + 226, + 544, + 271 + ], + "blocks": [ + { + "bbox": [ + 486, + 226, + 544, + 271 + ], + "lines": [ + { + "bbox": [ + 486, + 226, + 544, + 271 + ], + "spans": [ + { + "bbox": [ + 486, + 226, + 544, + 271 + ], + "type": "image", + "image_path": "faae30c40b0e6bf5f9603ed1f81264770393f9e285721d78245b02b8e9ed4fc3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 273, + 525, + 281 + ], + "lines": [ + { + "bbox": [ + 502, + 273, + 525, + 281 + ], + "spans": [ + { + "bbox": [ + 502, + 273, + 525, + 281 + ], + "type": "text", + "content": "(d) HS" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 282, + 367, + 327 + ], + "blocks": [ + { + "bbox": [ + 309, + 282, + 367, + 327 + ], + "lines": [ + { + "bbox": [ + 309, + 282, + 367, + 327 + ], + "spans": [ + { + "bbox": [ + 309, + 282, + 367, + 327 + ], + "type": "image", + "image_path": "59f89ce8ff5861b1f2e67b3c7b774e17d7937050e829b1dde0a52dcd66fa6d2e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 329, + 353, + 338 + ], + "lines": [ + { + "bbox": [ + 321, + 329, + 353, + 338 + ], + "spans": [ + { + "bbox": [ + 321, + 329, + 353, + 338 + ], + "type": "text", + "content": "(e) USPS" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 369, + 282, + 426, + 327 + ], + "blocks": [ + { + "bbox": [ + 369, + 282, + 426, + 327 + ], + "lines": [ + { + "bbox": [ + 369, + 282, + 426, + 327 + ], + "spans": [ + { + "bbox": [ + 369, + 282, + 426, + 327 + ], + "type": "image", + "image_path": "18359c3002312e97af235bd60b4545e7b71182ff888a9670f909bab56e9c6123.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 329, + 409, + 338 + ], + "lines": [ + { + "bbox": [ + 383, + 329, + 409, + 338 + ], + "spans": [ + { + "bbox": [ + 383, + 329, + 409, + 338 + ], + "type": "text", + "content": "(f) A2S" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 427, + 282, + 485, + 327 + ], + "blocks": [ + { + "bbox": [ + 427, + 282, + 485, + 327 + ], + "lines": [ + { + "bbox": [ + 427, + 282, + 485, + 327 + ], + "spans": [ + { + "bbox": [ + 427, + 282, + 485, + 327 + ], + "type": "image", + "image_path": "ecd4f700752bfb27c7755b432a22a01af5847be04bcbf8bb3af961fb482c5995.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 329, + 478, + 338 + ], + "lines": [ + { + "bbox": [ + 432, + 329, + 478, + 338 + ], + "spans": [ + { + "bbox": [ + 432, + 329, + 478, + 338 + ], + "type": "text", + "content": "(g) Activation" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 348, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 306, + 348, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 545, + 371 + ], + "type": "text", + "content": "Figure 1. Saliency maps generated by USOD methods. Our result is generated based on the activation map of a deep network." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 486, + 282, + 544, + 327 + ], + "blocks": [ + { + "bbox": [ + 486, + 282, + 544, + 327 + ], + "lines": [ + { + "bbox": [ + 486, + 282, + 544, + 327 + ], + "spans": [ + { + "bbox": [ + 486, + 282, + 544, + 327 + ], + "type": "image", + "image_path": "27c0cf6292ecebe1168c2ac0070e99bb3231f34e9bfbd1349693c44a92980d06.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 329, + 527, + 338 + ], + "lines": [ + { + "bbox": [ + 499, + 329, + 527, + 338 + ], + "spans": [ + { + "bbox": [ + 499, + 329, + 527, + 338 + ], + "type": "text", + "content": "(h) Ours" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 386, + 539, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 386, + 539, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 539, + 398 + ], + "type": "text", + "content": "ing conditions bring severe challenges to USOD methods." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 399, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 699 + ], + "type": "text", + "content": "Most Deep Learning-based (DL-based) methods [21, 29, 32, 55, 63, 66] base on the saliency cues extracted by traditional SOD methods (Fig. 1-c and 1-d). These handcrafted features related cues are employed as pseudo labels to train deep networks under certain constraints, e.g., binary cross-entropy (BCE) loss. However, saliency cues by traditional methods usually shift away from target objects, especially in complex scenes. Moreover, conventional constraints, such as BCE loss, works well on fully-supervised SOD methods, but is suboptimal when fitting the noisy labels for unsupervised methods (Fig. 1-e). Recently, Zhou et al. [72] addressed the first issue by extracting saliency cues (Fig. 1-f) based on a unsupervisedly pre-trained network (Fig. 1-g) instead of using traditional methods. During training, they focus on learning reliable saliency knowledge from easy samples, but ignore latent knowledge in hard samples. The main reason is that hard samples may be wrongly-labeled and corrupt the fragile saliency knowledge learned in the early training phase. Therefore, to leverage hard samples, we argue that all samples should be employed in a meaningful order (i.e., from high reliable to low reliable), which is crucial for mining accurate knowledge from noisy labels. Trained by such a strategy, the network can mine valuable knowledge from hard examples without corrupting the knowledge learned from easy samples." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "Deep networks can learn to localize salient regions from" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "text", + "content": "*Corresponding author. This project is supported by the Key-Area Research and Development Program of Guangdong Province (2019B010155003), the National Natural Science Foundation of China (U22A2095, 62072482, 62076258, 62206316), and the Guangdong NSF Project (2022A1515011254)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7257" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "noisy labels [72], but still struggle to find the precise boundaries of target objects. Generally, the appearance around saliency boundary has a similar texture as in saliency map. Therefore, matching the textures between different maps can serve as a guidance for producing reasonable saliency boundaries. We will demonstrate that above strategies are applicable to multimodal data besides RGB image, including depth map, thermal image, and optical flow." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 289, + 418 + ], + "type": "text", + "content": "Based on the above analysis, we propose a novel framework to tackle the Unsupervised Salient Object Detection (USOD) tasks. Specifically, two strategies are proposed to mine saliency knowledge from noisy saliency labels. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme that scores samples with noisy labels conditioned on samples' confidence. Then, our CSD guides the network to learn saliency knowledge from easy samples to more complex ones progressively by employing an adaptive loss conditioned on the training progress. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to refine the saliency boundaries of noisy labels by matching the textures around the predicted boundaries. During training, the predicted saliency boundaries are shifting toward surrounding edges in the appearance space of the whole image. Finally, guided by above two mechanisms, our method can produce high-quality pseudo labels to train generalized saliency detectors. Extensive experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method achieves state-of-the-art performance compared to existing USOD methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 420, + 286, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 420, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 59, + 420, + 286, + 431 + ], + "type": "text", + "content": "The main contributions of our novel USOD method are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 287, + 579 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 441, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 287, + 489 + ], + "type": "text", + "content": "1. We propose a Confidence-aware Saliency Distilling (CSD) to mine rich and accurate saliency knowledge from noisy labels, which breaks through the limitation that existing methods cannot utilize hard samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 287, + 534 + ], + "type": "text", + "content": "2. We propose a Boundary-aware Texture Matching (BTM) to refine the boundary of the predicted saliency maps by matching textures in different spaces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 543, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 287, + 579 + ], + "type": "text", + "content": "3. Extensive experiments on RGB, RGB-D, RGB-T and video SOD benchmarks prove that our method achieves state-of-the-art USOD performance." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 590, + 138, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 138, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 138, + 602 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 610, + 238, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 238, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 238, + 624 + ], + "type": "text", + "content": "2.1. Supervised Salient Object Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Researchers have developed a large family of fully-supervised Salient Object Detection (SOD) algorithms [5,7, 15,19,30,39,54,58,67-69,71,73] in the past decades. Ronneberger et al. [41] proposed a U-shape structure that progressively upsamples and concatenates the smaller features to the larger ones. To ease the annotation burden, Zhang et al. [65] relabeled the DUTS-TR dataset [52] with scribbles" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "and leveraged an edge detection model for boundary localization. Yu et al. [62] proposed a local coherence loss to find precise boundary based on scribbles annotations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 215 + ], + "type": "text", + "content": "Multimodal SOD tasks aim at using other modality data to improve the SOD performance, such as depth map, thermal image and optical flow. Recently, abundant methods [4,8,14,17,40,45-48,57,74] were proposed. Most of these methods employ a two-stream encoder-decoder structure to aggregate multi-level information in multimodal data. To reduce the annotating cost, Zhao et al. [70] proposed a video SOD dataset with scribbles annotations to indicate the location of salient objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 216, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 251 + ], + "type": "text", + "content": "Although the above methods have achieved significant performance, they require numerous human annotations for training, which are expensive to collect." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 258, + 509, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 509, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 509, + 271 + ], + "type": "text", + "content": "2.2. Unsupervised Salient Object Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 396 + ], + "type": "text", + "content": "Traditional SOD methods [22, 26, 60, 76] extracted saliency cues from images by modeling the correlation of hand-crafted features. Inspired by the centralization prior, Jiang et al. [22] considered the distances between boundary superpixels and non-boundary superpixels as saliency scores. Yan et al. [60] employed a tree-structure graphical model to compute the saliency results based on several over-segmented maps. The above methods fail to accurately localize salient objects because the global information in hand-crafted features are not representative enough." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 396, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 587 + ], + "type": "text", + "content": "Existing DL-based USOD methods can be divided into two pipelines based on the method used to extract saliency cues from images. First, most USOD methods [29, 32, 55, 63, 64, 66] focused on refining the coarse saliency cues extracted by several traditional SOD methods. For example, Zhang et al. [63] weighted these saliency cues by combining both intra-image and inter-image fusion streams. Zhang et al. [66] designed a noise modeling module to deal with noises in these saliency cues. Nguyen et al. [32] used these saliency cues as labels to train multiple deep networks. Ji et al. [21] refined the saliency maps extracted by traditional SOD methods to produce more accurate saliency predictions. Second, to prevent the localization errors caused by traditional SOD methods, Zhou et al. [72] proposed a novel framework that converts the activation maps of a pre-trained network to high-quality pseudo labels." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 624 + ], + "type": "text", + "content": "In this paper, we propose two novel mechanisms to tackle USOD tasks and explore the potential of improving the quality of pseudo labels by using multimodal data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 634, + 397, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 397, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 397, + 647 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": "In our method, we propose two novel strategies to mine accurate saliency knowledge based on the noisy activation maps generated by a deep network " + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 654, + 545, + 689 + ], + "type": "text", + "content": ", as shown in Fig. 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Activation map generation. Following previous work [72], we generate an activation map " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " for input image " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "X" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7258" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 266 + ], + "type": "image", + "image_path": "c9f3403048fd88e99f5dac79ba7f3639bd8d12656dd8267e399a241f37b53610.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed unsupervised framework. The network are guided by two mechanisms: (1) Conidence-aware Saliency Distilling mines rich and accurate saliency knowledge from noisy labels, and (2) Boundary-aware Texture Matching makes the boundary of the predicted saliency map more accurate. We pad some constant maps for the missing modalities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 315, + 187, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 187, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 187, + 327 + ], + "type": "text", + "content": "using a deep network " + }, + { + "bbox": [ + 47, + 315, + 187, + 327 + ], + "type": "inline_equation", + "content": "\\Phi(X) = Y" + }, + { + "bbox": [ + 47, + 315, + 187, + 327 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 337, + 190, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 337, + 190, + 350 + ], + "spans": [ + { + "bbox": [ + 72, + 337, + 190, + 350 + ], + "type": "interline_equation", + "content": "E _ {3}, E _ {4}, E _ {5} = \\operatorname {E n c o d e r} (X),", + "image_path": "0f934af806a2041b67956c2d7179aaa4d07853009db8f1393d2a9f310ce52031.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 352, + 287, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 352, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 108, + 352, + 287, + 371 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {i} = S E \\left(E _ {i}\\right), i \\in \\{3, 4, 5 \\}, \\\\ H = S E \\left(\\text {一} (E _ {i}, E _ {i}, E _ {i})\\right) \\end{array} \\tag {1}", + "image_path": "0b2688c059b07f459c5d021848d2d7bdb1375c0c6ba89138b1530729ab0d5042.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 367, + 236, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 367, + 236, + 380 + ], + "spans": [ + { + "bbox": [ + 109, + 367, + 236, + 380 + ], + "type": "interline_equation", + "content": "H = S E (\\text {c o n c a t} (F _ {3}, F _ {4}, F _ {5})),", + "image_path": "00b45d4ef82b7221ec654314414aca7d698b0beaa4f2665d480e30641beaaed4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 381, + 261, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 381, + 261, + 396 + ], + "spans": [ + { + "bbox": [ + 109, + 381, + 261, + 396 + ], + "type": "interline_equation", + "content": "Y = \\operatorname {i n v} (\\operatorname {S i g m o i d} (\\operatorname {s u m} (H - \\bar {H}))),", + "image_path": "b82cb5f81c00d0c246e48cd624c6dc440c50aaea3044a38a1896bf465f31b30a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\bar{H}" + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "content": " is the spatial mean of " + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "content": ". Specifically, we employ the ResNet-50 [18] pre-trained by MoCo-v2 [6] as our encoder, which is trained without extra manual annotations. In our decoder, four SE blocks [20] integrate multiple encoder features into " + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "content": ". After that, we set " + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "inline_equation", + "content": "H - \\bar{H}" + }, + { + "bbox": [ + 46, + 405, + 287, + 548 + ], + "type": "text", + "content": " to ensure that (1) noisy labels are adaptive to input images when using a fixed threshold and (2) coexisting of positive and negative samples. Next, features are summed over the channel dimension to produce one-channel activation map. Finally, the Sigmoid function produces the final saliency scores, and the inv function identifies the regions with more corner pixels as background." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": "Training strategy. The activation maps produced by network " + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": " perceive some discriminative regions in input images [72], but still are low-quality because of widespread noises. To improve the quality, we employ three strategies to train the network " + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 549, + 287, + 609 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 617, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 617, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 52, + 617, + 287, + 632 + ], + "type": "interline_equation", + "content": "L _ {\\Phi} = \\lambda_ {c} \\left(L _ {c s d} + \\hat {L} _ {c s d}\\right) + \\lambda_ {b} \\left(L _ {b t m} + \\hat {L} _ {b t m}\\right) + \\lambda_ {m} L _ {m s}, \\tag {2}", + "image_path": "924849cd142db75a6521824a70bce9d74bfe387c430a8d58424f441af4170ca8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "where all " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " are hyperparameters. " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " is the loss for the predictions of a different scale. First, we propose a Confidence-aware Saliency Distilling (CSD) scheme to excavate valuable saliency knowledge from simple examples to more complex ones. Second, we propose a Boundary-aware Texture Matching (BTM) strategy to align the boundaries in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 315, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 352 + ], + "type": "text", + "content": "appearance and the predicted saliency maps. In addition, a multi-scale consistency loss " + }, + { + "bbox": [ + 305, + 315, + 545, + 352 + ], + "type": "inline_equation", + "content": "L_{ms}" + }, + { + "bbox": [ + 305, + 315, + 545, + 352 + ], + "type": "text", + "content": " ensures our method to produce consistent predictions for multi-scale inputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 359, + 500, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 500, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 500, + 373 + ], + "type": "text", + "content": "3.1. Confidence-aware Saliency Distilling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 533 + ], + "type": "text", + "content": "The primary challenge for USOD methods is how to localize salient regions. Existing DL-based USOD methods [32,63,72] are based on noisy saliency cues from traditional SOD methods [22,26,76,77] or pre-trained networks [6,18]. Following [72], we employ the activation maps produced by a pre-trained network as our saliency cues instead of the saliency predictions of traditional SOD methods. The main reasons are two-fold: (1) traditional methods require extra computation loads; (2) network can focus on mining saliency knowledge instead of fitting the inductive bias of traditional methods. We binarize the saliency cues as the initial labels using a fixed threshold of 0.5, while Eqn. 1 ensures that these labels remain adaptive to the input image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 712 + ], + "type": "text", + "content": "To generate high-quality pseudo labels, USPS [32] uses these initial labels to train deep networks like other fully-supervised methods. However, the simple training strategy in fully-supervised methods is suboptimal for unsupervised methods. Specifically, we define pixels with saliency scores close to 0.5 as hard examples because of their low confidences. In the noisy labels, hard examples are likely to be wrongly labeled. Therefore, it is difficult to learn robust saliency knowledge from these hard samples using traditional loss functions (e.g., BCE loss). To this end, A2S [72] learns reliable saliency knowledge from easy samples. However, for hard examples, the saliency knowledge hidden in noises is not fully explored. In summary, a customized strategy that organizes the samples in a more meaningful order is crucial for USOD methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7259" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 293, + 148 + ], + "type": "image", + "image_path": "d5307d04d37199010b5171636846edb834e620c762c96ef3e77705d129c57532.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 145, + 184, + 154 + ], + "lines": [ + { + "bbox": [ + 167, + 145, + 184, + 154 + ], + "spans": [ + { + "bbox": [ + 167, + 145, + 184, + 154 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 167, + 145, + 184, + 154 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 71, + 532, + 148 + ], + "blocks": [ + { + "bbox": [ + 298, + 71, + 532, + 148 + ], + "lines": [ + { + "bbox": [ + 298, + 71, + 532, + 148 + ], + "spans": [ + { + "bbox": [ + 298, + 71, + 532, + 148 + ], + "type": "image", + "image_path": "45f3d08fa0bdf4f49ca0d8b86556e7ad8757057b1f3417f4417299ff59654473.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 145, + 426, + 154 + ], + "lines": [ + { + "bbox": [ + 405, + 145, + 426, + 154 + ], + "spans": [ + { + "bbox": [ + 405, + 145, + 426, + 154 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 405, + 145, + 426, + 154 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{\\mathrm{bce}}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 156, + 294, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 156, + 294, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 156, + 294, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 156, + 294, + 234 + ], + "type": "image", + "image_path": "952434bf3adb66ac367a13640254c1e139d3a85115fd9e637c9459873ab3dc4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 233, + 188, + 241 + ], + "lines": [ + { + "bbox": [ + 167, + 233, + 188, + 241 + ], + "spans": [ + { + "bbox": [ + 167, + 233, + 188, + 241 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 167, + 233, + 188, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{L_{adb}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 244, + 407, + 255 + ], + "lines": [ + { + "bbox": [ + 185, + 244, + 407, + 255 + ], + "spans": [ + { + "bbox": [ + 185, + 244, + 407, + 255 + ], + "type": "text", + "content": "Figure 3. The gradient landscapes of different loss functions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 156, + 532, + 234 + ], + "blocks": [ + { + "bbox": [ + 298, + 156, + 532, + 234 + ], + "lines": [ + { + "bbox": [ + 298, + 156, + 532, + 234 + ], + "spans": [ + { + "bbox": [ + 298, + 156, + 532, + 234 + ], + "type": "image", + "image_path": "c51815631348ac1e794cfc859084468d3de4b8c23d62a5131b7184473e736585.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 232, + 436, + 241 + ], + "lines": [ + { + "bbox": [ + 396, + 232, + 436, + 241 + ], + "spans": [ + { + "bbox": [ + 396, + 232, + 436, + 241 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 396, + 232, + 436, + 241 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_{\\mathrm{csd}}" + }, + { + "bbox": [ + 396, + 232, + 436, + 241 + ], + "type": "text", + "content": " (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "type": "text", + "content": "To address the above problem, we propose a Confidence-aware Saliency Distillation (CSD) scheme that scores samples with noisy labels conditional on their confidence and training progress. Concretely, easy samples contain reliable knowledge, and thus can assist our method to learn reliable saliency knowledge. On the contrary, the saliency knowledge in hard samples is hidden within noises, and may corrupt the fragile saliency patterns in the early stages of network training. Inspired by self-paced learning [2], we dynamically adjust the gradients for samples by introducing a factor " + }, + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "type": "text", + "content": ", which is linearly increasing from 0 to 1 as training proceeding. Our " + }, + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 269, + 290, + 413 + ], + "type": "text", + "content": " loss can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 419, + 287, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 419, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 91, + 419, + 287, + 452 + ], + "type": "interline_equation", + "content": "L _ {c s d} = - \\frac {1}{N} \\sum_ {i} ^ {N} | \\Phi (\\mathbf {x} _ {i}) - 0. 5 | ^ {(2 ^ {1 - \\rho})}, \\tag {3}", + "image_path": "3ad53b23b434876b9f0a1325b0d2897cfce63566bbc8aa6584bc3e34d5af3145.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " is the number of pixels, and " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " is the saliency prediction of pixel " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": ". Notice that noisy label " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " is omitted in Eqn. 3 because it is generated based on " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": ". Thus, we can calculate the confidence score " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "|k_{i}|" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " of pixel " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " in a simpler way as: " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "|k_{i}| = |\\Phi (\\mathbf{x}_{i}) - 0.5|" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": ", where the constant value 0.5 is related to " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\bar{H}" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " in Eqn. 1. The partial derivative of our " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\Phi (\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 459, + 287, + 544 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 84, + 550, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 550, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 84, + 550, + 287, + 574 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L _ {c s d}}{\\partial \\Phi (\\mathbf {x} _ {i})} = - \\operatorname {s i g n} \\left(k _ {i}\\right) 2 ^ {(1 - \\rho)} \\left| k _ {i} \\right| ^ {2 ^ {(1 - \\rho)} - 1}, \\tag {4}", + "image_path": "d6b1745940a70100b81078a5b7d3043651f9142db22d5b8c259422aec107897a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\text{sign}(k_i) \\in \\{-1, 1\\}" + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": " for negative and positive values, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "For an intuitive comparison, we draw the gradient landscapes of different losses in Fig. 3, including " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{bce}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{adb}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " [72] and our " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ". The gradient of other losses are consistent throughout the training process, while our " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " varies with the training process. Specifically, in the beginning, our " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " assigns low gradients to hard samples to learn reliable saliency knowledge from easy samples. As training proceeding, the gradients of hard samples are increasing to mine more valuable saliency knowledge." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 267, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 267, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 267, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 547, + 376 + ], + "type": "image", + "image_path": "3aa6303ecb981a1b07e5e5d25e23d2a66ed516e7077aa2f6638749a8c102f825.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 380, + 525, + 391 + ], + "lines": [ + { + "bbox": [ + 326, + 380, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 326, + 380, + 525, + 391 + ], + "type": "text", + "content": "Figure 4. Appearance information in multimodal data." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 410, + 495, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 495, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 495, + 423 + ], + "type": "text", + "content": "3.2. Boundary-aware Texture Matching" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 430, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 537 + ], + "type": "text", + "content": "In general, the appearance of saliency boundary has a similar texture as in saliency prediction. Therefore, matching the textures between different maps can guide our method to produce reasonable saliency scores. This strategy is also applicable to other modalities, such as depth map, thermal image, and optical flow, as shown in Fig. 4. By aggregating the rich appearance information in multimodal data, we can provide a more generalized guidance for our USOD method." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": "Based on the above analysis, we propose a Boundary-aware Texture Matching (BTM) strategy to align the saliency boundaries and image edges by matching their textures. First of all, similar to the LBP feature [33], we extract texture vectors from saliency predictions and appearance, respectively. For saliency predictions, the texture vector " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "T_{i}^{s} = [t_{i,1}^{s}, t_{i,2}^{s}, \\dots, t_{i,k^{2}}^{s}]" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": "-th pixel is computed by " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "t_{i,j}^{s} = |\\Phi(\\mathbf{x}_{i}) - \\Phi(\\mathbf{x}_{j})|, j \\in K_{i}" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": " neighborhoods around pixel " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i}" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": ". For appearance information, it may be one of RGB image, optical flow, depth map or thermal image. To extract more distinctive features from multimodal data, we formulate the texture vector " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "T_{i}^{a} = [t_{i,1}^{a}, t_{i,2}^{a}, \\dots, t_{i,k^{2}}^{a}]" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "t_{i,j}^{a} = \\exp(-\\alpha \\sum_{m} \\| \\mathbf{x}_{i}^{(m)} - \\mathbf{x}_{j}^{(m)} \\|^{2}), j \\in K_{i}" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "text", + "content": " is a hyperparameter and " + }, + { + "bbox": [ + 304, + 540, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i}^{(m)}" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7260" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 279, + 205 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 279, + 205 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 279, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 279, + 205 + ], + "type": "image", + "image_path": "944cfc938bb9fe20d6c3cfdd5d8358203ecf4da8f25e5226c99b8a85994c9163.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 208, + 266, + 220 + ], + "lines": [ + { + "bbox": [ + 68, + 208, + 266, + 220 + ], + "spans": [ + { + "bbox": [ + 68, + 208, + 266, + 220 + ], + "type": "text", + "content": "Figure 5. Graphical illustration of the proposed " + }, + { + "bbox": [ + 68, + 208, + 266, + 220 + ], + "type": "inline_equation", + "content": "L_{btm}" + }, + { + "bbox": [ + 68, + 208, + 266, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": "means the appearance data of modal " + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": ". It is noteworthy that a small difference between two pixels produces a small element in " + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "T^s" + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": ", but a large element in " + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "T^a" + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "T_i^s \\cdot (T_i^a)^T" + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": " is defined as the matching penalty for considering pixel " + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": " as saliency boundary, contrary to similarity. Finally, the formula of our " + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "inline_equation", + "content": "L_{btm}" + }, + { + "bbox": [ + 46, + 228, + 287, + 300 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 304, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 304, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 111, + 304, + 287, + 334 + ], + "type": "interline_equation", + "content": "L _ {b t m} = \\frac {\\sum_ {i} b _ {i} T _ {i} ^ {s} \\cdot \\left(T _ {i} ^ {a}\\right) ^ {T}}{\\sum_ {i} b _ {i}}, \\tag {5}", + "image_path": "b97a458cf4f87aae5b23967a77f8b85e92655e0a3c9d567c3eb403c995989956.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 338, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 286, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 338, + 286, + 350 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 47, + 338, + 286, + 350 + ], + "type": "text", + "content": " is the binary boundary mask of saliency prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 350, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 469 + ], + "type": "text", + "content": "An intuitive example is shown in Fig. 5. For the predicted boundary pixels, we find some nearby pixels that differ significantly in saliency scores. In general, the appearances between boundary and these pixels are also different because they locate within salient objects and backgrounds, respectively. If not the case, we expect the saliency score of this boundary pixel to be close to those nearby pixels, so that the boundary will shift in the opposite direction. After an iterative training process, our method can align the predicted saliency boundaries with image edges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 476, + 181, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 181, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 181, + 489 + ], + "type": "text", + "content": "3.3. Multi-scale Consistency" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 495, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 531 + ], + "type": "text", + "content": "Salient objects are consistent in multi-scale inputs. Thus, we resize input images to a reference scale and encourage our method to produce consistent predictions by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 537, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 537, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 103, + 537, + 287, + 562 + ], + "type": "interline_equation", + "content": "L _ {m s} = \\sum_ {i} \\left| y _ {i} - r e s i z e \\left(\\hat {y} _ {i}\\right) \\right| ^ {2}, \\tag {6}", + "image_path": "58808de7a36c89356342cc6e0843ca981ecad7554d8d0b828f736f8a1c5a79fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 47, + 569, + 286, + 581 + ], + "type": "text", + "content": " are saliency predictions of different scales." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 587, + 222, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 222, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 222, + 600 + ], + "type": "text", + "content": "3.4. Training a Generalized Detector" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Following previous USOD methods [29,32,72], we generate pseudo labels " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\ddot{Y}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " based on saliency predictions " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\ddot{Y} = CRF(Y)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " and use them to train an extra saliency detector with the IOU loss. For RGB SOD, we use the same detector as the previous method [72]. For multimodal SOD tasks, we employ the MIDD [47] network as our detector. To ensure a fair comparison with existing SOD methods, we train the extra detector only using task-specific data and corresponding pseudo labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": "Implementation details. All experiments are implemented on a single GTX 1080 Ti GPU. The batch size is 8 and input images are resized to " + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "inline_equation", + "content": "320^2" + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": ". The reference scale is randomly selected from " + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "inline_equation", + "content": "(192^2, 256^2, 384^2, 448^2)" + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": ". Only horizontal flipping is employed as our data augmentation strategy. We train our method for 20 epochs using the SGD optimizer with an initial learning rate of 0.1, which is decayed linearly. " + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "inline_equation", + "content": "\\lambda_c" + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "inline_equation", + "content": "\\lambda_b" + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "inline_equation", + "content": "\\lambda_m" + }, + { + "bbox": [ + 304, + 91, + 545, + 212 + ], + "type": "text", + "content": " are set to 200, 1, 0.05 and 1, respectively. We train the extra detectors for 10 epochs using the SGD optimizer with a learning rate of 0.005." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 212, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 546, + 415 + ], + "type": "text", + "content": "Datasets. In our experiments, we follow the prevalent settings of different SOD tasks. Specifically, for RGB SOD, we use the training subsets of MSRA-B [31] or DUTS [52] to train our method, respectively. ECSSD [42], PASCALS [27], HKU-IS [25], DUTS-TE [52], DUT-O [61], and the testing subset of MSRA-B are employed for evaluation. For RGB-D SOD, we choose 2185 samples from NLPR [36] and NJUD [23] as the training set. RGBD135 [9], SIP [13] and the testing subsets of NJUD and NLPR are employed for evaluation. For RGB-T SOD, 2500 images in VT5000 [49] are for training, while VT1000 [50], VT821 [51] and the rest 2500 images in VT5000 are for testing. For video SOD, we choose the training splits of DAVIS [37] and DAVSOD [14] to train our method. Moreover, we randomly select 5 frames of each video in DAVSOD to avoid overfitting. The testing splits of DAVIS, DAVSOD, SegV2 [24] and FBMS [3] are employed for evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": "Metrics. We adopt three criteria for evaluation, including ave-" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": ", Mean Absolute Error " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " and E-Measure " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "(E_{\\xi})" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " [12]. Specifically, " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "F_{\\beta} = \\frac{(1 + \\beta^{2}) \\times \\text{Precision} \\times \\text{Recall}}{\\beta^{2} \\times \\text{Precision} + \\text{Recall}}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\beta^{2}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " is set to 0.3 [1]. The ave-" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " scores by setting the threshold as two times mean values. " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " is the absolute error between predictions and ground truth. " + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "inline_equation", + "content": "E_{\\xi}" + }, + { + "bbox": [ + 304, + 416, + 545, + 505 + ], + "type": "text", + "content": " measures the global statistics and local pixel matching information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 514, + 429, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 429, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 429, + 526 + ], + "type": "text", + "content": "4.1. Results on RGB SOD" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "As shown in Tab. 1, we compare the proposed method with fully-supervised methods, MINet [34], LDF [56], and KRN [59], weakly-supervised methods, WSSA [65], MFNet [38], and SCW [62], and unsupervised methods, SBF [63], MNL [66], USPS [32], EDNS [64], A2S [72], DCFD [29], and SelfMask [43]. We list our results with different settings: (1) saliency results without extra detector and post-processing, denoted as \"Ours" + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "inline_equation", + "content": "_s1" + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "\"; (2) our full method as \"Ours\"; (3) training our method using multimodal data, while only DUTS-TR or MSRA-B datasets with pseudo labels are employed to train the extra detector (\"Ours" + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "inline_equation", + "content": "_{mm}" + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "\")." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Either training on MSRA-B or DUTS-TR, the proposed method achieves significant improvements compared to existing USOD methods. Moreover, our unsupervised method" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7261" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 95, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Experiment results on SOD benchmarks. \"Sup.\" indicates the supervised signals used to train SOD methods. \"F\", \"W\" and \"U\" mean fully-supervised, weakly-supervised and unsupervised, respectively. Best scores are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 95, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 48, + 95, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 545, + 365 + ], + "type": "table", + "html": "
MethodsYearSup.ECSSDMSRA-BDUT-OPASCAL-SDUTS-TEHKU-IS
\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta} \\uparrow\\)\\(E_{\\xi}\\uparrow\\)\\(\\mathcal{M}\\downarrow\\)
Trained on MSRA-B
SBF [63]2017U.812.878.087.867.929.058.611.771.106.711.795.131.627.785.105.805.895.074
MNL [66]2018U.874.906.069.881.932.053.683.821.076.792.846.091---.874.932.047
USPS [32]2019U.875.903.064.896.938.042.715.839.069.770.828.107.730.840.072.880.933.043
DCFD [29]2022U.880.900.064.903.938.041.731.838.064.773.830.105.744.832.068.887.926.044
A2S [72]2022U.888.911.064.902.941.041.719.841.069.790.838.106.750.860.065.887.937.042
Ours-U.902.923.056.912.948.036.731.851.065.803.848.099.767.871.061.891.939.041
Trained on DUTS-TR
MINet [34]2020F.924.953.033.903.948.038.756.873.055.842.899.064.828.917.037.908.961.028
LDF [56]2020F.930.951.034.902.944.037.773.881.052.853.903.062.855.929.034.914.960.028
KRN [59]2021F.931.951.032.911.950.036.793.893.050.851.894.068.865.934.033.920.961.027
WSSA [65]2020W.870.917.059.869.929.049.703.845.068.785.855.096.742.869.062.860.932.047
MFNet [38]2021W.844.889.084.872.923.059.621.784.098.756.824.115.693.832.079.839.919.058
SCW [62]2021W.900.931.049.898.940.040.758.862.060.827.879.080.823.890.049.896.943.038
EDNS [64]2020U.872.906.068.880.932.051.682.821.076.801.846.097.735.847.065.874.933.046
SelfMask [43]2022U.856.920.058.844.925.050.668.815.078.774.856.087.714.848.063.819.915.053
DCFD [29]2022U.888.915.059.888.930.045.710.837.070.795.860.090.764.855.064.889.935.042
Ourss1-U.847.912.057.849.925.054.622.773.111.763.840.093.676.814.082.819.914.053
Ours-U.916.938.044.904.944.039.745.863.061.830.882.074.810.901.047.902.947.037
Oursmm-U.913.942.043.899.945.039.737.855.064.831.890.073.803.899.048.898.946.037
", + "image_path": "57609085ff481549fdbc57373181b956526948d4c36d8c291049676eff579532.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 373, + 545, + 534 + ], + "blocks": [ + { + "bbox": [ + 48, + 373, + 545, + 534 + ], + "lines": [ + { + "bbox": [ + 48, + 373, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 545, + 534 + ], + "type": "image", + "image_path": "77e557d8a7e4f54dd74aff4836cab19cf2837d1a98c239306df109f9670f068b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 538, + 436, + 550 + ], + "lines": [ + { + "bbox": [ + 155, + 538, + 436, + 550 + ], + "spans": [ + { + "bbox": [ + 155, + 538, + 436, + 550 + ], + "type": "text", + "content": "Figure 6. Visual comparison between state-of-the-art SOD methods and ours." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "text", + "content": "is competitive to recent weakly-supervised SCW [62] and fully-supervised methods KRN [59]. Furthermore, the results of “" + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{s1}" + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "text", + "content": "” prove that our method extracts precise saliency knowledge from training samples, but somehow weaken when deploying on unseen images. In addition, the similar results of “" + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{mm}" + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "text", + "content": "” and “" + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}" + }, + { + "bbox": [ + 46, + 564, + 288, + 684 + ], + "type": "text", + "content": "” imply that extra multimodal data may not bring significant improvements to RGB SOD task. As proved in [53], the large-scale DUTS-TR dataset is saturated for training saliency detectors, even in the unsupervised case." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "A qualitative comparison is illustrated in Fig. 6. Overall, our saliency predictions are much more precise than other" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 564, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 612 + ], + "type": "text", + "content": "unsupervised methods. For example, in the first example, our method and two fully-supervised methods well segment target object with precise boundaries, while other methods fail to capture the tiny salient object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 622, + 460, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 460, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 460, + 634 + ], + "type": "text", + "content": "4.2. Results on Multimodal SOD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "We conduct more experiments on multimodal tasks, including RGB-D, RGB-T and video SOD. Noted that we precompute the optical flow for each video frame as an extra modality. To verify the improvements brought by multimodal data, we build two sets for training: (1) Task-specific data (\"Ours\"). For example, for RGB-T SOD task, we only" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7262" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 83, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 503, + 82 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 503, + 82 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 503, + 82 + ], + "type": "text", + "content": "Table 2. Results on RGB-D SOD benchmarks. \"F\" and \"U\" mean fully-supervised and unsupervised, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 83, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 54, + 83, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 54, + 83, + 541, + 175 + ], + "type": "table", + "html": "
MethodsYearSup.RGBD135NJUDNLPRSIP
\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}_{\\downarrow}\\)
DSA2F [45]2021F.899.958.021.901.937.039.897.953.024---
SPNet [74]2021F.927.984.013---.903.959.021.893.931.043
CCFE [28]2022F.911.964.020.914.953.032.907.962.021.889.923.047
DSU [21]2022U.767.895.061.719.797.135.745.879.065.619.774.156
Ours-U.834.945.037.787.829.109.834.924.043.716.784.124
\\(Ours_{mm}\\)-U.877.946.029.862.908.060.852.931.034.873.925.051
", + "image_path": "bffa937a8fae0bccead981a1e8d4cf159cd75fab408909995374457df3be89c4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 198, + 544, + 291 + ], + "blocks": [ + { + "bbox": [ + 55, + 186, + 537, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 537, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 537, + 197 + ], + "type": "text", + "content": "Table 3. Results on VSOD benchmarks. \"F\", \"W\" and \"U\" mean fully-supervised, weakly-supervised and unsupervised, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 198, + 544, + 291 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 544, + 291 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 544, + 291 + ], + "type": "table", + "html": "
MethodsYearSup.DAVSODDAVISSegV2FBMS
\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M}_{\\downarrow}\\)\\(F_{\\beta \\uparrow}\\)\\(E_{\\xi \\uparrow}\\)\\(\\mathcal{M} \\downarrow\\)
PCSA [17]2020F.556.749.077.794.922.022.789.931.019.783.868.042
TENet [40]2020F.595.773.067.821.941.017---.851.915.026
STVS [4]2021F.563.764.080.812.940.022.835.950.016.821.903.042
WSVSOD [70]2021W.492.710.103.731.900.036.711.909.031.736.840.084
Ours-U.534.747.084.751.913.042.751.914.033.732.794.100
\\(Ours_{mm}\\)-U.547.762.085.756.908.037.808.927.021.795.876.060
", + "image_path": "1f67e1e26141457ff93c686f278d4ab370adc1ec197d1418800ae0d377d4dccc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 64, + 313, + 526, + 405 + ], + "blocks": [ + { + "bbox": [ + 89, + 300, + 503, + 312 + ], + "lines": [ + { + "bbox": [ + 89, + 300, + 503, + 312 + ], + "spans": [ + { + "bbox": [ + 89, + 300, + 503, + 312 + ], + "type": "text", + "content": "Table 4. Results on RGB-T SOD benchmarks. \"F\" and \"U\" mean fully-supervised and unsupervised, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 313, + 526, + 405 + ], + "lines": [ + { + "bbox": [ + 64, + 313, + 526, + 405 + ], + "spans": [ + { + "bbox": [ + 64, + 313, + 526, + 405 + ], + "type": "table", + "html": "
MethodsYearSup.VT5000VT1000VT821
Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓Fβ↑Eξ↑M↓
MIED [46]2020F.761.880.050.853.928.030.760.877.050
MIDD [47]2021F.801.899.043.882.942.027.805.898.045
APNet [75]2021F.821.918.035.885.951.021.818.912.034
CCFE [28]2022F.859.937.030.906.963.018.857.934.027
Ours-U.810.904.046.885.939.031.805.900.043
Oursmm-U.807.903.047.881.939.032.805.899.044
", + "image_path": "3d6f45fdc655a127d50a089fcdfc0ff2658dd3d311562fe39a157f5d51e18a59.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 49, + 429, + 283, + 488 + ], + "blocks": [ + { + "bbox": [ + 72, + 416, + 261, + 427 + ], + "lines": [ + { + "bbox": [ + 72, + 416, + 261, + 427 + ], + "spans": [ + { + "bbox": [ + 72, + 416, + 261, + 427 + ], + "type": "text", + "content": "Table 5. Label quality on multimodal SOD datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 429, + 283, + 488 + ], + "lines": [ + { + "bbox": [ + 49, + 429, + 283, + 488 + ], + "spans": [ + { + "bbox": [ + 49, + 429, + 283, + 488 + ], + "type": "table", + "html": "
MethodsRGBRGB-DVSODRGB-T
Fβ↑M↓Fβ↑M↓Fβ↑M↓Fβ↑M↓
Ours.917.038.746.110.557.099.932.029
Oursrgb.917.040.804.068.613.089.920.030
Oursmm.918.040.826.062.637.079.923.029
", + "image_path": "dcd14aded3894013eba04c9f67dfa0275fb9b66b0f4ea6b6b496204a463c3c2b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 502, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 539 + ], + "type": "text", + "content": "use 2500 RGB-thermal image pairs for training as prevalent RGB-T SOD methods [46,47,75]; (2) Multimodal data from the training sets of four SOD tasks (\"Ours" + }, + { + "bbox": [ + 46, + 502, + 287, + 539 + ], + "type": "inline_equation", + "content": "_{mm}" + }, + { + "bbox": [ + 46, + 502, + 287, + 539 + ], + "type": "text", + "content": "\")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 539, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 660 + ], + "type": "text", + "content": "Overall, our method reports state-of-the-art performance on all multimodal SOD tasks. Specifically, for the RGB-D SOD task in Tab. 2, our method surpasses the latest unsupervised method DSU [21]. For video and RGB-T SOD tasks in Tab. 3 and 4, to the best of our knowledge, our method is the first unsupervised method and is competitive to supervised methods. Furthermore, our method trained on multimodal data (\"Oursmm\") produce high-quality pseudo labels for all multimodal SOD datasets simultaneously and has achieved performance improvements on most tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 670, + 141, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 141, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 141, + 683 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Label quality comparison. In Tab. 5, we exhibit the scores of pseudo labels on multimodal SOD datasets. Ex" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "content": "cept for the above two training sets, we employ an additional set that collects all RGB images from four tasks to train our method, denoted as “" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{rgb}" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "content": "”. Overall, “" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{mm}" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "content": "” reports more generalized results on all datasets. Moreover, training on task-specific data (“Ours”) slightly surpasses the performance of “" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{mm}" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "content": "” on the training sets of RGB and RGB-T tasks. Since our method is trained without ground truth, fitting a hybrid dataset may cause a slight performance drop on some subsets. In addition, the comparison between “" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{rgb}" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "content": "” and “" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Ours}_{mm}" + }, + { + "bbox": [ + 304, + 418, + 545, + 622 + ], + "type": "text", + "content": "” prove that multimodal data significantly improves the quality of the generated pseudo labels. For an intuitive comparison, we show some examples in Fig. 7. Trained with task-specific data, our method precisely localizes salient regions but fails to segment the whole objects. With more training data, the network perceives fine-grained concepts of target objects, resulting in more complete segmentations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Impact of initial saliency cues. Our method is based on the activation maps of a pre-trained network [6, 18] instead of traditional methods [22,26,76] as other DL-based USOD methods [29,32,55,63,66]. We compare the performance of these two types of designs in Tab. 6. Overall, the activation map is worse before training, whereas better than traditional methods [22,26,76] after training. The activation maps are" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7263" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 72, + 288, + 187 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 187 + ], + "type": "image", + "image_path": "d682fdaa7ecd89d795a228c17a8adb0068f9c14ce1cc576ee58a8f9be5f3eb53.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 194, + 287, + 218 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 218 + ], + "type": "text", + "content": "Figure 7. Examples of the generated pseudo labels. The \"D/T/O\" means depth map, thermal image or optical flow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 244, + 286, + 320 + ], + "blocks": [ + { + "bbox": [ + 69, + 232, + 264, + 243 + ], + "lines": [ + { + "bbox": [ + 69, + 232, + 264, + 243 + ], + "spans": [ + { + "bbox": [ + 69, + 232, + 264, + 243 + ], + "type": "text", + "content": "Table 6. Comparison between different saliency cues." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 244, + 286, + 320 + ], + "lines": [ + { + "bbox": [ + 48, + 244, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 286, + 320 + ], + "type": "table", + "html": "
Saliency cuesBefore trainingAfter training
\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}\\downarrow\\)\\(F_{\\beta}\\)↑\\(E_{\\xi}\\)↑\\(\\mathcal{M}\\downarrow\\)
MC [22].810.877.145.889.914.053
DSR [26].780.867.118.886.901.059
RBD [76].803.883.109.891.912.053
Ours.451.659.353.914.940.039
", + "image_path": "e2e6a6b80621a3dfad72362ac05eeb1461a6fcca6b43e1fe724495303b1ed899.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 337, + 286, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 286, + 432 + ], + "type": "text", + "content": "dynamic during training, such that the network strengthens the learned saliency knowledge incrementally. On the contrary, using traditional methods, the network learns from fixed labels throughout the training process, and thus fits the biased knowledge in those traditional methods. In addition, traditional methods introduce extra computation loads, while the pre-trained network is indispensable for all DL-based USOD methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 434, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 469 + ], + "type": "text", + "content": "Effectiveness of loss functions. We conduct ablation studies on various combinations of loss functions in Tab. 7 and more details are introduced as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": "In ablation study A, our method receives continuous improvements by appending new losses (" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{ms}" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{btm}" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": ") or replacing the original " + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{adb}" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": " to our " + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": ". This experiment proves that these three losses can assist the network mine more detailed saliency knowledge from different perspectives. More importantly, these losses are supplementary to each other, such that their combination can collaborate to produce high-quality pseudo labels for SOD datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": "In ablation study B, our " + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "inline_equation", + "content": "L_{csd}" + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": " loss can provide precise saliency localization information and thus achieves the best performance among all competitors. Specifically, " + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "inline_equation", + "content": "L_{bce}" + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": " fail to mine detailed and reliable saliency knowledge from noisy saliency cues. " + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "inline_equation", + "content": "L_{adb}" + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": " [72] improves the robustness of the learned saliency knowledge by focusing on easy samples, and thus surpasses the " + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": " loss. " + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "inline_equation", + "content": "L_{spl}" + }, + { + "bbox": [ + 46, + 568, + 287, + 675 + ], + "type": "text", + "content": " [2] also employs a dynamic schedule, but its binary weights filter out latent saliency knowledge in hard samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "In ablation study C, we compare our " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{btm}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{lsc}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " [62] and two variants: (1) using L1 distance for the texture features in appearance space, denoted as " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{c1}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "; (2) removing the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 83, + 539, + 255 + ], + "blocks": [ + { + "bbox": [ + 346, + 71, + 504, + 80 + ], + "lines": [ + { + "bbox": [ + 346, + 71, + 504, + 80 + ], + "spans": [ + { + "bbox": [ + 346, + 71, + 504, + 80 + ], + "type": "text", + "content": "Table 7. Ablation studies on loss functions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 83, + 539, + 255 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 539, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 539, + 255 + ], + "type": "table", + "html": "
TagLossFβ↑Eξ↑M↓
A1Ladb [72] (Baseline).882.915.071
A2Ladb [72] + Lms.891.921.066
A3Lcsd + Lms.908.937.050
A4Lcsd + Lbtm + Lms (Ours).917.945.038
B1Lbce + Lbtm + Lms.558.540.208
B2L1 + Lbtm + Lms.895.928.044
B3Ladb [72] + Lbtm + Lms.903.934.042
B4Lspl [2] + Lbtm + Lms.910.938.041
B5Lcsd + Lbtm + Lms (Ours).917.945.038
C1Lcsd + Lc1 + Lms.735.819.120
C2Lcsd + Llsc [62] + Lms.906.923.052
C3Lcsd + Lc2 + Lms.907.926.051
C4Lcsd + Lbtm + Lms (Ours).917.945.038
", + "image_path": "f099e270fed56ec94845ecf1b0ab4fa63e40237951015c0c0dcfa8f87d65dac1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 278, + 534, + 318 + ], + "blocks": [ + { + "bbox": [ + 344, + 265, + 506, + 276 + ], + "lines": [ + { + "bbox": [ + 344, + 265, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 344, + 265, + 506, + 276 + ], + "type": "text", + "content": "Table 8. Label quality on DUTS-TR dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 278, + 534, + 318 + ], + "lines": [ + { + "bbox": [ + 318, + 278, + 534, + 318 + ], + "spans": [ + { + "bbox": [ + 318, + 278, + 534, + 318 + ], + "type": "table", + "html": "
Pre-trainingFβ↑Eξ↑M↓
Supervised.915.942.039
Unsupervised (Ours).917.945.038
", + "image_path": "167caabc64bf43724b073b64d8a7eae12805f26f2e68b28ef0a3f5b41733e96a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": "boundary masks in " + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "inline_equation", + "content": "L_{btm}" + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "inline_equation", + "content": "L_{c2}" + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": ". In general, our " + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "inline_equation", + "content": "L_{btm}" + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": " outperforms other variants. Specifically, the L1 distance in " + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "inline_equation", + "content": "L_{c1}" + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": " causes the texture features to be not distinctive enough, resulting in ambiguous saliency boundaries. " + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "inline_equation", + "content": "L_{lsc}" + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": " focuses more on the adjacent pixels so that the boundaries are more susceptible to a limited region rather than a larger patch. Moreover, using " + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "inline_equation", + "content": "L_{c2}" + }, + { + "bbox": [ + 304, + 335, + 545, + 432 + ], + "type": "text", + "content": ", edges within objects or backgrounds may corrupt the learned saliency knowledge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 432, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 552 + ], + "type": "text", + "content": "Supervised or unsupervised pre-training? Under the unsupervised setting, whether we can use a encoder with supervised pre-training is controversial. For the ImageNet dataset [10], the class labels of images usually indicate the category of the most salient object. It means that the supervised encoder receives extra saliency knowledge from these manual labels. Thus, for completely unsupervised SOD, we employ unsupervised MoCo-v2 to initialize our encoder. As listed in Tab. 8, the performance of our method when using supervised encoder is comparable." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose an Unsupervised Salient Object Detection (USOD) method guided by two novel mechanisms. First, we propose a Confidence-aware Saliency Distilling (CSD) to learn saliency knowledge from easy samples to hard ones with noisy labels progressively. Second, we propose a Boundary-aware Texture Matching (BTM) to make the location of saliency prediction more accurate. As a result, the proposed method produces high-quality pseudo labels to train saliency detectors. Experiments on RGB, RGB-D, RGB-T, and video SOD benchmarks prove that our method outperforms existing USOD methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7264" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "text", + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 223 + ], + "type": "text", + "content": "[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295. Springer, 2010. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 278 + ], + "type": "text", + "content": "[4] Chenglizhao Chen, Guotao Wang, Chong Peng, Yuming Fang, Dingwen Zhang, and Hong Qin. Exploring rich and efficient spatial temporal interactions for real-time video salient object detection. IEEE Transactions on Image Processing, 30:3995-4007, 2021. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "text", + "content": "[5] Peijia Chen, Jianhuang Lai, Guangcong Wang, and Huajun Zhou. Confidence-guided adaptive gate and dual differential enhancement for video salient object detection. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "text", + "content": "[7] Zixuan Chen, Huajun Zhou, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. Contour-aware loss: Boundary-aware learning for salient object segmentation. IEEE Transactions on Image Processing, 30:431-443, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[8] Xiaolong Cheng, Xuan Zheng, Jialun Pei, He Tang, Zehua Lyu, and Chuanbo Chen. Depth-induced gap-reducing network for rgb-d salient object detection: An interaction, guidance and refinement approach. IEEE Transactions on Multimedia, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 513 + ], + "type": "text", + "content": "[9] Yupeng Cheng, Huazhu Fu, Xingxing Wei, Jiangjian Xiao, and Xiaochun Cao. Depth enhanced saliency detection method. In Proceedings of international conference on internet multimedia computing and service, pages 23-27, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[11] Wenhui Dao, Xian Sun, Xinwei Zheng, Fangzheng Dou, Hongqi Wang, and Kun Fu. Efficient saliency-based object detection in remote sensing images using deep belief networks. IEEE Geoscience and Remote Sensing Letters, 13(2):137-141, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 677 + ], + "type": "text", + "content": "[12] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In International Joint Conference on Artificial Intelligence (IJCAI), pages 698-704, 2018. http://dpfan.net/e-measure/.5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Deng-Ping Fan, Zheng Lin, Zhao Zhang, Menglong Zhu, and Ming-Ming Cheng. Rethinking rgb-d salient object detection: Models, data sets, and large-scale benchmarks." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE Transactions on neural networks and learning systems, 32(5):2075-2089, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 151 + ], + "type": "text", + "content": "[14] Deng-Ping Fan, Wenguan Wang, Ming-Ming Cheng, and Jianbing Shen. Shifting more attention to video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8554-8564, 2019. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 153, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 198 + ], + "type": "text", + "content": "[15] Mengyang Feng, Hutchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1623-1632, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 199, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 243 + ], + "type": "text", + "content": "[16] Carola Figueroa Flores, Abel Gonzalez-Garcia, Joost van de Weijer, and Bogdan Raducanu. Saliency for fine-grained object recognition in domains with scarce training data. Pattern Recognition, 94:62-73, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 245, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 300 + ], + "type": "text", + "content": "[17] Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming-Ming Cheng, and Shao-Ping Lu. Pyramid constrained self-attention network for fast video salient object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 10869–10876, 2020. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 302, + 545, + 346 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 348, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 403 + ], + "type": "text", + "content": "[19] Qibin Hou, Ming-Ming Cheng, Xiaowei Hu, Ali Borji, Zhuowen Tu, and Philip HS Torr. Deeply supervised salient object detection with short connections. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3203-3212, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 405, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 438 + ], + "type": "text", + "content": "[20] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 440, + 545, + 483 + ], + "type": "text", + "content": "[21] Wei Ji, Jingjing Li, Qi Bi, Chuan Guo, Jie Liu, and Li Cheng. Promoting saliency from depth: Deep unsupervised rgb-d saliency detection. arXiv preprint arXiv:2205.07179, 2022. 1, 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 485, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 485, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 485, + 545, + 529 + ], + "type": "text", + "content": "[22] Bowen Jiang, Lihe Zhang, Huchuan Lu, Chuan Yang, and Ming-Hsuan Yang. Saliency detection via absorbing markov chain. In Proceedings of the IEEE international conference on computer vision, pages 1665-1672, 2013. 2, 3, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 532, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 585 + ], + "type": "text", + "content": "[23] Ran Ju, Ling Ge, Wenjing Geng, Tongwei Ren, and Gangshan Wu. Depth saliency based on anisotropic center-surround difference. In 2014 IEEE international conference on image processing (ICIP), pages 1115-1119. IEEE, 2014. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 632 + ], + "type": "text", + "content": "[24] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 656 + ], + "type": "text", + "content": "[25] Guanbin Li and Yizhou Yu. Visual saliency based on multiscale deep features. In CVPR, 2015. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] Xiaohui Li, Huchuan Lu, Lihe Zhang, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via dense and sparse reconstruction. In Proceedings of the IEEE international conference on computer vision, pages 2976-2983, 2013. 2, 3, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7265" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[27] Yin Li, Xiaodi Hou, Christof Koch, James M. Rehg, and Alan L. Yuille. The secrets of salient object segmentation. In CVPR, 2014. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[28] Guibiao Liao, Wei Gao, Ge Li, Junle Wang, and Sam Kwong. Cross-collaborative fusion-encoder network for robust rgb-thermal salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[29] Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causal debiasing framework for unsupervised salient object detection. In Thirty-sixth AAAI conference on artificial intelligence, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 242 + ], + "type": "text", + "content": "[30] Nian Liu, Junwei Han, and Ming-Hsuan Yang. Picanet: Learning pixel-wise contextual attention for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3089-3098, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 286 + ], + "type": "text", + "content": "[31] Tie Liu, Zejian Yuan, Jian Sun, Jingdong Wang, Nanning Zheng, Xiaou Tang, and Heung-Yeung Shum. Learning to detect a salient object. IEEE Transactions on Pattern analysis and machine intelligence, 33(2):353-367, 2010. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 288, + 353 + ], + "type": "text", + "content": "[32] Duc Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Thi Phuong Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction with self-supervision. arXiv preprint arXiv:1909.13055, 2019. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[33] Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on pattern analysis and machine intelligence, 24(7):971-987, 2002. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "text", + "content": "[34] Youwei Pang, Xiaoji Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9413-9422, 2020. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 511 + ], + "type": "text", + "content": "[35] Prashant W Patil, Subrahmanyam Murala, Abhinav Dhall, and Sachin Chaudhary. Msednet: multi-scale deep saliency learning for moving object detection. In 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 1670-1675. IEEE, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "type": "text", + "content": "[36] Houwen Peng, Bing Li, Weihua Xiong, Weiming Hu, and Rongrong Ji. Rgbd salient object detection: A benchmark and algorithms. In European conference on computer vision, pages 92-109. Springer, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 621 + ], + "type": "text", + "content": "[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[38] Yongri Piao, Jian Wang, Miao Zhang, and Huchuan Lu. Mfnet: Multi-filter directive network for weakly supervised salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4136-4145, 2021. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[40] Sucheng Ren, Chu Han, Xin Yang, Guoqiang Han, and Shengfeng He. Tenet: Triple excitation network for video salient object detection. In European Conference on Computer Vision, pages 212-228. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[42] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4):717-729, 2015. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 296 + ], + "type": "text", + "content": "[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 331 + ], + "type": "text", + "content": "[44] Ali Shokoufandeh, Ivan Marsic, and Sven J Dickinson. View-based object recognition using saliency maps. Image and Vision Computing, 17(5-6):445-460, 1999. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 332, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 388 + ], + "type": "text", + "content": "[45] Peng Sun, Wenhu Zhang, Huanyu Wang, Songyuan Li, and Xi Li. Deep rgb-d saliency detection with depth-sensitive attention and automatic multi-modal fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1407-1417, 2021. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[46] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive encoder-decoder network for rgbt salient object detection. arXiv e-prints, pages arXiv-2005, 2020. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[47] Zhengzheng Tu, Zhun Li, Chenglong Li, Yang Lang, and Jin Tang. Multi-interactive dual-decoder for rgb-thermal salient object detection. IEEE Transactions on Image Processing, 30:5678–5691, 2021. 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[48] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[49] Zhengzheng Tu, Yan Ma, Zhun Li, Chenglong Li, Jieming Xu, and Yongtao Liu. Rgbt salient object detection: A largescale dataset and benchmark. IEEE Transactions on Multimedia, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "text", + "content": "[50] Zhengzheng Tu, Tian Xia, Chenglong Li, Xiaoxiao Wang, Yan Ma, and Jin Tang. Rgb-t image saliency detection via collaborative graph learning. IEEE Transactions on Multimedia, 22(1):160-173, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 669 + ], + "type": "text", + "content": "[51] Guizhao Wang, Chenglong Li, Yunpeng Ma, Aihua Zheng, Jin Tang, and Bin Luo. RGB-t saliency detection benchmark: Dataset, baselines, analysis and a novel approach. In Chinese Conference on Image and Graphics Technologies, pages 359-369. Springer, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[52] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In CVPR, 2017. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7266" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. arXiv preprint arXiv:1904.09146, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[54] Wenguan Wang, Shuyang Zhao, Jianbing Shen, Steven CH Hoi, and Ali Borji. Salient object detection with pyramid attention and salient edges. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1448-1457, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[55] Yifan Wang, Wenbo Zhang, Lijun Wang, Ting Liu, and Huchuan Lu. Multi-source uncertainty mining for deep unsupervised saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11727-11736, 2022. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 284 + ], + "type": "text", + "content": "[56] Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13025-13034, 2020. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "text", + "content": "[57] Hongfa Wen, Chenggang Yan, Xiaofei Zhou, Runmin Cong, Yaoqi Sun, Bolun Zheng, Jiyong Zhang, Yongjun Bao, and Guiguang Ding. Dynamic selective network for rgb-d salient object detection. IEEE Transactions on Image Processing, 30:9179-9192, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "text", + "content": "[58] Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3907-3916, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "text", + "content": "[59] Binwei Xu, Haoran Liang, Ronghua Liang, and Peng Chen. Locate globally, segment locally: A progressive architecture with knowledge review network for salient object detection. In Proceedings. of the AAAI Conference On Artificial Intelligence, pages 1-9, 2021. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "text", + "content": "[60] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 521 + ], + "type": "text", + "content": "[61] Chuan Yang, Lihe Zhang, Hutchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[62] Siyue Yu, Bingfeng Zhang, Jimin Xiao, and Eng Gee Lim. Structure-consistent weakly supervised salient object detection with local saliency coherence. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). AAAI Palo Alto, CA, USA, 2021. 2, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[63] Dingwen Zhang, Junwei Han, and Yu Zhang. Supervision by fusion: Towards unsupervised learning of deep salient object detector. In Proceedings of the IEEE International Conference on Computer Vision, pages 4048-4056, 2017. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[64] Jing Zhang, Jianwen Xie, and Nick Barnes. Learning noise-aware encoder-decoder from noisy labels by alternating back-propagation for saliency detection. In European conference on computer vision, pages 349-366. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[65] Jing Zhang, Xin Yu, Aixuan Li, Peipei Song, Bowen Liu, and Yuchao Dai. Weakly-supervised salient object detection via" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "scribble annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12546-12555, 2020. 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[66] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9029-9038, 2018. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[67] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Xiang Ruan. Amulet: Aggregating multi-level convolutional features for salient object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 202-211, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[68] Pingping Zhang, Dong Wang, Huchuan Lu, Hongyu Wang, and Baocai Yin. Learning uncertain convolutional features for accurate saliency detection. In ICCV, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[69] Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 714-722, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "text", + "content": "[70] Wangbo Zhao, Jing Zhang, Long Li, Nick Barnes, Nian Liu, and Junwei Han. Weakly supervised video salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16826-16835, 2021. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 363, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 407 + ], + "type": "text", + "content": "[71] Xiaoqi Zhao, Youwei Pang, Lihe Zhang, Huchuan Lu, and Lei Zhang. Suppress and balance: A simple gated network for salient object detection. In European Conference on Computer Vision, pages 35-51. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 408, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 461 + ], + "type": "text", + "content": "[72] Huajun Zhou, Peijia Chen, Lingxiao Yang, Xiaohua Xie, and Jianhuang Lai. Activation to saliency: Forming high-quality labels for unsupervised salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 463, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 518 + ], + "type": "text", + "content": "[73] Huajun Zhou, Xiaohua Xie, Jian-Huang Lai, Zixuan Chen, and Lingxiao Yang. Interactive two-stream decoder for accurate and fast saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9141-9150, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "text", + "content": "[74] Tao Zhou, Huazhu Fu, Geng Chen, Yi Zhou, Deng-Ping Fan, and Ling Shao. Specificity-preserving rgb-d saliency detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4681-4691, 2021. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 564, + 545, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 618 + ], + "type": "text", + "content": "[75] Wujie Zhou, Yun Zhu, Jingsheng Lei, Jian Wan, and Lu Yu. Apnet: Adversarial learning assistance and perceived importance fusion network for all-day rgb-t salient object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 619, + 545, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 619, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 307, + 619, + 545, + 663 + ], + "type": "text", + "content": "[76] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 2, 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 664, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 664, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 307, + 664, + 545, + 708 + ], + "type": "text", + "content": "[77] Wenbin Zou and Nikos Komodakis. Harf: Hierarchy-associated rich features for salient object detection. In Proceedings of the IEEE international conference on computer vision, pages 406-414, 2015. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7267" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_content_list.json b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aadc3baf62b97258aa512d008b16efbb42c5b388 --- /dev/null +++ b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_content_list.json @@ -0,0 +1,1479 @@ +[ + { + "type": "text", + "text": "The Best Defense is a Good Offense: Adversarial Augmentation against Adversarial Attacks", + "text_level": 1, + "bbox": [ + 125, + 130, + 843, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Iuri Frosio, NVIDIA", + "bbox": [ + 285, + 204, + 452, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ifrosio@nvidia.com", + "bbox": [ + 289, + 224, + 452, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jan Kautz, NVIDIA", + "bbox": [ + 521, + 204, + 681, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jkautz@nvidia.com", + "bbox": [ + 524, + 224, + 676, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 312, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many defenses against adversarial attacks (e.g. robust classifiers, randomization, or image purification) use countermeasures put to work only after the attack has been crafted. We adopt a different perspective to introduce $A^5$ (Adversarial Augmentation Against Adversarial Attacks), a novel framework including the first certified preemptive defense against adversarial attacks. The main idea is to craft a defensive perturbation to guarantee that any attack (up to a given magnitude) towards the input in hand will fail. To this aim, we leverage existing automatic perturbation analysis tools for neural networks. We study the conditions to apply $A^5$ effectively, analyze the importance of the robustness of the to-be-defended classifier, and inspect the appearance of the robustified images. We show effective on-the-fly defensive augmentation with a robustifier network that ignores the ground truth label, and demonstrate the benefits of robustifier and classifier co-training. In our tests, $A^5$ consistently beats state of the art certified defenses on MNIST, CIFAR10, FashionMNIST and Tinyimagenet. We also show how to apply $A^5$ to create certifiably robust physical objects. Our code at https://github.com/NVlabs/A5 allows experimenting on a wide range of scenarios beyond the man-in-the-middle attack tested here, including the case of physical attacks.", + "bbox": [ + 75, + 305, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 694, + 207, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since Deep Neural Networks (DNNs) have been found vulnerable to adversarial attacks [11, 30], researchers studied various protection strategies [4, 12, 20, 42, 43]. For instance, adversarial training [11, 30] generates attacks while asking a DNN for the correct output in training; it is simple, partially effective and widely adopted. Certified methods (e.g., IBP [12], CROWN [43], CROWN-IBP [42]) do a step more by estimating correct (although often pessimistic) output bounds (Fig. 1, a) used for training. Adversarial training regularizes the classification landscape against the attacks (Fig. 1, b), but high protection often produces a loss in clean accuracy. Other partially effective defenses are based", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on randomness [35, 35] or removal of the adversarial signal [13, 18, 27, 40], by moving the input back to the space of natural (non-attacked) data before classification (Fig. 1, c).", + "bbox": [ + 498, + 275, + 892, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "All the aforementioned strategies activate the defense mechanism only after the attack has been crafted. However, when dealing with adversarial attacks, the first actor to move has a significant advantage. For instance, perturbing an image can avoid online person identification and preserve privacy [6, 19, 26]. Acting first is particularly suitable against Man in the Middle (MitM) attacks that may prac", + "bbox": [ + 498, + 323, + 892, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c048a3d55e74570bce821571e1968792f8820461b031e5f3f4a643956839de52.jpg", + "image_caption": [ + "(a) certified bounds" + ], + "image_footnote": [], + "bbox": [ + 547, + 446, + 669, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1cd31a8efbd7ed677e5f28d56f613a27f651ef891869d2a46ef985db4fb8328.jpg", + "image_caption": [ + "(b) adversarial training" + ], + "image_footnote": [], + "bbox": [ + 715, + 448, + 833, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f457675ae7502def028e1a2fefac5fd56cdc88638b2780215c892bee0d6194bb.jpg", + "image_caption": [ + "(c) image purification" + ], + "image_footnote": [], + "bbox": [ + 550, + 521, + 669, + 573 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e94738566a2130b22e9623e5135caff2f0dfb8262984d2437da2d5e8ef2b21b9.jpg", + "image_caption": [ + "(d) $A^5$" + ], + "image_footnote": [], + "bbox": [ + 717, + 521, + 834, + 573 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cbd9e9047d0c7d1cb305fdf4c976c459c03dd15c5fa9f99f8eedd445a435327.jpg", + "image_caption": [ + "Base: airplane (69.8%) \n[Verified: ship (100.0%)]", + "Figure 1. (a) An input $\\pmb{x} \\in y^*$ is correctly classified. The grey box (certified bounds [12, 42, 43]) shows that, under an attack $\\delta \\pmb{x}_A$ , $||\\delta \\pmb{x}_A||_{\\infty} < \\epsilon_A$ , misclassification $((\\pmb{x} + \\delta \\pmb{x}_A) \\in y_0)$ is possible. (b) Adversarial training [4, 20] creates robust DNNs with regular classification landscapes: misclassification is less likely. (c) Image purification [27, 40] moves the attacked input $\\pmb{x} + \\delta \\pmb{x}_A$ by $\\delta \\pmb{x}_P$ back to correct classification. (d) $A^5$ preemptively moves $\\pmb{x}$ into a non-attackable position. The original CIFAR10 image classified as airplane (69.8% confidence) can be misclassified under attack (ship, 100% confidence). Once robustified through $A^5$ (right), misclassification does not occur anymore." + ], + "image_footnote": [], + "bbox": [ + 531, + 617, + 679, + 733 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b0c10474ff1425b84189394b598963fce90bb20601108c5301961cbf6dfcfe0.jpg", + "image_caption": [ + "Robustified: airplane $(82.6\\%)$ Verified: airplane $(82.5\\%)$" + ], + "image_footnote": [], + "bbox": [ + 704, + 617, + 854, + 732 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4067", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f26b1262cb261d31b022106422cef8ec31998a1481eef1f5cf6b919c6374a667.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
vA vectorvjVector j-th element
CClassifierRRobustifier
xDataδxAAttacking perturbation
δxDDefensive perturbationx + δxAData under attack
x + δxDRobustified datax + δxD + δxARobustified data under attack
εDDefense magnitude, ||δxD||p < εDεRAttack magnitude, ||δxA||p < εR_A while training R
εAAttack magnitude ||δxA||p < εA while testingεCAttack magnitude, ||δxA||p < εC_A while training C
", + "bbox": [ + 81, + 88, + 893, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Notation for data $\\mathbf{x}$ ,that are the inputs of the classifier $C$ . We adopt an equivalent notation for physical objects $\\mathbf{w}$ .", + "bbox": [ + 119, + 210, + 846, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cally arise in automotive [32], audio processing [13, 22, 38], or while communicating with a remote machine [39]. Our idea spouses this reasoning line: we investigate a novel way to augment the data to preemptively certify that it cannot be attacked (Fig. 1, d). This fits in a recent research area that up to now has received less attention than adversarial training. Researchers explored image and real object augmentation to guarantee a high recognition rate in presence of noise or geometric distortions, but not in adversarial scenarios [24]. Encryption schemes coupled with deep learning [1, 28] or watermarking [16] only partially protect against MitM attacks. The few existing preemptive robustification methods [23] include a procedure [17] that first runs a classifier on clean data and achieves preemptive robustification against MitM attacks through iterative optimization; these are partially effective and do not provide any certification. Our novel framework encompasses most of the aforementioned cases while also introducing for the first time the concept of certified robustification. More in detail, our manifold contributions are:", + "bbox": [ + 75, + 251, + 472, + 554 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) We introduce $A^5$ (Adversarial Augmentation Against Adversarial Attacks), a comprehensive framework for preemptive augmentation to make data and physical objects certifiably robust against MitM adversarial attacks. As far as we know, this is the first time certified defense is achieved in this way. Since we provide certified robustness, we guarantee protection against any form of white, grey or black box attack.", + "(ii) We test different flavours of $A^5$ on standard datasets. By doing so, we study the connection between the robustness of the legacy classifier, the magnitude of the defensive augmentation, and the protection level delivered by $A^5$ . We show $A^5$ achieving state-of-the-art certified protection against MitM attacks by training a robustifier DNN coupled with the legacy classifier, and even better results for co-training of the robustifier and classifier. We perform a critical, visual inspection of the robustified images to answer an interesting theoretical question (how does a non-attackable image look like?) and potentially provide directions for the acquisition of inherently robust images.", + "(iii) Using Optical Character Recognition (OCR) as an ex" + ], + "bbox": [ + 76, + 568, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ample, we show the application of $A^5$ for the design of certifiably robust physical objects, which extends [24] to the case of certified defense against MitM attacks.", + "bbox": [ + 529, + 251, + 890, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(iv) We share our code at https://github.com/NVlabs/A5, to allow replicating our results or testing $A^5$ in scenarios not considered here, for instance on other datasets or for protection against physical adversarial attack (e.g., adversarial patches).", + "bbox": [ + 501, + 297, + 890, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work and Background", + "text_level": 1, + "bbox": [ + 500, + 388, + 785, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify three main philosophies for the development of defenses against adversarial attacks. They are not mutually exclusive and can be adopted together.", + "bbox": [ + 498, + 416, + 890, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first approach aims at creating DNNs that are robust to adversarial attacks: e.g., in adversarial training, attacks are generated while training the DNN that should process them correctly [4, 11, 20]. This regularizes the classification landscape and makes attacks less effective (Fig. 1, b). Adversarial training is simple but has one drawback: as attack generation is computationally demanding (even after speed up [25]), attacks are randomly sampled and cannot cover the entire input space, thus impacting the final accuracy. Certified defense [12, 42, 43] can be seen as a state of the art refined form of adversarial training: robustness is obtained by propagating through the DNN an input interval that encompasses all the attacks with magnitude up to $\\epsilon_A^C$ . The corresponding set of output bounds allows estimating the worst case scenario and derive the training cost function (e.g., worst case entropy [12, 42]). Research focuses on finding bounds that are both tight and computationally light. IBP [12] computes (and backwards) the bounds at the cost of an additional forward and backward pass. Since the bounds are loose at the beginning of training, a complex schedule is required to guarantee stability and convergence. Linear relaxation [43] delivers tighter bounds leveraging the correlation between different layers to linearly propagate signal and bounds, when possible, and eventually put in linear relation the input interval and output bounds. CROWN [43] bounds are often tighter than IBP, but at an impractical computational cost for large DNNs. Among the attempts to speed up CROWN [33, 37], CROWN-IBP [42] strikes the right balance of computa", + "bbox": [ + 496, + 463, + 892, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4068", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tional complexity and tightness. It is implemented in the auto_LiRPA $^{1}$ library [36, 37] adopted here. One limitation of adversarial training, certified methods, and $A^{5}$ as well, is the trade-off between clean and certified (worst case underattack) accuracy: increasing the latter leads to a decrease in the former [4, 12, 20, 42]. Among the architecture solutions against adversarial attacks, we also mention the recent development of $\\ell_{\\infty}$ -dist neurons that are inherently robust [41], and could be leveraged by $A^{5}$ as well in future.", + "bbox": [ + 75, + 90, + 467, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second defense philosophy is based on active countermeasures taken after the attack is crafted. Image purification [18, 27, 40] projects an attacked image onto the manifold of natural images through a generative model (Fig. 1, c). It is less effective than adversarial training [7], but does not make any assumption on the classifier. Other forms of signal pre-processing have been proposed in the field of audio processing and automatic speech recognition, where MitM physical attacks can be easily implemented [22, 38] and partially defeated by randomized smoothing, DefenseGAN, variational autoencoder (VAE), and Parallel WaveGAN vocoder [13]. Both certified defenses and pre-processing methods do not assume any specific attack form, but the latter are less effective in white box scenarios [2, 31]. Randomization can also be used after the attack, e.g. by changing resolution or padding an image [35], or by picking random experts in a mixture, that is proved to be more robust than deterministic DNNs [21].", + "bbox": [ + 75, + 226, + 467, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$A^5$ belongs to a third class of methods, that put in place countermeasures before the attack (Fig. 1, d). The roots of $A^5$ lie in two ideas that have been only partially investigated so far. The first is that, in adversarial scenarios, the first actor to move has a significant advantage. This rule is already leveraged for online privacy, where a preemptive adversarial attack prevents the identification of the framed subject [6, 19]. Since adversarial attacks often generalize to many classifiers, this ploy is particularly effective [26]. Preemptive protection is also a natural form of defense against MitM attacks that can be easily created in contexts like automotive [32], speech and audio processing [13, 22, 38], or anytime the capturing device communicates with a server [39]. Encryption [1, 28] and watermarking [16] coupled with deep learning work against MitM attacks, but do not achieve full protection, are not certified, require the key to be hidden (grey box scenario) and may significantly alter the appearance of the image. Blindpreprocessing [23] applies a tanh transform, normalization, quantization and thermometer encoding to achieve partial and uncertified defense. The closest algorithm to $A^5$ is preemptive robustification [17]: given an image $\\pmb{x}$ and a classifier $C$ that can be fine-tuned, it computes $y_0 = C(\\pmb{x})$ and randomly samples attacks $\\delta \\pmb{x}_A$ against $\\pmb{x} + \\delta \\pmb{x}_D$ to find a defensive perturbation $\\delta \\pmb{x}_D$ , such that $(\\pmb{x} + \\delta \\pmb{x}_D + \\delta \\pmb{x}_A) \\in y_0$", + "bbox": [ + 75, + 498, + 467, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(correct classification under attack). Since $A^5$ uses certified bounds, it results in simpler training and higher protection, and does not require running $C$ first, as our robustifier DNN generalizes to all the inputs. The second idea behind $A^5$ comes from physical adversarial examples, i.e., 3d physical objects [3] or 2d patches [5] that consistently fool a DNN under a wide range of viewing conditions. Reverting the attacker and defender roles, the complementary cost function leads to the creation of unadversarial examples: patches or textures that, once applied to a real 3d object, reinforce (instead of degrading) the DNN desired behaviour [24]. However, these are not designed to be inherently robust to adversarial attacks. We exploit the idea in $A^5$ in all those situations where system designers control, to some extent, the inputs fed to the DNN (e.g. in OCR, robotic systems or during the design of the road infrastructure). We show as an example the use of $A^5$ to design certifiably protected characters for OCR, something unexplored both in [17] and [24].", + "bbox": [ + 500, + 90, + 890, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 376, + 589, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Threat models We summarize our notation in Table 1, whereas Fig. 2 illustrates the full $A^5$ framework including trainable DNNs in blue, training algorithms in green, physical objects $\\boldsymbol{w}$ , and acquired data $\\boldsymbol{x}$ . MitM attacks are crafted while transmitting data from the acquisition device to the classifier $C$ , while physical attacks (not tested here) are crafted onto physical objects $\\boldsymbol{w}$ , before data acquisition. $A^5$ can be used in a scenario where the attacker has full access to the $C$ : since it uses certified bounds, it is agnostic to the specific nature of the attack. In other words, we certify protection for any $\\delta x_A$ , $||\\delta x_A|| < \\epsilon_A$ and the attacker can use any white-, grey- or black-box attacking algorithm. We assume that $A^5$ can run in a protected environment, not accessible to the attacker, to perform robustification (light blue rectangles in Fig. 2). More details and limitations of our threat model are discussed in the Supplementary, whereas here we first introduce useful definitions and then provide several recipes for the use of $A^5$ in different scenarios.", + "bbox": [ + 498, + 402, + 890, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definitions For a classifier $C$ and input $\\mathbf{x}$ , $\\mathbf{y} = C(\\mathbf{x})$ is the logit vector output by $C$ . Under attack, we have:", + "bbox": [ + 500, + 693, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} + \\delta \\boldsymbol {y} = C (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {A}), \\| \\left| \\delta \\boldsymbol {x} _ {A} \\right| | _ {p} < \\epsilon_ {A}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 736, + 890, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{A}$ is the magnitude of the attack $\\delta \\pmb{x}_A$ in $p$ -norm and $\\delta \\pmb{y}$ is the logit perturbation. Linear relaxation [43] estimates the lower $(\\pmb{y}^l)$ and upper $(\\pmb{y}^u)$ bounds of $\\pmb{y} + \\delta \\pmb{y}$ , for instance using IBP [12], CROWN [43], CROWN-IBP [42], or a combination of them. The margin for the class $j$ is the difference between the lower bound of the ground truth class $j^{*}$ and the upper bound of $j$ , while it is zero by definition for $j^{*}$ . More formally, for $M$ classes the elements of the margin vector $\\pmb{m} = [m_0, m_1, \\dots, m_{M - 1}]$ are: $m_{j^{*}} = 0$", + "bbox": [ + 498, + 763, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Automatic Linear Relaxation based Perturbation Analysis", + "bbox": [ + 94, + 886, + 406, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/735e07d238056c1a05cd65e211b875af78b4d89cc3cfb1e9ab64a214dc1f33dd.jpg", + "image_caption": [ + "Figure 2. Schematic representation of the full $A^5$ framework. It allows the robust augmentation of both physical objects $\\boldsymbol{w}$ and acquired data $\\boldsymbol{x}$ to make them certifiably robust against MitM and physical (not tested here) adversarial attacks." + ], + "image_footnote": [], + "bbox": [ + 181, + 88, + 794, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $m_j = \\pmb{y}_{j^*}^l -\\pmb{y}_j^u$ $\\forall j\\neq j^{*}$ . The confidence is obtained by applying softmax to $\\pmb{m}$ ..", + "bbox": [ + 75, + 238, + 468, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {j} (\\boldsymbol {x}) = e ^ {- m _ {j}} / \\sum_ {k = 0} ^ {M - 1} e ^ {- m _ {k}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 281, + 468, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and thus the worst case cross entropy $E(\\pmb{x})$ is:", + "bbox": [ + 76, + 335, + 383, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE (\\boldsymbol {x}) = - \\sum_ {j = 0} ^ {M - 1} \\hat {y} _ {j} \\log \\left(p _ {j} (\\boldsymbol {x})\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 363, + 468, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{y}_j$ is the one-hot-encoding of $j^*$ . To formally introduce $A^5$ , we first give the following definitions:", + "bbox": [ + 76, + 417, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 ( $\\epsilon_A$ -robustness). Given an input $\\pmb{x}$ and a classifier $C$ , we say that $\\pmb{x}$ is $\\epsilon_A$ -robust with respect to $C$ in norm $p$ if $\\forall \\delta \\pmb{x}_A, ||\\delta \\pmb{x}_A||_p < \\epsilon_A$ , $C(\\pmb{x} + \\delta \\pmb{x}_A)$ generates the correct classification output.", + "bbox": [ + 76, + 460, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 2 (robustification). For a given input $\\pmb{x}$ and classifier $C$ , robustification is the process of finding a defensive augmentation $\\delta \\pmb{x}_D$ , $||\\delta \\pmb{x}_D||_p < \\epsilon_D$ , such that $\\pmb{x} + \\delta \\pmb{x}_D$ is $\\epsilon_A$ -robust with respect to $C$ in norm $p$ .", + "bbox": [ + 76, + 532, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compute certified CROWN-IBP bounds through auto_LiRPA [37] and optimize $E(\\pmb{x} + \\delta \\pmb{x}_D)$ to find preemptive, defensive perturbations $\\delta \\pmb{x}_D$ that make $\\pmb{x} + \\delta \\pmb{x}_D \\in \\epsilon_A$ -robust, following the training recipes provided here. We also show how to find defensive perturbations $\\delta \\pmb{w}_D$ for physical objects $\\pmb{w}$ using the same approach.", + "bbox": [ + 75, + 604, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Offline robustification with ground truth $(A^5 /O)$ : Consider a legacy classifier $C$ that has been certifiably trained (e.g. through CROWN-IBP [42]) against attacks of magnitude $\\epsilon_A^C$ . Given an input $\\pmb{x}$ and its class $j^{*}$ , we want to find a defensive perturbation $\\delta \\pmb{x}_D,||\\delta \\pmb{x}_D||_p < \\epsilon_D$ , such that $\\pmb{x} + \\delta \\pmb{x}_D$ is $\\epsilon_A^R$ -robust with respect to $C$ in $p$ -norm. Our solution to this problem is $A^5$ /Offline (or $A^5 /O$ ).", + "bbox": [ + 75, + 715, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To force $||\\delta \\pmb{x}_D||_p < \\epsilon_D$ , we parameterize $\\delta \\pmb{x}_D$ via a vector $\\pmb{z}$ to which we apply an element-wise sigmoid:", + "bbox": [ + 76, + 821, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\boldsymbol {x} _ {D} = \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) = [ \\delta x _ {D} (\\boldsymbol {z}) _ {0}, \\dots , \\delta x _ {D} (\\boldsymbol {z}) _ {K - 1} ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 864, + 468, + 882 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta x _ {D} (\\boldsymbol {z}) _ {j} = 2 \\epsilon_ {D} \\left[ 1 / \\left(1 + e ^ {- z _ {j}}\\right) - 0. 5 \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of elements in $\\pmb{x}$ . This is a robust classification problem where $\\pmb{z}$ is the only unknown. For any $\\pmb{x}$ , we minimize the worst case cross entropy:", + "bbox": [ + 500, + 238, + 890, + 285 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} = \\operatorname {a r g m i n} _ {\\tilde {\\boldsymbol {z}}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\tilde {\\boldsymbol {z}}) ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 297, + 890, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using RMSProp. Notice that the magnitude of the attack used during robustification $(\\epsilon_A^R)$ and that adopted to train $C$ $(\\epsilon_A^C)$ can be different. Furthermore, similarly to adversarial training and certified methods, these can also be different from the target attack magnitude, $\\epsilon_{A}$ .", + "bbox": [ + 498, + 325, + 890, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$A^5 / O$ is similar in spirit to preemptive image robustification [17], but does not need $C$ to run on the clean input before robustification. Although $A^5 / O$ could be used in a similar way to robustify $\\pmb{x}$ accordingly to the class $C(\\pmb{x})$ , running $C$ first and then iteratively solving the optimization problem in Eq. 6 makes it hardly usable in practice. Therefore, it does not directly serve any practical purpose, but it helps answering important theoretical questions (e.g., under which conditions can we robustify $\\pmb{x}$ ? What does a robustified image looks like?) and establishing a baseline with known ground truth and no computational constraints.", + "bbox": [ + 498, + 402, + 892, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Online robustification with robustifier $R$ , legacy $C$ ( $A^5 / R$ ): We introduce a second recipe, $A^5 / R$ , where we train a robustifier DNN $R$ that takes in input $\\mathbf{x}$ , no ground truth label, and outputs $\\mathbf{z}$ , from which a defensive perturbation $\\delta \\mathbf{x}_D$ , $||\\delta \\mathbf{x}_D||_p < \\epsilon_D$ , is computed as in Eq. 5. We assume again that a robust legacy $C$ is given and we train $R$ to make each $\\mathbf{x} + \\delta \\mathbf{x}_D$ in the training dataset $\\epsilon_A^R$ -robust with respect to $C$ in $p$ -norm. More formally, the weights $\\theta_R$ of $R$ are found by minimization of the average worst case entropy on the training dataset $E(\\theta_R)$ , defined as:", + "bbox": [ + 498, + 589, + 892, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} = R (\\boldsymbol {x} \\mid \\boldsymbol {\\theta} _ {R}), E (\\boldsymbol {\\theta} _ {R}) = (1 / N) \\sum_ {\\boldsymbol {x}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 753, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the dataset size. Unlike preemptive robustification [17], $R$ does not iteratively solve a complex optimization problem, neither we need running $C$ before $R$ . Therefore, we can use $A^5 / R$ to protect $\\pmb{x}$ against $\\mathrm{MitM}$ adversarial attacks on-the-fly, ignoring its class, soon after its acquisition, and before its transmission to $C$ (see $A^5 / R$ in Supplementary). Furthermore, since $R$ is trained for a specific", + "bbox": [ + 498, + 794, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4070", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$C$ , it can be deployed on the field (e.g., into the firmware of the acquisition devices) without changing (and thus maintaining the legacy of) the classifier infrastructure.", + "bbox": [ + 76, + 90, + 470, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Online robustification with robustifier $R$ , retrained $C$ ( $A^5 / RC$ ): The $A^5 / RC$ recipe is equivalent to $A^5 / R$ , but it leverages the co-adaptation of $C$ and $R$ by joint training (see $A^5 / RC$ in Supplementary). The cost function is again the average worst case entropy $E(\\theta_C, \\theta_R)$ , defined now as:", + "bbox": [ + 76, + 154, + 468, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} = R (\\boldsymbol {x} | \\boldsymbol {\\theta} _ {R}), \\quad \\boldsymbol {y} = C (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} | \\boldsymbol {\\theta} _ {C}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 241, + 468, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE \\left(\\boldsymbol {\\theta} _ {C}, \\boldsymbol {\\theta} _ {R}\\right) = (1 / N) \\sum_ {\\boldsymbol {x}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) ], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 258, + 468, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{C}$ are the parameters of $C$ . Training $R$ and $C$ from scratch is often unstable and requires careful tuning to achieve convergence (likewise robust classifiers [12, 42]). Therefore, we first train $R$ with $A^5 / C$ and then fine tune $R$ and $C$ together. During the tuning phase, we can use random input augmentation to prevent overfitting (see Section 4 and Supplementary for details). Differently from $A^5 / R$ , $A^5 / RC$ does not preserve the legacy classifier $C$ , but it achieves the same scope (with more effectiveness) in terms of protection against MitM attacks.", + "bbox": [ + 76, + 297, + 468, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Offline physical robustification with $(A^5 /PC)$ and without $(A^5 /P)$ retraining $C$ , with ground truth: This recipe aims at producing physical objects that are certifiably robust against MitM attacks (see $A^5 /P$ , $A^5 /PC$ in Supplementary). Similarly to unadversarial objects [24], we assume that the system creator can design the entire infrastructure and the attacker cannot interfere in this phase. The problem is formally defined for $A^5 /PC$ as:", + "bbox": [ + 76, + 467, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} = A (\\boldsymbol {w} + \\delta \\boldsymbol {w} _ {D}), \\| \\delta \\boldsymbol {w} _ {D}) \\| < \\epsilon_ {D}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 599, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE \\left(\\boldsymbol {\\theta} _ {C}, \\delta \\boldsymbol {w} _ {D}\\right) = (1 / N) \\sum_ {\\boldsymbol {x}} E (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D}) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 618, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{w}$ is the object appearance, whereas $\\pmb{x} = A(\\pmb{w})$ is the data acquisition model (e.g., image formation). $N$ is here the number of random samples generated as $\\pmb{x} + \\delta \\pmb{x}_D = A(\\pmb{w} + \\delta \\pmb{w}_D)$ ; it is important noticing in fact that the acquisition process is not deterministic: as an exemplary case we mention OCR, where $\\pmb{w}$ is the shape of a character, while $A$ may include random perspective transformations, contrast and brightness changes, noise addition and blurring, that may occur on a real camera. The defensive perturbation $\\delta \\pmb{w}_D$ represents a change in the physical character appearance $\\pmb{w}$ that decreases the success rate of an attack towards the acquired image $\\pmb{x}$ (therefore after framing the character with $A$ ) and the classifier $C$ .", + "bbox": [ + 76, + 659, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In $A^5 / P$ we change the character shapes and use the legacy classifier $C$ ; in $A^5 / PC$ we also train $C$ while estimating the robust characters $\\boldsymbol{w} + \\delta \\boldsymbol{w}$ . In both cases $C$", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ede81c608ee7ac20dc79d9244f93dcd2c39ce9c0cb68e372430531962e071ec3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
εAεAC=0.0εAC=0.1εAC=0.2εAC=0.3εAC=0.4
0.1——1.17 [—, 2.24] [42]——
0.10.81 [89.02, 100.00]0.92 [2.00, 5.41]1.32 [2.14, 2.60]1.44 [2.33, 2.63]2.42 [3.36, 3.41]
0.3———1.82 [—, 7.02] [42]—
0.30.69 [100.00, 100.00]0.94 [98.41, 100.00]1.30 [99.62, 100.00]1.41 [7.06, 9.98]2.58 [7.03, 7.68]
", + "bbox": [ + 501, + 87, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST classifiers trained by us with CROWN-IBP [42], under attack $\\epsilon_{A} = \\{0.1,0.3\\}$ ; we also show the metrics reported in [42]. A $100\\%$ certified error rate identifies DNNs that are not protected against attacks.", + "bbox": [ + 498, + 145, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/08815ecb0e201d5f8d7ac9af25174fee5770a5dd3df6f3b992b1666792f726b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.79 [87.09, 100.00]0.82 [1.51, 2.89]1.22 [1.64, 1.86]1.34 [1.76, 1.87]2.38 [2.91, 2.92]
0.10.100.79 [86.51, 100.00]0.72 [1.21, 1.92]1.15 [1.43, 1.49]1.34 [1.50, 1.58]2.41 [2.71, 2.74]
0.10.200.89 [86.40, 100.00]0.56 [0.78, 1.04]1.13 [1.18, 1.21]1.35 [1.37, 1.38]2.40 [2.55, 2.55]
0.10.300.93 [86.44, 100.00]0.51 [0.57, 0.70]1.14 [1.16, 1.17]1.34 [1.37, 1.37]2.38 [2.52, 2.52]
0.10.400.99 [85.93, 100.00]0.52 [0.53, 0.60]1.14 [1.16, 1.16]1.35 [1.38, 1.38]2.44 [2.51, 2.51]
0.30.050.71 [100.00, 100.00]0.93 [97.18, 100.00]1.34 [99.17, 100.00]1.35 [5.16, 6.97]2.55 [6.00, 6.42]
0.30.100.72 [100.00, 100.00]0.88 [95.32, 100.00]1.36 [98.58, 100.00]1.33 [4.12, 5.19]2.49 [5.54, 5.85]
0.30.200.72 [100.00, 100.00]0.88 [90.62, 100.00]1.41 [96.21, 100.00]1.32 [3.21, 3.77]2.35 [5.02, 5.22]
0.30.300.73 [100.00, 100.00]0.89 [86.28, 100.00]1.41 [94.18, 100.00]1.31 [2.91, 3.40]2.32 [4.66, 4.81]
0.30.400.81 [100.00, 100.00]0.85 [83.49, 100.00]1.42 [92.63, 100.00]1.32 [2.74, 3.20]2.28 [4.16, 4.29]
", + "bbox": [ + 501, + 228, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, $A^5 / O$ , under attack $\\epsilon_A = \\{0.1, 0.3\\}$ . During training we use $\\epsilon_A^R = \\epsilon_A$ .", + "bbox": [ + 498, + 324, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/09a44271a78f72cfa48ea272b6121265766c2001d0bbd38cd6a6284714cc8d35.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.82 [85.50, 100.00]0.95 [1.71, 3.58]1.30 [1.82, 2.02]1.40 [1.86, 2.03]2.43 [2.96, 2.97]
0.10.100.81 [82.87, 100.00]1.00 [1.48, 2.54]1.19 [1.62, 1.73]1.40 [1.68, 1.74]2.38 [2.77, 2.79]
0.10.200.88 [78.48, 100.00]0.93 [1.33, 1.77]1.20 [1.42, 1.52]1.31 [1.53, 1.59]2.26 [2.55, 2.57]
0.10.300.99 [75.89, 100.00]0.91 [1.16, 1.54]1.14 [1.31, 1.36]1.26 [1.43, 1.47]2.20 [2.42, 2.43]
0.10.401.04 [75.69, 100.00]0.87 [1.05, 1.35]1.14 [1.27, 1.33]1.24 [1.43, 1.46]2.22 [2.41, 2.43]
0.30.050.69 [100.00, 100.00]0.94 [96.93, 100.00]1.33 [99.11, 100.00]1.41 [5.13, 6.91]2.58 [5.77, 6.09]
0.30.100.71 [100.00, 100.00]0.91 [95.09, 100.00]1.33 [98.47, 100.00]1.38 [3.86, 4.81]2.55 [4.71, 4.83]
0.30.200.72 [100.00, 100.00]0.92 [90.13, 100.00]1.37 [96.08, 100.00]1.41 [2.26, 2.63]2.55 [3.33, 3.40]
0.30.300.75 [100.00, 100.00]0.93 [85.54, 100.00]1.35 [93.36, 100.00]1.32 [1.86, 1.99]2.37 [2.86, 2.88]
0.30.400.83 [100.00, 100.00]0.96 [82.48, 100.00]1.36 [92.15, 100.00]1.29 [1.69, 1.83]2.35 [2.67, 2.69]
", + "bbox": [ + 501, + 378, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, $A^5 / R$ , under attack $\\epsilon_A = \\{0.1, 0.3\\}$ . During training we use $\\epsilon_A^R = \\epsilon_A$ .", + "bbox": [ + 498, + 474, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is previously trained to be robust, e.g., through CROWN-IBP [42]. The $A^5 / P$ and $A^5 / PC$ recipes find application in all those situations where the developer can design the entire infrastructure; beyond OCR, other examples are road sign or robotic tool design and classification.", + "bbox": [ + 498, + 544, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results and discussion", + "text_level": 1, + "bbox": [ + 500, + 633, + 710, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We test $A^5 / O$ , $A^5 / R$ , and $A^5 / RC$ on MNIST [10], CIFAR10 [14], FashionMNIST [34], and Tinyimageenet [15], that have been widely used to establish significant milestones in the adversarial defense field. Training schedules and DNN architectures are given in the Supplementary. In all cases we use a $p = \\infty$ norm for the attack and defense. Our intention is to show the improvement achieved by $A^5$ over traditional, certified methods like CROWN-IBP. The reader should anyway keep in mind that the direct comparison of $A^5$ and CROWN-IBP is partially unfair, as CROWN-IBP does not apply any preemptive defense.", + "bbox": [ + 496, + 657, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first use CROWN-IBP to train five certifiably robust MNIST classifiers with varying levels of robustness, for training attacks $\\epsilon_A^C = 0.0$ (unprotected $C$ ) and $\\epsilon_A^C = \\{0.1, 0.2, 0.3, 0.4\\}$ . Their clean and certified errors are reported in Table 2, for attacks $\\epsilon_A = \\{0.1, 0.3\\}$ . We also", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e3ff4516848f66abe0c02a37a7aac280b601ef872b1eb8cb825f02b4b315b5c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
εA=εCεDεCA=0.0εCA=0.1εCA=0.2εCA=0.3εCA=0.4
0.10.0588.66[88.66, 88.66]1.00[1.38, 2.35]1.16[1.61, 2.46]1.14[1.54, 2.29]1.37[1.72, 2.33]
0.10.1088.66[88.66, 88.66]0.77[0.95, 1.40]0.99[1.16, 1.56]0.86[1.03, 1.42]1.13[1.31, 1.72]
0.10.2088.66[88.66, 88.66]0.69[0.80, 1.05]0.84[0.94, 1.25]0.96[1.07, 1.33]0.88[1.10, 1.54]
0.10.3088.66[88.66, 88.66]0.64[0.70, 1.02]0.91[0.99, 1.16]0.79[0.88, 1.21]1.00[1.12, 1.43]
0.10.4088.66[88.66, 88.66]0.61[0.68, 0.88]0.93[0.99, 1.18]0.69[0.76, 1.10]1.00[1.03, 1.30]
0.30.0588.66[88.66, 88.66]88.66[88.66, 88.66]2.57[8.35, 10.86]3.42[9.51, 11.72]8.00[18.05, 19.77]
0.30.1088.66[88.66, 88.66]88.66[88.66, 88.66]2.07[5.66, 7.56]2.35[6.18, 8.07]2.35[6.31, 7.93]
0.30.2088.66[88.66, 88.66]1.68[3.11, 4.79]1.40[2.42, 3.62]1.28[2.82, 4.01]1.88[3.53, 4.90]
0.30.3088.66[88.66, 88.66]0.84[0.96, 1.38]1.00[1.14, 1.36]0.94[1.15, 1.59]1.07[1.39, 1.90]
0.30.4088.66[88.66, 88.66]0.72[0.80, 1.03]0.75[0.82, 1.06]0.91[1.14, 1.68]0.87[1.14, 1.50]
", + "bbox": [ + 78, + 87, + 467, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "report the error rates estimated with autoattack [8, 9], an automatic toll based on an ensemble of four attacks to evaluate the practical DNN robustness. These error rates are in the same ballpark of those reported in the CROWN-IBP paper [42]: they represent the baseline for the analysis of the different $A^5$ recipes considered here.", + "bbox": [ + 75, + 268, + 467, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each $C$ we run $A^5 /O$ on the MNIST test set, for defense magnitudes $\\epsilon_{D} = \\{0.05,0.1,0.2,0.3,0.4\\}$ . During the optimization of the entropy in Eq. 6, we use attacks $\\epsilon_A^R = \\{0.1,0.3\\}$ , not necessarily equal to $\\epsilon_A^C$ . Table 3 reports the clean, autoattack, and certified errors for a testing attack $\\epsilon_{A} = \\epsilon_{A}^{R}$ . When $C$ is not robust ( $\\epsilon_{A}^{C} = 0.0$ ), $A^5 /O$ does not reduce the certified error², otherwise it generally overcomes the base CROWN-IBP classifier, as shown by comparing against the error rates in Table 2. As a rule of thumb, the best clean / certified error trade off is often achieved for $\\epsilon_{A}^{R} = \\epsilon_{A}^{C} = \\epsilon_{A}$ . Results also improve, as expected, for larger defenses $\\epsilon_{D}$ , when $A^5 /O$ beat CROWN-IBP by a significant margin on clean and certified errors. However, $A^5 /O$ does not work well when the legacy $C$ is trained with a large $\\epsilon_{A}^{C}$ , possibly because of its initial low clean accuracy. The results are consistent with our interpretation of the effects of adversarial training and robustification on the classification landscape (Fig. 1): training $C$ with a large $\\epsilon_{A}^{C}$ creates classification valleys wide enough to accommodate $\\pmb{x} + \\delta \\pmb{x}_D + \\delta \\pmb{x}_A$ , rendering the adversarial attacks ineffective. $A^5$ moves any $\\pmb{x}$ towards the centers of the valleys; a large $\\epsilon_{D}$ allows moving them more easily. Too large $\\epsilon_{A}^{C}$ remains however detrimental for the clean accuracy.", + "bbox": [ + 75, + 359, + 467, + 707 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 reports results for $A^5 / R$ , demonstrating that training a DNN robustifier $R$ for on-the-fly, certified robustification, ignoring the ground truth, is feasible. Likewise $A^5 / O$ , $A^5 / R$ beats the base CROWN-IBP classifier by a significant margin. Sometimes $A^5 / O$ does better by leveraging the knowledge of the ground truth class, whereas the regular robustification signal generated by $A^5 / R$ is more effective in other cases.", + "bbox": [ + 75, + 708, + 467, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our last MNIST experiment investigates $A^5 /RC$ that does not preserve the legacy $C$ like $A^5 /R$ , but achieves bet", + "bbox": [ + 76, + 829, + 467, + 861 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/abea04267df01868f912526fdaf5e4bc00886cf477a56bfa0d65a5a94527531c.jpg", + "table_caption": [ + "Table 5. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST, $A^5 / RC$ , under attack $\\epsilon_A = \\{0.1, 0.3\\}$ . During training we use $\\epsilon_A^R = \\epsilon_A$ , and data augmentation." + ], + "table_footnote": [], + "table_body": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2554/25552.75 [63.47, 64.09]53.34 [62.42, 62.79]53.75 [64.92, 65.49]
8/2558/25551.06 [60.04, 60.73]52.36 [59.24, 59.58]53.16 [63.50, 64.03]
8/25516/25548.72 [54.89, 55.25]50.60 [55.22, 55.38]52.23 [60.79, 61.14]
8/25532/25545.68 [49.51, 49.74]48.06 [50.86, 50.94]50.99 [57.37, 57.61]
", + "bbox": [ + 500, + 87, + 890, + 145 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6de69708d9c905b2a0e61320a73dd5616fa84ec7cb74ec82430c6ca1ebd81e0d.jpg", + "table_caption": [ + "Table 6. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, $A^5 / O$ , under attack $\\epsilon_A = 8/255$ . During training we use different $\\epsilon_A^R$ and $\\epsilon_D$ . The robust classifier trained with CROWN-IBP has clean [and certified] errors equal to 54.02 [66.94] or 45.47 [69.55], depending on the CROWN-IBP training configuration [42]." + ], + "table_footnote": [], + "table_body": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25554.60 [66.34, 67.11]54.60 [66.34, 67.11]54.60 [66.31, 67.11]
8/2554/25553.73 [64.07, 64.57]54.12 [63.32, 63.74]54.37 [63.58, 64.03]
8/2558/25552.91 [62.59, 63.12]53.24 [61.38, 61.93]54.54 [60.85, 61.16]
8/25516/25551.58 [60.40, 61.02]51.95 [59.21, 59.67]54.44 [59.87, 60.11]
8/25532/25550.91 [57.26, 57.75]51.70 [56.91, 57.43]54.05 [58.39, 58.73]
", + "bbox": [ + 501, + 253, + 890, + 325 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc002a5be2e733ed779ebfc231fa28b42a6c9665f9c5e8f6c8c155d33b2000b5.jpg", + "table_caption": [ + "Table 7. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, $A^5 / R$ , under attack $\\epsilon_A = 8 / 255$ . During training we use different $\\epsilon_A^R$ and $\\epsilon_D$ . The robustifier $R$ in the first row uses $\\epsilon_D = 0$ and thus it does not provide any form of defense — the error reported here is equivalent to that of a robust classifier trained with CROWN-IBP [42]." + ], + "table_footnote": [], + "table_body": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25550.24 [69.17, 73.65]58.63 [69.25, 70.63]74.28 [77.82, 77.91]
8/2554/25546.76 [63.03, 70.99]54.05 [62.80, 64.44]67.59 [71.20, 71.32]
8/2558/25541.24 [54.01, 64.31]50.74 [56.63, 59.49]65.96 [67.63, 67.78]
8/25516/25536.96 [42.45, 52.90]39.15 [41.33, 44.22]58.17 [58.67, 58.85]
8/25532/25535.26 [37.13, 42.76]41.59 [43.01, 45.55]42.25 [43.02, 44.08]
", + "bbox": [ + 500, + 433, + 890, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, $A^5 / RC$ , under attack $\\epsilon_A = 8/255$ . During training we use different $\\epsilon_A^R$ and $\\epsilon_D$ . The robustifier $R$ in the first row uses $\\epsilon_D = 0$ and thus it does not provide any form of defense, whereas the corresponding robust CROWN-IBP classifier is fine tuned using $\\epsilon_A^C = \\epsilon_R$ during training.", + "bbox": [ + 498, + 516, + 890, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ter clean accuracy and protection (Table 5). In the Supplementary we discuss the role of data augmentation to achieve this result. Interestingly, the best results are obtained when $C$ is initially trained for $\\epsilon_A^C \\leq \\epsilon_A$ , probably because $C$ has an initial high clean accuracy and subsequently co-adapts with $R$ to also guarantee a small certified error.", + "bbox": [ + 498, + 628, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 shows examples of robustification on MNIST. At visual inspection, $A^5 / O$ , $A^5 / R$ , and $A^5 / RC$ act similarly on these images: they add contrast and enhance the high frequency features (i.e., edges) of the digits. The few failure cases are often pictures that look ambiguous even to a human observer (right panels in Fig. 3).", + "bbox": [ + 496, + 719, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform tests on CIFAR10 and compare $A^5$ against a state of the art CROWN-IBP classifiers $C$ , trained with $\\epsilon_A^C = 8.8 / 255$ , that we also adopt as legacy $C$ for $A^5$ . To test the effect of the magnitude of the attack used for robustification (with $A^5 / O$ ) or for training $R$ (for $A^5 / R, A^5 / RC$ ), we use $\\epsilon_A^R = \\{4 / 255, 8 / 255, 16 / 255\\}$ that are", + "bbox": [ + 498, + 809, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Coherently with [24], the worst case cross entropy improves, but not enough to affect the ranking of the classified classes under attack.", + "bbox": [ + 75, + 875, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4072", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/468d7a738172efaae82413600154de03b76c286f7eebc1e91f303ccfbbeae23e.jpg", + "image_caption": [ + "Figure 3. Successful and failed robustifications for $A^5 / O$ , $A^5 / R$ , and $A^5 / RC$ , for MNIST ( $\\epsilon_A = \\epsilon_A^C = \\epsilon_A^R = \\epsilon_D = 0.3$ ) and CIFAR10 ( $\\epsilon_A = 8.8 / 255$ , $\\epsilon_A^C = 8 / 255$ , $\\epsilon_A^R = 4 / 255$ , and $\\epsilon_D = 32 / 255$ ). Each triplet shows the original (left) and robustified (center) images; the rightmost panel is the defensive augmentation. All the $A^5$ recipes consistently increase the contrast and the high frequency content on MNIST, where failure cases are ambiguous even for a human observer. For CIFAR10, $A^5 / O$ leverage isolated pixel changes, whereas $A^5 / R$ and $A^5 / RC$ consistently increase the color contrast and saturation, similarly to results in literature [17]." + ], + "image_footnote": [], + "bbox": [ + 147, + 87, + 823, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5eecde7c36eb1cb17bf359374e103bf358884f2393484dc265a9e25f36e525f8.jpg", + "image_caption": [ + "Figure 4. The first, second and third row show respectively the robustified $(\\boldsymbol{w} + \\delta \\boldsymbol{w}_D)$ , nominal $(\\boldsymbol{w})$ and robustification $(\\delta \\boldsymbol{w}_D)$ sets of 62 letters considered for testing $A^5 / PC$ . The last row shows examples of the robustified letters after framing them, $\\boldsymbol{x} + \\delta \\boldsymbol{x}_D = A(\\boldsymbol{w} + \\delta \\boldsymbol{w}_D)$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 750, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "smaller, equal or greater than $\\epsilon_A^C$ . Results in Table 6 are consistent with those measured for MNIST: $A^5 / O$ signifi", + "bbox": [ + 75, + 857, + 468, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cantly improves the clean and certified error rates over the base CROWN-IBP classifier. Training $R$ through $A^5 / R$", + "bbox": [ + 500, + 857, + 890, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cd25bd1a42ee0c2e3970a287b7eeaefe6cf55494066ff38969dfbbb89e4d3ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AlgoError [certified error]
Vanilla0.89% [100.00%]
CROWN-IBP3.85% [13.85%]
A5/P3.08% [11.84%]
A5/PC0.73% [4.20%]
", + "bbox": [ + 138, + 88, + 410, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Error and certified error for the classification of the 62 characters in Fig. 4, for $\\epsilon_A = 0.1$ and different algorithms. $A^5 /P$ changes the physical shape of the letters to achieve robustness after framing them; $A^5 /PC$ also trains the associated classifier.", + "bbox": [ + 76, + 176, + 470, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "produces a smaller improvement, still sufficient to beat the CROWN-IBP classifier by a significant margin (Table 7). The value of $\\epsilon_A^R$ has a significant impact on the final result: if too large ( $\\epsilon_A^R > \\epsilon_A$ ), the clean accuracy is penalized, because $A^5$ looks for very large plateaus in the classification landscape, that are hard or impossible to find. In other words, we observe for $A^5$ a trade-off between the clean and certified accuracy, likewise adversarial training and certified methods. Best results on CIFAR10 are achieved again by $A^5 / RC$ , when $\\epsilon_A^R$ is equal or slightly smaller than $\\epsilon_A = \\epsilon_A^C$ . The improvements in clean and certified error are in this case in the impressive range of $19\\%$ and $24\\%$ respectively (Table 8). We also notice here that the theoretical certified error rates measured with auto_LiRPA [37] in $A^5$ and those measured experimentally by autoattack [8, 9] are close for $A^5$ , whose working point at convergence may favor the consistency between these two estimates.", + "bbox": [ + 75, + 263, + 468, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also for CIFAR10 we perform a visual inspection of defensive augmentation (Fig. 3). $A^5 / O$ exploits a slight contrast increase and modifies isolated pixels: likewise adversarial attacks with small $p$ norm (up to single pixel attacks [29]), the same seems possible when crafting defensive perturbations. Preemptive robustification [17] shows a similar pattern (see their Fig. 6): non natural textures and enhanced color saturation emerge in their images for large defensive perturbations. The robust images crafted by $R$ in $A^5 / R$ or $A^5 / RC$ , however, are different: they generally show increased contrast and color saturation, without the textures generated by $A^5 / O$ and [17]. We speculate that these textures may be associated with the iterative optimization in $A^5 / O$ and [17], that may both suffer some form of obfuscated gradient [2]; a more regular defense is instead generated by $R$ without resorting to any iterative process.", + "bbox": [ + 75, + 521, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional results on FashionMNIST and Tinyimagenet, overall consistent with the ones presented here and showing the quantitative advantages provided by $A^5$ on these datasets, are reported in the Supplementary. These experiments highlight two important facts. The first one is that the integration of the design philosophy of $A^5$ with that of other recent ideas in the field of adversarial defense (like the $\\ell_{\\infty}$ -dist neurons that implement 1-Lipschitz functions and are inherently robust to adversarial attacks) may lead to", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "even bigger improvements in terms of certified robustness. The second is that the integration of different defense strategies may indeed be strictly needed to guarantee further progresses, as scaling to very large dataset remains problematic even for a preemptive robustification algorithms like $A^5$ .", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, for $A^5 / P$ and $A^5 / PC$ , we test the classification of an alphabet of 62 characters (Fig. 4), after random rotation, shift, perspective distortion, noise addition, blurring and color jittering (details in Supplementary), that simulate the scanning of a document for OCR. Table 9 shows the clean and certified errors for: a vanilla, non robust classifier; a CROWN-IBP classifier; $A^5 / P$ ; and $A^5 / PC$ . $A^5$ achieves again a significant improvement both on clean and certified errors: $A^5 / PC$ has better (0.73%) accuracy of the vanilla $C$ (0.89%) and can be attacked in only 4.20% of the cases. This results is obtained with a slight modification of the shape of the prototype characters shown in Fig. 4.", + "bbox": [ + 496, + 167, + 892, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and conclusion", + "text_level": 1, + "bbox": [ + 500, + 369, + 741, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce $A^5$ , the first framework that leverages adversarial augmentation to preemptively modify the input of a DNN or a physical object, to make it certifiable robust against adversarial attacks. $A^5$ is not simply complementary to other defense methods: its tights with them are strict and they can be used together to achieve higher robustness. For instance, one of the limitations we found is that $A^5 / R$ requires a somehow robust initial classifier $C$ . Co-training $R$ and a poorly robust $C$ easily compensates for the initially low robustness of $C$ and boosts its high clean accuracy while also achieving state of the art results in terms of certified accuracy; finding the optimal initial robustness of $C$ remains an open question though.", + "bbox": [ + 496, + 397, + 892, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we do not investigate all the possible $A^5$ recipes: in its most general implementation $(A^5 / PRC)$ the physical objects, $R$ and $C$ could all be optimized at the same time. This opportunity can be explored through the released code at https://github.com/NVlabs/A5. This will also serve answering other questions that do not find space here: for instance, we do not know if robustification generalizes to other classifiers as adversarial attacks do, how $A^5$ works in norms other than $p = \\infty$ , and how much it can further improve if coupled with state-of-the-art solutions like the recently proposed $\\ell_{\\infty}$ -dist neurons [41].", + "bbox": [ + 496, + 595, + 890, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different $A^5$ recipes find applications in scenarios where the user can control the acquisition device, equip it with $R$ and guarantee its protection (e.g., image acquisition on a phone before server communication) or while designing the infrastructure (e.g., robust road signs creation); it may be not suitable for more general adversarial scenarios. Overall, we believe that the practical deployment of robust systems may benefit from (or even require) methods for both robust classifiers and preemptive robustification.", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4074", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maungmaung Aprilpyone and Hitoshi Kiya. Encryption inspired adversarial defense for visual classification. 2020 IEEE International Conference on Image Processing (ICIP), pages 1681-1685, 2020. 2, 3", + "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 274-283. PMLR, 10-15 Jul 2018. 3, 8", + "[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples, 2018. 3", + "[4] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 4312-4321. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Survey Track. 1, 2, 3", + "[5] Tom Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. 2017. 3", + "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1, 3", + "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020. 3", + "[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 5, 6, 8", + "[9] Francesco Croce and Matthias Hein. Mind the box: $l_{1}$ -apgd for sparse adversarial attacks on image classifiers. In ICML, 2021. 5, 6, 8", + "[10] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 5", + "[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1, 2", + "[12] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Mann, and Pushmeet Kohli. On the effectiveness of interval bound propagation for training verifiably robust models, 2018. 1, 2, 3, 5", + "[13] Sonal Joshi, Jesús Villalba, Piotr Želasko, Laureano MoroVelázquez, and Najim Dehak. Study of pre-processing defenses against adversarial attacks on state-of-the-art speaker recognition systems. IEEE Transactions on Information Forensics and Security, 16:4811-4826, 2021. 1, 2, 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5", + "[15] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 5", + "[16] Xiaoting Li, Lingwei Chen, Jinquan Zhang, James Larus, and Dinghao Wu. Watermarking-based defense against adversarial attacks on deep neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2021. 2, 3", + "[17] Seungyong Moon, Gaon An, and Hyun Oh Song. Preemptive image robustification for protecting users against man-in-the-middle adversarial attacks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 4, 7, 8", + "[18] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning (ICML), 2022. 1, 3", + "[19] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. pages 1491-1500, 10 2017. 1, 3", + "[20] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 1, 2, 3", + "[21] Rafael Pinot, Raphael Ettedgui, Geovani Rizk, Yann Chevaleyre, and Jamal Atif. Randomization matters how to defend against strong adversarial attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7717-7727. PMLR, 13-18 Jul 2020. 3", + "[22] Yao Qin, Nicholas Carlini, Ian J. Goodfellow, G. Cottrell, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. ArXiv, abs/1903.10346, 2019. 2, 3", + "[23] Adnan Siraj Rakin, Zhezhi He, Boqing Gong, and Deliang Fan. Blind pre-processing: A robust defense method against adversarial examples. ArXiv, abs/1802.01549, 2018. 2, 3", + "[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 3, 5, 6", + "[25] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2", + "[26] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Fawkes: Protecting Privacy against Unauthorized Deep Learning Models. USENIX Association, USA, 2020. 1, 3", + "[27] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Jinmyeong Shin, Seok-Hwan Choi, and Yoon-Ho Choi. Is homomorphic encryption-based deep learning secure enough? Sensors, 21(23), 2021. 2, 3", + "[29] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 8", + "[30] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. Jan. 2014. 2nd International Conference on Learning Representations, ICLR 2014; Conference date: 14-04-2014 Through 16-04-2014. 1", + "[31] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1633-1645. Curran Associates, Inc., 2020. 3", + "[32] Derui Wang, Chaoran Li, Sheng Wen, Surya Nepal, and Yang Xiang. Man-in-the-middle attacks against machine learning classifiers via malicious generative models. IEEE Transactions on Dependable and Secure Computing, 18(5):2074-2087, sep 2021. 2, 3", + "[33] Shiqi Wang, Huan Zhang, Kaidi Xu, Xue Lin, Suman Jana, Cho-Jui Hsieh, and J Zico Kolter. Beta-CROWN: Efficient bound propagation with per-neuron split constraints for complete and incomplete neural network verification. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[34] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5", + "[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In International Conference on Learning Representations, 2018. 1, 3", + "[36] Kaidi Xu, Zhouxing Shi, Huan Zhang, Yihan Wang, Kai-Wei Chang, Minlie Huang, Bhavya Kailkhura, Xue Lin, and Cho-Jui Hsieh. Automatic perturbation analysis for scalable certified robustness and beyond. arXiv: Learning, 2020. 3", + "[37] Kaidi Xu, Huan Zhang, Shiqi Wang, Yihan Wang, Suman Jana, Xue Lin, and Cho-Jui Hsieh. Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers. In International Conference on Learning Representations, 2021. 2, 3, 4, 8", + "[38] Hiromu Yakura and Jun Sakuma. Robust audio adversarial example for a physical attack. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5334–5341. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 3", + "[39] Jianfei Yang, Han Zou, and Lihua Xie. Robustsense: Defending adversarial attack for secure device-free human activity recognition. ArXiv, abs/2204.01560, 2022. 2, 3", + "[40] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of Machine Learning Research, pages 12062-12072. PMLR, 18-24 Jul 2021. 1, 3", + "[41] Bohang Zhang, Tianle Cai, Zhou Lu, Di He, and Liwei Wang. Towards certifying 1-infinity robustness using neural networks with 1-inf-dist neurons. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12368-12379. PMLR, 18-24 Jul 2021. 3, 8", + "[42] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane Boning, and Cho-Jui Hsieh. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 5, 6", + "[43] Huan Zhang, Tsui-Wei Weng, Pin-Yu Chen, Cho-Jui Hsieh, and Luca Daniel. Efficient neural network robustness certification with general activation functions. In Advances in Neural Information Processing Systems (NIPS), arXiv preprint arXiv:1811.00866, dec 2018. 1, 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 357 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4076", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_model.json b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6c622b6733f92989fc51ff1a511c79be383d5d4b --- /dev/null +++ b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.844, + 0.175 + ], + "angle": 0, + "content": "The Best Defense is a Good Offense: Adversarial Augmentation against Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.205, + 0.454, + 0.221 + ], + "angle": 0, + "content": "Iuri Frosio, NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.226, + 0.453, + 0.239 + ], + "angle": 0, + "content": "ifrosio@nvidia.com" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.205, + 0.683, + 0.221 + ], + "angle": 0, + "content": "Jan Kautz, NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.226, + 0.678, + 0.239 + ], + "angle": 0, + "content": "jkautz@nvidia.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.275, + 0.313, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.474, + 0.669 + ], + "angle": 0, + "content": "Many defenses against adversarial attacks (e.g. robust classifiers, randomization, or image purification) use countermeasures put to work only after the attack has been crafted. We adopt a different perspective to introduce \\( A^5 \\) (Adversarial Augmentation Against Adversarial Attacks), a novel framework including the first certified preemptive defense against adversarial attacks. The main idea is to craft a defensive perturbation to guarantee that any attack (up to a given magnitude) towards the input in hand will fail. To this aim, we leverage existing automatic perturbation analysis tools for neural networks. We study the conditions to apply \\( A^5 \\) effectively, analyze the importance of the robustness of the to-be-defended classifier, and inspect the appearance of the robustified images. We show effective on-the-fly defensive augmentation with a robustifier network that ignores the ground truth label, and demonstrate the benefits of robustifier and classifier co-training. In our tests, \\( A^5 \\) consistently beats state of the art certified defenses on MNIST, CIFAR10, FashionMNIST and Tinyimagenet. We also show how to apply \\( A^5 \\) to create certifiably robust physical objects. Our code at https://github.com/NVlabs/A5 allows experimenting on a wide range of scenarios beyond the man-in-the-middle attack tested here, including the case of physical attacks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.208, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Since Deep Neural Networks (DNNs) have been found vulnerable to adversarial attacks [11, 30], researchers studied various protection strategies [4, 12, 20, 42, 43]. For instance, adversarial training [11, 30] generates attacks while asking a DNN for the correct output in training; it is simple, partially effective and widely adopted. Certified methods (e.g., IBP [12], CROWN [43], CROWN-IBP [42]) do a step more by estimating correct (although often pessimistic) output bounds (Fig. 1, a) used for training. Adversarial training regularizes the classification landscape against the attacks (Fig. 1, b), but high protection often produces a loss in clean accuracy. Other partially effective defenses are based" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.276, + 0.893, + 0.322 + ], + "angle": 0, + "content": "on randomness [35, 35] or removal of the adversarial signal [13, 18, 27, 40], by moving the input back to the space of natural (non-attacked) data before classification (Fig. 1, c)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.893, + 0.43 + ], + "angle": 0, + "content": "All the aforementioned strategies activate the defense mechanism only after the attack has been crafted. However, when dealing with adversarial attacks, the first actor to move has a significant advantage. For instance, perturbing an image can avoid online person identification and preserve privacy [6, 19, 26]. Acting first is particularly suitable against Man in the Middle (MitM) attacks that may prac" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.448, + 0.671, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.504, + 0.676, + 0.517 + ], + "angle": 0, + "content": "(a) certified bounds" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.449, + 0.834, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.503, + 0.851, + 0.518 + ], + "angle": 0, + "content": "(b) adversarial training" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.522, + 0.67, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.578, + 0.683, + 0.592 + ], + "angle": 0, + "content": "(c) image purification" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.522, + 0.836, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.576, + 0.796, + 0.591 + ], + "angle": 0, + "content": "(d) \\(A^5\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.602, + 0.652, + 0.618 + ], + "angle": 0, + "content": "Base: airplane (69.8%) \n[Verified: ship (100.0%)]" + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.602, + 0.836, + 0.618 + ], + "angle": 0, + "content": "Robustified: airplane \\((82.6\\%)\\) Verified: airplane \\((82.5\\%)\\)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.618, + 0.68, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.618, + 0.855, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.744, + 0.894, + 0.897 + ], + "angle": 0, + "content": "Figure 1. (a) An input \\(\\pmb{x} \\in y^*\\) is correctly classified. The grey box (certified bounds [12, 42, 43]) shows that, under an attack \\(\\delta \\pmb{x}_A\\), \\(||\\delta \\pmb{x}_A||_{\\infty} < \\epsilon_A\\), misclassification \\(((\\pmb{x} + \\delta \\pmb{x}_A) \\in y_0)\\) is possible. (b) Adversarial training [4, 20] creates robust DNNs with regular classification landscapes: misclassification is less likely. (c) Image purification [27, 40] moves the attacked input \\(\\pmb{x} + \\delta \\pmb{x}_A\\) by \\(\\delta \\pmb{x}_P\\) back to correct classification. (d) \\(A^5\\) preemptively moves \\(\\pmb{x}\\) into a non-attackable position. The original CIFAR10 image classified as airplane (69.8% confidence) can be misclassified under attack (ship, 100% confidence). Once robustified through \\(A^5\\) (right), misclassification does not occur anymore." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4067" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.201 + ], + "angle": 0, + "content": "
vA vectorvjVector j-th element
CClassifierRRobustifier
xDataδxAAttacking perturbation
δxDDefensive perturbationx + δxAData under attack
x + δxDRobustified datax + δxD + δxARobustified data under attack
εDDefense magnitude, ||δxD||p < εDεRAttack magnitude, ||δxA||p < εR_A while training R
εAAttack magnitude ||δxA||p < εA while testingεCAttack magnitude, ||δxA||p < εC_A while training C
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.211, + 0.848, + 0.226 + ], + "angle": 0, + "content": "Table 1. Notation for data \\( \\mathbf{x} \\) ,that are the inputs of the classifier \\( C \\) . We adopt an equivalent notation for physical objects \\( \\mathbf{w} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.473, + 0.555 + ], + "angle": 0, + "content": "cally arise in automotive [32], audio processing [13, 22, 38], or while communicating with a remote machine [39]. Our idea spouses this reasoning line: we investigate a novel way to augment the data to preemptively certify that it cannot be attacked (Fig. 1, d). This fits in a recent research area that up to now has received less attention than adversarial training. Researchers explored image and real object augmentation to guarantee a high recognition rate in presence of noise or geometric distortions, but not in adversarial scenarios [24]. Encryption schemes coupled with deep learning [1, 28] or watermarking [16] only partially protect against MitM attacks. The few existing preemptive robustification methods [23] include a procedure [17] that first runs a classifier on clean data and achieves preemptive robustification against MitM attacks through iterative optimization; these are partially effective and do not provide any certification. Our novel framework encompasses most of the aforementioned cases while also introducing for the first time the concept of certified robustification. More in detail, our manifold contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.569, + 0.47, + 0.689 + ], + "angle": 0, + "content": "(i) We introduce \\( A^5 \\) (Adversarial Augmentation Against Adversarial Attacks), a comprehensive framework for preemptive augmentation to make data and physical objects certifiably robust against MitM adversarial attacks. As far as we know, this is the first time certified defense is achieved in this way. Since we provide certified robustness, we guarantee protection against any form of white, grey or black box attack." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.689, + 0.47, + 0.886 + ], + "angle": 0, + "content": "(ii) We test different flavours of \\( A^5 \\) on standard datasets. By doing so, we study the connection between the robustness of the legacy classifier, the magnitude of the defensive augmentation, and the protection level delivered by \\( A^5 \\). We show \\( A^5 \\) achieving state-of-the-art certified protection against MitM attacks by training a robustifier DNN coupled with the legacy classifier, and even better results for co-training of the robustifier and classifier. We perform a critical, visual inspection of the robustified images to answer an interesting theoretical question (how does a non-attackable image look like?) and potentially provide directions for the acquisition of inherently robust images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "(iii) Using Optical Character Recognition (OCR) as an ex" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.569, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.252, + 0.892, + 0.298 + ], + "angle": 0, + "content": "ample, we show the application of \\( A^5 \\) for the design of certifiably robust physical objects, which extends [24] to the case of certified defense against MitM attacks." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.299, + 0.892, + 0.375 + ], + "angle": 0, + "content": "(iv) We share our code at https://github.com/NVlabs/A5, to allow replicating our results or testing \\( A^5 \\) in scenarios not considered here, for instance on other datasets or for protection against physical adversarial attack (e.g., adversarial patches)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.787, + 0.408 + ], + "angle": 0, + "content": "2. Related Work and Background" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.892, + 0.462 + ], + "angle": 0, + "content": "We identify three main philosophies for the development of defenses against adversarial attacks. They are not mutually exclusive and can be adopted together." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The first approach aims at creating DNNs that are robust to adversarial attacks: e.g., in adversarial training, attacks are generated while training the DNN that should process them correctly [4, 11, 20]. This regularizes the classification landscape and makes attacks less effective (Fig. 1, b). Adversarial training is simple but has one drawback: as attack generation is computationally demanding (even after speed up [25]), attacks are randomly sampled and cannot cover the entire input space, thus impacting the final accuracy. Certified defense [12, 42, 43] can be seen as a state of the art refined form of adversarial training: robustness is obtained by propagating through the DNN an input interval that encompasses all the attacks with magnitude up to \\(\\epsilon_A^C\\). The corresponding set of output bounds allows estimating the worst case scenario and derive the training cost function (e.g., worst case entropy [12, 42]). Research focuses on finding bounds that are both tight and computationally light. IBP [12] computes (and backwards) the bounds at the cost of an additional forward and backward pass. Since the bounds are loose at the beginning of training, a complex schedule is required to guarantee stability and convergence. Linear relaxation [43] delivers tighter bounds leveraging the correlation between different layers to linearly propagate signal and bounds, when possible, and eventually put in linear relation the input interval and output bounds. CROWN [43] bounds are often tighter than IBP, but at an impractical computational cost for large DNNs. Among the attempts to speed up CROWN [33, 37], CROWN-IBP [42] strikes the right balance of computa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.226 + ], + "angle": 0, + "content": "tional complexity and tightness. It is implemented in the auto_LiRPA\\(^{1}\\) library [36, 37] adopted here. One limitation of adversarial training, certified methods, and \\(A^{5}\\) as well, is the trade-off between clean and certified (worst case underattack) accuracy: increasing the latter leads to a decrease in the former [4, 12, 20, 42]. Among the architecture solutions against adversarial attacks, we also mention the recent development of \\(\\ell_{\\infty}\\)-dist neurons that are inherently robust [41], and could be leveraged by \\(A^{5}\\) as well in future." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.468, + 0.499 + ], + "angle": 0, + "content": "The second defense philosophy is based on active countermeasures taken after the attack is crafted. Image purification [18, 27, 40] projects an attacked image onto the manifold of natural images through a generative model (Fig. 1, c). It is less effective than adversarial training [7], but does not make any assumption on the classifier. Other forms of signal pre-processing have been proposed in the field of audio processing and automatic speech recognition, where MitM physical attacks can be easily implemented [22, 38] and partially defeated by randomized smoothing, DefenseGAN, variational autoencoder (VAE), and Parallel WaveGAN vocoder [13]. Both certified defenses and pre-processing methods do not assume any specific attack form, but the latter are less effective in white box scenarios [2, 31]. Randomization can also be used after the attack, e.g. by changing resolution or padding an image [35], or by picking random experts in a mixture, that is proved to be more robust than deterministic DNNs [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.468, + 0.877 + ], + "angle": 0, + "content": "\\(A^5\\) belongs to a third class of methods, that put in place countermeasures before the attack (Fig. 1, d). The roots of \\(A^5\\) lie in two ideas that have been only partially investigated so far. The first is that, in adversarial scenarios, the first actor to move has a significant advantage. This rule is already leveraged for online privacy, where a preemptive adversarial attack prevents the identification of the framed subject [6, 19]. Since adversarial attacks often generalize to many classifiers, this ploy is particularly effective [26]. Preemptive protection is also a natural form of defense against MitM attacks that can be easily created in contexts like automotive [32], speech and audio processing [13, 22, 38], or anytime the capturing device communicates with a server [39]. Encryption [1, 28] and watermarking [16] coupled with deep learning work against MitM attacks, but do not achieve full protection, are not certified, require the key to be hidden (grey box scenario) and may significantly alter the appearance of the image. Blindpreprocessing [23] applies a tanh transform, normalization, quantization and thermometer encoding to achieve partial and uncertified defense. The closest algorithm to \\(A^5\\) is preemptive robustification [17]: given an image \\(\\pmb{x}\\) and a classifier \\(C\\) that can be fine-tuned, it computes \\(y_0 = C(\\pmb{x})\\) and randomly samples attacks \\(\\delta \\pmb{x}_A\\) against \\(\\pmb{x} + \\delta \\pmb{x}_D\\) to find a defensive perturbation \\(\\delta \\pmb{x}_D\\), such that \\((\\pmb{x} + \\delta \\pmb{x}_D + \\delta \\pmb{x}_A) \\in y_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.364 + ], + "angle": 0, + "content": "(correct classification under attack). Since \\( A^5 \\) uses certified bounds, it results in simpler training and higher protection, and does not require running \\( C \\) first, as our robustifier DNN generalizes to all the inputs. The second idea behind \\( A^5 \\) comes from physical adversarial examples, i.e., 3d physical objects [3] or 2d patches [5] that consistently fool a DNN under a wide range of viewing conditions. Reverting the attacker and defender roles, the complementary cost function leads to the creation of unadversarial examples: patches or textures that, once applied to a real 3d object, reinforce (instead of degrading) the DNN desired behaviour [24]. However, these are not designed to be inherently robust to adversarial attacks. We exploit the idea in \\( A^5 \\) in all those situations where system designers control, to some extent, the inputs fed to the DNN (e.g. in OCR, robotic systems or during the design of the road infrastructure). We show as an example the use of \\( A^5 \\) to design certifiably protected characters for OCR, something unexplored both in [17] and [24]." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.377, + 0.59, + 0.393 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Threat models We summarize our notation in Table 1, whereas Fig. 2 illustrates the full \\( A^5 \\) framework including trainable DNNs in blue, training algorithms in green, physical objects \\( \\boldsymbol{w} \\), and acquired data \\( \\boldsymbol{x} \\). MitM attacks are crafted while transmitting data from the acquisition device to the classifier \\( C \\), while physical attacks (not tested here) are crafted onto physical objects \\( \\boldsymbol{w} \\), before data acquisition. \\( A^5 \\) can be used in a scenario where the attacker has full access to the \\( C \\): since it uses certified bounds, it is agnostic to the specific nature of the attack. In other words, we certify protection for any \\( \\delta x_A \\), \\( ||\\delta x_A|| < \\epsilon_A \\) and the attacker can use any white-, grey- or black-box attacking algorithm. We assume that \\( A^5 \\) can run in a protected environment, not accessible to the attacker, to perform robustification (light blue rectangles in Fig. 2). More details and limitations of our threat model are discussed in the Supplementary, whereas here we first introduce useful definitions and then provide several recipes for the use of \\( A^5 \\) in different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.694, + 0.891, + 0.724 + ], + "angle": 0, + "content": "Definitions For a classifier \\(C\\) and input \\(\\mathbf{x}\\), \\(\\mathbf{y} = C(\\mathbf{x})\\) is the logit vector output by \\(C\\). Under attack, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.737, + 0.892, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} + \\delta \\boldsymbol {y} = C (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {A}), \\| \\left| \\delta \\boldsymbol {x} _ {A} \\right| | _ {p} < \\epsilon_ {A}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon_{A}\\) is the magnitude of the attack \\(\\delta \\pmb{x}_A\\) in \\(p\\)-norm and \\(\\delta \\pmb{y}\\) is the logit perturbation. Linear relaxation [43] estimates the lower \\((\\pmb{y}^l)\\) and upper \\((\\pmb{y}^u)\\) bounds of \\(\\pmb{y} + \\delta \\pmb{y}\\), for instance using IBP [12], CROWN [43], CROWN-IBP [42], or a combination of them. The margin for the class \\(j\\) is the difference between the lower bound of the ground truth class \\(j^{*}\\) and the upper bound of \\(j\\), while it is zero by definition for \\(j^{*}\\). More formally, for \\(M\\) classes the elements of the margin vector \\(\\pmb{m} = [m_0, m_1, \\dots, m_{M - 1}]\\) are: \\(m_{j^{*}} = 0\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.408, + 0.9 + ], + "angle": 0, + "content": "Automatic Linear Relaxation based Perturbation Analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4069" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.795, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.895, + 0.214 + ], + "angle": 0, + "content": "Figure 2. Schematic representation of the full \\(A^5\\) framework. It allows the robust augmentation of both physical objects \\(\\boldsymbol{w}\\) and acquired data \\(\\boldsymbol{x}\\) to make them certifiably robust against MitM and physical (not tested here) adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.47, + 0.271 + ], + "angle": 0, + "content": "and \\(m_j = \\pmb{y}_{j^*}^l -\\pmb{y}_j^u\\) \\(\\forall j\\neq j^{*}\\) . The confidence is obtained by applying softmax to \\(\\pmb{m}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.282, + 0.47, + 0.325 + ], + "angle": 0, + "content": "\\[\np _ {j} (\\boldsymbol {x}) = e ^ {- m _ {j}} / \\sum_ {k = 0} ^ {M - 1} e ^ {- m _ {k}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.384, + 0.352 + ], + "angle": 0, + "content": "and thus the worst case cross entropy \\( E(\\pmb{x}) \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.364, + 0.469, + 0.407 + ], + "angle": 0, + "content": "\\[\nE (\\boldsymbol {x}) = - \\sum_ {j = 0} ^ {M - 1} \\hat {y} _ {j} \\log \\left(p _ {j} (\\boldsymbol {x})\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.419, + 0.47, + 0.451 + ], + "angle": 0, + "content": "where \\(\\hat{y}_j\\) is the one-hot-encoding of \\(j^*\\). To formally introduce \\(A^5\\), we first give the following definitions:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.461, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Definition 1 (\\(\\epsilon_A\\)-robustness). Given an input \\(\\pmb{x}\\) and a classifier \\(C\\), we say that \\(\\pmb{x}\\) is \\(\\epsilon_A\\)-robust with respect to \\(C\\) in norm \\(p\\) if \\(\\forall \\delta \\pmb{x}_A, ||\\delta \\pmb{x}_A||_p < \\epsilon_A\\), \\(C(\\pmb{x} + \\delta \\pmb{x}_A)\\) generates the correct classification output." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Definition 2 (robustification). For a given input \\(\\pmb{x}\\) and classifier \\(C\\), robustification is the process of finding a defensive augmentation \\(\\delta \\pmb{x}_D\\), \\(||\\delta \\pmb{x}_D||_p < \\epsilon_D\\), such that \\(\\pmb{x} + \\delta \\pmb{x}_D\\) is \\(\\epsilon_A\\)-robust with respect to \\(C\\) in norm \\(p\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.697 + ], + "angle": 0, + "content": "We compute certified CROWN-IBP bounds through auto_LiRPA [37] and optimize \\( E(\\pmb{x} + \\delta \\pmb{x}_D) \\) to find preemptive, defensive perturbations \\( \\delta \\pmb{x}_D \\) that make \\( \\pmb{x} + \\delta \\pmb{x}_D \\in \\epsilon_A \\)-robust, following the training recipes provided here. We also show how to find defensive perturbations \\( \\delta \\pmb{w}_D \\) for physical objects \\( \\pmb{w} \\) using the same approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Offline robustification with ground truth \\((A^5 /O)\\): Consider a legacy classifier \\(C\\) that has been certifiably trained (e.g. through CROWN-IBP [42]) against attacks of magnitude \\(\\epsilon_A^C\\). Given an input \\(\\pmb{x}\\) and its class \\(j^{*}\\), we want to find a defensive perturbation \\(\\delta \\pmb{x}_D,||\\delta \\pmb{x}_D||_p < \\epsilon_D\\), such that \\(\\pmb{x} + \\delta \\pmb{x}_D\\) is \\(\\epsilon_A^R\\)-robust with respect to \\(C\\) in \\(p\\)-norm. Our solution to this problem is \\(A^5\\)/Offline (or \\(A^5 /O\\))." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.469, + 0.854 + ], + "angle": 0, + "content": "To force \\(||\\delta \\pmb{x}_D||_p < \\epsilon_D\\), we parameterize \\(\\delta \\pmb{x}_D\\) via a vector \\(\\pmb{z}\\) to which we apply an element-wise sigmoid:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.866, + 0.469, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\delta \\boldsymbol {x} _ {D} = \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) = [ \\delta x _ {D} (\\boldsymbol {z}) _ {0}, \\dots , \\delta x _ {D} (\\boldsymbol {z}) _ {K - 1} ] \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\delta x _ {D} (\\boldsymbol {z}) _ {j} = 2 \\epsilon_ {D} \\left[ 1 / \\left(1 + e ^ {- z _ {j}}\\right) - 0. 5 \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.239, + 0.892, + 0.286 + ], + "angle": 0, + "content": "where \\(K\\) is the number of elements in \\(\\pmb{x}\\). This is a robust classification problem where \\(\\pmb{z}\\) is the only unknown. For any \\(\\pmb{x}\\), we minimize the worst case cross entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.299, + 0.892, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} = \\operatorname {a r g m i n} _ {\\tilde {\\boldsymbol {z}}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\tilde {\\boldsymbol {z}}) ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.892, + 0.401 + ], + "angle": 0, + "content": "using RMSProp. Notice that the magnitude of the attack used during robustification \\((\\epsilon_A^R)\\) and that adopted to train \\(C\\) \\((\\epsilon_A^C)\\) can be different. Furthermore, similarly to adversarial training and certified methods, these can also be different from the target attack magnitude, \\(\\epsilon_{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.893, + 0.569 + ], + "angle": 0, + "content": "\\(A^5 / O\\) is similar in spirit to preemptive image robustification [17], but does not need \\(C\\) to run on the clean input before robustification. Although \\(A^5 / O\\) could be used in a similar way to robustify \\(\\pmb{x}\\) accordingly to the class \\(C(\\pmb{x})\\), running \\(C\\) first and then iteratively solving the optimization problem in Eq. 6 makes it hardly usable in practice. Therefore, it does not directly serve any practical purpose, but it helps answering important theoretical questions (e.g., under which conditions can we robustify \\(\\pmb{x}\\)? What does a robustified image looks like?) and establishing a baseline with known ground truth and no computational constraints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.893, + 0.741 + ], + "angle": 0, + "content": "Online robustification with robustifier \\( R \\), legacy \\( C \\) (\\( A^5 / R \\)): We introduce a second recipe, \\( A^5 / R \\), where we train a robustifier DNN \\( R \\) that takes in input \\( \\mathbf{x} \\), no ground truth label, and outputs \\( \\mathbf{z} \\), from which a defensive perturbation \\( \\delta \\mathbf{x}_D \\), \\( ||\\delta \\mathbf{x}_D||_p < \\epsilon_D \\), is computed as in Eq. 5. We assume again that a robust legacy \\( C \\) is given and we train \\( R \\) to make each \\( \\mathbf{x} + \\delta \\mathbf{x}_D \\) in the training dataset \\( \\epsilon_A^R \\)-robust with respect to \\( C \\) in \\( p \\)-norm. More formally, the weights \\( \\theta_R \\) of \\( R \\) are found by minimization of the average worst case entropy on the training dataset \\( E(\\theta_R) \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.754, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} = R (\\boldsymbol {x} \\mid \\boldsymbol {\\theta} _ {R}), E (\\boldsymbol {\\theta} _ {R}) = (1 / N) \\sum_ {\\boldsymbol {x}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(N\\) is the dataset size. Unlike preemptive robustification [17], \\(R\\) does not iteratively solve a complex optimization problem, neither we need running \\(C\\) before \\(R\\). Therefore, we can use \\(A^5 / R\\) to protect \\(\\pmb{x}\\) against \\(\\mathrm{MitM}\\) adversarial attacks on-the-fly, ignoring its class, soon after its acquisition, and before its transmission to \\(C\\) (see \\(A^5 / R\\) in Supplementary). Furthermore, since \\(R\\) is trained for a specific" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4070" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.138 + ], + "angle": 0, + "content": "\\(C\\), it can be deployed on the field (e.g., into the firmware of the acquisition devices) without changing (and thus maintaining the legacy of) the classifier infrastructure." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.47, + 0.232 + ], + "angle": 0, + "content": "Online robustification with robustifier \\( R \\), retrained \\( C \\) (\\( A^5 / RC \\)): The \\( A^5 / RC \\) recipe is equivalent to \\( A^5 / R \\), but it leverages the co-adaptation of \\( C \\) and \\( R \\) by joint training (see \\( A^5 / RC \\) in Supplementary). The cost function is again the average worst case entropy \\( E(\\theta_C, \\theta_R) \\), defined now as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.242, + 0.469, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} = R (\\boldsymbol {x} | \\boldsymbol {\\theta} _ {R}), \\quad \\boldsymbol {y} = C (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} | \\boldsymbol {\\theta} _ {C}), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.26, + 0.469, + 0.291 + ], + "angle": 0, + "content": "\\[\nE \\left(\\boldsymbol {\\theta} _ {C}, \\boldsymbol {\\theta} _ {R}\\right) = (1 / N) \\sum_ {\\boldsymbol {x}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) ], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.299, + 0.47, + 0.451 + ], + "angle": 0, + "content": "where \\(\\theta_{C}\\) are the parameters of \\(C\\). Training \\(R\\) and \\(C\\) from scratch is often unstable and requires careful tuning to achieve convergence (likewise robust classifiers [12, 42]). Therefore, we first train \\(R\\) with \\(A^5 / C\\) and then fine tune \\(R\\) and \\(C\\) together. During the tuning phase, we can use random input augmentation to prevent overfitting (see Section 4 and Supplementary for details). Differently from \\(A^5 / R\\), \\(A^5 / RC\\) does not preserve the legacy classifier \\(C\\), but it achieves the same scope (with more effectiveness) in terms of protection against MitM attacks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.47, + 0.59 + ], + "angle": 0, + "content": "Offline physical robustification with \\((A^5 /PC)\\) and without \\((A^5 /P)\\) retraining \\(C\\), with ground truth: This recipe aims at producing physical objects that are certifiably robust against MitM attacks (see \\(A^5 /P\\), \\(A^5 /PC\\) in Supplementary). Similarly to unadversarial objects [24], we assume that the system creator can design the entire infrastructure and the attacker cannot interfere in this phase. The problem is formally defined for \\(A^5 /PC\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.6, + 0.469, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} = A (\\boldsymbol {w} + \\delta \\boldsymbol {w} _ {D}), \\| \\delta \\boldsymbol {w} _ {D}) \\| < \\epsilon_ {D}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.619, + 0.469, + 0.65 + ], + "angle": 0, + "content": "\\[\nE \\left(\\boldsymbol {\\theta} _ {C}, \\delta \\boldsymbol {w} _ {D}\\right) = (1 / N) \\sum_ {\\boldsymbol {x}} E (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D}) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.47, + 0.854 + ], + "angle": 0, + "content": "where \\(\\pmb{w}\\) is the object appearance, whereas \\(\\pmb{x} = A(\\pmb{w})\\) is the data acquisition model (e.g., image formation). \\(N\\) is here the number of random samples generated as \\(\\pmb{x} + \\delta \\pmb{x}_D = A(\\pmb{w} + \\delta \\pmb{w}_D)\\); it is important noticing in fact that the acquisition process is not deterministic: as an exemplary case we mention OCR, where \\(\\pmb{w}\\) is the shape of a character, while \\(A\\) may include random perspective transformations, contrast and brightness changes, noise addition and blurring, that may occur on a real camera. The defensive perturbation \\(\\delta \\pmb{w}_D\\) represents a change in the physical character appearance \\(\\pmb{w}\\) that decreases the success rate of an attack towards the acquired image \\(\\pmb{x}\\) (therefore after framing the character with \\(A\\)) and the classifier \\(C\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In \\( A^5 / P \\) we change the character shapes and use the legacy classifier \\( C \\); in \\( A^5 / PC \\) we also train \\( C \\) while estimating the robust characters \\( \\boldsymbol{w} + \\delta \\boldsymbol{w} \\). In both cases \\( C \\)" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.136 + ], + "angle": 0, + "content": "
εAεAC=0.0εAC=0.1εAC=0.2εAC=0.3εAC=0.4
0.1——1.17 [—, 2.24] [42]——
0.10.81 [89.02, 100.00]0.92 [2.00, 5.41]1.32 [2.14, 2.60]1.44 [2.33, 2.63]2.42 [3.36, 3.41]
0.3———1.82 [—, 7.02] [42]—
0.30.69 [100.00, 100.00]0.94 [98.41, 100.00]1.30 [99.62, 100.00]1.41 [7.06, 9.98]2.58 [7.03, 7.68]
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.146, + 0.892, + 0.216 + ], + "angle": 0, + "content": "Table 2. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST classifiers trained by us with CROWN-IBP [42], under attack \\(\\epsilon_{A} = \\{0.1,0.3\\}\\); we also show the metrics reported in [42]. A \\(100\\%\\) certified error rate identifies DNNs that are not protected against attacks." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.229, + 0.892, + 0.314 + ], + "angle": 0, + "content": "
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.79 [87.09, 100.00]0.82 [1.51, 2.89]1.22 [1.64, 1.86]1.34 [1.76, 1.87]2.38 [2.91, 2.92]
0.10.100.79 [86.51, 100.00]0.72 [1.21, 1.92]1.15 [1.43, 1.49]1.34 [1.50, 1.58]2.41 [2.71, 2.74]
0.10.200.89 [86.40, 100.00]0.56 [0.78, 1.04]1.13 [1.18, 1.21]1.35 [1.37, 1.38]2.40 [2.55, 2.55]
0.10.300.93 [86.44, 100.00]0.51 [0.57, 0.70]1.14 [1.16, 1.17]1.34 [1.37, 1.37]2.38 [2.52, 2.52]
0.10.400.99 [85.93, 100.00]0.52 [0.53, 0.60]1.14 [1.16, 1.16]1.35 [1.38, 1.38]2.44 [2.51, 2.51]
0.30.050.71 [100.00, 100.00]0.93 [97.18, 100.00]1.34 [99.17, 100.00]1.35 [5.16, 6.97]2.55 [6.00, 6.42]
0.30.100.72 [100.00, 100.00]0.88 [95.32, 100.00]1.36 [98.58, 100.00]1.33 [4.12, 5.19]2.49 [5.54, 5.85]
0.30.200.72 [100.00, 100.00]0.88 [90.62, 100.00]1.41 [96.21, 100.00]1.32 [3.21, 3.77]2.35 [5.02, 5.22]
0.30.300.73 [100.00, 100.00]0.89 [86.28, 100.00]1.41 [94.18, 100.00]1.31 [2.91, 3.40]2.32 [4.66, 4.81]
0.30.400.81 [100.00, 100.00]0.85 [83.49, 100.00]1.42 [92.63, 100.00]1.32 [2.74, 3.20]2.28 [4.16, 4.29]
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.325, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Table 3. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, \\( A^5 / O \\), under attack \\( \\epsilon_A = \\{0.1, 0.3\\} \\). During training we use \\( \\epsilon_A^R = \\epsilon_A \\)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.38, + 0.892, + 0.466 + ], + "angle": 0, + "content": "
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.82 [85.50, 100.00]0.95 [1.71, 3.58]1.30 [1.82, 2.02]1.40 [1.86, 2.03]2.43 [2.96, 2.97]
0.10.100.81 [82.87, 100.00]1.00 [1.48, 2.54]1.19 [1.62, 1.73]1.40 [1.68, 1.74]2.38 [2.77, 2.79]
0.10.200.88 [78.48, 100.00]0.93 [1.33, 1.77]1.20 [1.42, 1.52]1.31 [1.53, 1.59]2.26 [2.55, 2.57]
0.10.300.99 [75.89, 100.00]0.91 [1.16, 1.54]1.14 [1.31, 1.36]1.26 [1.43, 1.47]2.20 [2.42, 2.43]
0.10.401.04 [75.69, 100.00]0.87 [1.05, 1.35]1.14 [1.27, 1.33]1.24 [1.43, 1.46]2.22 [2.41, 2.43]
0.30.050.69 [100.00, 100.00]0.94 [96.93, 100.00]1.33 [99.11, 100.00]1.41 [5.13, 6.91]2.58 [5.77, 6.09]
0.30.100.71 [100.00, 100.00]0.91 [95.09, 100.00]1.33 [98.47, 100.00]1.38 [3.86, 4.81]2.55 [4.71, 4.83]
0.30.200.72 [100.00, 100.00]0.92 [90.13, 100.00]1.37 [96.08, 100.00]1.41 [2.26, 2.63]2.55 [3.33, 3.40]
0.30.300.75 [100.00, 100.00]0.93 [85.54, 100.00]1.35 [93.36, 100.00]1.32 [1.86, 1.99]2.37 [2.86, 2.88]
0.30.400.83 [100.00, 100.00]0.96 [82.48, 100.00]1.36 [92.15, 100.00]1.29 [1.69, 1.83]2.35 [2.67, 2.69]
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "Table 4. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, \\( A^5 / R \\), under attack \\( \\epsilon_A = \\{0.1, 0.3\\} \\). During training we use \\( \\epsilon_A^R = \\epsilon_A \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.545, + 0.892, + 0.62 + ], + "angle": 0, + "content": "is previously trained to be robust, e.g., through CROWN-IBP [42]. The \\( A^5 / P \\) and \\( A^5 / PC \\) recipes find application in all those situations where the developer can design the entire infrastructure; beyond OCR, other examples are road sign or robotic tool design and classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.712, + 0.65 + ], + "angle": 0, + "content": "4. Results and discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We test \\( A^5 / O \\), \\( A^5 / R \\), and \\( A^5 / RC \\) on MNIST [10], CIFAR10 [14], FashionMNIST [34], and Tinyimageenet [15], that have been widely used to establish significant milestones in the adversarial defense field. Training schedules and DNN architectures are given in the Supplementary. In all cases we use a \\( p = \\infty \\) norm for the attack and defense. Our intention is to show the improvement achieved by \\( A^5 \\) over traditional, certified methods like CROWN-IBP. The reader should anyway keep in mind that the direct comparison of \\( A^5 \\) and CROWN-IBP is partially unfair, as CROWN-IBP does not apply any preemptive defense." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We first use CROWN-IBP to train five certifiably robust MNIST classifiers with varying levels of robustness, for training attacks \\(\\epsilon_A^C = 0.0\\) (unprotected \\(C\\)) and \\(\\epsilon_A^C = \\{0.1, 0.2, 0.3, 0.4\\}\\). Their clean and certified errors are reported in Table 2, for attacks \\(\\epsilon_A = \\{0.1, 0.3\\}\\). We also" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4071" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.088, + 0.468, + 0.173 + ], + "angle": 0, + "content": "
εA=εCεDεCA=0.0εCA=0.1εCA=0.2εCA=0.3εCA=0.4
0.10.0588.66[88.66, 88.66]1.00[1.38, 2.35]1.16[1.61, 2.46]1.14[1.54, 2.29]1.37[1.72, 2.33]
0.10.1088.66[88.66, 88.66]0.77[0.95, 1.40]0.99[1.16, 1.56]0.86[1.03, 1.42]1.13[1.31, 1.72]
0.10.2088.66[88.66, 88.66]0.69[0.80, 1.05]0.84[0.94, 1.25]0.96[1.07, 1.33]0.88[1.10, 1.54]
0.10.3088.66[88.66, 88.66]0.64[0.70, 1.02]0.91[0.99, 1.16]0.79[0.88, 1.21]1.00[1.12, 1.43]
0.10.4088.66[88.66, 88.66]0.61[0.68, 0.88]0.93[0.99, 1.18]0.69[0.76, 1.10]1.00[1.03, 1.30]
0.30.0588.66[88.66, 88.66]88.66[88.66, 88.66]2.57[8.35, 10.86]3.42[9.51, 11.72]8.00[18.05, 19.77]
0.30.1088.66[88.66, 88.66]88.66[88.66, 88.66]2.07[5.66, 7.56]2.35[6.18, 8.07]2.35[6.31, 7.93]
0.30.2088.66[88.66, 88.66]1.68[3.11, 4.79]1.40[2.42, 3.62]1.28[2.82, 4.01]1.88[3.53, 4.90]
0.30.3088.66[88.66, 88.66]0.84[0.96, 1.38]1.00[1.14, 1.36]0.94[1.15, 1.59]1.07[1.39, 1.90]
0.30.4088.66[88.66, 88.66]0.72[0.80, 1.03]0.75[0.82, 1.06]0.91[1.14, 1.68]0.87[1.14, 1.50]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.24 + ], + "angle": 0, + "content": "Table 5. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST, \\( A^5 / RC \\), under attack \\( \\epsilon_A = \\{0.1, 0.3\\} \\). During training we use \\( \\epsilon_A^R = \\epsilon_A \\), and data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.468, + 0.359 + ], + "angle": 0, + "content": "report the error rates estimated with autoattack [8, 9], an automatic toll based on an ensemble of four attacks to evaluate the practical DNN robustness. These error rates are in the same ballpark of those reported in the CROWN-IBP paper [42]: they represent the baseline for the analysis of the different \\( A^5 \\) recipes considered here." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.468, + 0.708 + ], + "angle": 0, + "content": "For each \\(C\\) we run \\(A^5 /O\\) on the MNIST test set, for defense magnitudes \\(\\epsilon_{D} = \\{0.05,0.1,0.2,0.3,0.4\\}\\). During the optimization of the entropy in Eq. 6, we use attacks \\(\\epsilon_A^R = \\{0.1,0.3\\}\\), not necessarily equal to \\(\\epsilon_A^C\\). Table 3 reports the clean, autoattack, and certified errors for a testing attack \\(\\epsilon_{A} = \\epsilon_{A}^{R}\\). When \\(C\\) is not robust (\\(\\epsilon_{A}^{C} = 0.0\\)), \\(A^5 /O\\) does not reduce the certified error², otherwise it generally overcomes the base CROWN-IBP classifier, as shown by comparing against the error rates in Table 2. As a rule of thumb, the best clean / certified error trade off is often achieved for \\(\\epsilon_{A}^{R} = \\epsilon_{A}^{C} = \\epsilon_{A}\\). Results also improve, as expected, for larger defenses \\(\\epsilon_{D}\\), when \\(A^5 /O\\) beat CROWN-IBP by a significant margin on clean and certified errors. However, \\(A^5 /O\\) does not work well when the legacy \\(C\\) is trained with a large \\(\\epsilon_{A}^{C}\\), possibly because of its initial low clean accuracy. The results are consistent with our interpretation of the effects of adversarial training and robustification on the classification landscape (Fig. 1): training \\(C\\) with a large \\(\\epsilon_{A}^{C}\\) creates classification valleys wide enough to accommodate \\(\\pmb{x} + \\delta \\pmb{x}_D + \\delta \\pmb{x}_A\\), rendering the adversarial attacks ineffective. \\(A^5\\) moves any \\(\\pmb{x}\\) towards the centers of the valleys; a large \\(\\epsilon_{D}\\) allows moving them more easily. Too large \\(\\epsilon_{A}^{C}\\) remains however detrimental for the clean accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.468, + 0.829 + ], + "angle": 0, + "content": "Table 4 reports results for \\( A^5 / R \\), demonstrating that training a DNN robustifier \\( R \\) for on-the-fly, certified robustification, ignoring the ground truth, is feasible. Likewise \\( A^5 / O \\), \\( A^5 / R \\) beats the base CROWN-IBP classifier by a significant margin. Sometimes \\( A^5 / O \\) does better by leveraging the knowledge of the ground truth class, whereas the regular robustification signal generated by \\( A^5 / R \\) is more effective in other cases." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.468, + 0.862 + ], + "angle": 0, + "content": "Our last MNIST experiment investigates \\(A^5 /RC\\) that does not preserve the legacy \\(C\\) like \\(A^5 /R\\), but achieves bet" + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.088, + 0.892, + 0.146 + ], + "angle": 0, + "content": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2554/25552.75 [63.47, 64.09]53.34 [62.42, 62.79]53.75 [64.92, 65.49]
8/2558/25551.06 [60.04, 60.73]52.36 [59.24, 59.58]53.16 [63.50, 64.03]
8/25516/25548.72 [54.89, 55.25]50.60 [55.22, 55.38]52.23 [60.79, 61.14]
8/25532/25545.68 [49.51, 49.74]48.06 [50.86, 50.94]50.99 [57.37, 57.61]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.157, + 0.892, + 0.24 + ], + "angle": 0, + "content": "Table 6. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, \\( A^5 / O \\), under attack \\( \\epsilon_A = 8/255 \\). During training we use different \\( \\epsilon_A^R \\) and \\( \\epsilon_D \\). The robust classifier trained with CROWN-IBP has clean [and certified] errors equal to 54.02 [66.94] or 45.47 [69.55], depending on the CROWN-IBP training configuration [42]." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.254, + 0.892, + 0.327 + ], + "angle": 0, + "content": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25554.60 [66.34, 67.11]54.60 [66.34, 67.11]54.60 [66.31, 67.11]
8/2554/25553.73 [64.07, 64.57]54.12 [63.32, 63.74]54.37 [63.58, 64.03]
8/2558/25552.91 [62.59, 63.12]53.24 [61.38, 61.93]54.54 [60.85, 61.16]
8/25516/25551.58 [60.40, 61.02]51.95 [59.21, 59.67]54.44 [59.87, 60.11]
8/25532/25550.91 [57.26, 57.75]51.70 [56.91, 57.43]54.05 [58.39, 58.73]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Table 7. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, \\( A^5 / R \\), under attack \\( \\epsilon_A = 8 / 255 \\). During training we use different \\( \\epsilon_A^R \\) and \\( \\epsilon_D \\). The robustifier \\( R \\) in the first row uses \\( \\epsilon_D = 0 \\) and thus it does not provide any form of defense — the error reported here is equivalent to that of a robust classifier trained with CROWN-IBP [42]." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.434, + 0.892, + 0.507 + ], + "angle": 0, + "content": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25550.24 [69.17, 73.65]58.63 [69.25, 70.63]74.28 [77.82, 77.91]
8/2554/25546.76 [63.03, 70.99]54.05 [62.80, 64.44]67.59 [71.20, 71.32]
8/2558/25541.24 [54.01, 64.31]50.74 [56.63, 59.49]65.96 [67.63, 67.78]
8/25516/25536.96 [42.45, 52.90]39.15 [41.33, 44.22]58.17 [58.67, 58.85]
8/25532/25535.26 [37.13, 42.76]41.59 [43.01, 45.55]42.25 [43.02, 44.08]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.517, + 0.892, + 0.601 + ], + "angle": 0, + "content": "Table 8. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, \\( A^5 / RC \\), under attack \\( \\epsilon_A = 8/255 \\). During training we use different \\( \\epsilon_A^R \\) and \\( \\epsilon_D \\). The robustifier \\( R \\) in the first row uses \\( \\epsilon_D = 0 \\) and thus it does not provide any form of defense, whereas the corresponding robust CROWN-IBP classifier is fine tuned using \\( \\epsilon_A^C = \\epsilon_R \\) during training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.891, + 0.719 + ], + "angle": 0, + "content": "ter clean accuracy and protection (Table 5). In the Supplementary we discuss the role of data augmentation to achieve this result. Interestingly, the best results are obtained when \\( C \\) is initially trained for \\( \\epsilon_A^C \\leq \\epsilon_A \\), probably because \\( C \\) has an initial high clean accuracy and subsequently co-adapts with \\( R \\) to also guarantee a small certified error." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Fig. 3 shows examples of robustification on MNIST. At visual inspection, \\( A^5 / O \\), \\( A^5 / R \\), and \\( A^5 / RC \\) act similarly on these images: they add contrast and enhance the high frequency features (i.e., edges) of the digits. The few failure cases are often pictures that look ambiguous even to a human observer (right panels in Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We perform tests on CIFAR10 and compare \\( A^5 \\) against a state of the art CROWN-IBP classifiers \\( C \\), trained with \\( \\epsilon_A^C = 8.8 / 255 \\), that we also adopt as legacy \\( C \\) for \\( A^5 \\). To test the effect of the magnitude of the attack used for robustification (with \\( A^5 / O \\)) or for training \\( R \\) (for \\( A^5 / R, A^5 / RC \\)), we use \\( \\epsilon_A^R = \\{4 / 255, 8 / 255, 16 / 255\\} \\) that are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "2Coherently with [24], the worst case cross entropy improves, but not enough to affect the ranking of the classified classes under attack." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.088, + 0.825, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.664, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Figure 3. Successful and failed robustifications for \\( A^5 / O \\), \\( A^5 / R \\), and \\( A^5 / RC \\), for MNIST (\\( \\epsilon_A = \\epsilon_A^C = \\epsilon_A^R = \\epsilon_D = 0.3 \\)) and CIFAR10 (\\( \\epsilon_A = 8.8 / 255 \\), \\( \\epsilon_A^C = 8 / 255 \\), \\( \\epsilon_A^R = 4 / 255 \\), and \\( \\epsilon_D = 32 / 255 \\)). Each triplet shows the original (left) and robustified (center) images; the rightmost panel is the defensive augmentation. All the \\( A^5 \\) recipes consistently increase the contrast and the high frequency content on MNIST, where failure cases are ambiguous even for a human observer. For CIFAR10, \\( A^5 / O \\) leverage isolated pixel changes, whereas \\( A^5 / R \\) and \\( A^5 / RC \\) consistently increase the color contrast and saturation, similarly to results in literature [17]." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.75, + 0.891, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.802, + 0.895, + 0.832 + ], + "angle": 0, + "content": "Figure 4. The first, second and third row show respectively the robustified \\((\\boldsymbol{w} + \\delta \\boldsymbol{w}_D)\\), nominal \\((\\boldsymbol{w})\\) and robustification \\((\\delta \\boldsymbol{w}_D)\\) sets of 62 letters considered for testing \\(A^5 / PC\\). The last row shows examples of the robustified letters after framing them, \\(\\boldsymbol{x} + \\delta \\boldsymbol{x}_D = A(\\boldsymbol{w} + \\delta \\boldsymbol{w}_D)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.858, + 0.47, + 0.889 + ], + "angle": 0, + "content": "smaller, equal or greater than \\(\\epsilon_A^C\\). Results in Table 6 are consistent with those measured for MNIST: \\(A^5 / O\\) signifi" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.858, + 0.892, + 0.889 + ], + "angle": 0, + "content": "cantly improves the clean and certified error rates over the base CROWN-IBP classifier. Training \\( R \\) through \\( A^5 / R \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4073" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.139, + 0.089, + 0.411, + 0.168 + ], + "angle": 0, + "content": "
AlgoError [certified error]
Vanilla0.89% [100.00%]
CROWN-IBP3.85% [13.85%]
A5/P3.08% [11.84%]
A5/PC0.73% [4.20%]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.178, + 0.472, + 0.234 + ], + "angle": 0, + "content": "Table 9. Error and certified error for the classification of the 62 characters in Fig. 4, for \\(\\epsilon_A = 0.1\\) and different algorithms. \\(A^5 /P\\) changes the physical shape of the letters to achieve robustness after framing them; \\(A^5 /PC\\) also trains the associated classifier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.47, + 0.52 + ], + "angle": 0, + "content": "produces a smaller improvement, still sufficient to beat the CROWN-IBP classifier by a significant margin (Table 7). The value of \\(\\epsilon_A^R\\) has a significant impact on the final result: if too large (\\(\\epsilon_A^R > \\epsilon_A\\)), the clean accuracy is penalized, because \\(A^5\\) looks for very large plateaus in the classification landscape, that are hard or impossible to find. In other words, we observe for \\(A^5\\) a trade-off between the clean and certified accuracy, likewise adversarial training and certified methods. Best results on CIFAR10 are achieved again by \\(A^5 / RC\\), when \\(\\epsilon_A^R\\) is equal or slightly smaller than \\(\\epsilon_A = \\epsilon_A^C\\). The improvements in clean and certified error are in this case in the impressive range of \\(19\\%\\) and \\(24\\%\\) respectively (Table 8). We also notice here that the theoretical certified error rates measured with auto_LiRPA [37] in \\(A^5\\) and those measured experimentally by autoattack [8, 9] are close for \\(A^5\\), whose working point at convergence may favor the consistency between these two estimates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Also for CIFAR10 we perform a visual inspection of defensive augmentation (Fig. 3). \\( A^5 / O \\) exploits a slight contrast increase and modifies isolated pixels: likewise adversarial attacks with small \\( p \\) norm (up to single pixel attacks [29]), the same seems possible when crafting defensive perturbations. Preemptive robustification [17] shows a similar pattern (see their Fig. 6): non natural textures and enhanced color saturation emerge in their images for large defensive perturbations. The robust images crafted by \\( R \\) in \\( A^5 / R \\) or \\( A^5 / RC \\), however, are different: they generally show increased contrast and color saturation, without the textures generated by \\( A^5 / O \\) and [17]. We speculate that these textures may be associated with the iterative optimization in \\( A^5 / O \\) and [17], that may both suffer some form of obfuscated gradient [2]; a more regular defense is instead generated by \\( R \\) without resorting to any iterative process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Additional results on FashionMNIST and Tinyimagenet, overall consistent with the ones presented here and showing the quantitative advantages provided by \\( A^5 \\) on these datasets, are reported in the Supplementary. These experiments highlight two important facts. The first one is that the integration of the design philosophy of \\( A^5 \\) with that of other recent ideas in the field of adversarial defense (like the \\( \\ell_{\\infty} \\)-dist neurons that implement 1-Lipschitz functions and are inherently robust to adversarial attacks) may lead to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "even bigger improvements in terms of certified robustness. The second is that the integration of different defense strategies may indeed be strictly needed to guarantee further progresses, as scaling to very large dataset remains problematic even for a preemptive robustification algorithms like \\( A^5 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.893, + 0.351 + ], + "angle": 0, + "content": "Finally, for \\( A^5 / P \\) and \\( A^5 / PC \\), we test the classification of an alphabet of 62 characters (Fig. 4), after random rotation, shift, perspective distortion, noise addition, blurring and color jittering (details in Supplementary), that simulate the scanning of a document for OCR. Table 9 shows the clean and certified errors for: a vanilla, non robust classifier; a CROWN-IBP classifier; \\( A^5 / P \\); and \\( A^5 / PC \\). \\( A^5 \\) achieves again a significant improvement both on clean and certified errors: \\( A^5 / PC \\) has better (0.73%) accuracy of the vanilla \\( C \\) (0.89%) and can be attacked in only 4.20% of the cases. This results is obtained with a slight modification of the shape of the prototype characters shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.371, + 0.743, + 0.387 + ], + "angle": 0, + "content": "5. Discussion and conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.893, + 0.595 + ], + "angle": 0, + "content": "We introduce \\( A^5 \\), the first framework that leverages adversarial augmentation to preemptively modify the input of a DNN or a physical object, to make it certifiable robust against adversarial attacks. \\( A^5 \\) is not simply complementary to other defense methods: its tights with them are strict and they can be used together to achieve higher robustness. For instance, one of the limitations we found is that \\( A^5 / R \\) requires a somehow robust initial classifier \\( C \\). Co-training \\( R \\) and a poorly robust \\( C \\) easily compensates for the initially low robustness of \\( C \\) and boosts its high clean accuracy while also achieving state of the art results in terms of certified accuracy; finding the optimal initial robustness of \\( C \\) remains an open question though." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Here we do not investigate all the possible \\( A^5 \\) recipes: in its most general implementation \\( (A^5 / PRC) \\) the physical objects, \\( R \\) and \\( C \\) could all be optimized at the same time. This opportunity can be explored through the released code at https://github.com/NVlabs/A5. This will also serve answering other questions that do not find space here: for instance, we do not know if robustification generalizes to other classifiers as adversarial attacks do, how \\( A^5 \\) works in norms other than \\( p = \\infty \\), and how much it can further improve if coupled with state-of-the-art solutions like the recently proposed \\( \\ell_{\\infty} \\)-dist neurons [41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Different \\(A^5\\) recipes find applications in scenarios where the user can control the acquisition device, equip it with \\(R\\) and guarantee its protection (e.g., image acquisition on a phone before server communication) or while designing the infrastructure (e.g., robust road signs creation); it may be not suitable for more general adversarial scenarios. Overall, we believe that the practical deployment of robust systems may benefit from (or even require) methods for both robust classifiers and preemptive robustification." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4074" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Maungmaung Aprilpyone and Hitoshi Kiya. Encryption inspired adversarial defense for visual classification. 2020 IEEE International Conference on Image Processing (ICIP), pages 1681-1685, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 274-283. PMLR, 10-15 Jul 2018. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.298 + ], + "angle": 0, + "content": "[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[4] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 4312-4321. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Survey Track. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[5] Tom Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.59 + ], + "angle": 0, + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.634, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[9] Francesco Croce and Matthias Hein. Mind the box: \\( l_{1} \\)-apgd for sparse adversarial attacks on image classifiers. In ICML, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.717 + ], + "angle": 0, + "content": "[10] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.83 + ], + "angle": 0, + "content": "[12] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Mann, and Pushmeet Kohli. On the effectiveness of interval bound propagation for training verifiably robust models, 2018. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[13] Sonal Joshi, Jesús Villalba, Piotr Želasko, Laureano MoroVelázquez, and Najim Dehak. Study of pre-processing defenses against adversarial attacks on state-of-the-art speaker recognition systems. IEEE Transactions on Information Forensics and Security, 16:4811-4826, 2021. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[14] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[16] Xiaoting Li, Lingwei Chen, Jinquan Zhang, James Larus, and Dinghao Wu. Watermarking-based defense against adversarial attacks on deep neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[17] Seungyong Moon, Gaon An, and Hyun Oh Song. Preemptive image robustification for protecting users against man-in-the-middle adversarial attacks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning (ICML), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[19] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. pages 1491-1500, 10 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[20] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[21] Rafael Pinot, Raphael Ettedgui, Geovani Rizk, Yann Chevaleyre, and Jamal Atif. Randomization matters how to defend against strong adversarial attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7717-7727. PMLR, 13-18 Jul 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[22] Yao Qin, Nicholas Carlini, Ian J. Goodfellow, G. Cottrell, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. ArXiv, abs/1903.10346, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[23] Adnan Siraj Rakin, Zhezhi He, Boqing Gong, and Deliang Fan. Blind pre-processing: A robust defense method against adversarial examples. ArXiv, abs/1802.01549, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Fawkes: Protecting Privacy against Unauthorized Deep Learning Models. USENIX Association, USA, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4075" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[28] Jinmyeong Shin, Seok-Hwan Choi, and Yoon-Ho Choi. Is homomorphic encryption-based deep learning secure enough? Sensors, 21(23), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[29] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[30] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. Jan. 2014. 2nd International Conference on Learning Representations, ICLR 2014; Conference date: 14-04-2014 Through 16-04-2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[31] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1633-1645. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[32] Derui Wang, Chaoran Li, Sheng Wen, Surya Nepal, and Yang Xiang. Man-in-the-middle attacks against machine learning classifiers via malicious generative models. IEEE Transactions on Dependable and Secure Computing, 18(5):2074-2087, sep 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.468, + 0.488 + ], + "angle": 0, + "content": "[33] Shiqi Wang, Huan Zhang, Kaidi Xu, Xue Lin, Suman Jana, Cho-Jui Hsieh, and J Zico Kolter. Beta-CROWN: Efficient bound propagation with per-neuron split constraints for complete and incomplete neural network verification. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[34] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.533, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In International Conference on Learning Representations, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.591, + 0.468, + 0.645 + ], + "angle": 0, + "content": "[36] Kaidi Xu, Zhouxing Shi, Huan Zhang, Yihan Wang, Kai-Wei Chang, Minlie Huang, Bhavya Kailkhura, Xue Lin, and Cho-Jui Hsieh. Automatic perturbation analysis for scalable certified robustness and beyond. arXiv: Learning, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.648, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[37] Kaidi Xu, Huan Zhang, Shiqi Wang, Yihan Wang, Suman Jana, Xue Lin, and Cho-Jui Hsieh. Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers. In International Conference on Learning Representations, 2021. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[38] Hiromu Yakura and Jun Sakuma. Robust audio adversarial example for a physical attack. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5334–5341. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[39] Jianfei Yang, Han Zou, and Lihua Xie. Robustsense: Defending adversarial attack for secure device-free human activity recognition. ArXiv, abs/2204.01560, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[40] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of Machine Learning Research, pages 12062-12072. PMLR, 18-24 Jul 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[41] Bohang Zhang, Tianle Cai, Zhou Lu, Di He, and Liwei Wang. Towards certifying 1-infinity robustness using neural networks with 1-inf-dist neurons. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12368-12379. PMLR, 18-24 Jul 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[42] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane Boning, and Cho-Jui Hsieh. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[43] Huan Zhang, Tsui-Wei Weng, Pin-Yu Chen, Cho-Jui Hsieh, and Luca Daniel. Efficient neural network robustness certification with general activation functions. In Advances in Neural Information Processing Systems (NIPS), arXiv preprint arXiv:1811.00866, dec 2018. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4076" + } + ] +] \ No newline at end of file diff --git a/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_origin.pdf b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..19a9f4a3e881cfcc1230c41c8f2985240a4d43ff --- /dev/null +++ b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/2480b99e-0eac-4b8f-a85f-c4137baac3df_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47e4aba02c68d1d94f5e130430bd99b9c5194792e97192f5a33deb816f38c258 +size 2735262 diff --git a/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/full.md b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7c1a1024f44f87a17a89d07af8590cf85d346ad6 --- /dev/null +++ b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/full.md @@ -0,0 +1,294 @@ +# The Best Defense is a Good Offense: Adversarial Augmentation against Adversarial Attacks + +Iuri Frosio, NVIDIA + +ifrosio@nvidia.com + +Jan Kautz, NVIDIA + +jkautz@nvidia.com + +# Abstract + +Many defenses against adversarial attacks (e.g. robust classifiers, randomization, or image purification) use countermeasures put to work only after the attack has been crafted. We adopt a different perspective to introduce $A^5$ (Adversarial Augmentation Against Adversarial Attacks), a novel framework including the first certified preemptive defense against adversarial attacks. The main idea is to craft a defensive perturbation to guarantee that any attack (up to a given magnitude) towards the input in hand will fail. To this aim, we leverage existing automatic perturbation analysis tools for neural networks. We study the conditions to apply $A^5$ effectively, analyze the importance of the robustness of the to-be-defended classifier, and inspect the appearance of the robustified images. We show effective on-the-fly defensive augmentation with a robustifier network that ignores the ground truth label, and demonstrate the benefits of robustifier and classifier co-training. In our tests, $A^5$ consistently beats state of the art certified defenses on MNIST, CIFAR10, FashionMNIST and Tinyimagenet. We also show how to apply $A^5$ to create certifiably robust physical objects. Our code at https://github.com/NVlabs/A5 allows experimenting on a wide range of scenarios beyond the man-in-the-middle attack tested here, including the case of physical attacks. + +# 1. Introduction + +Since Deep Neural Networks (DNNs) have been found vulnerable to adversarial attacks [11, 30], researchers studied various protection strategies [4, 12, 20, 42, 43]. For instance, adversarial training [11, 30] generates attacks while asking a DNN for the correct output in training; it is simple, partially effective and widely adopted. Certified methods (e.g., IBP [12], CROWN [43], CROWN-IBP [42]) do a step more by estimating correct (although often pessimistic) output bounds (Fig. 1, a) used for training. Adversarial training regularizes the classification landscape against the attacks (Fig. 1, b), but high protection often produces a loss in clean accuracy. Other partially effective defenses are based + +on randomness [35, 35] or removal of the adversarial signal [13, 18, 27, 40], by moving the input back to the space of natural (non-attacked) data before classification (Fig. 1, c). + +All the aforementioned strategies activate the defense mechanism only after the attack has been crafted. However, when dealing with adversarial attacks, the first actor to move has a significant advantage. For instance, perturbing an image can avoid online person identification and preserve privacy [6, 19, 26]. Acting first is particularly suitable against Man in the Middle (MitM) attacks that may prac + +![](images/c048a3d55e74570bce821571e1968792f8820461b031e5f3f4a643956839de52.jpg) +(a) certified bounds + +![](images/d1cd31a8efbd7ed677e5f28d56f613a27f651ef891869d2a46ef985db4fb8328.jpg) +(b) adversarial training + +![](images/f457675ae7502def028e1a2fefac5fd56cdc88638b2780215c892bee0d6194bb.jpg) +(c) image purification + +![](images/e94738566a2130b22e9623e5135caff2f0dfb8262984d2437da2d5e8ef2b21b9.jpg) +(d) $A^5$ + +![](images/3cbd9e9047d0c7d1cb305fdf4c976c459c03dd15c5fa9f99f8eedd445a435327.jpg) +Base: airplane (69.8%) +[Verified: ship (100.0%)] +Figure 1. (a) An input $\pmb{x} \in y^*$ is correctly classified. The grey box (certified bounds [12, 42, 43]) shows that, under an attack $\delta \pmb{x}_A$ , $||\delta \pmb{x}_A||_{\infty} < \epsilon_A$ , misclassification $((\pmb{x} + \delta \pmb{x}_A) \in y_0)$ is possible. (b) Adversarial training [4, 20] creates robust DNNs with regular classification landscapes: misclassification is less likely. (c) Image purification [27, 40] moves the attacked input $\pmb{x} + \delta \pmb{x}_A$ by $\delta \pmb{x}_P$ back to correct classification. (d) $A^5$ preemptively moves $\pmb{x}$ into a non-attackable position. The original CIFAR10 image classified as airplane (69.8% confidence) can be misclassified under attack (ship, 100% confidence). Once robustified through $A^5$ (right), misclassification does not occur anymore. + +![](images/0b0c10474ff1425b84189394b598963fce90bb20601108c5301961cbf6dfcfe0.jpg) +Robustified: airplane $(82.6\%)$ Verified: airplane $(82.5\%)$ + +
vA vectorvjVector j-th element
CClassifierRRobustifier
xDataδxAAttacking perturbation
δxDDefensive perturbationx + δxAData under attack
x + δxDRobustified datax + δxD + δxARobustified data under attack
εDDefense magnitude, ||δxD||p < εDεRAttack magnitude, ||δxA||p < εR_A while training R
εAAttack magnitude ||δxA||p < εA while testingεCAttack magnitude, ||δxA||p < εC_A while training C
+ +Table 1. Notation for data $\mathbf{x}$ ,that are the inputs of the classifier $C$ . We adopt an equivalent notation for physical objects $\mathbf{w}$ . + +cally arise in automotive [32], audio processing [13, 22, 38], or while communicating with a remote machine [39]. Our idea spouses this reasoning line: we investigate a novel way to augment the data to preemptively certify that it cannot be attacked (Fig. 1, d). This fits in a recent research area that up to now has received less attention than adversarial training. Researchers explored image and real object augmentation to guarantee a high recognition rate in presence of noise or geometric distortions, but not in adversarial scenarios [24]. Encryption schemes coupled with deep learning [1, 28] or watermarking [16] only partially protect against MitM attacks. The few existing preemptive robustification methods [23] include a procedure [17] that first runs a classifier on clean data and achieves preemptive robustification against MitM attacks through iterative optimization; these are partially effective and do not provide any certification. Our novel framework encompasses most of the aforementioned cases while also introducing for the first time the concept of certified robustification. More in detail, our manifold contributions are: + +(i) We introduce $A^5$ (Adversarial Augmentation Against Adversarial Attacks), a comprehensive framework for preemptive augmentation to make data and physical objects certifiably robust against MitM adversarial attacks. As far as we know, this is the first time certified defense is achieved in this way. Since we provide certified robustness, we guarantee protection against any form of white, grey or black box attack. +(ii) We test different flavours of $A^5$ on standard datasets. By doing so, we study the connection between the robustness of the legacy classifier, the magnitude of the defensive augmentation, and the protection level delivered by $A^5$ . We show $A^5$ achieving state-of-the-art certified protection against MitM attacks by training a robustifier DNN coupled with the legacy classifier, and even better results for co-training of the robustifier and classifier. We perform a critical, visual inspection of the robustified images to answer an interesting theoretical question (how does a non-attackable image look like?) and potentially provide directions for the acquisition of inherently robust images. +(iii) Using Optical Character Recognition (OCR) as an ex + +ample, we show the application of $A^5$ for the design of certifiably robust physical objects, which extends [24] to the case of certified defense against MitM attacks. + +(iv) We share our code at https://github.com/NVlabs/A5, to allow replicating our results or testing $A^5$ in scenarios not considered here, for instance on other datasets or for protection against physical adversarial attack (e.g., adversarial patches). + +# 2. Related Work and Background + +We identify three main philosophies for the development of defenses against adversarial attacks. They are not mutually exclusive and can be adopted together. + +The first approach aims at creating DNNs that are robust to adversarial attacks: e.g., in adversarial training, attacks are generated while training the DNN that should process them correctly [4, 11, 20]. This regularizes the classification landscape and makes attacks less effective (Fig. 1, b). Adversarial training is simple but has one drawback: as attack generation is computationally demanding (even after speed up [25]), attacks are randomly sampled and cannot cover the entire input space, thus impacting the final accuracy. Certified defense [12, 42, 43] can be seen as a state of the art refined form of adversarial training: robustness is obtained by propagating through the DNN an input interval that encompasses all the attacks with magnitude up to $\epsilon_A^C$ . The corresponding set of output bounds allows estimating the worst case scenario and derive the training cost function (e.g., worst case entropy [12, 42]). Research focuses on finding bounds that are both tight and computationally light. IBP [12] computes (and backwards) the bounds at the cost of an additional forward and backward pass. Since the bounds are loose at the beginning of training, a complex schedule is required to guarantee stability and convergence. Linear relaxation [43] delivers tighter bounds leveraging the correlation between different layers to linearly propagate signal and bounds, when possible, and eventually put in linear relation the input interval and output bounds. CROWN [43] bounds are often tighter than IBP, but at an impractical computational cost for large DNNs. Among the attempts to speed up CROWN [33, 37], CROWN-IBP [42] strikes the right balance of computa + +tional complexity and tightness. It is implemented in the auto_LiRPA $^{1}$ library [36, 37] adopted here. One limitation of adversarial training, certified methods, and $A^{5}$ as well, is the trade-off between clean and certified (worst case underattack) accuracy: increasing the latter leads to a decrease in the former [4, 12, 20, 42]. Among the architecture solutions against adversarial attacks, we also mention the recent development of $\ell_{\infty}$ -dist neurons that are inherently robust [41], and could be leveraged by $A^{5}$ as well in future. + +The second defense philosophy is based on active countermeasures taken after the attack is crafted. Image purification [18, 27, 40] projects an attacked image onto the manifold of natural images through a generative model (Fig. 1, c). It is less effective than adversarial training [7], but does not make any assumption on the classifier. Other forms of signal pre-processing have been proposed in the field of audio processing and automatic speech recognition, where MitM physical attacks can be easily implemented [22, 38] and partially defeated by randomized smoothing, DefenseGAN, variational autoencoder (VAE), and Parallel WaveGAN vocoder [13]. Both certified defenses and pre-processing methods do not assume any specific attack form, but the latter are less effective in white box scenarios [2, 31]. Randomization can also be used after the attack, e.g. by changing resolution or padding an image [35], or by picking random experts in a mixture, that is proved to be more robust than deterministic DNNs [21]. + +$A^5$ belongs to a third class of methods, that put in place countermeasures before the attack (Fig. 1, d). The roots of $A^5$ lie in two ideas that have been only partially investigated so far. The first is that, in adversarial scenarios, the first actor to move has a significant advantage. This rule is already leveraged for online privacy, where a preemptive adversarial attack prevents the identification of the framed subject [6, 19]. Since adversarial attacks often generalize to many classifiers, this ploy is particularly effective [26]. Preemptive protection is also a natural form of defense against MitM attacks that can be easily created in contexts like automotive [32], speech and audio processing [13, 22, 38], or anytime the capturing device communicates with a server [39]. Encryption [1, 28] and watermarking [16] coupled with deep learning work against MitM attacks, but do not achieve full protection, are not certified, require the key to be hidden (grey box scenario) and may significantly alter the appearance of the image. Blindpreprocessing [23] applies a tanh transform, normalization, quantization and thermometer encoding to achieve partial and uncertified defense. The closest algorithm to $A^5$ is preemptive robustification [17]: given an image $\pmb{x}$ and a classifier $C$ that can be fine-tuned, it computes $y_0 = C(\pmb{x})$ and randomly samples attacks $\delta \pmb{x}_A$ against $\pmb{x} + \delta \pmb{x}_D$ to find a defensive perturbation $\delta \pmb{x}_D$ , such that $(\pmb{x} + \delta \pmb{x}_D + \delta \pmb{x}_A) \in y_0$ + +(correct classification under attack). Since $A^5$ uses certified bounds, it results in simpler training and higher protection, and does not require running $C$ first, as our robustifier DNN generalizes to all the inputs. The second idea behind $A^5$ comes from physical adversarial examples, i.e., 3d physical objects [3] or 2d patches [5] that consistently fool a DNN under a wide range of viewing conditions. Reverting the attacker and defender roles, the complementary cost function leads to the creation of unadversarial examples: patches or textures that, once applied to a real 3d object, reinforce (instead of degrading) the DNN desired behaviour [24]. However, these are not designed to be inherently robust to adversarial attacks. We exploit the idea in $A^5$ in all those situations where system designers control, to some extent, the inputs fed to the DNN (e.g. in OCR, robotic systems or during the design of the road infrastructure). We show as an example the use of $A^5$ to design certifiably protected characters for OCR, something unexplored both in [17] and [24]. + +# 3. Method + +Threat models We summarize our notation in Table 1, whereas Fig. 2 illustrates the full $A^5$ framework including trainable DNNs in blue, training algorithms in green, physical objects $\boldsymbol{w}$ , and acquired data $\boldsymbol{x}$ . MitM attacks are crafted while transmitting data from the acquisition device to the classifier $C$ , while physical attacks (not tested here) are crafted onto physical objects $\boldsymbol{w}$ , before data acquisition. $A^5$ can be used in a scenario where the attacker has full access to the $C$ : since it uses certified bounds, it is agnostic to the specific nature of the attack. In other words, we certify protection for any $\delta x_A$ , $||\delta x_A|| < \epsilon_A$ and the attacker can use any white-, grey- or black-box attacking algorithm. We assume that $A^5$ can run in a protected environment, not accessible to the attacker, to perform robustification (light blue rectangles in Fig. 2). More details and limitations of our threat model are discussed in the Supplementary, whereas here we first introduce useful definitions and then provide several recipes for the use of $A^5$ in different scenarios. + +Definitions For a classifier $C$ and input $\mathbf{x}$ , $\mathbf{y} = C(\mathbf{x})$ is the logit vector output by $C$ . Under attack, we have: + +$$ +\boldsymbol {y} + \delta \boldsymbol {y} = C (\boldsymbol {x} + \delta \boldsymbol {x} _ {A}), \| \left| \delta \boldsymbol {x} _ {A} \right| | _ {p} < \epsilon_ {A}, \tag {1} +$$ + +where $\epsilon_{A}$ is the magnitude of the attack $\delta \pmb{x}_A$ in $p$ -norm and $\delta \pmb{y}$ is the logit perturbation. Linear relaxation [43] estimates the lower $(\pmb{y}^l)$ and upper $(\pmb{y}^u)$ bounds of $\pmb{y} + \delta \pmb{y}$ , for instance using IBP [12], CROWN [43], CROWN-IBP [42], or a combination of them. The margin for the class $j$ is the difference between the lower bound of the ground truth class $j^{*}$ and the upper bound of $j$ , while it is zero by definition for $j^{*}$ . More formally, for $M$ classes the elements of the margin vector $\pmb{m} = [m_0, m_1, \dots, m_{M - 1}]$ are: $m_{j^{*}} = 0$ + +![](images/735e07d238056c1a05cd65e211b875af78b4d89cc3cfb1e9ab64a214dc1f33dd.jpg) +Figure 2. Schematic representation of the full $A^5$ framework. It allows the robust augmentation of both physical objects $\boldsymbol{w}$ and acquired data $\boldsymbol{x}$ to make them certifiably robust against MitM and physical (not tested here) adversarial attacks. + +and $m_j = \pmb{y}_{j^*}^l -\pmb{y}_j^u$ $\forall j\neq j^{*}$ . The confidence is obtained by applying softmax to $\pmb{m}$ .. + +$$ +p _ {j} (\boldsymbol {x}) = e ^ {- m _ {j}} / \sum_ {k = 0} ^ {M - 1} e ^ {- m _ {k}}, \tag {2} +$$ + +and thus the worst case cross entropy $E(\pmb{x})$ is: + +$$ +E (\boldsymbol {x}) = - \sum_ {j = 0} ^ {M - 1} \hat {y} _ {j} \log \left(p _ {j} (\boldsymbol {x})\right), \tag {3} +$$ + +where $\hat{y}_j$ is the one-hot-encoding of $j^*$ . To formally introduce $A^5$ , we first give the following definitions: + +Definition 1 ( $\epsilon_A$ -robustness). Given an input $\pmb{x}$ and a classifier $C$ , we say that $\pmb{x}$ is $\epsilon_A$ -robust with respect to $C$ in norm $p$ if $\forall \delta \pmb{x}_A, ||\delta \pmb{x}_A||_p < \epsilon_A$ , $C(\pmb{x} + \delta \pmb{x}_A)$ generates the correct classification output. + +Definition 2 (robustification). For a given input $\pmb{x}$ and classifier $C$ , robustification is the process of finding a defensive augmentation $\delta \pmb{x}_D$ , $||\delta \pmb{x}_D||_p < \epsilon_D$ , such that $\pmb{x} + \delta \pmb{x}_D$ is $\epsilon_A$ -robust with respect to $C$ in norm $p$ . + +We compute certified CROWN-IBP bounds through auto_LiRPA [37] and optimize $E(\pmb{x} + \delta \pmb{x}_D)$ to find preemptive, defensive perturbations $\delta \pmb{x}_D$ that make $\pmb{x} + \delta \pmb{x}_D \in \epsilon_A$ -robust, following the training recipes provided here. We also show how to find defensive perturbations $\delta \pmb{w}_D$ for physical objects $\pmb{w}$ using the same approach. + +Offline robustification with ground truth $(A^5 /O)$ : Consider a legacy classifier $C$ that has been certifiably trained (e.g. through CROWN-IBP [42]) against attacks of magnitude $\epsilon_A^C$ . Given an input $\pmb{x}$ and its class $j^{*}$ , we want to find a defensive perturbation $\delta \pmb{x}_D,||\delta \pmb{x}_D||_p < \epsilon_D$ , such that $\pmb{x} + \delta \pmb{x}_D$ is $\epsilon_A^R$ -robust with respect to $C$ in $p$ -norm. Our solution to this problem is $A^5$ /Offline (or $A^5 /O$ ). + +To force $||\delta \pmb{x}_D||_p < \epsilon_D$ , we parameterize $\delta \pmb{x}_D$ via a vector $\pmb{z}$ to which we apply an element-wise sigmoid: + +$$ +\delta \boldsymbol {x} _ {D} = \delta \boldsymbol {x} _ {D} (\boldsymbol {z}) = [ \delta x _ {D} (\boldsymbol {z}) _ {0}, \dots , \delta x _ {D} (\boldsymbol {z}) _ {K - 1} ] \tag {4} +$$ + +$$ +\delta x _ {D} (\boldsymbol {z}) _ {j} = 2 \epsilon_ {D} \left[ 1 / \left(1 + e ^ {- z _ {j}}\right) - 0. 5 \right], \tag {5} +$$ + +where $K$ is the number of elements in $\pmb{x}$ . This is a robust classification problem where $\pmb{z}$ is the only unknown. For any $\pmb{x}$ , we minimize the worst case cross entropy: + +$$ +\boldsymbol {z} = \operatorname {a r g m i n} _ {\tilde {\boldsymbol {z}}} E [ \boldsymbol {x} + \delta \boldsymbol {x} _ {D} (\tilde {\boldsymbol {z}}) ], \tag {6} +$$ + +using RMSProp. Notice that the magnitude of the attack used during robustification $(\epsilon_A^R)$ and that adopted to train $C$ $(\epsilon_A^C)$ can be different. Furthermore, similarly to adversarial training and certified methods, these can also be different from the target attack magnitude, $\epsilon_{A}$ . + +$A^5 / O$ is similar in spirit to preemptive image robustification [17], but does not need $C$ to run on the clean input before robustification. Although $A^5 / O$ could be used in a similar way to robustify $\pmb{x}$ accordingly to the class $C(\pmb{x})$ , running $C$ first and then iteratively solving the optimization problem in Eq. 6 makes it hardly usable in practice. Therefore, it does not directly serve any practical purpose, but it helps answering important theoretical questions (e.g., under which conditions can we robustify $\pmb{x}$ ? What does a robustified image looks like?) and establishing a baseline with known ground truth and no computational constraints. + +Online robustification with robustifier $R$ , legacy $C$ ( $A^5 / R$ ): We introduce a second recipe, $A^5 / R$ , where we train a robustifier DNN $R$ that takes in input $\mathbf{x}$ , no ground truth label, and outputs $\mathbf{z}$ , from which a defensive perturbation $\delta \mathbf{x}_D$ , $||\delta \mathbf{x}_D||_p < \epsilon_D$ , is computed as in Eq. 5. We assume again that a robust legacy $C$ is given and we train $R$ to make each $\mathbf{x} + \delta \mathbf{x}_D$ in the training dataset $\epsilon_A^R$ -robust with respect to $C$ in $p$ -norm. More formally, the weights $\theta_R$ of $R$ are found by minimization of the average worst case entropy on the training dataset $E(\theta_R)$ , defined as: + +$$ +\boldsymbol {z} = R (\boldsymbol {x} \mid \boldsymbol {\theta} _ {R}), E (\boldsymbol {\theta} _ {R}) = (1 / N) \sum_ {\boldsymbol {x}} E [ \boldsymbol {x} + \delta \boldsymbol {x} _ {D} (\boldsymbol {z}) ], \tag {7} +$$ + +where $N$ is the dataset size. Unlike preemptive robustification [17], $R$ does not iteratively solve a complex optimization problem, neither we need running $C$ before $R$ . Therefore, we can use $A^5 / R$ to protect $\pmb{x}$ against $\mathrm{MitM}$ adversarial attacks on-the-fly, ignoring its class, soon after its acquisition, and before its transmission to $C$ (see $A^5 / R$ in Supplementary). Furthermore, since $R$ is trained for a specific + +$C$ , it can be deployed on the field (e.g., into the firmware of the acquisition devices) without changing (and thus maintaining the legacy of) the classifier infrastructure. + +Online robustification with robustifier $R$ , retrained $C$ ( $A^5 / RC$ ): The $A^5 / RC$ recipe is equivalent to $A^5 / R$ , but it leverages the co-adaptation of $C$ and $R$ by joint training (see $A^5 / RC$ in Supplementary). The cost function is again the average worst case entropy $E(\theta_C, \theta_R)$ , defined now as: + +$$ +\boldsymbol {z} = R (\boldsymbol {x} | \boldsymbol {\theta} _ {R}), \quad \boldsymbol {y} = C (\boldsymbol {x} + \delta \boldsymbol {x} _ {D} | \boldsymbol {\theta} _ {C}), \tag {8} +$$ + +$$ +E \left(\boldsymbol {\theta} _ {C}, \boldsymbol {\theta} _ {R}\right) = (1 / N) \sum_ {\boldsymbol {x}} E [ \boldsymbol {x} + \delta \boldsymbol {x} _ {D} (\boldsymbol {z}) ], \tag {9} +$$ + +where $\theta_{C}$ are the parameters of $C$ . Training $R$ and $C$ from scratch is often unstable and requires careful tuning to achieve convergence (likewise robust classifiers [12, 42]). Therefore, we first train $R$ with $A^5 / C$ and then fine tune $R$ and $C$ together. During the tuning phase, we can use random input augmentation to prevent overfitting (see Section 4 and Supplementary for details). Differently from $A^5 / R$ , $A^5 / RC$ does not preserve the legacy classifier $C$ , but it achieves the same scope (with more effectiveness) in terms of protection against MitM attacks. + +Offline physical robustification with $(A^5 /PC)$ and without $(A^5 /P)$ retraining $C$ , with ground truth: This recipe aims at producing physical objects that are certifiably robust against MitM attacks (see $A^5 /P$ , $A^5 /PC$ in Supplementary). Similarly to unadversarial objects [24], we assume that the system creator can design the entire infrastructure and the attacker cannot interfere in this phase. The problem is formally defined for $A^5 /PC$ as: + +$$ +\boldsymbol {x} + \delta \boldsymbol {x} _ {D} = A (\boldsymbol {w} + \delta \boldsymbol {w} _ {D}), \| \delta \boldsymbol {w} _ {D}) \| < \epsilon_ {D}, \tag {10} +$$ + +$$ +E \left(\boldsymbol {\theta} _ {C}, \delta \boldsymbol {w} _ {D}\right) = (1 / N) \sum_ {\boldsymbol {x}} E (\boldsymbol {x} + \delta \boldsymbol {x} _ {D}) \tag {11} +$$ + +where $\pmb{w}$ is the object appearance, whereas $\pmb{x} = A(\pmb{w})$ is the data acquisition model (e.g., image formation). $N$ is here the number of random samples generated as $\pmb{x} + \delta \pmb{x}_D = A(\pmb{w} + \delta \pmb{w}_D)$ ; it is important noticing in fact that the acquisition process is not deterministic: as an exemplary case we mention OCR, where $\pmb{w}$ is the shape of a character, while $A$ may include random perspective transformations, contrast and brightness changes, noise addition and blurring, that may occur on a real camera. The defensive perturbation $\delta \pmb{w}_D$ represents a change in the physical character appearance $\pmb{w}$ that decreases the success rate of an attack towards the acquired image $\pmb{x}$ (therefore after framing the character with $A$ ) and the classifier $C$ . + +In $A^5 / P$ we change the character shapes and use the legacy classifier $C$ ; in $A^5 / PC$ we also train $C$ while estimating the robust characters $\boldsymbol{w} + \delta \boldsymbol{w}$ . In both cases $C$ + +
εAεAC=0.0εAC=0.1εAC=0.2εAC=0.3εAC=0.4
0.1——1.17 [—, 2.24] [42]——
0.10.81 [89.02, 100.00]0.92 [2.00, 5.41]1.32 [2.14, 2.60]1.44 [2.33, 2.63]2.42 [3.36, 3.41]
0.3———1.82 [—, 7.02] [42]—
0.30.69 [100.00, 100.00]0.94 [98.41, 100.00]1.30 [99.62, 100.00]1.41 [7.06, 9.98]2.58 [7.03, 7.68]
+ +Table 2. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST classifiers trained by us with CROWN-IBP [42], under attack $\epsilon_{A} = \{0.1,0.3\}$ ; we also show the metrics reported in [42]. A $100\%$ certified error rate identifies DNNs that are not protected against attacks. + +
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.79 [87.09, 100.00]0.82 [1.51, 2.89]1.22 [1.64, 1.86]1.34 [1.76, 1.87]2.38 [2.91, 2.92]
0.10.100.79 [86.51, 100.00]0.72 [1.21, 1.92]1.15 [1.43, 1.49]1.34 [1.50, 1.58]2.41 [2.71, 2.74]
0.10.200.89 [86.40, 100.00]0.56 [0.78, 1.04]1.13 [1.18, 1.21]1.35 [1.37, 1.38]2.40 [2.55, 2.55]
0.10.300.93 [86.44, 100.00]0.51 [0.57, 0.70]1.14 [1.16, 1.17]1.34 [1.37, 1.37]2.38 [2.52, 2.52]
0.10.400.99 [85.93, 100.00]0.52 [0.53, 0.60]1.14 [1.16, 1.16]1.35 [1.38, 1.38]2.44 [2.51, 2.51]
0.30.050.71 [100.00, 100.00]0.93 [97.18, 100.00]1.34 [99.17, 100.00]1.35 [5.16, 6.97]2.55 [6.00, 6.42]
0.30.100.72 [100.00, 100.00]0.88 [95.32, 100.00]1.36 [98.58, 100.00]1.33 [4.12, 5.19]2.49 [5.54, 5.85]
0.30.200.72 [100.00, 100.00]0.88 [90.62, 100.00]1.41 [96.21, 100.00]1.32 [3.21, 3.77]2.35 [5.02, 5.22]
0.30.300.73 [100.00, 100.00]0.89 [86.28, 100.00]1.41 [94.18, 100.00]1.31 [2.91, 3.40]2.32 [4.66, 4.81]
0.30.400.81 [100.00, 100.00]0.85 [83.49, 100.00]1.42 [92.63, 100.00]1.32 [2.74, 3.20]2.28 [4.16, 4.29]
+ +Table 3. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, $A^5 / O$ , under attack $\epsilon_A = \{0.1, 0.3\}$ . During training we use $\epsilon_A^R = \epsilon_A$ . + +
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.82 [85.50, 100.00]0.95 [1.71, 3.58]1.30 [1.82, 2.02]1.40 [1.86, 2.03]2.43 [2.96, 2.97]
0.10.100.81 [82.87, 100.00]1.00 [1.48, 2.54]1.19 [1.62, 1.73]1.40 [1.68, 1.74]2.38 [2.77, 2.79]
0.10.200.88 [78.48, 100.00]0.93 [1.33, 1.77]1.20 [1.42, 1.52]1.31 [1.53, 1.59]2.26 [2.55, 2.57]
0.10.300.99 [75.89, 100.00]0.91 [1.16, 1.54]1.14 [1.31, 1.36]1.26 [1.43, 1.47]2.20 [2.42, 2.43]
0.10.401.04 [75.69, 100.00]0.87 [1.05, 1.35]1.14 [1.27, 1.33]1.24 [1.43, 1.46]2.22 [2.41, 2.43]
0.30.050.69 [100.00, 100.00]0.94 [96.93, 100.00]1.33 [99.11, 100.00]1.41 [5.13, 6.91]2.58 [5.77, 6.09]
0.30.100.71 [100.00, 100.00]0.91 [95.09, 100.00]1.33 [98.47, 100.00]1.38 [3.86, 4.81]2.55 [4.71, 4.83]
0.30.200.72 [100.00, 100.00]0.92 [90.13, 100.00]1.37 [96.08, 100.00]1.41 [2.26, 2.63]2.55 [3.33, 3.40]
0.30.300.75 [100.00, 100.00]0.93 [85.54, 100.00]1.35 [93.36, 100.00]1.32 [1.86, 1.99]2.37 [2.86, 2.88]
0.30.400.83 [100.00, 100.00]0.96 [82.48, 100.00]1.36 [92.15, 100.00]1.29 [1.69, 1.83]2.35 [2.67, 2.69]
+ +Table 4. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, $A^5 / R$ , under attack $\epsilon_A = \{0.1, 0.3\}$ . During training we use $\epsilon_A^R = \epsilon_A$ . + +is previously trained to be robust, e.g., through CROWN-IBP [42]. The $A^5 / P$ and $A^5 / PC$ recipes find application in all those situations where the developer can design the entire infrastructure; beyond OCR, other examples are road sign or robotic tool design and classification. + +# 4. Results and discussion + +We test $A^5 / O$ , $A^5 / R$ , and $A^5 / RC$ on MNIST [10], CIFAR10 [14], FashionMNIST [34], and Tinyimageenet [15], that have been widely used to establish significant milestones in the adversarial defense field. Training schedules and DNN architectures are given in the Supplementary. In all cases we use a $p = \infty$ norm for the attack and defense. Our intention is to show the improvement achieved by $A^5$ over traditional, certified methods like CROWN-IBP. The reader should anyway keep in mind that the direct comparison of $A^5$ and CROWN-IBP is partially unfair, as CROWN-IBP does not apply any preemptive defense. + +We first use CROWN-IBP to train five certifiably robust MNIST classifiers with varying levels of robustness, for training attacks $\epsilon_A^C = 0.0$ (unprotected $C$ ) and $\epsilon_A^C = \{0.1, 0.2, 0.3, 0.4\}$ . Their clean and certified errors are reported in Table 2, for attacks $\epsilon_A = \{0.1, 0.3\}$ . We also + +
εA=εCεDεCA=0.0εCA=0.1εCA=0.2εCA=0.3εCA=0.4
0.10.0588.66[88.66, 88.66]1.00[1.38, 2.35]1.16[1.61, 2.46]1.14[1.54, 2.29]1.37[1.72, 2.33]
0.10.1088.66[88.66, 88.66]0.77[0.95, 1.40]0.99[1.16, 1.56]0.86[1.03, 1.42]1.13[1.31, 1.72]
0.10.2088.66[88.66, 88.66]0.69[0.80, 1.05]0.84[0.94, 1.25]0.96[1.07, 1.33]0.88[1.10, 1.54]
0.10.3088.66[88.66, 88.66]0.64[0.70, 1.02]0.91[0.99, 1.16]0.79[0.88, 1.21]1.00[1.12, 1.43]
0.10.4088.66[88.66, 88.66]0.61[0.68, 0.88]0.93[0.99, 1.18]0.69[0.76, 1.10]1.00[1.03, 1.30]
0.30.0588.66[88.66, 88.66]88.66[88.66, 88.66]2.57[8.35, 10.86]3.42[9.51, 11.72]8.00[18.05, 19.77]
0.30.1088.66[88.66, 88.66]88.66[88.66, 88.66]2.07[5.66, 7.56]2.35[6.18, 8.07]2.35[6.31, 7.93]
0.30.2088.66[88.66, 88.66]1.68[3.11, 4.79]1.40[2.42, 3.62]1.28[2.82, 4.01]1.88[3.53, 4.90]
0.30.3088.66[88.66, 88.66]0.84[0.96, 1.38]1.00[1.14, 1.36]0.94[1.15, 1.59]1.07[1.39, 1.90]
0.30.4088.66[88.66, 88.66]0.72[0.80, 1.03]0.75[0.82, 1.06]0.91[1.14, 1.68]0.87[1.14, 1.50]
+ +report the error rates estimated with autoattack [8, 9], an automatic toll based on an ensemble of four attacks to evaluate the practical DNN robustness. These error rates are in the same ballpark of those reported in the CROWN-IBP paper [42]: they represent the baseline for the analysis of the different $A^5$ recipes considered here. + +For each $C$ we run $A^5 /O$ on the MNIST test set, for defense magnitudes $\epsilon_{D} = \{0.05,0.1,0.2,0.3,0.4\}$ . During the optimization of the entropy in Eq. 6, we use attacks $\epsilon_A^R = \{0.1,0.3\}$ , not necessarily equal to $\epsilon_A^C$ . Table 3 reports the clean, autoattack, and certified errors for a testing attack $\epsilon_{A} = \epsilon_{A}^{R}$ . When $C$ is not robust ( $\epsilon_{A}^{C} = 0.0$ ), $A^5 /O$ does not reduce the certified error², otherwise it generally overcomes the base CROWN-IBP classifier, as shown by comparing against the error rates in Table 2. As a rule of thumb, the best clean / certified error trade off is often achieved for $\epsilon_{A}^{R} = \epsilon_{A}^{C} = \epsilon_{A}$ . Results also improve, as expected, for larger defenses $\epsilon_{D}$ , when $A^5 /O$ beat CROWN-IBP by a significant margin on clean and certified errors. However, $A^5 /O$ does not work well when the legacy $C$ is trained with a large $\epsilon_{A}^{C}$ , possibly because of its initial low clean accuracy. The results are consistent with our interpretation of the effects of adversarial training and robustification on the classification landscape (Fig. 1): training $C$ with a large $\epsilon_{A}^{C}$ creates classification valleys wide enough to accommodate $\pmb{x} + \delta \pmb{x}_D + \delta \pmb{x}_A$ , rendering the adversarial attacks ineffective. $A^5$ moves any $\pmb{x}$ towards the centers of the valleys; a large $\epsilon_{D}$ allows moving them more easily. Too large $\epsilon_{A}^{C}$ remains however detrimental for the clean accuracy. + +Table 4 reports results for $A^5 / R$ , demonstrating that training a DNN robustifier $R$ for on-the-fly, certified robustification, ignoring the ground truth, is feasible. Likewise $A^5 / O$ , $A^5 / R$ beats the base CROWN-IBP classifier by a significant margin. Sometimes $A^5 / O$ does better by leveraging the knowledge of the ground truth class, whereas the regular robustification signal generated by $A^5 / R$ is more effective in other cases. + +Our last MNIST experiment investigates $A^5 /RC$ that does not preserve the legacy $C$ like $A^5 /R$ , but achieves bet + +Table 5. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST, $A^5 / RC$ , under attack $\epsilon_A = \{0.1, 0.3\}$ . During training we use $\epsilon_A^R = \epsilon_A$ , and data augmentation. + +
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2554/25552.75 [63.47, 64.09]53.34 [62.42, 62.79]53.75 [64.92, 65.49]
8/2558/25551.06 [60.04, 60.73]52.36 [59.24, 59.58]53.16 [63.50, 64.03]
8/25516/25548.72 [54.89, 55.25]50.60 [55.22, 55.38]52.23 [60.79, 61.14]
8/25532/25545.68 [49.51, 49.74]48.06 [50.86, 50.94]50.99 [57.37, 57.61]
+ +Table 6. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, $A^5 / O$ , under attack $\epsilon_A = 8/255$ . During training we use different $\epsilon_A^R$ and $\epsilon_D$ . The robust classifier trained with CROWN-IBP has clean [and certified] errors equal to 54.02 [66.94] or 45.47 [69.55], depending on the CROWN-IBP training configuration [42]. + +
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25554.60 [66.34, 67.11]54.60 [66.34, 67.11]54.60 [66.31, 67.11]
8/2554/25553.73 [64.07, 64.57]54.12 [63.32, 63.74]54.37 [63.58, 64.03]
8/2558/25552.91 [62.59, 63.12]53.24 [61.38, 61.93]54.54 [60.85, 61.16]
8/25516/25551.58 [60.40, 61.02]51.95 [59.21, 59.67]54.44 [59.87, 60.11]
8/25532/25550.91 [57.26, 57.75]51.70 [56.91, 57.43]54.05 [58.39, 58.73]
+ +Table 7. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, $A^5 / R$ , under attack $\epsilon_A = 8 / 255$ . During training we use different $\epsilon_A^R$ and $\epsilon_D$ . The robustifier $R$ in the first row uses $\epsilon_D = 0$ and thus it does not provide any form of defense — the error reported here is equivalent to that of a robust classifier trained with CROWN-IBP [42]. + +
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25550.24 [69.17, 73.65]58.63 [69.25, 70.63]74.28 [77.82, 77.91]
8/2554/25546.76 [63.03, 70.99]54.05 [62.80, 64.44]67.59 [71.20, 71.32]
8/2558/25541.24 [54.01, 64.31]50.74 [56.63, 59.49]65.96 [67.63, 67.78]
8/25516/25536.96 [42.45, 52.90]39.15 [41.33, 44.22]58.17 [58.67, 58.85]
8/25532/25535.26 [37.13, 42.76]41.59 [43.01, 45.55]42.25 [43.02, 44.08]
+ +Table 8. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, $A^5 / RC$ , under attack $\epsilon_A = 8/255$ . During training we use different $\epsilon_A^R$ and $\epsilon_D$ . The robustifier $R$ in the first row uses $\epsilon_D = 0$ and thus it does not provide any form of defense, whereas the corresponding robust CROWN-IBP classifier is fine tuned using $\epsilon_A^C = \epsilon_R$ during training. + +ter clean accuracy and protection (Table 5). In the Supplementary we discuss the role of data augmentation to achieve this result. Interestingly, the best results are obtained when $C$ is initially trained for $\epsilon_A^C \leq \epsilon_A$ , probably because $C$ has an initial high clean accuracy and subsequently co-adapts with $R$ to also guarantee a small certified error. + +Fig. 3 shows examples of robustification on MNIST. At visual inspection, $A^5 / O$ , $A^5 / R$ , and $A^5 / RC$ act similarly on these images: they add contrast and enhance the high frequency features (i.e., edges) of the digits. The few failure cases are often pictures that look ambiguous even to a human observer (right panels in Fig. 3). + +We perform tests on CIFAR10 and compare $A^5$ against a state of the art CROWN-IBP classifiers $C$ , trained with $\epsilon_A^C = 8.8 / 255$ , that we also adopt as legacy $C$ for $A^5$ . To test the effect of the magnitude of the attack used for robustification (with $A^5 / O$ ) or for training $R$ (for $A^5 / R, A^5 / RC$ ), we use $\epsilon_A^R = \{4 / 255, 8 / 255, 16 / 255\}$ that are + +![](images/468d7a738172efaae82413600154de03b76c286f7eebc1e91f303ccfbbeae23e.jpg) +Figure 3. Successful and failed robustifications for $A^5 / O$ , $A^5 / R$ , and $A^5 / RC$ , for MNIST ( $\epsilon_A = \epsilon_A^C = \epsilon_A^R = \epsilon_D = 0.3$ ) and CIFAR10 ( $\epsilon_A = 8.8 / 255$ , $\epsilon_A^C = 8 / 255$ , $\epsilon_A^R = 4 / 255$ , and $\epsilon_D = 32 / 255$ ). Each triplet shows the original (left) and robustified (center) images; the rightmost panel is the defensive augmentation. All the $A^5$ recipes consistently increase the contrast and the high frequency content on MNIST, where failure cases are ambiguous even for a human observer. For CIFAR10, $A^5 / O$ leverage isolated pixel changes, whereas $A^5 / R$ and $A^5 / RC$ consistently increase the color contrast and saturation, similarly to results in literature [17]. + +![](images/5eecde7c36eb1cb17bf359374e103bf358884f2393484dc265a9e25f36e525f8.jpg) +Figure 4. The first, second and third row show respectively the robustified $(\boldsymbol{w} + \delta \boldsymbol{w}_D)$ , nominal $(\boldsymbol{w})$ and robustification $(\delta \boldsymbol{w}_D)$ sets of 62 letters considered for testing $A^5 / PC$ . The last row shows examples of the robustified letters after framing them, $\boldsymbol{x} + \delta \boldsymbol{x}_D = A(\boldsymbol{w} + \delta \boldsymbol{w}_D)$ . + +smaller, equal or greater than $\epsilon_A^C$ . Results in Table 6 are consistent with those measured for MNIST: $A^5 / O$ signifi + +cantly improves the clean and certified error rates over the base CROWN-IBP classifier. Training $R$ through $A^5 / R$ + +
AlgoError [certified error]
Vanilla0.89% [100.00%]
CROWN-IBP3.85% [13.85%]
A5/P3.08% [11.84%]
A5/PC0.73% [4.20%]
+ +Table 9. Error and certified error for the classification of the 62 characters in Fig. 4, for $\epsilon_A = 0.1$ and different algorithms. $A^5 /P$ changes the physical shape of the letters to achieve robustness after framing them; $A^5 /PC$ also trains the associated classifier. + +produces a smaller improvement, still sufficient to beat the CROWN-IBP classifier by a significant margin (Table 7). The value of $\epsilon_A^R$ has a significant impact on the final result: if too large ( $\epsilon_A^R > \epsilon_A$ ), the clean accuracy is penalized, because $A^5$ looks for very large plateaus in the classification landscape, that are hard or impossible to find. In other words, we observe for $A^5$ a trade-off between the clean and certified accuracy, likewise adversarial training and certified methods. Best results on CIFAR10 are achieved again by $A^5 / RC$ , when $\epsilon_A^R$ is equal or slightly smaller than $\epsilon_A = \epsilon_A^C$ . The improvements in clean and certified error are in this case in the impressive range of $19\%$ and $24\%$ respectively (Table 8). We also notice here that the theoretical certified error rates measured with auto_LiRPA [37] in $A^5$ and those measured experimentally by autoattack [8, 9] are close for $A^5$ , whose working point at convergence may favor the consistency between these two estimates. + +Also for CIFAR10 we perform a visual inspection of defensive augmentation (Fig. 3). $A^5 / O$ exploits a slight contrast increase and modifies isolated pixels: likewise adversarial attacks with small $p$ norm (up to single pixel attacks [29]), the same seems possible when crafting defensive perturbations. Preemptive robustification [17] shows a similar pattern (see their Fig. 6): non natural textures and enhanced color saturation emerge in their images for large defensive perturbations. The robust images crafted by $R$ in $A^5 / R$ or $A^5 / RC$ , however, are different: they generally show increased contrast and color saturation, without the textures generated by $A^5 / O$ and [17]. We speculate that these textures may be associated with the iterative optimization in $A^5 / O$ and [17], that may both suffer some form of obfuscated gradient [2]; a more regular defense is instead generated by $R$ without resorting to any iterative process. + +Additional results on FashionMNIST and Tinyimagenet, overall consistent with the ones presented here and showing the quantitative advantages provided by $A^5$ on these datasets, are reported in the Supplementary. These experiments highlight two important facts. The first one is that the integration of the design philosophy of $A^5$ with that of other recent ideas in the field of adversarial defense (like the $\ell_{\infty}$ -dist neurons that implement 1-Lipschitz functions and are inherently robust to adversarial attacks) may lead to + +even bigger improvements in terms of certified robustness. The second is that the integration of different defense strategies may indeed be strictly needed to guarantee further progresses, as scaling to very large dataset remains problematic even for a preemptive robustification algorithms like $A^5$ . + +Finally, for $A^5 / P$ and $A^5 / PC$ , we test the classification of an alphabet of 62 characters (Fig. 4), after random rotation, shift, perspective distortion, noise addition, blurring and color jittering (details in Supplementary), that simulate the scanning of a document for OCR. Table 9 shows the clean and certified errors for: a vanilla, non robust classifier; a CROWN-IBP classifier; $A^5 / P$ ; and $A^5 / PC$ . $A^5$ achieves again a significant improvement both on clean and certified errors: $A^5 / PC$ has better (0.73%) accuracy of the vanilla $C$ (0.89%) and can be attacked in only 4.20% of the cases. This results is obtained with a slight modification of the shape of the prototype characters shown in Fig. 4. + +# 5. Discussion and conclusion + +We introduce $A^5$ , the first framework that leverages adversarial augmentation to preemptively modify the input of a DNN or a physical object, to make it certifiable robust against adversarial attacks. $A^5$ is not simply complementary to other defense methods: its tights with them are strict and they can be used together to achieve higher robustness. For instance, one of the limitations we found is that $A^5 / R$ requires a somehow robust initial classifier $C$ . Co-training $R$ and a poorly robust $C$ easily compensates for the initially low robustness of $C$ and boosts its high clean accuracy while also achieving state of the art results in terms of certified accuracy; finding the optimal initial robustness of $C$ remains an open question though. + +Here we do not investigate all the possible $A^5$ recipes: in its most general implementation $(A^5 / PRC)$ the physical objects, $R$ and $C$ could all be optimized at the same time. This opportunity can be explored through the released code at https://github.com/NVlabs/A5. This will also serve answering other questions that do not find space here: for instance, we do not know if robustification generalizes to other classifiers as adversarial attacks do, how $A^5$ works in norms other than $p = \infty$ , and how much it can further improve if coupled with state-of-the-art solutions like the recently proposed $\ell_{\infty}$ -dist neurons [41]. + +Different $A^5$ recipes find applications in scenarios where the user can control the acquisition device, equip it with $R$ and guarantee its protection (e.g., image acquisition on a phone before server communication) or while designing the infrastructure (e.g., robust road signs creation); it may be not suitable for more general adversarial scenarios. Overall, we believe that the practical deployment of robust systems may benefit from (or even require) methods for both robust classifiers and preemptive robustification. + +# References + +[1] Maungmaung Aprilpyone and Hitoshi Kiya. Encryption inspired adversarial defense for visual classification. 2020 IEEE International Conference on Image Processing (ICIP), pages 1681-1685, 2020. 2, 3 +[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 274-283. PMLR, 10-15 Jul 2018. 3, 8 +[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples, 2018. 3 +[4] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 4312-4321. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Survey Track. 1, 2, 3 +[5] Tom Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. 2017. 3 +[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1, 3 +[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020. 3 +[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 5, 6, 8 +[9] Francesco Croce and Matthias Hein. Mind the box: $l_{1}$ -apgd for sparse adversarial attacks on image classifiers. In ICML, 2021. 5, 6, 8 +[10] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 5 +[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1, 2 +[12] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Mann, and Pushmeet Kohli. On the effectiveness of interval bound propagation for training verifiably robust models, 2018. 1, 2, 3, 5 +[13] Sonal Joshi, Jesús Villalba, Piotr Želasko, Laureano MoroVelázquez, and Najim Dehak. Study of pre-processing defenses against adversarial attacks on state-of-the-art speaker recognition systems. IEEE Transactions on Information Forensics and Security, 16:4811-4826, 2021. 1, 2, 3 + +[14] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5 +[15] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 5 +[16] Xiaoting Li, Lingwei Chen, Jinquan Zhang, James Larus, and Dinghao Wu. Watermarking-based defense against adversarial attacks on deep neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2021. 2, 3 +[17] Seungyong Moon, Gaon An, and Hyun Oh Song. Preemptive image robustification for protecting users against man-in-the-middle adversarial attacks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 4, 7, 8 +[18] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning (ICML), 2022. 1, 3 +[19] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. pages 1491-1500, 10 2017. 1, 3 +[20] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 1, 2, 3 +[21] Rafael Pinot, Raphael Ettedgui, Geovani Rizk, Yann Chevaleyre, and Jamal Atif. Randomization matters how to defend against strong adversarial attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7717-7727. PMLR, 13-18 Jul 2020. 3 +[22] Yao Qin, Nicholas Carlini, Ian J. Goodfellow, G. Cottrell, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. ArXiv, abs/1903.10346, 2019. 2, 3 +[23] Adnan Siraj Rakin, Zhezhi He, Boqing Gong, and Deliang Fan. Blind pre-processing: A robust defense method against adversarial examples. ArXiv, abs/1802.01549, 2018. 2, 3 +[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 3, 5, 6 +[25] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2 +[26] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Fawkes: Protecting Privacy against Unauthorized Deep Learning Models. USENIX Association, USA, 2020. 1, 3 +[27] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1, 3 + +[28] Jinmyeong Shin, Seok-Hwan Choi, and Yoon-Ho Choi. Is homomorphic encryption-based deep learning secure enough? Sensors, 21(23), 2021. 2, 3 +[29] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 8 +[30] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. Jan. 2014. 2nd International Conference on Learning Representations, ICLR 2014; Conference date: 14-04-2014 Through 16-04-2014. 1 +[31] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1633-1645. Curran Associates, Inc., 2020. 3 +[32] Derui Wang, Chaoran Li, Sheng Wen, Surya Nepal, and Yang Xiang. Man-in-the-middle attacks against machine learning classifiers via malicious generative models. IEEE Transactions on Dependable and Secure Computing, 18(5):2074-2087, sep 2021. 2, 3 +[33] Shiqi Wang, Huan Zhang, Kaidi Xu, Xue Lin, Suman Jana, Cho-Jui Hsieh, and J Zico Kolter. Beta-CROWN: Efficient bound propagation with per-neuron split constraints for complete and incomplete neural network verification. Advances in Neural Information Processing Systems, 34, 2021. 2 +[34] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5 +[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In International Conference on Learning Representations, 2018. 1, 3 +[36] Kaidi Xu, Zhouxing Shi, Huan Zhang, Yihan Wang, Kai-Wei Chang, Minlie Huang, Bhavya Kailkhura, Xue Lin, and Cho-Jui Hsieh. Automatic perturbation analysis for scalable certified robustness and beyond. arXiv: Learning, 2020. 3 +[37] Kaidi Xu, Huan Zhang, Shiqi Wang, Yihan Wang, Suman Jana, Xue Lin, and Cho-Jui Hsieh. Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers. In International Conference on Learning Representations, 2021. 2, 3, 4, 8 +[38] Hiromu Yakura and Jun Sakuma. Robust audio adversarial example for a physical attack. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5334–5341. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 3 +[39] Jianfei Yang, Han Zou, and Lihua Xie. Robustsense: Defending adversarial attack for secure device-free human activity recognition. ArXiv, abs/2204.01560, 2022. 2, 3 +[40] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of + +Proceedings of Machine Learning Research, pages 12062-12072. PMLR, 18-24 Jul 2021. 1, 3 +[41] Bohang Zhang, Tianle Cai, Zhou Lu, Di He, and Liwei Wang. Towards certifying 1-infinity robustness using neural networks with 1-inf-dist neurons. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12368-12379. PMLR, 18-24 Jul 2021. 3, 8 +[42] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane Boning, and Cho-Jui Hsieh. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 5, 6 +[43] Huan Zhang, Tsui-Wei Weng, Pin-Yu Chen, Cho-Jui Hsieh, and Luca Daniel. Efficient neural network robustness certification with general activation functions. In Advances in Neural Information Processing Systems (NIPS), arXiv preprint arXiv:1811.00866, dec 2018. 1, 2, 3 \ No newline at end of file diff --git a/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/images.zip b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e72dfecabe54adf133d6331261d93efdaa1041a7 --- /dev/null +++ b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42af3d17155406e7560395316cc827582ff6f87a822e603f71c9a027fb253989 +size 584388 diff --git a/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/layout.json b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c0ca2a651b99fa1501eb5c39652b248c4fad4ef --- /dev/null +++ b/2023/The Best Defense Is a Good Offense_ Adversarial Augmentation Against Adversarial Attacks/layout.json @@ -0,0 +1,13480 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 516, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 516, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 516, + 138 + ], + "type": "text", + "content": "The Best Defense is a Good Offense: Adversarial Augmentation against Adversarial Attacks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 162, + 277, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 162, + 277, + 175 + ], + "spans": [ + { + "bbox": [ + 175, + 162, + 277, + 175 + ], + "type": "text", + "content": "Iuri Frosio, NVIDIA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 178, + 277, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 178, + 277, + 189 + ], + "spans": [ + { + "bbox": [ + 177, + 178, + 277, + 189 + ], + "type": "text", + "content": "ifrosio@nvidia.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 319, + 162, + 417, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 162, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 319, + 162, + 417, + 175 + ], + "type": "text", + "content": "Jan Kautz, NVIDIA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 321, + 178, + 414, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 178, + 414, + 189 + ], + "spans": [ + { + "bbox": [ + 321, + 178, + 414, + 189 + ], + "type": "text", + "content": "jkautz@nvidia.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 217, + 191, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 191, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 191, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "text", + "content": "Many defenses against adversarial attacks (e.g. robust classifiers, randomization, or image purification) use countermeasures put to work only after the attack has been crafted. We adopt a different perspective to introduce " + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "text", + "content": " (Adversarial Augmentation Against Adversarial Attacks), a novel framework including the first certified preemptive defense against adversarial attacks. The main idea is to craft a defensive perturbation to guarantee that any attack (up to a given magnitude) towards the input in hand will fail. To this aim, we leverage existing automatic perturbation analysis tools for neural networks. We study the conditions to apply " + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "text", + "content": " effectively, analyze the importance of the robustness of the to-be-defended classifier, and inspect the appearance of the robustified images. We show effective on-the-fly defensive augmentation with a robustifier network that ignores the ground truth label, and demonstrate the benefits of robustifier and classifier co-training. In our tests, " + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "text", + "content": " consistently beats state of the art certified defenses on MNIST, CIFAR10, FashionMNIST and Tinyimagenet. We also show how to apply " + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 242, + 290, + 529 + ], + "type": "text", + "content": " to create certifiably robust physical objects. Our code at https://github.com/NVlabs/A5 allows experimenting on a wide range of scenarios beyond the man-in-the-middle attack tested here, including the case of physical attacks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Since Deep Neural Networks (DNNs) have been found vulnerable to adversarial attacks [11, 30], researchers studied various protection strategies [4, 12, 20, 42, 43]. For instance, adversarial training [11, 30] generates attacks while asking a DNN for the correct output in training; it is simple, partially effective and widely adopted. Certified methods (e.g., IBP [12], CROWN [43], CROWN-IBP [42]) do a step more by estimating correct (although often pessimistic) output bounds (Fig. 1, a) used for training. Adversarial training regularizes the classification landscape against the attacks (Fig. 1, b), but high protection often produces a loss in clean accuracy. Other partially effective defenses are based" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 218, + 546, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 218, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 546, + 255 + ], + "type": "text", + "content": "on randomness [35, 35] or removal of the adversarial signal [13, 18, 27, 40], by moving the input back to the space of natural (non-attacked) data before classification (Fig. 1, c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 256, + 546, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 546, + 340 + ], + "type": "text", + "content": "All the aforementioned strategies activate the defense mechanism only after the attack has been crafted. However, when dealing with adversarial attacks, the first actor to move has a significant advantage. For instance, perturbing an image can avoid online person identification and preserve privacy [6, 19, 26]. Acting first is particularly suitable against Man in the Middle (MitM) attacks that may prac" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 335, + 354, + 410, + 396 + ], + "blocks": [ + { + "bbox": [ + 335, + 354, + 410, + 396 + ], + "lines": [ + { + "bbox": [ + 335, + 354, + 410, + 396 + ], + "spans": [ + { + "bbox": [ + 335, + 354, + 410, + 396 + ], + "type": "image", + "image_path": "c048a3d55e74570bce821571e1968792f8820461b031e5f3f4a643956839de52.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 399, + 413, + 409 + ], + "lines": [ + { + "bbox": [ + 334, + 399, + 413, + 409 + ], + "spans": [ + { + "bbox": [ + 334, + 399, + 413, + 409 + ], + "type": "text", + "content": "(a) certified bounds" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 438, + 355, + 510, + 396 + ], + "blocks": [ + { + "bbox": [ + 438, + 355, + 510, + 396 + ], + "lines": [ + { + "bbox": [ + 438, + 355, + 510, + 396 + ], + "spans": [ + { + "bbox": [ + 438, + 355, + 510, + 396 + ], + "type": "image", + "image_path": "d1cd31a8efbd7ed677e5f28d56f613a27f651ef891869d2a46ef985db4fb8328.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 398, + 520, + 410 + ], + "lines": [ + { + "bbox": [ + 428, + 398, + 520, + 410 + ], + "spans": [ + { + "bbox": [ + 428, + 398, + 520, + 410 + ], + "type": "text", + "content": "(b) adversarial training" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 337, + 413, + 410, + 454 + ], + "blocks": [ + { + "bbox": [ + 337, + 413, + 410, + 454 + ], + "lines": [ + { + "bbox": [ + 337, + 413, + 410, + 454 + ], + "spans": [ + { + "bbox": [ + 337, + 413, + 410, + 454 + ], + "type": "image", + "image_path": "f457675ae7502def028e1a2fefac5fd56cdc88638b2780215c892bee0d6194bb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 457, + 417, + 468 + ], + "lines": [ + { + "bbox": [ + 329, + 457, + 417, + 468 + ], + "spans": [ + { + "bbox": [ + 329, + 457, + 417, + 468 + ], + "type": "text", + "content": "(c) image purification" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 439, + 413, + 511, + 454 + ], + "blocks": [ + { + "bbox": [ + 439, + 413, + 511, + 454 + ], + "lines": [ + { + "bbox": [ + 439, + 413, + 511, + 454 + ], + "spans": [ + { + "bbox": [ + 439, + 413, + 511, + 454 + ], + "type": "image", + "image_path": "e94738566a2130b22e9623e5135caff2f0dfb8262984d2437da2d5e8ef2b21b9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 456, + 487, + 468 + ], + "lines": [ + { + "bbox": [ + 460, + 456, + 487, + 468 + ], + "spans": [ + { + "bbox": [ + 460, + 456, + 487, + 468 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 460, + 456, + 487, + 468 + ], + "type": "inline_equation", + "content": "A^5" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 325, + 489, + 416, + 581 + ], + "blocks": [ + { + "bbox": [ + 340, + 476, + 399, + 489 + ], + "lines": [ + { + "bbox": [ + 340, + 476, + 399, + 489 + ], + "spans": [ + { + "bbox": [ + 340, + 476, + 399, + 489 + ], + "type": "text", + "content": "Base: airplane (69.8%) \n[Verified: ship (100.0%)]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 489, + 416, + 581 + ], + "lines": [ + { + "bbox": [ + 325, + 489, + 416, + 581 + ], + "spans": [ + { + "bbox": [ + 325, + 489, + 416, + 581 + ], + "type": "image", + "image_path": "3cbd9e9047d0c7d1cb305fdf4c976c459c03dd15c5fa9f99f8eedd445a435327.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "lines": [ + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": "Figure 1. (a) An input " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in y^*" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " is correctly classified. The grey box (certified bounds [12, 42, 43]) shows that, under an attack " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_A" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "||\\delta \\pmb{x}_A||_{\\infty} < \\epsilon_A" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": ", misclassification " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "((\\pmb{x} + \\delta \\pmb{x}_A) \\in y_0)" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " is possible. (b) Adversarial training [4, 20] creates robust DNNs with regular classification landscapes: misclassification is less likely. (c) Image purification [27, 40] moves the attacked input " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_A" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_P" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " back to correct classification. (d) " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " preemptively moves " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " into a non-attackable position. The original CIFAR10 image classified as airplane (69.8% confidence) can be misclassified under attack (ship, 100% confidence). Once robustified through " + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 589, + 547, + 710 + ], + "type": "text", + "content": " (right), misclassification does not occur anymore." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 431, + 489, + 523, + 580 + ], + "blocks": [ + { + "bbox": [ + 441, + 476, + 511, + 489 + ], + "lines": [ + { + "bbox": [ + 441, + 476, + 511, + 489 + ], + "spans": [ + { + "bbox": [ + 441, + 476, + 511, + 489 + ], + "type": "text", + "content": "Robustified: airplane " + }, + { + "bbox": [ + 441, + 476, + 511, + 489 + ], + "type": "inline_equation", + "content": "(82.6\\%)" + }, + { + "bbox": [ + 441, + 476, + 511, + 489 + ], + "type": "text", + "content": " Verified: airplane " + }, + { + "bbox": [ + 441, + 476, + 511, + 489 + ], + "type": "inline_equation", + "content": "(82.5\\%)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 489, + 523, + 580 + ], + "lines": [ + { + "bbox": [ + 431, + 489, + 523, + 580 + ], + "spans": [ + { + "bbox": [ + 431, + 489, + 523, + 580 + ], + "type": "image", + "image_path": "0b0c10474ff1425b84189394b598963fce90bb20601108c5301961cbf6dfcfe0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4067" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "type": "table", + "html": "
vA vectorvjVector j-th element
CClassifierRRobustifier
xDataδxAAttacking perturbation
δxDDefensive perturbationx + δxAData under attack
x + δxDRobustified datax + δxD + δxARobustified data under attack
εDDefense magnitude, ||δxD||p < εDεRAttack magnitude, ||δxA||p < εR_A while training R
εAAttack magnitude ||δxA||p < εA while testingεCAttack magnitude, ||δxA||p < εC_A while training C
", + "image_path": "f26b1262cb261d31b022106422cef8ec31998a1481eef1f5cf6b919c6374a667.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "lines": [ + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "spans": [ + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "text", + "content": "Table 1. Notation for data " + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "text", + "content": " ,that are the inputs of the classifier " + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "text", + "content": " . We adopt an equivalent notation for physical objects " + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 73, + 167, + 518, + 178 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 199, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 289, + 439 + ], + "type": "text", + "content": "cally arise in automotive [32], audio processing [13, 22, 38], or while communicating with a remote machine [39]. Our idea spouses this reasoning line: we investigate a novel way to augment the data to preemptively certify that it cannot be attacked (Fig. 1, d). This fits in a recent research area that up to now has received less attention than adversarial training. Researchers explored image and real object augmentation to guarantee a high recognition rate in presence of noise or geometric distortions, but not in adversarial scenarios [24]. Encryption schemes coupled with deep learning [1, 28] or watermarking [16] only partially protect against MitM attacks. The few existing preemptive robustification methods [23] include a procedure [17] that first runs a classifier on clean data and achieves preemptive robustification against MitM attacks through iterative optimization; these are partially effective and do not provide any certification. Our novel framework encompasses most of the aforementioned cases while also introducing for the first time the concept of certified robustification. More in detail, our manifold contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 450, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 52, + 450, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 450, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 52, + 450, + 287, + 545 + ], + "type": "text", + "content": "(i) We introduce " + }, + { + "bbox": [ + 52, + 450, + 287, + 545 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 52, + 450, + 287, + 545 + ], + "type": "text", + "content": " (Adversarial Augmentation Against Adversarial Attacks), a comprehensive framework for preemptive augmentation to make data and physical objects certifiably robust against MitM adversarial attacks. As far as we know, this is the first time certified defense is achieved in this way. Since we provide certified robustness, we guarantee protection against any form of white, grey or black box attack." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "text", + "content": "(ii) We test different flavours of " + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "text", + "content": " on standard datasets. By doing so, we study the connection between the robustness of the legacy classifier, the magnitude of the defensive augmentation, and the protection level delivered by " + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "text", + "content": ". We show " + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 50, + 545, + 287, + 701 + ], + "type": "text", + "content": " achieving state-of-the-art certified protection against MitM attacks by training a robustifier DNN coupled with the legacy classifier, and even better results for co-training of the robustifier and classifier. We perform a critical, visual inspection of the robustified images to answer an interesting theoretical question (how does a non-attackable image look like?) and potentially provide directions for the acquisition of inherently robust images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "(iii) Using Optical Character Recognition (OCR) as an ex" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 199, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 199, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 324, + 199, + 545, + 236 + ], + "type": "text", + "content": "ample, we show the application of " + }, + { + "bbox": [ + 324, + 199, + 545, + 236 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 324, + 199, + 545, + 236 + ], + "type": "text", + "content": " for the design of certifiably robust physical objects, which extends [24] to the case of certified defense against MitM attacks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 236, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 297 + ], + "type": "text", + "content": "(iv) We share our code at https://github.com/NVlabs/A5, to allow replicating our results or testing " + }, + { + "bbox": [ + 307, + 236, + 545, + 297 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 307, + 236, + 545, + 297 + ], + "type": "text", + "content": " in scenarios not considered here, for instance on other datasets or for protection against physical adversarial attack (e.g., adversarial patches)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 308, + 481, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 481, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 481, + 323 + ], + "type": "text", + "content": "2. Related Work and Background" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 330, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 545, + 365 + ], + "type": "text", + "content": "We identify three main philosophies for the development of defenses against adversarial attacks. They are not mutually exclusive and can be adopted together." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 367, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 714 + ], + "type": "text", + "content": "The first approach aims at creating DNNs that are robust to adversarial attacks: e.g., in adversarial training, attacks are generated while training the DNN that should process them correctly [4, 11, 20]. This regularizes the classification landscape and makes attacks less effective (Fig. 1, b). Adversarial training is simple but has one drawback: as attack generation is computationally demanding (even after speed up [25]), attacks are randomly sampled and cannot cover the entire input space, thus impacting the final accuracy. Certified defense [12, 42, 43] can be seen as a state of the art refined form of adversarial training: robustness is obtained by propagating through the DNN an input interval that encompasses all the attacks with magnitude up to " + }, + { + "bbox": [ + 304, + 367, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C" + }, + { + "bbox": [ + 304, + 367, + 546, + 714 + ], + "type": "text", + "content": ". The corresponding set of output bounds allows estimating the worst case scenario and derive the training cost function (e.g., worst case entropy [12, 42]). Research focuses on finding bounds that are both tight and computationally light. IBP [12] computes (and backwards) the bounds at the cost of an additional forward and backward pass. Since the bounds are loose at the beginning of training, a complex schedule is required to guarantee stability and convergence. Linear relaxation [43] delivers tighter bounds leveraging the correlation between different layers to linearly propagate signal and bounds, when possible, and eventually put in linear relation the input interval and output bounds. CROWN [43] bounds are often tighter than IBP, but at an impractical computational cost for large DNNs. Among the attempts to speed up CROWN [33, 37], CROWN-IBP [42] strikes the right balance of computa" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4068" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": "tional complexity and tightness. It is implemented in the auto_LiRPA" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": " library [36, 37] adopted here. One limitation of adversarial training, certified methods, and " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "A^{5}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": " as well, is the trade-off between clean and certified (worst case underattack) accuracy: increasing the latter leads to a decrease in the former [4, 12, 20, 42]. Among the architecture solutions against adversarial attacks, we also mention the recent development of " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": "-dist neurons that are inherently robust [41], and could be leveraged by " + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "inline_equation", + "content": "A^{5}" + }, + { + "bbox": [ + 46, + 72, + 286, + 178 + ], + "type": "text", + "content": " as well in future." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 286, + 395 + ], + "type": "text", + "content": "The second defense philosophy is based on active countermeasures taken after the attack is crafted. Image purification [18, 27, 40] projects an attacked image onto the manifold of natural images through a generative model (Fig. 1, c). It is less effective than adversarial training [7], but does not make any assumption on the classifier. Other forms of signal pre-processing have been proposed in the field of audio processing and automatic speech recognition, where MitM physical attacks can be easily implemented [22, 38] and partially defeated by randomized smoothing, DefenseGAN, variational autoencoder (VAE), and Parallel WaveGAN vocoder [13]. Both certified defenses and pre-processing methods do not assume any specific attack form, but the latter are less effective in white box scenarios [2, 31]. Randomization can also be used after the attack, e.g. by changing resolution or padding an image [35], or by picking random experts in a mixture, that is proved to be more robust than deterministic DNNs [21]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " belongs to a third class of methods, that put in place countermeasures before the attack (Fig. 1, d). The roots of " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " lie in two ideas that have been only partially investigated so far. The first is that, in adversarial scenarios, the first actor to move has a significant advantage. This rule is already leveraged for online privacy, where a preemptive adversarial attack prevents the identification of the framed subject [6, 19]. Since adversarial attacks often generalize to many classifiers, this ploy is particularly effective [26]. Preemptive protection is also a natural form of defense against MitM attacks that can be easily created in contexts like automotive [32], speech and audio processing [13, 22, 38], or anytime the capturing device communicates with a server [39]. Encryption [1, 28] and watermarking [16] coupled with deep learning work against MitM attacks, but do not achieve full protection, are not certified, require the key to be hidden (grey box scenario) and may significantly alter the appearance of the image. Blindpreprocessing [23] applies a tanh transform, normalization, quantization and thermometer encoding to achieve partial and uncertified defense. The closest algorithm to " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " is preemptive robustification [17]: given an image " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " that can be fine-tuned, it computes " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "y_0 = C(\\pmb{x})" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " and randomly samples attacks " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_A" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " against " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_D" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": " to find a defensive perturbation " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_D" + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 46, + 395, + 286, + 694 + ], + "type": "inline_equation", + "content": "(\\pmb{x} + \\delta \\pmb{x}_D + \\delta \\pmb{x}_A) \\in y_0" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": "(correct classification under attack). Since " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " uses certified bounds, it results in simpler training and higher protection, and does not require running " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " first, as our robustifier DNN generalizes to all the inputs. The second idea behind " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " comes from physical adversarial examples, i.e., 3d physical objects [3] or 2d patches [5] that consistently fool a DNN under a wide range of viewing conditions. Reverting the attacker and defender roles, the complementary cost function leads to the creation of unadversarial examples: patches or textures that, once applied to a real 3d object, reinforce (instead of degrading) the DNN desired behaviour [24]. However, these are not designed to be inherently robust to adversarial attacks. We exploit the idea in " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " in all those situations where system designers control, to some extent, the inputs fed to the DNN (e.g. in OCR, robotic systems or during the design of the road infrastructure). We show as an example the use of " + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 306, + 72, + 545, + 288 + ], + "type": "text", + "content": " to design certifiably protected characters for OCR, something unexplored both in [17] and [24]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 298, + 361, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 361, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 361, + 311 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": "Threat models We summarize our notation in Table 1, whereas Fig. 2 illustrates the full " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": " framework including trainable DNNs in blue, training algorithms in green, physical objects " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": ", and acquired data " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": ". MitM attacks are crafted while transmitting data from the acquisition device to the classifier " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": ", while physical attacks (not tested here) are crafted onto physical objects " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": ", before data acquisition. " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": " can be used in a scenario where the attacker has full access to the " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": ": since it uses certified bounds, it is agnostic to the specific nature of the attack. In other words, we certify protection for any " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\delta x_A" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "||\\delta x_A|| < \\epsilon_A" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": " and the attacker can use any white-, grey- or black-box attacking algorithm. We assume that " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": " can run in a protected environment, not accessible to the attacker, to perform robustification (light blue rectangles in Fig. 2). More details and limitations of our threat model are discussed in the Supplementary, whereas here we first introduce useful definitions and then provide several recipes for the use of " + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 305, + 319, + 545, + 534 + ], + "type": "text", + "content": " in different scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "text", + "content": "Definitions For a classifier " + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "text", + "content": " and input " + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = C(\\mathbf{x})" + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "text", + "content": " is the logit vector output by " + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 549, + 545, + 573 + ], + "type": "text", + "content": ". Under attack, we have:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 342, + 583, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 583, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 342, + 583, + 545, + 597 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} + \\delta \\boldsymbol {y} = C (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {A}), \\| \\left| \\delta \\boldsymbol {x} _ {A} \\right| | _ {p} < \\epsilon_ {A}, \\tag {1}", + "image_path": "23e41931bf5b565efe7634f6cad87a2156a64fb7fce30ceb7976fd0e2f816c85.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " is the magnitude of the attack " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_A" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": "-norm and " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{y}" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " is the logit perturbation. Linear relaxation [43] estimates the lower " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "(\\pmb{y}^l)" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " and upper " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "(\\pmb{y}^u)" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " bounds of " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y} + \\delta \\pmb{y}" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": ", for instance using IBP [12], CROWN [43], CROWN-IBP [42], or a combination of them. The margin for the class " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " is the difference between the lower bound of the ground truth class " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "j^{*}" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " and the upper bound of " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": ", while it is zero by definition for " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "j^{*}" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": ". More formally, for " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " classes the elements of the margin vector " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{m} = [m_0, m_1, \\dots, m_{M - 1}]" + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "text", + "content": " are: " + }, + { + "bbox": [ + 305, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "m_{j^{*}} = 0" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 249, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 249, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 249, + 712 + ], + "type": "text", + "content": "Automatic Linear Relaxation based Perturbation Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4069" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 70, + 486, + 140 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 486, + 140 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 486, + 140 + ], + "type": "image", + "image_path": "735e07d238056c1a05cd65e211b875af78b4d89cc3cfb1e9ab64a214dc1f33dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "text", + "content": "Figure 2. Schematic representation of the full " + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "text", + "content": " framework. It allows the robust augmentation of both physical objects " + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "text", + "content": " and acquired data " + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 46, + 146, + 547, + 169 + ], + "type": "text", + "content": " to make them certifiably robust against MitM and physical (not tested here) adversarial attacks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "inline_equation", + "content": "m_j = \\pmb{y}_{j^*}^l -\\pmb{y}_j^u" + }, + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\forall j\\neq j^{*}" + }, + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "text", + "content": " . The confidence is obtained by applying softmax to " + }, + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\pmb{m}" + }, + { + "bbox": [ + 46, + 189, + 287, + 214 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 223, + 287, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 223, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 110, + 223, + 287, + 257 + ], + "type": "interline_equation", + "content": "p _ {j} (\\boldsymbol {x}) = e ^ {- m _ {j}} / \\sum_ {k = 0} ^ {M - 1} e ^ {- m _ {k}}, \\tag {2}", + "image_path": "0b2ab3b7fd8fedfd17d821a782edb72bf7d7cb23ad54d7c79a375f349b42e3ed.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 266, + 235, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 235, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 235, + 278 + ], + "type": "text", + "content": "and thus the worst case cross entropy " + }, + { + "bbox": [ + 47, + 266, + 235, + 278 + ], + "type": "inline_equation", + "content": "E(\\pmb{x})" + }, + { + "bbox": [ + 47, + 266, + 235, + 278 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 288, + 287, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 288, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 287, + 322 + ], + "type": "interline_equation", + "content": "E (\\boldsymbol {x}) = - \\sum_ {j = 0} ^ {M - 1} \\hat {y} _ {j} \\log \\left(p _ {j} (\\boldsymbol {x})\\right), \\tag {3}", + "image_path": "7b9be30744eefbb0b676696536b1e6adfb1569cff8ec4f2779df072e3a2d3c8d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\hat{y}_j" + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "text", + "content": " is the one-hot-encoding of " + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "inline_equation", + "content": "j^*" + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "text", + "content": ". To formally introduce " + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 47, + 331, + 287, + 357 + ], + "type": "text", + "content": ", we first give the following definitions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": "Definition 1 (" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\epsilon_A" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": "-robustness). Given an input " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": ", we say that " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\epsilon_A" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": "-robust with respect to " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": " in norm " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\forall \\delta \\pmb{x}_A, ||\\delta \\pmb{x}_A||_p < \\epsilon_A" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "inline_equation", + "content": "C(\\pmb{x} + \\delta \\pmb{x}_A)" + }, + { + "bbox": [ + 47, + 365, + 287, + 413 + ], + "type": "text", + "content": " generates the correct classification output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": "Definition 2 (robustification). For a given input " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": ", robustification is the process of finding a defensive augmentation " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_D" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "||\\delta \\pmb{x}_D||_p < \\epsilon_D" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_D" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\epsilon_A" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": "-robust with respect to " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": " in norm " + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 422, + 287, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": "We compute certified CROWN-IBP bounds through auto_LiRPA [37] and optimize " + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "inline_equation", + "content": "E(\\pmb{x} + \\delta \\pmb{x}_D)" + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": " to find preemptive, defensive perturbations " + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_D" + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": " that make " + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_D \\in \\epsilon_A" + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": "-robust, following the training recipes provided here. We also show how to find defensive perturbations " + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{w}_D" + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": " for physical objects " + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 46, + 479, + 287, + 552 + ], + "type": "text", + "content": " using the same approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": "Offline robustification with ground truth " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "(A^5 /O)" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": ": Consider a legacy classifier " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": " that has been certifiably trained (e.g. through CROWN-IBP [42]) against attacks of magnitude " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": ". Given an input " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": " and its class " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "j^{*}" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": ", we want to find a defensive perturbation " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_D,||\\delta \\pmb{x}_D||_p < \\epsilon_D" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_D" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": "-robust with respect to " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": "-norm. Our solution to this problem is " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": "/Offline (or " + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "inline_equation", + "content": "A^5 /O" + }, + { + "bbox": [ + 46, + 567, + 287, + 651 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": "To force " + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "inline_equation", + "content": "||\\delta \\pmb{x}_D||_p < \\epsilon_D" + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": ", we parameterize " + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{x}_D" + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": " via a vector " + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 47, + 651, + 287, + 676 + ], + "type": "text", + "content": " to which we apply an element-wise sigmoid:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 685, + 287, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 685, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 59, + 685, + 287, + 699 + ], + "type": "interline_equation", + "content": "\\delta \\boldsymbol {x} _ {D} = \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) = [ \\delta x _ {D} (\\boldsymbol {z}) _ {0}, \\dots , \\delta x _ {D} (\\boldsymbol {z}) _ {K - 1} ] \\tag {4}", + "image_path": "8bc9613d45bda73b346f41d9da4f2b666fa6eebcc7ed1b3bd7094ce325026d78.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 99, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 99, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\delta x _ {D} (\\boldsymbol {z}) _ {j} = 2 \\epsilon_ {D} \\left[ 1 / \\left(1 + e ^ {- z _ {j}}\\right) - 0. 5 \\right], \\tag {5}", + "image_path": "49903c152c9c70bae22e9f023ca3c1c794288ad1072f4000ab9442d7fe511232.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "text", + "content": " is the number of elements in " + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "text", + "content": ". This is a robust classification problem where " + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "text", + "content": " is the only unknown. For any " + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 306, + 189, + 545, + 226 + ], + "type": "text", + "content": ", we minimize the worst case cross entropy:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 362, + 236, + 545, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 236, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 362, + 236, + 545, + 249 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} = \\operatorname {a r g m i n} _ {\\tilde {\\boldsymbol {z}}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\tilde {\\boldsymbol {z}}) ], \\tag {6}", + "image_path": "ade34407370a77f9022d262d971eb9e6af05c024fa30f12190a8948674642e4e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "text", + "content": "using RMSProp. Notice that the magnitude of the attack used during robustification " + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "inline_equation", + "content": "(\\epsilon_A^R)" + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "text", + "content": " and that adopted to train " + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "inline_equation", + "content": "(\\epsilon_A^C)" + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "text", + "content": " can be different. Furthermore, similarly to adversarial training and certified methods, these can also be different from the target attack magnitude, " + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}" + }, + { + "bbox": [ + 305, + 258, + 545, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": " is similar in spirit to preemptive image robustification [17], but does not need " + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": " to run on the clean input before robustification. Although " + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": " could be used in a similar way to robustify " + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": " accordingly to the class " + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "C(\\pmb{x})" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": ", running " + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": " first and then iteratively solving the optimization problem in Eq. 6 makes it hardly usable in practice. Therefore, it does not directly serve any practical purpose, but it helps answering important theoretical questions (e.g., under which conditions can we robustify " + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 319, + 546, + 450 + ], + "type": "text", + "content": "? What does a robustified image looks like?) and establishing a baseline with known ground truth and no computational constraints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": "Online robustification with robustifier " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", legacy " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": "): We introduce a second recipe, " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", where we train a robustifier DNN " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " that takes in input " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", no ground truth label, and outputs " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", from which a defensive perturbation " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{x}_D" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "||\\delta \\mathbf{x}_D||_p < \\epsilon_D" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", is computed as in Eq. 5. We assume again that a robust legacy " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " is given and we train " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " to make each " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\delta \\mathbf{x}_D" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " in the training dataset " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": "-robust with respect to " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": "-norm. More formally, the weights " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\theta_R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": " are found by minimization of the average worst case entropy on the training dataset " + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "inline_equation", + "content": "E(\\theta_R)" + }, + { + "bbox": [ + 305, + 467, + 546, + 586 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 597, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} = R (\\boldsymbol {x} \\mid \\boldsymbol {\\theta} _ {R}), E (\\boldsymbol {\\theta} _ {R}) = (1 / N) \\sum_ {\\boldsymbol {x}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) ], \\tag {7}", + "image_path": "71be73d399903a90403eb4ff502da351e87fdc5393c842b277e3853fdb4c10bb.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " is the dataset size. Unlike preemptive robustification [17], " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " does not iteratively solve a complex optimization problem, neither we need running " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " before " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": ". Therefore, we can use " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " to protect " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " against " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{MitM}" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " adversarial attacks on-the-fly, ignoring its class, soon after its acquisition, and before its transmission to " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " (see " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " in Supplementary). Furthermore, since " + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 629, + 545, + 714 + ], + "type": "text", + "content": " is trained for a specific" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4070" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": ", it can be deployed on the field (e.g., into the firmware of the acquisition devices) without changing (and thus maintaining the legacy of) the classifier infrastructure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": "Online robustification with robustifier " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": ", retrained " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": "): The " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": " recipe is equivalent to " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": ", but it leverages the co-adaptation of " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": " by joint training (see " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": " in Supplementary). The cost function is again the average worst case entropy " + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "inline_equation", + "content": "E(\\theta_C, \\theta_R)" + }, + { + "bbox": [ + 47, + 122, + 287, + 183 + ], + "type": "text", + "content": ", defined now as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 191, + 287, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 191, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 69, + 191, + 287, + 205 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} = R (\\boldsymbol {x} | \\boldsymbol {\\theta} _ {R}), \\quad \\boldsymbol {y} = C (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} | \\boldsymbol {\\theta} _ {C}), \\tag {8}", + "image_path": "5bf380e639ab4ba76226154307924d5f35a765eb04d60ae84cd1fc68c6ea8d56.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 205, + 287, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 205, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 73, + 205, + 287, + 230 + ], + "type": "interline_equation", + "content": "E \\left(\\boldsymbol {\\theta} _ {C}, \\boldsymbol {\\theta} _ {R}\\right) = (1 / N) \\sum_ {\\boldsymbol {x}} E [ \\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} (\\boldsymbol {z}) ], \\tag {9}", + "image_path": "49e1d5f2c1a34078035d26d96f17f1530aa34e63a39d4bbfd6a270778001fb33.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\theta_{C}" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " are the parameters of " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": ". Training " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " from scratch is often unstable and requires careful tuning to achieve convergence (likewise robust classifiers [12, 42]). Therefore, we first train " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "A^5 / C" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " and then fine tune " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " together. During the tuning phase, we can use random input augmentation to prevent overfitting (see Section 4 and Supplementary for details). Differently from " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": " does not preserve the legacy classifier " + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 236, + 287, + 357 + ], + "type": "text", + "content": ", but it achieves the same scope (with more effectiveness) in terms of protection against MitM attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": "Offline physical robustification with " + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "inline_equation", + "content": "(A^5 /PC)" + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": " and without " + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "inline_equation", + "content": "(A^5 /P)" + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": " retraining " + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": ", with ground truth: This recipe aims at producing physical objects that are certifiably robust against MitM attacks (see " + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "inline_equation", + "content": "A^5 /P" + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "inline_equation", + "content": "A^5 /PC" + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": " in Supplementary). Similarly to unadversarial objects [24], we assume that the system creator can design the entire infrastructure and the attacker cannot interfere in this phase. The problem is formally defined for " + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "inline_equation", + "content": "A^5 /PC" + }, + { + "bbox": [ + 47, + 370, + 287, + 467 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 475, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 475, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 66, + 475, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D} = A (\\boldsymbol {w} + \\delta \\boldsymbol {w} _ {D}), \\| \\delta \\boldsymbol {w} _ {D}) \\| < \\epsilon_ {D}, \\tag {10}", + "image_path": "5fdaa37d056a2a581ddaf791372ceda2d2d395156817e50cad6f32fb9822a251.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 490, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 490, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 82, + 490, + 287, + 514 + ], + "type": "interline_equation", + "content": "E \\left(\\boldsymbol {\\theta} _ {C}, \\delta \\boldsymbol {w} _ {D}\\right) = (1 / N) \\sum_ {\\boldsymbol {x}} E (\\boldsymbol {x} + \\delta \\boldsymbol {x} _ {D}) \\tag {11}", + "image_path": "144f86fa36e72bc6ec650caca9a65618544166cf38b0e939b7d7e6b6b7cf473d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " is the object appearance, whereas " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{x} = A(\\pmb{w})" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " is the data acquisition model (e.g., image formation). " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " is here the number of random samples generated as " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_D = A(\\pmb{w} + \\delta \\pmb{w}_D)" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": "; it is important noticing in fact that the acquisition process is not deterministic: as an exemplary case we mention OCR, where " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " is the shape of a character, while " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " may include random perspective transformations, contrast and brightness changes, noise addition and blurring, that may occur on a real camera. The defensive perturbation " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\delta \\pmb{w}_D" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " represents a change in the physical character appearance " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " that decreases the success rate of an attack towards the acquired image " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": " (therefore after framing the character with " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": ") and the classifier " + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 522, + 287, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^5 / P" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " we change the character shapes and use the legacy classifier " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "; in " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "A^5 / PC" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " we also train " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " while estimating the robust characters " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w} + \\delta \\boldsymbol{w}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ". In both cases " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 69, + 545, + 107 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 107 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 107 + ], + "type": "table", + "html": "
εAεAC=0.0εAC=0.1εAC=0.2εAC=0.3εAC=0.4
0.1——1.17 [—, 2.24] [42]——
0.10.81 [89.02, 100.00]0.92 [2.00, 5.41]1.32 [2.14, 2.60]1.44 [2.33, 2.63]2.42 [3.36, 3.41]
0.3———1.82 [—, 7.02] [42]—
0.30.69 [100.00, 100.00]0.94 [98.41, 100.00]1.30 [99.62, 100.00]1.41 [7.06, 9.98]2.58 [7.03, 7.68]
", + "image_path": "ede81c608ee7ac20dc79d9244f93dcd2c39ce9c0cb68e372430531962e071ec3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "type": "text", + "content": "Table 2. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST classifiers trained by us with CROWN-IBP [42], under attack " + }, + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "type": "inline_equation", + "content": "\\epsilon_{A} = \\{0.1,0.3\\}" + }, + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "type": "text", + "content": "; we also show the metrics reported in [42]. A " + }, + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 115, + 545, + 171 + ], + "type": "text", + "content": " certified error rate identifies DNNs that are not protected against attacks." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 181, + 545, + 248 + ], + "blocks": [ + { + "bbox": [ + 307, + 181, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 307, + 181, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 181, + 545, + 248 + ], + "type": "table", + "html": "
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.79 [87.09, 100.00]0.82 [1.51, 2.89]1.22 [1.64, 1.86]1.34 [1.76, 1.87]2.38 [2.91, 2.92]
0.10.100.79 [86.51, 100.00]0.72 [1.21, 1.92]1.15 [1.43, 1.49]1.34 [1.50, 1.58]2.41 [2.71, 2.74]
0.10.200.89 [86.40, 100.00]0.56 [0.78, 1.04]1.13 [1.18, 1.21]1.35 [1.37, 1.38]2.40 [2.55, 2.55]
0.10.300.93 [86.44, 100.00]0.51 [0.57, 0.70]1.14 [1.16, 1.17]1.34 [1.37, 1.37]2.38 [2.52, 2.52]
0.10.400.99 [85.93, 100.00]0.52 [0.53, 0.60]1.14 [1.16, 1.16]1.35 [1.38, 1.38]2.44 [2.51, 2.51]
0.30.050.71 [100.00, 100.00]0.93 [97.18, 100.00]1.34 [99.17, 100.00]1.35 [5.16, 6.97]2.55 [6.00, 6.42]
0.30.100.72 [100.00, 100.00]0.88 [95.32, 100.00]1.36 [98.58, 100.00]1.33 [4.12, 5.19]2.49 [5.54, 5.85]
0.30.200.72 [100.00, 100.00]0.88 [90.62, 100.00]1.41 [96.21, 100.00]1.32 [3.21, 3.77]2.35 [5.02, 5.22]
0.30.300.73 [100.00, 100.00]0.89 [86.28, 100.00]1.41 [94.18, 100.00]1.31 [2.91, 3.40]2.32 [4.66, 4.81]
0.30.400.81 [100.00, 100.00]0.85 [83.49, 100.00]1.42 [92.63, 100.00]1.32 [2.74, 3.20]2.28 [4.16, 4.29]
", + "image_path": "08815ecb0e201d5f8d7ac9af25174fee5770a5dd3df6f3b992b1666792f726b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "text", + "content": "Table 3. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, " + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "text", + "content": ", under attack " + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\epsilon_A = \\{0.1, 0.3\\}" + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "text", + "content": ". During training we use " + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R = \\epsilon_A" + }, + { + "bbox": [ + 305, + 257, + 545, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 307, + 300, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 307, + 300, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 307, + 300, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 545, + 369 + ], + "type": "table", + "html": "
εA^R = εAεDεC^A = 0.0εC^A = 0.1εC^A = 0.2εC^A = 0.3εC^A = 0.4
0.10.050.82 [85.50, 100.00]0.95 [1.71, 3.58]1.30 [1.82, 2.02]1.40 [1.86, 2.03]2.43 [2.96, 2.97]
0.10.100.81 [82.87, 100.00]1.00 [1.48, 2.54]1.19 [1.62, 1.73]1.40 [1.68, 1.74]2.38 [2.77, 2.79]
0.10.200.88 [78.48, 100.00]0.93 [1.33, 1.77]1.20 [1.42, 1.52]1.31 [1.53, 1.59]2.26 [2.55, 2.57]
0.10.300.99 [75.89, 100.00]0.91 [1.16, 1.54]1.14 [1.31, 1.36]1.26 [1.43, 1.47]2.20 [2.42, 2.43]
0.10.401.04 [75.69, 100.00]0.87 [1.05, 1.35]1.14 [1.27, 1.33]1.24 [1.43, 1.46]2.22 [2.41, 2.43]
0.30.050.69 [100.00, 100.00]0.94 [96.93, 100.00]1.33 [99.11, 100.00]1.41 [5.13, 6.91]2.58 [5.77, 6.09]
0.30.100.71 [100.00, 100.00]0.91 [95.09, 100.00]1.33 [98.47, 100.00]1.38 [3.86, 4.81]2.55 [4.71, 4.83]
0.30.200.72 [100.00, 100.00]0.92 [90.13, 100.00]1.37 [96.08, 100.00]1.41 [2.26, 2.63]2.55 [3.33, 3.40]
0.30.300.75 [100.00, 100.00]0.93 [85.54, 100.00]1.35 [93.36, 100.00]1.32 [1.86, 1.99]2.37 [2.86, 2.88]
0.30.400.83 [100.00, 100.00]0.96 [82.48, 100.00]1.36 [92.15, 100.00]1.29 [1.69, 1.83]2.35 [2.67, 2.69]
", + "image_path": "09a44271a78f72cfa48ea272b6121265766c2001d0bbd38cd6a6284714cc8d35.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "text", + "content": "Table 4. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for MNIST, " + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "text", + "content": ", under attack " + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon_A = \\{0.1, 0.3\\}" + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "text", + "content": ". During training we use " + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R = \\epsilon_A" + }, + { + "bbox": [ + 305, + 376, + 545, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "type": "text", + "content": "is previously trained to be robust, e.g., through CROWN-IBP [42]. The " + }, + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "type": "inline_equation", + "content": "A^5 / P" + }, + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "type": "inline_equation", + "content": "A^5 / PC" + }, + { + "bbox": [ + 305, + 431, + 545, + 491 + ], + "type": "text", + "content": " recipes find application in all those situations where the developer can design the entire infrastructure; beyond OCR, other examples are road sign or robotic tool design and classification." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 502, + 435, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 435, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 435, + 514 + ], + "type": "text", + "content": "4. Results and discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": "We test " + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": " on MNIST [10], CIFAR10 [14], FashionMNIST [34], and Tinyimageenet [15], that have been widely used to establish significant milestones in the adversarial defense field. Training schedules and DNN architectures are given in the Supplementary. In all cases we use a " + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "inline_equation", + "content": "p = \\infty" + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": " norm for the attack and defense. Our intention is to show the improvement achieved by " + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": " over traditional, certified methods like CROWN-IBP. The reader should anyway keep in mind that the direct comparison of " + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 521, + 545, + 653 + ], + "type": "text", + "content": " and CROWN-IBP is partially unfair, as CROWN-IBP does not apply any preemptive defense." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We first use CROWN-IBP to train five certifiably robust MNIST classifiers with varying levels of robustness, for training attacks " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C = 0.0" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " (unprotected " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C = \\{0.1, 0.2, 0.3, 0.4\\}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Their clean and certified errors are reported in Table 2, for attacks " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_A = \\{0.1, 0.3\\}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". We also" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4071" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 286, + 137 + ], + "type": "table", + "html": "
εA=εCεDεCA=0.0εCA=0.1εCA=0.2εCA=0.3εCA=0.4
0.10.0588.66[88.66, 88.66]1.00[1.38, 2.35]1.16[1.61, 2.46]1.14[1.54, 2.29]1.37[1.72, 2.33]
0.10.1088.66[88.66, 88.66]0.77[0.95, 1.40]0.99[1.16, 1.56]0.86[1.03, 1.42]1.13[1.31, 1.72]
0.10.2088.66[88.66, 88.66]0.69[0.80, 1.05]0.84[0.94, 1.25]0.96[1.07, 1.33]0.88[1.10, 1.54]
0.10.3088.66[88.66, 88.66]0.64[0.70, 1.02]0.91[0.99, 1.16]0.79[0.88, 1.21]1.00[1.12, 1.43]
0.10.4088.66[88.66, 88.66]0.61[0.68, 0.88]0.93[0.99, 1.18]0.69[0.76, 1.10]1.00[1.03, 1.30]
0.30.0588.66[88.66, 88.66]88.66[88.66, 88.66]2.57[8.35, 10.86]3.42[9.51, 11.72]8.00[18.05, 19.77]
0.30.1088.66[88.66, 88.66]88.66[88.66, 88.66]2.07[5.66, 7.56]2.35[6.18, 8.07]2.35[6.31, 7.93]
0.30.2088.66[88.66, 88.66]1.68[3.11, 4.79]1.40[2.42, 3.62]1.28[2.82, 4.01]1.88[3.53, 4.90]
0.30.3088.66[88.66, 88.66]0.84[0.96, 1.38]1.00[1.14, 1.36]0.94[1.15, 1.59]1.07[1.39, 1.90]
0.30.4088.66[88.66, 88.66]0.72[0.80, 1.03]0.75[0.82, 1.06]0.91[1.14, 1.68]0.87[1.14, 1.50]
", + "image_path": "e3ff4516848f66abe0c02a37a7aac280b601ef872b1eb8cb825f02b4b315b5c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 286, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 286, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 286, + 284 + ], + "type": "text", + "content": "report the error rates estimated with autoattack [8, 9], an automatic toll based on an ensemble of four attacks to evaluate the practical DNN robustness. These error rates are in the same ballpark of those reported in the CROWN-IBP paper [42]: they represent the baseline for the analysis of the different " + }, + { + "bbox": [ + 46, + 213, + 286, + 284 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 213, + 286, + 284 + ], + "type": "text", + "content": " recipes considered here." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": "For each " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " we run " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "A^5 /O" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " on the MNIST test set, for defense magnitudes " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{D} = \\{0.05,0.1,0.2,0.3,0.4\\}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ". During the optimization of the entropy in Eq. 6, we use attacks " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R = \\{0.1,0.3\\}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ", not necessarily equal to " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ". Table 3 reports the clean, autoattack, and certified errors for a testing attack " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{A} = \\epsilon_{A}^{R}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " is not robust (" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}^{C} = 0.0" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "A^5 /O" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " does not reduce the certified error², otherwise it generally overcomes the base CROWN-IBP classifier, as shown by comparing against the error rates in Table 2. As a rule of thumb, the best clean / certified error trade off is often achieved for " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}^{R} = \\epsilon_{A}^{C} = \\epsilon_{A}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ". Results also improve, as expected, for larger defenses " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{D}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ", when " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "A^5 /O" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " beat CROWN-IBP by a significant margin on clean and certified errors. However, " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "A^5 /O" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " does not work well when the legacy " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " is trained with a large " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}^{C}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ", possibly because of its initial low clean accuracy. The results are consistent with our interpretation of the effects of adversarial training and robustification on the classification landscape (Fig. 1): training " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " with a large " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}^{C}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " creates classification valleys wide enough to accommodate " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\pmb{x} + \\delta \\pmb{x}_D + \\delta \\pmb{x}_A" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": ", rendering the adversarial attacks ineffective. " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " moves any " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " towards the centers of the valleys; a large " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{D}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " allows moving them more easily. Too large " + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "inline_equation", + "content": "\\epsilon_{A}^{C}" + }, + { + "bbox": [ + 46, + 285, + 286, + 560 + ], + "type": "text", + "content": " remains however detrimental for the clean accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": "Table 4 reports results for " + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": ", demonstrating that training a DNN robustifier " + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": " for on-the-fly, certified robustification, ignoring the ground truth, is feasible. Likewise " + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": " beats the base CROWN-IBP classifier by a significant margin. Sometimes " + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": " does better by leveraging the knowledge of the ground truth class, whereas the regular robustification signal generated by " + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 46, + 561, + 286, + 656 + ], + "type": "text", + "content": " is more effective in other cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "text", + "content": "Our last MNIST experiment investigates " + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "inline_equation", + "content": "A^5 /RC" + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "text", + "content": " that does not preserve the legacy " + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "text", + "content": " like " + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "inline_equation", + "content": "A^5 /R" + }, + { + "bbox": [ + 47, + 657, + 286, + 682 + ], + "type": "text", + "content": ", but achieves bet" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 69, + 545, + 115 + ], + "blocks": [ + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "text", + "content": "Table 5. Error on clean data and (within brackets) under autoattack [8,9] and certified, for MNIST, " + }, + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "text", + "content": ", under attack " + }, + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_A = \\{0.1, 0.3\\}" + }, + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "text", + "content": ". During training we use " + }, + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R = \\epsilon_A" + }, + { + "bbox": [ + 46, + 144, + 287, + 190 + ], + "type": "text", + "content": ", and data augmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 69, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 545, + 115 + ], + "type": "table", + "html": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2554/25552.75 [63.47, 64.09]53.34 [62.42, 62.79]53.75 [64.92, 65.49]
8/2558/25551.06 [60.04, 60.73]52.36 [59.24, 59.58]53.16 [63.50, 64.03]
8/25516/25548.72 [54.89, 55.25]50.60 [55.22, 55.38]52.23 [60.79, 61.14]
8/25532/25545.68 [49.51, 49.74]48.06 [50.86, 50.94]50.99 [57.37, 57.61]
", + "image_path": "abea04267df01868f912526fdaf5e4bc00886cf477a56bfa0d65a5a94527531c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 201, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "text", + "content": "Table 6. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, " + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "text", + "content": ", under attack " + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_A = 8/255" + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "text", + "content": ". During training we use different " + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_D" + }, + { + "bbox": [ + 305, + 124, + 545, + 190 + ], + "type": "text", + "content": ". The robust classifier trained with CROWN-IBP has clean [and certified] errors equal to 54.02 [66.94] or 45.47 [69.55], depending on the CROWN-IBP training configuration [42]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 201, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 258 + ], + "type": "table", + "html": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25554.60 [66.34, 67.11]54.60 [66.34, 67.11]54.60 [66.31, 67.11]
8/2554/25553.73 [64.07, 64.57]54.12 [63.32, 63.74]54.37 [63.58, 64.03]
8/2558/25552.91 [62.59, 63.12]53.24 [61.38, 61.93]54.54 [60.85, 61.16]
8/25516/25551.58 [60.40, 61.02]51.95 [59.21, 59.67]54.44 [59.87, 60.11]
8/25532/25550.91 [57.26, 57.75]51.70 [56.91, 57.43]54.05 [58.39, 58.73]
", + "image_path": "6de69708d9c905b2a0e61320a73dd5616fa84ec7cb74ec82430c6ca1ebd81e0d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 343, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": "Table 7. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, " + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": ", under attack " + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\epsilon_A = 8 / 255" + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": ". During training we use different " + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\epsilon_D" + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": ". The robustifier " + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": " in the first row uses " + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "inline_equation", + "content": "\\epsilon_D = 0" + }, + { + "bbox": [ + 305, + 266, + 545, + 332 + ], + "type": "text", + "content": " and thus it does not provide any form of defense — the error reported here is equivalent to that of a robust classifier trained with CROWN-IBP [42]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 343, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 306, + 343, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 545, + 401 + ], + "type": "table", + "html": "
εC_AεDεR_A=4/255εR_A=8/255εR_A=16/255
8/2550/25550.24 [69.17, 73.65]58.63 [69.25, 70.63]74.28 [77.82, 77.91]
8/2554/25546.76 [63.03, 70.99]54.05 [62.80, 64.44]67.59 [71.20, 71.32]
8/2558/25541.24 [54.01, 64.31]50.74 [56.63, 59.49]65.96 [67.63, 67.78]
8/25516/25536.96 [42.45, 52.90]39.15 [41.33, 44.22]58.17 [58.67, 58.85]
8/25532/25535.26 [37.13, 42.76]41.59 [43.01, 45.55]42.25 [43.02, 44.08]
", + "image_path": "fc002a5be2e733ed779ebfc231fa28b42a6c9665f9c5e8f6c8c155d33b2000b5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": "Table 8. Error on clean data and (within brackets) under autoattack [8, 9] and certified, for CIFAR10, " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": ", under attack " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_A = 8/255" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": ". During training we use different " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_D" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": ". The robustifier " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": " in the first row uses " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_D = 0" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": " and thus it does not provide any form of defense, whereas the corresponding robust CROWN-IBP classifier is fine tuned using " + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C = \\epsilon_R" + }, + { + "bbox": [ + 305, + 409, + 545, + 475 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "text", + "content": "ter clean accuracy and protection (Table 5). In the Supplementary we discuss the role of data augmentation to achieve this result. Interestingly, the best results are obtained when " + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "text", + "content": " is initially trained for " + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C \\leq \\epsilon_A" + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "text", + "content": ", probably because " + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "text", + "content": " has an initial high clean accuracy and subsequently co-adapts with " + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 498, + 545, + 569 + ], + "type": "text", + "content": " to also guarantee a small certified error." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": "Fig. 3 shows examples of robustification on MNIST. At visual inspection, " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 304, + 570, + 545, + 641 + ], + "type": "text", + "content": " act similarly on these images: they add contrast and enhance the high frequency features (i.e., edges) of the digits. The few failure cases are often pictures that look ambiguous even to a human observer (right panels in Fig. 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": "We perform tests on CIFAR10 and compare " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": " against a state of the art CROWN-IBP classifiers " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": ", trained with " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C = 8.8 / 255" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": ", that we also adopt as legacy " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": ". To test the effect of the magnitude of the attack used for robustification (with " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": ") or for training " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "A^5 / R, A^5 / RC" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": "), we use " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R = \\{4 / 255, 8 / 255, 16 / 255\\}" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": " that are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "text", + "content": "2Coherently with [24], the worst case cross entropy improves, but not enough to affect the ranking of the classified classes under attack." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4072" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 69, + 504, + 517 + ], + "blocks": [ + { + "bbox": [ + 90, + 69, + 504, + 517 + ], + "lines": [ + { + "bbox": [ + 90, + 69, + 504, + 517 + ], + "spans": [ + { + "bbox": [ + 90, + 69, + 504, + 517 + ], + "type": "image", + "image_path": "468d7a738172efaae82413600154de03b76c286f7eebc1e91f303ccfbbeae23e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": "Figure 3. Successful and failed robustifications for " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ", for MNIST (" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_A = \\epsilon_A^C = \\epsilon_A^R = \\epsilon_D = 0.3" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ") and CIFAR10 (" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_A = 8.8 / 255" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C = 8 / 255" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R = 4 / 255" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_D = 32 / 255" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": "). Each triplet shows the original (left) and robustified (center) images; the rightmost panel is the defensive augmentation. All the " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": " recipes consistently increase the contrast and the high frequency content on MNIST, where failure cases are ambiguous even for a human observer. For CIFAR10, " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": " leverage isolated pixel changes, whereas " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 46, + 525, + 545, + 583 + ], + "type": "text", + "content": " consistently increase the color contrast and saturation, similarly to results in literature [17]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 594, + 545, + 628 + ], + "blocks": [ + { + "bbox": [ + 48, + 594, + 545, + 628 + ], + "lines": [ + { + "bbox": [ + 48, + 594, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 594, + 545, + 628 + ], + "type": "image", + "image_path": "5eecde7c36eb1cb17bf359374e103bf358884f2393484dc265a9e25f36e525f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "lines": [ + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "text", + "content": "Figure 4. The first, second and third row show respectively the robustified " + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w} + \\delta \\boldsymbol{w}_D)" + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "text", + "content": ", nominal " + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w})" + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "text", + "content": " and robustification " + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "inline_equation", + "content": "(\\delta \\boldsymbol{w}_D)" + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "text", + "content": " sets of 62 letters considered for testing " + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "inline_equation", + "content": "A^5 / PC" + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "text", + "content": ". The last row shows examples of the robustified letters after framing them, " + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} + \\delta \\boldsymbol{x}_D = A(\\boldsymbol{w} + \\delta \\boldsymbol{w}_D)" + }, + { + "bbox": [ + 46, + 635, + 547, + 658 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "type": "text", + "content": "smaller, equal or greater than " + }, + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "type": "inline_equation", + "content": "\\epsilon_A^C" + }, + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "type": "text", + "content": ". Results in Table 6 are consistent with those measured for MNIST: " + }, + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 679, + 287, + 704 + ], + "type": "text", + "content": " signifi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 679, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 679, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 306, + 679, + 545, + 704 + ], + "type": "text", + "content": "cantly improves the clean and certified error rates over the base CROWN-IBP classifier. Training " + }, + { + "bbox": [ + 306, + 679, + 545, + 704 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 306, + 679, + 545, + 704 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 306, + 679, + 545, + 704 + ], + "type": "inline_equation", + "content": "A^5 / R" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4073" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 251, + 133 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 251, + 133 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 251, + 133 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 251, + 133 + ], + "type": "table", + "html": "
AlgoError [certified error]
Vanilla0.89% [100.00%]
CROWN-IBP3.85% [13.85%]
A5/P3.08% [11.84%]
A5/PC0.73% [4.20%]
", + "image_path": "5cd25bd1a42ee0c2e3970a287b7eeaefe6cf55494066ff38969dfbbb89e4d3ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "text", + "content": "Table 9. Error and certified error for the classification of the 62 characters in Fig. 4, for " + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "inline_equation", + "content": "\\epsilon_A = 0.1" + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "text", + "content": " and different algorithms. " + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "inline_equation", + "content": "A^5 /P" + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "text", + "content": " changes the physical shape of the letters to achieve robustness after framing them; " + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "inline_equation", + "content": "A^5 /PC" + }, + { + "bbox": [ + 47, + 140, + 288, + 185 + ], + "type": "text", + "content": " also trains the associated classifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": "produces a smaller improvement, still sufficient to beat the CROWN-IBP classifier by a significant margin (Table 7). The value of " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " has a significant impact on the final result: if too large (" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R > \\epsilon_A" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": "), the clean accuracy is penalized, because " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " looks for very large plateaus in the classification landscape, that are hard or impossible to find. In other words, we observe for " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " a trade-off between the clean and certified accuracy, likewise adversarial training and certified methods. Best results on CIFAR10 are achieved again by " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": ", when " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon_A^R" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " is equal or slightly smaller than " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\epsilon_A = \\epsilon_A^C" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": ". The improvements in clean and certified error are in this case in the impressive range of " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " respectively (Table 8). We also notice here that the theoretical certified error rates measured with auto_LiRPA [37] in " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": " and those measured experimentally by autoattack [8, 9] are close for " + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 209, + 287, + 411 + ], + "type": "text", + "content": ", whose working point at convergence may favor the consistency between these two estimates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": "Also for CIFAR10 we perform a visual inspection of defensive augmentation (Fig. 3). " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " exploits a slight contrast increase and modifies isolated pixels: likewise adversarial attacks with small " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " norm (up to single pixel attacks [29]), the same seems possible when crafting defensive perturbations. Preemptive robustification [17] shows a similar pattern (see their Fig. 6): non natural textures and enhanced color saturation emerge in their images for large defensive perturbations. The robust images crafted by " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "A^5 / RC" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": ", however, are different: they generally show increased contrast and color saturation, without the textures generated by " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " and [17]. We speculate that these textures may be associated with the iterative optimization in " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "A^5 / O" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " and [17], that may both suffer some form of obfuscated gradient [2]; a more regular defense is instead generated by " + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 413, + 287, + 605 + ], + "type": "text", + "content": " without resorting to any iterative process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Additional results on FashionMNIST and Tinyimagenet, overall consistent with the ones presented here and showing the quantitative advantages provided by " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " on these datasets, are reported in the Supplementary. These experiments highlight two important facts. The first one is that the integration of the design philosophy of " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " with that of other recent ideas in the field of adversarial defense (like the " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "-dist neurons that implement 1-Lipschitz functions and are inherently robust to adversarial attacks) may lead to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "even bigger improvements in terms of certified robustness. The second is that the integration of different defense strategies may indeed be strictly needed to guarantee further progresses, as scaling to very large dataset remains problematic even for a preemptive robustification algorithms like " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": "Finally, for " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "A^5 / P" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "A^5 / PC" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": ", we test the classification of an alphabet of 62 characters (Fig. 4), after random rotation, shift, perspective distortion, noise addition, blurring and color jittering (details in Supplementary), that simulate the scanning of a document for OCR. Table 9 shows the clean and certified errors for: a vanilla, non robust classifier; a CROWN-IBP classifier; " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "A^5 / P" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "A^5 / PC" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": " achieves again a significant improvement both on clean and certified errors: " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "A^5 / PC" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": " has better (0.73%) accuracy of the vanilla " + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 133, + 546, + 277 + ], + "type": "text", + "content": " (0.89%) and can be attacked in only 4.20% of the cases. This results is obtained with a slight modification of the shape of the prototype characters shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 293, + 454, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 454, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 454, + 306 + ], + "type": "text", + "content": "5. Discussion and conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": "We introduce " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": ", the first framework that leverages adversarial augmentation to preemptively modify the input of a DNN or a physical object, to make it certifiable robust against adversarial attacks. " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": " is not simply complementary to other defense methods: its tights with them are strict and they can be used together to achieve higher robustness. For instance, one of the limitations we found is that " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "A^5 / R" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": " requires a somehow robust initial classifier " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": ". Co-training " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": " and a poorly robust " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": " easily compensates for the initially low robustness of " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": " and boosts its high clean accuracy while also achieving state of the art results in terms of certified accuracy; finding the optimal initial robustness of " + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 315, + 546, + 471 + ], + "type": "text", + "content": " remains an open question though." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": "Here we do not investigate all the possible " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": " recipes: in its most general implementation " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "(A^5 / PRC)" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": " the physical objects, " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": " could all be optimized at the same time. This opportunity can be explored through the released code at https://github.com/NVlabs/A5. This will also serve answering other questions that do not find space here: for instance, we do not know if robustification generalizes to other classifiers as adversarial attacks do, how " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": " works in norms other than " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "p = \\infty" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": ", and how much it can further improve if coupled with state-of-the-art solutions like the recently proposed " + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 304, + 472, + 545, + 604 + ], + "type": "text", + "content": "-dist neurons [41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Different " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "A^5" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " recipes find applications in scenarios where the user can control the acquisition device, equip it with " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " and guarantee its protection (e.g., image acquisition on a phone before server communication) or while designing the infrastructure (e.g., robust road signs creation); it may be not suitable for more general adversarial scenarios. Overall, we believe that the practical deployment of robust systems may benefit from (or even require) methods for both robust classifiers and preemptive robustification." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4074" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Maungmaung Aprilpyone and Hitoshi Kiya. Encryption inspired adversarial defense for visual classification. 2020 IEEE International Conference on Image Processing (ICIP), pages 1681-1685, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 212 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 274-283. PMLR, 10-15 Jul 2018. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 236 + ], + "type": "text", + "content": "[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "type": "text", + "content": "[4] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 4312-4321. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Survey Track. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 312, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 286, + 335 + ], + "type": "text", + "content": "[5] Tom Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gilmer. Adversarial patch. 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 390 + ], + "type": "text", + "content": "[6] Valeriia Cherepanova, Micah Goldblum, Harrison Foley, Shiyuan Duan, John P Dickerson, Gavin Taylor, and Tom Goldstein. Lowkey: Leveraging adversarial attacks to protect social media users from facial recognition. In International Conference on Learning Representations, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 467 + ], + "type": "text", + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 286, + 501 + ], + "type": "text", + "content": "[8] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "type": "text", + "content": "[9] Francesco Croce and Matthias Hein. Mind the box: " + }, + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 53, + 502, + 286, + 534 + ], + "type": "text", + "content": "-apgd for sparse adversarial attacks on image classifiers. In ICML, 2021. 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 567 + ], + "type": "text", + "content": "[10] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 601 + ], + "type": "text", + "content": "[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 286, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 657 + ], + "type": "text", + "content": "[12] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Mann, and Pushmeet Kohli. On the effectiveness of interval bound propagation for training verifiably robust models, 2018. 1, 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[13] Sonal Joshi, Jesús Villalba, Piotr Želasko, Laureano MoroVelázquez, and Najim Dehak. Study of pre-processing defenses against adversarial attacks on state-of-the-art speaker recognition systems. IEEE Transactions on Information Forensics and Security, 16:4811-4826, 2021. 1, 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[14] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "type": "text", + "content": "[15] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[16] Xiaoting Li, Lingwei Chen, Jinquan Zhang, James Larus, and Dinghao Wu. Watermarking-based defense against adversarial attacks on deep neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 172, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 216 + ], + "type": "text", + "content": "[17] Seungyong Moon, Gaon An, and Hyun Oh Song. Preemptive image robustification for protecting users against man-in-the-middle adversarial attacks. In AAAI Conference on Artificial Intelligence, 2022. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[18] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning (ICML), 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "text", + "content": "[19] Seong Joon Oh, Mario Fritz, and Bernt Schiele. Adversarial image perturbation for privacy protection a game theory perspective. pages 1491-1500, 10 2017. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 327 + ], + "type": "text", + "content": "[20] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 403 + ], + "type": "text", + "content": "[21] Rafael Pinot, Raphael Ettedgui, Geovani Rizk, Yann Chevaleyre, and Jamal Atif. Randomization matters how to defend against strong adversarial attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 7717-7727. PMLR, 13-18 Jul 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[22] Yao Qin, Nicholas Carlini, Ian J. Goodfellow, G. Cottrell, and Colin Raffel. Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. ArXiv, abs/1903.10346, 2019. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[23] Adnan Siraj Rakin, Zhezhi He, Boqing Gong, and Deliang Fan. Blind pre-processing: A robust defense method against adversarial examples. ArXiv, abs/1802.01549, 2018. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 547 + ], + "type": "text", + "content": "[24] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 624 + ], + "type": "text", + "content": "[25] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[26] Shawn Shan, Emily Wenger, Jiayun Zhang, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Fawkes: Protecting Privacy against Unauthorized Deep Learning Models. USENIX Association, USA, 2020. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[27] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervised learning. In International Conference on Learning Representations, 2021. 1, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4075" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[28] Jinmyeong Shin, Seok-Hwan Choi, and Yoon-Ho Choi. Is homomorphic encryption-based deep learning secure enough? Sensors, 21(23), 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[29] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5):828-841, 2019. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 207 + ], + "type": "text", + "content": "[30] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. Jan. 2014. 2nd International Conference on Learning Representations, ICLR 2014; Conference date: 14-04-2014 Through 16-04-2014. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 274 + ], + "type": "text", + "content": "[31] Florian Tramer, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1633-1645. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 330 + ], + "type": "text", + "content": "[32] Derui Wang, Chaoran Li, Sheng Wen, Surya Nepal, and Yang Xiang. Man-in-the-middle attacks against machine learning classifiers via malicious generative models. IEEE Transactions on Dependable and Secure Computing, 18(5):2074-2087, sep 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 332, + 286, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 286, + 386 + ], + "type": "text", + "content": "[33] Shiqi Wang, Huan Zhang, Kaidi Xu, Xue Lin, Suman Jana, Cho-Jui Hsieh, and J Zico Kolter. Beta-CROWN: Efficient bound propagation with per-neuron split constraints for complete and incomplete neural network verification. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 388, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 420 + ], + "type": "text", + "content": "[34] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms, 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 422, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 286, + 465 + ], + "type": "text", + "content": "[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In International Conference on Learning Representations, 2018. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 468, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 468, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 468, + 286, + 510 + ], + "type": "text", + "content": "[36] Kaidi Xu, Zhouxing Shi, Huan Zhang, Yihan Wang, Kai-Wei Chang, Minlie Huang, Bhavya Kailkhura, Xue Lin, and Cho-Jui Hsieh. Automatic perturbation analysis for scalable certified robustness and beyond. arXiv: Learning, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 513, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 513, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 513, + 286, + 567 + ], + "type": "text", + "content": "[37] Kaidi Xu, Huan Zhang, Shiqi Wang, Yihan Wang, Suman Jana, Xue Lin, and Cho-Jui Hsieh. Fast and complete: Enabling complete neural network verification with rapid and massively parallel incomplete verifiers. In International Conference on Learning Representations, 2021. 2, 3, 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 568, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 286, + 632 + ], + "type": "text", + "content": "[38] Hiromu Yakura and Jun Sakuma. Robust audio adversarial example for a physical attack. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5334–5341. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "type": "text", + "content": "[39] Jianfei Yang, Han Zou, and Lihua Xie. Robustsense: Defending adversarial attack for secure device-free human activity recognition. ArXiv, abs/2204.01560, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[40] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 283 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of Machine Learning Research, pages 12062-12072. PMLR, 18-24 Jul 2021. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 171 + ], + "type": "text", + "content": "[41] Bohang Zhang, Tianle Cai, Zhou Lu, Di He, and Liwei Wang. Towards certifying 1-infinity robustness using neural networks with 1-inf-dist neurons. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 12368-12379. PMLR, 18-24 Jul 2021. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "text", + "content": "[42] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane Boning, and Cho-Jui Hsieh. Towards stable and efficient training of verifiably robust neural networks. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 283 + ], + "type": "text", + "content": "[43] Huan Zhang, Tsui-Wei Weng, Pin-Yu Chen, Cho-Jui Hsieh, and Luca Daniel. Efficient neural network robustness certification with general activation functions. In Advances in Neural Information Processing Systems (NIPS), arXiv preprint arXiv:1811.00866, dec 2018. 1, 2, 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4076" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file